paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1706.07155 | 4 | 1706 | 2018-04-05T01:49:23 | Topological conjugacy of topological Markov shifts and Ruelle algebras | [
"math.OA",
"math.DS"
] | We will characterize topologically conjugate two-sided topological Markov shifts $(\bar{X}_A,\bar{\sigma}_A)$ in terms of the associated asymptotic Ruelle $C^*$-algebras ${\mathcal{R}}_A$ with its commutative $C^*$-subalgebras $C(\bar{X}_A)$ and the canonical circle actions. We will also show that extended Ruelle algebras ${\widetilde{\mathcal{R}}}_A$, which are purely infinite version of the asymptotic Ruelle algebras, with its commutative $C^*$-subalgebras $C(\bar{X}_A)$ and the canonical torus actions $\gamma^A$ are complete invariants for topological conjugacy of two-sided topological Markov shifts. We then have a computable topological conjugacy invariant, written in terms of the underlying matrix, of a two-sided topological Markov shift by using K-theory of the extended Ruelle algebra. The diagonal action of $\gamma^A$ has a unique KMS-state on ${\widetilde{\mathcal{R}}}_A$, which is an extension of the Parry measure on $\bar{X}_A$. | math.OA | math |
Topological conjugacy of topological Markov shifts and
Ruelle algebras
Kengo Matsumoto
Department of Mathematics
Joetsu University of Education
Joetsu, 943-8512, Japan
November 18, 2018
Abstract
We will characterize topologically conjugate two-sided topological Markov shifts
( ¯XA, ¯σA) in terms of the associated asymptotic Ruelle C ∗-algebras RA with its commu-
tative C ∗-subalgebras C( ¯XA) and the canonical circle actions. We will also show that
extended Ruelle algebras eRA, which are unital, purely infinite version of the asymp-
totic Ruelle algebras, with its commutative C ∗-subalgebras C( ¯XA) and the canonical
torus actions γ A are complete invariants for topological conjugacy of two-sided topo-
logical Markov shifts. We then have a computable topological conjugacy invariant,
written in terms of the underlying matrix, of a two-sided topological Markov shift
by using K-theory of the extended Ruelle algebra. The diagonal action of γ A has a
unique KMS-state on eRA, which is an extension of the Parry measure on ¯XA.
1
Introduction
A Smale space (X, φ) is a hyperbolic dynamical system having a local product structure (cf.
[2], [35]). A two-sided topological Markov shift ( ¯XA, ¯σA) gives a typical example of Smale
space. D. Ruelle in [33], [34] introduced C ∗-algebras from a Smale space (X, φ). After the
Ruelle's work, I. Putnam in [22], [23] has initiated to study structure of these C ∗-algebras
by using groupoid technique (for further studies, see [13], [24], [25], [26], [36], etc. ). For a
Smale space (X, φ), Putnam considered three kinds of C ∗-algebras S(X, φ), U (X, φ) and
A(X, φ) and their crossed products S(X, φ)⋊Z, U (X, φ)⋊Z and A(X, φ)⋊Z induced by the
original homeomorphisms φ, respectively. The algebras S(X, φ), U (X, φ) and A(X, φ) are
the C ∗-algebras of the groupoids of stable equivalence relation on X, unstable equivalence
relation on X and asymptotic equivalence relation on X, respectively.
I. Putnam has
pointed out that if the Smale space (X, φ) is a two-sided topological Markov shift ( ¯XA, ¯σA)
defined by an irreducible matrix A, the C ∗-algebras S(X, φ), U (X, φ) are isomorphic to
AF-algebras FA ⊗ K,FAt ⊗ K, and S(X, φ) ⋊ Z, U (X, φ) ⋊ Z are isomorphic to OA ⊗
K,OAt ⊗ K where OA,FA are the Cuntz -- Krieger algebra, the canonical AF-subalgebra of
OA, respectively for the matrix A, and FAt ,OAt are those ones for the transposed matrix
At of A, and K is the C ∗-algebra of compact operators on separable infinite dimensional
Hilbert space ℓ2(N).
1
In [18], the author has introduced notions of asymptotic continuous orbit equivalence
and asymptotic conjugacy in Smale spaces, and studied relationship with the crossed
product A(X, φ) ⋊ Z of the asymptotic Ruelle C ∗-algebra. In this paper, we will restrict
our interest in Smale spaces to two-sided topological Markov shifts. Let A be an N × N
irreducible non-permutation matrix with entries in {0, 1}. The shift space ¯XA of the
two-sided topological Markov shift ( ¯XA, ¯σA) is defined by the compact metric space of
bi-infinite sequences (xi)i∈Z satisfying A(xi, xi+1) = 1, i ∈ Z with shift transformation
¯σA((xi)i∈Z) = (xi+1)i∈Z, where metric d on ¯XA is defined by
d((xn)n∈Z, (yn)n∈Z) =
0
1
(λ0)k+1
if (xn)n∈Z = (yn)n∈Z,
if x0 6= y0,
if k = Max{n xi = yi for all i with i ≤ n}
for some fixed real number 0 < λ0 < 1. Let Ga
( ¯XA, ¯σA) defined by the asymptotic equivalence relation
A be the asymptotic ´etale groupoid for
A = {(x, z) ∈ ¯XA × ¯XA
Ga
lim
n→∞
d(¯σn
A(x), ¯σn
A(z)) = lim
n→∞
d(¯σ−n
A (x), ¯σ−n
A (z)) = 0}.
A with topology which makes the groupoid
A ´etale (see [22], [23]). For a general theory for ´etale groupoids, see [1], [27], [28], [29],
⋊ Z)
There are natural groupoid operations on Ga
Ga
etc. As in [23], the C ∗-algebra A( ¯X¯σA, ¯σA) ⋊ Z is realized as the C ∗-algebra C ∗(Ga
of the ´etale groupoid
A
Ga
A
⋊ Z = {(x, n, z) ∈ ¯XA × Z × ¯XA (¯σk
A(x), ¯σl
A(z)) ∈ Ga
A, n = k − l}.
The C ∗-algebra is denoted by RA and called the asymptotic Ruelle algebra in this paper.
Let dA : Ga
⋊ Z −→ Z be the groupoid homomorphism defined by dA(x, n, z) = n. As
A
⋊ Z is homeomorphic to ¯XA, the commutative C ∗-algebra
the unit space (Ga
A
C( ¯XA) is naturally regarded as a subalgebra of RA. As the algebra RA is a crossed
product A( ¯X¯σA , ¯σA) ⋊ Z, it has the dual action ρA
t of t ∈ T. In [18], an extended version
Gs,u
A
⋊ Z2 of the groupoid Ga
A
⋊ Z is introduced by setting
⋊ Z)◦ of Ga
A
⋊ Z2 = {(x, p, q, z) ∈ ¯XA × Z × Z × ¯XA (¯σp
A(x), z) ∈ Gs
A, (¯σq
A(x), z) ∈ Gu
A}
Gs,u
A
where
A ={(x, z) ∈ ¯XA × ¯XA
Gs
A ={(x, z) ∈ ¯XA × ¯XA
Gu
lim
n→∞
lim
n→∞
d(¯σn
A(x), ¯σn
A (x), ¯σ−n
A(z)) = 0},
A (z)) = 0}.
d(¯σ−n
⋊ Z2 with topology which makes Gs,u
A
There are natural groupoid operations on Gs,u
A
´etale (see [18]). Let cA : Gs,u
⋊ Z2 −→ Z2 be the groupoid homomorphism defined by
A
cA(x, p, q, z) = (p, q). The groupoid C ∗-algebra C ∗(Gs,u
⋊ Z2) is denoted by eRA which
A
was denoted by Rs,u
⋊ Z2)◦ of Gs,u
⋊ Z2 is {(x, 0, 0, x) ∈
A
Gs,u
⋊ Z2 x ∈ ¯XA}, which is regarded as ¯XA, the algebra eRA includes C( ¯XA) as a
A
subalgebra in natural way. There is a projection EA in the tensor product OAt ⊗OA such
that eRA is naturally isomorphic to EA(OAt ⊗ OA)EA. Hence the algebra eRA might be
regarded as a bilateral Cuntz -- Krieger algebra. The tensor product αAt
s of the gauge
A in [18]. Since the unit space (Gs,u
r ⊗ αA
⋊ Z2
A
2
t = γA
δA
actions αAt
r on OAt and αA
(t,t), t ∈ T is isomorphic to RA.
s on OA gives rise to an action γA
(r,s) of (r, s) ∈ T2 on eRA. It has
been shown in [18] that the fixed point algebra (eRA)δA
of eRA under the diagonal action
In [8] (cf. [6]), Cuntz -- Krieger proved that the stabilized Cuntz -- Krieger algebra OA ⊗
K with its diagonal C ∗-subalgebra DA ⊗ C of OA ⊗ K, where C denotes the maximal
abelian C ∗-subalgebra of K consisting of diagonal operators on ℓ2(N), and the stabilized
gauge action αA ⊗ id is invariant under topological conjugacy of the two-sided topological
Markov shift ( ¯XA, ¯σA) for irreducible non-permutation matrix A. T. M. Carlsen and
J. Rout have recently proved in [5] that the converse also holds even for more general
matrices without irreducibility and non-permutation. As a consequence, the stabilized
Cuntz -- Krieger algebra OA ⊗K with its diagonal C ∗-subalgebra DA ⊗C and the stabilized
gauge action αA ⊗ id is a complete invariant of the topological conjugacy of the two-sided
topological Markov shift. Inspired by this fact, we will in this paper show that the Ruelle
algebra RA with its subalgebra C( ¯XA) and the dual action ρA is a complete invariant
of the two-sided topological Markov shift ( ¯XA, ¯σA). We will also see that the C ∗-algebra
eRA with its subalgebra C( ¯XA) and the action γA of T2 is also a complete invariant of
topological conjugacy of ( ¯XA, ¯σA). We will show the following theorem.
Theorem 1.1 (Theorem 4.2). Let A, B be irreducible, non-permutation matrices with
entries in {0, 1}. The following six conditions are equivalent.
(i) Topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are topologically conjugate.
(ii) Topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are asymptotically conjugate.
(iii) There exists an isomorphism ϕ : Ga
A
dB ◦ ϕ = dA.
(iv) There exists an isomorphism ϕ : Gs,u
A
cB ◦ ϕ = cA.
⋊ Z −→ Ga
B
⋊ Z2 −→ Gs,u
B
⋊ Z of ´etale groupoids such that
⋊ Z2 of ´etale groupoids such that
t = ρB
t ◦ Φ for t ∈ T.
C( ¯XB) and Φ ◦ ρA
(v) There exists an isomorphism Φ : RA −→ RB of C ∗-algebras such that Φ(C( ¯XA)) =
(vi) There exists an isomorphism Φ : eRA −→ eRB of C ∗-algebras such that Φ(C( ¯XA)) =
C( ¯XB) and Φ ◦ γA
The equivalences among (ii), (iii) and (v) come from [18]. The main assertion is the
implication (ii) =⇒ (i) which will be proved in Theorem 3.3. Other implications will be
seen in the proof of Theorem 4.2, which are not tough tasks.
(r,s) ◦ Φ for (r, s) ∈ T2.
(r,s) = γB
UCT, its isomorphism class is completely determined by its K-theory date by a general
Since the algebra eRA is a unital, simple, purely infinite, nuclear C ∗-algebra satisfying
classification theorem ([15], [21], [31]). The K-groups K∗(eRA) are seen by the Kunneth
formulas and the universal coefficient theorem such that
3
K0(eRA) ∼= KK 1(OAt ,OA),
K1(eRA) ∼= KK(OAt,OA).
class [1 eRA
As a corollary of Theorem 1.1, we know that the group K0(eRA) and the position of the
of eRA in K0(eRA) is invariant under topological conjugacy of
] of the unit 1 eRA
( ¯XA, ¯σA). We see that (K0(eRA), [1 eRA
]) is isomorphic to (K0(OAt ⊗ OA), [EA]) and the
class [EA] of the projection EA actually lives in the group K0(OAt)⊗ K0(OA). We set the
vector ei = [0, . . . , 0,
1, 0, . . . , 0] ∈ ZN for i = 1, . . . , N . We have the following theorem.
i
Theorem 1.2 (Theorem 5.3). Suppose that A is an N × N irreducible, non-permutation
matrix with entries in {0, 1}. The position [EA] of the projection EA in K0(OAt )⊗K0(OA)
i=1[ei] ⊗ [ei] in the group ZN /(id − A)ZN ⊗ ZN /(id − At)ZN . Hence it is invariant
under topological conjugacy of the two-sided topological Markov shift ( ¯XA, ¯σA).
is PN
We put
eA =
NXi=1
[ei] ⊗ [ei]
in ZN /(id − A)ZN ⊗ ZN /(id − At)ZN .
i,j=1, B = [B(i, j)]M
We will actually see that the pair (ZN /(id − A)ZN ⊗ ZN /(id − At)ZN , eA) is a shift
equivalence invariant (Proposition 5.5, Proposition 5.8). We will present an example of
i,j=1 such that K0(OA) ∼= K0(OB), det(id − A) =
matrices A = [A(i, j)]N
det(id − B), but the invariants (ZN /(id − A)ZN ⊗ ZN /(id − At)ZN , eA) and (ZM /(id −
B)ZM⊗ZM /(id−Bt)ZM , eB) are different (Proposition 5.6). This shows that the invariant
(ZN /(id − A)ZN ⊗ ZN /(id − At)ZN , eA) is strictly stronger than the Bowen-Franks group
ZN /(id − A)ZN and not invariant under flow equivalence.
J. Cuntz in [7] studied the homotopy groups πn(End(OA ⊗K)) of the space End(OA ⊗
K) of endomorphisms of the C ∗-algebras OA ⊗ K. He proved that natural maps ǫn :
πn(End(OA⊗K)) −→ KK n(OA,OA) yield isomorphisms, and defined an element denoted
by ǫ1(λA) in Ext(OA) ⊗ K0(OA), where λA denotes the gauge action αA on OA. His
observation shows that the element ǫ1(λA) is noting but the above element eA under the
natural identification between Ext(OA)⊗ K0(OA) and ZN /(id− A)ZN ⊗ ZN /(id− At)ZN .
He already states in [7] that the position ǫ1(λA) in ZN /(id − A)ZN ⊗ ZN /(id − At)ZN is
invariant under topological conjugacy of the topological Markov shift ( ¯XA, ¯σA).
(t,t) on eRA, and
eRA for the action δA at the inverse temperature log γ exists if and only if γ is the Perron --
Frobenius eigenvalue β of A. The admitted KMS state is unique. The restriction of the
admitted KMS state to the subsalgebra C( ¯XA) is the state defined by the Parry measure
on ¯XA.
Theorem 1.3 (Theorem 6.14). Assume that the matrix A is aperiodic. A KMS state on
We will finally study that KMS states for the diagonal action δA
prove the following theorem.
t = γA
The Parry measure is the measure of maximal entropy (cf.
[37]). Since log β is the
topological entropy of the Markov shift ( ¯XA, ¯σA), the inverse temperature expresses the
entropy. This exactly corresponds to the result obtained by Enomoto -- Fujii -- Watatani in
[10] on KMS states for the gauge action on the Cuntz -- Krieger algebras OA.
set of positive integers.
Throughout the paper, we denote by Z+ the set of nonnegative integers and by N the
This paper is a continuation of the paper [18].
4
2 Preliminaries
We fix an irreducible, non-permutation matrix A = [A(i, j)]N
i,j=1 with entries in {0, 1}.
Let OA and OAt be the Cuntz -- Krieger algebras for the matrices A and its transpose
At, respectively. We may take generating partial isometries Si, i = 1, . . . , N of OA and
Ti, i = 1, . . . , M of OAt such that
SiS∗
i = 1,
S∗
i Si =
TiT ∗
i = 1,
T ∗
i Ti =
NXi=1
NXi=1
A(i, j)Sj S∗
j ,
At(i, j)Tj T ∗
j .
NXj=1
NXj=1
(2.1)
(2.2)
In the C ∗-algebra OAt ⊗OA of tensor product, let us denote by EA the projection defined
by
EA =
T ∗
i Ti ⊗ SiS∗
i .
(2.3)
NXi=1
the C ∗-algebra ([18])
By using the relations (2.1) and (2.2), it is easy to see that EA =PN
C ∗-algebra eRA is defined as the groupoid C ∗-algebra C ∗(Gs,u
The C ∗-algebra eRA was denoted by Rs,u
simple and purely infinite, and eRA ⊗K is isomorphic to OAt ⊗OA⊗K, the C ∗-algebra eRA
is simple and purely infinite if A is irreducible and non-permutation (cf. [26, Proposition
5.5]).
eRA = EA(OAt ⊗ OA)EA.
A in [18]. Since both the algebras OAt,OA are
i Si. The
⋊ Z2) which is realized as
i ⊗ S∗
i=1 TiT ∗
A
Let Bn( ¯XA) be the set of admissible words in ¯XA of length n. We set B∗( ¯XA) =
n=0Bn( ¯XA), where B0( ¯XA) denotes the empty word. For a word ξ = (ξ1, . . . , ξk), µ =
∪∞
(µ1, . . . , µm) ∈ B∗( ¯XA), we denote by ¯ξ = (ξk, . . . , ξ1) ∈ Bk( ¯XAt ) and set T ¯ξ = Tξk ··· Tξ1
and Sµ = Sµ1 ··· Sµm. Let αA, αAt
be the gauge actions of OA and OAt, respectively, which
are defined by
t (Si) = exp(√−1t)Si,
The fixed point algebras (OA)αA
are known to be AF-algebras, which are denoted by FA,FAt , respectively.
i = 1, . . . , N, t ∈ R/2πZ = T.
of OA,OAt under the gauge actions αA, αAt
t (Ti) = exp(√−1t)Ti,
αAt
, (OAt )αAt
We first note the following facts which were seen in [18].
αA
Proposition 2.1.
(i) The groupoid C ∗-algebra C ∗(Ga
A) of the groupoid Ga
A is isomorphic to the C ∗-subalgebra
of FAt ⊗ FA defined by
C ∗( T ¯ξT ∗
¯η ⊗ SµS∗
ν ∈ OAt ⊗ OA
µ = (µ1, . . . , µm), ν = (ν1, . . . , νn) ∈ B∗( ¯XA),
¯ξ = (ξk, . . . , ξ1), ¯η = (ηl, . . . , η1) ∈ B∗( ¯XAt),
A(ξk, µ1) = A(ηl, ν1) = 1, k = l, m = n ).
5
(ii) The C ∗-algebra RA is isomorphic to the C ∗-subalgebra of OAt ⊗ OA defined by
C ∗( T ¯ξT ∗
¯η ⊗ SµS∗
ν ∈ OAt ⊗ OA
µ = (µ1, . . . , µm), ν = (ν1, . . . , νn) ∈ B∗( ¯XA),
¯ξ = (ξk, . . . , ξ1), ¯η = (ηl, . . . , η1) ∈ B∗( ¯XAt),
A(ξk, µ1) = A(ηl, ν1) = 1, k − l = n − m ).
(iii) The C ∗-algebra eRA is isomorphic to the C ∗-subalgebra of OAt ⊗ OA defined by
C ∗( T ¯ξT ∗
ν ∈ OAt ⊗ OA
¯η ⊗ SµS∗
µ = (µ1, . . . , µm), ν = (ν1, . . . , νn) ∈ B∗( ¯XA),
¯ξ = (ξk, . . . , ξ1), ¯η = (ηl, . . . , η1) ∈ B∗( ¯XAt),
A(ξk, µ1) = A(ηl, ν1) = 1
).
We note that for i = 1, . . . , N the identity
Ti ⊗ S∗
i =
NXj,k=1
TiTjT ∗
j ⊗ SkS∗
kS∗
i =
NXj,k=1
A(j, i)A(i, k)Tij T ∗
j ⊗ SkS∗
ik
j ⊗ SkS∗
ik belongs to RA, we see that Ti ⊗ S∗
i and hence
holds. Since A(j, i)A(i, k)Tij T ∗
T ∗
i ⊗ Si belong to RA.
Define the diagonal action δA on eRA by setting
t ⊗ αA
t ,
t (EA) = EA, the automorphisms δA
t = αAt
δA
Since δA
t ∈ R/2πZ = T.
t , t ∈ T define an action of T on eRA. For
ν = (ν1, . . . , νn) ∈ B∗( ¯XA),
¯η = (ηl, . . . , η1) ∈ B∗( ¯XAt )
µ = (µ1, . . . , µm),
¯ξ = (ξk, . . . , ξ1),
satisfying A(ξk, µ1) = A(ηl, ν1) = 1, we see that
δA
t (T ¯ξT ∗
¯η ⊗ SµS∗
ν ) = exp(√−1(k − l + m − n)t)T ¯ξT ∗
¯η ⊗ SµS∗
ν
so that the following lemma holds.
Hence we have
¯η ⊗ SµS∗
the asymptotic Ruelle algebra RA.
Lemma 2.2. Keep the above notation. The element T ¯ξT ∗
if and only if k − l = n − m.
ν in eRA belongs to RA
Proposition 2.3 ([18, Theorem 9.6]). The fixed point algebra (eRA)δA
of eRA under δA is
As in [18, Lemma 9.5], the C ∗-subalgebra of C ∗(Ga
¯ξ ⊗
µ, ¯ξ = (ξk, . . . , ξ1) ∈ B∗( ¯XAt), µ = (µ1, . . . , µm) ∈ B∗( ¯XA) with A(ξk, µ1) = 1 is
SµS∗
canonically isomorphic to the commutative C ∗-algebra C( ¯XA) of continuous functions on
¯XA. In what follows, we identify the subalgebra with the algebra C( ¯XA) so that C( ¯XA)
is a C ∗-subalgebra of RA and eRA.
A) generated by elements T ¯ξT ∗
6
3 Asymptotic conjugacy and topological conjugacy
For x = (xn)n∈Z ∈ ¯XA, we set x+ = (xn)∞
n=0. Let us denote by XA
the compact Hausdorff space of right infinite sequences (xi)i∈Z+ ∈ {1, . . . , N}Z+ satisfying
A(xi, xi+1) = 1, i ∈ Z+. The right one-sided topological Markov shift (XA, σA) is defined
by a topological dynamical system of shift transformation σA((xi)i∈Z+ ) = (xi+1)i∈Z+ on
XA. For x = (xi)i∈Z+ ∈ XA and k ∈ Z+, we set x[k,∞) = σk
the notion for topological Markov shifts and rephrase it in the following way.
In [18], a notion of asymptotic conjugacy in Smale spaces were introduced. We apply
A(x) = (xk, xk+1, . . . ) ∈ XA.
n=0 and x− = (x−n)∞
Definition 3.1 ([18]). Two topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are said to
be asymptotically conjugate if there exists a homeomorphism h : ¯XA −→ ¯XB satisfying
the following three conditions
(i) There exists a nonnegative integer K ∈ Z+ such that
¯σK+1
B
¯σ−K+1
B
¯σK+1
A
¯σ−K+1
A
(h(x))+ = ¯σK
(h(x))− = ¯σ−K
(h−1(y))+ = ¯σK
(h−1(y))− = ¯σ−K
B (h(¯σA(x)))+
B (h(¯σA(x)))−
A (h−1(¯σB(y)))+
A (h−1(¯σB(y)))−
for x ∈ ¯XA,
for x ∈ ¯XA,
for y ∈ ¯XB,
for y ∈ ¯XB.
(3.1)
(3.2)
(3.3)
(3.4)
(ii) There exists a continuous function m1 : Ga
¯σm1(x,z)
B
¯σ−m1(x,z)
B
(h(x))+ = ¯σm1(x,z)
(h(x))− = ¯σ−m1(x,z)
B
B
(h(z))−
A −→ Z+ such that
(h(z))+
for (x, z) ∈ Ga
A,
for (x, z) ∈ Ga
A.
(iii) There exists a continuous function m2 : Ga
¯σm2(y,w)
A
¯σ−m2(y,w)
A
(h−1(y))+ = ¯σm2(y,w)
(h−1(y))− = ¯σ−m2(y,w)
A
A
B −→ Z+ such that
(h−1(w))+
(h−1(w))−
for (y, w) ∈ Ga
B,
for (y, w) ∈ Ga
B.
Let A = [A(i, j)]N
i,j=1, B = [B(i, j)]M
The following proposition is key in this section.
i,j=1 be irreducible matrices with entries in {0, 1}.
Proposition 3.2. If the topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are asymptoti-
cally conjugate, then they are topologically conjugate.
Proof. Let h : ¯XA −→ ¯XB be a homeomorphism and K ∈ Z+ a nonnegative integer
satisfying (3.1), (3.2), (3.3), (3.4). We define two continuous maps h+ : ¯XA −→ XB and
+ : ¯XB −→ XA by setting
h−1
h+(x) = ¯σK
h−1
+ (y) = ¯σK
B (h(x))+,
A (h−1(y))+,
x ∈ ¯XA,
y ∈ ¯XB.
7
It then follows that by (3.1),
h+(¯σA(x)) = ¯σK
B (h(¯σA(x)))+
B
= ¯σK+1
= [¯σK+1
(h(x))+
(h(x))][0,∞)
= [¯σB(h(x))][K,∞).
B
On the other hand,
σB(h+(x)) = σB([¯σK
B (h(x))][0,∞))
= [¯σK
B (h(x))][1,∞)
= [h(x)][K+1,∞)
= [¯σB(h(x))][K,∞).
Therefore we have
h+(¯σA(x)) = σB(h+(x))
for x ∈ ¯XA.
Hence the continuous map h+ : ¯XA −→ XB is a sliding block code (cf. [16]) so that there
exists a block map Φ : Bm+n+1( ¯XA) −→ {1, 2, . . . , M} for some m, n ∈ Z+ such that
h+((xi)i∈Z) = Φ([xi−m, . . . , xi+n])i∈Z+
for x = (xi)i∈Z ∈ ¯XA.
+ : ¯XB −→ XA satisfies h−1
+ (¯σB(y)) =
+ (y)) for y ∈ ¯XB so that there exists a block map Ψ : Bm′+n′+1( ¯XB) −→ {1, 2, . . . , N}
Similarly we know that the continuous map h−1
σA(h−1
for some m′, n′ ∈ Z+ such that
h−1
+ ((yi)i∈Z) = Ψ([yi−m′, . . . , yi+n′])i∈Z+
for y = (yi)i∈Z ∈ ¯XB.
By using these block maps Φ : Bm+n+1( ¯XA) −→ {1, 2, . . . , M} and Ψ : Bm′+n′+1( ¯XB) −→
{1, 2, . . . , N}, we define two sliding block codes Φ∞ : ¯XA −→ ¯XB and Ψ∞ : ¯XB −→ ¯XA
by setting
Φ∞((xi)i∈Z) = Φ([xi−m, . . . , xi+n])i∈Z ∈ ¯XB
Ψ∞((yi)i∈Z) = Ψ([yi−m′, . . . , yi+n′])i∈Z ∈ ¯XA
for x = (xi)i∈Z ∈ ¯XA,
for y = (yi)i∈Z ∈ ¯XB.
We note that
Φ∞((xi)i∈Z)+ = h+((xi)i∈Z) ∈ XB
Ψ∞((yi)i∈Z)+ = h−1
+ ((yi)i∈Z) ∈ XA
for x = (xi)i∈Z ∈ ¯XA,
for y = (yi)i∈Z ∈ ¯XB.
For y = (yi)i∈Z ∈ ¯XB, we have
[Ψ∞(y)][K,∞) =[¯σK
A (Ψ∞(y))][0,∞)
=[Ψ∞(¯σK
B (y))][0,∞)
=h−1
+ (¯σK
B (y))
A (h−1(¯σK
=[¯σK
=[h−1(¯σK
B (y))][K,∞).
B (y)))][0,∞)
8
As Φ∞ is a sliding block code with memory m, the condition [Ψ∞(y)][K,∞) = [h−1(¯σK
implies
B (y)))][K,∞)
[Φ∞(Ψ∞(y))][K+m,∞) = [Φ∞(h−1(¯σK
B (y)))][K+m,∞).
It then follows that
[Φ∞(Ψ∞(y))][K+m,∞) =[Φ∞(h−1(¯σK
=[h+(h−1(¯σK
=[(¯σK
=[¯σK
=[¯σ2K
B ◦ h)(h−1(¯σK
B (¯σK
B (y)][K+m,∞)
B (y))][K+m,∞)
B (y)))][K+m,∞)
B (y)))][K+m,∞)
B (y)))][K+m,∞)
so that
[Φ∞(Ψ∞(y))][K+m,∞) = [¯σ2K
B (y)][K+m,∞)
for y ∈ ¯XB.
Since Φ∞ ◦ Ψ∞ is a sliding block code, we obtain that
Φ∞ ◦ Ψ∞ = ¯σ2K
B .
Hence Φ∞ is surjective. Similarly we know that Ψ∞ ◦ Φ∞ = ¯σ2K
Therefore we have a topological conjugacy Φ∞ : ¯XA −→ ¯XB.
A so that Φ∞ is injective.
We remark that the above proof needs only the equalities (3.1) and (3.3).
We thus conclude the following.
Theorem 3.3. Two topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are asymptotically
conjugate if and only if they are topologically conjugate.
Proof. It is direct to see that topological conjugacy implies asymptotic conjugacy. Hence
the assertion follows from the preceding proposition.
4 Conjugacy, groupoid isomorphism and C∗-algebras
We consider the groupoid Gs,u
A
γA of T2 on eRA = EA(OAt ⊗ OA)EA is defined by setting
⋊ Z2 and its C ∗-algebra written eRA. Recall that an action
r ⊗ αA
on OAt ⊗ OA for (r, s) ∈ T2.
(r,s) = αAt
γA
s
Since γA
(r,s)(EA) = EA, we have an action γA of T2 on eRA, which defines two kinds of
actions of T on eRA such that
δA
t = γA
(t,t)
and
ρA
t = γA
(− t
2 , t
2 )
for t ∈ T.
We regard the groupoid C ∗-algebra C ∗(Ga
A
in a natural way. Let us denote by σA the dual action on C ∗(Ga
⋊ Z) as the C ∗-crossed product C ∗(Ga
A) ⋊ Z
A) ⋊ Z. In the following
lemma, the C ∗-algebra eRA is regarded as a C ∗-subalgebra of OAt ⊗ OA as in Proposition
2.1 (ii).
9
Lemma 4.1. There exists an isomorphism Ψ : C ∗(Ga
A) ⋊ Z −→ RA such that
t = ρA
t ∈ T.
i ⊗ Si. As in [18, Proposition
9.9], Ad(UA) corresponds to the shift operation on C( ¯XA). Since
Ψ(C( ¯XA)) = C( ¯XA)
and Ψ ◦ σA
Proof. Let UA be the unitary in RA defined by UA =PN
NXi=1
(Si) = exp(√−1t)
ρA
t (UA) =
i ) ⊗ α t
NXi=1
t ◦ Ψ,
i=1 T ∗
α− t
(T ∗
2
2
T ∗
i ⊗ Si = exp(√−1t)UA,
we have the assertion.
We have the following main result of the paper.
Theorem 4.2. The following six conditions are equivalent.
(i) Topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are topologically conjugate.
(ii) Topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are asymptotically conjugate.
(iii) There exists an isomorphism ϕ : Ga
A
dB ◦ ϕ = dA.
(iv) There exists an isomorphism ϕ : Gs,u
A
cB ◦ ϕ = cA.
⋊ Z −→ Ga
B
⋊ Z2 −→ Gs,u
B
⋊ Z of ´etale groupoids such that
⋊ Z2 of ´etale groupoids such that
C( ¯XB) and Φ ◦ ρA
(v) There exists an isomorphism Φ : RA −→ RB of C ∗-algebras such that Φ(C( ¯XA)) =
(vi) There exists an isomorphism Φ : eRA −→ eRB of C ∗-algebras such that Φ(C( ¯XA)) =
(r,s) ◦ Φ for (r, s) ∈ T2.
C( ¯XB) and Φ ◦ γA
t ◦ Φ for t ∈ T.
(r,s) = γB
t = ρB
Proof. The equivalence between (i) and (ii) is proved in Theorem 3.3.
The equivalences among (ii), (iii) and (v) are shown in [18].
We will prove the three implications (i) =⇒ (iv), (iv) =⇒ (vi), (vi) =⇒ (v).
(i) =⇒ (iv): Suppose that there exists a topological conjugacy h : ¯XA −→ ¯XB so
that h ◦ ¯σA = ¯σB ◦ h. For (x, p, q, z) ∈ Gs,u
A(x), z) ∈ Gs
A and
(¯σq
A(x), z) ∈ Gu
B(h(x)), h(z)) ∈ Gs
B, so that we have
(h(x), p, q, h(z)) ∈ Gs,u
⋊ Z2. It is routine to show that the correspondence
⋊ Z2, the conditions (¯σp
A(h(x)), h(z)) ∈ Gu
A imply (¯σp
B and (¯σq
B
A
ϕ : (x, p, q, z) ∈ Gs,u
A
⋊ Z2 −→ (h(x), p, q, h(z)) ∈ Gs,u
B
⋊ Z2
yields an isomorphism of ´etale groupoids. It is then clear that cB ◦ ϕ = cA. This shows
the condition (iv).
⋊ Z2 of
⋊ Z2
(iv) =⇒ (vi): Suppose that there exists an isomorphism ϕ : Gs,u
´etale groupoids such that cB ◦ ϕ = cA. Since the both groupoids Gs,u
are amenable and ´etale by [18, Proposition 7.2 and Lemma 7.3], the C ∗-algebras eRA
and eRB are represented on the Hilbert C ∗-modules ℓ2(Gs,u
⋊ Z2),
⋊ Z2 is an isomorphism of ´etale groupoids,
respectively as in [18]. As ϕ : Gs,u
A
⋊ Z2 −→ Gs,u
⋊ Z2 and Gs,u
B
⋊ Z2) and ℓ2(Gs,u
B
⋊ Z2 −→ Gs,u
B
B
A
A
A
10
there exist a homeomorphism h : ¯XA −→ ¯XB and a continuous groupoid homomorphism
c : Gs,u
A
⋊ Z2 −→ Z2 such that
ϕ(x, p, q, z) = (h(x), c(x, p, q, z), h(z)),
(x, p, q, z) ∈ Gs,u
A
⋊ Z2.
The condition cB ◦ ϕ = cA forces us to hold the equality c(x, p, q, z) = (p, q) so that we
have
ϕ(x, p, q, z) = (h(x), p, q, h(z)),
(x, p, q, z) ∈ Gs,u
A
⋊ Z2.
Let us consider the unitaries Vh : ℓ2(Gs,u
B
Z2) −→ ℓ2(Gs,u
⋊ Z2) by setting
B
⋊ Z2) −→ ℓ2(Gs,u
A
⋊ Z2) and Vh−1 : ℓ2(Gs,u
A
⋊
[Vhζ](x, p, q, z) =ζ(h(x), p, q, h(z)) for ζ ∈ ℓ2(Gs,u
B
⋊ Z2), (x, p, q, z) ∈ Gs,u
A
[Vh−1ξ](y, m, n, w) =ξ(h−1(y), m, n, h−1(w)) for ξ ∈ ℓ2(Gs,u
⋊ Z2)) = Cc(Gs,u
Put Φ = Ad(Vh) which satisfies Φ(Cc(Gs,u
B
A
Since ¯XA, ¯XB are identified with the unit spaces
A
⋊ Z2,
⋊ Z2), (y, m, n, w) ∈ Gs,u
⋊ Z2) so that Φ(eRA) = eRB.
B
⋊ Z2.
(Gs,u
A
(Gs,u
B
⋊ Z2)◦ ={(x, 0, 0, x) ∈ Gs,u
⋊ Z2)◦ ={(y, 0, 0, y) ∈ Gs,u
B
A
⋊ Z2 x ∈ ¯XA},
⋊ Z2 y ∈ ¯XB},
(r,s) = γB
respectively, we easily knows that Φ(C( ¯XA)) = C( ¯XB). It is also direct to see that the
(r,s) ◦ Φ for (r, s) ∈ T2 holds, because of the equality cB ◦ ϕ = cA.
identity Φ ◦ γA
(vi) =⇒ (v): Suppose that there exists an isomorphism Φ : eRA −→ eRB of C ∗-algebras
such that Φ(C( ¯XA)) = C( ¯XB) and Φ◦ γA
(r,s) ◦ Φ for (r, s) ∈ T2. As the action δA
t =
(t,t) of t ∈ T act on eRA and its fixed point algebra (eRA)δA
γA
is RA. Let us denote by Φ the
restriction of Φ to the fixed point algebra RA. It induces an isomorphism Φ : RA −→ RB.
Then it is clear that the action ρA
t ◦ Φ. This shows
the condition (v).
on RA satisfies Φ ◦ ρA
(r,s) = γB
t = γA
t = ρB
2 , t
2 )
(− t
5 K-theoretic invariants
ant under topological conjugacy of two-sided topological Markov shifts. Concerning the
asymptotic Ruelle algebra RA, its K-group formula has been obtained by Putnam [22,
p.129] (cf.
By using Theorem 4.2, the isomorphism classes of the C ∗-algebras RA and eRA are invari-
[11], [14]). We focus on studying the K-group K0(eRA) of the latter algebra
eRA. Under the assumption that the matrix A is irreducible and non-permutation, the
algebra eRA is a unital, simple, purely infinite, nuclear C ∗-algebra satisfying UCT, so that
its isomorphism class is completely determined by its K-theory date by a general classifi-
cation theory of Kirchberg ([15]) and Phillips ([21]). Hence the following is a corollary of
Theorem 4.2.
Proposition 5.1. The pair (K0(eRA), [1 eRA
]) of the K0-group of eRA and the position of the
of eRA in K0(eRA) is invariant under topological conjugacy of two-sided topological
unit 1 eRA
Markov shift ( ¯XA, ¯σA).
11
Recall that the projection EA is defined in (2.3). We have
Proposition 5.2. There exists an isomorphism Φ : eRA ⊗ K −→ OAt ⊗ OA ⊗ K such that
the induced isomorphism Φ∗ : K0(eRA) −→ K0(OAt ⊗ OA) satisfies Φ∗([1 eRA
Proof. Since the C ∗-algebra OAt ⊗ OA is unital and simple, the projection EA in (2.3) is
a full projection in OAt ⊗ OA, Brown's theorem [4] tells us that there exists an isometry
vA in the multiplier algebra M (OAt ⊗ OA ⊗ K) of OAt ⊗ OA ⊗ K such that v∗
AvA = 1 and
vAv∗
A).
Let p0 be a rank one projection in K. We then have
A = EA ⊗ 1. Define an isomorphism Φ : eRA ⊗ K −→ OAt ⊗ OA ⊗ K by Φ = Ad(v∗
]) = [EA].
]) = Φ∗([EA ⊗ p0]) = [v∗
A(EA ⊗ p0)vA] = [EA ⊗ p0] = [EA]
Φ∗([1 eRA
in K0(OAt ⊗ OA).
Hence the position [EA] in K0(OAt ⊗OA) as well as the group K0(OAt ⊗OA) is invari-
ant under topological conjugacy of topological Markov shift ( ¯XA, ¯σA). By the Kunneth
formulas [32] of the K-groups of the tensor product C ∗-algebras, we know that
K0(OAt ⊗ OA) ∼= (K0(OAt) ⊗ K0(OA)) ⊕ (K1(OAt ) ⊗ K1(OA)),
K1(OAt ⊗ OA) ∼= (K0(OAt ) ⊗ K1(OA)) ⊕ (K1(OAt) ⊗ K0(OA)) ⊕ TorZ
1 (K0(OAt), K0(OA)).
By the universal coefficient theorem for KK-groups, the K-group Ki(OAt ⊗OA) is isomor-
phic to the KK-group KK i+1(OAt ⊗ OA) for i = 0, 1, we see that
K0(eRA) ∼= KK 1(OAt ,OA),
K1(eRA) ∼= KK(OAt,OA).
Since K0(OAt ) is isomorphic to K0(OA) and K1(OA) is the torsion free part of K0(OA),
the groups Ki(OAt ⊗ OA), i = 0, 1 do not have any further information than the group
K0(OA) by the above Kunneth formulas. As K0(OA) = ZN /(id − At)ZN , it is a direct
sum Zn ⊕ TA of its torsion free part Zn and its torsion part TA = Z/m1Z ⊕ ··· ⊕ Z/mkZ,
where mimi+1 with mi ≥ 2, i = 1, . . . , k − 1. It is easy to see that
ZN /(id − A)ZN ⊗ ZN /(id − At)ZN
∼=Zn2
⊕ (TA)n ⊕ (TA)n ⊕ (TA ⊗ TA)
∼=Zn2
⊕ (Z/m1Z)2n+2k−1 ⊕ (Z/m2Z)2n+2k−3 ⊕ ··· ⊕ (Z/mkZ)2n+1.
Hence the groups Ki(OAt ⊗ OA), i = 0, 1 give us the same information as the group
K0(OA).
The position [EA] in K0(OAt ⊗OA) however gives us more information than the group
In the above Kunneth formula for K0(OAt ⊗ OA), the element [EA] lives in
K0(OA).
i ] by definition of EA. Therefore the
position [EA] of the projection EA in K0(OAt) ⊗ K0(OA) is invariant under topological
i
conjugacy of ( ¯XA, ¯σA). We set the vector ei = [0, . . . , 0,
1, 0, . . . , 0] for i = 1, . . . , N . We
rephrase the above fact with the following theorem.
K0(OAt)⊗ K0(OA) as the elementPN
i Ti]⊗ [SiS∗
i=1[T ∗
Theorem 5.3. The position [EA] of the projection EA in K0(OAt)⊗ K0(OA) is invariant
under topological conjugacy of ( ¯XA, ¯σA). Hence the position PN
i=1[ei] ⊗ [ei] in the group
ZN /(id − A)ZN ⊗ ZN /(id − At)ZN is invariant under topological conjugacy of ( ¯XA, ¯σA).
12
Proof. We give its precise proof by matrix method. Let A = [A(i, j)]N
i,j=1
be irreducible non-permutation matrices such that the two-sided topological Markov shifts
( ¯XA, ¯σA) and ( ¯XA, ¯σA) are topological conjugate. By William's theorem [38], the matrices
A, B are strong shift equivalent, and hence we may assume that there exists two rectangular
nonnegative integer matrices C, D such that A = CD, B = DC. By [17, Theorem 4.6],
there exists an isomorphism Φ : OA ⊗K −→ OB ⊗K of C ∗-algebras such that the diagram
i,j=1, B = [B(i, j)]M
K0(OA)
Φ∗−−−−→
K0(OB)
ZN /(id − At)ZN
−−−−→ ZM /(id − Bt)ZM
mCt
commutes, where mC t is the isomorphism induced by multiplying the matrix C t from the
left and ǫA : K0(OA) → ZN /(id − At)ZN is an isomorphism defined by ǫA([SiS∗
i ]) = [ei]
the class of the vector ei in ZN . Since the identities At = DtC t, Bt = C tDt also hold, we
similarly have an isomorphism Φt : OAt ⊗ K −→ OBt ⊗ K of C ∗-algebras such that the
diagram
yǫB
K0(OBt )
yǫBt
ǫAy
K0(OAt )
ǫAty
Φt
∗−−−−→
Φt
∗⊗Φ∗
−−−−→
mD ⊗mCt
commutes. We then have a commutative diagram:
ZN /(id − A)ZN mD−−−−→ ZM /(id − B)ZM
ZN /(id − A)ZN ⊗ ZN /(id − At)ZN
−−−−−−→ ZM /(id − B)ZM ⊗ ZM /(id − Bt)ZM .
K0(OBt ) ⊗ K0(OB)
yǫBt ⊗ǫB
K0(OAt ) ⊗ K0(OA)
ǫAt ⊗ǫAy
We note that
NXi=1
ǫAt([T ∗
i Ti]) ⊗ ǫA([SiS∗
i ]) =
=
and set the specific element as
NXi=1
NXi=1
ǫAt([TiT ∗
i ]) ⊗ ǫA([SiS∗
i ])
[ei] ⊗ [ei]
in ZN /(id − A)ZN ⊗ ZN /(id − At)ZN ,
eA =
NXi=1
[ei] ⊗ [ei]
in ZN /(id − A)ZN ⊗ ZN /(id − At)ZN .
(5.1)
We will show that (mD ⊗ mC t)(eA) = eB. In the computation below, the vectors ei, and
fj denote the N × 1 matrix in ZN whose ith component is one and zero elsewhere, and
the M × 1 matrix in ZM whose jth component is one and zero elsewhere, respectively. We
13
Dei ⊗ C tei
D(1, i)
D(2, i)
have
NXi=1
NXi=1
=
=
NXi=1
=
MXj=1
Hence we have
so that
MXj=1
...
...
NXi=1
D(M, i)
C(i, M )
D(M, i)
⊗
...
⊗
...
C(i, 1)
C(i, 2)
D(1, i)
D(2, i)
...
...
PN
PN
MXj=1
PN
=
NXi=1
⊗ fj =
MXj=1
Bfj ⊗ fj.
D(1, i)C(i, j)
D(2, i)C(i, j)
D(M, i)C(i, j)
B(1, j)
B(2, j)
B(N, j)
⊗ fj =
C(i, j)fj
MXj=1
i=1 D(1, i)C(i, j)
i=1 D(2, i)C(i, j)
i=1 D(N, i)C(i, j)
⊗ fj
(5.2)
Dei ⊗ C tei −
(B − id)fj ⊗ fj
fj ⊗ fj =
MXj=1
MXj=1
NXi=1
[Dei] ⊗ [C tei] =
MXj=1
[fj] ⊗ [fj] = eB
(mD ⊗ mC t)(eA) =
thus proving the theorem.
Remark 5.4.
(i) The pair (ZN /(id − A)ZN ⊗ ZN /(id − At)ZN , eA) is a complete invariant for the
isomorphism class of the C ∗-algebra eRA, because the group structure of ZN /(id −
A)ZN ⊗ ZN /(id − At)ZN determines the groups Ki(OA), Ki(OAt ), i = 0, 1 and also
the pair determines the position [EA] in K0(OAt ⊗ OA). Hence by Proposition
5.2, the pair (K0(eRA), [EA]) and the group K1(eRA) are determined by the pair
(ZN /(id − A)ZN ⊗ ZN /(id − At)ZN , eA).
(ii) Since the projection EA is regarded as an element of the C ∗-algebra FAt ⊗ FA such
that C ∗(Ga
A) = EA(FAt ⊗ FA)EA, we have another topological conjugacy invariant
(K0(FAt ) ⊗ K0(FA), [EA]) the position [EA] in the group K0(FAt ) ⊗ K0(FA). We
will discuss this kind of invariants in [19].
(iii) J. Cuntz in [7] studied the homotopy groups πn(End(OA ⊗ K)) of the space of
endomorphisms End(OA ⊗ K) of the C ∗-algebras OA ⊗ K. He proved that natural
maps ǫn : πn(End(OA ⊗ K)) −→ KK n(OA,OA) yield isomorphisms, and defined an
element denoted by ǫ1(λA) in Ext(OA)⊗K0(OA), where λA denotes the gauge action
αA on OA. By the Kaminker -- Putnam's K-theoretic duality between Ext(OA) and
14
K0(OAt ) ([12]), the element ǫ1(λA) is regarded as an element in K0(OAt)⊗ K0(OA).
Cuntz's observation in [7] shows that the element ǫ1(λA) is noting but the above
element eA under the identification between Ext(OA)⊗K0(OA) and ZN /(id−A)ZN⊗
ZN /(id−At)ZN . He already states in [7] that the position ǫ1(λA) in ZN /(id−A)ZN ⊗
ZN /(id − At)ZN is invariant under topological conjugacy of the topological Markov
shift ( ¯XA, ¯σA).
In [38], Williams introduced an equivalence relation in matrices called shift equiva-
lence. It is weaker than strong shift equivalence. The shift equivalence relation has been
playing crucial role in the classification theory of symbolic dynamical systems (cf.
[16]).
Two matrices A, B are said to be shift equivalent if there exist a positive integer ℓ and
rectangular nonnegative integer matrices R, S such that
AR = RB,
SA = BS,
Aℓ = RS,
Bℓ = SR.
(5.3)
In the proof of the above theorem, we notice that the following proposition holds.
Proposition 5.5. The pair (ZN /(id − A)ZN ⊗ ZN /(id − At)ZN , eA) is invariant under
shift equivalence.
i,j=1, B = [B(i, j)]M
Proof. Suppose that matrices A = [A(i, j)]N
i,j=1 are shift equivalent.
Let ℓ be a positive integer and R, S rectangular nonnegative integer matrices satisfying
(5.3). Then the map mS : ZN /(id − A)ZN −→ ZM /(id − B)ZM defined by the left
multiplication of the matrix S yields an isomorphism of the abelian groups. We similarly
see that mRt : ZN /(id − At)ZN −→ ZM /(id − Bt)ZB defined by the left multiplication
of the matrix Rt yields an isomorphism of the abelian groups. A similar computation
proving the equality (5.2) in the proof of the preceding theorem shows that the equality
NXi=1
Sei ⊗ Rtei −
MXj=1
fj ⊗ fj =
MXj=1
(SR − id)fj ⊗ fj =
MXj=1
(Bℓ − id)fj ⊗ fj
holds. As Bℓ − id = (B − id)(Bℓ−1 + ··· + B + id), we know that (mS ⊗ mRt)(eA) = eB
so that the map
mS ⊗ mRt : ZN /(id − A)ZN ⊗ ZN /(id − At)ZN −→ ZM /(id − B)ZM ⊗ ZM /(id − Bt)ZM
gives rise to an isomorphism between (ZN /(id−A)ZN⊗ZN /(id−At)ZN , eA) and (ZM /(id−
B)ZM ⊗ ZM /(id − Bt)ZM , eB).
We will present an example showing that the invariant in the group (ZN /(id− A)ZN ⊗
ZN /(id − At)ZN , eA) is strictly finer than the K-group K0(OA). We note that Enomoto --
Fujii -- Watatani in [9] listed a complete classification table of Cuntz -- Krieger algebras OA
in terms of its K-groups for which its sizes of matrices are three.
Let A =
1 1 1
1 1 1
1 1 1
l
m
n
=
−m − n
−l − n
−l − m
. Since (id − A)
ϕ :
, the map
∈ Z3 −→ [a + b + c] ∈ Z/2Z
a
b
c
15
induces an isomorphism ¯ϕ : Z3/(id − A)Z3 −→ Z/2Z. Hence we have an isomorphism
ϕ := ¯ϕ ⊗ ¯ϕ : Z3/(id − A)Z3 ⊗ Z3/(id − At)Z3 −→ Z/2Z ⊗ Z/2Z ∼= Z/2Z.
Since ϕ(ei ⊗ ei) = ¯ϕ(ei) ⊗ ¯ϕ(ei) = 1 ⊗ 1, we then have
ϕ(eA) = [1 ⊗ 1] + [1 ⊗ 1] + [1 ⊗ 1] = [1]
in Z/2Z
so that
−l
l
m
n
−m − n
−l − m + n
(Z3/(id − A)Z3 ⊗ Z3/(id − At)Z3, eA) ∼= (Z/2Z, [1]).
1 1 1
1 1 1
1 0 0
1 1 1
1 1 0
1 1 0
On the other hand, let B =
=
and hence Bt =
. Since
,
=
(id − Bt)
,
(id − B)
ψ :
ψt :
∈ Z3 −→ [a + b + c] ∈ Z/2Z,
∈ Z3 −→ [b + c] ∈ Z/2Z
ψ((id − B)
ψt((id − Bt)
) = 2(−l − m),
) = −2l
−m − n
−l − n
−l + n
l
m
n
l
m
n
l
m
n
a
b
c
the maps
satisfy
a
b
c
so that they induce isomorphisms
¯ψ : Z3/(id − B)Z3 −→ Z/2Z,
¯ψt : Z3/(id − Bt)Z3 −→ Z/2Z
ψ := ¯ψ ⊗ ¯ψt : Z3/(id − B)Z3 ⊗ Z3/(id − Bt)Z3 −→ Z/2Z ⊗ Z/2Z ∼= Z/2Z.
Proposition 5.6. Let A =
1 1 1
1 1 1
1 1 1
and B =
K0(OB)(∼= Z/2Z) and det(id − A) = det(id − B)(= −2). However
1 1 1
1 1 0
1 1 0
. They satisfy K0(OA) ∼=
(Z3/(id − A)Z3 ⊗ Z3/(id − At)Z3, eA) ∼= (Z/2Z, [1]),
(Z3/(id − B)Z3 ⊗ Z3/(id − Bt)Z3, eB) ∼= (Z/2Z, [0]).
16
and
Since
we then have
so that
ψ(ei ⊗ ei) = ¯ψ(ei) ⊗ ¯ψt(ei) =([1 ⊗ 0] = [0]
[1 ⊗ 1] = [1]
ψ(eA) = [1 ⊗ 0] + [1 ⊗ 1] + [1 ⊗ 1] = [0]
(Z3/(id − B)Z3 ⊗ Z3/(id − Bt)Z3, eB) ∼= (Z/2Z, [0]).
in Z/2Z
if i = 1,
if i = 2, 3,
In the rest of this section, we will deal with square matrices with entries in nonnegative
integers. Such matrices are called nonnegative integral matrices. A nonnegative integral
matrix is said to be essential if none of its rows or columns is zero vector. Let A =
[A(i, j)]N
i,j=1 be an N × N essential nonnegative integral matrix. The matrix defines
a finite directed graph GA = (VA, EA) with N vertices VA = {v1, . . . , vN} and A(i, j)
directed edges from the vertex vi to the vertex vj for i, j = 1, . . . , N . The directed edges
are denoted by {a1, . . . , aNA} = EA. For an edge ak ∈ EA, denote by s(ak), t(ak) its source
vertex, terminal vertex, respectively. The directed graph GA has the NA × NA transition
matrix AG = [AG(i, j)]NA
i,j=1 of edges defined by
AG(i, j) =(1
0
if t(ai) = s(aj),
otherwise,
i, j = 1, . . . , NA.
As in [8, Remark 2.16] and [30, Section 4], the Cuntz -- Krieger algebra OA for the nonneg-
ative integral matrix A is defined to be the Cuntz -- Krieger algebra OAG for the matrix AG
with entries in {0, 1}. It is well-known that there exist rectangular nonnegative integral
matrices R, S such that A = RS, AG = SR (cf.
[16]). As in [17, Lemma 4.5], the left
multiplication of the matrix St induces an isomorphism mSt : ZNA/(id − (AG)t)ZNA −→
ZN /(id − At)ZN such that mSt([1NA ]) = [1N ], where 1NA = [1, . . . , 1] ∈ ZNA, 1N =
[1, . . . , 1] ∈ ZN . Let 1OA be the unit of the Cuntz -- Krieger algebra OA. By [6, Proposition
3.1], there exists an isomorphism from K0(OAG ) to ZNA/(id − (AG)t)ZNA that sends the
class [1OA] of 1OA to the class [1NA] of 1NA. Hence for a nonnegative integral matrix
A, there exists an isomorphism from K0(OA) to ZN /(id − At)ZN that sends the class of
the unit [1OA] of OA to the class [1N ] of 1N . We define the element [eA] in the group
ZN /(id − A)ZN ⊗ ZN /(id − At)ZN by the same formula (5.1) as that for matrices with
entries in {0, 1}. We notice the following lemma.
Lemma 5.7. There exists an isomorphism Φ of groups from ZNA/(id−AG)ZNA⊗ZNA/(id−
(AG)t)ZNA onto ZN /(id − A)ZN ⊗ ZN /(id − At)ZN such that Φ(eAG) = eA.
Proof. Let R, S be rectangular nonnegative integral matrices R, S satisfying A = RS, AG =
SR. As in the proof of Theorem 5.3, the isomorphism mR ⊗ mSt : ZNA/(id − AG)ZNA ⊗
ZNA/(id − (AG)t)ZNA −→ ZN /(id − A)ZN ⊗ ZN /(id − At)ZN satisfies mR ⊗ mSt(eAG) =
eA.
We may obtain the following proposition in a similar way to the proof of Proposition
5.5.
i,j=1 be an N×N essential nonnegative integral matrix.
We will present an example of nonnegative integral matrix A such that the two C ∗-
Proposition 5.8. Let A = [A(i, j)]N
The pair (ZN /(id − A)ZN ⊗ ZN /(id − At)ZN , eA) is invariant under shift equivalence.
algebras eRA and OAt ⊗ OA are not isomorphic.
Let A =(cid:20)4 1
m(cid:21) =(cid:20)−3l − m
Z/4Z induces isomorphisms ¯ϕ : Z2/(id − A)Z2 −→ Z/4Z and
−l + m(cid:21) , the map ϕ :(cid:20) l
1 0(cid:21) . Since (id−A)(cid:20) l
m(cid:21) ∈ Z2 −→ [l+m] ∈
ϕ := ¯ϕ ⊗ ¯ϕ : Z2/(id − A)Z2 ⊗ Z2/(id − At)Z2 −→ Z/4Z ⊗ Z/4Z ∼= Z/4Z.
17
Since ϕ(ei ⊗ ei) = ¯ϕ(ei) ⊗ ¯ϕ(ei) = 1 ⊗ 1, we have
ϕ(eA) = [1 ⊗ 1] + [1 ⊗ 1] = [2]
in Z/4Z.
On the other hand, we have ϕ([12] ⊗ [12]) = ϕ((cid:20)1
1(cid:21)) ⊗ ϕt((cid:20)1
1(cid:21)) = [2⊗ 2] = [0] in Z/4Z. We
thus have
(Z2/(id − A)Z2 ⊗ Z2/(id − At)Z2, eA) ∼= (Z/4Z, [2]),
(Z2/(id − A)Z2 ⊗ Z2/(id − At)Z2, [12] ⊗ [12]) ∼= (Z/4Z, [0]),
unital, purely infinite, simple nuclear C ∗-algebras ([15], [21]).
so that the algebras eRA and OAt ⊗ OA are not isomorphic by classification theorem of
6 KMS states on eRA
In this section, we will study KMS states on the C ∗-algebra eRA for the diagonal action δA.
Following after [3], we will define KMS states in the following way. For a one-parameter
automorphism group αt, t ∈ R on a C ∗-algebra A and a real number γ ∈ R, a state ψ on
A is called a KMS state for the action α if ψ satisfies
ψ(Xαiγ (Y )) = ψ(Y X)
(6.1)
for all X, Y in a norm dense α-invariant ∗-subalgebra of the set of entire analytic elements
for α in A. The value γ is called the inverse temperature and the condition (6.1) is called
the KMS condition.
Let β be the Perron -- Frobenius eigenvalue for an irreducible matrix A with entries in
{0, 1}.
It has been shown in [10] that KMS states for gauge action on Cuntz -- Krieger
algebra OA exists if and only if its inverse temperature is log β, and the admitted KMS
state is unique. Let us denote by ϕ the unique KMS state for gauge action on OA.
Similarly we denote by ϕt the unique KMS state for gauge action on OAt As in [10], the
ϕ(S1S∗
1)
...
ϕ(SN S∗
N )
vector
β
gives rise to the unique normalized positive eigenvector of A for the
=
ϕ(S∗
1 S1)
...
ϕ(S∗
N SN )
=
··· A(N, N )
··· A(1, N )
ϕ(S1S∗
1)
ϕ(SN S∗
N )
A(N, 1)
...
A(1, 1)
...
...
ϕ(S1S∗
1)
...
ϕ(SN S∗
N )
eigenvalue β. Hence we have
so that βϕ(SiS∗
i ) = ϕ(S∗
i Si), i = 1, . . . , N and more generally
βmϕ(Sµ1 ···µmS∗
µ1···µm) = ϕ(S∗
µ1···µmSµ1···µm),
(µ1, . . . , µm) ∈ Bm( ¯XA).
Therefore we have
ϕ(Sµm S∗
µm) =
1
β
ϕ(S∗
µm Sµm) =
1
β
ϕ(S∗
µ1 ···µm Sµ1···µm) = βm−1ϕ(Sµ1 ···µm S∗
µ1···µm)
(6.2)
18
and similarly
ϕt(Tξ1 T ∗
ξ1) = βk−1ϕt(Tξk ···ξ1T ∗
ξk ···ξ1),
(ξk, . . . , ξ1) ∈ Bk( ¯XAt ).
(6.3)
Let [ai]N
i=1 and [bi]N
respectively satisfying
i=1 be the positive eigenvectors of A and At for the eigenvalue β,
aibi = 1.
NXi=1
For admissible words ξ = (ξ1, . . . , ξk) ∈ Bk( ¯XA) and ν = (ν1, . . . , νn) ∈ Bn( ¯XA), put
ξν = (ξ1, . . . , ξk, ν1, . . . , νn) ∈ Bk+n( ¯XA). For i ∈ Z, let us denote by U[ξν]i+k+n−1
the
cylinder set of ¯XA such that
i
i
U[ξν]i+k+n−1
= {(xj )j∈Z ∈ ¯XA xi = ξ1, . . . , xi+k−1 = ξk, xi+k = ν1, . . . , xi+k+n−1 = νn}.
In [20], W. Parry proved that there exists a unique invariant measure µ on ¯XA of maximal
entropy. It is called the Parry measure, which satisfies the following equality
µ(U[ξν]i+k+n−1
i ∈ Z.
Let C ∗(Ga
A. As in the Putnam's paper
[22] and his lecture note [23], the algebra is an AF-algebra with a tracial state Tr defined
by
A) be the groupoid C ∗-algebra for the groupoid Ga
) = bξ1aνnβ−(k+n−1),
(6.4)
i
Tr(f ) =Z ¯XA
f (x, x)dµ(x)
for f ∈ Cc(GA).
ϕ =
Let us define a state ϕ on eRA by setting
Since (ϕt ⊗ ϕ)(EA) =PN
j=1 ϕt(TjT ∗
j=1 ϕt(TjT ∗
PN
We know more about ϕ in the following way.
1
j Sj)
j )ϕ(S∗
j )ϕ(S∗
on eRA = EA(OAt ⊗ OA)EA.
ϕt ⊗ ϕ
j Sj), we know that ϕ gives rise to a state on eRA.
(i) The state ϕ is a KMS state on eRA for the diagonal action δA at
Proposition 6.1.
the inverse temperature log β.
(ii) The restriction of ϕ to the subalgebra C( ¯XA) coincides with the Parry measure µ on
¯XA.
(iii) The formula
holds.
Proof. (i) For
ϕ(Y ) = Tr(cid:18)ZZT2
γA
r,s(Y )drds(cid:19)
for Y ∈ eRA
(6.5)
µ = (µ1, . . . , µm), ν = (ν1, . . . , νn), µ′ = (µ′
¯ξ = (ξk, . . . , ξ1), ¯η = (ηl, . . . , η1), ¯ξ′ = (ξ′
1, . . . , µ′
k′, . . . , ξ′
m′ ), ν ′ = (ν ′
1), ¯η′ = (η′
1, . . . , ν ′
l′, . . . , η′
n′) ∈ B∗( ¯XA),
1) ∈ B∗( ¯XAt)
19
with A(ξk, µ1) = A(ηl, ν1) = A(ξ′
k′, µ′
1) = A(η′
l′, ν ′
1) = 1, put
x = T ¯ξT ∗
¯η ⊗ SµS∗
ν ,
It then follows that
x′ = T ¯ξ ′T ∗
¯η′ ⊗ Sµ′S∗
ν ′ ∈ eRA.
¯η ) ⊗ αA
i log β(SµS∗
i log β(SµS∗
ν ))
ν )))
¯η′ ⊗ Sµ′S∗
(ϕt ⊗ ϕ)(EA) · ϕ(x′δA
=(ϕt ⊗ ϕ)((T ¯ξ ′ T ∗
¯η′αAt
=ϕt(T ¯ξ ′T ∗
=ϕt(T ¯ξT ∗
¯η T ¯ξ ′T ∗
=(ϕt ⊗ ϕ)((T ¯ξT ∗
=(ϕt ⊗ ϕ)(xx′)
=(ϕt ⊗ ϕ)(EA) · ϕ(xx′),
i log β(T ¯ξT ∗
¯η′)ϕ(SµS∗
¯η ⊗ SµS∗
i log β(T ¯ξT ∗
ν ′αA
i log β(x))
ν ′)(αAt
¯η ))ϕ(Sµ′ S∗
ν Sµ′S∗
ν ′)
¯η′ ⊗ Sµ′S∗
ν )(T ¯ξ ′T ∗
ν ′))
thus proving that ϕ is a KMS state eRA for the diagonal action δA at the inverse temper-
ature log β.
(ii) Put
¯ai =
so that
It then follows that
PN
NXi=1
ai
i=1 ai
= ϕ(SiS∗
i ),
ϕ(SiS∗
i )ϕt(TiT ∗
i ) =
µ(U[ξν]i+k+n−1
i
NXi=1
) =¯bξ1 · (
=ϕt(Tξ1T ∗
bi) · ¯aνn(
NXi=1
ξ1) · (
1
By using (6.2) and (6.3) we thus have
i=1 ϕt(TiT ∗
i )ϕ(SiS∗
=
PN
= ϕt(TiT ∗
i )
.
i=1 bi)
¯bi =
bi
i=1 bi
1
PN
(PN
i=1 ai) · (PN
ai) · β−(k+n−1)
NXi=1
i ) · ϕt(Tξ1T ∗
νn)(
NXi=1
bi) · ϕ(SνnS∗
ai) · β−(k+n−1)
ξ1)ϕ(SνnS∗
νn) · β−(k+n−1).
µ(U[ξν]i+k+n−1
i
) =
ξk···ξ1)ϕ(Sν1···νnS∗
ν1···νn)
1
β · ϕt(Tξk···ξ1T ∗
¯ξ )ϕ(Sν S∗
ν )
1
1
=
i )ϕ(SiS∗
i=1 ϕt(TiT ∗
PN
PN
i=1 ϕt(TiT ∗
(ϕt ⊗ ϕ)(EA) · ϕt(T ¯ξT ∗
= ϕ(T ¯ξT ∗
i ) ·
i Si) · ϕt(T ¯ξT ∗
¯ξ )ϕ(Sν S∗
ν )
i )ϕ(S∗
=
¯ξ ⊗ SνS∗
ν ).
1
20
(iii) For
µ = (µ1, . . . , µm),
¯ξ = (ξk, . . . , ξ1),
ν = (ν1, . . . , νn) ∈ B∗( ¯XA),
¯η = (ηl, . . . , η1) ∈ B∗( ¯XAt),
satisfying A(ξk, µ1) = A(ηl, ν1) = 1, it is direct to see the following equalities
ϕ(T ¯ξT ∗
¯η ⊗ SµS∗
ν ) =ϕt(T ¯ξT ∗
¯η )ϕ(SµS∗
ν )
¯ξ )ϕ(Sν S∗
ν )
0
=(ϕt(T ¯ξT ∗
=(µ(U[ξν]i+k+n−1
0
i
if ¯ξ = ¯η, µ = ν,
otherwise
)
if ¯ξ = ¯η, µ = ν,
otherwise.
Since the above value coincides with
Tr(cid:18)ZZT2
we know the formula (6.5).
γA
r,s(T ¯ξT ∗
¯η ⊗ SµS∗
ν )drds(cid:19) ,
We finally prove that a KMS state on eRA for the diagonal action δA exists only if
at the inverse temperature log β. We will further know that the admitted KMS state is
unique. In order to avoid non essential difficulty, we assume that the irreducible matrix
A with entries in {0, 1} is aperiodic so that there exists n0 ∈ N such that An0(i, j) ≥ 1
for all i, j = 1, . . . , N . Let ψ be a KMS state on eRA for the diagonal action δA at the
inverse temperature log γ for 1 < γ ∈ R. We will prove that γ = β: the Perron -- Frobenius
eigenvalue of A and ψ = ϕ.
For i, j ∈ {1, . . . , N} and µ = (µ1, . . . , µm), ν = (ν1, . . . , νn) such that (i, µ1, . . . , µm, j) ∈
Bm+2( ¯XA), (i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA), we set a partial isometry
j ⊗ SiSµ1 ··· SµmSj
Vν,µ(i, j) = T ∗
ν1 ··· T ∗
νnT ∗
i T ∗
(6.6)
Since T ∗
identities
i ⊗ Si, T ∗
j ⊗ Sj ∈ eRA, we know that Vν,µ(i, j) belongs to eRA. We then have the
Vν,µ(i, j)Vν,µ(i, j)∗ =T ∗
=T ∗
νnT ∗
ν1 ··· T ∗
i T ∗
i Ti ⊗ SiSµ1 ··· SµmSjS∗
j S∗
µm ··· S∗
µ1S∗
i
j TjTνn ··· Tν1Ti ⊗ SiSµ1 ··· SµmSjS∗
j S∗
µm ··· S∗
µ1S∗
i
and
i T ∗
i T ∗
ν1 ··· T ∗
ν1 ··· T ∗
Vν,µ(i, j)∗Vν,µ(i, j) =TjTνn ··· Tν1TiT ∗
=TjTνn ··· Tν1TiT ∗
A (p) the pth spectral subspace of eRA for the action δA.
For p ∈ Z, denote by eRδA
Lemma 6.2. Suppose that X ∈ eRA belongs to eRδA
j ⊗ S∗
j ⊗ S∗
µm ··· S∗
νnT ∗
νnT ∗
j S∗
j Sj
ψ(X) = 0
µ1S∗
A (p) for some p 6= 0. Then we have
i SiSµ1 ··· SµmSj
21
Proof. We may assume p > 0. For i, j = 1, . . . , N , let µ = (µ1, . . . , µn0+p) be an ad-
missible word such that (i, µ1, . . . , µn0+p, j) ∈ Bn0+p+2( ¯XA). Take ν = (ν1, . . . , νn0) with
(i, ν1, . . . , νn0, j) ∈ Bn0( ¯XA) and consider the partial isometry
i T ∗
Vν,µ(i, j) = T ∗
ν1 ··· T ∗
The partial isometry Vν,µ(i, j) belongs to eRδA
Vν,µ(i, j)Vν,µ(i, j)∗ = T ∗
νn0
A (p) and satisfies
i Ti ⊗ SiSµ1 ··· Sµn0+pSjS∗
T ∗
j ⊗ SiSµ1 ··· Sµn0+pSj.
j S∗
µn0+p ··· S∗
µ1S∗
i .
We then have
EA =
NXi,j=1 Xµ∈Bn0 +p( ¯XA)
Vν,µ(i, j)Vν,µ(i, j)∗.
It then follows that
ψ(X) =ψ(EAX)
ψ(Vν,µ(i, j)Vν,µ(i, j)∗X)
ψ(Vν,µ(i, j)∗XδA
i log γ(Vν,µ(i, j)))
1
γp
NXi,j=1 Xµ∈Bn0+p( ¯XA)
NXi,j=1 Xµ∈Bn0+p( ¯XA)
NXi,j=1 Xµ∈Bn0 +p( ¯XA)
NXi,j=1 Xµ∈Bn0 +p( ¯XA)
NXi,j=1 Xµ∈Bn0 +p( ¯XA)
1
γp
1
γp
1
γp ψ(EAX).
=
=
=
=
=
=
ψ(Vν,µ(i, j)∗XVν,µ(i, j))
ψ(Vν,µ(i, j)δA
i log γ(Vν,µ(i, j)∗X))
ψ(Vν,µ(i, j)Vν,µ(i, j)∗X)
Since γ > 1, we have ψ(X) = 0.
Since RA is the fixed point algebra (eRA)δA
expectation EA : eRA −→ RA by
δA
t (X)dt,
EA(X) =ZT
The preceding lemma implies the following lemma.
of eRA under δA, we may define a conditional
X ∈ eRA.
(6.7)
Lemma 6.3. Let ψ0 be the restriction of ψ to the subalgebra (eRA)δA
state on RA such that ψ = ψ0 ◦ EA.
. Then ψ0 is a tracial
22
Hence the value of KMS state is determined on the subalgebra RA. Recall that UA
i=1 T ∗
i ⊗ Si which belongs to RA.
denotes the unitary UA =PN
Lemma 6.4. ψ(UAXU ∗
Proof. Since UA is fixed under the action δA, we have
A) = ψ(X) for all X ∈ eRA.
ψ(X) = ψ(U ∗
AUAX) = ψ(UAXδA
i log γ(U ∗
A)) = ψ(UAXU ∗
A).
As in [18, Proposition 9.9], the automorphism Ad(UA) behaves like the shift on eRA.
Lemma 6.4 tells us that the KMS state is invariant nuder the shift. The following lemma
is crucial in our discussions.
Lemma 6.5. Let X = T ¯ξT ∗
¯η ⊗ SµS∗
ν ∈ RA where
µ = (µ1, . . . , µm), ν = (ν1, . . . , νn) ∈ B∗( ¯XA),
¯ξ = (ξk, . . . , ξ1), ¯η = (ηl, . . . , η1) ∈ B∗( ¯XAt ).
Suppose that ψ(X) 6= 0. Then we have k = l, m = n and µ = ν, ¯ξ = ¯η.
Proof. Since X belongs to RA, we have A(ξk, µ1) = A(ηl, ν1) = 1 and k − l = n − m. We
may assume that k ≥ l and hence n ≥ m. It then follows that
ψ(X) =ψ(Tξk ··· Tξ1T ∗
=ψ((Tξk ··· Tξ1T ∗
ηl ⊗ Sν1S∗
=ψ((Tηl T ∗
γk+m−l−n ψ((Tηl T ∗
=ψ(TηlT ∗
νn ··· S∗
ν1)
νn ··· S∗
ν1) · (TηlT ∗
ηl ⊗ Sν1S∗
ηl ⊗ Sµ1 ··· SµmS∗
η1 ··· T ∗
ν1) · (Tξk ··· Tξ1T ∗
η1 ··· T ∗
ηl ⊗ Sν1S∗
η1 ··· T ∗
η1 ··· T ∗
ν1) · δA
ηl ⊗ Sν1S∗
η1 ··· T ∗
ηl ⊗ Sµ1 ··· SµmS∗
ηl ⊗ Sµ1 ··· SµmS∗
i log γ(Tξk ··· Tξ1T ∗
ηl ⊗ Sµ1 ··· SµmS∗
ν1))
νn ··· S∗
ηl · Tξk ··· Tξ1T ∗
νn ··· S∗
By the assumption ψ(X) 6= 0, we get ηl = ξk and ν1 = µ1, and we have
νn ··· S∗
ψ(X) = ψ(Tηl Tξk−1 ··· Tξ1T ∗
ηl ⊗ Sµ1 ··· SµmS∗
ν1 · Sµ1 ··· SµmS∗
ν1))
νn ··· S∗
η1 ··· T ∗
ν1))
ν1).
ν1).
=
1
As
UA · TηlTξk−1 ··· Tξ1T ∗
=Tξk−1 ··· Tξ1T ∗
η1 ··· T ∗
Lemma 6.4 shows us the equality
ηl ⊗ Sµ1 ··· SµmS∗
η1 ··· T ∗
ηl−1 ⊗ SηlSµ1 ··· SµmS∗
νn ··· S∗
νn ··· S∗
ν1 · U ∗
ν1S∗
ηl,
A
ψ(X) = ψ(Tξk−1 ··· Tξ1T ∗
η1 ··· T ∗
ηl−1 ⊗ SηlSµ1 ··· SµmS∗
νn ··· S∗
ν1S∗
ηl).
(6.8)
We apply the same argument above to the right hand side of (6.8), and continue these
procedures so that we finally get
ηl−1 = ξk−1, ηl−2 = ξk−2, . . . , η1 = ξk−l+1
23
and the identity
algebra eRA such that
ψ(X) = ψ(Tξk−l ··· Tξ1T ∗
νn ··· S∗
As ξk−l+1 = η1, we see that A(ξk−l, η1) = 1 and hence Tξk−l ··· Tξ1T ∗
η1Tη1 ⊗ Sη1Sη2 ··· SηlSµ1 ··· SµmS∗
ν1S∗
ηl ··· S∗
η2S∗
η1).
η1 ⊗ Sη1 belongs to the
δA
i log γ(Tξk−l ··· Tξ1T ∗
η1 ⊗ Sη1) =
1
γk−l Tξk−l ··· Tξ1T ∗
η1 ⊗ Sη1.
Hence we have
ψ(X) =ψ((Tξk−l ··· Tξ1T ∗
=ψ((Tη1 ⊗ Sη2 ··· SηlSµ1 ··· SµmS∗
1
γk−l ψ(Tη1 Tξk−l ··· Tξ1T ∗
=
η2S∗
η2, we have
η1Sη1 = S∗
Since S∗
η1 ⊗ Sη1) · (Tη1 ⊗ Sη2 ··· SηlSµ1 ··· SµmS∗
η1) · δA
η2S∗
ηl ··· S∗
νn ··· S∗
η1 ⊗ Sη2 ··· SηlSµ1 ··· SµmS∗
νn ··· S∗
ν1S∗
ν1S∗
ηl ··· S∗
νn ··· S∗
η2S∗
i log γ(Tξk−l ··· Tξ1T ∗
ν1S∗
η1Sη1).
η2S∗
ηl ··· S∗
η1))
η1 ⊗ Sη1))
ψ(X) =
=
=
1
γk−l ψ(Tη1 Tξk−l ··· Tξ1T ∗
1
γk−l ψ(U ∗
1
γk−l ψ((Tηl ··· Tη2Tη1Tξk−l ··· Tξ1T ∗
l−1(Tη1Tξk−l ··· Tξ1T ∗
A
η1 ⊗ Sη2 ··· SηlSµ1 ··· SµmS∗
νn ··· S∗
η1 ⊗ Sη2 ··· SηlSµ1 ··· SµmS∗
ν1S∗
ηl ··· S∗
η2)
ν1S∗
νn ··· S∗
η1T ∗
η2 ··· T ∗
ηl ⊗ Sµ1 ··· SµmS∗
νn ··· S∗
η2U l−1
A )
ηl ··· S∗
ν1).
Since (ηl, . . . , η1) = (ξk, . . . , ξk−l+1), we finally obtain that
ψ(X) =
1
γk−l ψ(X)
so that k = l and hence η = ξ. We similarly see that µ = ν.
Since any element X of RA is approximated by finite linear combinations of elements
ν ∈ RA, we have the following proposition by using Lemma 6.3.
¯η ⊗ SµS∗
of the form T ¯ξT ∗
Proposition 6.6. If an element X ∈ eRA satisfies ψ(X) 6= 0, then X belongs to C( ¯XA).
We will next show that the restriction of the KMS state ψ to the commutative subal-
gebra C( ¯XA) coincides with the state defined by the Parry measure on ¯XA.
Recall that the partial isometry Vν,µ(i, j) for i, j ∈ {1, . . . , N} and µ = (µ1, . . . , µm), ν =
(ν1, . . . , νn) such that (i, µ1, . . . , µm, j) ∈ Bm+2( ¯XA), (i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA) is de-
fined by (6.6). We set
pm,µ(i, j) = ψ(TiT ∗
i ⊗ Sµ1 ··· SµmSjS∗
j S∗
µm ··· S∗
µ1).
The following lemma holds.
Lemma 6.7.
(i) ψ(Vν,µ(i, j)Vν,µ(i, j)∗) = pm,µ(i, j).
24
(ii) ψ(Vν,µ(i, j)∗Vν,µ(i, j)) = pn,ν(i, j).
(iii) ψ(Vν,µ(i, j)Vν,µ(i, j)∗) = γn−mψ(Vν,µ(i, j)∗Vν,µ(i, j)).
i log γ(T ∗
i ⊗ Si) = T ∗
Proof. (i) Since T ∗
i Ti ⊗ SiSµ1 ··· SµmSjS∗
j S∗
µm ··· S∗
i ⊗ Si) · (Ti ⊗ Sµ1 ··· SµmSjS∗
j S∗
µm ··· S∗
µ1S∗
j S∗
µm ··· S∗
i ⊗ Si belongs to eRA such that δA
ψ(Vν,µ(i, j)Vν,µ(i, j)∗) =ψ(T ∗
=ψ((T ∗
=ψ((Ti ⊗ Sµ1 ··· SµmSjS∗
j S∗
=ψ(TiT ∗
i ⊗ Sµ1 ··· SµmSjS∗
=pm,µ(i, j).
i ⊗ Si, we have
µ1S∗
i )
µm ··· S∗
i ) · (T ∗
µ1S∗
i Si)
µ1S∗
i ))
i ⊗ Si))
(ii) We have Vν,µ(i, j)∗Vν,µ(i, j) = TjTνn ··· Tν1TiT ∗
A Vν,µ(i, j)∗Vν,µ(i, j)U ∗n+1
U n+1
A
= TiT ∗
i T ∗
ν1 ··· T ∗
νnT ∗
i ⊗ Sν1 ··· SνnSjS∗
j ⊗ S∗
j S∗
νn ··· S∗
ν1.
j Sj and hence
By Lemma 6.4, we have the desired identity.
(iii) As δA
i log γ(Vν,µ(i, j)∗) = γn−mVν,µ(i, j)∗, the KMS condition for ψ ensures us the
desired identity.
The preceding lemma tells us that the values pm,µ(i, j) and pn,ν(i, j) coincide each other
for m = n as long as (i, µ1, . . . , µm, j) ∈ Bm+2( ¯XA), (i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA). Hence
the value pn,ν(i, j) does not depend on the choice of the word ν as long as the length of ν is n
and (i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA). We may thus define pn(i, j) by pn,ν(i, j) for some ν with
(i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA). If there are no word ν such as (i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA),
then we define pn(i, j) to be zero.
Lemma 6.8. Let i, j = 1, . . . , n and n ∈ Z+.
(i) Assume An+1(i, j) > 0 and An+2(i, j) > 0. Then we have pn(i, j) = γpn+1(i, j).
(ii) Assume An+1(i, j) > 0. Then we have
pn(i, j) =
NXk=1
A(j, k)pn+1(i, k) =
A(h, i)pn+1(h, j).
NXh=1
(iii) Assume An(i, j) > 0 and An+1(i, j) > 0. Then we have
γpn(i, j) =
NXk=1
A(j, k)pn(i, k) =
A(h, i)pn(h, j).
NXh=1
Proof. (i) Since An+1(i, j), An+2(i, j) > 0, we may find ν = (ν1, . . . , νn), µ = (µ1, . . . , µn+1)
such that (i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA), (i, µ1, . . . , µn+1, j) ∈ Bn+3( ¯XA). Consider Vν,µ(i, j) =
T ∗
i T ∗
j ⊗ SiSµ1 ··· Sµn+1Sj. It then follows that
ν1 ··· T ∗
νnT ∗
pn(i, j) =pn,ν(i, j)
=ψ(Vν,µ(i, j)∗Vν,µ(i, j))
=ψ(Vν,µ(i, j)δA
=γψ(Vν,µ(i, j)Vν,µ(i, j)∗)
=γpn+1,µ(i, j) = γpn+1(i, j).
i log γ(Vν,µ(i, j)∗))
25
(ii) Since An+1(i, j) > 0, we may find ν = (ν1, . . . , νn) such that (i, ν1, . . . , νn, j) ∈
Bn+2( ¯XA). It then follows that
pn(i, j) =pn,ν(i, j)
=ψ(TiT ∗
i ⊗ Sν1 ··· SνnSjS∗
A(j, k)ψ(TiT ∗
j S∗
νn ··· S∗
ν1)
i ⊗ Sν1 ··· SνnSjSkS∗
kS∗
j S∗
νn ··· S∗
ν1)
A(j, k)pn+1,νj (i, k) =
A(j, k)pn+1(i, k).
NXk=1
=
=
NXk=1
NXk=1
We also see that
pn(i, j) =
=
NXh=1
NXh=1
At(i, h)ψ(TiThT ∗
i T ∗
j S∗
νn ··· S∗
ν1)
A(h, i)pn+1,iνj (h, j) =
A(h, i)pn+1(h, j).
h ⊗ Sν1 ··· SνnSjS∗
NXh=1
The assertion (iii) follows from (i) and (ii).
Lemma 6.9. For i = 1, . . . , N and n ∈ Z+, we have
(i) PN
(ii) PN
j=1 An+1(i, j)pn(i, j) = ψ(T ∗
i Ti ⊗ SiS∗
i ⊗ S∗
Proof. (i) We have the following identities
j=1 An+1(j, i)pn(i, j) = ψ(TiT ∗
i ) and hencePN
i Si) and hencePN
i,j=1 An+1(i, j)pn(i, j) = 1.
i,j=1 An+1(j, i)pn(i, j) = 1.
A(i, µ1)ψ(T ∗
i Ti ⊗ SiSµ1S∗
µ1S∗
i )
ψ(T ∗
i Ti ⊗ SiS∗
i ) =
=
=
NXµ1=1
NXj=1
NXµ1,...,µn=1
NXj=1
An+1(i, j)pn(i, j).
A(i, µ1)A(µ1, µ2)··· A(µn, j)ψ(T ∗
i Ti ⊗ Siµ1···µnjS∗
iµ1···µnj)
We also havePN
i Ti ⊗ SiS∗
(ii) is similarly shown to (i) .
i=1 ψ(T ∗
i ) = ψ(EA) = 1.
i ) = ψ(TiT ∗
i Ti ⊗ SiS∗
We notice that ψ(T ∗
i ⊗
Si) = T ∗
i ⊗ Si and the KMS condition for ψ. Recall that we are assuming the matrix A
is aperiodic so that there exists n0 ∈ N such that An(i, j) > 0 for all i, j = 1, . . . , N and
n ≥ n0.
Lemma 6.10. γ = β.
i Si) because of the equality δA
i ⊗ S∗
i log γ(T ∗
26
Proof. Lemma 6.8 together with Lemma 6.9 implies that the vector [pn(i, k)]N
k=1 is a
nonnegative eigenvector of the matrix A of eigenvalue γ for each n ∈ N and i = 1, . . . , N .
Since A is aperiodic, [pn(i, k)]N
k=1 is actually a positive eigenvector of eigenvalue γ. By
Perron -- Frobenius theorem, γ coincides with the Perron-Frobenius eigenvalue β.
We have seen that γ must be the Perron-Frobenius eigenvalue of the matrix A by
Lemma 6.10. Its proof does not need the assumption γ > 1 that we had first assumed.
Now the matrix A is aperiodic and not any permutation so that its Perron-Frobenius
eigenvalue is always greater than one. Hence γ(= β) becomes greater than one without
assumption γ > 1.
Recall that [aj]N
j=1, [bi]N
β respectively such thatPN
Lemma 6.11. For n ≥ n0 and i, j = 1, . . . , N , we have
i=1 be the positive eigenvectors of A and At for the eigenvalue
i=1 aibi = 1. We have the following lemma.
pn(i, j) =
biaj
(PN
h=1 bh) · (PN
k=1 ak)
NXh,k=1
pn(h, k).
(6.9)
Proof. We fix n ≥ n0. For a fixed i = 1, . . . , N, the vector [pn(i, k)]N
k=1 is a positive
eigenvector of the matrix A for the eigenvalue β. By the uniqueness of the positive
eigenvector of A, we may find a positive real number cn,i such that
pn(i, j) = cn,iaj
for j = 1, . . . , N.
(6.10)
By Lemma 6.8, we know that the vector [PN
matrix At for the eigenvalue β. Hence the normalized positive eigenvectors [
j=1 pn(i, j)]N
i=1 is a positive eigenvector of the
]N
i=1
PN
j=1 pn(i,j)
PN
h,k=1 pn(h,k)
and [
biPN
k=1 bk
]N
i=1 coincide, so that we have
NXj=1
pn(i, j) = biPN
PN
h,k=1 pn(h, k)
k=1 bk
By (6.10) and (6.11), we have
for i = 1, . . . , N.
(6.11)
cn,i = PN
PN
j=1 pn(i, j)
=
j=1 aj
so that we know (6.9) by using (6.10) again.
h,k=1 pn(h, k)
biPN
(PN
k=1 bk)(PN
j=1 aj)
We thus obtain the following lemma.
Lemma 6.12. For n ≥ n0 and i, j = 1, . . . , N , we have
pn(i, j) =
1
βn+1 biaj.
27
Proof. We fix n ≥ n0. By Lemma 6.11 together with Lemma 6.9, we have
1 =
NXi,j=1
An+1(i, j)pn(i, j) =
NXi,j=1
NXh,k=1
pn(h, k).
k=1 ak)
As [aj]N
j=1 is a positive eigenvector of A for the eigenvalue β, we have
NXi,j=1
so that the equalities
An+1(i, j)biaj =
βn+1biai = βn+1
An+1(i, j)biaj
(PN
h=1 bh) · (PN
NXi=1
and
βn+1
k=1 ak)
1 =
(PN
h=1 bh) · (PN
NXh,k=1
pn(h, k) =
1
βn+1 (
NXi=1
NXh,k=1
pn(h, k)
bi) · (
NXj=1
aj)
(6.12)
hold. By (6.9) and (6.12), we get the desired equality.
Consequently, we know the following proposition.
Proposition 6.13. The restriction of a KMS state ψ on eRA to the commutative C ∗-
subalgebra C( ¯XA) coincides with the state defined by the Parry measure on ¯XA.
Proof. For n ≥ n0 and ξ = (i, ν1, . . . , νn, j) ∈ Bn+2( ¯XA), Lemma 6.12 shows that
ψ(TiT ∗
i ⊗ Sν1···νnjS∗
ν1···νnj) =
1
βn+1 biaj.
Let µ be the Parry measure on ¯XA. Since the Parry measure of the cylinder set U[ξ]m+n+1
¯XA, m ∈ Z for the word ξ is given by
m
⊂
µ(U[ξ]m+n+1
m
) =
1
βn+1 biaj
[ξ]m+n+1
m
be the charachteristic function of the cylinder set
by the formula (6.4). Let χU
U[ξ]m+n+1
. Since
m
ψ(TiT ∗
i ⊗ Sν1···νnjS∗
ν1···νnj) = ψ(χU
),
[ξ]m+n+1
m
we obtain
µ(U[ξ]m+n+1
m
) = ψ(χU
).
[ξ]m+n+1
m
Any cylinder set on ¯XA is a finite union of cylinder sets of words having its length greater
than n0 + 1. Hence we may conclude that the restriction of ψ to the commutative C ∗-
subalgebra C( ¯XA) of eRA coincides with the state defined by the Parry measure on ¯XA.
Therefore we reach the following theorem.
28
Theorem 6.14. Assume that the matrix A is aperiodic. A KMS state on eRA for the
action δA at the inverse temperature log γ exists if and only if γ is the Perron -- Frobenius
eigenvalue of A. The admitted KMS state is unique. The restriction of the admitted KMS
state to the subsalgebra C( ¯XA) is the state defined by the Parry measure on ¯XA.
Acknowledgments: This work was supported by JSPS KAKENHI Grant Number
15K04896.
References
[1] C. Anantharaman-Delaroche and J. Renault, Amenable Groupoids,
L'Enseignement Math´ematique Gen´eve, 2000.
[2] R. Bowen, Markov partitions for axiom A diffeomorphisms, Amer. J. Math.
92(1970), pp. 725 -- 747
[3] O. Bratteli and D. W. Robinson, Operator algebras and quantum statistical
mechanics II, Springer-Verlag, New York, Berlin, Heidelberg (1981).
[4] L. G. Brown, Stable isomorphism of hereditary subalgebras of C ∗-algebras, Pacific
J. Math. 71(1977), pp. 335 -- 348.
[5] T. M. Carlsen and J. Rout, Diagonal-preserving gauge invariant isomorphisms
of graph C ∗-algebras, preprint, arXiv: 1610.00692 [math. OA].
[6] J. Cuntz, A class of C ∗-algebras and topological Markov chains II: reducible chains
and the Ext-functor for C ∗-algebras, Invent. Math. 63(1980), pp. 25 -- 40.
[7] J. Cuntz, On the homotopy groups of the space of endomorphisms of a C ∗-algebra
(with applications to topological Markov chains), Operator algebras and group repre-
sentations, vol. I (Neptun, 1980), 124 -- 137, Monogr. Stud. Math., 17, Pitman, Boston,
MA, 1984.
[8] J. Cuntz and W. Krieger, A class of C ∗-algebras and topological Markov chains,
Invent. Math. 56(1980), pp. 251 -- 268.
[9] M. Enomoto, M. Fujii and Y. Watatani, K0-groups and classifications of Cuntz --
Krieger algebras, Math. Japon. 26(1981), pp. 443 -- 460.
[10] M. Enomoto, M. Fujii and Y. Watatani, KMS states for gauge action on OA,
Math. Japon 29(1984), pp. 607 -- 619.
[11] C. G. Holton, The Rohlin property for shifts of finite type, J. Funct. Anal.
229(2005), pp. 277 -- 299.
[12] J. Kaminker and I. F. Putnam, K-theoretic duality of shifts of finite type, Comm.
Math. Phys. 187(1997), pp. 509 -- 522.
[13] J. Kaminker, I. F. Putnam and J. Spielberg, Operator algebras and hyperbolic
dynamics, Operator algebras and quantum field theory (Rome, 1996), 525 -- 532, Int.
Press, Cambridge, MA, 1997.
29
[14] D. B. Killough and I. F. Putnam, Ring and module structures on dimension
groups
associated with a shift of finite type, Ergodic Theory Dynam. Systems 32(2012), pp.
1370 -- 1399.
[15] E. Kirchberg, The classification of purely infinite C ∗-algebras using Kasparov's
theory, preprint,1994.
[16] D. Lind and B. Marcus, An introduction to symbolic dynamics and coding,
Cambridge University Press, Cambridge (1995).
[17] K. Matsumoto, Topological conjugacy of topological Markov shifts and Cuntz --
Krieger algebras, Documenta Math. 22(2017), pp. 873 -- 915.
[18] K. Matsumoto, Asymptotic continuous orbit equivalence of Smale spaces and Ruelle
algebras, preprint, arXiv: 1703.07011, to appear in Can. J. Math..
[19] K. Matsumoto, Flow equivalence of topological Markov shifts and Ruelle algebras,
preprint, arXiv: 1801.09198.
[20] W. Parry, Intrinsic Markov chains, Trans. Amer. Math. Soc. 112(1964), pp. 55-66.
[21] N. C. Phillips, A classification theorem for nuclear purely infinite simple C ∗-
algebras, Doc. Math. 5(2000), pp. 49 -- 114.
[22] I. F. Putnam, C ∗-algebras from Smale spaces, Can. J. Math. 48(1996), pp. 175 -- 195.
[23] I. F. Putnam, Hyperbolic systems and generalized Cuntz -- Krieger algebras, Lecture
Notes, Summer School in Operator Algebras, Odense August 1996.
[24] I. F. Putnam, Functoriality of the C ∗-algebras associated with hyperbolic dynamical
systems, J. London Math. Soc. 62(2000), pp. 873 -- 884.
[25] I. F. Putnam, A homology theory for Smale spaces, Memoirs of Amer. Math. Soc.
232(2014), No. 1094.
[26] I. F. Putnam and J. Spielberg, The structure of C ∗-algebras associated with
hyperbolic dynamical systems, J. Func. Anal. 163(1999), pp. 279 -- 299.
[27] J. Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Math. 793,
Springer-Verlag, Berlin, Heidelberg and New York (1980).
[28] J. Renault, Cartan subalgebras in C ∗-algebras, Irish Math. Soc. Bull. 61(2008), pp.
29 -- 63.
[29] J. Renault, Examples of masas in C ∗-algebras, Operator structures and dynamical
systems, pp. 259 -- 265, Contemp. Math., 503, Amer. Math. Soc., Providence, RI, 2009.
[30] M. Rørdam, Classification of Cuntz-Krieger algebras, K-theory 9(1995), pp. 31 -- 58.
[31] M. Rørdam, Classification of purely infinite simple C ∗-algebras I, J. Func. Anal.
131(1995), pp. 415 -- 458.
30
[32] J. Rosenberg and C. Schochet, The Kunneth theorem and the universal coef-
ficient theorem for Kasparov's generalized K-functor, Duke Math. J. 55(1987) pp.
431 -- 474.
[33] D. Ruelle, Thermodynamic formalism, Addison-Wesley, Reading (Mass.) (1978).
[34] D. Ruelle, Non-commutative algebras for hyperbolic diffeomorphisms, Invent. Math.
93(1988), pp. 1 -- 13.
[35] S. Smale Differentiable dynamical systems, Bull. Amer. Math. Soc. 73(1967), pp.
747 -- 817.
[36] K. Thomsen, C ∗-algebras of homoclinic and heteroclinic structure in expansive dy-
namics, Memoirs of Amer. Math. Soc. 206(2010), No 970.
[37] P. Walters, An Introduction to Ergodic Theory, Springer Graduate Texts in Math
79, New York, Heidelberg, 1982.
[38] R. F. Williams, Classification of subshifts of finite type, Ann. Math. 98(1973), pp.
120 -- 153. erratum, Ann. Math. 99(1974), pp. 380 -- 381.
31
|
1911.05008 | 1 | 1911 | 2019-11-12T16:48:24 | Curvature of differentiable Hilbert modules and Kasparov modules | [
"math.OA",
"math-ph",
"math.DG",
"math.KT",
"math-ph",
"math.QA"
] | In this paper we introduce the curvature of densely defined universal connections on Hilbert $C^{*}$-modules relative to a spectral triple (or unbounded Kasparov module), obtaining a well-defined curvature operator. Fixing the spectral triple, we find that modulo junk forms, the curvature only depends on the represented form of the universal connection. We refine our definition of curvature to factorisations of unbounded Kasparov modules. Our refined definition recovers all the curvature data of a Riemannian submersion of compact manifolds, viewed as a $KK$-factorisation. | math.OA | math |
Curvature of differentiable Hilbert modules and
Kasparov modules
Bram Mesland§∗, Adam Rennie†, Walter D. van Suijlekom‡∗
§Mathematisch Instituut, Universiteit Leiden, The Netherlands
†School of Mathematics and Applied Statistics, University of Wollongong
Wollongong, Australia
‡IMAPP Mathematics, Faculty of Science
Radboud University Nijmegen, The Netherlands
November 13, 2019
Abstract
In this paper we introduce the curvature of densely defined universal connec-
tions on Hilbert C∗-modules relative to a spectral triple (or unbounded Kasparov
module), obtaining a well-defined curvature operator. Fixing the spectral triple,
we find that modulo junk forms, the curvature only depends on the represented
form of the universal connection. We refine our definition of curvature to factori-
sations of unbounded Kasparov modules. Our refined definition recovers all the
curvature data of a Riemannian submersion of compact manifolds, viewed as a
KK-factorisation.
Contents
1 Universal differential forms for C 2-Kasparov modules
1.1 Operator ∗-algebras and differential structures . . . . . . . . . . . . . . .
1.2 Universal and represented differential forms for C 2-spectral triples . . . .
1.2.1 Rough algebraic outline
. . . . . . . . . . . . . . . . . . . . . . .
1.2.2 The formal definitions of represented forms . . . . . . . . . . . . .
1.3 Second derivatives and junk . . . . . . . . . . . . . . . . . . . . . . . . .
2 Curvature in unbounded KK-theory
2.1 The definition of C 1 and C 2-connections
. . . . . . . . . . . . . . . . . .
2.2 The represented curvature of a C 2-connection on C 1-module . . . . . . .
2.2.1 The curvature operator of a C 2-connection on a C (1,2)-module . .
2.3 The curvature of a C 2-correspondence . . . . . . . . . . . . . . . . . . . .
2.4 Universal connections on locally convex modules . . . . . . . . . . . . . .
∗email: [email protected], [email protected], [email protected]
5
6
8
9
10
12
13
14
19
21
23
26
1
3 Finitely generated projective modules over spectral triples
3.1 Connections on finite projective C 2-modules . . . . . . . . . . . . . . . .
3.2 The curvature operator for finitely generated projective modules . . . . .
4 Grassmann connections
4.1 Differentiable stabilisation . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 C 2-Grassmann connections . . . . . . . . . . . . . . . . . . . . . . . . . .
5 The curvature operator of a Riemannian submersion
5.1 Geometric setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Local expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 The stabilisation isometry . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 The universal lift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Curvature of ∇ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
28
29
30
31
32
34
34
37
38
41
43
Introduction
This paper offers a new approach to defining and effectively computing curvature of
Hilbert modules and unbounded Kasparov modules. We will introduce a notion of cur-
vature directly at the operator-algebraic level, thus sidestepping some of the difficulties
imposed by the absence of a differential graded algebra. Our approach does not rely
on the heat kernel coefficient analogy, and so our results differ from the recent work
of [14, 11, 12, 20, 22, 34, 35]. Rather, we provide a complementary point of view.
Numerous other approaches to curvature have appeared independently in many works
[1, 5, 6, 16, 17, 40].
The usual theory of curvature of Z2-graded right modules X over associative algebras B
relies, in its most algebraic formulation, on the existence of a differential graded algebra
(Ω∗(B), δ). The first step, existence of connections
∇ : X → X ⊗B Ω1(B)
∇(xb) = ∇(x)b + γ(x) ⊗ δ(b),
x ∈ X, b ∈ B
where γ is the grading, was settled by Cuntz and Quillen [15]. The curvature R∇ of
(X,∇) is then defined as the composition
∇2 : XB
∇→ X ⊗B Ω1(B)
∇⊗1+1⊗δ−→ X ⊗B Ω2(B).
(0.1)
Thus as soon as we have a differential graded algebra and a connection we obtain an
endomorphism-valued two-form ∇2. The full details of this approach in noncommutative
geometry, pioneered by Connes and Rieffel [13], appear in the book of Connes [8, Sect.
VI.1] (see also [33, Sect. 7.2]).
Instead, the route we take to curvature is inspired by tools and ideas from unbounded
KK-theory, and the unbounded version of the internal Kasparov product in particu-
lar. As we will see, the above algebraic notion of curvature appears naturally in this
functional analytical framework. Let us sketch the main idea.
2
Suppose we are given two (suitably differentiable) unbounded KK-cycles (A, X, S) and
(B, Y, T ) and a (suitable) connection ∇ on X. We may consider an unbounded rep-
resentative of the internal Kasparov product given by the essentially self-adjoint and
regular operator S ⊗ 1 + 1 ⊗∇ T , defined on the appropriate domain in X ⊗B Y (see
[7, 28, 37, 38, 39] for more details). Our definition of curvature is given by the formula
R(S,∇T ) = (S ⊗ 1 + 1 ⊗∇ T )2 − (S2 ⊗ 1 + 1 ⊗∇ T 2).
(0.2)
We will make precise sense of this unbounded operator on X ⊗B Y in due course, but
let us highlight some of its features:
• the curvature R(S,∇T ) can be interpreted as a measure of the "defect" of the internal
Kasparov product to respect the taking of squares of the operators S and T . This
is in line with the notion of curvature for linear maps used in cyclic theory (see
[15, page 255]).
• it vanishes for a direct product of spaces: for the external Kasparov product we
have for the tensor sum (on graded modules):
(D1 ⊗ 1 + 1 ⊗ D2)2 − D2
1 ⊗ 1 − 1 ⊗ D2
2 = 0
• the "geometric" information described by R(S,∇T ) is only accessible at the un-
bounded level, thus forming a refinement of the topological information described
at the level of (bounded) KK-theory.
Our main task is now to make sense of formula (0.2) so let us see what it says alge-
braically. Upon expanding the brackets we observe that the curvature can be understood
in terms of the following two operators
(1 ⊗∇ T )2 − 1 ⊗∇ T 2,
[S ⊗ 1, 1 ⊗∇ T ] .
The approach we will take is to first make sense of the above two operators and then
define R(S,∇T ) in terms of them in Section 2.3. Intriguingly, the well-definedness of the
operator (1⊗∇ T )2−1⊗∇ T 2 in Definition 2.19 relies heavily on the existence of a relative
S-bound on the commutator [S ⊗ 1, 1 ⊗∇ T ] (cf. Definition 2.10 below), which in turn
is a sufficient condition for representing the Kasparov product (see [31, 37]). Moreover,
it turns out that the operator (1 ⊗∇ T )2 − 1 ⊗∇ T 2 is of interest in itself, and we will
call it the curvature operator of the (C 2-) connection ∇ on the module X.
This terminology is justified by the result (Theorem 2.21) that (1 ⊗∇ T )2 − 1 ⊗∇ T 2 is
given by a represented square πT (∇2) of a universal connection ∇ as we now explain.
Unbounded Kasparov modules like (B, Y, T ) provide the basic geometric objects of non-
commutative geometry. One feature of the Kasparov module (B, Y, T ) is the bimodule
of differential one-forms [8, Chapter VI]
T (B) := span{a[T, b] : a, b ∈ B}.
Ω1
3
T (B) consists of operators on the Hilbert C∗-module Y . The differential
The bimodule Ω1
graded algebra of universal differential forms Ω∗
u(B) can then be represented as operators
on Y by taking products, but the image of this representation does not carry the structure
of a differential graded algebra anymore.
The obstruction to defining a differential is the existence of junk-forms [8, Chapter VI].
Although quotienting out the junk forms yields a differential graded algebra, it can no
longer be represented on Y . Our represented curvature
πT (∇2) = 1 ⊗∇ T 2 − (1 ⊗∇ T )2,
is well-defined up to junk forms, and so connects to the existing literature on curvature
of connections. Of course the challenge is to make sense of this square, which we do in
Section 2.2.1.
It is also useful to illustrate our notion of curvature for finitely-generated projective
modules X over B (see Section 3 below for full details). So, consider computing the
Kasparov product (C, XB, 0) ⊗B (B, H, D). In this case, a smooth submodule XB ⊂ XB
is guaranteed to exist. Then realising X ∼= pBN for some projection p ∈ MN (B),
all compatible connections are of the form p ◦ d + A for A ∈ X ⊗B Ω1
D(B) ⊗B X∗ an
endomorphism-valued one-form. For any (Hermitian) connection on the module X we
obtain a representative (C, X ⊗B H, 1 ⊗∇ D) of the Kasparov product with operator
1 ⊗∇ D : X ⊗B Dom(D) → X ⊗B H 1 ⊗∇ D(x ⊗ ξ) := γ(x) ⊗ Dξ + ∇D(x)ξ.
(0.3)
The key observation is then that the curvature operator is given by
(1 ⊗∇ D)2 − 1 ⊗∇ D2 = p[D, p][D, p]p + A2 + dA.
(0.4)
Here dA = πD(δAu) indicates an operator defined through choosing a lift Au of A to the
universal calculus and is independent of the choice of lift up to junk forms. Nevertheless,
the left hand side of Equation (0.4) is a well-defined, direct and constructive way of
representing the curvature of a module XB: all we require is the differential structure
provided by a spectral triple or unbounded Kasparov module.
Going beyond finitely generated modules, for countably generated C∗-B-modules we not
only require a differentiable structure induced from a differentiable structure on B, but
we also need to fix a regular operator S on X. This operator should be thought of as
defining a vertical differential structure on X. Hilbert C∗-modules are generalisations
of continuous fields of Hilbert spaces, and such fields are trivial if and only if they are
locally trivial. Thus, in order to detect nontrivial topological content, working at the
continuous level will not suffice. Differential structures on continuous fields of Hilbert
spaces are not naturally given, and need to be prescribed. This phenomena requires us
to talk about both horizontal and vertical differentiability, and as already noted, these
considerations are compatible with KK-factorisation.
Examples where curvature appears in the context of unbounded Kasparov theory is in the
factorisation of Dirac operators on Riemannian submersions and G-spectral triples [7, 9,
29]. We will review and illustrate our notion of curvature for Riemannian submersions in
4
Section 5. As we will see, the curvature operator contains the information of the second
fundamental form of the Riemannian submersion, the mean curvature associated to it, as
well as the curvature of the metric connection used in the Kasparov product. Note once
again that all this geometric information becomes available only at the unbounded level,
thus refining the topological information present at the level of bounded KK-theory.
Section 1 outlines both the algebraic and analytic aspects of differential forms. Section
2 outlines the analysis required to make sense of the curvature operator, and especially
second derivatives. Section 3 outlines the consequences for finitely generated projective
modules, Section 4 gives results for the special case of Grassmann connections, and in
Section 5 we outline the application to Riemannian submersions.
Notation
All algebras denoted by symbols A, B, C are assumed to be unital and trivally graded.
All modules denoted by symbols X, Y are assumed to be Z/2-graded, with grading
operator γX, γY or simply γ. Consequently, algebras of operators on such modules are
Z/2-graded as well. Homomorphisms between graded algebras are assumed to respect
the grading. All commutators [a, b] are graded commutators, which, for homogenous
elements a, b with degrees ∂a, ∂b respectively are defined to be
[a, b] := ab − (−1)∂a∂bba,
and extended by linearity. Sometimes we write [·,·]+ for anticommutators for emphasis.
We use various completed tensor products. The algebraic tensor product of modules will
be denoted ⊗alg, the completed tensor product of Hilbert C∗-modules will be denoted ⊗,
the completed projective tensor product of locally convex vector spaces will be denoted
(cid:98)⊗, and Haagerup tensor product of operator spaces will be denoted ⊗h.
Acknowledgements The authors acknowledge the support of the Erwin Schrodinger
Institute during the Thematic Programme Bivariant K-theory in geometry and physics
in November 2018, when a substantial part of this work was conducted. AR and BM
thank the Gothenburg Centre for Advanced Studies in Science and Technology for fund-
ing and the University of Gothenburg and Chalmers University of Technology for their
hospitality in November 2017 when this project took shape. We thank Alan Carey, Jens
Kaad and Giovanni Landi for inspiring feedback in the course of this work. BM thanks
Matthias Lesch for numerous conversations related to some of the technical aspects of
this work.
1 Universal differential forms for C 2-Kasparov modules
This section develops the necessary tools to talk about universal differential forms in the
context of unbounded Kasparov modules. We do not need forms of all degrees: in order
to talk about curvature, degrees one and two suffice and we restrict attention to these
degrees.
5
We carefully capture the corresponding C 1-and C 2-topologies in terms of suitable oper-
ator ∗-algebras (see [3]), which we will first introduce.
1.1 Operator ∗-algebras and differential structures
Throughout this section we will develop our approach to differential structure relative
to a fixed unbounded Kasparov module (B, Y, T ). This Kasparov module consists of
a complex ∗-algebra B, a Z2-graded Hilbert module over a C∗-algebra C, and an odd
self-adjoint operator T : Dom T → Y . This data satisfies the requirements:
1. there is an injective ∗-representation B → B(Y );
2. T : Dom T → H is a self-adjoint operator and b(T ± i)−1 ∈ K(Y ) for all b ∈ B;
3. for all b ∈ B it holds that b : Dom T → Dom T and [T, b] extends to a bounded
operator.
Remark 1.1. We note that the property of locally compact resolvent in point 2. is not
used anywhere in connection with the differential structure provided by T . So for the
purposes of differential structure, we may use any unbounded operator on the Hilbert
module Y . In particular, one may use indefinite Kasparov modules (non-commutative
analogues of pseudo-Riemannian geometries, [19]) to define differential structures.
We can always reduce to the case of operators on a Hilbert space. Given a Kasparov
module (B, YC, T ), we may choose an injective Hilbert space representation C → B(H).
Then we obtain the injective ∗-homomorphism K(Y ) → B(Y ⊗C H) by [32, Proposition
4.7] and since B(Y ) = M (K(Y )) this extends to an injective ∗-homomorphism B(Y ) →
B(Y ⊗C H). We may therefore consider (B, Y ⊗C H, T ⊗ 1).
We write B for the C∗-closure of B in the norm it inherits as an algebra of operators
on B(Y ). An operator space is a closed subspace of B(H), an operator algebra is an
operator space that is closed under multiplication in B(H), and an operator ∗-algebra is
an operator algebra that carries a completely isometric involution, [3, Definition 1.4].
Consider the algebra representation
T : B (cid:51) b (cid:55)→
π1
(cid:18) b
[T, b]
(cid:19)
0
b
∈ B(Y ⊕ Y ).
The representation π1
(cid:107)π1
T extends to matrices and satisfies the properties
b ∈ Mn(B),
T (b)∗(cid:107) = (cid:107)π1
(cid:107)b(cid:107) ≤ (cid:107)π1
T (b∗)(cid:107),
T (b)(cid:107),
and thus defines an operator ∗-algebra structure on the closure B1 of B in the norm
(cid:107)b(cid:107)1 := (cid:107)π1
T (b)(cid:107),
which is compatible with the C∗-norm on B in the sense that the inclusion B → B is a
completely contractive homomorphism of operator ∗-algebras.
We now present the additional C 2-condition the unbounded Kasparov module (B, Y, T )
is required to satisfy.
6
Definition 1.2. An unbounded Kasparov module (B, Y, T ) satisfies the C 2-condition if
there is a core C ⊂ Dom T 2 for T 2 such that for all b ∈ B we have
b : C → Dom T 2,
and the densely defined operator
[T 2, b](T ± i)−1 : (T ± i)C → Y,
extends to a bounded operator on Y .
(cid:18) a
Note that any core for T 2 is a core for T and that [T, b] extends to a bounded operator
for all b ∈ B since (B, Y, T ) is an unbounded Kasparov module.
The C 2-topology on B is defined as in [38]. A concrete description as an operator
∗-algebra comes from the algebra representations
T : a (cid:55)→
π2
(cid:18)(T + i)a(T + i)−1 0
∈ B(Y ⊕ Y ),
∈ B(Y ⊕ Y ).
(1.1)
T extends to matrices but is not directly compatible with the ∗-
[T 2, a](T + i)−1
The representation π2
structure (as in [38]), so we define the operator ∗-norm
0
[T, a] a
T : a (cid:55)→
π1
(cid:19)
(cid:19)
a
(cid:107)b(cid:107)2 := max(cid:8)(cid:107)π1
T (b)(cid:107),(cid:107)π2
T (b)(cid:107),(cid:107)π2
b ∈ Mn(B).
T (b∗)(cid:107)(cid:9),
The norm (cid:107)b(cid:107)2 is realised concretely in the representation
π(a) := π1
T (a) ⊕ π2
T (a) ⊕ π2
T (a∗)∗.
(1.2)
This gives the completion B2 of B in the norm (cid:107) · (cid:107)2 the structure of an operator ∗-
algebra [3]. By construction, the inclusions B2 → B1 → B are completely contractive
operator ∗-algebra homomorphisms. We now discuss several realisations of bimodules
of universal differential forms over B.
Remark 1.3. The norm (cid:107)·(cid:107)2 gives an analogue of a C 2-norm, and to define this norm we
needed the additional smoothness of the Kasparov module described in Definition 1.2.
In this light, an unbounded Kasparov module is a C 1-Kasparov module, having enough
smoothness to define the "C 1-norm" (cid:107) · (cid:107)1.
In the sequel we will make frequent use of the Haagerup tensor product for operator
spaces (see [4]). Given two operator spaces X and Y their Haagerup tensor product is
the completion of X ⊗alg Y in the norm
xix∗
(cid:88)
xi ⊗ yi
(cid:107)z(cid:107)2
(cid:111)
(cid:13)(cid:13)(cid:13) : z =
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:88)
(cid:110)(cid:13)(cid:13)(cid:13)(cid:88)
h := inf
which is denoted X⊗hY and can be shown to be an operator space again. It follows quite
directly from the definition that for closed subspaces A, B ⊂ B(H), the multiplication
map
y∗
i yi
,
i
m : A ⊗h B → B(H),
a ⊗ b (cid:55)→ ab,
is completely contractive, and this property motivates the definition of the Haagerup
norm (see [4, Theorem 2.3.2]) . For our purposes the following characterisation of ele-
ments in X ⊗h Y is of crucial importance.
7
Proposition 1.4 (Proposition 1.5.6 of [4]). Let X and Y be operator spaces, and X⊗h Y
the Haagerup tensor product.
1. Let z ∈ X ⊗h Y with (cid:107)z(cid:107)h < 1. Then z can be written as a norm convergent series
kyk are norm convergent series of norm
z =(cid:80) xk ⊗ yk, where (cid:80)
2. If the sequences (xk) ⊂ X and (yk) ⊂ Y are such that(cid:80)
and supN (cid:107)(cid:80)k≤N xkx∗
k and (cid:80) y∗
k(cid:107) < ∞, then(cid:80) xk ⊗ yk is norm convergent in X ⊗h Y .
kyk is norm convergent
< 1 in X and Y respectively;
k xkx∗
k y∗
If B is an algebra and X is a right- and Y a left- B module, the Haagerup module
tensor product X ⊗h
B Y is the quotient of X ⊗h Y by the closed subspace generated by
xb ⊗ y − x ⊗ by (see [4, Section 3.4.2]). Hilbert C∗-modules carry a natural operator
space structure inherited from the embedding into their linking algebra. The following
theorem characterizes the Haagerup module tensor product for Hilbert C∗-modules.
Theorem 1.5 (Theorem 4.3 of [2]). Let X and Y be Hilbert C∗-modules over C∗-algebras
B and C respectively, and suppose B → B(Y ) is a ∗-homomorphism. Then the Hilbert
C∗-module tensor product X⊗B Y is completely isometrically isomorphic to the Haagerup
module tensor product X ⊗h
We will use Theorem 1.5 result freely in the sequel. We use the symbol ⊗B for the
C∗-module tensor product, ⊗h
B for
the balanced algebraic tensor product.
B for the Haagerup module tensor product and ⊗alg
B Y .
1.2 Universal and represented differential forms for C 2-spectral triples
We wish to define bimodules of universal 1-forms and 2-forms associated to a C 2-
Kasparov module (B, Y, T ). To this end we use the Haagerup tensor product for the
unital operator ∗-algebras B2, B1 and B. To define universal one-forms we need to
consider the kernel of the multiplication map m : B × B → B restricted to suitable
subalgebras. We define three spaces of universal one-forms for the Kasparov module
(B, Y, T ).
m−→ B(cid:1)
m−→ B(cid:1)
m−→ B1
(cid:1),
Ω1
Ω1
Ω1
u(B, B2) := ker(cid:0)B ⊗h B2
u(B, B1) := ker(cid:0)B ⊗h B1
u(B1, B2) := ker(cid:0)B1 ⊗h B2
u(B, B1) (for instance) as(cid:80)
8
with m the multiplication map. As m is a complete contraction on each of these spaces,
the respective modules of forms are operator bimodules for the respective algebras.
i aiδ(bi) where δ(b) = 1⊗ b− b⊗ 1.
We denote elements of Ω1
Here we should take ai ∈ B and bi ∈ B1, with similar descriptions of the other bimodules
of one-forms.
1.2.1 Rough algebraic outline
Let us give a brief algebraic sketch of what we need our forms to do, so that the purpose
of the analysis to follow is clear. There is a map
πT : Ω1
u(B, B1) → B(Y ),
πT (aδ(b)) = a[T, b].
T (B, B1), and these are called the represented one forms. Re-
The range is denoted Ω1
stricting πT to the other bimodules gives different spaces of one-forms. To discuss cur-
vature we need to consider two-forms. Universally we have a few options, for instance,
Ω2
u(B, B2) := Ω1
u(B, B1) ⊗h
B1 Ω1
u(B1, B2).
The common factor of B1 allows us to use the Leibniz rule
aδ(b1)b2δ(c) = aδ(b1b2)δ(c) − ab1δ(b2)δ(c),
to see that all two-forms can be represented as sums (cid:80)
algebra elements. The universal differential δ : Ω1
is defined by
a ∈ B, b1, b2 ∈ B1, c ∈ B2,
u(B1, B2) → Ω1
i aiδ(bi)δ(ci) for appropriate
u(B1, B2)
u(B, B1)⊗h
Ω1
B1
aδ(b) = a ⊗ b − ab ⊗ 1 (cid:55)→ (1 ⊗ a − a ⊗ 1) ⊗ (1 ⊗ b − b ⊗ 1) = δ(a)δ(b),
= 1 ⊗ a ⊗ b − 1 ⊗ ab ⊗ 1 − a ⊗ 1 ⊗ b + a ⊗ b ⊗ 1
and satisfies δ2(a) = 0 for all a ∈ B2.
By declaring the symbol δ to be odd, we obtain a Z2-grading on the various spaces of
universal one-forms Ω1
Since the map πT is B-bilinear, and the Haagerup tensor product linearises operator
multiplication, we can also represent our two-forms in B(Y ) via
u. Since B is trivially graded, all elements of Ω1
u are odd.
m ◦ (πT ⊗ πT ) : Ω2
u(B, B2) (cid:51) aδ(b)δ(c) (cid:55)→ πT (aδ(b))πT (δ(c)) = a[T, b][T, c].
The map m◦(πT ⊗πT ) is compatible with ∗-structures as well if we define δ(a)∗ = −δ(a∗)
for a ∈ B1.
As is well-known, there is typically no differential d : Ω1
that πT ◦ δ = d ◦ πT . The naive formula
T (B1, B2) → Ω2
T (B, B2) such
(cid:33)
(cid:88)
i bi[T, ci] = 0 while (cid:80)
is not well-defined. It can happen that (cid:80)
(cid:32)(cid:88)
[T, bi][T, ci]
bi[T, ci]
=
i[T, bi][T, ci] (cid:54)= 0. The
T := πT (δ(ker(πT ))) are known as junk forms.
forms in J2
Below we will identify analytic spaces of represented one- and two-forms which will serve
as suitable receptacles for curvature.
d
i
i
9
1.2.2 The formal definitions of represented forms
We again fix a C 2-Kasparov module (B, Y, T ).
By construction, the operator ∗-algebra B1 acts completely contractively on the Hilbert
space Dom T equipped with the graph norm. Via this representation the C∗-algebra
B(Dom T ) becomes an operator B1-bimodule. Furthermore recall that the operator
norm on B(Dom T ) can be expressed as
(cid:107)R(cid:107)B(Dom T ) = (cid:107)(T + i)R(T + i)−1(cid:107)B(Y ),
R ∈ B(Dom T ),
which we will exploit in the proof of the following Lemma.
Lemma 1.6. The map b (cid:55)→ [T, b] defines a completely bounded derivation δT : B2 →
B(Dom T ). Hence there is a completely bounded map
πT : Ω1
u(B1, B2) → B(Dom T ),
a ⊗ b (cid:55)→ a[T, b].
The map b (cid:55)→ [T 2, b] extends to a completely bounded derivation
δT 2 : B2 → B(Dom T, Y ),
on B2. Hence there is a completely bounded map
u(B, B2) → B(Dom T, Y ),
πT 2 : Ω1
a ⊗ b (cid:55)→ a[T 2, b].
We can extend δT to a completely bounded map
δT : πT (Ω1
u(B1, B2)) → B(Dom T, Y )
and the composition δT ◦ πT : Ω1
δT (a[T, b]) := T (a[T, b]) + (a[T, b])T = [T, a][T, b] + a[T 2, b]
(1.3)
u(B1, B2) → B(Dom T, Y ) is a completely bounded map.
Proof. The estimate
ai[T, bi]
=
(cid:13)(cid:13)(cid:13)(cid:88)
i
ai[T, bi]
(cid:13)(cid:13)(cid:13)B(Dom T )
(cid:13)(cid:13)(cid:13)(T + i)
(cid:16)(cid:88)
≤(cid:13)(cid:13)(cid:13)(cid:88)
(cid:13)(cid:13)(cid:13)B(Y )
(T + i)−1(cid:105)(cid:13)(cid:13)(cid:13)B(Y )
(cid:17)
≤(cid:13)(cid:13)(cid:13)(cid:88)
(cid:13)(cid:13)(cid:13)B(Y )
(cid:13)(cid:13)(cid:13)B(Y )
ai[T 2, bi](T + i)−1(cid:13)(cid:13)(cid:13)B(Y )
(cid:13)(cid:13)(cid:13)(cid:88)
(T + i)−1(cid:13)(cid:13)(cid:13)B(Y )
(cid:17)
(cid:13)(cid:13)(cid:13)(cid:104)
(cid:16)(cid:88)
(cid:13)(cid:13)(cid:13)(cid:88)
i ai[T, bi](cid:107)B(Dom T ) ≤ 3(cid:107)(cid:80) ai ⊗ bi(cid:107)B1⊗hB2.
[T, ai][T, bi]
ai[T, bi]
ai[T, bi]
+
T,
ai[T, bi]
+
i
i
i
i
+
i
i
,
shows that (cid:107)(cid:80)
For b ∈ B2, the operator [T 2, b] is defined on Dom T 2 and [T 2, b](T + i)−1 : Dom T → Y
extends to a bounded operator on Y . Thus if ξn ∈ Dom T 2 → ξ ∈ Dom T in the graph
norm of T , then [T 2, b]ξn = [T 2, b](T + i)−1(T + i)ξn is convergent. The estimate
(cid:107)[T 2, b]ξ(cid:107) ≤ (cid:107)[T 2, b](T + i)−1(cid:107)(cid:107)(T + i)ξ(cid:107),
10
holds on the T -graph-norm dense subspace Dom T 2 ⊂ Dom T , and shows that the
derivation δT 2 is completely bounded as a map B2 → B(Dom T, Y ). We denote the cb-
norm of δT 2 by (cid:107)δT 2(cid:107)cb. The second statement now follows from the standard Haagerup
estimate
(cid:13)(cid:13)(cid:13)(cid:88)
i
ai[T 2, bi]
(cid:13)(cid:13)(cid:13)2 ≤(cid:13)(cid:13)(cid:13)(cid:88)
≤(cid:13)(cid:13)δT 2
(cid:13)(cid:13)2
cb
aia∗
(cid:13)(cid:13)(cid:13)(cid:88)
(cid:13)(cid:13)(cid:13)B
(cid:13)(cid:13)(bi)(cid:13)(cid:13)2
(cid:13)(cid:13)(ai)t(cid:13)(cid:13)2
i
i
B
.
B2
[T 2, bi]∗[T 2, bi]
(cid:13)(cid:13)(cid:13)B(Dom T,Y )
where (ai)t is a row vector and (bi) a column. Since πT : B1 ⊗h B2 → B(Y ), this proves
that δT ◦ πT (ω) is defined as an operator B(Dom T, Y ).
The appearance of commutators [T 2, b] for b ∈ B is a consequence of the natural norm
on the domain of T . Somewhat more heuristically, it can be considered as a consequence
of trying to extend the derivation b (cid:55)→ [T, b] to one-forms via the graded commutator
a[T, b] (cid:55)→ δT (a[T, b]) =(cid:2)T, a[T, b](cid:3)
+ = [T, a][T, b] + a[T 2, b].
Thus the graded commutator implements the ill-defined "differential" a[T, b] (cid:55)→ [T, a][T, b],
up to the unwanted unbounded term a[T 2, b], as in Equation (1.3).
Definition 1.7. For a C 2-Kasparov module (B, Y, T ) we denote by Ω1
closure of the image of the map
T (B1, B2) the
πT : Ω1
u(B1, B2) → B(Dom T ),
and by Ω1
T (B, B1) the closure of the image of the map
u(B, B1) → B(Y ).
πT : Ω1
T (B, B1) contains the module of differential forms Ω1
Note that Ω1
T defined by Connes
in [8, Chapter VI] as a dense subspsace (for which only finite linear combinations of
the a[T, b] are allowed). However, in the case that Ω1
T is a finitely generated projective
(right) B-module, they coincide.
Definition 1.8. For a C 2-Kasparov module (B, Y, T ) we denote by Ω2
of the image of the map m ◦ (πT ⊗ πT ), defined as the composition
u(B, B1) ⊗h
Ω1
u(B, B1) πT ⊗πT−−−−→ B(Y ) ⊗h
u(B1, B2) → Ω1
u(B, B1) ⊗h
B1 Ω1
B1 Ω1
B1
T (B1) the closure
B(Y ) m−→ B(Y ).
(1.4)
Lemma 1.9. The closure of the range of the map
m ◦ (πT ⊗ πT ) : Ω1
coincides with the space of two-forms
u(B, B1) ⊗h
B1 Ω1
Ω2
T (B1) =
over B1.
u(B1, B2) → B(Y )
(cid:111)
ai[T, bi][T, ci] : ai ∈ B, bi, ci ∈ B1
,
(cid:110)(cid:88)
11
(cid:88)
since(cid:80)
Proof. To see that the closure of the range coincides with Ω2
ai[T, bi]cj[T, dj] =
(cid:88)
(cid:88)
i aibi = 0. Hence we have an inclusion ran m ◦ (πT ⊗ πT ) ⊂ Ω2
ai[T, bicj][T, dj] − aibi[T, cj][T, dj] =
i,j
i,j
i,j
T (B1). Since any
c ∈ B1 is a limit of ci ∈ B2 such that [T, ci] → [T, c], all expressions a[T, b][T, c] are in
the range of m ◦ (πT ⊗ πT ) and the reverse inclusion follows as well.
T (B1), consider a finite sum
ai[T, bicj][T, dj],
1.3 Second derivatives and junk
We now provide a discussion of junk forms for Kasparov modules in our analytic context.
The rather strange representation of forms given by πT 2 turns out to be precisely what
is required to cancel out the unwanted term in the anticommutator [T, a[T, b]]+. In turn,
this cancellation allows us to both represent curvature and capture junk.
Proposition 1.10. For any ω ∈ Ω1
u(B1, B2) we have that
δT ◦ πT (ω) − πT 2(ω) = m ◦ (πT ⊗ πT )(δω) ∈ Ω2
T (B1) ⊂ B(Y ).
T (B1) is completely contractive. In particular
u(B1, B2) such that πT (ω) = 0, the operator πT 2(ω) is well-defined and bounded
u(B1, B2) → Ω2
The map δT ◦ πT − πT 2 : Ω1
for ω ∈ Ω1
on Y .
Proof. Let α = (αij) ∈ MN (Ω1
u(B1, B2)). Approximate α by a series
u(B1, B2)(cid:1),
(cid:0)(B1 ⊗alg B2) ∩ Ω1
(cid:32) n(cid:88)
n(cid:88)
(1.5)
(cid:33)
using Proposition 1.4.1. By abuse of notation we denote the self-adjoint regular operator
diag T : (Dom T )N → Y N by T and we identify MN (B(Y )) with B(Y N ). The identity
[T, πT (αn)] =
T,
aik[T, bkj]
=
[T, aik][T, bkj] +
k=1
k=1
aik[T 2, bkj]
,
ij
is valid in B(Dom T, Y N ). By continuity of δT ◦πT the left hand side of this equation con-
verges (as n → ∞) in B(Dom T, Y N ). By continuity of πT 2 we have πT 2(αn) → πT 2(α) in
k=1[T, aik][T, bkj])ij is convergent in B(Dom T, Y N ). Since
there is a complete contraction B2 → B1 we have the estimate
(cid:13)(cid:13)(bkj)(cid:96)
(cid:13)(cid:13)B1
≤(cid:13)(cid:13)(aik)(cid:96)
(cid:13)(cid:13)B1
(cid:13)(cid:13)(bkj)(cid:96)
(cid:13)(cid:13)B2
,
k=m
k=m
k=m
B(Dom T, Y N ). Therefore ((cid:80)n
(cid:13)(cid:13)(cid:13)B(Y N )
(cid:13)(cid:13)(cid:13)(cid:16) m(cid:88)
≤(cid:13)(cid:13)(aik)(cid:96)
(cid:13)(cid:13)B1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)B(Y N )
(cid:33)
(cid:17)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:32) m(cid:88)
from which we infer that
[T, aik][T, bkj]
[T, aik][T, bkj]
k=m
k=(cid:96)
k=(cid:96)
ij
ij
≤ (cid:107)αk − αm(cid:107)MN (Ω1
u(B1,B2)) → 0.
αn :=
aik ⊗ bkj
∈ MN
k=1
ij
(cid:32) n(cid:88)
(cid:32)(cid:104)
n(cid:88)
k=1
(cid:33)
(cid:33)
(cid:105)
+
ij
12
It thus follows that the series (cid:32)(cid:88)
(cid:33)
[T, aik][T, bkj]
,
is norm convergent in B(Y N ), proving that δT ◦ πT (ω) − πT 2(ω) ∈ MN (Ω2
The above estimates show that
T ) ⊂ B(Y N ).
k
ij
(cid:107)(δT ◦ πT − πT 2)(ω)(cid:107)B(Y N ) ≤ (cid:107)ω(cid:107)MN (Ω1
u(B1,B2)), ω ∈ MN (Ω1
u(B1, B2)),
which proves that δT ◦ πT − πT 2 is completely contractive. In case ω ∈ Ω1
u(B1, B2) and
πT (ω) =(cid:80)
(cid:88)
(cid:88)
i
i ai[T, bi] = 0 we find that
πT 2(ω) =
=
ai[T 2, bi] =
[ai, T ][T, bi] +
aiT [T, bi] + ai[T, bi]T
T ai[T, bi] = −(cid:88)
i
i
i
(cid:88)
(cid:88)
i
[T, ai][T, bi] ∈ B(Dom T, Y ).
This proves that πT 2(ω) extends to a bounded operator on Y .
T 2(B1, B2) and πT : Ω1
Corollary 1.11. Let πT 2 : Ω1
be the universal maps. Then ker πT is a closed subbimodule of Ω1
bimodule of junk forms
u(B1, B2) → Ω1
(cid:88)
(cid:88)
(cid:40)(cid:88)
J2
T (B1) :=
[T, ai][T, bi] :
aibi =
ai[T, bi] =
u(B1, B2) → Ω1
T (B1, B2)
u(B1, B2) and the closed
(cid:41)
(cid:88)
[T, ai]bi = 0
,
is equal to the closure of the space πT 2(ker πT ).
i
2 Curvature in unbounded KK-theory
We now come to the main construction in this paper, which is the notion of curvature
in the context of the unbounded Kasparov product. As mentioned in the introduction,
it is our goal to make sense of the following formula
R(S,∇T ) = (S ⊗ 1 + 1 ⊗∇ T )2 − S2 ⊗ 1 − 1 ⊗∇ T 2.
For this we again fix an unbounded Kasparov module (B, Y, T ) which will provide our
reference 'horizontal' differential structure. For the most part we make no use of the
(locally) compact resolvent of T , only occasionally (and always explicitly) requiring that
(B, Y, T ) defines a KK-class. More important are the various modules of forms, junk
and representations
Ω1
T (B∗, B∗),
JT ,
πT ,
πT 2
defined as in the last section.
In order to define the curvature we need to introduce a suitable notion of C 2-connection,
along with some 'vertical' differentiability conditions on the C∗-module X. As we will
see, the latter is phrased naturally in terms of the self-adjoint regular operator S on X.
13
2.1 The definition of C 1 and C 2-connections
In order to define curvature we need a suitable notion of C 2-connection. We require the
notion of "form-valued inner products". If we have x, y ∈ X and ω ∈ Ω1
u(B1, B2) (for
instance), we define pairings
(cid:104)x, y ⊗ ω(cid:105) := (cid:104)x, y(cid:105) ⊗ ω,
and (cid:104)x ⊗ ω, y(cid:105) := ω∗ ⊗ (cid:104)x, y(cid:105),
(2.1)
which are compatible with the balancing relation over B1.
Definition 2.1. Let X be a Z2-graded C∗-module over B and (B, Y, T ) a C k- un-
bounded Kasparov module (k = 1, 2). A dense Bk-submodule X ⊂ X is horizontally
C k-differentiable with respect to (B, Y, T ) if for all x, y ∈ X it holds that (cid:104)x, y(cid:105) ∈ Bk.
Remark 2.2. We will develop curvature with the minimal smoothness assumptions that
we can, working with horizontally differentiable C 1-modules where possible, and impos-
ing further C 2-structure as we need it.
Definition 2.3. A universal connection on X is a linear map, so if γ is the grading
operator of X then
∇ : X → X ⊗h
u(B, B1), ∇(γ(x)) = (γ ⊗ 1)∇(x),
B Ω1
which satisfies the Leibniz rule
∇(xb) = ∇(x)b + γ(x) ⊗ δ(b),
∀x ∈ X b ∈ B1.
If in addition ∇ satisfies
(cid:104)γ(x),∇(y)(cid:105) − (cid:104)∇(γ(x)), y(cid:105) = δ((cid:104)x, y(cid:105)),
∀x, y ∈ X,
then ∇ is said to be Hermitian or compatible.
We write ∇T := πT ◦ ∇ : X → X ⊗h
connection induced by ∇.
Lemma 2.4. Let X be a horizontally differentiable C 1-module for the unbounded Kas-
parov module (B, Y, T ) and ∇ : X → X ⊗h
T (B1) and call the composition the represented
u(B, B1) a Hermitian connection. Then
B Ω1
1 ⊗∇ T : X ⊗alg
B Ω1
B Dom T → X ⊗B Y,
x ⊗ y (cid:55)→ γ(x) ⊗ T y + ∇T (x)y,
is a densely-defined odd symmetric operator in X ⊗B Y .
Proof. On X ⊗B Ω1
γ ⊗ γ = (γ ⊗ 1)(1 ⊗ γ), for x ⊗ y ∈ X ⊗alg
B Dom T we have
u(B, B1) it holds that (1 ⊗ γ)(x ⊗ ωT )y = −x ⊗ ωT γ(y). Since
(γ ⊗ γ)(1 ⊗∇ T )(x ⊗ y) = (γ ⊗ γ)(γ(x) ⊗ T y + ∇T (x)y)
= −x ⊗ T γ(y) − (γ ⊗ 1)∇T (x)γ(y)
= −x ⊗ T γ(y) − ∇T (γ(x))γ(y)
= −(1 ⊗∇ T )(γ ⊗ γ)(x ⊗ y),
14
so 1 ⊗∇ T is an odd operator. For symmetry, we write
(cid:10)(1 ⊗∇ T )(x ⊗ y), x ⊗ y(cid:11) =(cid:10)(cid:104)x, γ(x)(cid:105)T y, y(cid:105) + (cid:104)∇T (x)y, x ⊗ y(cid:11)
= −(cid:10) [T,(cid:104)γ(x), x(cid:105)] y, y(cid:105) + (cid:104)(cid:104)γ(x), x(cid:105)y, T y(cid:105) + (cid:104)∇T (x)y, x ⊗ y(cid:11)
=(cid:10)(cid:104)∇T (x), x(cid:105)y, y(cid:105) − (cid:104)(cid:104)x,∇T (x)(cid:105)y, y(cid:105) + (cid:104)y,(cid:104)x, γ(x)(cid:105)T y(cid:105) + (cid:104)∇T (x)y, x ⊗ y(cid:11)
=(cid:10)y,(cid:104)x,∇T (x)(cid:105)y(cid:105) + (cid:104)y,(cid:104)x, γ(x)(cid:105)T y(cid:11)
=(cid:10)x ⊗ y, γ(x) ⊗ T y(cid:105) + (cid:104)x ⊗ y,∇T (x)y(cid:11)
=(cid:10)x ⊗ y, (1 ⊗∇ T )(x ⊗ y)(cid:11).
Polarisation completes the proof.
A horizontal C 1-structure is all one requires for finitely generated modules, but it is not
sufficient to describe the connections and curvature of countably generated modules. In
order to define appropriate notions of connections, we need to introduce some further
smoothness on the C∗-module X.
Let S : Dom S → X be a self-adjoint regular operator on X. We think of S as defining
a vertical differential structure on X. The presence of a compatible horizontal differ-
entiable structure for a Kasparov module (B, Y, T ) then provides us with the correct
notion of differentiable submodule.
We do not require that the data (X, S) define an unbounded Kasparov module, although
this is often the case in examples.
Definition 2.5. Let (B, Y, T ) be a C 1-Kasparov module. A C 1-module for (B, Y, T ) is
a pair (X, S) such that
1. X is a horizontally differentiable C 1-module with C∗-closure X;
2. S : X → X is an essentially self-adjoint and regular operator on X.
A C 1-connection on a C 1-module (X, S) is a Hermitian connection
∇ : X → X ⊗B Ω1
Dom T → X ⊗B Y is essentially self-adjoint and regular.
u(B, B1),
such that S⊗ 1 + 1⊗∇ T : X⊗alg
Remark 2.6. The choice S = 0 is allowed for defining a vertical differential structure.
Given a C 1-Hermitian connection ∇ : X → X ⊗h
of X by
X∇T :=(cid:8)x ∈ X : ∃xn ∈ X, xn → x ∈ X, ∇T (xn) → ∇T (x) ∈ X ⊗h
T (B1)(cid:9). (2.2)
u(B, B1) we can define a completion
B Ω1
B1
B Ω1
Proposition 2.7 ([3] Section 3.6, Lemma 3.4). Let (B, Y, T ) be a C 1-Kasparov module
and (X, S) a C 1-module. The space X∇T is an operator ∗-module over B1. Moreover,
for every x ∈ X∇T the linear map x(cid:105) : Y → X ⊗B Y , y (cid:55)→ x ⊗ y satisfies
(cid:0)γ(x)(cid:105)T − 1 ⊗∇ Tx(cid:105)(cid:1) ∈ B(Y, X ⊗B Y ).
x(cid:105) : Dom T → Dom 1 ⊗∇ T,
and
15
We define X S to the completion of X in the norm
(cid:107)x(cid:107)S := (cid:107)(S + i)−1x(cid:107) = (cid:107)(S − i)−1x(cid:107).
The norm on X S is induced from the inner product
(cid:104)x, y(cid:105)S := (cid:104)(S + i)−1x, (S + i)−1y(cid:105),
and X S is a Hilbert C∗-module in this inner product. It is appropriate to think of X S
as a degree −1 Sobolev space associated to S, as the following observation shows.
Lemma 2.8. For x ∈ X the maps x (cid:55)→ (S + i)−1x, x (cid:55)→ (S − i)−1x can be viewed as
densely defined maps X S → X and these extend to unitary isomorphisms X S → X.
Remark 2.9. One naturally expects that defining second derivatives would require two
1 → L2. In order to accommodate countably gen-
Sobolev spaces, normally W 2
erated modules and Kasparov modules with only a C 1-structure, we take a slightly
that behaves like a −1 Sobolev space in the
different route and introduce a space X S∇T
vertical direction but a +1 Sobolev in the horizontal direction.
2 → W 2
to be the completion of X in the operator space topology induced by the
We define XS∇T
norm
(cid:107)x(cid:107)S∇T
:= max
(cid:110)(cid:107)(S + i)−1x(cid:107)X,(cid:107)(S + i)−1∇T (x)(cid:107)X⊗h
(cid:111)
BΩ1
T (B,B1)
.
(2.3)
, the identity map on X extends
Since the norm on X∇T dominates the norm on XS∇T
to a complete contraction ιS : X∇T → XS∇T
. To define the appropriate notion of C 2-
connection, we need the operators S ⊗ 1 and 1 ⊗∇ T to be compatible in a more precise
way. We introduce the following definition.
Definition 2.10 (cf.
Hilbert C∗-module E. We say that (s, t) is a vertically anticommuting pair if
[27, 37, 39]). Let (s, t) be self-adjoint regular operators in the
1. (s ± i)−1 : Dom t → F(s, t) := {e ∈ Dom s ∩ Dom t : se ∈ Dom t, te ∈ Dom s};
2. st + ts : F(s, t) → E extends to Dom s.
Remark 2.11. The definition of vertically anticommuting pair is an asymmetric version
of [37, Definition 2.1] and was used in [27, 39]. The main result is that s+t is self-adjoint
and regular on Dom s∩ Dom t. In this paper we require the more restrictive asymmetric
version, which is sufficient to cover many geometric examples, is compatible with the
unbounded Kasparov product and seems to be necessary for technical reasons.
Lemma 2.12. Let (X, S) be a C 1-module over the Kasparov module (B, Y, T ) and ∇ :
X → X⊗h
T (B, B1) a connection such that (S⊗1, 1⊗∇T ) is a vertically anticommuting
→ X S.
pair. Then the identity map extends to a completely contractive injection XS∇T
B Ω1
Proof. Since the norm on XS∇T
dominates the norm on X S, the fact that the identity
map on X extends to a complete contraction is immediate. To see that this map is
16
injective, let xn ∈ X be a sequence such that (S + i)−1xn → 0 and (S + i)−1∇T (xn) is
convergent. For x ∈ X and y ∈ Dom T it holds that
∇T (x)y = (1 ⊗∇ T )(x ⊗ y) − γ(x) ⊗ T y.
Using this and the fact that (S ⊗ 1, 1 ⊗∇ T ) is a vertically anticommuting pair we write
(S + i)−1∇T (xn)y = (S + i)−1(1 ⊗∇ T )(xn ⊗ y) − (S + i)−1γ(xn) ⊗ T y
= −(1 ⊗∇ T )(S − i)−1(xn) ⊗ y + γ((S − i)−1xn) ⊗ T y
+ (S + i)−1[S ⊗ 1, 1 ⊗∇ T ](S − i)−1xn ⊗ y.
Since (S − i)−1xn is convergent to 0 and (S + i)−1[S ⊗ 1, 1 ⊗∇ T ] is bounded, it follows
that (1 ⊗∇ T )(S − i)−1(xn) ⊗ y is convergent. Since 1 ⊗∇ T is closed, the limit of the
latter sequence must be 0. Thus (S + i)−1∇T (xn)y converges to 0 for y ∈ Dom T . Since
(S + i)−1∇T (xn) is convergent in B(Y, X ⊗B Y ), its limit must be 0.
Before introducing the curvature operator we require some technical domain results.
Lemma 2.13. Let (X, S) be a C 1-module over the Kasparov module (B, Y, T ) and let
T (B, B1) be a connection such that (S ⊗ 1, 1 ⊗∇ T ) is a vertically
∇ : X → X ⊗h
anticommuting pair. Then:
B Ω1
1. (1⊗∇ T )(S + i)−1 ⊗ 1 is defined on Dom 1⊗∇ T and closable. There is an equality
Dom(cid:0)(1 ⊗∇ T )(S ± i)−1 ⊗ 1(cid:1) = Dom(cid:0)(S ∓ i)−1 ⊗ 1(1 ⊗∇ T )(cid:1),
of domains
of closed operators in X ⊗B Y .
→ X S ⊗B Ω1
2. ∇ : X → X ⊗B Ω1
3. (S+i)−1⊗1 : X S⊗BY → X⊗BY restricts to a map (S+i)−1⊗1 : XS∇T
T (B, B1) extends to a map ∇ : XS∇T
T (B, B1).
⊗h
B1
Dom T →
Dom(1 ⊗∇ T ).
Proof. In the following we will frequently write S for S ⊗ 1. We first prove that
Dom(cid:0)(1 ⊗∇ T )(S + i)−1(cid:1) = Dom(cid:0)(S − i)−1(1 ⊗∇ T )(cid:1).
Since (S ⊗ 1, 1 ⊗∇ T ) is a vertically anticommuting pair, both operators are initially
defined on Dom 1 ⊗∇ T and
(S − i)−1(1 ⊗∇ T ) + (1 ⊗∇ T )(S + i)−1 = (S − i)−1[1 ⊗∇ T, S](S + i)−1,
is a bounded operator. Hence if ξn → ξ in X ⊗B Y then (1 ⊗∇ T )(S + i)−1ξn converges
if and only if (S − i)−1(1 ⊗∇ T )ξn converges. So both operators are closable and have
the same closure.
2. Suppose that xn ∈ X is a sequence converging to x ∈ XS∇T
, that is (S + i)−1xn and
(S + i)−1∇T (xn) are both Cauchy sequences. Thus ∇T (xn) converges to an element
17
B Ω1
T (B, B1) is well defined for x ∈ XS∇T
T (B, B1). By Lemma 2.12, if xn → 0 ∈ XS∇T
z ∈ X S ⊗h
x (cid:55)→ ∇T (x) := z ∈ X S ⊗B Ω1
3. We use that (S⊗1, 1⊗∇ T ) is a vertically anticommuting pair. Suppose that x ∈ XS∇T
which by construction is a submodule of X S, so that (S + i)−1x ∈ X.
Take a finite row ξ = (x1, . . . , xn) with xi ∈ X∇T and a finite column η = (y1, . . . , yn)t
with yj ∈ Dom T . Observe that
then z = 0 and the map
.
,
(S + i)−1xi ⊗ yi ∈ F(S ⊗ 1, 1 ⊗∇ T ) ⊂ Dom 1 ⊗∇ T,
i = 1, . . . , n,
and write ξ ⊗ η =(cid:80)n
i=1 xi ⊗ yi ∈ XS∇T
⊗alg
Dom T . Now consider
1 ⊗∇ T (S + i)−1ξ ⊗ η = (S − i)−1[S ⊗ 1, 1 ⊗∇ T ](S + i)−1ξ ⊗ η
B1
− (S − i)−1(1 ⊗∇ T )(ξ ⊗ η)
= (S − i)−1[S ⊗ 1, 1 ⊗∇ T ](S + i)−1ξ ⊗ η
− (S − i)−1γ(ξ) ⊗ T η − (S − i)−1∇T (ξ)η
= (S − i)−1[S ⊗ 1, 1 ⊗∇ T ](S + i)−1ξ ⊗ η
+ γ((S + i)−1ξ) ⊗ T η − (S − i)−1∇T (ξ)η,
from which, using Theorem 1.5, we obtain the estimate
(cid:13)(cid:13)1 ⊗∇ T (S + i)−1ξ ⊗ η(cid:13)(cid:13)X⊗BY ≤ C(cid:107)(S+i)−1ξ(cid:107)X(cid:107)η(cid:107)Y + (cid:107)(S + i)−1ξ(cid:107)X(cid:107)T η(cid:107)Y
+ (cid:107)(S − i)−1∇T (ξ)(cid:107)X⊗h
T (B,B1)(cid:107)η(cid:107)Y .
BΩ1
(2.4)
The estimate (2.4) implies that
and thus that
(cid:13)(cid:13)1 ⊗∇ T (S + i)−1ξ ⊗ η(cid:13)(cid:13)X⊗BY ≤ C(cid:107)ξ(cid:107)XS∇T
(cid:13)(cid:13)1 ⊗∇ T (S + i)−1ξ ⊗ η(cid:13)(cid:13) ≤ C(cid:107)ξ ⊗ η(cid:107)XS∇T
(cid:107)η(cid:107)Dom T ,
⊗h
B1
Dom T .
Since X∇T ⊗alg
B1
Dom T is dense in XS∇T
⊗h
B1
Dom T , the result follows.
Suppose now that we are given a map
∇S : X → XS∇T
⊗h
B1 Ω1
u(B1, B2),
satisfying the Leibniz rule ∇S(xb) = ∇S(x)b + γ(x)⊗ δ(b) for all b ∈ B2. Then we obtain
a well-defined operator
1 ⊗∇S T : X ⊗alg
T (x)y.
By composing 1 ⊗∇S T with the resolvent (S + i)−1 : X S → X, which is defined on all
of X S by Lemma 2.8, we obtain a well-defined map
x ⊗ y (cid:55)→ γ(x) ⊗ T y + ∇S
B1 Y → X S ⊗B Y,
⊗h
Dom T → XS∇T
B1
(S + i)−1 · 1 ⊗∇S T : X ⊗alg
B1
Dom T → X ⊗B Y.
18
Definition 2.14. Let (X, S) be a C 1-module over the C 1-Kasparov module (B, Y, T ).
By a C 2-connection on X we mean a pair (∇,∇S) of connections
∇ : X → X ⊗h
⊗h
B1 Ω1
with ∇ a Hermitian connection and ∇S a connection, such that
u(B, B1), ∇S : X → XS∇T
B Ω1
u(B1),
1. (S ⊗ 1, 1 ⊗∇ T ) is a vertically anticommuting pair;
2. for all x ∈ X and y ∈ Dom T we have
(S + i)−1∇T (x)y = (S + i)−1 · ∇S
T (x)y ∈ X ⊗B Y.
Note that this definition implies that for all x ∈ X and y ∈ Dom T it holds that
(S + i)−1 · (1 ⊗∇S T )(x ⊗ y) ∈ Dom S and thus that 1 ⊗∇S T (x ⊗ y) = 1 ⊗∇ T (x ⊗ y) is
in fact an element of X ⊗B Y viewed as a subspace of X S ⊗B Y via the dense inclusion
X → X S.
2.2 The represented curvature of a C 2-connection on C 1-module
Now that we have a clear picture of how represented C 2-connections determine the
domains of induced operators, we can introduce the represented curvature of a C 2-
connection on a C 1-module (X, S). To make appropriate sense of πT (∇2) a little care
is required and this operator will more correctly be written πT ((1 ⊗∇T δ) ◦ ∇S), which
yields a well-defined operator.
Proposition 2.15. Let (∇,∇S) be a C 2-connection on a C 1-module (X, S) over a C 1-
Kasparov module (B, Y, T ). The map
1 ⊗∇T δ : XS∇ ⊗h
B1 Ω1
u(B1) → X S ⊗h
x ⊗ ω (cid:55)→ ∇T (x) ⊗ ω + γ(x) ⊗ (πT ⊗ 1)(δω),
T (B, B1) ⊗h
u(B1),
B1 Ω1
B Ω1
is well-defined and satisfies (1 ⊗∇T δ)(ηb) = (1 ⊗∇T δ)(η)b − (γ ⊗ πT )(η) ⊗ δ(b), for all
η ∈ XS∇ ⊗h
u(B1) and b ∈ B1.
Ω1
Proof. The map ∇T : X → X ⊗h
T (B, B1) extends to a map
B1
B Ω1
∇T : XS∇T
by part 2 of Lemma 2.13. The maps
→ X S ⊗h
B Ω1
T (B, B1),
δ : Ω1
u(B1) → Ω1
u(B, B1) ⊗h
B1 Ω1
u(B1)
and
πT ⊗ 1 : Ω1
u(B, B1) ⊗h
B1 Ω1
u(B1) → Ω1
T (B, B1) ⊗h
B1 Ω1
u(B1)
19
are completely bounded, and therefore
(1 ⊗ πT ⊗ 1) ◦ (∇ ⊗ 1 + γ ⊗ δ) : XS∇ ⊗hC Ω1
u(B1) → X S ⊗h
x ⊗ ω (cid:55)→ ∇T (x) ⊗ ω + γ(x) ⊗ (πT ⊗ 1)(δω),
(2.5)
T (B, B1) ⊗h
u(B1)
B1 Ω1
B Ω1
is well-defined and completely bounded. Since also
∇T (xb) ⊗ ω + γ(x) ⊗ (πT ⊗ 1)(bδω) = ∇T (x)bω + γ(x) ⊗ (δT b) ⊗ ω
+ γ(x) ⊗ (πT ⊗ 1)(bδω)
= ∇T (x)bω + γ(x) ⊗ (πT ⊗ 1)(δ(bω)),
the map (2.5) is compatible with the balancing relation and descends to a completely
bounded map
1 ⊗∇T δ : XS∇T
⊗h
B1 Ω1
u(B1, Bk) → X S ⊗h
B Ω1
T (B, B1) ⊗h
B1 Ω1
u(B1),
which is the desired statement. The (graded) Leibniz rule for 1 ⊗∇T δ follows directly
from the defining formula and the graded Leibniz rule for δ, δ(ωb) = δ(ω)b − ω ⊗ δ(b),
for ω ∈ Ω1
u(B1) and b ∈ B1.
By Proposition 2.15, we can consider the composition
(1 ⊗∇T δ) ◦ ∇S : X → X S ⊗h
T (B1) ⊗h
B Ω1
B1 Ω1
u(B1).
The map (1 ⊗∇T δ) ◦ ∇S is right B1-linear. This follows from the computation
(1 ⊗∇T δ) ◦ ∇S(xb) = (1 ⊗∇T δ)(∇S(x)b) + (1 ⊗∇T δ)(γ(x) ⊗ δ(b))
= (1 ⊗∇T δ)(∇S(x)b) + ∇T (γ(x)) ⊗ δ(b)
= (1 ⊗∇T δ)(∇S(x))b − (γ ⊗ πT ⊗ 1)(∇S(x) ⊗ δb) + ∇T (γ(x)) ⊗ δ(b)
= (1 ⊗∇T δ)(∇S(x))b,
where the last line holds since ∇T (x) = ∇S
Definition 2.16. Let (X, S) be a C 1-module over a C 1-Kasparov module (B, Y, T ) and
(∇,∇S) a C 2-connection. We define the represented curvature of (∇,∇S) to be the
operator πT (∇ ◦ ∇S) : X ⊗alg
Y → X S ⊗B Y defined on x ⊗ y ∈ X ⊗alg
T (x) in X S ⊗h
T (B, B1).
Y by
πT (∇ ◦ ∇S)(x ⊗ y) := (1 ⊗ m)(1 ⊗ 1 ⊗ πT )((1 ⊗∇T δ) ◦ ∇S)(x)y.
(2.6)
The notation πT (∇◦∇S) is a convenient shorthand for (1⊗m)(1⊗1⊗πT )((1⊗∇T δ)◦∇S).
B Ω1
B1
B1
20
2.2.1 The curvature operator of a C 2-connection on a C (1,2)-module
Additional smoothness simplifies the domain considerations of the previous section. In
order to denote the differentiability properties of modules (X, S) in the horizontal and
vertical directions we use pairs (n, k) where k corresponds to the horizontal and n to the
vertical direction.
Definition 2.17. Let (B, Y, T ) be a C 2-Kasparov module. A C (1,2)-module for (B, Y, T )
is a pair (X, S) such that such that
1. X is a horizontally differentiable C 2-module with C∗-closure X;
2. X ⊂ Dom S.
∇ : X → X ⊗h
B Ω1
u(B, B2), ∇S : X → XS∇T
A C 2-connection on a C (1,2)-module (X, S) is a pair (∇,∇S) of connections
u(B1, B2),
⊗h
B1 Ω1
such that (S ⊗ 1, 1 ⊗∇ T ) is a vertically anticommuting pair.
First observe that if (X, S) is a C (1,2)-module then (X ⊗alg
B1, S) is a C 1-module for
(B, Y, T ). For a C 2-connection on a C (1,2)-module (X, S), we wish to compute πT (∇◦∇S),
for y ∈ Dom T 2. By Lemmas 1.6 and 2.12 we may define
1 ⊗∇S T 2 : X ⊗alg
T 2(x)y,
in the spirit of Lemma 2.4. Since T 2 is an even operator, the grading γ does not appear
in this formula.
x ⊗ y (cid:55)→ x ⊗ T 2(y) + ∇S
⊗h
B1 Y ⊂ X S ⊗B Y,
Dom T 2 → XS∇T
B2
B2
Lemma 2.18. Let (B, Y, T ) be a C 2 Kasparov module, (X, S) a C (1,2)-module and
(∇,∇S) a C 2-connection. Then the operator
1 ⊗∇ T : X ⊗alg
B1
Dom T → X ⊗B Y,
Dom T 2 into Dom(S + i)−1(1 ⊗∇ T ).
maps X ⊗alg
Proof. For x ∈ X and y ∈ Dom T 2 we have
B2
(1 ⊗∇ T )(x ⊗ y) = γ(x) ⊗ T y + ∇T (x)y,
and clearly γ(x)⊗T y ∈ X⊗alg
Dom T ⊂ Dom(1⊗∇T ) ⊂ Dom(S +i)−1(1⊗∇T ). Consider
B1
(S + i)−1∇T (x)y = (S + i)−1 · ∇S
T (x)y,
and observe that ∇S
T (x)y ∈ XS∇T
⊗h
B1
(S + i)−1∇S
Dom T. By part 3 of Lemma 2.13, it follows that
T (x)y ∈ Dom 1 ⊗∇ T,
from which we conclude that (S + i)−1∇T (x)y ∈ Dom 1 ⊗∇ T . Thus
∇T (x)y ∈ Dom(1 ⊗∇ T )(S + i)−1 = Dom(S + i)−1(1 ⊗∇ T ),
by Lemma 2.13.
21
It follows by Lemma 2.13.2 that the operator
1 ⊗∇S T : X ⊗alg
B2
Dom T 2 → XS∇T
⊗h
B1 Dom T ⊂ Dom(1 ⊗∇ T ) ⊂ X S ⊗B Y,
is well defined and maps X ⊗alg
Definition 2.19. Let (B, Y, T ) be a C 2-Kasparov module, (X, S) a C (1,2)-module and
(∇,∇S) a C 2-connection. The curvature operator of (∇,∇S) is defined to be the map
Dom T 2 into Dom(1 ⊗∇ T ) ⊂ X S ⊗B Y .
B2
R∇T : X ⊗alg
Dom T 2 → X S ⊗B Y, R∇T := (1 ⊗∇ T ) ◦ (1 ⊗∇S T ) − 1 ⊗∇S T 2.
By Lemma 2.18, composition with the resolvent (S + i)−1 : X S → X yields the map
B2
(S + i)−1R∇T : X ⊗alg
B2
Dom T 2 → X ⊗B Y,
which admits the expression (S + i)−1R∇T = (S + i)−1((1 ⊗∇ T )2 − 1 ⊗∇S T 2).
Our goal is to identify the curvature operator R∇T from Definition 2.19 with the repre-
sented curvature πT (∇T ◦ ∇S) of Equation (2.6) in Definition 2.16.
Lemma 2.20. Let (∇,∇S) be a C 2-connection on a C (1,2)-module (X, S). For η ∈
XS∇T
u(B1, B2) and y ∈ Dom T it holds that
Ω1
⊗h
(1 ⊗ m)(1 ⊗ 1 ⊗ πT )((1 ⊗∇T δ)(η))y = (1 ⊗∇ T )(1 ⊗ πT )(η)y
B1
+ (γ ⊗ πT )(η)T y − (γ ⊗ πT 2)(η)y,
as elements of X S ⊗B Y .
Proof. Let η := x ⊗ ω ∈ XS∇T
1.10
⊗h
B1
Ω1
u(B1, B2) be an elementary tensor. By Proposition
m ◦ (πT ⊗ πT )(δω) = δT ◦ πT (ω) − πT 2(ω) = [T, ωT ] − ωT 2,
and we compute for y ∈ Dom T 2:
(1 ⊗ m)(1 ⊗ 1 ⊗ πT )((1 ⊗∇T δ)(x ⊗ ω))y = (1 ⊗ m)(∇T (x)ωT + γ(x) ⊗ (πT ⊗ πT )(δω))y
= ∇T (x)ωT y + γ(x) ⊗ [T, ωT ]y − γ(x) ⊗ ωT 2y
= ∇T (x)ωT y + γ(x) ⊗ T ωT y + γ(x) ⊗ ωT T y − γ(x) ⊗ ωT 2y
= (1 ⊗∇ T )(x ⊗ ωT y) + γ(x) ⊗ ωT T y − γ(x) ⊗ ωT 2y
= (1 ⊗∇ T )(1 ⊗ πT )(x ⊗ ω)y + (γ ⊗ πT )(x ⊗ ω)T y − (γ ⊗ πT 2)(x ⊗ ω)y. (2.7)
By Lemma 1.6, πT defines a continuous map
u(B1, B2) → B(Dom T, XS∇T
⊗h
1 ⊗ πT : XS∇T
B1 Ω1
and a continuous map
⊗h
B1 Dom T ) ⊂ B(Dom T, Dom 1 ⊗∇ T ),
u(B1, B2) → B(Dom T, X ⊗B Y ).
Invoking Lemma 2.12 as well, we see that πT 2 defines a continuous map
γ ⊗ πT : XS∇T
⊗h
B1 Ω1
γ ⊗ πT 2 : XS∇T
⊗h
B1 Ω1
u(B1, B2) → B(Dom T, XS∇T
⊗h
B1 Y ) → B(Dom T, X S ⊗B Y ).
Therefore by Equation (2.7) the operator (1 ⊗ 1 ⊗ πT )(1 ⊗∇T δ) extends by continuity
to all η ∈ XS∇T
u(B1, B2) and, as Dom T 2 is a core for T , to all y ∈ Dom T .
Ω1
⊗h
B1
22
⊗h
For a C 2-connection (∇,∇S) on a C (1,2)-module (X, S) and x ∈ X we have ∇S
XS∇T
Lemma 2.20 the operator πT (∇ ◦ ∇S) : X ⊗alg
We then come to our main result.
u(B1, B2), which maps completely contractive to XS∇T
Ω1
Y → X S ⊗B Y is defined.
T (x) ∈
u(B1). Hence by
⊗h
Ω1
B1
B1
B1
Theorem 2.21. Let (B, Y, T ) be a C 2-Kasparov module, (X, S) a C (1,2)-module and
(∇,∇S) a C 2-connection. Let πT (∇ ◦ ∇S) be the represented curvature from Definition
2.16 and R∇T be the curvature operator from Definition 2.19. Then there is an equality
R∇T = πT (∇ ◦ ∇S) : X ⊗alg
B2
Dom T 2 → X S ⊗B Y.
T (B1) ⊂ B(Y, X S⊗B Y ) for all x ∈ X and R∇T extends
T , then the difference
T
B2
B Ω2
)(x) ∈ X S ⊗h
Y . If (∇(cid:48),∇(cid:48)S) is another C 2-connection with ∇T = ∇(cid:48)
Consequently, R∇T (x) ∈ X S⊗h
to X ⊗alg
(R∇T − R∇(cid:48)
Proof. For x ∈ X, ∇S(x) ∈ XS∇T
y ∈ Dom T 2:
πT (∇ ◦ ∇S)(x ⊗ y) = (1 ⊗ m)(1 ⊗ 1 ⊗ πT )((1 ⊗∇T δ) ◦ ∇S)(x)y
T 2(x)y
J2
T (B1).
⊗h
Ω1
B1
B
u(B1, B2), so we compute using Lemma 2.20, with
T (x)T y − ∇S
T (x)T y + x ⊗ T 2y − ∇S
T (x)y + ∇S
T (x)y + ∇S
T (x)y + x ⊗ T y) − (1 ⊗∇S T 2)(x ⊗ y)
= (1 ⊗∇ T )∇S
= (1 ⊗∇ T )∇S
= (1 ⊗∇ T )(∇S
= ((1 ⊗∇ T ) ◦ (1 ⊗∇S T ) − 1 ⊗∇S T 2)(x ⊗ y)
= R∇T (x)y
(2.8)
T 2(x)y − x ⊗ T 2y
as desired. Now if (∇(cid:48),∇(cid:48)S) is such that ∇T = ∇(cid:48)
(R∇T − R∇(cid:48)
T
)(x ⊗ y) = (∇S
T 2 − ∇(cid:48)S
By Corollary 1.11 (1 ⊗ πT 2)(∇S − ∇(cid:48)S)(x) ∈ XS∇T
T then Equation (2.8) gives
T 2)(x)y = (1 ⊗ πT 2)(∇S − ∇(cid:48)S)(x)y.
T (B1).
J2
T (B1) = X S ⊗h
⊗h
J2
B
B1
2.3 The curvature of a C 2-correspondence
So far we have focused on giving meaning to the curvature of a C 2-connection on X
under minimal differentiability assumptions. The represented curvature exists on C 1-
modules, and it coincides with the curvature operator on C (1,2)-modules. It is now time
to give meaning to Equation (0.2) and define the curvature associated to a correspon-
dence. Briefly, a correspondence is a Kasparov module together with a connection on
the module, and we define this in detail below. Before doing so, we describe how we
impose additional C 2-conditions on our modules and connections.
Definition 2.22. Let (B, Y, T ) be a C 2-Kasparov module. A C 2-module for (B, Y, T )
is a pair (X, S) such that
1. X is a horizontally differentiable C 2-module with C∗-closure X;
23
2. X ⊂ Dom S2.
A C 2-connection on a C 2-module (X, S) is a pair (∇,∇1) of connections
∇ : X → X ⊗h
u(B, B2), ∇1 : X → X∇T ⊗h
such that (S ⊗ 1, 1 ⊗∇ T ) is a vertically anticommuting pair.
B Ω1
B1 Ω1
u(B1, B2),
Let us clarify the modifications of the above definition relative to Definitions 2.14 and
is that the module X is now assumed to be
2.22. The content of conditions 1. and 2.
both horizontally and vertically C 2 as opposed to just vertically C 1. The curvature of
the C 2-connection is now viewed, using our Sobolev space analogy from Remark 2.9, as
a map W 2
(−1,0). The
next Lemma makes this statement precise.
Lemma 2.23. Let (X, S) be a C 2-module and (∇,∇1) a C 2-connection. The connection
(2,2) → L2, whereas with only a C 1-structure it is a map W 2
(1,2) → W 2
∇S := (ιS ⊗ 1) ◦ ∇1 : X → XS∇T
⊗h
B1 Ω1
u(B1, B2),
makes (∇,∇S) into a C 2-connection on (X, S) viewed as a C (1,2)-module.
Proof. The identity map on X extends to a complete contraction ιS : X∇T → XS∇T
.
The only thing to check from Definition 2.14 is the compatibility (S + i)−1∇T (x) =
(S + i)−1 · ∇S
T (x), which holds automatically.
The next proposition shows that the curvature operator of a C 2-connection is well-
defined.
Proposition 2.24. Let (X, S) be a C 2-module and (∇,∇1) a C 2-connection. Then
B2
X ⊗alg
Dom T 2 ⊂ Dom(S ⊗ 1 + 1 ⊗∇ T )2 ⊂ F(S ⊗ 1, 1 ⊗∇ T ),
and R∇T := (1⊗∇ T )2 − 1⊗∇ T 2 is a symmetric operator defined on X⊗alg
extends to X ⊗alg
Proof. As (S ⊗ 1, 1⊗∇ T ) is a vertically anticommuting pair [37, Theorem 5.1] gives that
Dom(S2 ⊗ 1)∩ Dom(1⊗∇ T )2 = Dom(S ⊗ 1 + 1⊗∇ T )2 ⊂ F(S ⊗ 1, 1⊗∇ T ). By Lemma
2.18
Y . Here the set F is as in Definition 2.10.
Dom T 2 that
B2
B2
X ⊗alg
B2
Dom T 2 ⊂ Dom(1 ⊗∇ T )2,
so condition 2 of Definition 2.22 implies that X⊗alg
Definition 2.25. A C 2-correspondence between a C 2-Kasparov A-C module (A, E, D)
and a C 2-Kasparov B-C module (B, Y, T ) is a quintuple (A, X, S, (∇,∇1)) such that:
Dom T 2 ⊂ Dom(S⊗1+1⊗∇ T )2.
B2
1. (A, X, S) is an unbounded (A, B) Kasparov module;
2. (X, S) is a C 2-module;
24
3. (∇,∇1) is a C 2-connection;
4. there is a unitary isomorphism u : E → X⊗B Y intertwining the A-representations
and such that
Dom D ∩ u∗(Dom S ⊗ 1 ∩ Dom 1 ⊗∇ T ),
is dense in E and
Z := D − u∗(S ⊗ 1 + 1 ⊗∇ T )u : Dom D ∩ u∗(Dom S ⊗ 1 ∩ Dom 1 ⊗∇ T ) → E,
is bounded and preserves Dom D.
As Z preserves Dom D, (D, Z) is a vertically anticommuting pair. By [37, Theorem 5.1]
Dom D2 = Dom D2 ∩ Dom Z 2 = Dom(D − Z)2 = Dom u∗(S ⊗ 1 + 1 ⊗∇ T )2u.
Then by Definiton 2.22, for a C 2-correspondence we have X ⊗alg
As mentioned, correspondences are more-or-less unbounded Kasparov modules with ad-
ditional connection data. The additional connection data allows one to construct repre-
sentatives of the Kasparov product [7, 28, 37, 38, 39], as described in the next result.
Proposition 2.26. Let (A, X, S, (∇,∇1)) be a C 2-correspondence for (A, E, D) and
(B, Y, T ). Then
Dom T 2 ⊂ Dom D2.
B2
[(A, E, D)] = [(A, X, S)] ⊗B [(B, Y, T )] ∈ KK0(A, C),
where ⊗B denotes the Kasparov product.
Proof. Since Z is bounded it follows from [18, Theorem 4.2] that D and D− Z represent
the same KK-class. Since (S⊗1, 1⊗∇ T ) form a weakly anti-commuting pair and 1⊗∇ T
commutes boundedly with A, it follows from [37, Theorem 7.2] that the product relation
holds.
The reason for introducing correspondences is the more refined curvature data that
becomes available.
Definition 2.27. Let (A, E, D) and (B, Y, T ) be two C 2-Kasparov modules and let
(A, X, S, (∇,∇1)) a C 2-correspondence for them. We define the curvature operator of
the correspondence (A, X, S, (∇,∇1)) to be the symmetric operator
Dom T 2 → X ⊗B Y.
R(S,∇T ) := (S ⊗ 1 + 1 ⊗∇ T )2 − S2 ⊗ 1 − 1 ⊗∇ T 2 : X ⊗alg
(2.9)
By Proposition 2.24 it holds that R(S,∇T ) = R∇T + [S ⊗ 1, 1⊗∇ T ]+, by a straightforward
algebraic calculation on the domain X ⊗alg
Note that it is not necessary to specify D in order to define the curvature operator, since
we could just as well take D to be the tensor sum S ⊗ 1 + 1 ⊗∇ T itself. However, in
examples it turns out that the bounded operator Z appearing as their difference contains
geometric information as well (see for instance Equation (5.1)).
B2
Dom T 2.
B2
25
2.4 Universal connections on locally convex modules
In the applications of our theory (cf. Section 5 below) one typically finds that the
modules actually come equipped with more differentiable structure, beyond the C 2-
stucture described above. For instance, in the category of smooth manifolds the Hilbert
modules are typically based on Fr´echet modules, and Fr´echet continuous maps. Let us
describe here how to incorporate such locally convex spaces and algebras in the above
C 2-context.
Let (B, Y, T ) be a C 2-Kasparov module and assume that B carries the structure of a
complete locally convex m-∗-algebra for which there is a continuous inclusion B → B2,
see [21]. Denote by B the C∗-closure of B in the norm coming from B(Y ). The Haagerup
tensor norm is a cross-norm, that is, it satisfies (cid:107)x⊗y(cid:107)h = (cid:107)x(cid:107)(cid:107)y(cid:107) (see [4, Section 1.5.4]).
that the identity map on B ⊗alg B extends to continuous maps
Denoting the projective tensor product by (cid:98)⊗, [42, Propositions 43.4 and 43.12.a)] prove
B(cid:98)⊗B → B(cid:98)⊗Bk → B ⊗h Bk, B(cid:98)⊗B → B1(cid:98)⊗Bk → B1 ⊗h Bk,
k = 1, 2.
By continuity of the multiplication maps, these inclusions restrict to continuous maps
u(B) → Ω1
Ω1
u(B, Bk), Ω1
u(B) → Ω1
u(B1, B2).
(2.10)
Let X be a locally convex topological vector space which is a right B-module such that
the module multiplication defines a continuous map X(cid:98)⊗B → X. Moreover assume that
there is a continuous inner product
X × X → B,
giving X the structure of a pre-Hilbert C∗-module over the pre-C∗-algebra B and denote
by X the C∗-module closure of X. The identity on X induces a continuous map X → X,
and thus by [42, Proposition 43.4] we obtain continuous maps
ι0 : X(cid:98)⊗BΩ1
ι1 : X(cid:98)⊗BΩ1
u(B) → X(cid:98)⊗BΩ1
u(B) → X(cid:98)⊗B1Ω1
u(B, B1) → X ⊗h
B Ω1
u(B1, B2) → X∇T ⊗h
u(B, B1)
B1 Ω1
u(B1, B2).
(2.11)
(2.12)
The maps ι0, ι1 are well-defined because the Haagerup norm is a cross-norm and the
projective norm is the largest cross-norm [42, Proposition 43.12.a)]. Therefore
ιT := (1 ⊗ πT ) ◦ ι0 : X(cid:98)⊗BΩ1
u(B) → X ⊗h
B Ω1
T (B, B1),
X(cid:98)⊗BΩ1
is continuous as well.
Proposition 2.28. Let (B, Y, T ) be a C 2-Kasparov module. Assume that ∇u : X →
u(B) is a Hermitian connection, S : X → X an essentially self-adjoint and regular
operator and (S ⊗ 1, 1 ⊗∇ T ) is a vertically anticommuting pair. If the map
(S + i)−1 ◦ ιT ◦ ∇u : X → X ⊗h
B Ω1
T (B, B1),
26
is continuous, then the identity map X → XS∇T
map on the algebraic tensor product extends to a continous map
is continuous. Consequently the identity
ιS : X(cid:98)⊗BΩ1
u(B) → XS∇T
⊗h
B1 Ω1
u(B1, B2).
B B2, S).
With ∇ := ι0 ◦ ∇u and ∇S := ιS ◦ ∇u, the pair (∇,∇S) is a C 2-connection on the
C (1,2)-module (X ⊗alg
Remark 2.29. When S = 0 the map ιS reduces to the map ι1 defined above.
Proof. Since we have assumed that (S ⊗ 1, 1 ⊗∇ T ) is a vertically anti-commuting pair,
it suffices to define ∇S and verify that ∇S satisfies condition 2. of Definition 2.14.
Continuity of the map (S+i)−1◦ιT◦∇u means that for the Haagerup norm (cid:107)·(cid:107)X⊗h
there is a continuous seminorm p on X such that
(cid:107)(S + i)−1 ◦ ιT ◦ ∇u(x)(cid:107)X⊗h
T (B,B1) ≤ p(x).
T (B,B1)
(2.13)
Thus we obtain a continuous inclusion ιS : X → XS∇T
. The remaining statements
now follow by functoriality of the projective tensor product for continuous maps [42,
Proposition 43.4]. For the pair (∇,∇S) := (ι0 ◦ ∇u, ιS ◦ ∇u) and x ∈ X it holds that
BΩ1
BΩ1
(S + i)−1∇(x) = (S + i)−1(ι0 ◦ ∇u)(x)
= (S + i)−1 · (ιS ◦ ∇u)(x) = (S + i)−1∇S(x),
and thus condition 2. of Definition 2.14 is satisfied.
Corollary 2.30. Assume that ∇u : X → X(cid:98)⊗BΩ1
u(B) is a universal Hermitian connection
such that the connection ιT ◦ ∇u : X → X ⊗h
T (B, B1) is continuous. Applying
B Ω1
Proposition 2.28 with S = 0 yields the C 2-connection
∇1 := ι1 ◦ ∇u : X → X∇T ⊗h
u(B1, B2),
B1 Ω1
on the C 2-module (X, S). Assume further that
1. X ⊂ Dom S2;
2. (A, X, S) is an unbounded Kasparov module;
3. (S ⊗ 1, 1 ⊗∇ T ) is a vertically anti-commuting pair;
4. for all a ∈ A, a : Dom 1 ⊗∇ T → Dom 1 ⊗∇ T and [1 ⊗∇ T, a] is bounded.
Then (A, X ⊗alg
1 ⊗∇ T ) and (B, Y, T ). On X ⊗alg
B B2, S, (ι0 ◦ ∇u, ι1 ◦ ∇u)) is a C 2-correspondence for (A, X ⊗B Y, S ⊗ 1 +
B Dom T 2 there are equalities
R∇T = (1 ⊗∇ T )2 − 1 ⊗∇ T 2 = (1 ⊗ m)(1 ⊗ πT ⊗ πT )(∇2
u),
R(S,∇T ) = R∇T + [S ⊗ 1, 1 ⊗∇ T ]+,
of symmetric operators in X ⊗B Y .
(2.14)
(2.15)
Proof. The equality (2.14) is proved analogously to Lemma 2.20 and Theorem 2.21.
27
3 Finitely generated projective modules over spectral triples
In this section we assume that B is a unital C∗-algebra and X is a finitely generated full
Hilbert C∗-module over B. Then X is algebraically finitely generated and projective.
We assume that X is Z/2-graded with grading γ. We describe the construction of the
curvature operator explicitly in this case. Our results are in complete agreement with
the purely algebraic approach described, for instance, in [33].
3.1 Connections on finite projective C 2-modules
We fix a C 2-spectral triple (B, H, D) in the sense of Definition 1.2. Consider a Z2-graded
inner product module X which is finitely generated and projective over B2, together with
an inner-product preserving injection v : X → B2N
of right B2-modules. The map v
2 , where p ∈ M2N (B2) is a projection and
induces a module isomorphism v : X → pB2N
extends to an isometry v : X → pB2N on the C∗-module level.
The main simplification of the construction in Section 2 is that we take the vertical
operator S = 0. Most importantly, for a Hermitian connection ∇ : X → X ⊗h
u(B, B1)
1 via the inclusion B2 → B1,
we have XS∇D
we define the norm (cid:107)x(cid:107)v := (cid:107)v(x)(cid:107)B2N
on X and denote the completion of X in this norm
by Xv.
Let ei denote the standard basis of B2N
2
{xi}1≤i≤N ⊂ X is a frame for X, that is IdX =(cid:80)
= X∇D. Considering v as a map v : X → B2N
:= v∗(ei). Then the finite set
and set xi
B Ω1
2
1
The Grassmann connection
∇v : X → X ⊗alg
B2
u(B1, B2), ∇v(x) :=
Ω1
γ(xi) ⊗ δ((cid:104)xi, x(cid:105)),
1≤i≤N xi(cid:105)(cid:104)xi.
(cid:88)
1≤i≤N
}.
is a well-defined Hermitian connection. Given any other connection ∇ on X we define
the connection one-form ω(x) := ∇(x) − ∇v(x).
Recall from Equation (2.3) that for a Hermitian connection ∇ and S = 0 we have
(cid:107)x(cid:107)∇D = max{(cid:107)x(cid:107)X,(cid:107)∇D(x)(cid:107)X⊗h
BΩ1
D
Lemma 3.1. Let ∇ : X → X ⊗B Ω1
operator space norms (cid:107) · (cid:107)v and (cid:107) · (cid:107)∇D are cb-equivalent.
Proof. It follows from [39, Lemma 3.6] that the norms (cid:107)x(cid:107)v = (cid:107)v(x)(cid:107)B2N
and (cid:107)x(cid:107)∇v are
cb-equivalent. Since ω(x) := ∇(x) − ∇v(x) is B2-linear it is bounded for the C∗-norm.
It follows that the norm (cid:107)x(cid:107)∇D = max{(cid:107)x(cid:107)X,(cid:107)∇D(x)(cid:107)X⊗h
} is equivalent to the norm
(cid:107)x(cid:107)v.
u(B, B1) be a Hermitian connection on X. Then the
BΩ1
D
1
This means that for any two connections X∇ = X∇(cid:48) = Xv. The operator module Xv is
finitely generated and projective over B1, as is X over B. For any w ∈ Xv the map
(cid:104)w : Xv → B1,
x (cid:55)→ (cid:104)w, x(cid:105),
28
Z we then have ω =(cid:80)
is completely bounded by [39, Proposition 3.7.2]. Hence for any left operator B1-module
Z and ω ∈ Xv ⊗h
Z.
Similarly X ⊗h
B Z. We summarise the simplifications in the following
definition.
1≤i≤N xi ⊗ (cid:104)xi, ω(cid:105) so Xv ⊗h
B Z = X ⊗alg
Z = Xv ⊗alg
B1
B1
B1
Definition 3.2. Let (B, H, D) a C 2-spectral triple. For k = 1, 2 a C k-submodule of X
is a dense Bk submodule X ⊂ X such that for all x1, x2 ∈ X we have (cid:104)x1, x2(cid:105) ∈ Bk. A
C 2-connection is a linear map ∇ : X → Xv ⊗alg
u(B1, B2) satisfying the Leibniz rule.
Given a C 2-connection ∇ on the finite projective module X, we obtain the represented
connection
Ω1
B1
∇D : X → X ⊗alg
B Ω1
D(B, B1), ∇D(x) = (1 ⊗ πD)(∇(x)).
Similarly we obtain
defined by ∇D2(x) := (1 ⊗ πD2)(∇(x)).
∇D2 : X → Xv ⊗alg
B1
Ω1
D2(B2, B1),
3.2 The curvature operator for finitely generated projective modules
We are now in a position to apply our general formalism to compute the curvature
operator of a finite projective module. We first recall the following well-known result.
Proposition 3.3. Suppose that X is algebraically finitely generated and projective over
B2. The operator
1 ⊗∇ D : X ⊗alg
B2
Dom D → X ⊗B H,
x ⊗ h (cid:55)→ γ(x) ⊗ Dh + ∇D(x)h,
is well-defined and essentially self-adjoint. Moreover X ⊗alg
and the operator
B2
Dom D2 ⊂ Dom(1 ⊗∇ D)2
1 ⊗∇ D2 : X ⊗alg
B2
Dom D2 → X ⊗B H,
x ⊗ h (cid:55)→ x ⊗ D2h + ∇D2(x)h,
is well-defined and symmetric. The curvature operator
R∇D := 1 ⊗∇ D2 − (1 ⊗∇ D)2 : X ⊗alg
B2
Dom D2 → X ⊗B H,
is a densely defined symmetric operator.
Proof. The first statement is proved in [8, 10, 36] and several subsequent works [7, 28,
38, 39]. The second statement follows from Lemma 2.18 and the third statement from
Theorem 2.21 both with S = 0.
Proposition 3.4. Suppose that X is finitely generated and projective over B2 and that
∇ : X → Xv ⊗alg
u(B1, B2) is a C 2-connection. Then R∇ extends to a bounded operator
on X ⊗B H. Moreover if ∇v is the Grassmann connection of the frame {xi} and
Ω1
B1
ω : X → X ⊗alg
B1
u(B1, B2), ω(x) := ∇(x) − ∇v(x),
Ω1
the connection form of ∇, then R∇D = v∗[D, p][D, p]v + ω2
D + v∗πD(δ(vωv∗))v.
29
Proof. Theorem 2.21 with S = 0 gives us that R∇D(x) ∈ X ⊗h
R∇D we then find
B Ω2
D. By B2-linearity of
(cid:88)
R∇D(x) =
R∇D(xi)(cid:104)xi, x(cid:105),
and since there are finitely many elements R∇D(xi) ∈ X ⊗B Ω2
a bounded operator. Now we have
D, it follows that R∇D is
i
1 ⊗∇ D = v∗Dv + ωD,
1 ⊗∇ D2 = v∗D2v + ωD2,
and
(1 ⊗∇ D)2 = (v∗Dv + ωD)2 = v∗Dvv∗Dv + v∗[D, vωDv∗]v + ω2
D.
Using p[D, p]p = 0 we compute v∗Dvv∗Dv− v∗D2v = v∗[D, p][D, p]v where p = vv∗, and
v∗[D, vωDv∗]v − ωD2 = v∗([D, πD(vωv∗)] − πD2(vωv∗))v = v∗πD(δ(vωv∗))v,
by Proposition 1.10. Thus it now follows that
R∇D = v∗[D, p][D, p]v + v∗πD(δ(vωv∗))v + ω2
D,
as claimed.
Although the curvature operator of a finitely generated projective module is bounded,
there is no uniform bound on its norm, in the following sense. As an illustrative example,
consider the module L1 of sections of the tautological line bundle L1 → P1(C) over the
two sphere and define Ln := L⊗n
1 . Then the calculations in [7, Section 6] show that
(cid:107)R∇n(cid:107) ∼ n for a natural family of Grassmann connections ∇n. The infinite direct
Ln can be given the structure of a C 1-module, whose curvature operator is
sum(cid:76)∞
n=1
unbounded.
4 Grassmann connections
The Kasparov stabilisation theorem shows that every countably generated Hilbert mod-
ule is a complemented submodule of the standard module and thus admits a frame.
Modulo differentiability, every isometric inclusion in the standard module yields a con-
nection, called a Grassmann connection. In purely noncommutative settings, one often
has access to a frame but not necessarily much else. In this section we provide sufficient
differentiability conditions on frames and modules for C 2-Grassmann and their curva-
ture to exist. A range of examples comes from Cuntz-Pimsner algebras [7, 23, 24, 41]
including the θ and q-deformed 3-spheres, as well as abstract constructions of connec-
tions in KK-theory [26, 39]. Interestingly, we will see in Section 5 that our sufficient
conditions are met for Riemannian submersions as well.
To handle Z2-graded modules we need a Z2-graded standard module, which we take to be
(the completion of) HB = H ⊗ B = (cid:96)2(Z)⊗ B. Here Z = Z\{0}, and the ±-homogenous
30
subspaces are those indexed by positive n ∈ Z and those indexed by negative n ∈ Z. A
basis {ei}i∈Z for HB is homogenous if ei has positive degree if and only if i is positive.
Recall that a countable frame for a Hilbert module X is a sequence {xi} ⊂ X such
i xi(cid:104)xi, x(cid:105) as a norm convergent series. That is, the sum
that for all x ∈ X, x = (cid:80)
IdX =(cid:80)
i xi(cid:105)(cid:104)xi converges strictly.
4.1 Differentiable stabilisation
Recall from Definition 2.1 that for k = 1, 2, a horizontally C k-submodule is a Bk-
submodule X ⊂ X for which (cid:104)x, y(cid:105) ∈ Bk.
Definition 4.1. Let k = 1, 2 and X a C∗-module over B and X ⊂ X a horizontally C k-
submodule. An even stabilisation isometry v : X → HB horizontally C k-differentiable
with respect to (B, Y, T ) if
1. v : X → HB restricts to a map v : X → H ⊗h Bk;
2. there is a dense graded subspace H ⊂ H such that v∗ : HB → X restricts to an
even map v∗ : H ⊗alg Bk → X.
Given a homogenous orthonormal basis {ei} ⊂ H ⊂ H, we say that the frame {xi =
v∗(ei ⊗ 1)} is a C k-frame.
That {xi = v∗(ei ⊗ 1)} is a frame for the Hilbert C∗-module X is a short computation.
Proposition 4.2. Let (B, Y, T ) be a C k unbounded Kasparov module, X a C∗-B-module,
X a horizontal C k-submodule and v : X → H⊗hBk a C k-stabilisation. For any countable
homogenous basis {ei} ⊂ H for the Hilbert space H, the elements xi := v∗(ei ⊗ 1) ∈ X
form a frame for X with the property that for all x ∈ X the series
[T,(cid:104)xi, x(cid:105)]∗[T,(cid:104)xi, x(cid:105)],
(T − i)−1[T 2,(cid:104)xi, x(cid:105)]∗[T 2,(cid:104)xi, x(cid:105)](T + i)−1,
i∈Z
i∈Z
are norm convergent in End∗
Proof. Let {ei} ⊂ H be any countable homogenous basis for the Hilbert space H. Then
xi := v∗(ei ⊗ 1) ∈ X by condition 2 of Definition 4.1. Since
C(Y ) (if k = 1, only the first series converges).
(cid:88)
((cid:104)xi, x(cid:105))i∈Z = ((cid:104)v∗(ei ⊗ 1), x(cid:105))i∈Z = ((cid:104)ei ⊗ 1, v(x)(cid:105))i∈Z,
the column ((cid:104)xi, x(cid:105))i∈Z satisfies
(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:88)
i≥n
[T,(cid:104)xi, x(cid:105)]∗[T,(cid:104)xi, x(cid:105)]
[T,(cid:104)ei ⊗ 1, v(x)(cid:105)]∗[T,(cid:104)ei ⊗ 1, v(x)(cid:105)]
T ((cid:104)ei ⊗ 1, v(x)(cid:105))∗π1
π1
T ((cid:104)ei ⊗ 1, v(x)(cid:105))
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) → 0,
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =
=
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:88)
(cid:88)
i≥n
i≥n
31
(cid:88)
as claimed. To prove the norm convergence of
(T − i)−1[T 2,(cid:104)xi, x(cid:105)]∗[T 2,(cid:104)xi, x(cid:105)](T + i)−1,
i∈Z
we use the same argument, estimating with the representation π2
T of Equation (1.1).
4.2 C 2-Grassmann connections
Lemma 4.3. Let X be a horizontally C 1-module for (B, Y, T ) and (v, HB) a C 1-stabilisation.
The Grassmann connection
∇v : X → X ⊗h
u(B1, B), ∇v(x) := (γ(v)∗ ⊗ 1)(1 ⊗ δ)v(x),
B Ω1
is defined on X. For any homogenous orthonormal basis {ei} ⊂ H ⊂ H, we can use the
C 1-frame {xi = v∗(ei ⊗ 1)}i∈Z to express the Grassmann connection as
(cid:88)
∇v(x) =
γ(xi) ⊗ δ((cid:104)xi, x(cid:105)),
(4.1)
as a norm convergent series. Consequently
i∈Z
∇v
T := πT (∇v) : X → X ⊗h
B Ω1
T (B1), ∇v
T (x) =
(cid:88)
γ(xi) ⊗ [T,(cid:104)xi, x(cid:105)],
i∈Z
is well-defined and independent of the choice of orthonormal basis in H ⊂ H.
Proof. The derivation δ : B1 → Ω1(B1, B) is completely contractive, hence
1 ⊗ δ : H ⊗h B1 → H ⊗h Ω1(B1, B)
∼−→ H ⊗h B1 ⊗h
B1 Ω1(B1, B),
is defined. Thus the composition ∇v := (v∗ ⊗ 1)(1 ⊗ δ)v is defined on X. Choose an
orthonormal basis {ei} ⊂ H ⊂ H and form the C 1-frame {xi = v∗(ei ⊗ 1)}i∈Z. Then
∇v(x) = (γ(v)∗ ⊗ 1)(1 ⊗ δ)v(x) = γ(v)∗(1 ⊗ δ)
ei ⊗ (cid:104)ei ⊗ 1, v(x)(cid:105)
= γ(v)∗(1 ⊗ δ)
ei ⊗ (cid:104)xi, x(cid:105)
ei ⊗ δ((cid:104)xi, x(cid:105))
(cid:88)
(cid:88)
= γ(v)∗
i∈Z
i∈Z
(cid:88)
i∈Z
(cid:88)
=
γ(xi) ⊗ δ((cid:104)xi, x(cid:105)).
i∈Z
Hence the represented connection ∇v
and independent of the choice of orthonormal basis.
T = (1⊗πT )◦∇v : X → X⊗h
B Ω1
T (B1) is well-defined
32
By Lemma 2.4 we obtain a densely-defined symmetric operator
1 ⊗∇v T : X ⊗alg
B Dom T → X ⊗B H,
x ⊗ h (cid:55)→ γ(x) ⊗ T h + ∇v
T (x)h,
in the Hilbert C∗-module X ⊗B Y .
Proposition 4.4. Let (B, Y, T ) be a C k-Kasparov module, (X, S) a C 1-module if k = 1
and a C (1,2)-module if k = 2, and (v, H) a C k-stabilisation. Suppose that
Dom T → HY ,
(γ ⊗ T )(v ⊗ 1) − (v ⊗ 1)(1 ⊗∇v T ) : X ⊗alg
(4.2)
extends to Dom S ⊗ 1. If (S ⊗ 1, 1 ⊗∇v T ) is a vertically anticommuting pair, then
B1
∇v,S : X → XS∇v
T
⊗h
B1 Ω1
u(B1, B2), ∇v,S(x) := (γ(v)∗ ⊗ 1)(1 ⊗ δ)v(x),
is well-defined and the pair (∇v,∇v,S) defines a C 2-connection on (X, S).
u(B1, Bk) is given by δ(a) = 1 ⊗ a − a ⊗ 1.
Proof. The universal differential δ : Bk → Ω1
Choose an orthonormal basis {ei} ⊂ H ⊂ H and form the C k-frame {xi = v∗(ei⊗1)}i∈Z.
As a C k-frame is in particular a C 1-frame, by Equation (4.1) we have
(cid:88)
i∈Z
∇v(x) =
γ(xi) ⊗ δ((cid:104)xi, x(cid:105)).
It suffices to show that this series is convergent in the Haagerup norm of the tensor
product XS∇v
⊗h
T
B1
Ω1
∇v,S(x) =
u(B1, Bk), for then we can define
γ(xi) ⊗ δ((cid:104)xi, x(cid:105)) ∈ XS∇v
(cid:88)
⊗h
B1 Ω1
T
u(B1, Bk).
In order to prove norm-convergence of ∇v,S(x) using the Haagerup tensor norm, by part
2 of Proposition 1.4, we need to address summability of the column with entries
δ((cid:104)xi, x(cid:105)) = 1 ⊗ (cid:104)xi, x(cid:105) − (cid:104)xi, x(cid:105) ⊗ 1,
. Now C k-column finiteness guarantees
u(B1,Bk). It thus remains to show that the row (γ(xi)) is bounded
as well as boundedness of the row (γ(xi)) in XS∇v
that this column is in HΩ1
in XS∇v
By Equation 4.2, (1 − vv∗)(γ ⊗ T )v(S + i)−1 is a bounded operator. Computing the
norm of the row γ(xi) in XS∇v
.
T
T
The second term has norm 1 and the first term is estimated by
T
(cid:13)(cid:13)(γ(xi)(cid:105))t
(cid:13)(cid:13)XS∇v
(cid:13)(cid:13)(cid:13)(cid:0)(S + i)−1∇v
T (γ(xi))(cid:105)(cid:1)t
T
i
i
i
gives
≤(cid:13)(cid:13)(cid:13)(cid:0)(S + i)−1∇v
(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(γ(xi)(cid:105))t
(cid:13)(cid:13) .
T (γ(xi))(cid:105)(cid:1)t
(cid:13)(cid:13)(cid:13)B(HY ,X⊗BY )
(cid:13)(cid:13)(cid:13) =
(cid:13)(cid:13)(cid:13)(cid:0)(S + i)−1γ(v)∗[γ ⊗ T, v(γ(xi))](cid:1)t
(cid:13)(cid:13)B(X⊗BY,HY )
=(cid:13)(cid:13)(cid:0)[γ ⊗ T,(cid:104)xiv∗]v(S − i)−1(cid:1)
=(cid:13)(cid:13)(cid:0)(γ ⊗ T(cid:104)xi − (cid:104)xi1 ⊗∇v T )(S − i)−1(cid:1)
(cid:13)(cid:13)B(X⊗BY,HY )
≤(cid:13)(cid:13)(cid:0)((1 − vv∗)(γ ⊗ T )v)(S − i)−1(cid:1)(cid:13)(cid:13) ,
i
i
i
i
which remains bounded since Equation (4.2) tells us that ((1 − vv∗) ⊗ 1)(γ ⊗ T )(v ⊗ 1)
extends to Dom(S ⊗ 1).
33
Remark 4.5. In [26, Lemma 4.2, Theorem 4.7] it is proved that, given X and (B, Y, T ),
one can find a dense B1-submodule X ⊂ X and an isometry v : X → (cid:96)2(Z)⊗hB such that
v restricts to a map v : X → (cid:96)2(Z)⊗h B1 and v∗ restricts to a map v∗ : Cc(Z)⊗alg B1 → X.
Furthermore, using [26, Theorem 3.9] it can be shown that (v⊗1)(1⊗∇v T )−(γ⊗T )(v⊗1)
is defined on the range of a certain explicit positive compact operator K. The inverse
of K, made odd in an appropriate way, is a natural candidate for a vertical operator
S. It is unclear however whether such S ⊗ 1 and 1 ⊗∇v T can be made to vertically
anticommute, so that Proposition 4.4 can be applied. This is subject of future research.
5 The curvature operator of a Riemannian submersion
We will now illustrate our notion of curvature for a large class of examples given by
Riemannian submersions of closed spinc manifolds M → B that were analysed using
techniques from unbounded KK-theory in [29]. The main result therein ([29, Theorem
23]) was a factorisation of essentially self-adjoint operators of the form
DM = DV ⊗ 1 + 1 ⊗∇ DB + c(Ω).
(5.1)
where DM is the Dirac operator on the total space, DV a vertical family of Dirac oper-
ators, DB the Dirac operator on the base manifold lifted to an operator 1 ⊗∇ DB on M
using the connection ∇ and, finally, c(Ω) is (Clifford multiplication by) the curvature of
the Riemannian submersion (cf. Definition 5.1 below).
As we will see, in this case the curvature operator of Definition 2.19 -- for which we will
use the short-hand R(DV ,∇) := R(DV ,∇DB ) -- indeed captures curvature of the connection
∇ on the vertical Hilbert module of the submersion, as well as other geometric infor-
mation such as the mean curvature. We will check that the conditions that enter in our
general framework are indeed fulfilled in this concrete geometric context. But first we
give a summary of the geometric setup.
5.1 Geometric setup
Let us start by recalling from [29] the relevant ingredients, refering to that paper for all
details. Thus, we consider a Riemannian submersion of closed Riemannian manifolds
π : M → B. Recall the following tensors that are associated to this structure.
Definition 5.1.
1. The second fundamental form, defined for real vertical vector
fields X, Y and real horizontal vector fields Z on M by
(cid:0)Z((cid:104)X, Y (cid:105)M ) − (cid:104)[Z, X], Y (cid:105)M − (cid:104)[Z, Y ], X(cid:105)M
(cid:1)
(5.2)
Sπ(X, Y, Z) :=
1
2
2. The mean curvature k ∈ π∗Ω1(B) is given as the trace
k = (tr⊗1)(Sπ) .
34
3. The curvature of the fibre bundle π : M → B is given by the element Ω in
Ω2(M ) ⊗C∞(M ) Ω1(M ) given by
Ω(X, Y, Z) := −(cid:104)[(1 − P )X, (1 − P )Y ], P Z(cid:105)M
where P is the orthogonal projection onto vertical vector fields.
If M and B are Riemannian spin manifolds, we may introduce a vertical spinor module
EV , defined in terms of the spinor modules EM and EB on the given spinc manifolds M
and B, respectively [29, Section 3]. We will not dwell on the precise definition here, but
merely recall that EV is a finitely-generated projective C∞(M )-module which satisfies
∼= EM where π∗EB ≡ EB ⊗C∞(B) C∞(M ) denotes
the crucial property that EV ⊗ π∗EB
the pullback. Moreover, Clifford multiplication cV by vertical vector fields is defined on
EV .
The module EV has a (Clifford) connection ∇EV defined in terms of the spinor con-
nections ∇EM and ∇EB . The connection ∇EV is Hermitian for the natural Hermitian
structure (cid:104)·,·(cid:105)EV on EV . The smooth sections of EV have a natural locally convex struc-
ture coming from the usual C∞-topology, which can be defined using ∇EV .
We now define the pre-Hilbert module X over C∞(B) to be EV where the right action
of C∞(B) is defined via the inclusion C∞(B) → C∞(M ) dual to π : M → B. The
C∞(B)-valued inner product is defined by
(cid:104)s, t(cid:105)X(b) :=
(cid:104)s, t(cid:105)EV (x) dµπ−1(b)(x),
s, t ∈ EV .
(5.3)
(cid:90)
π−1(b)
Here dµπ−1(b) is the Riemannian volume form on the submanifold π−1(b). As b (cid:55)→ dµπ−1(b)
is smooth (the volume form on M decomposes locally as a product), this inner product
does in fact take values in C∞(B).
Proposition 5.2. [29, Proposition 16] Let {ej} be a local orthonormal frame of verti-
cal vector fields on M . Then the following local expression defines an odd symmetric
unbounded operator (DV )0 : X → X:
dim(F )(cid:88)
(DV )0(ξ) = i
cV (ej)∇EV
ej
(ξ)
The closure DV : dom(DV ) → X of (DV )0 is regular and self-adjoint.
j=1
In order to form the unbounded Kasparov product of the vertical and the horizontal
components we need to lift the Dirac operator DB on the base manifold to an essentially
self-adjoint unbounded operator on the Hilbert space X ⊗C(B) L2(EB). It turns out that
the Hermitian Clifford connection ∇EV on EV (Hermitian with respect to the C∞(M )-
valued inner product) does not define a metric connection on EV ⊆ X with the C∞(B)-
valued inner product, due to correction terms that come from the measure on the fibres
Mb, b ∈ B. However, these can be nicely absorbed in an additional term proportional
to the mean curvature.
35
Definition 5.3. The metric connection ∇X : X → X ⊗C(B) Ω1
cont(B) is defined by
Z(ξ) = ∇EV
∇X
ZH
(ξ) +
k(ZH) · ξ
1
2
in terms of a vector field Z on B and corresponding horizontal lift ZH.
Proposition 5.4. Let {fi} be a local orthonormal frame of vector fields on B. The local
expression
(1 ⊗∇ DB)0(ξ ⊗ r) := ξ ⊗ DB(r) + i
∇X
fi
(ξ) ⊗ cB(fi)(r)
ξ ∈ EV , r ∈ EB
(cid:88)
defines an essentially self-adjoint unbounded operator
i
(1 ⊗∇ DB)0 : EV ⊗alg
C∞(B)
EB → X ⊗C(B) L2(EB) .
We denote its closure by 1 ⊗∇ DB : Dom(1 ⊗∇ DB) → X ⊗C(B) L2(EB).
Remark 5.5. In [29] it is only shown that (1 ⊗∇ DB)0 is a symmetric operator, whose
closure was then used in the tensor sum factorization (5.1) of DM . However, we may
consider the operator (1 ⊗∇ DB)0 as a differential operator of order 1 on the finitely-
generated projective C∞(M )-module EV ⊗C∞(M ) π∗EB. It is then a classical result [25,
Corollary 10.2.6] that such an operator is essentially self-adjoint.
The main result of [29] is a factorisation of the Dirac operator DM on M in terms of a
vertical family of Dirac operators DV on X and the Dirac operator DB on B. Explicitly,
(up to conjugation by a unitary operator) DM is given by the tensor sum (5.1):
DM = DV ⊗ 1 + 1 ⊗∇ DB + c(Ω).
The last term c(Ω) is Clifford multiplication by the curvature Ω of the fibration π : M →
B. Thus the curvature of the fibration appears as an obstruction to the realisation of
DM as an (unbounded) internal Kasparov product. We also record the following result,
which is Lemma 17 of [30].
Lemma 5.6. Suppose that ξ ∈ X and r ∈ EB ⊆ L2(EB). Let {ej} denote a local
orthonormal frame of vertical vector fields on M and {fi} a local orthonormal frame of
vector fields on B. Then we have the local expression
[DV ⊗ 1, 1 ⊗∇ DB]+(ξ ⊗ r)
Sπ(ek, ej, (fi)H)(cid:0)cV (ej)∇EV
ek
= −(cid:88)
−(cid:88)
i,j,k
i,j
(cid:16)
(cid:1)(ξ) ⊗ cB(fi)(r)
(cid:0)k((fi)H)(cid:1)(cid:17)
cV (ej)
ΩEV (ej, (fi)H) +
1
2
ej
(ξ) ⊗ cB(fi)(r) ,
where ΩEV : EV → EV ⊗C∞(M ) Ω2(M ) is the curvature form of the Hermitian connection
∇EV .
As a consequence we obtain that the anti-commutator [DV ⊗ 1, 1 ⊗∇ DB]+ is relatively
bounded by DV ⊗ 1 and makes DV ⊗ 1 and 1 ⊗∇ DB a vertically anticommuting pair in
the sense of Definition 2.10.
36
5.2 Local expressions
Let us first take a typical fibre F0 of π : M → B and consider trivialisations
ρα : π−1(Wα) → F0 × Wα
for charts Wα ⊆ B. It is then possible to choose so-called fibration charts Vα × Wα ⊂
F0 × B so that with Uα = ρ−1
α (Vα × Wα) we have that
ρα : Uα → Vα × Wα
α χ2
covering {Uα}, so that (cid:80)
α} a partition of unity subordinate to the
is a diffeomorphism. Let us denote by {χ2
α = 1. Note that EV as a (finitely-generated projective)
C∞(M )-module admits a (finite) local orthonomal frame supported on Uα; we denote
such a frame by {xα,n}.
On the base manifold B we choose local coordinates σα : Wα → Rdim B whose components
will be denoted by σµ
It is convenient to write the inner product on X as an integral over the typical fibre F0.
We denote by µb the measure on F0 that corresponds to µπ−1(b) through the identification
π−1(b) ∼= F0. We then obtain
(cid:104)s, t(cid:105)X(b) =
α : Wα → R for µ = 1, . . . , dim B.
(cid:1) (y, b) dµb(y)
(cid:0)(ρ−1
(cid:88)
α )∗(cid:104)s, t(cid:105)EV
(cid:90)
χ2
α
α
F0
for all b ∈ Wα. As a special case, if (y, b0) ∈ ρα0(Uα0) where π−1(b0) = F0 is our typical
fibre then we set dµ0 := dµb0.
The Radon-Nikodym derivative of µb with respect to µ0 gives a function on F0×B which
we will denote by
dµ
dµ0
(y, b) :=
(y)
dµb
dµ0
Remark 5.7. As in [29] we may combine a choice of coordinates on each of the fibration
charts {Uα} with the Riemannian metric g on M to obtain a positive invertible matrix
of smooth functions
gα : Uα → GLdim(M )(R)+.
Furthermore, letting Q : Rdim(M ) → Rdim(M ) denote the projection
Q : (t1, . . . , tdim(M )) (cid:55)→ (t1, . . . , tdim(F ), 0, . . . , 0)
onto the first dim(F ) copies of R in Rdim(M ), we obtain a positive matrix of smooth
functions
QgαQ : Uα → GLdim(F )(R)+,
Suppose that (y, b) ∈ ρα(Uα) and let y1
α
Then we may write the volume form on F0 as
dµb(y) =(cid:112)det QgαQ(ρ−1
α, . . . , ydim F0
α (y, b))dy1
QgαQ(x) = Qgα(x)Q.
denote local coordinates on Vα ⊂ F0.
α ∧ ··· ∧ dydim F0
α
.
37
Let us check, for completeness, that this expression does not depend on the choice of
trivialisation. In fact, if (y, b) ∈ ρα(Uα) ∩ ρβ(Uβ) then the transition functions ρα ◦ ρ−1
β, the map y → y(cid:48) corresponds
map (y, b) to (y(cid:48), b). In terms of the coordinates yk
to an orthogonal transformation T kl(y) := ∂yk
α/∂yl
α and yl
β in Rdim(F ). Hence we have
β
det QgβQ(ρ−1
β (y, b)) = det T 2 · det QgαQ(ρ−1
α (y, b))
while at the same time
β ∧ ··· ∧ dydim F0
dy1
β
= det
∂yk
β
∂yl
α
· dy1
α ∧ ··· ∧ dydim F0
α
= det T −1 · dy1
α ∧ ··· ∧ dydim F0
α
so that the terms involving det T cancel in the definition of dµb.
The Radon-Nikodym derivative can now be written unambiguously on Vα × Wα by
α (y, b))
α0 (y, b0))
This is a smooth and nowhere vanishing function on Vα × Wα.
dµ
dµ0
(y, b) =
.
(cid:112)det QgαQ(ρ−1
(cid:112)det Qgα0Q(ρ−1
5.3 The stabilisation isometry
on X where it will map to C∞(F0)N(cid:98)⊗C∞(B) (in terms of the projective tensor product
We are now ready to define the crucial technical ingredient in our approach to curvature,
to wit, a stabilising isometry v : X → L2(F0)N ⊗h C(B). In fact, we will realise this map
of Fr´echet spaces). We let N denote the product of the cardinalities of the sets {χα}
and {xα,n} for the partition of unity and the frame of EV , respectively, and will consider
elements in L2(F0)N as column vectors.
Lemma 5.8. The map
v : X → C∞(F0)N(cid:98)⊗C∞(B) (cid:39) C∞(F0 × B)N
(cid:33)
(cid:32)(cid:115)
s (cid:55)→
· (ρ−1
α )∗((cid:104)χαxα,n, s(cid:105)EV )
dµ
dµ0
αn
is a continuous map of Fr´echet spaces and furthermore extends to an isometry X →
L2(F0)N ⊗h C(B).
Proof. First observe that as each ρα is a diffeomorphism, the map ρ∗
C∞-topology. Likewise the derivatives(cid:112)dµ/dµ0 are (uniformly bounded) C∞ functions,
α is continuous in the
and so multiplication by them is continuous. Now for every section s we have
(cid:88)
(cid:88)
s =
χ2
αs =
χαxα,n(cid:104)χαxα,n, s(cid:105),
α
α,n
and each term in the sum depends continuously on s. Taking the inner product with
χβxβ,k is C∞ continuous, as is (ρ−1
α )∗. Hence the map v is Fr´echet continuous.
38
to check that for all s, t ∈ X we have (cid:80)
By a standard density argument , to show that v extends to an isomtery, it is sufficient
αn(cid:104)v(s)αn, v(t)αn(cid:105)L2(F0)⊗hC(B) = (cid:104)s, t(cid:105)X. We
(cid:90)
compute(cid:88)
(cid:88)
(cid:104)v(s)αn,v(t)αn(cid:105)L2(F0)⊗C(B)(b) =
v(s)αn(y, b)v(t)αn(y, b) dµ0(y)
α,n
=
α,n
F0
α )∗((cid:104)s, χαxα,n(cid:105)EV (cid:104)χαxα,n, t(cid:105)EV )(y, b)
(ρ−1
dµ
dµ0
(y, b) dµ0(y)
(cid:90)
(cid:90)
(cid:88)
(cid:88)
α,n
F0
α
F0
=
α(ρ−1
χ2
α )∗ ((cid:104)s, t(cid:105)) (y, b) dµb(y) ≡ (cid:104)s, t(cid:105)X
using completeness of the frame {xα,n} and(cid:80)
Lemma 5.9. For all F = (Fαn) ∈ C∞(F0)N(cid:98)⊗C∞(B) we have
(cid:32)(cid:115)
(cid:32)
(cid:33)
(cid:33)
α χ2
α = 1 in the last equality.
v∗(F ) =
(ρ−1
α )∗
dµ0
dµ
Fαn
χαxα,n
αn
(5.4)
(cid:88)
αn
Consequently, the adjoint v∗ : L2(F0)N ⊗h C(B) → X restricts to a map
v∗ : C∞(F0)N(cid:98)⊗C∞(B) → X,
and we have v∗v = 1.
Proof. We check that the formula (5.4) does indeed provide the adjoint of v by computing
(cid:104)s, v∗(F )(cid:105)X(b) =
α )∗ ((cid:104)s, χαxα,n(cid:105)EV ) (y, b)
(ρ−1
Fαn
(y, b) dµb(y)
(cid:90)
(cid:90)
=
= (cid:104)v(s), F(cid:105)L2(F0)N⊗hC(B)
F0
F0
(cid:33)
(cid:33)
(cid:32)(cid:115)
(cid:32)(cid:115)
dµ0
dµ
dµ
dµ0
α )∗ ((cid:104)s, χαxα,n(cid:105)EV ) (y, b)
(ρ−1
Fαn
(y, b) dµ0(y)
The identity v∗v = 1 holds true by construction.
Thus, the operator v defines a C 2-stabilisation v : X ⊗alg
in the sense of Definition 4.1.
Proposition 5.10. The operator
((v ⊗ 1)(1 ⊗∇ DB) − (γ ⊗ DB)(v ⊗ 1)) : X⊗alg
C∞(B)
is DV -bounded.
39
C∞(B) C 2(B) → L2(F0) ⊗h C 2(B)
EB → C∞(F0)(cid:98)⊗EB ⊂ L2(F0)⊗h C(B)
Proof. We start by computing the first term on s ⊗ ψ ∈ X ⊗ EB, say, with supp s ⊆ Uα:
(5.5)
(v ⊗ 1)(1 ⊗∇ DB)(s ⊗ ψ) = (v ⊗ 1)(∇X
(s) ⊗ γµψ + s ⊗ DBψ)
∂/∂σµ
α
=
On the other hand, we have
(DB)(v ⊗ 1)(s ⊗ ψ) =
=
(cid:32)
(cid:32)
(cid:18)(cid:115)
DB
∂
∂σµ
α
(cid:115)
dµ
dµ0
+
(cid:32)(cid:115)
(cid:32)(cid:115)
(cid:115)
+
· (ρ−1
α )∗((cid:104)χαxα,n,∇X
∂/∂σµ
α
s(cid:105)EV )γµψ
· (ρ−1
dµ
dµ0
α )∗((cid:104)χαxα,n, s(cid:105)EV ) · DBψ
(cid:33)(cid:33)
dµ
dµ0
α )∗((cid:104)χαxα,n, s(cid:105)EV ) · ψ
(ρ−1
(cid:33)
α )∗((cid:104)χαxα,n, s(cid:105)EV ) · γµψ
(ρ−1
(cid:33)
α )∗((cid:104)χαxα,n, s(cid:105)EV ) · DBψ
(ρ−1
dµ
dµ0
dµ
dµ0
αn
(cid:19)
(5.6)
αn
(5.7)
αn
The first term in this last expression is bounded as it is a derivative of the Radon-
Nikodym derivative, while the last term cancels against the corresponding term in (5.6).
We are thus left to consider
α )∗((cid:104)χαxα,n,∇X
(ρ−1
∂/∂σµ
α
= −(ρ−1
α )∗((cid:104)∇EV
+ (ρ−1
α )∗(cid:18)(cid:18) ∂
(∂/∂σµ
(cid:0)(ρ−1
α )∗((cid:104)χαxα,n, s(cid:105)EV )(cid:1)
s(cid:105)EV ) − ∂
∂σµ
α
(cid:19)
(cid:19)
(χαxα,n), s(cid:105)EV ) +
(cid:0)(ρ−1
α )∗((cid:104)χαxα,n, s(cid:105)EV )(cid:1) .
(cid:104)χαxα,n, s(cid:105)EV
1
(ρ−1
α )∗ (k((∂/∂σµ
2
− ∂
∂σµ
α
∂σµ
α
α)H
α)H)(cid:104)χαxα,n, s(cid:105)EV )
H
The first two terms on the right-hand side (involving the derivative on the frame and
the mean curvature) is bounded, and we claim that the remaining terms combine to
give only vertical derivatives, and can thus be relatively bounded with respect to the
vertically elliptic DV when acting on s. In order to see that the combination is a vertical
derivative, let us consider the more general expression
α )∗ZH(f ) − Z((ρ−1
(ρ−1
α )∗(f ))
for a vector field Z on B and a function f on M (supported in Uα). Here we understand
Z to act on a function on F0× B by only deriving in the second coordinate. For f = π∗g
one finds that
α )∗ZH(π∗g) − Z((ρ−1
(ρ−1
by definition of the horizontal lift
α )∗(π∗g)) = (ρ−1
α )∗ZH(g ◦ π) − Z(g ◦ π ◦ ρ−1
α ) = 0
α )∗ZH(g ◦ π) = (ρ−1
(ρ−1
α )∗π∗Z(g) = Z(g)
40
in combination with the identity Z(g ◦ π ◦ ρ−1
α )∗) is a vertical vector field, as desired.
Z((ρ−1
α ) = Z(g). We conclude that (ρ−1
α )∗ZH −
Thus, the stabilisation map v satisfies the hypotheses of Proposition 4.4 and the asso-
ciated Grassmann connection constitutes and example of a C 2-connection on a C (1,2)-
module.
5.4 The universal lift
We may use the isometry v to obtain a convenient expression for the connection ∇X. In
particular, we can obtain a lift of ∇X to a universal connection ∇X
u , where by 'lift' we
mean that πDB ◦ ∇X
C∞(F0)N(cid:98)⊗C∞(B). For any s ∈ X there exist functions fk ∈ C∞(F0)N , gk ∈ C∞(B)
Since v∗v = 1 any s ∈ X can be written as s = v∗(F (s)) where F (s) = v(s) ∈
u = c ◦ ∇X where c is Clifford multiplication on spinors.
such that
Then we have
F (s) =
fk ⊗ gk.
(cid:88)
k
dµ
dµ0
(cid:115)
(cid:88)
(cid:0)(ρ−1
s = v∗(F (s)) =
α )∗ (fk ⊗ gk) χαxα,n
(ρ−1
(5.8)
α,n,k
so that with respect to local coordinates σµ
(cid:88)
∇X(s) =
∇X
∂/∂σµ
α
α on B we have
α )∗ (fk ⊗ gk) χαxα,n
(cid:1) ⊗C∞(B) dσµ
α.
α,n,k
Using the Leibniz rule and the fact that the derivative on the base commutes with the
functions fk in the fibre direction, we find that
∇X(s)
=
(5.9)
α )∗ (fk ⊗ 1) ⊗ dgk(s).
α )∗ (fk ⊗ 1) ⊗ gkdσµ
χαxα,n · (ρ−1
(χαxα,n) (ρ−1
(cid:88)
(cid:88)
∇X
α +
∂/∂σµ
α
α,n,k
α,n,k
Lemma 5.11. The connection ∇X can be lifted to a universal connection
u : X → X(cid:98)⊗Ω1
∇X
u(C∞(B)),
in the sense that πDB ◦ ∇X
u = c ◦ ∇X where c denotes Clifford multiplication.
(cid:88)
Proof. From Equation (5.9) we identify a candidate universal connection as
∇X
u (s) =
α )∗ (fk ⊗ 1)⊗δ(σµ
(χαxα,n) (ρ−1
χαxα,n·(ρ−1
α)gk+
∇X
α )∗ (fk ⊗ 1)⊗δ(gk).
∂/∂σµ
α
α,n,k
α,n,k
(5.10)
41
(cid:88)
First of all, the right hand side of (5.10) makes sense in the projective tensor product
topology. The first sum is readily compared to (5.8) since there are only finitely many
terms in the α, n sums. The second term can similarly be compared to (5.8) using the
fact that δ : C∞(B) → Ω1
Multiplication of a section s ∈ X by a function g ∈ C∞(B) via pullback along π amounts
to multiplying each gk by g. Applying the Leibniz rule for δ to the right-hand side of
Equation (5.10) proves that ∇X
u (s)g + s⊗ δ(g). It is then clear by construction
that πDB ◦ ∇X
u(C∞(B)) is completely bounded.
u coincides with c ◦ ∇X.
u (sg) = ∇X
In the next few statements we compare projective tensor products and Haagerup tensor
products and so need the notation introduced in Equations (2.11), (2.12), Section 2.4.
Lemma 5.12. The map
ιDB ◦ ∇X
u : X → X ⊗h
C1(B) Ω1
DB
(C 1(B)),
is continuous.
Proof. For x ∈ X we have an equality
ιDB ◦ ∇X
u (x) = ∇DB (x) = (1 ⊗∇ DBx(cid:105) − γ(x)(cid:105)DB)
= v∗(v(1 ⊗∇ DB) − (γ ⊗ DB)v)(x) + v∗ [γ ⊗ DB, v(x)] ,
of operators Dom DB → X ⊗h
C(B) L2(EB). Since
v∗(v(1 ⊗∇ DB) − (γ ⊗ DB)v) = v∗(v(1 ⊗∇ DB) − (γ ⊗ DB)v)(DV + i)−1(DV + i),
and DV : X → X is continuous, this operator is continuous by Proposition 5.10. Fur-
thermore,
v∗ [γ ⊗ DB, v(x)] = v∗(γ ⊗ c(dv(x))),
and x (cid:55)→ c(dv(x)) is a composition of continuous maps
X v−→ C∞(F0)(cid:98)⊗C∞(B) 1⊗d−−→ C∞(F0)(cid:98)⊗Ω1(B) c−→ L2(F0)N ⊗h Ω1
(C 1(B)).
DB
Since v∗ : L2(F0)N ⊗h Ω1
proved.
T (B1) → X ⊗h
B Ω1
DB
(B1) is continuous as well, the lemma is
Corollary 5.13. The pair (X, DV ) is a C 2-module relative to (C 2(B), L2(EB), DB) and
the universal connection
u : X → X(cid:98)⊗C∞(B)Ω1
∇X
u(C∞(B)),
defines a C 2-connection (∇,∇1) on (X, DV ), where ∇ = ι0 ◦ ∇X
quintuple (C 2(M ), X⊗alg
(C 2(M ), L2(EM ), DM ) and (C 2(B), L2(EB), DB).
u . The
B B2, DV , (∇,∇1)) is a C 2-correspondence for the spectral triples
u and ∇1 = ι1 ◦ ∇X
42
Proof. Since (cid:104)X, X(cid:105) ⊂ C∞(B) and DV : X → X ⊂ X is essentially self-adjoint and
regular, (X, DV ) is a C 2-module (see Definition 2.5). Next we show how to obtain a pair
(∇,∇1) satisfying Definition 2.14. As the inclusions X → X and C∞(B) → Lip(DB) are
u : X → X ⊗B Ω1(C 1(B)). By Lemma
continuous, we obtain a connection ∇ := ι0 ◦ ∇X
5.6 the pair (DV ⊗ 1, 1 ⊗∇ DB) is vertically anti-commuting and by Lemma 5.12
ιDB ◦ ∇X
u : X → X ⊗h
C1(B) Ω1
DB
(C 1(B), C(B)),
is continuous. The conclusion now follows from Equation (5.1) and Corollary 2.30.
5.5 Curvature of ∇
In this section we compute the curvature R∇X
u
of the connection ∇X, which is given by
= (1 ⊗∇X
u
B) − (1 ⊗∇X
D2
u
R∇X
u
DB)2
in terms of the Dirac operator DB on the base manifold of the submersion.
Proposition 5.14.
1. The curvature operator R∇X ≡ πDB ((∇X
given by Clifford multiplication with the curvature of ∇X. More precisely, in terms
of a local orthonormal frame {fj} of vector fields on B we have the equality
= c ◦(cid:0)(∇X)2(cid:1) ≡(cid:88)
(cid:16)(cid:104)∇X
fj
(cid:105) − ∇X
,∇X
fk
R∇X
u
u )2) on X(cid:98)⊗C∞(B)EB is
(cid:17)
[fj ,fk]
γjγk
as skew-symmetric operators from X ⊗alg
γj = c((fj)H) are flat Dirac matrices.
C∞(B)
EB to X ⊗C(B) L2(EB) and where
j,k
2. There is the following local expression for the curvature in terms of the curvature
Ω of the submersion and the connection one-form AX of ∇X:
R∇X
u
(ξ) =
Ω(·,·, ei)ei + dAX + (AX)2
((fj)H, (fk)H)γjγkξ
j,k
i
with ξ supported in a suitable coordinate chart of M and where {ei} is a local
orthonormal frame of vertical vector fields on M .
3. The curvature operator R∇X
u
is relatively DV -bounded.
Proof. In view of Corollary 2.30 we have that R∇X ≡ πDB ((∇X
u )2). By Lemma 5.11
and the fact that the Clifford representation of universal forms factors through the
DeRham calculus, we may compute the right-hand side by working with πD-represented
de Rham differential forms and thus exploit local expressions. Let us start by writing
the connection ∇X in terms of a connection one-form: using Definition 5.3 we have for
ξ ∈ X supported in a trivializing chart (for EV ) on M :
∇X
Z(ξ) = ZH(ξ) + AEV (ZH)(ξ) +
k(ZH) · ξ
1
2
43
(cid:32)(cid:88)
(cid:88)
(cid:33)
where we have written ∇EV = dM + AEV in terms of a (locally-defined) connection
one-form AEV ∈ EndC∞(M )(EV ) ⊗C∞(M ) Ω1(M ).
2k ∈ EndC∞(M )(EV )⊗C∞(M )
Let us define a combined connection one-form as AX := AEV + 1
Ω1(M ) so that ∇X
Z = ZH + AX(ZH). Note that since M is compact we can assume that
there is a finite number of such trivialising charts. Hence, for the relative bounds of the
curvature operator that we are after here we may just as well work on a single chart.
With these preparations, we compute the curvature operator acting on a ξ ∈ X supported
in a single chart, finding
c ◦(cid:0)(∇X)2(cid:1) (ξ) =
(cid:16)(cid:104)∇X
fj
,∇X
fk
(cid:105) − ∇X
(cid:17)
[fj ,fk]
γjγk(ξ)
=
([(fj)H, (fk)H] − [fj, fk]H) γjγk(ξ) + cH ◦ (dAX + (AX)2)(ξ),
where cH denotes Clifford multiplication only in the horizontal direction (involving the
γj and γk). This last term satisfies
(cid:88)
(cid:88)
j,k
j,k
(cid:90)
(cid:104)cH(dAX + (AX))ξ, cH(dAX + (AX))ξ(cid:105)X(b)
(cid:104)cH(dAX + (AX))ξ, cH(dAX + (AX))ξ(cid:105)EV (x)dµπ−1(b)(x)
π−1(b)
=
≤ (cid:107)cH(dAX + (AX))(cid:107)2
EndC∞(M )(EV )(cid:104)ξ, ξ(cid:105)(b)
The relevant and potentially unbounded term in the curvature is thus [(fj)H, (fk)H] −
[fj, fk]H. But this difference of commutators is a vertical vector field and, in fact, it is
precisely the one described by the curvature Ω of π as defined in Definition 5.1. Indeed,
the horizontal lift of a commutator is the horizontal part of the commutator of the lifted
vector fields and hence
[XH, YH] − [X, Y ]H =
Ω(XH, YH, ei)ei
(cid:88)
for vector fields X, Y on B and an orthonormal frame {ei} of vertical vector fields. We
conclude that the curvature is given locally by a vertical vector field plus bounded terms,
and since DV is vertically elliptic we find the desired relative bound.
i
References
[1] J. Arnlind, M. Wilson, Riemannian curvature of the noncommutative 3-sphere,
J. Noncommut. Geom. 11 (2), (2017), 507 -- 536.
[2] D. Blecher, A new approach to Hilbert C∗-modules, Math. Ann. 307, 253 -- 290
(1997).
[3] D. Blecher, J. Kaad and B. Mesland, Operator ∗-correspondences in analysis and
geometry, Proc. Lond. Math. Soc.117 (2) (2018), 303 -- 344.
44
[4] D. Blecher and M. LeMerdy, Operator algebras and their modules, an opera-
tor space approach, London Mathematical Society Monographs 30 (New Series),
Clarendon Press, Oxford (2004).
[5] J. Bhowmick, D. Goswami and S. Joardar, A new look at Levi-Civita connection
in noncommutative geometry, arXiv:1606.08142v3.
[6] J. Bhowmick, D. Goswami and S. Mukhopadhyay, Levi-Civita connections for a
class of spectral triples, arXiv:1809.06721v1.
[7] S. Brain, B. Mesland and W. D. van Suijlekom, Gauge theory for spectral triples
and the unbounded Kasparov product, J. Noncommut. Geom. 10 (1) (2016), 131-
202.
[8] A. Connes, Noncommutative Geometry, Academic Press, 1994.
[9] B. ´Ca´cic and B. Mesland, Gauge theory for noncommutative Riemannian princi-
pal bundles, to appear.
[10] A. Connes, Gravity coupled with matter and foundation of noncommutative geo-
metry, Commun. Math. Phys. 182 (1996), 155 -- 176.
[11] A. Connes and F. Fathizadeh, The term a4 in the heat kernel expansion of non-
commutative tori, arXiv:1611.09815.
[12] A. Connes and H. Moscovici, Modular curvature for noncommutative two-tori, J.
Amer. Math. Soc. 27 (2014), 639 -- 684.
[13] A. Connes and M. Rieffel, Yang-Mills for noncommutative two-tori, Contemp.
Math. 62 (1987), 237 -- 266.
[14] A. Connes and P. Tretkoff, The Gauss-Bonnet Theorem for the noncommutative
two torus, Noncommutative geometry, arithmetic, and related topics, 141 -- 158,
Johns Hopkins Univ. Press, Baltimore, MD, 2011.
[15] J. Cuntz and D. Quillen, Algebra extensions and nonsingularity, J. Amer. Math.
Soc. 8 (1995), 251 -- 289.
[16] L. Dabrowski, P. Hajac, G. Landi and P. Siniscalco, Metrics and pairs of left and
right connections on bimodules, J. Math. Phys. 37 (9) (1996), 4635 -- 4646.
[17] M. Dubois-Violette, J. Madore and T. Masson, On curvature in noncommutative
geometry, J. Math. Phys. 37 (8) (1996) 4089 -- 4102.
[18] K. van den Dungen, Locally bounded perturbations and (odd) unbounded KK-
theory, J. Noncommut. Geom. 12 (2018), 1445 -- 1467.
[19] K. van den Dungen and A. Rennie, Indefinite Kasparov modules and pseudo-
Riemannian manifolds, Annales Henri Poincar´e 17 (2016), 3255 -- 3286.
45
[20] R. Floricel, A. Ghorbanpour and M. Khalkhali, The Ricci curvature in noncom-
mutative geometry, to appear in J. Noncommut. Geom., arXiv:1612.06688v1.
[21] M. Fragoulopoulou, Topological algebras with involution, North-Holland Math.
Studies 200, 2005.
[22] A. Ghorbanpour and M. Khalkhali, Spectral geometry of functional metrics on
noncommutative tori, arXiv:1811.04004v1.
[23] M. Goffeng and B. Mesland, Spectral triples and finite summability on Cuntz-
Krieger algebras, Doc. Math. 20 (2015), 89 -- 170.
[24] M. Goffeng and B. Mesland, A.Rennie, Shift-tail equivalence and an unbounded
representative of the Cuntz-Pimsner extension, Ergodic Th. Dynam. Sys. 38
(2018), 1389 -- 1421.
[25] N. Higson and J. Roe, Analytic K-Homology, OUP, 2001.
[26] J. Kaad, Differentiable absorption of Hilbert C -modules, connections, and lifts
of unbounded operators, J. Noncommut. Geom. 11 (2017), 1037 -- 1068.
[27] J. Kaad and M. Lesch, A local-global principle for regular operators on Hilbert
modules, J. Funct. Anal. 262 (2012), 4540 -- 4569.
[28] J. Kaad and M. Lesch, Spectral flow and the unbounded Kasparov product, Adv.
Math. 248 (2013), 495 -- 530.
[29] J. Kaad and W. D. van Suijlekom. Riemannian submersions and factorization of
Dirac operators, J. Noncommut. Geom. 12 (2018), 1133 -- 1159.
[30] J. Kaad and W. D. van Suijlekom. Factorization of Dirac operators on almost-
regular fibrations of spinc manifolds, to appear in Doc. Math., arXiv:1710.03182.
[31] D. Kucerovsky, The KK-product of unbounded modules, K-Theory 11 (1997),
17 -- 34.
[32] E. C. Lance, Hilbert C∗-modules: a toolkit for operator algebraists, London Math-
ematical Society Lecture Notes 210, Cambridge University Press, Cambridge,
1995.
[33] G. Landi, An Introduction to Noncommutative Spaces and their Geometry, Lec-
ture Notes in Physics: Monographs 51, Springer-Verlag, Berlin Heidelberg, 1997.
[34] M. Lesch and H. Moscovici, Modular curvature and Morita equivalence, Geom.
Funct. Anal. 26 (3), (2016), 818 -- 873.
[35] M. Lesch and H. Moscovici, Modular Gaussian curvature, arXiv:1810.10394.
[36] S. Lord, A. Rennie and J. V´arilly, Riemannian manifolds in noncommutative
geometry, J. Geom. Phys. 62 (2012), 1611 -- 1638.
46
[37] M. Lesch and B. Mesland, Sums of regular self-adjoint operators in Hilbert C∗-
modules, J. Math. Anal. Appl. 472 (1) (2019), 947 -- 980.
[38] B. Mesland, Unbounded bivariant K-theory and correspondences in noncommu-
tative geometry, J. Reine Angew. Math. 69 (2014), 101 -- 172.
[39] B. Mesland and A. Rennie, Nonunital spectral triples and metric completeness in
unbounded KK-theory, J. Funct. An. 271 (9) (2016), 2460 -- 2538.
[40] J. Rosenberg, Levi-Civita's theorem for noncommutative tori, SIGMA, 9 (2013)
071, 9 pages.
[41] R. Senior, Modular spectral triples and KMS states in noncommutative geometry,
Thesis, Australian National University (2011).
[42] F. Tr`eves, Topological vector spaces, distributions and kernels, Academic Press,
1967.
47
|
1012.3494 | 3 | 1012 | 2013-07-12T15:01:05 | Almost commuting self-adjoint matrices --- the real and self-dual cases | [
"math.OA"
] | We show that a pair of almost commuting self-adjoint, symmetric matrices is close to a pair of commuting self-adjoint, symmetric matrices (in a uniform way). Moreover we prove that the same holds with self-dual in place of symmetric. The notion of self-dual Hermitian matrices is important in physics when studying fermionic systems that have time reversal symmetry. Since a symmetric, self-adjoint matrix is real, we get a real version of Huaxin Lin's famous theorem on almost commuting matrices. Similarly the self-dual case gives a version for matrices over the quaternions.
We prove analogous results for element of real C^*-algebras of "low rank." In particular, these stronger results apply to paths of almost commuting Hermitian matrices that are real or self-dual. Along the way we develop a theory of semiprojectivity for real C^*-algebras. | math.OA | math |
ALMOST COMMUTING SELF-ADJOINT MATRICES - THE
REAL AND SELF-DUAL CASES
TERRY A.LORING AND ADAM P. W.SØRENSEN
Abstract. We show that a pair of almost commuting self-adjoint, symmetric
matrices is close to a pair of commuting self-adjoint, symmetric matrices (in a
uniform way). Moreover we prove that the same holds with self-dual in place of
symmetric. The notion of self-dual Hermitian matrices is important in physics
when studying fermionic systems that have time reversal symmetry. Since a
symmetric, self-adjoint matrix is real, we get a real version of Huaxin Lin's
famous theorem on almost commuting matrices. Similarly the self-dual case
gives a version for matrices over the quaternions.
We prove analogous results for element of real C ∗-algebras of "low rank."
In particular, these stronger results apply to paths of almost commuting Her-
mitian matrices that are real or self-dual. Along the way we develop a theory
of semiprojectivity for real C ∗-algebras.
1. Introduction
In 1997 Lin proved an important theorem about almost commuting matrices [27].
Nowadays it is know as Lin's theorem. Loosely speaking it states that two almost
commuting self-adjoint matrices are close to commuting self-adjoint matrices, and
in a way that is uniform over all dimensions. Formally it says:
Theorem 1.1. (Lin) For all ε > 0 there exists a δ > 0 such that for all n ∈ N the
following holds: Whenever A, B ∈ Mn(C) are two self-adjoint, contractive matrices
such that kAB − BAk < δ there exists self-adjoint matrices A′, B′ ∈ Mn(C) such
that A′B′ = B′A′ and
kA − A′k,kB − B′k < ε.
Friis and Rørdam [16] generalized Lin's theorem to hold for elements in a variety
of C∗-algebras of "low dimension." In particular, their result applies to paths of
almost commuting matrices.
As one might expect, such a fundamental result in linear algebra as Lin's The-
orem has had applications outside of C∗-algebras. We believe that many more
applications are possible if Lin's theorem is generalized in a few directions.
Applications of almost commuting matrices to quantum mechanics were consid-
ered by von Neumann [44]. The claim made by von Neumann was not carefully
stated, but his underlying assertion regarding macrocopic observables was proven
recently by Ogata [37], using the techniques in Lins' proof. The applications of Lin's
theorem, and related results on almost commuting unitary matrices, was observed
by Hastings [20] and then pursued in a series of papers [22, 23, 29].
More recently, M.M. Bronstein and his coauthors [25, 17] have been applying
Lin's theorem, including the real version, to problems in machine learning. Lin's
2000 Mathematics Subject Classification. Primary: 46L85; Secondary: 15B33.
1
2
TERRY A.LORING AND ADAM P. W.SØRENSEN
theorem is used there to prove results regarding commutators measured by the
Frobenius norm
kAkF =Xj Xk
ajk2
1
2
.
Computer scientist and physicists are less familiar with Lin's theorem than they
are with its close relative called "Joint Approximate Diagonalization" (JADE). This
algorithm, developed by Cardoso and Souloumiac for use in blind source separation
[8], takes two or more matrices and finds a unitary change of basis to make both
matrices approximately diagonal. Equivalently [17], the algorithm finds commuting
matrices as close as possible to the original matrices, as measured in the Frobenius
norm.
This is closely related to the problem of finding a small perturbation of an almost
commuting pair of matrices to a commuting pair. JADE minimizes off-diagonal
parts in a sense other than as measured by the operator norm. There is a real
version of the algorithm that is probably used far more often than the complex
version. The real version of the Cardoso and Souloumiac algorithm is used in
computer vision [14]. It has also been used in first principal molecular dynamics
simulations [19]. In that case, it is to produce (generalized) Wannier functions.
In the present context, Wannier functions are not really functions at all, but
rather joint approximate eigenvectors. The translation from joint approximate
eigenvectors to Wannier functions is discussed at length in [23]. Suffice it to say,
the smaller the error in the eigenequation in the finite model, the more localized
the corresponding Wannier wave function in the valence band.
We see three directions to go in generalizing Lin's theorem so that it might be
have broad applications. The same goes for applications of theorems related to
other approximate relations, especially theorems about almost commuting unitary
matrices.
The first direction is to deal with anti-unitary symmetries. The role of anti-
unitary symmetries in physics has greatly expanded in the past decade. Topological
insulators were predicted in 2005 and soon thereafter realized experimentally in
the form of the quantum spin Hall effect [39]. Critical here is the time-reversal
invariance of the system, where mathematically time-reversal is implemented by an
anti-unitary operator, such as the transpose. Time reversal symmetry differs for
Bosons and Fermions, and it can be absent. The matrices relevant to physics then
at least include the real, complex and quaternionic, corresponding to what Dyson
called the three-fold way [11].
The present paper takes Lins' theorem down two additional paths in the three-
fold way. We note, however, that there can be a second anti-unitary symmetry to
be dealt with in some quantum systems, and one needs really to be doing linear
algebra with respect to the ten-fold way [1] of Atland and Zirnbauer. A good place
to learn about the ten-fold way and its relation to KO-theory is [15]. However, the
three-fold way is quite enough for this one paper.
The second direction is to find an algorithm to implement Lin's theorem, in any
form. The operator norm leads to the rather unfamiliar Finsler geometry [3] so
this may be very difficult. However in a one-dimensional setting, where the almost
commuting Hermitian matrices X and Y also approximately satisfy X 2 + Y 2 = I,
there are algorithms available, as discussed in [30] and [45, §4].
ALMOST COMMUTING SELF-ADJOINT MATRICES
3
The third direction is to deal with continuously parametrized sets of almost
commuting matrices. From K-theoretical considerations, we must consider only
one-dimensional parameter spaces. An emerging area in topological insulators are
the Floquet topological insulators [28]. These have a Hamiltonian that varies peri-
odically in time. Theorem 7.11 may lead to applications regarding the existence of
continuously varying (generalized) Wannier functions in Floquet systems in three
symmetry classes. However, it probably needs to be generalized to deal with paths
unitary matrices in place of Hermitian matrices before it finds direct applications.
Some results about almost commuting unitary matrices in the presence of anti-
unitary symmetries will be addressed in [35, 31].
Let us focus again on the present paper. Two paths on the three-fold way
involve real and complex matrices. The third path is seen by the physics world are
involving structured complex matrices, but can just as well be seen as involving
matrices over the quaternions. Our results about self-dual Hermitian matrices thus
can be reinterpreted as results about Hermitian matrices of quaternions. This is
discussed leisurely in [34].
We have two main theorems, which we state as one.
Theorem 1. For all ε > 0 there exists a δ > 0 such that for all n ∈ N the following
holds: Whenever A, B are two n-by-n, contractive, self-adjoint, real (resp. self-dual)
matrices such that kAB − BAk < δ there exists n-by-n, self-adjoint, real (resp. self-
dual) matrices A′, B′ such that A′B′ = B′A′ and
kA − A′k,kB − B′k < ε.
There are essentially three known proofs of (the complex case of) Lin's theorem:
Lin's original proof and the two proofs given by Friis-Rørdam [16]. These later
proofs generalize Lin's theorem beyond matrices. Working solely with the matrix
case, Hastings has derived a quantitative version of Lin's theorem [21].
The second Friis-Rørdam proof achieves a greater generalization of Lin's the-
orem than their first proof.
It is the only proof that shows that Lin's theorem
remains true for almost commuting elements of C∗-algebras of stable rank one.
This proof depends on the semiprojectivity of C(X) for X a one-dimensional, finite
CW complex [32].
We modeled our proof on the second Friis-Rørdam proof. This requires that we
develop semiprojectivity of real C∗-algebras to to the point of proving the semipro-
jectivity of C(X, R) for X a one-dimensional, finite CW complex.
We get then a stronger version of our main result, which we present in Section 7.
This applies to paths of almost commuting matrices.
2. Real C∗-algebras
2.1. Two types of real C∗-algebras. In the past, there have been two ways
to talk about real C∗-algebras. There have been real C∗-algebras (that is, with
lowercase r) and Real C∗-algebras (with uppercase R). For general background on
real/Real C∗-algebras, see [18] and [26]. Real and real C∗-algebras are different
objects, and even though they are closely related the similar names cause confusion
(especially in verbal communication). We are not the first to feel this way. See, for
example, [36, page 698] regarding Atiyah's [2] use of Real and real as distinct terms.
Adding to the confusion is that fact the Real C∗-algebras have C as their scalar
field. In fairness to Atiyah we should mention that the category of real spaces sits
4
TERRY A.LORING AND ADAM P. W.SØRENSEN
nicely inside the category of Real spaces, thus reducing potential confusion. This,
however, is not true for noncommutative C∗-algebras. To minimize confusion we
suggest new names. Using these names we present some basic facts. A lot of this
is either known (though not necessarily written down) or not so hard to prove, so
we will omit some proofs.
First we describe a class of algebras with scalar field R.
Definition 2.1. Given a real Banach ∗-algebra we let AC be the set of formal sums
a1 ∔ i · a2, a1, a2 ∈ A. Letting a1, a2, b1, b2 ∈ A and α, β ∈ R we define algebraic
operations on AC by:
(a1 ∔ i · a2) + (b1 ∔ i · b2) = (a1 + b1) ∔ i · (a2 + b2),
(a1 ∔ i · a2)(b1 ∔ i · b2) = (a1b1 − a2b2) ∔ i · (a2b1 + a1b2),
(a1 ∔ i · a2)∗ = a∗1 ∔ i · (−a∗2),
(α + βi)(a1 ∔ i · a2) = (αa1 − βa2) ∔ i · (αa2 + βa1).
With those operations AC is a complex ∗-algebra. It is called the complexification
of A.
Definition 2.2. A real Banach ∗-algebra A is called an R∗-algebra if there exist a
norm on AC such that AC becomes a C∗-algebra, and the norm on AC extends the
norm on A.
Remark 2.3. An R∗-algebra is known in the literature as a real C∗-algebra [38].
We have the obvious morphisms, and with those we have a category.
Definition 2.4. A map φ : A → B between two R∗-algebras is called an R∗-
homomorphism if it is R-linear, multiplicative and ∗-preserving.
Definition 2.5. Denote by R∗ the category with objects all R∗-algebras and mor-
phisms all R∗-homomorphisms. Denote by R∗1 the category of unital R∗-algebras
and unital morphisms.
We will also define a class of algebras that is seemingly closer to C∗-algebras.
The motivation for this is that the real matrices can be described as those complex
matrices where the adjoint and the transpose agree. We define something similar
to the transpose in a more general setting.
Definition 2.6. Let A be a C∗-algebra. A linear and ∗-preserving map τ : A → A
such that τ (ab) = τ (b)τ (a), and τ (τ (a)) = a for all a, b ∈ A is called a reflection
on A.
Remark 2.7. A reflection is an order two anti-automorphism of A. Thus it is
automatically norm preserving and continuous. Furthermore, the 0 element in A
must be mapped to 0 by τ , if A has a unit it too must be mapped to itself by τ , and
for any a ∈ A the spectrum of a equals that of τ (a).
Definition 2.8. A C∗,τ -algebra is a pair (A, τ ) where A is a C∗-algebra and τ is
a reflection of A. We will often write τ (a) as aτ .
Similar to how the letter d is almost always used to represent a generic metric,
we will write (A, τ ) when we do not know anything special about τ .
Remark 2.9. The Real C∗-algebras correspond to C∗,τ -algebras.
ALMOST COMMUTING SELF-ADJOINT MATRICES
5
We also have morphisms between C∗,τ -algebras, and so we also get a category.
Definition 2.10. By a C∗,τ -homomorphism (or ∗-τ -homomorphism) we mean a
map φ : (A, τ ) → (B, τ ) such that φ is a ∗-homomorphism from A to B and φ(aτ ) =
φ(a)τ for all a ∈ A.
Definition 2.11. Let C∗,τ be the category with objects all C∗,τ -algebras and mor-
phisms all ∗-τ -homomorphisms. Let C∗,τ
be the category of unital C∗,τ -algebras
and unital morphisms.
1
2.2. Connections between R∗ and C∗,τ . We will now consider the close rela-
tionship between R∗-algebras and C∗,τ -algebras. We have a notion of real elements
inside a C∗,τ -algebra.
Definition 2.12. Given a ∈ (A, τ ) we let ℜτ (a) = (a + a∗τ )/2.
We will say that a is a real element or that it is in the real part of (A, τ ) if
ℜτ (a) = a. This happens precisely when a∗ = aτ .
Lemma 2.13. If a ∈ (A, τ ) then
a = ℜτ (a) − iℜτ (ia).
Lemma 2.14. If a ∈ (A, τ ) and we can write a = a1 + ia2 with a1 and a2 in the
real part of A then a1 = ℜτ (a) and a2 = ℜτ (−ia).
We use this new-found knowledge to show that inside all C∗,τ -algebras lives an
R∗-algebra.
Proposition 2.15. If (A, τ ) is a C∗,τ -algebra then {a ∈ A a∗ = aτ} is an R∗-
algebra.
Proof. Let A0 = {a ∈ A a∗ = aτ}. The map from A to (A0)C sending a ∈ A to
ℜτ (a) ∔ i · ℜτ (−ia) is an R∗-isomorphism.
(cid:3)
We now define a functor from C∗,τ to R∗.
Definition 2.16. Define ℜ : C∗,τ → R∗ on objects by
ℜ((A, τ )) = {a ∈ A a∗ = aτ},
and if φ : (A, τ ) → (B, τ ) we let
where we co-restrict the right hand side to ℜ((B, τ )).
We also wish to have a functor from R∗ to C∗,τ .
ℜ(φ) = φℜ(A,τ ),
Lemma 2.17. If A is an R∗-algebra then ¯∗ : AC → AC given by
(a1 ∔ i · a2)¯∗ = a∗1 ∔ i · a∗2,
is a reflection on AC. Furthermore ℜ(AC, ¯∗) ∼= A.
Definition 2.18. Define ⋆ to be the functor from R∗ to C∗,τ that maps an R∗-
algebra A to (AC, ¯∗) and an R∗-homomorphism φ : A → B to ⋆(φ) : (AC, ¯∗) →
(BC, ¯∗) given by
⋆(φ)(a1 ∔ i · a2) = φ(a1) ∔ i · φ(a2).
It is not obvious that ⋆ is a functor, but on the other hand it is not hard to
prove.
6
TERRY A.LORING AND ADAM P. W.SØRENSEN
Remark 2.19. The functor ⋆ maps surjections to surjections and injections to
injections.
It can be shown that our two functors are almost inverses. That is, if A is an
R∗-algebra and (B, τ ) is a C∗,τ -algebra, then
⋆(ℜ(B, τ )) ∼= (B, τ ),
and ℜ(⋆(A)) ∼= A.
In fact it is know that they both yield categorical equivalences. As such a lot of
the study of R∗-algebras can be done using C∗,τ -algebras. This is the approach
we will take throughout this paper. The reasoning behind this choice is that the
C∗,τ -algebras lets us utilize a lot of our C∗-algebra knowledge. Hence there is less
reproving of theorems.
2.3. Two examples.
Example 2.20. We modeled a reflection on the transpose so of course it is a
reflection, and ℜ(Mn(C), T ) = Mn(R).
blocks and define
There is another reflection on M2n(C). If A ∈ M2n(C) we let Aij be the n × n
A21 A22(cid:19)♯
(cid:18)A11 A12
=(cid:18) AT
−AT
22 −AT
21 AT
12
11 (cid:19) .
This is a reflection, and ℜ(M2n(C), ♯) = Mn(H), where H is the quaternions. This is
an important operation in physics, as is discussed in the survey [46] of applications
of random matrices in physics.
Example 2.21. Consider the C∗-algebra of continuous complex-valued functions
on the circle C(S1). Since C(S1) is abelian, a reflection is just an order-two isomor-
phism. Hence any reflection will come from an order-two homeomorphism of the
circle. From [10] we glean that there are only three such maps (up to conjugation),
namely:
(1) z 7→ −z,
(2) z 7→ z, and,
(3) z 7→ z.
Each gives rise to a C∗,τ -algebra by defining for instance f τ (z) = f (−z). The real
parts will be
(1) {f ∈ C(S1, C) f (z) = f (−z) for all z ∈ S1},
(2) {f ∈ C(S1, C) f (z) = f (z) for all z ∈ S1}, and,
(3) {f ∈ C(S1, C) f (z) = f (z)} ∼= C(S1, R).
As there are two essentially distinct reflections on M2n(C) and three on C(S1),
we immediately find six replacements in the real case for
U2n(A) ∼= hom(cid:0)C(S1), M2n(A)(cid:1) .
Hence we get six objects that could correspond to the 6 odd real K-groups, counting
degrees 1, 3, 5, and 7 in KO and degrees 1 and 3 in self-conjugate K-theory [2, 6].
2.4. Ideals in and operations on. We wish to study ideals in C∗,τ -algebras. In
C∗-algebras the ideals are precisely the kernels of ∗-homomorphisms. The kernel
of a C∗,τ -homomor-phism will be self-τ (that is, if x ∈ ker φ then xτ ∈ ker φ), but
there are C∗-ideals that need not be self-τ . We wish to eliminate those ideals, and
so we give the following definition.
ALMOST COMMUTING SELF-ADJOINT MATRICES
7
Definition 2.22. Let (A, τ ) be a C∗,τ -algebra. We say that I ⊆ A is an ideal in
(A, τ ) if I is a C∗-ideal in A and I is self-τ . We will sometimes write I ⊳τ A or
I ⊳ (A, τ ).
With ideals at hand, we can define quotients.
Lemma 2.23. If I ⊳ (A, τ ) then (I, τI ) is a C∗,τ -algebra. Let π : A → A/I be the
quotient map. The map π(a)τ 7→ π(aτ ) defines a reflection on A/I. Thus, A/I is
naturally a C∗,τ -algebra and π is C∗,τ -homomorphism.
We note that we now have obtained what we wanted: The C∗,τ ideals are pre-
cisely the kernels of the C∗,τ -homomorphisms.
The following lemma and theorem tells us that we have direct sums and pullbacks
in the category C∗,τ .
Lemma 2.24. Given two C∗,τ -algebras (A, τ ) and (B, τ ) the map τ ⊕ τ : A⊕ B →
A ⊕ B will be a reflection.
Theorem 2.25. Suppose ϕ1 : (A1, τ ) → (C, τ ) and ϕ2 : (A2, τ ) → (C, τ ) are ∗-τ -
homomorphisms, and form the pull-back C∗-algebra
A1 ⊕C A2 = {(a1, a2) ∈ A1 ⊕ A2 ϕ1(a1) = ϕ2(a2)} .
This becomes a C∗,τ -algebra with
(a1, a2)τ = (aτ
1 , aτ
2)
and it gives us the pull-back of the given C∗,τ -algebras, where we are using the
restricted projection maps πj : (A1, τ ) ⊕(C,τ ) (A2, τ ) → (Aj , τ ).
We can also define what it means to unitize a C∗,τ -algebra.
Lemma 2.26. Let (A, τ ) be a C∗,τ algebra. The formula
(a + λ1)σ = aτ + λ1,
a ∈ A, λ ∈ C,
defines a reflection on A. Thus ( A, σ) is a C∗,τ -algebra. And it is the only way to
unitize (A, τ ) while preserving the reflection on A.
Definition 2.27. If (A, τ ) is a C∗,τ -algebra we will also denote by τ the extension
of τ to A given in lemma 2.26 (this should cause no confusion, as the lemma shows
this extension is unique). The C∗,τ -algebra ( A, τ ) we denoted by ^(A, τ ) or (A, τ )∼,
and call it the unitization of (A, τ ).
Example 2.28. The unitization of the C∗,τ -algebra C0((0, 1), id) is C(S1, id).
3. (Semi) Projective real C∗-algebras
The definition of a semiprojective C∗-algebra commonly used today was given by
Blackadar in [4]. We will modify that definition so we can use it for C∗,τ -algebras.
The theory of semiprojective C∗-algebras is well developed; for good resources on
the subject see [5], [33], and the references therein.
In what follows we try to
develop some theory of semiprojective C∗,τ -algebras.
Semiprojectivity is the first key ingredient in shape theory [12]. The other key
ingredient is K-theory, which we do not see here. The K-theory is uninteresting
when working with one-dimensional spaces, as in this section. In section 6 we are
working with the disk or square so the K-theory is still not a factor, this time due
8
TERRY A.LORING AND ADAM P. W.SØRENSEN
to the contractability of these space. See [23] for how K-theory will make life more
interesting when we get to the sphere and torus.
A semiprojective C∗-algebra is an analog of an absolute neighborhood retract
(ANR). Indeed, a necessary, but far from sufficient condition, for A to be semipro-
jective is that its abelianization have spectrum an ANR. Semiprojective C∗-algebras
interact well with inductive limits. One approach to shape theory (take by Effros
and Kaminker in [12]) is to write two C∗-algebras of interest as inductive limits of
semiprojective C∗-algebras, then intertwine the semiprojective C∗-algebras to pro-
duce a substitute for a homotopy equivalence when a true homotopy equivalence is
not possible.
It was later discoverer, by Dadarlat [9], that shape theory is closely related to
E-theory. In Dadarlat's approach to shape theory, semiprojectivity plays only a
supporting role.
Neither shape theory nor E-theory have been worked out for real C∗-algebras,
but questions of semiprojectivity and the lack-thereof in real C∗-algebras are now
seen to arise in physics [23]. Without prejudice, we walk around those subjects and
attack semiprojectivity for real C∗-algebras head-on.
Our task is to incorporate reflections in the subject of semiprojectivity. Our
goal is to prove that C(X, id) is semiprojective for any finite one dimensional CW
complex. We will only achieve that goal once we reach the end of section 5.
In this and the next two sections we will borrow proof techniques from across
the field of semiprojectivity in C∗-algebras without further references.
3.1. Definitions and the basics. We give the obvious definitions of projectivity
and semiprojectivity in the categories C∗,τ and C∗,τ
1 .
Definition 3.1. Let C be one of the categories C∗,τ or C∗,τ
1 . An object A in C
is said to be projective, if whenever J is an ideal in B, another object in C, and
we have a morphism φ : A → B/J in C, we can find a morphism ψ : A → B in C
such that π ◦ ψ = φ, where π it the quotient map from B to B/J.
Definition 3.2. Let C be one of the categories C∗,τ or C∗,τ
1 . An object A in C
is said to be semiprojective, if whenever J1 ⊆ J2 ⊆ ··· is an increasing sequence
of ideals in B, another object in C, and we have a morphism φ : A → B/J, J =
∪nJn, in C, we can find an m ∈ N and morphism ψ : A → B/Jm in C such that
πm,∞ ◦ ψ = φ, where πm,∞ it the quotient map from B/Jm to B/J.
Notation. Whenever we have a C∗,τ -algebra B containing an increasing sequence
of τ -invariant ideals J1 ⊆ J2 ⊆ ··· we denote the quotient maps as follows:
πn : B ։ B/Jn,
πn,m : B/Jn ։ B/Jm,
πm,∞ : B/Jm ։ B/J,
π∞ : B ։ B/J,
where n < m are natural numbers and J = ∪nJn.
ALMOST COMMUTING SELF-ADJOINT MATRICES
9
Pictorially, (A, τ ) is semiprojective, if we can always solve the following lifting
problem
(B, τ )
(B/Jn, τ )
9s
s
(B/J, τ )
s
s
s
φ
(A, τ )
Of course one could just as easily define semiprojective R∗-algebras. Studying
how the functors ℜ and ⋆ behave with respect to ideals and lifting problems, the
following two propositions can be proved. For reasons of brevity we have chosen
not to include proofs of these propositions.
Proposition 3.3. Let A, B be R∗-algebras, let J be an ideal in B, and let φ : A →
B/J be an R∗-homomorphism. There exists an R∗-homomorphism ψ : A → B such
that π ◦ ψ = φ if and only if there exists a C∗,τ -homomorphism χ : ⋆(A) → ⋆(B)
such that ⋆(π) ◦ χ = ⋆(φ).
Proposition 3.4. If A is an R∗-algebra then A is (semi-) projective if and only if
⋆(A) is. If (B, τ ) is a C∗,τ -algebra then (B, τ ) is (semi-) projective if and only if
ℜ(B, τ ) is.
Just as in the C∗-case we can somewhat simplify the task of proving semipro-
jectivity.
Proposition 3.5. To show that a C∗,τ -algebra (A, τ ) is semiprojective it suffices
to solve lifting problems where φ : (A, τ ) → (B/J, τ ) is either injective, surjective
or both.
Proof. This is well know in the C∗-case [33], and is no harder in the C∗,τ -case.
To get injectivity we replace φ with φ ⊕ id : (A, τ ) → (B/J ⊕ A, τ ⊕ τ ). To get
(cid:3)
surjectivity we focus on the image of φ.
Functional calculus is indispensable when working with lifting problems. The
following lemma tells some of the story about C∗,τ -algebras and functional calculus.
Lemma 3.6. Suppose a is a normal element in a C∗,τ -algebra (A, τ ). If f is a
continuous function from σ(a) to C then f (a)τ = f (aτ ).
If b ∈ (A, τ ) is a normal and self-τ element and σ(b) ⊆ X ⊆ C then the C∗-
homomorphism φ : C0(X) → A given by f 7→ f (b) is a C∗,τ -homomorphism from
C(X, id) to (A, τ ).
Proof. We remind the reader that σ(a) = σ(aτ ). Since τ is linear and a is normal
we have p(a)τ = p(aτ ), for any polynomial p in a and a∗. By continuity of τ we
now get f (a)τ = f (aτ ) for any function f ∈ C(σ(a)).
For any function f ∈ C0(X) we have
f (b)τ = f (bτ ) = f (b) = (f ◦ id)(b).
(cid:3)
With that lemma at our disposal, we can give some basic examples of (semi-)
projective C∗,τ -algebras.
/
/
9
10
TERRY A.LORING AND ADAM P. W.SØRENSEN
Example 3.7. The C∗,τ -algebra C0((0, 1], id) is projective. To see this, suppose
we are given the following lifting problem:
(B, τ )
π
C0((0, 1], id)
φ
/ (B/J, τ )
Let h = φ(t 7→ t). Then h is a self-τ positive contraction. Let x be a positive
contractive lift of h, and let k = (x+ xτ )/2. Then k is a self-τ , positive contraction,
and π(x) = h. By Lemma 3.6 the map f 7→ f (k) is a C∗,τ -homomorphism. It is a
lift of φ by standard C∗-theory.
Example 3.8. The C∗,τ -algebra (C, id) is semiprojective. To prove this, suppose
we are given the following lifting problem:
(B, τ )
πn
(B/Jn, τ )
πn,∞
(C, id)
/ (B/J, τ )
φ
Let p = φ(1). Then p is a self-τ projection. Let y ∈ (B, τ ) be any self-adjoint lift
of p. If we let x = (y + yτ )/2 then x is a self-τ and self-adjoint lift of p. Since
πn(x2− x) → 0 as n → ∞ we can find some m ∈ N such that 1/2 /∈ σ(πm(x)). Now
let f be the function that is 0 on (−∞; 1/2) and 1 on (1/2;∞). Then q = f (πm(x))
is a projection and a lift of p. Since x is self-τ , q will be self-τ . We can now define
a C∗,τ -homomorphism from C to B/Jm by λ 7→ λq (Lemma 3.6). It is a lift of φ.
3.2. Closure results.
3.2.1. Unitizing. We aim to get the C∗,τ equivalent of C∗ result that A is semipro-
jective if and only if A is. First we show that if A is unital it suffices to solve unital
lifting problems.
Lemma 3.9. A unital C∗,τ -algebra is semiprojective in C∗,τ if and only if it is
semiprojective in C∗,τ
1 .
Proof. Let (A, τ ) be a unital C∗,τ -algebra.
The proof that (A, τ ) semiprojective in C∗,τ implies that it is semiprojective in
Suppose that (A, τ ) is semiprojective in C∗,τ
1 . Let (B, τ ) be a C∗,τ -algebra
containing an increasing sequence of C∗,τ ideals J1 ⊆ J2 ⊆ ··· , let J = ∪nJn,
and let φ : (A, τ ) → (B/J, τ ) be a C∗,τ -homomorphism. Put p = φ(1A). Then p
is a self-τ projection in (B/J, τ ). Since (C, id) is semiprojective, we can find some
n0 ∈ N and self-τ projection q ∈ B/Jn0 such that πn0,∞(q) = p. For each n ≥ n0
define qn = πn0,n(q). Since all the qn are self-τ all the corners qn(B/Jn)qn are
self-τ . Hence for all n ≥ n0 we have that qn(B/Jn)qn ∼= (qBq)/(qJnq) and that by
C∗,τ
1
is precisely the same as in the C∗-case.
/
/
ALMOST COMMUTING SELF-ADJOINT MATRICES
11
restricting the τ 's we get the following commutative diagram of C∗,τ -algebras:
(qn0 (B/Jn0 )qn0 , τ ) ֒
/ (B/Jn0, τ )
(qn(B/Jn)qn, τ ) ֒
/ (B/Jn, τ )
(A, τ )
φ
/ (p(B/J)p, τ ) ֒
/ (B/J, τ )
In the two left most columns there are only unital maps and algebras, so since
(A, τ ) is semiprojective in the unital category, we can find a lift for some n ≥ n0.
This lifting combines with the inclusion qn(B/Jn)qn ֒→ B/Jn to show that (A, τ )
(cid:3)
is semiprojective.
The lemma is a stepping stone towards a goal, but it also has its own applications.
Example 3.10. The C∗,τ -algebras C(S1, id), C(S1, z 7→ z), and C(S1, z 7→ −z)
are all semiprojective. We will only show the first one, but the remaining proofs are
similar. By Lemma 3.9 it suffices to solve lifting problems of the form:
(B, τ )
πn
(B/Jn, τ )
πn,∞
(C(S1), id)
/ (B/J, τ )
φ
where everything is unital. Let u = φ(z 7→ z). Then u is a self-τ unitary. Let y be
any self-τ lift of u. We can find an m such that x = πm(x) satisfies that xx∗ and
x∗x are invertible. Now define v = x(x∗x)−1/2. The v is a unitary lift of u and, by
Lemma 3.6 and a standard functional calculus trick,
vτ = ((x∗x)−1/2)τ xτ = ((x∗x)τ )−1/2x = (xx∗)−1/2x = x(x∗x)−1/2 = v.
There is C∗-homomorphism from C(S1) to B/Jm given by ψ(f ) = f (v). Since v is
self-τ and every element in C(S1, id) is self-τ , this is actually a C∗,τ -homomorphism
from C(S1, id) to (B/Jm, τ ). Because v is a lift of u, ψ is a lift of φ.
We omit the proof of the next lemma, as the standard C∗ proof carries over
easily.
Lemma 3.11. A C∗,τ -algebra (A, τ ) is semiprojective if and only if ^(A, τ ) is
semiprojective in the unital C∗,τ category.
Corollary 3.12. A C∗,τ -algebra is semiprojective if and only if its unitization is.
Example 3.13. Since C((0, 1), id)∼ ∼= C(S1, id) and the latter is semiprojective,
C((0, 1), id) is semiprojective.
/
/
/
/
/
12
TERRY A.LORING AND ADAM P. W.SØRENSEN
3.2.2. Direct sums. In this section we aim to show the following.
Proposition 3.14. If (A, τ ), (B, τ ) are separable and semiprojective C∗,τ -algebras,
then (A ⊕ B, τ ⊕ τ ) is a semiprojective C∗,τ -algebra.
Before we can do that however, we need to set up some theory.
Lemma 3.15. The relations 0 ≤ h, k ≤ 1, h = hτ , k = kτ , hk = 0 are liftable.
Proof. Suppose we are given a τ -invariant ideal J in a C∗,τ -algebra B, and suppose
h, k ∈ B/J satisfy the relations. Let a = h − k. Then a is a a self-τ self-adjoint
contraction. Thus we can lift it to a self-adjoint self-τ contraction in B, a say.
Define f : R → R by f (x) = (x + x)/2. Then we know from C∗-algebra theory
that f (a) is a positive contractive lift of h, that f (−a) is a positive contractive
lift of k, and that f (a)f (−a) = 0. Lemma 3.6 tells us that f (a) and f (−a) are
(cid:3)
self-τ .
Lemma 3.16. Let (B, τ ) be a C∗,τ -algebra. If h ∈ (B, τ ) is strictly positive in B
then so is hτ . Hence ℜτ (h) is strictly positive.
Proof. Let φ : B → C be a linear positive functional. Then we have, writing τ as a
function,
φ(hτ ) = φ(τ (h)) = (φ ◦ τ )(h).
Since τ is linear and maps positive elements to positive elements φ ◦ τ is a positive
linear functional. But then if φ is non-zero we have
φ(hτ ) = (φ ◦ τ )(h) > 0.
(cid:3)
Corollary 3.17. If (B, τ ) is a separable C∗,τ -algebra then it contains a self-τ
positive element h such that hBh = B.
A discussion of hereditary subalgebras in the context of real C∗-algebras can be
found in [43].
We are now ready to prove Proposition 3.14.
Proof of Proposition 3.14. Since both (A, τ ) and (B, τ ) are separable we can use
Corollary 3.17 to find h ∈ A and k ∈ B, positive contractions such that hτ = h,
kτ = k, hAh = A and kBk = B. Suppose we are given a C∗,τ -algebra (D, τ )
containing an increasing sequence of τ -invariant ideals J1 ⊆ J2 ⊆ ··· and a C∗,τ -
homomorphism
φ : (A ⊕ B, τ ⊕ τ ) → (D/J, τ ),
where J = ∪nJn. Let h = φ((h, 0)) and k = φ((k, 0)). Since h and k are orthogonal
positive self-τ contractions we can, by Lemma 3.15, find positive orthogonal self-
τ contractive lifts h, k of them in D. For each n ∈ N ∪ {∞} let hn = πn(h),
kn = πn(k), An = hn(D/Jn)hn, and Bn = kn(D/Jn)kn. For each n ∈ N∪{∞} the
map γn = τAn ⊕ τBn is a reflection since hn and kn are self-τ . Observe that we
have
h(D/J)h = h∞(D/J)h∞ = A∞ and
k(D/J)k = k∞(D/J)k∞) = B∞.
Define for each n ∈ N ∪ {∞} a map
αn : (An ⊕ Bn, γn) → (D/Jn, τ ),
ALMOST COMMUTING SELF-ADJOINT MATRICES
13
by α((x, y)) = x + y. It will be an C∗,τ -homomorphism since hnkn = 0. Noticing
that
π(hDh) = A∞ and
we see there must be a C∗,τ -homomorphism
π(kDk) = B∞,
ψ : (A ⊕ B, τ ⊕ σ) → (A∞ ⊕ B∞, γ∞),
such that φ : α∞◦ ψ. Hence we get the following commutative diagram for all n ∈ N
(An ⊕ Bn, γn)
αn
/ (D/Jn, τ )
(A ⊕ B, τ ⊕ σ)
ψ
/ (A∞ ⊕ B∞, γ∞)
α∞ /
/ (D/J, τ )
φ
Since γ is a direct sum of two reflections, we can use the semiprojective of (A, τ )
and (B, σ), one at a time, to show that (A ⊕ B, τ ⊕ σ) is semiprojective.
(cid:3)
Remark 3.18. We observe that we only used (A, τ ) and (B, τ ) separable to get
strictly positive real elements h, k. So we might as well have assumed that A and
B were σ-unital. Lemma 3.16 tells us that whether we define (A, τ ) to be σ-unital
when A is or when (A, τ ) contains a strictly positive real element, we get the same
class of algebras.
The knowledge we have accumulated so far lets us take a small step towards
showing that if X is a finite one-dimensional CW-complex then C(X, id) is semipro-
jective.
Proposition 3.19. If X is a wedge of circles (a bouquet) then the C∗,τ -algebra
C(X, id) is semiprojective.
Proof. By assumption
C(X, id) ∼= nMi=1
for some n ∈ N. By Proposition 3.14 Ln
C0((0, 1), id)! f
,
each summand is. So by Corollary 3.11 C(X, id) is semiprojective.
i=1 C0((0, 1), id) is semiprojective since
(cid:3)
The above proposition will later be the basis step of an induction proof.
Remark 3.20. If X is a wedge of two circles, then we can put a reflection on
C(X) by mapping one circle to the other. This reflection is not a direct sum of
two reflections on the circle. Hence showing that the C∗,τ -algebra it defines is
semiprojective requires different techniques than the ones we have just used.
4. Multiplier algebras
In this section we will study multiplier and corona algebras of C∗,τ -algebras.
The idea is that we already have multiplier algebras at our disposal. So the main
body of work lies in showing that we can extend a reflection on A to a reflection
on M (A).
/
6
6
/
14
TERRY A.LORING AND ADAM P. W.SØRENSEN
4.1. A reflection on M (A). The following theorem is in [24]. We present it here
with a few more details.
Theorem 4.1. Suppose (A, τ ) is a C∗,τ -algebra. There is an operation τ on M (A)
defined by
mτ a = (aτ m)τ ,
and
amτ = (maτ )τ
for a in A and m in M (A), and (M (A), τ ) is a C∗,τ -algebra, and the C∗-inclusion
ι : A → M (A),
is also a C∗,τ -homomorphism.
Proof. Consider for a moment a fixed m in M (A). Define L : A → A and R : A → A
by
L(a) = (aτ m)τ
and R(a) = (maτ )τ .
For all a and b in A,
L(ab) = (bτ aτ m)τ = (aτ m)τ b = L(a)b,
R(ab) = (mbτ aτ )τ = a (mbτ )τ = aR(b),
and
R(a)b = (maτ )τ b = (bτ maτ )τ = a (bτ m)τ = aL(b),
so (L, M ) is an element of M (A), which we denote mτ . Notice mτ is specified
within all multipliers by either one of the formulas
mτ a = (aτ m)τ ,
or
amτ = (maτ )τ .
We claim that the operation defined above, on all multipliers m 7→ mτ , makes
M (A) a C∗,τ -algebra. For any multiplier m, and any a in A,
mτ τ a = (aτ mτ )τ = (maτ τ )τ τ = ma,
so τ ◦ τ = id. For n in M (A) and α in C,
(αm + n)τ a = (aτ (αm + n))τ = αmτ a + nτ a = (αmτ + nτ )a,
(mn)τ a = (aτ mn)τ = ((mτ a)τ n)τ = nτ (mτ a) = (nτ mτ )a,
and
(m∗)τ a = (aτ m∗)τ = (ma∗τ )∗τ = (a∗mτ )∗ = (mτ )∗a,
which means τ commutes with ∗, is anti-multiplicative, and C-linear.
If a is in A, then for any other b in A,
ι(a)τ b = (bτ ι(a))τ = (bτ a)τ = aτ b = ι(aτ )b,
so ι(a)τ = ι(aτ ).
(cid:3)
We now observe that a few more standard constructions automatically respect
reflections.
Lemma 4.2. Suppose A is a C∗,τ -subalgebra of B, where (B, τ ) is a given C∗,τ -
algebra. The idealizer
I(A : B) = {b ∈ B bA + Ab ⊆ A} ,
is self-τ, and so a C∗,τ -subalgebra of B containing A as a self-τ ideal.
Proof. Suppose b is in the idealizer and a is in A. Then aτ ∈ A and so (bτ a)τ = aτ b
is in A. Hence bτ a ∈ A. Likewise (abτ )τ ∈ A so abτ ∈ A.
(cid:3)
ALMOST COMMUTING SELF-ADJOINT MATRICES
15
Theorem 4.3. Suppose (B, τ ) is a C∗,τ -algebra and A ⊳ (B, τ ) is a self-τ ideal.
The unique ∗-homomorphism θ : B → M (A) for which θ(a) = ι(a) for all a in A,
is automatically a ∗-τ -homomorphism.
Proof. We know θ(b)a = ba defines the only possible ∗-homomorphism from B to
M (A) satisfying θ(a) = ι(a). For b in B and a in A we compute
θ(b)τ a = (aτ θ(b))τ = (θ(aτ b))τ = θ((aτ b)τ ) = θ(bτ a) = θ(bτ )a,
which proves θ is τ -preserving.
(cid:3)
Lemma 4.4. Let ϕ : (A, τ ) → (B, τ ) be a proper ∗-τ -homomorphism between σ-
ϕ is actually a ∗-τ -homomorphism.
Proof. The fact that ϕ is proper tells us B = ϕ(A)B = Bϕ(A). The defining
unital C∗,τ -algebras. The unique ∗-homomorphism bϕ : M (A) → M (B) that extends
formulas for bϕ are
and
Therefore
bϕ(m)ϕ(a)b = ϕ(ma)b
bϕ(a)bϕ(m) = bϕ(am).
bϕ(m)τ ϕ(a)b = ((ϕ(a)b)τbϕ(m))τ
= (bτ ϕ(aτ )bϕ(m))τ
= bϕ(mτ )ϕ(a)b.
= (bτ ϕ(aτ m))τ
= ϕ(mτ a)b
(cid:3)
We get "multiplier realization" for free.
Theorem 4.5. Let C(E) denote the corona of a σ-unital C∗,τ -algebra (E, τ ), and
let D and N be separable C∗,τ -subalgebras of C(E). Suppose
A ⊆ C(E) ∩ D′ ∩ N⊥,
is a σ-unital C∗,τ -subalgebra. Then the ∗-τ -homomorphism
θ : I(A : C(E) ∩ D′ ∩ N⊥) → M (A)
is onto.
Proof. We know that θ is onto, by Corollary 3.2 of [13]. All we are asserting here
is that this map is now a morphism in the category of C∗,τ -algebras.
(cid:3)
4.2. Corona extendible morphisms.
Definition 4.6. We say a morphism of C∗,τ -algebras γ : (A, τ ) → (B, τ ) is corona
extendible if, for every ∗-τ -homomorphism ϕ : A → C(E) with E a σ-unital C∗,τ -
Theorem 4.7. Suppose 0 → A → X → P → 0 is a short-exact sequence of σ-unital
C∗,τ -algebras If P is projective then the inclusion A → X is corona extendible.
algebra, there exists a ∗-τ -homomorphism bϕ : B → C(E) so that bϕ ◦ γ = ϕ.
Moreover, the unitization of this map eA → eX is also corona extendible.
16
TERRY A.LORING AND ADAM P. W.SØRENSEN
Proof. Except for the ∗-τ -homomorphism claim, this is Theorem 3.4 of [13] com-
bined with the usual universal property of a split extension, as in Theorem 7.3.6 of
[33]. We summarize those proofs and verify that various maps can be selected to
be ∗-τ -homomorphisms.
Since P is projective, the exact sequence has a splitting by a C∗,τ -homomorphism
λ : P → X. We assume we are given a ∗-τ -homomorphism ϕ : A → C(E) with E
being σ-unital. As in the proof of Theorem 3.4 of [13], we have the commutative
diagram, ignoring for now ψ0,
0
/ A
A
λ
/ X
/ P
0
M (A)
ϕ(A)
/ M (ϕ(A))
ψ0
ϕ(A)
/ I(ϕ(A) : C(E))
where the map A → ϕ(A) is the co-restriction of ϕ making it onto. The essential fact
that the arrow up from the idealizer to the multiplier algebra is both surjective and
a ∗-τ -homomorphism is Theorem 4.5. The map from X to M (A) in the top square
is a ∗-τ -homomorphism by Theorem 4.3. The map from M (A) to M (ϕ(A)) in the
middle square is a ∗-τ -homomorphism by Lemma 4.4. We use the projectivity, in
the ∗-τ -sense, of P to get a ∗-τ -homomorphism ψ0 making the diagram commute.
Following ψ0 by the inclusion into the corona algebra give us a ∗-τ -homomor-
phism ψ : P → C(E) such that
ψ(p)ϕ(a) = ϕ(λ(p)a)
for all p in P and a in A. This induces a ∗-homomorphism
Ψ : X → C(E)
extending ϕ by
Ψ(a + λ(p)) = ϕ(a) + ψ(p)
which is evidently a ∗-τ -homomorphism.
To get the last claim, we must use more of the power of Theorem 4.5. We are
given eA → C(E) which we regard as a ∗-τ -homomorphism ϕ : A → C(E) together
with a projection p in C(E) such that pϕ(a) = ϕ(a) for all a in A. We can replace
I(ϕ(A) : C(E)) in the big diagram by
I(ϕ(A) : C(E)) ∩ (1 − p)⊥.
We still have the needed surjectivity onto M (ϕ(A)) and end up with Ψ : B → C(E)
(cid:3)
with the property pΨ(b) = Ψ(b) for all b in B.
Corollary 4.8. Suppose X is a compact metrizable space and Y ⊆ X is a closed
subset of X homeomorphic to the closed interval [0; 1]. Let X1 be the quotient of X
obtained by collapsing Y to a point. The inclusion C(X1, id) ֒→ C(X, id) of abelian
C∗,τ -algebras is corona extendible.
/
/
/
/
/
s
s
/
/
/
/
O
O
O
O
ALMOST COMMUTING SELF-ADJOINT MATRICES
17
Proof. Let y0 denote the point in Y associated to 0 in [0; 1]. Let y∗ be the point in
X1 that is the image of Y in the quotient map. We have an exact sequence
0
/ C0(X1 \ {y∗})
/ C0(X \ {y0})
/ C0(0; 1]
/ 0
where all C∗-algebras are equipped with the trivial τ operation. Thus we are done
(cid:3)
by Theorem 4.7 and Example 3.7.
4.3. Corona (semi-) projective. Just as in the C∗-case, the work of showing
semiprojectivity can be reduced using corona algebras. Most of the proof of the
following two theorems can be copied from the proof of [33, Theorem 14.1.7] if one
only remembers to change category. The only change is that we have not studied
the Calkin algebra in a C∗,τ setting. To avoid using that, use the corona algebra
^(A, τ ).
ofL∞n=1
Theorem 4.9. Suppose A is a separable C∗,τ -algebra. The following are equivalent:
(1) A is projective;
(2) we can solve the lifting problem for A whenever ρ is the quotient map
(3) we can solve the lifting problem for A whenever ρ is the quotient map
M (E) → C(E) for a separable C∗,τ -algebra E and ϕ is injective;
M (E) → C(E) for a separable C∗,τ -algebra E;
B/I for a separable C∗,τ -algebra B and closed τ -closed ideal I.
(4) we can solve the lifting problem for A whenever ρ is the quotient map B →
Theorem 4.10. Suppose A is a separable C∗,τ -algebra. The following are equiva-
lent:
ideals of E and ϕ is injective;
(1) A is semiprojective;
(2) we can solve the partial lifting problem for A whenever B = M (E) for
(3) we can solve the partial lifting problem for A whenever B = M (E) for
a separable C∗,τ -algebra E and S Ek = E for some chain of τ -invariant
a separable C∗,τ -algebra E and S Ek = E for some chain of τ -invariant
(4) we can solve the partial lifting problem for A whenever B is separable.
Just as in the C∗ case, the corona algebra description of (semi-) projectivity let
ideals of E;
us prove that (semi-) projectivity passes to ideals in certain cases.
Theorem 4.11. Suppose 0 → I → A → B → 0 is an exact sequence of separable
C∗,τ -algebras. If A and B are projective then I is projective.
Proof. We need only lift morphisms of the form I → C(E). These extend to mor-
phisms A → C(E), and those morphisms lift.
(cid:3)
Theorem 4.12. Suppose 0 → I → A → B → 0 is an exact sequence of separable
C∗,τ -algebras. If A is semiprojective and B is projective then I is semiprojective.
5. Functions on graphs
In this section we show semiprojectivity of continuous functions on finite one-
dimensional CW-complexes with the trivial reflection. We will sometimes refer to
finite one-dimensional CW-complexes as graphs. The proof follows the ideas put
forth in [32]. In that paper semiprojectivity of "dimension drop graphs" is shown.
Since we have a specific goal in mind, we have chosen to discard the matrix algebras.
/
/
/
/
18
TERRY A.LORING AND ADAM P. W.SØRENSEN
Theorem 5.1. If X is a finite one-dimensional CW complex, then C(X, id) is a
semiprojective C∗,τ -algebra.
Proof. Since semiprojectivity is closed under direct sums, we can assume that X is
connected. We will do the proof by induction on the number of vertices in X.
The case where X has only one vertex is Proposition 3.19.
Suppose now any one-dimensional CW complex with k vertices gives rise to
a semiprojective C∗,τ -algebra. Let X be a one-dimensional CW complex with
k + 1 vertices. Fix two vertices, v1 and v2 say. Let X be a topological copy of
X. Denote the copies of v1 and v2 in X by w1 and w2 respectively. Choose a
continuous function h0 : X → [−1; 2] such that h−1
0 ([−1; 0]) consists of the union
of closed subintervals, containing w1, of each of the edges adjacent to w1, and
such that h−1
0 ([1, 2]) consists of the same for the edges adjacent to w2, and also
h−1
0 ({−1}) = {w1} and h−1
0 ({2}) = {w2}. We will identify X with the quotient of
X obtained by collapsing h−1
0 ([−1; 0]) to one point and h−1
0 ([1; 2]) to another. Let
γX : X ։ X be the quotient map. Collapsing v1 and v2 to one point we obtain a
space, Y say. Let η : X → Y be the quotient map. Collapsing w1 and w2 in X we
get a space Y , call the quotient map η. And we can collapse arcs in Y to obtain
Y , with quotient map γY say. Thus we have a nice commuting square of quotient
maps
X
γX /
X
η
Y
η
/ Y
γY
We will view C(X, id), C( Y , id) and C(Y, id) as subalgebras of C( X, id) using the
following identifications:
C(X, id) ∼=(cid:26)f ∈ C( X, id)
f (x) = f (w1)
f (x) = f (w2)
C( Y , id) ∼= {f ∈ C( X, id) f (w1) = f (w2)},
C(Y, id) ∼= {f ∈ C( X, id) f (x) = f (w1) if h0(x) ≤ 0 or h0(x) ≥ 1}.
if h0(x) ≥ 1 (cid:27) ,
if h0(x) ≤ 0
Define h1 : X → [0, 1] by
h1(x) =
0,
h0(x),
1,
h0(x) ≤ 0
0 ≤ h0(x) ≤ 1
1 ≤ h0(x)
.
Note that h1 and C(Y, id) generate C(X, id).
Suppose now that we are given a C∗,τ -algebra (E, τ ) containing an increasing
sequence of τ -invariant ideals E1 ⊆ E2 ⊆ ··· such that ∪nEn = E, and an injective
C∗,τ -homomorphism φ : C(X, id) → (C(E), τ ). Putting some of the quotient maps
and φ into one diagram, we have the following.
C( Y , id)
(η)∗
C( X, id)
γX ∗
C(X, id)
φ
/ (C(E), τ )
/
/
/
/
/
/
o
o
/
ALMOST COMMUTING SELF-ADJOINT MATRICES
19
where −∗ denotes the induced maps. Using Corollary 4.8 repeatedly we get a
C∗,τ -homomorphism φ : C( X, id) → (C(E), id) such that φ = φ ◦ γX∗. Using that
Y is a one-dimensional CW complex with one vertex less than X we get that
C( Y , id) is semiprojective, so we can find an n ∈ N and a C∗,τ -homomorphism
ψ : C( Y , id) → (M (E)/En, τ ) such that πn,∞ ◦ ψ = φ ◦ (η)∗. All in all we have the
following commutative diagram
ψ
C( Y , id)
(η)∗
(M (E)/En, τ )
πn,∞
C( X, id)
(γX )∗
C(X, id)
φ
/ (C(E), τ )
φ
We will now find a lift of φ(h1) in (M (E)/En, τ ) that is positive contractive self-
τ and commutes with (ψ◦ (γY )∗)(C(Y, id)). Since φ(h1) is positive and contractive,
we can find a positive and contractive lift. Averaging this lift with τ of it, we
get a self-τ positive contractive lift of φ(h1). Let us call it H. Define functions
l, m, k : [−1; 2] → [0; 1] by
l(t) =
m(t) =
k(t) =
,
,
−1 ≤ t ≤ 0,
0,
0 ≤ t ≤ 1,
t,
2 − t, 1 ≤ t ≤ 2
−t,
−1 ≤ t ≤ 0,
0 ≤ t ≤ 1,
0,
t − 1, 1 ≤ t ≤ 2
0, −1 ≤ t ≤ 0,
0 ≤ t ≤ 1,
t,
1, 1 ≤ t ≤ 2
.
Observe that l + mk = k, that k ◦ h0 = h1, and that l ◦ h0 and m ◦ h0 both are in
C( Y , id). Hence we can define
H = ψ(l ◦ h0) + ψ((m ◦ h0)1/2)Hψ((m ◦ h0)1/2).
Since all the functions are real valued we get that H is self-adjoint. Since every
thing else is self-τ so is H. It is a lift of φ(h1) since
πn,∞( H) = φ(l ◦ h0) + φ((m ◦ h0)1/2) φ(h1) φ((m ◦ h0)1/2)
= φ((l ◦ h0) + (m ◦ h0)h1) = φ((l ◦ h0) + (m ◦ h0)(k ◦ h0))
= φ((l + mk) ◦ h0) = φ(k ◦ h0) = φ(h1).
By functional calculus one can replace H with H = k( H) to obtain a positive
contractive lift of φ(h1). By Lemma 3.6 H is self-τ . To show that this lift commutes
with (ψ ◦ (γY )∗)(C(Y, id)) it suffices to show that H does. Let f ∈ C(Y, id). Then
f (m ◦ h0) = 0 so we must have
Hence ψ((γY )∗(f )) commutes with H.
ψ((γY )∗(f ))ψ(m ◦ h0) = 0.
/
/
3
3
o
o
/
20
TERRY A.LORING AND ADAM P. W.SØRENSEN
Let D = C(Y × [0, 1]). We have shown that given a C∗,τ -homomorphism
φ : C(X, id) → (C(E), τ ) we can find an n0 ∈ N and a C∗-homomorphism χ : D →
M (E)/En0 such that the following diagram commutes
D
β
χ
M (E)/En0
πn0 ,∞
C(X)
/ C(E)
φ
Here β denotes the map induced by sending C(Y, R) (inside D) to C(Y, R) (inside
C(X, R)) and h to h1. Since H is self-τ Lemma 3.6 gives that χ is actually τ -
preserving. Hence we can view the above diagram as being a commutative diagram
in the C∗,τ category.
For each n ≥ n0 define χn = πn0,n ◦ χ and let Dn = D/ ker χn. Then if
n0 ≤ n ≤ n′ we have a surjection Dn ։ Dn′ . Since D = C(Y × [0; 1]) and each Dn,
n ≥ n0, is a quotient of D, there must be spaces Yn, n ≥ n0, such that Dn ∼= C(Yn).
Thus we have an inductive system
D ։ C(Yn0 ) ։ C(Yn0+1) ։ C(Yn0+2) ։ ···
Call the bonding maps δn,n′. This is an inductive system in the category of C∗-
algebras, so we can compute the limit as
D/ ker(πn0,∞ ◦ χ) ∼= (πn0,∞ ◦ χ)(D) = (φ ◦ β)(D) = φ(C(X)) ∼= C(X).
Since X is an ANR we can find an n1 ≥ n0 and a C∗-homomorphism λ : C(X) →
C(Yn1 ) such that δn1,∞ ◦ λ = id. Clearly λ and δn1,∞ are C∗,τ -homomorphisms if
we equip all the commutative algebras with the identity reflection.
Consider the the following commutative diagram.
(D, id)
χ
(M (E)/En0 , τ )
δn1
πn0,n1
(C(Yn1 ), id)
/❴❴❴
(M (E)/En1 , τ )
δn1 ,∞
πn1,∞
C(X, id)
/ (C(E), τ )
φ
Since all the vertical maps are quotient maps, we can fit a C∗,τ -homomor-phism on
the dashed arrow in such a way that the diagram continues to commute. Call this
homomorphism µ. We claim that µ ◦ λ is a lift of φ. To see that, we compute
πn1,∞ ◦ µ ◦ λ = φ ◦ δn1,∞ ◦ λ = φ.
(cid:3)
6. Variations on Lin's theorem
From here on out we more less just follow the proof in [16], modifying their
techniques to keep track of reflections.
In this section we write Mn for Mn(C).
/
/
/
/
/
/
/
ALMOST COMMUTING SELF-ADJOINT MATRICES
21
6.1. Approximating normal elements. The following two lemmas gives self-τ
versions of known facts. One that Mn has stable rank one, and the second that we
can do polar decomposition of invertible elements.
Lemma 6.1. Let τ be a reflection on Mn. For any ε > 0 and any self-τ matrix
A ∈ (Mn, τ ) we can find a self-τ invertible matrix B such that kA − Bk < ε.
Proof. If A is invertible there is nothing to prove. So suppose A is not invertible.
Consider the path of self-τ matrices Bt = (1−t)A+tI. Define a function p : [0, 1] →
C by
p(t) = det(Bt).
By definition of det and Bt the function p is a polynomial in t. Since p(0) =
det(B0) = det(A) = 0 and p(1) = det(B1) = det(I) = 1, p is not constant. Hence
it has only finitely many zeros. Thus for any ε > 0 we can find a t0 such that
0 < t0 < ε/(kA − Ik) and p(t0) 6= 0. Then Bt0 is self-τ and invertible, and
kA − Bt0k = kAt0 − It0k ≤ t0kA − Ik < ε.
(cid:3)
Lemma 6.2. Let a be a self-τ invertible element in a unital C∗,τ -algebra (A, τ ).
Then a can be written as a = up where u is a self-τ -unitary and p = (a∗a)1/2.
Proof. Since a is invertible so is a∗a. Therefore we can define u = a(a∗a)−1/2 and
p = (a∗a)1/2. Then u is a unitary and up = a. By using Lemma 3.6 and standard
functional calculus tricks, we get
uτ = ((a∗a)−1/2)τ aτ = ((a∗a)τ )−1/2a = (aa∗)−1/2a = a(a∗a)−1/2 = u
(cid:3)
Let (nj) be a sequence of natural numbers and let τj be a reflection on Mnj .
Define
(M, τ ) =Yj
(A, τ ) =Mj
(Mnj , τj),
(Mnj , τj).
Let π : (M, τ ) → (M/A, τ ) denote the quotient map.
The remainder of this section is devoted to showing that normal self-τ elements
in (M/A, τ ) are close to normal self-τ elements with nice spectrum. Each lemma
gradually improves the niceness of the spectrum, until proposition 6.7 gives us what
we want.
Lemma 6.3. For any self-τ element a ∈ (M/A, τ ) there exists a self-τ unitary
u ∈ (M/A, τ ) such that a = up, where p = (a∗a)1/2.
Proof. Let x = (xj ) be any self-τ lift of a. Using Lemma 6.1 we can for all j ∈ N
find an invertible self-τ element yj ∈ (Mnj , τj) such that kxj − yjk < 1/j. Then the
sequence (yj) is in (M, τ ) and π(y) = π(x) = a. By Lemma 6.2 we can, for each
j ∈ N, find a self-τ unitary vj ∈ (Mnj , τj) such that yj = vjqj, where qj = (y∗j yj)1/2.
If we let v = (vj) and q = (qj) then y = vq and v is a self-τ unitary. Now put
u = π(v) and p = π(q). Then a = π(y) = π(v)π(q) = up, u is a self-τ unitary, and
p = π(q) = π((y∗y)1/2) = (π(y)∗π(y))1/2 = (a∗a)1/2.
(cid:3)
Lemma 6.4. If x ∈ (M/A, τ ) is normal and self-τ then for every ε > 0 there is a
normal self-τ invertible element y ∈ (M/A, τ ) such that kx − yk < ε.
22
TERRY A.LORING AND ADAM P. W.SØRENSEN
Proof. By Lemma 6.3 we can write x = up where u is a self-τ unitary and p =
(x∗x)1/2. Since we assumed x to be normal u and p commute by standard func-
tional calculus. Define y = u(p + (ε/2)I), where I is the unit in M/A. Since y
is the product of two commuting normal and invertible elements it is normal and
invertible. By Lemma 3.6 we have
From that it follows that y is self-τ . Finally we see that
pτ =(cid:0)(x∗x)τ(cid:1)1/2
=(cid:0)xx∗(cid:1)1/2
=(cid:0)x∗x(cid:1)1/2
= p.
kx − yk = kup − (up + (ε/2)u)k = k(ε/2)uk = ε/2 < ε.
(cid:3)
Lemma 6.5. Let λ ∈ C be given. If x ∈ (M/A, τ ) is normal and self-τ then for
every ε > 0 there is a normal self-τ element y ∈ (M/A, τ ) with λ /∈ σ(y), and such
that kx − yk < ε.
Proof. Let x = x − λI. Then x is normal and self-τ so by Lemma 6.4 we can find
a normal, self-τ and invertible y ∈ M/A such that ky − xk < ε. Let y = y + λI.
Then y is normal and self-τ , and
ky − xk = ky + λI − xk = ky − (x − λI)k = ky − xk < ε.
(cid:3)
We note that since 0 is not in the spectrum of y we have λ /∈ σ(y).
Lemma 6.6. Let F be an at most countable subset of C. If x ∈ (M/A, τ ) is normal
and self-τ then for every ε > 0 there is a normal self-τ element y ∈ (M/A, τ ) with
F ∩ σ(y) = ∅, and such that kx − yk < ε.
Proof. Let X be the set of normal and self-τ elements in (M/A, τ ). This is a closed
subset of M/A, so it is a complete metric space. Let F = {λ1, λ2, . . .}. For each
n ∈ N let Un be the set of self-τ normal elements in (M/A, τ ) that do not have
λn in their spectrum. By Lemma 6.5 all the Un are dense in X. Since the set
of invertible elements in a C∗-algebra is open all the Un are open in the relative
topology of X. By Baire's theorem the setTn Un is dense in X. That is, the set
of normal self-τ elements whose spectrum does not contain F is dense in the set of
(cid:3)
normal self-τ elements.
For any complex number z we denote by ℜ(z) and ℑ(z) the real and imaginary
parts of z. For all ε > 0 define
Γε = {z ∈ C ℜ(z) ∈ εZ or ℑ(z) ∈ εZ},
Σε = {z ∈ C ℜ(z) ∈ ε(Z +
1
2
) and ℑ(z) ∈ ε(Z +
1
2
)}.
Proposition 6.7. If x ∈ (M/A, τ ) is normal and self-τ then for every ε > 0 there
is a normal self-τ element y ∈ (M/A, τ ) with σ(y) ⊆ Γε, and such that kx− yk < ε.
Proof. By Lemma 6.6 we can find a normal and self-τ element y ∈ (M/A, τ ) with
σ(y) ∩ Σε = ∅,
and
ky − xk < 1 −
√2
2 ! ε.
√2
There is a continuous retraction f : C \ Σε → Γε with f (z) − z < (1 −
2 )ε for all
z. Let y = f (y). Then y is normal, has the right spectrum, and is no more than ε
away from x. By Lemma 3.6 we have
yτ = f (y)τ = f (yτ ) = f (y) = y.
(cid:3)
ALMOST COMMUTING SELF-ADJOINT MATRICES
23
6.2. The proof of Theorem 1. We begin by using our knowledge of semiprojec-
tive C∗,τ -algebras.
Proposition 6.8. Suppose (An, τn) is a sequence of C∗,τ -algebras. If x is a normal
self-τ element in
with spectrum contained in some finite one dimensional CW complex, then there is
(Q, τ ) =
(An, τn), ∞Mn=1
∞Yn=1
a lift of x to a normal self-τ element inQ∞n=1(An, τn).
(An, τn)
Proof. Let Γ be a finite graph such that σ(x) ⊆ Γ. By Lemma 3.6 the map f 7→ f (x)
is a C∗,τ -homomorphism from C(Γ, id) to (Q, τ ). Since C(Γ, id) is semiprojective,
we can find an m ∈ N and a normal self-τ element
such that y is a lift of x. Identifying
(An, τ ) ,
(An, τ )
y ∈
(An, τn), mMn=1
∞Yn=1
(An, τn), mMn=1
∞Yn=1
∞Yn=m
(An, τ ),
with
Q∞n=1(An, τn).
we see that if we pad y with leading zeros we get a self-τ and normal lift of x in
(cid:3)
We are now ready to prove real versions of Lin's theorem. First we do the case
of normal matrices.
Theorem 6.9. For every ε > 0 there is a δ > 0 such that for any n ∈ N, any
reflection τ on Mn and self-τ matrix X ∈ (Mn, τ ) with kXk ≤ 1 and
kX∗X − XX∗k < δ,
there exists a normal self-τ matrix X′ ∈ (Mn, τ ) with
kX − X′k < ε.
Proof. Suppose there was an ε that had no accompanying δ. Then there must exist
a sequence (nj) of natural numbers, reflections τj on Mnj , and self-τ contractive
matrices Xj ∈ (Mnj , τj) such that
kX∗j Xj − XjX∗j k → 0,
but every Xj is at least ε away from all normal self-τ matrices in (Mnj , τj).
Let, as in Section 6.1,
(M, τ ) =Yj
(Mnj , τj),
(Mnj , τj).
(A, τ ) =Mj
Let x = (Xj) and let y = π(x), where π is the quotient map from (M, τ ) to
(M/A, τ ). Then y is a normal and self-τ element. By Proposition 6.7 we can find
a normal self-τ element z ∈ (M/A, τ ) with spectrum contained in a finite graph
and ky − zk < ε/4. Using Proposition 6.8 we can find a normal self-τ element
24
TERRY A.LORING AND ADAM P. W.SØRENSEN
x′ ∈ (M, τ ) such that π(x′) = z. The definition of the norm in (M/A, τ ) tells us
that there exists (Aj ) = a ∈ A such that
k(x − x′) − ak = ky − zk + ε/4 < ε/2.
Now pick a j0 such that kAj0k < ε/2. Then we have
kXj0 − X′j0k ≤ k(Xj0 − X′j0 ) − Aj0k + kAj0k < k(x − x′) − ak + ε/2 < ε.
Which contradicts our assumption about all the Xj being at least ε away from any
(cid:3)
normal self-τ element.
Theorem 6.10. For every ε > 0 there is a δ > 0 such that for any n ∈ N, any
reflection τ on Mn and any pair A, B ∈ (Mn, τ ) of self-adjoint, self-τ matrices such
that kAk,kBk ≤ 1 and
kAB − BAk < δ,
there exists a commuting pair A′, B′ ∈ (Mn, τ ) of self-adjoint and self-τ matrices
with
kA − A′k + kB − B′k < ε.
Proof. Let ε > 0 be given. Use Theorem 6.9 to find a δ matching ε/2. Let A, B ∈
(Mn, τ ) be given as in the theorem. Define X = (A + iB)/2. Then
kXX∗ − X∗Xk =
1
2kAB − BAk < δ.
Hence we can find a normal self-τ matrix X′ ∈ Mn such that kX − X′k < ε/2.
Now let
A′ = X′ + X′∗
and B′ = −i(X′ − X′∗).
Then A′ and B′ are self-adjoint and self-τ . Since X′ is normal they commute. As
kA − A′k = k(X + X∗) − (X′ + X′∗)k ≤ kX − X′k + kX∗ − X′∗k < ε,
and likewise for kB − B′k, A′ and B′ show that we can approximate A and B by
commuting self-adjoint, self-τ matrices.
(cid:3)
Setting τ equal the transpose in Theorem 6.10 we obtain the extension of Lin's
theorem to real matrices. Using the dual operation, τ = ♯, we obtain the extension
of Lin's theorem to self-dual matrices.
7. TR rank one
The proof of Theorem 1 generalizes to real C∗-algebras beyond the matrices. In
particular, we get versions of Theorem 1 that apply to paths of almost commuting
matrices.
We will need a replacement for Lemma 6.1 regarding the approximation of self-
dual matrices by invertible self-dual matrices. We must introduce a concept along
the lines of stable rank and real rank.
ALMOST COMMUTING SELF-ADJOINT MATRICES
25
7.1. TR rank one. Recall that real rank zero [7] and stable rank one [40] are
defined in terms of density of the invertibles within a class of elements. For stable
rank one the class is all elements, while for real rank zero the class is restricted
by requiring a∗ = a. When (A, τ ) is a C∗,τ -algebra we can look at many more
symmetry restrictions, such as a∗ = aτ = a or a∗ = −aτ . Rank based on the
elements with a∗ = aτ was considered in [42].
Looking at Lemma 6.1 we see we need to consider invertibles within the class
of elements with aτ = a. To the resulting definition of rank we give a name that
reflects the role of time reversal symmetry in motivating this work.
Definition 7.1. A unital C∗,τ -algebra A is said to have TR rank one if the invert-
ible elements of A such that aτ = a are dense in the set of elements of A such that
aτ = a. A non-unital C∗,τ -algebra is said to have TR rank one when its unitization
has TR rank one. An R∗-algebra has TR rank one when its complexification has
TR rank one.
It is obvious that a direct sum of unital C∗,τ -algebras has TR rank one if and
only if each factor has TR rank one. Therefore there is no conflict in the first and
second sentences of the definition.
We give a few examples of TR rank one algebras.
Lemma 7.2. The C∗,τ -algebra C(S1) with f τ = f has TR rank one.
Proof. This is just the usual assertion that a function from the circle to the plane
(cid:3)
can be perturbed to miss the origin.
Lemma 7.3. The C∗,τ -algebra C(S1) ⊗ M2(C) with f τ (x) = (f (x))♯ has TR rank
one.
Proof. The only two-by-two matrices that are self-dual are the scalar matrices, so
(cid:3)
this follows from the previous lemma.
It is useful for us that TR rank one is preserved when adding matrices.
Proposition 7.4 (See [41, Proposition 3]). Let (A, τ ) be a C∗,τ -algebra with iden-
tity. If (A, τ ) has TR rank one then (Mn(A), τ ⊗ T ) has TR rank one.
Proof. As an induction hypothesis assume that Mn(A) has TR rank one (it is the
assumption of the proposition that the basis of the induction is true). Suppose we
have X = X τ⊗T in Mn+1(A). In block form
X =(cid:18) a
Bτ T D(cid:19) ,
B
where aτ = a in A and Dτ = D in Mn(D). Here the meaning of Bτ T is to apply τ
componentwise and then take matrix transpose. Let
0 = a0 being invertible and close to a. Since (a−1
0 we can find an
0 B that is invertible and with Eτ = E.
0 )τ = a−1
B
Bτ T D(cid:19) ,
X0 =(cid:18) a0
with aτ
E ∈ Mn(A) arbitrarily close to D − Bτ T a−1
Let D0 = E + Bτ T a−1
X1 =(cid:18) a0
0 B and notice that
B
Bτ T D0(cid:19)
26
TERRY A.LORING AND ADAM P. W.SØRENSEN
will be close to X and invariant under τ ⊗ T . We are done if we show that X1 is
invertible. To this end note that
are invertible. Since
0
0
and
0 B
(cid:18)1 −a−1
1 (cid:19)
X1(cid:18)1 −a−1
1 (cid:19)τ⊗T
(cid:18)a0
0 E(cid:19)
1 (cid:19) =(cid:18)a0
0 B
0 B
0
(cid:18)1 −a−1
0
0
0 E(cid:19) ,
we can write X1 as a product of invertibles. Hence it is invertible.
(cid:3)
Although we have no need for it here we expand slightly on how TR rank one
behaves with respect to matrices.
Lemma 7.5 (see [40, lemma 3.4]). Let (A, τ ) be a unital C∗,τ -algebra and let n ∈ N.
If (Mn+1(A), τ ⊗ T ) has TR rank one, then (A, τ ) has TR rank one.
Proof. Let a self-τ element t ∈ A and an 0 < ε < 1 be given. Define
X =(cid:18)t
0
0
In(cid:19) ∈ Mn+1(A),
where In is the identity element in Mn(A). By assumption we can find an invertible
Y ∈ Mn+1(A) within ε of X and such that Y = Y τ⊗T . In block form Y looks like
Y =(cid:18) a
Bτ T D(cid:19)
B
Notice that aτ = a and Dτ⊗T = D. Computing as in [40] we see that D and
t0 = a − BD−1Bτ T are invertible and that t0 is within ε(1 − ε)−1 of t. Hence
we are done if we can show that tτ
0 = t0. Since a is self-τ it suffices to show
that BD−1Bτ T is self-τ . For this write D−1 = (dij ) and B = (b1b2 . . . bn). The
symmetry on D−1 translates to dτ
ij = dji. We compute
(BD−1Bτ T )τ =
nXi,j=1
nXi,j
=
bidij bτ
τ
j
=
nXi,j
bjdτ
ij bτ
i
bjdjibτ
i = BD−1Bτ T .
(cid:3)
Combining the two previous results we get:
Theorem 7.6. Let (A, τ ) be a unital C∗,τ -algebra. The following are equivalent:
(1) (A, τ ) has TR rank one.
(2) (Mn(A), τ ⊗ T ) has TR rank one for all n ∈ N.
(3) (Mn(A), τ ⊗ T ) has TR rank one for some n ∈ N.
Corollary 7.7. Both C(S1, Mn(C)) with τ -operation
f τ (x) = (f (x))T
and C(S1, M2N (C)) with τ -operation
f τ (x) = (f (x))♯
have TR rank one.
ALMOST COMMUTING SELF-ADJOINT MATRICES
27
Remark 7.8. We could have added the τ -operation that exchanges two summands
that are each C(S1, Mn(C)). In terms of the underlying R∗-algebras, we have thus
the fact that
all have TR rank one.
C(cid:0)S1, Mn(R)(cid:1) , C(cid:0)S1, Mn(C)(cid:1) , C(cid:0)S1, Mn(H)(cid:1)
7.2. Generalizing the Friis-Rørdam Theorem. In the terms of this section,
Lemma 6.1 just says that (Mn(C), T ) and (M2n(C), ♯) have TR rank one. All of
section 6.1 was built on that lemma and the structured polar decomposition. Hence
one can redo the work starting with the assumption of TR rank 1 instead of Lemma
6.1. That work yields the following.
Proposition 7.9. Let (An, τn) be a sequence of unital TR rank one C∗,τ -algebras.
Define (M, τ ) = (Q An,Q τn) and A = L An. If x ∈ (M/A, τ ) is normal and
self-τ then for every ε > 0 there is a normal self-τ element y ∈ (M/A, τ ) with
σ(y) ⊆ Γε, and such that kx − yk < ε, where
Γε = {z ∈ C ℜ(z) ∈ εZ or ℑ(z) ∈ εZ}.
With that at hand, we get a much more general theorem.
Theorem 7.10. For all ε > 0 there exists a δ > 0 such that for all C∗,τ -algebras
of TR rank one the following holds: Whenever a, b are in A with a∗ = a = aτ ,
b∗ = b = bτ and kab − bak < δ there exists a′, b′ in A such that a′∗ = a′ = a′τ ,
b′∗ = b′ = b′τ , a′b′ = b′a′, and
ka − a′k,kb − b′k < ε.
Proof. Just as it was for matrices, it suffices to prove that a self-τ almost normal
element is close to a self-τ normal element.
We reduce to the case where A is unital. To do this, let an element x in a non-
unital C∗,τ -algebra B be given. Suppose we can find a normal and self-τ element
y ∈ B that is close to x, say ky − xk < η. Write y = α1 + z with α ∈ C and
z ∈ B. Now we notice that z must be normal and self-τ . Further we have that
ky − zk = α < η. Hence z is a normal and self-τ element that is at most 2η away
from x.
It remains to prove the theorem for A unital. The proof of this is just as the
proof for matrices given when we proved Theorem 6.9, only we should reference
(cid:3)
Proposition 7.9 instead of Proposition 6.7.
As an application, we get nice perturbation properties for paths of matrices.
Theorem 7.11. For all ε > 0 there exists a δ > 0 such that for all n ∈ N the
following holds: Whenever At, Bt are two closed paths of n-by-n, contractive, self-
adjoint, real (resp. self-dual) matrices such that kAtBt − BtAtk < δ for all t there
exists closed paths of n-by-n, self-adjoint, real (resp. self-dual) matrices A′t, B′t such
that A′tB′t = B′tA′t for all t and
for all t.
kAt − A′tk,kBt − B′tk < ε
Proof. Combine Corollary 7.7 with Theorem 7.10.
(cid:3)
Remark 7.12. Theorem 7.11 is also true for paths that need not be closed.
28
TERRY A.LORING AND ADAM P. W.SØRENSEN
References
[1] Alexander Altland and Martin R Zirnbauer, Nonstandard symmetry classes in mesoscopic
normal-superconducting hybrid structures, Physical Review B 55 (1997), no. 2, 1142.
[2] M. F. Atiyah, K-theory and reality, Quart. J. Math. Oxford Ser. (2) 17 (1966), 367–386.
MR 0206940 (34 #6756)
[3] Rajendra Bhatia, Spectral variation, normal matrices, and finsler geometry, Mathematical
Intelligencer 29 (2007), no. 3, 41–46.
[4] Bruce Blackadar, Shape theory for C ∗-algebras, Math. Scand. 56 (1985), no. 2, 249–275.
MR 813640 (87b:46074)
[5]
, Semiprojectivity in simple C ∗-algebras, Operator algebras and applications, Adv.
Stud. Pure Math., vol. 38, Math. Soc. Japan, Tokyo, 2004, pp. 1–17. MR 2059799
(2005g:46101)
[6] Jeffrey L. Boersema, Real C ∗-algebras, united K-theory, and the Kunneth formula, K-Theory
26 (2002), no. 4, 345–402. MR 1935138 (2004b:46072)
[7] Lawrence G. Brown and Gert K. Pedersen, C ∗-algebras of real rank zero, J. Funct. Anal. 99
(1991), no. 1, 131–149. MR 1120918 (92m:46086)
[8] J.F. Cardoso, Blind signal separation: statistical principles, Proceedings of the IEEE 86
(1998), no. 10, 2009–2025.
[9] Marius Dadarlat, Shape theory and asymptotic morphisms for C ∗-algebras, Duke Math. J.
73 (1994), no. 3, 687–711. MR 1262931 (95c:46117)
[10] Alessandro Derango, C*-algebras associated with homeomorphisms of
the unit circle,
ProQuest LLC, Ann Arbor, MI, 2000, Thesis (Ph.D.)–University of Toronto (Canada).
MR 2701546
[11] Freeman J Dyson, The threefold way. algebraic structure of symmetry groups and ensembles
in quantum mechanics, Journal of Mathematical Physics 3 (1962), no. 1199.
[12] E. G. Effros and J. Kaminker, Homotopy continuity and shape theory for C ∗-algebras, Geo-
metric methods in operator algebras (Kyoto, 1983), Pitman Res. Notes Math. Ser., vol. 123,
Longman Sci. Tech., Harlow, 1986, pp. 152–180. MR 866493 (88a:46082)
[13] Søren Eilers, Terry A. Loring, and Gert K. Pedersen, Morphisms of extensions of C ∗-algebras:
pushing forward the Busby invariant, Adv. Math. 147 (1999), no. 1, 74–109. MR 1725815
(2000j:46104)
[14] Davide Eynard, Klaus Glashoff, Michael M Bronstein, and Alexander M Bronstein,
laplacians, arXiv preprint
Multimodal diffusion geometry by joint diagonalization of
arXiv:1209.2295 (2012).
[15] Daniel S Freed and Gregory W Moore, Twisted equivariant matter, Annales Henri Poincar´e,
Springer, pp. 1–97.
[16] Peter Friis and Mikael Rørdam, Almost commuting self-adjoint matrices-a short proof of
Huaxin Lin's theorem, J. Reine Angew. Math. 479 (1996), 121–131. MR 1414391 (97i:46097)
as-
diagonalization,
and Michael M Bronstein, Matrix
[17] Klaus Glashoff
commutators:
approximate
their
ymptotic metric
www.cs.technion.ac.il/ mbron/publications.html, 2013.
properties
relation
and
to
joint
[18] K. R. Goodearl, Notes on real and complex C ∗-algebras, Shiva Mathematics Series, vol. 5,
Shiva Publishing Ltd., Nantwich, 1982. MR 677280 (85d:46079)
[19] F. Gygi, J.L. Fattebert, and E. Schwegler, Computation of maximally localized wannier
functions using a simultaneous diagonalization algorithm, Computer physics communications
155 (2003), no. 1, 1–6.
[20] M. B. Hastings, Topology and phases in fermionic systems, J. Stat. Mech. Theory Exp.
(2008), no. 1, L01001, 8. MR 2373523 (2009d:81434)
[21]
, Making almost commuting matrices commute, Comm. Math. Phys. 291 (2009),
no. 2, 321–345. MR 2530163 (2010m:15032)
[22] M.B. Hastings and T.A. Loring, Almost commuting matrices, localized wannier functions,
and the quantum hall effect, Journal of Mathematical Physics 51 (2010), 015214.
[23]
, Topological insulators and C ∗-algebras: Theory and numerical practice, Arxiv
preprint arXiv:1012.1019 (2010).
[24] G. G. Kasparov, Hilbert C ∗-modules: theorems of Stinespring and Voiculescu, J. Operator
Theory 4 (1980), no. 1, 133–150. MR 587371 (82b:46074)
ALMOST COMMUTING SELF-ADJOINT MATRICES
29
[25] Artiom Kovnatsky, Michael M Bronstein, Alexander M Bronstein, Klaus Glashoff, and Ron
Kimmel, Coupled quasi-harmonic bases, Computer Graphics Forum, vol. 32, Wiley Online
Library, 2013, pp. 439–448.
[26] Bingren Li, Real operator algebras, World Scientific Publishing Co. Inc., River Edge, NJ,
2003. MR 1995682 (2004k:46100)
[27] Huaxin Lin, Almost commuting selfadjoint matrices and applications, Operator algebras and
their applications (Waterloo, ON, 1994/1995), Fields Inst. Commun., vol. 13, Amer. Math.
Soc., Providence, RI, 1997, pp. 193–233. MR 1424963 (98c:46121)
[28] Netanel H Lindner, Gil Refael, and Victor Galitski, Floquet topological insulator in semicon-
ductor quantum wells, Nature Physics 7 (2011), no. 6, 490–495.
[29] T.A. Loring and M.B. Hastings, Disordered topological insulators via C ∗-algebras, EPL (Eu-
rophysics Letters) 92 (2010), 67004.
[30] Terry A. Loring, Computing a logarithm of a unitary matrix with general spectrum,
arXiv:1203.6151.
[31]
[32]
[33]
[34]
, Quantitative k-theory and spin chern numbers, arXiv:1302.0349.
, Stable relations. II. Corona semiprojectivity and dimension-drop C ∗-algebras, Pacific
J. Math. 172 (1996), no. 2, 461–475. MR 1386627 (97c:46070)
, Lifting solutions to perturbing problems in C ∗-algebras, Fields Institute Monographs,
vol. 8, American Mathematical Society, Providence, RI, 1997. MR 1420863 (98a:46090)
, Factorization of matrices of quaternions, Exposition. Math. 30 (2012), no. 3, 250–
267.
[35] Terry A. Loring and Adam P. W. Sørensen, Almost commuting unitary matrices related to
time reversal, Comm. Math. Phys. (to appear), arXiv:1107.4187.
[36] J. P. May, Stable algebraic topology, 1945–1966, History of topology, North-Holland, Ams-
terdam, 1999, pp. 665–723. MR 1721119 (2000m:55002)
[37] Yoshiko Ogata, Approximating macroscopic observables in quantum spin systems with com-
muting matrices, Journal of Functional Analysis (2013).
[38] T. W. Palmer, Real C ∗-algebras, Pacific J. Math. 35 (1970), 195–204. MR 0270162 (42
#5055)
[39] Xiao-Liang Qi and Shou-Cheng Zhang, The quantum spin hall effect and topological insula-
tors, Physics Today 63 (2010), no. 1, 33–38.
[40] Marc A. Rieffel, Dimension and stable rank in the K-theory of C ∗-algebras, Proc. London
Math. Soc. (3) 46 (1983), no. 2, 301–333. MR 693043 (84g:46085)
[41] A. Guyan Robertson, Stable range in C ∗-algebras, Math. Proc. Cambridge Philos. Soc. 87
(1980), no. 3, 413–418. MR 556921 (82h:46079)
[42] P. J. Stacey, Stability of involutory ∗-antiautomorphisms in UHF algebras, J. Operator The-
ory 24 (1990), no. 1, 57–74. MR 1086544 (92g:46077)
[43]
, Real structure in purely infinite C ∗-algebras, J. Operator Theory 49 (2003), no. 1,
77–84. MR 1978322 (2004c:46111)
[44] J. von Neumann, Beweis des ergodensatzes und des H-theorems in der neuen mechanik,
Zeitschrift f ur Physik A Hadrons and Nuclei 57 (1929), 30–70, 10.1007/BF01339852.
[45] Daniel Weigand, Entanglement entropy in 1d noninteracting fermionic systems), Master's
thesis, 2012, http://www.physik.rwth-aachen.de/institute/institut-fuer-quanteninformation/student-projects/past-theses/.
[46] A. Zabrodin, Matrix models and growth processes: from viscous flows to the quantum Hall
effect, Applications of random matrices in physics, NATO Sci. Ser. II Math. Phys. Chem.,
vol. 221, Springer, Dordrecht, 2006, pp. 261–318. MR 2232116 (2007d:82081)
University of New Mexico, Department of Mathematics and Statistics & the Center
for Advanced Research Computing, Albuquerque, New Mexico, 87131, USA
E-mail address: [email protected]
Department of Mathematical Sciences, University of Copenhagen, Universitetspar-
ken 5, DK-2100, Copenhagen Ø, Denmark
Current address: University of New Mexico, Department of Mathematics and Statistics, Al-
buquerque, New Mexico, 87131, USA
E-mail address: [email protected]
|
1403.3799 | 2 | 1403 | 2015-05-19T16:29:43 | $K$-theory and homotopies of 2-cocycles on higher-rank graphs | [
"math.OA",
"math.KT"
] | This paper continues our investigation into the question of when a homotopy $\omega = \{\omega_t\}_{t \in [0,1]}$ of 2-cocycles on a locally compact Hausdorff groupoid $\mathcal{G}$ gives rise to an isomorphism of the $K$-theory groups of the twisted groupoid $C^*$-algebras: $K_*(C^*(\mathcal{G}, \omega_0)) \cong K_*(C^*(\mathcal{G}, \omega_1)).$ In particular, we build on work by Kumjian, Pask, and Sims to show that if $\mathcal{G} = \mathcal{G}_\Lambda$ is the infinite path groupoid associated to a row-finite higher-rank graph $\Lambda$ with no sources, and $\{c_t\}_{t \in [0,1]}$ is a homotopy of 2-cocycles on $\Lambda$, then $K_*(C^*(\mathcal{G}_\Lambda, \sigma_{c_0})) \cong K_*(C^*(\mathcal{G}_\Lambda, \sigma_{c_1})),$ where $\sigma_{c_t}$ denotes the 2-cocycle on $\mathcal{G}_\Lambda$ associated to the 2-cocycle $c_t$ on $\Lambda$. We also prove a technical result (Theorem 3.3), namely that a homotopy of 2-cocycles on a locally compact Hausdorff groupoid $\mathcal{G}$ gives rise to an upper semi-continuous $C^*$-bundle. | math.OA | math |
K-theory and Homotopies of 2-Cocycles on
Higher-Rank Graphs
Elizabeth Gillaspy∗
October 14, 2018
Abstract
This paper continues our investigation into the question of when a ho-
motopy of 2-cocycles on a locally compact Hausdorff groupoid gives rise
to an isomorphism of the K-theory groups of the twisted groupoid C ∗-
algebras. Our main result, which builds on work by Kumjian, Pask, and
Sims, shows that a homotopy of 2-cocycles on a row-finite higher-rank
graph Λ gives rise to twisted groupoid C ∗-algebras with isomorphic K-
theory groups. (Here, the groupoid in question is the path groupoid of Λ.)
We also establish a technical result: any homotopy of 2-cocycles on a lo-
cally compact Hausdorff groupoid G gives rise to an upper semicontinuous
bundle of C ∗-algebras.
Keywords: Higher-rank graph, twisted groupoid C ∗-algebra, K-theory,
twisted k-graph C ∗-algebra, upper semicontinuous C ∗-bundle, C0(X)-
algebra, groupoid, 2-cocycle.
MSC (2010): 46L05, 46L80.
1 Introduction
Higher-rank graphs, or k-graphs, provide a k-dimensional analogue of directed
graphs. They were introduced by Kumjian and Pask in [9] to provide a combi-
natorial model for the higher-rank Cuntz-Krieger algebras studied by Robertson
and Steger in [19]. Much of the interest in the C∗-algebras C∗(Λ) associated to
k-graphs Λ stems from the multiple ways one can model C∗(Λ) -- the k-graph
Λ reflects many of the properties of C∗(Λ), but we can also describe C∗(Λ) as
a universal C∗-algebra for certain generators and relations, or as a groupoid
C∗-algebra C∗(Λ) ∼= C∗(GΛ).
The class of groupoids includes groups, group actions, equivalence relations,
and group bundles. Renault initiated the study of groupoid C∗-algebras in
[18], and the theory and applications of groupoid C∗-algebras have since been
developed by many researchers. Given a 2-cocycle ω ∈ Z 2(G, T) on a groupoid
∗Department of Mathematics, University of Colorado - Boulder, Campus Box 395, Boulder,
CO 80309-0395, [email protected]
1
G, Renault also explains in [18] how to construct the twisted groupoid C∗-
algebra C∗(G, ω). These objects have received relatively little attention until
quite recently, but it has now become clear that twisted groupoid C∗-algebras
can help answer many questions about the structure of untwisted groupoid C∗-
algebras (cf. [14, 13, 2, 7, 1]), as well as classifying those C∗-algebras which
admit diagonal subalgebras (also known as Cartan subalgebras) -- cf. [8]. In
another direction, [20] establishes a connection between the K-theory of twisted
groupoid C∗-algebras and the classification of D-brane charges in string theory.
Two recent papers have explored the effect of a homotopy {ωt}t∈[0,1] of 2-
cocycles on the K-theory of the twisted groupoid C∗-algebras. First, Echterhoff,
Luck, Phillips, and Walters showed in Theorem 1.9 of [3] that if G is a group that
satisfies the Baum-Connes conjecture with respect to the coefficient algebras K
and C([0, 1], K), and {ωt}t∈[0,1] is a homotopy of 2-cocycles on G, then the K-
theory groups of the reduced twisted group C∗-algebras are unperturbed by the
homotopy:
K∗(C∗
r (G, ω0)) ∼= K∗(C∗
r (G, ω1)).
(1)
In particular, taking G = Z2, we obtain another proof of the fact, established
in 1980 by Pimsner and Voiculescu in [15], that all of the rotation algebras
{Aθ}θ∈[0,1] have isomorphic K-theory groups.
Kumjian, Pask, and Sims also studied the effect of a homotopy of 2-cocycles
on K-theory in [10]. Theorem 5.4 of [10] establishes that if Λ is a row-finite
source-free k-graph and c is a 2-cocycle on Λ such that c(λ, µ) = e2πiσ(λ,µ) for
some R-valued 2-cocycle σ, then K∗(C∗(Λ)) ∼= K∗(C∗(Λ, c)). Defining ct(λ, µ) =
e2πitσ(λ,µ) for t ∈ [0, 1] gives us a homotopy of 2-cocycles linking c and the trivial
2-cocycle. Moreover, Corollary 7.8 of [11] tells us that C∗(Λ, c) is isomorphic to
a twisted groupoid C∗-algebra C∗(GΛ, ωc). Thus, we can view [10] Theorem 5.4
as a result about homotopic 2-cocycles on groupoids.
Inspired by the above-mentioned results, we have begun exploring the ques-
tion of when a homotopy of 2-cocycles on a locally compact Hausdorff groupoid
G induces an isomorphism of the K-theory groups of the (full or reduced)
twisted groupoid C∗-algebras. In a previous article [5], we extended the above-
mentioned Theorem 1.9 of the paper [3] by Echterhoff et al. to the case when
G = G ⋉ X is a transformation group, where X is locally compact Hausdorff
and G satisfies the Baum-Connes conjecture with coefficients.
In this article, we prove the following generalization of [10] Theorem 5.4:
Theorem 4.1. Let Λ be a row-finite k-graph with no sources and let {ct}t∈[0,1]
be a homotopy of 2-cocycles in Z2(Λ, T). Then {ct}t∈[0,1] gives rise to a homo-
topy {σct}t∈[0,1] of 2-cocycles on GΛ such that
K∗(C∗(GΛ, σc0)) ∼= K∗(C∗(GΛ, σc1 )).
As of this writing, we are unaware of any examples of groupoids G and ho-
motopies ω = {ωt}t∈[0,1] of 2-cocycles on G where the homotopy does not induce
an isomorphism of the K-theory groups of the twisted groupid C∗-algebras.
2
1.1 Outline
This paper begins by recalling the definitions of a higher-rank graph and of a
groupoid in Section 2, as well as the definition of a 2-cocycle in each category,
and sketching the procedure by which we can construct a C∗-algebra from these
objects.
In Section 3 we define a homotopy of 2-cocycles on a k-graph and
on a groupoid, and show that the definitions are compatible. We also prove a
technical result (Theorem 3.3), namely, that a homotopy {ωt}t∈[0,1] of 2-cocycles
on a groupoid G gives rise to a C([0, 1])-algebra with fiber algebra C∗(G, ωt) at
t ∈ [0, 1]. We expect that this result will prove useful in future work, as we
search for more classes of groupoids where a homotopy of 2-cocycles induces an
isomorphism of the K-theory groups of the twisted groupoid C∗-algebras.
In Section 4 we begin the proof of Theorem 4.1. Our proof technique consists
of proving a stronger version of Theorem 4.1 in a simple case, and then showing
how to use this simple case to obtain our desired result for general k-graphs.
To be precise, Proposition 4.2 shows that when the degree map d on Λ satisfies
d(λ) = b(s(λ)) − b(r(λ)) for all λ ∈ Λ, the C([0, 1])-algebra associated to a
homotopy of 2-cocycles on Λ is actually a trivial continuous field. Section 4.2
shows how to exploit the triviality of the continuous field in this special case
to see that a homotopy of 2-cocycles on any row-finite, source-free k-graph Λ
induces an isomorphism K∗(C∗(Λ, c0)) ∼= K∗(C∗(Λ, c1)). The argument in this
section closely parallels Section 5 of [10].
Acknowledgments: This research constitutes part of my PhD thesis, and
I would like to thank my advisor, Erik van Erp, for his encouragement and
assistance. I would also especially like to thank Prof. Jody Trout for his help
and support throughout my PhD work.
2 Groupoids and k-graphs
Definition 2.1 (cf. [9] Definition 1.1). A higher-rank graph of degree k, or a
k-graph, is a nonempty countable small category Λ equipped with a functor
d : Λ → Nk (the degree map) satisfying the following factorization property:
Given a morphism λ ∈ Λ with d(λ) = m + n, there exist unique µ, ν ∈ Λ such
that λ = µν and d(µ) = m, d(ν) = n.
The simplest example of a k-graph is Nk, equipped with the identity mor-
phism id : Nk → Nk.
In this article we will use the arrows-only picture of category theory, so that
we think of the objects of a category Λ as identity morphisms. Hence, λ ∈ Λ
means that λ is a morphism in Λ. Given an element λ in a category Λ, write
s(λ) for the domain, or source, of the morphism λ, and r(λ) for its target, or
range. We say k-graph Λ is row-finite if, for any v ∈ Obj(Λ) and any n ∈ Nk,
the set
vΛn := {λ ∈ Λ : r(λ) = v, d(λ) = n}
is finite. We say Λ has no sources if vΛn 6= ∅ for every v ∈ Obj(Λ) and every
n ∈ Nk. We will only consider k-graphs which are row-finite and have no sources,
3
since these are the k-graphs which we can study via the groupoid method that
was introduced in [9] and which we will explain in Section 2.1.
Definition 2.2 (cf. [18] Definition I.1.12, [11] Section 3). For a category Λ, let
Λ∗2 = {(λ1, λ2) ∈ Λ × Λ : s(λ1) = r(λ2)}. A function c : Λ∗2 → T is called a
2-cocycle on Λ if
c(λ, µν)c(µ, ν) = c(λµ, ν)c(λ, µ)
(2)
whenever (λ, µ), (µ, ν) ∈ Λ∗2, and c(λ, s(λ)) = c(r(λ), λ) = 1 ∀ λ ∈ Λ. We write
Z2(Λ, T) for the set of 2-cocycles on Λ.
If c, c are two 2-cocycles on Λ, we say that c, c are cohomologous if there
exists a function b : Λ → T such that
c(µ, ν) := b(µ)b(ν)b(µν)−1c(µ, ν) = δb(µ, ν)c(µ, ν) ∀ (µ, ν) ∈ Λ∗2.
We note that cohomologous 2-cocycles give rise to isomorphic twisted C∗-
algebras (cf. [11] Proposition 5.6, [18] Proposition II.1.2).
The only cocycles we will consider in this paper are 2-cocycles, so we will
occasionally drop the 2 and refer to them simply as cocycles.
Definition 2.3 ([11] Definition 5.2). The twisted higher-rank-graph algebra
C∗(Λ, c) associated to a k-graph Λ and a 2-cocycle c on Λ is the universal
C∗-algebra generated by a collection {sλ}λ∈Λ of partial isometries satisfying
the following twisted Cuntz-Krieger relations:
(CK1) {sv}v∈Obj(Λ) is a collection of mutually orthogonal projections;
(CK2) sµsν = c(µ, ν)sµν whenever s(µ) = r(ν);
(CK3) s∗
µsµ = ss(µ) for all µ ∈ Λ;
(CK4) sv =Pµ∈vΛn sµs∗
µ for all v ∈ Obj(Λ) and all n ∈ Nk.
Note that every k-graph Λ admits at least one 2-cocycle: the trivial cocycle,
obtained by setting c(λ, µ) = 1 for all (λ, µ) ∈ Λ∗2. In this case, the definition
above of C∗(Λ, c) agrees with that of C∗(Λ) given in [9] Definitions 1.5. For
example, if Λ = Nk and c is the trivial cocycle, then C∗(Λ, c) ∼= C(Tk). More
generally, if Λ = N2, let cθ : Λ∗2 → T be given by cθ((m, n), (j, k)) = e2πiθnj.
Then cθ is a 2-cocycle on Λ and C∗(Λ, cθ) is isomorphic to the rotation algebra
Aθ.
2.1 Groupoids
In this section, we review the construction of a twisted groupoid C∗-algebra
set forth in [18], as well as the procedure given in the seminal article [9] for
associating a groupoid to a k-graph. Theorem 3.3 in Section 3 applies to arbi-
trary locally compact Hausdorff groupoids, so we present in full generality all
the definitions necessary for the construction of a twisted groupoid C∗-algebra.
4
A groupoid G is a small category with inverses. We use the notation of [18]
to denote groupoid elements and operations; for example, G(2) ⊆ G × G denotes
the set of composable pairs and G(0) denotes the unit space. If u ∈ G(0), write
Gu = {x ∈ G : s(x) = u}
Gu = {x ∈ G : r(x) = u}.
In this article, we restrict our attention to groupoids which admit a locally
compact Hausdorff topology in which the operations of composition (or multi-
plication) and inversion are continuous.
In addition to the groupoids associated to k-graphs, examples of groupoids
include groups, vector bundles, and transformation groups. For more details
and examples, see [6, 12].
Given a row-finite, source-free k-graph Λ, Section 2 of [9] describes how to
form the associated path groupoid GΛ:
Definition 2.4 ([9] Examples 1.7(ii)). We define the k-graph Ωk to be the
category with Obj(Ωk) = Nk, and morphisms Ωk = {(m, n) ∈ Nk ×Nk : n ≥ m}.
We have r(m, n) = m, s(m, n) = n, d(m, n) = n−m. Composition in Ωk is given
by (m, n)(n, ℓ) = (m, ℓ).
For a k-graph Λ, let Λ∞ denote the set of degree-preserving functors x :
Ωk → Λ. When k = 1, the elements x ∈ Λ∞ are the infinite paths in Λ.
Given p ∈ Nk, define σp : Λ∞ → Λ∞ by σp(x)(m, n) = x(m+p, n+p). When
Λ is row-finite and source-free, Proposition 2.3 in [9] shows that if λ ∈ Λ, x ∈ Λ∞
satisfy s(λ) = x(0), there is a unique y ∈ Λ∞ such that σd(λ)(y) = x; we often
write y = λx.
Definition 2.5 ([9] Definition 2.1). Given a row-finite, source-free k-graph Λ,
the associated path groupoid GΛ is the groupoid associated to the equivalence
relation on Λ∞ of "shift equivalence with lag." In other words,
GΛ := {(x, n − m, y) ∈ Λ∞ × Zk × Λ∞ : n, m ∈ Nk, σn(x) = σm(y)},
Λ = Λ∞, with r(x, ℓ, y) = x, s(x, ℓ, y) = y, and multiplication and inver-
and G(0)
sion in GΛ given by (x, ℓ, y)(y, m, z) = (x, ℓ + m, z); (x, ℓ, y)−1 = (y, −ℓ, x).
When Λ is a row-finite, source-free k-graph, Proposition 2.8 in [9] tells us
that the sets
Z(µ, ν) := {(µx, d(µ) − d(ν), νx) : x(0) = s(µ) = s(ν)}
form a basis of compact open sets for a locally compact Hausdorff topology on
GΛ (in fact, with this topology GΛ is an ample ´etale groupoid).
To build a C∗-algebra out of a groupoid G we will start by putting a ∗-
algebra structure on Cc(G), and to do this we will need to integrate over the
groupoid G. A Haar system {λu}u∈G(0) (the groupoid analogue of Haar measure
for groups, cf. Definition I.2.2 in [18]) will allow us to do this. Unlike in the
group case, one cannot make existence or uniqueness statements about Haar
5
systems for groupoids, so one usually starts by hypothesizing the existence of a
fixed Haar system. For example, we obtain a Haar system {λx}x∈Λ∞ on GΛ by
setting
λx(E) = #{e ∈ E : e = (x, n, y) for some n ∈ Zk, y ∈ Λ∞}.
We will always use this Haar system on GΛ in this paper.
Definition 2.6. Let G be a locally compact Hausdorff groupoid equipped with
a Haar system {λu}u∈G(0) and a continuous 2-cocycle ω. We define a ∗-algebra
structure on Cc(G) as follows: for f, g ∈ Cc(G) let
f ∗ω g(a) =ZGs(a)
f (ab)g(b−1)ω(ab, b−1)dλs(a)(b),
f ∗(a) = f (a−1)ω(a, a−1).
One can check (cf.
[18] Proposition II.1.1) that the multiplication is well
defined (that is, that f ∗ω g ∈ Cc(G) as claimed) and associative, and that
(f ∗)∗ = f so that the involution is involutive. The proof of associativity relies
on the cocycle condition (2).
Given the fundamental role that the cocycle ω plays in the multiplication
and involution on Cc(G), we will often write Cc(G, ω) to denote the set Cc(G)
equipped with the ∗-algebra structure of Definition 2.6. We define C∗(G, ω) to
be the completion of Cc(G, ω) in the maximal C∗-norm, as described in Chapter
II of [18].
Definition 2.7. When G = GΛ is the groupoid associated to a row-finite k-graph
Λ with no sources, Lemma 6.3 of [11] explains how, given a cocycle c ∈ Z2(Λ, T),
we can construct a cocycle σc ∈ Z 2(GΛ, T). Then Corollary 7.8 of [11] shows
that C∗(Λ, c) ∼= C∗(GΛ, σc). The construction of σc is rather technical, but since
we will need the details later, we present it here.
Lemma 6.6 of [11] establishes the existence of a subset
P ⊆ {Z(µ, ν) : s(µ) = s(ν)}
that partitions GΛ. In other words, every a ∈ GΛ has exactly one representation
of the form a = (µax, d(µa) − d(νa), νax) with Z(µa, νa) ∈ P. Note that if
(a, b) ∈ G(2)
Λ , we need not have µa = µab or νb = νab. However, given (a, b) ∈
G(2)
Λ , Lemma 6.3(i) of [11] shows that we can always find y ∈ Λ∞ and α, β, γ ∈ Λ
such that
a = (µaαy, d(µa) − d(νa), νaαy)
b = (µbβy, d(µb) − d(νb), νbβy)
ab = (µabγy, d(µab) − d(νab), νabγy).
Then, given a 2-cocycle c on Λ, we define a 2-cocycle σc on GΛ by
σc(a, b) = c(µa, α)c(µb, β)c(νab, γ)c(νa, α)c(νb, β)c(µab, γ).
6
Since c satisfies the cocycle condition (2), it's straightforward to check that σc
does also. Lemma 6.3 of [11] checks that σc is well-defined and continuous, so we
can construct the groupoid C∗-algebra C∗(GΛ, σc) as outlined above. Corollary
7.8 of [11] tells us that C∗(GΛ, σc) ∼= C∗(Λ, c).
Theorem 6.5 of [11] establishes that different choices of partitions P will
give rise to cohomologous groupoid cocycles, and hence to isomorphic twisted
groupoid C∗-algebras.
3 Homotopies of Cocycles
In order to define a homotopy of groupoid 2-cocycles, we begin by observing
that, given any locally compact Hausdorff groupoid G, we can make G × [0, 1]
into a LCH groupoid by equipping it with the product topology and setting
(G × [0, 1])(2) := G(2) × [0, 1]. In other words, (G × [0, 1])(0) = G(0) × [0, 1], and
r(γ, t) = (r(γ), t), s(γ, t) = (s(γ), t).
Moreover, if G has a Haar system {λu}u∈G(0) , then setting λu,t := λu for every
t ∈ [0, 1] gives rise to a Haar system on G × [0, 1]. We will always use this Haar
system on G × [0, 1] in this paper.
Definition 3.1 ([5] Definition 2.12). A homotopy of (2-)cocycles on a locallpy
compact Hausdorff groupoid G is a 2-cocycle ω ∈ Z 2(G × [0, 1], T). We say
that two cocycles ω0, ω1 ∈ Z 2(G, T) are homotopic if there exists a homotopy
ω ∈ Z 2(G × [0, 1], T) such that ωi = ωG×{i} for i = 0, 1.
If ω is a homotopy of cocycles on G linking ω0, ω1, Theorem 3.3 below tells
us that C∗(G, ω0) and C∗(G, ω1) are quotients of C∗(G × [0, 1], ω). This will be
fundamental to the proof of our main result, Theorem 4.1.
Definition 3.2 (cf. [21] Definition C.1). Let X be a locally compact Hausdorff
space. A C∗-algebra A is a C0(X)-algebra if we have a ∗-homomorphism Φ :
C0(X) → ZM (A) such that
A = span{Φ(f )a : f ∈ C0(X), a ∈ A}.
We usually write f · a for Φ(f )a.
If A is a C0(X)-algebra, then for any x ∈ X, spanC0(X\x) · A is an ideal
Ix. We call Ax := A/Ix the fiber of A at x ∈ X.
Theorem 3.3. Let ω be a homotopy of cocycles on a locally compact Hausdorff
groupoid G with Haar system {λu}u∈G(0) . Then C∗(G × [0, 1], ω) is a C([0, 1])-
algebra, with fiber C∗(G, ωt) at t ∈ [0, 1].
Proof. We begin by checking that C∗(G × [0, 1], ω) is a C([0, 1])-algebra. For
f ∈ C([0, 1]), φ ∈ Cc(G × [0, 1], ω), define
f · φ(a, t) = f (t)φ(a, t).
7
It's not difficult to check that this action extends to a ∗-homomorphism
Φ : C([0, 1]) → ZM (C∗(G × [0, 1], ω))
such that Φ(f )φ ≤ f ∞φ, or that
Φ(C([0, 1])) · Cc(G × [0, 1], ω) = Cc(G × [0, 1], ω)
is dense in C∗(G × [0, 1], ω). In other words, Φ makes C∗(G × [0, 1], ω) into a
C([0, 1])-algebra as claimed.
Fix t ∈ [0, 1] and denote by qt : Cc(G × [0, 1], ω) → Cc(G, ωt) the evaluation
map. Then qt is bounded by the I-norm (cf. [18] Section II.1), and hence
extends to a surjective ∗-homomorphism qt : C∗(G × [0, 1], ω) → C∗(G, ωt). In
other words, C∗(G, ωt) is a quotient of C∗(G × [0, 1], ω). To see that C∗(G, ωt) ∼=
C∗(G × [0, 1], ω)t, we need to check that ker qt = It. A standard approximation
argument will show that ker qt ⊇ It: thus, we will only detail the proof that
ker qt ⊆ It.
Note that the fiber algebra C∗(G × [0, 1], ω)t ∼= C∗(G × [0, 1], ω)/It can be
calculated as a completion Cc(G × [0, 1], ω) with respect to the norm given by
kf kt := sup{kL(f )k : L(It) = 0, L is an I-norm-bounded representation}.
Thus, to show that ker qt ⊆ It, we will show that each such representation L
factors through qt.
Given such a representation L : Cc(G × [0, 1], ω) → B(H), define L′
:
Cc(G, ωt) → B(H) by L′(qt(f )) := L(f ). We claim that L′ is an I-norm-bounded
representation of Cc(G, ωt). To see this, it suffices to check that L′ is well-defined
and bounded.
Lemma 3.4. If f, g ∈ Cc(G × [0, 1], ω) satisfy qt(f ) = qt(g), then the function
h = f − g ∈ Cc(G × [0, 1], ω) lies in It. Consequently, L(f ) = L(g) and L′ is
well defined on Cc(G, ωt).
Proof. Let {fi}i∈I be an approximate unit for C0([0, 1]\t) such that fi(s) ր 1
for every s 6= t, and moreover that for each i there exists δi > 0 such that
fi(s) = 1 if s − t ≥ δi. We will show that the I-norm kh − fihkI → 0.
Consequently, h = limi fih in C∗(G × [0, 1], ω), so h ∈ It.
For any k ∈ Cc(G × [0, 1], ω), the axioms of a Haar system tell us that the
function (u, t) 7→ R k(a, t)dλu,t(a) is in C0(G(0) × [0, 1]). In particular, if we
take k to be a function that equals 1 where h is nonzero, and vanishes rapidly
off supp h, this shows us that φ(u, t) := λu,t(supp h) is a pointwise limit of
functions in C0(G(0) × [0, 1]), and hence is bounded. Let K = max φ.
Let ǫ > 0 be given. Since h is compactly supported and h(a, t) = 0 ∀ a ∈ G,
we can choose δ > 0 such that h(a, s) < ǫ/K ∀ a ∈ G whenever s − t < δ, and
choose j such that i ≥ j means δi < δ. Then, if s − t < δ,
ZGu
h(a, s) − fi(s)h(a, s) dλu(a) = (1 − fi(s))ZGu
h(a, s) dλu(a) < 1 · ǫ.
8
On the other hand, if s − t ≥ δ > δi, then fi(s) = 1 and
h(a, s) − fi(s)h(a, s) dλu(a) = 0
ZGu
for any u ∈ G(0). In either case, given any ǫ > 0 we can always choose j such
that i ≥ j implies
kh − fihkI = max( sup
s∈[0,1]
sup
s∈[0,1]
< ǫ.
sup
u∈G(0)ZGu
u∈G(0)ZGu
sup
h(a, s) − fi(s)h(a, s)dλu(a),
h(a−1, s) − fi(s)h(a−1, s)dλu(a))
Since kh − fihk ≤ kh − fihkI and It is closed, it follows that h ∈ It as desired,
and so L(h) = 0.
Having seen that L′ is well defined, we now proceed to show that it is
bounded.
Lemma 3.5. For any fixed f ∈ Cc(G × [0, 1], ω), the map s 7→ kqs(f )kI is
continuous.
Proof. Fix f ∈ Cc(G × [0, 1], ω) and fix t ∈ [0, 1]. As in the proof of Lemma 3.4,
let K denote the supremum of the function (u, s) 7→ λu,s(supp f ). Since f has
compact support, given ǫ > 0 we can choose δ such that
s − t < δ ⇒ f (a, t) − f (a, s) <
ǫ
2K
∀ a ∈ G.
Now, by definition of the I-norm, there exists u ∈ G(0) such that either
kqs(f )kI <ZGu
f (a, s) dλu(a) +
ǫ
2
, or kqs(f )kI <ZGu
f (a−1, s) dλu(a) +
ǫ
2
.
It follows that either
kqs(f )kI <ZGu
or kqs(f )kI <ZGu
Thus,
f (a, t) +
ǫ
2K
dλu(a) +
ǫ
2
f (a−1, t) +
ǫ
2K
dλu(a) +
≤ZGu
≤ZGu
ǫ
2
f (a, t) dλu(a) + ǫ,
f (a−1, t) dλu(a) + ǫ.
kqs(f )kI < max(cid:26)ZGu
≤ max( sup
u∈G(0)ZGu
f (a, t) dλu(a),ZGu
f (a−1, t) dλu(a)(cid:27) + ǫ
u∈G(0)ZGu
f (a−1, t) dλu(a)) + ǫ
f (a, t) dλu(a), sup
= kqt(f )kI + ǫ
9
if s − t < δ. Reversing the roles of s and t in the above argument tells us that
s − t < δ ⇒ kqs(f )kI − kqt(f )kI < ǫ,
as desired.
Now we can complete the proof of Theorem 3.3. Set St = {ψ ∈ C ([0, 1]) :
ψ(t) = 1}; for any ψ ∈ St and any f ∈ Cc(G × [0, 1], ω), we have
L(ψ · f ) = L′(qt(ψ · f )) = L′(qt(f )).
Consequently,
L′(qt(f )) = inf
ψ∈S
L(ψ · f ) ≤ inf
ψ
ψ · f I
= inf
ψ
max( sup
s∈[0,1]
sup
u∈G(0)Z ψ(s)f (a, s) dλu(a),
sup
s∈[0,1]
sup
u∈G(0)Z ψ(s)f (a−1, s)dλu(a))
= inf
ψ
sup
s∈[0,1]
qs(ψ · f )I .
Let ǫ > 0 be given. Choose δ such that s−t < δ ⇒ kqs(f )kI − kqt(f )kI <
ǫ; choose ψǫ ∈ C([0, 1]) such that ψǫ(t) = 1 and s − t ≥ δ ⇒ ψǫ(s) = 0. Then,
since ψǫ ∈ St,
qs(ψǫ · f )I = ψǫ(s)qs(f )I < ψǫ(s) (qt(f )I + ǫ) ≤ qt(f )I + ǫ
(3)
if s − t < δ; otherwise we have qs(ψǫ · f )I = 0, and (3) still holds.
Since we can find such a ψǫ for any ǫ > 0, it follows that
L′(qt(f )) ≤ inf
ψ∈St
sup
s∈[0,1]
qs(ψ · f )I ≤ inf
ǫ
sup
qs(ψǫ · f )I
s
≤ inf
ǫ
qt(f )I + ǫ
= qt(f )I .
The fact that qt is onto now tells us that L′ is a bounded representation of
Cc(G, ωt) as claimed. In other words, every representation L of Cc(G × [0, 1], ω)
that kills It also factors through qt, so ker qt ⊆ It. This completes the proof that
the fiber algebra C∗(G × [0, 1], ω)/It of the C([0, 1])-algebra C∗(G × [0, 1], ω) is
simply C∗(G, ωt).
In order to apply Theorem 3.3 to a homotopy of cocycles on a k-graph, we
first need to define such a homotopy. Unlike for groupoids, there is no obvious
way to make Λ × [0, 1] into a higher-rank graph, so our definition of a homotopy
of k-graph cocycles will look rather different than Definition 3.1 above. However,
Proposition 3.8 below will show that the two definitions are compatible.
10
Definition 3.6. Let Λ be a k-graph. A family {ct}t∈[0,1] of 2-cocycles in
Z2(Λ, T) is a homotopy of (2-)cocycles on Λ if for each pair (λ, µ) ∈ Λ∗2 the
function t 7→ ct(λ, µ) ∈ T is continuous.
Definition 3.7. Let {ct}t∈[0,1] be a homotopy of cocycles on a k-graph Λ.
Define ω ∈ Z 2(GΛ × [0, 1], T) by
ω ((a, t), (b, t)) = σct (a, b),
where σct is the cocycle on GΛ associated to ct as in Definition 2.7.
A moment's thought will reveal that ω satisfies the cocycle condition (2),
since each σct is a cocycle. Thus, in order to see that ω is a homotopy of cocycles
on GΛ, we merely need to check that ω : (GΛ × [0, 1])(2) → T is continuous.
Proposition 3.8. The cocycle ω described in Definition 3.7 is continuous, and
hence is a homotopy of groupoid cocycles on GΛ.
Proof. We will show that if {(ai, bi, ti)}i∈I ⊆ G(2)
to (a, b, t), then
Λ ×[0, 1] is a net which converges
ω ((ai, ti), (bi, ti)) := σcti (ai, bi) = σcti (a, b)
(4)
for large enough i. Recall from Definition 2.7 that σcti (a, b) is a finite product
of terms of the form cti(µ, ν) and their inverses, where the elements µ, ν depend
only on the elements a, b and on the choice of partition P of GΛ -- but not on the
2-cocycle cti. Thus, Equation (4), and the continuity of the maps t 7→ ct(µ, ν),
will imply that ω ((ai, ti), (bi, ti)) → σct (a, b) = ω ((a, t), (b, t)) .
In what follows, we will use the notation of Definition 2.7. If (ai, bi, ti) →
(a, b, t), then for large enough i we have ai ∈ Z(µa, νa), bi ∈ Z(µb, νb), and
aibi ∈ Z(µab, νab) as well. In other words, we can write
a = (µaαy, d(µa) − d(νa), νaαy),
b = (µbβy, d(µb) − d(νb), νbβy),
ab = (µabγy, d(µab) − d(νab), νabγy),
ai = (µaαiyi, d(µa) − d(νa), νaαiyi)
bi = (µbβiyi, d(µb) − d(νb), νbβiyi)
aibi = (µabγiyi, d(µab) − d(νab), νabγiyi)
for some α, β, γ, αi, βi, γi ∈ Λ and y, yi ∈ Λ∞.
Since ai → a we must also have αiyi → αy in Λ∞. Thus, for large enough
i, αiyi ∈ Z(α) := {αy : y ∈ Λ∞, y(0) = s(α)} (cf. Proposition 2.8 of [9]). It
follows that
ai = (µaαy′
bi = (µbβz′
i, d(µb) − d(νb), νbβz′
i)
i, d(µa) − d(νa), νaαy′
aibi = (µabγw′
i),
i, d(µab) − d(νab), νabγw′
i),
where (since each pair (ai, bi) ∈ G(2)
Λ by hypothesis)
νaαy′
i = µbβz′
i; µaαy′
i = µabγw′
i;
νbβz′
i = νabγw′
i.
11
Now, νaα = µbβ by [11] Lemma 6.3, and thus y′
i = z′
i. A similar argument
i. In other words, for large enough i,
gives z′
i = w′
i as well, so y′
i = z′
i = w′
σcti (ai, bi) = cti (µa, α)cti (µb, β)cti (νab, γ)cti(νa, α)cti (νb, β)cti (µab, γ)
= σcti (a, b),
as claimed. As observed in the first paragraph of the proof, it now follows that
ω is a homotopy of cocycles on GΛ as desired.
Corollary 3.9. Let {ct} be a homotopy of cocycles on a k-graph Λ, and define
a cocycle ω on GΛ × [0, 1] as in Definition 3.7. Then C∗(GΛ × [0, 1], ω) is a
C([0, 1])-algebra with fiber algebra C∗(GΛ, σct ) ∼= C∗(Λ, ct) at t ∈ [0, 1].
Proof. Proposition 3.8 tells us that ω is a homotopy of cocycles on GΛ, and
Theorem 3.3 tells us that the fiber over t ∈ [0, 1] of the C([0, 1])-algebra C∗(GΛ ×
[0, 1], ω) is C∗(GΛ, σct ). The final isomorphism is provided by Corollary 7.8 of
[11].
4 The main theorem
Our goal in this section is to prove the following:
Theorem 4.1. Let Λ be a row-finite k-graph with no sources and let {ct}t∈[0,1]
be a homotopy of cocycles on Λ. Then
K∗(C∗(Λ, c0)) ∼= K∗(C∗(Λ, c1)).
Moreover, this isomorphism preserves the K-theory class of the vertex projection
sv for each v ∈ Obj(Λ).
We begin by proving a stronger version of Theorem 4.1 in the simpler case
when the degree functor d satisfies d(λ) = δb(λ) := b(s(λ)) − b(r(λ)) for some
function b : Obj(Λ) → Zk; this is Proposition 4.2 below. We then combine
Proposition 4.2 with techniques from [10] to prove Theorem 4.1 in full generality.
4.1 The AF case
If (Λ, d) is a k-graph such that d = δb, then Lemma 8.4 of [11] tells us that
C∗(Λ, c) and C∗(Λ) are both AF-algebras, with the same approximating subal-
gebras and multiplicities of partial inclusions. Consequently, C∗(Λ, c) ∼= C∗(Λ).
In order to fix notation for what follows, we describe this isomorphism in some
detail.
Lemma 3.1 of [9] shows that if Λ is a row-finite, source-free k-graph, then
µ : s(λ) = s(µ)} spans a dense ∗-subalgebra of C∗(Λ). Moreover, when
µ : b(s(λ)) = b(s(µ)) = n} forms a
{sλs∗
d = δb, Lemma 5.4 of [9] tells us that {sλs∗
collection of matrix units for the subalgebra
An = span{sλs∗
µ : b(s(λ)) = b(s(µ)) = n} ∼= Mb(v)=n
K(ℓ2(s−1(v))).
12
Observe that we can think of An as a subalgebra of C∗(Λ) or of C∗(Λ, c). In
fact, these subalgebras allow us to exhibit C∗(Λ, c) and C∗(Λ) as AF algebras:
C∗(Λ, c) = lim−→(An, φc
where the connecting maps φm,n, φc
m,n) and C∗(Λ) = lim−→(An, φm,n),
m,n : An → Am are given by
m,n(sλs∗
φc
µ) =
φm,n(sλs∗
µ) =
Xr(α)=s(λ),b(s(α))=m
Xr(α)=s(λ),b(s(α))=m
c(λ, α)c(µ, α)sλαs∗
µα
sλαs∗
µα.
We can now describe explicitly the isomorphism C∗(Λ, c) ∼= C∗(Λ). As in
Theorem 4.2 of [10], write 1 for (1, . . . , 1) ∈ Nk, and define κ : Λ → T by
κ(λ) =(cid:26)
1,
d(λ) 6≥ 1
κ(µ)c(µ, α), d(α) = 1 and λ = µα.
For n ∈ Zk, let Un = Pb(s(λ))=n κ(λ)sλs∗
will show that for any λ, µ with sλs∗
λ ∈ U (M (An)). A quick computation
µ ∈ An,
Ad Un(sλs∗
µ) = κ(λ)κ(µ)sλs∗
µ.
(5)
Moreover, the factorization property tells us that for any h ∈ Z,
φc
(h+1)1,h1 ◦ Ad Uh1 = Ad U(h+1)1 ◦ φ(h+1)1,h1.
In other words, Ad U∗ intertwines the connecting maps φc
implements the isomorphism C∗(Λ) → C∗(Λ, c).
m,n, φm,n, and hence
We can now use this isomorphism to prove that a homotopy of cocycles on
Λ gives rise to a trivial continuous field when d = δb:
Proposition 4.2. Let (Λ, d) be a row-finite, source-free k-graph such that d = δb
for some function b : Obj(Λ) → Zk; let {ct}t∈[0,1] be a homotopy of cocycles on
Λ; and let ω be the cocycle on GΛ × [0, 1] associated to {ct}t∈[0,1] as in Definition
3.7. We have an isomorphism of C([0, 1])-algebras
C∗(GΛ × [0, 1], ω) ∼= C∗(GΛ × [0, 1]) ∼= C([0, 1]) ⊗ C∗(Λ).
Proof. Recall that
C∗(GΛ × [0, 1])t ∼= C∗(GΛ, σct ) ∼= C∗(Λ, ct) ∼= C∗(Λ)
if d = δb. Thus, the C([0, 1])-algebras C∗(GΛ × [0, 1], ω) and C([0, 1]) ⊗ C∗(Λ)
have isomorphic fibers over each point t ∈ [0, 1].
In order to prove the Proposition, we need to show that these isomorphisms
C∗(GΛ, σct ) ∼= C∗(Λ) vary continuously in t, so that they patch together to give
us an isomorphism of C([0, 1])-algebras C∗(GΛ × [0, 1], ω) ∼= C([0, 1]) ⊗ C∗(Λ).
13
For each t ∈ [0, 1], let πt : C∗(Λ, ct) → C∗(GΛ, σct ) denote the isomorphism
described in Theorem 6.7 of [11]. Let π : C∗(Λ) → C∗(GΛ) denote the equivalent
isomorphism for the case of a trivial cocycle c. For each n ∈ Zk, write U t
n for
the unitary U t
n : An → An associated to the cocycle ct as above. Setting
Ψt := πt ◦ Ad U t
∗ ◦ π−1
consequently gives an isomorphism of C∗-algebras Ψt : C∗(GΛ) → C∗(GΛ, σct).
We claim that Ψ := {Ψt}t∈[0,1] defines an isomorphism of C([0, 1])-algebras
Ψ : C∗(GΛ × [0, 1]) → C∗(GΛ × [0, 1], ω).
In order to prove this assertion, we begin by writing down an explicit formula for
Ψt on the characteristic functions 1Z(µ,ν) ∈ Cc(GΛ) where Z(µ, ν) ∈ P, where
P is the partition of GΛ described in Lemma 6.6 of [11].
Recall that the value of σct(a, b) depends only on the sets Z(µ, ν) ∈ P
containing the points a, b, and ab in GΛ. Moreover, the proof of [11] Theorem 6.7
establishes that, if 1Z(µ,ν) denotes the characteristic function on Z(µ, ν) ⊆ GΛ,
and we write a ∈ Z(µ, ν) as a = bd where b ∈ Z(µ, s(µ)), d ∈ Z(s(ν), ν),
πt(sµs∗
ν)(a) = 1Z(µ,ν)(a)σct (b, d)σct (d−1, d) = 1Z(µ,ν)(a)σct (bd, d−1).
Moreover, we have Z(µ, s(µ)) ∈ P ∀ µ ∈ Λ by Lemma 6.6 of [11]. If we also
have Z(µ, ν) ∈ P, then the elements α, β, γ in the formula for σct (bd, d−1) given
in Definition 2.7 are all units, so for any t, σct (bd, d−1) = 1 by our hypothesis
that any cocycle c satisfy c(λ, s(λ)) = c(r(λ), λ) = 1. Thus,
Z(µ, ν) ∈ P ⇒ πt(sµs∗
ν) = 1Z(µ,ν) ⇒ Ψt(1Z(µ,ν)) = κt(µ)κt(ν)1Z(µ,ν).
Now, observe that each f ∈ Cc(GΛ × [0, 1]) can be written as a finite sum
f (a, t) =Pi∈N fi(a, t), where, for all i, fi ∈ C(Z(µi, νi) × [0, 1]) and Z(µi, νi) ∈
P. Consequently, on Cc(GΛ × [0, 1]), our map Ψ becomes
Ψ Xi∈N
fi! (a, t) = Xi∈N
Ψt(fi(·, t))(a) = Xi∈N
κt(µi)κt(νi)fi(a, t);
(6)
the fact that all the sums are finite implies that Ψ takes Cc(GΛ × [0, 1]) onto
Cc(GΛ × [0, 1]).
Since Ψ is evidently C([0, 1])-linear and is a ∗-isomorphism in each fiber,
Proposition C.10 of [21] tells us that Ψ is norm-preserving. Moreover, Ψ is a
∗-homomorphism since the operations in Cc(GΛ × [0, 1]) preserve the fiber over
t ∈ [0, 1], and each Ψt is a ∗-homomorphism.
In other words, Ψ extends to an isomorphism of C([0, 1])-algebras
Ψ : C∗(GΛ × [0, 1]) ∼= C∗(GΛ × [0, 1], ω).
A straightforward check will establish that the identity map on Cc(GΛ × [0, 1])
induces an isomorphism id : C∗(GΛ × [0, 1]) → C([0, 1], C∗(GΛ)) of C([0, 1])-
algebras; the isomorphism C∗(GΛ) ∼= C∗(Λ) of [9] Corollary 3.5(i) now finishes
the proof.
14
Remark 4.3. Note that Ψ induces an isomorphism Φ : C([0, 1]) ⊗ C∗(Λ) →
C∗(GΛ × [0, 1], ω) as follows. If Z(µ, ν) ∈ P and f ∈ C([0, 1]), then
Φ(f ⊗ sµs∗
ν)(x, t) = f (t)1Z(µ,ν)(x)κt(µ)κt(ν).
(7)
Remark 4.4. Since evaluation at t ∈ [0, 1] induces a homotopy equivalence be-
tween C([0, 1], C∗(Λ)) and C∗(Λ), the isomorphism established in the previous
Proposition implies that evaluation at t also induces a homotopy equivalence
between C∗(GΛ × [0, 1], ω) and its fiber algebra C∗(GΛ, σct ) when d = δb.
4.2 Proof of Theorem 4.1
To leverage Proposition 4.2 into the proof of Theorem 4.1, we will use the skew-
product k-graphs Λ ×d Zk:
Definition 4.5 ([9] Definition 5.1). Given a k-graph (Λ, d), the skew-product
k-graph Λ ×d Zk is the set Λ × Zk, with the structure maps
r(λ, n) = (r(λ), n);
s(λ, n) = (s(λ), n + d(λ));
d(λ, n) = d(λ),
and multiplication given by (λ, n)(µ, n + d(λ)) = (λµ, n) for (λ, µ) ∈ Λ∗2.
Observe that the function b : (Λ ×d Zk)(0) = Λ(0) × Zk → Zk given by
b(v, n) = n satisfies δb = d on Λ ×d Zk. Moreover, if Λ is row-finite and source-
free, then so is Λ ×d Zk.
We can now complete the proof of Theorem 4.1.
Proof of Theorem 4.1: Let φ : Λ ×d Zk → Λ be the projection onto the first
coordinate: φ(λ, n) = λ. A cocycle c on Λ induces a cocycle c ◦ φ on the skew
product k-graph Λ ×d Zk:
c ◦ φ ((λ, n), (µ, n + d(λ))) := c(λ, µ)
whenever (λ, µ) ∈ Λ∗2. Note that if {ct}t∈[0,1] is a homotopy of cocycles on Λ
then {ct ◦ φ}t is also a homotopy of cocycles on Λ ×d Zk.
If ω is the homotopy of cocycles on GΛ×d Zk associated to the homotopy
{ct}t∈[0,1] of cocycles on Λ, then Proposition 4.2 tells us that
C∗(GΛ×d Zk × [0, 1], ω) ∼= C([0, 1]) ⊗ C∗(Λ ×d Zk).
Now, we define an action of Zk on C([0, 1]) ⊗ C∗(Λ ×d Zk) by setting
f ⊗ sλ,n · m := f ⊗ sλ,n+m.
(8)
To see that this formula gives us a well-defined action of Zk on C([0, 1]) ⊗
C∗(Λ ×d Zk), one checks first that for each m ∈ Zk, {sλ,m+n : λ ∈ Λ, n ∈ Zk} is
a collection of partial isometries satisfying the defining axioms (CK1)-(CK4) for
C∗(Λ ×d Zk). Consequently, the universal property of C∗(Λ ×d Zk) implies that
for each fixed m ∈ Zk, the map sλ,n 7→ sλ,n+m determines a ∗-homomorphism
αm : C∗(Λ ×d Zk) → C∗(Λ ×d Zk).
15
Each αm is invertible with inverse α−m; it follows that m 7→ αm defines a group
action of Zk on C∗(Λ ×d Zk). Thus, Equation (8) describes a well-defined action
id ⊗ α of Zk on C([0, 1]) ⊗ C∗(Λ ×d Zk), given by m 7→ id ⊗ αm. The fact that
the degree map on Λ ×d Zk is a coboundary now allows us to combine the action
id⊗α with the isomorphism Φ : C([0, 1])⊗C∗(Λ×d Zk) → C∗(GΛ×d Zk ×[0, 1], ω)
of Remark 4.3 to obtain an action β of Zk on C∗(GΛ×d Zk × [0, 1], ω):
βn(cid:16)Φ(f ⊗ sµ,ms∗
ν,m+d(µ)−d(ν))(cid:17) := Φ(id ⊗ αn(f ⊗ sµ,ms∗
ν,m+d(µ)−d(ν))).
Moreover, since both id⊗α and Φ (and hence β) fix C([0, 1]) by construction,
Lemma 5.3 of [10] tells us that the crossed product
C∗(GΛ×d Zk × [0, 1], ω) ⋊β Zk ∼=(cid:0)C([0, 1]) ⊗ C∗(Λ ×d Zk)(cid:1) ⋊id⊗α Zk
is a C([0, 1])-algebra with fiber C∗(GΛ×d Zk , σct◦φ) ⋊βt
Zk, where
(βt)n(Φt(sµ,ms∗
ν,m+d(µ)−d(ν))) = Φt(αn(s(µ,m)s∗
(ν,m+d(µ)−d(ν))))
= κt(µ)κt(ν)1Z((µ,m+n),(ν,m+n+d(µ)−d(ν)))
whenever Z((µ, m + n), (ν, m + n + d(µ) − d(ν))) is in the partition P of GΛ×d Zk
that we used in the proof of Proposition 4.2.
Recall that we have a homotopy equivalence qt : C∗(GΛ×d Zk × [0, 1], ω) →
C∗(GΛ×d Zk , σct ). A computation will show that qt is equivariant with respect
to the actions β, βt of Zk; thus, Theorem 5.1 of [10] tells us that
K∗(C∗(GΛ×d Zk × [0, 1], ω) ⋊β Zk) ∼= K∗(C∗(GΛ×d Zk , σct◦φ) ⋊βt
Zk).
(9)
Thanks to Lemma 5.2 of [10], we know that C∗(Λ ×d Zk, ct ◦ φ) ⋊lt Zk ∼M E
C∗(Λ, ct), where ltm(sλ,n) = sλ,n+m. To make use of this result, we need to
show that βt induces the action lt on C∗(Λ ×d Zk, ct ◦ φ).
Recall from the proof of Proposition 4.2 that πt(sλ,m) = 1Z((λ,m),(s(λ),m+s(λ))),
since Z((λ, m), (s(λ), m + s(λ))) ∈ P always. Observe that
C∗(GΛ×d Zk , σct◦φ) ⋊βt
Zk ∼= C∗(Λ ×d Zk, ct ◦ φ) ⋊γt
Zk, where
(10)
(γt)n(sλ,m) := (πt)−1((βt)n(πt(sλ,m))) = (πt)−1(βt)n(1Z((λ,m),(s(λ),m)))
= (πt)−1(βt)n(Φt(κt(λ)s(λ,m))) = (πt)−1(cid:16)Φt(αn(κt(λ)sλ,m))(cid:17)
= (πt)−1(cid:16)Φt(κt(λ)sλ,m+n)(cid:17) = (πt)−1(cid:0)1Z((λ,m+n),(s(λ),m+n))(cid:1)
= sλ,m+n.
It follows that the action (γt) induced by βt agrees with lt, as desired. Now,
the Morita equivalence of Lemma 5.2 of [10] and Equation (10) tell us that
C∗(GΛ×d Zk , σct◦φ) ⋊βt
Zk ∼M E C∗(Λ, ct).
(11)
16
Combining Equations (9) and (11) now yields
K∗(C∗(Λ, ct)) ∼= K∗(C∗(GΛ×d Zk × [0, 1], ω) ⋊β Zk)
for any t ∈ [0, 1]. It follows that, if {ct}t∈[0,1] is a homotopy of cocycles on a
row-finite k-graph Λ with no sources, then for any s, t ∈ [0, 1],
K∗(C∗(Λ, ct)) ∼= K∗(C∗(Λ, cs)).
It remains to show that this isomorphism preserves the K-theory class of
each vertex projection sv. Essentially, this follows because the cocycles ct, and
thus the functions κt, are all trivial on any v ∈ Obj(Λ).
To be precise, let v ∈ Obj(Λ) and define fv ∈ Cc(Zk, Cc(GΛ×d Zk × [0, 1]))) ⊆
C∗(GΛ×d Zk × [0, 1], ω) ⋊β Zk by
fv(n)(a, t) =(cid:26) 1, a ∈ Z(v,0),(v,0) and n = 0
else.
0,
Then the projection qt ⋊id(fv) of fv onto the fiber algebra C∗(GΛ×d Zk , ωt)⋊βt
is independent of the choice of t ∈ [0, 1]:
Zk
qt ⋊ id(fv)(n)(a) =(cid:26) 1, a ∈ Z(v,0),(v,0) and n = 0
else
0,
for any t ∈ [0, 1]. Moreover, the isomorphism Φt : C∗(Λ ×d Zk, ct ◦ φ) →
C∗(GΛ×d Zk , σct◦φ) of Remark 4.3 satisfies
Φt ⋊ id(j(s(v,0))) = qt ⋊ id(fv),
(12)
where j : C∗(Λ ×d Zk, ct ◦ φ) → C∗(Λ ×d Zk, ct ◦ φ) ⋊lt Zk is the canonical
embedding of C∗(Λ ×d Zk, ct ◦ φ) into the crossed product.
The fact that the Morita equivalence C∗(Λ, ct) ∼M E C∗(Λ×d Zk, ct◦φ)⋊lt Zk
takes sv ∈ C∗(Λ, ct) to j(s(v,0)) (cf. Lemma 5.2 in [10]) thus implies that our
K-theoretic isomorphism K∗(C∗(GΛ×d Zk × [0, 1], ω) ⋊β Zk) → K∗(C∗(Λ, ct)),
which is given by the composition of the Morita equivalence (11) with the ∗-
homomorphism
qt ⋊ id : C∗(GΛ×d Zk × [0, 1], ω) ⋊β Zk → C∗(GΛ×d Zk , ωt) ⋊βt
Zk
∼= C∗(Λ ×d Zk, ct ◦ φ) ⋊γt
Zk,
takes [fv] to [sv] for any v ∈ Obj(Λ) and any t ∈ [0, 1]. Consequently, the iso-
morphism K∗(C∗(Λ, ct)) ∼= K∗(C∗(Λ, cs)) preserves the class of sv, as claimed.
This finishes the proof of Theorem 4.1.
Remark 4.6. It's tempting to think that since C∗(Λ ×d Zk, c ◦ φ) ∼= C∗(Λ ×d Zk)
and C∗(Λ, c) ∼M E C∗(Λ ×d Zk, c ◦ φ) ⋊lt Zk for any cocycle c on Λ, any two
twisted k-graph C∗-algebras should be Morita equivalent. This statement is
false, however (the rotation algebras provide a counterexample). The flaw lies
17
in the fact that the isomorphism Ad U∗ : C∗(Λ×d Zk, c◦φ) → C∗(Λ×d Zk) is not
equivariant with respect to the left-translation action of Zk, so the isomorphism
C∗(Λ ×d Zk, c ◦ φ) ∼= C∗(Λ ×d Zk)
does not pass to an isomorphism C∗(Λ ×d Zk, c ◦ φ) ⋊lt Zk → C∗(Λ ×d Zk) ⋊lt Zk.
In other words, a K-theoretic equivalence of twisted k-graph C∗-algebras is the
best result we can hope for in general.
5 Future work
The standing hypotheses of this paper, that our k-graphs be row-finite and
source-free, are slightly more restrictive than the current standard for k-graphs.
Thus, we would like to extend Theorem 4.1 to apply to all finitely aligned k-
graphs. Finitely aligned k-graphs were introduced in [16, 17], and it seems that
they constitute the largest class of k-graphs to which one can profitably as-
sociate a C∗-algebra. However, the Kumjian-Pask construction of a groupoid
GΛ associated to a k-graph Λ, which we described in Section 2 and which we
use throughout the proof of Theorem 4.1, only works when Λ is row-finite and
source-free. In [4], Farthing, Muhly, and Yeend provide an alternate construc-
tion of a groupoid G which can be associated to an arbitrary finitely-aligned
k-graph, and we hope that this approach will allow us to apply groupoid results
such as Theorem 3.3 to study the effect on K-theory of homotopies of cocycles
for finitely-aligned k-graphs.
References
[1] J.H. Brown and A. an Huef, Decomposing the C∗-algebras of groupoid ex-
tensions, Proceedings of the American Mathematical Society 142 (2014),
1261 -- 1274.
[2] L.O. Clark and A. an Huef, The representation theory of C∗-algebras as-
sociated to groupoids, Mathematical Proceedings of the Cambridge Philo-
sophical Society 153 (2012), 167 -- 191.
[3] S. Echterhoff, W. Luck, N.C. Phillips, and S. Walters, The structure
of crossed products of irrational rotation algebras by finite subgroups of
SL2(Z), Journal fur die reine und angewandte Mathematik 639 (2010),
173 -- 221.
[4] C. Farthing, P.S. Muhly, and T. Yeend, Higher-rank graph C∗-algebras:
an inverse semigroup and groupoid approach, Semigroup Forum 71 (2005),
no. 2, 159 -- 187.
[5] E. Gillaspy, K-theory and homotopies of 2-cocycles on transformation
groups, Journal of Operator Theory (to appear).
18
[6] G. Goehle, Groupoid crossed products, Ph.D. thesis, Dartmouth College,
May 2009.
[7] A. an Huef, A. Kumjian, and A. Sims, A Dixmier-Douady theorem for Fell
algebras, Journal of Functional Analysis 260 (2011), 1543 -- 1581.
[8] A. Kumjian, On C∗-diagonals, Canadian Journal of Mathematics 38
(1986), 969 -- 1008.
[9] A. Kumjian and D. Pask, Higher rank graph C∗-algebras, New York Journal
of Mathematics 6 (2000), 1 -- 20.
[10] A. Kumjian, D. Pask, and A. Sims, On the K-theory of twisted higher-rank-
graph C∗-algebras, Journal of Mathematical Analysis and Applications 401
(2013), no. 1, 104 -- 113.
[11]
, On twisted higher-rank graph C∗-algebras, Transactions of the
American Mathematical Society (2013).
[12] P.S. Muhly, Coordinates in operator algebra.
[13] P.S. Muhly, J.N. Renault, and D.P. Williams, Continuous-trace groupoid
C∗-algebras. III, Transactions of the American Mathematical Society 348
(1996), 3621 -- 3641.
[14] P.S. Muhly and D.P. Williams, Continuous trace groupoid C∗-algebras II,
Mathematica Scandinavica 70 (1992), 127 -- 145.
[15] M. Pimsner and D.-V. Voiculescu, Exact sequences for K-groups and Ext-
groups of certain crossed-product C∗-algebras, Journal of Operator Theory
4 (1980), 93 -- 118.
[16] I. Raeburn and A. Sims, Product systems of graphs and the Toeplitz algebras
of higher-rank graphs, Journal of Operator Theory 53 (2005), no. 2, 399 --
429.
[17] Iain Raeburn, Aidan Sims, and Trent Yeend, The C∗-algebras of finitely
aligned higher-rank graphs, Journal of Functional Analysis 213 (2004), 206 --
240.
[18] J. Renault, A groupoid approach to C∗-algebras, Lecture Notes in Mathe-
matics, vol. 793, Springer-Verlag, 1980.
[19] G. Robertson and T. Steger, Affine buildings, tiling systems and higher rank
Cuntz-Krieger algebras, Journal fur die reine und angewandte Mathematik
513 (1999), 115 -- 144.
[20] J.-L. Tu, P. Xu, and C. Laurent-Gengoux, Twisted K-theory of differ-
entiable stacks, Annales Scientifiques de l'´Ecole Normale Sup´erieure 37
(2004), 841 -- 910.
19
[21] D.P. Williams, Crossed products of C∗-algebras, Mathematical Surveys &
Monographs, vol. 134, AMS, 2007.
20
|
1604.03032 | 1 | 1604 | 2016-04-11T17:05:39 | Elementary constructions of non-discrete C*-simple groups | [
"math.OA",
"math.GR"
] | Recently Raum has given the first examples of locally compact non-discrete groups with the simple reduced group C*-algebra, answering a question of de la Harpe. Here we construct such groups whose proof relies only on results in the discrete case. | math.OA | math |
ELEMENTARY CONSTRUCTIONS OF NON-DISCRETE C∗-SIMPLE
GROUPS
YUHEI SUZUKI
Abstract. Recently Raum has given the first examples of locally compact non-discrete
groups with the simple reduced group C∗-algebra, answering a question of de la Harpe.
Here we construct such groups whose proof relies only on results in the discrete case.
Let G be a locally compact group. Recall that the left regular representation λ of G is
the unitary representation of G on the Hilbert space L2(G, µ) acting by the left translation.
Here µ denotes the left Haar measure of G. This representation induces a ∗-representation
of the group algebra Cc(G) on L2(G, µ). The reduced group C∗-algebra C∗
λ(G) of G is
the operator norm closure of the image of Cc(G) under this representation. This provides
basic and important examples of C∗-algebras.
A locally compact group is said to be C∗-simple if its reduced group C∗-algebra has no
proper closed two-sided ideal. A basic question asks when a given group is C∗-simple. The
first such a group was given by Powers [8], by showing that the free groups are C∗-simple.
His strategy is quite powerful, and until the recent breakthrough result of Kalantar and
Kennedy [5], his method was basically the only way to show C∗-simplicity. Now in the
discrete case, fairly satisfactory characterizations of C∗-simplicity are obtained [3], [5], [6].
Recently, among other things, Raum [9] has constructed the first examples of non-
discrete C∗-simple groups, based on properties of groups acting on trees. The existence of
such a group was asked by de la Harpe ([4], Question 5).
In this paper, we establish a quite elementary method to construct non-discrete C∗-
simple groups. Our result gives explicit examples, and we only use previously known
results in the discrete case. (In fact, Powers's original result is enough to construct such a
group.) Furthermore, we show that these groups have the unique trace property. Here we
say a locally compact group has the unique trace property if its reduced group C∗-algebra
has a unique lower semicontinuous semifinite trace up to scaling.
For a compact open subgroup K of a locally compact group G, let pK denote the image
λ(G). Then pK is the projection
of the normalized characteristic function µ(K)−1χK in C∗
onto the subspace L2(G)K of K-invariant functions.
To provide non-discrete C∗-simple groups, we first establish a criterion for C∗-simplicity.
Proposition. Let G be a locally compact group. Assume we have a decreasing sequence
(Kn)∞
n=1 of clopen
subgroups of G with the following properties.
n=1 of compact open subgroups of G and an increasing sequence (Ln)∞
• Each Ln contains Kn and normalizes it.
• The quotient groups Ln/Kn are C∗-simple.
• The intersection T∞
• The union S∞
n=1 Ln is equal to G.
n=1 Kn is the trivial subgroup {e}.
Then G is C∗-simple and has the unique trace property.
2000 Mathematics Subject Classification. Primary 22D25, Secondary 46L05.
Key words and phrases. C∗-simplicity, locally compact group, group von Neumann algebra.
1
2
YUHEI SUZUKI
Proof. We only prove simplicity. The unicity of trace is similarly shown. (We recall that
for discrete groups, C∗-simplicity implies the unicity of tracial states on the reduced group
C∗-algebra [1], [3].) For each n, let An denote the C∗-subalgebra of C∗
λ(G) generated by
the set {λgpKn : g ∈ Ln}. Since Ln normalizes Kn and Kn is compact open, the map
g ∈ Ln 7→ λgpKn ∈ An defines a unitary representation π of Ln/Kn on L2(G)Kn . The
canonical isomorphism L2(G)Kn ∼= ℓ2(Kn\G) yields the unitary equivalence of the left
translation Ln/Kn-actions. Since the left translation action of Ln/Kn on Kn\G is free,
this shows the unitary equivalence of π and a multiple of the left regular representation.
Hence An is isomorphic to C∗
λ(Ln/Kn), which is simple by our assumption. The inclusions
Kn+1 ⊂ Kn ⊂ Ln ⊂ Ln+1 imply An ⊂ An+1. Hence the norm closure A of the increasing
union S∞
increases to G, we conclude A = C∗
n=1 decreases to the trivial subgroup and (Ln)∞
n=1 An is simple. Since (Kn)∞
λ(G).
n=1
(cid:3)
Theorem. For each n ∈ N, let Γn be a discrete group and let Fn be a finite group acting
on the group Γn whose semidirect product Γn ⋊ Fn is C∗-simple. Set G := (L∞
n=1 Γn) ⋊
Q∞
n=1 Fn is the compact group
equipped with the product topology, and the action Q∞
n=1 Γn is the product of
given actions. Then G is C∗-simple and has the unique trace property.
n=1 Γn is regarded as a discrete group, Q∞
n=1 Fn. Here L∞
n=1 Fn y L∞
Before the proof, we note that Powers's result [8] already gives groups and actions
satisfying the conditions in Theorem. For instance, consider the free product Z2 ∗ Z. Let
Γ be the kernel of the quotient homomorphism Z2 ∗ Z → Z2 given by sending the second
free product component to 0. Then, since this map has a homomorphism lifting, we have a
semidirect product decomposition Z2 ∗Z = Γ⋊Z2. Powers's proof [8] shows that C∗
λ(Z2∗Z)
is simple and has a unique tracial state.
Proof of Theorem. For each n ∈ N, set Kn := Q∞
k=1 Fk.
(Both are regarded as a clopen subgroup of G in the canonical way.) Then it is not hard to
check that the sequences (Kn)∞
n=1 satisfy the conditions in Proposition. (cid:3)
k=n+1 Fk and Ln := (Ln
n=1 and (Ln)∞
k=1 Γk)⋊Q∞
Group von Neumann algebras. For a group as in Theorem, in a similar way, it can
be shown that its group von Neumann algebra is a factor of type II∞. This factor is not
injective, as a finite corner has a non-injective subfactor. By modifying our construction,
for any rational number 0 < q ≤ 1, we can construct a C∗-simple group whose group von
Neumann algebra is a factor of type IIIq as follows.
Sketch of the construction. We only show the case q < 1. The case q = 1 then follows
by taking an appropriate direct product of such groups (cf.
[2]). Take n1, n2 ∈ N with
q = n1/n2. For i = 1, 2, let Γi be a discrete group and let Fi be a finite group of
order ni acting on Γi such that the semidirect product Γi ⋊ Fi is C∗-simple. (Such ones
are easily found in a similar way to that in the remark below Theorem.) Set Hi
:=
(Ln∈Z Γi) ⋊ ((Ln≤0 Fi) × (Qn≥1 Fi)) for i = 1, 2. Then Hi is naturally regarded as a
locally compact group. It is easy to check that Hi satisfies the conditions in Proposition.
Now let αi be the automorphism of Hi given by the forward shift of the indices n ∈ Z. Set
H := H1 × H2. Then α := α1 × α−1
2 defines an automorphism of H. Put G := H ⋊α Z.
Obviously α induces automorphisms of C∗
λ(H) and L(H). We denote them by the same
λ(H) ⋊α Z and L(G) ∼= L(H) ¯⋊αZ.
λ(G) ∼= C∗
symbol α. Then we have isomorphisms C∗
Since L(H) is a type II∞ factor and α scales the trace on L(H) at the rate q, Connes's
theorem ([2] Theorem 4.4.1) shows that L(G) is a type IIIq factor. Also, since α scales a
semifinite trace on C∗
λ(H), nonzero powers of α are outer. Now Kishimoto's theorem ([7],
Theorem 3.1) shows the C∗-simplicity of G.
(cid:3)
ELEMENTARY CONSTRUCTIONS OF NON-DISCRETE C∗-SIMPLE GROUPS
3
Remark. Raum [9] has already constructed C∗-simple groups whose group von Neumann
algebras are factors of types as above. Also, for any λ ∈ [0, 1], Sutherland [10] has
constructed a factorial group von Neumann algebra of type IIIλ.
Acknowledgement. This work was carried out while the author was staying at Mittag-
Leffler institute for the program "Classification of operator algebras: complexity, rigidity,
and dynamics". He acknowledges the organizers and the institute for the invitation and
the kind hospitality. He also thanks Sven Raum for helpful comments on the first draft
and letting him know the reference [10]. This work was supported by JSPS Research
Fellowships for Young Scientists (PD 28-4705).
References
[1] E. Breuillard, M. Kalantar, M. Kennedy, N. Ozawa, C∗-simplicity and the unique trace property for
discrete groups. arXiv:1410.2518.
[2] A. Connes, Une classification des facteurs de type III. Ann. Sci. ´Ec. Norm. Supr. (4) 6 (1973), 133 -- 252.
[3] U. Haagerup, A new look at C∗-simplicity and the unique trace property of a group. arXiv:1509.05880.
[4] P. de la Harpe, On simplcity of reduced C∗-algebras of groups. Bull. Lond. Math. Soc. 39 (2007), 1 -- 26.
[5] M. Kalantar, M. Kennedy, Boundaries of reduced C∗-algebras of discrete groups. To appear in J. reine
angew. Math., arXiv:1405.4359.
[6] M. Kennedy, Characterizations of C∗-simplicity. arXiv:1509.01870.
[7] A. Kishimoto, Outer automorphisms and reduced crossed products of simple C∗-algebras. Comm. Math.
Phys. 81 (1981), no. 3, 429 -- 435.
[8] R. T. Powers, Simplicity of the C∗-algebra associated with the free group on two generators. Duke
Math. J. 42 (1975), 151 -- 156.
[9] S. Raum, C∗-simplicity of locally compact Powers groups. To appear in J. reine angew. Math.,
arXiv:1505.07793.
[10] C. E. Sutherland, Type analysis of the regular representation of a non-unimodular group. Pacific J.
Math. 79 (1978), 225 -- 250.
Graduate School of Science, Chiba University, Inage-ku, Chiba 263-8522, Japan
E-mail address: [email protected]
|
1507.07610 | 1 | 1507 | 2015-07-27T23:58:22 | Realising the Toeplitz algebra of a higher-rank graph as a Cuntz-Krieger algebra | [
"math.OA"
] | For a row-finite higher-rank graph {\Lambda}, we construct a higher-rank graph T{\Lambda} such that the Toeplitz algebra of {\Lambda} is isomorphic to the Cuntz-Krieger algebra of T{\Lambda}. We then prove that the higher-rank graph T{\Lambda} is always aperiodic and use this fact to give another proof of a uniqueness theorem for the Toeplitz algebras of higher-rank graphs. | math.OA | math |
REALISING THE TOEPLITZ ALGEBRA OF A HIGHER-RANK
GRAPH AS A CUNTZ-KRIEGER ALGEBRA
YOSAFAT E. P. PANGALELA
Abstract. For a row-finite higher-rank graph Λ, we construct a higher-rank graph T Λ
such that the Toeplitz algebra of Λ is isomorphic to the Cuntz-Krieger algebra of T Λ. We
then prove that the higher-rank graph T Λ is always aperiodic and use this fact to give
another proof of a uniqueness theorem for the Toeplitz algebras of higher-rank graphs.
1. Introduction
Higher-rank graphs and their Cuntz-Krieger algebras were introduced by Kumjian and
Pask in [4] as a generalisation of the Cuntz-Krieger algebras of directed graphs. Kumjian
and Pask proved an analogue of the Cuntz-Krieger uniqueness theorem for a family of
aperiodic higher-rank graphs [4, Theorem 4.6]. Aperiodicity is a generalisation of Condi-
tion (L) for directed graphs and comes in several forms for different kinds of higher-rank
graphs (see [1, 4, 5, 9, 10, 11, 12, 13]).
The Toeplitz algebra of a directed graph is an extension of the Cuntz-Krieger algebra in
which the Cuntz-Krieger equations at vertices are replaced by inequalities. An analogous
family of Toeplitz algebras for higher-rank graph was introduced and studied by Raeburn
and Sims [8]. They proved a uniqueness theorem for Toeplitz algebras [8, Theorem 8.1],
generalising a previous theorem for directed graphs [3, Theorem 4.1].
For a directed graph E, the Toeplitz algebra of E is canonically isomorphic to the
Cuntz-Krieger algebra of a graph T E (see [6, Theorem 3.7] and [14, Lemma 3.5]). Here
we provide an analogous construction for a row-finite higher-rank graph Λ. We build a
higher-rank graph T Λ, and show that the Toeplitz algebra of Λ is canonically isomorphic
to the Cuntz-Krieger algebra of T Λ (Theorem 4.1). Our proof relies on the uniqueness
theorem of [8]. However, it is interesting to observe that the higher-rank graph T Λ is
always aperiodic. Hence our isomorphism shows that the uniqueness theorem of [8] is a
consequence of the general Cuntz-Krieger uniqueness theorem of [10] (see Remark 4.3).
2. Higher-rank graphs
Let k be a positive integer. We regard Nk as an additive semigroup with identity 0.
For m, n ∈ Nk, we write m ∨ n for their coordinate-wise maximum.
A higher-rank graph or k-graph is a pair (Λ, d) consisting of a countable small category
Λ together with a functor d : Λ → Nk satisfying the factorisation property:
for every
λ ∈ Λ and m, n ∈ Nk with d (λ) = m + n, there are unique elements µ, ν ∈ Λ such that
λ = µυ and d (µ) = m, d (ν) = n. We then write λ (0, m) for µ and λ (m, m + n) for ν.
Date: April 29, 2015.
This research is part of the author's Ph.D. thesis, supervised by Professor Iain Raeburn and Dr. Lisa
Orloff Clark.
1
2
YOSAFAT E. P. PANGALELA
We regard elements of Λ0 as vertices and elements of Λ as paths. For detailed explanation
and examples, see [7, Chapter 10].
For v ∈ Λ0 and E ⊆ Λ, we define vE := {λ ∈ E : r (λ) = v} and m ∈ Nk, we write
Λm := {λ ∈ Λ : d (λ) = m}.We use term edge to denote a path e ∈ Λei where 1 ≤ i ≤ k,
and write
Λ1 := [1≤i≤k
Λei
for the set of all edges. We say that Λ is row-finite if for every v ∈ Λ0, the set vΛei is
finite for 1 ≤ i ≤ k. Finally, we say v ∈ Λ0 is a source if there exists m ∈ Nk such that
vΛm = ∅.
For a row-finite k-graph Λ, we shall construct a k-graph T Λ which is row-finite and
always has sources. Our k-graph T Λ is typically not locally convex in the sense of [9,
Definition 3.9] (see Remark 3.3), so the approriate definition of Cuntz-Krieger Λ-family is
the one in [10]. For detailed discussion about row-finite k-graphs and their generalisations,
see [15, Section 2].
From now on, we focus on a row-finite k-graph Λ. For λ, µ ∈ Λ, we say that τ is a
minimal common extension of λ and µ if
d (τ ) = d (λ) ∨ d (µ) , τ (0, d (λ)) = λ and τ (0, d (µ)) = µ.
Let MCE (λ, µ) denote the collection of all minimal common extensions of λ and µ. Then
we write
Λmin (λ, µ) := {(λ′, µ′) ∈ Λ × Λ : λλ′ = µµ′ ∈ MCE (λ, µ)} .
A set E ⊆ vΛ1 is exhaustive if for all λ ∈ vΛ, there exists e ∈ E such that Λmin (λ, e) 6= ∅.
A Toeplitz-Cuntz-Krieger Λ-family is a collection {tλ : λ ∈ Λ} of partial isometries in
a C ∗-algebra B satisfying:
(TCK1) {tv : v ∈ Λ0} is a collection of mutually orthogonal projections;
(TCK2) tλtµ = tλµ whenever s (λ) = r (µ); and
(TCK3) t∗
µ′ for all λ, µ ∈ Λ.
λtµ =P(λ′,µ′)∈Λmin(λ,µ) tλ′t∗
follows from (TCK1-3), and hence our definition is basically same as that of [8].
Meanwhile, based on [10, Proposition C.3], a Cuntz-Krieger Λ-family is a Toeplitz-
Remark 2.1. In [8, Lemma 9.2], Raeburn and Sims required also that "for all m ∈ Nk\ {0},
λ". However, by [10, Lemma 2.7 (iii)], this
v ∈ Λ0, and every set E ⊆ vΛm, tv ≥Pλ∈E tλt∗
(CK) Qe∈E(cid:0)tr(E) − tet∗
Cuntz-Krieger Λ-family {tλ : λ ∈ Λ} which satisfies
e(cid:1) = 0 for all v ∈ Λ0 and exhaustive E ⊆ vΛ1.
Raeburn and Sims proved in [8, Section 4] that there is a C ∗-algebra T C ∗ (Λ) generated
by a universal Toeplitz-Cuntz-Krieger Λ-family {tλ : λ ∈ Λ}. If {Tλ : λ ∈ Λ} is a Toeplitz-
Cuntz-Krieger Λ-family in C ∗-algebra B, we write φT for the homomorphism of T C ∗ (Λ)
into B such that φT (tλ) = Tλ for λ ∈ Λ. The quotient of T C ∗ (Λ) by the ideal generated
by
{Ye∈E(cid:0)tr(E) − tet∗
e(cid:1) = 0 : v ∈ Λ0, E ⊆ vΛ1 is exhaustive}
is generated by a universal family of Cuntz-Krieger Λ-family {sλ : λ ∈ Λ}, and hence we
can identify it with the C ∗-algebra C ∗ (Λ). For a Cuntz-Krieger Λ-family {Sλ : λ ∈ Λ} in
C ∗-algebra B, we write πS for the homorphism of C ∗ (Λ) into B such that πS (sλ) = Sλ
for λ ∈ Λ. Furthermore, we have sv 6= 0 for v ∈ Λ0 [10, Proposition 2.12].
(*)
(Tv − TeT ∗
e ) 6= 0
Ye∈vΛ1
THE TOEPLITZ ALGEBRA OF A HIGHER-RANK GRAPH
3
As for directed graphs, we have uniqueness theorems for the Toeplitz algebra [8, Theo-
rem 8.1] and the Cuntz-Krieger algebra [10, Theorem 4.5]. The former does not need any
hypothesis on the k-graph as stated in the following theorem.
Theorem 2.2. Let Λ be a row-finite k-graph. Let {Tλ : λ ∈ Λ} be a Toeplitz-Cuntz-
Krieger Λ-family in a C ∗-algebra B. Suppose that for every v ∈ Λ0,
(where this includes Tv 6= 0 if vΛ1 = ∅). Suppose that φT : T C ∗ (Λ) → B is the homo-
morphism such that φT (tλ) = Tλ for λ ∈ Λ. Then φT : T C ∗ (Λ) → B is injective.
Remark 2.3. Every k-graph Λ gives a product system of graphs over Nk and a Toeplitz-
Cuntz-Krieger Λ-family gives a Toeplitz Λ-family of the product system [8, Lemma 9.2].
Lemma 9.3 of [8] shows that, if the Toeplitz-Cuntz-Krieger Λ-family satisfies (*), then the
Toeplitz Λ-family satisfies the hypothesis of [8, Theorem 8.1].
Remark 2.4. In the actual hypothesis, we need to verify whetherQ1≤i≤k(Tv−Pe∈Gi
e ) 6=
0 for every v ∈ Λ0, 1 ≤ i ≤ k, and finite set Gi ⊆ vΛei. However, since we only consider
row-finite k-graphs, then for every v ∈ Λ0 and 1 ≤ i ≤ k, the set vΛei is finite. Thus for
a row finite k-graph, we can simplify Lemma 9.3 of [8] as Theorem 2.2.
TeT ∗
On the other hand, the Cuntz-Krieger uniqueness theorem only holds for k-graphs
satisfying a special Condition (B) [10]. Later, Lewin and Sims in [5, Proposition 3.6]
proved that Condition (B) is equivalent to the following aperiodicity condition: for every
pair of distinct paths λ, µ ∈ Λ with s (λ) = s (µ), there exists η ∈ s (λ) Λ such that
MCE (λη, µη) = ∅ [5, Definition 3.1]. (For discussion about the equivalence of various
aperiodicity definitions, see [5, 11, 12, 13].) Therefore we get the following theorem.
Theorem 2.5 ([10, Theorem 4.5]). Suppose that Λ is aperiodic row-finite k-graph and
{Sλ : λ ∈ Λ} is a Cuntz-Krieger Λ-family in a C ∗-algebra B such that Sv 6= 0 for v ∈ Λ0.
Suppose that πS : C ∗ (Λ) → B is the homomorphism such that πS (sλ) = Sλ for λ ∈ Λ for
λ ∈ Λ. Then πS is an injective homomorphism.
3. The k-graph T Λ
Suppose that Λ is a row-finite k-graph Λ. In this section, we define a k-graph T Λ; later
we show that T C ∗ (Λ) ∼= C ∗ (T Λ) (Theorem 4.1). Interestingly, our k-graph T Λ is always
aperiodic (Proposition 3.4).
Proposition 3.1. Let Λ = (Λ, d, r, s) be a row-finite k-graph. Then define sets T Λ0 and
T Λ as follows:
T Λ0 :=(cid:8)α (v) : v ∈ Λ0(cid:9) ∪(cid:8)β (v) : vΛ1 6= ∅(cid:9) ;
T Λ := {α (λ) : λ ∈ Λ} ∪(cid:8)β (λ) : λ ∈ Λ, s (λ) Λ1 6= ∅(cid:9) .
Define functions r, s : T Λ\T Λ0 → T Λ0 by
r (α (λ)) = α (r (λ)), s (α (λ)) = α (s (λ)) ,
r (β (λ)) = α (r (λ)), s (β (λ)) = β (s (λ))
(r, s are the identity on T Λ0). We also define a partially defined product (τ , ω) 7→ τ ω
from
{(τ , ω) ∈ T Λ × T Λ : s (τ ) = r (ω)}
4
YOSAFAT E. P. PANGALELA
to T Λ, where
(α (λ) , α (µ)) 7→ α (λµ)
(α (λ) , β (µ)) 7→ β (λµ)
and a function d : T Λ → Nk where
d (α (λ)) = d (β (λ)) = d (λ) .
Then (T Λ, d) is a k-graph.
Proof. First we claim that T Λ is a countable category. Note that T Λ is a countable since
Λ is countable.
Now we show that for all paths η, τ , ω in T Λ where s (η) = r (τ ) and s (τ ) = r (ω), we
have s (τ ω) = s (ω), r (τ ω) = r (τ ), and (ητ ) ω = η (τ ω). If one of τ , ω is a vertex then
we are done. So assume otherwise, and we have η = α (λ), τ = α (µ), and ω is either
α (ν) or β (ν) for some paths λ, µ, ν in Λ. In both cases, we always have s (λ) = r (µ),
s (µ) = r (ν), and (λµ) ν = λ (µν). If ω = α (ν), we have
s (τ ω) = s (α (µ) α (ν)) = s (α (µν)) = α (s (µν)) = α (s (ν)) = s (α (ν)) = s (ω) ,
r (τ ω) = r (α (µ) α (ν)) = r (α (µν)) = α (r (µν)) = α (r (µ)) = r (α (µ)) = r (τ ) , and
(ητ ) ω =(cid:0)α (λ) α (µ)(cid:1)α (ν) = α (λµ) α (ν) = α ((λµ) ν)
= α (λ (µν)) = α (λ) α (µν) = α (λ)(cid:0)α (µ) α (ν)(cid:1) = η (τ ω) .
On the other hand, if ω = β (ν), then
s (τ ω) = s (α (µ) β (ν)) = s (β (µν)) = β (s (µν)) = β (s (ν)) = s (β (ν)) = s (ω) ,
r (τ ω) = r (α (µ) β (ν)) = r (β (µν)) = α (r (µν)) = α (r (µ)) = r (α (µ)) = r (τ ) , and
(ητ ) ω =(cid:0)α (λ) α (µ)(cid:1)β (ν) = α (λµ) β (ν) = β ((λµ) ν)
= β (λ (µν)) = α (λ) β (µν) = α (λ)(cid:0)α (µ) β (ν)(cid:1) = η (τ ω) .
Thus, T Λ is a countable category, as claimed.
Now we show that d is a functor. Note that both T Λ and Nk are categories. First
take object x ∈ T Λ0, then d (x) = 0 is an object in category Nk. Next take morphisms
τ , ω ∈ T Λ with s (τ ) = r (ω). Then by definition of d,
d (τ ω) = d (τ ) + d (ω) .
Hence, d is a functor.
To show that d satisfies the factorisation property, take ω ∈ T Λ and m, n ∈ Nk such
that d (ω) = m + n. By definition, ω is either α (λ) or β (λ) for some path λ in Λ. In
both cases, there exist paths µ, ν in Λ such that λ = µν, d (µ) = m, and d (ν) = n. Then,
we have d (α (µ)) = m, d (α (ν)) = d (β (ν)) = n, and ω is either equal to α (µ) α (ν) or
α (µ) β (ν). Therefore, the existence of factorisation is guaranteed.
Now we show that the factorisation is unique. First suppose ω = α (µ) α (ν) =
α (µ′) α (ν′) where d (α (µ)) = d (α (µ′)) and d (α (ν)) = d (α (ν ′)). We consider paths
λ = µν and λ′ = µ′ν′. Since α (λ) = ω = α (λ′), then λ = λ′. This implies µ = µ′
and ν = ν′ based on the uniquness of factorisation in Λ. Then α (µ) = α (µ′) and
α (ν) = α (ν ′). For the case ω = α (µ) β (ν), we get the same result by using the same
argument. The conclusion follows.
(cid:3)
THE TOEPLITZ ALGEBRA OF A HIGHER-RANK GRAPH
5
Remark 3.2. For a directed graph E (that is, for k = 1), the graph T E was constructed
by Muhly and Tomforde [6, Definition 3.6] (denoted EV ), and by Sims [14, Section 3]
(denoted eE). Our notation follows that of Sims because we want to distinguish between
paths in T Λ (denoted α (λ) and β (λ)) and those in Λ (denoted λ).
Remark 3.3. Every vertex β (v) satisfies β (v) T Λ1 = ∅. Then if Λ has a vertex v which
receives edges e, f with d (e) 6= d (f ), then there is no edge g ∈ β (s (e)) Λd(f ) (or g ∈
α (s (e)) Λd(f ) if s (e) Λ = ∅), and hence Λ is not locally convex.
The following lemma tells about properties of the k-graph T Λ.
Proposition 3.4. Let Λ be a row-finite k-graph and T Λ be the k-graph as in Proposition
3.1. Then,
(a) T Λ is row-finite.
(b) T Λ is aperiodic.
Proof. To show part (a), take x ∈ T Λ0. If x = β (v) for some v ∈ Λ0, then xT Λ1 = ∅ by
Remark 3.3. Suppose x = α (v) for some v ∈ Λ0. If vΛ1 = ∅, then xT Λ1 = ∅. Otherwise,
for 1 ≤ i ≤ k such that vΛei 6= ∅, we have
xT Λei ≤ 2 vΛei ,
which is finite.
For part (b), take τ , ω ∈ T Λ such that τ 6= ω and s (τ ) = s (ω). We have to show
there exists η ∈ s (τ ) T Λ such that MCE (τ η, ωη) = ∅. If s (τ ) = β (v) for some v ∈ Λ0,
then choose η = β (v) and MCE (τ η, ωη) = ∅. So suppose s (τ ) = α (v) for some v ∈ Λ0.
If vΛ1 = ∅, then choose η = α (v) and MCE (τ η, ωη) = ∅. Suppose vΛ1 6= ∅. Take
e ∈ vΛ1.
If s (e) Λ1 = ∅, then choose η = α (e) and MCE (τ η, ωη) = ∅. Otherwise,
we have s (e) Λ1 6= ∅. Then choose η = β (e) and MCE (τ η, ωη) = ∅. Hence, T Λ is
aperiodic.
(cid:3)
4. Realising T C ∗ (Λ) as a Cuntz-Krieger algebra
Let Λ be a row-finite k-graph and T Λ be the k-graph as in Lemma 3.1. In this Section,
we show that T C ∗ (Λ) is isomorphic to C ∗ (T Λ).
Theorem 4.1. Let Λ be a row-finite k-graph and T Λ be the k-graph as in Proposition
3.1. Let {tλ : λ ∈ Λ} be the universal Toeplitz-Cuntz-Krieger Λ-family and {sω : ω ∈ T Λ}
be the universal Cuntz-Krieger T Λ-family. For λ ∈ Λ, let
Tλ :=(sα(λ) + sβ(λ)
sα(λ)
if s (λ) Λ1 6= ∅
if s (λ) Λ1 = ∅.
Then there is an isomorphism φT : T C ∗ (Λ) → C ∗(T Λ) satisfying φT (tλ) = Tλ for every
λ ∈ Λ.
Furthermore, sα(λ) = φT (tλ) if s (λ) Λ1 = ∅. Meanwhile, if s (λ) Λ1 6= ∅, we have
sα(λ) = φT(cid:0)tλ − tλQe∈vΛ1(tv − tet∗
e)(cid:1) and sβ(λ) = φT(cid:0)tλQe∈vΛ1(tv − tet∗
e)(cid:1).
Proof that {Tλ : λ ∈ Λ} is a Toeplitz-Cuntz-Krieger Λ-family. To avoid an argument by
cases, for λ ∈ Λ with s (λ) Λ1 = ∅, we write sβ(λ) := 0, so that
Tλ = sα(λ) + sβ(λ).
6
YOSAFAT E. P. PANGALELA
First, we want to show {Tλ : λ ∈ Λ} is a Toeplitz-Cuntz-Krieger Λ-family in C ∗(T Λ).
For (TCK1), take v ∈ Λ0. Since (cid:8)sα(v)(cid:9) ∪(cid:8)sβ(v)(cid:9) are mutually orthogonal projections,
then Tv is a projection. Meanwhile, for v, w ∈ Λ0 with v 6= w,
TvTw = sα(v)sα(w) + sα(v)sβ(w) + sβ(v)sα(w) + sβ(v)sβ(w) = 0.
Next we show (TCK2). Take µ, ν ∈ Λ where s (µ) = r (ν). Then
TµTν = sα(µ)sα(ν) + sα(µ)sβ(ν) + sβ(µ)sα(ν) + sβ(µ)sβ(ν).
If ν is a vertex, the middle terms vanish and we get
as required. Otherwise, the last two terms vanish and we get
TµTν = sα(µ) + sβ(µ) = Tµ,
TµTν = sα(µ)sα(ν) + sα(µ)sβ(ν) = sα(µν) + sβ(µν) = Tµν,
which is (TCK2).
To show (TCK3), take λ, µ ∈ Λ. Then
α(λ)sα(µ) + s∗
T ∗
λ Tµ = s∗
(4.1)
α(λ)sβ(µ) + s∗
β(λ)sα(µ) + s∗
β(λ)sβ(µ).
We give separate arguments for Λmin (λ, µ) = ∅ and Λmin (λ, µ) 6= ∅. For case Λmin (λ, µ) =
∅, we have
∅ = T Λmin (α (λ) , α (µ)) = T Λmin (α (λ) , β (µ))
= T Λmin (β (λ) , α (µ)) = T Λmin (β (λ) , β (µ)) .
Hence, s∗
α(λ)sα(µ) = s∗
α(λ)sβ(µ) = s∗
β(λ)sα(µ) = s∗
β(λ)sβ(µ) = 0 and then Equation 4.1 becomes
T ∗
λ Tµ = 0 = X(λ′,µ′)∈Λmin(λ,µ)
Tλ′T ∗
µ′.
Now suppose Λmin (λ, µ) 6= ∅. Take (a, b) ∈ Λmin (λ, µ). We consider several cases:
whether a equals s (λ) and/or b equals s (µ). First suppose a = s (λ) and b = s (µ). So
λ = λs (λ) = µs (µ) = µ. Because α (λ) and β (λ) are paths with the same degree and
different sources, then T Λmin (α (λ) , β (λ)) = ∅. Thus,
s∗
β(λ)sα(λ) = 0 = s∗
α(λ)sβ(λ)
and Equation 4.1 becomes
T ∗
λ Tλ = s∗
α(λ)sα(λ) + s∗
β(λ)sβ(λ)
= ss(α(λ)) + ss(β(λ)) = sα(s(λ)) + sβ(s(λ))
= Ts(λ) = Ts(λ)T ∗
s(λ)
= X(λ′,µ′)∈Λmin(λ,λ)
Tλ′T ∗
µ′ (since Λmin (λ, λ) = s (λ) ).
Next suppose a = s (λ) and b 6= s (µ). Then λ = µb and
T Λmin (α (λ) , β (µ)) = ∅ = T Λmin (β (λ) , β (µ))
since s (β (µ)) T Λ1 = ∅. Hence
α(λ)sβ(µ) = 0 = s∗
s∗
β(λ)sβ(µ)
THE TOEPLITZ ALGEBRA OF A HIGHER-RANK GRAPH
7
and Equation 4.1 becomes
T ∗
λ Tµ = s∗
α(λ)sα(µ) + s∗
β(λ)sα(µ).
Every (α (s (λ)) , η) ∈ T Λmin (α (λ) , α (µ)) has η = α (µ′) with (s (λ) , µ′) ∈ Λmin (λ, µ).
Similarly, every (β (s (λ)) , η) ∈ T Λmin (β (λ) , α (µ)) has η = β (µ′) with (s (λ) , µ′) ∈
Λmin (λ, µ). Thus, by using (TCK3) in C ∗ (T Λ),
T ∗
λ Tµ = s∗
α(λ)sα(µ) + s∗
β(λ)sα(µ)
η +
sα(s(λ))s∗
X
α(µ′) + X(s(λ),µ′)∈Λmin(λ,µ)
(β(s(λ)),η)∈T Λmin(β(λ),α(µ))
sβ(s(λ))s∗
β(µ′)
sβ(s(λ))s∗
η
(α(s(λ)),η)∈T Λmin(α(λ),α(µ))
sα(s(λ))s∗
=
X
= X(s(λ),µ′)∈Λmin(λ,µ)
= X(s(λ),µ′)∈Λmin(λ,µ)
= X(s(λ),µ′)∈Λmin(λ,µ)
= X(s(λ),µ′)∈Λmin(λ,µ)
(sα(s(λ))s∗
α(µ′) + sβ(s(λ))s∗
β(µ′))
(sα(s(λ)) + sβ(s(λ)))(s∗
α(µ′) + s∗
β(µ′))
Ts(λ)T ∗
µ′ = X(λ′,µ′)∈Λmin(λ,µ)
Tλ′T ∗
µ′.
By taking adjoints, we deduce (TCK3) when a 6= s (λ) and b = s (µ).
Now we consider the last case, which is a 6= s (λ) and b 6= s (µ). This means we have
neither λ = µb nor µ = λa. Hence,
T Λmin (α (λ) , β (µ)) = T Λmin (β (λ) , α (µ)) = T Λmin (β (λ) , β (µ)) = ∅
since s (β (λ)) T Λ1 = ∅ = s (β (µ)) T Λ1 = ∅. Hence,
β(λ)sα(µ) = s∗
s∗
α(λ)sβ(µ) = s∗
β(λ)sβ(µ) = 0.
On the other hand, we have
T Λmin (α (λ) , α (µ)) =(cid:8)(α (λ′) , α (µ′)) , (β (λ′) , β (µ′)) : (λ′, µ′) ∈ Λmin (λ, µ)(cid:9) .
Therefore, Equation 4.1 becomes
T ∗
λ Tµ = s∗
α(λ)sα(µ) =
= X(λ′,µ′)∈Λmin(λ,µ)
= X(λ′,µ′)∈Λmin(λ,µ)
= X(λ′,µ′)∈Λmin(λ,µ)
X(ω,η)∈T Λmin(α(λ),α(µ))
sωs∗
η
(sα(λ′)s∗
α(µ′) + sβ(λ′)s∗
β(µ′))
(sα(λ′) + sβ(λ′))(s∗
α(µ′) + s∗
β(µ′))
Tλ′T ∗
µ′.
T ∗
λ Tµ = X(λ′,µ′)∈Λmin(λ,µ)
Tλ′T ∗
µ′
So for all cases, we have
and {Tλ : λ ∈ Λ} satisfies (TCK3).
(cid:3)
8
YOSAFAT E. P. PANGALELA
Proof that φT is injective. Now the universal property of T C ∗ (Λ) gives a homomorphism
φT : T C ∗ (Λ) → C ∗(T Λ) satisfying φT (tλ) = Tλ for every λ ∈ Λ.
We show the injectivity of φT by using Theorem 2.2. Take v ∈ Λ0. We show
First suppose vΛ1 6= ∅. Take 1 ≤ i ≤ k such that vΛei 6= ∅. We claim
(Tv − TeT ∗
e ) 6= 0.
(Tv − TeT ∗
e ) ≥ sβ(v).
Since vΛei 6= ∅, then α (v) T Λei 6= ∅ and by [10, Lemma 2.7 (iii)],
sgs∗
g
sα(e)s∗
Ye∈vΛ1
Ye∈vΛei
sα(v) ≥ Xg∈α(v)T Λei
= Xe∈vΛei
= Xe∈vΛei
s(e)Λ16=∅(cid:0)sα(e)s∗
= Xe∈vΛei
= Xe∈vΛei
Ye∈vΛei
(Tv − TeT ∗
TeT ∗
e .
TeT ∗
s(e)Λ16=∅
α(e) + Xe∈vΛei
s(e)Λ16=∅
sβ(e)s∗
β(e)
α(e) + sβ(e)s∗
sα(e)s∗
α(e)
β(e)(cid:1) + Xe∈vΛei
s(e)Λ1=∅
e + Xe∈vΛei
s(e)Λ1=∅
TeT ∗
e
TeT ∗
e
e ) = Tv − Xe∈vΛei
=(cid:0)sα(v) + sβ(v)(cid:1) − Xe∈vΛei
= sβ(v) +(cid:16)sα(v) − Xe∈vΛei
≥ sβ(v),
TeT ∗
e
TeT ∗
e(cid:17)
Meanwhile, since every e ∈ vΛei has the same degree,
as claimed. This claim implies
Ye∈vΛ1
since vΛ1 6= ∅, as required.
(Tv − TeT ∗
e ) ≥ Y{i:vΛei 6=∅}
sβ(v) = sβ(v) 6= 0
Finally, for v ∈ Λ0 with vΛ1 = ∅, we have
Tv = sα(v) 6= 0.
Hence, by Theorem 2.2, φT is injective.
Proof that φT is surjective. Now we show the surjectivity of φT . Since C ∗(T Λ) is gener-
ated by {sτ : τ ∈ T Λ}, then it suffices to show that for every τ ∈ T Λ, sτ ∈ im (φT ). Recall
that for every τ ∈ T Λ, sτ is either sα(λ) or sβ(λ) for some λ ∈ Λ.
(cid:3)
THE TOEPLITZ ALGEBRA OF A HIGHER-RANK GRAPH
9
Take v ∈ Λ0. First we show sα(v) and sβ(v) (if it exists) belong to im (φT ). If vΛ1 = ∅,
then
Next suppose vΛ1 6= ∅. First we show that sβ(v) = Qe∈vΛ1 (Tv − TeT ∗
every f ∈ α (v) T Λ1, the projection sα(v) − sf s∗
e ). Note that for
f ≤ sα(v) is othogonal to sβ(v). This implies
sα(v) = Tv ∈ im (φT ) .
((sα(v) + sβ(v)) − sf s∗
(sα(v) − sf s∗
f )
Yf ∈α(v)T Λ1
f ) = sβ(v) + Yf ∈α(v)T Λ1
= sβ(v),
since vΛ1 is an exhaustive set. Hence,
(Tv − sβ(e)s∗
β(e))
(Tv − sα(e)s∗
α(e))(Tv − sβ(e)s∗
β(e))
(Tv − (sα(e)s∗
α(e) + sβ(e)s∗
β(e)))
(Tv − TeT ∗
e )
((sα(v) + sβ(v)) − sf s∗
f )
s(e)Λ16=∅
s(e)Λ1=∅
s(e)Λ16=∅
(Tv − sα(e)s∗
(Tv − sα(e)s∗
(Tv − sα(e)s∗
α(e)) Ye∈vΛ1
α(e)) Ye∈vΛ1
α(e)) Ye∈vΛ1
e ) Ye∈vΛ1
sβ(v) = Yf ∈α(v)T Λ1
= Ye∈vΛ1
= Ye∈vΛ1
= Ye∈vΛ1
= Ye∈vΛ1
= Ye∈vΛ1
sα(v) = Tv − sβ(v) = Tv − Ye∈vΛ1
(Tv − TeT ∗
(Tv − TeT ∗
e ) ,
s(e)Λ1=∅
s(e)Λ16=∅
s(e)Λ1=∅
s(e)Λ16=∅
as required, and sβ(v) belongs to im (φT ). Furthermore,
(Tv − TeT ∗
e ) ∈ im (φT ) ,
as required.
Now take λ ∈ Λ. We have to show sα(λ) and sβ(λ) (if it exists) belong to im (φT ). If
s (λ) Λ1 = ∅, then
sα(λ) = sα(λ)sα(s(λ)) = TλTs(λ) = Tλ ∈ im (φT ) .
Next suppose s (λ) Λ1 6= ∅. Then sβ(λ)sα(s(λ)) = 0 and sα(λ)sβ(s(λ)) = 0. Hence,
(Ts(λ) − TeT ∗
sα(λ) = sα(λ)sα(s(λ)) =(cid:0)sα(λ) + sβ(λ)(cid:1) sα(s(λ))
e )(cid:17)
= Tλ(cid:16)Ts(λ) − Ye∈s(λ)Λ1
= Tλ − Tλ Ye∈s(λ)Λ1
sβ(λ) = sβ(λ)sβ(s(λ)) =(cid:0)sα(λ) + sβ(λ)(cid:1) sβ(s(λ))
(Ts(λ) − TeT ∗
e ) ∈ im (φT )
and
tλ
Sτ :=
tλ − tλQe∈vΛ1(tv − tet∗
tλQe∈vΛ1(tv − tet∗
e)
e)
if τ = α (λ) with s (λ) Λ1 = ∅
if τ = α (λ) with s (λ) Λ1 6= ∅
if τ = β (λ) with s (λ) Λ1 6= ∅.
10
YOSAFAT E. P. PANGALELA
= Tλ Ye∈s(λ)Λ1
(Ts(λ) − TeT ∗
e ) ∈ im (φT ) .
Therefore, φT is surjective and an isomorphism.
(cid:3)
Corollary 4.2. Let Λ be a row-finite k-graph and T Λ be the k-graph as in Proposition
3.1. Let {tλ : λ ∈ Λ} be the universal Toeplitz-Cuntz-Krieger Λ-family and {sω : ω ∈ T Λ}
be the universal Cuntz-Krieger T Λ-family. For τ ∈ T Λ, define
Suppose that φT : T C ∗ (Λ) → C ∗(T Λ) is the isomorphism as in Theorem 4.1 and πS :
C ∗ (T Λ) → T C ∗ (Λ) is the homomorphism such that πS (sτ ) = Sτ for τ ∈ T Λ. Then
φ−1
T = πS.
Proof. Take λ ∈ Λ. By Theorem 4.1, we get φ−1
if s (λ) Λ1 6= ∅, by Theorem 4.1, we have φ−1
φ−1
e). Hence, φ−1
{Sτ : τ ∈ T Λ} is a Cuntz-Krieger T Λ-family, and then φ−1
T (cid:0)sα(λ)(cid:1) = tλ if s (λ) Λ1 = ∅. Meanwhile,
T (cid:0)sα(λ)(cid:1) = tλ − tλQe∈vΛ1(tv − tet∗
e) and
T (sτ ) = Sτ for τ ∈ T Λ. This implies that
(cid:3)
T (cid:0)sβ(λ)(cid:1) = tλQe∈vΛ1(tv − tet∗
T = πS.
Remark 4.3. Proposition 3.4 says that T Λ is always aperiodic, and hence the Cuntz-
Krieger uniqueness theorem always applies to T Λ. This helps explain why no hypothesis
on Λ is required in the uniquness theorem of [8, Theorem 8.1]. Indeed, we could have
deduced that theorem by applying the Cuntz-Krieger uniqueness theorem to T Λ. With
our current proof of Theorem 4.1, this argument would be circular, since we used [8,
Theorem 8.1] in the proof of Theorem 4.1. However, we could prove Corollary 4.2 directly
by showing that {Sτ : τ ∈ T Λ} is a Cuntz-Krieger T Λ-family in T C ∗ (Λ), hence gives
a homomorphism πS : C ∗ (T Λ) → T C ∗ (Λ), and using the Cuntz-Krieger uniqueness
theorem to see that πS is injective. Then we could deduce [8, Theorem 8.1] from Corollary
4.2, and this would be a legitimate proof. We worked out the details of this approach, but
it seemed to require an extensive cases argument, and hence became substantially more
complicated.
References
[1] C. Farthing, P. S. Muhly and T. Yeend, Higher-rank graph C ∗-algebras: an inverse semigroup and
groupoid approach, Semigroup Forum 71 (2005), 159 -- 187.
[2] N. J. Fowler, M. Laca and I. Raeburn, The C ∗-algebras of infinite graphs, Proc. Amer. Math. Soc.
128 (2000), 2319 -- 2327.
[3] N. J. Fowler and I. Raeburn, The Toeplitz algebra of a Hilbert bimodule, Indiana Univ. Math. J. 48
(1999), 155 -- 181.
[4] A. Kumjian and D. Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20.
[5] P. Lewin and A. Sims, Aperiodicity and cofinality for finitely aligned higher-rank graphs, Math. Proc.
Cambridge Philos. Soc. 149 (2010), 333 -- 350.
[6] P. Muhly and M. Tomforde, Adding tails to C ∗-correspondences, Documenta Math. 9 (2004), 79 -- 106.
[7] I. Raeburn, Graph Algebras, CBMS Regional Conference Series in Math., vol. 103, American Math-
ematical Society, 2005.
[8] I. Raeburn and A. Sims, Product systems of graphs and the Toeplitz algebras of higher-rank graphs,
J. Operator Theory 53 (2005), 399 -- 429.
THE TOEPLITZ ALGEBRA OF A HIGHER-RANK GRAPH
11
[9] I. Raeburn, A. Sims and T. Yeend, Higher-rank graphs and their C ∗-algebras, Proc. Edinb. Math.
Soc. 46 (2003), 99 -- 115.
[10] I. Raeburn, A. Sims and T. Yeend, The C ∗-algebras of finitely aligned higher-rank graphs, J. Funct.
Anal. 213 (2004), 206 -- 240.
[11] D. Robertson and A. Sims, Simplicity of C ∗-algebras associated to higher-rank graphs, Bull. London
Math. Soc. 39 (2007), 337 -- 344.
[12] D. Robertson and A. Sims, Simplicity of C ∗-algebras associated to row-finite locally convex higher-
rank graphs, Israel J. Math. 172 (2009), 171 -- 192.
[13] J. Shotwell, Simplicity of finitely-aligned k-graph C ∗-algebras, J. Operator Theory 67 (2012), 335 --
347.
[14] A. Sims, The co-universal C ∗-algebra of a row-finite graph, New York J. Math. 16 (2010), 507 -- 524.
[15] S. B. G. Webster, The path space of a higher-rank graph, Studia Math., 204 (2011), 155 -- 185.
Department of Mathematics and Statistics, University of Otago, PO Box 56, Dunedin
9054, New Zealand
E-mail address: [email protected]
|
1709.08211 | 1 | 1709 | 2017-09-24T15:38:40 | Maximally unitarily mixed states on a C*-algebra | [
"math.OA"
] | We investigate the set of maximally mixed states of a C*-algebra, extending previous work by Alberti on von Neumann algebras. We show that, unlike for von Neumann algebras, the set of maximally mixed states of a C*-algebra may fail to be weak* closed. We obtain, however, a concrete description of the weak* closure of this set, in terms of tracial states and states which factor through simple traceless quotients. For C*-algebras with the Dixmier property or with Hausdorff primitive spectrum we are able to advance our investigations further. In the latter case we obtain a concrete description of the set of maximally mixed states in terms of traces and extensions of the states of a closed two-sided ideal. We pose several questions. | math.OA | math | MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
ROBERT ARCHBOLD, LEONEL ROBERT, AND AARON TIKUISIS
Abstract. We investigate the set of maximally mixed states of a C*-algebra, ex-
tending previous work by Alberti on von Neumann algebras. We show that, unlike
for von Neumann algebras, the set of maximally mixed states of a C*-algebra may fail
to be weak* closed. We obtain, however, a concrete description of the weak* closure
of this set, in terms of tracial states and states which factor through simple traceless
quotients. For C*-algebras with the Dixmier property or with Hausdorff primitive
spectrum we are able to advance our investigations further.
In the latter case we
obtain a concrete description of the set of maximally mixed states in terms of traces
and extensions of the states of a closed two-sided ideal. We pose several questions.
7
1
0
2
p
e
S
4
2
]
.
A
O
h
t
a
m
[
1
v
1
1
2
8
0
.
9
0
7
1
:
v
i
X
r
a
1. Introduction
Investigations into the entropy and irreversibility of the states of a physical system
lead to the consideration of the "more mixed than" or "more chaotic than" pre-order on
the space of states of a C*-algebra. This pre-order, first introduced by Uhlmann, has
been investigated for the state spaces of matrix algebras and, more generally, of von
Neumann algebras, by Alberti, Uhlmann, Wehrl, and others ([2, 3, 12, 13]). Uhlmann
also introduced a distinguished collection of states: the maximally mixed states. These
are the maximal elements in the "more mixed than" pre-order. In [1, Theorem 5.2],
Alberti gave a complete description of the maximally mixed states of a von Neumann
algebra. In this paper we undertake the study of the maximally mixed states of a C*-
algebra. In particular, we probe the extent to which Alberti's theorem can be extended
to arbitrary C*-algebras.
Let us be more specific. Let A be a C*-algebra. Given two states ϕ and ψ on A, let's
say that ψ is more (unitarily) mixed than ϕ if ψ belongs to the weak* closure of the
convex hull of the unitary conjugates of ϕ. A state ϕ is maximally (unitarily) mixed
if whenever ψ is more mixed than ϕ then ϕ is also more mixed than ψ. Maximally
mixed states are guaranteed to exist by weak* compactness and Zorn's lemma (in fact,
given any state ϕ there exist maximally mixed states that are more mixed than ϕ). We
denote the set of maximally mixed states of A by S∞(A).
The main question that we address here is "can the set S∞(A) be described more
concretely?". The tracial states on A are obviously maximally mixed. Another source of
maximally mixed states on A is the quotients A/M that are simple and have no bounded
traces. The states (on A) that factor through these quotients are also maximally mixed.
2010 Mathematics Subject Classification. 46L30.
Key words and phrases. States of C*-algebras, Unitary mixings, Dixmier property.
A.T. was partially supported by an NSERC Postdoctoral Fellowship and through the EPSRC grant
EP/N00874X/1.
1
2
ARCHBOLD, ROBERT, AND TIKUISIS
Alberti showed that if A is a von Neumann algebra then S∞(A) is the weak* closure of
the convex hull of the tracial states and the states that factor through simple traceless
quotients (see [1], though it is not quite stated this way). We demonstrate below with
natural examples that the set S∞(A) need not always be weak* closed. It is the case,
however, that the weak* closure of S∞(A) is precisely the weak* closure of the convex
hull of the tracial states and the states factoring through simple traceless quotients
(Theorem 3.10). We leave open the question of the convexity of S∞(A).
For C*-algebras with the Dixmier property we are able to advance our understanding
of S∞(A) further. Recall that A is said to have the Dixmier property if for every a ∈ A
the norm closure of the convex hull of the unitary conjugates of a intersects the center
of A. Von Neumann algebras have the Dixmier property (by Dixmier's approximation
theorem), but the class of C*-algebras with the Dixmier property is much larger (see
[5]). We show that if A has the Dixmier property then S∞(A) is convex and weakly
closed. Further, we obtain a necessary and sufficient condition (involving the primitive
spectrum) for S∞(A) to be weak* closed (Theorem 4.7).
The paper is organized as follows: In Section 2 we introduce notation. In Section 3 we
embark on the investigation of the maximally mixed states of an arbitrary C*-algebra.
In Section 4 we consider C*-algebras with the Dixmier property. In Section 5 we rely
on the results from the previous sections to obtain a concrete description of the set of
maximally mixed states of a C*-algebra with Hausdorff primitive spectrum.
2. Preliminaries on Dixmier sets
Let A be a C*-algebra. We denote by Asa the set of self-adjoint elements of A and
by A+ the set of positive elements of A. If A is unital we denote by U(A) the group of
unitary elements of A.
We denote by A∼ the minimal unitization of A, i.e., A itself if A is unital and the
unitization A + C1 if A is non-unital.
Let A∗ denote the dual of A. We denote by A∗
sa the set of self-adjoint functionals in
A∗ and by A∗
+ the set of positive functionals in A∗.
We denote the convex hull of a set S (in an affine space) by co(S).
2.1. Dixmier sets in A and A∗. We call a set C ⊆ A a Dixmier set if it is convex,
norm-closed, and invariant under unitary conjugation. The latter means that uCu∗ ⊆ C
for all unitaries u ∈ U(A∼). We will largely work with singly generated Dixmier sets.
Given a ∈ A we denote by DA(a) the smallest Dixmier set containing a.
We let A, and more generally M(A) (the multiplier algebra of A), act on A∗ in the
usual way: if a ∈ M(A) and ϕ ∈ A∗ then
aϕ(x) := ϕ(ax),
(ϕa)(x) := ϕ(xa)
(x ∈ A).
A set C ⊆ A∗ is called a Dixmier set if it is convex, weak* closed, and invariant
under unitary conjugation. The latter condition means that uCu∗ ⊆ C for all unitaries
u ∈ U(A∼). Given ϕ ∈ A∗ we denote by DA(ϕ) the Dixmier set generated by ϕ, i.e.,
the smallest Dixmier set containing ϕ. Since DA(ϕ) is weak* closed and bounded, it is
weak* compact.
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
3
We shall make frequent use of the standard fact that A is the dual of A∗ when the
latter is endowed with the weak* topology. This, combined with the Hahn-Banach
theorem, implies that elements of A separate disjoint weak* compact convex sets in A∗.
Let V be a subgroup of the unitary group U(M(A)) of M(A). On some occasions
we will need more general versions of the sets defined above where the unitaries range
through V rather than U(A∼). Thus, given a ∈ A we define DA(a, V) as the small-
est norm-closed convex subset of A containing a and invariant under conjugation by
unitaries in V. Similarly, given ϕ ∈ A∗ we define DA(ϕ, V) as the the smallest weak*
closed convex subset of A∗ containing ϕ and invariant under conjugation by unitaries
in V.
2.2. Mixing operators. Let V be a subgroup of the unitary group U(M(A)) of M(A).
We call a linear operator T : A → A a V-mixing operator if it is defined by an equation
of the form
T a =
λjujau∗
j
(a ∈ A),
nXj=1
where n ∈ N, λj > 0, uj ∈ V (1 6 j 6 n), and Pn
j=1 λj = 1. Elementary properties of
such operators are described in [4, 2.2]. We denote by Mix(A, V) the set of V-mixing
operators on A. If V = U(A∼) we simply write Mix(A). Notice that
DA(a, V) = {T a : T ∈ Mix(A, V)}
k·k
.
We also call an operator T : A∗ → A∗ a V-mixing operator if it is the adjoint of a
V-mixing operator on A. In this case T has the form
T ϕ =
nXj=1
λjujϕu∗
j
(ϕ ∈ A∗),
where n ∈ N, λj > 0, uj ∈ V (1 6 j 6 n), and Pn
j=1 λj = 1. Observe that T is positive
(T ϕ > 0 for all ϕ > 0) and contractive. We denote the set of V-mixing operators on
A∗ by Mix(A∗, V) or simply by Mix(A∗) if V = U(A∼). Notice that
DA(ϕ, V) = {T ϕ : T ∈ Mix(A∗, V)}
weak∗
.
Lemma 2.1. Let a ∈ A and ϕ ∈ A∗. Then
(2.1)
DA(ϕ, V)(a) = ϕ(DA(a, V)).
Proof. Since DA(ϕ, V) is weak∗-compact, DA(ϕ, V)(a) is a closed subset of C. To prove
the lemma it suffices to show that ϕ(DA(a, V)) is a dense subset of DA(ϕ, V)(a). Let
T ∈ Mix(A, V). Then (T ∗ϕ)(a) = ϕ(T a). Letting T range through all Mix(A, V) the
left side is dense in DA(ϕ, V)(a) while the right side is dense in ϕ(DA(a, V)).
(cid:3)
We will find it convenient to work with more general unitary mixing operators on
A∗. We let Mix(A∗, V) denote the closure of Mix(A∗, V) in the point-weak∗ topology on
B(A∗) (the bounded linear operators on A∗). If V = U(A∼) we simply write Mix(A∗).
Since a limit in the point-weak∗ topology of positive contractions is again a positive
contraction, all T ∈ Mix(A∗, V) are positive contractions. Since the unit ball of B(A∗)
is compact in the point-weak∗ topology, Mix(A∗, V) is a compact set in this topology.
4
ARCHBOLD, ROBERT, AND TIKUISIS
Lemma 2.2. Let ϕ ∈ A∗. Then DA(ϕ, V) = {T ϕ : T ∈ Mix(A∗, V)}.
Proof. Clearly, T ϕ ∈ DA(ϕ, V) for all T ∈ Mix(A∗, V). Suppose that ψ ∈ DA(ϕ, V).
Then Tiϕ → ψ in the weak* topology for some net of V-mixing operators (Ti)i on A∗.
Passing to a subnet of (Ti)i convergent in the point-weak* topology we get that ψ = T ϕ
for some T ∈ Mix(A∗, V).
(cid:3)
3. Maximally mixed functionals
Let ϕ ∈ A∗. If ψ ∈ DA(ϕ) we say that ψ is more unitarily mixed than ϕ. We say
that ϕ is maximally (unitarily) mixed if DA(ϕ) is minimal with respect to the order by
inclusion in the lattice of weak∗-compact Dixmier subsets of A∗. Thus ϕ is maximally
mixed if and only if for all ψ ∈ DA(ϕ) we have DA(ψ) = DA(ϕ).
It follows from Zorn's lemma that any weak* compact Dixmier set contains a max-
imally mixed functional. In particular, DA(ϕ) contains a maximally mixed functional
for all ϕ ∈ A∗. Note also that (i) the zero functional is maximally mixed, (ii) if ϕ is
tracial then DA(ϕ) = {ϕ} and hence ϕ is maximally mixed, and (iii) if ϕ is maximally
mixed and λ ∈ C then λϕ is maximally mixed.
Theorem 3.1. Let A be a C*-algebra and let ϕ ∈ A∗ be maximally mixed. Then the
self-adjoint and skew-adjoint parts of ϕ are maximally mixed. If ϕ is self-adjoint, then
its positive and negative parts are maximally mixed.
Proof. Let ϕsa denote the self-adjoint part of ϕ. Let ψ ∈ DA(ϕsa). Then ψ = T ϕsa
for some T ∈ Mix(A∗) (Lemma 2.2). Mixing operators in Mix(A∗) preserve the self-
adjoint part. So ψ is the self-adjoint part of T ϕ. Since ϕ is maximally mixed and
T ϕ ∈ DA(ϕ), there exists S ∈ Mix(A∗) such that ST ϕ = ϕ. Taking self-adjoint parts
we get Sψ = ϕsa. Thus, ϕsa ∈ DA(ψ), as desired. The same argument applies to the
skew-adjoint part.
Suppose now that ϕ is self-adjoint (and maximally mixed). Let us show first that
(T ϕ)+ = T ϕ+ and (T ϕ)− = T ϕ− for any T ∈ Mix(A∗). Observe that kψk 6 kϕk for
all ψ ∈ DA(ϕ). But, since ϕ is maximally mixed, we must have that kψk = kϕk for all
ψ ∈ DA(ϕ). That is, all the functionals in DA(ϕ) have the same norm. Applying T on
both sides of ϕ = ϕ+ − ϕ− we get T ϕ = T ϕ+ − T ϕ−. Then,
kT ϕ+k + kT ϕ−k 6 kϕ+k + kϕ−k = kϕk = kT ϕk.
It follows that T ϕ+ and T ϕ− are orthogonal ([11, Lemma 3.2.3]). By the uniqueness
of the Jordan decomposition ([11, Theorem 3.2.5]), (T ϕ)+ = T ϕ+ and (T ϕ)− = T ϕ−.
That ϕ+ and ϕ− are maximally mixed is now straightforward. For suppose that
ψ ∈ DA(ϕ+). By Lemma 2.2, there exists T ∈ Mix(A∗) such that ψ = T ϕ+. Further,
since ϕ is maximally mixed, there exists S ∈ Mix(A∗) such that ST ϕ = ϕ. Then
Sψ = ST ϕ+ = (ST ϕ)+ = ϕ+. Thus, ϕ+ is maximally mixed. The same argument
shows that ϕ− is maximally mixed.
(cid:3)
Due in part to the previous theorem, in the sequel our focus will be on the positive
maximally mixed functionals. We warn however that it is not true that a self-adjoint
functional whose positive and negative parts are maximally mixed is itself maximally
mixed: see Example 4.10.
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
5
Theorem 3.2. Let A be a C*-algebra. The set of maximally mixed functionals is a
norm-closed subset of A∗.
Proof. Let ϕ ∈ A∗ be in the norm-closure of the set of maximally mixed functionals.
Let ψ ∈ DA(ϕ). By Lemma 2.2, there exists T ∈ Mix(A∗) such that ψ = T ϕ. Let
ε > 0. Then there exists a maximally mixed ϕ such that kϕ − ϕk < ε. Since T is a
contraction,
kψ − T ϕk = kT ϕ − T ϕk 6 kϕ − ϕk < ε.
Since ϕ is maximally mixed, there exists S ∈ Mix(A∗) such that ST ϕ = ϕ. Then,
kSψ − ϕk = kSψ − ST ϕk 6 kψ − T ϕk < ε.
So kϕ − Sψk < 2ε. Since DA(ψ) is norm-closed, we have ϕ ∈ DA(ψ) and hence
DA(ψ) = DA(ϕ). Thus, ϕ is maximally mixed.
(cid:3)
We will show in Examples 3.16 and 3.17 that the set of maximally mixed functionals
is not always weak* closed. We do have the following:
Proposition 3.3. Let A be a unital C*-algebra and let ϕ ∈ A∗
+.
(i) Suppose that for every a ∈ Asa and ε > 0 there exists a maximally mixed ϕ′ ∈ A∗
+
such that ϕ′ 6 ϕ and ϕ(a) − ϕ′(a) < ε. Then ϕ is maximally mixed.
(ii) Suppose that for every a ∈ Asa and ε > 0 there exists a maximally mixed ϕ′ ∈ A∗
+
such that ϕ′ > ϕ and ϕ(a) − ϕ′(a) < ε. Then ϕ is maximally mixed.
(iii) Suppose that (ϕi)i is a norm-bounded net of maximally mixed functionals in A∗
+
which is either upward directed or downward directed relative to the order in A∗
+.
Then the net is convergent and the limit is maximally mixed.
Proof. (i) Let ψ ∈ DA(ϕ) and suppose that ψ = T ϕ, where T ∈ Mix(A∗). Suppose,
towards a contradiction, that ϕ /∈ DA(ψ). Then by the Hahn-Banach theorem there
exist a ∈ Asa, t ∈ R and ε > 0 such that ρ(a) 6 t for all ρ ∈ DA(ψ) but ϕ(a) > t + ε.
Replacing a by a + kak1 and t by t + kakkϕk, we may assume that a > 0.
By hypothesis, there exists a maximally mixed functional ϕ′ ∈ A∗
+ such that ϕ′ 6 ϕ
and ϕ′(a) > t + ε/2. Let ψ′ = T ϕ′. Note that, since T is positive, ψ′ 6 ψ. Since ϕ′ is
maximally mixed, ϕ′ ∈ DA(ψ′). Thus, there exists S ∈ Mix(A∗) such that Sψ′ = ϕ′.
Let ρ = Sψ. Then ϕ′ 6 ρ and so ϕ′(a) 6 ρ(a) 6 t since a > 0. This contradicts the
fact that ϕ′(a) > t + ε/2. Thus ϕ ∈ DA(ψ) and hence DA(ψ) = DA(ϕ).
(ii) This is similar to (i).
(iii) The convergence of the net follows from weak∗-compactness, monotonicity and
the fact that A is the linear span of A+. The limit is maximally mixed by (i) and
(ii).
(cid:3)
Next we prepare to examine the relation of the maximally mixed functionals of A
with those of its ideals and quotients. Theorem 3.6 will tell us that, given an ideal J of
A, maximal mixedness of a functional can be read off by its decomposition with respect
to A/J and J. Part (i) of the following proposition is a classical key result used to
prove permanence of the Dixmier property under suitable extensions; we use part (ii)
in an analogous way to handle Dixmier sets of functionals.
6
ARCHBOLD, ROBERT, AND TIKUISIS
Proposition 3.4. Let A be a C*-algebra, let a ∈ A and let ϕ ∈ A∗. The following are
true:
(i) DA(a) is equal to the norm-closure of co{eihae−ih : h ∈ Asa}.
(ii) DA(ϕ) is equal to the weak* closure of co{eihϕe−ih : h ∈ Asa}.
Proof. (i) For unital A, the result is given in [4, Proposition 2.4]. For non-unital A, we
apply this result to A∼ and use the fact that if h ∈ Asa and t ∈ R then ei(h+t1) = eiteih.
(ii) This follows from (i) and the Hahn-Banach theorem. Indeed, if (ii) fails to hold
then there is a unitary conjugate of ϕ which does not belong to the weak* closure of
co{eihϕe−ih : h ∈ Asa}. Since A∗ with the weak∗-topology has dual space A, it follows
by the Hahn-Banach separation theorem that there exists u ∈ U(A∼), a ∈ A and t ∈ R
such that Re(ϕ(uau∗)) > t and Re(ϕ(eihae−ih)) 6 t for all h ∈ Asa. It follows from the
last inequality and part (i) that Re(ϕ(x)) 6 t for all x ∈ DA(a). This contradicts the
fact that Re(ϕ(uau∗)) > t.
(cid:3)
Proposition 3.5. Let J be a proper, closed two-sided ideal of a unital C*-algebra A.
Let ιJ : J → A and qJ : A → A/J denote the inclusion and quotient maps.
(i) The adjoint map ι∗
(ii) We have DA(ϕ) = DA(ϕ, U(J + C1)) for all ϕ ∈ A∗
(iii) The adjoint map q∗
J : A∗ → J ∗ maps DA(ϕ) onto DJ (ϕJ ) for all ϕ ∈ A∗
+.
J : (A/J)∗ → A∗ maps DA/J (ϕ) bijectively to DA(ϕ ◦ qJ ) for all
+ such that kϕk = kϕJk.
ϕ ∈ (A/J)∗
+.
Proof. If the ideal J is a unital C*-algebra then A ∼= J ⊕ A/J and all three results
(i)-(iii) have a straightforward proof. We thus assume that J is non-unital. Note then
that J + C1 may be regarded as the unitization of J.
(i) Let us first show that ρ
ι∗
J
7−→ ρJ maps DA(ϕ) into DJ (ϕJ). Let ψ ∈ DA(ϕ) and
suppose that ψJ /∈ DJ (ϕJ). Then, by the Hahn-Banach theorem, there exist a ∈ Jsa
and t ∈ R such that ψ(a) > t and ρ(a) 6 t for all ρ ∈ DJ (ϕJ ). It follows from Lemma
2.1 applied to ϕJ and a that ϕ(b) 6 t for all b ∈ DJ (a). But, by [4, Remark 2.6],
DJ (a) = DA(a) (since a ∈ J). Hence ϕ(b) 6 t for all b ∈ DA(a). Lemma 2.1, applied
now to ϕ and a, implies that ρ(a) 6 t for all ρ ∈ DA(ϕ). Since ψ ∈ DA(ϕ), we obtain
that ψ(a) 6 t which gives a contradiction. Thus ι∗
J maps DA(ϕ) into DJ (ϕJ).
Let us prove surjectivity. Since ι∗
J is weak∗-continuous, the image of DA(ϕ) is a
weak∗-compact convex subset of DJ (ϕJ). For every T ∈ Mix(A, U(J + C1)) we have
(ϕ ◦ T )J = ϕJ ◦ T J. Clearly, every mixing operator in Mix(J) has the form T J for
some T ∈ Mix(A, U(J + C1)). Thus, letting T range through Mix(A, U(J + C1)) the
functionals ϕJ ◦ T J range through a dense subset of DJ (ϕJ ). This shows that the
image of DA(ϕ) by ι∗
J is also dense in DJ (ϕJ).
(ii) Clearly DA(ϕ, U(J + C1)) ⊆ DA(ϕ). To prove the opposite inclusion it suffices
to show that uϕu∗ ∈ DA(ϕ, U(J + C1)) for all u ∈ U(A). Let u ∈ U(A) and set
ψ = uϕu∗. By (i), ψJ ∈ DJ (ϕJ ), so there exists a net of mixing operators (Ti)i in
Mix(A, U(J + C1)) such that
(ϕ ◦ Ti)J = (ϕJ) ◦ (TiJ ) weak∗
−→ ψJ.
Passing to a subnet if necessary, we may assume that ϕ ◦ Ti → ψ′ ∈ DA(ϕ, U(J + C1)).
Then ψ′J = ψJ . Moreover, kψ′k 6 kϕk = kϕJ k = kψJk. By the uniqueness of the
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
7
norm-preserving positive extension of a positive functional, we get that ψ′ = ψ. Thus,
ψ ∈ DA(ϕ, U(J + C1)).
(iii) The image of DA/J (ϕ) by q∗
J is the set {ρ ◦ qJ : ρ ∈ DA/J (ϕ)}. This set is convex,
weak* compact, and contains ϕ ◦ qJ . Moreover, for u ∈ U(A) and ρ ∈ DA/J (ϕ) we have
u(ρ ◦ qJ )u∗ = (vρv∗) ◦ qJ , where v = qJ (u) ∈ U(A/J). Hence {ρ ◦ qJ : ρ ∈ DA/J (ϕ)} is
invariant under unitary conjugations. It follows that
DA(ϕ ◦ qJ ) ⊆ {ρ ◦ qJ : ρ ∈ DA/J (ϕ)}.
To prove the reverse inclusion it suffices to show that the left side is dense in the right
side (since the left side is weak* compact). By Proposition 3.4 (ii) (applied in A/J),
it suffices to show that eikϕe−ik ◦ qJ belongs to DA(ϕ ◦ qJ ) for all k ∈ (A/J)sa. But if
k ∈ (A/J)sa then we may find h ∈ Asa such that qJ (h) = k, from which it follows that
(eikϕe−ik) ◦ qJ = eih(ϕ ◦ qJ )e−ih ∈ DA(ϕ ◦ qJ ),
as desired.
We have shown that q∗
J maps DA/J (ϕ) onto DA(ϕ ◦ qJ ). Since q∗
J is also injective, the
(cid:3)
result follows.
Let J ⊆ A be as above a proper closed two-sided ideal of A. Let (A∗
+ such that kϕk = kϕJk. Let (A∗
+ such that ϕ(J) = {0}. Recall then that every ϕ ∈ A∗
set of functionals ϕ ∈ A∗
ϕ ∈ A∗
form ϕ = ϕ1 + ϕ2, with ϕ1 ∈ (A∗
unique (see, for example, [6, 2.11.7]).
+)J and ϕ2 ∈ (A∗
+)J denote the
+)J denote the functionals
+ can be expressed in the
+)J and that this decomposition is
+ and write ϕ = ϕ1 + ϕ2, where ϕ1, ϕ2 ∈ A∗
Theorem 3.6. Let A be a unital C*-algebra and let J be a proper closed ideal of A.
Let ϕ ∈ A∗
+)J and
ϕ2 ∈ (A∗
+)J .
(i) ϕ1 is maximally mixed if and only if ϕ1J ∈ J ∗
(ii) ϕ2 is maximally mixed if and only if the functional that it induces on A/J is
+ are such that ϕ1 ∈ (A∗
+ is maximally mixed.
maximally mixed.
(iii) ϕ is maximally mixed if and only if both ϕ1 and ϕ2 are maximally mixed. Moreover,
in this case DA(ϕ) = DA(ϕ1) + DA(ϕ2).
Proof. (i) Suppose first that ϕ1 is maximally mixed. Let ψ′ ∈ DJ (ϕ1J ). By Proposition
3.5 (i), there exists ψ ∈ DA(ϕ1) such that ψJ = ψ′. Since ϕ1 is maximally mixed, ϕ1 ∈
DA(ψ). Then, again by Proposition 3.5 (i), ϕ1J ∈ DJ (ψ′). Thus, ϕ1J is maximally
mixed.
Let us prove the converse. Let ψ ∈ DA(ϕ1). Then ψJ ∈ DJ (ϕ1J ) by Proposition 3.5
(i). Since ϕ1J is maximally mixed, ϕ1J ∈ DJ (ψJ ). By Proposition 3.5 (i), there exists
ϕ′
1k 6 kψk 6 kϕ1k. By the uniqueness
of the norm-preserving extension of a positive functional, ϕ′
1 = ϕ1. So ϕ1 ∈ DA(ψ), as
desired.
1J = ϕ1J. Moreover, kϕ′
1 ∈ DA(ψ) such that ϕ′
(ii) This is a rather straightforward consequence of Proposition 3.5 (iii). Let ϕ ∈
(A/J)∗ be such that ϕ = ϕ ◦ qJ . Suppose that ϕ is maximally mixed. By Proposition
3.5 (iii), if ψ ∈ DA(ϕ) then ψ = ψ ◦ qJ for some ψ ∈ DA/J ( ϕ). Since ϕ is maximally
mixed, ϕ ∈ DA/J ( ψ). Again by Proposition 3.5 (iii), ϕ ∈ DA(ψ) as desired. Suppose
8
ARCHBOLD, ROBERT, AND TIKUISIS
on the other hand that ϕ is maximally mixed. Let ψ ∈ DA/J ( ϕ). Then ψ ◦ qJ ∈ DA(ϕ).
Hence, ϕ ∈ DA( ψ ◦ qJ ). By Proposition 3.5 (iii), ϕ ∈ DA/J ( ψ) as desired.
+)J and T ϕ2 ∈ (A∗
(iii) Suppose that ϕ is maximally mixed. Let T ∈ Mix(A∗). Let us show first that
T ϕ1 ∈ (A∗
+)J
and (A∗
+)J is a Dixmier set. Thus, restricting to J in T ϕ = T ϕ1 + T ϕ2 we obtain
that (T ϕ)J = (T ϕ1)J . Since ϕ is maximally mixed, ϕ ∈ DA(T ϕ), and therefore
ϕJ ∈ DJ ((T ϕ)J) by Proposition 3.5 (i). Hence,
It is clear that T ϕ2 ∈ (A∗
+)J , since ϕ2 ∈ (A∗
+)J .
kϕ1k = kϕJk 6 k(T ϕ)Jk = k(T ϕ1)Jk.
So kT ϕ1k 6 kϕ1k 6 k(T ϕ1)J k, which shows that T ϕ1 ∈ (A∗
(A∗
+)J ).
To prove that ϕ1 and ϕ2 are maximally mixed we proceed as follows: Since ϕ is
maximally mixed, there exists S ∈ Mix(A∗) such that ST ϕ = ϕ. We thus have that
ϕ = ST ϕ1 + ST ϕ2. Using the last paragraph with ST in place of T , we have that
+)J . By the uniqueness of the decomposition of ϕ into
ST ϕ2 ∈ (A∗
a functional in (A∗
+)J we conclude that ST ϕ1 = ϕ1 and ST ϕ2 = ϕ2.
Thus, for any T ∈ Mix(A∗) there exists S ∈ Mix(A∗) such that ST ϕ1 = ϕ1 and
ST ϕ2 = ϕ2. In view of Lemma 2.2, this shows that ϕ1 and ϕ2 are maximally mixed.
+)J and ST ϕ1 ∈ (A∗
+)J and one in (A∗
+)J (by the definition of
Suppose now that both ϕ1 and ϕ2 are maximally mixed. Let us show first that
DA(ϕ) = DA(ϕ1) + DA(ϕ2). The inclusion DA(ϕ) ⊆ DA(ϕ1) + DA(ϕ2) is clear, for
if T ∈ Mix(A∗) then T ϕ = T ϕ1 + T ϕ2, which belongs to DA(ϕ1) + DA(ϕ2), and by
Lemma 2.2 T ϕ ranges through all of DA(ϕ). Let ϕ′
2 ∈ DA(ϕ2) and
1 + ϕ′
let us show that ϕ′
2, so
2. Recall that, as shown above, operators in Mix(A∗) preserve the
that T ϕ = T ϕ1 + ϕ′
decomposition of a maximally mixed functional into functionals in (A∗
+)J .
1 ∈ DA(T ϕ1), there exists S ∈ Mix(A∗) such that
Hence, T ϕ1 ∈ (A∗
ST ϕ1 = ϕ′
1. Moreover, by Proposition 3.5 (ii), we can choose S ∈ Mix(A∗, U(J + C1)).
Observe then that Sϕ′
2, as
desired.
2 ∈ DA(ϕ). Choose T ∈ Mix(A∗) such that T ϕ2 = ϕ′
2 vanishes on J). Hence, ST ϕ = ϕ′
1 ∈ DA(ϕ1) and ϕ′
+)J . Since ϕ′
+)J and (A∗
2 (since ϕ′
2 = ϕ′
1 + ϕ′
Continue to assume that ϕ1 and ϕ2 are maximally mixed and let us show that ϕ
1 ∈ DA(ϕ1) and
is maximally mixed. Let ϕ′ ∈ DA(ϕ). Then ϕ′ = ϕ′
ϕ′
2, where ϕ′
1 + ϕ′
2 ∈ DA(ϕ2). So
DA(ϕ) = DA(ϕ1) + DA(ϕ2) = DA(ϕ′
1) + DA(ϕ′
2) = DA(ϕ′),
where we use the fact that ϕ′
previous paragraph, for the final equality. Hence, ϕ is maximally mixed.
2 are maximally mixed, and the result of the
(cid:3)
1 and ϕ′
Corollary 3.7. Let A be a non-unital C*-algebra and ϕ ∈ A∗
+. Then ϕ is maximally
mixed if and only if its norm preserving positive extension to A∼ is maximally mixed.
In view of the previous corollary in the sequel we focus our attention on unital
C*-algebras. Further, since the scalar multiples of a maximally mixed functional are
maximally mixed, we work with states. We denote by S(A) the state space of A and
by S∞(A) the set of maximally mixed states of A.
Let A be a unital C*-algebra. Consider states ϕ ∈ S(A) of the following two types:
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
9
(A) ϕ is tracial,
(B) ϕ factors through a simple quotient A/M without bounded traces.
Not much effort is needed to see that the states of these types are maximally mixed
(for tracial states, this is obvious, whereas for type (B) states, it follows from a short
argument in Lemma 3.8 below); this prompts us to ponder whether all maximally mixed
states can be described in terms of these ones. We show in Theorem 3.10 that we are
close to getting all maximally mixed states by taking the convex hull of these ones
-- although we don't know whether the set of maximally mixed states is convex, see
Question 3.13 below.
Lemma 3.8. If B is a simple unital C*-algebra with no bounded traces, then for ev-
ery state ϕ ∈ S(B), DB(ϕ) = S(B), and thus every state of B is maximally mixed.
Therefore every type (B) state of a unital C*-algebra is maximally mixed.
Proof. Suppose for a contradiction that there exists ψ ∈ S(B) \ DB(ϕ). By the Hahn-
Banach theorem, there exist a ∈ Asa and t ∈ R such that DA(ϕ)(a) 6 t (i.e., s 6 t for all
s ∈ DA(ϕ)(a)) and ψ(a) > t. Translating by a scalar, we may assume that a is positive.
We then know that ϕ(DA(a)) 6 t (Lemma 2.1) and ψ(a) > t. But kak · 1 ∈ DA(a) (by
[7, Théorème 4]), and so kak 6 t, which contradicts that ψ(a) > t.
The final statement now follows by Theorem 3.6 (ii).
(cid:3)
In the following proposition (and henceforth, where appropriate) by a type (B) pos-
itive functional we mean a positive scalar multiple of a type (B) state.
Proposition 3.9. Let A be a unital C*-algebra. Let ϕ ∈ A∗
and let ψ ∈ A∗
DA(ϕ + ψ) = DA(ϕ) + DA(ψ).
+ be maximally mixed
+ be either tracial or type (B). Then ϕ + ψ is maximally mixed and
Proof. If ψ is tracial then DA(ϕ + ψ) = DA(ϕ) + ψ, from which the result follows at
once. Suppose then that ψ is type (B), i.e., it factors through a simple quotient A/M
without bounded traces. Let ϕ = ϕ1 + ϕ2, where ϕ1 ∈ (A∗
+)M . Then
+)M and ϕ2 ∈ (A∗
ϕ + ψ = ϕ1 + (ϕ2 + ψ).
By Theorem 3.6 (iii), ϕ1 is maximally mixed. On the other hand, ϕ2 + ψ is type (B) (it
factors through A/M), so by Lemma 3.8, it is maximally mixed. Hence, by Theorem
3.6 (iii), ϕ + ψ = ϕ1 + (ϕ2 + ψ) is maximally mixed. Moreover, Theorem 3.6 (iii) also
shows that DA(ϕ+ψ) = DA(ϕ1)+DA(ϕ2 +ψ). But DA(ϕ2 +ψ) = (ϕ2(1)+ψ(1))S(A)M ,
where S(A)M = S(A/M) ◦ qM (i.e., all states that factor through A/M). So
DA(ϕ + ψ) = DA(ϕ1) + ϕ2(1)S(A)M + ψ(1)S(A)M
= DA(ϕ1) + DA(ϕ2) + DA(ψ)
= DA(ϕ) + DA(ψ),
using Theorem 3.6 (iii) again for the last equality.
(cid:3)
Given a C*-algebra A, we denote by T (A) the set of tracial states on A.
10
ARCHBOLD, ROBERT, AND TIKUISIS
Theorem 3.10. Let A be a unital C*-algebra. Let S(B)(A) denote the set of states on
A of type (B). Then
co(T (A) ∪ S(B)(A))
k·k
⊆ S∞(A) ⊆ co(T (A) ∪ S(B)(A))
weak∗
.
Examples 3.16, 3.17, and 4.9 show that both inclusions in the above theorem can be
strict.
Proof. It follows by Proposition 3.9 that co(T (A) ∪ S(B)(A)) ⊆ S∞(A), and so by
Theorem 3.2
k·k
co(T (A) ∪ S(B)(A))
⊆ S∞(A).
weak∗
To show that S∞(A) is contained in the weak* closure of co(T (A)∪S(B)(A)), it suffices
to show that for any ϕ ∈ S(A) the Dixmier set DA(ϕ) has nonempty intersection with
co(T (A) ∪ S(B)(A))
. Suppose, for the sake of contradiction, that this is not the
case for some ϕ ∈ S(A). Then, by the Hahn-Banach theorem, there exists a self-adjoint
element a and real numbers t1 < t2 such that ψ(a) 6 t1 for all ψ ∈ co(T (A) ∪ S(B)(A))
and ψ′(a) > t2 for all ψ′ ∈ DA(ϕ). Translating a by a multiple of the unit we can assume
that it is positive. Since DA(ϕ)(a) = ϕ(DA(a)) (Lemma 2.1), we have that ϕ(a′) > t2
for all a′ ∈ DA(a). On the other hand, ψ(a) 6 t1 for every tracial state and every state
that factors through a simple quotient without bounded traces. By [5, Theorem 4.12],
the distance from DA(a) to 0 is at most t1. Thus, there exists a′ ∈ DA(a) such that
ka′k < t2. This contradicts that ϕ(a′) > t2.
(cid:3)
Condition (i) of the following corollary has appeared in several papers previously
(e.g. [9], [7], [10]). In fact, an improved version of this corollary is [7, Theorem 5].
Corollary 3.11. Let A be a unital C*-algebra. The following are equivalent:
(i) Every simple quotient of A has a bounded trace.
(ii) All the maximally mixed states of A are tracial.
(iii) For every ϕ ∈ S(A) the Dixmier set DA(ϕ) contains a tracial state.
Proof. (i)⇒(ii) follows from the previous theorem, since (i) implies that co(T (A) ∪
S(B)(A)) = T (A), which is already a weak* closed set. As remarked at the beginning
of this section, every Dixmier set DA(ϕ) must contain maximally mixed functionals.
With this in mind, (ii)⇒(iii) is obvious. Finally, assume (iii), and let A/M be a given
simple quotient of A. Choose any ϕ ∈ S(A) that factors through A/M. Then any trace
in DA(ϕ) factors through A/M. We thus have (i).
(cid:3)
In the case of simple C*-algebras we obtain a complete description of the maximally
mixed positive functionals:
Corollary 3.12. Let A be a simple C*-algebra.
(i) If A is unital and has at least one non-zero bounded trace then every maximally
mixed positive functional on A is tracial.
(ii) If A is unital and has no bounded traces then all the positive functionals on A are
maximally mixed.
(iii) If A is non-unital then every maximally mixed positive functional on A is tracial.
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
11
Proof. (i) follows from Corollary 3.11, while (ii) is Lemma 3.8. For (iii), note that A∼
has only one simple quotient, namely C, and it has a bounded trace. Hence by Corollary
3.11, every maximally mixed state of A∼ is tracial, and then (iii) follows from Theorem
3.6 (i).
(cid:3)
Question 3.13. Let A be a unital C*-algebra. Is the set S∞(A) of maximally mixed
states convex?
A closely related question is the following:
Question 3.14. Given maximally mixed functionals ϕ, ψ ∈ A∗
DA(ϕ) + DA(ψ)?
+, do we have DA(ϕ+ψ) =
An affirmative answer to Question 3.14 for all ϕ, ψ ∈ S∞(A) also answers affirmatively
Question 3.13. Indeed, suppose that Question 3.14 has an affirmative answer for all
ϕ, ψ ∈ S∞(A). Say we are given ϕ, ψ ∈ S∞(A) and ϕ′ ∈ DA(ϕ) and ψ′ ∈ DA(ψ). Then
DA(ϕ + ψ) = DA(ϕ) + DA(ψ) = DA(ϕ′) + DA(ψ′) = DA(ϕ′ + ψ′).
Observe that Proposition 3.9 answers Question 3.14 affirmatively in the case that ψ is
either tracial or type (B).
Turning to the question of whether the containment S∞(A) ⊆ co(T (A) ∪ S(B)(A))
weak∗
is strict (where S(B)(A) is as defined in Theorem 3.10), it is evident from that theorem
that (non-)strictness of this inequality is equivalent to the natural question of whether
S∞(A) is weak* closed. The next proposition gives an obstruction to S∞(A) being
weak* closed -- in fact, it is the only obstruction we have been able to find, see Question
3.18.
Proposition 3.15. Let A be a unital C*-algebra such that S∞(A) is a weak* closed
subset of S(A). Then the set X ⊆ Prim(A) of all maximal ideals M such that A/M
either is isomorphic to C or has no bounded traces is a closed subset of Prim(A).
Proof. We may assume that X is non-empty. Let J = TM ∈X M. Let N ∈ Prim(A)
be an adherence point of X, i.e, J ⊆ N. Then every state on A that factors through
A/N is a weak* limit of convex combinations of states that factor through A/M, with
M ∈ X ([6, Proposition 3.4.2 (i)]). Notice that S∞(A/M) = S(A/M) for all M ∈ X.
Thus, all the states of A that factor through A/M, with M ∈ X, are maximally mixed.
It follows that all states factoring through A/N are maximally mixed, and so all states
of A/N are maximally mixed by Theorem 3.6 (ii).
Since N is primitive, let ϕ ∈ S(A/N) be a pure state whose GNS representation πϕ is
faithful. Then any pure state ψ on A/N is a weak* limit of vector states (with respect
to πϕ) by [6, Corollary 3.4.3]. By the unitary version of Kadison's Transitivity Theorem
([6, Theorem 2.8.3 (iii)]), each of these vector states is in fact unitarily equivalent to
ϕ, and thus ψ is a weak* limit of unitary conjugates of ϕ. By approximating arbitrary
states on A/N by convex combinations of pure states, we find that S(A/N) = DA/N (ϕ).
This implies that A/N is simple, for otherwise the states factoring through a non-trivial
quotient would form a proper Dixmier subset of DA/N (ϕ) (recall that ϕ is maximally
mixed). From Corollary 3.12 we see that A/N must either be isomorphic to C or
without bounded traces. Thus, N ∈ X.
(cid:3)
12
ARCHBOLD, ROBERT, AND TIKUISIS
The examples below show that S∞(A) may fail to be weak* closed.
Example 3.16. Fix a simple unital C*-algebra B without bounded traces (e.g., the
Cuntz algebra O2). Let A be the C*-subalgebra of C([0, 1], M2(B)) of functions f
such that f (1) ∈ M2(C) ⊆ M2(B). For each t ∈ [0, 1] let It = {f ∈ A : f (t) = 0}.
∼= M2(B) for all 0 6 t < 1. So It is a maximal ideal such that A/It is
Then A/It
simple without bounded traces. The maximal ideal I1 is an adherence point of the set
∼= M2(C) has a bounded trace and is not isomorphic
{It : 0 6 t < 1}. However, A/I1
to C. Thus, S∞(A) is not weak* closed, by Proposition 3.15.
Example 3.17. Again fix a simple unital C*-algebra B without bounded traces. Let A
be the C*-subalgebra of C({1, 2, . . . , ∞}, (B ⊗ K)∼) of f such that f (n) ∈ Mn(B) + C1
for all n ∈ N, where we regard Mn(B) embedded in B ⊗ K as the top left corner.
For each n ∈ N define In = {f ∈ A : enf (n) = 0}, where en is the unit of Mn(B).
∼= Mn(B) has no bounded
Then In is a maximal ideal for all n = 1, 2, . . . and A/In
traces. Since Tn In = {0}, the set {In : n ∈ N} is dense in Prim(A). Consider the
ideal I∞ = {f : f (∞) = 0}. Since A/I∞ = (B ⊗ K)∼ is a primitive C*-algebra,
I∞ ∈ Prim(A). But I∞ is not maximal. By Proposition 3.15, S∞(A) is not weak*
closed.
If one wanted an algebra A with no bounded traces in which S∞(A) is not weak*
closed, one can simply tensor the example just given with a nuclear, unital, simple,
traceless C*-algebra (this operation does not change the ideal lattice, so the same
obstruction applies).
Question 3.18. Is the converse of Proposition 3.15 true? That is, let A be unital.
Suppose that the set of maximal ideals M such that A/M either is isomorphic to C or
has no bounded traces is a closed subset of Prim(A). Is S∞(A) weak* closed?
The previous question admits a reduction to the special case of characterizing when
S∞(A) is all of S(A) (Remark 3.21 below). Before explaining this, we point out the
following result:
Proposition 3.19. Let A be a unital C*-algebra. Let X ⊆ Prim(A) be as in Proposition
3.15. The following are equivalent:
(i) X = Prim(A);
(ii) every pure state of A is either multiplicative or type (B);
(iii) S∞(A) contains the norm-closed convex hull of the pure states.
Proof. (i)⇒(ii): If ϕ is a pure state then ϕ factors through a primitive quotient A/N.
By (i), this quotient is simple and either isomorphic to C or without traces. In the first
case ϕ is muliplicative and in the second case it is of type (B).
(ii)⇒(iii): This follows from Theorem 3.10.
(iii)⇒(i): Let N ∈ Prim(A) and choose a pure state ϕ ∈ S(A/N) whose GNS
representation is faithful. Then, since ϕ is maximally mixed, it follows as in the proof
of Proposition 3.15 that A/N is simple. Since every pure state of A/N is maximally
mixed, it follows from Corollary 3.12 that A/N must either be C or have no bounded
traces. Thus, N ∈ X.
(cid:3)
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
13
Question 3.20. Let A be a unital C*-algebra that satisfies the equivalent conditions of
Proposition 3.19. Does it follow that S∞(A) = S(A)?
Remark 3.21. Questions 3.18 and 3.20 are equivalent. For if Question 3.18 has been
answered affirmatively and A satisfies the equivalent conditions of Proposition 3.19,
then S∞(A) is weak* closed, and so by Proposition 3.19 (iii) S∞(A) = S(A). Suppose
conversely that Question 3.20 has been answered affirmatively. Let A be a unital C*-
algebra. Let X ⊆ Prim(A) be as in Proposition 3.15 and suppose that X is a closed
set. If X = ∅ then, by Corollary 3.11, S∞(A) = T (A) (the set of tracial states on
A), which is weak∗ closed. Assume X 6= ∅. Let J = TM ∈X M. Then the C*-algebra
A/J satisfies the equivalent conditions in Proposition 3.19 (check (i)). Hence, all of its
states are maximally mixed. By Theorem 3.6 (ii), S(A)J ⊆ S∞(A). Observe also that
every state of A of type (B) is in S(A)J . Then,
co(T (A), S(A)J) ⊆ S∞(A) ⊆ co(T (A), S(A)J )
weak∗
,
where the first inclusion follows from Proposition 3.9 and the second inclusion from
Theorem 3.10. But co(T (A), S(A)J) is weak* closed. So S∞(A) = co(T (A), S(A)J), is
weak* closed.
In the next section we answer affirmatively Questions 3.13, 3.14, and 3.18 for C*-
algebras with the Dixmier property.
4. C*-algebras with the Dixmier property
In this section, we find further properties of the set of maximally mixed states in the
case of C*-algebras with the Dixmier property (defined below).
Let A be a unital C*-algebra, let a ∈ A and let ϕ ∈ S(A). Acting by conjugation, the
unitary group U(A) induces a group of isometric affine transformations of the convex
set DA(a), and similarly for DA(ϕ). An element z ∈ DA(a) is a fixed point for the
group of conjugations of DA(a) if and only if it belongs to the centre Z(A). An element
τ ∈ DA(ϕ) is a fixed point for the group of conjugations on DA(ϕ) if and only if it is a
tracial state, i.e., τ ∈ T (A).
The C*-algebra A is said to have the Dixmier property if DA(a) ∩ Z(A) is non-empty
for all a ∈ A, and A is said to have the singleton Dixmier property if DA(a) ∩ Z(A) is
a singleton set for all a ∈ A (see [5] and the many references cited therein). On the
other hand, we have seen in Corollary 3.11 that DA(ϕ) ∩ T (A) is non-empty for all
ϕ ∈ S(A) if and only if every simple quotient of A has a tracial state. Since a unital
simple C*-algebra has the Dixmier property if and only if it has at most one tracial
state [7], we see that the Dixmier property is neither necessary nor sufficient for the
equivalent properties of Corollary 3.11 to hold. Indeed, it is shown in [5, Proposition
1.4] that A has the Dixmier property and also satisfies the conditions of Corollary 3.11
if and only if it has the singleton Dixmier property.
If a Dixmier set DA(a) does not contain a central element then there is no natural
"second prize" at which to aim. In contrast, if a Dixmier set DA(ϕ) does not contain
a tracial state then we may nevertheless study the maximally mixed states, which
we have already seen to be guaranteed to exist in DA(ϕ).
In this section we study
14
ARCHBOLD, ROBERT, AND TIKUISIS
the maximally mixed states in the case where A has the Dixmier property but not
necessarily the singleton Dixmier property.
Henceforth in this section we assume that A is a unital C*-algebra with the Dixmier
property.
Let Z denote the spectrum of Z(A). Since C*-algebras with the Dixmier property
are weakly central (e.g., see [5]), we can identify Z with the set of maximal ideals of A.
We denote the latter set by Max(A).
To analyze the maximally mixed states for such A, we will make frequent use of a
description of DA(a) ∩ Z(A) (for a self-adjoint) found in [5] (see [5, Corollary 4.5] and
the discussion between Theorem 2.6 and Corollary 2.7 of [5], with details in the proof
of Theorem 2.6). For a ∈ A self-adjoint, define fa, ga : Z → R by
min sp(qM (a)),
τM (a),
if A/M has no bounded traces;
otherwise,
where τM is the (necessarily unique) tracial state on A which factors through A/M.
Likewise,
fa(M) :=
ga(M) :=
(4.1)
max sp(qM (a)),
τM (a),
if A/M has no bounded traces;
otherwise.
Then fa is upper semicontinuous, ga is lower semicontinuous, fa 6 ga, and, identifying
Z(A) = C( Z) now,
DA(a) ∩ Z(A) = {z ∈ C( Z) : z = z∗ and fa 6 z 6 ga}.
Let us say that two maximally mixed bounded functionals ϕ and ψ are equivalent if
they generate the same Dixmier set, i.e., DA(ϕ) = DA(ψ).
Proposition 4.1. Let A be a unital C*-algebra with the Dixmier property. The equiv-
alence classes of maximally mixed, bounded functionals on A are in bijective correspon-
dence with the bounded functionals on the center of A. The correspondence is given by
the restriction map ϕ 7→ ϕZ(A), for ϕ maximally mixed.
Proof. Any two equivalent functionals agree on the center, so the mapping is well defined
on equivalence classes. To see that it is onto, fix a functional µ ∈ Z(A)∗. The set of all
ϕ ∈ A∗ whose restriction to Z(A) is µ is a weak* compact Dixmier set. It thus must
contain maximally mixed functionals.
Let us now show that the mapping is injective. Let ϕ, ψ ∈ A∗ be two maximally mixed
functionals that agree on Z(A). Suppose for a contradiction that DA(ϕ) 6= DA(ψ).
Then DA(ϕ) and DA(ψ) are disjoint. By the Hahn-Banach theorem, we can find a ∈ A
and real numbers t1 < t2 such that Re(ϕ′(a)) 6 t1 for all ϕ′ ∈ DA(ϕ) and Re(ψ′(a)) > t2
for all ψ′ ∈ DA(ψ). By Lemma 2.1, Re(ϕ(a′)) 6 t1 and Re(ψ(a′)) > t2 for all a′ ∈ DA(a).
This holds in particular for a′ ∈ DA(a) ∩ Z(A). This contradicts that ϕ and ψ agree
on Z(A).
(cid:3)
Remark 4.2. The previous proposition implies that if A has the Dixmier property then
DA(ϕ), for ϕ ∈ S(A), contains a unique equivalence class of maximally mixed states;
namely, the maximally mixed states that agree with ϕ on Z(A). This is in general
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
15
not true for C*-algebras without the Dixmier property. Take for example A to be a
simple unital C*-algebra with at least two tracial states and let ϕ be a pure state of
A. Then DA(ϕ) is the set of all states, so it contains distinct tracial states (which are
inequivalent maximally mixed states).
We need the following little lemma in the proceeding theorem.
Lemma 4.3. Let X be a Hausdorff topological space, let µ be a Radon probability
measure on X, and let f : X → R be a bounded lower semicontinuous function. Then
ZX
f dµ = supZX
g dµ,
where the supremum is taken over upper semicontinuous functions g : X → R which
are (pointwise) dominated by f .
Proof. Without loss of generality, f > 0. We may approximate f uniformly by simple
lower semicontinuous functions, i.e., positive scalar linear combinations of characteristic
functions of open sets. Thus, it suffices to handle the case that f is the characteristic
function of an open set, say f = χU .
In this case, since µ is inner regular, µ(X) is the supremum of measures of compact
sets K contained in U, so
ZX
f dµ = µ(X)
= sup
K
µ(K)
= sup
K ZX
χK dµ,
where the suprema are taken over compact sets contained in U; but now we are done,
since each χK is upper semicontinuous.
(cid:3)
Theorem 4.4. Let A be a unital C*-algebra with the Dixmier property. Let ϕ ∈ A∗
+.
The following are equivalent:
(i) ϕ satisfies that
(4.2)
ϕ(a) 6 sup{ϕ(z) : z ∈ DA(a) ∩ Z(A)}
(a ∈ A+).
(ii) ϕ is maximally mixed.
Proof. (i)⇒(ii). Suppose for a contradiction that there exists ψ ∈ DA(ϕ) such that
ϕ /∈ DA(ψ). Then there exist a self-adjoint element a and t ∈ R separating DA(ψ)
and ϕ. That is, ψ′(a) 6 t for all ψ′ ∈ DA(ψ) and ϕ(a) > t. Translating a by a scalar
multiple of the unit we may assume that it is positive. By Lemma 2.1, we get that
ψ(a′) 6 t for all a′ ∈ DA(a). From ψ ∈ DA(ϕ) we deduce that ψ(a′) = ϕ(a′) for all
a′ ∈ Z(A). Hence
ϕ(a) 6 sup{ϕ(a′) : a′ ∈ DA(a) ∩ Z(A)}
= sup{ψ(a′) : a′ ∈ DA(a) ∩ Z(A)} 6 t.
This contradicts that ϕ(a) > t.
16
ARCHBOLD, ROBERT, AND TIKUISIS
(ii)⇒(i). We may assume that ϕ 6= 0 and then, multiplying it by a scalar, that it is
a state. First, let us show that if a maximally mixed state ϕ satisfies (4.2) then so do
all the states equivalent to it. Let ϕ be a state that satisfies (4.2) and let ψ ∈ DA(ϕ).
Say ψ = limi ϕ ◦ Ti, where (Ti)i is a net of mixing operators in Mix(A). Let a ∈ A+.
Since DA(Tia) ⊆ DA(a),
Hence
ϕ(Tia) 6 sup{ϕ(z) : z ∈ DA(Tia) ∩ Z(A)}
6 sup{ϕ(z) : z ∈ DA(a) ∩ Z(A)}.
ψ(a) = lim
i
ϕ(Ti · a)
6 sup{ϕ(z) : z ∈ DA(a) ∩ Z(A)}
= sup{ψ(z) : z ∈ DA(a) ∩ Z(A)},
where the last equality is valid since ϕ and ψ agree on Z(A).
By Proposition 4.1, it now suffices to show that every probability (Radon) measure
µ on the center can be extended to a state ϕ on A satisfying (4.2). We do this next.
For each a ∈ Asa let us define pµ(a) ∈ [0, ∞) by
pµ(a) := Z
bZ
ga(M) dµ(M),
where ga : bZ → [0, ∞) is the lower semicontinuous function on the spectrum of the
center associated to a (as in (4.1) with a in place of a). Let us show that pµ is a
seminorm. Clearly pµ(ta) = tpµ(a) for any t ∈ R. To prove the triangle inequality it
suffices to show that ga+b 6 ga + gb for all a, b ∈ Asa. Let us evaluate both sides of
this inequality on an ideal M ∈ Max(A) such that A/M has no bounded traces. Set
¯a = qM (a) and ¯b = qM (b) (the images of a and b in A/M). Then we must show that
k¯a + ¯bk 6 k¯ak + k¯bk. But this is clear from the triangle inequality for k · k and the
fact that the norm of an element is equal to the norm of its absolute value. Suppose
now that M is such that A/M has bounded traces. Let τM be the unique tracial state
on A factoring through A/M. Then we must show that
τM (a + b) 6 τM (a) + τM (b).
(4.3)
Let p ∈ A∗∗ be a projection such that p(a + b)p = (a + b)+. Multiplying by p on the
left and on the right of a + b 6 a+ + b+ we get (a + b)+ 6 pa+p + pb+p. Evaluating τM
(extended to a normal trace on A∗∗) on both sides we get
The same inequality, applied to −a and −b, yields that
τM ((a + b)+) 6 τM (a+) + τM (b+).
τM ((a + b)−) 6 τM (a−) + τM (b−).
Now adding both inequalities we get (4.3), as desired. Thus, pµ is a seminorm. Since
ga 6 kak, we also have that pµ(a) 6 kak for all a ∈ Asa.
For any self-adjoint central element z we have that
(cid:12)(cid:12)(cid:12)(cid:12)
Z z(M) dµ(M)(cid:12)(cid:12)(cid:12)(cid:12) 6 Z z(M) dµ(M) = pµ(z).
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
17
So we can extend µ by the Hahn-Banach extension theorem to a self-adjoint functional
ϕ on A such that
ϕ(a) 6 pµ(a)
(a ∈ Asa).
Notice that ϕ(1) = 1 and that kϕk 6 1, since pµ(a) 6 kak for all a ∈ Asa. Hence, ϕ is
a state.
Let a ∈ A+. To establish (4.2), we will show that pµ(a) is dominated by the right-
hand side of (4.2) (though we don't need it, in fact this implies that these two quantities
are equal, as the reverse inequality is straightforward). Let ε > 0. Since ga is lower
semicontinuous, it follows from Lemma 4.3 that we may find an upper semicontinuous
function w ∈ C( Z) such that w 6 ga and R w(M) dµ(M) > R ga(M) dµ(M) − ε. By the
Katetev-Tong insertion theorem, we may find a continuous function z0 ∈ C( Z)+ such
that
fa
w
6 z0 6 ga,
and therefore
Z z0(M) dµ(M) > Z w(M) dµ(M) > Z ga(M) dµ(M) − ε = pµ(a) − ε.
Thus
as required.
pµ(a) 6 sup{ϕ(z) : z ∈ DA(a) ∩ Z(A)},
(cid:3)
Corollary 4.5. Let A be a unital C*-algebra with the Dixmier property.
(i) S∞(A) is convex and weakly closed.
(ii) We have DA(ϕ + ψ) = DA(ϕ) + DA(ψ) for all ϕ, ψ ∈ A∗
+ maximally mixed.
Proof. (i) Let us show first that S∞(A) is convex. Since a scalar multiple of a maximally
mixed functional is again maximally mixed, it suffices to show that if ϕ, ψ ∈ A∗
+ are
maximally mixed, then ϕ + ψ is maximally mixed. So let ϕ, ψ ∈ A∗
+ be maximally
mixed. We show that ϕ + ψ satisfies (4.2). Let a ∈ A+ and ε > 0. Since ϕ and ψ
satisfy (4.2), there exist x, y ∈ DA(a) ∩ Z(A) such that
ϕ(a) 6 ϕ(x) + ε and ψ(a) 6 ψ(y) + ε.
By the structure of DA(a) ∩ Z(A) for self-adjoint a we know that it is a lattice. So we
can choose z ∈ DA(a) ∩ Z(A) such that x, y 6 z. Then (ϕ + ψ)(a) 6 (ϕ + ψ)(z) + 2ε.
This shows that ϕ + ψ satisfies (4.2) and is therefore maximally mixed.
Since S∞(A) is convex and norm closed (Theorem 3.2), it is also weakly closed (i.e.,
closed in the σ(A∗, A∗∗) topology).
(ii) The inclusion DA(ϕ + ψ) ⊆ DA(ϕ) + DA(ψ) is straightforward: if T ∈ Mix(A∗)
then
T (ϕ + ψ) = T ψ + T ψ ∈ DA(ϕ) + DA(ψ),
and letting T range through Mix(A∗), T (ϕ+ψ) ranges through all of DA(ϕ+ψ) (Lemma
2.2).
Let ϕ, ψ ∈ A∗
+ be maximally mixed and suppose, for a contradiction, that there exist
ϕ′ ∈ DA(ϕ) and ψ′ ∈ DA(ψ) such that ϕ′ + ψ′ /∈ DA(ϕ + ψ). Then there exist a ∈ Asa
and t ∈ R such that ρ(a) 6 t for all ρ ∈ DA(ϕ + ψ) while (ϕ′ + ψ′)(a) > t. Translating
18
ARCHBOLD, ROBERT, AND TIKUISIS
a by a scalar multiple of the unit, and changing t accordingly, we may assume that a
is positive. By Lemma 2.1, (ϕ + ψ)(b) 6 t for all b ∈ DA(a). Since ϕ + ψ and ϕ′ + ψ′
agree on Z(A), we obtain that (ϕ′ + ψ′)(b) 6 t for all b ∈ DA(a) ∩ Z(A). But ϕ′ + ψ′ is
maximally mixed by the proof of (i). It follows by Theorem 4.4 that (ϕ′ + ψ′)(a) 6 t,
which contradicts our choice of a and t.
(cid:3)
Remark 4.6. The C*-algebras in Examples 3.16 and 3.17 both have the Dixmier property
(this can be deduced from [5, Theorem 1.1]). So S∞(A) may fail to be weak* closed for
C*-algebras with the Dixmier property.
Theorem 4.7. Let A be a unital C*-algebra with the Dixmier property. The following
are equivalent.
(i) The set S∞(A) is weak* closed;
(ii) The set of maximal ideals M such that A/M either is isomorphic to C or has no
bounded traces is a closed subset of Prim(A);
(iii) For each self-adjoint a ∈ A, the set DA(a) ∩ Z(A) contains a maximal element.
Proof. (i)⇒(ii): This is Proposition 3.15 (no Dixmier property required).
(ii)⇒(iii): By translating, we may assume that a > 0. Let X denote the set of
maximal ideals M ∈ Max(A) such that A/M either is isomorphic to C or has no
bounded traces, and we assume that this set is closed in Prim(A). It is evident from
the description of DA(a) ∩ Z(A), at the beginning of this section, that we need only
show that the function ga : Z → R from (4.1) is continuous. Since ga is always lower
semicontinuous, it remains to show that it is upper semicontinuous. Let t > 0. Set
Y := {M ∈ Max(A) : T (A/M) 6= ∅},
which is closed in Max(A) by [5, Theorem 2.6]; for M ∈ Y , A has a unique tracial
state τM that factors through A/M. Since τM depends weak* continuously on M ∈ Y
([5, Theorem 2.6]),
{M ∈ Y : τM (a) > t}
is closed in Max(A). Also, {M ∈ Prim(A) : kqM (a)k > t} is a compact subset of
Prim(A) ([6, Proposition 3.3.7]), from which (along with that X is closed in Prim(A))
we deduce that
{M ∈ Prim(A) : kqM (a)k > t} ∩ X
is compact. Since Max(A) is Hausdorff, the set above is also closed in Max(A). There-
fore,
{M ∈ Y : τM (a) > t} ∪ ({M ∈ Prim(A) : kqM (a)k > t} ∩ X)
is closed in Max(A). But this set is g−1
uous.
a ([t, ∞)), and therefore, ga is upper semicontin-
(iii)⇒(i): For each self-adjoint element a ∈ A, let za denote the maximal element of
DA(a) ∩ Z(A), which exists since we are assuming (iii). Given a state ϕ, the inequality
(4.2) is equivalent to ϕ(a) 6 ϕ(za) for all a ∈ A+. The latter inequality is clearly
preserved under weak* limits. By Theorem 4.4, S∞(A) is weak* closed.
(cid:3)
We recover as a corollary Alberti's theorem on the maximally mixed states of a von
Neumann algebra ([1, Theorem 5.2] and [3, Theorem 4-12]):
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
19
Corollary 4.8. Let A be a von Neumann algebra. Then S∞(A) agrees with the weak*
closure of the convex hull of the set of tracial states and type (B) states.
Proof. Let A = Af L Api be the decomposition of A into a finite and a properly infinite
von Neumann algebra. Let J ⊆ Api denote the strong Jacobson radical of Api, i.e, the
intersection of all maximal ideals of Api. By [8, Proposition 2.3], N ∈ Prim(Api) is a
maximal ideal of Api if and only if J ⊆ N. It follows that N ∈ Prim(A) is maximal
(in A) and such that A/N has no bounded traces if and only if Af ⊕ J ⊆ N. This set
is thus a closed subset of Prim(A). On the other hand, the set of M ∈ Prim(A) such
that A/M ∼= C is a closed subset of Prim(A) (indeed, these are the M ∈ Prim(A) that
contain the ideal generated by the commutators of A, which is the smallest ideal the
quotient by which is abelian). So the union of these two sets is closed. Moreover, A has
the Dixmier property (by Dixmier's approximation theorem). Thus, by the previous
theorem, S∞(A) is weak* closed. The result then follows from Theorem 3.10.
(cid:3)
We end this section by taking advantage of the insight we have gained in the case of
the Dixmier property, to provide some examples alluded to earlier. The first example
shows that the set of maximally mixed states may be larger than the norm-closed convex
hull of the tracial states and type (B) states.
Example 4.9. Let B be a simple unital C*-algebra with no bounded traces, and set
A := C([0, 1], B). If ϕ is in the norm-closed convex hull of the type (B) states, then the
state that ϕ induces on the centre is in the norm-closed convex hull of point-masses,
and therefore corresponds to a discrete measure on [0, 1]. However, A has the Dixmier
property by [5, Theorem 2.6] and hence S∞(A) is weak* closed by Theorem 4.7 ((ii)
⇒ (i)). Since every pure state of A is of type (B), it follows from Theorem 3.10 that
S∞(A) = S(A). So the norm-closed convex hull of the type (B) states (and tracial
states, as there are none) is only a small part of S∞(A) in this case.
The next example addresses the converse to Theorem 3.1.
Example 4.10. Let A be a simple unital C*-algebra with no bounded traces. Then A
has the Dixmier property ([7]). Let ϕ be a nonzero self-adjoint functional on A such
that ϕ(1) = 0. Then ϕ is not maximally mixed, because if it were, then since the
zero functional is maximally mixed, it would follow by Proposition 4.1 that DA(ϕ) =
DA(0) = {0}. However, by Corollary 3.12 (ii), both the positive and negative parts of
ϕ are maximally mixed.
5. Hausdorff primitive spectrum
Here we impose a different property -- Hausdorffness of the primitive ideal space -- to
make the study of the structure of S∞(A) tractable.
Given a C*-algebra A, we continue to denote by T (A) the set of tracial states on A.
Theorem 5.1. Let A be a unital C*-algebra with Hausdorff primitive spectrum.
(i) Suppose that T (A) = ∅. Then every state of A is maximally mixed.
(ii) Suppose that T (A) 6= ∅. Then the set
Y := {M ∈ Max(A) : T (A/M) 6= ∅}
20
ARCHBOLD, ROBERT, AND TIKUISIS
is non-empty and closed in Max(A) and
S∞(A) = co(T (A) ∪ S(A)J ),
where J := TM ∈Y M is a proper closed ideal of A, and S(A)J consists of all states
in S(A) which arise as extensions of states in S(J).
(iii) Questions 3.13 and 3.14 have affirmative answers for A.
Proof. Observe first that, since Prim(A) is Hausdorff, Prim(A) = Max(A) = Glimm(A),
and these spaces are all homeomorphic to Max(Z(A)) via the assignment M 7→ M ∩
Z(A). For each maximal ideal N of Z(A), let ϕN be the unique pure state of Z(A)
with kernel equal to N.
(i) Since the continuous functions on the compact Hausdorff space Prim(A) separate
the points, it follows from the Dauns-Hofmann theorem that A is a central C*-algebra.
Combining this with the fact that T (A) is empty, we obtain from [5, Theorem 2.6] that
A has the Dixmier property. Every pure state of A is of type (B), so by Theorem 3.10,
S∞(A) is weak* dense in S(A). It follows from Theorem 4.7 that every state of A is
maximally mixed.
(ii) Since T (A) is non-empty, it contains an extreme point τ by the Krein-Milman
theorem. By [5, Lemma 2.4], τ Z(A) is a pure state of Z(A) and hence annihilates
M ∩ Z(A) for some M ∈ Max(A). By the Cauchy -- Schwartz inequality for states, τ
annihilates the Glimm ideal (M ∩ Z(A))A. But, as noted above, (M ∩ Z(A))A = M
and so τ induces a tracial state of A/M. Thus Y is non-empty. Moreover, τ (J) = {0}
and so, by the Krein-Milman theorem, every tracial state of A annihilates J.
To show that Y is closed, suppose that (Mi) is a net in Y that is convergent to
M ∈ Max(A). For each i, let τi be a tracial state of A that vanishes on Mi. Since T (A)
is weak∗-compact, there exist τ ∈ T (A) and a subnet (τij ) such that τij →j τ . Then
τ Z(A) = lim
j
ϕMij ∩Z(A) = ϕM ∩Z(A).
It follows from the Cauchy -- Schwartz inequality for states that τ annihilates the Glimm
ideal (M ∩ Z(A))A and so τ induces a tracial state of A/M as before. Thus M ∈ Y ,
as required.
Since Y is closed, every maximal ideal of A/J has the form M/J for some M ∈ Y
and hence every simple quotient of A/J has a tracial state. It follows by Corollary 3.11
that S∞(A/J) = T (A/J). Letting S2 be the set of maximally mixed states of S∞(A)
which factor through A/J, it follows by Theorem 3.6 (ii) that
S2 = S∞(A/J) ◦ qJ = T (A/J) ◦ qJ = T (A).
Under the Dauns -- Hofmann isomorphism between Z(A) and C(Prim(A)), Z(J) cor-
responds to C0(Prim(J)), where Prim(J) is identified with an open subset of Prim(A)
(namely Prim(A) \ Y ) in the usual way. It follows that Z(J) separates the primitive
ideals of J + C1 and hence J + C1 is a central C*-algebra. Since J has no tracial states
(this follows from [5, Lemma 2.2]), J + C1 has a unique tracial state, namely the one
factoring through the quotient (J + C1)/J. Hence by [5, Theorem 2.6], J + C1 has
the Dixmier property. We also have that Prim(J + C1) = Max(J + C1), with every
simple quotient being either traceless or isomorphic to C, and thus by Theorem 4.7,
MAXIMALLY UNITARILY MIXED STATES ON A C*-ALGEBRA
21
S∞(J + C1) is weak* closed. Since every pure state is either type (B) or tracial, it now
follows from Theorem 3.10 that S∞(J + C1) = S(J + C1). By Theorem 3.6 (i) (used
once with J ⊳ J + C1 and again with J ⊳ A), S(A)J ⊆ S∞(A). Thus, letting S1 be
the set of maximally mixed states of A which are extensions of states from J, we have
By Theorem 3.6 (iii), we have
S∞(A) = co(S1 ∪ S2) = co(S(A)J ∪ T (A)),
S1 = S(A)J .
as required.
(iii) It is evident in both the cases covered by (i) and (ii) that S∞(A) is convex. Now
let ϕ, ψ ∈ A∗
+ be maximally mixed, and let's argue that DA(ϕ + ψ) = DA(ϕ) + DA(ψ).
In case (i), we saw that A has the Dixmier property, so this holds by Corollary 4.5 (ii).
In case (ii), write ϕ = ϕ1 + ϕ2 and ψ = ψ1 + ψ2 where ϕ1, ψ1 are positive tracial
functionals and ϕ2, ϕ2 are non-negative scalar multiples of states in S(A)J ; by Theorem
3.6 (i), ϕ2J and ψ2J are maximally mixed functionals on J. Thus, so are their norm-
preserving positive extensions to J +C1 (Theorem 3.6 (i)). Since J +C1 has the Dixmier
property (seen in the proof of (ii)), we have by Corollary 4.5 (ii) that
DJ+C1((ϕ2 + ψ2)J+C1) = DJ+C1(ϕ2J+C1) + DJ+C1(ψ2J+C1).
Further, by Proposition 3.5 (i), the same holds restricting ϕ2 and ψ2 to J:
DJ ((ϕ2 + ψ2)J ) = DJ (ϕ2J ) + DJ (ψ2J ).
Then we have
DA(ϕ + ψ) = DA(ϕ1 + ϕ2 + ψ1 + ψ2)
= DA(ϕ1 + ψ1) + DA(ϕ2 + ψ2)
= DA(ϕ1) + DA(ψ1) + DJ ((ϕ2 + ψ2)J ) ◦ ι∗
J
= DA(ϕ1) + DA(ψ1) + (DJ (ϕ2J ) + DJ (ψ2)) ◦ ι∗
J
= DA(ϕ1) + DA(ψ1) + DA(ϕ2) + DA(ψ2)
= DA(ϕ) + DA(ψ)
where we used Proposition 3.5 (i) in the third and fifth equalities, and Proposition
3.9 (the case that one of the functionals is tracial) in the second, third, and final
equalities.
(cid:3)
References
[1] Alberti, P. M. On maximally unitarily mixed states on W ∗-algebras. Math. Nachr. 91 (1979),
423 -- 430.
[2] Alberti, P. M.; Uhlmann, A. The order structure of states in C ∗ and W ∗-algebras. Proceed-
ings of the International Conference on Operator Algebras, Ideals, and their Applications in
Theoretical Physics (Leipzig, 1977), pp. 126 -- 135, Teubner, Leipzig, 1978
[3] Alberti, Peter M.; Uhlmann, Armin. Stochasticity and partial order. Doubly stochastic maps
and unitary mixing. With a preface by Michiel Hazewinkel. Mathematics and its Applications,
9. D. Reidel Publishing Co., Dordrecht-Boston, Mass., 1982. 123 pp.
[4] R.J. Archbold, An averaging process for C*-algebras related to weighted shifts, Proc. London
Math. Soc. (3), 35 (1977) 541 -- 554.
22
ARCHBOLD, ROBERT, AND TIKUISIS
[5] R. J. Archbold, L. Robert and A. Tikuisis, The Dixmier property and tracial states for C*-
algebras. J. Funct. Anal., to appear, arXiv:1611.08263v2 (2017) 55 pp.
[6] J. Dixmier, C*-algebras, North-Holland, Amsterdam, 1977.
[7] U. Haagerup and L. Zsidó, Sur la propriété de Dixmier pour les C*-algèbres, C. R. Acad. Sci.
Paris Sér. I Math., 298 (1984) 173 -- 176.
[8] Halpern, Herbert. Commutators in properly infinite von Neumann algebras. Trans. Amer.
Math. Soc. 139 1969 55 -- 73.
[9] G.J. Murphy, Uniqueness of the trace and simplicity, Proc. Amer. Math. Soc., 128 (2000)
3563 -- 3570.
[10] N. Ozawa, Dixmier approximation and symmetric amenability for C*-algebras, J. Math. Soc.
Univ. Tokyo, 20 (2013) 349 -- 374.
[11] G.K. Pedersen, C*-algebras and their automorphism groups, Academic Press, London, 1979.
[12] Uhlmann, A. Sätze über Dichtematrizen. (German) Wiss. Z. Karl-Marx-Univ. Leipzig Math.-
Natur. Reihe 20 (1971), 633 -- 637.
[13] Wehrl, A. How chaotic is a state of a quantum system? Rep. Math. Phys. 6 (1974), 15 -- 28.
Robert Archbold, Institute of Mathematics, University of Aberdeen, King's Col-
lege, Aberdeen AB24 3UE, Scotland, United Kingdom
E-mail address: [email protected]
Leonel Robert, Department of Mathematics, University of Louisiana at Lafayette,
Lafayette, 70504-3568, USA
E-mail address: [email protected]
Aaron Tikuisis, Department of Mathematics and Statistics, University of Ottawa,
585 King Edward, Ottawa, ON K1N 6N5, Canada
E-mail address: [email protected]
|
1410.7767 | 4 | 1410 | 2015-03-24T01:15:44 | Erratum to "Full and reduced C*-coactions". Math. Proc. Camb. Phil. Soc. 116 (1994), 435--450 | [
"math.OA"
] | This short note amends Proposition 2.5 of the named article, which states that a full coaction of a locally compact group on a C*-algebra is nondegenerate if and only if its normalization is. The proof given there of the reverse implication is incorrect and, unfortunately, we have been unable to find a correct proof. Instead, we summarize the current state of our knowledge of the relationships between nondegeneracy of normal, reduced, and generic C*-coactions. | math.OA | math |
ERRATUM TO "FULL AND REDUCED
C ∗-COACTIONS". MATH. PROC. CAMB. PHIL. SOC.
116 (1994), 435 -- 450.
S. KALISZEWSKI AND JOHN QUIGG
Proposition 2.5 of [Qui94] states that a full coaction of a locally
compact group on a C ∗-algebra is nondegenerate if and only if its nor-
malization is. Unfortunately, the proof there only addresses the for-
ward implication, and we have not been able to find a proof of the
opposite implication. This issue is important because the theory of
crossed-product duality for coactions requires implicitly that the coac-
tions involved be nondegenerate. Moreover, each type of coaction --
full, reduced, normal, maximal, and (most recently) exotic -- has its
own distinctive properties with respect to duality, making it crucial to
be able to convert from one to the other without losing nondegeneracy.
While it is generally believed that all coactions of all types are non-
degenerate, in this note we summarize what little is actually known
about nondegeneracy of C ∗-coactions. We also hope to caution the
reader that the error in [Qui94] has propagated widely, and sometimes
invisibly, in the literature. For example, a normal coaction is nonde-
generate if and only if the associated reduced coaction is nondegenerate
([Qui94, Proposition 3.3], which is independent of Proposition 2.5). So
since reduced coactions of discrete groups are automatically nondegen-
erate ([BS89, Corollaire 7.15]), it is often mistakenly assumed (as in
[Qui96] and [EKQR06]) that every full coaction of a discrete group
is also nondegenerate. An equivalent assumption (as in [KS13, Sec-
tion 2.4]) is that every C ∗-algebra that carries a coaction of a discrete
group is the closed span of its spectral subspaces.
An overview of the definitions and basic results concerning full coac-
tions, their normalizations, and their reductions can be found in Ap-
pendix A of [EKQR06].
Date: August 28, 2018.
2000 Mathematics Subject Classification. Primary 46L05.
Key words and phrases. full coaction, reduced coaction, nondegeneracy, normal-
ization, crossed product duality.
1
2
KALISZEWSKI AND QUIGG
Theorem 1. Let (A, δ) be a full coaction of a locally compact group G.
Then among the following conditions, we have the implications
(1) ⇒ (2) ⇔ (3):
(1) (A, δ) is nondegenerate.
(2) The normalization (An, δn) is nondegenerate.
(3) The reduction (Ar, δ r) is nondegenerate.
Lemma 2. Let (A, δ) and (B, ε) be full coactions of a locally compact
group G. If (A, δ) is nondegenerate and there exists a δ − ε equivariant
surjection ϕ : A → B, then (B, ε) is also nondegenerate.
Proof. By equivariance, ε(B) = ε(ϕ(A)) = (ϕ ⊗ id)(δ(A)), so
span ε(B)(1 ⊗ C ∗(G)) = span(ϕ ⊗ id)(δ(A))(1 ⊗ C ∗(G))
= span(ϕ ⊗ id)(cid:0)δ(A)(1 ⊗ C ∗(G))(cid:1)
= (ϕ ⊗ id)(cid:0)span δ(A)(1 ⊗ C ∗(G))(cid:1).
Since (A, δ) is nondegenerate, span δ(A)(1 ⊗ C ∗(G) = A ⊗ C ∗(G), so
span ε(B)(1 ⊗ C ∗(G)) = (ϕ ⊗ id)(A ⊗ C ∗(G)) = B ⊗ C ∗(G).
Thus (B, ε) is nondegenerate as well.
(cid:3)
Proof of Theorem 1. Since An is by definition a quotient of A, and δn
is defined so that the quotient map is δ − δn equivariant, (1) ⇒ (2) is
immediate from Lemma 2.
By Definition 3.5 of [Qui94], the reduction (Ar, δ r) of an arbitrary
coaction (A, δ) coincides with the reduction ((An)r, (δn)r) of the nor-
malization (An, δn). By Proposition 3.3 of [Qui94], (An, δn) is nonde-
generate if and only if ((An)r, (δn)r) is, and (2) ⇔ (3) follows.
(cid:3)
We reiterate that it is still an open question whether or not (2) im-
plies (1). We are grateful to Iain Raeburn for pointing out the error in
the original proof, and to Alcides Buss for drawing our attention to a
problem with our initial attempt to correct it.
References
[BS89]
S. Baaj and G. Skandalis, C ∗-alg`ebres de Hopf et th´eorie de Kasparov
´equivariante, K-Theory 2 (1989), 683 -- 721.
[EKQR06] S. Echterhoff, S. Kaliszewski, J. Quigg, and I. Raeburn, A Categorical
Approach to Imprimitivity Theorems for C*-Dynamical Systems, vol.
180, Mem. Amer. Math. Soc., no. 850, American Mathematical Society,
Providence, RI, 2006.
B. K. Kwa´sniewski and W. Szyma´nski, Topological aperiodicity for prod-
uct systems over semigroups of Ore type, preprint, 2013.
[KS13]
[Lan79] M. B. Landstad, Duality theory for covariant systems, Trans. Amer.
Math. Soc. 248 (1979), 223 -- 267.
ERRATUM TO "FULL AND REDUCED C ∗-COACTIONS"
3
[Qui94]
[Qui96]
[Rae92]
J. C. Quigg, Full and reduced C ∗-coactions, Math. Proc. Cambridge
Philos. Soc. 116 (1994), 435 -- 450.
, Discrete C ∗-coactions and C ∗-algebraic bundles, J. Austral.
Math. Soc. Ser. A 60 (1996), 204 -- 221.
I. Raeburn, On crossed products by coactions and their representation
theory, Proc. London Math. Soc. 64 (1992), 625 -- 652.
School of Mathematical and Statistical Sciences, Arizona State
University, Tempe, Arizona 85287
E-mail address: [email protected]
School of Mathematical and Statistical Sciences, Arizona State
University, Tempe, Arizona 85287
E-mail address: [email protected]
|
1107.5244 | 4 | 1107 | 2012-09-17T19:10:37 | Completely positive multipliers of quantum groups | [
"math.OA",
"math.FA"
] | We show that any completely positive multiplier of the convolution algebra of the dual of an operator algebraic quantum group $\G$ (either a locally compact quantum group, or a quantum group coming from a modular or manageable multiplicative unitary) is induced in a canonical fashion by a unitary corepresentation of $\G$. It follows that there is an order bijection between the completely positive multipliers of $L^1(\G)$ and the positive functionals on the universal quantum group $C_0^u(\G)$. We provide a direct link between the Junge, Neufang, Ruan representation result and the representing element of a multiplier, and use this to show that their representation map is always weak$^*$-weak$^*$-continuous. | math.OA | math |
Completely positive multipliers of quantum groups
Matthew Daws
November 2, 2018
Abstract
We show that any completely positive multiplier of the convolution algebra of the
dual of an operator algebraic quantum group G (either a locally compact quantum group,
or a quantum group coming from a modular or manageable multiplicative unitary) is
induced in a canonical fashion by a unitary corepresentation of G. It follows that there is
an order bijection between the completely positive multipliers of L1(G) and the positive
functionals on the universal quantum group C u
0 (G). We provide a direct link between the
Junge, Neufang, Ruan representation result and the representing element of a multiplier,
and use this to show that their representation map is always weak∗-weak∗-continuous.
Keywords: Locally compact quantum group, manageable multiplicative unitary, com-
pletely bounded multiplier, completely positive multiplier, corepresentation.
MSC classification (2010): 20G42, 22D10, 22D15, 43A22, 46L07, 46L89, 81R50.
1
Introduction
Multipliers arise throughout the study of algebras in analysis, as a useful tool for embedding a
non-unital algebra into its "largest" unitisation. In abstract harmonic analysis, the convolution
algebra L1(G) of a locally compact group G is unital only when G is discrete; Wendel's Theorem
tells us that the multiplier algebra of L1(G) is the measure algebra M(G). On the "Fourier
transform" side, the Fourier algebra A(G) is unital only for compact G. Here it seems most
profitable to study the completely bounded multipliers, as these are more tractable, have better
functorial properties, and can capture interesting geometric aspects of the group G, see [7, 11,
28] for example.
The convolution algebras of locally compact quantum groups G form a generalisation of
both L1(G) and A(G), together with genuinely "non-commutative" examples, such as compact
quantum groups. Again, completely bounded multipliers are profitable to study, and in par-
ticular, it was shown by Junge, Neufang and Ruan in [17] that there is a bijection between
"abstract" completely bounded (left) multipliers of the dual L1( G) (that is, right module maps
on L1( G)) and "concrete" multipliers -- elements of L∞(G) which "multiply" the image of L1( G)
into itself. This mirrors the Fourier algebra picture, where there is a bijection between com-
pletely bounded multipliers of A(G), thought of as module maps on A(G), and continuous
functions f on G which multiply the function algebra A(G) into itself.
It was shown by De Canniere and Haagerup, [11, Proposition 4.2], that completely positive
multipliers of A(G) biject with positive definite functions on G. Positive definite functions arise
from unitary group representations, equivalently, ∗-representations of the full group C∗-algebra.
In this light, we can also observe that positive multipliers of L1(G) arise from positive measures
in M(G), or equivalently, ∗-representations of C0(G). Similarly, the use of completely positive
multipliers of quantum groups has arisen in the study of various approximation properties
associated to quantum group algebras; see for example [14], and in relation to the Haagerup
approximation property, [4, 5].
1
The main objective of this paper is to establish the same result for quantum groups: any
completely positive multiplier of L1( G) comes from a unitary corepresentation of G, or equiv-
0 ( G) (to be thought of as
alently, from a ∗-representation of the universal quantum group C u
generalising the role of C ∗(G).) We remark that such a result cannot in general be extended
to completely bounded multipliers, as unpublished work of Losert, extending his paper [23],
shows that G is amenable if and only if the Fourier-Stieltjes algebra B(G) = C ∗(G)∗ agrees with
the algebra of completely bounded multipliers of A(G). (This was shown for various explicit
examples in [7, 11] and for all discrete groups in [3].)
A key tool for us is the representation result of [17], where it is shown that a completely
bounded multiplier of L1( G) extends to a normal, completely bounded map on B(L2(G)) with
certain properties. In the completely positive case, applying the Stinespring construction to
this map yields a Hilbert space, and it is this space which will be the carrier for our unitary
corepresentation.
Thus a completely bounded (left) multiplier gives rise to both a normal map Φ on B(L2(G))
and a "representing element" a0 ∈ L∞(G). We show that these are linked by the relation
(cid:0)Φ(θξ,η)α(cid:12)(cid:12)β(cid:1) = h∆(a0), ωα,η ⊗ ω♯∗
ξ,βi.
Here θξ,η is the compact operator α 7→ (αη)ξ, and ♯ denotes the (in general unbounded) ∗-
operation on L1(G). See Proposition 6.1 for a precise statement. We use this result to show
that the representation map of [17] is always weak∗-weak∗ continuous, which answers in the
affirmative the question asked before [16, Theorem 4.7]. Furthermore, this suggests a possible
definition for a "positive definite function" on a quantum group: namely, those a0 such that
Φ can be chosen to be completely positive.
In joint work with Salmi, [10], we verify this
conjecture -- namely, it is indeed the case that if Φ can be chosen to be completely positive,
then automatically Φ must come from a completely positive multiplier, and so a0 comes from
a corepresentation of G.
The organisation of the paper is as follows. In Section 2 we quickly survey the Operator
Algebraic quantum groups which we shall study, and we prove some useful auxiliary results.
In this paper, all our arguments work in the slightly more general setting of quantum groups
coming from manageable (or modular) multiplicative unitaries (that is, without assuming in-
variant weights). In Section 3 we recall the construction of [17]. As the actual constructions
involved, and not just the results of that paper, are indispensable to us, we sketch some of the
proofs, and take the opportunity to show how the arguments can be made to work for quantum
groups coming from manageable multiplicative unitaries.
In Section 4 we show how certain
corepresentations of G canonically give rise to completely bounded multipliers of L1( G). We
also show how unitary corepresentations, equivalently, ∗-representations of the universal dual
0 ( G), give rise to completely positive multipliers of L1( G). The following section is
algebra C u
the main part of the paper, and proves the converse -- any completely positive multiplier of
L1( G) comes from a unitary corepresentation of G in the manner discussed in Section 4. As
mentioned above, we construct the Hilbert space for our corepresentation from the Stinespring
dilation construction. We can view this as a completion of L1(G) with respect to a certain
inner-product; then the corepresentation U is built by actually forming U ∗, which is linked to
the anti-representation of L1(G) given by right multiplication of L1(G) on itself. We then pro-
ceed to show that U is in fact unitary, and then that we can recover the original multiplier from
U. In the final section, we explore, as discussed above, the link between multipliers as maps on
L1(G) (or, via [17], as maps on B(L2(G))) and their "representing elements" in L∞(G).
A final word on notation. Our Hilbert space inner products shall be linear in the first
variable, and we write (··) for an inner product (or more generally, a sesquilinear form). We
write h·, ·i for the bilinear pairing between a Banach space and its dual. For a Hilbert space
H, we write B(H) for the algebra of all bounded operators on H, write B(H)∗ for its predual
2
(the trace class operators) and write B0(H) for the ideal of compact operators. For ξ, η ∈ H,
we denote by ωξ,η the normal functional in B(H)∗, and by θξ,η the rank one operator in B0(H),
which are defined by
hT, ωξ,ηi = (T ξη),
θξ,η(γ) = (γη)ξ
(T ∈ B(H), γ ∈ H).
Given a normal map T on a von Neumann algebra M, we write T∗ for the pre-adjoint of T
acting on the predual M∗. We write ⊗ to mean a completed tensor product, either of Hilbert
spaces, or the minimal C∗-algebraic tensor product. We write ⊗ for the von Neumann algebraic
tensor product, and ⊙ for the purely algebraic tensor product. We write Σ for the tensor swap
map of Hilbert spaces, say Σ : H ⊗ H → H ⊗ H; ξ ⊗ η 7→ η ⊗ ξ.
We use the basic theory of Operator Spaces without comment; see, for example, [12] for
further details.
Acknowledgements: The author wishes to thank Michael Brannan for asking the initial
question which lead to this paper, and to thank Zhong-Jin Ruan for pointing out that the
argument at the end of Proposition 3.2 needed more justification. The anonymous referee of
an earlier version of this paper made many helpful comments;
in particular suggesting the
argument of Theorem 2.2.
2 Operator algebraic quantum groups
In this paper, we shall be concerned with quantum groups in the operator algebraic setting -- to
be precise, either locally compact quantum groups, in the Kustermans, Vaes sense [20, 21, 22,
30], or C∗-algebraic quantum groups built from manageable or modular multiplicative unitaries,
in the So ltan, Woronowicz sense [25, 26, 33] (the latter generalising the former). In fact, for
many of our results, we shall need remarkably little -- our main tool being that "invariants are
constant" (see below -- our inspiration here is [24, Section 2]).
A locally compact quantum group in the von Neumann algebraic setting is a Hopf-von Neu-
mann algebra (M, ∆) equipped with left and right invariant weights. As usual, we use ∆ to
turn M∗ into a Banach algebra, and we write the product by juxtaposition. We shall "work on
the left"; so using the left invariant weight, we build the GNS space H, and a multiplicative
unitary W acting on L2(G) ⊗ L2(G) (of course, the existence of a right weight is needed to
show that W is unitary). There is a (in general unbounded) antipode S which admits a "polar
decomposition" S = Rτ−i/2, where R is the unitary antipode, and (τt) is the scaling group.
There is a nonsingular positive operator P which implements (τt) as τt(x) = P itxP −it. Then
W is manageable with respect to P .
A manageable multiplicative unitary W acting on H ⊗ H has, by definition, a nonsingular
positive operator P , and an operator W acting on H ⊗ H such that
(cid:0)W (ξ ⊗ α)(cid:12)(cid:12)η ⊗ β(cid:1) = (cid:0) W (P −1/2ξ ⊗ β)(cid:12)(cid:12)P 1/2η ⊗ α(cid:1),
for all α, β ∈ H and ξ ∈ D(P −1/2), η ∈ D(P 1/2). A word on notation: we work with left
multiplicative unitaries, whereas So ltan and Woronowicz, in the conventions of [22], work with
right multiplicative unitaries, and so we have translated everything to the left.
Given such a W , the space {(ι ⊗ ω)W : ω ∈ B(H)∗} is an algebra, and its closure is a C∗-
algebra, say A. There is a coassociative map ∆ : A → M(A ⊗ A) given by ∆(a) = W ∗(1 ⊗ a)W .
If we formed W from (M, ∆) with invariant weights, then A is σ-weakly dense in M, and
the two definitions of ∆ agree. Similarly, {(ω ⊗ ι)W : ω ∈ B(H)∗} is norm dense in a C∗-
algebra A, and defining ∆(a) = W ∗(1 ⊗ a) W , we get a non-degenerate ∗-homomorphism
∆ : A → M( A ⊗ A), where here W = ΣW ∗Σ.
If we started with (M, ∆) having invariant
3
weights, then we can construct invariant weights on ( A′′, ∆). The unitary W is in the multiplier
algebra M(A ⊗ A) ⊆ B(H ⊗ H).
When W is a manageable multiplicative unitary, we can still form S, R and (τt) with the
usual properties. The antipode S has elements of the form (ι⊗ω)W as a core, and S((ι⊗ω)W ) =
(ι ⊗ ω)(W ∗). Then also S = Rτ−i/2 and (τt) is again implemented by P (the same map which
appears in the definition of "manageable").
There is a more general notation of a modular multiplicative unitary, see [26], which is more
natural in certain examples. However, at the cost of changing our space H, we can recover
(A, ∆) from a different, but related, manageable multiplicative unitary. Indeed, in [25], it is
shown that if (A, ∆) is given by some modular multiplicative unitary, then ( A, ∆), the σ-weak
topologies on A and A, the image of W in M(A ⊗ A), and all the maps S, R, (τt), S, R and
(τt), are independent of the particular choice of modular multiplicative unitary giving (A, ∆).
For this reason, we shall henceforth work only with manageable multiplicative unitaries (but
of course all our results hold in the modular case as well).
We write G for an abstract object to be thought of as a quantum group. We write
C0(G), L∞(G) and L1(G) for A, M and M∗ (and similar for the dual objects); as mentioned
in the previous paragraph, these are well-defined. We also write L2(G) for H, but be aware
that if G is given by a modular or manageable multiplicative unitary, then there is some arbi-
trary choice involved in L2(G). If G has invariant weights, then these weights unique up to a
constant, and so L2(G) is unique.
This concludes our brief summary; we shall develop further theory as and when we need it.
We finish this section with one of our major tools -- that "invariants are constant". Notice that,
by using the unitary antipode, we could replace y13 by y23 in the following; but we shall have
no need of this variant.
Theorem 2.1. For any G and a von Neumann algebra N, if x, y ∈ L∞(G)⊗N satisfy (∆⊗ι)x =
y13, then x = y ∈ C⊗N.
Proof. We shall prove this when N = C, the general case comes from considering (ι ⊗ ω)x
and (ι ⊗ ω)y, as ω ∈ N∗ varies. For locally compact quantum groups, this was shown in [1,
Lemma 4.6], compare also [21, Result 5.13]. For general G, [24, Theorem 2.6] shows that if
a, b ∈ B(L2(G)) with W ∗(1 ⊗ a)W = b ⊗ 1 (working with left multiplicative unitaries) then
a = b ∈ C1, and this immediately implies the result.
The following result is, when we have invariant weights, well-known to experts, but proofs
can be hard to find (for example, [17] cites [32, Proposition 3.5], which uses Tomita-Takesaki
theory; compare also the proof of [31, Proposition 5.13]). The argument here was suggested to
us by the anonymous referee of an earlier version of this paper.
Theorem 2.2. Let G be a quantum group. Then the linear span of L∞( G)L∞(G)′ = {xy′ :
x ∈ L∞( G), y′ ∈ L∞(G)′} is σ-weakly dense in B(L2(G)).
Proof. Let N be the σ-weak closed linear span of L∞( G)L∞(G)′, and let P be the von Neumann
algebra generated by N. We first claim that N = P ; we need only show that P ⊆ N. In fact,
it suffices to show that L∞(G)′L∞( G) ⊆ N, as then N will be a bimodule over both L∞( G)
and L∞(G)′, and hence will contain the algebra generated by L∞( G) and L∞(G)′.
Let y′ ∈ L∞(G)′, let ξ, η ∈ L2(G) and set x = (ι ⊗ ωξ,η)( W ) ∈ L∞( G). Letting (ei) be an
orthonormal basis for L2(G), we have that
= X
i
y′x = y′(ι ⊗ ωξ,η)( W ) = (ι ⊗ ωξ,η)( W W ∗(y′ ⊗ 1) W )
(ι ⊗ ωei,η)( W )(ι ⊗ ωξ,ei)( W ∗(y′ ⊗ 1) W ),
4
the sum converging σ-weakly. However, for ω ∈ B(L2(G))∗ and y ∈ L∞(G),
(ι ⊗ ω)( W ∗(y′ ⊗ 1) W )y = (ω ⊗ ι)(W (1 ⊗ y′)W ∗(1 ⊗ y)) = (ω ⊗ ι)(W (1 ⊗ y′)∆(y)W )
= (ω ⊗ ι)(W ∆(y)(1 ⊗ y′)W ) = (ω ⊗ ι)((1 ⊗ y)W (1 ⊗ y′)W )
= y(ι ⊗ ω)( W ∗(y′ ⊗ 1) W ).
It follows that (ι ⊗ ω)( W ∗(y′ ⊗ 1) W ) ∈ L∞(G)′, and so y′(ι ⊗ ωξ,η)( W ) ∈ N. As such x are
σ-weakly dense in L∞( G), the claim follows.
Now let z ∈ P ′ = L∞( G)′ ∩ L∞(G). Then ∆(z) = W ∗(1 ⊗ z)W = 1 ⊗ z as W ∈
L∞(G)⊗L∞( G). By applying Theorem 2.1 to R(z) it follows that z ∈ C1. Thus N = P =
(C1)′′ = B(L2(G)) as required.
3 Multipliers of quantum groups
In this section, we review some of the ideas used by Junge, Neufang and Ruan in [17]. We shall
actually need some constructions coming from the proofs in [17] (and not just the statements
of the results). Rather than just give sketch proofs, we instead give quick, full proofs, and
take the opportunity to show that some of their results also hold for quantum groups coming
from manageable multiplicative unitaries. Further details and related ideas can be found in
[8, 9, 16, 17].
Definition 3.1. A completely bounded left multiplier of L1(G) is a completely bounded map
L∗ : L1(G) → L1(G) with L∗(ω1ω2) = L∗(ω1)ω2 for ω1, ω2 ∈ L1(G).
Such maps are also often called "centralisers" in the literature (and in particular, in [17]).
A simple calculation shows that a completely bounded map L∗ : L1(G) → L1(G) is a left
multiplier if and only if its adjoint L = (L∗)∗ satisfies (L ⊗ ι)∆ = ∆L.
Let us make a few remarks about normal completely bounded maps. As explained, for
example, in the proof of [15, Theorem 2.5], as L is normal, we can find a normal ∗-representation
π : L∞(G) → B(H) for some Hilbert space H, and bounded maps P, Q : L2(G) → H, with
L(x) = P ∗π(x)Q for each x ∈ L∞(G). By the structure theory for normal ∗-representations (see
[29, Theorem 5.5, Chapter IV]) by adjusting P and Q, and may suppose that H = L2(G) ⊗ H ′
for some Hilbert space H ′, and that π(x) = x ⊗ 1. For example, it then follows that
(L ⊗ ι)( W ) = (P ∗ ⊗ 1) W13(Q ⊗ 1).
As W ∈ M(B0(L2(G))⊗C0(G)), it follows easily from this that also (L⊗ι)( W ) ∈ M(B0(L2(G))⊗
C0(G)), a fact we shall use in the following proof.
The following is a short unification of (the left version of) [17, Corollary 4.4] (compare [17,
Theorem 4.10]) and [8, Theorem 4.2]; we make use of the "invariants are constant" technique.
Remember that the left regular representation of G is the injective homomorphism λ :
L1(G) → C0( G); ω 7→ (ω ⊗ ι)(W ). By duality we define λ : L1( G) → C0(G); ω 7→ (ω ⊗ ι)( W ) =
(ι ⊗ ω)(W ∗).
Proposition 3.2. Let L∗ be a completely bounded left multiplier of L1( G). There is a ∈
M(C0(G)) with aλ(ω) = λ(L∗(ω)) for ω ∈ L1(G), or equivalently, with (1 ⊗ a) W = (L ⊗ ι)( W ).
Proof. That aλ(ω) = λ(L∗(ω)) for each ω ∈ L1(G), if and only if (1 ⊗ a) W = (L ⊗ ι)( W )
follows easily from the definition of λ. Consider now
( ∆ ⊗ ι)(cid:0)(L ⊗ ι)( W ) W ∗(cid:1) = (cid:0)(L ⊗ ι ⊗ ι)( ∆ ⊗ ι)( W )(cid:1)( ∆ ⊗ ι)( W ∗)
W ∗
13
= (cid:0)(L ⊗ ι ⊗ ι)( W13 W23)(cid:1) W ∗
= (L ⊗ ι)( W )13 W ∗
13,
23
5
where we have used that ∆L = (L ⊗ ι) ∆, that ∆ is a ∗-homomorphism, and that ( ∆ ⊗ ι)( W ) =
W13 W23. By Theorem 2.1, it follows that there is a ∈ L∞(G) with (L ⊗ ι)( W ) W ∗ = 1 ⊗ a.
However, as W ∈ M(B0(L2(G)) ⊗ C0(G)), we see that
1 ⊗ a = (L ⊗ ι)( W ) W ∗ ∈ M(B0(L2(G)) ⊗ C0(G)),
from which it follows immediately that a ∈ M(C0(G)).
In the language of [8], the previous lemma says that L∗ is "represented"; in the language of
[17], the element a is the "multiplier" associated to the "centraliser" L∗.
Proposition 3.3. Let L∗ be a completely bounded left multiplier of L1( G). There is a completely
bounded, normal map Φ : B(L2(G)) → B(L2(G)) which extends L, and which is a L∞(G)′-
bimodule map. Indeed, Φ satisfies
1 ⊗ Φ(x) = W(cid:0)(L ⊗ ι)( W ∗(1 ⊗ x) W )(cid:1) W ∗
(x ∈ B(L2(G))).
Proof. We closely follow [17, Proposition 4.3], while translating "to the left" and using that
"invariants are constant". Define
T : B(L2(G)) → L∞( G)⊗B(L2(G)),
T (x) = W(cid:0)(L ⊗ ι)( W ∗(1 ⊗ x) W )(cid:1) W ∗.
We now perform a similar calculation to that in the previous lemma:
( ∆ ⊗ ι)T (x) = W13 W23(L ⊗ ι ⊗ ι)(cid:0)( ∆ ⊗ ι)( W ∗(1 ⊗ x) W )(cid:1) W ∗
23
W ∗
13
= W13 W23(L ⊗ ι ⊗ ι)(cid:0) W ∗
= W13(L ⊗ ι ⊗ ι)(cid:0) W ∗
23
= T (x)13.
W ∗
13(1 ⊗ 1 ⊗ x) W13 W23(cid:1) W ∗
23
W ∗
13
13(1 ⊗ 1 ⊗ x) W13(cid:1) W ∗
13
So by Theorem 2.1, there is Φ(x) ∈ B(L2(G)) with T (x) = 1 ⊗ Φ(x). It is easy to see that Φ is
completely bounded and normal.
For x ∈ L∞( G)
1 ⊗ Φ(x) = T (x) = W ((L ⊗ ι) ∆(x)) W ∗ = W ∆(L(x)) W ∗ = 1 ⊗ L(x),
and so Φ extends L. For y, z ∈ L∞(G)′, as W ∈ L∞( G)⊗L∞(G), it is easy to see that
T (yxz) = (1 ⊗ y)T (x)(1 ⊗ z)
(x ∈ B(L2(G))),
and so Φ(yxz) = yΦ(x)z as required.
In the language of [17], we have thus constructed a map from the set of completely bounded
left multipliers of L1( G) to CBσ,L∞( G)
L∞(G)′ (B(L2(G))). It seems that, to continue with the arguments
of [17], we start to need to use arguments that involve the relative position of L∞(G) and its
commutant in B(L2(G)). In particular, to show that every Φ ∈ CBσ,L∞( G)
L∞(G)′ (B(L2(G))) comes
from a left multiplier would require us to know that L∞(G) ∩ L∞( G) = C (at least if one is
following the proof of [17, Proposition 3.2]), and we have no proof of this in the Manageable
Multiplicative Unitary setting.
6
4 Multipliers coming from invertible corepresentations
In this section, we show, rather explicitly, how corepresentations and "universal" quantum
groups give rise to completely bounded and completely positive multipliers.
A corepresentation of G shall be, for us, an element U ∈ L∞(G)⊗B(H) with (∆ ⊗ ι)(U) =
U13U23 (so, we don't assume that U is unitary). We state the following in a little generality,
but note that it obviously applies to unitary corepresentations. Similar ideas are explored in
[8, Section 6], and for Kac algebras, with much less emphasis on corepresentation theory, see
[18].
Proposition 4.1. Let U be a corepresentation of G, and suppose there is V ∈ B(L2(G) ⊗ H)
with V U ∗ = 1 (that is, U has a right inverse). For each α, β ∈ H, there is a completely bounded
left multiplier of L1( G) represented by a = (ι ⊗ ωα,β)(U ∗). If U ∗ is an isometry (so we may
take V = U) and α = β, then the multiplier is completely positive.
Proof. We have that (∆ ⊗ ι)(U ∗) = U ∗
V23W ∗
define L : L∞( G) → B(L2(G)) by
23 = U ∗
13W ∗
12U ∗
23 = U ∗
12, and using that W = ΣW ∗Σ, it follows that V13 W12U ∗
13, or equivalently, W ∗
12U ∗
23U ∗
13W ∗
12. Thus also
W12. Thus
13 = U ∗
23
23U ∗
L(x) = (ι ⊗ ωα,β)(V (x ⊗ 1)U ∗)
(x ∈ L∞( G)).
Clearly L is a normal, completely bounded map. Then immediately we see that (L ⊗ ι)( W ) =
(1 ⊗ a) W , and it is now easy to see (compare [8, Proposition 2.3]) that L maps into L∞( G),
and that L is the adjoint of a left multiplier on L1( G), represented by a.
When U ∗ is an isometry, V = U and α = β, clearly L is completely positive.
For U, V as in the proposition, we could weaken the condition on U to asking that U ∈
12U23W12 = U13U23. Then, arguing as in [33, Page 142], we see that
23 ∈ M(C0(G) ⊗ B0(L2(G)) ⊗ B0(H)), and so U ∈ M(C0(G) ⊗ B0(H)), in
B(L2(G) ⊗ H) with W ∗
U13 = W ∗
particular, U is a corepresentation in our sense.
12U23W12V ∗
Let us just remark that if also V is a corepresentation, then consider forming Φ as in
Section 3, using the L given as in the proposition. So, for x ∈ B(L2(G)),
1 ⊗ Φ(x) = (ι ⊗ ωα,β ⊗ ι)(cid:0) W13V12 W ∗
= (ι ⊗ ι ⊗ ωα,β)(cid:0) W12V13 W ∗
13(1 ⊗ 1 ⊗ x) W13U ∗
12(1 ⊗ x ⊗ 1) W12U ∗
13
12
W ∗
W ∗
13(cid:1)
12(cid:1).
Now, W (a ⊗ 1) W ∗ = ΣW ∗(1 ⊗ a)W Σ = Σ∆(a)Σ for a ∈ L∞(G), and so
1 ⊗ Φ(x) = (ι ⊗ ι ⊗ ωα,β)(cid:0)V23V13(1 ⊗ x ⊗ 1)U ∗
13U ∗
23(cid:1) = 1 ⊗ (ι ⊗ ωα,β)(cid:0)V (x ⊗ 1)U ∗(cid:1).
Hence Φ has the same "defining formula" as L.
4.1 Links with universal quantum groups
Universal quantum groups are constructed in [19] and [26, Section 5]. We write C u
universal dual of C0(G). For us, the important properties are that:
0 ( G) for the
• There is a coassociative non-degenerate ∗-homomorphism ∆u : C u
0 ( G) → M(C u
0 ( G) ⊗
0 ( G));
C u
• There is a surjective ∗-homomorphism πu : C u
0 ( G) → C0( G) with ∆πu = (πu ⊗ πu) ∆u;
• There is a unitary corepresentation W ∈ M(C0(G) ⊗ C u
0 ( G)) of C0(G) such that (ι ⊗
πu)W = W and (ι ⊗ ∆u)W = W13W12.
7
• The space {(ω ⊗ ι)(W) : ω ∈ L1(G)} is dense in C u
0 ( G).
• There is a bijection between unitary corepresentations U of C0(G) and non-degenerate
∗-homomorphisms π : C u
0 ( G) → B(H) given by the relation that U = (ι ⊗ π)(W).
Note that our W is denoted by V in the notation of [19]; and is the "left analogue" of W in the
notation of [26].
The map π∗
u : L1( G) → C u
for example [9, Proposition 8.3]) that this identifies L1( G) with an ideal in C u
that members of C u
0 ( G)∗ is an isometry and an algebra homomorphism. We know (see
0 ( G)∗, and hence
0 ( G)∗ induce multipliers on L1( G). Let us make links with Proposition 4.1.
Proposition 4.2. Let U be a unitary corepresentation of G on H, and let α, β ∈ H. Let π
0 ( G) on H associated with U. Then the multiplier represented by
be the ∗-representation of C u
(ι ⊗ ωα,β)(U ∗) is given by left multiplication by µ = ωα,β ◦ π ∈ C u
0 ( G)∗.
Proof. Let L : L∞( G) → L∞( G) be the adjoint of the completely bounded left multiplier
represented by a = (ι ⊗ ωα,β)(U ∗), as constructed in Proposition 4.1. Then (L ⊗ ι)( W ) =
(1 ⊗ a) W , or equivalently, (ι ⊗ L)(W ∗) = (a ⊗ 1)W ∗.
Define L†(x) = L(x∗)∗ for x ∈ L∞( G), so L† is a normal, completely bounded map on
L∞( G). For any von Neumann algebra M and X ∈ M⊗L∞( G), we see that (ι ⊗ L†)(X ∗) =
In particular, it follows that (L† ⊗ ι) ∆ = ∆L†, and so L† is the adjoint of a
(ι ⊗ L)(X)∗.
completely bounded left multiplier of L1( G), represented by b say. The proof of Proposition 4.1
shows that b = (ι ⊗ ωβ,α)(U ∗).
Given ω ∈ L1( G), we wish to show that µπ∗
u(ω) = π∗
u(L∗(ω)). Let ω ∈ L1(G) and set
x = (ω ⊗ ι)W ∈ C u
0 ( G). Then
hµπ∗
u(ω), xi = hµ ⊗ π∗
u(ω), ∆u((ω ⊗ ι)W)i = hω ⊗ µ ⊗ π∗
u(ω), W13W12i
= hω ⊗ ωα,β ⊗ ω, (ι ⊗ πu)(W)13(ι ⊗ π)(W)12i = hW13U12, ω ⊗ ωα,β ⊗ ωi,
and also
hπ∗
u(L∗(ω)), xi = h(ι ⊗ πu)W, ω ⊗ L∗(ω)i = hW, ω ⊗ L∗(ω)i = h(ι ⊗ L)(W ), ω ⊗ ωi.
Now, (ι ⊗ L)(W ) = (ι ⊗ L†)(W ∗)∗ = ((b ⊗ 1)W ∗)∗ = W (b∗ ⊗ 1), and so, using that b∗ =
(ι ⊗ ωα,β)(U), we have that
hπ∗
u(L∗(ω)), xi = hW (b∗ ⊗ 1), ω ⊗ ωi = hW13U12, ω ⊗ ωα,β ⊗ ωi.
As such x are dense in C u
0 ( G), the proof is complete.
In particular, taking U = W, we see that every positive functional on C u
0 ( G) induces a
completely positive left multiplier of L1( G). The main result of this paper is to show that the
converse is also true.
5 Completely positive multipliers
In this section, we study completely positive multipliers of L1( G). Motivated by Proposition 3.3,
we will first study completely positive normal maps on B(L2(G)). As B0(L2(G)) is σ-weakly-
dense in B(L2(G)), it suffices to consider completely positive maps B0(L2(G)) → B(L2(G)).
The following is simply the Stinespring construction, tailored to this specific situation. We give
the details, as they are central to our argument. For ξ, η ∈ L2(G), let θξ,η ∈ B0(L2(G)) be the
rank-one operator α 7→ (αη)ξ.
8
Let Φ : B0(L2(G)) → B(L2(G)) be a completely positive map. We remark that Φ has a
unique normal extension to a normal completely positive map B(L2(G)) → B(L2(G)), which we
shall also denote by Φ. Let H be the completion of the algebraic tensor product L2(G) ⊙ L2(G)
for the pre-inner-product
(cid:0)ξ ⊗ α(cid:12)(cid:12)η ⊗ β(cid:1)H = (cid:0)Φ(θη,ξ)αβ(cid:1).
That this is a positive sesquilinear form follows from the fact that Φ is completely positive
(compare with [29, Theorem 3.6] for example). We shall write να,ξ for the equivalence class of
ξ ⊗ α in H; see the end of the following paragraph for an explanation of this notation.
Let (ei) be an orthonormal basis of L2(G), and define
V : L2(G) → L2(G) ⊗ H; α 7→ X
ei ⊗ να,ei.
i
This makes sense, that is, the sum converges, as
X
i
kνα,eik2
H = X
i
(Φ(θei,ei)αα) = (Φ(1)αα).
Then, for α, β, ξ, η ∈ L2(G),
(V ∗(θξ,η ⊗ 1)V αβ) = X
= X
i,j
i
(θξ,η(ei) ⊗ να,eiej ⊗ νβ,ej )H
(eiη)(ξej)(Φ(θej ,ei)αβ) = (Φ(θξ,η)αβ).
Thus we have a Stinespring dilation of Φ. Now let (fi) be an orthonormal basis of H, and
define a family (ai) in B(L2(G)) by setting
V (α) = X
i
ai(α) ⊗ fi ∈ L2(G) ⊗ H
(α ∈ L2(G)).
It follows that Φ(x) = Pi a∗
(να,ξνβ,η)H = X
i xai for each x ∈ B(L2(G)). Furthermore, for ξ, η, α, β ∈ L2(G),
i θη,ξaiαβ) = X
(a∗
(aiαξ)(ηaiβ) = X
hai, ωα,ξihai, ωβ,ηi,
i
i
i
and so να,ξ = Pihai, ωα,ξifi in H. This explains the choice of notation να,ξ, which is deliberately
reminiscent of ωα,ξ.
Proposition 5.1. In the above setting, suppose further that M is a von Neumann algebra on
L2(G), and that Φ is an M-bimodule map. Then (x ⊗ 1)V = V x for each x ∈ M, and ai ∈ M ′
for each i.
Proof. For readability, we drop the ν notation in this proof. For x ∈ M and ξ, η, α, β ∈ L2(G),
(x∗(ξ) ⊗ αη ⊗ β)H = (Φ(θη,ξx)αβ) = (Φ(θη,ξ)xαβ) = (ξ ⊗ x(α)η ⊗ β)H.
Thus x∗(ξ) ⊗ α = ξ ⊗ x(α) in H. It follows that
V (x(α)) = X
= X
i
ei ⊗ (ei ⊗ x(α)) = X
(x(ej)ei)ei ⊗ (ej ⊗ α) = X
i
ei ⊗ (x∗(ei) ⊗ α) = X
i,j
i,j
j
x(ej) ⊗ (ej ⊗ α) = (x ⊗ 1)V (α),
ei ⊗ ((x∗(ei)ej)ej ⊗ α)
remembering that H is the completion of L2(G) ⊗ L2(G). It now follows that xai = aix for
each i, and so as x ∈ M was arbitrary, ai ∈ M ′ for each i.
The previous result (in the more general completely bounded setting) is well-known, see for
example [27, Theorem 3.1] and unpublished work of Haagerup. However, the actual construc-
tion will be central to our arguments.
9
5.1 Constructing a corepresentation
For the remainder of this section, fix a completely positive left multiplier on L1( G). Form
Φ : B(L2(G)) → B(L2(G)) using Proposition 3.3 applied to this multiplier, and apply the
construction of the previous section to find H and V : L2(G) → L2(G) ⊗ H. Fixing an
i xai for each x ∈ B(L2(G)). By
Proposition 5.1, we see that (x ⊗ 1)V = V x for each x ∈ L∞(G)′, equivalently, that ai ∈ L∞(G)
for each i.
orthonormal basis (fi) for H, we find ai such that Φ(x) = Pi a∗
Proposition 5.2. There is a unique isometry U ∗ on L2(G) ⊗ H which satisfies
U ∗(cid:0)ξ ⊗ να,η(cid:1) = X
i
(ωα,η ⊗ ι)∆(ai)ξ ⊗ fi,
for all ξ, η, α ∈ L2(G).
Proof. We know that Φ(x) = Pi a∗
tells us that
i xai for x ∈ L∞( G). We now use Proposition 3.3, which
1 ⊗ Φ(x) = X
i
W (a∗
i ⊗ 1) W ∗(1 ⊗ x) W (ai ⊗ 1) W ∗
(x ∈ B0(L2(G))).
For ξ1, η1, α1, ξ2, η2, α2 ∈ L2(G), we hence have that
i ⊗ 1) W ∗(1 ⊗ θη2,η1) W (ai ⊗ 1) W ∗(ξ1 ⊗ α1)(cid:12)(cid:12)ξ2 ⊗ α2(cid:1)
(cid:0)ξ1 ⊗ να1,η1(cid:12)(cid:12)ξ2 ⊗ να2,η2(cid:1)L2(G)⊗H = (cid:0)(1 ⊗ Φ(θη2,η1))(ξ1 ⊗ α1)(cid:12)(cid:12)ξ2 ⊗ α2(cid:1)
(cid:0) W (a∗
(cid:0)(1 ⊗ θη2,η1)Σ∆(ai)Σ(ξ1 ⊗ α1)(cid:12)(cid:12)Σ∆(ai)Σ(ξ2 ⊗ α2)(cid:1)
(cid:0)(θη2,η1 ⊗ 1)∆(ai)(α1 ⊗ ξ1)(cid:12)(cid:12)∆(ai)(α2 ⊗ ξ2)(cid:1)
(cid:0)(ωα1,η1 ⊗ ι)∆(ai)ξ1(cid:12)(cid:12)(ωα2,η2 ⊗ ι)∆(ai)ξ2(cid:1),
= X
= X
= X
= X
i
i
i
i
using that W (a ⊗ 1) W ∗ = Σ∆(a)Σ for a ∈ L∞(G).
It follows immediately that U ∗ exists
and is an isometry; uniqueness follows as vectors of the form ξ ⊗ να,η are linearly dense in
L2(G) ⊗ H.
As να,η = Pihai, ωα,ηifi, by using linearity and continuity, we see that also
U ∗(cid:16)ξ ⊗X
hai, ωifi(cid:17) = X
(ω ⊗ ι)∆(ai)ξ ⊗ fi
(ξ ∈ L2(G), ω ∈ L1(G)).
i
i
Proposition 5.3. The operator U is a member of L∞(G)⊗B(H), and is a corepresentation,
that is, (∆ ⊗ ι)U = U13U23.
Proof. Let x ∈ L∞(G)′, so for ξ, α, β ∈ L2(G),
U ∗(xξ ⊗ να,β) = X
i
(ωα,β ⊗ ι)∆(ai)xξ ⊗ fi = X
i
= (x ⊗ 1)U ∗(ξ ⊗ να,β).
x(ωα,β ⊗ ι)∆(ai)ξ ⊗ fi
Thus U ∗ ∈ (L∞(G)′⊗C)′ = L∞(G)⊗B(H), and of course the same is true of U.
10
We shall prove that (∆ ⊗ ι)(U ∗) = U ∗
It is easy to see that this is equivalent to
π : L1(G) → B(H); ω 7→ (ω ⊗ ι)(U ∗) being an anti-homomorphism of the Banach algebra
L1(G). However, notice that for ω1, ω2 ∈ L1(G) and ξ, η ∈ L2(G),
23U ∗
13.
(cid:16)π(ωξ,η)X
i
hai, ω1ifi(cid:12)(cid:12)(cid:12)X
j
haj, ω2ifj(cid:17) = (cid:16)U ∗(cid:16)ξ ⊗X
i
j
hai, ω1ifi(cid:17)(cid:12)(cid:12)(cid:12)η ⊗X
(ω1 ⊗ ι)∆(ai)ξ ⊗ fi(cid:12)(cid:12)(cid:12)η ⊗X
hai, ω1ωξ,ηifi(cid:12)(cid:12)(cid:12)X
haj, ω2ifj(cid:17).
j
j
= (cid:16)X
= (cid:16)X
i
i
haj, ω2ifj(cid:17)
haj, ω2ifj(cid:17)
Thus
π(ω)(cid:16)X
i
hai, ω′ifi(cid:17) = X
i
hai, ω′ωifi
(ω, ω′ ∈ L1(G)),
and it is now immediate that π is an anti-homomorphism.
We remark that we can view H as being a completion of L1(G), where we identify ω ∈ L1(G)
with Pihai, ωifi ∈ H. Then π in the above proof (that is, the anti-homomorphism from L1(G)
to B(H) induced by U ∗) is simply the map π(ω) : ω′ 7→ ω′ω.
We now finish the argument by showing that U is unitary.
Lemma 5.4. The closed image of U ∗ is equal to the closed linear span of {(a ⊗ 1)V (ξ) : ξ ∈
L2(G), a ∈ C0( G)}. In particular, the image of U ∗ contains the image of V , and so U ∗U V = V .
Proof. Let ξ1, ξ2, η ∈ L2(G), and let Pi ξi ⊗ fi ∈ L2(G) ⊗ H, and observe that
(cid:16)U ∗(cid:0)ξ1 ⊗X
i
(aiξ2η)fi(cid:1)(cid:12)(cid:12)(cid:12)X
j
ξj ⊗ fj(cid:17) = X
= X
i
i
(cid:0)(ωξ2,η ⊗ ι)∆(ai)ξ1(cid:12)(cid:12)ξi(cid:1)
(cid:0)(1 ⊗ ai)W (ξ2 ⊗ ξ1)(cid:12)(cid:12)W (η ⊗ ξi)(cid:1).
We will compute the image of U ∗ by taking linear combinations of ξ1, ξ2 and η. In particular,
as W is unitary, we may replace W (ξ2 ⊗ ξ1) by ξ2 ⊗ ξ1 in the above expression. It follows that
the image of U ∗ is the closed linear span of vectors of the form
X
i
(ω ⊗ ι)(W )∗ai(ξ) ⊗ fi
(ω ∈ L1(G), ξ ∈ L2(G)).
Now, {(ω ⊗ ι)(W )∗ : ω ∈ L1(G)} is dense in C0( G), and so the result follows, as V (ξ) =
Pi ai(ξ) ⊗ fi. As C0( G) contains a bounded approximate identity, clearly the image of U ∗
contains the image of V . As U ∗U is the orthogonal projection onto the image of U ∗ (as U ∗ is
an isometry) it follows immediately that U ∗U V = V .
Proposition 5.5. The corepresentation U is unitary.
Proof. Suppose that Pi ξi ⊗ fi ∈ L2(G) ⊗ H is orthogonal to the image of U ∗. Let ξ ∈ L2(G)
and x′ ∈ L∞(G)′, so by Lemma 5.4, for any a ∈ C0( G),
As C0( G) is σ-weakly dense in L∞( G), it follows that
i
0 = (cid:16)(a ⊗ 1)V x′ξ(cid:12)(cid:12)(cid:12)X
X
(cid:0)xx′aiξ(cid:12)(cid:12)ξi(cid:1) = 0
i
ξi ⊗ fi(cid:17) = X
i
(cid:0)aaix′ξ(cid:12)(cid:12)ξi(cid:1) = X
i
(cid:0)ax′aiξ(cid:12)(cid:12)ξi(cid:1).
(ξ ∈ L2(G), x ∈ L∞( G), x′ ∈ L∞(G)′)),
11
By Theorem 2.2, this implies that
0 = X
i
(cid:0)T aiξ(cid:12)(cid:12)ξi(cid:1) = (cid:16)(T ⊗ 1)V ξ(cid:12)(cid:12)(cid:12)X
i
ξi ⊗ fi(cid:17)
(ξ ∈ L2(G), T ∈ B(L2(G))).
However, we know that {(x⊗1)V (ξ) : x ∈ B0(L2(G)), ξ ∈ L2(G)} is linearly dense in L2(G)⊗H.
Hence Pi ξi ⊗ fi = 0, and so U ∗ has dense range, as required.
Remark 5.6. We started with a completely positive left multiplier of L1( G); let L be the
adjoint, a normal completely positive map on L∞( G). We immediately used Proposition 3.3 to
extend L to a normal completely positive map Φ on all of B(L2(G)). Remember that the repre-
sentation Φ(x) = V ∗(x⊗1)V is unique (up to unitary isomorphism), as this dilation is minimal.
This is equivalent to the non-degeneracy condition that {(x⊗1)V ξ : x ∈ B0(L2(G)), ξ ∈ L2(G)}
is linearly dense in L2(G) ⊗ H.
As Φ extends L, we hence have a normal Stinespring representation of L, as L(x) = V ∗(x ⊗
1)V . The previous results show that {(x ⊗ 1)V (ξ) : x ∈ L∞( G), ξ ∈ L2(G)} is also linearly
dense in L2(G) ⊗ H. It follows that we also have a minimal Stinespring dilation of the original
multiplier L.
5.2 Recovering the multiplier
In the previous section, we showed how to construct a unitary corepresentation of G from a
completely positive left multiplier of L1( G). We now show how to recover the multiplier from
the corepresentation.
ξ ∈ L2(G).
Proposition 5.7. There is α0 ∈ H such that U ∗(ξ ⊗ α0) = V (ξ) = Pi ai(ξ) ⊗ fi for all
Proof. Let our left multiplier be represented by a0 ∈ M(C0(G)), so that (1 ⊗ a0) W = (Φ ⊗
ι)( W ) = Pi(a∗
i )∆(ai) = a0 ⊗ 1. For ω ∈ L1(G) and
i ⊗ 1) W (ai ⊗ 1). Equivalently, Pi(1 ⊗ a∗
ξ, η ∈ L2(G), observe that
(cid:16)U ∗(cid:0)ξ ⊗X
i
hai, ωifi(cid:1)(cid:12)(cid:12)(cid:12)V (η)(cid:1) = X
i
(cid:0)(ω ⊗ ι)∆(ai)ξ(cid:12)(cid:12)ai(η)(cid:1)
= (cid:16)(ω ⊗ ι)(cid:0)(1 ⊗ a∗
i )∆(ai)(cid:1)ξ(cid:12)(cid:12)η(cid:1) = ha0, ωi(ξη).
As this holds for all ξ, η, it follows that the map Pihai, ωifi 7→ ha0, ωi is bounded, and so the
Riesz representation theorem for Hilbert spaces provides α0 ∈ H such that
i
(cid:16)X
hai, ωifi(cid:12)(cid:12)(cid:12)α0(cid:17) = ha0, ωi
(cid:0)U ∗(ξ ⊗ α)(cid:12)(cid:12)V (η)(cid:1) = (ξ ⊗ αη ⊗ α0)
(ω ∈ L1(G)).
(ξ, η ∈ L2(G), α ∈ H),
By continuity,
that is, U V (η) = η ⊗ α0 for all η ∈ H. By Lemma 5.4, as U ∗U V = V , it follows that
V (η) = U ∗U V (η) = U ∗(η ⊗ α0) as required.
We now take slices of U against this vector α0, and find that this constructs our original
multiplier, in the sense of Proposition 4.1.
Theorem 5.8. Let L∗ be a completely positive left multiplier of L1( G). There is a unitary
corepresentation U of G on H, such that L = (L∗)∗ is induced by U, using α0 ∈ H.
12
Proof. Form U as above and form α0 as in the previous proposition.
It is immediate that
ai = (ι ⊗ αα0,fi)(U ∗) for all i. So the multiplier constructed by Proposition 4.1 for α0 is, on
L∞( G), the map x 7→ a∗
i xai, which is just L, as required.
Theorem 5.9. Let G be a locally compact quantum group. There is an isometric, order pre-
serving bijection between the completely positive multipliers of L1( G) and C u
0 ( G)∗
+.
0 ( G)∗
+ induce completely positive left multipliers of L1( G) in the sense
Proof. Members of C u
discussed in Section 4.1. Conversely, any completely positive left multiplier comes from a
unitary corepresentation, and this is associated to a member of C u
+ by Proposition 4.2.
That this procedure gives a bijection follows as L1( G) is an essential ideal in C u
0 ( G)∗, see [9,
Proposition 8.3].
0 ( G)∗
0 ( G) ⊆ B(H) is the universal representation,
so µ = ωα,α for some α ∈ H. Then W can be identified with a member of B(L2(G) ⊗ H), and
Proposition 4.2 and Proposition 4.1 show that left multiplication by µ induces the completely
positive multiplier L, where in particular,
+ is a state, then suppose that C u
If µ ∈ C u
0 ( G)∗
L(1) = (ι ⊗ ωα,α)(WW ∗) = 1hµ, 1i = 1.
So kLk = 1, and hence our bijection is an isometry.
Finally, if µ ≤ λ in C u
+ then form the associated completely positive multipliers Lµ
and Lλ. Let L be the multiplier formed from λ − µ, so by uniqueness, L = Lλ − Lµ. As L is
completely positive, Lλ ≥ Lµ. The converse is simply a case of reversing the argument. Thus
our bijection is order preserving.
0 ( G)∗
We remark that it is completely obvious from these results that for any completely positive
left multiplier L, there is a completely positive right multiplier L′ such that (L, L′) forms a
0 ( G)∗
double multiplier (simply let L′ be induced by right multiplication by the element of C u
associated to L).
It seems to be unknown if a similar result holds for completely bounded
multipliers.
6 Representing elements for completely bounded multi-
pliers
Notice that while Proposition 3.2 shows that all completely bounded multipliers are "repre-
sented", we didn't use this fact until Proposition 5.7. Here we show how to use the representing
element more directly.
Recall, from [19] for example, that L1(G) contains a dense ∗-subalgebra L1
♯ (G); we define
♯ (G) if and only if there is τ ∈ L1(G) with hx, τ i = hS(x)∗, ωi for all x ∈ D(S), and in
♯ (G) can be developed
ω ∈ L1
this case, denote ω♯ = τ . We note that the elementary properties of L1
mutatis mutandis for G coming from manageable multiplicative unitaries.
Recall that the scaling group (τt) is implemented as τt(x) = P itxP −it, where P is a certain
positive injective operator. As R and τt commute for all t, and S = Rτ−i/2, it follows that R
leaves D(S) invariant, and RS = SR. It is then easy to see that R∗ leaves L1
♯ (G) invariant,
and R∗(ω♯) = R∗(ω)♯ for ω ∈ L1
♯ (G). Given β ∈ D(P −1/2) and ξ ∈ D(P 1/2), we have that for
x ∈ D(S) = D(τ−i/2),
hx, ωP −1/2β,P 1/2ξi = (cid:0)P 1/2xP −1/2β(cid:12)(cid:12)ξ(cid:1) = hτ−i/2(x), ωβ,ξi = hS(R(x)), ω∗
ξ,βi = hx, R∗(ω♯
ξ,β)i,
and so ωξ,β ∈ L1
♯ (G) with ω♯
ξ,β = R∗(ωP −1/2β,P 1/2ξ).
13
Proposition 6.1. Let L be a completely bounded left multiplier of L1( G), represented by a0 ∈
M(C0(G)). For ξ, η ∈ D(P 1/2) and α, β ∈ D(P −1/2), we have that
(cid:0)Φ(θξ,η)α(cid:12)(cid:12)β(cid:1) = h∆(a0), ωα,η ⊗ ω♯∗
ξ,βi.
Proof. Let ξ0 ∈ L2(G) be a unit vector, let (ei) be an orthonormal basis for L2(G), and let
i ⊗ ei. For ǫ > 0, we can find a family (βi) in
D(P −1/2) with
W (α ⊗ ξ0) = Pi αi ⊗ ei and W (β ⊗ ξ0) = Pi β′
(cid:13)(cid:13)(cid:13)W (β ⊗ ξ0) −X
i
βi ⊗ ei(cid:13)(cid:13)(cid:13) < ǫ.
Using Proposition 3.3, and that W = ΣW ∗Σ, we see that
(cid:0)Φ(θξ,η)α(cid:12)(cid:12)β(cid:1) = (cid:0)(ι ⊗ L)(W (θξ,η ⊗ 1)W ∗)W (α ⊗ ξ0)(cid:12)(cid:12)W (β ⊗ ξ0)(cid:1)
⊗ ι)(ι ⊗ L)(W (θξ,η ⊗ 1)W ∗)ei(cid:12)(cid:12)ej(cid:1)
⊗ ι)(W )(ωαi,η ⊗ ι)(W ∗))ei(cid:12)(cid:12)ej(cid:1).
(cid:0)(ωαi,β ′
(cid:0)L((ωξ,β ′
= X
= X
i,j
i,j
j
j
A similar calculation establishes that if
x = X
i,j
(cid:0)L((ωξ,βj ⊗ ι)(W )(ωαi,η ⊗ ι)(W ∗))ei(cid:12)(cid:12)ej(cid:1),
then
That is, we may replace (β′
(cid:12)(cid:12)(cid:0)Φ(θξ,η)α(cid:12)(cid:12)β(cid:1) − x(cid:12)(cid:12) < ǫkLkcbkαkkξkkηk.
j) by (βj), at the cost of a small error term.
As (ω ⊗ ι)(W )∗ = (ω♯ ⊗ ι)(W ) for ω ∈ L1
ξ,βj ⊗ ι)(W ∗). This makes sense, as βj ∈ D(P −1/2) and ξ ∈ D(P 1/2). Thus
♯ (G), we see that (ωξ,βj ⊗ ι)(W ) = (ω♯
(ω♯∗
ξ,βj
⊗ ι)(W )∗ =
(ωξ,βj ⊗ ι)(W )(ωαi,η ⊗ ι)(W ∗) = (ω♯∗
ξ,βj
⊗ ι)(W ∗)(ωαi,η ⊗ ι)(W ∗) = (ωαi,ηω♯∗
ξ,βj
⊗ ι)(W ∗).
Recall that (ι ⊗ L)(W ∗) = (a0 ⊗ 1)W ∗, and that (∆ ⊗ ι)(W ∗) = W ∗
23W ∗
13, and so
i,j
x = X
= X
= X
i,j
j
Let a ∈ D(S)∗, so that
(cid:0)(ωαi,ηω♯∗
ξ,βj
⊗ ι)((ι ⊗ L)(W ∗))ei(cid:12)(cid:12)ej(cid:1)
h(∆(a0) ⊗ 1)W ∗
23W ∗
13, ωαi,η ⊗ ω♯∗
ξ,βj
⊗ ωei,ej i
h(∆(a0) ⊗ 1)W ∗
23, ωα,η ⊗ ω♯∗
ξ,βj
⊗ ωξ0,ej i.
X
j
h(a ⊗ 1)W ∗, ω♯∗
ξ,βj
⊗ ωξ0,ej i = X
hS((ι ⊗ ωξ0,ej )(W ))∗a∗, ω♯
j
haS((ι ⊗ ωξ0,ej )(W )), ω♯∗
ξ,βj
i = X
ξ,βj
i
j
h(ι ⊗ ωξ0,ej )(W )S(a∗)∗, ωξ,βji
hW (S(a∗)∗ ⊗ 1), ωξ,βj ⊗ ωξ0,ej i = (cid:16)W (S(a∗)∗ ⊗ 1)(ξ ⊗ ξ0)(cid:12)(cid:12)X
j
βj ⊗ ej(cid:17).
= X
= X
j
j
14
By comparison,
(cid:16)W (S(a∗)∗ ⊗ 1)(ξ ⊗ ξ0)(cid:12)(cid:12)X
j
β′
j ⊗ ej(cid:17) = (cid:0)(S(a∗)∗ ⊗ 1)(ξ ⊗ ξ0)(cid:12)(cid:12)W ∗W (β ⊗ ξ0)(cid:1)
= hS(a∗)∗, ωξ,βi = ha∗, ω♯
ξ,βi = ha, ω♯∗
ξ,βi.
If it so happens that a = (ωα,η ⊗ ι)∆(a0) is in D(S)∗, then we have
However, observe that for this choice of a,
(cid:12)(cid:12)x − ha, ω♯∗
ξ,βi(cid:12)(cid:12) ≤ ǫkξkkS(a∗)k.
ha, ω♯∗
ξ,βi = h∆(a0), ωα,η ⊗ ω♯∗
ξ,βi,
and so as ǫ > 0, this gives the required result.
So it remains to show that a = (ωα,η ⊗ ι)∆(a0) ∈ D(S)∗. By [8, Theorem 5.9], we know
that a0 ∈ D(S)∗, and by hypothesis, ωη,α ∈ L1
♯ (G). Thus, for ω ∈ L1
♯ (G),
ha∗, ω♯i = h(ωη,α ⊗ ι)∆(a∗
0), ω♯i = ha∗
0, ωη,αω♯i = hS(a∗
0)∗, ωω♯
η,αi = h(ι ⊗ ω♯
η,α)∆(S(a∗
0)∗), ωi.
This is enough to show that a∗ ∈ D(S) with S(a∗) = (ι ⊗ ω♯
σ-weakly closed operator; for details see for example [6, Appendix A].
η,α)∆(S(a∗
0)∗), given that S is a
6.1 Weak∗-continuity of the Junge, Neufang, Ruan representation
As explained in Section 3 above, [17] shows that for a locally compact quantum group G, there
is a bijection between the completely bounded left multipliers of L1( G), say M l
cb(L1( G)), and
CBσ,L∞( G)
L∞(G)′ (B(L2(G))). In [16], it is shown that this map is weak∗-weak∗ continuous, at least
when G has the left co-AP property, see [16, Corollary 4.10] (and [16, Theorem 4.7] for the
version for right multipliers).
In this final section of the paper, we apply the result of the
previous section to show that this weak∗-continuity result is true for all G.
Firstly, we recall from [16] the proof that M l
cb(L1( G)) is a dual space. Proposition 3.2 shows
cb(L1( G)) → L∞(G) (actually, this maps into M(C0(G)), but this is
that we have a map Λ : M l
unimportant here) which satisfies
(L ⊗ ι)(W )W ∗ = 1 ⊗ Λ(L∗),
Λ(L∗)λ(ω) = λ(cid:0)L∗(ω)(cid:1)
(ω ∈ L1( G)).
It follows that Λ is a contractive algebra homomorphism. Then [16, Proposition 3.4] shows
cb(L1( G)), then
that if we denote by X the image of Λ, equipped with the norm coming from M l
cb(L1( G)) its canonical
the closed unit ball of X is weak∗-closed in L∞(G). Indeed, giving M l
operator space structure, the closed unit ball of Mn(X) is weak∗-closed in Mn(L∞(G)). Using
cb(L1( G)) be the closure in M l
cb(L1( G))∗ of the
this, [16, Theorem 3.5] shows that if we let Ql
cb(L1( G))∗ is completely isometrically isomorphic
image of L1(G) under the adjoint of Λ, then Ql
to M l
cb(L1( G)). Thus we get a weak∗-topology on M l
In [9, Section 8] we independently gave an analogous construction of a weak∗-topology on
the space of double multipliers.
In fact, the first part of the proof of [9, Proposition 8.11]
already works for merely left multipliers, and then one can apply the abstract result which is
cb(L1( G)). In [9] we found a very "Banach algebraic" way
[9, Proposition 8.12] to construct Ql
to construct preduals for double multiplier algebras (see [9, Theorem 7.7] for example), but it
seems that at several crucial points, it really is necessary to work with double multipliers. It
would be interesting to know how to adapt these ideas to one-sided multipliers.
cb(L1( G)).
15
For us, the important point is that if (Lα) is a bounded net in M l
cb(L1( G)), then (Lα) is
weak∗-null with respect to Ql
cb(L1( G)) if and only if (Λ(Lα)) is weak∗-null in L∞(G).
We next consider the space CBσ,L∞( G)
L∞(G)′ (B(L2(G))). Firstly, we consider the larger space
CBσ(B(L2(G))) which can be identified with CB(B0(L2(G)), B(L2(G))). This in turn is the dual
space of B0(L2(G))b⊗B(L2(G))∗, the operator space projective tensor product of the compact
operators B0(L2(G)) with the trace-class operators B(L2(G))∗. By restriction, we have a weak∗-
topology on CBσ,L∞( G)
L∞(G)′ (B(L2(G))). Again, for us the important point is that a bounded net (Φα)
in CBσ,L∞( G)
L∞(G)′ (B(L2(G))) is weak∗-null if and only if (Φα(θ)) is a weak∗-null net in B(L2(G)), for
each θ ∈ B0(L2(G)). All this is explained in [16, Section 4] and the references therein.
The following improves [16, Theorem 4.7] (which is stated for right multipliers) in that we
need make no approximation property type assumptions.
Theorem 6.2. For any G, the map M l
L∞(G)′ (B(L2(G))) is weak∗-weak∗-
continuous. If G is a locally compact quantum group, this correspondence is a weak∗-weak∗-
continuous homeomorphism.
cb(L1( G)) → CBσ,L∞( G)
cb(L1( G)) → CBσ,L∞( G)
Proof. Denote by φ the map M l
L∞(G)′ (B(L2(G))). To show that φ is weak∗-
cb(L1( G)), then
continuous, it suffices to show that if (Li) is a bounded, weak∗-null net in M l
the corresponding bounded net, say (Φi), in CBσ(B(L2(G))) is weak∗-null. When G is a locally
compact quantum group, we know from [17] that φ is a completely isometric isomorphism, and
then if φ is weak∗-continuous, it is automatically a weak∗-weak∗-continuous homeomorphism.
This is perhaps not well-known (in the operator space setting) but see [9, Lemma 10.1] for
example.
We fix a bounded weak∗-null net (Li) of left multipliers, with corresponding net (Φi). For
each i let Li be represented by ai ∈ L∞(G). That (Li) is weak∗-null means that (ai) is weak∗-
null in L∞(G). As explained above, as Λ is a contraction, (ai) is also a bounded net. By
Proposition 6.1, we have that
(Φi(θξ,η)αβ) = hai, ωα,ηω♯∗
ξ,βi
(ξ, η ∈ D(P 1/2), α, β ∈ D(P −1/2)).
As D(P 1/2) and D(P −1/2) are dense in L2(G), we immediately see that
lim
i
hΦi(θ), ωi = 0
for a dense collection of θ ∈ B0(L2(G)) and ω ∈ B(L2(G))∗. As (Φi) is a bounded net, this is
enough to show that (Φi) is weak∗-null, as required.
As remarked on in the proof of [16, Theorem 4.7], we can equivalently state this result in
terms of the Haagerup tensor product (see [12, Chapter 9]). We have a completely isometric
isomorphism
B0(L2(G))b⊗B(L2(G))∗ → B(L2(G))∗
h
⊗ B(L2(G))∗;
θξ,η ⊗ ωα,β 7→ ωξ,β ⊗ ωα,η.
See also the discussion after [16, Remark 4.6]. The adjoint gives a normal completely isometric
isomorphism
B(L2(G))
eh
⊗ B(L2(G)) → CBσ(B(L2(G)));
x ⊗ y 7→ Tx,y.
Here we use the extended (or weak∗) Haagerup tensor product, see [2, 13], and Tx,y is the
eh
operator z 7→ xzy. This isomorphism restricts to an isomorphism between L∞(G)
⊗ L∞(G)
and CBσ
L∞(G)′(B(L2(G))), and the predual of L∞(G)
eh
⊗ L∞(G) is L1(G)
h
⊗ L1(G).
16
We can hence restate Theorem 6.2 as saying that there is a completely bounded map
φ∗ : L1(G)
h
⊗ L1(G) → Ql
cb(L1( G)),
the adjoint of which is our map M l
cb(L1( G)) → CBσ
L∞(G)′(B(L2(G))).
References
[1] O. Yu. Aristov, "Amenability and compact type for Hopf-von Neumann algebras from the ho-
mological point of view", in "Banach algebras and their applications", Contemp. Math. 363, pp.
15 -- 37, (Amer. Math. Soc., Providence, RI, 2004).
[2] D. P. Blecher, R. R. Smith, "The dual of the Haagerup tensor product", J. London Math. Soc.
(2) 45 (1992) 126 -- 144.
[3] M. Bozejko, "Positive definite bounded matrices and a characterization of amenable groups",
Proc. Amer. Math. Soc. 95 (1985) 357 -- 360.
[4] M. Brannan, "Approximation properties for free orthogonal and free unitary quantum groups",
to appear in Journal fur die Reine und Angewandte Mathematik, see arXiv:1103.0264 [math.OA].
[5] M. Brannan, "Reduced operator algebras of trace-preserving quantum automorphism groups",
preprint, see arXiv:1202.5020v2 [math.OA].
[6] M. Brannan, M. Daws, E. Samei, "Completely bounded representations of convolution algebras
of locally compact quantum groups", preprint, see arXiv:1107.2094v2 [math.OA]
[7] M. Cowling, U. Haagerup, "Completely bounded multipliers of the Fourier algebra of a simple
Lie group of real rank one", Invent. Math. 96 (1989) 507 -- 549.
[8] M. Daws, "Multipliers of locally compact quantum groups via Hilbert C∗-modules", J. London
Math. Soc. 84 (2011) 385 -- 407.
[9] M. Daws, "Multipliers, Self-Induced and Dual Banach Algebras", Dissertationes Math. 470 (2010)
62pp.
[10] M. Daws, P. Salmi, "Completely positive definite functions and Bochner's theorem for locally
compact quantum groups", in preparation.
[11] J. De Canni´ere, U. Haagerup, "Multipliers of the Fourier algebras of some simple Lie groups and
their discrete subgroups", Amer. J. Math. 107 (1985) 455 -- 500.
[12] E. G. Effros, Z.-J. Ruan, "Operator spaces", London Mathematical Society Monographs, New
Series, 23 (Oxford University Press, New York, 2000).
[13] E. G. Effros, Z.-J. Ruan, "Operator space tensor products and Hopf convolution algebras", J.
Operator Theory 50 (2003) 131 -- 156.
[14] A. Freslon, "A Note on weak amenability for free products of discrete quantum groups", C. R.
Math. Acad. Sci. Paris 350 (2012) 403 -- 406.
[15] U. Haagerup, M. Musat, "Classification of hyperfinite factors up to completely bounded isomor-
phism of their preduals", J. Reine Angew. Math. 630 (2009) 141 -- 176.
[16] Z. Hu, M. Neufang, Z.-J. Ruan, "Completely bounded multipliers over locally compact quantum
groups", Proc. London Math. Soc. 103 (2011) 1 -- 39.
17
[17] M. Junge, M. Neufang, Z.-J. Ruan, "A representation theorem for locally compact quantum
groups", Internat. J. Math. 20 (2009) 377 -- 400.
[18] J. Kraus, Z.-J. Ruan, "Multipliers of Kac algebras", Internat. J. Math. 8 (1997) 213 -- 248.
[19] J. Kustermans, "Locally compact quantum groups in the universal setting", Internat. J. Math.
12 (2001) 289 -- 338.
[20] J. Kustermans, "Locally compact quantum groups" in Quantum independent increment processes.
I, Lecture Notes in Math. 1865, pp. 99 -- 180 (Springer, Berlin, 2005).
[21] J. Kustermans, S. Vaes, "Locally compact quantum groups", Ann. Sci. ´Ecole Norm. Sup. 33
(2000) 837 -- 934.
[22] J. Kustermans, S. Vaes, "Locally compact quantum groups in the von Neumann algebraic set-
ting", Math. Scand. 92 (2003) 68 -- 92.
[23] V. Losert, "Properties of the Fourier algebra that are equivalent to amenability", Proc. Amer.
Math. Soc. 92 (1984) 347 -- 354.
[24] R. Meyer, S. Roy, S. L. Woronowicz, "Homomorphisms of quantum groups", to appear in Mnster
J. Math., see arXiv:1011.4284v2 [math.OA]
[25] P. M. So ltan, S. L. Woronowicz, "A remark on manageable multiplicative unitaries", Lett. Math.
Phys. 57 (2001) 239 -- 252.
[26] P. M. So ltan, S. L. Woronowicz, "From multiplicative unitaries to quantum groups. II", J. Funct.
Anal. 252 (2007) 42 -- 67.
[27] R. R. Smith, "Completely bounded module maps and the Haagerup tensor product", J. Funct.
Anal. 102 (1991) 156 -- 175.
[28] N. Spronk, "Measurable Schur multipliers and completely bounded multipliers of the Fourier
algebras", Proc. London Math. Soc. 89 (2004) 161 -- 192.
[29] M. Takesaki, "Theory of operator algebras. I.", Encyclopaedia of Mathematical Sciences, 124.
Operator Algebras and Non-commutative Geometry, 5. (Springer-Verlag, Berlin, 2002).
[30] S. Vaes, "Locally compact quantum groups", PhD. thesis, Katholieke Universiteit Leuven, 2001.
Available from http://wis.kuleuven.be/analyse/stefaan/
[31] S. Vaes, A. Van Daele, "Hopf C ∗-algebras", Proc. London Math. Soc. 82 (2001) 337 -- 384.
[32] S. Vaes, A. Van Daele, "The Heisenberg commutation relations, commuting squares and the Haar
measure on locally compact quantum groups", in Operator algebras and mathematical physics
(Constant¸a, 2001), 379 -- 400 (Theta, Bucharest, 2003).
[33] S. L. Woronowicz, "From multiplicative unitaries to quantum groups", Internat. J. Math. 7 (1996)
127 -- 149.
Matthew Daws
School of Mathematics,
University of Leeds,
LEEDS LS2 9JT
United Kingdom
Email: [email protected]
18
|
1001.2982 | 2 | 1001 | 2010-12-14T12:05:55 | $C^*$-algebras associated to $C^*$-correspondences and applications to mirror quantum spheres | [
"math.OA"
] | The structure of the $C^*$-algebras corresponding to even-dimensional mirror quantum spheres is investigated. It is shown that they are isomorphic to both Cuntz-Pimsner algebras of certain $C^*$-correspondences and $C^*$-algebras of certain labelled graphs. In order to achieve this, categories of labelled graphs and $C^*$-correspondences are studied. A functor from labelled graphs to $C^*$-correspondences is constructed, such that the corresponding associated $C^*$-algebras are isomorphic. Furthermore, it is shown that $C^*$-correspondences for the mirror quantum spheres arise via a general construction of restricted direct sum. | math.OA | math | C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES AND
APPLICATIONS TO MIRROR QUANTUM SPHERES
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
Abstract. The structure of the C ∗-algebras corresponding to even-dimensional mirror
quantum spheres is investigated. It is shown that they are isomorphic to both Cuntz-
Pimsner algebras of certain C ∗-correspondences and C ∗-algebras of certain labelled
graphs. In order to achieve this, categories of labelled graphs and C ∗-correspondences
are studied. A functor from labelled graphs to C ∗-correspondences is constructed, such
that the corresponding associated C ∗-algebras are isomorphic. Furthermore, it is shown
that C ∗-correspondences for the mirror quantum spheres arise via a general construction
of restricted direct sum.
0
1
0
2
c
e
D
4
1
]
.
A
O
h
t
a
m
[
2
v
2
8
9
2
.
1
0
0
1
:
v
i
X
r
a
MSC 2010: 46L08, 46L65, 46L85
Keywords: C*-correspondence, labelled graph, quantum sphere, Cuntz-Pimsner algebra
This work has been supported in part by: the FNU Rammebevilling grant 'Operator alge-
bras and applications' (2009 -- 2011), the Marie Curie Research Training Network MRTN-
CT-2006-031962 EU-NCG, the NordForsk Research Network 'Operator algebra and dy-
namics', the Marie Curie International Staff Exchange Scheme PIRSES-GA-2008-230836,
and the Polish Government grant N201 1770 33.
Date: 10 January 2010.
1
2
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
1. Introduction
Our original motivation for this study came from the desire to better understand the
C ∗-algebraic structure of mirror quantum spheres. They were first defined in dimension
2 in [11], and then in full generality in [15, 16]. It was noted from the beginning that
in even dimensions these differ on the C ∗-algebraic level from the Euclidean quantum
spheres arising from quantum groups.
In particular, the Euclidean quantum spheres
correspond to certain graph C ∗-algebras, [12, 14], but such a convenient description was
lacking for the even dimensional mirror quantum spheres.
In the present paper, we show that the C ∗-algebras of even dimensional mirror quan-
tum spheres are naturally isomorphic to both Cuntz-Pimsner algebras of certain C ∗-
correspondences (in the sense of [18]) and C ∗-algebras of certain labelled graphs (in the
sense of [4, 5]). We arrive at these two realisations in the following way. At first, we
define a category of C ∗-correspondences and show that the process of associating C ∗-
algebras to them is functorial. Furthermore, this functor sends restricted direct sums of
C ∗-correspondences to pull-backs of their C ∗-algebras. Restricted direct sums of Hilbert
modules were introduced in [2], and this construction is extended to C ∗-correspondences
in the present article. We then use a realisation of the mirror quantum spheres as pull-
backs of quantum discs to obtain the appropriate C ∗-correspondences. To this end, we
must first find suitable C ∗-correspondences for the quantum discs.
A close examination of the above mentioned C ∗-correspondences results in finding
labelled graphs for the mirror quantum spheres. The explicit forms of the isomorphisms
are obtained by comparing the defining relations.
Our paper is organised as follows. We begin Section 2 by recalling the definition of
C ∗-correspondences and the C ∗-algebra associated to them, following the approach of
Katsura [18]. We also show that the class of C ∗-correspondences, together with appro-
priately defined morphisms forms a category. Thus we obtain a functor which maps
a C ∗-correspondence (X, A) to the C ∗-algebra OX, and sends morphisms between C ∗-
correspondences to C ∗-homomorphisms.
In Section 3, we consider restricted direct sums of C ∗-correspondences, which were first
defined on the level of Hilbert modules by Baki´c and Guljas in [2]. We show that this
construction lifts to a pull-back on the level of corresponding C ∗-algebras.
We use this result in Section 4 in order to obtain a new representation of the even
dimensional mirror quantum spheres. The C ∗-algebra of such a quantum sphere is defined
as the pull-back of two copies of the 2n dimensional quantum disc algebra C(D2n
q ) over
the obvious surjection to the 2n − 1 dimension quantum sphere algebra C(S2n−1
), with
one copy of the disc pre-composed with a 'flip' automorphism. We show that they can be
realised as C ∗-algebras associated to certain explicitly constructed C ∗-correspondences.
In Section 5, we first introduce a category of labelled graphs and construct faithful
representations of C ∗-algebras associated to them. Then we show existence of a functor
q
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
3
from this category of labelled graphs to the category of C ∗-correspondences which pre-
serves the associated C ∗-algebras. Finally, we explicitly construct labelled graphs giving
rise to the C ∗-algebras of the even dimensional mirror quantum spheres.
2. Category of C ∗-correspondences
In this section, we construct a category of C ∗-correspondences in such a way that the
association of C ∗-algebras to the correspondences becomes functorial. These C ∗-algebras
first arose in the work of Pimsner in [25]. However, in this original definition the left
action was required to be injective as non-injective left actions often led to degenerate
C ∗-algebras. This definition included a large class of C ∗-algebras, for example crossed
product C ∗-algebras and Cuntz-Krieger algebras. However, the above mentioned restric-
tion did lead to the exclusion of many interesting C ∗-algebras, notably graph algebras
where the underlying directed graph had sinks. This problem was resolved by Katsura in
[18], and his definition reduces to that of Pimsner when the left action is injective. This
larger class of C ∗-algebras now contains examples such as crossed products by partial
automorphisms, graph algebras associated to graphs with sinks, and as we shall show in
this paper, the labelled graph algebras of Bates and Pask, [4].
We begin this section by recalling the definition of a C ∗-correspondence, e.g.
see
[23, 18].
For a C ∗-algebra A, a right Hilbert A-module is a Banach space X equipped with a
right action of A, and an A-valued inner-product satisfying
(1) hξ, ηai = hξ, ηia;
(2) hη, ξi = hξ, ηi∗; and
(3) hξ, ξi ≥ 0 and kξk =pkhξ, ξik
for all ξ, η ∈ X and a ∈ A.
We denote by L(X) the set of all adjointable operators on X; that is, linear operators
T : X → X such that there exists a linear operator T ∗ called the adjoint of T such that
hT ξ, ηi = hξ, T ∗ηi
for all ξ, η ∈ X. If the adjoint T ∗ exists, it is unique. With the usual operator norm
kT k = sup{kT xk : kxk ≤ 1}, L(X) is a C ∗-algebra
For ξ, η ∈ X, define θξ,η to be the operator satisfying
θξ,η(ζ) = ξhη, ζi.
This is an adjointable operator with (θξ,η)∗ = θη,ξ. We call
K(X) = span{θξ,η : ξ, η ∈ X}
the set of compact operators. It is a closed two-sided ideal in L(X).
Definition 2.1. We say that a right Hilbert A-module X is a C ∗-correspondence over
A when a ∗-homomorphism φX : A → L(X) is given. We call φX the left action of the
C ∗-correspondence.
4
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
Example 2.2. Let D be a C ∗-algebra. Then there is a natural way to realise (D, D) as
a C ∗-correspondence by defining the left and right actions by multiplication and inner-
product ha, bi = a∗b for a, b ∈ D. With a C ∗-correspondence defined in this manner,
there is a canonical isomorphism
satisfying ιD(θa,b) = ab∗.
ιD : K(D) → D
For C ∗-correspondences (X, A) and (Y, B) and a continuous linear map ψX : X → Y
a ∗-homomorphism ψ+
X : K(X) → K(Y ) is defined satisfying
ψ+
X (θξ,η) = θψX (ξ),ψX (η).
To see that this map is well-defined we refer the reader to [25, 17] and [2, Proposition
3.10]. If (Z, C) is another C ∗-correspondences and ψY : Y → Z is a continuous map,
then we have the following composition law: ψ+
X = (ψY ◦ ψX )+.
Now we aim to define the category of C ∗-correspondences and introduce a covariant
functor F from the category of C ∗-correspondences to the category of C ∗-algebras, sat-
isfying F (X, A) = OX where OX is the C ∗-algebra associated to (X, A) in the sense of
Katsura [18]. To this end, we first define after Katsura ([18]) an ideal JX of A by
Y ◦ ψ+
JX := {a ∈ A : φX (a) ∈ K(X) and ab = 0 for all b ∈ ker(φX)}.
X (K(X)) whenever φX is injective.
Note that JX = φ−1
Definition 2.3. Let (X, A) and (Y, B) be C ∗-correspondences. Let ψX : X → Y be a
linear map and ψA : A → B be a C ∗-homomorphism. We say that the pair (ψX, ψA) is
a morphism of C ∗-correspondences if the following conditions hold.
(C1) hψX(ξ), ψX(η)i = ψA(hξ, ηi) for all ξ, η ∈ X,
(C2) ψX(φX(a)ξ) = φY (ψA(a))ψX (ξ) for all ξ ∈ X and a ∈ A,
(C3) ψA(JX) ⊂ JY , and
(C4) φY (ψA(a)) = ψ+
X(φX(a)) for all a ∈ JX .
Proposition 2.4. Let (ψX , ψA) : (X, A) → (Y, B) be a morphism of C ∗-correspondences.
Then ψX (ξa) = ψX (ξ)ψA(a) for all a ∈ A. Furthermore, if the image ψX (X) ⊂ Y is
dense, then condition (C4) automatically follows from (C1) -- (C3).
Proof. The first part follows from [1, Theorem 2.3]. For the second property, suppose
the image ψX (X) ⊂ Y is dense. It follows from the first property and (C1) that for any
θξ,η ∈ K(X) and ψX(ζ) ∈ ψX (X) we have
ψ+
X (θξ,η)ψX(ζ) = ψX(ξ)hψX(η), ψX(ζ)i = ψX (θξ,ηζ).
Therefore ψ+
have
X (K)ψX(ζ) = ψX (Kζ) for all K ∈ K(X). In particular, for all a ∈ JX we
ψ+
X (φX(a))ψX (ζ) = ψX (φX(a)ζ) = φY (ψA(a))ψX (ζ)
by (C2), since φX(a) ∈ K(X). By density of ψX (X) ⊂ Y , we may now conclude that
ψ+
(cid:3)
X (φX(a) = φY (ψA(a)).
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
5
The following simple example illustrates the fact that (C4) may fail if the image of ψX
is not dense.
Example 2.5. Let X be the one dimensional Hilbert space with generator e, and Y
be the two dimensional Hilbert space with generators f1 and f2. Let A = B = C.
Then by defining left and right actions as multiplication, (X, A) and (Y, B) are C ∗-
correspondences. Define a pair of maps (ψX, ψA) : (X, A) → (Y, B) by ψX (e) = f1 and
ψA = id. Clearly, the image of ψX is not dense in Y . It is easily shown that conditions
(C1), (C2) and (C3) hold. But (C4) says that we should have
X (φX(1)) = φY (ψA(1))
X (θe,e) = φY (1)
ψ+
⇐⇒ ψ+
⇐⇒ θf1,f1 = θf1,f1 + θf2,f2
which would imply that the generator f2 is zero. So (C4) does not hold.
We can now construct the desired category, which we call C. The objects are given by
Obj(C) = {(X, A) : X is a C ∗-correspondence over A}
and morphisms Mor(C) as in Definition 2.3.
There are well defined domain and codomain maps Mor(C) → Obj(C), namely for
(ψX , ψA) : (X, A) → (Y, B), we have dom(ψX , ψA) = (X, A) and cod(ψX, ψA) = (Y, B).
It is clear from the definition of morphisms that the composition of two composable
morphisms will result in another morphism. For an object (X, A) the identity morphism
is id(X,A) = (idX, idA) which is also clearly a morphism of C ∗-correspondences.
The next step is to define the C ∗-algebra associated to a C ∗-correspondence, and show
that this process is naturally implemented by a functor between the categories. In order
to do this, we first need to define what we mean by a covariant representation of a
C ∗-correspondence on a C ∗-algebra. We do this via morphisms of C ∗-correspondences.
Definition 2.6. Let (X, A) be a C ∗-correspondence and let D be an arbitrary C ∗-algebra.
A covariant representation of (X, A) on D is a morphism (ρX, ρA) : (X, A) → (D, D),
where (D, D) is the C ∗-correspondence introduced in Example 2.2.
Remark 2.7. It is easy to see that this definition is equivalent to that of a covariant rep-
resentation, given in [18]. If we consider morphisms that only satisfy conditions (C1) and
(C2), then we recover the original definition of a representation of a C ∗-correspondence.
Furthermore, it follows from the somorphism K(D) ∼= D that JD = D, and hence con-
dition (C3) is automatic for morphisms of this form. Condition (C4) is the covariance
condition of [18]; i.e. it says that
X(φX(a)) = ι−1
ρ+
D (a)
where ιD : K(D) → D is the isomorphism from example 2.2.
Definition 2.8. [18, Definition 2.6] Define the OX to be the universal C ∗-algebra gen-
erated by the image of (X, A) under the universal covariant representation (πX , πA).
6
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
It is easy to show the existence of such a universal representation, which in fact is
always injective.
Now, we want to construct a functor F from our category C to the category of C ∗-
algebras, satisfying F (X, A) = OX .
In order to do this, we must first see that any
morphism (ψX , ψA) : (X, A) → (Y, B) extends to a unique C ∗-homomorphism, which we
denote by Ψ : OX → OY .
Proposition 2.9. Let (ψX , ψA) : (X, A) → (Y, B) be a morphism of C ∗-correspondences.
Then this morphism extends to a unique C ∗-homomorphism Ψ : OX → OY such that the
following diagram commutes.
(X, A)
(ψX , ψA)
............
.......................................................................................................................................................................................
(Y, B)
(πX , πA)
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
OX
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
(πY , πB)
OY
Ψ
............
................................................................................................................................................................................................................................................
Proof. It follows from the definition that (πY ◦ψX , πB ◦ψA) is a covariant representation of
(X, A) on OY . Hence the universal property of OX implies the existence of the required
∗-homomorphism Ψ : OX → OY such that the diagram commutes.
(cid:3)
Definition 2.10. Define a map F from the category C of C ∗-correspondences to the
category of C ∗-algebras which satisfies
F (X, A) = OX
and for a morphism (ψX , ψA) : (X, A) → (Y, B) we let F (ψX, ψA) be the unique map
Ψ : OX → OY satisfying the conditions of Proposition 2.9.
Proposition 2.11. The map F is a covariant functor from the category of C ∗-correspon-
dences to the category of C ∗-algebras.
Proof. First fix C ∗-correspondences (X, A), (Y, B) and (Z, C), and morphisms of C ∗-
correspondences (ψX, ψA) : (X, A) → (Y, B) and (ωY , ωB) : (Y, B) → (Z, C). Since
idX is linear, idA is a homomorphism and idA(JX) = JX , we know that (idX, idA) is a
morphism of C ∗-correspondences. So in order to see that F is a covariant functor, we
need to show that
(1) idF (X,A) = F (id(X,A)); and
(2) F (ψX, ψA) ◦ F (ωY , ωB) = F ((ψX, ψA) ◦ (ωY , ωB)).
For (1), we need only show that this holds on the generators {πX(ξ) : ξ ∈ X} and
{πA(a) : a ∈ A} and this follows from the commutativity of the diagram in Proposition
2.9.
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
7
For (2), again we only see that it holds on generators, so fix ξ ∈ X and a ∈ A. Then
by three applications of the commutativity of the diagram in Proposition 2.9 and the
fact that the composition (ψX , ψA) ◦ (ωY , ωB) is a morphism of C ∗-correspondences we
have
F (ωY , ωB) ◦ F (ψX , ψA)(πX(ξ)) = F (ωY , ωB) ◦ πY ((ψX, ψA)(ξ))
= πZ((ωY , ωB) ◦ (ψX , ψA)(ξ))
= F ((ωY , ωB) ◦ (ψX, ψA))(πX (ξ))
and similarly we can show that
F (ωY , ωB) ◦ F (ψX, ψA)(πX (a)) = F ((ωY , ωB) ◦ (ψX , ψA))(πX (a))
as required. So F is a covariant functor from the category of C ∗-correspondences to the
category of C ∗-algebras.
(cid:3)
In what follows we will just write OX for F (X, A) and use capitalised Greek characters
for induced homomorphisms between C ∗-algebras.
3. Gluing C ∗-correspondences
The purpose of this section is to show that the functor F constructed above is well-
behaved with respect to taking pullbacks. This will be useful in applications when we
consider noncommutative spaces as being constructed by gluing two underlying spaces
together over a common boundary. The idea of taking pullbacks on the level of C ∗-
correspondences motivates the following definition, which is due to Baki´c and Guljas,
[2].
Definition 3.1. Given (X, A), (Y, B) and (Z, C) ∈ C, and morphisms of C ∗-correspon-
dences (ψX , ψA) : (X, A) → (Z, C), (ωY , ωB) : (Y, B) → (Z, C), define the restricted
direct sum
X ⊕Z Y := {(ξ, η) ∈ X ⊕ Y : ψX (ξ) = ωY (η)}.
Proposition 3.2. The restricted direct sum X ⊕Z Y is a C ∗-correspondence over the
C ∗-algebra A ⊕C B defined to be the pullback C ∗-algebra of A and B along ψA and ωB.
Proof. It follows from [2, Lemma 2.1] that X ⊕Z Y is a Hilbert A ⊕C B-module.
In
order to prove that it is also a C ∗-correspondence we need a left action φX⊕Z Y . For
(ξ, η) ∈ X ⊕Z Y and (a, b) ∈ A ⊕C B we define
φX⊕Z Y (a, b)(ξ, η) := (φX(a)ξ, φY (b)η).
To see that this is an element of X ⊕Z Y , first recall that (πZ ◦ ψA, tZ ◦ ψX ) is a covariant
representation of (X, A) on OZ and (πZ ◦ ωY , πC ◦ ωB) is a covariant representation of
8
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
(Y, B) on OZ. Then
πC ◦ ψA(φX(a)ξ) = (πC ◦ ψA)(a)(πZ ◦ ψX )(ξ)
= (πC ◦ ωB)(b)(πZ ◦ ωY )(η)
= πC(ωB(φY (b)η))
and then since πZ is injective, it follows that ψA(φX(a)ξ) = ωB(φY (b)η).
It is clear
from the defintion that it is a ∗-homomorphism so φX⊕Z Y : A ⊕C B → L(X ⊕Z Y ) is a
left-action and (X ⊕C Y, A ⊕C B) is a C ∗-correspondence.
(cid:3)
The following theorem is the main result of this section.
Theorem 3.3. Let (X, A), (Y, B) and (Z, C) be C ∗-correspondences. Fix morphisms of
C ∗-correspondences (ψX , ψA) : (X, A) → (Z, C), (ωY , ωB) : (Y, B) → (Z, C) satisfying
(1) (ψX, ψA) and (ωY , ωB) are surjective morphisms with ψA(ker(φX)) = ωB(ker(φY )),
(2) φX(A) ⊂ K(X) and φY (B) ⊂ K(Y ), and
(3) the ideals ker(φX) and ker(φY ) are complemented; i.e. there exist ideals JA E A
and JB E B such that A = JA ⊕ ker(φX) and B = JB ⊕ ker(φY ).
Then
OX⊕Z Y
∼= OX ⊕OZ OY
where OX ⊕OZ OY is the pullback C ∗-algebra of OX and OY along Ψ and Ω.
We begin by showing that there is a covariant representation of the restricted direct
sum correspondence (X ⊕Z Y, A ⊕C B) on the pullback C ∗-algebra OX ⊕OZ OY . First,
we need a definition.
Definition 3.4. Let (ρX , ρA) be a covariant representation of a C ∗-correspondence (X, A)
on a C ∗-algebra D. Then we say that (ρX , ρA) admits a gauge action if for z ∈ T there
exists a ∗-homomorphism αz : C ∗(ρX , ρA) → C ∗(ρX , ρA) such that αz(ρA(a)) = ρA(a)
and αz(ρX(ξ)) = z ρX(ξ) for all a ∈ A and ξ ∈ X. We say an ideal I ⊂ C ∗(ρX , ρA) is
gauge invariant if αz(I) ⊂ I for all z ∈ T.
It is well-known that the universal covariant representation (πX , πA) of (X, A) admits
a gauge action.
Lemma 3.5. Let (ψX, ψA) be a morphism of C ∗-correspondences, with associated C ∗-
homomorphism Ψ : OX → OY . Then ker(Ψ) is a gauge invariant ideal of OX . Further-
more, if we assume that (ψX , ψA) is surjective, φX(A) ⊂ K(X) and ker(φX) is comple-
mented then ker(Ψ) is the ideal generated by πX (ker(ψA)).
Proof. Let γ : T → Aut(OX ) be the gauge action. To see that ker(Ψ) is a gauge invariant
ideal it is enought to show that Ψ commutes with the automorphism γz for any z ∈ T.
This is easily seen to hold the generators {tX(ξ) : ξ ∈ X} and {πX(a) : a ∈ A}.
Now suppose the morphism satisfies the extra hypotheses. We begin by showing that
the ideal ker(ψA) is X-invariant and X-saturated in the sense of [24]. That is, we need
to show that
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
9
(1) φX(ker(ψA))X ⊂ X ker(ψA), and
(2) a ∈ JX and φX(a)X ⊂ X ker(ψA) =⇒ a ∈ ker(ψA).
For (1), suppose ξ ∈ X ker(ψA). Then [20, Proposition 1.3] implies that hξ, ξi ∈ ker(ψA),
so we have ψX(ξ) = 0 by definition of a morphism. Likewise, if ξ ∈ ker(ψX ), then hξ, ξi ∈
ker(ψA) so [20, Proposition 1.3] implies ξ ∈ X ker(ψA). So we have ker(ψX ) = X ker(ψA)
and condition (1) easily follows.
For condition (2), fix a ∈ JX and suppose φX (a)X ⊂ X ker(ψA). From the argument
above this is equivalent to ψX (φX(a)X) = {0}. Since ψX is surjective, this means we
must have φY (ψA(a)) = 0 and hence ψA(a) = 0 since ψA(a) ∈ JY and φY is injective on
JY . So a ∈ ker(ψA).
Finally, [24, Theorem 6.4] implies that ker(Ψ) is the ideal generated by πX (π−1
X (ker(Ψ))) =
πX (ker(ψA)) as required.
(cid:3)
Definition 3.6. Define a pair of maps (ρX⊕Z Y , ρA⊕C B) : (X ⊕Z Y, A⊕C B) → OX ⊕OZ OY
as follows. The map ρA⊕C B satisfies
and ρX⊕Z Y satisfies
ρA⊕C B(a, b) = (πA(a), πB(b))
ρX⊕Z Y (ξ, η) = (πX (ξ), πY (η)).
Proposition 3.7. The pair (ρX⊕Z Y , ρA⊕C B) is an injective covariant representation of
the restricted direct sum (X ⊕Z Y, A ⊕C B) on the pullback C ∗-algebra OX ⊕OZ OY .
Furthermore, this representation admits a gauge action.
Proof. It is easy to see from the definition that ρA⊕C B is a ∗-homomorphism and that
ρX⊕Z Y is a linear map. Furthermore, routine calculations show that conditions (C1) and
(C2) are satisfied. We know ρA⊕C B is injective because the universal homomorphisms πX
and πY are injective. Remark 2.7 implies that (C3) is automatic in this case. For (C4),
since πX and πY are covariant representations, it is enough to show that (a, b) ∈ JX⊕Z Y
implies a ∈ JX and b ∈ JY . First notice that we have an inclusion K(X ⊕Z Y ) ⊂
K(X) ⊕ K(Y ) given on generators by
θ(ξ,η),(ξ′,η′) 7→ (θξ,ξ′, θη,η′).
So (a, b) ∈ JX⊕Z Y implies φX(a) ∈ K(X) and φY (b) ∈ K(Y ). Now, fix c ∈ ker(φX). Then
since ψA(ker(φX)) = {0} = ωB(ker(φY )) we can find an element d ∈ ker(φY ) so that
(c, d) ∈ ker(φX⊕Z Y ). Hence
(a, b)(c, d) = 0 =⇒ ac = 0
and we see that a ∈ JX. Likewise, we may show that b ∈ JY and we have an injective
covariant representation as required.
Finally, it is easy to show that (ρX⊕Z Y , ρA⊕C B) admits a gauge action simply because
(cid:3)
both of the universal representations πX and πY admit gauge actions.
10
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
So we have an injective covariant representation of (X ⊕Z Y, A ⊕C B) on the pullback
C ∗-algebra OX ⊕OZ OY so in particular, the universality of OX⊕Z Y implies that there
exists a homomorphism P : OX⊕Z Y → OX ⊕OZ OY . We want to see that this is an
isomorphism. To do this, we use the gauge-invariant uniqueness theorem, originally
proved for the injective left-action case in [10] and then in the general case by Katsura,
[19, Theorem 6.4]. It is restated without proof here for the readers convenience.
Theorem 3.8. [19, Theorem 6.4] For a covariant representation (ρX, ρA) of a C ∗-
correspondence (X, A), the ∗-homomorphism P : OX → C ∗(ρX, ρA) is an isomorphism if
and only if (ρX, ρA) is injective and admits a gauge action.
We now have the required results to begin the proof of Theorem 3.3.
Proof of Theorem 3.3. We have an injective, covariant representation ρ = (ρX⊕Z Y , ρA⊕C B)
of X ⊕Z Y on the pullback C ∗-algebra OX ⊕OZ OY . So the universality of OX⊕Z Y induces
a homomorphism φ : OX⊕Z Y → OX ⊕OZ OY . Let α, β be the gauge actions on OX and
OY respectively. Then clearly we get a gauge action α ⊕ β on the pullback OX ⊕OZ OY
that is compatible with ρ. So the gauge invariant uniqueness theorem implies that φ is
in fact an injection. It only remains to be seen that this map is a surjection.
Fix (x, y) ∈ OX ⊕OZ OY . Since the C ∗-correspondence morphisms (ψX, ψA) and
(ωY , ωB) are surjective, we can find an element (x, y′) ∈ OX⊕Z Y with y′ ∈ Ω−1(Ψ(x)). By
definition of this element, we must have Ω(y) = Ω(y′) = Ψ(x) and hence y − y′ ∈ ker(Ω).
Hence, we can write (x, y) as a sum
(x, y) = (x, y′) + (0, w)
with (x, y′) ∈ OX⊕Z Y and w = y − y′ ∈ ker(Ω). So to see that (x, y) ∈ OX⊕Z Y , we need
only show that (0, w) ∈ OX⊕Z Y . It is enough to show that ker(Ω) is the ideal generated
by πY (ker(ωB)), and this follows from Lemma 3.5.
So we have the required isomorphism.
(cid:3)
4. Construction of 2n-dimensional mirror quantum spheres
As an application of Theorem 3.3, in this section we present a realisation of the 2n
q,β) where q ∈ (0, 1), as a C ∗-algebra associated
dimensional mirror quantum sphere C(S2n
to a C ∗-correspondence. For the original construction and results on this object we refer
the reader to [16]. We recall the basic definition here for the convenience of the reader.
Firstly, fix n ∈ N. We construct the 2n dimensional quantum sphere by gluing two 2n
dimensional quantum discs along their boundary, a 2n − 1 dimensional quantum sphere.
The 'mirroring' is obtained by pre-composing one of the discs with an automorphism that
'flips' it prior to gluing. More concretely, we construct it as the pullback of the following
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
11
diagram
C(D2n
q )
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
β ◦ π
π
where π : C(D2n
q ) → C(S2n−1
q
............
.......................................................................................................................................................................................
C(D2n
q )
) is the natural surjection and β ∈ Aut(C(S2n−1
C(S2n−1
)
q
)).
q
The first step is to represent each C ∗-algebra as a C ∗-algebra generated by a C ∗-
q ) is isomorphic to the graph algebra
n = {v1, . . . , vn+1}, edge set
n = {ei,j : 1 ≤ i ≤ n, i ≤ j ≤ n + 1} and range and source maps r and s satisfying
correspondence. We know from [16] that C(D2n
C ∗(Mn), where Mn is the graph with n + 1 vertices M 0
M 1
s(ei,j) = vi and r(ei,j) = vj.
Now, consider the vector space X = span{wi,j : i = 1, . . . , n, j = i, . . . , n + 1} and C ∗-
algebra generated by mutually orthogonal projections A = span{Pi : i = 1, . . . , n + 1}.
Define an A valued inner-product on X satisfying
hwi,j, wk,li = δi,kδj,lPj
and a right action of A on X satisfying
Define a left action φX : A → L(X) satisfying
wi,jPk = δj,kwi,j.
φX(Pk)wi,j = δi,kwi,j.
Then it is easily checked that X ∼= Cc(M 1
n) and the C ∗-correspondence
structure matches that defined in [18, Section 3.4], so [18, Proposition 3.10] implies that
we have an isomorphism OX
n) and A ∼= C0(M 0
∼= C(D2n
q ).
Similarly for C(S2n
q ), consider the vector space
Z = span{zi,j : i = 1, . . . , n, j = i, . . . , n}
and C ∗-algebra generated by mutually orthogonal projections
C = span{Si : i = 1, . . . , n} ∼= Cn.
Then OZ
that we are using to construct Z and C.
∼= C(S2n−1
q
). We refer the reader to [16] for the details of the directed graph
Now, we want to see that there exists a morphism of C ∗-correspondences (ψX , ψA) :
(X, A) → (Z, C) such that the induced homomorphism Ψ : OX → OZ and the surjec-
tion π : C(D2n
q ) and the
isomorphism OZ
q ) are intertwined by the isomorphism OX
q ). Define
q ) → C(S2n
∼= C(S2n
∼= C(D2n
ψX(wi,j) =(cid:26) zi,j
0
for 1 ≤ i ≤ n and i ≤ j ≤ n
if i = n, j = n + 1
12
and
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
ψA(Pi) =(cid:26) Si
for 1 ≤ i ≤ n
0 if i = n + 1
Then it is routine to show that this is a morphism of C ∗-correspondences and the induced
map Ψ : OX → OZ is the desired homomorphism. So we have
Ψ(πA(wi,j)) = πZ(zi,j) for i = 1, . . . , n, j = i, . . . , n
Ψ(πA(wn,n+1)) = 0
Ψ(πA(Pi)) = πC(Si) for i = 1, . . . , n
Ψ(πA(Pn+1)) = 0.
For the mirroring, we compose with the automorphism β ∈ Aut(OZ) satisfying
β(πZ(zi,j)) = πZ(zi,j) for i = 1, . . . , n − 1, j = i, . . . , n − 1
β(πZ(zn,n)) = πZ(zn,n)∗
β(πC(Si)) = πC(Si)
However, we are not yet ready to use Theorem 3.3 because the map β ◦Ψ does not come
from a C ∗-correspondence morphism from (X, A) to (Z, C). So we must find another C ∗-
correspondence (Y, B) and a C ∗-correspondence morphism (ωY , ωB) : (Y, B) → (Z, C)
q ) and the extension Ω : OY → OZ of (ωY , ωB) is the C ∗-
such that OY
homomorphism β ◦ Ψ.
∼= C(D2n
Let Y1 := {xi,j : 1 ≤ i ≤ n − 1, i ≤ j ≤ n + 1}, Y2 := {xi,n,j : 1 ≤ i ≤ n − 1, j ≥ 1} and
Y3 := {y, y′, yi : i ≥ 1} be sets of generators and define a vector space
Y := span(Y1 ∪ Y2 ∪ Y3).
Let B be the C ∗-algebra generated by a set of n + 1 mutually orthogonal non-zero
projections {Ri
: j ≥ 1} of mutually orthogonal
projections satisfying Qj ≤ Rn for all j ≥ 1.
Define a B valued inner-product on Y by
: i = 1, . . . , n + 1} and a set {Qj
hxi,j, xk,li = δi,kδj,lRj
hxi,j, yi = 0
hxi,n,j, yi = 0
hxi,n,j, xk,n,li = δi,kδj,lQj
hy, y′i = Q1
hy′, y′i = Q1
hyi, yji = δi,jQi+1
hxi,j, xk,n,li = δi,kδj,nQl
hxi,j, yki = 0
hxi,n,j, yki = 0
hy, yi = Rn
hy, yii = Qi+1
hy′, yii = 0
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
13
and a right action of B on Y given by
xi,jRk = δj,kxi,j
xi,n,jRk = δn,kxi,n,j
xi,jQk = δj,nxi,n,k
xi,n,jQk = δj,kxi,n,j
yRk = δn,ky
y′Rk = δn,ky′
yiRk = δn,kyi
yk−1 if k ≥ 2
y′ if k = 1
0 otherwise
yQk =
y′Qk = δ1,ky′
yiQk = δi+1,kyi.
We can define a left-action φY : B → L(Y ) by
φY (Rk)xi,j = δi,kxi,j
φY (Rk)y = δn,k(y − y′) + δn+1,ky′
φY (Rk)yi = δn,kyi
φY (Qk)xi,n,j = 0
φY (Qk)y′ = 0
φY (Rk)xi,n,j = δi,kxi,n,j
φY (Rk)y′ = δn+1,ky′
φY (Qk)xi,j = 0
φY (Qk)y = yk
φY (Qk)yi = δi,kyi.
Tedious, but straightforward calculations show that (Y, B) is a C ∗-correspondence.
Theorem 4.1. The C ∗-algebras OX and OY are isomorphic.
Before we can prove Theorem 4.1 we need some preliminary results.
Lemma 4.2. The ideal JX is generated by the projections Pi for 1 ≤ i ≤ n. Furthermore,
we have
(1) φX(Pi) =Pn
(2) φX(Pn) = θwn,n,wn,n + θwn,n+1,wn,n+1.
j=i θwi,j ,wi,j for 1 ≤ i ≤ n − 1; and
Proof. We can immediately see from the definition of φX that ker(φX) is generated by
Pn+1, so it will suffice to show (1) and (2). For 1 ≤ i ≤ n − 1 have
n
Xj=i
θwi,j ,wi,j! (wk,l) =
=
= δi,kwk,l
= φX (Pi)wk,l.
wi,jhwi,j, wk,li
δi,kδj,lwi,jPj
n
n
Xj=i
Xj=i
Similarly, (θwn,n,wn,n + θwn,n+1,wn,n+1)wi,j = φX(Pn)wi,j as required.
j=i θwi,j ,wi,j agree on generators, they must be the same operator.
(cid:3)
So since φX(Pi) and Pn
Lemma 4.3. The ideal JY is equal to B. Furthermore we have the following relations:
(1) For 1 ≤ i ≤ n − 1 we have φY (Ri) =
θxi,j ,xi,j
n
Xj=i
14
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
(2) φY (Rn) = θy−y′,y−y′
(3) φY (Rn+1) = θy′,y′
(4) For i ≥ 1 we have φY (Qi) = θyi,yi.
Proof. It is not hard to see from the definition that φY is injective, so to see that JY = B
we need only show that each generator is an element of K(Y ). So it suffices to show that
the relations (1) to (4) hold. This is straightforward verification again.
(cid:3)
Proof of Theorem 4.1. We construct an isomorphism ΠX : OX → OY by showing that
there exists an injective covariant representation (ρX , ρA) of (X, A) on OY and an injective
covariant representation (ρY , ρB) of (Y, B) on OX such that the induced homomorphisms
are bijective and mutually inverse.
Define the linear map ρX on the generators of X by
ρX(wi,j) = πY (xi,j) for i = 1, . . . , n − 1, j = i, . . . , n + 1
ρX (wn,n) = πY (y)∗ − πY (y′)∗
ρX (wn,n+1) = πY (y′)∗
and define the homomorphism ρA on the generators of A by
ρA(Pi) = πB(Ri) for 1 ≤ i ≤ n − 1,
ρA(Pn) = πB(Rn),
ρA(Pn+1) = πB(Rn+1)
Routine caculations show that (ρX , ρA) : (X, A) → OY satisfy (C1) and (C2). We must
show (C4). We know from Lemma 4.2 that JX is generated by the projections Pi where
1 ≤ i ≤ n. So we need to show that we have
ρA(Pi) = ρ+
X (φX(Pi))
for 1 ≤ i ≤ n. This also follows easily from Lemma 4.2. So we have a covariant
representation (ρX , ρA) of (X, A) on OY .
Now, we want to construct another covariant representation of (Y, B) on OX . Define
a linear map ρY : Y → OX by
ρY (xi,j) = πX (wi,j)
ρY (xi,n,j) = πX (wi,n)aj
ρY (y) = πX (wn,n)∗ + πX(wn,n+1)∗
ρY (y′) = πX (wn,n+1)∗
ρY (yi) = ai(πX (wn,n)∗ + πX (wn,n+1)∗)
and a ∗-homomorphism ρB : B → OX satisfying
ρB(Ri) = Pi ρB(Qi) = ai
where ai is the projection
ai = (πX (wn,n) + πX (wn,n+1))iπA(Pn+1)(πX (wn,n)∗ + πX (wn,n+1)∗)i ∈ OX.
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
15
We can easily show from Lemma 4.3 that (ρY , ρB) : (Y, B) → OX is also a covariant
representation.
Now, each representation induces an injective map, ΠX : OX → OY and ΠY : OY →
OX so all that remains to be seen is that ΠX ◦ ΠY = 1OY and ΠY ◦ ΠX = 1OX . This is
straightforward, and the result follows.
(cid:3)
Now let ωY : Y → Z be the unique linear map satisfying πZ(ωY (ξ)) = Ψ(ΠY (πY (ξ)))
and ωB : B → C be the unique homomorphism satisfying πC(ωB(b)) = Ψ(ΠY (πB(b))).
Then it is easily checked that (ωY , ωB) is a C ∗-correspondence morphism, and furthermore
Ω = β ◦ Ψ ◦ ΠY .
Before we can apply Theorem 3.3 we need to show that these C ∗-correspondences sat-
isfy the hypotheses. We have both ψX and ωY surjective, and also ψA and ωB surjective.
Lemma 4.2 and Lemma 4.3 imply that both left actions are by compact operators. Fur-
thermore, ψA(ker(φX)) = {0} = ωB(ker(φY )). If we set JA := C ∗(P1, . . . Pn), then it is
clear that JA is an ideal in A and A = JA ⊕ ker(φX), so ker(φX) is complemented. The
ideal ker(φY ) is just the zero ideal, so it is also trivially complemented. Hence we can
use Theorem 3.3. There is an isomorphism between the pullback OX ⊕OZ OY and the
C ∗-algebra OX⊕Z Y , where the underlying Banach space satisfies
X ⊕Z Y = span({(wi,j, xi,j : 1 ≤ i ≤ n − 1, i ≤ j ≤ n + 1
∪ {(0, xi,n,j) : 1 ≤ i ≤ n, j ≥ 1} ∪ {(0, yi) : i ≥ 1}
∪ {(wn,n, y), (wn,n+1, 0), (0, y′)})
and the pullback C ∗-algebra A ⊕C B is generated by projections
{(Pi, Ri : 1 ≤ i ≤ n)} ∪ {(0, Qi : i ≥ 1} ∪ {(Pn, Rn), (Pn+1, 0), (0, Rn+1)}
where (0, Qi) ≤ (Pn, Rn), and all other projections are orthogonal. So we have shown
that C(S2n
q,β) ∼= OX⊕Z Y is a C ∗-algebra associated to a C ∗-correspondence.
5. C ∗-algebras of Labelled Graphs
Now, we can use this characterisation to prove that C(S2n
q,β) is actually a C ∗-algebra
associated to a labelled graph, first introduced by Bates and Pask in [4]. We begin by
recalling basic definitions and results from [4], and then showing that with an appropriate
notion of morphisms between labelled spaces, the class of labelled spaces forms a category.
Furthermore, there is a functor from this category to the category of C ∗-correspondences
constructed earlier.
Definition 5.1. A labelled graph (E, L) over an alphabet A is a directed graph E
together with a labelling map L : E1 → A which assigns to each edge e ∈ E1 a label
a ∈ A.
It is worth noting at this point that there are no conditions imposed on the map L;
it is not assumed to be either injective or surjective. However, in practice we generally
16
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
make L surjective by simply discarding any elements of A which do not lie in the range
of L.
Define a word to be a finite string a1a2 . . . an, with each ai ∈ A. We write A∗ for the
collection of all words in A and then for any n ∈ N, the labelling can be extended to En
by defining L(e1e2 . . . en) = L(e1)L(e2) . . . L(en) ∈ A∗. We write L∗(E) = ∪n≥1L(En).
An element α ∈ L∗(E) is called a labelled path.
Definition 5.2. Let (E, L) be a labelled graph. We say that (E, L) is left-resolving if
for all v ∈ E0, the labelling L restricted to r−1(v) is injective.
In other words, a labelled graph is left-resolving if all edges entering a particular vertex
carry different labels.
Definition 5.3. Let (E, L) be a labelled graph, let A ⊂ E0 and let α ∈ L(E∗) be a
labelled path. The relative range of α in A, denoted r(A, α) is defined to be the set
r(A, α) := {r(λ) : λ ∈ E∗, L(λ) = α and s(λ) ∈ A}
Now, let B ⊂ 2E 0 be a collection of subsets of E0. We say that B is closed under
relative ranges if for any A ∈ B and α ∈ L∗(E), we have r(A, α) ∈ B. We say that
B is accommodating for (E, L) if it is closed under relative ranges, contains r(α) for all
α ∈ L∗(E), contains {v} whenever v is a sink, and is also closed under finite unions and
intersections.
Definition 5.4. A labelled space is a triple (E, L, B) where (E, L) is a labelled graph
and B is accommodating for (E, L).
Definition 5.5. Let (E, L, B) be a labelled space. We say (E, L, B) is left-resolving
if (E, L) is left resolving and we say that (E, L, B) is weakly left-resolving if for every
A, B ∈ B and α ∈ L∗(E), we have
r(A, α) ∩ r(B, α) = r(A ∩ B, α)
Note that we will need to assume our labelled spaces are weakly left-resolving in order
for the associated C ∗-algebras to be non-degenerate.
We now have the required definitions to define the C ∗-algebra associated to a labelled
space.
Definition 5.6. Let (E, L, B) be a weakly left-resolving labelled space. A representation
of (E, L, B) is a collection {pA : A ∈ B} of projections and a collection {sa : a ∈ L(E1)}
of partial isometries such that:
(1) For A, B ∈ B, we have pApB = pA∩B and pA∪B = pA + pB − pA∩B where p∅ = 0;
(2) For a ∈ L(E1) and A ∈ B, we have pAsa = sapr(A,a);
(3) For a, b ∈ L(E1), we have s∗
asb = 0 unless a = b; and
asa = pr(a) and s∗
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
17
(4) For A ∈ B define L1(A) := {a ∈ L(E1) : s(a) ∩ A 6= ∅}. Then if L1(A) is finite
and non-empty, we have
pA = Xa∈L1(A)
sapr(A,a)s∗
a + Xv∈A:v is a sink
p{v}.
As noted in Remark 3.2 of [5], there was an error in the original definition of C ∗(E, L, B),
where projections at sinks would be degenerate. Hence the proof of the existence of the
C ∗-algebra associated to a labelled graph given in [4] doesn't hold when the underlying
graph contains sinks. Since we will be looking specifically at a labelled graph with sinks,
we reprove the result in the required generality here. The proof closely follows that of
the proof of the existence of C ∗(E) when E is a directed graph with sinks.
Proposition 5.7. Let (E, L, B) be a weakly left-resolving labelled space. Then there
exists a C ∗-algebra B generated by a universal representation of {sa, pA} of (E, L, B).
Furthermore, the sa's are nonzero and every pA with A 6= ∅ is nonzero.
Proof. Fix a weakly left-resolving labelled space (E, L, B). We define a new directed
graph F by
F 0 = E0 ∪ {vi : v is a sink, i ∈ N} and F 1 = E1 ∪ {fv,j : v is a sink, j ∈ N}
and extend the range and source maps on E to F by
s(fv,i) =(cid:26) v if i = 1
vi−1 if i > 1
and r(fv,i) = vi.
We also extend the labelling L on E to a labelling on F by L(F 1\E 1 = id. Define
the set BF ⊂ 2F 0 to be the smallest subset containing B ∪ {{vi} : vi ∈ F 0\E0} that is
accommodating for (F, L). Then it is easily checked that (F, L, BF ) is a labelled space,
and weakly left resolving if and only if (E, L, B) is. Furthermore, since F has no sinks it
follows from [4, Theorem 4.5] that C ∗(F, L, BF ) exists.
Let {sa : a ∈ L(F 1)}, {pA : A ∈ BF } be the universal representation of (F, L, BF ).
Then we claim that the restriction {sa : a ∈ L(E1)}, {pA : A ∈ B} is a representation
of (F, L, BF ).
Indeed, condition (1) - (3) hold in (E, L, B) as they do in (F, L, BF ).
However condition (4) is not so clear as any sink in E is not a sink in F . So fix a set
AE ∈ B satisfying the conditions of (4). If AE does not contain any sinks then the result
is obvious, so assume that AE contains finitely many, but at least one sink.
18
DAVID ROBERTSON AND WOJCIECH SZYMA ´NSKI
Denote by AF the identical set considered as an element of BF . Then it is easy to see
that L1
AE ∪ {{fv,1} : v ∈ AE is a sink} = L1
pAE = pAF
AF . So
pAF sas∗
a
pAE sas∗
a +
pAE sas∗
a +
pAE sas∗
pAE sas∗
AF
AE
AE
= Xa∈L1
= Xa∈L1
= Xa∈L1
= Xa∈L1
= Xa∈L1
AE
AE
X{{fv,1}:v∈AE is a sink}
X{{fv,1}:v∈AE is a sink}
Xb∈L1
a + X{v∈AE is a sink}
a + X{v∈AE is a sink}
p{v}
{v}
pAE sfv,1s∗
fv,1
p{v}sfv,1s∗
fv,1
b
p{v}sbs∗
since L1
{v} = {fv,1}
since {v} ∈ BF satisfies (4)
as required. So we have the required result.
We can now make the following definition.
(cid:3)
Definition 5.8. Let (E, L, B) be a weakly left-resolving labelled space. Define C ∗(E, L, B)
to be the universal C ∗-algebra generated by representations of (E, L, B).
Now that we have the definitions in place, we are ready to prove that the labelled
spaces form a category and that there is a functor from this category to the category of
C ∗-correspondences such that the associated C ∗-algebras are isomorphic.
We define the objects of our category to be labelled spaces (E, L, B). A morphism
ψE : (E, LE, BE) → (F, LF , BF ) is a pair of maps
ψE 0 : E0 ∪ {0} → F 0 ∪ {0} and ψE 1 : E1 ∪ {0} → F 1 ∪ {0}
satisfying
(L1) ψE 0(0) = ψE 1(0) = 0,
(L2) ψE 0(u) = ψE 0(v) 6= 0 implies v = w,
(L3) LF (ψE 1(e)) = LF (ψE 1(f )) 6= 0 implies LE(e) = LE(f ) where LE(0) = LF (0) = 0,
(L4) {ψE 0(v) : v ∈ r(a)} = {r(ψE 1(e)) : LE(e) = a} for all a ∈ LE(E1),
(L5) {ψE 0(v) : v ∈ A} \ {0} ∈ BF for all A ∈ BE and 0 < L1(A) < ∞ implies
0 < L1(ψE 0(A)) < ∞.
Properties (L3) and (L5) ensure that we can extend these maps to LE(E1) and BE so we
define ψE(a) := LF (ψE 1(e)) where LE(e) = a and ψE(A) := {ψE 0(v) : v ∈ A} \ {0} for
all a ∈ L(E1) and A ∈ BE.
C ∗-ALGEBRAS ASSOCIATED TO C ∗-CORRESPONDENCES
19
It is not hard to see that this gives a category with the obvious domain and range
maps, identity morphism and composition of morphisms.
Now that we want to show that C ∗(E, L, B) can be naturally realised as a C ∗-algebra
associated to a C ∗-correspondence, and that there is a functor between the category
of labelled spaces and the category of C ∗-correspondences that preserves the associated
C ∗-algebras up to isomorphism.
Proposition 5.9. Let (E, L, B) be a weakly left-resolving labelled space. Then there exists
a C ∗-correspondence (X(E), A(E)) such that OX(E)
∼= C ∗(E, L, B).
Proof. Let A(E) be the C ∗-subalgebra of C ∗(E, L, B) generated by the set of projections
{pA : A ∈ B} and X(E) be the Banach subspace of C ∗(E, L, B) spanned by the elements
{sapB : a ∈ L(E1), B ∈ B}. Define the right action of A(E) on X(E) simply by
multiplication, and the inner-product by'
hsapB, scpDi := (sapB)∗scpD = δa,c pB∩D∩r(a).
Similarly, define the left action φX(E) by
φX(E)(pC)sapB := sapB∩r(a,C).
Then we have a C ∗-correspondence. In order to see that there is an isomorphism between
OX(E) and C ∗(E, L, B), we show that there is a covariant representation of (E, L, B) inside
OX(E). Indeed, by construction of X(E) and A(E) it is easily show that
{πA(E)(pA) : A |
1008.2603 | 2 | 1008 | 2011-07-24T09:33:35 | The L^p-Fourier transform on locally compact quantum groups | [
"math.OA",
"math.QA"
] | Using interpolation properties of non-commutative L^p-spaces associated with an arbitrary von Neumann algebra, we define a L^p-Fourier transform 1 <= p <= 2 on locally compact quantum groups. We show that the Fourier transform determines a distinguished choice for the interpolation parameter as introduced by Izumi. We define a convolution product in the L^p-setting and show that the Fourier transform turns the convolution product into a product. | math.OA | math |
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM
GROUPS
MARTIJN CASPERS
Abstract. Using interpolation properties of non-commutative Lp-spaces associated with an
arbitrary von Neumann algebra, we define a Lp-Fourier transform 1 ≤ p ≤ 2 on locally
compact quantum groups. We show that the Fourier transform determines a distinguished
choice for the interpolation parameter as introduced by Izumi. We define a convolution
product in the Lp-setting and show that the Fourier transform turns the convolution product
into a product.
1. Introduction
The Fourier transform is one of the most powerful tools coming from abstract harmonic
analysis. Many classical applications, in particular in the direction of Lp-spaces, can be found
in for example [6]. Here we extend this tool by giving a definition of a Fourier transform on
the non-commutative Lp-spaces associated with a locally compact quantum group. This gives
a link between quantum groups and non-commutative measure theory.
Recall that the Fourier transform on locally compact abelian groups can be defined in an
Lp-setting for p any real number between 1 and 2. This is done in the following way. Let G
be a locally compact group and let G be its Pontrjagin dual. For a L1-function f on G, we
define its Fourier transform f to be the function on G, which is defined by
(1.1)
f (π) =Z f (x)π(x)dlx,
π ∈ G.
Then f is a continuous function on G vanishing at infinity. So we can consider this transform
as a bounded map F1 : L1(G) → L∞( G). The Plancherel theorem yields that if f is moreover
a L2-function on G, then f is a L2-function on G and this map extends to a unitary map
F2 : L2(G) → L2( G).
It is known that the Fourier transform can be generalized in a Lp-setting by means of the
Riesz-Thorin theorem, see [1]. The statement of this theorem directly implies the following.
For any p, with 1 ≤ p ≤ 2, the linear map L1(G) ∩ L2(G) → L2( G) ∩ L∞( G) : f 7→ f extends
uniquely to a bounded map
Fp : Lp(G) → Lq( G),
1
p
+
1
q
= 1.
This map Fp is known as the Lp-Fourier transform.
Date: May 10th, 2011.
Keywords: Locally compact quantum groups, Fourier transform, Interpolation spaces.
2000 Mathematics Subject Classification numbers: 20G42, 20N99, 47N99.
1
2
MARTIJN CASPERS
Quantum groups have been around for quite some time and have been studied in many
different guises. From the 80'ies onwards quantum groups are studied in an operator algebraic
approach. In particular, a satisfactory C∗-algebraic definition of a compact quantum group
has been given by Woronowicz [21].
Around 2000 locally compact quantum groups were introduced by Kustermans and Vaes
[13], [14], see also [19]. Their definitions include a C∗-algebraic one and a von Neumann
algebraic one. Since their introduction many aspects of abstract harmonic analysis have
been given a suitable extension for quantum groups.
In particular, the Pontrjagin duality
theorem has been generalized to locally compact quantum groups. In fact, this was one of
the main motivations for their definition. So every locally compact quantum group admits
a (Pontrjagin) dual quantum group such that the double dual is isomorphic to the originial
quantum group.
Since we now have a von Neumann algebraic interpretation for quantum groups at hand,
it is natural to ask if the Lp-Fourier transform can be defined in this context. This is for
two reasons. First of all, this framework studies quantum groups in a measurable setting
which appeals to a more general interest: what links can be found between on one hand
non-commutative measure spaces, in particular non-commutative Lp-spaces, and on the other
hand the theory of quantum groups. The Lp-Fourier transform studied in the present paper
establishes such a link. Secondly, the existence of a Pontrjagin dual is always guaranteed in
the Kustermans-Vaes setting. This is an essential ingredient for defining Fourier transforms.
The L1- and L2-Fourier transform already appear in the present theory of quantum groups.
In fact, they are implicitly used to define duals of quantum groups. Let us comment on this.
First of all, the L2-Fourier transform is implicitly used in the construction of the Pontrjagin
dual of a (von Neumann algebraic) quantum group. For the classical case of a locally compact
abelian group G, let (L∞(G), ∆G) be the usual quantum group associated with it. Its dual is
given by (L(G), ∆G), where L(G) is the group von Neumann algebra of G. This structure is
spatially isomorphic to (L∞( G), ∆ G) be means of the L2-Fourier transform. That is L∞( G) =
F2L(G)F −1
Secondly, in a paper by Van Daele [20] he explains how the Fourier transform should be
interpreted on the algebraic level of quantum groups. In his concluding remarks he suggests to
study this transform in the operator algebraic framework. Here, this investigation is carried
out. We take Van Daele's definition, which agrees with the classical transform (1.1), as a
starting point for defining a L2-Fourier transform in the operator algebraic framework.
and similarly the coproduct and other concepts translate.
2
Finally, an operator algebraic interpretation of the Fourier transform can be found in [10].
The main ideas for our L2-Fourier transform first appear here. However, the suggested Fourier
transform [10, Definition 3] is well-defined only if the Haar weights of a quantum group are
states, i.e. if the quantum group is compact. In the more general situation, one has to give a
more careful definition, which we work out in Section 5.
The present paper is related to a collection of papers studying module structures of Lp-
spaces associated to the Fourier algebra of locally compact groups [3], [4], [5]. These papers
are based on the theory of non-commutative Lp-spaces associated to arbitrary (not necessarily
semi-finite) von Neumann algebras, which we recall below.
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
3
When dealing with these spaces, we are confronted with the following obstruction. For
classical Lp-spaces associated with a measure space X, there is a clear understanding of the
the intersections of Lp-spaces by means of disjunction of sets. So Lp(X) ∩ Lp′(X) gives the
intersection of Lp(X) and Lp′(X). For non-commutative Lp-spaces it is more difficult to find
the intersection of two such spaces. In fact, there is a choice which determines the intersection
and which depends on a complex interpolation parameter z ∈ C. In [5] the parameter z = −1/2
is used, whereas [3] focuses on the case z = 0 in order to define module actions. In the final
remarks of [4], it is questioned which parameter would fit best for quantum groups.
One of the results of the present paper is that to define a Lp-Fourier transform, one is obliged
to choose the parameter z = −1/2. We also determine intersections of the L1- and L2-space
and of the L2- and L∞-space associated with a von Neumann algebra for this parameter,
which are natural spaces.
Structure of the paper. In Section 2 we recall the definition of non-commutative Lp-spaces
and introduce the complex interpolation parameter z ∈ C. We specialize the theory for
z = −1/2 and introduce short hand notation. In Sections 3 - 6, we only work with Lp-spaces
with respect to this parameter. The justification for this specialization is given in the final
chapter.
As indicated, the study of the intersections of Lp-spaces becomes more intricate in the non-
commutative setting. In Section 3 we determine the intersections of L1- and L2-space and of
the L2- and L∞-space associated with a von Neumann algebra. These intersections turn out
to be well-known spaces in the theory of quantum groups. This gives a confirmation that our
choice for the interpolation parameter made at the beginning is a natural one. Moreover, it
gives the necessary ammunition to apply the re-iteration theorem. We warn the reader that
the contents of Section 3 are relatively technical and if one is more interested in Fourier theory
on quantum groups, one can skip Section 3 at first reading.
Section 4 recalls the definition of a locally compact quantum group as given by Kustermans
and Vaes. We give the von Neumann algebraic definition.
In Section 5 we define the Lp-Fourier transform. We start with the L1- and L2-theory and
then obtain the Lp-Fourier transform (1 ≤ p ≤ 2) through the complex interpolation method,
a method similar to the Riesz-Thorin theorem mentioned in the introduction.
In Section 6 we define a convolution product in the Lp-setting and show that the Fourier
transform turns the convolution product into a product.
Finally, in Section 7, we prove that the interpolation parameter used in Sections 3 - 6 is
distinguished. That is, we prove that given the L2-Fourier transform, there is only one choice
for the interpolation parameter that allows an Lp-Fourier transform. This justifies our choice
for this parameter made in the beginning.
Notations and conventions. Throughout this paper, let M be a von Neumann algebra and
ϕ a normal, semi-finite, faithful weight on M. We adopt the standard notations from [15].
So nϕ = {x ∈ M ϕ(x∗x) < ∞}, mϕ = n
ϕnϕ. We denote ∇, J, σ for the modular operator,
modular conjugation and modular automorphism group associated with ϕ. We denote Tϕ for
the Tomita algebra defined by
∗
Tϕ =(cid:8)x ∈ M x is analytic w.r.t. σ and ∀z ∈ C : σz(x) ∈ nϕ ∩ n
∗
ϕ(cid:9) .
4
MARTIJN CASPERS
Let (H, π, Λ) be the GNS-representation of M with respect to ϕ. Note that M can be consid-
ered as acting on H and therefore we omit the map π if possible. For x ∈ mϕ and for a ∈ M
analytic with respect to σ, we have ax ∈ mϕ and xa ∈ mϕ and
ϕ(ax) = ϕ(xσ−ia).
For a subset A ⊆ M, we denote A+ for the positive elements in A. Similarly, M +
∗ denotes
the space of positive normal functionals on M. Let ω ∈ M∗. We donote ω ∈ M∗ for the
functional defined by ω(x) = ω(x∗), x ∈ M. For y ∈ M, we denote yω and ωy for the normal
functionals defined respectively by (yω)(x) = ω(xy) and (ωy)(x) = ω(yx) with x ∈ M. Inner
products on a Hilbert spaces are linear in the first entry and anti-linear in the second. Suppose
that M acts on a Hilbert space H and let ξ, η ∈ H. We denote ωξ,η for the normal functional
defined by ωξ,η(x) = hxξ, ηi. The character ι will always stand for the identity homomorphism.
If x is a preclosed operator, we use [x] for its closure.
2. Non-commutative Lp-spaces
To any von Neumann algebra M, there is a way to associate a non-commutative Lp-space
to it. In fact there are many ways to do this. If M is semi-finite, i.e.
it admits a normal,
semi-finite, faithful trace τ , then one can define Lp(M) as the space of closed, densely defined
operators x affiliated with M for which if x = R[0,∞) λdEλ is the spectral decomposition of
x, then
kxkp :=(cid:18)sup
n∈N
τ(cid:18)Z[0,n]
λpdEλ(cid:19)(cid:19)1/p
< ∞.
If M is abelian, we recover the classical spaces Lp(X) for a certain measure space X.
Since the introduction of Tomita-Takesaki theory, Lp-spaces have been defined for von
Neumann algebras that are not semi-finite. Definitions of non-commutative Lp-spaces have
been given by Haagerup [7], [16], Hilsum [8], Terp [17] and Izumi [9]. The definitions can
be shown te be equivalent. That is, the Lp-spaces obtained by the various definitions are
isometrically isomorphic Banach spaces. For a good introduction to this theory, we refer to
[16], where a comparison of Haagerup's definition and Hilsum's definition is made.
Here we mainly use Izumi's definition [9] which is abstract in nature. He defines Lp-spaces
associated with M by means of the complex interpolation method; a method that admits a
property that is reminiscent of the Riesz-Thorin theorem, see the introduction. It is for this
reason that Izumi's definition is the most suitable context to work in.
Drawback of this context is that the more concrete approach of the other definitions is
absent. Whenever it feels appropriate we comment on this.
2.1. The complex interpolation method. We recall the complex interpolation method as
explained in [1, Section 4.1].
Definition 2.1. Let E0, E1 be Banach spaces. The couple (E0, E1) is called a compatible
couple (of Banach spaces) if E0 and E1 are continuously embedded into a Banach space E.
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
5
Note that we suppress E in the notation (E0, E1). We can consider the spaces E0 ∩ E1 and
E0 + E1 interpreted within E and equip them with norms
kxkE0∩E1 = max{kxkE0,kxkE1},
x ∈ E0 ∩ E1,
kxkE0+E1 = inf{kx0kE0 + kx1kE1 x0 + x1 = x}, x ∈ E0 + E1,
which make them Banach spaces. In that case we can consider E0 and E1 as subspaces of
E0 + E1.
Definition 2.2. A morphism between compatible couples (E0, E1) and (F0, F1) is a bounded
map T : E0 + E1 → F0 + F1 such that for any j ∈ {0, 1}, T (Ej) ⊆ Fj and the restriction
T : Ej → Fj is bounded.
Remark 2.3. Let (E0, E1) and (F0, F1) be compatible couples. If T0 : E0 → F0, T1 : E1 → F1
are bounded maps such that T0 and T1 agree on E0 ∩ E1, then we call T0 and T1 compatible
morphisms. In this case, there is a unique bounded map T : E0 + E1 → F0 + F1. This gives
a way to find morphisms of compatible couples.
Now we describe the complex interpolation method. Let (E0, E1) be a compatible couple.
Let S = {z ∈ C 0 ≤ Re(z) ≤ 1} and let S ◦ denote its interior. Let G(E0, E1) be the set of
functions f : S → E0 + E1 such that
(1) f is bounded and continuous on S and analytic on S ◦;
(2) For t ∈ R, j ∈ {0, 1}, f (it + j) ∈ Ej and t 7→ f (it + j) is continuous and bounded with
(3) For j ∈ {0, 1}, kf (it + j)kEj → 0 as t → ∞.
respect to the norm on Ej;
Note that at this point our notation is different from [1] and [9], where G is denoted by F ,
which we reserve for the Fourier transform. For f ∈ G(E0, E1), we define a norm
f = max{sup kf (it)kE0, supkf (it + 1)kE1}.
Let θ ∈ [0, 1]. We define (E0, E1)[θ] ⊆ E to be the space {f (θ) f ∈ G(E0, E1)} with norm
With this norm, (E0, E1)[θ] is a Banach space [1, Theorem 4.1.2].
kxk[θ] = inf{f f (θ) = x, f ∈ G(E0, E1)}.
Definition 2.4. The assignment from compatible couples of Banach spaces to Banach spaces
which is given by Cθ : (E0, E1) → (E0, E1)[θ] is called the complex interpolation method (at
parameter θ ∈ [0, 1]). (E0, E1)[θ] is called a complex interpolation space.
The following Riesz-Thorin-like theorem plays a central role in the present paper. It gives
the functorial property of the complex interpolation method.
Theorem 2.5 (Theorem 4.1.2 of [1]). Let θ ∈ [0, 1]. Let T be a morphism between compatible
couples (E0, E1) and (F0, F1). Then, it restricts to a bounded linear map T : (E0, E1)[θ] →
(F0, F1)[θ]. The norm is bounded by kTk ≤ kT : E0 → F0k1−θkT : E1 → F1kθ.
If we let Cθ of Definition 2.4 act on the morphisms T : (E0, E1) → (F0, F1) of compatible
couples by assigning its restriction T : (E0, E1)[θ] → (F0, F1)[θ] to it, we see that Cθ is a functor.
6
MARTIJN CASPERS
Remark 2.6. Using the notation of Remark 2.3, the compatible morphisms T0, T1 give rise to
a morphism Cθ(T ) : (E0, E1)[θ] → (F0, F1)[θ] on the interpolation spaces with norm kCθ(T )k ≤
kT0k1−θkT1kθ.
We will need the following useful fact, see [1, Theorem 4.2.2].
Lemma 2.7. Let (E0, E1) be a compatible couple and θ ∈ [0, 1]. E0∩E1 is dense in (E0, E1)[θ].
2.2. Izumi's Lp-spaces. In [17] Terp shows that the Lp-spaces as introduced by Hilsum can
be obtained by applying the complex interpolation method to a specific compatible couple
(M, M∗), see [17, Theorem 36]. Izumi [9] realized that there is more than one way to turn
(M, M∗) into a compatible couple in order to obtain the Lp-spaces through interpolation.
His idea is to define non-commutative Lp-spaces as complex interpolation spaces of certain
compatible structures. It is this definition which we recall here.
Here, we present the general picture. However, in the larger part of the present paper, we
only work with the complex interpolation parameters z = −1/2 and z = 1/2 (we introduce
the parameter in a minute). We will specialize the theory for these parameters in Sections 2.3
and 2.4 and introduce short hand notation there. The more general theory is used in Section
7, where we prove that there is in principle only one interpolation parameter that allows a
Lp-Fourier transform, namely z = −1/2.
construction of Lp-spaces can be found in [9].
Definition 2.8. For z ∈ C, we put
Fix a von Neumann algebra M with normal, semi-finite, faithful weight ϕ. The following
L(z) =(cid:8)x ∈ M ∃ϕ(z)
x ∈ M∗ s.t. ∀a, b ∈ Tϕ : ϕ(z)
x (a∗b) = hxJ∇¯zΛ(a) J∇−zΛ(b)i(cid:9) .
The number z ∈ C will be called the complex interpolation parameter.
Remark 2.9. We will mainly be dealing with the cases z = −1/2 and z = 1/2. Note that if
ϕ is a state, then for any x ∈ M, we see that for a, b ∈ Tϕ,
hxJ∇−1/2Λ(a) J∇1/2Λ(b)i = hxJ∇1/2Λ(σi(a)) J∇1/2Λ(b)i = ϕ(bxσ−i(a∗)) = ϕ(a∗bx),
and hence L(−1/2) = M and ϕ(−1/2)
x
= xϕ. Similarly, L(1/2) = M and ϕ(1/2)
x
= ϕx.
The following proposition implies that there are plenty of elements contained in L(z).
Proposition 2.10 (Propostion 2.3 of [9]). T 2
ϕ = {ab a, b ∈ Tϕ} is contained in L(z).
We are now able to construct Izumi's Lp-spaces using the complex interpolation method.
First, we define a compatible couple. For x ∈ L(z), we define a norm:
We define norm-decreasing injections:
kxkL(z) = max{kxk,kϕ(z)
x k}.
i1
(z) : L(z) → M∗ : x 7→ ϕ(z)
x .
Using the duals of the maps, we obtain the following diagram. Note that (i∞
(−z))∗ : M ∗ → L∗
is restricted to M∗.
i∞
(z) : L(z) → M : x 7→ x;
(−z)
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
7
(2.1)
L(z)
M∗
(−z))∗
(i∞
%JJJJJJJJJJ
9tttttttttt
(i1
(−z))∗
i1
(z)
ip
(z)
:vvvvvvvvv
$IIIIIIIIII
/ Lp
i∞
(z)
M
(z)(M)
/ L∗
(−z).
Now, [9, Theorem 2.5] yields that the outer rectangle of (2.1) commutes. This turns (M, M∗)
into a compatible couple of Banach spaces.
Definition 2.11. For p ∈ (1,∞), we define Lp
(M, M∗)[1/p]. We set L1
(z)(M) to be the complex interpolation space
(z)(M) = M∗ and L∞
(z)(M) = M .
By Lemma 2.7, L(z) can be embedded in Lp
(z)(M). This map is denoted by ιp
(z). Note that
by definition of the complex interplation method Lp
(z)(M) is a linear subspace of L∗
(−z).
Notation 2.12. The map i∞
it is convenient to omit the map i∞
role in the statement. Similarly, we do not introduce notation for the inclusion of Lp
L∗
(−z), where p ∈ (1,∞).
A priori one could think that Lp
(z) : L(z) → M is basically the inclusion of a subspace. Therefore,
(z) in our notation if the norms of the spaces do not play a
(z)(M) in
(z ′)(M) with z 6= z′, are different as Banach
spaces. However, Izumi proves that they are isometrically isomorphic. Terp [17] considers
the case z = 0. The main result of [17] is that Lp
(0)(M) is isometrically isomorphic to the
Lp-spaces by Hilsum [8]. We will come back to this Section 2.4.
(z)(M) and Lp
Theorem 2.13 (Theorem 3.8 of [9]). For z, z′ ∈ C, there is an isometric isomorphism
Up,(z ′,z) : Lp
(z)(M) → Lp
(z ′)(M),
such that for a ∈ T 2
ϕ ,
(2.2)
Up,(z ′,z)(ip
(z)(a)) = ip
(z ′)(σi r′ −r
p
p ∈ (1,∞),
−(s′−s)(a)),
where z = r + is and z′ = r′ + is′, r, r′, s, s′ ∈ R.
We emphasize, that although the Lp-spaces appearing in (2.1) are isomorphic for different
complex interpolation parameters, the intersections defined by this figure may be different. In
any case, by [9, Corollary 2.13],
(2.3)
(i1
(−z))∗(L(z)) = (i1
(−z))∗(M) ∩ (i∞
(−z))∗(M∗),
i.e. if one consideres L(z), M, M∗ as subspaces of L∗
(−z), then L(z) = M ∩ M∗.
r
%
,
:
r
$
/
/
,
9
8
MARTIJN CASPERS
2.3. Specializations for the complex interpolation parameters. In the present paper
In order to study these spaces also
we will mainly work with the parameter z = −1/2.
the parameter z = 1/2 will play a role. In this section, we specialize the theory for these
parameters. The following proposition shows that L(−1/2) and L(1/2) can be described by a
condition that is in generally more easy to check. If ϕ is a state it reduces to Remark 2.9.
Proposition 2.14. We have the following alternative descriptions:
(1) Let L = {x ∈ nϕ ∃ xϕ ∈ M∗ s.t. ∀y ∈ nϕ : xϕ(y∗) = ϕ(y∗x)}. Then, L = L(−1/2).
(2) Let R = {x ∈ n
ϕ ∃ϕx ∈ M∗ s.t. ∀y ∈ nϕ : ϕx(y) = ϕ(xy)}. Then, R = L(1/2).
∗
Proof. We only give the proof of (1), since (2) can be proved similarly. We first prove ⊆. For
x ∈ L, a, b ∈ Tϕ,
xϕ(a∗b) = ϕ(a∗bx) = ϕ(bxσ−i(a∗)) = hxΛ(σ−i(a∗)), Λ(b∗)i
=hx∇J∇1/2Λ(a), J∇1/2Λ(b)i = hxJ∇−1/2Λ(a), J∇1/2Λ(b)i.
x
.
To prove ⊇, we first prove that MT 2
ϕ ⊆ L(−1/2). Indeed, let x ∈ M and let c, d ∈ Tϕ. The
Hence x ∈ L(−1/2) and xϕ = ϕ(−1/2)
functional M ∋ y 7→ ϕ(σi(d)yxc) is normal. Furthermore, for a, b ∈ Tϕ,
hxcdJ∇−1/2Λ(a), J∇1/2Λ(b)i = hΛ(xcdσ−i(a∗)), Λ(b∗)i = ϕ(bxcdσ−i(a∗)) = ϕ(σi(d)a∗bxc).
Hence, xcd ∈ L(−1/2).
Next, we prove that L(−1/2) ⊆ nϕ. Take x ∈ L(−1/2) and let (ej)j∈J be a bounded net in
Tϕ such that σi(ej) is bounded and such that ej → 1 σ-weakly, see [17, Lemma 9]. Then,
xej → x σ-weakly. Furthermore,
kΛ(xej)k2 = ϕ(e∗
(2.4)
j x∗xej) = ϕ(−1/2)
j )(x∗) ≤ kϕ(−1/2)
xejσ−i(e∗
j )kkxk,
xejσ−i(e∗
where the second equality is due to the previous paragraph. By [9, Proposition 2.6],
(2.5)
ϕ(−1/2)
xejσ−i(e∗
j ) = ϕ(−1/2)
x
σi(ej)e∗
j ,
where for ω ∈ M∗, y ∈ M, ωy is the normal functional defined by (ωy)(a) = ω(ya), a ∈ M.
From (2.4) and (2.5) it follows that (Λ(xej))j∈J is a bounded net. Furthermore, for a, b ∈ Tϕ,
hΛ(xej), Λ(ab)i = ϕ(b∗a∗xej) = ϕ(a∗xejσ−i(b∗)) → ϕ(a∗xσ−i(b∗)).
Since (Λ(xej))j∈J is bounded, this proves that (Λ(xej))j∈J is weakly convergent. Since Λ is
σ-weak/weak closed, this implies that x ∈ Dom(Λ) = nϕ. So L(−1/2) ⊆ nϕ.
hΛ(x), Λ(ab)i. The proposition then follows since by Lemma A.2, T 2
for Λ. The proposition follows from:
To finish the proof, let again x ∈ L(−1/2) and let a, b ∈ Tϕ. We prove that ϕ(−1/2)
((ab)∗) =
ϕ is a σ-weak/weak-core
x
hΛ(x), Λ(ab)i = ϕ(b∗a∗x) = ϕ(a∗xσ−i(b∗)) = hxJ∇−1/2Λ(b), J∇1/2Λ(a∗)i = ϕ(−1/2)
x
(b∗a∗).
(cid:3)
In particular, it follows from Proposition 2.14 that for y ∈ nϕ,
xϕ(y∗) = ϕ(y∗x),
ϕx(y) = ϕ(xy),
x ∈ L,
x ∈ R.
(2.6)
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
9
We emphasize that one has to be careful that (2.6) does not make sense for every x, y ∈ M.
Also, (2.6)justifies why (2.1) is also called the left injection for z = −1/2 and the right injection
for z = 1/2.
Part of the next Corollary is already proved in [9]. Using the alternative descriptions of
Proposition 2.14, it is easy to prove the remaining statements.
Corollary 2.15. We have inclusions MT 2
ϕ M ⊆ R, T 2
ML ⊆ L and RM ⊆ R. Moreover, R = {x∗ x ∈ L} and for x ∈ L, ϕx∗ = xϕ.
Proof. The first inclusion has already been proved in the proof of Proposition 2.14. Here we
have proved that for x ∈ M, a, b ∈ Tϕ, xabϕ(z) = ϕ(σi(b) z xa), z ∈ M. Similarly, one can
prove that for x, z ∈ M, a, b ∈ Tϕ, yl ∈ L, yr ∈ R,
ϕ ⊆ L∩R, LTϕ ⊆ L, TϕR ⊆ R,
ϕ ⊆ L, T 2
ϕabx(z) = ϕ(bx z σ−i(a)); ϕab(z) = ϕ(b z σ−i(a));
ylaϕ(z) =yl ϕ(σi(a)z);
ϕyrx(z) = ϕyr (xz);
ϕayr (z) = ϕ(1/2)
ϕx∗ = xϕ.
yr
(zσ−i(a));
abϕ(z) = ϕ(σi(b) z a);
xylϕ(z) =yl ϕ(zx);
(cid:3)
Since we are mainly dealing with complex interpolation parameter z = −1/2 and z = 1/2,
it is more convenient to adapt our notation.
Notation 2.16. We use the following short hand notations. For p ∈ [1,∞],
Lp(M)left = Lp
Lp(M)right = Lp
(−1/2)(M), L = L(−1/2),
(1/2)(M), R = L(1/2),
lp = ip
rp = ip
xϕ = ϕ(−1/2)
(−1/2),
(1/2), ϕx = ϕ(1/2)
x
x
for x ∈ L,
for x ∈ R.
Recall that by definition M∗ = L1(M)left and M = L∞(M)left. From now on, we consider M∗
and M as subspaces of R∗ by means of the respective maps r∗
1 and it is convenient to
omit these maps in the notation. So the identifications of M∗ and M in R∗ are given by the
pairings:
∞ and r∗
hω, yiR∗,R = ω(y),
hx, yiR∗,R = ϕy(x),
(2.7)
(2.8)
The norm on L will be denoted by k · kL.
2.4. Comparison with Hilsum's Lp-spaces. Here, we recall the definition of non-commuta-
tive Lp-spaces given in [8], see also [17]. We need these spaces for two reasons.
ω ∈ M∗, y ∈ R,
x ∈ M, y ∈ R.
First of all, many of the objects we introduce are constructed by means of Theorem 2.5.
For that reason the structures are abstract in nature. The advantage of the Hilsum approach
is that it is much more concrete. Hence, also the objects defined in Section 5 have a more
concrete meaning when they are considered in the Hilsum setting.
Secondly, a non-commutative L2-space associated with a von Neumann algebra M with
weight ϕ can be identified with the GNS-space H of the weight. In [17, Theorem 23] this
identification is given for Hilsum's definition. In [9], Izumi does not explicitly keep track of
an isomorphism between L2(M)left with H. Here we make this isomorphism explicit. This
is useful for the Lp-Fourier transform.
In particular, Corollary 5.5 relies heavily on this
identification.
10
MARTIJN CASPERS
We refer to the original paper [8] for Hilsum's Lp-spaces. The following is also nicely
summarized in [16, Sections III and IV]. Fix a normal, semi-finite, faithful weight φ on the
commutant M ′. Let σφ be its modular automorphism group.
Definition 2.17. A closed, densely defined operator x on H is called γ-homogeneous, with
γ ∈ R if the following skew commutation relation holds
xa ⊆ aσφ
iγ(x),
for all a ∈ M ′ analytic w.r.t. σφ.
The following theorem requires the spatial derivate [2], [15]. We will not recall this con-
struction, but rather cite its properties. For a good introduction, we refer to [17, Section III].
The spatial derivative construction gives a passage between M∗ and the (−1)-homogeneous
operators. The following theorem can be found under the given references in [16]. It can be
derived from [2, Theorem 13].
Theorem 2.18 (Theorem 29, Definition 33 and Corollary 34 of [16]). Let x be a closed,
densely defined operator with polar decomposition x = ux. Let p ∈ [1,∞]. The following are
equivalent:
(1) x is (−1/p)-homogeneous;
(2) u ∈ M and xp is (−1)-homogeneous;
(3) u ∈ M and there is a normal, semi-finite weight ψ on M such that xp equals the
spatial derivative dψ/dφ.
Definition 2.19. Let p ∈ [1,∞). The Hilsum Lp-space Lp(φ) is defined as the space of
closed, densely defined operators x on the GNS-space H of ϕ such that if x = ux is the polar
decomposition, then xp is the spatial derivative of a positive ω ∈ M∗ and u ∈ M. It carries
the norm kxkp = (ω(1))1/p. We set L∞(M) = M.
In particular, every operator in Lp(φ) is closed, densely defined and (−1/p)-homogeneous.
This includes p = ∞. By Theorem 2.18, the spatial derivative gives an isometric isomorphism
between M∗ and L1(φ).
We introduce notation for the distinguished spatial derivative
d is a strictly positive, self-adjoint operator acting on the GNS-space H. We need the fact
that it implements the modular automorphism group of ϕ and φ, i.e.
d = dϕ/dφ.
σt(x) = ditxd−it, x ∈ M
σφ
t (x) = d−itydit, y ∈ M ′.
Using this, one can prove that d is (−1)-homogeneous, see [17, Lemma 22]. The operator d
forms a handy tool to find elements of Lp(φ).
Lemma 2.20 (Theorem 26 of [17]). Let p ∈ [2,∞] and let x ∈ nϕ. Then, xd1/p is preclosed and
its closure [xd1/p] is in Lp(φ). Moreover, there is an isometric isomorphism P : H → L2(φ)
given by [xd1/2] 7→ Λ(x).
We use this result to prove the following.
Proposition 2.21. Let p ∈ [1,∞].
(1) Let a, b ∈ Tϕ. Then, abd1/p is preclosed and its closure [abd1/p] is in Lp(φ).
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
11
(2) There is an isometric isomorphism Φp : Lp(φ) → Lp(M)left such that
a, b ∈ Tϕ.
(3) There is a unitary map Ul : L2(M)left → H determined by
Φp : [abd1/p] 7→ lp(ab),
(4) More general, there is a unitary map U(z) : L2
Ul : l2(a) 7→ Λ(a),
a ∈ T 2
ϕ .
(z)(M) → H determined by
U(z) : i2
(z)(a) 7→ Λ(σ−i(z/2+1/4)(a)),
a ∈ T 2
ϕ .
Proof. (1) First note that using [17, Lemma 22] for the first inclusion, Lemma 2.20 for the
third and [8, Theorem 4 (3)] for the last,
abd1/p ⊆ d2/pσ2i/p(ab)d2/p ⊆ d2/pσ2i/p(a) · [σ2i/p(b)d2/p] ∈ Lp/2(φ) · Lp/2(φ) ⊆ Lp(φ).
Hence,
(2.9)
(abd1/p)∗ ⊇ (d2/pσ2i/p(a) · [σ2i/p(b)d2/p])∗ ∈ Lp(φ).
So that (abd1/p)∗ is densely defined. Hence abd1/p is preclosed and by (the proof of) [8, Theorem
4 (1)], (abd1/p)∗ = (d2/pσ2i/p(a) · [σ2i/p(b)d2/p])∗, hence [abd1/p] = d2/pσ2i/p(a) · [σ2i/p(b)d2/p].
compatible couple as considered in [17]. First note that by [17, Eqn. (50)],
(2) It is argued in the remarks following [9, Proposition 2.4] that (2.1) for z = 0 equals the
[abd1/p] = d2/pσ2i/p(a) · [σ2i/p(b)d2/p] = µp(σ2i/p(ab)),
where µp is the embedding of L(0) in Lp(φ), see [17, Theorem 27]. The main result of [17]
is that Lp(φ) is isometrically isomorphic to Lp
(0)(M). The isomorphism is given by the map
νp : Lp(φ) → Lp
(0) by
commutativity of [17, Eqn. (55)]. In turn we have (i1
(0) by commutativity of (2.1).
Hence, we have an isometric isomorphism Lp(φ) → Lp
(0)(M) of [17, Theorem 30]. Moreover, we see that νpµp = (i1
(0))∗i∞
(0)(M) for which
(0) = ip
(0))∗i∞
[abd1/p] 7→ νp([abd1/p]) = νpµp(σ2i/p(ab)) = ip
(0)(σ2i/p(ab)).
We conclude the proof by applying the isometric isomorphism U(−1/2,0) of Theorem 2.13, so
that we we get an isometric isomorphism Φp : Lp(φ) → Lp
(−1/2)(M) = Lp(M)left, such that
Φp : [abd1/p] 7→ U(−1/2,0)ip
(0)(σ2i/p(ab)) = ip
(−1/2)(ab) = lp(ab),
(3) This follows from (2) by applying Lemma 2.20 and the fact that Λ(T 2
So Ul = Φ−1
p P −1. (4) U(z) = UlU2,(−1/2,z).
a, b ∈ Tϕ.
ϕ ) is dense in H.
(cid:3)
Recall that L2(M)left is by definition a subspace of R∗. Therefore, we can pair elements of
L2(M)left with elements of R.
Proposition 2.22. For ξ ∈ H, y ∈ R,
hU ∗
l ξ, yiR∗,R = hξ, Λ(y∗)i.
12
MARTIJN CASPERS
Proof. First assume that ξ = Λ(x) = Ull2(x), x ∈ L. Using the commutativity of (2.1) in the
second equality,
hU ∗
l ξ, yiR∗,R = hl2(x), yiR∗,R = hl1(x), yiR∗,R
=(2.7)(xϕ)(y) =(2.6)ϕ(yx) = hΛ(x), Λ(y∗)i = hξ, Λ(y∗)i.
ϕ ) ⊆ Λ(L) is dense in H.
The proposition follows by the fact that Λ(T 2
Notation 2.23. From now on, we will identify H and L2(M)left and consider it as a subspace
of R∗. The identification is given via the unitary Ul. Under this identification the map
l2 becomes the GNS-map Λ, see Proposition 2.21. By Proposition 2.22 we see that H is
identified as a subspace of R∗ by means of the pairing
(cid:3)
(2.10)
hξ, yiR∗,R = hξ, Λ(y∗)i
ξ ∈ H, y ∈ R.
3. Intersections of Lp-spaces
As indicated in the Section 2, the intersections of the various Lp-spaces depend on the
interpolation parameter z of Definition 2.8. Here we study the intersections of L1(M)left and
L2(M)left, as well as the intersections of L2(M)left and L∞(M)left. The spaces turn out to
be natural and well-known in the theory of locally compact quantum groups. We use the
intersections in order to apply the re-iteration theorem, see [1].
Notation 3.1. In this section, any interpolation space should be understood with respect
to the diagram in (2.1) for the parameter z = −1/2. Recall that we introduced short hand
notation for this diagram in Notations 2.16 and 2.23. Moreover, we identified M∗,H and M
as subspaces of R∗ by means of the pairings (2.7), (2.10) and (2.8). Similarly, any intersection
of two such spaces should be understood as an intersection within R∗. The notation can be
summarized by means of the non-dotted arrows in the following diagram. The dotted part of
the diagram is the main topic of the present section.
M∗
(2.7)
l1
6llllllllll
/_______
RR
ξ
RRRR
lp
Λ
Lp(M)left
(2.10)
)TTTTTTTTTTTTTTTTT
5kkkkkkkkkkkkkkkkk
/ R∗;
L
>
>
>
I
?
l1
>
nϕ
RRR
(R
H
Λ
lq
6llllllllll
)RRRRRRRRRR
/_______
l∞
Lq(M)left
M
(2.8)
6
(
/
)
8
8
.
.
&
&
0
0
/
/
?
/
6
)
/
5
J
J
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
13
3.1. The intersection of M∗ and H. The following set defines the intersection of M∗ and
H.
Definition 3.2. We set:
I = {ω ∈ M∗ Λ(x) 7→ ω(x∗), x ∈ nϕ is bounded } .
By the Riesz theorem, for every ω ∈ I, there exists a ξ(ω) ∈ H such that hξ(ω), Λ(x)i = ω(x∗).
Theorem 3.3. We have I = H ∩ M∗, where the equality should be interpreted within R∗, see
Notation 3.1. Within R∗, ω ∈ I equals ξ(ω) ∈ H.
Proof. We first prove ⊇. Let ξ ∈ H and ω ∈ M∗ be such that ξ = ω in R∗. For y ∈ R,
ω(y) =(2.7)hω, yiR∗,R = hξ, yiR∗,R =(2.10)hξ, Λ(y∗)i.
ϕ is a σ-strong-∗/norm core for Λ, see Lemma A.2. Hence, it
L contains T 2
follows that ω ∈ I.
ϕ . Moreover, T 2
To prove ⊆, let ω ∈ I. For y ∈ R,
hξ(ω), yiR∗,R =(2.10)hξ(ω), Λ(y∗)i = ω(y) =(2.7)hω, yiR∗,R.
Hence, ξ(ω) = ω in R∗.
(cid:3)
Note that (M∗,H) forms a compatible couple. As explained in Section 2.1, the intersection
of these two spaces carries a natural norm for which it is a Banach space. So, for ω ∈ I we
define
kωkI = max{kωk,kξ(ω)k}.
Proposition 3.4. The map k : L → I : x 7→xϕ is injective, norm-decreasing and has dense
range. In fact, k(T 2
Proof. Suppose that x ∈ L and xϕ = 0, then 0 = (xϕ)(x∗) = ϕ(x∗x). So x = 0 and hence k is
injective. For x ∈ L, kxϕk ≤ kxkL and
ϕ ) is k · kI-dense in I.
kξ(xϕ)k = kΛ(x)k = kxϕ(x∗)k1/2 ≤ kxϕk1/2kx∗k1/2 ≤ kxkL,
so that k is norm-decreasing. Now we prove that the range of k is dense in I. We identify
I with the subspace {(ω, ξ(ω)) ω ∈ I} ⊆ M∗ × H. We equip M∗ × H with the norm
k(ω, ξ)kmax = max{kωk,kξk}. The norm coincides with k · kI on I. The dual of (M∗ × H,k ·
kmax) can be identified with (M×H∗,k·ksum), where k(x, ξ)ksum = kxk+kξk. Let N ⊆ M×H∗
be the space of all (y, η) such that h(ω, ξ(ω)), (y, η)iM∗×H,M ×H∗ = 0 for all ω ∈ I. The dual of
I is given by (M × H)/N equipped with the quotient norm.
Now, let (y, η) ∈ M × H be such that
h(xϕ, Λ(x)), (y, η)iM∗×H,M ×H∗ = (xϕ)(y) + hΛ(x), ηi = 0
ϕ . The proof is finished if we can show that (y, η) ∈ N. In order to do this,
(ej). By the assumptions on (y, η), for
for all x ∈ T 2
let (ej)j∈J be a net as in Lemma A.1. Put aj = σ− i
x ∈ Tϕ,
(3.1)
2
(xaj ϕ)(y) = −hΛ(xaj), ηi.
14
MARTIJN CASPERS
For the left hand side we find by Corollary 2.15,
(3.2)
(xaj ϕ)(y) = ϕ(σi(aj)yx) = hΛ(x), Λ(y∗σi(aj)∗)i,
where the first equality follows from [9, Proposition 2.3]. For the right hand side of (3.1) we
find
(3.3)
hΛ(xaj), ηi = hJσ− i
2
(a∗
j )JΛ(x), ηi = hΛ(x), Jσ i
2
(aj)Jηi.
Hence (3.1) together with (3.2) and (3.3) yield
Λ(y∗σi(aj)∗) = −Jσ i
2
(aj)Jη.
Hence, since σ i
2
(aj) = ej → 1 strongly, Λ(y∗σi(aj)∗) → −η weakly. For ω ∈ I,
j∈Jhξ(ω), Λ(y∗σi(aj)∗)i = − lim
ω(σi(aj)y) = − lim
ω(σ i
j∈J
j∈J
2
hξ(ω), ηi = − lim
(ej)y) = −ω(y).
Thus (y, η) ∈ N.
3.2. The intersection of H and M . It turns out that nϕ is the intersection of M and H.
Theorem 3.5. We have nϕ = H ∩ M, where the equality should be interpreted within R∗, see
Notation 3.1. Within R∗, x ∈ nϕ equals Λ(x) ∈ H.
(cid:3)
Moreover, let L be the closure of l∞(L) in M. Then nϕ = H ∩ L.
Proof. First we prove that nϕ = H ∩ M in R∗. For x ∈ nϕ, y ∈ R,
hΛ(x), yiR∗,R =(2.10)hΛ(x), Λ(y∗)i = ϕ(yx) =(2.6)ϕy(x) =(2.8)hx, yiR∗,R,
so Λ(x) = x in R∗. Hence the inclusion ⊆ follows.
Now let x ∈ M, ξ ∈ H be such that x = ξ in R∗. For y ∈ R,
ϕy(x) =(2.8)hx, yiR∗,R = hξ, yiR∗,R =(2.10)hξ, Λ(y∗)i.
For a ∈ T 2
ϕ , y ∈ nϕ, using Corollary 2.15 for the third, fourth and fifth equality,
hΛ(xa), Λ(y)i = ϕ(y∗xa) =a ϕ(y∗x) = ϕσi(a)(y∗x)
=ϕσi(a)y∗ (x) = hξ, Λ(yσi(a)∗)i = hJσi/2(a)∗Jξ, Λ(y)i.
ϕ , Λ(xa) = Jσi/2(a)∗Jξ. Let (ej)j∈J be a net as in Lemma A.1. Put aj = e2
So for a ∈ T 2
j .
Then xaj → x σ-weakly. Furthermore, Jπ(σi/2(aj)∗)Jξ → ξ weakly, hence Λ(xaj) converges
weakly. Since Λ is σ-weak/weak closed, x ∈ Dom(Λ) = nϕ and ξ = Λ(x). This proves ⊇.
Recall the complex interpolation method from Definition 2.4. Recall that in this section
every interpolation space should be interpreted with respect to (2.1) for parameter z = −1/2.
Note that [1, Theorem 4.2.2] gives the second equality in
(3.4)
We now prove that
H = (M, M∗)[1/2] = (L, M∗)[1/2] ⊆ L + M∗.
(3.5)
Take any s ∈ M ∩ (L + M∗) ⊆ R∗. Since s ∈ L + M∗, we can choose representatives
x ∈ L, ω ∈ M∗ such that s = x + ω in R∗. Since s ∈ M, we can find a representative
M ∩ (L + M∗) = L.
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
15
y ∈ M such that s = y in R∗. Then ω = y − x is both in M∗ and M, and hence by (2.3) in
M∗ ∩ M = L. Hence we see that s = x + ω ∈ L + L = L. This proves ⊆, the other inclusion
is trivial.
Now, (3.4) and (3.5) imply:
H ∩ M = H ∩ M ∩ (L + M∗) = H ∩ L.
(cid:3)
Again, we introduce the norm on an intersection of a compatible couple as in Section 2.1.
For x ∈ nϕ, we put
kxknϕ = max{kxk,kΛ(x)k}.
Again, we can prove a density result similar to Propostion 3.4
Proposition 3.6. The map k′ : L → nϕ : x 7→ x is injective, norm-decreasing and has dense
range.
Proof. The non-trivial part is that k′(L) is dense in nϕ with respect to k · knϕ. To prove
this, we identify nϕ with the subspace {(x, Λ(x)) x ∈ nϕ} ⊆ M × H. For (x, ξ) ∈ M × H,
we set k(x, ξ)kmax = max{kxk,kξk}. So k · kmax coincides with k · knϕ on nϕ. The dual of
(M × H,k · kmax) is given by (M ∗ × H∗,k · ksum), where k(θ, ξ)ksum = kθk + kξk.
Let (θ, ξ) ∈ M ∗ × H∗ be such that for all x ∈ L,
(3.6)
We must prove that (3.6) holds for all x ∈ nϕ. The proof proceeds in several steps.
θ(x) + hΛ(x), ξi = 0.
Claim I: There exists an ω ∈ M∗ such that for x ∈ L ∩ R, ω(x) = θ(x).
Proof of the claim. From Corollary 2.15 it follows that
L ∩ R(cid:16)= l∞(L) ∩ r∞(R)(cid:17)
is a C∗-algebra. Here and in the rest of this proof the closure has to be interpreted within
M. Let (uj)j∈J be an approximate unit for the C∗-algebra L ∩ R. We may assume that
uj ∈ (L ∩ R)+. Set ωj(x) = −hxΛ(uj), ξi, x ∈ M. So ωj ∈ M∗. Moreover, by (3.6) and
Corollary 2.15,
ωj(x) = −hxΛ(uj), ξi = θ(xuj).
Let ρ be a representation of L ∩ R on a Hilbert space Hρ such that θ(x) = hρ(x)ξ, ηi
for certain vectors ξ, η ∈ Hρ. Then ωj(x) = hρ(x)ρ(uj)ξ, ηi. Since ρ(uj) → 1 strongly,
kωjL∩R − θL∩Rk → 0. L ∩ R(⊇ T 2
ϕ ) is σ-weakly, hence strongly dense in M so that by
Kaplansky's density theorem kωjk = kωjL∩Rk. Hence (ωj)j∈J is a Cauchy net in M∗. Let
ω ∈ M∗ be its limit. This proves the first the claim.
Claim II: For x ∈ nϕ, we find ω(x) = −hΛ(x), ξi.
Proof of the claim. Note that L ∩ R is a σ-weak/weak core for Λ. Indeed, T 2
L ∩ R so that we can apply Lemma A.2.
Now, if x ∈ L ∩ R, the claim follows by the first claim and the properties of θ, i.e. ω(x) =
θ(x) = −hΛ(x), ξi. Let x ∈ nϕ and, by the previous paragraph, let (xi)i∈I be a net in L ∩ R
ϕ is contained in
16
MARTIJN CASPERS
converging σ-weakly to x such that Λ(xi) → Λ(x) weakly. Then, we arrive at the following
equation:
(3.7)
ω(x) = lim
i∈I
ω(xi) = − lim
i∈I hΛ(xi), ξi = −hΛ(x), ξi.
This proves the second claim.
Claim III: L ∩ R = nϕ ∩ n∗
ϕ = L ∩ R, where the closures are interpreted with respect to
ϕ. By Theorem 3.5, we see
the norm on M.
Proof of the claim. Note that by Proposition 2.14, L ∩ R ⊆ nϕ ∩ n∗
that nϕ ⊆ L. Since R = {x∗ x ∈ L}, see Corollary 2.15, we also have n
(closures in M).
L ∩ R ⊆ nϕ ∩ n∗
ϕ ⊆ L ∩ R
∗
ϕ ⊆ R. Hence,
The inclusions are in fact equalities. Indeed, let x ∈ L ∩ R be positive. Let xn and yn be
sequences in L, respectively R, converging in norm to x. Then, by Corollary 2.15, ynxn ∈
LR ⊆ L∩R. ynxn is norm convergent to x2. So x2 ∈ L ∩ R, hence x ∈ L ∩ R. From Corollary
2.15 it follows that L ∩ R and L ∩ R are C∗-algebras. Hence, L ∩ R = nϕ ∩ n∗
ϕ = L ∩ R.
ϕ.
Claim IV: Equation (3.6) holds for x ∈ nϕ ∩ n
∗
∗
ϕ and by the third claim, let xn ∈ L ∩ R be a sequence
Proof of the claim. Let x ∈ nϕ ∩ n
converging in norm to x. Then, using the first claim in the second equality and the third
claim in the fourth equality,
θ(x) = lim
n→∞
θ(xn) = lim
n→∞
ω(xn) = ω(x) = −hΛ(x), ξi.
∗
ϕ.
Hence (3.6) follows for x ∈ nϕ ∩ n
Proof of the proposition. Let x ∈ nϕ and let x = ux be its polar decomposition. Since (3.6)
holds for y ∈ L, we find for y ∈ L that uy ∈ L by Corollary 2.15 and,
(3.8)
(θu)(y) + hΛ(y), u∗ξi = 0,
where we defined θu ∈ M by (θu)(a) = θ(ua), a ∈ M.
If we apply the arguments in the
∗
previous paragraphs to the pair (θu, u∗ξ), we see that actually (3.8) holds for all y ∈ nϕ ∩ n
ϕ.
In particular, putting y = x, the required equation (3.6) follows.
(cid:3)
3.3. Re-iteration. Here we apply the re-iteration theorem, see [1, Theorems 4.6.1], for the
complex interpolation method to obtain Lp(M)left, p ∈ (1, 2] as an interpolation space of H
and M∗. Similarly, Lp(M)left, p ∈ [2,∞) as an interpolation space of H and M. Recall that
in this section every intersection and interpolation is understood with respect to (2.1) for the
parameter z = −1/2.
Theorem 3.7. We have the following interpolation properties:
(1) For p ∈ (1, 2], (H, M∗)[ 2
(2) For q ∈ [2,∞), (H, M)[1− 2
−1] = Lp(M)left.
q ] = (M,H)[ 2
p
q ] = Lq(M)left.
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
17
Proof. (1) Recall that L = l∞(L) denotes the closure of L in M. Recall from (2.3) that
M∗ ∩ M = L. By [1, Theorem 4.2.2 (b)] we get the first equality in of:
(3.9)
(L, M∗)[ 1
2 ] = (M, M∗)[ 1
2 ] = L2(M)left ≃ H.
The latter isomorphism is the identification in Notation 2.23. On the other hand, we find:
(3.10)
(L, M∗)[1] = (M, M∗)[1]
Since l1(L) is dense in M∗, see [9, Proposition 2.4], we find that (3.10) in turn equals M∗ by
[1, Proposition 4.2.2].
Note that the following three density assumptions are satisfied:
(i) l1(L) is dense in M∗, see [9, Proposition 2.4].
(ii) l∞(L) is dense in L (trivial).
(iii) l1(L) is k ·kI-dense in I by Proposition 3.4. Moreover I is the intersection of M∗ and H
Hence, we have checked the assumptions of the re-iteration theorem [1, Theorems 4.6.1] which
is used in the third equality,
Lp(M)left = (M, M∗)[ 1
in R∗, see Theorem 3.3.
−1],
p ] = (L, M∗)[ 1
p ] = ((L, M∗)[ 1
2 ], (L, M∗)[1])[ 2
p
−1] = (H, M∗)[ 2
p
(here the second equality follows again by [1, Theorem 4.2.2]).
(2) Completely analogously, using Theorem 3.5 and Proposition 3.6, one proves that
Lq(M)left = (H, M)[1− 2
q ],
which in turn equals (M,H)[ 2
q ] by [1, Theorem 4.2.1].
(cid:3)
4. Locally compact quantum groups
We now recall the Kustermans-Vaes definition of a locally compact quantum groups, see
[13] and [14]. Since we will be dealing with non-commutative Lp-spaces, we stick to the von
Neumann algebra setting. For an introduction to the theory of locally compact quantum
groups we refer to [12] or [18], where the results below are summarized. See also [19] were a
simple von Neumann algebraic approach to quantum groups is presented.
4.1. Von Neumann algebraic quantum groups.
Definition 4.1. A locally compact quantum group (M, ∆) consists of the following data:
(1) A von Neumann algebra M;
(2) A unital, normal ∗-homomorphism ∆ : M → M ⊗ M satisfying the coassociativity
(3) Two normal, semi-finite, faithful weights ϕ, ψ on M so that
relation (∆ ⊗ ι) ◦ ∆ = (ι ⊗ ∆) ◦ ∆, where ι : M → M is the identity;
∗ , ∀ x ∈ m
ϕ ((ω ⊗ ι)∆(x)) = ϕ(x)ω(1),
ψ ((ι ⊗ ω)∆(x)) = ψ(x)ω(1),
∗ , ∀ x ∈ m
ϕ is the left Haar weight and ψ the right Haar weight.
∀ ω ∈ M +
∀ ω ∈ M +
+
ϕ
+
ψ
(left invariance);
(right invariance).
18
MARTIJN CASPERS
Note that we suppress the Haar weights in the notation. The triple (H, π, Λ) denotes the
GNS-construction with respect to the left Haar weight ϕ. We may assume that M acts on
the GNS-space H.
following example.
In order to reflect to the classical situation of a locally compact group, we include the
Example 4.2. Let G be a locally compact group. Consider M = L∞(G) and define the
coproduct ∆G : L∞(G) → L∞(G) ⊗ L∞(G) ≃ L∞(G × G) by putting
(∆G(f ))(x, y) = f (xy).
ϕ and ψ are given by integrating against the left and right Haar weight respectively. In this
way (L∞(G), ∆G) is a locally compact quantum group.
4.2. Multiplicative unitary. There exists a unique unitary operator W ∈ B(H⊗H) defined
by
W ∗ (Λ(a) ⊗ Λ(b)) = (Λ ⊗ Λ) (∆(b)(a ⊗ 1)) ,
a, b ∈ nϕ.
W is known as the multiplicative unitary. It satisfies the pentagonal equation W12W13W23 =
W23W12 in B(H ⊗ H ⊗ H). Furthermore, ∆(x) = W ∗(1 ⊗ x)W, x ∈ M.
4.3. The dual quantum group. In [13], [14], it is proved that there exists a dual locally
∆) = (M, ∆). The dual left and right Haar
compact quantum group ( M , ∆), so that (
weight are denoted by ϕ and ψ. Similarly, all other dual objects will be denoted by a hat. By
construction,
M,
M = {(ω ⊗ ι)(W ) ω ∈ B(H)∗}
σ−strong−∗
.
Furthermore, W = ΣW ∗Σ, where Σ denotes the flip on H⊗H. This implies that W ∈ M ⊗ M
and
M = {(ι ⊗ ω)(W ) ω ∈ B(H)∗}
σ−strong−∗
.
The dual coproduct can be given by the dualized formula ∆(x) = W ∗(1 ⊗ x) W , x ∈ M . For
ω ∈ M∗, we use the standard notation
λ(ω) = (ω ⊗ ι)(W ).
Finally, we introduce the dual left Haar weight ϕ. Recall from Definition 3.2, that we let I
be the set of ω ∈ M∗, such that Λ(x) 7→ ω(x∗), x ∈ nϕ extends to a bounded functional on H.
By the Riesz theorem, for every ω ∈ I, there is a unique vector denoted by ξ(ω) ∈ H such
that ω(x∗) = hΛ(x), ξ(ω)i, x ∈ nϕ.
Definition 4.3. The dual left Haar weight ϕ is defined to be the unique normal, semi-finite,
faithful weight on M, with GNS-construction (H, ι, Λ) such that λ(I) is a σ-strong-∗/norm
core for Λ and Λ(λ(ω)) = ξ(ω), ω ∈ I.
Since we do not need it, we merely mention that there also exists a dual right Haar weight.
The following example gives the dual structure in the classical situation.
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
19
Example 4.4. Let G → B(L2(G)) : x 7→ λx be the left regular representation. For (M, ∆G)
as in Example 4.2, one finds that H ⊗ H = L2(G) ⊗ L2(G) ≃ L2(G × G) and
W f (x, y) = f (x, x−1y).
For f ∈ L1(G), let ωf be the functional on L∞(G) defined by ωf (g) =RG f (x)g(x)dlx. Then,
λ(ωf ) = (ωf ⊗ ι)(W ) =ZG
f (x)λxdlx,
where the integral is in the σ-strong-∗ topology. So λ is the left regular representation. We
find that M is given by the group von Neumann algebra M = L(G).
For completeness, we mention that ∆(λx) = λx ⊗ λx. The dual left Haar weight is given by
the Plancherel weight [15]. For a continous, compactly supported function f on G, one finds
ϕ(λ(f )) = f (e), where e is the identity of G.
If G is abelian, conjugation with the L2-Fourier transform shows that this structure is
isomorphic to (L∞( G), ∆ G).
5. Fourier theory
In this Section we define an Lp-Fourier transform. Our strategy is similar to the one defining
the classical Lp-Fourier transform on locally compact abelian groups. We first define a L1- and
L2-Fourier transform and show that they form a compatible pair of morphisms, see Remark
2.3. Then we apply the complex interpolation method to get a bounded Lp-Fourier transform
for p ∈ [1, 2],
Fp : Lp(M)left → Lq(M)left,
+
= 1.
1
p
1
q
The crucial property of the complex interpolation method is the one given by Theorem 2.5.
The theorem gives the non-commutative analogue of the Riesz-Thorin theorem as mentioned in
the introduction. It is for this reason that we have approached Lp-spaces from the perspective
of interpolation spaces and that we have used Izumi's definition.
( M, ∆) is the Pontrjagin dual.
Notation 5.1. From now on, let (M, ∆) be a locally compact quantum group with left Haar
In this section all Lp-spaces we encounter are
weight ϕ.
'left' Lp-spaces which are defined with respect to the (dual) left Haar weight. More precisely,
we stick to Notation 3.1. We equip the objects introduced in Section 2 with a hat if they
are associated with the dual quantum group. So we get L, R, Lp( M )left, . . . Recall that by
construction H = H.
Theorem 5.2 (L1- and L2-Fourier transform). We can define compatible Fourier transforms
in the following way:
(1) There exists a unique unitary map F2 : H → H, which is determined by:
(5.1)
Λ(x) 7→ Λ(λ(xϕ)),
x ∈ L.
(2) There exists a bounded map F1 : M∗ → M : ω 7→ λ(ω). Moreover, kF1k = 1.
20
MARTIJN CASPERS
(3) F2 : H → H and F1 : M∗ → M are compatible in the sense of Remark 2.3, i.e. the
following diagram commutes:
(5.2)
L
l1
?
Λ
M∗
H
(2.7)
BBBBBBBB
(2.10)
R∗
F1
F2
L
3 M
l∞
? Λ
/ H
(2.8)
AAAAAAAA
(2.10)
/ R∗.
Proof. (1) By Proposition 3.3, we see that for x ∈ L, we have xϕ ∈ I. By definition of Λ, we
have Λ(λ(xϕ)) = ξ(xϕ). Since by (2.6),
xϕ(y∗) = ϕ(y∗x) = hΛ(x), Λ(y)i,
y ∈ nϕ,
we see that by definition of ξ(xϕ), we have ξ(xϕ) = Λ(x). So (5.1) is the identity map. Since
T 2
ϕ ⊆ L and Λ(T 2
ϕ ) is dense in H, see the much stronger result of Lemma A.2, this determines
a map on H.
(2) The norm bound follows, since:
kλ(ω)k = k(ω ⊗ ι)(W )k ≤ k(ω ⊗ ι)kkWk ≤ kωk.
(3) For y ∈ R, x ∈ L, we find:
hF2Λ(x), yi R∗, R = hΛ(λ(xϕ)), yi R∗, R =(2.10)hΛ(λ(xϕ)), Λ(y∗)i
= ϕ(yλ(xϕ)) =(2.6) ϕy(λ(xϕ)) =(2.8)hλ(xϕ), yi R∗, R = hF1(xϕ), yi R∗, R,
which proves the commutativity of the diagram.
(cid:3)
Remark 5.3. Kahng [10] defines an operator algebraic Fourier transform and in principle the
idea behind Theorem 5.2 can also be found here. However, [10, Definition 3] has to be given a
more careful interpretation, since, if ϕ is not a state, the expression (ϕ⊗ι)(W (a⊗1)), a ∈ λ(I),
is in general undefined. In case ϕ is a state, our definition of F2 equals Kahng's by Remark
2.9.
We comment on the classical situation. As is shown in Example 4.4, the Fourier transform
is implicitly used to define the dual quantum group of a classical abelian group. It is for this
reason that the L2-Fourier transform F2 trivializes on the level of GNS-spaces. We work this
out in the next example.
Example 5.4. Let G be a locally compact abelian group. For f ∈ L1(G), let ωf be the the
normal functional on L∞(G) given by ωf (g) =RG f (x)g(x)dlx. Then,
f (x)λxdlx ∈ L(G),
where x 7→ λx is the left-regular representation. On the other hand, using the direct integral
decomposition L∞( G) =R ⊕
f =Z ⊕
G ZG
λ((ωf )) = (ωf ⊗ ι)(W ) =ZG
f (x)π(x)dx dπ =ZG
f (x)Z ⊕
Cdπ, we find for (1.1),
π(x)dπ dx.
G
G
+
/
/
?
/
/
4
<
?
/
/
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
21
The left regular representation x 7→ λx is unitarily equivalent toR ⊕
G πdπ, where the intertwiner
is given by the (classical) L2-Fourier transform. Since the dual quantum group associated to
a classical group is given by conjugating L∞( G) with the classical L2-Fourier transform, see
Example 4.4, the indentification of the dual GNS-space L2( G) with the GNS-space L2(G) = H
is given by applying the classical (inverse) L2-Fourier transform. Hence, using these identi-
factions, we see that the transform defined in (1.1) is the quantum group analogue of the
transform of Theorem 5.2.
Note that it is due to the identifications L2(M)left with H and L2( M)left with H = H that
the L2-Fourier transform becomes the identity map. If we had not made these identifications
the map would be less trivial. It is for this reason that we have choosen to write unitary map
in the first statement of Theorem 5.2 instead of identity map.
Note that moreover, our transform coincides with the definition given in [20, Definition 1.3].
To comment on this, suppose that (M, ∆) is compact, i.e. ϕ is a state. Let A ⊆ M be the
Hopf algebra of the underlying algebraic quantum group. We mention that A is the Hopf
algebra of matrix coefficients of irreducible, unitary corepresentations of M and refer to [18]
for more explanation. Let A be its dual, which is the space of linear functionals on A of the
form ϕ( · x), where x ∈ A, [20, Theorem 1.2]. Van Daele defines the transform by
A → A : x 7→ ϕ( · x),
x ∈ A.
Since this map is on the GNS-level given by the identity, we automatically find the following
corollary.
Corollary 5.5. We have F −1
2 = F2.
Next, we apply the complex interpolation method to define Lp-Fourier transforms.
Theorem 5.6. Let p ∈ [1, 2] and set q by 1/p + 1/q = 1. There exists a unique bounded linear
map Fp : Lp(M)left → Lq( M )left such that Fp is compatible with F1 and F2 in the sense of
On the other hand, by Remark 2.9, the normal functional xϕ, with x ∈ L is given by xϕ,
since we assumed that ϕ is a state. Here xϕ ∈ M∗ is defined by (xϕ)(y) = ϕ(yx), y ∈ M (we
use this notation to distinguish it from the algebraic map ϕ( · x)). So the Fourier transform
is defined by:
x 7→ λ(xϕ),
x ∈ M.
Since in the transition of compact algebraic quantum groups to compact von Neumann al-
gebraic quantum groups, the element ϕ( · x) ∈ A corresponds to λ(xϕ) ∈ M, this shows
the correspondence. Here, we refer to [18] for compact algebraic quantum groups and their
relations to locally compact quantum groups.
By Pontrjagin duality, one also find dual Fourier transforms. At every point in Theorem
5.2 where one of the objects L, R, M, ϕ, Λ, λ,F2,F1 appears, one should replace the object by
the same object equipped with a hat and vise versa. In that way, we get a dual L2-Fourier
transform F2 : H → H, determined by
Λ(x) 7→ Λ(λ(x ϕ)),
x ∈ L.
22
MARTIJN CASPERS
Remark 2.3, i.e. the following diagram commutes:
F1
Fp
F2
L
3 M
lq
l1
;wwwwwwwwwww
#HHHHHHHHHHH
3 H
Λ
(2.8)
$HHHHHHHHHHH
:uuuuuuuuuuu
(2.10)
Lq( M )left
/ R∗
(5.3)
L
l1
;wwwwwwwwww
#HHHHHHHHHHH
lp
/ Lp(M)left
Λ
(2.7)
$HHHHHHHHHH
:uuuuuuuuuuu
(2.10)
R∗
M∗
H
Moreover, kFpk ≤ 1.
Proof. We can apply the complex interpolation method with parameter θ = 2/p− 1 = 1− 2/q
to the pairs (H, M∗) and (H, M ) which are compatible couples as in (5.2). By Theorem 3.7
the corresponding interpolation spaces are respectively Lp(M)left and Lq( M )left.
Since by Theorem 5.6, F1 : M∗ → M and F2 : H → H are compatible, with respect to
diagram (5.2), we can use Remark 2.3 to obtain a map Fp : Lp(M)left → Lq( M )left with the
desired properties.
(cid:3)
We conclude this section by giving the Fourier transform explicitly in terms of Hilsum's
Lp-spaces. We omit the proof and merely give a few comments. The result relies on some
technicalities involving Hilsum's Lp-spaces, which was not our focus. The theorem is not
needed for the the subsequent sections.
Theorem 5.7. Let p ∈ [1, 2] and set q by 1/p + 1/q = 1. Fix a normal, semi-finite, faithful
weight φ on M ′ and φ on M ′. Set the corresponding spatial derivatives d = dϕ/dφ and
d = d ϕ/d φ. Then,
q FpΦp : Lp(φ) → Lq( φ) : [ad1/p] 7→ [λ(aϕ) d1/q],
Φ−1
a ∈ T 2
ϕ .
Note that in Theorem 5.7, we see that [ad1/p] is in Lp(φ) by Proposition 2.21. Moreover,
since aϕ ∈ I, we see that λ(aϕ) ∈ n ϕ. Therefore, [λ(aϕ) d1/q] is in Lq( φ) by Lemma 2.20. The
theorem follows by a careful analyis of (5.3) involving Proposition 2.21. The proof then relies
on the following fact. For x ∈ n ϕ, one can consider x as an element of Lq( M)left, see Section
3, and one can prove that Φqx = [x d1/q].
6. Convolution product
We define convolutions of elements in L1(M)left = M∗ with elements in Lp(M)left. We
prove that the Fourier transform transfers the convolution product into a product on the dual
quantum group.
Notation 6.1. We keep the notation as in Section 5, c.f. Notation 5.1.
Note that since Izumi's Lp-spaces are defined by means of complex interpolation, there is
a priori no multiplication on these spaces. Therefore, we extend the multiplication of M to
the Lp-setting in the following proposition. This seems to be the most natural definition of a
multiplication in the Lp-setting. The proof of the following proposition is completely similar
to the one of Theorem 5.2 (3) and 5.6.
$
+
$
#
;
/
/
/
0
8
;
/
/
#
/
:
+
:
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
23
Proposition 6.2. We extend the product of M to the Lp-setting.
(1) Let x ∈ M. The maps
m∞
x : M → M : y 7→ xy,
m1
x : M∗ → M∗ : ω 7→ xω,
are compatible in the sense of Remark 2.3, i.e. the non-dotted arrows in the following
diagram commute:
(6.1)
M∗
m1
x
3 M∗
l1
l1
L
(2.7)
:vvvvvvvvvv
$IIIIIIIIIII
:vvvvvvvvvv
$IIIIIIIIII
$IIIIIIIIIII
:tttttttttt
UU WW XX YY ZZ [[ \\ \\ ]] ^^ ^^ __ `` `` aa bb bb cc dd ee ffgg
(2) Let p ∈ (1,∞). There is a unique bounded map mp
lp
/ Lp(M)left
l∞
(2.8)
R∗
m∞
x
M
mp
x
L
l∞
lp
is compatible with m∞
commutative.
x and m1
x, i.e.
Lp(M)left
/ R∗.
(2.7)
$JJJJJJJJJJ
9tttttttttt
(2.8)
3 M
x : Lp(M)left → Lp(M)left that
the dotted arrow in (6.1) makes the diagram
Definition 6.3. Let x ∈ M and let y ∈ Lp(M)left. We will write xy for mp
x(y).
For ω1, ω2 ∈ M∗, we define the convolution product,
ω1 ∗ ω2 = (ω1 ⊗ ω2) ◦ ∆.
This product is well-known in the theory of l.c. quantum groups. We show that it is possible
to extend it to the Lp-setting for p ∈ [1, 2]. Moreover, the convolution product is turned into
the product of Definition 6.3 by the Fourier transfrom.
Theorem 6.4. Let p ∈ [1, 2] and set q ∈ [2,∞] by 1/p + 1/q = 1.
(1) Let x ∈ L and let ω ∈ M∗. Then,
and
ω ∗ (xϕ) ∈ I
ξ(ω ∗ (xϕ)) = λ(ω)Λ(x).
(2) Let ω ∈ M∗. We denote ω∗2 for the bounded operator λ(ω) : H → H. Furthermore,
we define ω∗1 : M∗ → M∗ : θ 7→ ω ∗ θ. Then, ω∗1 and ω∗2 are compatible, i.e. the
non-dotted arrows in following diagram commute:
(6.2)
L
M∗
ω∗1
3 M∗
l1
:vvvvvvvvvv
$IIIIIIIIIII
lp
/ Lp(M)left
Λ
l1
R∗
(2.7)
:vvvvvvvvvv
$IIIIIIIIII
$IIIIIIIIIII
:ttttttttttt
UU WW XX YY ZZ [[ \\ \\ ]] ^^ ^^ __ `` `` aa bb bb cc dd ee ffgg
(2.10)
3 H
ω∗2
ω∗p
L
lp
Λ
H
(2.7)
$JJJJJJJJJJ
9ttttttttttt
(2.10)
Lp(M)left
/ R∗.
(3) There is a unique bounded operator ω∗p : Lp(M)left → Lp(M)left that is compatible with
ω∗1 and ω∗2, i.e. (6.2) commutes.
$
+
$
$
:
/
/
/
0
8
:
/
/
$
/
:
+
9
$
+
$
$
:
/
/
/
0
8
:
/
/
$
/
:
+
9
24
MARTIJN CASPERS
(4) For ω ∈ M∗, a ∈ Lp(M)left,
where the left hand side uses Definition 6.3 for Lq( M)left.
F1(ω)Fp(a) = Fp(ω ∗p a),
Proof. Let θ ∈ I and put y = λ(θ) = (ι ⊗ θ)(W ∗). Now, (2) follows from,
(ω ∗ (xϕ))(y∗) = (ω ⊗ (xϕ))∆((ι ⊗ θ)(W ∗)∗) = (ω ⊗ (xϕ) ⊗ θ)(W13W23)
=θ(( (ω ⊗ ι)(W )((xϕ) ⊗ ι)(W ) )∗) = hΛ ((ω ⊗ ι)(W )((xϕ) ⊗ ι)(W )) , ξ(θ)i
=h(ω ⊗ ι)(W )ξ(xϕ), Λ(y)i = hλ(ω)Λ(x), Λ(y)i,
and the fact that {Λ((ι ⊗ θ)(W ∗)) θ ∈ I} is dense in H.
Theorem 3.7 to (2). (4) For ω1, ω2 ∈ M∗, note that
F1(ω1 ∗ ω2) = (ω1 ⊗ ω2 ⊗ ι)(∆ ⊗ ι)(W ) = (ω1 ⊗ ω2 ⊗ ι)W13W23 = (ω1 ⊗ ι)(W )(ω2 ⊗ ι)(W ).
For x ∈ L, ω ∈ M∗,
The compatibility in (2) follows directly from (1) using Theorem 3.3. (3) follows by applying
Fp(ω ∗p lp(x)) = F1(ω ∗ (xϕ)) = F1(ω)F1(xϕ) = F1(ω)Fp(lp(x)).
Here, the first and last equality follows from commutativity of (5.3), (6.1) and (6.2). Since
the range of lp is dense in Lp(M)left, see Lemma 2.7, (4) follows.
(cid:3)
7. A distinguished choice for the interpolation parameter
Recall that in Sections 3 to 6 we considered the compatible couple (M, M∗) for the inter-
polation parameter z = −1/2, see Definition 2.8. For this parameter one is able to define a
Lp-Fourier transform. In this section we show that the real part of the parameter is distin-
guished. More precisely, we investigate the example of (M, ∆) = SUq(2) and show that given
the fact that
(7.1)
is the L1-Fourier transform the only interpolation parameters z that allows a passage to a
Lp-Fourier transform are z = −1/2 + it, where t ∈ R.
F1 : M∗ → M : ω 7→ (ω ⊗ ι)(W ),
The importance of this result is strengthened by the final remark of [4]. For classical,
locally compact groups there is an approximation property called Reiter's property (Pp), where
p ∈ [1,∞). The definition assumes the existence of a net of functions in Lp(G) satisfying the
approximation axiom of [4, Definition 1.2]. Daws and Runde show that (P1) and (P2) can be
defined for quantum groups as well and they use them to study (co-)amenability properties
of quantum groups.
In the final remark of [4], Daws and Runde mention that it remains to be seen if there is
a property (Pp) for any p ∈ [1,∞). In particular, they mention that it remains unclear how
the Lp-space associated with a quantum group should be turned into a L1-module. In [5] this
is done using Izumi's Lp-spaces for the complex interpolation parameter z = −1/2, whereas
in [3] a similar, but not identical construction was used for the parameter z = 0. We believe
that the Fourier transform indicates that the most natural choice would be z = −1/2.
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
25
From now on we let (M, ∆) be the quantum group SUq(2), see [21], [22], [11]. See also [12]
for a concise introduction. We recall its most important properties.
We set H = L2(N) ⊗ L2(T). Let (ei)i∈N be the canonical orthonormal basis of L2(N) and
let (fk)k∈Z be the canonical orthonormal basis for L2(Z) (so fk = ζ k, where ζ is the identity
function on the complex unit circle T). Define operators α, γ given by:
(7.2)
Then, M is the von Neumann algebra generated by α and γ, i.e.
α ei ⊗ fk =p1 − q2iei−1 ⊗ fk,
γ ei ⊗ fk = qiei ⊗ fk+1.
M = B(L2(N)) ⊗ L∞(T) ≃ L∞(T, B(L2(N))).
For x = x(t) ∈ L∞(T, B(L2(N))), both the left and right Haar weight are given by the state
ϕ(x) =
(1 − q2)
2π
ZT
∞
Xi=0
q2ihx(t)ei, eiidt.
Next, we need Peter-Weyl theory for SUq(2). Recall [11] that for every l ∈ 1
N, there exists
a unique irreducible corepresentation t(l) ∈ M ⊗M2l+1(C). In fact, these are all the irreducible
corepresentations of (M, ∆) and we have a Peter-Weyl decomposition
2
W ≃ Ml∈ 1
2
N
t(l) ⊗ 12l+1
( ∈ M ⊗Ml∈N
M2l+1(C) ⊗ M2l+1(C)
).
So every corepresentation t(l) appears 2l + 1 times in the multiplicative unitary.
−l, g(l)
Let g(l)
−l+1, . . . , g(l)
matrix elements (ι ⊗ ωg(l)
operator Q(l) such that we have orthogonality relations between matrix coefficients.
l denote the standard basis vectors of C2l+1. Let us denote t(l)
i,j for the
N, there exists a unique strictly positive
)(t(l)). For every l ∈ 1
j ,g(l)
2
i
(7.3)
In fact, with respect to the basis g(l)
i
⊕l∈ 1
(7.4)
NM2l+1(C). We put Q = ⊕l∈ 1
2
2
ϕ((t(l)
i,j)∗t(l′)
i′,j ′) = δl,l′δj,j ′hQ(l)g(l)
, the matrix Q(l) is diagonal.
, g(l)
i′ i
i
NQ(l), so that Q is affiliated with M . Moreover,
It follows that M ≃
σt(x) = Q−itxQit,
x ∈ M .
Finally, the following two matrix coefficients will play an essential role in the proof of the
main theorem of this section. It follows from [11, Chapter 4] that for n ∈ N,
t(n/2)
n/2,n/2 = αn,
t(n/2)
−n/2,−n/2 = (α∗)n.
Recall the notational conventions from Section 2.
Theorem 7.1. Consider (M, ∆) = SUq(2) and let z, z′ ∈ C. Let F1 : M∗ → M be defined as
(z ′)( M) making the following
in (7.1). Suppose that there is bounded map F2 : L2
(z)(M) → L2
26
MARTIJN CASPERS
diagram commutative
(7.5)
L(z)
M∗
i1
(z)
;wwwwwwwwww
i2
(z)
(−z))∗
(i∞
$IIIIIIIIII
(z)(M) ⊆
L2
L∗
(−z)
F1
F2
L(z ′)
3 M
i∞
(z′)
:vvvvvvvvvv i2
(z′)
)∗
(−z′)
(i1
%JJJJJJJJJJ
⊆
/ L2
(z ′)( M )
/ L∗
(−z ′).
Then, z = −1/2 + it for some t ∈ R.
Proof. We will prove that F2 is unbounded unless z = −1/2 + it for some t ∈ R. We need
three preparations.
Firstly, the modular automorphism group of ϕ is given by σt(x) = (γγ∗)itx(γγ∗)−it. Hence,
it follows from (7.2) that α ∈ Tϕ and
σz(α) = q−2izα,
σz(α∗) = q2izα∗.
Secondly, for a, b ∈ Tϕ, x ∈ T 2
ϕ , z ∈ C, we find
x (a∗b) =hxJ∇¯zΛ(a), J∇−zΛ(b)i = h∇z+ 1
ϕ(z)
2 x∇−z− 1
2∇J∇1/2Λ(a), J∇1/2Λ(b)i
2 )(x)σ−i(a∗)) = ϕ(a∗bσ−i(z+ 1
2 )(x)).
=hσ−i(z+ 1
So we conclude that ϕ(z)
2 )(x)Λ(σ−i(a∗)), Λ(b∗)i = ϕ(bσ−i(z+ 1
x = σ−i(z+ 1
Thirdly, we identify M with ⊕l∈ 1
NM2l+1(C). Let e(n/2)
2 )(x)ϕ.
2
−n/2,−n/2 ) be the element
of M, with matrix elements equal to zero everywhere, except for the summand with index
n/2, where it has a 1 on the upper left (respectively lower right) corner. Using the Peter-Weyl
orthogonality relations, we see that for every n ∈ N:
(ϕ ⊗ ι)(t(l)(t(n/2)∗
−n/2,−n/2 ⊗ 1)) =ϕ((α∗)nαn)e(n/2)
−n/2,−n/2,
n/2,n/2 (and e(n/2)
(7.6)
(ϕ ⊗ ι)(W (αn ⊗ 1)) = Ml∈ 1
(ϕ ⊗ ι)(W ((α∗)n ⊗ 1)) =Ml∈N
N
2
(ϕ ⊗ ι)(t(l)(t(n/2)∗
n/2,n/2 ⊗ 1)) =ϕ(αn(α∗)n)e(n/2)
n/2,n/2,
ϕ((α∗)nαn) =
ϕ(αn(α∗)n) =
(1 − q2)q2n
1 − q2n+2 ,
(1 − q2)
1 − q2n+2 .
By (7.4) and the fact that Q is diagonal we see that (7.6) are in T 2
(7.7)
ϕ and
n/2,n/2) = e(n/2)
−n/2,−n/2) = e(n/2)
−n/2,−n/2,
σz(e(n/2)
σz(e(n/2)
n/2,n/2.
Now we prove that F2 must be unbounded by proving that the map U(z ′)F2U ∗
unbounded. Recall that U(z) was defined in Proposition 2.21.
(z)(αn) = (i1
(z)Λ(σ−i(z+1/2)/2(αn)) = F2i2
F2U ∗
(−z ′))∗F1i1
(z)(αn)
(−z ′))∗(ϕ(z)
=(i1
=q−2n(z+1/2)(i1
αn ⊗ ι)(W ) = (i1
(−z ′))∗(ϕ ⊗ ι)(W (αn ⊗ 1)).
(−z ′))∗(ϕ ⊗ ι)(W (σ−i(z+1/2)(αn) ⊗ 1))
(z) : H → H is
$
+
%
/
/
;
/
/
/
7
:
/
/
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
27
Since (ϕ ⊗ ι)(W (αn ⊗ 1)) ∈ T 2
ϕ , we see that by commutativity of the right triangle in (7.5),
F2U ∗
(z)Λ(σ−i(z+1/2)/2(αn)) = q−2n(z+1/2)(i2
(z ′))(ϕ ⊗ ι)(W (αn ⊗ 1)).
Hence,
U(z ′)F2U ∗
(z)Λ(σ−i(z+1/2)/2(αn)) = q−2n(z+1/2) Λ(σ−i(z ′+1/2)/2(ϕ ⊗ ι)(W (αn ⊗ 1)))
=(7.6),(7.7)q−2n(z+1/2) Λ((ϕ ⊗ ι)(W (αn ⊗ 1))) = q−2n(z+1/2)ξ(αnϕ) = q−2n(z+1/2)Λ(αn).
Hence,
U(z ′)F2U ∗
(z) : Λ(αn) = qn(z+1/2)Λ(σ−i(z+1/2)/2(αn)) 7→ q−n(z+1/2)Λ(αn).
which is unbounded in case Re(z) > −1/2.
By a similar computation, one finds that
In this case we see that U(z ′)F2U ∗
U(z ′)F2U ∗
(z) : Λ((α∗)n) 7→ qn(z+1/2)Λ((α∗)n).
(z) is unbounded for Re(z) < −1/2.
(cid:3)
Remark 7.2. By Pontrjagin duality also the dual statement holds. So in order to get a proper
Fourier theory on quantum groups with Fourier transforms and inverse Fourier transforms,
one is obliged to take the interpolation parameter on both M and M to be −1/2.
Appendix A.
We have not found an explicit proof of the following lemma's in the literature. For com-
pleteness, we prove them here. Here M is a von Neumann algebra with normal, semi-finite,
faithful weight ϕ.
The following lemma is a variant of [17, Lemma 9].
Lemma A.1. Let δ > 0. There exists a net (ej)j∈J in Tϕ such that (1) kσz(ej)k ≤ eδIm(z)2,
(2) ej → 1 strongly and (3) σi/2(ej) → 1 σ-weakly.
Proof. Let (fj)j∈J and (ej)j∈J be nets as in [17, Lemma 9]. This lemma proves already that
(ej)j∈J satisfies (1) and (2). Now, (3) follows, since for ξ ∈ H,
σt(fj)dt! = r δ
πZ ∞
(ωξ,ξ ◦ σt)dt! (1) =r δ
πZ ∞
(ej)ξ, ξi =ωξ,ξ r δ
π Z ∞
→ r δ
πZ ∞
(ωξ,ξ ◦ σt)dt! (fj)
dthξ, ξi = hξ, ξi,
e−δ(t− i
e−δ(t− i
e−δ(t− i
e−δ(t− i
2 )2
hσ i
2
2 )2
2 )2
2 )2
−∞
−∞
−∞
−∞
where the last equality is obtained by means of the residue formula for meromorphic functions.
So σϕ
(cid:3)
(ek) is bounded and converges weakly, hence σ-weakly to 1.
i
2
ϕ is a σ-strong-∗/norm core for Λ.
Lemma A.2. T 2
Proof. It is enough to prove that T 2
ϕ is a σ-weak/weak core for Λ, since the σ-weak/weak
continuous functionals on the graph of Λ equal the σ-strong-∗/norm continuous functionals.
28
MARTIJN CASPERS
It is not too hard to show that nϕ ∩ n
Put
∗
ϕ is a σ-weak/weak core for Λ. Now, let x ∈ nϕ ∩ n
ϕ.
√π Z ∞
n
σt(x)dt,
e−(nt)2
−∞
∗
xn =
n
Λ(xn) =
where the integral is taken in the σ-strong-∗ sense. By standard techniques (c.f. the proof of
[17, Lemma 9]), xn ∈ Tϕ and xn converges σ-weakly to x. Moreover, using the the fact that
Λ is σ-strong-∗/norm closed, we obtain
√πZ ∞
∇itΛ(x)dt → Λ(x)
where the integral is a Bochner integral, c.f. [15, Chapter VI, Lemma 2.4]. Hence Tϕ is a core
for Λ. Now, let x ∈ Tϕ and let (ej)j∈J be a net in Tϕ such that ej → 1 σ-weakly, c.f. Lemma
A.1. Then, ejx ∈ T 2
Acknowledgement. The author thanks Erik Koelink for useful discussions and also many
thanks to the referee for his comments and suggestions.
ϕ and ejx → x σ-weakly and Λ(ejx) = ejΛ(x) → Λ(x) weakly.
e−(nt)2
weakly,
−∞
(cid:3)
References
[1] J. Bergh, J. Lofstrom, Interpolation spaces, Springer 1976.
[2] A. Connes, On the spatial theory of von Neumann algebras., J. Funct. Anal. 35 (1980), 153 -- 164.
[3] M. Daws, Representing multipliers of the Fourier algebra on non-commutative Lp spaces, Canadian J.
Math. 63 (2011), 798 -- 825.
[4] M. Daws, V. Runde, Reiter's properties (P1) and (P2) for locally compact quantum groups. Math. Anal.
Appl. 2 (2010), 352 -- 365.
[5] B. Forrest, H.H. Lee, E. Samei, Projectivity of modules over Fourier algebras, To appear in Proc. London
Math. Soc.
[6] L. Grafakos, Classical and modern Fourier analysis, Pearson Education 2004.
[7] U. Haagerup, Lp-spaces associated with an arbitrary von Neumann algebra, Alg`ebres d'op´erateurs et leurs
applications en physique math´ematique, Proc. Colloq., Marseille 1977, 175 -- 184.
[8] M. Hilsum, Les espaces Lp d'une alg`ebre de von Neumann d´efinies par la deriv´ee spatiale, J. Funct. Anal.
40 (1981), 151 -- 169.
[9] H. Izumi, Constructions of non-commutative Lp-spaces with a complex parameter arising from modular
actions, Internat. J. Math. 8 (1997), 1029 -- 1066.
[10] B.J. Kahng, Fourier transform on locally compact quantum groups, J. Operator Theory 64 (2010), 69 -- 87.
[11] A. Klimyk, K. Schmudgen, Quantum groups and their representations, Springer 1997.
[12] J. Kustermans, Locally compact quantum groups, LNM 1865, Springer 2005, 99 -- 180.
[13] J. Kustermans, S. Vaes, Locally compact quantum groups, Ann. Scient. ´Ec. Norm. Sup. 33 (2000), 837 -- 934.
[14] J. Kustermans, S. Vaes, Locally compact quantum groups in the von Neumann algebraic setting, Math.
Scand. 92 (2003), 68 -- 92.
[15] M. Takesaki, Theory of operator algebras II, Springer 2000.
[16] M. Terp, Lp spaces associated with von Neumann algebras. Notes, Report No. 3a + 3b, Københavns
Universitets Matematiske Institut, Juni 1981.
[17] M. Terp, Interpolation spaces between a von Neumann algebra and its predual, J. Operator Theory 8
(1982), 327 -- 360.
[18] T. Timmermann, An invitation to quantum groups and duality, EMS 2008.
[19] A. Van Daele,
quantum groups. A von Neumann
compact
Locally
algebra
approach,
http://arxiv.org/abs/math/0602212 (2006).
[20] A. Van Daele, The Fourier transform in quantum group theory, New techniques in Hopf algebras and
graded ring theory, K. Vlaam. Acad. Belgie Wet. Kunsten (KVAB), Brussels 2007, 187 -- 196.
THE Lp-FOURIER TRANSFORM ON LOCALLY COMPACT QUANTUM GROUPS
29
[21] S.L. Woronowicz, Compact matrix pseudogroups, Commun. Math. Phys. 111 (1987), 613 -- 665.
[22] S.L. Woronowicz, Twisted SU (2) group. An example of a non-commutative differential calculus, Publ.
RIMS, Kyoto University, 23 (1987), 117 -- 181.
Radboud Universiteit Nijmegen, IMAPP, FNWI, Heyendaalseweg 135, 6525 AJ Nijmegen,
the Netherlands
E-mail address: [email protected]
|
1706.00515 | 1 | 1706 | 2017-06-01T22:44:39 | Scale invariant transfer matrices and Hamiltionians | [
"math.OA",
"cond-mat.stat-mech",
"math-ph",
"math.DS",
"math-ph",
"math.QA"
] | Given a direct system of Hilbert spaces $s\mapsto \mathcal H_s$ (with isometric inclusion maps $\iota_s^t:\mathcal H_s\rightarrow \mathcal H_t$ for $s\leq t$) corresponding to quantum systems on scales $s$, we define notions of scale invariant and weakly scale invariant operators. Is some cases of quantum spin chains we find conditions for transfer matrices and nearest neighbour Hamiltonians to be scale invariant or weakly so. Scale invariance forces spatial inhomogeneity of the spectral parameter. But weakly scale invariant transfer matrices may be spatially homogeneous in which case the change of spectral parameter from one scale to another is governed by a classical dynamical system exhibiting fractal behaviour. | math.OA | math |
SCALE INVARIANT TRANSFER MATRICES AND
HAMILTIONIANS.
VAUGHAN F. R. JONES
Abstract. Given a direct system of Hilbert spaces s (cid:55)→ (cid:72)s (with isometric
s : (cid:72)s → (cid:72)t for s ≤ t) corresponding to quantum systems on
inclusion maps ιt
scales s, we define notions of scale invariant and weakly scale invariant operators.
Is some cases of quantum spin chains we find conditions for transfer matrices
and nearest neighbour Hamiltonians to be scale invariant or weakly so. Scale
invariance forces spatial inhomogeneity of the spectral parameter. But weakly
scale invariant transfer matrices may be spatially homogeneous in which case the
change of spectral parameter from one scale to another is governed by a classical
dynamical system exhibiting fractal behaviour.
1. Introduction
According to dogma, critical phenomena in physics are accompanied by scale
invariance-patterns repeat on all scales- and attendant long-range interactions. In
this paper we explore states and observables of a quantum spin chain that exhibit
very strict forms of scale invariance, without passing to a continuum limit. The un-
derlying philosophy is that elements of the Thompson groups ([6]) express local scale
transformations on a lattice, which must be supposed infinite for the transformation
to exist mathematically. The Thompson group is a replacement for the diffeomor-
phism group Diff(S1) (Virasoro algebra) whose presence in the continuum limit is a
consequence of local scale invariance at criticality of a 2-dimensional system. Thus
one the details of the Thompson group representations occurring in various models
should supply both qualitative and quantitative information about physics on the
lattice at a critical point. By [2] we do not expect critical behaviour at non-zero
temperature so the best place to look for Thompson group symmetry is at a quan-
tum phase transition where changing some physical parameter besides temperature
causes an abrupt change of behaviour.
Indeed we will observe three distinct types of behaviour manifested in the asymp-
totics of the correlation of states with themselves under lattice rotation/translation
by one lattice spacing, as the lattice tends to infinity. This correlation can be as
simple as an alternation between two values, but most often it tends rapidly to zero.
In this case it is possible to rescale the correlation so that it has limits which exist
as sesquilinear forms on the pre-Hilbert space of states. In the model we investigate
V.J. is supported by the NSF under Grant No. DMS-1362138 and grant DP140100732, Symme-
tries of subfactors.
1
2
VAUGHAN F. R. JONES
there are two sesquilinear forms S1 and S2 and the rescaled correlation tends to
one or the other according to parity. But as the quantum parameter in the model
increases, S1 and S2 coalesce at a certain critical value after which the convergence
is to the common sesquilinear.
Our approach is wide open to criticism. The states of our "infinite tensor prod-
uct" ([29]) have a built in long range correlation which forces the impossibility of a
continuum limit. This has already been observed by others and Evenbly and Vidal
in [9] proposed their MERA precisely to overcome this problem. But here we are no
longer trying to construct a continuum limit.
A perhaps more serious criticism is that the model we present uses exclusively
spin-doubling renormalisation which does not appear particularly physical. To obtain
model independent results we should investigate many different models to see if there
are phenomena common to them all. Or introduce bigger groups than the Thompson
groups which allow more general local scale invariance.
There have been interesting mathematical developments coming out of this progam-
see [18], but the physical relevance of our approach will ultimately be decided by the
existence or otherwise of states with scaling properties in actual physical systems.
The spin-doubling operators are no more complicated than some of the "gates" in the
world of quantum computing ([25]) so one could in principle prepare scale invariant
states with a machine. But the number of gates required would be rather large.
It is interesting that the calculation of correlation asymptotics becomes the it-
eration of a purely classical dynamical system that may be as simple as a rational
function on CP 1. Fixed points, periodic points and their stability properties thus
become the critical values for the quantum system.
Let us give a more precise account of our results. In [17] we proposed the construc-
tion of a Hilbert space for the continuum limit of a quantum spin chain by reversing
the process of block spin renormalisation. We thus obtained Hilbert spaces (cid:72)n for a
chain of 2n spins with each (cid:72)n embedded isometrically in (cid:72)n+1 by replacing each spin
so that
by 2 copies of itself. The spin doubling isometry is symbolically denoted
the isometry (cid:72)n (cid:44)→ (cid:72)n+1 is represented by
. We called the
inductive limit pre-Hilbert space (cid:72) = lim→ (cid:72)n the semicontinuous limit. The Thomp-
son groups F (for an open spin chain) and T (for periodic boundary conditions) act
on (cid:72) by unitary transformations which implement local scale transformations.
The result of [17] showed that the continuum limit does not naturally live on the
semicontinuous limit or its completion. We calculated that for an SO(3) invariant
2n are hopelessly discontinuous as n → ∞ for the topology
spin 1 chain rotations by 1
induced on rotations by the circle. (The advantage of this particular spin chain is
that
is unique up to an irrelevant scalar, though in unpublished calculations we
have shown the same discontinuity for a family of spin-tripling systems where the
spin-tripling operator is not at all unique.)
···
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
3
In the algebraic Bethe ansatz or quantum inverse scaterring method (QISM, [12,
10]) one starts with a transfer matrix T (λ) depending on a spectral parameter λ
then obtains a nearest neighbour Hamiltonian as the logarithmic derivative of T (λ)
at some value of λ , and other Hamiltonians, and conserved quantities, by further
manipulation of T (λ).
The calculation of [17] did display a feature common, albeit in a topsey-turvey
way, to the semicontinuous limit and the QISM. Namely the infinitesimal behaviour
of space translation is determined by a transfer matrix with spectral parameter. (In
[17] this was only shown for rotations by 1
2n but for general rotations there is a more
complicated way to manipulate the transfer matrix.) In the QISM the infinitesimal
time translation (given by a local Hamiltonian) is also given by a transfer matrix with
spectral parameter. These considerations have led us to treat the transfer matrix
itself as being of fundamental physical significance, being the generator of space
translation on the one hand and of many Hamiltonians and constants of the motion
on the other. It thus becomes attractive to look for transfer matrices that are defined
on the semicontinuous limit. For this we will introduce two notions of scale invariant
operators on (cid:72), the first kind being operators on the (cid:72)n which commute with the
inclusion maps ιn+1
and the second kind, the weakly scale invariant operators, which
in the sense of their matrix coefficients. It was the weakly
commute with the ιn+1
scale invariant operators that arose in the calculation of [17].
In a simple model
coming from the Temperley-Lieb algebra [28], we show that scale invariant transfer
matrices exist in both senses, and that scale invariant nearest neighbour Hamiltonians
exist in the sense of their matrix coefficients.
We should perhaps end by saying that this work began as an attempt to construct
chiral conformal field theory on a circle directly from a subfactor ([13]), and thus
hopefully extending the correspondence begun in [30] to include "exotic" subfactors
as in [3], [8],[19]. This has not worked but the intriguing question arises from this
paper as to whether these exotic subfactors have attendant solutions of our ABC
equation of this paper.
n
n
2. Scale invariant transfer matrices.
2.1. Definition. Suppose we are given a directed set (D,≤) (thought of as defining
various scales of quantum systems) and a direct system A of Hilbert spaces s (cid:55)→
(cid:72)s = A(s) for s ∈ D with corresponding isometric inclusions ιt
s : (cid:72)s → (cid:72)t for s ≤ t
satisfying the usual direct system conditions -[17]. We let (cid:72) = (cid:72)A be the direct
limit of the A(s). By definition (cid:72) is the set of ordered pairs (d, ξ) with ξ ∈ (cid:72)d,
modulo the equivalence relation (d, ξ) ∼ (e, η) iff there is an f with f ≥ e and f ≥ d
with ιf
e (η). Since the ι's are linear isometries, (cid:72) inherits the structure of a
pre-Hilbert space in the obvious way. Each (cid:72)s will be identifed with a subspace of
(cid:72).
d(ξ) = ιf
Definition 2.1. With notation as above,
4
VAUGHAN F. R. JONES
(1) a scale invariant operator on A will be a family Ts : (cid:72)s → (cid:72)s of linear maps
(2) a weakly scale invariant operator on A will be a family Ts : (cid:72)s → (cid:72)s of linear
Tt ◦ ιt
s = ιt
s ◦ Ts
such that
maps such that
(cid:104)Tt(ιt
sξ), ιt
sη(cid:105) = (cid:104)Ts(ξ), η(cid:105) for all ξ, η ∈ (cid:72)s.
Remark 2.2. Scale invariance implies weak scale invariance but not the other way
round since the weak condition does not force Tt to preserve (cid:72)s.
Proposition 2.3. (1)A scale invariant operator Ts determines, and is determined
by, an operator T : (cid:72) → (cid:72) satisfying T(cid:72)s = Ts for all s.
sense of [27]) form [, ]on (cid:72) by
(2) A weakly scale invariant operator determines a sesquilinear ("quadratic" in the
[ξ, η] = (cid:104)Ts(ξ), η(cid:105) for any s with ξ, η ∈ (cid:72)s
(cid:3)
Proof. These follow immediately from the definition.
Remark 2.4. If n (cid:55)→ s(n) is a cofinal sequence in D then we may choose an or-
thonormal basis ξi of (cid:72) with ξ1, ξ2,··· , ξdim(cid:72)s(n) being a basis of (cid:72)s(n) for all n. The
square matrix [ξi, ξj] for 1 ≤ i, j ≤ dim(cid:72)s(n) is the matrix of Ts(n) for this basis.
The form [, ] may not define an operator on (cid:72) since there is no a priori control
of the size of the matrix entries.
2.2. Examples from spin chains. We will adopt the "direct limit over trees" ap-
proach of [17] to the semicontinuous limit. This allows us to double a single spin at
a time. Thus we consider the directed set T of planar binary rooted trees with s ≤ t
iff s is a rooted subtree of t.
Definition 2.5. For each n let (cid:84)n be the tree with 2n leaves all at the same distance
from the root.
Remark 2.6. The (cid:84)n form a cofinal sequence in T.
If (cid:104) is a (usually finite dimensional) Hilbert space of spin states for a single spin we
begin with a "spin-doubling" operator Y : (cid:104) → (cid:104) ⊗ (cid:104). We suppose it is an isometry,
i.e. Y ∗Y = id.
In Penrose tensor notation this condition can be drawn as
= where
stands for the tensor Y and when it is upside down it represents its adjoint. Such a
diagram is read from bottom to top, starting with a vector ξ ∈ (cid:104) which is sent by Y
to (cid:104) ⊗ (cid:104) then back to (cid:104) by Y ∗. This kind of notation is very common,see e.g. [7, 26],
and is known as "tensor networks", and we will generalise it to planar algebras later
on.
Now form the direct system AY on T where if t has k leaves, AY (t) = ⊗k(cid:104). To
s note that any ≤ in T decomposes into a sequence of ≤'s where
define the maps ιt
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
5
one leaf is added a time. So it suffices to define ιi : (cid:72)s → (cid:72)t where t is obtained
from s by doubling the ith leaf (from the left). We set
ιi(η1 ⊗ η2 ⊗ η3 ··· ⊗ ηi ⊗ ··· ηk) = η1 ⊗ η2 ⊗ η3 ··· ⊗ Y (ηi) ⊗ ··· ⊗ ηk
or, in tensor network language: ιi =
We leave it to the reader to check that these elementary ι's consistently define a
direct system.
Note that the (cid:72)(cid:84)n are the Hilbert spaces of quantum spin chains with 2n spins
each having Hilbert space (cid:104). They are embedded one into the next by doubling all
the spins with Y .
.
The concept of transfer matrix for a spin chain is well known: if we are given a
tensor L in ⊗4(cid:104) we denote it by
.
Remark 2.7. The placement of the L indicates that the indices for the tensor L
should begin on the string immediately following L in clockwise order, and continue
in clockwise order.
A transfer matrix is then an operator of the form
T (L1, L2,··· , Lk) =
for some choices Li. We need to do something about the horizontal boundary to
make this picture an operator from ⊗k(cid:104) to itself. Let us assume periodic boundary
conditions, i.e. we identify the first and last horizontal edges in the picture which may
then be thought of as living in an annulus. We want to find when T (L1, L2,··· , Lk)
defines a scale invariant operator. For physics we only need to define it on the (cid:72)(cid:84)n
and we could then deduce its values on all the (cid:72)s by restriction. But it will be just
as easy define transfer matrices on each (cid:72)t.
To proceed we introduce an equation which we call the ABC equation, which
allows us to extend transfer matrices when a single spin is doubled.
Definition 2.8. The ABC equation is the following equation in ⊗5(cid:104).
(1)
D(A, B) = Y (C)
where A, B and C are unknown tensors in ⊗4(cid:104) and
iLLL1L 2L3L45Lk 6
VAUGHAN F. R. JONES
D(A, B) =
Y (C) =
.
Proposition 2.9. Suppose (Ai, Bi, Ci) satisfy the ABC equation for i = 1, 2,··· , k.
Then if T = T (A1, B1, A2, B2,··· , Ak, Bk),
T (cid:22)⊗k(cid:104)= T (C1, C2,··· Ck)
Proof. Here ⊗k(cid:72) is embedded in ⊗k+1(cid:72) by applying Y to each tensor product
component. So the proof is simply a matter of applying the ABC equation at every
(cid:3)
trivalent vertex in the diagram for this embedding.
Given C, the ABC equation may or may not have solutions for A and B and if
it does it may have many. So in order to define a scale invariant transfer matrix we
need to choose solutions if possible. Given such a choice it is easy to define Ts for
any s ∈ T provided we set up a little notation.
Definition 2.10. For each tree t ∈ T and leaf l of t, let w(l) be the word on {0, 1}
read from the path on t up from root to leaf, with 0 on left turns at trivalent vertices
and 1 at right turns.
Definition 2.11. For any function w (cid:55)→ Lw ∈ ⊗4(cid:104) from all words w on {0, 1}
and t ∈ T with k leaves, define T L
to be the transfer matrix (on (cid:72)t) with periodic
boundary conditions:
t
T L
t =
where µi = Lwi, where the leaves of t are numbered from left to right and wi is the
word coding for the ith. leaf.
In the special case t = (cid:84)n where all the branches of t have the same length (so t
t . Thus
has 2k leaves for some k), and Lw takes the same value L, we use TL for T L
for k = 3, TL =
.
Definition 2.12. A coherent choice Lw of tensors, for all words w on {0, 1} will be
one such that for each w ,
i.e. putting A = LLw0, B = Lw1, and C = Lw gives a solution of the ABC equation.
D(Lw0, Lw1) = Y (Lw)
ABCkµµµµµ1µ3 245LLLLLLLL t
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
7
Proposition 2.13. Suppose w (cid:55)→ Lw is a coherent choice of tensors, then t (cid:55)→ T L
defines a scale invariant operator on the direct system AY .
Proof. Just apply the ABC equation every time t differs from s by doubling a single
vertex. The formulas defining w and Lw take care of the book-keeping for the values
(cid:3)
of the spectral parameter.
Thus in a particular model, to exhibit scale invariant transfer matrices it will
suffice to exhibit coherent choices of tensors.
Remark 2.14. In fact, because we are dealing with the direct limit, a coherent choice
of tensors does not have to be defined for all words. Given t ∈ T one can choose
a solution to the ABC equation for every leaf of t, form the corresponding transfer
matrix on (cid:72)t and extend it using coherent choices for each leaf of t. This will define
an operator on (cid:72) which should be considered scale invariant. This requires a slight
but obvious modificaiton of definition 2.1 which we have not given to avoid confusion.
The restriction of this operator to (cid:72)s for trees not containing t will not in general
preserve the subspace (cid:72)s so is not a transfer matrix on (cid:72)s.
For the convenience of the reader we exhibit a pair (t,⊗7(cid:104)) whose tree has 7 leaves,
and the corresponding transfer matrix for some choice of L's:
(AB)i,j,k,(cid:96) =
Ai,p,q,(cid:96)Bp,j,k,q
which we will call "multiplication" and under
(cid:88)
(cid:88)
p,q
Before exploring explicit solutions to the ABC equation we point out a few general
features. We will exhibit a symmetry of the ABC equation that uses operations α
and β on certain subsets of ⊗4(cid:104). Note that ⊗4(cid:104) becomes a unital associative algebra
under (see remark 2.7)
(A.B)i,j,k,(cid:96) =
Ap,q,k,(cid:96)Bi,j,q,p
which we will abusively call "comultiplication". Both multiplications have obvious
diagrammatic representations. F will be the rotation by π/2:
p,q
F (A)i,j,k,(cid:96) = A(cid:96),i,j,k
LL00L01100111LL1010L1011L110 8
VAUGHAN F. R. JONES
The two multiplications are conjugate:
F (AB) = F (A).F (B)
Or, F gives an isomorphism from ⊗4(cid:104) under multiplication to ⊗4(cid:104) under comulti-
plication.
We will use X τ for the inverse of X for the comultiplication structure.
Lemma 2.15. We have
(1) If X−1 exists then F (X)τ does also and
F (X)τ = F (X−1)
.
(2) If X τ exists then F (X)−1 does also and
F (X)−1 = F (X τ )
.
(3) The same assertions hold with F replaced by F −1.
Proof.
(1) follows simply from the fact that F is an isomorphism from multiplication
to comultiplication.
(2) is a bit more subtle. Since F 2 is a comultiplication antiautomorphism,
F (X τ ) = F −1(F 2(X τ )) = F −1(F 2(X)τ )) = F (X)−1 , the last equality
being because F −1 is an isomorphism from comultiplication to multiplica-
tion.
(3) follows by applying F −2 = F 2 to both sides and using its antiautomorphism
properties.
(cid:3)
Definition 2.16. Let D = {XX−1 and X τ exist }.
Note that the set D is NOT invariant under taking inverses and coinverses. For
instance the element (d2 − 1)
considered below is invertible and coin-
vertible but its inverse is not coinvertible. But we can correct inverse and coinverse
by F to form α and β whose domain and range behave appropriately.
Definition 2.17. Let X ∈ ⊗4(cid:104) be such that both X and F −1(X) are invertible.
Then define
+ d
α(X) = F −1(X−1) and β(X) = F (X τ )
.
Observe that the domain of α is the set of invertibles and the domain of β is the
set of coinvertibles.
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
9
Lemma 2.18. If X ∈ Dom(α) then α(x) ∈ Dom(β) and β(α(X)) = X. Also
if X ∈ Dom(β) then β(x) ∈ Dom(α) and α(β(X)) = X. And α and β commute
with F 2.
Proof. The first two assertions follow easily from the previous lemma. The last
(cid:3)
assertion is trivial.
We shall now see how α and β arise in the ABC equation.
For simplicity we assume from now on that Y is invariant under rotation:
Yi,j,k = Yj,k,i.
Proposition 2.19. Suppose A ∈ domain(α), B ∈ domain(β). Then
D(A, B) = Y (C) ⇐⇒ D(β(B), F 2(C)) = Y (F 2(A)) ⇐⇒ D(F 2(C), α(A)) = Y (F 2(B))
Proof. Suppose D(A, B) = Y (C). Attach Bτ to the diagrams of D(A, B) and Y (C)
to obtain the following equality:
=
=
.
Rotate by 2π/3 and isotope a little to obtain:
=
.
Y (F 2(A)).
Rotating the appropriate tensors we get Y (C) = D(A, B) =⇒ D(β(B), F 2(C)) =
The process is clearly reversible so D(A, B) = Y (C) ⇐⇒ D(β(B), F 2(C)) =
(cid:3)
Y (F 2(A)), and the other equivalence is proved similarly.
BτCτABBAAτCB 10
VAUGHAN F. R. JONES
Unfortunately domain(αn) is not invariant under α for n ∈ Z.
Definition 2.20. Let us say that (A, B, C) ∈ ⊗4(cid:104) "scales" if it satisfies the ABC
equation and A, B, C ∈ domain(αn) for all n ∈ Z.
Theorem 2.21. Suppose F 2 = id and (A, B, C) scales. Then define Lw inductively
on words on {0, 1} by L0 = A, L1 = B,
αk+1(B)
αk(C)
αk−1(A)
αk(B)
if Lw = αk(A)
if Lw = αk(B)
if Lw = αk(C)
if Lw = αk(A)
if Lw = αk(B)
if Lw = αk(C)
and
Lw0 =
Lw1 =
αk(C)
αk(A)
Then Lw is a coherent choice of tensors so the Lw determine a scale invariant
transfer matrix by proposition 2.13.
Proof. This follows immediately by induction from 2.19. This formal proof some-
what obscures what is going on. The idea is that, once ABC is a solution, so are
α(B)CA, α(C)Aα(B), α(A)α(B)α(C), Cα−1(A)B and so on.
(cid:3)
Thus provided we have a solution to the ABC equation (with F 2 = id) the only
problem in using it to construct a scale invariant transfer matrix is the problem of
the domains of α and β. We will solve this completely in a special model.
2.3. A concrete example: the (quantum)SO(3)-invariant spin 1 chain. It
was explained in [15] how subfactors and bimodules provide quantum spin chains
more elaborate than with ordinary spins. The spin state space may fractional dimen-
sion. In this sense there is for instance a spin chain for the Andrews-Baxter-Forrester
models of [1], on which the transfer matrix acts. Subfactors/bimodules are known
to be described by planar algebras ([14, 16]) and all the calculations we have done
in this paper are diagrammatic, so extend without alteration to (unshaded) positive
definite planar algebras. The planar algebra is a graded vector space (cid:80) = (Pn) on
which the planar operad acts. Elements of Pn are called n-boxes and they may be
inserted into the internal discs of a planar tangle. Each Pn is equipped with an
antilinear involution ∗ for which the sesquilinear form
(cid:104)R, S(cid:105) =
is positive definite. (Illustrated here for P5.) The internal discs of planar tangles
have been reduced to points and the labels have been placed in regions near that
*RS SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
11
point which correspond to the distinguished interval on the boundary of the disc.
The output discs of all tangles have been eliminated but are of course implicit in the
diagrams.
The preceeding constructions for tensor networks work equally well for any (cid:80). One
chooses any Y =
from P3 satisfying the isometry condition. The direct system
of Hilbert spaces is defined by AY (t) = (cid:72)t = Pk if t is a tree with k leaves, and the
s are diagrammatic. The direct limit is again denoted (cid:72). Or (cid:72)(cid:80),Y if necessary. (In
ιt
fact all that is needed is an "annular" or "affine" representation of the planar algebra
in question-see [11, 16, 20]-here we are just using the "trivial" representation.)
We choose a planar algebra having the advantage that the 3-box space is one
dimensional.
It is the planar algebra for the quantum group UqSO(3) in its 3-
dimensional representation. Alternatively it can be obtained from the Temperley-
Lieb algebra TL(δ) ([21]) by doubling all strings and reducing by the JW idempotent
in the 4-box space of TL. Obtained in this way, its loop parameter is d = δ2 − 1. It
is positive definite when d = 4cos2π/n− 1, n = 6, 7, 8,··· or d ∈ R,≥ 3. We will call
this planar algebra (cid:81) = (Qn). (cid:81) is described in considerable detail in [23] (which
actually gives a list of all the simplest planar algebras generated by a single 3-box).
Remark 2.22. We should point out that in this case dim Q1 = 0 so the image of
in Q2 is zero dimensional and the spaces of the direct system really only begin
at Q2. Thus in forming a scale invariant transfer matrix from a solution of the ABC
equation there is no constraint on L0 and L1 which we can choose arbitrarily as "C"'s
in solutions of the ABC equation and then extend to all of (cid:72)(cid:81),Y as in theorem 2.21.
See remark 2.14
There is an "a priori" solution of the ABC equation given by the braiding (which
can be established by viewing (cid:81) as a cabled Temperley-Lieb planar algebra). We
draw the picture:
The braid solution A = B = C =
:
The trivalent vertex
is the unique (up to an irrelevant sign) non-zero self-
adjoint element of Q3. It is rotationally invariant and it is shown in [23] that the
following "skein" relations suffice to do all calculations in (cid:81) (along with positive
definiteness):
= 0,
= 1
d−1 (
) , and of course unitarity,
.
= 12
VAUGHAN F. R. JONES
These structure constants are all real so we will be mainly thinking of real solutions
to the ABC equation. Our answers will apply to complex solutions though and the
braid solutions are in fact complex for d < 3.
Since F 2 = id on Q4, we know that a scale-invariant transfer matrix can be
formed from a solution of the ABC equation provided all powers αn(A) (or B or C)
are invertible for both algebra structures. If X = p
we record
the formula
+ r
+ q
α(X) = −
p
q(p + q)
dqr − pq − pr − qr
q(d − 1)(p + q)(q + dr)
+
(d − 2)p + (d − 1)q
(d − 1)q(p + q)
+
+ a3
+ c2
+ c3
We now solve the ABC equation using this formula for α to optimise reduction of
the number of unknowns from 9 to 7.
+ b2
+ b3
+ a2
, B = b1
Let A = a1
and C =
, which we will represent in the more compact vector
c1
notation A = (a1, a2, a3), B = (b1, b2, b3) and C = (c1, c2, c3). Since A and B have
to be in the domain of α, it must be true, by our formula for α, that a2 and b2 are
non-zero. Thus all solutions that we are considering come from solutions in which
a2 = 1 and b2 = 1.
Let's do a count of equations and unknowns. The equations happen in the 5-box
space which in this case is (at most) 6 dimensional. It is spanned by 6 explicit tangles,
and the
5 of which are the rotations of a tangle with a single trivalent vertex
other is any connected tangle with three instances of
. The inner product on
Q5 is positive definite since it is the restriction of the Temperley-Lieb inner product.
So by taking inner products with the 6 spanning elements we obtain 6 equations in
the 9 coefficients, now 7 after putting a2 and b2 = 1, of A, B and C. So we have 6
equations in 7 unknowns and expect the solutions to depend on a single parameter.
to the top of the ABC
equation we note the appearance of a tangle from [17] which will be used frequently
later.
Now let us turn to the equations in detail. Attaching a
Definition 2.23. The renormalisation tangle below defines the nonassociative alge-
bra structure ! on Q4:
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
13
x!y =
Three of the seven equations for A, B and C are consequences of A!B = C so
we are left with a system of 3 equations in 4 unknowns for A and B. So we expect
solutions for three of the variables depending on a fourth.
Here are the three equations (before choosing a2 = b2 = 1):
(1)
c1 + c2 = (a2 + da3)(
d − 2
d − 1
b1 + b3)
c1 + c3 = (a1 + a2)(b1 + db2 + b3)
(2)
d − 2
d − 1
(3)
(a1 + da2 + a3)(b2 + db3) + (
d − 2
d − 1
a1 + a3)(b1 + b3) =
1
d(d − 1)
{ d(d − 2)
d − 1
a1b1 + d(a2b2 + a3b3 + a1b3 + a3b1 + d(a2b3 + b2a3))}
We illustrate with the first equation which comes from taking the inner product
of both diagrams in the ABC equation with
. Doing this we get the
equality:
=
The equations are all of the form "(cid:80)
which readily yields equation (1) by applying the skein relations of (cid:81).
and z (remember C = A!B). Thus in principle they are easy to solve.
i,j yi,jaibj = z" for three sets of constants yi,j
yxCAB 14
VAUGHAN F. R. JONES
We have solved the equations for A, B and C in terms of a1 which we call a, and
found (for d (cid:54)= 3) the unique solution:
A = (a, 1,
(d − 3)a
a + d2 − 3d + 2
B = (− (d − 1)2(a + 1)
a
, 1,
)
(d2 − 4d + 3)(a + 1)
a + d − 1
)
C = (
(d − 1)2(a + 1)((d − 2)a + d − 1)
−a(a + d − 1))
(d − 1)(a + 1)
a
,
,
(a + 1)((d − 2)a + d − 1)
−a
)
These equations are still somewhat messy. The key to further progress is to observe
that the transformation α preserves the solutions. In fact
α(A) = const(
(d − 1)((d2 − 3d + 1)a + d2 − 3d + 2)
((d − 2)a + d − 1)
, 1,
((d2 − 3d + 1)a + d2 − 3d + 2)
a
)
where const =
so the effect on the variable a is the linear fractional transformation:
(1 + a)(d − 1)((d2 − 3d + 1)a + d2 − 3d + 2)
a((d − 2)a + d − 1)
σ(a) = − (d − 1)((d2 − 3d + 1)a + d2 − 3d + 2)
((d − 2)a + d − 1)
The case d = 3 is special and we will deal with it in the next subsection.
So if we assume d (cid:54)= 3 and change variables to
with d = ω + ω−1 + 1, then
a = − (1 + ω2)
ω2
(z + ω2)
z + 1
, 1,− (ω − 1)(z + ω2)
ω(ωz − 1)
),
A = (− (1 + ω2)(z + ω2)
ω2(ω3 − z)(ω2 + z)
(ω5 − z)(ω4 + z)
ω2(z + 1)
−z
ω3 ), B = A(
A(
z
ω2 ) and A =
ω(z − ω)
z − ω3 C(ωz).
α(A(z)) =
Theorem 2.24. Suppose d (cid:54)= 3. Let A, B and C satisfy equation 1 for some C(z).
Then (A, B, C) scales in the sense of 2.20 provided z (cid:54)= ±ωn for any n ∈ Z.
Proof. The spectrum of A is easy to work out by changing to the basis of minimal
idempotents for Q4. It is
{
z − ω5
ω2(ωz − 1)
,− z + ω4
ω2(z + 1)
, 1}
So provided z (cid:54)= ±ωn for any n, the above formulae show that none of A, B or C
(cid:3)
nor any power of α applied to them, has zero in its spectrum.
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
15
Corollary 2.25. For any two complex numbers z0 and z1 satisfying zi (cid:54)= ±ωn for
i = 0, 1, there are scale invariant transfer matrices T (z0, z1)t on (cid:72)(cid:81),Y such that, on
(cid:72)
they are:
T (z0, z1)
=
Moreover L0w is of the form κA(εωpz0) where κ ∈ C, ε ∈ {±1} and p ∈ Z are all
determined by w, and similarly for L1w
Proof. By the formula before theorem 2.24 we can begin with either A, B or C and
theorem 2.21 shows how to extent to all the (cid:72)t. Explicit formulae for κ ∈ C, ε ∈ {±1}
and p ∈ Z could be constructed from theorem 2.21 but we will only need the fact
(cid:3)
that they are determined by the leaf coded for by w.
We have thus determined all values of z for which this procedure gives a scale
invariant transfer matrix. Note that a priori we could have asserted that, for d =
4 cos2 π/n − 1 and n = 6, 7, 8,··· there is a non-empty open interval of real values
of a for which the semicontinuous limit transfer matrix exists. This is because the
transformation α is then periodic so at most a finite set of values of a can be bad.
Remark 2.26. There are only isolated values of C for which A = B. Thus the
scale invariant transfer matrix is not spatially homogeneous for an interval's worth
of spectral parameter values.
2.4. Commuting transfer matrices. We are unable to think of an a priori reason
why the solutions of the ABC equations should have anything to do with the Yang-
Baxter equation ([4, 31]). We will see, however that not only are the solutions we have
obtained a well known solution of YBE, but also the spatially inhomogeneous transfer
matrices on the semicontinuous limit (on the circle-periodic boundary conditions)
commute for different values of the parameter. Note that this is not automatic-
given a solution R(λ) of the YBE , one may not choose λi arbitrarily and expect
T (R(λ1), R(λ2),··· , R(λn)) to commute with each other.
We will identify our solutions with the "Izergin-Korepin" model. In order to do
this we will express the solutions of the ABC equation that we found in section 2.3
in terms of another basis of Q4- the "braid basis" consisting of a crossing, its inverse
and the identity. If we set
R = (−ω − ω−1)
+ ω−1
+ (ω − 1)
(recall d = 1 + ω + ω−1) then the braid equation holds and moreover R + R−1 =
−(w+1/w)(
) so we are dealing with representations of the BMW algebra
[5, 24].
+
A (z )A(z )01 16
VAUGHAN F. R. JONES
If we rewrite A(z, ω) from subsection 2.3 in the braid basis we obtain:
A =
1
(1 + z)(ω − 1)
R + z
1 + ω + ω3 + ω4
(1 + z)ω(z + ω2)
id +
zω
(1 + z)(1 − ω)
R−1
Multiplying by (1 + z)(ω − 1)ω−3/2(1 + ω2
z ) and putting z = ω2
x one gets
A(cid:48) =
1 + x
ω3/2
R +
(ω2 − 1)(ω3 + 1)
ω5/2
id − (1 + x)ω3/2
x
R−1.
Putting ω = e−η/2 and x = −e−λ and R(λ, η) = A(cid:48), one obtains exactly twice for-
mula (III.9) for R(λ, η) in [22]. This is the Izergin-Korepin R matrix [12]. It thus
satisfies the Yang-Baxter equation ([4, 31]):
=
We will use this in the following form which is easily deduced from the previous
equations:
=
Definition 2.27. For all λ and t put T λ = T (−e−λ,−e−λ)t (see 2.25).
Theorem 2.28. For all t ∈ T and all λ and µ,
t T µ
T λ
t = T µ
t T λ
t
Proof. We begin by reviewing the well known diagrammatic argument for commut-
ing transfer matrices with periodic boundary conditions and spatially homogeneous
R( )λR( )λ+µµR( )R( )λ+µR( )λµR( )λ−µ−η/2)λµR( )R( )R( λ−µ−η/2λµR( )R( )R( ) SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
17
dependence on the spectral parameter (see [15]). One starts with
T (λ)T (µ)
Then one attaches R(λ − µ − η/2) and its inverse (for horizontal multiplication)
to the right of the picture so as not to change the operator T (λ)T (µ). Then suc-
cessively apply YBE, moving R(λ − µ − η/2) to the left one step at a time, each
time interchanging an R(λ) with an R(µ). Thus an intermediate step would look like:
the inverse.
where we have used a * to indicate the spectral parameter value corresponding to
When R(λ − µ − η/2) has reached the left side of the picture it meets its inverse
because of the periodic boundary condtions, and disappears, to leave
T (µ)T (λ) =
This argument goes through almost without alteration for any T λ
t . As
the R(λ − µ − η/2) goes to the left, it will meet the leaf coded for by the word w
t and T λ
(on {0, 1}), the pair
where L is, by corollary 2.25, κR(λ + φ) and M is
κR(µ + φ) for some κ and φ depending only on w. The κ factors are the same on
λµµµµµµλλλλλµλλµλ−µ−η/2(λ−µ−η/2)*λµµλλµλµµµµµµλλλλλLM 18
both sides of the YBE and the φ's cancel in the YBE so that R(λ − µ − η/2) goes
past the pair, swapping L and M.
(cid:3)
VAUGHAN F. R. JONES
2.5. The case d = 3, i.e. ω = 1. Note that in this case the planar algebra (cid:81) is the
tensor planar algebra based on a 3-dimensional Hilbert space so the vector spaces
(cid:72)t are ordinary spin chains.
If we solve equations (1) and (2) of the previous subsection for a1 and a3 and
substitute in equation 3 we find that the solution curve for d (cid:54)= 3 degenerates into a
line pair b1 = −2 and b3 = 1. These two lines are interchanged by α. Here are the
two solutions which we parametise by v and w to avoid confusion:
A1 = (v, 1, 0), B1 = (−4
1 + v
v
, 1, 0), and C1 = 2
1 + v
v
(−2, 1,−(1 +
v
2
))
and
A2 = (−2, 1, w), B2 = (−2, 1,
w
), and C2 = (−2
w + 1
2w + 1
w + 1
, 1, 0)
The transformation α is a little complicated with this parametrisation so we in-
troduce the changes of variables
Then α(A1(s)) =
3s − 1
3s
v(s) =
and w(t) =
6s − 2
−3s + 2
1
.
3t − 4
3t − 2
3t − 1
A2(s + 1) and α(A2(t)) =
A1(t). So
α2(A1(s)) =
3s
3s − 1
3s + 1
3s + 2
A1(s + 1).
We have the following other relations among the A's, B's and C's:
(1) B1(s) = A1(
3s + 1
)
(2) C1(s) =
3s − 1
3s
(3) B2(t) = A2(
3
A2(
3s + 2
3
)
)
3t + 1
3t − 1
3
)
3
(4) C3(t) = A1(
Thus any one of A1, A2, B1, B2, C1, C2 determines all the others via linear frac-
tional transformations of the parameters, and multiples which are themselves linear
fractional transformations of the parameters.
We finally turn to the question of when the solution A1, B1, C1 scales in the sense
of 2.20. This will be true provided two conditions are satisfied:
(1) The constant factors introduced when applying powers of α must never be
zero or infinity.
(2) The spectrum, for both multiplication and comultiplication, of all powers αn
of all A's B's and C's must not contain zero.
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
19
The three values of the spectrum of an element (x, y, z) are determined by linear
functions with constant coefficients of x, y, z. Since applying α just adds 1 to the
argument s We see that there are a finite number of linear fractional transformations
βis+γi
δis+i
and constants κi so that A1(s), B1(s), C1(s) scales provided
βi(s + n) + γi
δi(s + n) + i
(cid:54)= κi,
which can be written s /∈ zi + Z for any n and some finite set of numbers zi. Clearly
there are non-empty open intervals of s-values for which this is true. The same
argument applies to A2(s), B2(s), C2(s),
We have thus proved, by theorem 2.21 :
Theorem 2.29. Suppose d = 3. There are non-empty open intervals of s values for
which the transfer matrix
on Q2 = (cid:72)
boundary conditions as usual.)
extends to a scale-invariant transfer matrix on (cid:72)(cid:81),Y .
(Periodic
3. Spatially homogeneous scale invariant transfer matrix and
Hamiltonian via quadratic forms.
3.1. Generalities. It is easy to argue on physical grounds that scale invariance
for an observable should be expressed in terms of its expected values. This leads
immediately to the notion of weak scale invariance in definition 2.1. The word
"weak" is unfortunate in this context but is in universal use in functional analysis to
describe properties determined by matrix coefficients rather than an operator as a
whole. A strong motivation for considering weakly scale invariant transfer matrices
is that they arose inevitably in [17] in the calculation of the correlation of a state
with its rotation by a single lattice site.
Indeed our construction of weakly scale
invariant transfer matrices will allow us to deepen the study of the behaviour of the
rotation by one lattice site.
We immediately give the criterion for weak scale invariance, invoking the algebra
structure x!y of definition 2.23. Let L, t and w be as in definitions 2.10 and 2.11.
Proposition 3.1. The map t (cid:55)→ T L
direct system AY provided
t defines a weakly scale invariant operator on the
for all w.
Proof. This follows just as in proposition 2.13
Lw0!Lw1 = Lw
(cid:3)
1A (s)B (s)1 20
VAUGHAN F. R. JONES
At the risk of labouring the point, here is the picture of the equation
with A = Lw0, B = Lw1 and C = Lw:
Lw0!Lw1 = Lw
=
This makes it obvious that a solution of the ABC equation provides one of the
equation A!B = C (scale invariance implies weak scale invariance). But there are in
general many more solutions and weakly scale invariant transfer matrices can exhibit
quite arbitrary dependence on the lattice point. We shall thus focus on spatially
homogeneous ones which are provided by solutions of the equation X!X = X.
Definition 3.2. If (cid:80) is a planar algebra and Y =
map (cid:82) : P4 → P4,
(cid:82)(X) = X!X
∈ P3 we call the quadratic
the renormalisation dynamical system of (cid:80), Y .
Corollary 3.3. If X = {Xi, i = 1, 2, 3,···} is a sequence of elements of P4 obtained
by back-iterating (cid:82), i.e. such that (cid:82)(Xi+1) = Xi, then we get a weakly scale invariant
transfer matrix by putting
for all leaves (cid:108) of all trees.
Lw = Xlength(w(cid:108))
In this case the sesquilinear form is spatially homogeneous since if all the leaves of
a tree have the same length, the same value of the spectral parameter is used.
Definition 3.4. We call [, ]X the sesquilinear form on (cid:72) determined by this weakly
invariant transfer matrix (see proposition 2.3). If X is a fixed point for (cid:82) we let X
mean the constant sequence with all terms equal to X.
Let us make a couple of easy general remarks. From now on we will use (cid:82)(cid:48) to
denote (cid:82) on the projective space PP4 and, for X ∈ P4, (cid:82)(cid:48)(X) to denote (cid:82)(cid:48) of the
class of X in PP4.
i of back iterates of (cid:82)(cid:48) can always be lifted to back iterates Xi
(1) A sequence X(cid:48)
of (cid:82). The liftings are unique up to signs.
ABC SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
21
(2) A sequence Xi as above is by definition an element of the projective or
inverse limit of the inverse system over N all of whose spaces are P4 and
whose connecting maps are all (cid:82). Thus up to signs the space of spatially
homogeneous scale invariant transfer matrices is a subspace of the inverse
limit of PP4 with connecting maps (cid:82)(cid:48), a compact space.
(3) Back iterating from a given X1 may or may not be possible since (cid:82) may
not be surjective.A special case where back iteration is always possible is
obtained from a fixed point or more generally a periodic point W for (cid:82) (i.e.
there exists a p ∈ N for which (cid:82)p(W ) = W ).One may then put X1 = W and
choose the back iterates Xi to be (cid:82)j(W ) for the appropriate choice of W .
We let [, ]W be the sesquilinear form for this choice of back iterates.
(4) In an entire neighbourhood of a repelling fixed point of (cid:82) back iteration is
always possible:
Definition 3.5. A fixed point X for (cid:82) is called repelling if there is a norm
− on P4 and > 0 such that (cid:82)(Y )−X > Y −X whenever Y −X <
.
Theorem 3.6. If X is a repelling fixed point for (cid:82) there is a neighbourhood
V of X such that for all Y ∈ V, there are Yi ∈ V, i = 1, 2,··· , with Y1 =
Y, (cid:82)(Yi+1) = Yi for all i.
The corresponding weakly invariant invariant transfer matrix is weakly an-
alytic as a function of Y .
Proof. By the repelling property (cid:82) can be inverted locally to give (cid:82)−1
which takes the ball of radius inside itself. So choose the back iterates
by iterating (cid:82)−1. The inverse is analytic so that if ξ and η are fixed
[ξ, η]{Yi} = (cid:104)((cid:82)−1)nξ, η(cid:105) for some fixed n which is as analytic as (cid:82)−1.
(cid:3)
3.2. Topsey-turvey momenta. One can rephrase Stone's theorem on one param-
eter unitary groups as follows: "Given a strongly continuous one parameter unitary
group on Hilbert space, t (cid:55)→ ut, (cid:104)utξ, η(cid:105) − (cid:104)ξ, η(cid:105) tends to zero as t → 0, but one
may renormalise (cid:104)utξ, η(cid:105) by dividing by t to obtain a sesquiliear form [, ] on a dense
subspace such that
(cid:104)utξ, η(cid:105) − (cid:104)ξ, η(cid:105)
→ [ξ, η].
t
Thus in quantum mechanics one obtains energy and momentum from time evolu-
If ρn is the rotation (=translation) by 1
tion and space translation.
2n on the direct limit Hilbert space we saw
in [17] that (cid:104)ρnξ, η(cid:105) is determined by iterating (cid:82) starting with
. (cid:82) being a
homogeneous quadratic, there is always a neighbourhood of zero within which all
points iterate rapidly towards zero under it. The result of [17] followed simply by
showing that
is in such a neighbourhood after a few iterates. In this paper we
will take inspiration from Stone's theorem and renormalise (cid:104)ρnξ, η(cid:105) so that it has
limits as n → ∞. These limits will be expressed in terms of [ξ, η]X for some X's.
Unfortunately we may not always know the exact rate at which (cid:104)ρnξ, η(cid:105) approaches
22
VAUGHAN F. R. JONES
zero but it will be possible to create the quadratic form without that knowledge.
The situation is a little more complicated than in Stone's theorem as there will be
two quadratic forms rather than one that control the behaviour of (cid:104)ρnξ, η(cid:105).
Since ρn is a unitary given by spatial translation (by a single lattice site) we will
call these quadratic forms the topsey turvey momenta.
Lemma 3.7. Suppose X is a fixed point for (cid:82) and that Xn is a sequence in P4 with
limn→∞ Xn = [X]. Then there exist constants cn ∈ C such that, for any ξ, η ∈ (cid:72),
n→∞ cn(cid:104)TXnξ, η(cid:105) = [ξ, η]X
lim
Proof. The action of C× on P4 \ {0} gives a locally trivial fibre bundle with base
space PP4. Thus in a trivialising neighbourhood of X one can just lift the Xn to
cnXn so they have the same vertical coordinate as X in this trivialisation. Fix ξ and
η in some P2k. By the estimate of theorem 4.0.1 of [17], λnXn − X tends to zero
so cn(cid:104)TXnξ, η(cid:105) → (cid:104)TX ξ, η(cid:105) = [ξ, η]X.
(cid:3)
Now we can get the result we want:
Theorem 3.8. Suppose X is a fixed point for (cid:82)p and that W ∈ P4 is such that
limn→∞ (R(cid:48))np+i(W ) = ((cid:82)(cid:48))i(X) for i = 0, 1,··· , p − 1. Then there exist constants
cm ∈ C such that, for any ξ, η ∈ (cid:72), and i = 0, 1,··· , p − 1,
n→∞ cnp+i(cid:104)T(R)np+i(W )ξ, η(cid:105) = [ξ, η](cid:82)i(X)
lim
Proof. Apply the argument of the previous lemma to the sequence (cid:82)np(W ) then
apply (cid:82) p − 1 times to the conclusion.
(cid:3)
Remark 3.9. An explicit choice of cn can generally be made by choosing ξ = η =
Ω =some unit vector in P1 (or P2 if dim(P1) = 0) to obtain the choice
cn =
[Ω, Ω]X
(cid:104)TXnΩ, Ω(cid:105)
but it would be better to have an exact expression for this, at least asymptotically.
(Note that the numerator is just a simple quadratic in the coefficients of X.
As an immediate corollary we get the topsey-turvey momentum operators:
Theorem 3.10. Suppose there is an X ∈ P4 so that (cid:82)2 converges to X on iteration
. Then there are two sesquilinear forms [, ]±, and numbers cn and
starting at
dn such that, for ξ, η ∈ P22k,
and
lim
n→∞ cn(cid:104)ρ2n(ξ), η(cid:105) = [ξ, η]+
n→∞ cn(cid:104)ρ2n+1(ξ), η(cid:105) = [ξ, η]−
lim
We will see in the next subsection that the existence of X holds in every example
that we examine. It might be true universally.
SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
23
3.3. The case (cid:80) = (cid:81). In this case (cid:80) = (cid:81), the dynamical system is a quadratic
map from C3 to C3 so should properly be thought of as a dynamical system on CP2.
Once again the braid solution is an important one which will be missed over the
reals, at least for d < 3, but in keeping with this paper we will tend to think of
(cid:82) as a map on RP2 which means we can fix a circle at infinity and draw pictures
in the plane. Beware that this is not the plane of CP1 much beloved of dynamical
systems people. In particular (cid:82) is not surjective. We have seen that both iteration
and back-iteration of this dynamical system are relevant. Note that fixed points on
CP2 and C3 are the same apart from 0 and ∞ since a point in C3 that is fixed up to
a scalar can be rescaled to an absolute fixed point.
Let
X = p
+ q
+ r
= (p, q, r)
be an arbitrary element of Q4.
Proposition 3.11.
(cid:82)(X) = (
d2 − 5d + 7
(d − 1)2 p2 + 2pq + 2
d − 2
d − 1
(d − 1)3 p2 − 1
+(−
d − 1
d2 − 3d + 3
(d − 1)3 p2 +
d − 1
+
1
1
pr + q2 + r2)
(2pq + q2))
(2pq + q2))
Proof. This is just a calculation using the skein theory of (cid:81).
(cid:3)
We want to understand the dynamical systems (cid:82) for different values of d. At
this stage there is insufficient justification for doing this for all values of d so we
will content ourselves with presenting representative results for 4 different regimes
of behaviour: d = 2, 2 < d < 3, d = 3 and d > 3. For a value chosen in each of these
cases we give a portrait of the significant points in RP2 with values rounded off to
a few decimal places. We have chosen to send the circle p = 0 to infinity since this
makes
the origin in the portrait.
√
(1) d = golden ratio, 1+
2
5
.
there. In this case Q4 has dimension equal to 2 and is spanned by
This case was not covered in [17] though it is actually easier than anything
and
holds. It is extremely
(see [23]). The relation
− 1
=
d
easy to calculate
(cid:82)(q
+ r
) = (r2 − q2
d
)
+ (q2 − r2
2
)
. One checks immediately that (cid:104)ρnξ, η(cid:105) → 0 very rapidly as n → ∞ just as
in [17]. Since dim(Q4) = 2, the projective version of (cid:82) is actually a rational
dynamical system on CP1 but it is rather boring-the Julia set is the unit circle
24
VAUGHAN F. R. JONES
(2) d = 2 This case does not merit a portrait since the entire plane is mapped to
is a fixed point for (cid:82)2 so that the sesquilinear
. The only fixed points
(cid:82) interchanges the inside and outside. It has an attracting orbit of period
two to which all points not on the unit circle converge under iteration.
the single line r = −q and
forms [, ]± are just given by [, ]
for (cid:82)(cid:48) are the braid crossings and the real point (−1/2, 1/2).
and [, ]
(cid:82)(
)
√
(3) d = 1 +
2
Portrait:
Legend:
The element
The element
of P4 = (0, 0)
of P4 = (−0.707, 1)
Stable points of period 2. (−0.825, 1.022) and (0.315,−0.118)
All the real fixed points for (cid:82). (−1.557,−0.850), repelling, (−1.332, 0.332),repelling,
(−0.61, 0.098), unstable, (−0.375,−0.625), repelling and (−0.186, 0.521), un-
stable.
Iterates to.
Backiterates to.
Missing are two complex fixed points, the positive and negative braid cross-
ings.
q1−1r−1+1 SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
25
(4) d = 3
Portrait:
Legend:
The element
The element
of P4 = (0, 0)
of P4 = (−0.5, 0.5)
Stable points of period 2. (−0.5, 0.781) and (0.281, 0)
All the real fixed points for (cid:82). (−1.675,−1.175), repelling, (−1.309, 0.309),repelling,
(−0.5, 0), neutral-this is the braid solution, (−0.191,−0.809), repelling and
(−0.075, 0.425), unstable.
Iterates to.
Backiterates to.
No complex fixed points this time. The braid solution is symmetric and
real.
(5) For d between 3 and just under 3.52783, the braid solution at d = 3 splits
into three fixed points, two of which are the braid solutions, but apart from
this the picture looks much like at d = 3.
q1−1r−1+1 26
VAUGHAN F. R. JONES
(6) Something more interesting happens at 3.52783. The two periodic points
that are the limits of
and
Here is the portrait for large d:
under iteration coalesce.
d = 26.04
Legend:
Backiterates to.
The element
The element
Unique attracting fixed point. Limit of
All the real fixed points for (cid:82)
of P4 = (0, 0)
of P4 = (−0.0399, 0.0399)
and
under iteration.
Remark 3.12. The fixed and periodic points other than the limits under iteration
and backiteration do not yet have any interest for the scale invariant physics but
they might be of interest for topology as they provide coefficient-like functions on
the Thompson group.
q1r−1−1+1 SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
27
4. Weakly scale invariant nearest neighbour Hamiltonians.
By a nearest neighbour Hamiltonian we mean an operator of the form:
n(cid:88)
id ⊗ id ⊗ ··· ⊗ hi ⊗ ··· ⊗ id
i=1
where hi are self-adjoint linear maps on (cid:104)⊗ (cid:104) and hi comes after the (i− 1)th. tensor
product symbol, with periodic boundary conditions. It is spatially homogeneous if
hi is independent of i. We will call this map Hn, being deliberately ambiguous about
the tensor power of (cid:72) on which it acts. (The extension to general planar algebras is
obvious.)
The analogue of the ABC equation has no solutions for (cid:81), spatially inhomogeneous
or not, but if we only require weak scale invariance there are solutions. We restrict
our attention to the spatially homogeneous case. We will see that the equation only
invokes a linear renormalisation so is very easy to solve.
Definition 4.1. The scale invariance map (cid:83) : End((cid:104)⊗ (cid:104)) → End((cid:104)⊗ (cid:104)) is the map
(cid:83)(h) =
+
being a tree with n leaves) Y being used to define the embedding maps ιt
We assume the notation of section 2.2 with the direct system t (cid:55)→ A(t) = ⊗n(cid:104), (t
s for s ≤ t.
We will content ourselves to use the scale invariance map to construct nearest
neighbour Hamiltonians on for the cofinal sequence (cid:84)n for which the vector space is
A((cid:84)n) = ⊗2n
(cid:104), leaving it to the reader to sort out the restrictions to an arbitrary
A(t).
Recall that for s ≤ t ∈ T, the inclusion
Theorem 4.2. Given h ∈ End((cid:104)⊗ (cid:104)) consider Hh on the Hilbert space A((cid:84)n). Then
for ξ, η ∈ A((cid:84)n),
so that (cid:104)Hhξ, η(cid:105) on ⊗2n
for all k.
(ξ)), ι(cid:84)n+1
< Hh(ι(cid:84)n+1
(cid:104) extends to a sesquilinear form on (cid:72) provided (cid:83)−n({h}) (cid:54)= ∅
(η) >=< H(cid:83)(h)ξ, η >
(cid:84)n
(cid:84)n
Proof. In evaluating < Hh(ι(cid:84)n+1
(η) > there are two kinds of terms ac-
cording to the parity of i in the sum defining Hh. In one kind h is surrounded by
(ξ)), ι(cid:84)n+1
(cid:84)n
(cid:84)n
hh (d−2)2
(d−1)2
0
d−2
d−1
0
0
2
d−3
d−1
1
(d−1)2
d−2
(d−1)2
}
(cid:3)
28
VAUGHAN F. R. JONES
and in the other kind by
. Grouping the terms together and
using periodic boundary conditions one obtains the result.
(cid:3)
neighbour hamiltonians.
Thus the eigenvectors of (cid:83) can be used to construct very scale-invariant nearest
Finally we calculate (cid:83) : Q4 → Q4 in our example to make sure it is generic enough
for our usual linear algebra intuition to hold.
Proposition 4.3. The matrix of (cid:83) with respect to the basis {
of Q4 is
,
,
Proof. Just calculate using the skein relations of (cid:81).
The eigenvalues of (cid:83) are thus 2,
(d − 2)2
(d − 1)2 and
,
1
(d − 1)2 with eigenvectors
(d2 − 2)
−(2d2 − 4d + 1)
+ (d − 1)(d − 2)
+ (2d2 − 4d + 1)
, and
+ (d − 2)2
respectively.
would get the identity Hamiltonian.
So provided d (cid:54)= 2 (cid:83) is invertible.
The eigenvalue 2 is of no interest since the eigenvector is just the identity and we
For d < 3 the largest eigenvalue of (cid:83)−1 is (d−2)2
(d−1)2 so any scale invariant spatially
homogeneous nearest neighbour Hamiltonian will tend, modulo scalars, to that with
h = (d2 − 2)
+ (d − 1)(d − 2)
And for d > 3 the same limit Hamiltonian will have h = −(2d2 − 4d + 1)
+
.
(2d2 − 4d + 1)
+ (d − 2)2
.
For both of these possibilities for h, as elements of the algebra Q4, their eigenvalues
are both positive and negative. One can force the spectrum of the Hamiltonian to be
positive on a given Qn by adding a large multiple of the identity, but on applying (cid:83)−1
enough times, as required by scale invariance, the spectrum will eventually contain
negative numbers as well as positive.
hh SCALE INVARIANT TRANSFER MATRICES AND HAMILTIONIANS.
29
Thus, not surprisingly for topsey turvey systems, the spectrum of a scale invariant
Hamiltonian can only be positive for the trivial Hamiltonian.
References
[1] Andrews, G. E., Baxter, R. J. and Forrester, P.J. (1984). Eight vertex SOS model and gener-
alized Rogers–Ramanujan type identities. Journal of Statistical Physics, 35, 193–266.
[2] Araki, H. (1969). Gibbs states of a one dimensional quantum lattice. Commun. Math Phys.
14 120–157.
√
13)/2 and (5 +
√
(5 +
[3] Asaeda, M. and Haagerup, U. (1999). Exotic subfactors of finite depth with Jones indices
[4] Baxter, R. J. (1982). Exactly solved models in statistical mechanics. Academic Press, New York.
[5] Birman, J. S. and Wenzl, H. (1989). Braids, link polynomials and a new algebra. Transactions
17)/2. Communications in Mathematical Physics, 202, 1–63.
of the American Mathematical Society, 313, 249–273.
[6] Cannon, J.W., Floyd,W.J. and Parry, W.R.(1996) Introductory notes on Richard Thompson's
groups. L'Enseignement Mathématique 42 215–256
[7] Cirac, J. I. and Verstraete, F. (2009) Renormalization and tensor product states in spin chains
and lattices. JOURNAL OF PHYSICS A-MATHEMATICAL AND THEORETICAL 42 (50)
[8] David E. Evans and Terry Gannon. The exoticness and realisability of twisted Haagerup-Izumi
modular data. Comm. Math. Phys., 307(2):463–512, 2011.
[9] G. Evenbly, G. Vidal, Tensor Network Renormalization, arXiv:1412.0732
[10] Fadeev, l. (1999) Instructive History of the Quantum Inverse Scattering Method Quantum
Field Theory: Perspective and Prospective Volume 530 of the series NATO Science Series pp
161-177
[11] J. J. Graham and G.I. Lehrer, The representation theory of affine Temperley Lieb algebras,
L'Enseignement Mathématique 44 (1998), 1–44.
[12] A. G. Izergin and V. E. Korepin, The Inverse Scattering Method Approach to the Quantum
[20] Jones, V. and Reznikoff, S. (2006) Hilbert Space representations of the annular Temperley-Lieb
algebra. Pacific Math Journal, 228, 219–250
[21] Kauffman, L. (1987). State models and the Jones polynomial. Topology, 26, 395–407.
[22] Symmetries of spin systems and the Birman Wenzl Murakami algebra
P. P. Kulish, N. Manojlovic, and Z. Nagy Journal of Mathematical Physics 51 (2010)
[23] S. Morrison, E.Peters, N. Snyder,
(2015) Categories generated by a trivalent vertex.
[24] Murakami, J. (1987). The Kauffman polynomial of lins and representation theory. Osaka Jour-
arXiv:1501.06869
nal of Mathematics, 24, 745–758.
Shabat-Mikhailov Model
Commun. Math. Phys. 79 (1981), 303-316
[13] V.F.R. Jones, Index for subfactors, Invent. Math. 72 (1983), 1–25.
[14] V.F.R. Jones, Planar Algebras I, preprint. math/9909027
[15] Jones, V. F. R. In and around the origin of quantum groups. Prospects in mathematical physics.
Contemp. Math., 437 Amer. Math. Soc. (2007) 101–126. math.OA/0309199.
[16] V.F.R. Jones, The annular structure of subfactors, in "Essays on geometry and related topics",
Monogr. Enseign. Math. 38 (2001), 401–463.
[17] V.F.R. Jones, A no-go theorem for the continuum limit of a periodic quantum spin chain.
[18] V.F.R. Jones (2014) Some unitary representations of Thompson's groups F and T,
[19] Jones, V. Morrison, S. and Snyder, N. (2013). The classification of subfactors of index ≤ 5. To
arXiv:1412.7740
arXiv:1607.08769
appear.
[26] Penrose, R. (1971). Applications of negative dimensional tensors. Applications of Combinatorial
[27] Simon, B (1970) Hamiltonians defined as quadratic forms Comm. Math. Phys. Volume 21,
bridge University Press.
Mathematics, Academic Press, 221–244
Number 3 (1971), 192-210.
30
VAUGHAN F. R. JONES
[25] Nielsen, M. and Chuang, I. (2011). Quantum Computation and Quantum Information Cam-
[28] Temperley, H. N. V. and Lieb. E. H. (1971). Relations between the "percolation" and "colouring"
problem and other graph-theoretical problems associated with regular planar lattices: some
exact results for the "percolation" problem. Proceedings of the Royal Society A, 322, 251–280.
[29] von Neumann, J. (1939) On infinite direct products. Compositio Mathematica, 6 1–77
[30] Wassermann, A. (1998). Operator algebras and conformal field theory III: Fusion of positive
energy representations of LSU (N ) using bounded operators. Inventiones Mathematicae, 133,
467–538.
[31] Yang CN (1968) S Matrix for the One-Dimensional N-Body Problem with Repulsive or At-
tractive δ-Function Interaction. Phys. Rev. 167, 1920–1923.
|
1712.09386 | 1 | 1712 | 2017-12-26T19:58:52 | On the geometry of idempotents in von Neumann algebras | [
"math.OA"
] | We consider the general linear group as an invariant of von Neumann factors. We prove that up to complement, a set consisting of all idempotents generating the same right ideal admits a characterisation in terms of properties of the general linear group of a von Neumann factor. We prove that for two Neumann factors, any bijection of their general linear groups induces a bijection of their idempotents with the following additional property: If two idempotents or their two complements generate the same right ideal, then so does their image. This generalises work on regular rings, such include von Neumann factors of type $I_{n}$, $n < \infty$. | math.OA | math |
ON THE GEOMETRY OF IDEMPOTENTS IN VON
NEUMANN ALGEBRAS
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
Abstract. We consider the general linear group as an invariant of von
Neumann factors. We prove that up to complement, a set consisting of
all idempotents generating the same right ideal admits a characterisa-
tion in terms of properties of the general linear group of a von Neumann
factor. We prove that for two Neumann factors, any bijection of their
general linear groups induces a bijection of their idempotents with the
following additional property: If two idempotents or their two comple-
ments generate the same right ideal, then so does their image. This
generalises work on regular rings, such include von Neumann factors of
type In, n ă 8.
Introduction
This project is our first contribution in an ongoing classification of von
Neumann factors. Here we consider the general linear group as an invariant.
We study how the following geometry of idempotents can be characterised
by the general linear group. Recall that an idempotent, also called a gener-
alised projection, is an element satisfying e2 " e. Let N be a von Neumann
factor. For each idempotent e P N let res denote the equivalence class of all
idempotents f satisfying f N " eN . We consider the following question:
Question A. Is there a way to characterise the set res of idempotents gen-
erating the same right ideal in terms of properties of the general linear group
of N ?
When considering the unitary group as an invariant Dye [2] proved the
following result back in 1954: Let N and M be two von Neumann factors
not of type I2n and ϕ a group isomorphism between their unitary groups,
then there exist a linear or conjugate linear -isomorphism of N and M
whose restriction to the unitary group agrees up to character with ϕ.
Our results follows closely the work by Baer [1] and Ehrlich [3]. Bear
considered the finite dimensional case and Ehrlich studied regular rings,
these include precisely the von Neumann factors of finite dimension, this
was pointed out by von Neumann in [5].
Date: July 2, 2018.
2010 Mathematics Subject Classification. 46L35, 46L05, 46L80.
1Communicating author.
1
2
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
Let N be a von Neumann algebra. Recall that there is a canonical bijec-
tion ιN , e ÞÑ 2e ´ 1 from the set of idempotents IpN q in N into the set of
involutions InvpN q in N . For each e P IpN q we let ∆`peq denote the image
of res via ιN and ∆´peq the set ´∆`peq. Two idempotents generate the
same right ideal precisely when their associated ∆`-sets are equal. Ques-
tion A is therefore equivalent to asking for a characterisation of a ∆`-set in
term of properties of the general linear group GLpN q. Our main theorem
states as follows:
Theorem B. Let N be a von Neumann factor. Let φ be a nonempty subset
of InvpN q. Then φ is a ∆`-set or a ∆´-set if and only if φ is a maximal
set among the nonempty subset of involutions in N satisfying (1)-(4):
(1) If u, v, w P φ then uvw " wuv P φ.
(2) If u, v P φ then there exist a unique w P φ s.t. wvw " u.
(3) If u P InvpN q, then uφ " φu iff uw " wu for some w P φ.
(4) If t P φ2 then pt ´ 1q2 " 0.
The final item (4) is not stated as a property of GLpN q, but this can be
done as follows: For each element t P N let Cptq denote the centraliser of t
in GLpN q and C 2ptq the second centraliser of t in GLpN q. We have
Theorem C. Let N be a von Neumann factor containing 1 ‰ s P GLpN q.
Then ps ´ 1q2 " 0 if and only if the following (1)-(4) holds:
(1) If t P GLpN q, then Cpsq Ĺ Cptq iff GLpN q " Cptq.
(2) There exists u P InvpN q such that usu " s´1, and
(3) an element r P C 2puq such that rsr´1 " s2.
(4) s3 ‰ 1.
Having established a characterisation of ∆`-sets (up to a sign) we apply
this result to bijections between the general linear groups of von Neumann
factors. For e, f P IpN q, write e « f whenever eN " f N . This is an
equivalence relation on IpN q. Let res def" tf P IpN q : f « eu denote the
equivalence class containing e and IpN q{« def" tres : e P IpN qu the set of
equivalence classes.
Theorem D. Let N and M be two von Neumann factors and let ϕ a group
isomorphism of their general linear groups. Let θ be the bijection of idem-
potents induced by ϕ, i.e., θ " ι´1
M ϕ ιN . Then there exist a partitioning
of the nontrivial elements of IpN q{« into two set Io, I¯o, such that
θpresq "
θpresq,
θpr1 ´ esq,
r1s,
r0s,
if e P Io
if e P I¯o
if e " 1
if e " 0
$''&
''%
is a bijection of IpN q{« and IpM q{«.
It is possible that Io or I¯o is the empty set, consider for example the
bijections ϕpuq " u or ϕpuq " u.
ON THE GEOMETRY OF IDEMPOTENTS IN VON NEUMANN ALGEBRAS
3
1. Elements of class 2 and the proof of Theorem C.
In this section we give a characterisation of elements of class 2 (see Defi-
nition 1.1) in terms of properties of the general linear group. We start with
a few definitions. Let N be a von Neumann algebra. We let GLpN q de-
note the set of invertible elements in N , IpN q the set of idempotents in N ,
InvpN q the set of involutions in N , and ZpN q the center of N , i.e.,
GLpN q def" tu P N : uv " vu " 1 for some v P N u, IpN q def" te P N : e2 " eu,
InvpN q def" tu P N : u2 " 1u, ZpN q def" tx P N : xy " yx for all y P N u.
The inverse of u P GLpN q is unique and is denoted u´1. Two idempotents
are orthogonal if they commute and their product is zero. For u P N set
Cpuq def" tt P GLpN q : ut " tuu,
C 2puq def" tv P GLpN q : tv " vt for all t P Cpuqu.
Definition 1.1 (cf. [1, 3]). Let N be a von Neumann algebra. An element
t P N is of class 1 if t " 1 and of class 2 if t ‰ 1 and pt ´ 1q2 " 0.
We start with a few standard facts on von Neumann factors. Most of the
properties are trivial and are included merely as a reference.
Lemma 1.2. Let N be a von Neumann factor. Then the following holds:
(1) For each x P N , x
2 is the unique y P N satisfying 2y " x).
(2) For each 0 ‰ n P N such that n2 " 0 there exists e, f P IpN q such
3 P N ( x
2 , x
that n " enf and f e " 0.
(3) For each 0 ‰ n P N , n2 " 0 there exists idempotents e, g P IpN q and
k P N such that n " eng, ge " eg " 0, e " nk and kn " g.
(4) For each u " 2e ´ 1 P InvpN q and r P C 2puq there exists elements
z1, z2 P ZpN q such that r " z1e ` z2p1 ´ eq.
(5) For each e P IpN q and d1 P eN e we have that if d1x " xd1 for all
x P GLpeN eq, then d1 " ze for some z P ZpN q.
(6) For e, g P IpN q from (3), f def" 1 ´ e ´ g and a3 P eN f , a4 P f N f ,
a5 P f N g we have a3pf N gq " 0 ñ a3 " 0, a4pf N gq " 0 ñ a4 " 0
and peN f qa5 " 0 ñ a5 " 0.
(7) If t P N , Cptq " GLpN q and pt ´ 2qpt ` 1q " 0 then t " 2 or t " ´1.
Proof. Left to the reader.
(cid:3)
We now present a few lemmata constituting the proof of Theorem C. No-
tice, each class 2 element s is invertible with inverse 2 ´ s, see Definition 1.1.
Lemma 1.3. Let N be a von Neumann factor. Then for each s P N of class
2 there exists u P InvpN q and r P C 2puq such that usu " s´1, rsr´1 " s2
and s3 ‰ 1.
Proof. Fix s " n ` 1 and select e, f as in Lemma 1.2(2). Since f e " 0 we get
f " f p1 ´ eq, so n " enf " enp1 ´ eq. It follows that u def" e ´ p1 ´ eq satisfies
4
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
u P InvpN q and usu " 1 ` unu " 1 ` uenp1 ´ equ " 1 ´ enp1 ´ eq " s´1.
Set r def" 1 ` e P GLpN q with r´1 " 1 ´ e
2 . Now, if t P Cpuq then tu " ut,
te " et and tr " rt so r P C 2puq. Moreover, rsr´1 " 1 ` renp1 ´ eqr´1 "
1` 2enp1´ eq " 1` 2n " s2. Since s3 " 1` 3n, and n ‰ 0 we get s3 ‰ 1. (cid:3)
Remark 1.4. Recall two idempotents e and g are orthogonal if eg " ge " 0.
For two such idempotents f def" 1 ´ e ´ g is also an idempotent and one can
write each element t P N as the sum t " ete ` etg ` etf ` get ` gtg ` gtf `
f te ` f tg ` f tf . Notice that ete P eN e and so on. This may be recorded as
t " ete etg etf
f te f tg f tf ,
gte gtg gtf
with respect to the ordering e, g, f . Since the idempotents are pairwise
orthogonal one can perform classical matrix multiplication to get products.
Lemma 1.5. Let N be a von Neumann factor. Then for each s " 1 ` n
of class 2 and t P GLpN q we have that t P Cpsq iff there exists pairwise
orthogonal idempotents e, g, f P IpN q and elements k, t1, . . . , t5 P N such
that e ` g ` f " 1, e " nk, g " kn, n " eng and
(1.1)
t " t1
t5
t2
t3
0 kt1n 0
0
t4 ,
with respect to the order e, g, f (so t2 P eN g and so on).
Proof. Fix s " 1 ` n of class 2 and t P GLpN q. By assumption on n there
exist e, g, k as in Lemma 1.2(3). Define f def" 1 ´ e ´ g.
Suppose t P Cpsq. Then ts " st and tn " nt. Using gn " gen " 0
we get gte " gtnk " gntk " 0. Similarly f te " f tnk " f ntk " 0 and
gtf " kntf " ktnf " 0. Finally, using en " n " ng we get the desired form
of t via kpeteqn " ketpenqg " ketng " kentg " kntg " gtg.
Conversely, for any t P GLpN q of the form as in (1.1) with e " nk, g " kn
and n " eng. We get st " ts from
n " ´ 0 eng 0
0 0 0¯ , nt " ´ 0 nkt1n 0
0¯ ,
0 0 0
0
0
0
0
0
tn " ´ 0 t1n 0
0 0 0¯ , nkt1n " et1n " t1n.
0 0 0
(cid:3)
Lemma 1.6. Let N be a von Neumann factor. Then for each s " 1 ` n of
class 2 and d P C 2psq there exists elements z1, z2 P ZpN q such that
d " z11 ` z2n.
Proof. Fix s " 1 ` n of class 2 and d P C 2psq. Since d P C 2psq Ď Cpsq,
there exists e, g, f P IpN q and elements k, d1 . . . , d5 P N representing d as
in (1.1). Using Lemma 1.5 on the following t1, t2, t3 P GLpN q,
t1
0
0 kxn 0
0
def" ´ x 0
0 f ¯ , t2
def" e 0 0
0 y f , t3
0 g 0
def" e 0 z
0 0 f , d " d1 d2 d3
d5 d4 ,
0 kd1n 0
0
0 g 0
with (any) x P GLpeN eq, y P f N g, z P eN f we have that each ti P Cpsq.
Since d P C 2psq, d commutes with each ti P Cpsq. In particular, dt1 " t1d,
ON THE GEOMETRY OF IDEMPOTENTS IN VON NEUMANN ALGEBRAS
5
so d1x " xd1 in eN e for all x P GLpeN eq. Using Lemma 1.2(5) we know
d1 " z1e for some z1 P ZpN q. Now, since kd1n " kpz1eqn " z1kn " z1g,
d " d1 d2 d3
d5 d4 " z1e d2 d3
d5 d4 " z11 ` d1
0 kd1n 0
0
0 z1g 0
0
for d1 def" 0 d2
0 d5 d4´z1f .
d3
0
0 0
Successively using d1t2 " t2d1 and d1t3 " t3d1 we get d3 " 0, d4 ´ z1f " 0
and d5 " 0 (by Lemma 1.2(6)). For r def" n ` k ` f the equality d1t1r " t1d1r
reduces to xpd2kq " pd2kqx in eN e. By Lemma 1.2(5), d2k " z2e for
some z2 P ZpN q. Consequently d1 " d2 " d2g " d2kn " z2en " z2n, so
d " z11 ` z2n.
(cid:3)
Proposition 1.7. Let N be a von Neumann factor. Then for each s P N ,
the element s is of class 2 if and only if (1)-(4) is satisfied:
(1) If t P GLpN q, then Cpsq Ĺ Cptq ô GLpN q " Cptq.
(2) There exists u P InvpN q such that usu " s´1, and
(3) an elements r P C 2puq such that rsr´1 " s2.
(4) s3 ‰ 1.
Proof. Fix s " 1 ` n P GLpN q. Suppose s is of class 2. Then (2)-(4) holds
by Lemma 1.3. To show (1) fix any t P GLpN q.
"ð": Since Cptq " GLpN q it suffices to show Cpsq ‰ Cptq. Assuming
Cpsq " Cptq we have GLpN q " Cptq " Cpsq " Cpnq, but r R Cpnq because
rnr´1 " s2 ´ 1 " p1 ` nq2 ´ 1 " 2n ‰ n.
"ñ" Since Cpsq Ď Cptq, we get t P C 2psq " tt1 : xt1 " t1x for all x P Cpsqu.
By Lemma 1.6, t " z11 ` z2n for some z1, z2 P ZpN q. Assume z2n ‰ 0.
Then x P GLpN q commutes with t iff it commutes with s " 1 ` n. This
implies Cpsq " Cptq, giving a contradiction. We deduce that z2n " 0, so
t " z11 P ZpN q and GLpN q " Cptq.
Suppose s " 1 ` n P GLpN q satisfy (1)-(4). Set t " s ` s´1. By (2)
we get u P InvpN q such that utu " t. By (3), r P C 2puq. Hence property
Lemma 1.2(4) provides elements z1, z2 P ZpN q such that r " z1e ` z2p1 ´ eq
for e def" u`1
2 . We know that t commutes with u, e, and r. It follows (by (3))
that t " rtr´1 " rps ` s´1qr´1 " s2 ` s´2 " ps ` s´1q2 ´ 2 " t2 ´ 2, so
tpt ´ 1q " 2 implying t P GLpN q.
We now show Cpsq Ĺ Cptq: "Cpsq Ď Cptq": If x P Cpsq then s´1pxsqs´1 "
s´1psxqs´1 so x P Cps´1q and x P Cptq. "Cpsq ‰ Cptq": We know u P Cptq.
Assuming u P Cpsq, we get s " s´1 by (2). Using (3), 1 " s2 " rsr´1, so
s " 1. Hence GLpN q " Cpsq and by (1), Cpsq Ĺ Cpsq (false). So u R Cpsq.
Knowing Cpsq Ĺ Cptq we get GLpN q " Cptq by (1). Since we already
established pt ´ 2qpt ` 1q " 0, we get t " 2 or t " ´1 by Lemma 1.2(7).
If t " ´1 then ´1 " s ` s´1, s3 " p1 ` s´1qsp1 ` s´1q " ps ` 1qp1 `
s´1q " 2 ` t " 1, contradicting (4). Therefore t " 2. Hence s ‰ 1 and
ps ´ 1q2 " s2 ` 1 ´ 2s " sps ` s´1 ´ 2q " spt ´ 2q " 0. So s is of class 2. (cid:3)
Proof of Theorem C: Since 1 ‰ s P GLpN q, s is of class 2 iff ps ´ 1q2 " 0.
(cid:3)
The result now follows from Proposition 1.7.
6
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
Remark 1.8. Theorem C remains valid in greater generality. In fact, for any
unital ring N satisfying properties (1)-(7) of Lemma 1.2, we have the same
characterisation of elements of class 2 in terms of the general linear group.
This is because no other properties of N as a unital ring were used in the
proof.
2. ∆-sets and the proof of Theorem B
In this section we prove the main result. This is a characterisation of a
natural equivalence relation on the set of idempotents, but expressed via
properties of the general linear group.
Let N be a von Neumann algebra. We let P pN q denote the set of projec-
tion elements in N , i.e.,
P pN q def" tp P N : p " p " p2u.
Definition 2.1 (cf. [1, 3]). Let N be a von Neumann algebra containing an
idempotent e, an involution u and a nonempty subset φ of involutions. Set
φ2 def" tuv : u, v P φu and
I puq def" tx P N : ux " xu, I `pφ2q def" tx P N : tx " x for all t P φ2u,
N pφq def" tv P InvpN q : vφ " φvu, ∆peq def" tv P InvpN q : I pvq " eN u.
Definition 2.2 (cf. [1, 3]). Let N be a von Neumann algebra containing a
nonempty subset φ of involutions. We say φ is a ∆`-set (resp. ∆´-set) if
φ " ∆`peq (resp. ∆´peq) for some idempotent e. We call φ a ∆-set if it is
a ∆`-set or a ∆´-set for some e P IpN q.
As in the previous section we include a few standard facts on von Neu-
mann factors. Two idempotents e, f are similar if e " uf u´1 for some
invertible element u.
Lemma 2.3. Let N be a von Neumann factor. Then the following holds:
2 P N ( x
2 is the unique y P N satisfying y ` y " x).
def" inftp P P pN q : x " pxu P P pN q and x " pxx.
def" inftq P P pN q : y " yqu P P pN q and y " yqy.
(1) For each x P N , x
(2) P pN q is partially ordered via p ď q iff pq " qp " p.
(3) If x P N then px
(4) If y P N then qy
(5) If X Ď N then qX
(6) If yx " 0 for y, x P N then qypx " 0.
(7) If y, x P N are both nonzero then yN x ‰ t0u.
(8) If x P N then px " uu and qx " uu for some u " uuu P N .
(9) Every idempotent in N is similar to a projection in N .
def" suptqx : x P Xu P P pN q.
Proof. Left to the reader.
(cid:3)
We recall a few standard facts. These are purely algebraic observations.
The proofs are included for completeness.
Lemma 2.4 (cf. [1, 3, 5]). Let N be a von Neumann algebra containing an
idempotent e. Then
ON THE GEOMETRY OF IDEMPOTENTS IN VON NEUMANN ALGEBRAS
7
(1) If u def" 2e ´ 1 then u P InvpN q, I `puq " eN and I ´puq " p1 ´ eqN .
(2) If u P InvpN q then I `puq ` I ´puq " N and I `puq X I ´puq " t0u.
(3) If u, v P InvpN q satisfy I puq " I pvq then u " v.
(4) If e, f P IpN q then eN " f N ô f " e ` eyp1 ´ eq for some y P N .
Proof. (1): If e P IpN q then p2e´1q2 " 1, so 2e´1 P InvpN q. Set u def" 2e´1.
Then I `puq " tx P N : p2e ´ 1qx " xu " tx P N : 2ex " 2xu " eN and
I ´puq " tx P N : 2ex " 0u " tx P N : p1 ´ eqx " xu " p1 ´ eqN .
(2): If x P N , set x
P N . Then ux " x, so x " x` ` x´,
x` P I `puq and x´ P I ´puq giving N " I `puq ` I ´puq. If x P I `puq X I ´puq
then ux " x " ´p´xq " ´ux so 2ux " 0. Using u2 " 1 we get x " 0.
def" xux
2
(3): For suitable e, f such that u " 2e´1 and v " 2f ´1 we get eN " f N
and p1 ´ eqN " p1 ´ f qN . Hence ef " f and p1 ´ eqp1 ´ f q " 1 ´ f . The
latter gives 1 ´ e ´ f ` ef " 1 ´ f . So ef " e, f " ef " e and u " v.
(4): 'ð' f N " pe ` eyp1 ´ eqqN Ď eN , and f e " e so eN " f eN Ď f N .
'ñ' Using e " ee " f x for some x, f e " f pf xq " f x " e. Similarly ef " f .
One now verify f " e ` eyp1 ´ eq for y def" f ´ e.
(cid:3)
We are now in position to establish properties of a ∆`-set which (up to
maximality) characterise the ∆`-set.
Lemma 2.5. Let N be a von Neumann factor. Then every ∆`-set φ satis-
fies the following properties:
(1) If u, v, w P φ then uvw " wuv P φ.
(2) If u, v P φ then there exist a unique w P φ such that wvw " u.
(3) If u P InvpN q, then u P N pφq iff wuw " u for some w P φ.
(4) If t P N is an elements of φ2 then t is of class 1 or class 2.
Proof. (4): Fix t P φ2. Select u, v P φ such that t " uv. Since elements of a
∆`-set have the same I `-set, I `puq " I `pvq. For u " 2e´ 1 and v " 2f ´ 1,
eN " I `puq " I `pvq " f N by Lemma 2.4(1). If follows that ef " f and
f e " e, see Lemma 2.4(4). Hence uv " p2e´1qp2f ´1q " 4ef ´2f ´2e`1 "
2pf ´ eq ` 1 and pt ´ 1q2 " 4pf ´ eq2 " 0. So t is of class 1 or class 2.
(1): Using u, v, w P φ, I `puq " I `pvq " I `pwq. Similarly to the proof of
(4), uvw " p2e ´ 1qp2f ´ 1qp2g ´ 1q " 2pe ´ f ` gq ´ 1 " u ´ v ` w for suitable
e, f, g P IpN q (using eN " f N ñ ef " f , etc.). This implies uvw " wvu, so
uvw P InvpN q. Now I `puvwq " pe´f `gqN " pe´ef `egqN Ď eN " I `puq
by Lemma 2.4, and conversely if x P I `puq " I `pvq " I `pwq, then uvwx " x
and x P I `puvwq. Since u P φ and I `puq " I `puvwq, also uvw P φ.
(2): Fix any u, v P φ. As in the proof of (4), uv " 2pf ´ eq ` 1. By
symmetry vu " 2pe ´ f q ` 1, so uv ` vu " 2 and uvu ` v " 2u (using
u2 " 1). For w def" u`v
2 we get 4wvw " puv ` 1qpu ` vq " uvu ` 2u ` v " 4u
(using uvu ` v " 2u), so wvw " u. Similarly 4w2 " pu ` vqpu ` vq " 4
(using uv ` vu " 2), so w P InvpN q. Since 2w " u ` v " 2pe ` f q ´ 2 it
follows that I `pwq " pe ` f qN " pe ` ef qN Ď eN " I `puq. Conversely if
x P I `puq " I `pvq, then pu ` vqx " 2x and x P I `pwq. Therefore w P φ.
8
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
2
For the uniqueness take u, w, v P φ such that wvw " u. As in (1) we get
wvw " w ´ v ` w. Hence u " w ´ v ` w and w " u`v
2 .
(3): Fix u P InvpN q. If u P N pφq, then uw " w‹u for some w, w‹ P φ,
so uwu " w‹. Using (2), w:ww: " w‹ for some w: P φ. As in (1) we
get w:ww: " 2w: ´ w. Hence uwu " w‹ " w:ww: " 2w: ´ w, so w: "
uwu`w
commutes with u. Conversely, suppose vu " uv for some v P φ. To
show uφ " φu, fix any w P φ. Since v and w belong to the same ∆`-set
I `pvq " I `pwq. For w " 2e ´ 1, v " 2f ´ 1, eN " I `pwq " I `pvq " f N by
Lemma 2.4. We have w‹ def" 2pueuq ´ 1 P InvpN q using ueu P IpN q. Since
uv " vu we get uf " f u and I `pw‹q " pueuqN " ueN " uf N " f N by
Lemma 2.4. Now v P φ and I `pw‹q " I `pvq giving that w‹ P φ. Using
uwu " w‹ we get uw " w‹u. It follows that uφ " φu, i.e., u P N pφq.
(cid:3)
The following lemmata relies heavily on the properties listed in Lemma 2.3.
In [3] similar results were established using dimension theory, irreducibility,
regularity and lattice properties of continuous rings, see [3, Proposition 3
and Proposition 8].
Lemma 2.6. Let N be a von Neumann factor. Suppose e P IpN q, B Ď N e
and A def" tx P eN : bx " 0 for all b P Bu satisfy A ‰ t0u and vA " A for all
v P InvpeN eq. If e " e, then A " eN .
Proof. Using Lemma 2.3(5) set q def" qB P N . The proof consist of five steps.
Step 1. "A " tx P N : x " ex " p1 ´ qqxu": Fix x P N such that
x " ex " p1´qqx. Using Lemma 2.3(2)-(3), px ď 1´q. Using Lemma 2.3(4),
b " bqb. Using Lemma 2.3(3), x " pxx. Using Lemma 2.3(5), qb ď q for
b P B. So bx " bqbqp1 ´ qqpxx " 0. Hence x P eN and bx " 0 for all b P B,
i.e., x P A. Conversely, fix x P eN such that bx " 0 for all b P B. By
Lemma 2.3(6), qbpx " 0 for b P B, so px ď 1 ´ qb. Then qb ď 1 ´ px for
b P B. Using Lemma 2.3(5), q ď 1 ´ px, so p1 ´ qqpx " px. It follows that
x " pxx " p1 ´ qqpxx " p1 ´ qqx and x " xe.
Assume q ‰ 0. Using that A ‰ t0u select a nonzero z P A. Using
Lemma 2.3(4), qz ‰ 0. Using Lemma 2.3(7), there exists a nonzero element
y P qzN q.
Step 2. "y P eN e": For each b P B Ď N e we have b " be. We now
have that bp1 ´ eq " 0, so qbp1´e " 0 by Lemma 2.3(6).
It follows that
qb ď 1 ´ p1´e, q ď 1 ´ p1´e, qp1´e " 0, qp1´ep1 ´ eq " 0, qp1 ´ eq " 0, so
q " qe. Consequently, y P qzN q " N qe Ď N e. Since z P A Ď eN and
e " e we have z " pezq " ze, so zp1 ´ eq " 0. Using Lemma 2.3(6),
qzp1 ´ eq " 0, so p1 ´ eqqz " 0 and eqz " qz. So y P qzN q Ď eN . We
conclude y P eN e.
Step 3. "py K qy": Using Lemma 2.3(3) on y " qzy (y P qzN q), py ď qz.
Using Lemma 2.3(4) on y " yq (y P qzN q), qy ď q. Since z " p1 ´ qqz we
have qz " 0 and zq " 0. Using Lemma 2.3(6), qzq " 0, so qz ď 1 ´ q.
Hence py ď qz ď 1 ´ q and qy ď q. We conclude py K qy (i.e., pyqy " 0).
ON THE GEOMETRY OF IDEMPOTENTS IN VON NEUMANN ALGEBRAS
9
By Lemma 2.3(8), there exists u " uuu P N such that py " uu and
qy " uu. Set
v def" epu ` u ` 1 ´ py ´ qyqe.
Step 4. "v2 " e": First we show u, u, py, qy P eN e. Using u " uuu,
py " uu, py ď qz and eqz " qz we have u " eqzpyu " eu. Using
u " uuu, qy " uu, qy ď q and qe " q we have u " uqyqe " ue. So
eue " u. Using e " e, eue " u. We have eqye " qy using that qy ď q ď e
(from q " qe and y " yq). We have epye " py using that py ď qz ď e
(from y " qzy and eqz " qz). So v " u ` u ` e ´ py ´ qy. It is clear
that pu ` uq is orthogonal to e ´ ppy ` qyq because pu ` uqe " pu ` uq and
pu ` uqppy ` qyq " uqy ` upy " pu ` uq using that py K qy and u " pyuqy.
Similary we deduce uu " uqypyu " 0, so u2 " 0, puq2 " 0. We conclude
v2 " pu ` uq2 ` pe ´ ppy ` qyqq2 " uu ` uu ` e ´ ppy ` qyq " e.
Step 5. "y " 0": By assumption applied to v P InvpeN eq we get vA " A.
Hence vz P A so vz " p1 ´ qqvz giving qvz " 0. Hence uqvz " 0. Using
u " uqy and qy ď q, uq " puqyqq " upqyqq " uqy " u, so uvz " 0. Using
u " uqy " ue and v " pyu`qyu`e´py ´qy we get uv " 0`uu`u´0´u "
py. So pyz " 0 and hence zpy " 0. Using Lemma 2.3(6), qzpy " 0. But
py ď qz, so py " 0. We conclude y " pyy " 0y " 0.
But this contradicts the fact that y ‰ 0. We conclude q " 0, A " eN . (cid:3)
Lemma 2.7. Let N be a von Neumann factor. Suppose e P IpN q, B Ď N e
and A def" tx P eN : bx " 0 for all b P Bu satisfy A ‰ t0u and vA " A for all
v P InvpeN eq. Then A " eN .
Proof. By Lemma 2.3(9) any idempotent is similar to a projection. Hence
we can find u P GLpN q such that p " ueu´1 P P pN q. Set
(2.1)
B1 def" uBu´1.
Using B Ď Be we get B1 " uBu´1 Ď uBeu´1 " uBu´1ueu´1 " B1p. Set
(2.2)
A1 def" uAu´1.
Since A ‰ t0u we get A1 ‰ t0u. Since upeN qu´1 " ueN " puN " pN we
get
A1 " tuxu´1 P ueN u´1 : bx " 0 for all b P Bu
" tuxu´1 P pN : ubu´1uxu´1 " 0 for all b P Bu
" ty P pN : ubu´1y " 0 for all b P Bu
" ty P pN : ubu´1y " 0 for all ubu´1 P uBu´1u
" ty P pN : b1y " 0 for all b1 P B1u
We claim that wA1 " A1 for all w P InvppN pq. To see this fix w P InvppN pq.
Notice that InvppN pq " uInvpeN equ´1 because if v P InvpeN eq, then
uvu´1 " ueveu´1 " pueu´1quvu´1pueu´1q P InvppN pq,
10
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
and similarly for the converse containment. Hence w " uvu´1 for some
v P InvpeN eq. By assumption we know that vA " A.
It follows that
wA1 " uvu´1uAu´1 " A1.
Having that p P IpN q, B1 Ď N p and A1 " ty P pN : b1y " 0 for all b1 P B1u
satisfy wA1 " A1 ‰ t0u for all w P InvppN pq, and p " p, we can now apply
Lemma 2.6 to deduce A1 " pN . It now follows that
A " u´1uAu´1u " u´1pN u " u´1pueu´1qN " eN.
(cid:3)
Lemma 2.8. Let N be a von Neumann factor. Suppose u P InvpN q, B Ď N
and A def" tx P N : bx " 0 for all b P Bu satisfy vA " A for all v P
Cpuq X InvpN q. Then A P tt0u, I `puq, I ´puq, N u.
Proof. Define A def" A X I puq and e def" 1`u
eN " I `puq.
2 . By Lemma 2.4, e P IpN q and
Fix any v P InvpeN eq. Define w def" v ` p1 ´ eq. Since v2 " e and v " eve
we get w2 " v2 ` p1 ´ eq2 " 1, so w P InvpN q. Also, using ew " we, we get
uw " wu (recall u " 2e ´ 1), so w P Cpuq X InvpN q. Hence, by assumption
wA " A. Notice, if x P eN then x " ex, so wpeN X Aq " wepeN X Aq. Now
using ew " v " we, we get
vA` " vpeN X Aq " wepeN X Aq " wpeN X Aq " weN X wA
" ewN X A " eN X A " A`.
Recalling A " tx P N : bx " 0 for all b P Bu and A` " eN X A it follows
that A` " tx P eN : bx " 0 for all b P Bu. Moreover, we can replace B by
Be without enlarging A`: indeed if x P eN satisfies bx " 0 for all b P Be,
then every b1 P B satisfies b1x " b1pexq " pb1eqx " 0. Consequently,
A` " tx P eN : bx " 0 for all b P Beu.
Assuming A` ‰ t0u (and using that vA` " A` for each v P InvpeN eq) we
can now apply Lemma 2.7 to A`, Be in place of A, B to deduce A` " eN .
Including the trivial case we have A` P tt0u, I `puqu.
In a similar fashion (using I ´puq " p1 ´ eqN , see Lemma 2.4) we can
prove that vA´ " A´ for each v P Invpp1 ´ eqN p1 ´ eqq and use Lemma 2.7
on A´, Bp1 ´ eq to deduce A´ P tt0u, I ´puqu.
By Lemma 2.4, I `puq`I ´puq " N and I `puqXI ´puq " t0u. Hence A` `
A´ " A and A` X A´ " t0u. If follows that A P tt0u, I `puq, I ´puq, N u. (cid:3)
Lemma 2.9. Let N be a von Neumann factor containing a subset φ of
involutions satisfying (1)-(4) of Lemma 2.5 and φ ą 1. Define A def" I `pφ2q.
Then A " eN for some e P IpN q and φ Ď ∆`peq or φ Ď ∆´peq.
Proof. The proof consist of five steps.
Step 1. "A ‰ N ": If A " N then 1 P N " tx P N : tx " x for all t P φ2u,
so t1 " 1 for each t P φ2. But φ2 " t1u implies φ ď 1 (if uv " 1 ñ u " v).
Step 2. "A ‰ t0u": Clearly 1 P φ2 using u P φ ñ 1 " u2 P φ2. Also
φ2 ‰ t1u, see Step 1. One can therefore select t0 P φ2 such that t0 ‰ 1. By
ON THE GEOMETRY OF IDEMPOTENTS IN VON NEUMANN ALGEBRAS
11
property Lemma 2.5(4), t0 " 1 ` n0 is of class 2, so n0 ‰ 0 and n2
0 " 0.
Select any t " 1 ` n P φ2. Then t " uv, t0 " u0v0 for some u, v, u0, v0 P φ.
By property Lemma 2.5(1), tt0 " uvu0v0 " uv0u0v " u0v0uv " t0t P φ2.
Now property Lemma 2.5(4) ensures tt0 is of class 1 or 2, so ptt0 ´ 1q2 " 0.
Using n, n0 commute (recall tt0 " t0t, t " 1 ` n, t0 " 1 ` n0) gives
0 " ptt0 ´ 1q2 " ppn ` 1qpn0 ` 1q ´ 1q2 " pn ` n0 ` nn0q2 " 2nn0.
We deduce that tn0 " p1 ` nqn0 " n0 for all t P φ2. Consequently, n0 is an
element of tx P N : tx " x for all t P φ2u " A.
Step 3. "sA " A for all s P N pφq": Using sφ " φs we get
sA " tsx P N : tx " x for all t P φ2u
" tsx P N : uvs2x " s2x for all u, v P φu
" tsx P N : su1v1sx " s2x for all u1, v1 P φu
" tsx P N : u1v1sx " sx for all u1, v1 P φu " A.
Step 4. "A " eN for some e P IpN q": Fix any u def" 2f ´ 1 P φ.
If
s P Cpuq X InvpN q then we know s P InvpN q and wsw " s for some w P φ.
By property Lemma 2.5(3), s P N pφq. Hence sA " A by Step 3. Define
B def" tb : 1 ` b P φ2u. We get A " tx P N : bx " 0 for all b P Bu.
Applying Lemma 2.8 to u, B, A (recall sA " A for all s P Cpuq X InvpN q)
we get A P tt0u, I `puq, I ´puq, N u. By Lemma 2.4, A " I `puq " f N or
A " I ´puq " p1 ´ f qN .
Step 5. "φ Ď ∆`peq or φ Ď ∆´peq". Suppose A " I `puq " f N with
u " 2f ´ 1 from Step 4. Take any v P φ. By property Lemma 2.5(2) there
exsits w P φ such that wvw " u. Since w P φ, property Lemma 2.5(3) gives
w P N pφq (because w‹ww‹ " w for some w‹ P φ). By Step 3, wA " A, so
I `pvq " twx P N : vwx " wxu " twx P N : wux " wxu
" twx P N : ux " xu " wI `puq " wA " A.
It follows that v P ∆`pf q from I `pvq " f N . Hence φ Ď ∆`pf q. Similarly, if
A " I ´puq " p1´f qN with u " 2f ´1 from Step 4, then φ Ď ∆´p1´f q. (cid:3)
Lemma 2.10. Let N be a von Neumann factor containing a nonzero idem-
potent e. Set A def" eN and φ def" ∆`peq. Then A " I `pφ2q.
Proof. If e " 1 then A " N , φ " ∆`p1q " t2f ´ 1 : f N " N u " t1u and
I `pφ2q " I `pt1uq " tx P N : tx " x for all t P t1uu " N , so A " I `pφ2q.
We may assume 1 ´ e ‰ 0. Using property Lemma 2.3(7) select x P N such
that exp1 ´ eq ‰ 0. Then f " e ` exp1 ´ eq P IpN q and eN " f N , see
Lemma 2.4(4). Since f ‰ e, 2f ´ 1 and 2e ´ 1 are distinct elements of φ.
Since φ is a ∆`-set it satisfies properties (1)-(4) of Lemma 2.5. Define
A1 def" I `pφ2q. By Lemma 2.9, A1 " f N for some f P IpN q and φ Ď ∆`pf q
or φ Ď ∆´pf q.
Suppose φ Ď ∆´pf q. Select distinct u, v P φ. Using u, v P φ Ď ∆´pf q
we get I ´puq " I ´pvq " f N . Using u, v P φ " ∆`peq we get I `puq "
12
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
I `pvq " eN . By Lemma 2.4(3), u " v giving a contradiction. Consequently
φ Ď ∆`pf q. Select u P φ. Using u P φ Ď ∆`pf q we get I `puq " f N . Using
u P φ " ∆`peq we get I `puq " eN . We conclude
A " eN " I `puq " f N " A1 " I `pφ2q. (cid:3)
Lemma 2.11. Let N be a von Neumann factor. Then every ∆`-set is
maximal among nonempty sets φ Ď InvpN q satisfying (1)-(4) of Lemma 2.5.
Proof. Fix φ def" ∆`peq for e P IpN q. Then φ satisfies (1)-(4) of Lemma 2.5.
To verify maximality of φ select any φ1 Ď InvpN q satisfying (1)-(4) (with
φ1 in place of φ) of Lemma 2.5 and contaning φ. Assume φ1 ‰ φ. We derive
a contradiction.
Since φ is nonempty and φ Ĺ φ1 we get φ1 ą 1. Define A1 def" I `pφ12q. By
Lemma 2.9, A1 " f N for some f P IpN q and φ1 Ď ∆`pf q or φ1 Ď ∆´pf q.
Suppose φ1 Ď ∆`pf q. Select u P φ. Using u P φ Ď φ1 Ď ∆`pf q we get
I `puq " f N . Using u P φ " ∆`peq we get I `puq " eN , so f N " eN and
φ1 Ď ∆`pf q " ∆`peq " φ Ĺ φ1. Consequently φ1 Ď ∆´pf q. Select u P φ.
Using u P φ Ď φ1 Ď ∆´pf q we get I ´puq " f N . Using u P φ " ∆`peq we
get I `puq " eN . We consider two cases:
Case 1. Suppose e ‰ 0: By Lemma 2.10, eN " I `pφ2q. We also have
I `pφ12q Ď I `pφ2q because if x P I `pφ12q then tx " x for all t P φ12, hence
tx " x for all t P φ2 (Ď φ12), so x P I `pφ2q. It follows that
I ´puq " f N " A1 " I `pφ12q Ď I `pφ2q " eN " I `puq.
By Lemma 2.4(2), I ´puq " t0u, so f " 0 and φ1 Ď ∆´pf q " ∆´p0q " t1u.
But this contradicts φ1 ą 1.
Case 2. Suppose e " 0: Then I `puq " eN " 0 and, using Lemma 2.4(2),
f N " I ´puq " N , so f " 1 and φ1 Ď ∆´pf q " ∆´p1q " t´1u. This also
contradicts φ1 ą 1.
We deduce that φ1 " φ, so φ is maximal.
(cid:3)
Proposition 2.12. Let N be a von Neumann factor. Then every ∆-set is
maximal among nonempty sets φ Ď InvpN q satisfying (1)-(4) of Lemma 2.5.
Proof. Lemma 2.5 and Lemma 2.11 provides the desired result for ∆`-sets,
so we only need to consider ∆´-sets. Notice that any ∆´-set φ satisfies the
properties (1)-(4) of Lemma 2.5 because the properties are independent of
the sign of φ.
Maximality of a ∆´-set, say φ def" ∆´peq, can be verified by modifying the
proof of Lemma 2.11: The proof is unchanged until reaching A1 " f N for
f P IpN q and φ1 Ď ∆`pf q or φ1 Ď ∆´pf q. Then, if u P φ Ď φ1 Ď ∆´pf q one
shows f N " I ´puq " eN and φ1 Ĺ φ1. Moreover, if u P φ Ď φ1 Ď ∆`pf q, one
shows that f N " I `puq Ď I ´puq and f " 0 (in Case 1) or I ´puq " eN " 0,
I `puq " f N and f " 1 (in Case 2). This implies φ1 ď 1, but φ1 ą 1. (cid:3)
Proof of Theorem B: Let φ be a nonempty subset of InvpN q. If φ is a
∆`-set or a ∆´-set, then φ is a maximal set among the nonempty subset of
involutions in N satisfying (1)-(4) of Lemma 2.5, see Proposition 2.12.
ON THE GEOMETRY OF IDEMPOTENTS IN VON NEUMANN ALGEBRAS
13
Conversely, assume φ is a maximal set among the nonempty subset of
involutions in N satisfying (1)-(4) of Lemma 2.5. We show φ is a ∆-set.
Suppose φ ą 1. Define A def" I `pφ2q. By Lemma 2.9, A " f N for some
f P IpN q and φ Ď ∆`pf q or φ Ď ∆´pf q. By maximality of φ we get that
φ " ∆`pf q or φ " ∆´pf q, so φ is a ∆-set.
Suppose φ " 1.
If φ " t1u or φ " t´1u, then φ is a ∆-set because
∆`p1q " t1u and ∆´p1q " t´1u. We claim no other option is possible.
To see this assume φ " tuu for u def" 2e ´ 1 such that u ‰ 1 and u ‰ ´1.
Using that I `puq " eN (Lemma 2.4(1)) we get that φ " tuu Ď ∆`peq. By
maximality of φ, tuu " φ " ∆`peq. Since u differs from 1, ´1, both e and
1 ´ e are nonzero. Using property Lemma 2.3(7) select x P N such that
exp1 ´ eq ‰ 0. If follows that f " e ` exp1 ´ eq P IpN q and f N " eN , see
Lemma 2.4(4). Therefore u ‰ 2f ´ 1 P ∆`peq " tuu. Contradiction.
(cid:3)
Remark 2.13. It is unclear to which extend the characterisation in Theo-
rem B remains valid for unital rings N which are neither von Neumann
factors nor regular rings. We suspect that this should be true for any uni-
tal rings N satisfying properties (1)-(7) of Lemma 1.2 and the following
additional property
(8) Suppose e P IpN q, B Ď N e and A def" tx P eN : bx " 0 for all b P Bu
satisfy A ‰ t0u and vA " A for all v P InvpeN eq. Then A " eN .
3. Idempotents and the proof of Theorem D.
In this section we prove Theorem D. First we establish some notation.
Given two von Neumann factors N and M , and a group isomorphism ϕ : GLpN q Ñ
GLpM q between their general linear groups, the formula
1 ´ 2θpeq " ϕp1 ´ 2eq,
e P IpN q,
induces a bijection θ : IpN q Ñ IpBq between the set of idempotents of N
and M . To simplify notation we record the following:
Notation 3.1. (i) The quadruple pN, M, ϕ, θq will denote a pair of von
Neumann factors N and M , a group isomorphism ϕ : GLpN q Ñ GLpM q,
and the induced bijection θ : IpN q Ñ IpM q where 1 ´ 2θpeq " ϕp1 ´ 2eq.
(ii) With pN, M, ϕ, θq as above and e P IpN q we write ∆`peq Ñ ∆`
(resp. ∆`peq Ñ ∆´) to indicate that ϕ maps the ∆`-set ∆`peq in GLpN q
into a ∆`-set (resp. ∆´-set) in GLpM q.
(iii) We implicitly consider the quadruple pM, N, ϕ´1, θ´1q. We will also
write ∆`pf q Ñ ∆` and ∆`pf q Ñ ∆´ for f P IpM q to indicate what ϕ´1
does to the ∆`-set ∆`pf q.
Remarks 3.2. (i) Let pN, M, ϕ, θq be as in (3.1). By Theorem B the bijection
ϕ maps each ∆-set into a ∆-set. In particular for e P IpN q
∆`peq Ñ ∆` or ∆`peq Ñ ∆´.
14
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
(ii) No ∆-set except for t1u and t´1u is both a ∆`-set and ∆´-set: To see
this assume
φ " ∆`peq " ∆´pf q,
for some e, f P IpN q. For u def" 2e ´ 1 and v def" 2f ´ 1 we have that u, ´v P φ,
so I ´puq " f N and I `p´vq " eN . Using Lemma 2.4(1), I ´p´vq " f N
and I `puq " eN . Hence u " ´v and e " 1 ´ f (see Lemma 2.4(3)). We
have ∆`peq X ∆´p1 ´ eq " tuu (because if w def" 2g ´ 1 P ∆`peq X ∆´p1 ´ eq
then I pwq " I puq, so w " u), so φ " tuu. The proof of Theorem B (last
paragraph) gives that u is 1 or ´1, as claimed. In particular for nontrivial
e P IpN q
either ∆`peq Ñ ∆` or ∆`peq Ñ ∆´.
(iii) pN, M, ϕ, θq be as in (3.1). Then ϕp´1q " ´1, see [4].
We now establish what is the image of each ∆`-set via the bijection ϕ.
Using ϕp´1q " ´1 this also gives the image of each ∆´-set.
Lemma 3.3. Let pN, M, ϕ, θq be as in (3.1). Then for each e P IpN q
ϕp∆`peqq "
∆`pθpeqq,
∆´pθp1 ´ eqq,
∆`p1q,
∆`p0q,
if e ‰ 0, 1 and ∆`peq Ñ ∆`
if e ‰ 0, 1 and ∆`peq Ñ ∆´
if e " 1
if e " 0
$''&
''%
Proof. Fix any e P IpN q. Since ϕp´1q " ´1, ϕp∆`p1qq " ϕp1q " 1 " ∆`p1q
and ϕp∆`p0qq " ϕp´1q " ´1 " ∆`p0q. We may therefore assume that e is
nontrivial. By Remark 3.2(ii) either ∆`peq Ñ ∆` or ∆`peq Ñ ∆´.
Suppose ∆`peq Ñ ∆`. Select g P IpM q such that ϕp∆`peqq " ∆`pgq. By
Lemma 2.4(1), I `p2e´1q " eN , so 2e´1 P ∆`peq. Hence ϕp2e´1q P ∆`pgq.
It follows that 2θpeq ´ 1 P ∆`pgq, so θpeqM " I `p2θpeq ´ 1q " gM . We
deduce ϕp∆`peqq " ∆`pgq " tv P InvpM q : I `pvq " gM u " ∆`pθpeqq.
Suppose ∆`peq Ñ ∆´. Select g P IpM q such that ϕp∆`peqq " ∆´pgq.
By Lemma 2.4(1), 2θpeq ´ 1 P ∆´pgq, so θp1 ´ eqM " I ´p2θpeq ´ 1q " gM .
So ϕp∆`peqq " ∆´pgq " tv P InvpM q : I ´pvq " gM u " ∆´pθp1 ´ eqq.
(cid:3)
Lemma 3.4. Let pN, M, ϕ, θq be as in (3.1). Let e P N be a nontrivial
idempotent and f def" θpeq its image via θ. Then the following holds:
(1) We have ∆`peq X ∆´p1 ´ eq " t2e ´ 1u.
(2) If ∆`peq Ñ ∆`, then ∆`pf q Ñ ∆`.
(3) If ∆`peq Ñ ∆´, then ∆`p1 ´ eq, ∆`p1 ´ f q, ∆`pf q Ñ ∆´.
Proof. "(1)": Fix any w def" 2g ´ 1 P InvpN q. Set u def" 2e ´ 1. Using
Lemma 2.4(1) we obtain I `puq " eN and I ´puq " p1 ´ eqN . Now take
any w P ∆`peq X ∆´p1 ´ eq. Then I `pwq " eN and I ´puq " p1 ´ eqN ,
so I puq " I pwq. By Lemma 2.4(3), w " u, so g " e. We conclude
∆`peq X ∆´p1 ´ eq " tuu.
"(2)": Suppose ∆`peq Ñ ∆`. By Lemma 3.3, ϕp∆`peqq " ∆`pθpeqq.
Hence ϕ´1p∆`pf qq " ϕ´1p∆`pθpeqq " ∆`peq, so ∆`pf q Ñ ∆`.
ON THE GEOMETRY OF IDEMPOTENTS IN VON NEUMANN ALGEBRAS
15
"(3)": Suppose ∆`peq Ñ ∆´. Assume that ∆`p1 ´ eq Ñ ∆`, we derive
a contradiction. By Lemma 3.3, we have ϕp∆`peqq " ∆´pθp1 ´ eqq and
ϕp∆`p1 ´ eqq " ∆`pθp1 ´ eqq. Using ϕp´1q " ´1 if follows that
ϕp∆`peqq " ϕp´1qϕp∆`p1 ´ eqq " ϕp∆´p1 ´ eqq.
By (1) we get t2e ´ 1u Ď ∆`peq Ď ∆`peq X ∆´p1 ´ eq " t2e ´ 1u. But we
know ∆`peq ą 1 for nontrivial e (see Lemma 2.3(7) and Lemma 2.4(4)).
Contradiction. Therefore ∆`p1 ´ eq Ñ ∆´, cf. Remark 3.2(ii).
Since ∆`peq Ñ ∆´, we get that ∆´peq Ñ ∆` and ∆´p1 ´ eq Ñ ∆` (by
the preceding paragraph and by ϕp´1q " ´1). Using that ∆´peq Ñ ∆`,
ϕp∆´peqq " ∆`pθp1 ´ eqq " ∆`p1 ´ f q, so ∆`p1 ´ f q Ñ ∆´. Using
that ∆´p1 ´ eq Ñ ∆`, ϕp∆´p1 ´ eqq " ∆`pθp1 ´ p1 ´ eqqq " ∆`pf q, so
∆`pf q Ñ ∆´.
(cid:3)
Recall we write e «N f (or e « f ) whenever e, f P IpN q and eN " f N .
This is an equivalence relation on IpN q. Let res denote the equivalence class
containing e and IpN q{« the set of equivalence classes, i.e.,
res def" tf P IpN q : f « eu,
IpN q{« def" tres : e P IpN qu.
Proof of Theorem D: Let pN, M, ϕ, θq be as in (3.1). Define the map
θ : IpN q{« Ñ IpM q{« as follows:
θpresq "
$''&
''%
rθpeqs,
rθp1 ´ eqs,
r1s,
r0s,
if e ‰ 0, 1 and ∆`peq Ñ ∆`
if e ‰ 0, 1 and ∆`peq Ñ ∆´
if e " 1
if e " 0
We show θ is well defined. Fix any equivalence class φ of idempotents
generating the same right ideal.
Suppose φ " 1. By Lemma 2.3(7) and Lemma 2.4(4), φ " r1s or φ " r0s,
so there is nothing to prove (as φ has precisely one representative).
Suppose φ ą 1. Select any e, f P φ. We have res " rf s. Recall that (by
definition) res " rf s iff ∆`peq " ∆`pf q. Suppose ∆`peq Ñ ∆`. Lemma 3.3
ensures that ∆`pθpeqq " ∆`pθpf qq, so rθpeqs " rθpf qs. Consequently,
θpresq " rθpeqs " rθpf qs " θprf sq
Suppose ∆`peq Ñ ∆´. By Lemma 3.3, rθp1 ´ eqs " rθp1 ´ f qs, hence
θpresq " rθp1 ´ eqs " rθp1 ´ f qs " θprf sq. So the map θ is well-defined.
We show θ is a bijection. Define the map ψ : IpM q{« Ñ IpN q{« by
ψprf sq "
rθ´1pf qs,
rθ´1p1 ´ f qs,
r1s,
r0s,
if f ‰ 0, 1 and ∆`pf q Ñ ∆`
if f ‰ 0, 1 and ∆`pf q Ñ ∆´
if f " 1
if f " 0
$''&
''%
The map is well-defined by arguments analogues to those establishing well-
definiteness of θ. We show ψ is the inverse for θ. Fix a nontrivial idempotent
e P N and set f def" θpeq.
16
THIERRY GIORDANO AND ADAM SIERAKOWSKI1
Suppose ∆`peq Ñ ∆`. Then ∆`pf q Ñ ∆` (by Lemma 3.4). By defini-
tion of θ and ψ, θpresq " rθpeqs (so θpresq " rf s) and ψprf sq " rθ´1pf qs (so
ψprf sq " res). We conclude
ψ θpresq " ψprf sq " res,
θ ψprf sq " θpresq " rf s.
Suppose ∆`peq Ñ ∆´. Using Lemma 3.4, ∆`p1 ´ f q Ñ ∆´. We get that
θpresq " rθp1 ´ eqs (so θpresq " r1 ´ f s) and ψpr1 ´ f sq " rθ´1p1 ´ p1 ´ f qqs
(so ψpr1 ´ f sq " res). Consequently
ψ θpresq " ψpr1 ´ f sq " res.
Knowing that ∆`peq Ñ ∆´, we also get ∆`p1 ´ eq Ñ ∆´ and ∆`pf q Ñ ∆´
(by Lemma 3.4). Hence we both have that θpr1 ´ esq " rθp1 ´ p1 ´ eqqs (so
θpr1 ´ esq " rf s) and ψprf sq " rθ´1p1 ´ f qs (so ψprf sq " r1 ´ es). Therefore
θ ψprf sq " θpr1 ´ esq " rf s.
We conclude that θ has an inverse, hence it is a bijection.
(cid:3)
Acknowledgements
Part of this research was conducted while the authors were participat-
ing in the research program Classification of operator algebras: complexity,
rigidity, and dynamics at the Mittag-Leffler Institute, January -- April 2016,
and the Intensive Research Program Operator algebras: dynamics and in-
teractions at the Centre de Recerca Matem`atica, March -- July 2017. This
research was supported by EIS Distinguished Visitors Program, by Aus-
tralian Research Council grant DP150101598 and by a grant from NSERC
Canada.
References
[1] Reinhold Baer, Linear algebra and projective geometry, Academic Press Inc., New
York, N. Y., 1952. MR 0052795
[2] H. A. Dye, On the geometry of projections in certain operator algebras, Ann. of Math.
(2) 61 (1955), 73 -- 89. MR 0066568 (16,598a)
[3] Gertrude Ehrlich, Characterization of a continuous geometry within the unit group,
Trans. Amer. Math. Soc. 83 (1956), 397 -- 416. MR 0081885
[4] Thierry Giordano and Adam Sierakowski, The general
linear group as a complete
invariant for C -algebras, J. Operator Theory 76 (2016), no. 2, 249 -- 269. MR 3552377
[5] John von Neumann, Continuous geometry, Foreword by Israel Halperin. Princeton
Mathematical Series, No. 25, Princeton University Press, Princeton, N.J., 1960.
MR 0120174
Department of Mathematics and Statistics, University of Ottawa, 585 King
Edward Avenue, Ottawa, Ontario, KiN 6N5, Canada
E-mail address: [email protected]
School of Mathematics and Applied Statistics, Building 39C, University of
Wollongong, Wollongong NSW, 2522, Australia
E-mail address: [email protected]
|
1511.00939 | 1 | 1511 | 2015-11-03T15:13:51 | Partial actions and subshifts | [
"math.OA",
"math.DS"
] | Given a finite alphabet $\Lambda$, and a not necessarily finite type subshift $X\subseteq \Lambda^\infty$, we introduce a partial action of the free group $F(\Lambda)$ on a certain compactification $\Omega_X$ of $X$, which we call the spectral partial action.
The space $\Omega_X$ has already appeared in many papers in the subject, arising as the spectrum of a commutative C*-algebra usually denoted by ${\cal D}_X$. Since the descriptions given of $\Omega_X$ in the literature are often somewhat terse and obscure, one of our main goals is to present a sensible model for it which allows for a detailed study of its structure, as well as of the spectral partial action, from various points of view, including topological freeness and minimality.
We then apply our results to study certain C*-algebras associated to $X$, introduced by Matsumoto and Carlsen. Most of the results we prove are already well known, but our proofs are hoped to be more natural and more in line with mainstream techniques used to treat similar C*-algebras. The clearer understanding of $\Omega_X$ provided by our model in turn allows for a fine tuning of some of these results, including a necessary and sufficient condition for the minimality of the Carlsen-Matsumoto C*-algebra ${\cal O}_X$, generalizing a similar result of Thomsen. | math.OA | math |
PARTIAL ACTIONS AND SUBSHIFTS
M. Dokuchaev and R. Exel
Given a finite alphabet Λ, and a not necessarily finite type subshift X ⊆ Λ∞, we
introduce a partial action of the free group F(Λ) on a certain compactification ΩX of
X, which we call the spectral partial action.
The space ΩX has already appeared in many papers in the subject, arising as the
spectrum of a commutative C*-algebra usually denoted by DX . Since the descriptions
given of ΩX in the literature are often somewhat terse and obscure, one of our main goals
is to present a sensible model for it which allows for a detailed study of its structure, as
well as of the spectral partial action, from various points of view, including topological
freeness and minimality.
We then apply our results to study certain C*-algebras associated to X, introduced
by Matsumoto and Carlsen. Most of the results we prove are already well known, but
our proofs are hoped to be more natural and more in line with mainstream techniques
used to treat similar C*-algebras. The clearer understanding of ΩX provided by our
model in turn allows for a fine tuning of some of these results, including a necessary
and sufficient condition for the minimality of the Carlsen-Matsumoto C*-algebra OX ,
generalizing a similar result of Thomsen.
1. Introduction.
The theory of C*-algebras associated to subshifts has a long and exciting history, having
been initiated by Matsumoto in [24], later receiving invaluable contributions from many
other authors, notably Carlsen, Silvestrov, and Thomsen. Accounts of this history may be
found in [9] and [7], to which the interested reader is referred.
For now it suffices to say that, given a finite alphabet Λ, and a closed subset X ⊆ Λ
which is invariant under the left shift
S : Λ → Λ,
in which case one says that the pair (X, S) is a (one sided) subshift, Matsumoto initiated
a study of certain C*-algebras associated to X, whose algebraic properties reflect certain
important dynamical properties of the subshift itself, and whose K-theory groups provide
new invariants for subshifts.
The extensive literature in this field (see for instance the list of references in [8])
contains a lot of information about the structure of these algebras, such as faithful repre-
sentations, nuclearity, characterization of simplicity, computation of K-theory groups and
a lot more. It is therefore a perilous task to attempt to add anything else to the wealth of
results currently available, a task we hope to be undertaking in a responsible manner.
The motivation that brought us to revisit the theory of C*-algebras associated to
subshifts, and the justification for writing the present paper, is twofold: firstly we are
able to offer a sensible description of a certain topological space, which we will denote
by ΩX (for the cognoscenti, we are referring to the spectrum of the ill-fated commutative
algebra DX appearing in most papers on the subject), and which has evaded all attempts
at analysis, except maybe for some somewhat obscure projective limit descriptions given
in [25] and [6: Section 2.1]. See also [30].
2
m. dokuchaev and r. exel
Secondly we will introduce a partial action of the free group F = F(Λ) on ΩX , called
the spectral partial action, whose associated crossed product is the Carlsen-Matsumoto
C*-algebra OX . Given that description, we may recover many known results for OX , as
well as give the first necessary and sufficient condition for simplicity which applies for all
subshifts, including those where the shift map is not surjective.
Our study of the space ΩX is perhaps the single most important contribution we have
to offer. The method we adopt is essentially the same one used by the second named
author and M. Laca in the analysis of Cuntz-Krieger algebras for infinite matrices [17],
the crucial insight being the introduction of a partial action of the free group F. To be
more precise, for each letter a in the alphabet Λ, consider the subsets Fa and Za of X,
given by
Fa = {y ∈ X : ay ∈ X}, and
Za = {x ∈ X : x = ay, for some infinite word y}.
It is then evident that the assignment
θa : y ∈ Fa → ay ∈ Za
is a continuous bijective map. By iterating the θa and their inverses, we obtain partially
defined continuous bijective maps on X, thus forming what is known as a partial action of
the group F on X.
Partial actions on topological spaces are often required to map open sets to open sets,
but except for the case of shifts of finite type, the above partial action does not satisfy this
requirement since the Fa may fail to be open. One may nevertheless use this badly behaved
partial action to build a bona fide partial representation on the (typically non-separable)
Hilbert space ℓ2(X). At this point it is evident that we have grossly violated the topology
of X, but alas a new commutative C*-algebra is born, generated by the range projections
of all partial isometries in our partial representation, and with it a new topological space
is also born, namely the spectrum of said algebra.
This algebra, usually denoted DX , turns out to be as well known for experts in the
field, as it is dreaded. Nevertheless it plays a well known role in the theory of partial repre-
sentations and, like many other commutative algebras arising from partial representations,
its spectrum ΩX may be seen as a subspace of the Bernoulli space 2F, which is moreover
invariant under the well known Bernoulli partial action [16: 5.10].
Based on the algebraic relations possessed by this representation we may deduce cer-
tain special properties enjoyed by the elements of 2F which lie in ΩX , and if we see 2F as
the set of all subsets of the Cayley graph of F, such properties imply that these elements
must have the aspect of a river basin, in the sense that there is a main river consisting
of an infinite (positive) word, together with its tributaries. Not only must the main river
partial actions and subshifts
3
...........
...........
...........
..◦
..◦
................................................
................................................
.....◦..◦..◦
..◦
..◦ ..◦
.....◦..◦
..◦
...............• ...............•
...............• ...............•
...............• ...............•
...............• ...............•
•..◦
•..◦
•..◦
•..◦
...............................
.......................
...............................
.......................
.................
.................
.........
.........
•.....◦..◦..◦
•.....◦..◦..◦
..............................................
..............................................
..◦
..◦
...........
...........
.....................
.....................
•............◦.....◦..◦..◦
......................................................................................
..◦ ..◦
.....◦..◦
..◦
......................................................................................................................................................................
...............• ...............•
...............• ...............•
...............• ...............•
...............• ...............•
•..◦
•..◦
•..◦
•..◦
...............................
.......................
...............................
.......................
.................
.................
.........
.........
•.....◦..◦..◦
•.....◦..◦..◦
..............................................
..............................................
..◦
..◦
...........
...........
.....................
.....................
•............◦.....◦..◦..◦
......................................................................................
..◦ ..◦
.....◦..◦
.....◦..◦..◦
..◦
•.........................◦............◦.....◦..◦..◦
..◦ ..◦
..◦ ..◦
.....◦..◦
.....◦..◦
............◦.....◦..◦..◦
..◦
............◦.....◦..◦..◦
..◦ ..◦
.....◦..◦
.....◦..◦..◦
.....◦..◦..◦
..◦
..◦
...............• ...............•
...............• ...............•
...............• ...............•
...............• ...............•
•..◦
•..◦
•..◦
•..◦
...............................
...............................
.......................
.......................
.................
.................
.........
.........
•.....◦..◦..◦
•.....◦..◦..◦
.............................................. ...........
..............................................
..◦
..◦
................................
.....................
•............◦.....◦..◦..◦
...................................................................................... ...........
..◦ ..◦
.................................................................................................................•...................................................................................... ...........
.....◦..◦
.....◦..◦..◦
..◦
.........................◦............◦.....◦..◦..◦
...............• ...............•
...............• ...............•
..◦ ..◦
..◦ ..◦
•..◦
•..◦
.....◦..◦
.....◦..◦
...............................
.......................
.................
.........
•.....◦..◦..◦
............◦.....◦..◦..◦
..............................................
..◦
..◦
................................................•..............................................
...........
.....................
............◦.....◦..◦..◦
............◦.....◦..◦..◦
...............• ...............•
..◦ ..◦
..◦ ..◦
•..◦
.....◦..◦
.....◦..◦
.......................
......................•.....◦..◦..◦
.................
.....◦..◦..◦
.....◦..◦..◦
..◦
...............•
..◦
..◦
.................•..◦...............•
......................................................................................................
......................................................................................................................................................................
..◦ ..◦
.....◦..◦
..◦
.....◦..◦..◦
..◦
................................................
..◦
..◦
..◦
.......................
...........
...........
...........
...........
..◦
................................................
................................................
......................................................................................................
...............• ...............•
...............• ...............•
...............• ...............•
...............• ...............•
•..◦
•..◦
•..◦
•..◦
...............................
.......................
...............................
.......................
.................
.................
.........
.........
•.....◦..◦..◦
•.....◦..◦..◦
..............................................
..............................................
..◦
..◦
...........
...........
.....................
.....................
•............◦.....◦..◦..◦
......................................................................................
..◦ ..◦
.....◦..◦
.....◦..◦..◦
..◦
......................................................................................................................................................................
...............• ...............•
...............• ...............•
...............• ...............•
...............• ...............•
•..◦
•..◦
•..◦
•..◦
...............................
.......................
...............................
.......................
.................
.................
.........
.........
•.....◦..◦..◦
•.....◦..◦..◦
..............................................
..............................................
..◦
..◦
...........
...........
.....................
.....................
•............◦.....◦..◦..◦
......................................................................................
..◦ ..◦
.....◦..◦
.....◦..◦..◦
..◦
..◦
•.........................◦............◦.....◦..◦..◦
..◦ ..◦
..◦ ..◦
.....◦..◦
.....◦..◦
............◦.....◦..◦..◦
..◦
............◦.....◦..◦..◦
..◦ ..◦
..◦ ..◦
.....◦..◦
.....◦..◦
.....◦..◦..◦
.....◦..◦..◦
..◦
..◦
..◦
•....................................................◦.........................◦............◦.....◦..◦..◦
..◦ ..◦
..◦ ..◦
..◦ ..◦
..◦ ..◦
.....◦..◦
.....◦..◦
.....◦..◦
.....◦..◦
............◦.....◦..◦..◦
............◦.....◦..◦..◦
..◦
..◦
.........................◦............◦.....◦..◦..◦
..◦ ..◦
.....◦..◦
.....◦..◦..◦
..◦
..◦
..◦ ..◦
..◦ ..◦
.....◦..◦
.....◦..◦
............◦.....◦..◦..◦
..◦
..◦ ..◦
.....◦..◦
..◦
.........................◦............◦.....◦..◦..◦
..◦ ..◦
..◦ ..◦
.....◦..◦
.....◦..◦
............◦.....◦..◦..◦
..◦
............◦.....◦..◦..◦
..◦ ..◦
.....◦..◦
.....◦..◦..◦
..◦
............◦.....◦..◦..◦
..◦ ..◦
.....◦..◦
.....◦..◦..◦
..◦
..◦ ..◦
.....◦..◦
..◦
.....◦..◦..◦
..◦
.....◦..◦..◦
..◦
..◦
..◦
..◦
The Cayley graph of the free group picturing a river basin.
The positive generators are pointing downwards, towards
southwest and southeast, to give the idea of downstream.
form an element of the subshift X, but if we start anywhere in any tributary, and if we
decide to travel downstream (i.e. following the edges of the Cayley graph corresponding to
positive generators), we will pick up another infinite word, which will merge into the main
river, and which will also consist of an infinite word belonging to X.
Although we do not believe it is possible to find a complete set of properties charac-
terizing ΩX, we at least know that it contains a dense copy of X (not necessarily with the
same topology), which may be precisely characterized and which allows for a reasonably
good handle on the other, more elusive elements of ΩX . Seen from this perspective, ΩX
appears slightly friendlier and we are in turn able to explore it quite efficiently.
While the methods most often used in the literature for analyzing OX are based on
a rather technical study of a certain AF-subalgebra (see e.g. [24: Corollary 3.9] and [31
: Theorem 4.14]), our arguments are rooted in the dynamical properties of the spectral
partial action. In particular our description of ΩX is concrete enough to allow us to find
sensible necessary and sufficient conditions for this partial action to be topologically free
and minimal. Our condition for minimality, for instance, is a lot similar to the cofinality
4
m. dokuchaev and r. exel
condition which has played an important role in characterizing simplicity for graph algebras
[21: Theorem 6.8], [20: Corollary 3.11]. These two crucial properties, namely topological
freeness and minimality, have been extensively used to characterize simplicity, and thus
our treatment of OX is done in the same footing as for several other better behaved C*-
algebras.
We believe this new picture for the hitherto intractable spectrum of DX will allow
for further advances in the understanding of subshifts as well as of Carlsen-Matsumoto
C*-algebras.
2. Subshifts.
We begin by fixing a nonempty finite set Λ which will henceforth be called the alphabet,
and whose elements will be called letters.
Any finite sequence of letters will be called a finite word , including the empty word ,
namely the word with length zero, which will be denoted by ∅. The set of all finite words
will be denoted by Λ∗. Infinite sequences of letters will also be considered and we shall
call them infinite words.
The best way to formalize the notion of sequences, twice referred to above, is by
resorting to the Cartesian product Λ × Λ × . . . × Λ, whose elements are therefore denoted
by something like
(x1, x2, . . . , xn).
We shall however choose a more informal notation, denoting such a sequence by
or by
x1x2 . . . xn,
x1x2x3 . . .
in the infinite case. This is compatible with our point of view according to which sequences
are viewed as words.
If α is a finite word and x is a finite or infinite word, then we will write αx for the
concatenation of α and x, namely the word obtained by juxtaposing α and x together.
The length of a finite word α, denoted α, is the number of letters in it.
Assigning the discrete topology to Λ, the set of all infinite words, namely Λ, becomes
a compact topological space with the product topology by Tychonoff's Theorem. The map
S : Λ → Λ
given by
S(x1x2x3x4 . . .) = x2x3x4x5 . . . ,
(2.1)
for every x = x1x2x3x4 . . . ∈ Λ, is called the (left) shift.
continuous.
Given a nonempty closed subset X ⊆ Λ such that
It is easy to see that S is
S(X) ⊆ X,
partial actions and subshifts
5
we may consider the restriction of S to X, and then the pair (X, SX) is called a (one-sided)
subshift. Sometimes we will also say that X itself is a subshift, leaving the shift map to
be deduced from the context.
There are many concrete situations in Mathematics where subshifts arise naturally
such as in dynamical systems, Markov chains, maps of the interval, billiards, geodesic
flows, complex dynamics, information theory, automata theory and matrix theory. The
present work is dedicated to studying subshifts from the point of view or partial dynamical
systems [16].
In what follows let us give an important example of subshifts. A finite word α is said
to occur in an infinite word x = x1x2x3x4 . . . , if there are integer numbers n ≤ m, such
that
α = xnxn+1 . . . xm.
In other words, α may be found within x as a contiguous block of letters. By default we
consider the empty word ∅ as occurring in any infinite word.
Given an arbitrary subset F ⊆ Λ∗, appropriately called the set of forbidden words, let
X = XF
be the set of all infinite words x such that no member of F occurs in x. It is easy to see
that S(XF ) ⊆ XF , and a simple argument shows that XF is closed in Λ, hence XF is a
subshift. If a subshift X coincides with XF , for some finite set F ⊆ Λ∗, then we say that
X is a subshift of finite type.
A notable example of a shift which is not of finite type, and which will often be used as
a counter-example below, is as follows: over the alphabet Λ = {0, 1}, consider as forbidden
all words of the form
where n ≥ 0. Thus any odd string of 1's delimited by two 0's is forbidden. The subshift
defined by this set of forbidden words is called the even shift. Clearly, an infinite word x
lies in the space of the even shift if and only if, anytime a contiguous block of 1's occurring
in x is delimited by 0's, there is an even number of said 1's. It should be noted that an
infinite word beginning with an odd (sic) number of ones and followed by a zero is not
immediately excluded.
Given a subshift X, the language of X is defined to be the subset
LX ⊆ Λ∗
(2.2)
formed by all finite words which occur in some x ∈ X.
For future reference we cite here a well known result in Symbolic Dynamics:
2.3. Proposition. [23: Proposition 1.3.4] For any subshift X one has that X = XF ,
where F = Λ∗ \ LX .
The following notions will prove to be of utmost importance in what follows:
2n+1
01
0 = 0 1 . . . 1
0,
2n+1
{z }
6
m. dokuchaev and r. exel
2.4. Definition. Let Λ be a finite alphabet, and let X ⊆ Λ be a subshift. For each α
in Λ∗, the follower set Fα, and the cylinder set Zα are defined by
Fα = {y ∈ X : αy ∈ X}, and
Zα = {x ∈ X : x = αy, for some infinite word y},
It is well known that the Zα form a basis for the product topology on X, consisting
of compact open subsets.
Thus the follower set of α is the set of all infinite words which are allowed to follow
α, while the cylinder Zα is the set of words which begin with the prefix α. If α = n, then
clearly Sn(Zα) = Fα, and the restriction of Sn to Zα gives a bijective map
Sn : Zα → Fα.
Since Zα is compact, the above is a homeomorphism from Zα to Fα. In particular it
follows that Fα is also compact, although it is not necessarily open as we shall now see in
the following well known result whose precise statement we have not been able to locate
in the literature.
2.5. Proposition. Given a finite alphabet Λ, and a subshift X ⊆ Λ, the following are
equivalent:
(i) Fα is open in X for every finite word α in Λ∗,
(ii) Fa is open in X for every letter a in Λ,
(iii) S is an open mapping on X,
(iv) X is a subshift of finite type.
Proof. (i) ⇒ (ii). Obvious.
(ii) ⇒ (iii). Given an open subset U ⊆ X, we have for every a in Λ, that U ∩ Za is open
relative to Za, and since S is a homeomorphism from Za to Fa, we conclude that S(U ∩Za)
is open relative to Fa, which in turn is open relative to X by hypothesis. So S(U ∩ Za) is
open relative to X, whence
S(U ) = S(cid:0) Sa∈A
U ∩ Za(cid:1) = Sa∈A
S(U ∩ Za)
is open in X.
(iii) ⇒ (i). As already seen Zα is always open and Fα = Sn(Zα), where n = α.
(iii) ⇔ (iv) This is a well known classical result in Symbolic Dynamics. See [28: Theorem
1], [19: Theorem 1] and [22: Theorem 3.35].
(cid:3)
partial actions and subshifts
7
3. Circuits.
One of the most important aspects of subshifts to be discussed in this work is the question
of topological freeness (to be defined later), and which requires a careful understanding
of circuits. We will therefore set this section aside to discuss this concept. Besides the
essential facts about circuits we shall need later, we will also present some interesting, and
perhaps unknown facts which came up in our research.
In order to motivate the concept of circuits let us first discuss the case of Markov
subshifts. So, for the time being, we will let X be a Markov subshift, which means that
X = XF , where the set F of forbidden words contains only words of length two. Setting
if the word 'ij' lies in F ,
aij =(cid:26) 0,
1, otherwise,
for every i, j ∈ Λ, the resulting 0-1 matrix A = (aij)i,j∈Λ is called the transition matrix
for X. Excluding the uninteresting case in which some letter of the alphabet is never used
in any infinite word in X, the transition matrix has no zero rows.
Let Gr(A) be the graph having A as its adjacency matrix, so that its vertices are the
elements of Λ, while there is one edge from vertex i to vertex j when aij = 1, and none
otherwise. The elements of X may then be thought of as infinite paths in Gr(A), namely
infinite sequences of vertices (as opposed to edges, another popular concept) in which two
successive vertices are joined by an edge.
3.1. Definition. Given a Markov subshift X, with transition matrix A, we will say that
a circuit 1 in Gr(A) is a finite path
γ = x1x2 . . . xn,
such that axn,x1 = 1.
For each circuit γ, the infinite periodic path
γ∞ = γγγ . . .
lies in X. The question of topological freeness, already alluded to (but not yet defined),
will be seen to be closely related the the non-existence of circuits γ such that γ∞ is an
isolated point of X. If indeed γ∞ is not isolated, that is, if {γ∞} is not an open set, we
have in particular that,
{γ∞} 6= Zγ.
This is to say that Zγ must contain at least one point other than γ∞. In other words, it
must be possible to prolong the finite word γ in such a way as to obtain an infinite path
distinct from γ∞. Thus the relevance of an exit for γ, namely a letter xi in the above
expression for γ which may be followed by at least one letter other than xi+1 (in case
i = n, we take xi+1 to mean x1).
In a general subshift, not necessarily of finite type, the above considerations do not
make any sense since there is no underlying graph, but they may nevertheless be reinter-
preted, as we will now see.
1 Also known as a loop, a cycle, or a closed path.
8
m. dokuchaev and r. exel
3.2. Definition. Let Λ be a finite alphabet, and let X ⊆ Λ be a subshift.
(a) We will say that a finite word γ ∈ Λ∗ is a circuit (relative to X), provided the infinite
periodic word γ∞ = γγγ . . .
lies in X.
(b) We will say that an infinite word y 6= γ∞ is an exit for a given circuit γ, if γy lies in
X.
Thus, to say that a circuit γ has an exit is equivalent to saying that the follower set
Fγ has at least one element other than γ∞. If γ is a circuit then
γ n := γ . . . γ
is also a circuit for any n ∈ . However, even if γ has an exit, there is no reason why γ n
would also have an exit (unless X is a Markov subshift).
{z }n
An example of this situation is the shift (of finite type) XF , on the alphabet {0, 1},
where F consists of a single forbidden word, namely '001'. Evidently γ = '0' is a circuit,
which admits the word y = '111 . . .' as an exit. However γ2 = '00' clearly has no exit.
Having an exit is therefore no big deal. A much more impressive property of a circuit
γ is for γ n to have an exit for every n.
3.3. Proposition. Let Λ be a finite alphabet and let X ⊆ Λ be a subshift. Given a
circuit γ, the following are equivalent:
(i) For each n ∈ , one has that γ n has an exit,
(ii) There is an infinite word z which is an exit for γ n, for all n ∈ .
Proof. The crucial difference between (i) and (ii), as the careful reader would have already
noticed, is that in (i) it is OK for each γ n to have a different exit, while in (ii) it is required
that there is one single exit which works for all γ n.
It is obvious that (ii) ⇒ (i), so let us focus on the converse. Assuming that γ n has an
exit for each n, we have that the follower set Fγn contains at least one element besides
(γ n)∞ = γ∞. We claim that in fact Fγn contains at least one point z which does not lie
in the cylinder Zγ.
Notice that, even though γ∞ lies in Fγn , it cannot be taken as the z above since it
lies in the cylinder Zγ, while z should not.
To prove the claim, choose y ∈ Fγn , with y 6= γ∞, and write
y = α1α2α3 . . . ,
where each αi is a finite word with the same length as γ. If α1 6= γ, then y is not in Zγ,
and there is nothing to be done. Otherwise let k be the smallest index such that αk 6= γ,
so k > 1, and α1 = α2 = · · · = αk−1 = γ. We may then write
y = γ k−1z,
where z = αkαk+1 . . .
of γ n, we then have that
is therefore an infinite word not in Zγ. Since y is in the follower set
X ∋ γ ny = γ nγ k−1z = γ k−1γ nz.
partial actions and subshifts
9
Recall that X is invariant under the shift. So, after applying S to the above element
γ(k − 1) times, we conclude that γ nz lies in X, so z ∈ Fγn \ Zγ, as desired.
Since Fγn is closed and Zγ is open, we have that Fγn \ Zγ is closed, and nonempty as
seen above. Observing that the Fγn are decreasing with n, we have by compactness of X
that
Any element z chosen in the above intersection is therefore not equal to γ∞, because
(cid:3)
it is not in Zγ, and it is therefore an exit for γ n, for all n ∈ .
Fγn \ Zγ 6= ∅.
Tn∈
3.4. Definition. Let Λ be a finite alphabet, and let X ⊆ Λ be a subshift. If γ is a
circuit relative to X, we will say that an infinite word y 6= γ∞ is a strong exit for γ, if
γ ny ∈ X, for all n ∈ .
It follows from (3.3) that a subshift in which all circuits have an exit, also satisfies the
apparently stronger property that all circuits have a strong exit (think about it).
An interesting example of such a subshift is as follows:
3.5. Proposition. Let X be the even shift. Then any circuit γ relative to X has a strong
exit.
Proof. Let us first assume that
for some k ≥ 1. Then, no matter how big is n, we may always exit γ n via an infinite string
of 0's. All other circuits γ have '0' somewhere, say γ = γ′0γ′′. Observing that
γ = 1k = 1 . . . 1
,
{z }k
γ nγ′011111 . . .
lies in the space of the even shift, one may always exit γ n via the infinite word γ′011111 . . . ,
which is therefore a strong exit for γ.
(cid:3)
In section (12) we will carefully study topological freeness for the partial dynamical
systems we shall encounter along the way. However we warn the reader that the nice
property of the even shift proven above (existence of strong exits for all circuits) will be
seen to be still insufficient for our purposes.
4. The standard partial action associated to a subshift.
Throughout this section we will fix a finite alphabet Λ and a subshift X ⊆ Λ.
As already seen, the shift restricts to a homeomorphism from Za to Fa, for each a in
Λ. The inverse of this homeomorphism is clearly given by the map
defined by
θa : Fa → Za,
θa(y) = ay,
∀ y ∈ Fa.
10
m. dokuchaev and r. exel
4.1. Definition. Let F = F(Λ) be the free group on Λ and let
θ =(cid:0){Xg}g∈F, {θg}g∈F(cid:1)
be the unique semi-saturated partial action of F on X assigning θa to each a in Λ, given
by [16: 4.10]. Henceforth θ will be referred to as the standard partial action associated to
X.
Incidentally, to say that θ is semi-saturated is to say that
θgh = θg ◦ θh,
whenever gh = g + h, where · is the usual length function on F. See [16: 4.9].
For the case of Markov subshifts, in fact for a generalization thereof, the standard
partial action was first studied in [17].
Recall from [16: 5.1] that, for a partial action on a topological space, it is usually
required that the Xg be open sets. However, since
Xa−1 = Domain(θa) = Fa,
∀ a ∈ Λ,
that requirement is not fulfilled for our θ unless X is a subshift of finite type by (2.5).
Rather than a nuisance, this is the first indication that a non-finite type subshift conceals
another partial action which will be seen to be crucial for the analysis we will carry out
later.
When considering θ, it is therefore best to think of it as a partial action in the category
of sets (as opposed to topological spaces).
In what follows let us give a simple description for θ, but first let us introduce some
notation. Denote by F+ the subsemigroup of F generated by Λ ∪ {1}, so that F+ may be
identified with the set Λ∗ of all finite words in Λ. Under this identification we shall see the
empty word ∅ as the unit of F.
4.2. Proposition. For every g in F one has:
−1
(i) If g 6∈ F+F
+ , then Xg is the empty set.
(ii) If g = αβ−1, with α, β ∈ F+, and g = α + β, then
Xg = {αy ∈ X : y ∈ Fα ∩ Fβ}.
(iii) If g is as in (ii), and x ∈ Xg−1, write x = βy, for some y ∈ Fβ ∩ Fα. Then
Proof. Given g in F, write
θg(x) = θαβ−1(βy) = αy.
g = c1c2 . . . cn,
with ci ∈ Λ ∪ Λ−1 in reduced form, meaning that ci+1 6= c−1
semi-saturated we have that
i
, for every i. Since θ is
θg = θc1 ◦ θc2 ◦ · · · ◦ θcn.
partial actions and subshifts
11
If g /∈ F+F
−1
+ then there is some i such that ci ∈ Λ−1, and ci+1 ∈ Λ, say ci = a−1
and ci+1 = b, for some a, b ∈ Λ. Therefore a 6= b, and then
θci ◦ θci+1 = θ−1
a θb
is the empty map, because after inserting b as the prefix of an infinite word x, we are left
with a word which does not begin with the letter a! Under these conditions we then have
that θg is the empty map, hence Xg is the empty set. This proves (i).
Given a finite word α, it is easy to see that Xα−1 = Fα, Xα = Zα, and
θα(x) = αx,
∀ x ∈ Xα−1.
If g is as in (ii) we have, again by semi-saturatedness that θg = θα ◦ θ−1
and (iii) follow easily.
β , from where (ii)
(cid:3)
We should remark that the sets Xg, appearing in (4.2.ii) above, have also played an
important role in Carlsen's study of subshifts [8: Definition 1.1.3].
Due to (4.2.ii) there will be numerous situations below in which we will consider group
elements of the form g = αβ−1, with α, β ∈ F+, and
g = α + β.
To express this fact we will simply say that g is in reduced form. This is clearly equivalent
to saying that the last letter of α is distinct from the last letter of β, so that no cancellation
takes place.
Observe also that "reduced form" is an attribute of the presentation of g as a product
−1
+ may be written as g = αβ−1,
of two elements, rather than of g itself. Any element of F+F
with α, β ∈ F+, and upon canceling as many final letters of α and β as necessary, one is
left with a reduced form presentation of g.
5. A partial representation associated to the subshift.
Throughout this section we will again fix a finite alphabet Λ and a subshift X ⊆ Λ.
Recall from [16: 9.1 & 9.2] that a *-partial representation of a group G in a unital
*-algebra B is a map u : G → B satisfying
(PR1) u1 = 1,
(PR2) uguhuh−1 = ughuh−1.
(PR3) ug−1 = (ug)∗,
for every g and h in G.
Consider the complex Hilbert space ℓ2(X), with its canonical orthonormal basis
{δx}x∈X . Here we shall consider *-partial representations of F = F(Λ) in the algebra
of bounded linear operators on ℓ2(X), and we will refer to these simply as partial repre-
sentations of F on ℓ2(X).
12
m. dokuchaev and r. exel
5.1. Proposition. For each g in F, denote by ug the unique bounded linear operator
such that for each x in X,
ug : ℓ2(X) → ℓ2(X),
ug(δx) =( δθg (x),
0,
if x ∈ Xg−1,
otherwise,
where θ is the standard partial action associated to X. Then the correspondence g 7→ ug
is a semi-saturated partial representation of F on ℓ2(X).
Proof. Given g and h in F, and x in X, we must prove that
uguhuh−1 (δx) = ughuh−1(δx).
(5.1.1)
Suppose first that x ∈ Xh ∩ Xg−1. Then the left-hand-side above equals
Observing that
the right-hand-side of (5.1.1) equals
uguhuh−1(δx) = uguh(cid:0)δθh−1 (x)(cid:1) = ug(δx) = δθg (x).
θh−1 (x) ∈ θh−1(cid:0)Xh ∩ Xg−1(cid:1) = Xh−1 ∩ X(gh)−1,
ughuh−1 (δx) = ugh(cid:0)δθh−1 (x)(cid:1) = δθgh(θh−1 (x)) = δθg (x),
proving (5.1.1) in the present case.
Suppose now that x /∈ Xh ∩ Xg−1. In case x /∈ Xh, then uh−1 (δx) = 0, and (5.1.1)
follows trivially. So we are left with the case that x ∈ Xh \ Xg−1. Under this assumption
we have
Moreover notice that y := θh−1(x) /∈ X(gh)−1, since otherwise
uguhuh−1 (δx) = uguh(cid:0)δθh−1 (x)(cid:1) = ug(δx) = 0.
x = θh(y) ∈ θh(cid:0)Xh−1 ∩ X(gh)−1(cid:1) = Xh ∩ Xg−1,
contradicting our assumptions. Therefore,
0 = ugh(δy) = ugh(δθh−1 (x)) = ughuh−1 (δx),
showing that the right-hand-side of (5.1.1) also vanishes. This proves (5.1.1), also known
as (PR2), and we leave the easy proofs of (PR1) and (PR3), as well as the fact that S is
semi-saturated, for the reader.
(cid:3)
partial actions and subshifts
13
Either analyzing the definition of ug directly, or as a consequence of the above result
(see [16: 9.8.i]), we have that ug is a partial isometry for every g in F, with initial space
ℓ2(Xg−1) and with final space ℓ2(Xg).
More specifically, recall from (4.2.i) that θg is the empty map when g is not in F+F
−1
+ ,
in which case
ug = 0.
On the other hand, assuming that g lies in F+F
form, with α, β ∈ F+, whence
ug = uαu∗
β,
(5.2)
−1
+ , we may write g = αβ−1, in reduced
by semi-saturatedness. Based on the description of θ given in (4.2.ii & iii), we then have
for every x in X that
uαβ−1 (δx) =(cid:26) δαy,
0,
if x = βy, for some y ∈ Fα ∩ Fβ,
otherwise.
Still under the assumption that g = αβ−1, in reduced form, we have by (4.2.ii) that
the final space of ug is
ℓ2(Xg) = span{δαy : y ∈ Fα ∩ Fβ}.
(5.3)
A relevant remark is that the final projections ugu∗
g are then seen to be diagonal
operators relative to the canonical orthonormal basis. In particular these commute with
each other, a well known fact from the general theory of partial representations [16: 9.8.iv].
A special case of interest is when g = α ∈ F+, in which case we have that
uα(δx) =(cid:26) δαx,
0,
if x ∈ Fα,
otherwise.
From this we see that the initial space of uα is ℓ2(Fα), and its final space is ℓ2(Zα).
The adjoint of uα may be described by
u∗
α(δy) = uα−1 (δy) =(cid:26) δx,
0,
if y ∈ Zα, and y = αx, with x ∈ Fα,
otherwise.
5.4. Proposition. Let T be the operator on ℓ2(X) defined by T =Pa∈Λ u∗
T (δx) = δS(x),
a. Then
for every x in X, where S is the shift map introduced in (2.1).
Proof. Given any y in X, let b be the first letter of y, and write y = bx, for some infinite
word x, necessarily in Fb. Then u∗
a(δy) = 0, for all a 6= b. This implies
that
b (δy) = δx, while u∗
u∗
a(δy) = δx = δS(y).
(cid:3)
T (δy) = Pa∈Λ
The operator T may therefore be interpreted as the manifestation of the shift S at the
level of operators on ℓ2(X). Since each ua is a partial isometry, it is clear that kT k ≤ Λ.
14
m. dokuchaev and r. exel
6. C*-algebras associated to subshifts.
Throughout this section we will fix a finite alphabet Λ and a subshift X ⊆ Λ. Our
goal here is to describe two important C*-algebras that have been extensively studied in
association with a subshift.
6.1. Proposition. Let π be the representation of C(X) on ℓ2(X) defined on the canonical
orthonormal basis by
π(f )δx = f (x)δx,
∀ f ∈ C(X),
∀ x ∈ X.
Then the following three sets generate the same C*-algebra of operators on ℓ2(X):
(i) u(F) = {ug : g ∈ F},
(ii) u(Λ) = {ua : a ∈ Λ},
(iii) π(cid:0)C(X)(cid:1) ∪ {T }.
Proof. Denote the C*-algebras generated by the sets in (i), (ii) and (iii) by BF, BΛ and
BT , respectively.
Since u is semi-saturated, for every g in F, one has that ug may be written as a
product of elements in u(Λ) ∪ u(Λ)∗. Therefore BF ⊆ BΛ.
For every α in Λ∗, the cylinder Zα is a clopen subset of X, so its characteristic
function, which we denote by 1α, is a continuous function. Moreover π(1α) is the orthogonal
projection onto ℓ2(Zα), hence it coincides with the final projections of uα, so
π(1α) = uαu∗
α ∈ BF.
It is evident that the set {1α : α ∈ Λ∗} separates points of X, so by virtue of the
Stone-Weierstrass Theorem, it generates C(X), as a C*-algebra. Therefore
Since T ∈ BF, by definition, we then conclude that
π(cid:0)C(X)(cid:1) ⊆ BF.
Given a in Λ, notice that
BT ⊆ BF.
π(1a)T ∗ = Pb∈Λ
uau∗
aub = ua,
proving that ua lies in the C*-algebra generated by π(cid:0)C(X)(cid:1) and T , whence BΛ ⊆ BT .
We have therefore shown that
whence the conclusion.
(cid:3)
BF ⊆ BΛ ⊆ BT ⊆ BF,
partial actions and subshifts
15
6.2. Definition. The Matsumoto C*-algebra associated to a given subshift X, henceforth
denote by MX , is the closed *-algebra of operators on ℓ2(X) generated by any one of the
sets described in the statement of (6.1).
The algebra defined above was first introduced by Matsumoto in [26: Lemma 4.1]
under the notation OX . However, due to the existence of several different algebras asso-
ciated with subshifts usually denoted by OX in the literature, we would rather use a new
notation. See also remark (6.5), below.
Let λ denote the left regular representation of the free group F = F(Λ) on ℓ2(F).
Thus, denoting the canonical orthonormal basis of ℓ2(F) by {δg}g∈F, one has that
λg(δh) = δgh,
∀ g, h ∈ F.
Regarding the partial representation u introduced in (5.1), we may define a new partial
representation u of F on ℓ2(X) ⊗ ℓ2(F), by tensoring u with λ, namely
ug = ug ⊗ λg,
∀ g ∈ F.
6.3. Proposition. Let T be the operator on ℓ2(X) ⊗ ℓ2(F) given by T =Pa∈Λ u∗
the following three sets generate the same C*-algebra of operators on ℓ2(X) ⊗ ℓ2(F):
(i) u(F),
(ii) u(Λ),
a. Then
(iii) (cid:0)π(cid:0)C(X)(cid:1) ⊗ 1(cid:1) ∪ { T }.
Proof. Denote the C*-algebras generated by the sets in (i), (ii) and (iii) by BF, BΛ and
BT , respectively. That BF ⊆ BΛ follows, as above, from the fact that u is semi-saturated.
For α in Λ∗ one has
π(1α) ⊗ 1 = uαu∗
α ⊗ 1 = (uα ⊗ λα)(uα ⊗ λα)∗ = uα u∗
α ∈ BF.
This, plus the fact that C(X) is generated by {1α : α ∈ Λ∗}, gives π(cid:0)C(X)(cid:1) ⊗ 1 ⊆ BF.
Since T ∈ BF, by definition, we then conclude that BT ⊆ BF. Given a in Λ, notice that
(cid:0)π(1a) ⊗ 1(cid:1) T ∗ = (uau∗
a ⊗ 1) Pb∈Λ
ub ⊗ λb = Pb∈Λ
uau∗
aub ⊗ λb = ua ⊗ λa = ua,
proving that ua lies in the C*-algebra generated by π(cid:0)C(X)(cid:1) ⊗ 1 and T , whence BΛ ⊆ BT ,
concluding the proof.
(cid:3)
6.4. Definition. The Carlsen-Matsumoto C*-algebra associated to a given subshift X,
henceforth denote by OX , is the closed *-algebra of operators on ℓ2(X) ⊗ ℓ2(F) generated
by any one of the sets described in the statement of (6.3).
The algebra defined above was studied in [10], [7], [8] (where it was also denoted by
OX ). Its definition in the above mentioned references is not the one given in (6.4), but we
will prove in (9.5) that the two definitions lead to the same algebra. We nevertheless note
that the description of OX given by (6.3.iii) is closely related to that given in [10].
16
m. dokuchaev and r. exel
6.5. Remark. Let us clarify the relationship between our notation and the one used in
the literature. According to [10] there are at least three possibly non isomorphic C*-
algebras associated to a subshift in the literature, some of them defined only for two-sided
subshifts. These are:
(a) the C*-algebra OΛ defined in [24],
(b) the C*-algebra OΛ defined in [9],
( c ) the C*-algebra OX defined in [7].
We shall not be concerned with the first algebra above, while the second and third ones
are respectively being denoted by MX and OX in this work.
Notice that, although the notation OX emphasizes the dynamical system (S, X), the
construction of this algebra was actually made based on the specific way in which X is
represented as a space of infinite words, as well as on the syntactic rules of inserting a
letter ahead of an infinite word.
It is then legitimate to ask what is the precise relationship between OX1 and OX2 ,
in case X1 and X2 are conjugate subshifts. This question has been addressed for the first
time in [27] where it was proved that MX1 is Morita equivalent to MX2 under suitable
hypothesis. Later Carlsen [8: Theorems 1.3.1 & 1.3.5] found a proof of the fact that OX1
and OX2 are Morita equivalent without any additional hypotheses. In [7: Section 8], having
realized OX as a Cuntz-Pimsner algebra, Carlsen finally proved that OX1 and OX2 are
actually isomorphic. A further, much simpler proof of this result was found by Carlsen
and Silvestrov [10: Section 11], based on the description of OX as a crossed product by an
endomorphism and a transfer operator [14].
The following proof, based on [10], is perhaps the simpler possible proof of the invari-
ance of Matsumoto's algebras, given that it only needs the original description of these
algebras.
6.6. Proposition. Suppose that X1 and X2 are conjugate subshifts. Then
MX1 ≃ MX2.
Proof. Let ϕ : X1 → X2 be a homeomorphism such that ϕ◦S1 = S2◦ϕ, where S1 and S2 are
the shift maps on X1 and X2, respectively. Define a unitary operator U : ℓ2(X1) → ℓ2(X2),
by setting
U (δx) = δϕ(x),
∀ x ∈ X1.
Decorating all of the ingredients of (6.1) with subscripts to indicate whether we are
speaking of X1 or X2, it is easy to see that
U π1(cid:0)C(X1)(cid:1)U ∗ = π2(cid:0)C(X2)(cid:1),
and U T1U ∗ = T2.
By (6.1) we then see that UMX1 U ∗ = MX2 , so MX1 and MX2 are in fact spatially
isomorphic.
(cid:3)
Even though the above result is probably well known to the experts, we have not been
able to locate it in the literature with this exact formulation.
Given our reliance on the free group, whose rank is definitely not an invariant of the
subshift, our method does not seem appropriate to prove Carlsen's invariance Theorem [7
: Theorem 8.6].
partial actions and subshifts
17
7. The spectrum.
As always we fix a finite alphabet Λ and a subshift X ⊆ Λ. There is an important subal-
gebra of MX which has played a crucial role in virtually every attempt to study subshifts
from the point of view of C*-algebras (see the references given in the introduction), and
which we would now like to describe.
7.1. Definition. We will denote by DX the closed *-algebra of operators on ℓ2(X) gen-
erated by the final projections
eg := ugu∗
g,
for all g in F.
Recall from [16: 9.8.iv] that the eg commute with each other, so DX is a commutative
C*-algebra, which is moreover unital because e1 = 1.
Since u is also a partial representation, the final projections
eg := ug u∗
g,
likewise commute with each other, and hence generate a commutative C*-algebra. However
notice that
eg = (ug ⊗ λg)(ug ⊗ λg)∗ = ugu∗
g ⊗ 1 = eg ⊗ 1,
(7.2)
from where one concludes that the C*-algebra generated by the eg is nothing but DX ⊗ 1.
This section is dedicated to studying the spectrum of DX , henceforth denoted by
sp(DX).
The full description of this space requires some machinery still to be developed, but
we may easily give examples of some of its elements, as follows. Observing that the
multiplication of two diagonal operators is done by simply multiplying the corresponding
diagonal entries, we see that the assignment of a given diagonal entry to an operator defines
a multiplicative linear functional on the set of all diagonal operators. In what follows we
will refer to diagonal entries indirectly, as eigenvalues relative to eigenvectors taken from
the canonical basis.
7.3. Definition. Given any x in X, let ϕx be the unique linear functional on DX such
that
a(δx) = ϕx(a)δx,
∀ a ∈ DX.
As observed above, each ϕx is a character on DX , hence an element of sp(DX). We
will see that not every element of sp(DX) is of the form ϕx, but the ϕx nevertheless form
a large subset of sp(DX) in the following sense:
7.4. Proposition. The subset of sp(DX) formed by all the ϕx is a dense set.
18
m. dokuchaev and r. exel
Proof. Assume by way of contradiction that the closure of
{ϕx : x ∈ X},
which we denote by C, is a proper subset of sp(DX). Picking a point ϕ outside C we
may invoke Urysohn's Lemma to find a continuous complex valued function f on sp(DX)
which vanishes on C, and such that f (ϕ) = 1. By Gelfand's Theorem we have that f is
the Gelfand transform of some a in DX, and then for every x in X we have
ϕx(a) = f (ϕx) = 0.
This implies that all diagonal entries of a are zero, and since a is itself a diagonal operator,
we deduce that a = 0, and hence also that f = 0, a contradiction.
(cid:3)
We thus get a map
Φ : x ∈ X 7→ ϕx ∈ sp(DX),
whose range is dense in sp(DX).
A crucial question in this subject is whether or not Φ is continuous. Should this be
the case, the compactness of X would imply that Φ is onto, and the mystery surrounding
sp(DX) would be immediately dispelled. However, we will see later in (7.19) that for
subshifts not of finite type sp(DX) is strictly bigger than the range of Φ and this can only
happen if Φ is discontinuous!
In order to describe the whole of sp(DX) it is useful to recall that DX is generated,
as a C*-algebra, by the projections
eg := ugu∗
g,
so a character ϕ on DX is pinned down as soon as we know the numbers ϕ(eg), which
necessarily lie in {0, 1}, for every g in F.
To be precise, given ϕ ∈ sp(DX), consider the element ξϕ ∈ 2F given by
ξϕ(g) = ϕ(eg),
∀ g ∈ F.
Identifying 2F with the set of all subsets of F as usual, we may think of ξϕ as the
subset of F given by
ξϕ = {g ∈ F : ϕ(eg) = 1}.
(7.5)
7.6. Proposition. Considering 2F as a topological space with the product topology, the
mapping
Ψ : ϕ ∈ sp(DX) 7→ ξϕ ∈ 2F,
is a homeomorphism from sp(DX) onto its range.
Proof. As already observed, each ϕ in sp(DX) is characterized by its values on the gener-
ating idempotents, so Ψ is seen to be one-to-one. It is evident that Ψ is continuous, and
since sp(DX) is compact, we have that Ψ is a homeomorphism onto its image.
(cid:3)
partial actions and subshifts
19
7.7. Definition. The range of Ψ, which will henceforth be denoted by
ΩX = Ψ(sp(DX )),
will be referred to as the spectrum of the subshift X.
As seen in (7.6), we have that ΩX is homeomorphic to sp(DX), and we will take the
former as our main model to study the latter. Since sp(DX) contains a dense copy of the
set2 X by (7.4), we have that ΩX also contains a dense copy of X. However we will see
that for subshifts not of finite type, ΩX is strictly bigger than X.
We will generally prefer to regard a given ξ in 2F as a subset of F, in the spirit of
(7.5), rather than as a {0, 1}-valued function on F.
Here is the list of properties, alluded to in the introduction, that every ξ in ΩX must
satisfy:
7.8. Proposition. For any ξ in ΩX one has that:
(i) 1 ∈ ξ,
(ii) ξ is convex,
(iii) for every g ∈ ξ, there exists a unique a in Λ, such that ga ∈ ξ,
(iv) if g ∈ ξ, and α is a finite word in Λ such that gα ∈ ξ, then α lies in LX.
(v) ξ ⊆ F+F
−1
+ .
Proof. Let ϕ be the character on DX such that ξ = ξϕ, so that
ξ = {g ∈ F : ϕ(eg) = 1}.
Since e1 = 1, one has that ϕ(e1) = 1, so 1 ∈ ξ, proving (i).
Recall from [16: 14.19] that to say that ξ is convex is to say that, whenever g, h ∈ ξ,
then the segment joining g and h, namely
gh := {k ∈ F : g−1h = g−1k + k−1h}
is contained in ξ. Assuming that g and h are in ξ, and that k is in gh, set
s = g−1k,
and
t = k−1h.
We then have that st = s + t, so ust = usut, by semi-saturatedness. Employing [16:
14.5] we conclude that est ≤ es, which is to say that
eg−1h ≤ eg−1k.
Conjugating the left-hand-side of this inequation by ug, we get by [16: 9.8.iii] that
ugeg−1hug−1 = ehugug−1 = eheg.
2 We use set as opposed to topological space to highlight the fact that it is not necessarily homeomorphic
to X, as Φ may be discontinuous.
20
m. dokuchaev and r. exel
Doing the same relative to the right-hand-side leads to
ugeg−1kug−1 = ekeg,
so we deduce that eheg ≤ ekeg, whence
ϕ(eheg) ≤ ϕ(ekeg).
Having assumed that g, h ∈ ξ, we see that ϕ(eg) = ϕ(eh) = 1, so the left-hand-side
above evaluates to 1. The same is therefore true for the right-hand-side, which implies
that ϕ(ek) = 1, from where it follows that k ∈ ξ, as desired. This proves (ii).
In order to prove (iii), pick any g in ξ. Observing that X is the disjoint union of the
cylinders Za, as a range in Λ, and that ea is the orthogonal projection onto ℓ2(Za), we see
that
Conjugating the above identity by ug, we deduce that
eg = ugug−1 = ug(cid:0) Pa∈Λ
Since g ∈ ξ, we have that ϕ(eg) = 1, so
ea = 1.
Pa∈Λ
ea(cid:1)ug−1 = Pa∈Λ
1 = ϕ(eg) = Pa∈Λ
egaugug−1 = Pa∈Λ
egaeg.
ugeaug−1 = Pa∈Λ
ϕ(egaeg) = Pa∈Λ
ϕ(ega).
Each ega is idempotent so ϕ(ega) is either 0 or 1. We then see that there exists a
unique a in Λ such that ϕ(ega) = 1, meaning that ga ∈ ξ, hence proving (iii).
Supposing, as in (iv), that g and gα lie in ξ, observe that
egegα = ugug−1egα = ugeαug−1.
Since ϕ(egegα) = 1, by hypothesis, it follows that egegα 6= 0, whence also eα 6= 0. Observ-
ing that eα is the orthogonal projection onto ℓ2(Zα), one deduces that Zα is nonempty,
which implies that α is a word in the language LX .
Regarding (v), let g ∈ ξ. Then ϕ(eg) = 1, so eg is nonzero and hence neither is ug. It
(cid:3)
then follows from (5.2) that g lies in F+F
−1
+ .
It should be stressed that properties (7.8.i-v), which we have seen to hold for every ξ
in ΩX , are not enough to characterize the elements in ΩX . Although it would be highly
desirable to find a set of properties giving such a precise characterization, we have not
been able to succeed in this task.
Nevertheless, recalling that the image of X under Φ is dense in sp(DX) by (7.4), we
already have a somewhat satisfactory description of ΩX, as the closure of X, or rather, of
its image under the following composition of maps
X Φ−→ sp(DX)
Ψ−→ ΩX ⊆ 2F.
The map described above will acquire a special relevance in what follows, so it deserves
a special notation:
partial actions and subshifts
21
7.9. Definition. We will denote by
Ξ : X → ΩX
the map given by the above composition, namely Ξ = Ψ ◦ Φ.
Unraveling the appropriate definitions it is easy to see that
Ξ(x) = {g ∈ F : eg(δx) = δx}.
Notice that to say that eg(δx) = δx is the same as saying that that δx lies in the final space
of ug, which we have seen to be ℓ2(Xg). Thus we may alternatively describe Ξ(x) as
Ξ(x) = {g ∈ F : x ∈ Xg}.
(7.10)
We will now give a further, more detailed, description of Ξ(x).
7.11. Proposition. Given x in X, let
ξx = Ξ(x).
Then ξx consists precisely of the elements g in F such that the following conditions hold:
(i) g may be written in reduced form as αβ−1, with α, β ∈ F+,
(ii) α is a prefix of x and, writing x = αy, one has that y ∈ Fα ∩ Fβ.
Proof. Given g ∈ ξx we may use (7.8.v) to write g = αβ−1, and we may clearly assume
that (i) above holds. By (7.10) we have that x ∈ Xg, and hence (ii) follows from (4.2.ii).
Conversely, assuming that g satisfies (i) and (ii), we have again by (4.2.ii) that x lies
(cid:3)
in Xg, proving that g ∈ ξx.
We shall often use the above result in the special case that α = ∅, in which case it
reads:
7.12. Corollary. Given x in X, and β in Λ∗, one has that
β−1 ∈ ξx ⇐⇒ x ∈ Fβ.
It is instructive to view these elements within the Cayley graph of F.
α
y
z
e
x1
x2
................................................•.........................................................•
.........................................................•
.........................................................•
.........................................................•
xn+1
.........................................................•
xn+2
.........................................................•
.........................................................•. . .
}
{
xn
{
.........................................................•
βm
z
...
}
.........................................................•
β2
.........................................................•
...
•.........................................................•
β1
................................................
g = αβ−1
22
m. dokuchaev and r. exel
Given any x in X, the goal is to mark the vertices of the Cayley graph of F corresponding
to the elements of ξx. Due to (7.8.i), we must always mark the unit group element.
Thereafter, beginning at the unit group element we mark all vertices according to the
successive letters of x = x1x2x3 . . . , thus forming the stem of ξx. We then choose an
integer n and focus on the nth vertex along this path, letting
α = x1x2 . . . xn,
and
y = xn+1xn+2 . . . ,
so that x = αy, and y ∈ Fα. Choosing a finite word β, we then back up starting at the
vertex chosen above, along the letters of β. However, before we do this, we must make
sure that the infinite word βy lies in X, which the same as saying that y ∈ Fβ. It is also
best to choose β such that βm 6= xn, since this will avoid stepping on a vertex we have
already traversed, guaranteeing (7.11.i).
The group element g = αβ−1 is therefore an element of ξx and, as seen in (7.11), all
elements of ξx arise in this way.
One may also think of the stem of ξx as a river , the β's considered in the above
diagram being its tributaries, while ξx consists of the whole river basin.
In fact, not only the ξx, but every ξ in ΩX has an interpretation as a river basin, but
first we need to identify the appropriate rivers.
7.13. Proposition. Let ξ be in ΩX and let g ∈ ξ. Then there exists a unique x in X
such that
ξ ∩ gF+ = {gα : α is a prefix of x}.
Proof. By induction and (7.8.iii), there exists an infinite sequence x = x1x2x3x4 . . . , with
xi ∈ Λ, such that
gx1x2 . . . xn ∈ ξ,
∀ n ∈ .
Using (7.8.iv) we have that x1x2 . . . xn ∈ LX, for every n, so it follows from (2.3) that
x ∈ X. The inclusion "⊇" relative to the sets in the statement then clearly holds.
On the other hand, given any gα ∈ ξ ∩ gF+, we claim that α is a prefix of x. In order
to prove this, suppose otherwise, and let α′ be the shortest prefix of α which is not a prefix
of x. By (7.8.ii) we have that gα′ ∈ ξ, which is to say that we may assume without loss of
generality that α is already minimal. Write α = y1y2 . . . yn, with yi in Λ, so that
β := y1y2 . . . yn−1
is a prefix of x by minimality. We then have that βxn is a prefix of x, whence gβxn lies in
ξ. But gα = gβyn also lies in ξ, so xn = yn, by (7.8.iii), whence
α = βyn = βxn
is a prefix of x, a contradiction.
(cid:3)
The following concept is reminiscent of [17: Definition 5.5].
7.14. Definition. Given ξ in ΩX and g in ξ, the unique x in X satisfying the conditions
of (7.13) will be called the stem of ξ at g, and it will be denoted by σg(ξ). In the special
case that g = 1, we will refer to x simply as the stem of ξ, denoting it by σ(ξ).
partial actions and subshifts
23
Given x in X, we have that all prefixes of x lie in ξx by (7.11). From this it immediately
follows that the stem of ξx is precisely x. In symbols
σ(ξx) = x.
This shows the following:
7.15. Proposition. The stem, viewed as a map
σ : ΩX 7→ X,
is a left inverse for Ξ. Consequently Ξ is one-to-one and σ is onto.
We have already hinted at the fact that the map
Φ : x ∈ X 7→ ϕx ∈ sp(DX)
may not be continuous, in which case neither is Ξ. Fortunately, not all of the maps in
sight are discontinuous:
7.16. Proposition. The stem defines a continuous mapping from ΩX to X.
Proof. Given a net {ξi}i in ΩX converging to some ξ, let
xi = σ(ξi),
and
x = σ(ξ).
In order to prove the statement we need to show that {xi}i converges to x.
By the definition of the product topology on X ⊆ Λ, given any neighborhood U of
x, we may find a cylinder Zα, with
x ∈ Zα ⊆ U.
It follows that α is a prefix of x, whence α belongs to ξ.
By the definition of the product topology on ΩX ⊆ 2F, for every g in F, the function
η ∈ ΩX 7→ [g ∈ η] ∈ {0, 1},
(7.16.1)
where the brackets correspond to Boolean value, is continuous. Therefore
1 = [α ∈ ξ] = lim
i
[α ∈ ξxi],
which means that α ∈ ξxi for all sufficiently large i. Consequently α is a prefix of xi, so
xi ∈ Zα ⊆ U.
This concludes the proof.
(cid:3)
24
m. dokuchaev and r. exel
Although σ is onto, it might not be one-to-one. This is to say that there may be many
different elements in ΩX with the same stem. An example of this situation is obtained by
taking any ξ together with ξx, where x = σ(ξ). The following result further explores the
relationship between these two elements.
7.17. Proposition. Given ξ in ΩX , let x = σ(ξ). Then ξ ⊆ ξx.
Proof. Given g in ξ, we may write g = αβ−1, with α, β ∈ F+ by (7.8.v), and we may clearly
assume that g is in reduced form, so (7.11.i) holds. By (7.8.ii) we have that α ∈ ξ ∩ F+,
so α is a prefix of x.
Write α = x1x2 . . . xn, and x = αxn+1xn+2 . . ., and notice that, for every integer
k > n, one has that αxn+1 . . . xk is a prefix of x, so
ξ ∋ αxn+1 . . . xk = αβ−1βxn+1 . . . xk = gβxn+1 . . . xk.
From (7.8.iv) it follows that βxn+1 . . . xk lies in LX , for every k, whence the infinite
word
lies in X by (2.3). We deduce that the infinite word
z = βxn+1xn+2 . . .
y = xn+1xn+2 . . .
lies in Fβ and clearly also in Fα. This proves (7.11.ii), so
g = αβ−1 ∈ ξx.
(cid:3)
As a consequence we see that the river basin picture of the ξx also applies to a general
element ξ in ΩX . That is, if x = σ(ξ), then ξ contains the whole "river" x, while it is
contained in the "river basin" ξx by (7.17).
In particular, conditions (7.11.i-ii), which are necessary and sufficient for a given
group element g to lie in ξx, are seen to still be necessary for membership in ξ, as long as
we take x to be the stem of ξ. By this we mean that when g ∈ ξ, then necessarily g ∈ ξx,
whence said conditions hold.
Incidentally, there is a special case in which these conditions are also sufficient, as we
shall now see.
7.18. Proposition. Given ξ in ΩX , suppose that α and β are elements of F+ satisfying
conditions (7.11.i-ii) relative to x = σ(ξ). Suppose, in addition, that the element y referred
to in condition (7.11.ii) actually lies in the interior of Fβ. Then αβ−1 ∈ ξ.
Proof. Since Φ(X) is dense in sp(DX), we have that Ξ(X) is dense in ΩX , so we may write
ξ = limi ξxi, with xi in X. By hypothesis we have that α is a prefix of x, so α ∈ ξ, and by
the continuity of the maps described in (7.16.1), we have that α ∈ ξxi, for all sufficiently
large i. It follows that α is a prefix of xi, so
xi = αyi,
partial actions and subshifts
25
for some infinite word yi, necessarily belonging to Fα. Since xi converges to x by (7.16),
we have that yi converges to y, which belongs to the interior of Fβ, by hypothesis. So
the yi lie in Fβ, again for all sufficiently large i. We conclude that αβ−1 satisfies (7.11.ii)
relative to all such xi, whence αβ−1 ∈ ξxi . By continuity of Boolean values we then have
[αβ−1 ∈ ξ] = lim
i
[αβ−1 ∈ ξxi] = 1,
so αβ−1 ∈ ξ.
(cid:3)
This result has the following important consequence (see also [8: Remark 1.1.5]):
7.19. Theorem. Given a subshift X, consider the mapping
Ξ : x ∈ X 7→ ξx ∈ ΩX,
already defined in (7.9). Then the following are equivalent:
(i) X is a subshift of finite type,
(ii) Ξ is onto,
(iii) Ξ is continuous,
(iv) Ξ is a homeomorphism.
Proof. (i) ⇒ (ii). Given ξ in ΩX , let x be its stem, and we claim that fact ξ = ξx. On the
one hand we have that ξ ⊆ ξx by (7.17). In order to prove the reverse inclusion, pick g
in ξx. By (7.11) we may find a decomposition g = αβ−1 satisfying conditions (7.11.i-ii) so
that, among other things, x = αy, with y ∈ Fβ.
Since X is of finite type, we have by (2.5) that Fβ is open, so y automatically lies in
the interior of Fβ, and we deduce from (7.18) that g = αβ−1 ∈ ξ. This shows that ξ = ξx,
and hence that Ξ is onto.
(ii) ⇒ (iii). Recall from (7.15) that σ is one-to-one. If we assume, in addition, that Ξ is
onto, then Ξ is an invertible map whose left-inverse σ is also its two-sided inverse and
hence invertible. Since σ is a continuous invertible map defined on a compact set, it must
be a homeomorphism, so its inverse, namely Ξ, is then seen to be continuous.
(iii) ⇒ (i). Observe that (7.12) allows for the following description of Fβ:
Fβ = {x ∈ X : β−1 ∈ ξx} = Ξ−1(cid:0){ξ ∈ ΩX : β−1 ∈ ξ}(cid:1),
which is then the inverse image of an open set under the continuous mapping Ξ. We
conclude that Fβ is open and then (i) follows from (2.5).
We have not taken (iv) into account so far, but it is clear that it implies either (ii) or
(cid:3)
(iii), while (ii)+(iii) easily implies (iv) since Ξ is one-to-one by (7.15).
26
m. dokuchaev and r. exel
8. The spectral partial action.
Recall from (4.1) that θ is a partial action of F on X which might however not be a topo-
logical partial action in the sense that not all of the Xg are open sets. To supersede this
badly behaved partial action we will now show that there exists a fully compliant topolog-
ical partial action of F on ΩX extending θ. The idea will be to build an algebraic partial
action on DX and then consider the corresponding action at the level of the spectrum.
For each g in F, let Dg be the two-sided ideal of DX generated by eg, namely
and let τg be the map from Dg−1 to Dg given by
Dg = egDX ,
τg(a) = ugaug−1,
∀ a ∈ Dg−1 .
As in [16: 10.1] one may show that
(8.1)
(8.2)
τ =(cid:0){Dg}g∈F, {τg}g∈F(cid:1)
is a C*-algebraic partial action (see [16: 11.4]) of F on DX . In fact [16: 10.1] is proved
in a purely algebraic context, but the proof given there carries over to the C*-algebraic
setting.
By [16: 11.6], there exists a topological partial action
ϑ =(cid:0){Ωg}g∈F, {ϑg}g∈F(cid:1)
of F on ΩX (here identified with the spectrum of DX ) linked to α via the fact that Dg
consists of the elements of DX (here identified with the algebra C(ΩX ) of all continuous
complex valued functions on ΩX by Gelfand's Theorem) vanishing off Ωg, plus the relation
for every g ∈ F, f ∈ Dg−1 , and ξ ∈ Ωg.
τg(f )ξ = f(cid:0)ϑg−1(ξ)(cid:1),
8.3. Definition. We will refer to the partial action ϑ introduced above as the spectral
partial action associated to X.
As we shall see, the spectral partial action may be seen as a the restriction of the
partial Bernoulli action of F [16: 5.12].
8.4. Proposition. Regarding the spectral partial action associated to X, one has that
(i) Ωg = {ξ ∈ ΩX : g ∈ ξ},
(ii) ϑg(ξ) = gξ = {gh : h ∈ ξ},
for each g in F, and for each ξ ∈ Ωg−1.
partial actions and subshifts
27
Proof. Under the well known correspondence between closed two-sided ideals in C(ΩX )
and open sets in ΩX , notice that the ideal generated by an idempotent element corresponds
to the support of the latter. In case of the idempotent eg, its support, initially viewed in
sp(DX), corresponds to the set of all characters ϕ such that ϕ(eg) = 1. Identifying sp(DX)
and ΩX by (7.7), and noting that
ϕ(eg) = 1
(7.5)
⇐⇒ g ∈ ξϕ,
for all characters ϕ, one sees that (i) follows.
Given g in F, and ξ ∈ Ωg−1, let ϕ be the character on DX corresponding to ξ under Ψ,
namely such that ξ = ξϕ. Since eg−1 ∈ ξ, by (8.4.i), we have that ϕ(eg−1) = 1. Moreover,
ϑg(ξ) will correspond to a character ψ, such that ψ(eg) = 1, and
For any a in DX , regardless of whether a is in Dg or not, we have that aeg ∈ Dg, whence
ψ(a) = ϕ(cid:0)τg−1(a)(cid:1),
∀ a ∈ Dg.
ψ(a) = ψ(eg)ψ(a) = ψ(ega) = ϕ(cid:0)τg−1(ega)(cid:1).
We then have that ϑg(ξ) = ξψ, so for any h in F, one has that
h ∈ ϑg(ξ) ⇐⇒ h ∈ ξψ
(7.5)
⇐⇒ ψ(eh) = 1.
(8.4.1)
Incidentally notice that
ψ(eh) = ϕ(cid:0)τg−1(eheg)(cid:1) = ϕ(cid:0)ug−1 ehegug)(cid:1) = ϕ(cid:0)eg−1heg−1(cid:1) = ϕ(eg−1h),
because ϕ(eg−1) = 1. Focusing on (8.4.1), we then have that
ψ(eh) = 1 ⇐⇒ ϕ(eg−1h) = 1 ⇐⇒ g−1h ∈ ξ ⇐⇒ h ∈ gξ,
proving that ϑg(ξ) = gξ, as desired.
We then have partial dynamical systems
θ =(cid:0){Xg}g∈F, {θg}g∈F(cid:1),
and
ϑ =(cid:0){Ωg}g∈F, {ϑg}g∈F(cid:1)
(cid:3)
(8.5)
on X and ΩX, respectively, and it is interesting to notice that these sets are related to
each other by the maps
Ξ : X → ΩX,
and
σ : ΩX → X,
introduced in (7.9) and (7.14), respectively.
8.6. Proposition. Both Ξ and σ are equivariant maps [16: 2.7] relative to the partial
dynamical systems in (8.5).
28
m. dokuchaev and r. exel
Proof. Let us first prove that
Ξ(Xg) ⊆ Ωg,
∀ g ∈ F.
Given x ∈ Xg, we have by (7.10) that g ∈ Ξ(x), whence Ξ(x) ∈ Ωg, by (8.4.i), proving the
above inclusion. We next must show that
(8.6.1)
for all g in F, and all x ∈ Xg−1. Given such an x, let y = θg(x). Then evidently y ∈ Xg,
and for any h ∈ F, one has that
Ξ(cid:0)θg(x)(cid:1) = ϑg(cid:0)Ξ(x)(cid:1),
h ∈ Ξ(y)
(7.10)
⇐⇒ y ∈ Xh ⇐⇒ y ∈ Xh ∩ Xg ⇐⇒ θg(x) ∈ Xh ∩ Xg ⇐⇒
⇐⇒ x ∈ θg−1(Xh ∩ Xg) = Xg−1h ∩ Xg−1 ⇐⇒ x ∈ Xg−1h
(7.10)
⇐⇒ g−1h ∈ Ξ(x).
From this we see that
so
g−1Ξ(y) = {g−1h : h ∈ Ξ(y)} = Ξ(x),
showing (8.6.1), and hence concluding the proof of the equivariance of Ξ.
Ξ(cid:0)θg(x)(cid:1) = Ξ(y) = gΞ(x)
(8.4.ii)
= ϑg(cid:0)Ξ(x)(cid:1),
To prove that
σ(Ωg) ⊆ Xg,
∀ g ∈ F,
−1
+ , then g /∈ ξ for every ξ in ΩX , by (7.8.v). It follows that Ωg
notice that if g is not in F+F
is empty and then there is nothing to prove. Otherwise write g = αβ−1, with α, β ∈ F+,
in reduced form. Given ξ in Ωg, we have that
g ∈ ξ
(7.17)
⊆ ξx,
where x = σ(ξ). It then follows from (7.11) that x = αy, where y ∈ Fα ∩ Fβ, which implies
that x ∈ Xg, by (4.2.ii). This proves that σ(Ωg) ⊆ Xg, as desired.
We must finally prove that
σ(cid:0)ϑg(ξ)(cid:1) = θg(cid:0)σ(ξ)(cid:1),
for all g in F, and all ξ in Ωg−1. Again there is nothing to do unless g lies in F+F
we may assume that g = αβ−1, with α, β ∈ F+, in reduced form.
Since g−1 = βα−1 ∈ ξ, we have that β ∈ ξ, by (7.8.ii) so also β ∈ ξ ∩ F+. By (7.13)
the latter set consists precisely of the prefixes of the stem of ξ, which we will denote by x
from now on. In particular we have that x = βy, for some infinite word y, and we have by
(4.2.iii) that
−1
+ , so
On the other hand, observe that if γ is any prefix of y, then βγ is a prefix of x, so βγ
lies in ξ. Consequently
Since αγ ∈ F+ for any such γ, one sees that αγ is a prefix of the stem of ϑg(ξ), from where
it follows that
(cid:3)
In case X is a subshift of finite type the above result may be combined with (7.19)
leading up to the following:
θg(cid:0)σ(ξ)(cid:1) = θαβ−1(cid:0)βy(cid:1) = αy.
αγ = gβγ ∈ gξ = ϑg(ξ).
σ(cid:0)ϑg(ξ)(cid:1) = αy = θg(cid:0)σ(ξ)(cid:1).
partial actions and subshifts
29
8.7. Proposition. If X is a subshift of finite type then the maps
σ : ΩX → X,
and
Ξ : X → ΩX
are mutually inverse equivariant homeomorphisms, whence the spectral partial action and
the standard partial action are equivalent.
Let us conclude this section with an important technical result, regarding membership
of an element of the form gβ−1, when we already know that g is a member of a given ξ.
This result actually belongs in section (7), but it was delayed up to now since its proof is
greatly facilitated by the existence of the spectral action.
8.8. Proposition. Given ξ in ΩX , pick g ∈ ξ, and let β be a finite word. Regarding the
statements:
(a) gβ−1 ∈ ξ, and
(b) σg(ξ) ∈ Fβ,
.........................................................•. . .
.........................................................•
................................................•.........................................................•
.........................................................•
.........................................................•
βm
g
x1
x2
x3
...
.........................................................•
β2
.........................................................•
hβ1
•.........................................................•
β1
................................................
h=gβ−1
one has that (a) ⇒ (b). Moreover, if ξ = ξx for some x in X, then (b) ⇒ (a) as well.
Proof. Let h = gβ−1, let x = σg(ξ), and write
x = x1x2x3 . . . ,
and
β = β1β2 . . . βm.
Notice that hβ1 lies in the segment joining h and g so, supposing (a), we have by
(7.8.ii) that hβ1 ∈ ξ. Likewise
hβ1β2 . . . βi ∈ ξ,
for all i. Furthermore, for every integer j we have that
hβ1β2 . . . βmx1x2 . . . xj = gx1x2 . . . xj ∈ ξ,
by the definition of the stem of ξ at g. It then follows that σh(ξ) = βx, so in particular
βx ∈ X, whence x ∈ Fβ, proving (b).
In order to prove the last sentence of the statement, let ξ = ξx for some x in X, and
suppose that (b) holds. In the special case that g = 1, we have that
x = σ(ξx) = σg(ξ) ∈ Fβ,
so β−1 belongs to ξ by (7.12), proving (a).
Dropping the assumption that g is the unit group element, let us deal with the general
case. Observing that g ∈ ξx, we have by (7.10) that x ∈ Xg, so
y := θg−1(x)
30
m. dokuchaev and r. exel
is well defined. Moreover, by (8.6)
gξy = ϑg(ξy) = ξθg (y) = ξx.
In other words, ξx is obtained by left-translating ξy by g. It then easily follows that
σ(ξy) = σg(ξx) ∈ Fβ,
so the first case treated above (i.e. g = 1) applies for ξy, and we deduce that β−1 ∈ ξy,
whence
gβ−1 ∈ gξy = ξx.
(cid:3)
The above result plays a crucial role in understanding the elements of ΩX from the
point of view of their stem. By this we mean that, once the stem of ξ is marked in the
Cayley graph of F, and we wish to mark the remaining group elements in ξ, we know from
(7.8.v) that we need only worry about elements of the form g = αβ−1. If we are careful to
take the reduced form of g, then a necessary condition for it to be marked is that α also
be marked, in which case α must be a prefix of the stem.
α
z
e
x1
x2
................................................•.........................................................•
.........................................................•
.........................................................•
.........................................................•
xn+1
.........................................................•
xn+2
.........................................................•
.........................................................•. . .
xn
{
.........................................................◦
βm
...
}
.........................................................◦
β2
.........................................................◦
...
◦.........................................................◦
β1
g = αβ−1
We then must decide whether or not to mark g itself, and this is precisely where (8.8)
intervenes: in case ξ is some of the ξx, then we should mark g if and only if the resulting
stem at g, namely
β1β2 . . . βmxn+1xn+2 . . .
lies in X (which is to say that xn+1xn+2 . . .
lies in Fβ). When ξ is not necessarily a ξx,
then (8.8) does not give a definite answer, except that marking g is forbidden in case the
above infinite word does not lie in X.
At this point it is perhaps useful to discuss an example:
it is well known that the
even shift is not of finite type3, and hence (7.19) predicts the existence of elements in ΩX
beyond the range of Ξ, meaning not of the form ξx. In what follows we will concretely
exhibit an example of such anomalous elements.
Let X be the even shift, and for each n, consider the infinite word
3 This will also follow from the analysis we are about to undertake.
xn = 1
2n+1
∞
0
= 1 . . . 1
0000 . . .
2n+1
{z }
partial actions and subshifts
31
Since ΩX is compact, there exists a subsequence, say {yk}k = {xnk }k, such that {ξyk }k,
converges to some ξ ∈ ΩX. Our next goal will be to prove that ξ is not of the form ξx, for
any x in X.
The fact that {xn}n, and hence also {yk}k converges to the infinite word
1∞ = 1111 . . .
relative to the topology of X, does not imply that ξyk converges to ξ1∞, as the correspon-
dence x → ξx is not known to be continuous (it will soon be evident that it is discontinuous
at 1∞). Nevertheless, the continuity of the stem (7.16) implies that
σ(ξ) = lim
k
σ(ξyk) = lim
k
yk = 1∞.
Therefore, if ξ = ξx, for some x, then
1∞ = σ(ξ) = σ(ξx) = x.
So, to prove that ξ is not equal to any ξx we therefore only need to verify that ξ 6= ξ1∞.
Observing that yk is not in the follower set of the finite word
we have by (7.12) that β−1 is not in ξyk . By (7.16.1) it follows that
[β−1 ∈ ξ] = lim
k
[β−1 ∈ ξyk ] = 0,
β = '0',
so β−1 /∈ ξ. Nevertheless, 1∞ does belong to the follower set of β, hence β−1 ∈ ξ1∞, again
by (7.12). This proves that ξ 6= ξ1∞, so ξ is not in the range of Ξ, whence Ξ is not onto,
and we then deduce from (7.19) that Ξ is not continuous. We also recover the well known
fact that the even shift is not of finite type.
This example also illustrates that the implication "(b) ⇒ (a)" in (8.8) may indeed fail,
since the stem of ξ lies in Fβ, and yet β−1 is not in ξ.
9. Partial crossed product description of OX .
As always we fix a finite alphabet Λ and a subshift X ⊆ Λ. In this section we plan to
prove that OX is isomorphic to the crossed product of C(ΩX ) (also known as DX) by the
spectral partial action of the free group F. In symbols
OX ≃ C(ΩX ) ⋊ϑ F.
(9.1)
We will also show that the associated semi-direct product Fell bundle is amenable
[12], whence the full and reduced crossed products coincide.
Regarding the partial action τ of F on DX introduced in (8.2), recall that ϑ was
defined as the partial action on the spectrum of DX induced by τ . Thus, moving in the
opposite direction, the partial dynamical system induced by ϑ on C(ΩX ) is equivalent to
τ . In order to prove (9.1), it therefore suffices to prove that
OX ≃ DX ⋊τ F.
9.2. Proposition. There exists a surjective *-homomorphism
ϕ : DX ⋊τ F → OX ,
such that ϕ(aδg) = (a ⊗ 1)ug = aug ⊗ λg, for all g in F, and every a ∈ Dg.
32
m. dokuchaev and r. exel
Proof. We claim that the pair (j, u) is a covariant representation (see [16: 9.10]) of τ in
OX , where
j : a ∈ DX 7→ a ⊗ 1 ∈ OX .
To see this we pick any g in F, and a in Dg−1, and compute
ugj(a)ug−1 = (ug ⊗ λg)(a ⊗ 1)(ug−1 ⊗ λg−1 ) = ugaug−1 ⊗ 1
(8.1)
= τg(a) ⊗ 1 = j(cid:0)τg(a)(cid:1).
This shows that indeed (j, u) is a covariant representation, so the existence of ϕ follows
from [16: 13.1]. Given α in Λ∗, notice that eα ∈ Dα, so
ϕ(eαδα) = eαuα ⊗ λα = uα ⊗ λα = uα.
This shows that every uα lies in the range of ϕ, whence ϕ is onto.
(cid:3)
In order to find a map in the opposite direction, let us consider the standard condi-
tional expectation E from the algebra of all bounded operators on ℓ2(X) ⊗ ℓ2(F) onto the
subalgebra of all diagonal operators relative to the standard orthonormal basis. Thus, if
t is any bounded operator on ℓ2(X) ⊗ ℓ2(F), then E(t) is the operator whose off diagonal
entries are zero, and whose diagonal entries are the same as those of t.
9.3. Lemma. Let t = eh1 eh2 . . . ehn ug, where h1, h2, . . . , hn, g ∈ F. Then
E(t) =(cid:26) t,
if g = 1,
0, otherwise.
Proof. We have
E(t) = E(eh1 eh2 . . . ehn ug) = eh1 eh2 . . . ehnE(ug),
because the eh's are diagonal operators by (7.2). When g = 1 we have that ug = 1, so it
is easy to see that E(t) = t. When g 6= 1, notice that for all x ∈ X, and h ∈ F,
hug(δx ⊗ δh), δx ⊗ δhi = hug(δx), δxihδgh, δhi = 0,
since gh 6= h. Thus all diagonal entries for ug vanish, whence E(ug) = 0, concluding the
proof.
(cid:3)
Observe that the elements of the form eh1 eh2 . . . ehn ug span a *-subalgebra of OX by
[16: 9.8]. It is therefore a dense subalgebra, and (9.3) then implies that OX is invariant
under E. Again by (9.3) one sees that the
E(OX) ⊆ DX ⊗ 1,
so we may see E as a conditional expectation from OX to DX ⊗ 1.
In our next result we will refer to the reduced partial crossed product of DX by F
under τ (see [16: 17.10]), which we will denote by DX ⋊red
τ F.
partial actions and subshifts
33
9.4. Proposition. There exists a *-homomorphism
ψ : OX → DX ⋊red
τ F,
such that ψ(uα) = eαδα, for all α in Λ∗.
Proof. For each g in F, let Bg be the closed linear subspace of OX given by
Bg = (Dg ⊗ 1)ug = (Dgug ⊗ λg).
We claim that BgBh ⊆ Bgh, for all g, h ∈ F. In fact, given a ∈ Dg, and b ∈ Dh, we
have
while
(aug ⊗ λg)(buh ⊗ λh) = augbuh ⊗ λgh,
(9.4.1)
augbuh = egaugbeg−1 uh = ugug−1 augbug−1uguh
[16:9.1.iii]
= ugτg−1(a)bug−1ugh =
= τg(τg−1(a)b(cid:1)ugh ∈ τg(Dg−1 ∩ Dh)ugh ⊆ Dghugh.
Therefore the element described in (9.4.1) lies in Bgh, proving the claim. We leave it for
Bg is a dense subspace of
OX , which, when combined with the conditional expectation E, provided above, verifies
all of the assumptions of [16: 19.1], which in turn provides for the desired map ψ.
(cid:3)
the reader to prove that (Bg)∗ = Bg−1, as well as that Pg∈F
Combining the above with (9.2) we arrive at the main result of this section:
9.5. Theorem. Let Λ be a finite alphabet, and let X ⊆ Λ be a subshift. Then:
(i) The semi-direct product bundle [16: 16.6] corresponding to the spectral partial action
satisfies the approximation property [16: 20.4], and hence is amenable [16: 20.1].
(ii) The Carlsen-Matsumoto C*-algebra OX is naturally isomorphic to both the full and
the reduced crossed product of C(ΩX) by the free group F(Λ) under the spectral
partial action. In symbols
OX ≃ C(ΩX ) ⋊ϑ F
≃ C(ΩX ) ⋊red
ϑ F.
Proof. The first point follows from [13: Theorems 4.1 & 6.3] (see also [16: 20.13]).
As for (ii), regarding the maps ϕ and ψ provided by (9.2) and (9.4), it is easy to see that
the composition ψ ◦ ϕ is the regular representation [16: 17.6] of C(ΩX ) ⋊ϑ F, which is an
isomorphism by (i). In particular ϕ is one-to-one, but since we already saw that it is onto in
(9.2), we deduce that it is an isomorphism. If both ϕ and the composition ψ ◦ϕ are isomor-
phisms, then so is ψ, whence OX is isomorphic to the reduced crossed product as well. (cid:3)
In the introduction of [7] Carlsen argues that, contrary to OX , Matsumoto's algebra
MX does not have good universal properties and hence they should be considered as the
class of reduced C*-algebras associated with subshifts.
Interpreting the term reduced as
one usually does when speaking of crossed products, the above result seems to indicate
that OX should be seen as both full and reduced C*-algebras and that MX is just an
epimorphic image of OX .
34
m. dokuchaev and r. exel
10. Comparison with Carlsen's description of OX .
As always we fix a finite alphabet Λ and a subshift X ⊆ Λ.
In this section we will prove the fact, already hinted at after (6.4), that OX is isomor-
phic to the algebra introduced by Carlsen in [7: Definition 5.1].
We have already observed that the eg are diagonal operators relative to the canonical
orthonormal basis of ℓ2(X). On the other hand, the algebra of all diagonal operators is
clearly isomorphic to ℓ∞(X), so we may see DX as a subalgebra of ℓ∞(X).
10.1. Proposition. For all finite words α and β in Λ∗, let
C(β, α) = {αy ∈ X : y ∈ Fα ∩ Fβ}.
Then DX coincides with the closed *-subalgebra of ℓ∞(X) generated by the characteristic
functions 1C(β,α), as α and β range in Λ∗.
Proof. In case α and β are such that αβ−1 = α + β−1, that is, when g := αβ−1 is in
reduced form, we have seen in (4.2.ii) that Xg = C(β, α). Since eg is the final projection
of ug, whose range is ℓ2(Xg) by (5.3), we have that eg coincides with 1C(β,α) (up to the
above identification of diagonal operators and bounded functions). For general α and β,
let γ be the longest common suffix of α and β, so we may find α′ and β′ such that
α = α′γ,
β = β′γ,
and α′β′−1 is the reduced form for the group element g := αβ−1. We leave it for the reader
to check that
C(β, α) = C(β′, α′) ∩ C(∅, α),
from where it follows that
1C(β,α) = 1C(β ′,α′)1C(∅,α) = egeα = eαβ−1 eα.
(10.1.1)
We therefore conclude that the algebra generated by all of the 1C(β,α) is the same as the
algebra generated by all of the eg.
(cid:3)
As a consequence we see that DX is the same as the algebra denoted DX studied in
[7: Definition 4.1], or the algebra denoted DX studied in [10: Lemma 7], and which also
appears in many other papers dealing with C*-algebras associated with subshifts (see the
references given in the introduction).
10.2. Theorem. For every subshift X, denote by O′
Carlsen in [7: Definition 5.1]. Then there is an isomorphism
X the C*-algebra introduced by
ϕ : O′
X → OX ,
such that ϕ(Sα) = uα, for all α in Λ∗, where Sα is the partial isometry given in [7:
Definition 5.3].
partial actions and subshifts
35
Proof. For each α ∈ Λ∗, let
Tα = uα = uα ⊗ λα,
and notice that for any given α, β ∈ Λ∗, we have
TαT ∗
β TβT ∗
α = uαeβ−1 uα−1
[16:9.8.iii]
= eαβ−1 uα uα−1 =
= eαβ−1 eα = (eαβ−1eα) ⊗ 1
(10.1.1)
= 1C(β,α) ⊗ 1.
Using the universal property of O′
X [7: Remark 7.3], we see that there exists a *-
X → OX sending each Sα to Tα, and which is therefore necessarily
homomorphism ϕ : O′
onto.
In order to complete the proof it now suffices to prove that ϕ is injective, and for this
we will employ [10: Theorem 13], which demands that we build a suitable action of the
circle group T on OX . Consider the unique group homomorphism
ε : F → Z,
such that ε(a) = 1, for every a ∈ Λ, so that in particular ε(α) = α, if α is a finite
word. Moreover, for each z in T, let Vz be the unitary operator defined on the canonical
orthonormal basis of ℓ2(X) ⊗ ℓ2(F) by
Vz(δx ⊗ δg) = zε(g)δx ⊗ δg,
∀ x ∈ X,
∀ g ∈ F.
We then claim that
VzTαVz−1 = zαTα,
∀ z ∈ T,
∀ α ∈ Λ∗.
(10.2.1)
To prove it we compute on a general element δx ⊗ δg of the canonical orthonormal basis:
VzTαVz−1 (δx ⊗ δg) = z−ε(g)Vz(uα ⊗ λα)(δx ⊗ δg) = z−ε(g)Vz(cid:0)uα(δx) ⊗ δαg(cid:1) =
= z−ε(g)+ε(αg)(cid:0)uα(δx) ⊗ δαg(cid:1) = zε(α)Tα(δx ⊗ δg).
This proves (10.2.1), so it follows that the formula
γz(a) = VzaVz−1,
∀ x ∈ T,
∀ a ∈ OX ,
defines a strongly continuous action of T on OX satisfying point (2) of [10: Theorem 13],
whence ϕ is injective.
(cid:3)
36
m. dokuchaev and r. exel
11. The topology on the spectrum.
The topology on ΩX , being induced from the product topology of 2F, admits a basis
formed by the open sets
Ug1,g2,...,gn; h1,h2,...,hm =(cid:26) η ∈ ΩX : g1 ∈ η, g2 ∈ η,
h1 /∈ η, h2 /∈ η, . . . , hm /∈ η(cid:27) ,
. . . , gn ∈ η
(11.1)
where g1, g2, . . . , gn; h1, h2, . . . , hm range in F. Given the special nature of elements of
ΩX , we may restrict to sets of a somewhat special nature, as follows.
11.2. Proposition. Given α, β1, β2, . . . , βn, γ1, γ2, . . . , γm ∈ F+, consider the subset of
ΩX given by
Vα; β1,β2,...,βn; γ1,γ2,...,γm =
η ∈ ΩX :
α
i
αβ−1
αγ−1
j
∈ η,
∈ η,
/∈ η,
for i = 1, . . . , n,
for j = 1, . . . , m
,
Then the collection consisting of all sets of the above form is a basis for the topology of
ΩX .
Proof. Let A be an open subset of ΩX, and let ξ ∈ A. It suffices to prove that there is an
open set V of the form described in the statement with ξ ∈ V ⊆ A.
By the definition of the product topology on 2F there are g1, g2, . . . , gn; h1, h2, . . . ,
hm ∈ F, as in (11.1), such that
ξ ∈ Ug1,g2,...,gn; h1,h2,...,hm ⊆ A.
(11.2.1)
Since the elements of ΩX only contain group elements of the form αβ−1, with α, β ∈
F+ by (7.8.v), we may ignore the gi and the hj which are not of this form without affecting
(11.2.1). We may therefore assume that
gi = αiβ−1
i
,
and
hj = µjγ−1
j
,
in reduced form, with αi, βi, µj, γj ∈ F+. Let α be any prefix of the stem of ξ, long enough
so that
and observe that since α ∈ ξ, then
α ≥ max(cid:8)α1, α2, . . . , αn, µ1, µ2, . . . , µm(cid:9),
(11.2.2)
ξ ∈ Uα,g1,g2,...,gn; h1,h2,...,hm ⊆
⊆ U g1,g2,...,gn; h1,h2,...,hm ⊆ A.
The reader is asked to compare the display above with (11.2.1), paying special atten-
tion to the important detail that α was inserted ahead of the gi's in the subscripts of the
first occurrence of U above. The proof will consist in showing that this occurrence of U
coincides with the set displayed in the statement for suitable choices of βi and γj.
partial actions and subshifts
37
Notice that, by assumption the gi ∈ ξ, so (7.8.ii) implies that αi ∈ ξ, and then the αi
are necessarily prefixes of the stem of ξ by (7.13). Consequently the αi are also prefixes of
the long α chosen above, and we may find suitable elements δi in F+, such that
for all i. Letting β′
i = βiδi, observe that
α = αiδi,
αβ′
i
−1 = αδ−1
i β−1
i = αiβ−1
i = gi,
(11.2.3)
so, upon replacing each βi by β′
i
noticed that this presentation of gi is no longer in reduced form.
i, we may suppose that gi = αβ−1
. It should however be
Next we should treat the hj. Firstly, let us consider those hj whose corresponding µj
is not a prefix of the stem of ξ. We then claim that, for any η in ΩX, one has that
α ∈ η ⇒ µj /∈ η.
Otherwise, if both α and µj lie in η, then both would be prefixes of the stem of η, in which
case µj would be a prefix of α, since the former it is shorter than the latter by (11.2.2).
This would entail that µj is a prefix of the stem of ξ, contradicting our assumptions, and
hence proving our claim. By (7.8.ii) we have that
µj /∈ η ⇒ µjγ−1
j
/∈ η,
so a combination of these two implications yields
α ∈ η ⇒ h−1
j
/∈ η.
We then see that hj may be deleted from the list of subscripts of our
Uα,g1,g2,...,gn; h1,h2,...,hm,
(11.2.4)
since the condition "α ∈ η" in the first line of (11.1) already gives "h−1
/∈ η", in the second.
After deleting such hj 's, we may assume that µj is a prefix of the stem of ξ, for every
j, and, again by (11.2.2), µj is necessarily a prefix of α. Arguing as in (11.2.3) we may
then stretch each µj and γj by the same amount, and hence assume that the µj all coincide
with α, so that hj = αγ−1
.
j
j
The description of the set in (11.2.4) is then identical to the description of the set
(cid:3)
displayed in the statement, so we have concluded the proof.
Up to the statement of the above result, whenever we considered an expression of the
form "αβ−1", this was supposed to be in reduced form. However the reader should be
warned that this is no longer the case, especially after (11.2.3), were we deliberately gave
up on reduced forms in exchange for working with a single α.
There is a further simplification which may be bestowed upon the general form of the
open sets described in (11.2), provided we are concerned with neighborhoods of points in
the range of Ξ, namely we may do away with the γj. Making this idea precise is our next
goal.
38
m. dokuchaev and r. exel
11.3. Proposition. Given x in X, let ξx = Ξ(x). Then the collection of all sets of the
form
Vα; β1,β2,...,βn =(cid:26) η ∈ ΩX :
α
αβ−1
i
∈ η,
∈ η,
for i = 1, . . . , n(cid:27) ,
for α, β1, β2, . . . , βn ∈ F+, which moreover contains ξx, forms a neighborhood base for ξx.
Proof. Needless to say, the above sets are special cases of the sets in (11.2), corresponding
to taking m = 0, meaning that the conditions "αγ−1
/∈ η" are now absent.
j
In order to prove the statement, we must show that for every open set U containing
ξx, there are α, β1, β2, . . . , βn ∈ F+ such that
Using (11.2) we may clearly suppose that
ξx ∈ Vα; β1,β2,...,βn ⊆ U.
U = Vα; β1,β2,...,βn; γ1,γ2,...,γm,
for suitable α, β1, β2, . . . , βn, γ1, γ2, . . . , γm ∈ F+. Observing that ξx lies in U , we have
that α ∈ ξx, while αγ−1
/∈ ξx, for every j = 1, . . . , m.
j
α
y
xn+1
.........................................................•
xn+2
.........................................................•
.........................................................•. . .
}
{
z
e
x1
x2
................................................•.........................................................•
.........................................................•
{
z
}
...
xn
.........................................................•
.........................................................•
◦.........................................................◦.........................................................◦.........................................................◦.........................................................◦
αγ−1
j
It follows that α is a prefix of the stem of ξx, also known as x, so we may write x = αy
for some infinite word y. The stem of ξx at α is therefore y and the fact, noted above, that
αγ−1
/∈ ξx, together with (8.8) leads to the conclusion that y is not in the follower set of
γj, which is to say that
j
γjy /∈ X.
By (2.3) we then deduce that γjy has some prefix which is a forbidden word, namely a
finite word not in the language LX. By increasing its length we may suppose that this
forbidden prefix is of the form
γjδj
where δj is a prefix of y. Denote by δ the longer among the δj, and set
α′ = αδ,
and
β′
i = βiδ,
for every i = 1, . . . , n. It follows that
α′β′
i
−1 = αβ−1
i
,
partial actions and subshifts
39
and the proof will be concluded once we show that
ξx ∈ Vα′; β ′
1,β ′
2,...,β ′
n
⊆ Vα; β1,β2,...,βn; γ1,γ2,...,γm.
(11.3.1)
We leave the easy "∈" for the reader to check and concentrate on the "⊆". We thus
pick η in the set appearing in the left-hand-side above, and we note that
(a) α′ ∈ η,
(b) α′β′
j
−1 ∈ η,
for every j = 1, . . . , m. By (a) and (7.8.ii) we have that α ∈ η, and clearly
αβ−1
i = α′β′
j
−1 ∈ η.
Thus, in order to prove that η lies in the set appearing in the right-hand-side of
(11.3.1), we must only check that αγ−1
/∈ η. For this, observe that
j
αγ−1
j = αδ(γjδ)−1 = α′(γjδ)−1.
Assuming by contradiction that this element belongs to η, we have by (a) and (8.8) that
σα′ (η) ∈ Fγj δ.
Letting z = σα′ (η), this means that γjδz ∈ X, but γjδ admits the forbidden word
γjδj as a prefix, a contradiction. This proves our claim that αγ−1
/∈ η, and hence that
j
η ∈ Vα; β1,β2,...,βn; γ1,γ2,...,γm,
showing (11.3.1).
(cid:3)
Speaking of a neighborhood of the form Vα; β1,β2,...,βn , as above, notice that for every
η in this set, one automatically has that
αβ−1 ∈ η,
for β = ∅, as well as for β = α, regardless of whether or not the words ∅ and α are among
the βi's. Therefore, should one so wish, these two words may be added to the βi's without
altering the resulting neighborhood. That is
Vα; β1,β2,...,βn = Vα; ∅,α,β1,β2,...,βn.
40
m. dokuchaev and r. exel
12. Topological freeness.
Recall from [18: Definition 2.1] (see also [3] and [16: Section 29]) that a topological partial
action
of a group G on a space Y is said to be topologically free if, for every g 6= 1, the set of fixed
points for ρg, namely
ρ =(cid:0){Yg}g∈G, {ρg}g∈G(cid:1)
Fixg = {x ∈ Yg−1 : ρg(x) = x},
(12.1)
has empty interior.
In this section we will give necessary and sufficient conditions for a general subshift
to have a topologically free spectral partial action. However, let us begin by reviewing the
well known Markov case.
Observing that Markov subshifts are of finite type, we see that the spectral partial
action is equivalent to the standard partial action by (8.7).
The following result, inspired by condition (I) of [11], characterizes topological freeness
in terms of circuits and exits. It was first proved for row-finite matrices in [20: Lemma
3.4] in the language of groupoids. A generalization for infinite matrices was given in [17:
Proposition 12.2].
12.2. Proposition. Let X be a Markov subshift. Then the standard partial action θ of
F on X (which, by (8.7) is equivalent to the spectral partial action) is topologically free if
and only if every circuit has an exit.
For subshifts not of finite type, the situation is a lot more delicate. Even if all circuits
have strong exits (see (3.4)), the spectral partial action may fail to be topologically free.
To see this, consider the even shift. In section (2) we have proven that all of its circuits
have strong exits, and yet its associated spectral partial action is not topologically free for
the following reason: consider the infinite word
1∞ = 11111 . . . ,
which is fixed by θg, where g is the element4 of F corresponding to the word '1'. By (8.6)
we have that Ξ is equivariant, so ξ1∞ is fixed by ϑg.
We will show that ϑ is not topologically free by showing that ξ1∞ is an isolated point
in ΩX (even though 1∞ is not isolated in X). In fact, let
β1 = 01,
and
β2 = 011,
and consider the open set V∅;β1,β2 described in (11.3). Notice that for any x in X one has
that
ξx ∈ V∅;β1,β2 ⇐⇒ β−1
1 , β−1
2 ∈ ξx
(8.8)
⇐⇒ x ∈ Fβ1 ∩ Fβ2 .
4 Warning: this is one of the generators of F and not the unit group element!
partial actions and subshifts
41
We leave it for the reader to check that the rules of the even shift imply that Fβ1 ∩Fβ2 =
{1∞}, so the only ξx in V∅;β1,β2 is ξ1∞. Since the set of all ξx is dense in ΩX by (7.4), and
since V∅;β1,β2 is open, a simple exercise in Topology gives that
V∅;β1,β2 = {ξ1∞}.
(12.3)
The above is then an open set of fixed points, whence ϑ is not topologically free.
We have therefore proven:
12.4. Proposition. The spectral partial action associated to the even shift is not topo-
logically free.
One might wonder if the standard (as opposed to spectral) partial action θ for the
even shift is topologically free. Although we believe this is not a well posed question, since
the standard partial is not topological (the Xg are not all open), one might decide to ignore
this and insist in checking whether the interior of any set of fixed points is empty. In this
case the answer is easily seen to be positive, so in this sense the standard partial of the
even shift is topologically free.
Before we give the appropriate characterization of topological freeness for the spectral
partial action of general subshifts, let us understand their fixed points a little better. The
following result is entirely similar to known results for Markov subshifts, but we give a full
proof, which takes no more than a few lines, for the convenience of the reader:
12.5. Proposition. Let Λ be a finite alphabet, let X ⊆ Λ be a subshift, and denote by
θ the standard partial action of F on X. Given g ∈ F \ {1}, let x ∈ X be a fixed point
for θg. Then
(i) g admits a decomposition in reduced form as να±1ν−1, with α, ν ∈ F+,
(ii) x = να∞, whence α is a circuit, and x is the unique fixed point for g.
−1
Proof. Since x lies in the domain of θg, that domain is nonempty, whence g ∈ F+F
+
by (4.2.i). We may therefore write g in reduced form as µν−1, with µ, ν ∈ F+. Since
x ∈ Xg−1, we have by (4.2.ii) that x = νy, for some y ∈ Fν ∩ Fµ, and then
νy = x = θµν−1(x) = µy.
Therefore either ν is a prefix of µ, or vice-versa. We assume without loss of generality
that µ > ν (one cannot have µ = ν because g 6= 1), and so we may write µ = να, for
some α ∈ Λ∗. It then follows from the above that νy = ναy, so y = αy, whence y = α∞.
Summarizing we have
g = µν−1 = ναν−1,
and
x = νy = να∞.
completing the proof.
(cid:3)
We may now present the main result of this section. Contrary to one might expect,
this result does not explicitly mention the existence of exits for circuits, but please see the
remarks after the proof below for an interpretation in terms of exits for circuits.
42
m. dokuchaev and r. exel
12.6. Theorem. Let Λ be a finite alphabet and let X ⊆ Λ be a subshift. Then the
following are equivalent:
(i) the spectral partial action ϑ associated to X is topologically free,
(ii) for every β1, β2, . . . , βn in Λ∗, and for every circuit γ such that
γ∞ ∈
Fβi ,
n
Ti=1
one has thatTn
observe that β−1
i ∈ ξγ∞ , for all i, by (7.12), so
i=1 Fβi contains some element other than γ∞.
Proof. (i) ⇒ (ii). Pick β1, β2, . . . , βn in Λ∗, and let γ be a circuit such that γ∞ ∈Tn
i=1 Fβi.
Notice that γ∞ is a fixed point for θγ, so ξγ∞ is a fixed point for ϑγ by (8.6). Also
ξγ∞ ∈ V∅;β1,β2,...,βn .
Should ξγ∞ be the only element of this set, the singleton {ξγ∞ } would be an open set
consisting of fixed points for ϑγ, contradicting (i). Thus
V∅;β1,β2,...,βn \ {ξγ∞ }
is not empty, hence it contains some ξx, for x in X, as Ξ(X) is dense in ΩX. Thus
necessarily x 6= γ∞, and since each β−1
belongs to ξx, we have that x ∈ Fβi , again by
(7.12). This proves (ii).
i
(ii) ⇒ (i). Arguing by contradiction, let g ∈ F \ {1}, and let U be a nonempty open subset
of Ωg−1 consisting of fixed points for ϑg. Since Ξ(X) is dense in ΩX by (7.4), there exists
some x in X such that ξx ∈ U , whence ξx is fixed by ϑg.
By (8.6) we conclude that x = σ(ξx) is fixed by θg, so (12.5.i) provides α, ν ∈ F+, such
that g = να±1ν−1, in reduced form. Upon replacing g by g−1 if necessary (and keeping U
unaltered), we may assume without loss of generality that g = ναν−1. Observing that ϑ
is semi-saturated, we have that
ϑg = ϑν ◦ ϑα ◦ ϑν−1,
so in particular Ωg−1 ⊆ Ων . Setting W = ϑν−1(U ) we then have that W ⊆ Ωα−1 , and it is
clear that ϑα is the identity on W .
The upshot of the above argument is that, if ϑ is not topologically free, then there
exists a finite word α, and a nonempty open set W consisting of fixed points for ϑα.
Employing (7.4) once more, there exists some x in X such that ξx ∈ W . Therefore ξx is
fixed by ϑα, whence x = σ(ξx) is fixed by θα, and we deduce from (12.5.ii) that x = α∞.
Using the special neighborhood base of ξx provided by (11.3), we may then find finite
words µ, β1, β2, . . . , βn ∈ Λ∗, such that
ξx ∈ Vµ; β1,β2,...,βn ⊆ W.
(12.6.1)
partial actions and subshifts
43
In particular we have that µ lies in ξx, which in turn implies that µ is a prefix of x by
(7.11). Being a prefix of x = α∞ therefore implies that µ is a prefix of αn, for some n,
whence there exists a finite word γ such that αn = µγ. Setting µ′ = µγ, and β′
i = βiγ, we
have that
ξx ∈ Vµ′; β ′
1,β ′
2,...,β ′
n ⊆ Vµ; β1,β2,...,βn ⊆ W.
Replacing µ by µ′, and each βi by the corresponding β′
of generality that (12.6.1) reads
i, we then may assume without loss
ξx ∈ Vαn; β1,β2,...,βn ⊆ W.
Among other things we then have that both αn and αnβ−1
i
lie in ξx, for all i, so (8.8)
gives
σαn (ξx) ∈ Fβi .
Recalling that x = α∞, and staring at the definition of the stem, will make it clear
that σαn(ξx) = α∞. So
α∞ ∈
Fβi .
Noticing that α∞ ∈ Fαn , we may soup up the above conclusion by writing
α∞ ∈
Fβi ∩ Fαn,
n
Ti=1
Ti=1
n
z = αny
and then we may use hypothesis (ii) to produce an infinite word y 6= α∞ belonging to
i=1 Fβi, as well as to Fαn. The infinite word
Tn
thus lies in X and the stem of ξz at αn is clearly y. For that reason, and using (8.8), we
have that αnβ−1
i ∈ ξz, for all i, whence
ξz ∈ Vαn; β1,β2,...,βn ⊆ W,
so ξz is a fixed point for ϑα, whence z is a fixed point for θα. However the only such fixed
point is x, which is manifestly different from z. We have thus reached a contradiction,
hence proving that ϑ is topologically free.
(cid:3)
When one compares our last result to the well known characterization of topological
freeness for Markov subshifts given in (12.2), one might wonder what happened to the
role of exits for circuits. Although property (12.6.ii) does not explicitly mention exits, it is
closely related to that concept. To see this let γ be a circuit. Then evidently γ∞ ∈ Fγn , for
every n, so the existence of an element x in Fγn , other than γ∞, as provided by (12.6.ii),
gives an exit for γ n. If this indeed holds for every n, then γ has a strong exit by (3.3).
We therefore see that (12.6.ii) implies that all circuits have strong exits. However we
should not forget that the existence of strong exits by itself is not enough to guarantee
topological freeness for the spectral partial action, as the example of the even shift in (3.5)
above shows.
44
m. dokuchaev and r. exel
13. Minimality.
Recall from [18: Definition 2.8] that a topological partial action
ρ =(cid:0){Yg}g∈G, {ρg}g∈G(cid:1)
of a group G on a space Y is said to be minimal if there are no nontrivial ρ-invariant closed
subsets. This is clearly equivalent to saying that for every y in Y , the orbit of y, namely
the set
Orb(y) = {ρg(y) : g ∈ G, Yg−1 ∋ y},
is dense in Y .
Minimality is well understood for Markov subshifts as well as for many other partial
dynamical systems related to it. See, for example [20] and [21].
For the purpose of comparison, lets us state the following well known result.
13.1. Proposition. Let X be a Markov subshift with transition matrix A. Then the
standard partial action θ of F on X (which, by (8.7) is equivalent to the spectral partial
action) is minimal if and only if, for any vertex a, and any infinite path x in Gr(A), there
exists a finite path starting in a and ending in some vertex of x.
The property described in the above result is sometimes referred to as cofinality.
Motivated by this concept it is natural to consider the following property applicable for
general subshifts.
13.2. Definition. Let Λ be a finite alphabet and let X ⊆ Λ be a subshift. We shall
say that an infinite word x in X may be reached from a finite word β in LX , when there
exists a finite word γ, and a prefix α of x, such that, upon writing x = αy, one has that
βγy ∈ X.
α
y
xn
{
.........................................................•
xn+1
................................................
.........................................................•
.........................................................•
γm
z
xn+2
.........................................................•
}
.........................................................•. . .
{
z
e
x1
x2
................................................•.........................................................•
.........................................................•
.........................................................•
...
}
.........................................................•
γ2
.........................................................•
...
•.........................................................•
γ1
................................................
.........................................................•
...
.........................................................•
β2
•.........................................................•
β1
Diagram (13.3)
We shall moreover say that X is cofinal if, for every β ∈ LX, and every x in X, one has
that x may be reached from β.
One could think of γ as being the bridge allowing one to travel from β in order to
reach x at α. A quicker way to express (13.2) is to say that
∃γ ∈ Λ∗, ∃n ∈ : Sn(x) ∈ Fβγ.
partial actions and subshifts
45
Substituting the above notion of cofinality for the classical notion, one might be
tempted to conjecture a generalization of Proposition (13.1) above to arbitrary subshifts,
but once more the even shift stands as a counter-example. It is easy to see that the even
shift satisfies the property just mentioned, because, given any x and β, as in the above
diagram, one could simply take α = ∅, and then either
γ = ∅,
or γ = '1'
(13.4)
will always work as a bridge, as one of them will correctly adjust the parity of the amount
of 1's between the last '0' of β and the first '0' of x.
However the spectral partial action for the even shift is not minimal. To see this let
us return to the situation presented in (12.3), when we have verified that the open set
V∅;β1,β2 consists of a single point, namely ξ1∞. Should the spectral partial action for the
even shift be minimal, then Orb(ξx) should intersect V∅;β1,β2, for every x in X, but this is
not the case, e.g. for the element x = '00000 . . .'. In fact, if
ϑg(ξ0∞) ∈ V∅;β1,β2 ,
for some g in F, then ϑg(ξ0∞) = ξ1∞, whence by (8.6) one has that θg(0∞) = 1∞, and this
is manifestly impossible.
For future reference let us highlight the conclusion reached above:
13.5. Proposition. The spectral partial action associated to the even shift is not mini-
mal.
As in the previous section, one might decide to ignore that the standard partial action
for the even shift is not a topological partial action and ask whether or not it is minimal
in the sense that all orbits are dense. The answer is then easily seen to be positive, so in
this sense the standard partial for the even shift is minimal.
Even though cofinality for a subshift does not imply minimality for the spectral par-
tial action, as in the Markov case, the former concept is quite relevant in our study of
minimality, having the following dynamical interpretation.
13.6. Proposition. Let X be a subshift. Then the following are equivalent:
(i) X is cofinal,
(ii) for every x in X, and for every β in LX , there is some g in F such that both g and
gβ lie in ξx,
(iii) for every x in X, and for every β in LX, one has that Orb(ξx) ∩ Ωβ 6= ∅.
Proof. (i) ⇒ (ii). Given x and β, as in (ii), choose α, γ and y, as in (i), so that x = αy,
and βγy ∈ X. We then have that σα(ξx) = y, so the fact that y ∈ Fγ yields αγ−1 ∈ ξx,
by (8.8). Similarly, the fact that y ∈ Fβγ yields α(βγ)−1 ∈ ξx.
It is then clear that
g = α(βγ)−1 satisfies the conditions of (ii).
(ii) ⇒ (i). Given β in LX , and x in X, choose g in F such that g, gβ ∈ ξx. By (7.11) we
may write gβ = αγ−1 in reduced form, with α, γ ∈ F+, such that x = αy, for some infinite
word y ∈ Fα ∩ Fγ. Observing that ξx contains both α and
g = gββ−1 = αγ−1β−1 = α(βγ)−1,
46
m. dokuchaev and r. exel
we deduce from (8.8) that y = σα(ξx) ∈ Fβγ, whence βγy ∈ X. This shows that x may be
reached from β.
The equivalence of (ii) and (iii) follows from the next simple result which we will also
(cid:3)
use later in a slightly more general situation.
13.7. Proposition. Let X be a subshift and let ξ ∈ ΩX . Given g ∈ F, and β ∈ LX, the
following are equivalent:
(i) both g and gβ lie in ξ,
(ii) ξ ∈ Ωg, and ϑg−1(ξ) ∈ Ωβ.
Proof. Notice that g ∈ ξ, if and only if ξ ∈ Ωg, by (8.4.i), and in this case
gβ ∈ ξ ⇐⇒ β ∈ g−1ξ = ϑg−1(ξ) ⇐⇒ ϑg−1(ξ) ∈ Ωβ.
(cid:3)
Recall that ϑ is minimal if and only if, for every ξ in ΩX , and for every nonempty
open set U ⊆ ΩX , one has that
Orb(ξ) ∩ U 6= ∅.
(13.8)
Expressing cofinality in terms of (13.6.iii) may thus be interpreted as a weak form
of minimality in the sense that the above holds when ξ has the form ξx, and U has the
form Ωβ. One therefore has a clear perspective that cofinality is a weaker property than
minimality, and in fact strictly weaker, as the example of the even shift shows.
If we are to find a set of properties related to cofinality, characterizing minimality
for the spectral partial action, we must therefore significantly strengthen the notion of
cofinality. We will in fact do this in two directions, introducing the notions of collective
cofinality and strong cofinality. The first will lead to Theorem (13.12), ensuring that (13.8)
holds for an arbitrary U , but still under the restriction that ξ = ξx, while the second will
show up in Theorem (13.16), yielding (13.8) for an arbitrary ξ, but under the restriction
that U = Ωβ. Together these two will eventually prove ϑ to be minimal.
13.9. Definition. Let Λ be a finite alphabet and let X ⊆ Λ be a subshift. Given a
finite set of finite words B ⊆ LX, and an infinite word x ∈ X, we shall say that x may be
collectively reached from B, when there exists a finite word γ, and a prefix α of x, such
that, upon writing x = αy, one has that βγy ∈ X, for all β ∈ B.
partial actions and subshifts
47
α
y
z
e
x1
x2
................................................•.........................................................•
.........................................................•
.........................................................•
xn+1
.........................................................•
xn+2
.........................................................•
.........................................................•. . .
}
{
xn
{
.........................................................•
.........................................................•
γm
z
.........................................................•
...
...
}
.........................................................•
γ2
•.........................................................•
β1
1
γ1
β1
2
•.........................................................•
................................................
.........................................................•
..........................................................•
..........................................................•
..........................................................•
.........................................................•
.........................................................•
•.........................................................•
•.........................................................•
β3
2
β2
1
β3
1
β2
2
It should be stressed that the bridge γ above is supposed to be the same for all β.
In fact an infinite word x may be reached individually from both β1 and β2, but not
collectively by the set {β1, β2}, as is the case of the even shift with
β1 = 01,
β2 = 011,
and
x = 00000 . . .
In case a word x is collectively reached from a subset B ⊆ LX, as in the diagram
above, notice that γy lies in the follower set Fβ, for all β in B, and in particular the
intersection of the Fβ is nonempty. This motivates an extension of the notion of follower
sets:
13.10. Definition. Given a finite subset B ⊆ LX , we will say that the follower set of B
is the set FB defined by
Fβ.
FB = Tβ∈B
13.11. Definition. We will say that a subshift X is collectively cofinal if, for every finite
set B ⊆ LX, with FB nonempty, and every x in X, one has that x may be collectively
reached from B.
The first relationship between collective cofinality and minimality is as follows:
13.12. Theorem. Let X be a subshift. Then the following are equivalent:
(i) for every x in X, the orbit of ξx is dense in ΩX,
(ii) X is collectively cofinal.
Proof. (i) ⇒ (ii). Given any finite subset B ⊆ LX, such that FB is nonempty, say B =
{β1, β2, . . . , βn}, we claim that
V∅; β1,β2,...,βn 6= ∅.
To see this, pick y in FB, and notice that β−1 ∈ ξy, for every β in B, by (7.12), whence ξy
is an element in the above set, proving it to be nonempty.
48
m. dokuchaev and r. exel
Having already chosen B, let us also choose any x in X, and our task is to show that
x may be collectively reached from B. By hypothesis the orbit of ξx is dense in ΩX , so
there exists g in F such that ξx ∈ Ωg−1, and
It follows that g−1 ∈ ξx, by (8.4.i), whence we may use (7.11) to write g−1 = αγ−1, in
reduced form, with α, γ ∈ F+, and moreover write x = αy, where y ∈ Fα ∩ Fγ. Therefore
ϑg(ξx) ∈ V∅; β1,β2,...,βn .
(13.12.1)
ϑg(ξx)
(8.6)
= ξθg(x) = ξγy.
It then follows from (13.12.1) that β−1
i ∈ ξγy, for every i, so γy ∈ Fβi , by (7.12), which is
the same as saying that βiγy ∈ X. This shows that x may be collectively reached from B,
and hence proves (ii).
(ii) ⇒ (i). Given any x in X, we must show that the orbit of ξx is dense. Since Ξ(X) is
already known to be dense in ΩX, it is enough to prove that
ξz ∈ Orb(ξ),
∀ z ∈ X.
Given any z in X, let U be an arbitrary neighborhood of ξz. By (11.3) there are finite
words α, β1, β2, . . . , βn, such that
ξz ∈ Vα; β1,β2,...,βn ⊆ U,
and all we need to do is prove that
Observing that Vα; β1,β2,...,βn ⊆ Ωα, and that
Orb(ξx) ∩ Vα; β1,β2,...,βn 6= ∅.
ϑα−1 (Vα;
β1,β2,...,βn ) =
V∅; α,β1,β2,...,βn ,
(13.12.2)
(13.12.3)
notice that any orbit intersecting V∅; α,β1,β2,...,βn will also intersect Vα; β1,β2,...,βn. So we
may assume without loss of generality that α = ∅.
have that β−1
Letting B = {β1, β2, . . . , βn}, we claim that FB is nonempty. In fact, by (13.12.2) we
i ∈ ξz, for all i, so z ∈ Fβi by (7.12), whence also z ∈ FB, proving our claim.
By hypothesis x may be collectively reached from B, so there are finite words µ and
γ, such that x = µy, and βγy ∈ X, for all β in B. Noticing that for all such β,
we have that
σµ(ξx) = y ∈ Fβγ,
µ(βγ)−1 ∈ ξx,
(13.12.4)
by (8.8). The fact that βγy ∈ X implies that γy ∈ X as well, so a similar argument gives
µγ−1 ∈ ξx, whence ξx ∈ Ωµγ−1 , and
ϑγµ−1 (ξx) = γµ−1ξx
(13.12.4)
∋
β−1.
Consequently ϑγµ−1 (ξx) ∈ V∅; β1,β2,...,βn , proving (13.12.3).
(cid:3)
partial actions and subshifts
49
The reader is invited to check that the only reason why (13.12.i) must be stated just
for ξx, rather than for a general ξ in ΩX , is that the above proof has used the crucial
implication "(b) ⇒ (a)" of (8.8), which does not hold in general.
In the presence of collective cofinality, topological freeness for the spectral partial
action may be characterized in a very simple way:
13.13. Proposition. Let X be a subshift, and consider the following statements:
(a) The spectral partial action associated to X is topologically free,
(b) X has at least one point which is not eventually periodic (by eventually periodic we
mean a word of the form αγ∞),
Then (a) implies (b). If X is collectively cofinal, then also (b) implies (a).
Proof. (a) ⇒ (b).
It is easy to see that there is only a countable number of eventually
periodic infinite words (even within the full shift Λ). Assuming by contradiction that
every x in X is eventually periodic, one then has that X is countable, and a standard
application of Baire's category Theorem implies that X has at least one isolated point, say
x. Since x is eventually periodic let us write x = αγ∞, for some circuit γ.
To say that x is isolated (relative to the topology of X), is to say that {x} coincides
with some cylinder Zµ, where µ is necessarily a prefix of x = αγ∞. By increasing the size
of µ up to α plus some multiple of γ, we may suppose that µ = αγ n, for some integer n.
Thus the only infinite word in X that one may produce by starting with αγ n is αγ∞,
which may also be expressed by saying that Fαγn = {γ∞}. This goes against (12.6.ii),
and hence also against (a), thus bringing about a contradiction, proving that X must have
some non eventually periodic point.
(b) ⇒ (a). We will prove (a) by verifying (12.6.ii). So let ν be a circuit and let B be a
finite set of finite words such that ν∞ ∈ FB. Our task is then to find some element in FB
other than ν∞.
For this, pick any non eventually periodic element x of X and, using that X is collec-
tively cofinal, choose a prefix α of x, and a finite word γ, such that, upon writing x = αy,
one has that βγy ∈ X, for all β in B. This implies that γy lies in FB, and since it is clearly
not eventually periodic, one necessarily has that γy 6= ν∞, thus verifying (12.6.ii).
(cid:3)
Having understood the dynamical meaning of collective cofinality in (13.12), let us
now discuss yet another version of cofinality. In order to motivate this notion let us refer
back to diagram (13.3). One might see the question of reaching x from β as an attempt
to bringing x into the follower set of β by deleting a few letters from the beginning of x,
namely the prefix α, and then inserting new letters in its place, namely γ, after which the
resulting word γy is supposed to lie in Fβ. If we have to pay a price for each deleted, as
well as for each inserted letter, then the cheapest situation is obviously when x itself lies
in the follower set of β, namely when α = γ = ∅ are enough to do the job. Beyond this
ideal situation we have:
13.14. Definition. Let β ∈ LX , and let x ∈ X. We shall say that the cost of reaching x
from β is the minimum value of α + γ, where α and γ are as in (13.2). In symbols,
If no such α and γ exist, we shall say that the cost if infinite.
Cost(β, x) = min{α + γ : x = αy, βγy ∈ X}.
50
m. dokuchaev and r. exel
In a cofinal subshift, given β in LX , one may reach any x in X, but perhaps at an
increasingly high cost. This motivates the following:
13.15. Definition. Let X be a subshift. We shall say that X is strongly cofinal if
Cost(β, x) < ∞,
sup
x∈X
for every β in LX.
Let us pause for a moment to give an example of a cofinal subshift which is not
strongly cofinal. Our staple counter-example, namely the even shift, will not serve us now
since it is both cofinal and strongly cofinal. In fact, as we already mentioned in (13.4),
when trying to reach an infinite word x from a finite word β, it suffices to take α = ∅, and
either γ = ∅ or γ = '1', which means that Cost(β, x) is at most 1.
However, a subshift based on a similar principle as the even shift will provide the
counter-example sought. Consider the alphabet Λ = {0, 1}, and let us take the following
set F of forbidden words:
F = {01n0 : n is not a power of 2}.
If X = XF is the corresponding subshift, then an infinite word x lies in X if and only
if, anytime a contiguous block of 1's occurring in x is delimited by 0's, the amount of said
1's is a power of 2.
We leave it for the reader to check that X is a duly cofinal subshift, but that reaching
the infinite word x = 1n0∞ from the finite word β = '0' involves unbounded costs. Hence
X is not strongly cofinal.
Exploring the relationship between strong cofinality and the dynamical properties of
the spectral partial action is our next goal.
13.16. Theorem. Let X be a subshift. Then the following are equivalent:
(i) For every β in LX , and for every ξ in ΩX, one has that Orb(ξ) ∩ Ωβ is nonempty,
(ii) X is strongly cofinal.
Proof. (i) ⇒ (ii). We will in fact prove that the negation of (ii) implies the negation of (i).
We therefore suppose that
Cost(β, x) = ∞,
sup
x∈X
for some β in LX. Then, for every natural number n, there is some zn in X, with
Cost(β, zn) ≥ n. By compactness we may then find a subsequence xk = znk , such that
{ξxk }k converges to some ξ in ΩX. We will accomplish our task by proving that
Orb(ξ) ∩ Ωβ = ∅.
(13.16.1)
Arguing by contradiction, suppose that there exists some h in F such that ξ ∈ Ωh−1,
and ϑh(ξ) ∈ Ωβ. By (13.7) we deduce that both h−1 and h−1β lie in ξ. In order to simplify
our notation, we shall make the change of variables g = h−1β, whence
g, gβ−1 ∈ ξ.
partial actions and subshifts
51
Using (7.8.v) we may write g = αγ−1 in reduced form, so that α ∈ ξ, by (7.8.ii).
Employing (7.16.1) we then have that
α, g, gβ−1 ∈ ξxk ,
for all large enough k. Focusing on the fact that
αγ−1 = g ∈ ξxk ,
if follows from (7.11) that xk = αyk, where yk ∈ Fα ∩ Fγ. Observing that the stem of ξxk
at g is γyk, and that gβ−1 ∈ ξxk , we conclude from (8.8) that γyk ∈ Fβ, which is to say
that βγyk ∈ X. This implies not only that xk may be reached from β, but also that
Cost(β, xk) ≤ α + γ.
Since neither α nor γ depend on k, we arrive at a contradiction with the fact that
limk→∞ Cost(β, xk) = ∞. This proves (13.16.1), as desired.
(ii) ⇒ (i). Given β in LX , and ξ in ΩX , we must prove that Orb(ξ) ∩ Ωβ 6= ∅. Writing
ξ = limn ξxn , with xn ∈ X, by hypothesis we may reach every xn from β with bounded
cost, meaning that for every n, there are finite words αn and γn, such that xn = αnyn,
and βγnyn ∈ X, and moreover αn + γn ≤ M , where M is a fixed constant.
As we are working with a finite alphabet, there are finitely many words of any given
length. Therefore the αn and γn just obtained must necessarily repeat infinitely many
often. We may then choose a subsequence {xnk }k, such that αnk = α, and γnk = γ, for
all k. Therefore xnk = αynk , and βγynk ∈ X. We then have that
α ∈ ξxnk
,
and
σα(ξxnk
) = ynk ∈ Fβγ,
whence α(βγ)−1 ∈ ξxnk
in Fγ, gives αγ−1 ∈ ξxnk
by (8.8). A similar reasoning, based on the fact that ynk also lies
, so by the continuity of Boolean values (7.16.1), we have that
αγ−1 ∈ ξ,
and
α(βγ)−1 ∈ ξ.
Setting g = α(βγ)−1, we have that
gβ = αγ−1β−1β = αγ−1 ∈ ξ,
whence ϑ−1
the proof.
g (ξ) ∈ Ωβ by (13.7). This shows that the orbit of ξ intersects Ωβ, concluding
(cid:3)
From (13.12) and (13.16) it is clear that a subshift whose associated spectral partial
action is minimal must necessarily be both collectively cofinal and strongly cofinal. Our
next major goal will be to prove that these two properties in turn characterize minimality
for the spectral partial action. However there is an important technical tool we still need
to develop before proving this main result.
52
m. dokuchaev and r. exel
13.17. Lemma. Let X be a collectively cofinal subshift. Then for every finite set B ⊆
LX, with FB nonempty, there are µ and ν in LX, such that
where Bν evidently means the set {βν : β ∈ B}.
Fµ ⊆ FBν ,
Proof. Let B be as in the statement. By hypothesis any x in X may be collectively reached
from B, so we may pick α, γ and y, as in (13.9), so that y ∈ Fα ∩ FBγ , and x = θα(y),
whence x ∈ θα(Fα ∩ FBγ ). This implies that
X = Sα,γ∈Λ∗
θα(Fα ∩ FBγ ).
Observing that Fα ∩ FBγ is compact, and hence that the sets in the above union are
closed in X, we may employ Baire's category Theorem producing α and γ in Λ∗ such that
θα(Fα ∩ FBγ ) has a nonempty interior. En passant we stress that both α and γ must lie in
LX, or else Fα ∩ FBγ is the empty set. Therefore, by the definition of the product topology
on X, there exists some µ in LX such that
Zµ ⊆ θα(Fα ∩ FBγ ).
Assuming without loss of generality that µ > α, and noticing that the range of θα is
contained in Zα, we have that Zµ ⊆ Zα, so α must be a prefix of µ, and we may then write
µ = αδ, for some finite word δ ∈ LX. We will now conclude the proof by showing that the
inclusion in the statement holds once we choose ν = γδ.
To prove this, let y ∈ Fµ, so
x := µy ∈ Zµ ⊆ θα(Fα ∩ FBγ ),
and then we may write x = αz, for some z ∈ Fα ∩ FBγ. Since
it follows that δy = z, so for any β in B, we have that
αδy = µy = x = αz,
βνy = βγδy = βγz ∈ X.
This shows that y ∈ Fβν, as desired, concluding the proof.
(cid:3)
Let us now give a dynamical interpretation of the conclusion of the above result.
13.18. Proposition. Let B = {β1, β2, . . . , βn} ⊆ LX, be a nonempty finite set and let
µ, ν ∈ LX, be such that Fµ ⊆ FBν , precisely as in the conclusion of (13.17). Then Ωµ−1 is
contained in the domain of ϑν (also known as Ων−1), and
ϑν(Ωµ−1 ) ⊆ V∅; β1,β2,...,βn.
partial actions and subshifts
53
Proof. We will first prove that
Ξ(X) ∩ Ωµ−1 ⊆ Ων−1.
(13.18.1)
We thus pick any x in X such that ξx ∈ Ωµ−1 . Notice that
ξx ∈ Ωµ−1
(8.4.i)
⇐⇒ µ−1 ∈ ξx
(7.12)
⇐⇒ x ∈ Fµ.
So x ∈ Fµ, and by hypothesis we then have that x ∈ FBν . Since B is nonempty, we
may pick any β in B, and so deduce that x ∈ Fβν , whence βνx ∈ X. This implies that
x ∈ Fν , and a reasoning similar to the above shows that ξx ∈ Ων−1 , proving (13.18.1). A
trivial exercise in Topology, using that Ξ(X) is dense, Ωµ−1 is open, and Ων−1 is closed,
now shows that Ωµ−1 ⊆ Ων−1, as required.
We will next show that
We thus again pick any x in X such that ξx ∈ Ωµ−1 . Then
ϑν(cid:0)Ξ(X) ∩ Ωµ−1(cid:1) ⊆ V∅; β1,β2,...,βn .
ϑν (ξx)
(8.6)
= ξθν (x) = ξνx.
(13.18.2)
In addition, as seen above, x ∈ FBν , so βνx ∈ X, for every β in B, and we see that
νx ∈ Fβ. We then have by (7.12) that β−1 ∈ ξνx, whence ξνx ∈ V∅; β1,β2,...,βn , proving
(13.18.2). The same "dense plus open plus closed" argument above, but now also using
that ϑν is continuous, leads one from (13.18.2) to the last conclusion in the statement. (cid:3)
We have now come to the main result of this section:
13.19. Theorem. Let Λ be a finite alphabet and let X ⊆ Λ be a subshift. Also let ϑ
be the spectral partial action of the free group F(Λ) on ΩX introduced in (8.3). Then a
necessary and sufficient condition for ϑ to be minimal is that X be both collectively cofinal
and strongly cofinal.
Proof. As already observed, the necessity of the above conditions follows immediately from
(13.12) and (13.16).
Conversely, suppose that X is both collectively cofinal and strongly cofinal. In order
to prove that ϑ is minimal, given any ξ in ΩX , we must show that the orbit of ξ is
dense. Arguing as in the proof of "(ii) ⇒ (i)" in (13.12), it suffices to show that if B =
{β1, β2, . . . , βn} is a subset of LX such that FB is nonempty, then
Orb(ξ) ∩ V∅; β1,β2,...,βn 6= ∅.
Using (13.17) and (13.18), pick µ and ν in LX , such that
ϑν(Ωµ−1 ) ⊆ V∅; β1,β2,...,βn.
Thus, to show that the orbit of ξ intersects V∅; β1,β2,...,βn, it is enough to show that it
intersects Ωµ−1 , or even Ωµ, since ϑµ−1 (Ωµ) = Ωµ−1 .
In order to prove the latter fact, that is, that Orb(ξ) ∩ Ωµ is nonempty, we just have
(cid:3)
to note that it follows immediately from strong cofinality and (13.16).
54
m. dokuchaev and r. exel
Of course one could put together the notions of collective cofinality and strong cofi-
nality, to form a new high powered notion, as follows:
13.20. Definition. Let X be a subshift.
(a) Given a finite subset B of LX, we shall say that the cost of reaching a given x in X
from B is the minimum value of α + γ, where α and γ are as in (13.9). In symbols,
Cost(B, x) = min{α + γ : x = αy, βγy ∈ X, for all β ∈ B}.
If no such α and γ exist, we shall say that the cost if infinite.
(b) We shall say that X is hyper cofinal if
Cost(B, x) < ∞,
sup
x∈X
for every finite subset B of LX such that FB is nonempty.
It is evident that hyper cofinality implies both collective cofinality and strong cofinal-
ity, but the reverse implication is not so obvious to see. Nevertheless it is true, as we shall
now prove.
13.21. Proposition. Let X be a subshift. Then the following are equivalent:
(i) X is is both collectively cofinal and strongly cofinal,
(ii) X is hyper cofinal,
(iii) the spectral partial action associated to X is minimal.
Proof. The equivalence of (i) and (iii), repeated here just for emphasis, is precisely the
content of (13.19). As we have already seen (ii) implies (i) for obvious reasons, so it
suffices to show that (iii) implies (ii).
So, assuming that ϑ is minimal, let B = {β1, β2, . . . , βn} ⊆ LX be such that FB is
nonempty. Consider the open subset
V := V∅; β1,β2,...,βn ⊆ ΩX ,
which is nonempty by the short argument used in the beginning of the proof of (13.12).
By minimality, for every ξ in ΩX, one has that Orb(ξ) is dense, so it must intersect V ,
meaning that there exists some g in F, such that ξ ∈ Ωg−1 , and ϑg(ξ) ∈ V . It follows that
ξ ∈ ϑg−1(V ∩ Ωg), so we conclude that
ϑg−1(V ∩ Ωg).
ΩX = Sg∈F
This is an open covering of the compact space ΩX , so there exists a finite collection
g1, g2, . . . , gm of elements in F such that
ΩX =
m
Si=1
ϑg−1
i
(V ∩ Ωgi).
partial actions and subshifts
55
Given any x in X, we may then find some i ≤ m, such that ξx ∈ ϑg−1
particular ξx ∈ Ωg−1
that αi is a prefix of x and, upon writing x = αiy, one has that
and ϑgi(ξx) ∈ V . Writing gi = γiα−1
i
i
(V ∩ Ωgi), so in
, in reduced form, we then have
i
ϑgi(ξx) = ξθgi (x) = ξγiy ∈ V = V∅; β1,β2,...,βn .
Consequently β−1
βjγiy lies in X. This shows that x may be reached from B, and moreover that
j ∈ ξγiy, for all j ≤ n, whence γiy ∈ Fβj by (7.12), which is to say that
Cost(B, x) ≤ αj + γj.
Since the j above may take only finitely many values, we see that the cost of reaching any
x from the given B is bounded, whence X is hyper cofinal.
(cid:3)
Not all subshifts are surjective5, but when they are, there is another property (to be
described in our next result) even stronger than hyper cofinality, which is still equivalent
to minimality. It says that one may reach infinite words at their source with bounded cost.
This property is essentially the same as the condition appearing in [31: Theorem 4.20].
See section (14) for a more thorough discussion of this condition.
13.22. Proposition. Let X be a subshift and suppose that the shift map S : X → X
is surjective. Then the spectral partial action associated to X is minimal if and only if,
for every finite subset B ⊆ Λ∗, with FB nonempty, there exists a constant M > 0, such
that, for every x in X, one may find γ ∈ Λ∗ with γ ≤ M , and such that βγx ∈ X, for all
β ∈ B. In other words,
min{γ : βγx ∈ X, for all β ∈ B} < ∞.
sup
x∈X
x1
x2
x3
...
•.........................................................•
.........................................................•
.........................................................•
.........................................................•
.........................................................•. . .
.........................................................•
γm
.........................................................•
...
.........................................................•
γ2
•.........................................................•
β1
1
γ1
β1
2
•.........................................................•
................................................
.........................................................•
..........................................................•
..........................................................•
..........................................................•
.........................................................•
.........................................................•
•.........................................................•
•.........................................................•
β3
2
β3
1
β2
1
β2
2
5 If a subshift is built from a two-sided shift, as in many papers on the subject (e.g. [9]), then the shift
map S is surjective, but there are many non-surjective (one-sided) subshifts as well.
56
m. dokuchaev and r. exel
Proof. It is evident that the condition given implies hyper cofinality, and hence minimality
by (13.21). Conversely, suppose that X is minimal, hence also hyper cofinal by (13.21).
Given any finite subset B of LX, let
n = sup
x∈X
Cost(B, x).
Fixing x in X, we may use the fact that S is surjective to find z in X such that
Sn(z) = x. It is then clear that z is of the form µx, with µ = n. Spelling out the fact
that z may be reached from B with cost no more than n, there are α and γ in Λ∗, such
that α is a prefix of z, and, upon writing z = αy, one has that βγy ∈ X, for every β in B,
and also α + γ ≤ n.
α
y
µ1
µ2
•.........................................................•
.........................................................•
.........................................................•
µk
...
µn
x1
x2
...
.........................................................•
.........................................................•
.........................................................•
.........................................................•
.........................................................•
.........................................................•
.........................................................•. . .
{
z
{
.........................................................•
µk−1
.........................................................•
γm
.........................................................•
...
...
}
.........................................................•
γ2
z
}
δ
{z
}
•.........................................................•
β1
1
γ1
β1
2
•.........................................................•
................................................
..........................................................•
.........................................................•
..........................................................•
..........................................................•
.........................................................•
.........................................................•
•.........................................................•
•.........................................................•
β2
1
β3
1
β3
2
β2
2
Of crucial importance is whether we have reached z before or after the end of the
prefix µ. Observing that
α ≤ α + γ ≤ n = µ,
the answer to the above dilemma is before! Since both α and µ are prefixes of z, we deduce
that α is a prefix of µ, and we may then write µ = αδ, for some finite word δ. It is then
clear that y = δx, and
X ∋ βγy = βγδx = βγ′x,
where γ′ = γδ. Noticing that
γ′ = γ + δ ≤ α + γ + µ ≤ 2n,
the proof is concluded.
(cid:3)
partial actions and subshifts
57
14. Applications.
In this short section we will draw a few conclusions about the Carlsen-Matsumoto C*-
algebra which may be derived from the work we did so far. Some of these results are
already known, but we may recover them easily given the many tools available to treat
partial crossed product algebras.
14.1. Theorem. (cf. [10: Theorem 17]) For every subshift X, one has that OX is a
nuclear C*-algebra.
Proof. By (9.5) we have that OX is the reduced crossed product relative to a partial dy-
namical system satisfying the approximation property. The result is then an immediate
application of [16: 25.10].
(cid:3)
The first description of OX as a groupoid C*-algebra was given in [6]. We may recover
it here due to its description as a partial crossed product:
14.2. Theorem. Let GX be the transformation groupoid associated to the spectral par-
tial action for a given subshift X. Then GX is a second countable, Hausdorff, ´etale,
amenable groupoid and
C∗(GX ) ≃ OX .
Proof. We refer the reader to [1: Section 2] for the construction of the transformation
groupoid relative to a partial action. Since the acting group, namely F, is discrete, it is
clear that GX is ´etale. It is also easy to see that GX is Hausdorff and, based on the fact
that ΩX is metrizable and F is countable, GX is seen to be second countable. That C∗(GX)
is isomorphic to the crossed product C(ΩX) ⋊ϑ F, and hence also to OX , follows from [1
: Theorem 3.3]. Finally, since OX is nuclear, we deduce from [5: Theorem 5.6.18] that GX
is amenable.
(cid:3)
Our next result involves a weaker alternative to the condition (I) introduced by Mat-
sumoto in [25]. Under condition (I) the result below has essentially been proved by Mat-
sumoto and Carlsen in [9: Lemma 2.3]. See also [10: Theorem 16].
14.3. Theorem. Let X be a subshift. Then the following are equivalent:
(i) X satisfies (12.6.ii), that is, for every β1, β2, . . . , βn in Λ∗, and for every circuit γ such
i=1 Fβi contains some element other than γ∞,
that γ∞ ∈Tn
i=1 Fβi , one has that Tn
(ii) the spectral partial action associated to X is topologically free,
(iii) GX is an essentially principal groupoid (see [29: Definition 3.1]),
(iv) every nontrivial closed two-sided ideal in OX has a nontrivial intersection with DX ,
(v) every *-homomorphism defined on OX is injective, provided it is injective on DX .
In case the equivalent conditions above hold, then OX is naturally isomorphic to MX.
Proof. The equivalence between (i) and (ii) is precisely the content of (12.6), while the
equivalence between (ii) and (iii) is evident. That (ii) implies (iv) was proved in [18
: Theorem 2.6] (see also [15: Theorem 4.4]), for the reduced crossed product, which is
isomorphic to OX by (9.5). A standard argument equating an ideal with the kernel of a
*-homomorphism proves that (iv) and (v) are equivalent.
58
m. dokuchaev and r. exel
To close the cycle it suffices to prove that (iv) implies (iii), and this follows from [4:
Proposition 5.5].
Addressing the last sentence in the statement, it is easy to see that the pair (i, u)
is a covariant representation of the spectral partial action in MX, where i denotes the
inclusion of C(ΩX ) = DX in MX. By the universal property of the crossed product [16:
13.1] there is an obviously surjective *-homomorphism
ϕ : C(ΩX ) ⋊ϑ F → MX,
extending the above covariant pair. Since ϕ restricts to the identity map on C(ΩX ), it
is injective there. The kernel of ϕ therefore has trivial intersection with C(ΩX) = DX,
whence this kernel itself is trivial by (iv). Therefore ϕ is one-to-one and the result follows
from the identification between the crossed product and OX given by (9.5).
(cid:3)
It is curious to compare (12.6.ii) with condition (I) from [25] already referred to.
Observe that the former is essentially saying that there is no l-past equivalence class
containing a single periodic point, while the latter says that there is no such class containing
a single point (periodic or not). Seen from this point of view, one realizes that condition
(I) is stronger than (12.6.ii), and therefore (14.3) is a stronger result than [10: Theorem
16], say. The following example shows that there is a real difference between the above
conditions.
14.4. Example. There exists a subshift satisfying condition (12.6.ii) but not condition
(I).
Proof. Let Σ = {1, 2} and let
Y = Σ = {1, 2}
be the full shift. Choose any non-periodic point z in Y , and let us construct another
subshift on the new alphabet Λ = {0, 1, 2}. For this we let F be the following set of
forbidden words:
F = {a0 : a ∈ Λ} ∪ {0α : α ∈ Y, α is not a prefix of z},
and we claim that the subshift
X = YF
satisfies the requirements in the statement.
Since all forbidden words have a '0' somewhere, any infinite word not involving '0' is
allowed, meaning that Y ⊆ X. Another interesting element of X is the word 0z, which
narrowly escapes being ruled out! Other than that, there is nothing else in X, meaning
that
as the reader may easily verify. Denoting by
X = Y ∪ {0z},
Λl(x) = {α ∈ Λ∗ : α = l, αx ∈ X},
∀ x ∈ X,
∀ l ∈ ,
partial actions and subshifts
59
notice that
Λ1(z) = {0, 1, 2},
and
Λ1(x) = { 1, 2}, ∀x 6= z.
So the 1-past equivalence class of z, in the sense of [25], consists only of z, and we thus
see that X does not satisfy condition (I).
In order to prove that X satisfies condition (12.6.ii), let γ be a circuit in X, and
suppose that γ∞ lies in FB, for some finite set B of finite words. Since we have chosen z
not periodic, we necessarily have that z 6= γ∞.
Given β in B we then have that βγ∞ is in X, so β cannot involve '0'. It then follows
that x ∈ Fβ, for every x in Y , whence Y ⊆ FB. So there are plenty of elements in FB
other than γ∞ to choose from, proving that X satisfies condition (12.6.ii).
(cid:3)
Our next Theorem is related to various known simplicity results for C*-algebras of
subshifts.
14.5. Theorem. Let X be a subshift. Then the following conditions are equivalent:
(i) OX is simple,
(ii) X is both collectively cofinal and strongly cofinal, and it satisfies condition (12.6.ii),
(iii) X is hyper cofinal and contains at least one non eventually periodic point.
Proof. (i) ⇒ (ii). By (14.2) and [4: Theorem 5.1] one has that GX is an essentially principal
and minimal groupoid. It then follows that the spectral partial action is topologically free
and minimal, so the conclusion follows from (12.6) and (13.19).
(ii) ⇒ (i). The spectral partial action is topologically free by (12.6), and minimal by (13.19).
Therefore the reduced crossed product DX ⋊red
τ F is simple by [18: Corollary 2.9], and hence
so is OX by (9.5).
(ii) ⇔ (iii) We have seen in (13.21) that X is both collectively cofinal and strongly cofinal
if and only if it is hyper cofinal. In this case the existence of a non eventually periodic
point is equivalent to topological freeness of the spectral partial action by (13.13), hence
also to condition (12.6.ii).
(cid:3)
Of course the last part of (14.5.ii) could be interchanged with the last part of (14.5.iii),
the above choice happening to be purely personal.
Similar simplicity results for C*-algebras associated to subshifts are to be found in
many works in the literature, such as [25: Corollary 6.11], [9: Proposition 2.6], [2: Propo-
sition 4.3], and [31: Theorem 4.20].
We nevertheless feel that the result above is worth the trouble of proving it, mainly
because it gives necessary and sufficient conditions for simplicity. These conditions also
have a distinctively geometrical flavor at the same time that they are closer to the cofi-
nality hypothesis of classical results about graph algebras, especially if compared with the
intricate condition of "irreducibility in l-past equivalence" present in [25] and [9].
Regarding [2: Proposition 4.3], one should be aware that it only holds for subshifts
of finite type because otherwise the shift map is not open, as we have seen in (2.5). In
particular the statement that OX is simple when X is the even shift, given in the second
60
m. dokuchaev and r. exel
paragraph of page 222, is not correct. As we have seen in (13.5), the spectral partial action
for the even shift is not minimal, so it fails to be hyper cofinal by (13.21), whence OX is
not simple by (14.5).
Among the references cited above, the only one providing for necessary and suffi-
cient conditions is [31: Theorem 4.20], which however only applies to surjective subshifts.
The condition given there is essentially the same as the one described in (13.22), hence
equivalent to minimality in the surjective case, but it is much too strong in general, even
for Markov subshifts. For example, the Markov subshift on the alphabet Λ = {1, 2, 3},
corresponding to the graph
.............
..........
........................................................................................................................................................
...................................................................................................................................................................................
........................................................................................................................................................
...................................................................................................................................................................................
............. ..........
............
..........
1
•
............ ..........
................................................................................................................................................................................................................
•
•
................................................................................................................................................................................................................
2
3
......................
..........
............
(recall that our subshifts consist of infinite paths formed by vertices, as opposed to edges)
is non-surjective: since '1' is a source, any infinite word beginning with '1' is not in the
range of S. The associated Carlsen-Matsumoto algebra (which in this case is well known
to be the Cuntz-Krieger algebra for the adjacency matrix of the above graph) is simple
by (14.5), as the reader may easily verify. However its associated subshift does not satisfy
Thomsen's condition, namely the condition given in (13.22), e.g. for
B = {'2'},
and
x = 12222 . . .,
simply because x is not in any follower set whatsoever.
References
[1] F. Abadie, "On partial actions and groupoids", Proc. Amer. Math. Soc., 132, (2003), 1037 -- 1047.
[2] C. Anantharaman-Delaroche, "Purely infinite C*-algebras arising form dynamical systems", Bull.
Soc. Math. France, 125 (1997), 199 -- 225.
[3] R. J. Archbold and J. Spielberg, "Topologically free actions and ideals in discrete dynamical systems",
Proc. Edinburgh Math. Soc., 37 (1993), 119 -- 124.
[4] J. Brown, L. O. Clark, C. Farthing and A. Sims, "Simplicity of algebras associated to ´etale groupoids",
Semigroup Forum, 88 (2014), 433 -- 452.
[5] N.P. Brown and N. Ozawa, "C*-algebras and finite-dimensional approximations", Graduate Studies
in Mathematics, 88, American Mathematical Society, Providence, RI, 2008.
[6] T. M. Carlsen, "Operator algebraic applications in symbolic dynamics", Ph.D. thesis, University of
Copenhagen (2004).
[7] T. M. Carlsen, "Cuntz-Pimsner C*-algebras associated with subshifts", Internat. J. Math., 19 (2008),
47 -- 70.
[8] T. M. Carlsen, "C*-algebras associated to shift spaces", Notes for the Summer School Symbolic
Dynamics and Homeomorphisms of the Cantor set, University of Copenhagen, 23 -- 27 June 2008.
[9] T. M. Carlsen and K. Matsumoto, "Some remarks on the C*-algebras associated with subshifts",
Math. Scand., 95 (2004), 145 -- 160.
partial actions and subshifts
61
[10] T. M. Carlsen and S. Silvestrov, "C*-crossed products and shift spaces", Expo. Math., 25 (2007),
275 -- 307.
[11] J. Cuntz and W. Krieger, "A class of C*-algebras and topological Markov chains", Invent. Math., 63
(1981), 25 -- 40.
[12] R. Exel, "Amenability for Fell bundles", J. reine angew. Math., 492 (1997), 41 -- 73.
[13] R. Exel, "Partial representations and amenable Fell bundles over free groups", Pacific J. Math., 192
(2000), 39 -- 63.
[14] R. Exel, "A new look at the crossed-product of a C*-algebra by an endomorphism", Ergodic Theory
Dynam. Systems, 23 (2003), 1733 -- 1750.
[15] R. Exel, "Non-Hausdorff ´etale groupoids", Proc. Amer. Math. Soc., 139 (2011), 897 -- 907.
[16] R. Exel, "Partial Dynamical Systems, Fell Bundles and Applications", to be published in a forthcom-
ing NYJM book series. Available from http://mtm.ufsc.br/∼exel/papers/pdynsysfellbun.pdf.
[17] R. Exel and M. Laca, "Cuntz-Krieger algebras for infinite matrices", J. reine angew. Math., 512
(1999), 119 -- 172.
[18] R. Exel, M. Laca and J. Quigg, "Partial dynamical systems and C*-algebras generated by partial
isometries", J. Operator Theory, 47 (2002), 169 -- 186.
[19] S. Ito and Y. Takahashi, "Markov subshifts and realization of β-expansions", J. Math. Soc. Japan,
26 (1974), 33 -- 55.
[20] A. Kumjian, D. Pask and I. Raeburn, "Cuntz-Krieger algebras of directed graphs", Pacific J. Math,
184 (1998), 161 -- 174.
[21] A. Kumjian, D. Pask, I. Raeburn, and J. Renault, "Graphs, groupoids, and Cuntz-Krieger algebras",
J. Funct. Anal., 144 (1997), 505 -- 541.
[22] P. K
◦
urka, "Topological and symbolic dynamics", Cours Sp´ecialis´es, 11. Soci´et´e Math´ematique de
France, Paris, 2003. xii+315 pp.
[23] D. Lind and B. Marcus, "An introduction to symbolic dynamics and coding", Cambridge Univ. Press,
1999.
[24] K. Matsumoto, "On C*-algebras associated with subshifts", Internat. J. Math., 8 (1997), 357 -- 374.
[25] K. Matsumoto, "Dimension groups for subshifts and simplicity of the associated C*-algebras", J.
Math. Soc. Japan, 51 (1999), 679 -- 698.
[26] K. Matsumoto, "On automorphisms of C*-algebras associated with subshifts", J. Operator Theory,
44 (2000), 91 -- 112.
[27] K. Matsumoto, "Stabilized C*-algebras constructed from symbolic dynamical systems", Ergodic The-
ory Dynam. Systems, 20 (2000), 821 -- 841.
[28] W. Parry, "Symbolic dynamics and transformations of the unit interval", Trans. Amer. Math. Soc.,
122 (1966), 368 -- 378.
[29] J. Renault, "Cartan subalgebras in C*-algebras", Irish Math. Soc. Bull., 61 (2008), 29 -- 63.
[30] D. Royer, "Representa¸coes parciais de grupos", Master Thesis, Universidade Federal de Santa Cata-
rina, 2001.
[31] K. Thomsen, "Semi-´etale groupoids and applications", Ann. Inst. Fourier (Grenoble), 60 (2010),
759 -- 800.
|
1106.1484 | 3 | 1106 | 2013-05-16T17:43:04 | Group actions on labeled graphs and their C*-algebras | [
"math.OA",
"math.FA"
] | We introduce the notion of the action of a group on a labeled graph and the quotient object, also a labeled graph. We define a skew product labeled graph and use it to prove a version of the Gross-Tucker theorem for labeled graphs. We then apply these results to the $C^*$-algebra associated to a labeled graph and provide some applications in nonabelian duality. | math.OA | math |
GROUP ACTIONS ON LABELED GRAPHS AND THEIR
C∗-ALGEBRAS
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
Abstract. We introduce the notion of the action of a group on a labeled graph and the
quotient object, also a labeled graph. We define a skew product labeled graph and use it
to prove a version of the Gross-Tucker theorem for labeled graphs. We then apply these
results to the C∗-algebra associated to a labeled graph and provide some applications in
nonabelian duality.
1. Introduction
A labeled graph (E,L) is a directed graph E = (E0, E1, r, s) together with a function
L : E1 → A where A is called the alphabet. Labeled graphs are a model for studying
symbolic dynamical systems; the labeled path space is a shift space whose properties
may be inferred from the labeled graph presentation (cf. [15]). Labeled graph algebras
were introduced in [2, 3], their theory has been developed in [1, 9, 10] and has found
applications in mirror quantum spheres in [21].
The main purpose of this paper is to introduce the notion of a group action on a labeled
graph and study the crossed products formed by the induced action on the associated C∗-
algebra. Before we do this we update the definition of the C∗-algebra associated to a
labeled graph. In order to circumvent a technical error in the literature we add a new
condition to ensure that the resulting C∗-algebra satisfies a version of the gauge-invariant
uniqueness Theorem. Since a directed graph is a labeled graph where L is injective,
we will be generalizing a suite of results for directed graphs and their C∗-algebras (see
[6, 13, 11]). This is not as straightforward as it may seem since two distinct edges may
carry the same label, so new techniques will be needed to prove our results.
An action of a group G on a labeled graph (E,L) is an action of G on E together
with a compatible action of G on A so that we may sensibly define the quotient object
(E/G,L/G) as a labeled graph. In [8] Gross and Tucker introduce the notion of a skew
product graph E ×c G formed from a map c : E1 → G and show that G acts freely
on E ×c G with quotient E. The Gross-Tucker Theorem [8, Theorem 2.1.2] takes a free
action of G on E and recovers (up to equivariant isomorphism) the original graph and
action from the quotient graph E/G. One might speculate that a similar result holds
for free actions on labeled graphs. In section 4 we describe a skew product construction
for labeled graphs and prove a version of the Gross-Tucker theorem for free actions on
labeled graphs (Theorem 5.10). Since a group action on a labeled graph is a pair of
Date: April 22, 2019.
2010 Mathematics Subject Classification. Primary: 46L05, Secondary: 37B10.
Key words and phrases. C∗-algebra, labeled graph, group action, skew product, nonabelian duality.
The second author was supported by the Australian Research Council. The third author was sup-
ported by the NSF Mathematical Sciences Postdoctoral Fellowship DMS-1004675, the University of Iowa
Graduate College Fellowship as part of the Sloan Foundation Graduate Scholarship Program, and the
University of Iowa Department of Mathematics NSF VIGRE grant DMS-0602242.
1
2
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
compatible actions, a new approach is needed: In Definition 4.1 we define a skew product
labeled graph (E ×c G,Ld) to be a skew-product graph E ×c G together with a labeling
Ld : (E ×c G)1 → A × G which is defined using a new function d : E1 → G. The purpose
of the new function d is to accommodate the possibility that two edges carry the same
label. In Remark 5.11 we discuss the importance of d.
We then turn our attention to applications of our results on labeled graph actions to
the C∗-algebras, C∗(E,L) we have associated to labeled graphs.
A function c : E1 → G on a directed graph gives rise to a coaction δ of G on C∗(E)
such that C∗(E) ×δ G ∼= C∗(E ×c G) (cf. [11]). In Proposition 6.2 we show that a skew
product labeled graph (E ×c G,Ld) gives rise to a coaction δ of G on C∗(E,L) provided
that c : E1 → G is consistent with the labeling map L : E1 → A. Then in Theorem 6.7 we
show that C∗(E,L)×δ G ∼= C∗(E×c G,L1) where 1 : E1 → G is given by 1(e) = 1G for all
e ∈ E1. Since this isomorphism is equivariant for the dual action of G on C∗(E,L) ×δ G
and the action of G on C∗(E ×c G,L1) induced by left translation of G on (E ×c G,L1),
Takai duality then gives us
C∗(E ×c G,L1) ×τ,r G ∼= C∗(E,L) ⊗ K((cid:96)2(G))
Indeed if d is consistent with the labeling map L : E1 → A, then
in Corollary 6.8.
C∗(E ×c G,Ld) is equivariantly isomorphic to C∗(E ×c G,L1) (see Proposition 6.3).
For a directed graph E a function c : E1 → Z given by c(e) = 1 for all e ∈ E1 gives rise
to a skew product graph E ×c G whose C∗-algebra which is strongly Morita equivalent
to the fixed point algebra C∗(E)γ for the gauge action. In the case of labelled graphs if
c, d : E1 → Z are given by c(e) = 1, d(e) = 0 for all e ∈ E1, then C∗(E ×c G,Ld) is
strongly Morita equivalent to C∗(E,L)γ (see Theorem 6.10).
An action α of G on a directed graph E induces an action of G on C∗(E), moreover if
the action is free, then using the Gross -- Tucker Theorem we have
C∗(E) ×α,r G ∼= C∗(E/G) ⊗ K((cid:96)2(G))
(1.1)
by [13, Corollary 3.10]. In Theorem 3.2 we show that an action of G on (E,L) induces an
action of G on C∗(E,L). If we wish to use the Gross-Tucker Theorem for labeled graphs
to prove the labeled graph analog (1.1) we need to know when the maps c, d : (E/G)1 → G
provided by Theorem 5.10 are consistent with the quotient labeling L/G. The answer to
this question is provided by Theorem 7.3: It happens precisely when the action α has a
fundamental domain. Hence, if the free action of G on (E,L) has a fundamental domain,
then in Corollary 7.4 we show that
C∗(E,L) ×α,r G ∼= C∗(E/G,L/G) ⊗ K((cid:96)2(G)).
2. Labeled Graphs and their C∗-algebras
We begin with a collection of definitions, which are taken from [2]. A directed graph
E = (E0, E1, r, s) consists of a vertex set E0, an edge set E1, and range and source maps
r, s : E1 → E0. We shall assume throughout this paper that E is row-finite and essential,
that is
r−1(v) (cid:54)= ∅ and 1 ≤ #s−1(v) < ∞
for all v ∈ E0. We let En denote the set of paths of length n and set E+ = ∪n≥1En.
Definition 2.1. A labeled graph (E,L) over an alphabet A consists of a directed graph
E together with a labeling map L : E1 → A.
GROUP ACTIONS
3
We may assume that L : E1 → A is surjective. Let A∗ be the collection of all words in
the symbols of A. For n ≥ 1 the map L extends naturally to a map L : En → A∗: for
λ = λ1 ··· λn ∈ En we set L(λ) = L(λ1)···L(λn) and we say that λ is a representative of
the labeled path L(λ). Let L(En) denote the collection of all labeled paths in (E,L) of
length n. Then L+(E) = ∪n≥1L(En) denotes the collection of all labeled paths in (E,L),
that is all words in the alphabet A which may be represented by paths in E.
Examples 2.2. (a) Every directed graph E gives rise to a labeled graph (E,Lτ ) over the
(b) The directed graph E whose edges e, f, g have been labeled using the alphabet {0, 1}
alphabet E1 where Lτ : E1 → E1 is the identity map.
as shown below is an example of a labeled graph
(E,L):=
1 e
. v
0
f
g
0
. w
Let (E,L) be a labeled graph. Then for β ∈ L+(E) we set
s(β) = {s(λ) : L(λ) = β}.
For A ⊆ E0 and β ∈ L+(E) the relative range of β with respect to A is
r(β) = {r(λ) : L(λ) = β},
r(A, β) = {r(λ) : λ ∈ E+,L(λ) = β, s(λ) ∈ A}.
The labeled graph (E,L) is left-resolving, if for all v ∈ E0 the map L restricted to
r−1(v) is injective. The labeled graph (E,L) is weakly left-resolving if for all A, B ⊆ E0
and β ∈ L+(E) we have
r(A ∩ B, β) = r(A, β) ∩ r(B, β).
If (E,L) is left-resolving then it is weakly left-resolving. Examples 2.2 (a) and (b) are
examples of left-resolving labeled graphs.
A collection B ⊆ 2E0 of subsets of E0 is closed under relative ranges for (E,L) if for
all A ∈ B and β ∈ L+(E) we have r(A, β) ∈ B. If B is closed under relative ranges for
(E,L), contains r(β) for all β ∈ L+(E) and is also closed under finite intersections and
unions, then B is accommodating for (E,L) and the triple (E,L,B) is called a labeled
space. Let E 0.− be the smallest accommodating collection of subsets of E0 for (E,L).
Definition 2.3. For A ⊆ E0 and n ≥ 1, let Ln
those labeled paths of length n whose source intersects A nontrivially.
Though E is row finite it is possible for L1
L1
E0 = E1, which is infinite if E1 is infinite. A labeled space (E,L,B) is set-finite if L1
finite for all A ∈ B. The following definition is given in [2].
Definition 2.4. A representation of a weakly left-resolving, set-finite labeled space (E,L,B)
consists of projections {pA : A ∈ B} and partial isometries {sa : a ∈ A} such that
(i) If A, B ∈ B, then pApB = pA∩B and pA∪B = pA + pB − pA∩B, where p∅ = 0.
(ii) If a ∈ A and A ∈ B, then pAsa = sapr(A,a).
(iii) If a, b ∈ A, then s∗
(iv) For A ∈ B we have
A to be infinite; for example if L is trivial, then
A is
A := {β ∈ L(En) : A ∩ s(β) (cid:54)= ∅} denote
asa = pr(a) and s∗
asb = 0 unless a = b.
(cid:88)
a∈L1
A
pA =
sapr(A,a)s∗
a.
4
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
C∗(E,L,B) is the universal C∗-algebra generated by a representation of (E,L,B). Let
γ : T → Aut C∗(E,L,B) be the gauge action determined by
γzpA = pA, γzsa = zsa for A ∈ B, a ∈ A.
Remark 2.5. The gauge invariant uniqueness Theorem for C∗(E,L,B) as stated in [2,
Theorem 5.3] is incorrect. The authors are grateful to Gow for pointing out the error.
The problem arises in [2, Lemma 5.2 (ii)] as it not possible to prove that the projection
r is nonzero under the hypotheses used in [2]. We are also grateful to Jeong and Kim for
pointing out an mistake in the formula [3, Remark 3.5] and in [9, Example 2.4] which is
a direct result of the error discovered by Gow.
The problem in [2, Lemma 5.2 (ii)] arises because, under the hypotheses on a labeled
space used in [2], it is possible to have A (cid:41) B ∈ B with pA = pB in C∗(E,L,B). To
rectify this problem we must assume that B is closed under relative complements; that is if
A, B ∈ B are such that A (cid:41) B, then A\B ∈ B. If B is closed under relative complements
then we also recover the formula in [3, Remark 3.5].
Before stating the Gauge Invariant Uniqueness Theorem we give a corrected version of [2,
Lemma 5.2] using the new hypothesis.
Lemma 2.6. Let (E,L,B) be a weakly left-resolving, set-finite labeled space where B is
closed under relative complements and {sa, pA} be a representation (E,L,B). Let Y =
: i = 1, . . . , N} be a set of partial isometries in C∗(E,L,B) which is closed
{sαipAis∗
If q is a minimal projection in C∗(Y ) then
under multiplication and taking adjoints.
either
for some 1 ≤ i ≤ N
βi
l=1 sαk(l)pAk(l)s∗
and 1 ≤ i ≤ N ; moreover there
(i) q = sαipAis∗
(ii) q = sαipAis∗
αi
αi
is a nonzero r = sαiβpr(Ai,β)s∗
αiβ ∈ C∗(E,L,B) such that q(cid:48)r = 0 and q ≥ r.
αk(l)
Proof. By [2, Lemma 4.4] any projection in C∗(Y ) may be written as
− q(cid:48) where q(cid:48) =(cid:80)m
n(cid:88)
sαi(j)pAi(j)s∗
j=1
αi(j)
− m(cid:88)
l=1
sαk(l)pAk(l)s∗
αk(l)
If q =(cid:80)n
where the projections in each sum are mutually orthogonal and for each l there is a unique
j such that sαi(j)pAi(j)s∗
l=1 sαk(l)pAk(l)s∗
we must have n = 1. If m = 0 then q = sαipAis∗
is a minimal projection in C∗(Y ) then
for some 1 ≤ i ≤ N . If m (cid:54)= 0 then
≥ sαk(l)pAk(l)s∗
j=1 sαi(j)pAi(j)s∗
αk(l)
αk(l)
αi(j)
αi(j)
.
αi
−(cid:80)m
q = sαipAis∗
αi
sαk((cid:96))pAk((cid:96))s∗
αk((cid:96))
,
where Ai, Ak((cid:96)) ∈ B for 1 ≤ (cid:96) ≤ m. If we apply Definition 2.4 (iv) we may write
q =
sαiβj pr(Ai,βj )s∗
αiβj
sαk((cid:96))κhpr(Ak((cid:96)),κh)s∗
αk((cid:96))κh
j=1
h=1
(cid:96)=1
where all αiβj and αk((cid:96))κh have the same length. Since q is a nonzero projection there is
1 ≤ j ≤ n and Hj ⊆ {1, . . . , t} × {1, . . . , m} such that αiβj = αk((cid:96))κh for all (h, (cid:96)) ∈ Hj
and
r(Ak((cid:96)), κh) (cid:40) r(Ai, βj).
(cid:91)
Yj :=
(h,(cid:96))∈Hj
− m(cid:88)
− t(cid:88)
(cid:96)=1
m(cid:88)
n(cid:88)
GROUP ACTIONS
5
αi
αiβj
is nonzero and q ≥ r since Xj ⊂ r(Ai, βj).
Since B is closed under finite unions we have Yj ∈ B. Then for this j define Xj =
(cid:54)= ∅, then Xj ∈ B since B is closed under relative complements. Hence
r(Ai, βj)\Yj
the projection r = sαiβj pXj s∗
If we set
− q then since Xj ∩ Yj = ∅ we have q(cid:48)r = 0 as required.
(cid:3)
q(cid:48) = sαipAis∗
Theorem 2.7 (Gauge invariant uniqueness Theorem). Let (E,L,B) be a weakly left-
resolving, set-finite labeled space where B is closed under relative complements and {Sa, PA}
be a representation (E,L,B) on Hilbert space. Take πS,P to be the representation of
C∗(E,L,B) satisfying πS,P (sa) = Sa and πS,P (pA) = PA. Suppose that PA (cid:54)= 0 for all
∅ (cid:54)= A ∈ B and that there is a strongly continuous action γ(cid:48) of T on C∗({Sa, PA}) such
that for all z ∈ T, γ(cid:48)
Proof. The proof is the same as given in [2, Theorem 5.3], using Lemma 2.6 instead of [2,
(cid:3)
Lemma 5.2].
Definition 2.8. Let (E,L) be a weakly left-resolving, set-finite labeled graph, then we
define E(r,L) to be the smallest accommodating collection of subsets of E0 which is closed
under relative complements.
Remark 2.9. Every A ∈ E 0,− can be written as A = ∪n
i ) and
i ∈ L+(E) for all i, j. Hence, by applications of de Morgan's laws we may show that
βj
i )\r(βj
every A ∈ E(r,L) can be written in the form A = ∪n
i )
where r(αj
j=1Aj where Aj = ∩m(j)
j=1Aj where Aj = ∩m(j)
z ◦ πS,P = πS,P ◦ γz. Then πS,P is faithful.
i ∈ L+(E) for all i, j
i ) (cid:41) r(βj
i=1 r(αj
i ) and αj
i , βj
i=1 r(βj
This Remark motivates the following definition.
Definition 2.10. Let (E,L) be a weakly left-resolving, set-finite labeled graph. A Cuntz-
Krieger (E,L)-family consists of commuting projections {pr(β) : β ∈ L+(E)} and partial
isometries {sa : a ∈ A} with the properties that:
(CK1a) For all β, ω ∈ L+(E), pr(β)pr(ω) = 0 if and only if r(β) ∩ r(ω) = ∅.
(CK1b) For all β, ω, κ ∈ L+(E), if r(β)∩r(ω) = r(κ), then pr(β)pr(ω) = pr(κ), if r(β)∪r(ω) =
r(κ), then pr(β) + pr(ω)− pr(β)pr(ω) = pr(κ) and if r(β) (cid:41) r(ω), then pr(β)− pr(ω) (cid:54)= 0.
(CK2) If a ∈ A and β ∈ L+(E), then pr(β)sa = sapr(βa).
(CK3) If a, b ∈ A, then s∗
(CK4) For β ∈ L+(E), if L1
asa = pr(a) and s∗
r(β) is finite and non-empty, then we have
asb = 0 unless a = b
(2.1)
pr(β) =
sapr(βa)s∗
a.
(cid:88)
a∈L1
r(β)
Let C∗(E,L) be the universal C∗-algebra generated by a Cuntz-Krieger (E,L)-family.
Let γ(cid:48) : T → Aut C∗(E,L) be the gauge action determined by
γ(cid:48)
zpr(β) = pr(β), γ(cid:48)
zsa = zsa for β ∈ L+(E), a ∈ A.
Theorem 2.11. Let (E,L) be a weakly left-resolving, set-finite labeled graph. Then
C∗(E,L) is isomorphic to C∗(E,L,E(r,L)); moreover
C∗(E,L) = span{sαpAs∗
β : α, β ∈ L+(E), A ∈ E(r,L)}.
Proof. Let {sa, pr(β)} be a universal Cuntz-Krieger (E,L)-family and {ta, qA} be a uni-
versal representation of the labeled space (E,L,E(r,L)). For a ∈ A, set Ta = sa.
By (CK1a) we may define Q∅ = 0. For α, β ∈ L+(E) we may define Qr(α)∩r(β) =
Qr(α)Qr(β) and Qr(α∪r(β) = Qr(α) + Qr(β) − Qr(α)∩r(β) in C∗(E,L). If r(α) (cid:41) r(β) then
6
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
we may define Qr(α)\r(β) = Qr(α) − Qr(β) (cid:54)= 0 in C∗(E,L).By Remark 2.9 and using the
inclusion/exclusion law we may define QA in C∗(E,L) for all A ∈ E(r,L).
It is a routine calculation to show that {Ta, QA} is a representation of the labeled space
(E,L,E(r,L)) in C∗(E,L). By the universal property of C∗(E,L,E(r,L)) there exists a
homomorphism Φ : C∗(E,L,E(r,L)) → C∗(E,L) such that Φ(ta) = Ta
and Φ(qA) =
z ◦ Φ = Φ ◦ γz for z ∈ T. The first statement then
QA. It is straightforward to see that γ(cid:48)
follows by Theorem 2.7, and the final statement follows by applying Φ to an arbitrary
element of C∗(E,L,E(r,L)) (see [2, Lemma 4.4]).
(cid:3)
3. Automorphisms of Labeled graphs and their C∗-algebras
We begin by defining what a labeled graph morphism is and use the definition to define
a labeled graph automorphism. Then in Theorem 3.2 we show that a labeled graph
automorphism of (E,L) induces an automorphism of C∗(E,L).
Definition 3.1. Let (E,L) and (F,M) be labeled graphs over alphabets AE and AF
respectively. A labeled graph morphism is a triple φ := (φ0, φ1, φAE ) : (E,L) → (F,M)
such that
(a) For all e ∈ E1 we have φ0(r(e)) = r(φ1(e)) and φ0(s(e)) = s(φ1(e));
(b) φAE : AE → AF is a map such that M ◦ φ1 = φAE ◦ L.
If the maps φ0, φ1, φAE are bijective, then the triple φ := (φ0, φ1, φAE ) is called a labeled
graph isomorphism. In the case that F = E, AE = AF and L = M we call (φ0, φ1, φA) a
labeled graph automorphism.
For a labeled graph morphism φ = (φ0, φ1, φAE ) we shall omit the superscripts on φ when
the context in which it is being used is clear.
The set Aut(E,L) := {φ : φ is a labeled graph automorphism of (E,L)} forms a group
under composition. The following result follows easily from the universal definition of
C∗(E,L).
Theorem 3.2. Let φ be an automorphism of a weakly left-resolving, set-finite labeled graph
(E,L) and {sa, pr(β)} be a universal Cuntz-Krieger (E,L)-family. The maps sa (cid:55)→ sφ(a)
and pr(β) (cid:55)→ pφ(r(β)) induce an automorphism of C∗(E,L).
4. Skew product labeled graphs and group actions
In this section we shall define a skew product labeled graph and define what it means for
a group to act on a labeled graph.
Definition 4.1. Let (E,L) be a labeled graph and let c, d : E1 → G be functions. The
skew product labeled graph (E ×c G,Ld) over alphabet A× G consists of the skew product
graph (E0 × G, E1 × G, rc, sc) where
rc(e, g) = (r(e), gc(e))
sc(e, g) = (s(e), g)
together with the labeling Ld : (E ×c G)1 → A × G given by Ld(e, g) := (L(e), gd(e)).
Since the labels received by (v, g) ∈ (E ×c G)0 are in one-to-one correspondence with the
labels received by v ∈ E0 it follows that if (E,L) is left-resolving, then so is (E ×c G,Ld).
Example 4.2. For the labeled graph (E,L) of Examples 2.2(b) let c, d : E1 → Z be given
by c(e) = 1 and d(e) = 0 for all e ∈ E1. Then
GROUP ACTIONS
7
(E ×c Z,Ld) :=
. . .
(v,0)
.
(1,0)
(v,1)
.
(1,1)
(v,2)
.
(1,2)
(v,3)
.
. . .
(0,0)
(0,0)
. . .
.
(w,0)
(0,1)
(0,1)
.
(w,1)
(0,2)
(0,2)
.
(w,2)
.
. . .
(w,3)
Remark 4.3. We shall use the following simpler description of the path space of E ×c G.
For v ∈ E0, e ∈ E1, g ∈ G set vg = (v, g), eg = (e, g). Then for µ ∈ En where n ≥ 2 and
g ∈ G set
For µ ∈ E∗ the map (µ, g) (cid:55)→ µg identifies E∗×G with (E×c G)∗. Then for (µ, g) ∈ E∗×G
we have
µg = (µ1, g)(µ2, gc(µ1))··· (µn, gc(µ(cid:48))) ∈ (E × G)n.
s(µ, g) = (s(µ), g) and r(µ, g) = (r(µ), gc(µ)).
(4.1)
Let (E,L) be a labeled graph over the alphabet A. A labeled graph action of G on (E,L) is
a triple ((E,L), G, φ) where φ : G → Aut(E,L) is a group homomorphism. In particular,
for all e ∈ E1 and g ∈ G we have
(4.2)
L(φg(e)) = φg(L(e)).
If we ignore the label maps, a labeled graph action ((E,L), G, φ) restricts to a graph
action of G on E; we denote this restricted action by (E, G, φ). The labeled graph action
((E,L), G, α) is free if φg(v) = v for some v ∈ E0, then g = 1G and if φg(a) = a some
a ∈ A, then g = 1G.
The following lemma shows that skew product labeled graphs provide a rich source of
g , τ 1
phism.
g , τA
g ) is a labeled graph automorphism.
τA
g (a, h) = (a, gh). Then τg = (τ 0
examples of free labeled graph actions. As the proof is routine, we omit it.
Lemma 4.4. Let (E,L) be a labeled graph, c, d : E1 → G be functions and (E ×c G,Ld)
be the associated skew product labeled graph. Then
(i) For (x, h) ∈ (E ×c G)i, (a, h) ∈ A × G, g ∈ G and i = 0, 1 let τ i
g(x, h) = (x, gh) and
(ii) The map τ = (τ 0, τ 1, τA) : G → Aut(E ×c G,Ld) defined by g (cid:55)→ τg is a homomor-
(iii) The triple ((E ×c G,Ld), G, τ ) is a free labeled graph action.
Definition 4.5. The map τ = (τ 0, τ 1, τA) : G → Aut(E×cG,Ld) as given in Lemma 4.4 (ii)
is called the left labeled graph translation map, and the action ((E ×c G,Ld), G, τ ) the left
labeled graph translation action.
Two labeled graph actions ((E,L), G, φ) and ((F,M), G, ψ) are isomorphic if there is a
labeled graph isomorphism ϕ : (E,L) → (F,M) which is equivariant in the sense that
ϕ ◦ φg = ψg ◦ ϕ for all g ∈ G.
Theorem 4.6. Let (E,L) be a weakly left-resolving, set-finite labeled graph, and ((E,L), G, α)
be a labeled graph action. Let {sa, pr(β)} be a universal Cuntz-Krieger (E,L)-family. Then
for h ∈ G the maps
αhsa = sαha and αhpr(β) = pαhr(β)
8
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
determine an an action of G on C∗(E,L). If ((E,L), G, φ) and ((F,M), G, ψ) are iso-
morphic then C∗(E,L) ×φ G ∼= C∗(F,M) ×ψ G.
Proof. Follows by a straightforward application of Theorem 3.2 and the universal property
(cid:3)
of crossed products.
5. Gross-Tucker Theorem
In this section we prove a version of the Gross-Tucker theorem for labeled graphs. For
directed graphs, the Gross-Tucker theorem says, roughly speaking, that up to equivariant
isomorphism, every free action α of a group G on a directed graph E is a left translation
automorphism τ on a skew product graph (E/G)×c G built from the quotient graph E/G.
Our aim is to prove a similar result for labeled graphs. The new ingredient is the map
d : E1 → G found in the definition of a skew product labeled graph for labeled graphs.
Before giving our main result, Theorem 5.10, we introduce some notation.
Definitions 5.1. Let ((E,L), G, α) be a labeled graph action. For i = 0, 1 and x ∈ Ei let
g (a) : g ∈ G}
Gx := {αi
and A/G = {Ga : a ∈ A}.
The proof of the following lemma is straightforward, so we omit it.
Lemma 5.2. Let ((E,L), G, α) be a labeled graph action. The maps r, s : (E/G)1 →
(E/G)0 given by
g(x) : g ∈ G} and (E/G)i = {Gx : x ∈ Ei}. For a ∈ A let Ga = {αA
r(Ge) = Gr(e) and s(Ge) = Gs(e) for Ge ∈ (E/G)1
(5.1)
and the map L/G : (E/G)1 → A/G given by (L/G)(Ge) = GL(e) are well-defined.
Consequently, (E/G,L/G) is a labeled graph over the alphabet A/G.
The map q = (q0, q1, qA) : (E,L) → (E/G,L/G) given by qi(x) = Gx for i = 0, 1, x ∈
Ei and qA(a) = Ga for a ∈ A is a surjective labeled graph morphism.
Definition 5.3. Let ((E,L), G, α) be a labeled graph action. The quotient labeled graph
(E/G,L/G) is the labeled graph described in Lemma 5.2, the map q : (E,L) → (E/G,L/G)
is the quotient labeled map.
The following Proposition is an analog of [8, Theorem 2.2.1] whose proof is routine, and
so we omit it.
Proposition 5.4. Let (E,L) be a labeled graph, c, d : E1 → G be functions and (E ×c
G,Ld) be the associated skew product labeled graph. Let ((E ×c G,Ld), G, τ ) be the left
labeled graph translation action. Then
((E ×c G)/G,Ld/G) ∼= (E,L).
Example 5.5. Recall the labeled graphs (E,L) and (E ×c Z,Ld) from Example 4.2. For
the left labeled graph translation action ((E×cZ,Ld), Z, τ ) we have ((E×cZ)/Z,Ld/Z) ∼=
(E,L) by Proposition 5.4.
The Gross-Tucker theorem is a converse to Proposition 5.4. It states that if we have a
free action of a group on a labeled graph, then we can recover the original graph from the
quotient via a skew product. Recall the following definition for directed graphs.
Definition 5.6. Let F, E be directed graphs. A surjective graph morphism p : F → E
has the unique path lifting property if given u ∈ F 0 and e ∈ E1 with s(e) = p0(u) there is
a unique edge f ∈ F 1 with s(f ) = u and p1(f ) = e.
GROUP ACTIONS
9
Remark 5.7. Let (E, G, α) be a free graph action. Then the quotient map q : E → E/G
has the unique path lifting property (see [13, §5] or [8, p.67] for instance).
Definitions 5.8. Let ((E,L), G, α) be a labeled graph action and q = (q0, q1, qA) :
(E,L) → (E/G,L/G) be the quotient labeled map. A section for qi is a map ηi
:
(E/G)i → Ei for i = 0, 1 such that qi ◦ ηi = id(E/G)i. A section for qA is ηA : A/G → A
such that qA ◦ ηA = idA/G.
Lemma 5.9. Let (E, G, α) be a graph action and q = (q0, q1) : E → E/G be the quotient
map. Given a section η0 for q0 there is a unique section η1 for q1 such that
s(η1(Ge)) = η0(s(Ge)) for all e ∈ E1.
(5.2)
Proof. By Remark 5.7 the quotient map q : E → E/G has the unique path lifting property.
Hence if we fix Gv ∈ (E/G)0, then for each Ge ∈ (E/G)1 with s(Ge) = Gv there is a
unique f ∈ E1 with q1(f ) = Ge = Gf and s(f ) = η0(Gv). Put η1(Ge) = f , then
η1 : (E/G)1 → E1 is well-defined and the source map on (E/G)1 is well-defined. Since
q1(η1(Ge)) = q1(f ) = Ge it follows that η1 is a section satisfying (5.2). Uniqueness of η1
(cid:3)
follows from the unique path lifting property of q.
The following is a version of the Gross-Tucker Theorem (cf. [8, Theorem 2.2.2]) for labeled
graphs.
Theorem 5.10. Let ((E,L), G, α) be a free labeled graph action. Let η0, ηA be sections
for q0, qA respectively. There are functions c, d : (E/G)1 → G such that ((E,L), G, α) is
isomorphic to ((E/G ×c G, (L/G)d), G, τ ).
Proof. Fix a section η0 : (E/G)0 → E0 for q0. By Lemma 5.9 there is a section η1 for
q1satisfying (5.2). For Ge ∈ (E/G)1 set f = η1(Ge), then
q0(r(η1(Ge))) = q0(r(f )) = Gr(f ) = r(Gf ) = r(Ge) = q0(η0(r(Ge))).
As (E, G, α) is free, there is a unique h ∈ G such that α0
may set c(Ge) = h. Define φ : E/G ×c G → E by
gη0(Gv) and φ1
φ0
c(Gv, g) = α0
c(Ge, g) = α1
gη1(Ge)
hη0(r(Ge)) = r(η1(Ge)) and we
for (Gv, g) ∈ (E/G×c G)0 and (Ge, g) ∈ (E/G×c G)1. One checks that φc : (E/G×c G) →
E is an isomorphism of directed graphs.
We claim that φc is equivariant. Notice that for all (Gv, h) ∈ (E/G ×c G)0 and g ∈ G
we have
c(Gv, gh) = α0
ghη0(Gv) = α0
gα0
hη0(Gv) = α0
gφ0
c(Gv, h)
φ0
c(τ 0
c ◦ τ 0
g (Gv, h)) = φ0
g = α0
g ◦ φ0
c for all g ∈ G. The argument for φ1
A/G×G
We now construct an equivariant bijection φ
d
and so φ0
c is similar and our claim
follows.
: A/G × G → A which satisfies
condition (b) of Definition 3.1. Fix a section ηA : A/G → A for qA. We now define a
map d : (E/G)1 → G. Fix Ge ∈ (E/G)1 and set f = η1(Ge) so that q1(f ) = Ge. Since
qAηA(L/G(Ge)) = qAηA(GL(f )) = qALη1(Ge)
and the graph action ((E,L), G, α) is free, there is a unique k ∈ G such that αA
L(η1(Ge)) and we may define d(Ge) = k. The function d : (E/G)1 → G described in this
way is such that d(Ge) is the unique element of G with the property that
k ηA((L/G)(Ge)) =
(5.3)
d(Ge)ηA((L/G)(Ge)) = L(η1(Ge)).
αA
10
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
For each (Ga, g) ∈ A/G × G we define φ
g ηA(Ga). We claim that φ
αA
(Ge, h) ∈ (E/G ×c G)1 we have
A/G×G
d
A/G×G
d
A/G×G
satisfies φ
d
: A/G × G → A by φ
◦ (L/G)d = L ◦ φ1
A/G×G
d
(Ga, g) =
c: By (5.3) for all
A/G×G
φ
d
◦ (L/G)d(Ge, h) = αA
h αA
d(Ge)ηA(L/G(Ge)) = L(α1
hη1(Ge)) = L ◦ φ1
c(Ge, h)
as required.
It is straightforward to see that φ
A/G×G
d
A/G×G
is bijective. To see that φ
d
is equivariant
notice that we have
(τA/G×G
A/G×G
φ
d
g
A/G×G
(Ge, h)) = φ
d
(Ge, gh) = αA
g αA
for all (Ge, h) ∈ (E/G × G)1 and g ∈ G. Thus φc,d = (φ0
labeled graph isomorphism.
A/G×G
h ηA(Ge) = αA
g φ
d
A/G×G
c, φ
d
c, φ1
(Ge, h)
) is the required
(cid:3)
Remark 5.11. The possibility that two edges in the quotient graph have the same label
means that we must choose a separate section ηA for qA. In turn means that the function
d given in the definition of a skew product labeled graph plays a crucial role in the
reconstruction of the labeled graph action in Theorem 5.10.
Example 5.12. Recall from Example 5.5 the labeled graph (E ×c Z,Ld) has a free action
of Z such that the quotient labeled graph is (E,L). We use this example to illustrate the
point made in Remark 5.11:
Suppose we choose a section η0 : E0 → (E ×c Z)0 such that η0(v) = (v, 0) and η0(w) =
(w, 2), then the section η1 : E1 → (E ×c Z)1 as defined in Lemma 5.9 is given by η1(e) =
(e, 0), η1(f ) = (f, 0), and η1(g) = (g, 2) whose image in (E ×c Z,Ld) is as shown below.
. . .
(v,0)
.
(v,1)
.
(1,0)
(e,0)
(v,2)
.
(v,3)
.
. . .
. . .
.
(w,0)
(f,0)
(0,0)
.
(w,1)
(0,2)
(g,2)
.
.
. . .
(w,2)
(w,3)
Note that c(e) = 1, c(f ) = −1, and c(g) = 3.
Observe that f, g ∈ E1 are such that L(f ) = L(g) = 0 however,
L(η1(f )) = L(f, 0) = (0, 0) (cid:54)= (0, 2) = L(g, 2) = L(η1(g))
The function d accounts for this difference. By Equation (5.3) we have d(g) = 2, since
2 (0, 0) = (0, 2), whereas d(f ) = 0. Observe that d(g) (cid:54)= d(f ) even though L(g) = L(f ).
αA
6. Coactions on Labeled Graph Algebras
In [11] it is shown that a function c : E1 → G induces a coaction δ of G on the
graph algebra C∗(E) such that C∗(E)×δ G ∼= C∗(E ×c G). One should expect, therefore,
that the functions c, d : E1 → G would induce a coaction δ of G on C∗(E,L) such
that C∗(E,L) ×δ G ∼= C∗(E ×c G,Ld). However in order to obtain such a result we must
assume that both functions c, d are label consistent (see Definition 6.1 below). For further
information about coactions of discrete groups see [18], amongst others.
Definition 6.1. Let (E,L) be a labeled graph over alphabet A. A function c : E1 → G
is label consistent if there is a function C : A → G such that c = C ◦ L.
GROUP ACTIONS
11
For any labeled graph (E,L) the function 1 : E1 → G given by 1(e) = 1G for all e ∈ E1
is label consistent. Firstly we show that if c is label consistent then there is a coaction of
G on C(E,L).
Proposition 6.2. Let (E,L) be a weakly left-resolving, set-finite labeled graph, G be a
discrete group, and c : E1 → G be a label consistent function. Then there is a maximal,
normal coaction δ : C∗(E,L) → C∗(E,L) ⊗ C∗(G) such that
(6.1)
where {sa, pr(β)} is a universal Cuntz-Krieger (E,L)-family and {ug : g ∈ G} are the
canonical generators of C∗(G).
δ(sa) = sa ⊗ uC(a) and δ(pr(β)) = pr(β) ⊗ u1G
Proof. The first part of the result follows by the same argument given in [11, Lemma 3.2].
That the coaction δ is normal and maximal follows by essentially the same arguments as
(cid:3)
the ones given in [6, Lemma 3.3] and [16, Theorem 7.1 (v)].
The next result shows that if d is label consistent then we may as well assume that d = 1.
Proposition 6.3. Let (E,L) be a weakly left-resolving, set-finite labeled graph and c :
E1 → G a function.
If d1, d2 : E1 → G are label consistent functions, then ((E ×c
G,Ld1), G, τ ) ∼= ((E ×c G,Ld2), G, τ ) where τ is the left translation action. Hence if
d : E1 → G is a label consistent function then there is an isomorphism from C∗(E×cG,Ld)
to C∗(E ×c G,L1) which is equivariant for the G -- action induced by τ .
Proof. For the first statement let φi : (E ×c G)i → (E ×c G)i be the identity map for
i = 0, 1 and define φA×G : A × G → A × G by
For (e, g) ∈ (E ×c G)1, after a short calculation we have
φA×G(a, g) = (a, gD−1
1 (a)D2(a)).
φA×GLd1(e, g) = (L(e), d2(e)) = Ld2(e, g).
It is then straightforward to check that φ = (φ0, φ1, φA×G) is a labeled graph isomorphism.
Since for all h ∈ G we have
τh(φA×G(a, g)) = (a, hgD−1
1 (a)D2(a)) = φA×G(τh(a, g))
it follows that ((E ×c G,Ld1), G, τ ) ∼= ((E ×c G,Ld2), G, τ ).
The final statement follows from Theorem 4.6.
(cid:3)
Remark 6.4. Thanks to Proposition 6.3 we may, without loss of generality, assume that
d = 1 when we are working with label consistent d-functions. On the other hand it is not
hard to see that a different choice of label consistent functions c will yield non-isomorphic
skew-product graphs.
Next we shall show that if d = 1 then there is a natural identification L+
labeled path space of (E ×c G,L1) with L+(E) × G.
Lemma 6.5. Let (E,L) be a labeled graph and c : E1 → G label consistent. For µ ∈ E+
and g ∈ G the map
1 (E ×c G), the
establishes a bijection from L+
L1(µ, g) (cid:55)→ (L(µ), g)
1 (E ×c G) to L+(E) × G.
12
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
Proof. From Remark 4.3 it follows that for n ≥ 1 every path in (E ×c G)n has the form
(µ, g) = (µ1, g)(µ2, gc(µ1))··· (µn, gc(µ(cid:48))), for some µ ∈ En and g ∈ G. Then by definition
we have
L1(µ, g) = (L(µ1), g)(L(µ2), gc(µ1))··· (L(µn), gc(µ(cid:48))).
(6.2)
If we define the right hand side of (6.2) to be (L(µ), g) the result follows.
The following Lemma indicates the behavior of the range map under the identification of
L+
1 (E ×c G) with L+(E) × G.
Lemma 6.6. Let (E,L) be a labeled graph and c : E1 → G be a label consistent function.
Let a ∈ A, β ∈ L+(E), and g ∈ G. Then under the identification of L+(E) × G with
L+
1 (E ×c G) we have r(β, g) = (r(β), gC(β)) ∈ E(r,L) × G.
Proof. Observe that for (β, g) ∈ L+(E) × G we have
(cid:3)
r(β, g) = {r(µ, g) : (µ, g) ∈ E∗ × G,L(µ) = β} by (6.2)
= {(r(µ), gC(β)) : L(µ) = β} by (4.1)
(6.3)
since the function c : E1 → G is label consistent. Hence we may identify r(β, g) with
(r(β), gC(β)) ∈ E(r,L) × G.
(cid:3)
With the above identifications in mind, we turn our attention to the main result of this
section. By Theorem 4.6 the left labeled graph translation action ((E ×c G,L1), G, τ )
defined in Definition 4.5 induces an action τ : G → Aut C∗(E ×c G,L1). When we
identify L+
1 (E ×c G) with L+(E) × G this action may be described on the generators of
C∗(E ×c G,L1) as follows: For h, g ∈ G, a ∈ A, and β ∈ L+(E) we have
(6.4)
τh(s(a,g)) = s(a,hg) and τh(p(r(β),g)) = p(r(β),hg).
The method of proof for the next result closely follows that of [11, Theorem 2.4], however
we give some of the details as they rely heavily on the identification we made in Lemma 6.6.
Theorem 6.7. Let (E,L) be a weakly left-resolving, set-finite labeled graph. Suppose that
G is a discrete group, c : E1 → G is a label consistent function, and δ is the coaction
from Proposition 6.2. Let jC∗(E,L), jG denote the canonical covariant homomorphisms of
C∗(E,L) and C∗(G) into M (C∗(E,L) ×δ G) and {s(a,g), p(r(β),g)} be the canonical gener-
ating set of C∗(E ×c G,L1). Then the map φ : C∗(E ×c G,L1) → C∗(E,L) ×δ G given
by
φ(s(a,g)) = jC∗(E,L)(sa)jG(χC(a)−1) φ(p(r(β),g)) = jC∗(E,L)(pr(β))jG(χg−1)
is an isomorphism.
Sketch of proof. For each g ∈ G, let C∗(E,L)g = {b ∈ C∗(E,L) : δ(b) = b ⊗ ug} denote
the corresponding spectral subspace; we write bg to denote a generic element of C∗(E,L)g.
Then C∗(E,L)×δ G is densely spanned by the set {(bg, h) : bg ∈ C∗(E,L)g and g, h ∈ G},
and the algebraic operations are given on this set by
(bg, x)(bh, y) = (bgbh, y) if y = h−1x (and 0 if not), and (bg, x)∗ = (b∗
g, gx).
If (jC∗(E,L), jG) denotes the canonical covariant homomorphism of C∗(E,L) into the mul-
tiplier algebra of C∗(E,L) ×δ G, then (bg, x) is by definition (jC∗(E,L)(bg)jG(χ{x})).
Using Lemma 6.6 we may show that for (a, g) ∈ A × G, β ∈ L+(E) and g ∈ G
t(a,g) = (sa, C(a)−1g−1) and q(r(β),g) = (pr(β), g−1)
is a Cuntz-Krieger (E ×c G,L1)-family in C∗(E,L) ×δ G.
GROUP ACTIONS
13
By universality of C∗(E ×c G,L1) there is a homomorphism πt,q from C∗(E ×c G,L1)
to C∗(E,L) ×δ G such that πt,q(s(a,g)) = t(a,g) and πt,q(p(r(β),g)) = q(r(β),g) which we may
show is injective using the argument from [11, Theorem 2.4] and Theorem 2.7.
Next we show that πt,q is surjective. Observe that C∗(E,L) ×δ G is generated by
(sa, g) and (pr(β), h). Since πt,q(s(a,g−1C(a)−1)) = t(a,g−1C(a)−1) = (sa, C(a)−1C(a)g), and
πt,q(p(r(β),h−1)) = (pr(β), h) we see that πt,q is surjective. Hence πt,q is the desired isomor-
phism.
all g ∈ G. It is enough to check on generators: Notice that for all s(a,h) ∈ C∗(E ×c G,L1)
We need to check that πt,q is equivariant for the G actions, that is πt,q ◦ τg = (cid:98)δg ◦ πt,q for
πt,q ◦ τg(s(a,h)) = πt,q(s(a,gh)) = (sa, C(a)−1h−1g−1) =(cid:98)δg(sa, C(a)−1h−1) = (cid:98)δg ◦ πt,q(s(a,h))
and similarly πt,q ◦ τg(p(r(β),h)) = (cid:98)δg ◦ πt,q(p(r(β),h)) for p(r(β),h) ∈ C∗(E ×c G,L1).
We claim that πt,q is equivariant for the T actions, that is πt,q◦ γz = (γz× G)◦ πt,q for all
z ∈ T. It is enough to check this on generators: Notice that for all s(a,h) ∈ C∗(E ×c G,L1)
and z ∈ T we have
πt,q ◦ γz(s(a,h)) = πt,q(zs(a,h)) = (zsa, C(a)−1h−1) = (γz × G)(sa, C(a)−1h−1)
= (γz ×δ G) ◦ πt,q(s(a,h)).
Similarly πt,q ◦ γz(p(r(β),h)) = (γz × G) ◦ πt,q(p(r(β),h)) for all p(r(β),h) ∈ C∗(E ×c G,L1). (cid:3)
Corollary 6.8. Let (E,L) be a weakly left-resolving, set-finite labeled graph. Suppose that
G is a discrete group, c : E1 → G be a label consistent function, and τ the induced action
of G on C∗(E ×c G,L1). Then
C∗(E ×c G,L1) ×τ,r G ∼= C∗(E,L) ⊗ K((cid:96)2(G)).
Proof. Since the isomorphism of C∗(E ×c G,L1) with C∗(E,L)×δ G is equivariant for the
G-actions τ,(cid:98)δ, respectively, it follows that
C∗(E ×c G,L1) ×τ,r G ∼= C∗(E,L) ×δ G ×(cid:98)δ,r G.
Following the argument in [11, Corollary 2.5], Katayama's duality theorem [12] gives us
(cid:3)
that C∗(E,L) ×δ G ×(cid:98)δ,r G is isomorphic to C∗(E,L) ⊗ K((cid:96)2(G)), as required.
In order to provide a version of Corollary 6.8 for group actions we must first characterise
when the functions c, d in the Gross-Tucker Theorem 5.10 are label consistent maps. We
will do this in the next section.
Recall from [18, p.209] that a coaction δ of a discrete group G on a C∗-algebra A
is saturated if for each s ∈ G we have AsA∗
s = Aδ where As is the spectral subspace
As = {b ∈ A : δ(b) = b ⊗ us} and Aδ is the fixed point algebra for δ
Aδ := {b ∈ A : δ(a) = a ⊗ u1G}.
Lemma 6.9. Let (E,L) be a weakly left-resolving, set-finite labeled graph and c : E1 → Z
be given by c(e) = 1 for all e ∈ E1. Then the coaction δ of Z on C∗(E,L) induced by c is
saturated.
Proof. The coaction δ of Z on C∗(E,L) defined in Proposition 6.2 is such that the fixed
point algebra C∗(E,L)δ is precisely the fixed point algebra C∗(E,L)γ for the canonical
gauge action of T on C∗(E,L) by the Fourier transform (cf. [5, Corollary 4.9]. By an
argument similar to that in [17, §2] we have
C∗(E,L)γ = span{sαpAs∗
β : α, β ∈ Ln(E), A ∈ E(r,L)}.
14
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
Since E has no sinks it follows by a similar argument to that in [17, Lemma 4.1.1] that
C∗(E,L) is saturated.
(cid:3)
Theorem 6.10. Let (E,L) be a weakly left-resolving, set-finite labeled graph. Then
C∗(E,L)γ is strongly Morita equivalent to C∗(E ×c Z,L1) where c : E1 → Z is given
by c(e) = 1 for all e ∈ E1.
Proof. Since c is label consistent it follows by Theorem 6.7 that
C∗(E ×c Z,L1) ∼= C∗(E,L) ×δ Z.
By Lemma 6.9 the coaction is δ is saturated and since C∗(E,L)δ ∼= C∗(E,L)γ the result
(cid:3)
follows.
7. Free group actions on labeled graphs
Recall that a fundamental domain for a graph action (E, G, α) is a subset T of E0 such
gw. Every
In this section we examine conditions on the free labeled graph action ((E,L), G, α) which
ensure that the functions c, d from Theorem 5.10 are label consistent.
that for every v ∈ E0 there exists g ∈ G and a unique w ∈ T such that v = α0
free graph action has a fundamental domain.
Definition 7.1. Let ((E,L), G, α) be a free labeled graph action. A fundamental domain
for ((E,L), G, α) is a fundamental domain T ⊆ E0 for the restricted graph action such
that for every e, f ∈ E1 we have
(a) if r(e), r(f ) ∈ T and GL(e) = GL(f ), then L(e) = L(f ) and
(b) if s(e), s(f ) ∈ T and GL(e) = GL(f ), then L(e) = L(f ).
In Examples 7.2 (i) below we see that not every free action of a group on a labeled graph
has a fundamental domain.
Examples 7.2.
(i) Consider the following labeled graph
(v,−1)
.
(1,−1)
(v,0)
.
. . .
(1,0)
f
(v,1)
.
(1,1)
(v,2)
.
. . .
(E,L) :=
(0,−1)
(0,−2)
. . .
.
(w,−1)
(1,1)
(0,0)
(0,−1)
.
(w,0)
(0,1)
(0,0)
.
(w,1)
e
(1,3)
.
. . .
(w,2)
(1,2)
The group Z acts freely on (E,L) by addition in the second coordinate of the ver-
tices, edges and labels as indicated in the picture above; call this action α. Let
T = {(v, 0), (w, 1)}, then T is a fundamental domain for the restricted graph action
(E, Z, α). However when considering the labeled graph action ((E,L), Z, α) the set
T does not satisfy Definition 7.1 (b).
Consider the edges e, f as shown above with L(e) = (1, 3) and L(f ) = (1, 0)
respectively. We have s(e) = (w, 1) ∈ T and s(f ) = (v, 0) ∈ T and ZL(e) =
ZL(f ) = {(1, n) : n ∈ Z}, however L(e) = (1, 3) (cid:54)= (1, 0) = L(f ).
Indeed any
fundamental domain for the restricted action (E, Z, α) will also fail Definition 7.1
(b).
(ii) Let c, d : E1 → G be label consistent functions and ((E ×c G,Ld), G, τ ) be the
associated left labeled graph translation action. Then one checks that T = {(v, 1G) :
v ∈ E0} is a fundamental domain for ((E ×c G,Ld), G, τ ).
GROUP ACTIONS
15
The following result shows that when we add the fundamental domain hypothesis to the
free labeled graph action, the functions c, d : (E/G)1 → G in the labeled graph version of
the Gross-Tucker Theorem (Theorem 5.10) may be chosen to be label consistent.
Theorem 7.3. Let ((E,L), G, α) be a free labeled graph action with a fundamental do-
main. Then there are label consistent functions c, d : (E/G)1 → G such that ((E,L), G, α) ∼=
((E/G) ×c G, (L/G)d), G, τ ).
Proof. Let T be a fundamental domain for ((E,L), G, α). For every Gv ∈ (E/G)0 there
exists a unique w ∈ T such that Gw = Gv. Hence if we define η0(Gv) = w, then η0 :
(E/G)0 → T is a section for q0. Then we may define η1, c, d, and ηA as in Theorem 5.10.
It suffices to show that c and d are label consistent. To see that d is label consistent
suppose Ge, Gf ∈ (E/G)1 are such that (L/G)(Ge) = (L/G)(Gf ) = Ga ∈ A/G. Let
b = ηA(Ga) ∈ A, d(Ge) = k ∈ G, and d(Gf ) = l ∈ G. Then by the definition of d we
have
L(η1(Ge)) = αA
L(η1(Gf )) = αA
k ηA(L/G)(Ge) = αA
k b
l ηA(L/G)(Gf ) = αA
l b.
(7.2)
This implies that GL(η1(Ge)) = Ga = GL(η1(Gf )) and so L(η1(Ge)) = L(η1(Gf )) since
s(η1(Ge)), s(η1(Gf )) ∈ T . From Equations (7.1) and (7.2) we have αA
l b and so
k = l since the G action on A is free. Therefore d is label consistent.
To see that c is label consistent suppose that Ge, Gf ∈ (E/G)1 are such that (L/G)(Ge) =
(L/G)(Gf ) = Ga ∈ A/G, say. Let b = ηA(Ga) ∈ A, cη(Ge) = k ∈ G, and c(Gf ) = l ∈ G.
Then by the definition of c we have
k b = αA
(7.1)
(7.3)
r(η1(Ge)) = α0
r(η1(Gf )) = α0
kη0(r(Ge))
l η0(r(Gf )).
(7.4)
Then if we let e = α1−k(η1(Ge)) and f = α1−l(η1(Gf )) we have e, f ∈ E1 with r(e) =
η0(r(Ge)), r(f ) = η0(r(Gf )) ∈ T and GL(e) = GL(f ). Since T is a fundamental domain
−l(L(η1(Gf ))).
we have L(e) = L(f ) and hence αA
Since L(η1(Ge)) = L(η1(Gf )) we can conclude that k = l as in the previous paragraph.
(cid:3)
Therefore c is label consistent and our result is established.
Corollary 7.4. Let (E,L) be a weakly left-resolving, set-finite labeled graph. Suppose that
((E,L), G, α) is a free labeled graph action which admits a fundamental domain. Then
−k(L(η1(Ge))) = L(e) = L(f ) = αA
C∗(E,L) ×α,r G ∼= C∗(E/G,L/G) ⊗ K((cid:96)2(G)).
Proof. By Theorem 7.3 there are label consistent functions c, d : E1/G → G such that
((E,L), G, α) ∼= ((E/G ×c G, (L/G)d), G, τ ),
so we have
C∗(E,L) ×α,r G ∼= C∗(E/G ×c G, (L/G)d) ×τ,r G.
By Proposition 6.3 and Corollary 6.8 we have
C∗(E/G ×c G, (L/G)d) ×τ,r G ∼= C∗(E/G ×c G, (L/G)1) ×τ,r G
∼= C∗(E/G,L/G) ⊗ K((cid:96)2(G))
which gives the desired result.
(cid:3)
16
TERESA BATES, DAVID PASK, AND PAULETTE WILLIS
1203.3072.v1 [math.OA]
References
[1] T. Bates, T. Carlsen & D. Pask. C∗-algebras of
[2] T. Bates & D. Pask, C∗-algebras of labeled graphs, J. Operator Theory, 57 (2007), 207 -- 226.
[3] T. Bates & D. Pask C∗-algebras of labeled graphs II -- Simplicity Results, Math. Scand., 104 (2009),
[4] T. Bates, D. Pask, I. Raeburn & W. Szyma´nski. The C∗-algebras of row-finite graphs, New York J.
labelled graphs III -- K-Theory. arXiv:
249 -- 274.
Math., 6 (2000), 307 -- 324.
[5] T. Crisp. Corners of Graph Algebras, J. Operator Theory, 60 (2008), 253 -- 271.
[6] K. Deicke, D. Pask, & I. Raeburn, Coverings of directed graphs and crossed products of C∗-algebras
[7] S. Echterhoff, S. Kaliszewski, J. Quigg & I. Raeburn, A categorical approach to imprimitivity theorems
by coactions of homogenous spaces, Internat. J. Math., 14 (2003), 773 -- 789.
for C∗-algebras, Mem. Amer. Math. Soc. 180 (2006), no. 850.
Wiley International, 1987.
[8] J. Gross & T. Tucker, Topological graph theory, Series in Discrete Mathematics and Optimization,
[9] J. Jeong & S. Kim. On simple labelled graph C∗-algebras, J. Math. Anal. Appl. 386 (2012), 631 -- 640.
[10] J. Jeong, S. Kim & G. Park The structure of gauge invariant ideals of labelled graph C∗-algebras, J.
Funct. Anal., 262 (2012), 1759 -- 1780.
[11] S. Kaliszewski, J. Quigg, & I. Raeburn, Skew products and crossed products by coactions, J. Operator
Theory, 46 (2001), 411 -- 433.
[12] Y. Katayama, Takesaki's duality for a non-degenerate co-action, Math. Scand., 55 (1985), 141 -- 151.
[13] A. Kumjian & D. Pask C∗-algebras of directed graphs and group actions, Ergod. Th. & Dynam. Sys.,
19 (1999), 1503 -- 1519.
[14] A. Kumjian, D. Pask & I. Raeburn, Cuntz-Krieger algebras of directed graphs, Pacific J. Math., 184
(1998), 161 -- 174.
[15] D. Lind & B. Marcus. An Introduction to Symbolic Dynamics and Coding. Cambridge University
Press, Cambridge, 1995.
415 -- 443.
[16] D. Pask, J. Quigg & I. Raeburn, Coverings of k-graphs, J. Algebra, 289 (2005), 161 -- 191.
[17] D. Pask & I. Raeburn. On the K-Theory of Cuntz-Krieger algebras, Publ. RIMS Kyoto, 32 (1996),
[18] J. Quigg. Discrete C∗-coactions and C∗-algebraic bundles, J. Austral. Math. Soc. Ser. A, 60 (1996),
[19] I. Raeburn. Crossed products of C∗-algebras by coactions of locally compact groups, Operator Algebras
204 -- 221.
and Quantum Field Theory (Rome 1996), International Press, 1997, pages 74-84.
[20] I. Raeburn, Graph Algebras, CBMS Regional Conference Series in Mathematics, 103, Amer. Math.
[21] D. Robertson & W. Szyma´nski. C∗-algebras associated to C∗-correspondences and applications to
Soc., (2005).
mirror quantum spheres. arXiv: 1001.2982.v1 [math.OA]
Teresa Bates, School of Mathematics and Statistics,, The University of NSW, UNSW
Sydney 2052, Australia
E-mail address: [email protected]
David Pask, School of Mathematics and Applied Statistics,, University of Wollon-
gong,, NSW 2522, Australia
E-mail address: [email protected]
Paulette Willis, Department of Mathematics 651 PGH,, University of Houston,, Hous-
ton, TX 77204-3008
E-mail address: [email protected]
|
1108.2184 | 1 | 1108 | 2011-08-10T13:59:05 | Spectral geometry of the Moyal plane with harmonic propagation | [
"math.OA",
"hep-th"
] | We construct a `non-unital spectral triple of finite volume' out of the Moyal product and a differential square root of the harmonic oscillator Hamiltonian. We find that the spectral dimension of this triple is d but the KO-dimension is 2d. We add another Connes-Lott copy and compute the spectral action of the corresponding U(1)-Yang-Mills-Higgs model. We find that the `covariant coordinate' involving the gauge field combines with the Higgs field to a unified potential, yielding a deep unification of discrete and continuous parts of the geometry. | math.OA | math |
Spectral geometry of the Moyal plane with
harmonic propagation
Victor Gayral1 and Raimar Wulkenhaar2
1Laboratoire de Math´ematiques de l'Universit´e de Reims and Laboratoire de Math´ematiques
et Applications de l'Universit´e de Metz, France
2Mathematisches Institut der Westfalischen Wilhelms-Universitat, Munster, Germany
Abstract
We construct a 'non-unital spectral triple of finite volume' out of the Moyal
product and a differential square root of the harmonic oscillator Hamiltonian.
We find that the spectral dimension of this triple is d but the KO-dimension
is 2d. We add another Connes-Lott copy and compute the spectral action of
the corresponding U(1)-Yang-Mills-Higgs model. We find that the 'covariant
coordinate' involving the gauge field combines with the Higgs field to a unified
potential, yielding a deep unification of discrete and continuous parts of the
geometry.
1
Introduction
Unlike the compact (unital) case [7] and until now, there is for complete non-compact
Riemannian spin manifolds no proper reconstruction theorem from a spectral point of view.
Thus the question of the defining 'axioms' for non-unital spectral triples is not yet fully
answered. However, the basic and most important ideas of modifications for the locally
compact case are clear and appeared already in Connes' founding paper [5]. The case of the
[email protected]
[email protected]
1
ordinary Dirac operator of a locally compact complete Riemannian spin manifold manifests
that one cannot assume the resolvent of the Dirac type operator underlying a locally compact
(non-unital) spectral triple to be a compact operator. The natural replacement is to ask that
the 'localized resolvent', i.e. the resolvent multiplied with an element of the algebra, is a
compact operator. Another issue, explained in depth in [12], is the choice of a unitization of
the algebra. This choice is constrained by the orientability condition, which in the unital case
(and with an integral metric dimension) is the question of the existence of an Hochschild cycle
defining a volume form through the noncommutative integral given by the Dixmier trace.
Again, commutative but locally compact examples show that this has to be a Hochschild
cycle on a specific unitization of the algebra we start with, but not on the algebra itself.
With these two main modifications (compactness of the localized resolvent and existence
of a preferred unitization), most of the conditions for a non-unital spectral triple are easy
to spell out. Only the Poincar´e duality remains unclear to formulate. To help the reader
with this discussion, in the appendix we have reproduced the modified conditions for non-
unital spectral triples, as in given in [12] with the only modification that the metric and
KO-dimensions do not have to coincide, according to the recent formulation of the standard
model [6] and the Podle´s quantum sphere [9]. Note that these conditions are not far away
from those given in [26]. However, in [12, 26] there is an extra assumption of existence of a
system of local (or quasi-local) units, akin to the local structure of a non-compact manifold.
These locality assumptions have been fully removed in a more recent joint work of one of
us [2]. However, in that work, the focus is on the index theoretical side of the notion of a
spectral triple, not on the noncommutative generalization of a spin manifold. The definition
for a non-unital spectral triple given in [2] is the minimal one, ensuring a well-posed Fredholm
index problem with a numerical index computable by means of a local representative of the
Chern character in cyclic cohomology.
The present article is devoted to the study of a situation somehow in between the compact
(unital) and non-compact (non-unital) setting. Indeed, our Dirac-type operator has compact
resolvent alone, but it does not reflect the metric dimension. It is only the localized resolvent
which exhibits the correct metric dimension. We term this weird situation as 'non-unital
spectral triple of finite volume'. This has, at least, one very nice feature, namely that the
spectral action can be defined and computed in the usual way. The main motivation for this
example comes from noncommutative quantum field theory.
Because of easy computability, quantum field theory on the Moyal plane is the most-
studied toy model for noncommutative quantum field theories. The ultra-violet/infra-red
mixing problems arising in these models have been solved by one of us in [18, 19] by the
introduction of a modified propagator associated with the harmonic oscillator Hamiltonian.
See also [23, 24, 29] for different renormalization proofs. From a physics point of view, the
most fascinating property of this model is the behavior of its β-function [10, 11, 17, 20],
which makes it a candidate for non-perturbatively renormalizable quantum field theory in
dimension four [22,25,28]. We recommend [27] for review and introduction to the literature.
In [21], one of us has sketched a possible spectral triple for Moyal space with harmonic
oscillator potential. However, it became clear very soon that working out the mathematical
2
details is a non-trivial issue so that the simpler commutative case was studied first [31]. In
this paper, we achieve the construction of a spectral triple for a suitable algebra of functions
on Rd endowed with the Moyal product, together with a Dirac operator which is a square root
of the d-dimensional harmonic oscillator Hamiltonian. Few remarks are in order. Firstly, in
the same way as to find a differential (and not pseudo-differential) square root of the ordinary
Laplacian on Rd where one has to go 2⌊d/2⌋ × 2⌊d/2⌋ matrices, to find a differential square root
of the d-dimensional harmonic oscillator Hamiltonian one has to go 2d × 2d matrices-this is
the main observation in [21]. The second important remark has to do with the choice of the
function algebra with Moyal product. Indeed, there are many non-unital Fr´echet algebras
of functions with Moyal product that one may use while respecting most of the non-unital
spectral triple conditions. But there is only one for which the finiteness axiom is satisfied for
a Dirac-type operator given by a square root of the harmonic oscillator Hamiltonian, namely
the algebra of Schwartz functions S(Rd). A similar phenomenon appeared in [12] where it
has been shown that with the ordinary Dirac operator of Rd, there is only one choice of
algebra of functions with Moyal product for which the finiteness axiom is satisfied, namely
the L2-Sobolev space W 2,∞(Rd). Lastly, the construction of a Hochschild cycle satisfying
the orientability axiom requires (see also [31]) two different differential square roots of the
harmonic oscillator Hamiltonian, not only one.
The paper is organized as follow. In Section 2, we introduce two spectral triples with
common algebra A⋆ given by the set of Schwartz functions S(Rd) with Moyal product,
and two different differential square roots of the harmonic oscillator Hamiltonian acting
densely on H := L2(Rd) ⊗ C2d. The rest of the section is then devoted to prove that these
spectral triples are regular, that the metric dimension is d, the KO-dimension is 2d and that
the dimension spectrum Sd is d − N.
In section 3, we specialize to the case d = 4 and
after having proven a heat-kernel expansion result adapted to our particular situation, we
explicitly compute the spectral action for a U(1)-Higgs model.
2 The harmonic oscillator spectral triple for Moyal space
We consider two Moyal-type deformations (A⋆,D•,H), • = 1, 2, of the (commutative) d-
dimensional harmonic oscillator spectral triple introduced in [31]. In order to implement the
Moyal product, the dimension d must be even.
2.1 An isospectral deformation
On L2(Rd), we introduce the (unbounded) bosonic creation and annihilation operators
aµ := ∂µ + Ω xµ ,
µ = 1, . . . , d ,
µ := −∂µ + Ω xµ ,
a∗
µ, a∗
satisfying the commutation relations [aµ, aν] = [a∗
ν] = 0 and [aµ, a∗
ν] = 2 Ω δµν. Here,
Ω > 0 is a frequency parameter. On the exterior algebra V(Cd), we introduce fermionic
∗} = 0 and
∗ which fulfill the anticommutation relations {bµ, bν} = {bµ
partners bµ, bµ
∗, bν
3
{bµ, bν
∗} = δµν. Then, on the Hilbert space
these operators give rise to two selfadjoint operators
H := L2(Rd) ⊗^(Cd) ≃ L2(Rd) ⊗ C2d
1 , D2 := iQ2 − iQ∗
D1 := Q1 + Q∗
2 ,
,
(1)
constructed out of the supercharges
Q1 := aµ ⊗ bµ∗ ,
Q2 := aµ ⊗ bµ ,
where Einstein's summation convention is used. Indices are raised or lowered by the Eu-
clidean metric δµν or δµν, respectively. The (anti-)commutation relations imply for • ∈ {1, 2}
(2)
• = H ⊗ 1 − (−1)• Ω ⊗ Σ , H := 1
D2
µ, bµ] .
2{aµ, a∗
µ} = −∂µ∂µ + Ω2 xµxµ , Σ := [b∗
We identify H as the Hamiltonian of the d-dimensional harmonic oscillator with frequency Ω.
Its spectrum is {λn = Ω( d
2 + n) , n ∈ N}, where the eigenvalue λn appears with multiplicity
(cid:0)n+d−1
d−1 (cid:1). It then follows that (D• + 1)−z, • = 1, 2, is trace-class for ℜ(z) > 2d.
Remark 1. Our choice of D2 differs from [31]. One should take D2 from (1) also for the
commutative case to view our spectral triple as isospectral deformation. As seen in the next,
the choice (1) is required by the orientability axiom. The commutative version is somehow
degenerate and does not detect the sign of the frequency in (2).
We now wish to implement the Moyal product ⋆ in this picture:
f ⋆ g(x) =ZRd×Rd
dy dk
(2π)d f (x+ 1
2Θ · k) g(x+y) eihk,yi ,
(3)
parametrized by an invertible skew-symmetric matrix Θt = −Θ ∈ Md(R). We first need to
find out which algebra A⋆ of functions (or distribution) with Moyal product to use. For that
aim, observe that the finiteness condition alone dictates the choice of the topological vector
space underlying the algebra A⋆. Indeed, from (2) we conclude for • = 1, 2
) ,
H∞ := \m≥0
dom(Dn
• ) = S(Rd) ⊗^(Cd) ≃ S(Rd, C2d
which is required to be a finitely generated projective module over the algebra of the spectral
triple. Thus this naturally leads us to the choice A⋆ :=(cid:0)S(Rd), ⋆(cid:1), with even d.
Remark 2. For the ordinary Dirac operator on the trivial spin bundle of Rd, the set of
smooth spinors is isomorphic to W 2,∞(Rd) ⊗ C2⌊d/2⌋.
In this case, the topological vector
space underlying the choice of the algebra is the L2-Sobolev space W 2,∞(Rd). Note that the
latter is stable under the Moyal product too, and that this is the choice made in [12].
4
Here are the main properties of the Moyal product we will use latter on (for more infor-
mation see [12]). First is strong closedness
Z f ⋆ g(x) dx =Z f (x) g(x) dx =Z g ⋆ f (x) dx ,
then, we have the Leibniz rule
∀f, g ∈ L2(R4) ,
and the following identities
∂µ(f ⋆ g) = ∂µf ⋆ g + f ⋆ ∂µg ,
{f, xµ}⋆ := xµ ⋆ f + f ⋆ xµ = 2xµf,
[xµ, f ]⋆ := xµ ⋆ f − f ⋆ xµ = iΘµν∂νf ,
both holding for f, g ∈ A⋆. Last is the (non-unique) factorization property [15, p. 877]
(4)
(5)
(6)
∀f ∈ A⋆ ,
∃ g, h ∈ A⋆ : f = g ⋆ h .
(7)
Following [12], we then specify the preferred unitization B⋆ of A⋆, as the space of smooth
bounded functions on Rd with all partial derivatives bounded. The Moyal product (3) ex-
tends to B⋆, and A⋆ ⊂ B⋆ is an essential two-sided ideal, [12, Theorem 2.21], but is not
dense. The reason why we chose this particular unitization is that B⋆ contains the plane
waves and constant functions (but no other non-constant polynomials) and this is crucial
for the orientability condition (see subsection 2.3). According to [12, Theorem 2.21], the
C ∗-completion of B⋆ is
A⋆ :=(cid:8)T ∈ S ′(Rd) : T ⋆ f ∈ L2(Rd) for all f ∈ L2(Rd)(cid:9) .
Therefore, A⋆ acts on H by componentwise left Moyal multiplication, that we denote by L⋆:
L⋆ : A⋆ × H → H ,
(f, ψ ⊗ m) 7→ (f ⋆ ψ) ⊗ m ,
for ψ ∈ L2(Rd) and m ∈V(Cd). In particular, we have the bounds [12]:
kL⋆(f )k ≤ C1(Θ)kfk2,
kL⋆(f )k ≤ C2(Θ)
f ∈ A⋆ ,
α≤d+1k∂αfk∞,
sup
f ∈ B⋆ .
We also define the (anti-)action R⋆ of A⋆ on H by componentwise right Moyal multiplication:
R⋆ : A⋆ × H → H ,
(f, ψ ⊗ m) 7→ (ψ ⋆ f ) ⊗ m .
Since the complex conjugation is an involution of the algebra A⋆, and from the traciality of
the Moyal product (4), we get L⋆(f )∗ = L⋆( ¯f ), R⋆(f )∗ = R⋆( ¯f ). Moreover, this also shows
that the two representations L⋆ and R⋆ are isometric:
kL⋆(f )k = kR⋆(f )k ,
∀f ∈ A⋆ .
To avoid too many notations, L⋆, R⋆ will also denote the left and right actions of A⋆ and B⋆
on L2(Rd).
We now check that our spectral triple (A⋆,H,D•), • = 1, 2, defines a non-unital spectral
triple with spectral dimension d and KO-dimension 2d, in the sense of Definition 25 in the
Appendix.
5
2.2 Boundedness and compactness
From (6), we obtain for f ∈ B⋆ on dom(D•):
[D1, L⋆(f )] = L⋆(i∂µf ) ⊗ Γµ ,
[D2, L⋆(f )] = L⋆(i∂µf ) ⊗ Γµ+d ,
Γµ := (ibµ − ib∗µ) − 1
Γµ+d := (bµ + b∗µ) − 1
Ω Θµν (bν + b∗
Ω Θµν (ibν − ib∗
2
2
ν) ,
ν) .
(8)
As ∂µf ∈ B⋆, the commutator [D•, L⋆(f )] extends to a bounded operator. It is a remarkable
property of the Moyal algebra that just the d-dimensional differential of f appears, no x-
multiplication.
For the compactness condition, there is not much to say as (D• + λ)−1 is already a
compact operator on H. Then, L⋆(f )(D• + λ)−1 is compact for any f ∈ A⋆, even for f ∈ B⋆.
2.3 Orientability
Note first that the operators Γµ, Γµ+d defined by (8) satisfy the anticommutation relations
{Γµ, Γν} = {Γµ+d, Γν+d} = 2 (g−1)µν ,
where the symmetric matrix g ∈ GL(d, R) is defined by
{Γµ, Γν+d} = 0 ,
4
g :=(cid:0)Idd − 1
Ω2 Θ2(cid:1)−1 ,
Ω2kΘk2)−1Idd ≤ g ≤ Idd .
(1 + 1
4
and plays the role of a effective metric. Note that Θ2 = −ΘtΘ is negative definite so that
(9)
We will frequently use that Θ, g, g−1 commute with each other. Raising and lowering of
summation indices will always be performed with the Euclidean metric δµν, δµν.
Thus the {Γ1, . . . , Γ2d} generate a Clifford algebra of double dimension 2d. The inverse
transformation of (8) reads
ibν − ib∗
ν = gνµ(cid:0)Γµ + 1
2
Ω Θµρ Γρ+d(cid:1) ,
bν + b∗
ν = gνµ(cid:0)Γµ+d + 1
2
Ω Θµρ Γρ(cid:1) .
Therefore, we can express D1, in terms of {Γ1, . . . , Γ2d}, but not in terms of the half set of
operators {Γ1, . . . , Γd} produced by the commutator of D1 with B⋆. Similar comments apply
for D2. In conclusion, we get
D1 = i∂ν ⊗ gνµ(cid:0)Γµ + 1
D2 = i∂ν ⊗ gνµ(cid:0)Γµ+d + 1
Ω Θµρ Γρ+d(cid:1) + Ω xν ⊗ gνµ(cid:0)Γµ+d + 1
Ω Θµρ Γρ(cid:1) + Ω xν ⊗ gνµ(cid:0)Γµ + 1
Ω Θµρ Γρ(cid:1) ,
Ω Θµρ Γρ+d(cid:1) .
2
2
2
2
General results for Clifford algebras then show that any element of the Clifford algebra which
anticommutes with every Γµ and Γµ+d is a multiple of the anti-symmetrized product of the
generators {Γ1, . . . , Γ2d}. Therefore, a grading operator commutating with D1 cannot be
found in the algebra generated by L⋆(f ), R⋆(f ) and [D1, L⋆(f )], so that an implementation
of the orientability axiom requires both Dirac-type operators D1,D2.
(10)
6
Let uµ := e−ixµ ∈ B⋆. We know from [12] that the element
i
ǫ(σ)
c := Xσ∈Sd
d(d−1)
2 √det g
d!
(cid:0)(u1 ⋆ · · · ⋆ ud)−1 ⊗ 1(cid:1) ⊗ uσ(1) ⊗ · · · uσ(d) ,
is a Hochschild d-cycle for the algebra B⋆, with values in B⋆⊗Bo
⋆. (In the expression of c, the
inverse is with respect to the ⋆-product and the scaling by √det g is irrelevant for cyclicity.)
For πD• defined in the Appendix, we then obtain from (8)
γ1 := πD1(c) =
γ2 := πD2(c) =
√det g
d!
√det g
d!
i
i
d(d−1)
2 Xσ∈Sd
2 Xσ∈Sd
d(d−1)
ǫ(σ) ⊗ Γσ(1) · · · Γσ(d) ,
ǫ(σ) ⊗ Γσ(1)+d · · · Γσ(d)+d ,
(11)
and they satisfy the relations:
2
1 = 1 = γ
2
2 ,
γ
∗
1 = γ1 , γ
∗
2 = γ2 ,
γ
γ1γ2 = (−1)d
γ2γ1 .
Thus we define
Since γ1, γ2 commute with every element of A⋆ or B⋆, Γ does too and the discussion above
shows that
Γ := (−i)d
γ1γ2 .
(12)
Γ2 = 1 ,
{D•, Γ} = 0 ,
• = 1, 2 ,
so that Γ defines the grading operator for the two spectral triples (A⋆,H,D•), • = 1, 2. We
stress that the necessity of the two Dirac operators D1,D2, is quite different from conventional
spectral triples [7] where a single operator is needed.
Note also that from the explicit formulae of D•, (up to a possible sign) one has the
relation Γ = 1 ⊗ (−1)Nf in terms of the fermionic number operator Nf = b∗
2.4 KO-dimension and other algebraic conditions
The real structure is an anti-linear isometry J on H. We assume that for d even the KO-
dimension k is even, too. Then, according to the sign table in the Appendix we have
µbµ.
JD• = D•J ,
• = 1, 2 .
This is achieved by the following non-trivial action on the matrix part of H:
JaµJ −1 = aµ ,
Ja∗
µJ −1 = a∗
µ ,
JbµJ −1 = b∗
µ ,
Jb∗
µJ −1 = bµ .
(13)
V(Cd) as generated by repeated action of {b†
In particular, conjugation by J preserves the (anti-)commutation relations. We can view
µ} on the vacuum vector 0i defined by bµ0i = 0.
7
It then follows that, up to a prefactor of modulus 1, which cancels in every relation of the
dimension table, J is the Hodge-∗ operator onV(Cd), i.e. is uniquely defined by
2 · · · b∗
1b∗
J0i = b∗
d0i ,
together with (13) and the anti-linearity J(zψ) = ¯zJψ. In particular, J◦L⋆(f )◦J −1 = R⋆(f ),
which implements the opposite algebra and achieves the order-one condition:
[JL⋆(f1)J −1, L⋆(f2)] = 0 ,
[JL⋆(f1)J −1, [D•, L⋆(f2)]] = 0 ,
for all f1, f2 ∈ B⋆ .
(14)
To compute J 2 we consider, for µ1 < µ2 < · · · < µk,
J(b∗
µ1 · · · b∗
µ1...µk
. . . means that b∗
1b∗
µk0i) = bµ1 · · · bµk b∗
2 · · · b∗
µ1, . . . , b∗
The notation
d0i = (−1)Pk
j=1(µj −1)b∗
1
µ1...µk
. . . b∗
d0i .
µk are missing. We apply J again, to get:
J 2(b∗
µ1 · · · b∗
µk0i) = (−1)Pk
j=1(µj −1)J(b∗
1
µ1...µk
. . . b∗
d0i) = (−1)Pd
j=1(j−1)b∗
µ1 · · · b∗
µk0i ,
which means
J 2 = (−1)
d(d−1)
2
.
(15)
From (8) it follows that J commutes with Γµ and Γµ+d. From (12) we then conclude JΓ =
(−1)dΓJ. Comparing these results with the dimension table in the Appendix, we have
proven:
Proposition 3. The spectral geometries (A⋆,H,D•, Γ, J), • = 1, 2, for the d-dimensional
Moyal algebra A⋆ are of KO-dimension 2d mod 8.
2.5 Metric dimension
Since D•, • = 1, 2, squares (up to matrices) to the d-dimensional harmonic oscillator Hamil-
tonian, we already now that (1 + D2
In this
subsection we are going to prove that for the localized operators, the critical dimension is
reduced by a factor of 2, that is for all f ∈ A⋆, the operators L⋆(f )(1 + D2
•)−d/2 belong to
L1,∞(H) and that any of its Dixmier traces is a constant multiple of the integral of f . To
obtain both Dixmier traceability and the value of the Dixmier trace, we will use the results
of [1]. In order to do this, we need some preliminary Lemmas (which will also be needed
to check the regularity condition, to obtain the dimension spectrum and to compute the
spectral action).
•)−d belongs to the Dixmier ideal L1,∞(H).
Lemma 4. Introducing the operators on L2(Rd):
∇µ := ∂µ + 1
2i Ω2 Θµν xν ,
∇µ := 1
2(cid:0)∂µ − 2i (Θ−1)µν xν(cid:1) ,
8
µ = 1,· · · , d ,
we have the following relations for f ∈ B⋆:
[H, L⋆(f )] = −L⋆(cid:0)(g−1)µν∂µ∂νf(cid:1) − 2L⋆(∂µf )∇µ ,
[∇µ, L⋆(f )] = L⋆(cid:0)(g−1)µν∂νf(cid:1) ,
[H, ∇µ] = 2i(Θ−1)µν∇ν ,
[H,∇µ] = −2i Ω2Θµν ∇ν ,
[ ∇µ, L⋆(f )] = L⋆(∂µf ) ,
[ ∇µ, ∇ν] = i(Θ−1)µν ,
[∇µ,∇ν] = −i Ω2Θµν ,
Proof. This follows from the relations (2), (5) and (6).
Corollary 5. Let Pα( ∇) be an element of order α of the polynomial algebra generated by
∇, ∇. Then Pα( ∇)(1 + H)−α/2 extends to a bounded operator.
Proof. From the operator inequalities (no summations on µ but summation on ν)
[∇µ, ∇ν] = i(g−1)µρ(Θ−1)ρν .
(1 + H)−1/2∂µ2 = −∂µ(1 + H)−1∂µ ≤ −∂µ(1 − ∂ν∂ν)−1∂µ ,
(1 + H)−1/2xµ2 = xµ(1 + H)−1xµ ≤ xµ(1 + Ω2xνxν)−1xµ ,
we see that ∇µ(1 + H)−1/2 is bounded. Then, the general case follows by induction using
∇µ1
∇µ2(1 + H)−1 = ∇µ1(1 + H)−1 ∇µ2 + ∇µ1[ ∇µ2, (1 + H)−1] ,
and
and
∇µ1[ ∇µ2, (1 + H)−1] = − ∇µ1(1 + H)−1[ ∇µ2, H](1 + H)−1 ,
which is bounded, too, according to Lemma 4.
The following Proposition will be crucial for the computation of the spectral action, the
estimate we need to evaluate the Dixmier trace and to compute the dimension spectrum and
the residues of the associated zeta functions.
Proposition 6. For f ∈ B⋆, define Tµ1...µk (f ) := Tr(cid:0)L⋆(f )∇µ1 . . .∇µk e−tH(cid:1). Then one has
Tµ1,...µk (f ) = X1≤j1<j2<···<j2a≤k(cid:16)
2π sinh(2 Ωt)(cid:17) d
Ω
2
×ZRd
dzpdet g f (z) e− Ω tanh( Ωt)hz,gzi (Zµ1
j1...j2a
. . . Zµk )(Nµj1 µj2 · · ·Nµj2a−1 µj2a
) ,
where
and
that g is a constant metric so that √det g can also be taken in front of the integral).
, . . .Zµj2a} are missing in the product Zµ1 · · ·Zµk . (Remember
Ω3 coth( Ωt)(ΘgΘ)µν − 1
2 i Ω2 Θµν ,
Zµ := − Ω tanh( Ωt)zµ + i Ω2(Θgz)µ ,
Nµν := − 1
. . . means that {Zµj1
Ω (coth( Ωt) + tanh( Ωt))(g−1)µν − 1
2
2
j1...j2a
In particular,
Tr(cid:0)e−tH(cid:1) =(cid:0)2 sinh( Ωt)(cid:1)−d
9
.
Proof. Since L⋆(f )∇µ1 . . .∇µk e−tH is trace-class (because ∇µ1 . . .∇µk e−tH/2 is bounded by
Corollary 5 and e−tH/2 is trace-class), the trace can be evaluated as the integral of the kernel
on the diagonal. Thus, in integral kernel representation, we have to compute
Tµ1,...µk (f ) =
ZRd×Rd
dx dy (L⋆(f ))(x, y)(cid:16) ∂
∂yµ1
+
i
2
Ω2Θµ1ν1yν1(cid:17)· · ·(cid:16) ∂
∂yµk
The operator kernel of e−tH is the Mehler kernel
+
i
2
Ω2Θµkνkyνk(cid:17)(cid:0)e−tH (y, x)(cid:1) .
while the operator kernel of L⋆(f ) is readily identified to be
e−tH (x, y) =(cid:16)
L⋆(f )(x, y) =
Ω
e−
Ω
4 tanh( Ωt)kx+yk2
Ω
4 coth( Ωt)kx−yk2−
2π sinh(2 Ωt)(cid:17)d/2
πd det ΘZ dz f (z) eihx−y,Θ−1(x+y)i+2ihz,Θ−1(x−y)i .
1
,
(16)
(17)
We introduce u = x − y and v = x + y and
Ω
coth( Ωt)uµ −
2
Ω
(coth( Ωt) + tanh( Ωt))δµν −
2
Yµν := −
Dµ(u, v) :=
tanh( Ωt)vµ +
Ω
2
i
4
i
2
Ω2Θµα(vα − uα) ,
Ω2Θµν ,
to obtain
Ω
2π sinh(2 Ωt)(cid:17) d
2
j1...j2a
1
(2π)d det Θ
Ω
2π sinh(2 Ωt)(cid:17) d
2
1
det Θ√det Q
Tµ1,...µk (f )
= X1≤j1<j2<···<j2a≤k(cid:16)
×Z du dv dz f (z)Dµ1(u, v)
= X1≤j1<j2<···<j2a≤k(cid:16)
×ZRd
where E := e− 1
dz f (z)Dµ1( i∂
∂ξ , i∂
∂η )
. . . Dµk (u, v)Yµj1 µj2 · · ·Yµj2a−1 µj2a
e− 1
2 h(u,v),Q(u,v)i−h(u,v),(2iΘ−1 z,0)i
∂ξ , i∂
j1...j2a
. . . Dµk ( i∂
∂η )Yµj1 µj2 · · ·Yµj2a−1 µj2aE(cid:12)(cid:12)(cid:12)ξ=η=0
2 h(−2zΘ−1+ξ,η),Q−1(2Θ−1z+ξ,η)i and Q ∈ M2d(C) is given by
2 tanh( Ωt) Idd! .
Q = Ω
2 coth( Ωt) Idd
−iΘ−1
iΘ−1
Ω
,
10
Recalling that g−1 = 1 − Ω2
matrices
4 Θ2 and gΘ = Θg, we find by writing Q as product of triangle
so that
1
det Q =
2 tanh( Ωt) g Θ2
det g (det Θ)2 ,
Q−1 = − Ω
2 coth( Ωt) g Θ2! ,
− Ω
E = expn − Ω tanh( Ωt)hz, gzi − Ω tanh( Ωt)hz, gΘξi − 2ihz, gηi
−ig Θ
ig Θ
Ω
4
+
tanh( Ωt)hξ, ΘgΘξi + ihξ, Θgηi +
∂η )E =(cid:16)i Ω2(Θgz)µ − Ω tanh( Ωt)zµ + i
Ω2
2
Dµ( i∂
∂ξ , i∂
Ω
4
coth( Ωt)hη, ΘgΘηio ,
(ΘgΘξ)µ
− Ω coth( Ωt)(gΘη)µ +
coth( Ωt)(Θη)µ −
Ω
2
Ω
2
tanh( Ωt)(Θξ)µ(cid:17)E .
Then, the functions
Zµ := E −1Dµ( i∂
∂ξ , i∂
Nµν := Yµν + Dµ( i∂
,
∂η )E(cid:12)(cid:12)(cid:12)ξ=η=0
∂η )(cid:0)E −1Dν( i∂
∂ξ , i∂
∂ξ , i∂
∂η )E(cid:1) ,
take the values given in the Lemma, and the assertion follows.
A very nice feature of the results of [1] is that both the questions of the Dixmier trace-
ability and of the value of the Dixmier trace of an operator of the form aGk are reduced
to the value of the Hilbert-Schmidt norm of the heat-type operator ae−tG−1. In our context
a = L⋆(f ), G = (1 + D2
•)−1, and all we need to do is to evaluate the Hilbert-Schmidt norm
of L⋆(f )e−tD2
•.
Lemma 7. If f ∈ A⋆, then we have:
Ωd/2
πd/2 tanhd/2(2 Ωt)Z dzpdet g ¯f ⋆ f (z)e− Ω tanh(2 Ωt)hz,gzi .
kL⋆(f )e−tD2
•k2
2 =
Proof. Since D2
• = H ⊗ 1 − (−1)• Ω ⊗ Σ, we have
and thus
For the matrix trace, we have
• = e−tH ⊗ e(−1)•t ΩΣ ,
0 ≤ e−tD2
•k2 = kL⋆(f )e−tHk2 tr(cid:0)e(−1)•2t ΩΣ(cid:1)1/2 .
kL⋆(f )e−tD2
tr(e(−1)•t ΩΣ) = tr(e(−1)•t Ω Pd
µ=1(b∗
µbµ−bµb∗
µ)) = tr(cid:16) dYµ=1
e(−1)•t Ω(b∗
µbµ−bµb∗
µ)(cid:17) .
11
In the basis s1, . . . , sdi := (b∗
1)s1 · · · (b∗
d)sd0,· · · , 0i of C2d, with si ∈ {0, 1}, we have
−(b∗
µbµ − bµb∗
µ)s1, . . . , sdi = (−1)sµs1, . . . , sdi ,
and therefore, for both • = 1, 2,
tr(e(−1)•t ΩΣ) = 2d coshd( Ωt) .
2 = Tr(cid:0)e−tH L⋆( ¯f ⋆ f )e−tH(cid:1) = Tr(cid:0)L⋆( ¯f ⋆ f )e−2tH(cid:1), has been
The other bit, kL⋆(f )e−tHk2
computed in Proposition 6.
Remark 8. Since for f ∈ A⋆, ¯f ⋆f is a priori not a positive function, in the previous Lemma,
one may wonder whyR dz √det g ¯f ⋆ f (z) exp{− Ω tanh(2 Ωt)hz, gzi} is positive, as it should
be. This follows from the following facts: For A a positive definite matrix commuting with
Θ, set gA(x) := e−<x,Ax>. Then a computation gives
gA ⋆ gA = (det(1 + ΘtA2Θ))−1/2 gB , with B =
2A
1 + ΘtA2Θ
.
It follows
for
exp{− Ω tanh(2 Ωt)hz, gzi} = (det(1 + ΘtA2Θ))1/2 gA ⋆ gA ,
A =
g−1 −(cid:0)g−2 − Ω2 tanh(2 Ωt)2ΘtΘ(cid:1)1/2
Ω tanh(2 Ωt) ΘtΘ
.
Note that g−2 − Ω2 ΘtΘ = (1 − Ω2 ΘtΘ/4)2 so that A exists for all t. Using the traciality
of the Moyal product (4), we then get for the matrix A given above and up to a positive
constant:
Z dz ¯f ⋆ f (z) exp{− Ω tanh(2 Ωt)hz, g−1zi} = C(Θ, Ω, t)Z dz ¯f ⋆ f (z) gA ⋆ gA(z)
= C(Θ, Ω, t)Z dz ¯f ⋆ f ⋆ gA ⋆ gA(z)
= C(Θ, Ω, t)Z dz f ⋆ gA ⋆ f ⋆ gA(z)
= C(Θ, Ω, t)Z dz f ⋆ gA(z) f ⋆ gA(z)
= C(Θ, Ω, t)kf ⋆ gAk2 ≥ 0 .
• is Hilbert-Schmidt also for f in B⋆, since A⋆ is an ideal
Moreover, it explains why L⋆(f )e−tD2
of B⋆ and A⋆ ⊂ L2(Rd).
Lemma 9. For all t > 0 and • = 1, 2, we have
Tr(cid:0)e−tD2
•(cid:1) = cothd( Ωt) .
12
Proof. This is a corollary of Lemma 7 and the Remark which follows it, by letting f going
to the constant unit function.
Lemma 10. If f ∈ A⋆ and t > 0, then we have the bound
where the constant depends only on Ω and Θ.
kL⋆(f )e−tD2
•k2 ≤ C k ¯f ⋆ fk1/2
1 max(1, t−d/4) ,
Proof. This is a direct consequence of Lemma 7.
Lemma 11. There is C ′ > 0 such that for any f1, f2 ∈ A⋆ and t > 0 one has
•/4]k1 .
kL⋆(f1)[L⋆(f2), e−tD2
kf1k2 kL⋆(∂µf )e−tD2
• is trace class for t > 0. We use the identity
Proof. By Lemma 9, e−tD2
[eA, B] =Z 1
0
to get
ds
L⋆(f1)[L⋆(f2), e−tD2
Hence we have
kL⋆(f1)[L⋆(f2), e−tD2
•]k1 ≤ tkL⋆(f1)kZ 1
0
d
•]k1 ≤ C ′t1/2
dXµ=1
ds(cid:0)esABe(1−s)A(cid:1) =Z 1
•] = −tL⋆(f1)Z 1
ds(cid:16)ke−tsD2
+ ke−tsD2
0
0
•[D•, L⋆(f2)]e−t(1−s)D2
•k = (2et)−1/2. Thus, using the relation
By spectral theory, kD•e−tD2
[D•, L⋆(f2)] =(iL⋆(∂µf2) ⊗ Γµ ,
iL⋆(∂µf2) ⊗ Γµ+4 ,
we get with kL⋆(f1)k ≤ Ckf1k2 and C ′ := C√2πe−1 sup2d
• = 1 ,
• = 2 ,
µ=1 kΓµk
ds esA[A, B]e(1−s)A ,
(18)
ds e−tsD2
•[D2
•, L⋆(f2)]e−t(1−s)D2
• .
•/2D•kke−tsD2
•/2[D•, L⋆(f2)]e−t(1−s)D2
•k1
•/2k(cid:17) .
•/2k1kD•e−t(1−s)D2
kL⋆(f1)[L⋆(f2), e−tD2
t1/2kf1k2
C ′
π
≤
•]k1
dXµ=1Z 1
0
Estimating
ds s−1/2(1 − s)−1/2ke−tsD2
•/2L⋆(∂µf2)e−t(1−s)D2
•/2k1 .
ke−tsD2
•/2L⋆(∂µf2)e−t(1−s)D2
•/2k1 ≤(kL⋆(∂µf2)e−tD2
•/4k1,
•/4L⋆(∂µf2)k1,
ke−tD2
if
if
s ∈ [0, 1/2] ,
s ∈ [1/2, 1] ,
the result follows.
13
Lemma 12. Let f ∈ A⋆. Then, there exists a finite constant C(f ) such that for all t > 0:
kL⋆(f )e−tD2
•k1 ≤ C(f ) max(t−d/2, td/2) .
Proof. Our strategy is to iterate a combination of the factorization property (7) with Lemma
10 and Lemma 11 far enough so that we can bound e−ǫtD2
• alone in trace-norm (i.e. without
element of the algebra of both sides).
According to (7), for all f ∈ A⋆ there exist f1, f2 ∈ A⋆ such that f = f1 ⋆ f2, giving
L⋆(f )e−tD2
• = L⋆(f1)e−tD2
From Lemma 10 and Lemma 11 we conclude
kL⋆(f )e−tD2
•k1 ≤ kL⋆(f1)e−tD2
≤ k ¯f1 ⋆ f1k1/2
•/2k2 ke−tD2
1 k ¯f2 ⋆ f2k1/2
•(cid:3) .
•L⋆(f2) + L⋆(f1)(cid:2)L⋆(f2), e−tD2
•(cid:3)k2
•/2L⋆(f2)k2 + kL⋆(f1)(cid:2)L⋆(f2), e−tD2
1 max(t−d/2, 1) + C ′t1/2
(19)
dXµ=1
kf1k2 kL⋆(∂µf2)e−tD2
•/4k1 .
Iterating d-times the estimate (19) with the repeated factorization
∂µ1f2 = f1,µ ⋆ f2,µ ,
. . . ,
∂µk+1f2,µ1...µk = f1,µ1...µk+1 ⋆ f2,µ1...µk+1 ,
with f1,µ, f2,µ, . . . , f1,µ1...µk+1 ⋆ f2,µ1...µk+1 ∈ A⋆, we get for some constants C0(f ), . . . , Cd(f )
depending on f and on the choice of factorization at each step:
Ck(f )tk/2 max(t−d/2, 1) + Cd(f )td/2
•k1 ≤
kL⋆(∂µf )e−tD2
d−1Xk=0
Using Lemma 9 we get
•/4dk1 ≤ kL⋆(f2,µ1...,µd)kke−tD2
kL⋆(f2,µ1...,µd)e−tD2
dXµ1,...,µd=1
kL⋆(f2,µ1...µd)e−tD2
•/4d
k1 .
•/4dk1 ≤ C ′′kf2,µ1,...,µdk2 max(t−d, 1) ,
which completes the proof.
Corollary 13. For any f ∈ A⋆, the operator [(D2
Proof. By factorization f = f1 ⋆ f2, f1, f2 ∈ A⋆ and Leibniz rule:
• + 1)−d/2, L⋆(f )] is of trace class.
• + 1)−d/2, L⋆(f2)] is of trace class for arbitrary f1, f2 ∈ A⋆.
• + 1)−d/2, L⋆(f 1)](cid:1)∗ ,
[(D2
• + 1)−d/2, L⋆(f )] = L⋆(f1)[(D2
it suffices to show that L⋆(f1)[(D2
By spectral theory,
• + 1)−d/2, L⋆(f2)] −(cid:0)L⋆(f 2)[(D2
Γ(d/2)Z ∞
• + 1)−d/2, L⋆(f2)]k1 ≤ C(f )Z ∞
• + 1)−d/2, L⋆(f2)] =
1
0
0
L⋆(f1)[(D2
kL⋆(f1)[(D2
As the integral converges, we are done.
14
dt td/2−1L⋆(f1)[e−t(D2
•+1), L⋆(f2)] .
dt e−t td/2−1 · t1/2 · max(t−d/2, td/2) .
Combining Lemma 11 with Lemma 12 we obtain for a finite constant depending only on f :
We have arrived at the main result of this subsection, that the spectral triple (A⋆,H,D•)
has metric dimension d and not 2d (remember that d is even).
Theorem 14. For f ∈ A⋆, • = 1, 2, the operator L⋆(f )(1 +D2
for any Dixmier trace Trω, we have
•)−d/2 belongs to L1,∞(H) and
Trω(cid:0)L⋆(f )(1 + D2
•)−d/2(cid:1) =
1
πd/2(d/2)!Z dxpdet g f (x) .
Proof. We use the factorization property (7) to write f = f1 ⋆ f2 with f1, f2 ∈ A⋆, which
gives
•)−d/2L⋆(f2) + L⋆(f1)(cid:2)L⋆(f2), (1 + D2
•)−d/2(cid:3) .
•)−d/2] is trace class, and combining Lemma 7 and
•)−d/2L⋆(f2) belongs to L1,∞(H). Therefore,
•)−d/2 is Dixmier-trace-class too, and any of its Dixmier trace coincides with
L⋆(f )(1 + D2
•)−d/2 = L⋆(f1)(1 + D2
By Corollary 13, L⋆(f1)[L⋆(f2), (1 + D2
[1, Proposition 4.8] we get that L⋆(f1)(1 + D2
L⋆(f )(1 + D2
those of L⋆(f1)(1 + D2
that
•)−d/2L⋆(f2).
Using a polarization identity, it suffices to compute Trω(cid:0)L⋆( ¯f )(1 + D2
(s − 1)Tr(cid:0)L⋆( ¯f )(1 + D2
lim
s→1+
•)−d/2L⋆(f )(cid:1). Note
•)−ds/2L⋆(f )(cid:1)
dt e−t tds/2−1Tr(cid:0)L⋆( ¯f )e−tD2
dt e−t td(s−1)/2−1
= lim
s→1+
= lim
s→1+
s − 1
Γ(ds/2)Z ∞
Γ(ds/2)Z ∞
s − 1
0
0
=
1
πd/2 lim
s→1+
(s − 1)Γ(d(s − 1)/2)
Γ(ds/2)
( Ωt)d/2
•L⋆(f )(cid:1)
πd/2 tanhd/2( Ωt)Z dzpdet g ¯f ⋆ f (z)e− Ω tanh( Ωt)hz,gzi
Z dzpdet g ¯f ⋆ f (z) .
Now [1, Proposition 5.13], which relies on Corollary 13, gives for any Dixmier trace
Trω(cid:0)L⋆( ¯f )(1 + D2
•)−d/2L⋆(f )(cid:1) =
This all what we needed to prove.
1
πd/2(d/2)!Z dxpdet g ¯f ⋆ f (x) .
(20)
2.6 Regularity and dimension spectrum
Our next task is to check the regularity condition.
Proposition 15. For any f ∈ B⋆ and • = 1, 2, both L⋆(f ) and [D•, L⋆(f )] belong to
n=1 dom δn
• , where δ•(T ) := [hD•i, T ] and hD•i := (D2
• + 1)
2 .
1
T∞
15
Proof. It is well known (see for example [2]) thatT∞
R• and L• are the unbounded linear operators given by
• =T∞
•, T ]hD•i−1, L•(T ) := hD•i−1[D2
R•(T ) := [D2
n=1 dom δn
•, T ] .
n,m=1 dom Rm
• ◦ Ln
• where
By an easy inductive argument, we see that for T ∈ B⋆ ∪ [D•,B⋆], we have
• ◦ Ln
Rm
•)(cid:1)n+m(T )hD•i−m .
• = H − (−1)• ΩΣ with Σ bounded and [H, Σ] = 0, we get
• (T ) = hD•i−n(cid:0)ad(D2
k (cid:19)(cid:0)(−1)•+1 Ω ad(Σ)(cid:1)n+m−k(cid:18)hD•i−n(cid:0)ad(H)(cid:1)k(T )hD•i−m(cid:19) . (21)
n+mXk=0(cid:18)n + m
Since D2
Rm
• ◦ Ln
• (T ) =
We treat the worst case only, which is when k = n + m (for the other values of k, one may
use similar but simpler arguments). That is, we need to show the boundedness of
hD•i−n(cid:0)ad(H)(cid:1)n+m
(T )hD•i−m, n, m ∈ N .
The preceding expression applied to T = [D1, L⋆(f )] = iΓµL⋆(∂µf ) gives
hD1i−n Γµ(cid:0)ad(H)(cid:1)n+m(L⋆(i∂µf ))hD1i−m
=(cid:0)Γµ + hD1i−n[Γµ,hD1in](cid:1)hD1i−n(cid:0)ad(H)(cid:1)n+m(L⋆(i∂µf ))hD1i−m ,
and using [Γµ,hD1in] = [Γµ, ( ΩΣ)n] it is enough to treat the case T ∈ B⋆. Similarly for D2.
Now, Lemma 4 shows that (ad(H))n+m(T ) can be written as a sum of terms of the form
∇kL⋆(f ) ∇l with f ∈ B⋆ and k (resp.
l) not exceeding n (resp. m), where ∇ is ∇ or ∇.
Then, one concludes using Corollary 5.
In the next and in analogy with the regularity condition, we will prove that one can
determine the dimension spectrum with the derivations R• and L•, instead of δ•.
for all M ∈ R, ζb is a finite sum of terms of the form Tr(cid:0)Rn1
Proposition 16. Let b belonging to the polynomial algebra generated by δn
• (A⋆) and
δn
• ([D•,A⋆]). Let also ζb(z) := Tr(bhD•i−z), defined on the open half plane ℜ(z) > 2d. Then,
nj, k, m ∈ N, bj ∈ A⋆ ∪ [D•,A⋆], plus a function holomorphic on the half plane ℜ(z) > M.
Proof. For ℜ(z) > 2d, hD•i−z is a trace class operator. Since the algebra generated by
δn
• (L⋆(f )) and δn
• ([D•, L⋆(f )]) consists by Proposition 15 of bounded operators, bhD•i−z is
trace class for ℜ(z) > 2d and any b in this polynomial algebra.
• (bk)hD•i−z−m(cid:1),
So let b ∈ A⋆ ∪ [D•,A⋆]. From the spectral representation of a positive operator A:
• (b1)· · · Rnk
A =
1
πZ ∞
0
dλ λ−1/2 A2
A2 + λ
,
16
Commuting [hD•i2, b] with (hD•i2 + λ)−1 to the left, we get after some re-arrangements and
dλ λ1/2
1
hD•i2 + λ
[hD•i2, b]
1
.
hD•i2 + λ
we get
1
δ•(b) =
πZ ∞
usingR dλ λ1/2t(t2 + λ)−2 = π/2:
0
δ•(b) = 1
2 R•(b) −
0
2
1
πZ ∞
πZ ∞
πZ ∞
2
0
0
dλ λ1/2
1
hD•i2 + λ
R2
•(b)
hD•i2
(hD•i2 + λ)2 .
This suggests to introduce the map T• : B(H) → B(H) given by
A 7→ T•(A) :=
dλ λ1/2
1
hD•i2 + λ
A
hD•i2
(hD•i2 + λ)2 .
Note that this operator is contractive. Indeed
kT•k ≤
λ1/2
(1 + λ)2 dλ = 1 .
Thus 2δ• = R• − T• ◦ R2
operators of left and right multiplications by functions of hD•i), we get
• and since T• commutes with R• (because R• commutes with the
2nδn
• =
nXk=0(cid:18)n
k(cid:19) (−1)kT k
• ◦ Rn+k
•
.
Hence, a typical element of the algebra generated by δn
sum of elements of the form
• (L⋆(f )) and δn
• ([D•, L⋆(f )]), is a finite
kYj=1
• ◦ Rmj
T nj
• (bj),
bj ∈ A⋆ ∪ [D•,A⋆], nj < mj ∈ N .
From the same reasoning as at the beginning of the proof, the function
ζR,T (b1, n1, m1;· · · ; bk, nk, mk; z) := Tr(cid:16) kYj=1
T nj
• ◦ Rmj
• (bj)hD•i−z(cid:17) ,
is holomorphic on the open half plane ℜ(z) > 2d. Starting from the definition, we have
T nj
• ◦ Rmj
• (bj) =Z[0,∞]n
dλn
kYj=1
kYj=1(cid:16) njYrj=1
(2/π)λ1/2
rj
hD•i2 + λrj(cid:17) Rmj
• (bj)(cid:16) njYsj=1
hD•i2
(hD•i2 + λsj )2(cid:17) .
The next step consists in commuting for each j the Rmj
• (bj) to the left of (hD•i2 + λrj )−1:
h
1
hD•i2 + λj
, Rmj
• (bj)i = −
Rmj +1
•
(bj)
hD•i
hD•i2 + λj
1
hD•i2 + λj
17
=
•
NjXpj=1
(−1)pj Rmj +pj
(−1)Nj +1
hD•i2 + λj
•
+
Rmj +Nj+1
(bj)
hD•ipj
(hD•i2 + λj)pj+1
hD•i
(bj)(cid:16)
hD•i2 + λj(cid:17)Nj+1
.
(22)
Any of the resulting λ-integrals is convergent, and Rn
• (b) is bounded for all n ∈ N. Choosing
the Nj large enough, we generate as much negative powers of hD•i as necessary to make the
product of the remainder with hD•i−z a trace-class operator for any given z with ℜ(z) > M
(if M gets more and more negative we need larger and larger Nj). The other terms integrate
to
j
(2/π)λ1/2
hD•ipj+2
(hD•i2 + λrj )pj+3 =
Γ( 3
2 + pj)hD•i−1−pj
√πΓ(3 + pj)
,
• (bj) is a finite linear combination of
dλj
Z[0,∞]
j=1 T nj
(b1)hD•i−n1−q1Rm2+q2
• Rmj
•
plane ℜ(z) > M,Qk
Rm1+q1
•
so that, up to the remainder term, which is easily seen to be holomorphic on the open half
(b2)hD•i−n2−q2 · · · Rmk+qk
•
(bk)hD•i−nk−qk .
The final step consists in commuting the hD•i−nj−qj to the right. If nj + qj = 2lj is even, we
use the (λj = 0)-case of (22). If nj + qj = 2lj − 1 is odd,
j+1
j+1
m′
•
m′
•
(cid:2)hD•i1−2lj , R
plane ℜ(z) > M,Qk
(bj+1)(cid:3) = hD•i−2lj δ•(R
Using δ• = 1
2(R• − T• ◦ R2
•), this case is reduced to the first one. Eventually, we conclude
that, up to a remainder term, which again is easily seen to be holomorphic on the open half
• ◦ Rmj
j=1 T nj
Rm′
(bj+1)) +(cid:2)hD•i−2lj , R
(bj+1)(cid:3)hD•i .
• (bj) is a finite linear combination of
• (b1)Rm′
m′
•
j+1
1
2(b2)· · · Rmk(bk)hD•i−m .
This concludes the proof.
We can now state the main result of this section, namely:
Theorem 17. For • = 1, 2, the spectral triple (A⋆,H,D•) has dimension spectrum Sd =
d − N. Moreover, all poles of ζb(z) at z ∈ Sd are simple with local residues, i.e. for b =
• L⋆(f1)· · · δnk
δn1
• L⋆(fk), any residue resz∈Sdζb(z) is a finite sum of terms of the form
ZRd
dx xα0 ⋆ (∂α1f1) ⋆ · · · ⋆ (∂αk fk) ,
where αi ∈ Nd. An analogous result holds when the L⋆(fk)'s in b are replaced by [D•, L⋆(fk)]'s.
Proof. According to Proposition 16, it is equivalent to consider the functions
Tr(cid:0)Rm1
• (b1)· · · Rmk
• (bk)hD•i−z(cid:1),
18
bi ∈ A⋆ ∪ [D•,A⋆] ,
instead of ζb(z). These functions are well defined for ℜ(z) > 2d, since Rm
Since [H, Σ] = 0, we get from (21)
• (b) is bounded.
Rmj
• (bj) =
Since bj is either L⋆(fj) or iΓµL⋆(∂µfj) for • = 1 and iΓµ+dL⋆(∂µfj) for • = 2, we get
k (cid:19)(cid:0)ad(H)(cid:1)k(cid:18)(cid:0)ad( ΩΣ)(cid:1)mj −k(bj)(cid:19)hD•i−mj .
δk,mj L⋆(fj)
mjXk=0(cid:18)mj
ad( ΩΣ)(cid:1)mj −k(bj) =
1)(cid:1) M1 hD•i−m1 · · ·(cid:0)ad(H)(cid:1)nk(cid:0)L⋆(f ′
ad( ΩΣ)(cid:1)mj −kiΓµL⋆(∂µfj)
ad( ΩΣ)(cid:1)mj −kiΓµ+dL⋆(∂µfj) .
• (b1)· · · Rmk
We may therefore assume that bj = MjL⋆(fj) with fj ∈ A and Mj ∈ Mat2d(C). Thus, the
product Rm1
• (bk) can be expressed as a finite sum of terms of the form
Using the table given in Lemma 4 we can express Rm1
of terms
(cid:0)ad(H)(cid:1)n1(cid:0)L⋆(f ′
L⋆(∂α1f1) M1 Pα1( ∇)hD•i−m1 · · · L⋆(∂αk fk) Mk Pαk( ∇)hD•i−mkhD•i−z
k)(cid:1) Mk hD•i−mk, nj ≤ mj .
• (b1)· · · Rmk
• (bk)hD•i−z as a finite sum
=
Γ( m1
1
)Γ( mk+z
2 )· · · Γ( mk−1
× L⋆(∂α1f1) M1 Pα1( ∇) e−t1(D2
2
2
)Z dt1 · · · dtk t
1
m1
2 −1
mk +z
2 −1
· · · t
k
•+1) ,
where Pαj ( ∇) is a polynomial in ∇, ∇ of degree αj ≤ mj (the αj are multi-indices).
•+1) · · · L⋆(∂αk fk) Mk Pαk( ∇) e−tk(D2
Using
[e−tj (D2
•+1), T ] = −tjZ 1
dsje−tj sj(D2
•+1)[D2
•, T ]e−tj(1−sj )(D2
•+1) ,
0
we commute all heat operators e−tj (D2
•+1) to the right, producing in each step a factor of tj.
The commutators [H, T ] are expressed by Lemma 4 and produce in each step at most one
derivative ∇. In the terms with all heat operators already on the right we then commute
the derivatives ∇ to the right of all functions fj but left of all heat operators. The result is
a finite sum of terms (with redefined fj, Mj)
X =
Γ( m1
R[0,1]N ds P (s)
2 )· · · Γ( mk−1
2
)Γ( mk+z
2
)Z dt1 · · · dtk t
1
m1
2 +β1−1
mk−1
2 +βk−1−1
· · · t
k−1
mk +z
2 +βk−1
t
k
e−t
×(cid:0)M1 · · · Mke(−)• Ω(t1+···+tk)Σ(cid:1) L⋆(cid:0)∂γ1f1 ⋆ · · · ⋆ ∂γk fk(cid:1)Pγ1,...,γk( ∇) e−(t1+···+tk)H ,
(23)
with m + β ≥ γ and β ≤ K (K can be chosen as big as one wishes by pushing the
expansion far enough), plus a finite sum of remainders (with redefined fj, Mj)
Y =
1
)Γ( mk+z
2
2
Γ( m1
2 )· · · Γ( mk−1
×Z[0,1]N′
ds P ′(s)
kYj=1
)Z dt1 · · · dtk t
1
m1
2 +β′
1−1
mk−1
2 +β′
k−1−1
· · · t
k−1
mk +z
2 +β′
k−1
t
k
L⋆(∂γ′
j fj)MjPγ′
( ∇)e−τj (D2
•+1) ,
(24)
j
19
Pk
with γ′ ≤ m + β′ and β′ > K, where τj are positive functions of {s}, t1, . . . , tk with
j=1 τj = t1 + · · · + tk.
In (23) we can use Lemma 4 to express Pγ( ∇) as a finite sum of Pγ′(x)Pγ′′(∇). The
polynomial P ′
γ(x) can (under the trace) be moved into L⋆(f ). We change the variables
t1 = t(1 − u2)(1 − u3)· · · (1 − uk), t2 = tu2(1 − u3)· · · (1 − uk), t3 = tu3(1 − u4)· · · (1 −
uk),. . . tk−1 = tuk−1(1 − uk) and tk = tuk with Jacobian tk−1(1 − u3)(1 − u4)2 · · · (1 − uk)k−2
and obtain
Tr(X) = R[0,1]N ds P (s)
Γ( m+z
Γ( mj
2 + βj)
Γ( mj
2 + β)(cid:16) −1Yj=1
× trC2d(cid:0)M1 · · · Mke(−1)• ΩtΣ(cid:1) Tr(cid:16)L⋆(cid:0)xγ0 ⋆ ∂γ1f1 ⋆ · · · ⋆ ∂γk fk(cid:1)Pγ1,...,γk (∇) e−tH(cid:17) .
2 ) (cid:17)Γ( mk+z
2 + βk)
Γ( mk+z
dt e−t t
2 +β−1
m+z
)
2
(25)
Z ∞
0
Note that Γ(
Γ(
2 +βj)
mk +z
)
mk +z
2
is a polynomial in z of degree βk.
The traces Tr(L⋆(f )∇µ1 · · ·∇µr e−tH ) are computed in Proposition 6. Accordingly, they
have, up to a remainder which leads to a holomorphic function in z, an asymptotic expansion
Tr(L⋆(f )∇µ1 · · ·∇µγe−tH ) =
t−d/2−[ γ
2 ]+aZRd
N ′′Xa=0
dx f (x)P a
γ(x) ,
where P a
γ(x) is a ⋆-polynomial of degree ≤ γ + 2a. Inserted into (25), the t-integral of any
such term yields (together with trC2d(cid:0)M1 · · · Mke(−1)• ΩtΣ(cid:1)) a linear combination of
Γ( m+2β−2[γ/2]+2a+z−d
)
2
Γ( m+2β+z
2
)
ZRd
dx f (x)P a
γ(x) .
As a function of z ∈ C, the latter is trivially holomorphic in C \ Z. As m + β ≥ γ, this
function is also holomorphic for z > d. If z = d − N there is a finite number of parameters
is a non-positive integer smaller than m+2β+d−N
βj, mj, a, γj for which m+2β−2[γ/2]+2a−N
.
Precisely these parameters yield simple poles at z = d − N. Each residue has the claimed
structure.
2
2
We estimate the remainders (24) in trace norm by:
kM1k· · ·kMkk
2 )· · · Γ( mk+z
Γ( m1
2
1
m1
2 +β′
)Z dt1 · · · dtk t
×Z[0,1]N′
ds P ′(s)(cid:13)(cid:13)(cid:13)
1−1
mk−1
2 +β′
k−1−1
· · · t
k−1
mk +z
2 +β′
k−1
t
k
L⋆(∂γ′
j fj)e(−1)•τj ΩΣPγ′
j
kYj=1
(∇)e−τj (H+1)(cid:13)(cid:13)(cid:13)1
.
By spectral theory and Corollary 5, we get for p ∈ [1,∞):
j /2ke−τ ′
(∇)e−τj (H+1)kp ≤ C(ǫτj)−γ′
kPγ′
j
j (H+1)(1−ǫ)kp .
20
j=1 tj, the Holder inequality
Since the τj are linear in tj withPk
gives
L⋆(∂γ′
j fj)e(−1)•τj ΩΣPγ′
(cid:13)(cid:13)(cid:13)
kYj=1
≤
j=1 tj, setting t =Pk
j=1 τj =Pk
j (H+1)(cid:13)(cid:13)(cid:13)1
(∇)e−τ ′
j
(∇)e−τj (H+1)kt/τ ′
j
kYj=1
j fj)k ke(−1)•τj ΩΣkt/τjkPγ′
kL⋆(∂γ′
kYj=1
j/2kL⋆(∂γ′
(ǫtj)−γ′
j
≤ Ce−t(1−ǫ)
= Ce−t(1−ǫ− Ωǫd)(cid:16) kYj=1
(ǫtj)−γ′
j /2kL⋆(∂γ′
j fj)k (2 cosh( Ωt))
dτj
t
dτj
t (2 sinh( Ωt(1 − ǫ))−
cosh( Ωt)
j fj)k(cid:17)(cid:16)
sinh( Ωt(1 − ǫ))e Ωtǫ(cid:17)d
,
where results of Proposition 6 and Lemma 7 have been used. Hence for ǫ′ := (1−ǫ− Ωǫd) > 0
the remainders (24) are bounded in trace norm by
C ′
)Z dt1 · · · dtk t
m1
2 +β′
1−1−
γ′
1
2
mk +z
2 +β′
k−1−
γ′
k
2
2
1
k
· · · t
2 )· · · Γ( mk+z
(t1 + · · · +tk)−de−ǫ′(t1+···+tk) .
Γ( m1
Remember that γ′ ≤ β′ +m and β′ > K and that K can be chosen as big as one wishes
(by pushing the expansion over and over). So given M ≤ 2d, by choosing K > M/2 + d,
we see that the remainder terms are well defined as a trace-class operators for ℜ(z) > M.
A similar analysis involving the z-derivative of the remainders can be done, showing that
by pushing the expansion far enough, the remainders yield holomorphic contributions for
ℜ(z) > M, with M ∈ R arbitrary.
3 The spectral action
3.1 Generalities
For a unital spectral triple with real structure (A,H,D, J), according to the spectral action
principle [3, 5], the bosonic action should depend only on the spectrum of the fluctuated
Dirac operator
D 7→ DA := D + A + ε′JAJ −1 ,
where A = P ai[D, bi], ai, bi ∈ A (finite sum), is a self-adjoint one-form and ε′ = ±1
depending on the KO-dimension of the triple. Ideally, such an action functional (of D and
of A) should be defined as the number of eigenvalues of D2
A smaller than a given scale Λ > 0:
or akin to the same and with χ the characteristic function of the interval [0, 1]:
SΛ(DA) = ♯(cid:8)λn : λn ∈ Spect(D2
A), λn ≤ Λ(cid:9) ,
A/Λ2)(cid:1) .
SΛ(DA) = Tr(cid:0)χ(D2
21
(26)
Then, the diverging part of SΛ(DA) in the limit Λ → ∞ should give access to an effective
action describing low energy physics. The problem is that with the characteristic function,
the expression (26) may not have a well-defined power series expansion in the limit Λ → ∞.
To overcome this difficulty one uses, instead of the characteristic function, a smooth one
approximating it and being the inverse Laplace transform of a Schwartz function on R∗
+. By
Laplace transformation one then has
SΛ(DA) =Z ∞
0
dt Tr(e−tD2
A/Λ2
) χ(t) ,
where χ is the inverse Laplace transform of χ. Assuming that the trace of the heat kernel
has an asymptotic expansion
Tr(e−tD2
A) =
∞Xk=−n
ak(D2
A) tk ,
n ∈ N ,
we obtain
One easily finds
Z ∞
0
A) Λ−2k Z ∞
0
ak(D2
SΛ(DA) =
∞Xk=−n
dt tk χ(t) =( 1
Γ(−k)R ∞
(−1)kχ(k)(0) ,
0 ds s−k−1χ(s) ,
for k /∈ N ,
for k ∈ N .
dt tk χ(t) .
(27)
If one only wants to keep the non-vanishing terms in the power-Λ expansion as Λ → ∞,
it is therefore sufficient to identify the the non-vanishing terms in the power-t expansion of
Tr(e−tD2
A) as t → 0.
If the spectral geometry (A,H,D, J) is non-unital, then the expression (26) becomes ill-
defined. There are different ways to 'regularize' the spectral action in this case. For instance
one may consider instead
SΛ(DA) := Tr(cid:0)χ(D2
A/Λ2) − χ(D2/Λ2)(cid:1) .
But the problem is then that one loses a lot of physical information since one cannot access
in this way the Einstein-Hilbert action. Another possibility, used in [13], is to introduce a
supplementary (adynamic) scalar field ρ ∈ A and to define
SΛ(DA, ρ) := Tr(cid:0)ρ χ(D2
A/Λ2)(cid:1) .
The advantage of this scheme is that one keeps the physical interpretation by performing
on the field equations the adiabatic limit ρ → 1 and that one can choose the one-form A
coming from the unitization of A and not necessarily from A itself. However, in the case of
the Moyal spectral triple with ordinary Dirac operator, treated in [13], the full computation
22
with the real structure was not possible; only the spectral action for partially fluctuated
Dirac operator D 7→ D + A was evaluated. The last possibility spelled out in [4] is to replace
the scale Λ by a dilaton field. That is, one performs the replacement
and one leaves (26) as it was:
Λ 7→ e−φ, φ∗ = φ ∈ A ,
Sφ(DA) := Tr(cid:0)χ(eφ D2
A eφ)(cid:1) .
Although this expression is analytically well-defined and conceptually perfect, the explicit
computation of such a functional seems to be fairly inaccessible, except for the commutative
(manifold) case.
In our setting of a spectral triple for Moyal plane with harmonic propagation, the question
of the definition and the computation of the spectral action is way more easy. This is because
even if the spectral triple (A⋆,H,D•), • = 1, 2, is non-unital, the heat operator e−tD2
• is trace-
class for all t > 0 (see Lemma 9). This means that the definition (26) of the spectral action
for unital spectral triple is still adapted to our situation.
In the next subsections we will perform a complete computation of the spectral action
for a U(1)-Higgs model for d = 4. Before this, we will derive a generic heat kernel type
expansion when one tensorizes (A⋆,H,D•) with a finite spectral triple.
3.2 Heat kernel expansion in dimension four
We derive here a short-time heat-kernel expansion for the semi-group generated by the
square of a twisted harmonic Dirac operator, for the algebra of Schwartz functions with
Moyal product. We start with preliminary results on Schatten norm estimates, using the
estimate of Lemma 12 together with complex interpolation methods. Here we specify to the
case d = 4.
Proposition 18. Let f ∈ A⋆. Then for all 1 ≤ p ≤ ∞ and t ∈ (0, 1], we have
kL⋆(f )e−tD2
•kp ≤ kfk1−1/p
2
C(f )1/p p−2/p t−2/p ,
where C(f ) is the constant appearing in Lemma 12.
Proof. For f ∈ A⋆, t ∈ (0, 1] and 1 ≤ p < ∞, consider on the strip S := {z ∈ C : ℜ(z) ∈
[0, 1]} the operator-valued function
Fp : z 7→ L⋆(f )e−tpzD2
• .
The function Fp is continuous on S, holomorphic on its interior and by Lemma 12 it satisfies
for y ∈ R:
kFp(iy)k ≤ kfk2,
kFp(1 + iy)k1 ≤ C(f )(pt)−2 .
23
Then, by standard complex interpolation methods (see for example [30]) we have Fp(z) ∈
L1/ℜ(z)(H) for all z ∈ S with
kFp(z)k1/ℜ(z) ≤ kFp(0)k1−ℜ(z)
∞
Applying this for z = 1/p, we get
kFp(1)kℜ(z)
1 ≤ kfk1−ℜ(z)
2
C(f )ℜ(z)(pt)−2ℜ(z) .
kL⋆(f )e−tD2
•kp ≤ kfk1−1/p
2
C(f )1/p p−2/p t−2/p ,
as needed.
Remark 19. For f ∈ A⋆, making a recursive chose of factorization as follows:
f = f1 ⋆ f2,
∂µ1f2 = f1,µ1 ⋆ f2,µ1,
· · ·
∂µ4f2,µ1µ2µ3 = f1,µ1µ2µ3µ4 ⋆ f2,µ1µ2µ3µ4 ,
the constant C(f ) appearing in Lemma 12 (for d = 4) and Proposition 18 is a finite multiple
(depending only on Ω and Θ) of
k ¯f1 ⋆ f1k1/2
1 k ¯f2 ⋆ f2k1/2
1 + kf1k2
4Xµ1=1(cid:16)k ¯f1,µ1 ⋆ f1,µ1k1/2
4Xµ2=1(cid:16)
1 + kf1,µ1k2
1 k ¯f2,µ1 ⋆ f2,µ1k1/2
4Xµ3=1(cid:16)k ¯f1,µ1µ2µ3 ⋆ f1,µ1µ2µ3k1/2
1
k ¯f1,µ1µ2 ⋆ f1,µ1µ2k1/2
1 k ¯f2,µ1µ2 ⋆ f2,µ1µ2k1/2
1 + kf1,µ1µ2k2
× k ¯f2,µ1µ2µ3 ⋆ f2,µ1µ2µ3k1/2
× k ¯f2,µ1µ2µ3µ4 ⋆ f2,µ1µ2µ3µ4k1/2
1 + kf1,µ1µ2µ3k2
4Xµ4=1(cid:16)k ¯f1,µ1µ2µ3µ4 ⋆ f1,µ1µ2µ3µ4k1/2
1
1 + kf1,µ1µ2µ3µ4k2kf2,µ1µ2µ3µ4k2(cid:17)(cid:17)(cid:17)(cid:17) .
Lemma 20. Let f ∈ A⋆, 1 ≤ p ≤ ∞, t ∈ (0, 1] and k ∈ N. Then, there exists a finite
constant Cp,k(f ) such that
kL⋆(f )Dk
•e−tD2
•kp ≤ Cp,k(f ) t−2/p−k/2 .
Proof. By spectral theory, we have kDk
Proposition 18.
• e−tD2
•k = (k/2et)k/2, so the proof is a consequence of
The next Lemma will explain why there is a major difference in the spectral action when
perturbing D by A + JAJ −1 or simply by A.
Lemma 21. For f, g ∈ A⋆, the operator L⋆(f ) R⋆(g) is of trace class on H.
Proof. By factorization, we can find f1, f2, g1, g2 ∈ A⋆ such that
f = f1 ⋆ f2 ,
g = g1 ⋆ g2 .
24
Hence
L⋆(f ) R⋆(g) = L⋆(f1 ⋆ f2) R⋆(g1 ⋆ g2) = L⋆(f1)L⋆(f2) R⋆(g1)R⋆(g2) .
But since the left and right regular representations commute (by associativity of the Moyal
product), we get
L⋆(f ) R⋆(g) = L⋆(f1)R⋆(g1) L⋆(f2)R⋆(g2) ,
so that it suffices to show that L⋆(f ) R⋆(g) is Hilbert-Schmidt for all f, g ∈ A⋆. From the
operator kernel formula (17) of L⋆(f ), and a similar one for R⋆(g), one easily deduces the
operator kernel for the product L⋆(f ) R⋆(g), and after a few lines of computations, we get
for a suitable constant depending only on det(Θ):
kL⋆(f ) R⋆(g)k2
2 =Z dx dy(cid:12)(cid:12)[L⋆(f ) R⋆(g)](x, y)(cid:12)(cid:12)2 = Ckfk2
2 kgk2
2 .
This completes the proof.
Corollary 22. Let ∇a
µ, µ = 1,· · · , 4, be the operators on L2(R4) given by
µ := i∂µ + aµνxµ ,
∇a
a ∈ M4(R) .
Then for f, g ∈ A⋆, t ∈ (0, 1] and Pα( ∇) a polynomial of order α in the operators ∇a
exists a finite constant C(f, g, α) such that
µ, there
(cid:13)(cid:13)L⋆(f )R⋆(g)Pα( ∇)e−tH(cid:13)(cid:13)1 ≤ C(f, g, α) t−α/2 .
Proof. From Lemma 21, it suffices to show that kPα( ∇)e−tHk ≤ Ct−α/2, which will follow
by spectral theory if Pα( ∇)(1 + H)−α/2 is bounded. But this is a slight generalization of
Corollary 5.
We can now deduce the germ of the asymptotic expansion formula we need.
Proposition 23. Let (Af ,Hf ,Df , Jf ) be a finite spectral triple. Let D := D• ⊗ 1 + Γ ⊗ Df ,
• = 1, 2, be the Dirac operator of the product spectral triple (A⊗Af ,H⊗Hf ,D•⊗1+Γ⊗Df ).
Let also DA := D + A + JAJ−1 be the fluctuated Dirac operator. Here J := J ⊗ Jf and A
A = D2 +F0 +F1 +J(F0 +F1)J−1+2AJAJ−1,
ai, bi ∈ A⊗Af . In terms of the decomposition D2
where F0 is a bounded operator and F1 is linear in the operators ∇a
µ of Corollary 22, the
following holds:
is a self-adjoint one-form, that is A = A∗ := Pi ai[D, bi], where the sum is finite and
Tr(cid:0)e−tD2
0 + F1F0 + F0F1 + F 2
A(cid:1) = Tr(cid:16)n1 − 2t(F0 + F1) + t2(cid:0)F 2
1(cid:1)
3(cid:0)F0[D2, F1] − [D2, F1]F0 + F1[D2, F1] + F0F 2
12(cid:0)F1[D2, [D2, F1]] + 2F 2
1 [D2, F1] + F1 [D2, F1]F1 + F 4
−
+
t3
t4
1 + F1F0F1 + F 2
1 F0 + F 3
1(cid:1)
1(cid:1)oe−tD2(cid:17) + O(√t) .
25
Proof. First, it is clear that e−tD2
A is of trace-class for all t > 0. Indeed, since the eigenvalues
of D• behave like n−1/8, its resolvent belongs to the Schatten ideal L8+ε(H) for all ε > 0.
From the relation
1
D + i
=
1
D• ⊗ 1 + i(cid:16)1 − (A + JAJ−1 + Γ ⊗ Df )
1
D + i(cid:17) ,
and the fact that A + JAJ−1 + Γ⊗ Df is bounded, we see that the resolvent of D belongs to
L8+ε(H ⊗ Hf ) for all ε > 0 too. Accordingly, e−tD2
Note also that the bounds of Lemmas 10, 12, 20 and Proposition 18, 22 remain valid
with D instead of D•. Indeed, since {D•, Γ} = 0 and Γ2 = 1, we get D2 = D2
• ⊗ 1 + 1 ⊗ D2
and thus
A is of trace-class for all t > 0.
f
Dke−tD2
=
Ck,jΓk−jDj
•e−tD2
• ⊗ Dk−j
f
e−tD2
f .
kXj=0
This implies for f, g ∈ B⋆, a, b ∈ Af and 1 ≤ p ≤ ∞,
kL⋆(f ) ⊗ a JL⋆(g) ⊗ bJ−1 Dke−tD2kp
≤
kXj=0
Ck,jkL⋆(f ) R⋆(g)Dj
•e−tD2
•kpkaJf bJ −1
f Dk−j
f
e−tD2
f kp .
Thus, we may assume without loss of generality that there is no finite spectral triple in the
picture.
We are going to deduce the expansion from the Duhamel principle:
e−t(A+B) = e−tA − tZ 1
0
e−st(A+B) B e−(1−s)tA ds .
A = D2 + F0 + F1, with F0 := F0 + JF0J−1 + 2AJAJ and F1 := F1 + JF1J−1. The
We write D2
operator F0 is bounded, whereas F1 is unbounded but relatively D-bounded. The Duhamel
expansion allows us to write (formally first):
e−tD2
A =
∞Xj=0
(−t)j Ej(t) ,
(28)
where E0(t) := e−tD2 and for j > 0:
Ej(t) := Xi1,··· ,ij∈{0,1}Z△j
and △j denotes the ordinary j-simplex:
e−s0tD2 Fi1 e−s1tD2 · · · Fij e−sjtD2
djs ,
△j :=(cid:8)s ∈ Rj+1; sk ≥ 0,
jXk=0
sk = 1(cid:9) .
26
We first show that the sum (28) converges in the trace norm for small values of t > 0. We
only treat the case j ≥ 1, the case j = 0 being covered by Lemma 10. For that we use the
Holder inequality (sincePj
kEj(t)k1 ≤ Xi1,··· ,ij∈{0,1}Z△j ke−s0tD2 Fi1 e−s1tD2
k=0 sk = 1):
k(s0+s1)−1k Fi2 e−s2tD2
ks−1
2
· · ·k Fij e−sjtD2
ks−1
j
djs .
Then we use the estimate of Proposition 18 and Lemma 21 for k = 2,· · · , j (see the Remark
19 for the precise value of the constants):
k Fik e−sktD2ks−1
k ≤(4k Fk2C1( F )sk t−2sk ,
2k FkC2( F )sk t−2sk (tsk)−1/2 ,
if
if
ik = 0 ,
ik = 1 .
For the case ik = 1, we need to use the factorization property of the algebra of Schwartz
functions with Moyal product (as in the proof of Lemma 12), to expand A as a finite sum
of products of elements in A⋆ ⊗ M16(C), and then we can proceed as for the other factors.
Taking into account that there are 2j such terms and that
Z∆j
jYi=0
s−1/2
i
djs ≤ 2j ,
we get, sincePj
Thus the sumP∞
k=0 sk = 1, the rough estimate
kEj(t)k1 ≤ 2j(cid:0)4k Fk2 + 4k Fk(cid:1)j
t−j/2−2 .
These estimates also show that
j=0(−t)j Ej(t) converges absolutely in the trace-norm for small values of t.
t → 0 ,
∞Xj=5
Note first that
(cid:12)(cid:12)(cid:12)Tr(cid:0)e−tD2
A(cid:1) −
Tr(cid:16) − t e−stD2
(−t)j Tr(cid:0)Ej(t)(cid:1)(cid:12)(cid:12)(cid:12) = O(t1/2) ,
and accordingly, we only need to consider the terms (−t)jTr(cid:0)Ej(t)(cid:1) for j = 0, 1, 2, 3 and 4.
Tr(−tE1(t)) =Z 1
( F0 + F1) e−(1−s)tD2(cid:17) ds = Tr(cid:16) − t ( F0 + F1) e−tD2(cid:17) .
E2(t) =Z△2
=Z△2
For j = 2, 3 and 4, we use the relation (18) to collect the heat operators as follows:
ds1ds2 − tZ△2Z 1
dr (s2 − s1) e−s1tD2
( F0 + F1)2e−(1−s1)tD2
e−s1tD2
( F0 + F1) e−(s2−s1)tD2
( F0 + F1) e−(1−s2)tD2
ds1ds2
( F0 + F1)
0
e−s1tD2
0
27
e−(1−s2)tD2
ds1ds2
× e−(s2−s1)rtD2
[D2, ( F0 + F1)]e−(s2−s1)(1−r)tD2
0
( F0 + F1)
dr1dr2 r1(s2 − s1)2 e−s1tD2
[D2, [D2, ( F0 + F1)]]e−(s2−s1)(1−r1r2)tD2
=Z△2
ds1ds2 e−s1tD2n( F0 + F1)2 − t(s2 − s1)( F0 + F1) [D2, ( F0 + F1)]oe−(1−s1)tD2
ds1ds2Z 1
+ t2Z△2
× e−(s2−s1)r1r2tD2
=Z△2
ds1ds2 e−s1tD2n( F0 + F1)2 − t(s2 − s1)( F0 + F1) [D2, ( F0 + F1)]
ds1ds2Z 1
− t3Z△2
× [D2, [D2, [D2, ( F0 + F1)]]]]e−(s2−s1)(1−r1r2r3)tD2
(s2 − s1)2( F0 + F1)[D2, [D2, ( F0 + F1)]]oe−(1−s1)tD2
1r2(s2 − s1)3 e−s1tD2
( F0 + F1) e−(s2−s1)r1r2r3tD2
dr1dr2dr3 r2
e−(1−s2)tD2
e−(1−s2)tD2
t2
2
+
.
0
(29)
Since the principal symbol of D2 is scalar, we see that [D2, [D2, [D2, F1]]]] has order 4. Thus,
Lemma 20 shows that the last integral (multiplied by its t2 global prefactor) gives rise
to a trace-class operator which trace is of order t1/2.
Integrating the trace over ∆2 and
disregarding the terms that vanish when t → 0, we find
Tr(t2E2(t)) = Tr(cid:16)nt2
2
( F0 + F1)2 −
+
t3
6(cid:0) F0[D2, F1] + F1[D2, F0] + F1[D2, F1](cid:1)
F1[D2, [D2, F1]]oe−tD2(cid:17) + O(√t) .
t4
24
It will be more convenient to write Tr(cid:16)F1[D2, F0]e−tD2(cid:17) = Tr(cid:16)−[D2, F1]F0e−tD2(cid:17). By similar
arguments one finds
Tr(−t3E3(t)) = Tr(cid:16)n −
+
and lastly
t3
6
t4
24
( F0 F 2
1 + F1 F0 F1 + F 2
1
F0 + F 3
1 )
(2 F 2
1 [D2, F1] + F1 [D2, F1] F1)o e−tD2(cid:17) + O(√t) ,
In summary, we have
Tr(cid:0)e−tD2
24
F 4
Tr(t4E4(t)) = Tr(cid:16) t4
2(cid:0) F 2
1 e−tD2(cid:17) + O(√t) .
1(cid:1)
6(cid:0) F0[D2, F1] − [D2, F1] F0 + F1[D2, F1] + F0 F 2
0 + F1 F0 + F0 F1 + F 2
A(cid:1) = Tr(cid:16)n1 − t( F0 + F1) +
−
t3
t2
28
1 + F1 F0 F1 + F 2
1
(30)
F0 + F 3
1(cid:1)
+
t4
24(cid:0) F1[D2, [D2, F1]] + 2 F 2
1 [D2, F1] + F1 [D2, F1] F1 + F 4
1(cid:1)oe−tD2(cid:17) + O(√t) .
Now, we can take into account the result of Lemma 21, which says in this context that
mixed products FiJFjJ−1 are already trace-class. Since JL⋆(g)J −1 = R⋆(¯g), with R⋆ the
right regular representation, we see by Lemma 4 that all terms in (30) with products of Fi
and JFjJ−1 are (up to matrices) of the form
L⋆(f )R⋆(g)
αµe−tH ,
( ∇µ)
4Yµ=1
with α not exceeding the number of F1 plus the number of commutators by D2. This
argument also relies on the fact that J commutes with ∇µ according to (13). Thus, Corollary
22 shows that the cross-terms, i.e. the terms with powers of both Fi and JFjJ−1 resulting
from products of F0 = F0 + JF0J−1 + 2F−1JF−1J−1 and F1 = F1 + JF1J−1, where F−1 := A,
give rise to vanishing contributions in the limit t → 0. Thus, only the terms with either
powers of Fi or powers of JFiJ−1 do contribute to the diverging part of this asymptotic the
expansion. Since moreover J commutes with D (we are in even KO-dimension), the trace
property shows that both terms (with only A or only JAJ−1) give the same contribution
and we get the announced result.
Remark 24. A very important feature of Proposition 23 is that if the heat-trace of the
partially fluctuated Dirac operator eDA := D + A, A = Pi ai[D, bi], has an asymptotic
expansion
Tr(cid:0)e−t eD2
A(cid:1) = a0 t−4 +
4Xk=1
ak t−2+k/2 + O(√t) ,
then the heat-trace of the fully fluctuated Dirac operator DA := D + A + JAJ−1 has the
asymptotic expansion
Tr(cid:0)e−tD2
A(cid:1) = a0 t−4 + 2
4Xk=1
ak t−2+k/2 + O(√t) ,
for the same coefficients a0,· · · , ak. Also, this shows that the asymptotic expansion of the
heat-trace of the fully fluctuated Dirac operator is independent of the choice of the real
structure Jf of the finite spectral triple (Af ,Hf ,Df ). This fact holds for Moyal spectral
triples with harmonic propagation in any (even) dimension.
3.3 Application: the spectral action for the U(1)-Higgs model
In the Connes-Lott spirit [8] we take the tensor product of the 4-dimensional spectral triple
(A⋆,H,D•, Γ, J), • = 1, 2, with the finite Higgs spectral triple (C ⊕ C, C2, Mσ1, Jf ), where
M > 0 and Jf is any real structure. The Dirac operator D = D• ⊗ 1 + Γ ⊗ Mσ1 of the
product triple becomes
D =(cid:18) D• MΓ
MΓ D• (cid:19) .
29
In this representation, the algebra is A⋆ ⊕ A⋆ ∋ (f, g), which acts on H ⊕ H by diagonal
left Moyal multiplication. The commutator of D with (f, g) is, in case that D1 is chosen,
according to (8) given by
[D, (f, g)] =(cid:18) iΓµL⋆(∂µf ) MΓL⋆(g − f )
iΓµL⋆(∂µg) (cid:19) .
MΓL⋆(f − g)
(If we choose D2 instead, then Γµ has to be replaced by Γµ+4 everywhere.) This shows
D + A + JAJ−1, J = J ⊗ Jf , is of the form
that the selfadjoint fluctuation A = Pi ai[D, bi] of the fluctuated Dirac operators DA =
A =(cid:18) ΓµL⋆(Aµ)
ΓL⋆( ¯φ)
ΓL⋆(φ)
ΓµL⋆(Bµ) (cid:19) ,
for real two real one-forms Aµ, Bµ ∈ A⋆ and a one complex field φ ∈ A⋆. Again, this holds
for D1; for D2 we have to replace Γµ by Γµ+4.
In terms of the connection introduced in Lemma 4 and using (10) we identify the relevant
• + M 2 + (F0 + F1) + J(F0 + F1)J−1 + 2AJAJ−1
operators arising in the expansion D2
A = D2
of Proposition 23 as follows:
F0 = L⋆(VA,φ)1 + i
F1 =(cid:18) 2iL⋆(Aµ)∇µ
0
iΓµΓL⋆(Dµφ)
0
2iL⋆(Bµ)∇µ (cid:19) ,
µν) ! ,
4 [Γµ, Γν]L⋆(F A
µν)
iΓµΓL⋆(Dµφ)
L⋆(VB,φ)1 + i
4 [Γµ, Γν]L⋆(F B
where
We have used
VA,φ := φ ⋆ ¯φ + M(φ + ¯φ) + (g−1)µν(i∂µAν + Aµ ⋆ Aν) ,
VB,φ := ¯φ ⋆ φ + M(φ + ¯φ) + (g−1)µν(i∂µBν + Bµ ⋆ Bν) ,
F A
µν := ∂µAν − ∂νAµ − i(Aµ ⋆ Aν − Aν ⋆ Aµ) ,
F B
µν := ∂µBν − ∂νBµ − i(Bµ ⋆ Bν − Bν ⋆ Bµ) ,
Dµφ := ∂µφ − iAµ ⋆ φ + iφ ⋆ Bµ − iM(Aµ − Bµ) .
D1ΓσL⋆(Aσ) + L⋆(Aσ)D1Γσ = −Γσ[D1, L⋆(Aσ)] + 2i(g−1)µνL⋆(∂µAν) + 2iL⋆(Aµ)∇µ .
According to the general asymptotic expansion we have obtained in Proposition (23), the
only further commutators we need are [D2, F1] and [D2, [D2, F1]]. Their expression will easily
follows from the following computation which relies on the relations in Lemma 4:
[D2, 2iL⋆(Aµ)∇µ] = −2iL⋆(cid:0)(g−1)ρσ∂ρ∂σAµ(cid:1)∇µ − 4iL⋆(∂νAµ)∇ν∇µ + 4 Ω2ΘµνL⋆(Aµ) ∇ν ,
[D2, [D2, 2iL⋆(Aµ)∇µ] = 8iL⋆(∂ρ∂νAµ)∇ρ∇ν∇µ + lower order ,
30
We have already computed the matrix trace of e−t ΩΣ in Lemma 10, and the result is
16 cosh4( Ωt). Using
the other matrix traces follow from the Clifford algebra:
−iΣ = (ibµ − ib∗
Ω
2
Ω
2
Θστ Γτ(cid:17) ,
ΘνρΓρ+4(cid:17)gµσ(cid:16)Γσ+4 +
µ)(bµ + bµ∗) = gµν(cid:16)Γν +
[Γµ, Γν] · e−t ΩΣ(cid:17) = −8 ΩΘµνt + O(t2) ,
trC16(cid:16) i
[Γρ, Γσ] · e−t ΩΣ(cid:17) = 4(g−1)µρ(g−1)νσ − 4(g−1)µσ(g−1)νρ(cid:1) + O(t)
[Γµ, Γν] ·
trC16(cid:0)iΓµΓ · iΓνΓ · e−t ΩΣ(cid:1) = 16(g−1)µν + O(t) .
4
i
4
trC16(cid:16) i
4
In terms of the functionals Tµ1...µk (f ) := TrL2(R4)(cid:0)L⋆(f )∇µ1 . . .∇µke−tH(cid:1) on A⋆ introduced
and computed in Proposition 6 and the similar functional
we obtain from Proposition 23 the trace Tr(e−tD2
A) as follows:
Tµν(f ) := TrL2(R4)(cid:0)L⋆(f )∇µ ∇νe−tH(cid:1)
Tr(cid:0)e−tD2
+ 32i(g−1)µνAµ ⋆ ∂νVA,φ + 16 ΩΘµνF A
A(cid:1) = e−tM 2(cid:26)16 cosh4( Ωt)Tr(e−tH ) − 2tT(cid:0)16VA,φ(cid:1) − 2tTµ(cid:0)32iAµ(cid:1)
+ t2 T(cid:16)16VA,φ ⋆ VA,φ + 16(g−1)µνDµφ ⋆ Dνφ + 8(g−1)µρ(g−1)νσF A
+ t2 Tµ(cid:16)32iAµ ⋆ VA,φ + 32iVA,φ ⋆ Aµ − 64(g−1)νρAν ⋆ ∂ρAµ(cid:17)
+ t2 Tµν(cid:0) − 64Aµ ⋆ Aν(cid:1)
3 Tµν(cid:16) − 4i((VA,φ ⋆ ∂µAν − ∂µAν ⋆ VA,φ)
+ 64Aµ(g−1)ρσ∂ρ∂σAν + 128Aρ(g−1)ρσ∂σ∂µAν
µν(cid:17)
−
t3
µνF A
ρσ
t3
− 64(cid:0)VA,φ ⋆ Aµ ⋆ Aν + Aµ ⋆ VA,φ ⋆ Aν + Aµ ⋆ Aν ⋆ VA,φ(cid:1)
− 128i(g−1)ρσ(cid:0)Aρ⋆(∂σAµ)⋆Aν + Aρ ⋆ Aµ⋆(∂σAν) + Aµ⋆Aρ⋆(∂σAν)(cid:1)(cid:17)
3 Tµνρ(cid:16)128Aµ ⋆ ∂νAρ − 128iAµ ⋆ Aν ⋆ Aρ(cid:17)
Tµν(cid:0)128i Ω2ΘρνAµ ⋆ Aρ(cid:1)
12Tµνρσ(cid:16) − 256Aµ ⋆ ∂ν∂ρAσ + 512iAµ ⋆ Aν ⋆ ∂ρAσ
+ 256iAµ ⋆ (∂νAρ) ⋆ Aσ + 256Aµ ⋆ Aν ⋆ Aρ ⋆ Aσ(cid:17)(cid:27)
t3
3
t4
−
−
+
31
+(cid:26)Aµ 7→ Bµ , F A
µν 7→ F B
µν , VA,φ 7→ VB,φ , Dµφ ↔ Dµφ(cid:27) + O(√t) .
(31)
The relevant traces have been computed in Proposition 6. A similar procedure gives
Ω
2π sinh(2 Ωt)(cid:17)2ZR4
Tµν(f ) =(cid:16)
:= −2i(Θ−1z)ν − 2 Ω tanh( Ωt)(gz)ν and Nµν
dzpdet g f (z)(cid:16) Nµν + Zµ Zν(cid:17)e− Ω tanh( Ωt)hz,gzi ,
with Zν
Ω tanh( Ωt)δµν. This shows that the contribution of Tµν(f ) is suppressed with O(t).
serting these traces into (31) we arrive at
:= 2i(Θ−1g−1)µν + i Ω2(gΘ)µν −
In-
Tr(e−tD2
A)
=
2
Ω4
t−4 −
2M 2
Ω4
− t−1 2 − 2M 2t
π2
8
+
+
Ω4
3 Ω4
8M 2
3 Ω2(cid:17)t−1 +(cid:16) 52
3 Ω2(cid:17)t−2 −(cid:16) M 6
t−3 +(cid:16)M 4
Z d4xpdet g(cid:26)φ ⋆ ¯φ + M(φ + ¯φ) + Ω2(cid:0)hXA, gXAi⋆ − hx, gxi⋆(cid:1)
+ ¯φ ⋆ φ + M(φ + ¯φ) + Ω2(cid:0)hXB, gXBi⋆ − hx, gxi⋆(cid:1)(cid:27)
M 8
12 Ω4
45
+
+
(32)
4M 4
3 Ω2(cid:17)
1
+
π2Z d4xpdet g(cid:26)2(g−1)µνDµφ ⋆ Dνφ
+(cid:0)φ ⋆ ¯φ + M(φ + ¯φ) + Ω2hXA, gXAi⋆(cid:1)2 −(cid:0) Ω2hx, gxi⋆(cid:1)2
+(cid:0) ¯φ ⋆ φ + M(φ + ¯φ) + Ω2hXB, gXBi⋆(cid:1)2 −(cid:0) Ω2hx, gxi⋆(cid:1)2
+(cid:16) 1
+ O(√t) ,
(g−1 + Ω2ΘgΘ)µρ(g−1 + Ω2ΘgΘ)νσ(cid:17)(cid:0)F A
(g−1)µρ(g−1)νσ −
1
6
2
µν ⋆ F A
ρσ + F B
µν ⋆ F B
ρσ(cid:1)(cid:27)
where
X µ
A(x) := xµ + ΘµνAν ,
hX, gY i⋆ := gµνX µ ⋆ Y ν .
To reduce (31) to (32) we have used
– the traciality (4) of the Moyal product and the resulting cyclicity under the integral,
– integration by parts where appropriate,
– xµ ⋆ f = 1
2[xµ, f ]⋆ = 1
2{xµ, f}⋆ + 1
– symmetries and antisymmetries in the indices.
2{xµ, f}⋆ + iΘµν∂νf where appropriate,
The matrix g−1 + Ω2ΘgΘ appearing in front of the curvature term in (32) can be equivalently
written as
g−1 + Ω2ΘgΘ = g−1 + 4g(1 − g−1) = g−1(1 − 2g)2 .
32
Since g−1 ≥ 1 according to (9), hence 0 ≤ g ≤ 1, we have 0 ≤ g−1 + Ω2ΘgΘ ≤ g−1, showing
that the matrix in front of the curvature term in (32) is strictly positive. It is minimal for
2, i.e. Ω2(ΘtΘ)µν = 4δµν. The curvature Fµν can also be expressed in terms of the
g = 1
covariant coordinates, [X µ
A]⋆ = iΘµν + iΘµρΘνσF A
ρσ.
0 ds sn−1χ(s) of the "characteristic function" and χ0 = χ(0),
A, X ν
With the moments χ−n =R ∞
we identify the spectral action (27) as
8M 2Λ2
χ0
+
2
2M 2Λ6
Ω4
4M 4
+
45
+
3 Ω4
8Λ4
+
Ω4
+
2Λ8
Ω4
χ−4 −
M 8
12 Ω4
SΛ(DA) =
χ−1
χ0
χ−1
χ0
+(cid:16)52
3 Ω2 (cid:17)χ−1
3 Ω2(cid:17)χ−2 −(cid:16) M 6Λ2
χ−3 +(cid:16) M 4Λ4
3 Ω2(cid:17)χ0
π2Z d4xpdet g(cid:26)2(g−1)µνDµφ ⋆ Dνφ
+(cid:16)φ ⋆ ¯φ + M(φ + ¯φ) + Ω2hXA, gXAi⋆ + M 2 −
+(cid:16) ¯φ ⋆ φ + M(φ + ¯φ) + Ω2hXB, gXBi⋆ + M 2 −
+(cid:16)1
+ O(Λ−1) .
The final result (33) for the spectral action agrees, up to typos, with the result obtained
in [21]. We recall that with the cumbersome computational method of [21] it was only
possible to identify the part of the spectral action at most bilinear in the gauge fields A, B.
By gauge-invariant completion it was argued that the total spectral action has to be (33).
(g−1 + Ω2ΘgΘ)µρ(g−1 + Ω2ΘgΘ)νσ(cid:17)(cid:0)F A
−(cid:16) Ω2hx, gxi⋆ + M 2 −
−(cid:16) Ω2hx, gxi⋆ + M 2 −
Λ2(cid:17)2
Λ2(cid:17)2
ρσ(cid:1)(cid:27)
In [21] the Θ-matrix was chosen as Θ = θ(cid:18) iσ2
2 this choice
leads to (g−1)µν = (1 + Ω2)δµν and √det g =
(1+Ω2)2 . Up to the global factor of 2 due to the
real structure, the few differences in the prefactors3 are easily identified as typos in [21]. We
finish by a brief discussion of the spectral action:
iσ2 (cid:19). In terms of Ω := θ Ω
(g−1)µρ(g−1)νσ −
Λ2(cid:17)2
Λ2(cid:17)2
χ−1
χ0
χ−1
χ0
µν ⋆ F A
ρσ + F B
µν ⋆ F B
0
0
1
1
6
(33)
• The square of covariant coordinates XA, XB combines with the Higgs field φ to a non-
trivial potential. This was not noticed in [14, 16]. We observe here a much deeper
unification of the continuous geometry described by Yang-Mills fields and discrete
geometry described by the Higgs field than previously in almost-commutative geometry.
• The coefficient in front of the Yang-Mills action is strictly positive for any real-valued
Ω. In the bosonic model of [14, 16] there was only the analogue of the negative part,
which leads to problems with the field equations.
3These are(cid:16) (1+Ω2)2
1
2 in [21] in front of (Dφ)2.
2 − (1−Ω2)4
6(1+Ω2)2(cid:17) versus(cid:16) (1−Ω2)2
2 − (1−Ω2)4
3(1+Ω2)2(cid:17) in [21] in front of Fµν F µν and (1+Ω2)
2
versus
33
Unlike the scalar model renormalized in [19] where Ω = θ Ω
2 can by Langmann-Szabo
duality be restricted to Ω ∈ [0, 1], the full spectral action (33) does not have a distin-
guished frequency parameter Ω > 0.
• The action (33) is invariant under gauge transformations
φ + M 7→ uA ⋆ (φ + M) ⋆ uB , XAµ 7→ uA ⋆ XAµ ⋆ uA , XBµ 7→ uB ⋆ XBµ ⋆ uB ,
where uA, uB ∈ U(A⋆ ⊕ C) are unital elements of the minimal unitization.
• For any value of the free parameter M 2χ0
, the action contains (A, B, φ)-linear terms
which lead to a complicated vacuum which is not attained at vanishing A, B, φ. Since
A, B, φ are Schwartz functions, the formal vacuum solution XA = 0 = XB and φ+M =
q χ−1
Λ is excluded. An enlargement of Schwartz class function to e.g. polynomially
bounded functions does not help either, because then we are not allowed to expand
the Gaussian e− Ω tanh( Ωt)hx,gxi⋆ in t, making the spectral action different from (33).
Λ2χ−1
χ0
• If we formally regard φ + M, XA, XB as dynamical variables of the model, then (33)
can be viewed as translation-invariant with respect to
φ(x) + M 7→ φ(x+a) + M ,
XA(x) 7→ XA(x+a) ,
XB(x) 7→ XB(x+a) .
This would clear away a frequent objection against the renormalizable φ4
4-models,
breaking of translation invariance. However, this transformation leaves the space of
Schwartz class functions for A, B, φ, so that translation invariance remains broken in
the consistent spectral action.
• The vacuum part of the spectral action is finite. In general, the heat kernel expansion
for non-compact spectral triples is ill-defined, so that a spatial regularization of the
operator trace is unavoidable. See e.g. [13]. The oscillator potential is one of many
possibilities. We want to advertise the point of view that if one takes the spectral
action principle serious, the spatial regularization is part of the geometry. The removal
of the spatial regularization must be carefully studied. In general, we should expect
that other limiting procedures such as those of quantum field theory make it impossible
to remove the regularization (UV/IR).
A Locally compact noncommutative spin manifolds
Definition 25. A non-compact spectral triple is given by the data (A,B,H,D, J, Γ, c)
satisfying conditions 0-6 given below. The data consist of a non-unital algebra A acting
faithfully (via a representation denoted by π) by bounded operators on the Hilbert space H; a
preferred unitization B of A acting by bounded operators on the same Hilbert space; and an
essentially self-adjoint unbounded operator D on H such that [D, π(a)] extends to a bounded
operator for any a ∈ B. The spectral triple is said to be even if there exists a Z2-grading
operator Γ on H satisfying Γ2 = 1, for which B is even and D is odd. The spectral triple is
34
said to be real if there exists an antiunitary operator J on H which satisfies conditions 4
and 5 below.
0. Compactness.
1
2 .
The operator π(a)(D − λ)−1 is compact for all a ∈ A and λ in the resolvent set of D.
n=1 dom(δn), with δ(T ) := [hDi, T ]
For any a ∈ B, both π(a) and [D, π(a)] belong toT∞
and hDi := (D2 + 1)
For any element b of the algebra Ψ0(A) generated by δn(π(A)) and δn([D, π(A)]),
the function ζb(z) := Tr(bhDi−z) is well defined and holomorphic for ℜ(z) large and
analytically continues to C\Sd for some discrete set Sd ⊂ C (the dimension spectrum).
Moreover the dimension spectrum is said to be simple if all the poles are simples, finite
if there is k ∈ N such that all the poles are order at most k and if not, infinite.
1. Metric dimension.
For the metric dimension d := sup{ℜ(z), z ∈ Sd}, the operator π(a)hDi−d belongs to
the Dixmier ideal L1,∞(H) for any a ∈ A. Moreover, for any Dixmier trace, the map
A+ ∋ a 7→ Trω(π(a)hDi−d) is non-vanishing.
2. Finiteness.
The algebra A and its preferred unitization B are pre-C ⋆-algebras, i.e. each one is a
⋆-subalgebra of some C ∗-algebra and stable under holomorphic functional calculus.
∞\k=0
Hk, with Hk := dom(Dk) completed with norm
The space of smooth spinors H∞ :=
kξk2
k := kξk2 + kDkξk2, is a finitely generated projective A-module pAm, for some
m ∈ N and some projector p = p2 = p∗ ∈ Mm(B). The composition of the Dixmier
trace with the induced hermitian structure h , iA : H∞ × H∞ → A coincides with the
scalar product ( , ) on H∞,
(ξ, η) = Trω(cid:16)hξ, ηiA hDi−d(cid:17) ,
ξ, η ∈ H∞ .
3. Reality.
The operator J defines a real structure of KO-dimension k ∈ Z8. This means
(even case)
J 2 = ε ,
(34)
JD = ε′DJ ,
JΓ = ε′′ΓJ
with signs ε, ε′, ε′′ ∈ {−1, 1} given as a function of k mod 8 by
6 7
1 1
1 1
−1
1
5
1 −1 −1 −1 −1
1 −1
1
k
ε
ε′ 1 −1
ε′′ 1
1
−1
0
1
2
4
3
1
Additionally, the action π of B on H satisfies the commutation rule [π(f ), πo(g)] = 0
for all f, g ∈ B, where πo(g) = Jπ(g∗)J −1 is the action of the opposite algebra Bo.
35
4. First order.
[[D, π(f )], πo(g)] = 0 for all f, g ∈ B.
5. Orientability.
Whenever the metric dimension d is an integer, there is a Hochschild d-cycle c on B
with values in B⊗Bo, i.e. a finite sum of terms (a0⊗b0)⊗a1⊗· · ·⊗ad. Its representation
πD(c) with πD((a0⊗b0)⊗a1⊗· · ·⊗ad) := π(a0)Jπ(b∗
0)J −1[D, π(a1)]· · · [D, π(ad)] satisfies
πD(c)2 = 1 and defines the volume form on A, i.e.
φc(f0, . . . , fd) = Trω(cid:0)πD(c)π(f0)[D, π(f1)]· · · [D, π(fd)]hDi−d(cid:1)
provides a non-vanishing Hochschild d-cocycle φd on A.
Acknowledgments
R.W. would like to thank Harald Grosse for the long-term collaboration which initiated this
paper through the preprint [21]. Both authors would like to thank Alan Carey for stimulating
discussions.
References
[1] A. Carey, V. Gayral, A. Rennie and F. Sukochev, "Integration on locally compact noncommu-
tative spaces," [arXiv:math.OA/0912.2817].
[2] A. Carey, V. Gayral, A. Rennie and F. Sukochev, "Index theory for locally compact noncom-
mutative geometries," [arXiv:math.OA/1107.0805].
[3] A. H. Chamseddine and A. Connes, "The spectral action principle," Commun. Math. Phys.
186 (1997) 731.
[4] A. H. Chamseddine and A. Connes, "Scale invariance in the spectral action," J. Math. Phys.
47 (2006) 063504.
[5] A. Connes, "Gravity coupled with matter and the foundation of non-commutative geometry,"
Commun. Math. Phys. 182 (1996) 155.
[6] A. Connes, "Noncommutative geometry and the standard model with neutrino mixing," JHEP
0611 (2006) 081.
[7] A. Connes, "On the spectral characterization of manifolds," [arXiv:math.OA/0810.2088].
[8] A. Connes and J. Lott, "Particle models and noncommutative geometry (expanded version),"
Nucl. Phys. Proc. Suppl. 18B (1991) 29.
[9] L. D¸abrowski and A. Sitarz, "Dirac Operator on the Standard Podle´s Quantum Sphere,"
Center Banach Center Publ. 61 (2003) 49.
[10] M. Disertori and V. Rivasseau, "Two and three loops beta function of noncommutative φ4
4
theory," Eur. Phys. J. C 50 (2007) 661.
[11] M. Disertori, R. Gurau, J. Magnen and V. Rivasseau, "Vanishing of beta function of noncom-
mutative φ4
4 theory to all orders," Phys. Lett. B 649 (2007) 95.
[12] V. Gayral, J. M. Gracia-Bond´ıa, B. Iochum, T. Schucker and J. C. V´arilly, "Moyal planes are
spectral triples," Commun. Math. Phys. 246 (2004) 569.
36
[13] V. Gayral and B. Iochum, "The spectral action for Moyal planes," J. Math. Phys. 46 (2005)
043503.
[14] A. de Goursac, J. C. Wallet and R. Wulkenhaar, "Noncommutative induced gauge theory,"
Eur. Phys. J. C 51 (2007) 977.
[15] J. M. Gracia-Bond´ıa and J. C. V´arilly, "Algebras of distributions suitable for phase-space
quantum mechanics I," J. Math. Phys. 29 (1988) 869.
[16] H. Grosse and M. Wohlgenannt, "Induced Gauge Theory on a Noncommutative Space," Eur.
Phys. J. C 52 (2007) 435.
[17] H. Grosse and R. Wulkenhaar, "The beta-function in duality-covariant noncommutative φ4-
theory," Eur. Phys. J. C 35 (2004) 277.
[18] H. Grosse and R. Wulkenhaar, "Power-counting theorem for non-local matrix models and
renormalisation," Commun. Math. Phys. 254 (2005) 91.
[19] H. Grosse and R. Wulkenhaar, "Renormalisation of φ4-theory on noncommutative R4 in the
matrix base," Commun. Math. Phys. 256 (2005) 305.
[20] H. Grosse and R. Wulkenhaar, "Renormalisation of φ4-theory on non-commutative R4 to all
orders," Lett. Math. Phys. 71 (2005) 13.
[21] H. Grosse and R. Wulkenhaar, "8D-spectral triple on 4D-Moyal space and the vacuum of
noncommutative gauge theory," [arXiv:hep-th/0709.0095].
[22] H. Grosse and R. Wulkenhaar, "Progress in solving a noncommutative quantum field theory
in four dimensions," [arXiv:hep-th/0909.1389].
[23] R. Gurau, J. Magnen, V. Rivasseau and F. Vignes-Tourneret, "Renormalization of non-
commutative φ4
4 field theory in x-space," Commun. Math. Phys. 267 (2006) 515.
[24] R. Gurau and V. Rivasseau, "Parametric representation of noncommutative field theory,"
Commun. Math. Phys. 272 (2007) 811.
[25] J. Magnen and V. Rivasseau, "Constructive φ4 field theory without tears," Ann. Henri Poincar´e
9 (2008) 403.
[26] A. Rennie, "Smoothness and locality for nonunital spectral triples," K-theory 28 (2003) 127.
[27] V. Rivasseau, "Non-commutative renormalization," in Quantum spaces, Prog. Math. Phys. 53
(2007), Birkhauser, Basel.
[28] V. Rivasseau, "Constructive Matrix Theory," J. High Energy Phys. 9 (2007) 008.
[29] V. Rivasseau, F. Vignes-Tourneret and R. Wulkenhaar, "Renormalization of noncommutative
φ4-theory by multi-scale analysis," Commun. Math. Phys. 262 (2006) 565.
[30] B. Simon, Trace Ideals and Their Applications, London Math. Soc. Lecture Notes 35, Cam-
bridge University Press, Cambridge, 1979.
[31] R. Wulkenhaar, "Non-compact spectral triples with finite volume," in Quanta of Maths, Clay
Math. Proc. 11 (2010) 617, Amer. Math. Soc., Providence, RI.
37
|
1806.00410 | 5 | 1806 | 2019-07-01T19:16:13 | Algebras of noncommutative functions on subvarieties of the noncommutative ball: the bounded and completely bounded isomorphism problem | [
"math.OA"
] | Given a noncommutative (nc) variety $\mathfrak{V}$ in the nc unit ball $\mathfrak{B}_d$, we consider the algebra $H^\infty(\mathfrak{V})$ of bounded nc holomorphic functions on $\mathfrak{V}$. We investigate the problem of when two algebras $H^\infty(\mathfrak{V})$ and $H^\infty(\mathfrak{W})$ are isomorphic. We prove that these algebras are weak-$*$ continuously isomorphic if and only if there is an nc biholomorphism $G : \widetilde{\mathfrak{W}} \to \widetilde{\mathfrak{V}}$ between the similarity envelopes that is bi-Lipschitz with respect to the free pseudo-hyperbolic metric. Moreover, such an isomorphism always has the form $f \mapsto f \circ G$, where $G$ is an nc biholomorphism. These results also shed some new light on automorphisms of the noncommutative analytic Toeplitz algebras $H^\infty(\mathfrak{B}_d)$ studied by Davidson--Pitts and by Popescu. In particular, we find that $\operatorname{Aut}(H^\infty(\mathfrak{B}_d))$ is a proper subgroup of $\operatorname{Aut}(\widetilde{\mathfrak{B}}_d)$.
When $d<\infty$ and the varieties are homogeneous, we remove the weak-$*$ continuity assumption, showing that two such algebras are boundedly isomorphic if and only if there is a bi-Lipschitz nc biholomorphism between the similarity envelopes of the nc varieties. We provide two proofs. In the noncommutative setting, our main tool is the noncommutative spectral radius, about which we prove several new results. In the free commutative case, we use a new free commutative Nullstellensatz that allows us to bootstrap techniques from the fully commutative case. | math.OA | math | ALGEBRAS OF NONCOMMUTATIVE FUNCTIONS ON SUBVARIETIES OF THE
NONCOMMUTATIVE BALL: THE BOUNDED AND COMPLETELY BOUNDED
ISOMORPHISM PROBLEM
GUY SALOMON, ORR M. SHALIT, AND ELI SHAMOVICH
ABSTRACT. Given a noncommutative (nc) variety V in the nc unit ball Bd, we consider the
algebra H∞(V) of bounded nc holomorphic functions on V. We investigate the problem
of when two algebras H∞(V) and H∞(W) are isomorphic. We prove that these algebras
are weak-∗ continuously isomorphic if and only if there is an nc biholomorphism G :(cid:102)W →
(cid:101)V between the similarity envelopes that is bi-Lipschitz with respect to the free pseudo-
Popescu. In particular, we find that Aut(H∞(Bd)) is a proper subgroup of Aut((cid:101)Bd).
hyperbolic metric. Moreover, such an isomorphism always has the form f (cid:55)→ f ◦ G, where
G is an nc biholomorphism. These results also shed some new light on automorphisms of
the noncommutative analytic Toeplitz algebras H∞(Bd) studied by Davidson -- Pitts and by
When d < ∞ and the varieties are homogeneous, we remove the weak-∗ continuity
assumption, showing that two such algebras are boundedly isomorphic if and only if there
is a bi-Lipschitz nc biholomorphism between the similarity envelopes of the nc varieties. We
provide two proofs. In the noncommutative setting, our main tool is the noncommutative
spectral radius, about which we prove several new results. In the free commutative case,
we use a new free commutative Nullstellensatz that allows us to bootstrap techniques from
the fully commutative case.
9
1
0
2
l
u
J
1
]
.
A
O
h
t
a
m
[
5
v
0
1
4
0
0
.
6
0
8
1
:
v
i
X
r
a
1. INTRODUCTION
The study of holomorphic functions in one and several complex variables is an old and
well developed subject with countless applications. One fruitful venue of research is the
interplay between complex analysis and operator algebras, exemplified in the Sz. Nagy-
Foias bounded analytic functional calculus [47] and in Taylor's functional calculus of sev-
eral commuting operators [48]. The main limitation of this approach lies in the fact that
in general operators do not commute. This led Taylor and Voiculescu to study analytic
functions of several noncommuting variables [49, 50, 53 -- 56].
In fact, classical analytic
functions can be viewed as shadows of their noncommutative (nc, for short) counterparts,
the so-called nc holomorphic functions, under an appropriate quotient map. We will defer
the formal definition of nc holomorphic functions to the next section. For now it suffices
to view them as a generalization of polynomials in several noncommuting variables, i.e,
elements of a free associative algebra.
A natural choice for the domain of nc holomorphic functions in d noncommuting vari-
ables is the nc universe Md := (cid:116)∞
n=1Mn(C)d, the graded set of all d-tuples of complex square
matrices. One can view the matrix levels as capturing the noncommutative nature of our
2010 Mathematics Subject Classification. 47L80, 46L07,47L25.
The first author was partially supported by the Clore Foundation. The second author was partially sup-
ported by Israel Science Foundation Grant no. 195/16. The third author was partially supported by the
Fields Institute for Research in the Mathematical Sciences.
1
d < 1.
1 + ··· + XdX∗
Ω ⊆ Md, its similarity orbit(cid:101)Ω, which is just the orbit of Ω under the levelwise GLn-action.
functions, in analogy with Kaplanski's theorem [25, Theorem 2] that states that elements
of the free associative algebra C(cid:104)z1, . . . , zd(cid:105) are determined by their values on Md.
Not surprisingly, however, Md is in a sense too big to have a rich theory of holomorphic
functions, so just like in the classical case, analysts usually consider only certain subsets of
it. Every classical domain, such as a ball or a polydisc admits natural "quantizations". In
particular, in this paper we will focus on the nc ball Bd: the set of all d-tuples (X1, . . . Xd) ∈
Md satisfying X1X∗
It is worth noting that for every n, the nth level of the nc universe admins a natural GLn-
action given by S · (X1, . . . , Xd) = (S−1X1S, . . . , S−1XdS). Unfortunately, most domains,
including our nc ball, are not invariant under this action. We therefore define, for every set
The algebra of bounded nc holomorphic functions on the nc ball, H∞(Bd) turns out to
be the free semigroup algebra Ld studied by Arias and Popescu and Davidson and Pitts,
see for example [3, 11, 12, 30, 34, 38]. The free semigroup algebra is the universal weak-∗
closed algebra generated by a pure row contraction. Its quotients by weak-∗ closed two
sided ideals are thus universal weak-∗ closed algebras generated by pure row contractions
satisfying prescribed algebraic relations. We would like to understand when such algebras
are isomorphic. Though isomorphism can be understood in many ways we will focus on
continuous and completely bounded isomorphisms. Such a question of course begs the
introduction of an invariant. An immediate candidate for an invariant is the vanishing
locus inside the nc ball of a weak-∗ closed two-sided ideal of H∞(Bd). We will call these
subsets V ⊆ Bd nc varieties. Each variety V comes equipped with its algebra of functions
H∞(V), namely the quotient of the free semigroup algebra by the ideal of functions that
vanish on the variety.
It turns out that the information contained in our nc subvariety of the ball is not at
all sufficient to answer the (completely) bounded isomorphism question. The question
forces us to treat the geometry of (cid:101)V, the similarity envelope V. Several delicate issues
immediately arise. The first obstacle is the fact that the similarity envelope is, in general,
unbounded and thus classical results on bounded domains are not readily available. Fur-
thermore, there is an algebraic complication since similarity orbits of even single points
can be quite complicated.
The "geometry" required for the classification theorem, is encoded in certain pseudo-
metrics defined on the similarity envelope. More precisely, we define two free pseudo-
hyperbolic distances δb and δcb on the similarity envelope of the closed nc ball Bd that
measure the difference between point evaluations: the first in terms of the usual operator
norm and the second in terms of the complete bounded norm.
We can now state the classification theorem for homogeneous nc varieties (i.e. nc vari-
eties that are cut out by homogeneous nc functions).
Theorem 1.1 (Theorem 7.8 and Corollary 6.3). Let V ⊆ Bd and W ⊆ Be be two homoge-
neous nc varieties. The following statements are equivalent:
(a) H∞(V) and H∞(W) are weak-∗ continuously isomorphic.
(b) H∞(V) and H∞(W) are boundedly isomorphic.
(c) H∞(V) and H∞(W) are completely boundedly isomorphic.
(d) There exists a δb-bi-Lipschitz nc biholomorphism mapping(cid:102)W onto (cid:101)V.
2
(e) There exists a δcb-bi-Lipschitz nc biholomorphism mapping(cid:102)W onto (cid:101)V.
(f) There exists a δb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
(g) There exists a δcb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
In addition, any isomorphism that appears in (a) -- (c) can be viewd as a precompostion with
a δcb-bi-Lipschitz nc biholomorphism between the similarity envelopes.
In the not necessarily homogeneous case, we lose the convenience of having a linear
map -- meaning, we no longer have (f) and (g) -- and the rest of the equivalence list
splits into two as follows: (a), (a)+(b), and (d) are equivalent, and (a)+(c) and (e) are
equivalent (see Corollary 6.3).
One may rightly ask what about classifying these algebras up to an isometric or com-
pletely isometric isomorphism. This question is somewhat easier to attack.
In [43] we
showed that when it comes to homogeneous varieties such an isomorphism exists if and
only if one of the varieties is the image of the other, under an nc automorphism of the nc
ball Bmax{d,e}, and any such isomorphism is given by a precomposition with such an nc
automorphism. The group of nc automorphisms of the nc ball is the well-known group
of Mobius transformations [38], so the class of isometric or completely isometric isomor-
phisms between the algebras H∞(V) and H∞(W) is somewhat poor.1
The situation in this paper is markedly different. For example, while the nc automor-
phism group of the nc ball is crystal-clear, the group of bi-Lipschitz nc automorphisms of
the similarity envelope of the nc ball is far from being well understood. In fact, Theorem
1.1 together with the fact that algebraic automorphisms of Ld are automatically weak-∗
continuous [11, Theorem 4.6] imply that it can be identified with the group of algebraic
∼= H∞(Bd), which is mysterious in many ways (for ex-
automorphisms of the algebra Ld
ample, it is not clear whether there are quasi-inner automorphisms which are not inner).
Some of the main tools that are developed to obtain Theorem 1.1 are interesting in
their own right. For example, in Section 4 we prove an nc counterpart of the well-known
Schwarz Lemma of the disc.
Theorem 1.2 (Lemma 4.9 and Proposition 4.14). Let f : D → (cid:102)Bd be a holomorphic function
mapping 0 to 0, and let ρ denote the joint spectral radius of a d-tuple of matrices. Then
(a) ρ(f (z)) ≤ z for every z ∈ D and ρ(f(cid:48)(0)) ≤ 1; and
(b) if f(cid:48)(0) is an irreducible coisometry, then f (z) is similar to zf(cid:48)(0) for very z ∈ D.
The case where the varieties V and W contain only commuting tuples is of special inter-
est. The isomorphism problem in the radical commutative case was treated by Davidson,
Ramsey and the second author in [13] and [14]. We develop new machinery to deal with
such "commutative noncommutative" varieties even in the non-reduced case, which also
gives rise to a different proof of Theorem 1.1, this time with a commutative flavor. One
main ingredient in this machinery is a certain type of a Nullstellensatz.
Theorem 1.3 (Theorem 9.7). Let V ⊆ Bd be a homogeneous nc variety containing only
commuting d-tuples and let V = V(1) be the scalar level of V. Then there exists an integer
1In [43], we also examined the nonhomogeneous case, and we showed that these algebras are completely
isometrically isomorphic if and only if the varieties V and W are nc biholomorphic. The main result of [45]
then shows that, at least when the varieties contain a scalar point, such an nc biholomorphism is just a
restriction of an nc automorphism of the nc ball.
3
N such that for every nc function f ∈ H∞(Bd) that vanishes on V , the Nth power f N of f
vanishes on the whole nc variety V.
At the end of this paper we turn to another operator algebra related to an nc variety V:
the algebra A(V) of all bounded analytic functions that extend to uniformly continuous
functions on V. We prove the following analog of Theorem 1.1.
Theorem 1.4 (Theorems 10.2 and 10.3). Let V ⊆ Bd and W ⊆ Be be two homogeneous nc
varieties. The following statements are equivalent:
(a) A(V) and A(W) are boundedly isomorphic.
(b) A(V) and A(W) are completely boundedly isomorphic.
(c) There exists a δb-bi-Lipschitz nc biholomorphism mapping(cid:102)W onto (cid:101)V.
(d) There exists a δcb-bi-Lipschitz nc biholomorphism mapping(cid:102)W onto (cid:101)V.
(e) There exists a δb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
(f) There exists a δcb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
In addition, any isomorphism that appears in (a) -- (b) can be viewd as a precompostion with
a δb-bi-Lipschitz or a δcb-bi-Lipschitz nc biholomorphism mapping one similarity envelope of
the variety's clousre onto the other.
We note that in order to treat the algebras A(V) for nonhomogeneous varieties one
would have to face some nontrivial nc-function-theoretic issues, which are beyond the
scope of this paper.
Besides the introduction, this paper consists of nine sections, and we shall now describe
their content. We start with a brief digression in which we treat the purely algebraic case,
which quickly illustrates the utillity of the nc point of view. In Section 2 we prove algebraic
analogues of Theorems 1.1 and 1.4, which we believe are interesting in their own right.
preliminaries. In Section 4 we give two alternative descriptions of (cid:102)Bd and present the nc
In Section 3 we lay the foundations of the rest of the paper by providing the necessary
counterpart of the well-known Schwartz lemma of the disc as presented in Theorem 1.2.
Section 5 is dedicated to two pseudo-hyperbolic distances on the similarity envelope of the
free ball, and in Section 6 we use these distances to state and prove the isomorphism theo-
rem for general nc varieties. In Section 7 we specialise to the homogeneous case, in which
we obtain the sharper results described in Theorem 1.1. Section 8 contains a detailed
study of the family of nc varieties determined by the q-commutation relations. In Section
9 the results of Section 7 are derived again, this time in the case of commutative varieties,
using commutative techniques that includes the Nullstellensatz presented in Theorem 1.3.
Then, in the las section, Section 10, we turn our attention to the norm closed (instead
of WOT-closed) algebras generated by the free polynomial functions on homogeneous nc
varieties and prove Theorem 1.4.
2. A DIGRESSION -- THE PURELY ALGEBRAIC CASE
As motivation for our main investigations, we consider the purely algebraic analogues
of our problems. Let C[z] := C[z1, . . . , zd] denote the algebra of complex polynomials in d
commuting variables (here d < ∞). With every ideal I (cid:47) C[z] one naturally associates the
4
corresponding affine variety
Z(I) = ZCd(I) = {z ∈ Cd : p(z) = 0 for all p ∈ I}.
Together with this geometric object, there are two natural algebraic objects: the quotient
C[z]/I -- which is the universal unital algebra generated by d commuting elements satis-
fying the relations in I -- and the algebra of regular functions:
(cid:110)
(cid:111)
p(cid:12)(cid:12)Z(I) : p ∈ C[z]
.
C[Z(I)] =
Consider two ideals I, J (cid:47) C[z]. One may ask when are the quotients C[z]/I and C[z]/J
isomorphic, as algebras. When I and J are radical, then it follows from Hilbert's Nullstel-
lensatz that C[z]/I ∼= C[Z(I)] and C[z]/J ∼= C[Z(J)], and it is then not hard to show that
C[z]/I and C[z]/J are isomorphic if and only if there exist polynomial maps F, G : Cd → Cd
that restrict to mutually inverse bijections between Z(J) and Z(I).
What happens if I and J are not radical ideals? The concrete and geometric object
Z(I) is no longer a complete invariant for the quotient algebra C[z]/I. Algebraic geometry
offers some elaborate but opaque "geometric" replacements for the variety. A more simple-
minded (and perhaps more satisfying) alternative is suggested to us by nc function theory.
We can consider C[z] as an algebra of nc functions on CMd (recall that CMd is the set of
all commuting d-tuples of matrices, of all sizes). Given an ideal I (cid:47) C[z], let
ZCMd(I) = {X ∈ CMd : p(X) = 0 for all p ∈ I}.
Points in ZCMd(I) correspond bijectively to all finite dimensional representations of C[z]/I,
via the map π that sends every finite dimensional representation ρ to its image on the
coordinate functions in the quotient:
π(ρ) = (ρ(z1 + I), . . . , ρ(zd + I)).
The inverse of π is given by
for all X ∈ ZCMd(I), where ρX is evaluation at X:
π−1 : X (cid:55)→ ρX,
ρX(p + I) = p(X).
Suppose we are given a homomorphism α : C[z]/I → C[z]/J. Then α gives rise to a map
between the spaces of representations, by α∗ : ρ (cid:55)→ ρ ◦ α. Now, as
(α(z1 + I), . . . , α(zd + I)) ∈ C[z]/J × ··· × C[z]/J,
there exists F = (F1, . . . , Fd) ∈ C[z] × ··· × C[z] such that α(zi + I) = Fi + J for every i.
Then, if X ∈ ZCMd(J),
F (X) = (ρX(F1 + J), . . . , ρX(Fd + J))
= (ρX(α(z1 + I)), . . . , ρX(α(zd + I)))
= (α∗(ρX)(z1 + I), . . . , α∗(ρX)(zd + I))
= π(α∗(ρX)).
Thus, we see that a homomorphism α : C[z]/I → C[z]/J gives rise to a polynomial map
F ∈ C[z]d mapping ZCMd(J) into ZCMd(I).
5
On the other hand, suppose we are given a polynomial map that restricts to a map from
ZCMd(J) into ZCMd(I). Let
(cid:110)
p(cid:12)(cid:12)ZCM
(cid:111)
.
C[ZCMd(I)] =
(I) : p ∈ C[z]
d
Then F clearly gives rise, via pre-composition, to a homomorphism from C[ZCMd(I)] to
C[ZCMd(J)].
We therefore see that a homomorphism C[z]/I → C[z]/J always gives rise to a polyno-
mial map that restricts to a map from ZCMd(J) into ZCMd(I), and such a map always gives
rise to a homomorphism C[ZCMd(I)] → C[ZCMd(J)]. To close the loop, we need a link, a
Nullstellensatz, between the quotient C[z]/I and the function algebra C[ZCMd(I)].
Given a set S ⊆ CMd, let
I(S) = IC[z](S) = {p ∈ C[z] : p(X) = 0 for all X ∈ S}.
In [43, Corollary 11.7] we obtained that I(ZCMd(J)) = J for every J (cid:47) C[z]. We call this
the commutative free Nullstellensatz, and it has been known to algebraists in one form or
another (see [16]). The commutative free Nullstellensatz implies at once that C[z]/I is
isomorphic to the function algebra C[ZCMd(I)].
This shows, additionally, that homomorphisms α : C[ZCMd(I)] → C[ZCMd(J)] are neces-
sarily pre-composition with a polynomial map F mapping ZCMd(J) into ZCMd(I). Indeed,
after identifying C[z]/I ∼= C[ZCMd(I)], we saw before that the relation between α and F is
given by F (X) = π(α∗(ρX)) for all X ∈ ZCMd(J). Applying π−1 to this equality, we obtain
that α∗(ρX) = ρF (X), and therefore
α(p) = p ◦ F ,
for all p ∈ C[ZCMd(I)].
We summarize the conclusion of the above discussion, in the case of an isomorphism, as
follows.
Theorem 2.1. Let I and J be two ideals in C[z]. The algebras C[z]/I and C[z]/J are iso-
morphic if and only if ZCMd(J) and ZCMd(I) are isomorphic, in the sense that there exists
polynomial maps F and G that restrict to bijections between ZCMd(J) and ZCMd(I). More-
over, every homomorphism from C[z]/I to C[z]/J is implemented by a polynomial map
F : ZCMd(J) → ZCMd(I).
One can consider ideals inside the algebra C(cid:104)z(cid:105) := C(cid:104)z1, . . . , zd(cid:105) of free polynomials
in d noncommuting variables, and given such an ideal I (cid:47) C(cid:104)z(cid:105), one can consider the
noncommutative variety
ZMd(I) = {X ∈ Md : p(X) = 0 for all p ∈ I}.
If I is a homogeneous ideal, then there is an appropriate noncommutative homogeneous
Nullstellensatz [43, Theorem 7.3], which says that (with obvious notation)
IC(cid:104)z(cid:105)(ZMd(J)) = J for every homogeneous J (cid:47) C(cid:104)z(cid:105).
If one replaces the commutative free Nullstellensatz with the noncommutative homoge-
neous Nullstellensatz, then the same argument as above (where polynomials are replaced
by free polynomials) gives the following theorem:
6
Theorem 2.2. Let I and J be two homogeneous ideals in C(cid:104)z(cid:105). The algebras C(cid:104)z(cid:105)/I and
C(cid:104)z(cid:105)/J are isomorphic if and only if ZMd(J) and ZMd(I) are isomorphic, in the sense that there
exists free polynomial maps F and G that restrict to bijections between ZMd(J) and ZMd(I).
Moreover, every homomorphism C(cid:104)z(cid:105)/I and C(cid:104)z(cid:105)/J is implemented by a free polynomial map
F : ZMd(J) → ZMd(I).
Our main goal in the remainder of this paper is to understand the analogue of the above
results for algebras of bounded analytic nc functions. As such algebras are not finitely
generated in an algebraic sense by the coordinate functions, there are interesting technical
issues to overcome.
In passing, we are happy to note that it was by considering the operator algebraic prob-
lems that the above purely algebraic results crystallized for us, and they seem to have been
overlooked. It is worth noting that Theorem 2.1 can be restated in the free setting as a the-
orem on ideals I, J of C(cid:104)z(cid:105) that contain the commutant ideal. This point of view, together
with Theorem 2.2, suggests that there might be a general theorem regarding any pair of
ideals I, J (cid:47)C(cid:104)z(cid:105). Such a generalization, however, fails to be true (consider the trivial ideal
and the ideal generated by the nc polynomial z1z2 − z2z1 − 1).
3. PRELIMINARIES
3.1. Nc functions and nc varieties. We study noncommutative (nc) function theory in d
complex variables, where d ∈ N or d = ∞. Let Mn = Mn(C) denote the set of all n × n
n be the set of all d-tuples X = (X1, X2, . . .) of n × n matrices,
matrices over C, and let M d
such that the row X determines a bounded operator from Cn ⊕ Cn ⊕ . . . to Cn. We norm
n with the induced
M d
topology. We define (the nc universe)
n with the row operator norm (cid:107)X(cid:107) = (cid:107)(cid:80)
j (cid:107)1/2, and endow M d
j XjX∗
Md = (cid:116)∞
n=1M d
n,
and (the commutative nc universe)
CMd = {X ∈ Md : XiXj = XjXi
for all i, j}.
A set Ω ⊆ Md is said to be an nc set if it is closed under direct sums.
If Ω is an nc
set, we denote Ω(n) = Ω ∩ M d
n. A subset Ω ⊆ Md is said to be open/closed if for all n,
Ω(n) is open/closed. This collection of open sets gives rise to a topology on Md and its
subsets, called the disjoint union topology. The boundary of Ω, denoted ∂Ω, is defined to
be (cid:116)∞
n=1∂Ω(n). The principle nc open set that we shall consider is the (d-dimensional) open
matrix unit ball Bd, which is defined to be
an nc function (with values in V) if
Let V be a vector space. A function f from an nc set Ω ⊆ Md to (cid:116)∞
(1) f is graded: X ∈ Ω(n) ⇒ f (X) ∈ Mn(V);
(2) f respects direct sums: f (X ⊕ Y ) = f (X) ⊕ f (Y ); and
(3) f respects similarities: if X ∈ Ω(n) and S ∈ Mn is invertible, and if S−1XS ∈ Ω(n),
n=1Mn(V) is said to be
then f (S−1XS) = S−1f (X)S.
7
(cid:110)
Bd =
X ∈ Md : (cid:107)X(cid:107)2 =
(cid:13)(cid:13)(cid:13)(cid:88)
(cid:13)(cid:13)(cid:13) < 1
(cid:111)
.
XjX∗
j
An nc function with values in C is said to be a scalar-valued nc function. In this paper, we
shall deal only with scalar-valued nc functions.
A function f defined on an nc open set Ω is said to be nc holomorphic if it is an nc function
It turns out that an nc holomorphic function is
and, in addition, it is locally bounded.
really a holomorphic function when considered as a function f : Ω(n) → Mn, for all n, and
moreover it has a "Taylor series" at every point (see [24]). We therefore allow ourselves
to use the terms holomorphic and analytic interchangeably.
We let
H∞(Bd) = {f : Bd → Md : f is a bounded nc function} ,
denote the algebra of all bounded nc holomorphic functions on the nc unit ball. This
algebra (with the sup norm) can be shown to be the algebra of multipliers of the noncom-
mutative Drury-Arveson space, and as an operator algebra it is unitarily equivalent to the
noncommutative analytic Toeplitz algebra studied by Davidson and Pitts, which was also
studied by Popescu under the name noncommutative Hardy algebra (see [43, Section 3]
and [11, 12, 32, 34]).
A noncommutative (nc) variety in the unit ball is a set of the form
VBd(S) = {X ∈ Bd : f (X) = 0 for all f ∈ S},
where S is a set of bounded nc holomorphic functions. We emphasize that although it
makes sense to consider varieties determined by arbitrary nc holomorphic functions, such
generality is beyond our scope, and we will assume that every variety is given by S ⊆
H∞(Bd). Note that nc varieties are nc sets.
A free polynomial is an element in C(cid:104)z1, . . . , zd(cid:105) (the free algebra in d noncommuting
variables). Let Wd be the free monoid on d generators {g1, . . . , gd}. If k = gi1 ··· gin ∈ Wd
(in which case we write k = n), we define the free monomial zk = zi1 ··· zin. A polynomial
akzk can be written in a unique way as p(z) =(cid:80)
p(z) =(cid:80)
n∈N pn(z) where
k∈Wd
(cid:88)
k∈Wd,k=n
pn(z) =
akzk.
The polynomial pn is called the homogeneous component of degree n of p. Every free poly-
nomial is a (scalar-valued) nc function on Md in a natural way, by evaluation; moreover,
this nc function is bounded on every uniformly bounded subset of Md. The inclusion of
C(cid:104)z1, . . . , zd(cid:105) in H∞(Bd) is injective (this might not be entirely obvious, but follows from
the known fact the matrix algebra Mn satisfies no polynomial identity of degree less than
2n; see [2]).
Given an nc variety V ⊆ Bd, we define H∞(V) to be the algebra of bounded nc func-
tions on V, and A(V) to be the algebra of bounded nc functions that extend to uniformly
continuous functions on V (the closure is taken in the disjoint union topology). We
give H∞(V) and A(V) the obvious operator algebra structure, where the matrix norm
of F ∈ Mn(H∞(V)) is given by
(cid:107)F(cid:107) = sup
z∈V
(cid:107)F (z)(cid:107) = sup
k∈N
sup
z∈V(k)
(cid:107)F (z)(cid:107)Mnk.
It follows from [5, Corollary 5.6] (see also [43, Theorem 4.7]) that if Ω = Bd, and if
f : V → M1 is an nc function that is bounded on V, then there exists a bounded nc
holomorphic function F on Bd such that (cid:107)F(cid:107) = (cid:107)f(cid:107) and f = F(cid:12)(cid:12)V.
8
The algebra H∞(V) can be identified with the multiplier algebra of an nc reproducing
kernel Hilbert space [43, Theorem 5.4]. Given an nc variety V, we define
IV = {f ∈ A(Bd) : f (Z) = 0 for all Z ∈ V}
and
JV = {f ∈ H∞(Bd) : f (Z) = 0 for all Z ∈ V}.
Then we have that H∞(V) is completely isometrically isomorphic to H∞(Bd)/JV [43,
Theorems 5.2 and 5.4], and when V is homogeneous, A(V) is completely isometrically
isomorphic to A(Bd)/IV [43, Proposition 9.7].
3.2. The nc Szego kernel and the nc Drury -- Arveson space. The nc Szego kernel K(Z, W )
on the nc ball Bd of Md defined by
K(Z, W )(T ) =
Z kT W ∗k,
k∈Wd
for Z ∈ Bd(n), W ∈ Bd(m) and T ∈ Mn×m(C) [5].
Consider the nc function KW,v,y for W ∈ Bd(m), v, y ∈ Cm, defined for Z ∈ Bd(n):
KW,v,y(Z)u = K(Z, W )(uv∗)y =
Z kuv∗W ∗ky =
(cid:104)y, W kv(cid:105)Z ku.
(cid:88)
k∈Wd
Thus KW,v,y is an nc function given by the power series
KW,v,y(Z) =
(cid:104)y, W kv(cid:105)Z k.
(cid:88)
(cid:88)
(cid:88)
k∈Wd
k∈Wd
In [43, Section 3] we saw that MultH2
The nc Drury -- Arveson space H2
d is the nc reproducing kernel Hilbert space determined by
the nc Szego kernel K, in the sense of [6].
d =
H∞(Bd). Moreover, we observed there that H∞(Bd) can be naturally identified with the
noncommutative analytic Toeplitz algebra Ld that Davidson and Pitts treated in [11], or,
what is the same, Popescu's noncommutative Hardy algebra [31], which he denotes by
F ∞
d .
Recall the Bunce -- Frazho -- Popescu dilation theorem [8, 19, 30], which says that if T =
(T1, . . . , Td) is a pure row contraction on a Hilbert space H, then there is a Hilbert space
E of dimension d, an auxiliary Hilbert space D, and an isometry V : H → F(E) ⊗ D such
that V H is a co-invariant subspace for the free shift L ⊗ ID, and such that
Ti = V ∗(Li ⊗ ID)V
,
i = 1, 2, . . . , d.
Identifying H∞(Bd) with Ld, this gives rise to a functional calculus: for every pure row
contraction T , there is a weak-operator continuous, unital, completely contractive homo-
morphism
given by ΦT (f ) = V ∗(f (L) ⊗ ID)V (where f (L) is the image of f in Ld under the isomor-
phism H∞(Bd) ∼= Ld). If T is a strict contraction ((cid:107)T(cid:107) := (cid:107)T(cid:107)row < 1) then it is not hard
to see that ΦT becomes the evaluation at T , that is
ΦT : H∞(Bd) → alg
wot
(T ),
(cid:32)(cid:88)
k∈Wd
ΦT
(cid:33)
akzk
=
9
(cid:88)
k∈Wd
akT k.
This gives rise to a functional calculus for multiplier algebras on nc varieties, versions of
which were observed in [35, 44].
In a previous work [43, Corollary 2.6], we showed that when Ω ⊆ Md is an nc set
and k is a completely positive nc kernel, then the nc multiplier algebra M(k) of the nc
RKHS H(k) associated to k is weak-∗ closed and is therefore a dual algebra. In addition,
to every nc subvariety V ⊆ Bd we associated an nc RKHS H2
V -- a subspace of the nc
Drury -- Arveson space H2
V, and we showed that
MultH2
V is completely isometrically isomorphic to H∞(V) (this relies on the deep result --
originally due to Agler and McCarthy [1, Theorem 1.5] for an algebraic variety, and later
extended to a much general setting by Ball, Marx and Vinnikov [5, Theorem 3.1] -- that
every bounded nc function on V extends to a bounded nc function of Bd with the same
norm); see [43, Theorem 5.4]. Thus, H∞(V) has a natural weak-∗ topology.
Davidson and Pitts showed that when the variety V is the whole nc ball Bd, then the
weak-operator and weak-∗ topologies of H∞(V) = H∞(Bd) coincide [12]. Since for every
nc variety V ⊆ Bd the algebra H∞(V) is a quotient of H∞(Bd) by the WOT-closed ideal
d -- and its algebra of multipliers MultH2
JV = {f ∈ H∞(Bd) : f (Z) = 0 for all Z ∈ V},
the weak-operator and weak-∗ topologies coincide on H∞(V) as well. We record this for a
later use.
Theorem 3.1. Let V ⊆ Bd be an nc variety. Then the weak-operator and weak-∗ topologies
coincide on H∞(V).
As a corollary we obtained that for every nc subvariety V ⊆ Bd and every a pure row
contraction T satisfying JV ⊆ ker ΦT (in particular, if T ∈ V), there is a weak-operator
continuous, unital completely contractive homomorphism from MultH2
(T ) map-
ping Mz to T [43, Corollary 5.3] (see also [3]).
3.3. Bounded finite dimensional representations of H∞(V). Let Repk(A) denote the
space of all unital bounded representations of an operator algebra A on Ck, and let
Repcc
p denote the set of all pure T ∈ Bd such that JV ⊆ ker ΦT . The
completely contractive representations of H∞(V) were determined in [43, Theorem 6.3]:
Theorem 3.2. Let V ⊆ Bd be an nc variety. For very k ∈ N, there is a natural continuous
projection πk of Repcc
k (A) denote the completely contractive representations in Repk(A).
For V ⊆ Bd, we let V
k (H∞(V)) into the closed unit ball Bd(k), given by
V to alg
wot
πk(Φ) = (Φ(z1), . . . , Φ(zd)).
p, there is a unique weak-∗ continuous representation ΦT ∈ π−1
For every T ∈ V
these are the only weak-∗ continuous elements in Repcc
k (T ) is the singleton {ΦT}. Moreover, if d < ∞, then
π−1
k (T ), and
k (H∞(V)). If d < ∞ and T ∈ V, then
πk(Repcc
k (H∞(V))) ∩ Bd(k) = V(k).
We can now describe the bounded finite dimensional representations. All that is required
to pass from completely contractive to bounded representations, is to pass to the similarity
invariant envelope of the variety.
Lemma 3.3. (cid:102)V
p
= (cid:101)V for every nc variety V ⊆ Bd.
10
p is the set of all pure T ∈ Bd such that JV ⊆ ker ΦT . It suffices to
Proof. Recall that V
p, then X is similar to a strict row contraction Y ∈ V. By Proposition
show that if X ∈ V
4.3 below, since X is pure, it is similar to some strict row contraction Y . Being similar to
X, we see that Y annihilates every f ∈ JV, so Y ∈ V.
(cid:3)
Theorem 3.4. Let V ⊆ Bd be an nc variety. For very k ∈ N, there is a natural continuous
projection πk of Repk(H∞(V)) into the similarity invariant envelope (cid:102)Bd(k) of the closed unit
For every T ∈ (cid:102)V
these are the only weak-∗ continuous elements in Repk(H∞(V)). If d < ∞ and T ∈ (cid:101)V, then
k (T ) is the singleton {ΦT}. Moreover, if d < ∞, then
π−1
= (cid:101)V, there is a unique weak-∗ continuous representation ΦT ∈ π−1
πk(Φ) = (Φ(z1), . . . , Φ(zd)).
ball Bd(k), given by
k (T ), and
p
πk(Repk(H∞(V))) ∩ (cid:94)Bd(k) = (cid:93)V(k).
Proof. Let Φ ∈ Repk(H∞(V)). By a theorem of Smith, Φ, being a bounded map into Mn(C),
is completely bounded (see [29], pp. 113 -- 114). By a theorem of Paulsen [29, Theorem
9.1], there exists a completely contractive ρ ∈ Repcc
k (H∞(V)) and an invertible matrix S
such that Φ(·) = Sρ(·)S−1. By Theorem 3.2, πk(ρ) ∈ Bd(k). Thus,
πk(Φ) := (Φ(z1), . . . , Φ(zd)) = Sπk(ρ)S−1 ∈ (cid:102)Bd(k).
Moreover, ρ is weak-∗ continuous if and only if Φ is, so the rest follows from Theorem 3.2
(cid:3)
and Lemma 3.3.
Henceforth, we will identify (cid:101)V with the weak-∗ continuous finite dimensional represen-
tations of H∞(V), via the identification
T ↔ ΦT .
Question 3.5. Are finite dimensional representations automatically continuous?
4. THE SIMILARITY INVARIANT ENVELOPE AND THE JOINT SPECTRAL RADIUS
Recall that Md = (cid:116)∞
n=1M d
n, the set of all d tuples of all matrices, of all sizes. Unless we
(cid:101)Ω(n) = {SXS−1 : X ∈ Ω(n), S ∈ GLn(C)}.
make the explicit assumption that d is finite, we shall treat d ∈ N ∪ {∞}.
4.1. The similarity invariant envelope. Let Ω ⊆ Md. The similarity invariant envelope
(or simply the similarity envelope) of Ω is defined to be the set(cid:101)Ω given by
Clearly, if Ω is open, then so is(cid:101)Ω. In the appendix of [24], it is shown that(cid:101)Ω is an nc set if
Ω is, and that every nc function f on Ω extends uniquely to an nc function (cid:101)f on(cid:101)Ω.
Every f ∈ H∞(Bd) extends uniquely to an nc holomorphic function (cid:101)f on (cid:102)Bd, and like-
phic function (cid:101)f on (cid:101)V. Of course, (cid:101)f need not be bounded. The nc holomorphic functions on
(cid:102)Bd which are extensions of functions in H∞(Bd) are precisely those whose restriction to
wise, if V ⊆ Bd is an nc variety, then every f ∈ H∞(V) extends uniquely to an nc holomor-
Bd is bounded, or, equivalently, whose restriction to every set bounded in the completely
bounded norm, is bounded.
11
The similarity envelope of Bd will be of particular interest in this paper and thus we
will provide two descriptions of this set. For every n ∈ N and X ∈ Md(n) we have the
associated completely positive map on Mn, given by
d(cid:88)
j=1
ΨX(T ) =
XjT X∗
j .
A point is in Bd if and only if ΨX(I) ≤ I. Recall that a point X ∈ Md(n) is called pure if
X(I) = 0. We will first show that every pure point is in fact similar to a strict row
limk→∞ Ψk
contraction. To prove this claim we need two lemmas. The first lemma, Lemma 4.1, was
proved in greater generality by Popescu in [39, Theorem 3.8] (see also [33, Section 5]),
and so were Lemma 4.5 and Proposition 4.3; see Remark 4.6.
Lemma 4.1 (Theorem 3.8 of [39]). For every X ∈ Md, X is similar to a strict row contrac-
tion if and only if there exists a strictly positive A, such that ΨX(A) < A.
Proof. First let us assume that such an A exists. If S is the unique positive square root of
A, then by assumption we have
(cid:0)S−1XjS(cid:1)(cid:0)S−1XjS(cid:1)∗
d(cid:88)
j=1
= S−1ΨX(A)(S−1)∗ < I.
In other words S−1XS is a strict row contraction.
Conversely, if X is similar to a strict row contraction, then there exists S invertible, such
that S−1XS is a strict row contraction. Using the same consideration as above we deduce
that
ΨX(SS∗) < SS∗.
(cid:3)
Now set A = SS∗.
Lemma 4.2. Suppose that X ∈ Md(n) and Y ∈ Md(m) are similar to strict row contractions.
Then, for every d-tuple of n × m matrices Z we have that the point
is similar to a
strict row contraction.
Proof. If S−1XS and T −1Y T are strict row contractions, then by conjugating by S ⊕ T we
may assume that in fact X and Y are strict row contractions, since Z will be replaced by
S−1ZT .
(cid:20)X Z
Let t > 0 and consider the conjugation of our point by t−1In ⊕ tIm
0 Y
(cid:21)
(cid:21)(cid:20)X Z
(cid:21)(cid:20)t−1In
0
(cid:20)X t2Z
(cid:21)
(cid:21)
0
tIm
0
0
0
=
0 Y
t−1Im
= X⊕Y ∈ Bd(n+m). Since Bd(n+m) is open, there exists t > 0
Note that limt→0
such that t−1In⊕tIm implements the similarity of our matrix to a strict row contraction. (cid:3)
A d-tuple X = (X1, . . . , Xd) ∈ Md(n) is called irreducible (or sometimes generic) if the
only nontrivial subspace K satisfying alg(X1, . . . , Xd)K ⊆ K is Cd, or equivalently, if the
X1, . . . , Xd have no joint nontrivial proper invariant subspace. If X is not irreducible, then
it is called reducible.
Y
Y
12
(cid:20)tIn
(cid:21)
(cid:20)X t2Z
0
(cid:21)
(cid:20)X(cid:48)
It is not hard to see that every X ∈ Md(n) is similar to a block upper triangular d-tuple
whose diagonal blocks are all irreducible. Indeed, if X is reducible, let 0 (cid:40) K (cid:40) Cn be
a joint invariant subspace of the Xis. Then there exists an invertible matrix S such that
S−1XS =
where both X(cid:48) and X(cid:48)(cid:48) are of size smaller than that of X. An induction
argument now completes the proof.
Alternatively, since it is finite dimensional, the alg(X1, . . . , Xd)-module Cn is both Noe-
therian and Artinian. Such modules (i.e., Nothereian and Artinian modules) are exactly
the modules that have composition series [22, Proposition 3.2.2].
∗
0 X(cid:48)(cid:48)
One way or another, the Jordan -- Holder theorem [22, Theorem 3.2.1], implies that if
Y (1)
0
...
∗
Y (k)
and
Z (1)
0
...
∗
Z (l)
are two "decompositions" as above of X -- namely, both are similar to X and each of the
diagonal blocks is irreducible -- then k = l, and there exists a permutation σ of {1, . . . , k}
such that Y (i) is similar to Z (σ(i)) for all 1 ≤ i ≤ k. Thus, up to similarity Y (1), . . . , Y (k) are
unique. We call them the Jordan -- Holder components of X.
Proposition 4.3 (Theorem 3.8 of [39]). A point X ∈ Md is pure if and only if it is similar
to a strict row contraction.
Proof. Assume first that X is similar to a strict row contraction, i.e., there exists an invert-
ible S, such that S−1XS is a strict row contraction. Thus ΨS−1XS(I) < I and we conclude
that limk→∞ Ψk
S−1XS(I) = 0. One checks that for A = SS∗, we have the identity
S−1XS(I) = S−1Ψk
Ψk
X(A)(S−1)∗.
It follows that limk→∞ Ψk
that cI ≤ A and applying Ψk we conclude that limk→∞ Ψk
is pure.
X(A) = 0. Since A is strictly positive, we can choose c > 0, such
X(I) = 0 or in other words that X
Now assume that X is pure. First, let us assume that X is also irreducible. Let r denote
the spectral radius of ΨX. By [18, Theorem 2] the map ΨX is irreducible in the sense
of [17]. By Theorems 2.3 and 2.5 in [17], there is a strictly positive A ∈ Mn, such that
ΨX(A) = rA. Since ΨX is completely positive, (cid:107)Ψk
X = 0
in norm. We can conclude that r < 1, so ΨX(A) = rA < A. Thus by Lemma 4.1 we are
done in the case that X is irreducible.
X(I)(cid:107) and thus limk→∞ Ψk
To prove the general case we proceed by induction on the number of Jordan -- Holder
components of X. If the number is 1, then X is irreducible and we have proved it. Now
, where X(cid:48) is irreducible and X(cid:48)(cid:48) has
if the number is k, then we can write X =
k − 1 Jordan -- Holder components. Note that since X is pure, X(cid:48) and X(cid:48)(cid:48) must also be pure.
By induction, X(cid:48) and X(cid:48)(cid:48) are similar to strict row contractions. It remains to apply Lemma
(cid:3)
4.2.
4.2. The joint spectral radius. Let X ∈ Md(n). Given k ∈ N, write X (k) for the row
comprising of the free monomials of degree k in the entries of X. Following Popescu [37],
0 X(cid:48)(cid:48)
X(cid:107) = (cid:107)Ψk
(cid:21)
(cid:20)X(cid:48) Z
13
ρ(X) = lim
k→∞(cid:107)Ψk
X(I)(cid:107)1/2k.
we define the joint spectral radius of the d-tuple X as
X is a completely positive map we have that (cid:107)Ψk
X(I)(cid:107)1/k is the spectral radius of ΨX. Note that (cid:107)Ψk
Since ΨX (k) = Ψk
limk→∞ (cid:107)Ψk
particular this notion of spectral radius reduces to the usual one when d = 1.
X(cid:107) = (cid:107)Ψk
X(I)(cid:107) and thus
X(I)(cid:107)1/2k = (cid:107)X (k)(cid:107)1/k; in
The joint spectral radius will be crucial tool for showing that every bounded isomor-
phism gives rise to an nc biholomorphism between the similarity envelopes of varieties.
The following lemma lists some properties of the joint spectral radius.
Lemma 4.4. For every X ∈ Md(n) we have that:
(1) for every S ∈ GLn, ρ(X) = ρ(S−1XS);
(2) if X is irreducible, then ρ(X) = min{(cid:107)S−1XS(cid:107) S ∈ GLn}; and
(3) if X is reducible and X1, . . . , Xk are its Jordan -- Holder components, then ρ(X) =
max{ρ(X1), . . . , ρ(Xk)}, and thus it is the same as the joint spectral radius of the
semi-simple elements in the closure of the similarity orbit of X.
Proof. Let X ∈ Bd(n) and let S ∈ GLn. Observe that for every k ∈ N, (S−1XS)(k) =
S−1X (k)S. Therefore, ρ(S−1XS) ≤ ρ(X). Applying the same consideration we see that
ρ(X) = ρ(S(S−1XS)S−1) ≤ ρ(S−1XS). Hence we have that similarities preserve the
joint spectral radius. In particular we have that ρ(X) = ρ(S−1XS) ≤ (cid:107)S−1XS(cid:107) for every
S ∈ GLn.
On the other hand, if X is irreducible, then -- as noted in the proof of Proposition 4.3
-- there exists a strictly positive A ∈ Mn, such that ΨX(A) = ρ(ΨX)A = ρ(X)2A. Letting
√
A, we put Y = T −1XT , and we find that ΨY (I) = T −1ΨX(A)(T −1)∗ = ρ(X)2I. On
T =
the other hand, ΨY (I) = (cid:107)Y (cid:107)2I, and ρ(X)2I = ρ(Y )2I. We therefore get that ρ(Y ) = (cid:107)Y (cid:107)
and additionally (cid:107)Y (cid:107) = min{(cid:107)S−1XS(cid:107) S ∈ GLn}.
k XSk = Xss. Since ρ(S−1
For the last claim observe that applying a similarity we may assume that X is block
upper triangular, or in other words, that X is similar to a sum Xss + N of block diagonal
d-tuple Xss and jointly nilpotent d-tuple N. Now we can choose Sk ∈ GLn, such that
limk→∞ S−1
radius of an operator on a finite dimensional space is a continuous function, we get that
ρ(X) = ρ(Xss). It is obvious that ρ(Xss) = max{ρ(X1), . . . , ρ(Xk)}.
(cid:3)
Lemma 4.5 (Theorem 3.8 of [39]). Let X ∈ Md and k ∈ N. Then, ρ(X) < 1 if and only if
k XSk) = ρ(X) = (cid:112)ρ(ΨX), and since the spectral
limk→∞ (cid:107)X (k)(cid:107) = 0, and this happens if and only if X ∈ (cid:102)Bd.
X ∈ (cid:102)Bd. Finally, let X be a strict row contraction. Then ΨX is a strict contraction, so
Proof. If ρ(X) < 1, then for k ∈ N sufficiently large (cid:107)X (k)(cid:107)1/k ≤ r < 1, and thus (cid:107)X (k)(cid:107) ≤
rk → 0. If limk→∞ (cid:107)X (k)(cid:107) = 0, then limk→∞ Ψk
X(I) = 0, so X is pure. By Proposition 4.3,
ρ(X) = ρ(ΨX)1/2 < 1. Since similar tuples have the same joint spectral radius, the proof is
(cid:3)
complete.
Remark 4.6. In [39, Theorem 3.8], Popescu proves in greater generality the equivalence
of the first and the third conditions in Lemma 4.5, the equivalence in Lemma 4.1 and the
one in Proposition 4.3. More precisely, in his notations, if one sets k = 1, n1 = d, m1 = 1,
q1 = Z1,1 +···+Z1,d, and Q = {0}, then one obtains Φqi,Xi(Y ) = ΨX(Y ), ∆p
q,X = (id−ΨX)p,
14
q (H) = (B(H)d)1, where (B(H)d)1 := {X ∈ B(H)d :(cid:80)d
and Dm
unit ball of B(H)d.
i=1 XiX∗
i < IH} is the open
X(I)(cid:107) = 0, and
In this case, Theorem 3.8 in his paper becomes the equivalence of the following five
conditions:
(i) X is similar to some Y ∈ (B(H)d)1,
(ii) there exists A ≥ 0 such that ΨX(A) < A,
(iii) ρ(X) < 1,
(iv) limk→∞ (cid:107)Ψk
(v) ΨX is power-bounded and pure, and there exists an invertible positive B ∈ B(H)
Consider the case H = Cn. Then the equivalence (i) ⇐⇒ (ii) is our Lemma 4.1, the
equivalence (i) ⇐⇒ (iv) is our Proposition 4.3, and the equivalence (i) ⇐⇒ (iii) is the
equivalence of the first and the third conditions in Lemma 4.5. Condition (v) is just an
alternative version of condition (ii). Nevertheless, our methods are different than those of
Popescu.
such that the equation ΨX(Z) = Z − B has a positive solution Z ∈ B(H).
The goal of the following discussion is to obtain a Schwarz type lemma for the joint
spectral radius, by proving that it is "subharmonic on discs". We will follow the ideas of
Vesentini [51] and [52].
Lemma 4.7. If f : D → Md(n) is an analytic function, then the function log ρ(f (z)) is sub-
harmonic.
Proof. It is obvious that the function fk(z) = f (z)(k) is holomorphic. Therefore, since the
norm of a holomorphic Banach-valued function is subharmonic (this follows from Cauchy's
formula) the function uk(z) = (cid:107)fk(z)(cid:107) is continuous and subharmonic. As in the proof
of [51, Theorem 1] we use the fact that a function u is log-subharmonic in D if and only if
for every a ∈ C the function eazu(z) is subharmonic. Therefore log uk(z) is subharmonic.
Let us write Tn = I2n ⊗ X (2n), then X (2n+1) = X (2n)Tn and thus (cid:107)X (2n+1)(cid:107) ≤ (cid:107)X (2n)(cid:107)2.
Therefore, u2n+1(z)1/2n+1 ≤ u2n(z)1/2n for every z ∈ D. Thus the sequence of functions
u2n(z)1/2n monotonically pointwise decreases to ρ(f (z)). As in part B of the proof of [51,
Theorem 1] we conclude that log ρ(f (z)) is subharmonic.
(cid:3)
Corollary 4.8. The function ρ(f (z)) is subharmonic for every f : D → Md(n) analytic.
The following lemma is a version of the Schwarz lemma for the joint spectral radius.
Lemma 4.9 (Vesentini -- Schwarz Lemma). If f : D → (cid:102)Bd(n) is an analytic function and
f (0) = 0, then ρ(f (z)) ≤ z for every z ∈ D and ρ(f(cid:48)(0)) ≤ 1.
Proof. Note that for every z ∈ D, ρ(f (z)) < 1. Since the function g(z) = f (z)
is analytic,
by Corollary 4.8 we get that ρ(g(z)) is subharmonic of the disc. Now for every z0 ∈ D and
every z0 < r < 1, since the joint spectral radius is homogeneous we get that ρ(g(z0)) ≤ 1
r .
Passing to the limit we see that ρ(g(z0)) ≤ 1. Thus ρ(f (z)) ≤ z for every z ∈ D.
Finally, note that g(0) = f(cid:48)(0) and thus ρ(f(cid:48)(0)) = ρ(g(0)) ≤ 1.
(cid:3)
We would like to obtain a spectral counterpart of the statement in the classical Schwartz
lemma that if f : D → D is such that f (0) = 0 and f(cid:48)(0) = 1, then f (z) = zf(cid:48)(0). We will
break the preparation for this result into several lemmas.
z
15
√
A, we get a row coisometry.
Lemma 4.10. Let X ∈ Md(n) be an irreducible point, such that ρ(X) = 1. Then X is similar
to a coisometry.
Proof. By Lemma 4.4 we know that X is similar to a point Y , such that (cid:107)Y (cid:107) = ρ(Y ) = 1,
hence Y is an irreducible row contraction. As in the proof of Proposition 4.3 we note that
ΨY is irreducible and thus there exists a strictly positive A, such that ΨY (A) = A. Taking
(cid:3)
the similarity with
Lemma 4.11. Let g : D → Bd(n) be an analytic function, such that g(z) is a coisometry for
some z0 ∈ D. Then g is constant.
Proof. Consider every X ∈ Bd(n) as a linear map X : Cd ⊗ Cn → Cn. Let ξ ∈ Cn be a
unit vector and define gξ(z) = (cid:104)g(z)g(z0)∗ξ, ξ(cid:105). Note that gξ is an analytic function on
D. Furthermore, by Cauchy -- Schwarz gξ(z) ≤ 1 on D and gξ(z0) = 1, since g(z0) is a
coisometry. Conclude that gξ is the constant function 1. Since this is true for every ξ we
see that for every z ∈ D, the numerical range of the operator g(z)g(z0)∗ is the singleton
{1}, and thus this operator is the identity. Thus,
0 ≤ (g(z) − g(z0)) (g(z)∗ − g(z0)∗) = g(z)g(z)∗−g(z)g(z0)∗−g(z0)g(z)∗+I = g(z)g(z)∗−I ≤ 0.
(cid:3)
Therefore, g(z) = g(z0).
Lemma 4.12. Let V be a finite dimensional real vector space, and let z (cid:55)→ Tz be a smooth
function from D to L(V ) the space of real linear transformations over V . Assume that 0 is
a simple eigenvalue (i.e., of algebraic multiplicity 1) of Tz for every z ∈ D. Then for every
0 (cid:54)= v0 ∈ ker T0, there exists s > 0 and a smooth function v : sD → V satisfying v(0) = v0,
such that
ker Tz = span{v(z)}
for all z ∈ sD.
∂v
∂g
∂v
=
α
vT
0
0
(cid:104) f (v,λ,z)
(cid:104) Tzv−λv
(cid:105)
equivalently that its kernel is trivial. Let(cid:2) u
Let F (v, λ, z) :=
0 v−1
vT
(cid:105)
g(v,λ,z)
:=
∂f
∂λ
∂g
∂λ
(cid:105)
(cid:104) ∂f
(cid:104) Tz−λ −v
. Then, F (cid:48)(v, λ, z) :=
Proof. We may identify V as Rn and L(V ) as Mn(R). We first show that there exists s > 0
and smooth functions v : sD → V and λ : sD → R satisfying v(0) = v0 and λ(0) = 0, such
that Tzv(z) = λ(z)v(z) for all z ∈ sD. This is likely to be very well known and also appears
in an unpublished note of Kazdan [26], but we include it here for completeness.
(cid:105)
(cid:3) ∈ ker F (cid:48)(v(0), λ(0), 0). Then T0u − αv0 = 0
0 u = 0. If α (cid:54)= 0, then by the first equality T 2
. To
apply the implicit function theorem we must show that F (cid:48)(v(0), λ(0), 0) is invertible, or
0 u = αT0v0 = 0 while T0u (cid:54)= 0,
and also vT
which is impossible as 0 is a simple eigenvalue of T0. Thus α = 0, so by the first equality
we have u ∈ ker T0 = span{v0}. Therefore, the second equality implies that u = 0. The
implicit function theorem now gives a neighborhood sD of the origin and smooth functions
v : sD → V and λ : sD → R satisfying v(0) = v0 and λ(0) = 0, such that Tzv(z) = λ(z)v(z)
for all z ∈ sD.
Finally, we need to show that λ must be the zero function. Let pTz (x) := det(xI −
Tz) be the characteristic polynomial of Tz, and set p(z, x) := pTz (x). Since 0 is a simple
∂x(0, 0) (cid:54)= 0. By the implicit function theorem, there exists a
eigenvalue of T0 we have that ∂p
neighborhood sD of the origin and a function µ : sD → R such that inside a neighborhood
of (0, 0) ∈ D × R the curve (z, µ(z)) contains all solutions to p(z, x) = 0, so λ(z) = µ(z) in
a neighborhood of 0. In particular, as we know that (z, 0) is a solution we find that µ = λ
(cid:3)
must be zero near the origin.
16
In [17, Theorem 2.3], the authors show -- as part of their generalization of the Perron --
Frobenius theorem -- that for a positive irreducible linear map Θ on Mn with spectral
radius r, there exists a unique strictly positive matrix A such that Θ(A) = rA. It is inter-
esting (and also useful, as we shall see in the next Lemma) to note that r is not only of
geometric multiplicity 1, but also of algebraic multiplicity 1 (i.e., simple).
To see this, equip Mn with the Hilbert -- Schmidt inner product, namely set (cid:104)Z, W(cid:105) :=
tr(W ∗Z), and let Θ∗ be the adjoint of Θ with respect to this inner-product. Since Θ∗ is
positive and irreducible as well [17], there exists a unique strictly positive matrix B such
that Θ∗(B) = r(cid:48)B where r(cid:48) is the spectral radius of Θ∗. Note that B⊥ := {Z ∈ Mn :
(cid:104)Z, B(cid:105) = 0} is Θ invariant.
If we show that Mn is the direct sum of the two Θ-invariant spaces CA and B⊥, then
we are clearly done. To this end, note that (cid:104)A, B(cid:105) = tr(B∗A) = tr(cid:0)(A 1
2(cid:1). Since B is
2 )∗BA 1
17
strictly positive, all its eigenvalues are positive. By the complex version of Sylvester's law
of inertia, all the eigenvalues of the latter matrix are positive, so that (cid:104)A, B(cid:105) > 0. Thus,
A (cid:54)∈ B⊥.
Lemma 4.13. Let g : D → Md(n) be an analytic function, such that g(z) is irreducible and
similar to a coisometry for every z ∈ D, then there exists an irreducible coisometry X, such
that g(z) is similar to X, for every z ∈ D.
Proof. Since for every z ∈ D, g(z) is irreducible and ρ(g(z)) = 1 we know that Ψg(z) is
irreducible with spectral radius 1. The discussion prior to this Lemma now implies that 1 is
a simple eigenvalue of Ψg(z) with a positive eigenvector. Note that Ψg(z)−idMn maps the real
vector space of self-adjoint matrices (Mn)sa into itself, and set Tz := (Ψg(z) − idMn)(Mn)sa.
Choose a positive A0 ∈ ker T0. By Lemma 4.12, there exists s > 0 and a smooth function
a : sD → (Mn)sa satisfying a(0) = A0 such that
Ψg(z)a(z) = a(z)
for all z ∈ sD.
Since a(0) > 0 there exists some r > 0 such that a(z) > 0 for every z ∈ rD. Since the
minimal eigenvalue of a is bounded away from 0 on the circle z = r, by [9, Corollary
III.2.1] we can find a holomorphic function h on the disc rD and continuous on the circle,
such that h(reit)h(reit)∗ = a(reit) for every t ∈ [0, 2π). There are two things to note
regarding the result cited. First, the result discusses right factorization, but it is equivalent
to the left factorization by just factoring a(z)T as Clancey and Gohberg indicate. Secondly,
it is immediate that a(z) has entries in the Wiener algebra since it is smooth.
Define(cid:101)g(z) = h(z)−1g(z)h(z). Note that for z = r,
so(cid:101)g(z) is a coisometry. Now think of(cid:101)g as a function taking values in n × nd matrices, and
complete(cid:101)g to a function F with values in nd× nd matrices by adding rows of zeroes. Note
that for every z ∈ rD, the n largest singular values of F (z) are the singular values of(cid:101)g(z).
function on rD that we shall denote by σ(z). Since each singular value of(cid:101)g(z) is bounded
(cid:101)g(z) is similar to a coisometry, we know that its singular values are non-zero (for otherwise,
(cid:101)g(z)∗ would have had a kernel, and similarity is implemented by a base change of specific
By [4, Theorem 1] the product of the n largest singular values of F (z) is a subharmonic
by 1 (the norm is the largest singular value), 0 ≤ σ(z) ≤ 1. Furthermore, for every
t ∈ [0, 2π) we have σ(reit) = 1, since the boundary values are coisometries. Since each
(cid:101)g(z)(cid:101)g(z)∗ = h(z)−1Ψg(z)(a(z))(h(z)∗)−1 = In,
form in the domain and range of the linear map). We conclude that σ is bounded away
from zero on rD. Therefore the function 1/σ is also subharmonic, since a composition
of a subharmonic function with a convex function is subharmonic. However, 1/σ is also
bounded by 1 and thus σ is identically 1 on rD.
Now recall that for each z ∈ rD, the singular values of (cid:101)g are non-negative numbers
decomposition,(cid:101)g(z) is a coisometry. Now it remains to apply Lemma 4.11 to conclude that
(cid:101)g is constant.
To conclude the proof, note that since X = (cid:101)g(0) is irreducible its similarity orbit is an
bounded from above by 1, and their product is 1, hence they are all 1, and by singular value
algebraic subvariety of Md(n), hence there is a finite number of polynomials p1, . . . , pk,
that cut out this orbit. By the above discussion the functions p1(g(z)), . . . , pk(g(z)) are
identically zero on rD and thus on all of D. Hence, we can conclude that g lands in the
(cid:3)
similarity orbit of X.
The next proposition is a version of the equality clause in the classical Schwartz lemma.
Proposition 4.14. If f : D → (cid:102)Bd(n) is an analytic function is such that f (0) = 0 and f(cid:48)(0) is
an irreducible coisometry, then f (z) is similar to zf(cid:48)(0), for every z ∈ D.
Proof. Since f(cid:48)(0) is a coisometry, then ρ(f(cid:48)(0)) = 1 and thus equality holds in the Vesentini --
Schwartz lemma. Let g(z) = f (z)/z, then ρ(g(z)) = 1 for every z ∈ D and g(0) = f(cid:48)(0) = X
is an irreducible coisometry. Since the irreducible points are open there exists r > 0, such
that for every z < r we have g(z) is irreducible and thus similar to a coisometry by
Lemma 4.10. Now by Lemma 4.13 g takes values in the similarity orbit of X. Now since
(cid:3)
f (z) = zg(z) we are done.
The following theorem is a spectral version of the Cartan uniqueness theorem. Let
X ∈ Bd. we will define the spectrum of X to be its Jordan -- Holder componenets and
denote it by σJH(X).
Theorem 4.15. Let G : (cid:102)Bd → (cid:102)Bd be an analytic nc map, such that G(0) = 0 and ∆G(0, 0) =
Consider the function f : D → (cid:102)Bd, defined by f (z) = G(zX/(cid:107)X(cid:107)). By our assumption
I. Then σJH(G(X)) = σJH(X) for every X ∈ Bd. In particular, if X is irreducible, then G(X)
is similar to X.
Proof. Assume first that X ∈ Bd is an irreducible point, such that X/(cid:107)X(cid:107) is a coisometry.
f(cid:48)(0) = ∆G(0, 0)(X/(cid:107)X(cid:107)) = X/(cid:107)X(cid:107). Applying Proposition 4.14 we have that f (z) is
similar to zX/(cid:107)X(cid:107).
Now if X is an irreducible point, by Lemmas 4.4 and 4.10, we know that X is similar to
a point Y , such that Y /(cid:107)Y (cid:107) is coisometric. Since G is nc G(X), is similar to G(Y ), but by
the previous paragraph G(Y ) is similar to Y and we are done with the case of irreducible
points.
Now for arbitrary X, we know that X is similar to a block upper triangular form with
irreducible blocks on the diagonal. Since G is nc the result follows from the irreducible
(cid:3)
case.
Remark 4.16. The theorem implies that such a G preserves the similarity orbit of the
semi-simple part of every X and thus induces the identity map on the GIT quotients
18
d = 1 was proved in [41] in wider generality, namely without assuming equivariance.
(cid:102)Bd(n)// PGLn. Hence the resemblance with Cartan's uniqueness theorem. The case for
Let X, Y ∈ (cid:102)Bd(n). We define the free pseudo-hyperbolic distance between X and Y by
representations of A(Bd) on Cn associated to X and Y , respectively. If X ∈ (cid:102)Bd(n) and
Y ∈ (cid:102)Bd(m), then we set δcb(X, Y ) = δcb(X⊕m, Y ⊕n).
δcb(X, Y ) = (cid:107)ΦX − ΦY (cid:107)cb, where ΦX and ΦY are norm continuous completely contractive
5. THE FREE PSEUDO-HYPERBOLIC DISTANCE
One can extend this distance to representations of H∞(Bd) on finite dimensional Hilbert
spaces. For the points in the interior of the ball that correspond to weak-∗ continuous
representations, these distances coincide. Therefore, for the interior of the ball we may
consider either H∞(Bd) or A(Bd).
(i) The free pseudo-hyperbolic distance is a pseudo-metric on (cid:102)Bd that is
independent of the choice of the level.
Proposition 5.1.
(ii) For every two points X, Y ∈ Bd we have δcb(X, Y ) < 2.
(iii) We have the following nc properties of the pseudo-hyperbolic distance:
• δcb(X(cid:48) ⊕ X(cid:48)(cid:48), Y (cid:48) ⊕ Y (cid:48)(cid:48)) = max{δcb(X(cid:48), Y (cid:48)), δcb(X(cid:48)(cid:48), Y (cid:48)(cid:48))}.
• If X, Y ∈ (cid:102)Bd(n) and S ∈ GLn, such that S−1XS, S−1Y S ∈ Bd, then:
δcb(S−1XS, S−1Y S) ≤ (cid:107)S(cid:107)(cid:107)S−1(cid:107)δcb(X, Y ).
Proof. (i) We think of Bd(n) as a subset of CB(A(Bd), Mn), the set of all completely
bounded homomorphism from A(Bd) into Mn, via the correspondence X (cid:55)→ ΦX. Thus,
in fact, each level is identified with a subset of the unit sphere in this Banach space and
we use the operator space structure on the operator space dual of A(Bd) to induce δcb.
By [24, Proposition 7.14] we immediately see that it is a pseudo-metric, with only the
direct sum ampliations of the same point identified.
(ii) For every f ∈ Mn(H∞(Bd)), such that f (0) = 0 and (cid:107)f(cid:107)∞ < 1, Popescu's free
Schwarz lemma [36] says that (cid:107)f (X)(cid:107) ≤ (cid:107)X(cid:107). This implies immediately that δcb(0, X) =
(cid:107)X(cid:107) for every X ∈ Bd. Now from the triangle inequality we conclude that for every two
points X, Y ∈ Bd we have δcb(X, Y ) < 2.
space structure on the dual of A(Bd).
(iii) This is immediate from the fact that the completely bounded norm is an operator
(cid:3)
Proposition 5.2. Let V ⊆ Bd and W ⊆ Be be nc varieties, and G :(cid:102)W → (cid:101)V an nc holomor-
phic function. Then the following statements are equivalent:
(i) G is Lipschitz with respect to δcb,
(iii) the mapping f (cid:55)→ (cid:101)f ◦ G for f ∈ H∞(V) is a well-defined completely bounded homomor-
(ii) supW∈W (cid:107)ΦG(W )(cid:107)cb < ∞, and
phism α : H∞(V) → H∞(W).
Furthermore, in this case
(a) (cid:107)α(cid:107)cb = supW∈W (cid:107)ΦG(W )(cid:107)cb,
(b) the value in (a) is a Lipschitz constant for G, and
(c) α is weak-∗ continuous.
19
Proof. (i) =⇒ (ii). If G is K-Lipschitz, choose some W0 ∈ W. Then
(cid:107)ΦG(W )(cid:107)cb ≤ (cid:107)ΦG(W )−ΦG(W0)(cid:107)cb+(cid:107)ΦG(W0)(cid:107)cb ≤ K(cid:107)ΦW−ΦW0(cid:107)cb+(cid:107)ΦG(W0)(cid:107)cb ≤ 2K+(cid:107)ΦG(W0)(cid:107)cb.
Thus, supW∈W (cid:107)ΦG(W )(cid:107)cb < ∞.
(ii) =⇒ (iii). If supW∈W (cid:107)ΦG(W )(cid:107)cb < ∞, then for every f =(cid:0)fij
(cid:13)(cid:13)(cid:13)(cid:0)(cid:101)fij(G(W ))(cid:1)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)ΦG(W )
(cid:1)(cid:13)(cid:13)
(cid:0)fij
(cid:107)((cid:101)f ◦ G)(W )(cid:107) = sup
(cid:1) ∈ Mn(H∞(V))
sup
W∈W
W∈W
= sup
W∈W
≤ sup
W∈W
(cid:107)ΦG(W )(cid:107)cb(cid:107)f(cid:107)Mn(H∞(V)).
Thus, (cid:101)f ◦ G is in Mn(H∞(W)), the map α : f (cid:55)→ (cid:101)f ◦ G is a completely bounded homomor-
phism, and (cid:107)α(cid:107)cb ≤ supW∈W (cid:107)ΦG(W )(cid:107)cb.
(iii) =⇒ (i). As ΦG(W ) = α∗(ΦW ) and α is completely bounded, G is (cid:107)α(cid:107)cb-Lipschitz.
The last argument shows, in addition, that supW∈W (cid:107)ΦG(W )(cid:107)cb ≤ (cid:107)α(cid:107)cb, so we have (a)
and (b). To show (c), we need to use following two facts: (1) by Theorem 3.1 the WOT and
weak-∗ topologies coincide in H∞(V) and H∞(W) and (2) for bounded nets in H∞(V),
WOT-convergence and pointwise convergence are the same [43, Lemma 2.5]. From the
first fact, we see (using the Krein-Smulian Theorem) that it suffices to check that α is
weak-∗/WOT continuous on the closed unit ball of H∞(V). This follows readily from the
(cid:3)
second fact, since α is implemented by composition.
Corollary 5.3. Let V ⊆ Bd and W ⊆ Be be nc varieties, and G : W → V an nc holomorphic
map. Then G is a contraction with respect to the pseudo-hyperbolic metric. In particular, if G
is a biholomorphism, then it is isometric.
Proof. Since G(W) ⊆ V, we have that supW∈W (cid:107)ΦG(W )(cid:107)cb ≤ 1. By Proposition 5.2, G is
(cid:3)
1-Lipschitz, so it is a contraction.
Proposition 5.4. Let G : Bd → Bd be an nc holomorphic map, and assume d < ∞. Then G
is an automorphism of Bd if and only if it is isometric with respect to the pseudo-hyperbolic
distance.
Proof. If G is an automorphism, by Corollary 5.3, it is isometric. On the other hand,
suppose that G is an isometry. G maps scalar points to scalar points, so by composing
with an nc automorphism of the ball Bd, we may assume that G(0) = 0. To see that there
is indeed no harm in this assumption, recall that Aut(Bd) = Aut(Bd) (see [38, Theorem
2.8]) and every such automorphism gives rise to a completely isometric automorphism of
H∞(Bd) (see Propositions 6.5 and 6.8 in [43]), therefore an automorphism of the ball is
isometric with respect to the pseudo-hyperbolic distance.
So suppose that G is an isometry and that G(0) = 0. Since δcb is a metric on every level
and G is graded, we see that G restricts to an isometry of the compact metric space rBd(n)
into itself, for every n and every r < 1. As such, the restriction of G to rBd(n) is also
surjective for all n and r, and we conclude that G is bijective. It is then straightforward
(cid:3)
that G−1 must also be free holomorphic.
20
Evidently, δb ≤ δcb, and all results above have counterpart versions for δb. It is interesting
to note that, since Proposition 5.4 holds with δb instead of δcb, it follows that a self map of
Bd is isometric with respect to δcb if and only if it is isometric with respect to δb. We record
the δb-version of Proposition 5.2 for later use. In fact, when the distance δb is used, we get
a somewhat sharper result.
Proposition 5.5. Let V ⊆ Bd and W ⊆ Be be nc varieties, and G :(cid:102)W → (cid:101)V an nc holomor-
phic function. Then the following statements are equivalent:
(i) G is Lipschitz with respect to δb,
(iii) the mapping f (cid:55)→ (cid:101)f ◦ G for f ∈ H∞(V) is a well-defined bounded homomorphism
(ii) supW∈W (cid:107)ΦG(W )(cid:107) < ∞,
(iii)' the mapping f (cid:55)→ (cid:101)f ◦ G for f ∈ H∞(V) is a well-defined homomorphism α : H∞(V) →
α : H∞(V) → H∞(W), and
H∞(W).
Furthermore, in this case
(a) (cid:107)α(cid:107) = supW∈W (cid:107)ΦG(W )(cid:107),
(b) the value in (a) is a Lipschitz constant for G, and
(c) α is weak-∗ continuous.
Proof. The proof of (i) =⇒ (ii) =⇒ (iii) =⇒ (i), and the proof of (a), (b), and (c) (in
case (i) -- (iii) holds, is the same as in Proposition 5.2. We also obviously have (iii) =⇒
(iii)'. We shall show (iii)' =⇒ (ii). Suppose that α is a well defined map from H∞(V) into
function on W for all f ∈ H∞(V), so we have
H∞(W) (it is clear in this case that it is a unital homomorphism). Then (cid:101)f ◦ G is a bounded
(cid:107)ΦG(W )(f )(cid:107) = sup
W∈W
sup
W∈W
(cid:107)(cid:101)f (G(W ))(cid:107) = (cid:107)(cid:101)f ◦ G(cid:107)∞ < ∞.
One can define another free pseudo-hyperbolic distance:
δb(X, Y ) = (cid:107)ΦX − ΦY (cid:107).
Since this holds for every f ∈ H∞(V), the uniform boundedness principle implies (ii). (cid:3)
Remark 5.6. It is tempting to consider a version of Gleason parts with respect to the free
pseudo-hyperbolic metric. The immediate question is whether the relation δcb(X, Y ) < 2 is
an equivalence relation. Classically, one can define the Gleason parts of a function algebra
using either a variant of the pseudo-hyperbolic metric or a form of Harnack inequality.
Harnack domination was defined by Suciu ( [46]) for the case d = 1 and by Popescu for
d > 1 ( [36]). Therefore, it is also natural to ask whether there is a connection between
the Harnack parts as defined by Popescu and the pseudo-hyperbolic metric. Unfortunately,
the answer to both questions is negative.
To see that the "Gleason parts" are not parts at all consider the following example. Let
(cid:1). Let f ∈ Mk(A(D)), then
0 1 ), then S−1AS =(cid:0) 0 1/2
0 0
A = ( 0 1
0 0 ) and S = ( 2 0
(cid:107)f (S−1AS) − f (A)(cid:107) = (cid:107)IMk ⊗ (TS − IM2)(A)(cid:107) ≤ (cid:107)TS − IM2(cid:107).
Here TS is the operator on M2 defined by TS(B) = S−1BS. Note that since S is diagonal,
TS is in fact a Schur multiplier, so is TS − IM2 is too. In particular, (TS − IM2)(B) = C ◦ B,
21
where C =(cid:0) 0 −1/2
(cid:1). Therefore, (cid:107)TS − IM2(cid:107) = 1, and we can conclude that for every k ∈ N
and f ∈ Mk(A(D)) we get
1
0
(cid:107)ΦS−1AS(f ) − ΦA(f )(cid:107) ≤ (cid:107)TS − IM2(cid:107) < 2.
α∗ : (cid:116)n Repn(H∞(W)) → (cid:116)n Repn(H∞(V)),
22
So A is "equivalent" to an interior point and so is −A, however, the distance between them
is at least 2, as can be observed by taking f (z) = z. Therefore, this is not an equivalence
relation.
Harnack domination is an equivalence relation, as proved by Popescu, and the open ball
is a part. One can show, just as in the classical case, that if X Harnack dominates Y , then
δcb(X, Y ) < 2, however as we have seen above, the converse is false.
Remark 5.7. Recently, a more general approach to noncommutative hyperbolic geometry
was formulated by Belinschi and Vinnikov in [7]. They define a noncommutative version
of the Lempert function for an nc domain Ω. Let X ∈ Ω(n) and Y ∈ Ω(m) and Z ∈
Mn,m(C) ⊗ Cd, then
L(X, Y )(Z)−1 = sup
t ∈ [0,∞]
∈ Ω, for all s ∈ [0, t]
.
:
0
Y
Unfortunately, in the case of Ω = (cid:101)Bd we get that L is identically 0, since the similarity
envelope of the ball consists precisely of points of spectral radius strictly less than 1 and
by Lemma 4.4 we know that the spectral radius of an block upper triangular matrix is the
same as the maximum of the spectral radii of its diagonal blocks. This observation also
tells us that there are many copies of C embedded into every level except the first. This
is an indication of the fact that the nice hyperbolic structure of the ball that is used in the
previous works cannot be extended to the similarity envelope. That being said, the results
of Section 4 show that a modicum of hyperbolic-like properties is preserved, when one
passes to the similarity envelope, if one uses the joint spectral radius.
(cid:26)
(cid:20)X sZ
(cid:21)
(cid:27)
6. CLASSIFICATION UP TO WEAK-∗ CONTINUOUS ISOMORPHISM
Let V ⊆ Bd and W ⊆ Be be nc varieties.
N ∪ {∞}, unless we state otherwise. An nc holomorphic map G : (cid:102)Be → Md is called an nc
biholomorphism from(cid:102)W onto (cid:101)V if there exists an nc holomorphic map F : (cid:102)Bd → Me such
In this section, we allow for any d, e ∈
that
F ◦ G(cid:101)W = id(cid:101)W and G ◦ F(cid:101)V = id(cid:101)V.
In this case, and whenever there is no confusion, we let G−1 denote the map F . The goal
of this section is to classify the algebras of bounded analytic functions on nc subvarieties
of Bd in terms of biholomorphisms between the respective similarity invariant envelopes
of the varieties.
If f is an nc holomorphic function on V, then we write (cid:101)f for the unique extension of f
to its similarity envelope (cid:101)V.
Remark 6.1. If α : H∞(V) → H∞(W) is a unital and bounded homomorphism, then the
adjoint α∗ : H∞(W)∗ → H∞(V)∗ is also bounded. Moreover, there is also a natural adjoint
mapping α∗ between representation spaces,
α(f ) = (cid:101)f ◦ G.
given by α∗(ρ) = ρ ◦ α. If α is completely contractive/bounded, then α∗ preserves com-
pletely contractive/bounded representations.
Proposition 6.2. Let V ⊆ Bd and W ⊆ Be be nc varieties. Let α : H∞(V) → H∞(W)
be a unital bounded homomorphism. If α∗ maps weak-∗ continuous finite dimensional rep-
resentations to weak-∗ continuous representations, then there exists an nc holomorphic map
G :(cid:102)W → (cid:101)V which implements α by the formula
In this case, (cid:107)α(cid:107) = supW∈W (cid:107)ΦG(W )(cid:107). If α is, in addition, completely bounded, then (cid:107)α(cid:107)cb =
supW∈W (cid:107)ΦG(W )(cid:107)cb.
Proof. Define G :(cid:102)W → (cid:102)Bd by
By the assumption on α∗ and Theorem 3.4, we find that G maps(cid:102)W into (cid:101)V. Note that
for all W ∈ W, so that α(f ) = (cid:101)f ◦ G. Propositions 5.2 and 5.5 show that (cid:107)α(cid:107) =
supW∈W (cid:107)ΦG(W )(cid:107) and that (cid:107)α(cid:107)cb = supW∈W (cid:107)ΦG(W )(cid:107)cb, if α is completely bounded.
(cid:101)f (G(W )) = ΦG(W )(f ) = α∗(ΦW )(f ) = ΦW (α(f )) = α(f )(W )
G(W ) = πk(α∗(ΦW )), W ∈(cid:102)W(k).
(cid:3)
We record the isomorphism equivalence implied by Propositions 5.2, 5.5 and 6.2.
Corollary 6.3. Let V ⊆ Bd and W ⊆ Be be nc varieties. Then the following statements are
equivalent:
(i) H∞(V) and H∞(W) are weak-∗ continuously and completely boundedly isomorphic,
(ii) the similarity envelopes (cid:101)V and(cid:102)W are nc biholomorphic via a δcb-bi-Lipschitz biholomor-
(iii) the similarity envelopes (cid:101)V and(cid:102)W are nc biholomorphic via an nc biholomorphism G :
(cid:102)W → (cid:101)V satisfying
phism,
(cid:107)ΦG(W )(cid:107)cb < ∞ and
(cid:107)ΦG−1(V )(cid:107)cb < ∞.
sup
V ∈V
sup
W∈W
Furthermore, in this case supW∈W (cid:107)ΦG(W )(cid:107)cb = (cid:107)α(cid:107)cb and supW∈W (cid:107)ΦG−1(W )(cid:107)cb = (cid:107)α−1(cid:107)cb,
and
(cid:107)α−1(cid:107)−1
cb δcb(W1, W2) ≤ δcb(G(W1), G(W2)) ≤ (cid:107)α(cid:107)cbδcb(W1, W2).
Likewise, the following statements are equivalent:
(i) H∞(V) and H∞(W) are weak-∗ continuously and boundedly isomorphic,
(i)' H∞(V) and H∞(W) are weak-∗ continuously isomorphic,
(ii) the similarity envelopes (cid:101)V and(cid:102)W are nc biholomorphic via a δb-bi-Lipschitz biholomor-
(iii) the similarity envelopes (cid:101)V and(cid:102)W are nc biholomorphic via an nc biholomorphism G :
(cid:102)W → (cid:101)V satisfying
phism,
(cid:107)ΦG(W )(cid:107) < ∞ and
(cid:107)ΦG−1(V )(cid:107) < ∞.
sup
W∈W
sup
V ∈V
23
Furthermore, in this case supW∈W (cid:107)ΦG(W )(cid:107) = (cid:107)α(cid:107) and supW∈W (cid:107)ΦG−1(W )(cid:107) = (cid:107)α−1(cid:107), and
(cid:107)α−1(cid:107)−1δb(W1, W2) ≤ δb(G(W1), G(W2)) ≤ (cid:107)α(cid:107)δb(W1, W2)
Remark 6.4. Note that when α is a completely isometric isomorphism, or equivalently
when (cid:107)α(cid:107)cb = (cid:107)α−1(cid:107)cb = 1, then Corollary 6.3 implies that supW∈W (cid:107)ΦG(W )(cid:107)cb = 1 and
supV ∈V (cid:107)ΦG−1(V )(cid:107)cb = 1. Since (cid:107)T(cid:107) ≤ (cid:107)ΦT(cid:107)cb for every T , we obtain that G must map
W injectively into V and G−1 maps V injectively into W. Recalling the nc maximum
principle [43, Lemma 6.11], it follows that G is an nc biholomorphism between W and
V. Conversely, whenever G is an nc biholomorphism between W and V, we have that
supW∈W (cid:107)ΦG(W )(cid:107)cb = 1 and supV ∈V (cid:107)ΦG−1(V )(cid:107)cb = 1. Corollary 6.3 yields a completely
bounded isomorphism α with (cid:107)α(cid:107)cb = (cid:107)α−1(cid:107)cb = 1, so that α is in fact a completely
isometric isomorphism. This recovers [43, Corollary 6.14] (in [43] it was shown that
when d, e < ∞, then the weak-∗ continuity condition is redundant).
Example 6.5. In [10, Example 6.6] the authors show that given a separated, but not
strongly separated Blaschke sequence V ⊆ D, there exists a bi-Lipschitz biholomorphism
of V onto W , where W is an interpolating sequence. Since V is not interpolating, the
associated multiplier algebras are not isomorphic. The above theorem demonstrated what
fails in this case, since bi-Lipschitz on the first level is not enough to conclude that the
multiplier algebras are isomorphic.
The following example shows that the assumption supW∈W (cid:107)ΦG(W )(cid:107) < ∞ does not hold
in fact, it may fail even for automorphisms of (cid:101)Bd. In particular, it shows that not every
for every holomorphic map G between similarity invariant envelopes of varieties, and,
automorphism of (cid:101)Bd gives rise to an automorphism of H∞(Bd).
is a well defined unbounded nc function, and it extends to (cid:101)B2. Define G : (cid:101)B2 → (cid:101)B2 by
It is easy to check that G is biholomorphism of (cid:101)Bd.
Example 6.6. On B2, consider the function f (x) = f (x1, x2) = (1 − x1)2. The inverse of f
G(x) = f (x)−1xf (x).
Now we consider pairs of matrices X = (X1, X2) ∈ B2(2), given by
(cid:20)λ 0
(cid:21)
0 µ
√
X1 =
, X2 =
(cid:20)0 a
(cid:21)
0 0
,
where all entries are assumed positive. In order to be a strict row contraction, it is neces-
1 − λ2 >
1 − λ2. For definiteness, choose a = 1
sary and sufficient that λ, µ < 1 and a <
2(1 − λ).
1
Now we observe G(X) = f (X)−1Xf (X), the second component of which is
2
√
(cid:20)(1 − λ)−2
(cid:20)0 a(cid:0) 1−µ
0
1−λ
0
=
0
(1 − µ)−2
(cid:1)2
(cid:21)
.
(cid:21)(cid:20)0 a
(cid:21)(cid:20)(1 − λ)2
0 0
0
(cid:21)
0
(1 − µ)2
f (X)−1X2f (X) =
0
By the choice of a, we find that
(cid:107)f (X)−1X2f (X)(cid:107) >
24
(1 − µ)2
1 − λ
1
2
.
Fixing µ small and letting λ → 1, we see that G is not bounded on B2(2).
We record a consequence of our results that adds to what is known (see, for example,
[11, 38]) about the automorphism group of H∞(Bd).
Theorem 6.7. Let d ∈ N. Every automorphism α of H∞(Bd) is given by
where G : (cid:101)Bd → (cid:101)Bd is an nc biholomorphism of (cid:101)Bd such that
α(f ) = (cid:101)f ◦ G,
sup
W∈Bd
(cid:107)ΦG(W )(cid:107) < ∞ and
(cid:107)ΦG−1(V )(cid:107) < ∞.
sup
V ∈Bd
Proof. By [11, Theorem 4.6], every automorphism of H∞(Bd) is WOT-continuous. Thus
(cid:3)
Corollary 6.3 applies.
Recall that in [11] an automorphism of H∞(Bd) was called quasi-inner, if it is in the
kernel of the homomorphism to Aut(Bd(1)).
In terms of the preceding theorem, this
means that the induced nc biholomorphism of Bd is the identity when restricted to the first
level. We can therefore apply the spectral Cartan theorem (Theorem 4.15) to obtain the
following corollary, that can be interpreted as stating that it is very difficult to distinguish
a quasi-inner automorphism from an inner one.
Corollary 6.8. If α : H∞(Bd) → H∞(Bd) is a quasi-inner automorphism, then there exists
that for every irreducible X ∈ Bd, G(X) is similar to X.
an nc biholomorphism G : (cid:101)Bd → (cid:101)Bd such that α(f ) = (cid:101)f ◦ G for all f ∈ H∞(Bd), and such
7. THE HOMOGENEOUS CASE
In this section, we focus our attention on the case where the underlying varieties are
homogeneous. Our first result shows that when V and W are homogeneous, the weak-∗
continuity requirement can be dropped in Proposition 6.2 and Corollary 6.3. Throughout
this section, we shall assume that d < ∞.
Theorem 7.1. Let V ⊆ Bd and W ⊆ Be be two homogeneous nc varieties. Let α : H∞(V) →
H∞(W) be a bounded isomorphism. Then the induced nc function G = π ◦ α∗ : (cid:102)W → (cid:102)Bd
maps (cid:102)W into (cid:101)V. Consequently, every (completely) bounded isomorphism between H∞(V)
and H∞(W) arises as the composition with an nc biholomorphism G :(cid:102)W → (cid:101)V.
Proof. We first note that if V = 0 or W = 0 (understood as the nc variety over the scalar
point 0), then the problem trivializes. Thus we assume henceforth that neither variety
consists of just the point 0.
For every n ∈ N we have that α∗ : Repn(H∞(W)) → Repn(H∞(V)) is a bijection and
that G = π ◦ α∗. Suppose towards a contradiction that for some X ∈ W, α∗(ΦX) is not
weak-∗ continuous. Then G(X) ∈ (cid:102)Bd \ (cid:102)Bd, so ρ(G(X)) = 1 by Lemma 4.5. Consider
the function f (X)(z) = G(zX/(cid:107)X(cid:107)) defined and analytic on the disc. By Corollary 4.8 the
function ρ(f (X)(z)) is subharmonic and thus by the maximum modulus principle it is the
constant function 1. In particular ρ(G(0)) = 1, i.e, (cid:107)G(0)(cid:107) = 1 as a point in Bd. This implies
that G is constant on W as follows.
25
1
1
1
d
(z), . . . , f (W )
(0) = In. Next, since G(W) ⊆ (cid:102)Bd, we have that f (W )
Applying a unitary we may assume that G(0) = (1, 0, . . . , 0). Now, let n ∈ N, let W be
a point in W(n), and consider f (W )
(z) -- the first coordinate of the function f (W )(z) =
(z)), where f (W )(z) = G(zW/(cid:107)W(cid:107)). Note that since G respects direct
(f (W )
sums f (W )
(z) is similar to a contraction
for every z ∈ D. Thus, for each z ∈ D, the eigenvalues of f (W )
(z) are all of absolute value
1
less than or equal to 1, so that tr(f (W )
(0)) = tr(In) = n, the classical
(z)) = n for every z ∈ D, and thus all of the
maximum modulus implies that tr(f (W )
eigenvalues of f (W )
(z) is similar
to a contraction, it is power bounded. Since all of its eigenvalues are 1, its Jordan form
(z) = In for every z ∈ D. Since this
cannot contain a non-trivial Jordan block. Thus, f (W )
holds for every point W ∈ W, we have that the function G1 = α(z1V) is identically equal
to the constant nc function 1. Furthermore, for every n ∈ N and every W ∈ W(n), we have
that G(W ) is similar to a row contraction, i.e, there exists S ∈ GLn(C), such that
(z) are all equal to 1. Since for every z ∈ D the matrix f (W )
(z)) ≤ n. As tr(f (W )
1
1
1
1
1
1
1
0 ≤ d(cid:88)
(S−1Gj(W )S)(S−1Gj(W )S)∗ ≤ In.
j=1
However, G1(W ) = I and thus this forces Gj(W ) = 0, for j = 2, . . . , d. Therefore, G is
constant, as we claimed.
Now recall that from the definition of G as π ◦ α∗, the nc function Gj is equal to α(zjV)
for every j = 1, . . . , d. It follows from the previous paragraph that α(zjV) is equal to a
constant nc function on W, for j = 1, . . . , d. Since α is an isomorphism and preserves the
unit, we find that the tuple of coordinate functions (z1V, . . . , zdV) is equal to the constant
It follows that G maps(cid:102)W into (cid:101)V, so by Theorem 3.4, α∗ maps weak-∗ continuous rep-
tuple G(0). This implies that V is the nc variety over the scalar point G(0), which is a
contradiction.
resentations in Repk(H∞(W)) to weak-∗ continuous representations in Repk(H∞(V)). By
argue similarly about α−1, the map G must be a biholomorphism between(cid:102)W and (cid:101)V. (cid:3)
Proposition 6.2, the isomorphism α is given by composition with G. Finally, since we can
the similarity envelopes (cid:101)V and(cid:102)W are nc biholomorphic, then (cid:101)V is the image of(cid:102)W under
Lemma 7.2. Let V ⊆ Bd and W ⊆ Be be two nc homogeneous varieties such that (cid:101)V and(cid:102)W
are nc biholomorphic. If 0 is not mapped to 0, then there exist two discs D1 ⊆ V(1) = (cid:101)V(1)
and D2 ⊆ W(1) =(cid:102)W(1), both containing 0, such that D1 is mapped by the biholomorphism
Proof. Let G : (cid:101)V →(cid:102)W be an nc biholomorphism mapping (cid:101)V onto(cid:102)W. If 0 is not mapped
to 0, then both V := V(1) = (cid:101)V(1) and W := W(1) = (cid:102)W(1) are non-trivial homogeneous
We shall now show that for any two homogeneous nc varieties V ⊆ Bd and W ⊆ Be, if
an invertible linear transformation. We start with the following lemma.
varieties and GV : V → W is a biholomorphism. The first paragraph of the proof of [42,
Lemma 5.9] (see also [13]) shows -- using an analysis of the singular nuclei of V and
W -- that there are two discs D1 ⊆ V and D2 ⊆ W , both containing 0, such that D1 is
mapped by GV onto D2.
(cid:3)
onto D2.
26
D2. Define
and
O(0;(cid:101)V) := {z ∈ D1 : z = F (0) for some automorphism F of (cid:101)V},
(cid:27)
O(0;(cid:101)V,(cid:102)W) :=
z = F (0) for some nc biholomorphism
z ∈ D2 :
(cid:26)
F of (cid:101)V onto(cid:102)W
.
Proposition 7.3. Let V ⊆ Bd and W ⊆ Be be two nc homogeneous varieties such that (cid:101)V
and(cid:102)W are nc biholomorphic. Then there exists an nc biholomorphism F of (cid:101)V onto(cid:102)W that
maps 0 to 0.
Proof. We can import the "disc trick" used in [13, Proposition 4.7] to the current setting
(see also [42, Lemma 5.9]). Since the argument is just a couple of paragraphs long, we
include it for completeness.
done. Assume that G(0) (cid:54)= 0. We will prove that there exists an nc biholomorphism F ,
Let G be an nc biholomorphism mapping (cid:101)V onto(cid:102)W. If 0 is mapped by G to 0, we are
mapping (cid:101)V onto(cid:102)W, such that F (0) = 0.
Lemma 7.2 implies there exist two discs D1 ⊆ (cid:101)V(1) and D2 ⊆(cid:102)W(1) such that G(D1) =
Since homogeneous varieties are invariant under multiplication by complex numbers, it is
easy to check that these sets are circular, that is, for every µ ∈ O(0;(cid:101)V) and ν ∈ O(0;(cid:101)V,(cid:102)W),
it holds that Cµ,D1 := {z ∈ D1 : z = µ} ⊆ O(0;(cid:101)V) and Cν,D2 := {z ∈ D2 : z = ν} ⊆
O(0;(cid:101)V, W).
Now, as G(0) belongs to O(0;(cid:101)V,(cid:102)W), we obtain that C := CG(0),D2 ⊆ O(0;(cid:101)V,(cid:102)W). There-
fore, the circle G−1(C) is a subset of O(0;(cid:101)V); note that this circle passes through the point
0 = G−1(G(0)). As O(0;(cid:101)V) is circular, every point of the interior of the circle G−1(C) lies in
O(0;(cid:101)V). Thus, the interior of the circle C must be a subset of O(0;(cid:101)V,(cid:102)W). But the interior
of C contains 0. We conclude that 0 ∈ O(0;(cid:101)V,(cid:102)W).
Proposition 7.4. Let V ⊆ Bd and W ⊆ Be be two homogeneous nc varieties. Let G : (cid:101)V →(cid:102)W
be an nc biholomorphism, such that G(0) = 0. Then for every X ∈ (cid:101)V we have that ρ(X) =
(cid:3)
ρ(G(X)).
Proof. Since similarities preserve the joint spectral radius, it suffices to prove that for every
X ∈ V, ρ(G(X)) = ρ(X). Let X ∈ V be irreducible. By Lemma 4.4 we know that
ρ(X) = min{(cid:107)S−1XS(cid:107) S ∈ GLn}. For the same reason, we may choose T ∈ GLn that
realizes the minimum and replace X by T −1XT and assume that ρ(X) = (cid:107)X(cid:107).
Now consider the function f (z) = G(zX/(cid:107)X(cid:107)) defined on the disc. Since G(0) = 0,
by Lemma 4.9 we know that ρ(G(X)) = ρ(f ((cid:107)X(cid:107))) ≤ (cid:107)X(cid:107) = ρ(X). Applying the same
consideration for G−1 we get that for every irreducible X ∈ Bd, ρ(X) = ρ(G(X)).
By Lemma 4.4 we know that if X is reducible, then ρ(X) = max{ρ(X1), . . . , ρ(Xk)},
where X1, . . . , Xk are the Jordan -- Holder components of X. Since G is an nc automor-
phism the Jordan -- Holder components of G(X) are precisely G(X1), . . . , G(Xk) and since
(cid:3)
we know that for each of them the spectral radius is preserved, we are done.
The following is the free analogue of [13, Lemma 7.5].
27
Corollary 7.5. Let V, W ⊆ Bd be two homogeneous nc varieties. Let A : Cd → Cd be an
invertible linear map, such that A maps (cid:101)V bijectively onto (cid:102)W. Then for every X ∈ V,
ρ(X) = ρ(A(X)).
Before stating the next proposition, we require some notation. For an ideal J(cid:47)C(cid:104)z1, . . . , zd(cid:105),
Z(J) = ZMd(J) = {X ∈ Md : p(X) = 0 for all p ∈ J}.
we write
Given a homogeneous variety V ⊆ Bd, we let JV be the corresponding homogeneous ideal
in C(cid:104)z1, . . . , zd(cid:105) consisting of polynomials vanishing on V.
Proposition 7.6. Let V ⊆ Bd and W ⊆ Be be two homogeneous nc varieties, and let G :
(cid:102)W → (cid:101)V be a 0-preserving nc holomorphic map. Then the linear transformation ∆G(0, 0)
maps(cid:102)W into (cid:101)V.
Proof. First, we prove that ∆G(0, 0) maps Z(JW) into Z(JV). For every W ∈ W(n) define
γW : D → Md(n) by
0 as well. Thus, there exists a holomorphic map βW : D → Md(n) such that
Then γW is a holomorphic map satisfying γW (D) ⊆ (cid:101)V(n). As G maps 0 to 0, γW maps 0 to
Now let f ∈ JV be an m-homogeneous function. As γW (D) ⊆ (cid:101)V(n) we have
γW (z) = z (γ(cid:48)
W (0) + zβW (z)) .
γW (z) := G(zW ).
0 = f (γW (z)) = f (z (γ(cid:48)
W (0) + zβW (z))) = zmf (γ(cid:48)
W (0) + zβW (z))
z
G(zW )
= ∆G(0, 0)(W ),
γ(cid:48)
W (0) = lim
z→0
z ∈ D. By Lemma 4.9 we get that
so ∆G(0, 0) maps W, and hence Z(JW), to Z(JV).
W (0) + zβW (z)) = 0 for every z ∈ D. Evaluating this at 0, we
W (0)) = 0. Since this is true for every homogeneous f ∈ JV, we must have that
for every z ∈ D, and thus f (γ(cid:48)
have f (γ(cid:48)
W (0) ∈ Z(JV). But
γ(cid:48)
Now, (cid:101)V = Z(JV) ∩ (cid:101)Bd, so in order to prove that ∆G(0, 0) maps (cid:102)W into (cid:101)V, it suffices
to show that ∆G(0, 0) maps(cid:102)W into (cid:101)Bd. Let X ∈ W(n) and consider the function f (z) =
G(zX/(cid:107)X(cid:107)). Since G(zX) ∈ (cid:101)V for every z ∈ D we know that ρ(f (z)) < 1 for every
ρ(∆G(0, 0)(X)) < 1. Thus by Lemma 4.5 we conclude that ∆G(0, 0)(X) ∈ (cid:101)Bd.
1(cid:107)X(cid:107)ρ(∆G(0, 0)(X)) = ρ(f(cid:48)(0)) ≤ 1, and, in particular,
(cid:3)
ible linear transformation mapping(cid:102)W onto (cid:101)V. Then the following statements are equivalent:
Proposition 7.7. Let V ⊆ Bd and W ⊆ Be be homogeneous varieties and let A be an invert-
(i) A is bi-Lipschitz with respect to δcb;
(iii) the mapping f (cid:55)→ (cid:101)f ◦ A, for f ∈ H∞(V), is a completely bounded isomorphism
(ii) supW∈W (cid:107)ΦA(W )(cid:107)cb < ∞ and supV ∈V (cid:107)ΦA−1(V )(cid:107)cb < ∞;
(iv) A is bi-Lipschitz with respect to δb;
(vi) the mapping f (cid:55)→ (cid:101)f ◦ A, for f ∈ H∞(V), is a bounded isomorphism H∞(V) → H∞(W);
(v) supW∈W (cid:107)ΦA(W )(cid:107) < ∞ and supV ∈V (cid:107)ΦA−1(V )(cid:107) < ∞;
H∞(V) → H∞(W);
and
28
(vii) the mapping f (cid:55)→ (cid:101)f ◦ A, for f ∈ H∞(V), is an algebraic isomorphism H∞(V) →
H∞(W).
Proof. The statements (i) -- (iii) are equivalent by Proposition 5.2, and the statements (iv) --
(vii) are equivalent by Proposition 5.5.
In addition, the equivalent statements (i) -- (iii)
clearly imply (iv) -- (vii), so we must show the opposite implication. Let α : H∞(V) →
H∞(W) be the bounded homomorphism of item (vi); we will show it is completely bounded.
Recall that by [43, Lemma 7.10], every nc holomorphic function f on an nc variety V
which is n-homogeneous -- namely, f (λZ) = λnf (Z) for every Z ∈ V and λ ∈ D -- is
a restriction to the variety of an n-homogeneous polynomial p, and (cid:107)f(cid:107)H∞(V) = (cid:107)f(cid:107)HV =
(cid:107)p(cid:107)H∞(Bd) = (cid:107)p(cid:107)H2
Let T be the densely-defined linear map HV → HW defined on a polynomial f by T f =
α(f ). As A is linear, T must be graded, namely, it maps n-homogeneous functions to n-
n fn.
homogeneous functions. Let f ∈ HV be a finite sum of homogeneous functions f =(cid:80)
Then obviously(cid:80)
n T fn is the homogeneous decomposition of T f and
.
d
(cid:107)T f(cid:107)2HW
(cid:88)
≤(cid:88)
=
n
(cid:88)
H∞(V) = (cid:107)α(cid:107)2(cid:88)
(cid:107)T fn(cid:107)2
=
n
H∞(W)
(cid:107)T fn(cid:107)2HW
(cid:107)α(cid:107)2(cid:107)fn(cid:107)2
(cid:107)fn(cid:107)2HV
n
= (cid:107)α(cid:107)2(cid:107)f(cid:107)2HV
n
.
Thus, T extends to a well defined bounded map HV → HW.
Now let kW,v,y be a kernel function in HW. Then, for every g ∈ HV we have
(cid:104)g, T ∗kW,v,y(cid:105) = (cid:104)(T g)(W )v, y(cid:105) = (cid:104)g(AW )v, y(cid:105) = (cid:104)g, kAW,v,y(cid:105).
Thus, T ∗kW,v,y = kAW,v,y.
Now, as H∞(V) ∼= Mult(HV) for every f ∈ H∞(V) there exists a bounded operator
Mf ∈ B(HV) such that f g = Mf g for every g ∈ HV. So for every kernel function kW,v,y ∈
HW we have
(T ∗)−1M∗
f T ∗kW,v,y = (T ∗)−1M∗
f kAW,v,y = kW,v,f (AW )∗y = M∗
α(f )kW,v,y.
Therefore
Mα(f ) = T Mf T −1.
Thus, α is given by similarity and is therefore completely bounded.
(cid:3)
Combining Theorem 7.1 and Propositions 7.6 and 7.7, we obtain the main result of this
section.
Theorem 7.8. Let V ⊆ Bd and W ⊆ Be be two homogeneous nc varieties. The following
statements are equivalent:
(i) H∞(V) and H∞(W) are weak-∗ continuously isomorphic.
(ii) H∞(V) and H∞(W) are boundedly isomorphic.
(iii) H∞(V) and H∞(W) are completely boundedly isomorphic.
(iv) There exists a δb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
(v) There exists a δcb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
29
M := sup
W∈W
(cid:107)ΦG(W )(cid:107) < ∞.
0. By Proposition 5.5, we have
Proof. By Corollary 6.3 and Proposition 7.7, we need to prove only one implication.
If α is a bounded isomorphism, then by Theorem 7.1 it induces a bi-Lipschitz biholo-
morphism G : (cid:102)W → (cid:101)V. By applying the analogue of Proposition 7.3 in the bi-Lipschitz
category, we may assume that G :(cid:102)W → (cid:101)V is a bi-Lipschitz biholomorphism that takes 0 to
Let A = ∆G(0, 0). By Proposition 7.6, A maps (cid:102)W into (cid:101)V . Our goal is to show that
supW∈W (cid:107)ΦA(W )(cid:107) < ∞; in fact we will show that supW∈W (cid:107)ΦA(W )(cid:107) ≤ M.
for every z ∈ D we have that zW ∈ W(n), and therefore, G(zW ) ∈ (cid:101)V ⊆ (cid:101)Bd, so by Lemma
r G(zW ) is an analytic function mapping D into (cid:101)Bd and 0 to 0, by the
4.5 ρ(G(zW )) < 1. Since D is compact s := supz∈D ρ(G(zW )) < 1, so we can choose some
s < r < 1. As z (cid:55)→ 1
r < 1. Thus, by Lemma 4.5 we see that u(z) ∈ (cid:102)Bd for all z ∈ D. Since (cid:101)V is homogeneous,
r G(zW )) ≤ z so that ρ(u(z)) <
Vesentini -- Schwarz Lemma (Lemma 4.9) we have that ρ( 1
u(z) = G(zW )/z belongs to Z(JV), so we conclude that u(z) ∈ Z(JV) ∩ (cid:102)Bd = (cid:101)V for all
Fix W ∈ W(n). We define u : D → Md by u(z) = G(zW )/z. Since W is homogeneous,
Next, we define a function F from D to the Banach space B(H∞(V), Mn) by F (z) =
Φu(z). We will show that this function is holomorphic. By applying some Mobius trans-
formation it suffices to show that F is differentiable at the origin. Let f ∈ H∞(V) with
(cid:107)f(cid:107) ≤ 1, and set g(z) := 1
2 (f (u(z)) − f (u(0))). Then g : D → Mn is a holomorphic function
with (cid:107)g(cid:107) ≤ 1 and g(0) = 0. By the Schwarz lemma, (cid:107)g(z)(cid:107) ≤ z, so we obtain
z ∈ D.
(cid:107)Φu(z) − Φu(0)(cid:107) ≤ 2z.
Now, the function f ◦ u is holomorphic in D so by Cauchy's integral formula for every
where dζ is the normalized Lebesgue measure of T. Since this is true for every (cid:107)f(cid:107) ≤ 1,
we have that for every z, w ∈ 1
D
and in particular limz→0
F (z)−F (0)
z
exists, so F is differentiable at the origin.
30
z (cid:54)= w ∈ 1
D
2
1
z − w
− Φu(w)(f ) − Φu(0)(f )
z
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Φu(z)(f ) − Φu(0)(f )
(cid:13)(cid:13)(cid:13)(cid:13)f (u(z)) − f (u(0))
(cid:16) 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:82)
ζ−z − 1
ζ=1 f (u(ζ))
z
z − w
z − w
=
=
w
1
1
z
(cid:18)
f (u(ζ))
ζ(ζ − z)(ζ − w)
ζ=1
(cid:13)(cid:13)(cid:13)(cid:13)
w
dζ
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
− f (u(w)) − f (u(0))
(cid:17)
(cid:82)
(cid:13)(cid:13)(cid:13)(cid:13) ≤ 4,
(cid:19)
(cid:13)(cid:13)(cid:13)(cid:13)Φu(z) − Φu(0)
−
dζ
w
1
z
ζ
ζ=1 f (u(ζ))
w
− Φu(w) − Φu(0)
− F (w) − F (0)
2
w
(cid:13)(cid:13)(cid:13)(cid:13) =
=
(cid:13)(cid:13)(cid:13)(cid:13)(cid:90)
(cid:13)(cid:13)(cid:13)(cid:13) F (z) − F (0)
z
ζ
dζ
(cid:17)
(cid:16) 1
ζ−w − 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13) ≤ 4z − w,
Now let βθ ∈ Aut(H∞(V)) be the (completely isometric) rotation automorphism
βθ(f )(V ) := f (e−iθV ),
f ∈ H∞(V), V ∈ V.
Then F (eiθ) = Φe−iθG(eiθW ) = ΦG(eiθW ) ◦ βθ. Thus, supz=1 (cid:107)F (z)(cid:107) ≤ M. But as F is holo-
morphic in D, it satisfies Cauchy's integral formula in the closed disc, which implies a
maximum principle:
(cid:107)F (0)(cid:107) ≤ maxz=1
(cid:107)F (z)(cid:107) ≤ M.
However, since F (0) = ΦA(W ), we have that (cid:107)ΦA(W )(cid:107) ≤ M for all W ∈ W. This completes
(cid:3)
the proof.
Question 7.9. Is a linear map between similarity envelopes of nc homogeneous varieties
automatically bi-Lipschitz?
Example 7.10. Let us consider a simple example that illuminates the rigidity of the non-
commutative case. Let V ⊆ B2 be the subvariety cut out by the single polynomial x2. Let
A be a 2× 2 matrix mapping (cid:101)V isomorphically onto some subvariety of (cid:101)B2. Note that V(1)
is the single line corresponding to the y-axis. Therefore, by [13, Lemma 7.5] A maps the
y-axis isometrically onto some line. Multiplying A on the left by a unitary we may assume
that A has the form A = ( a 0
c 1 ). We will show that in this case A is a diagonal unitary.
This point is a coisometry and thus has spectral radius 1. By Corollary 7.5 we know that
A preserves the spectral radius and thus the image must also have spectral radius 1. The
matrix of the associated completely positive map with respect to the standard basis of M2
is
.
Z =
0 0
1 0
(cid:21)
(cid:21)
, W =
First consider the point
(cid:20)0 1
(cid:20)0 0
0 0 0 a2 + c2
.
Its spectral radius is(cid:112)a2 + c2 and thus we have that a2 + c2 = 1.
(cid:20)0 δ
(cid:20)0 e−iθ
It is immediate that XX∗ + Y Y ∗ = (cid:0) 1 0
0 0 c
0 c 0
1 0 0
X =
0
0
0
0
0
0 δ2
(cid:21)
(cid:21)
(cid:1) and therefore this point lies on the boundary
, Y =
δ 0
.
Now, let , δ > 0, be such that 2 + δ2 = 1. Write c = reiθ and consider the point
of the ball. Furthermore, a short computation of the associated completely positive map
shows that the spectral radius of this point is 0 <
δ < 1. By Corollary 7.5 A maps the
point (X, Y ) to a point with spectral radius
√
δ. The image of this point is
√
A(X, Y ) =
(cid:21)
(cid:21)(cid:19)
(cid:20)0 δ + r
(cid:18)(cid:20)0 ae−iθ
δ 4(cid:112)(δ + r)2 + a22. Thus (δ + r)2 + a22 = 1.
0
√
0
0
δ
.
,
31
Computing again the associated completely positive map and its spectral radius we get
that the spectral radius of the point is
Opening brackets and using our assumptions and the fact that a2 + r2 = 1 we get that:
1 = (δ + r)2 + a22 = 1 − 2 + 2rδ + r22 + a22 = 1 + 2rδ.
This is possible if and only if r = 0 and thus A is a diagonal unitary.
Combining this with Theorem 7.8 we see that the only varieties that are isomorphic to
the variety cut out by x2 are those that are conformally equivalent to it via a unitary.
8. EXAMPLE: THE q-COMMUTATION VARIETIES
We will first show that unless p ∈ {q, 1
q}, there is no 2 × 2 matrix mapping(cid:102)Vq to (cid:102)Vp, and
In quantized versions of classical algebras, relations such as xy = qyx -- also known
as q-commutation -- naturally appear. Let q ∈ C and let Vq ⊆ B2 be the homogeneous
variety generated by the ideal (cid:104)z1z2 − qz2z1(cid:105).
therefore for q ∈ D ∪ {eiθ : 0 ≤ θ ≤ π} we get a continuum of mutually non-isomorphic
Banach algebras H∞(Vp).
Theorem 8.1. Let q (cid:54)= p be two complex numbers. If p = 1
(i) the varieties Vq and Vp are conformally equivalent, and
(ii) the operator algebras H∞(Vq) and H∞(Vp) are unitarily equivalent;
and otherwise,
(i) the varieties (cid:102)Vq and (cid:102)Vp not biholomorphically equivalent, and
q , then
(ii) the Banach algebras H∞(Vq) and H∞(Vp) are not boundedly isomorphic.
Proof. For the first case, note that for every q (cid:54)= 0, the unitary matrix
morphism of B2 -- maps Vq onto V 1
as a result H∞(Vq) and H∞(V 1
. Thus, Vq and V 1
) are unitarily equivalent.
1 0
-- an auto-
are conformally equivalent, and
(cid:20)0 1
(cid:21)
q
q
We now move to the second case. Note that if q = 1, then the first level of Vq -- a
variety that is considered in the next section and is called the commutative nc ball -- is B2
while the first level of every other Vq is only the cross C := {(x, y) ∈ B2 : xy = 0}. So we
may assume that q (cid:54)= 1.
Now let p, q ∈ D ∪ {eiθ : 0 < θ ≤ π}, and assume there is a 2 × 2 invertible matrix A
mapping Vq onto Vp. A simple computation shows that a matrix A, mapping the cross C
q
onto itself, must be of the form (cid:20)λ1
(cid:21)
for some λ1, λ2 ∈ T. The first option maps Vq onto itself, and the second option maps Vq
onto V 1
, so we only need to show that Vq (cid:54)= Vp for q (cid:54)= p. Indeed, the point
q
(cid:21)
λ2
(cid:20) 0 λ1
(cid:21)(cid:19)
(cid:20)0 1
0
0
0 λ2
or
(cid:18)(cid:20)q 0
(cid:21)
2(cid:112)q2 + 1
1
,
0 1
0 0
∈ B2
(cid:3)
belongs to Vq but not to Vp.
We will now examine the existence of irreducible points in the varieties Vq. This will
later help us understand maps of the corresponding operator algebras.
32
Proposition 8.2. If q is not a root of unity, then Vq has no non-scalar irreducible points. If
q is a root of unity, primitive of order k, then it has non-scalar irreducible points only in the
k-th level. In the latter case, they are all similar to a point of the form
λ
1
0
q
...
0
...
...
0 ···
···
...
q2
...
···
···
...
...
0
,
0
...
...
0
qk−1
0 ···
...
1
...
0
...
...
0 ···
···
0 µ
...
...
0
...
1
0
...
...
0
for some non zero λ and µ.
Proof. Before we start, for A ∈ Mn and v ∈ Cn, let C(A; v) denote the A-cyclic subspace of
Cn, namely, the linear subspace generated by {v, Av, A2v, . . . , An−1v}.
Let n > 1 and (X, Y ) ∈ Vq(n). First note that if (X, Y ) is irreducible, then X cannot
Indeed, if v (cid:54)= 0 is an eigenvector of X corresponding to an
have 0 as an eigenvalue.
eigenvalue 0, then C(Y ; v) ⊆ ker X. So if dimC(Y ; v) = n, then X = 0, and otherwise
C(Y ; v) is a nontrivial invariant subspace for both X and Y ; in any case, this contradicts
the irreducibility of (X, Y ). If q (cid:54)= 0, a symmetric argument shows that Y cannot have 0 as
an eigenvalue too.
If q = 0, then as X is invertible, the equality XY = 0 implies Y = 0, which contradicts
the irreducibility of (X, Y ). Thus, the only irreducible points of V0 are the ones in the first
level. So henceforward, we assume that q (cid:54)= 0.
Suppose first that 1, q, q2, . . . , qn are n + 1 distinct numbers, namely, that q is not a root
of unity of order less than or equal to n. Let λ be a nonzero eigenvalue of X with an
eigenvector v. Then it is easy to check inductively that for every m, Y mv is an eigenvector
of X for the eigenvalue λqm, which is of course impossible.
In particular, if q is not a root of unity, then the only irreducible points of Vq are the
ones in the first level, and if q is a root of unity, say primitive of order k > 1, then for every
1 < l < k, the l'th level Vq(k) does not contain irreducible points.
We will now show that in the latter case -- where q is a root of unity primitive of order
k -- there exist irreducible points in the k-th level, and we will give a full description of
them. Let λ be an eigenvalue of X, and recall that for every 0 ≤ m ≤ k − 1, Y mv is an
eigenvector of X with an eigenvalue λqm. Thus, the vectors v, Y v, Y v, . . . , Y k−1v form a
basis. Let S be the invertible matrix whose rows are the latter vectors (in the order they
were written). Additionally, note that as Y kv is an eigenvector of X with an eigenvalue
λqk = λ, there must be a scalar µ such that Y Y k−1v = µv. Thus,
1
0
q
...
S−1XS = λ
0
...
...
0 ···
and µ must be non-zero.
···
...
q2
...
···
···
...
...
0
0
...
...
0
qk−1
and
S−1Y S =
33
0 ···
...
1
...
0
...
...
0 ···
,
···
...
...
0
0 µ
0
...
...
0
...
1
On the other hand, it is evident that the above pair of matrices constitutes a point in
Z(JVq ). To see that (X, Y ) is irreducible one may argue as follows. Begin by observing
that det(XY − Y X) = (q − 1)k det X det Y (cid:54)= 0. Since for each two upper triangular
matrices R and T the matrix RT − T R is strictly upper triangular its determinant is 0, so
the pair (X, Y ) cannot be similar to a triangular pair (R, T ). As there are no irreducible
points in levels 1 < m < k, the point (X, Y ) itself must be irreducible.
Finally, we will show -- still in the case that q is a primitive root of unity of order k --
that levels higher than k contain no irreducible points.
Indeed, assume that n > k and (X, Y ) ∈ Vq(n) is irreducible. Let λ be an eigenvalue of
X (so λ (cid:54)= 0) with an eigenvector v. As before, consider the Y -cyclic subspace C(Y, v). If
its dimension is not n, then this subspace is a nontrivial invariant subspace for both X and
Y . Otherwise, the vectors v, Y v, Y 2v, . . . , Y n−1v form a basis. Rearrange them such that
all the eigenvectors in the list corresponding to the eigenvalue λ (i.e., v, Y kv, . . . ) comes
first, then the ones corresponding to the eigenvalue λq (i.e., Y v, Y k+1v, . . . ) , and so on.
For each eigenvalue, keep them arranged in their original increasing form. Let S be the
invertible matrix whose rows are the latter vectors. Note also that Y nv is an eigenvector
of X with an eigenvalue λqn. Thus, if k does not divide n, the first row of S−1Y S must be
zero, which is impossible. Thus, kn.
Let l = n/k. Since Y nv is an eigenvector of X with an eigenvalue λ, there are scalars
µ0, . . . µl−1 ∈ C such that Y nv =(cid:80)l−1
Now let 0 (cid:54)= v ∈ Cn be some eigenvector of A. Then, S(Cv)⊕k ⊆ Cn is a k-dimensional
(cid:3)
subspace invariant under both X and Y .
Let us now consider a change of angle in the varieties Vq. More precisely, let Wq ⊆ B2
be the nc variety generated by the ideal (cid:104)(z1 − z2)z2 − qz2(z1 − z2)(cid:105), and let
Then obviously, for every q (cid:54)= 1, A maps Vq(1) onto Wq(1) and Z(JVq ) onto Z(JWq ).
in (cid:102)Vq are only the scalar points. Now, since every (X, Y ) ∈ M2(n) is similar to a block
Suppose first that q is not a root of unity. By Proposition 8.2, the irreducible points
and
(cid:20)1
0
(cid:21)
.
1√
2
1√
2
A =
34
···
j=0 µjY jk+1v. Let
0 ···
µ0
...
1
...
0
...
...
0 ···
0
...
...
...
... 0
0
µ1
...
...
1 µl−1
A :=
···
...
q2Il
...
···
···
...
...
0
0
...
...
0
qk−1Il
∈ Ml(C).
S−1Y S =
.
0 A
...
...
0
Il
0
...
...
0
0 ···
...
Il
...
0
...
...
0 ···
···
...
...
0
Then
S−1XS = λ
Il
0
qIl
...
0
...
...
0 ···
This argument works in reverse, and we conclude that for every q which is not a root
upper triangular pair whose diagonal blocks are all irreducible, every irreducible (X, Y ) ∈
(cid:102)Vq must be similar to an upper triangular matrix whose diagonal entries are in Vq(1).
Therefore, as A maps Vq(1) onto Wq(1), Lemma 4.4 implies that for every (X, Y ) ∈ (cid:102)Vq,
we have ρ(A(X, Y )) < 1, so in view of Lemma 4.5, A(X, Y ) ∈ (cid:102)B2, so A maps (cid:102)Vq into (cid:102)Wq.
of unity A maps (cid:102)Vq bijectively onto (cid:102)Wq. Unfortunately, we couldn't prove that A is bi-
in the first level, it might happen that A does not map (cid:102)Vq bijectively onto (cid:102)Wq.
Lipschitz; if we could, this would have shown that H∞(Vq) and H∞(Wq) are completely
boundedly isomorphic.
If q is a root of unity, say primitive of order k, then since irreducible points exist not only
Indeed, consider the case q = −1. Let
(cid:20)1
(cid:21)
X0 =
0
0 −1
and
Y0 =
(cid:20)0 i
(cid:21)
,
1 0
ρ(X0, Y0). Corollary 7.5 would then imply that (cid:103)V−1 is not mapped bijectively onto (cid:103)W−1.
which by Proposition 8.2, is an irreducible point in Z(JVq ). We will show that ρ(A(X0, Y0)) (cid:54)=
To this end, recall that for a pair of two n × n matrices (A, B), the linear transformation
on Θ(A,B) : Mn → Mn, defined by Θ(A,B)(T ) = AT B, is represented, with respect to the
standard basis E1,1, E1,2, . . . , En,n, by the matrix A ⊗ BT . Thus, the matrix representing
Ψ(X,Y ) -- which is by definition Θ(X,X∗) + Θ(Y,Y ∗) -- is
X ⊗ X + Y ⊗ Y .
The square root of the largest number among all eigenvalue absolute values of the latter
n2 × n2 matrix is its spectral radius, and hence is equal to ρ(X, Y ).
Now, the matrix representing Ψ(X0,Y0) is
1
0
i
0
1
0 −1
0
0 −i −1 0
1
1
0
0
,
1 − 1√
−1
1√
−i
2
1√
1 − 1√
2
i
2
2
.
i
1
− 1√
1√
2
1
2
i
i
1√
2
i
−1
− 1√
2
X0 ⊗ X0 + Y0 ⊗ Y0 =
and the matrix representing ΨA(X0,Y0) is
(X0 +
Y0) ⊗ (X0 +
1√
2
1√
2
Y0) +
1√
2
Y0 ⊗ 1√
2
Y0 =
√
The largest number among all eigenvalue absolute values of the first matrix is 2, and of
3. Thus, ρ(A(X0, Y0)) (cid:54)= ρ(X0, Y0). We conclude that H∞(V−1) is not
the second --
boundedly isomorphic to H∞(W−1)
Similar computations show that V
spectral radius is not preserved under A. Thus, H∞(V
e
isomorphic to H∞(W
and Vi also contain irreducible points whose joint
) and H∞(Vi) are not boundedly
) and H∞(Wi).
2πi
2πi
e
2πi
3
3
e
3
35
9. THE COMMUTATIVE CASE
In this section, we study the isomorphism problem for the algebras of bounded analytic
functions on commutative nc varieties. This problem was treated extensively in the fully
commutative case, that is, when the algebra H∞(V) lives on a variety V which is minimal,
in the sense that it is the minimal nc variety in Bd which contains V = V(1); this is referred
to as the isomorphism problem for complete Pick algebras, see [10, 13, 14, 21, 27, 28, 42]. In
[43, Section 11] we explained how this problem can be investigated in the nc commutative
setting (cf. Remark 9.6 below).
As in Section 7, we assume d < ∞. Let us define
CMd = {X ∈ Md : XiXj = XjXi , i, j = 1, . . . , d}
and
CBd = {X ∈ Bd : XiXj = XjXi , i, j = 1, . . . , d}.
Our goal is, as before, to classify the algebras H∞(V), but this time only for nc subvarieties
V ⊆ CBd; we shall refer to such nc varieties as commutative nc varieties. Obviously,
the operator algebra H∞(V) of a commutative nc variety V is commutative. Note that
H∞(CBd) = Md -- the multiplier algebra on the Drury -- Arveson space; see [43, Section
11] for a detailed explanation of this fact.
In the setting of commutative nc varieties, we relax somewhat the weak-∗ continuity
conditions of our basic classification results from Section 6. When we specialize further
to homogeneous commutative nc varieties, we will recover the results of Section 7 by
other techniques. An important ingredient for the proof of the classification result is a
Nullstellensatz for homogeneous commutative nc varieties, which may be of independent
interest.
Let us define a new nc set (cid:100)CBd by(cid:100)CBd := {X ∈ CMd : σ(X) ⊆ Bd}.
Here σ(X) could mean Taylor spectrum, but since X is a tuple acting on a finite dimen-
sional space, σ(X) can be defined by any other reasonable definition of spectrum. Our
working definition for the joint spectrum σ(X) will be the set of d-tuples in Cd made from
taking elements from the diagonal of (X1, . . . , Xd) after they are realized in a simultaneous
upper triangular form.
to a continuous representation of the topological algebra O(Bd) consisting of all analytic
functions on Bd [48, Theorem 4.3].
For every X ∈ (cid:100)CBd, the mapping f (cid:55)→ f (X), which is well defined on C[z], extends
Theorem 9.1. Every X ∈ (cid:100)CBd is simultaneously similar to some Y ∈ CBd. In other words,
(cid:100)CBd = (cid:103)CBd.
Fix X ∈ (cid:100)CBd(n). Since the map EVX : O(Bd) → Mn given by evaluating at X is
Proof. This theorem follows from the results of Section 4.2, but we wish to present a dif-
ferent proof that is also interesting.
continuous, there must be a compact set K ⊆ Bd and a constant C for which
(cid:107)f (X)(cid:107) ≤ C(cid:107)f(cid:107)∞,K,
36
where (cid:107)f(cid:107)∞,K = sup{f (x) : x ∈ K}. Considering A, the closure of (O(Bd),(cid:107) · (cid:107)∞,K), as a
subalgebra of C(K), it is evident that it is a unital operator algebra. The evaluation map
EVX : O(Bd) → Mn extends to a bounded unital map from A into Mn. Since Mn is finite
dimensional, the evaluation map must be completely bounded [15, Corollary 2.2.4]. By
Paulsen's Theorem [29, Theorem 9.1], EVX is similar to a completely contractive homo-
morphism ρ. Since (C[z],(cid:107) · (cid:107)∞,K) is dense in A, the homomorphism ρ must be of the form
EVY for Y = (ρ(z1), . . . , ρ(zd)) ∈ CMd. Finally, to see that Y must be in CBd, we note that
the row norm of (z1, . . . , zd) is
(cid:113)(cid:88)xi2 < 1,
(cid:107)(z1, . . . , zd)(cid:107) = sup
x∈K
0
.
t
so, because ρ is a complete contraction, (cid:107)(Y1, . . . , Yd)(cid:107) < 1.
Remark 9.2. The above proof exposes the connection between the operator space com-
pletions of O(Bd) and the representations at hand. However, one can provide a more
elementary proof by induction as follows. For every X ∈ (cid:100)CBd(1) = Bd the claim is trivial.
Now let X ∈ (cid:100)CBd(n) and choose a unitary, such that U∗XU is upper triangular. Write
0 X(cid:48) ), where α ∈ Bd, v is some vector and X(cid:48) ∈ (cid:100)CBd(n − 1). By induction, there
matrix Tt = (cid:0) t
(cid:1). Since α ⊕ (S−1X(cid:48)S) is a strict
X = ( α v
exists a matrix S, such that S−1X(cid:48)S is a strict contraction. Take for t > 0 the invertible
contraction, for t small enough, so is TtU∗XU T −1
9.1. The isomorphism problem in the commutative case.
Proposition 9.3. Let V ⊆ CBd and W ⊆ CBe be two commutative nc varieties, and let
α : H∞(V) → H∞(W) be a bounded homomorphism. Suppose that α∗ maps every evaluation
functional Φw : f (cid:55)→ f (w) (f ∈ H∞(W), w ∈ W = W(1)) to an evaluation functional on
(cid:1), then TtU∗XU T −1
t = (cid:0) α t2vS
0 S−1X(cid:48)S
0 t−1S−1
G(T ) = πk(α∗(ΦT )) ,
Since α∗ is a similarity preserving and a direct sum preserving map between the nc spaces
H∞(V). Then α∗ maps(cid:102)W into (cid:101)V.
Proof. Define G :(cid:102)W → π((cid:116)k Repk(H∞(V))) by
of representations, it follows that G is an nc map from(cid:102)W into π((cid:116)k Repk(H∞(V))) = (cid:101)V.
Let T ∈ (cid:102)W. We know that G(T ) ∈ (cid:101)V, and we need to show that G(T ) ∈ (cid:101)V. Since T is
But as T ∈ (cid:102)W(k), we know that σ(T ) ⊆ W = W(1). By assumption, G(σ(T )) ⊆ V =
V(1) ⊆ Bd, so Theorem 9.1 implies that G(T ) ∈ (cid:103)CBd. Thus α∗(ΦT ) = ΦG(T ), and the proof
simultaneously upper-triangular in some basis, and since nc maps act entry-wise on the
diagonal, we find that
T ∈(cid:102)W(k).
σ(G(T )) = G(σ(T )).
(cid:3)
(cid:3)
is complete.
Recall that a character of an algebra is a non-zero linear multiplicative functional.
Proposition 9.4. Let V ⊆ CBd and W ⊆ CBe be two commutative nc varieties, and let
α : H∞(V) → H∞(W) be a unital and bounded homomorphism. Suppose that α∗ maps weak-
∗ continuous characters to weak-∗ continuous characters. Then there exists an nc holomorphic
37
map G :(cid:102)W → (cid:101)V which implements α by the formula
α(f ) = (cid:101)f ◦ G
for all f ∈ H∞(V) where (cid:101)f denotes the unique extension of f to (cid:101)V.
G :(cid:102)W → (cid:101)V. The conclusion now follows from Proposition 6.2 and Theorem 3.4.
Proof. Applying Proposition 9.3 to α, we find that α∗ restricts to an nc holomorphic map
(cid:3)
Putting the conclusion of the previous Proposition with Propositions 5.5 and 6.2, we
obtain the following strengthening of Corollary 6.3.
Corollary 9.5. Let d, e ∈ N, and V ⊆ CBd and W ⊆ CBe be commutative nc varieties. Then
H∞(V) and H∞(W) are continuously isomorphic via an isomorphism that preserves weak-∗
continuous characters if and only if the similarity envelopes (cid:101)V and(cid:102)W are nc biholomorphic
via an nc biholomorphism G :(cid:102)W → (cid:101)V satisfying
(cid:107)ΦG(W )(cid:107) < ∞ and
(cid:107)ΦG−1(V )(cid:107) < ∞.
sup
V ∈V
sup
W∈W
Likewise, H∞(V) and H∞(W) are completely boundedly isomorphic via an isomorphism
Furthermore, in this case supW∈W (cid:107)ΦG(W )(cid:107) = (cid:107)α(cid:107).
that preserves weak-∗ continuous characters if and only if the similarity envelopes (cid:101)V and(cid:102)W
are nc biholomorphic via an nc biholomorphism G :(cid:102)W → (cid:101)V satisfying
(cid:107)ΦG(W )(cid:107)cb < ∞ and
(cid:107)ΦG−1(V )(cid:107)cb < ∞,
sup
V ∈V
sup
W∈W
and in this case supW∈W (cid:107)ΦG(W )(cid:107)cb = (cid:107)α(cid:107)cb and supV ∈V (cid:107)ΦG−1(V )(cid:107)cb = (cid:107)α−1(cid:107)cb.
Remark 9.6. Let Bd and Be be the commutative (classical) open unit balls in d and e
dimensions, respectively. Let V ⊆ Bd and W ⊆ Be be two subvarieties, meaning that V and
W are both given as zero sets of multipliers of the Drury -- Arveson space. Let V and W be
the minimal nc varieties containing V and W as their first level. The "quantized" varieties
V and W are clearly commutative nc varieties, and the algebras H∞(V) and H∞(W) are
the multiplier algebras MV and MW studied in [13] and [14]. In [13] the authors show
that if MV and MW are isomorphic via a map that preserves weak-∗ continuous characters,
then the varieties V and W are multiplier biholomorphic, in the sense that the varieties
are biholomorphic via maps G and G−1 whose coordinates are multipliers. Multiplier
biholomorphism, however, is not even an equivalence relation; see [10, Remark 6.7]. This
raises the interesting open problem of describing an equivalence relation on varieties that
encodes the isomorphism classes of algebras of the form MV ; see [42, Subsection 7.3].
Before handling this problem from our perspective, we wish to understand how "variety
quantization" (i.e., the passage from a variety W ⊆ Bd to the smallest nc variety W ⊆ Bd
containing it as its first level W(1) = W ) behaves with respect to maps between varieties.
(It is perhaps worth stressing that here, as in the rest of the paper, by nc variety we mean
an nc variety that is cut out by bounded nc holomorphic functions.) Let g : W → V be a
holomorphic map. Since the spectrum σ(X) of every X ∈(cid:102)W is a subset of W , the function
g extends to an nc holomorphic map G : (cid:102)W → (cid:103)CBd satisfying σ(G((cid:102)W )) ⊆ V . We claim
G maps(cid:102)W into (cid:101)V, where V is the smallest nc variety which has V(1) = V . To see why this
that if the latter map G is δb-Lipschitz, or equivalently, if supX∈W (cid:107)ΦG(X)(cid:107) < ∞, then in fact
38
is true, consider X ∈ (cid:102)W. We wish to show that f (G(X)) = 0 every bounded nc function
and vanishes on W (because G(W ) = g(W ) ⊆ V ). From the definition of(cid:102)W it follows that
f ◦ G vanishes on(cid:102)W, whence f (G(X)) = 0, as required.
f ∈ H∞(Bd) that vanishes on V . But for every such f, the composition f ◦ G is in H∞(W)
Categorically speaking, if the morphisms in the category of commutative radical varieties
are assumed to be Lipschitz on a quantized variety, then the quantized mapping maps the
quantized variety to the quantization of the range, and so this quantization becomes a
functor into the category of nc varieties with Lipschitz nc holomorphic maps.
We now see, due to Corollary 9.5, that the "right" equivalence relation between varieties
V and W corresponding to weak-∗ isomorphism between the algebras MV and MW , is a
bi-Lipschitz biholomorphism of their quantizations. We do not know whether there exists
a characterization of this equivalence relation that can be read directly from the original
varieties V and W .
9.2. A Nullstellensatz for homogeneous commutative nc varieties. We now prove a
Nullstellensatz that will be important for our classification result. These results were es-
sentially obtained already in [40, Section 2.1.6] (and in [13, Section 6] in the norm closed
setting), but there are some modifications we need, so we expand. The reader is referred
to [43, Section 7] for a similar discussion in the noncommutative setting.
We define, given a subset S ⊆ CBd,
JS = {f ∈ H∞(Bd) : f (X) = 0 for all X ∈ S}.
Since we are dealing with the free commutative case, it is convenient to define
J c
S = {f ∈ Md = H∞(CBd) : f (X) = 0 for all X ∈ S},
as well as the polynomial ideal
I(S) = {p ∈ C[z1, . . . , zd] : p(X) = 0 for all X ∈ S}.
Theorem 9.7. Let V ⊆ CBd be a homogeneous commutative nc variety and let V = V(1).
There exists an integer N such that for every f ∈ Md,
Likewise, for every f ∈ H∞(Bd)
f(cid:12)(cid:12)V = 0 =⇒ f N(cid:12)(cid:12)V = 0.
f(cid:12)(cid:12)V = 0 =⇒ f N(cid:12)(cid:12)V = 0.
In other words, f ∈ J c
Proof. The ideal J c
V is homogeneous. Therefore, if f ∈ J c
is in the polynomial ideal I(V) (cid:47) C[z1, . . . , zd]. It holds that
f (eintX)e−intdt,
V implies that f N ∈ J c
V and similarly for JV and JV.
(cid:90) 2π
V and f =(cid:80)∞
fn(X) =
n=0 fn, then every fn
(cid:80) fn are bounded and converge pointwise to f. Thus every f ∈ J c
and a familiar argument using the Fej´er kernel shows that the Ces`aro sums of the series
V is the weak-∗ limit of
0
a bounded sequence in I(V).
39
By Hilbert's Basis Theorem, the radical I(V ) =(cid:112)I(V) of the ideal I(V) (cid:47) C[z] is finitely
generated. Combining this with Hilbert's Nullstellensatz, it is easy to see that there is some
N such that for every p ∈ C[z1, . . . , zd],
p(cid:12)(cid:12)V = 0 =⇒ pN ∈ I(V).
V) = JV.
If f ∈ J c
V , that is, if f ∈ Md vanishes on V , then, by the first paragraph of the proof, f is
the pointwise limit of a bounded sequence of polynomials qk ∈ I(V ). It follows that f N is
the bounded pointwise limit of qN
That proves the statement for f ∈ Md. Now let f ∈ H∞(Bd), and suppose that f(cid:12)(cid:12)V = 0.
Letting q : H∞(Bd) → Md = H∞(CBd) be the quotient map, we find that q(f )(cid:12)(cid:12)V = 0, so
k ∈ I(V), and therefore f N ∈ J c
V.
V. Therefore, f N ∈ q−1(J c
q(f N ) = q(f )N ∈ J c
9.3. Classification of homogeneous commutative nc varieties. The following theorem
is a special case of Theorem 7.1, when attention is restricted to commutative homoge-
neous nc varieties; we believe that it deserves to be presented separately, because of the
importance of the commutative case, and also because we have a different proof.
Theorem 9.8. Let V ⊆ CBd and W ⊆ CBe be two homogeneous commutative nc varieties,
and let α : H∞(V) → H∞(W) be a bounded isomorphism. Then there exists an nc biholo-
morphism G :(cid:102)W → (cid:101)V which implements α by the formula
for all f ∈ H∞(V) where (cid:101)f denotes the unique extension of f to (cid:101)V. Moreover,
α(f ) = (cid:101)f ◦ G
(cid:3)
(cid:107)ΦG(W )(cid:107) = (cid:107)α(cid:107),
sup
W∈W
and if α is completely bounded, then
sup
W∈W
(cid:107)ΦG(W )(cid:107)cb = (cid:107)α(cid:107)cb.
Proof. By Corollary 9.5, it suffices to prove that α∗ restricts to a bijection between the
weak-∗ continuous functionals. In fact, it suffices to show that α∗ maps W = W(1) into
V = V(1), since by symmetry this will also be true for α−1.
Suppose, therefore, that α∗ maps an evaluation functional Φw0 to a functional ρ =
α∗(Φw0) which is not an evaluation functional. By Theorem 3.4, ρ lies in a fiber over a
point of the boundary of the ball, that is ρ ∈ π−1
1 (v) for some v ∈ ∂Bd ∩ V . Since W is
connected, we can use the proof of [14, Lemma 5.3] to obtain that α∗(Φw) ⊆ π−1
1 (v) for all
w ∈ W . We will now show that this leads to a contradiction.
Without loss of generality, we may assume that v = (1, 0, . . . , 0). Now let g = z1 − 1,
which can be considered as a function in Md and also as a function in H∞(W). Then
gN does not vanish on V for any N ∈ N, since A − I is invertible for any matrix A with
(cid:107)A(cid:107) < 1. On the other hand, for every w ∈ W ,
α(g)(w) = α∗(Φw)(g) = α∗(Φw)(z1) − α∗(Φw)(1) = 1 − 1 = 0.
Note that this does not imply yet that α(g) is zero on W, just that it vanishes on the first
level. However, by Theorem 9.7, there exists N such that α(gN ) = α(g)N = 0, and this
contradicts the fact that α is an isomorphism. This contradiction shows that α∗ must map
(cid:3)
W into V , and by the first paragraph of the proof, we are done.
40
Combining Theorem 9.8 with Propositions 7.3, 7.6 and 7.7, we recover Theorem 7.8
for the case of commutative nc varieties (where Theorem 9.8 replaces Theorem 7.1 in the
proof). Proposition 7.6 can be given a slightly different proof in the commutative case,
using the joint spectrum instead of the joint spectral radius (we omit the details), and with
this approach the proof of the commutative case becomes different in a meaningful way.
We know that for homogeneous nc varieties, and in particular for commutative homo-
geneous nc varieties, the algebras H∞(V) and H∞(W) are boundedly isomorphic if and
only if there exists a δb-bi-Lipschitz linear map mapping (cid:102)W onto (cid:101)V. In Question 7.9 we
asked whether every bijective linear map taking(cid:102)W onto (cid:101)V is automatically δb-bi-Lipschitz.
We were unable to resolve the question. The answer is known only when the varieties are
commutative and minimal in a certain sense, as explained in the following example.
Example 9.9. The answer to Question 7.9 is affirmative in the case that one (and hence
both) of the varieties are the nc zero sets of a radical homogeneous ideal. This follows from
a difficult result of Hartz [20] (see Remark 9.6 or [43, Section 11] for the connection). In
our notation, Hartz's theorem says that if A : W → V a bijective linear map between
two homogeneous varieties in the unit ball Bd ⊆ Cd, and if V and W are the minimal nc
varieties in the nc ball Bd containing V and W as their first levels, then pre-composing
with A gives rise to a bounded isomorphism from H∞(V) onto H∞(W). Recall that every
nc holomorphic map(cid:102)W → (cid:101)V that gives rise to a homomorphism H∞(V) → H∞(W) by
bijective linear map A :(cid:102)W → (cid:101)V is a bijective linear map A : W → V .
pre-composition is Lipschitz (Proposition 5.5). Thus, Hartz's result provides a solution to
Question 7.9 in the commutative, radical case, because the restriction to the first level a
10. ALGEBRAS OF CONTINUOUS MULTIPLIERS
Let V ⊆ Bd be an nc variety. Recall that we let A(V) denote the algebra of all bounded
analytic functions that extend to uniformly continuous functions on V, in the sense that for
every > 0, there exists a δ > 0, such that for every n ∈ N we have that if X, Y ∈ V(n) are
such that (cid:107)X − Y (cid:107) < δ, then (cid:107)f (X) − f (Y )(cid:107) < (in other words, these are the bounded
analytic functions on V that are uniformly continuous on V(n), for all n, uniformly in
n). We give A(V) the sup norm (which coincides with the multiplier norm), and this
gives A(V) the structure of an operator algebra. By [43, Section 9], these algebras can be
considered to be continuous multipliers on the nc reproducing kernel Hilbert spaces HV.
The algebras of the form A(V) include among them many algebras that have been in-
vestigated before. For example, if V is a disc, then A(V) = A(D) is the disc algebra. If
V = Bd is an nc ball, then A(V) = A(Bd) is the noncommutative disc algebra [31]. If
V is a homogeneous ideal, then A(V) is the tensor algebra associated with a subproduct
system (with Hilbert space fibers), as studied in [13], [23] and [44].
A natural question, in the spirit of our investigations, is: how does the (operator/Banach)
algebraic structure of A(V) reflect the geometric structure of V? Other natural algebras
one might consider are (i) the closure AV of the algebra generated by the polynomials,
with respect to the supremum norm and (ii) the quotient algebra A(Bd)/IV, where
IV = {f ∈ A(Bd) : f (X) = 0 for all X ∈ V}.
In the general case, several technical difficulties immediately arise. First, it is not true
that every f ∈ A(V) has an extension F ∈ A(Bd) such that (cid:107)F(cid:107) = (cid:107)f(cid:107). Moreover, it
41
might happen that IV = {0} for a nontrivial V (see [14, Theorem 8.1]).
In this case,
the restriction map A(Bd) → A(V) is an isomorphism, hence the algebraic structure is
preserved while the geometry is dramatically changed. It is also not a simple matter to
determine when A(V) = AV. We refer the reader to Section 7 of the paper [14] for some
discussion of the kind of subtleties involved.
To recap, in the setting of multiplier algebras, we had the natural and completely iso-
metric identifications
H∞(V) = H∞(Bd)(cid:12)(cid:12)V = H∞(Bd)/JV = alg
A(V) = A(Bd)(cid:12)(cid:12)V = A(Bd)/IV = AV := alg
(cid:107)·(cid:107)
where the last algebra denotes the weak-operator (or weak-∗) closure of the unital algebra
generated by the free polynomials in MultH2
V; whereas, when passing to the norm closed
algebras, we do not know in general whether
(z1, . . . , zd)
(10.1)
holds. In particular, we cannot identify the spaces Repk(A(V)) of finite dimensional repre-
sentations, which in the previous context has been the starting point of the classification.
In [43, Section 9], we proved that the identifications in (10.1) all hold completely iso-
metrically when V is a homogeneous variety (with the provision that the appearance of
A(Bd)(cid:12)(cid:12)V in the identifications should be interpreted only as equality of sets).
In the rest of this section we shall consider the isomorphism problem for the algebras
of the kind A(V) for a homogeneous nc variety V ⊆ Bd, where d = ∞ is included. We
leave the investigation of the function theory, representation theory, and classification of
the algebras A(V) associated with general varieties for future work.
In Section 9.3 of [43] we observed that when V is homogeneous, every X ∈ V gives
w∗
(z1, . . . , zd),
rise to completely contractive unital representation
f (cid:55)→ f (X).
Further, for every Φ ∈ Repcc
k (A(V)) there exists X ∈ V such that Φ = ΦX. Using -- as
is done in the proof of Theorem 3.4 -- the fact that bounded representations into Mn are
completely bounded, together with the fact that completely bounded representations are
similar to completely contractive ones, we obtain the following description of the bounded
finite dimensional representations of A(V).
Proposition 10.1. Let V ⊆ Bd be a homogeneous nc variety. For very k ∈ N, the natural
projection πk of Repk(A(V)) into (cid:102)Bd given by
is a bijection onto the similarity invariant envelope (cid:101)V of V.
by evaluation at a point of (cid:101)V, we can use the methods of this paper (with significantly
A(W) be a unital bounded homomorphism. Then there exists an nc map G : (cid:102)W → (cid:101)V such
simpler proofs) to prove the following classification results.
Theorem 10.2. Let V ⊆ Bd and W ⊆ Be be homogeneous nc varieties. Let α : A(V) →
Once we know that every bounded finite dimensional representation of A(V) is given
πk(Φ) = (Φ(z1), . . . , Φ(zd))
42
that G((cid:102)W) = (cid:101)V, which implements α by the formula
α(f ) = (cid:101)f ◦ G.
In this case, (cid:107)α(cid:107) = supW∈W (cid:107)ΦG(W )(cid:107), and (cid:107)α(cid:107)cb = supW∈W (cid:107)ΦG(W )(cid:107)cb if α is completely
bounded. Moreover, if we write G = (G1, . . . , Gd), then the component Gi is in A(W) for all
i = 1, . . . , d.
Consequently, A(V) and A(W) are (completely) boundedly isomorphic, if and only if there
exist (δcb-bi-Lipschitz) δb-bi-Lipschitz nc maps G : (cid:102)W → (cid:101)V with components in A(W), and
H : (cid:101)V →(cid:102)W with components in A(V), which are mutual inverses of each other.
Theorem 10.3. Let V ⊆ Bd and W ⊆ Be be two homogeneous nc varieties. The following
are equivalent:
(1) A(V) and A(W) are boundedly isomorphic.
(2) A(V) and A(W) are completely boundedly isomorphic.
(3) There exists a δb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
(4) There exists a δcb-bi-Lipschitz linear map mapping(cid:102)W onto (cid:101)V.
Remark 10.4. When V and W are varieties determined by monomials in (finitely many)
noncommuting variables, the algebras A(V) and A(W) are precisely the tensor algebras
studied in [23] (the algebra A(V) appeared in [23] as AX). In [23, Theorem 9.2] it was
shown that in this case A(V) and A(W) are algebraically isomorphic if and only if they are
completely isometrically isomorphic, and this happens if and only if there is a permutation
of the variables, such that the defining monomials (and hence the varieties) are the same.
REFERENCES
[1] J. Agler and J. E. McCarthy. Pick interpolation for free holomorphic functions. Amer. J. Math.,
137(6):1685 -- 1701, 2015.
[2] A. S. Amitsur and J. Levitzki. Minimal identities for algebras. Proc. Amer. Math. Soc., 1:449 -- 463, 1950.
[3] A. Arias and G. Popescu. Noncommutative interpolation and Poisson transforms. Israel J. Math.,
115:205 -- 234, 2000.
[4] B. Aupetit. On log-subharmonicity of singular values of matrices. Studia Math., 122(2):195 -- 200, 1997.
[5] J. A. Ball, G. Marx, and V. Vinnikov. Interpolation and transfer-function realization for the noncommu-
tative schur-agler class. arXiv, 1602.00762, 2015.
[6] J. A. Ball, G. Marx, and V. Vinnikov. Noncommutative reproducing kernel Hilbert spaces. J. Funct. Anal.,
271(7):1844 -- 1920, 2016.
[7] S. Belinschi and V. Vinnikov. Noncommutative hyperbolic metrics. arXiv, 1707.09762, 2018.
[8] J. W. Bunce. Models for n-tuples of noncommuting operators. J. Funct. Anal., 57(1):21 -- 30, 1984.
[9] K. F. Clancey and I. Gohberg. Factorization of matrix functions and singular integral operators, volume 3
of Operator Theory: Advances and Applications. Birkhauser Verlag, Basel-Boston, Mass., 1981.
[10] K. R. Davidson, M. Hartz, and O. M. Shalit. Multipliers of embedded discs. Complex Anal. Oper. Theory,
9(2):287 -- 321, 2015.
[11] K. R. Davidson and D. R. Pitts. The algebraic structure of non-commutative analytic Toeplitz algebras.
Math. Ann., 311(2):275 -- 303, 1998.
[12] K. R. Davidson and D. R. Pitts. Invariant subspaces and hyper-reflexivity for free semigroup algebras.
Proc. London Math. Soc. (3), 78(2):401 -- 430, 1999.
[13] K. R. Davidson, C. Ramsey, and O. M. Shalit. The isomorphism problem for some universal operator
algebras. Adv. Math., 228(1):167 -- 218, 2011.
43
[14] K. R. Davidson, C. Ramsey, and O. M. Shalit. Operator algebras for analytic varieties. Trans. Amer. Math.
Soc., 367(2):1121 -- 1150, 2015.
[15] E. G. Effros and Z.-J. Ruan. Operator Spaces, volume 23 of London Mathematical Society Monographs.
New Series. The Clarendon Press, Oxford University Press, New York, 2000.
[16] D. Eisenbud and M. Hochster. A Nullstellensatz with nilpotents and Zariski's main lemma on holomor-
[17] D. E. Evans and R. Hø egh Krohn. Spectral properties of positive maps on C∗-algebras. J. London Math.
phic functions. J. Algebra, 58(1):157 -- 161, 1979.
Soc. (2), 17(2):345 -- 355, 1978.
[18] D. R. Farenick. Irreducible positive linear maps on operator algebras. Proc. Amer. Math. Soc.,
124(11):3381 -- 3390, 1996.
[19] A. E. Frazho. Models for noncommuting operators. J. Funct. Anal., 48(1):1 -- 11, 1982.
[20] M. Hartz. Topological
isomorphisms for some universal operator algebras. J. Funct. Anal.,
[21] M. Hartz. On the isomorphism problem for multiplier algebras of Nevanlinna-Pick spaces. Canad. J.
263(11):3564 -- 3587, 2012.
Math., to appear.
[22] M. Hazewinkel, N. Gubareni, and V. V. Kirichenko. Algebras, rings and modules. Vol. 1, volume 575 of
Mathematics and its Applications. Kluwer Academic Publishers, Dordrecht, 2004.
[23] E. T. A. Kakariadis and O. M. Shalit. Operator algebras associated with monomial ideals in noncom-
muting variables. J. Math. Anal. Appl., 472(1):738 -- 813, 2019.
[24] D. S. Kaliuzhnyi-Verbovetskyi and V. Vinnikov. Foundations of Free Noncommutative Function Theory,
volume 199 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI,
2014.
[25] I. Kaplansky. Rings with a polynomial identity. Bull. Amer. Math. Soc., 54:575 -- 580, 1948.
[26] J. L. Kazdan. Matrices A(t) depending on a parameter
t. Unpublished note, url
link
https://www.math.upenn.edu/ kazdan/504/eigenv.pdf.
[27] M. Kerr, J. E. McCarthy, and O. M. Shalit. On the isomorphism question for complete Pick multiplier
algebras. Integral Equations Operator Theory, 76(1):39 -- 53, 2013.
[28] J. E. McCarthy and O. M. Shalit. Spaces of Dirichlet series with the complete Pick property. Israel J.
Math., to appear.
[29] V. Paulsen. Completely Bounded Maps and Operator Algebras, volume 78 of Cambridge Studies in Ad-
vanced Mathematics. Cambridge University Press, Cambridge, 2002.
[30] G. Popescu. Isometric dilations for infinite sequences of noncommuting operators. Trans. Amer. Math.
Soc., 316(2):523 -- 536, 1989.
1 . Math. Scand., 68(2):292 -- 304, 1991.
[31] G. Popescu. von Neumann inequality for B(H)n
[32] G. Popescu. Multi-analytic operators on Fock spaces. Math. Ann., 303(1):31 -- 46, 1995.
[33] G. Popescu. Similarity and ergodic theory of positive linear maps. J. Reine Angew. Math., 561:87 -- 129,
2003.
[34] G. Popescu. Free holomorphic functions on the unit ball of B(H)n. J. Funct. Anal., 241(1):268 -- 333,
2006.
[35] G. Popescu. Operator theory on noncommutative varieties. Indiana Univ. Math. J., 55(2):389 -- 442,
[36] G. Popescu. Noncommutative hyperbolic geometry on the unit ball of B(H)n. J. Funct. Anal.,
2006.
256(12):4030 -- 4070, 2009.
[37] G. Popescu. Unitary invariants
in multivariable operator
theory. Mem. Amer. Math. Soc.,
200(941):vi+91, 2009.
[38] G. Popescu. Free holomorphic automorphisms of the unit ball of B(H)n. J. Reine Angew. Math.,
638:119 -- 168, 2010.
[39] G. Popescu. Similarity problems in noncommutative polydomains. J. Funct. Anal., 267(11):4446 -- 4498,
2014.
[40] C. Ramsey. Maximal ideal space techniques in non-selfadjoint operator algebras. PhD thesis, University of
[41] T. J. Ransford and M. C. White. Holomorphic self-maps of the spectral unit ball. Bull. London Math.
Waterloo, 2013.
Soc., 23(3):256 -- 262, 1991.
44
[42] G. Salomon and O. M. Shalit. The isomorphism problem for complete Pick algebras: a survey. In Opera-
tor Theory: Advances and Application, Proceedings of the International Workshop on Operator Theory and
Applications (IWOTA 2014), pages 168 -- 198. Operator Theory: Advances and Applications, Vol. 255.
Birkhauser, Basel, 2016.
[43] G. Salomon, O. M. Shalit, and E. Shamovich. Algebras of bounded noncommutative analytic functions
on subvarieties of the noncommutative unit ball. Trans. Amer. Math. Soc., 370(12):8639 -- 8690, 2018.
[44] O. M. Shalit and B. Solel. Subproduct systems. Doc. Math., 14:801 -- 868, 2009.
[45] E. Shamovich. On fixed points of self maps of the free ball. J. Funct. Anal., 275(2):422 -- 441, 2018.
[46] I. Suciu. Analytic formulas for the hyperbolic distance between two contractions. Ann. Polon. Math.,
66:239 -- 252, 1997. Volume dedicated to the memory of Włodzimierz Mlak.
[47] B. Sz.-Nagy, C. Foias, H. Bercovici, and L. K´erchy. Harmonic Analysis of Operators on Hilbert Space.
Universitext. Springer, New York, enlarged edition, 2010.
[48] J. L. Taylor. The analytic-functional calculus for several commuting operators. Acta Math., 125:1 -- 38,
1970.
[49] J. L. Taylor. A general framework for a multi-operator functional calculus. Advances in Math., 9:183 --
252, 1972.
[50] J. L. Taylor. Functions of several noncommuting variables. Bull. Amer. Math. Soc., 79:1 -- 34, 1973.
[51] E. Vesentini. Maximum theorems for spectra. In Essays on Topology and Related Topics (M´emoires d´edi´es
`a Georges de Rham), pages 111 -- 117. Springer, New York, 1970.
68, 1979.
[52] E. Vesentini. Variations on a theme of Carath´eodory. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4), 6(1):39 --
[53] D.-V. Voiculescu. Symmetries of some reduced free product C∗-algebras. In Operator algebras and their
connections with topology and ergodic theory (Bus¸teni, 1983), volume 1132 of Lecture Notes in Math.,
pages 556 -- 588. Springer, Berlin, 1985.
[54] D.-V. Voiculescu. Addition of certain noncommuting random variables. J. Funct. Anal., 66(3):323 -- 346,
[55] D.-V. Voiculescu. Free analysis questions. I. Duality transform for the coalgebra of ∂X : B. Int. Math. Res.
1986.
Not., (16):793 -- 822, 2004.
[56] D.-V. Voiculescu. Free analysis questions II: the Grassmannian completion and the series expansions at
the origin. J. Reine Angew. Math., 645:155 -- 236, 2010.
DEPARTMENT OF PURE MATHEMATICS, UNIVERSITY OF WATERLOO, WATERLOO, ON, CANADA
E-mail address: [email protected]
DEPARTMENT OF MATHEMATICS, TECHNION -- ISRAEL INSTITUTE OF TECHNOLOGY, HAIFA, ISRAEL
E-mail address: [email protected]
DEPARTMENT OF PURE MATHEMATICS, UNIVERSITY OF WATERLOO, WATERLOO, ON, CANADA
E-mail address: [email protected]
45
|
1101.0790 | 2 | 1101 | 2011-07-22T00:57:10 | Operator system quotients of matrix algebras and their tensor products | [
"math.OA"
] | An operator system modulo the kernel of a completely positive linear map of the operator system gives rise to an operator system quotient. In this paper, operator system quotients and quotient maps of certain matrix algebras are considered. Some applications to operator algebra theory are given, including a new proof of Kirchberg's theorem on the tensor product of B(H) with the group C*-algebra of a countable free group. We also show that an affirmative solution to the Connes Embedding Problem is implied by various matrix-theoretic problems, and we give a new characterisation of unital C*-algebras that have the weak expectation property. | math.OA | math |
OPERATOR SYSTEM QUOTIENTS OF MATRIX ALGEBRAS
AND THEIR TENSOR PRODUCTS
DOUGLAS FARENICK AND VERN I. PAULSEN
Abstract. If φ : S → T is a completely positive (cp) linear map of operator
systems and if J = ker φ, then the quotient vector space S/J may be endowed
with a matricial ordering through which S/J has the structure of an operator
system. Furthermore, there is a uniquely determined cp map φ : S/J → T
such that φ = φ ◦ q, where q is the canonical linear map of S onto S/J . The cp
map φ is called a complete quotient map if φ is a complete order isomorphism
between the operator systems S/J and T . Herein we study certain quotient
maps in the cases where S is a full matrix algebra or a full subsystem of
tridiagonal matrices.
Our study of operator system quotients of matrix algebras and tensor prod-
ucts has applications to operator algebra theory. In particular, we give a new,
simple proof of Kirchberg's Theorem C∗(F∞)⊗minB(H) = C∗(F∞)⊗max B(H),
show that an affirmative solution to the Connes Embedding Problem is im-
plied by various matrix-theoretic problems, and give a new characterisation of
unital C∗-algebras that have the weak expectation property.
Introduction
The C∗-algebra B(H) of bounded linear operators acting on a Hilbert space H
and the group C∗-algebra C∗(F∞) of the free group F∞ with countably infinitely
many generators are both universal objects in operator algebra theory. Therefore, it
is a remarkable fact that C∗(F∞) ⊗min B(H) = C∗(F∞) ⊗max B(H), which is a well
known and important theorem of Kirchberg [12, Corollary 1.2]. Kirchberg's proof
was achieved by first showing that the C∗-algebra C∗(F∞) has the lifting property
[12, Lemma 3.3] and by then invoking his theorem [12, Theorem 1.1] that states
A ⊗min B(H) = A ⊗max B(H) for every separable unital C∗-algebra A that has
the lifting property. A more direct proof of Kirchberg's theorem on the uniqueness
of the C∗-norm for C∗(F∞) ⊗ B(H) was later found by Pisier [17],[18, Chapter 13]
using a mix of operator algebra and operator space theory. Boca [3] made a further
extension, replacing C∗(F∞) with a free product of C∗-algebras Ai in which each
of the maps idAi is locally liftable. An exposition of the results of Kirchberg and
Pisier can be found in [13].
One of our main results in this paper is a new proof of Kirchberg's theorem (see
Corollary 3.2), obtained herein by reducing Kirchberg's theorem to the verification
of a certain property (Theorem 3.1) of a finite-dimensional quotient operator system
Wn whose C∗-envelope is C∗(Fn−1).
Our study of operator system quotients of the matrix algebra Mn and of the
full operator subsystem Tn ⊆ Mn of tridiagonal matrices (full in the sense that
C∗(Tn) = Mn) allows us to formulate matrix-theoretic questions in Section §5
1
2
DOUGLAS FARENICK AND VERN I. PAULSEN
whose resolution in the affirmative would result in a solution to the Connes Em-
bedding Problem. Our approach to this is by way of another celebrated theorem of
Kirchberg [11, Proposition 8]: C∗(F∞) ⊗min C∗(F∞) = C∗(F∞) ⊗max C∗(F∞) if and
only if every separable II1-factor can be embedded as a subfactor of the ultrapower
of the hyperfinite II1-factor. It is of course an open problem whether these equiva-
lent statements are true. The problem of whether or not C∗(F∞) ⊗min C∗(F∞) =
C∗(F∞) ⊗max C∗(F∞) is known as the Kirchberg Problem, while the latter problem
involving separable II1-factors is the Connes Embedding Problem.
The methods we use in this paper draw upon recent work of the second au-
thor and others [10, 9, 16] on matricially ordered vector spaces, tensor products of
operator systems, and quotients of operator systems.
An operator system is a triple consisting of: (i) all ∗-vector spaces Mn(S) of n×n
matrices over a fixed ∗-vector space S; (ii) distinguished cones Mn(S)+ in Mn(S)
that give rise to a matricial ordering of S; and (iii) a distinguished element e ∈ S
which is an Archimedean order unit. The axioms for operator systems are given in
[5, 14], and so we will not repeat them here. However, recall a very important and
fundamental fact [5]: every operator system S arises as an operator subsystem of
B(H), for some Hilbert space H. This result is known as the Choi -- Effros Theorem.
Two important C∗-algebras that arise in connection with a given operator system
S are the C∗-envelope C∗
e (S) and the injective envelope I(S) of S. The latter
algebra I(S) is very large (usually nonseparable) and there exist embeddings so
that S ⊆ C∗
e (S) is the C∗-subalgebra of I(S) generated
by S. These are "enveloping" algebras in the sense that C∗
e (S) is a quotient of every
C∗-algebra generated by (a copy of) S and I(S) rigidly contains S, which is to say
that if a linear unital completely positive (ucp) map ω : I(S) → I(S) satisfies
ωS = idS, then necessarily ω = idI(S). See [14, Chapter15] for further details on
C∗- and injective envelopes.
e(S) ⊆ I(S) and such that C∗
In contrast to progress on tensor products and quotients in operator space theory
(see [18], for example), an analogous theory in the category of operator systems is
only now emerging. In §1 we review those aspects of the theory that are required
for our work herein.
In Sections §2 and §4 we study certain quotient operator
systems that arise from quotients of operator systems of finite matrices, and in
Sections §3 and §5 we consider tensor products of these quotients and obtain the
earlier mentioned applications to Kirchberg's Theorem and the Connes Embedding
Problem. In Section §6 we use complete quotient maps to characterise unital C∗-
algebras that possess the weak expectation property (WEP), and in §7 the injective
envelope of C∗(Fn) is determined.
1. Tensor Products, Quotients, and Duals of Operator Systems: A
Review
In this section we review some of the fundamental facts, established in [10, 9],
concerning tensor products, quotients, and duals of operator systems, and introduce
the notion of a complete quotient map.
Some basic notation: (i) the Archimedean order unit e of an operator system
S is generally denoted by 1, but we will sometimes revert to the use of e in cases
where the order unit is not canonically given (for example, when considering duals
of operator systems); (ii) for a linear map φ : S → T , the map φ(n) : Mn(S) →
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
3
Mn(T ) is defined by φ(n) ([xij ]i,j) = [φ(xij )]i,j; (iii) for any operator systems S
and T , S ⊗ T shall denote their algebraic tensor product.
1.1. Quotients of operator systems. Assume that S is an operator system and
that J ⊂ S is norm-closed ∗-subspace that does not contain 1S. Let q : S → S/J
denote the canonical linear map of S onto the vector space S/J . The vector space
S/J has an induced involution defined by q(X)∗ = q(X ∗).
We will denote elements of S by uppercase letters, such as X, and elements of
S/J by lowercase letters, as in x = q(X), or by "dots" as in X = q(X).
For each n ∈ N, there is a linear isomorphism
(S/J ) ⊗ Mn ∼= (S ⊗ Mn)/(J ⊗ Mn) .
The canonical linear surjection S ⊗ Mn → (S ⊗ Mn)/(J ⊗ Mn) is denoted by q(n)
and we shall view q(n)(X) as an n×n matrix x = X of
Xij ∈ S/J , thereby identify-
ing (S/J )⊗Mn with Mn(S/J ). Further, each matrix space Mn(S/J ) is an invo-
lutive vector space under the involution h = [hij ]1≤i,j≤n 7→ h∗ = [h∗
ji]1≤i,j≤n. The
real vector space of hermitian (or selfadjoint elements) is denoted by Mn(S/J )sa.
Consider the subset Dn(S/J ) of Mn(S/J )sa defined by
Dn(S/J ) = { H : ∃ K ∈ Mn(J )sa such that H + K ∈ Mn(S)+} .
The collection {Dn(S/J )}n∈N is a family of cones that endow S/J with the struc-
ture of a matricially ordered space. However, this ordering will not be an operator
system on S/J in general.
Definition 1.1. A subspace J ⊂ S is a kernel if there are an operator system T
and a completely positive linear map φ : S → T such that J = ker φ.
If J ⊂ S is a kernel, then define a subset Cn(S/J ) ⊂ Mn(S/J )sa by
Cn(S/J ) = { H : ∀ ε > 0 ∃ Kε ∈ Mn(J )sa such that ε1 + H + Kε ∈ Mn(S)+} .
That is,
Cn(S/J ) = { H : ε 1 + H ∈ Dn(S/J ), ∀ ε > 0} .
The collection {Cn(S/J )}n∈N is a family of cones that endow S/J with the struc-
ture of an operator system with (Archimedean) order unit 1 = q(1).
Definition 1.2. The operator system (S/J , {Cn(S/J )}n∈N, q(1)) is called a quo-
tient operator system.
The first basic result concerning quotient operator systems is the First Isomor-
phism Theorem [9, Proposition 3.6].
Theorem 1.1. (First Isomorphism Theorem) If φ : S → T is a nonzero com-
pletely positive linear map, then there is a unique completely positive linear map
φ : S/ ker φ → T such that φ = φ ◦ q.
Associated with Dn(S/J ) and Cn(S/J ) are the subsets Dn(S/J ) and Cn(S/J )
of (S ⊗ Mn)sa defined by
Dn(S/J ) = q−1
n (Dn(S/J ))
and Cn(S/J ) = q−1
n (Cn(S/J )) .
Clearly Dn(S/J ) ⊆ Cn(S/J ). Examples in which the inclusion Dn(S/J ) ⊆
Cn(S/J ) is proper are given in [9].
Definition 1.3. A kernel J ⊂ S is completely order proximinal if Dn(S/J ) =
Cn(S/J ), for every n ∈ N.
4
DOUGLAS FARENICK AND VERN I. PAULSEN
This notion of a completely order proximinal kernel will be of importance for the
quotient operator systems we study in later sections.
Example 1.2. If A is a unital C∗-algebra and if J ⊂ A is the kernel of a ∗-
homomorphism, then J is completely order proximinal and
((A/J ) ⊗ Mn)+ = Dn(A/J ) = Cn(A/J ) , for all n ∈ N .
Proof. Let J = ker π, where π is a ∗-homomorphism which, without loss of gen-
erality, we assume to be unital. Note that Dn(A/J ) is precisely the positive cone
(A ⊗ Mn)+ of the quotient C∗-algebra A/J . Because any positive element h = H
of (A/J ) ⊗ Mn ∼= (A ⊗ Mn)/(J ⊗ Mn) lifts to an element of the form H + K,
where K ∈ J ⊗ Mn and H ∈ (A ⊗ Mn)+,
Dn(A/J ) = {H + K : H ∈ (A ⊗ Mn)+, K ∈ (J ⊗ Mn)sa} .
Suppose now that H ∈ Cn(A/J ). Then, for every ε > 0, ε1 + H ∈ Dn(A/J ) and
so every εqn(1) + qn(H) = ε 1 + h is positive in (A/J ) ⊗ Mn. But as 1 ∈ A/J is
an Archimedean order unit for the quotient C∗-algebra A/J , ε 1 + h is positive for
all ε > 0 only if h is positive, which is to say that H ∈ Dn(A/J ).
(cid:3)
The proof above shows that, in the case of C∗-agebras A and ideals J ⊂ A,
Dn(A/J ) = Cn(A/J ) for every n ∈ N. This in fact leads to the following useful
criterion for completely order proximinal kernels in operator systems.
Proposition 1.3. Let J be a kernel in an operator system S. Then
(1) Dn(S/J ) = Mn(J ) + Mn(S)+,
(2) Cn(S/J ) is the norm closure of Mn(J ) + Mn(S)+, and
(3) J is completely order proximinal if and only if Mn(J ) + Mn(S)+ is closed
for every n ∈ N.
Proof. The first statement is obvious. To see the second statement, first note
that Cn(S/J ) = {H : ∀ǫ > 0, ǫ1 + H ∈ Dn(S/J )} = {H : ∀ǫ > 0, ǫ1 + H ∈
Mn(S)+ + Mn(J )}. Thus, if H ∈ Cn(S/J ) then for every ǫ > 0 there exist
Pǫ ∈ Mn(S)+ and Kǫ ∈ Mn(J ), such that ǫ1 + H = Pǫ + Kǫ and it follows that
H is in the norm closure of Mn(S)+ + Mn(J ).
For the converse it is sufficient to assume that H = H ∗ is in the closure of
Mn(S)+ + Mn(J )sa and prove that ǫ1 + H ∈ Mn(S)+ + Mn(J ) for every ǫ > 0.
Since H is in the closure and Mn(J )∗ = Mn(J ), for every ǫ > 0 we may choose
Pǫ ∈ Mn(S)+ and Kǫ ∈ Mn(J )sa such that kH − (Pǫ + Kǫ)k < ǫ. This implies
that Qǫ = ǫ1 + H − Pǫ − Kǫ ∈ Mn(S)+, and hence ǫ1 + H = (Qǫ + Pǫ) + Kǫ ∈
Mn(S)+ + Mn(J ). This proves the second statement.
The third statement follows by combining the first and second statements. (cid:3)
1.2. Definitions: complete quotient maps and complete order injections.
Definition 1.4. A linear map φ : S → T of operator systems is a complete order
isomorphism if
(1) φ is a linear isomorphism and
(2) φ and φ−1 are completely positive.
Definition 1.5. A linear map φ : S → T of operator systems is a complete order
injection, or a coi map, if φ is a complete order isomorphism between S and the
operator system φ(S) with order unit φ(1S).
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
5
Definition 1.6. A completely positive linear map φ : S → T of operator systems
is a complete quotient map if the uniquely determined completely positive linear
map φ : S/ ker φ → T is a complete order isomorphism.
A basic fact about complete quotient maps that we shall use repeatedly is:
Proposition 1.4. If φ : S → T is a complete quotient map and if ker φ is com-
pletely order proximinal, then φ(n) maps Mn(S)+ onto Mn(T )+, for every n ∈ N.
The next fact is also basic.
Proposition 1.5. If T is an operator subsystem of S, and if J is a kernel such
that J ⊂ T ⊆ S, then there is a complete order embedding T /J → S/J .
1.3. Tensor products of operator systems.
1.3.1. The minimal tensor product. The matricial state space of an operator system
S is the set
S∞(S) = [n∈N
For every p ∈ N, let
Sn(S) , where Sn(S) = {all ucp maps S → Mn} .
Cmin
p
(S, T ) = {x ∈ Mp(S ⊗ T ) : (φ ⊗ ψ)(p)(x) ∈ Mp(Mk ⊗ Mm)+,
∀(φ, ψ) ∈ Sk(S) × Sm(T ), ∀ (k, m) ∈ N × N} .
p
The collection {Cmin
(S, T )}p∈N induces an operator system structure on S ⊗ T
with order unit 1S ⊗ 1T . The operator system that arises from this matricial order
structure is denoted by S ⊗min T and is called the minimal tensor product of S and
T .
1.3.2. The maximal tensor product. Let Dmax
subset of all elements of the form
p
(S, T ) ⊂ Mp(S ⊗ T ) denote the
α(s ⊗ t)α∗ , where s ∈ Mk(S)+ , t ∈ Mm(T )+ , α : Ck ⊗ Cm → Cp is linear .
For every p ∈ N, let
Cmax
p
(S, T ) = {h ∈ Mp(S ⊗ T ) : ε1Mp + h ∈ Dmax
p
(S, T ) for all ε > 0} .
p
The collection {Cmax
(S, T )}p∈N induces an operator system structure on S ⊗ T
with order unit 1S ⊗ 1T . The operator system that arises from this matricial order
structure is denoted by S ⊗max T and is called the maximal tensor product of S
and T .
Proposition 1.6. If φ : S → T is a complete quotient map, then φ ⊗ idR :
S ⊗max R → T ⊗max R is a complete quotient map for every operator system R.
Proof. Let J = ker φ. Because φ : S/J → T is a complete order isomorphism,
so is φ ⊗ idR : (S/J ) ⊗max R → T ⊗max R. Therefore, φ ⊗ idR is a complete
quotient map if (S ⊗max R)/ ker(φ ⊗ idR) and (S/J ) ⊗max R are completely order
isomorphic.
Let q : S ⊗max R → (S ⊗max R)/ ker(φ ⊗ idR) be the canonical surjective cp map
and define a bilinear map
σ : (S/J ) × R → (S ⊗max R)/ ker(φ ⊗ idR)
by σ(x, y) = q(X ⊗ y), where X ∈ S is any element for which q(X) = x. Note that
σ is well defined. Furthermore, σ is jointly completely positive and "unital," and
6
DOUGLAS FARENICK AND VERN I. PAULSEN
so by the universal property of the max tensor product [10, Theorem 5.8], there is
a ucp extension σ : (S/J ) ⊗max R → (S ⊗max R)/ ker(φ ⊗ idR). We claim that σ
is a complete order isomorphism.
By linear algebra, the restriction of σ to the algebraic tensor product is a linear
isomorphism between (S/J ) ⊗ R and (S ⊗ R)/ ker(φ ⊗ idR), and therefore σ is a
linear isomorphism of (S/J ) ⊗max R and (S ⊗max R)/ ker(φ ⊗ idR).
To show that σ−1 is completely positive, select a positive element b in the p × p
matrices over (S ⊗max R)/ ker(φ ⊗ idR). Hence, for every ε > 0 there is are P ∈
Mk(S)+, Q ∈ Mm(R)+ and α : Ck ⊗ Cm → Cp linear such that ε 1 + b = α(q(P ) ⊗
Q)α∗. Hence, εσ−1(p)( 1) + σ−1(p)(b) ∈ Dmax
(S/J , R) for all ε > 0. Because σ−1
is unital, this implies that σ−1(p)(b) ∈ Mp((S/J ) ⊗max R)+.
(cid:3)
p
Corollary 1.7. If φ : S → T is a complete quotient map, then φ ⊗ φ : S ⊗max S →
T ⊗max T is a complete quotient map.
Proof. Because φ⊗φ = (φ⊗idT )◦(idT ⊗φ) and a composition of complete quotient
maps is a complete quotient map, the result follows.
(cid:3)
1.3.3. The commuting tensor product. Let cp(S, T ) denote the set of all pairs (φ, ψ)
of completely positive linear maps that map S and T , respectively, into a common
B(H) and such that φ(x)ψ(y) = ψ(y)φ(x), for all (x, y) ∈ S × T . Such a pair
determines a bilinear map φ · ψ : S × T → B(H) via φ · ψ(x, y) = φ(x)ψ(y), and so
there exists a unique linear map, denoted again by φ · ψ, of S ⊗ T into B(H) and
for which φ · ψ(x ⊗ y) = φ(x)ψ(y), for all elementary tensors x ⊗ y ∈ S ⊗ T . Let
Ccomm
p
(S, T ) = {x ∈ Mp(S ⊗ T ) : (φ · ψ)(p)(x) ∈ Mp(B(H))+,
∀ (φ, ψ) ∈ cp(S, T )} .
p
The collection {Ccomm
(S, T )}p∈N induces an operator system structure on S ⊗ T
with order unit 1S ⊗ 1T . The operator system that arises from this matricial order
structure is denoted by S ⊗c T and is called the commuting tensor product of S
and T .
1.3.4. Properties of tensor products. Evidently,
Cmax
p
(S, T ) ⊆ Ccomm
p
(S, T ) ⊆ Cmin
p
(S, T ) ,
and so the maps S ⊗max T → S ⊗c T and S ⊗c T → S ⊗min T arising from the
identity map of S ⊗ T are ucp; we denote this by
S ⊗max T ⊆ S ⊗c T ⊆ S ⊗min T .
Definition 1.7. Assume that S is an operator system.
(1) S is said to be nuclear or (min,max)-nuclear if S ⊗min T = S ⊗max T for
every operator system T .
(2) S is (min,c)-nuclear if S ⊗min T = S ⊗c T for every operator system T .
All unital C∗-algebras that are nuclear in the sense of C∗-algebraic tensor prod-
ucts are also (min,max)-nuclear [10, Proposition 5.15]. Although finite-dimensional
C∗-algebras are, therefore, nuclear, it is not the case that every finite-dimesional
operator system is nuclear [10, Theorem 5.18].
By definition, an operator system S is endowed with a specific matricial ordering
in which the positive cone in the vector space Mp(S) of p × p matrices is denoted
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
7
by Mp(S)+. Expressed as tensor products, we have
Mn(S)+ = (S ⊗max Mn)+ = (S ⊗c Mn)+ = (S ⊗min Mn)+ .
1.4. Duals of finite-dimensional operator systems. For any finite-dimensional
operator system S, let Sd denote the dual of S. For every n ∈ N, there are the
following natural identifications:
Mn(Sd) ∼= L(Sdd, Mn) = L(S, Mn) ∼= Mn(S)d .
In the first of these identifications, and using ↔ to denote "is identified with,"
G = [gij]i,j ∈ Mn(Sd) ←→ G ∈ L(S, Mn), G(s) = [gij(s)]i,j .
In the second of the identifications,
G ∈ L(S, Mn) ←→ ϕG ∈ Mn(S)d, ϕG ([sij ]i,j) = Xi,j
gij(sij) .
A linear functional on Mn(S) of the form ϕG satisfies ϕG(x) ≥ 0, for all x ∈
Mn(S)+, if and only if G : S → Mn is completely positive. Hence, Mn(Sd)+ is
defined to be:
Mn(Sd)+ = {G ∈ Mn(Sd) : G : S → Mn is completely positive}
= {G ∈ Mn(Sd) : ϕG(x) ≥ 0, ∀ x ∈ Mn(S)+} .
One can define these cones for any operator system S and they endow Sd with a
matricial ordering. However, because S has finite dimension, this matricial ordering
on Sd gives rise to an operator system structure on Sd in which an Archimedean
order unit is given by e = ϕ, for some faithful state ϕ of S [5, §4].
For every linear map φ : S → T , let φd denote the adjoint of φ as a linear map
T d → Sd. That is, φd(ψ)[s] = ψ(φ(s)), for all ψ ∈ T d, s ∈ S.
Proposition 1.8. The following statements are equivalent for a linear map φ :
S → T of finite-dimensional operator systems:
(1) φ is a complete quotient map;
(2) φd is a complete order injection.
Proof. Assume that φ is a complete quotient map and fix n ∈ N. Recall that φd is
completely positive since, for every G ∈ Mn(T d), φd(n)(G) is identified with G ◦ φ,
a composition of completely positive linear maps; thus, φd(n)(G) ∈ Mn(Sd)+.
To show that φd is a complete order injection, it is sufficient to show that if
φd(n)(G) ∈ Mn(Sd)+, for some G ∈ Mn(T d), then necessarily G ∈ Mn(T d)+.
Therefore, it is necessary to show that G is completely positive. Let p ∈ N and
y ∈ Mp(T )+. By hypothesis, φ is a complete quotient map; thus, there is a x ∈
Mp(S)+ such that φ(p)(x) = y (Proposition 1.4). Thus, G(p)(y) = G(p) ◦ φ(p)(x) ∈
Mp(Mn)+, which shows that G ∈ Mn(T d)+. This proves that φd is a complete
order injection.
Conversely, assume that φd is a complete order injection. Let E denote an
arbitrary operator system. By way of the identification of Mn(E d)+ with the
linear functionals ϕ on Mn(E) for which ϕ(x) ≥ 0 for all x ∈ Mn(E)+, we identify
Mn(E d)+ with the dual cone [Mn(E)+]# of the cone Mn(E)+:
Mn(E d)+ = [Mn(E)+]#
= {ψ ∈ Mn(E)d : ψ(x) ≥ 0, ∀ x ∈ Mn(E)+} .
8
DOUGLAS FARENICK AND VERN I. PAULSEN
By hypothesis, ψ ∈ Mn(T d)+ if and only if φd(n)(ψ) ∈ Mn(Sd)+. The dual cone
of the cone φd(n) (Mn(S)+) in Mn(T ) is given by
(cid:2)φd(n) (Mn(S)+)(cid:3)#
= {ψ ∈ Mn(T )d : ψ(y) ≥ 0, ; ∀ y ∈ φd(n) (Mn(S)+)}
= {ψ ∈ Mn(T )d : ψ ◦ φ(n)(x) ≥ 0, ∀ x ∈ Mn(S)+}
= {ψ ∈ Mn(T )d : φd(n)(ψ) ∈ Mn(Sd)+}
= [Mn(T )+]# (as φ is coi) .
Hence, by the Bidual Theorem,
φd(n) (Mn(S)+) = hφd(n) (Mn(S)+)i##
which implies that φ is a complete order isomorphism.
= [Mn(T )+]## = Mn(T )+ ,
(cid:3)
Remark 1. The use of finite-dimensional operator systems in Proposition 1.8 is not
essential, as the same arguments can be applied to the case of arbitrary S and T .
What is lost is the fact that the duals Sd and T d are operator systems. However,
duals of operator systems are matricially ordered spaces and one can speak of
completely positive maps between such spaces. Thus, Proposition 1.8 holds in the
category whose objects are matricially normed spaces and whose morphisms are
completely positive linear maps.
Proposition 1.9. If S and T are finite-dimensional operator systems, then the
operator systems Sd ⊗max T d and (S ⊗min T )d are completely order isomorphic.
Proof. Let δ and ζ be faithful states on S and T , respectively, so that eS d = δ and
eT d = ζ are the Archimedean order units for the operator systems Sd and T d.
We first show that there is a completely positive embedding of ι1 : Sd ⊗max T d →
(S ⊗min T )d. Let P ∈ Mk(Sd)+ and Q ∈ Mm(T d)+; we aim to show that P ⊗ Q
is a completely positive map S ⊗min T → Mk ⊗ Mm. Any completely positive map
of an operator system into a matrix algebra is a matricial convex combination of
matricial states; thus,
α∗
µφµαµ ,
P = Xµ
β∗
ν ψν βν ,
Q = Xν
µαµ = 1k, Pν β∗
(αµ ⊗ βµ)∗(φµ ⊗ φν )(αµ ⊗ βν)
where φµ and ψν are matricial states on S and T , respectively, and αµ and βν are
ν βν = 1m. Hence,
rectangular matrices for which Pν α∗
P ⊗ Q = Xµ,ν
is completely positive. Hence, every α( P ⊗ Q)α∗ is completely positive, for every
linear map α : Ck ⊗ Cm → Cn. Identifying P ⊗ Q and P ⊗ Q, this proves that
Therefore, if G ∈ Mn(Sd ⊗max T d)+, then for every ε > 0 there are P ∈ Mk(Sd)+,
Q ∈ Mm(T d)+, and α : Ck ⊗ Cm → Cn linear such that
α(P ⊗ Q)α∗ ∈ Mn(cid:0)(S ⊗min T )d(cid:1)+ .
G + ε(eS d ⊗ eT d ) = α(P ⊗ Q)α∗ ∈ Mn(cid:0)(S ⊗min T )d(cid:1)+ .
Because G + ε(eS d ⊗ eT d ) ∈ Mn(cid:0)(S ⊗min T )d(cid:1)+ for every ε > 0, the Archimedean
property implies that G ∈ Mn(cid:0)(S ⊗min T )d(cid:1)+. Hence, the canonical linear em-
bedding ι1 : (Sd ⊗ T d) → (S ⊗min T )d is completely positive.
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
9
We now show that there is a completely positive embedding of (Sd ⊗max T d)d
into S ⊗min T . To this end, fix n and let G ∈ Mn(cid:0)(Sd ⊗max T d)d(cid:1)+. Because
Sdd = S and T dd = T , G = [gij]i,j for some gij ∈ S ⊗ T . Having identified G
as an element of Mn(S ⊗ T ), we need only show that G ∈ Mn(S ⊗min T )+. To
do so, let ψ : S → Mk and ψ : T → Mm be any pair of matricial states. Thus,
φ ∈ Mk(Sd)+ and ψ ∈ Mm(T d)+. By hypothesis, G : Sd ⊗max T d → Mn is
completely positive, and so G(km)(φ ⊗ ψ) is positive in Mn(Mk ⊗ Mm). In writing
φ and ψ as matrices of linear functionals φkℓ, ψµν , we note that
G(km)(φ ⊗ ψ) = [G(φkℓ ⊗ ψµν )] = [φkℓ ⊗ ψµν )(gij )] = φ ⊗ ψ (n)(G) ,
which implies that φ ⊗ ψ (n)(G) ∈ Mn(Mk ⊗ Mm)+. Therefore, the canonical
linear embedding ι : (Sd ⊗max T d)d → S ⊗min T is completely positive.
To complete the proof, the completely positive embedding ι : (Sd ⊗max T d)d →
S ⊗min T leads to a completely positive embedding
ι2 = ιd : (S ⊗min T )d → (Sd ⊗max T d)dd = Sd ⊗max T d .
In combination with the embedding ι1 and because ι1 = ι−1
order isomorphism of Sd ⊗max T d and (S ⊗min T )d.
Remark 2. Since Sd is an operator system with Archimedian order unit δ and T d
is an operator system with Archimedian order unit ζ, the image of δ ⊗ ζ under the
above complete order isomorphism is an Archimedean order unit for (S ⊗min T )d.
Thus, in particular, δ ⊗ ζ is a faithful state on S ⊗min T .
2 , we obtain a complete
(cid:3)
2. The Operator System Mn/Jn
Assume that n ≥ 2 and that w2, . . . , wn are n−1 universal unitaries that generate
the full group C∗-algebra C∗(Fn−1), where Fn−1 is the free group on n−1 generators.
Let w1 = 1 ∈ C∗(Fn−1). Throughout, tr denotes the standard (non-normalised)
trace functional on a full matrix algebra.
Definition 2.1. For each n ≥ 2, let
(1) Jn ⊂ Mn be the vector subspace of all diagonal matrices D ∈ Mn with
tr(D) = 0, and
(2) Wn be the operator system in C∗(Fn−1) spanned by {wiw∗
j : 1 ≤ i, j ≤ n}.
We first show that Jn is a kernel. To do so, consider the unital linear map
φ : Mn → C∗(Fn−1) defined on the matrix units of Mn by
(1)
φ(Eij ) =
1
n
wiw∗
j , 1 ≤ i, j ≤ n .
The Choi matrix corresponding to φ is
[φ(Eij )]1≤i,j≤n =
1
n
[wiw∗
j ]1≤i,j≤n =
1
n
W ∗Y W ∈ (C∗(Fn−1) ⊗ Mn)+ ,
where W = Pi w∗
is a completely positive linear map of Mn onto Wn. It is clear that Jn ⊆ ker φ.
Conversely, suppose that A ∈ ker φ. Then,
i ⊗ Eii and Y = Pi,j 1 ⊗ Eij ∈ (C∗(Fn−1) ⊗ Mn)+. Hence, φ
and so aij = 0 for all j 6= i and tr(A) = 0. That is, ker φ ⊆ Jn.
aij(wiw∗
j ) ,
1
nXj6=i
0 = φ(A) = (
aii)1 +
n
Xi=1
10
DOUGLAS FARENICK AND VERN I. PAULSEN
Because Jn is a kernel, we form and study the quotient operator system Mn/Jn.
In this section we will show that the ucp map φ : Mn → Wn is a complete quotient
map and that the C∗-envelope of the operator system Mn/Jn is C∗(Fn−1).
Notation 2.1. The order unit q(1) of Mn/Jn is denoted by 1 and eij denotes
q(Eij ) for every matrix unit Eij of Mn.
Lemma 2.1. For every i and j, eii = 1
n
1 and keijk = 1
n .
and therefore q(2)(Q) is positive in M2(Mn/Jn). However, the matrix
Eji Ejj (cid:21) is positive in M2(Mn),
Proof. Suppose that j 6= i. The matrix Q =(cid:20) Eii Eij
1 (cid:21)
q(2)(Q) = (cid:20) 1
n . Consider now the density matrix ρ ∈ Mn with 1
is positive only if keijk ≤ 1
n
in every entry and let sρ be the state on Mn defined by sρ(X) = tr(ρX). Then
sρ(Jn) = {0} and so we obtain a well-defined state s on Mn/Jn via s( X) = sρ(X).
With this state s, we have s(eij ) = tr(ρEij ) = 1
n . (cid:3)
n , which implies that keijk ≥ 1
1
n
eji
eij
1
n
The result above shows that the norm on the operator system quotient Mn/Jn
is quite different from the quotient norm of Mn by the subspace Jn. Indeed, in the
usual quotient norm, one has keijkq ≡ inf{kEij + Kk : K ∈ Jn} = 1, when i 6= j.
Lemma 2.2. Jn ⊆ Mn is completely order proximinal.
Proof. By Proposition 1.3, we need to prove that for every p ∈ N, Mp(Mn)+ +
Mp(Jn) is closed. To this end assume that Pk = (P k
ij ) ∈ Mp(Mn)+ and Kk =
(K k
ij) ∈ Mp(Jn) are sequences such that kH − Pk − Kkk → 0, where H = (Hij ).
Applying the partial trace idp⊗tr leads to k(tr(Hij )−(tr(P k
ij ))k → 0. Consequently,
the set {(tr(P k
ij ) ∈ Mp : k ∈ N} is norm bounded. But this implies that the set
of traces of this set of matrices is bounded, which in turn implies that the set of
positive matrices Pk has bounded trace and hence is a norm bounded set. Thus, by
dropping to a subsequence if necessary, we may assume that Pk converges in norm
to some P that is necessarily in Mp(Mn)+, since the latter set is closed. But then
for this same subsequence, we have that Kk converges to an element K ∈ Mp(Jn)
and so H = P + K ∈ Mp(Mn)+ + Mp(Jn) and the result follows.
(cid:3)
The following result gives a variety of characterisations of positivity for the ma-
tricial ordering of the operator system Mn/Jn.
Proposition 2.3. The following statements are equivalent for matrices A11, Aij ∈
Mp, with 1 ≤ i, j ≤ n and j 6= i:
eij ⊗ Aij is positive in (Mn/Jn) ⊗ Mp;
ψ(eij ) ⊗ Aij is positive in Mr ⊗ Mp for every r ∈ N and
every ucp map ψ : Mn/Jn → Mr;
ψ(Eij ) ⊗ Aij is positive in Mr ⊗ Mp for every r ∈ N and
(1) 1 ⊗ A11 +Xj6=i
(2) 1 ⊗ A11 +Xj6=i
(3) 1r ⊗ A11 +Xj6=i
every ucp map ψ : Mn → Mr such that ψ(Jn) = {0};
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
11
ψ(Eij ) ⊗ Aij is positive in Mr ⊗ Mp for every r ∈ N and
(4) 1r ⊗ A11 +Xj6=i
(5) 1r ⊗ A11 +Xj6=i
every ucp map ψ : Mn → Mr such that ψ(Eii) = 1
n 1r, for all 1 ≤ i ≤ n;
Bij ⊗ Aij is positive in Mr ⊗ Mp for every r ∈ N and every
collection of matrices Bij ∈ Mr, j 6= i, with the property that
is positive in Mr ⊗ Mn;
(6) 1r ⊗ A11 +Xj6=i
1
n
Cij ⊗ Aij is positive in Mr ⊗ Mp for every r ∈ N and
every collection of matrices Cij ∈ Mr, j 6= i, with the property that
1
n 1r B12
1
n 1r
B21
...
Bn1
· · ·
. . .
. . . Bn,n−1
B1n
...
...
1
n 1r
1r C12
· · ·
C21
...
Cn1
1r
. . .
. . . Cn,n−1
C1n
...
...
1r
is positive in Mr ⊗ Mn;
1
n
n
(
(7) 1 ⊗ A11 +Xj6=i
Xi=1
(8) 1 ⊗ (nA11) +
B1, . . . , Bn ∈ Mp such that
Eii ⊗ Bi + Xj6=i
Xi=1
n
wiw∗
j ) ⊗ Aij is positive in C∗(Fn−1) ⊗ Mp;
Eij ⊗ Aij is positive in Mn ⊗ Mp for some
Bi = (n − n2)A11.
Rij = Aij for i 6= j and R11 + · · · + Rnn = nA11.
Proof. We shall prove that (1) ⇒ (9) ⇒ (8) ⇒ ... ⇒ (1).
i,j=1 Eij ⊗ Rij is positive in Mn ⊗ Mp for some Rij ∈ Mp such that
(9) Pn
The hypothesis (1) is that r = 1⊗A11+Xj6=i
such that q ⊗ idp(R) = r. In writing R as R =Pi Eii ⊗ Rii +Pj6=i Eij ⊗ Rij, we
obtain from r =Pi,j eij ⊗Rij that Rij = Aij for all j 6= i and that A11 = 1
nPi Rii.
Thus (9) follows. To see (8), set Rii = nA11 + Bi, where Bi = −Pk6=i Rkk, and
so nA11 = Pi Rii = n2A11 +Pi Bi implies that Pi Bi = (n − n2)A11, which is
Because Jn ⊆ Mn is completely order proximinal, there is a positive R ∈ Mn ⊗Mp
eij ⊗Aij is positive in (Mn/Jn)⊗Mp.
statement (8).
n
Now assume (8). Because 1 ⊗ nA11 +
Eij ⊗ Aij is positive in
Xi=1
Eii ⊗ Bi + Xj6=i
Mn ⊗Mp, the image of this matrix under the map φ⊗ idp is positive in C∗(Fn−1)⊗
Mp, where φ : Mn → C∗(Fn−1) is the ucp map (1) in Lemma 2.1. That is, the
12
DOUGLAS FARENICK AND VERN I. PAULSEN
following element is positive:
n
1
n
1 ⊗ (nA11) +
Xi=1
= 1 ⊗ nA11 + 1
Xi=1
n
n
1
n
wiw∗
j ⊗ Aij
wiw∗
j ⊗ Aij
1 ⊗ Bi +Xj6=i
Bi! +Xj6=i
1
n
= 1 ⊗ A11 +Xj6=i
1
n
wiw∗
j ⊗ Aij ,
n
Bi = (1 − n)A11. This establishes (7).
because 1
n
Xi=1
Next assume (7): namely, 1 ⊗ A11 +Xj6=i
Mp. Let r ∈ N and suppose that the matrices Cij ∈ Mr, j 6= i, satisfy
(
1
n
wiw∗
j ) ⊗ Aij is positive in C∗(Fn−1) ⊗
(2)
∈ (Mr ⊗ Mn)+ .
Let A = C1/2, the positive square root of C. We may express A in two ways:
C =
A =
1r C12
· · ·
C21
...
Cn1
1r
. . .
. . . Cn,n−1
C1n
...
...
1r
a1
a2
...
an
1
= A∗ = (cid:2) a∗
C = AA∗ = (cid:2)aia∗
a∗
2
. . .
a∗
n (cid:3) ,
where each aj is a rectangular block matrix, or an n-tuple, of r × r matrices. Hence,
Because aja∗
we may dilate a∗
.
j = 1 for each j, the rectangular matrix a∗
j(cid:3)1≤i,j≤n
j to a unitary which we denote by u∗
j is an isometry. Therefore,
j by way of the Halmos dilation:
u∗
j = (cid:20) a∗
j
0
(1 − a∗
j aj)1/2
−aj
(cid:21) ∈(cid:20) Mnr,n Mnr
Mn Mn,nr (cid:21) = M(nr+n) .
By the Universal Property, there is a homomorphism π : C∗(Fn−1) → M(nr+n)
such that π(wj ) = u∗
j for each j. If j 6= i, then
i a∗
j
∗
j ) = (cid:20) a∗
π(wiw∗
∗
∗ (cid:21) .
Therefore, if ψ : C∗(Fn−1) → M(nr+n) is the ucp map ψ = v∗πv, where v : Cr →
Cnr+n embeds Cr into the first r-coordinates of Cnr+n, then ψ(wiw∗
j ) = Cij . Thus,
ψ(p)
1 ⊗ A11 +Xj6=i
wiw∗
j ) ⊗ Aij
(
1
n
= 1r ⊗ A11 +Xj6=i
1
n
Cij ⊗ Aij
is positive in Mr ⊗ Mp, which yields statement (6).
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
13
Assume (6) holds. For any set of matrices Bij ∈ Mr, j 6= i, for which
1
n 1r B12
1
B21
n 1r
...
Bn1
B =
· · ·
. . .
. . . Bn,n−1
B1n
...
...
1
n 1r
(3)
is positive in Mr ⊗ Mn, let Cij = nBij. Thus, the matrix B above is 1
C is the matrix in (2). Therefore, from statement (6) one obtains (5).
n C, where
Assuming (5), let ψ : Mn → Mr be a ucp map such that ψ(Eii) = 1
n 1r for all
1 ≤ i ≤ n. Let Bij = ψ(Eij ), j 6= i. Therefore, the Choi matrix of ψ, namely
[ψ(Eij )] ∈ Mr ⊗ Mn, is precisely the matrix B in equation (3). Because ψ is
completely positive, the Choi matrix B of ψ is positive in Mr ⊗ Mn. Hence, (4) is
implied by (5).
Next assume (4). If ψ : Mn → Mr is ucp and has the property that ψ(Jn) =
{0}, then Eii − Ejj ∈ Jn, for all j 6= i. Because ψ is unital, this implies that
ψ(Eii) = 1
n 1r, for each i, which yields (3).
Assume (3) and let ψ : Mn/Jn → Mr be a ucp map. The linear map ψ : Mn →
Mr defined by ψ = ψ ◦ q is unital, annihilates Jn, and is the composition of cp
maps ψ and q. Therefore, (2) follows directly from (3).
Assume (2): that is,
ψ(p)(h) is positive in Mr ⊗ Mp, for every r ∈ N and
every ucp map ψ : Mn/Jn → Mr, and where h ∈ (Mn/Jn) ⊗ Mp is given by
eij ⊗Aij. In any operator system S, an element x ∈ S is positive if
h = 1⊗A11 +Xj6=i
and only if ω(x) is positive in Mr, for every r ∈ N and every ucp map ω : S → Mr
(see, for example, [5, p. 178]). Now let S = (Mn/Jn)⊗Mp and choose an arbitrary
ucp ω : S → Mr. Thus, there are ucp maps ψ1, . . . , ψm : Mn/Jn → Mr and linear
maps γ1, . . . , γm : Cp → Cr ⊗ Cp such that
ω =
m
Xj=1
γ∗
j
ψ(p)
j γj .
Hence, (1) is implied by (2).
(cid:3)
Theorem 2.4. The ucp surjection φ : Mn → Wn is a complete quotient map and
the C∗-envelope of Mn/Jn is C∗(Fn−1).
Proof. By the First Isomorphism Theorem, there exists a ucp map φ : Mn/Jn →
C∗(Fn−1) such that φ = φ ◦ q, where φ : Mn → C∗(Fn−1) is the ucp map defined
j , for all 1 ≤ i, j ≤ n, and so φ maps
in equation (1). Therefore, φ(eij ) = 1
Mn/Jn surjectively onto Wn. Hence, as Mn/Jn and T are vector spaces of equal
finite dimension, the surjection φ is a linear isomorphism. The equivalence of
statements (1) and (7) in Proposition 2.3 shows, further, that φ is a complete order
isomorphism of Mn/Jn and Wn.
n wiw∗
We now show that the C∗-envelope of Mn/Jn is C∗(Fn−1). Because there is a
completely isometric embedding of Mn/Jn into its C∗-envelope C∗
e (Mn/Jn) [14,
Chapter 15], we assume without loss generality that Mn/Jn is an operator system
in C∗
e (Mn/Jn). We also
suppose that C∗(Fn−1) has been represented faithfully on a Hilbert space H. By
e (Mn/Jn) and that 1 is the multiplicative identity of C∗
14
DOUGLAS FARENICK AND VERN I. PAULSEN
Arveson's extension theorem, the ucp map φ : Mn/Jn → C∗(Fn−1) ⊂ B(H) has a
e (Mn/Jn) → B(H). Let = v∗πv denote a minimal Stinespring
ucp extension : C∗
decomposition of . With respect to the decomposition of the representing Hilbert
space Hπ as Hπ = ran v ⊕ (ran v)⊥, every operator π(neij) has the form
π(neij ) = (cid:20) wiw∗
∗
j
∗
∗ (cid:21) .
Because kneijk = 1 and π is a contraction, the operator matrix above has norm
exactly equal to 1. And since wiw∗
j is unitary, if either of the (1,2)- or (2,1)-entries
of π(neij ) were nonzero, then the norm of π(neij ) would exceed 1. Thus,
π(neij ) = (cid:20) wiw∗
0
j
0
∗ (cid:21) ,
which implies that is multiplicative on the generators eij of the C∗-algebra
C∗
e (Mn/Jn)
to the generators of C∗(Fn−1), which implies that is surjective.
e (Mn/Jn). Hence, is a homomorphism that maps the generators of C∗
e (Mn/Jn) → C∗
Because φ is a unital complete order isomorphism, there is a unital isomorphism
e (Wn) such that αMn/Jn = πe ◦ φ, where πe is the unique sur-
α : C∗
jective unital homomorphism C∗(Fn−1) = C∗(Wn) → C∗
e (Wn) arising from the Uni-
versal Property of the C∗-envelope. Therefore, πe(wj ) is unitary, for each 2 ≤ j ≤ n,
and hence so are e1j in C∗
e (Mn/Jn), for 2 ≤ j ≤ n. By the Universal Property of
C∗(Fn−1), there is a unital homomorphism β : C∗(Fn−1) → C∗
e (Mn/Jn) such that
β(wj ) = e1j, 2 ≤ j ≤ n. Thus, β ◦ (wj ) = wj, for every j, which implies that is
in fact an isomorphism.
(cid:3)
Corollary 2.5. If u1, . . . , un are universal unitaries, then the C∗-envelope of the
operator system Span{uiu∗
j : 1 ≤ i, j ≤ n} is C∗(Fn−1).
Proof. Set wi = u∗
1ui, and note that w2, . . . , wn is a set of n − 1 universal unitaries.
(cid:3)
Another application of Proposition 2.3 is the following theorem concerning the
factorisation of positive elements in the C∗-algebra C∗(Fn) ⊗ Mp.
Theorem 2.6. Let u1, . . . , un be universal unitaries and let Aij ∈ Mp. Then P =
j ⊗ Aij is positive in C∗(Fn) ⊗ Mp if and only if there exist matrices
i,j=1 uiu∗
k=1 YkY ∗
Proof. We have that the span of uiu∗
via the map uiu∗
Pn
Xik ∈ Mp, such that P =Pn
Pi6=j eij ⊗ Aij. Apply Proposition 2.3(9) to lift this element to a positive element
R =Pn
i,j=1 Eij ⊗Rij of Mn⊗Mp. Now factor R = XX ∗ with X =Pn
j is completely order isomorphic to Mn/Jn
j → eij with eii = (1/n) 1. Thus, P → 1 ⊗ [1/n(A11 + · · · + Ann)] +
k , where Yk =Pn
and define Yk as above.
i,j=1 Eij ⊗Xij
i=1 ui ⊗ Xik.
We have that
YkY ∗
k =
n
Xk=1
n
Xi,j,k=1
and the result follows.
uiu∗
j ⊗ XikX ∗
jk =
uiu∗
j ⊗ Rij = P ,
n
Xi,j=1
(cid:3)
To conclude this section, we determine below the dual operator system of Wn.
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
15
Proposition 2.7. Consider the operator subsystem En ⊆ Mn defined by
(4)
Then the operator systems W d
En = {A ∈ Mn : aii = ajj , 1 ≤ i, j ≤ n} .
n and En are completely order isomorphic.
Proof. The function φ : Mn → Wn defined in (1) is a complete quotient map.
n → Md
Therefore, the dual map φd : W d
n is a complete order injection (Proposition
1.8). The operator systems Md
n and Mn are completely order isomorphic via the
map Sij 7→ Eij, where {Sij}1≤ı,j≤n is the basis of Md
n that is dual to the basis
{Eij}1≤i,j≤n of Mn [16]. By Banach space duality, φd(W d
n) is the annihilator of
Jn, namely En.
(cid:3)
3. Tensor Products with Mn/Jn
Definition 3.1. Let S be an operator system.
(1) For a fixed n ∈ N, S is said to have property (Wn) if, for every p ∈ N and
all S11, Sij ∈ Mp(S), where 1 ≤ i, j ≤ n and j 6= i, for which
1 ⊗ S11 + Xj6=i
(
1
n
wiw∗
j ) ⊗ Sij ≥ 0 in C∗(Fn−1) ⊗min Mp(S) ,
ij ∈ Mp(S), 1 ≤ i, j ≤ n, such that
then for every ε > 0 there exist Rε
(a) Rε = [Rε
(b) Rε
ij = Sij for all j 6= i, and
n
ij]1≤i,j≤n is positive in Mn(Mp(S)),
(c)
Rε
ii = n(S11 + ε1Mp(S)).
Xi=1
(2) We say that S has property (W) if S has property (Wn) for every n ∈ N.
Property (W) is suggested by Proposition 2.3 (equivalent statements (7) and
(9)), which shows that every matrix algebra Mp has property (W). To say that
an operator system S has property (Wn) is equivalent to saying that the map
φ ⊗ idS : Mn ⊗min S → Wn ⊗min S is a complete quotient map (see Proposition 6.1
for a detailed argument).
The main results of this section are summarised by the following theorem.
Theorem 3.1. (Operator Systems with Property (W))
(1) An operator system S has property (Wn) if and only if (Mn/Jn) ⊗min S =
(Mn/Jn) ⊗max S.
(2) Every (min,max)-nuclear operator system has property (W).
(3) B(H) has property (W).
(4) If an operator system S has property (Wn), then C∗(Fn−1) ⊗min S =
C∗(Fn−1) ⊗max S.
Corollary 3.2. (Kirchberg's Theorem) C∗(F∞) ⊗min B(H) = C∗(F∞) ⊗max B(H).
Proof. Recall that the free groups embed into one another as subgroups; in particu-
lar, if n ≥ 2, then Fn−1 ⊂ Fn and F∞ ⊂ Fn as subgroups. These subgroup emebed-
dings lead to C∗-algebra embeddings C∗(Fn−1) ⊂ C∗(Fn) and C∗(F∞) ⊂ C∗(Fn),
for n ≥ 2. Hence, every representation of the algebraic tensor product C∗(Fn) ⊗
B(H) induces (by restriction) a representation of C∗(F∞) ⊗ B(H), and conversely
every representation of C∗(F∞) ⊗ B(H) induces a representation of C∗(Fn) ⊗ B(H).
Thus, in consideration of the definition of the max-norm, if a ∈ C∗(Fn) ⊗ B(H),
16
DOUGLAS FARENICK AND VERN I. PAULSEN
then the max-norm of a in C∗(F∞) ⊗max B(H) coincides with the max-norm of
a in C∗(Fn) ⊗max B(H). That is, C∗(Fn) ⊗max B(H) ⊂ C∗(F∞) ⊗max B(H) as a
C∗-subalgebra. As C∗(F∞) ⊗max B(H) ⊆ C∗(F∞) ⊗min B(H), we show below that
there is a monomorphism of ρ : C∗(F∞) ⊗min B(H) → C∗(F∞) ⊗max B(H).
By (3) and (4) of Theorem 3.1, C∗(Fn) ⊗min B(H) = C∗(Fn) ⊗max B(H) ⊂
C∗(F∞) ⊗max B(H) for every n ∈ N, and so for each n there is a monomorphism
ρn : C∗(Fn) ⊗min B(H) → C∗(F∞) ⊗max B(H) such that ρn−1 = ρn ◦ ιn−1, where
ιn−1 : C∗(Fn−1)⊗min B(H) → C∗(Fn)⊗min B(H) is the canonical inclusion. Because
C∗(F∞)⊗minB(H) is the direct limit of the directed system (C∗(Fn)⊗minB(H), ιn)n,
there is a unique monomorphism ρ : C∗(F∞) ⊗min B(H) → C∗(F∞) ⊗max B(H) such
that ρn = ρ ◦ in, where in is the embedding of C∗(Fn) ⊗min B(H) into C∗(F∞) ⊗min
B(H) that is compatible with the inclusions ιn. Hence, C∗(F∞) ⊗min B(H) =
C∗(F∞) ⊗max B(H).
(cid:3)
The following proposition yields assertion (1) of Theorem 3.1.
Proposition 3.3. (Mn/Jn) ⊗min S = (Mn/Jn) ⊗max S if and only if S has
property (Wn).
Proof. Observe that if p ∈ N and H ∈ Mp(Mn/Jn ⊗ S), then H = [hkℓ]1≤k,ℓ≤p
and
hkℓ = 1 ⊗ S(k,ℓ)
11 + Xℓ6=k
ekℓ ⊗ S(k,ℓ)
ij
for some S(k,ℓ)
11
, S(k,ℓ)
ij
∈ S, where 1 ≤ k, ℓ ≤ p and j 6= i. Therefore, we write H as
where
H = 1 ⊗ S11 + Xℓ6=k
eij ⊗ Sij
S11 = [S(k,ℓ)
11
]1≤k,ℓ≤p
and Sij = [S(k,ℓ)
ij
]1≤k,ℓ≤p .
Hence, to prove that S has property (Wn), any argument that shows the property
for p = 1 also shows the property for arbitrary p ∈ N.
Assume that (Mn/Jn) ⊗min S = (Mn/Jn) ⊗max S. We may also assume p = 1
by the observation above. Suppose that
1 ⊗ S11 + Xj6=i
(
1
n
wiw∗
j ) ⊗ Sij ∈ (C∗(Fn−1) ⊗min S)+ .
The ucp map φ : Mn → Wn given by (1) is a complete quotient map, and therefore
the map
φ ⊗ idS : (Mn/Jn) ⊗max S → Wn ⊗max S
is a complete order isomorphism. Hence, by the assumption that (Mn/Jn)⊗minS =
(Mn/Jn) ⊗max S, we deduce that
h = 1 ⊗ S11 + Xj6=i
eij ⊗ Sij
is positive in (Mn/Jn) ⊗min S. By hypothesis, h ∈ ((Mn/Jn) ⊗max S)+ and so,
for every ε > 0,
h + ε1 =
1 ⊗ (S11 + ε1) +Xj6=i
eij ⊗ Sij
1
∈ Dmax
(Mn/Jn, S) .
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
17
Thus, there are A ∈ Mk(Mn/Jn)+, C ∈ Mm(S)+, and α : Ck ⊗ Cm → C linear
such that
h + ε1 = α(A ⊗ C)α∗ .
By Proposition 2.3, the fact that A ∈ Mk(Mn/Jn)+ = (Mn/Jn ⊗ Mk)+ implies
that there exist R = [Rij] ∈ (Mn ⊗ Mk)+ such that q(k)(R) = A. Thus,
h + ε1 = α(q(R) ⊗ C)α∗ = q ⊗ idS (α(R ⊗ C)α∗) ∈ Dmax
1
(Mn, S) .
That is, h+ε1 = q⊗idS(Rε), where Rε = α(R⊗C)α∗ ∈ (Mn ⊗maxS)+ = Mn(S)+.
Therefore, we have proved that for every ε > 0 there exists Rε ∈ Mn(S)+ such
that
1 ⊗ (S11 + ε1) +Xj6=i
eii ⊗ Rε
eij ⊗ Rε
ij ,
ii +Xj6=i
ij = Sij for j 6= i and 1
ii = S11 + ε1. Hence, S has property
eij ⊗ Sij = Xi
nPi Rε
Conversely, suppose that S has property (Wn). To prove that Cmin
(Mn/Jn, S) ⊆
(Mn/Jn, S) for all p ∈ N, we may, as mentioned earlier, restrict ourselves to
p
which yields Rε
(Wn).
p
Cmax
the case p = 1.
Suppose that h ∈ (Mn/Jn ⊗min S)+. Thus,
h = 1 ⊗ S11 + Xj6=i
eij ⊗ Sij
for some S11, Sij ∈ S, where 1 ≤ i, j ≤ n and j 6= i. Let φ : Mn → C∗(Fn−1) be
the ucp map defined by (1). Thus,
φ ⊗ idS : Mn/Jn ⊗min S → C∗(Fn−1) ⊗min S
sends h to
1 ⊗ S11 + Xj6=i
(
1
n
wiw∗
j ) ⊗ Sij ∈ (C∗(Fn−1) ⊗min S)+ .
By hypothesis, S has property (W), and so for every ε > 0 there exist Rε
1 ≤ i, j ≤ n, such that Rε = [Rε
n
ij ∈ S,
ij = Sij for all j 6= i, and
ij]1≤i,j≤n ∈ Mn(S)+, Rε
Rε
ii = n(S11 + ε1S). The cone Mn(S)+ is (Mn ⊗max S)+ and therefore the
Xi=1
ucp map
q ⊗ idS : Mn ⊗max S → (Mn/Jn) ⊗max S
sends Rε to the positive element
n
eij ⊗ Rε
ij = ε( 1 ⊗ 1S) + h ∈ ((Mn/Jn) ⊗max S)+ .
Xi,j=1
As this holds for all ε > 0, this implies, by the Archimedean property of the positive
cone in operator systems, that h ∈ ((Mn/Jn) ⊗max S)+.
(cid:3)
By definition, an operator system S is (min,max)-nuclear if S ⊗min T = S ⊗max T
for every operator system T . Thus, statement (2) of Theorem 3.1 follows immedi-
ately:
Corollary 3.4. If S is (min,max)-nuclear, then S has property (W).
We now prove statement (3) of Theorem 3.1.
18
DOUGLAS FARENICK AND VERN I. PAULSEN
Proposition 3.5. B(H) has property (W).
Proof. Suppose that S11, Sij ∈ B(H), 1 ≤ i, j ≤ n and j 6= i, and that
1 ⊗ S11 + Xj6=i
(
1
n
wiw∗
j ) ⊗ Sij ∈ (C∗(Fn−1) ⊗min B(H))+ .
(We are assuming, as in the proof of Proposition 3.3, that p = 1.) Let ε > 0.
Because B(K) is nuclear for all finite-dimensional Hilbert spaces K, if pL ∈ B(H)
ij ∈
ij = pLSijpL
is a projection onto a finite-dimensional subspace L ⊆ H, then there exist Rε,L
B(H), 1 ≤ i, j ≤ n, such that Rε,L = [Rε,L
ij ]1≤i,j≤n ∈ Mn(B(H))+, Rε,L
n
for all j 6= i, and
Xi=1
Rε,L
ii = n(pLS11pL + εpL). Therefore, F = {Rε,L}L is a
n
net in Mn(B(H))+ under subspace inclusion. Moreover, the fact that
Rε,L
ii =
Xi=1
n(pLS11pL + εpL) ≤ n(S11 + ε1) implies that the diagonal operators of each matrix
Rε,L are bounded. By the positivity of Rε,L, the off-diagonal operators are also
bounded, and hence F is a bounded net of positive operators. Let Rε = [Rε
ij] ∈
Mn(B(H))+ be the limit of some weakly convergent subnet F ′ = {Rε,L′ }L′ of F .
n
Then pL′ → 1 strongly and, therefore,
for all j 6= i, the proof is complete.
Xi=1
Rε
ii = n(S11 + ε1). Because Rε
ij = Sij
(cid:3)
The proof below of statement (4) of Theorem 3.1 is the final step in the proof of
this theorem.
Proposition 3.6. If S has property (Wn), then C∗(Fn−1)⊗minS = C∗(Fn−1)⊗max
S.
Proof. By [9, Theorem 6.7], C∗(Fn−1) ⊗c S = C∗(Fn−1) ⊗max S; therefore, we aim
to show that if x ∈ Mn((C∗(Fn−1) ⊗min S))+, then φ · ψ(n)(x) ∈ Mn(B(H))+ for
every pair of completely positive maps φ : C∗(Fn−1) → B(H) and ψ : S → B(H)
with commuting ranges. Let φ = v∗πv be a minimal Stinespring representation of φ,
where π : C∗(Fn−1) → B(Hπ) is a unital representation. By the Commutant Lifting
Theorem [2, Theorem 1.3.1], there is a unital homomorphism δ : φ(C∗(Fn−1))′ →
B(Hπ) such that δ(z)v = vz. for all z ∈ φ(C∗(Fn−1))′. Because ψ(S) lies within
the commutant of φ(C∗(Fn−1)), we obtain a completely positive map ψ = δ ◦ ψ :
S → B(Hπ) such that the range of ψ commutes with the range of π.
By hypothesis, S has property (Wn) and so, by Proposition 3.3 and by the fact
that Mn/Jn and Wn are completely order isomorphic, Wn ⊗min S = Wn ⊗max S.
Consider the ucp map π · ψ : Wn ⊗min S → B(Hπ). Because Wn ⊗min S is ucp
mapped into C∗(Fn−1) ⊗min C∗
e (S) and because B(Hπ) is injective, there is a ucp
e (S) → B(Hπ) of π · ψ. For each i, γ(wi ⊗ 1) =
extension γ : C∗(Fn−1) ⊗min C∗
π · ψ(wi ⊗ 1) = π(wi), and so wi ⊗ 1 is in the multiplicative domain Mγ of γ. Hence,
C∗(Fn−1) ⊗ {1} = C∗(Wn) ⊗ {1} ⊆ Mγ and, therefore,
γ(a ⊗ s) = γ((a ⊗ 1)(1 ⊗ s)) = γ(a ⊗ 1)γ(1 ⊗ s) = π(a) ψ(s) ,
for all a ∈ C∗(Fn−1), s ∈ S. Thus, on elementary tensors,
φ · ψ(a ⊗ s) = φ(a)ψ(s) = v∗π(s)vψ(s) = v∗π(s) ψ(s)v = v∗γ(a ⊗ s)v .
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
Therefore, φ · ψ = Adv ◦ γC∗(Fn−1)⊗minS and so (φ · ψ)(n)(x) ∈ Mn(B(H))+.
19
(cid:3)
4. The Operator System Tn/Jn
Definition 4.1. Let Tn = {A ∈ Mn : Aij = 0, ∀ i − j ≥ 2}, the operator system
of tridiagonal matrices.
Observe that Tn = Span {Eij
i − j ≤ 1} and that Tn contains the kernel
Jn = {D ∈ Mn : D is diagonal and tr D = 0} of the ucp map φ : Tn → C∗(Fn−1)
defined in (1) by φ(Eij ) = 1
j . Our interest in this section is with the quotient
operator system Tn/Jn.
:
n wiw∗
We begin with a useful fact about the operator system Tn itself.
Proposition 4.1. Tn ⊗min S = Tn ⊗c S for every operator system S. That is, Tn
is (min,c)-nuclear.
Proof. Let G = (V, E) be the graph with vertex set V = {1, . . . , n} and edge set
E = {(i, j) : i − j = 1}. Thus, G is simply a line segment from vertex 1 through
to vertex n. The operator system SG ⊂ Mn of the graph G is the span of matrix
units Eij ∈ Mn for which (i, j) ∈ E, and so SG = Tn. Because G is a chordal
graph, Tn = SG is (min,c)-nuclear [10, Proposition 6.10].
(cid:3)
Theorem 4.2. Assume that u1, . . . , un−1 are n−1 universal unitaries that generate
the full group C∗-algebra C∗(Fn−1). Let u0 = 1, u−j = u∗
j , and
Sn−1 = Span {1, uj, u∗
j : 1 ≤ j ≤ n − 1} = Span {uj : −n + 1 ≤ j ≤ n − 1} .
Consider the function φ : Tn → Sn−1 defined by
(5)
Then:
φ(Eij ) =
1
n
uj−i, ∀ i − j ≤ 1 .
(1) ker φ = Jn, the vector subspace of all diagonal matrices D ∈ Mn with
tr(D) = 0,
(2) Jn ⊂ Tn is completely order proximinal, and
(3) φ is a complete quotient map.
Proof. Because Jn is a kernel in Mn and Jn ⊂ Tn ⊂ Mn, Proposition 1.5 shows
that we may assume without loss of generality that Tn/Jn ⊂ Mn/Jn. The function
φ defined in (5) above is simply the function φTn , the restriction of the the function
φ defined previously in (1) to the operator subsystem Tn, but where we identify
each product wiw∗
i+1 with ui, for all 1 ≤ i ≤ (n − 1). (Assuming that w2, . . . , wn
n − 1 are universal unitaries, then so are u1, . . . , un−1.) Therefore, the restriction of
φ from Mn to Tn preserves complete positivity. The proof of Lemma 2.2 shows that
Jn ⊂ Tn is completely order proximinal, and it is evident that the restriction of φ
from Mn/Jn to Tn/Jn is a complete order isomorphism of Tn/Jn and Sn−1. (cid:3)
Corollary 4.3. The following statements are equivalent for Ai ∈ Mp, −n + 1 ≤
i ≤ n − 1:
(1)
n−1
Xi=1−n
ui ⊗ Ai ∈ Mp(Sn−1)+;
20
DOUGLAS FARENICK AND VERN I. PAULSEN
(2) there exists R = [Rij ] ∈ Mp(Mn)+ such that Rij = 0 if i − j ≥ 2,
n
Ri,i+1 = Ai and Ri+1,i = A−i, for all 1 ≤ i ≤ (n − 1), and
Rii = A0.
Xi=1
Proof. By statements (7) and (9) of Proposition 2.3, the element
n−1
Xi=1−n
ui ⊗ Ai = 1 ⊗ A0 +
1
n
n−1
(
Xi=1
wi+1w∗
i ) ⊗ (nA−i) +
1
n
n−1
(
Xi=1
wiw∗
i+1) ⊗ (nAi)
is positive in C∗(Fn−1) ⊗ Mp if and only if there is a R = [ Rij ]i,j ∈ Mn(Mp)+
with Rij = 0 if i − j ≥ 2, Ri,i+1 = nAi and Ri+1,i = nA−i, for all 1 ≤ i ≤ (n − 1),
n
and
Rii = nA0. This last statement is equivalent to the matrix R = 1
n
R having
Xi=1
the properties in (2) above.
Theorem 4.4. Let Un and Vn be the operator subsystems of
(cid:3)
M2 defined by
n−1
Mk=1
Vn = (n−1
Mk=1(cid:20) ak
11 ak
12
21 ak
ak
22 (cid:21) : ak
22 = ak+1
11 , ∀ 1 ≤ k ≤ (n − 1))
Un = (n−1
Mk=1(cid:20) ak
11 ak
12
21 ak
ak
22 (cid:21) : ak
ii = aℓ
jj , ∀ 1 ≤ k, ℓ ≤ (n − 1) , i, j = 1, 2) .
(6)
and
(7)
Then
(1) T d
(2) Sd
n and Vn are completely order isomorphic, and
n−1 and Un are completely order isomorphic.
Proof. The canonical linear basis of Tn is given by the subset {Eij }i−j≤1 of the
matrix units of Mn. Let {Sij}i−j≤1 ⊂ T d
n that is dual to this
canonical basis of Tn.
n be the basis of T d
n−1
Consider the operator system
n−1
Mk=1
M2 and for each 1 ≤ k ≤ (n − 1) define a
ucp map ρk :
M2 → M2 by projection onto the k-th coordinate. Let {fj}2n−2
j=1
and {e1, e2} be the canonical orthonormal bases of C2n−2 and C2, respectively, and
let γk : C2n−2 → C2 be the linear map defined by γk(f2k) = e1, γk(f2k+1) = e2,
Mk=1
M2 → Tn be the ucp map defined by
n−1
Mk=1
and γk(fj) = 0 for all other j. Let θ :
kρkγk. Thus,
θ =Pk γ∗
θ ((Ak)k) =
n−1
Xk=1
2
Xi,j=1
ak
ij Ei+k−1,j+k−1 ,
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
21
n−1
where Ak = [ak
ij ]i,j ∈ M2. The dual map θd : T d
n →
θd
Xi−j≤1
aijSij
=
n−1
Mk=1(cid:20)
akk
ak+1 k
M2 is given by
Mk=1
ak+1 k+1 (cid:21) .
ak k+1
Now let ψ = θd and observe that the range of ψ is precisely Vn. Thus, ψ is a
completely positive linear isomorphism between T d
n and Vn. We now show that
ψ−1 is completely positive.
To this end, suppose that X ∈ Mp(T d
n ) and Y = ψ(p)(X) ∈ Mp(Vn)+. Our aim
is to prove that X is positive, which is to say that X is a positive linear functional
on Mp(Tn). The matrix Y is a p × p matrix of n × n tridiagonal matrices; hence,
y is an n × n tridiagonal matrix of p × p matrices yk
ij ∈ Mp:
Y =
n−1
Mk=1(cid:20)
ykk
y∗
k k+1
yk k+1
yk+1 k+1 (cid:21) .
The pull back of Y to X ∈ Mp(T d
n ) is
X =
y11
y∗
12
y12
y22
y∗
23
y23
y33
y34
. . .
ynn
.
View the matrix X above as a partial matrix in the sense that outside the tridiagonal
band the entries are not specified. The only fully specified square submatrices of
yk k+1
yk+1 k+1 (cid:21),
X are the ones corresponding to 2 × 2 principal submatrices (cid:20)
ykk
y∗
k k+1
each of which is a direct summand of the positive matrix Y . Hence, the partially
specified matrix X can be completed to a positive matrix X (see, for example,
[15, Theorem 4.3]). The action of X on Mn(Mp) as a linear functional is given
by Z 7→ tr(Z X), and so X is a restriction of X as a linear functional on Tn.
Therefore, X is also positive, which completes the proof that ψ is a complete order
isomorphism.
By Thorem 4.2, φ : Tn → Sn−1 is a complete quotient map; therefore, ψ ◦ φd :
Sd
n−1 → Vn is a complete order injection. We need only identify the range of
ψ ◦ φd. Since Sn−1 is completely order isomorphic to Tn/Jn, the dual of Sn−1 is
the annihilator of Jn in Md
(cid:3)
n. Hence, ψ ◦ φd(Sn−1) = Un.
5. The Kirchberg Problem and the Connes Embedding Problem
Recall that the operator systems En, Vn, and Un of n × n matrices are defined
in (4), (6), and (7) respectively. In this section we show that some matrix theory
problems involving these operator systems, if solved affirmatively, would lead to an
affirmative solution to Kirchberg's Problem and, hence, to the Connes Embedding
Problem as well.
The following lemma is a proof technique that we shall require.
22
DOUGLAS FARENICK AND VERN I. PAULSEN
Lemma 5.1. Suppose that R, S, T , and U are operator systems and that, for
linear transformations ψ, θ, µ, and ν, where ν is a complete quotient map, µ is
a complete order isomorphism, θ is a linear isomorphism, and θ−1 is completely
positive, the following diagram is commutative:
R
µ
−−−−→ U
(8)
ψy
ν
y
S −−−−→
T .
θ
Then ψ is a complete quotient map if and only if θ is a complete order isomorphism.
Proof. Suppose that ψ is a complete quotient map. Thus, if y ∈ Mp(S)+, then
there is a hermitian x ∈ Mp(R) such that for every ε > 0 there is a kε ∈ ker ψ such
that x′ = ε1R + x + kε ∈ Mp(R)+. We have ker ψ ⊆ ker(ν ◦ µ), as θ ◦ ψ = ν ◦ µ;
thus, εν ◦ µ(1) + ν ◦ µ(x) = ν ◦ µ(x′) ∈ Mp(T )+. As this is true for every ε > 0,
ν ◦ µ(x) = θ ◦ ψ(x) = θ(s) ∈ Mp(T )+, which proves that θ is cp and, hence, that θ
is a complete order isomorphism.
Conversely, if θ is a complete order isomorphism, then ψ = θ−1 ◦ ν ◦ µ. Because
ν is a complete quotient map and θ−1 and µ are complete order isomorphisms, one
easily deduces that ψ is a complete quotient map.
(cid:3)
We show in Theorems 5.2(3) and 5.4(2) below that an affirmative solution of
two problems concerning the tensor products of operator systems of matrices will
result in an affirmative solution to the Kirchberg Problem.
Consider the following commutative diagram of vector spaces and linear trans-
formations:
(9)
Mn ⊗min Mn
∼=−−−−→ Mn ⊗max Mn
φ⊗φy
Wn ⊗min Wn −−−−−−→
idWn⊗Wn
y
Wn ⊗max Wn .
φ⊗φ
(complete quotient map)
Above, the identity map on Mn ⊗ Mn is a complete order isomorphism between
Mn ⊗min Mn and Mn ⊗max Mn, and the map φ ⊗ φ : Mn ⊗max Mn → Wn ⊗max
Wn is a complete quotient map by Corollary 1.7. The map θ = idWn⊗Wn is a
linear isomorphism of Wn ⊗min Wn and Wn ⊗max Wn with θ−1 completely positive.
Therefore, Lemma 5.1 asserts that leftmost arrow on the diagram is a complete
quotient map if and only if the bottom arrow is a complete order isomorphism.
The following result captures this fact and two additional equivalences.
Theorem 5.2. The following statements are equivalent:
(1) the map φ ⊗ φ : Mn ⊗min Mn → Wn ⊗min Wn is a complete quotient map;
(2) Wn ⊗min Wn = Wn ⊗max Wn;
(3) En ⊗min En = En ⊗max En;
(4) Wn has property (Wn).
If any of these equivalent statements is true for every n ∈ N, then the Kirchberg
Problem has an affirmative solution.
Proof. The equivalence of (1) and (2) is, as mentioned above, a consequence of
Corollary 1.7 and Lemma 5.1.
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
23
Statements (2) and (3) are equivalent by duality. Indeed, by Proposition 2.7,
the operator systems W d
n and En are completely order isomorphic. Therefore, the
hypothesis En⊗minEn = En⊗maxEn is equivalent to W d
n and,
by passing to duals (Proposition 1.9), is equivalent to Wn ⊗max Wn = Wn ⊗min Wn.
Theorem 3.1(4) asserts that Wn has property (Wn) if and only if (Mn/Jn) ⊗min
Wn = (Mn/Jn) ⊗max Wn. Since Mn/Jn is completely order isomorphic to Wn,
we deduce that statements (2) and (4) are equivalent.
n⊗maxW d
n⊗minW d
n = W d
Suppose now that any one of the equivalent conditions holds, for every n ∈ N.
By Theorem 3.1(4), C∗(Fn−1) ⊗min Wn = C∗(Fn−1) ⊗max Wn. But this means that
C∗(Fn−1) has property (Wn) and so again Theorem 3.1(4) is invoked to conclude
that C∗(Fn−1) ⊗min C∗(Fn−1) = C∗(Fn−1) ⊗max C∗(Fn−1). As this is true for all
n, we obtain C∗(F∞) ⊗min C∗(F∞) = C∗(F∞) ⊗max C∗(F∞) by the direct limit
argument used in the proof of Corollary 3.2.
(cid:3)
In moving from the operator system Wn to the operator system Wn−1, it it is
interesting to contrast Theorem 5.2 above with Theorem 5.3 below.
Theorem 5.3. ([9]) The Kirchberg Problem has an affirmative solution if and only
if Sn ⊗min Sn = Sn ⊗c Sn for every n ∈ N.
Proof. If Sn ⊗min Sn = Sn ⊗c Sn for every n ∈ N, then C∗(F∞) has the weak
expectation property [9, Theorem 9.14] which in turn implies that C∗(F∞) ⊗min
C∗(F∞) = C∗(F∞) ⊗max C∗(F∞) [11, Proposition 1.1(i)].
Conversely, if C∗(F∞) ⊗min C∗(F∞) = C∗(F∞) ⊗max C∗(F∞), then C∗(F∞) has
WEP [11, Proposition 1.1(iii)], and so Sn ⊗min Sn = Sn ⊗c Sn for every n ∈ N [9,
Theorem 9.14].
(cid:3)
We now turn to our second multilinear algebra problem.
Theorem 5.4. The following statements are equivalent:
(1) φ ⊗ φ : Tn ⊗min Tn → Sn−1 ⊗min Sn−1 is a complete quotient map;
(2) Un ⊗max Un ⊆coi Vn ⊗max Vn.
If either of these equivalent statements is true for every n ∈ N, then the Kirchberg
Problem has an affirmative solution.
Proof. Assume (1) and let µ = φ ⊗ φ. Thus, µ is a complete quotient map. By
Theorem 4.4, T d
n−1 = Vn are complete order isomorphisms. Thus, if
µ is a complete quotient map, then
n = Un and Sd
µd : (Sn−1 ⊗min Sn−1)d → (Tn ⊗min Tn)d
is a complete order injection, but
and
(Sn−1 ⊗min Sn−1)d = Un ⊗max Un
(Tn ⊗min Tn)d = Vn ⊗max Vn .
Conversely, assume (2). That is, Un ⊗max Un ⊆coi Vn ⊗max Vn, and so
n ⊗max T d
n .
n−1 = Un ⊗max Un ⊆coi Vn ⊗max Vn = T d
n−1 ⊗max Sd
Sd
Denote this complete order injection Sd
in passing to duals, we obtain a complete quotient map
n−1 ⊗max Sd
ϑd : Tn ⊗min Tn = (T d
n−1 ⊗max Sd
n ⊗max T d
n )d → (Sd
n−1)d = Sn−1 ⊗min Sn−1.
n−1 → T d
n ⊗max T d
n by ϑ. Hence,
24
DOUGLAS FARENICK AND VERN I. PAULSEN
Since ϑd = µ, statement (1) follows.
Assume that one of (1) or (2) is true for all n ∈ N. By hypothesis, each y ∈
Mp(Sn−1 ⊗min Sn−1)+ is given by µ(p)(x) for some x ∈ Mp(Tn ⊗min Tn)+. Recall
from Proposition 4.1 that Tn is (min,c)-nuclear; therefore, in particular, Tn ⊗min
Tn = Tn ⊗c Tn. Thus, since Mp(Tn ⊗min Tn)+ = Mp(Tn ⊗c Tn)+ and because
µ : Tn ⊗c Tn → Sn−1 ⊗c Sn−1 is completely positive, y ∈ Mp(Sn−1 ⊗c Sn−1)+,
which implies that Sn−1 ⊗min Sn−1 = Sn−1 ⊗c Sn−1. As this is true for all n ∈ N,
Theorem 5.3 completes the argument.
(cid:3)
6. C∗-Algebras with the Weak Expectation Property
By considering the positive liftings in Proposition 2.3 in the case of the operator
system Tn, we are led to consider a special case of property (Wn), which we call
property (Sn).
Definition 6.1. An operator system S is said to have property (Sn) for a fixed
n ∈ N if, for every p ∈ N and all Si ∈ Mp(S), where 1 − n ≤ i ≤ n − 1 and j 6= i,
for which
n−1
ui ⊗ Si ≥ 0 in C∗(Fn−1) ⊗min Mp(S) ,
Xi=1−n
Xi=1−n
then for every ε > 0 there exist Rε
ij ∈ Mp(S), 1 ≤ i, j ≤ n, such that
(1) Rε = [Rε
(2) Rε
ij = 0 for all i − j ≥ 2, Rε
n
ij]1≤i,j≤n is positive in Mn(Mp(S)),
i,i+1 = Si, and Rε
i+1,i = S−i for all i, and
(3)
Rε
ii = S0 + ε1Mp(S).
Xi=1
We say that S has property (S) if it has property (Sn) for every n ∈ N.
Proposition 6.1. The following statements are equivalent for an operator system
S:
(1) S has property (Sn);
(2) φ ⊗ idS : Tn ⊗min S → Sn−1 ⊗min S is a complete quotient map.
Proof. To set up the argument, note that Mp(R ⊗min T ) and R ⊗min Mp(T )
are completely order isomorphic operator systems for any p ∈ N and operator
n−1
systems R and T .
If Z ∈ Sn−1 ⊗ Mp(S) is arbitrary, then Z =
ui ⊗ Si
for some Si ∈ Mp(S).
X ∈ Tn ⊗ Mp(S) is given by
In this case Z = [φ ⊗ idS ](p)(X), where (one choice of)
n
n
n
X =
Ejj ⊗ S0 +
Ei,i+1 ⊗ nSi +
Ei+1,i ⊗ nS−i .
Xj=1
Xi=1
Xi=1
Via the First Isomorphism Theorem, Z is the image of the quotient element X.
element X is positive if and only if for every ε > 0 there is a Kε = [K ε
Choose p and suppose that Z ∈ Sn−1 ⊗min Mp(S) is positive. By definition, the
i,j=1 ∈ Tn ⊗
ii = 0, and (iii) Rε = ε1 + X + Kε
Rε is positive.
is positive in Tn ⊗min Mp(S). But Rε is positive if and only if Rε = 1
n
ii = ε1Mp(S) + S0, Rε
Note that Rε = [Rε
i,i+1 = Si, and
Rε
Mp(S) such that (i) Kε is diagonal, (ii) Pn
ij ]i,j is tridiagonal withPi Rε
i+1,i = S−i.
i=1 K ε
ij]n
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
25
Thus, if S has property (Sn), then the positivity of Rǫ for every ε > 0 implies
X is positive. Conversely, if φ ⊗ idS is a complete quotient map, then the
X and, hence, the positivity of Rε also, for every
(cid:3)
that
positivity of Z implies that of
ε > 0.
Theorem 6.2. The following statements are equivalent for a unital C∗-algebra A:
(1) A has property (S);
(2) A has WEP.
Proof. Assume (1) holds and consider the following commutative diagram of vector
spaces and linear transformations:
Tn ⊗min A
∼=−−−−→ Tn ⊗max A
φ⊗idy
Sn−1 ⊗min A −−−−−−→
idSn−1⊗A
φ⊗id
y
Sn−1 ⊗max A .
The top arrow is a complete order isomorphism by Proposition 4.1 and [10, Theorem
6.7], and the leftmost arrow (φ ⊗ id) is a complete quotient map, by hypothesis. By
Proposition 1.6, the righthand arrow is also a quotient map. Now apply Lemma 5.1
to conclude that the bottom arrow is a complete order isomorphism.
Thus, Sn−1 ⊗min A = Sn−1 ⊗max A, for all n ∈ N. By the direct limit argument
of Corollary 3.2 we arrive at C∗(F∞) ⊗min A = C∗(F∞) ⊗max A, which is equivalent
to A having WEP. Thus, statement (1) is established.
Next assume (2). Thus, C∗(F∞) ⊗min A = C∗(F∞) ⊗max A, by [11, Proposition
1.1(iii)], and so C∗(Fn) ⊗min A = C∗(Fn) ⊗max A, for every n [9, Lemma 7.5].
Therefore, for every n,
Sn ⊗min A ⊆coi C∗(Fn) ⊗min A = C∗(Fn) ⊗max A .
By the inclusion and equality above,
Sn ⊗min A = Sn ⊗c A = Sn ⊗max A .
Every unital C∗-algebra is (c,max)-nuclear [10, Theorem 6.7]. Thus, using the fact
that Tn is (min,c)-nuclear, we deduce that
Tn ⊗min A = Tn ⊗c A = Tn ⊗max A .
We are therefore led to the following commutative diagram:
Tn ⊗min A
∼=−−−−→ Tn ⊗max A
φ⊗idy
y
φ⊗id
(complete quotient map)
Sn−1 ⊗min A −−−−→
Sn−1 ⊗max A .
∼=
Hence, by Lemma 5.1 the map φ ⊗ id : Tn ⊗min A → Sn−1 ⊗min A is a complete
quotient map for every n, and so A has property (S), which proves (1).
(cid:3)
By similar methods we can also prove:
Theorem 6.3. If Sn−1 has property (Sn) for every n ∈ N, then Kirchberg's Prob-
lem has an affirmative solution.
26
DOUGLAS FARENICK AND VERN I. PAULSEN
7. Injective Envelopes
In the earlier sections we have seen that operator system quotients of matrix
In this section, we examine their injective
spaces can have large C∗-envelopes.
envelopes.
Recall that the injective envelope [14, Chapter 15] of a unital C∗-algebra A is
denoted by I(A). An AW∗-algebra is a unital C∗-algebra B with the property
that, for any nonempty subset X ⊂ B, there is a projection e ∈ B such that
annR(X ) = {ey y ∈ B}, where
annR(X ) = {b ∈ B xb = 0, ∀ x ∈ X } .
Like von Neumann algebras, AW∗-algebras admit a decomposition into direct sums
of AW∗-algebras of types I, II, and III. An AW∗-factor is an AW∗-algebra with triv-
ial centre. The theory of AW∗-algebras is relevant to the study of injective envelopes
because every injective C∗-algebra is monotone complete and every monotone com-
plete C∗-algebra is an AW∗-algebra.
Theorem 7.1. If A 6= C is a unital, separable, prime C∗-algebra with only trivial
projections, then its injective envelope I(A) is a type III AW∗-factor with no normal
states.
Proof. Under the hypothesis given, the injective envelope of A cannot be a W∗-
algebra (that is, a von Neumann algebra): for if I(A) were a von Neumann algebra,
then the separability of A would imply that A has a nontrivial projection (see [1,
Theorem 2.2(v)] or [8, §3]), contrary to the hypothesis.
As A is separable, A has a faithful representation as a unital C∗-subalgebra of
B(H) for some separable Hilbert space H. The injective envelope of A is obtained as
follows [14]: there is a ucp projection φ : B(H) → B(H) mapping onto an operator
system I that contains A as a subsystem and such that I is an injective C∗-algebra
under the Choi -- Effros product x◦y = φ(xy), x, y ∈ I [5, 14]. The injective envelope
I(A) of A, when I(A) is viewed as a C∗-algebra, is precisely the operator system
I with the Choi -- Effros product ◦.
Because H is separable, B(H) has a faithful state ω. Let ϕ be the state on I
given by ϕ(x) = ω (φ(x)). By the Schwarz inequality, ϕ(x∗ ◦ x) = ω (φ(x∗x)) ≥
ω (φ(x)∗φ(x)), which implies that ϕ is a faithful state on the AW∗-algebra I(A).
The hypothesis that A is prime (that is, no two nonzero ideals can have zero
intersection) implies that the injective envelope I(A) is an AW∗-factor [7, Theorem
7.1]. We show that the only type of factor that I(A) could possibly be is a factor
of type III.
If I(A) were a factor of type I, then I(A) would be coincide with B(H) for some
Hilbert space H, implying that I(A) is a von Neumann algebra, which we have
argued is not possible. If I(A) were a factor of type II (finite) or II∞, then the fact
that I(A) admits a faithful state implies that I(A) is a von Neumann algebra [6],
which again is not possible. Therefore, I(A) must be of type III.
Finally, if I(A) were to have a normal state, then I(A) would have a direct
summand which is a von Neumann algebra. But the fact that I(A) is a factor
discounts the existence of a nonzero direct summand, and because I(A) is not a
W∗-algebra, we conclude that I(A) has no normal states.
(cid:3)
Corollary 7.2. The injective envelope of C∗(Fn), where n ≥ 2, is a type III AW∗-
factor with no normal states.
OPERATOR SYSTEM QUOTIENTS AND TENSOR PRODUCTS
27
Proof. If n ≥ 2, then C∗(Fn) is unital, separable, primitive (and therefore prime),
and has only trivial projections [4].
(cid:3)
Corollary 7.3. The injective envelopes of Mn/Jn and Tn/Jn coincide. When
n ≥ 2, their injective envelope is a type III AW∗-factor with no normal states.
Proof. The injective envelope of an operator system S and its C∗-envelope C∗
coincide. Because C∗
follows.
e (S)
e (Mn/Jn) = C∗(Fn−1) (Theorem 2.4), the result
(cid:3)
e (Tn/Jn) = C∗
Acknowledgement
This work was undertaken at Institut Mittag-Leffler (Djursholm, Sweden) in
autumn 2010. The authors wish to acknowledge the support and hospitality of the
institute during their stay there. The authors also wish to thank Ali S. Kavruk for
pointing out a mistake in an earlier version of this paper, Ivan Todorov for useful
commentary, and the referee for a very careful review of the manuscript and for
drawing our attention to reference [3]. The work of the first author is supported in
part by NSERC.
References
[1] M. Argerami and D. R. Farenick, Local multiplier algebras, injective envelopes, and type I
W ∗-algebras, J. Operator Theory 59 (2008), no. 2, 237 -- 245.
[2] W. Arveson, Subalgebras of C ∗-algebras, Acta Math. 123 (1969), 141 -- 224.
[3] F. P. Boca, A note on full free product C ∗-algebras, lifting and quasidiagonality, Operator
theory, operator algebras and related topics (Timi¸soara, 1996), Theta Found., Bucharest,
1997, pp. 51 -- 63.
[4] M. D. Choi, The full C ∗-algebra of the free group on two generators, Pacific J. Math. 87
(1980), no. 1, 41 -- 48.
[5] M. D. Choi and E. G. Effros, Injectivity and operator spaces, J. Functional Analysis 24
(1977), no. 2, 156 -- 209.
[6] G. A. Elliott, K. Sait¯o, and J. D. M. Wright, Embedding AW ∗-algebras as double commutants
in type I algebras, J. London Math. Soc. (2) 28 (1983), no. 2, 376 -- 384.
[7] M. Hamana, Regular embeddings of C ∗-algebras in monotone complete C ∗-algebras, J. Math.
Soc. Japan 33 (1981), no. 1, 159 -- 183.
[8]
, The centre of the regular monotone completion of a C ∗-algebra, J. London Math.
Soc. (2) 26 (1982), no. 3, 522 -- 530.
[9] A. S. Kavruk, V. I. Paulsen, I. G. Todorov, and M. Tomforde, Quotients, exactness and
nuclearity in the operator system category, arXiv:1008.2811v2 (2010), preprint.
[10]
[11] E. Kirchberg, On nonsemisplit extensions, tensor products and exactness of group C ∗-
, Tensor products of operator systems, J. Funct. Anal. 261 (2011), no. 2, 267 -- 299.
algebras, Invent. Math. 112 (1993), no. 3, 449 -- 489.
[12]
, Commutants of unitaries in UHF algebras and functorial properties of exactness, J.
Reine Angew. Math. 452 (1994), 39 -- 77.
[13] N. Ozawa, About the QWEP conjecture, Internat. J. Math. 15 (2004), no. 5, 501 -- 530.
[14] V. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Advanced
Mathematics, vol. 78, Cambridge University Press, Cambridge, 2002.
[15] V. I. Paulsen, S. C. Power, and R. R. Smith, Schur products and matrix completions, J.
Funct. Anal. 85 (1989), no. 1, 151 -- 178.
[16] V. I. Paulsen, I. G. Todorov, and M. Tomforde, Operator system structures on ordered spaces,
Proc. Lond. Math. Soc. (3) 102 (2011), no. 1, 25 -- 49.
[17] G. Pisier, A simple proof of a theorem of Kirchberg and related results on C ∗-norms, J.
Operator Theory 35 (1996), no. 2, 317 -- 335.
[18]
, Introduction to operator space theory, London Mathematical Society Lecture Note
Series, vol. 294, Cambridge University Press, Cambridge, 2003.
28
DOUGLAS FARENICK AND VERN I. PAULSEN
Department of Mathematics and Statistics, University of Regina, Regina, Saskatchewan
S4S 0A2, Canada
Department of Mathematics, University of Houston, Houston, Texas 77204-3476,
U.S.A.
|
0905.0169 | 3 | 0905 | 2015-09-09T20:53:24 | On Matrix-Valued Square Integrable Positive Definite Functions | [
"math.OA",
"math.FA",
"math.GR",
"math.RT"
] | In this paper, we study matrix valued positive definite functions on a unimodular group. We generalize two important results of Godement on square integrable positive definite functions to matrix valued square integrable positive definite functions. We show that a matrix-valued continuous $L^2$ positive definite function can always be written as a convolution of a $L^2$ positive definite function with itself. We also prove that, given two $L^2$ matrix valued positive definite functions $\Phi$ and $\Psi$, $\int_G Trace(\Phi(g) \bar{\Psi(g)}^t) d g \geq 0$. In addition this integral equals zero if and only if $\Phi * \Psi=0$. Our proofs are operator-theoretic and independent of the group. | math.OA | math | 1 2 3
ON MATRIX VALUED SQUARE INTEGRABLE POSITIVE DEFINITE
FUNCTIONS
HONGYU HE
Abstract. In this paper, we study matrix valued positive definite functions on a unimodular group.
We generalize two important results of Godement on L2 positive definite functions. We show that a
matrix-valued continuous L2 positive definite function can always be written as the convolution of an
matrix-valued L2 positive definite function with itself. We also prove that, given two L2 matrix valued
positive definite functions Φ and Ψ, RG T r(Φ(g)Ψ(g)
)dg ≥ 0. In addition this integral equals zero if
and only if Φ ∗ Ψ = 0. Our proofs are operator-theoretic and independent of the group.
t
5
1
0
2
p
e
S
9
]
.
A
O
h
t
a
m
[
3
v
9
6
1
0
.
5
0
9
0
:
v
i
X
r
a
1. Introduction
About 60 years ago, Godement published a paper on square integrable positive definite functions on
a locally compact group ( [1]). In his paper, Godement proved that every continuous square integrable
positive definite function has an L2-positive definite square root. He also proved, among others, that the
inner product between two positive definite L2-functions must be nonnegative. Godement's results and
proofs were quite elegant. The purpose of this paper is to extend Godement's theorem to matrix-valued
positive definite functions on unimodular groups. Obviously, the diagonal of matrix-valued positive
definite functions must all be positive definite. Yet, there is not much to say about the off-diagonal
entries and their relationship with diagonal entries. So Godement's results do not carry easily to the
matrix-valued case. In this paper, we generalize Godement's theorems to matrix-valued positive definite
functions ([1] and Ch 13.[2]). Our results, we believe, are new.
Let Mn(C) be the set of n × n matrices. For a matrix A, let [A]ij be the (i, j)-th entry of A. Let
G be a unimodular group. A continuous function Φ : G → Mn(C) is said to be positive definite if for
any {Ci ∈ Cn}l
i=1 and {xi ∈ G}l
i=1,
l
Xi,j=1
(Ci)tΦ(x−1
i xj)Cj ≥ 0.
t
Take x1 = e and x2 = g. The above inequality implies that Φ(g) = Φ(g−1)
(See for example, Prop.
2.4.6 [7]). When n = 1, our definition agrees with the definition of continuous positive definite functions.
We denote the set of continuous matrix-valued positive definite functions by P(G, Mn).
Definition 1.1. Let L1
loc(G, Mn) act on u ∈ Cc(G, Cn) by [λ(Φ)(u)(x)]i = Pn
L1
loc(G, Mn) be the set of Mn-valued locally integrable functions on G. Let Φ ∈
j=1R [Φ(g)]ij [u(g−1x)]jdg. We write λ(Φ)(u)(x) =
1This research is partially supported by NSF grant DMS-0700809.
2AMS Subject Classification: 43A35, 43A15, 22D
3Key word: Positive definite function, unimodular group, moderated functions, square integrable functions, convolution
algebra, unitary representations.
1
2
HONGYU HE
R Φ(g)u(g−1x)dg. Clearly, λ(Φ)u is continuous. We say that Φ is positive definite if
n
hλ(Φ)(u), ui =
for all u ∈ Cc(G, Cn).
ZG
Xi=1
[λ(Φ)u(g)]i[u(g)]idg ≥ 0
We denote the set of matrix-valued positive definite functions by P(G, Mn). Clearly, P(G, Mn) ⊃
P(G, Mn) (see Prop 13.4.4 [2]).
Definition 1.2. A matrix-valued function Φ(x) is said to be square integrable, or simply L2 if [Φ]ij is in
L2(G) for all (i, j). We denote the set of matrix-valued square integrable function by L2(G, Mn). Define
hΦ, Ψi = ZG
t
T rΦΨ
dg.
Put P 2(G, Mn) = L2(G, Mn) ∩ P(G, Mn) and P2(G, Mn) = L2(G, Mn) ∩ P(G, Mn).
Let Φ, Ψ ∈ L1
loc(G, Mn). Define the convolution
[Φ ∗ Ψ]ij =
n
Xk=1
[Φ]ik ∗ [Ψ]kj
whenever the right hand side is well-defined, i.e., the convolution integral converges absolutely.
Theorem [A] Let G be a unimodular locally compact group. Let Φ ∈ P 2(G, Mn). Then there ex-
ists a Ψ ∈ P2(G, Mn) such that Φ = Ψ ∗ Ψ.
Theorem [B] Let G be a unimodular locally compact group. Let Φ, Ψ ∈ P2(G, Mn). Then hΦ, Ψi ≥ 0.
Theorem [C] Let G be a unimodular locally compact group. Let Φ, Ψ ∈ P2(G, Mn). Then hΦ, Ψi = 0
if and only of Φ ∗ Ψ = 0.
Our motivation comes from the theory of unitary representations of Lie groups. There are represen-
tations that appear as a space of "invariant distributions "in a unitary representation (π,Hπ). To
construct a Hilbert inner product for the invariant distributions, one is often led to investigate whether
(1)
ZG
(π(g)u, u)dg ≥ 0
when (π(g)u, u) is L1, u ∈ Hπ and G is unimodular ([6] [4]). For G = R, the affirmative answer to
this question is a direct consequence of Bochner's theorem, namely, the integral of a L1 positive definite
function on R is nonnegative. In his thesis [1], Godement raised this question for G unimodular. It is
known that for G amenable, the Inequality (1) is always true (Prop 18.3.6 [2]). An amenable group is
characterized by the fact that the unitary dual is weakly contained in L2(G). Therefore for G nilpotent,
the Inequality (1) holds.
Consider the other extreme, namely, G semisimple and noncompact. The inequality (1) is false in
its full generality. Yet, applying the result of this paper, we show that Inequality (1) holds if Hπ can be
written as a tensor product of two L2-representations of G and u is finite in the tensor decomposition.
See Theorem 7.1. In Cor. 7.1, we give a result about a certain integral related to Howe's correspondence
([5]). It is more general than the results given in [4].
ON MATRIX VALUED SQUARE INTEGRABLE POSITIVE DEFINITE FUNCTIONS
3
2. Convolution Algebras
Let G be a unimodular locally compact group. A matrix-valued function on G is said to be in Lp if
each entry is in Lp(G). Let Φ ∈ L1(G, Mn). Define kΦkL1 = P k[Φ]ijkL1. Then L1(G, Mn) becomes a
Banach algebra. We have
L1(G, Mn) ∗ L1(G, Mn) ⊆ L1(G, Mn);
Cc(G, Mn) ∗ Cc(G, Mn) ⊆ Cc(G, Mn).
L2(G, Mn) ∗ L2(G, Mn) ⊆ BC(G, Mn).
Here BC(G, Mn) is the space of bounded continuous functions.
For each u, v ∈ L2(G, Cn), define the standard inner product hu, vi = RGPi[u(g)]i[v(g)]idg. Obviously
L1(G, Mn) acts on L2(G, Cn) by λ. Then map
λ : L1(G, Mn) → B(L2(G, Cn))
defines a bounded Banach algebra isomorphism. Notice that if Φ ∈ L2(G, Mn) and u ∈ L1(G, Cn), then
λ(Φ)u ∈ L2(G, Cn).
We define the ∗ operation on Lloc(G, Mn) by letting
[Φ∗(g)]ij = [Φ(g−1)]ji.
For u, v ∈ Cc(G, Cn), we have hλ(Φ)u, vi = hu, λ(Φ∗)vi. If Φ is positive definite in Lloc(G, Mn), then
hλ(Φ)u, ui = hλ(Φ)u, ui = hu, λ(Φ)ui = hλ(Φ∗)u, ui.
Hence hλ(Φ − Φ∗)u, ui = 0. Let u = v + tw (t ∈ R). Then
0 = hλ(Φ − Φ∗)(v + tw), v + twi = hλ(Φ − Φ∗)v, twi + hλ(Φ − Φ∗)tw, vi.
We obtain hλ(Φ − Φ∗)v, wi = hλ(−Φ + Φ∗)w, vi = hw, λ(−Φ∗ + Φ)vi = hλ(Φ − Φ∗)v, wi. Hence hλ(Φ −
Φ∗)v, wi must be real for all v, w ∈ Cc(G, Cn). It follows that Φ∗ = Φ. We have
Lemma 2.1. Let Ψ, Φ ∈ P2(G, Mn). Then Φ = Φ∗, Ψ = Ψ∗ and hΦ, Ψi = T r(Φ ∗ Ψ(e)).
Proof: The last statement follows from
(2)
ZG
T r(Φ ∗ Ψ(e)) = Xi,j
Definition 2.1 (Ch 13. [2]). Let Φ ∈ L2(G, Mn). We say that Φ is moderated if λ(Φ) on Cc(G, Cn) is
a bounded operator in the L2 norm, i.e., there is a M such that
[Φ]ij(g)[Ψ]ji(g−1)dg = Xi,j
T rΦ(g)Ψ(g)tdg = hΦ, Ψi.
[Φ]ij (g)[Ψ]ij(g)dg = ZG
ZG
for any u ∈ Cc(G, Cn).
kλ(Φ)uk ≤ Mkuk
When Φ is moderated, λ(Φ)Cc(G,Cn) can be extended to a bounded operator on L2(G, Cn) which
coincides with the operator λ(Φ)L2(G,Cn). To see this, let ui → u under the L2-norm with ui ∈
i=1 yields a Cauchy sequence in L2(G, Mn).
Cc(G, Cn). Since λ(Φ) is bounded on Cc(G, Mn), {λ(Φ)ui}∞
Therefore λ(Φ)ui converges to an L2-function v ∈ L2(G, Cn) . In particular, (λ(Φ)ui)(g) converges in
L2-norm to v(g) on any compact subset K. On the other hand, (λ(Φ)ui)(g) converges to λ(Φ)u(g)
uniformly on G, in particular, on K. Hence (λ(Φ)u)(g) = v(g) for g ∈ K almost everywhere. It follows
that v(g) = λ(Φ)u(g) almost everywhere. Therefore λ(Φ)u ∈ L2(G, Cn). In short, if Φ is moderated,
4
HONGYU HE
and u ∈ L2(G, Cn), then λ(Φ)(u) ∈ L2(G, Cn). We retain λ(Φ) to denote the bounded operator on
L2(G, Cn). The following is obvious.
Lemma 2.2. Φ is moderated if and only if [Φ]ij are all moderated in L2(G).
Let M (G, Mn) be the space of moderated L2 functions on G. Lemma 13.8.4 [2] asserts that M (G) ∗
M (G) ⊆ M (G) and λM(G) : M (G) → B(L2(G)) is an algebra homomorphism. Therefore, we obtain
Lemma 2.3. Let Φ, Ψ ∈ L2(G, Mn). If Φ and Ψ are moderated, then Φ ∗ Ψ is also moderated. In
addition λ(Φ ∗ Ψ) = λ(Φ)λ(Ψ).
3. Matrix-Valued Positive Definite Functions
Let G be a unimodular locally compact group. Let us recall some basic result from [2]. Let Φ, Ψ ∈
P(G, Mn). We define an ordering Φ (cid:22) Ψ if Ψ − Φ ∈ P(G, Mn). An immediate consequence is that
T r(Φ)(e) ≤ T r(Ψ)(e). Clearly, Φ (cid:22) Ψ if and only if
hλ(Φ)u, ui ≤ hλ(Ψ)u, ui
(∀ u ∈ Cc(G, Cn)).
For two bounded operators X and Y in B(H), we say that X (cid:22) Y if Y − X is positive (Ch 2.4 [7]).
Y − X is positive implies that Y − X is self-adjoint (Prop. 2.4.6 [7]). If Φ and Ψ are moderated, then
Φ (cid:22) Ψ if and only if λ(Φ) (cid:22) λ(Ψ).
Theorem 3.1 ( Prop. 16 [1]). Let Φ, Ψ be two moderated elements in P2(G, Mn). Suppose that Φ∗ Ψ =
Ψ ∗ Φ. Then hΦ, Ψi ≥ 0. Let
be an increasing sequence of moderated positive definite functions in L2(G, Mn). Suppose that Φi mutu-
ally commute. If supi kΦikL2 < ∞, then Φ = lim Φi exists in P2(G, Mn).
The n = 1 case is proved as Prop. 16 in [1]. See also 13.8.5, 13.8.4 [2].
Φ1 (cid:22) Φ2 (cid:22) . . . (cid:22) Φn (cid:22) . . .
Proof of Theorem 3.1: Φ ∗ Ψ = Ψ ∗ Φ implies that λ(Φ)λ(Ψ) = λ(Ψ)λ(Φ) as bounded operators on
L2(G, Cn). Since λ(Φ) and λ(Ψ) are both positive, they must be self-adjoint. Hence λ(Φ)λ(Ψ) must be
positive and self-adjoint. In other words, λ(Φ ∗ Ψ) is positive on L2(G, Cn). In particular, it is positive
with respect to Cc(G, Cn). Hence Φ ∗ Ψ, as a matrix-valued continuous function, is positive definite.
Φ ∗ Ψ(e) must be a positive semi-definite matrix. By Lemma 2.1 hΦ, Ψi = T r(Φ ∗ Ψ)(e) ≥ 0.
Let Φ1 (cid:22) Φ2 (cid:22) . . . (cid:22) Φn (cid:22) . . . be an increasing sequence of moderated positive definite functions
in L2(G, Mn). For j ≥ i, notice that kΦjk2 = kΦik2 +hΦi, Φj − Φii +hΦj − Φi, Φii +kΦj − Φik2 ≥ kΦik2.
Since supi kΦikL2 < ∞, the sequence {kΦik} is an increasing sequence bounded from above. In partic-
ular, it is a Cauchy sequence. Notice that for j ≥ i
kΦj − Φik2 = kΦjk2 − kΦik2 − hΦi, Φj − Φii − hΦj − Φi, Φii ≤ kΦjk2 − kΦik2.
This implies that {Φi} is a Cauchy sequence in L2(G, Mn). Let Φ be the L2-limit of {Φi}. For every
u = (up) ∈ Cc(G, Cn), since λ(up) is a bounded operator on L2(G), we obtain
It follows that
([Φi]p,q ∗ uq, up) → ([Φ]p,q ∗ uq, up).
0 ≤ hλ(Φi)u, ui → hλ(Φ)u, ui.
ON MATRIX VALUED SQUARE INTEGRABLE POSITIVE DEFINITE FUNCTIONS
5
Therefore Φ is positive definite. (cid:3)
Theorem 3.2 (Thm. 17 [1]). Suppose that Φ is a moderated element in P2(G, Mn) such that Φ (cid:22) Θ with
Θ a continuous positive definite function. Then there is a unique moderated element Ψ in P2(G, Mn)
such that Φ = Ψ ∗ Ψ in L2.
In particular, Φ equals a continuous positive definite function almost
everywhere and kΨk2 ≤ T r(Θ(e)).
In particular, if Φ is continuous and moderated in P2(G, Mn) , its square root Ψ exists and is unique.
The n = 1 case is established by Godement. Our proof follows from the proof of Theorem 13.8.6
in [2] for the scalar-valued positive definite L2 functions. The original idea of Godement is to construct
an increasing sequence of positive definite moderated elements Ψk in L2(G) that approaches the square
root. In Dixmier's book, Ψk = kλ(Φ)kpk( Φ
kλ(Φ)k ). Here {pk(t)} is an increasing sequence of nonnegative
polynomials on [0, 1] such that pk(t) → √t on [0, 1] and pk(0) = 0. We shall supply a proof of this fact
before we carry out the proof of Theorem 3.2.
Lemma 3.1. There exists a sequence of polynomials
0 ≤ p1(t) ≤ p2(t) ≤ . . . ≤ pk(t) ≤ . . . ≤
√t
(t ∈ [0, 1]),
such that pk(t) → √t uniformly on [0, 1].
Proof: Consider the function t− 1
Clearly
2 , (t ∈ [0, 1]). Let qk(t) be the k-th Taylor polynomial at t = 1.
qk+1(t) = qk(t) +
2 )( 3
( 1
2 ) . . . ( 2k+1
(k + 1)!
2
)
(1 − t)k+1.
Let pk(t) = tqk(t). Clearly pk(t) is an increasing sequence of non-negative continuous functions with
limit √t over the interval [0, 1]. By Taylor's theorem, pk(t) → √t uniformly on [ǫ, 1]. On [0, ǫ],
√t − pk(t) < √t ≤ √ǫ. Hence pk(t) → √t uniformly on [0, 1]. (cid:3)
Proof of Theorem 3.2: Let Φ be a moderated element in P2(G, Mn) such that Φ (cid:22) Θ with Θ a
continuous positive definite function. Without loss of generality, suppose the operator norm kλ(Φ)k = 1.
For any polynomial p(t) = Pr
i=0 aiti, define
i
p(Φ) =
r
Xi=0
ai
z
{
Φ ∗ Φ ∗ . . . ∗ Φ .
}
Let Ψk = pk(Φ) with pk(t) defined in the last lemma. Essentially by functional calculus, we will have
λ(Ψk) (cid:22) λ(Ψk+1),
λ(Ψk)λ(Ψk) (cid:22) λ(Φ).
It follows that Ψk (cid:22) Ψk+1 and Ψk ∗ Ψk (cid:22) Φ (cid:22) Θ. By taking the value at e, we have T r(Ψk ∗ Ψk(e)) ≤
T r(Θ(e)). By Lemma 2.1, kΨkk is bounded by pT rΘ(e). Since {Ψk} mutually commutes and is an
increasing sequence, by Theorem 3.1, the L2-limit of Ψk exists. Put Ψ = limk→∞ Ψk. By Theorem 3.1,
Ψ ∈ P2(G, Mn). We have
Ψ ∗ Ψ(g) = lim
k→∞
Ψk ∗ Ψk(g),
pointwise. Since limk→∞ λ(Ψk ∗ Ψk − Φ) = 0 in the operator norm, for u ∈ Cc(G, Cn), we have
lim
k→∞
λ(Ψk ∗ Ψk)u = λ(Φ)u
6
HONGYU HE
in L2(G, Cn). However, the pointwise limit of the left hand side is obviously λ(Ψ ∗ Ψ)u. Hence
λ(Ψ ∗ Ψ − Φ)u = 0 for every u ∈ Cc(G, Cn). Hence Ψ ∗ Ψ(g) = Φ(g) almost everywhere.
In par-
ticular, Φ(g) is equal to a continuous positive definite function almost everywhere.
Now λ(Φ) = λ(Ψ)2 on Cc(G, Cn). For any u ∈ Cc(G, Cn), we have
kλ(Ψ)uk2 = hλ(Ψ)u, λ(Ψ)ui = hλ(Ψ)2u, ui = hλ(Φ)u, ui ≤ kλ(Φ)kkuk2.
Hence λ(Ψ) is bounded on Cc(G, Cn) and the function Ψ is moderated. By Lemma 2.3, λ(Φ) = λ(Ψ)2,
as bounded self-adjoint operators on L2(G, Cn). Since λ(Ψ) is positive, λ(Ψ) is unique as a bounded
operator on L2(G, Cn). In particular, λ(Ψ) is uniquely defined on Cc(G, Cn). Then Ψ must be unique.
(cid:3)
4. Square Roots: Proof of Theorem A
Let G be a unimodular locally compact group. Let Φ ∈ P 2(G, Mn). Now we would like to give a proof
of Theorem A. Our proof is somewhat different from the proof of Theorem 13.8.6 given in [2]. The basic
idea is the same, namely, to construct a sequence of moderated continuous positive definite functions
Φk → Φ. Let Ψk be the square root of Φk. Then the square root of Φ can be obtained as the L2-limit
of Ψk. The construction is canonical. In our proof, the continuity of Φk is given by Theorem 3.2. We
do not use Cor. 13.7.11 in [2] which requires several more pages of argument. We also wish to point out
a major difference. In the scalar case λ(Φk) acts on L2(G) and in our case λ(Φk) acts on L2(G, Cn) not
on L2(G, Mn).
Proof of Theorem [A]: Let x ∈ G. Let ρ(x) act on L2(G, Cn) by (ρ(x)u)(g) = u(gx). The action ρ
is simply the right regular action. Hence ρ(x) is a unitary operator on L2(G, Cn). If Φ ∈ Lloc(G, Mn),
then obviously
(3)
on Cc(G, Cn).
ρ(x)λ(Φ)ρ(x−1) = λ(Φ)
Let Φ ∈ P 2(G, Mn). Then λ(Φ)Cc(G,Cn) is a positive symmetric operator densely define on L2(G, Cn),
by the definition of positive definiteness of Φ. Let Λ(Φ) be the Friedrichs extension of λ(Φ)Cc(G,Cn).
Then Λ(Φ) is an (unbounded) positive and self-adjoint operator (Ch. 5.6. [7]). By Equation (3), we
must have ρ(x)Λ(Φ)ρ(x−1) = Λ(Φ).
0
0
Let Λ(Φ) = R ∞
Then ρ(x)Λ(Φ)ρ(x−1) = R ∞
tdP be the spectral decomposition. Here P is a projection-valued measure on the Borel
subsets of R. In other words, for every B a Borel subset of R, there is a projection P (B) on L2(G, Cn).
td[ρ(x)P ρ(x−1)]. Notice here that ρ(x) is unitary. Hence ρ(x)P (B)ρ(x−1)
remains a projection. The uniqueness of the spectral decomposition of self-adjoint operators implies that
ρ(x)P (B)ρ(x−1) = P (B). Since P (B) is bounded, we have ρ(x)P (B) = P (B)ρ(x) for any Borel subset
B and for any x ∈ G.
Let [Φ]∗j be the j-th column vector of Φ. Fix a Borel subset B. Define ΦB by letting the j-th col-
umn vector to be [ΦB]∗j = P (B)[Φ]∗j. Clearly ΦB ∈ L2(G, Mn).
Claim 1: λ(ΦB) = P (B)λ(Φ) on Cc(G, Cn).
ON MATRIX VALUED SQUARE INTEGRABLE POSITIVE DEFINITE FUNCTIONS
7
Proof: Let u ∈ Cc(G, Cn). Then
Zx∈G
[λ(Φ)u](g) = Xj
[Φ]∗j(gx−1)[u]j(x)dx = Xj
Zx∈G
(ρ(x−1)[Φ]∗j )(g)[u]j(x)dx.
For any x ∈ G, we have
(P (B)ρ(x−1)[Φ]∗j)(g) = (ρ(x−1)P (B)[Φ]∗j)(g) = (P (B)[Φ]∗j )(gx−1).
Since P (B) is a bounded operator on L2(G, Cn), [Φ]∗j ∈ L2(G, Cn) and [u]j(x) ∈ L1(G), we have
[P (B)(λ(Φ)u)](g) =P (B)Z Xj
(ρ(x−1)[Φ]∗j)(g)[u]j(x)dx
(4)
(P (B)ρ(x−1)[Φ]∗j)(g)[u]j(x)dx
(ρ(x−1)P (B)[Φ]∗j )(g)[u]j(x)dx
=Z Xj
=Z Xj
=Z Xj
=Z Xj
=Z ΦB(gx−1)u(x)dx
(P (B)[Φ]∗j )(gx−1)[u]j(x)dx
([ΦB]∗j)(gx−1)[u]j(x)dx
Our claim is proved.
=(λ(ΦB)u)(g)
Observe that P (B)λ(Φ) = P (B)Λ(Φ) on Cc(G, Cn) and P (B)Λ(Φ) is positive and bounded. There-
fore λ(ΦB) = P (B)λ(Φ) is bounded on Cc(G, Cn) and positive with respect to Cc(G, Cn). Hence ΦB
is moderated and positive definite. We must have λ(ΦB) = P (B)Λ(Φ) on L2(G, Cn).
In addition if
B1 ⊃ B2
on Cc(G, Cn) and the right hand side is positive and self adjoint. Hence ΦB1 (cid:23) ΦB2 . Similarly ΦB1 (cid:22) Φ.
For each positive integer k, define Φk = Φ[0,k]. We then obtain an increasing sequence of moderated
positive definite functions
λ(ΦB1 − ΦB2 ) = (P (B1) − P (B2))Λ(Φ)
Due to the way [Φk]∗j are defined, Φk → Φ in L2-norm. We have
Lemma 4.1. Let G be a unimodular group. Every Φ ∈ P2(G, Mn) is a L2-limit of an increasing sequence
of mutually commutative moderated elements in P2(G, Mn).
Φ1 (cid:22) Φ2 (cid:22) . . . (cid:22) Φk (cid:22) . . . ((cid:22) Φ).
The n = 1 case was proved by Godement as Prop. 14 in [1].
Since Φk is moderated in P2(G, Mn) with Φk (cid:22) Φ, by Theorem 3.2, there is a moderated element
Ψk ∈ P2(G, Mn) such that Φk = Ψk ∗ Ψk almost everywhere. Without loss of generality, suppose that
8
HONGYU HE
Φk = Ψk ∗ Ψk pointwise. Notice that both λ(Φk) and λ(Ψk) can be regarded as positive bounded self-
adjoint operators on the Hilbert space L2(G, Cn). By Lemma 2.3, as bounded self-adjoint operators on
L2(G, Cn), λ(Φk) = λ(Ψk)2. We have
(5)
λ(Ψk) = Z k
0
√tdP (t).
In particular, λ(Ψk) is uniquely defined on Cc(G, Cn). Therefore Ψk is unique and satisfies Equation 5.
By functional calculus, {λ(Ψk)} mutually commute and yield an increasing sequence of positive bounded
self-adjoint operators on L2(G, Cn). Restricted to Cc(G, Cn), it is easy to see that {Ψk} must mutually
commute and
Ψ1 (cid:22) Ψ2 (cid:22) . . . (cid:22) Ψk (cid:22) . . . .
Observe that kΨkk2 = T r(Ψk ∗ Ψk(e)) ≤ T r(Φ(e)). By Theorem 3.1, {Ψk} converges in L2(G, Mn).
Let Ψk → Ψ in L2(G, Mn). By Theorem 3.1, Ψ ∈ P2(G, Mn). Notice that Ψk ∈ L2(G, Mn). Then
Φk = Ψk ∗ Ψk converges uniformly to Ψ ∗ Ψ. Since ΦkK → ΦK in L2(K, Mn) for any compact set K,
ΦK = Ψ ∗ ΨK almost everywhere. Therefore Φ = Ψ ∗ Ψ almost everywhere. Since Φ is continuous,
Φ = Ψ ∗ Ψ. Theorem A is proved. (cid:3)
5. Nonnegative Integral: Proof of Theorem B
Let G be a unimodular group. Let Φ, Γ ∈ P2(G, Mn). We want to prove that
hΦ, Γi ≥ 0.
The main idea of the proof here is essentially due to Godement ( Prop.18 [1]). We start with the follow-
ing lemma.
Lemma 5.1. Let G be a unimodular group. Every Φ in P2(G, Mn) is a limit of an increasing sequence
of moderated elements in P 2(G, Mn) under the L2 norm.
Proof: By Lemma 4.1, it suffices to show that every moderated element Φ in P2(G, Mn) is the L2
limit of an increasing sequence of moderated elements in P 2(G, Mn). Without loss of generality, suppose
that kλ(Φ)k = 1. Let rk(t) be the k−th Taylor polynomial of 1
t at t = 1. We define qk(t) = t2rk(t).
Then qk(t) is an increasing sequence of nonnegative polynomial functions on [0, 1] such that qk(t) → t
uniformly on [0, 1] (c.f. Lemma 3.1).
Let Φk = qk(Φ). Then λ(Φk) = qk(λ(Φ)) is an increasing sequence of positive self-adjoint operators
that approaches λ(Φ). Obviously, Φk(g) is positive definite. Since λ(Φ) extends to a bounded operator
on L2(G, Cn), λ(Φk) also extends to a bounded operator on L2(G, Cn). Hence Φk is moderated. Since
Φ ∗ Φ is continuous, Φk = qk(Φ) is always continuous. Therefore, {Φk} is an increasing sequence of
continuous moderated positive definite functions.
Notice that λ(Φk), λ(Φ) all mutually commute. Since λ(Φk) (cid:22) λ(Φ), (λ(Φk))k (cid:22) (λ(Φ))k. Hence
Φk ∗ Φk (cid:22) Φ ∗ Φ. This implies T r(Φk ∗ Φk(e)) ≤ T r(Φ ∗ Φ(e)). By a similar argument in the proof
of Theorem 3.1, kΦkk ≤ kΦk. By Theorem 3.1, let Ψ be the L2-limit of Φk. For any u ∈ Cc(G, Mn),
λ(Ψ)u = limk→∞ λ(Φk)u pointwise, and limk→∞ λ(Φk)u = λ(Φ)u in L2-norm. It follows that Ψ = Φ
almost everywhere. Therefore Φk → Φ in L2-norm.
ON MATRIX VALUED SQUARE INTEGRABLE POSITIVE DEFINITE FUNCTIONS
9
We have obtained an increasing sequence of moderated elements in P 2(G, Mn) such that Φk → Φ
in L2-norm. (cid:3)
Lemma 5.2. Let Φ1 be a moderated element in P2(G, Mn) and Φ2 ∈ P 2(G, Mn). We have
hΦ1, Φ2i ≥ 0.
Proof: Suppose that Φ2 = Ψ ∗ Ψ with Ψ ∈ P2(G, Mn). Then
hΦ1, Φ2i = T r(Φ1 ∗ Φ2(e)) = T r(Φ1 ∗ Ψ ∗ Ψ(e)) = hλ(Φ1)Ψ, Ψi =
n
Xi=1
hλ(Φ1)[Ψ]∗i, [Ψ]∗ii.
Notice that λ(Φ1) is a bounded positive self adjoint operator. We have hΦ1, Φ2i ≥ 0. (cid:3)
Proof of Theorem B: For Φ, Γ ∈ P2(G, Mn), let Φα be a sequence of moderated element in P2(G, Mn)
with L2-limit Φ and Γβ be a sequence of elements in P 2(G, Mn) with L2-limit Γ. Then we have
Theorem B is proved. (cid:3)
hΦ, Γi = lim
α,β→∞hΦα, Γβi ≥ 0.
6. Zero Integral
Let Φ, Ψ ∈ P2(G, Mn). If Φ ∗ Ψ = 0, we have hΦ, Ψi = T r(Φ ∗ Ψ(e)) = 0. Now we would like to show
that the converse is also true.
Theorem 6.1. Let G be a unimodular locally compact group. Let Φ, Ψ ∈ P2(G, Mn). If hΦ, Ψi = 0,
then Φ ∗ Ψ = 0.
Proof: By Lemma 4.1, let Φm be an increasing sequence of moderated elements in P2(G, Mn) such that
m . By Lemma 5.1, let Ψp be an increasing sequence in P 2(G, Mn) such that kΨp− Ψk ≤ 1
p .
kΦm− Φk ≤ 1
Then
0 = hΦ, Ψi = hΦ − Φm, Ψi + hΦm, Ψi ≥ hΦm, Ψi ≥ hΦm, Ψpi ≥ 0.
Hence all the inequalities here must be equalities. Suppose that Ψp = Θp ∗ Θp with Θp ∈ P2(G, Mn).
Then
0 = hΦm, Ψpi = T r(Φm ∗ Θp ∗ Θp(e)) = Xi
hλ(Φm)[Θp]∗i, [Θp]∗ii.
Since λ(Φm) is a positive operator on L2(G, Cn), hλ(Φm)[Θp]∗i, [Θp]∗ii = 0. Thus λ(Φm)[Θp]∗i = 0
in L2(G, Cn). It follows that Φm ∗ Θp = 0 in L2(G, Mn). Since Φm ∗ Θp(g) is a continuous function,
Φm ∗ Θp(g) = 0 for all g ∈ G. Hence Φm ∗ Ψp(g) = Φm ∗ Θp ∗ Θp(g) = 0. Since Φm → Φ and Ψp → Ψ in
L2(G, Mn), we have Φm ∗ Ψp(g) → Φ ∗ Ψ(g). Therefore Φ ∗ Ψ(g) = 0 for all g. (cid:3)
Corollary 6.1. Let G be a locally compact unimodular group. Let Φ, Ψ ∈ P2(G, Mn). Then hΦ, Ψi ≥ 0
and hΦ, Ψi = 0 if and only if Φ ∗ Ψ = 0.
10
HONGYU HE
7. Applications in Representation Theory
Let G be a unimodular group. We call a unitary representation (π,H) of G Lp if there is a cyclic
vector u in H such that (π(g)u, u) is Lp. A Lp unitary representation has a G-invariant dense subspace
with Lp-matrix coefficients.
Theorem 7.1. Let G be a unimodular locally compact group and (π,H) be a unitary representation of
G. Suppose that (π1,H1) and (π2,H2) are two L2-unitary representations of G such that
(π,H) ∼= (π1 ⊗ π2,H1 ⊗H2).
Let u = Pn
i=1 u(i)
2
1 ⊗ u(i)
ZG
such that matrix coefficients with respect to {u(i)
2 , u(j)
1 )(π2(g)u(i)
(π1(g)u(i)
1 , u(j)
(π(g)u, u)dg =
1 } and {u(i)
2 )dg ≥ 0.
n
Xi,j=1
ZG
2 } are all L2. Then
Proof: Observe that Φ1 defined by [Φ1]ij = (π1(g)u(i)
Similarly, Φ2 ∈ L2(G, Mn) defined by [Φ2]ij = (u(i)
This theorem follows easily from Theorem B. (cid:3)
1 , u(j)
2 , π2(g)u(j)
1 ) is square integrable and positive definite.
2 ) is square integrable and positive definite.
Now we shall apply our result to Howe's correspondence ([5]). Let (G(m), G′(n)) be a dual reduc-
tive pair in Sp. Let (G′(n1), G′(n2)) be two G′-groups diagonally embedded in G′(n) with n1 + n2 = n.
Then (G(m), G′(ni)) is a dual reductive pair in some Sp(i) such that (Sp(1), Sp(2)) are diagonally em-
]
Sp(i). Let ω be the oscillator representation of
bedded in Sp. Let ωi be the oscillator representation of
Sp. Then ω can be identified with ω1⊗ω2. This identification preserves that actions of G(m) and G′(ni).
Now suppose that the matrix coefficients of ω1 G(m) with respect to the Schwartz space are L2. Let π be
an irreducible unitary representation of G(m). Suppose that the matrix coefficients for ω∞
2 G(m) ⊗ π∞
are all square integrable. Then for any v ∈ π∞, u(j)
1 ⊗ u(j)
1 , u(k)
1 , u(j)
1 ∈ ω∞
2 ), (X u(k)
1 )(ω2(g)u(j)
2 with j ∈ [1, N ], we have
2 ))(π(g)v, v)dg
2 ∈ ω∞
1 ⊗ u(k)
2 , u(k)
2 )(π(g)v, v)dg.
(ω(g)(X u(j)
Z G(m)
(ω1(g)u(j)
Z G(m)
=Xj,k
(6)
By Theorem 7.1, this integral must be nonnegative.
Corollary 7.1. Consider a dual reductive pair (G(m), G′(n)) in Sp. Let n = n1+n2. Let (G(m), G′(ni))
be a dual reductive pair in Sp(i). Let ωi be the oscillator representation of Sp(i). Let π be an irre-
ducible unitary representation of G(m). Suppose that the matrix coefficients with respect to ω∞
1 G(m)
and ω∞
2 G(m) ⊗ π∞ are square integrable. Let ξ ∈ ω∞
2 and u ∈ π∞, then
1 ⊗ ω∞
Z G(m)
(ω(g)ξ, ξ)(π(g)u, u)dg ≥ 0.
This Corollary holds for both p-adic groups and real groups. See [6] [3] [4] for the importance of this
integral in Howe's correspondence ([5]). In particular, under the hypothesis of the Corollary, Howe's
correspondence preserves unitarity.
ON MATRIX VALUED SQUARE INTEGRABLE POSITIVE DEFINITE FUNCTIONS
11
References
[1] R. Godement, "Les fonctions de type positif et la thorie des groupes, "Trans. Amer. Math. Soc. 63, (1948). 1-84
[2] J. Dixmier, C ∗-Algebra, North-Holland Publishing Company, 1969.
[3] H. He, "Theta Correspondence I-Semistable Range: Construction and Irreducibility ", Communications in Contem-
porary Mathematics (Vol 2), 2000, (255-283).
[4] H. He, "Unitary representations and theta correspondence for type I classical groups,"J. Funct. Anal. 199 (2003), no.
1, 92-121.
[5] R. Howe, "Transcending Classical Invariant Theory ", Journal of Amer. Math. Soc.(Vol. 2), 1989 (535-552).
[6] J-S. Li, "Singular Unitary Representation of Classical Groups"Inventiones Mathematicae (V. 97),1989 (237-255).
[7] R. Kadison and Ringrose, Fundamentals of the Theory of Operator Algebras, Academic Press, 1983.
Department of Mathematics, Louisiana State University, Baton Rouge, LA 70803
E-mail address: [email protected]
|
1304.2432 | 2 | 1304 | 2013-04-10T02:54:37 | A note on two-sided ideals in locally C*-algebras | [
"math.OA",
"math.FA"
] | In the present note we show that if A is a locally C*-algebra, and I and J are closed two-sided ideals in A, then the positive part of (I+J) is equal to the sum of positive parts of I and J. | math.OA | math |
A NOTE ON TWO-SIDED IDEALS IN LOCALLY C ∗-ALGEBRAS
ALEXANDER A. KATZ
Abstract. In the present note we show that if A is a locally C ∗-algebra, and
I and J are closed two-sided ideals in A, then (I + J)+ = I + + J +.
1. Introduction
Let A be a C ∗-algebra, and I and J be closed ∗-ideals in A. Let I + (resp. J +)
In 1964 in the first French
denotes the set of positive elements in I (resp. J).
edition of [3] Dixmier has formulated a problem whether or not
(I + J)+ = I + + J +?
Using the results of Effros [4] and Kadison [7],[8], Størmer was able in 1967 to settle
this problem in affirmative in his paper [14]. In 1968 Pedersen has noticed that it
was Combes who obtained a different proof of the aforementioned result of Størmer
[14] as a corrolary of Pedersen's Decomposition Theorem for C ∗-algebra (see [11]
for details). In 1971 Bunce has given in [2] yet another a very short and elegant
proof of the same result of Størmer from [14].
The Hausdorff projective limits of projective families of Banach algebras as nat-
ural locally-convex generalizations of Banach algebras have been studied sporadi-
cally by many authors since 1952, when they were first introduced by Arens [1] and
Michael [10]. The Hausdorff projective limits of projective families of C ∗-algebras
were first mentioned by Arens [1]. They have since been studied under various
names by many authors. Development of the subject is reflected in the monograph
of Fragoulopoulou [5]. We will follow Inoue [6] in the usage of the name locally
C ∗-algebras for these algebras.
The purpose of the present notes is to extend the aforementioned result of
Størmer from [14] to locally C ∗-algebras.
2. Preliminaries
First, we recall some basic notions on topological ∗-algebras. A ∗-algebra (or
involutive algebra) is an algebra A over C with an involution
such that
and
∗ : A → A,
(a + λb)∗ = a∗ + λb∗,
(ab)∗ = b∗a∗,
Date: April 7, 2013.
Key words and phrases. C ∗-algebras, locally C ∗-algebras, projective limit of projective family
of C ∗-algebras.
2010 AMS Subject Classification: Primary 46K05, Secondary 46K10.
1
2
ALEXANDER A. KATZ
for every a, b ∈ A and λ ∈ C.
A seminorm k.k on a ∗-algebra A is a C ∗-seminorm if it is submultiplicative, i.e.
kabk ≤ kak kbk ,
and satisfies the C ∗-condition, i.e.
ka∗ak = kak2 ,
for every a, b ∈ A. Note that the C ∗-condition alone implies that k.k is submulti-
plicative, and in particular
ka∗k = kak ,
for every a ∈ A (cf. for example [5]).
When a seminorm k.k on a ∗-algebra A is a C ∗-norm, and A is complete in in
the topology generated by this norm, A is called a C ∗-algebra.
A topological ∗-algebra is a ∗-algebra A equipped with a topology making the op-
erations (addition, multiplication, additive inverse, involution) jointly continuous.
For a topological ∗-algebra A, one puts N (A) for the set of continuous C ∗-seminorms
on A. One can see that N (A) is a directed set with respect to pointwise ordering,
because
max{k.kα , k.kβ} ∈ N (A)
for every k.kα , k.kβ ∈ N (A), where α, β ∈ Λ, with Λ being a certain directed set.
For a topological ∗-algebra A, and k.kα ∈ N (A), α ∈ Λ,
ker k.kα = {a ∈ A : kakα = 0}
is a ∗-ideal in A, and k.kα induces a C ∗-norm (we as well denote it by k.kα) on
the quotient Aα = A/ ker k.kα, and Aα is automatically complete in the topology
generated by the norm k.kα , thus is a C ∗-algebra (see [5] for details). Each pair
k.kα , k.kβ ∈ N (A), such that
β (cid:23) α,
α, β ∈ Λ, induces a natural (continuous) surjective ∗-homomorphism
Let, again, Λ be a set of indices, directed by a relation (reflexive, transitive,
gβ
α : Aβ → Aα.
antisymmetric) " (cid:22) ". Let
be a family of C ∗-algebras, and gβ
α be, for
{Aα, α ∈ Λ}
the continuous linear ∗-mappings
so that
for all α ∈ Λ, and
whenever
α (cid:22) β,
gβ
α : Aβ −→ Aα,
gα
α(xα) = xα,
α ◦ gγ
gβ
β = gγ
α,
α (cid:22) β (cid:22) γ.
TWO-SIDED IDEALS IN LOCALLY C ∗-ALGEBRAS
3
Let Γ be the collections {gβ
of the direct product algebra
α} of all such transformations. Let A be a ∗-subalgebra
so that for its elements
for all
where
and
Aα,
Y
α∈Λ
xα = gβ
α(xβ ),
α (cid:22) β,
xα ∈ Aα,
xβ ∈ Aβ.
Definition 1. The ∗-algebra A constructed above is called a Hausdorff projective
limit of the projective family
relatively to the collection
and is denoted by
{Aα, α ∈ Λ},
Γ = {gβ
α : α, β ∈ Λ : α (cid:22) β},
α ∈ Λ, and is called the Arens-Michael decomposition of A.
lim←−Aα,
It is well known (see, for example [15]) that for each x ∈ A, and each pair
α, β ∈ Λ, such that α (cid:22) β, there is a natural projection
πβ : A −→ Aβ,
defined by
α(πβ(x)),
and each projection πα for all α ∈ Λ is continuous.
πα(x) = gβ
Definition 2. A topological ∗-algebra (A, τ ) over C is called a locally C ∗-algebra
if there exists a projective family of C ∗-algebras
{Aα; gβ
α; α, β ∈ Λ},
so that
A ∼= lim←−Aα,
α ∈ Λ, i.e. A is topologically ∗-isomorphic to a projective limit of a projective family
of C ∗-algebras, i.e.
there exits its Arens-Michael decomposition of A composed
entirely of C ∗-algebras.
A topological ∗-algebra (A, τ ) over C is a locally C ∗-algebra iff A is a complete
Hausdorff topological ∗-algebra in which the topology τ is generated by a saturated
separating family of C ∗-seminorms (see [5] for details).
Example 1. Every C ∗-algebra is a locally C ∗-algebra.
Example 2. A closed ∗-subalgebra of a locally C ∗-algebra is a locally C ∗-algebra.
Example 3. The product Y
Aα of C ∗-algebras Aα, with the product topology, is
a locally C ∗-algebra.
α∈Λ
4
ALEXANDER A. KATZ
Example 4. Let X be a compactly generated Hausdorff space (this means that a
subset Y ⊂ X is closed iff Y ∩ K is closed for every compact subset K ⊂ X).
Then the algebra C(X) of all continuous, not necessarily bounded complex-valued
functions on X, with the topology of uniform convergence on compact subsets, is a
locally C ∗-algebra. It is well known that all metrizable spaces and all locally compact
Hausdorff spaces are compactly generated (see [9] for details).
Let A be a locally C ∗-algebra. Then an element a ∈ A is called bounded, if
kak∞ = {sup kakα , α ∈ Λ : k.kα ∈ N (A)} < ∞.
The set of all bounded elements of A is denoted by b(A).
It is well-known that for each locally C ∗-algebra A, its set b(A) of bounded
elements of A is a locally C ∗-subalgebra, which is a C ∗-algebra in the norm k.k∞ ,
such that it is dense in A in its topology (see for example [5]).
2.1. The cone of positive elements in a locally C*-algebra. If (A, τ ) is a
unital topological ∗-algebra, then the spectrum spA(a) of an element a ∈ A is the
set
spA(a) = {z ∈ C : zeA − a /∈ GA},
where eA is a unital element in A, and GA is the group of invertable elements in
A.
An element a ∈ A in a unital topological ∗-algebra (A, τ ) is called positive, and
we write
if a ∈ Asa, and
a ≥ 0A,
spA(a) ⊆ [0, ∞).
We denote the set of positive elements is (A, τ ) by A+.
Let A,τ be a locally C ∗-algebra, and
A ∼= lim←−Aα,
α ∈ Λ, be its Arens-Michael decomposition. Then
a ≥ 0A ←→ aα ≥ 0Aα, for all α ∈ Λ,
where aα = πα(a).
A positive part A+ of any locally C ∗-algebra is not empty, and, together with
each a ∈ Asa, it contains positive elements a+, a− and a , such that
a = a+ − a−,
a+a− = 0A = a−a+,
a = a+ + a−.
For any positive a ∈ A+ in any locally C ∗-algebra A there exists a unique
positive squire root b ∈ A+, such that
(see [5] for details).
The following theorem is valid:
a = b2.
Theorem 1 (Inoue-Schmudgen). Let (A, τ ) be a locally C ∗-algebra. The following
are equivalent:
(1) a ≥ 0A;
TWO-SIDED IDEALS IN LOCALLY C ∗-ALGEBRAS
(2) a = b∗b, for some b ∈ A;
(3) a = c2, for some c ∈ Asa.
Proof. See [6] and [13] for details.
5
(cid:3)
As a corollary from this theorem one can see that A+ = {a∗a : a ∈ A} in each
locally C ∗-algebra A.
The following theorem is valid:
Theorem 2 (Inoue-Schmudgen). Let (A, τ ) be a locally C ∗-algebra. Then A+ is a
closed convex cone such that A+ ∩ (−A+) = {0A}.
Proof. See [6] and [13] for details.
3. Positive parts of two-sided ideals in locally C*-algebras
We start by recalling the following well-known result:
Proposition 1. Let (A, τ A) and (B, τ B) be two locally C ∗-algebras, and
be a ∗-homomorphism. Then
ϕ : A → B,
ϕ(A+) = B+ ∩ ϕ(A).
Proof. See for example [5] for details.
As a corrolary from that theorem one gets:
(cid:3)
(cid:3)
Corollary 1. Let (B, τ B) be a locally C ∗-algebra, and (A, τ A) be its closed *-
subalgebra (with τ A = τ BA). Then
A+ = B+ ∩ A.
Proof. See for example [5] for details.
(cid:3)
Now, let (A, τ A) be an unital locally C ∗-algebra, I and J be closed ∗-ideals of
(A, τ A), B be a C ∗-algebra, and
ϕ : A → B,
be a surjective continuous ∗-homomorphism from A onto B. In the following se-
ries of lemmata we analyse what ϕ(I), ϕ(J), ϕ(A+), ϕ(I +), ϕ(J +), ϕ(I + + J +) and
ϕ((I + J)+) are in B.
Lemma 1. Let A, B, I (J), ϕ be as above. Then ϕ(I) (resp. ϕ(J)) is a closed
∗-ideal in B.
Proof. Because ϕ is continuous surjection from A onto B, and I is a ∗-subalgebra
(as a ∗-ideal), it follows that ϕ(I) is a closed ∗-subalgebra of B. It remained to
show that if x ∈ B, then
Let a be a fixed arbitrary element
x · ϕ(I) ⊂ ϕ(I).
a ∈ ϕ−1(x) ⊂ A
(it means that ϕ(a) = x), and b ∈ I. Because I is an ∗-ideal in A, it follows that
ab = c ∈ I.
6
ALEXANDER A. KATZ
Therefore, from
it follows that
for all b ∈ I, Q.E.D.
ϕ(a) · ϕ(b) = ϕ(ab) = ϕ(c) ∈ ϕ(I)
x · ϕ(b) ∈ ϕ(I),
Lemma 2. Let A, B, A+, B+ and ϕ be as above. Then
ϕ(A+) = B+.
Proof. From Proposition 1 it follows that
ϕ(A+) = B+ ∩ ϕ(A) ⊂ B+.
Therefore, it is enough to show that
B+ ⊂ ϕ(A+).
(cid:3)
Let x be an arbitrary element in B+. Then (see for example [12]) there exists a
positive y ∈ B+, such that
Let b ∈ A be an arbitrary element in
x = y2.
ϕ−1(y) ⊂ A
(it means that ϕ(b) = y). From Inoue-Schmudgen Theorem 1 above it follows that
and
Q.E.D.
b2 ∈ A+,
ϕ(b2) = (ϕ(b))2 = y2 = x,
(cid:3)
Lemma 3. Let A, B, I (resp. J), I + (resp. J +) and ϕ be as above. Then
(resp. ϕ(J +) = (ϕ(J))+).
ϕ(I +) = (ϕ(I))+
Proof. Because I is a locally C ∗-algebra, and ϕ(I) is a C ∗-algebra, the statement
is immediately follows from Lemma 2 above.
(cid:3)
Lemma 4. Let A, B, I, J, I +, J + and ϕ be as above. Then
ϕ(I + + J +) = (ϕ(I))+ + ((ϕ(J))+
Proof. Because ϕ is a ∗-homomorphism, from Lemma 2 above it follows that
ϕ(I + + J +) = ϕ(I +) + ϕ(J +) = (ϕ(I))+ + ((ϕ(J))+,
Q.E.D.
(cid:3)
Lemma 5. Let A, B, I, J, I +, J + and ϕ be as above. Then
ϕ((I + J)+) = (ϕ(I) + ϕ(J))+
TWO-SIDED IDEALS IN LOCALLY C ∗-ALGEBRAS
7
Proof. Let c be an arbitrary element in (I + J)+. From Inoue-Schmudgen Theorem
1 above it follows that there exist a ∈ I and b ∈ J, such that
c = (a + b)∗(a + b).
x = ϕ(a), and y = ϕ(b).
ϕ(c) = ϕ((a + b)∗(a + b)) = ϕ((a + b)∗)ϕ(a + b) =
= (ϕ(a) + ϕ(b))∗(ϕ(a) + ϕ(b)) ∈ (ϕ(I) + ϕ(J))+,
Let us denote
Then
thus
Inversly, let t be an arbitrary element
ϕ((I + J)+) ⊂ (ϕ(I) + ϕ(J))+.
t ∈ (ϕ(I) + ϕ(J))+.
From the basic properties of the cone of positive elements in a C ∗-algera (see for
example [12]) if follows that there exist
such that
x ∈ ϕ(I), and y ∈ ϕ(J),
t = (x + y)∗(x + y).
Let a be an arbitrary element in
and b be an arbitrary element in
ϕ−1(x) ⊂ I,
ϕ−1(y) ⊂ J.
It means that
x = ϕ(a), and y = ϕ(b).
Then from Inoue-Schmudgen Theorem 1 above it follows that the element
(a + b)∗(a + b) ∈ (I + J)+,
and
ϕ((a + b)∗(a + b)) = ϕ((a + b)∗)ϕ(a + b) = (ϕ(a) + ϕ(b))∗(ϕ(a) + ϕ(b)) =
thus
Q.E.D.
= (x + y)∗(x + y) = t,
ϕ((I + J)+) ⊃ (ϕ(I) + ϕ(J))+,
(cid:3)
Now we are ready to present the main theorem of the current notes.
Theorem 3. Let (A, τ A) be an unital locally C ∗-algebra, and (I, τ I ) and (J, τ J ) be
two ∗-ideals in A, such that
Then
τ I = τ AI and τ J = τ AJ .
(I + J)+ = I + + J +.
8
ALEXANDER A. KATZ
Proof. Let now (A, τ ) be a locally C ∗-algebra, and let
A = lim←−Aα,
α ∈ Λ, be its Arens-Michael decomposition, built using the family of seminorms
k.kα , α ∈ Λ, that defines the topology τ . Let
πα : A → Aα,
α ∈ Λ, be a projection from A onto Aα, for each α ∈ Λ. Each πα is an injective
∗-homomorphism from A onto Aα, α ∈ Λ, thus, we can apply to A, Aα, and πα
lemmata 3-5 above for all α ∈ Λ. Let
Iα = πα(I)
(resp. Jα = πα(J)). Therefore, for all α ∈ Λ, Størmer's result from [14] for C ∗-
algebras implies that
(Iα + Jα)+ = I +
α + J +
α .
Thus, because
and
α ∈ Λ, we get
Q.E.D.
(I + J)+ = lim
(I + J)+
α ,
←−
I + = lim
I +
α ,
←−
J + = lim
J +
α ,
←−
(I + J)+ = I + + J +,
References
(cid:3)
[1] Arens, R., A generalization of normed rings. (English) Pacific J. Math., Vol. 2 (1952), pp.
455 -- 471.
[2] Bunce, J., A note on two-sided ideals in C ∗-algebras. (English) Proc. Amer. Math. Soc.
Vol. 28 (1971), pp. 635.
[3] Dixmier, J., C ∗-algebras. (Engliah) Translated from the French by Francis Jellett. North-
Holland Mathematical Library, Vol. 15. North-Holland Publishing Co., Amsterdam-New
York-Oxford (1977), 492 pp.
[4] Effros, E.G., Order ideals in a C ∗-algebra and its dual. (English) Duke Math. J. Vol. 30
(1963), pp. 391 -- 411.
[5] Fragoulopoulou, M., Topological algebras with involution. (English) North-Holland Math-
ematics Studies, Vol. 200, Elsevier Science B.V., Amsterdam (2005), 495 pp.
[6] Inoue, A., Locally C ∗-algebra. (English), Mem. Fac. Sci. Kyushu Univ. Ser. A , Vol. 25
(1971), pp. 197 -- 235.
[7] Kadison, R.V., A representation theory for commutative topological algebra. (English)
Mem. Amer. Math. Soc. (1951), No. 7, 39 pp.
[8] Kadison, R.V., Transformations of states in operator theory and dynamics. (English)
Topology Vol. 3 (1965) suppl. No. 2, pp. 177 -- 198.
[9] Kelley, J.L., General topology. (English) Reprint of the 1955 edition [Van Nostrand,
Toronto, Ont.]. Graduate Texts in Mathematics, No. 27. Springer-Verlag, New York-Berlin
(1975), 298 pp
[10] Michael, E.A., Locally multiplicatively-convex topological algebras. (English) Mem. Amer.
Math. Soc., No. 11 (1952), 79 pp.
[11] Pedersen, G.K., A decomposition theorem for C ∗ algebras. (English) Math. Scand. Vol.
22 (1968), pp. 266 -- 268
[12] Pedersen, G.K., C ∗-algebras and their automorphism groups. (English) London Math-
ematical Society Monographs, Vol. 14. Academic Press, Inc. [Harcourt Brace Jovanovich,
Publishers], London-New York (1979), 416 pp.
TWO-SIDED IDEALS IN LOCALLY C ∗-ALGEBRAS
9
[13] Schmudgen, K., Uber LM C ∗-Algebren. (German) Math. Nachr. Vol. 68 (1975), pp. 167 --
182.
[14] Størmer, E., Two-sided ideals in C ∗-algebras. (English) Bull. Amer. Math. Soc. Vol. 73
(1967), pp. 254 -- 257.
[15] Tr`eves, F., Topological vector spaces: Distributions and Kernels. (English), New York-
London: Academic Press. (1967), 565 pp.
Dr. Alexander A. Katz, Department of Mathematics and Computer Science, St.
John's College of Liberal Arts and Sciences, St. John's University, 8000 Utopia Park-
way, St. John's Hall 334-i, NY 11439, USA
E-mail address: [email protected]
|
1605.06395 | 4 | 1605 | 2017-11-30T18:31:35 | C*-simplicity of free products with amalgamation and radical classes of groups | [
"math.OA",
"math.GR"
] | We give new characterizations to ensure that a free product of groups with amalgamation has a simple reduced group C*-algebra, and provide a concrete example of an amalgam with trivial kernel, such that its reduced group C*-algebra has a unique tracial state, but is not simple. Moreover, we show that there is a radical class of groups for which the reduced group C*-algebra of any group is simple precisely when the group has a trivial radical corresponding to this class. | math.OA | math |
C∗-SIMPLICITY OF FREE PRODUCTS WITH AMALGAMATION AND
RADICAL CLASSES OF GROUPS
NIKOLAY A. IVANOV AND TRON OMLAND
Abstract. We give new characterizations to ensure that a free product of groups with
amalgamation has a simple reduced group C∗-algebra, and provide a concrete example
of an amalgam with trivial kernel, such that its reduced group C∗-algebra has a unique
tracial state, but is not simple.
Moreover, we show that there is a radical class of groups for which the reduced group C∗-
algebra of any group is simple precisely when the group has a trivial radical corresponding
to this class.
1. Introduction
Recall that a discrete group G is called C∗-simple if its reduced group C∗-algebra C∗
Groups have been among the most studied objects in connection with operator algebras,
and one of the natural questions to consider in this regard is whether group C∗-algebras are
simple. While full group C∗-algebras can never be simple (unless the group is trivial), the
problem for reduced group C∗-algebras has been of great interest ever since the work of Powers
[28], showing that nonabelian free groups have a simple reduced group C∗-algebra. Larger
classes of groups with properties similar to the one Powers described were later studied, and
several results and open problems on this topic are discussed by de la Harpe [18].
r (G)
is simple, and G is said to have the unique trace property if C∗
r (G) has a unique tracial
state. A long-standing open problem was whether the two properties were equivalent, but
fairly recently it has been shown that C∗-simplicity is strictly stronger than the unique trace
property. The "stronger" part is due to Breuillard, Kalantar, Kennedy, and Ozawa [7, 22]
and the "strictly" part is due to Le Boudec [5]. The former showed that the unique trace
property is equivalent to having a trivial amenable radical, a property already known to be
weaker than C∗-simplicity, see [18]. Both of the works [7, 22] use extensively the theory of
boundary actions developed by Furstenberg and Hamana. Even more recently, some other
more operator-theoretical characterizations have been obtained by Haagerup [15] and Kennedy
[23]. However, the conditions from [15, 23] that are equivalent with C∗-simplicity, are not
always easy to check in concrete situations, for example by combinatorial group properties.
We also note that the unique trace property for G always implies that G is icc, that is, every
nontrivial conjugacy class in G is infinite, and it is well-known that the group von Neumann
algebra associated to G is a factor if and only if G is icc.
is C∗-simple was first considered by Bédos [1], and mentioned as Problem 27 in [18].
The problem of finding conditions to ensure that a free product of groups with amalgamation
In this article, we first make a detailed study of amalgamated free products, inspired by
work of de la Harpe and Préaux [20]. By making use of a few new observations, we are able
to improve some of the results in [20]. Then we present in Section 4 a concrete example of an
amalgam that has the unique trace property, but is not C∗-simple.
Date: November 30, 2017.
2010 Mathematics Subject Classification. 22D25, 20E06 (Primary) 46L05, 43A07 (Secondary).
Key words and phrases. C∗-simplicity, free product of groups with amalgamation, radical classes of groups.
1
2
NIKOLAY A. IVANOV AND TRON OMLAND
In particular, our example shows that a free product with amalgamation can fail to be
C∗-simple when it has a trivial kernel, but has "one-sided" kernels that are nontrivial (and
amenable). Every amalgam G0 ∗H G1 acts on its Bass-Serre tree, and in this setup the
one-sided kernels consist of the elements that fix all vertices of the "half-trees" obtained by
removing the edge corresponding to H. This has similarities with [5].
We keep the first sections mostly to "elementary" proofs, restricting to combinatorial group
properties, and delay the involvement of boundary actions to Section 5.
In the final two sections, we first recall the definitions of radical and residual classes of
groups and prove several statements that will be of later use. The result of [7] implies that
the class of groups with the unique trace property is the residual class "dual" to the radical
class of amenable groups. We then show that the class of C∗-simple groups is also a residual
class, giving rise to a "predual" radical class of groups that contains all the amenable groups,
and we will call a group "amenablish" if it belongs to this class. The example we provide
of an amalgam that is not C∗-simple, but has the unique trace property, turns out to be a
(nonamenable) amenablish group. Moreover, for every group there is an amenablish radical
such that C∗-simplicity of the group is equivalent to this radical being trivial, giving an analog
of the relation between the amenable radical and the unique trace property.
2. Preliminaries
We consider only discrete groups in this article.
Let G be a group. As usual, we equip the Hilbert space '2(G) with the standard orthonormal
basis {δg}g∈G, and define the left regular representation λ of G on '2(G) by λ(g)δh = δgh.
r (G), is the C∗-subalgebra of B('2(G))
Then the reduced group C∗-algebra of G, denoted by C∗
generated by λ(G). The group G is called C∗-simple if C∗
r (G) is simple, that is, if it has no
nontrivial proper two-sided closed ideals.
A state on a unital C∗-algebra A is a linear functional φ: A → C that is positive, i.e.,
φ(a) ≥ 0 whenever a ∈ A and a ≥ 0, and unital, i.e., φ(1) = 1. A state φ is called tracial if it
satisfies the additional property that φ(ab) = φ(ba) for all a, b ∈ A. There is a canonical faithful
r (G), namely the vector state associated with δe, that is, τ(a) = haδe, δei
tracial state τ on C∗
for all a ∈ C∗
r (G). The group G is said to have the unique trace property if τ is the only tracial
state on C∗
r (G).
Recall that a group G is amenable if there exists a state on '∞(G) which is invariant under
the left translation action by G. It is explained by Day [10] that every group G has a unique
maximal normal amenable subgroup, called the amenable radical of G. Then [7, Theorem 1.3]
shows that G has the unique trace property if and only if the amenable radical of G is trivial,
so C∗-simplicity is stronger than the unique trace property by [18, Proposition 3].
The first larger class of groups that were shown to be C∗-simple with the unique trace
property, was the so-called Powers groups introduced by de la Harpe [17] (see also [19]).
Definition 2.1. A group G is called a Powers group if for any finite subset F of G \ {e} and
any integer k ≥ 1 there exists a partition G = D t E and elements g1, g2, . . . , gk in G such
that
f D ∩ D = ∅ for all f ∈ F,
giE ∩ gjE = ∅ for all distinct 1 ≤ i, j ≤ k.
Moreover, a group G is called a weak Powers group if it satisfies the above definition for
all finite sets F that are contained in a nontrivial conjugacy class. Finally, G is called a
weak∗ Powers group if it satisfies the above definition for all nontrivial one-element sets F.
Clearly, every Powers group is a weak Powers group and every weak Powers group is a
weak∗ Powers group. It is known that the weak Powers property implies C∗-simplicity [3],
while the weak∗ Powers property implies the unique trace property [31, Section 5.2].
siE0 ∩ sjE0 = g−1
⊆ g−1
1 gif g−1
1 giE ∩ g−1
i g1E0 ∩ g−1
1 gjE = ∅.
1 gjf g−1
j g1E0 ⊆ g−1
1 gif g−1
i g1E ∩ g−1
1 gjf g−1
j g1E
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
3
Lemma 2.2. Assume that N is a normal subgroup of a group G. If G is a Powers group,
then N is a Powers group. If G is a weak Powers group, then N is a weak Powers group. If
G is a weak∗ Powers group, then N is a weak∗ Powers group.
Proof. The first statement is a (seemingly unnoticed) result by Kim [24, Theorem 1], and the
last two statements are identically proven. For the convenience of the reader, we present the
argument of the middle one:
Let k ≥ 1, and suppose that F is a finite set and C0 a nontrivial conjugacy class in N such
that F ⊆ C0. If two elements are conjugate in N they are also conjugate in G, so there is a
nontrivial conjugacy class C in G such that F ⊆ C. Since G is weak Powers, there exists a
partition G = Dt E and g1, g2, . . . , gk such that f D∩ D = ∅ for all f ∈ F and giE ∩ gjE = ∅
whenever i 6= j.
If k = 1, set D0 = {e}, E0 = N \ {e}, and s1 = e.
If k ≥ 2, set D0 = D ∩ N and E0 = E ∩ N, so N = D0 t E0 and f D0 ∩ D0 ⊆ f D ∩ D = ∅
1 giE are mutually disjoint for i ≥ 1,
Furthermore, fix an f ∈ F and note that the sets g−1
1 giE ⊆ G \ E = D. This again
i g1 for i ≥ 2. Since N is normal in G, we have si ∈ N for all
for all f ∈ F.
so that in particular, for i ≥ 2, then g−1
implies that f g−1
i, and if i 6= j, then
1 giE ∩ D ⊆ f D ∩ D = ∅, so f g−1
1 giE ∩ E = ∅, i.e., g−1
Set s1 = e and si = g−1
1 giE ⊆ G \ D = E.
1 gif g−1
Kim also proves that D0 and E0 are nonempty, but that does not seem to be necessary. (cid:3)
In [20] a group is said to be "strongly Powers" if every subnormal subgroup is a Powers
group. A consequence of the above lemma is that the notion of strongly Powers coincides with
Powers.
We conclude this section by mentioning two definitions. First, a group G has the free
semigroup property if for any finite subset F of G there exists g in G such that gF is semifree,
that is, the subsemigroup generated by gF in G is free over gF. Finally, a group G is said to
have stable rank one if its reduced group C∗-algebra has stable rank one, that is, the invertible
elements of C∗
r (G) are dense in C∗
r (G).
3. Free products of groups with amalgamation
Recall that (see e.g. [30]) a free product of groups G0 and G1 with amalgamation over a
common subgroup H (embedded via injective homomorphisms H → Gi for i = 0, 1) is a group
G together with homomorphisms φi : Gi → G for i = 0, 1, that agree on H, universal in the
i : Gi → G0 that agree on H, there
sense that for any other group G0 with homomorphisms φ0
is a unique homomorphism φ: G → G0 such that φ0
The amalgamated free products G = G0 ∗H G1 that we consider in this article are always
assumed to be nondegenerate in the sense that ([G0 : H] − 1) · ([G1 : H] − 1) ≥ 2, otherwise
the situation is very different.
g∈G gHg−1 denote the kernel of the amalgam. It coincides with the normal
core of H in G, i.e., the largest subgroup of H which is normal in G (and thus in G0 and G1).
If N is any subgroup of H which is normal in G, then the quotient G/N is isomorphic to
(G0/N) ∗H/N (G1/N). In particular, this holds when N = ker G. Note that we always have
ker(G/ ker G) = {e}. Indeed, if ker(G/ ker G) = N ⊆ H/ ker G, let N0 be the inverse image
of N under the quotient map G → G/ ker G. Then N0 is a normal subgroup of G containing
ker G and sitting inside H, so we must have N0 = ker G by maximality of ker G.
Let ker G =T
i = φ ◦ φi.
4
NIKOLAY A. IVANOV AND TRON OMLAND
If G is nondegenerate, then G/ ker G is nondegenerate, as [Gi : H] = [Gi/ ker G : H/ ker G].
Let F C(G) denote the normal subgroup of G consisting of elements with finite conjugacy
class in G. Clearly, if g ∈ Gi \ H, then the conjugacy class of g in G must be infinite. Hence,
F C(G) is contained in H.
Let now AR(G) denote the amenable radical of G and let N F(G) denote the largest normal
subgroup of G that does not contain any nonabelian free subgroup. By [9, Proposition 7],
these groups fit in general into a sequence of subgroups, namely
(1)
In particular, N F(G) (cid:40) G, so G always contains a nonabelian free group. Moreover, there is
an even longer sequence
F C(G) ⊆ AR(G) ⊆ N F(G) ⊆ ker G ⊆ H.
F C(G) ⊆ F C(ker G) ⊆ AR(ker G) = AR(G) ⊆ N F(G) = N F(ker G) ⊆ ker G.
The first containment is clear, the second holds since F C(ker G) is an amenable normal
subgroup of ker G, and the two equalities follow from Examples 6.4, 6.6, and Lemma 6.7.
If H is finite, then F C(G) = ker G, so the sequence collapses. Indeed, if h ∈ ker G, then
h ∈ gHg−1 for all g ∈ G, so g−1hg ∈ H for all g ∈ G. Hence, the conjugacy class of h is
contained in H, which is finite, so h ∈ F C(G). See Theorem 3.7 for more on this case.
Proposition 3.1. A nondegenerate amalgamated free product G has the unique trace property
if and only if ker G has the unique trace property.
Proof. Since AR(G) = AR(ker G) as explained above, this follows from [7, Theorem 1.3]. (cid:3)
Let us now recall the normal form for an element in G. First, for i = 0, 1, we choose sets Si
so that Si∪{e} form systems of representatives for the left cosets of Gi/H. Then every element
of G can be written uniquely as either s0s1 ··· sn−1snh or s1 ··· sn−1snh, where si ∈ Si (mod 2)
and h ∈ H. To avoid division into separate cases, the notation (s0)s1s2 ··· sn−1snh for the
normal form of an element of G is often used, especially in Section 5.
Since we will not always assume that elements are on normal form, the following observation
is sometimes useful: for i = 0, 1, if gi ∈ Gi \ H and h ∈ H, then there exists g0
i ∈ Gi \ H such
i. In other words, we can "cycle" a letter from H through a word (g0)g1 ··· gn
that gih = hg0
in G, with gi ∈ Gi (mod 2) \ H, without changing the length of the word.
Let G be any nondegenerate free product with amalgamation. We decompose G as follows.
For j = 0, 1 and k ≥ 1, let
Tj,k = {g0 ··· gk−1 : gi ∈ Gi+j (mod 2) \ H},
i.e., Tj,k consists of all words of length k starting with a letter from Gj \ H. To simplify
notation, we set T0,0 = T1,0 = H. For j = 0, 1 and every k ≥ 0, we now define the set
(2)
gHg−1,
and note that C0,0 = C1,0 = H and that H ∩ Cj,k is always a normal subgroup of H. Then
Cj,k = \
ker G = \
g∈Tj,k
K0 = \
K0 = \
k≥0
k≥0
Cj,k,
k≥0
j=0,1
and K1 = \
and K1 = \
k≥0
k≥0
which, as mentioned above, is always normal in G. Next, we set
(3)
C0,k
C1,k.
Both K0 and K1 are normal subgroups of H. Remark that it actually follows that
(4)
C1,2k.
C0,2k
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
5
Indeed, if h /∈ K0, then g−1hg /∈ H for some g ∈ T0,n with n ≥ 1. If n is odd, take any
g1 ∈ G1 \ H, and note that gg1 ∈ T0,n+1 and g−1
1 g−1, hence,
k≥0 C0,2k. A similar argument works for K1.
1 g−1hgg1 /∈ H, so h /∈ gg1Hg−1
h /∈T
Moreover, it should be clear that ker G = K0 ∩ K1 and that
K0 = H ∩ \
g0∈G0\H
g0K1g−1
0
and K1 = H ∩ \
g1K0g−1
1 .
g1∈G1\H
(5)
(6)
If K0 = ker G, then K1 ⊆ ker G since ker G is normal, so K0 = K1 (similarly if K1 = ker G).
Also, we have that K0 = {e} if and only if K1 = {e}.
Finally, using the notation from [20, (i) p. 2-3] we set
Ck = \
Cj,n.
0≤n≤k
j=0,1
Note that for any g ∈ T0,k+1, we have K0 ⊆ gCkg−1, and for any g ∈ T1,k+1, we have
K1 ⊆ gCkg−1. Thus, if Ck is trivial for some k ≥ 0, then K0 = K1 = {e}. Indeed, pick k ≥ 0,
g ∈ T0,k+1, h ∈ K0, and let s be an arbitrary element of length ≤ k. Then gs ∈ T0,n for some
n ≥ 1, so h ∈ gsHs−1g−1. Since, this holds for all such s, we have h ∈ gCkg−1.
Moreover, it is worth noticing the difference between K0, K1, and Ck:
• the intersections defining the Ck's involve conjugation of elements of finite length
(bounded by some k), while the first letter of the words can come from any of G0, G1.
• the intersections defining the Ki's involve conjugation of elements of arbitrary length,
where the first letter of the words comes from the same group.
Theorem 3.2. Let G = G0 ∗H G1 be a nondegenerate free product with amalgamation, and
let K0, K1 be as defined above. Then the following are equivalent:
(i) K0 = K1 = {e}.
(ii) for every finite F ⊆ G \ {e}, there exists g ∈ G such that gF g−1 ∩ H = ∅.
(iii) for every finite F ⊆ H \ {e}, there exists g ∈ G such that gF g−1 ∩ H = ∅.
Moreover, any one of these equivalent conditions implies that G is a Powers group and has
the free semigroup property.
Proof. It is obvious that (ii) implies (iii).
To see that (iii) implies (i) suppose first that ker G 6= {e}. Then pick f ∈ ker G \ {e} and
set F = {f} and clearly gf g−1 ∈ H for all g ∈ G, i.e., gf g−1 ∈ gF g−1 ∩ H. Next, assume
that ker G = {e}, but K0 6= K1. Recall from (5) and the subsequent note that this means
both K0 6= {e} and K1 6= {e}. Thus we pick fi ∈ Ki \ {e} for i = 0, 1 and set F = {f0, f1}.
Let g ∈ G be arbitrary. If g ∈ H there is nothing to show, so assume that g ends with a letter
from Gi \ H. But then gfig−1 ∈ H, i.e., gfig−1 ∈ gF g−1 ∩ H.
Finally we prove that (i) implies (ii). Choose an arbitrary finite set F ⊆ G \ {e}. The idea
is to conjugate the elements from F out of H one by one, and at the same time make sure we
do not conjugate any elements back into H.
Assume first there is an element f1 ∈ F ∩ H (else there is nothing to show). Because K0 is
trivial, then (4) means that all elements in H can be conjugated out of H by an element of
even length starting with a letter from G0 \ H. That is, we can find r1 = g0g1 ··· g2n1−1 such
that gi ∈ Gi (mod 2) \ H and r−1
1 f r1 : f ∈ F}. Assume that there is an element f2 ∈ F,
1 ··· g0
1 f2r1 ∈ H (otherwise we are done). Then there exists r2 = g0
so that r−1
0g0
2n2−1 such
1 f1r1r2 /∈ H.
1 f2r1r2 /∈ H. This also means that r−1
2 r−1
that g0
0 f1g0 ··· gj /∈ H, i.e., it belongs to
Indeed, let j be the smallest number such that g−1
0 f g0 ··· gjgj+1 ∈ Tj+1 (mod 2),3, and as we
Gj (mod 2) \ H = Tj (mod 2),1. Then g−1
··· g−1
Next, consider the set F1 = {r−1
2 r−1
i ∈ Gi (mod 2) \ H and r−1
1 f1r1 /∈ H.
j+1g−1
··· g−1
j
j
2 r−1
2 r−1
NIKOLAY A. IVANOV AND TRON OMLAND
6
continue to conjugate by elements alternating between G0 \ H and G1 \ H this product will
only increase in length.
1 f r1r2 : f ∈ F}. If all elements of F2 are outside H we are done,
Now we set F2 = {r−1
so assume that there is some f3 ∈ F such that r−1
It should be clear how this process continues, and since F is finite, we take r to be the
product of the ri's, and then r−1f r /∈ H for every f ∈ F.
and Remark 4.5].
Remark 3.3. The above result shows that [20, Theorem 3 (i)] can be slightly improved, as
Lemma 2.2 shows that "strongly Powers" is the same as Powers. In fact, using the comment
following (6), one can replace "Ck = {e} for some k" with any of the equivalent properties of
Theorem 3.2. Additionally, the countability assumption is no longer needed.
We will come back to the geometric interpretation of K0 and K1 in Section 5. In particular,
The last two observations follow from [17, Proposition 10] and [13, Example 4.4 (iii),(iv),
(cid:3)
1 f3r1r2 ∈ H.
Proposition 5.3 below gives more properties equivalent to those in Theorem 3.2.
Remark 3.4. In Section 4, we give an explicit example of a group Γ for which ker Γ = {e},
while K0 and K1 are both nontrivial. We show that the group has the unique trace property, but
is not C∗-simple. This gives a counterexample to the first author's statement [21, Corollary 4.7],
which the second author noticed to be incorrect.
Proposition 3.5. If G is countable, H is amenable, and K0 or K1 is nontrivial, then G is
not C∗-simple.
This result is generalized in Theorem 5.9 below, but in the context of recent works, we
provide two other short proofs of the statement.
First proof. We will show that H is recurrent in G in the sense of [23, Definition 5.1]. Let
(gn) be a sequence in G. Then at least one of the following holds:
(a) infinitely many elements from (gn) belong to H
(b) infinitely many elements from (gn) start with a letter from G0 \ H
(c) infinitely many elements from (gn) start with a letter from G1 \ H
, if it is (b) then K0 ⊆ T
it is (a) then H = T
K1 ⊆T
nk
nk
k gnk Hg−1
k gnk Hg−1
Pick a subsequence (gnk) of (gn) with elements from the one of (a),(b),(c) that holds. If
, and if it is (c) then
. If one of K0 and K1 is nontrivial, then both the other one and H are
(cid:3)
nontrivial as well. Hence, it follows from [23, Theorem 1.1] that G is not C∗-simple.
Second proof. If K0 = ker G, then K1 = ker G by the comment following (5). Thus, by
assumption, ker G is a nontrivial normal amenable subgroup of G, and hence G cannot be
C∗-simple. The similar argument holds if K1 = ker G, so we may assume that both K0 and
K1 are different from ker G.
k gnk Hg−1
nk
Choose a ∈ K0 \ ker G and b ∈ K1 \ ker G. Then
{gH : gH 6= agH} ⊆ {gH : g ∈ T1,k for some k ≥ 1}
{gH : gH 6= bgH} ⊆ {gH : g ∈ T0,k for some k ≥ 1},
and
which are clearly disjoint. By using the technique from [16, Proposition 5.8], explained in [27,
p. 12], the action of G on G/H gives rise to a unitary representation π : G → '2(G/H), that
r (G). It follows that (1− λ(a))(1− λ(b)) generates
extends to a continuous representation of C∗
(cid:3)
a proper two-sided closed ideal of C∗
Example 3.6. For any triple of groups H, G0, and G1, we have
r (G). Hence, G is not C∗-simple.
G = (G0 × H) ∗H (H × G1) ∼= (G0 ∗ G1) × H.
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
7
In this case H = ker G = K0 = K1, and G is C∗-simple if and only if H is C∗-simple. In
particular, this means that G can be C∗-simple even if ker G is nontrivial.
We now consider the special case where H is finite.
Theorem 3.7. Let G = G0 ∗H G1 be a nondegenerate free product with amalgamation, and
assume that H is finite.
Then the following are equivalent:
(i) G is icc
(ii) ker G = {e}
(iii) K0 = K1 = {e}
(iv) G is Powers
(v) G is C∗-simple
(vi) G has the unique trace property
(vii) there exists g ∈ G such that H ∩ gHg−1 = {e}
(viii) G has the free semigroup property
Proof. (i) =⇒ (ii): Since ker G ⊆ H, it is a finite normal subgroup of G, so it must be trivial
when G is icc.
(ii) =⇒ (iii): Let Ck be defined as in (6), so that ker G is the intersection of the decreasing
chain C0 ⊇ C1 ⊇ C2 ⊇ ··· . Since all the Ck's are subgroups of H, they are finite, so if
ker G = {e}, we must have Ck = {e} for some k. Hence, the remark following (6) explains
that K0 = K1 = {e}.
(iii) =⇒ (iv) follows from Theorem 3.2.
(iv) =⇒ (v) =⇒ (vi) =⇒ (i) is known to hold for all groups.
(iii) =⇒ (vii) is also a consequence of Theorem 3.2, by taking F = H \ {e}.
(vii) =⇒ (viii) follows from [13, Proposition 5.1].
(viii) =⇒ (ii): If H is finite and ker G 6= {e}, then there is no g ∈ G such that g ker G is
(cid:3)
semifree, i.e., G does not have the free semigroup property.
Remark 3.8. If G is an amalgam with finite H satisfying the equivalent conditions of
Theorem 3.7, then G has stable rank one by [13, Theorem 1.6] (where the argument is based
on [12, Corollary 3.9] for trivial H). However, there also exists amalgams G with finite H
such that G has stable rank one, but G does not satisfy the conditions of Theorem 3.7, as
explained in the paragraph following [13, Theorem 1.6].
Moreover, it follows from condition (vii) and [21, Corollary 3.6] that for G = G0 ∗H G1, the
positive cone of the K0-group of C∗
r (G) is not perforated, i.e.,
K0(C∗
r (G))+ = {γ ∈ K0(C∗
r (G)) : K0(τ)(γ) > 0} ∪ {0},
where τ is the canonical tracial state on C∗
easily seen to hold from the fact that for a finite group H, C∗
algebras and those have trivial K1-groups.
r (G). Indeed, the assumption K1(C∗
r (H)) = 0 is
r (H) is a direct sum of matrix
4. An example
denote the positive integers.
In this section ⊕, when applied to indices, will denote summation modulo 2, and N will
The group Γ = G0 ∗H G1 is defined as follows: first H is given by the set of generators
{h(i1, i2, . . . , in) : n ∈ N and ik ∈ {0, 1} for all k ∈ {1, . . . , n}},
with the relations
h(i1, i2, . . . , in)2 = e for all n ∈ N, k ∈ {1, . . . , n}, and ik ∈ {0, 1},
8
and for n ≥ k
h(i1, . . . , ik)h(j1, . . . , jn)h(i1, . . . , ik) =
(
NIKOLAY A. IVANOV AND TRON OMLAND
h(j1, . . . , jk, jk+1 ⊕ 1, . . . , jn)
h(j1, . . . , jk, jk+1, . . . , jn)
if n > k and i' = j' for all 1 ≤ ' ≤ k,
otherwise.
Then define
with the additional relations
Γ = hH ∪ {g0, g1}i,
g2
0 = g2
1 = e,
(g0h(1))3 = e,
(g1h(0))3 = e,
g0h(1, 0, i3, . . . , in)g0 = h(0, i3, . . . , in),
g1h(0, 0, i3, . . . , in)g1 = h(0, 0, i3, . . . , in),
g0h(1, 1, i3, . . . , in)g0 = h(1, 1, i3, . . . , in),
g1h(0, 1, i3, . . . , in)g1 = h(1, i3, . . . , in).
Finally, let
G0 = hH ∪ {g0}i and G1 = hH ∪ {g1}i,
so that Γ = G0 ∗H G1. Note also that
H = h{h(0, . . . , 0
n times
{z }
), n ∈ N} ∪ {h(1, 0, . . . , 0
n−1 times
{z }
), n ∈ N}i.
Lemma 4.1. For every g ∈ G0\H, there exists h ∈ H such that either g = g0h or g = h(1)g0h,
and for every g ∈ G1 \ H, there exists h ∈ H such that either g = g1h or g = h(0)g1h.
Consequently, [G0 : H] = [G1 : H] = 3, and the sets
S0 = {g0, h(1)g0}
and S1 = {g1, h(0)g1}.
(7)
provide representatives for the nontrivial left cosets of G0/H and G1/H, respectively.
Proof. First, if h is one of the generators of H and h 6= h(1), then the defining relations above
show that there exists an h0 such that hg0 = g0h0. Moreover, for h ∈ H, with h 6= h(1), we
have hh(1) = h(1) · h(1)hh(1) = h(1)h0 for some h0 6= h(1), that is, we can move the h(1)'s
to the left in the product. Hence, for any element g = (g0)h1g0h2g0 ··· g0hn(g0) ∈ G0, where
hi ∈ H, we can combine the two observations above to find the required h.
(cid:3)
A similar argument holds in G1.
Let us now denote the subgroups of H with elements determined by a prescribed start by
H(j1, . . . , j') = h{h(j1, . . . , j', i'+1, . . . , in) : n ∈ N, n ≥ ', i'+1, . . . , in ∈ {0, 1}}i
and the subgroups of H(j1, . . . , j') of elements having arguments of minimum length k by
Hk(j1, . . . , j') = h{h(j1, . . . , j', i'+1, . . . , in) : n ∈ N, n ≥ k ≥ ', i'+1, . . . , in ∈ {0, 1}}i.
Remark 4.2. Let i, j be two sequences such that h(i) and h(j) commute. For any k, ' greater
than the respective lengths of i, j, we have
hHk(i) ∪ H'(j)i = Hk(i)H'(j) = H'(j)Hk(i).
Let k ≥ 1 and note that g0h(1)g0 /∈ H. By using the defining relations for Γ, we compute
g0Hk(0)g0 = Hk+1(1, 0)
h(1)g0Hk(0)g0h(1) = h(1)Hk+1(1, 0)h(1) = Hk+1(1, 1)
H ∩ g0H(1)g0 = H(0)H(1, 1)
H ∩ h(1)g0H(1)g0h(1) = h(1)(H ∩ g0H(1)g0)h(1) = h(1)H(0)H(1, 1)h(1) = H(0)H(1, 0),
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
9
and then
H ∩ g0Hg0 = H(0)H2(1)
H ∩ h(1)g0Hg0h(1) = h(1)(H ∩ g0Hg0)h(1) = h(1)H(0)H2(1)h(1) = H(0)H2(1).
Finally,
H ∩ g0Hk(0)H(1)g0 = g0Hk(0)g0
(cid:0)H ∩ g0H(1)g0
(cid:1) = Hk+1(1, 0)H(0)H(1, 1),
where the first equality follows from Dedekind's modular law for groups, and
H ∩ h(1)g0Hk(0)H(1)g0h(1) = h(1)(H ∩ g0Hk(0)H(1)g0)h(1)
= h(1)Hk+1(1, 0)H(0)H(1, 1)h(1) = Hk+1(1, 1)H(0)H(1, 0).
Similarly,
H ∩ g1Hg1 = H ∩ h(0)g1Hg1h(0) = H(1)H2(0)
H ∩ g1H(0)Hk(1)g1 = Hk+1(0, 1)H(1)H(0, 0)
H ∩ h(0)g1H(0)Hk(1)g1h(0) = Hk+1(0, 0)H(1)H(0, 1).
Lemma 4.3. We have K0 = H(0), K1 = H(1), and ker Γ = K0 ∩ K1 = H(0) ∩ H(1) = {e}.
Proof. We use the notation for Cj,k from (2), so that C0,0 = C1,0 = H, and remark that for
every j, k we have
gC0,kg−1.
g∈G1\H
In the following we will make use of Remark 4.2, the coset representatives from (7), and the
fact that every Cj,k is invariant under conjugation by elements of H. We compute that
gHg−1 = H ∩ g0Hg0 ∩ h(1)g0Hg0h(1) = H(0)H2(1)
and C1,k+1 = \
and
g∈G0\H
gC1,kg−1
C0,k+1 = \
H ∩ C0,1 = H ∩ \
H ∩ C1,1 = H ∩ \
Next, let k ∈ N and assume kT
C0,i = H(0)Hk+1(1) and kT
C0,i = H ∩ k\
\
(cid:16) k\
(cid:16) k\
gC1,ig−1
k+1\
g∈G0\H
g∈G0\H
g∈G1\H
(cid:17)
i=0
i=0
gHg−1 = H ∩ g1Hg1 ∩ h(0)g1Hg1h(0) = H(1)H2(0).
C1,i = H(1)Hk+1(0). Then
i=0
(cid:17)
C1,i
g0 ∩ h(1)g0
i=0
= H ∩ g0
= H ∩ g0H(1)Hk+1(0)g0 ∩ h(1)g0H(1)Hk+1(0)g0h(1)
= Hk+2(1, 0)H(0)H(1, 1) ∩ Hk+2(1, 1)H(0)H(1, 0)
= H(0)Hk+2(1),
g0h(1)
C1,i
i=0
i=0
C1,i = H(1)Hk+2(0).
k+1\
C0,k = \
K0 = \
i=0
k≥0
k≥0
H(0)Hk+1(1) = H(0)
and similarly
Hence, we get that
10
and
NIKOLAY A. IVANOV AND TRON OMLAND
K1 = \
C1,k = \
k≥0
k≥0
H(1)Hk+1(0) = H(1),
and thus ker Γ = K0 ∩ K1 = H(0) ∩ H(1) = {e}.
(cid:3)
Theorem 4.4. The group Γ defined above has the unique trace property, but is not C∗-simple.
Proof. By Lemma 4.3 we have ker Γ = {e}, so Proposition 3.1 gives that Γ has the unique
trace property. Moreover, Γ is countable, H is amenable since it is locally finite, and K0 and
(cid:3)
K1 are nontrivial. Hence, it follows from Proposition 3.5 that Γ is not C∗-simple.
Note that Γ contains uncountably many amenable (abelian) subgroups, for example the
ones generated by subsets of
{z }
{h(0, . . . , 0
n times
, 1) n ∈ N)},
so [7, Theorem 1.7] does not apply.
isomorphic to Z3 ∗ Z3, which clearly contains a nonabelian free group as a subgroup.
We can find an explicit free nonabelian subgroup of Γ by checking that hg0h(1), g1h(0)i is
Finally, we remark that hK0 ∪ K1i = H, and that the normal closure of H is all of Γ.
Proposition 4.5. Let
θ: Γ → Z/2Z × Z/2Z
be the group homomorphism defined on generators by
θ(h(0)) = (−1, 1),
θ(h(1)) = (1,−1),
θ(g0) = (1,−1),
θ(g1) = (−1, 1).
Set Γ0 = ker θ. Then Γ0 is simple.
Proof. Below we only provide the idea of the argument and omit most of the technicalities.
Observe that since t−1h(0)t = h(0, i1, . . . , i2k) for a suitable element t ∈ T0,2k, it follows
that θ(h(0, i1, . . . , i2k)) = θ(h(0)) = (−1, 1). Therefore
h(0)h(0, i1, . . . , i2k) ∈ Γ0
for all k ∈ N, j ∈ {1, . . . , 2k}, and ij ∈ {0, 1}. Analogous arguments yield
h(0)h(1, i1, . . . , i2k−1), h(1)h(1, i1, . . . , i2k), h(1)h(0, i1, . . . , i2k−1) ∈ Γ0
for all k ∈ N, j ∈ {1, . . . , 2k}, and ij ∈ {0, 1}.
Notice also that g1h(0), h(0)g1, g0h(1), h(1)g0 ∈ Γ0. One can now check that
Γ0 = h{h(0)h(0, i1, . . . , i2k), h(0)h(1, i1, . . . , i2k−1), h(1)h(1, i1, . . . , i2k),
h(1)h(0, i1, . . . , i2k−1),∀k ∈ N,∀ij ∈ {0, 1},∀j} ∪ {h(0)g1, h(1)g0}i.
Let N 6= {e} be a normal subgroup of Γ0 and pick an element a ∈ N \ {e}. The remainder
of the proof is about showing that each of the generators of Γ0 listed above can be described by
a suitable product of conjugates of a by elements of Γ0. Since these computations are rather
(cid:3)
tedious, we leave them out.
Remark 4.6. If G is an exact group with stable rank one and the unique trace property, then
G is C∗-simple by [2, Theorem 2.1]. Let Γ = G0 ∗H G1 be the group defined above. Since H
is amenable and has finite index in G0 and G1, both these groups are amenable, and therefore
Γ is exact by [11, Corollary 3.3]. Hence, Γ does not have stable rank one by Theorem 4.4.
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
11
Remark 4.7. It was pointed out to us by Adrien Le Boudec that the group Γ is isomorphic
to the group G(A3, S3)?, which is one of the examples from [5, Section 5].
Let S3 and A3 denote the symmetric and alternating group, respectively, on a three-element
set. Consider a 3-regular tree T and color the edges {1, 2, 3}, so that neighboring edges have
different colors. Let σ(g, v) ∈ S3 be the permutation of the three colors induced in the natural
way by the element g ∈ Aut(T). Then G(A3, S3) < Aut(T) is the group of all automorphisms
g of T such that σ(g, v) ∈ A3 for all but finitely many vertices v. Then G(A3, S3)? is the
subgroup of G(A3, S3) of index two preserving the natural bipartition of vertices of T.
To see this, remark that by Bass-Serre's theory (cf. [30, I.4 Theorem 6]) this index two
subgroup is an amalgamated product ¯G0 ∗ ¯H
¯G1, where ¯G0 and ¯G1 are the stabilizers of two
adjacent vertices (we will call them v0 and v1), and ¯H is the stabilizer of the edge between
them (we will call it e). Then, using the notation from [4, Subsection 3.1], we may write ¯Gi
as an increasing union
G(S3)vi = [
n≥1
Kn(vi),
i = 0, 1,
where Kn(vi) is the subgroup of G(A3, S3) that fixes the vertex vi and has σ(g, v) ∈ A3
for all g ∈ Kn(vi) and all v that are at a distance larger than n from vi. With this nota-
tion, the element h(0, i2, . . . , in) acts on T by swapping the two half-trees, emanating from
a vertex (we will call it v(0, i2, . . . , in)) that is at a distance n − 1 from v0 and at a dis-
tance n from v1, and not intersecting the geodesic between v0 and v(0, i2, . . . , in). Clearly,
σ(h(0, i2, . . . , in), v(0, i2, . . . , in)) /∈ A3, because it leaves one edge (therefore one color) fixed.
The matching of the other vertices of the half-trees is defined so the local permutations belong
to A3. This matching is just a matter of orientation of the tree. Adding the element g1 to the
picture, we see that Kn(v0) is isomorphic to the wreath product Z2 o ··· o Z2 o S3 (n − 1 factors
of Z2), where the top elements beneath the S3 factor are h(1), h(0, 0), and h(0, 1). Likewise
Kn(v1) is isomorphic to Z2 o ··· o Z2 o S3 (n − 1 factors of Z2), where the top elements beneath
the S3 factor are h(0), h(1, 0) and h(1, 1). In this way, we see that ¯Gi is isomorphic to Gi,
i = 0, 1, and therefore Γ is isomorphic to G(A3, S3)?.
The simplicity of the group Γ0 is not covered by [4, Corollary 4.20], because A3 is not
We finally note that G(A3, S3) is isomorphic to Γ(cid:111)Z2, where Z2 acts on Γ by interchanging
generated by its point stabilizers.
the indices 0 and 1 of all generating elements of Γ.
5. Actions of free products with amalgamation
Let G be any group acting on a space X, and let us first recall some notation. The stabilizer
subgroup of an element x ∈ X is Gx = {g ∈ G : gx = x}, and the fixed-point set of g ∈ G is
X g = {x ∈ X : gx = x}. The kernel of the action is the set of elements in G acting trivially
on X, namely
ker(G (cid:121) X) = \
x∈X
Gx = {g ∈ G : X g = X}.
Note that for every x ∈ X and all s, g ∈ G, we have sGxs−1 = Gsx and X sgs−1 = sX g, so the
kernel is a normal subgroup of G. The action is called faithful when the kernel is trivial, i.e.,
if for every g ∈ G \ {e} there exists x ∈ X such that gx 6= x, and strongly faithful if for every
finite set F ⊆ G \ {e} there exists x ∈ X such that gx 6= x for all g ∈ F.
Furthermore, the action of G on X is called free if whenever g ∈ G, x ∈ X, and gx = x,
then g = e. Since hGx : x ∈ Xi is invariant under conjugation, it is a normal subgroup of G,
which coincides with the subgroup of G generated by {g ∈ G : X g 6= ∅}. This subgroup is the
so-called "join" of {Gx : x ∈ X}, while ker(G (cid:121) X) is the "meet" of {Gx : x ∈ X}. Obviously,
G acts freely on X if and only if this subgroup is trivial.
12
NIKOLAY A. IVANOV AND TRON OMLAND
int(G (cid:121) X) = hGo
Let X be a topological space, and suppose that G acts continuously on X, that is, by
x as the subgroup of Gx consisting of all elements
x for some x ∈ X if and only if
homeomorphisms. For every x ∈ X define Go
that fix a neighborhood of x pointwise. We notice that g ∈ Go
X g has nonempty interior, and we define the interior of the action as
(8)
Then, by using the identity X sgs−1 = sX g, we see that {g ∈ G : X g has nonempty interior}
is invariant under conjugation. Indeed, if X g contains a nonempty open subset V , then sV is
a nonempty open subset of sX g = X sgs−1. Therefore, int(G (cid:121) X) is automatically a normal
sx for every x ∈ X and s ∈ G. We say
subgroup of G. One may also check that sGo
that the action is topologically free if X g has empty interior for every g ∈ G \ {e}, that is, if
the interior is trivial. Finally, we note that for a topological space X, the interior and kernel
of an action corresponds to the join and meet, respectively, of {Go
x : x ∈ Xi = h{g ∈ G : X g has nonempty interior}i.
xs−1 = Go
x : x ∈ X}.
Now, fix a nondegenerate amalgam G = G0 ∗H G1 and let T denote its Bass-Serre tree
(cf. [30], see also [20] and references therein). Then T has vertex set G/G0 t G/G1 and
(geometric) edge set G/H. Two vertices in T are adjacent if either of the form
(g0)g1 ··· g2n−1G0
(g0)g1···g2n−1g2nH
−−−−−−−−−−−−→ (g0)g1 ··· g2n−1g2nG1
or of the form (for a transversal edge)
(g0)g1 ··· g2nG1
(g0)g1···g2ng2n+1H
−−−−−−−−−−−−→ (g0)g1 ··· g2ng2n+1G0
n=0 and (yn)∞
for gi ∈ Gi (mod 2) \ H.
Let V be the set of vertices and E the set of edges of T, and let s, r: E → V denote the
source and range maps. Given any two vertices v, w, there are exactly two paths between
them (one starting in v and ending in w, and one in the opposite direction), and the length of
these paths is the combinatorial distance d(v, w). A ray in T is a sequence (xn)∞
n=0 of vertices,
which is geodesic in the sense that d(xm, xn) = m − n for all m, n (i.e., xn+2 6= xn for all
n). Moreover, given two rays (xn)∞
n=0 in T, we say they are cofinal, and write
n=0 ∼ (yn)∞
(xn)∞
n=0 if there exist integers k and N such that yn = xn+k for all n > N. Define
the boundary ∂T of the Bass-Serre tree T as the set of equivalence classes of cofinal rays.
For e ∈ E, define Z0(e) = {v ∈ V : d(v, s(e)) > d(v, r(e))}, i.e., the set of all vertices that
are closer to the endpoint of e than the starting point. Moreover, define Z∞(e) as the set of all
n=0 such that xj = s(e) and xj+1 = r(e) for some j ≥ 0, and then define ZB(e) ⊂ ∂T
rays (xn)∞
as Z∞(e)/ ∼. Finally, set Z(e) = Z0(e) ∪ ZB(e). The family of all finite intersections of
sets from the collection {Z(e) : e ∈ E} forms a base of compact clopen sets for a totally
disconnected compact Hausdorff topology on V ∪ ∂T, sometimes called the "shadow topology"
on V ∪ ∂T. We refer to [25, Section 4, especially Proposition 4.4] in this regard (there it is
assumed that T is countable, but their proofs hold also without this hypothesis, although then
the topology is not metrizable).
Moreover, by removing an edge from T, we get two components, so-called half-trees. An
extended half-tree is a half-tree together with all its associated boundary points.
In this
terminology, as explained in [6, Section 4.3], the shadow topology is generated by all the
extended half-trees of V ∪ ∂T.
Next, define F ⊆ V as the set of all vertices v such that s−1(v) = r−1(v) is finite, i.e., only
finitely many edges start and end in v. The following can be deduced from sections of [6, 25]
mentioned above:
Proposition (new). The closure ∂T of ∂T in V ∪ ∂T is (V \ F) ∪ ∂T, and is compact,
minimal, and G-invariant. Moreover, ∂T is closed in V ∪ ∂T if and only if F = V , if and
only if T is locally finite, if and only if H has finite index in both G0 and G1.
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
13
Henceforth, we fix sets S0 and S1 of representatives for the nontrivial left cosets of G/G0
and G/G1, respectively. Then every x ∈ ∂T can be uniquely represented by an infinite word
(g0)g1g2 ··· , where gi ∈ Si (mod 2), that is, we take either x0 = g0G1, x1 = g0g1G0 etc. or
x0 = G1, x1 = g1G0, etc., i.e., every class of cofinal rays will be represented by the unique ray
in the class starting with gG1 for g ∈ S0 ∪ {e}.
The boundary ∂T becomes a totally disconnected locally compact Hausdorff space when
equipped with the subspace topology coming from the shadow topology on V ∪ ∂T. This
topology is generated by basic clopen sets U((g0)g1 ··· gn), where n ≥ 0 and gi ∈ Si (mod 2),
consisting of all equivalence classes of cofinal rays that are identified with infinite words
starting with (g0)g1 ··· gn.
The amalgam G acts on its Bass-Serre tree T by left translation, that is, the action of
s ∈ G on vertices is given by s · gG0 = sgG0, s · gG1 = sgG1 and on edges by s · gH = sgH.
Clearly, this also induces an action of G on the boundary of its Bass-Serre tree ∂T.
Lemma (new). Let g, s ∈ G, and suppose that s fixes U(g) pointwise. Then sgH = gH.
Consequently, s fixes every vertex in any ray coming from an infinite word starting with g.
For a complete proof that holds in a more general case, see [8, Lemma 3.6].
Sketch of proof. The idea is to assume that sgH 6= gH, and then construct an infinite word
starting with g, that gives rise to a ray that is not fixed (up to cofinality) by s. The argument
involves division into several subcases, where the infinite word depends upon the last letter of
(cid:3)
g, and the first and last letter of g−1sg.
n ··· g−1
1 (g−1
0 ),
Recall (3) and define the set
K((g0)g1 ··· gn) = (g0)g1 ··· gnKn+1 (mod 2)g−1
(9)
where gi ∈ Si (mod 2). This is the subgroup of G consisting of all elements that fix the basic
open set U((g0)g1 ··· gn) pointwise.
Lemma 5.1. We have that ker G = ker(G (cid:121) T) = ker(G (cid:121) ∂T) = ker(G (cid:121) ∂T), and that
int(G (cid:121) ∂T) equals int(G (cid:121) ∂T) and coincides with the normal closure of K0 ∪ K1.
Proof. It should be clear that ker G = ker(G (cid:121) T) ⊆ ker(G (cid:121) ∂T) ⊆ H. Indeed, the latter
inclusion follows from the new lemma above. Suppose that h ∈ H \ ker G. Then there exists g
such that ghg−1 /∈ H. Moreover, we can find a g of the form (g0)g1 ··· gn with gi ∈ Si (mod 2)
with this property. Let x be any ray starting with the corresponding vertices. Then hx 6∼ x,
so h /∈ ker(G (cid:121) ∂T) ⊆ H The equality ker(G (cid:121) ∂T) = ker(G (cid:121) ∂T) follows from continuity
of the action.
To see that int(G (cid:121) ∂T) = int(G (cid:121) ∂T), note first that if (∂T)g has nonempty interior,
i.e., there exists open nonempty V ⊆ (∂T)g, then V ∩ ∂T ⊆ (∂T)g is open nonempty in ∂T.
Next, if (∂T)g has nonempty interior, i.e., there exists open nonempty U ⊆ (∂T)g, then there
exists open V ⊆ ∂T such that U = V ∩ ∂T. Using density and continuity of the action, it
follows that V ⊆ U ⊆ (∂T)g.
Next, K0 fixes U(g0), i.e., all sequences of vertices starting with g0G1 for any g0 ∈ S0,
pointwise, and K1 fixes U(g1), i.e., all sequences of vertices starting with G1, g1G0 for any
g1 ∈ S1, pointwise. Hence, K0 ∪ K1 ⊆ int(G (cid:121) ∂T). Therefore, as the latter is normal, the
same inclusion holds for the normal closure of K0 ∪ K1.
Pick g ∈ G and suppose that (∂T)h has nonempty interior. Then h must fix some basic
open set pointwise, say U(g) for g = (g0)g1 ··· gn. This means that h ∈ K(g), as defined in
(9), so h ∈ gKn+1 (mod 2)g−1, that is, h belongs to the normal closure of Kn+1 (mod 2). Since
int(G (cid:121) ∂T) is generated by {h ∈ G : (∂T)h has nonempty interior}, the conclusion follows.
Note that h also fixes U((g0)g1 ··· gn+1), so h belongs to the normal closure of the other
Ki as well. In fact, the normal closures of K0, K1, and K0 ∪ K1 are all the same.
(cid:3)
NIKOLAY A. IVANOV AND TRON OMLAND
14
Definition 5.2. The subgroup int(G (cid:121) ∂T) = h{g ∈ G : (∂T)g has nonempty interior}i of
G, or equivalently, the normal closure of K0 ∪ K1 in G, will be called the interior of G and
denoted int G.
In Proposition 5.4 below, we show that G is C∗-simple if and only if int G is C∗-simple,
giving an analog of Proposition 3.1 (however int G can be all of G).
Proposition 5.3. The following are equivalent:
(i) int G = {e},
(ii) G (cid:121) T is strongly faithful,
(iii) G (cid:121) ∂T is strongly faithful,
(iv) G (cid:121) ∂T is topologically free, i.e., G (cid:121) T is slender in the sense of [20].
Proof. The equivalences between (i), (iv), and condition (i) of Theorem 3.2 follow directly
from Lemma 5.1 and Definition 5.2. Moreover, condition (ii) of Theorem 3.2 coincides with
condition (SF) from [17, p. 245]. Remark that in [17], the notations Edg+X and Y are used
for T and ∂T, respectively. In particular, (ii) is the same as condition (SF), and thus the
above means that (ii) is equivalent with (i) and (iv). Finally, it follows from [17, Lemma 9]
that (ii) implies (iii), and it is stated immediately after the proof of [17, Lemma 9] that its
(cid:3)
converse holds as well.
Using the terminology of [5], we see from the proof of Lemma 5.1 that K0 and K1 are
precisely the fixators of the half-trees of T obtained by removing the edge H.
Proposition 5.4. Let G be a nondegenerate free product with amalgamation. Then G is
C∗-simple if and only if int G is C∗-simple.
Proof. Suppose that int G is C∗-simple.
First, assume that int G ⊆ H. For the moment, write Ki(G) and Ki(G/ ker G) for the
Ki's corresponding to the amalgams G and G/ ker G, respectively. Clearly, a word in G
starts with a letter in Gi \ H if and only if its image in G/ ker G starts with a letter in
(Gi/ ker G) \ (H/ ker G). Thus, we see that Ki(G)/ ker G ∼= Ki(G/ ker G).
Because int G is a normal subgroup of G contained in H, we must have int G = ker G, and
then also K0 = K1 = ker G. Thus, G/ int G is C∗-simple by Theorem 3.2. Since we have
assumed that int G is C∗-simple, it follows from [7, Theorem 1.4] that G is C∗-simple.
Next, let CG(int G) denote the centralizer of int G in G, and suppose that g ∈ CG(int G)\ H.
In particular, this means that g commutes with all elements in K0 and K1, so gKig−1 = Ki
i ⊇ Ki. From
for i = 0, 1. Moreover, for all gi ∈ Gi \ H, we always have that giKi+1 (mod 2)g−1
this it follows that K0 = K1, which means that ker G = int G.
Indeed, let g have length n and denote by gn the last letter of g. If n is odd, then
Kn (mod 2) = gKn (mod 2)g−1 ⊆ ··· ⊆ gnKn (mod 2)g−1
n ⊆ Kn+1 (mod 2),
which easily implies K0 = K1, and hence ker G = int G. If n is even, then
Kn+1 (mod 2) = gKn+1 (mod 2)g−1 ⊆ ··· ⊆ gnKn (mod 2)g−1
n ⊆ Kn+1 (mod 2),
and therefore all containments are equalities. Assume that [Gj : H] ≥ 3, for j = 0 or 1, pick a
letter gj ∈ Gj \ H of g, and choose another element g0
j ∈ Gj \ H.
Then
j ∈ Gj \ H such that g−1
j g0
j)−1gjKj (mod 2)g−1
(g0
j g0
j = (g0
j)−1Kj+1 (mod 2)g0
j ⊇ Kj (mod 2).
Moreover,
meaning that ker G = int G.
j)−1gjKj (mod 2)g−1
(g0
j g0
j ⊆ Kj+1 (mod 2),
so Kj (mod 2) ⊆ Kj+1 (mod 2),
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
15
Finally, if CG(int G) ⊆ H, then CG(int G) ⊆ ker G, since it is normal in G and contained in
H. Then CG(int G) = Z(int G), since it is contained in int G, so it must be trivial since int G
is assumed to be C∗-simple. Hence, it follows from [7, Theorem 1.4] that G is C∗-simple.
(cid:3)
The converse holds by [7, Theorem 1.4] because int G is a normal subgroup of G.
Proposition 5.5. Suppose that ker G is trivial. Then G is a weak∗ Powers group.
Proof. The action of G on the boundary of its Bass-Serre tree is always minimal and strongly
hyperbolic (see [20, Proposition 19]). If ker G is trivial, then the action is also faithful by
Lemma 5.1. In [17, Lemma 4] we now replace "strongly faithful" by "faithful", and the first
part of the proof still works, under the assumption that F 6= {e} is a one-element set. The
(cid:3)
rest of the argument goes along the same lines.
Example 5.6. The group Γ from Section 4 is a weak∗ Powers group that is not C∗-simple.
We complete this section by some facts about boundary actions and refer to [7, 22, 27]
for further details. An action of G on a compact space X is called a boundary action if it
is minimal (i.e., the orbits are dense) and strongly proximal (i.e., the orbit-closure in P(X)
of every probability measure on X contains a point mass). In this case we also say that the
space X is a G-boundary.
For every group G there is a universal G-boundary called the Furstenberg boundary and
denoted ∂F G. If X is any other G-boundary, then there exists a continuous surjective G-
equivariant map ∂F G → X. Moreover, we remark that for every g ∈ G, the set (∂F G)g is
always clopen, cf. [7, Lemma 3.3].
In Section 7 we will often make use of the following observation from [7].
Lemma 5.7. Let G be a group, N a normal subgroup of G, and L a normal subgroup of N.
Suppose that g ∈ N is such that (∂F G)g 6= ∅. Then (∂F (N/L))gL 6= ∅.
Proof. First, the quotient map N → N/L gives rise to an action of N on ∂F (N/L), which
makes ∂F (N/L) an N-boundary. Thus, there exists an N-equivariant continuous surjective
map ψ : ∂F N → ∂F (N/L). Next, [7, Lemma 5.2] says that the N-action on ∂F N extends to
an action of G, such that ∂F N becomes a G-boundary. Therefore, there is a G-equivariant
continuous surjective map φ: ∂F G → ∂F N, so altogether we have surjections
φ−→ ∂F N
ψ−→ ∂F (N/L).
∂F G
Now, if g ∈ N and (∂F G)g 6= ∅, there exists x ∈ ∂F G such that gx = x. It follows that
gψ(φ(x)) = ψ(φ(gx)) = ψ(φ(x)), so ψ(φ(x)) ∈ (∂F (N/L))g = (∂F (N/L))gL, which is therefore
nonempty. In particular, we have ψ(φ((∂F G)g)) ⊆ (∂F (N/L))gL.
(cid:3)
Lemma 5.8. Suppose that G is a nondegenerate amalgam. Then ∂T is a G-boundary.
Proof. The action of a nondegenerate amalgam G = G0 ∗H G1 on its Bass-Serre Tree T is
minimal and strongly hyperbolic (see [20, Proposition 19]), that is, of "general type" in the sense
of [6, Section 4.3]. It follows that the action of G on ∂T is minimal by [20, Proposition 19] and
Lemma 5.1, extremely proximal by [6, Proposition 4.26] (see [6, Section 2.1] for terminology),
(cid:3)
and thus strongly proximal by [14, Theorem 2.3 (3.3)]. Hence, ∂T is a G-boundary.
Theorem 5.9. Suppose that K0 or K1 is amenable. Then G is C∗-simple if and only if both
K0 and K1 are trivial.
Proof. Let x ∈ ∂T be represented by a sequence of vertices g0G1, g0g1G0, etc., where
gi ∈ Si (mod 2) (of course, the argument is similar if it starts with G1, g1G0, etc). Then Go
x is
the direct limit of the sequence K(g0) ⊆ K(g0g1) ⊆ ··· , that is, of
0 ⊆ ··· .
0 ⊆ g0g1K0g−1
K0 ⊆ g0K1g−1
1 g−1
16
NIKOLAY A. IVANOV AND TRON OMLAND
x is amenable.
Clearly, K0 is amenable if and only if K1 is amenable, since either of them is a subgroup of a
conjugate of the other, see (5). Therefore, all K((g0)g1 ··· gn) are also amenable, since they
are conjugates of K0 or K1, see (9). As the class of amenable groups is closed under direct
limits, we have that Go
Assume that G is C∗-simple. Since ∂T is a G-boundary by Lemma 5.8 and G is assumed to
be C∗-simple, we may use [7, Corollary 7.5] to say that C(∂T) (cid:111)r G is simple. Then it follows
from [27, Theorem 14 (2)] that G (cid:121) ∂T is topologically free, so G (cid:121) ∂T is topologically free
by Lemma 5.1, that is, int G = {e}. Hence, Proposition 5.3 gives that both K0 and K1 are
trivial.
(cid:3)
x is nonamenable for all x ∈ ∂T if K0 or K1 is nonamenable,
Conversely, if K0 = K1 = {e}, then G is C∗-simple by Theorem 3.2.
A similar argument shows that Go
and it seems likely that this implies C∗-simplicity of G.
Remark 5.10. Any amalgamated free product where K0 and K1 are nontrivial, amenable,
and K0∩K1 = {e} is not C∗-simple, but has the unique trace property. As noted in Remark 4.7,
the group Γ of Section 4 is isomorphic to one of Le Boudec's examples from [5, Section 5].
However, if the groups G0 and G1 are nonisomorphic, then such an amalgamated product will
not be covered by [5, Section 5].
6. Radical and residual classes of groups
In this section, by a class of groups, we will always mean a class X of groups that contains
the trivial group and is closed under isomorphisms, i.e., if G ∈ X and H ∼= G, then H ∈ X.
In [29], this is called a group theoretical class.
Let X be a class of groups and let G be any group. Define ρ(G) as the normal subgroup of
G generated by all normal subgroups of G that belong to X, and ρ∗(G) as the intersection of
all normal subgroups of G with quotient belonging to X, i.e.,
ρ(G) =Y{N : N / G and N ∈ X},
ρ∗(G) =\{N : N / G and G/N ∈ X}.
These are both normal subgroups of G, called join and meet of the respective families.
Whenever more than one class is around, we will often write ρX and ρ∗
X.
Definition 6.1. A class of groups X is called a radical class if it is closed under quotients
(i.e., closed under homomorphic images), and if for any group G we have
(i) ρ(G) ∈ X,
(ii) ρ(G/ρ(G)) = {e}.
A class of groups X is called a residual (or coradical or semisimple) class if it is closed under
normal subgroups, and if for any group G we have
(i*) G/ρ∗(G) ∈ X,
(ii*) ρ∗(ρ∗(G)) = ρ∗(G).
Clearly, if X is radical, then for any G we have ρ(ρ(G)) = ρ(G), and G ∈ X if and only if
ρ(G) = G. Moreover, if X is residual, then for any G we have ρ∗(G/ρ∗(G)) = {e}, and G ∈ X
if and only if ρ∗(G) = {e}.
The above definitions are not completely consistent within references; some say that a class
is radical if it is closed under quotients and (i) holds, and strict radical if (ii) holds as well
(similarly for residual), and there are possibly other variations.
Proposition 6.2. A class of groups is radical if and only if it is closed under quotients,
extensions, and satisfies (i). Moreover, a class of groups is residual if and only if it is closed
under normal subgroups, extensions, and satisfies (i*).
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
17
Proof. This is explained in [29, Theorems 1.32 and 1.35], where (i) and (i*) are equivalent to
the properties N X = X and RX = X of [29, p. 20 and p. 23], respectively, see [29, p. 19]. (cid:3)
Let X be a class of groups closed under quotients, and define X∗ to be all groups satisfying
ρ(G) = {e}, that is, the class of groups with no nontrivial normal subgroups in X, i.e.,
X∗ = {G : ρ(G) = {e}} = {G : N / G, N 6= {e} ⇒ N /∈ X}.
Similarly, if X is a class of groups closed under normal subgroups, define X∗ to be all groups
satisfying ρ∗(G) = G, that is, the class of groups with no nontrivial quotients in X, i.e.,
X∗ = {G : ρ∗(G) = G} = {G : N / G, N 6= G ⇒ G/N /∈ X}.
Proposition 6.3. Let X be a class of groups closed under quotients. Then X is radical if
and only if X = (X∗)∗ if and only if X = Y∗ for some class of groups Y that is closed under
normal subgroups.
Let X be a class of groups closed under normal subgroups. Then X is residual if and only
if X = (X∗)∗ if and only if X = Y ∗ for some class of groups Y that is closed under quotients.
(cid:3)
Proof. This follows from [29, Theorems 1.38 and 1.39], see [29, p. 6-7] for notation.
Example 6.4. Let X be the class of amenable groups, which is known to be a radical class.
Then [7, Theorem 1.3] shows that the class of groups with the unique trace property coincides
with X∗, and is therefore residual by Proposition 6.3. Thus, a group is amenable if and only
if it does not have any nontrivial quotient with the unique trace property.
Example 6.5. Let X be the class of Powers (resp. weak Powers, weak∗ Powers) groups,
which is not closed under extensions, so it is not a residual class. However, X is closed under
normal subgroups by Lemma 2.2, meaning that we can define the class X∗ of groups with no
nontrivial quotient which is a Powers (resp. weak Powers, weak∗ Powers) group. Moreover,
X∗ is a radical class, but ρX∗(G) = {e} does not imply that G ∈ X.
If X is the class of all Powers groups, then P = (X∗)∗ is the residual closure of X, that is,
the smallest residual class containing all Powers groups. In Section 7 we will see that the class
of C∗-simple groups is residual, and hence it contains P (it could possibly coincide with P).
Example 6.6. Some radical classes of groups are e.g. locally finite groups, elementary
amenable groups, amenable groups, and groups that do not contain any nonabelian free
subgroup. We denote the latter class by N F, which gives rise to the residual class AF = (N F)∗,
consisting of all groups for which every nontrivial normal subgroup contains a free nonabelian
subgroup, and G ∈ N F if and only if ρ∗
AF (G) = G. Since every amenable group belongs
to N F, every group in AF has the unique trace property. Moreover, (1) gives that every
amalgamated free product with trivial kernel belongs to AF.
Both the class of amenable groups and N F are closed under subgroups, but in general,
radical classes are not necessarily closed even under normal subgroups. A radical class that is
closed under normal subgroups is sometimes called hereditary.
Lemma 6.7. Let X be a class of groups satisfying (i). Then X is closed under normal
subgroups if and only if for any group G and normal subgroup N of G we have
ρ(N) = ρ(G) ∩ N.
In particular, this implies that ρ(N) is a normal subgroup of G.
Proof. If G ∈ X and N is normal in G, then ρ(N) = ρ(G) ∩ N = G ∩ N = N, so N ∈ X.
The converse is explained in [29, Lemma 1.31, Corollaries 1 and 2].
(cid:3)
18
NIKOLAY A. IVANOV AND TRON OMLAND
A direct proof in the special case of countable amenable groups is given in [31, Corollary B.4].
The result below is similar to [31, Corollary B.6], using Lemma 6.7.
For a group G and a subset N, we let ZG(N) and Z(G) denote the centralizer of N in G,
and the center of G, respectively.
Lemma 6.8. Let X be a class of groups that satisfies (i) and is closed under normal subgroups.
Assume that G is any group and N is a normal subgroup of G such that ρ(N) = {e}.
Then ρ(G) = ρ(ZG(N)).
Proof. Since {e} = ρ(N) = ρ(G) ∩ N, the normal subgroups ρ(G) and N commute, so
ρ(G) ⊆ ZG(N). Hence, ρ(ZG(N)) = ρ(G) ∩ ZG(N) = ρ(G).
(cid:3)
Lemma 6.9. Let X be a class of groups that satisfies (i), is closed under normal subgroups,
and contains all abelian groups. Assume that G is any group and N is a normal subgroup of
G such that both ρ(N) and ρ(G/N) are trivial. Then ρ(G) = {e}.
Proof. First, Z(N) is normal in N, so Lemma 6.7 gives that ρ(Z(N)) ⊆ ρ(N) = {e}, and since
Z(N) ∈ X, we must have Z(N) = {e}. Thus, the map ZG(N) → G/N, x 7→ xN is injective,
and ρ(G/N) = {e} implies ρ(ZG(N)) = {e}. By Lemma 6.8, we thus get ρ(G) = {e}.
(cid:3)
Lemma 6.10. Let X be a residual class, and suppose that X∗ is closed under normal subgroups
and contains all abelian groups. Let G be any group and N a normal subgroup of G. Then G
belongs to X if and only if both N and ZG(N) belong to X.
Proof. Clearly, X∗ satisfies the conditions of Lemmas 6.8 and 6.9, and by Proposition 6.3 we
know that G ∈ X if and only if ρX∗(G) = {e}.
(cid:3)
The above result is an analog of [7, Theorem 1.4], while the result below shows that when
X is a residual class and G is a group, then ρX∗(G) is the smallest normal subgroup of G that
produces a quotient in X.
Lemma 6.11. Let X be a class of groups closed under normal subgroups and suppose that G
is a group with a normal subgroup N such that G/N ∈ X. Then ρX∗(G) ⊆ N.
Proof. If G/N ∈ X, then ρX∗(G/N) = {e}. Suppose that ρX∗(G) is not contained in N. Then
ρX∗(G)/(ρX∗(G) ∩ N) is isomorphic to (ρX∗(G)N)/N, which is a normal nontrivial subgroup
(cid:3)
of G/N that belongs to X∗. This is a contradiction.
Lemma 6.12. Let X be a class of groups that is closed under normal subgroups, extensions,
and contains all finite groups. Assume that G is any group and H is a subgroup of G of finite
index. Then G ∈ X if and only if H ∈ X.
Proof. Let N be the normal core of H in G, that is, the largest normal subgroup of G contained
in H. It is well-known that N also has finite index in G. Suppose first that G ∈ X. Then
N ∈ X and H/N is finite, so H/N ∈ X, and hence H ∈ X. The converse is similar; if H ∈ X,
then N ∈ X and G/N is finite, so G ∈ X.
(cid:3)
7. Amenablish groups and the amenablish radical
Since the class of C∗-simple groups is closed under normal subgroups [7, Theorem 1.4], the
following definition makes sense in light of Proposition 6.3.
Definition 7.1. We call a group amenablish if it has no nontrivial C∗-simple quotients. The
class of all amenablish groups is radical, so every group G has a unique maximal normal
amenablish subgroup, which will be called the amenablish radical of G.
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
19
Then G/T
α∈Λ Nα is C∗-simple.
It is clear that every amenable group is amenablish, but not all amenablish groups are
amenable, as explained in Corollary 7.11 below.
We will now show that the class of C∗-simple groups is residual, which will imply that a
group is C∗-simple precisely when its amenablish radical is trivial. Since the class of C∗-simple
groups is known to be closed under normal subgroups and extensions by [7, Theorem 1.4], we
only have to prove that (i*) from Definition 6.1 holds.
Proposition 7.2. Suppose that G is a group and {Nα}α∈Λ is a family of normal subgroups
of G such that G/Nα is C∗-simple for all α.
∼= Nα/(Nα ∩ Nβ) is a normal
Proof. For any two indicies α and β, the group (NαNβ)/Nβ
subgroup of G/Nβ, so it is C∗-simple again by [7, Theorem 1.4]. Moreover, G/(Nα ∩ Nβ) →
G/Nα is surjective with kernel Nα/(Nα ∩ Nβ). Hence, applying [7, Theorem 1.4] once more
gives that G/(Nα ∩ Nβ) is C∗-simple. We may therefore assume that the family {Nα}α∈Λ is
closed under finite intersections.
It is easy to see, using transfinite induction and the Axiom of Choice, that we can obtain a
well-ordered set {Nβ}β∈I of normal subgroups of G with the property
\
β∈I
Nβ = \
α∈Λ
Nα
def= N.
After factoring the whole family by N, we deduce the following equivalent reformulation of
the above statement: Suppose G is a group with a decreasing (transfinite) sequence
G = N0 (cid:41) N1 (cid:41) N2 (cid:41) ··· (cid:41) Nα (cid:41) ··· (cid:41) Nβ (cid:41) ···
that satisfies
(ii) T
(i) Nα is normal in G for all α ∈ I,
(iii) G/Nα is C∗-simple for all α ∈ I.
α∈I Nα = {e},
Then G is C∗-simple.
To prove the latter statement, set Xα = ∂F (Nα/Nα+1), where ∂F denotes the Furstenberg
boundary. Then Xα is a G-boundary for all α. Indeed, by [7, Lemma 5.2] and the normality
assumption, the action of Nα/Nα+1 on Xα extends uniquely to a boundary action of G/Nα+1
on Xα. Then, composition with the quotient map gives a boundary action of G on Xα
(compare with [7, Proof of Theorem 1.4]).
Now, set X =Q
Next, take an arbitrary point x = (xα) ∈ X and an arbitrary basic open set U =Q
α Xα (with the usual product topology). We wish to show that the action
of G on X is a boundary action, that is, strongly proximal and minimal. First, [27, Lemma 3]
gives that the action of G on X is strongly proximal, since it is strongly proximal on each factor.
α Uα,
where we have Uα = Xα except for finitely many sets Uα1 , . . . , Uαn, with α1 < ··· < αn. Note
that for all α, the set (Nα \ Nα+1)xα is dense in Xα and also that Nα acts minimally on
Xα, while Nα+1 acts trivially on Xα. Therefore there exists a group element g1 ∈ Nα1, such
that g1xα1 ∈ Uα1. Also there exists g2 ∈ Nα2, such that g2g1xα2 ∈ Uα2. We continue the
argument and finally deduce that there exists gn ∈ Nαn with gn ··· g1xαn ∈ Uαn. This shows
that gn ··· g1x ∈ U. We conclude that Gx is dense in X, hence, X is a G-boundary. Note
that the last argument does not require the Axiom of Choice, since we just need the existence
of gi's and not their concrete choice.
We wish to show that this action is free, i.e., that for all nontrivial g ∈ G, the set
X g = {x : gx = x} is empty. So pick an arbitrary g ∈ G \ {e}. From the assumption it follows
that there must be a unique α such that g ∈ Nα \ Nα+1. Clearly, Nα/Nα+1 is C∗-simple, since
20
it is a normal subgroup of G/Nα+1, which is C∗-simple by assumption, so [22, Theorem 6.2]
and [7, Lemma 3.3] imply that Nα/Nα+1 acts freely on ∂F (Nα/Nα+1), that is,
NIKOLAY A. IVANOV AND TRON OMLAND
Hence, X g =Q
α X g
α = ∅.
α = ∂F (Nα/Nα+1)gNα+1 = ∅
X g
(cid:3)
The result now follows from [22, Theorem 6.2].
In Proposition 7.2 the use of the Axiom of Choice can be avoided for groups that are
concretely given, e.g. if G is given by a totally ordered set of generators and relations, or if G
is given by some concrete dynamical properties.
Corollary 7.3. The class of C∗-simple groups is a residual class.
the smallest normal subgroup of G that produces a C∗-simple quotient.
Hence, a group G is C∗-simple if and only if its amenablish radical N is trivial, and N is
We will now describe the amenablish radical in terms of the Furstenberg boundary.
Definition 7.4. Let G be any group. Set N0 = {e} and N1 = int(G (cid:121) ∂F G), recall (8), and
moreover, for every ordinal α, define a normal subgroup Nα+1 of G by
Nα+1/Nα = int(G/Nα (cid:121) ∂F (G/Nα)),
and for every limit ordinal β, set Nβ =S
AH(G) =S
α Nα.
α<β Nα, which is clearly also normal in G. Then
{Nα}α is an ascending normal series of G which eventually stabilizes, and we finally set
Lemma 7.5. For any group G, the quotient G/AH(G) is C∗-simple.
Proof. Let {Nα}α be as in Definition 7.4. Then there exists some ordinal β such that
AH(G) = Nβ. If G/Nβ is not C∗-simple, then Nβ+1/Nβ = (∂F (G/Nβ))gNβ 6= ∅ is nontrivial.
(cid:3)
Hence, Nβ (cid:40) Nβ+1, contradicting the definition of AH(G).
Note that G/ int(G (cid:121) ∂F G) is not necessarily C∗-simple, i.e., AH(G) is in general bigger
than int(G (cid:121) ∂F G). Indeed, it was explained to us by Adrien Le Boudec that by applying
[6, Theorem 1.11], one can construct an example G = G(F, F 0) such that int(G (cid:121) ∂F G) has
index two in G (the condition is that F 0 is generated by its point stabilizers). We refer to
[4, 5, 6] for more about this construction.
Lemma 7.6. Suppose that N is a normal subgroup of a group G such that G/N is C∗-simple.
Then AH(G) ⊆ N.
Proof. The action of G/N on X = ∂F (G/N) is free by [22, Theorem 6.2]. Pick g ∈ int(G (cid:121)
∂F G) such that (∂F G)g 6= ∅. By Lemma 5.7, we have X gN 6= ∅, which means that gN is
trivial in G/N, i.e., g ∈ N. Since the set of all g with (∂F G)g 6= ∅ generates int(G (cid:121) ∂F G),
it follows that int(G (cid:121) ∂F G) ⊆ N.
We continue by transfinite induction. Let {Nα}α be the series from Definition 7.4 associated
with G. We have shown that N1 ⊆ N. Assume next that Nα ⊆ N for some ordinal α and note
that there is a quotient map G/Nα → G/N. Choose g ∈ Nα+1 such that (∂F (G/Nα))gNα 6= ∅.
Then the same argument as above gives that X gN 6= ∅, so g ∈ N. Hence, we conclude that
Nα+1 ⊆ N. Finally, if β is a limit ordinal and Nα ⊆ N for all α < β, then clearly Nβ ⊆ N. (cid:3)
Lemma 7.7. Let N be a normal subgroup of G. Then AH(N) = AH(G) ∩ N.
Proof. Pick g ∈ int(G (cid:121) ∂F G) ∩ N such that (∂F G)g 6= ∅. It follows from Lemma 5.7 that
(∂F N)g 6= ∅, so g ∈ int(N (cid:121) ∂F N). Since the set of all g with (∂F G)g 6= ∅ generates
int(G (cid:121) ∂F G), we get that int(G (cid:121) ∂F G) ∩ N ⊆ int(N (cid:121) ∂F N).
We continue by transfinite induction. Let {Nα}α and {Hα}α be the series from Definition 7.4
associated with G and N, respectively. We have shown that N1 ∩ N ⊆ H1. Let α be an ordinal
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
21
number and assume that Nα ∩ N ⊆ Hα. Note that N/(N ∩ Nα) ∼= (N Nα)/Nα is a normal
subgroup of G/Nα, and that N/Hα is a quotient of N/(N ∩ Nα). Choose g ∈ Nα+1 ∩ N
such that ∂F (G/Nα)gNα 6= ∅. Then by Lemma 5.7 we have (∂F (N/Hα))gHα 6= ∅. Hence,
gHα ∈ int(N/Hα (cid:121) ∂F (N/Hα)) = Hα+1/Hα, and it follows that Nα+1 ∩ N ⊆ Hα+1. Finally,
if β is a limit ordinal and Nα ∩ N ⊆ Hα for all α < β, then clearly Nβ ∩ N ⊆ Hβ. Thus, we
have AH(G) ∩ N ⊆ AH(N).
For the opposite inclusion, note that G/AH(G) is C∗-simple by Lemma 7.5, and that
(AH(G)N)/AH(G) is normal in G/AH(G), so N/(AH(G) ∩ N) ∼= (AH(G)N)/AH(G) is
C∗-simple by using [7, Theorem 1.4]. Hence, Lemma 7.6 gives that AH(N) ⊆ AH(G)∩ N. (cid:3)
Proposition 7.8. For any group G, the amenablish radical of G coincides with AH(G).
Proof. We need to show that AH(G) is amenablish, and that it contains all normal amenablish
subgroups of G.
Set M = int(G (cid:121) ∂F G). Suppose first that L is a normal subgroup of M such that L 6= M.
Pick g ∈ M \ L so that (∂F G)g 6= ∅. It follows from Lemma 5.7 that (∂F (M/L))gL 6= ∅, so
M/L is not C∗-simple. Hence, M is amenablish.
Let {Nα}α be the series from Definition 7.4 associated with G. Then it follows that
Nα+1/Nα is amenablish for every ordinal α, by using the same argument as above with G/Nα
in place of G. Since the class of amenablish groups is radical, it is closed under extensions
and under increasing unions of normal subgroups. Since N1 is amenablish, an argument
by transfinite induction gives that Nα is amenablish for every ordinal α. Hence, AH(G) is
amenablish.
Next, let L be an amenablish normal subgroup of G, and assume that L is not contained in
AH(G). Set K = L ∩ AH(G), then K 6= L and K = AH(L) by Lemma 7.7. Hence, L/K is
(cid:3)
C∗-simple by Lemma 7.5.
Lemma 7.9. The class of amenablish groups is closed under normal subgroups.
Proof. This is a direct consequence of Lemma 6.7, Lemma 7.7, and Proposition 7.8.
(cid:3)
Lemma 7.10. Let G be any group and H a subgroup of finite index. Then G is amenablish
if and only if H is amenablish.
Proof. The class of amenablish groups is closed under normal subgroups, extensions, and
(cid:3)
contains all finite groups. Hence, Lemma 6.12 applies.
Corollary 7.11. The group Γ of Section 4 is amenablish, but not amenable (and has trivial
amenable radical).
Proof. By Theorem 4.4, the group Γ is not C∗-simple, but has the unique trace property, so
it has trivial amenable radical and is icc. The normal subgroup Γ0 from Proposition 4.5 is not
C∗-simple either, because it has finite index in Γ (see [18, Proposition 19 (iv)]). Since Γ0 is
simple and AH(Γ0) 6= {e}, we must have AH(Γ0) = Γ0, that is, Γ0 is amenablish. Hence, it
(cid:3)
follows from Lemma 7.10 that Γ is amenablish.
Remark 7.12. The class of amenablish groups is not closed under subgroups. Indeed, by
Corollary 7.11 the group Γ is amenablish, but it contains as subgroup a nonabelian free group,
which is C∗-simple, and thus not amenablish.
Moreover, AH(Γ) = Γ and N F(Γ) = {e}, while recent work by Olshanskii and Osin [26]
presents a group G with the property AH(G) = {e} and N F(G) = G. Hence, it seems to be
no relation between the class of amenablish groups and N F (except that both contain all
amenable groups).
22
NIKOLAY A. IVANOV AND TRON OMLAND
Remark 7.13. Let G be a countable group, and let N be the subgroup of G generated by all
recurrent amenable subgroups in G (see [23, Definition 5.1]). Does N coincide with AH(G)?
It is mentioned in [23, Remark 5.4] that for every x ∈ ∂F G the subgroup Gx is recurrent
amenable in G, so int(G (cid:121) ∂F G) = hGx : x ∈ ∂F Gi ⊆ N. Moreover, [23, Theorem 1.1] says
that N is trivial if and only if G is C∗-simple, that is, if and only if int(G (cid:121) ∂F G) is trivial.
Note also that a recent paper [6, Section 4] introduces a unique maximal amenable uniformly
recurrent subgroup AG of G, and [6, Proposition 2.21 (ii)] states that hH : H ∈ AGi = int(G (cid:121)
∂F G), which in general is smaller than AH(G), cf. comment after Lemma 7.5.
Example 7.14. If G is a simple group, then G is either C∗-simple or amenablish. E.g.
Thompson's group T is known to be simple, so it follows directly from [6, 16] that T is
amenablish if and only if Thompson's group F is amenable.
Acknowledgements
The second author is funded by the Research Council of Norway through FRINATEK,
project no. 240913. Part of the work of the second author was conducted during a research stay
at Arizona State University, Fall 2015, and another part when attending the research program
"Classification of operator algebras: complexity, rigidity, and dynamics" at the Mittag-Leffler
Institute outside Stockholm, January -- February 2016. In both cases, he would like to thank
the hosts for providing excellent working conditions. The second author is also grateful to
Erik Bédos for useful feedback on an early version of the article, and to Adrien Le Boudec for
helpful e-mail correspondence at a later stage. Finally, the authors would like to thank the
referee for several valuable comments.
Update November 2017
This paper was published in Journal of Functional Analysis in 2017 (272(9):3712 -- 3741).
After that, it was pointed out to us by Rasmus Sylvester Bryder that we had incorrectly
assumed that the boundary of the Bass-Serre tree is always compact, and therefore the
statement of Lemma 5.8 was incorrect. This did not cause any major problems, and the proof
of the only result that depended on it, Theorem 5.9, was easily fixable (it even turned out that
the result can be generalized, see [8, Theorem 3.9]). We still decided to reformulate certain
paragraphs of Section 5 to accomodate for this mistake, by inserting a new proposition and
modify Lemma 5.8 and the proof of Theorem 5.9. At the same time, we also inserted a new
lemma used to clarify the proof of Lemma 5.1, fixed an inaccuracy in the proof of Theorem 3.2,
corrected some typos, and updated the reference list.
References
266(3):279 -- 286, 1984.
[1] Erik Bédos. Operator algebras associated with free products of groups with amalgamation. Math. Ann.,
[2] Erik Bédos and Tron Omland. On reduced twisted group C∗-algebras that are simple and/or have a
[3] Florin P. Boca and Viorel Niţică. Combinatorial properties of groups and simple C∗-algebras with a
unique trace. J. Noncommut. Geom., to appear.
unique trace. J. Operator Theory, 20(1):183 -- 196, 1988.
91(2):253 -- 293, 2016.
[4] Adrien Le Boudec. Groups acting on trees with almost prescribed local action. Comment. Math. Helv.,
[5] Adrien Le Boudec. C∗-simplicity and the amenable radical. Invent. Math., 209(1):159 -- 174, 2017.
[6] Adrien Le Boudec and Nicolás Matte Bon. Subgroup dynamics and C∗-simplicity of groups of homeomor-
[7] Emmanuel Breuillard, Mehrdad Kalantar, Matthew Kennedy, and Narutaka Ozawa. C∗-simplicity and
[8] Rasmus Sylvester Bryder, Nikolay A. Ivanov, and Tron Omland. C∗-simplicity of HNN extensions and
the unique trace property for discrete groups. Publ. Math. Inst. Hautes Études Sci., to appear.
phisms. Ann. Sci. Éc. Norm. Supér. (4), to appear.
groups acting on trees. Preprint arXiv:1711.10442, 2017.
C∗-SIMPLICITY, AMALGAMS, AND RADICAL CLASSES
23
[9] Yves de Cornulier. Infinite conjugacy classes in groups acting on trees. Groups Geom. Dyn., 3(2):267 -- 277,
180, 2004.
[10] Mahlon M. Day. Amenable semigroups. Illinois J. Math., 1:509 -- 544, 1957.
[11] Kenneth J. Dykema. Exactness of reduced amalgamated free product C∗-algebras. Forum Math., 16(2):161 --
[12] Kenneth J. Dykema, Uffe Haagerup, and Mikael Rørdam. The stable rank of some free product C∗-algebras.
[13] Kenneth J. Dykema and Pierre de la Harpe. Some groups whose reduced C∗-algebras have stable rank
Duke Math. J., 91(1):95 -- 121, 1997.
one. J. Math. Pures Appl. (9), 78(6):591 -- 608, 1999.
[14] Shmuel Glasner. Topological dynamics and group theory. Trans. Amer. Math. Soc., 187:327 -- 334, 1974.
[15] Uffe Haagerup. A new look at C∗-simplicity and the unique trace property of a group. In Operator algebras
and applications, vol. 12 of Abel Symp., pp. 161 -- 170. Springer, Heidelberg, 2016.
Funct. Anal., 272(11):4838 -- 4852, 2017.
[16] Uffe Haagerup and Kristian Knudsen Olesen. Non-inner amenability of the Thompson groups T and V . J.
[17] Pierre de la Harpe. Reduced C∗-algebras of discrete groups which are simple with a unique trace. In
Operator algebras and their connections with topology and ergodic theory (Buşteni, 1983), vol. 1132 of
Lecture Notes in Math., pp. 230 -- 253. Springer, Berlin, 1985.
[18] Pierre de la Harpe. On simplicity of reduced C∗-algebras of groups. Bull. Lond. Math. Soc., 39(1):1 -- 26,
extensions, and fundamental groups of 3-manifolds. J. Topol. Anal., 3(4):451 -- 489, 2011.
[19] Pierre de la Harpe and K. Jhabvala. Quelques propriétés des algèbres d'un groupe discontinu d'isométries
hyperboliques. In Ergodic theory (Sem., Les Plans-sur-Bex, 1980), vol. 29 of Monograph. Enseign. Math.,
pp. 47 -- 55. Univ. Genève, Geneva, 1981.
[20] Pierre de la Harpe and Jean-Philippe Préaux. C∗-simple groups: amalgamated free products, HNN
[21] Nikolay A. Ivanov. On the structure of some reduced amalgamated free product C∗-algebras. Internat. J.
[22] Mehrdad Kalantar and Matthew Kennedy. Boundaries of reduced C∗-algebras of discrete groups. J. Reine
[23] Matthew Kennedy. An intrinsic characterization of C∗-simplicity. Preprint arXiv:1509.01870, 2017.
[24] Sang Og Kim. Powers groups and crossed product C∗-algebras. Commun. Korean Math. Soc., 5(2):147 -- 149,
Angew. Math., 727:247 -- 267, 2017.
Math., 22(2):281 -- 306, 2011.
2009.
2007.
1990.
spaces. J. Differential Geom., 67(3):395 -- 455, 2004.
[25] Nicolas Monod and Yehuda Shalom. Cocycle superrigidity and bounded cohomology for negatively curved
[26] Alexander Yu. Olshanskii and Denis V. Osin. C∗-simple groups without free subgroups. Groups Geom.
[27] Narutaka Ozawa. Lecture on the Furstenberg boundary and C∗-simplicity, 2014. http://www.kurims.
[28] Robert T. Powers. Simplicity of the C∗-algebra associated with the free group on two generators. Duke
kyoto-u.ac.jp/~narutaka/notes/yokou2014.pdf.
Dyn., 8(3):933 -- 983, 2014.
Math. J., 42:151 -- 156, 1975.
[29] Derek J. S. Robinson. Finiteness conditions and generalized soluble groups. Part 1, volume 62 of Ergeb.
Math. Grenzgeb. (2). Springer-Verlag, New York-Berlin, 1972.
[30] Jean-Pierre Serre. Arbres, amalgames, SL2, volume 46 of Astérisque. Soc. Math. France, Paris, 1977.
[31] Robin D. Tucker-Drob. Shift-minimal groups, fixed price 1, and the unique trace property. Preprint
arXiv:1211.6395, 2012.
Faculty of Mathematics and Informatics, University of Sofia, blvd. James Bourchier 5, BG-1164
Sofia, Bulgaria
E-mail address: [email protected]
Department of Mathematics, University of Oslo, P.O.Box 1053 Blindern, NO-0316 Oslo, Norway
E-mail address: [email protected]
|
1012.4259 | 3 | 1012 | 2011-09-12T02:02:45 | On commutative, operator amenable subalgebras of finite von Neumann algebras | [
"math.OA",
"math.FA"
] | An open question, raised independently by several authors, asks if a closed amenable subalgebra of ${\mathcal B}({\mathcal H})$ must be similar to an amenable C*-algebra; the question remains open even for singly-generated algebras. In this article we show that any closed, commutative, operator amenable subalgebra of a finite von Neumann algebra ${\mathcal M}$ is similar to a commutative C*-subalgebra of ${\mathcal M}$, with the similarity implemented by an element of ${\mathcal M}$. Our proof makes use of the algebra of measurable operators affiliated to ${\mathcal M}$. | math.OA | math |
On commutative, operator amenable subalgebras
of finite von Neumann algebras
Yemon Choi
31st August 2011
Abstract
An open question, raised independently by several authors, asks if a closed amenable
subalgebra of B(H) must be similar to an amenable C∗-algebra; the question remains open
even for singly-generated algebras. In this article we show that any closed, commutative,
operator amenable subalgebra of a finite von Neumann algebra M is similar to a commutative
C∗-subalgebra of M, with the similarity implemented by an element of M. Our proof makes
use of the algebra of measurable operators affiliated to M.
MSC 2010: 46J40, 47L75 (primary), 47L60 (secondary).
1
Introduction
The notion of amenability for Banach algebras, and its variants for particular classes of Banach
algebra, has been much studied. By the work of several authors, we know that a C∗-algebra is
amenable as a Banach algebra if and only if it is nuclear; one striking feature of this result is
that the definition of amenability does not use any involution on the algebra. In contrast, to
date there is no comparable characterization of amenable, not necessarily self-adjoint, operator
algebras. (Here, and throughout this paper, the term operator algebra will always mean a norm-
closed subalgebra of B(H) for some Hilbert space H; the Hilbert space H is part of the data,
even if we don't mention it explicitly.)
One way to generate examples is as follows. If A is an amenable C∗-algebra, represented faith-
fully on a Hilbert space H, and R ∈ B(H) is invertible, then the algebra R−1AR = {R−1aR : a ∈
A} will be an amenable operator algebra. It has been asked by several authors, in various forms
and in various contexts, if every amenable operator algebra arises in this way. As yet, the
question remains unresolved -- even in the singly generated case! -- but there have been notable
partial results:
-- Gifford showed in his PhD thesis (see also [9]) that any amenable subalgebra of K(H) is
similar inside B(H) to a C∗-subalgebra; this had previously been shown for singly generated
subalgebras of K(H) by Willis [25].
-- Marcoux [11] showed that an amenable, commutative operator algebra which has enough
Hilbert-space representations of a certain form must be similar to a C∗-algebra (and con-
versely, that such representations are easily obtained for any operator algebra which is
similar to a commutative C∗-algebra).
1
Also worth mentioning is an older result of Curtis and Loy [4], showing that any amenable
operator algebra which is generated by its normal elements is automatically self-adjoint, and
hence an (amenable) C∗-algebra. This extends a theorem of Sheinberg [21], who had proved
that every amenable uniform algebra must be self-adjoint. In all these results, it turns out that
one can replace amenability with operator amenability (to be defined below); while this is not
always noted explicitly, it follows easily upon inspection of the proofs.
In this article we make a contribution in similar vein to those mentioned above, but restrict
our attention to commutative, closed subalgebras of finite von Neumann algebras. (Note that
several classical examples of Banach function algebras, such as Lipschitz algebras and certain
algebras of differentiable functions, fall into this class.) Our main result is as follows.
Theorem 1.1. Let M be a finite von Neumann algebra and let A be a closed, commutative
subalgebra. Suppose A is operator amenable with constant ≤ K. Then there exists a positive
invertible R ∈ M, with kRkkR−1k ≤ (1 + 2K)2, such that RAR−1 is a self-adjoint subalgebra
of M. Moreover, if the identity of M lies in the σ(M, M∗)-closure of A, then we can take
kRkkR−1k ≤ K 2.
Since an amenable Banach algebra is operator amenable, we have the following corollary.
Corollary 1.2. If A is a closed, commutative, amenable subalgebra of a finite von Neumann
algebra, then A is similar to an abelian C∗-algebra.
The corollary appears to be new even in the case where A is singly generated, i.e. when we
have T ∈ M which is an amenable operator in the sense of [25].
Ideas behind the proof of Theorem 1.1
The underlying strategy behind the proof of Theorem 1.1 is to find a family of commuting
idempotents in the relative bicommutant M ∩ A′′, linear combinations of which can be used to
approximate the elements of A in a suitable sense; for then we can apply existing techniques
to find a similarity that simultaneously renders all the idempotents self-adjoint. Making this
idea precise, and setting up enough machinery to make it work, will take up most of the present
paper.
Our task would be easier if we had a priori knowledge of some kind of "positivity" in the given
algebra A. We get round this by showing that the Gelfand transform of A must be injective with
dense range (see Corollaries 2.6 and 4.3), so that A may be identified with a dense subalgebra
of some C0(X); we then show that one can, by approximation arguments, realize characteristic
functions of open subsets of X as idempotents in M of the desired form. (The approximation
argument relies on a version of Grothendieck's theorem, whose proof ultimately rests on the
order structure available inside C0(X); so in some sense we are still exploiting positivity, but
not a priori inside A.) A technical complication is that we have to first look outside M for these
us, it turns out that operator amenability of A is such a strong hypothesis that the idempotents
idempotents, in the larger algebra fM of measurable operators affiliated to M; fortunately for
we create in fM will be forced to actually lie in M.
Let us make some additional remarks to put Theorem 1.1 in further context. Firstly, note
that it suffices to prove the theorem in the special case where M has a faithful, normal, finite
trace (which is guaranteed if we know that M is σ-finite and finite). This is because a given finite
2
von Neumann algebra can always be decomposed as an ℓ∞-direct sum of σ-finite ones (see [23,
Corollary V.2.9]), and it is straightforward to check that if Theorem 1.1 is known to hold for
each of these summands, then it will hold for the original von Neumann algebra. Secondly, note
that if we could replace 'finite' with 'semifinite' in Theorem 1.1, this would prove that every
commutative, operator amenable, operator algebra is similar to a C∗-algebra. Such a result,
even if true, seems to lie beyond the reach of the present article's techniques. Nevertheless, we
hope that by using more refined tools one could make improvements.
Overview of the contents
The main part of the proof of Theorem 1.1 will be given in Section 4, once we have established
some preliminary results in Sections 2 and 3. Some of these preliminaries concern calculations
in the algebra of measurable operators affiliated to M: for various technical details concerning
this algebra, our main source is Nelson's article [13]. Otherwise, we have tried to make the
presentation in this article fairly self-contained.
In Section 5, for sake of contrast, we give
examples of commutative semisimple subalgebras of finite von Neumann algebras which are
weakly amenable but non-amenable. Finally, we offer some closing remarks.
2 Preliminaries
Since the audience for this paper may include those who -- like the author -- have more experience
working with Banach-algebraic problems than operator-algebraic ones, we have tried to make
this article accessible to researchers in both areas. This has led us to include material in this
preliminary section which will doubtless be already familiar to specialists.
We start with some standard notation and terminology. Throughout, the dual of a Banach
space E will be denoted by E∗. If H is a Hilbert space and S ⊆ B(H) is an arbitrary subset,
then S′ will denote the commutant of S in B(H). If E and F are vector spaces then E ⊗ F
will denote their algebraic tensor product. The projective tensor product of Banach spaces will
the operator space projective tensor product, see [8, §7].
be denoted by b⊗ and the operator space projective tensor product of operator spaces will be
denoted by b⊗op. For background material on operator spaces, completely bounded maps, and
Throughout this paper, an idempotent in an algebra is merely an element equal to its own
square; the term projection will always be reserved for a self-adjoint idempotent in a C∗-algebra.
2.1 Operator amenability and quantitative variants
Let us briefly give some background to set up what we need in this paper.
The notion of amenability for Banach algebras was introduced by B. E. Johnson in the 1970s,
motivated by problems concerning the L1-convolution algebras of locally compact groups. (For
a short and reader-friendly account of the basic definitions, see [3, §43].) Johnson's original
definition was phrased in terms of derivations; we shall instead use the characterization found
in his later article [10] as a working definition.
Definition 2.1. A Banach algebra A is amenable if and only if there exists M ∈ (Ab⊗ A)∗∗ such
that a · M = M · a and a · π∗∗(M) = κ(a) for all a ∈ A, where the bimodule action on (Ab⊗ A)∗∗
3
where κ : A → A∗∗ is the natural embedding of A in its bidual.
is the natural one induced from Ab⊗ A, the map π : Ab⊗ A → A is defined by π(a ⊗ b) = ab, and
The element M in Definition 2.1 is called a virtual diagonal for A, and the amenability
constant of A is defined to be the infimum of kMk for all possible virtual diagonals M. (By
weak∗-compactness, the infimum is actually attained; and we may speak of A being amenable
with constant ≤ K, meaning that it has a virtual diagonal of norm ≤ K.)
Example 2.2. C(X) is amenable for any compact Hausdorff X (see [3, §43, Theorem 12]).
Furthermore, if A ⊆ C(X) is a closed amenable subalgebra which separates points and contains
the constant functions, then by a theorem of Sheinberg [21], A is self-adjoint and hence all of
C(X) by the Stone-Weierstrass theorem.
Certain natural Banach algebras possess an operator space structure for which the multipli-
cation map is jointly completely contractive -- such objects are said to be completely contractive
Banach algebras. For these, a richer theory can be obtained by using the notion of operator
amenability. This was first formally introduced by Ruan in [15], and has a characterization
analogous to that in Definition 2.1 -- the difference is that we only require a virtual diagonal
M ∈ (Ab⊗op A)∗∗, see [15, Proposition 2.4] or [8, Theorem 16.1.4]. Every amenable Banach alge-
bra is operator amenable (no matter what operator space structure is put on it) but the converse
is not true. We have the analogous notion of A being operator amenable with constant ≤ K.
The precise nature of operator amenability will in fact not concern us much, and we will
avoid overt discussion of operator space techniques, except in the next two results (Lemma 2.3
and Theorem 2.4).
hold:
Lemma 2.3. Let A ⊆ B(H) be an operator algebra that is operator amenable with constant ≤ K
when equipped with the operator space structure induced from B(H). Then there exists a net
i , then the following properties
(∆α) ⊂ A ⊗ A such that, if we write ∆α as a finite sumPi cα
(i) The elementary operator Eα : B(H) → B(H) defined by Eα(T ) = Pi cα
(ii) The elementary operator Fα : B(H) → B(H) defined by Fα(T ) = Pi dα
i has norm
i has norm
i ⊗ dα
i T dα
i T cα
≤ K.
≤ K.
(iii) For each T ∈ B(H) and a ∈ A, kaEα(T ) − Eα(T )ak → 0.
(iv) For each T ∈ B(H) and a ∈ A, kFα(aT − T a)k → 0.
i dα
i , the net (uα) ⊂ A is a bounded approximate identity for A.
(v) Putting uα =Pi cα
Proof of Lemma 2.3. By hypothesis there exists M ∈ (Ab⊗ A)∗∗ with (c.b.) norm ≤ K, satisfy-
[3, §43, Lemma 8]. -- we obtain a net (Mα)α ⊂ Ab⊗op A which satisfies kMαk ≤ K for all α and
ing a · M − M · a = 0 and a · π∗∗(M) = κ(a) for each a ∈ A. As in the proof of [8, Theorem
16.1.4] -- itself patterned on the standard argument for Banach algebras, see [10, Lemma 1.2] or
(aπ(Mα) − a) = 0
for each a ∈ A.
a · Mα − Mαa = lim
α
lim
α
4
Furthermore, since the algebraic tensor product A ⊗ A is dense in Ab⊗op A, and since kMαk is
bounded away from zero for sufficiently large α, we can by truncating and rescaling obtain a net
(∆α) ⊂ A ⊗ A with k∆αk ≤ K for all α and k∆α − Mαk → 0; clearly
lim
α
a · ∆α − ∆αa = lim
α
(aπ(∆α) − a) = 0
for each a ∈ A.
(2.1)
([8, Proposition 7.1.4]). Also:
it is easily checked that for each T ∈ B(H), the bilinear map
A × A → B(H) defined by (a, b) 7→ aT b is completely bounded with norm ≤ kT k, and hence
The 'flip map' σ : Ab⊗op A → Ab⊗op A defined by a ⊗ b 7→ b ⊗ a is a complete isometry
([8, Proposition 7.1.2]) induces a completely contractive linear map Ab⊗op A → B(H), also of
norm ≤ kT k. Therefore, we have a contractive linear map Φ : Ab⊗op A → B(B(H)) which satisfies
Φ(a ⊗ b)(T ) = aT b for all a, b ∈ A and T ∈ B(H). We put Eα = Φ(∆α) and Fα = Φ(σ(∆α));
both are elementary operators with (c.b.) norm ≤ K, and this proves (i) and (ii).
Finally: for a ∈ A and T ∈ B(H) we have
aEα(T ) − Eα(T )a = Φ(a · ∆α − ∆α · a)(T ) and Fα(aT − T a) = Φ(σ(a · ∆α − ∆α · a)(T )
Using (2.1) we therefore get (iii) and (iv), and (v) also follows.
The next result we shall need is a slightly more precise version of a theorem of Runde,
generalizing Sheinberg's theorem to the setting of completely contractive Banach algebras.
Theorem 2.4 (Runde). Let X be a locally compact Hausdorff space, and let B be a closed
subalgebra of C0(X), which separates points and which is operator amenable when given its
minimal operator space structure. Then B = C0(X).
The result follows from the arguments given on [18, p. 634], and we omit the details. One
note of caution: the argument there takes it as read that the unitization of an operator amenable
algebra is itself operator amenable; this is true, and standard folklore, but locating an explicit
statement of this somewhat tricky. (Here is a sketch of a proof: if A is operator amenable with
constant ≤ K, then as noted in the proof of Lemma 2.3, there is a net (Mα)α∈Λ ⊂ Ab⊗op A
which satisfies kMαk ≤ K for all α and
a · Mα − Mαa = 0 and
aπ(Mα) = a
for each a ∈ A.
lim
α
lim
α
Let A# denote the forced unitization of A, with 1 denoting the adjoined unit, and define
∆α = (1 − π(Mα)) ⊗(1 − π(Mα)) + Mα + Mα · (1 − π(Mα)) ∈ A#b⊗op A# .
Straightforward calculations now show that, by taking any cluster point of the net (∆α) in in
the second dual of A#b⊗op A#, we obtain an operator virtual diagonal for A#.)
Remark 2.5. The initial part of the proof in [18] can be sketched as follows: starting with an
operator amenable algebra B, construct a certain Hilbertian representation B → B(H); take a
suitable (i.e. non-degenerate) B-invariant subspace V ⊆ H; and then use operator amenability
to construct a B-module map T : H → V satisfying T (v) = v for all v ∈ V. In the context of
Lemma 2.3, such a map T arises as a weak∗-limit of the net (Eα(P )), where P is the orthogonal
projection of H onto V. We will see a slightly more complicated version of the same idea later
on, in the proof of Proposition 3.5.
5
Amenability of a Banach algebra A implies the amenability of θ(A) whenever θ : A → B is a
continuous homomorphism to another Banach algebra B. As observed in [16, Proposition 2.2],
the same remains true if A is operator amenable and θ is completely bounded (with respect to
given operator space structures on A and B). Now, given an operator space V , any bounded
linear map from the underlying Banach space of V to a commutative C∗-algebra is automatically
completely bounded, when the latter is given its minimal operator space structure: see [8,
Proposition 2.2.6]. We therefore have the following consequence of Theorem 2.4, which will be
needed later.
Corollary 2.6. Let A be a commutative Banach algebra, with non-empty character space X
(equipped with the Gelfand topology), and let G : A → C0(X) be the Gelfand transform. If A
is operator amenable (when equipped with some operator space structure), then G(A) is dense
in C0(X).
2.2 The algebra fM and closed densely defined operators
We start with some notation. Given a von Neumann algebra M which has a faithful normal
tracial state τ , define kxkL1(τ )
def= τ (x) for x ∈ M. Then (see [23, §V.2, p. 320])
kxkL1(τ ) = sup{τ (xy) : y ∈ M, kyk ≤ 1},
(2.2)
showing that k · kL1(τ ) is a norm on M; the corresponding completion of M will be denoted by
L1(τ ). We write ı1 for the canonical embedding M → L1(τ ); this is continuous as a linear map
between Banach spaces.
Note that by Equation (2.2) and the faithfulness of τ , the map M → M∗ which is defined
by y 7→ y · τ extends to an isometric isomorphism from L1(τ ) onto M∗ ([23, Theorem V.2.18]).
For our proof of Theorem 1.1, we shall find ourselves wanting to say that certain elements of
L1(τ ) are idempotents; and to do this, we need to embed L1(τ ) into some ring. Now in the case
of a commutative von Neumann algebra, say L∞[0, 1] with Lebesgue measure as a tracial normal
state, a convenient way to do this is to identify the completion of L∞[0, 1] in the L1-norm with
a space of (equivalence classes of) measurable functions on [0, 1]. In our setting, where M need
by Nelson in [13]. (See also the start of [20, Section 2] for an excellent precis of those parts of
[13] which are most relevant to the present article.)
not be commutative, we use the algebra fM of 'measurable operators' affiliated to M, introduced
For our purposes, fM can be treated as a 'black box', which has enough good properties that
we can perform various algebraic calculations and approximation arguments inside it. Let us
briefly sketch how it is defined. For each ε, δ > 0 we define N (ε, δ) to be the set of all x ∈ M
for which there exists a projection p ∈ M that satisfies τ (1 − p) ≤ δ and kxpk ≤ ε; then the
family of all such N (ε, δ) may be taken as the neighbourhood base at 0 of a translation-invariant
if M is represented faithfully and non-degenerately on a Hilbert space H, then we
Similarly:
may define O(ε, δ) to be the set of all ξ ∈ H for which there exists a projection p ∈ M that
satisfies τ (1 − p) ≤ δ and kpξk ≤ ε, and use this as the neighbourhood base at 0 for a translation
topology on M; and fM is defined to be the completion of M with respect to this topology.
invariant topology on H; the completion of H in this topology will be denoted by eH.
We now summarize the properties of fM and eH which we need in Theorems 2.7 and 2.9
below.
6
Theorem 2.7. Let M be a von Neumann algebra equipped with a faithful normal tracial state τ ,
and fix a Hilbert space H on which M is faithfully and non-degenerately represented. Let fM
and eH be defined as above, let ı : M → fM denote the canonical (linear, continuous) map, and
denote the canonical (linear, continuous) map H → eH by ξ 7→ ξ♯.
(i) The multiplication map M × M → M extends to a separately continuous bilinear map
Moreover, the bilinear map M× H → H given by (T, ξ) 7→ T ξ extends to a separately continuous
(iii) ı(T )[ξ♯] = (T ξ)♯ for all T ∈ M and ξ ∈ H;
fM × fM → fM, making fM into a metrizable semitopological algebra.
(ii) The map ı : M → fM and the map H → eH, ξ 7→ ξ♯ are both injective with dense range.
bilinear map fM × eH → eH; denoting this extended map by (x, η) 7→ x[η], we have:
(iv) x[y[η]] = xy[η] for all x, y ∈ fM and η ∈ eH.
(v) there is a continuous, injective, linear map ı0 : L1(τ ) → fM such that ı = ı0ı1.
Proof. These all follow from the results of [13]. In more detail: part (i), and the claim that the
action of M on H is separately continuous in the measure topologies of M and H, follow from
[13, Theorem 1]. Part (ii) follows from [13, Theorem 2]; parts (iii) and (iv) follow because the
map (x, η) 7→ x[η] is an extension by continuity of the given action of M on H. Finally, part (v)
is the case p = 1 of [13, Theorem 5].
Finally,
Definition 2.8. For each t ∈ fM, define
D(Mt)
def
= {ξ ∈ H : t[ξ♯] ∈ H♯} ⊆ H;
and for each ξ ∈ D(Mt), define Mtξ to be the unique vector in H satisfying (Mtξ)♯ = t[ξ♯]
It is straightforward to
(uniqueness following from injectivity of the map H → eH, η 7→ η♯).
check that the function Mt : D(Mt) → H is linear.
Recall that a closed densely-defined operator T on H, with domain D(T ), is said to be
affiliated to M when, for each unitary u ∈ M′, we have u(D(T )) = D(T ) and T u(ξ) = uT (ξ)
for all ξ ∈ D(T ).
Theorem 2.9. D(Mt) is a dense linear subspace of H, and Mt : D(Mt) → H is a closed linear
operator on H, affiliated to M. Moreover the domain of Mt is D(Mt) and its range is
Ran(Mt)
def
= Mt(D(Mt)) = {Mt(ξ) : ξ ∈ D(Mt)}.
Proof. See [13, Theorem 4].
The following is really just an observation, but has been isolated as a lemma for convenient
reference.
7
Lemma 2.10. Let S ∈ M, t ∈ fM, and suppose ı(S)t = tı(S). Then S(D(Mt)) ⊆ D(Mt) and
S(Ran(Mt)) ⊆ Ran(Mt). Moreover, MtS(ξ) = SMt(ξ) for each ξ ∈ D(Mt).
Proof. Let ξ ∈ D(Mt). Then
t[(Sξ)♯] = t[ı(S)[ξ♯]]
= tı(S)[ξ♯]
= ı(S)t[ξ♯]
= ı(S)[(Mtξ)♯]
= (SMtξ)♯
(by Theorem 2.7(iii))
(by Theorem 2.7(iv))
(by the definition of Mt)
(by Theorem 2.7(iii)).
This shows that S(D(Mt)) ⊆ D(Mt), and (by definition of Mt) that MtS agrees with SMt on
D(Mt). The rest follows.
Of particular importance in the proof of Theorem 1.1 is the ability to manipulate idempotents
in fM. The following lemma is essentially taken from parts of [20, Section 2].
Lemma 2.11 (cf. [20, Proposition 2.4]). Let e ∈ fM be an idempotent. Then Ran(Me) is a closed
subspace of H, satisfying
Ran(Me) = {ξ ∈ H : e[ξ♯] = ξ♯} ⊆ D(Me)
(2.3)
so that Meη = η for all η ∈ Ran(Me). Moreover, the orthogonal projection onto Ran(Me) lies
in M.
Proof. Let e ∈ fM be idempotent. Firstly, if η ∈ D(Me) then Meη ∈ H and
e[(Meη)♯] = e[e[η♯]] = e[η♯] = (Meη)♯
(∗)
thus Ran(Me) ⊆ {ξ ∈ H : e[ξ♯] = ξ♯} ⊆ D(Me). We also have {ξ ∈ H : e[ξ♯] = ξ♯} ⊆
Me(D(Me)) = Ran(Me) and thus (2.3) holds. It also follows from (∗) that MeMeξ = Meξ for
all ξ ∈ D(e).
The fact that Ran(Me) is closed is a special case of [20, Proposition 2.4]. We give a slightly
different argument. Let (ξn) ⊂ D(e) be a sequence such that (Meξn) converges to some η ∈ H.
Then e[ξn
♯] → η♯ in eH; but since the action of fM on eH is separately continuous as a bilinear
map fM × eH → eH, we have
lim
n
♯]] = e[lim
n
Hence η♯ = e[η♯], so by (2.3) η ∈ Ran(Me) as required.
♯] = lim
n
e[e[ξn
e[ξn
e[ξn
♯]] = e[η♯].
Finally, let p be the orthogonal projection from H onto Ran(Me). We wish to show p ∈ M;
while this seems to be well known, I was unable to find a precise reference. The following
argument was communicated to me by M. Argerami ([1]).
It suffices (by the bicommutant
theorem) to show that pu = up for all u ∈ M′. But since Me is affiliated to M, for each such u
we have u(D(Me)) = D(Me) and uMe = Meu on D(Me). Therefore
u(Ran(Me)) = uMe(D(Me)) = Meu(D(Me)) = Me(D(Me)) = Ran(Me) ,
showing that pup = up in B(H). Repeating this argument with u replaced by u∗ gives pu∗p = u∗p
in B(H), and taking adjoints we get pup = pu, so that up = upu = pu as required.
8
3 An automatic boundedness result for certain idempotents
In our eventual proof of Theorem 1.1, a key idea will be the following result.
Definition 3.1 (cf. [11, Lemma 3.7]). Let B be an algebra and let F be a set of mutually
commuting idempotents in B. We say that F is closed under symmetric differences if E + F −
2EF ∈ F for every E, F ∈ F.
Lemma 3.2 (Unitarization lemma). Let H be a Hilbert space, and let F ⊆ B(H) be a set
of mutually commuting idempotents which is closed under symmetric differences. Suppose that
k2e−Ik ≤ C for all e ∈ F. Then there exists R ∈ B(H), positive and invertible with kRkkR−1k ≤
C 2, such that ReR−1 is self-adjoint for each e ∈ F. Moreover, R lies in the von Neumann algebra
generated by I and F.
Lemma 3.2 is by no means a new observation, but it semems hard to pin down a precise
citation. Without the constants and the statement about von Neumann algebras, it can be
found on [7, pp. 222 -- 223]; there, Dixmier remarks that the result is related to previous work of
Lorch (see also the remarks in [25]). A quantitative version can be found in [11, Lemma 3.8],
but once again there is no mention that R lies in a certain von Neumann algebra. Therefore,
we shall prove Lemma 3.2 from scratch, using a more general result.
Theorem 3.3 (Day; Dixmier). Let G be a locally compact amenable group, and let π : G →
B(H) be a uniformly bounded representation of G on some Hilbert space H. Then there exists
a positive, invertible operator R in the von Neumann algebra generated by π(G), satisfying
kRkkR−1k ≤ (supx kπ(x)k)2, such that Rπ(x)R−1 is unitary for each x ∈ G.
For sake of completeness we provide a full proof, using a clasical averaging argument which
is independently due to Day [5, Theorem 8] and Dixmier (ibid.). Our approach follows Dixmier's
closely, and is guided by an outline given in [14].
Proof. Let C
define fξ : G → C by fξ(g) = kπ(g)ξk2; then fξ ∈ L∞(G), with
def
= supg∈G kπ(g)k, and fix a left-invariant mean Λ on L∞(G). For each ξ ∈ H
C −2kξk2 ≤ fξ(g) ≤ C 2kξk
for all g ∈ G.
Define F (ξ) def= Λ(fξ); then since Λ is positive, F is a positive quadratic form on H satisfying
C −2kξk2 ≤ F (ξ) ≤ C 2kξk
for all ξ ∈ H.
(3.1)
and hence (Riesz -- Fischer plus polarization) there exists a unique positive linear operator T :
H → H such that F (ξ) = hT ξ, ξi for all ξ ∈ H. By (3.1), T is invertible with kT k ≤ C 2 and
kT −1k ≤ C 2. Moreover, left-invariance of Λ, together with the fact that h · fξ = fπ(h)ξ for all
h ∈ G and ξ ∈ H, yields
hT π(h)ξ, π(h)ξi = Λ(fπ(h)ξ) = Λ(fξ) = hT ξ, ξi
for all ξ ∈ H.
(3.2)
We therefore take R = T −1/2; clearly it satisfies the required norm bounds, and it follows easily
from (3.2) that Rπ(h)R−1 is unitary for each h ∈ G. Finally, if U ∈ π(G)′ is a unitary operator
commuting with all π(g), then for each ξ ∈ H we have fU ξ = fξ and hence hU ∗T U ξ, ξi =
F (U ξ) = F (ξ) = hT ξ, ξi. Therefore U ∗T U = T and so U ∗RU = R, implying that R commutes
with every unitary in π(G)′, and hence by the double commutant theorem that R lies in the von
Neumann algebra generated by π(G).
9
Lemma 3.2 now follows from Theorem 3.3 by taking G = {I −2E : E ∈ F}; for the conditions
on F ensure that G is a bounded commutative subgroup of B(H), and if R(I −2E)R−1 is unitary
then RER−1 must be self-adjoint.
Remark 3.4. Neither of the articles [5, 7] state explicitly that the operator R implementing the
similarity can be taken to lie in the von Neumann algebra generated by π(G). The observation
may well be folklore: it is explicitly mentioned in the introductory parts of [14].
In view of Lemma 3.2, we have a strategy for showing that a given commutative operator
algebra is similar to a self-adjoint one: try to show that its WOT-closure contains a WOT-
dense, bounded set of idempotents that is closed under symmetric differences, and then apply
Lemma 3.2 (note that the idempotents need not lie in the original algebra). The problem is that
in our setting, it is not clear how to produce such idempotents directly; we shall instead first
that in fact they must lie in ı(M), using the following technical result.
Proposition 3.5. Let A be a subalgebra of M which is operator amenable with constant ≤ K.
produce a suitable family of commuting idempotents in the larger algebra fM, and then show
Let e ∈ fM be an idempotent, and suppose there is a sequence (bn) in the centre of A such that
ı(bn) → e in fM. Then kMe(ξ)k ≤ Kkξk for all ξ ∈ D(Me), and we may therefore identify e
with an idempotent in M of norm at most K. Moreover, if I lies in the WOT-closure of A,
then k2e − Ik ≤ K.
Proposition 3.5 is based on an argument of Gifford ([9, Lemmas 1.5 and 4.4]), and the core
idea in our proof is the same one underlying his result. However, new complications arise since
we are dealing with potentially unbounded operators, which are only densely defined; we have
therefore chosen a slightly different approach to Gifford's.
Proof of Proposition 3.5. Since multiplication in fM is separately continuous,
ı(bna) = eı(a)
for all a ∈ A.
ı(a)e = lim
n
ı(abn) = lim
n
Hence by Lemma 2.10, both D(Me) and Ran(Me) are A-invariant linear subspaces of H. More-
over, by Lemma 2.11, Ran(Me) is closed in H, and the orthogonal projection P from H onto
Ran(Me) is an element of M.
We now use Lemma 2.3. Let Eα be the elementary operator described there, and define
def= Eα(P ) ∈ M. The net (Qα) is bounded and hence has a σ(M, M∗)-cluster point Q ∈ M,
i = Eα(I), and recall (Lemma 2.3(v)) that (uα) is a BAI for
i dα
Qα
with kQk ≤ K. Put uα
A of norm ≤ K.
def= Pi cα
In the case where I is in the WOT-closure of A, we observe that (uα) converges σ(M, M∗)
to I. (Let e0 be any σ(M, M∗)-cluster point in M of the net (uα)α; then ae0 = a for all a ∈ A,
and since there is by assumption a net (ai) ⊂ A WOT-converging to I, for each ξ, η ∈ H we
have
he0ξ, ηi = lim
i
haie0ξ, ηi = lim
i
haiξ, ηi = hξ, ηi
so that e0 = I.) Thus, in this case, Eα(2P − I) → 2Q − I in the σ(M, M∗) topology, yielding
the improved estimate k2Q − Ik ≤ Kk2P − Ik = K.
Lemma 2.3.
Now Q ∈ A′, since for each b ∈ A we have kbQα − Qαbk = kbEα(P ) − Eα(P )bk → 0 by
It follows by continuity of multiplication in fM and the assumption on e that
10
ı(Q)e = eı(Q). By Lemma 2.10,
Q(D(Me)) ⊆ D(Me)
and MeQ(ξ) = Q(Meξ) for all ξ ∈ D(Me).
(3.3)
We now make the following claims:
(i) MeQξ = Qξ for all ξ ∈ D(Me);
(ii) Q(Meξ) = Meξ for all ξ ∈ D(Me).
Combining (i) and (ii) with (3.3) gives Meξ = QMeξ = MeQξ = Qξ for each ξ ∈ D(Me), which
would imply the desired result (since D(Me) is dense in H, by Theorem 2.9, and kQξk ≤ Kkξk).
It therefore merely remains to prove (i) and (ii).
Proof of (i). Since P (H) = Ran(Me) is A-invariant, P aP = aP for all a ∈ A. Hence
P Qα =Xi
P cα
i P dα
i =Xi
cα
i P dα
i = Qα
for all α,
which implies P Q = Q by σ(M, M∗)-continuity. On the other hand, since P (H) = Ran(Me),
Lemma 2.11 implies that MeP = P , so that MeQ = MeP Q = P Q = Q as required.
Proof of (ii). We have P Me(ξ) = Me(ξ) for all ξ ∈ D(Me). If b ∈ A then b(Ran(Me)) ⊆ Ran(Me),
and so P b and b agree on Ran(Me). Therefore, for each η ∈ Ran(Me),
Qα(η) =Xi
cα
i P dα
i (η) =Xi
cα
i dα
i (η) = uα(η) ∈ H for all α.
(3.4)
Let e0 ∈ M be a σ(M, M∗)-cluster point of the bounded net (uα); this is also a cluster point
in the WOT of B(H), hence Q(η) = e0(η) by (3.4). But since (uα) is an approximate identity
for A we also have e0bn = bn for all n, which by (separate) continuity of multiplication in fM
implies that ı(e0)e = e. Therefore e0Me agrees with Me on D(Me), and so for each ξ ∈ D(Me)
we have Meξ = e0Meξ = QMeξ as required.
4 The main proof
Throughout this section M denotes a von Neumann algebra, faithfully and non-degenerately
represented on a Hilbert space H, and equipped with a faithful, finite, normal tracial state τ .
We fix a norm-closed subalgebra A ⊆ M which is operator amenable with constant ≤ K, but
for the moment is not assumed to be commutative. Given x ∈ M, we write ρ(x) for its spectral
radius.
Our starting point for the proof of Theorem 1.1 is the following result, which is the key place
where we require τ to be finite rather than merely semifinite.
Proposition 4.1. kakL1(τ ) ≤ Kρ(a) for each a in the centre of A.
The proof of Proposition 4.1 is inspired by arguments from [25, p. 242], which in effect
prove the following result: if a nonzero compact operator on Hilbert space generates an amenable
operator algebra, it must have a non-zero eigenvalue. As in Willis' proof, we exploit the presence
of a trace and the submultiplicativity of spectral radius on commutative algebras, but replace
his use of Lidskii's trace theorem by the bound given in the following lemma.
11
Lemma 4.2. Let B be a unital C∗-algebra and let τ : B → C be a finite trace on B. Then
τ (b) ≤ kτ kρ(b) for each b ∈ B.
Lemma 4.2 follows immediately from the fact (see [12, Exercise 2.6]) that in any unital
C∗-algebra B we have ρ(b) = inf{kt−1btk : t ∈ B, t > 0}.
Proof of Proposition 4.1. Let a be in the centre of A and let y ∈ M. With the definitions of
Lemma 2.3, put zα
i . The net (zα) is bounded in norm by Kkyk, and so
has a σ(M, M∗)-cluster point, z say, which satisfies kzk ≤ Kkyk. Since a is central in A,
i ycα
def= Fα(y) =Pi dα
dα
i yacα
i = Fα(ya),
(4.1)
azα =Xi
dα
i aycα
i = Fα(ay) and zαa =Xi
and so kazα − zαak → 0 by Lemma 2.3. By continuity az = za, and therefore
ρ(az) ≤ ρ(a)ρ(z) ≤ Kρ(a)kyk.
(4.2)
On the other hand, since τ is a trace, using (4.1) gives
τ (azα) =Xi
τ (dα
i aycα
i ) =Xi
τ (cα
i dα
i ay) = τ (uαay)
and it follows (since τ · a ∈ M∗ and (uα) is a BAI for A) that
τ (az) = lim
α
τ (azα) = lim
α
τ (auαy) = τ (ay) .
Combining this with (4.2) and Lemma 4.2, we obtain
τ (ay) = τ (az) ≤ ρ(az) ≤ Kρ(a)kyk.
In particular, τ (a) ≤ Kρ(a) as required.
Since τ is faithful, Proposition 4.1 has the following corollary, which we would like to high-
light.
Corollary 4.3. If B is an (operator) amenable closed subalgebra of a finite von Neumann algebra,
then the centre of B is semisimple.
For the remainder of this section, we shall make the extra assumption that A is commutative
(as in the statement of Theorem 1.1) and non-zero. Recall that our goal is to prove the following
Theorem (Theorem 1.1, reprise.). Under the assumptions made on M and A, there exists a
positive invertible operator R ∈ M such that RAR−1 is a self-adjoint subalgebra of M.
Proof of the theorem. Let X be the character space of A, equipped with the Gelfand topology,
Gelfand theory for non-unital, commutative Banach algebras.) By Corollary 4.3, A is semisimple,
so X is non-empty and G is a norm-decreasing, injective algebra homomorphism.
and let G : A → C0(X) denote the Gelfand transform a 7→ba. (See [3, §17] for a basic account of
ρ(a) = kbak∞ for all a ∈ A, there is a unique continuous linear map θ : C0(X) → L1(τ ), of norm
By Corollary 2.6, G(A) is dense in C0(X). Hence, by Proposition 4.1 and the fact that
≤ K, which makes the diagram in Figure 1 commute.
12
A ⊂
✲ M
G
❄
C0(X)
ı1
❄
✲ L1(τ )
θ
Figure 1: Extending to C0(X)
Let S be the ring of subsets of X generated by the open sets (so we allow complements, finite
unions, and finite intersections). We write χU for the indicator function of U ∈ S and define B
to be the finite linear span in CX of {χU : U ∈ S}; then take eB to be the closure of B in the
uniform norm on X. It is straightforward to check that C0(X) ⊆eB. We shall now extend θ to
a continuous linear map eθ :eB → L1(τ ); the images of characteristic functions under ı0eθ will be
idempotents in fM, and with a little book-keeping we can then apply our version of Gifford's
We obtain this extension of θ using a result of Tomczak-Jaegermann, which can be thought
of as an intermediate stage between (one version of) the classical Grothendieck inequality and
the later, noncommutative versions of Pisier and Haagerup.
argument (Proposition 3.5).
Theorem 4.4 (Tomczak-Jaegermann). Let M be a von Neumann algebra, let Ω be a locally
compact Hausdorff space, and let T : C0(Ω) → M∗ be a bounded linear map. Then there exists
a finite, positive, regular Borel measure µ on Ω, such that
kT (f )k ≤(cid:18)ZΩ
f (x)2 dµ(x)(cid:19)1/2
for all f ∈ C0(Ω).
(4.3)
Proof. First suppose that Ω is compact. Then the result is given by [24, Theorem 4.2]. In slightly
more detail: it is shown in [24] that the predual of any von Neumann algebra is a Banach space
of cotype 2; hence by a variant of Grothendieck's theorem, the map T is 2-summing, and the
existence of a measure µ with the stated properties now follows from the Pietsch domination
theorem. For further details see [6, Corollary 2.15 and Theorem 11.14].
In the general case, where Ω need not be compact, the simplest approach is to observe that
T : C0(Ω) → M∗ can always be extended continuously to T ♯ : C(Ω♯) → M∗, where Ω♯ is the
one-point compactification of Ω; the result for the compact case yields a (finite, positive, regular)
measure on Ω♯, whose restriction to Ω has the required properties. (Alternatively, one could
adapt the proofs of the results used in the compact case to the locally compact setting, but this
is rather time-consuming.)
By Theorem 4.4 there exists a positive finite regular Borel measure µ on X, such that
kθ(f )kL1(τ ) ≤ kf kL2(µ)
for all f ∈ C0(X).
(4.4)
lemma is a slightly more precise version of the following assertion: there is a continuous linear
(Note that, a priori , k · kL2(µ) might only be a seminorm on eB, rather than a norm.) Our next
map eθ :eB → L1(τ ) which makes the diagram in Figure 2 commute.
13
ı
ı0
✲
✲ fM
A ⊂
✲ M
G
❄
C0(X)
❄
eB
ı1
❄
✲ L1(τ )
✲
θ
eθ
Figure 2: A further extension
Lemma 4.5. There exists a unique linear map eθ :eB → L1(τ ) with the following properties:
(i) eθ(f ) = θ(f ) for all f ∈ C0(X);
(ii) eθ takes k · kL2(µ)-Cauchy sequences in eB to Cauchy sequences in L1(τ ).
Moreover, ı0eθ :eB → fM is a homomorphism.
Proof. Clearly the restriction of µ to X is finite, Borel and regular, so by Lusin's theorem (see
e.g. [17, Theorem 2.4]) the canonical map C0(X) → L2(X, µ) has dense range. In particular,
(†)
easily verified.
for each h ∈eB there exists a sequence (fn) ⊂ C0(X) such that kh − fnkL2(µ) → 0.
(We can bypass the explicit use of Lusin's theorem to prove (†), at the expense of a slightly
longer argument; see Remark 4.6.) Now put eθ(h) = limn θ(fn), the limit existing by (4.4). It is
routine to check that eθ is well-defined, linear, and satisfies (i) and (ii); uniqueness can also be
It remains to show that ı0eθ : eB → fM is a homomorphism. We shall proceed in some
detail. Let g, h ∈eB and let (an), (bm) be sequences in A such that limn kcan − gkL2(µ) = 0 and
limm kcbm − hkL2(µ) = 0. Then from (4.4) we have
kı1(an) −eθ(g)kL1(τ ) = 0 and
kı1(bm) −eθ(h)kL1(τ ) = 0,
lim
m
lim
n
(4.5)
keθ(cancbm −canh)kL1(τ ) = 0 for each n.
(4.6)
(by (4.6))
Therefore, inside L1(τ ) we have
lim
n
lim
m
keθ(canh − gh)kL1(τ ) = 0 and
m eθ(cancbm)
n eθ(canh) = lim
m eθ([anbm) = lim
= lim
n
lim
n
n
eθ(gh) = lim
lim
lim
m
ı1(anbm)
(since eθG = ı)
so that, by continuity of ı0, we have ı0eθ(gh) = limn limm ı(anbm). On the other hand
where in each equality, we used (4.5) and separate continuity of multiplication in fM. Since
ı : A → fM is a homomorphism we are done.
[ı(an)ı0eθ(h)] = lim
ı0eθ(g)ı0eθ(h) = lim
ı(an)ı(bm)
lim
m
n
n
14
Remark 4.6. The claim in (†) can be proved directly by using some of the ideas that go into
proving Lusin's theorem. For the reader's convenience we sketch an argument for doing this.
Since convergence in the uniform norm implies k · kL2(µ)-convergence, it suffices to prove that
(†) holds for each h ∈ B; then, by linearity, we can reduce further to the case where h = χU for
some open U ⊆ X. Fixing an open subset U ⊆ X, regularity of µ (as a measure on X) implies
we can find an increasing sequence K1 ⊆ K2 ⊆ · · · ⊆ U of compact subsets with µ(Kn) ր µ(U ).
For each n, Urysohn's lemma ensures there exists hn ∈ C0(X) with 0 ≤ hn ≤ 1, hn(x) = 0 for
all x ∈ X \ U , and hn(x) = 1 for all x ∈ Kn; it follows that khn − χU kL2(µ) ≤ µ(U \ Kn)1/2 → 0,
and so (†) is proved.
Proposition 4.7. There exists a commuting family (eU )U ∈S of idempotents in M, such that
eθ(χU ) = ı1(eU ) for all U ∈ S. We have supU keU k ≤ K, and if I is in the WOT-closure of A
we have supU k2eU − Ik ≤ K. Moreover, eU eV = eU ∩V for all U, V ∈ S.
Consequently, there exists R ∈ M, positive and invertible with kRkkR−1k ≤ (1 + 2K)2,
such that ReU R−1 is self-adjoint for each U ∈ S. If I is in the WOT-closure of A we have
kRkkR−1k ≤ K 2.
Proof. Let U ∈ S. By Lemma 4.5, ı0eθ is a homomorphism, and so ı0eθ(χU ) is an idempotent in
fM. By using the claim (†) from the proof of Lemma 4.5, and the density of G(A) in C0(X), we
can extract a sequence (bn) ⊂ A such that
ı0eθ(χU ) = ı0(lim
n eθ(bbn)) = lim
n
ı0eθ(bbn) = lim
n
ı(bn).
keU k ≤ K, and if I is in the WOT-closure of A we have k2eU − Ik ≤ K. Finally, if U, V ∈ S
Hence by Proposition 3.5 there is a unique idempotent eU ∈ M with ı(eU ) = ı0eθ(χU ); moreover,
then ı(eU eV ) = ı0eθ(χU )ı0eθ(χV ) = ı0eθ(χU ∩V ), and thus eU eV = eU ∩V = eV eU .
This last identity implies that the family (eU )U ∈S is closed under symmetric differences, and
therefore by applying our unitarization lemma (Lemma 3.2), we see that there exists R ∈ M,
positive and invertible, such that ReU R−1 is self-adjoint for each U ∈ S and kRkkR−1k ≤
supU k2eU − Ik2 ≤ (1 + 2K)2. As observed earlier in the proof, if I lies in the WOT-closure of
A then this can be improved to K 2.
Proposition 4.8. There exists a (norm-)continuous algebra homomorphism φ : eB → M such
that ıφ = ı0eθ. For each h ∈eB, we have kRφ(h)R−1k ≤ khk∞ and
(Rφ(h)R−1)∗ = Rφ(h)R−1 .
(4.7)
Proof of Proposition 4.8. We shall initially define φ on the dense subalgebra B and then show
it can be extended by continuity.
Thus, given f ∈ B = lin{χU : U ∈ S}, we know by Proposition 4.7 that ı0eθ(f ) ∈ ı(M). Let
φ(f ) be the unique element of M satisfying ı(φ(f )) = ı0eθ(f ). It is easily checked that φ : B → M
is an algebra homomorphism, since ı0eθ is an algebra homomorphism and ı is injective.
Given c1, . . . , cn ∈ C and pairwise disjoint U1, . . . , Un ∈ S, we have
Rφ nXi=1
ciχUi! R−1 =
ciReUiR−1,
nXi=1
(∗)
15
and since ReU R−1 is self-adjoint for each U ∈ S, it follows from Equation (∗) that
(Rφ(f )R−1)∗ = Rφ(f )R−1
for all f ∈ B.
(∗∗)
Moreover, if f ∈ B we may write it as f =Pm
i=1 ciχUi, as in Equation (∗), so that
kRφ(f )R−1kM = k
ciReUiR−1k = max
1≤i≤m
ci = kf k∞ ,
mXi=1
using the fact that the ReUiR−1 form a family of pairwise orthogonal projections. Thus φ is
continuous as a linear map from the normed space (B, k · k∞) to the Banach space (M, k · k),
φB is an algebra homomorphism, so is φ. Equation (4.7) now follows by continuity from the
and hence has a unique (norm-)continuous extensioneB → M, which we also denote by φ. Since
special case (∗∗). It remains only to note that if h ∈ eB, and (fn) is any sequence in B with
kfn − hk∞ → 0, then
ıφ(h) = ı(cid:16)lim
n
φ(fn)(cid:17) = lim
n
ı (φ(fn)) = lim
n
ı0eθ(fn) = ı0eθ(h);
inclusion map A → M factors as φG, and so G is bounded below as a linear map; since G has
dense range by Corollary 2.6, this forces it to be surjective. Hence, given a ∈ A, there exists
thus ıφ = ı0eθ, and the proof of the proposition is complete.
We can now complete the proof of Theorem 1.1. For each a ∈ A, we know that ı(a) = ı0eθ(ba);
but by Proposition 4.8 we also have ı0eθ(ba) = ıφ(ba). Since ı is injective this implies that the
b ∈ A such thatbb =ba, and applying (4.7) gives
Thus RAR−1 is a self-adjoint subalgebra of fM, as required.
(RaR−1)∗ = (Rφ(ba)R−1)∗ = Rφ(bb)R−1 = RbR−1 ∈ RAR−1 .
Remark 4.9. Once we know A is similar to a self-adjoint subalgebra of M, it is immediate that
the Gelfand transform is bijective. However, at present I don't see how to show that A is similar
to a self-adjoint subalgebra without first proving that the Gelfand transform is surjective.
5 An example for contrast
WOT
has a large supply of idempotents, as it is similar to
It follows from Theorem 1.1 that A
If we had been able to find such idempotents in advance,
an abelian von Neumann algebra.
then the proof of Theorem 1.1 would have been easier and shorter. However, there are exam-
ples of semisimple, regular, commutative subalgebras of 2-homogeneous von Neumann algebras,
whose WOT-closures contains no idempotents other than 0 or I. This suggests that, given a
commutative subalgebra A of a finite von Neumann algebra, any attempt to produce non-trivial
idempotents in A
has to use some reasonably strong hypotheses on A.
WOT
Such an example may be found within the class of (isomorphic images of) little Lipschitz
algebras on compact metric spaces, whose definition we briefly recall. Let (X, d) be a compact,
16
infinite metric space; let Y = {(x, y) ∈ X × X : x 6= y}; and let 0 < α < 1. The little Lipschitz
algebra lipα(X, d) is the space of all functions f : X → C for which the associated function
∆αf : Y → C;
(x, y) 7→
f (x) − f (y)
d(x, y)α
def
= kf k∞ + k∆αf k∞, lipα(X, d) is a regular,
belongs to C0(Y ). Equipped with the norm kf kα
semisimple, commutative Banach algebra [22, §2]; it is shown in [2] that lipα(X, d) is weakly
amenable (see that paper for the definition) when α < 1/2, yet is never amenable.
We write M2 for the algebra of 2 × 2 complex matrices, and define M to be the (2-
homogeneous) von Neumann algebra
ℓ∞(Y, M2) ≡ ℓ∞(Y ) ⊗ M2 ≡ Y(x,y)∈Y
M2 .
If T ∈ M and (x, y) ∈ Y , we write T (x, y) for the 2 × 2 matrix that occurs as the "(x, y)th
coordinate" of T . The predual of M is ℓ1(Y, M2) ≡ ℓ1(Y )b⊗ M2, with duality implemented by
tr(S(x, y)T (x, y))
the pairing
(S, T ) 7→ X(x,y)∈Y
The following observation appears to be folklore (the author learned of it from M. C. White).
Lemma 5.1. Let 0 < α < 1. There is an injective, continuous algebra homomorphism θ :
lipα(X, d) → M with norm-closed range, which satisfies
θ(f )(x, y) =(cid:20)f (x) ∆αf (x, y)
(cid:21)
f (y)
0
for all (x, y) ∈ Y .
(5.1)
Outline of proof. Given (x, y) ∈ Y , define θx,y : lipα(X) → M2 by setting θx,y(f ) to be the right-
hand side of (5.1). Straightforward calculations show that θx,y is a homomorphism of norm ≤ 2,
and that 2 sup(x,y)∈Y kθx,y(f )k ≥ kf k∞ + k∆αf k∞α.
We let A denote θ(lipα(X, d)). This is a commutative operator algebra contained in M. and
we define B to be its σ(M, M∗)-closure.
Proposition 5.2. Let T ∈ B. Then there exists a unique function h : X → C such that
T (x, y) = θx,y(h) for all (x, y) ∈ Y . Moreover, h is continuous.
Proof. If a ∈ A = θ(lipα([0, 1])), then
a(x, y)11 = a(x, z)11 = a(y, x)22 = a(z, x)22 whenever x 6= y and x 6= z
a(x, y)21 = 0 whenever x 6= y
a(x, y)11 − a(x, y)22 = d(x, y)αa(x, y)21 whenever x 6= y
(5.2)
Hence, by σ(M, M∗)-continuity, these identities also hold for T ∈ B. We may therefore
define h : X → C by setting h(x) = T (x, y)11 (this does not depend on the choice of y). Since
T satisfies (5.2), it can be checked that θx,y(h) = T (x, y) whenever x 6= y; and clearly h is the
unique function with this property. Finally, for each x 6= y we have
h(x) − h(y) = d(x, y)αT (x, y)12 ≤ kT kd(x, y)α
and thus h is continuous.
17
It follows that if X is connected, then B contains no non-trivial idempotents.
6 Some closing remarks
Attempting to use the total reduction property The articles [9] and [11] work with
a property that is weaker than operator amenability, namely the total reduction property (see
the first of these articles for the definition and further discussion). The second half of the
argument in Section 4 (once we have set up the map ı0eθ : C0(X) → fM) might remain valid
if we only assume the total reduction property, especially since Proposition 3.5 is modelled on
arguments from [9]. However, the first half of our argument seems to crucially use some version
of amenability rather than merely the total reduction property, because we are looking at a
certain A-bimodule action on M, and not just at actions of A on Hilbert space.
Operator-Connes-amenable cases For certain completely contractive Banach algebras
which have natural preduals, one can consider the notion of operator-Connes-amenability. (This
notion seems to have been first formally introduced in the article [19], to which we refer the
reader for the context, definition, and further references to the literature). I suspect that one
could extend the method of proof in this article, to show that if a weak∗-closed, commutative
subalgebra of a finite von Neumann algebra is operator-Connes-amenable, then it is similar inside
M to a self-adjoint subalgebra. On the other hand, such a proof would require extra technical
baggage, and does not seem to lead to greater generality in practice: for in all examples I know
of, an (operator-)Connes-amenable dual Banach algebra has a norm-closed and weak∗-dense
subalgebra which is (operator) amenable, and Theorem 1.1 may be applied directly to that
subalgebra to obtain the desired similarity.
A possible alternative proof Inspection of our proof of Theorem 1.1 shows that a key
step involved, in passing, the construction of a large supply of closed A-invariant subspaces in
M -- namely, the ranges of the idempotents eU that were discussed in Proposition 4.7 -- which
give a kind of spectral decomposition for elements of A. With this in mind, it seems plausible
that the spectral subspaces and unbounded-idempotent-valued-measure discussed in [20] might,
in combination with some variant of Proposition 3.5, provide another way to obtain our result.
However, those constructions rest on deep and hard results of Haagerup and Schultz on (hy-
per)invariant subspaces for arbitrary operators in II1 factors, whose proofs appear to require a
formidable amount of work. The approach taken in Section 4 seems to use less machinery and
is relatively self-contained, save for Theorem 4.4. (It is worth remarking that the algebra fM
is also needed in the aforementioned work of Haagerup and Schultz, for similar reasons to the
present article; one is trying to construct idempotents using some kind of functional calculus,
but without a priori resolvent estimates one has to look in a larger space than M for the limits
of approximating sequences.)
Acknowledgements
The work here was partly supported by an internal grant for new faculty from the University of
Saskatchewan, and partly by an NSERC Discovery Grant.
18
I would like to thank Martin Argerami for kindly supplying part of the argument in the proof
of Lemma 2.11, and Serban Belinschi, Ebrahim Samei and Stuart White for their constructive
feedback on earlier versions of this article. Matthew Daws was a careful reader and made several
useful suggestions and corrections; I would also like to thank him for valuable discussions of [25],
back in 2006, and for later bringing [13] to my attention.
The diagrams in this article were prepared using Paul Taylor's diagrams.sty macros.
References
[1] M. Argerami, personal communication, http://mathoverflow.net/questions/36605
(version: 2010-08-26).
[2] W. G. Bade, P. C. Curtis, Jr., and H. G. Dales, Amenability and weak amenability for
Beurling and Lipschitz algebras, Proc. London Math. Soc. (3) 55 (1987), no. 2, 359 -- 377.
[MR 88f:46098]
[3] F. F. Bonsall and J. Duncan, Complete normed algebras, Ergebnisse der Mathematik und
[MR 54 #11013]
ihrer Grenzgebiete, Band 80, Springer-Verlag, New York, 1973.
[4] P. C. Curtis, Jr. and R. J. Loy, A note on amenable algebras of operators, Bull. Austral.
[MR 96f:47093]
Math. Soc. 52 (1995), no. 2, 327 -- 329.
[5] M. M. Day, Means for the bounded functions and ergodicity of the bounded representations
[MR 13,357e]
of semi-groups, Trans. Amer. Math. Soc. 69 (1950), 276 -- 291.
[6] J. Diestel, H. Jarchow, A. Tonge, Absolutely summing operators, Cambridge Studies in
Advanced Mathematics, vol. 43, Cambridge University Press, Cambridge, 1995.
[MR 96i:46001]
[7] J. Dixmier, Les moyennes invariantes dans les semi-groups et leurs applications, Acta Sci.
[MR 12,267a]
Math. Szeged 12 (1950), 213 -- 227.
[8] E. G. Effros and Z.-J. Ruan, Operator spaces, London Mathematical Society Monographs.
New Series, vol. 23, The Clarendon Press Oxford University Press, New York, 2000.
[MR 2002a:46082]
[9] J. A. Gifford, Operator algebras with a reduction property, J. Aust. Math. Soc. 80 (2006),
[MR 2007b:46087]
no. 3, 297 -- 315.
[10] B. E. Johnson, Approximate diagonals and cohomology of certain annihilator Banach alge-
[MR 47 #5598]
bras, Amer. J. Math. 94 (1972), 685 -- 698.
[11] L. W. Marcoux, On abelian, triangularizable, total reduction algebras, J. Lond. Math. Soc.
[MR 2009k:47235]
(2) 77 (2008), no. 1, 164 -- 182.
[12] G. J. Murphy, C ∗-algebras and operator theory, Academic Press Inc., Boston, MA, 1990.
[MR 91m:46084]
[13] E. Nelson, Notes on non-commutative integration, J. Functional Analysis 15 (1974), 103 --
[MR 50 #8102]
116.
19
[14] G. Pisier, Simultaneous similarity, bounded generation and amenability, Tohoku Math. J.
[MR 2008d:46075]
(2) 59 (2007), no. 1, 79 -- 99.
[15] Z.-J. Ruan, The operator amenability of A(G), Amer. J. Math. 117 (1995), no. 6, 1449 --
[MR 96m:43001]
1474.
[16]
, Amenability of Hopf von Neumann algebras and Kac algebras, J. Funct. Anal. 139
[MR 98e:46077]
(1996), no. 2, 466 -- 499.
[17] W. Rudin, Real and complex analysis, third ed., McGraw-Hill Book Co., New York, 1987.
[MR 88k:00002]
[18] V. Runde, The operator amenability of uniform algebras, Canad. Math. Bull. 46 (2003),
[MR 2004h:46050]
no. 4, 632 -- 634.
[19] V. Runde and N. Spronk, Operator amenability of Fourier-Stieltjes algebras, Math. Proc.
[MR 2005b:46106]
Cambridge Philos. Soc. 136 (2004), no. 3, 675 -- 686.
[20] H. Schultz, Brown measures of sets of commuting operators in a type II1 factor, J. Funct.
[MR 2008k:46191]
Anal. 236 (2006), no. 2, 457 -- 489.
[21] M. V. Seınberg, A characterization of the algebra C(Ω) in terms of cohomology groups,
[MR 57 #1150]
Uspehi Mat. Nauk 32 (1977), no. 5(197), 203 -- 204.
[22] D. R. Sherbert, The structure of ideals and point derivations in Banach algebras of Lipschitz
[MR 28 #4385]
functions, Trans. Amer. Math. Soc. 111 (1964) 240-272.
[23] M. Takesaki, Theory of operator algebras. I, Encyclopaedia of Mathematical Sciences, vol.
[MR 2002m:46083]
124, Springer-Verlag, Berlin, 2002, Reprint of the first (1979) edition.
[24] N. Tomczak-Jaegermann, The moduli of smoothness and convexity and the Rademacher
averages of trace classes Sp(1 ≤ p < ∞), Studia Math. 50 (1974), 163 -- 182. [MR 50 #8141]
[25] G. A. Willis, When the algebra generated by an operator is amenable, J. Operator Theory
[MR 97f:47044]
34 (1995), no. 2, 239 -- 249.
Address:
Department of Mathematics and Statistics
McLean Hall
University of Saskatchewan
106 Wiggins Road, Saskatoon, SK
Canada, S7N 5E6
Email: [email protected]
20
|
1507.04824 | 2 | 1507 | 2015-11-04T17:08:26 | An application of free transport to mixed $q$-Gaussian algebras | [
"math.OA"
] | We consider the mixed $q$-Gaussian algebras introduced by Speicher which are generated by the variables $X_i=l_i+l_i^*,i=1,\ldots,N$, where $l_i^* l_j-q_{ij}l_j l_i^*=\delta_{i,j}$ and $-1<q_{ij}=q_{ji}<1$. Using the free monotone transport theorem of Guionnet and Shlyakhtenko, we show that the mixed $q$-Gaussian von Neumann algebras are isomorphic to the free group von Neumann algebra $L(\mathbb{F}_N)$, provided that $\max_{i,j}|q_{ij}|$ is small enough. The proof relies on some estimates which are generalizations of Dabrowski's results for the special case $q_{ij}\equiv q$. | math.OA | math |
AN APPLICATION OF FREE TRANSPORT TO MIXED q-GAUSSIAN
ALGEBRAS
BRENT NELSON• AND QIANG ZENG◦
Abstract. We consider the mixed q-Gaussian algebras introduced by Speicher which
are generated by the variables Xi = li + l∗
i = δi,j
and −1 < qij = qji < 1. Using the free monotone transport theorem of Guionnet and
Shlyakhtenko, we show that the mixed q-Gaussian von Neumann algebras are isomorphic
to the free group von Neumann algebra L(FN ), provided that maxi,j qij is small enough.
The proof relies on some estimates which are generalizations of Dabrowski's results for
the special case qij ≡ q.
i , i = 1, . . . , N , where l∗
i lj − qij lj l∗
1. Introduction
A fundamental problem in the theory of operator algebras is whether two algebras
are isomorphic. The operator algebra (both the (reduced) C ∗-algebra and von Neumann
algebra) of the free group FN with N generators has been a central object to study. In
particular, the von Neumann algebras of FN are isomorphic to those generated by N
free semi-circular variables (Si)i=1,...,N due to Voiculescu; see [VDN92]. Motivated from
mathematical physics, Bozejko and Speicher introduced the q-Gaussian variables [BS91],
which can be regarded as a deformation of the free semi-circular system. Since then,
the q-Gaussian algebras have been extensively studied. For an incomplete list of results,
see [BKS97, Shl04, Nou04, ´Sni04, Ric05, KN11, Avs11] among others. More recently, using
estimates of Dabrowski [Dab14], Guionnet and Shlyakhtenko [GS14] have shown that
the q-Gaussian von Neumann algebras are isomorphic to those generated from the free
semi-circular variables for q small enough. This result was proved using the powerful
free monotone transport theorem. The first named author [Nel15] adapted this to the
non-tracial setting and showed that the finitely generated q-deformed free Araki-Woods
algebras are isomorphic to the finitely generated free Araki-Woods factor for q small
enough (cf. [Shl97], [Hia03]). In this paper, we give another application of Guionnet and
Shlyakhtenko's theory.
The q-Gaussian variables and q-commutation relations were further generalized with
the motivation from physics. In [Spe93], Speicher introduced the commutation relation
(1)
l∗
i lj − qijljl∗
i = δi,j
2010 Mathematics Subject Classification. 46L54, 81S05.
• Research supported by the NSF awards DMS-1161411 and DMS-1502822.
1
2
BRENT NELSON AND QIANG ZENG
where Q = (qij)N
i,j=1 is a symmetric matrix with qij ≤ 1, and δi,j is the Kronecker
delta function. It was shown [Spe93, BS94] that (1) can be represented as left creation and
annihilation operators on a certain Fock space. We will always use this Fock representation
of (1) in this paper. We call the operator algebras generated by Xi = li + l∗
i the mixed q-
Gaussian algebras and call Xi's the mixed q-Gaussian variables. In fact, the so-called braid
relations (a.k.a. Yang -- Baxter equation), which are more general than (1), were also studied
by Bozejko, Speicher, Nou, and Kr olak in [BS94, Nou04, Kr o00, Kr´o05], among others. As
for (1), Lust-Piquard [LP99] showed the Lp boundedness of the Riesz transforms associated
to the number operator of the system. More recently, Junge and the second named author
[JZ15] studied various properties of the mixed q-Gaussian von Neumann algebras and in
particular proved that they have the complete metric approximation property and are
strongly solid in the sense of Ozawa and Popa [OP10] as long as max1≤i,j≤N qij < 1.
In the present paper, we show that if max1≤i,j≤N qij is small enough then the mixed
q-Gaussian algebras are isomorphic to the algebras generated from free semi-circular vari-
ables. To state the result precisely, let us denote by Γq(RN ) the q-Gaussian von Neu-
mann algebra of N generators, L(FN ) the von Neumann algebra generated from FN , and
C ∗(Y1, ..., YN ) the C ∗-algebra generated by operators Y1, ..., YN .
Theorem 1. Let Q = (qij) be a symmetric N × N matrix with N ∈ {2, 3, . . .} and
qij ∈ (−1, 1). Let ΓQ be the von Neumann algebra generated by the mixed q-Gaussian
variables X1, . . . , XN . Then there exists a q0 = q0(N) > 0 depending only on N such
∼= Γ0(RN ) ∼= L(FN ) and C ∗(X1, . . . , XN ) ∼= C ∗(S1, ..., SN ) for all Q satisfying
that ΓQ
maxi,j qij < q0.
The proof of this theorem relies on the construction of the conjugate variables and
potentials for ΓQ. To this end, we follow the idea of Dabrowski [Dab14] and obtain some
estimates which are generalized from similar ones for the qij ≡ q case.
2. The Mixed q-Gaussian Algebra
We refer the readers to [BS94, LP99, JZ15] for unexplained preliminary facts for the
i=1 be an orthonormal basis of RN . The Fock space
mixed q-Gaussian variables. Let (ei)N
associated with the mixed q-Gaussian variables is defined as FQ = ⊕∞
Q is
isomorphic to (CN )⊗n as a vector space and H 0
Q = CΩ with Ω being the vacuum state.
Let Sn denote the symmetric group on n elements and write i = (i1, . . . , in) for a vector
in [N]n := {1, . . . , N}n. The inner product of FQ is given by
Q, where H n
n=0H n
hei1 ⊗ · · · ⊗ eim, ej1 ⊗ · · · ⊗ ejniQ = δm,n Xσ∈Sn
a(σ, j)hei1, ejσ−1(1)
i · · · heim, ejσ−1(n)
i.
Here a(σ, j) is a product of (qkl) defined as follows: We write τ1 = (12), τ2 = (23), . . . , τn =
(n1) for transpositions. It is well known that (τi)n
i=1 is a generating set of Sn and that the
AN APPLICATION OF FREE TRANSPORT TO MIXED q-GAUSSIAN ALGEBRAS
3
number of inversions of σ ∈ Sn is given by
For σ ∈ Sn, assume σ = k and σ = τm1 · · · τmk . Then (see [BS94, LP99])
σ = min{k ∈ N : σ = τi1 · · · τik}.
a(σ, i) =
k−1
Yj=1
q(iσj(mk−j ), iσj (mk−j +1))q(imk , imk+1),
where σj = τmk−j+1 · · · τmk and we have written qi1i2 = q(i1, i2). By definition,
li(ej1 ⊗ · · · ⊗ ejn) = ei ⊗ ej1 ⊗ · · · ⊗ ejn,
ri(ej1 ⊗ · · · ⊗ ejn) = ej1 ⊗ · · · ⊗ ejn ⊗ ei,
n
l∗
i (ej1 ⊗ · · · ⊗ ejn) =
δi,jkqij1 · · · qijk−1ej1 ⊗ · · · ⊗ ejk−1 ⊗ ejk+1 ⊗ · · · ⊗ ejn.
Here li = l(ei) is the left creation operator and l∗
i the left annihilation operator. One
can check that l∗
i is the adjoint operator of li with respect to the inner product h·, ·iQ of
L2(ΓQ, τQ). Similarly, ri and r∗
i are the right creation and annihilation operator, respec-
tively. Let Xi = li + l∗
i be the mixed q-Gaussian variables. Let ΓQ denote the mixed
q-Gaussian von Neumann algebra generated by Xi, i = 1, . . . , N. By [BS94], there is a
normal faithful tracial state τQ on ΓQ defined as τQ(X) = hXΩ, ΩiQ for X ∈ ΓQ.
If
maxij qij < 1, then there is a canonical unitary isomorphism between L2(ΓQ, τQ) and FQ
given by
X 7→ XΩ, for X ∈ ΓQ,
which extends continuously to L2(ΓQ). From time to time this identification will be used
implicitly in the following and we write h·, ·iτQ for the inner product of L2(ΓQ, τQ). Given
a finite-length tensor ξ ∈ FQ, there is a unique element W (ξ) in ΓQ such that W (ξ)Ω = ξ,
and W (ei1 ⊗ · · · ⊗ ein) is called the Wick word (a.k.a. Wick product in the literature) of
ei1 ⊗ · · · ⊗ ein.
Following [GS14, Dab14], we consider ChY1, . . . , YN i, the algebra of noncommutative
polynomials in N self-adjoint variables. Given a noncommutative power series
Xk=1
F (Y1, . . . , YN ) =Xi,p
ai,pYi1 · · · Yip ⊗ Yip+1 · · · Yin
whose radius of convergence is greater than R > 1, we define the norm kF kR =Pi,p ai,pRn.
Similarly, for
F (Y1, . . . , YN ) =Xi
aiYi1 · · · Yin,
with radius of convergence greater than R > 1 we define kF kR =Pi aiRn. For an algebra
A, we write Aop for the opposite algebra of A, and write a◦ ∈ Aop whenever a ∈ A.
4
BRENT NELSON AND QIANG ZENG
3. The Derivation ∂(Q)
j
and Ξj
Consider the linear map
∂(Q)
j
: ChX1, . . . , XN i → B(L2(ΓQ)),
∂(Q)
j
(X) = [X, rj] := Xrj − rjX.
For i = 1, . . . , N, define
Ξi : FQ → FQ, Ξi(ej1 ⊗ · · · ⊗ ejn) = qij1 · · · qijnej1 ⊗ · · · ⊗ ejn.
We also write qi(j) = qij1 · · · qijn for short.
For each n ≥ 1, we consider the following equivalence relation on [N]n: i ∼ j if ∃σ ∈ Sn
such that
i = σ · j = (jσ(1), . . . , jσ(n)).
Let [i] denote the equivalence class of i ∈ [N]n. Note that qk(j) = qk(i) for each j ∈ [i]
and each k = 1, . . . , N; consequently, we may at times denote qk(i) by qk([i]). For each
equivalence class [i] we define the subspace
F[i] := span(cid:8)ej1 ⊗ · · · ⊗ ejn : j ∈ [i](cid:9) ,
and denote by p[i] the orthogonal projection onto F[i]. It is easy to see that H 0
Q along with
the subspaces F[i] (ranging over all equivalence classes and all n ≥ 1) offers an orthogonal
decomposition of FQ, and consequently
pΩ +Xn≥1 X[i]∈[N ]n/∼
p[i] = 1,
where pΩ is the projection onto the vacuum vector. For notational consistency, we will
often denote pΩ = p[(∅)] ∈ [N]0/ ∼.
For each j = 1, . . . , N it follows that
(2)
Ξj =Xn≥0 X[i]∈[N ]n/∼
qj(i)p[i].
Since qj(i) are real numbers, Ξj is a self-adjoint operator. Moreover, if q := max1≤i,j≤N qij
satisfies q2N < 1 then Ξj ∈ HS(FQ), the Hilbert -- Schmidt operators on FQ, since for each
n ≥ 1
X[i]∈[N ]n/∼
kp[i]k2
HS = Xk1+···+kN =n(cid:18)
n
k1, . . . , kN(cid:19) = N n.
Noting that [li, rj] = 0, we see that
∂(Q)
j
(Xi)(ei1 ⊗ · · · ⊗ ein) = δi,jqii1 · · · qiinei1 ⊗ · · · ⊗ ein,
and hence ∂(Q)
ideal in B(FQ), the Leibniz rule implies ∂(Q)
(Xi) = δi,jΞj. As the space of Hilbert -- Schmidt operators is a two-sided
j maps ChX1, . . . , XN i into HS(FQ) for each
j
AN APPLICATION OF FREE TRANSPORT TO MIXED q-GAUSSIAN ALGEBRAS
5
j = 1, . . . , N whenever Ξj ∈ HS(FQ). When this is the case, we think of ∂(Q)
defined derivation
j
as a densely
Recall that L2(ΓQ ¯⊗Γop
Q , τQ ⊗ τ op
: L2(ΓQ, τQ) → HS(FQ).
∂(Q)
j
Q ) is isomorphic to HS(FQ) via the map
a ⊗ b◦ 7→ h·, b∗ΩiaΩ.
In particular, 1 ⊗ 1◦ 7→ pΩ. We will usually think of ∂(Q)
τ op
Q ).
Proposition 2. Suppose Ξj ∈ HS(FQ). Then ∂(Q)∗
j
j
(1 ⊗ 1◦) = Xj.
as having range L2(ΓQ ¯⊗Γop
Q , τQ ⊗
Proof. Fix i ∈ [N]n and let π1 ∈ B(FQ) denote the projection onto tensors of length one.
Then there exist scalars c1, . . . , cn such that
n
where we are summing over which operator Xi1, . . . , Xin created the vector eit. We claim
π1Xi1 · · · XinΩ =
cteit,
Xt=1
qit(j)hXi1 · · · Xit−1p[j]Xit+1 · · · XinΩ, ΩiQ
ct =Xd≥0 X[j]∈[N ]d/∼
= hXi1 · · · Xit−1ΞitXit+1 · · · XinΩ, ΩiQ.
First note that the second equality is immediate from (2). Now, the only terms from
π1Xi1 · · · XinΩ which contribute to ct are those where Xit creates eit; that is, ones where
the creation operator rather than the annihilation operator in Xit acts. Hence towards
computing ct we may replace Xit with lit and compute
π1Xi1 · · · Xit−1litXit+1 · · · XinΩ.
Recall that we have the partition of unity {p[j] : d ≥ 0,
and [j] ∈ [N]d/ ∼, let {ζ
[j]
ℓ } be an orthonormal basis for F[j]. Then we have
[j] ∈ [N]d/ ∼}. For each d ≥ 0
π1Xi1 · · · Xit−1litXit+1 · · · XinΩ
=Xd≥0 X[j]∈[N ]d/∼
=Xd≥0 X[j]∈[N ]d/∼Xℓ
π1Xi1 · · · Xit−1litp[j]Xit+1 · · · XinΩ
π1Xi1 · · · Xit−1eit ⊗ ζ
[j]
ℓ DXit+1 · · · XinΩ, ζ
[j]
ℓ EQ
.
Furthermore, of the above terms the only ones which contribute to ct are those where
eit survives; that is, where none of the operators Xi1, . . . , Xit−1 annihilate eit. And yet,
[j]
ℓ must be completely annihilated by Xi1 · · · Xit−1. The
to survive the action of π1, ζ
annihilation operators from Xi1 · · · Xit−1 tasked with this must each skip over eit at a
6
BRENT NELSON AND QIANG ZENG
scalar cost qitk for some k ∈ [N]. Since ζ
is a linear combination of ek1 ⊗ · · · ⊗ ekd,
k ∈ [j], the total scalar cost will be qit(j). The remaining actions of Xi1 · · · Xit−1 (any
creation operators and any annihilation operators acting on vectors left of eit in the tensor
product) are unaffected by the presence of eit. In summary, the contribution to ct from
the terms in the sum above is as follows:
[j]
ℓ
Xd≥0 X[j]∈[N ]d/∼Xℓ
qit(j)eitDXi1 · · · Xit−1ζ
[j]
ℓ , ΩEQDXit+1 · · · XinΩ, ζ
[j]
ℓ EQ
.
Noting that
Xℓ DXi1 · · · Xit−1ζ
[j]
ℓ , ΩEQDXit+1 · · · XinΩ, ζ
[j]
ℓ EQ
=DXi1 · · · Xit−1p[j]Xit+1 · · · XinΩ, ΩEQ
,
we see that ct has the claimed value.
Thus for s ∈ [N] we have
hXs, Xi1 · · · XiniτQ = hes, π1Xi1 · · · XinΩiQ
n
=
hes, eitiQhΩ, Xi1 · · · Xit−1ΞitXit+1 · · · XinΩiQ
Xt=1
= hpΩ, ∂(Q)
= h1 ⊗ 1◦, ∂(Q)
s
(Xi1 · · · Xin)iHS
s
(Xi1 · · · Xin)iτQ⊗τ op
Q
.
Extending this via linearity from monomials to the dense subset ChX1, . . . , XNi in the
domain of ∂(Q)
(cid:3)
concludes the proof.
s
Corollary 3. Suppose Ξj ∈ HS(FQ). Then
ChX1, . . . , XN i ⊗ ChX1, . . . , XNiop ⊂ Dom ∂(Q)∗
j
.
In particular, for a, b ∈ ChX1, . . . , XN i
(3)
∂(Q)∗
j
(a ⊗ b◦) = aXjb − m ◦ (1 ⊗ τQ ⊗ 1) ◦ (1 ⊗ ∂(Q)
j + ∂(Q)
j ⊗ 1)(a ⊗ b◦),
where m(a ⊗ b◦) = ab. Consequently, ∂(Q)
j
is closable.
Proof. The formula is a simple computation (cf. Proposition 4.1 in [Voi98], the proof
of Theorem 34 in [Dab14], or Corollary 2.4 in [Nel15]). The closability of ∂(Q)
then
follows because this formula holds on the dense subset ChX1, . . . , XN i⊗ChX1, . . . , XNiop ⊂
L2(ΓQ ¯⊗Γop
(cid:3)
Q , τQ ⊗ τ op
Q ).
j
Let us update the notation ∂(Q)
j
derivation.
so that from now on it denotes the closure of this
AN APPLICATION OF FREE TRANSPORT TO MIXED q-GAUSSIAN ALGEBRAS
7
Let φ : Sn → B(H n
Q) be the quasi-multiplicative function defined in [BS94] and define
P (n) =Pσ∈Sn
φ(σ). According to [BS94], we have
hξ, ηiQ = δn,mhξ, P (n)ηi0, for ξ ∈ H n
Q, η ∈ H m
Q .
Here h·, ·i0 is the inner product associated to (Γ0(RN ), τ0). Let q = max1≤i,j≤N qij.
Assume q < 1. By [Boz98, Theorem 2], we find
k(P (n))−1k ≤h(1 − q)
Using the Gauss identity, we have the estimate
∞
Yk=1
1 + qk
1 − qkin
.
(4)
k(P (n))−1k ≤h(1 − q)(cid:16)
∞
Xk=−∞
(−1)kqk2(cid:17)−1in
≤(cid:16) 1 − q
1 − 2q(cid:17)n
.
Lemma 4. If ε > 0 and q(3 − 2q + (3 + ε)2N 2) < 1, then there exists a noncommutative
power series representation of Ξi with radius of convergence greater than R = 2+ε
1−q > kXik
such that
kΞi − 1 ⊗ 1◦kR ≤
for i = 1, . . . , N .
qN 2(3 + ε)2
1 − q(3 − 2q + (3 + ε)2N 2)
=: π(q, N)
Proof. Following the argument of [Dab14], let Gn denote the Gram matrix of the inner
product on (ΓQ, τQ) from the natural basis (ei1 ⊗ · · · ⊗ ein) of H n
Q, where i ∈ [N]n. Namely,
Gn is the matrix of P (n) in the basis (ei1 ⊗ · · · ⊗ ein). We write ψi = W (ei1 ⊗ · · · ⊗ ein) for
the Wick word. From the isomorphism L2(ΓQ, τQ) ∼= FQ, we can also write
(Gn)ij = hei1 ⊗ · · · ⊗ ein, ej1 ⊗ · · · ⊗ ejniQ = hψi, ψjiτQ.
Let us define inductively the noncommutative polynomials, ψε = 1 for the empty word ε
and
n
j−1
(5)
ψi1,...,in(Y1, . . . , YN ) = Yi1ψi2,...,in −
δi1,ij
qi1ikψi2,...,ij−1,ij+1,...,in(Y1, . . . , YN ),
Xj=2
Yk=2
where the product over empty set is understood to be 1. It can be checked that ψi =
ψi(X1, . . . , XN ); cf. [Kr o00]. Let us define B = G−1/2
. Note that B is a positive-definite
symmetric N n × N n matrix and that Bij = 0 unless i ∼ j. For each i = n let
n
(6)
pi(Y1, . . . , YN ) = Xj=n
Bijψj(Y1, . . . , YN ).
Then {pi(X1, . . . , XN )Ω}i=n is an orthonormal basis of H n
Q, and {pk(X1, . . . , XN )Ω}k∈[i] is
an orthonormal basis of F[i]. We want to write Ξi as a sum of tensors. Unlike the qij ≡ q
8
BRENT NELSON AND QIANG ZENG
case considered in [Dab14], Ξi behaves more like a multiplier instead of a projection.
Consider
∞
Ξj(Y1, . . . , YN ) =
One can check that
qj(i)pi(Y1, . . . , YN ) ⊗ p∗
i (Y1, . . . , YN ).
Xn=0Xi=n
which means that Ξj can be identified as Ξj(X1, . . . , XN ) via the isomorphism FQ
L2(ΓQ, τQ). By the change of basis formula (6), writing wj = ψj(Y1, . . . , YN ), we have
∼=
Ξj(X1, . . . , XN )ψi = qj(i)ψi,
∞
Ξj(Y1, . . . , YN ) =
=
qj(i) Xj,k=n
BijBkiwj ⊗ w∗
k
qj(k)(B2)kjwj ⊗ w∗
k,
∞
Xn=0 Xi=n
Xn=0 Xj,k=n
where we have used in the second equality that Bki = 0 unless k ∼ i, in which case
qj(i) = qj(k). Taking the norm, we have
By (5), we find in the same way as the proof of [Dab14, Corollary 29] that
qj(k)(B2)kjwj ⊗ w∗
(cid:13)(cid:13)(cid:13) Xj,k=n
≤ qn Xk=n
k(cid:13)(cid:13)(cid:13)R
kwkkR(cid:13)(cid:13)(cid:13)Xj=n
1 − q(cid:19)n
1
.
(B2)kjwj(cid:13)(cid:13)(cid:13)R
Using the triangle inequality, we have
sup
i=n
kwikR ≤(cid:18)R +
≤ Xj=n
(G−1
n )kj sup
i=n
kwikR ≤ N nk(P (n))−1k(cid:18)R +
1
1 − q(cid:19)n
.
Combining with (4), we have
(cid:13)(cid:13)(cid:13)Xj=n
(B2)kjwj(cid:13)(cid:13)(cid:13)R
(cid:13)(cid:13)(cid:13) Xj,k=n
qj(k)(B2)kjwj ⊗ w∗
≤ qnN 2n(cid:18)R +
1
1 − q(cid:19)2n
(cid:16) 1 − q
1 − 2q(cid:17)n
.
k(cid:13)(cid:13)(cid:13)R
Plugging in R = 2+ε
1−q and summing over all n ≥ 1, we complete the proof.
(cid:3)
4. Proof of the Main Theorem
Let us write A = ChY1, . . . , YNi. Suppose Ξi ∈ ΓQ ¯⊗Γop
Q and is invertible in this algebra.
Let ∂j : A → A ⊗ Aop denote the j-th free difference quotient with the property ∂jP =
j = δi,j1 ⊗ 1◦, we have
PP =AYjB A ⊗ B for a monomial P ∈ A. Since ∂jXi = δi,jΞj#Ξ−1
, where # is the multiplication in ΓQ ¯⊗Γop
Q .
∂j = ∂(Q)
j #Ξ−1
j
AN APPLICATION OF FREE TRANSPORT TO MIXED q-GAUSSIAN ALGEBRAS
9
Proposition 5. Assume π(q, N) < 1. Then we have:
(i) There exist noncommutative power series ξj(Y1, . . . , YN ) of convergence radius R =
2+ε
1−q > kXik such that {ξj(X1, . . . , XN )}N
j=1 are the conjugate variables of X1, . . . , XN .
(ii) There exists a self-adjoint potential V (Y1, . . . , YN ) which is also a noncommutative
power series of convergence radius R such that DiV (Y1, . . . , YN ) = ξi(Y1, . . . , YN )
(iii) limq→0 kξi(Y1, . . . , YN ) − YikR = 0 for i = 1, . . . , N .
where Di is the cyclic gradient, i.e., DiP =PP =AYiB BA for P ∈ A.
Proof. By Lemma 4, Ξ−1
and we can define a noncommutative power series
j = Ξ−1
j (X1, . . . , XN ) for a noncommutative power series Ξ−1
j (Y1, . . . , YN )
ξj(Y1, . . . , YN ) := (Ξ−1
j )∗(Y1, . . . , YN )#Yj
− m ◦ (1 ⊗ τQ ⊗ 1) ◦ (1 ⊗ ∂(Q)
j + ∂(Q)
j ⊗ 1)((Ξ−1
j )∗(Y1, . . . , YN )) ∈ A,
where (a ⊗ b◦)#x = axb and m(a ⊗ b◦) = ab. Then by (3) we have
ξj := ξj(X1, . . . , XN ) = ∂(Q)∗
j
((Ξ−1
j )∗).
Consequently for P ∈ ChX1, . . . , XN i we have
hξj, P iτQ = hΞ−1
j
, ∂(Q)
j
(P )iHS = h1 ⊗ 1◦, ∂j(P )iτQ⊗τ op
Q
;
that is, ξj is a conjugate variable.
Let N be the number operator acting on ChY1, . . . , YNi; that is, N is defined by N P =
dP for any monomial P of degree d. Let Σ denote the inverse of N restricted to polynomials
with no degree zero term. Define
V (Y1, . . . , YN ) = Σ 1
2
N
Xi=1
ξi(Y1, . . . , YN )Yi + Yiξi(Y1, . . . , YN )! .
Then by precisely the same arguments as in Step 4 of the proof of Theorem 34 in [Dab14],
one can see that DiV (Y1, . . . , YN ) = ξi(Y1, . . . , YN ). Indeed, thanks to Proposition 2 and
part (i) above, Lemma 36 in [Dab14] can be verified using Lemma 12 in [Dab14] in our
setting. The rest argument of Step 4 is algebraic, and does not use our particular inner
product of FQ.
Finally, Lemma 4 implies that Ξ−1
j (Y1, . . . , YN ) converges to 1 ⊗ 1◦ with respect to the
R-norm as q → 0. By an argument similar to that of Lemma 4.3 in [Nel15], it is easy to
see that this implies limq→0 kξj(Y1, . . . , YN ) − YjkR = 0.
(cid:3)
Proof of Theorem 1. This follows from Proposition 5 and the free monotone transport
result of Guionnet and Shlyakhtenko [GS14, Corollary 4.3].
(cid:3)
10
BRENT NELSON AND QIANG ZENG
Acknowledgements
B.N. would like to thank Dimitri Shlyakhtenko for his comments about the paper, and is
grateful for the support from the UCLA Dissertation Year Fellowship and the NSF Math-
ematical Sciences Postdoctoral Research Fellowship. Q.Z. would like to thank Michael
Brannan, Alice Guionnet, and Marius Junge for helpful conversations. He also thanks the
financial support from Prof. Horng-Tzer Yau and the Center of Mathematical Sciences
and Applications at Harvard University. Both authors would like to thank NCGOA 2015
for providing the occasion for them to collaborate.
References
[Avs11] S. Avsec, Strong Solidity of the q-Gaussian Algebras for all −1 < q < 1, ArXiv e-prints (October
2011), available at 1110.4918.
[BKS97] M. Bozejko, B. Kummerer, and R. Speicher, q-Gaussian processes: non-commutative and clas-
sical aspects, Comm. Math. Phys. 185 (1997), no. 1, 129 -- 154. MR1463036 (98h:81053)
[Boz98] M. Bozejko, Completely positive maps on Coxeter groups and the ultracontractivity of the
q-Ornstein-Uhlenbeck semigroup, Quantum probability (Gda´nsk, 1997), 1998, pp. 87 -- 93.
MR1649711 (2000k:20053)
[BS91] M. Bozejko and R. Speicher, An example of a generalized Brownian motion, Comm. Math. Phys.
137 (1991), no. 3, 519 -- 531. MR1105428 (92m:46096)
[BS94]
, Completely positive maps on Coxeter groups, deformed commutation relations, and
operator spaces, Math. Ann. 300 (1994), no. 1, 97 -- 120. MR1289833 (95g:46105)
[Dab14] Y. Dabrowski, A free stochastic partial differential equation, Ann. Inst. Henri Poincar´e Probab.
Stat. 50 (2014), no. 4, 1404 -- 1455. MR3270000
[GS14] A. Guionnet and D. Shlyakhtenko, Free monotone transport, Invent. Math. 197 (2014), no. 3,
613 -- 661. MR3251831
[Hia03] F. Hiai, q-deformed Araki-Woods algebras, Operator algebras and mathematical physics
(Constant¸a, 2001), 2003, pp. 169 -- 202. MR2018229 (2004j:46086)
[JZ15] M. Junge and Q. Zeng, Mixed q-Gaussian algebras, ArXiv e-prints (May 2015), available at
1505.07852.
[KN11] M. Kennedy and A. Nica, Exactness of the Fock space representation of the q-commutation
relations, Comm. Math. Phys. 308 (2011), no. 1, 115 -- 132. MR2842972
[Kr o00] I. Kr olak, Wick product for commutation relations connected with Yang-Baxter operators and
new constructions of factors, Comm. Math. Phys. 210 (2000), no. 3, 685 -- 701. MR1777345
(2001i:46103)
[Kr´o05] I. Kr´olak, Contractivity properties of Ornstein-Uhlenbeck semigroup for general commutation
relations, Math. Z. 250 (2005), no. 4, 915 -- 937. MR2180382 (2006i:81105)
[LP99] F. Lust-Piquard, Riesz transforms on deformed Fock spaces, Comm. Math. Phys. 205 (1999),
no. 3, 519 -- 549. MR1711277 (2001j:46104)
[Nel15] B. Nelson, Free monotone transport without a trace, Comm. Math. Phys. 334 (2015), no. 3,
1245 -- 1298. MR3312436
[Nou04] A. Nou, Non injectivity of the q-deformed von Neumann algebra, Math. Ann. 330 (2004), no. 1,
17 -- 38. MR2091676 (2005k:46187)
[OP10] N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra, Ann. of
Math. (2) 172 (2010), no. 1, 713 -- 749. MR2680430 (2011j:46101)
AN APPLICATION OF FREE TRANSPORT TO MIXED q-GAUSSIAN ALGEBRAS
11
[Ric05] ´E. Ricard, Factoriality of q-Gaussian von Neumann algebras, Comm. Math. Phys. 257 (2005),
no. 3, 659 -- 665. MR2164947 (2006e:46069)
[Shl04] D. Shlyakhtenko, Some estimates for non-microstates free entropy dimension with applications
to q-semicircular families, Int. Math. Res. Not. 51 (2004), 2757 -- 2772. MR2130608 (2006e:46074)
, Free quasi-free states, Pacific J. Math. 177 (1997), no. 2, 329 -- 368. MR1444786
[Shl97]
(98b:46086)
[´Sni04] P. ´Sniady, Factoriality of Bozejko-Speicher von Neumann algebras, Comm. Math. Phys. 246
(2004), no. 3, 561 -- 567. MR2053944 (2005d:46130)
[Spe93] R. Speicher, Generalized statistics of macroscopic fields, Lett. Math. Phys. 27 (1993), no. 2,
97 -- 104. MR1213608 (94c:81096)
[VDN92] D. V. Voiculescu, K. J. Dykema, and A. Nica, Free random variables, CRM Monograph Series,
vol. 1, American Mathematical Society, Providence, RI, 1992. A noncommutative probability
approach to free products with applications to random matrices, operator algebras and harmonic
analysis on free groups. MR1217253 (94c:46133)
[Voi98] D. Voiculescu, The analogues of entropy and of Fisher's information measure in free probabil-
ity theory. V. Noncommutative Hilbert transforms, Invent. Math. 132 (1998), no. 1, 189 -- 227.
MR1618636 (99d:46087)
• Department of Mathematics, University of California, Berkeley, CA 94709
E-mail address: [email protected]
◦ Center of Mathematical Sciences and Applications, Harvard University, Cambridge,
MA 02138
E-mail address: [email protected]
|
1605.02800 | 2 | 1605 | 2017-01-29T05:34:01 | Around Property (T) for quantum groups | [
"math.OA",
"math.DS",
"math.FA"
] | We study Property (T) for locally compact quantum groups, providing several new characterisations, especially related to operator algebraic ergodic theory. Quantum Property (T) is described in terms of the existence of various Kazhdan type pairs, and some earlier structural results of Kyed, Chen and Ng are strengthened and generalised. For second countable discrete unimodular quantum groups with low duals Property (T) is shown to be equivalent to Property (T)$^{1,1}$ of Bekka and Valette. This is used to extend to this class of quantum groups classical theorems on 'typical' representations (due to Kerr and Pichot), and on connections of Property (T) with spectral gaps (due to Li and Ng) and with strong ergodicity of weakly mixing actions on a particular von Neumann algebra (due to Connes and Weiss). Finally we discuss in the Appendix equivalent characterisations of the notion of a quantum group morphism with dense image. | math.OA | math |
AROUND PROPERTY (T) FOR QUANTUM GROUPS
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
Abstract. We study Property (T) for locally compact quantum groups, providing sev-
eral new characterisations, especially related to operator algebraic ergodic theory. Quan-
tum Property (T) is described in terms of the existence of various Kazhdan type pairs, and
some earlier structural results of Kyed, Chen and Ng are strengthened and generalised.
For second countable discrete unimodular quantum groups with low duals Property (T) is
shown to be equivalent to Property (T)1
1 of Bekka and Valette. This is used to extend to
this class of quantum groups classical theorems on 'typical' representations (due to Kerr
and Pichot), and on connections of Property (T) with spectral gaps (due to Li and Ng) and
with strong ergodicity of weakly mixing actions on a particular von Neumann algebra (due
to Connes and Weiss). Finally we discuss in the Appendix equivalent characterisations of
the notion of a quantum group morphism with dense image.
,
Introduction
The discovery of Property (T) by Kazhdan in [Kaz] was a major advance in group theory.
It has numerous applications, in particular in abstract harmonic analysis, ergodic theory
and operator algebras, which can be found in the recent book [BHV] of Bekka, de la Harpe
and Valette dedicated to this property. Recall that a locally compact group G has Property
(T) if every unitary representation of G that has almost-invariant vectors actually has a
non-zero invariant vector. Property (T) is understood as a very strong type of rigidity. It
is 'antipodal' to softness properties such as amenability -- indeed, the only locally compact
groups that have both properties are the ones that have them trivially, namely compact
groups. Property (T) for G has many equivalent conditions (some under mild hypotheses),
among them are the following: the trivial representation is isolated in bG; every net of
normalised positive-definite functions on G, converging to 1 uniformly on compact sets,
converges uniformly on all of G; there exists a compact Kazhdan pair for G; there exists a
compact Kazhdan pair with 'continuity constants'; for every unitary representation π of G,
the first cohomology space H 1(G, π) of 1-cocycles vanishes; and all conditionally negative
definite functions on G are bounded.
In the context of von Neumann algebras, Property (T) was employed by Connes in
[Con1] to establish the first rigidity phenomenon of von Neumann algebras, namely the
countability of the fundamental group. Property (T) and its relative versions have been
introduced for von Neumann algebras in [Con2, CoJ, Pop1, AD, PeP, Pop2]. The various
forms of Property (T) have been a key component of Popa's deformation/rigidity theory
(see for example [Pop2] and references therein). Thus now Property (T) appears in operator
1
2
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
algebras both as an important ingredient related to the groups and their actions, and as a
notion intrinsic to the theory.
It is therefore very natural to consider Property (T) for locally compact quantum groups
in the sense of Kustermans and Vaes. The definition is the same as for groups, using the
quantum notion of a unitary (co-) representation. Property (T) was first introduced for
Kac algebras by Petrescu and Joit¸a in [PeJ] and then for algebraic quantum groups by
B´edos, Conti and Tuset [BCT]. Property (T) for discrete quantum groups in particular
was studied in depth by Fima [Fim], and later by Kyed [Kye] and Kyed with So ltan [KyS].
They established various characterisations of Property (T), extending the ones mentioned
above and others, produced examples, and demonstrated the usefulness of Property (T) in
the quantum setting. Property (T) for general locally compact quantum groups was first
formally defined by Daws, Fima, Skalski and White [DFSW] and then studied by Chen
and Ng [ChN]. Recently Arano [Ara] and Fima, Mukherjee and Patri [FMP] developed
new methods of constructing examples of Property (T) quantum groups; the first of these
papers showed also that for unimodular discrete quantum groups Property (T) is equivalent
to Central Property (T), which will be of importance for the second half of our paper. It
is also worth noting that these results spurred interest in studying Property (T) also in
the context of general C∗-categories ([PoV, NY]).
Most of our results establish further characterisations of Property (T), which can be
roughly divided into two classes. The first is based on conditions that do not involve
representations directly. These are interesting in their own right, and also provide a tool
to approach the conditions in the second class. The latter includes conditions that are
a formal weakening of Property (T) obtained either by restricting the class of considered
representations or by replacing lack of ergodicity by a weaker requirement. Some of these
results extend celebrated theorems about groups that have proven to admit many appli-
cations. The most crucial single result along these lines is a generalisation of a theorem of
Bekka and Valette [BV1], [BHV, Theorem 2.12.9], stating that for σ-compact locally com-
pact groups Property (T) is equivalent to the condition that every unitary representation
with almost-invariant vectors is not weakly mixing.
The paper is structured as follows. Its first part, consisting of Sections 1 -- 3, is devoted
to background material, and the main results are presented in its second part, consisting
of Sections 4 -- 8. Section 1 discusses preliminaries on locally compact quantum groups, in
particular their representations. Sections 2 and 3 contain more preliminaries as well as
new results on (almost-) invariant vectors and actions, respectively. Section 4 introduces
various notions of what a 'Kazhdan pair' for a locally compact quantum group should be,
and shows that each of these is equivalent to Property (T). It also proves that for locally
compact quantum groups H and G, if H has Property (T) and there is a morphism from H
to G 'with dense range', then G has Property (T) -- this substantially strengthens one of
the results of [ChN]. In Section 5, Kazhdan pairs are applied to present characterisations of
Property (T) in terms of positive-definite functions and generating functionals of particular
types, at least for particular classes of locally compact quantum groups.
For a discrete quantum group, Fima showed in [Fim] that Property (T) implies that the
quantum group is finitely generated, thus second countable, and unimodular. In Section
AROUND PROPERTY (T) FOR QUANTUM GROUPS
3
6 we introduce a further notion for a discrete quantum group -- that of having a 'low
dual'. We then provide the above-mentioned generalisation of the Bekka -- Valette theorem
for second countable unimodular discrete quantum groups with low duals (Theorem 6.3);
its consequences -- Theorems 6.6, 7.6 and 8.3 -- will also have these hypotheses on the
quantum group. Our proof is closer to those of [Jol2, Theorem 1.2] (see also [Jol1, Lemma
4.4]), [Bek, Theorem 9] and [PeP] than to the original one [BV1]. The low dual condition is a
significant restriction, but nevertheless still allows the construction of non-trivial examples
-- see Remark 1.6. The attempts to drop it led us to believe that in fact it might be
necessary for the quantum Bekka -- Valette theorem to hold; however we do not have a
theorem or a counterexample which would confirm or disprove this intuition.
As a first consequence of our version of the Bekka -- Valette theorem we establish that lack
of Property (T) is equivalent to the genericity of weak mixing for unitary representations
(Theorem 6.6). This extends a remarkable theorem of Kerr and Pichot [KP], which goes
back to a famous result of Halmos about the integers [Hal]. Next, in Sections 7 and 8 we
prove several results roughly saying that to deduce Property (T) it is enough to consider
certain actions of the quantum group. Since, as is the case for groups, every action of a
locally compact quantum group is unitarily implemented on a suitable Hilbert space, the
conditions obtained this way are indeed formally weaker than Property (T).
Motivated by Li and Ng [LiN], Section 7 deals with spectral gaps for representations
and actions of locally compact quantum groups. After proving a few general results, we
show that one can characterise Property (T) by only considering actions of the (discrete)
quantum group on B(K) with K being a Hilbert space (Theorem 7.6). The purpose of
Section 8 is to present a non-commutative analogue of the influential theorem of Connes
and Weiss [CoW], [BHV, Theorem 6.3.4], originally stated only for discrete groups. The
latter asserts that a second countable locally compact group G has Property (T) if and only
if every ergodic (or even weakly mixing) measure-preserving action of G on a probability
space is strongly ergodic; note that strong ergodicity is weaker than the absence of almost-
invariant vectors. Probabilistic tools play a central role in the proof of the Connes -- Weiss
theorem. Naturally, in the more general, highly non-commutative framework of locally
compact quantum groups, actions on probability spaces will not do. Instead, our result
(Theorem 8.3) talks about actions on a particular case of Shlyakhtenko's free Araki -- Woods
factors, namely VN(F∞), that preserve the trace. Thus, probability theory is essentially
replaced by free probability theory. What makes it possible is a result of Vaes [Va3] that
facilitates the construction of such actions.
The Appendix treats our notion of a morphism between two locally compact quantum
It presents several equivalent conditions, one of which is
groups having dense range.
required in Section 4.
Acknowledgements. The work on this paper was initiated during the visit of AS to the
University of Leeds in June 2012, funded by the EPSRC grant EP/I026819/1. We thank
Stuart White and Jan Cameron for valuable input at that time; we also acknowledge
several later discussions with Stuart White without which this paper would have never
been written. AV thanks Orr Shalit and Baruch Solel for discussions on issues related
4
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
to this work. Finally, we are grateful to the referees for their remarks. AS was partially
supported by the NCN (National Centre of Science) grant 2014/14/E/ST1/00525.
1. Preliminaries
We begin by establishing some conventions and notation to be used throughout the
paper. Inner products will be linear on the right side. For a Hilbert space H the symbols
B(H),K(H) will denote the algebras of bounded operators and compact operators on H,
respectively, and if ξ, η ∈ H, then ωξ,η ∈ B(H)∗ will be the usual vector functional, T 7→
hξ, T ηi. We will also write simply ωξ for ωξ,ξ.
The symbols ⊗,⊗ will denote the spatial/minimal tensor product of C∗-algebras and the
normal spatial tensor product of von Neumann algebras, respectively. If A is a C∗-algebra
then M(A) denotes its multiplier algebra. For ω ∈ A∗, we denote by ω ∈ A∗ its adjoint
given by ω(x) := ω(x∗), x ∈ A.
A morphism between C∗-algebras A and B is a non-degenerate ∗-homomorphism from
A to M(B), and the set of all such maps is denoted by Mor(A, B). Representations of
C∗-algebras in this paper are always non-degenerate ∗-representations. We will tacitly
identify elements of A∗ and Mor(A, B) with their unique extensions to M(A) that are strictly
continuous on the closed unit ball [Lan, Proposition 2.5 and Corollary 5.7]. This allows
composition of morphisms in the usual manner. We will freely use the basic facts of the
theory of multiplier algebras and Hilbert modules -- again, see [Lan] for details. We will
also occasionally use the 'leg' notation for operators acting on tensor products of Hilbert
spaces or C∗-algebras.
For a normal, semi-finite, faithful (n.s.f.) weight θ on a von Neumann algebra N,
we denote by (L2(N, θ), ηθ) the associated GNS construction and by Nθ the left ideal
{x ∈ N : θ(x∗x) < ∞}. We write ∇θ, Jθ and Tθ for the modular operator and modular
conjugation of θ and for the closure of the anti-linear map ηθ(x) 7→ ηθ(x∗), x ∈ Nθ ∩ N ∗
θ ,
respectively, all acting on L2(N, θ) [Str, Tak]. In particular, Tθ = Jθ∇1/2
. For a locally
compact quantum group G, we write ∇, J, T, η for the objects associated with the left Haar
weight of G (see the following subsection).
θ
1.1. Quantum groups. The fundamental objects studied in this paper will be locally
compact quantum groups in the sense of Kustermans and Vaes. We refer the reader to
[KV1, KV2, VD2] and to the lecture notes [Ku2] for an introduction to the general theory
-- we will in general use the conventions of these papers and also those of [DFSW].
A locally compact quantum group is a pair G = (L∞(G), ∆), with L∞(G) being a von
Neumann algebra and ∆ : L∞(G) → L∞(G) ⊗ L∞(G) a unital normal ∗-homomorphism
called the comultiplication (or coproduct) that is coassociative:
(∆ ⊗ id)∆ = (id ⊗ ∆)∆,
AROUND PROPERTY (T) FOR QUANTUM GROUPS
5
such that there exist n.s.f. weights ϕ, ψ on L∞(G), called the left and right Haar weights,
respectively, that satisfy
ϕ((ω ⊗ id)∆(x)) = ϕ(x)ω(1)
ψ((id ⊗ ω)∆(x)) = ψ(x)ω(1)
for all ω ∈ L∞(G)+
for all ω ∈ L∞(G)+
∗ , x ∈ L∞(G)+ with ϕ(x) < ∞,
∗ , x ∈ L∞(G)+ with ψ(x) < ∞.
For a locally compact quantum group G, the predual of L∞(G) will be denoted by L1(G)
and the corresponding algebra of 'continuous functions on G vanishing at infinity', which is
a weakly dense C∗-subalgebra of L∞(G), will be denoted by C0(G). The comultiplication
reduces to a map ∆ ∈ Mor(cid:0)C0(G), C0(G) ⊗ C0(G)(cid:1). The dual locally compact quantum
group of G will be denoted by bG. The GNS constructions with respect to the left Haar
weights of G,bG give the same Hilbert space. Therefore, we usually assume that both
L∞(G) and L∞(bG) act standardly on the Hilbert space L2(G). The comultiplication is
implemented by the multiplicative unitary W ∈ M(C0(G) ⊗ C0(bG)): ∆(x) = W ∗(1⊗ x)W
for all x ∈ L∞(G). The multiplicative unitary of bG, to be denotedcW , is equal to σ(W )∗,
where σ : M(C0(G)⊗ C0(bG)) → M(C0(bG)⊗ C0(G)) is the flip map. The 'universal' version
∆u ∈ Mor(cid:0)Cu
0(G), with the (coassociative) comultiplication
0(G) → C0(G)
The antipode S is an ultraweakly closed, densely defined, generally unbounded linear
operator on L∞(G). It maps a dense subspace of C0(G) into C0(G) and admits a 'polar
decomposition' S = R◦τ−i/2, where R is an anti-automorphism of L∞(G) called the unitary
antipode and (τt)t∈R is a group of automorphisms of L∞(G) called the scaling group. In
general the respective maps related to the dual quantum group will be adorned with hats,
and the 'universal versions' of objects we introduce will be adorned with the index u, so
0(G)(cid:1), the canonical reducing morphism Λ : Cu
of C0(G) (see [Ku1]) will be denoted by Cu
0(G), Cu
and the counit ǫ : Cu
0 (G) ⊗ Cu
0(G) → C.
Both R and τt leave C0(G) invariant and are implemented on L2(G) as follows (recall
antipode of Cu
the quantum group to which they are associated.
that for example the right invariant weight on bG will be denoted by bψ and the unitary
0(bG) by bRu; if necessary we will also decorate the symbols with the index of
the convention introduced above: b∇ is the modular operator of (L∞(bG),bϕ)):
∀x∈L∞(G) ∀t∈R.
Using the inclusion C0(G) ⊆ L∞(G) and the epimorphism Λ : Cu
0(G)∗.
The comultiplication induces the convolution product on Cu
(1.1)
0(G) → C0(G), one
obtains the natural embeddings L1(G) ֒→ C0(G)∗ ֒→ Cu
0(G)∗ by the formula ω1∗ω2 :=
0 (G)∗, C0(G)∗ and L1(G) into completely con-
(ω1⊗ω2)◦ ∆u for ω1, ω2 ∈ Cu
tractive Banach algebras. The embeddings from the last paragraph respect the convolution
product and make the smaller spaces closed ideals in the larger ones. The subspace
τt(x) = b∇itxb∇−it
R(x) = bJx∗bJ,
0(G)∗, turning Cu
is a dense subalgebra of L1(G). For ω ∈ L1
such that ρ(x) = ω(S(x)) for each x ∈ D(S). Then ω 7→ ω♯ is an involution on L1
♯ (G), let ω♯ be the unique element ρ ∈ L1(G)
♯ (G).
L1
♯ (G) :=(cid:8)ω ∈ L1(G) : ∃ρ∈L1(G) ∀x∈D(S)
ρ(x) = ω(S(x))(cid:9)
6
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
Sometimes we will assume that the locally compact quantum groups we study are second
countable, by which we mean that C0(G) is separable.
If the left and right Haar weights of G coincide, we say that G is unimodular. In the
case when G is compact (so that C0(G) is unital, and we denote it simply by C(G)) recall
The multiplicative unitary W admits 'semi-universal' versions (see [Ku1]), namely uni-
that G is of Kac type if its antipode S is bounded, or equivalently, bG is unimodular.
taries W ∈ M(Cu
properties. For every C∗-algebra B, there is a bijection between
0(G)⊗ C0(bG)) and W ∈ M(C0(G)⊗ Cu
• unitary elements U ∈ M(B ⊗ C0(bG)) with (id ⊗b∆)(U) = U13U12 and
• non-degenerate ∗-homomorphisms φ = φU : Cu
0(bG)), characterised by the following
0(G) → M(B),
given by the relation (φ ⊗ id)(W) = U.
Similarly, for every C∗-algebra B, there is a bijection between
• unitary elements U ∈ M(C0(G) ⊗ B) with (∆ ⊗ id)(U) = U13U23 and
• non-degenerate ∗-homomorphisms φ = φU : Cu
given by the relation (id ⊗ φ)(W) = U.
equalities W = (id ⊗ Λ bG)(VV) and W = (ΛG ⊗ id)(VV).
level, see [Ku1, Proposition 6.2], in that
There is also a truly universal bicharacter VV ∈ M(Cu
We shall use repeatedly that W implements 'one-half' of the coproduct at the universal
0(bG)).
It satisfies the
0(bG) → M(B),
0(G) ⊗ Cu
(id ⊗ ΛG)∆u(x) = W∗(1 ⊗ ΛG(x))W (x ∈ Cu
(1.2)
An element a ∈ M(C0(G)) is said to be a positive-definite function/element if there
exists µ ∈ Cu
+ such that a = (id⊗ µ)(W∗). We say that a is normalised if µ is a state.
For more information on positive-definite functions, including equivalent definitions, see
[Daw, DaS].
0(bG)∗
0(G)).
1.2. Representations of quantum groups. In this subsection we recall basic facts con-
cerning unitary representations of quantum groups.
Definition 1.1. A unitary representation of G (or a unitary corepresentation of C0(G))
on a Hilbert space H is a unitary U ∈ M(C0(G) ⊗ K(H)) with (∆ ⊗ id)U = U13U23. We
will often write HU for the Hilbert space upon which U acts. The trivial representation of
G, i.e. U = 1 ⊗ 1 ∈ M(C0(G) ⊗ C), will be denoted simply by 1.
In the above definition, it suffices to require that U ∈ L∞(G)⊗B(H), as this automatically
implies that U ∈ M(C0(G)⊗K(H)). This is folklore, see for example [BDS, Theorem 4.12].
As in this paper we will only consider unitary representations, we will most of the time
simply talk about 'representations of G'.
Definition 1.2. The contragradient representation of a representation U ∈ M(C0(G) ⊗
K(H)) is defined to be U c = (R⊗⊤)U ∈ M(C0(G)⊗K(H)). Here H is the complex conjugate
Hilbert space and ⊤ : K(H) → K(H) is the 'transpose' map, defined by ⊤(x)(ξ) = x∗(ξ).
AROUND PROPERTY (T) FOR QUANTUM GROUPS
7
Evidently, if J is an anti-unitary from H onto another Hilbert space J H and we consider
the ∗-anti-isomorphism j : K(H) → K(J H) given by j(x) := J x∗J ∗, x ∈ K(H), then the
representation (R ⊗ j)(U) is unitarily equivalent to U c.
Recall that an anti-unitary on a Hilbert space is involutive if its square is equal to 1.
Definition 1.3. Say that U satisfies condition R if there exists an involutive anti-unitary
J on H such that, with j : K(H) → K(H) as above, we have (R ⊗ j)(U) = U.
By the foregoing, if U satisfies condition R then U c is unitarily equivalent to U.
Every representation U ∈ M(C0(G)⊗K(H)) satisfies the formal condition (S ⊗ id)(U) =
U ∗, which means that for all ω ∈ B(H)∗,
(id ⊗ ω)(U) ∈ D(S) and S((id ⊗ ω)(U)) = (id ⊗ ω)(U ∗).
(1.3)
Due to the polar decomposition of the antipode and (1.1), this is equivalent to (see the
paragraph before Subsection 1.1 for the notation)
(id ⊗ ω)(U)bT ⊆ bT (id ⊗ ω)(U),
(1.4)
The map L1(G) → B(H) given by ω 7→ (ω ⊗ id)(U) is an algebra homomorphism, and its
restriction to L1
♯ (G) is a ∗-homomorphism. Thus, the norm closure of {(ω ⊗ id)(U) : ω ∈
L1(G)} is a C∗-algebra.
We define the tensor product of two representations in two ways: namely as the repre-
sentations
∀ω∈B(H)∗.
U
⊤ V = U12V13
and
⊥ V := V13U12,
U
both acting on HU ⊗ HV . (Note that the notation '
⊥ ' is not consistent with [Wor1].) There
is also a natural notion of the direct sum of two representations U and V , which is the
obvious representation acting on HU ⊕ HV . A representation of G is called irreducible if it
is not (unitarily equivalent to) a direct sum of two non-zero representations.
We will often use the fact (a consequence of the correspondence described in the previous
subsection) that there is a natural bijection between representations of G and representa-
0(bG), implemented by the semi-universal multiplicative unitary.
tions of the C∗-algebra Cu
It interacts naturally with the passing to the contragredient and the tensor product oper-
ations, as the two following lemmas show.
0(bG)∗ be a state, let
Lemma 1.4. Let G be a locally compact quantum group. Let µ ∈ Cu
0(bG) → B(K) be the GNS representation of µ and let U be the representation of G on
π : Cu
K associated with π. If µ is bRu-invariant then U satisfies condition R.
0(bG), µ). The map η(x) 7→ η(bRu(x∗)),
0(bG), extends to an involutive anti-unitary J on K by invariance of µ and be-
cause bRu is an anti-automorphism of Cu
0(bG) satisfying (bRu)2 = id. Now, defining an
anti-automorphism of B(K) by j(x) := J x∗J , x ∈ B(K), we have j ◦ π = π ◦ bRu. As a
Proof. Consider the GNS construction (K, π, η) of (Cu
x ∈ Cu
result,
(cid:3)
(R ⊗ j)(U) = (id ⊗ π)(R ⊗ bRu)(W) = (id ⊗ π)(W) = U.
8
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
and σ : M(Cu
0(bG) → B(HU ), φV : Cu
Lemma 1.5. Let U, V be representations of a locally compact quantum group G on Hilbert
spaces HU , HV , and let φU : Cu
0(bG) → B(HV ) be the corresponding
representations of Cu(bG). Then φU
⊤ V = φU ⋆ φV , where φU ⋆ φV = (φU ⊗ φV ) ◦ σ ◦ b∆u,
0(bG)) → M(Cu
0(bG) ⊗ Cu
Proof. Since (id ⊗ b∆u)(W ) = W 13W 12, the non-degenerate ∗-homomorphism φU ⋆ φV :
0(bG) → B(HU ⊗ HV ) satisfies
(id ⊗ (φU ⋆ φV ))(W) = (id ⊗ (φU ⊗ φV )σ)(W13W12) = (id ⊗ (φU ⊗ φV ))(W12W13)
0(bG)) is the tensor flip.
0(bG) ⊗ Cu
Cu
= (id ⊗ φU )(W)12(id ⊗ φV )(W)13 = U12V13 = U
⊤ V.
By the universal property of W, we have φU
⊤ V = φU ⋆ φV .
(cid:3)
We will also sometimes use another 'picture' of representations, as follows. For i = 1, 2,
let Ai be a C∗-algebra and Xi a Hilbert Ai-module. The exterior tensor product X1 ⊗ X2
[Lan, Chapter 4] is the Hilbert A1 ⊗ A2-module obtained from the algebraic tensor product
X1 ⊙ X2 using the A1 ⊙ A2-valued inner product
hx1 ⊗ x2, x′
1 ⊗ x′
2i := hx1, x′
1i ⊗ hx2, x′
2i
after completion. Now, we deduce from [Lan, Theorem 2.4 and p. 37] that
L(X1 ⊗ X2) ∼= M(K(X1 ⊗ X2)) ∼= M(K(X1) ⊗ K(X2))
(1.5)
canonically as C∗-algebras. In the particular case when X1 is a C∗-algebra A and X2 is
a Hilbert space H, the Hilbert A-module A ⊗ H satisfies L(A ⊗ H) ∼= M(A ⊗ K(H)) by
(1.5) since K(A) ∼= A. This ∗-isomorphism takes U ∈ M(A ⊗ K(H)) to the unique element
U ∈ L(A ⊗ H) characterised by
hb ⊗ η,U(a ⊗ ζ)i = b∗(id ⊗ ωη,ζ)(U)a
∀a,b∈A ∀ζ,η∈H.
When G is a locally compact quantum group and U ∈ M(C0(G)⊗K(H)) is a representation
of G on H, we henceforth use the correspondences U ↔ U ∈ L(C0(G) ⊗ H) ↔ φU ∈
Mor(Cu
0(bG),K(H)) without further comment.
1.3. Compact/discrete quantum groups. Recall that a locally compact quantum group
is called compact if the algebra C0(G) is unital (we then denote it simply by C(G)), or,
equivalently, the Haar weight is in fact a bi-invariant state.
It is said to be discrete if
C0(G) is a direct sum of matrix algebras (and is then denoted c0(G), and likewise we write
definitions and [Run] for the equivalent characterisations.
ℓ∞(G), ℓ1(G), ℓ2(G)), or, equivalently, bG is compact. See [Wor2, EfR, VD1] for the original
For a compact quantum group G the symbol Irr(G) will denote the family of all equiv-
alence classes of (necessarily finite-dimensional) irreducible unitary representations of G.
The trivial representation will be denoted simply by 1. We will always assume that for
each α ∈ Irr(G) a particular representative has been chosen and moreover identified with
a unitary matrix U α = (uα
ij)nα
AROUND PROPERTY (T) FOR QUANTUM GROUPS
9
i,j=1 ∈ Mnα(Cu(G)). So for all α ∈ Irr(G) and 1 ≤ i, j ≤ nα,
(1.6)
∆u(uα
ij) =
ik ⊗ uα
uα
kj.
nαXk=1
ij is a dense (Hopf) ∗-subalgebra of Cu(G), denoted Pol(G).
The span of all coefficients uα
The algebra of (vanishing at infinity) functions on the dual discrete quantum group is
given by the equality c0(bG) =Lα∈Irr(G) Mnα. Thus the elements affiliated to c0(bG) can be
identified with functionals on Pol(G). Note that as the Haar state of G is faithful on Pol(G)
we can also view the latter algebra as a subalgebra of C(G). The universal multiplicative
unitary of G is then given by the formula
W = Xα∈Irr(G)
ij ⊗ eα
uα
ij ∈ Yα∈Irr(G)
Cu(G) ⊗ Mnα = M(Cu(G) ⊗ c0(bG)).
(1.7)
The following definition will play an important role later on: a compact quantum group
G is low if the set {nα : α ∈ Irr(G)} is bounded in N. The dual of every discrete group
satisfies this trivially. Finally note that a compact quantum group G is second countable
if and only if Irr(G) is countable.
Remark 1.6. A classical locally compact group is low if and only if it has an abelian
subgroup of finite index, as shown by Moore ([Moo]). We do not know if an analogous
result holds for quantum groups: one would of course need to replace the 'abelian group' in
the statement above by a 'cocommutative quantum group'. In Sections 6 -- 8 of our paper a
key role will be played by 'second countable discrete unimodular quantum groups with low
duals'. Non-trivial examples of such objects can be produced via the bi-crossed product
construction of [FMP]. Indeed, if we start with a matched pair of a finite group G and
a countable discrete group Γ, Theorem 3.4 of [FMP] produces a quantum group G in the
class above (and Theorem 4.3 of the same paper shows that G has Property (T) -- to be
discussed in Section 4 -- if and only if Γ has Property (T)). Another construction of the
quantum groups in the class above is as follows: take any discrete (quantum) group G
in the class (so for example the dual of a second countable compact group which has an
abelian subgroup of finite index) and an action of a countable discrete group Γ on the dual
of G by automorphisms. Then the crossed product construction of Wang gives another
quantum group in the same class (see Theorem 6.1 of [FMP]). We refer for the details to
[FMP].
By a state on Pol(G) we mean a linear functional µ : Pol(G) → C which is positive in the
sense that µ(a∗a) ≥ 0 for all a ∈ Pol(G). The algebra Cu(G) is the enveloping C∗-algebra
of Pol(G) and there is a bijection between states of Cu(G) and states of Pol(G) by [BMT,
Theorem 3.3]. Note that any functional µ on Pol(G) can be identified with a sequence of
matrices (µα)α∈Irr(G), with µα ∈ Mnα defined by (µα)i,j = µ(uα
A generating functional on a compact quantum group G is a functional L : Pol(G) → C
which is selfadjoint (L(a∗) = L(a) for all a ∈ Pol(G)), vanishes at 1 and is conditionally
if a ∈ Pol(G) and
negative definite, i.e. negative on the kernel of the counit (formally:
ij), i, j = 1, . . . , nα.
10
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
ǫ(a) = 0, then L(a∗a) ≤ 0). Note a sign difference with, for example, [DFKS] -- here
we choose to work with conditionally negative definite functions, as in [Kye, DFSW]. A
(weakly continuous) convolution semigroup of states on Cu(G) is a family (µt)t≥0 of states
of Cu(G) such that
(i) µs+t = µs ∗ µt for all s, t ≥ 0;
(ii) µ0 = ǫ;
t→0+
(iii) µt
We will need at some point the next lemma (following from [LS1] -- see also [Sch]).
−−−→ ǫ in the weak∗ topology of Cu(G)∗.
Lemma 1.7. Let G be a compact quantum group. There is a one-to-one correspondence
between
(a) convolution semigroups of states (µt)t≥0 of Cu(G);
(b) generating functionals L on G.
It is given by the following formulas: for each a ∈ Pol(G) ⊆ Cu(G) we have
L(a) = lim
t→0+
ǫ(a) − µt(a)
,
µt(a) = exp∗(−tL)(a) :=
(−t)n
n!
L∗n(a).
t
∞Xn=0
We say that a generating functional L : Pol(G) → C is strongly unbounded if for each
M > 0 there exists α ∈ Irr(G) such that Lα is a positive matrix of norm larger than M.
Note that if L is strongly unbounded then it is also unbounded with respect to the universal
norm on Cu(G) (if L were bounded we could identify L = (Lα)α∈Irr(G) ∈ M(c0(G)) with
l = (L ⊗ id)(W); but then l would have to be bounded).
1.4. Morphisms of quantum groups. Let G, H be locally compact quantum groups. As
shown in [MRW] (see also [Ku1]), there is a natural correspondence between the following
classes of objects, each of which should be thought of as representing a quantum group
morphism from H to G:
(b) bicharacters (from H to G), i.e. unitaries
(a) morphisms
such that
such that
π ∈ Mor(cid:0)Cu
(π ⊗ π) ◦ ∆u
0(G), Cu
G = ∆u
0(H)(cid:1)
H ◦ π;
V ∈ M(cid:0)C0(H) ⊗ C0(bG)(cid:1)
(idC0(H) ⊗ ∆ bG)(V ) = V13V12,
(∆H ⊗ idC0( bG))(V ) = V13V23.
AROUND PROPERTY (T) FOR QUANTUM GROUPS
11
The correspondence is given by
The apparent difference with [MRW] stems from the fact that we are using the conventions
of [Ku1] and [DFSW] rather than these of [MRW] or [DKSS].
V = (ΛHπ ⊗ id)(WG).
(1.8)
(1.9)
the equality
To each morphism from H to G as above corresponds a dual morphism from bG to bH,
0(bH), Cu
0(bG)) determined uniquely by
described (for example) by the morphismbπ ∈ Mor(Cu
(π ⊗ id)(VVG) = (id ⊗bπ)(VVH).
G is said to have dense image if the map L1(bG) → M(Cu
the associated morphism π ∈ Mor(Cu
Definition 1.8. Let G and H be locally compact quantum groups. A morphism from H to
0(H)) induced by the restriction of
0(H)), namely ω 7→ π((id⊗ω)WG), is injective.
For a detailed discussion of this notion and several equivalent formulations we refer to
0(G), Cu
the Appendix.
2. Invariant and almost invariant vectors for representations
In this section we recall and expand some facts concerning the notions of invariant and
almost invariant vectors (as defined for example in [DFSW]) and connect these to the
classical Hilbert space convexity arguments. In particular, we extend a classical result of
Godement about an invariant mean on the Fourier -- Stieltjes algebra of a locally compact
group.
Let U ∈ M(C0(G) ⊗ K(H)) be a representation of a locally compact quantum group G
on a Hilbert space H.
Lemma 2.1 ([DFSW, Proposition 3.4]). For ζ ∈ H, the following conditions are equivalent:
(a) U(η ⊗ ζ) = η ⊗ ζ for every η ∈ L2(G), when considering U as acting on L2(G)⊗ H;
(b) (ω ⊗ id)(U)ζ = ω(1)ζ for every ω ∈ L1(G);
(c) φU (·)ζ = ǫ(·)ζ.
Remark 2.2. Conditions (a) and (b) are clearly equivalent for every U ∈ B(L2(G) ⊗ H).
When U is unitary, they are thus equivalent to the same conditions with U ∗ in place of U.
Definition 2.3. A vector that satisfies the equivalent conditions of the previous lemma
is said to be invariant under U. Denote by Inv(U) the closed subspace of all vectors in
H that are invariant under U, and by pU the projection of H onto Inv(U). If U has no
non-zero invariant vectors, it is called ergodic.
Note that the restriction U(1 ⊗ (1 − pU )) of U to Inv(U)⊥ is a representation of G on
Inv(U)⊥.
Lemma 2.4. For a net (ζi)i∈I of unit vectors in H, the following conditions are equivalent:
(I.a) for every η ∈ L2(G), kU(η ⊗ ζi) − η ⊗ ζik i∈I−→ 0;
(I.b) the net (id ⊗ ωζi)(U) converges in the weak∗ topology of L∞(G) to 1;
12
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
(I.c) for every a ∈ C0(G), kU(a ⊗ ζi) − a ⊗ ζik i∈I−→ 0;
0(bG), kφU (a)ζi −bǫ(a)ζik i∈I−→ 0.
(I.d) for every a ∈ Cu
Furthermore, the following conditions are equivalent:
(II.a) there exists a net satisfying the above equivalent conditions;
(II.b) the trivial representation of G is weakly contained in U, that is, there exists a state
(II.c) there exists a state Ψ on B(H) such that (id⊗Ψ)(U) = 1, or equivalently, (ΨIm φU )◦
A state Ψ as in (II.b) or (II.c) satisfies xpU = Ψ(x)pU for all x ∈ Im φU .
Ψ on Im φU such that Ψ ◦ φU =bǫ;
φU =bǫ.
Proof. Apart from the last sentence, everything is taken from [DFSW, Proposition 3.7 and
Corollary 3.8].
As for all a ∈ Cu
0 (bG) and ζ ∈ Inv(U), we have φU (a)ζ =bǫ(a)ζ, if Ψ is as in (II.b) or
(II.c), so that Ψ ◦ φU =bǫ, we get xpU = Ψ(x)pU for all x ∈ Im φU .
Definition 2.5. A net satisfying the equivalent conditions (I.a) -- (I.d) of the previous
lemma is said to be almost invariant under U. If such a net exists (namely, (II.a) -- (II.c)
are fulfilled), U is said to admit almost-invariant vectors.
(cid:3)
The next corollary will not be required later, but should be mentioned in this context.
For analogous results in the classical context see [LiN], [Pet, Proposition 1.5.5], and for
quantum groups see [BMT, Theorem 2.3].
Corollary 2.6. Let U be a representation of a discrete quantum group G on a Hilbert
space H and φU be the associated representation of Cu(bG). The following conditions are
equivalent:
(a) U has almost-invariant vectors;
(b) there exists a state Ψ of B(H) such that
Ψ(φU (uγ
ik)) = δik
∀γ∈Irr( bG) ∀1≤i,k≤nγ ;
F Pγ∈F
1
nγPnγ
ii) satisfies
i=1 φU (uγ
1 ∈ σ(Re aF ).
(c) for every finite F ⊆ Irr(bG), the operator aF := 1
Proof. (a) ⇐⇒ (b) is immediate from Lemma 2.4 because the operators of the form φU (uγ
ik)
are linearly dense in φU (Cu(bG)).
(a) =⇒ (c): let (ζι)ι∈I be a net of unit vectors that is almost-invariant under U. Then
ι∈I−→ 0. Hence
for every finite F ⊆ Irr(bG), we have aF ζι − ζι
(c) =⇒ (a): suppose that a finite F ⊆ Irr(bG) satisfies 1 ∈ σ(Re aF ). Then there is a
n→∞−→ 0. Since each φU (uγ
ii) is
n→∞−→ 0 for every γ ∈ F, 1 ≤ i ≤ nγ.
n→∞−→ 0 for
1 ∈ σ(Re aF ).
sequence (ζn)n∈N of unit vectors in H such that (Re aF )ζn − ζn
contractive, by uniform convexity we get φU (uγ
Since φ(nγ )
γ ∈ F, 1 ≤ i, k ≤ nγ. As F was arbitrary, U has almost-invariant vectors.
U (uγ) is unitary, thus contractive, we infer that φU (uγ
ι∈I−→ 0 and (Re aF )ζι − ζι
ik)ζn − δikζn
ii)ζn − ζn
(cid:3)
AROUND PROPERTY (T) FOR QUANTUM GROUPS
13
Definition 2.7. Let G be a locally compact quantum group.
• A finite-dimensional representation u ∈ M(C0(G) ⊗ Mn) of G is admissible [So l,
Definition 2.2] if, when viewed as a matrix (uij)1≤i,j≤n of elements in M(C0(G)), its
transpose (uji)1≤i,j≤n is invertible.
• A unitary representation U of G is weakly mixing [Vis, Definition 2.9] if it admits
no non-zero admissible finite-dimensional sub-representation.
Let G be a locally compact quantum group with trivial scaling group. The characterisa-
tions of weak mixing in [Vis, Theorem 2.11] become much simpler under this assumption,
as follows. For starters, every finite-dimensional (unitary) representation of G is admis-
sible. In [Vis], a '(unitary) representation of G on H' (called there a 'corepresentation')
is a unitary X ∈ B(H) ⊗ L∞(G) such that (id ⊗ ∆)(X) = X12X13. Such X is weakly
mixing (namely, it has no non-zero finite-dimensional sub-representation) if and only if
(⊤ ⊗ R)(X)13X23 is ergodic, if and only if Y13X23 is ergodic for all representations Y of
G as above. Let U ∈ L∞(G) ⊗ B(H) be unitary. Then U is a representation of G in
our sense if and only if X := σ(U ∗) is a representation of the opposite locally compact
quantum group Gop [KV2, Section 4] in the above sense. In this case, since Rop = R, we
have (⊤ ⊗ Rop)(σ(U ∗)) = σ(U c∗). Also, for any other representation V ∈ L∞(G) ⊗ B(K)
of G on a Hilbert space K, setting Y := σ(V ∗) we get σ12,3(Y13X23)∗ = U13V12 = V
⊥ U.
Notice also that a representation is ergodic, respectively weakly mixing, if and only if its
contragradient has this property. Putting all this together, weak mixing of U is equivalent
to ergodicity of any of U c
⊥ U c, and also to ergodicity --
and thus, a posteriori, to weak mixing -- of any of V
⊥ V
for every unitary representation V of G (for instance, V
⊤ U is weakly mixing because
(V
⊤ U is ergodic). Similar results can be also found
in [ChN].
⊥ U and U
⊥ U and U
⊤ U c, U c
⊤ (V
⊤ U) = ((V
⊤ U, U
⊤ V , V
⊤ U, U
⊤ U)c
⊤ U)c
⊤ V )
13U u
The notions of (almost-) invariant vectors discussed above are framed in terms of the
0(G). Recall from [Ku1, Proposition 6.6] that there is a bijection
0(G) ⊗ K(H))
23. The bijection is given by
algebra C0(G) and not Cu
between representations U ∈ M(C0(G) ⊗ K(H)) and unitaries U u ∈ M(Cu
satisfying the 'representation equation' (∆u ⊗ id)U u = U u
the relation (ΛG ⊗ id)(U u) = U; thus we have U u = (id ⊗ φU )(VV).
Definition 2.8. Let V ∈ M(Cu
0(G) ⊗ K(H)) be a representation in the sense discussed
above. Then ξ ∈ H is invariant for V if (µ ⊗ id)(V )ξ = µ(1)ξ for all µ ∈ Cu
0(G)∗. A net
of unit vectors (ξi)i∈I in H is almost invariant for V if k(µ ⊗ id)(V )ξi − ξik i∈I−→ 0 for each
state µ ∈ Cu
Proposition 2.9. With notation as above, Inv(U) = Inv(U u) and a net (ξi)i∈I of unit
vectors in H is almost invariant for U if and only if it is almost invariant for U u.
Proof. Applying the reducing morphism shows that Inv(U u) ⊆ Inv(U). Conversely, let ξ
be invariant for U. By (1.2)
0(G)∗.
U u
13U23 = ((id ⊗ ΛG)∆u ⊗ id)(U u) = W∗
12U23W12.
14
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
Fix a state ω ∈ L1(G) , and let µ ∈ Cu
0(G)∗. Then
(µ ⊗ ω ⊗ id)(U u
13U23)ξ = (µ ⊗ id)(U u)(ω ⊗ id)(U)ξ = (µ ⊗ id)(U u)ξ,
as ξ is invariant. Let π : Cu
0(G) → B(K) be a representation such that there are α, β ∈ K
with µ = ωα,β ◦ π. Let ω = ωγ for some γ ∈ L2(G). Then, with X = (π ⊗ id)(W) and
ζ ∈ H,
hζ, (µ ⊗ id)(U u)ξi = hζ, (µ ⊗ ω ⊗ id)(W∗
12U23W12)ξi = hα ⊗ γ ⊗ ζ, X ∗
12U23X12(β ⊗ γ ⊗ ξ)i
= hα ⊗ γ ⊗ ζ, X ∗
12X12(β ⊗ γ ⊗ ξ)i = hζ, µ(1)ξi,
using that ξ is invariant, so U(η ⊗ ξ) = η ⊗ ξ for all η ∈ L2(G), and that X is unitary.
Thus ξ is invariant for U u as claimed.
If (ξi)i∈I is a net of almost invariant vectors for U u, then let ω ∈ L1(G) be a state and set
µ = ω ◦ ΛG. It follows that k(ω ⊗ id)(U)ξi − ξik = k(µ⊗ id)(U u)ξi − ξik i∈I−→ 0. By linearity
k(ω ⊗ id)(U)ξi − ω(1)ξik i∈I−→ 0 for all ω ∈ L1(G) and so it follows that (id⊗ ωξi)(U) i∈I−→ 1
in the weak∗ topology of L∞(G), thus verifying condition (I.b) of Lemma 2.4. So (ξi)i∈I is
almost invariant for U.
Conversely, we use the same argument as above, but with more care. Let µ = ωβ ◦ π be
0(G) → B(K) a representation,
0(G), and ω = ωγ a state in L1(G), with π : Cu
a state of Cu
β ∈ K and γ ∈ L2(G). If (ξi)i∈I is almost invariant for U,
i∈I k(µ ⊗ ω ⊗ id)(U u
lim
13U23)ξi − ξik = lim
= lim
i∈I k(µ ⊗ id)(U u)(ω ⊗ id)(U)ξi − ξik
i∈I k(µ ⊗ id)(U u)ξi − ξik.
Let (ej)j∈J be an orthonormal basis for K, let X be as above and let X(β⊗γ) =Pj∈J ej ⊗
γj, soPj∈J kγjk2 = 1. Then, for η ∈ H,
(cid:12)(cid:12)hη, (µ ⊗ ω ⊗ id)(W∗
=(cid:12)(cid:12)hβ ⊗ γ ⊗ η, X ∗
=(cid:12)(cid:12)(cid:12)Xj∈J
12U23W12)ξii − hη, ξii(cid:12)(cid:12)
12U23X12(β ⊗ γ ⊗ ξi)i − hη, ξii(cid:12)(cid:12)
hγj ⊗ η, U(γj ⊗ ξi)i − hη, ξii(cid:12)(cid:12)(cid:12).
For ǫ > 0 there is a finite set F ⊆ J withPj6∈F kγjk2 < ǫ, and then there is i0 ∈ I so that
if i ≥ i0 and j ∈ F then kU(γj ⊗ ξi) − γj ⊗ ξik < ǫ/pF. Thus for such i,
hγj ⊗ η, γj ⊗ ξii(cid:12)(cid:12)(cid:12)
hγj ⊗ η, U(γj ⊗ ξi)i − hη, ξii(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)Xj∈F
(cid:12)(cid:12)(cid:12)Xj∈J
hγj ⊗ η, U(γj ⊗ ξi)i −Xj∈F
hγj ⊗ η, U(γj ⊗ ξi)i(cid:12)(cid:12)(cid:12)
hγj ⊗ η, γj ⊗ ξii − hη, ξii(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)Xj6∈F
+(cid:12)(cid:12)(cid:12)Xj∈F
kγjk2(cid:12)(cid:12)(cid:12) +Xj6∈F
ǫF−1/2kγj ⊗ ηk + hη, ξii(cid:12)(cid:12)(cid:12)1 −Xj∈F
≤Xj∈F
≤ kηk(cid:16)ǫ + ǫ + ǫ(cid:17) ≤ 3ǫkηk.
kγjk2kηk
AROUND PROPERTY (T) FOR QUANTUM GROUPS
15
So limi∈I k(µ⊗ ω ⊗ id)(W∗
as required.
12U23W12)ξi − ξik = 0 and hence limi∈I k(µ⊗ id)(U u)ξi − ξik = 0,
(cid:3)
We will later need to see invariant vectors of a given representation U as arising from a
version of an averaging procedure. Similar arguments can be found in [DaD, Section 4].
Definition 2.10. Let U be representation of a locally compact quantum group G on a
Hilbert space H. A subset S ⊆ B(H) is an averaging semigroup for U if S is a semigroup
of contractions such that
(a) each T ∈ S leaves Inv(U) pointwise invariant;
(b) each T ∈ S leaves Inv(U)⊥ invariant;
(c) if ξ ∈ H is such that T (ξ) = ξ for all T ∈ S, then ξ ∈ Inv(U).
Given such a semigroup, the S-average of ξ ∈ H is the unique vector of minimal norm in
the closed convex hull of {T (ξ) : T ∈ S}.
We will now give two examples of averaging semigroups.
Lemma 2.11. For any representation U of G the family S = {(ω ⊗ id)(U) : ω ∈
L1(G) is a state} is an averaging semigroup for U.
Proof. The fact that U is a representation shows that S is a semigroup of contractions.
We now check the conditions: (a) and (c) hold more or less trivially. Now let β ∈ Inv(U)⊥
and ω ∈ L1(G). For α ∈ Inv(U), we find that
hα, (ω ⊗ id)(U)βi = h(ω ⊗ id)(U ∗)α, βi = hω(1)α, βi = hα, βiω(1) = 0
using that α is invariant, so that (ω ⊗ id)(U ∗)α = ω(1)α by Remark 2.2. It follows that
(ω ⊗ id)(U)β ∈ Inv(U)⊥, as required to show (b).
Lemma 2.12. For any representation U of G the family S = {(ω ⊗ id)(U ∗) : ω ∈
L1(G) is a state} is an averaging semigroup for U.
Proof. As ω 7→ (ω ⊗ id)(U ∗) is an anti-homomorphism, S is a semigroup of contractions.
Conditions (a) and (c) hold by Remark 2.2, and the argument used in the previous proof
is easily adapted to show (b).
(cid:3)
Proposition 2.13. Let U be a representation of G on a Hilbert space H, and let S be an
averaging semigroup for U. Then:
(cid:3)
contains 0;
(a) U is ergodic if and only if, for all ξ ∈ H, the closed convex hull of {T (ξ) : T ∈ S}
(b) for ξ ∈ H, the S-average of ξ, the unique vector of minimal norm in the closed
convex hull of {T (ξ) : T ∈ S}, is equal to the orthogonal projection of ξ onto
Inv(U).
Proof. For (b), let ξ ∈ H and let C be the closed convex hull of {T (ξ) : T ∈ S}. As S is a
semigroup, it follows that T (C) ⊆ C for each T ∈ S. If η denotes the S-average of ξ, then
as each T ∈ S is a contraction, kT (η)k ≤ kηk and so by uniqueness, T (η) = η. By the
16
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
assumptions on S, it follows that η ∈ Inv(U). Now let ξ = ξ0 + ξ1 ∈ Inv(U) ⊕ Inv(U)⊥ be
the orthogonal decomposition of ξ. For T ∈ S, we have T (ξ0) = ξ0, and so
{T (ξ) : T ∈ S} = {ξ0 + T (ξ1) : T ∈ S}.
If we denote by C0 the closed convex hull of {T (ξ1) : T ∈ S}, then C = ξ0 + C0 and so
η = ξ0 + η0 for some η0 ∈ C0. As Inv(U)⊥ is invariant for S, it follows that C0 ⊆ Inv(U)⊥,
and so η0 ∈ Inv(U)⊥. We hence conclude that η − ξ0 = η0 ∈ Inv(U) ∩ Inv(U)⊥ = {0},
and so η = ξ0 as required. (We remark that actually, by Pythagoras's Theorem, η0 is the
unique vector of minimal norm in C0. It follows that 0 ∈ C0.)
(a) now follows immediately, as U is ergodic if and only if Inv(U) = {0}.
From the last result we obtain the following quantum analogue of the ergodic theorem
(cid:3)
in [RSN, Section 146].
Proposition 2.14. For a representation U of G and an averaging semigroup S for U, the
orthogonal projection onto Inv(U) belongs to the strong closure of the convex hull of S.
Proof. The convex hull of S is also an averaging semigroup for U. Thus, we can and do
assume the S is convex. Endow S with a preorder '(cid:22)' by saying that T1 (cid:22) T2 if T1 = T2
or there exists T ∈ S such that T2 = T T1.
Let ζ ∈ Inv(U)⊥ and ε > 0. By Proposition 2.13 there exists T0 ∈ S such that kT0ζk < ε.
Since S consists of contractions, it follows that kT ζk < ε for all T0 (cid:22) T ∈ S.
Fix ζ1, . . . , ζn ∈ Inv(U)⊥ and ε > 0. By the foregoing, there is T1 ∈ S such that
kT ζ1k < ε for all T1 (cid:22) T ∈ S. Since T1ζ2 ∈ Inv(U)⊥ there is, again by the foregoing,
T ′
2 ∈ S such that kT T1ζ2k < ε for all T ′
2T1, we have kT ζjk < ε
for all j ∈ {1, 2} and T2 (cid:22) T ∈ S. Proceeding by induction, there is Tn ∈ S such that
kT ζjk < ε for all j ∈ {1, . . . , n} and Tn (cid:22) T ∈ S, and in particular for T = Tn. This
proves the assertion because S leaves Inv(U) pointwise invariant.
2 (cid:22) T ∈ S. Setting T2 := T ′
(cid:3)
As another consequence of Proposition 2.13, we generalise a classical result of Godement
[G, Section 23] about the existence of an invariant mean on the Fourier -- Stieltjes algebra of
an arbitrary (not necessarily amenable) locally compact group. This should be compared
with Section 4, and in particular Theorem 4.4, of [DaD].
Recall that the module actions of the Banach algebra L1(G) on its dual L∞(G) are given
by
ω · a = (id ⊗ ω)(∆(a)),
a · ω = (ω ⊗ id)(∆(a))
Write E for the set of all a ∈ L∞(G) such that each of the sets
(a ∈ L∞(G), ω ∈ L1(G)).
{ω · a : ω ∈ L1(G) is a state}
intersects C1.
k·k
,
{a · ω : ω ∈ L1(G) is a state}
k·k
For a ∈ E, the intersections of the above sets with C1 are equal, and are a singleton.
n→∞−−−→ λ21 for
n=1 are states in L1(G) and ω1
n)∞
n
n)∞
n=1, (ω2
Indeed, if (ω1
some λ1, λ2 ∈ C, then
n · a n→∞−−−→ λ11, a · ω2
n) n→∞−−−→ λ21,
λ11
n→∞←−−− (ω1
n · a) · ω2
n = ω1
n · (a · ω2
AROUND PROPERTY (T) FOR QUANTUM GROUPS
17
so that λ1 = λ2. Let us denote this common scalar by M(a). It follows readily that E is
norm closed and selfadjoint, and that for every a ∈ E and λ, µ ∈ C, we have λa + µ1 ∈ E,
M(λa + µ1) = λM(a) + µ, M(a∗) = M(a), M(a) ≤ kak, and if ρ ∈ L1(G) is such that
ρ · a ∈ E (respectively, a · ρ ∈ E), then M(ρ · a) = ρ(1)M(a) (respectively, M(a · ρ) =
ρ(1)M(a)).
Denote by B(G) the Fourier -- Stieltjes algebra of G, namely the subalgebra of M(C0(G))
consisting of the coefficients of all representations of G:
B(G) := {(id ⊗ ω)(U) : U is a representation of G on a Hilbert space H and ω ∈ B(H)∗}
= {(id ⊗ ωζ,η)(U) : U is a representation of G on a Hilbert space H and ζ, η ∈ H} .
Write B(G)∗ for {a∗ : a ∈ B(G)}. We were informed that a result essentially equivalent to
the following proposition had been recently obtained by B. Das.
k·k
k·k
k·k
Proposition 2.15. The subalgebras B(G)
of M(C0(G)) are contained in
E, and M is linear on each of them. Thus, M restricts to an invariant mean on both
B(G)
Proof. Let U be a representation of G on a Hilbert space H. Fix ζ, η ∈ H, and consider
a := (id ⊗ ωη,ζ)(U). For every ρ ∈ L1(G),
and B(G)∗
and B(G)∗
k·k
.
ρ · a = (id ⊗ ρ)(∆(a)) = (id ⊗ ρ)(id ⊗ id ⊗ ωη,ζ)(U13U23)
= (id ⊗ ωη,(ρ⊗id)(U )ζ )(U)
a · ρ = (ρ ⊗ id)(id ⊗ id ⊗ ωη,ζ)(U13U23) = (id ⊗ ω(ρ⊗id)(U ∗)η,ζ )(U).
Denote by p the orthogonal projection onto Inv(U). By Lemmas 2.11, 2.12 and Proposition
2.13, there are sequences (ρ1
pζ and (ρ2
n=1 of states in L1(G) such that (ρ1
n ⊗ id)(U)ζ n→∞−−−→
n ⊗ id)(U ∗)η n→∞−−−→ pη. Consequently,
n=1, (ρ2
n)∞
n)∞
and
and
n · a n→∞−−−→ (id ⊗ ωη,pζ)(U) = hη, pζi 1
ρ1
n→∞−−−→ (id ⊗ ωpη,ζ)(U) = hpη, ζi 1
a · ρ2
n
in norm. This implies that a ∈ E. Hence, B(G) ⊆ E. Furthermore, the above calculation
shows that ρ· a, a· ρ ∈ B(G) ⊆ E, and thus M(ρ· a) = M(a) = M(a· ρ), for every a ∈ B(G)
and every state ρ ∈ L1(G).
Let a1, a2 ∈ B(G). Given ε > 0, find states ρ1, ρ2 ∈ L1(G) such that kρ1 · a1 − M(a1)1k ≤
ε and ka2 · ρ2 − M(a2)1k ≤ ε. Then
M(a1 + a2) = M(ρ1 · (a1 + a2) · ρ2),
so that
M(a1 + a2) − (M(a1) + M(a2)) = M(cid:0)(ρ1 · a1 − M(a1)1) · ρ2 + ρ1 · (a2 · ρ2 − M(a2)1)(cid:1) .
But the norm of the element to which M is applied on the right-hand side does not exceed
2ε. As a result, M(a1 + a2) − (M(a1) + M(a2)) ≤ 2ε. Therefore, M is additive, hence
18
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
linear, on B(G). This entails that M is norm-continuous, and so linear, on B(G)
also on B(G)∗
k·k
.
k·k
, thus
(cid:3)
The set of all representations of a second countable locally compact quantum group G
on a fixed infinite-dimensional separable Hilbert space H, denoted RepG
H, has a natural
Polish topology. This is the topology it inherits as a closed subset from the unitary group
of M(C0(G) ⊗ K(H)) equipped with the strict topology, which itself is a Polish space since
C0(G) ⊗ K(H) is a separable C∗-algebra. For more details, see [DFSW, Section 5].
Once we fix a unitary u : H → H ⊗ H, we can equip RepG
H with the product
U ⊠ V = (1 ⊗ u∗)(U
The lemmas below will be needed in Section 6.
⊤ V )(1 ⊗ u).
Notice that if Un
H → RepG
H, U 7→ U ⊠ U c.
H × RepG
H → RepG
H, (U, V ) 7→ U ⊠ V and RepG
H →
H, U 7→ U c are continuous (here U c is formed using some fixed anti-unitary from H
Lemma 2.16. The maps RepG
RepG
onto itself ). Hence, so is the map RepG
Proof. The second map is continuous by the definitions of U c and of the strict topology.
H, that is, in the strict topology on M(C0(G) ⊗ K(H)),
then (Un)12 − U12 → 0 in the strict topology on M(C0(G) ⊗ K(H ⊗ H)) as (Un)n∈N is a
n∈N−→ V in
sequence of unitaries, so that in particular it is bounded. Let Un
n∈N−→ U ⊠ V , we should show that
RepG
⊤ Vn)A − (U
n∈N−→ U in RepG
n∈N−→ U and Vn
H. To prove that Un ⊠ Vn
for all A ∈ C0(G) ⊗ K(H ⊗ H). However, this follows from the inequality
⊤ V )A(cid:13)(cid:13) n∈N−→ 0
(cid:13)(cid:13)(Un
⊤ V )A(cid:13)(cid:13)
⊤ Vn)A − (U
(cid:13)(cid:13)(Un
≤(cid:13)(cid:13)(Un)12(Vn)13A − (Un)12V13A(cid:13)(cid:13) +(cid:13)(cid:13)(Un)12V13A − U12V13A(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:0)(Vn)13 − V13(cid:1)A(cid:13)(cid:13) +(cid:13)(cid:13)(cid:0)(Un)12 − U12(cid:1)V13A(cid:13)(cid:13),
where both summands tend to zero as explained above.
(cid:3)
Lemma 2.17. The collection of all ergodic representations is a Gδ subset of RepG
H. If
the scaling group of G is trivial, so is the collection of all weakly mixing representations.
Proof. By Lemma 2.11 and Proposition 2.13 if S denotes the collection of states in L1(G),
then U is ergodic if and only if for every non-zero ξ ∈ H the closure of {(ω⊗id)(U)ξ : ω ∈ S}
contains 0. Let (ξn)n∈N be a dense sequence in the unit ball of H. As each operator
(ω ⊗ id)(U) is a contraction, we see that U is ergodic if and only if
U ∈ \n,m∈N[ω∈S(cid:8)V ∈ RepG
H : k(ω ⊗ id)(V )ξnk < 1/m(cid:9).
So the proof of the assertion on ergodic representations will be complete if we show that
for each n, m ∈ N and ω ∈ S the set {V ∈ RepG
H : k(ω ⊗ id)(V )ξnk < 1/m} is open,
H : k(ω ⊗ id)(V )ξnk ≥ 1/m} is closed. However, if (Vn)n∈N is a
equivalently, if {V ∈ RepG
AROUND PROPERTY (T) FOR QUANTUM GROUPS
19
n=1 that is dense in RepG
H whose equivalence class in RepG
H, namely {(1 ⊗
H.
sequence in this set converging strictly to V , then, as slice maps are strictly continuous,
(ω ⊗ id)(Vn) → (ω ⊗ id)(V ) strictly in B(H) = M(K(H)), and hence converges strongly,
showing the result. Combining the foregoing with the discussion after Definition 2.7 and
with Lemma 2.16 implies the assertion on weakly mixing representations.
(cid:3)
Lemma 2.18. There exists U ∈ RepG
u∗)U(1 ⊗ u) : u ∈ B(H) is unitary}, is dense in RepG
Proof. Choose a sequence (Un)∞
an element U in RepG
n=1 Un as
H by fixing a unitary v : H → H ⊗ ℓ2(N) and letting U := (1 ⊗
n=1 Un(1 ⊗ v), where the direct sum is calculated according to an orthonormal basis
(ηi)∞
H, ε > 0, a1, . . . , am ∈ C0(G) and k1, . . . , km ∈ K(H)
be such that the operators ki are of finite rank. By definition, there is n ∈ N such
i=1 ai ⊗ ki) ∈ C0(G) ⊗ K(H), we can
of finite rank operators in K(H) such that
v∗)L∞
that k(Un − V )(Pm
find finite collections (bj)j in C0(G) and(cid:0)k′
j(cid:1)j
(cid:13)(cid:13)(cid:13)Un(Pi ai ⊗ ki) −Pj bj ⊗ k′
j(cid:13)(cid:13)(cid:13) < ε/3. Write H1 for the (finite-dimensional) linear span of
j. Let u be a unitary on H such that uζ = v∗(ζ ⊗ ηn)
i=1 ai ⊗ ki)k < ε/3. Since Un(Pm
H. One can view L∞
the ranges of all operators kj and k′
for all ζ ∈ H1. Then
i=1 of ℓ2(N). Let V ∈ RepG
and so
ai ⊗ ki)(cid:13)(cid:13)(cid:13) < 2ε/3,
(cid:13)(cid:13)(cid:13)[(1 ⊗ u∗)U(1 ⊗ u) − Un] (
mXi=1
ai ⊗ ki)(cid:13)(cid:13)(cid:13) < ε.
(cid:13)(cid:13)(cid:13)[(1 ⊗ u∗)U(1 ⊗ u) − V ] (
mXi=1
The same technique works also for finite families of the formPm
and to the right of U, V . We deduce that the equivalence class of U is dense in RepG
i=1 ai⊗ki, applied to the left
H. (cid:3)
3. Actions of locally compact quantum groups on von Neumann algebras
In this section we discuss actions of locally compact quantum groups and their properties.
After quoting the basic definitions we discuss the canonical unitary implementation of such
actions due to Vaes, various notions of ergodicity, and (almost) invariant states. We obtain
in particular a new characterisation of the canonical unitary implementation of discrete
quantum group actions (Proposition 3.11, (c) and Corollary 3.12). These technical results
will be of crucial use in Sections 7 and 8.
Let G be a locally compact quantum group and let N be a von Neumann algebra. By
an action of G on N we understand an injective normal unital ∗-homomorphism α : N →
L∞(G) ⊗ N satisfying the action equation
The crossed product of N by the action α is then the von Neumann subalgebra of B(L2(G))⊗
(∆ ⊗ idN ) ◦ α = (idL∞(G) ⊗ α) ◦ α.
N generated by L∞(bG) ⊗ 1 and α(N).
20
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
We say that α is implemented by a unitary V ∈ B(L2(G) ⊗ K) if N ⊆ B(K) and
α(x) = V ∗(1 ⊗ x)V, x ∈ N.
In [Va1] Vaes shows that every action of a locally compact quantum group is implemented
in a canonical way. We use the notation of that paper, but our version of a representation
(so also what we call the unitary implementation of an action) is the adjoint of the one in
[Va1]. Let us then fix an n.s.f. weight θ on N. Define
and consider the linear map η0 : D0 → L2(G) ⊗ L2(N, θ) given by
D0 := span {(ba ⊗ 1)α(x) :ba ∈ N bϕ, x ∈ Nθ} ,
η0((ba ⊗ 1)α(x)) :=bη(ba) ⊗ ηθ(x)
(ba ∈ N bϕ, x ∈ Nθ).
Then η0 is indeed well-defined, and is ∗-ultrastrong -- norm closable. Denote its closure by η :
D → L2(G)⊗ L2(N, θ). Now D is a weakly dense left ideal in the crossed product G α⋉N,
and there exists a (unique) n.s.f. weight θ0 on G α⋉N such that (L2(G) ⊗ L2(N, θ), id, η)
is a GNS construction for θ0 (see [Va1, Lemma 3.3 and the preceding discussion, as well as
Definition 3.4]). We let J, ∇ stand for the corresponding modular conjugation and modular
operator, respectively, and set T := J ∇1/2. The unitary implementation of α is then the
unitary U := (bJ ⊗ Jθ) J. It satisfies
∀x∈N
(3.1)
and
α(x) = U ∗(1 ⊗ x)U
U(bJ ⊗ Jθ) = (bJ ⊗ Jθ)U ∗.
(3.2)
Furthermore, U ∈ M(C0(G) ⊗ K(L2(N, θ))) is a representation of G on L2(N, θ) [Va1,
Definition 3.6, Proposition 3.7, Proposition 3.12 and Theorem 4.4]. The choice of the
weight θ is of no importance up to unitary equivalence by [Va1, Proposition 4.1]. It is easy
to see that U satisfies condition R of Definition 1.3 with respect to Jθ.
bϕ, we see that the subspace
Lemma 3.1 ([Va1, Lemma 3.11]). Denoting by bT the closure of the anti-linear mapbη(bx) 7→
bη(bx∗),bx ∈ N bϕ ∩ N ∗
is a core for T and for every x, y ∈ Nθ and η ∈ D(bT ),
span(cid:8)α(x∗)(η ⊗ ηθ(y)) : x, y ∈ Nθ, η ∈ D(bT )(cid:9)
T α(x∗)(η ⊗ ηθ(y)) = α(y∗)(bT η ⊗ ηθ(x)).
(3.3)
Definition 3.2. Let α : N → L∞(G)⊗N be an action of a locally compact quantum group
G on a von Neumann algebra N. A (not necessarily normal) state θ of N is invariant under
α if (id ⊗ θ)α = θ(·)1, or equivalently, if θ(ω ⊗ id)α = ω(1)θ(·) for every ω ∈ L1(G).
Remark 3.3. Suppose that G = G, a locally compact group. The above notion of invari-
ance, which perhaps should have been called topological invariance, is stronger than the
classical one, i.e., θ ◦ αt = θ for all t ∈ G. Indeed, it is not difficult to observe that the
former entails the latter. The converse, however, is not always true: for every compact
(quantum) group G, there exists only one ∆-invariant state of L∞(G), namely the Haar
AROUND PROPERTY (T) FOR QUANTUM GROUPS
21
state h, as any ∆-invariant state θ satisfies h(·) = θ(h(·)1) = θ(h ⊗ id)∆ = θ(·); but, for
instance, taking G to be the complex unit circle, h is not the only state that is classically
invariant under the translation action, represented in our context by ∆ [Lub, Proposition
2.2.11]. Nevertheless, if G is discrete or θ is normal the two notions of invariance are easily
seen to be equivalent.
When α is an action of G on a von Neumann algebra N that leaves invariant a faithful
normal state θ, its unitary implementation takes a particularly simple form. To elaborate,
one verifies that the formula
(id ⊗ ωηθ(y),ηθ(x))(U ∗) = (id ⊗ θ)((1 ⊗ y∗)α(x))
or equivalently
(x, y ∈ N),
(3.4)
(ω ⊗ id)(U ∗)ηθ(x) = ηθ((ω ⊗ id)(α(x)))
(3.5)
defines (uniquely) an isometry U ∗, and that its adjoint U is a representation of G on
L2(N, θ), which is indeed a unitary by [BDS, Corollary 4.15], and which satisfies (3.1).
One can essentially repeat the argument in the proof of [Va1, Proposition 4.3], with 1 in
(x ∈ N, ω ∈ L1(G)),
place of δ−1 and the fact that b∇⊗∇θ commutes with U (see the proof of [RuV1, Theorem
A.1], and take into account the difference in the terminology) in lieu of [Va1, Proposition
2.4, last formula], to infer that U is the unitary implementation of α.
Definition 3.4. Let α : N → L∞(G) ⊗ N be an action of a locally compact quantum
group G on a von Neumann algebra N leaving invariant a faithful normal state θ of N.
We say that α is ergodic (respectively, weakly mixing) if its implementing unitary, when
restricted to L2(N, θ) ⊖ Cηθ(1), is ergodic (respectively, weakly mixing).
For an arbitrary action α of G on N, one normally defines ergodicity of α as the equality
of its fixed-point algebra N α := {a ∈ N : α(a) = 1 ⊗ a} and C1. In the context of Defi-
nition 3.4, this definition is known to be equivalent to ours when G is a group: see [Jad,
Lemma 2.2] or [HeT, Section 2, Theorem], where the abelianness assumption is unnec-
essary. We now show that this holds for all locally compact quantum groups (Corollary
3.6).
The next result is inspired by [KS] as presented in [DKS, Theorem 1] and [Jad, Section
2]; compare [Duv] and [RuV1, Theorem 2.2].
Proposition 3.5. Let α : N → L∞(G) ⊗ N be an action of G on a von Neumann
algebra N. Assume that α preserves a faithful normal state θ of N. Write U ∈ L∞(G) ⊗
B(L2(N, θ)) for the unitary implementation of α, and p for the orthogonal projection of
L2(N, θ) onto Inv(U). Then using the notation of Proposition 2.15, there exists a faithful
normal conditional expectation E of N onto N α given by
ω(E(a)) = M ((id ⊗ ω)α(a))
(3.6)
Moreover, E satisfies E(a)p = pap for all a ∈ N. This identity determines E uniquely.
Additionally, E is α-invariant: E ((ω ⊗ id)α(a)) = E(a) for every a ∈ N and every state
ω ∈ L1(G).
(a ∈ N, ω ∈ N∗).
22
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
k·k
Proof. First, let us explain why the right-hand side of (3.6) makes sense by proving that
(id ⊗ ω)α(a) ∈ B(G)∗
It suffices to prove that (id ⊗
ωJθ ηθ(c),Jθ ηθ(b))α(a) ∈ B(G)∗ for every b, c ∈ N. But since Jθηθ(b) = JθbJθηθ(1), JθbJθ ∈ N ′
and α(a) ∈ L∞(G) ⊗ N, we get (id ⊗ ωJθ ηθ(c),Jθ ηθ(b))α(a) = (id ⊗ ωJθ ηθ(b∗c),ηθ(1))α(a), which
by (3.4) equals (id ⊗ ωJθ ηθ(b∗c),ηθ(a))(U ∗) ∈ B(G)∗.
for each a ∈ N, ω ∈ N∗.
By Proposition 2.15, M is a linear contraction on B(G)∗
. Hence, (3.6) determines a
well-defined contraction E : N → N. To show that E maps into N α, pick a ∈ N and
states ω ∈ L1(G), ρ ∈ N∗. We should show that (ω ⊗ ρ)α(E(a)) = ρ(E(a)). Setting ν :=
(ω ⊗ ρ) ◦ α, we have (ω ⊗ ρ)α(E(a)) = M ((id ⊗ ν)α(a)) and ρ(E(a)) = M ((id ⊗ ρ)α(a))
by the definition of M. However,
k·k
(id ⊗ ν)(α(a)) = (id ⊗ ω ⊗ ρ)(id ⊗ α)(α(a)) = (id ⊗ ω ⊗ ρ)(∆ ⊗ id)(α(a))
= (id ⊗ ω)∆((id ⊗ ρ)α(a)) = ω · ((id ⊗ ρ)α(a)).
As a result, (ω ⊗ ρ)α(E(a)) = M ((id ⊗ ν)α(a)) = M ((id ⊗ ρ)α(a)) = ρ(E(a)), proving
that E(a) ∈ N α. It is also clear that E(b) = b for every b ∈ N α. In conclusion, E is a
conditional expectation of N onto N α.
Indeed, it suffices to show that if b ∈ N α then
Notice that N α commutes with p.
pbp = bp. But α(b) = 1 ⊗ b if and only if 1 ⊗ b commutes with U, if and only if b
commutes with (ω ⊗ id)(U) for every ω ∈ L1(G). Let ζ ∈ L2(N, θ). Then pbpζ is in the
closure of {(ω ⊗ id)(U)bpζ : ω ∈ L1(G) is a state} by Lemma 2.11 and Proposition 2.13.
Now (ω ⊗ id) (U)bpζ = b (ω ⊗ id) (U)pζ = bpζ for every ω, showing that pbpζ = bpζ.
Let a ∈ N. To observe that E(a)p = pap, take ζ, η ∈ L2(N, θ). Then
ωζ,η(E(a)p) = ωpζ,pη(E(a)) = M ((id ⊗ ωpζ,pη)α(a))
= M ((id ⊗ ωpζ,pη)(U ∗(1 ⊗ a)U)) = M(ωζ,η(pap)1) = ωζ,η(pap),
proving the desired equality. Since pηθ(1) = ηθ(1) and this vector is separating for N, the
equality E(a)p = pap determines E uniquely, and E is faithful and normal.
Let a ∈ N and let ω ∈ L1(G), ρ ∈ N∗ be states. Then
(id⊗ρ)α ((ω ⊗ id)α(a)) = (ω⊗id⊗ρ)(id⊗α)α(a) = (ω⊗id⊗ρ)(∆⊗id)α(a) = ((id⊗ρ)α(a))·ω.
Hence
ρ [E ((ω ⊗ id)α(a))] = M [(id ⊗ ρ)α ((ω ⊗ id)α(a))]
= M (((id ⊗ ρ)α(a)) · ω) = M ((id ⊗ ρ)α(a)) = ρ(E(a)),
implying that E ((ω ⊗ id)α(a)) = E(a). That is, E is α-invariant.
We now present an alternative proof of the existence and properties of E not using the mean M.
By Lemma 2.12 and Proposition 2.14, there is a net (ωi)i∈I of states in L1(G) such that
((ωi ⊗ id)(U ∗))i∈I converges strongly to p. Let a ∈ N. Then since U(1 ⊗ p) = 1 ⊗ p by
Lemma 2.1, we have
(cid:3)
pap = lim
i∈I
(ωi ⊗ id)(U ∗)ap = lim
i∈I
(ωi ⊗ id) (U ∗(1 ⊗ a)U) p = lim
i∈I
(ωi ⊗ id) (α(a)) p
(3.7)
AROUND PROPERTY (T) FOR QUANTUM GROUPS
23
strongly. The bounded net [(ωi ⊗ id) (α(a))]i∈I in N has a subnet [(ωj ⊗ id) (α(a))]j∈J
that converges in the weak∗ topology to an element b ∈ N. By (3.7), pap = bp. This
equality determines b uniquely in N since pηθ(1) = ηθ(1) and this vector is separating for
N. Moreover, as pηθ(a) = ηθ(b), it follows from (3.4) that b ∈ N α. Consequently, there
exists a linear map E : N → N α given by E(a)p = pap. From (3.7) it is clear that E is a
contractive projection, thus a conditional expectation, which is faithful and normal as in
the first proof. Also, similarly to (3.7),
E((ω ⊗ id)α(a))p = p(ω ⊗ id)α(a)p = pap = E(a)p, thus E((ω ⊗ id)α(a)) = E(a),
(cid:3)
for every a ∈ N and every state ω ∈ L1(G), so E is α-invariant.
Corollary 3.6. Under the assumptions of Proposition 3.5, α is ergodic (namely:
restriction of U to L2(N, θ) ⊖ Cηθ(1) is ergodic) if and only if N α = C1.
Proof. The implication ( =⇒ ) is trivial by (3.4). To prove the converse, suppose that
0 6= ζ ∈ L2(N, θ) ⊖ Cηθ(1) is invariant under U. Let (an)∞
n=1 be a sequence in N such
that ηθ(an) n→∞−−−→ ζ. Then ηθ(E(an)) = pηθ(an) n→∞−−−→ pζ = ζ. Since hηθ(1), ζi = 0, we
necessarily have N α % C1.
the
(cid:3)
Definition 3.7. Let α be an action of a locally compact quantum group G on a von
Neumann algebra N with an invariant faithful normal state θ. A bounded net (xi)i∈I in N
is called asymptotically invariant under α if for every normal state ω of L∞(G), we have
i∈I−→ 0
(ω ⊗ id)α(xi) − xi
strongly. We say that α is strongly ergodic if all its asymptotically invariant nets are trivial.
i∈I−→ 0 strongly. Such a net is said to be trivial if xi − θ(xi)1
Strong ergodicity evidently implies ergodicity.
In the setting of Definition 3.7, denote by U ∈ M(C0(G) ⊗ K(L2(N, θ))) the unitary
implementation of α. Observe that when (xi)i∈I is asymptotically invariant and does
not converge strongly to 0, we may assume, by passing to a subnet if necessary, that
(kηθ(xi)k)i∈I is bounded from below, and the normalised net (
kηθ(xi)k ηθ(xi))i∈I is then
almost invariant under U, because by (3.5), for every normal state ω of L∞(G), since α
preserves θ,
1
(ω ⊗ id)(U ∗)ηθ(xi) − ηθ(xi) = ηθ((ω ⊗ id)α(xi) − xi) i∈I−→ 0.
For the next definition, let N be a von Neumann algebra in standard form on L2(N)
with modular conjugation J and let n ∈ N. We will consider Mn ⊗ N as acting standardly
on L2(Mn, tr) ⊗ L2(N) which, as a vector space, is just Mn ⊗ L2(N). The positive cone
[Haa] of Mn ⊗ N in this representation is the closure of the set of sums of vectors of the
form (ξiJξj)n
Definition 3.8 ([Sau, 2.5]). Let N be a von Neumann algebra and u = (uij)n
Mn ⊗ B(L2(N)). We say that u preserves the positive cone if for every matrix (ζij)n
L2(Mn, tr)⊗ L2(N) in the positive cone of Mn ⊗ N, the matrix (uijζij)n
the positive cone of Mn ⊗ N.
i,j=1 for ξ1, . . . , ξn in the left Hilbert algebra of N inside L2(N).
i,j=1 ∈
i,j=1 ∈
i,j=1 also belongs to
24
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
Remark 3.9. By self-duality of the positive cone, (uij)n
i,j=1 in Mn ⊗ B(L2(N)) preserves
it if and only if(cid:0)u∗
ij(cid:1)n
preserves it.
i,j=1
We will require a generalisation of several results of Sauvageot [Sau]. Since the proof
of [Sau, Remarque 4.6, 2] is not explicit, and, as mentioned in [Va1], [Sau, Lemme 4.1] is
incorrect, we provide full details.
Proposition 3.10. Let G be a locally compact quantum group acting on a von Neumann
algebra N by an action α : N → L∞(G)⊗N. Let U ∈ M(C0(G)⊗K(L2(N))) be the unitary
id)(U ∗))1≤i,j≤n and ((ωξj, b∇−1/2ξi ⊗ id)(U))1≤i,j≤n preserve the positive cone.
implementation of α. Then for every ξ1, . . . , ξn ∈ D(b∇−1/2), the matrices ((ω b∇−1/2ξi,ξj ⊗
Proof. Fix a ∈ N, b, c, d ∈ Nθ, ξ ∈ D(b∇−1/2) and η ∈ L2(G). Note that bJθηθ(d) =
hη ⊗ bJθηθ(d), U ∗(ξ ⊗ aJθηθ(c))i = hη ⊗ JθdJθηθ(b), α(a)U ∗(ξ ⊗ Jθηθ(c))i
= hα(a∗)(η ⊗ ηθ(b)), (1 ⊗ Jθd∗Jθ)U ∗(ξ ⊗ Jθηθ(c))i .
JθdJθηθ(b) with JθdJθ ∈ N ′. By (3.1), we have
(3.8)
(3.9)
Using (3.1), (3.2) and (3.3) we obtain, as bJ ξ = bTb∇−1/2ξ,
(1 ⊗ Jθd∗Jθ)U ∗(ξ ⊗ Jθηθ(c)) = (bJ ⊗ Jθ)(1 ⊗ d∗)U(bJ ξ ⊗ ηθ(c))
= (bJ ⊗ Jθ)Uα(d∗)(bJξ ⊗ ηθ(c))
= J T α(c∗)(b∇−1/2ξ ⊗ ηθ(d))
= ∇1/2α(c∗)(b∇−1/2ξ ⊗ ηθ(d)).
To conclude,
(3.10)
hη ⊗ bJθηθ(d), U ∗(ξ ⊗ aJθηθ(c))i =Dα(a∗)(η ⊗ ηθ(b)), ∇1/2α(c∗)(b∇−1/2ξ ⊗ ηθ(d))E .
Take now a1, . . . , an, b1, . . . , bn ∈ Nθ and ξ1, . . . , ξn ∈ D(b∇−1/2). From (3.10),
nXi,j=1DbiJθηθ(bj), (ω b∇−1/2ξi,ξj ⊗ id)(U ∗)aiJθηθ(aj)E
nXi,j=1Db∇−1/2ξi ⊗ biJθηθ(bj), U ∗(ξj ⊗ aiJθηθ(aj))E
nXi,j=1Dα(a∗
i )(b∇−1/2ξi ⊗ ηθ(bi)), ∇1/2α(a∗
i )(b∇−1/2ξi ⊗ ηθ(bi))(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
nXi=1
∇1/4
j )(b∇−1/2ξj ⊗ ηθ(bj))E
≥ 0.
α(a∗
=
=
2
AROUND PROPERTY (T) FOR QUANTUM GROUPS
25
The desired conclusion follows from the self-duality of the positive cone and Remark 3.9.
(cid:3)
Proposition 3.11. Let α : N → ℓ∞(G) ⊗ N be an action of a discrete quantum group G
on a von Neumann algebra N. Denote its unitary implementation by U.
(a) Assume that a state m of N is invariant under α. Then there exists a net of unit
vectors (ζι)ι∈I in the positive cone P of N in L2(N) that is almost invariant under
U and such that ωζι
ι∈I−→ m in the weak∗ topology of N ∗.
(b) If a normal positive functional ρ of N is invariant under α, then the unique vector
ζ ∈ P with ρ = ωζ is invariant under U.
(c) The unitary implementation of α is the unique unitary U in M(c0(G) ⊗ K(L2(N)))
that satisfies (3.1) and such that for every ξ1, . . . , ξn ∈ D(b∇−1/2), the matrix ((ωξj, b∇−1/2ξi⊗
id)(U))1≤i,j≤n preserves the positive cone.
Proof. (a) From the standard convexity argument it follows that one can find a net (mι)ι∈I
of normal states of N converging in the weak∗ topology to the state m such that
k(ωη′,η ⊗ mι)α − hη′, ηi mιk ι∈I−→ 0
∀η,η′∈ℓ2(G).
(3.11)
and ε > 0. Let (ηi)n
For every ι ∈ I, write ζι for the unique vector in P such that mι = ωζι.
Recall that ℓ2(G) decomposes asLγ∈Irr( bG) ℓ2(G)γ. Moreover, for every γ ∈ Irr(bG), the
operator b∇ restricts to a bijection over ℓ2(G)γ. Fix a non-empty finite set F ⊆ Irr(bG)
i=1 be an orthonormal basis of ℓ2(G)F := Lγ∈F ℓ2(G)γ and set u :=
((ωηj , b∇−1/2ηi ⊗ id)(U))1≤i,j≤n and c := kb∇1/2ℓ2(G)F k. Assume that ρ is a normal state of N
with
(cid:13)(cid:13)(cid:13)(ω b∇−1/2ηj , b∇−1/2ηi ⊗ ρ)α − hb∇−1/2ηj,b∇−1/2ηiiρ(cid:13)(cid:13)(cid:13) ≤
and write ζ for the unique vector in P such that ρ = ωζ. For every η ∈ ℓ2(G)F ,
ε2
n4c2
∀1≤i,j≤n,
(3.12)
U(η ⊗ ζ) =
ηk ⊗ (ωηk,η ⊗ id)(U)ζ.
nXk=1
Therefore, for every 1 ≤ i, j ≤ n and x ∈ N,
(ω b∇−1/2ηj , b∇−1/2ηi ⊗ ρ)α(x) =DU(b∇−1/2ηj ⊗ ζ), (1 ⊗ x)U(b∇−1/2ηi ⊗ ζ)E =
so that
nXk=1
hujkζ, xuikζi ,
ε2
n4c2 kxk .
(3.13)
nXk=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
hujkζ, xuikζi − hb∇−1/2ηj,b∇−1/2ηii hζ, xζi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
Set Z1 := (hηj,b∇−1/2ηiiζ)1≤i,j≤n and Z2 := (uijζ)1≤i,j≤n. Both belong to the positive
cone of Mn ⊗ N in L2(Mn, tr) ⊗ L2(N), the second by Proposition 3.10. Consider the
26
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
functionals ωZ1, ωZ2 over Mn ⊗ N. For every X = (xij)1≤i,j≤n ∈ Mn ⊗ N, we have
ωZ1(X) =
and ωZ2(X) =Pn
nXi,j,k=1Dhηk,b∇−1/2ηjiζ, xjihηk,b∇−1/2ηiiζE =
i,j,k=1 hujkζ, xjiuikζi. Hence, by (3.13),
ε2
n4c2 kxjik ≤
(ωZ1 − ωZ2)(X) ≤
nXi,j=1
nXi,j=1
hb∇−1/2ηj,b∇−1/2ηii hζ, xjiζi
ε2
n2c2 kXk .
In conclusion, kωZ1 − ωZ2k ≤ ε2
it follows that kZ1 − Z2k ≤ ε
n2c2 . By the Powers -- Størmer inequality [Haa, Lemma 2.10],
nc. As a result, for each 1 ≤ i, j ≤ n,
A simple calculation thus shows that for every α, β ∈ ℓ2(G)F ,
(cid:13)(cid:13)(cid:13)(ωηj, b∇−1/2ηi ⊗ id)(U − 1L∞(G) ⊗ 1B(L2(N )))ζ(cid:13)(cid:13)(cid:13) = k(Z1)ij − (Z2)ijk ≤
(cid:13)(cid:13)(ωα,β ⊗ id)(U − 1L∞(G) ⊗ 1B(L2(N )))ζ(cid:13)(cid:13) ≤
To complete the proof, notice that by (3.11), there exists ι0 ∈ I such that ρ := mι
nckαkkβk = ε kαkkβk .
ε
nc
ε
nc
.
satisfies (3.12) for every ι0 ≤ ι ∈ I. Since ε was arbitrary, we get
(ωα,β ⊗ id)(U − 1L∞(G) ⊗ 1B(L2(N )))ζι = 0
lim
ι∈I
for all α, β ∈ ℓ2(G)F . Letting F vary, this holds for all α, β ∈ ℓ2(G) by density. From
Lemma 2.4, (I.b), the net (ζι)ι∈I is almost invariant under U.
(b) Repeat the proof of (a) with ε = 0.
(c) Suppose that U, V ∈ ℓ∞(G) ⊗ B(L2(N)) satisfy the given requirements. Fix γ ∈
i=1 be an orthonormal basis of ℓ2(G)γ, and set u := ((ωηj, b∇−1/2ηi⊗id)(U))1≤i,j≤n,
v := ((ωηj , b∇−1/2ηi ⊗ id)(V ))1≤i,j≤n. As in the proof of (a), for every x ∈ N and ζ ∈ L2(N),
Irr(bG), let (ηi)n
nXk=1
hujkζ, xuikζi = (ω b∇−1/2ηj , b∇−1/2ηi ⊗ ωζ)α(x) =
hvjkζ, xvikζi .
nXk=1
Consequently, if ζ ∈ P and we put Zu := (uijζ)1≤i,j≤n and Zv := (vijζ)1≤i,j≤n, then for
i,j,k=1 hvjkζ, xjivikζi =
ωZv(X). Since Zu, Zv belong to the positive cone of Mn ⊗ N in L2(Mn, tr) ⊗ L2(N) by
assumption, this entails Zu = Zv. As P is total in L2(N), we get u = v. Then it is easy
i=1 was a basis of ℓ2(G)γ.
Proposition 3.10 ends the proof.
(cid:3)
all X = (xij)1≤i,j≤n we have ωZu(X) =Pn
i,j,k=1 hujkζ, xjiuikζi =Pn
to see that in fact U = V , as γ ∈ Irr(bG) was arbitrary and (ηi)n
Corollary 3.12 (cf. [Sau, Lemme 4.4 and Remarque 4.6]). Let G be a locally compact
quantum group with trivial scaling group, acting on a von Neumann algebra N by an action
α : N → L∞(G) ⊗ N. Let U ∈ M(C0(G) ⊗ K(L2(N))) be the unitary implementation of
α. Then for every ξ1, . . . , ξn ∈ L2(G), the matrices ((ωξi,ξj ⊗ id)(U ∗))1≤i,j≤n and ((ωξj,ξi ⊗
id)(U))1≤i,j≤n preserve the positive cone. If G is discrete, then the unitary implementation
AROUND PROPERTY (T) FOR QUANTUM GROUPS
27
of α is the unique unitary U in M(c0(G)⊗K(L2(N))) that satisfies (3.1) and such that for
every ξ1, . . . , ξn ∈ ℓ2(G), the matrix ((ωξj,ξi ⊗ id)(U))1≤i,j≤n preserves the positive cone.
Proof. Let ξ1, . . . , ξn ∈ D(b∇−1/2). Since the scaling group of G is trivial, b∇ is affiliated
preserves the positive cone by Proposition 3.10. Since b∇−1/4D(b∇−1/2) is dense in L2(G),
with the commutant L∞(G)′. Hence, as U ∈ L∞(G) ⊗ B(L2(N)), we have (ω b∇−1/2ξi,ξj ⊗
id)(U ∗) = (ω b∇−1/4ξi, b∇−1/4ξj ⊗ id)(U ∗) for every i, j. Thus ((ω b∇−1/4ξi, b∇−1/4ξj ⊗ id)(U ∗))1≤i,j≤n
((ωξi,ξj ⊗ id)(U ∗))1≤i,j≤n preserves the positive cone for all ξ1, . . . , ξn ∈ L2(G). The second
statement follows similarly by using Proposition 3.11, (c).
(cid:3)
For the next lemma, recall that for a Hilbert space K, B(K) is standardly represented
(e.g., by using the trace on B(K)) on K ⊗ K by πK : B(K) ∋ x 7→ x ⊗ 1 with conjugation
J : K ⊗ K → K ⊗ K, ζ ⊗ η 7→ η ⊗ ζ. The positive cone of Mn ⊗ B(K) in Mn ⊗ K ⊗ K is
k ∈ K (1 ≤ i ≤ n,
the closure of the set of vectors of the form (Pm
1 ≤ k ≤ m).
Lemma 3.13. Let V ∈ M(c0(G) ⊗ K(K)) be a representation of a discrete quantum group
G on a Hilbert space K. The unitary implementation of the action α of G on B(K) given
by α(x) := V ∗(1 ⊗ x)V , x ∈ B(K), on K ⊗ K is V
Proof. Observe that by the definition of V c and (1.1), the operator V
k)1≤i,j≤n with ζ i
⊥ V c = V c
k ⊗ ζ j
k=1 ζ i
⊥ V c.
13V12 acting
The representation V
on ℓ2(G) ⊗ K ⊗ K using id ⊗ πK.
⊥ V c obviously implements α. By Proposition 3.11, (c) it suffices to
⊥ V c))1≤i,j≤n
on ℓ2(G)⊗ K⊗ K is equal to (bJ ⊗J )V ∗(bJ ⊗J )V , where V ∈ ℓ∞(G)⊗B(K) is represented
prove that for every α1, . . . , αn ∈ D(b∇−1/2), the matrix ((ωαj , b∇−1/2αi ⊗ id)(V
preserves the positive cone. For α ∈ D(b∇−1/2), β ∈ ℓ2(G) and ζ1, ζ2, η1, η2 ∈ K, we have
Dβ ⊗ ζ2 ⊗ η2, (V
=Dβ ⊗ ζ2 ⊗ η2, (bJ ⊗ J )V ∗(bJ ⊗ J )V (b∇−1/2α ⊗ ζ1 ⊗ η1)E
=D(bJ ⊗ J )V (bJβ ⊗ η2 ⊗ ζ2), V (b∇−1/2α ⊗ ζ1 ⊗ η1)E
=DbJ(id ⊗ ωη1,η2)(V )bJβ, (id ⊗ ωζ2,ζ1)(V )b∇−1/2αE .
⊥ V c)(b∇−1/2α ⊗ ζ1 ⊗ η1)E
Thus further, by (1.4),
Dβ ⊗ ζ2 ⊗ η2, (V
⊥ V c)(b∇−1/2α ⊗ ζ1 ⊗ η1)E =DbJ(id ⊗ ωη1,η2)(V )bJβ,bT (id ⊗ ωζ1,ζ2)(V )bJαE
=Db∇1/2(id ⊗ ωζ1,ζ2)(V )bJα, (id ⊗ ωη1,η2)(V )bJβE .
(3.14)
Take n, m ∈ N, α1, . . . , αn ∈ D(b∇−1/2) and ζ i
(3.14), applying ((ωαj, b∇−1/2αi ⊗ id)(V
k, ξi
⊥ V c))1≤i,j≤n to (Pk ζ i
k ∈ K (1 ≤ i ≤ n, 1 ≤ k ≤ m). By
k)1≤i,j≤n component-wise
k ⊗ ζ j
28
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
ℓ , (V
ℓ ⊗ ξj
and taking the inner product with (Pℓ ξi
ℓ ⊗ ξj
Xi,j,k,ℓDαj ⊗ ξi
⊥ V c)(b∇−1/2αi ⊗ ζ i
= Xi,j,k,ℓDb∇1/2(id ⊗ ωζ i
mXk,ℓ=1*b∇1/2 nXi=1
ℓ )1≤i,j≤n, we obtain
k)E
k ⊗ ζ j
)(V )bJαjE
)(V )bJαi, (id ⊗ ωζ j
)(V )bJαi! ,
nXj=1
(id ⊗ ωζ j
)(V )bJ αj+ ≥ 0
by the positivity of b∇1/2. By the self-duality of the positive cone, it is preserved by
⊥ V c))1≤i,j≤n, and the proof is complete.
((ωαj, b∇−1/2αi ⊗ id)(V
(id ⊗ ωζ i
k,ξj
k,ξi
ℓ
k,ξi
ℓ
k,ξj
ℓ
ℓ
(cid:3)
=
4. Property (T)
In this section we recall the fundamental notion of Property (T) for locally compact
quantum groups. We define different notions of having a 'Kazhdan pair' in this setting,
and show that they are all equivalent to Property (T). We also show that if a locally
compact quantum group G admits a morphism with a dense image from a quantum group
with Property (T), then also G itself has Property (T).
Definition 4.1. A locally compact quantum group has Property (T) if each of its repre-
sentations which has almost-invariant vectors has a non-trivial invariant vector.
Definition 4.2. A locally compact quantum group G has Property (T)1,1 if for every
⊤ U c has a non-
representation U of G with almost invariant vectors the representation U
zero invariant vector. Further if r, s ∈ N then we say that G has Property (T)r,s if for every
⊤ s has
representation U of G with almost invariant vectors the representation U
a non-zero invariant vector.
⊤ (U c)
⊤ r
Note that by the discussion after Definition 2.7 if G has a trivial scaling group, then it
has Property (T)1,1 if and only if each of its representations which has almost-invariant
vectors is not weakly mixing.
Proposition 4.3. If a locally compact quantum group G has Property (T), then it also
has Property (T)r,s for every r, s ∈ N.
Proof. This is trivial, for if a representation U admits a non-zero invariant vector, so does
⊤ r
U
(cid:3)
Definition 4.4. Let G be locally compact quantum group.
⊤ (U c)
⊤ s.
(a) A Kazhdan pair of type 1 for G is a finite set Q ⊆ L2(G) and ε > 0, such that
for any representation U ∈ M(C0(G) ⊗ K(H)), if there is a unit vector ξ ∈ H with
kU(η ⊗ ξ) − η ⊗ ξk < ε for each η ∈ Q, then U has a non-zero invariant vector.
(b) A Kazhdan pair of type 2 for G is a finite set Q ⊆ C0(G) and ε > 0, such that
for any representation U ∈ L(C0(G) ⊗ H), if there is a unit vector ξ ∈ H with
kU(a ⊗ ξ) − a ⊗ ξk < ε for each a ∈ Q, then U has a non-zero invariant vector.
AROUND PROPERTY (T) FOR QUANTUM GROUPS
29
for any representation φ : Cu
(c) A Kazhdan pair of type 3 for G is a finite set Q ⊆ Cu
0(bG) and ε > 0, such that
0(bG) → B(H), if there is a unit vector ξ ∈ H with
kφ(x)ξ −bǫ(x)ξk < ε for each x ∈ Q, then there is a non-zero vector ξ′ with φ(y)ξ′ =
bǫ(y)ξ′ for all y ∈ Cu
For a Kazhdan pair of any type (Q, ε), a vector ξ ∈ H that satisfies the inequalities in the
respective definition above is said to be (Q, ε)-invariant.
Theorem 4.5. G has Property (T) if and only if G has a Kazhdan pair of any (or equiv-
alently, all) type(s).
0(G).
Proof. By the various equivalent notions of having almost-invariant vectors in Lemma 2.4,
it is clear that if G has a Kazhdan pair of any type, then it has Property (T).
Conversely, suppose G does not have a Kazhdan pair of type 2, but that G has Property
(T). Denote by F the family of all finite subsets of the unit ball of C0(G). For each F ∈ F
and ε > 0, we can find a representation UF,ε with no non-zero invariant vector, but which
admits a unit vector ξF,ε with kUF,ε(a⊗ ξF,ε)− a⊗ ξF,εk < ε for a ∈ F . Ordering F × (0,∞)
in the obvious way, we find that
lim
(F,ε)(cid:13)(cid:13)UF,ε(a ⊗ ξF,ε) − a ⊗ ξF,ε(cid:13)(cid:13) = 0.
It follows easily that U =LUF,ε has almost invariant vectors, and hence has a non-zero
invariant vector, say η = (ηF,ε). However, then if ηF,ε 6= 0, then ηF,ε is a non-zero invariant
vector for UF,ε, a contradiction.
Exactly the same argument applies to Kazhdan pairs of type 1; and a very similar
argument applies to Kazhdan pairs of type 3.
(cid:3)
The following is valid for a Kazhdan pair of any type; we prove the version most suited to
0(bG).
if φ(x)ξ =bǫ(x)ξ for all x ∈ Cu
our needs. If we fix a representation φ : Cu
via the associated representation U = (id⊗ φ)(W). By Lemma 2.1, ξ ∈ Inv(U) if and only
Lemma 4.6. Let (Q, ε) be a Kazhdan pair of type 3 for G, and let δ > 0. For any
representation φ of Cu
P : H → Inv(U) is the orthogonal projection.
0(bG),
Proof. Let α ∈ Inv(U)⊥ and β ∈ Inv(U). For x ∈ Cu
0(bG) → B(H), we will define the space Inv(U) ⊆ H
0(bG) on H, if ξ ∈ H is (Q, εδ)-invariant, then kξ − P ξk ≤ δkξk where
hβ, φ(x)αi = hφ(x∗)β, αi =bǫ(x∗)hβ, αi = 0,
so φ(x)α ∈ Inv(U)⊥. Thus φ restricts to the subspace Inv(U)⊥.
Let ξ = P (ξ) + ξ0 where ξ0 ∈ Inv(U)⊥. As Inv(U)⊥ contains no non-zero invariant
vectors, there is x ∈ Q with kφ(x)ξ0 −bǫ(x)ξ0k ≥ εkξ0k. By assumption, kφ(x)ξ −bǫ(x)ξk ≤
εδkξk, and as P (ξ) is invariant, φ(x)P (ξ) =bǫ(x)P (ξ). Thus
kφ(x)ξ0 −bǫ(x)ξ0k = kφ(x)ξ0 − φ(x)ξ + φ(x)ξ −bǫ(x)ξ0k = k − φ(x)P (ξ) + φ(x)ξ −bǫ(x)ξ0k
= k −bǫ(x)P (ξ) + φ(x)ξ −bǫ(x)ξ0k = kφ(x)ξ −bǫ(x)ξk ≤ εδkξk.
We conclude that εkξ0k ≤ εδkξk, showing that kξ − P ξk = kξ0k ≤ δkξk as claimed.
(cid:3)
30
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
4.1. Hereditary property. Classically if a locally compact group G has Property (T)
and K is a closed normal subgroup of G then G/K has Property (T). In fact even if there
is only a morphism G → H with dense range, where H is another locally compact group,
then Property (T) passes from G to H.
In [ChN] it is shown under some additional assumptions that the existence of a surjective
0(bG) to Cu
∗-homomorphism from Cu
we strengthen this result, at the same time dropping the extra assumptions.
0(bH) implies that Property (T) passes from G to H. Here
Theorem 4.7. Let H, G be locally compact quantum groups. If there is a morphism from
H to G with dense image, and H has Property (T), then so does G.
Proof. Let the relevant morphism be given by π ∈ Mor(Cu
0(G), Cu
0 (H)). Let U be a rep-
resentation of G on a Hilbert space H which has almost-invariant vectors, (ξi)i∈I. They
are also almost invariant for U u by Proposition 2.9. Then V := (ΛH ◦ π ⊗ id)(U u) is a
representation of H. Let ω ∈ L1(H), and note that
lim
i∈I
ω((id ⊗ ωξi)(V )) = lim
i∈I hξi, (ω ◦ ΛH ◦ π ⊗ id)(U u)ξii = ω(1)
Let ω ∈ L1(H) and y = (ω ⊗ id)(WH). Then using (1.9) we see that
by the definition of what it means for (ξi)i∈I to be invariant for U u. Hence (id⊗ωξi)(V ) i∈I−−→
1 in the weak∗ topology of L∞(H), so by Lemma 2.4 the net (ξi)i∈I is almost invariant for
V . As H has Property (T), V has a non-zero invariant vector, say ξ0 ∈ H. Thus, Lemma
0(bH).
2.1 implies that φV (a)ξ0 =bǫ(a)ξ0 for all a ∈ Cu
φU (bπ(y)) = φU(cid:0)(ω ⊗bπ)(WH)(cid:1) = φU(cid:0)(ωΛHπ ⊗ id)(VVG)(cid:1) = (ωΛHπ ⊗ id)(U u) = (ω ⊗ id)(V )
0(bH). The same calculation yields thatbǫG ◦bπ =bǫH.
0(bH). By condition (d)
Hence φU (bπ(y))ξ0 = φV (y)ξ0 =bǫH(y)ξ0 =bǫG(bπ(y))ξ0 for all y ∈ Cu
0(bG) → B(H) = M(K(H))
in Theorem A.1,bπ has strictly dense range. In addition, φU : Cu
is strictly continuous, hence strictly-strongly continuous. It follows that φU (x)ξ0 =bǫG(x)ξ0
0 (bG)). In particular, ξ0 is invariant for U, as required.
for all x ∈ M(Cu
= φV(cid:0)(ω ⊗ id)(WH)(cid:1) = φV (y).
By density, this holds for all y ∈ Cu
5. Applications of Kazhdan pairs
(cid:3)
The main aim of this section is to apply the results on Kazhdan pairs to give an alter-
native proof and strengthen in a crucial way Theorem 3.1 of [Kye]. In connection with
the recent results of [Ara] this will enable us in the next section to establish equivalence of
Property (T) and Property (T)1,1 for a class of discrete quantum groups.
The equivalence of (a) and (b) of the next result was established in [Kye, Theorem 3.1]
under the additional assumption that G is a second countable discrete quantum group.
However, the following proof, and especially (c) =⇒ (a), is quite different: not having
the particular structure of discrete quantum groups at our disposal, we use the averaging
semigroups machinery of Section 2. Note that outside of the Property (T) context [RuV2,
AROUND PROPERTY (T) FOR QUANTUM GROUPS
31
Theorem 4.6] provides an equivalence between various modes of convergence of normalised
positive-definite elements on locally compact quantum groups.
Theorem 5.1. Let G be a locally compact quantum group. The following are equivalent:
(a) G has Property (T);
(b) if (µi)i∈I is a net of states of Cu
(c) if (ai)i∈I is a net of normalised positive-definite functions in M(C0(G)), which con-
0(bG) converging in the weak∗ topology to the counit
bǫ, then (µi)i∈I converges in norm tobǫ;
verges strictly to 1, then limi∈I kai − 1k = 0.
lim
i∈I kφ(a)ξi −bǫ(a)ξik2 = lim
Proof. (a) =⇒ (b): If G has Property (T) then for each i ∈ I let (φi, Hi, ξi) be the GNS
construction for µi. Then let φ :=Li∈I φi, a representation of Cu
0(bG) on H =Li∈I
Hi. By
0(bG),
a slight abuse of notation, we treat each ξi as a vector in H. Then for a ∈ Cu
i∈I kφi(a)ξi −bǫ(a)ξik2
i∈I(cid:2)hξi, φi(a∗a)ξii +bǫ(a∗a) − 2 Re(cid:10)bǫ(a)ξi, φi(a)ξi(cid:11)(cid:3)
i∈I(cid:2)µi(a∗a) +bǫ(a∗a) − 2 Re(cid:0)bǫ(a∗)µi(a)(cid:1)(cid:3)
= 2bǫ(a∗a) − 2 Rebǫ(a∗a) = 0.
Let (Q, ε) be a Kazhdan pair of type 3 for G. For δ > 0, if i ∈ I sufficiently large, ξi
is (Q, δε)-invariant. By Lemma 4.6 it follows that kξi − P (ξi)k ≤ δ, and ηi = P (ξi) is
0(bG),
invariant. Then, for a ∈ Cu
= lim
= lim
(b) =⇒ (c): Let (ai)i∈I be as in the hypotheses, so for each i ∈ I there is a state µi on
which follows as φ(a)∗ηi = φ(a∗)ηi =bǫ(a∗)ηi. Thus
µi(a) −bǫ(a) = hξi, φ(a)ξi −bǫ(a)ξii = hξi − ηi, φ(a)ξii − hξi − ηi,bǫ(a)ξii ,
µi(a) −bǫ(a) ≤ 2δkak,
i∈I−→bǫ in norm.
0(bG) with ai = (id ⊗ µi)(W ∗). Let ω ∈ L1(G) and set a = (ω ⊗ id)(W ∗); we note that
Cu
0(bG) and that such elements are norm dense in Cu
0(bG). By Cohen's factorisation
a ∈ Cu
theorem, there are elements ω′ ∈ L1(G) and x ∈ C0(G) with ω = xω′. Thus
As {µi : i ∈ I} is a bounded net, this shows that µi →bǫ in the weak∗ topology, and hence
by hypothesis, µi →bǫ in norm. It follows immediately that ai → 1 in norm, as claimed.
(c) =⇒ (a): Let U be a representation of G on H which has almost-invariant vectors.
Let U ∈ L(C0(G) ⊗ H) be the associated adjointable operator. By Lemma 2.4, there is a
net (ξi)i∈I of unit vectors with
ω′(aix) = ω′(x) = ω(1) = ω((id ⊗bǫ)(W∗)) =bǫ(a).
ω(ai) = lim
i∈I
µi(a) = lim
i∈I
so µi
lim
i∈I
lim
i∈I(cid:13)(cid:13)U(a ⊗ ξi) − a ⊗ ξi(cid:13)(cid:13) = 0
∀a∈C0(G).
32
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
lim
i∈I(cid:13)(cid:13)(id ⊗ ξ∗
For ξ ∈ H let (id ⊗ ξ∗) : C0(G) ⊗ H → H be the adjointable map defined on elementary
tensors by (id⊗ ξ∗) : a⊗ η 7→ hξ, ηia. As each ξi is a unit vector and U is unitary, it follows
that
i )U ∗(a ⊗ ξi) − a(cid:13)(cid:13) = 0
∀a∈C0(G).
Now, the relation between U and U implies that (id ⊗ ξ∗
i )U ∗(a ⊗ ξi) = (id ⊗ ωξi)(U ∗)a.
Thus, if ai := (id ⊗ ωξi)(U ∗), then ai is a normalised positive-definite element, and the net
(ai)i∈I is a left approximate identity for C0(G). By repeating the same argument with U
instead of U ∗ and then taking adjoints, we deduce that (ai)i∈I is also a right approximate
identity for C0(G). Thus by hypothesis, ai → 1 in norm.
For each i ∈ I let ηi be the orthogonal projection of ξi onto Inv(U). By combining
Lemma 2.12 with Proposition 2.13, we see that ηi is the unique vector of minimal norm in
the closed convex hull of
C = {(ω ⊗ id)(U ∗)ξi : ω ∈ L1(G) is a state}.
For ε > 0 choose i ∈ I sufficiently large so that kai − 1k < ε. Then, for ω a state in L1(G),
we see that as ξi is a unit vector,
Thus
(cid:13)(cid:13)(ω ⊗ id)(U ∗)ξi − ξi(cid:13)(cid:13)2 =(cid:13)(cid:13)(ω ⊗ id)(U ∗)ξi(cid:13)(cid:13)2 + 1 − 2 Rehξi, (ω ⊗ id)(U ∗)ξii
(cid:12)(cid:12)hξi, ξi − (ω ⊗ id)(U ∗)ξii(cid:12)(cid:12) = h1 − ai, ωi < ε.
≤ 2(cid:0)1 − Rehξi, (ω ⊗ id)(U ∗)ξii(cid:1)
≤ 2(cid:12)(cid:12)hξi, ξi − (ω ⊗ id)(U ∗)ξii(cid:12)(cid:12) < 2ε.
So every vector in C is at most √2ε from ξi, and hence the same is true of ηi. In particular,
if ε < 1/2 then ηi is non-zero, and invariant, showing that U has non-zero invariant vectors,
as required.
(cid:3)
We will now show that in fact we can have better control over the elements appearing
in the above theorem.
Lemma 5.2. Let (ai)i∈I be a net of Hilbert space contractions which does not converge to
1. Then (Re(ai))i∈I also does not converge to 1.
Proof. It is enough to show the statement for sequences, as its sequential version is easily
seen to be equivalent to the following fact:
∀ε>0 ∃δ>0 ∀a∈B(H) (kak ≤ 1,ka − 1k > ε) =⇒ k Re(a) − 1k > δ.
So suppose then that the sequential statement fails, so that we have ε > 0 and (an)∞
sequence of contractions such that kan− 1k > ε and k Re(an)− 1k ≤ 1
n=1 a
n for all n ∈ N. Then
ε < kan − 1k = k Re(an) − 1 + iIm(an)k ≤
1
n
+ kIm(an)k
AROUND PROPERTY (T) FOR QUANTUM GROUPS
33
so for n big enough kIm(an)k > ε
hξn, Im(an)ξni > ε
2 . But then
2. Thus there are vectors ξn ∈ H of norm 1 such that
1 ≥ hξn, anξni = hξn, ξni + hξn, (Re(an) − 1)ξni + ihξn, Im(an)ξni = 1 + αn + iβn,
where αn, βn are real numbers, αn ≤ 1
proof is finished.
n, βn > ε
2 . This is clearly contradictory and the
(cid:3)
Of course for normal operators it would have been enough to use the functional calculus.
This is how one can prove in elementary fashion the next lemma.
Lemma 5.3. Let (ai)i∈I be a net of selfadjoint Hilbert space contractions which does not
converge to 1. Then the net (of positive operators) (exp(ai − 1))i∈I does not converge to
1.
Proof. Via the spectral theorem it suffices to note the following elementary fact:
∀ε>0 ∃δ>0 ∀a∈R (a ≤ 1,a − 1 > ε) =⇒ exp(a − 1) − 1 > δ.
(cid:3)
We are now able to strengthen Theorem 5.1 under the assumption of triviality of the
scaling group.
Proposition 5.4. Let G be a locally compact quantum group with trivial scaling group.
Then the following conditions are equivalent:
(a) G has Property (T);
(b) if (ai)i∈I is a net of normalised positive-definite elements in M(C0(G))+ which con-
verges strictly to 1, then limi∈I kai − 1k = 0.
For the converse direction we begin by observing the following:
0(bG), and bG has a bounded antipode (as we assume here), then so is µ ◦bSu.
Proof. The implication (a)=⇒(b) is contained in Theorem 5.1.
if a is a normalised
positive-definite element of M(C0(G)), then so is a∗ -- this follows from the fact that if µ
is a state of Cu
Further also Re(a) is a normalised positive-definite element, as a convex combination of
elements of this type. Additionally, if a ∈ M(C0(G)) is a normalised positive-definite
element, then so is exp(a − 1) -- this time we use [LS3, Theorem 6.3] and [LS2, Theorem
3.7], which imply that if µ is a state of Cu
Assume then that G does not have Property (T). By Theorem 5.1 there exists a net
(ai)i∈I of normalised positive-definite elements of M(C0(G)) which converges strictly to 1,
but does not converge to 1 in norm. By the considerations above and Lemmas 5.2 and 5.3
the net (exp(Re(ai)− 1))i∈I is a net of normalised positive-definite elements of M(C0(G))+
which does not converge to 1 in norm. It is elementary to verify that it converges to 1
strictly.
(cid:3)
0(bG), then so is exp∗(µ −bǫ) =P∞
(µ−bǫ)∗k
k=0
k!
.
The following definition appears in [Ara] (for discrete quantum groups). Recall that a
functional in Cu
respect to the convolution product.
0(bG)∗ is said to be central if it commutes with all elements of Cu
0(bG)∗ with
34
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
Definition 5.5. A locally compact quantum group G has Central Property (T) if for any
net (µi)i∈I of central states of Cu
net (µi)i∈I actually converges in norm tobǫ.
It is clear from Theorem 5.1 that Property (T) implies Central Property (T). The follow-
ing result of Arano establishes the converse implication for discrete unimodular quantum
groups.
0(bG) converging in the weak∗ topology to the counitbǫ, the
Theorem 5.6 ([Ara, Proposition A.7]). A discrete unimodular quantum group has Property
(T) if and only if it has Central Property (T).
Note that for a discrete G a state µ ∈ Cu(bG)∗ is central if and only if all the matrices
µα (α ∈ Irr(bG) -- see comments after Remark 1.6) are scalar multiples of the identity, if
and only if the corresponding normalised positive-definite element (which in this case can
be identified simply with (µα)α∈Irr(G) ∈ M(c0(G))) is in the centre of M(c0(G)). Using this
fact we can easily show, following the lines above, the next corollary.
Corollary 5.7. Let G be a unimodular discrete quantum group. Then the following con-
ditions are equivalent:
(a) G has Property (T);
(b) if (ai)i∈I is a net of central normalised positive-definite elements in M(c0(G))+ which
converges strictly to 1, then limi∈I kai − 1k = 0.
The following result now extends [Kye, Theorem 5.1] by adding to it another equivalent
condition, namely (c) below.
Theorem 5.8. Let G be a second countable discrete quantum group. The following are
equivalent:
(a) G has Property (T);
generating functional.
(b) G is unimodular and Pol(bG) admits no unbounded generating functional;
(c) G is unimodular and Pol(bG) admits no central strongly unbounded bSu-invariant
Proof. (a) =⇒ (b): this is the implication (a) =⇒ (b) in [Kye, Theorem 5.1]; the first part
dates back to [Fim].
(b) =⇒ (c): trivial.
(c) =⇒ (a): suppose that (c) holds and yet (T) fails. Let (Kn)∞
n=1 be an increasing
proof of the implication (a4)=⇒(b3) in [Jol2], attributed there to [AkW].
Corollary 5.7 implies the existence of a net of central normalised positive-definite con-
tractions (ai)i∈I in M(c0(G))+ which is strictly convergent to 1, but not norm convergent
family of finite subsets of Irr(bG) such that S∞
n=1 Kn = Irr(bG). We will argue as in the
to 1. As Irr(bG) is assumed to be countable, we can replace this net by a sequence with
exists ǫ > 0 such that for each k ∈ N there exists l ≥ k and αl ∈ Irr(bG) such that
kInαl − aαl
k ∈ Mnα. There
l k ≥ ǫ. Similarly for each p ∈ N there exists N ∈ N such that for all k ≥ N we
the same properties. Consider the corresponding positive scalar matrices aα
AROUND PROPERTY (T) FOR QUANTUM GROUPS
35
have supα∈Kp kInα − aα
kk ≤ ǫ
4p . Using these two facts we can, by passing to a subsequence
constructed inductively, assume that in fact for each l ∈ N there exists αl ∈ Irr(bG) such
(5.1)
that
(5.2)
kInαl − aαl
α∈Kl kInα − aα
sup
l k ≥ ǫ,
l k ≤
ǫ
4l .
∞Xl=1
2l(Inα − aα
l ).
L =
∞Xl=1
2l(bǫ − µl),
Define for each α ∈ Irr(bG)
Lα =
Condition (5.2) assures that each Lα is well defined. For each l ∈ N we have Lαl ≥
2l(Inαl − aαl
l ) ≥ 0, and thus condition (5.1) implies that kLαlk ≥ 2lǫ. Finally it is easy to
verify that the functional L induced on Pol(bG) via matrices Lα is a generating functional,
as it is defined via (pointwise convergent) series
where µl denotes the state of Pol(bG) corresponding to al, which satisfies µl ◦bSu = µl. This
ends the proof, as L is clearly strongly unbounded, central and bSu-invariant.
6. Implication (T)1,1 =⇒ (T) for discrete unimodular quantum groups
(cid:3)
with low duals
In this section we establish one of the main results of the whole paper, that is we show
that for discrete unimodular quantum groups with low duals Property (T) is equivalent
to Property (T)1,1 (and in fact to Property (T)r,s for any r, s ∈ N). The key tool will be
Theorem 5.8, and we begin with a series of technical lemmas regarding the behaviour of
convolution semigroups of states at infinity implied by the properties of their generators.
Lemma 6.1. Let G be a discrete unimodular quantum group and let L be a central strongly
unbounded bSu-invariant generating functional on Pol(bG) (so that we have Lα = cαInα
for all α ∈ Irr(bG), where cα ∈ R+) and suppose that (γl)l∈N is a sequence of elements
of Irr(bG) such that we have Lγl = dlIMnγl
numbers increasing to infinity. Let α, β ∈ Irr(bG) and define for each l ∈ N the matrix
V (l) ∈ Mnα ⊗ Mnγl ⊗ Mnβ by the formula
(l ∈ N), with (dl)∞
l=1 a sequence of positive
V (l)
(i,j,k),(p,r,s) = L(cid:16)(uα
ip)∗uγl
jruβ
ks(cid:17) ,
i, p = 1, . . . , nα, j, r = 1, . . . , nγl, k, s = 1, . . . , nβ. Then all matrices V (l) are selfadjoint,
and there exists a sequence (el)∞
l=1 of positive numbers converging to infinity such that for
each l ∈ N
σ(V (l)) ⊆ {λ ∈ C : Re λ ≥ el}.
36
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
Proof. The matrices V (l) are selfadjoint, because L being bSu-invariant and bG being of Kac
type imply that for all suitable i, j, k, p, r, s,
(V (l)∗)(i,j,k),(p,r,s) = (V (l)
(p,r,s),(i,j,k)) = L((uα
rj)∗uα
sk)∗(uγl
pi)∗uγl
pi) = L((uα
sk) = L((uβ
jruβ
ip)∗uγl
sk)∗(uγl
rj)∗uα
pi)
ks) = V (l)
rjuβ
(i,j,k),(p,r,s).
= (L ◦bSu)((uβ
c(ab) = ρ(a)c(b) + c(a)bǫ(b).
The 'conditional' GNS construction for the generating functional L (originally due to
Schurmann, we refer for example to [DFSW, Subsection 7.2] for details) yields the existence
Recall that if ρ is a representation of Cu(bG) on a Hilbert space H, a linear map c :
Pol(bG) → H is a cocycle for ρ if it is a ρ −bǫ-derivation, i.e., for all a, b ∈ Pol(bG),
of a Hilbert space H, a representation ρ : Cu(bG) → B(H), and a cocycle c : Pol(bG) → H for
ρ such that for all a, b ∈ Pol(bG) we have
c is real as L is bSu-invariant, see [DFSW, Proposition 7.21, (7.10), and the succeeding
paragraph].) Expanding this equality and using the ρ −bǫ derivation property yields for all
a, b, d ∈ Pol(bG)
L(a∗bd) = L(a)bǫ(bd) +bǫ(a)L(bd) − hc(a), c(bd)i
L(a∗b) = L(a)bǫ(b) +bǫ(a)L(b) − hc(a), c(b)i.
(Remark that the scalars cl from the lemma's statement are indeed non-negative since
= L(a)bǫ(bd) +bǫ(a)(cid:16)L(b∗)bǫ(d) +bǫ(b∗)L(d) − hc(b∗), c(d)i(cid:17) − hc(a), ρ(b)c(d)i −bǫ(d)hc(a), c(b)i
= L(a∗)bǫ(bd) + L(b)bǫ(a∗d) + L(d)bǫ(a∗b) −bǫ(a∗)hc(b∗), c(d)i − hc(a), ρ(b)c(d)i −bǫ(d)hc(a), c(b)i.
Using now the fact that L is central and properties of the counit we obtain
V (l)
(i,j,k),(p,r,s) =
= δi,pδj,rδk,s(cα+cl+cβ)−δi,phc((uγl
jr)c(uβ
ks)i.
The trick is now based on expressing the last equality in a matrix form. We will do it
: Cnγl → H ⊗ Cnγl ,
ks)i−δk,shc(uα
jr)i−hc(uα
jr)∗), c(uβ
ip), ρ(uγl
ip), c(uγl
Tl(ej) =
step by step. Define the following Hilbert space operators: Tl,eTl
fTα : Cnα → H ⊗ Cnα, and Tβ : Cnβ → H ⊗ Cnβ :
nγlXa=1
nγlXa=1
eTl(ej) =
nαXb=1
fTα(ei) =
aj)∗) ⊗ ea,
ja) ⊗ ea,
bi) ⊗ eb,
c((uγl
c(uγl
c(uα
AROUND PROPERTY (T) FOR QUANTUM GROUPS
37
Tβ(ek) =
nβXb=1
c(uβ
kb) ⊗ eb,
for j ∈ {1, . . . , nγl}, i ∈ {1, . . . , nα}, k ∈ {1, . . . , nβ}, where (ej)
orthonormal bases of the respective spaces.
nγl
j=1, (ei)nα
i=1 and (ek)nβ
k=1 are
To simplify the notation we will now write Iα for IMnα , Il for IMnγl
and Iβ for IMnβ
and denote the Hilbert space tensor flips by Σ (with the position of the flipped arguments
understood from the context). Then we claim the following:
V (l) =(cα + cl + cβ)Iα ⊗ Il ⊗ Iβ − Iα ⊗(cid:16)(eTl
⊗ Iβ −h(fTα
⊗ Il)Σ(Iα ⊗ Tl)(cid:17)t
−(cid:16)(fTα
∗
∗
⊗ Iβ)Σ(Il ⊗ Tβ)(cid:17)t
⊗ Il ⊗ Iβ)Σ(Iα ⊗ ρ(nγl )(uγl)t ⊗ Iβ)(Iα ⊗ Il ⊗ Tβ)it
∗
,
where ρ(nγl )(uγl) is the unitary operator in Mnγl ⊗ B(H) obtained via the matrix lifting
of the representation ρ and interpreted as an operator from Cnγl ⊗ H to Cnγl ⊗ H, and t
denotes matrix transposition. Note that (uγl)t is unitary because G is of Kac type.
. It is equal to
∗
⊗ Iβ)Σ(Il ⊗ Tβ)(cid:17)t
(j,k),(r,s)
Analyse for example the expression(cid:16)(eTl
her ⊗ es, (eTl
hc((uγl
∗
⊗ Iβ)Σ(Il ⊗ Tβ)(ej ⊗ ek)i = h(eTl ⊗ Iβ)(er ⊗ es), Σ(Il ⊗ Tβ)(ej ⊗ ek)i
kb) ⊗ ej ⊗ ebi = hc((uγl
ar)∗) ⊗ ea ⊗ es, Σ(ej ⊗ c(uβ
ar)∗) ⊗ ea ⊗ es, c(uβ
kb) ⊗ eb)i
hc((uγl
=Xa,b
=Xa,b
jr)∗), c(uβ
ks)i,
and similar computations allow us to verify the 'matricial' formula for V (l) given above.
Recalling that the sequence (cl)∞
of the lemma it suffices to note that the norms of Tl and eTl are equal to (2cl)
compute for example
l=1 converges monotonically to infinity, to finish the proof
2 . Indeed,
1
(T ∗
l Tl)i,j = hTlei, Tleji =
hc(uγl
ia) ⊗ ea, c(uγl
jb) ⊗ ebi
nγlXa,b=1
=
nγlXa=1
hc(uγl
ia), c(uγl
= −L(δij 1) + L(uγl
nγlXa=1
ia)∗uγl
ja)i =
ij ) + L(uγl
(−L((uγl
ji) = 2clδi,j,
ja) +bǫ(uγl
ja)L(uγl
ia) + L(uγl
ia))
ja)bǫ(uγl
38
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
where we use the unitarity of (unγ )t and the fact that L(1) = 0. Similarly we compute
(eTl
∗eTl)i,j = heTlei,eTleji =
nγlXa,b=1
nγlXa=1
hc((uγl
ai)∗) ⊗ ea, c((uγl
bj)∗) ⊗ ebi
nγlXa=1
(cid:3)
=
ai))
ai(uγl
hc((uγl
ai)∗), c((uγl
aj)L(uγl
ai) + L(uγl
= −L(δij 1) + L(uγl
aj)∗) +bǫ(uγl
t > 0, πt ⋆ πt does not contain non-zero invariant vectors.
aj)∗)i =
ji) + L(uγl
(−L(uγl
ij ) = 2clδi,j.
Proof. Fix t > 0. Denote the GNS triple of µt by (πt, Ht, Ωt). Then the representation
Lemma 6.2. Let G be a second countable discrete unimodular quantum group with a
aj)bǫ(uγl
low dual and let L be a central strongly unbounded bSu-invariant generating functional on
Pol(bG). Consider the convolution semigroup of states (µt)t≥0 on Cu(bG) generated by L (see
Lemma 1.7). Let πt : Cu(bG) → B(Ht) be the corresponding GNS representation. Then for
πt ⋆ πt acts on the Hilbert space Ht ⊗ Ht and H0 := span{πt(a)Ωt ⊗ πt(b)Ωt : a, b ∈ Pol(bG)}
is in fact equal to 1). To this end consider ζ ∈ H0, ζ =Pk
k ∈ N, a1, . . . , ak, b1, . . . , bk ∈ Pol(bG), kζk = 1.
Let (γl)l∈N be a sequence of elements of Irr(bG) such that we have Lγl = clIMnγl
k(πt ⋆ πt)(zl)ζ −bǫ(zl)ζk2 = k(πt ⋆ πt)(zl)ζ − ζk2
is a dense subspace of Ht ⊗ Ht. It suffices then to show that the distance of the unit sphere
of H0 from the set of invariant vectors for πt ⋆ πt of norm 1 is non-zero (we will show it
i=1 πt(ai)Ωt ⊗ πt(bi)Ωt for some
with (cl)∞
interested in the following expression, calculated using Lemma 1.5 and (1.6):
l=1 a sequence of positive numbers increasing to infinity. Define zl := uγl
= kζk2 + k(πt ⋆ πt)(zl)ζk2 − 2 Rehζ, (πt ⋆ πt)(zl)ζi
≥ 1 − 2
Rehπt(aj)Ωt ⊗ πt(bj)Ωt, (πt ⋆ πt)(zl)[πt(ai)Ωt ⊗ πt(bi)Ωt]i
(l ∈ N),
11. We are
kXi,j=1
kXi,j=1
kXi,j=1
= 1 − 2
= 1 − 2
Rehπt(aj)Ωt ⊗ πt(bj)Ωt, (πt(uγl
r1) ⊗ πt(uγl
1r))[πt(ai)Ωt ⊗ πt(bi)Ωt]i
Re[µt(a∗
j uγl
r1ai)µt(b∗
j uγl
1rbi)].
nγlXr=1
nγlXr=1
r=1 µt(a∗
j uγl
r1ai)µt(b∗
j uγl
1rbi) is a finite (independent of l, as we work with fixed ai
and bi) linear combination of the terms of the form
The sumPnγl
µt((uα
ip)∗uγl
r1uβ
ks)µt((uα′
i′p′)∗uγl
1ruβ′
k′s′),
nγlXr=1
AROUND PROPERTY (T) FOR QUANTUM GROUPS
39
each of which can by (1.6) be in turn expressed as
(e−tV (l)
)(i,r,k),(p,1,s)(e−t(V ′)(l)
)(i′,1,k′),(p′,r,s′),
nγlXr=1
where the matrices V (l) and (V ′)(l) are of the type introduced in the last lemma.
Here we can finally use the lowness assumption: as there exists N such that nγl ≤ N
for all l ∈ N, it suffices to show that in fact each suitable matrix entry of the matricial
sequence (e−tV (l)) (indexed by l) tends to 0. We know however that this sequence even
converges to 0 in norm by Lemma 6.1.
(cid:3)
We are now ready to formulate the main result of this section, generalising the result of
Bekka and Valette [BV1], [BHV, Theorem 2.12.9] to discrete unimodular quantum groups
with low duals.
Theorem 6.3. Let G be a second countable discrete unimodular quantum group with a low
dual. Then G has Property (T) if and only if G has Property (T)1,1.
Proof. The forward implication does not require discreteness or unimodularity, and was
noted in Proposition 4.3.
Assume then that G does not have Property (T). We will show that there exists a weakly
mixing representation of G which has almost-invariant vectors. Indeed, by Theorem 5.8
there exists a central strongly unbounded bSu-invariant generating functional L on Pol(bG).
Let (µt)t≥0 be the convolution semigroup of states on Cu(bG) generated by L and for each
t > 0 let πt : Cu(bG) → B(Ht) denote the GNS representation of µt, and Ut the associated
representation of G. Since L is bSu-invariant, each of the states µt is also bSu-invariant. By
convergence of µt as t −→ 0+ we can verify that the representation V := Ln∈N Utn has
Lemma 1.4, each of the representations Ut satisfies condition R of Definition 1.3, and is
thus unitarily equivalent to its contragredient.
n=1 of positive real numbers convergent to 0. Then using pointwise
Choose a sequence (tn)∞
almost-invariant vectors (recall that the passage between U and π respects direct sums).
⊤ V c would not be ergodic.
We claim that V is weakly mixing. If that was not the case, V
⊤ U c
This implies that for some n, m ∈ N the representation Utn
tm is not ergodic. Hence,
⊤ Utn admits a non-zero invariant vector
Utn is not weakly mixing, that is, Utn
(see Section 2 for all this), and thus so does φUtn
. By Lemma 1.5 the last map is
⊤ Utn
exactly πtn ⋆ πtn, and Lemma 6.2 yields a contradiction.
(cid:3)
tn ∼= Utn
⊤ U c
Remark 6.4. In the above proof, since each of the representations Ut, t > 0, satisfies
condition R, the same is true for V . Therefore, for a discrete unimodular quantum group
that satisfies the assumptions of Theorem 6.3, Property (T) is equivalent to the following
weakening of Property (T)1,1: each of its representations which satisfies condition R and
has almost-invariant vectors is not weakly mixing.
Remark 6.5. Note that the low dual assumption was only used in the last part of the
proof of Lemma 6.2, and that in fact an analogous proof shows that for second countable
40
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
discrete unimodular quantum group with a low dual Property (T) is equivalent to Property
(T)r,s for any (equivalently, all) r, s ∈ N.
Observe also that the original definition of Property (T)r,s for locally compact groups in
[BV1, p. 294] translates in our setting to the following condition: for every representation
⊤ s has a non-
U of G such that U
zero invariant vector. Clearly, if U has almost invariant vectors, so does U c, and thus
⊤ s. Consequently, if G has Property (T)r,s in the sense of [BV1] as
so does U
explained above, it has Property (T)r,s in the sense of Definition 4.2. Therefore, for a
discrete unimodular quantum group that satisfies the assumptions of Theorem 6.3 and
r, s ∈ N, Property (T) is also equivalent to Property (T)r,s in the sense of [BV1].
⊤ s has almost invariant vectors, U
⊤ (U c)
⊤ (U c)
⊤ (U c)
⊤ r
⊤ r
⊤ r
As a corollary, we obtain the following generalisation of part of [KP, Theorem 2.5]
to discrete unimodular quantum groups with low duals.
It says that lack of Property
(T) is equivalent to the genericity of weak mixing for representations on a fixed infinite-
dimensional separable Hilbert space.
H.
H.
Theorem 6.6. Let G be a second countable discrete unimodular quantum group with a
low dual. Then G does not have Property (T) if and only if the weakly mixing represen-
If G has Property (T), then the ergodic
tations form a dense set in the space RepG
representations (hence also the weakly mixing ones) form a nowhere dense closed set in
RepG
Proof. If G fails (T) then G fails (T)1,1 by Theorem 6.3 and so there is a representation
U of G which has almost-invariant vectors but which is weakly mixing. Recall from Sub-
section 1.2 that weak mixing of representations of G is stable under tensoring by arbitrary
representations. In conclusion, by [DFSW, Lemma 5.3], the collection of weakly mixing
representations is dense in RepG
H, Un → U and all Un are ergodic.
Suppose that G has Property (T). Then the set of ergodic representations is closed (thus,
so is the set of weakly mixing representations, by Lemma 2.16). Indeed, let (Q, ε) be a
Kazhdan pair of type 1 for G (see Theorem 4.5). Suppose that (Un)∞
n=1 and U are in
If ξ is a unit vector invariant under U, then
RepG
H implies strong convergence on L2(G)⊗ H, there is n such that
since convergence in RepG
kUn(η ⊗ ξ) − η ⊗ ξk < ε for every η ∈ Q, contradicting the ergodicity of Un.
Furthermore, using a suitable Kazhdan pair of type 2 for G, we deduce that there exists
H all of whose elements are not ergodic. Assume by
a neighbourhood A1 of 1 in RepG
H contains an open subset A2.
contradiction that the set of ergodic representations in RepG
From Lemma 2.18 it follows that there exists U ∈ RepG
H whose equivalence class is dense
in RepG
(cid:3)
H. In particular, it intersects both A1 and A2, which is absurd.
H.
7. Spectral gaps
In this section we define the notion of a spectral gap for representations and actions of
locally compact quantum groups and relate it to ergodic properties of the relevant actions.
In conjunction with the results of the previous section it gives a characterisation of Property
(T) for a second countable discrete unimodular quantum group with a low dual in terms
AROUND PROPERTY (T) FOR QUANTUM GROUPS
41
of the relation between arbitrary invariant states and normal invariant states for actions
of G, which extends the analogous classical result of [LiN].
Definition 7.1. We say that a representation U of a locally compact quantum group has
spectral gap if the restriction of U to Inv(U)⊥ does not have almost-invariant vectors. An
action of a locally compact quantum group on a von Neumann algebra is said to have
spectral gap if its implementing unitary has spectral gap.
Evidently, a locally compact quantum group has Property (T) if and only if each of its
representations has spectral gap. Recall that given a representation U, pU denotes the
orthogonal projection onto the invariant vectors of U.
0 (bG). Then U does not have spectral
0(bG). Then U ∼= 1B(Inv(U )) ⊕ U1. By Lemma 2.4, U does
Lemma 7.2. Let U be a representation of a locally compact quantum group G on a Hilbert
space H and φU be the associated representation of Cu
gap if and only if there exists a state Ψ of B(H) satisfying (id⊗ Ψ)(U) = 1 and Ψ(pU ) = 0,
and if pU 6= 0, this is equivalent to the condition pU /∈ Im φU .
Proof. Write U1 for the restriction of U to Inv(U)⊥, and let φU1 = φU (·)Inv(U )⊥ be the
associated representation of Cu
not have spectral gap if and only if there exists a state Ψ of B(Inv(U)⊥) with (id ⊗
Ψ)(U1) = 1L∞(G). This is plainly the same as the existence of a state state Ψ of B(H) with
(id ⊗ Ψ)(U) = 1L∞(G) and Ψ(pU ) = 0.
Assume now that such a state Ψ exists. If pU ∈ Im φU , then 0 = Ψ(pU )pU = pU pU = pU
by Lemma 2.4. Hence pU = 0.
Conversely, if pU /∈ Im φU , consider the linear map j : Im φU → B(H), j(φU (a)) :=
follows from pU /∈ Im φU ) whose image is (Im φU )(1 − pU ), the subspace we identify with
Im φU1. Therefore, fixing a unit vector ζ ∈ Inv(U), the state Ψ1 := ωζ ◦ j−1 of Im φU1
U does not have spectral gap.
(cid:3)
0(bG). It is an injective ∗-homomorphism (injectivity
φU (a)(1−pU ) = φU (a)−bǫ(a)pU , a ∈ Cu
satisfies Ψ1 ◦ φU1 =bǫ. Hence, from Lemma 2.4, φU1 has almost-invariant vectors, namely,
Lemma 7.3. Let α be an action of a locally compact quantum group G on a von Neumann
algebra N ⊆ B(K). If α is implemented by a unitary V ∈ B(L2(G) ⊗ K) and a state Ψ of
B(K) satisfies (id ⊗ Ψ)(V ) = 1, then the restriction ΨN is invariant under α.
Proof. The argument is as in [BeT, Theorem 3.2]. Observe that for a state Ψ of B(K) as
above, V belongs to the multiplicative domain of the unital completely positive map id⊗Ψ
[Pau, Theorem 3.18]. Consequently,
(id⊗ΨN )α(x) = (id⊗Ψ)(V ∗(1⊗x)V ) = (id⊗Ψ)(V ∗)(id⊗Ψ)(1⊗x)(id⊗Ψ)(V ) = ΨN (x)1
for every x ∈ N.
(cid:3)
We are ready for the first of the main results of this section.
Theorem 7.4. Let G be a locally compact quantum group acting on a von Neumann
algebra N by an action α : N → L∞(G) ⊗ N. Write U ∈ M(C0(G) ⊗ K(L2(N))) for the
42
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
implementing unitary of α, and let φU be the associated representation of Cu
the following conditions:
0 (bG). Consider
(a) α has spectral gap;
(b) every state Ψ of B(L2(N)) with (id ⊗ Ψ)(U) = 1 satisfies Ψ(pU ) 6= 0;
(c) pU ∈ Im φU ;
(d) every state of N that is invariant under α is the weak∗-limit of a net of normal
states of N that are invariant under α.
Then (a) is equivalent to (b), and they are equivalent to (c) when pU 6= 0. If G is a
discrete quantum group, then (a) =⇒ (d), and (d) =⇒ (a) when dk·k(pU , N) < 1
2.
Proof. The statements about the implications between (a), (b) and (c) are just Lemma
7.2.
Assume henceforth that G is a discrete quantum group. For (a) =⇒ (d), suppose that
α has spectral gap and m is an α-invariant state of N. By Proposition 3.11, (a), there
is a net of unit vectors (ζi)i∈I in L2(N) that is almost invariant under U and such that
ωζi
to zero for U has spectral gap -- otherwise, by normalising and passing to a subnet, we get
a net in Inv(U)⊥ that is almost invariant under U. Therefore, letting ζ ′
kpU ζik pU ζi for
of normal states of N that are invariant
i∈I−→ m in the weak∗ topology of N ∗. The net(cid:0)(1 − pU )ζi(cid:1)i∈I in Inv(U)⊥ must converge
i(cid:1)i∈I
every i, m is the weak∗-limit of the net(cid:0)ωζ ′
i := 1
under α.
(d) =⇒ (b) when dk·k(pU , N) < 1
2: a vector ζ in the positive cone P of N in L2(N) is
invariant under U if and only if ωζ is invariant under α. Indeed, the forward implication
follows from the definitions for arbitrary ζ ∈ L2(N), while the backward implication follows
from Proposition 3.11, (b). If Ψ is a state of B(L2(N)) with (id ⊗ Ψ)(U) = 1, then from
Lemma 7.3 and (d), ΨN is the weak∗-limit of a net of normal states of N that are invariant
under α, say (ωζi)i∈I with ζi ∈ P for every i ∈ I. So for every i ∈ I, ζi is invariant under
2 . Since ΨN = weak∗−limi∈I ωζi,
we have Ψ(q) − 1 = limi∈I ωζi(q − pU ), hence Ψ(q) − 1 < 1
U, i.e., ωζi(pU ) = 1. Let q ∈ N be such that(cid:13)(cid:13)pU − q(cid:13)(cid:13) < 1
2 . We conclude that
(cid:12)(cid:12)Ψ(pU )(cid:12)(cid:12) ≥ 1 −(cid:12)(cid:12)Ψ(pU − q)(cid:12)(cid:12) − Ψ(q) − 1 > 1 −
1
2 −
1
2
= 0,
yielding (b).
Remark 7.5. In the proof of (a) =⇒ (d) we were limited to discrete quantum groups
only because of Proposition 3.11. However, Proposition 3.11 holds more generally than it
is stated (although we do not know precisely to what extent). This does not contradict
the example at the top of [LiN, p. 4919], explaining that the analogue of (a) =⇒ (d) fails
when α is the left translation action of the circle group, because the notion of invariance
of (non-normal) states under an action used in [LiN] is weaker than ours; see Remark 3.3.
Using our terminology, (a) =⇒ (d) holds trivially whenever G is a compact quantum group
and α := ∆.
(cid:3)
The next theorem characterises Property (T) of certain discrete unimodular quantum
groups in terms of their actions on von Neumann algebras.
AROUND PROPERTY (T) FOR QUANTUM GROUPS
43
Theorem 7.6. Let G be a second countable discrete unimodular quantum group with a low
dual. Then the following are equivalent:
(a) G has Property (T);
(b) for every action α of G on a von Neumann algebra N, every state Ψ of B(L2(N))
with (id ⊗ Ψ)(U) = 1 satisfies Ψ(pU ) 6= 0, where U is the unitary implementation
of α;
(c) for every action α of G on a von Neumann algebra N, every state of N that is
invariant under α is the weak∗-limit of a net of normal states of N that are invariant
under α;
(d) every action α of G on B(K), for some Hilbert space K, having no α-invariant
normal states of B(K), has no α-invariant states of B(K).
Proof. By the definition of Property (T) and Theorem 7.4, we have (a) =⇒ (b) =⇒ (c).
Also (c) implies (d).
(d) =⇒ (a): If G does not have Property (T), then by Theorem 6.3, G does not have
Property (T)1,1. Let then V ∈ M(c0(G) ⊗ K(K)) be a representation of G on a Hilbert
space K which has an almost-invariant net (ζi)i∈I of unit vectors but is weakly mixing.
Consider the action α of G on B(K) given by α(x) := V ∗(1 ⊗ x)V , x ∈ B(K). Then any
weak∗-cluster point of (ωζi)i∈I is an α-invariant state of B(K).
Nevertheless, α does not admit any normal invariant state. Indeed, by Lemma 3.13, the
⊥ V c ∈ M(c0(G)⊗K(K⊗ K)). Weak
(canonical) unitary implementation of α on K⊗ K is V
⊥ V c. By Proposition 3.11, (b), any normal
mixing of V is equivalent to ergodicity of V
⊥ V c-invariant unit vector in K ⊗ K. Hence, such a state
α-invariant state would give a V
cannot exist.
(cid:3)
8. A non-commutative Connes -- Weiss theorem
In this short last section we provide a non-commutative version of a classical result of
Connes and Weiss from [CoW], showing that for a certain class of discrete quantum groups
G, Property (T) may be characterised by the properties of the trace-preserving actions
of G on the von Neumann algebra VN(F∞). The key role in this result is played by the
construction of 'canonical actions' of quantum groups on the free Araki -- Woods factors,
due to Vaes.
Let us then recall the construction of Vaes [Va3, Section 3], yielding canonical actions,
induced by certain representations, of locally compact quantum groups on the free Araki --
Woods factors of Shlyakhtenko [Shl] (see also [Va2]). Let T be an involution on a Hilbert
space K, that is, a closed, densely-defined, injective anti-linear operator on K that satisfies
T = T−1, and let T = J Q1/2 be its polar decomposition. On the (full) Fock space
F (K) := CΩ ⊕
K⊗n
∞Mn=1
44
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
(the unit vector Ω is called the vacuum vector) consider the left creation (shift) operators
ℓ(ζ), ζ ∈ K, given by Ω 7→ ζ and K⊗n ∋ η 7→ ζ ⊗ η ∈ K⊗(n+1), and the operators
s(ζ) := ℓ(ζ) + ℓ(Tζ)∗,
ζ ∈ D(T).
The von Neumann algebra Γ(KJ , Qit)′′ := {s(ζ) : ζ ∈ D(T)}′′ on F (K) is called the free
Araki -- Woods von Neumann algebra associated with T. The vacuum vector Ω is generating
and separating for Γ(KJ , Qit)′′. Thus, the free quasi-free state ωΩ of Γ(KJ , Qit)′′ is faithful,
and the above representation of Γ(KJ , Qit)′′ on F (K) can be seen as its GNS representation
with respect to ωΩ.
When the dimension dim KJ of the real Hilbert space KJ := {ζ ∈ K : J ζ = ζ} is at
least 2, the von Neumann algebra Γ(KJ , Qit)′′ is a factor. In particular, when Q = 1 and
n := dim KJ ≥ 2, we have (Γ(KJ , Qit)′′, ωΩ) ∼= (VN(Fn), τ ), where Fn is the free group on
n generators, VN(Fn) is its group von Neumann algebra and τ is the (unique) tracial state
of VN(Fn).
Proposition 8.1 ([Va3, Proposition 3.1]). In the above setting, let U ∈ M(C0(G) ⊗ K(K))
be a representation of the locally compact quantum group G on K that satisfies
(ω ⊗ id)(U ∗)T ⊆ T(ω ⊗ id)(U ∗)
∀ω∈L1(G).
(8.1)
Then the representation of G on F (K) given by
∞Mn=0
⊥ 0 := 1 ∈ M(C0(G) ⊗ K(CΩ)) and U
F (U) :=
U
⊥ n,
where U
M(C0(G) ⊗ K(K⊗n)), n ≥ 1, induces an action α of G on Γ(KJ , Qit)′′ given by
Furthermore, the free quasi-free state is invariant under αU .
αU (x) := F (U)∗(1 ⊗ x)F (U),
(8.2)
⊥ n = U
⊥ U
= U1(n+1) · · · U13U12 ∈
n times
⊥ · · ·
{z
}
x ∈ Γ(KJ , Qit)′′.
We will need the following simple observation: when the assumptions of Proposition 8.1
hold, the implementing unitary of α on F (K) with respect to the vacuum vector Ω is F (U).
Indeed, by the foregoing and (3.4), one should show that
(id ⊗ ωyΩ,xΩ)(F (U)∗) = (id ⊗ ωyΩ,Ω)(α(x))
for every x, y ∈ Γ(KJ , Qit)′′. But this is clear from the definition of F (U).
Lemma 8.2. Let G be a locally compact quantum group with trivial scaling group and U
a representation of G on a Hilbert space K. Then for an involutive anti-unitary J on
K, U satisfies condition R of Definition 1.3 with respect to J if and only if U fulfils the
assumptions of Proposition 8.1 with T := J .
Proof. As before, write j for the ∗-anti-isomorphism on K(K) given by j(x) := J x∗J ,
x ∈ K(K). Now (R ⊗ j)(U) = U if and only if
(ω ⊗ ρ)(U) = (ω ◦ R)[(id ⊗ (ρ ◦ j))(U)]
∀ω∈L1(G)∀ρ∈B(K)∗,
AROUND PROPERTY (T) FOR QUANTUM GROUPS
45
while U satisfies (8.1) for T := J if and only if (ω ⊗ id)(U) = J (ω ⊗ id)(U)J for all
ω ∈ L1(G), if and only if
(ω ⊗ ρ)(U) = ω[(id ⊗ (ρ ◦ j))(U ∗)]
∀ω∈L1(G)∀ρ∈B(K)∗ .
Since the scaling group of G is trivial, we have S = R, so by (1.3) the right-hand sides of
the last two equations are equal.
(cid:3)
The following is a non-commutative analogue of the Connes -- Weiss theorem on discrete
groups [CoW], [BHV, Theorem 6.3.4], in which VN(F∞) := VN(Fℵ0) substitutes commu-
tative von Neumann algebras. Recall Definitions 3.4 and 3.7.
Theorem 8.3. Let G be a second countable discrete unimodular quantum group with a low
dual. Then the following conditions are equivalent:
(a) G has Property (T);
(b) for every von Neumann algebra N with a faithful normal state θ, every ergodic
θ-preserving action of G on N is strongly ergodic;
(b′) condition (b) is satisfied for (N, θ) = (VN(F∞), τ );
(c) for every von Neumann algebra N with a faithful normal state θ, every weakly mixing
θ-preserving action of G on N is strongly ergodic;
(c′) condition (c) is satisfied for (N, θ) = (VN(F∞), τ ).
Proof. (a) =⇒ (b): if (xι)ι∈I is a net of elements of N which is asymptotically invariant
and not trivial, then letting yι := xι − θ(xι)1, we get ηθ(yι) ∈ ηθ(1)⊥ for all ι and yι
X−→ 0
strongly. By the paragraph succeeding Definition 3.7, the restriction of the implementing
unitary of α to ηθ(1)⊥, which is ergodic by assumption, has almost-invariant vectors,
contradicting Property (T).
ι∈I
(b) implies (b′) and (c), and either of them implies (c′).
(c′) =⇒ (a): suppose that G does not have Property (T). By Theorem 6.3, G does not
have Property (T)1,1. So there exists a representation U ∈ M(c0(G) ⊗ K(K)) of G on a
Hilbert space K that satisfies condition R, has a net (ζi)i∈I of unit vectors that is almost-
invariant under U and is weakly mixing (see Remark 6.4). Since G is unimodular, the
assumptions of Proposition 8.1 are fulfilled with T being a suitable involutive anti-unitary
J on K by Lemma 8.2. Consider the induced action α := αU of G on the free Araki -- Woods
factor Γ(KJ , id)′′ ∼= VN(F∞) given by (8.2), and the canonical free quasi-free state (the
vector state ωΩ of the vacuum vector Ω), which is invariant under α, and which, in this
case, is just the canonical trace τ on VN(F∞).
Consider the bounded net (s(ζi))i∈I in Γ(KJ , id)′′. For every i ∈ I we have s(ζi)Ω = ζi,
X−→ 0 strongly. In the proof of [Va3, Proposition 3.1] it is
so that τ (s(ζi)) = 0 but s(ζi) i∈I
observed that for all ω ∈ ℓ1(G) and ζ ∈ K,
Therefore, by almost invariance of (ζi)i∈I under U, we have
(ω ⊗ id)α(s(ζ)) = s((ω ⊗ id)(U ∗)ζ).
(ω ⊗ id)α(s(ζi)) − s(ζi) = s((ω ⊗ id)(U ∗)ζi − ζi) i∈I−→ 0
46
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
in norm for every normal state ω of L∞(G).
invariant under α and is non-trivial, and thus α is not strongly ergodic.
In conclusion, (s(ζi))i∈I is asymptotically
we get the operator Y :=L∞
It is left to prove that the action α is weakly mixing. By the observation succeeding
Proposition 8.1, F (U) is the implementing unitary of α. Restricting F (U) to Ω⊥ ⊆ F (K),
⊥ (n−1)
⊥ (U
⊥ Y c). But U
is weakly mixing, which, since G is of Kac type, is equivalent to U
⊥ Z being ergodic for
⊥ Y c is ergodic, i.e., Y is weakly mixing
every representation Z of G. Consequently, Y
(again, G being of Kac type).
(cid:3)
⊥ Y c =L∞
⊥ n. Now Y
n=1 U
n=1 U
Note that to the best of our knowledge the equivalence between (a) and (c′) above is
new also for classical discrete groups.
Appendix
In this appendix we discuss the notion of a morphism between locally compact quantum
groups having dense image, as proposed in Definition 1.8. Begin by the following easy
observation (see [DFSW, Theorem 1.1]): for locally compact Hausdorff spaces X and Y ,
there is a one-to-one correspondence between continuous maps φ : X → Y and morphisms
Φ : C0(Y ) → Cb(X) given by Φ(f ) = f ◦ φ, f ∈ C0(Y ); a map φ has dense image if and
only if the associated Φ is injective. Considering this fact one might expect that for locally
compact quantum groups G and H, a morphism from H to G should be viewed as having
a dense image if the associated morphism π ∈ Mor(Cu
0(H)) is injective. This is
however not satisfactory, as there exists a lattice Γ in a Lie group G such that the natural
'inclusion' morphism in Mor(C∗(Γ), C∗(G)) is not injective ([BV2]) (and we would like to
view the natural morphism from bG to bΓ as having dense image; this is indeed the case --
This motivates the apparently more complicated Definition 1.8; in this appendix we
0(G), Cu
see below).
present its several equivalent characterisations.
0(G), Cu
0(H)), ω 7→ π((id ⊗ ω)WG), is injective;
(a) the morphism in question has a dense image in the sense of the Definition 1.8, i.e.
Theorem A.1. Let G, H be locally compact quantum groups and consider a morphism
from H to G represented by π ∈ Mor(Cu
0(H)) intertwining the respective coproducts,
with associated bicharacter V ∈ M(C0(H) ⊗ C0(bG)). The following are equivalent:
the map α : L1(bG) → M(Cu
(b) the map β : L1(bG) → M(C0(H)), ω 7→ ΛH(α(ω)) = (id ⊗ ω)(V ), is injective;
(c) AV := {(ζ ⊗ id)(V ) : ζ ∈ L1(H)} is dense in the weak∗ topology of L∞(bG);
0(bG)) has strictly dense range;
(d) the dual morphismbπ : Cu
0(bG)∗ → M(C0(H)), µ 7→ ΛHπ(cid:0)(id ⊗ µ)VVG(cid:1) is injective;
0(bH) → M(C0(bG)) has strictly dense range.
(f) the map Λ bGbπ : Cu
Let G, H be locally compact groups. In the commutative case, let a morphism from H to
G be represented by π ∈ Mor(C0(G), C0(H)) with associated continuous map φ : H → G.
As explained above, if φ has dense image, then the morphism has dense image in our sense.
Conversely, if φ does not have dense image, let H ′ := φ(H), and view VN(H ′) as a proper
0(bH) → M(Cu
(e) the map γ : Cu
AROUND PROPERTY (T) FOR QUANTUM GROUPS
47
weak∗-closed subspace (indeed, a von Neumann subalgebra) of VN(G) = L∞(bG). By
Hahn -- Banach separation, there is a function f 6= 0 in the Fourier algebra A(G) ∼= VN(G)∗
(see [Eym]), canonically embedded in C0(G), with f (H ′) = {0}. Hence π(f ) = f ◦ φ = 0.
Therefore, the morphism does not have dense image in our sense.
let a morphism from bG to bH be represented by π ∈
Mor(C∗(H), C∗(G)). The dual morphism bπ ∈ Mor(C0(G), C0(H)) is associated with a
continuous map from H to G, which is injective if and only if bπ has strictly dense image
(again, see [DFSW, Theorem 1.1]), if and only if the original morphism has dense image
in our sense by Theorem A.1.
In the cocommutative case,
We postpone the proof of Theorem A.1 to present first some technical results. Note
that the condition (c) in the above Theorem implies that the notion of the dense image
is compatible with the concept of the closure of the image of a quantum group morphism,
as defined recently in [KKS] -- indeed, according to (c) a morphism between H and G has
dense image in the sense of Definition 1.8 if and only if the closure of its image in the sense
of [KKS] is equal to G.
Let us now turn to look at the strict topology on multiplier algebras.
Lemma A.2. Let A, B be C∗-algebras and let θ : A → B be a surjective ∗-homomorphism.
Writing θ for the strict extension of θ, if X ⊆ M(A) is a strictly dense C∗-algebra, then
also θ(X) is strictly dense in M(B).
Proof. Easy consequence of the strict version of the Kaplansky density theorem.
Proposition A.3. Let G be a locally compact quantum group, let (xi)i∈I be a bounded
net in M(Cu
0(G)) be such that ΛG(xi) → ΛG(x) in L∞(G) for the
∗-strong topology. Fix ω0 ∈ L1(G) and set yi = (id ⊗ ω0 ◦ ΛG)∆G
0(G)) and
y = (id ⊗ ω0 ◦ ΛG)∆G
0(G)).
Proof. Recall that by [Ku1, Proposition 6.2],
i∈I−→ y in the strict topology on M(Cu
0(G)), and let x ∈ M(Cu
u (xi) ∈ M(Cu
u (x) ∈ M(Cu
0(G)). Then yi
(cid:3)
(id ⊗ ΛG)∆G
u (x) = W∗(1 ⊗ ΛG(x))W (x ∈ M(Cu
0(G))).
We may suppose that ω0 = ωξ0,η0 for some ξ0, η0 ∈ L2(G). Fix ε > 0, fix a ∈ Cu
0(G),
and choose a norm one compact operator θ0 ∈ K(L2(G)) with θ0(η0) = η0. As W ∈
M(Cu
j=1 bj ⊗ θj ∈ Cu
0(G) ⊗ K(L2(G)) with
0(G) ⊗ K(L2(G))) we can findPn
(cid:13)(cid:13)(cid:13)W(a ⊗ θ0) −
We may suppose that Cu
with kξkkηk = 1, i ∈ I,
nXj=1
bj ⊗ θj(cid:13)(cid:13)(cid:13) ≤ ε.
0(G) acts faithfully on a Hilbert space H. Then, for any ξ, η ∈ H
hξ, (yi − y)aηi = hξ ⊗ ξ0, W∗(1 ⊗ ΛG(xi − x))W(aη ⊗ θ0η0)i.
Up to an error not greater than εkξ0kkη0kkxi − xk, this is
nXj=1
hξ ⊗ ξ0, W∗(1 ⊗ ΛG(xi − x))(bjη ⊗ θjη0)i.
48
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
If i is sufficiently large, then for each j = 1, . . . , n we have that kΛG(xi − x)θjη0k ≤
εn−1kbjk−1, and so the absolute value of the sum is dominated by
kξ0kkbjkkΛG(xi − x)θjη0k ≤ kξ0kε.
nXj=1
Thus
hξ, (yi − y)aηi ≤ εkξ0k(kη0kkxi − xk + 1)
(cid:3)
Lemma A.2.
i ) → ΛG(x∗) strongly, shows that ka(yi −
i∈I−→ y in the strict topology, as required.
we have that (id⊗ω)(V ) = 0 if and only if ωAV = 0. The implication (b) =⇒ (a) is obvious.
with the adjoint of the reducing morphism Λ bG : Cu
and so k(yi − y)ak is small for sufficiently large i, as required.
An analogous argument, using that ΛG(x∗
y)k i∈I−→ 0 as well. Hence yi
Proof of Theorem A.1. The equivalence (b) ⇐⇒ (c) follows as for any element ω ∈ L1(bG)
The canonical embedding of L1(bG) into Cu
0(bG)∗ is the composition of the restriction map
0(bG) → C0(bG). Thus β is the restriction
of γ to L1(bG) (see (1.8)), and hence (e) =⇒ (b). The implication (d) =⇒ (f) follows from
0(bH), and as (id ⊗
If (d) holds, then as {(ω ◦ ΛH ⊗ id)VVH : ω ∈ L1(H)} is dense in Cu
bπ)VVH = (π ⊗ id)VVG, we see that
0(bG)∗∗. So for
{bπ((ω ◦ ΛH ⊗ id)VVH) : ω ∈ L1(H)} = {(ω ◦ ΛH ◦ π ⊗ id)VVG) : ω ∈ L1(H)}
0(bG)), and thus also weak∗-dense as a subspace of Cu
0 6= µ(cid:0)(ω ◦ ΛH ◦ π ⊗ id)VVG(cid:1) = γ(µ)(ω),
and so γ(µ) 6= 0. Hence (d) =⇒ (e) (so by the earlier reasoning (d) =⇒ (a)). In a similar
way we can prove that (f) =⇒ (b), arguing via a 'reduced' version of the map γ appearing
in (e).
0(bG)∗, there is ω ∈ L1(H) with
is strictly dense in M(Cu
non-zero µ ∈ Cu
12V ∗
12)∗V23W H
Now let µ ∈ Cu
12V13) = (ζ ⊗ ω0)((x ⊗ 1)V ) = (ζx ⊗ ω0)(V ) = 0,
23) = (ζ ⊗ ζ ′ ⊗ ω0)(W H
12 = (∆H ⊗ id)(V ) = V13V23 we see that for ζ ∈ L1(H)
Suppose now that (c) does not hold, so there is a non-zero ω0 ∈ L1(bG) with (ζ⊗ω0)(V ) =
0 for all ζ ∈ L1(H). As (W H
and ζ ′ ∈ B(L2(H))∗,
(ζ ⊗ ζ ′ ⊗ ω0)(V23W H
where in the above computation x = (id ⊗ ζ ′)(W H) ∈ C0(H).
find a net (ζi)i∈I in L1(H) such that (ζi ⊗ id)(W H) → x in the weak∗ topology. It follows
that for any ζ ′ ∈ B(L2(H))∗,
(µ ⊗ ζ ′ ⊗ ω0)(V23WH
As WH is a unitary, this also shows that that for any µ ∈ Cu
have
23) = 0.
0(H)∗ and set x = (µ ⊗ id)(WH) ∈ M(C0(bH)) ⊆ L∞(bH). Then we can
(ζi ⊗ ζ ′ ⊗ ω0)(V23W H
23) = 0.
0(H)∗ and ζ ′ ∈ B(L2(H))∗ we
23) = (ζ ′ ⊗ ω0)(V (x ⊗ 1)V ∗) = lim
12)∗V23WH
12V ∗
i∈I
12V ∗
12V ∗
(µ ⊗ ζ ′ ⊗ ω0)((WH
AROUND PROPERTY (T) FOR QUANTUM GROUPS
49
Let U = (π ⊗ id)(WG) so that V = (ΛH ⊗ id)(U), and so
(WH
12)∗V23WH
12 = (id ⊗ ΛH ⊗ id)(∆H
u ⊗ id)(U) = (id ⊗ ΛH ⊗ id)(U13U23) = U13V23.
0(H)∗,
Thus, for any state ζ ′ ∈ B(L2(H))∗, and any µ ∈ Cu
0 = (µ⊗ ζ ′⊗ ω0)(U13) = (µ⊗ ω0)(U) = µ((id⊗ ω0)(U)) = µ(π((id⊗ ω0)(WG))) = µ(α(ω0)).
Hence α(ω0) = 0 and so (a) does not hold. Thus (a) =⇒ (c).
Thus it remains to show (c) =⇒ (d). We assume that AV is weak∗-dense in L∞(bG), and
aim to show that the morphismbπ has strictly dense range. Let now
U := (ΛHπ ⊗ id)(VVG) = (id ⊗bπ)(WH)
denote the 'other-sided' (compared to U above) lift of the bicharacter V , so that U ∈
0(bG)) and (id ⊗ Λ bG)( U ) = V . As {(ζ ⊗ id)(W H) : ζ ∈ L1(H)} is dense in
M(C0(H) ⊗ Cu
0(bH), we see that
Cu
is dense in the image ofbπ. We conclude thatbπ has strictly dense range if and only if A U is
0(bG)). Notice
u )( U) = U13 U12, so applying (id ⊗ id ⊗ Λ bG) by (1.2) we see that V13 U12 =
strictly dense, and that the (norm) closure of A U is a C∗-subalgebra of M(Cu
also that AV = Λ bG(A U ).
A U := {(ζ ⊗ id)( U ) : ζ ∈ L1(H)} ⊆ M(Cu
0(bG))
23, so also U12 = V ∗
bG
13(W
bG
23)∗V13W
bG
23. It follows that
bG
Now, (id ⊗ ∆
bG
23)∗V13W
(W
bG
u (·).
again using that (W
bG)∗(1 ⊗ Λ bG(·))W
bG = (id ⊗ Λ bG)∆
0(bG). As (c) holds, the norm closure of
Pick ω1 ∈ L1(G) and set y = (ω1 ⊗ id)(WG) ∈ Cu
AV , which is a C∗-algebra, is weak∗-dense in L∞(bG). By the Kaplansky density theorem,
i∈I−→ Λ bG(y) in the ∗-strong topology. We
0(bG)) with Λ bG(yi) = xi for each i.
For any ω0 ∈ L1(bG), by Proposition A.3, we know that
we can find a bounded net (xi)i∈I in AV with xi
can 'lift' (xi)i∈I to find a bounded net (yi)i∈I in M(Cu
u (yi) i∈I−→ (id ⊗ ω0 ◦ Λ bG)∆
bG
(id ⊗ ω0 ◦ Λ bG)∆
bG
u (y)
in the strict topology on M(Cu
0(bG)).
As V is unitary, the closure of A U equals the closed linear span of
A U = {(ζ ⊗ id ⊗ µ)(cid:0)V ∗
{(ζ ⊗ id ⊗ µ)(cid:0)(W
bG
23)∗V13W
= {(id ⊗ µ)(cid:0)(W
13(W
bG
23)∗V13W
bG
23(cid:1) : µ ∈ B(L2(G))∗, ζ ∈ L1(H)}.
bG
23(cid:1) : µ ∈ B(L2(G))∗, ζ ∈ L1(H)}
bG)∗(1 ⊗ x)W
bG(cid:1) : µ ∈ B(L2(G))∗, x ∈ AV },
0(bG)), Λ bG(y) ∈ AV , ω ∈ L1(bG)},
and further is equal to the closed linear span of
u (y) : y ∈ M(Cu
bG
{(id ⊗ (ω ◦ Λ bG))∆
50
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
bG
By the above, we know that (id ⊗ ω0 ◦ Λ bG)∆
u (yi) is a member of the closure of A U for
each i ∈ I. Furthermore,
bG
bG
u )(WG)
u (y) = (ω1 ⊗ id ⊗ ω0 ◦ Λ bG)(id ⊗ ∆
(id ⊗ ω0 ◦ Λ bG)∆
WG
12)
13
= (ω1 ⊗ id ⊗ ω0 ◦ Λ bG)(WG
= (ω1z ⊗ id)(WG),
say, where z = (id ⊗ ω0 ◦ Λ bG)(WG) = (id ⊗ ω0)(W G) ∈ C0(G). Thus, as ω0 and ω1 vary,
bG
u (y) takes values in a dense subset of Cu
we see that (id ⊗ ω0 ◦ Λ bG)∆
strictly dense in M(Cu
We conclude that the strict closure of A U contains all of Cu
0(bG).
0(bG), and hence that A U is
(cid:3)
0(bG)), as required.
References
[AkW] C. A. Akemann and M. E. Walter. Unbounded negative definite functions. Canad. J. Math.,
[AD]
[Ara]
33(4):862 -- 871, 1981. 34
C. Anantharaman-Delaroche. On Connes' property T for von Neumann algebras. Math. Japon.,
32(3):337 -- 355, 1987. 1
Y. Arano. Unitary spherical representations of Drinfeld doubles. J. Reine Angew. Math. To ap-
pear, doi: 10.1515/crelle-2015-0079, arXiv:1410.6238. 2, 30, 33, 34
[BCT] E. B´edos, R. Conti, and L. Tuset. On amenability and co-amenability of algebraic quantum
groups and their corepresentations. Canad. J. Math., 57(1):17 -- 60, 2005. 2
[BMT] E. B´edos, G. J. Murphy, and L. Tuset. Co-amenability of compact quantum groups. J. Geom.
[BeT]
Phys., 40(2):130 -- 153, 2001. 9, 12
E. B´edos and L. Tuset. Amenability and co-amenability for locally compact quantum groups.
Internat. J. Math., 14(8):865 -- 884, 2003. 41
B. Bekka. Property (T) for C ∗-algebras. Bull. London Math. Soc., 38(5):857 -- 867, 2006. 3
[Bek]
[BHV] B. Bekka, P. de la Harpe, and A. Valette. Kazhdan's property (T), volume 11 of New Mathematical
Monographs. Cambridge University Press, Cambridge, 2008. 1, 2, 3, 39, 45
[BV1] M. E. B. Bekka and A. Valette. Kazhdan's property (T) and amenable representations. Math.
Z., 212(2):293 -- 299, 1993. 2, 3, 39, 40
[BV2] M. E. B. Bekka and A. Valette. Lattices in semi-simple Lie groups, and multipliers of group C ∗-
algebras. Ast´erisque, (232):67 -- 79, 1995. Recent advances in operator algebras (Orl´eans, 1992).
46
[BDS] M. Brannan, M. Daws, and E. Samei. Completely bounded representations of convolution algebras
of locally compact quantum groups. Munster J. Math., 6:445 -- 482, 2013. 6, 21
[ChN] X. Chen and C.-K. Ng. Property T for locally compact quantum groups. Internat. J. Math.,
26(3):1550024, 13 pp., 2015. 2, 13, 30
[Con1] A. Connes. A factor of type II1 with countable fundamental group. J. Operator Theory, 4(1):151 --
153, 1980. 1
[Con2] A. Connes. Classification des facteurs. In Operator algebras and applications, Part 2 (Kingston,
Ont., 1980), volume 38 of Proc. Sympos. Pure Math., pages 43 -- 109. Amer. Math. Soc., Provi-
dence, R.I., 1982. 1
A. Connes and V. Jones. Property T for von Neumann algebras. Bull. London Math. Soc.,
17(1):57 -- 62, 1985. 1
[CoJ]
[CoW] A. Connes and B. Weiss. Property T and asymptotically invariant sequences. Israel J. Math.,
37(3):209 -- 210, 1980. 3, 43, 45
AROUND PROPERTY (T) FOR QUANTUM GROUPS
51
[DaD]
B. Das and M. Daws. Quantum Eberlein compactifications and invariant means. Indiana Univ.
Math. J., 65(1):307 -- 352, 2016. 15, 16
[DFKS] B. Das, U. Franz, A. Kula, and A. Skalski. One-to-one correspondence between generating func-
tionals and cocycles on quantum groups in presence of symmetry. Math. Z., 281(3-4):949 -- 965,
2015. 10
[Daw] M. Daws. Completely positive multipliers of quantum groups. Internat. J. Math., 23(12):1250132,
23 pp., 2012. 6
[DFSW] M. Daws, P. Fima, A. Skalski, and S. White. The Haagerup property for locally compact quantum
groups. J. Reine Angew. Math., 711:189 -- 229, 2016. 2, 4, 10, 11, 12, 18, 36, 40, 46, 47
[DKSS] M. Daws, P. Kasprzak, A. Skalski, and P. M. So ltan. Closed quantum subgroups of locally
compact quantum groups. Adv. Math., 231(6):3473 -- 3501, 2012. 11
[DaS] M. Daws and P. Salmi. Completely positive definite functions and Bochner's theorem for locally
[DKS]
[Duv]
[EfR]
[Eym]
[Fim]
compact quantum groups. J. Funct. Anal., 264(7):1525 -- 1546, 2013. 6
S. Doplicher, D. Kastler, and E. Størmer. Invariant states and asymptotic abelianness. J. Funct.
Anal., 3:419 -- 434, 1969. 21
R. Duvenhage. A mean ergodic theorem for actions of amenable quantum groups. Bull. Aust.
Math. Soc., 78(1):87 -- 95, 2008. 21
E. G. Effros and Z.-J. Ruan. Discrete quantum groups. I. The Haar measure. Internat. J. Math.,
5(5):681 -- 723, 1994. 8
P. Eymard. L'alg`ebre de Fourier d'un groupe localement compact. Bull. Soc. Math. France,
92:181 -- 236, 1964. 47
P. Fima. Kazhdan's property T for discrete quantum groups. Internat. J. Math., 21(1):47 -- 65,
2010. 2, 34
[FMP] P. Fima, K. Mukherjee, and I. Patri. On compact bicrossed products. J. Noncommut. Geom. To
[G]
[Haa]
[Hal]
[HeT]
[Jad]
[Jol1]
[Jol2]
[KKS]
[Kaz]
[KP]
[KS]
[Ku1]
appear, arXiv:1504.00092. 2, 9
R. Godement. Les fonctions de type positif et la theorie des groupes. Trans. Amer. Math. Soc.,
63(1):1 -- 84, 1948. 16
U. Haagerup. The standard form of von Neumann algebras. Math. Scand., 37(2):271 -- 283, 1975.
23, 26
P. R. Halmos. In general a measure preserving transformation is mixing. Ann. of Math. (2),
45:786 -- 792, 1944. 3
R. H. Herman and M. Takesaki. States and automorphism groups of operator algebras. Comm.
Math. Phys., 19:142 -- 160, 1970. 21
A. Z. Jadczyk. On some groups of automorphisms of von Neumann algebras with cyclic and
separating vector. Comm. Math. Phys., 13:142 -- 153, 1969. 21
P. Jolissaint. Property T for discrete groups in terms of their regular representation. Math. Ann.,
297(3):539 -- 551, 1993. 3
P. Jolissaint. On property (T) for pairs of topological groups. Enseign. Math. (2), 51(1-2):31 -- 45,
2005. 3, 34
P. Kasprzak, F. Khosravi, and P. M. So ltan. Integrable actions and quantum subgroups. Int.
Math. Res. Not. IMRN. To appear, arXiv:1603.06084. 47
D. A. Kazdan. On the connection of the dual space of a group with the structure of its closed
subgroups. Funkcional. Anal. i Prilozen., 1:71 -- 74, 1967. 1
D. Kerr and M. Pichot. Asymptotic abelianness, weak mixing, and property T. J. Reine Angew.
Math., 623:213 -- 235, 2008. 3, 40
I. Kov´acs and J. Szucs. Ergodic type theorems in von Neumann algebras. Acta Sci. Math. (Szeged),
27:233 -- 246, 1966. 21
J. Kustermans. Locally compact quantum groups in the universal setting. Internat. J. Math.,
12(3):289 -- 338, 2001. 5, 6, 10, 11, 13, 47
52
[Ku2]
[KV1]
[KV2]
[Kye]
[KyS]
[Lan]
[LiN]
[LS1]
[LS2]
[LS3]
[Lub]
MATTHEW DAWS, ADAM SKALSKI, AND AMI VISELTER
J. Kustermans. Locally compact quantum groups. In Quantum independent increment processes.
I, volume 1865 of Lecture Notes in Math., pages 99 -- 180. Springer, Berlin, 2005. 4
J. Kustermans and S. Vaes. Locally compact quantum groups. Ann. Sci. ´Ecole Norm. Sup. (4),
33(6):837 -- 934, 2000. 4
J. Kustermans and S. Vaes. Locally compact quantum groups in the von Neumann algebraic
setting. Math. Scand., 92(1):68 -- 92, 2003. 4, 13
D. Kyed. A cohomological description of property (T) for quantum groups. J. Funct. Anal.,
261(6):1469 -- 1493, 2011. 2, 10, 30, 34
D. Kyed and P. M. So ltan. Property (T) and exotic quantum group norms. J. Noncommut.
Geom., 6(4):773 -- 800, 2012. 2
E. C. Lance. Hilbert C ∗-Modules. A Toolkit for Operator Algebraists, volume 210 of London
Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1995. 4, 8
H. Li and C.-K. Ng. Spectral gap actions and invariant states. Int. Math. Res. Not. IMRN,
(18):4917 -- 4931, 2014. 3, 12, 41, 42
J. M. Lindsay and A. G. Skalski. Quantum stochastic convolution cocycles. II. Comm. Math.
Phys., 280(3):575 -- 610, 2008. 10
J. M. Lindsay and A. G. Skalski. Convolution semigroups of states. Math. Z., 267(1-2):325 -- 339,
2011. 33
J. M. Lindsay and A. G. Skalski. Quantum stochastic convolution cocycles III. Math. Ann.,
352(4):779 -- 804, 2012. 33
A. Lubotzky. Discrete groups, expanding graphs and invariant measures, volume 125 of Progress
in Mathematics. Birkhauser Verlag, Basel, 1994. 21
[MRW] R. Meyer, S. Roy, and S. L. Woronowicz. Homomorphisms of quantum groups. Munster J. Math.,
[Moo]
[NY]
[Pau]
[Pet]
[PeP]
[PeJ]
[Pop1]
[Pop2]
[PoV]
[RSN]
[Run]
5:1 -- 24, 2012. 10, 11
C. C. Moore. Groups with finite dimensional irreducible representations. Trans. Amer. Math.
Soc., 166:401 -- 410, 1972. 9
S. Neshveyev and M. Yamashita. Drinfeld center and representation theory for monoidal cate-
gories. Comm. Math. Phys., 345(1):385 -- 434, 2016. 2
V. Paulsen. Completely bounded maps and operator algebras, volume 78 of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, Cambridge, 2002. 41
J. Peterson. Lecture notes on ergodic theory. http://www.math.vanderbilt.edu/~peters10/teaching/Spring2011/math390Blecturenotes.html.
12
J. Peterson and S. Popa. On the notion of relative property (T) for inclusions of von Neumann
algebras. J. Funct. Anal., 219(2):469 -- 483, 2005. 1, 3
S. Petrescu and M. Joit¸a. Property (T) for Kac algebras. Rev. Roumaine Math. Pures Appl.,
37(2):163 -- 178, 1992. 2
S. Popa. Correspondences. INCREST preprint, unpublished, 1986. 1
S. Popa. On a class of type II1 factors with Betti numbers invariants. Ann. of Math. (2),
163(3):809 -- 899, 2006. 1
S. Popa and S. Vaes. Representation theory for subfactors, λ-lattices and C∗-tensor categories.
Comm. Math. Phys., 340(3):1239 -- 1280, 2015. 2
F. Riesz and B. Sz.-Nagy. Functional analysis. Dover Books on Advanced Mathematics. Dover
Publications, Inc., New York, 1990. 16
V. Runde. Characterizations of compact and discrete quantum groups through second duals. J.
Operator Theory, 60(2):415 -- 428, 2008. 8
[RuV1] V. Runde and A. Viselter. Ergodic theory for quantum semigroups. J. Lond. Math. Soc. (2),
89(3):941 -- 959, 2014. 21
[RuV2] V. Runde and A. Viselter. On positive definiteness over locally compact quantum groups. Canad.
J. Math., 68(5):1067 -- 1095, 2016. 31
AROUND PROPERTY (T) FOR QUANTUM GROUPS
53
[Sau]
[Sch]
[Shl]
[So l]
[Str]
J.-L. Sauvageot. Impl´ementation canonique pour les co-actions d'un groupe localement compact
et les syst`emes dynamiques g´en´eralis´es. Math. Ann., 270(3):325 -- 337, 1985. 23, 24, 26
M. Schurmann. White noise on bialgebras, volume 1544 of Lecture Notes in Mathematics.
Springer-Verlag, Berlin, 1993. 10
D. Shlyakhtenko. Free quasi-free states. Pacific J. Math., 177(2):329 -- 368, 1997. 43
P. M. So ltan. Quantum Bohr compactification. Illinois J. Math., 49(4):1245 -- 1270, 2005. 13
S¸. Stratila. Modular theory in operator algebras. Abacus Press, Tunbridge Wells, England, 1981.
4
[Tak] M. Takesaki. Theory of operator algebras. II, volume 125 of Encyclopaedia of Mathematical Sci-
[Va1]
[Va2]
[Va3]
[VD1]
[VD2]
[Vis]
[Wor1]
[Wor2]
ences. Springer-Verlag, Berlin, 2003. 4
S. Vaes. The unitary implementation of a locally compact quantum group action. J. Funct. Anal.,
180(2):426 -- 480, 2001. 20, 21, 24
S. Vaes. ´Etats quasi-libres libres et facteurs de type III (d'apr`es D. Shlyakhtenko). Ast´erisque,
(299):Exp. No. 937, ix, 329 -- 350, 2005. S´eminaire Bourbaki. Vol. 2003/2004. 43
S. Vaes. Strictly outer actions of groups and quantum groups. J. Reine Angew. Math., 578:147 --
184, 2005. 3, 43, 44, 45
A. Van Daele. Discrete quantum groups. J. Algebra, 180(2):431 -- 444, 1996. 8
A. Van Daele. Locally compact quantum groups. A von Neumann algebra approach. SIGMA
Symmetry Integrability Geom. Methods Appl., 10:Paper 082, 41 pp., 2014. 4
A. Viselter. Weak mixing for locally compact quantum groups. Ergodic Theory Dynam. Systems.
To appear, doi: 10.1017/etds.2015.115, arXiv:1504.01292. 13
S. L. Woronowicz. Compact matrix pseudogroups. Comm. Math. Phys., 111(4):613 -- 665, 1987. 7
S. L. Woronowicz. Compact quantum groups. In Sym´etries quantiques (Les Houches, 1995), pages
845 -- 884. North-Holland, Amsterdam, 1998. 8
Leeds, United Kingdom
E-mail address: [email protected]
Institute of Mathematics of the Polish Academy of Sciences, ul. ´Sniadeckich 8, 00-656
Warszawa, Poland
E-mail address: [email protected]
Department of Mathematics, University of Haifa, 31905 Haifa, Israel
E-mail address: [email protected]
|
1811.00545 | 1 | 1811 | 2018-11-01T14:17:51 | A new type of Numerical radius of operators on Hilbert $C^*$-module | [
"math.OA",
"math.FA"
] | In this paper, we define a new concept of numerical range $W_{o}(\cdot)$ and prove its basic results. We also define the numerical radius $\omega_{o}(\cdot)$ and prove that $$\omega_{o}(T)\leq||| T|||\leq 2\omega_{o}(T).$$ | math.OA | math |
A NEW TYPE OF NUMERICAL RADIUS OF OPERATORS ON
HILBERT C ∗-MODULE
MARZIEH MEHRAZIN, MARYAM AMYARI AND MOHSEN ERFANIAN OMIDVAR∗
Abstract. In this paper, we define a new concept of numerical range Wo(·) and prove its
basic results. We also define the numerical radius ωo(·) and prove that
ωo(T ) ≤ 9T 9 ≤ 2ωo(T ).
1. Introduction and preliminaries
Suppose that B(H) is the set of all bounded linear operators on a complex Hilbert space
H equipped with the operator norm k · k. The numerical range and the numerical radius are
defined by
and
W (T ) = {hT x, xi : x ∈ H, kxk = 1},
ω(T ) = sup{hT x, xi : x ∈ H, kxk = 1},
respectively. In fact, ω(.) defines a norm on B(H).
It is known that
kT k = sup{hT x, yi : x, y ∈ H, kxk = kyk = 1}
for each T ∈ B(H), see [8, theorem 2.4.1]
If T ∈ B(H) is a self-adjoint operator, then
kT k = sup{hT x, xi : x ∈ H, kxk = 1}
see [2, Theorem 4.4.14]. In this case kT k = ω(T ).
Gustafson [6, theorem 1.3.1] showed that
ω(T ) ≤ kT k ≤ 2ω(T ).
2010 Mathematics Subject Classification. Primary 46L08; 47A12 .
Key words and phrases. Hilbert A -module; numerical range; numerical radius.
∗Corresponding author.
1
(1.1)
(1.2)
2
M. MEHRAZIN, M. AMYARI AND M. E. OMIDVAR
This result show that k.k and ω(.) are equivalent.
By using (1.1), Kittaneh [9, Theorem 1] proved that
1
4
kT ∗T + T T ∗k ≤ (ω(T ))2 ≤
1
2
kT ∗T + T T ∗k.
(1.3)
There are several numerical inequalities in the literatur related to inequalities above, see e.g.
[1, 3, 4, 5, 11, 12, 13].
In this paper, we define a new norm, a new concept of numerical range and a new notion
of numerical radius for operators on Hilbert A -modules, where A is an abelian C ∗-algebra.
We investigate the above inequalities in the framework of Hilbert A -modules.
Recall that a right pre-Hilbert C ∗-module E over a C ∗-algebra A (or a right pre Hilbert
A - module) is a linear space which is right A -module equipped with an A -valued inner
product h·, ·i : E × E → A that satisfies the following properties:
(i)hx, αy + βzi = αhx, yi + βhx, zi
(ii)hx, yai = hx, yia
(iii)hx, yi∗ = hy, xi
(iv)hx, xi ≥ 0; if hx, xi = 0 then x = 0
for each x, y, z ∈ E , a ∈ A and α, β ∈ C.
A pre Hilbert A -module which is complete with respect to the norm kxk = khx, xik
2 is
called a Hilbert C ∗-module over A , or a Hilbert A -module. Suppose that E and F are
Hilbert A-modules. We define L(E , F ) to be the set of all maps T : E → F for which there
is a map T ∗ : F → E such that hT x, yi = hx, T ∗yi for all x ∈ E , y ∈ F . It is known that T
must be a bounded A-linear map (that is, T is bounded linear map and T (xa) = T (x)a for
all x ∈ E , a ∈ A). If E = F , then L(E) is a C ∗-algebra together with the operator norm.
1
Suppose that A is an abelian C ∗-algebra. Recall that a character ϕ on A is a non-zero
∗-homomorphism ϕ : A → C such that kϕk = 1. We denote the set of all characters on A
by τ (A ).
2. Main results
In the rest of the paper we assume that A is an abelian C ∗-algebra. We start this section
with the following definition.
A NEW TYPE OF NUMERICAL RADIUS OF OPERATORS ON HILBERT C ∗-MODULE
3
Definition 2.1. Let T ∈ L(E).
9 T 9 :≡ sup{ϕ(T x) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1},
(2.1)
where x = hx, xi
1
2 .
First we show that 9 · 9 is a norm on E .
If T = 0, it is obvious that 9T 9 = 0.
If 9T 9 = 0, then for every ϕ ∈ τ (A ) and each x ∈ E such that ϕ(x) = 1, we have
ϕ(T x) = 0. We want to show that T x = 0 for each x ∈ E .
Fix x ∈ E,
(i) If ϕ(x) = 0, then by the Cauchy-Schwarz inequality we have
ϕ(hT x, T xi) = ϕ(hT ∗T x, xi) ≤ ϕ(hT ∗T x, T ∗T xi)
1
2 ϕ(hx, xi)
1
2 ,
thus ϕ(T x) = 0.
(ii) If ϕ(x) 6= 0, then by taking y = x
ϕ(x), we get ϕ(y) = 1. By definition 2.1, ϕ(T y) = 0
ϕ(x) ϕ(T x) = 0. Thus ϕ(T x) = 0. Since for every ϕ ∈ τ (A ), we have ϕ(T x) = 0.
and so
1
We conclude that T x = 0 for each x ∈ E . So T = 0.
On the other hand A is an abelian C ∗-algebra, then by [7, Theorem 3.6], x + y ≤ x + y
for each x, y ∈ E.
Thus
for each T, S ∈ L(E) and x ∈ E. Hence
T (x) + S(x) ≤ T (x) + S(x)
ϕ((T + S)x) = ϕ(T (x) + S(x) ≤ ϕ(T (x)) + ϕ(S(x))
Now by taking the supremum over x ∈ E and ϕ ∈ τ (A ) with ϕ(x) = 1, we get
9T + S9 ≤ 9T 9 + 9 S 9 .
Clearly 9αT 9 = α 9 T 9, for α ∈ C.
Remark 2.1. If E is Hilbert space, then
ϕ(T x) = ϕ(hT x, T xi
1
2 ) = ϕ(kT xk) = kT xkϕ(1) = kT xk.
Similary ϕ(x) = kxk. Hence kT k = 9T 9.
4
M. MEHRAZIN, M. AMYARI AND M. E. OMIDVAR
Theorem 2.2. If E is a Hilbert A -module, then
9T 9 = sup{ϕ(hT x, yi) : x, y ∈ E , ϕ ∈ τ (A ) & ϕ(x) = ϕ(y) = 1}.
Proof. Let β = sup{ϕ(hT x, yi) : x, y ∈ E , ϕ ∈ τ (A ) & ϕ(x) = ϕ(y) = 1}.
sufficient to prove that 9T 9 = β.
It is
If ϕ ∈ τ (A ) and x, y ∈ E , such that ϕ(x) = ϕ(y) = 1, then by using the Cauchy-Schwarz
inequality, we get
ϕ(hT x, yi) ≤ ϕ(hT x, T xi)
1
2 ϕ(hy, yi)
1
2
= (ϕ(T x)2)
1
2 (ϕ(y)2)
1
2
= ϕ(T x)
≤ 9T 9 .
Hence β ≤ 9T 9 .
For every ϕ ∈ τ (A ) and x ∈ E, with ϕ(x) = 1, we have
ϕ(T x)2 = ϕ(T x2) = ϕ(hT x, T xi) = ϕ(T x)ϕ(cid:18)hT x,
T x
ϕ(T x)
i(cid:19) ,
where we assume that ϕ(T x) 6= 0. Thus
ϕ(T x) = ϕ(cid:18)hT x,
T x
ϕ(T x)
i(cid:19)
≤ sup{ϕ(hT x, yi) : x, y ∈ E , ϕ ∈ τ (A ) & ϕ(x) = ϕ(y) = 1}.
Therefore, ϕ(T x) ≤ β. Hence 9T 9 ≤ β.
(cid:3)
Theorem 2.3. If T ∈ L(E) is self-adjoint, then
9 T 9 = sup{ϕ(hT x, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}.
(2.2)
Proof. Let M = sup{ϕ(hT x, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}. If ϕ ∈ τ (A ) and
T ∈ L(E) is self-adjoint, then by using the Cauchy-Schwartz inequality
ϕ(hT x, xi) ≤ ϕ(hT x, T xi)
1
2 ϕ(hx, xi)
1
2
= (ϕ(T x)2)
1
2 (ϕ(x)2)
1
2 .
A NEW TYPE OF NUMERICAL RADIUS OF OPERATORS ON HILBERT C ∗-MODULE
5
If ϕ(x) = 1, then
ϕ(hT x, xi) ≤ ϕ(T x)
ϕ(hT x, xi) ≤ sup{ϕ(T x) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
ϕ(hT x, xi) ≤ 9T 9 .
(2.3)
By taking supremum over ϕ(x) = 1, we get
M ≤ 9T 9 .
Conversely, let ϕ ∈ τ (A ) and x, y ∈ E. Then
ϕ(hT (x + y), x + yi) − ϕ(hT (x − y), x − yi) = 4ϕ(RehT x, yi).
Hence
ϕ(RehT x, yi) =
≤
=
1
4
1
4
1
4
+
ϕ(hT (x + y), x + yi) − ϕ(hT (x − y), x − yi)
ϕ(hT (x + y), x + yi) +
1
4
ϕ(hT (x − y), x − yi)
ϕ(x + y2)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ϕ(cid:18)h
ϕ(x − y2)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ϕ(cid:18)h
1
4
T (x + y)
ϕ(x + y)
,
x + y
ϕ(x + y)
T (x − y)
ϕ(x − y)
,
x − y
ϕ(x − y)
i(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
.
x+y
ϕ(x+y)(cid:12)(cid:12)(cid:12)
since ϕ((cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)
ϕ(h T (x+y)
ϕ(x+y),
whence
) = ϕ(x+y)
ϕ(x+y) = 1, we obtain
x+y
ϕ(x+y)i)(cid:12)(cid:12)(cid:12)
ϕ(h T (x−y)
ϕ(x−y),
≤ M and (cid:12)(cid:12)(cid:12)
≤ M,
x−y
ϕ(x−y)i)(cid:12)(cid:12)(cid:12)
ϕ(RehT x, yi) ≤
=
1
4
1
4
M(ϕ(x + y2) + ϕ(x − y2)) =
1
4
M ϕ(x + y2 + x − y2)
M ϕ(2x2 + 2y2) =
1
2
M ϕ(x2 + y2).
6
M. MEHRAZIN, M. AMYARI AND M. E. OMIDVAR
If y = T x
ϕ(T x) and ϕ(x) = 1, then
T x
ϕT x
(cid:12)(cid:12)ϕ(RehT x,
i)(cid:12)(cid:12) ≤
=
=
M
2
M
2
ϕ(cid:18)x2 +(cid:12)(cid:12)
ϕ (x2 +
2 ϕ(x2) +
M
T x
T x2
2(cid:19)
ϕ(T x)(cid:12)(cid:12)
ϕ(T x2)!
ϕ(T x2)!
ϕ(T x2)
Hence
= M.
(cid:12)(cid:12)(cid:12)(cid:12)
1
ϕ(T x)
ϕ(cid:0)RehT x, T xi(cid:1)(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)
1
ϕ(T x)
1
ϕ(T x)
Re(ϕ(T x2))(cid:12)(cid:12)(cid:12)(cid:12)
ϕ(T x2)(cid:12)(cid:12)(cid:12)(cid:12)
= ϕ(T x) ≤ M.
(cid:3)
3. Numerical range and Numerical radius
In this section, we define the numerical range and numerical radius for operators on L(E),
according to the definition of 9 · 9 on L(E).
Definition 3.1. Let T ∈ L(E). Then the numerical range of T is defined by
Wo(T ) = {ϕ(hT x, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}.
(3.1)
The next result represent some of the basic properties for the numerical range.
Theorem 3.2. If T, S ∈ L(E), then
(i) Wo(T ∗) = Wo(T ), where Wo(T ) is conjugate of Wo(T ).
(ii) If α, β ∈ C, then Wo(αT + βIE) = αWo(T ) + β.
(iii) If U ∈ L(E) is unitary, then Wo(U T U ∗) = Wo(T ).
(iv) Wo(T ) ⊆ R if and only if T is self-adjoint.
(v) Wo(T + S) ⊆ Wo(T ) + Wo(S).
A NEW TYPE OF NUMERICAL RADIUS OF OPERATORS ON HILBERT C ∗-MODULE
7
Proof. (i)
Wo(T ∗) = {ϕ(hT ∗x, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
= {ϕ(hx, T xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
= {ϕ(hT x, xi) : x ∈ E , ϕ ∈ (τ (A ) & ϕ(x) = 1}
= {¯λ : λ ∈ Wo(T )} = Wo(T ).
(ii) It is clear.
(iii) Since ϕ(x) = ϕ(U x) = 1, we have
Wo(U ∗T U) = {ϕ(hU ∗T U x, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
= {ϕ(hT U x, U xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
= {ϕ(hT y, yi) : y ∈ E , ϕ ∈ τ (A ) & ϕ(y) = 1}
= Wo(T ).
(iv) If T ∈ L(E) is self-adjoint, then
ϕ(hT x, xi) = ϕ(hx, T xi) = ϕ(hT x, xi)∗ = ϕ(hT x, xi),
which is equivalent to ϕ(hT x, xi) ∈ R.
Conversely, if ϕ(hT x, xi) ∈ R, then ϕ(hT x, xi) = ϕ(hT ∗x, xi) i.e. ϕ(h(T − T ∗)x, xi) = 0.
Hence h(T − T ∗)x, xi = 0 for every x ∈ E . Thus T = T ∗.
(v) Since
Wo(T + S) = {ϕ(h(T + S)x, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
= {ϕ(hT x, xi) + ϕ(hSx, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1},
and ϕ(hT x, xi) ∈ Wo(T ), ϕ(hSx, xi) ∈ Wo(S), we arrive at the result.
(cid:3)
Definition 3.3. Let T ∈ L(E). Then the numerical radius of T is defined by
ωo(T ) = sup{ϕ(hT x, xi) : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}.
(3.2)
It is easy to show that ωo(.) is a norm on L(E).
Lemma 3.4. If E is a Hilbert A -module, then for every ϕ ∈ τ (A ), x ∈ E ,
ϕ(hT x, xi) ≤ (ϕ(x2)ωo(T )
(3.3)
8
M. MEHRAZIN, M. AMYARI AND M. E. OMIDVAR
Proof. For each ϕ ∈ τ (A ) and x ∈ E we have
1
ϕ(x2)(cid:12)(cid:12)ϕ(hT x, xi)(cid:12)(cid:12) =(cid:12)(cid:12)ϕ(h
T x
ϕ(x)
,
x
ϕ(x)
i)(cid:12)(cid:12)
ϕ(hT x, xi) ≤ ωo(T ) =⇒ ϕ(hT x, xi) ≤ (ϕ(x2)ωo(T ),
hence
1
ϕ(x2)
since ϕ(
x
ϕ(x)) = 1.
In the next result, we show that 9 · 9 and ωo(·) are equivalent.
Theorem 3.5. If T ∈ L(E), then
ωo(T ) ≤ 9T 9 ≤ 2ωo(T ).
(cid:3)
(3.4)
Proof. For every ϕ ∈ τ (A ) and x ∈ E such that ϕ(x) = 1, by Theorem (2.2), we have
ϕ(hT x, xi) ≤ 9T 9,
By getting supremum, we obtain
Fix x, y ∈ E and ϕ ∈ τ (A ).
ωo(T ) ≤ 9T 9 .
(3.5)
4ϕ(hT x, yi) = ϕ(hT (x + y), x + yi − hT (x − y), x − yi
+ ihT (x + iy), x + iyi − ihT (x − iy), x − iyi)
≤ ϕ(hT (x + y), x + yi) + ϕ(h(x − y), x − yi)
+ ϕ(hT (x + iy), x + iyi) + ϕ(hT (x − iy, x − iyi),
Thus
ϕ(hT x, yi) ≤
=
=
1
4
+ ϕ(x + iy2)ωo(T ) + ϕ(x − iy2)ωo(T ))
1
4
1
4
(ωo(T )ϕ(2x2 + 2y2 + 2x2 + 2iy2))
= ωo(T )ϕ(x2 + y2).
(ϕ(x + y2)ωo(T ) + ϕ(x − y2)ωo(T )
(3.3)
ωo(T )(ϕ(x + y2) + ϕ(x − y2) + ϕ(x + iy2) + ϕ(x − iy2))
A NEW TYPE OF NUMERICAL RADIUS OF OPERATORS ON HILBERT C ∗-MODULE
9
If ϕ(x) = ϕ(y) = 1, then
Hence
(cid:12)(cid:12)ϕ(hT x, yi)(cid:12)(cid:12) ≤ 2ωo(T ).
9 T 9 ≤ 2ωo(T ).
(3.6)
(cid:3)
We use some similar strategies as in [9, Theorem 1] to prove the next result.
Theorem 3.6. If T ∈ L(E), then
1
4
9 T ∗T + T T ∗9 ≤ (ωo(T ))2 ≤
1
2
9 T ∗T + T T ∗ 9 .
(3.7)
Proof. Let T = M + iN, where M and N are self-adjoint and T ∗T + T T ∗ = 2(M 2 + N 2).
Let x ∈ E. From convexity of the function f (t) = t2, we have
ϕhT x, xi2 = ϕh(M + iN)x, xi2 = ϕhM x, xi + iϕhN x, xi2
= (ϕhM x, xi)2 + (ϕhN x, xi)2
≥
≥
=
=
1
2
1
2
1
2
1
2
(ϕhM x, xi + ϕhN x, xi)2
ϕhM x, xi ± ϕhN x, xi2
ϕ(hM x, xi ± hN x, xi)2
ϕhM ± N)x, xi2.
Hence
(ωo(T ))2 = sup{ϕhT x, xi2 : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
≥
=
1
2
1
2
sup{ϕh(M ± N)x, xi2 : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
9 (M ± N)2 9 .
10
So
Therefore
M. MEHRAZIN, M. AMYARI AND M. E. OMIDVAR
2(ωo(T ))2 ≥
≥
9 (M + N)2 9 +
1
2
9 (M − N)29
9 (M + N)2 + (M − N)29
1
2
1
2
= 9M 2 + N 29
=
1
2
9 T ∗T + T T ∗ 9 .
(ωo(T ))2 ≥
1
4
9 T ∗T + T T ∗ 9 .
To prove the right hand inequality, let ϕ ∈ τ (A) and x ∈ E such that ϕx = 1. From the
Cauchy -- Schwartz inequality, we have
ϕhT x, xi2 = (ϕhM x, xi)2 + (ϕhN x, xi)2
≤ ϕhM x, M xiϕhx, xi + ϕhN x, N xiiϕhx, xi
≤ ϕhM x, M xiϕx2 + ϕhN x, N xiiϕx2
= ϕhM x, M xi + ϕhN x, N xi
= ϕhM 2x, xi + ϕhN 2x, xi
= ϕh(M 2 + N 2)x, xi.
Hence
(ωo(T ))2 = sup{ϕhT x, xi2 : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
≤ sup{ϕh(M 2 + N 2)x, xi : x ∈ E , ϕ ∈ τ (A ) & ϕ(x) = 1}
= 9M 2 + N 29
=
1
2
9 T ∗T + T T ∗9,
which complete the proof.
(cid:3)
Example 3.7. Let X be a compact Hausdorff space and E = A = C(X). Then C(X) is a
Hilbert C(X)-module, such that hf, gi = ¯f g for each f, g ∈ C(X). Let ϕ ∈ τ (C(X)). Then
by Theorem 2.1.15 in [10], there exists x ∈ X such that ϕ = ϕx, where ϕx(f ) = f (x) for
A NEW TYPE OF NUMERICAL RADIUS OF OPERATORS ON HILBERT C ∗-MODULE
11
each f ∈ C(X).
Thus
9T 9 = sup{ϕ(T f ) : f ∈ A , ϕ ∈ τ (A ) & ϕ(f ) = 1}
= sup{T f (x) : f ∈ A , ϕ ∈ τ (A ) & f (x) = 1}
= sup{T f (x) : f ∈ A , ϕ ∈ τ (A ) & f (x) = 1}.
sup{ϕ(hT f, gi) : f, g ∈ A , ϕ ∈ τ (A ) & ϕ(f ) = ϕ(g) = 1}
= sup{ϕ((T f )g) : f, g ∈ A , ϕ ∈ τ (A ) & ϕ(f ) = ϕ(g) = 1}
= sup{T f g(x) : x ∈ E , f, g ∈ A , ϕ ∈ τ (A ) & f (x) = g(x) = 1}
= sup{T f (x)g(x) : x ∈ E , f, g ∈ A , ϕ ∈ τ (A ) & f (x) = g(x) = 1}
= sup{T f (x) : x ∈ E , f ∈ A , ϕ ∈ τ (A ) & f (x) = 1}
= sup{T f (x) : x ∈ E , f ∈ A , ϕ ∈ τ (A ) & f (x) = 1}.
Also
Hence
9T 9 = sup{ϕhT f, gi : f, g ∈ A , ϕ ∈ τ (A ) & ϕ(f ) = ϕ(g) = 1}
= sup{T f (x) : x ∈ E , f, g ∈ A , ϕ ∈ τ (A ) & f (x) = 1}.
If T is self-adjoint, then
9T 9 = sup{ϕ(hT f, f i) : f ∈ A , ϕ ∈ τ (A ) & ϕ(f ) = 1},
whence
ωo(T ) = 9T 9 .
References
[1] M. Bakherad and kh. Shebrawi, Upper bounds for numerical radius inequalities involving off-diagonal
operator matrices, Ann. Funct. Anal. 9 (2018), no. 3, 297 -- 309.
[2] F. F. Bonsall and J. Duncan, Numerical ranges of operators on normed spaces and of elements of normed
algebras, London Math. Soc. Lecture Note Series 2, Cambridge, 1971.
12
M. MEHRAZIN, M. AMYARI AND M. E. OMIDVAR
[3] S. S. Dragomir, Inequalities for the numerical radius of linear operators in Hilbert spaces, SpringerBriefs
in Mathematics. Springer, Cham, 2013.
[4] S. S. Dragomir, A survey of some recent inequalities for the norm and numerical radius of operators in
Hilbert spaces, Banach J. Math. Anal. 1 (2007), no. 2, 154 -- 175.
[5] R. Golla, On the numerical radius of a quaternionic normal operator, Adv. Oper. Theory 2 (2017), no.
1, 78 -- 86.
[6] K. E. Gustafson and D. K. M. Rao, Numerical range, Springer, New York, 1997.
[7] R. Jiang, A note on the tringular inequality for the C ∗-valued norm on the Hilbert C ∗-modules, Math.
Inequal. Appl. 16 (2013), no. 3, 743 -- 749.
[8] R. V. Kadison and J. R. Ringrose, Fundamentals of the theory of operator algebras, Academic Press
Inc, London, 1983.
[9] F. Kittaneh, Numerical radius inequalities for Hilbert space operators, Studia Math. 168 (2005), no. 1,
13 -- 80.
[10] G. J. Murphy, C ∗-algebras and Operator theory, Academic Press Inc, London, 1990.
[11] M. Sattari, M. S. Moslehian and T. Yamazaki, Some generalized numerical radius inequalities for Hilbert
space operators, Linear Algebra Appl. 470 (2015), 216 -- 227.
[12] J. Youqing and L. Bin, On operators with closed numerical ranges. Ann. Funct. Anal. 9 (2018), no. 2,
233 -- 245
[13] A. Zamani, Some lower bounds for the numerical radius of Hilbert space operators, Adv. Oper. Theory
2 (2007), no. 2, 98 -- 107.
Department of Mathematics, Mashhad Branch, Islamic Azad University, Mashhad, Iran
E-mail address: marzie [email protected]
E-mail address: maryam [email protected]; [email protected]
E-mail address: mn [email protected]
|
1910.13304 | 1 | 1910 | 2019-10-29T14:51:07 | Representations of C*-algebras of row-countable graphs and unitary equivalence | [
"math.OA"
] | In this article we show that there are branching systems (which induce representations of the graph algebra $C^*(E)$) associated to each row-countable graph $E$. For row-countable graphs, we characterize the condition $(L)$ via branching systems. Moreover, we show that each permutative representation in Hilbert spaces operators is unitarily equivalent to one induced by a branching system, even the spaces being not separable. Furthermore, under some hypothesis on the graph, we show that each representation of the graph C*-algebra is permutative. | math.OA | math |
Representations of C*-algebras of row-countable graphs and unitary
equivalence
Ben-Hur Eidt∗and Danilo Royer
October 30, 2019
Abstract
In this article we generalize the main results of [3] and [2]. More specifically, we show that there are
branching systems (which induce representations of the graph C∗(E)) associated to each row-countable
graph E. For row-countable graphs, we characterize the condition (L) via branching systems. Moreover,
we show that each permutative representation in Hilbert spaces operators is unitarily equivalent to one
induced by a branching system, even the spaces being not separable. Furthermore, under some hypothesis
on the graph, we show that each representation of the graph C*-algebra is permutative.
MSC 2010: 47L30
Keywords: Graph C∗-algebras, branching systems, representation theory, unitary equivalence.
Introduction
The concept of a graph C*-algebra was first developed, in [4], by considering row-finite countable graphs
(recall that a graph E = (E0, E1, r, s) is countable if E0 and E1 are both countable and is row-finite if s−1(v)
is finite for each vertex v), and have been extensively explored since then.
Ideas related to branching systems have been studied in some areas like random walks, symbolic dynamics,
scientific computing and operator theory, see [2] for references.
In this paper we deal with branching systems in row-countable graphs, that is, graphs with the property
that s−1(v) is at most countable for each vertex v. In [3] the authors define a structure called branching
system for graphs and show how to obtain a representation of C∗(E) through a branching system of a graph
E. Moreover, there is proved a result that ensures the existence of a branching system for all countable
graphs. We prove this theorem for a larger class of graphs, the row-countable graphs. In [2], it is proved
that each permutative representation ϕ : C∗(E) → B(H) (with H separable) is unitarily equivalent to a
representation arising from a branching system. We prove this result even H being not separable. Moreover,
in [2], the authors find a class of graphs where each representation ϕ : C∗(E) → B(H) (with H separable) is
permutative. We find a larger class where this result remains to be true.
The paper is organized as follows. In the first chapter we introduce branching systems and recall from [3]
how to obtain representations through this structure. After this we show how to obtain branching systems for
row-countable graphs, and for graphs of this class, we characterize the condition (L) via branching systems.
In the second chapter, we consider a permutative representations ϕ : C∗(E) → B(H) and show that this
representations are unitarily equivalent to one induced by a branching system, even H not being separable.
In the last chapter, we prove that for graphs in a certain class, each representation is permutative.
In this work, following [5], given an arbitrary graph E we define the algebra C∗(E) as being the universal
C∗-algebra generated by {Pv}v∈E0∪{Se}e∈E1 with the following relations: {Pv}v∈E0 are mutually orthogonal
projections, {Se}e∈E1 are partial isometries with orthogonal ranges and
(CK1) S∗
e Se = Pr(e) ∀e ∈ E1,
∗This author is partially supported by Conselho Nacional de Desenvolvimento Cient´ıfico e Tecnol´ogico - CNPq.
1
(CK2) Ps(e)SeS∗
(CK3) Pv = (cid:80)
e = SeS∗
SeS∗
e ∀e ∈ E1,
e provided that v ∈ E0 is such that 0 < #s−1(v) < ∞.
e∈s−1(v)
1 Representations arising from branching systems
In this section we define E-branching systems and recall from [3] a theorem that shows how obtain a
representation induced from a branching system. After that we prove that for row-countable graphs graphs
there always exists a branching system. For graphs in this class, we also characterize the condition (L) via
representations induced from branching systems.
Definition 1.1. [2.1:[3]] Let E = (E0, E1, r, s) be a graph, (X,M, µ) a measure space and {Re}e∈E1 ,
{Dv}v∈E0 a collection of measurable subsets of X such that:
1. Re ∩ Rf
2. Dv ∩ Dw
µ−a.e.
= ∅, ∀e, f ∈ E1 with e (cid:54)= f .
µ−a.e.
= ∅, ∀v, w ∈ E0 with v (cid:54)= w.
3. Re
4. Dv
µ−a.e.⊆ Ds(e), ∀e ∈ E1.
µ−a.e.
(cid:83)
=
e∈s−1(v)
Re, ∀v ∈ E0 such that 0 < #s−1(v) < ∞.
5. For each e ∈ E1 there exist functions fe : Dr(e) → Re, f−1
µ−a.e.
= Re, f−1
µ−a.e.
= IdDr(e) e fe ◦ f−1
6. For each e ∈ E1 there exist the Radon-Nikodym derivatives
that fe(Dr(e))
◦ fe
e
e
e
µ−a.e.
: Re → Dr(e), both measurable and such
= IdRe .
dµ ◦ fe
dµ
, denoted by Φfe and
dµ ◦ f−1
and
dµ
e
Φf
−1
e
, respectively . Furthermore,
(Φf
−1
e
◦ fe).Φfe = 1 = (Φfe ◦ f−1
e
).Φf
−1
e
µ − a.e.
The measure space (X,M, µ), with the collections {Re}e∈E1, {Dv}v∈E0 and the functions fe, f−1
, satisfying the items above is called an E-branching system.
e
Φf
−1
e
, Φfe ,
In the measure spaces Dr(e) and Re we consider the σ-algebras induced by X, moreover, since the Radon-
> 0 µ − a.e. Below we show
Nikodym derivative is a positive function, it follows from item 6 that Φfe , Φf
a sufficient condition that ensures the equality of item 6.
Proposition 1.2. Let (X,M, µ) a measure space and suppose that the items 1 until 5 from definition 1.1 are
. If for each e ∈ E1 the measures µ : Dr(e) → [0,∞] and µ : Re → [0,∞] are
satisfied and exist Φfe and Φ−1
µ − a.e. In particular, (X,M, µ) is a E-branching
semi-finite then (Φf
system.
◦ fe).Φfe = 1 = (Φfe ◦ f−1
).Φf
−1
e
−1
e
−1
e
fe
e
Proof. For each measurable set E,
◦ fe).Φfe dµ =
(Φf
−1
e
(cid:90)
(cid:90)
(cid:90)
E
=
(cid:90)
(cid:90)
χE.(Φf
−1
e
◦ fe).Φfe dµ =
χE.(Φf
−1
e
◦ fe) d(µ ◦ fe) =
Dr(e)
(χE ◦ f−1
e
).Φf
−1
e
dµ =
Dr(e)
(χE ◦ f−1
e
) d(µ ◦ f−1
e
) =
(cid:90)
(cid:90)
χE dµ =
1 dµ.
Re
Re
Dr(e)
E
Since µ is semi-finite then (Φf
−1
e
µ − a.e.
◦ fe).Φfe
µ−a.e.
= 1. The same argument shows that (Φfe ◦ f−1
e
).Φf
−1
e
= 1
(cid:4)
2
In the article [3], two questions were developed, the first is the connection between a branching system
and representations of C∗(E). The main result in this way is the following theorem.
Theorem 1.3. [2.2: [3]] Let E = (E0, E1, r, s) be a graph and (X,M, µ) a E-branching system. Then exists
a representation π : C∗(E) → B(L2(X,M, µ)) such that
.(φ ◦ f−1
π(Pv)(φ) = χDv .φ
π(Se)(φ) = χRe .Φ
e
)
e
1
2
−1
f
e
The second question is: given a graph E, there exists always an E-branching system? There is shown
that if E is countable then the answer is positive, that is, there always exists an E-branching system (see [3,
Theorem 3.1]). However, the hypothesis can be weakened and that is the main goal of this section. For this,
we need some preliminary results.
Let E = (E0, E1, r, s) be a graph. Let (0, 1] with the Borel σ-algebra and the Lebesgue measure and let
and Λ := E0 ∪ E1. Define for each A ⊆ (0, 1] × Λ and for each λ ∈ Λ the set A(λ) := {t ∈ (0, 1] (t, λ) ∈ A}
and define the collection M = {A ⊆ (0, 1] × Λ A(λ) is Borel measurable ∀λ ∈ Λ}. Moreover, define the
m(A(λ)) where m is the Lebesgue measure.
Proposition 1.4. ((0, 1] × Λ,M, µ) is a measure space.
Proof. The reader can check that M is a σ-algebra. Let {An}n∈N ⊆ M be a family of disjoints subsets and
function µ : M → [0,∞] given by µ(A) = (cid:80)
A = (cid:83)
An. We will consider two cases.
λ∈Λ
n∈N
Case 1 : Suppose that m(A(λ)) > 0 for a uncountable number of elements λ ∈ Λ. Then
(cid:33)
(cid:32)(cid:91)
n∈N
(cid:88)
λ∈Λ
µ
An
= µ(A) =
m(A(λ)) = ∞.
(cid:18)(cid:83)
On the other hand, there exist n0 ∈ N with the property that there exists an uncountable number of elements,
λ ∈ Λ, such that m(A(λ)
n0 ) > 0. In fact, otherwise, for each N, there exist at most a countable number of
n ) > 0 and then there exists a countable number of elements λ ∈ Λ such that
elements λ ∈ Λ such that m(A(λ)
n we conclude that m(A(λ)) > 0 for a countable number of elements
m
λ ∈ Λ, and this is a contradiction with the hypothesis assumed in the Case 1. Therefore,
> 0. Since A(λ) = (cid:83)
(cid:19)
A(λ)
n
A(λ)
n∈N
n∈N
and as (cid:80)
n∈N
µ(An) ≥ µ(An0) = ∞ then
(cid:88)
λ∈Λ
µ(An0) =
m(A(λ)
n0
) = ∞
(cid:88)
n∈N
(cid:33)
.
(cid:88)
λ∈Λ
An
n∈N
(cid:32)(cid:91)
(cid:33)(λ) =
(cid:88)
µ(An) = µ
(cid:32)(cid:91)
(cid:88)
n∈N
m
An
(cid:32)(cid:91)
µ
n∈N
=
(cid:88)
λ∈Λ
(cid:33)
(cid:88)
(cid:88)
=
An
λ∈Λ
n∈N
(cid:33)
A(λ)
n
=
(cid:32)(cid:91)
(cid:88)
n∈N
m
n∈N
Case 2 : Suppose that m(A(λ)) > 0 for, at most, a countable number of elements λ ∈ Λ. In this case,
m(A(λ)
n ) =
m(A(λ)
n ) =
µ(An).
n∈N
λ∈Λ
(cid:4)
This finishes the proof.
Definition 1.5. Let E = (E0, E1, r, s) be a graph. We say that E is row-countable if s−1(v) is countable for
every v ∈ E0.
3
We are now ready to prove the main theorem of this section.
Theorem 1.6. Let E = (E0, E1, r, s) be a row-countable graph. Then there exists an E-branching system.
Proof. Let ((0, 1] × Λ,M, µ) be the measure space as in the previous proposition. Note that if I ⊆ (0, 1] is
Borel-measurable and (λn)n∈N ⊆ Λ then I × {λ1, λ2, . . .} ∈ M. For each e ∈ E1 define Re = (0, 1] × {e}, let
W = {v ∈ E0 v is a sink }. For each v ∈ W define Dv = (0, 1] × {v} and for each v ∈ E0 − W define
(cid:91)
(cid:91)
Dv :=
Re =
e∈s−1(v)
e∈s−1(v)
(0, 1] × {e} = (0, 1] × s−1(v).
Now we verify the first four conditions of Definition 1.1, and after that we define the maps fe : Dr(e) → Re.
For e, d ∈ E1 with d (cid:54)= e it holds that
Rd ∩ Re = (0, 1] × {d} ∩ (0, 1] × {e} = (0, 1] × {e} ∩ {d} = ∅.
Fore v, u ∈ E0 with v (cid:54)= u there are three cases:
• If u, v ∈ W then
Dv ∩ Du = (0, 1] × {v} ∩ (0, 1] × {u} = (0, 1] × {v} ∩ {u} = ∅.
• If u ∈ W and v /∈ W then
Dv ∩ Du = (0, 1] × {v} ∩ (0, 1] × s−1(u) = (0, 1] × {v} ∩ s−1(u) = ∅.
• If u /∈ W and v /∈ W then
Dv ∩ Du = (0, 1] × s−1(v) ∩ (0, 1] × s−1(u) = (0, 1] × s−1(v) ∩ s−1(u) = ∅.
Moreover, given e ∈ E1, since s(e) is not a sink we have Re ⊆
0 < #s−1(v) < ∞, v is not a sink and then Dv = (cid:83)
Re.
e∈s−1(v)
Now we define the functions fd : Dr(d) → Rd for each d ∈ E1. Fix d ∈ E1.
• Case 1: r(d) ∈ W . In this case Dr(d) = (0, 1] × {r(d)} e Rd = (0, 1] × {d}. Define
Rd = Ds(e). If v ∈ E0 is such that
(cid:83)
d∈s−1(s(e))
fd : (0, 1] × {r(d)} → (0, 1] × {d}
(t, r(d)) (cid:55)→ (t, d).
◦ fd = IdDr(d) and
Clearly fd is a bijection and, if f−1
f (Dr(d)) = Rd. Note that fd are measurable: let B ∈ MRd then B = A ∩ Rd = A ∩ (0, 1] × {d} for
some A ∈ M and then B = A(d) × {d}. Therefore
is the inverse of fd then fd ◦ f−1
d = IdRd , f−1
d
d
d (A(d) × {d}) = A(d) × {r(d)} ∈ M.
f−1
d (B) = f−1
d (B) ∈ MDr(d) and then fd is measurable. The same argument shows that
d (B) ⊆ Dr(d) then f−1
is measurable.
As f−1
f−1
Moreover, as µ(Dr(d)) = µ((0, 1] × {r(d)}) = m(0, 1] = 1 < ∞ and µ ◦ fd(Dr(d)) = µ(Rd) = µ((0, 1] ×
{d}) = m(0, 1] = 1 < ∞ then the measures are σ-finite. If B ∈ MDr(d) is such that µ(B) = 0 then
B = A ∩ (0, 1] × {r(d)} = A(r(d)) × {r(d)} for some A ∈ M and then 0 = µ(B) = m(A(r(d)). Therefore
d
µ ◦ fd(B) = µ ◦ fd(A(r(d)) × {r(d)}) = µ(A(r(d)) × {d}) = m(A(r(d))) = 0
and then µ◦fd (cid:28) µ. By the Radon-Nikodym Theorem there exists
can be used for Φf
=: Φfd . The same argument
. By Proposition 1.2 we conclude that the sixth condition of Definition 1.1 is true.
−1
d
dµ ◦ fd
dµ
4
• Case 2: 0 < #s−1(r(d)) < ∞.
N(cid:83)
Write s−1(r(d)) = {e1, . . . , eN} for some N ∈ N. Note that (0, 1] =
Ij = ( j−1
N , j
(0, 1) → ( j−1
Ij where, for each j = 1, . . . , N
N ]. Fix j ∈ {1, . . . , N}. Let ϕj : (0, 1] → Ij be a homeomorphism such that ϕj(0,1) :
N , j
N ) is a diffeomorphism (for example, the linear increasing homeomorphism) and define
j=1
fd,j : (0, 1] × {ej} → Ij × {d}
(t, ej) (cid:55)→ (ϕj(t), d)
j (t), ej) for each t ∈ Ij. Moreover,
Notice that fd,j is a bijection with inverse given by f−1
fd,j is measurable (this follows from a similar argument that was used in Case 1 and from the fact that
ϕj is continuous). Define
d,j (t, d) = (ϕ−1
fd : (0, 1] × {e1, . . . , eN} → (0, 1] × {d}
(t, ej) (cid:55)→ fd,j(t, ej)
N(cid:83)
for each j = 1, . . . , N . Note that fd is bijetction; as (0, 1]×{e1, . . . , eN} = (0, 1]×{e1}∪. . .∪(0, 1]×{en}
and (fd)(0,1]×{ej} = fd,j for each j = 1, . . . , N the function fd is measurable. The same argument ensure
that f−1
is measurable.
d
i=1
((0, 1] × {ei}), then µ ((0, 1] × {ei}) = 1 for each i = 1, . . . , N and µ ◦ fd(Dr(d)) =
As Dr(d) =
µ((0, 1] × {d}) = 1. Therefore both the measures are σ-finite.
If B ∈ MDr(d) then B = A ∩ Dr(d) = A(e1) × {e1} ∪ . . . ∪ A(eN ) × {eN} for some A ∈ M. If B satisfies
µ(B) = 0 then
m(A(ei)) = 0 and so m(A(ei)) = 0 for each i = 1, . . . , N . Moreover,
N(cid:80)
i=1
fd(B) =
N(cid:91)
N(cid:91)
fd(A(ei) × {ei}) =
fd,i(A(ei) × {ei}).
For each i = 1, . . . , N ,
i=1
i=1
fd,i(A(ei) × {ei}) = {fd,i(t, ei) t ∈ A(ei)} = {(ϕi(t), d) t ∈ A(ei)}
and then
µ(fd,i(A(ei) × {ei})) = m(ϕi(A(ei))) = m(ϕi(A(ei) ∩ (0, 1) ∪ A(ei) ∩ {1})) =
= m(ϕi(A(ei) ∩ (0, 1)) + m(ϕi(A(ei) ∩ {1}))
Note that A(ei) ∩ (0, 1) ⊆ A(ei) is µ-null. As ϕi is a diffeomorphism, it follows from [6, pg. 153] that
m(ϕi(A(ei) ∩ (0, 1)) = 0. Clearly m(ϕi(A(ei) ∩ {1})) = 0. Therefore µ(fd,i(A(ei) × {ei})) = 0 for each
i = 1, . . . , N and this shows that µ(fd(B)) = 0. Therefore µ◦fd (cid:28) µ. By the Radon-Nikodym Theorem
and by Proposition 1.2 we conclude this case.
• Case 3: #s−1(r(d)) = ∞.
∞(cid:83)
By hypothesis s−1(r(d)) is countable and then we can write s−1(r(d)) = {e1, e2, . . .}. Moreover, (0, 1] =
j ]. Fix j ∈ N and let ϕj : (0, 1] → Ij a homeomorphism such that ϕj(0,1) :
Ij where Ij = ( 1
j−1 , 1
i=1
(0, 1) → ( 1
j−1 , 1
j ) is a diffeomorphism (for example, the linear homeomorphism). Define
fd,j : (0, 1] × {ej} → Ij × {d}
(t, ej) (cid:55)→ (ϕj(t), d)
5
and
fd : Dr(d) = (0, 1] × {e1, e2, e3, . . .} → (0, 1] × {d}
(t, ej) (cid:55)→ (ϕj(t), d).
The same arguments used in Case 2 can be used in Case 3 and so we get an E-branching system.
(cid:4)
Note that if a E graph is countable or row-finite then, in particular, E is row-countable. Therefore for
this kind of graphs we can ensure the existence of a branching system.
Corollary 1.7. Let E = (E0, E1, r, s) be a row-countable graph. Then, for every v ∈ E0 and e ∈ E1 it holds
that Pv (cid:54)= 0 and Se (cid:54)= 0.
Proof. As s−1(v) is countable for every v ∈ E0, let X a E-branching system. Note that the representation
π : C∗(E) → B(L2(X)) induced by Theorems 1.6 and 1.3 is such that π(Pv)(φ) = χDv .φ and by the proof
of Theorem 1.6 we get that µ(Dv) > 0. Therefore π(Pv) (cid:54)= 0 and then Pv (cid:54)= 0. From S∗
e Se = Pr(e) (cid:54)= 0 it
follows that Se (cid:54)= 0.
(cid:4)
As a consequence of the Corollary, we see that for every path α in E, the element Sα ∈ C∗(E) is nonzero,
moreover, if α and β are paths in E such that r(α) = r(β) then SαS∗
Corollary 1.8. Let E = (E0, E1, r, s) be a row-countable graph. Then E satisfies the condition (L) if and
only if every representation π : C∗(E) → B(L2(X)) induced from a branching system is faithful.
Proof. For the direct implication, as π(Pv) and Pv are nonzero elements, the result follows from [1, Theorem
(2)]. For the converse, let's show that if α = e1 . . . en is a path without exit then exists a representation ϕ
such that ϕ is not injective. Let X be the branching system as in the proof of Theorem 1.6. We redefine
only the maps fei for 1 ≤ i ≤ n. From the proof of Theorem 1.6 we get Rei = (0, 1] × {ei} and Dr(ei) =
(0, 1] × s−1(r(ei)) = (0, 1] × {ei+1} for each 1 ≤ i ≤ n − 1 and Dr(en) = (0, 1] × s−1(r(en)) = (0, 1] × {e1}.
√
t, e1) and for i = n, define fen : Dr(en) → Ren
For i = 1, define fe1 : Dr(e1) → Re1 by fe1(t, e2) = (
√
t, en). For each i = 2, . . . , n − 1 define fei : Dr(ei) → Rei by fei (t, ei+1) = (t, ei). So
by fen(t, e1) = (
we get a new E-branching system. For this branching system, by Theorems 1.3 we get a representation
ψ : C∗(E) → B(L2(X)) such that ψ(Se1 . . . Sen) (cid:54)= ψ(Ps(e1)). In particular, Se1 . . . Sen (cid:54)= Ps(e1).
t, e1), fen (t, e1) = (t2, en) and fei(t, ei+1) = (t, ei) for each
1 < i < n, getting another E-branching system Y . Let ϕ : C∗(E) → B(L2(Y )) be the induced representation.
For φ ∈ L2(Y ) it holds that
Moreover, we can choose fe1 = (t, e2) = (
β (cid:54)= 0.
√
ϕ(Se1 . . . Sen )(φ) = χRe1
.(Φf
−1
en ◦...◦f
−1
e1
).(φ ◦ f−1
en
◦ . . . ◦ f−1
e1
) =
and then ϕ(Se1 . . . Sen ) = ϕ(Ps(e1)). Since Se1 . . . Sen (cid:54)= Ps(e1) then ϕ is not injective.
= χRe1
.φ = χDr(en) .φ = χDs(e1) .φ = ϕ(Ps(e1))(φ)
(cid:4)
2 Unitary equivalence and permutative representations
In this section we show that each permutative representation of graph C*-algebras in Hilbert spaces
operators, even the spaces being non-separable, are unitarily equivalent to representations induced from
branching systems. Moreover, we find a class of graphs where each representation is permutative.
Let ϕ : C∗(E) → B(H) a representation of C∗(E). The relations that define the universal C*-algebra
C∗(E) and the fact that ϕ is a ∗-homomorphism ensure that {ϕ(Pv)}v∈E0 and {ϕ(Se)ϕ(S∗
e )}e∈E1 are families
of mutually orthogonal projections. For each edge e and vertex v let
Hv := ϕ(Pv)(H)
He = ϕ(Se)ϕ(Se)∗(H).
As ϕ(Pv) and ϕ(Se)ϕ(Se)∗ are projections, Hv and He are closed subspaces of H. Moreover, it holds
that:
6
1. If v (cid:54)= w then Hv ∩ Hw = {0H},
2. If e (cid:54)= f then He ∩ Hf = {0H},
3. The restriction of ϕ(Se) given by ϕ(Se) : Hr(e) → He is a surjective, isometric and unitary operator,
He and if #s−1(v) = ∞ then we may write Hv =
(cid:32) (cid:76)
He
e∈s−1(v)
(cid:33)(cid:76) Vv
4. If 0 < #s−1(v) < ∞ then Hv = (cid:76)
(cid:32) (cid:76)
(cid:32) (cid:76)
e∈s−1(v)
v∈E0
He
Hv
.
e∈s−1(v)
(cid:33)⊥
(cid:33)(cid:76) V where V =
(cid:18)(cid:83)
(cid:19)
Aλ
where Vv =
5. We write H =
Recall that (cid:76)
(cid:32) (cid:76)
v∈E0
(cid:33)⊥
Hv
.
Bλ and
(cid:4)
λ∈Λ
λ∈Λ
provided that {Aλ}λ∈Λ is a family of mutually orthogonal subspaces
of a Hilbert space (that's the case). The proof of properties above follows from the relations that define C∗(E).
Other interesting fact is the relation between total orthonormal sets in Aλ with total orthonormal sets in
Aλ := span
Aλ; this relation is given by the following proposition.
λ∈Λ
Proposition 2.1. Let X be a Hilbert space and {Aλ}λ∈Λ a collection of mutually orthogonal subspaces of
Bλ is a total orthonormal set in
(cid:76)
X. If for each λ ∈ Λ, Bλ ⊆ Aλ is a total orthonormal set in Aλ then (cid:83)
H := (cid:76)
Proof. It is easy to see that (cid:83)
Bλ is an orthonormal set. We show that (cid:83)
Fix h ∈ H and > 0. As H = (cid:76)
Aλ, there exists a = aλ1 + . . . + aλn ∈ span (cid:83)
Aλ (where n ∈ N and
Bλ is total in H.
λ∈Λ
λ∈Λ
λ∈Λ
λ∈Λ
Aλ.
λ∈Λ
λi ∈ Aλi for each i = 1, . . . , n) such that (cid:107)h − a(cid:107) <
exists bλi ∈ span Bλi such that (cid:107)aλi − bλi(cid:107) <
moreover
2n
2
. Define b = bλ1 + . . . + bλn , note that b ∈ span (cid:83)
. As Bλi is a total orthonormal set in Aλi then there
λ∈Λ
λ∈Λ
(cid:107)h − b(cid:107) ≤ (cid:107)h − a(cid:107) + (cid:107)a − b(cid:107)
<
2
≤
2
2
<
+ (cid:107)aλ1 − bλ1 + . . . + aλn − bλn(cid:107)
+ (cid:107)aλ1 − bλ1(cid:107) + . . . + (cid:107)aλn − bλn(cid:107)
+ n
2n
= .
We show some consequences of these properties.
Proposition 2.2. Let E be a graph and ϕ : C∗(E) → B(H) a representation.
1. For each x ∈ C∗(E), ϕ(x) vanishes at V (where V is as above).
2. If H is separable then ϕ(v) (cid:54)= 0 for, at most, a countable number of vertices v ∈ E0.
(cid:32) (cid:76)
Hv
v∈E0
(cid:33)(cid:76) V . Let y ∈ V and µ, ν paths in E such that
Proof. First we prove 1. We know that H =
r(µ) = r(ν). If a = SµS∗
ν then
(cid:107)ϕ(a)(y)(cid:107)2 = (cid:104)ϕ(a)(y), ϕ(a)(y)(cid:105) = (cid:104)y, ϕ(a∗a)(y)(cid:105) =
7
= (cid:104)y, ϕ(SνS∗
µSµS∗
ν )(y)(cid:105) = (cid:104)y, ϕ(Ps(ν))ϕ(SνS∗
µSµS∗
ν )(y)(cid:105) = 0.
the the continuity of the inner product in H.
total orthonormal subset Bv ⊆ Hv. Then (cid:83)
H is separable, (cid:76)
As each element of C∗(E) may be approximated by elements of the form SµSν the result follows from
Now we prove 2. If ϕ(v) (cid:54)= 0 for a uncountable number of vertices v ∈ E0 we choose, for every Hv, a
Hv. Since
Hv ⊆ H is separable. This is a contradiction because every total orthonormal subset of
(cid:4)
a separable Hilbert space is countable.
Definition 2.3. Let ϕ : C∗(E) → B(H) a representation. We say that ϕ is permutative if for every e ∈ E1,
v ∈ E0 there exist total orthonormal sets Be ⊆ He and Bv ⊆ Hv such that
Bv is a uncountable total orthonormal set of (cid:76)
v∈E0
v∈E0
v∈E0
• If e ∈ s−1(v) then Be ⊆ Bv,
• If 0 < #s−1(v) < ∞ then Bv = (cid:83)
Be,
e∈s−1(v)
• ϕ(Se)(Br(e)) = Be.
As ϕ(Se) : Hr(e) → He is unitary, the third condition of the definition above is equivalent to Br(e) =
ϕ(Se)∗(Be). Furthermore, ϕ(Se)(Br(e)) is always a total orthonormal set in He because ϕ(Se) : Hr(e) → He
is isometric and surjective.
Example 2.4. Here we show an example of a permutative representation. Let E be the graph as follows.
e2
e1
.
.
v
. .
ek
For each i = 1, . . . , k define the operator Ui : l2 → l2 given by
Ui ((xn)n∈N) = (0, . . . , 0,
i(cid:122)(cid:125)(cid:124)(cid:123)x1 , 0, . . . , 0,
i+k(cid:122)(cid:125)(cid:124)(cid:123)x2 , 0, . . . , 0,
i+2k(cid:122)(cid:125)(cid:124)(cid:123)x3 , . . .),
and notice that
U∗
i ((yn)n∈N) = (yi, yi+k, yi+2k, . . .).
By the universal property of C∗(E) there exists a ∗-homomorphism ϕ : C∗(E) → B(l2) such that ϕ(Sei) =
Ui and ϕ(Pv) = Id. Let's show that ϕ is permutative. As
UiU∗
i (y) = (0, . . . , 0,
i(cid:122)(cid:125)(cid:124)(cid:123)yi , 0, . . . , 0,
i+k(cid:122)(cid:125)(cid:124)(cid:123)yi+k, . . .))
then
where (δn)n∈N is the canonical basis of l2. For each i = 1, . . . , k choose Bei = {δi, δi+k, δi+2k . . .} ⊆ Hei =
UiU∗
i (l2) and define:
Hei = UiU∗
(cid:91)
i (l2) = span{δi, δi+k, δi+2k . . .}.
k(cid:91)
Bei = {δ1, δ2, δ3, . . .} = (δn)n∈N.
Bv :=
Be =
e∈s−1(v)
i=1
Note that π(Sei )(Bv) = π(Sei)((δn)n∈N) = Bei because π(Sei)(δ1) = δi, π(Sei)(δ2) = δi+k, π(Sei)(δ3) = δi+2k
and so on. So ϕ is permutative.
Example 2.5. Now we will show an example of a non-permutative representation. Let E be the graph
8
e
v
Define U : C2 → C2 by U (z, w) = (iz, iw). By the universal property of C∗(E) there exists a ∗-homomorphism
ϕ : C∗(E) → B(C2) such that ϕ(Se) = U and ϕ(Pv) = Id. Suppose ϕ permutative. Then there exist an
orthonormal total set Be ⊆ He = U U∗(C2) = C2 and as orthonormal total set Bv ⊆ Hv = C2 such that
Bv = Be and ϕ(Se)(Br(e)) = U (Br(e)) = Be. As Be is an orthonormal total set in C2 we can write
Be = {(z1, z2), (w1, w2)}. Thus
ϕ(Se)(Br(e)) = U (Br(e)) = {(iz1, iz2), (iw1, iw2)} (cid:54)= Be.
That's a contradiction. So ϕ isn't permutative.
Now we prove the main theorem of this section. This theorem is a generalization of Theorem 2.1 of [2].
Theorem 2.6. Let ϕ : C∗(E) → B(H) be a representation and suppose that ϕ is permutative. Then there
exists a representation π : C∗(E) → B(l2(Λ)), arising from a branching system, such that ϕ and π are
unitarily equivalent.
Proof. We use the same notations as in the beginning of this section. Recall that H =
Choose Bv ⊆ Hv a total orthonormal set. Moreover, defining B(cid:48) = (cid:83)
set in (cid:76)
Bv we obtain a total orthonormal
Hv. Also, choosing D a orthonormal total set in V we define B = B(cid:48) ∪ D and write B = {bλ}λ∈Λ,
which is a total orthonormal set in H. Let (Λ, η) be the measure space where η is the counting measure. For
each e ∈ E1, v ∈ E0 define the sets
v∈E0
v∈E0
v∈E0
Hv
(cid:32) (cid:76)
(cid:33)(cid:76) V .
Re = {λ ∈ Λ bλ ∈ Be}
and
Dv = {λ ∈ Λ bλ ∈ Bv}.
Now we define the desired branching system. If e (cid:54)= f , Be ∩ Bf = ∅ (because He ∩ Hf = {0H}) and then
Re ∩ Rf = ∅. The same argument shows that Dv ∩ Dw = ∅ (for v (cid:54)= w). For an edge e and λ ∈ Re, as
e ∈ s−1(s(e)) and ϕ is permutative then Be ⊆ Bs(e), and so bλ ∈ Bs(e). This means that λ ∈ Ds(e) and so
Re ⊆ Ds(e). Moreover, if v ∈ E0 is such that 0 < #s−1(v) < ∞ then Bv = (cid:83)
λ ∈ Dv ⇔ bλ ∈ Bv ⇔ bλ ∈ Be(cid:48) for some e(cid:48) ∈ s−1(v) ⇔ λ ∈ (cid:91)
Be and so
e∈s−1(v)
Re.
So Dv = (cid:83)
e∈s−1(v)
e∈s−1(v)
Re. For ech edge e we define the function fe : Dr(e) → Re by the following rule: given
λ0 ∈ Dr(e) we have bλ0 ∈ Br(e) and as ϕ is permutative ϕ(Se)(bλ0 ) ∈ Be. Therefore ϕ(Se)(bλ0) = bµ0 for
some µ0 ∈ Re. Define fe(λ0) = µ0.
the inverse of fe. It's clear that fe and f−1
measure and fe, f−1
As ϕ is permutative, fe is surjective and since ϕ(Se) : Hr(e) → He is injective so is fe; we choose f−1
as
are both measurable functions. Moreover, as η is the counting
. Then (Λ, η) is a branching system.
By Theorem 1.3 there exists a representation π : C∗(E) → B(L2(Λ, η)) = B(l2(Λ)) such that
are bijections we have Φfe = 1 = Φf
−1
e
e
e
e
π(Se)(φ) = χRe.Φ
.(φ ◦ f−1
e
) = χRe .(φ ◦ f−1
e
)
1
2
−1
f
e
and
π(Pv)(φ) = χDv .φ.
We define U : span{bλ}λ∈Λ → l2(Λ) given by
U (α1bλ1 + . . . + αnbλn ) = α1χ{λ1} + . . . + αnχ{λn}.
Its clear that U is linear and as {bλ}λ∈Λ is a orthonormal set in H and {χ{λ}}λ∈Λ is a orthonormal set in
αi2 = (cid:107)U (x)(cid:107)2. This shows that U is an isometric operator. Now we can extend the
l2(Λ) then (cid:107)x(cid:107)2 =
n(cid:80)
i=1
9
operator U to a operator (which we also call U ) from H = span({bλ}λ∈Λ) to l2(Λ). It is easy to see that U
is a unitary operator.
Claim 1. U∗π(Se)U = ϕ(Se) for each e ∈ E1.
For bλ0 ∈ B it holds that π(Se)U (bλ0 ) = π(Se)(χ{λ0}) = χRe (χ{λ0} ◦ f−1
1, if µ ∈ Re and µ = fe(λ0).
0, otherwise.
π(Se)U (bλ0)(µ) =
(cid:40)
), and so
e
If µ ∈ Λ is such that µ ∈ Re e µ = fe(λ0) then
Suppose bλ0 /∈ Br(e).
(µ) = λ0 ∈ Dr(e), that's a contradiction. Then π(Se)U (bλ0) = 0 and U∗π(Se)U (bλ0) = 0. Furthermore,
f−1
ϕ(Se)(bλ0) = 0 because if bλ0 /∈ Br(e), bλ0 ∈ Bv for some v (cid:54)= r(e) or bλ0 ∈ D. In the first case bλ0 ∈ Hv and
then bλ0 = ϕ(Pv)(bλ0). So,
In this situation λ0 /∈ Dr(e).
e
ϕ(Se)(bλ0 ) = ϕ(Se)ϕ(Pr(e))ϕ(Pv)(bλ0) = ϕ(Se)ϕ(Pr(e)Pv)(bλ0 ) = 0.
(cid:33)⊥
(cid:32) (cid:76)
Hv
v∈E0
In the second case bλ0 ∈ D ⊆
⊆ H⊥
r(e) = Ker ϕ(Pr(e)).
It follows that ϕ(Se)(bλ0 ) =
ϕ(Se)ϕ(Pr(e))(bλ0) = 0. In both situations it holds that
U∗π(Se)U (bλ0) = 0 = ϕ(Se)(bλ0).
Now let bλ0 ∈ Br(e).
fe(λ0) = µ0. Then,
In this case as ϕ is permutative then ϕ(Se)(bλ0) = bµ0 for some bµ0 ∈ Be, so
(cid:40)
(cid:40)
1, if µ ∈ Re and µ = fe(λ0).
0, otherwise.
=
1, if µ = µ0.
0, otherwise.
= χ{µ0}.
π(Se)U (bλ0 ) =
Therefore,
U∗π(Se)U (bλ0) = U∗(χ{µ0}) = bµ0 = ϕ(Se)(bλ0).
By linearity and continuity follows Claim 1.
Claim 2. U∗π(Pv)U = ϕ(Pv) for each v ∈ E0.
Let bλ0 ∈ B. Then π(Pv)U (bλ0) = χDv χ{λ}, and so
U∗π(Pv)U (bλ0) = U∗(χDv χ{λ0}) = U∗(χDv∩{λ0}) = χBv (U∗(χ{λ0})).
Moreover, ϕ(Pv)(bλ0) = χBv (bλ0 ) and then
U∗π(Pv)U (bλ0 ) = χBv (U∗(χ{λ0})) = χBv (bλ0) = ϕ(Pv)(bλ0).
By linearity and continuity it holds that for every h ∈ H, and so U∗π(Pv)U (h) = ϕ(Pv)(h), a so Claim 2 is
proved.
Finally, since π and ϕ are homomorphisms then for every x ∈ C∗(E) it holds that U∗π(x)U = ϕ(x). (cid:4)
For separable Hilbert spaces H, since each orthonormal total set of such spaces are countable, we get
from the previous theorem the following corollary.
Corollary 2.7 (2.1,[2]). Let ϕ : C∗(E) → B(H) be a representation, suppose that H is separable and ϕ is
permutative. Then exists a representation π : C∗(E) → B(l2(N)), induced by a branching system, such that
ϕ and π are unitarily equivalent.
3 Graphs whose all representations are permutative
In this section we prove that permutative representations are unitarily equivalent to representations
induced by branching systems. Now, our goal is to find a class of graphs such that every representation is
permutative. For this, we need some language that was developed in [2].
10
Definition 3.1 (3.1, 3.3: [2]). Let E be a graph.
1. For f ∈ E1 and v ∈ E0 we say that f and v are adjacent if r(f ) = v or s(f ) = v.
2. For u, v ∈ E0 such that u (cid:54)= v, we say that u and v are adjacent if exists a edge f ∈ E1 such that f to
u and f is adjacent to v.
3. For f, g ∈ E1 with f (cid:54)= g, we say that f, g are adjacent if there exists a vertex v ∈ E0 such that v is
adjacent to f and v is adjacent to g.
4. A path between u, v ∈ E0 is a pair (u0u1 . . . un, e1e2 . . . en); where ui ∈ E0 for each i = 0, . . . , n, ei ∈ E1
for each i = 1, . . . , n; and:
(a) u0 = u and un = v,
(b) ei (cid:54)= ej if i (cid:54)= j,
(c) for each i = 1, . . . , n it holds that r(ei) = ui−1 and s(ei) = ui or r(ei) = ui and s(ei) = ui−1.
5. A cycle is a path (u0u1 . . . un, e1e2 . . . en) such that u0 = un.
6. We say that E is P -simple if E has not loops and for every v, u ∈ r(E1)∪ s(E1) with v (cid:54)= u there exists
at most one path between v and u.
7. We say that a vertex v in E is a extreme vertex of E if s−1(v) ∪ r−1(v) = 1 and v is not a basis of a
loop. In this case, the only edge adjacent to v is called an extreme edge.
The reader can check that E is P -simple if, and only if, E has no cycles.
Let E be a graph. We define E1 as being the subgraph E1 = (E0 − X1, E1 − Y1, r1, s1) of E where X1 is
the set of extreme vertices of E, Y1 is the set of extreme edges of E and r1, s1 denote the restrictions of r
and s to E1 − Y1. The vertices in X1 are called the level 1 vertices of E and the edges in Y1 are called the
level 1 edges of E.
More generally, for each i ∈ N we define Ei as the subgraph Ei = (E0−X1∪. . .∪Xi, E1−Y1∪. . .∪Yi, ri, si)
of Ei−1 where Xi is the set of extreme vertices of Ei−1, Yi is the set of extreme edges of Ei−1 and ri, si denote
the restrictions of r and s to E1 − Y1 ∪ . . .∪ Yi. The vertices in Xi are called the level i vertices of E and the
edges in Yi are called the level i edges of E.
Note that the extreme vertices (edges) of Ei−1 are exactly the level i − 1 vertices (edges) of E.
Definition 3.2 (3.2: [2]). Let E = (E0, E1, r, s) be a graph and V ⊆ E0. We say that V is conneceted in E
if for all u, v ∈ V there exists a path (in E) between u e v. If V = E0 is connected in E we say that E is
connected.
Fix a graph E. Given u, v ∈ r(E1) ∪ s(E1) we say that u ∼ v if u = v or there exists a path between u
and v. Note that ∼ is a equivalence relation in Z := r(E1)∪ s(E1). Let ∆ be a set with exactly one member
of each equivalence class. Then we can write Z =
Zvi where Zvi denote the equivalence class of vi. As
consequence if R = E0 − Z then E0 =
.(cid:83)
.(cid:83) R. The elements of R are called isolated vertices. It's easy
.(cid:83)
vi∈∆
Zvi
vi∈∆
to see that
is a subgraph of E.
(s−1(Zv), Zv, ss−1(Zv), rs−1(Zv)) = (r−1(Zv), Zv, sr−1(Zv), rr−1(Zv))
The following proposition are very useful in the next results.
Proposition 3.3. Let E be a graph. Then
a) Suppose that Ei is defined for i ∈ N. If Z := r(E1) ∪ s(E1) is connected in E then Z − X1 ∪ . . . ∪ Xi is
connected in Ei.
b) [3.4: [2]] If v ∈ Xn for some n ∈ N then exists at most one vertex w in E such that w is adjacent to v
and the level of w is greater or equal to n.
11
c) [3.4: [2]]If Z =
m(cid:83)
i=1
Xi, for some m ∈ N and Z is connected then
I) Given v ∈ Xn with n < m exists one, and only one vertex w with level greater than n such that w
is adjacent to v.
II) The set Xm has exactly two vertices and exists exactly one edge adjacent to the two vertices.
m(cid:83)
.(cid:83){v} for each m ∈ N and Z is connected then:
d) [3.4: [2]] If Z = (
Xi)
i=1
I) If v ∈ Xn and n < m then v is adjacent to v or exists exactly one vertex w with level greater than
n such that w is adjacent to v.
II) For each v ∈ Xm exists exactly one edge adjacent to v and v.
Proof. We only will prove a). The proofs of the other items can be found in [2]. Suppose that i = 1. We will
prove that Z − X1 is connected in E1. Let u, v ∈ Z − X1. As Z is connected in E then there exists a path (in
E), (u0 . . . up, e1 . . . ep), between u and v. Notice that #s−1(ui)∪ r−1(ui) ≥ 2 for each i = 2, . . . , p− 1. Then
ui isn't a extreme vertex of E; therefore ui ∈ Z − X1 for each i = 1, . . . , p. Moreover, if ei /∈ E1 − Y1 then
ei ∈ Y1 and ei is a extreme edge of E. Thus s(ei) or r(ei) are extreme edges of E, that's a contradiction.
Therefore ei ∈ E1 − Y1 e (u0 . . . up, e1 . . . ep) is a path in E1 between u and v. Now the proof follows by
(cid:4)
inductive arguments.
For more details about adjacency, extreme vertices and connected graphs we recommend [2]. The next
m(cid:83)
m(cid:83)
Xi or Z = (
theorem is a generalization of [3.4 d), [2]].
Theorem 3.4. Let E be a P -simple graph. If Z := r(E1) ∪ s(E1) is connected and there exists n ∈ N such
that En exists and has finitely many vertices then Z =
Proof. Let n0 the smallest number such that En0 has finite vertices. Let m the biggest number such that Em
is defined. Suppose that Z := r(E1) ∪ s(E1) (cid:54)=
Xi. So v
is a vertex of Em and we can suppose that Em has N (N ∈ N) vertices.
We claim that v is a isolated vertex of the graph Em. Otherwise, there exists an edge e1 in the graph
Em such that e1 is adjacent to v. Suppose that r(e1) = v. Of course s(e1) (cid:54)= v because E is P -simple. Let
v0 = s(e1) and as Em is a subgraph, v0 is a vertex of Em.
As v is not a extreme vertex of Em then #s−1(v) ∪ r−1(v) ≥ 2. Let e2 be another edge adjacent to v in
Em. Without loss of generality, we assume s(e2) = v and r(e2) = v2. Note that v2 (cid:54)= v0, v2 (cid:54)= v. Proceeding
inductively we get a contradiction because Em has finite vertices.
.(cid:83){v} for some m ∈ N.
Xi ⊆ Z then there exists v ∈ Z − m(cid:83)
Xi. As
m(cid:83)
m(cid:83)
Xi)
i=1
i=1
i=1
i=1
i=1
e1
e2
e3
v0
Finally, suppose that Z − m(cid:83)
in the graph Em but this contradicts the fact that Em is connected. Therefore Z − m(cid:83)
v3...
v2
i=1
v
Xi has two or more elements. By the claim above the vertices are isolated
Xi has exactly one
i=1
element.
(cid:4)
Theorem 3.5. Let E = (E0, E1, r, s) a P -simple graph. Suppose that Z := r(E1) ∪ s(E1) is connected and
suppose that there exists n ∈ N such that the graph En exists and has finitely many vertices. If ϕ : C∗(E) →
B(H) is a representation then ∀v ∈ E0 and ∀e ∈ E1 then there exists total orthonormal sets Bv and Be from
Hv and He, respectively, such that:
1. If e ∈ s−1(v) then Be ⊆ Bv and if 0 < #s−1(v) < ∞ then Bv = (cid:83)
Be.
e∈s−1(v)
2. If e ∈ r−1(v) then ϕ(Se)(Bv) = Be.
12
m(cid:83)
i=1
m(cid:83)
i=1
m(cid:83)
i=1
Proof. The proof of this theorem consists in two inductive processes in the level of the vertices. By the
previous theorem, Z =
Xi or Z =
Xi ∪ {v}. Suppose that Z =
Xi.
.
.
1
1
2
1
.
2
2
2
2
1
1
(cid:83)
and s(e) /∈ X V F
Step 1. For each v ∈ X V F
we choose a total orthonormal set Bv ⊆ Hv. For each e ∈ r−1(v) we
define Be := π(Se)(Bv). For all the other vertices and edges we choose arbitrary orthonormal total subsets
Bv ⊆ Hv and Be ⊆ He. Then the conditions 1 and 2 are satisfied for all v ∈ X V F
and for all e ∈ E1 such
that r(e) ∈ X V F
If s−1(v) = ∅ we choose Bv := Bv, if 0 < #s−1(v) < ∞ we choose Bv :=
Be and if #s−1(v) = ∞ we define Bv as a total orthonormal set of Hv such that Be ⊆ Bv for each
e∈s−1(v)
e ∈ s−1(v). Thus, if e ∈ r−1(v) define Be := π(Se)(Bv). For the other vertices and edges define Bv := Bv e
Be := Be.
Step 2. Let v ∈ X V F
Claim 2.: If e ∈ r−1(v) and v ∈ X V F
Of course r(e) /∈ X V F
. If s(e) ∈ X V F
then r(e) /∈ X V F
then s(e) is a vertex of the graph E1 as well as r(e); so e is a edge
of this graph. As s(e) ∈ X V F
then there exists a vertex w with level greater than 2 and a edge f in the
graph E1 such that r(f ) = s(e). Note that e (cid:54)= f . Then s(e) is a adjacent to e and f in the graph E1. That
is a contradiction because s(e) is a extreme vertex of E1. Therefore s(e) /∈ X V F
get total orthonormal sets Bv (v ∈ E0) e Be (e ∈ E1) such that 1 and 2 are satisfied for all v ∈ X V F
2 are satisfied for all v ∈ X V F
The claim above ensures that the Step 2 doesn't modifies the previous choices. Thus, after the Step 2 we
1 ∪ X V F
.
We proceed inductively until the step m − 1. So, we get total orthonormal sets Bv e Be such that 1 and
Step m. Let v ∈ X V F
m . If s−1(v) = ∅ we define Bv := Bv, if 0 < #s−1(v) < ∞ we define Bv := (cid:83)
Claim m. If e ∈ r−1(v) and v ∈ X V F
Of course r(e) (cid:54)∈ X V F
Be
and if #s−1(v) = ∞ we choose Bv a orthonormal total set of Hv such that Be ⊆ Bv for each e ∈ s−1(v). If
e ∈ r−1(v) define Be := π(Se)(Bv). For the all the other vertices and edges we define Bv := Bv e Be := Be.
m−1. Suppose that s(e) ∈ X V F
for some i = 2, . . . , m. If 2 ≤ i ≤ m− 1 the
argument is the same as in Claim 2. If i = m then s(e) ∈ X V F
i
m as well as r(e), thus r(e), s(e) are distinct
vertices of Em−1 and e is a edge of this graph. As s(e) is a final vertex of Xm exists a edge f in this graph
such that r(f ) = s(e). Of course f (cid:54)= e. As Xm has exactly two elements (Proposition 3.3) then s(f ) = s(e)
or s(f ) = r(e). In both cases we get a cycle, which is impossible, since E is P -simple.
After this step we get total orthonormal sets Bv and Be such that 1 and 2 are satisfied for all v ∈
1 ∪ . . . ∪ X V F
X V F
m .
Step m + 1. Let v ∈ X V I
m . If 0 < #s−1(v) < ∞ we define
Bv := (cid:83)
m−1 and s(e) /∈ X V F
m then r(e) /∈ X V F
1 ∪ . . . ∪ X V F
m−1.
2 ∪ . . . ∪ X V F
m .
1 ∪ . . . ∪ X V F
1 ∪ . . .∪ X V F
e∈s−1(v)
∼
.
2
2
Be and if #s−1(v) = ∞
Bv for each e ∈ s−1(v). If e ∈ r−1(v) define
e∈s−1(v)
we choose
∼
Be = π(Se)(
∼
Bv a orthonormal total set of Hv such that Be ⊆ ∼
∼
∼
Bv). For all the other vertices and edges define
Bv := Bv e
∼
Be := Be.
m(cid:83)
Bv := (cid:83)
(cid:94)
With analogue arguments as used above we conclude that this construction don't change our previous
choices. Now, with inductive arguments we conclude the proof in the case Z =
m(cid:83)
.(cid:83) v. The steps 1 until m are the same. After this, we need a extra step as
Xi.
i=1
Now, suppose that Z =
Xi
follows:
i=1
(cid:94)
(cid:94)
we choose
Extra Step. We need deal with v in a extra step because v has not level. The argument is similar as in the
Be and if #s−1(v) = ∞
Bv := Bv; if 0 < #s−1(v) < ∞ define
e∈s−1(v)
Bv for every e ∈ s−1(v). For each e ∈ r−1(v) define
(cid:94)
Be := ϕ(Se)(
other steps. If s−1(v) = ∅ define
Bv such that Be ⊆ (cid:94)
(cid:94)
Bv) and for
the others vertices and edges define
(cid:94)
(cid:94)
Bv := Bv and
Be := Be.
Extra Claim. If e ∈ r−1(v) then r(e) /∈ X V F
1 ∪ . . . ∪ X V F
Of course r(e) /∈ X V F
m and if s(e) ∈ X V F
2 ∪ . . . ∪ X V F
(that's possible because v ∈ Ei for each i = 1, . . . , m − 1). If s(e) ∈ X V F
graph Em−1 such that r(f ) = s(e) and s(f ) = v, then (vs(e)s(f ), ef ) is a cycle. That's a contradiction.
m−1 we repeat the previous arguments
m then there exists an edge f in the
1 ∪ . . . ∪ X V F
m and s(e) /∈ X V F
2 ∪ . . . ∪ X V F
m .
13
The next steps are the same as in the case Z =
m(cid:83)
Xi. So we conclude the case Z =
m(cid:83)
.(cid:83){v} and the
Xi
(cid:4)
proof of the theorem.
Corollary 3.6. Let E = (E0, E1, r, s) be P -simple graph. Suppose that Z := r(E1) ∪ s(E1) is connected
and exists n ∈ N such that the graph En exits and has finitely many vertices. Then every representation
ϕ : C∗(E) → B(H) is unitarily equivalent to a representation arising from a E-branching system.
Proof. It follows from Theorems 3.5 e 2.6.
(cid:4)
i=1
i=1
Example 3.7. Due to the previous corollary every representation of the two graphs below is unitary equivalent
to a representation arising from a branching system
v1
e2
e3
... eN ...
v2
v3
vN
...
Given a graph E = (E0, E1, r, s) we know that E0 =
(cid:32) .(cid:83)
Zvi
vi∈∆
v1
e−1
v2
e0
v3
v4
e1e2
v5
(cid:33) .(cid:83) R. Let us consider the connected
subgraphs Evi = (Zvi, s−1(Zvi), rs−1(Zvi ), ss−1(Zvi )).
Corollary 3.8. Let E = (E0, E1, r, s) be a P -simple graph. Suppose that for each i ∈ I there exists ni ∈ N
has finitely many vertices. Then, every representation ϕ : C∗(E) → B(H) of C∗(E)
such that the graph E vi
is unitarily equivalent to a representation arising from a E-branching system.
ni
Proof. It follows by applying the previous result to each subgraph Evi.
(cid:4)
Example 3.9. By the previous corollary, each representation of the graph below is unitary equivalent to a
representation induced by a branching system.
v1
e2
e3
e1
v2
v3
v4
e4
v5
e6
e5
v9
v8
v7
v6
The converse of the previous corollaries are note true. There are graphs such that every representation of
their C*-algebras are unitarily equivalent to representations induced by a branching system but the graphs
do not satisfy the hypothesis of the previous corolaries. We will show two examples.
Example 3.10 (2.2, [2]). In [2], has been shown that every representation of the graph
e−1
...
v−1
e0
v0
e1
v1
e2
...
is permutative. Then, by Theorem 2.6 every representation of this graph is unitarily equivalent to a repre-
sentation arising from a branching system. However, notice that there are no extreme edges and extreme
vertices, so that no En does exist.
Example 3.11. Let E be the following graph
e1
v1
e8
v2
e2
v3
e3
v4
v8
e7
v6
e6
v7
14
e4
v5
e5
and let ϕ : C∗(E) → B(H) an arbitrary representation. First choose arbitrary orthonormal total sets for the
vertices v such that #r−1(v) ≥ 2, that is, the vertices v1, v3 and v5. For this vertices choose Bv1 ⊆ Hv1 ,
Bv3 ⊆ Hv3, Bv5 ⊆ Hv5. Now define Bei := ϕ(Sei)(Br(ei)) for each i = 1, 2, 3, 4, 5, 8. After this, choose
Bv2 := Be1 ∪ Be2, Bv4 := Be3 ∪ Be4 , Bv8 := Be8 and Bv6 := Be5; finally define Be7 := π(Se7)(Bv8) and
Be6 := π(Se6)(Bv6). Finally define Bv7 = Be7 ∪ Be6. With this choices is clear that ϕ is permutative. Of
course this graphs doesn't satisfies the hypothesis of corollaries because E isn't P -simple.
Remark 3.12. The previous example may be generalized in a natural way, with analogous arguments, for
every graph E which is a cycle with the property that {v ∈ E0 #r−1(v) = 2} (cid:54)= ∅ for some vertex v. The
hypothesis {v ∈ E0 #r−1(v) = 2} (cid:54)= ∅ is important. For example, in Example 2.5 the graph is a cycle,
with {v ∈ E0 #r−1(v) = 2} = ∅ for each vertex, and in this example there is shown a non permutative
representation.
References
[1] Neal Fowler, Marcelo Laca, and Iain Raeburn. The C*-algebras of infinite graphs. Proceedings of the
American Mathematical Society, 128(8):2319 -- 2327, 2000.
[2] Daniel Gon¸calves and Danilo Royer. Unitary equivalence of representations of graph algebras and branch-
ing systems. Functional Analysis and Its Applications, 45(2):117 -- 127, 2011.
[3] Daniel Gon¸calves and Danilo Royer. Graph C*-algebras, branching systems and the Perron -- Frobenius
operator. Journal of Mathematical Analysis and Applications, 391(2):457 -- 465, 2012.
[4] Alex Kumjian, David Pask, and Iain Raeburn. Cuntz -- krieger algebras of directed graphs. pacific journal
of mathematics, 184(1):161 -- 174, 1998.
[5] Iain Raeburn. Graph algebras. Number 103. American Mathematical Soc., 2005.
[6] Walter Rudin. Real and complex analysis. Tata McGraw-hill education, 1987.
Ben-Hur Eidt, Departamento de Matem´atica, Universidade Federal de Santa Catarina, Florian´opolis,
88040-900, Brasil
Email: [email protected]
Danilo Royer, Departamento de Matem´atica, Universidade Federal de Santa Catarina, Florian´opolis,
88040-900, Brasil
Email: [email protected]
15
|
1804.06591 | 1 | 1804 | 2018-04-18T07:48:17 | Efficient Presentations of Relative Cuntz-Krieger Algebras | [
"math.OA"
] | In this article, we present a new method to study relative Cuntz-Krieger algebras for higher-rank graphs. We only work with edges rather than paths of arbitrary degrees. We then use this method to simplify the existing results about relative Cuntz-Krieger algebras. We also give applications to study ideals and quotients of Toeplitz algebras. | math.OA | math |
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER
ALGEBRAS
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
Abstract. In this article, we present a new method to study relative Cuntz-Krieger
algebras for higher-rank graphs. We only work with edges rather than paths of arbitrary
degrees. We then use this method to simplify the existing results about relative Cuntz-
Krieger algebras. We also give applications to study ideals and quotients of Toeplitz
algebras.
1. Introduction
For a directed graph E, Fowler and Raeburn introduced the Toeplitz algebra T C ∗(E)
[1]. The usual graph algebra C ∗(E) (or Cuntz-Krieger algebra) is the quotient of T C ∗(E)
in which the Cuntz-Krieger relation
pv = Xr(e)=v
ses∗
e
is imposed at every regular vertex v; that is, at every vertex that receives only finitely
many edges. Muhly and Tomforde described the quotients of C ∗(E) as relative graph
algebras: for a set V of regular vertices, the relative graph algebra C ∗(E; V ) is the quotient
of T C ∗(E) in which the Cuntz-Krieger relation is imposed at every v ∈ V [5]. These
relations are independent of each other: if v /∈ V , then pv 6=Pe∈vE 1 ses∗
The higher-rank graphs or k-graphs of Kumjian and Pask [4] are higher-dimensional
analogues of directed graphs, and they also have both a Cuntz-Krieger algebra [4, 8, 9] and
a Toeplitz algebra [7]. Here we consider the class of finitely aligned k-graphs discovered
in [7] and studied in [9]. For such a k-graph Λ, the Cuntz-Krieger algebra C ∗(Λ) is the
quotient of the Toeplitz algebra T C ∗(Λ) in which the Cuntz-Krieger relation
e in C ∗(E; V ).
(1.1)
(tv − tλt∗
λ) = 0
Yλ∈E
is imposed at every finite exhaustive set E of every vΛ := r−1(v). However, the relations
(1.1) are not independent of each other: imposing the relations for some exhaustive sets
automatically imposes others. For a collection E of finite exhaustive sets of Λ, Sims
introduced the relative Cuntz-Krieger algebras C ∗(Λ; E) to be the quotient of T C ∗(Λ) in
which the relation (1.1) is imposed for every E ∈ E. He also identified satiated collections
E of finite exhaustive sets which describe the possible quotients C ∗(Λ; E) [10].
Sims's satiated sets are huge, and his exhaustive sets can include paths of arbitrary
degrees. However, while it has been standard since the beginning of the subject [4] to
2010 Mathematics Subject Classification. 46L05.
Key words and phrases. higher-rank graph, relative graph algebra, graph C ∗-algebra.
This research was supported by Marsden grant 15-UOO-071 from the Royal Society of New Zealand.
Thank you to Iain Raeburn for sharing his insights.
1
2
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
work with Cuntz-Krieger relations of all degrees, we know from [9, Appendix C] that it is
possible to work only with sets of edges, as one does for directed graphs (the 1-graphs),
and that it is then easier to see what is happening. Here we describe and study a family
of efficient collections in which the exhaustive sets contain only a minimal number of
edges. This simplifies the description of the relations being imposed when passing from
the Toeplitz algebra to a relative Cuntz-Krieger algebra.
This paper is organised as follows. In Section 2, we give the definition of higher-rank
graphs and establish our notation.
In Section 3, we introduce efficient sets and give
examples. For a directed graph E, every set of regular vertices of E can be viewed as
an efficient set (Example 3.5). Hence relative graph algebras for directed graphs [5] were
defined using efficient sets rather than Sims's satiated sets.
In Section 4, we introduce E-boundary paths (Definition 4.1). We use these paths to
establish properties of the universal relative Cuntz-Krieger algebras (Proposition 4.5). In
Section 5, we discuss the satiated sets of Sims [10] and show that efficient sets are in
bijective correspondence with satiated sets (Theorem 5.3).
Finally, we discuss applications in Section 6. The first application is a new version of the
gauge-invariant uniqueness theorem for relative Cuntz-Krieger algebras of [10] (Theorem
6.1). The second application is to simplify a complete listing of the gauge-invariant
ideals in a relative Cuntz-Krieger algebra of [12] (Theorem 6.4). The authors of [10]
and [12] use satiated sets to formulate both results. By translating these into efficient
sets, we provide alternative versions, which are considerably more checkable. In the last
application, we investigate the relationship among k-graphs Toeplitz algebras and their
ideals and quotients (Proposition 6.10).
2. Preliminaries
Let k be a positive integer. We regard Nk as an additive semigroup with identity 0. We
write n ∈ Nk as (n1, . . . , nk) and define n :=P1≤i≤k ni. We denote the usual basis of Nk
by {ei}. For m, n ∈ Nk, we write m ≤ n to denote mi ≤ ni for 1 ≤ i ≤ k. We also write
m ∨ n for their coordinate-wise maximum and m ∧ n for their coordinate-wise minimum.
A higher-rank graph or k-graph is a countable category Λ endowed with a functor d :
Λ → Nk satisfying the factorisation property: for λ ∈ Λ and m, n ∈ Nk with d(λ) = m+n,
there are unique elements µ, ν ∈ Λ such that λ = µυ, d(µ) = m and d(ν) = n. We then
write λ(0, m) for µ and λ(m, m + n) for ν.
For n ∈ Nk, we define
Λn := {λ ∈ Λ : d(λ) = n}
and call the elements λ of Λn paths of degree n. For 1 ≤ i ≤ k, a path e ∈ Λei is an edge,
and we write
for the set of all edges. We regard elements of Λ0 as vertices. For v ∈ Λ0, λ ∈ Λ and
E ⊆ Λ, we define
Λ1 := [1≤i≤k
Λei
vE := {µ ∈ E : r(µ) = v},
λE := {λµ ∈ Λ : µ ∈ E, r(µ) = s(λ)},
Eλ := {µλ ∈ Λ : µ ∈ E, s(µ) = r(λ)}.
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
3
A k-graph Λ is row-finite if for v ∈ Λ0 and 1 ≤ i ≤ k, the set vΛei is finite. A vertex
v ∈ Λ0 is a source if there exists m ∈ Nk such that vΛm = ∅.
To visualise k-graphs, we use coloured graphs of [2]. For a k-graph Λ, we choose k-
different colours c1, . . . , ck and associate each edge e ∈ Λei to an edge of colour ci. We
call this coloured graph the skeleton of Λ.
Convention. We draw
m
•v
•w
in the skeleton of a 2-graph to denote that there are m (1, 0)-edges from w to v and 1
(0, 1)-edge from v to w.
For λ, µ ∈ Λ, we define
MCE(λ, µ) :={τ ∈ Λd(λ)∨d(µ) : τ (0, d(λ)) = λ and τ (0, d(µ)) = µ},
Λmin(λ, µ) := {(λ′, µ′) ∈ Λ × Λ : λλ′ = µµ′ ∈ MCE(λ, µ)}.
We say that Λ is finitely aligned if Λmin(λ, µ) is finite (possibly empty) for all λ, µ ∈ Λ.
For v ∈ Λ0, E ⊆ vΛ is exhaustive if for λ ∈ vΛ, there exists µ ∈ E with Λmin(λ, µ) 6= ∅.
We write FE(Λ) to denote the collection of finite exhaustive sets E ⊆ vΛ\{v}, and
FE(Λ1) := FE(Λ) ∩ Λ1.
For E ∈ FE(Λ), we write r(E) for the vertex v ∈ Λ0 such that E ⊆ vΛ.
A Toeplitz-Cuntz-Krieger Λ-family is a collection {Tλ : λ ∈ Λ} of partial isometries in
a C ∗-algebra B satisfying:
(TCK1) {Tv : v ∈ Λ0} is a collection of mutually orthogonal projections;
(TCK2) TλTµ = Tλµ whenever s(λ) = r(µ); and
(TCK3) T ∗
µ′ for all λ, µ ∈ Λ.
λ Tµ =P(λ′,µ′)∈Λmin(λ,µ) Tλ′T ∗
Λ-family {Tλ : λ ∈ Λ} which satisfies the Cuntz-Krieger relations:
For E ⊆ FE(Λ), a relative Cuntz-Krieger (Λ; E)-family is a Toeplitz-Cuntz-Krieger
(CK) Qλ∈E(Tr(E) − TλT ∗
λ ) = 0 for all E ∈ E.
In [7, Section 4], Raeburn and Sims proved that there is a C ∗-algebra T C ∗(Λ) generated
by a universal Toeplitz-Cuntz-Krieger Λ-family {tλ : λ ∈ Λ}. For E ⊆ FE(Λ), the quotient
C ∗(Λ; E) of T C ∗(Λ) by the ideal generated by
nYλ∈E
(tr(E) − tλt∗
λ) : E ∈ Eo
is generated by a universal relative Cuntz-Krieger (Λ; E)-family {sE
λ : λ ∈ Λ}. For a
relative Cuntz-Krieger (Λ; E)-family {Sλ : λ ∈ Λ} in C ∗-algebra B, we write πS for the
homomorphism of C ∗(Λ; E) into B such that πS(sE
λ) = Sλ for λ ∈ Λ.
3. Efficient sets
In this section, we introduce efficient sets and study their properties.
Imposing the
Cuntz-Krieger relations on a higher-rank graph has domino effects, which are described
in Proposition 3.1 and Proposition 3.3, and these effects motivate Definition 3.4.
4
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
Proposition 3.1. Let v ∈ Λ0, E ∈ v FE(Λ1) and f ∈ r(E)Λ1\E. Suppose that {Tλ : λ ∈
Λ} is a Toeplitz-Cuntz-Krieger Λ-family such thatQe∈E(Tv − TeT ∗
Then ExtΛ(f ; E) ∈ s(f ) FE(Λ1) andQg∈ExtΛ(f ;E)(Ts(f ) − TgT ∗
ExtΛ(f ; E) := {g ∈ s(f )Λ : f g ∈ MCE(f, e) for some e ∈ E}.
g ) = 0.
Proof. Since E ∈ FE(Λ), [9, Lemma C.5] gives ExtΛ(f ; E) ∈ s(f ) FE(Λ). We claim that
ExtΛ(f ; E) ⊆ s(f )Λ1. To prove the claim, let g ∈ ExtΛ(f ; E). Then there exists an
edge e ∈ E such that f g ∈ MCE(f, e). First we show that d(e) 6= d(f ). Suppose for a
contradiction that d(f ) = d(e). Since f g ∈ MCE(f, e),
e ) = 0. Define
d(f ) + d(g) = d(f g) = d(f ) ∨ d(e) = d(f ).
Hence d(g) = 0 and g = s(f ). Because f g ∈ MCE(f, e), f = e ∈ E, which contradicts
f /∈ E. Thus d(f ) 6= d(e). Now since d(f ) = 1 = d(e), d(f ) ∨ d(e) = d(f ) + d(e). So
d(f ) + d(g) = d(f g) = d(f ) ∨ d(e) = d(f ) + d(e)
and hence d(g) = d(e) and d(g) = 1 proving the claim. Thus ExtΛ(f ; E) ∈ s(f ) FE(Λ1).
To showQg∈ExtΛ(f ;E)(Ts(f ) − TgT ∗
g ) = 0, we prove
(T ∗
Yg∈ExtΛ(f ;E)
(Ts(f ) − TgT ∗
g ) =Ye∈E
e )Tf ) =Ye∈E(cid:16)Ts(f ) −(cid:16)
e )Tf ) =Ye∈E(cid:16)Ts(f ) −(cid:16) X(g,e′)∈Λmin(f,e)
First by (TCK3), we have
(T ∗
f (Tv − TeT ∗
Ye∈E
(T ∗
f (Tv − TeT ∗
Ye∈E
=Ye∈E Y(g,e′)∈Λmin(f,e)
f(cid:16)Ye∈E
e )Tf ) = T ∗
Ye∈E
f (Tv − TeT ∗
e )Tf ) = 0.
TgT ∗
Tg(T ∗
X(g,e′),(g′′,e′′)∈Λmin(f,e)
g(cid:17)(cid:17)
g ) = Yg∈ExtΛ(f ;E)
e )(cid:17)(Tf T ∗
e′Te′′)T ∗
g′′(cid:17)(cid:17).
For e′ 6= e′′, d(e′) = d(e′′) and by (TCK3), T ∗
e′Te′′ = 0. If e′ = e′′, then T ∗
e′Te′ = Ts(e′). So
(Ts(f ) − TgT ∗
(Ts(f ) − TgT ∗
g ).
On the other hand, since {TλT ∗
λ : λ ∈ Λ} is a commuting family [9, Lemma 2.7(i)],
(T ∗
f (Tv − TeT ∗
(Tv − TeT ∗
f )E−1Tf = 0.
(cid:3)
Corollary 3.2. Let v ∈ Λ0 and E ∈ v FE(Λ1). For λ ∈ r(E)Λ\ EΛ, the set
ExtΛ(λ; E) := {g ∈ s(λ)Λ : λg ∈ MCE(λ, e) for some e ∈ E}
belongs to s(λ) FE(Λ1).
Proof. For w ∈ Λ0, F ⊆ wΛ, λ1 ∈ wΛ and λ2 ∈ s(λ1)Λ, Lemma 4.9 of [10] tells that
Hence by induction on λ, ExtΛ(λ; E) ∈ FE(Λ1) follows from Proposition 3.1.
(cid:3)
ExtΛ(λ1λ2; F ) = ExtΛ(λ2; ExtΛ(λ1; F )).
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
5
Proposition 3.3. Let v ∈ Λ0 and E ∈ v FE(Λ1). Suppose that {Tλ : λ ∈ Λ} is a Toeplitz-
g ) = 0. Also suppose that e ∈ E and
Cuntz-Krieger Λ-family such that Qg∈E(Tv − TgT ∗
F ∈ s(e) FE(Λ1) satisfiesQf ∈F (Ts(e) − Tf T ∗
belongs to v FE(Λ1) andQg∈EF
(Tv − TgT ∗
g ) = 0.
f ) = 0. Then
EF := (E\{e}) ∪ {(ef )(0, d(f )) : f ∈ F }
Proof. Since E and F are nonempty and finite, so is EF . To show that EF is exhaustive,
take λ ∈ vΛ. We give a separate argument for Λmin(λ, e) = ∅ and Λmin(λ, e) 6= ∅. First
suppose Λmin(λ, e) = ∅. Because E ∈ v FE(Λ1) and Λmin(λ, e) = ∅, there exists g ∈ E\{e}
with Λmin(λ, g) 6= ∅. By definition of EF , g ∈ EF as needed.
Next suppose Λmin(λ, e) 6= ∅. Take (λ′, e′) ∈ Λmin(λ, e). So λλ′ = ee′. Since F ∈
r(e′) FE(Λ1), there exists f ∈ F with Λmin(e′, f ) 6= ∅. Take (e′′, f ′′) ∈ Λmin(e′, f ). Then
e′e′′ = f f ′′, λλ′e′′ = ee′e′′ = ef f ′′, and Λmin(λ, ef ) 6= ∅. Let g := (ef )(0, d(f )). Then
g ∈ EF and Λmin(λ, g) 6= ∅. So EF is exhaustive and EF ∈ FE(Λ1).
f ) = 0, we have Tv − TeT ∗
ef ) by [9, Lemma
SinceQf ∈F (Ts(e) − Tf T ∗
C.7]. Hence
Yg∈E\{e}
g )Yf ∈F
(Tv − TgT ∗
(Tv − Tef T ∗
e =Qf ∈F (Tv − Tef T ∗
ef ) =Yg∈E
(Tv − TgT ∗
g ) = 0.
On the other hand, for f ∈ F , Tef T ∗
ef = (Tef T ∗
ef )(T(ef )(0,d(f ))T ∗
(ef )(0,d(f ))). So by (TCK3),
Yg∈EF
(Tv − TgT ∗
g ) = Yg∈E\{e}
g )Yf ∈F
(Tv − TgT ∗
((Tv − Tef T ∗
ef )(Tv − T(ef )(0,d(f ))T ∗
(ef )(0,d(f )))) = 0.
(cid:3)
Definition 3.4. We call a subset E of FE(Λ1) efficient if the following three conditions
are satisfied:
(E1) if E, F ∈ E and E ⊆ F , then E = F ;
(E2) if E ∈ E and f ∈ r(E)Λ1\E, then there exists F ∈ E such that F ⊆ ExtΛ(f ; E);
and
(E3) if E ∈ E, e ∈ E, and F ∈ s(e)E, then there exists G ∈ E with
G ⊆ (E\{e}) ∪ {(ef )(0, d(f )) : f ∈ F }.
Example 3.5. Suppose that Λ is a 1-graph. Let V ⊆ Λ0 be a nonempty subset of regular
vertices; that is, 0 < vΛ1 < ∞ for v ∈ V . Then the set E := {vΛ1 : v ∈ V } is
efficient: Properties (E1) and (E2) are trivially true. To show (E3), let E ∈ E, e ∈ E and
F ∈ s(e)E. (Since s(e) is a regular vertex, s(e)E is nonempty.) Then
(E\{e}) ∪ {(ef )(0, d(f )) : f ∈ F } = (E\{e}) ∪ {e} = E.
Choose G := E and (E3) follows. So one can translate the results about subsets of regular
vertices of [5] into results about efficient sets.
6
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
Example 3.6. Suppose that Λ is a 2-graph with the following skeleton:
n
•v
m
We show {vΛe1} is efficient. Condition (E1) is trivial. Notice that for any g ∈ vΛe2, we
have ExtΛ(g; vΛe1) = vΛe1 so (E2) holds. For (E3), let e ∈ vΛe1. Then
(vΛe1\{e}) ∪ {(ef )(0, d(f )) : f ∈ vΛe1} = (vΛe1\{e}) ∪ {e} = vΛe1
since all edges of vΛe1 have the same degree. A similar argument shows {vΛe2} and
{vΛe1 ∪ vΛe2} are also efficient.
Remark 3.7. For a row-finite k-graph Λ with no sources and a nonempty subset K of
{1, . . . , k},
EK :=n[i∈K
vΛei : v ∈ Λ0o and Ei := {vΛei : v ∈ Λ0} for 1 ≤ i ≤ k
are all efficient. So, for example, the sets {vΛe1 : v ∈ Λ0}, {vΛe2 : v ∈ Λ0} and {vΛe1 ∪
vΛe2 : v ∈ Λ0} are always efficient. However the set {vΛe1, vΛe2 : v ∈ Λ0} might not be.
Consider the 2-graph Λ with skeleton
...
e
. . .
•
•
•
g
•
v
•
...
. . .
•
•
Then
(vΛe1\{e}) ∪ {(ef )(0, d(f )) : f ∈ xΛe2} = (vΛe1\{e}) ∪ {g}
contains neither vΛe1 nor vΛe2, so (E3) fails.
Now we study properties of efficient sets.
Definition 3.8. Let E ⊆ FE(Λ1). Then
min(E) := {E ∈ E : F ⊆ E and F ∈ E imply F = E},
and the edge satiation of E is
Remark 3.9. Using the edge satiation, we provide an alternate characterisation of efficient.
A subset E of FE(Λ1) is efficient if it satisfies (E1) and
bE := {E ∈ FE(Λ1) : there exists F ∈ E with F ⊆ E}.
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
7
(E2′) if E ∈ E and f ∈ r(E)Λ1\E, then ExtΛ(f ; E) ∈ bE;
(E3′) if E ∈ E, e ∈ E, and F ∈ s(e)E, then (E\{e}) ∪ {(ef )(0, d(f )) : f ∈ F } ∈ bE.
Lemma 3.10. Suppose that E ⊆ FE(Λ1) is efficient. Then E = min(bE).
Proof. To show E ⊆ min(bE), take E ∈ E. Then E ∈ bE. To show E ∈ min(bE), take F ∈ bE
such that F ⊆ E. Since F ∈ bE, there exists F ′ ∈ E such that F ′ ⊆ F . So F ′ ⊆ F ⊆ E
and by (E1), F ′ = E and E = F . Therefore E ∈ min(bE) and E ⊆ min(bE).
To show min(bE) ⊆ E, take E ∈ min(bE). Then E ∈ bE and there exists F ∈ E with
F ⊆ E. So F ∈ bE. Since F ⊆ E and E ∈ min(bE), F = E and E ∈ E. So min(bE) ⊆ E. (cid:3)
The next Proposition shows that the edge satiation gives the same relative Cuntz-
Krieger algebra.
Proposition 3.11. Let Λ be a finitely aligned k-graph. Suppose that E ⊆ FE(Λ1). Then
(Λ; E)-family. Now suppose that {Sλ : λ ∈ Λ} is a relative Cuntz-Krieger (Λ; E)-family.
C ∗(Λ; E) = C ∗(Λ;bE).
Proof. Since E ⊆ bE, a relative Cuntz-Krieger (Λ;bE)-family is a relative Cuntz-Krieger
For E ∈ bE, there exists F ∈ E with F ⊆ E, soQe∈F (Sr(E) − SeS∗
e ) Ye∈ E\F
Thus {Sλ : λ ∈ Λ} is also a relative Cuntz-Krieger (Λ;bE)-family. The universal property
of C ∗(Λ; E) and C ∗(Λ;bE) implies the two algebras coincide.
e ) =Ye∈F
4. E -boundary paths
(Sr(E) − SeS∗
e ) = 0.
e ) = 0 and
(Sr(E) − SeS∗
(Sr(E) − SeS∗
Ye∈E
(cid:3)
In this section, we discuss E-boundary paths and investigate properties of relative
Cuntz-Krieger algebras (Proposition 4.5). For k ∈ N and m ∈ (N∪{∞})k, Ωk,m is
the k-graph which has vertices {n ∈ Nk : n ≤ m}, morphisms {(n1, n2) : n1, n2 ∈
Nk, n1 ≤ n2 ≤ m}, degree map d((n1, n2)) = n2 − n1 and range and source maps
r((n1, n2)) = n1, s((n1, n2)) = n2 (see [8, Section 2]).
Definition 4.1. Suppose that Λ is a finitely aligned k-graph and that E ⊆ FE(Λ1) is
efficient. A path x : Ωk,m → Λ is an E-boundary path of Λ if for n ∈ Nk such that n ≤ m,
and E ∈ x(n)E, there exists e ∈ E such that x(n, n + d(e)) = e. We denote the collection
of all E-boundary paths of Λ by ∂(Λ; E). We write d(x) for m and r(x) for x(0).
The next two lemmas use similar arguments to Lemma 4.4 and Lemma 4.7 of [10] (so
we omit the proofs).
Lemma 4.2. Let Λ be a finitely aligned k-graph and E ⊆ FE(Λ1) be efficient. Suppose
that x ∈ ∂(Λ; E).
(a) If n ∈ Nk with n ≤ d(x), then x(n, d(x)) ∈ ∂(Λ; E).
(b) If λ ∈ Λr(x), then λx ∈ ∂(Λ; E).
Lemma 4.3. Let Λ be a finitely aligned k-graph, E be efficient and v ∈ Λ0. Then
v∂(Λ; E) 6= ∅ and for E ∈ v FE(Λ1)\bE, v∂(Λ; E)\ E∂(Λ; E) 6= ∅.
8
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
Now we give a concrete example of a relative Cuntz-Krieger (Λ; E)-family. We use
this family to prove Proposition 4.5, which establishes properties of the universal relative
Cuntz-Krieger (Λ; E)-family.
Example 4.4. For a finitely aligned k-graph Λ and an efficient set E ⊆ FE(Λ1), we define
a partial isometries {SE
λ : λ ∈ Λ} ⊆ B(l2(∂(Λ; E))) by
λ (ex) :=(eλx
0
SE
if s(λ) = r(x),
otherwise.
Then an argument similar to the proof of [10, Lemma 4.6] shows {SE
λ : λ ∈ Λ} is a relative
Cuntz-Krieger (Λ; E)-family. We call this family the E-boundary path representation of
C ∗(Λ; E).
Proposition 4.5. Let Λ be a finitely aligned k-graph and E ⊆ FE(Λ1) be efficient. Suppose
that {sE
λ : λ ∈ Λ} is the universal relative Cuntz-Krieger (Λ; E)-family.
(1) For v ∈ Λ0, sE
v 6= 0.
e sE∗
v − sE
Proof. We use {SE
take v ∈ Λ0 and x ∈ v∂(Λ; E). Then SE
universal property of C ∗(Λ; E) shows that sE
λ : λ ∈ Λ} the E-boundary path representation of C ∗(Λ; E). For (1)
v is nonzero and the
(2) For v ∈ Λ0 and E ⊆ vΛ1, E ∈ bE if and only ifQe∈E(sE
For (2), take v ∈ Λ0 and E ⊆ vΛ1. Suppose E /∈ bE. We give separate arguments for
E /∈ FE(Λ1) and E ∈ FE(Λ1)\bE. First suppose E /∈ FE(Λ1). Then there exists g ∈ vΛ1
with ExtΛ(g; E) = ∅. For e ∈ E, we have Λmin(g, e) = ∅ and sE∗
g = sE
hand, by part (1), sE∗
Thus
g sE
g is nonzero, sE
e = 0. On the other
6= 0.
v (ex) = ex 6= 0. Therefore SE
g 6= 0. Since sE
s(g) 6= 0 and sE
v is nonzero.
e ) = 0.
g = sE
g sE∗
g
g sE∗
g sE
g sE
g Ye∈E
(SE
sE
g sE∗
(sE
v − sE
e sE∗
e ) = sE
g sE∗
g
6= 0
e sE∗
e SE∗
v − sE
v − SE
e ∈ E, (SE
e )(ex) = SE
e ) is nonzero.
and henceQe∈E(sE
Next suppose E ∈ FE(Λ1)\bE. By Lemma 4.3, there exists x ∈ v∂(Λ; E)\ E∂(Λ; E). For
SoQe∈E(SE
For the reverse implication suppose E ∈ bE. Then there exists F ∈ E with F ⊆ E. So
Qf ∈F (sE
e ) 6= 0 and thenQe∈E(sE
f ) = 0 and since F ⊆ E,Qe∈E(sE
v (ex) = ex and
e SE∗
e )(ex) = ex 6= 0.
5. Relationship between efficient and satiated sets
Ye∈E
e ) 6= 0.
e ) = 0.
v − SE
v − SE
v − sE
v − sE
v − sE
e SE∗
f sE∗
e sE∗
e sE∗
(cid:3)
Now we discuss satiated set and show that for a finitely aligned k-graph, there exists a
bijection between its efficient sets and its satiated sets (Theorem 5.3). As in [10, Definition
4.1], a subset F ⊆ FE(Λ) is satiated if it satisfies
(S1) if E ∈ F and F ∈ FE(Λ) with E ⊆ F , then F ∈ F ;
(S2) if E ∈ F and λ ∈ r(E)Λ\EΛ, then ExtΛ(λ; E) ∈ F ;
(S3) if E ∈ F and 0 < nλ ≤ d(λ) for λ ∈ E, then {λ(0, nλ) : λ ∈ E} ∈ F ;
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
9
(S4) if E ∈ F , E′ ⊆ E and for each λ ∈ E′, E′
λ ∈ s(λ)F , then
(cid:16)(E\E′) ∪(cid:16) [λ∈E ′
λE′
λ(cid:17)(cid:17) ∈ F .
For F ⊆ FE(Λ), Sims writes F for the smallest satiated subset of FE(Λ) which contains
F , and call it the satiation of F [10, Section 5]. He also shows how to construct the
satiation of F .
Remark 5.1. Suppose F ⊆ FE(Λ). Sims defines maps Σ1 to Σ4 [10, Definition 5.2] and
shows that the iterated application of these maps produces F [10, Proposition 5.5]. The
set
Σ1(F ) := {F ∈ FE(Λ) : there exists G ∈ F with G ⊆ F }
is contained in F . So we get the following result: Suppose that E ⊆ FE(Λ1). Then the
set bE is contained in Σ1(E) and bE ⊆ E. So E ⊆ bE ⊆ E and E ⊆ bE ⊆ E = E giving
bE = E.
3.11, C ∗(Λ; E), C ∗(Λ;bE) and C ∗(Λ; E) all coincide.
Now we state the main result of this section.
Remark 5.2. Corollary 5.6 of [10] shows that C ∗(Λ; E) = C ∗(Λ; E). So by Proposition
Theorem 5.3. For a finitely aligned k-graph Λ, the map E 7→ E is a bijection between
efficient sets of Λ and satiated sets of Λ, with inverse given by F 7→ min(F ∩ FE(Λ1)).
The rest of this section is devoted to proving Theorem 5.3. First we establish some
preliminary results (Proposition 5.4 and Proposition 5.11). Proposition 5.4 describes the
relationship between an efficient set with its satiation. On the other hand, given a satiated
set, we construct an efficient set in Proposition 5.11.
Proposition 5.4. Suppose that Λ is a finitely aligned k-graph and that E ⊆ FE(Λ1) is
efficient. Then bE = E ∩ FE(Λ1).
Proof. To show bE ⊆ (E ∩ FE(Λ1)), take E ∈ bE. It is clear that E ∈ FE(Λ1). On the other
hand, by Remark 5.1, bE ⊆ E and E ∈ E. So bE ⊆ (E ∩ FE(Λ1)).
For (E ∩ FE(Λ1)) ⊆ bE, take E ∈ (E ∩ FE(Λ1)). Let {sE
relative Cuntz-Krieger (Λ; E)-family. So Qe∈E(sE
Corollary 4.9]. Since E ∈ FE(Λ1), Proposition 4.5(2) implies E ∈ bE.
λ : λ ∈ Λ} be the universal
e ) = 0 since E ∈ E and [10,
(cid:3)
Before stating Proposition 5.11, we establish some results that we use in the proof.
r(E) − sE
e sE∗
Lemma 5.5. Let Λ be a finitely aligned k-graph. Suppose that λ ∈ Λ, that m ∈ Nk,
and that E ⊆ r(λ)Λm. If ExtΛ(λ; E) 6= ∅, then there exists a unique n ∈ Nk such that
ExtΛ(λ; E) ⊆ s(λ)Λn and n ≤ m.
Proof. Take ν1, ν2 ∈ ExtΛ(λ; E). We show d(ν1) = d(ν2). So there exist µ1, µ2 ∈ E with
ν1 ∈ ExtΛ(λ; {µ1}) and ν2 ∈ ExtΛ(λ; {µ2}). Since E ⊆ r(λ)Λm, then d(µ1) = d(µ2) and
d(λ) + d(ν1) = d(λν1) = d(λ) ∨ d(µ1) = d(λ) ∨ d(µ2) = d(λν2) = d(λ) + d(ν2).
So d(ν1) = d(ν2). Then there exists a unique n ∈ Nk such that ExtΛ(λ; E) ⊆ s(λ)Λn.
10
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
To show n ≤ m, take ν ∈ ExtΛ(λ; E). There exists µ ∈ E ⊆ r(λ)Λm with ν ∈
ExtΛ(λ; {µ}). Then d(µ) = m and
d(λ) + n = d(λ) + d(ν) = d(λ) ∨ d(µ) ≤ d(λ) + d(µ) = d(λ) + m .
Thus n ≤ m.
Corollary 5.6. For E ⊆ Λ, we define
L(E) := max
µ∈E
d(µ)
(cid:3)
(L(E) := 0 if E = ∅). Then for v ∈ Λ0, λ ∈ vΛ and E ⊆ vΛ, L(ExtΛ(λ; E)) ≤ L(E).
Proof. If ExtΛ(λ; E) = ∅, then we are done. Suppose ExtΛ(λ; E) 6= ∅. Take µ ∈
ExtΛ(λ; E) with d(µ) = L(ExtΛ(λ; E)). So there exists γ ∈ E such that µ ∈ ExtΛ(λ; {γ}).
By Lemma 5.5, d(µ) ≤ d(γ) ≤ L(E). Therefore L(ExtΛ(λ; E)) ≤ L(E).
(cid:3)
Lemma 5.7. Suppose that Λ is a finitely aligned k-graph and that F is a satiated set.
Then F ⊆ (F ∩ FE(Λ1)).
Proof. For E ⊆ Λ and l ∈ N, we define
N(E; l) :=(cid:12)(cid:12){m ∈ Nk : m = l and there exists µ ∈ E with d(µ) = m}(cid:12)(cid:12) .
With a slight abuse of notation, N(E) := N(E; L(E)). To prove the lemma, we show
that for E ∈ FE(Λ),
(5.1)
E ∈ F implies E ∈ (F ∩ FE(Λ1)).
We use nested induction arguments on pairs in (L(E), N(E)). Our strategy is as follows:
We start out by proving that (5.1) is true for (L(E), N(E)) = (l, j) for j ∈ N by induction
on l. So step 1 is to show that (5.1) is true for (L(E), N(E)) = (1, j). Then for the
inductive step, we assume that l ≥ 2 and (5.1) is true for (L(E), N(E)) = (l − x, j)
for all j and 1 ≤ x ≤ l − 1. We prove that (5.1) is true for (L(E), N(E)) = (l, j) by
induction on l. Thus in step 2 we show (5.1) is true for (L(E), N(E)) = (l, 1) (using the
inductive hypothesis for l). Then we assume (5.1) is true for (L(E), N(E)) = (l, j − y)
with 1 ≤ y ≤ j − 1. Finally, step 3 is to show that, with these assumptions in place, (5.1)
holds for (L(E), N(E)) = (l, j).
Step 1: If E ∈ F with L(E) = 1, then E ∈ FE(Λ1) and E ∈ (F ∩ FE(Λ1)). So
E ∈ (F ∩ FE(Λ1)), as required.
Since the argument for Step 2 and Step 3 is similar, to save from repeating things we
take E ∈ F with (L(E), N(E)) = (l, j) where either j = 1 (Step 2) or j ≥ 2 (Step 3). To
show E ∈ (F ∩ FE(Λ1)), we prove that E can be constructed by applying processes in
(S1-4) to certain elements of (F ∩ FE(Λ1)).
Since (L(E), N(E)) = (l, j), there exist m1, . . . , mj ∈ Nk such that for 1 ≤ i ≤ j, we
have mi = l and there exists λi ∈ E with d(λi) = mi. Define
Emi := {λ ∈ E : d(λ) = mi} for 1 ≤ i ≤ j,
E′ := {λ ∈ Em1 : there exists νλ ∈ E with νλ 6= λ and νλν′
λ = λ},
E′′ := Em1\E′.
For λ ∈ E′′, we choose iλ with d(λ) ≥ eiλ and define λ′ := λ(0, d(λ) − eiλ). Since l ≥ 2,
then λ′ /∈ Λ0. Now we establish the following claims:
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
11
Claim 5.8. E\ E′ ∈ F .
Proof of Claim 5.8. For λ ∈ E′, we get λ ∈ Em1 ⊆ r(λ)Λm1, νλ /∈ E′ and νλ ∈ E\ E′. So
E\E′ = E\E′ ∪ {νλ : λ ∈ E′} = {λ : λ ∈ E\E′} ∪ {λ(0, d(νλ)) : λ ∈ E′}.
Since E ∈ F and F is satiated, then by (S3), E\ E′ ∈ F .
(cid:3) Claim 5.8
Claim 5.9. For λ ∈ E′′, we have λ′ /∈ EΛ and ExtΛ(λ′; E\ E′) ∈ (F ∩ FE(Λ1)).
Proof of Claim 5.9. Take λ ∈ E′′. Since λ ∈ E′′ = Em1\E′, then λ /∈ EΛ\E. Suppose for
contradiction that λ′ ∈ EΛ. Write λ′ := eµ with e ∈ E and µ ∈ Λ. Then
λ = λ′ [λ(d(λ) − eiλ, d(λ))] = eµ [λ(d(λ) − eiλ, d(λ))]
and λ ∈ EΛ\E, which contradicts λ /∈ EΛ\E. Thus λ′ /∈ EΛ.
Since λ′ /∈ EΛ, we have λ′ /∈ (E\E′)Λ and λ′ ∈ r(E\E′)Λ\(E\E′)Λ. Since E\E′ ∈ F
(Claim 5.8), by (S2), ExtΛ(λ′; E\E′) ∈ F . To show ExtΛ(λ′; E\E′) ∈ (F ∩ FE(Λ1)), we
now give separate arguments for j = 1 (step 2) and j ≥ 1 (step 3).
(Step 2) Suppose j = 1. Then for ν ∈ E\Em1, we have d(ν) ≤ l − 1 and by Corollary 5.6,
L(ExtΛ(λ′; {ν})) ≤ l − 1. Since (E\E′)\Em1 = E\Em1, this implies
L(ExtΛ(λ′; (E\E′)\Em)) ≤ l − 1.
For µ ∈ Em1 ⊆ r(E)Λm1, we have d(λ′) = m1 − eiλ = d(µ) − eiλ, so (d(λ′) ∨ d(µ)) −
d(λ′) = eiλ and L(ExtΛ(λ′; {µ})) ≤ 1. Then L(ExtΛ(λ′; Em)) ≤ 1 and
L(ExtΛ(λ′; E\E′)) = max{L(ExtΛ(λ′; (E\E′)\Em))), L(ExtΛ(λ′; Em))} ≤ l − 1.
Since ExtΛ(λ′; E\E′) ∈ F and L(ExtΛ(λ′; E\E′)) ≤ l − 1, by the inductive hy-
pothesis for l, we have ExtΛ(λ′; E\E′) ∈ (F ∩ FE(Λ1)), as required.
(Step 3) Since we have now verified both bases cases, we have both l ≥ 2 and j ≥ 2.
Take 2 ≤ i ≤ j. For ν ∈ Emi, we have d(ν) = mi = l and by Corollary 5.6,
L(ExtΛ(λ′; {ν})) ≤ d(ν) = l. So L(ExtΛ(λ′; Emi)) ≤ l. If ExtΛ(λ′; Emi) = ∅,
then N(ExtΛ(λ′; Emi); l) = 0, otherwise, since Emi ⊆ s(λ)Λmi, by Lemma 5.5,
there exists a unique n ∈ Nk with n ≤ mi = l such that ExtΛ(λ; Emi) ⊆ s(λ)Λn.
Hence in either case, N(ExtΛ(λ′; Emi); l) ≤ 1. Therefore
(5.2)
(5.3)
L(ExtΛ(λ′; [2≤i≤j
Emi)) ≤ l and N(ExtΛ(λ′; [2≤i≤j
Emi); l) ≤ X2≤i≤j
1 = j − 1.
On the other hand, for µ ∈ Em1 ⊆ r(E)Λm1, we have d(λ′) = m1 −eiλ = d(µ)−eiλ,
so (d(λ′) ∨ d(µ)) − d(λ′) = eiλ and L(ExtΛ(λ′; {µ})) ≤ 1. Thus
L(ExtΛ(λ′; Em1)) ≤ 1 and N(ExtΛ(λ′; Em1); l) = 0
since l ≥ 2. Now note that for ν ∈ (E\E′)\S1≤i≤j Emi, we have d(ν) ≤ l − 1
and by Corollary 5.6, L(ExtΛ(λ′; {ν}) ≤ l − 1. Hence
(5.4) L(ExtΛ(λ′; (E\E′)\ [1≤i≤j
Therefore by (5.2), (5.3), and (5.4), we get
L(ExtΛ(λ′; E\E′)) = max{
Emi)) ≤ l − 1 and N(ExtΛ(λ′; (E\E′)\ [1≤i≤j
L(ExtΛ(λ′;S2≤i≤j+1 Emi)), L(ExtΛ(λ′; Em1)),
L(ExtΛ(λ′; (E\E′)\S1≤i≤j+1 Emi))
}
≤ max{l, 1, l − 1} = l,
Emi); l) = 0.
12
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
N(ExtΛ(λ′; E\E′); l) = N(ExtΛ(λ′; [2≤i≤j+1
Emi); l) + N(ExtΛ(λ′; Em1); l)
+ N(ExtΛ(λ′; (E\E′)\ [1≤i≤j+1
≤ (j − 1) + 0 + 0 = j − 1.
Emi); l)
Hence L(ExtΛ(λ′; E\E′)) is either equal to l with N(ExtΛ(λ′; E\E′); l) ≤ j − 1; or
strictly less than l. In either case, by the inductive hypotheses, ExtΛ(λ′; E\E′) ∈
(F ∩ FE(Λ1)) since ExtΛ(λ′; E\E′) ∈ F .
Therefore, for λ ∈ E′′, we have ExtΛ(λ′; E\E′) ∈ (F ∩ FE(Λ1)).
Claim 5.10. G := ((E\E′)\E′′ ∪Sλ∈E ′′{λ′}) ∈ (F ∩ FE(Λ1)).
Proof of Claim 5.10. Note that E\ E′ ∈ F (Claim 5.8). Then by (S3),
(cid:3) Claim 5.9
G = {λ : λ ∈ (E\E′)\E′′} ∪ {λ(0, d(λ) − eiλ) : λ ∈ E′′} ∈ F .
If j = 1, then Em1 contains all paths in E with absolute length l and L(G) ≤ l − 1. If
j ≥ 2, then Em1 contains E ∩ Λm1 with m1 = l, so L(G) = l and N(G) = j − 1. In either
case, by the inductive hypothesis, G ∈ (F ∩ FE(Λ1)).
(cid:3) Claim 5.10
Now we show E ∈ (F ∩ FE(Λ1)). For every λ ∈ E′′, we have λ′ /∈ E (Claim 5.9)
and then (E\E′)\E′′ = (G\Sλ∈E ′′ λ′). Because G ∈ (F ∩ FE(Λ1)) (Claim 5.10) and for
λ ∈ E′′, ExtΛ(λ′; E\E′) ∈ (F ∩ FE(Λ1)) (Claim 5.9), by (S4),
λ′ ExtΛ(λ′; E\E′), there exists nν ∈ Nk such that
ν(0, d(nν)) ∈ E\E′. For ν ∈ (E\E′)\E′′, set nν := d(ν). By (S3), (5.5) implies
{ν(0, nν) : ν ∈ F } ∈ (F ∩ FE(Λ1)).
Note that {ν(0, nν) : ν ∈ F } ⊆ E\E′ ⊆ E and by (S1), E ∈ (F ∩ FE(Λ1)). So (5.1) is
true for (L(E), N(E)) = (l, j).
(cid:3)
Proposition 5.11. Suppose that Λ is a finitely aligned k-graph and that F is a satiated
set. Then min(F ∩ FE(Λ1)) is efficient and min(F ∩ FE(Λ1)) = F .
Proof. We show that min(F ∩ FE(Λ1)) is efficient. To show (E1), take E, F ∈ min(F ∩
FE(Λ1)) with E ⊆ F . By definition of min(F ∩ FE(Λ1)), we have E = F and (E1)
holds. For (E2), take E ∈ min(F ∩ FE(Λ1)) and g ∈ r(E)Λ1\E. Since E ∈ FE(Λ1),
by Corollary 3.2, ExtΛ(g; E) ∈ FE(Λ1). Since E ∈ F , by (S2), ExtΛ(g; E) ∈ F . Then
ExtΛ(g; E) ∈ (F ∩ FE(Λ1)) and there exists F ∈ min(F ∩ FE(Λ1)) with F ⊆ ExtΛ(g; E).
So (E2) also holds. To show (E3), take E, F ∈ min(F ∩ FE(Λ1)) and g ∈ Er(F ). Define
EF := (E\{g}) ∪ {(gf )(0, d(f )) : f ∈ F }.
Since E, F ∈ F , by (S4), EF ∈ F and EF ∈ FE(Λ). Then EF ∈ FE(Λ1) since EF ⊆
r(EF )Λ1. So EF ∈ (F ∩ FE(Λ1)) and there exists G ∈ min(F ∩ FE(Λ1)) with G ⊆ EF .
Therefore (E3) holds and min(F ∩ FE(Λ1)) is efficient.
(5.5)
F := (E\E′)\E′′ ∪ [λ∈E ′′
= ((G\ [λ∈E ′′
λ′) ∪ [λ∈E ′′
On the other hand, for ν ∈ Sλ∈E ′
m1
λ′ ExtΛ(λ′; E\E′))
λ′ ExtΛ(λ′; E\E′)) ∈ (F ∩ FE(Λ1)).
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
13
Now we show min(F ∩ FE(Λ1)) = F . Since both min(F ∩ FE(Λ1)) and F ∩ FE(Λ1)
have the same edge satiation, then by Remark 5.1, min(F ∩ FE(Λ1)) = (F ∩ FE(Λ1)). So
it suffices to show (F ∩ FE(Λ1)) = F . Note that (F ∩ FE(Λ1)) ⊆ F . So (F ∩ FE(Λ1)) ⊆
F = F since F is satiated. The other inclusion follows from Lemma 5.7.
(cid:3)
We are finally ready to prove Theorem 5.3.
Proof of Theorem 5.3. To show injectivity, take efficient sets E1 and E2 with E1 = E2. By
Proposition 5.4, bE1 = E1 ∩ FE(Λ1) = E2 ∩ FE(Λ1) = bE2. By Lemma 3.10, E1 = E2. For
surjectivity, take a satiated set F . By Proposition 5.11, min(F ∩ FE(Λ1)) is efficient and
F = min(F ∩ FE(Λ1)), as required.
(cid:3)
A direct consequence of Theorem 5.3 is:
Corollary 5.12. The map θ : F 7→ F ∩ FE(Λ1) is a bijective map between satiated sets
and edge satiations of efficient sets. Furthermore θ preserves containment in the sense
that F1 F2 implies (F1 ∩ FE(Λ1)) (F2 ∩ FE(Λ1)).
Proof. The bijectivity of θ follows Lemma 3.10 and Theorem 5.3. Take satiated sets F1, F2
⊆ FE(Λ) such that F1 F2. We trivially have (F1 ∩ FE(Λ1)) ⊆ (F2 ∩ FE(Λ1)). Suppose
for contradiction that (F1 ∩ FE(Λ1)) = (F2 ∩ FE(Λ1)). Then
F1 = θ−1(F1 ∩ FE(Λ1)) = θ−1(F2 ∩ FE(Λ1)) = F2,
which contradicts F1 6= F2. The conclusion follows.
(cid:3)
6. Applications
6.1. The gauge-invariant uniqueness theorem for relative Cuntz-Krieger alge-
bras. In [10], Sims introduced two uniqueness theorems for relative Cuntz-Krieger alge-
bras, namely the gauge-invariant uniqueness theorem [10, Theorem 6.1] and the Cuntz-
Krieger uniqueness theorem [10, Theorem 6.3].
In this subsection, we show how our
Theorem 5.3 simplifies the hypothesis of [10, Theorem 6.1] as follows:
Theorem 6.1 (The gauge-invariant uniqueness theorem). Let Λ be a finitely aligned
k-graph such that E ⊆ FE(Λ1) is efficient. Suppose that {Sλ : λ ∈ Λ} is a relative
Cuntz-Krieger (Λ; E)-family in a C ∗-algebra B which satisfies:
(G1) Sv 6= 0 for all v ∈ Λ0;
(G2) Qe∈E(Sr(E) − SeS∗
(G3) there exists an action θ : Tk → Aut(B) with θz(Sλ) = zd(λ)Sλ for z ∈ Tk,λ ∈ Λ.
The homomorphism πS obtained from the universal property of C ∗(Λ; E), is injective.
e ) 6= 0 for all E ∈ FE(Λ1)\bE; and
The only difference between our Theorem 6.1 and Sims's gauge-invariant uniqueness
theorem [10, Theorem 6.1] is (G2), which in [10], (G2) is replaced by the following condi-
tion:
Hence in order to show the two uniqueness theorem are identical, it suffices to show that
(G2) is equivalent to (6.1).
(6.1)
(Sr(E) − SλS∗
λ) 6= 0 for all E ∈ FE(Λ)\E.
Yλ∈E
14
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
Lemma 6.2. Suppose that {Sλ : λ ∈ Λ} is a relative Cuntz-Krieger (Λ; E)-family in a
C ∗-algebra B. Then the following two conditions are equivalent:
λ) 6= 0 for all E ∈ FE(Λ)\ E.
(a) Qλ∈E(Sr(E) − SλS∗
(b) Qe∈E(Sr(E) − SeS∗
e ) 6= 0 for all E ∈ FE(Λ1)\bE.
Proof. To show (a)⇒(b), it suffices to show FE(Λ1)\bE ⊆ FE(Λ)\E. Take E ∈ FE(Λ1)\bE.
Then E ∈ FE(Λ). Since bE = E ∩ FE(Λ1) (Proposition 5.4), then E /∈ bE implies E /∈ E.
Therefore E ∈ FE(Λ)\ E and FE(Λ1)\bE ⊆ FE(Λ)\E.
Qe∈E(Sr(E) − SeS∗
a relative Cuntz-Krieger (Λ; E)-family and Qe∈E(Sr(E) − SeS∗
Qf ∈F (Sr(F ) − Sf S∗
For (b)⇒(a), suppose for contradiction that there exists E ∈ FE(Λ)\E such that
e ) = 0. Consider the set E1 := E ∪ E. Since {Sλ : λ ∈ Λ} is
e ) = 0, then for F ∈ E1,
f ) = 0. Following the argument of [10, Corollary 5.6], this implies
(6.2)
(Sr(F ) − Sf S∗
f ) = 0 for F ∈ E1.
Yf ∈F
Since E ∈ FE(Λ)\E and E1 = E ∪ E, we have E E1. So E E1 since E1 ⊆ E1.
By Proposition 5.4 and Corollary 5.12, bE = (E ∩ FE(Λ1)) (E1 ∩ FE(Λ1)). Take F ∈
E1 ∩ FE(Λ1) with F /∈ bE. Since F ∈ E1, by (6.2), Qf ∈F (Sr(F ) − Sf S∗
contradicts Condition (b) since F ∈ FE(Λ1)\bE. The conclusion follows.
An advantage of our version of the theorem is that our condition (G2) is more checkable;
f ) = 0, which
(cid:3)
there are fewer sets to consider.
6.2. Gauge-invariant ideals in a relative Cuntz-Krieger algebra. Another applica-
tion of Theorem 5.3 is to give a complete listing of the gauge-invariant ideals in a relative
Cuntz-Krieger algebra (Theorem 6.4). This is a simplification of Theorem 4.6 of [12] (see
Remark 6.7). First we give some preliminary notation and results.
Suppose that Λ is a finitely aligned k-graph and that E ⊆ FE(Λ1). We define a relation
≥ on Λ0 by w ≥ v if and only if vΛw 6= ∅. A subset H ⊆ Λ0 is hereditary if v ∈ H
and w ≥ v imply w ∈ H; and H is E-saturated if, whenever v ∈ Λ0 and E ∈ vE with
s(E) ⊆ H, we have v ∈ H. For a hereditary subset H ⊆ Λ0, the subcategory
is a finitely aligned k-graph (see [11, Lemma 4.1]).
Λ\ΛH := {λ ∈ Λ : s(λ) /∈ H}
Lemma 6.3. Suppose that Λ is a finitely aligned k-graph, that E ⊆ FE(Λ1) is an efficient
set, and that H ⊆ Λ0 is a E-saturated hereditary set. Then
is a subset of FE((Λ\ΛH)1).
EH := {E\EH : E ∈ E}
Proof. Take E ∈ EH. Write E = F \F H with F ∈ E. Since F ⊆ r(E)Λ1 and F is finite,
we have E ⊆ r(E)(Λ\ΛH)1 and E is finite.
To show that E is exhaustive, take λ ∈ r(E)(Λ\ΛH). If λ ∈ F Λ, then s(λ) /∈ H implies
λ ∈ EΛ and ExtΛ\ΛH (λ; E) 6= ∅, as required. So suppose λ ∈ r(F )Λ\F Λ. Suppose for
contradiction that (Λ\ΛH)min(λ, e) = ∅ for every e ∈ E. Hence Λmin(λ, e) ⊆ ΛH ×ΛH for
e ∈ E, so ExtΛ(λ; E) ⊆ ΛH. Note that ExtΛ(λ; F ) = ExtΛ(λ; E)∪ExtΛ(λ; F H) and since
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
15
H is hereditary, ExtΛ(λ; F H) ⊆ ΛH. So ExtΛ(λ; F ) ⊆ ΛH and s(ExtΛ(λ; F )) ⊆ H. Be-
cause λ ∈ r(F )Λ\F Λ and F ∈ E, by (E2), there exists F ′ ∈ E with F ′ ⊆ ExtΛ(λ; F ). Be-
cause H is E-saturated, s(F ′) ⊆ s(ExtΛ(λ; F )) ⊆ H and F ′ ∈ E, then s(λ) = r(F ′) ∈ H,
which contradicts λ ∈ r(E)(Λ\ΛH). So there exists e ∈ E such that (Λ\ΛH)min(λ, e) 6= ∅.
Thus E is exhaustive and E ∈ FE((Λ\ΛH)1).
(cid:3)
Now we state our classification theorem.
Theorem 6.4. Let Λ be a finitely aligned k-graph such that E ⊆ FE(Λ1) is an efficient
set. Suppose that {sE
λ : λ ∈ Λ} is the universal relative Cuntz-Krieger (Λ; E)-family. For
H ⊆ Λ0 and B ⊆ FE((Λ\ΛH)1), define IH,B to be the ideal generated by
{sE
v : v ∈ H} ∪nYe∈E
(sE
r(E) − sE
e sE∗
e ) : E ∈ Bo.
Then Φ : (H, B) 7→ IH,B is a bijection between the set of pairs (H, B) such that H is a
E-saturated hereditary set and B ⊆ FE((Λ\ΛH)1) is an efficient set such that EH ⊆ bB,
and the gauge-invariant ideals in C ∗(Λ; E).
For any gauge-invariant ideal I, define
HI := {v ∈ Λ0 : sE
v ∈ I},
BI := min(cid:16)nE ∈ FE((Λ\ΛHI)1) :Ye∈E
(sE
r(E) − sE
e sE∗
e ) ∈ Io(cid:17).
Then the inverse of Φ is given by I 7→ (HI, BI).
The rest of this subsection is devoted to proving Theorem 6.4.
Lemma 6.5. Suppose that I is an ideal in C ∗(Λ; E). Then, HI is E-saturated hereditary
and BI is an efficient set with EHI ⊆cBI .
Proof. To show HI is hereditary, take v ∈ HI and w ∈ Λ0 with w ≥ v. Take λ ∈ vΛw. So
v ∈ HI ⇒ sE
v ∈ I ⇒ sE
λ = sE
v sE
λ ∈ I ⇒ sE
w = sE∗
λ sE
λ ∈ I ⇒ w ∈ HI
and HI is hereditary.
To show that HI is E-saturated, take v ∈ Λ0 and E ∈ vE with s(E) ⊆ HI. For e ∈ E,
s(e) ∈ HI ⇒ sE
s(e) ∈ I ⇒ sE
e = sE
e sE
s(e) ∈ I ⇒ sE
e sE∗
e ∈ I.
Since E ∈ E and {sE
λ : λ ∈ Λ} is a commuting family (see [9, Lemma 2.7(i)]),
λsE∗
0 =Ye∈E
(sE
v − sE
e sE∗
e ) = sE
v + XF ⊆E,F 6=∅
(−1)F Ye∈F
(sE
e sE∗
e )
and since sE
e sE∗
e ∈ I for each e ∈ E, sE
v ∈ I. Therefore v ∈ HI and HI is E-saturated.
C ∗(Λ; E) → C ∗(Λ; E)/I. We show that BI is efficient. Write Γ := Λ\ΛHI. (E1) follows
r(E) −
To show that BI is an efficient set with EHI ⊆ cBI, consider the quotient map q :
from definition of BI . For (E2′), take E ∈ BI and f ∈ r(E)Γ1\E. So q(Qe∈E(sE
e )) = 0. By Proposition 3.1,
e sE∗
sE
(sE
s(f ) − sE
g sE∗
g )) = 0.
q( Yg∈ExtΓ(f ;E)
So ExtΓ(f ; E) ∈cBI and (E2′) holds.
16
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
For (E3′), take E ∈ BI , e ∈ E, and F ∈ s(e)BI . Define
EF := (E\{e}) ∪ {(ef )(0, d(f )) : f ∈ F }.
Note that
By Proposition 3.3,
q(Yf ∈E
(sE
r(E) − sE
f sE∗
f )) = 0 = q(Yf ∈F
(sE
s(E) − sE
f sE∗
f )).
q(Yf ∈EF
(sE
r(E) − sE
f sE∗
f )) = 0.
(sE
(sE
r(E))
e sE∗
e sE∗
e sE∗
e sE∗
e sE∗
r(E)) and
r(E) − sE
r(E) − sE
r(E) − sE
r(E) − sE
e )) = q(sE
s(e) ∈ I and sE
q(cid:16)Ye∈E
So EF ∈cBI. Thus (E3′) holds and BI is efficient.
We show that cBI contains EHI . Take E ∈ EHI , we show q(Qe∈E(sE
e )) = 0.
Since E ∈ EHI , there exists E′ ∈ E such that E = E′\E′HI. Then E′ = E ∪ E′HI. For
e ∈ E′HI, we have sE
e ∈ I. So q(Qe∈E ′HI
e )(cid:17) = q(cid:16)Ye∈E
= q(cid:16)Ye∈E
= q(cid:16)Ye∈E ′
since E′ ∈ E. So E ∈cBI and EHI ⊆cBI.
HI is E-saturated hereditary and BI is efficient such that EHI ⊆cBI . We show I = IHI ,BI .
Proof of Theorem 6.4. We use a similar argument to [6, Theorem 4.9]. Write Γ := Λ\ΛH.
To show that Φ is surjective, take a gauge-invariant ideal I in C ∗(Λ; E). By Lemma 6.5,
Since all the generators of IHI ,BI are in I, we have IHI ,BI ⊆ I. To show the reverse
inclusion, consider the quotient maps
e )(cid:17)q(sE
e ) Ye∈E ′HI
e )(cid:17) = q(0) = 0
e )(cid:17)
r(E) − sE
r(E) − sE
r(E) − sE
e sE∗
e sE∗
e sE∗
(sE
(sE
(sE
(sE
(cid:3)
qI : C ∗(Λ; E) → C ∗(Λ; E)/I, qIHI ,BI
: C ∗(Λ; E) → C ∗(Λ; E)/IHI ,BI , and
: C ∗(Λ; E)/IHI,BI → C ∗(Λ; E)/I = (C ∗(Λ; E)/IHI ,BI )/(I/IHI ,BI ).
qI/IHI ,BI
So qI = qI/IHI ,BI
relative Cuntz-Krieger (Γ; BI)-families. Since qI, qIHI ,BI
0 = qIHI ,BI
Cuntz-Krieger Γ-families. For E ∈ BI , based on definition of BI and IHI ,BI , we have
λ) : λ ∈ Γ} are
λ) =
λ) : λ ∈ Γ} are Toeplitz-
λ) for λ ∈ ΛH, both {qI(sE
are quotient maps and qI(sE
λ) : λ ∈ Γ} and {qIHI ,BI
λ) : λ ∈ Γ} and {qIHI ,BI
. We claim both {qI(sE
◦ qIHI ,BI
(sE
(sE
(sE
qI(cid:16)Ye∈E
(sE
r(E) − sE
e sE∗
e )(cid:17) = 0 = qIHI ,BI(cid:16)Ye∈E
(sE
r(E) − sE
e sE∗
e )(cid:17).
Let πHI ,BI
So both families satisfy (CK) and are relative Cuntz-Krieger (Γ; BI)-families, as claimed.
: C ∗(Γ; BI) → C ∗(Λ; E)/IHI ,BI and πI : C ∗(Γ; BI) → C ∗(Λ; E)/I be the
homomorphisms obtained from the universal property of C ∗(Γ; BI). So πI and qIHI ,BI
◦
πHI ,BI are homomorphisms which agree on {sE
µ : λ, µ ∈ Γ}, which is the generators of
C ∗(Γ; BI) (see [9, Lemma 2.7(iv)]), and hence are equal. We use Theorem 6.1 to show
λsE∗
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
17
that πI is injective. By definition of HI, qI(sE
v ) 6= 0 for all v ∈ Γ0, so {qI(sE
λ) : λ ∈ Γ}
satisfies (G1). If E ∈ FE(Γ1) such that E /∈cBI , then
r(E) − sE
e sE
(sE
e∗)) 6= 0
qI(Ye∈E
by definition of BI. So {qI(sE
λ) : λ ∈ Γ} satisfies (G2). Since I is gauge-invariant, then
the gauge action β on C ∗(Λ; E) descends to an action θ on C ∗(Λ; E)/I which satisfies
(G3). Therefore, by Theorem 6.1, πI is injective. Since πHI ,BI
is surjective and πI =
qI/IHI ,B
is also injective. Thus IHI ,BI = I,
as required.
◦πHI ,BI , the injectivity of πI implies that qI/IHI ,B
I
I
Next we show the injectivity of Φ. Take a E-saturated hereditary set H and an efficient
set B ⊆ FE(Γ1) such that EH ⊆ bB. Define
HIH,B := {v ∈ Λ0 : sE
v ∈ IH,B},
BIH,B := min(cid:16)nE ∈ FE(Γ1) :Ye∈E
(sE
r(E) − sE
e sE∗
e ) ∈ IH,Bo(cid:17).
To show that Φ is injective, we show H = HIH,B and B = BIH,B. We trivially have
H ⊆ H1 and bB ⊆ [BIH,B. To prove the reverse inclusion, consider the universal relative
λ : λ ∈ Γ}. For λ ∈ Λ, we define
Cuntz-Krieger (Γ; B)-family {sB
Sλ :=(sB
λ
0
if s(λ) /∈ H,
if s(λ) ∈ H.
We show that {Sλ : λ ∈ Λ} is a relative Cuntz-Krieger (Λ; E)-family. It is clear that the
family satisfies (TCK1-3). To show (CK), take E ∈ E. Then
Ye∈E
(Sr(E) − SeS∗
(Sr(E) − SeS∗
(Sr(E) − SeS∗
e )
e ) = Ye∈EH
= Sr(E) Ye∈E\EH
e ) Ye∈E\EH
r(E) Ye∈E\EH
(Sr(E) − SeS∗
e ) = sB
(sB
r(E) − sB
e sB∗
e ) = 0
Krieger (Λ; E)-family and the universal property of C ∗(Λ; E) gives a homomorphism πS
of C ∗(Λ; E) into C ∗(Γ; B) with πS(sE
v ) = 0 for v ∈ H and
since E\EH ∈ EH ⊆ bB and Proposition 3.11. So {Sλ : λ ∈ Λ} is a relative Cuntz-
πS(Qe∈E(sE
e )) = 0 for E ∈ B, we have IH,B ⊆ ker πS. By Proposition 4.5,
v 6= 0 ⇒ Sv 6= 0 ⇒ πS(sE
λ) = Sλ for λ ∈ Λ. Since πS(sE
e sE∗
r(E) − sE
v /∈ H ⇒ sB
v /∈ ker πS ⇒ sE
v ) 6= 0 ⇒ sE
v /∈ IH,B,
r(E) − sB
e sB∗
(Sr(E) − SeS∗
e ) 6= 0
(sB
E ∈ FE(Γ1)\bB ⇒Ye∈E
⇒ πS(cid:16)Ye∈E
⇒Ye∈E
(sE
e ) 6= 0 ⇒Ye∈E
e )(cid:17) 6= 0 ⇒Ye∈E
e sE∗
r(E) − sE
e sE∗
e ) /∈ IH,B.
(sE
r(E) − sE
(sE
r(E) − sE
e sE∗
e ) /∈ ker πS
Then HIH,B ⊆ H and [BIH,B ⊆ bB. Hence H = HIH,B and bB = [BIH,B. By Lemma 3.10,
B = BIH,B. Therefore Φ is injective and hence an isomorphism.
(cid:3)
18
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
Remark 6.6. From the proof of Theorem 6.4, we get:
(a) For a gauge-invariant ideal I, we have IHI ,BI = I.
(b) For a hereditary set H and an efficient set B ⊆ FE((Λ\ΛH)1) such that EH ⊆ bB,
Part (a) is from the surjectivity of Φ. The injectivity of Φ implies part (b).
we have HIH,B = H and BIH,B = B.
Remark 6.7. In [12], Sims, Whitehead and Whittaker gives a complete listing of the gauge-
invariant ideals in a twisted C ∗-algebra associated to a higher-rank graph [12, Theorem
4.6]. Since their twisted C ∗-algebras can be viewed as a generalisation of relative Cuntz-
Krieger algebras, they actually have an alternative version of Theorem 6.4, which uses
satiated sets rather than efficient sets. Indeed, we could have shown Theorem 6.4 as a
consequence of Theorem 4.6 of [12]. However, the direct argument above takes about the
same amount of effort.
6.3. Toeplitz algebras and their quotient algebras. Throughout this subsection,
suppose that Λ is a row-finite k-graph with no sources. In this subsection, we study the
relationship between Toeplitz algebras and their ideals and quotient algebras.
Historically, one was forced to consider satiations whenever working with ideals and
quotients of higher-rank graph Toeplitz algebras. For example in [3, Appendix A], an
Huef, Kang and Raeburn must prove results about satiations even though they are really
only interested in E := {Si∈K vΛei : v ∈ Λ0} where K is a nonempty subset of {1, . . . , k}.
Remark 3.7 tells us that {Si∈K vΛei : v ∈ Λ0} is efficient and so we have established tools
Our next theorem is a special case of Theorem 6.4, which lists all the gauge-invariant
ideals in a higher-rank graph Toeplitz algebra. Here we adjust some notation as explained
in Remark 6.9.
that allow us to avoid these unruly satiations.
Theorem 6.8. Suppose that Λ is a row-finite k-graph with no sources and that {tλ : λ ∈
Λ} is the universal Toeplitz-Cuntz-Krieger Λ-family.
(a) Suppose that H ⊆ Λ0 and that E ⊆ FE((Λ\ΛH)1). The ideal IH,E, as defined in
Theorem 6.4, is a gauge-invariant ideal in T C ∗(Λ).
(b) Suppose that I is a gauge-invariant ideal of T C ∗(Λ). Suppose that HI and EI are
as defined in Theorem 6.4. Then HI is a hereditary subset of Λ, and EI is an
efficient subset of FE((Λ\ΛHI)1).
(c) Suppose that I is a gauge-invariant ideal of T C ∗(Λ). If E ∈ FE((Λ\ΛHI)1) such
thatQe∈E(tr(E) − tet∗
e) ∈ I, then there exists F ∈ EI such that F ⊆ E.
(d) For a gauge-invariant ideal I, we have IHI ,EI = I.
(e) For a hereditary set H of Λ and an efficient set E ⊆ FE((Λ\ΛH)1), we have
HIH,E = H and EIH,E = E.
Remark 6.9. A number of aspects are worth commenting on:
(i) Since Toeplitz-Cuntz-Krieger Λ-families coincide with relative Cuntz-Krieger (Λ; ∅)-
families, then to simplify the notation, we replace B of Theorem 6.4 with E.
(ii) Part (c) holds since EI is efficient. Part (d) and (e) follows from Remark 6.6.
Here for B ⊆ T C ∗(Λ), we write hBi to denote the ideal of T C ∗(Λ) generated by the
elements of B. Note that hBi is the smallest ideal which contains B.
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
19
Proposition 6.10. Suppose that Λ is a row-finite k-graph with no sources and that {tλ :
λ ∈ Λ} is the universal Toeplitz-Cuntz-Krieger Λ-family. Suppose that K, L are nonempty
subsets of {1, . . . , k}. Define
A :=nYi∈K(cid:16)tv − Xe∈vΛei
tet∗
e(cid:17) : v ∈ Λ0o and B :=nYj∈L(cid:16)tv − Xe∈vΛej
tet∗
e(cid:17) : v ∈ Λ0o.
Then the diagram
0
↓
0
↓
↓
↓
0 → hABi → hAi → hA∪Bi
0 → hBi → T C ∗(Λ) → T C ∗(Λ)
0 → hA∪Bi
↓
hAi → T C ∗(Λ)
↓
0
↓
hAi → T C ∗(Λ)
↓
0
↓
0
0
↓
hBi → 0
↓
hBi → 0
↓
hA∪Bi → 0
is commutative and all the rows and columns are exact.
Before giving the proof, we establish a stepping stone result.
Lemma 6.11. Suppose that K, L are nonempty subsets of {1, . . . , k} and that A, B are
as in Theorem 6.10. Then hAi + hBi = hA ∪ Bi and hAi ∩ hBi = hABi = hBAi.
Proof. We have A ⊆ hA ∪ Bi and B ⊆ hA ∪ Bi. Because hAi is the smallest ideal which
contains A, hAi ⊆ hA ∪ Bi and similarly, hBi ⊆ hA ∪ Bi. So hAi + hBi ⊆ hA ∪ Bi.
We show the reverse inclusion. Since A ⊆ hAi + hBi and B ⊆ hAi + hBi, we have
A ∪ B ⊆ hAi + hBi. So hA ∪ Bi ⊆ hAi + hBi since hAi + hBi is a closed ideal, as required.
Now we show hAi ∩ hBi = hABi. Write I := hAi ∩ hBi. First we show that I is a
gauge-invariant ideal. Define E := {Si∈K vΛei : v ∈ Λ0} and we have
(tr(E) − tet∗
(tv − tet∗
hAi =DnYi∈K Ye∈vΛei
e) : v ∈ Λ0oE =DnYe∈E
e) : E ∈ EoE.
By Remark 3.7, E is efficient and by Theorem 6.8(a), we have
(6.3)
is a gauge-invariant ideal. Similarly, hBi is a gauge-invariant ideal. For z ∈ Tk, we have
γz(hAi) ⊆ hAi and γz(hBi) ⊆ hBi where γ is the gauge action of Tk on T C ∗(Λ). Then
hAi = I∅,E
γz(I) ⊆ γz(hAi) ⊆ hAi and γz(I) ⊆ γz(hBi) ⊆ hBi .
So γz(I) ⊆ hAi ∩ hBi = I and I is a gauge-invariant ideal.
Now we investigate HI and EI. For v ∈ Λ0, we have
(6.4)
(tv − tet∗
e) ∈ hAi ∩ hBi = I.
Yi∈K∪L Ye∈vΛei
Thus I 6= 0. Since I is a gauge invariant ideal, by Theorem 6.8(b), HI is hereditary and
EI is efficient . By Theorem 6.8(e), (6.3) implies HI∅,E = ∅ and HhAi = ∅. Thus for v ∈ Λ0,
tv /∈ hAi and since I ⊆ hAi, tv /∈ I. Hence
(6.5)
HI = ∅.
20
LISA ORLOFF CLARK AND YOSAFAT E. P. PANGALELA
e) ∈ I ⊆
e) ∈ I
(see (6.4)), by Theorem 6.8(c), there exists F ∈ EI such that F ⊆ G. So F ⊆ G ⊆ E. Since
We claim EI = {Si∈K∪L vΛei : v ∈ Λ0}. To show EI ⊆ {Si∈K∪L vΛei : v ∈ Λ0}, take
E ∈ EI. Write G :=Si∈K∪L r(E)Λei. We claim E = G. We haveQe∈E(tr(E) − tet∗
hAi = I∅,E (see (6.3)). Since E ∈ FE(Λ1), by Theorem 6.8(c),Si∈K r(E)Λei ⊆ E. Using a
similar argument to hBi,Sj∈L r(E)Λej ⊆ E. Hence G ⊆ E. SinceQe∈G(tr(E) − tet∗
E, F ∈ EI, by (E1), F = E and then E = G. Therefore EI ⊆ {Si∈K∪L vΛei : v ∈ Λ0}.
To show {Si∈K∪L vΛei : v ∈ Λ0} ⊆ EI, take v ∈ Λ0 and write E :=Si∈K∪L vΛei. We
show E ∈ EI. By (6.4), Qe∈E(tv − tet∗
F ∈ vEI such that F ⊆ E. Since EI ⊆ {Si∈K∪L vΛei : v ∈ Λ0}, we have F =Si∈K∪L vΛei.
EI =n [i∈K∪L
Since HI = ∅ and EI = {Si∈K∪L vΛei : v ∈ Λ0} ((6.5) and (6.6)), by Theorem 6.8(d),
vΛei : v ∈ Λ0o.
e) : v ∈ Λ0oE = I = hAi ∩ hBi .
e) ∈ I and then by Theorem 6.8(c), there exists
Dn Yi∈K∪L Ye∈vΛei
Hence E = F ∈ EI. Therefore
(tv − tet∗
(6.6)
On the other hand, we have
AB = {ab : a ∈ A, b ∈ B}
=n(cid:16)Yi∈K Ye∈vΛei
(tv − tet∗
e)(cid:17)(cid:16)Yj∈L Ye∈vΛej
(tv − tet∗
e)(cid:17) : v ∈ Λ0o
and then hAi ∩ hBi = hABi. Using a similar argument, hAi ∩ hBi = hBAi.
(cid:3)
Proof of Proposition 6.10. The second row and the second column are exact. On the
other hand, using the third isomorphism theorem, we get the exactness of the third row
and the third column. Next we show that the first row and the first column are exact.
Lemma 6.11 tells hAi + hBi = hA ∪ Bi and hAi ∩ hBi = hABi = hBAi. Then by the
second isomorphism theorem,
hAi
hABi
hBi
hBAi
=
=
hAi
hAi ∩ hBi
hBi
hAi ∩ hBi
∼=
∼=
hAi + hBi
hBi
hAi + hBi
hAi
=
=
hA ∪ Bi
hBi
hA ∪ Bi
hAi
,
.
References
(cid:3)
[1] N.J. Fowler and I. Raeburn, The Toeplitz algebra of a Hilbert bimodule, Indiana Univ. Math. J. 48
(1999), 155 -- 181.
[2] R. Hazlewood, I. Raeburn, A. Sims and S.B.G. Webster, Remarks on some fundamental results about
higher-rank graphs and their C ∗-algebras, Proc. Edinb. Math. Soc. 56 (2013), 575 -- 597.
[3] A. an Huef, S. Kang and I. Raeburn, Spatial realisations of KMS states on the C ∗-algebras of
higher-rank graphs, J. Math. Anal. Appl. 427 (2015), 977 -- 1003.
[4] A. Kumjian and D. Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20.
[5] P. Muhly and M. Tomforde, Adding tails to C ∗-correspondences, Documenta Math. 9 (2004), 79 -- 106.
[6] I. Raeburn, Graph Algebras, CBMS Regional Conference Series in Math., vol. 103, American Math-
ematical Society, 2005.
[7] I. Raeburn and A. Sims, Product systems of graphs and the Toeplitz algebras of higher-rank graphs,
J. Operator Theory 53 (2005), 399 -- 429.
EFFICIENT PRESENTATIONS OF RELATIVE CUNTZ-KRIEGER ALGEBRAS
21
[8] I. Raeburn, A. Sims and T. Yeend, Higher-rank graphs and their C ∗-algebras, Proc. Edinb. Math.
Soc. 46 (2003), 99 -- 115.
[9] I. Raeburn, A. Sims, and T. Yeend, The C ∗-algebras of finitely aligned higher-rank graphs, J. Funct.
Anal. 213 (2004), 206 -- 240.
[10] A. Sims, Relative Cuntz-Krieger algebras of finitely aligned higher-rank graphs, Indiana Univ. Math.
J. 55 (2006), 849 -- 868.
[11] A. Sims, Gauge-invariant ideals in the C*-algebras of finitely aligned higher-rank graphs, Canad. J.
Math. 58 (2006), 1268 -- 1290.
[12] A. Sims, B. Whitehead, and M. F. Whittaker, Twisted C*-algebras associated to finitely aligned
higher-rank graphs, Documenta Math. 19 (2014), 831 -- 866.
Lisa Orloff Clark, School of Mathematics and Statistics, Victoria University of
Wellington, PO Box 600, Wellington 6140, New Zealand
Yosafat E.P. Pangalela, Pinnacle Investment, Wisma GKBI 38th Floor, Suite 3805, Jl
Jendral Sudirman No 28, Jakarta 10210, Indonesia
E-mail address: [email protected]
|
1109.5803 | 1 | 1109 | 2011-09-27T08:40:41 | Recovering the Elliott invariant from the Cuntz semigroup | [
"math.OA",
"math.RA"
] | Let $A$ be a simple, separable C$^*$-algebra of stable rank one. We prove that the Cuntz semigroup of $\CC(\T,A)$ is determined by its Murray-von Neumann semigroup of projections and a certain semigroup of lower semicontinuous functions (with values in the Cuntz semigroup of $A$). This result has two consequences. First, specializing to the case that $A$ is simple, finite, separable and $\mathcal Z$-stable, this yields a description of the Cuntz semigroup of $\CC(\T,A)$ in terms of the Elliott invariant of $A$. Second, suitably interpreted, it shows that the Elliott functor and the functor defined by the Cuntz semigroup of the tensor product with the algebra of continuous functions on the circle are naturally equivalent. | math.OA | math |
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ
SEMIGROUP
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
ABSTRACT. Let A be a simple, separable C∗-algebra of stable rank one. We prove
that the Cuntz semigroup of C(T, A) is determined by its Murray-von Neumann
semigroup of projections and a certain semigroup of lower semicontinuous functions
(with values in the Cuntz semigroup of A). This result has two consequences. First,
specializing to the case that A is simple, finite, separable and Z-stable, this yields
a description of the Cuntz semigroup of C(T, A) in terms of the Elliott invariant of
A. Second, suitably interpreted, it shows that the Elliott functor and the functor de-
fined by the Cuntz semigroup of the tensor product with the algebra of continuous
functions on the circle are naturally equivalent.
INTRODUCTION
The Cuntz semigroup Cu(A) of a C∗-algebra A is intimately related to the classifi-
cation program of simple, separable, and nuclear algebras. This is a semigroup built
out of equivalence classes of positive elements in the stabilization of the algebra A
much in an analogous way as the projection semigroup V (A) is, and comes equipped
with an order that is not algebraic, except for finite dimensional algebras. One order
property – almost unperforation – plays a significant role in classification of such al-
gebras up to isomorphism (see [19]). This property is equivalent to strict comparison,
which allows to determine the order in the semigroup by means of traces.
The Elliott conjecture predicts the existence of a K-theoretic functor Ell such that,
for unital, simple, separable, nuclear C∗-algebras A and B in a certain class, isomor-
phism between Ell(A) and Ell(B) can be lifted to a ∗-isomorphism of the algebras.
The concrete form of the invariant (known as the Elliott invariant) for which this
conjecture has had tremendous success is the following:
Ell(A) = ((K0(A), K0(A)+, [1A]), K1(A), T(A), r) ,
consisting of (ordered) topological K-Theory, the trace simplex, and the pairing be-
tween K-Theory and traces given by evaluating a trace at a projection (see, e.g. [7])
(The category where the said invariant sits will be described later.)
It is possible (and generally agreed) that the largest class for which classification
in its original form (i.e. using the Elliott invariant as above) may hold consists of
those algebras that absorb the Jiang-Su algebra Z tensorially.
Indeed, Z-stability
springs into prominence as a necessary condition for classification to hold (under the
assumption of weak unperforation on K0; see [9]). This property of being Z-stable
Date: May 23, 2018.
1
2
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
stands out as a regularity property for C∗-algebras, together with finite decomposi-
tion rank and the condition of strict comparison alluded to above. Among separable,
simple, nuclear C∗-algebras, a conjecture of Toms and Winter (see [21], and also [23])
asserts that these three conditions are equivalent.
The linkage between the Elliott invariant and the Cuntz semigroup has been ex-
plored in a number of papers (see, e.g. [3], [6], [14], [18]). One of the main results in
[3] recovers the Cuntz semigroup from the Elliott invariant in a functorial manner,
for the class of simple, unital, Z-stable algebras. Tikuisis shows, in [18], that the El-
liott invariant is equivalent to the invariant Cu(C(T,·)), for simple, unital, non-type
I ASH algebras with slow dimension growth (which happen to be Z-stable, as fol-
lows from results of Toms and Winter ([20], [22])). One of our main results in this
paper confirms that this equivalence can be extended to all simple, separable, finite
Z-stable algebras. Thus, from a functorial point of view and related to the Elliott
conjecture, we prove the following:
Theorem. Let A be a simple, unital, nuclear, finite C∗-algebra that absorbs Z tensorially.
Then:
(i) There is a functor which recovers the Elliott invariant Ell(A) from the Cuntz semigroup
Cu(C(T, A)).
(ii) Viewing the Elliott invariant as a functor from the category of C∗-algebras to the cate-
gory Cu (where the Cuntz semigroup naturally lives), there is a natural equivalence of
functors between Ell(·) and Cu(C(T,·)).
Since the Cuntz semigroup is a natural carrier of the ideal structure of the algebra,
it is plausible to expect that the object Cu(C(T,·)) may be helpful in the classification
of non-simple algebras.
The natural transformation that yields the equivalence of functors in the theorem
above is described in Section 4, and is based on describing the Cuntz semigroup
of C(T, A) for any simple, separable, unital C∗-algebra of stable rank one. This is
carried out in Sections 2 and 3, and is done in terms of the Murray-von Neumann
semigroup of projections of C(T, A) together with a the subsemigroup of the so-called
non-compact lower semicontinuous functions with values in Cu(A). Some of the
methods used are similar to the ones in [2].
1. NOTATION AND PRELIMINARIES
We briefly recall the construction of the Cuntz semigroup and the main technical
aspects that we shall be using throughout the paper. As a blanket assumption, A will
be a separable C∗-algebra.
Given positive elements a, b in A, we say that a is Cuntz subequivalent to b, in sym-
n → a in norm. The antisym-
bols a - b, if there is a sequence (xn) in A such that xnbx∗
metrization ∼ of the relation - is referred to as Cuntz equivalence.
The Cuntz semigroup of A is defined as
Denote the class of a positive element a by [a]. Then Cu(A) is ordered by [a] ≤ [b] if
a - b, and it becomes an abelian semigroup with addition given by [a] + [b] = [( a 0
0 b )].
Cu(A) = (A ⊗ K)+/ ∼ .
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ SEMIGROUP
3
As it was proved in [4], there exists a category of ordered semigroups, termed Cu,
with an enriched structure, such that the assignment A → Cu(A) defines a sequen-
tially continuous functor. We define this category below.
In an ordered semigroup S, we say that x is compactly contained in y if, whenever
there is an increasing sequence (zn) with y ≤ sup zn, there is m such that x ≤ zm.
This is denoted by x ≪ y (see [8]). If x ≪ x, we say that x is compact. An increasing
sequence (xn) is termed rapidly increasing provided that xn ≪ xn+1 for every n.
Define Cu to be the category whose objects are positively ordered (abelian) semi-
groups for which: (i) every increasing sequence has a supremum; (ii) every element
is a the supremum of a rapidly increasing sequence; and (iii) suprema and ≪ are
compatible with addition. Maps in Cu will be those semigroup maps that preserve
addition, order, suprema, and ≪. We shall be using repeatedly the fact that, for any
positive element a,
[a] = sup
n→∞
[(a − 1/n)+] ,
as follows from [16, Proposition 2.4] (see also [12]). We shall be frequently using that,
if a and b are positive elements and ka− bk < ǫ, then there is a contraction c in A such
that (a − ǫ)+ = cbc∗, so in particular (a − ǫ)+ - b (see [13, Lemma 2.2] and also [16,
Proposition 2.2]).
For a compact space X and a semigroup S in the category Cu, we shall use Lsc(X, S)
to denote the ordered semigroup of all lower semicontinuous functions from f : X →
S, with pointwise order and operation. (Here, f is lower semicontinuous if, for any
x ∈ S, the set {t ∈ X x ≪ f (t)} is open in X.)
space, then there is a natural map:
If A is a C∗-algebra and X is a one dimensional compact Hausdorff topological
α : Cu(C(X, A)) −→ Lsc(X, Cu(A))
x
7−→
x
where, if x = [f ], then x(t) = [f (t)]. It is proved in [2, Theorem 5.15] that Lsc(X, Cu(A))
equipped with the point-wise order and addition is a semigroup in Cu, and that α is
a well defined map in Cu which is an order embedding in case A has stable rank one
and K1(I) = 0 for all ideals of A. Furthermore, α is surjective provided it is an order
embedding (and thus an order isomorphism).
2. THE CUNTZ SEMIGROUP OF C([0, 1], A) FOR A SIMPLE ALGEBRA A
In this section, we prove that if A is a simple C∗-algebra with stable rank one,
then the Cuntz semigroup of C([0, 1], A) is order-isomorphic to Lsc([0, 1], Cu(A)), thus
obtaining the same result as in [2, Theorem 2.1] for a simple algebra, but without
requiring that K1(A) = 0. The key point in the argument is based on the fact that, for
certain continuous fields of C∗-algebras, unitaries from fibres can be lifted to unitaries
in the algebra.
Lemma 2.1. Let A be a unital continuous field of C∗-algebras over X = [0, 1] and let u, v ∈
U(A). If u(t0) ∼h v(t0) for some t0 ∈ (0, 1), then there exists w ∈ U(A) such that w(0) =
u(0) and w(1) = v(1).
4
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
Proof. Since u(t0) ∼h v(t0), we have (vu∗)(t0) ∼h 1A(t0). Therefore, there exists a
unitary w ∈ U0(A), such that w(t0) = vu∗(t0). Consider a continuous path ws of
unitaries in U0(A) such that w0 = 1A and wt0 = w. Let us define the following element
w ∈ Qt∈[0,1] A(t) given by
w(s) := (cid:26) ws(s)u(s)
v(s)
if s ≤ t0.
otherwise
Clearly w(0) = u(0) and w(1) = v(1). Since A is a continuous field of C∗-algebras,
to prove w ∈ A it is enough to find, for each t ∈ [0, 1] and ǫ > 0, a neighborhood Vt of
t and an element z ∈ A such that kw(s) − z(s)k < ǫ for all s ∈ V .
This is obvious if t ∈ (t0, 1]. If t ∈ [0, t0), and ǫ > 0, there exists a neighborhood
Vt such that for all s, s′ ∈ Vt, kws − ws′k < ǫ (since ws is a continuous path). Hence,
considering the element z = wtu ∈ A, we have, for all s ∈ Vt,
kw(s) − (wtu)(s)k = kws(s)u(s) − wt(s)u(s)k ≤ kws(s) − wt(s)k · ku(s)k < ǫ.
Now for t = t0, since k(wt0u)(t0) − v(t0)k = 0 and by the continuity of the norm
in A, there exists a neighborhood Vt0 such that k(wt0u)(s) − v(s)k < ǫ for all s ∈ Vt0,
and furthermore we can choose Vt0 such that kws − wt0k < ǫ. Now, with a similar
argument as above, we are done taking z = wt0u.
(cid:3)
Given a C ∗-algebra A and a hereditary subalgebra B ⊆ C(X, A), B becomes a con-
tinuous field of C∗-algebras over X whose fibres Bx can be identified with hereditary
subalgebras of A. If A is simple, then for all x ∈ X such that Bx 6= 0, the inclusion
ix : Bx → A induces an isomorphism
(ix)∗ : K1(Bx) → K1(A).
If A has stable rank one, then K1(A) = U(A∼)/U0(A∼) and elements can be iden-
tified with connected components of unitaries in U(A∼), which we denote by [v]A.
Hence, for all x, Bx will also have stable rank one and (ix)∗([v]Bx) = [i∼
x (v)]A = [v]A
where i∼
Let DB = B + C(X) · 1C(X,A∼) ⊆ C(X, A∼). Then DB is a unital continuous field
of C∗-algebras whose fibres DB(x) ∼= Bx + C · 1C(X,A∼)(x). Assuming A is stable, we
have 1A∼ = 1C(X,A∼)(x) 6∈ Bx ⊆ A, hence DB(x) ∼= B∼
x . Observe furthermore that the
following diagram commutes:
x → A∼ denotes tha natural extension to the unitizations.
x : B∼
DB = B + C(X) · 1C(X,A∼)
/ C(X, A) + C(X) · 1C(X,A∼)
πx
πx
DB(x) = Bx + C · 1C(X,A∼)(x)
i∼
x
/ A + C · 1C(X,A∼)(x) = A∼
Hence we will assume, DB(x) = B∼
x ⊆ A∼ = C(X, A∼)(x).
Proposition 2.2. Let A be a simple C∗-algebra with stable rank one and let X be a finite
graph. Suppose B is a hereditary subalgebra of C(X, A) such that Bx 6= 0 for all x ∈ X. Let
(ix)∗ denote the induced isomorphisms. Let x0, . . . , xn ∈ X and ui ∈ U(B∼
xi) for i = 0, . . . , n.
If (ixk)∗([uk]) = (ixl)∗([ul]) for all k, l, then there exists u ∈ U(DB) such that u(xi) = ui for
i = 0, . . . , n.
/
/
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ SEMIGROUP
5
Proof. Let us view X as a 1-dimensional simplicial complex where its 0-skeleton is
X0 = {x0, . . . , xn, . . . , xm} (possibly adding vertices to X0 as new points xi). To define
a unitary u ∈ U(DB), it is enough to define it first in X0 and then in each of the edges
of the 1-skeleton, provided the values in the boundary match the corresponding val-
ues in X0. Since X0 is a finite set of points, u can be easily defined pointwise (see
below) choosing, for all i = 1, . . . , n, u(xi) = ui. Therefore, in order to define u for
the 1-skeleton we can reduce to the case X = [0, 1] and x0 = 0, x1 = 1 with unitaries
u0, u1 such that [u1] = (ix1)−1
∗ (ix0)∗[u0].
Let us choose, for the remaining x ∈ (0, 1), unitaries ux ∈ U(B∼
x ) such that
[ux]Bx = (ix)−1
∗ (ix0)∗[u0]Bx0 ,
and hence such that [ix0(u0)]A = [ix(ux)]A which means that ux and u0 are connected
in U(A∼).
For each x ∈ X we can find an open neighborhood Vx such that ux = vx(x) for some
vx ∈ DB and vxVx is a unitary. Since X is compact, we can find a finite number of
such neighborhoods Vx0 := V0, . . . , Vr := Vx1 covering X. Furthermore by restricting
the Vx's to be open intervals we can assume that the resulting cover has multiplicity
1 and denote Vi∩ Vi+1 = (ai, bi) for i = 1, . . . , r− 1 (ai < bi < ai+1). For i = 1, . . . , r− 1,
let us assume Vi = Vyi for some yi ∈ X, y0 = x0 = 0 and yr = x1 = 1.
For each i = 0, . . . , r − 1 choose zi ∈ (ai, bi). Since DB ⊆ C(X, A∼), both vyi Vi and
vyi+1 Vi+1 are paths of unitaries in A∼. Hence in A∼ we have
vyi(zi) ∼h vyi(yi) = uyi ∼h u0 ∼h uyi+1 = vyi+1(yi+1) ∼h vyi+1(zi).
This implies (izi)∗[vyi(zi)] = (izi)∗[vyi+1(zi)], but since (izi)∗ is an isomorphism, we
obtain vyi+1(zi) ∼h vyi(zi) in B(zi)∼. Now, using Lemma 2.1, we can construct a
unitary wi in DB(Vi ∩ Vi+1)∼ = DB([ai, bi])∼ such that wi(ai) = vyi(ai) and wi(bi) =
vyi+1(bi).
Therefore, defining v ∈ DB as the following element in Qx∈X B∼
x
v(x) = (cid:26) vyi(x)
wi(x)
if x ∈ Vi \ (Vi−1 ∪ Vi+1)
if x ∈ Vi ∩ Vi+1
we obtain an element in DB, which is furthermore a unitary and v(0) = u0, v(1) = u1.
(cid:3)
Remark 2.3. Observe that, in the particular case of only one point x0 ∈ X, the Proposition
states that the map U(DB) → U(B∼
x0) is surjective, and thus we can lift unitaries from each
fibre.
Let X be a locally compact, Hausdorff space. Suppose A is a continuous field of
C∗-algebras over X and a ∈ A. We denote by supp(a) = {x ∈ X a(x) 6= 0}. Observe
that, since the assignment x → ka(x)k is continuous, supp(a) is an open subset of X.
If Y ⊆ X is a closed subset of X, and a ∈ A, then aY denotes the image of a by the
projection πY : A → A(Y ).
Lemma 2.4. Let A be a continuous field of C∗-algebras over a space X and let a, b ∈ A+.
i=1Xi is a finite disjoint union of open sets, then a - b if and only if
(1) If X = ⊔r
aXi - bXi for i = 1, . . . , r.
6
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
(2) If bK - aK for some K such that supp(b) ⊆ K ⊆ supp(a), then a - b.
Proof. (i) Is clear since A ∼= ⊕r
i=1A(Xi). Let us prove (ii). Suppose bK - aK as in the
statement. Given ǫ > 0, we can find d ∈ A such that kb(x) − d(x)a(x)d∗(x)k < ǫ for
all x ∈ K, but since A is a continuous field of C∗-algebras, this is valid in an open set
K ⊆ U,
(1)
b(x) − dad∗(x) < ǫ
for all x ∈ U.
Now, since K ∩ U c = ∅ we can consider a continuous function λ : X → [0, 1] such
that λK = 1 and λU c = 0. If x ∈ supp(b) ⊆ K ⊆ U, then
kb(x) − (λd)a(λd)∗(x))k = kb(x) − d(x)a(x)d∗(x)k < ǫ,
by (1), and if x 6∈ U then b(x) − (λd)a(λd)∗(x) = 0. Finally, if x ∈ U \ supp(b), then
b(x) = 0, and
b(x) − (λd)a(λd)∗(x) = λ2b(x) − (λd)a(λd)∗(x) = λ2(x) · kb(x) − dad∗(x)k < ǫ,
again by (1). Hence, since b − (λd)a(λd)∗ = supx∈X b(x) − (λd)a(λd)∗(x) < ǫ, we
obtain b - a.
(cid:3)
Recall that if X is a locally compact Hausdorff topological space, then the set O(X)
consisting of open sets ordered by inclusion is a continuous lattice. In the case X
is second countable, we have that U ≪ V whenever there exists a compact set K
such that U ⊆ K ⊆ V (the countability condition is needed since our definition of
compact containment is only for increasing sequences, and not arbitrary nets). In
fact, O(X) with union as addition is a semigroup in Cu, which can be described as
Lsc(X,{0,∞}) through the assignment f → supp(f ) (since ∞ is a compact element
in {0,∞} and thus supp(f ) = f −1({∞}) is an open set, and, by the same argument,
the characteristic function (in {0,∞}) of any open set, is lower semicontinuous). The
following Lemma illustrates the relation of Cuntz order in a continuous field of C∗-
algebras over X with the ordered structure of O(X).
Lemma 2.5. Let A be a (stable) continuous field of C∗-algebras over a compact Hausdorff
space X and a, b ∈ A+ such that [b] ≤ [a]. Then supp(b) ⊆ supp(a) and, if [b] ≪ [a] we
have supp(b) ≪ supp(a).
Proof. The first statement is obvious, let us suppose [b] ≪ [a] for some a, b ∈ A+.
Since O(X) is in Cu, let us write supp(a) = ∪i≥0Ui for some Ui ≪ Ui+1 (hence Ui ⊆
Ui ⊆ Ui+1). We can find, by Urysohn's Lemma, continuous functions λn : X → [0, 1]
such that λn(Un) = 1 and λn(U c
n+1) = 0. Since X is compact we obtain λna → a
and λna ≤ λn+1a, thus [a] = supn[λna]. Now since [b] ≪ [a] we get [b] ≤ [λN a] for
some N > 0 and therefore supp(b) ⊆ supp(λN a) ⊆ UN +1 ⊆ UN +1 ⊆ supp(a). Hence
supp(b) ≪ supp(a).
(cid:3)
Theorem 2.6. Let A be a C∗-algebra which is separable, simple and has stable rank one.
Then, the map α : Cu(C([0, 1], A)) → Lsc([0, 1], Cu(A)) is an order isomorphism.
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ SEMIGROUP
7
Proof. We may assume A is stable. Suppose f, g ∈ C([0, 1], A) are such that f (t) - g(t).
It is enough to prove that (f−δ)+ - g for all δ > 0, so let us first assume that [f ] ≪ [g].
Hence, by Lemma 2.5 we have supp(f ) ≪ supp(g), and thus there exists a compact
set K such that supp(f ) ⊆ K ⊆ supp(g). Since finite unions of open intervals form a
dense subset of O([0, 1]), and K is compact, we may further assume that K is a finite
union of closed intervals. Now by virtue of Lemma 2.4 (i) and (ii), we may finally
assume that supp(g) = [0, 1].
Now the proof follows the lines of [2, Theorem 2.1]. In there, K1(A) = 0 was as-
sumed in order to lift unitaries from Her(g(t))∼ to DHer(g) = Her(g) + C(X)· 1M (C(X,A)).
From our argument in the previous paragraph we can reduce to the case where
Her(g(t))∼ 6= 0 for all t and then use Proposition 2.2 with B := Her(g) (see also
(cid:3)
Remark 2.3).
3. THE INVARIANT CuT(A)
In this section we give a complete description of the Cuntz semigroup of C(T, A),
for a simple, separable C∗-algebra A that has stable rank one. We also show that, in
the simple, Z-stable, finite case, the information it contains is equivalent to that of
the Elliott invariant (see next section and also [18]).
We start with the following:
Lemma 3.1. Let A be a simple C∗-algebra with stable rank one, and let y ∈ A be a contrac-
tion. Let ǫ > 0 be such that ǫ ∈ σ(yy∗), and let B be a hereditary subalgebra of A with
yy∗ ∈ B. If u, v are unitaries in B∼, then there is u0 ∈ U(B∼) with [u0] = [v] in K1(B),
and
kuy − u0yk < 5√ǫ .
Proof. We know that ǫ ∈ σ(yy∗), so there exists 0 < c with kck < 2ǫ and such that
(yy∗ − 2ǫ)+ ⊥ c , and (yy∗ − 2ǫ)+ + c ≤ yy∗ .
∼
Note that c ∈ B. Write d = (yy∗ − 2ǫ)+. As c 6= 0, inclusion induces an isomorphism
K1(cAc) ∼= K1(B), so there is w ∈ U(cAc
) of the form 1 + a, where a ∈ cAc such that
[w] 7→ [v] − [u].
Put u0 = uw, a unitary in B∼. Note that, in K1(B), we have [u0] = [u] + [w] =
[u] + [v] − [u] = [v].
ǫ/2, so kczck < ǫ/2 + 2. Compute that
Next, choose czc ∈ cAc such that ka−czck < ǫ/2. Then, kw−(1+czc)k = ka−czck <
u(1 + czc)(d + c) = u(d + c + czc2) = ud + uc + uczc2 ,
whence
ku(1 + czc)(d + c) − u(d + c)k = kuczc2k < (2 + ǫ/2)2ǫ = 4ǫ + ǫ2 .
8
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
Therefore
kuyy∗ − u0yy∗k ≤kuyy∗ − u(d + c)k + ku(d + c) − u0(d + c)k + ku0(d + c) − u0yy∗k
<4ǫ + ku(d + c) − u0(d + c)k
≤4ǫ + ku(d + c) − u(1 + czc)(d + c)k + ku(1 + czc)(d + c) − uw(d + c)k
<8ǫ + ǫ2 + ku(1 + czc − w)(d + c)k < 8ǫ + ǫ2 + ǫ/2 .
Thus
kuy − u0yk2 = k(u − u0)yy∗(u − u0)∗k < 2(8ǫ + ǫ2 + ǫ/2) < 19ǫ ,
so that kuy − u0yk < √19ǫ < 5√ǫ.
(cid:3)
Proposition 3.2. Let A be a simple, separable C ∗-algebra of stable rank one. Let f and g
be elements in C(T, A) such that f is not equivalent to a projection, and g is never zero. If
f (t) - g(t) for all t ∈ T, then f - g.
Proof. Since f is not equivalent to a projection, zero is an isolated point of σ(f ) and
this implies, as σ(f ) = ∪t∈Tσ(f (t)), that for every n, there is tn ∈ T and λn ∈ σ(f (tn))
with 0 < λn < 1/2n. By compactness, and passing to a subsequence if necessary, we
may assume that (tn) converges to a point t0. We shall assume that the sequence (tn)
is not eventually constant, since otherwise (that is, f (t0) itself is not equivalent to a
projection), the argument is similar, and easier.
Let ϕ : [0, 1] → T be the map that ϕ(0) = ϕ(1) = t0. Since f (t) - g(t) for all t, this
Let 0 < ǫ < 1. There exists d ∈ A such that k(f ◦ ϕ)(0) − d∗(g ◦ ϕ)(0)dk < ǫ. There is
also holds when composing with ϕ, so (f ◦ ϕ)(s) - (g ◦ ϕ)(s) for all s ∈ [0, 1].
then a neighbourhood U of 0 and 1 such that, with h(s) = d, we have
k(f ◦ ϕ − h∗(g ◦ ϕ)h)Uk < ǫ .
0, 1], with s0 < s′
0. Now, there exists ǫ′ < ǫ2, s1 ∈ U and
Write U = [0, s0) ∪ (s′
λϕ(s1) ∈ σ(f (ϕ(s1))) such that ǫ′ < λϕ(s1) < ǫ2, and we may assume (without loss of
generality) that 0 < s1 < s0. Choose also s′
0 < s2 < 1.
By Theorem 2.6, there exists c ∈ C([0, 1], A) such that kf ◦ ϕ− c∗(g ◦ ϕ)ck < ǫ′/2. By
[13, Lemma 2.2], there is a contraction e ∈ C([0, 1], A) such that, with y1 = (g◦ ϕ)1/2ce,
we have ((f ◦ ϕ) − ǫ′/2)+ = y∗
1y1. If we let y2 = (g ◦ ϕ)1/2h, we have
kf ◦ ϕ − y∗
1y1k ≤ ǫ′/2 < ǫ′ , kf ◦ ϕ − y∗
By evaluating at the si, for i = 1, 2, we get
2y2k < ǫ and yiy∗
i ∈ Her(g ◦ ϕ) for i = 1, 2 .
k(f ◦ ϕ)(si) − y∗
1y1(si)k < ǫ′ and k(f ◦ ϕ)(si) − y∗
2y2(si)k < ǫ ,
so we may apply [2, Lemma 1.4] to find unitaries
u′
1 ∈ Her((g ◦ ϕ)(s1))∼ and u2 ∈ Her((g ◦ ϕ)(s2))∼
such that
ku′
1y1(s1) − y2(s1)k < 9ǫ and ku2y1(s2) − y2(s2)k < 9ǫ .
Let u′′
1 be a unitary such that
[u′′
1] = (is1)−1
∗ ◦ (is2)∗([u2]) .
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ SEMIGROUP
9
Since λϕ(s1) ∈ σ((f ◦ ϕ)(s1)), we have that 0 < λϕ(s1) − ǫ′/2 ∈ σ(((f ◦ ϕ)(s1)− ǫ′/2)+) =
σ(y∗
1(s1)). By Lemma 3.1, there is a unitary u1 ∈
Her((g ◦ ϕ)(s1))∼ such that
1y1(s1)), so λϕ(s1) − ǫ′/2 ∈ σ(y1y∗
[u1] = [u′′
1] in K1(Her((g ◦ ϕ)(s1)))
1y1(s1) − u1y1(s1)k < 5qλϕ(s1) − ǫ′/2 < 5ǫ .
ku′
w(ϕ(s2)) = u2.
Thus
ku1y1(s1) − y2(s1)k ≤ ku1y1(s1) − u′
1y1(s1) − y2(s1)k < 5ǫ + 9ǫ = 14ǫ .
By Proposition 2.2 there is a unitary w ∈ DHer(g) such that w(ϕ(s1)) = u1 and
1y1k < ǫ′ < ǫ, and
Put y′
also that
ky′
1(s1) − y2(s1)k = k(w(ϕ(s1))y1(s1) − y2(s1)k = ku1y1(s1) − y2(s1)k < 14ǫ
1 = (w ◦ ϕ)y1, and notice that kf ◦ ϕ− (y′
1)k = kf ◦ ϕ− y∗
1y1(s1)k + ku′
1)∗(y′
and
and
ky′
1(s2) − y2(s2)k = kw(ϕ(s2))y1(s2) − y2(s2)k = ku2y1(s2) − y2(s2)k < 9ǫ .
Therefore, there exists a neighbourhood W ⊂ U of s1 and s2, that neither contains 0
nor 1, with
ky′
1(s) − y2(s)k < 14ǫ for all s ∈ W .
Let V = [0, s1)∪ (s2, 1], and let µ1, µ2 be a partition of unity associated to the covering
V ∪ W , V c ∪ W , and consider the element
z = µ1y′
1 + µ2y2 .
We need to estimate kf − z∗zk. It is enough to consider (f − z∗z)W . Since (y′
Note that z(0) = y2(0) = y2(1) = z(1), so z ∈ C(T, A). Also zz∗ ∈ Her(g).
z)W = (y′
1−z)Wk ≤ k(y′
14ǫ. Therefore, a standard argument shows that
1−y2)W , we see that k(y′
1−µ2y2)W = µ2(y′
1 −
1−y2)Wk <
1−µ1y′
k((y′
1)∗y′
1 − z∗z)Wk < 28ǫ√1 + ǫ < 42ǫ ,
whence
This implies that (f − 43ǫ)+ - g, and since ǫ > 0 is arbitrary, it follows that f - g in
C(T, A), as desired.
(cid:3)
k(f − z∗z)Wk < ǫ + 42ǫ = 43ǫ .
Remark 3.3. In view of the previous result, the reader may wonder whether if an
element f ∈ C(X, A) is not equivalent to a projection, then there is some point x ∈ X
such that f (x) is itself not equivalent to a projection. We remark this is not true, as is
seen by taking, e.g. X = [0, 1], p any non-zero projection in A, λ(t) = (1/2 − t)+, and
f = λp. Then, clearly f is equivalent to a projection pointwise, but not globally.
Proposition 3.4. Let A be a simple, separable C ∗-algebra of stable rank one. Let f and g be
elements in C(T, A) such that f is not equivalent to a projection. If f (t) - g(t) for all t ∈ T,
then f - g.
10
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
Proof. If g is never zero, then the result follows from Proposition 3.2. We may there-
fore assume that, without loss of generality, g(1) = 0 (and then also f (1) = 0).
kf ◦ ϕ − c(g ◦ ϕ)c∗k < ǫ/2 .
Let ϕ : [0, 1] → T be the map such that ϕ(0) = ϕ(1) = 1. Since f ◦ ϕ(s) - g ◦ ϕ(s)
for every s ∈ [0, 1], it follows from Theorem 2.6 that (f ◦ ϕ) - (g ◦ ϕ). Let ǫ > 0. Find
c ∈ C([0, 1], A) such that
Since f (1) = g(1) = 0, there is a neighbourhood U of 0 and 1 such that k(f ◦ ϕ)Uk < ǫ
and k(g ◦ ϕ)Uk < ǫ/(2kck2). Let λ : [0, 1] → C be a continuous function such that
0 ≤ λ ≤ 1, λU c = 1, and λ(0) = λ(1) = 0, and let d = λ1/2c, which defines an element
in C(T, A). Then (f ◦ ϕ − λc(g ◦ ϕ)c∗)U c = (f ◦ ϕ − c(g ◦ ϕ)c∗)U c, and
k(f ◦ ϕ− λc(g◦ ϕ)c∗)Uk ≤ k(f ◦ ϕ− c(g◦ ϕ)c∗)Uk +k(1− λ)Ukkgkkck2 < ǫ/2 + ǫ/2 = ǫ ,
whence kf − dgd∗k < ǫ so (f − ǫ)+ - g. Since ǫ > 0 is arbitrary, this implies that f - g,
(cid:3)
as was to be shown.
We are now ready to describe the Cuntz semigroup of C(T, A), whenever A is
simple and has stable rank one. As A is, in particular, stably finite, this is also the
case for C(T, A). Thus, upon identification of V(C(T, A)) with its image in CuT(A),
we have
where CuT(A)nc stands for the subsemigroup of non-compact elements.
CuT(A) = V (C(T, A)) ⊔ CuT(A)nc ,
Observe that CuT(A) → Lsc(T, Cu(A)) sends compact elements to compact ele-
ments. Using the arguments in [2, Corollary 3.8], those are the functions that take a
constant value in V(A).
If X is a compact Hausdorff, connected space, and S is a semigroup in the category
Cu, let us denote by Lscnc(X, S) the set of non-compact elements in Lsc(X, S).
Remark 3.5. Observe that, if X is a connected, compact Hausdorff space, A is a (sta-
ble) C∗-algebra and f ∼ p, for f ∈ C(X, A)+ and p a projection in C(X, A), then f is
pointwise equivalent to a projection q ∈ A. This is easy to verify by a direct argu-
ment, but can also be obtained as a consequence of the fact that, for a semigroup S
in Cu, the compact elements in Lsc(X, S) are precisely the constant, compact-valued
functions (see, e.g. the arguments in [2, Corollary 3.8]). In particular, such an f is
either identically zero or always non-zero.
Lemma 3.6. Let X be compact Hausdorff and connected, and let S be a semigroup in Cu
with cancellation of compact elements and such that the set of non-compact elements is closed
under addition. Then Lscnc(X, S) is a subsemigroup of Lsc(X, S).
Proof. Let f, g ∈ Lscnc(X, S) and assume that f + g is compact. The arguments in [2,
Corollary 3.8] show that there is a compact element c ∈ S such that (f + g)(t) = c for
every t ∈ X. By our assumptions on S, it follows that f (t) and g(t) are compact for
every t ∈ X.
Using that f (t) ≪ f (t) and g(t) ≪ g(t), and that f and g are lower semicontinuous,
find a neighbourhood Ut of t such that f (t) ≪ f (s) and g(t) ≪ g(s) for every s ∈ Ut
(see, e.g. [2, Lemma 5.1]). It then follows that
f (t) + g(s) ≪ f (s) + g(s) = c = f (t) + g(t) .
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ SEMIGROUP
11
By cancellation of compact elements, g(s) ≤ g(t) ≤ g(s) in Ut, so that g is constant in
a neighbourhood of t. Since X is connected, it follows that g is constant. Likewise, f
(cid:3)
is constant.
When S as above comes as a Cuntz semigroup of a C∗-algebra, then it satisfies the
additional axiom of having an "almost algebraic order" (see [17, Lemma 7.1 (i)], and
also [15]): if x ≤ y and x′ ≪ x, then there is z ∈ S such that x′ +z ≤ y ≤ x+z. One can
then prove that, if such an S has moreover cancellation of compact elements, then the
set Snc of non-compact elements is a subsemigroup of S. Indeed, if x + y is compact,
choose x′ ≪ x′′ ≪ x such that x′ + y = x′′ + y = x + y. By the almost algebraic order
axiom, there is z ∈ S with x′ + z ≤ x ≤ x′′ + z. Adding y to this inequality yields
(x + y) + z ≤ x + y, and since x + y is compact, it follows that z = 0, and this implies
that x ≤ x′′ ≪ x.
For a simple, separable C∗-algebra with stable rank one, consider the semigroup
equipped with addition that extends both the natural operations in both components,
and with
V(C(T, A)) ⊔ Lscnc(T, Cu(A)) ,
x + f = x + f , whenever x ∈ V(C(T, A)) and f ∈ Lscnc(T, Cu(A)) .
We can order this semigroup by taking the algebraic ordering in V(C(T, A)), the
pointwise ordering on Lscnc(T, Cu(A)), and we order mixed terms as follows:
(i) f ≤ x if f (t) ≤ x(t) for every t ∈ T.
(ii) x ≤ f if there is g ∈ Lscnc(T, Cu(A)) such that x + g = f .
That this ordering is transitive is not entirely trivial, but it follows from the argu-
ments in Theorem 3.7 below.
We may now define an ordered map in the category of semigroups:
α : CuT(A) −→ V(C(T, A)) ⊔ Lscnc(T, Cu(A))
7−→ (cid:26) x if x ∈ V(C(T, A))
otherwise
x
x
Theorem 3.7. If A is a simple C∗-algebra with stable rank one, then there is an order-
isomorphism
CuT(A) ∼= V(C(T, A)) ⊔ Lscnc(T, Cu(A)) .
Proof. We will show that the map α just defined is a surjective order-embedding.
First note that C(T, A) is the following pullback
C(T, A)
ev1
/ A
C([0, 1], A)
ev0,1
/ A ⊕ A
Since, by Theorem 2.6, the natural map Cu(C([0, 1], A) → Lsc([0, 1], Cu(A)) is an
order-embedding, we may use [2, Theorem 3.3] to conclude that the pullback map
CuT(A) → Cu(C([0, 1], A)) ⊕Cu(A⊕A) Cu(A)
/
/
12
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
is a surjective map in the category Cu. Upon identifying Cu(C([0, 1], A)) ⊕Cu(A⊕A)
Cu(A) with Lsc(T, Cu(A)), we obtain that the map
CuT(A) → Lsc(T, Cu(A)), given by x 7→ x ,
is also surjective. This implies in particular that the map α is surjective.
To prove that α is an order-embedding, let x, y ∈ CuT(A) and assume that α(x) ≤
α(y). There is nothing to prove if x, y ∈ V(C(T, A)).
If x /∈ V(C(T, A)), then write x = [f ], y = [g], and our assumption just means that
f (t) - g(t) for every t ∈ T. We may then apply Proposition 3.4 to conclude that
f - g.
Finally, assume that x ∈ V(C(T, A)) and y /∈ V(C(T, A)). Then α(x) ≤ α(y) means,
by definition, that there is g ∈ Lscnc(T, Cu(A)) with x + g = y. Let z ∈ CuT(A) be such
that z = g. Then
(x + z) = x + g = y .
Note that x + z /∈ V(C(T, A)), as otherwise (x + z) would be a compact element in
Lsc(T, Cu(A)). By Lemma 3.6 (or rather, its proof – see also Remark 3.5), g = z would
be constant (and compact), a contradiction.
The argument in the previous paragraph then shows that x + z = y, as wanted. (cid:3)
Theorem 3.8. Let A be a separable, finite Z-stable C∗-algebra. Then, there is an order-
isomorphism
CuT(A) ∼= ({0} ⊔ (V(A)∗ × K1(A))) ⊔ Lscnc(T, Cu(A)) ,
where V(A)∗ = V(A) \ {0}.
Proof. By Theorem 3.7, we only need to show that V(C(T, A)) ∼= {0} ⊔ (V(A)∗ ×
K1(A)). This follows once we notice that C(T, A)) has cancellation of projections
(see, e.g. [18]). Since A is Z-stable, then C(T, A) is also Z-stable, whence C(T, A)
has cancellation of full projections by [11, Theorem 1]. We have already observed
(see Remark 3.5) that every projection in (matrices over) C(T, A) is either identically
zero or always non-zero, and in that case it is a full projection as A is simple, by an
(cid:3)
application of [5, Lemma 10.4.2].
Remark 3.9. In light of these results, one might expect that the same description of
the Cuntz semigroup will hold for more general spaces (of dimension at most 1).
However, the following example provided by N. C. Phillips shows that this is not
the case.
Let A be a simple C∗-algebra with stable rank one, K1(A) 6= 0 and such that
V(C(T, A)) ∼= {0} ⊔ V(A)∗ × K1(A) (for example, A could be Z-stable as above).
Let X = T ∪ [1, 2], and take f ′, g′ ∈ C(T, A) be elements such that f ′(t) ∼ g′(t)
for all t ∈ T, yet f ′ and g′ are not comparable. For example, we could take a non-
zero element [p] ∈ V(A)∗, a non-trivial class [u] ∈ K1(A), and f ′ corresponding to
([p], [1]) and g′ corresponding to ([p], [u]). Define f, g ∈ C(X, A) as f ′, g′ over T, and
f (t) = (2 − t)f (1), g(t) = g(1) for t ∈ [1, 2]. Then clearly f (t) - g(t) for all t ∈ X, but
f 6- g.
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ SEMIGROUP
13
4. A CATEGORICAL APPROACH
As already shown in [18], the Elliott invariant and the invariant defined by CuT(−)
are equivalent in a functorial way, for simple, unital non-type I ASH algebras with
slow dimension growth. Because of Theorem 3.8, this is actually true in the more gen-
eral setting of separable Z-stable, simple C∗-algebras with stable rank one. Our aim
in this section is to develop a (somewhat) abstract approach that makes the functorial
equivalence explicit, thus also proving the Theorem announced in the Introduction.
Let S be a semigroup in Cu. Assume that the subset Snc of non-compact elements
is an absorbing subsemigroup, in the sense that Snc + S ⊆ Snc. Denote by Sc the
subsemigroup of compact elements and S ∗
c = Sc \{0}. Let G be an abelian group and
consider the semigroup
SG = ({0} ⊔ (G × S ∗
c )) ⊔ Snc ,
with natural operations in both components, and (g, x) + y = x + y whenever x ∈ S ∗
c ,
y ∈ Snc, and g ∈ G. This semigroup can be ordered by
(i) For x, y ∈ S ∗
c , and g, h ∈ G, (g, x) ≤ (h, y) if and only if x = y and g = h, or else
(ii) For x ∈ S ∗
c , y ∈ Snc, g ∈ G, (g, x) is comparable with y if x is comparable with y.
The proof of the following lemma is rather straightforward, hence we omit the de-
tails.
x < y.
Lemma 4.1. Let S be an object of Cu such that Snc is an absorbing subsemigroup. If G is an
abelian group, then SG is also an object of Cu.
As in [14], let us write I to denote the category whose objects are 4-tuples
I = ((G0, G+
0 , u), G1, X, r) ,
Maps between objects ((G0, G+
where (G0, G+
0 , u) is a (countable) simple partially ordered abelian group with order-
unit u, G1 is a (countable) abelian group, X is a (metrizable) Choquet simplex, and
r : X → S(G0, u) is an affine map, where S(G0, u) denotes the state space of (G0, u).
0 , v), H1, Y, s) of I are de-
scribed as 3-tuples (θ0, θ1, γ), where θ0 is a morphism of ordered groups with order
unit, θ1 is a morphism of abelian groups, and γ : Y → X is an affine and continuous
map such that r ◦ γ = θ∗
0 : S(H0, v) → S(G0, u) is the naturally induced
map.
Let Cs denote the class of simple, unital, separable and nuclear C∗-algebras. Then,
0 , u), G1, X, r) and ((H0, H +
0 ◦ s, where θ∗
the Elliott invariant defines a functor
Ell : Cs → I
by
Ell(A) = ((K0(A), K0(A)+, [1A]), K1(A), T(A), r) ,
where T(A) is the trace simplex and r is the pairing between K-Theory and traces.
Let us define a functor
F : I → Cu
14
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
as follows. If I = ((G0, G+
0 , u), G1, X, r) is an object of I, set
F (I) = ({0} ⊔ (G1 × G++
0 \ {0}.
0
)) ⊔ Lscnc(T, G+
0 ⊔ LAff(X)++) ,
where G++
Since G+
0 = G+
0 ⊔ LAff(X)++ is an object of Cu (see, e.g. [1, Lemma 6.3]), it follows from
0 ⊔ LAff(X)++
That F is a functor follows almost by definition. The only non-trivial detail that
Lemma 4.1 above that F (I) is also an object of Cu. (The addition on G+
is given by (g + f )(x) = r(x)(g) + f (x), where g ∈ G, f ∈ LAff(X) and x ∈ X.)
needs to be checked is that, if
(θ0, θ1, γ) : ((G0, G+
is a morphism in I and f : T → G+
H +
0 ⊔ LAff(Y )++ is also non-compact. Here
0 , u), G1, X, r) → ((H0, H +
0 ⊔LAff(X)++ is non-compact, then (θ0⊔γ∗)◦f : T →
0 , v), H1, Y, s)
is defined as θ0 on G+
θ0 ⊔ γ∗ : G+
0 and γ∗ on LAff(X)++.
0 ⊔ LAff(X)++ → H +
0 ⊔ LAff(Y )++
If (θ0 ⊔ γ∗) ◦ f is compact, then there is h ∈ H +
0 such that θ0(f (T)) = {h} and
f (T) ⊆ G+
0 . As f is non-compact and lower semicontinuous, there are s, t ∈ T with
f (t) < f (s), whence f (s) − f (t) ∈ G++
is an order-unit. Thus, there exists n ∈ N with
f (s) ≤ n(f (s) − f (t)). After applying θ0, we obtain that h ≤ 0, so that h = 0. But this
is not possible since, as f is not constant, it takes some non-zero value a, which will
be an order-unit with θ0(a) = 0, contradicting that θ0(u) = v.
0
Let us show that F : I → F (I) is a full, faithful and dense functor, so it yields
an equivalence of categories. Therefore, by standard category theory, there exists
a functor G : F (I) → I such that F ◦ G and G ◦ F are naturally equivalent to the
(respective) identities.
We only need to prove that F is a faithful functor. If
(θ0, θ1, γ) : ((G0, G+
0 , u), G1, X, r) → ((H0, H +
0 , v), H1, Y, s)
is a morphism in I, we shall write F ((θ0, θ1, γ)) = (θ1 × θ0) ⊔ (θ0 ⊔ (γ∗)∗), where
(θ0 ⊔ (γ∗))∗(f ) = (θ0 ⊔ (γ∗)) ◦ f ,
0 ⊔LAff(X)++). If now F ((θ0, θ1, γ)) = F ((θ′
1, whence θi = θ′
for f ∈ Lscnc(T, G+
1, γ′)), we readily see
i. It also follows that γ∗(h) = h◦γ = h◦γ′ = γ′∗(h),
that θ0×θ1 = θ′
0×θ′
for every affine continuous function h on X. Since X is homeomorphic to the state
space on Aff(X) (normalized at the constant function 1) via the natural evaluation
map ψ : X → S(Aff(X), 1) (e.g. [10, Theorem 7.1]), the compositions
0, θ′
∼= /
Y
/ S(Aff(Y ), 1)
γ ∗
γ ′∗
/ S(Aff(X), 1)
∼= /
/ X
yield that γ = γ′.
Assembling our observations above (together with Theorem 3.8 and [3, Corollary
5.7]), we get the following:
/
/
/
RECOVERING THE ELLIOTT INVARIANT FROM THE CUNTZ SEMIGROUP
15
Theorem 4.2. (Cf. [18]) Upon restriction to the class of unital, simple, separable and finite
Z-stable algebras, there are natural equivalences of functors
Therefore, for these algebras, Ell is a classifying functor if, and only if, so is CuT.
F ◦ Ell ≃ CuT and Ell ≃ G ◦ CuT .
ACKNOWLEDGEMENTS
This work was carried out at the Centre de Recerca Matem`atica (Bellaterra) during
the programme "The Cuntz Semigroup and the Classification of C∗-algebras" in 2011.
We gratefully acknowledge the support and hospitality extended to us. It is also a
pleasure to thank N. Brown, I. Hirshberg, N. C. Phillips and H. Thiel for interesting
discussions concerning the subject matter of this paper. The first, third and fourth
authors were partially supported by a MEC-DGESIC grant (Spain) through Project
MTM2008-06201-C02-01/MTM, and by the Comissionat per Universitats i Recerca
de la Generalitat de Catalunya. The second author was partially supported by NSF
grant #DMS–1101305.
REFERENCES
[1] R. Antoine, J. Bosa, and F. Perera. Completions of monoids with applications to the Cuntz semi-
group. Int. J. Math., 22(5):1–25, 2011.
[2] R. Antoine, F. Perera, and L. Santiago. Pullbacks, C(X)-algebras, and their Cuntz semigroup. J.
Funct. Anal., 260(10):2844–2880, 2011.
[3] N. P. Brown, F. Perera, and A. S. Toms. The Cuntz semigroup, the Elliott conjecture, and dimen-
sion functions on C ∗-algebras. J. Reine Angew. Math., 621:191–211, 2008.
[4] K. T. Coward, G. A. Elliott, and C. Ivanescu. The Cuntz semigroup as an invariant for C ∗-
algebras. J. Reine Angew. Math., 623:161–193, 2008.
[5] J. Dixmier, C∗-algebras, volume 15 of North-Holland Mathematical Library. North-Holland Publish-
ing Co., Amsterdam-New York-Oxford, 1977.
[6] G. A. Elliott, L. Robert, and L. Santiago. The cone of lower semicontinuous traces on a C ∗-algebra.
Amer. J. Math., 133(4):969–1005, 2011.
[7] G. A. Elliott and A. S. Toms. Regularity properties in the classification program for separable
amenable C ∗-algebras. Bull. Amer. Math. Soc. (N.S.), 45(2):229–245, 2008.
[8] G. Gierz, K. H. Hofmann, K. Keimel, J. D. Lawson, M. Mislove, and D. S. Scott. Continuous lattices
and domains, volume 93 of Encyclopedia of Mathematics and its Applications. Cambridge University
Press, Cambridge, 2003.
[9] G. Gong, X. Jiang, and H. Su. Obstructions to Z-stability for unital simple C ∗-algebras. Canad.
[10] K. R. Goodearl. Partially ordered abelian groups with interpolation, volume 20 of Mathematical Surveys
Math. Bull., 43(4):418–426, 2000.
and Monographs. American Mathematical Society, Providence, RI, 1986.
[11] X. Jiang. Non-stable K-theory for Z-stable C∗-algebras. arXiv:math/9707228v1 [math.OA], pages
1–18, 1997.
[12] E. Kirchberg and M. Rørdam. Non-simple purely infinite C ∗-algebras. Amer. J. Math., 122(3):637–
666, 2000.
Adv. Math., 167(2):195–264, 2002.
[13] E. Kirchberg and M. Rørdam. Infinite non-simple C ∗-algebras: absorbing the Cuntz algebras O∞.
[14] F. Perera and A. S. Toms. Recasting the Elliott conjecture. Math. Ann., 338(3):669–702, 2007.
[15] L. Robert. The cone of functionals on the Cuntz semigroup. arXiv:1102.1451v2 [math.OA], pages
1–22, 2011.
16
RAMON ANTOINE, MARIUS DADARLAT, FRANCESC PERERA, AND LUIS SANTIAGO
[16] M. Rørdam. On the structure of simple C ∗-algebras tensored with a UHF-algebra. II. J. Funct.
Anal., 107(2):255–269, 1992.
[17] M. Rørdam and W. Winter. The Jiang-Su algebra revisited. J. Reine Angew. Math., 642:129–155,
2010.
[18] A. Tikuisis. The Cuntz semigroup of continuous functions into certain simple C ∗-algebras. Int. J.
Math., 22(7):1–37, 2011.
[19] A. S. Toms. On the classification problem for nuclear C ∗-algebras. Ann. of Math. (2), 167(3):1029–
1044, 2008.
[20] A. S. Toms. K-theoretic rigidity and slow dimension growth. Invent. Math., 183(2):225–244, 2011.
[21] A. S. Toms. Characterizing classifiable AH algebras. arXiv:1102.0932v1 [math.OA], pages 1–3,
2011.
[22] W. Winter. Nuclear dimension and Z-stability of pure C∗-algebras. to appear in Invent. Math.;
[23] W. Winter and J. Zacharias. The nuclear dimension of C ∗-algebras. Adv. Math., 224(2):461–498,
arXiv:1006.2731v2 [math.OA], pages 1–65, 2011.
2010.
RA, FP & LS: DEPARTAMENT DE MATEM `ATIQUES, UNIVERSITAT AUT `ONOMA DE BARCELONA,
08193 BELLATERRA, BARCELONA, SPAIN
E-mail address: [email protected], [email protected], [email protected]
MD: DEPARTMENT OF MATHEMATICS, PURDUE UNIVERSITY, WEST LAFAYETTE, IN 47907, USA
E-mail address: [email protected]
Current address: (LS) Department of Mathematics, University of Oregon, Eugene OR 97403, USA
E-mail address: [email protected]
|
1302.1114 | 2 | 1302 | 2013-06-14T10:43:47 | A decomposition theorem in II_1-factors | [
"math.OA",
"math.FA"
] | Building on results of Haagerup and Schultz, we decompose an arbitrary operator in a diffuse, finite von Neumann algebra into the sum of a normal operator and an s.o.t.-quasinilpotent operator. We also prove an analogue of Weyl's inequality relating eigenvalues and singular values for operators in a diffuse, finite von Neumann algebra. | math.OA | math |
A DECOMPOSITION THEOREM IN II1 -- FACTORS
K. DYKEMA∗, F. SUKOCHEV§, AND D. ZANIN§
Abstract. Building on results of Haagerup and Schultz, we decompose an
arbitrary operator in a diffuse, finite von Neumann algebra into the sum of a
normal operator and an s.o.t.-quasinilpotent operator. We also prove an ana-
logue of Weyl's inequality relating eigenvalues and singular values for operators
in a diffuse, finite von Neumann algebra.
1. Introduction
The following result is due to Schur (see e.g. [17]) and is one of the cornerstones
of linear algebra.
Theorem 1. For every matrix T ∈ Mn(C), there is a unitary matrix U ∈ Mn(C)
such that U −1T U is an upper-triangular matrix.
Alternatively, there exists a basis in Cn such that the matrix of the operator
T with respect to this basis is upper-triangular. Taking the diagonal part N of
the operator T in this basis, we obtain a normal operator. The difference T − N
is, obviously, a strictly upper-triangular matrix. Every strictly upper-triangular
matrix is, clearly, nilpotent. Thus, an arbitrary matrix is a sum of a normal matrix
and nilpotent matrix.
The following theorem due to Ringrose [13] extends Schur's result to the realm
of compact operators in Hilbert space. Recall that an operator Q is quasinilpotent
if its spectrum is {0} and that, by the C∗ -- property, the functional calculus for
self-adjoint operators and the spectral radius formula (see [10]), for an operator Q
on Hilbert space we have
and quasinilpotency of Q is equivalent to
k(cid:0)(Q∗)nQn(cid:1)1/2n
k∞ = k(Q∗)nQnk1/2n
∞ = kQnk1/n
∞
(1)
lim
n→∞
k(cid:0)(Q∗)nQn)1/2nk∞ = 0.
(Here and throughout this paper, we use the notation kXk∞ for the operator norm
of a bounded operator X on Hilbert space.)
Theorem 2. For every compact operator T ∈ B(H), there exists an increasing net
of projections pλ, λ ∈ [0, 1], with p0 = 0 and p1 = 1, such that, letting pλ−0 =
∨µ<λ pµ,
(a) T pλ = pλT pλ for all λ ∈ [0, 1].
(b) for every λ ∈ (0, 1] either pλ = pλ−0 or pλ − pλ−0 is a one-dimensional projec-
tion.
2000 Mathematics Subject Classification. 47C15.
∗ Research supported in part by NSF grant DMS -- 1202660.
§ Research supported by ARC.
1
2
DYKEMA, SUKOCHEV, AND ZANIN
Furthermore, if for such a family we have T pλ = pλ−0T pλ for all λ ∈ (0, 1], then
T is quasinilpotent.
This yields as an immediate corollary the following decomposition result:
Corollary 3. For every compact operator T ∈ B(H), there exist a normal operator
N and a quasinilpotent operator Q such that T = N + Q.
Proof. Indeed, set
N = Xpλ6=pλ−0
(pλ − pλ−0)T (pλ − pλ−0), Q = T − N,
where the sum (of pairwise orthogonal 1−dimensional operators) converges in strong
operator topology. It is clear that N is normal and that Qpλ = pλ−0Qpλ. In par-
ticular, it follows from Theorem 2 that Q is quasinilpotent.
(cid:3)
It is quite natural to ask whether the latter decomposition remains true in other
settings. In this paper, we concentrate on II1 -- factors and, more generally, diffuse,
finite von Neumann algebras (see §2.1).
The following result due to Haagerup and Schultz [9] is of utmost importance
for our investigation. (See §2.3 below for description of the Brown measure.)
Theorem 4. Let M ⊆ B(H) be a II1 -- factor with tracial state τ and let T ∈ M.
For every Borel set B ⊂ C, there exists a unique projection pB ∈ M such that
(a) τ (pB) = νT (B), where νT is the Brown measure of T
(b) T pB = pBT pB
(c) if pB 6= 0, then the Brown measure of T pB considered as an element of pBMpB
is supported in B
(d) if pB 6= 1, then the Brown measure of (1 − pB)T considered as an element of
(1 − pB)M(1 − pB) is supported in C\B.
Moreover, pB is T -hyperinvariant, meaning that it is invariant under every S ∈
B(H) that commutes with T . If Borel sets B1, B2 ⊂ C are such that B1 ⊂ B2, then
pB1 ≤ pB2 .
This projection pB is called the Haagerup -- Schultz projection for the operator T
associated to the set B.
Remark 5. In the above theorem, the hypothesis that M be a II1 -- factor and τ its
tracial state can clearly be loosened to require only that M be a diffuse, finite von
Neumann algebra and τ be any normal, faithful, tracial state on it. This is because
(a) any such M can be embedded, via a normal, trace-preserving ∗-homomorphism,
into a II1 -- factor (to see this, use for example Lemma 2.3 and Theorem 2.5 from [2]
to see that the free product with respect to traces of M with L∞[0, 1] is a II1-factor)
and (b) the projection onto a hyperinvariant subspace of any operator belongs to
the von Neumann algebra generated by the operator (this is easy to show, but for
a proof see, for example, Proposition 2.1 of [3]).
The purpose of this paper is to use the Haagerup -- Schultz theorem above to ob-
tain the following finite von Neumann algebra version of the Ringrose result. We
note that the quasinilpotent operator in Ringrose's theorem is here replaced by
an s.o.t.-quasinilpotent operator, which is an operator Q such that ((Q∗)nQn)1/2n
converges in strong operator topology to 0; in this notation (which was introduced
A DECOMPOSITION THEOREM IN II1 -- FACTORS
3
in [4]) s.o.t. is an abbreviation for "strong operator topology;" that quasinilpo-
tent operators must be s.o.t. -- quasinilpotent follows from the characterization (1)
of quasinilpotency. By Theorem 8.1 of [9], in a finite von Neumann algebra the
s.o.t.-quasinilpotent operators are precisely those whose Brown measures are con-
centrated at {0}.
Theorem 6. Let M be a diffuse, finite von Neumann algebra with normal, faithful,
tracial state τ and let T ∈ M. Then there exist N, Q ∈ M such that
(a) T = N + Q,
(b) the operator N is normal and the Brown measure of N equals that of T ,
(c) the operator Q is s.o.t-quasinilpotent.
As in Ringrose's theorem, a family (qt)0≤t≤1 of T -- invariant (in fact, T -- hyper-
invariant) projections is involved in Theorem 6, and N and Q can be regarded
as diagonal and upper triangular, respectively, with respect to this family of pro-
jections. Indeed, N is obtained as the conditional expectation of T onto the von
Neumann algebra generated by {qt 0 ≤ t ≤ 1}. However, unlike in Ringrose's the-
orem, the dimensions (i.e., in our finite von Neumann algebra setting, the traces)
of the projections qt can take large jumps as t varies. Indeed, if T itself is s.o.t.-
quasinilpotent, then our construction yields N = 0 and {qt 0 ≤ t ≤ 1} ⊆ {0, 1}.
For compact operators in B(H), it was shown in [13] that in Corollary 3, the
eigenvalues of the operator N coincide with those of the operator T (counting the
algebraic multiplicities). The analogous result for the operators in diffuse, finite
von Neumann algebras is given by Theorem 6 (b).
The eigenvalues λ(k, T ), k ≥ 0, and singular values µ(k, T ), k ≥ 0, of a compact
operator T ∈ B(H) are related by means of the Weyl theorem [16]. See [7] for
detailed proof.
Theorem 7. Let T ∈ B(H) be a compact operator and let Φ be a real -- valued
increasing function on [0, ∞) such that Φ(0) = 0 and Φ ◦ exp is convex. Then
∞
Xk=0
Φ(λ(k, T )) ≤
∞
Xk=0
Φ(µ(k, T )).
We will also prove the following theorem, which extends Weyl's result to the
II1 -- setting:
Theorem 8. Let M be a diffuse finite von Neumann algebra with normal, faithful,
tracial state τ and let T ∈ M. For the normal operator N from Theorem 6 and for
every real-valued increasing function Φ on [0, ∞) such that Φ ◦ exp is convex, we
have
Z Φ(z) dµT (z) = τ (Φ(N )) ≤ τ (Φ(T )).
2. Preliminaries
2.1. Finite von Neumann algebras and II1 -- factors. A von Neumann algebra
M is called finite if it has a normal, faithful, tracial state τ . It is called diffuse if it
has no minimal (nonzero) projections and it is called a factor if its center is trivial.
The infinite dimensional finite von Neumann algebra factors are diffuse and are
called II1 -- factors, and each of these has a unique tracial state τ , which is normal
and faithful. See, e.g., [10] for details.
4
DYKEMA, SUKOCHEV, AND ZANIN
Throughout this paper, M will denote a diffuse, finite von Neumann algebra and
τ will be a normal, faithful, tracial state on M.
2.2. Singular value function. For every T ∈ M, the generalised singular value
function µ(T ), denoted t → µ(t, T ) for t ∈ (0, 1), is defined by the formula (see,
e.g., [5])
µ(t, T ) = inf{kT pk∞ : p ∈ Proj(M), τ (1 − p) ≤ t}.
It is continuous from the right in t. Equivalently, µ(T ) can be defined in terms of
the distribution function dT of the operator T . That is, setting
dT (s) = τ (ET (s, ∞)),
s ≥ 0,
we obtain
µ(t, T ) = inf{s ≥ 0 : dT (s) ≤ t},
t > 0.
Here, ET denotes the projection valued spectral measure of the operator T .
The following result is a widely known consequence of the spectral theorem.
Lemma 9. Let M be a diffuse, finite von Neumann algebra equipped with a normal,
faithful, tracial state τ and let 0 ≤ A ∈ M. Then there is an increasing net
(ps)0≤s≤1 of projections in M with τ (ps) = s with
A = Z 1
0
µ(s, A) dps.
Sketch of proof. If s is such that µ(s, A) is a point of continuity of the distribu-
tion function dA, then let ps be the spectral projection EA((µ(s, A), ∞)). At
any remaining points s, the projection EA({µ(s, A)}) is nonzero; there are at
most countably many values r where EA({r}) is nonzero; for each of them, let
a(r) = τ (EA({r}) and choose an increasing family (q(r)
)0≤t≤a(r) of projections
with τ (q(r)
a(r) = EA({r}); when µ(s, A) = r is one of these points, let
ps = EA((µ(s, A), ∞)) + q(r)
t with t chosen so that τ (ps) = s.
(cid:3)
) = t and q(r)
t
t
2.3. Fuglede-Kadison determinant and Brown measure. Fuglede and Kadi-
son [6] constructed a mapping ∆ : M → R+ which is a homomorphism with respect
to the multiplication. This mapping is defined by
(2)
∆(T ) = exp(τ (log(T ))),
T ∈ M.
For every operator T, the function
(3)
λ → τ (log(T − λ)),
λ ∈ C,
is shown to be subharmonic by Brown [1]. Using this fact, Brown constructed a
probability measure νT such that
(4)
τ (log(T − λ)) = ZC
log(z − λ)dνT (z) λ ∈ C.
This νT , (called the Brown measure of T ) can be viewed as the II1 -- analogue of the
spectral counting measure (according to algebraic multiplicity) on matrices. It can
be recovered by taking the Laplacian of the mapping in (3).
The following is Proposition 2.24 of [8] and a consequence of it.
A DECOMPOSITION THEOREM IN II1 -- FACTORS
5
Theorem 10. If T ∈ M and if p ∈ M is a projection such that T p = pT p, so that
we may write T = ( A B
0 C ) , where A = T p and C = (1 − p)T, then
∆M(T ) = ∆pMp(A)τ (p)∆(1−p)M(1−p)(C)τ (1−p)
(5)
and
νT = τ (p)νA + τ (1 − p)νC .
We will use the equation (5) also in the case of p = 0 or p = 1, by making the
convention ∆{0}(0)0 = 1.
2.4. Haagerup -- Schultz projections and s.o.t. -- quasinilpotent operators.
It is proved in [9] that, for every T ∈ M, ((T ∗)nT n)1/2n converges as n → ∞
in strong operator topology. The spectral projection of the limiting operator on
the interval [0, r] is exactly the Haagerup -- Schultz projection pBr from Theorem 4
corresponding to the ball Br = {z ≤ r}. Thus, an operator T has Brown measure
supported on {0} if and only if ((T ∗)nT n)1/2n converges in strong operator topology
to 0. We call such operators s.o.t.-quasinilpotent (this notation was introduced in
[4]).
Though aesthetically attractive, the above definition of the projection pBr is not
suitable for our purposes. We employ a different characterization also taken from
[9]. Define the subspace Hr of the Hilbert space H by setting
(6)
Hr = {ξ ∈ H : ∃ξn → ξ, with lim sup
n→∞
kT nξnk1/n ≤ r}.
The projection onto the subspace Hr is shown in [9] to be the Haagerup -- Schultz
projection pBr corresponding to the ball Br = {z ≤ r}.
We will not need the construction of Haagerup -- Schultz projections for sets other
than balls; see [9] for the construction in the general case.
2.5. Submajorization and logarithmic submajorization. The operator B ∈
M is said to be submajorized by the operator A ∈ M (written B ≺≺ A) if
Z t
0
µ(s, B)ds ≤ Z t
0
µ(s, A)ds,
0 < t < 1.
The importance of submajorization can be observed from the following theorem,
which is really a result about functions rather than operators and is essentially
an inequality of Hardy, Littlewood and Polya (see e.g Lemma II.3.4 of [7] for the
sequence version, or Proposition 14.H.1.a of [12] for a result that implies the fol-
lowing).
Theorem 11. If A, B ∈ M and if B ≺≺ A, then for every increasing convex
function Φ on [0, ∞), we have
τ (Φ(B)) ≤ τ (Φ(A)).
We also need the notion of logarithmic submajorization. The operator B ∈ M is
said to be logarithmically submajorized by the operator A ∈ M (written B ≺≺log
A) if
Z t
0
log(µ(s, B))ds ≤ Z t
0
log(µ(s, A))ds,
0 < t < 1.
We collect some easy observations into a lemma, for future use.
6
DYKEMA, SUKOCHEV, AND ZANIN
Lemma 12. If A, B ∈ M and if c > 0, then
B ≺≺ A
B ≺≺log A
⇔
⇔
cB ≺≺ cA,
cB ≺≺log cA.
Furthermore, if 1 ≤ A, B ∈ M, where 1 represents the identity operator, then
B ≺≺log A
⇔
log(B) ≺≺ log(A).
Proof. The first assertion follows from the fact that µ(s, cA) = cµ(s, A) and the
second from the fact that, for A ≥ 1, we have µ(s, log(A)) = log(µ(s, A)).
(cid:3)
2.6. Conditional expectation. If D is a von Neumann subalgebra of the finite
von Neumann algebra M with normal, faithful, tracial state τ , then there exists a
unique linear operator ExpD : M → D such that, for all A ∈ M and B ∈ D,
(a) ExpD(AB) = ExpD(A)B
(b) ExpD(BA) = BExpD(A)
(c) τ (ExpD(A)) = τ (A).
Furthermore, ExpD is positive (in fact, completely positive), of norm 1 and can
be realised as the orthogonal projection from L2(M) onto L2(D) restricted to M.
See, for example, [15] for these and other facts.
It is well known and not difficult to verify that for the action of M on L2(M),
the strong operator topology on bounded sets of M coincides with the topology
provided by the norm kxk2 = τ (x∗x)1/2. From this, it is easy to prove the following
well known lemma:
Lemma 13. Assume An, n ≥ 1, is a family of von Neumann subalgebras in M,
that is either increasing or decreasing in n. Let A be the strong operator closure of
n=1 An in the first case and A = T∞
S∞
ExpA(T ) = lim
n→∞
n=1 An in the second case. Then
ExpAn (T ),
(T ∈ M),
where the limit is taken in the strong operator topology.
3. Construction of the normal part
Throughout this section, M will be a diffuse, finite von Neumann algebra and
τ a normal, faithul, tracial state on M. Our plan for proving Theorem 6 is to
take as normal operator N = ExpD(T ) for a suitable commutative von Neumann
subalgebra D, namely, the one given below.
Construction 14. Let T ∈ M and let ρ : [0, 1] → {z ≤ kT k∞} be a Peano curve.1
(a) Set qt to be the Haagerup -- Schultz projection for T associated to the Borel
set ρ([0, t]). Then qt is increasing in t. Since τ (qt) = νT (ρ([0, t]), we have
qt = ∧t′>tqt′ ; i.e., qt is strong-operator-topology continuous from the right in t.
(b) Set D to be the von Neumann algebra generated by {qt t ∈ [0, 1]}.
(c) For every n ≥ 0, set Dn to be the algebra generated by qk/2n , 0 ≤ k ≤ 2n.
For technical convenience, we will assume that the Brown measure of T has no atom
at ρ(0) (i.e. νT (ρ(0)) = 0). This ensures q0 = 0 and it can always be arranged by
modification of ρ, if necessary.
1A continuous surjective mapping [0, 1] → [0, 1]2 was first constructed by Peano [14]. See,
[11] for details. We may take the ball rather than the square since they are
for example,
homeomorphic.
A DECOMPOSITION THEOREM IN II1 -- FACTORS
7
Since the function ρ is uniformly continuous, it follows that there exists a mono-
tone function ω : [0, 1] → R+ (called the modulus of continuity of ρ) such that
ω(+0) = 0 and such that
ρ(t1) − ρ(t2) ≤ ω(t1 − t2),
t1, t2 ∈ [0, 1].
Lemma 15. Let T ∈ M and let qt, D and Dn be as in Construction 14. Then
ExpDn (T ) converges in norm to ExpD(T ), and, in fact, we have
where ω is the modulus of continuity of ρ.
kExpDn(T ) − ExpD(T )k∞ ≤ ω(2−n),
Proof. By Theorem 4, the projections qk/2n , 0 ≤ k ≤ 2n, are increasing in k. Letting
f n
k = q(k+1)/2n − qk/2n for 0 ≤ k < 2n, we have
(7)
ExpDn(T ) = X0≤k<2n
f n
k 6=0
τ (f n
k T f n
k )
τ (f n
k )
f n
k .
k Mf n
k T f n
By Construction 14 and Theorem 4, when f n
k
2n ])\ρ([0, k
in f n
2n ])).
It follows from Theorem 4 (c),(d) combined with Brown's analogue of Lidskii's
theorem [1] that
k 6= 0, the Brown measure of f n
2n ]) (which is a subset of ρ(( k
2n , k+1
k is supported in ρ([0, k+1
(8)
τ (f n
k T f n
k )
τ (f n
k )
= Rρ([0, k+1
νT (ρ([0, k+1
2n ])\ρ([0, k
2n ]) zdνT (z)
2n ]))
2n ])\ρ([0, k
∈ conv(ρ((
k
2n ,
k + 1
2n ])).
Now take m > n and note that
f n
k =
2m−n(k+1)−1
Xj=2m−nk
f m
j .
For 2m−nk ≤ j < 2m−n(k + 1) such that f m
j
6= 0, we have
τ (f m
j T f m
j )
τ (f m
j )
∈ conv(ρ((
j
2m ,
j + 1
2m ])) ⊂ conv(ρ((
k
2n ,
k + 1
2n ])).
It follows that
(9)
kExpDn (T ) − ExpDm(T )k∞ ≤ max
0≤k<2n
diam(cid:0)ρ((
k
2n ,
k + 1
2n ])(cid:1) ≤ ω(2−n)
and this upper bound tends to 0 as n → ∞. By Lemma 13, ExpDn (T ) converges
in strong operator topology to ExpD(T ). By the above estimate, it is Cauchy in
the uniform norm, and, therefore, converges in that norm to ExpD(T ). Now letting
m → ∞ in (9) completes the proof of the lemma.
(cid:3)
Lemma 16. Let T ∈ M and let qt, D and Dn be as in Construction 14. For every
λ ∈ C, and ε > 0, we have
(10)
lim
n→∞
Proof. Letting f n
log ∆(ExpDn (T ) − λ2 + ε) = ZC
k be as in the proof of Lemma 15 and using (7), we have
log(z − λ2 + ε)dνT (z).
(11)
log ∆(ExpDn (T ) − λ2 + ε) = X0≤k<2n
f n
k 6=0
τ (f n
k ) log(
τ (f n
k T f n
k )
τ (f n
k )
− λ2 + ε).
8
DYKEMA, SUKOCHEV, AND ZANIN
We now define for each n a function xn on the disk {z ≤ kT k∞}. Letting z be a
complex number with z ≤ kT k∞ and z 6= ρ(0), we have z ∈ ρ([0, k+1
2n ])
for some unique k = k(z) ∈ {0, . . . , 2n − 1}. Indeed, selecting the minimal t such
that z = ρ(t), we take k such that k/2n < t ≤ (k + 1)/2n. In the case f n
k 6= 0, we
let
2n ])\ρ([0, k
(12)
xn(z) = log(
τ (f n
k T f n
k )
τ (f n
k )
− λ2 + ε),
while if z = ρ(0) or f n
however, that the set {ρ(0)} ∪ {z f n
k(z) = 0, then for specificity we set xn(z) = log ε. (Note,
k(z) = 0} of such exceptional z is a νT -- null set.)
2n ])). Hence, using (11),
k ) = νT (ρ([0, k+1
2n ])\ρ([0, k
By Theorem 4, we have τ (f n
we get
(13)
log ∆(ExpDn(T ) − λ2 + ε) = Z{z≤kT k∞}
xn(z)dνT (z).
Moreover, we clearly have
log(ε) ≤ xn(z) ≤ log(ε + (λ + kT k∞)2)
for every z ≤ kT k∞ and, therefore,
kxnk∞ ≤ max{ log(ε), log(ε + (λ + kT k∞)2)}.
Given z ∈ ρ([0, k+1
2n ])\ρ([0, k
2n ]) with τ (f n
k ) 6= 0, by Theorem 4 (c),(d) combined
with Brown's version of Lidskii theorem [1], we have
τ (f n
k T f n
k )
τ (f n
k )
∈ conv(ρ([0,
k + 1
2n ])\ρ([0,
k
2n ])) ⊂ conv(ρ((
k
2n ,
k + 1
2n ])).
Thus,
(14)
z −
τ (f n
k T f n
k )
τ (f n
k )
≤ diam(ρ((
k
2n ,
k + 1
2n ]))) ≤ ω(2−n).
Combining (14) and (12), we infer that xn(z) converges to log(z −λ2+ε) as n → ∞
on a set of full νT measure. The Dominated Convergence Principle now yields that
the right-hand-side of (13) tends to the right-hand-side of (10) as n → ∞.
(cid:3)
Note that one could not first prove Lemma 16 in the case λ = 0 and then infer
its assertion in full generality by applying this result to the operator T − λ. The
reason is that algebra D for the operator T differs from that fo r the operator T − λ.
The following proposition is central to this section. It proves Theorem 6 (b).
Proposition 17. Let T ∈ M, D and Dn be as in Construction 14. The Brown
measure of the normal operator ExpD(T ) equals that of T.
Proof. Fix ε > 0. By Lemma 15, we have that ExpDn(T ) converges in norm to
ExpD(T ) as n → ∞. Since the Fuglede-Kadison determinant is continuous with
respect to the uniform norm topology on the set of invertible elements (see [6]), for
every λ ∈ C we have
log ∆(ExpDn (T ) − λ2 + ε) → log(∆(ExpD(T ) − λ2 + ε)).
On the other hand, it follows from Lemma 16 that
log ∆(ExpDn (T ) − λ2 + ε) → ZC
log(z − λ2 + ε)dνT (z)
A DECOMPOSITION THEOREM IN II1 -- FACTORS
9
for every λ ∈ C. Hence,
log(∆(ExpD(T ) − λ2 + ε)) = ZC
log(z − λ2 + ε)dνT (z)
for every λ ∈ C. Letting ε → 0, we infer from the Monotone Convergence Principle
that
ZC
log(z − λ)dνExpD (T )(z) = ZC
log(z − λ)dνT (z)
for every λ ∈ C. The assertion follows by taking Laplacians of both sides.
(cid:3)
It is tempting to try to infer Theorem 6 from Proposition 17 by applying its
assertion to the operator T − ExpD(T ). This is impossible because the algebra D
for the operator T differs from that for the operator T − ExpD(T ).
4. Decomposition
The following submajorization result is related to the Weyl lemma stating that
λ(T ) ≺≺log µ(T ) for every compact operator T ∈ B(H) (see, e.g., Theorem II.3.1
of [7]).
We continue to assume that M is a diffuse, finite von Neumann algebra and that
τ is a normal, faithful, tracial state on M.
Lemma 18. Let T ∈ M and let p ∈ M be a projection such that T p = pT p. Then
T p + (1 − p)T ≺≺log T.
Proof. Let S = T p + (1 − p)T. Writing elements of M as matrices with respect to
projections p and (1 − p), we have
T = (cid:18)A B
0 C(cid:19) ,
S = (cid:18)A 0
0 C(cid:19) ,
with A = T p and C = (1 − p)T. By Lemma 9, there exist increasing nets ps ≤ p,
0 ≤ s ≤ τ (p), and qs ≤ 1 − p, 0 ≤ s ≤ τ (1 − p), of projections in M such that
τ (ps) = s, τ (qs) = s and such that
C = Z τ (supp(C))
0
µ(s, C)dqs,
A∗ = Z τ (supp(A))
0
µ(s, A)dps.
Choosing appropriate spectral projections of A and C, we see that for every
t > 0, there exist t1, t2 > 0 such that t1 + t2 = t and such that
(15)
Set
and
Note that
Z t
0
log(µ(s, S))ds = Z t1
0
log(µ(s, A))ds +Z t2
0
log(µ(s, C))ds.
A0 = Z t1
0
µ(s, A)dps, C0 = Z t2
0
µ(s, C)dqs
0 C0(cid:19) , S0 = (cid:18)A0
T0 = (cid:18)A0 B
Z t
log(µ(s, S))ds = Z t
0
log(µ(s, S0))ds.
0
0
0 C0(cid:19) .
10
DYKEMA, SUKOCHEV, AND ZANIN
We now claim that µ(T0) ≤ µ(T ). Indeed,
0 C(cid:19)∗(cid:18)A B
(cid:18)A B
0 C(cid:19) = (cid:18)A∗A
B∗A B∗B + C∗C(cid:19) ≥ (cid:18)A B
0 C0(cid:19)∗(cid:18)A B
0 C0(cid:19)
A∗B
and
0 C0(cid:19)(cid:18)A B
(cid:18)A B
0 C0(cid:19)∗
Therefore,
= (cid:18)AA∗ + BB∗ BC0
C0B
C2
0 (cid:19) ≥ (cid:18)A0 B
0 C0(cid:19)(cid:18)A0 B
0 C0(cid:19)∗
.
µ(T ) = µ((cid:18)A B
0 C(cid:19)) ≥ µ((cid:18)A B
0 C0(cid:19)) ≥ µ((cid:18)A0 B
0 C0(cid:19)) = µ(T0).
Let now r = pτ (supp(C0)) + qτ (supp(A0)). Using (2), we get
Z t
0
log(µ(s, S))ds = Z t
0
log(µ(s, S0))ds = log(∆rMr(S0))
and, since µ(rT0r) ≤ µ(T0) ≤ µ(T ), we get
log(∆rMr(rT0r))
(2)
= Z t
0
log(µ(s, rT0r))ds ≤ Z t
0
log(µ(s, T ))ds.
It follows now from Theorem 10 that ∆rMr(S0) = ∆rMr(rT0r) and, therefore,
Z t
0
log(µ(s, S))ds ≤ Z t
0
log(µ(s, T ))ds.
(cid:3)
Corollary 19. Let T ∈ M and let p ∈ M be a projection such that T p = pT p.
Then
∆(1 +(cid:12)(cid:12)T p + (1 − p)T )(cid:12)(cid:12)
2
) ≤ ∆(1 + T 2).
Proof. Set y = µ(T p + (1 − p)T ) and x = µ(T ). We may without loss of generality
assume y is not identically 0. By Lemma 18, we have y ≺≺log x. Set α to be
the infimum of the set y−1({0}). Then we must have inf x−1({0}) ≥ α. Let εn =
min{y(α − 1
n < α. Then the functions
n )} for all integers n so large that 1
n ), x(α − 1
yn = log(ε−1
n y)χ(0,α−1/n),
xn = log(ε−1
n x)χ(0,α−1/n),
when nonzero, take only values ≥ 0. Clearly, yn ≺≺ xn. Since the function Φn :
z → log(1 + ε2
ne2z) is convex on [0, ∞), it follows from Theorem 11 that
0
Z α−1/n
≤ Z α−1/n
0
log(1 + y2(s))ds = Z 1
0
Φn(yn(s))ds − (1 − α +
1
n
) log(1 + ε2
n)
Φn(xn(s))ds − (1 − α +
1
n
) log(1 + ε2
n) = Z α−1/n
0
log(1 + x2(s))ds.
Letting n → ∞, we obtain
Z 1
0
log(1 + y2(s))ds = Z α
0
log(1 + y2(s))ds
Now (2) finishes the proof.
(cid:3)
≤ Z α
0
log(1 + x2(s))ds ≤ Z 1
0
log(1 + x2(s))ds.
A DECOMPOSITION THEOREM IN II1 -- FACTORS
11
The following lemma is easy and we omit the proof.
Lemma 20. If a scalar sequence {an,m}n,m≥1 is decreasing in both arguments,
then
lim
n→∞
lim
m→∞
an,m = lim
m→∞
lim
n→∞
an,m.
The following lemmas make up the heart of the proof of Theorem 6. In the next
n := M ∩ (Dn)′ and D′ := M ∩ D′ mean the relative commutants
two lemmas, D′
of the respective algebras in M.
Lemma 21. Let pt ∈ M, t ∈ [0, 1], be an increasing net of projections with p0 = 0
and p1 = 1. Let m and n be positive integers and let Dn be the von Neumann
subalgebra generated by pk/2n , 0 ≤ k ≤ 2n. If T ∈ M and T pt = ptT pt for all
t ∈ [0, 1], then
(T )2 +
∆(ExpD′
1
m
n = p(k+1)/2n −pk/2n, 0 ≤ k < 2n and similarly f k
n+1 = p(k+1)/2n+1 −
) ≥ ∆(ExpD′
(T )2 +
1
m
).
n+1
n
Proof. We let f k
pk/2n+1, 0 ≤ k < 2n+1 . For an arbitrary X ∈ M, we have
ExpD′
n
(X) =
2n−1
Xk=0
n Xf k
f k
n,
ExpD′
n+1
(X) =
2n+1−1
Xk=0
f k
n+1Xf k
n+1.
Note that T is upper -- triangular with respect to the list of projections (f k
and (f k
n )0≤k<2n
n-invariant and we may write
n+1)0≤k<2n+1 and, in particular, f 2k
f k
nT f k
n+1 T f 2k+1
n+1
n = (cid:18)f 2k+1
0
n T f k
n+1 is f k
n+1T f 2k+1
f 2k
n+1 (cid:19) .
n+1
n+1T f 2k
f 2k
It follows from Corollary 19 and Theorem 10 that
∆f 2k+1
n+1 Mf 2k+1
n+1
(f 2k+1
n+1 T f 2k+1
n+1 2+
1
m
)τ (f 2k+1
n+1 )∆f 2k
n+1Mf 2k
n+1
(f 2k
n+1T f 2k
)τ (f 2k
n+1)
1
m
)τ (f k
n ).
n+12+
1
m
n2 +
≤ ∆f k
nMf k
n (f k
nT f k
It follows now from Theorem 10 that
∆(ExpD′
n
(T )2 +
1
m
) =
2n−1
Yk=0
∆f k
nMf k
n (f k
nT f k
n2 +
1
m
)τ (f k
n)
2n+1−1
≥
Yk=0
∆f k
n+1Mf k
n+1
(f k
n+1T f k
n+12 +
1
m
)τ (f k
n+1) = ∆(ExpD′
n+1
(T )2 +
1
m
)
(cid:3)
The next lemma shows that the Fuglede-Kadison determinant of the operator
T coincides with that of its expectation onto the commutant of a nest of invariant
projections.
Lemma 22. Let pt ∈ M, t ∈ [0, 1], be an increasing net of projections, continuous
from the right, with p0 = 0 and p1 = 1 and let D be the von Neumann subalgebra
generated by {pt t ∈ [0, 1]}. If T ∈ M and T pt = ptT pt for all t ∈ [0, 1], then
(16)
∆(T ) = ∆(ExpD′(T )).
12
DYKEMA, SUKOCHEV, AND ZANIN
Proof. Let Dn be the algebra generated by the projections pk/2n , 0 ≤ k ≤ 2n. Using
n and, by Lemma 13, we
the continuity from the right of pt, we get D′ = Tn≥1 D′
have
ExpD′(T ) = lim
n→∞
ExpD′
n
(T )
in the strong operator topology.
By Lemma 21, the quantity ∆(ExpD′
n
trivially, also decreasing in m. It follows from Lemma 20 that
(T )2 + 1
m ) is decreasing in n and it is,
(17)
lim
n→∞
lim
m→∞
∆(ExpD′
n
(T )2 +
1
m
) = lim
m→∞
lim
n→∞
∆(ExpD′
n
(T )2 +
1
m
).
Note that, by T -invariance of the projections pk/2n , using them to write T as
a block matrix of operators, yields an upper triangular matrix, and ExpD′
(T ) is
obtained by setting the non-diagonal blocks to zero. To compute the left hand side
of (17), Theorem 10 yields ∆(ExpD′
(T )) = ∆(T ) and, thereby, we have
n
n
lim
m→∞
∆(ExpD′
n
(T )2 +
1
m
) = 2∆(ExpD′
n
(T )) = 2∆(T ).
Letting n → ∞, we infer that the left hand side of (17) is 2∆(T ). To compute the
right hand side of (17), by Lemma 2.25 of [8], we have
lim
n→∞
∆(ExpD′
n
(T )2 +
1
m
) = ∆(ExpD′(T )2 +
1
m
).
Letting m → ∞, we infer that the right hand side of (17) is 2∆(ExpD′(T )). Thus,
we have (16).
(cid:3)
Lemma 23. Let T ∈ M and let ρ, D and Dn be as in Construction 14. If ω is the
modulus of continuity of ρ, then the Brown measure of T − ExpDn(T ) is supported
in the ball of radius ω(2−n) centered at the origin.
Proof. The operator T − ExpDn(T ) when written as a matrix with respect to the
nonzero projections from the list (f n
is upper triangular by Theorem 4 (b)
and has for diagonal entries
k )2n−1
k=0
(18)
f n
k T f n
k −
τ (f n
k T f n
k )
τ (f n
k )
f n
k .
k T f n
k in f n
k Mf n
Repeating the argument in Lemma 15 (see (8)), we obtain, when f n
k 6= 0, that the
Brown measure of f n
k )/τ (f n
k )
lies in the convex hull of this set. Hence, the Brown measure of the difference (18)
is supported in the ball centered at the origin of radius diam(ρ(( k
2n ])), which
is no greater than ω(2−n). Now applying Theorem 10 n times, we get that the
Brown measure of T − ExpDn(T ) lies in the ball of radius ω(2−n) centered at the
origin.
(cid:3)
k is supported in ρ(( k
2n ]) and τ (f n
2n , k+1
2n , k+1
k T f n
The following lemma gives the decomposition of Theorem 6 in a special case.
Lemma 24. Let T ∈ M and let D be as in Construction 14. If T ∈ D′, then
T − ExpD(T ) is s.o.t.-quasinilpotent.
Proof. We assume without loss of generality kT k∞ ≤ 1/2. For every n ≥ 0, let
Dn be the subalgebra of D generated by qk/2n , 0 ≤ k ≤ 2n. Fix n ∈ N and a unit
A DECOMPOSITION THEOREM IN II1 -- FACTORS
13
vector η ∈ H. By assumption, T commutes with ExpD(T ). Since T − ExpDn (T )
and ExpD(T ) − ExpDn(T ) commute, we have
(T − ExpD(T ))2m =
2m
Xk=0
(−1)k(cid:18)2m
k (cid:19)(ExpD(T ) − ExpDn (T ))2m−k(T − ExpDn (T ))k.
Since kT k∞ ≤ 1/2, it follows that both ExpD(T ) − ExpDn (T ) and T − ExpDn (T )
are contractions.
For k ≤ m, we have
k(ExpD(T ) − ExpDn(T ))2m−k(T − ExpDn (T ))kηk ≤ kExpD(T ) − ExpDn(T )km
∞.
For k > m, we have
k(ExpD(T ) − ExpDn (T ))2m−k(T − ExpDn(T ))kηk ≤ k(T − ExpDn (T ))mηk.
Hence, by Lemma 15, we have
(19)
k(T − ExpD(T ))2mηk ≤ 22m max{ω(2−n)m, k(T − ExpDn (T ))mηk}.
By Lemma 23, the Brown measure of T − ExpDn(T ) is contained in the ball
of radius ω(2−n) centered at 0. By the Haagerup -- Schultz characterization (6), for
every unit vector ξ ∈ H, there exists a sequence ξm → ξ such that kξmk = 1 and
lim sup
m→∞
k(T − ExpDn (T ))mξmk1/m ≤ ω(2−n).
Hence, there exists M (depending on n), such that
k(T − ExpDn(T ))mξmk ≤ (2ω(2−n))m, m > M.
Substituting ξm for η in (19), we obtain
k(T − ExpD(T ))2mξmk1/m ≤ 8ω(2−n), m > M.
Since ξ was any unit vector, it follows from the characterization (6) of the Haagerup --
Schultz subspace that the Brown measure of (T − ExpD(T ))2 is supported in the
ball centered at 0 of radius 8ω(2−n). Letting n → ∞, we obtain that the Brown
measure of T − ExpD(T ) is δ0.
(cid:3)
Proof of Theorem 6. Let qt, t ∈ [0, 1], and D be as in Construction 14. Applying
Lemma 22 to the operators T qt − λ and (1 − qt)T − λ, we obtain
∆(T qt − λ) = ∆(ExpD′(T qt) − λ),
∆((1 − qt)T − λ) = ∆(ExpD′((1 − qt)T − λ)).
Hence, by (4), the Brown measure of the operator T qt (respectively, of (1 − qt)T )
equals that of ExpD′(T )qt (respectively, of (1 − qt)ExpD′(T )). In particular, the
projection qt is the Haagerup -- Schultz projection for ExpD′(T ) associated to the
set ρ([0, t]). By definition of D′, the operator ExpD′(T ) commutes with every qt.
Hence, ExpD′(T ) satisfies the hypotheses of Lemma 24. Since D ⊆ D′ we have
ExpD(ExpD′(T )) = ExpD(T ) and by application of this lemma we infer that the
Brown measure of the operator ExpD′(T ) − ExpD(T ) is δ0.
Applying Lemma 22 to the operator T − ExpD(T ) − λ, we obtain
∆(T − ExpD(T ) − λ) = ∆(ExpD′(T ) − ExpD(T ) − λ).
Hence, by (4), the Brown measure of the operator T − ExpD(T ) equals that of
ExpD′(T ) − ExpD(T ), i.e., it is δ0.
14
DYKEMA, SUKOCHEV, AND ZANIN
Now letting Q = T − ExpD(T ) and N = ExpD(T ), we have T = N + Q with N
normal and Q s.o.t.-quasinilpotent. This proves Theorem 6 (c). Theorem 6 (b) is
already proved in Proposition 17.
(cid:3)
5. A Weyl inequality in the II1 -- setting
In the following assertion, log+(t) := max(log(t), 0). The analogue of the follow-
ing assertion for Hardy-Littlewood submajorization is well-known.
Lemma 25. We have B ≺≺log A if and only if for all t > 0 we have
(20)
τ (log+(
B
t
)) ≤ τ (log+(
A
t
)).
Proof. We prove the if part first. For a given u ∈ (0, 1), set t = µ(u, A). We have
Z u
0
log(
µ(s, B)
t
Hence, B ≺≺log A.
)ds ≤ τ (log+(
B
t
)) ≤ τ (log+(
A
t
)) = Z u
0
log(
µ(s, A)
t
)ds.
We now prove the only if part. If t > kBk∞, then (20) holds trivially, while if
t ≤ kBk∞ then we have
τ (log+(
B
t
)) = Z dB(t)
0
log(
µ(s, B)
t
)ds ≤ Z dB(t)
≤ Z dA(t)
0
0
log(
µ(s, A)
t
)ds
log(
µ(s, A)
t
)ds = τ (log+(
A
t
)).
(cid:3)
Lemma 26. Let M be a finite von Neumann algebra equipped with a faithful normal
tracial state τ. If 0 ≤ A, B ∈ M are such that B ≺≺log A, then B + 1 ≺≺log A + 1.
Proof. By enlarging M if necessary, we may without loss of generality assume it
is diffuse. Referring to Lemma 9, we infer the existence of increasing nets ps, qs
0 ≤ s ≤ 1, of projections in M such that τ (ps) = s, τ (qs) = s and such that
B = Z 1
0
µ(s, B)dqs, A = Z 1
0
µ(s, A)dps.
Fix t < τ (supp(B)) and define the operators
Bt = Z t
0
µ(s, B)dqs, Ct = Z t
0
µ(s, B)dps, At = Z t
0
µ(s, A)dps.
It is clear that µ(Bt) = µ(Ct), Ct ≺≺log At and the operators At, Ct are in the
von Neumann algebra ptMpt. Note that we have τ (supp(Bt)) = τ (supp(Ct)) ≤
τ (supp(At)). Since t < τ (supp(B)), it follows that At and Ct are invertible and
log(At), log(Ct) ∈ ptMpt. Let r = max{kA−1
t k∞}. By Lemma 12, from
Ct ≺≺log At we get rCt ≺≺log rAt and log(rCt) ≺≺ log(rAt). Since the mapping
Φ : z → log(1 + r−1ez) is convex, it follows from Theorem 11 that
t k∞, kC−1
Z t
0
log(1 + µ(s, B))ds = τ (Φ(log(rCt))) ≤ τ (Φ(log(rAt))) = Z t
0
log(1 + µ(s, A))ds.
Since t < τ (supp(B)) is arbitrary, the assertion follows.
(cid:3)
Now we prove Theorem 8 and some more.
A DECOMPOSITION THEOREM IN II1 -- FACTORS
15
Theorem 27. Let M be a diffuse, finite von Neumann algebra and let τ be a
normal, faithful, tracial state on M. Let T ∈ M. If N ∈ M is the normal operator
constructed by means of Theorem 6, then
(a) N ≺≺log T.
(b) for every increasing real valued function Φ on [0, ∞) so that Φ ◦ exp is convex,
we have
(21)
Z Φ(z) dµT (z) = τ (Φ(N )) ≤ τ (Φ(T )).
Proof. Since the Brown measure of N equals the distribution of N and equals the
Brown measure of T , we have
τ (log+(
N
t
)) = Zz≥t
log(
z
t
)dνN (z) = Zz≥t
log(
z
t
)dνT (z).
It follows from Lemma 2.20 in [8] that
Therefore,
Zz≥t
log(
z
t
)dνT (z) ≤ τ (log+(
T
t
)).
τ (log+(
N
t
)) ≤ τ (log+(
T
t
))
Assertion (a) follows now from Lemma 25.
The equality in (21) holds since µT = µN and N is normal. Assertion (a)
together with Lemma 26 and homogeneity of logarithmic submajorization shows
that
nN + 1 ≺≺log nT + 1, n ≥ 1.
Now Lemma 12 yields log(nN + 1) ≺≺ log(nT + 1). Since Φ ◦ exp is convex on
R, so is the function Φ( 1
n exp(·)) and from Theorem 11 we get
τ (Φ(N +
1
n
))) ≤ τ (Φ(T +
1
n
))), n ≥ 1.
We now infer from the Monotone Convergence Principle that
τ (Φ(N )) = lim
n→∞
τ (Φ(N +
1
n
)) ≤ lim
n→∞
τ (Φ(T +
1
n
)) = τ (Φ(T )).
This proves (b).
(cid:3)
References
[1] Brown L. Lidskii's theorem in the type II case. in Geometric Methods in Operator Algebras
(Kyoto 1983), pp. 1 -- 35, Pitman Res. Notes Math. Ser., vol. 123, Longman Sci. Tech., Harlow,
1986.
[2] Dykema, K., Factoriality and Connes' invariant T (M) for free products of von Neumann
algebras. J. reine angew. Math. 450 (1994), 159-180.
[3] Dykema, K., Hyperinvariant subspaces for some B-circular operators, Math. Ann. 333
(2005), 485-523.
[4] Dykema K., Fang J., Skripka A. Upper triangular Toeplitz matrices and real parts of quasi-
nilpotent operators. Indiana Univ. Math. J. (to appear), arXiv:1206.5080.
[5] Fack T., Kosaki H. Generalized s-numbers of τ -measurable operators Pacific J. Math. 123
(1986), no. 2, 269 -- 300.
[6] Fuglede B., Kadison R.V. Determinant theory in finite factors Ann. Math. 55, no. 3, (1952),
520-530.
16
DYKEMA, SUKOCHEV, AND ZANIN
[7] Gohberg I., Krein M. Introduction to the theory of linear nonselfadjoint operators. Trans-
lations of Mathematical Monographs, Vol. 18 American Mathematical Society, Providence,
R.I.
[8] Haagerup U., Schultz H. Brown measures of unbounded operators affiliated with a finite von
Neumann algebra. Math. Scand. 100 (2007), no. 2, 209 -- 263.
[9] Haagerup U., Schultz H. Invariant subspaces for operators in a general II1 -- factor. Publ.
Math. Inst. Hautes Etudes Sci. No. 109 (2009), 19 -- 111.
[10] Kadison R., Ringrose J. Fundamentals of the theory of operator algebras. Vol. I. Elemen-
tary theory. Reprint of the 1983 original. Graduate Studies in Mathematics, 15. American
Mathematical Society, Providence, RI, 1997.
[11] Mandelbrot B. The fractal geometry of nature. Schriftenreihe fur den Referenten. W. H.
Freeman and Co., San Francisco, Calif., 1982.
[12] Marshall A., Olkin I., Arnold B. Inequalities: Theory of Majorization and its Applictaions.,
2nd Ed., Springer, New York, 2011.
[13] Ringrose J. Super-diagonal forms for compact linear operators. Proc. London Math. Soc. (3)
12 1962 367-384.
[14] Peano G. Sur une courbe, qui remplit toute une aire plane. Math. Ann. XXXVI, 157-160
(1890).
[15] Sinclair A., Smith R. Finite von Neumann algebras and masas. London Math. Soc. Lecture
Note Series, vol. 351, Cambridge University Press, Cambridge, U.K., 2008.
[16] Weyl H. Inequalities between the two kinds of eigenvalues of a linear transformation. Proc.
Nat. Acad. Sci. U. S. A. 35, (1949). 408 -- 411.
[17] Zhang F. Matrix theory. Basic results and techniques. Second edition. Universitext. Springer,
New York, 2011.
Department of Mathematics, Texas A&M University, College Station, TX, USA.
E-mail address: [email protected]
School of Mathematics and Statistics, University of new South Wales, Kensington,
NSW, Australia.
E-mail address: [email protected]
School of Mathematics and Statistics, University of new South Wales, Kensington,
NSW, Australia.
E-mail address: [email protected]
|
1006.2278 | 1 | 1006 | 2010-06-11T12:26:17 | On projective representations for compact quantum groups | [
"math.OA",
"math.QA"
] | We study actions of compact quantum groups on type I factors, which may be interpreted as projective representations of compact quantum groups. We generalize to this setting some of Woronowicz' results concerning Peter-Weyl theory for compact quantum groups. The main new phenomenon is that for general compact quantum groups (more precisely, those which are not of Kac type), not all irreducible projective representations have to be finite-dimensional. As applications, we consider the theory of projective representations for the compact quantum groups associated to group von Neumann algebras of discrete groups, and consider a certain non-trivial projective representation for quantum SU(2). | math.OA | math |
On projective representations for compact quantum groups
Kenny De Commer
Dipartimento di Matematica, Universit`a degli Studi di Roma Tor Vergata
Via della Ricerca Scientifica 1, 00133 Roma, Italy
e-mail: [email protected]
Abstract
We study actions of compact quantum groups on type I factors, which may be interpreted as projective
representations of compact quantum groups. We generalize to this setting some of Woronowicz' results
concerning Peter-Weyl theory for compact quantum groups. The main new phenomenon is that for
general compact quantum groups (more precisely, those which are not of Kac type), not all irreducible
projective representations have to be finite-dimensional. As applications, we consider the theory of
projective representations for the compact quantum groups associated to group von Neumann algebras
of discrete groups, and consider a certain non-trivial projective representation for quantum SU 2.
Keywords: compact quantum group; projective representation; Galois (co-)object
AMS 2010 Mathematics subject classification: 17B37, 81R50, 16T15
Introduction
It is well-known that for compact groups, one can easily extend the main theorems of the Peter-Weyl
theory to cover also projective representations. In this article, we will see that if one tries to do the
same for Woronowicz' compact quantum groups, one confronts at least one surprising novelty: not all
irreducible projective representations of a compact quantum group have to be finite-dimensional. On
the other hand, one will still be able to decompose any projective representation into a direct sum of
irreducible ones, and to determine certain orthogonality relations between the matrix coefficients of
irreducible projective representations.
The main tool we will use in this article are the Galois co-objects which we introduced in [10]. Indeed,
we showed there that when one quantizes the notion of a projective representation, this structure
plays the role of a 'generalized 2-cocycle function'.
In the first section of this article, we will develop a structure theory for such Galois co-objects in the
setting of compact quantum groups. A lot of the techniques we use are directly inspired by the theory
of the compact quantum groups themselves.
Supported in part by the ERC Advanced Grant 227458 OACFT "Operator Algebras and Conformal Field Theory"
1
In the second section, we will show that such Galois co-objects can be dualized into Galois objects for
their dual discrete quantum groups, a concept which was introduced in [8].
In the third section, which, except for the last part, is independent from the more technical second
section, we present a Peter-Weyl theory for projective representations of compact quantum groups.
We also show how projective representations give rise to module categories over the tensor category of
the (ordinary) representations, and introduce the notion of fusion rules between irreducible projective
representations and (ordinary) irreducible representations.
In the fourth section, we will give some details on the 'reflection technique' introduced in [10]. We
showed there that from any Galois co-object for a given compact quantum group, one can create a
(possibly) new locally compact quantum group. We will show that the type of this quantum group
(namely whether it is compact or not) is intimately tied up with the behavior of the Galois co-object
itself.
In the fifth section, we will consider the special case of compact Kac algebras. We show that in
this case, all irreducible projective representations will be finite-dimensional, and the theory becomes
essentially algebraic.
In the sixth and seventh section, we further specialize. We first quickly consider the case of finite
quantum groups (i.e. finite-dimensional Kac algebras), for which we can mostly refer to the existing
literature. Then we will treat co-commutative compact Kac algebras, which correspond to group
von Neumann algebras of discrete groups. In this case, the projective representations turn out to be
classified by certain special 2-cohomology classes of finite subgroups of the associated discrete group.
In particular, we will be able to deduce that the group von Neumann algebra of a torsionless discrete
group admits no non-trivial 2-cocycles. These results will be proven using only the material in the
first section and the first part of the third section.
In the eighth section, we give a concrete example of what can happen in the non-Kac case by consid-
ering a particular non-trivial Galois co-object for the compact quantum group SUq 2. We compute
explicitly all its associated projective representations, provide the corresponding orthogonality rela-
tions and calculate the fusion rules.
Notations and conventions
We will assume that all our Hilbert spaces are separable, and we take the inner product to be conju-
gate linear in the second argument. We also assume that all our von Neumann algebras have separable
predual.
By ι, we denote the identity map on a set.
We denote by d the algebraic tensor product between vector spaces, by b the tensor product between
Hilbert spaces, and by b the spatial tensor product between von Neumann algebras.
By Σ we denote the flip map between a tensor product of Hilbert spaces:
Σ : H1 b H2 Ñ H2 b H1 : ξ1 b ξ2 Ñ ξ2 b ξ1.
When A „ B H1, H2 and B „ B H2, H3 are linear spaces of maps between certain Hilbert spaces,
we will denote B A (cid:16)
i(cid:16)1 biai n N0, bi B, ai A .
°n
2
We use the leg numbering notation for operators on tensor products of Hilbert spaces. E.g.,
Z : H b2
Z on the first and third factor, and as the identity on the second factor.
Ñ H b2 is a certain operator, we denote by Z13 the operator H b3
if
Ñ H b3 acting as
At certain points, we will need the theory of weights on von Neumann algebras, which is treated in
Ñ 0, 8 is a
detail in the first chapters of [27]. When M is a von Neumann algebra, and ϕ : M
normal semi-finite faithful (nsf) weight on M , we denote
Nϕ (cid:16) x M ϕxx (cid:160) 8
for the space of square integrable elements, M
ϕ .
for the linear span of M
ϕ for the space of positive integrable elements, and Mϕ
1 Galois co-objects for compact quantum groups
We begin with introducing the following concepts.
Definition 1.1. A von Neumann bialgebra M, ∆M consists of a von Neumann algebra M and a
faithful normal unital -homomorphism ∆M : M Ñ M ¯bM , satisfying the coassociativity condition
∆M b ι∆M (cid:16) ι b ∆M ∆M .
A von Neumann bialgebra M, ∆M is called a compact Woronowicz algebra ([23],[20]) if there exists
a normal state ϕM on M which is ∆M -invariant:
ϕM b ι∆M x (cid:16) ι b ϕM ∆M x (cid:16) ϕM x1
for all x M.
A compact Woronowicz algebra is called a compact Kac algebra if there exists a normal ∆M -invariant
tracial state τM on M .
Remarks:
1. von Neumann bialgebras are also referred to as Hopf-von Neumann algebras in the literature.
However, we prefer the above terminology, as for example a finite-dimensional Hopf-von Neu-
mann algebra is not necessarily a Hopf algebra. Admittedly, a finite-dimensional von Neumann
bialgebra is also not necessarily a bialgebra, as there could be no co-unit, but this seems a lesser
ambiguity.
2. It is easy to see that a normal ∆M -invariant state on a von Neumann bialgebra M, ∆M , when
it exists, is unique. One can moreover show that this state will automatically be faithful. We
will then always use the notation ϕM for it in the general setting, but use the notation τM in
the setting of compact Kac algebras to emphasize the traciality.
3. Compact Woronowicz algebras can be characterized as those von Neumann bialgebras arising
from Woronowicz' compact quantum groups in the C-algebra setting ([34],[36]), by performing
a GNS-type construction. However, we have decided to focus only on the von Neumann algebraic
picture in this paper.
Let us also introduce the following notations, which will be constantly used in the following.
Notation 1.2. Let M, ∆M be a compact Woronowicz algebra. We denote by L 2
GNS-construction of M with respect to ϕM . That is, L 2
a pre-Hilbert space with respect to the inner product structure
, ΛM the
M is the completion of M , considered as
M ,
xx, yy (cid:16) ϕM yx,
3
and ΛM is the natural inclusion M ãÑ L 2
of B L 2
M by letting x M corresponding to the (bounded) closure of the operator
M . We then identify M as a von Neumann subalgebra
ΛM M Ñ L 2
M : ΛM y Ñ ΛM xy,
for all y M.
We will further denote by ξM the cyclic and separating vector ΛM 1M in L 2
ΛM x for all x M .
M , so that xξM (cid:16)
The following two unitaries are of fundamental importance.
Definition 1.3. Let M, ∆M be a compact Woronowicz algebra.
The right regular corepresentation of M, ∆M is defined to be the unitary V B L 2
is uniquely determined by the formula
M ¯bM which
V ΛM x b η (cid:16) ∆M xξM b η,
for all x M, η L 2
M .
The left regular corepresentation of M, ∆M is defined to be the unitary W M ¯bB L 2
is uniquely determined by the fact that
M which
W η b ΛM x (cid:16) ∆M xη b ξM ,
for all x M, η L 2
M .
We will in the following always use the above notations for these corepresentations. Note that estab-
lishing the unitarity of these maps requires some non-trivial work! An approach to compact quantum
groups based on the properties of such unitaries can be found in [3], section 4.
Let us now introduce the notion of a Galois co-object for a compact Woronowicz algebra (see [10]).
Definition 1.4. Let M, ∆M be a compact Woronowicz algebra. A right Galois co-object for
M, ∆M consists of a Hilbert space L 2
N
and a normal linear map ∆N : N Ñ N ¯bN , such that the following properties hold: with N op denoting
the set
N , a σ-weakly closed linear space N „ B L 2
M , L 2
N op :(cid:16) x
x N „ B L 2
N , L 2
M ,
we should have
1. N L 2
M is norm-dense in L 2
N , and N op
L 2
N is norm-dense in L 2
M ,
2. the space N is a right M -module (by composition of operators),
3. for each x, y N , we have xy M ,
4. ∆N xy (cid:16) ∆N x∆M y for all x N and y M ,
5. ∆N x
∆N y (cid:16) ∆M xy for all x, y N ,
6. ∆N is coassociative: ∆N b ι∆N (cid:16) ι b ∆N ∆N , and
7. the linear span of the set ∆N xy b z x N, y, z M is σ-weakly dense in N ¯bN .
If N1, ∆N1 and N2, ∆N2 are two Galois co-objects for a von Neumann bialgebra M, ∆M , we call
them isomorphic if there exists a unitary u : L 2
N2 such that uN1 (cid:16) N2 and
N1 Ñ L 2
Remarks:
∆N2 ux (cid:16) u b u∆N1 x
for all x N1.
1. The previous definition can be shown to be equivalent with the one presented in [10], Definition
0.5. Also remark that the previous conditions can be grouped together as follows: a Galois
co-object is a right Morita (or imprimitivity) Hilbert M -module (conditions 1 to 3) with a ∆M -
compatible coalgebra structure (conditions 4 and 5 and condition 6) which is 'non-degenerate'
(condition 7).
4
2. A trivial example of a right M, ∆M -Galois co-object is M, ∆M itself.
Indeed, the final
condition even holds in a stronger form, as it can be shown that already ∆M x1by x, y M
is σ-weakly dense in M ¯bM for compact Woronowicz algebras.
It follows that this stronger
condition is then in fact true for all Galois co-objects for compact Woronowicz algebras.
3. A treatment of Galois co-objects in the setting of Hopf algebras can be found in [6].
One can similarly define the notion of a left Galois co-object. Left Galois co-objects can be created
from right ones in the following way.
Definition 1.5. Let N, ∆N be a right Galois co-object for the compact Woronowicz algebra M, ∆M .
We call the couple N op, ∆N op , consisting of
N op
(cid:16) x
x N „ B L 2
N , L 2
M ,
together with the coproduct
∆N op x :(cid:16) ∆N x
,
x N op,
the opposite (left) Galois co-object of N, ∆N . It is a left Galois co-object for the compact Worono-
wicz algebra M, ∆M .
We call the couple N cop, ∆N cop , where
N cop
(cid:16) N „ B L 2
M , L 2
N
and
∆N cop (cid:16) ∆op
N : N Ñ N ¯bN : x Ñ Σ∆N xΣ,
for all x N,
the co-opposite (right) Galois co-object of N, ∆N . It is a right Galois co-object for the compact
Woronowicz algebra M, ∆op
M .
The following notation will be useful.
Notation 1.6. Let M, ∆M be a compact Woronowicz algebra, and N, ∆N a right Galois co-object
for M, ∆M . We denote
ΛN : N Ñ L 2
N : x Ñ xξM .
Remark: By the second condition in Definition 1.4, we know that N is a right M -module, and then
we trivially have that
ΛN xy (cid:16) xΛM y,
for all x N, y M.
By the third condition in that definition, together with the faithfulness of ϕM , we see that ΛN is
injective and that
xΛN x, ΛN yy (cid:16) ϕM yx,
for all x, y N.
And finally, by the first (and second) condition in that definition, we see that ΛN has norm-dense range.
One can construct for a Galois co-object N, ∆N certain unitaries which are analogous to the regular
corepresentations for a compact Woronowicz algebra (and coincide with them in case N, ∆N (cid:16)
M, ∆M ).
Proposition 1.7. Let M, ∆M be a compact Woronowicz algebra, and N, ∆N a right Galois co-
object for M, ∆M .
5
1. There exists a unitary
V : L 2
N b L 2
M Ñ L 2
N b L 2
N
which is uniquely determined by the property that for all η L 2
M and x N , we have
V ΛN x b η (cid:16) ∆N xξM b η.
Similarly, there exists a unitary
€W : L 2
N b L 2
N Ñ L 2
M b L 2
N ,
uniquely determined by the property that for all η L 2
M and x N , we have
€W η b ΛN x (cid:16) ∆N xη b ξM .
2. We have V B L 2
N ¯bN and €W
N ¯bB L 2
N .
3. For x N , we have
∆N x (cid:16)
V x b 1V
€W
1 b xW.
(cid:16)
4. The following 'pentagonal identities' hold:
V12 V13V23 (cid:16)
V23 V12
as maps from L 2
N b L 2
M b L 2
M to L 2
N b L 2
N b L 2
N , and
W12€W13€W23 (cid:16)
€W23€W12
as maps from L 2
N b L 2
N b L 2
N to L 2
M b L 2
M b L 2
N .
5. The following identities hold:
ι b ∆N
V (cid:16)
V12 V13,
∆N b ι
€W
(cid:16)
€W
23€W
13.
Proof. The statements for €W follow immediately from the ones for V , by considering the co-opposite
Galois co-object.
We then refer to [10] for the proofs of the first four statements (Proposition 2.3 for the first and second
assertion, Proposition 2.4 for the third and fourth). The fifth statement follows immediately from
combining the three preceding ones.
Remark: Although we referred to [10], we want to stress that these assertions are quite straightforward
to prove. For example, the surjectivity of V follows quite immediately from the seventh condition in
Definition 1.4, combined with the surjectivity of V .
Definition 1.8. We call the unitary V appearing in the previous proposition the right regular N, ∆N -
corepresentation of M, ∆M . We call the unitary €W the left regular N op, ∆N op -corepresentation
of M, ∆M (where we recall that N op, ∆N op is the left Galois co-object opposite to N, ∆N , see
Definition 1.5).
6
Remark: The general notion of an 'N, ∆N -corepresentation' will be introduced in the third section.
For the rest of this section, we will fix a compact Woronowicz algebra M, ∆M and a right Galois
co-object N, ∆N for M, ∆M . We then further keep denoting by V and W the right and left
regular corepresentations of M, ∆M , and by V and €W the right regular N, ∆N - and left regular
N op, ∆N op -corepresentation of M, ∆M .
Our following lemma improves the second assertion in Proposition 1.7.
Lemma 1.9. The following equalities hold:
N (cid:16) ω b ι
V ω B L 2
N
σ-weak closure
(cid:16) ι b ω
€W
ω B L 2
N
σ-weak closure.
Proof. We will again only prove the first identity, as the second one then follows by symmetry.
For ξ, η L 2
for x B L 2
N , denote by ωξ,η the normal functional on B L 2
N . Then for x, y N , a straightforward computation shows that
N determined by ωξ,η x (cid:16) xxξ, ηy
ωΛN x,ΛN y
b ι
V (cid:16) ϕM b ιy
b 1∆N x.
It is thus enough to prove that the linear span of such elements is σ-weakly dense in N .
Suppose that this were not so. Then we could find a non-zero ω N such that
ϕM y
ι b ω∆N x (cid:16) 0
for all x, y N.
Taking y equal to ι b ω∆N x, we would have ι b ω∆N x (cid:16) 0 for all x N by faithfulness of
ϕM . But then also
ι b ω∆N xm b 1 (cid:16) 0
for all x N, m M.
Now the set ∆M m1 m2 b 1 m1, m2 M has σ-weakly dense linear span in M ¯bM . It then
follows, by the conditions 2, 4 and 7 in Definition 1.4 that
and so necessarily ω (cid:16) 0, a contradiction.
ι b ωz (cid:16) 0
for all z N ¯bN,
The following result will allow us to obtain a decomposition for €W and V .
Proposition 1.10. Denote by N „ B L 2
N the von Neumann algebra
N (cid:16) x B L 2
N
V
x b 1
V (cid:16) x b 1 .
Then N satisfies the following properties.
1. The von Neumann algebra N is an l8-sum of type I-factors.
2. The equality N (cid:16) ω b ι
€W
ω B L 2
M , L 2
N
σ-weak closure holds.
Remark: In the special case where N, ∆N equals M, ∆M considered as a right Galois co-object
over itself, one denotes the above von Neumann algebra as xM .
7
Proof. Consider the unital normal faithful -homomorphism
AdL : B L 2
N Ñ M ¯bB L 2
N : x Ñ Σ V
x b 1
V Σ.
Then by Proposition 1.7.5, it follows that AdL is a coaction by M, ∆M :
∆M b ιAdL (cid:16) ι b AdL AdL.
Hence N is precisely the set B L 2
elements satisfying AdL x (cid:16) 1 b x. It is well-known (and easy to see) that the map
AdL of AdL-fixed elements in B L 2
N
N , that is, the set of
E : B L 2
N Ñ B L 2
N : x Ñ ϕM b ιAdL x
is then a normal conditional expectation of B L 2
of type I-factors (see for example Exercise IX.4.1 in [27]).
N onto N . This forces N to be an l8-direct sum
We now prove the second point. First of all, remark that
V23€W
12 (cid:16)
€W
12 V23,
which follows from a straightforward computation. From this, it is easy to get that
ι b AdL
€W (cid:16)
€W13,
and so all elements of the form ω b ι
that all elements of N can be approximated σ-weakly by such elements.
€W with ω B L 2
N , L 2
M
lie in N . We next show
For ω1, ω2 N, denote
ω1 ω2 :(cid:16) ω1 b ω2 ∆N N.
For ξ, η B L 2
for the normal functional x Ñ xxξ, ηy. Choose b, x, y N , and denote
N , denote θξ,η for the rank one operator ζ Ñ xζ, ηyξ on L 2
N , and denote ωξ,η
a (cid:16) ωΛN x,ΛN y
b ι
V N,
SN a (cid:16) ωΛN x,ΛN y
b ι
V
N op,
where we recall that N op
(cid:16) x
x N „ B L 2
N , L 2
M . We will prove the identity
E θΛN a,ΛN b
(cid:16) ϕM b
ϕM SN a b ι
€W
,
(1)
where E is the conditional expectation defined in the first part of the proof, and where ϕM b
and ϕM SN a are the obvious normal functionals on N . As the linear span of the θΛN a,ΛN b is
σ-weakly dense in B L 2
N by Lemma 1.9, and as E is a normal map with N as its range, the
second point of the proposition will follow from this identity.
To prove the identity (1), choose further c, d N . It is sufficient to prove then that
xE θΛN a,ΛN b
ΛN c, ΛN dy (cid:16) xϕM b
ϕM SN a b ι
€W
ΛN c, ΛN dy.
(2)
We remark now that a and SN a can also be rewritten in the following form, by a simple computation
involving only the definition of V :
a (cid:16) ϕN b ιy
b 1∆N x,
SN a (cid:16) ϕM b ι∆N y
x b 1.
Using again the definition of V , the left hand of equation (2) then simplifies to
ϕM b ϕM b ϕM b ϕM y
b 1 b b
b 1∆N d
24∆N c34∆N x12 .
(3)
8
On the other hand, using the definition of €W , we get that the right hand side of equation (2) becomes
ϕM b ϕM b ϕM b ϕM 1 b b
b 1 b d
∆N y
13 x b ∆2
N c,
(4)
where ∆2
N c (cid:16) ι b ∆N ∆N c. In both expressions (3) and (4), we can write
v (cid:16) ϕM b ιb
b 1∆N c,
and we then have to prove that
ϕM b ϕM b ϕM y b ∆N d
∆N x b v (cid:16) ϕM b ϕM b ϕM ∆N y b d
x b ∆N v. (5)
Now by the final condition in Definition 1.4 (and the second remark following it), it is enough to show
that these two expression are equal when we replace xbv by ∆N z m b1 and y bd by ∆N wn b1,
where w, z N and m, n M . But then the left hand side of (5) becomes
ϕM b ϕM b ϕM n
b 1 b 1∆2
M wz ∆M m b 1,
which by invariance of ϕM collapses to ϕM nwzm. A similar computation shows that with this
replacement, also the right hand side expression in (5) collapses to ϕM nwzm. This concludes the
proof.
Of course, we then also have
N (cid:16) ω b ι
€W ω B L 2
N , L 2
M
σ-weak closure,
which follows immediately by applying the -operation to both sides of the identity in the second
point of the previous proposition.
IN , for
Notation 1.11. By Proposition 1.10.1, we may identify the center Z
some countable set IN . Denoting pr the minimal central projection in Z
N associated to the element
r IN , we may further identify pr N with B Hr for some separable Hilbert space Hr. We also denote
N of N with l8
Proposition 1.12. The unital normal faithful -homomorphism
nr :(cid:16) dimHr N0 Y 8 .
AdR : N Ñ B L 2
N ¯bM : x Ñ Σ€W 1 b x
€W Σ
restricts to a -homomorphism N Ñ
N ¯bM , and defines in this way a right coaction of M, ∆M on N .
Moreover, the set of fixed elements for AdR coincides with the center Z
N of N .
Proof. From Proposition 1.10.2, it follows that €W N op ¯b
property as N op is a corner of a von Neumann algebra). Hence AdR x
applying Proposition 1.7.5, we get
N (we may apply the weak slice map
N . By
N ¯bM for x
Hence the first part of the proposition follows.
AdR b ιAdR x (cid:16) ι b ∆M AdR x.
If further x
Proposition 1.10.2, we deduce that xy (cid:16) yx for all y
N is a fixed element for AdR, then it follows that 1 b x
N .
N , i.e. x Z
€W (cid:16)
€W 1 b x. Again by
9
Corollary 1.13. Using Notation 1.11 and the notation from the previous proposition, the coaction
AdR restricts to an ergodic coaction
for each r IN .
Adr
R : B Hr Ñ B Hr ¯bM
We recall that a coaction α is called ergodic if the only elements satisfying αx (cid:16) x b 1 are scalar
multiples of the unit element.
Proof. Clearly, as Z
that AdR indeed restricts to B Hr . If then x is a fixed element for Adr
previous proposition, that x Z
N consists of the fixed points of AdR by the previous proposition, it is immediate
R , we have, again by the
N X B Hr , and x is a scalar operator.
Now as each Adr
state φN,r on B Hr , determined by the formula
R appearing in the previous corollary is ergodic, there exists a unique Adr
R -invariant
φN,r x1B Hr
(cid:16) ι b ϕM Adr
R x,
for all x B Hr .
Notation 1.14. If Tr is the positive trace class operator associated to the state φN,r on B Hr in-
troduced above, we denote by Tr,0 ¥ Tr,1 ¥ . . . the descending sequence of its eigenvalues, counting
multiplicities. We further fix in Hr a basis er,i, with 0 ¤ i (cid:160) nr, such that er,i is an eigenvector for
Tr with eigenvalue Tr,i.
We denote by er,ij
normal functionals on N „ B `rIN
Hr :
N the matrix units associated to the basis er,i, and we denote by ωr,ij the following
ωr,ij x (cid:16) xxer,i, er,j y,
x
N .
In the special case where N, ∆N equals M, ∆M considered as a right Galois co-object over itself,
we will denote the nr as mr, the Tr,j as Dr,j and the Hr as Kr, but otherwise keep all notation as
above.
Theorem 1.15. Denote
€Wr,ij (cid:16) ι b ωr,ji
€W N op
„ B L 2
N , L 2
M .
Then the following statements hold.
1. The unitary €W equals the strong convergent sum °
°nr 1
i,j(cid:16)0
rIN
€Wr,ij b er,ij.
2. For each r IN and 0 ¤ i, j (cid:160) nr, we have
both sums converging strongly.
nr1
k(cid:16)0
nr1
k(cid:16)0
€Wr,ik
€W
r,jk (cid:16) δi,j1L 2
M ,
€W
r,ki
€Wr,kj (cid:16) δi,j1L 2
,
N
10
3. For each r IN and 0 ¤ i, j (cid:160) nr, we have
nr1
∆N
€W
r,ij (cid:16)
€W
r,ik b
€W
r,kj,
k(cid:16)0
the sum again being a strongly converging one.
4. The following orthogonality relations hold:
ϕM
€Wr,ij
€W
s,kl (cid:16) δr,sδi,kδj,l Tr,j,
for all r, s IN , 0 ¤ i, j (cid:160) nr, 0 ¤ k, l (cid:160) ns.
Proof. The first point is immediate, and also the second one follows straightforwardly from the uni-
13 in Proposition 1.7.5.
tarity of €W . The third point follows from the identity ∆N b ι
In the fourth point, the orthogonality relations for r (cid:16) s follow from writing out the identity
23€W
€W
€W
(cid:16)
ι b ϕM Adr
R er,ij (cid:16) φN,r er,ij (cid:16) δi,j Tr,j.
Thus the only thing left to show is that ϕM
s,kl (cid:16) 0 for r (cid:24) s. But also here, we can use a
standard technique (see e.g. [34]). For suppose that this were not so, and choose r (cid:24) s which violate
this condition. Consider, for x B Hs, Hr , the element
€Wr,ij
€W
where of course
F x (cid:16) ϕM b ι
€Wr 1 b x
€W
s B Hs, Hr ,
€Wr (cid:16) 1 b pr
€W (cid:16)
nr 1
i,j(cid:16)0
€Wr,ij b er,ij N op ¯bB Hr .
By assumption, there must exist an x with F x (cid:24) 0. Fixing such an x, denote y (cid:16) F x. Then it is
easy to see that
€Wr 1 b y
€W
s (cid:16) 1 b y,
using Proposition 1.7.5 and the ∆M -invariance of ϕM . This implies that yy, resp. yy, is a fixed
element for Ads
R . Since these coactions are ergodic, yy and yy must be (identical)
scalars, and so we can scale x such that y becomes a unitary u.
R , resp. Adr
We then find that
1 b u
€Ws 1 b u
(cid:16)
€Wr.
This implies that there exist two non-equal normal functionals ω1 and ω2 on N such that
ι b ω1
€W (cid:16) ι b ω2
€W .
As the set ω b ι
clearly gives a contradiction. Hence ϕM
€W ω B L 2
N , L 2
€Wr,ij
€W
s,kl (cid:16) 0 for r (cid:24) s.
M is σ-weakly dense in N by Proposition 1.10.2, this
Notation 1.16. By the final part of the previous proposition, we have a unitary transformation
by means of the map
L 2
N (cid:21)
Hr b Hr,
rIN
€W
r,ijξM Ñ T 1{2
r,j er,i b er,j.
In the following, we will then always identify L 2
example the elements x N act directly as linear operators
N and À
rIN
Hr b Hr in this way, so that for
Kr b Kr Ñ
Hr b Hr.
rIM
rIN
11
Lemma 1.17.
1. With Vr,ij :(cid:16) T 1{2
r,i T 1{2
r,j
€W
r,ij, we have the identity
nr 1
V (cid:16)
er,ij b
Vr,ij,
rIN
i,j(cid:16)0
the sum converging strongly.
2. The Vr,ij satisfy the following orthogonality relations:
ϕM
V
r,ij Vs,kl (cid:16) δr,sδi,kδj,lTr,i
for all r, s IN , 0 ¤ i, j (cid:160) nr, 0 ¤ k, l (cid:160) ns.
3. The following equalities hold:
N 1
(cid:16) ι b ω
(cid:16) x B L 2
V ω N
N
€W
1 b x
€W (cid:16) 1 b x .
Proof. Choose r IN , 0 ¤ i, j (cid:160) nr and η L 2
M . Then we compute
V er,i b er,j b η (cid:16) T 1{2
r,j
V
€W
r,ijξM b η
(cid:16) T 1{2
r,j
nr 1
k(cid:16)0
€W
r,ikξM b
€W
r,kjη
(cid:16)
nr1
k(cid:16)0
r,k T 1{2
T 1{2
r,j
er,i b er,k b
€W
r,kjη.
From this, the first point in the lemma follows.
The second point is of course just a reformulation of Theorem 1.15.4.
These orthogonality relations then immediately imply that
N 1
(cid:16) ι b ω
V ω N
.
Also the second equality of the third point follows straightforwardly: if x B L 2
N and
€W
1 b x
€W (cid:16) 1 b x,
then xω b ι
x
N 1. As €W N op
. From Proposition 1.10.2, we conclude that
N 1 satisfies €W
€W (cid:16) 1 b x.
1 b x
€W (cid:16) ω b ι
€W x for all ω N op
N , it is also clear that any x
b
Recall that we had introduced in Definition 1.5 the notion of the co-opposite Galois co-object
N cop, ∆N cop (cid:16) N, ∆op
N .
The following lemma gathers some transfer results between this structure and the original one.
Lemma 1.18.
1. The right regular N, ∆op
N -corepresentation for M, ∆op
M equals Σ€W Σ, while
the left regular N, ∆op
N -corepresentation equals Σ V Σ.
2. The dual von Neumann algebra N cop
^ equals N 1.
12
Proof. The two statements are easily verified (the second one follows from Lemma 1.17.3).
One can also relate the two adjoint coactions on respectively N and N 1, but this result requires some
more preparation. We will relegate this investigation to the end of the third section (see Proposition
3.11).
Let us end this section with some remarks on 2-cocycles.
Definition 1.19. ([12]) Let M, ∆M be a von Neumann bialgebra. A unitary element Ω M ¯bM is
called a unitary 2-cocycle if Ω satisfies the following identity, called the 2-cocycle identity:
Ω b 1∆M b ιΩ (cid:16) 1 b Ωι b ∆M Ω.
Example 1.20. Let M, ∆M be a compact Woronowicz algebra, and Ω a unitary 2-cocycle for
M, ∆M . Then if we put L 2
M , N (cid:16) M and
N (cid:16) L 2
∆N x (cid:16) Ω∆M x,
for all x M,
the couple N, ∆N is a Galois co-object for M, ∆M , called the Galois co-object associated to Ω.
It is easy to see that if Ω1 and Ω2 are two unitary 2-cocycles for M, ∆M , then their associated Galois
co-objects are isomorphic iff the unitary 2-cocycles are coboundary equivalent, that is, iff there exists
a unitary u M such that
Ω2 (cid:16) u
b u
Ω1∆M u.
In particular, the Galois co-object associated to a 2-cocycle Ω on M, ∆M is isomorphic to M, ∆M
as a right Galois co-object iff the 2-cocycle is a coboundary, i.e. is coboundary equivalent to 1 b 1.
Definition 1.21. Let M, ∆M be a compact von Neumann algebra, and N, ∆N a Galois co-object
for M, ∆M . Then N, ∆N is called cleft if there exists a unitary 2-cocycle Ω for M, ∆M such
that N, ∆N is isomorphic to the Galois co-object associated to Ω.
At the moment, we do not have any examples of non-cleft Galois co-objects for compact Woronowicz
algebras, although these do exist in the non-compact case. For example, in [4], non-cleft Galois co-
objects were (implicitly) constructed for discrete Woronowicz algebras (see Definition 2.1), the Galois
co-object being an l8-direct sum of rectangular matrix blocks. For commutative compact Woronowicz
algebras, that is, those arising from compact groups, it can be proven that all Galois co-objects are
necessarily cleft (that is, arise from a unitary (measurable) 2-cocycle function on the compact group).
We will later prove that this is also the case for co-commutative compact Woronowicz algebras (i.e.
group von Neumann algebras of discrete groups).
2 Galois objects for discrete Woronowicz algebras
In this section, we will make the connection with the theory of Galois objects from [10].
We first introduce the notion of the dual of a compact Woronowicz algebra.
Definition 2.1. Let M, ∆M be a compact Woronowicz algebra with regular left corepresentation W .
Define
xM (cid:16) ω b ιW ω M
σ-weak closure.
Then xM is a von Neumann algebra which can be endowed with a von Neumann bialgebra structure by
giving it the unique comultiplication ∆ xM such that
We will call the couple
xM , ∆ xM the discrete Woronowicz algebra dual to M, ∆M .
ι b ∆ xM W (cid:16) W13W12.
13
In fact, we had already introduced the notation xM in the remark after Proposition 1.10, as it can
be considered to be the space N in the special case where the right Galois co-object N, ∆N equals
M, ∆M . We then also remind that we had introduced some special notations for this case in the
Notation 1.14. The following proposition gathers some useful information which can be found in
the literature (for example, see the Remark 1.15 in [30], although we warn the reader that their
comultiplication on xM is opposite to ours).
Proposition 2.2. Let M, ∆M be a compact Woronowicz algebra, and
xM , ∆ xM its dual.
1. For all r IM , the number mr (cid:16) dimKr is finite.
2. There exists a left ∆M -invariant nsf weight ϕ xM on xM :
positive x
xM , we have
for all normal states on xM and all
A concrete formula for ϕ xM is given by
ϕ xM ω b ι∆ xM x (cid:16) ϕ xM x.
ϕ xM er,ij (cid:16) δi,jD1
r,j ,
for all r IM , 0 ¤ i, j (cid:160) mr.
3. On the other hand, define ψ xM to be the unique nsf weight on xM such that
ψ xM er,ij (cid:16) δi,j c2
r Dr,i,
where cr (cid:16) TrD1
1{2 (which is known as the quantum dimension of the irreducible corepre-
r
sentation corresponding to the index r IM ). Then ψ xM is right ∆ xM -invariant: for all normal
states on xM and all positive x
xM , we have
ψ xM ι b ω∆ xM x (cid:16) ψ xM x.
4. The Radon-Nikodym derivative between ψ xM and ϕ xM is given by the (possibly unbounded) positive,
non-singular operator
and δ xM is then a group-like element: for all t R, we have
δ xM (cid:16)
rD2
c2
r ,
rIM
∆ xM δit
xM (cid:16) δit
xM b δit
.
xM
The purpose of this section is to show that for an arbitrary Galois co-object N, ∆N , the comultipli-
xM , ∆ xM on N . This coaction then shares many
cation ∆ xM can be generalized to a coaction α N of
properties with the actual comultiplication ∆ xM .
For the rest of this section, we again fix a compact Woronowicz algebra M, ∆M and a right Galois
co-object N, ∆N for it. We keep using the notation from the previous section.
Notation 2.3. We denote by ϕ N the nsf weight on N which is uniquely determined by the fact that
all er,ij Mϕ
, with
xN
ϕ N er,ij (cid:16) δi,j T 1
r,j ,
where the Tr,j were introduced in Notation 1.14. We will then take the GNS-construction for ϕ N also
inside À
Hr b Hr, the GNS-map Λ N of ϕ N being determined by
rIN
The same notation will be used when N, ∆N equals M, ∆M considered as a right Galois co-object
over itself, taking however into consideration the special notations from Notation 1.14.
Λ N er,ij (cid:16) T 1{2
r,j
er,i b er,j.
14
Remarks:
S1{2
1. The fact that there exists a unique nsf weight with the above properties requires in fact a small
technical argument (at least in case the Hr are not finite-dimensional). The main observa-
tions to make are the well-known fact that any nsf weight ψ on a type I-factor is of the form
TrS1{2
for some non-singular positive (possibly unbounded) operator S (see [27], Lemma
VIII.2.8), and the fact that if ξ is a vector with ψθξ,ξ (cid:160) 8 (where we recall that θξ,ξ is the rank
one operator associated to ξ), then ξ D S1{2
(cid:16) ψθξ,ξ (this can, for example,
be pieced together from the results in [27], section IX.3). With this information, it should then
be easy to verify that the nsf weight ϕ N in the previous notation is indeed well-defined and
uniquely determined.
2
with }S1{2ξ }
2. It is easy to check, using the orthogonality relations between the €Wr,ij, that for r IN and
0 ¤ i, j (cid:160) nr, we have
with
ϕM
€W
r,ij b ι
€W Nϕ
,
xN
Λ N ϕM
€W
r,ij b ι
€W (cid:16)
€W
r,ijξM .
Hence our identifications of L 2
duality is defined in the setting of (locally) compact quantum groups (see [20]).
N coincide with the 'usual' way in which Pontryagin
N and L 2
Proposition 2.4. Denote by α N the unital normal faithful -homomorphism
α N : N Ñ
N ¯bB L 2
M : x Ñ Σ€W x b 1
€W Σ.
Then α N has range in N ¯b
xM , and determines an ergodic coaction of
xM , ∆ xM on N .
Before giving the proof, we first state a lemma that we will need in the course of it.
Lemma 2.5. Take r IN arbitrary. Then the linear span of the set
€W
r,ija 0 ¤ i, j (cid:160) nr, a M
is σ-weakly dense in N .
Proof. Denote by N the σ-weak closure of the linear span of
clearly N „ N . Now choose r IN fixed, and take an x N . By Proposition 1.15.2, we have
r,ija 0 ¤ i, j (cid:160) nr, a M . Then
€W
nr1
x (cid:16)
€W
r,k0
€Wr,k0x,
k(cid:16)0
the sum converging σ-weakly. As €Wr,k0x M for all 0 ¤ k (cid:160) nr, the sum on the right hand side lies
in N . So also N „
N .
Proof (of Proposition 2.4). Take ω N op
for €W (Proposition 1.7.4), we easily get that
, and denote x (cid:16) ω b ι
€W . By the pentagonal identity
α N x (cid:16) ω b ι b ιW13€W12
N ¯b
xM ,
and by an application of the formula ι b ∆ xM W (cid:16) W13W12, we find
α N b ια N x (cid:16) ι b ∆ xM α N x.
15
As elements of the form x constitute a σ-weakly dense subspace of N by Proposition 1.10.2, we have
proven that α N is a well-defined coaction.
We now show that it is ergodic. Take an element x
we get
N satisfying α N x (cid:16) x b 1. Then for all y
N ,
x b 1AdR y (cid:16) AdR yx b 1,
where we recall that AdR y (cid:16) Σ€W 1 b y
for any coaction of a compact Woronowicz algebra, we find x Z
€W Σ. As ι b ωAdR y ω M
N , the center of N .
2
(cid:16)
N , a general fact
°
Write then x (cid:16)
Then as α N x (cid:16) x b 1, we have
rIN
xrpr, where r Ñ xr l8
IN and pr the r-th minimal central projection of N .
xι b ω
€W Σ (cid:16) ι b ω
€W Σx
for all ω B L 2
M , L 2
N
. If we take ω (cid:16) ω
aξM ,€W
r,ij ξM
for some r IN , 0 ¤ i, j (cid:160) nr and
a M , and apply both sides of the above equality to €W
r,kjξM for some 0 ¤ k (cid:160) nr, we get, by using
the orthogonality relations for the €Wr,ij, that
x€W
r,kiaξM (cid:16) xr €W
r,kiaξM .
As, with r a fixed element of IN , we have that
previous lemma, we get that xξ (cid:16) xrξ for all ξ L 2
€W
r,kia a M is σ-weakly dense in N , by the
N , so x is a scalar multiple of the unit.
Our next goal is to show that α N is nicely behaved with respect to the weight structure on
xM , ∆ xM .
Proposition 2.6. For all x
N and all normal positive states ω on N , we have
In particular, α N is an integrable coaction.
ϕ xM ω b ια N x (cid:16) ϕ N x.
Remark: The fact that α N is integrable means that there exists a σ-weakly dense subspace of N
consisting of elements x for which all expressions ω b ια N x with ω
N lie in Mϕ
.
xM
Proof. We first recall a small technical result from [29], Proposition 1.3. Namely, as α N is an ergodic
on N , determined by
coaction, there exists a (not necessarily semi-finite) normal faithful weight ϕ1
N
the following formula: for all x
N , we have
ϕ1
N x (cid:16) ϕ xM ω b ια N x,
where ω is any normal state on N . Our job then is to prove that ϕ1
Notation 2.3, it is enough to prove that the er,ij are in M
with
N (cid:16) ϕ N . By the remark after
ϕ1
xN
N er,ij (cid:16) δi,j T 1
r,j .
ϕ1
16
Take r, s IN and 0 ¤ i, j (cid:160) nr, 0 ¤ k, l (cid:160) ns. Then we compute, using the GNS-construction for
ϕ N from Notation 2.3 and the functionals ωs,kl introduced in Notation 1.14, that
ι b ωs,kl
€W Λ N er,ij
(cid:16) T 1
r,j
€Ws,lk€W
r,ijξM
mt1
(cid:16) T 1
r,j
(cid:16) T 1
r,j
(cid:16) T 1
r,j
tIM
m,n(cid:16)0
mt1
tIM
m,n(cid:16)0
mt1
tIM
m,n(cid:16)0
€Ws,lk€W
r,ijξM ,
x
1
}W
t,mnξM }
W
t,mnξM y
1
}W
t,mnξM }
W
t,mnξM
ϕM Wt,mn €Ws,lk €W
r,ij D1
t,n W
t,mnξM
ϕM Wt,mn €Ws,lk €W
r,ij Λ xM et,mn ,
the latter sums converging in norm.
On the other hand,
ωs,kl b ια N er,ij (cid:16) T 1
(cid:16) T 1
r,j ωs,kl b ια N ϕM
r,j ϕM
€W
r,ij b ωs,kl b ιW13€W12
€W
r,ij b ι
€W
(cid:16) T 1
r,j
mt1
tIM
m,n(cid:16)0
ϕM Wt,mn€Ws,lk€W
r,ij et,mn,
where the latter sum now converges in the strong topology.
As Λ xM is a strong-norm closed map from Nϕ
xM
to L 2
xM , it follows that
ωs,kl b ια N er,ij Nϕ
,
xM
with
Λ xM ωs,kl b ια N er,ij (cid:16) ι b ωs,kl
€W Λ N er,ij .
Again by closedness, these assertions remain true when ωs,kl is replaced by an arbitrary normal func-
tional on N .
Let now ξr,i,j (cid:16) er,i b er,j for r IN and 0 ¤ i, j (cid:160) nr. For any finite subset J0 of the set
J (cid:16) r, i, j r IN , 0 ¤ i, j (cid:160) nr , denote by PJ0 the orthogonal projection onto the linear span
N and x (cid:16) yy in the linear span of the er,ij.
of the ξn with n J0. Take an arbitrary state ω
We remark then that there exists a unit vector ξ L 2
N with ω (cid:16) ωξ,ξ (cid:16) x ξ, ξ y. We can now
compute, using the normality of our weights, that
ϕ xM ω b ια N x (cid:16) lim
J0 „
fin
J
ϕ xM ωξ,ξ b ια N y
PJ0 b 1α N y
(cid:16) lim
J0 „
J
fin
(cid:16) lim
J
J0 „
fin
nJ0
nJ0
ϕ xM ωξ,ξn b ια N y
ωξ,ξn b ια N y
}ι b ωξ,ξn
€W Λ N y}
2
(cid:16) ϕ N yy (cid:16) ϕ N x,
by the unitarity of €W . From this, it immediately follows that all er,ij are integrable for ϕ1
N
N (cid:16) ϕ N on the linear span of the er,ij. This then concludes the proof.
ϕ1
, and that
17
We have shown so far that α N is an integrable, ergodic coaction. The final property of α N is that a
certain isometry which can be constructed from α N is in fact a unitary.
Proposition 2.7. Take x, y Nϕ
. Then α N yx b 1 Nϕ
xN
xN bϕ
xM
, and
Λ N b Λ xM αyx b 1 (cid:16) Σ€W Σ Λ N x b Λ N y.
Proof. The claim concerning the square integrability of α N yx b 1 follows immediately from the
fact that α N is integrable, with ι b ϕ xM α N (cid:16) ϕ N . Moreover, we can then define an isometry
G : L 2
N b L 2
N Ñ L 2
M b L 2
N
inside B L 2
N , L 2
M ¯b
N such that precisely
Λ N b Λ xM αyx b 1 (cid:16) Σ GΣ Λ N x b Λ N y
for all x, y Nϕ
xN
. We need to show that G (cid:16)
€W .
However, it is easily seen that for all ω B L 2
N
and x Nϕ
, we will have
xN
ι b ω
GΛ N x (cid:16) Λ xM ω b ια N x.
By the computations made in the previous proposition, it follows that ι b ω
ι b ω
€W on the linear span of the er,i b er,j for r IN , 0 ¤ i, j (cid:160) nr, and hence G (cid:16)
€W .
G coincides with
The three propositions above immediately show the following.
Theorem 2.8. Let M, ∆M be a compact Woronowicz algebra, N, ∆N a right Galois co-object for
M, ∆M . Then the couple
N , α N makes N into a right Galois object for the discrete Woronowicz
algebra
xM , ∆ xM , with corresponding Galois unitary G (cid:16)
€W .
For the terminology 'Galois object', we refer the reader to [8] (where the notations N and N are inter-
changed). In fact, it is simply defined to be an integrable ergodic coaction for which the map G, as we
constructed it in the course of the proof the previous proposition, is a unitary. This map G is in general
called the Galois unitary associated with the Galois object, and as we saw in the previous proposition,
it coincides precisely with €W in case the Galois object is constructed from a Galois co-object for a
compact Woronowicz algebra. Galois objects can also be defined as being ergodic, semi-dual coactions
(see [29], Proposition 5.12 for the terminology, and the remark under Proposition 3.5 of [8] for the
connection). We further remark that Galois objects for compact Woronowicz algebras were treated
in [4] (where they are termed 'actions of full quantum multiplicity'), and for commutative compact
Woronowicz algebras, that is for ordinary compact groups, in [33] and [21] (where they are termed 'ac-
tions of full multiplicity'). For Galois objects in the Hopf algebra setting, we refer to the overview [26].
We may now use the results from [8], which we gather in the following theorem.
Theorem 2.9. Let N, ∆N be a right Galois co-object for a compact Woronowicz algebra M, ∆M ,
xM , ∆ xM .
and let
Then the following statements hold.
N , α N be the associated right Galois object for the dual discrete Woronowicz algebra
1. There exists an nsf weight ψ N on N , unique up to scaling with a positive constant, which is
α N -invariant: for all states ω
xM and all x
N , we have
ψ N ι b ωα N x (cid:16) ψ N x.
18
2. The Radon-Nikodym derivative δ N η N of ψ N with respect to ϕ N satisfies
α N δit
N (cid:16) δit
N b δit
xM
for all t R.
3. The Radon-Nikodym derivative δ N is σ
parametergroup associated to ϕ N :
ϕ
t
xN
-invariant, where the latter denotes the modular one-
ϕ
t
σ
xN
δis
N (cid:16) δis
N
for all s, t R.
Proof. See [8], Theorem 3.18, Proposition 3.15 and Lemma 3.17.
Notation 2.10. We denote by Ar the non-singular (possibly unbounded) positive operator prδ N
B Hr , so that
δ N (cid:16)
Ar.
rIN
By the third item in the previous theorem, the operator Ar strongly commutes with Tr. In particular,
this means that Ar is diagonalizable, and that we may choose our er,i Hr so that they are also
eigenvectors for the Ar. We then write Ar,i for the eigenvalue of Ar with respect to the eigenvector er,i.
Remark: Let A be the Hopf -algebra associated to M, ∆M , consisting of all elements x M
with ∆M x M d M , the algebraic tensor product. Then it is well-known that A is a σ-weakly
dense sub--algebra of M , closed under the modular automorphism group σϕM
of ϕM . Let Br be the
sub--algebra of B Hr consisting of all elements x B Hr with Adr
R x B Hr d A . Again,
it is well-known that Br is a σ-weakly dense sub--algebra of B Hr (it is the linear span of the
coefficients of the spectral subspaces associated to Adr
R , see for example [4]). Then the operators Ar,
introduced in the above notation, turn out to be determined, up to a scalar, by the formula
t
Ait
r xAit
r (cid:16) ι b ε σϕM
t
Adr
R x,
for all x Br,
where ε denotes the counit of A . This formula can be derived from the way in which δ N was con-
structed in [8]. Hence, up to multiplication with a non-singular (possibly unbounded) positive element
in the center of N , the operator δ N can be recovered from the knowledge of all the Adr
R .
3 Projective representations of compact quantum groups
Using the results from the first section, we can easily develop a Peter-Weyl theory for projective
representations of compact quantum groups. We will in the following use again the notation which
we introduced in the first section.
We first define the notion of a projective representation relative to a fixed Galois co-object.
Definition 3.1. Let M, ∆M be a compact Woronowicz algebra, N, ∆N a right Galois co-object for
M, ∆M . A (left) N, ∆N -corepresentation of M, ∆M on a Hilbert space H consists of a unitary
map G N ¯bB H such that
∆N b ιG (cid:16) G13G23.
We call a Hilbert subspace K „ H invariant w.r.t. the N, ∆N -corepresentation when G restricts to
a unitary in N ¯bB K .
19
We call G irreducible if the only invariant Hilbert subspaces are 0 and H , and indecomposable when
H can not be written as the direct sum of two non-zero invariant subspaces.
We call two N, ∆N -corepresentations G1 and G2 on respective Hilbert spaces H1 and H2 unitarily
equivalent if there exists u B H1, H2 such that
G2 1 b u (cid:16) 1 b uG1.
Remark: When N, ∆N comes from a 2-cocycle Ω for M, ∆M , we will also simply speak of Ω-
corepresentations.
Theorem 3.2. Let M, ∆M be a compact quantum group, N, ∆N a right Galois co-object for
M, ∆M . Denote by V the right regular N, ∆N -corepresentation for M, ∆M , and let
Vr (cid:16) pr b 1
V N ¯bB Hr
be the components of V , where the pr denote the minimal projections of Z
N .
1. The unitaries Σ VrΣ are indecomposable left N, ∆N -corepresentations on the Hilbert spaces Hr.
2. Any indecomposable N, ∆N -corepresentation is unitarily equivalent with a unique Σ VrΣ.
3. Any N, ∆N -corepresentation splits as a direct sum of indecomposable N, ∆N -corepresentations.
4. Any indecomposable N, ∆N -corepresentation is irreducible.
Proof. As ι b ∆N
corepresentations. By Lemma 1.17.3, the space
V (cid:16)
V12 V13, we immediately get that the unitaries Σ VrΣ are left N, ∆N -
ι b ω
Vr ω B L 2
M , L 2
N
equals the whole of B Hr , from which it immediately follows that Σ VrΣ is indecomposable, and even
irreducible.
For the second statement, we use that the linear span of the matrix entries of the Vr's span a
norm-dense subset of L 2
N when applied to ξM . Hence, if G is an indecomposable left N, ∆N -
corepresentation of M, ∆M on a Hilbert space H , there must exist some r IN and an x
B H , Hr such that
ϕM b ιΣ VrΣ
1 b xG (cid:24) 0.
As in the proof of Theorem 1.15, this forces a scalar multiple of this expression to be a unitary in-
tertwiner, proving that G is isomorphic to Σ VrΣ. By the orthogonality relations between the Σ VrΣ,
these are all pairwise non-isomorphic. Hence the above r for G is uniquely determined.
To prove the third statement, consider the normal faithful unital -homomorphism
α : B H Ñ M ¯bB H : x Ñ G
1 b xG.
α, the set of fixed points for α. As
As ∆N b ιG (cid:16) G13G23, we see that α is a coaction. Let Z (cid:16) B H
in the proof of Proposition 1.10, we have that Z is the range of a normal conditional expectation on
B H . Hence Z is a von Neumann algebraic direct sum of type I-factors. Let then A be an atomic
maximal abelian von Neumann subalgebra of Z, and denote by pi the set of minimal projections in
A. Then it is clear that each piH is a fixed subspace for G. The spaces piH must further be inde-
composable: for if not, then we could find a pi and a non-zero projection p in B H with p strictly
smaller than pi and both pH and pi pH invariant under G. This would imply that p is a fixed
20
element for α, commuting with all x A. Hence p A by maximal abelianness. As pi was a minimal
projection in A, this gives a contradiction.
As for the fourth point, we may take our indecomposable N, ∆N -corepresentation to be some Σ VrΣ,
for which we have already proven irreducibility in the proof of the first point.
Corollary 3.3. Let N, ∆N be a right Galois co-object for a compact Woronowicz algebra M, ∆M .
Let Gi, i I, be a maximal set of non-isomorphic irreducible N, ∆N -corepresentations on Hilbert
spaces Hi. Then I (cid:21) IN , and N (cid:21) `iI B Hi .
Proof. This follows immediately from the second point of the previous proposition.
We can now pass to projective representations without reference to a fixed Galois co-object.
Definition 3.4. Let M, ∆M be a compact Woronowicz algebra. A projective (left) corepresentation
of M, ∆M on a Hilbert space H consists of a left coaction α of M, ∆M on B H ,
α : B H Ñ M ¯bB H .
Remarks:
1. Interpreting M, ∆M as the space of L 8-functions on some 'compact quantum group' G,
the above then corresponds to having a (necessarily continuous) action of G on B H . As
AutB H (cid:21) U H {S1, this indeed captures the notion of a projective representation when
G is an actual compact group.
2. One similarly has the notion of a projective right corepresentation of M, ∆M , for which we
replace the left coaction α above by a right coaction. For example, the coactions AdR and Adr
R
from the first section are then projective right corepresentations. At the end of this section, we
will show one can pass from left to right projective representations, so that one may essentially
restrict oneself to the study of projective left corepresentations, as we will do.
3. Some results on (special) coactions of compact Kac algebras on type I factors appear in [22].
In [10], we proved that from any projective corepresentation α, one can construct a Galois co-object
N, ∆N together with an N, ∆N -corepresentation G 'implementing' α. We will state the proposition
and give a sketch of the proof. For the full proof, we refer the reader to [10].
Proposition 3.5. Let M, ∆M be a compact Woronowicz algebra, H a Hilbert space, and α a
projective corepresentation of M, ∆M on H . Then there exists a right Galois co-object N, ∆N for
M, ∆M , together with an N, ∆N -corepresentation G of M, ∆M on H , such that
G
1 b xG (cid:16) αx,
for all x B H .
Sketch of proof. Choose a basis ei iI of H , and fix an element 0 I. We can then consider the
Hilbert space L 2
M b H . We can construct a unitary
N (cid:16) αe00 L 2
G : L 2
M b H Ñ L 2
N b H : ξ Ñ
αe0i ξ b ei.
iI
Denote by Gij the i, j-th component of G. Then Gij is an operator from L 2
M to L 2
N . We define
N (cid:16) Gij m i, j I, m M
σ-weakly closed linear span.
21
It is then possible to construct a map ∆N : N Ñ N ¯bN , uniquely determined by the properties that
∆N b ιG (cid:16) G
13
G
23
(where we have added brackets in the leg numbering notation to distinguish them from the indices
for matrix coefficients of G) and
∆N xy (cid:16) ∆N x∆M y,
for all x N, y M.
One proves that N, ∆N is a Galois co-object for M, ∆M , and then it immediately follows from the
above property that G is a left N, ∆N -corepresentation of M, ∆M on H . Finally, one proves that
G
1 b xG (cid:16) αx by direct computation.
Definition 3.6. Let α be a projective corepresentation of a compact Woronowicz algebra M, ∆M
on a Hilbert space H . Denote by N, ∆N the Galois co-object constructed from α as in the above
proposition, and denote by N, ∆N its isomorphism class. Then we say that α is an N, ∆N -
corepresentation.
N , ∆ N -corepresentation on H implementing α, then necessarily
It can be proven (see Proposition 3.4 in [10]) that if α is an N, ∆N -corepresentation of M, ∆M
on a Hilbert space H , and if there exists a Galois co-object
N , ∆ N for M, ∆M which possesses
an
N , ∆ N (cid:16) N, ∆N . One
may regard the isomorphism class of such a Galois co-object as a generalization of the notion of a
2-cohomology class. Also remark that if G N ¯bB H is a projective N, ∆N -corepresentation of
M, ∆M on a Hilbert space H , then the associated projective corepresentation
α : B H Ñ M ¯bB H : x Ñ G
1 b xG
is an N, ∆N -corepresentation by the above uniqueness result.
If N, ∆N is a Galois co-object for a compact Woronowicz algebra M, ∆M , and an N, ∆N -
corepresentation α for M, ∆M is given, then it is in general not true that all N, ∆N -corepresenta-
tions implementing α are isomorphic: consider for example ordinary one-dimensional representations.
We will see a further instance of this in the final section.
Definition 3.7. Let M, ∆M be a compact Woronowicz algebra, α a projective corepresentation of
M, ∆M on a Hilbert space H . We say that α is an irreducible projective corepresentation if α is
ergodic.
Proposition 3.8. Let N, ∆N be a right Galois co-object for a compact Woronowicz algebra M, ∆M ,
let α be an N, ∆N -corepresentation of M, ∆M on a Hilbert space H , and let G be an N, ∆N -
corepresentation implementing α.
Then there is a one-to-one correspondence between the set of α-fixed self-adjoint projections in B H
and the G-invariant subspaces K of H , given by the correspondence
p Ñ K (cid:16) pH .
In particular, α is irreducible iff G is irreducible.
Proof. By assumption, we have that
αx (cid:16) G
1 b xG
for all x B H .
22
So if p is a self-adjoint projection in B H with αp (cid:16) 1 b p, we have
and hence
G 1 b p (cid:16) 1 b pG,
G L 2
M b pH „ L 2
N b pH ,
which means pH is a G-invariant subspace.
Conversely, if K is a G-invariant subspace, and p the projection onto K , then also K is G-invariant
by Theorem 3.2. Hence we have
and so αp (cid:16) 1 b p, i.e. p is an α-fixed projection.
G 1 b p (cid:16) 1 b pG,
Remark: We in fact already used the above argument in the course of proving Theorem 3.2.3.
Our next proposition shows how projective corepresentations and ordinary corepresentations mesh
together.
Proposition 3.9. Let N, ∆N be a right Galois co-object for a compact Woronowicz algebra M, ∆M .
Let G be a projective left N, ∆N -corepresentation on a Hilbert space H , and let U be an ordinary
left corepresentation of M, ∆M on a Hilbert space K .
Then the following statements hold.
1. The unitary G12U13 N ¯bB H b K is a unitary N, ∆N -corepresentation on H b K .
2. If both G and U are irreducible, then G12U13 is a finite direct sum of irreducible N, ∆N -
corepresentations.
Proof. The fact that G12U13 is a unitary N, ∆N -corepresentation is trivial to verify. If further G
and U are irreducible, then we know already that G12U13 (cid:16) `iJ Gj for a certain set Gj of irreducible
N, ∆N -corepresentations indexed by a parameter set J. We have to prove that J is finite.
Let α be the projective corepresentation associated to G, so
αx (cid:16) G
1 b xG,
x B H .
By the previous proposition, we know that α is ergodic. Let B be the linear span of the spectral
subspaces inside B H , which is a σ-weakly dense sub--algebra of B H (see the remark following
Notation 2.10). If we then denote by Ur, r IM , a total set of representatives for the irreducible
corepresentations of M, ∆M on Hilbert spaces Kr, we know by [5] that
B, α (cid:21) `rIM
Kr b Ckr , `rIM Ur b 1
as a comodule over the Hopf algebra A „ M , where kr (cid:160) 8.
Now if β is the projective corepresentation associated to G12U13, then
β x (cid:16) U
13 α b ιxU13,
for all x B H b B K .
Hence if B is the linear span of the spectral subspaces of β, then as a comodule, we have
B (cid:21) U c
B U,
23
where U c denotes the contragredient of U and where we denote by the tensor product of corep-
resentations/comodules. But this means that the trivial corepresentation appears in B with multi-
grkr, where gr is the multiplicity of Ur „ U U c. Hence the fixed point algebra of β is
plicity °
finite-dimensional, and by the previous Proposition, J will have as its cardinality the dimension of a
maximal abelian subalgebra of the fixed point algebra of β. Hence J is finite.
The previous proposition leads to the following considerations. Let N, ∆N be a fixed right Galois
co-object for a compact Woronowicz algebra M, ∆M . Then we can make a W-category D by
considering as objects the N, ∆N -corepresentations which are (isomorphic to) finite direct sums of
irreducible N, ∆N -corepresentations, and as morphisms bounded intertwiners. Then if we denote
by C the tensor W-category of finite-dimensional M, ∆M -corepresentations, we can make D into
a right C -module by the natural composition introduced above:
D C Ñ D : G, U Ñ G12U13,
while the action of morphisms is simply by tensoring. We can then also turn FN :(cid:16) `rIN
module over the fusion ring FM :(cid:16) `rIM
construction.
Z into a
Z by means of the fusion rules associated to this categorical
But in fact, there is another different way in which to obtain these fusion rules, making use of the
theory developed in [28]. In that paper, Wassermann's multiplicity theory for ergodic compact Lie
group actions on C-and von Neumann algebras is extended to the setting of compact Woronowicz
algebras. Although the paper works in the C-algebraic realm and uses right coactions, the results
also apply in the von Neumann algebra setting and with left actions, and we make the transition
without further comment in explaining these ideas.
Let then M, ∆M be a compact Woronowicz algebra with an ergodic coaction α on a von Neumann
algebra A. It is well-known that the crossed product
M (cid:10) A (cid:16) x b 1αy x M, y A
2
„ B L 2
M b L 2
A
is a von Neumann algebraic direct sum of type I factors. Let Iα be the set of atoms of the center
of M (cid:10) A, and let Fα be the free abelian group generated by Iα. Then one can turn Fα into a right
FM -module by the following procedure. Let ps s Iα be the set of minimal projections in the
center of M (cid:10) A, and choose for each s Iα a minimal projection es ¤ ps in M (cid:10) A. We can equip
the corners es B L 2
M ¯bAet with a left M, ∆M -coaction αst by the formula
αst z (cid:16) Σ b 1V
12 ι b αz V12 Σ b 1.
For each r IM , s, t Iα, define M r
to be the dimension of the set of M, ∆M -intertwiners between
the corepresentation Ur associated to r and αst. Then the action of r IM on an element t Iα is
defined as
st
t r :(cid:16)
sIα
M r
ts
s.
Let now N, ∆N be a right Galois co-object for the compact Woronowicz algebra M, ∆M . Choose
r IN . Then we can apply the above ideas to the left coaction αr on B Hr , where αr is the coac-
tion associated to the irreducible projective N, ∆N -corepresentation Gr pertaining to r. We claim
that the resulting right FM -module is independent of the choice of r, and coincides precisely with the
right FM -module as constructed after Proposition 3.9. We will briefly indicate how this can be proven.
N ¯bB Hr Gr. Indeed, this follows by the characterization
We first observe that M (cid:10)B Hr equals G
r
of N as a fixed point space, by the pentagon identity for V (and the related pentagon identity for
24
Gr), by the fact that Gr implements αr, and finally by the characterization of M (cid:10) B Hr as the set
of elements z in B L 2
M ¯bB Hr satisfying
V12z13V
12 (cid:16) ι b αr z .
Hence we already see that Iαr (cid:16) IN .
Choose then for each s IN the element
es :(cid:16) es,00 b er,00
N ¯bB Hr (cid:21) M (cid:10) B Hr
M ¯bB Hr et is isomorphic to es,00B L 2
as a minimal projection. Then by transporting all structure with the aid of G, one sees that the corner
es B L 2
N et,00, equipped with the restriction of the
coaction AdL (which appears in the proof of Proposition 1.10). This may further be simplified to the
coaction
αst : B Hs, Ht Ñ M ¯bB Hs, Ht : x Ñ G
s 1 b xGt.
This final coaction may be interpreted as corresponding to the (ordinary) corepresentation 'Gc
s Gt'.
A Frobenius-type argument then shows that this corepresentation contains some Uu with u IM as
much times as Gt is contained in Gs Uu. This shows that the two mentioned fusion rules indeed
coincide, and ends our sketch of proof.
To end this section, let us come back to comparing the structures of N, ∆N and N, ∆op
started in Lemma 1.18. We begin by introducing a certain antipode on a subspace of N .
N which we
Proposition 3.10. Let N, ∆N be a right Galois co-object for a compact quantum group M, ∆M .
Denote by N the linear span of the Vr,ij in N (see Lemma 1.17). Denote by A the corresponding
subspace of M , which coincides with the Hopf -algebra associated with M, ∆M (see the remark after
Notation 2.10). Then the following statements hold.
1. The space N is a right A -module.
2. If we define the anti-linear map
SN
: N Ñ N : Vr,ij Ñ
Vr,ji,
then for all x N and y A , we have
SN xy
(cid:16) SN x
SM y
,
where SM denotes the antipode of the Hopf -algebra A .
3. For all r IN and 0 ¤ i, j (cid:160) nr, we have
ΛN SN
Vr,ij
(cid:16)
(cid:10)1{2
Tr,j
Tr,i
J N ΛN
Vr,ij ,
where J N denotes the modular conjugation for N, given by er,i b er,j Ñ er,j b er,i.
Proof. As the Vr,ij form a basis of N , it is easy to see that SN
is well-defined. Moreover, using
the formulas in Lemma 1.17.1, the third statement follows immediately. As for the first point, this is
an immediate consequence of Proposition 3.9.2.
So the only thing left to show is the second statement, which is at least meaningful by the first part
of the proposition.
25
on N 1. Let us call a normal
For ω in the predual of N 1, denote by ω the normal functional x Ñ ωx
functional on N 1 elementary when it is of the form er,ij Ñ ωer,ji for some ω in the linear span of
the ωs,kl (see Notation 1.14). Then with ω elementary, we immediately obtain the formula
Choose now normal functionals ω1 and ω2 on respectively B L 2
M which restrict
to elementary functionals on respectively N 1 and xM 1. Then by the pentagonal identity for V , we have
N and B L 2
SN ω b ι
V
(cid:16) ω b ι
V .
(6)
where ω is the functional
ω1 b ι
V ω2 b ιV (cid:16) ω b ι
V ,
x B L 2
N Ñ ω1 b ω2
V
12 1 b x
V12 .
By the first part of the proposition, the restriction of ω to N 1 is again elementary. Combining these
is indeed right
statements with equation (6) (and the corresponding one for SM ), we see that SN
SM
-linear.
We can now make the connection between the adjoint coactions of M, ∆M on N and M, ∆op
M on
N 1 respectively (see the remark after Lemma 1.18). Let us first recall that any compact Woronowicz
algebra M, ∆M is endowed with an involutive anti-comultiplicative anti--automorphism RM , given
by the formula
RM x (cid:16) J xM xJ xM
for all x M.
More concretely, we have RM Wr,ij (cid:16) Vr,ji for all r IM and 0 ¤ i, j (cid:160) mr. We will also denote
and use the same notation for its inverse.
C
N : N Ñ
N 1 : x Ñ J N xJ N (cid:16) x,
Proposition 3.11. Let N, ∆N be a right Galois co-object for a compact Woronowicz algebra M, ∆M .
Let N, ∆op
of M, ∆op
N be the co-opposite right Galois co-object for M, ∆op
M on N cop
M . Then the right adjoint coaction
N 1 is given by
^
(cid:16)
x Ñ CN b RM AdR CN x.
Proof. Note that the right adjoint coaction on B Hr is, by its definition and by Lemma 1.18.1, given
as
x Ñ
V
x b 1
V .
Denote now by J N the modular conjugation for N , which we recall is simply the map
L 2
N Ñ L 2
N : er,i b er,j Ñ er,j b er,j .
It is then easily seen that the proposition follows one we can prove the following identity:
V (cid:16) J N b J N Σ€W ΣJ N b J xM .
(7)
Now piecewise, the identity (7) corresponds to the identities
J N
€W
r,jiJ xM (cid:16)
Vr,ij,
for all r IN , 0 ¤ i, j (cid:160) nr.
But in [8], it was proven that J N xJ xM N for x N (see the remark just before Lemma 4.3 in that
paper). Hence we only have to check if
J N
€W
r,jiJ xM
ξM (cid:16)
Vr,ij ξM , for all r IN , 0 ¤ i, j (cid:160) nr.
This now follows from an easy computation using Lemma 1.17.
26
Remark: It seems nicer to treat the right adjoint M, ∆op
M -coaction on N 1 as a left M, ∆M -coaction:
AdL : N 1
Ñ M ¯b
N 1 : x Ñ Σ V
x b 1
V Σ.
These then localize to left adjoint coactions Adr
L on the B Hr . Note that the map AdL (as well as
the map SN ) in fact already appeared in the proof of Proposition 1.10, and that the Adr
L are nothing
but the N, ∆N -corepresentations of M, ∆M associated to the N, ∆N -corepresentations Σ VrΣ
from Theorem 3.2.
4 Reflecting a compact Woronowicz algebra across a Ga-
lois co-object
In this section, we will consider in the special case of compact Woronowicz algebras a technique which
was introduced in [10]. The following theorem was proven in [10], Proposition 2.1 and Theorem 0.7.
Theorem 4.1. Let M, ∆M be a compact Woronowicz algebra, and N, ∆N a right Galois co-object
for M, ∆M . Denote by P „ B L 2
N the von Neumann algebra which is generated by elements of
the form xy, where x, y N . Then P can be made into a von Neumann bialgebra, the comultiplication
∆P being uniquely determined by the fact that
∆P xy
(cid:16) ∆N x∆N y
for all x, y N.
The von Neumann bialgebra P, ∆P furthermore admits (not necessarily finite) left and right ∆P -
invariant nsf weights (i.e. is a von Neumann algebraic quantum group in the terminology of [20]).
The following theorem gives a concrete formula for the above invariant weights. We will use the
notations introduced in the first three sections (see Notation 1.14 and Notation 2.10).
Theorem 4.2. Let M, ∆M be a compact Woronowicz algebra, N, ∆N a right Galois co-object for
M, ∆M , and P, ∆P the von Neumann bialgebra introduced in Theorem 4.1. Then, up to a positive
scalar, the left invariant nsf weight ϕP satisfies
with
Vr,ij V
s,kl MϕP
ϕP
Vr,ij V
s,kl (cid:16) δr,sδi,kδj,l
Tr,j
Ar,j
,
while the right invariant nsf weight ψP satisfies, again up to a positive scalar,
and
€W
r,ij €Ws,kl MψP
ψP
€W
r,ij €Ws,kl (cid:16) δr,sδi,kδj,l
Tr,i
Ar,i
.
Proof. For the proof of the theorem, we have to explain first how the invariant weights ϕP and ψP
can be obtained. This goes back to Theorem 4.8 of [8].
Denote by ∇it
N,M the following one-parametergroup of unitaries on L 2
N :
∇it
N,M (cid:16) ∇it
N
J N δit
N
J N ,
27
where ∇
parametergroup is concretely given as
N is the modular operator associated the weight ϕ N on N . On basis vectors, this one-
∇it
N,M er,i b er,j (cid:16) Ait
r,j T it
r,j T it
r,i er,i b er,j.
We can then implement on N a one-parametergroup σN,M
t
, determined by the formula
σN,M
t
x (cid:16) ∇it
N,M x∇it
M ,
for all x N,
where ∇M is the modular operator on L 2
by using the commutation relation
M associated to ϕM . One verifies that this is well-defined
∇it
M b 1
€W ∇it
N,M b 1 (cid:16) 1 b ∇it
N
€W 1 b ∇it
N
δit
N ,
N ,
proven in Proposition 3.20 of [8] (we remark that the P -operator introduced there coincides with ∇
as the modular element of M, ∆M is trivial). This commutation relation also immediately shows
that
σN,M
t
Vr,ij (cid:16) T it
r,i T it
r,jAit
r,j
Vr,ij ,
using the notation from Lemma 1.17.1.
Now by construction (see the discussion preceding Lemma 4.4 in [8]), all elements x N which are
analytic with respect to σN,M
will lie in the space of square integrable elements of ϕP , by the formula
t
ϕP xx
(cid:16) ϕM σN,M
i{2 x
σN,M
i{2 x.
By polarity, we get for x, y N analytic w.r.t. σN,M
t
that
ϕP xy
(cid:16) ϕM σN,M
i{2 y
σN,M
i{2 x.
Applying this to x (cid:16)
immediately get the first formula in the statement of the theorem.
Vr,ij and y (cid:16)
Vs,kl, and using the orthogonality relations between the Vr,ij, we
For the second formula, we can use the expression ψP (cid:16) ϕP RP , where RP was an involutory
anti-automorphism on P determined by the formula
RP x (cid:16) J N xJ N ,
for all x P
(see Lemma 4.3 in [8]). In fact, the discussion before that lemma states that, for x, y N , we have
RP xy
RN x, where
(cid:16) RN y
RN : N Ñ N op : x Ñ J xM xJ N .
By applying RN
Vr,ij
to ξM , we find that
RN
Vr,ij (cid:16)
€Wr,ji
(see also the proof of Proposition 3.11). Applying then ϕP RP to €W
of the proof, the expression for ψP as in the statement of the theorem follows.
r,ij
€Ws,kl and using the first part
From the formulas in Theorem 4.2, we can draw the following conclusions.
28
Proposition 4.3. Let M, ∆M , N, ∆N and P, ∆P be as in the foregoing theorem.
1. If P, ∆P is a compact Woronowicz algebra (that is, if ϕP and ψP are finite), then all nr (cid:160) 8,
i.e. all irreducible N, ∆N -corepresentations of M, ∆M are finite-dimensional.
2. Conversely, if one of the irreducible N, ∆N -corepresentations for M, ∆M is finite-dimensional,
then they all are, and then P, ∆P is a compact Woronowicz algebra.
3. If P, ∆P is unimodular (that is, if ϕP is a multiple of ψP ), then there exist positive numbers
dr such that Ar (cid:16) d2
rT 2
r (where the Ar were introduced in Notation 2.10).
Remark:
If the condition in the third point is satisfied, then one could interpret dr as a (finite!)
relative quantum dimension of the irreducible N, ∆N -corepresentation corresponding to r, in anal-
ogy with the case of ordinary irreducible corepresentations (compare with Proposition 2.2). Here the
relativity refers to one irreducible N, ∆N -corepresentation w.r.t. another, as δ N is only determined
up to a positive scalar. We note that we do not know of any particular examples where P, ∆P is
not unimodular, so it could well be that this condition is always fulfilled.
Proof. The third point follows immediately from the formulas in the previous theorem combined with
Lemma 1.17, as there then exists a positive number c ¡ 0 such that
Ar,j
Ar,i
(cid:16) c
T 2
r,j
T 2
r,i
.
If P, ∆P is moreover compact, then for any r IN and 0 ¤ i (cid:160) nr, we get, by using the unitary of
€Wr (and the normality of ψP ), that
nr 1
ψP 1 (cid:16) ψP
€W
r,ij €Wr,ij
i(cid:16)0
(cid:16)
(cid:16)
nr 1
i(cid:16)0
ψP
€W
r,ij €Wr,ij
1
d2
r
nr 1
i(cid:16)0
1
Ti,r
(cid:160) 8.
As the Tr,i are summable, the final sum must necessarily be finite, i.e. nr (cid:160) 8.
Finally, suppose that M, ∆M has a finite-dimensional irreducible N, ∆N -corepresentation, say
corresponding to the index value r. Then as ψP 1 (cid:16) ψP
€Wr,ij , we see that ψP is finite,
and hence P, ∆P is a compact Woronowicz algebra. By the second point, also all other irreducible
N, ∆N -corepresentations of M, ∆M are finite-dimensional.
°nr 1
i(cid:16)0
€W
r,ij
Remark: In case the irreducible N, ∆N -corepresentations are finite-dimensional, the linear span of
r,ij generates inside N a purely algebraic Galois co-object N for the Hopf algebra A inside
the €W
M, ∆M . Conversely, if one starts with a Galois co-object for A , satisfying some suitable relations
with the -structure, we can in essence develop the whole theory so far in an algebraic way, and then
necessarily the reflection will correspond to a compact quantum group (this was essentially already
observed in [7], see also [32]). As it turns out, there do exist interesting Galois co-objects which are of
a non-algebraic type (see the final section), which was part of the motivation for writing this paper.
29
5 Galois co-objects and projective corepresentations for
compact Kac algebras
In this short section, we show that when one deals with compact Kac algebras (see Definition 1.1),
one is always in the algebraic setup (see the remark at the end of the previous section).
Proposition 5.1. Let N, ∆N be a Galois co-object for a compact Kac algebra M, ∆M , and let
P, ∆P be the reflected von Neumann bialgebra as obtained in Theorem 4.1. Then also P, ∆P is a
compact Kac algebra.
Proof. As we recalled in Theorem 2.9, the modular element δit
. However,
N
for a compact Kac algebra, δ xM (cid:16) 1. By ergodicity of α N , we then conclude that we can take δ N (cid:16) 1.
satisfies α N δit
N (cid:16) δit
N b δit
xM
From Theorem 4.2 and the normality of ψP , we then find
nr 1
ψP 1 (cid:16) ψP
€W
r,ij €Wr,ij
i(cid:16)0
(cid:16)
(cid:16)
nr 1
i(cid:16)0
nr 1
i(cid:16)0
ψP
€W
r,ij €Wr,ij
Ti,r (cid:16) 1,
so that ψP is finite, and P, ∆P thus a compact Woronowicz algebra.
But then P, ∆P is in particular unimodular, so that the third point of Proposition 4.3 gives us that Tr
is a scalar matrix, and hence just 1
times the unit matrix on Hr. This shows that the one-parameter-
nr
group ∇it
σN,M
t
N,M , which we introduced in the course of the proof of Theorem 4.2, is trivial. As σϕP
for all x, y N (which follows from the fact that σN,M
is actually the restriction
xσN,M
xy
t
y
t
t
(cid:16)
to N of the modular automorphism group of the balanced weight ϕP ` τM on
that σϕP
is trivial, and hence ϕP is a trace. This concludes the proof.
t
(cid:10)
N
P
N op M
), we get
Combining the previous proposition with Theorem 3.2 and Proposition 4.3, we obtain the following
corollary.
Corollary 5.2. Let M, ∆M be a compact Kac algebra. If H is a Hilbert space, and α : B H Ñ
M ¯bB H a coaction, then the following statements hold.
1. The trace on B H is α-invariant.
2. If α is ergodic, then H is finite-dimensional.
In particular, this says that the invariant state associated to an ergodic coaction of a compact Kac
algebra on a type I-factor is tracial. Note that this is not true for ergodic coactions of Kac algebras
on arbitrary von Neumann algebras (counterexamples can be found in [4]). It would be nice to have
a more direct proof of the above corollary, but we were not able to produce one.
30
6 Projective corepresentations of finite-dimensional Kac
algebras
In this section, we will briefly review what can be said concerning the situation of finite-dimensional
compact Woronowicz (and hence Kac) algebras.
Proposition 6.1. Let M, ∆M be a finite-dimensional Kac algebra, and N, ∆N a right Galois
co-object for M, ∆M . Then N is finite-dimensional with dimN (cid:16) dimM , and N, ∆N is cleft.
Proof. Choose r IN . Then as the right coaction Adr
R of M, ∆M on B Hr is ergodic, it is easy
to see immediately that nr (cid:16) dimHr is finite. Then take r IN fixed. As N is the σ-weak closure
of the set
r,ijm 0 ¤ i, j ¤ nr, m M by Lemma 2.5, we see that N is finite-dimensional. As €W
gives a unitary from L 2
N , necessarily dimN (cid:16) dimM .
M to L 2
€W
N b L 2
N b L 2
Now disregarding the -structures, we get that N, ∆N is a Galois co-object for the Hopf algebra
M, ∆M . It is then well-known that N, ∆N is cleft in this weaker form (see for example the remark
following Corollary 3.2.4 in [26]). But this means that N (cid:21) M as right M -modules. As M is a direct
sum of matrix algebras, it is easy to see that we can in fact find a unitary u : L 2
M
such that uN (cid:16) M . Hence we may identify L 2
M and N with M . We can then
consider Ω (cid:16) ∆N 1M . By right linearity of ∆N , we then get ∆N x (cid:16) Ω∆M x for all x N , and
by coassociativity of ∆N we have that Ω satisfies the 2-cocycle relation. Hence N, ∆N is cleft.
N with L 2
N Ñ L 2
Remark: Finite Galois co-objects (in the operator algebra context) have also been dealt with in the
papers [12], [31] and, in a more general setting, [17].
For later purposes, we also introduce the following definition of a non-degenerate 2-cocycle (see Defi-
nition 1.19 for the general notion of a unitary 2-cocycle).
Definition 6.2. Let M, ∆M be a finite-dimensional compact Kac algebra, and Ω a unitary 2-cocycle
for M, ∆M . We call Ω non-degenerate if the associated Galois object N is a (finite-dimensional)
type I-factor.
xM , ∆ xM , we can create from it a Galois co-object for
N , α N is then in fact also a projective
This terminology was introduced in [1]. One observes that, as
(right) corepresentation for
xM , ∆ xM , which
xM an implementing unitary 2-cocycle,
will then necessarily also be cleft. If we denote by Ω
Ω between cohomology
then Ω turns out to be non-degenerate again, and the correspondence Ω Ñ
classes of non-degenerate 2-cocycles of resp. M, ∆M and
xM , ∆ xM is a bijection. (For the proof of
this result, we refer again to [1].) To give a simple example, consider a finite abelian group G. Then
the bicharacter on G G given by evaluation gives a non-degenerate 2-cocycle function Ω on G
G
Ñ χh, and its dual is simply the same construction applied to the
by the formula g, χ, h, χ1
evaluation bicharacter on G G.
xM b
7 Projective corepresentations for cocommutative Kac
algebras
As a second special case, we will consider compact Woronowicz algebras M, ∆M which have a
cocommutative coproduct: ∆op
M (cid:16) ∆M . It is not so difficult to show that M, ∆M is then Kac, and
31
in fact that M (cid:16) L Γ for some (countable) discrete group Γ, the coproduct being given by
∆M λg (cid:16) λg b λg
for all g Γ,
where the λg denote the standard unitary generators in the group von Neumann algebra L Γ. We
will in the following denote L Γ as shorthand for L Γ, ∆L Γ
, and we denote the invariant trace
by τ . The dual discrete Woronowicz algebra
Γ,
equipped with the coproduct
xM , ∆ xM is then simply the function space l8
such that
∆ xM : l8
Γ Ñ l8
Γ ¯bl8
Γ (cid:21) l8
Γ Γ
∆ xM f g, h (cid:16) f gh
for all g, h Γ.
We will in the following write L 2
at the point g Γ, we have
L Γ (cid:16) l2
Γ of course, and then, with δg being the Dirac function
ΛL Γ
λg (cid:16) δg.
As group von Neumann algebras are in particular Kac algebras, we know from Corollary 5.2 that they
can only act ergodically on type I factors which are of finite type. Let us give a more immediate proof
of this fact in this particular case.
Lemma 7.1. Let Γ be a discrete group. Let B be a von Neumann algebra, and suppose that we have
given an ergodic coaction
α : B Ñ L Γ b B.
If we denote by φB the unique α-invariant state on B, then φB is a trace.
Proof. For each g Γ, denote Bg (cid:16) x B αx (cid:16) λg b x . As α is ergodic, it is easily seen that
each Bg is either zero- or one-dimensional. It is further immediate that Bg Bh „ Bgh for all g, h Γ,
g (cid:16) Bg1. Therefore, whenever Bg is not zero-dimensional, we may assume that Bg (cid:16) Cug
and that B
g (cid:16) ug1. When Bg (cid:16) 0, we will denote ug (cid:16) 0.
with ug a unitary. We may moreover assume that u
We claim that the linear span of the ug is σ-weakly dense in B. Indeed, if this was not the case, then
we could find a nonzero x B with φB xug (cid:16) 0 for all g Γ. But as φB (cid:16) τ b ια by definition,
this would imply that
τ b ιαxλg b 1ug (cid:16) 0
for all g Γ.
Now an easy computation shows that for all g Γ, we have
τ b ιαxλg b 1 B
g .
Hence we in fact have
This implies αx (cid:16) 0, and so x (cid:16) 0, which is a contradiction.
τ b ιαxλg b 1 (cid:16) 0
for all g Γ.
It is now enough to prove that φB uguh (cid:16) φB uhug for all g, h Γ. But the left hand side is a
multiple of φB ugh , which is zero in case g (cid:24) h1. Similarly, the right hand side is zero in case
g (cid:24) h1. As both sides equal 1 when g (cid:16) h1, we are done.
Remark: General coactions of group von Neumann algebras (or rather, of the associated group C-
algebras), have been studied in the theory of Fell bundles over groups (see for example [25]). The
intuitive connection between these notions is essentially contained the above proof.
32
Corollary 7.2. Let Γ be a discrete group, H a Hilbert space, and
α : B H Ñ L Γ ¯bB H
an ergodic coaction of L Γ on B H . Then the following statements hold.
1. The dimension of H is finite.
2. There exists a finite subgroup H of Γ such that
αB H „ L H b B H ,
and such that, denoting by β the coaction α with range restricted to L H b B H , the couple
B H , β is a (left) Galois object for L H .
Remark: The notion of a Galois object was introduced in the second section. In the finite-dimensional
setting, it may be defined as follows: let A be a finite-dimensional Hopf -algebra with a left coaction
β on a finite-dimensional -algebra B. Then B, β is called a left Galois object for A if the map
is a bijection.
B b B Ñ A b B : x b y Ñ β x1 b y
Proof. By the previous lemma, we know that B H admits a tracial state. Hence H must be finite.
As for the second point, this is rather a corollary of the proof of the previous proposition. For, using
the notation introduced there, denote by H the set of elements g in Γ for which ug (cid:24) 0. As the ug
are orthogonal to each other with respect to the α-invariant state, we must have that H is finite. As
ug uh is a non-zero multiple of ugh, and u
g (cid:16) ug1, we must have that H is a finite group. It is then
immediate that indeed αB H „ L H b B H .
The coaction β of L H on B H is then clearly also an ergodic action, with the ordinary (normal-
ized) trace tr as its invariant state. This implies that the map
L 2
B H , tr b L 2
B H , tr Ñ L 2
L H , τ b L 2
B H , tr : x b y Ñ β x1 b y
is isometric and thus injective. As the ug with g H form an orthonormal basis of B H , we have
that the order of H equals the square of the dimension of H . Hence a comparison of dimensions
shows that the above map is also surjective, which proves that B H , β is a left Galois object for
L H .
Proposition 7.3. Let Γ be a discrete group, and let N, ∆N be a right Galois co-object for L Γ.
Then there exists a finite subgroup H „ Γ and a non-degenerate 2-cocycle Ω for L H , such that
N, ∆N is isomorphic to the cleft Galois co-object induced by Ω (considered as a 2-cocycle for L Γ).
Proof. Let N, ∆N be a right Galois co-object for L Γ. Using the terminology introduced in
Definition 3.6, choose an irreducible N, ∆N -corepresentation
α : B H Ñ L Γ b B H
of L Γ on a Hilbert space H (for example, one of the Adr
L , see the end of the third section). By
the previous proposition, we know that H is finite-dimensional, and that we can choose a minimal
finite subgroup H „ Γ for which
αB H „ L H b B H .
33
We moreover know that the corresponding coaction β of L H on B H is then a Galois object.
This means that the Galois co-object NH , ∆NH which is associated to β (as a projective corep-
resentation) may be taken to be equal to L H , Ω∆L H
, where Ω L H b L H is a
non-degenerate unitary 2-cocycle (see the final remarks of the previous section). Denote further
N , ∆ N :(cid:16) L Γ, Ω∆L Γ
, which is a cleft Galois co-object for L Γ.
Let then G be a projective NH , ∆NH -corepresentation which implements β. As NH (cid:16) L H „
N (cid:16) L Γ, we may interpret G to be an element of N b B H . It is trivial to see that G is then
N , ∆ N as a
an
right Galois co-object for L Γ (see the remark after Definition 3.6), which proves the proposition.
N , ∆ N -corepresentation which implements α. Therefore N, ∆N is isomorphic to
Corollary 7.4. If Γ is a torsionless discrete group, or more generally, a group with no finite sub-
groups of square order, then any Galois co-object for L Γ, ∆L Γ
is isomorphic to L Γ, ∆L Γ
as a right Galois co-object.
In particular, any unitary 2-cocycle for L Γ is then a 2-coboundary (see the remark after Example
1.20).
Proof. The statement concerning torsionless discrete groups is of course an immediate consequence of
the previous proposition. As for the statement concerning the case when there are no finite subgroups
of square order, observe that if K is any finite group for which some B H allows a Galois object
structure for l8
K , then necessarily K (cid:16) dimH
2.
Remarks:
1. For finite groups, Proposition 7.3 was proven in [24] (see also [13]).
2. In [16], it is shown that for any group Γ, all quasi-symmetric 2-cocycles for L Γ, i.e. cocycles
which also satisfy ΣΩΣ (cid:16) ηΩ for some η S1, are coboundaries. (The authors weaken this to
allow unitaries satisfying the 2-cocycle condition up to a scalar, but it is possible to show that,
in any compact Woronowicz algebra, such unitaries are automatically 2-cocycles).
3. One can also easily describe the Galois objects dual to the Galois co-objects appearing in Propo-
sition 7.3. Namely, if we have given a discrete group Γ, a finite subgroup H and a non-degenerate
2-cocycle Ω for L H , let γ : B H Ñ B H b l8
H dual
to the Galois co-object associated with Ω. Then the dual of the Galois co-object for L Γ
associated to H, Ω is the induction of γ to Γ. The underlying von Neumann algebra of this
construction consists of the set of all elements x B H b l8
H be the Galois object for l8
Γ for which
γ b ιx (cid:16) ι b βl8
H
x,
where βl8
coaction α of l8
i.e. αx :(cid:16) ι b ∆l8
x.
Γ
H is the coaction associated to the left translation action of H on Γ. The right
Γ on this von Neumann algebra is then simply given by right translation,
The projective corepresentations associated to the Galois co-objects for L Γ can be determined as
follows.
Proposition 7.5. Let Γ be a discrete group with a finite subgroup H. Let Ω be a non-degenerate
2-cocycle for L H , and let G L H b B H be the associated irreducible Ω-corepresentation on
some Hilbert space H .
34
Then with N, ∆N the cleft Galois co-object for L Γ associated to Ω L ΓbL Γ, we have
IN (cid:21) H zΓ, and a maximal set of irreducible non-isomorphic N, ∆N -corepresentations is given by
the set
GHg :(cid:16) G λsHg
b 1 L Γ b B H ,
where s : H zΓ Ñ Γ is a fixed section for Γ Ñ H zΓ.
Remark: As Ω is assumed to be non-degenerate for L H , the unitary G is indeed the unique Ω-
corepresentation for L H , up to isomorphism.
Proof. It is immediately seen that the right regular N, ∆N -corepresentation for L Γ equals V (cid:16)
ΩV , while the left regular N op, ∆N op -corepresentation equals €W (cid:16) Ω
H W . For g Γ, let δHg
B l2
Γ be the indicator function for the coset Hg. Clearly, δHg commutes pointwise with L H .
Using then the definition of N as a fixed point set (see Proposition 1.10), it is easy to check that
N . Using the second definition of N as the closure of the right leg of €W (see again Proposition
δHg
1.10), we get that in fact δHg Z
N , the center of N .
We claim now that IN (cid:21) H zΓ, and that the VHg :(cid:16)
Indeed, as V (cid:16) `gH VHg, it is enough to show that each of the sets ι b ω
type I-factor. But denoting VH (cid:16)
V δHg b 1 are the irreducible components of V .
is a
hH δh b λh, an easy computation shows that
VHg ω L Γ
°
VHg (cid:16) ρg b 1ΩVH ρ
g b λg ,
where ρg is a right translation operator on l2
a non-degenerate 2-cocycle, we know that ι b ωΩVH ω L Γ
the claim.
Γ, i.e. ρgδk (cid:16) δkg for g, k Γ. As we assumed that Ω is
is a type I-factor. This proves
Now the irreducible Ω-corepresentation Σ VHgΣ of L Γ is immediately seen to be isomorphic to
the Ω-corepresentation GgH in the statement of the proposition, while VH is isomorphic to G as an
Ω-corepresentation for L H . This then finishes the proof.
Remarks:
1. With the help of this proposition, one can show that if H and K are two finite subgroups of
Γ, with respective non-degenerate 2-cocycles ΩH and ΩK, then the associated Galois co-objects
N, ∆N and
N , ∆ N for L Γ are isomorphic iff there exists g Γ and a unitary u L H
with g1Hg (cid:16) K and
λg b λg Ω2 λ
g b λ
g (cid:16) u
b u
Ω1∆L H
u.
2. One can also be more specific on when the projective corepresentations associated to two given
N, ∆N -corepresentations as above are actually isomorphic. Namely, using the notation as in
the statement of the proposition, let αg be the coaction of L Γ on B H implemented by GHg.
We may assume that αe is the 'extension' of the coaction β of L H on B H implemented by
G. Then we have αg (cid:21) αe iff gHg1
g b λ
g
(inside L H ).
(cid:16) H and Ω is coboundary equivalent to λg b λg Ωλ
35
8 A projective representation for SUq 2
In this section, we want to consider one special and non-trivial example of a projective representation
of the compact quantum group SUq 2. This projective representation will be nothing else but (a
completion of) its action on the standard Podle`s sphere.
Let us first recall the definition of SUq 2 on the von Neumann algebra level. For the rest of this
section, we fix a number 0 (cid:160) q (cid:160) 1.
Definition 8.1. Denote L 2
SUq 2 (cid:16) l2
N b l2
N b l2
Z. Consider on it the operators
a
a (cid:16)
1 q2k ek1,k b 1 b 1,
kN0
b (cid:16)
qk ekk b 1 b S,
kN
where S denotes the forward bilateral shift.
Then the compact Woronowicz algebra L 8
SUq 2, ∆
consists of the von Neumann algebra
L 8
SUq 2 (cid:16) B l2
I
¯b1 ¯bL Z „ B L 2
SUq 2,
equipped with the unique unital normal -homomorphism
∆ : L 8
SUq 2 Ñ L 8
SUq 2 ¯bL 8
SUq 2
which satisfies
"
∆
∆
a (cid:16) a b a qb
b (cid:16) b b a a
b b
b b.
Its invariant state ϕ is given by the formula
ϕ
eij b 1 b Sk
(cid:16) δi,j δk,0 1 q2
q2k,
for all i, j N, k Z,
and we may identify L 2
ξM (cid:16)
1 q2 °
a
SUq 2 with the Hilbert space in the GNS-construction for ϕ by putting
iN qi ei b ei b e0.
The definition of the standard Podle`s sphere and the associated action of SUq 2 takes the following
form on the von Neumann algebraic level.
Definition 8.2. Define L 8
SUq 2 generated by
q0 to be the von Neumann algebra inside L 8
X (cid:16) qba and Z (cid:16) bb. Then ∆ restricts to a left (ergodic) coaction α of L 8
SUq 2 on L 8
q0 .
We say that this coaction corresponds to 'the action of SUq 2 on the standard Podle`s sphere'.
S2
S2
One can show that L 8
way that
S2
q0 may be identified with the von Neumann algebra B l2
N, in such a
X Ñ
a
qk
1 q2k ek1,k
kN0
Z Ñ
q2k ekk.
kN
Under this correspondence, the α-invariant state on L 8
S2
q0 (cid:16) B l2
N equals
φα eij (cid:16) δij 1 q2
q2i.
In the terminology of the present paper, the coaction α is thus an irreducible projective representation
of SUq 2 on an infinite dimensional Hilbert space. In [9], we computed the associated Galois co-object.
To introduce it, let us first recall some notations from q-analysis (see [14]).
36
Notation 8.3. For n N Y 8 and a C, we denote
n1
a; qn (cid:16)
1 qka,
k(cid:16)0
which determines analytic functions in the variable a with no zeroes in the open unit disc.
For n N and a C, we denote by pn x; a, 0 q the Wall polynomial of degree n with parameter
value a; so
pn x; a, 0 q (cid:16) 2ϕ1 qn, 0; qa q, qx,
where 2ϕ1 denotes Heine's basic hypergeometric function.
Proposition 8.4. Let L 2
denotes the forward bilateral shift, and by L0 the operator such that
N (cid:16) l2
Z b l2
Z b l2
Z. Denote by v the operator S
b 1 b 1, where S
L0
en b em b ek (cid:16) q2n 2; q2
1{2
8 en b em b ek,
so L0
l2
the elements vnL0 , where n Z. (One easily shows that N equals the set B l2
Z. Denote N for the σ-weak closure of the right L 8
1{2
8 with u the canonical isometric inclusion of l2
(cid:16) uq2bb; q2
Z b l2
Z into
SUq 2-module generated by
Z ¯b1 b L Z.)
N b l2
N b l2
Z b l2
N, l2
Then there exists a unique Galois co-object structure N, ∆N on N for which
∆N vnL0
(cid:16) vn
b vn
q2; q2
1
p vpL0 bp
b vpL0
qb
p
,
8
p(cid:16)0
the right hand side converging in norm.
The coaction α of L 8
q0 is then an irreducible N, ∆N -corepresentation, and an
associated implementing N, ∆N -corepresentation G is determined by the following formula: denoting
G (cid:16)
N, we have, for 0 ¤ t ¤ s, that
SUq 2 on L 8
s,tN Gts b ets N ¯bB l2
°
S2
Gts (cid:16) qtts
(cid:10)1{2
q2; q2
q2; q2
s
t
q2; q2
1
st vs tL0 bst
pt bb; q2st, 0 q2
,
while for 0 ¤ s ¤ t, we have
Gts (cid:16) qsst
(cid:10)1{2
q2; q2
q2; q2
t
s
q2; q2
1
ts vt sL0
qb
ts
ps bb; q2ts, 0 q2
.
For the rest of this section, we take M, ∆M to be L 8
, and we fix the right
Galois co-object N, ∆N for M, ∆M introduced above. We then also keep using the notations intro-
duced above, as well as those from the first four sections.
SUq 2, ∆L 8
SUq 2
Our aim now is to describe in more detail the structure of the Galois co-object N, ∆N . In particular,
we want to find a complete set of irreducible N, ∆N -corepresentations. This is in fact not so difficult.
Proposition 8.5. Let N, ∆N be the Galois co-object and G the N, ∆N -corepresentation introduced
in Proposition 8.4.
Then the set of unitaries
G n :(cid:16) vn
b 1G,
n Z
37
forms a complete set of irreducible N, ∆N -corepresentations for L 8
SUq 2.
In particular, the set IN of isomorphism classes of irreducible N, ∆N -corepresentations can be nat-
urally identified with Z.
Proof. It is trivial to see that the G n are indeed irreducible N, ∆N -corepresentations, by the group-
like property of v. We then only need to show that the G n are mutually non-isomorphic and have
σ-weakly dense linear span in N .
We first prove that all G n are mutually non-isomorphic. As an isomorphism between G n and G m
would induce an isomorphism between G 0 and G mn, it is sufficient to show that G (cid:16) G 0 is not
isomorphic to G n for n (cid:24) 0. But for this, it is in turn sufficient to show that L0
(cid:16) G00 is orthogonal
to all G n
ts , by the orthogonality relations in Lemma 1.17.2 (and Theorem 3.2.2).
Now we remark that ϕ satisfies the property that ϕ
(cid:16) 0 whenever m (cid:24) 0 and k (cid:24) l,
and likewise with a replaced by a. Moreover, one easily computes that the commutation relation
(cid:16) L0 a holds. Using then the concrete form for the Grs in the previous Proposition, it is easy
vL0
to see that
b
ambk
l
ϕ
L
0
ts (cid:24) 0 ñ s (cid:16) t and n 2t (cid:16) 0.
G n
Hence the only thing left to do is to prove that L0 is orthogonal to G 2t
for t Z0. But suppose
this were not so. Then as G 2t and G 0 are irreducible, we would necessarily get that they are
isomorphic, again by the orthogonality relations. As the inner product of L0 with all G 2t
except
r, s (cid:16) t is zero, this would then imply that L0 must be a scalar multiple of G 2t
, which is equivalent
with saying that pt bb; 1, 0 q2
is a scalar multiple of the unit. As the spectrum of bb is not finite,
and pt x; 1, 0 q2
is a non-constant polynomial in x, we obtain a contradiction. Hence the G n are
mutually non-isomorphic.
rs
tt
tt
t,t k
ts have a σ-weakly dense linear span in N . Consider, for k, t N, the
. Then, up to a non-zero constant, this equals the element L0 bkpt bb; q2k, 0 q2
.
are polynomials of degree t, we conclude that the linear span of the G 2tk
t,t k
contain
We end by showing that the G n
element G 2tk
As the pt x; q2k, 0 q2
contains all elements of the form L0 bk t
all elements of the form L0 bs
contains all elements of the
form L0 bs
t for s, t N. As this linear span is closed under left multiplication with powers of v,
we conclude that this linear span contains all elements of the form vnL0 bs
t where n Z, s, t N.
As we can σ-weakly approximate elements of the form err b 1 b Sk by elements in the linear span of
the bs
indeed equals
b
N (cid:16) B l2
t. A similar argument shows that the G 2sk
t, it follows immediately that the σ-weak closure of the linear span of the G n
ts
N, l2
Z ¯b1 b L Z.
s k. Hence the linear span of all G n
ts
s k,s
b
b
b
b
Corollary 8.6.
1. Up to isomorphism, there is only one irreducible N, ∆N -corepresentation of
L 8
SUq 2.
2. The von Neumann algebra N (cf. Proposition 1.10) can be identified with `rZB l2
N.
Proof. The first statement follows immediately from the previous proposition, Theorem 3.2.2 and
Proposition 3.8, as any G n clearly implements the same irreducible N, ∆N -corepresentation. The
second statement follows from Corollary 3.3.
38
Proposition 8.7. Denote M (cid:16) L 8
SUq 2. The elements
sis of L 2
N , and, under the identification L 2
to en,t b en,s, the element
1
1q2
qt
N b l2
N Ñ `nZ l2
?
G n
ts ξM form an orthonormal ba-
N by sending
?
1
1q2
qt
G n
ts ξM
8
nZ
i,j(cid:16)0
1 b en,ts b G n
ts
N 1 ¯bN
equals V .
Proof. Remark first that ϕ
G n
ts
G n
ru (cid:16) ϕ
G
tsGru . But as we have that
αeij (cid:16)
G
ikGjl b ekl
k,lN
b ια (cid:16) φα, it follows immediately that ϕ
and ϕ
previous proposition, this proves that the
G n
G
ik
?
1
1q2
qt
Gjl (cid:16) δklδij 1 q2
q2i. Combined with the
ts ξM form an orthonormal basis.
Now
V G n
ij ξM b ξM (cid:16)
G n
ik ξM b G n
kj ξM .
kN
On the other hand, denoting V 1
(cid:16)
°
nZ
°
8
i,j(cid:16)0 1 b en,ts b G n
ts , we have
V 1 G n
ij ξM b ξM (cid:21) qi
a
1 q2
en,i b en,k b G n
kj ξM
kN
(cid:21)
G n
ik ξM b G n
kj ξM .
kN
As ξM is separating for N , this shows that V (cid:16)
V 1.
One may deduce from this that the ordinary matrix units en,ts in N (cid:16) `nNB l2
N are of the
form we required in the first section, namely their corresponding vectors en,t l2
N are eigenvectors
for the trace class operator T implementing φα (it should be remarked that we are implicitly using
Proposition 3.11 here).
We now remark that in [9], we had already computed that the reflection of L 8
(see section 4 for the terminology) may be identified with Woronowicz' quantum group L 8
(see [35]), which has as its associated von Neumann algebra
SUq 2 across N, ∆N
Eq 2
L 8
Eq 2 (cid:16) B l2
Z ¯b1 ¯bL Z „ B L 2
N .
Now it is known (see [2]) that this is a unimodular quantum group, with its invariant nsf weight ϕ0
determined by
ϕ0 eij b Sk
(cid:16) δij δk,0q2i.
Proposition 8.8. The modular element δ N (see Theorem 2.9.2) equals `nZq2nT 2
B l2
N is the operator T ei (cid:16) q2iei.
N , where T
Proof. Denote again by An the n-th component of δ N in N . Then by Proposition 4.3.3, we know
nT 2 for some dn ¡ 0. Moreover, by Proposition 8.7 and Theorem 4.2, we know that
that An (cid:16) d2
ϕ0 G n
00 G n
t v (cid:16)
q2itv, we have that ϕ0 vnL0 L
. As the dn are only determined up to a
0
fixed scalar multiple anyway, we see that we may take dn (cid:16) qn, which ends the proof.
. So to know dn, we should compute ϕ0 vnL0 L
0
(cid:16) q2nϕ0 L0 L
0
. But as σϕ0
1
1q2
v
v
00
d2
n
(cid:16)
n
n
39
Proposition 8.9. For all m, n Z and i, j, k, l N, we have
ϕ0 G n
ij G m
kl
(cid:16) δmnδikδjl
1
q2n 2j
.
Proof. By Proposition 4.2 and the previous proposition, we have that
ϕ0 G n
ij G m
kl
(cid:16) δmnδikδjl
c
q2n 2j
for a certain constant c. This constant is precisely the number ϕ0 L0 L
compute in the previous proposition. But
0
which we neglected to
ϕ0 L0 L0
(cid:16)
q2k
q2k 2; q2
8
kZ
(cid:16) q2; q2
8
kN
(cid:16) 1,
q2k
q2; q2
k
by the q-binomial theorem.
Remark: These orthogonality relations can also be written out in terms of the Wall polynomials pn.
Then one would simply get back the well-known orthogonality relations between these polynomials
(see e.g. [19], equation (2.6)), from which the above orthogonality relations can also be directly de-
duced.
As a final computation, let us determine the fusion rules between the G n and the irreducible corep-
resentations Ur of L 8
SUq 2 (where r
N).
1
2
Proposition 8.10. For all n Z and r
1
2
N, we have
G n
Ur (cid:21) `
2r
i(cid:16)0G n2r 2i.
Proof. By multiplying to the left with powers of v, it is easy to see that the fusion rules are invariant
under translation of n. We may therefore restrict to the case n (cid:16) 0.
1
2
It is then well-known that if we consider L 8
SUq 2-comodule, the coaction α splits
as `s
NU2s. From the proof of Proposition 3.9 and the remark preceding it, we obtain that G Ur
will then split as a direct sum of less than 2r 1 corepresentations. By the orthogonality relations
between the G n
ij , it then suffices to find in each G 2i2r, with 0 ¤ i ¤ 2r, a component which has
non-trivial scalar product with a matrix element of G Ur.
q0 as an L 8
S2
By looking at the border of Ur, and by using G0
00 (cid:16) L0 , we find in G Ur the elements L0 aib2ri
(up to a non-zero scalar), where i N with i ¤ 2r. It is then enough to find inside G 2i2r some
element which has non-trivial scalar product with L0 aib2ri. But by an easy computation, we
ib2ri. On the
have L0 a (cid:16) vL0
other hand, G 2i2r
L0 is a non-zero
positive operator, we find that indeed ϕ
0,2ri equals viL0 b2ri, up to a scalar. As q2ibb; q2
1 bb, and then by induction L0 aib2ri
(cid:24) 0, which then finishes the proof.
q2ibb; q2
(cid:16) viL0
2riL
0
L0 aib2ri
G 2i2r
0,2ri
i bb
40
Remarks:
1. By the discussion following Proposition 3.9, we could also have deduced these fusion rules directly
from [28], as the multiplicity diagram of the ergodic coaction of SUq 2 on the standard Podle`s
sphere is explictly computed there.
2. In [11], we discussed the 'reflection technique' (cf. section 4) with respect to another action of
SUq 2 on a type I factor, namely the von Neumann algebraic completion of its action on the
'quantum projective plane' (see e.g. [15]). We showed that the reflected quantum group in this
case was the extended (cid:129)SU q 1, 1 quantum group of Koelink and Kustermans ([18]). This shows
N , ∆ N constructed from the quantum projective plane
in particular that the Galois co-object
action is different from the one we considered in this section. In fact, as the multiplicity diagram
of this action was explicitly computed in [28], we see that the
N , ∆ N -projective representations
of SUq 2 are labeled by the forked half-line 0 , 0
Y N0 (again by the discussion following
Proposition 3.9). Now it can be shown that the quantum group (cid:129)SU q 1, 1 contains only two
N , ∆ N -
group-like unitaries. By Proposition 3.5 of [10], this implies that the associated
projective corepresentations still form a (countably) infinite family. We do not know if this
family of 'ergodic actions on type I factors' already occurs somewhere in the literature. At the
moment, we have not even succeeded in explicitly describing these ergodic coactions, except for
two cases. One case is the action on the quantum projective plane itself, which we will denote
as α. The other coaction α is obtained by amplifying with the spin 1/2-representation U1{2 of
SUq 2:
α : B Hα b M2 C Ñ M ¯bB Hα b M2 C : x Ñ U1{2
13 α b ιxU1{2 13.
Note that this is still irreducible, as the spectral decomposition of α only contains corepresenta-
tions corresponding to even integer spin by [28] (see the proof of Proposition 3.9). Remark that
the ergodic coaction α is not isomorphic to a coideal of SUq 2 (again by [28]).
References
[1] E. Aljadeff, P. Etingof, S. Gelaki and D. Nikshych, On twisting of finite-dimensional Hopf algebras,
Journal of Algebra 256 (2) (2002), 484 -- 501.
[2] S. Baaj, Repr´esentation r´eguliere du groupe quantique des d´eplacements de Woronowicz,
Ast´erisque, 232 (1995), 11 -- 49.
[3] S. Baaj and G. Skandalis, Unitaires multiplicatifs et dualit´e pour les produits crois´es de C-
alg`ebres, Ann. Sci. ´Ecole Norm. Sup., 4e s´erie, 26 (4) (1993), 425 -- 488.
[4] J. Bichon, A. De Rijdt and S. Vaes, Ergodic coactions with large quantum multiplicity and
monoidal equivalence of quantum groups, Comm. Math. Phys. 262 (2006), 703 -- 728.
[5] F. Boca, Ergodic actions of compact matrix pseudogroups on C-algebras, In: Recent advances in
operator algebras (Orl´eans, 1992), Ast´erisque 232, 93 -- 109 (1995).
[6] S. Caenepeel, Brauer groups, Hopf algebras and Galois theory, K-Monographs in Mathematics 4,
Kluwer Academic Publishers, Dordrecht (1998), 488+xvi p.
[7] K. De Commer, Galois objects for algebraic quantum groups, Journal of Algebra 321 (6) (2009),
1746-1785.
[8] K. De Commer, Galois objects and the twisting of locally compact quantum groups, to appear in
the Journal of Operator theory, preprint available at arXiv:math.OA/0804.2405v3.
[9] K. De Commer, On a Morita equivalence between the duals of quantum SU 2 and quantum E 2,
arXiv:math.QA/0912.4350v2.
41
[10] K. De Commer, Comonoidal W-Morita equivalence
for von Neumann bialgebras,
arXiv:math.OA/1004.0824v1.
[11] K. De Commer, On a correspondence between quantum SU 2, quantum E 2 and extended
quantum SU 1, 1, arXiv:math.QA/1004.4307v1.
[12] M. Enock and L. Vainerman, Deformation of a Kac algebra by an Abelian subgroup, Comm.
Math. Phys. 178 (1996), 571 -- 596.
[13] P. Etingof and S. Gelaki, The classification of triangular semisimple and cosemisimple Hopf
algebras over an algebraically closed field, Internat. Math. Res. Notices 5 (2000), 223-234.
[14] G. Gasper and M. Rahman, Basic Hypergeometric Series, Cambridge University Press, Cam-
bridge, U.K. (1990).
[15] P. Hajac, R. Matthes, W. Szyma´nski, Quantum Real Projective Space, Disc and Spheres, Algebras
and Representation Theory 6 (2) (2003), 169 -- 192.
[16] A. Ioana, S. Popa and S. Vaes, A class of superrigid group von Neumann algebras, in preparation.
[17] M. Izumi and H. Kosaki, On a subfactor analogue of the second cohomology, Rev. Math. Phys.
14 (2002), 733 -- 757.
[18] E. Koelink and J. Kustermans, A locally compact quantum group analogue of the normalizer of
SU 1, 1 in SL2, C, Comm. Math. Phys. 233 (2003), 231 -- 296.
[19] T. H. Koornwinder, The addition formula for little q-Legendre polynomials and the SU 2 quan-
tum group, SIAM J. Math. Anal. 22 (1) (1991), 295 -- 301.
[20] J. Kustermans and S. Vaes, Locally compact quantum groups in the von Neumann algebraic
setting, Mathematica Scandinavica 92 (1) (2003), 68 -- 92.
[21] M. Landstad, Operator algebras and compact groups, in Operator algebras and Group Represen-
tations, Vol. II, Monographs Stud. Math. 18, Pitman (1984), 33 -- 47.
[22] J. Lining, Representation and duality of unimodular C-discrete quantum groups, J. Korean
Math. Soc. 45 (2) (2008), 575-585.
[23] T. Masuda and Y. Nakagami, A von Neumann algebra framework for the duality of the quantum
groups, Publ. RIMS, Kyoto University 30 (1994), 799-850.
[24] M. Movshev, Twisting in group algebras of finite groups (Russian), Funktsional. Anal. i Prilozhen
27 (1993), 17-23, English translation in Funct. Anal. Appl. 27 (1993), 240-244.
[25] J. Quigg, Discrete C-coactions and C-algebraic bundles, J. Austral. Math. Soc. Ser. A 60 (2)
(1996), 204-221.
[26] P. Schauenburg, Hopf-Galois and Bi-Galois extensions, Galois theory, Hopf algebras, and semi-
abelian categories, Fields Inst. Commun. 43, AMS (2004), 469-515.
[27] M. Takesaki, Theory of Operator Algebras II, Springer, Berlin (2003).
[28] R. Tomatsu, Compact quantum ergodic systems, J. Funct. Anal. 254 (2008), 1 -- 83.
[29] S. Vaes, The unitary implementation of a locally compact quantum group action, Journal of
Functional Analysis. 180 (2001), 426-480.
[30] S. Vaes and N. Vander Vennet, Identification of the Poisson and Martin boundaries of orthogonal
discrete quantum groups, Journal of the Institute of Mathematics of Jussieu 7 (2008), 391 -- 412.
[31] L. Vainerman, 2-cocycles and twisting of Kac algebras, Comm. Math. Phys. 191 (1998), 697-721.
[32] A. Van Daele, Morita equivalence of algebraic quantum groups, in preparation.
[33] A. Wassermann, Ergodic actions of compact groups on operator algebras. II. Classification of full
multiplicity ergodic actions, Canad. J. Math. 40 (1988), 1482-1527.
42
[34] S.L. Woronowicz, Compact matrix pseudogroups, Comm. Math. Phys. 111 (1987), 613 -- 665.
[35] S.L. Woronowicz, Quantum E 2-group and its Pontryagin dual, Letters on Math. Phys. 23
(1991), 251 -- 263.
[36] S.L. Woronowicz, Compact quantum groups, in: Sym´etries quantiques (Les Houches, 1995),
North-Holland (1998), 845-884.
43
|
1510.05469 | 2 | 1510 | 2016-03-01T05:11:47 | Irreducible representations of nilpotent groups generate classifiable C*-algebras | [
"math.OA"
] | We show that C*-algebras generated by irreducible representations of finitely generated nilpotent groups satisfy the universal coefficient theorem of Rosenberg and Schochet. This result combines with previous work to show that these algebras are classifiable by their Elliott invariants within the class of unital, simple, separable, nuclear C*-algebras with finite nuclear dimension that satisfy the universal coefficient theorem. We also show that these C*-algebras are central cutdowns of twisted group C*-algebras with homotopically trivial cocycles. | math.OA | math |
Munster J. of Math. -- ( -- ), 999 -- 999
DOI
Munster Journal of Mathematics
c(cid:13) Munster J. of Math. --
Irreducible representations of nilpotent groups generate
classifiable C*-algebras
Caleb Eckhardt and Elizabeth Gillaspy
Abstract. We show that C*-algebras generated by irreducible representations of finitely generated nilpotent groups satisfy the
universal coefficient theorem of Rosenberg and Schochet. This result combines with previous work to show that these algebras
are classifiable by their Elliott invariants within the class of unital, simple, separable, nuclear C*-algebras with finite nuclear
dimension that satisfy the universal coefficient theorem. We also show that these C*-algebras are central cutdowns of twisted
group C*-algebras with homotopically trivial cocycles.
1. Introduction
This note concludes a long line of study into the C ∗-algebras generated by irreducible representations
of finitely generated nilpotent groups. Specifically, we prove that such C ∗-algebras satisfy the universal
coefficient theorem (UCT) of Rosenberg and Schochet [15]. We combine this with a slew of other results
to show that these algebras are classifiable by their Elliott invariant within the class C of unital, simple,
separable, nuclear C*-algebras with finite nuclear dimension that satisfy the UCT.
Let G be a finitely generated nilpotent group and π an irreducible unitary representation of G. Let C ∗
π(G)
be the C ∗-algebra generated by π(G). Tikuisis, White and Winter recently took the final step in a long
and beautiful journey of showing that two elements of C with the same Elliott invariant are isomorphic [17,
Corollary D]. Therefore our job consists of showing that C ∗
π(G) ∈ C.
We have known for a while that C ∗
π(G) is nuclear [8] and simple [12]. Recently, the first author together
with McKenney [7] showed that C ∗
π(G) has finite nuclear dimension (this work directly relied on a long list
of results including [5, 10, 11, 14, 18] -- see the introduction of [7] for the full story). As pointed out in [7,
Theorem 4.5] the only missing ingredient to show C ∗
π(G) ∈ C was the UCT. It was also pointed out in [7,
Theorem 4.6] that if G is torsion free and π is a faithful representation then C ∗
π(G) satisfies the UCT. This
observation has already found success in [6] where the authors calculated the Elliott invariant of C*-algebras
generated by faithful irreducible representations of the (torsion free) unitriangular group U T (4, Z), thus
classifying them within C.
Proving Theorem 3.0.9, that C ∗
π(G) satisfies the UCT, is the main goal of this note. In the course of
our investigations we noticed that C ∗
π(G) is isomorphic to a central cutdown of a twisted group C*-algebra
C ∗(G/N, σ) for some 2-cocycle σ, where N is a finite index subgroup of Z(G). Moreover, the cocycle σ is
homotopic to the trivial cocycle. We include this as Theorem 4.0.16 as it may be of independent interest
and useful for K-theory calculations.
2. Nilpotent lemmas
We first recall and prove necessary facts about nilpotent groups used in subsequent sections. We refer the
reader to Segal's book [16] for information about nilpotent and polycyclic groups. Throughout this section
G is a finitely generated nilpotent group. We let Z(G) denote the center of G and
Gf = {x ∈ G : the conjugacy class of x is finite }
C.E. was partially supported by a grant from the Simons Foundation.
1000
Caleb Eckhardt and Elizabeth Gillaspy
Clearly Z(G) ≤ Gf ≤ G. Define the torsion subgroup of G as
T (G) = {x ∈ G : x has finite order }.
Remark 2.0.1. In general, T (G) need not be a subgroup of G but it is a standard exercise to show that
T (G) ≤ G for nilpotent G (see [16, Corollary 1.B.10]).
Every finitely generated nilpotent group is polycyclic, that is, there is a normal series
(2.1)
{e} = Gn E Gn−1 E · · · E G1 E G0 = G
where Gi/Gi+1 is cyclic for each i = 0, . . . , n − 1. From this definition it follows that polycyclic groups are
finitely generated and that subgroups of polycyclic groups are polycyclic. Moreover, for finitely generated
nilpotent groups, T (G) is polycyclic and therefore finite, since it must satisfy (2.1). From this it easily follows
that T (G) ≤ Gf .
It is most likely well-known among group theorists that Z(G) has finite index in Gf and that G/Gf is
torsion free. We were not able to locate references for these facts (although they are more-or-less corollaries
of Baer's [1]) so we include the brief proofs.
Lemma 2.0.2. Let G be torsion free, finitely generated and nilpotent. Then Z(G) = Gf .
Proof. By [1, Lemma 3], the group Z(Gf ) has finite index in Gf .
Case 1. Z(Gf ) = Z(G). By [9], G/Z(G) is torsion free. Then Gf /Z(Gf ) is a finite, torsion free group,
i.e. Gf = Z(Gf ).
Case 2. Z(Gf ) 6= Z(G). By assumption there are elements x ∈ Z(Gf ) \ Z(G) and y ∈ G such that
yxy−1 6= x. Let α be the automorphism of Z(Gf ) induced by conjugation by y. Because x ∈ Gf we have
αn(x) = x for some n > 1. Moreover, Z(Gf ) is a finitely generated torsion free abelian group, since G is
nilpotent, torsion free, and finitely generated. Since the group generated by Z(Gf ) and y is nilpotent it
follows that the matrix for α is unipotent. In particular, 1 is the only eigenvalue of α, so α cannot have any
periodic, non-fixed points, a contradiction to the fact that αn(x) = x. Therefore Z(G) = Z(Gf ), so Case 2
never occurs and we are in Case 1.
(cid:3)
Lemma 2.0.3. Let G be a finitely generated nilpotent group, then Z(G) has finite index in Gf .
Proof. Let x ∈ Gf . By [1, Lemma 3], there is some n ≥ 1 such that z = xn ∈ Z(Gf ). Let y1, . . . , yk
generate G. Since zT (G) ∈ (G/T (G))f and G/T (G) is torsion free, by Lemma 2.0.2 it follows that zT (G) ∈
Z(G/T (G)), i.e.
Then for n ≥ 1 we have
(2.2)
[zn+1, yi] = zn[z, yi]z−n[zn, yi] = [z, yi][zn, yi].
[z, yi] ∈ T (G) for i = 1, . . . , k.
The first equality holds in every group and the second follows because z ∈ Z(Gf ) and [z, yi] ∈ T (G) ≤ Gf .
By induction, one uses (2.2) to show that [zn, yi] = [z, yi]n for all 1 ≤ i ≤ k and n ≥ 1.
Since [z, yi] ∈ T (G) it follows that there is some power of z such that [zd, yi] = e for all i = 1, . . . , k. In
particular xnd = zd ∈ Z(G). Therefore Gf /Z(G) is a torsion group. But Gf /Z(G) is polycyclic and therefore
finite.
(cid:3)
Lemma 2.0.4. Let G be a finitely generated nilpotent group. Then G/Gf is torsion free.
Proof. Suppose x ∈ G \ Gf and xn ∈ Gf for some n. By Lemma 2.0.3 we have xm ∈ Z(G) for some m. Since
G/T (G) is torsion free and xmT (G) ∈ Z(G/T (G)) we have xT (G) ∈ Z(G/T (G)). This means that for every
y ∈ G we have yxy−1x−1 ∈ T (G); equivalently, yxy−1 = xz for some z ∈ T (G). Since T (G) is finite, this
means x ∈ Gf , a contradiction.
(cid:3)
Munster Journal of Mathematics Vol. -- ( -- ), 999 -- 999
Irreducible representations of nilpotent groups generate classifiable C*-algebras
1001
3. Main result
Definition 3.0.5. For a group G and normalized positive definite function φ : G → C, let (πφ, Hφ) denote
πφ (G) denote the C ∗-algebra generated by πφ(G). A
the GNS representation of G associated with φ. Let C ∗
trace on G is a normalized, positive definite function that is constant on conjugacy classes. Notice that a
trace τ on G canonically induces a tracial state on C ∗(G), which we will also denote by the same symbol τ .
The following lemma is well-known and may be found for example in [2, Proposition 4.1.9].
Lemma 3.0.6. Let A be a C*-algebra and φ a faithful state on A. Let α be an automorphism of A. Let
u ∈ A ⋊α Z be the unitary implementing α. One extends φ to a faithful state on A ⋊α Z by setting φ(xun) = 0
when x ∈ A and n 6= 0.
Lemma 3.0.7. Let N be a discrete group and α an automorphism of N. Set G = N ⋊α Z. Let u ∈ G be the
element implementing α by conjugation. Let τ be a trace on G such that τ (x) = 0 for all x ∈ G \ N. Then
Proof. Let σ : C ∗
covariant representation which is the inclusion map on C ∗
τ ◦ σ defines a faithful tracial state on C ∗
πτ (N ) ⋊Adπτ (u) Z → C ∗
πτ (N ) ⋊Adπτ (u) Z
πτ (G) ∼= C ∗
C ∗
πτ (G) denote the surjective *-homomorphism corresponding to the
πτ (N ) and sends n ∈ Z to πτ (u)n. By Lemma 3.0.6,
(cid:3)
πτ (N ) ⋊Adπτ (u) Z. Therefore σ is injective.
Theorem 3.0.8. Let G be finitely generated and nilpotent. Let τ be a trace on G such that τ (x) = 0 for all
x ∈ G \ Gf . Then C ∗
πτ (G) is isomorphic to an iterated crossed product of C ∗
πτ (Gf ) by Z-actions, i.e.
(3.1)
and C ∗
πτ (G) satisfies the UCT.
πτ (G) ∼= C ∗
C ∗
πτ (Gf ) ⋊ Z ⋊ · · · ⋊ Z,
Proof. By Lemma 2.0.4, G/Gf is torsion free. Since G/Gf is finitely generated and nilpotent it is isomorphic
to an iterated semi-direct product of Z-actions. By repeatedly using the fact that the short exact sequence
of groups 0 → A → B → Z → 0 always splits we have
G ∼= Gf ⋊ Z ⋊ · · · ⋊ Z.
Now (3.1) follows from repeated applications of Lemma 3.0.7. By Lemma 2.0.3, the group C*-algebra C ∗(Gf )
is subhomogeneous, from which it follows that C ∗
πτ (Gf ) is
Type I, so it satisfies the UCT by [15, Theorem 1.17]. Finally, [15, Proposition 2.7, Theorem 4.1] shows that
the UCT is preserved by Z-actions, implying that C ∗
(cid:3)
πτ (Gf ) is also subhomogeneous. In particular, C ∗
πτ (G) satisfies the UCT.
Theorem 3.0.9. Let G be a finitely generated nilpotent group and π an irreducible representation of G.
Then C ∗
π(G) satisfies the UCT.
Proof. Suppose first that π is faithful on G. Since G is finitely generated, by [3, Proposition 5.1] there is
an extreme trace τ on G such that C ∗
πτ (G). Since τ is an extreme trace and π is faithful on G, it
follows from [3, Theorem 4.5] that τ (g) = 0 if g 6∈ Gf . By Theorem 3.0.8, C ∗
π(G) ∼= C ∗
πτ (G) satisfies the UCT.
If π is not faithful, then we replace G with G/ker(π) and π with eπ(gker(π)) := π(g). Then eπ is faithful
(cid:3)
π(G), so we simply apply the above proof to (G/ker(π), eπ).
eπ(G/ker(π)) ∼= C ∗
and C ∗
Corollary 3.0.10. Let G be a finitely generated nilpotent group and π an irreducible representation of G.
Then C ∗
π(G) is classified by its ordered K-theory within the class of simple, unital, nuclear C*-algebras with
finite nuclear dimension that satisfy the universal coefficient theorem.
Proof. This is Theorem 3.0.9 combined with [7, Theorem 4.5].
(cid:3)
4. A structure theorem for C ∗
π(G)
In the course of proving Theorem 3.0.9 we discovered a structure theorem (Theorem 4.0.16) for C ∗
π(G)
that is most likely unknown. We therefore include it as it may be of interest to those working with twisted
group C ∗-algebras and representation theory.
Let G be a discrete amenable group and N ≤ Z(G). Let ω ∈ bN (the dual group of N ) and also denote by ω
the trivial extension of ω to G (i.e., ω(x) = 0 for x 6∈ N.) Let c : G/N → G be a choice of coset representatives.
For t ∈ G, let tω ∈ Hω denote the canonical image of t associated with the GNS representation.
Munster Journal of Mathematics Vol. -- ( -- ), 999 -- 999
1002
Caleb Eckhardt and Elizabeth Gillaspy
Recall that, for any s, t ∈ G, we have htω, sωi = ω(s−1t). Hence if s−1t 6∈ N , the vectors sω and tω are
orthogonal. On the other hand, for s ∈ G we have hsω, c(sN )ωi = ω(c(sN )−1s) ∈ T, from which it follows
that
(4.1)
sω = ω(c(sN )−1s)c(sN )ω.
We deduce that {sω : s ∈ c(G/N )} is an orthonormal basis for Hω. Define W : Hω → ℓ2(G/N ) by
W (sω) = δsN for s ∈ c(G/N ). Then W is unitary. Moreover, by (4.1) we have that, for any y ∈ G and
s ∈ c(G/N ),
(4.2)
W πω(y)W ∗(δsN ) = ω(c(ysN )−1yc(sN ))δysN .
Now (as in [13]) define the 2-cocycle σ : G/N × G/N → T by
(4.3)
σ(xN, yN ) = ω(c(xN )c(yN )c(xyN )−1).
By an obvious adaptation of the proof of [6, Proposition 2.9] and the discussion preceding it, we have the
following
Lemma 4.0.11. Let G be a discrete amenable group and N ≤ Z(G). Let ω ∈ bN . Also let ω denote the
πω (G) is isomorphic to the twisted group C ∗-algebra
trivial extension to G and define σ as in (4.3). Then C ∗
C ∗(G/N, σ).
Let λG/N denote the left regular representation of G on ℓ2(G/N ).
Lemma 4.0.12. Let G be a discrete amenable group and N ≤ Z(G). Let ω ∈ bN and denote by ω the trivial
extension of ω to G. Let τ be a trace on G such that τ N = ωN . Then λG/N ⊗ πτ is unitarily equivalent to
πω ⊗ 1Hτ .
Writing ≺ to denote weak containment, it then follows that πτ ≺ πω; equivalently, C ∗
πτ (G) is a quotient
of C ∗
πω (G).
Proof. This is a slight modification of the proof of Fell's absorption principle. Define a unitary U on
ℓ2(G/N ) ⊗ Hτ by U (δtN ⊗ ξ) = δtN ⊗ πτ (c(tN ))ξ. For any y, t ∈ G,
U ∗(λG/N (y) ⊗ πτ (y))U (δtN ⊗ ξ) = U ∗(λG/N (y) ⊗ πτ (y))(δtN ⊗ πτ (c(tN ))ξ)
= U ∗(δytN ⊗ πτ (yc(tN ))ξ)
= δytN ⊗ πτ (c(ytN )−1yc(tN ))ξ
= ω(c(ytN )−1yc(tN ))δytN ⊗ ξ,
because τ = ω on N , both are multiplicative on N , and c(ytN )−1yc(tN ) ∈ N. Unitary equivalence now
follows from (4.2).
We now show weak containment. Let 1G denote the trivial representation of G on C. Since G/N is
amenable, λG/N contains an approximately fixed vector, and thus 1G ≺ λG/N . Then
πτ ∼ 1G ⊗ πτ ≺ λG/N ⊗ πτ ∼ πω ⊗ 1Hτ ≺ πω.
(cid:3)
The following is well-known (see for example [2, Corollary 2.5.12]).
Lemma 4.0.13. Let H ≤ G be discrete groups. Let C ∗
r (G) denote the reduced group C*-algebra of G. The
linear map from C[G] to C[H] defined by E(λs) = λs if s ∈ H and E(λs) = 0 if s 6∈ H extends to a
conditional expectation from C ∗
r (G) such that
ω(λs) = 0 when s 6∈ H, then ω ◦ E = ω.
r (H). In particular, if ω is a tracial state on C ∗
r (G) onto C ∗
Lemma 4.0.14. Let H ≤ G be amenable discrete groups. Let τ be a trace on G that vanishes on G \ H. Let
πτ (G) → C ∗
E : C ∗(G) → C ∗(H) be the conditional expectation from Lemma 4.0.13. The map Eτ : C ∗
πτ (H)
given by Eτ (πτ (x)) = πτ (E(x)) extends to a well-defined τ -preserving conditional expectation onto C ∗
πτ (H).
Munster Journal of Mathematics Vol. -- ( -- ), 999 -- 999
Irreducible representations of nilpotent groups generate classifiable C*-algebras
1003
Proof. Since τ is a tracial state, for any x ∈ C ∗(G), we have πτ (x) = 0 if and only if τ (x∗x) = 0. Con-
sequently, πτ (x) = 0 implies τ (E(x)∗E(x)) ≤ τ (E(x∗x)) = τ (x∗x) = 0, so πτ (E(x)) = 0. Therefore Eτ is
well-defined and τ -preserving by Lemma 4.0.13.
The map Eτ is clearly idempotent so we only need to check it is contractive. This proceeds as in the
case of building conditional expectations in the arena of finite von Neumann algebras. We include the proof
for the convenience of the reader.
For each x ∈ C ∗(G), let xτ ∈ Hτ be the canonical image of x. Since E is τ -preserving, the map P (xτ ) :=
E(x)τ extends to an orthogonal projection on Hτ . Since E(x) ∈ C ∗(H),
kπτ (E(x))k =
sup
hπτ (E(x))yτ , zτ i
y,z∈C ∗(H),τ (y∗y)=τ (z∗z)=1
=
=
=
=
≤
sup
τ (z ∗E(x)y)
y,z∈C ∗(H),τ (y∗y)=τ (z∗z)=1
sup
hP (xτ ), (zy∗)τ i
y,z∈C ∗(H),τ (y∗y)=τ (z∗z)=1
sup
hxτ , P ((zy∗)τ )i
y,z∈C ∗(H),τ (y∗y)=τ (z∗z)=1
sup
hxτ , (zy∗)τ i
y,z∈C ∗(H),τ (y∗y)=τ (z∗z)=1
sup
hxτ , (zy∗)τ i
y,z∈C ∗(G),τ (y∗y)=τ (z∗z)=1
= kπτ (x)k.
(cid:3)
Lemma 4.0.15. Let G be a finitely generated nilpotent group and τ an extreme trace on G such that τ (g) = 1
implies g = e. Let N ≤ Z(G) be a finite index subgroup of Z(G). Let ω be the trivial extension of τ N to G.
Then there is a central projection p ∈ C ∗
πω (Gf ) ∩ C ∗
πω (G)′ such that
πτ (G) ∼= pC ∗
C ∗
πω (G).
Proof. Our hypotheses, combined with [3, Theorem 4.5], imply that τ vanishes on G \ Gf . By Lemma 4.0.12,
the representation πω weakly contains πτ . Therefore the *-homomorphism σ : C ∗
πτ (G) given by
σ(πω(x)) = πτ (x) is well-defined. From Lemma 4.0.14 (with H = Gf ) we have trace-preserving conditional
expectations Eω : C ∗
πτ (Gf ). By the definitions of these expectations
in Lemma 4.0.14 we have, for any x ∈ C ∗(G),
πω (Gf ) and Eτ : C ∗
πω (G) → C ∗
πω (G) → C ∗
πτ (G) → C ∗
Eτ (σ(πω(x))) = Eτ (πτ (x)) = πτ (E(x)) = σ(πω(E(x))) = σ(Eω(πω(x))).
Equivalently, the following diagram commutes:
(4.4)
C ∗
πω (G)
Eω
C ∗
πω (Gf )
σ
σ
C ∗
πτ (G)
Eτ
/ C ∗
πτ (Gf )
By [3, Proposition 5.1] there is an irreducible representation π of G such that π(x) 7→ πτ (x) defines an
isomorphism from C ∗
πτ (G). The irreducibility of π then implies that πτ (z) ∈ C · 1Hτ for all
z ∈ Z(G).
π(G) onto C ∗
Using formula (4.2) we see that πω(n)sω = ω(n)sω for all s ∈ c(G/N ), hence πω(n) = ω(n) · 1Hω . By
πω (Gf ) is finite dimensional by Lemma 4.0.11.
Lemma 2.0.3, the group N has finite index in Gf . Therefore, C ∗
πω (Gf ) ∩ C ∗
πω (Gf )′ such that
Since C ∗
(4.5)
πω (Gf ) is finite dimensional there is a projection p ∈ C ∗
πω (Gf ) = (1 − p)C ∗
ker(σ) ∩ C ∗
πω (Gf ).
Munster Journal of Mathematics Vol. -- ( -- ), 999 -- 999
/
/
/
1004
Caleb Eckhardt and Elizabeth Gillaspy
Let x ∈ C ∗
Using this fact we obtain
πω (G). By Lemma 4.0.14, the conditional expectations Eω and Eτ preserve their respective traces.
x ∈ ker(σ) ⇐⇒ τ (σ(x∗x)) = 0
⇐⇒ τ (Eτ (σ(x∗x))) = 0
⇐⇒ Eτ (σ(x∗x)) = 0
⇐⇒ σ(Eω(x∗x)) = 0 By (4.4)
⇐⇒ pEω(x∗x) = 0 By (4.5)
⇐⇒ Eω(px∗xp) = 0
⇐⇒ ω(Eω(px∗xp)) = 0
⇐⇒ ω(px∗xp) = 0
⇐⇒ xp = 0
⇐⇒ x ∈ C ∗
πω (G)(1 − p).
It follows that (1 − p) is central. Indeed, for any x ∈ C ∗
(it follows from the above list of equivalences that C ∗
for some y ∈ C ∗
same argument applied to x∗ shows px(1 − p) = 0, in other words, that x commutes with p.
πω (G), we have (1 − p)x = (x∗(1 − p))∗ ∈ C ∗
πω (G)(1 − p)
πω (G)(1 − p) is self-adjoint). Then (1 − p)x = y(1 − p)
πω (G). Consequently, (1 − p)x(1 − p) = y(1 − p) = (1 − p)x. It follows that (1 − p)xp = 0. The
(cid:3)
Theorem 4.0.16. Let G be a finitely generated nilpotent group and π a faithful irreducible representation of
G. There is a torsion free, finite index subgroup N ≤ Z(G), a 2-cocycle σ on G/N , and a central projection
p ∈ C ∗(G/N, σ), such that C ∗
π(G) ∼= pC ∗(G/N, σ). Moreover, σ is homotopic to the trivial cocycle.
Proof. As in the proof of Theorem 3.0.9 there is an extreme trace τ on G satisfying the hypotheses of Lemma
4.0.15 such that C ∗
π(G) ∼= C ∗
πτ (G).
Since Z(G) is a finitely generated abelian group it contains a finite index torsion free subgroup N. Then
bN is a d-torus and in particular path-connected. It follows that the cocycle defined in (4.3) for any ω ∈ bN
is homotopic to the trivial cocycle. The conclusion now follows from Lemmas 4.0.11 and 4.0.15.
(cid:3)
Remark 4.0.17. Let C ∗(G/N, σ) be as in Theorem 4.0.16. Since G/N is amenable, [4, Theorem 1.9]
applies to C ∗(G/N, σ). In particular, since σ is homotopic to the trivial cocycle, we have K∗(C ∗(G/N, σ)) ∼=
K∗(C ∗(G/N )). Since K-theory is additive over direct sums, we feel that Theorem 4.0.16 may be beneficial
for K-theory calculations of C*-algebras generated by irreducible representations of nilpotent groups (see for
example [6] for this idea in action).
References
[1] Reinhold Baer. Finiteness properties of groups. Duke Math. J., 15:1021 -- 1032, 1948.
[2] Nathanial P. Brown and Narutaka Ozawa. C ∗-algebras and finite-dimensional approximations, volume 88 of Graduate
Studies in Mathematics. American Mathematical Society, Providence, RI, 2008.
[3] A. L. Carey and W. Moran. Characters of nilpotent groups. Math. Proc. Cambridge Philos. Soc., 96(1):123 -- 137, 1984.
[4] Siegfried Echterhoff, Wolfgang Luck, N. Christopher Phillips, and Samuel Walters. The structure of crossed products of
irrational rotation algebras by finite subgroups of SL2(Z). J. Reine Angew. Math., 639:173 -- 221, 2010.
[5] Caleb Eckhardt. Quasidiagonal representations of nilpotent groups. Adv. Math., 254:15 -- 32, 2014.
[6] Caleb Eckhardt, Craig Kleski, and Paul McKenney. Classification of C*-algebras generated by representations of the
unitriangular group U T (4, Z). arXiv:1506.01272 [math.OA], 2015.
[7] Caleb Eckhardt and Paul McKenney. Finitely generated nilpotent group C*-algebras have finite nuclear dimension.
arXiv:1409.4056 [math.OA], to appear in J. Reine Angew., 2015.
[8] Christopher Lance. On nuclear C ∗-algebras. J. Functional Analysis, 12:157 -- 176, 1973.
[9] A. I. Mal'cev. Nilpotent torsion-free groups. Izvestiya Akad. Nauk. SSSR. Ser. Mat., 13:201 -- 212, 1949.
[10] Hiroki Matui and Yasuhiko Sato. Strict comparison and Z-absorption of nuclear C ∗-algebras. Acta Math., 209(1):179 -- 196,
2012.
[11] Hiroki Matui and Yasuhiko Sato. Decomposition rank of UHF-absorbing C*-algebras. Duke Math. J., 163(14):2687 -- 2708,
2014.
[12] Calvin C. Moore and Jonathan Rosenberg. Groups with T1 primitive ideal spaces. J. Functional Analysis, 22(3):204 -- 224,
1976.
Munster Journal of Mathematics Vol. -- ( -- ), 999 -- 999
Irreducible representations of nilpotent groups generate classifiable C*-algebras
1005
[13] Judith A. Packer and Iain Raeburn. On the structure of twisted group C ∗-algebras. Trans. Amer. Math. Soc., 334(2):685 --
718, 1992.
[14] Mikael Rørdam. The stable and the real rank of Z-absorbing C ∗-algebras. Internat. J. Math., 15(10):1065 -- 1084, 2004.
[15] Jonathan Rosenberg and Claude Schochet. The Kunneth theorem and the universal coefficient theorem for Kasparov's
generalized K-functor. Duke Math. J., 55(2):431 -- 474, 1987.
[16] Daniel Segal. Polycyclic groups, volume 82 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge,
1983.
[17] Aaron Tikuisis, Stuart White, and Wilhelm Winter. Quasidiagonality of nuclear C*-algebras. arXiv:1509.08318 [math.OA],
2015.
[18] Wilhelm Winter. Nuclear dimension and Z-stability of pure C∗-algebras. Invent. Math., 187(2):259 -- 342, 2012.
Department of Mathematics, Miami University, Oxford, Ohio
E-mail: [email protected]
Department of Mathematics, University of Colorado - Boulder, Boulder, Colorado
E-mail: [email protected]
Munster Journal of Mathematics Vol. -- ( -- ), 999 -- 999
|
1605.06050 | 1 | 1605 | 2016-05-19T16:57:50 | Rokhlin Property for Group Actions on Hilbert $C^*$-modules | [
"math.OA",
"math.DS",
"math.FA"
] | We introduce Rokhlin properties for certain discrete group actions on $C^*$-correspondences as well as on Hilbert bimodules and analyze them. It turns out that the group actions on any $C^*$-correspondence $E$ with Rokhlin property induces group actions on the associated $C^*$-algebra $\mathcal O_E$ with Rokhlin property and the group actions on any Hilbert bimodule with Rokhlin property induces group actions on the linking algebra with Rokhlin property. Permanence properties of several notions such as nuclear dimension and $\mathcal D$-absorbing property with respect to crossed product of Hilbert $C^*$-modules with groups, where group actions have Rokhlin property, are studied. We also investigate a notion of outerness for Hilbert bimodules. | math.OA | math | Rokhlin Property for Group Actions
on Hilbert C ∗-modules
Santanu Dey, Hiroyuki Osaka and Harsh Trivedi
Abstract
We introduce Rokhlin properties for certain discrete group actions on C ∗-
correspondences as well as on Hilbert bimodules and analyze them. It turns out
that the group actions on any C ∗-correspondence E with Rokhlin property in-
duces group actions on the associated C ∗-algebra OE with Rokhlin property and
the group actions on any Hilbert bimodule with Rokhlin property induces group
actions on the linking algebra with Rokhlin property. Permanence properties of
several notions such as nuclear dimension and D-absorbing property with respect
to crossed product of Hilbert C ∗-modules with groups, where group actions have
Rokhlin property, are studied. We also investigate a notion of outerness for Hilbert
bimodules.
AMS 2010 Subject Classification: 46L08, 46L35, 46L40, 46L55.
Key words: crossed products, Cuntz-Pimsner algebra, group action, Hilbert C ∗-
modules, linking algebras, nuclear dimension, Rokhlin property, self-absorbing C ∗-
algebras.
6
1
0
2
y
a
M
9
1
]
.
A
O
h
t
a
m
[
1
v
0
5
0
6
0
.
5
0
6
1
:
v
i
X
r
a
1
Introduction
Elliott in [6] initiated the classification program for nuclear C ∗-algebras
based on the K-theory of C ∗-algebras. Many aspects of the modern approach to this
program are described in the monograph [32] by Rørdam.
In [7], Elliott and Toms
discussed the importance of the following regularity properties in the modern study of the
classification program: strict comparison, finite decomposition rank, and Z-absorbing
where Z is the Jiang-Su algebra which is the first object in the category of strongly self-
absorbing C ∗-algebras. Toms and Winter gave a plausible conjecture: If A is a unital
simple nuclear separable infinite dimensional C ∗-algebra, then the following statements
are equivalent:
(1) A has finite nuclear dimension;
(2) A is Z-stable;
(3) A has the strict comparison.
1
The implications (1) ⇒ (2) and (2) ⇒ (3) have been proved by Winter [36] and Rørdam
[33], respectively. The implication (3) ⇒ (2) has been proved by several authors under
certain assumption, see [35].
The classification of the crossed products of C ∗-algebras is a very challenging prob-
lem, i.e., if we start with a C ∗-algebra of a particular type and a group action on it, then
it is difficult to predict the properties of their crossed product. The Rokhlin property
for group actions on C ∗-algebras has significantly influenced the approaches taken in
the classification theory of C ∗-algebras. Kishimoto [21] showed that the reduced crossed
product of a simple C ∗-algebra, with respect to an outer action of a discrete group, is
simple. Herman and Ocneanu [10] defined the Rokhlin property of group actions on
C ∗-algebras in terms of projections and this notion is stronger than the outerness. The
modern definition of the Rokhlin property is due to Izumi [13]. Several classes includ-
ing AF-algebras, AI-algebras, AT-algebras are closed under finite group actions with
the Rokhlin property (cf.
[25]). An approach to the Rokhlin property, introduced by
Santiago [34], for finite group actions on not necessarily unital C ∗-algebras (cf. [12, 24])
use positive contractions instead of projections. Note that when A is unital, using [34,
Corollary 1], we can replace orthogonal positive contractions by orthogonal projections
and get Izumi's definition of Rokhlin property. We begin Section 2 with Santiago's def-
inition and list there some classes that are preserved under crossed products by finite
group actions with this Rokhlin property.
In [29], from C ∗-correspondences Pimsner constructed C ∗-algebras which are known
as Cuntz-Pimsner algebras. The class of Cuntz-Pimsner algebras includes Cuntz alge-
bras, Cuntz-Krieger algebras and the crossed products by Z. For every C ∗-correspondence
(E, A, φ), Katsura in [17] defined a C ∗-algebra OE. The algebra OE is the same as the
Cuntz-Pimsner algebra when φ is injective. The algebras OE also generalize the crossed
product, as defined in [1], by Hilbert bimodules and graph algebras (cf. [8]). The graph
C ∗-algebras of topological graphs can also be realized as OE for some C ∗-correspondence
(E, A, φ) (see [18] for details).
In Subsection 3.1 we define compatible action, (η, α), of a locally compact group
on a C ∗-correspondence. Hao and Ng in [9] proved that each compatible action of
a locally compact group G on a C ∗-correspondence (E, A, φ) induces an action of G
on the associated C ∗-algebra OE. It is of interest to determine at least the sufficient
conditions under which for a compatible action (η, α) of G, permanence property is
exhibited by several notions related to this C ∗-algebra, associated to the Hilbert module
with respect to the crossed product. We define the Rokhlin property for finite group
actions on C ∗-correspondences in Subsection 3.2 and provide an answer of the above
question regarding sufficient conditions in Subsection 3.3 when G is finite.
For any class of unital and separable C ∗-algebras C, Osaka and Phillips [25] intro-
duced the notion of local C-algebra. Santiago [34] extended this approach by considering
non-unital C ∗-algebras. A notion of closed under local approximation is defined (see page
104) in terms of local C-algebra. If C denotes certain class of C ∗-algebras such as purely
infinite C ∗-algebras, simple stably projectionless C ∗-algebras etc. listed in Theorem 2.3,
then C is closed under local approximation and under crossed product with a finite group
action with the Rokhlin property. As an application of the observation made in Section
4 we show that if an action (η, α) of a finite group G on (E, A, φ) has the Rokhlin
2
property and OE belongs to one of the classes mentioned above, then OE×ηG belongs to
the same class (see Corollary 3.8). At the end of Section 3, we point out that the gauge
action on the graph C ∗-algebra is saturated by [15, Theorem 6.3], but the corresponding
action on the C ∗-correspondence does not have the Rokhlin property.
We introduce, in Section 4, the Rokhlin property for compatible finite group actions
on Hilbert bimodules and prove the following: If we realize a Hilbert A-module E as
a Hilbert K(E)-A bimodule, and if A belongs to a class C in the previous paragraph
and if a group action η of a finite group G on the Hilbert A-module E has the Rokhlin
property as a certain compatible action on the bimodule, then the linking algebra of the
crossed product Hilbert A ×α G-module E ×η G belongs to the class C. To obtain this
result we first prove that any compatible action of a finite group G with the Rokhlin
property on a Hilbert bimodule induces an action of G with the Rokhlin property on
the linking algebra.
2 Rokhlin property for finite group actions on C ∗-
algebras
In this section, first we recall the definition of the Rokhlin property for finite group
actions on a C ∗-algebra, from [34], which involves positive contractions. The Rokhlin
property for finite group actions on C ∗-algebras was studied by several authors [13, 24,
25, 28, 34], and it was proved that several classes of C ∗-algebras are preserved under the
crossed product when the action of the group has the Rokhlin property. We list here
some such classes from [34].
Definition 2.1. Let α : G → Aut(A) be an action of a finite group G on a C ∗-algebra
A. We say that α has the Rokhlin property if for any ǫ > 0 and every finite subset S
of A there exist orthogonal positive contractions (fg)g∈G ⊂ A satisfying
(1) k(Pg∈G fg)a − ak < ǫ for all a ∈ S,
(2) kαh(fg) − fhgk < ǫ for h, g ∈ G,
(3) k[fg, a]k < ǫ for g ∈ G and a ∈ S.
The elements (fg)g∈G are called Rokhlin elements for α.
Remark 2.2. When A is unital, α has the Rokhlin property in the sense of Izumi [13].
That is, we can take a partition of unity (eg)g∈G consisting of projections in place of
(fg)g∈G (see [34, Corollary 1]).
The following notions are borrowed from [25, 34]: Let C be a class of C ∗-algebras.
A local C-algebra is a C ∗-algebra A such that for every finite set S ⊂ A and every
ǫ > 0, there exists a C ∗-algebra B in C and a ∗-homomorphism π : B → A such that
dist(a, π(B)) < ǫ for all a ∈ S. We say that C is closed under local approximation if every
local C-algebra belongs to C.
We recall the following result from [34] (cf. [26, 25, 28, 11, 20, 14]).
3
Theorem 2.3. [34] The following classes are closed under local approximation and under
crossed product with finite group actions with the Rokhlin property, respectively:
(i) purely infinite C ∗-algebras,
(ii) C ∗-algebras having stable rank one,
(iii) C ∗-algebras with real rank zero,
(iv) C ∗-algebras of nuclear dimension at most n, where n ∈ Z+,
(v) separable D-absorbing C ∗-algebras where D is a strongly self-absorbing C ∗-algebra,
(vi) simple C ∗-algebras,
(vii) simple C ∗-algebras that are stably isomorphic to direct limits of sequences of C ∗-
algebras, in a class S, where S is a class of finitely generated semiprojective C ∗-
algebras that is closed under taking tensor products by matrix algebras over C,
(viii) separable AF-algebras,
(ix) separable simple C ∗-algebras that are stably isomorphic to AI-algebras,
(x) separable simple C ∗-algebras that are stably isomorphic to AT-algebras,
(xi) C ∗-algebras that are stably isomorphic to sequential direct limits of one-dimensi-
onal noncommutative CW-complexes,
(xii) separable C ∗-algebras whose quotients are stably projectionless,
(xiii) simple stably projectionless C ∗-algebras,
(xiv) separable C ∗-algebras with almost unperforated Cuntz semigroup,
(xv) simple C ∗-algebras with strict comparison of positive elements,
(xvi) separable C ∗-algebras whose closed two-sided ideals are nuclear and satisfy the
Universal Coefficient Theorem.
3 Rokhlin property for finite group actions on C ∗-
correspondences
In this section we define and explore the Rokhlin property for a compatible group action
on a C ∗-correspondence when the group is finite.
4
3.1 C ∗-correspondence
Let E be a vector space which is a right module over a C ∗-algebra A and satisfying
α(xa) = (αx)a = x(αa) for x ∈ E, a ∈ A, α ∈ C. The module E is called an (right)
inner-product A-module if there exists a map h·, ·iA : E × E → A such that
(1) hx, xiA ≥ 0 for x ∈ E and hx, xiA = 0 only if x = 0,
(2) hx, yaiA = hx, yiAa for x, y ∈ E and for a ∈ A,
(3) hx, yiA = hy, xi∗
A for x, y ∈ E,
(4) hx, µy + νziA = µhx, yiA + νhx, ziA for x, y, z ∈ E and for µ, ν ∈ C.
An (right) inner-product A-module E is called (right) Hilbert A-module or (right)
Hilbert C ∗-module over A if it is complete with respect to the norm
kxk := khx, xiAk1/2 for x ∈ E.
If there is no ambiguity, we simply write h·, ·i instead of h·, ·iA. The notion of Hilbert C ∗-
module was introduced independently by Paschke [27] and Rieffel [31]. In [16], Kasparov
used Hilbert C ∗-modules as a tool to study a bivariate K-theory for C ∗-algebras. Below
we define the notion of C ∗-correspondences which will play an important role in this
article.
Definition 3.1. Let A be a C ∗-algebra. A right Hilbert A-module E is called a C ∗-
correspondence over A if there exists a ∗-homomorphism φ : A → Ba(E) where Ba(E)
is the set of all adjointable operators on E, which gives a left action of A on E as
ay := φ(a)y for all a ∈ A, y ∈ E.
We use notation (E, A, φ) for the C ∗-correspondence and denote by K(E) the C ∗-algebra
generated by maps {θx,y : x, y ∈ E} defined by
θx,y(z) := xhy, zi for x, y, z ∈ E.
In this article we work with a certain type of action, which we define below, of a
locally compact group on a C ∗-correspondence:
Definition 3.2. Let (G, α, A) be a C ∗-dynamical system of a locally compact group
G and let (E, A, φ) be a C ∗-correspondence over A. An α-compatible action η of G
on E is defined as a group homomorphism from G into the group of invertible linear
transformations on E such that
(1) ηg(φ(a)x) = αg(a)ηg(x) for a ∈ A, x ∈ E, g ∈ G,
(2) hηg(x), ηg(y)i = αg(hx, yi) for x, y ∈ E, g ∈ G,
5
and g 7→ ηg(x) is continuous from G into E for each x ∈ E. We denote an α-compatible
action η by (η, α). In this case, we define a ∗-isomorphism Adηs : Ba(E) → Ba(E) for
each s ∈ G by
Adηs(T )(x) := ηs(T (ηs−1(x)) for T ∈ Ba(E), x ∈ E.
(3.1)
Let G be a locally compact group and △ be the modular function of G. Let η be an
α-compatible action of G on the C ∗-correspondence (E, A, φ). Then the crossed product
E ×η G (cf. [16, 5, 9]) is a Hilbert A ×α G-module and is defined as the completion of
an inner-product Cc(G, A)-module Cc(G, E) where the module action and the Cc(G, A)-
valued inner-product are given by
l · g(s) =ZG
hl, miCc(G,A)(s) =ZG
l(t)αt(g(t−1s))dt,
αt−1(hl(t), m(ts)iA)dt,
(3.2)
(3.3)
respectively for g ∈ Cc(G, A) and l, m ∈ Cc(G, E). For each s ∈ G the ∗-isomorphism
Adηs defined by Equation 3.1 satisfies Adηs(θx,y) = θηs(x),ηs(y) for x, y ∈ E, and
(G, Adη, K(E)) becomes a C ∗-dynamical system. From Definition 3.2 (1) it follows that
φ : A → M(K(E)) is equivariant, i.e.,
φ(αs(a)) = Adηs(φ(a)) for all a ∈ A, s ∈ G.
Indeed, using Adηs we get another ∗-isomorphism Ξ : K(E ×η G) → K(E) ×Adη G (cf.
Section 3.11 of [16] and Section 2 of [9]) defined by
Ξ(θl,m)(s) :=ZG
θl(r),Adηs(m(s−1r)) △ (s−1r)dr where l, m ∈ Cc(G, E), s ∈ G.
From the fact that φ : A → M(K(E)) is equivariant we get an equivariant ∗-homomorphism
χ : A ×α G → M(K(E) ×Adη G) satisfying χ(f ⊗ a) = f ⊗ φ(a) for f ∈ Cc(G), a ∈ A.
We identify K(E ×η G) with K(E) ×Adη G and treat χ and Ξ−1 ◦ χ as same.
3.2 Rokhlin property for compatible finite group actions on
C ∗-correspondences
Definition 3.3. Let (G, α, A) be a C ∗-dynamical system of a finite group G on A and
let (E, A, φ) be a C ∗-correspondence. Let (η, α) be an α-compatible action of G on E.
Then we say that η has the Rokhlin property if for each ǫ > 0, and finite subsets S1 and
S2 of E and A respectively, there exists (ag)g∈G ⊂ A consisting of mutually orthogonal
positive contractions such that
(1) kPg∈G φ(ag)x − xk < ǫ, kPg∈G xag − xk < ǫ,
kPg∈G aga − ak < ǫ and kPg∈G aag − ak < ǫ for all x ∈ S1, a ∈ S2,
(2) kαh(ag) − ahgk < ǫ for h, g ∈ G,
6
(3) kxag − φ(ag)xk < ǫ and kaga − aagk < ǫ for all x ∈ S1, a ∈ S2 and g ∈ G.
The following example is based on the construction of an action of Z2 on C0(X)
where X is equipped with a homeomorphism of order 2 defined on it:
Example 3.4. Let X = { 1
ψ : X → X by
n : n ∈ N} and the topology on X be discrete. Define a map
ψ(1/(2n − 1)) := 1/(2n), ψ(1/(2n)) := 1/(2n − 1) for all n ∈ N.
Observe that ψ is a homeomorphism of order 2. Thus we obtain an automorphism
α : C0(X) → C0(X) such that
α(g)(x) := g(ψ−1(x)) for each x ∈ X, g ∈ C0(X).
Indeed, α2 = idC0(X) and this provides an action of Z2 on C0(X) which we denote by α.
Let H be a Hilbert space and let C0(X, H) be the space of continuous H-valued functions
on X vanishing at infinity. The space C0(X, H) becomes a Hilbert C0(X)-module where
module action and inner product are defined as follow:
f · g(x) := g(x)f (x); hf, f ′i(x) := hf (x), f ′(x)i
for all f, f ′ ∈ C0(X, H), g ∈ C0(X). In fact, C0(X, H) becomes a C ∗-corresponden-
ce over C0(X) with the left action φ defined by
φ(g)f := f · g for all f ∈ C0(X, H), g ∈ C0(X).
Define η : C0(X, H) → C0(X, H) by
η(f )(x) := f (ψ−1(x)) for all x ∈ X, f ∈ C0(X, H).
It follows that hη(f ), η(f ′)i = α(hf, f ′i) for all f, f ′ ∈ C0(X, H). Moreover, η2 =
idC0(X,H) and hence we get an induced α-compatible Z2 action on
(C0(X, H), C0(X), φ) say (η, α). The action (η, α) has the Rokhlin property in the
sense of Definition 3.3: Let a(0)
n ∈ C0(X) be the characteristic functions of the
sets {1/(2k − 1) : 1 ≤ k ≤ n} and {1/(2k) : 1 ≤ k ≤ n}, respectively for each n ∈ N.
Note that these functions are continuous (because the given sets are open), α(a(0)
n ) = a(1)
n ,
and (a(0)
n are
orthogonal. If (en)n∈N is an approximate unit for a C ∗-algebra A and E is a Hilbert
A-module, then (xen)n∈N converges to x for each x ∈ E. Hence (η, α) has the Rokhlin
property, for C0(X) is commutative.
n )n∈N is an approximate unit for C0(X). It is clear that a(0)
n , a(1)
n + a(1)
n and a(1)
For any subset S of a C ∗-algebra, we use symbol S∗ to denote the set {x∗ : x ∈ S}.
Example 3.5. Let l2(A) be the direct sum of a countable number of copies of a C ∗-
algebra A. The vector space l2(A) is known as the standard Hilbert C ∗-module where
the right A-module action and the A-valued inner-product is given by
(a1, a2, . . . , an, . . .)a := (a1a, a2a, . . . , ana, . . .) and
7
h(a1, a2, . . . , an, . . .), (a′
1, a′
2, . . . , a′
n, . . .)i :=
i a′
a∗
i for all a, a1, a′
1, a2, a′
2 . . . ∈ A.
∞Xi=1
It is easy to note that (l2(A), A, φ) is a C ∗-correspondence where the adjointable left
action φ : A → Ba(l2(A)) is defined as
φ(a)(a1, a2, . . . , an, . . .) = (aa1, aa2, . . . , aan, . . .) for all a, a1, a2, . . . ∈ A.
Let (G, α, A) be a finite group action. Define η : G → Aut l2(A) by
ηt(a1, a2, . . . , an, . . .) := (αt(a1), αt(a2), . . . , αt(an), . . .)
where t ∈ G and (a1, a2, . . . , an, . . .) ∈ l2(A). It is clear that η is an α-compatible action
of the group G on (l2(A), A, φ).
Next we show that if α has the Rokhlin property, then η has the Rokhlin property as
an α-compatible action of the group G on the C ∗-correspondence (l2(A), A, φ).
1, aj
2, . . . , aj
Let ǫ > 0 and let S1 = {(aj
such that kPn>N j aj∗
n, . . .) : j = 1, 2, . . . , N} and S2 be finite
subsets of l2(A) and A respectively. Thus for each j, there exist positive integers N j
n : n ≤ N j, 1 ≤ j ≤ N}
and let K = (maxjN j) + 1. Assume that α has the Rokhlin property for the finite set
S = S′
2, i.e., we get Rokhlin elements {fg : g ∈ G} consist of mutually
orthogonal positive contractions in A satisfying the following:
2(G2+2G+1). Fix S′
1 ∪ S2 ∪ S∗
1 := {aj
1 ∪ S′∗
n aj
2 <
nk
ǫ
1
(1) k(Pg∈G fg)a − ak < ǫ
(2) kαh(fg) − fhgk < ǫ
2K for all a ∈ S,
2K for h, g ∈ G,
(3) k[fg, a]k < ǫ
2K for g ∈ G and a ∈ S.
Now we check that the action (η, α) has the Rokhlin property with Rokhlin elements
{ag : g ∈ G} where ag := fg for each g ∈ G: Note that
fgaj
fgaj
1 − aj
φ(ag)(aj
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg∈G
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xg∈G
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
∞Xn=1"Xg∈G
N jXn=1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg∈G
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xn>N j"Xg∈G
fgaj
<
1, aj
2, . . . , aj
n, . . .) − (aj
1, aj
2, . . . , aj
n − aj
1
2
fgaj
fgaj
fgaj
n − aj
2 − aj
2, . . . ,Xg∈G
1,Xg∈G
n#(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n#∗"Xg∈G
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xn>N j"Xg∈G
n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n#∗"Xg∈G
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xn>N j
n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n "Xg∈G
n#∗
fgaj
fgaj
aj∗
aj
1
2
8
n − aj
n − aj
n, . . .)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n, . . .!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n#(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n#(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
fgaj
fgaj
1
2
1
2
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xn>N j
1
2
aj∗
n aj
n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
<
N jXn=1
ǫ
2K
+
ǫ
2
< ǫ, and
k(aj
= k(aj
1, aj
2, . . . , aj
1ag − agaj
n, . . .)ag − φ(ag)(aj
1, aj
2, . . . , aj
2ag − agaj
1, aj
2, . . . , aj
nag − agaj
n, . . .)k
n, . . .)k
<
<
N jXn=1
N jXn=1
kaj
nfg − fgaj
nk +
ǫ
2
ǫ
2K
+
ǫ
2
< ǫ for all (aj
1, aj
2, . . . , aj
n, . . .) ∈ S1 and g ∈ G.
It is easy to check other conditions of the definition of Rokhlin property and hence (η, α)
has the Rokhlin property.
3.3 Rokhlin property for induced actions on Cuntz-Pimsner
algebras
For a C ∗-correspondence, Katsura [17] introduced the following associated C ∗-algebra:
Definition 3.6. Let (E, A, φ) be a C ∗-correspondence over a C ∗-algebra A and B be a
C ∗-algebra.
(1) A pair (π, Ψ) is called covariant representation of (E, A, φ) on B if π : A → B is a
∗-homomorphism and Ψ : E → B is a bounded linear map satisfying
(a) Ψ(x)∗Ψ(y) = π(hx, yi) for all x, y ∈ E,
(b) π(a)Ψ(x) = Ψ(φ(a)x) for all a ∈ A and x ∈ E,
(c) π(b) = ΠΨ(φ(b)) for all b ∈ JE where
JE := φ−1(K(E)) ∩ (kerφ)⊥
and ΠΨ : K(E) → B is a ∗-homomorphism defined by
ΠΨ(θx,y) := Ψ(x)Ψ(y)∗ for x, y ∈ E.
The notation C ∗(π, Ψ) denotes the C ∗-algebra generated by the images of map-
pings π and Ψ in B.
(2) A covariant representation (πU , ΨU ) of a C ∗-correspondence (E, A, φ) is said to be
universal if for any covariant representation (π, Ψ) of (E, A, φ) on B, there exists a
natural surjection ψ : C ∗(πU , ΨU ) → C ∗(π, Ψ) such that π = ψ ◦ πU and Ψ = ψ ◦ ΨU .
We denote the C ∗-algebra C ∗(πU , ΨU ) by OE.
9
In Lemma 2.6 of [9] Hao and Ng proved that each action (η, α) of a locally compact
group G on (E, A, φ) induces a C ∗-dynamical system (G, γ, OE) such that γs(πU (a)) =
πU (αs(a)) and γs(ΨU (x)) = ΨU (ηs(x)) for all a ∈ A, x ∈ E and s ∈ G. The theorem
below shows that Definition 3.3 is the natural choice for Rokhlin property.
Theorem 3.7. Let (η, α) be an action of a finite group G on a C ∗-correspondence
(E, A, φ). The following statements are equivalent:
(a) The action (η, α) has the Rokhlin property.
(b) The induced action γ of G on OE as mentioned above has the Rokhlin property
with Rokhlin elements from πU (A).
Proof. Let ǫ > 0 and let S = {b1, b2, . . . , bn} be any finite subset of OE. For each
1 ≤ j ≤ n there exist finite sets {xl
j }1≤m≤mj ⊂ A such that
kbj − pj(ΨU (xl
3G where
j}1≤l≤lj ⊂ E and {am
j ))k < ǫ
j), πU (am
pj(ΨU (xl
j), πU (am
j )) =
λj,iuj,i,1uj,i,2 . . . uj,i,kj,i
njXi=1
is a finite linear combination of words uj,i,1uj,i,2 . . . uj,i,kj,i in the set
{ΨU (xl
j), ΨU (xl
j)∗, πU (am
j ), πU (am
j )∗ : 1 ≤ l ≤ lj, 1 ≤ m ≤ mj, 1 ≤ j ≤ n}.
exists (ag)g∈G ⊂ A consists of mutually orthogonal positive contractions such that
i=1 λj,ikuj,i,1kkuj,i,2k . . . kuj,i,kj,ik. Set
Kj(cid:19). Since (η, α) has the Rokhlin property, there
1≤j≤n
max
1≤i≤nj
Let S1 = {xl
j}l,j, S2 = {am
j }m,j and Kj = Pnj
K := 3(cid:18) max
kj,i(cid:19)(cid:18) max
K , kPg∈G xag − xk < ǫ
(1) kPg∈G φ(ag)x − xk < ǫ
kPg∈G aga − ak < ǫ
K and kPg∈G aag − ak < ǫ
(2) kαh(ag) − ahgk < ǫ
1≤j≤n
K for h, g ∈ G,
K and kaga − aagk < ǫ
K ,
K for all x ∈ S1, a ∈ S2,
(3) kxag − φ(ag)xk < ǫ
K for all x ∈ S1, a ∈ S2 and g ∈ G.
We show that γ has the Rokhlin property with respect to (fg)g∈G where for each
g ∈ G, fg := πU (ag). For each g ∈ G, we have kfgk ≤ 1, and fg's are mutually
orthogonal positive contractions. Further
(1) For each 1 ≤ j ≤ n,
kPg∈Gfgbj − bjk ≤ kPg∈GπU (ag)bj −Pg∈GπU (ag)pj(ΨU (xl
+ kPg∈GπU (ag)pj(ΨU (xl
+ kpj(ΨU (xl
j), πU (am
j), πU (am
j )) − bjk < ǫ.
j), πU (am
j ))k
j )) − pj(ΨU (xl
j), πU (am
j ))k
(2) For h, g ∈ G we have
kγh(πU (ag)) − πU (ahg)k = kπU (αh(ag)) − πU (ahg)k < ǫ.
10
(3) For 1 ≤ j ≤ n we get
kπU (ag)bj − bjπU (ag)k
= kπU (ag)bj − πU (ag)pj(ΨU (xl
+ kπU (ag)pj(ΨU (xl
+ kpj(ΨU (xl
j), πU (am
j), πU (am
j ))k
j )) − pj(ΨU (xl
j), πU (am
j ))πU (ag) − bjπU (ag)k < ǫ.
j), πU (am
j ))πU (ag)k
Thus γ has the Rokhlin property with Rokhlin elements from the C ∗-algebra πU (A).
Conversely, let S1 and S2 be finite subsets of E and A respectively. Fix ǫ > 0 and
S := {ΨU (x), πU (hx, xi), πU (a), πU (a∗) : x ∈ S1, a ∈ S2}.
Since γ has the Rokhlin property with Rokhlin elements from the C ∗-algebra πU (A),
there exist mutually orthogonal positive contractions (πU (fg))g∈G ⊂ πU (A) satisfying
(1) k(Pg∈G πU (fg))a − ak < ǫ′ for all a ∈ S,
(2) kαh(πU (fg)) − πU (fhg)k < ǫ′ for h, g ∈ G,
(3) k[πU (fg), a]k < ǫ′ for g ∈ G and a ∈ S,
where ǫ′ = min{ǫ,
Rokhlin elements (fg)g∈G:
ǫ2
G+1}. We show below that (η, α) has the Rokhlin property with
(1) For each x ∈ S1, a ∈ S2 we get
< ǫ,
2
φ(fg)x − x(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg∈G
xfg − x(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg∈G
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
φ(fg)x − x!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ΨU Xg∈G
πU (fg)ΨU (x) − ΨU (x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg∈G
hxfg − x, xfg′ − xi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg,g′∈G
fghx, xi − hx, xi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤Xg′∈G(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg∈G
G + 1(cid:19) = ǫ2,
<(G + 1)(cid:18) ǫ2
fga − ak =kπU (Xg∈G
afg − ak =kπU (Xg∈G
kXg∈G
kXg∈G
fga − a)k < ǫ and
afg − a)k < ǫ;
kfg′k +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xg∈G
fghx, xi − hx, xi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(2) kαh(fg) − fhgk = kπU (αh(fg) − fhg)k = kγh(πU (fg)) − πU (fhg)k < ǫ for h, g ∈ G;
11
(3) For all x ∈ S1, a ∈ S2 and g ∈ G we obtain
kxag − φ(fg)xk = kΨU (xfg − φ(fg)x)k
= kΨU (x)πU (fg) − πU (fg)ΨU (x)k < ǫ
and
kfga − afgk = kπU (fga − afg)k = kπU (fg)πU (a) − πU (a)πU (fg)k < ǫ.
Hence (η, α) has the Rokhlin property with Rokhlin elements (fg)g∈G.
3.4 Applications of our characterization
Katsura obtained several results about the nuclearity of the C ∗-algebra OE associated
to a C ∗-correspondence E in [19]. We discuss permanence properties of this notion and
several other notions for the C ∗-algebra, associated to a C ∗-correspondence, with respect
to the crossed product E ×η G of a C ∗-correspondence E for some action (η, α) of a finite
group G with Rokhlin property. The nuclear dimension of OE is estimated recently in
[4, Corollary 5.22].
Corollary 3.8. Assume (η, α) to be an action of a finite group G on a C ∗-correspondence
(E, A, φ). If (η, α) has the Rokhlin property and if OE belongs to any one of the classes,
say C, listed in Theorem 2.3, then OE×ηG also belongs to the same class C.
Proof. Suppose action (η, α) has the Rokhlin property. By Theorem 3.7 the induced
action γ of G on OE has the Rokhlin property. Since OE belongs to the class C from
Theorem 2.3, from the remarks made just before this corollary it follows that OE ×γ G
also belongs to the same class. The C ∗-algebras OE ×γ G and OE×ηG has been shown
to be isomorphic in [9, Theorem 2.10]. Hence OE×ηG also belongs to the same class.
A directed graph E = (E 0, E 1, r, s) consists of a countable vertex set E 0, and a count-
able edge set E 1, along with maps r, s : E 1 → E 0 describing the range and the source of
edges. We also assume that the directed graph is always row finite, i.e., for every vertex
v ∈ E0, the set s−1(v) is a finite subset of E1. Let A denote the C ∗-algebra C0(E 0). A
graph C ∗-algebra of the directed graph E (cf.
[22]) is a universal C ∗-algebra generated
by partial isometries {Se : e ∈ E 1} and projections {Pv : v ∈ E 0} such that
Se
∗Se = Pr(e) = Xs(f )=r(e)
Sf Sf
∗ for all e ∈ E 1.
Since the graph is row finite, the summation is finite. We use the symbol C ∗(E) to
denote the graph C ∗-algebra of a directed graph E. The vector space Cc(E 1) becomes
an inner-product A-module with the following inner-product and module action:
:= Xe∈r−1(v)
hf, gi(v)
(f h)(e)
f (e)g(e) for each v ∈ E 0;
:= f (e)h(r(e)) for all e ∈ E 1;
12
where f, g ∈ Cc(E 1) and h ∈ A. Let E(E) denote the completion of the inner-product
module Cc(E 1). Define φ : A → Ba(E(E)) by
φ(h)f (e) := h(s(e))f (e) for each e ∈ E 1; f ∈ Cc(E 1); h ∈ A.
Thus (E(E), A, φ) is a C ∗-correspondence and the graph C ∗-algebra C ∗(E) of a directed
graph E is always isomorphic to OE(E) (cf. [17, Proposition 3.10]).
Definition 3.9. Let E = (E 0, E 1, r, s) be a directed graph and let c from E 1 to a countable
group G be a mapping. The skew-product graph is the graph E(c) = (G×E 0, G×E 1, r′, s′)
where r′(g, e) := (gc(e), r(e)) and s′(g, e) := (g, s(e)) for all g ∈ G; e ∈ E 1.
For a given countable abelian group G and a function c : E 1 → G, we can define an
γc
action γc of bG on C ∗(E) (cf. [22, Corollary 2.5]) by
Let α be the trivial action of bG on A and let η be an action of bG on E(E) defined by
χ(Se) := hχ, c(e)iSe for each χ ∈ bG, e ∈ E 1.
ηχ(f )(e) := hχ, c(e)if (e) for each χ ∈ bG, e ∈ E 1, f ∈ Cc(E 1).
From [9, Corollary 2.11] it follows that γc coincides with the action of G on OE(E) induced
by the action (η, α).
Proposition 3.10. Let E = (E 0, E 1, r, s) be a directed graph, G be a finite abelian group
kαg(fs) − fgsk = kfs − fgsk = kfs + fgsk
E(E) defined in the previous paragraph. Then (η, α) does not have the Rokhlin property.
and c : E 1 → G be a function. Let (η, α) be an action of bG on the C ∗-correspondence
Proof. Since α is the trivial action of bG on A, we have
≤Xs∈ bG
≥ max {kfsk, kfgsk} for all s, g ∈ bG;
1 = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xs∈ bG
Thus α does not have the Rokhlin property. It follows that the condition (2) in Definition
3.3 is not satisfied, and hence the α-compatible action (η, α) does not have the Rokhlin
property.
Corollary 3.11. Let E = (E 0, E 1, r, s) be a directed graph, G be a finite abelian group
fs(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
kfsk .
Rokhlin elements from πU (A).
and c : E 1 → G be a function. Let (η, α) be an action of bG on the C ∗-correspondence
E(E). Then the induced action γc of bG on C ∗(E) does not have the Rokhlin property with
Proof. Suppose that the induced action γc of bG on C ∗(E) has the Rokhlin property with
Rokhlin elements from πU (A). From [9, Corollary 2.11] we have
γc
χ(πU (δv)) = πU (αχ(δv)) = πU (δv) for v ∈ E 0,
γc
χ(ΨU (δe)) = hχ, c(e)i(ΨU (δe)) = ΨU (ηχ(δe)) for e ∈ E 1,
for χ ∈ bG. Then Rokhlin elements belong to the fixed point algebra C ∗(E(c))γc. There-
fore the action γc does not satisfy the condition (2) in Definition 2.1, and we have a
contradiction.
13
4 Rokhlin property for group actions on Hilbert bi-
modules
Analogous to a right Hilbert A-module, a left Hilbert A-module is defined as a left A-
module with the positive definite form Ah·, ·i : E × E → A which is, conjugate-linear in
the second variable, linear in the first variable and we have
Ahax, yi = aAhx, yi for x, y ∈ E, a ∈ A.
Definition 4.1. Let A and B be two C ∗-algebras. A left Hilbert B-module E is called
Hilbert B-A bimodule if it is also a right Hilbert A-module satisfying
Bhx, yiz = xhy, ziA for x, y, z ∈ E.
On a Hilbert bimodule we consider actions of a locally compact group similar to
those introduced in Definition 3.2.
Definition 4.2. Let (G, α, A) and (G, β, B) be C ∗-dynamical systems of a locally com-
pact group G and let E be a B-A Hilbert bimodule. A β-compatible action (respectively
an α-compatible action) η of G on E is defined as a group homomorphism from G into
the group of invertible linear transformations on E such that
(1) ηg(bx) = βg(b)ηg(x) (respectively ηg(xa) = ηg(x)αg(a)),
(2) Bhηg(x), ηg(y)i = βg(Bhx, yi) (respectively hηg(x), ηg(y)iA = αg(hx, yiA))
for a ∈ A, b ∈ B, x, y ∈ E, g ∈ G; and g 7→ ηg(x) is continuous from G into E for
each x ∈ E. The combination of these two compatibility conditions will be simply called
(β, α)-compatibility.
Consider a (β, α)-compatible action η of a locally compact group G on the B-A
[16, 5]) is an B ×β G-
Hilbert bimodule E. The crossed product bimodule E ×η G (cf.
A ×α G Hilbert bimodule obtained by completion of Cc(G, E) such that
(lg)(s) =ZG
l(t)αt(g(t−1s))dt,
f (t)ηt(m(t−1s))dt,
(f m)(s) =ZG
αt−1(hl(t), m(ts)iA)dt,
Bhl(st−1), ηs(m(t−1))idt
hl, miA×α G(s) =ZG
B×β Ghl, mi(s) =ZG
for all f ∈ Cc(G, B), g ∈ Cc(G, A) and l, m ∈ Cc(G, E).
If E is a (right) Hilbert A-module, then E is a K(E)-A Hilbert bimodule with respect
to K(E)hx, yi = θx,y for all x, y ∈ E. Moreover, we can associate a C ∗-algebra called the
linking algebra, defined by
LE :=(cid:18)K(E) E
E∗ A(cid:19) ⊂ K(E ⊕ A)
14
(4.1)
(cf. [30, p. 50]), to each (right) Hilbert A-module E. Let (G, α, A) be a C ∗-dynamical
system and η be an α-compatible action of G on E. For each s ∈ G, let us define
s (x∗) := ηs(x)∗ for x ∈ E. Indeed, η is also
Adηs(t) := ηstηs−1 for t ∈ K(E) where η∗
an (Adη, α)-compatible action and we get the induced action θ of G on LE (cf. [5, 23])
defined by
y∗ a(cid:19) :=(cid:18)Adηst ηs(x)
θs(cid:18) t
s (y∗) αs(a)(cid:19) .
η∗
x
for all s ∈ G, t ∈ K(E), a ∈ A and x, y ∈ E. We denote this C ∗-dynamical system by
(G, θ, LE).
4.1 Rokhlin property for induced finite group actions on link-
ing algebras
Analogous to Definition 3.3 the Rokhlin property for finite group actions on Hilbert
bimodules is defined as follows:
Definition 4.3. Let (G, α, A) and (G, β, B) be C ∗-dynamical systems of a finite group
G and let E be a B-A Hilbert bimodule. Assume η to be a (β, α)-compatible action of G
on E. We say that η has the Rokhlin property if for each ǫ > 0 finite subsets S1 and S2
of E, and finite subsets S3 and S4 of B and A respectively, there are sets (ag)g∈G ⊂ A
and (bg)g∈G ⊂ B consisting of mutually orthogonal positive contractions such that
1. kPg∈G agu − uk < ǫ for all u ∈ S∗
2 ∪ S4, kPg∈G bgv − vk < ǫ for all v ∈ S1 ∪ S3.
2. kαh(ag) − ahgk < ǫ and kβh(bg) − bhgk < ǫ for h, g ∈ G,
3. kxag − bgxk < ǫ, ktbg − bgtk < ǫ and
kaga − aagk < ǫ, kagy∗ − y∗bgk < ǫ for all x ∈ S1, y ∈ S2, t ∈ S3, a ∈ S4 and
g ∈ G.
Following theorem justifies the choice of this version of Rokhlin property for group
actions on a bimodule:
Theorem 4.4. Let E be a Hilbert A-module where A is C ∗-algebra. Suppose α : G →
Aut(A) is an action of a finite group G and η is an α-compatible action of G on E.
Then the following statements are equivalent:
(a) η has the Rokhlin property as an (Adη, α)-compatible action.
(b) The action θ of G on LE induced by η has the Rokhlin property with Rokhlin
elements coming from the C ∗-subalgebra(cid:18)K(E)
0
0
A(cid:19) of LE.
Proof. Let ǫ > 0 be given and S = (cid:26)(cid:18) ti xi
y∗
i
i(cid:19) : i = 1, 2, . . . , n(cid:27) be any finite subset
of LE. Consider S1 = {x1, x2, . . . , xn}, S2 = {y1, y2, . . . , yn}, S3 = {t1, t2, . . . , tn} and
S4 = {a′
n}. Suppose η has the Rokhlin property as an (Adη, α)-compatible
action, there are sets (ag)g∈G ⊂ A and (bg)g∈G ⊂ K(E) consisting of mutually orthogonal
positive contractions such that
2, . . . , a′
1, a′
a′
15
y∗
i
(1) (cid:13)(cid:13)(cid:13)(cid:13)Pg∈G fg(cid:18) ti xi
=(cid:13)(cid:13)(cid:13)(cid:13)Pg∈G(cid:18)bg
a′
0
(2) For t, g ∈ G, we have
(1) kPg∈G agu − uk < ǫ
(2) kαh(ag) − ahgk < ǫ
4 for all u ∈ S∗
4 and kβh(bg) − bhgk < ǫ
2 ∪ S4, kPg∈G bgv − vk < ǫ
4 for h, g ∈ G,
4 for all v ∈ S1 ∪ S3.
(3) kxag − bgxk < ǫ
kaga − aagk < ǫ
g ∈ G.
4, ktbg − bgtk < ǫ
4 and
4, kagy∗ − y∗bgk < ǫ
4 for all x ∈ S1, y ∈ S2, t ∈ S3, a ∈ S4 and
We prove that the action θ of G on LE induced by η has the Rokhlin property with
respect to (fg)g∈G where
g, h ∈ G with g 6= h we get
fg :=(cid:18)bg
0
0 ag(cid:19). For each g ∈ G, kfgk = supk(x,a)k≤1(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)bg
0 ah(cid:19) = 0.
0
0
Now we verify conditions (1)-(3) of Definition 2.1:
0
0 ag(cid:19)(cid:18)x
a(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1. Further for
0
a′
a′
y∗
i
y∗
i
y∗
i
0 ag(cid:19)(cid:18)bh
fgfh =(cid:18)bg
i(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
i(cid:19) −(cid:18) ti xi
0 ag(cid:19)(cid:18) ti xi
i(cid:19) −(cid:18) ti xi
kθh(fg) − fhgk =(cid:13)(cid:13)(cid:13)(cid:13)θh(cid:18)bg
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)ηhbgη∗
i(cid:19)(cid:21)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)bg
0 ag(cid:19)(cid:18) ti xi
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:18) bgti
y∗ 0(cid:19) ,(cid:18)0 0
0 0(cid:19) ,(cid:18) 0
agxi
aga′
agy∗
i
y∗
i
0
0
0
0
h
0
0
0
0
a′
i(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < ǫ.
ahg(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
0 ag(cid:19) −(cid:18)bhg
ahg(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < ǫ.
αg(ag)(cid:19) −(cid:18)bhg
0 ag(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
i(cid:19)(cid:18)bg
iag(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < 4(ǫ/4) = ǫ.
0 a(cid:19) : x ∈ S1, y ∈ S2, t ∈ K(E), a ∈ A(cid:27) .
0 ag(cid:19) : g ∈ G(cid:27)
i(cid:19) −(cid:18) ti xi
i(cid:19) −(cid:18) tibg xiag
y∗
i bg a′
y∗
i
a′
a′
0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:20)fg,(cid:18) ti xi
y∗
i
a′
(3) For each g ∈ G and 1 ≤ i ≤ n, we get
Conversely, let ǫ > 0 and let S1, S2 be finite subsets of E. Let S3 and S4 be finite
subsets of K(E) and A, respectively. Choose
S =(cid:26)(cid:18)t 0
0 0(cid:19) ,(cid:18)0 x
Suppose θ has the Rokhlin property with Rokhlin elements(cid:26)(cid:18)bg
⊂ (cid:18)K(E)
A(cid:19) ⊂ LE with respect to the set S and ǫ > 0. Then it is easy to verify
0
0
0
16
that η has the Rokhlin property as an (Adη, α)-compatible action with the positive
contractions (ag)g∈G ⊂ A and (bg)g∈G ⊂ K(E) with respect to the sets S1, S2, S3 and
S4 on noting that the steps and arguments of the first part of this proof can be carried
out in the reverse order.
Let E be a Hilbert A-module. There are several examples of C ∗-algebras A for
which nuclear dimension of LE is at most n. One way to obtain it is by considering
A with nuclear dimension at most n, because from the fact that A is a full hereditary
C ∗-subalgebra of LE it follows that the nuclear dimension of LE is at most n (see [37,
Corollary 2.8]).
Corollary 4.5. Let (G, α, A) be a C ∗-dynamical system where G is finite group and
n ∈ N ∪ {0}. Let E be a Hilbert A-module. Assume η to be an action of G on E which
has the Rokhlin property as an (Adη, α)-compatible action and A belongs to any one of
the classes, say C, listed in Theorem 2.3, then LE×ηG also belongs to the same class C.
Proof. It follows from Theorem 4.4 that the induced action θ on LE has the Rokhlin
property. Since A is a full hereditary subalgebra of LE, the linking algebra LE belongs to
the class C from Theorem 2.3, then LE ×θ G also belongs to the same class. We identify
LE ×θ G and LE×ηG (cf. the proof of [5, Theorem 4.1]). Hence LE×ηG also belongs to
the same class.
4.2 Rokhlin property for induced actions of Z on linking alge-
bras
Next we investigate the case where action is of the infinite discrete group Z over a Hilbert
bimodule E instead of action of finite group on E. Indeed, we use notation η for η1 and
α for α1. We first recall the definition of the Rokhlin property for the automorphisms
on C ∗-algebras from [2].
Definition 4.6. Let A be a C ∗-algebra and α ∈ Aut(A). We say that α has the Rokhlin
property if for any positive integer p, any finite set S ⊂ A, and any ǫ > 0, there are
mutually orthogonal positive contractions e0,0, . . . , e0,p−1, e1,0, . . . , e1,p such that
r=0Pp−1+r
j=0
(2) k[er,j, a]k < ǫ for all r, j and a ∈ S,
(1) (cid:13)(cid:13)(cid:13)(cid:16)P1
er,j(cid:17) a − a(cid:13)(cid:13)(cid:13) < ǫ for all a ∈ S,
(3) kα(er,j)a − er,j+1ak < ǫ for all a ∈ S, r = 0, 1 and j = 0, 1, . . . , p − 2 + r,
(4) kα(e0,p−1 + e1,p)a − (e0,0 + e1,0)ak < ǫ for all a ∈ S.
We call elements e0,0, . . . , e0,p−1, e1,0, . . . , e1,p the Rokhlin elements for α.
Definition 4.7. Let (Z, α, A) and (Z, β, B) be C ∗-dynamical systems. Suppose E is a
Hilbert B-A bimodule and η is a (β,α)-compatible automorphism of E. We say that η
has the Rokhlin property if for any ǫ > 0, any positive integer p, any finite subsets S1
and S2 of E, and any finite subsets S3 and S4 of B and A respectively, there are sets
consisting of mutually orthogonal positive contractions {a0,0, . . . , a0,p−1, a1,0, . . . , a1,p} ⊂
A and {b0,0, . . . , b0,p−1, b1,0, . . . , b1,p} ⊂ B such that
17
1. bi,jbi′,j ′ = 0 and ai,jai′,j ′ = 0 if (i, j) 6= (i′, j′).
2. (cid:13)(cid:13)(cid:13)P1
u ∈ S∗
r=0Pp−1+r
j=0
2 ∪ S4.
br,jv − v(cid:13)(cid:13)(cid:13) < ǫ, (cid:13)(cid:13)(cid:13)P1
r=0Pp−1+r
j=0
ar,ju − u(cid:13)(cid:13)(cid:13) < ǫ for all v ∈ S1 ∪ S3,
3. kxar,j − br,jxk < ǫ, kar,jy∗ − y∗br,jk < ǫ, kbbr,j − br,jbk < ǫ, kar,ja − aar,jk < ǫ
for all x ∈ S1, y ∈ S2, b ∈ S3, a ∈ S4 and for all r, j.
4. kβ(br,j)v − br,j+1vk < ǫ , kα(ar,j)u − ar,j+1uk < ǫ
for all v ∈ S1 ∪ S3, u ∈ S∗
2 ∪ S4; r = 0, 1 and j = 0, . . . , p − 2 + r.
5. kβ(b0,p−1 + b1,p)v − (b0,0 + b1,0)vk < ǫ,
kα(a0,p−1 + a1,p)u − (a0,0 + a1,0)uk < ǫ for all v ∈ S1 ∪ S3, u ∈ S∗
2 ∪ S4.
The following observation is a justification for the choice of the above definition of
Rokhlin property for actions of Z on a bimodule:
Theorem 4.8. Suppose (Z, α, A) is a C ∗-dynamical system. Assume E to be a Hilbert
A-module and η to be an automorphism on E. The following statements are equivalent:
(a) η has the Rokhlin property as an (Adη,α)-compatible automorphism.
(b) The automorphism θ in Aut(LE) induced by η has the Rokhlin property with
Rokhlin elements coming from the C ∗-subalgebra(cid:18)K(E)
0
0
A(cid:19) of LE.
Proof. Let ǫ > 0 be given and S = (cid:26)(cid:18) ti xi
ai(cid:19) : i = 1, 2, . . . , n(cid:27) be any finite subset
of LE. Consider S1 = {x1, x2, . . . , xn}, S2 = {y1, y2, . . . , yn}, S3 = {t1, t2, . . . , tn} and
S4 = {a1, a2, . . . , an}. Suppose η has the Rokhlin property, there are sets consist of
mutually orthogonal positive contractions {a0,0, . . . ,
a0,p−1, a1,0, . . . , a1,p} ⊂ A and {b0,0, . . . , b0,p−1, b1,0, . . . , b1,p} ⊂ K(E) such that
y∗
i
(1) bi,jbi′,j ′ = 0 and ai,jai′,j ′ = 0 if (i, j) 6= (i′, j′).
(2) (cid:13)(cid:13)(cid:13)P1
u ∈ S∗
r=0Pp−1+r
j=0
2 ∪ S4.
br,jv − v(cid:13)(cid:13)(cid:13) < ǫ
4, (cid:13)(cid:13)(cid:13)P1
r=0Pp−1+r
j=0
ar,ju − u(cid:13)(cid:13)(cid:13) < ǫ
4 for all v ∈ S1 ∪ S3,
(3) kxar,j − br,jxk < ǫ
4, kar,jy∗ − y∗br,jk < ǫ
4, kbbr,j − br,jbk < ǫ
4 , kar,ja − aar,jk < ǫ
4
for all x ∈ S1, y ∈ S2, b ∈ S3, a ∈ S4 and for all r, j.
(4) kβ(br,j)v − br,j+1vk < ǫ
4 , kα(ar,j)u − ar,j+1uk < ǫ
4
for all v ∈ S1 ∪ S3, u ∈ S∗
2 ∪ S4; r = 0, 1 and j = 0, . . . , p − 2 + r.
(5) kβ(b0,p−1 + b1,p)v − (b0,0 + b1,0)vk < ǫ
4,
kα(a0,p−1 + a1,p)u − (a0,0 + a1,0)uk < ǫ
4 for all v ∈ S1 ∪ S3, u ∈ S∗
2 ∪ S4.
18
We verify that the action θ of G on LE induced by η has the Rokhlin property as an
(Adη, α)-compatible automorphism with respect to e0,0, . . . , e0,p−1, e1,0, . . . ,
e1,p where ei,j :=(cid:18)bi,j
0
For (i, j) 6= (i′, j′) we have
ei,jei′,j ′ =(cid:18)bi,j
0
0
kei,jk = sup
ai,j(cid:19). For each (i, j),
k(x,a)k≤1(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)bi,j
ai,j(cid:19)(cid:18)bi′,j ′
0
0
0
0
0
ai,j(cid:19)(cid:18)x
a(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1.
ai′,j ′(cid:19) =(cid:18)bi,jbi′,j ′
0
0
ai,jai′,j ′(cid:19) = 0.
We verify conditions (1)-(4) of Definition 4.6 below:
(1) For 1 ≤ i ≤ n we get
(2) For all r, j we have
(4) For 1 ≤ i ≤ n, we get
(3) For r = 0, 1 and j = 0, . . . , p − 2 + r, we have
ai(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < 4 × ǫ/4 = ǫ.
0
ar,j(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < 4 × ǫ/4 = ǫ.
y∗
i
ai(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
ai(cid:19) −(cid:18) ti xi
ar,j(cid:19)(cid:19)(cid:18) ti xi
ai(cid:19) −(cid:18) ti xi
y∗
i
y∗
i
0
0
j=0
y∗
i
y∗
i
(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)P1
er,j(cid:17)(cid:18) ti xi
r=0Pp−1+r
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)P1
j=0 (cid:18)br,j
r=0Pp−1+r
ai(cid:19)(cid:21)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:20)er,j,(cid:18) ti xi
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)br,j
ar,j(cid:19)(cid:18) ti xi
(cid:13)(cid:13)(cid:13)(cid:13)θ(er,j)(cid:18) ti xi
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)η(br,j)η∗
y∗
i
y∗
i
0
0
0
0
0
y∗
i
y∗
i
ai(cid:19)(cid:18)br,j
ai(cid:19) −(cid:18) ti xi
ai(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
ai(cid:19) − er,j+1(cid:18) ti xi
ai(cid:19) −(cid:18)br,j+1
α(ar,j)(cid:19)(cid:18) ti xi
(cid:13)(cid:13)(cid:13)(cid:13)θ(e0,p−1 + e1,p)(cid:18) ti xi
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)η(b0,p−1 + b1,p)η∗
−(cid:18)b0,0 + b1,0
a0,0 + a1,0(cid:19)(cid:18) ti xi
y∗
i
y∗
i
y∗
i
0
0
0
0
0
y∗
i
ai(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < ǫ.
ar,j+1(cid:19)(cid:18) ti xi
ai(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
ai(cid:19) − (e0,0 + e1,0)(cid:18) ti xi
α(a0,p−1 + a1,p)(cid:19)(cid:18) ti xi
ai(cid:19)
ai(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < ǫ.
y∗
i
y∗
i
0
19
Conversely, let S1 ∪ S2, S3 and S4 be finite subsets of E, K(E) and A, respectively.
Let
S =(cid:26)(cid:18)t 0
0 0(cid:19) ,(cid:18)0 x
0 0(cid:19) ,(cid:18) 0
y∗ 0(cid:19) ,(cid:18)0 0
0 a(cid:19) : x ∈ S1, y ∈ S2, t ∈ K(E), a ∈ A(cid:27) .
0
0
0
0
0
0
Suppose θ has the Rokhlin property with Rokhlin elements(cid:18)b0,0
a0,1(cid:19) , . . . ,(cid:18)b0,p−1
(cid:18)b0,1
from the C ∗-algebra(cid:18)K(E)
a1,p(cid:19) coming
a0,p−1(cid:19) , (cid:18)b1,0
A(cid:19) ⊂ LE with respect to the set S and any ǫ > 0. Then
a0,0(cid:19) ,
a1,1(cid:19) , . . . , (cid:18)b1,p
it is easy to check that η has the Rokhlin property as an (Adη, α)-compatible action
with the positive contractions {a0,0, . . . , a0,p−1, a1,0, . . . , a1,p} ⊂ A and {b0,0, . . . , b0,p−1,
b1,0, . . . , b1,p} ⊂ K(E) with respect to the sets S1 ∪ S2, S3 and S4 on observing that
the steps and arguments of the first part of this proof can be carried out in the reverse
order.
a1,0(cid:19) ,(cid:18)b1,1
0
0
0
0
0
0
0
0
0
We recall the definition of D-absorbing (cf. [11]).
Definition 4.9. A separable, unital C ∗-algebra D ≇ C is strongly self-absorbing if there
exists an isomorphism ϕ : D → D ⊗ D such that ϕ and idD ⊗ 1D are approximately
unitarily equivalent ∗-homomorphisms. If D is a strongly self-absorbing C ∗-algebra, we
say that a C ∗-algebra A is D-absorbing if A ∼= A ⊗ D.
In the following we observe a permanence property of the D-absorbing property with
respect to the crossed product E ×η Z of a bimodule E for an (Adη, α)-compatible action
with Rokhlin property:
Corollary 4.10. Assume (Z, α, A) to be a C ∗-dynamical system. Let η be an α-
compatible automorphism on a Hilbert A-module E and let D be a strongly self-absorbing
C ∗-algebra. If η ∈ Aut(E) has the Rokhlin property as an (Adη, α)-compatible action
and let A be separable and D-absorbing, then LE×ηZ is D-absorbing.
Proof. By Theorem 4.8 the induced automorphism θ on LE has the Rokhlin property.
Since A is a full hereditary C ∗-subalgebra of LE, LE is separable and D-absorbing,
LE ×θ Z is D-absorbing (cf. [2, Theorem 1.9]). We identify LE ×θ Z and LE×ηZ (cf. the
proof of [5, Theorem 4.1]) and hence LE×ηZ is D-absorbing.
5 Outerness for group actions on Hilbert bimodules
In this section we define and explore outer actions of a locally compact group on a
Hilbert bimodule.
Definition 5.1. Let (G, α, A) and (G, β, B) be unital C ∗-dynamical systems of a locally
compact group G and let E be a B-A Hilbert bimodule. Let u and u′ be two unitaries in
20
A and B, respectively. Define an (Ad(u′), Ad(u))-compatible automorphism Ad(u′, u) :
E → E by
Ad(u′, u)(x) = u′∗xu for each x ∈ E.
Let η be an (β, α)-compatible action of G on E. We say that η is outer if for each
t ∈ G \ {e} we have ηt 6= Ad(u′, u) for any unitaries u ∈ A and u′ ∈ B, where e denotes
the identity of G.
In a B-A Hilbert bimodule, it was pointed out in [3, p. 1152] that
Bhxa, yi = Bhx, ya∗i and hbx, yiA = hx, b∗yiA for all x, y ∈ E; a ∈ A; b ∈ B.
Indeed, using these conditions in the following computations we show that
Ad(u′, u) is an (Ad(u′), Ad(u))-compatible automorphism:
(1) For x ∈ E, a ∈ A and b ∈ B we get
Ad(u′, u)(x)Ad(u)(a) = u′∗xuu∗au = u′∗xau = Ad(u′, u)(xa),
Ad(u′)(b)Ad(u′, u)(x) = u′∗bu′u′∗xu = u′∗bxu = Ad(u′, u)(bx).
(2) For each x, y ∈ E we have
hAd(u′, u)(x), Ad(u′, u)(y)iA = hu′∗xu, u′∗yuiA = hxu, u′u′∗yuiA = hxu, yuiA
= u∗hx, yiAu = Ad(u)(hx, yiA),
BhAd(u′, u)(x), Ad(u′, u)(y)i = Bhu′∗xu, u′∗yui = Bhu′∗x, u′∗yuu∗i = Bhu′∗x, u′∗yi
= u′∗
Bhx, yiu′ = Ad(u′)(Bhx, yi).
We check our definition of outerness for group actions on Hilbert bimodules is com-
patible with the definition of outerness for group actions on C ∗-algebras as follows: If
A = B, then A naturally becomes an A-A Hilbert bimodule. In this case we fix (G, α, A)
and take η = β = α. If η is not outer, then ηs = Ad(u′, u) for some s ∈ G, u ∈ A, u′ ∈ B.
Further from the above computations, this gives u = u′ and βs = αs = Ad(u). Thus α
can not be outer. Hence α is outer implies η is outer.
The following proposition says that outerness for group actions on Hilbert bimodules
is a weaker notion than outerness for group actions on C ∗-algebras.
Proposition 5.2. Suppose (G, α, A) and (G, β, B) are unital C ∗-dynamical systems of a
locally compact group G and E is a B-A Hilbert bimodule. Let η be an (β, α)-compatible
action of G on E. If E is full with respect to both the inner products, and if α or β is
outer, then η is outer.
Proof. Let s ∈ G \ {e} and let us and u′
each x, y ∈ E we get
s be unitaries in A. If ηs = Ad(u′
s, us), then for
kαs(hx, yiA) − u∗
shx, yiAusk = khηs(x), ηs(y)iA − u∗
shx, yiAusk
21
= khηs(x), ηs(y)iA − hxus, yusiAk
= khηs(x), ηs(y)iA − hxus, u′
= khηs(x), ηs(y)iA − hu′∗
su′∗
s xus, u′∗
s yusiAk
s yusiAk = 0.
s for each x, y ∈
Similarly if ηs = Ad(u′
E.
s, us), then we obtain βs(Bhx, yi) = u′∗
s (Bhx, yi)u′
Corollary 5.3. Suppose (G, α, A) and (G, β, B) are unital C ∗-dynamical systems of a
finite group G and E is a B-A Hilbert bimodule. Let η be an (β, α)-compatible action
of G on E. If E is full with respect to both the inner products, and if η has the Rokhlin
property, then η is outer.
Proof. Since η is an (β, α)-compatible action of the finite group G on E with the Rokhlin
property, it follows from Definition 4.3 that β and α have the Rokhlin property. This im-
plies that β and α are outer (cf. [34, Proposition 2]). Therefore η is outer by Proposition
5.2.
Acknowledgements: The first author was supported by Seed Grant from IRCC, IIT
Bombay, the second author was supported by JSPS KAKENHI Grant Number 26400125,
and the third author was supported by CSIR, India.
References
[1] Beatriz Abadie, Søren Eilers, and Ruy Exel, Morita equivalence for crossed products
by Hilbert C ∗-bimodules, Trans. Amer. Math. Soc. 350 (1998), no. 8, 3043 -- 3054.
MR1467459 (98k:46109)
[2] Jonathan Brown and Ilan Hirshberg, The Rokhlin property for endomorphisms and
strongly self-absorbing C ∗-algebras, Illinois J. Math. 58 (2014), no. 3, 619 -- 627.
MR3395953
[3] Lawrence G. Brown, James A. Mingo, and Nien-Tsu Shen, Quasi-multipliers and
embeddings of Hilbert C ∗-bimodules, Canad. J. Math. 46 (1994), no. 6, 1150 -- 1174.
MR1304338 (95k:46091)
[4] N. P. Brown, Aaron Tikuisis, and A. M. Zelenberg, Rokhlin dimension for C ∗-
correspondences, http://www.math.psu.edu/nbrown/RokDim.pdf (2015).
[5] Siegfried Echterhoff, S. Kaliszewski, John Quigg, and Iain Raeburn, Naturality
and induced representations, Bull. Austral. Math. Soc. 61 (2000), no. 3, 415 -- 438.
MR1762638 (2001j:46101)
[6] George A. Elliott, On the classification of C ∗-algebras of real rank zero, J. Reine
Angew. Math. 443 (1993), 179 -- 219. MR1241132 (94i:46074)
22
[7] George A. Elliott and Andrew S. Toms, Regularity properties in the classifica-
tion program for separable amenable C ∗-algebras, Bull. Amer. Math. Soc. (N.S.)
45 (2008), no. 2, 229 -- 245. MR2383304 (2009k:46111)
[8] Neal J. Fowler, Marcelo Laca, and Iain Raeburn, The C ∗-algebras of infinite graphs,
Proc. Amer. Math. Soc. 128 (2000), no. 8, 2319 -- 2327. MR1670363 (2000k:46079)
[9] Gai Hao and Chi-Keung Ng, Crossed products of C ∗-correspondences by amenable
group actions, J. Math. Anal. Appl. 345 (2008), no. 2, 702 -- 707. MR2429169
(2009j:46160)
[10] Richard H. Herman and Adrian Ocneanu, Stability for integer actions on UHF
C ∗-algebras, J. Funct. Anal. 59 (1984), no. 1, 132 -- 144. MR763780 (86m:46057)
[11] Ilan Hirshberg and Wilhelm Winter, Rokhlin actions and self-absorbing C ∗-algebras,
Pacific J. Math. 233 (2007), no. 1, 125 -- 143. MR2366371
[12] Ilan Hirshberg, Wilhelm Winter, and Joachim Zacharias, Rokhlin dimension and
C ∗-dynamics, Comm. Math. Phys. 335 (2015), no. 2, 637 -- 670. MR3316642
[13] Masaki Izumi, Finite group actions on C ∗-algebras with the Rohlin property. I, Duke
Math. J. 122 (2004), no. 2, 233 -- 280. MR2053753 (2005a:46142)
[14] Ja A Jeong, Purely infinite simple C ∗-crossed products, Proc. Amer. Math. Soc.
123 (1995), no. 10, 3075 -- 3078. MR1273502
[15] Ja A. Jeong and Gi Hyun Park, Saturated actions by finite-dimensional Hopf ∗-
algebras on C ∗-algebras, Internat. J. Math. 19 (2008), no. 2, 125 -- 144. MR2384896
[16] G. G. Kasparov, Equivariant KK-theory and the Novikov conjecture, Invent. Math.
91 (1988), no. 1, 147 -- 201. MR918241 (88j:58123)
[17] Takeshi Katsura, A construction of C ∗-algebras from C ∗-correspondences, Advances
in quantum dynamics (South Hadley, MA, 2002), Contemp. Math., vol. 335, Amer.
Math. Soc., Providence, RI, 2003, pp. 173 -- 182. MR2029622 (2005k:46131)
[18] Takeshi Katsura, A class of C ∗-algebras generalizing both graph algebras and homeo-
morphism C ∗-algebras. I. Fundamental results, Trans. Amer. Math. Soc. 356 (2004),
no. 11, 4287 -- 4322 (electronic). MR2067120 (2005b:46119)
[19] Takeshi Katsura, On C ∗-algebras associated with C ∗-correspondences, J. Funct.
Anal. 217 (2004), no. 2, 366 -- 401. MR2102572 (2005e:46099)
[20] Akitaka Kishimoto, Simple crossed products of C ∗-algebras by locally compact
abelian groups, Yokohama Math. J. 28 (1980), no. 1-2, 69 -- 85. MR623751
[21] Akitaka Kishimoto, Outer automorphisms and reduced crossed products of simple
C ∗-algebras, Comm. Math. Phys. 81 (1981), no. 3, 429 -- 435. MR634163 (83c:46061)
23
[22] Alex Kumjian and David Pask, C ∗-algebras of directed graphs and group ac-
tions, Ergodic Theory Dynam. Systems 19 (1999), no. 6, 1503 -- 1519. MR1738948
(2000m:46125)
[23] Masaharu Kusuda, Duality for crossed products of Hilbert C ∗-modules, J. Operator
Theory 60 (2008), no. 1, 85 -- 112. MR2415558 (2009i:46111)
[24] Norio Nawata, Finite group actions on certain stably projectionless C∗-algebras
with the Rohlin property, Trans. Amer. Math. Soc. 368 (2016), no. 1, 471 -- 493.
MR3413870
[25] Hiroyuki Osaka and N. Christopher Phillips, Crossed products by finite group actions
with the Rokhlin property, Math. Z. 270 (2012), no. 1-2, 19 -- 42. MR2875821
[26] Hiroyuki Osaka and Tamotsu Teruya, Nuclear dimension for an inclusion of unital
C ∗-algebras, arXiv:1111.1808 (2011).
[27] William L. Paschke, Inner product modules over B∗-algebras, Trans. Amer. Math.
Soc. 182 (1973), 443 -- 468. MR0355613 (50 #8087)
[28] Cornel Pasnicu and N. Christopher Phillips, Permanence properties for crossed
products and fixed point algebras of finite groups, Trans. Amer. Math. Soc. 366
(2014), no. 9, 4625 -- 4648. MR3217695
[29] Michael V. Pimsner, A class of C ∗-algebras generalizing both Cuntz-Krieger algebras
and crossed products by Z, Free probability theory (Waterloo, ON, 1995), Fields
Inst. Commun., vol. 12, Amer. Math. Soc., Providence, RI, 1997, pp. 189 -- 212.
MR1426840 (97k:46069)
[30] Iain Raeburn and Dana P. Williams, Morita equivalence and continuous-trace C ∗-
algebras, Mathematical Surveys and Monographs, vol. 60, American Mathematical
Society, Providence, RI, 1998. MR1634408 (2000c:46108)
[31] Marc A. Rieffel, Induced representations of C ∗-algebras, Advances in Math. 13
(1974), 176 -- 257. MR0353003 (50 #5489)
[32] M. Rørdam, Classification of nuclear, simple C ∗-algebras, Classification of nuclear
C ∗-algebras. Entropy in operator algebras, Encyclopaedia Math. Sci., vol. 126,
Springer, Berlin, 2002, pp. 1 -- 145. MR1878882 (2003i:46060)
[33] Mikael Rørdam, The stable and the real rank of Z-absorbing C ∗-algebras, Internat.
J. Math. 15 (2004), no. 10, 1065 -- 1084. MR2106263
[34] Luis Santiago, Crossed products by actions of finite groups with the Rokhlin property,
Internat. J. Math. 26 (2015), no. 7, 1550042, 31. MR3357031
[35] Andrew S. Toms, Stuart White, and Wilhelm Winter, Z-stability and finite-
dimensional tracial boundaries, Int. Math. Res. Not. IMRN (2015), no. 10, 2702 --
2727. MR3352253
24
[36] Wilhelm Winter, Nuclear dimension and Z-stability of pure C∗-algebras, Invent.
Math. 187 (2012), no. 2, 259 -- 342. MR2885621
[37] Wilhelm Winter and Joachim Zacharias, The nuclear dimension of C ∗-algebras,
Adv. Math. 224 (2010), no. 2, 461 -- 498. MR2609012 (2011e:46095)
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai-
400076, India
Email address: [email protected]
Department of Mathematical Sciences, Ritsumeikan University, Kusatsu, Shiga, 525-
8577 Japan
Email address: [email protected]
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai-
400076, India
Email address: [email protected]
25
|
1805.08814 | 2 | 1805 | 2019-03-16T01:13:30 | An Elementary Approach to Free Entropy Theory for Convex Potentials | [
"math.OA"
] | We present an alternative approach to the theory of free Gibbs states with convex potentials. Instead of solving SDE's, we combine PDE techniques with a notion of asymptotic approximability by trace polynomials for a sequence of functions on $M_N(\mathbb{C})_{sa}^m$ to prove the following. Suppose $\mu_N$ is a probability measure on on $M_N(\mathbb{C})_{sa}^m$ given by uniformly convex and semi-concave potentials $V_N$, and suppose that the sequence $DV_N$ is asymptotically approximable by trace polynomials. Then the moments of $\mu_N$ converge to a non-commutative law $\lambda$. Moreover, the free entropies $\chi(\lambda)$, $\underline{\chi}(\lambda)$, and $\chi^*(\lambda)$ agree and equal the limit of the normalized classical entropies of $\mu_N$. | math.OA | math | AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR
CONVEX POTENTIALS
DAVID JEKEL
Abstract. We present an alternative approach to the theory of free Gibbs states with
convex potentials. Instead of solving SDE's, we combine PDE techniques with a notion of
asymptotic approximability by trace polynomials for a sequence of functions on MN (C)m
sa
to prove the following. Suppose µN is a probability measure on on MN (C)m
sa given by
uniformly convex and semi-concave potentials VN , and suppose that the sequence DVN is
asymptotically approximable by trace polynomials. Then the moments of µN converge to a
non-commutative law λ. Moreover, the free entropies χ(λ), χ(λ), and χ∗(λ) agree and equal
the limit of the normalized classical entropies of µN .
9
1
0
2
r
a
M
6
1
]
.
A
O
h
t
a
m
[
2
v
4
1
8
8
0
.
5
0
8
1
:
v
i
X
r
a
1. Introduction
1.1. Motivation and Main Ideas. Since Voiculescu introduced free entropy of a non-commutative
law in [38, 39, 40], a number of open problems have prevented a satisfying unification of the
theory (as explained in [41]). The free entropy χ was defined by taking the lim sup as N → ∞
of the normalized log volume of the space of microstates, where the microstates are certain
tuples of N × N self-adjoint matrices having approximately the correct distribution. It is un-
clear whether using the lim inf instead of lim sup would yield the same quantity. Voiculescu
also defined a non-microstates free entropy χ∗ by integrating the free Fisher information of
X + t1/2S where S is a free semicirulcar family free from X, and conjectured that χ = χ∗.
Biane, Capitaine, and Guionnet [6] showed that χ ≤ χ∗ as a consequence of their large
deviation principle for the GUE (see also [17]). The proof relied on stochastic differential equa-
tions relative to Hermitian Brownian motion and analyzed exponential functionals of Brownian
motion. Recent work of Dabrowski [13] combined these ideas with stochastic control theory
and ultraproduct analysis in order to show that χ = χ∗ for free Gibbs states defined by a
convex and sufficiently regular potential. This resolves this part of the unification problem for
a significant class of non-commutative laws.
This paper will prove a similar result to Dabrowski's using deterministic rather than stochas-
tic methods. We want to argue as directly as possible that the classical entropy and Fisher's
information of a sequence of random matrix models converge to their free counterparts. Let
us motivate and sketch the main ideas, beginning with the heuristics behind Voiculescu's non-
microstates entropy χ∗.
Consider a non-commutative law λ of an m-tuple and suppose λ is the limit of a sequence
of random N × N matrix distributions µN given by convex, semi-concave potentials VN :
MN (C)m
sa → R. Let σt,N be the distribution of m independent GUE matrices which each have
normalized variance t, and let σt be the non-commutative law of m free semicircular variables
which each have variance t. Let VN,t be the potential corresponding to the convolution µN∗σt,N .
1991 Mathematics Subject Classification. Primary: 46L53, Secondary: 35K10, 37A35, 46L52, 46L54, 60B20.
Key words and phrases. free entropy, free Fisher information, free Gibbs state, trace polynomials, invariant
ensembles.
1
2
DAVID JEKEL
The classical Fisher information I satisfies
d
dt
1
N 2 h(µN ∗ σt,N ) =
and from this we deduce that
1
N 2 h(µN ) +
m
2
log N =
1
N 3I(µN ∗ σt,N ) =Z kDVN,t(x)k2
2Z (cid:18) m
N 3I(µN ∗ σt,N )(cid:19) dt +
1 + t −
1
1
2 d(µN ∗ σt,N )(x),
m
2
log 2πe.
As N → ∞, we expect the left hand side to converge to the microstates free entropy χ(λ)
because the distribution µN should be concentrated on the microstate spaces of the law λ. On
the other hand, we expect the right hand side to converge to the Voiculescu's non-microstates
free entropy χ∗(λ) defined by
χ∗(λ) =
1
2Z (cid:18) m
1 + t − Φ∗(λ ⊞ σt)(cid:19) dt +
m
2
log 2πe,
where Φ∗ is the free Fisher information and λ ⊞ σt is the free convolution [40].
Under suitable assumptions on VN , the microstates free entropy χ(λ) is the lim sup of normal-
ized classical entropies of µN . On the right hand side, we want to show that N −3I(µN ∗σt,N ) →
Φ∗(λ ⊞ σt) for all t ≥ 0. Since the Fisher information is the L2(µN ) norm squared of the score
function or (classical) conjugate variable DVN,t(x), we want to prove that the classical conju-
gate variables DVN,t(x) behave asymptotically like the free conjugate variables for λ ⊞ σt for
all t.
This would not be surprising because classical objects associated to invariant random matrix
ensembles often behave asymptotically like their free counterparts. For instance, Biane showed
that the entrywise Segal-Bargman transform of non-commutative functions evaluated on N ×N
matrices can be approximated by the free Segal-Bargman transform computed through analytic
functional calculus [5]. Similarly, Guionnet and Shlyakhtenko showed that classical monotone
transport maps for certain random matrix models approximate the free monotone transport
[22, Theorem 4.7]. Moreover, Dabrowski's approach to prove χ = χ∗ involved constructing
solutions to free SDE as ultraproducts of the solutions to classical SDE [13].
In section 3.4, we make precise the idea that a sequence of functions on MN (C)m
sa has a
"well-defined, non-commutative asymptotic behavior" by defining asymptotic approximability
by trace polynomials (Definition 3.24). We assume that DVN at time zero has the approximation
property and must show that the same is true for DVN,t for all t.
First, we show that this property is preserved under several operations on sequences, includ-
ing composition and convolution with the Gaussian law σN,t (see §3.4). Then in §6 we analyze
the PDE that describes the evolution of VN,t. We show that for all t the solution can be ap-
proximated in a dimension-independent way by applying a sequence of simpler operations, each
of which preserves asymptotic approximability by trace polynomials. We conclude that if the
initial data DVN is asymptotically approximable by trace polynomials, then so is DVN,t, and
hence we obtain convergence of the classical Fisher information to the free Fisher information.
This proves the equality χ(λ) = χ∗(λ) whenever a sequence of log-concave random matrix
models µN converges to λ in an appropriate sense (Theorem 7.1). Another result (Theorem
4.1), proved by similar techniques, establishes sufficient conditions for a sequence of log-concave
random matrix models µN to converge in moments to a non-commutative law λ, so that The-
orem 7.1 can be applied. As a consequence, we show that χ = χ∗ for a class of free Gibbs
states.
1.2. Main Results. To fix notation, let MN (C)m
self-adjoint N×N matrices and let kxk2 = (Pj τN (x2
sa be space of m-tuples x = (x1, . . . , xm) of
j ))1/2, where τN = (1/N ) Tr. We denote by
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
3
kxk∞ the maximum of the operator norms kxjk. Recall that a trace polynomial f (x1, . . . , xm)
is a linear combination of terms of the form
p(x)
nYj=1
τ (pj (x)),
where p and pj are non-commutative polynomials in x1, . . . , xm.
Consider a sequence of potentials VN : MN (C)m
2 is convex
2 is concave for some 0 < c < C. Define the associated probability
and VN (x) − (C/2)kxk2
measure µN by
sa → R such that VN (x)− (c/2)kxk2
ZN =ZMN (C)m
e−N 2VN (x) dx.
sa
dµN (x) =
1
ZN
e−N 2VN (x) dx,
Assume that the sequence of normalized gradients DVN (x) = N∇VN (x) is asymptoticallly
approximable by trace polynomials in the sense that for every ǫ > 0 and R > 0, there exists a
trace polynomial f (x) such that
lim sup
N→∞
sup
kxk∞≤RkDVN (x) − f (x)k2 ≤ ǫ,
R (x − τN (x)) dµN (x) is bounded in operator norm as N → ∞ (it will be zero if µN is unitarily
where kxk∞ denotes the maximum of the operator norms of the xj 's. Also, assume that
invariant or has expectation zero). In this case, we have the following. (To clarify the larger
picture, we include statements of the concentration estimates (1) and (3), although these are
standard and not proved in this paper.)
(1) There exists a constant R0 such that µN (kxk∞ ≥ R0 + δ) ≤ me−cN δ2/2 for δ > 0.
(2) There exists a non-commutative law λ such that
N→∞Z τN (p(x)) dµN (x) = λ(p)
lim
for every non-commutative polynomial p.
(3) The measures µN exhibit exponential concentration around λ, in the sense that
lim
N→∞
1
N 2 log µN (kxk∞ ≤ R,τN (p(x)) − λ(p) ≥ δ) < 0
for every R > 0 and every non-commutative polynomial p.
(4) The law λ has finite free entropy and we have
χ(λ) = χ(λ) = χ∗(λ) = lim
N→∞
1
N 2(cid:16)h(µN ) +
m
2
log N(cid:17) ,
where χ and χ are respectively the lim sup and lim inf versions of microstates free
entropy, χ∗ is the non-microstates free entropy, and h is the classical entropy.
(5) The same holds for µN ∗ σt,N and λ ⊞ σt, where σt,N is the law of m independent GUE
matrices with variance t and σt is the law of m free semicircular variables with variance
t.
(6) The law λ has finite free Fisher information. If I is the classical Fisher information
and Φ∗ is the free Fisher information, then
lim
N→∞
1
N 3I(µN ∗ σt,N ) = Φ∗(λ ⊞ σt).
(7) The functions t 7→ 1
N 3I(µN ∗ σt,N ) and t 7→ Φ∗(λ ⊞ σt) are decreasing and Lipschitz in
t with the absolute value of the derivative bounded by C2m(1 + Ct)−2.
4
DAVID JEKEL
Here claims (1) through (3) come from Theorem 4.1, which is similar to the earlier results
[21, Theorem 4.4], [15, Proposition 50 and Theorem 51], [13, Theorem 4.4] plus standard results
on concentration of measure. Claims (4) through (7) come from Theorem 7.1, which is similar
to [13, Theorem A].
In particular, we recover Dabrowski's result [13, Theorem A] that χ(λ) = χ(λ) = χ∗(λ)
when the law λ is a free Gibbs state given by a sufficiently regular convex non-commutative
potential V (X), because taking VN = V will define a sequence of random matrix models µN
which concentrate around the non-commutative law λ.
Unlike Dabrowski, we do not provide an explicit formula for (d/dt)Φ(λ ⊞ σt). However,
we are able to prove that Φ(λ ⊞ σt) is Lipschitz in t rather than merely having a derivative
in L2(dt) (and hence being Holder 1/2 continuous) as shown by Dabrowski. Our results also
allow slightly more flexibility in the choice of random matrix models, so that we do not have to
assume that VN is given by exactly the same formula for every N or that VN is exactly unitarily
invariant.
1.3. Organization of Paper. Section 2 establishes notation and reviews basic facts from
non-commutative probability and random matrix theory.
Section 3 defines the algebra of trace polynomials and describes how they behave under differ-
entiation and convolution with Gaussians. We then introduce the notion that a sequence {φN}
of functions MN (C)m
sa or C is asymptotically approximable by trace polynomi-
als. We show that this approximation property is preserved under several operations including
composition and Gaussian convolution.
sa → MN (C)m
is the semigroup such that ut = T VN
t
t
Section 4 proves Theorem 4.1 concerning the convergence of moments for the measure µN
t u,
t u solves the equation ∂tut = (1/2N )∆ut −
by iterating simpler operations in order to show that if
t uN , and
(claims (1) - (3) of §1.2). We evaluate R u dµN for a Lipschitz function u as limt→∞ T VN
where T VN
(N/2)∇V · ∇ut. We approximate T VN
N∇Vn and uN are asymptotically approximable by trace polynomials, then so is T VN
hence that limN→∞R uN dµN exists.
Section 5 reviews the defintions of free entropy and Fisher's information. We also show that
the microstates free entropies χ(λ) and χ(λ) are the lim sup and lim inf of normalized classical
entropies of µN , provided that µN concentrates around λ and satisfies some mild operator norm
tail bounds, and that {VN} is asymptotically approximable by trace polynomials. Similarly, if
{N∇VN} is asymptotically approximable by trace polynomials, then the normalized classical
Fisher information converges to the free Fisher information.
Section 6 considers the evolution of the potential VN (x, t) corresponding to µN ∗ σt,N , where
σt,N is the law of m independent GUE of variance t. Our goal is to show that if N∇VN (x, 0)
is asymptotically approximable by trace polynomials, then so is N∇VN (x, t) for all t > 0, so
that we can apply our previous result that the classical Fisher information converges to the
free Fisher information. As in §4, we construct the semigroup Rt which solves the PDE as a
limit of iterates of simpler operations which are known to preserve asymptotic approximation
by trace polynomials.
Section 7 concludes the proof of our main theorem on free entropy and Fisher's information
(Theorem 7.1), which establishes claims (4) - (7) of §1.2, assuming a weakened version of the
hypothesis and conclusion of Theorem 4.1.
In section 8, we characterize the limiting laws λ which arise in Theorem 4.1 as the free Gibbs
states for a certain class of potentials. In particular, we apply Theorem 7.1 to show that χ = χ∗
for several types of free Gibbs states considered in previous literature.
1.4. Acknowledgements. I thank Timothy Austin, Guillaume C´ebron, Yoann Dabrowski,
Alice Guionnet, Benjamin Hayes, Dimitri Shlyakhtenko, Terence Tao, Yoshimichi Ueda, and
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
5
Dan Voiculescu for various useful conversations. I especially thank Shlyakhtenko for his men-
torship and ongoing conversations about free entropy, and Dabrowski for detailed discussions
of his own results and other recent literature. I acknowledge the support of the NSF grants
DMS-1344970 and DMS-1500035. Part of this research was conducted at the Institute for Pure
and Applied Mathematics (IPAM) during the long program on Quantitative Linear Algebra in
Spring 2018. IPAM provided hospitality and a stimulating research environment where many
of the conversations mentioned above took place. I also thank the referee for detailed feedback
that improved the clarity and correctness of the paper throughout.
The first subsection 2.1 fixes certain notations which will be used throughout the paper. The
2. Preliminaries
other subsections of §2 discuss background that they reader may refer to as needed.
2.1. Notation for Matrix Algebras. Let MN (C) denote the N × N matrices over C and
let MN (C)sa be the self-adjoint elements. Note that MN (C)m
sa is a real inner product space
j=1 Tr(xj yj) for x = (x1, . . . , xm) and y = (y1, . . . , ym).
Moreover, MN (C) can be canonically identified with the complexification C ⊗R MN (C)sa by
decomposing a matrix into its self-adjoint and anti-self-adjoint parts.
. An explicit choice of
with the inner product hx, yiTr :=Pm
Being a real-inner product space, MN (C)sa is isomorphic to RmN 2
coordinates can be made using the following orthonormal basis for MN (C)sa:
(2.1)
BN = {Ek,k}N
k=1 ∪(cid:26) 1
√2
Ek,ℓ +
1
√2
Eℓ,k(cid:27)k<ℓ ∪(cid:26) i
√2
Ek,ℓ −
i
√2
Eℓ,k(cid:27)k<ℓ
.
This basis has the property that for any x, y, z ∈ MN (C), we have
(2.2)
xbybz = xz Tr(y),
Xb∈BN
which follows from an elementary computation.
hx, yi2 =Pm
denote the normalized trace by τN = 1
We denote the norm corresponding to Tr by · (essentially the Euclidean norm). We
N Tr. We denote the corresponding inner product by
j=1 τN (xj yj) and the norm by k·k2. For x ∈ MN (C), we denote the operator norm
by kxk. Similarly, if x = (x1, . . . , xm) ∈ MN (C)m, we denote kxk∞ = maxjkxjk.
The symbols ∇ and ∆ will respresent the gradient and Laplacian operators with respect
to the coordinates of MN (C)sa in the non-normalized inner product h·,·iTr. The symbols
D and LN will denote the normalized versions N∇ and (1/N )∆ respectively, as well as the
corresponding linear transformations on the algebra of trace polynomials. This normalization
and notation will be explained and justified in §3.2.
2.2. Non-Commutative Probability Spaces and Laws. The following are standard defi-
nitions and facts in non-commutative probability. For further background, see [42, 27].
Definition 2.1. A von Neumann algebra is a unital C-algebra M of bounded operators on a
Hilbert space H which is closed under adjoints and closed in the weak operator topology.
Definition 2.2. A tracial von Neumann algebra or non-commutative probability space is a von
Neumann algebra M together with a bounded linear map τ : M → C which is continuous in
the weak operator topology and satisfies τ (1) = 1, τ (xy) = τ (yx), and τ (x∗x) ≥ 0. The map τ
is called a trace.
Definition 2.3. For m ≥ 1, we denote by NCPm = ChX1, . . . , Xmi the algebra of non-
commutative polynomials in X1, . . . , Xm. A non-commutative law (for an m-tuple) is a map
λ : NCPm → C such that
6
DAVID JEKEL
(1) λ is linear,
(2) λ is unital (that is, λ(1) = 1),
(3) λ is completely positive, that is, for any matrix P with entries in ChX1, . . . , Xmi, the
(4) λ is tracial, that is, λ(p(X)q(X)) = λ(q(X)p(X)).
matrix λ(P ∗P ) is positive semi-definite.
We denote by Σm the space of non-commutative laws equipped with the topology of pointwise
convergence on ChX1, . . . , Xmi, that is, convergence in non-commutative moments.
Definition 2.4. We say that a non-commutative law λ is bounded by R if we have
We denote the space of such laws by Σm,R.
λ(Xi1 , . . . , Xin ) ≤ Rn.
Definition 2.5. Suppose that x1, . . . xm are bounded self-adjoint elements of a tracial von
Neumann algebra (M, τ ). Then the law of x = (x1, . . . , xm) is the map
λx : ChX1, . . . , Xni → C : p(X) 7→ τ (p(x)).
Definition 2.6. Let MN (C) be the algebra of N × N matrices over C. Let τN = 1
N Tr be
the normalized trace. Then (MN (C), τN ) is a tracial von Neumann algebra, and hence for any
tuple of self-adjoint matrices x = (x1, . . . , xm), the law λx is defined by Definition 2.5.
Proposition 2.7. The space Σm,R is compact, separable, and metrizable. Moreover, every
µ ∈ Σm,R can be realized as λx for some tuple x = (x1, . . . , xm) of self-adjoint elements of a
tracial von Neumann algebra (M, τ ) with kxk∞ ≤ R.
2.3. Non-commutative Lα-norms. On several occasions, we will need to use the non-commutative
Lα norms for α ∈ [1, +∞]. (Here we use α rather than p since the letter p will often be used
for a polynomial.) If y is any element of a tracial von Neumann algebra (M, τ ), then we de-
fine y = (y∗y)1/2 defined using continuous functional calculus. For α ∈ [0, +∞), we define
kykα = τ (yα)1/α. We also define kyk∞ to be the operator norm.
Proposition 2.8. If (M, τ ) is a tracial von Neumann algebra and α ∈ [1, +∞], then k·kα
defines a norm. Moreover, we have the non-commutative Holder inequality
whenever
kx1 . . . xnkα ≤ kx1kα1
. . .kxnkαn
1
α1
+ ··· +
1
αn
=
1
α
.
Moreover, we have τ (y) ≤ kyk1.
A standard proof of the Holder inequality uses polar decomposition, complex interpolation,
and the three lines lemma. We will in fact only need this inequality for the trace τN on MN (C).
Modulo renormalization of the trace, the inequality for matrices follows from the treatment of
trace-class operators in [34]; see especially Thm. 1.15 and Thm. 2.8, as well as the references
cited on p. 31. For the setting of von Neumann algebras, a convenient proof can be found in
[12, Thm. 2.4 - 2.6]; for an overview and further history see [29, §2].
Remark 2.9. One can define the non-commutative Lα norm for a tuple (y1, . . . , ym) as
k(y1, . . . , ym)kα =(τ (y1α + ··· + ymα)1/α, α ∈ [1, +∞)
maxjkyjk,
α = +∞.
However, for tuples, we will only need to use the 2 and ∞ norms.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
7
2.4. Free Independence, Semicircular Law, and GUE. We will use the following standard
definitions and facts from free probability. For further background, refer to [36, 37, 42, 27, 1].
Let (M, τ ) be a tracial von Neumann algebra, and let A1, . . . , An be unital ∗-subalgebras
of M. Then we say that A1, . . . , Am are freely independent if given a1, . . . , ak with aj ∈ Aij
and ij 6= ij+1 and τ (aj ) = 0 for each j, we have also τ (a1 . . . ak) = 0.
In particular, if S1, . . . , Sn are subsets of M, then we say that they are freely independent
if the unit ∗-subalgebras they generate are freely independent. Thus, for instance, self-adjoint
elements x1, . . . , xm of M are freely independent if given polynomials f1, . . . , fk and indices
i1, . . . , ik with ij 6= ij+1 such that τ (fj (Xij )) = 0, we have also τ (f1(Xi1 ) . . . fk(Xik )) = 0.
The free convolution of two non-commutative laws µ and ν (of self-adjoint m-tuples) is
defined as the non-commutative law of (x1 + y1, . . . , xm + ym), given that {x1, . . . , xm} and
{y1, . . . , ym} are freely independent and the non-commutative law of (x1, . . . , xm) is µ, and the
non-commutative law of (y1, . . . , ym) is ν. Then ⊞ is well-defined, independent of the particular
choice of operators that realize the laws µ and ν. Moreover, ⊞ commutative and associative.
If X1, . . . , Xm are freely independent, then their joint law is determined by the individual
laws of the Xj's, each of which is represented by a compactly supported probability measure
on R. The semicircle law (of mean zero and variance 1) is the probability measure given by
density (1/2π)√4 − x21[−2,2](x) dx. We denote by σt the non-commutative law of m freely
independent semicircular random variables which each have mean zero and variance t (that is,
σt(Xj) = 0 and σt(X 2
j ) = t).
These free semicircular families play the role of multivariable Gaussians in free probability.
Moreover, they form a semigroup under free convolution, that is, σs ⊞ σt = σs+t.
We denote by σt,N the probability distribution on MN (C)m
sa for m independent GUE matrices
of normalized variance t, that is,
dσt,N (x) =
1
ZN,t
exp−N
mXj=1
Tr(x2
j )/2t dx,
where ZN,t is chosen so that σt,N is a probability measure. It is well known that the indepen-
dent GUE matrices behave in the limit like freely independent semicircular random variables,
although we shall directly prove the specific properties we use in §3.
2.5. Concentration and Operator Norm Tail Bounds. The following is a standard con-
centration estimate for uniformly log-concave random matrix models. The best known proof
goes through the log-Sobolev inequality and Herbst's argument (see [1, §4.4.2]), although it can
also be proved directly using the heat semigroup directly as in [26]. We state the theorem here
with free probabilistic normalizations.
Theorem 2.10. Suppose that V : MN (C)m
convex. Define
sa → R is a potential such that V (x) − (c/2)kxk2
2 is
dµ(x) =
Suppose that f : MN (C)m
Z =Z exp(−N 2V (x)) dx.
1
Z
sa → R is K-Lipschitz with respect to k·k2. Then
exp(−N 2V (x)) dx,
µ(x : f (x) −R f dµ ≥ δ) ≤ ecN 2δ2/2K 2
µ(x : f (x) −R f dµ ≥ δ) ≤ 2e−cN 2δ2/2K 2
,
.
and since the same estimate can be applied to −f , we have also
In particular, this concentration estimate applies to the GUE law σt,N with c = 1/t.
In
addition to the concentration estimate, we will also use the fact that such uniformly convex
8
DAVID JEKEL
random matrix models have subgaussian moments and therefore have good tail bounds on the
probability of large operator norm. The following theorem is a special case of [23, Theorem
1.1] and the application to random matrix models is taken from the proof of [20, Theorem 3.4].
Theorem 2.11. Let V and µ be as in Theorem 2.10, and suppose that f : MN (C)m
sa → R is
convex. Let a =R x dµ(x). Then
In particular, if kxkα denotes the Lα norm from §2.3, then for every α ∈ [1, +∞] and β ∈
[1, +∞), we have
Z f (x − a) dµ(x) ≤Z f (y) dσc−1,N (y).
Z kxj − ajkβ
α dµ(x) ≤Z kyjkβ
α dσc−1,N (y).
Proof. The convexity assumption on V means that µ has a log-concave density with respect to
the Gaussian measure σc−1,N (y). Therefore, the first claim follows from Harg´e's theorem [23,
Theorem 1.1]. The second claim follows because norms on vector spaces are convex functions,
and the function t 7→ tβ is convex for β ≥ 1.
(cid:3)
Corollary 2.12. Let VN : MN (C)sa → R be a function such that VN (x) − (c/2)kxk2
and let µN be the corresponding measure. Let aN,j =R xj dµN (x). Then
2 is convex
and
lim sup
N→∞ Z kxj − aN,jk dµN (x) ≤ 2c−1/2,
µN (x : kxjk ≥R kyjk dµN (yj) + δ) ≤ e−cδ2N/2.
Proof. In light of Theorem 2.11, for the first claim of the Corollary, it suffices to check the
special case σc−1,N . This special case is a standard result in random matrix theory; see for
instance the proof of [1, Theorem 2.1.22]. The second claim follows from Theorem 2.10 after
we observe that the function on MN (C)N
sa given by x 7→ kxjk∞ is N 1/2-Lipschitz with respect
to k·k2.
2.6. Semi-convex and Semi-concave functions. We recall the following terminology and
facts about semi-convex and semi-concave functions. These facts are typically applied to func-
tions from Rn → R, but of course they hold equally well if Rn is replaced by a finite-dimensional
real inner product space. In particular, we focus on the case of MN (C)m
sa.
sa → R is semi-convex if there exists some c ∈ R such that u(x) −
2 is convex. If this holds for some c > 0, then u is said to be uniformly convex. Similarly,
sa → R is said to be semi-concave if there exists C ∈ R such that u(x) − (C/2)kxk2
(c/2)kxk2
u : MN (C)m
is concave, and it is uniformly concave if this holds for some C < 0.
A function u : MN (C)m
(cid:3)
2
Fix m and N . For c ≤ C be real numbers. Then we define
sa → R : u(x)−(c/2)kxk2
Em,N (c, C) = {u : MN (C)m
2 is concave}.
We will often suppress m and N in the notation and simply write E(c, C). Throughout the
paper, we rely on the following basic properties of functions in E(c, C).
Proposition 2.13.
2 is convex and u(x)−(C/2)kxk2
(1) The space E(c, C) is closed under translation, averaging with respect to probability mea-
sures, and pointwise limits.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
9
(2) A function u is in E(c, C) if and only if for every point x0 ∈ MN (C)m
sa, there exists
some p ∈ MN (C)m
u(x0) + hp, x − x0i2 +
sa such that
c
2kx − x0k2
2 ≤ u(x) ≤ u(x0) + hp, x − x0i2 +
(3) In particular, if u ∈ E(c, C), then u is differentiable everywhere.
(4) If u ∈ E(c, C), then the gradient Du is max(c,C)-Lipschitz with respect to k·k2.
(5) If u ∈ E(c, C), then
C
2 kx − x0k2
2.
ckx − yk2
2 ≤ hDu(x) − Du(y), x − yi2 ≤ Ckx − yk2
2.
(6) If u ∈ E(c, C) for some c > 0, then u is bounded below and achieves a global minimum
at its unique critical point.
Sketch of proof. (1) One follows from elementary computation and the fact that the same holds
for the class of convex functions.
(2), (3) The convex functions u(x) − (c/2)kxk2
2 − u(x) must have supporting
hyperplanes at x0. This yields one vector p which satisfies the left inequality of (2) and another
vector p′ satisfying the right inequality. Then one checks that p must equal p′ and this implies
that u is differentiable at x0.
2 and (C/2)kxk2
(4), (5) One can check these inequalities for smooth functions in Ec,C directly using Taylor
expansions and calculus. Consider a general u ∈ Ec,C. Let un = u ∗ ρn, where ρn be a smooth
probability density supported in the ball of radius 1/2 around 0. Then un is smooth and
un → u locally uniformly. Also, un ∈ Ec,C by (1), hence Dun is max(c,C)-Lipschitz. By
the Arzela-Ascoli theorem, after passing to a subsequence, we may assume that Dun converges
locally uniformly to some F . It follows from local uniform convergence of Dun that F = Du.
Moreover, since (4) and (5) hold for Dun, they also hold for Du.
(cid:3)
3. Trace Polynomials
In this secion, we consider the algebra of trace polynomials in non-commutative variables X1,
. . . , Xm, first defined in [31], [32]. As in [30], [9], [16], we describe how trace polynomials behave
under differentiation (§3.2) and convolution with Gaussian (§3.3). Finally, in §3.4, we define
the property of asymptotic approximability by trace polynomials for a sequence of functions
on MN (C)m
sa, which is one of the key technical tools in our proof.
3.1. Definition.
Definition 3.1. We define the algebra of scalar-valued trace polynomials, or TrP0
Let V be the vector space NCPm / Span(pq − qp : p, q ∈ NCPm). We define the vector space
(3.1)
TrP0
m, as follows.
m =
∞Mn=0
V ⊙n,
where ⊙ is the symmetric tensor power over C. Then TrP0
m forms a commutative algebra
with the tensor operator ⊙ as the multiplication. We denote the element p1 ⊙ ··· ⊙ pn by
τ (p1) . . . τ (pn) where τ is a formal symbol.
To state the definition more suggestively, an element of TrP0
m is a linear combination of
terms of the form τ (p1(X)) . . . τ (pn(X)) where p1, . . . , pn are non-commutative polynomials
in X1, . . . , Xm and τ is a formal symbol thought of as the trace. By forming a quotient vector
space, we identify τ (pq) with τ (qp). The trace polynomials form a commutative ∗-algebra TrP0
over C where the ∗-operation is
(3.2)
(τ (p1(X)) . . . τ (pn(X)))∗ = τ (p1(X)∗) . . . τ (pn(X)∗)
m
10
DAVID JEKEL
and the multiplication operation is the one suggested by the notation.
We define TrPk
m to be the vector space
TrPk
m := TrP0
m ⊗ChX1, . . . , Xmi⊗k.
m forms a ∗-algebra because it is the tensor products of two ∗-algebras.
m operator-valued trace polynomials. We use the term trace poly-
m. Note
We call the elements of TrP1
nomials more generally to describe elements of TrPk
that TrP1
Definition 3.2. Suppose that M is a von Neumann algebra with trace σ. Given f ∈ TrP1
and a self-adjoint tuple x = (x1, . . . , xm) of elements of M, we define f (x) to the element of M
given by formally evaluating f with the formal symbol X repalce by x and the formal symbol
τ replaced with σ. For instance if f (X) = p0(X) ⊗ τ (p1(X)) . . . τ (pn(X)) in TrP1
m or tuples of elements from TrPk
m, then
m
f (x) = p0(x)σ(p1(x)) . . . σ(pn(x)).
In particular, we define f (x) when x is an m-tuple of self-adjoint N × N matrices by setting
τ = τN .
Definition 3.3. If f ∈ TrP0
m and λ is a non-commutative law, we define the evaluation λ(f )
to be the number obtained by replacing the symbol τ with λ everywhere in f . For example, if
f (X1, X2, X3) = τ (X1)τ (X2X3) + τ (X 2
2 ), then we define
λ(f ) = λ(f (X1, . . . , Xm)) = λ(X1)λ(X2X3) + λ(X 2
2 ).
Definition 3.4. We define the degree for elements of NCPm and TrPk
m as follows. If p ∈ NCPm
is a monomial p(X1, . . . , Xm) = Xi1 . . . Xid , then we define deg′(p) = d. If p1, . . . , pℓ and q1,
. . . , qk are non-commutative monomials, then consider the element τ (p1) . . . τ (pℓ)q1⊗···⊗ qk ∈
TrPk
m, and define
deg′(τ (p1) . . . τ (pℓ)q1 ⊗ ··· ⊗ qk) = deg′(p1) . . . deg′(pℓ) deg′(q1) . . . deg′(qk).
m, we define the degree deg(f ), as the infimum of max(deg′(f1), . . . , deg′(fℓ)),
For general f ∈ TrPk
where f = f1 + ··· + fℓ and each fj is a product of non-commutative monomials and traces
of non-commutative monomials as above. Similarly, for general f ∈ NCPm, we define deg(f )
as the infimum of max(deg′(f1), . . . , deg′(fℓ)) where f = f1 + ··· + fℓ and each fj is a non-
commutative monomial.
Remark 3.5. One can check that if f is a product of monomials as above, then deg(f ) = deg′(f ).
Moreover, the degree makes TrP0
m into graded algebras. Finally, we observe that
f ∈ TrP0
sa defined by x 7→ f (x) is a polynomial in the
entries of x1,. . . , xm, and the degree of x 7→ f (x) with respect to the entries is bounded above
by the degree of f in TrP0
m. None of these facts will be used in what follows, so we
omit the proofs.
m, then the function on MN (C)m
m and TrP1
m or TrP1
m or TrP1
We also observe that there is a composition operation (TrP1
m)m
defined just as one would expect from manipulations in MN (C). If f, g ∈ (TrP1
m)m, we define
f (g(x)) by substituting gj(x) as the j-th argument of f . Then we multiply elements out by
treating the terms of the form τ (p) like scalars. For instance, if f (Y1, Y2) = (τ (Y1Y2)Y2, Y1 +
τ (Y 2
1 )Y2) and g(X1, X2) = (τ (X1)X2 + X1, X1), then f ◦ g(X1, X2) = (Z1, Z2), where
m)m → (TrP1
m)m × (TrP1
Z1 = (τ [τ (X1)X2 + X1]X1)X1 = τ (X1)τ (X2X1)X1 + τ (X 2
1 )X1
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
11
and
Z2 = τ (X1)X2 + X1 + X1τ [(τ (X1)X2 + X1)2]
= τ (X1)X2 + X1 + τ [τ (X1)2X 2
= τ (X1)X2 + X1 + [τ (X1)2τ (X 2
= τ (X1)X2 + X1 + τ (X1)2τ (X 2
2 + τ (X1)X2X1 + τ (X1)X1X2 + X 2
2 ) + τ (X1)τ (X2X1) + τ (X1)τ (X1X2) + τ (X 2
2 )X1 + 2τ (X1)τ (X2X1)X1 + τ (X 2
1 )X1.
1 ]X1
1 )]X1
One can check that composition on (TrP1
adjoint elements of (TrP1
f ◦ g ∈ (TrP1
composition of the corresponding functions for f and g.
m)m, then they define functions MN (C)m
m)m defined abstractly will product a function MN (C)m
m)m is well-defined. Moreover, if f and g are self-
sa, and the element
sa which is the
sa → MN (C)m
sa → MN (C)m
3.2. Differentiation of Trace Polynomials. In this section, we give explicit formulas for
the gradient and Laplacian of trace polynomials and in particular show that these operations
have a well-defined limit as N → ∞ (see [30], [9], [16, §3]). We first recall the free difference
quotients of Voiculescu [40].
Definition 3.6. We define the free difference quotient (or simply non-commutative derivative)
Dj : NCPm → NCPm ⊗ NCPm by
Xi1 . . . Xik−1 ⊗ Xik+1 . . . Xin .
We also define Dj : NCP⊗n
Dj[Xi1 . . . Xin ] = Xk:ik=j
m → NCP⊗n+1
nXk=1
by
m
Dj[p1 ⊗ ··· ⊗ pn] =
p1 ⊗ ··· ⊗ pk−1 ⊗ Djpk ⊗ pk+1 ⊗ ··· ⊗ pn.
j is a well-defined map NCP⊗n
Then of course Dk
Remark 3.7. We caution the reader that the normalization for Dn
[40] by a factor of n!.
Definition 3.8. We define the cyclic derivative D◦
by
m → NCP⊗n+k
m
.
j [Xi1 . . . Xin ] = Xk:ik=j
D◦
Xik+1 . . . Xin Xi1 . . . Xik−1 .
j f here differs from that of
j : NCPm → NCPm as the linear map given
Definition 3.9. Given an algebra A (e.g. NCPm), we define the hash operation as the bilinear
map (A ⊗ A) × A given by (a1 ⊗ a2)#b = a1ba2.
Example 3.10. Let X = (X1, X2, X3) and define f (X) = X1X2X 2
1 X3X2. Then
D1f (X) = 1 ⊗ X2X 2
1f (X) = X2X 2
D◦
D1f (X)#Y = Y X2X 2
1 X3X2 + X1X2 ⊗ X1X3X2 + X1X2X1 ⊗ X3X2
1 X3X2 + X1X3X2X1X2 + X3X2X1X2X1,
1 X3X2 + X1X2Y X1X3X2 + X1X2X1Y X3X2.
To compute D2
1f (X) = D1[D1f (X)], we would add together the three terms
D1[1 ⊗ X2X 2
1 X3X2] = 1 ⊗ X2 ⊗ X1X3X2 + 1 ⊗ X2X1 × X3X2
D1[X1X2 ⊗ X1X3X2] = 1 ⊗ X2 ⊗ X1X3X2 + X1X2 ⊗ 1 ⊗ X3X2
D1[X1X2X1 ⊗ X3X2] = 1 ⊗ X2X1 ⊗ X3X2 + X1X2 ⊗ 1 ⊗ X3X2.
12
DAVID JEKEL
Now we can define derivatives for scalar-valued and non-commutative trace polynomials that
sa. We begin
correspond with differentiation with respect to the standard coordinates on MN (C)m
with the gradient.
To fix notation, recall that in §2.1 we gave a canonical orthonormal basis for MN (C)sa
with respect to the inner product hx, yi = Tr(x∗y). Using these coordinates, we may iden-
tify MN (C)sa with RN 2
. Similarly, we identify the
complexification C ⊗ MN (C)m
sa → C and
x = (x1, . . . , xm) ∈ MN (C)m
sa, we denote by ∇f (x) ∈ MN (C)m the gradient computed in these
coordinates; similarly, we denote by ∇jf (x) ∈ MN (C) the gradient with respect to xj computed
in these coordinates.
sa with MN (C)m and with CmN 2
and hence identify MN (C)m
. For f : MN (C)m
sa with RmN 2
Definition 3.11. Define the jth gradient operator TrP0
Dj" nYk=1
τ (pk)# =
m by
m → TrP1
j pkYℓ6=k
D◦
τ (pℓ).
nXk=1
(3.3)
Note that Dj is defined so as to obey the Leibniz rule (that is, it is a derivation).
Lemma 3.12. If f ∈ TrP0
(3.4)
m is viewed as a function MN (C)m
[Djf ](x).
sa → MN (C)m = C, then we have
∇j[f (x)] =
1
N
Similarly, for F : MN (C)m
sa → MN (C)m, let JjF denote the Jacobian linear transformation
(a.k.a. Fr´echet derivative) with respect to xj . Then for a non-commutative polynomial p, we
have
(3.5)
[Jjp(x)](y) = [Djp](x)#y,
and hence by the product rule for p ∈ NCPm and f ∈ TrP0
(3.6)
Proof. By standard computations, for a non-commutative polynomial p and y ∈ MN (C)sa, we
have
[Jj(pf )(x)](y) = ([Dj p](x)#y)f (x) + p(x)τN ([Djf ](x)y).
m, we have
[Jjp(x)](y) = [Djp](x)#y
j p](x).
∇j[τN (p)](x) =
[D◦
1
N
The claims (3.4) and (3.6) now follow from the product rule.
(cid:3)
Next, we can define the algebraic Laplacian operators on TrP0
m and TrP1
For f : MN (C)m
to computing the Laplacian on scalar-valued or vector-valued functions on MN (C)m
the coordinates given in §2.1.
matrix xj . Note that ∆f =Pm
sa → C, let ∆jf be the Laplacian with respect to the coordinates of the j-th
sa → MN (C) is an operator-
valued function, we define ∆jf and ∆f by applying ∆j and ∆ entrywise (as is standard notation
for the Laplacian of a vector-valued function).
j=1 ∆jf . Similarly, if f : MN (C)m
NCP⊗3
Motivated by (2.2) and the computation in Lemma 3.18 below, we define the map η :
m → TrP1
m by
m, which correspond
sa, still using
η(p1 ⊗ p2 ⊗ p3) = p1p3τ (p2).
Definition 3.13. We define Lj and LN,j : TrP0
such that
m → TrP0
m to be the unique linear operators
(3.7)
Lj[τ (p)] = LN,j[τ (p)] = τ ◦ η[D2
j p] for p ∈ NCPm .
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
13
and such that the following product rule is satisfied:
Lj[f · g] = Lj[f ] · g + f · Lj[g]
(3.8)
(3.9)
LN,j[f · g] = LN,j[f ] · g + f · LN,j[g] +
2
N 2 τ (Dj f · Djg).
Then we define L =Pm
j=1 Lj and LN =Pm
j=1 LN,j.
Remark 3.14. To show the existence of operators LN,j and Lj satisfying (3.7) and the product
rule, one can define LN,j more explicitly as the linear operator TrP0
m given by
LN,j[τ (p1) . . . τ (pn)] =
nXk=1
τ ◦ η[D2pk] ·Yi6=k
τ (pi) +
1
N 2
nXk=1Xℓ6=k
m → TrP0
τ (D◦
j pk · D◦
j pℓ) Yi6=k,ℓ
τ (pi),
and check that this operator satisfies the product rule. Moreover, the uniqueness of the operator
LN,j satisfying (3.7) and the product rule follows from the fact that TrP0
m is spanned by
products of terms of the form τ (p) for p ∈ NCPm. The argument for the existence and
uniqueness of Lj is the same.
Example 3.15. Let X = (X1, X2). Consider f (X) = τ (f1(X))τ (f2(X)) where f1(X) =
X1X2X1X3 and f2(X) = X 2
2 X1. Then
D1[τ (f1)] = D◦
D1[τ (f2)] = D◦
1f1 = X2X1X3 + X3X1X2
2f2 = X 2
2 ,
and
L1[τ (f1)] = LN,1[τ (f1)] = τ ◦ η[D2
L1[τ (f2)] = LN,1[τ (f2)] = 0.
1f1] = τ [η[1 ⊗ X2 ⊗ X3]] = τ [1 · X3] · τ [X2]
Therefore, we have
L1[f ] = L1[τ (f1)]τ (f2) + τ (f1)L1[τ (f2)] = τ (X3)τ (X2)τ (X 2
1 f1D◦
LN,1[f ] = LN,1[τ (f1)]τ (f2) + τ (f1)LN,1[τ (f2)] +
2
N 2 τ [D◦
1f2]
2 X1) + 0
= τ (X3)τ (X2)τ (X 2
2 X1) +
2
N 2 τ [(X2X1X3 + X3X1X2)X 2
2 ].
One can carry out a similar computation for L2[f ] and LN,2[f ] and thus find L[f ] and LN,2[f ].
Since we will also deal with the Laplacians of matrix-valued functions on matrices, we also
need to define the algebraic Laplacian on operator-valued trace polynomials.
m to be the unique linear operators
Definition 3.16. We also define Lj and LN,j : TrP1
on the operator-valued trace polynomials such that
Lj[p] = LN,j[p] = η[D2
(3.10)
and the following product rule is satisfied for p ∈ NCPm and f ∈ TrP0
m:
(3.11)
j p] for p ∈ NCPm
m → TrP1
Lj[p · f ] = Lj[p] · f + p · Lj[f ]
(3.12)
LN,j[p · f ] = LN,j[p] · f + p · LN,j[f ] +
2
N 2Djp#Djf,
where Lj[f ] and LN,j[f ] are given by Definition 3.13. Then we define L = Pm
LN =Pm
j=1 LN,j.
j=1 Lj and
14
DAVID JEKEL
m is similar to the argument for TrP0
Remark 3.17. The argument for the existence and uniqueness of the operators Lj and LN,j on
TrP1
m, only that it relies on the previous scalar-valued case
since the scalar-valued case was used in the product rule.
Lemma 3.18. Let f ∈ TrP0
(3.13)
The same formula holds if f ∈ TrP1
Proof. We begin with the special case of computing the Laplacian of p ∈ NCPm (as a matrix-
valued function). To differentiate, we use the basis BN given by (2.1). Note that
m and f is viewed as a function MN (C)m
m. Viewing f is a function MN (C)m
sa → C, we have
∆jf (x) = N [LN,jf ](x)
sa → MN (C).
∆f (x) = N [LN f ](x).
f (x1, . . . , xj−1, xj + tb, xj+1, . . . , xm)
d2
dt2(cid:12)(cid:12)(cid:12)t=0
∆jp(x) = Xb∈BN
= Xb∈BN
D2
j p(x)#(b ⊗ b)
= N [η(D2
= [LN,jp](x)
j p)](x),
where the second-to-last equality follows from (2.2).
Next, we consider the case of computing the Laplacian of τN (p) (as a scalar-valued function)
for p ∈ NCPm. Since τN is a linear map MN (C) → C, we have
∆j[τN (p(x))] = τN (∆j p(x)),
where the Laplacian ∆j on the left hand side is applied to a scalar-valued function and on the
right hand side it is applied to a matrix-valued function. Therefore, it follows from the previous
computation that
∆j[τN (p(x))] = N τN ([η(D2
j p)](x)) = [LN,j[τ (p)]](x).
For the general case of scalar-valued trace polynomials, recall that the vector space of trace
m. Let
m. The Laplacian ∆j of a product of a functions can be computed using the
polynomials is spanned by elements of the form f = τ (p1) . . . τ (pN ) where pj ∈ NCP0
fj = τ (pj ) ∈ TrP0
product rule of differentiation as
∆jf (x) =
nXk=1
∆j[f (x)]Yi6=k
fi(x) +
nXk=1Xℓ6=k
Tr(∇j fk(x)∇j fℓ(x)) Yi6=k,ℓ
fi(x).
The special cased proved above shows that ∆j[fk(x)] = N [LN,jf ](x). Moreover, by (3.4), we
have ∇j [fk(x))] = 1
N [Djfk](x). Thus, we have
∆jf (x) =
nXk=1
N [LN,jfk](x)Yi6=k
fi(x) +
1
N
nXk=1Xℓ6=k
τN ([Djfk](x)[Dj fℓ](x)) Yi6=k,ℓ
fi(x).
Because of the product rule in the definition of LN,j, the right hand side equals N [LN,jf ](x).
This completes the proof of (3.13) in the scalar-valued case. The proof for the operator-valued
case is similar, using the cases proved above, as well as (3.4) and (3.6).
(cid:3)
Corollary 3.19. Let f ∈ TrP0
is a trace polynomial of lower degree than f , and we have coefficient-wise
m. If we view f as a function on MN (C)m
m or TrP1
sa, then ∆f
lim
N→∞
1
N
∆f (x) = lim
N→∞
LN f (x) = Lf (x).
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
15
Remark 3.20. We have shown that if f is a scalar-valued trace polynomial, then viewed as a
map MN (C)m
sa → C, we have
Du = N∇f,
LN f =
1
N
∆f.
Therefore, in the rest of the paper, we will freely write Df and LN f for N∇f and (1/N )∆f
for general functions f : MN (C)m
sa → C. The same considerations apply to the Laplacian for
operator-valued trace polynomials, viewed as maps MN (C)m
m or f ∈ TrP1
3.3. Convolution of Trace Polynomials and Gaussians. Let f ∈ TrP0
m.
Then viewing f is a function defined on MN (C)m
sa, we may define the convolution of f with the
probability measure σt,N (the law of an m-tuple of independent GUE). This is equivalent to the
classical convolution of f with the function MN (C)m
sa → R giving the density of the measure
σt,N . Moreover, ft = f ∗ σt,N is the solution to the heat equation with initial condition f , or
more precisely
sa → MN (C).
∂tft =
1
2N
∆ft.
(The integral formula for the solution to the heat equation with the Laplacian ∆ is well-known
[19, §2.3], and to solve the equation with (1/2N )∆ one renormalizes time by a factor of 1/2N ,
and this corresponds precisely to our normalizations in the definition of σN,t. We leave this
computation to the reader.)
We showed in the last subsection that LN = 1
N ∆ on trace polynomials is given by a purely
algebraic computation. Moreover, examining the construction of LN , one can see that it maps
trace polynomials of degree ≤ d to trace polynomials of degree ≤ d. We can view LN and L
as linear transformations on the finite-dimensional vector space of trace polynomials of degree
≤ d and define exp(tLN /2) and exp(tL/2) by the matrix exponential.
Because this holds for any d, we know that exp(tLN /2) and exp(tL/2) define linear trans-
formations TrP0
m. Moreover, a standard computation shows that
ft = exp(tLN /2)f satisfies the heat equation ∂tft = (1/2)LN ft. These observations, together
with Corollary 3.19 yield the following.
m → TrP1
m → TrP0
m and TrP1
Lemma 3.21. Let f be a trace polynomial in TrP0
m or TrP1
m. Then we have
(3.14)
with deg(exp(tLN /2)f ) ≤ deg(f ), and we have
(3.15)
σt,N ∗ f (x) = [exp(tLN /2)f ](x),
lim
N→∞
exp(tLN /2)f = exp(tL/2)f coefficient-wise.
Example 3.22. Let X = (X1, . . . , Xm) and define f (X) = Pm
2(1 ⊗ 1 ⊗ 1) for each j, and hence L[τ (f )] = 2m = LN [τ (f )]. We also have D◦
j=1 X 2
j . Note that D2
j [f (X)] =
j f = 2Xj. Hence,
L[τ (f )2] = 2L[τ (f )]τ (f ) = 4mτ (f )
LN [τ (f )2] = 2L[τ (f )τ (f ) + 2
mXj=1
τ (D◦
j f · D◦
j f )
= 4mτ (f ) +
8m
N 2 τ (f ).
Therefore, (L/2)[τ (f )2] = 2mτ (f ) and (L/2)[τ (f )] = m. Thus, the span of τ (f )2, τ (f ), and 1
is invariant under the operator L/2, and L/2 is given by a nilpotent matrix on this subspace.
Direct computation then shows that
e−tL/2[τ (f )2] = τ (f )2 + 2mtτ (f ) + m2t2.
16
DAVID JEKEL
A similar computation shows that
e−tLN /2[τ (f )2] = τ (f )2 + 2m(1 + 2/N 2)tτ (f ) + 2m2(1 + 2/N 2)t2/2.
Thus, as N → +∞, we have e−tLN /2[τ (f )2] → e−tL/2[τ (f )2].
The probabilistic interpretation of f ∗ σt,N = exp(tLN /2)f , which follows from a standard
computation, is that σt,N ∗ f (x) is the expectation of f (x + t1/2Y ), where Y is an m-tuple of
independent GUE of variance 1. Moreover, for any probability measure µ on MN (C)m
sa with
finite moments, we have
(3.16)
Z f (x) d(µ ∗ σt,N )(x) =Z (σt,N ∗ f )(x) dµ(x) =Z [exp(tLN /2)f ](x) dµ(x).
In the free setting, the operator exp(tL/2) has a similar relationship with the free convolution
with σt. This fact is standard in free probability, but because we need it for Lemmas 3.28 and
7.4 below, we include a sketch of the proof here.
Lemma 3.23. Let λ ∈ Σn,R be a non-commutative law. Then for any trace polynomial f ∈
TrP0
m, we have
(3.17)
λ ⊞ σt(f ) = λ(exp(tL/2)f ).
Proof. Because free convolution with σt forms a semigroup and exp(tL/2) is also a semigroup,
it suffices to prove that
By the product rule, it suffices to handle the case of f = τ (p) for p ∈ NCPm by showing that
λ ⊞ σt(f ) =
1
2
λ(Lf ).
λ ⊞ σt(p) =
1
2
λ(η(D2
j p)).
d
dt(cid:12)(cid:12)(cid:12)t=0
dt(cid:12)(cid:12)(cid:12)t=0
d
Let X = (X1, . . . , Xm) be a random variable with law λ and let S = (S1, . . . , Sm) be a freely
independent tuple of semi-circulars realized together in a von Neumann algebra (M, τ ). We
want to compute d
τ (p(X + t1/2S)). But note that
dt(cid:12)(cid:12)(cid:12)t=0
p(X + t1/2S) = p(X) + t1/2
mXj=1
Djp(X)#Sj +
1
2
t
mXj,k=1
DjDkp(X)#(Sj ⊗ Sk) + O(t3/2)
A moment computation with free independence shows that the terms of order t1/2 have expec-
tation zero, and so do the terms of order t with j 6= k. We are left with
d
dt(cid:12)(cid:12)(cid:12)t=0
τ (p(X + t1/2S)) =
1
2
τ (D2
j p(X)#(Sj ⊗ Sj)),
nXj=1
which using freeness evaluates to (1/2)Pn
j=1 τ (ηD2
3.4. Asymptotic Approximation by Trace Polynomials. Now we are ready to define the
approximation property which captures the asymptotic behavior of functions on MN (C)m
sa.
Definition 3.24. A sequence of functions φN : MN (C)m
sa → MN (C)m is said to be asymp-
totically approximable by trace polynomials if for every ǫ > 0 and R > 0, there exists some
f ∈ (TrP1
m)m (an m-tuple of operator-valued trace polynomials) such that
lim sup
N→∞
kxk∞≤RkφN (x) − f (x)k2 ≤ ǫ.
sup
j p(X)) = τ ((1/2)Lp(X)).
(cid:3)
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
17
m) and apply the absolute value rather than the 2-norm.
In this case, we call f an (ǫ, R)-approximation of {φN}. We make the same definitions for
functions φN : MN (C)m
sa → C, except that we use scalar-valued trace polynomials (elements of
TrP0
Observation 3.25. If f ∈ (TrP1
sa → MN (C)m given
by x 7→ f (x), then fN is asymptotically approximable by trace polynomials. Also, asymptotically
approximable sequences form a vector space over C.
m)m and if fN denotes the map MN (C)m
N }N,ℓ∈N be a sequence of functions where φ(ℓ)
Observation 3.26. Let {φ(ℓ)
MN (C)m. Suppose that {φN} is another sequence such that for every R > 0,
N : MN (C)m
sa →
lim
ℓ→∞
lim sup
N→∞
sup
kxk∞≤Rkφ(ℓ)
N (x) − φN (x)k2 = 0.
If {φ(ℓ)
N }N ∈N is asymptotically approximable by trace polynomials for each ℓ, then so is {φN}N ∈N.
The same holds in the case of scalar-valued functions and scalar-valued trace polynomials.
Lemma 3.27. Let φN , ψN : MN (C)m
sa → MN (C)m
sa. Suppose that {φN} and {ψN} are both
asymptotically approximable by trace polynomials, and furthermore suppose that {φN}N ∈N is
uniformly Lipschitz in k·k2, that is, for some L > 0,
kφN (x) − φN (y)k2 ≤ Kkx − yk2 for all x, y, for all N.
Then {φN ◦ ψN} is asymptotically approximable by trace polynomials.
Proof. Choose ǫ > 0 and R > 0. Choose a trace polynomial g which is an (ǫ/2K, R)-
approximation of {ψN}. Since g is a trace polynomial, there exists some R′ > 0 such that
for any tuple x of self-adjoint matrices of any size, we have
kxk∞ ≤ R =⇒ kg(x)k∞ ≤ R′.
Now because φN is asymptotically approximable by trace polynomials, we can choose polyno-
mial f which is an (ǫ/2, R′)-approximation of {φN}. Now we observe that when kxk∞ ≤ R
(hence kg(x)k∞ ≤ R′), we have
kφN ◦ ψN (x) − f ◦ g(x)k2 ≤ kφN ◦ ψN (x) − φN ◦ g(x)k2 + kφN ◦ g(x) − f ◦ g(x)k2
kyk∞≤R′kφN (y) − f (y)k2.
kxk∞≤RkψN (x) − g(x)k2 + sup
≤ K sup
Therefore,
lim sup
N→∞
sup
kxk∞≤RkφN ◦ ψN (x) − f ◦ g(x)k2 ≤ K ·
ǫ
2K
+
ǫ
2
= ǫ.
(cid:3)
Lemma 3.28. Suppose that φN : MN (C)m
trace polynomials and that
sa → MN (C)m
sa is asymptotically approximable by
(3.18)
kφN (x)k2 ≤ A1 +Xj
τN (x2n
j )
for some A > 0 and some integer n ≥ 0. If {φN} is asymptotically approximable by trace
polynomials, then so is {φN ∗ σt,N}.
Proof. Fix R > 0 and ǫ > 0. Choose a trace polynomial f which is an (ǫ, R + 3t1/2) approxi-
mation for {φN}. Now for x with kxk∞ ≤ R, we estimate
kσt,N ∗ φN (x) − σt,N ∗ f (x)k2 ≤Z kφN (x + y) − f (x + y)k2 dσt,N (y).
18
DAVID JEKEL
We break this integral into two pieces: The integral over the region where kyk∞ ≤ 3t1/2 is
bounded by ǫ as N → ∞ by our choice of f . Furthermore, we claim that the integral over the
region where kyk∞ > 3t1/2 vanishes as N → ∞. Using assumption (3.18) and the fact that f
is a trace polynomial, we see that there exists a C > 0 and integer d > 0, depending only on
R, A, n, and f , such that
Therefore, we have
j ) .
[kφN (x + y)k2 + kf (x + y)k2] ≤ C1 +Xj
j ) dσt,N (y).
Zkyk∞≥3t1/2kφN (x + y) − f (x + y)k2 dσt,N (y) ≤ CZkyk∞≥3t1/21 +Xj
This vanishes as N → ∞ by Corollary 2.12 applied to the GUE. Therefore, we have
τN (y2d
τN (y2d
kxk∞≤R
sup
lim sup
N→∞
kxk∞≤Rkσt,N ∗ φN (x) − σt,N ∗ f (x)k2 ≤ ǫ.
sup
On the other hand, by Lemma 3.21, we have σt,N ∗ f = exp(tLN /2)f → exp(tL/2)f coefficient-
wise, and therefore,
lim sup
N→∞
kxk∞≤Rkσt,N ∗ f (x) − [exp(tL/2)f ](x)k2 = 0,
sup
so that
lim sup
N→∞
sup
kxk∞≤Rkσt,N ∗ φN (x) − [exp(tL/2)f ](x)k2 ≤ ǫ.
(cid:3)
Lemma 3.29. Suppose that φN : MN (C)m
sa → C and suppose that {DφN} = {N∇φN} is
asymptotically approximable by trace polynomials and that φN (0) = 0. Then {φN} is asymp-
totically approximable by trace polynomials.
Proof. Given a trace polynomial F ∈ (TrP1
f (X) =Z 1
m)m, we can define
τ (F (tX)X) dt
0
in TrP0
m. Then we have
sup
kxk∞≤RφN (x) − f (x) = sup
≤ R sup
kxk∞≤R(cid:12)(cid:12)(cid:12)(cid:12)Z 1
0 hDφN (tx) − F (tx), xi2 dt(cid:12)(cid:12)(cid:12)(cid:12)
kyk∞≤RkN∇φN (y) − F (y)k2.
(cid:3)
4. Convergence of Moments
Our goal in this section is prove the following theorem. The convergence of moments is
related to [21, Theorem 4.4], [15, Proposition 50 and Theorem 51], [13, Theorem 4.4], and we
include versions of standard concentration estimates (not proved in this paper) in the statement.
sa → R be a sequence of potentials such that VN (x)− (c/2)kxk2
2
2 is concave. Let µN be the associated measure. Suppose that
Theorem 4.1. Let VN : MN (C)m
is convex and VN (x) − (C/2)kxk2
the sequence {DVN} is asymptotically approximable by trace polynomials, and assume that
(4.1)
max
M = lim sup
N→∞
j (cid:13)(cid:13)(cid:13)(cid:13)Z (xj − τN (xj)1) dµN (x)(cid:13)(cid:13)(cid:13)(cid:13) < +∞,
where 1 denotes the N × N identity matrix.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
19
(1) We have the following bounds on the operator norm. If RN = maxjR kxjk dµN (x), then
lim sup
N→∞
RN ≤
lim sup
N→∞
max
j
2
c1/2 +
2
c1/2 +
1
c
1
c
≤
(cid:12)(cid:12)(cid:12)(cid:12)Z τN (xj ) dµN (x)(cid:12)(cid:12)(cid:12)(cid:12) + M
C − c
2c3/2 + M,
lim sup
N→∞ kDVN (0)k2 +
and as a consequence of concentration we have for each j that
µN (kxjk ≥ RN + δ) ≤ e−cN δ/2.
(2) There exists a non-commutative law λ ∈ Σm,R∗ , where R∗ = lim supN→∞ RN , such
that for every non-commutative polynomial p,
N→∞Z τN (p(x)) dµN (x) = λ(p).
lim
(3) The sequence {µN} exhibits exponential concentration around λ in the sense that for
every R > 0, and every neighborhood U of λ in Σm,
lim sup
N→∞
1
N 2 log µN (kxk∞ ≤ R, λx 6∈ U) < 0.
+∞ is trivially satisfied if either (1) µN has expectation zero or (2) µN is invariant under unitary
Remark 4.2. The rather artificial hypothesis that lim supN→∞ maxj(cid:13)(cid:13)R (xj − τN (xj )) dµk(x)(cid:13)(cid:13) <
conjugation and henceR xj dµN (x) is equal toR τN (xj) dµN (x) times the identity matrix.
We have already seen in §2.5 that concentration estimates and operator norm tail bounds
are standard. To prove that the moments converge, something more is needed; indeed, the
only assumption relating the measures µN for different values of N is the fact that DVN is
asymptotically approximable by trace polynomials. But even if DVN is given by the same
"trace analytic-function" for different values of N , it is not immediate that the measure would
concentrate in the same regions for different size matrices.
To prove convergence of moments, we want to expressR u dµN in terms of DVN for a Lipschitz
function u. One of the standard techniques is to show µN is the unique stationary distribution
for a process Xt that satisfies the SDE
dXt = dYt −
1
2
DVN (Xt) dt,
where Yt is a GUE Brownian motion. This machinery lies behind the log-Sobolev inequality
and concentration results, as well as earlier theorems about convergence of moments for general
convex potentials.
Specifically, Dabrowski, Guionnet, and Shylakhtenko used the free version of this SDE to
show that for a non-commutative potential V satisfying certain convexity assumptions, there
exists a free Gibbs law for V which is the unique stationary distribution [15, Proposition 5]. As
an application, they show convergence of moments for random matrix models given by VN = V
[15, Proposition 50 and Theorem 51], essentially a special case of our Theorem 4.1.
Dabrowski was able to show convergence of moments under weaker convexity assumptions
by constructing the solution to the free SDE as an ultralimit of the finite-dimensional solutions
[13, Theorem 4.4]. Our theorem has similar convexity assumptions to Dabrowski's, but we
consider a more general sequence of potentials VN . We also perform most of our analysis in the
finite-dimensional setting, but we use deterministic rather than stochastic methods.
Instead of the solving the SDE, we study the associated semigroup T VN
, acting on Lipschitz
t
functions u, given by
T VN
t u(x) = Ex[u(Xt)],
20
DAVID JEKEL
where Xt is the process solving the SDE with initial condition x. The semigroup provides the
solution to a certain PDE, that is, if u(x, t) = Ttu0(x), then we have
∂tu =
1
2N
∆u −
N
2 ∇VN · ∇u =
1
2
LN u −
1
2hDVN , Dui2.
The semigroup T VN
then T VN
t will decrease the Lipschitz norms of functions and thus, if u is Lipschitz,
terms of DVN . We will describe a construction of the semigroup T V
through iterating simpler
t
operations (§4.1), and then we will show (Lemma 4.10) that the iteration procedure preserves
t u will converge toR u dµN as t → ∞.
Solving the differential equation and taking t → ∞ provides a way to evaluate R u dµN in
approximability by trace polynomials and hence conclude that limN→∞R u dµN exists.
4.1. Iterative Construction of the Semigroup. To simplify notation in this section, we fix
N and fix a potential V : MN (C)m
2 is convex and V (x) −
(C/2)kxk2
We will construct Tt by combining two simpler semigroups corresponding to the stochastic
and deterministic terms of dYt − (1/2)DV (Xt) dt. Recall that the solution to the heat equation
∂tu = (1/2N )∆u with initial data u0 is given by the heat semigroup:
2 is concave, and we write Tt rather than T V
t .
sa → R such that V (x) − (c/2)kxk2
Ptu0(x) =Z u0(x + y) dσt,N (y).
Stu0(x) = u0(W (x, t)),
Meanwhile, the solution to ∂tu = −(1/2)hDV, Dui2 with initial data u0 is given by
where W (x, t) is the solution to the ODE
(4.2)
∂tW (x, t) = −
1
2
DV (W (x, t))
W (x, 0) = x.
We want to define Tt = limn→∞(Pt/nSt/n)n. This is motivated by Trotter's product formula
which asserts that et(A+B) = limn→∞(etA/netB/n)n for nice enough self-adjoint operators A
and B (see [35], [24], [33, pp. 4 - 6]). But of course, we must show that (Pt/nSt/n)n converges
and derive dimension-independent error bounds.
We use the following basic properties of the semigroups Pt and St. Here if u : MN (C)m
sa → C,
then kukLip denotes the Lipschitz norm with respect to the normalized L2 metric k·k2 on
MN (C)m
sa and kukL∞ denotes the standard L∞ norm. We are only concerned with Lipschitz
functions, so in the following estimates, the reader may always assume u is Lipschitz, but of
course kukL∞ may be infinite for Lipschitz functions.
Lemma 4.3.
(1) kPtukL∞ ≤ kukL∞ .
(2) kPtukLip ≤ kukLip.
(3) kPtu − ukL∞ ≤ m1/2t1/2kukLip.
Proof. (1) and (2) follow from the fact that Ptu is u convolved with a probability measure. To
prove (3), suppose kukLip < +∞. Then
Ptu(x) − u(x) =(cid:12)(cid:12)(cid:12)(cid:12)Z (u(x + y) − u(x)) dσt,N (y)(cid:12)(cid:12)(cid:12)(cid:12)
≤Z u(x + y) − u(x) dσt,N (y)
≤ kukLipZ kyk2 dσt,N (y).
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
21
Meanwhile,
Z kyk2 dσt,N (y) ≤(cid:18)Z 1 dσt,N (y)(cid:19)(cid:18)Z kyk2
2 dσt,N (y)(cid:19)1/2
j ) dσt,N (y) = t for each j.
since y an m-tuple (y1, . . . , ym) andR τN (y2
Lemma 4.4.
= (mt)1/2,
(cid:3)
(1) The solution W (x, t) to (4.2) exists for all t.
(2) kW (x, t) − W (y, t)k2 ≤ e−ct/2kx − yk2.
(3) kW (x, t) − xk2 ≤ (t/2)kDV (x)k2.
(4) k(W (x, t) − x) − (W (y, t) − y)k2 ≤ C
(5) kStukLip ≤ e−ct/2kukLip.
(6) kStukL∞ ≤ kukL∞.
c (1 − e−ct/2)kx − yk2.
Proof. (1) The convexity and semi-concavity assumptions on V imply that DV is C-Lipschitz
and therefore global existence of the solution follows from the Picard -- Lindelof Theorem.
(2) Let V (x) = V (x) − (c/2)kxk2
2. Because V is convex, we have
which translates to
Now observe that
hD V (x) − D V (y), x − yi2 ≥ 0,
hDV (x) − DV (y), x − yi2 ≥ ckx − yk2
2.
d
dtkW (x, t) − W (y, t)k2
2 = −hDV (W (x, t)) − DV (W (y, t)), W (x, t) − W (y, t)i2
≤ −ckW (x, t) − W (y, t)k2
2,
and hence by Gronwall's inequality, kW (x, t)−W (y, t)k2
yk2
2.
(3) Note that
2 ≤ e−ctkW (x, 0)−W (y, 0)k2
2 = e−ctkx−
d
dtkW (x, t) − xk2
2 = −hDV (W (x, t)), W (x, t) − xi2
= −hDV (W (x, t)) − DV (x), W (x, t) − xi2 − hDV (x), W (x, t) − xi2
≤ kDV (x)k2kW (x, t) − xk2.
Meanwhile, kW (x, t) − xk2 is Lipschitz in t and hence differentiable almost everywhere and we
have
d
dtkW (x, t) − xk2
2 = 2kW (x, t) − xk2
d
dtkW (x, t) − xk2.
Thus, we have
which proves (3).
d
dtkW (x, t) − xk2 ≤
1
2kDV (x)k2,
22
DAVID JEKEL
(4) We observe that
k(W (x, t) − x) − (W (y, t) − y)k2 ≤
≤
≤
=
0 kDV (W (x, s)) − DV (W (y, s))k2 ds
0 kW (x, s) − W (y, s)k2 ds
C
C
1
2Z t
2 Z t
2 Z t
e−cs/2kx − yk2 ds
(1 − e−ct/2)kx − yk2.
C
c
0
(5) follows from (2).
(6) is immediate because Stu is u precomposed with another function.
(cid:3)
Now we combine Pt and St as in Trotter's formula, except that for technical convenience
we define our approximations using dyadic time intervals rather than subdividing [0, t] into
intervals of size t/n.
Lemma 4.5. For dyadic t ∈ 2−ℓN, define
Tt,ℓu = (P2−ℓ S2−ℓ)2ℓtu.
Then Ttu := limℓ→∞ Tt,ℓu exists and we have
kTt,ℓu − TtukL∞ ≤
Cm1/2
c(2 − 21/2)
2−ℓ/2kukLip.
We also have kTtukLip ≤ e−ct/2kukLip.
Proof. We want to show that the sequence {Tt,ℓu}ℓ is Cauchy by estimating the difference
between consecutive terms. Suppose that t ∈ 2−ℓN and write t = n/2ℓ and δ = 2−ℓ−1. Note
the telescoping series identity
Tt,ℓ+1u − Tt,ℓu =
n−1Xj=0
(PδSδ)2jPδ(SδPδ − PδSδ)Sδ(P2δS2δ)n−1−ju.
Thus, we want to estimate SδPδ − PδSδ and then control the propagation of the errors through
the applications of the other operators. Note that for a Lipschitz function v, we have
SδPδv(x) − PδSδv(x) ≤Z v(W (x, δ) + y) − v(W (x + y, δ)) dσδ,N (y)
≤ kvkLipZ k(W (x, δ) − x) − (W (x + y, δ) − (x + y))k2 dσδ,N (y)
(1 − e−cδ/2)Z kyk2 dσδ,N (y)
(1 − e−cδ/2)(mδ)1/2,
C
c
C
c
≤ kvkLip
≤ kvkLip
where the last inequality follows by the same reasoning as Lemma 4.3 (3). Therefore,
kSδPδv − PδSδvkL∞ ≤
C
c
m1/2δ1/2(1 − e−cδ/2)kvkLip.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
23
Therefore, we can estimate a single term in the telescoping series identity by
k(PδSδ)2jPδ(SδPδ − PδSδ)Sδ(P2δS2δ)n−1−jukL∞ ≤ k(SδPδ − PδSδ)Sδ(P2δS2δ)n−1−jukL∞
C
c
C
c
≤
≤
m1/2δ1/2(1 − e−cδ/2)kSδ(P2δS2δ)n−1−jukLip
m1/2δ1/2(1 − e−cδ/2)e−cδ/2e−cδ(n−j−1)/2kukLip.
Here we have first applied the fact that Pδ and Sδ are contractions with respect to the L∞ norm
from Lemma 4.3 (1) and Lemma 4.4 (6); second, we used the above estimate on SδPδ − PδSδ;
and third we used the estimates kPδukLip ≤ kukLip and kSδukLip ≤ e−cδ/2kukLip found in
Lemma 4.3 (2) and Lemma 4.4 (5). Now summing up the telescoping series, we get
1
≤
=
C
c
C
c
C
c
kTt,ℓ+1u − Tt,ℓukL∞ ≤
m1/2δ1/2(1 − e−cδ/2)e−cδ/2e−cδ(n−j−1)/2kukLip
n−1Xj=0
m1/2δ1/2(1 − e−cδ/2)e−cδ/2
C
m1/2δ1/2e−cδ/2kukLip ≤
2c
In other words, we have kTt,ℓ+1u − Tt,ℓukL∞ ≤ Cm1/2
It follows that the
sequence is Cauchy with respect to k·kL∞ and we have the desired estimate on kTt,ℓu− TtukL∞
from summing the geometric series.
The estimate kTt,ℓukLip ≤ e−ct/2kukLip follows from Lemma 4.3 (2) and Lemma 4.4 (5), and
1 − e−cδ/2kukLip
m1/2δ1/2kukLip.
then by taking the limit as ℓ → +∞, we obtain kTtukLip ≤ e−ct/2kukLip.
Lemma 4.6. The semigroup Tt defined above extends to a semigroup defined for positive t such
that for s ≤ t,
(cid:3)
2c 2−(ℓ+1)/2kukLip.
Ttu(x) − Tsu(x) ≤ e−cs/2(cid:18) C
c
(6 + 5√2)(t − s)1/2 + kDV (x)k2(t − s)(cid:19)kukLip,
and kTtukLip ≤ e−ct/2kukLip.
Proof. We first prove the estimate on Ttu − Tsu for dyadic values of s and t. First, consider
the case where t = 2−ℓ and s = 0. Note that
(Tt − 1)u = (Tt − PtSt)u + (Pt − 1)Stu + (St − 1)u.
The first term can be estimated by Lemma 4.5 with ℓ = 1, the second term can be estimated
by Lemma 4.3 (3) and Lemma 4.4 (5) as
k(Pt − 1)StukL∞ ≤ m1/2t1/2kStukLip ≤ m1/2t1/2kukLip,
and the third term can be estimated by Lemma 4.4 (3). Altogether, we obtain
Ttu(x) − u(x) ≤(cid:18) Cm1/2
c(2 − 21/2)
t1/2 + m1/2t1/2 +
t
2kDV (x)k2(cid:19)kukLip.
In the case of general dyadic s and t, suppose t > s and write t − s in a binary expansion to
obtain
t = s +
aj2−j,
∞Xj=n+1
24
DAVID JEKEL
where aj ∈ {0, 1} and an+1 = 1. Note that 2−n−1 ≤ s− t ≤ 2−n. Let tk = s +Pk
Then
j=n+1 aj2−j.
Ttu(x) − Tsu(x) ≤
≤
2−j
c(2 − 21/2)
2−j/2 + m1/22−j/2 +
Ttj u(x) − Ttj−1 u(x)
∞Xj=n+1
∞Xj=n+1(cid:18) Cm1/2
≤(cid:18)(cid:18) Cm1/2
+ 1(cid:19)
≤(cid:18)(cid:18) Cm1/2
+ 1(cid:19) 21/2
≤ e−cs/2(cid:18) Cm1/2
(6 + 5√2)(t − s)1/2 + kDV (x)k2(t − s)(cid:19)kukLip,
2 kDV (x)k2(cid:19)kTtj−1 ukLip
1 − 2−1/2 · 2−n/2 + kDV (x)k2 · 2−n−1(cid:19)kTsukLip
1 − 2−1/2 (t − s)1/2 + kDV (x)k2 · 2−n−1(cid:19) e−cs/2kukLip
c(2 − 21/2)
c(2 − 21/2)
1
c
where we used the crude estimate that 1 ≤ Cm1/2/c to combine the first two terms. Because
this continuity estimate holds for dyadic values of s and t, we can extend the definition of Ttu
to all positive t. Furthermore, because kTtukLip ≤ e−ct/2kukLip for dyadic t, the same must
hold for real values of t.
Now let us verify that TsTt = Ts+t for all real t. Choose dyadic sn ց s and tn ց t and let u
be a Lipschitz function. We know that Tsn Ttnu = Tsn+tn u and that Tsn+tnu → Ts+tu locally
uniformly, so it suffices to show that Tsn Ttnu → TsTtu. Observe that
Tsn Ttnu − TsTtu ≤ (Tsn − Ts)Ttn u + Ts(Ttn − Tt)u.
The first term can be estimated by
(Tsn − Ts)Ttn u(x) ≤ e−s/2(cid:18) C
(Ttn − Tt)u(x) ≤ e−t/2(cid:18) C
c
c
(6 + 5√2)(sn − s)1/2 + kDV (x)k2(sn − s)(cid:19) kTtnukLip,
(6 + 5√2)(tn − t)1/2 + kDV (x)k2(tn − t)(cid:19)kukLip
which goes to zero as n → ∞. For the second term, we first note that
Let hn(x) be the right hand side. Note that u ≤ v implies that Tsu ≤ Tsv because this holds
for the operators Ps and Ss (since Ps is given by convolution and Ss is given by composition).
Therefore,
Ts(Ttn − Tt)u(x) ≤ Ts(Ttn − Tt)u(x) ≤ Tshn(x).
Because DV is C-Lipschitz, we know that hn is a e−t/2(tn − t)CkukLip-Lipschitz function and
hence
Tshn(x) ≤ hn(x) + (Ts − 1)hn(x)
≤ hn(x) + e−t/2(tn − t)CkukLip(cid:18) C
c
(6 + 5√2)s1/2 + kDV (x)k2s(cid:19) ,
which goes to zero as n → ∞.
Lemma 4.7. Let u(x) be Lipschitz. Then Ttu is a weak solution of the equation
(cid:3)
∂tTtu =
1
2N
∆(Ttu) −
N
2 ∇V · ∇(Ttu)
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
25
in the sense that for φ ∈ C∞
c (MN (C)m
sa), we have
ZMN (C)m
sa
[Tt1 u φ − Tt0u φ] =Z t1
sa(cid:20)−
t0 ZMN (C)m
1
2N ∇(Tsu) · ∇φ −
N
2
(∇V · ∇(Tsu))φ(cid:21) ds.
Proof. Recall that by Rademacher's theorem if u is Lipschitz, then ∇u exists almost everywhere
and it is in L∞. Moreover, because ∇V is Lipschitz, we also know that the second derivatives
of V exist almost everywhere and are in L∞.
sa → R and a φ ∈
We begin by considering R (SδPδ − 1)u · φ for a Lipschitz u : MN (C)m
sa) and δ > 0. Note that
c (MN (C)m
C∞
(SδPδ − 1)u = (Sδ − 1)Pδu + (Pδ − 1)u.
Now Pδu is the convolution of u with the Gaussian and so ∇(Pδu) = Pδ(∇u). Because the
gradient of the Gaussian is O(δ−1/2), we see that the first derivatives of Pδ(∇u) are O(δ−1/2)
in L∞. (Here our estimates may depend on N .)
Pδu(y) − Pδu(x) = ∇Pδu(x) · (x − y) + O(δ−1/2kx − yk2
2).
Now using equation (4.2) and Lemma 4.4 (3), we have W (x, δ) − x = N δ
uniformly on any compact set K. Therefore,
2 ∇V (x) + O(δ2)
(Sδ − 1)Pδu(x) = Pδu(W (x, δ)) − Pδu(x) = −
N δ
2 ∇(Pδu)(x) · ∇V (x) + O(δ3/2).
Now we have
Z (SδPδ − 1)u · φ =Z (Sδ − 1)Pδu φ +Z (Pδ − 1)u φ
N δ
2 Z [∇(Pδu) · ∇V ]φ +Z u (Pδ − 1)φ + O(δ3/2)
2 Z Pδu[(∆V )φ + ∇V · ∇φ] +Z u
2 Z u Pδ[(∆V )φ + ∇V · ∇φ] +
2N Z u ∆φ + O(δ3/2),
∆φ + O(δ3/2)
δ
2N
δ
N δ
N δ
= −
= −
= −
where the error estimates depend only on C, N , kukLip, the support of φ, and the L∞ norms
of its derivatives. We also know from the proof of Lemma 4.5 that (SδPδ − PδSδ)u is bounded
by kukLip(Cm1/2/c)(1 − e−cδ)δ1/2 which is O(δ3/2). Therefore,
Z (PδSδ − 1)u · φ = −
N δ
2 Z u Pδ[∆V φ + ∇V · ∇φ] +
δ
2N Z u ∆φ + O(δ3/2).
Now suppose that t is a dyadic rational and write t = nδ where δ = 2−ℓ for some integer ℓ.
Recall that Tt,ℓ = (PδSδ)n. Then by a telescoping series argument
Z (Tt,ℓ − 1)u · φ =
n−1Xj=0(cid:18)−
N δ
2 Z Tjδ,ℓu Pδ[(∆V )φ + ∇V · ∇φ] +
δ
2N Z Tjδ,ℓu ∆φ(cid:19) + O(δ1/2).
We fix a dyadic t and take ℓ → ∞ (and hence δ → 0). The above sum over j may be
viewed as a Riemann sum for an integral from 0 to t where δ is the mesh size. Using Lemma
4.6, we know that Ttu is Holder continuous in t. Also, by Lebesgue differentiation theory,
Pδ[(∆V )φ + ∇V · ∇φ] → (∆V )φ + ∇V · ∇φ in L1
loc. Thus, in the limit, we obtain
Z (Tt − 1)u · φ dx =Z t
0 Z (cid:18)−
N
2
Tsu[(∆V )φ + ∇V · ∇φ] +
1
2N
Tsu ∆φ(cid:19) dx ds.
26
DAVID JEKEL
We pass from dyadic t to all positive t using Lemma 4.6. Finally, after another integration
by parts (which is justified by approximation by smooth functions in the appropriate Sobolev
spaces), we have
Z (Tt − 1)u · φ dx =Z t
0 Z (cid:18)−
N
2
(∇Tsu · ∇V )φ −
1
2N ∇Tsu · ∇φ(cid:19) dx ds.
The asserted formula then follows by replacing u with Tt0 u and t with t1 − t0.
Lemma 4.8. If µ is the measure given by the potential V and if u is Lipschitz, then we have
(cid:3)
Proof. By applying Lemma 4.7 and approximating (1/Z) exp(−N 2V (x)) by compactly sup-
ported smooth functions, we see that
R Ttu dµ =R u dµ.
Z Ttu dµ −Z u dµ
0 (cid:20)−
ZZ Z t
(cid:12)(cid:12)(cid:12)(cid:12)Ttu(x) −Z u dµ(cid:12)(cid:12)(cid:12)(cid:12) ≤ e−ct/4(cid:18) 4Cm1/2
c2
=
1
1
Proof. Fix t and fix r ≥ t. Let n be an integer. Then
2N ∇(Tsu) · ∇[e−N 2V ] −
Lemma 4.9. We have Ttu(x) →R u dµ as t → ∞ and more precisely
2
(6 + 5√2)t−1/2 +
Tt+ru(x) − Ttu(x) ≤
Tt+r(j+1)/nu(x) − Tt+rj/nu(x)
N
2
(∇V · ∇(Tsu))e−N 2V(cid:21) ds dx = 0. (cid:3)
ckV (x)k2(cid:19)kukLip.
n−1Xj=0
n−1Xj=0
≤
≤ e−ct/2ecr/2n 2n
e−ct/2e−crj/2n(cid:18) Cm1/2
cr (cid:18) Cm1/2
Tt+ru(x) − Ttu(x) ≤ e−ct/4(cid:18) 4Cm1/2
c2
c
c
(6 + 5√2)(r/n)1/2 + kV (x)k2(r/n)(cid:19) kukLip
(6 + 5√2)(r/n)1/2 + kV (x)k2(r/n)(cid:19) kukLip
(6 + 5√2)t−1/2 +
ckV (x)k2(cid:19)kukLip.
2
Since r ≥ t, we can choose n such that t/4 ≤ r/n ≤ t/2. Then we have
Because this holds for all sufficiently large r, this shows that limt→∞ Ttu(x) exists. Because
kTtukLip ≤ e−ct/2kukLip, the limit must be constant and therefore equalsR u dµ. Moreover, we
have the asserted rate of convergence by taking r → ∞ in the above estimate.
4.2. Approximability and Convergence of Moments. Now we are ready to show that
the semigroup T VN
associated to a sequence of potentials VN will preserve asymptotic approx-
imability by trace polynomials and as a consequence we will show that the moments of the
associated measures µN converge.
(cid:3)
t
and T VN
sa → R be a sequence of potentials such that VN (x)− (c/2)kxk2
Lemma 4.10. Let VN : MN (C)m
2
is convex and VN (x)− (C/2)kxk2
2 is concave. For each N , let µN be the associated measure. Let
SVN
denote the semigroups defined in the previous section. Suppose that the sequence
{DVN} is asymptotically approximable by trace polynomials. Suppose that {uN} is a sequence
of scalar-valued K-Lipschitz functions which is asymptotically approximable by (scalar-valued)
trace polynomials. Then
t
t
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
27
t uN} is asymptotically approximable by trace polynomials for each t ≥ 0.
t uN} is asymptotically approximable by trace polynomials for each t ≥ 0.
(1) {SVN
(2) {T VN
(3) limN→∞R uN dµN exists.
Proof. (1) Recall that SVN
t uN = uN (WN (x, t)), where WN is the solution to (4.2). Thus,
by Lemma 3.27, it suffices to show that WN (x, t) is asymptotically approximable by trace
polynomials for each t. To this end, we write WN (x, t) as the limit as ℓ → ∞ of Picard iterates
WN,ℓ given by
WN,0(x, t) = x,
WN,ℓ+1(x, t) = x −
DVk(WN (x, s)) ds.
Because DVN is C-Lipschitz, the standard Picard-Lindelof arguments show that
1
2Z t
0
kWN,ℓ(x, t) − WN (x, t)k2 ≤
∞Xn=ℓ+1
Cn−1tn
2nn! kDVN (x)k2.
Because DVN is asymptotically approximable by trace polynomials, we know that kDVN (x)k2
is uniformly bounded on kxk ≤ R for any given R > 0, and therefore for each T and R > 0, the
convergence of WN,ℓ to WN as ℓ → ∞ is uniform for all kxk ≤ R and t ≤ T and N ∈ N. Thus,
by Observation 3.26, it suffices to show that each Picard iterate {WN,ℓ(x, t)}N is asymptotically
approximable by trace polynomials.
Fix T > 0. We claim that for every ℓ, for every R > 0 and ǫ > 0, there exists a trace
polynomial f (X, t) with coefficients that are polynomial functions of t, such that
lim sup
N→∞
sup
t∈[0,T ]
kxk∞≤RkWN,ℓ(x, t) − f (x, t)k2 ≤ ǫ.
sup
We proceed by induction on ℓ, with the base case ℓ = 0 being trivial. For the inductive step,
fix ǫ and R, and choose a trace polynomial f (X, t) which provides a (ǫ/CT, R) approximation
for WN,ℓ for all t ≤ T . Let
R′ = sup
t∈[0,T ]
sup
N
sup
sa:kxk∞≤Rkf (x, t)k < +∞.
x∈MN (C)m
Choose another trace polynomial g(X) which is an (ǫ/T, R′) approximation for {DVN}, and
2R t
let h(X, t) = X − 1
0 g(f (X, s)) ds. Then arguing as in Lemma 3.27, we have for kxk ≤ R and
t ∈ [0, T ] that
kWN,ℓ+1(x, t) − h(x, t)k ≤
1
2Z t
0 kDVN (WN,ℓ(x, s)) − g(f (x, s))k2 ds
sup
kyk≤R′kDVN (y) − g(y)k2 +
Ct
2
s∈[0,T ]
sup
t
2
≤
(2) We have shown that SVk
t
sup
kxk≤RkWN,ℓ(x, s) − f (x, s)k2.
Taking N → ∞, we see that h(x, t) is an (ǫ, R) approximation for {WN,ℓ(x, t)}N for all t ≤ T .
preserves asymptotic approximability. Moreover, if the se-
quence uN : MN (C)m
sa → C is asymptotically approximable by trace polynomials and uN is
K-Lipschitz, then the sequence PtuN is also asymptotically approximable by trace polynomials
by Lemma 3.28 (the hypothesis (3.18) is satisfied since uN (x) ≤ uN (0) + Lkxk2 and uN (0)
is bounded as N → +∞ because uN is asymptotically approximable by trace polynomials).
2−ℓ)2ℓt preserves asymptotic approximability for
Therefore, the iterated operator T Vk
dyadic values of t. Taking ℓ → ∞, we see by Observation 3.26 and Lemma 4.5 that T Vk
pre-
serves asymptotic approximability for dyadic values of t. Finally, we extend the approximability
property to T Vk
for all real t using Observation 3.26 and Lemma 4.6.
t,ℓ = (P2−ℓ SVk
t
t
Therefore,
C + c
2
(cid:13)(cid:13)(cid:13)(cid:13)DVN (x) −(cid:18)DVN (0) +
(cid:13)(cid:13)(cid:13)(cid:13)DVN (0) +
aN(cid:13)(cid:13)(cid:13)(cid:13)2 ≤
C + c
2
C − c
2
kxk2.
C − c
2
x(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)2
= kD VN (x) − D VN (0)k2 ≤
Z kxk2 dµN (x)
2 kaNk2 +(cid:18)Z kx − aNk2
(cid:16)kaNk2 + c−1/2(cid:17) ,
2
C − c
C − c
≤
≤
2 dµ(x)(cid:19)1/2!
28
DAVID JEKEL
(cid:3)
(3) We know by Lemma 4.9 that T VN
are independent of N . It follows by Observation 3.26 that the sequence of constant functions
t uN (x) → R uN dµN as t → ∞ with estimates that
{R uN dµN} is asymptotically approximable by trace polynomials. But since these functions
are constant, this simply means that the limit ofR uN dµN as N → ∞ exists.
Proof of Theorem 4.1. (1) Let aN =R x dµN (x) and aN,j =R xj dµN (x). Note that
j (cid:13)(cid:13)(cid:13)(cid:13)Z (xj − τN (xj )) dµ(x)(cid:13)(cid:13)(cid:13)(cid:13).
RN ≤ max
When we take the lim sup as N → ∞, the first term is bounded by 2/c1/2 by Corollary 2.12
j Z kxj − aN,jk dµN (x) + max
while the last term is bounded by M . It remains to estimateR τN (xj ) dµN (x).
On the other hand, we may estimate(cid:13)(cid:13)DVN (x) −(cid:0)DVN (0) + C+c
(cid:12)(cid:12)(cid:12)(cid:12)Z τN (xj ) dµN (x)(cid:12)(cid:12)(cid:12)(cid:12) + max
Z DVN (x) dµN (x) = 0
2 is convex and VN (x) − (C/2)kxk2
2 (C−c)kxk2
2 is convex and VN (x)− 1
that VN (x) − (c/2)kxk2
2 kxk2
2. Then VN (x)+ 1
D VN is (C − c)/2-Lipschitz with respect to k·k2. It follows that
2 x(cid:1)(cid:13)(cid:13)2 as follows. We assumed
2 is concave. Let VN (x) = VN (x) −
2 (C−c)kxk2
2 is concave. Therefore,
Using integration by parts, we see that
C+c
j
where the last step follows from Theorem 2.11. Altogether,
C + c
2
kaNk2 ≤
C − c
2
kaNk2 + kDVN (0)k2 +
C − c
2c1/2 .
Then we move (C − c)/2 · kaNk2 to the left hand side and divide the equation by c to obtain
(cid:12)(cid:12)(cid:12)(cid:12)Z τN (xj) dµN (x)(cid:12)(cid:12)(cid:12)(cid:12) ≤ kaNk2 ≤
1
ckDVN (0)k2 +
C − c
2c3/2 ,
which proves the asserted estimate on RN . The tail estimate on µN (kxjk ≥ RN + δ) follows
from Corollary 2.12.
(2) Fix a non-commutative polynomial p. Let R∗ = lim supN ∈N RN which we know is
finite because of (1) and suppose that R′ > R∗. Let ψ ∈ C∞
c (R) be such that ψ(t) = t for
t ≤ R′, and define Ψ(x1, . . . , xm) = (ψ(x1), . . . , ψ(xm)), where ψ(xj ) is defined through the
continuous functional calculus for self-adjoint operators. Now x 7→ ψ(x) is Lipschitz in k·k2
for x ∈ MN (C)sa with constants independent of N (see for instance Proposition 8.8 below). It
follows that p(Ψ(x)) is globally Lipschitz in k·k2 and it equals p(x) when kxk ≤ R′.
Furthermore, we claim that the sequence τN (p(Ψ(x))) is asymptotically approximable by
trace polynomials. To see this, choose some radius r and δ > 0. By the Weierstrass ap-
proximation theorem, there exists a polynomial bψ(t) such that ψ(t) − bψ(t) ≤ δ for t ∈
[−r, r]. By the spectral mapping theorem, we have kψ(y) − bψ(y)k ≤ δ if y ∈ MN (C)sa and
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
29
kyk ≤ r. In particular, if we let bΨ(x) = (bψ(x1), . . . ,bψ(xm)) for x ∈ MN (C)m
kΨ(x) − bΨ(x)k ≤ δ when kxk∞ ≤ r. Given ǫ > 0, we may choose δ small enough to guar-
antee that τN (p(Ψ(x))) − τN (p(bΨ(x))) ≤ ǫ for kxk∞ ≤ r, and clearly τN (p(bΨ(x))) is a trace
polynomial. Thus, τN (p(Ψ(x)) is asymptotically approximable by trace polynomials.
Therefore, by Lemma 4.10, the limit
sa, then we have
λ(p) = lim
N→∞Z τN (p(Ψ(x)) dµN (x)
exists. Clearly, λ satisfies all the conditions to be a non-commutative law. Furthermore,
because of the operator norm bounds (1), we know thatRkxk≥R′ τN (p(x)) dµN (x) is finite and
approaches zero as N → ∞ and the same holds for the integral of τN (p(Ψ(x))). Therefore,
N→∞Z τN (p(x)) dµN (x) = lim
N→∞Z τN (p(Ψ(x)) dµN (x) = λ(p).
lim
Also, we have λ(p) = limN→∞Rkxk≤R′ τN (p(x)) dµN (x) and hence λ ∈ Σm,R′ . But since this
holds for every R′ > R∗, we have λ ∈ Σm,R∗.
(3) It suffices to prove the concentration claim (3) for sufficiently large R, say R > 2R∗.
Because the topology of Σm,R is generated by the functions λ 7→ λ(p) for non-commutative
polynomials p, it suffices to consider the case where U = {λ′ : λ′(p) − λ(p) < ǫ} for some
non-commutative polynomial p. Choose a function ψ ∈ C∞
c (R) with ψ(t) = t for t ≤ R, and
let Ψ be as above. Then by Theorem 2.10,
µN(cid:18)(cid:12)(cid:12)(cid:12)(cid:12)τN (p(Ψ(x)) −Z τN (p ◦ Ψ) dµN(cid:12)(cid:12)(cid:12)(cid:12) ≥ ǫ/2(cid:19) ≤ 2e−N 2ǫ2/8kτN (p◦Ψ)k2
But by the same reasoning as in part (2), we know that large enough N , we have
Lip .
(cid:12)(cid:12)(cid:12)(cid:12)Z τN (p ◦ Ψ) dµN − λ(p)(cid:12)(cid:12)(cid:12)(cid:12) ≤
ǫ
2
,
and hence
lim sup
N→∞
1
N 2 log µN (kxk ≤ R, τN (p(x)) − λ(p) ≥ ǫ) < 0.
(cid:3)
5. Entropy and Fisher's Information
5.1. Classical Entropy. In this section, we will state sufficient conditions for the microstates
free entropies χ and χ to be evaluated as the lim sup and lim inf of renormalized classical
entropies. Recall that the (classical, continuous) entropy of a measure dµ(x) = ρ(x) dx on Rn
is defined as
h(µ) :=ZRn −ρ log ρ,
whenever the integral makes sense. We will later use the following basic facts about the classical
entropy, so for convenience we provide a proof.
Lemma 5.1.
(1) If µ has a density ρ andR x2 dµ(x) < +∞, then the positive part of −ρ log ρ has finite
integral and henceR −ρ log ρ is well-defined in [−∞, +∞).
(2) In fact, we have h(µ) ≤ (n/2) log 2πae, where a = R x2 dµ(x)/n, and equality is
pointwise almost everywhere, and suppose that R x2 dµk(x) → R x2 dµ(x) < +∞.
(3) Suppose {µk} is a sequence of probability measures with density ρk, suppose ρk → ρ
achieved in the case of a centered Gaussian with covariance matrix aI.
Then lim supk→∞ h(µk) ≤ h(µ).
(4) If µ and ν have finite second moments, then h(µ ∗ ν) ≥ max(h(µ), h(ν)).
30
DAVID JEKEL
Proof. (1) Let a =R x2 dµ(x)/n. Let g(x) = (2πa)−n/2e−x2/2a be the Gaussian of variance
a and let γ be the corresponding Gaussian measure. Let ρ = ρ/f be the density of µ relative
to the Gaussian. We write
−ρ(x) log ρ(x) = − ρ(x) log ρ(x) · g(x) − ρ(x) log g(x) · g(x)
= − ρ(x) log ρ(x) · g(x) +(cid:18) 1
2ax2 +
n
2
log 2πa(cid:19) ρ(x).
The second term has a finite integral by assumption. The function −t log t is bounded above
for t ∈ R, and g(x) is a probability density; thus, the positive part of − ρ log ρ · g has finite
(2) The function −t log t is concave and its tangent line at t = 0 is 1− t and hence −t log t ≤
integral. Hence,R −ρ log ρ is well-defined.
1 − t. Thus,
so
Z − ρ log ρ dγ ≤Z (1 − ρ) dγ = 0,
h(µ) ≤Z (cid:18) 1
2ax2 +
n
2
log 2πa(cid:19) ρ(x) dx =
n
2
n
2
+
log 2πa =
n
2
log 2πe.
In the case where µ = γ, we have ρ = 1 and henceR − ρ log ρ = 0.
(3) Let γ be the Gaussian of covariance matrix I and g its density. Let ρk = ρk/g. As before,
h(µk) =Z − ρk log ρk dγ +Z (cid:18) 1
By assumption, the second term converges toR ( 1
2 log 2π) dµ. Since the function −t log t
is bounded above and γ is a probability measure, the integral of the positive part of − ρk log ρk
converges to the corresponding quantity for ρ. For the negative part, we can apply Fatou's
lemma. This yields lim supk→∞ h(µk) = h(µ).
2x2 + n
log 2π(cid:19) dµk.
2x2 +
n
2
(4) We can assume without loss of generality that one of the measures, say µ, has finite
entropy and in particular has a density ρ. Then µ ∗ ν has a density given almost everywhere
− ρ(x) log ρ(x) ≥Z −ρ(x − y) log ρ(x − y) dν(y).
by ρ(x) =R ρ(x − y) dν(y). Since −t log t is concave, Jensen's inequality implies that
The right hand side isRR −ρ(x−y) log ρ(x−y) dν(y) dx =RR −ρ(x−y) log ρ(x−y) dx dν(y) =
h(µ) > −∞. Therefore, h(µ ∗ ν) =R − ρ log ρ ≥ h(µ).
h(µ), where the exchange of order is justified because we know that −ρ log ρ is integrable since
5.2. Microstates Free Entropy. Because there is no integral formula known for free entropy
of multiple non-commuting variables as in the classical case, Voiculescu defined the free analogue
of entropy [38, 39] using Boltzmann's microstates viewpoint on entropy.
(cid:3)
Definition 5.2. For U ⊆ Σm, we define the microstate space
sa : λx ∈ U}
sa : λx ∈ U,kxk∞ ≤ R}.
The microstates free entropy of a non-commutative law λ is defined as
ΓN (U) = {x ∈ MN (C)m
ΓN,R(U) = {x ∈ MN (C)m
N→∞ (cid:18) 1
lim sup
χR(λ) = inf
U ∋λ
χ(λ) = sup
R>0
χR(λ).
N 2 log vol ΓN,R(U) +
m
2
log N(cid:19)
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
31
Here U ranges over all open neighborhoods of λ in Σm. Similarly, we denote
N→∞ (cid:18) 1
lim inf
χR(λ).
χ
(λ) = inf
U ∋λ
R
χ(λ) = sup
R>0
N 2 log vol ΓN,R(U) +
m
2
log N(cid:19)
Definition 5.3. A sequence of probability measures µN on MN (C)m
sa is said to concentrate
around the non-commutative law λ if λx → λ in probability when x is chosen according to µN ,
that is, for any neighborhood U of λ in Σm, we have
µN (x ∈ ΓN (U)) = 1.
lim
k→∞
Proposition 5.4. Let VN : MN (C)m
and let µN be the associated measure. Assume:
(A) The sequence {µN} concentrates around a non-commutative law λ.
(B) The sequence {VN} is asymptotically approximable by scalar-valued trace polynomials.
j ).
(D) There exists R0 > 0 such that
sa → R be a potential with R exp(−N 2VN (x)) dx < +∞
(C) For some n ≥ 1 and a, b > 0 we have VN ≤ a + bPm
j=1 τN (x2n
lim
N→∞Zkxk∞≥R01 +
mXj=1
τN (x2n
where n is the same number as in (C).
It follows that λ(X 2n
for any n > 0. From here it is a standard fact that λ can be
realized by self-adjoint random variables in a tracial von Neumann algebra which have norm
≤ R0.
j ) ≤ R2n
0
Now let us evaluate χR and χ
for R ≥ R0. Recall that
R
dµN (x) =
1
ZN
and note that
ZN =Z exp(−N 2VN (x)) dx,
exp(−N 2VN (x)) dx,
h(µN ) = N 2Z VN (x) dµN (x) + log ZN .
The assumptions (C) and (D) imply that
N→∞Zkxk∞≥R VN (x) dµN (x) = 0 and lim
lim
N→∞
µN (x : kxk∞ ≥ R) = 0.
Therefore, if we let
1
dµN,R(x) =
ZN,R
1kxk∞≤R exp(−N 2VN (x)) dx,
ZN,R =Zkxk∞≤R
exp(−N 2VN (x)) dx.
j ) dµN (x) = 0,
log N(cid:19)
log N(cid:19) .
m
2
m
2
N→∞ (cid:18) 1
N→∞ (cid:18) 1
N 2 h(µN ) +
N 2 h(µN ) +
Then λ can be realized as the law of non-commutative random variables X = (X1, . . . , Xm) in
a von Neumann algebra (M, τ ) with kXjk ≤ R0. Moreover, we have
(5.1)
χ(λ) = χR0(λ) = lim sup
(5.2)
χ(λ) = χ
R0
(λ) = lim inf
Proof. It follows from assumptions (A) and (D) that for every non-commutative polynomial p,
N→∞Zkxk∞≤R0
lim
τN (p(x)) dµN (x) = λ(p).
32
DAVID JEKEL
then as N → ∞, we have
Z VN dµN −Z VN dµN,R → 0,
and hence
log ZN − log ZN,R → 0,
1
N 2 h(µN ) −
1
N 2 h(µN,R) → 0.
Fix ǫ > 0. By assumption (B), there is scalar-valued trace polynomial f such that VN (x) −
f (x) ≤ ǫ/2 for kxk∞ ≤ R and for sufficiently large N . Now because the trace polynomial
f is continuous with respect to convergence in non-commutative moments, the set U = {λ′ :
λ′(f ) − λ(f ) < ǫ/2} is open. Now suppose that V ⊆ U is a neighborhood of λ. Note that
lim
N→∞
µN,R(ΓN,R(V)) = lim
N→∞
ZN
ZN,R
µN (ΓN (V) ∩ {x : kxk∞ ≤ R}) = 1,
where we have used that ZN /ZN,R → 1 as shown above, that µN (ΓN (V)) → 1 by assumption
(A), and that µN (kxk∞ ≤ R) → 1 by assumption (D). Moreover, by our choice of f and U, we
have
x ∈ ΓN,R(V) =⇒ VN (x) − λ(f ) ≤ ǫ.
Therefore,
Thus,
ZN,RµN,R(ΓN,R(V)) =ZΓN,R(V)
exp(−N 2VN (x)) dx
= vol ΓN,R(V) exp(−N 2(λ(f ) + O(ǫ))).
Meanwhile, note that f is bounded by some constant K whenever kxk∞ ≤ R. Therefore,
log ZN,R + log µN,R(ΓN,R(V)) = log vol ΓN,R(V) − N 2(λ(f ) + O(ǫ)).
Z VN dµN,R =ZΓN,R(V)
=ZΓN,R(V)
= λ(f )µN,R(ΓN,R(V)) + O(ǫ) + O(cid:0)K µN (ΓN,R(V c))(cid:1).
VN dµN,R +ZΓN,R(V c)
λ[f ] dµN,R +ZΓN,R(V c)
λx[f ] dµN,R + O(ǫ)
VN dµN,R
Altogether,
1
N 2 h(µN,R) =Z VN dµN,R +
1
N 2 log ZN,R
=λ(f )(µN,R(ΓN,R(V)) − 1) +
+ O(ǫ) + O(cid:0)K µN (ΓN,R(V c))(cid:1) −
1
N 2 log vol ΓN,R(V)
1
N 2 log µN,R(ΓN,R(V)).
Now we apply the fact that µN,R(ΓN,R(V)) → 1 to obtain
lim sup
N→∞
1
N 2h(µN,R) − log vol ΓN,R(V) = O(ǫ).
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
33
Because this holds for all sufficiently small neighborhoods V with the error O(ǫ) only depending
on U, we have
χR(λ) = lim sup
N→∞ (cid:18) 1
N→∞ (cid:18) 1
= lim sup
N 2 h(µN,R) +
m
2
N 2 h(µN ) +
m
2
log N(cid:19) + O(ǫ)
log N(cid:19) + O(ǫ).
Next, we take ǫ → 0 and obtain χR(λ) = lim supN→∞(N −2 log h(µN ) + (m/2) log N ) for
R ≥ R0. Now χ(λ) = supR χR(λ) and χR(λ) is an increasing function of R. Since our claim
about χR(λ) holds for sufficiently large R, it also holds for χ(λ), so (5.1) is proved. The proof
of (5.2) is identical.
(cid:3)
5.3. Classical Fisher Information. The classical Fisher information of a probability measure
µ on Rn describes how the entropy changes when µ is convolved with a Gaussian. Suppose µ
is given by the smooth density ρ > 0 on Rn, and let γt be the multivariable Gaussian measure
on Rn with covariance matrix tI. Then the density ρt for µt = µ ∗ γt evolves according to the
(which we justify in more detail below).
heat equation ∂tρt = (1/2)∆ρt. Integration by parts shows that ∂th(µt) = (1/2)R ∇ρt/ρt2dµt
The Fisher information of µ represents the derivative at time zero and it is defined as
The Fisher information is the L2(µ) norm of the function −∇ρ(x)/ρ(x), which is known as the
score function. If X is a random variable with smooth density ρ, then the Rn-valued random
variable Ξ = −∇ρ(X)/ρ(X) satisfies the integration-by-parts relation
2
dµ.
∇ρ
ρ (cid:12)(cid:12)(cid:12)(cid:12)
I(µ) :=Z (cid:12)(cid:12)(cid:12)(cid:12)
f (x)ρ(x) dx =Z ρ(x)∇f (x) dx = E[∇f (X)] for f ∈ C∞
(5.3) E[Ξ· f (X)] = −Z ∇ρ(x)
ρ(x)
c (Rn),
or equivalently E[Ξjf (X)] = E[∂jf (X)] for each j.
In fact, the integration-by-parts relation E[Ξ · f (X)] = E[∇f (X)] makes sense even if we
do not assume that X has a smooth density. Following the terminology used by Voiculescu
in the free case, if X is an Rn-valued random variable on the probability space (Ω, P ), we say
that an Rn-valued random variable Ξ ∈ L2(Ω, P ) is a (classical) conjugate variable for X if
E[Ξ · f (X)] = E[∇f (X)] and if each Ξj is in the closure of {f (X) : f ∈ C∞
c (Rn)} in L2(Ω, P ).
In other words, this means that Ξ is a function of X (up to almost sure equivalence) and
satisfies the integration-by-parts relation. Since the integration-by-parts relation uniquely de-
c (Rn), it follows that the
termines the L2(Ω, P ) inner product of Ξj and f (X) for all f ∈ C∞
conjugate variable is unique (up to almost sure equivalence), and it is also independent of (Ω, P )
and only depends on the law of X. Thus, we may unambiguously define the Fisher information
I(µ) = E[Ξ2] if Ξ is a conjugate variable to X and I(µ) = +∞ if no conjugate variable exists.
The probabilistic viewpoint is useful because it enables us to produce conjugate variables
and estimate Fisher information using conditional expectation. (See [40, Proposition 3.7] for
the free case.)
Lemma 5.5. Suppose that X and Y are independent Rn-valued random variables with X ∼ µ
and Y ∼ ν. If Ξ is a conjugate variable for X, then E[ΞX + Y ] is a conjugate variable for
X + Y . In particular,
I(µ ∗ ν) ≤ min(I(µ),I(ν)).
34
DAVID JEKEL
Proof. Because X and Y are independent, we have for g ∈ C∞
E[∂Xj g(X, Y )]. In particular, if f ∈ C∞
c (Rn), then
c (Rn × Rn) that E[Ξj g(X, Y )] =
E[Ξjf (X + Y )] = E[∂Xj (f (X + Y ))] = E[(∂jf )(X + Y )].
But E[ΞjX +Y ] is the orthogonal projection onto the closed span of {f (X +Y ) : f ∈ C∞
and hence
c (Rn)}
So I(µ ∗ ν) = E[E[ΞX + Y ]2] ≤ E[Ξ2] = I(µ). By symmetry, I(µ ∗ ν) ≤ I(ν).
E [E[ΞjX + Y ]f (X + Y )] = E[∂jf (X + Y )].
(cid:3)
The entropy of a measure µ can be recovered by integrating the Fisher information of µ∗ γt.
The following integral formula was the motivation for Voiculescu's definition of non-microstates
free entropy χ∗. For the reader's convenience, we include a statement and proof in the random
matrix setting with free probabilistic normalizations. See also [4, Lemma 1] and [40, Proposition
7.6]. Recall that we identify MN (C)m
using the orthonormal basis given in §2.1
rather than entrywise coordinates (since some entries are real and some are complex).
sa with RmN 2
Lemma 5.6. Let µ be a probability measure on MN (C)m
sity ρ, and let σt,N be the law of m independent GUE's of normalized variance t.
sa with finite variance and with den-
If a =
(1/m)R kxk2
(5.4)
2 dµ(x) = (1/mN )R x2 dµ(x), then we have for t ≥ 0 that
N 3I(µ)(cid:19) .
m
a + t ≤
1
1
t
,
1
N 2 h(µ) +
m
2
log N =
m
2
log 2πe.
Proof. To prove (5.4), suppose t ≥ 0 and let X and Y be random variables with the laws µ and
σt,N respectively. The lower bound is trivial if I(µ ∗ σt,N ) = +∞, so suppose that X + Y has
a conjugate variable Ξ. Then after some computation, the integration-by-parts relation shows
that EhΞ, X + Y iTr = mN 2. Thus,
E[Ξ2] ≥ EhΞ, X + Y iTr2
EX + Y 2
=
(mN 2)2
N (ma + mt)
=
N 3
a + t
since the variance of Y with respect to the non-normalized inner product is N mt and the
variance of X is N a. The upper bound is trivial in the case where t = 0. If t > 0, then by the
previous lemma I(µ ∗ σt,N ) ≤ min(I(µ),I(σt,N ). Moreover, a direct computation shows that
if Y ∼ σt,N , then the conjugate variable is (N/t)Y and the Fisher information is mN 3/t.
Next, to prove (5.5), let µt := µ ∗ σt,N . By basic properties of convolving positive functions
with the Gaussian, µt has a smooth density ρt. We claim that if 0 < δ < t, then
h(µt) − h(µδ) =
1
2N Z t
δ I(µs) ds =
1
2N Z t
δ ZMN (C)m
sa
∇ρs(x)2
ρs(x)
dx ds.
This will follow from integration by parts, but to give a complete justification, we first introduce
a smooth compactly supported "cutoff" function ψR : MN (C)m
sa → R such that 0 ≤ ψR ≤ 1
Moreover,
(5.5)
and
(5.6)
1
N 2 h(µ ∗ σt,N ) −
N 3I(µ ∗ σt,N ) ≤ min(cid:18) m
2Z t
N 3I(µ ∗ σs,N )(cid:19) ds +
0 (cid:18) m
2Z ∞
1
N 3I(µ ∗ σs,N ) ds
1
N 2 h(µ) =
1 + s −
1
1
1
0
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
35
and ψR(x) = 1 when x ≤ R and ψR(x) = 0 when x ≥ 2R. Because of scale-invariance, we
can arrange that k∇ψR(x)k2 ≤ C/R. Because ∂sρs = (1/2N )∆ρs, we have
d
dt(cid:20)−Z ψRρs log ρs(cid:21) = −
1
2N Z ψR · (∆ρs log ρs + ∆ρs)
2N Z ψR ∇ρs2
+
1
1
2N Z ∇ψR · ∇ρs · (1 + log ρs),
=
sa with respect to dx. This is equal to
ρs
where all the integrals are taken over MN (C)m
1
2N Z ψR∇ρs/ρs2 dµs −
1
2N Z (∇ψR · ∇ρs/ρs)(1 + log ρs) dµs.
Of course, by the monotone convergence theorem
R→+∞Z t
δ Z ψR∇ρs/ρs2 dµs ds =Z t
lim
δ I(µs) ds.
The other term is an error which can be estimated as follows: Note that µs = µ ∗ σs,N and
sa. Therefore,
sa and hence log ρs is uniformly bounded
that σs,N has a density that is bounded uniformly for s ∈ [δ, t] and x ∈ MN (C)m
ρs is uniformly bounded for s ∈ [δ, t] and x ∈ MN (C)m
above. To obtain a lower bound on log ρs, first note that there is a K > 0 such that
µ(x : x ≤ K) ≥ 1/2.
sa and y ≤ K, then x−y ≥ x−K and hence x−y2 ≤ x2−2Kx+K 2 ≥
Now if x ∈ MN (C)m
2x2 +2K 2, where the last inequality follows because 2Kx ≤ (1/2)x2 +2K 2 by the arithmetic
geometric mean inequality. Therefore, letting Z be the normalizing constant for σt,N , we have
1
1
dµ(y)
ZZ e−(N/2t)x−y2
ZZy≤KZ e−(N/2t)x−y2
ZZy≤K
1
e−(N/t)(x2+K 2) dµ(y)
dµ(y)
e−N K 2/t
2Z
e−(N/t)x2
,
ρs(x) =
≥
≥
≥
so that log ρs ≥ K ′ − x2 for some constant K ′. In particular, combining our upper and lower
bounds, there is a constant α such that for sufficiently large x, we have 1 + log ρs ≤ αx2.
Recall that ∇ψR(x) is supported when R ≤ x ≤ 2R and thus is bounded by 1/R ∼ 1/x.
Altogether we have ∇ψR(1 + log ρs) ≤ βx for some constant β when x is large enough.
Thus, our error term is bounded by
Z t
δ Z (∇ψR · Ξs)(1 + log ρs) dµs ds ≤ βZ t
δ Zx≥R x∇ρs(x)/ρs(x) dµs(x) ds
βZ t
δ Zx≥R
Z t
δ Z (x2 + ∇ρs(x)/ρs(x)2) dµs(x) ds =Z t
[(a + ms) + I(µs)] ds < +∞,
The right hand is the tail of the convergent integral
≤
1
2
δ
(x2 + ∇ρs(x)/ρs(x)2) dµs(x) ds.
36
DAVID JEKEL
and therefore it goes to zero as R → +∞ by the dominated convergence theorem. We can
earlier estimate that ρs is subquadratic for each s. The result is that
also apply the dominated convergence theorem to −R ψRρt log ρt and −R ψRρδ log ρδ given our
h(µt) − h(µδ) =
1
2N Z t
δ Z ∇ρs/ρs2 dµs ds =
1
2N Z t
δ I(µs) ds.
To complete the proof of (5.5), we must take δ ց 0. We can take the limit of the right hand
side by the monotone convergence theorem. As for the left hand side, Lemma 5.1 (3) implies
that lim supδց0 h(µδ) ≤ h(µ) because ρδ → ρ almost everywhere by Lebesgue differentiation
theory. On the other hand, h(µδ) ≥ h(µ) by Lemma 5.1 (4), hence h(µδ) → h(µ), so (5.5) is
proved.
To prove (5.6), we follow [40, Proposition 7.6]. First, suppose that h(µ) > −∞. Note that
1 + s −
1
h(µ) =
0 (cid:18) mN 2
2Z t
0 (cid:16) mN 2
If h(µ) > −∞, thenR 1
1+s − 1
0 (cid:18) mN 2
2Z t
1 + s −
t→+∞
lim
1
1
N I(µs)(cid:19) ds −
mN 2
2
log(1 + t) + h(µt).
N I(µs)(cid:17) ds is finite. In light of (5.4), the integral from 1 to
N I(µs)(cid:19) ds =
N I(µs)(cid:19) ds.
0 (cid:18) mN 2
2Z ∞
1 + s −
1
1
1
+∞ is also finite and by the dominated convergence theorem
It remains to understand the behavior of h(µt) − (mN 2/2) log(1 + t). By Lemma 5.1 (4) and
(2),
h(µt) ≥ h(σt,N ) =
mN 2
2
log
2πet
N
=
mN 2
2
log
2πe
N
+
mN 2
2
log t.
On the other hand, by Lemma 5.1 (2), sinceR x2 dµt(x) = N (a + tm), we have
2πe(a + t)
mN 2
mN 2
mN 2
=
2
log
2πe
N
+
2
log(a + t).
h(µt) ≤
2
log
N
As t → ∞, we have log(1 + t) − log(a + t) → 0 and log(1 + t) − log t → 0 and therefore
mN 2
log(1 + t) →
2
mN 2
2
log
2πe
N
=
mN 2
log 2πe −
2
2
log N.
mN 2
h(µt) −
Hence,
h(µ) =
1
0 (cid:18) mN 2
2Z ∞
1 + s −
1
N I(µs)(cid:19) ds +
mN 2
2
log 2πe −
mN 2
2
log N,
0 (cid:16) mN 2
R 1
1+s − 1
which is equivalent to the asserted formula (5.6). In the case where h(µ) = −∞, we also have
N I(µs)(cid:17) ds = −∞ by (5.5), but the integral from 1 to ∞ is finite as shown above.
So both sides of (5.6) are −∞.
5.4. Free Fisher Information. The starting point for the definition of free Fisher information
is the integration-by-parts formula (5.3). Indeed, if we formally apply this to a non-commutative
polynomial p and renormalize, we obtain
(cid:3)
(5.7)
Z τN(cid:18) 1
N
Ξj(x)p(x)(cid:19) dµ(x) =Z τN ⊗ τN (Dj p(x)) dµ(x),
(and this integration by parts is justified under sufficient assumptions of finite moments).
Voiculescu therefore made the following definitions:
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
37
Definition 5.7 ([40, §3]). Let X = (X1, . . . , Xm) be a tuple of self-adjoint random variables
in a tracial von Neumann algebra (M, τ ) and assume that M is generated by X as a von
Neumann algebra. We say that ξ = (ξ1, . . . , ξm) ∈ L2(M, τ )m is the (free) conjugate variable
of X if
(5.8)
τ (ξj p(X)) = τ ⊗ τ (Dj p(X))
for every non-commutative polynomial p. In this case, we say that X (or equivalently the law
j ). We also
of X) has finite free Fisher information and define Φ∗(X) := Φ∗(λX ) := Pj τ (ξ2
denote the conjugate variable ξ by J(X).
Definition 5.8 ([40, Definition 7.1]). The non-microstates free entropy of a non-commutative
law λ is
χ∗(λ) :=
1
0 (cid:18) m
2Z ∞
1 + t − Φ∗(λ ⊞ σt)(cid:19) +
1
2
log 2πe.
Now we are ready to state conditions under which the classical Fisher information of a
sequence of measures µN converges to the free Fisher information of the law λ. First, to
clarify the normalization, note that if dµN (x) = (1/ZN ) exp(−N 2VN (x)) dx, then the classical
conjugate variable is given by ΞN = N 2∇VN . The normalized conjugate variable used in (5.7)
is (1/N )ΞN = N∇VN = DVN . The corresponding normalized Fisher information is then
1
N
which is the same normalization as in Lemma 5.6.
Z kDVNk2
N (cid:12)(cid:12)(cid:12)(cid:12)
2 dµN =Z 1
ΞN(cid:12)(cid:12)(cid:12)(cid:12)
2
dµ =
1
N 3I(µN ),
sa → R be a potential with R exp(−N 2VN (x)) dx < +∞
Proposition 5.9. Let VN : MN (C)m
and let µN be the associated measure. Assume:
(A) The sequence µN concentrates around a non-commutative law λ.
(B) The sequence {DVN} is asymptotically approximable by trace polynomials.
(C) For some n ≥ 0 and a, b > 0 we have kDVNk2
(D) There exists R0 > 0 such that
mXj=1
2 ≤ a + bPm
j ) dµN (x) = 0.
N→∞Zkxk∞≥R01 +
j=1 τN (x2n
τN (x2n
j ).
lim
Then
(1) The law λ can be realized by self-adjoint random variables X = (X1, . . . , Xm) in a
tracial von Neumann algebra (M, τ ) with kXjk ≤ R0.
(2) There exists a sequence of trace polynomials f (k) ∈ (TrP1
m)m such that
lim
k→∞
lim sup
N→∞
sup
kxk∞≤R0kDVN (x) − f (k)(x)k2 = 0.
(3) If {f (k)} is any sequence as in (2), then {fk(X)} converges in L2(M, τ ) and the limit
(4) The law λ has finite free Fisher information and N −3I(µN ) → Φ∗(λ) as N → ∞.
is the conjugate variable J(X).
Proof. (1) This follows from the same argument as Proposition 5.4.
(2) This follows from the definition of asymptotic approximability by trace polynomials.
(3) Let {f (k)} be a sequence as in (2). Because µN concentrates around λ and because
µN ({x : kxk∞ ≤ R0}) → 1 as N → +∞ by (4), we have
λ[(f (j) − f (k))∗(f (j) − f (k))] = lim
τN [(f (j) − f (k))∗(f (j) − f (k))(x)] dµN (x).
N→∞Zkxk∞≤R0
38
DAVID JEKEL
For every ǫ > 0, if j and N are large enough, then supkxk∞≤R0kDVN (x) − f (j)(x)k2 < ǫ by our
assumption on f (j). In particular, if j and k are sufficiently large, then λ[(f (j) − f (k))∗(f (j) −
f (k))] < (2ǫ)2. This shows that {f (k)(X)} is Cauchy in L2(M, λ) since X has the law λ.
Let ξ = limk→∞ f (k)(X). We must show that ξ is the conjugate variable for X. Let ψ ∈
C∞
c (R) such that ψ(y) = y when y ≤ R0. For x ∈ MN (C)m
sa, let Ψ(x) = (ψ(x1), . . . , ψ(xm)).
By (5.7), because DVN (x) is the classical conjugate variable for X, we have for every non-
commutative polynomial p that
Z τN [DjVN (x) · p(Ψ(x))] dµN (x) =Z Dj[τN (p(Ψ(x)))] dµN (x).
It follows from our assumptions (C) and (D) that
N→∞Zkxk∞≥R0kDVN (x)k2
lim
2 dµN (x) = 0.
Because p(Ψ(x)) and Dj[τN (p(Ψ(x)))] are globally bounded in operator norm, the integral of
these quantities over kxk∞ ≥ R0 will vanish as N → ∞ and therefore
But since p(Ψ(x)) = p(x) on this region, we have
Zkxk∞≤R0
Zkxk∞≤R0
τN [DjVN (x)p(Ψ(x))] dµN (x) −Zkxk∞≤R0
τN [DjVN (x)p(x)] dµN (x) −Zkxk∞<R0
Dj[τ (p(Ψ(x)))] dµN (x) → 0
τN ⊗ τN [Djp(x)] dµN (x) → 0.
Now the second term converges to λ ⊗ λ[Dj p] = τ ⊗ τ [Dj p(X)] by our concentration assump-
tion (A). For the first term, we can replace DjVN (x) by f (k)
(x) with an error bounded by
supkxk∞≤R0kf (k)(x)−DVN (x)k2. Then we apply concentration to conclude thatR τN [f (k)
)∗p]. Overall,
λ[(f (k)
j
j
j
(x)∗p(x)] dµN (x) →
Taking k → ∞, we obtain τ [ξj p(X)] − τ ⊗ τ [Dj p(X)] = 0 as desired.
that
(4) We know from (3) that λ has finite Fisher information. Assumptions (C) and (D) imply
sup
kxk∞≤R0kf (k)(x) − DVN (x)k2.
j
(cid:12)(cid:12)(cid:12)λ[(f (k)
N→∞
)∗p] − λ ⊗ λ[Dj p](cid:12)(cid:12)(cid:12) ≤ lim sup
N 3I(µN ) −Zkxk∞≤R0kDVN (x)k2
1
2 dµN (x) → 0.
mateRkxk∞≤R0kf (k)k2
By similar arguments as before, we can approximate DVN by f (k) on kxk∞ ≤ R0, approxi-
2 dµN by λ((f (k))∗f (k)), and then approximate λ((f (k))∗f (k)) by τ (ξ∗ξ) =
Φ∗(λ), where the error terms vanish as N → ∞ and then k → ∞. This implies that
N −3I(µN ) → Φ∗(λ).
(cid:3)
6. Evolution of the Conjugate Variables
6.1. Motivation and Statement of the Equation. In the last section, we stated conditions
under which the classical entropy and Fisher information of µN converge to their free counter-
parts for the limiting non-commutative law λ. In order to prove that χ(λ) = χ∗(λ), we want to
take the limit in the integral formula (5.6), and therefore, we want N −3I(µN∗σt,N ) → Φ∗(λ⊞σt)
for all t > 0. In order to apply Proposition 5.9 to µN ∗ σt,N , we need to show that {DVN,t}N is
asymptotically approximable by trace polynomials, where VN,t is the potential corresponding
to µN ∗ σt,N .
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
39
By adding a constant to each VN if necessary, we may assume without loss of generality that
ZN = 1. We call that VN,t(x) is given by
(6.1)
exp(−N 2VN,t(x)) =Z exp(−N 2VN (x + y)) dσt,N (y).
Then exp(−N 2VN,t(x)) solves the normalized heat equation
(6.2)
∂t[exp(−N 2VN,t(x))] =
1
2N
∆[exp(−N 2VN,t(x))],
where (1/N )∆ = LN is the normalized Laplacian. However, we do not know how to show
that DVN (·, t) is asymptotically approximable by trace polynomials from a direct analysis of
the heat equation because of the dimension-dependent factor of N 2 in the exponent. What we
want is a dimension-independent and "hands-on" way of producing VN,t from VN .
As in §4, we will analyze the PDE which describes the evolution of the function VN,t. We
first derive the equation by rewriting the (6.2) in terms of VN,t rather than e−N 2VN,t . By the
chain rule,
and
∂t[exp(−N 2VN,t)] = −N 2∂tVN,t · exp(−N 2VN,t)
∆[exp(−N 2VN,t)] = [∆(−N 2VN,t) + ∇(−N 2VN,t)2] exp(−N 2VN,t)
= (−N 2∆VN,t + N 4∇VN,t2) exp(−N 2V ),
where ∆ and ∇ denote the classical (non-normalized) Laplacian and gradient, where MN (C)m
has been identified with RmN 2
using the coordinates in 2.1. Thus, our equation becomes
sa
−N 2∂tVN,t =
∂tVN,t =
1
2N
1
2N
(−N 2∆VN,t + N 4∇VN,t2)
∆VN,t − N∇VN,t2.
Recall that (1/N )∆ is the normalized Laplacian discussed in §3.2. The normalized gradi-
ent is DVN,t = N∇VN,t, and the normalized Euclidean norm is kxk2
j ) =
NPm
2 = Pm
j=1 τN (x2
j=1 Tr(x2
j ) = 1
1
N N∇VN,t2 =
1
N DVN,t2 = kDVN,tk2
2.
1
N x2. Then
N∇VN,t2 =
and therefore we obtain the following equation that is normalized in a dimension-independent
way
(6.3)
∂tVN,t =
1
2
LN VN,t −
1
2kDVN,tk2
2.
In the remainder of this section, we study a semigroup Rt acting on convex and semi-concave
functions on MN (C)m
sa such that VN,t = RtVN (here Rt depends implicitly on N ). In §6.2 -
§6.6, we construct Rt from scratch by iterating the heat semigroup and Hopf-Lax semigroup.
Next, in §6.7, we verify that RtVN solves (6.3) in the viscosity sense (for background, see [10]),
and deduce that RtVN must agree with the smooth solution VN,t defined by (6.1). Finally, in
(§6.8), we show that if {DVN} is asymptotically approximable by trace polynomials, then so is
{D(RtVN )}.
40
DAVID JEKEL
(6.4)
6.2. Strategy to Approximate Solutions. To construct the semigroup Rt that solves (6.3),
we view the equation as a hybrid between the heat equation ∂tu = (1/2N )∆u and the Hamilton-
Jacobi equation with quadratic potential ∂tu = −(1/2)kDuk2
2. The heat equation can be solved
by the heat semigroup
Ptu(x) :=Z u(x + y) dσt,N (y),
Qtu(x) := inf(cid:20)u(x + y) +
2(cid:21)
1
2tkyk2
as a special case of the Hopf-Lax formula (see [19, Chapter 3.3]).
while the Hamilton-Jacobi equation can be solved using the inf-convolution semigroup
(6.5)
Dupuis and Ustunel as the infimum of E[u(x + Bt +R t
In Dabrowski's approach, the solution to (6.3) was expressed through a formula of Bou´e,
2 ds] over a certain
class of stochastic processes Yt adapted to a standard Brownian motion Bt (see [13, Theorem
3.1]). This formula, roughly speaking, combines the Gaussian convolution and inf-convolution
operations by replacing the y in the definition of Qt by a stochastic process and allowing it to
evolve with Bt. Dabrowski then identifies the minimizing process Yt as a Brownian bridge [13,
Section 5] and analyzes it using a forward-backward SDE. Through the Picard iteration solving
the SDE, he shows that the solution is well-approximated by non-commutative functions.
0 Ys ds) + (1/2)R t
0kYsk2
We instead give a deterministic proof following the same strategy as in §4 that is motivated
by Trotter's formula, we define a semigroup Rtu at dyadic times t by alternating between
P2−ℓ and Q2−ℓ and then letting ℓ → ∞. We establish convergence through a telescoping series
argument after showing that PtQt− QtPt = o(t). Then we show that Rtu depends continuously
on t in order to extend its definition to all positive real t.
In contrast to §4, we must understand how the semigroups Pt, Qt, and Rt affect Du as well
as u, and we want D(Rtu) to be Lipschitz for all t. We therefore view these operators as acting
on spaces of the form
E(c, C) =(cid:26)u : MN (C)m
sa → R, u(x) −
1
2
ckxk2
2 is convex and u(x) −
1
2
Ckxk2
2 is concave(cid:27) ,
where 0 ≤ c ≤ C < +∞, where we suppress the dependence on m and N in the notation. These
spaces have the virtue that if u ∈ E(c, C), then kDukLip ≤ C automatically (see Proposition
2.13 (3)).
At every step of the proof, we include estimates both for u and for Du. In addition, con-
trolling the error propagation requires more work because Qt and Rt are not contractions with
respect to kDukL∞ .
The following theorem summarizes the results of the construction.
Here, for a measurable function u : MN (C)m
sa (for instance F = Du for some u : MN (C)m
norm. If F : MN (C)m
kFkL∞ = supx∈MN (C)m
in both the domain and the target space.
sa → R, the notation kukL∞ is the standard L∞
sa → MN (C)m
sa → R, then
sakF (x)k2; similarly, kFkLip is the Lipschitz norm of F when using k·k2
Note that kFk2 does not denote the L2 norm of F with respect to any measure, but rather
j ))1/2, which is a function of x. We denote N = {1, 2, 3, . . .} and N0 = {0, 1, 2, . . .}.
j=1 τ (F 2
(Pm
2 =Sn≥0 2−nN0 the nonnegative dyadic rationals. Moreover, we assume
Theorem 6.1. There exists a semigroup of nonlinear operators Rt :SC>0 E(0, C) →SC>0 E(0, C)
We also denote by Q+
throughout the section that 0 ≤ c ≤ C < +∞.
with the following properties:
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
41
(1) Change in Convexity:
ct)−1, C(t + Ct)−1).
If u ∈ E(c, C) where 0 ≤ c ≤ C, then Rtu ∈ E(c(1 +
(2) Approximation by Iteration: For ℓ ∈ Z and t ∈ 2−ℓN0, denote Rt,ℓu = (P2−ℓQ2−ℓ)2ℓtu.
Suppose t ∈ Q+
(a) If 2−ℓ−1C ≤ 1, then
2 and u ∈ E(0, C).
C2mt
1 + Ct
Rtu − Rt,ℓu ≤(cid:18) 3
2
+ log(1 + Ct)(m + Cm + kDuk2
(3) Continuity in Time: Suppose s ≤ t ∈ R+ and u ∈ E(0, C).
2)(cid:19) 2−ℓ.
(b) kD(Rt,ℓu) − D(Rtu)kL∞ ≤ [t/2 + C(t/2)2]C2m1/2(2 · 2−ℓ/2 + 2−3ℓ/2C).
(a) Rtu ≤ Rsu + m
(b) Rtu ≥ Rsu − 1
(c) If C(t− s) ≤ 1, then kD(Rtu)− D(Rsu)k2 ≤ 5Cm1/221/2(t− s)1/2 + C(t− s)kDuk2.
(4) Error Estimates: Let t ∈ R+ and u, v ∈ E(0, C). Then
(a) kD(Rtu) − D(Rtv)kL∞ ≤ (1 + Ct)kDu − DvkL∞.
(b) If u ≤ v + a + bkDvk2
2 where a ∈ R and b ≥ 0, then
2 [log(1 + Ct) − log(1 + Cs)].
2 (t − s)(Cm + kDuk2
2).
Rtu ≤ Rtv + a + b
(c) We have
C2mt
1 + Ct
+ bkD(Rtv)k2
2.
kD(Rtu)k2
2 ≤
C2mt
1 + Ct
+ kDuk2
2.
Remark 6.2. Knowing that exp(−N 2(Rtu)) = Pt exp(−N 2u), one can deduce (1) from the
Braskamp-Lieb and Holder inequalities, as in [7, Theorem 4.3]. But the proof of (1) given here
is independent of [7].
We also point out that the ideas of semigroups and discrete-time approximation schemes
have been employed to study Hamilton-Jacobi equations in Hilbert space (e.g. by [3]).
6.3. The Hopf-Lax Semigroup, the Heat Semigroup, and Convexity. We remind the
reader of our standing assumption that 0 ≤ c ≤ C.
Lemma 6.3. Suppose u ∈ E(c, C). Then
(1) Ptu ∈ E(c, C).
(2) kD(Ptu) − DukL∞ ≤ Cm1/2t1/2.
Proof. (1) follows because E(c, C) is closed under translation and averaging, hence convolution
by a probability measure.
(2) We know that Du is C-Lipschitz and thus
kD(Ptu)(x) − Du(x)k2 ≤Z kDu(x + y) − Du(x)k2 dσt,N (y)
≤Z Ckyk2 dσt,N (y)
≤ Cm1/2t1/2.
(cid:3)
The following lemma gives basic properties of Qt from the PDE literature; see for instance
[18, p. 309-311], [25], [10, Lemma A.5], [19, Section 3.3.2]. For completeness and convenience,
we include a proof of all the facts we will use.
Lemma 6.4.
(1) If u, v : MN (C)m
sa → R and u ≤ v, then Ptu ≤ Ptv and Qtu ≤ Qtv.
42
DAVID JEKEL
(2) Suppose that v(x) = a+hp, xi2
semi-definite linear map MN (C)m
1
2hAx, xi2 where a ∈ R, p ∈ MN (C)m
sa, and A is a positive
sa. Then
sa → MN (C)m
1
2hAx, xi2,
Ptv(x) = a +
Qtv(x) = a −
t
2N 2 Tr(A) + hp, xi +
t
2kpk2 + hp, xi +
1
2hA(1 + tA)−1(x − tp), x − tpi.
sa to RmN 2
Remark 6.5. The meaning of Tr(A) In the above formula is as follows. Using the identification
of MN (C)m
given by §2.1, we can express A as an mN 2 × mN 2 and compute its trace
in this way. Alternatively, since A is a linear transformation of the real inner product space
MN (C)m
sa. Because the trace
is similarity-invariant, this answer is independent of the choice of basis (and also independent
of the choice of normalization for the inner product). Note that the trace of the identity is
mN 2, which makes the normalization in the above formula dimension-independent.
sa, we may compute Tr(A) using an orthonormal basis of MN (C)m
Proof. (1) is immediate to check from the definition. We leave the first formula of (2) as an
2tky − xk2
exercise. To prove the last formula, fix t > 0 and x ∈ MN (C)m
2
is a uniformly convex function of y and therefore it has a unique minimizer. The minimizer y
must be a critical point and hence
sa and note that u(y) + 1
0 = Du(y) +
1
t
(y − x) = p + Ay +
1
t
(y − x).
Thus, (1 + tA)y = x − tp and y − x = −t(p + Ay). Thus,
Qtu(x) = u(y) +
1
2tky − xk2
2
1
2hp + Ay, y − xi
1
= a + hp, yi +
2hAy, yi −
1
1
= a +
2hp, yi +
2hp + Ay, xi
1
1
2hAy, xi
2hp, y − xi +
= a + hp, xi +
1
t
= a + hp, xi −
2hp, p + Ayi +
2hAy, xi
1
= a −
2hAy, x − tpi
1
2hA(1 + tA)−1(x − tp), x − tpi
= a −
t
2kpk2 + hp, xi +
t
2kpk2 + hp, xi +
(cid:3)
Lemma 6.6. Let u ∈ E(c, C) and t ∈ R+,
sa, the infimum Qtu(x0) = inf y[u(y) + t
a unique point y0 satisfying y0 = x0 − tDu(y0).
(1) The operators {Qt}t≥0 form a semigroup, that is, QsQtu = Qs+tu.
(2) For each x0 ∈ MN (C)m
(3) If x0 ∈ MN (C)m
(4) We have Qtu ∈ E(c(1 + ct)−1, C(1 + Ct)−1).
(5) kD(Qtu)(x0)k2 = kDu(y0)k2 ≤ (1 + ct)−1kDu(x0)k2.
sa and y0 is the minimizer from (1), then D(Qtu)(x0) = Du(y0).
2ky − x0k2
2 is achieved at
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
43
Proof. (1) By definition
QsQtu(x) = inf
y
= inf
y
inf
= inf
2 +
[Qtu(y) +
1
2skx − yk2
2]
z (cid:20)u(z) +
1
2tky − zk2
y (cid:20) 1
z (cid:20)u(z) + inf
2tky − zk2
2 + (1/2s)kx − yk2
2, then by the previously lemma, we have
1
2
2(cid:21)
1
2skx − yk2
2(cid:21)(cid:21) .
1
2skx − yk2
2(s + t)kxk2
2.
1 + t−1skxk2
2 =
2 +
t−1
1
But note that inf y[(1/2t)ky − zk2
(1/2t)kx − zk2
2. If g(x) = (1/2t)kxk2
Qsg(x) =
2] is by definition Qsf (z), where f (x) =
QsQtu(x) = inf
z (cid:20)u(z) +
Since Qs is clearly translation-invariant, Qsf (x) = 1/2(s + t) · kx − zk2
2(cid:21) = Qs+tu(x).
2i is in E(c + 1/2t, C + 1/2t) and
(2) Fix x0. Note that the function y 7→hu(y) + 1
hence it achieves a global minimum at the unique critical point. Thus, the infimum is achieved
at the point y0 satisfying Du(y0) = (1/t)(y0 − x0), or in other words y0 = x0 − tDu(y0).
all x that
(3) and (4) Let x0 and y0 be as above. Let p = Du(y0). Because u ∈ E(c, C), we have for
2(s + t)kx − zk2
2tky − x0k2
2. Therefore,
1
u(y0) + hp, x − y0i2 +
c
2hx − y0i ≤ u(x) ≤ u(y0) + hp, x − y0i2 +
C
2 kx − y0k2
2
Let v(y) and v(y) be the functions on the left and right hand sides. Then by Lemma 6.4 (1),
we have Qtv ≤ Qtu ≤ Qtv. To compute Qtv, we apply Lemma 6.4 (2) with A = cI and with a
change of coordinates to translate y0 to the origin, and we obtain
Qtv(x) = u(y0) −
2 + hp, x − y0i +
Since y0 = x0 + tp and p = (y0 − x0)/t, this becomes
t
2kpk2
1
2
c(1 + ct)−1kx − y0 − tpk2.
Qtv(x) = u(y0) −
= u(y0) +
t
2kpk2
1
2tky0 − x0k2
2 + tkpk2 + hp, x − x0i +
1
2 + hp, x − x0i +
2
1
c(1 + ct)−1kx − x0k2
2
c(1 + ct)−1kx − x0k2
2
2
= Qtu(x0) + hp, x − x0i +
1
2
c(1 + ct)−1kx − x0k2
2.
The analogous computation holds for Qtv as well. Thus, we have
1
2
c(1+ct)−1kx−x0k2
Qtu(x0)+hp, x−x0i+
This inequality implies that D(Qtu)(x0) = p = Du(y0). Since the above inequality holds for
every x0, we see that Qtu ∈ E(c(1 + ct)−1, C(1 + Ct)−1).
2 ≤ Qtu(x) ≤ Qtu(x0)+hp, x−x0i+
C(1+Ct)−1kx−x0k2
2.
(5) Let x0, y0, and p be as above. Then we have
1
2
hDu(y0) − Du(x0), y0 − x0i2 ≥ cky0 − x0k2
2.
But recall that y0 − x0 = −tDu(y0) and hence
−thDu(y0) − Du(x0), Du(y0)i ≥ ct2kDu(y0)k2
2.
44
DAVID JEKEL
Rearranging produces
(1 + ct)kDu(y0)k2
2 ≤ hDu(x0), Du(y0)i2 ≤ kDu(x0)k2kDu(y0)k2,
and hence (1 + ct)kDu(y0)k2 ≤ kDu(x0)k2 as desired.
Corollary 6.7. Let u ∈ E(c, C) and s, t ≥ 0.
(cid:3)
(cid:3)
(1) For each x, the gradient D(Qtu)(x) is the unique vector p satisfying p = Du(x − tp).
(2) We have Qtu(x) = u(x − tD(Qtu)(x)) + t
(3) u(x) − t
2 (1 + Ct)kD(Qtu)(x)k2 ≤ Qtu(x) ≤ u(x) − t
2 (1 + ct)kD(Qtu)(x)k2
2.
2kD(Qtu)(x)k2
2.
Proof. (1) and (2) follow from Lemma 6.6 (2) and (3).
To prove (3), fix x and let y = x − tD(Qtu)(x). By Proposition 2.13 (2),
2 ≤ u(x) ≤ u(y) + hDu(y), x − yi +
u(y) + hDu(y), x − yi2 +
c
2kx − yk2
C
2 kx − yk2
2.
Hence,
u(x) − hDu(y), x − yi2 −
C
2 kx − yk2
2 ≤ u(y) ≤ u(x) − hDu(y), x − yi2 −
c
2kx − yk2
2.
But from the previous lemma, we know that Du(y) = D(Qtu)(x) and x − y = tD(Qtu)(x), so
that
u(x) − tkD(Qtu)(x)k2
Finally, we substitute Qtu(x) = u(y) + (t/2)kD(Qtu)(x)k2
6.4. Estimates for Error Propagation. To prepare for our iteration, we first prove some
estimates that will help control the propagation of errors.
2 ≤ u(y) ≤ u(x) − tkD(Qtu)(x)k2
2 and obtain (3).
c
2kD(Qtu)(x)k2
2.
t2kD(Qtu)(x)k2
2 −
C
2
2 −
Lemma 6.8. If u, v ∈ E(c, C), then we have
(1) kD(Ptu) − D(Ptv)kL∞ ≤ kDu − DvkL∞.
(2) kD(Qtu) − D(Qtv)kL∞ ≤ (1 + Ct)kDu − DvkL∞.
Proof. The first inequality follows because D(Ptu)− D(Ptv) is the convolution of Du− Dv with
the Gaussian density. To prove the second inequality, note that
kD(Qtu)(x) − D(Qtv)(x)k2 = kDu(x − tD(Qtu)(x)) − Dv(x − tD(Qtv)(x)k2
≤ kDu(x − tD(Qtv)(x)) − Dv(x − tD(Qtv)(x)k2
+ kDu(x − tD(Qtu)(x)) − Du(x − tD(Qtv)(x)k2
≤ kDu − DvkL∞ + CtkD(Qtu)(x) − D(Qtv)(x)k2,
where the last inequality follows because Du is C-Lipschitz. This implies that for t < 1/C,
kD(Qtu) − D(Qtv)kL∞ ≤ (1 − Ct)−1kDu − DvkL∞.
Now we improve the estimate using the semigroup property of Qt. Fix a positive integer k and
for j = 1, . . . , k, let tj = tj/k, and let Cj = C(1 − Ctj)−1. Then Qtj u and Qtj are in E(0, Cj ).
Thus, we have
kD(Qtj+1 u) − D(Qtj+1 v)kL∞ ≤ (1 − Cjt/k)−1kD(Qtj u) − D(Qtj v)kL∞,
and hence
kD(Qtu) − D(Qtv)kL∞ ≤ kDu − DvkL∞
k−1Yj=0
1
1 − Cj t/k
.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
45
log
k−1Yj=0
1
1 − Cjt/k
Now
Hence,
=
=
− log(1 − Cjt/k)
k−1Xj=0
k−1Xj=0(cid:0)Cjt/k + O(1/k2)(cid:1)
k−1Xj=0
=Z t
1 + Ctj
1 + Cs
= log(1 + Ct) + O(1/k).
=
C
C
ds + O(1/k)
0
(tj+1 − tj) + O(1/k)
kD(Qtu) − D(Qtv)kL∞ ≤ (1 + Ct + O(1/k))kDu − DvkL∞ ,
(cid:3)
sa → R be convex and let v ∈ E(c, C) and u ≤ v + a +
and the proof is completed by taking k → ∞.
Lemma 6.9. Suppose that u : MN (C)m
bkDvk2
2 for some a ∈ R and b ≥ 0.
(1) Ptu ≤ Ptv + a + bC2mt + bkD(Ptv)k2
2.
(2) Qtu ≤ Qtv + a + bkD(Qtv)k2
2.
Proof. (1) Using monotonicity and linearity of Pt, we have
So it suffices to show that
Ptu ≤ Ptv + a + bZ kDv(x + y)k2
2 dσt,N (y) − kD(Ptv)(x)k2
Z kDv(x + y)k2
2 dσ(y).
2 ≤ bC2mt.
In probabilistic terms, the left hand side is the variance of the random variable Dv(x + Y )
where Y ∼ σt,N . Since the variance is translation-invariant, this is the same as the variance of
Dv(x + Y ) − Dv(x), and this is bounded above by the second moment
2 = C2mt.
EkDv(x + Y ) − Dv(x)k2
2 ≤ C2 · EkY k2
(2) Note that
Qtu(x) = inf
y
[u(y) +
1
2tky − xk2
2]
≤ u(x − tD(Qtv)(x)) +
≤ v(x − tD(Qtv)(x)) +
= Qtv(x) + a + bkD(Qtv)(x)k2
2,
t
2kD(Qtv)(x)k2
t
2kD(Qtv)(x)k2
2
2 + a + bkDv(x − tD(Qtv)(x))k2
2
where the last equality follows from Corollary 6.7 (1) and (2).
(cid:3)
Lemma 6.10. Let u ∈ E(0, C). Then
2 ≤ kDuk2
2.
2 ≤ C2mt + kDuk2
2.
(1) kD(Qtu)k2
(2) kD(Ptu)k2
46
DAVID JEKEL
Proof. The first claim follows from Lemma 6.6 (5). To prove the second claim, note that by
Minkowski's inequality,
kD(Ptu)(x)k2
where the last inequality was shown in the proof of Lemma 6.9 (1).
2 =(cid:13)(cid:13)(cid:13)(cid:13)Z Du(x + y) dσt,N (y)(cid:13)(cid:13)(cid:13)(cid:13)
2 ≤Z kDu(x + y)k2
2
2 dσt,N (y) ≤ C2mt + kDu(x)k2
2,
(cid:3)
Next, we iterate the previous inequalities to obtain our main lemma on error propagation.
Lemma 6.11. Let t1,. . . ,tn > 0 and write
t∗ = t1 + ··· + tn
R = Ptn Qtn . . . Pt1 Qt1
Let u, v ∈ E(c, C).
(1) Ru, Rv ∈ E(c(1 + ct∗)−1, C(1 + Ct∗)−1).
(2) kD(Ru) − D(Rv)kL∞ ≤ (1 + Ct∗)kDu − DvkL∞.
(3) If u ≤ v + a + bkDvk2
2 with a ∈ R and b ≥ 0, then we have
C2mt∗
1 + Ct∗ + bkD(Rv)k2
Ru ≤ Rv + a + b
2.
In particular, u ≤ v implies Ru ≤ Rv.
(4) We have
kD(Ru)k2
C2mt∗
1 + Ct∗ + kDuk2
2 ≤
≤ Cm + kDuk2
2.
2
Proof. (1) Let u ∈ E(c, C). Let t∗
induction that uj ∈ E(c(1 + ct∗
induction step, note that
c(1 + ct∗
j )−1
1 + [c(1 + ct∗
j )−1]tj+1
=
c
j ) + ctj+1
(1 + ct∗
= c(1 + ct∗
j+1)−1
j = t1 + ··· + tj and uj = Psj Qtj . . . Ps1 Qt1 u. We show by
j )−1). The base case j = 0 is trivial. For the
j )−1, C(1 + Ct∗
and the same holds for C. Hence, by Lemma 6.6 (4), if uj ∈ E(c(1 + ct∗
then Qtj+1uj ∈ E(c(1 + ct∗
Ptj+1 Qtj+1 uj ∈ E(c(1 + ct∗
induction that kDuj − DvjkL∞ ≤ (1 + Ct∗
the induction step, recall that uj, vj ∈ E(c(1 + ct∗
and the induction hypothesis,
j )−1),
j+1)−1). By Lemma 6.3, this implies that uj+1 =
j+1)−1). The same argument of course applies to v.
j and uj be as in the proof of (1) and define vj similarly to uj. We show by
j )kDu − DvkL∞. The base case j = 0 is trivial. For
j )−1) and hence by Lemma 6.8
j+1)−1, C(1 + Ct∗
j+1)−1, C(1 + Ct∗
j )−1, C(1 + Ct∗
j )−1, C(1 + Ct∗
(2) Let t∗
kD(Qtj+1 uj) − D(Qtj+1 vj)kL∞ ≤ (1 + C(1 + Ct∗
≤ (1 + C(1 + Ct∗
= (1 + Ct∗
j )−1tj+1)kDuj − DvjkL∞
j )−1tj+1)(1 + Ct∗
j )kDu − DvkL∞
j+1)kDu − DvkL∞.
Then by Lemma 6.8 again, since uj+1 = Ptj+1 Qtj+1 uj and vj+1 = Ptj+1 Qtj+1vj , we have
kDuj+1 − Dvj+1kL∞ ≤ (1 + Ct∗
j+1)kDu − DvkL∞.
(3) First, we show by induction on j that
uj ≤ vj + a + b
C2mti
(1 + Ct∗
i )2 + bkDvjk2
2.
jXi=1
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
47
The base case j = 0 is trivial. If the claim holds for uj and vj, then it also holds for Qtj+1uj
and Qtj+1 vj by Lemma 6.9 (2). Then we apply Lemma 6.9 (1) together with the fact that
Qtj+1 uj and Qtj+1 vj are in E(c(1 + ct∗
j+1)−1) to conclude that
j+1)−1, C(t + Ct∗
uj+1 ≤ vj+1 + a + b
C2mti
(1 + Ct∗
i )2 + bkDvj+1k2
2.
j+1Xi=1
i )2 is the lower
Riemann sum for the function C2m/(1+Ct)2 on the interval [0, t∗] with respect to the partition
{0, t∗
This completes the induction. Finally, we observe that Pn
(1 + Ct)2 dt = Cm(cid:18)1 −
1 + Ct∗(cid:19) =
i )2 ≤Z t∗
i=1 C2mti/(1 + Ct∗
C2mt∗
1 + Ct∗ .
C2mti
(1 + Ct∗
1, . . . , t∗
n}. Thus,
1
C2m
0
nXi=1
This shows the main claim of (3), and the claim that u ≤ v implies Ru ≤ Rv is the special case
when a = 0 and b = 0.
(4) By Lemma 6.10, we have kD(Qtj+1 uj)k2
kDuj+1k2
2 ≤
C2mtj+1
1 + Ct∗
j+1
2 and
2 ≤ kDujk2
2 ≤
+ kD(Qtj+1 uj)k2
C2mtj+1
1 + Ct∗
j+1
+ kDujk2
2.
2 follows because C2mt/(1 + Ct) ≤ Cm.
We sum from j = 0, . . . , n − 1 and obtain the same lower Riemann sum as in the proof of (3).
The final estimate Cm + kDuk2
6.5. Iterative Construction of Rt for Dyadic t. We are now ready to carry out the Trotter's
formula strategy and construct the semigroup for dyadic values of t. The next step is to show
that the operators Pt and Qt almost commute when t is small.
Lemma 6.12. Let u ∈ E(c, C) and t > 0.
(cid:3)
Proof. (1) Note that
(1) kD(QtPtu) − D(PtQtu)kL∞ ≤ C2m1/2(2 + Ct)t3/2.
(2) PtQtu ≤ QtPtu.
(3) If Ct ≤ 1, then QtPtu ≤ PtQtu + 2C2mt2 + 2Ct2kD(PtQtu)k2
2.
D(QtPtu)(x) = D(Ptu)(x − tD(QtPtu)(x)) =Z Du(x + y − tD(QtPtu)(x)) dσt,n(y).
D(PtQtu)(x) =Z D(Qtu)(x + y) dσt,n(y) =Z Du(x + y − tD(Qtu)(x + y)) dσt,n(y).
kD(QtPtu)(x) − D(PtQtu)(x)k2 ≤ CtZ kD(Qtu)(x + y) − D(QtPtu)(x)k2 dσt,n(y).
Because Du is C Lipschitz, we have
On the other hand,
We can estimate the integrand by
kD(Qtu)(x + y) − D(Qtu)(x)k2 + kD(Qtu)(x) − D(QtPtu)(x)k2.
Integrating the first term and using the fact that D(Qtu) is C-Lipschitz (since u ∈ E(0, C) by
Lemma 6.6 (4)), we have
Z kD(Qtu)(x + y) − D(Qtu)(x)k2 dσt,n(y) ≤ CZ kyk2 dσt,n ≤ Cm1/2t1/2
48
DAVID JEKEL
Meanwhile, the second term is independent of y and thus it is unchanged when we integrate it
against the probability measure σt,N , and this quantity can be estimated using Lemma 6.8 (2)
and Lemma 6.3 (2) as
kD(Qtu)(x) − D(QtPtu)(x)k2 ≤ (1 + Ct)kDu − D(Ptu)kL∞ ≤ (1 + Ct)Cm1/2t1/2,
Altogether, we obtain
kD(QtPtu)(x) − D(PtQtu)(x)k2 ≤ C2m1/2(2 + Ct)t3/2.
(cid:3)
(2) The idea is that the average of the infimum is less than or equal to the infimum of the
average. More precisely,
[u(y) +
PtQtu(x) =Z inf
=Z inf
y Z [u(y − z) +
[u(y − z) +
≤ inf
1
2tk(x + z) − yk2
1
2tkx − yk2
1
2tkx − yk2
y
y
2] dσt,N (z)
2] dσt,N (z)
2] dσt,N (z)
= inf
y
[Ptu(y) +
= QtPtu(x).
1
2tkx − yk2
2]
(3) To prove the other inequality, note that by Corollary 6.7 (3),
(6.6)
Also by Corollary 6.7 (3),
QtPtu ≤ Ptu −
t
2kD(QtPtu)k2
2.
u ≤ Qtu +
t
2
(1 + Ct)kD(Qtu)k2
2.
Hence, by Lemma 6.9, since Qtu ∈ E(c(1 + ct)−1, C(1 + Ct)−1) ⊆ E(0, C), we have
(6.7)
(1 + Ct) +
C2mt2
t
2
(1 + Ct)kD(PtQt)k2
2.
Ptu ≤ PtQtu +
Plugging (6.7) into (6.6), we obtain
2
(6.8)
QtPtu ≤ PtQtu +
By using part (1), we have
kD(QtPtu)k2
C2mt2
2
(1 + Ct) −
t
2kD(QtPtu)k2
2 +
t
2
(1 + Ct)kD(PtQt)k2
2.
2 ≥ [kD(PtQt)uk2 − C2m1/2t3/2(2 + Ct)]2
≥ kD(PtQtu)k2
≥ kD(PtQtu)k2
2 − 2C2m1/2t3/2(2 + Ct)kD(PtQtu)k2
2 − (2 + Ct)[C3mt2 + CtkD(PtQtu)k2
2]
where the last step follows from the arithmetic-geometric mean inequality
2Cm1/2t1/2kD(PtQtu)k2 ≤ C2mt + kD(PtQtu)k2
2.
So substituting our estimate for kD(QtPtu)k2
by
2 into (6.8), we see that PtQtu− QtPtu is bounded
C2mt2
2
+
t
2
(2 + Ct)[C3mt2 + CtkD(PtQtu)k2
2] −
t
2kD(PtQtu)k2
2 +
t
2
(1 + Ct)kD(PtQt)k2
2
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
49
Now we cancel the first-order terms (t/2)kD(PtQtu)k2
assumption that Ct ≤ 1. Thus, this is bounded by
2 and we estimate 2 + Ct by 3 using our
C2mt2
2
+
3
2
t[C3mt2 + CtkD(PtQtu)k2
2] +
1
2
Ct2kD(PtQtu)k2
2
≤2C2mt2 + 2Ct2kD(PtQtu)k2
2,
where we have again used our assumption Ct ≤ 1 to cancel Ct from the term t · C3mt2.
Finally, we can construct the semigroup Rt for dyadic values of t. As in the statement of
Theorem 6.1, we define Rt,ℓu = (P2−ℓQ2−ℓ)2ℓtu whenever ℓ ∈ Z and t ∈ 2−ℓN0.
Lemma 6.13. Let C ≥ 0. For t ∈ Q+
Moreover, we have for t ∈ 2−ℓN0 that
2 and u ∈ E(0, C), the limit Rtu = limℓ→∞ Rt,ℓu exists.
(1) Rt,ℓu ≤ Rtu.
(2) If C/2ℓ+1 ≤ 1, then
Rtu ≤ Rt,ℓu +(cid:18) 3
2)(cid:19) 2−ℓ.
(3) kD(Rt,ℓu) − D(Rtu)kL∞ ≤ [t/2 + C(t/2)2]C2m1/2(2 · 2−ℓ/2 + 2−3ℓ/2C).
+ log(1 + Ct)(m + Cm + kDuk2
C2mt
1 + Ct
2
Proof. First, we prove some intermediate claims relating Rt,ℓu and Rt,ℓ+1u. To this end, we
fix ℓ ∈ Z and t = 2−ℓn for some n ∈ N0. Let δ = 2−ℓ−1. For j = 0, . . . , n, define
uj = (PδQδ)2(n−j)(P2δQ2δ)j u.
and note that
Let
Then for j = 1, . . . , n, we have
u0 = Rt,ℓ+1u
un = Rt,ℓu.
vj = Qδ(P2δQ2δ)ju.
uj−1 = [(PδQδ)2(n−j)Pδ](QδPδvj−1)
uj = [(PδQδ)2(n−j)Pδ](PδQδvj−1).
We also define for k = 1, . . . , 2n,
Ck = C(1 + Ckδ)−1,
ck = c(1 + ckδ)−1.
Thus, by Lemma 6.11 (1) and Lemma 6.6 (4), we have vj−1 ∈ E(c2j−1, C2j−1).
First, we claim that
(6.9)
Now by Lemma 6.12 (2), we have PδQδvj−1 ≤ QδPδvj−1. Hence, by monotonicity of Pt and
Qt (Lemma 6.11 (3)), we have uj ≤ uj−1. Hence, Rt,ℓu = un ≤ u0 = Rt,ℓ+1u, proving (6.9).
Rt,ℓu ≤ Rt,ℓ+1u.
For an inequality in the other direction, we claim that
(6.10)
By Lemma 6.12 (3), since vj−1 ∈ E(c2j−1, C2j−1), we obtain
Rt,ℓ+1u ≤ Rt,ℓu +(cid:18) 3
QδPδvj−1 ≤ PδQδvj−1 + 2C2
Cm + log(1 + Ct)(m + Cm + kDuk2
2)(cid:19) 2−ℓ−1
2j−1mδ2 + 2C2j−1δ2kD(PδQδvj−1)k2
2
2
Thus, by Lemma 6.9 (1), since QδPδvj−1 and PδQδvj−1 are in E(c2j , C2j), we have
PδQδPδvj−1 ≤ P2δQδvj−1 + 2C2j−1mδ2 + 2C2j−1δ2(cid:16)C2
2(cid:17)
2j mδ + kD(P2δQδvj−1)k2
50
DAVID JEKEL
Recalling that uj−1 and uj are obtained by applying (PδQδ)2(n−j) to PδQδPδvj−1 and P2δQδvj−1,
and that PδQδPδvj−1 and P2δQδvj−1 are in E(c2j, C2j ), we may apply Lemma 6.11 (3) to con-
clude that
uj−1 ≤ uj + 2C2j−1mδ2 + 2C2j−1δ2 C2
2j mδ +
C2
2jm(n − j)δ
1 + 2C2j(n − j)δ
2!
+ kDujk2
By our assumption, C2j δ ≤ Cδ ≤ 1, and thus
Therefore,
C2
2j mδ +
C2
2jm(n − j)δ
1 + 2C2j(n − j)δ ≤ C2jmδ +
uj−1 − uj ≤ 2C2j−1mδ2 + 3C2
1
2
C2jmδ =
3
2
C2j mδ ≤
3
2
C2j−1mδ.
2j−1mδ2 + 2C2j−1δ2kDujk2
2.
By Lemma 6.11 (4), we have kDujk2 ≤ Cm + kDuk2
2, and hence
Rt,ℓ+1u − Rt,ℓu ≤ 3mδ2
Therefore, summing from j = 1, . . . , n, we have
2j−1 + 2δ2(m + Cm + kDuk2
C2
uj−1 − uj ≤ 3C2
nXj=1
mδ
nXj=1
2j−1mδ2 + 2C2j−1δ2(m + Cm + kDuk2
2).
C2j−1
2)
nXj=1
2j−1(2δ) + δ(m + Cm + kDuk2
C2j−1(2δ)
2)
nXj=1
Recalling the definition of C2j−1, two times the first sum is Pn
j=1 C2(2δ)/(1 + C(2j − 1)δ)2,
which is the Riemann sum for the function φ(s) = C2/(1 + Cs)2 on the interval [0, t] = [0, 2nδ],
where we use a partition into subintervals of length 2δ and evaluate φ at the midpoint of each
interval. Because φ is convex, the value of φ at the midpoint is less than or equal to the average
value over the subinterval and therefore
C2
3
2
=
C2
(1 + Cs)2 ds =
C2t
1 + Ct
.
By similar reasoning,
Therefore,
0
C2(2δ)
nXj=1
(1 + C(2j − 1)δ)2 ≤Z t
nXj=1
nXj=1
Rt,ℓ+1u − Rt,ℓu ≤(cid:18) 3
C2mt
1 + Ct
C2j−1(2δ) =
Cδ
2
(1 + C(2j − 1)δ) ≤Z t
0
C
1 + Cs
ds = log(1 + Ct).
+ log(1 + Ct)(m + Cm + kDuk2
2)(cid:19) δ,
which proves (6.10).
Together, (6.9) and (6.10) show that
Rt,ℓ+1u − Rt,ℓu ≤(cid:18) 3
2
C2mt
1 + Ct
+ log(1 + Ct)(m + Cm + kDuk2
2)(cid:19) 2−ℓ−1.
Because the right hand side is summable in ℓ, we see that the sequence {Rt,ℓu(x)}ℓ∈N is Cauchy
and hence converges. Thus, limℓ→∞ Rt,ℓu exists. Also, by (6.9) the convergence is monotone
and thus Rt,ℓu ≤ Rtu, establishing (1). On the other hand, we obtain (2) by summing up the
estimate (6.10) from ℓ to ∞ using the geometric series formula.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
51
It remains to prove (3). We first claim that
kD(Rt,ℓ+1u) − D(Rt,ℓu)kL∞ ≤ [t/2 + C(t/2)2]C2m1/2(2 + 2−(ℓ+1)C)2−(ℓ+1)/2
(6.11)
By Lemma 6.11 (3), we have QδPδvj−1 and PδQδvj−1 are in E(c(1 + 2cjδ)−1, C(1 + 2Cjδ)−1),
hence in E(0, C). Therefore, by Lemma 6.11 (??) and Lemma 6.12 (1), we have
kDuj − Duj−1kL∞ ≤ [1 + 2C(n − j)δ]kD(QδPδvj) − D(PδQδvj)kL∞
≤ [1 + 2C(n − j)δ]C2m1/2(2 + Cδ)δ3/2.
Therefore,
kD(Rt,ℓ+1u) − D(Rt,ℓu)kL∞ ≤
kDuj − Duj−1kL∞
[1 + 2C(n − j)δ]C2m1/2(2 + Cδ)δ3/2
nXj=1
nXj=1
≤
= [n + Cn(n − 1)δ]C2m1/2(2 + Cδ)δ3/2
≤ [t/2 + C(t/2)2]C2m1/2(2 + Cδ)δ1/2
= [t/2 + C(t/2)2]C2m1/2(2 + 2−(ℓ+1)C)2−(ℓ+1)/2
since 2nδ = t. This proves (6.11).
Because [t/2 + C(t/2)2]C2m1/2(2 + 2−(ℓ+1)C)2−(ℓ+1)/2 is summable with respect to ℓ, we
see that {D(Rt,ℓu)}ℓ∈N is Cauchy with respect to the L∞ norm (even though the individual
functions may not be in L∞) and hence converges uniformly to some function. We already
know that Rt,ℓu converges to Rtu, so the limit of D(Rt,ℓu) must be D(Rtu). We obtain the
estimate (3) by summing the (6.11) from ℓ to ∞ using the geometric series formula.
(cid:3)
Corollary 6.14. Let 0 ≤ c ≤ C. Let u, v ∈ E(c, C) and let t ≥ 0 be a dyadic rational.
(1) Rtu, Rtv ∈ E(c(1 + ct)−1, C(1 + Ct)−1).
(2) kD(Rtu) − D(Rtv)kL∞ ≤ (1 + Ct)kDu − DvkL∞.
(3) If u ≤ v + a + bkDvk2
2 for some a ∈ R and b ≥ 0, then
Rtu ≤ Rtv + a + b
(4) kD(Rtu)k2
2 ≤ C 2mt
1+Ct + kDuk2
2.
C2mt
1 + Ct
+ bkD(Rtv)k2
2.
Proof. We know that these properties hold for Rt,ℓ by Lemma 6.11. By Lemma 6.13, they also
hold in the limit taking ℓ → ∞. (For (1), we use the fact that E(c′, C′) is closed under pointwise
limits for each c′ and C′.)
6.6. Continuity and Semigroup Property. In order to extend Rt to all real t ≥ 0, we prove
estimates that show that Rt depends continuously on t. We begin with some simple estimates
for Pt and Qt.
Lemma 6.15. Let ℓ ∈ Z and suppose that t ∈ 2−ℓN0 and u ∈ E(0, C). Then
(cid:3)
(1) u ≤ Ptu ≤ u + Cmt/2.
(2) u − (t/2)kDuk2
(3) kD(Qtu) − Duk2 ≤ CtkDuk2.
2 ≤ Qtu ≤ u.
Proof. (1) Because u is convex and u(x) − (C/2)kxk2
2 is concave, we have
u(x) + hDu(x), yi ≤ u(x + y) ≤ u(x) + hDu(x), yi +
C
2 kyk2
2.
52
DAVID JEKEL
Integrating with respect to dσt,N (y) yields
u(x) ≤ Ptu(x) ≤ u(x) +
Cmt
2
for u ∈ E(0, C).
(2) As for the operator Qt, it is immediate from the definition that Qtu ≤ u. On the other
hand,
Qtu(x) = u(x − tD(Qtu)(x)) +
t
2kD(Qtu)(x)k2
2
≥ u(x) − thD(Qtu)(x), Du(x)i2 +
≥ u(x) −
t
2kDu(x)k2
2,
t
2kD(Qtu)(x)k2
2
2kDu(x)k2
2.
(3) Using the fact that Du is C-Lipschitz, together with Corollary 6.7 (1) and Lemma 6.6
where the last inequality follows because hD(Qtu)(x), Du(x)i2 ≤ 1
(5)
2kD(Qtu)(x)k2
2 + 1
kD(Qtu)(x) − Du(x)k2 = kDu(x − tD(Qtu)(x)) − Du(x)k2
≤ CtkD(Qtu)(x)k2
≤ CtkDu(x)k2.
(cid:3)
Lemma 6.16. Let s ≤ t be two numbers in Q+
2 , and let u ∈ E(0, C).
2 [log(1 + Ct) − log(1 + Cs)].
2 (t − s)(Cm + kDuk2
2).
(1) Rtu ≤ Rsu + m
(2) Rtu ≥ Rsu − 1
(3) If C(t − s) ≤ 1, then kD(Rtu) − D(Rsu)k2 ≤ 5Cm1/221/2(t − s)1/2 + C(t − s)kDuk2.
Moreover, if ℓ ∈ Z and if s, t ∈ 2−ℓN0, then the same estimates hold with Rt replaced by Rt,ℓ.
Proof. (1) Fix ℓ ∈ Z and let δ = 2−ℓ. Suppose that s = nδ and t = n′δ where n, n′ ∈ N0. By
the previous lemma,
R(j+1)δ,ℓu = PδQδRjδ,ℓu
≤ QδRjδ,ℓu +
Cmδ
2(1 + C(j + 1)δ)
≤ Rjδ,ℓu +
Cmδ
2(1 + C(j + 1)δ)
,
where we have used the fact that QδRjδ,ℓu ∈ E(0, C(1 + C(j + 1)δ)−1). Therefore,
Rn′δ,ℓu ≤ Rnδ,u +
n′−1Xj=n
Cmδ
2(1 + C(j + 1)δ)
.
Since the sum on the right hand side is a lower Riemann sum for the function Cmδ/2(1 + Cτ )
for τ ∈ [s, t], we obtain
Rt,ℓu ≤ Rs,ℓu +
m
2
[log(1 + Ct) − log(1 + Cs)].
We obtain (1) by letting ℓ → +∞ and using Lemma 6.13.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
53
(2) Let ℓ, δ, s, t, n, n′ be as above. By the previous lemma,
R(j+1)δ,ℓu = PδQδRjδ,ℓu
≥ QδRjδ,ℓu
≥ Rjδ,ℓu −
≥ Rjδ,ℓu −
δ
2kD(Rjδ,ℓu)k2
δ
(Cm + kDuk2
2),
2
2
where the last inequality follows from Lemma 6.11 (4). So when we sum from j = n to n′ − 1,
we obtain
Rtu ≥ Rsu −
t − s
2
(Cm + kDuk2
2).
Then (2) follows by taking ℓ → +∞.
write t − s in a binary expansion
(3) Assume that s, t ∈ 2−ℓN0. Choose k ∈ Z such that 2−k−1 ≤ t − s ≤ 2−k. Then we may
t − s =
ℓXj=k+1
aj2−j,
where aj ∈ {0, 1} for each j and ak+1 = 1. Let tj = s+ak+12−k−1+···+aj2−j. Let uj = Rtj ,ℓu.
We will estimate kDuj(x) − Duj−1(x)k2 for each j. Of course, if aj = 0, then uj = uj−1, so
there is nothing to prove. On the other hand, suppose that aj = 1. Now we estimate (at our
given point x, suppressed in the notation)
(6.12)
kD(R2−j ,ℓuj−1) − Duj−1k2 ≤ kD(R2−j ,ℓuj−1) − D(P2−j Q2−j uj−1)k2
+ kD(P2−j Q2−j uj−1) − D(Q2−j uj−1)k2
+ kD(Q2−j uj−1) − Duj−1k2.
The first term on the right hand side may be estimated as follows. Recall that we proved
Lemma 6.13 (3) from the estimate from (6.11) by summing the geometric series. The same
reasoning shows that if ℓ ≥ j and δ ∈ 2−ℓN0, then
kD(Rδ,ℓuj−1) − D(Rδ,juj−1)kL∞ ≤ [δ/2 + C(δ/2)2]C2m1/2(2 · 2−j/2 + 2−3j/2C)
since uj−1 ∈ E(0, C). When we substitute δ = 2−j, R2−j ,j is simply equal to P2−j Q2−j . Thus,
at the point x,
kD(R2−j ,ℓuj−1) − D(P2−j Q2−j uj−1)k2 ≤ C2m1/2[2−j/2 + C2−2j/4][2 · 2−j/2 + 2−3j/2C].
By our assumption C2−j ≤ C(t − s) ≤ 1 and hence we may replace C2−2j/4 by 2−j/2 and
repalce 2−3j/2C by 2−j/2 and hence
kD(R2−j ,ℓuj−1) − D(P2−j Q2−j uj−1)k2 ≤ 3C2m1/22−3j/2 ≤ 3Cm1/22−j/2.
The second term on the right hand side of (6.12) can be estimated by Lemma 6.3 (2) by
kD(P2−j Q2−j uj−1) − D(Q2−j uj−1)k2 ≤ Cm1/22−j/2
since Q2−j uj−1 ∈ E(0, C). The third term on the right hand side of (6.12) can be estimated
using Corollary 6.15 (3) by
Meanwhile, by Lemma 6.11 (4) and the triangle inequality
kD(Q2−j uj−1) − Duj−1k2 ≤ C2−jkDuj−1k2.
kDuj−1k2 ≤qCm + kDuk2
2 ≤ C1/2m1/2 + kDuk2.
54
DAVID JEKEL
So using the fact C2−j ≤ 1, we have
kD(Q2−j uj−1) − Duj−1k2 ≤ C3/2m1/22−j + C2−jkDuk2 ≤ Cm1/22−j/2 + C(tj − tj−1)kDuk2.
Therefore, plugging all our estimates into (6.12), we get
kDuj − Duj+1k2 ≤ 5Cm1/22−j/2 + C(tj − tj−1)kDuk2.
Then summing from j = k + 1 to ℓ we obtain
kDuℓ − Dukk2 ≤ 5Cm1/22−k/2 + C(t − s)kDuk2
≤ 5Cm1/221/2(t − s)1/2 + C(t − s)kDuk2.
Because uℓ = Rt,ℓu and uk = Rs,ℓu, we have shown that (3) holds for Rs,ℓ and Rt,ℓ instead of
Rs and Rt. Thus, (3) follows by taking ℓ → +∞.
(cid:3)
Proof of Theorem 6.1. Lemma 6.16 shows that if t ≥ 0 and if tℓ is a sequence of dyadic ratio-
nals converging to t as ℓ → ∞, then Rtℓu converges to some function v and this function is
independent of the approximating sequence, so we define Rtu = v. Claim (1), (3), and (4) of
the Theorem were proved for dyadic t in Corollary 6.14 (1), Lemma 6.16, and Corollary 6.14
(2) - (4) respectively, and each of these claims can be extended to real t ≥ 0 in light of the
continuity estimates Lemma 6.16. Claim (3) of the theorem is Lemma 6.13.
Thus, it remains to show that Rt is a semigroup. That is, we must show that RsRtu = Rs+tu
for u ∈ E(0, C) (and we have not even checked this for dyadic s, t yet). First, we check this
property for real s, t ≥ 0 under the additional restriction that Ct ≤ 1/2. For each ℓ ∈ Z, there
exist sℓ and tℓ ∈ 2−ℓN0 such that s − 2−ℓ < sℓ ≤ s and t − 2−ℓ < tℓ ≤ t. By Lemma 6.16 (1)
and (2) we have
Rtℓu − Rtu ≤ tℓ − t
2
(Cm + kDuk2
2) ≤ 2−ℓ 1
2
(Cm + kDuk2
2),
since log(1 + Ctℓ)− log(1 + Ct) ≤ Ctℓ− t (from computation of the derivative of log(1 + Ct)).
By Lemma 6.13 (1) and (2), if C2−ℓ−1 ≤ 1, then
Rtℓ,ℓu − Rtℓu ≤ 2−ℓ(cid:18) 3
2
C2mt
1 + Ct
2)(cid:19) .
+ log(1 + Ctℓ)(m + Cm + kDuk2
Since tℓ ≤ t, we can replace tℓ by t on the right hand side. By the triangle inequality, we obtain
(6.13)
Rtℓ,ℓu − Rtu ≤ 2−ℓKt(1 + kDuk2
2)
for some constant Kt depending on t (and C). Using Lemma 6.16 (3), or rather its extension
to real values of t,
kD(Rtu) − Duk2 ≤ 5Cm1/221/2t1/2 + CtkDuk2
≤ 5Cm1/221/2t1/2 + CtkD(Rtu) − Duk2 + CtkD(Rtu)k2.
Hence,
kD(Rtu) − Duk2 ≤ (1 − Ct)−1[5Cm1/221/2t1/2 + CtkD(Rtu)k2],
so by the triangle inequality,
kDuk2 ≤ kD(Rtu)k2 + (1 − Ct)−1[5Cm1/221/2t1/2 + CtkD(Rtu)k2.
By squaring and applying the arithmetic-geometric mean inequality, we get
kDuk2
2 ≤ At + BtkD(Rtu)k2
2
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
55
for some constants A and B depending on t. The same reasoning applies to Rtℓ,ℓ since Lemma
6.16 (3) holds for Rtℓ,ℓ also. We thus obtain
kDuk2 ≤ kD(Rtℓ,ℓu)k2 + (1 − Ctℓ)−1[5Cm1/221/2t1/2
ℓ + CtℓkD(Rtℓ,ℓu)k2
≤ kD(Rtℓ,ℓu)k2 + (1 − Ct)−1[5Cm1/221/2t1/2 + CtkD(Rtℓ,ℓu)k2
and so
Overall,
kDuk2
2 ≤ At + BtkD(Rtℓ,ℓu)k2
2.
Rtu ≤ Rtℓ,ℓu + 2−ℓKt(1 + At + BtkD(Rtℓ,ℓu)k2
Rtℓ,ℓu ≤ Rtu + 2−ℓKt(1 + At + BtkD(Rtu)k2
2
2
So by Lemma 6.11 (3) and (4)
Rsℓ,ℓRtu ≤ Rsℓ,ℓRtℓ,ℓu + 2−ℓK1(1 + At + BtkD(Rsℓ,ℓRtℓ,ℓu)k2
2)
≤ Rsℓ,ℓRtℓ,ℓu + 2−ℓK1(1 + At + CmBt + BtkDuk2
2)
and the same holds with Rt and Rtℓ,ℓ switched, so that
(6.14)
Rsℓ,ℓRtu − Rsℓ+tℓ,ℓu ≤ 2−ℓK1(1 + At + CmBt + BtkDuk2
2),
where we have noted that Rsℓ+tℓ,ℓu = Rsℓ,ℓRtℓ,ℓu.
But then by Lemma 6.11. By the same token as (6.13), since Rtu ∈ E(0, C), we have
(6.15)
Similarly, since (s + t) − (sℓ + tℓ) ≤ 2 · 2−ℓ, we have
(6.16)
Rsℓ,ℓRtu − RsRtu ≤ 2−ℓKs(1 + kD(Rtu)k2
2).
Rsℓ+tℓ,ℓu − Rs+tu ≤ 2−ℓ · 2Ks+t(1 + kDuk2
2).
Combining these with (6.14) using the triangle inequality, we get
RsRtu − Rs+tu ≤ 2−ℓK1(1 + At + CmBt + BtkDuk2
2)
2) + 2−ℓ · 2Ks+t(1 + kDuk2
2).
Taking ℓ → ∞, we get RsRtu = Rs+tu as desired. This completes the case when Ct ≤ 1/2.
1/2. Then for j = 1, . . . , n− 1, we have Rn−j
In the general case, suppose s, t ≥ 0 and u ∈ E(0, C). Choose n large enough that Ct/n ≤
t/k u ∈ E(0, C). Therefore, by the previous argument
+ 2−ℓKs(1 + kD(Rtu)k2
Rs+jt/nRn−j
t/n u = (Rs+jt/nRt/n)(Rn−j−1
t/n
u) = Rs+(j+1)t/nRn−j−1
t/n
u,
so by induction Rs+tu = RsRn
Rtu. Thus, Rs+tu = RsRtu.
t/nu. Since this also holds with s replaced by 0, we have Rn
t/nu =
(cid:3)
6.7. Solution to the Differential Equation. It remains to show that the semigroup Rt
produces solutions to the differential equation ∂tu = (1/2N )∆u − (1/2)kDuk2
2, and that the
result agrees with the solution produced by solving the heat equation for exp(−N 2u). More
precisely, we will prove the following.
Theorem 6.17. Let u0 : MN (C)m
u(x, t) = Rtu(x). Then u is a smooth function on MN (C)m
∂tu = (1/2N )∆u − (1/2)kDuk2
sa → R be a given function in E(c, C) for some c ≥ 0. Let
sa×(0, +∞) and it solves the equation
2. Moreover, exp(−N 2 · Rtu0) = Pt[exp(−N 2u0)].
56
DAVID JEKEL
At this point, we have not proved enough smoothness for Rtu to show that it solves the
equation in the classical sense. Therefore, as an intermediate step, we show that u solves the
equation in the viscosity sense defined by [11]; see [10] and the references cited therein for
further background. We will then deduce that exp(−N 2u) is a viscosity solution of the heat
equation and hence show it agrees with the smooth solution of the heat equation.
The definition of viscosity solution for parabolic equations is as follows. Here we continue
sa with the normalized inner product (rather than Rn for some
sa → R, we denote by Du and Hu the gradient and Hessian with
sa, then Du(x0) is the vector
to use the vector space MN (C)m
n). For smooth u : MN (C)m
respect to the inner product h·,·i2; in other words, if x0 ∈ MN (C)m
in MN (C)m
sa and Hu(x0) is the linear transformation MN (C)m
u(x) = u(x0) + hDu(x0), x − x0i2 +
We denote the space of linear transformations MN (C)m
denote the self-adjoint elements by B(MN (C)m
Definition 6.18. Let F : B(MN (C)m
consider the partial differential equation
sa)sa.
sa → MN (C)m
sa such that
1
2hHu(x0)[x − x0], x − x0i2 + o(kx − x0k2
2).
sa by B(MN (C)m
sa → MN (C)m
sa), and we
sa)sa × MN (C)sa × R × MN (C)sa → R be continuous, and
(6.17)
We say that a function u : MN (C)m
semi-continuous and if the following condition holds: Suppose that
∂tu = F (Hu, Du, u, x).
sa × [0, +∞) → R is a viscosity subsolution if it is upper
x0 ∈ MN (C)m
sa,
and suppose that u satisfies
t0 > 0, A ∈ B(MN (C)m
sa)sa,
p ∈ MN (C)m
sa, α ∈ R,
(6.18) u(x, t) ≤ u(x0, t0)+α(t−t0)+hp, x−x0i2 +
Then we also have
1
2hA(x−x0), x−x0i2 +o(t−t0+kx−x0k2
2).
(6.19)
Definition 6.19. With the same setup as above, we say that u : MN (C)m
sa × [0, +∞) → R
is a viscosity supersolution if it is lower semi-continuous and the following condition holds: If
x0, t0, A, p, α are as above and if
α ≤ F (A, p, u(x0), x0).
(6.20) u(x, t) ≥ u(x0, t0)+α(t−t0)+hp, x−x0i2 +
then
1
2hA(x−x0), x−x0i2 +o(t−t0+kx−x0k2
2),
(6.21)
α ≥ F (A, p, u(x0), x0).
Definition 6.20. We say that u is a viscosity solution if it is both a subsolution and a super-
solution.
Remark 6.21. Roughly speaking, being a viscosity solution means that whenever there exist
upper or lower second-order Taylor approximations to u, then we can evaluate the differential
operator F on the Taylor approximation and get an inequality in one direction.
Example 6.22. The heat equation ∂tu = (1/2N )∆u is obtained by taking
F (A, p, u, x) =
1
2N 2 Tr(A).
To understand why 1/N 2 is the correct normalization on the right hand side, suppose that u
is smooth and A = Hu(x0) and p = Du(x0), so that
u(x) = u(x0) + hp, x − x0i2 +
1
2hA(x − x0), x − x0i2 + o(kx − x0k2
2).
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
57
In terms of the non-normalized inner product (which we denote by the dot product), this means
that
u(x) = u(x0) +
1
N
p · (x − x0) +
(A(x − x0)) · (x − x0).
1
2N
Thus, the Hessian with respect to the non-normalized inner product is (1/N )A. Hence,
(1/N )∆u(x0) = (1/N 2) Tr(A). Similarly, the equation ∂tu = (1/2N )∆u − (1/2)kDuk2
2 is
obtained by taking
F (A, p, u, x) =
1
2N 2 Tr(A) −
1
2kpk2
2.
Proposition 6.23. Let u0 ∈ E(0, C) and define u(x, t) = Rtu0(x). Then u is a viscosity
solution of the equation ∂tu = (1/2N )∆u − (1/2)kDuk2
2.
Proof. First, note that u is continuous. Indeed, by Theorem 6.1 (3), u is continuous in t with
a rate of continuity that is uniform for x in a bounded region (this follows because the term
kDu0k2
2 on the right hand side of Lemma 6.16 (2) is bounded on bounded regions since Du0 is
C-Lipschitz). Also, u(·, t) is continuous for each t since it is in E(0, C). Together, this implies
u is jointly continuous in (x, t).
To show that u is a viscosity subsolution, suppose that we have a lower second-order ap-
sa and t0 > 0, given by
1
2hA(x − x0), x − x0i2 + o(t − t0 + kx − x0k2
2).
Our first goal is to replace the soft bound o(t − t0 + kx − x0k2
proximation at the point (x0, t0), where x0 ∈ MN (C)m
u(x, t) ≥ u(x0, t0) + α(t − t0) + hp, x − x0i +
Then we must show that α ≤ (1/2N 2) Tr(A) − (1/2)kpk2
2.
2) by a more explicit error
bound, at the cost of modifying α and A by some positive ǫ. Pick ǫ > 0. Then there exists
r > 0 such that if t − t0 + kx − x0k2
1
2h(A − ǫI)(x − x0), x − x0i2.
(6.22) u(x, t) ≥ u(x0, t0) + α(t − t0) − ǫt − t0 + hp, x − x0i +
Let us assume that t0 − r < t ≤ t0, so that the above inequality holds for kx− x0k2 < r and we
have α(t − t0) − ǫt− t0 = (α + ǫ)(t− t0). For x such that kx − x0k2
2 ≥ r, we may use Theorem
6.1 (3b), the fact that Du is C-Lipschitz, and the convexity of u to conclude that
2 < 2r, then we have
u(x, t) ≥ u0(x) −
t
2
(Cm + kDuk2
2)
≥ u0(x0) + hDu(x0), x − x0i2 −
t
2
(Cm + (kDu(x0)k2 + Ckx − x0k2)2)
In other words, u is bounded below by a quadratic in x − x0, and the estimate holds uniformly
for t in a bounded interval. Moreover, the right hand side of (6.22) is also bounded by a
quadratic in x − x0 uniformly for t ∈ [t0 − r, t0 + r]. It follows that for a large enough constant
Kǫ, we have
1
2h(A − ǫI)(x − x0), x − x0i2 − u(x, t) ≤ Kǫkx − x0k4
u(x0, t0) + (α + ǫ)(t − t0) + hp, x − x0i +
whenever t ∈ (t0 − t, t0] and kx − x0k2 ≥ r. Therefore, overall, assuming that t ∈ (t0 − r, t0],
we have
2
(6.23) u(x, t) ≥ u(x0, t0)+(α+ǫ)(t−t0)+hp, x−x0i+
1
2h(A−ǫI)(x−x0), x−x0i2−Kǫkx−x0k4
For t ∈ R, let us denote ut(x) = u(x, t) = Rtu0(x). Now the strategy for proving α + ǫ ≤
(1/2N 2) Tr(A − ǫI) − (1/2)kpk2
2 is roughly to use the fact that ut0(x0) = Rδut0−δ(x0) and
estimate ut0−δ(x0) from above using the upper Taylor approximation for small δ > 0. However,
2
58
DAVID JEKEL
for the sake of computation, it is easier to estimate QδPδut0−δ rather than Rδ (and then we
will control the error between Rδ and QδPδ using Lemmas 6.12 and 6.13).
Let δ ∈ (0, r). Then using the above inequality and monotonicity of Pδ, we have
Pδut0−δ(x) ≥ ut0(x0) + (α − ǫ)δ + hp, x − x0i
1
2N 2 Tr(A − ǫI)δ +
+
− Kǫ(kx − x0k4
1
2h(A − ǫI)(x − x0), x − x0i
2 + 2(1 + 2/N 2)mδkx − x0k2
2 + 2m2(1 + 2/N 2)δ2).
Here we have evaluated Pδ applied to kx−x0k4
2 using Example 3.22 and translation-invariance of
Pδ. Now recall that QδPδut0−δ(x0) is obtained by evaluating Pδut0−δ at x0−δD(QδPδut0−δ)(x0).
Also, in light of Lemma 6.11 (4) and Corollary 6.14 (4), kD(QδPδut0−δ)(x0)k2
2 is bounded by
kDu0(x0)k2
2 plus a constant. In particular, kD(QδPδut0−δ)(x0)k2 is bounded as δ → 0. There-
fore,
(6.24)
QδPδut0−δ(x0) = Pδut0−δ(x0 − δD(QδPδut0−δ)(x0)) +
1
2
δkD(QδPδut0−δ)(x0)k2
2
≥ ut0(x0) +
1
2kD(QδPδut0−δ)(x0)k2
2δ
+ (α + ǫ)(−δ) − hp, D(QδPδut0−δ)(x0)iδ +
1
2N 2 Tr(A − ǫI) + O(δ2).
(Here the the implicit constant in O(δ2) depends on ǫ.)
Because ut0−δ ∈ E(0, C), Lemma 6.12 (2) and (3) imply that if Cδ ≤ 1, then
QδPδut0−δ(x0) − PδQδut0−δ(x0) ≤ 2C2mδ2 + 2Cδ2kD(PδQδut0−δ)(x0)k2.
Again by Lemma 6.11 (4) and Theorem 6.1 (4c), kD(QδPδut0−δ)(x0)k2
plus a constant, so that
2 is bounded by kDu0(x0)k2
2
QδPδut0−δ = PδQδut0−δ + O(δ2).
Also, if we let δℓ = 2−ℓ for ℓ ∈ Z, then Lemma 6.13 implies that when 2Cδℓ ≤ 1 and δℓ < r,
then
PδℓQδℓut0−δ(x0) − Rδℓ ut0−δℓ(x0) = Rδℓ,ℓut0−δℓ(x0) − Rδℓut0−δℓ(x0)
≤(cid:18) 3
2
= O(δ2
C2mδℓ
1 − Cδℓ
ℓ ).
+ log(1 + Cδℓ)(m + Cm + kDu(x0)k2
2(cid:19) 2−ℓ
So overall
(6.25)
QδℓPδℓ ut0−δℓ(x0) = Rδℓut0−δℓ(x0) + O(δ2
ℓ ) = ut0(x0) + O(δ2
ℓ ).
Using similar reasoning, Lemma 6.12 (1) shows that
D(QδℓPδℓut0−δℓ )(x0) = D(Pδℓ Qδℓut0−δℓ)(x0) + O(δ3/2
ℓ
).
Then using Lemma 6.13 (3), we obtain
D(Pδℓ Qδℓut0−δℓ)(x0) = D(Rδℓ ut0−δℓ)(x0) + O(δ3/2).
Finally, because ut0−δ ∈ E(0, C), it is differentiable everywhere; the upper Taylor approximation
(6.22) implies that ut0(x) ≤ ut0(x0) + hp, x − x0i2 + o(kx − x0k2) and therefore p must equal
Dut0(x0). Thus, overall
(6.26)
D(QδℓPδℓ ut0−δℓ)(x0) = p + O(δ3/2
ℓ
).
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
59
Substituting (6.25) and (6.26) into (6.24), we obtain
ut0(x0) ≥ ut0(x0) +
1
2kpk2
2δℓ + (α + ǫ)(−δℓ) − kpk2
2δℓ +
1
2N 2 Tr(A − ǫI) + O(δ2
ℓ ).
We cancel ut0 (x0) from both sides, divide by δℓ, and move α−ǫ to the left hand side to conclude
that
1
2N 2 Tr(A − ǫI) −
1
2kpk2
α + ǫ ≥
2 + O(δℓ).
Then taking ℓ → ∞, we get α + ǫ ≥ (1/2N 2) Tr(A − ǫI) − (1/2)kpk2
have α ≥ (1/2N 2) Tr(A) − (1/2)kpk2
To show that the u is a viscosity subsolution, the argument is symmetrical for the most part.
However, to obtain the constant Kǫ in (6.23), we used the one-sided estimate Theorem 6.1 (3b)
to show that u is bounded below by a quadratic in x − x0 that is independent of t, so long
as t ∈ (t0 − r, t0]. To show that u is a viscosity supersolution, we want to prove an analogous
quadratic upper bound. But by Theorem 6.1 (3a) and semi-concavity of u0, we have for t ≤ t0
that
2. This shows that u is a viscosity supersolution.
2. Since ǫ was arbitrary, we
ut(x) ≤ u0(x) +
m
2
log(1 + Ct0)
≤ u0(x0) + hDu0(x0), x − x0i +
C
2 kx − x0k2
2 +
m
2
log(1 + Ct0),
which is the desired upper bound. The rest of the argument is symmetrical except that α + ǫ
is replaced by α − ǫ and A − ǫI is replaced by A + ǫI.
(cid:3)
Lemma 6.24. Let u : MN (C)m
sa × [0, +∞) → R. Then u is a viscosity solution to ∂tu =
(1/2N )∆u − (1/2)kDuk2
2 if and only if exp(−N 2u) is a viscosity solution to ∂tu = (1/2N )∆u.
Proof. More precisely, we claim that u is a viscosity subsolution if and only if exp(−N 2u)
is viscosity supersolution and vice versa. Suppose that u is a subsolution, and let us show
that v = exp(−N 2u) is a supersolution. If u is upper semi-continuous, then v is lower semi-
continuous. Now suppose that we have a lower Taylor approximation at (x0, t0)
1
2hA(x − x0), x − x0i2 + o(t − t0 + kx − x0k2
2).
v(x, t) ≥ v(x0, t0) + α(t − t0) + hp, x − x0i2 +
Note that v > 0 and u = −1/N 2 log v. The function h 7→ log h is increasing and analytic for
h > 0 and we have
log(h + δ) = log(h) + log(1 + δ/h) = log(h) +
δ
h −
1
2(cid:18) δ
h(cid:19)2
+ O(δ3).
Substituting h = v(x0, t0) = exp(−N 2u(x0, t0)) and δ = v(x, t) − v(x0, t0) = α(t − t0) + hp, x −
x0i2 + 1
2hA(x − x0), x − x0i2 + o(t − t0 + x − x02), we get
(t − t0) +
−N 2u(x, t) ≥ −N 2u(x0, t0) +
v(x0, t0)
α
+
1
2v(x0, t0)hA(x − x0), x − x0i2 −
+ o(t − t0 + kx − x0k2
2),
1
v(x0, t0)hp, x − x0i2
2v(x0, t0)2hp, x − x0i2
1
2
since hp, x − x0i2/v(x0, t0)2 is the only term from (−1/2)(δ/h)2 + O(δ3) that is not o(t − t0 +
kx− x0k2
2). Let
2) (here we use the fact that t− t0kx − x0k2 ≤ (2/3)t− t03/2 + (1/3)kx − x0k3
60
DAVID JEKEL
us denote by P the linear map P (x − x0) = php, x − x0i2. Then the above inequality becomes
u(x, t) ≤ u(x0, t0) −
1
α
N 2v(x0, t0)
(t − t0) −
1
N 2v(x0, t0)hp, x − x0i2
2N 2v(x0, t0)hA(x − x0), x − x0i2 +
−
+ o(t − t0 + kx − x0k2
2).
Because u is a subsolution, we have
1
2N 2v(x0, t0)2 hP (x − x0), (x − x0)i
−
α
1
2N 4 Tr(A) +
N 2v(x0, t0) ≤ −
2N 4v(x0, t0)2 Tr(P ) −
2, so the last two terms cancel. Thus, α ≥ 1
1
1
2N 4v(x0, t0)2 kpk2
2.
2N 2 Tr(A) as desired. So v is a
But Tr(P ) = kpk2
supersolution.
A symmetrical argument shows that if v is a supersolution, then u is a subsolution. The other
two claims are proved in the same way except using the Taylor expansion of the exponential
function instead of the logarithm.
(cid:3)
Now we are ready to prove Theorem 6.17 in the special case where u0 is bounded below.
Lemma 6.25. Suppose that u0 ∈ E(0, C) is bounded below. Then exp(−N 2Rtu0) = Pt[exp(−N 2u0)].
Proof. Let v(x, t) = exp(−N 2Rtu0(x)) and let w(x, t) = Pt[exp(−N 2u0)](x). Since u0 is
bounded below by some constant K, we have Rtu0 ≥ K by monotonicity of Rt (see Corollary
6.14 (3)) and the fact that it does not affect constant functions (since the same is true of Pt
and Qt). Hence, v = exp(−N 2Rtu0) ≤ exp(−N 2K). We also have exp(−N 2u0) ≤ exp(−N 2K)
and hence w ≤ exp(−N 2K).
Thus, v and w are both bounded, w is a smooth solution to the heat equation, and v is a
viscosity solution by the previous lemma. We will conclude from this that v = w (and this is
nothing but a standard argument for the maximum principle together with the basic philosophy
of viscosity solutions).
To show that v ≤ w, choose ǫ > 0, and consider the function
φ(x, t) = w(x, t) − v(x, t) − ǫkxk2
2 − 2mǫt.
Suppose for contradiction that φ > 0 at some point. Since φ is continuous on MN (C)m
sa×[0, +∞)
and since w and v are bounded, φ achieves a maximum at some (x0, t0). Since the maximum
is strictly positive, we have t0 > 0. Let ψ(x, t) = w(x, t) − (1/2)ǫkxk2
2 − 2ǫt, so that φ(x, t) =
v(x, t) − ψ(x, t). Then φ(x, t) ≤ φ(x0, t0) implies that
v(x, t) ≥ v(x0, t0) + ψ(x, t) − ψ(x0, t0)
= v(x0, t0) + ∂tψ(x0, t0)(t − t0) + hDψ(x0, t0), x − x0i2
+
1
2hHψ(x0, t0)(x − x0), x − x0i2 + o(t − t0 + kx − x0k2
2),
where the last step follows because ψ is smooth. Because v is a viscosity supersolution,
∂tψ(x0, t0) ≥
1
2N
∆ψ(x0, t0).
However, this is a contradiction because at every point (x, t), we have
∂tψ = ∂tw − 2mǫ <
1
2N
∆w − mǫ =
1
2N
∆ψ,
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
61
by computation and the fact that w solves the heat equation. It follows that φ ≤ 0 and hence
v(x, t) ≥ w(x, t) − 1
2 − mǫt. Since ǫ was arbitrary v ≥ w. Then a symmetrical argument
shows that v ≤ w.
2 ǫkxk2
(cid:3)
Thus, to prove Theorem 6.17, it only remains to remove the boundedness assumption on u0.
We achieve this by replacing u0 with the function
(6.27)
u0(x) = u0(x) − hDu0(0), xi2,
which is nonnegative by convexity of u0 and hence bounded below.
Lemma 6.26. Let u0 ∈ E(0, C) and let u0 be given by (6.27). Let v0 = exp(−N 2u0) and
v0 = exp(−N 2 u0). Then the integral defining Pt exp(−N 2u0) is well-defined and also
2)Ptv0(x − tDu0(0))
2) dy. Also, denote p = Du0(0).
Ptv0(x) =
Ptv0(x) = exp(−N 2hDu0(0), xi +
N 2t
2 kDu0(0)k2
Proof. We can express dσt,N (y) = (1/ZN ) exp(−(N 2/2t)kyk2
Then
ZN Z exp(−N 2u0(x + y)) exp(−(N 2/2t)kyk2
ZN Z exp(−N 2 u0(x + y) − N 2hp, x + yi −
ZN Z exp(−N 2 u0(x + y) − N 2hp, xi +
ZN Z exp(−N 2 u0(x − tp + z) − N 2hp, xi +
= exp(−N 2hp, xi +
2)Ptv0(x − tp).
N 2t
2 kpk2
=
=
=
1
1
1
1
2) dy
N 2
2t kyk2
2) dy
N 2t
2 kpk2
2 −
N 2t
2 kpk2
N 2
2t ky + tpk2
N 2
2t kzk2
2 −
2) dy
2) dy
(cid:3)
Lemma 6.27. Let u0 ∈ E(0, C), let p ∈ MN (C)m
2kpk2
2.
2kpk2
2.
(1) Ptu0(x) = Pt u0(x) + hp, xi2.
(2) Qtu0(x) = Qt u0(x − tp) + hp, xi2 − t
(3) Rtu0(x) = Rt u0(x − tp) + hp, xi2 − t
sa, and let u0(x) = u0(x) − hp, xi2. Then
Proof. (1) holds because Pt is linear and it does not affect linear functions. To prove (2), fix x
and let y be the point where the infimum defining Qtu0(x) is achieved and let y be the point
where the infimum defining Qt u0(x − tp) is achieved. By Corollary 6.7 (1), the points y and y
are characterized respectively by the relations
y = x − tDu0(y),
y = x − tp − tD u0(y).
But D u0(y) = Du0(y) − p. Thus, x − tDu0(y) = y, so that y = y.
Qtu0(x) = u0(y) +
2
1
2tky − xk2
1
2tky − xk2
= u0(y) + hp, yi2 +
1
t
2tky − (x − tp)k2
2kpk2
= u0(y) + hp, xi2 −
2 +
t
2kpk2
2.
= Qt u0(x − tp) + hp, xi2 −
2
2
62
DAVID JEKEL
(3) It follows by iteration (after some computation) that for t ∈ 2−ℓN0, we have Rt,ℓu0(x) =
Rt u0(x − tp) + hp, xi2 − t
2. Then by Lemma 6.13, we may take ℓ → ∞, and by Theorem
6.1 (3), we may extend the inequality to all real t.
2kpk2
(cid:3)
Proof of Theorem 6.17. We have already proved the case where u0 is bounded. For the general
case, let u0 ∈ E(0, C). Define p = Du0(0) and u0(x) = u0(x) − hp, xi2. As remarked above, u0
is bounded below by zero. By Lemmas 6.26, the bounded case, and 6.27,
Pt exp(−N 2u0)(x) = exp(cid:18)−N 2hp, xi +
= exp(cid:18)−N 2hp, xi +
= exp(cid:18)−N 2(cid:18)Rt u0(x − tp) + hp, xi −
= exp(−N 2Rtu0(x)).
N 2t
2 kpk2
N 2t
2 kpk2
2(cid:19) [Pt exp(−N 2 u0)](x − tp)
2(cid:19) exp(−N 2Rt u0(x − tp))
2(cid:19)(cid:19)
t
2kpk2
In particular, since Pt exp(−N 2 u0) is smooth for t > 0, we see that all the functions in the
above equation are smooth for t > 0, and hence Rtu0(x) is smooth function of (x, t). Also,
Pt[exp(−N 2u0)] = exp(−N 2Rtu0) as desired.
(cid:3)
6.8. Approximation by Trace Polynomials. Now we are ready to prove that Rt preserves
asymptotic approximability by trace polynomials.
Proposition 6.28. Let {VN} be a sequence of functions MN (C)m
sa → R such that VN is convex
and VN (x) − (C/2)kxk2
2 is concave, and {DVN} is asymptotically approximable by trace poly-
nomials. Then for every t > 0, the sequences {D(PtVN )}, {D(QtVN )}, and {D(RtVN )} are
asymptotically approximable by trace polynomials.
Proof. The fact that {D(PtVN )} is asymptotically approximable by trace polynomials follows
from Lemma 3.28.
Now consider D(QtVN ). Note that by Corollary 6.7 (1), D(QtVN )(x) is the solution of the
fixed point equation
y = DVN (x − ty).
But if t < 1/C, then y 7→ DVN (x − ty) is a contraction and thus iterates of this function will
converge to the fixed point. Let us define φN,0(x) = 0 and φN,ℓ+1(x) = DVN (x− tφN,ℓ(x)). By
Lemma 6.6 (5), the distance from the fixed point D(QtVN )(x) from 0 is bounded by kDVN (x)k2,
hence,
kφN,ℓ(x) − D(QtVN )(x)k2 ≤ CℓtℓkDVN (x)k2.
Because DVN,t is C-Lipschitz, Lemma 3.27 implies that {φN,ℓ}N is asymptotically approximable
by trace polynomials.
Now kDVN (0)k2 is bounded by some constant A as N → ∞ because DVN is asymptotically
approximable by trace polynomials. Since DVN is also C-Lipschitz, kDVN (x)k2 ≤ A + Ckxk2.
In particular, kφN,ℓ(x)−D(QtVN )(x)k2 ≤ Cℓtℓ(A+Ckxk2). Thus, by Lemma 3.26, {D(QtVN )}
is asymptotically approximable by trace polynomials.
t/n where n is large
enough that t/n < 1/C, and then iterating the previous statement shows that {QtVN} is
asymptotically approximable by trace polynomials.
2 , we know {D(Rt,ℓVN )} is asymp-
totically approximable by trace polynomials. By Theorem 6.1 (1c) and Lemma 3.26, the se-
quence {D(RtVN )} is asymptotically approximable by trace polynomials for t ∈ Q+
2 . Finally,
For the sequence {D(RtVN )}, first note that when t ∈ Q+
This holds whenever t < 1/C. But for general t, we can write Qt = Qn
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
63
by Theorem 6.1 (2d) and Lemma 3.26, the sequence {D(RtVN )} is asymptotically approximable
by trace polynomials for all t ∈ R+.
(cid:3)
7. Main Theorem on Free Entropy
We are now ready to prove the following theorem which shows that χ = χ∗ for a law which
is the limit of log-concave random matrix models.
pendent of N .
Theorem 7.1. Let µN be a sequence of probability measures on MN (C)m
VN . Assume
(A) The potential VN (x) is convex and VN (x) − (C/2)kxk2
(B) The sequence µN concentrates around some non-commutative law λ with λ(X 2
(C) For some R0 > 0, we have limN→∞Rkxk≥R0
(D) The sequence {DVN} is asymptotically approximable by trace polynomials.
Then λ ∈ Σm,R0 and moreover
(1) The law λ has finite Fisher information Φ∗(λ), and for all t ≥ 0, we have
(1 + kxk2
2) dµN (x) = 0.
j ) > 0.
sa given by the potential
2 is concave for some C > 0 inde-
lim
N→∞
1
N 3I(µN ∗ σt,N ) → Φ∗(λ ⊞ σt).
(2) We have for all t ≥ 0,
χ(λ ⊞ σt) = χ(λ ⊞ σt) = lim
N→∞
1
N 2(cid:16)h(µN ∗ σt,N ) +
m
2
log N(cid:17) = χ∗(λ ⊞ σt).
(3) The functions t 7→ 1
N 3I(µN ∗ σt,N ) and t 7→ Φ∗(λ ⊞ σt) are decreasing and Lipschitz and
and the absolute value of the derivative (where defined) is bounded by C2m(1 + Ct)−2.
Remark 7.2. If VN (x)− (c/2)kxk2
2 is concave and if {DVN} is
asymptotically approximable by trace polynomials, then Theorem 4.1 implies that µN satisfies
the hypotheses of Theorem 7.1 for some law λ.
2 is convex and VN (x)− (C/2)kxk2
However, Theorem 7.1 holds in a slightly more general situation than Theorem 4.1 in that
we do not have to assume uniform convexity, finite moments, or exponential concentration.
In preparation for the proof of the Theorem 7.1, we have already verified that the hypotheses
(A), (C), and (D) are preserved under Gaussian convolution. Now we show that (B) is preserved
in Lemma 7.4. This is straightforward apart from one subtlety -- although we have assumed
that for every non-commutative polynomial p, the non-commutative moment τN (p(x)) concen-
trates around λ(p) under µN , we have not assumed that τN (p(x)) has finite expectation. To
deal with this issue, we first prove an auxiliary lemma.
Lemma 7.3. Let λ be a non-commutative law in Σm, let p(X, Y ) = p(X1, . . . , Xm, Y1, . . . , Ym)
be a non-commutative polynomial of 2m variables, and let R > 0. Then there exists a neighbor-
hood V of λ in Σm and a constant L such that, for all N ∈ N, for all x ∈ ΓN (V), the function
y 7→ τN (p(x, y)) is L-Lipschitz with respect to k·k2 for self-adjoint tuples y in the operator-norm
ball {y : kyjk ≤ R}.
Proof. To prove the lemma, it suffices to consider the case of a non-commutative monomial.
j=1 pj where pj is a monomial, and if we find neighborhoods Vj and Lipschitz
j=1 Lj.
Thus, assume without loss of generality that p(X, Y ) is a non-commutative monomial. Then
Indeed, if p =Pn
constants Lj for each pj, then the result will also hold for p with V =Tn
j=1 Vj and L =Pn
it can be written in the form
p(X, Y ) = q0(X)Yi1 q1(X)Yi2 . . . qℓ−1(X)Yiℓ qℓ(X).
64
DAVID JEKEL
where ij ∈ {1, . . . , m} and qj(X) is a non-commutative monomial in X (which of course is
allowed to be 1). Consider x, y, y′ ∈ MN (C)m
ik∞ ≤ R
for each i. Then
sa, and suppose that kyik∞ ≤ R and ky′
Recalling the non-commutative Lα norms and Holder's inequality (see 2.3), we have
q0(x)yi1 . . . yij−1 qj−1(x)(yij − y′
ij )qi(x)yij+1 . . . yiℓqℓ(x).
p(x, y) − p(x, y′) =
ℓXj=1
kp(x, y) − p(x, y′)k1 ≤
ℓXj=1Yk6=j
τN (p(x, y)) − τN (p(x, y′)) ≤
This implies that
ky′
kqj(x)k2(ℓ+1)Yk<j
ℓXj=1Yk6=j
ikk∞kyj − y′
kyikk∞Yk>j
kqj(x)k2(ℓ+1) Rℓ−1ky − y′k2.
jk2.
Now
We can define
kqj(x)k2(ℓ+1) =(cid:0)τN [(qj(x)∗qj(x))ℓ+1](cid:1)1/2(ℓ+1)
.
V = {λ′ : λ′[(q∗
j qj)ℓ+1] < λ[(q∗
j qj)ℓ+1] + 1 for j = 0, . . . , ℓ}.
Then kqj(x)k2(ℓ+1) is uniformly bounded for x ∈ ΓN (V) for each j = 0, . . . , ℓ. Suppose that
each of these quantities is bounded by K. Then the above estimate shows that
τN (p(x, y)) − τN (p(x, y)) ≤ ℓK ℓ+1Rℓ−1ky − y′k2
whenever x ∈ ΓN (V) and y, y′ are in the operator-norm ball of radius R.
Lemma 7.4. Suppose that {µN} concentrates around a non-commutative law λ. Then {µN ∗
σt,N} concentrates around λ ⊞ σt for every t > 0.
Proof. Fix t. Let XN = (XN,1, . . . , XN,m) and YN = (YN,1, . . . , YN,m) be independent random
variables with the laws µN and σt,N respectively. Because the topology on the space Σm of
non-commutative laws is generated by non-commutative moments, it suffices to show that for
each non-commutative polynomial p and δ > 0,
(cid:3)
lim
N→∞
P (τN (p(XN + YN )) − λ ⊞ σt(p) ≥ δ) = 0.
Fix p and let d be its degree. By the previous lemma, there is a neighborhood V of λ and
a constant K such that for every x ∈ ΓN (V), the function y 7→ τN (p(x + y)) is K-Lipschitz
with respect to k·k2 on the operator-norm ball {y = (y1, . . . , ym) : kyjk ≤ 4t1/2}. By shrinking
V if necessary, we may also assume that on τN (q(x)) is uniformly bounded for every non-
commutative monomial q(x) of degree less than or equal to d.
function ψ : R → R such that ψ(z) = z for z ≤ 3t1/2 and ψ(z) ≤ 4t1/2. Then
Ψ : (y1, . . . , ym) 7→ (ψ(y1), . . . , ψ(ym)) is globally Lipschitz in k·k2 and it also maps MN (C)m
into the operator norm ball of radius 4t1/2 (which is the region where z 7→ τN (p(x, z)) was
assumed to K-Lipschitz with respect to k·k2 whenever x ∈ ΓN (V)). This implies that there is
some constant K ′ such that y 7→ τN (p(x, Ψ(y))) is K ′-Lipschitz for all x ∈ ΓN (V). Let
Choose a C∞
c
sa
αN (x) = E[τN (p(x + Ψ(YN ))]
βN (x) = E[τN (p(x + Y + N ))] = exp(tLN /2)[τ (p)](x)
β(x) = exp(tL/2)[τ (p)](x)
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
65
Therefore, by Theorem 2.10 applied to YN ,
x ∈ ΓN (V) =⇒ P (τN (p(x + Ψ(YN ))) − αN (x) ≥ δ/3) ≤ 2e−δ2N 2/18t(K ′)2
.
On the other hand, we know by standard tail estimates on the GUE (see Corollary 2.12) that
lim
N→∞
E[τN (q(YN )1kYN k≥3t1/2 ] = 0
for every non-commutative polynomial q. This implies that αN (x)− βN (x) → 0 uniformly for
x ∈ ΓN (V). On the other hand, by Lemma 3.21,
βN (x) = exp(tLN /2)[τ (p)](x) → exp(tL/2)[τ (p)](x) = β(x)
where the convergence occurs coefficient-wise. Now exp(tLN /2)[τ (p)] is a sum of products of
traces of non-commutative monomials q of degree ≤ d and for every such q, we know τN (q(x))
is uniformly bounded on ΓN (V) by our choice of V. Thus, coefficient-wise convergence of
βN → β implies uniform convergence for x ∈ ΓN (V). Therefore, for sufficiently large N we
have βN (x) − βN (x) ≤ δ/3 for x ∈ ΓN (V), and hence
P(cid:16)(cid:12)(cid:12)τN (p(XN + YN )) − τ (β(XN ))(cid:12)(cid:12) ≥ 2δ/3, XN ∈ ΓN (V), kYNk ≤ 3t1/2(cid:17) ≤ 2e−δ2N 2/18t(K ′)2
where we have applied the Fubini-Tonelli theorem for the product measure µN ⊗ σt,N . By our
concentration assumption,
,
P(cid:16)(cid:12)(cid:12)τN (β(XN )) − λ[β](cid:12)(cid:12) ≥ δ/3(cid:17) → 0,
and by Corollary 2.12 also P (kYkk ≥ 3t1/2) → 0. Altogether, we have
P (XN ∈ ΓN (V)) → 1,
P(cid:16)(cid:12)(cid:12)τN (p(XN + YN )) − λ[β](cid:12)(cid:12) ≥ δ(cid:17) → 0.
But note that λ[β] = λ[exp(tL/2)[τ (p)]] = λ ⊞ σt[p] by Lemma 3.23. Thus, the proof is
complete.
(cid:3)
Proof of Theorem 7.1. Let VN,t = RtVN be the potential associated to µN ∗ σt,N . Let us verify
that VN,t satisfies the assumptions (A) - (D) for every t > 0.
(A) This follows from Theorem 6.1 (1) because VN,t = RtVN , hence VN,t ∈ E(0, C).
(B) This follows from Lemma 7.4.
(C) This follows from tail bounds on the GUE (Corollary 2.12).
(D) This follows from Proposition 6.28.
Next, the fact that λ ∈ Σm,R0 follows from Proposition 5.4 (1) with n = 1.
Claim (1) of the theorem follows by applying Proposition 5.9 to µN ∗ σt,N with n = 1.
For claim (2), recall that by Lemma 5.6, equation (5.6),
(7.1)
1
N 2 h(µN ) +
m
2
log N =
1
0 (cid:18) m
2Z ∞
1 + t −
1
N 3I(µN ∗ σt,N )(cid:19) ds +
m
2
log 2πe.
Because N −3I(µN ) converges as N → ∞, there is some constant K with N −3I(µN ) ≤ K for
j ) > 0.
2 dµN (x) ≥ ma for large enough N . Thus, (5.4),
all N . Also, because of assumptions (B) and (C), we haveR kxk2
2 dµN (x) →Pm
Therefore, there is a constant a such thatR kxk2
t (cid:17) .
N 3I(µN ∗ σt,N ) ≤ min(cid:16)M,
we have for sufficiently large N that
m
a + t ≤
j=1 λj (X 2
1
m
66
DAVID JEKEL
Thus, we can apply the dominated convergence theorem to take the limit as N → ∞ inside the
integral on the right hand side of (7.1) and apply claim (1) to conclude that
N→∞(cid:18) 1
lim
N 2 h(µN ) +
m
2
log N(cid:19) → χ∗(λ).
On the left hand side of (7.1), we will apply Proposition 5.4 with n = 1. We may replace VN by
VN−VN (0) without changing µN (because the definition of µN includes the normalizing constant
ZN anyway). Then because {DVN} is asymptotically approximable by trace polynomials, we
know that {VN} is asymptotically approximable by trace polynomials (Lemma 3.29). Therefore,
the hypotheses of Proposition 5.4 are satisfied and so
χ(λ) = lim sup
N 2 h(µN ) +
N→∞ (cid:18) 1
m
2
log N(cid:19) = χ∗(λ)
and the same holds for χ(λ). Moreover, this holds for µN ∗ σt,N just as well as µN because
µN ∗ σt,N satisfies the same assumptions (A) - (D).
For claim (3), first fix N and let X be a random variable with law µN , and let Yt be an
independent Hermitian Brownian motion (here Yt ∼ σt,N ). Let Ξt = DVN,t(X + Yt), which is
the conjugate variable of X + Yt. Then
1
N 3I(µN ∗ σt,N ) = EkΞtk2
2
Suppose 0 ≤ s ≤ t ≤ T . Then X + Yt is the sum of the independent random variables X + Ys
and Yt − Ys, and thus Ξt = E[ΞsX + Yt] by Lemma 5.5. In other words, Ξt is the orthogonal
projection of DVN,s(X + Ys) onto the space of L2 random variables that are functions of X + Yt,
or in other words it is the closest function of X + Yt to Ξs in L2. This implies that
hkΞs − Ξtk2
2i ≤ EhkDVN,s(X + Ys) − DVN,s(X + Yt)k2
2i
≤ E(cid:20)
2(cid:21)
(1 + Cs)2 kYs − Ytk2
C2
C2
=
(1 + Cs)2 m(t − s)
using the fact that VN,s ∈ E(0, C(1 + Ct)−1) and hence DVN,s is C(1 + Cs)−1-Lipschitz. Since
Ξt is the orthogonal projection of Ξs onto this subspace, we know Ξs − Ξt is orthogonal to Ξt
and hence
EhkΞsk2
2i − EhkΞtk2
2i = EhkΞs − Ξtk2
2i .
Overall,
1
N 3I(µN ∗ σs,N ) −
1
N 3I(µN ∗ σt,N ) ≤
0 ≤
C2
(1 + Cs)2 m(t − s).
This immediately proves that t 7→ N −3I(µN ∗ σt,N ) is decreasing function of t, it is Lipschitz,
and the absolute value of the derivative is bounded by C2m/(1 + Ct)2. The same holds for
Φ∗(λ ⊞ σt) by taking the limit as N → ∞.
(cid:3)
8. Free Gibbs Laws
In the situation of Theorem 4.1, we want to interpret the law λ as the free Gibbs state for a
potential which is the limit of the VN 's. To this end, we will define a non-commutative function
space where each point is a limit of functions on MN (C)m
sa. We will then give several charac-
terizations of the closure of trace polynomials in this space, as well as the class of potentials to
which our previous results apply.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
67
8.1. Asymptotic Approximation and Function Spaces. Let Y• = {YN} be a sequence of
normed vector spaces. We define a (possibly infinite) semi-norm on sequences φ• = {φN} of
functions MN (C)m
sa → YN by
kφ•kR,Y•
= lim sup
N→∞
sup
kxk≤RkφN (x)kYN
.
Let Fm(Y•) be the vector space
{φ• : kφ•kR,Y•
< +∞ for all R}/{φ• : kφ•kR,Y•
= 0 for all R}.
For a sequence φ•, we denote its equivalence class by [φ•].
We equip Fm(Y•) with the topology generated by the seminorms k·kR,Y•
given by the metric
, or equivalently
(8.1)
dFm(Y•)(φ•, ψ•) =
1
2n min(kφ• − ψ•kn,Y , 1).
∞Xn=1
The vector space of scalar-valued trace polynomials TrP0
Note that Fm(Y•) is a complete metric space in this metric and is a locally convex topological
vector space.
m := Fm(C) by
the map that sends a trace polynomial to the corresponding sequence of functions it defines on
MN (C)m
sa. A sequence φ• is asymptotically approximable by trace polynomials if and only if
[φ•] is in the closure of TrP0
m embeds into F 0
m in F 0
m, which we will denote by T 0
m.
The spaces T 0
Similarly, let M•(C)m be the sequence {MN (C)m} equipped with k·k2. The vector space
TrP1
m := Fm(M•(C)). A sequence φ• of functions MN (C)m
m embeds into F 1
sa → MN (C)sa is
asymptotically approximable by trace polynomials if and only if [φ•] is in the closure of TrP1
m,
which we denote by T 1
m.
m can be viewed as non-commutative function spaces through the
following alternative characterization. Here R denotes the hyperfinite II1 factor and Rω denotes
its ultrapower. For background, see [2, §1.6 and §5.4] or [8, p. 5 - 7].
Lemma 8.1. Let f ∈ TrP0
m. Then we have
sup
(8.2)
f (x) = sup
f (x) = sup
m and T 1
f (x).
lim sup
N→∞
sup
N
x∈MN (C)m
sa
kxk∞≤R
x∈MN (C)m
sa
kxk∞≤R
x∈(Rω
sa)m
kxk∞≤R
If we denote the common value by kfkT 0
topology on TrP0
metric. The same result holds for T 1
(8.3)
m with the metric given as in (8.1), and T 0
m using the seminorm
sup
sup
m,R, then this family of seminorms defines a metrizable
m in this
m is the completion of TrP0
lim sup
N→∞
x∈MN (C)m
sa
kxk∞≤R
kf (x)k2 = sup
N
x∈MN (C)m
sa
kxk≤R
kf (x)k2 = sup
x∈(Rω
sa)m
kxk∞≤R
kf (x)k2.
Proof. Fix f and let A, B, and C be the three quantities in (8.2) from left to right.
It is
clear that A ≤ B. Moreover, B ≤ C because there is an isometric trace-preserving embedding
of MN (C) into Rω. To show that C ≤ A, pick x ∈ (Rω
sa)m with kxk ≤ R. Then there
exists xn ∈ Rm
sa with kxnk ≤ R and x = limn→ω xn. For each n, we can choose an Nn, an
embedding MNn(C) → R and a yn ∈ MNn(C) such that kynk ≤ R and kxn − ynk2 ≤ 1/2n and
limn→∞ Nn = +∞. Then x = limn→ω yn and f (x) = limn→ω f (yn) ≤ A. This shows that
the three seminorms in (8.2) are equal, and the other claims follow because these seminorms
are the same as the seminorms for F 0
m.
(cid:3)
From this point of view, any f ∈ T 0
m has a canonical sequence that represents its equivalence
m, constructed as follows. If we write f as the limit of a sequence of trace polynomials
class in F 0
68
DAVID JEKEL
f (k), then f(k)MN (C)m
independent of the approximating sequence f (k). We can therefore define fMN (C)m
limit.
converges locally uniformly on MN (C)m
sa as k → ∞ and the limit is
sa to be this
sa
Similarly, f defines a function on (Rω
sa)m. Moreover, if (M, τ ) is a tracial von Neumann
algebra and there is a trace-preserving embedding ι : M → Rω, then we may define fM = f ◦ι.
It is easy to see that this is independent of the choice of trace-preserving embedding if f is a
trace polynomial, and this holds for general f ∈ T 0
m or T 1
m by density of trace polynomials. In
this sense, T 0
m represent spaces of universal scalar- or operator-valued functions that
can be applied to self-adjoint operators in every Rω
sa-embeddable tracial von Neumann algebra.
m and T 1
In the scalar-valued case, we have yet another characterization of T 0
m:
Lemma 8.2. Let Σm,bdd =SR>0 Σm,R. Let C(Σm,bdd) be the space of functions g : Σm,bdd → C
such that g ∈ C(Σm,R) for all R, equipped with the family of seminorms k·kC(Σm,R). Then T 0
is isomorphic to C(Σm,bdd) as a topological vector space.
m
Proof. For a scalar-valued trace polynomial f , the value f (x) only depends on the law of x, so
that f (x) = g(λx) for some function g : Σm → R such that g ∈ C(Σm,R) for all R, and we have
kfkT 0
m,R = kgkC(Σm,R).
Passing to the completion with respect to the metric defined as in (8.1), we have a map ι :
T 0
m → C(Σm,bdd) which is an isomorphism onto its image. To show that ι is surjective, note
the algebra of trace polynomials is self-adjoint and separates points in Σm,R, and hence by the
Stone-Weierstrass theorem, trace polynomials are dense in C(Σm,R) for every R. Therefore, if
g ∈ C(Σm,R), we can choose a trace polynomial g(k)(λx) = f (k)(x) such that kg−g(k)kC(Σm,k) ≤
1/2k. Then f (k) converges to some f in T 0
8.2. Convex Differentiable Functions. Now we are ready to characterize the type of convex
functions which occur in Theorem 7.1. First of all, we let T 0,1
m be the completion of the trace
polynomials with respect to the metric
m, and we have ι(f ) = g.
(cid:3)
d(f, g) =
1
2n [min(1,kf − gkT 0
m,n) + min(1,kDf − Dgk(T 1
m)m,n)].
∞Xn=1
Observe that if f ∈ T 0,1
as k → ∞, then Df (k) converges in (T 1
approximating sequence. We denote this limit by Df .
m and f (k) is a sequence of trace polynomials converging to f in T 0,1
m
m)m and the limit is independent of the choice of
Remark 8.3. If f and f (k) are as above, then since Df (k) is a tuple of trace polynomials, it is
continuous on the operator norm ball {y ∈ MN (C)m
sa : kyk∞ ≤ R} with a modulus of continuity
that only depends on R and does not depend on N . Because Df (k) → Df uniformly on the
operator-norm ball (with rate of convergence independent of N ), then Df is also continuous
on this operator-norm ball with modulus of continuity independent of N .
It follows that for every x, y ∈ MN (C)m
sa with kxk,kyk ≤ R, we have
f (y) − f (x) = hDf (x), y − xi2 + o(ky − xk2),
where the error estimate only depends on R and not on N . In particular, this shows Df is
uniquely determined by f . Also, it shows that DfMN (C)m
sa is equal to the normalized gradient
sa ∼= RmN 2
in the ordinary sense of functions on MN (C)m
of fMN (C)m
Lemma 8.4. Let f ∈ T 0,1
m be real-valued. The following are equivalent:
sa
.
(1) The function fMN (C)m
sa is convex for every N .
(2) The function f is convex as a function on (Rω
sa)m.
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
69
(3) There exists a sequence of differentiable convex functions VN : MN (C)m
sa → R such that
[V•] = f and [DV•] = Df . (Here DV• denotes the sequence (DVN )N ∈N, where D is the
normalized gradient understood in the standard sense of calculus.)
Proof. The equivalence between (1) and (2) follows from similar argument to the proof of
Lemma 8.1.
(1) =⇒ (3) because we can take VN = fMN (C)m
sa.
(3) =⇒ (1). Fix N . To prove that fMN (C)m
Df (y), x − yi2 ≥ 0 for every x, y ∈ MN (C)m
MN k(C)m
sa. Then as k → ∞,
sa is convex, it suffices to show that hDf (x) −
sa. For k ∈ N, consider x ⊗ Ik and y ⊗ Ik in
hDf (x) − Df (y), x − yi2 = hDf (x ⊗ Ik) − Df (y ⊗ Ik), x ⊗ Ik − y ⊗ Oki2;
meanwhile, if R = max(kxk,kyk), then since DVN − Df → 0 in k·k2 uniformly on the operator
norm ball of radius R, we have as k → ∞ that
hDf (x⊗Ik)−Df (y⊗Ik), x⊗Ik −y⊗Iki2−hDVN k(x⊗Ik)−DVN k(y⊗Ik), x⊗Ik−y⊗Iki2 → 0.
Because VN k is convex, the second inner product is ≥ 0 and therefore hDf (x)−Df (y), x−yi2 ≥
0.
(cid:3)
Let Em(c, C)0,1 denote the class of V ∈ T 0,1
2 is concave. If 0 < c < C and if V ∈ Em(c, C)0,1, if VN = V MN (C)m
2 is convex and
V (x)− (C/2)kxk2
, then the
sequence of normalized gradients DVN is asymptotically approximable by trace polynomials. If
we let µN be the corresponding measure, then Theorem 4.1 (the hypothesis (4.1) being trivially
satisfied by unitary invariance) implies that µN concentrates around a non-commutative law
λV , which we will call the free Gibbs state for the potential V .
m such that V (x) − (c/2)kxk2
sa
Furthermore, the free Gibbs state λV is independent of the choice of representative sequence
in the following sense. Let µN be the measure on MN (C)m
sa given by the potential VN =
V MN (C)m
sa. Let WN be another sequence of potentials satisfying the hypotheses of Theorem
4.1 such that [W•] = V in T 0,1
m , and let νN be the sequence of random matrix measures given
by WN . By Theorem 4.1 νN concentrates around some non-commutative law λ. We claim that
λ = λV . To prove this, consider the sequence VN which equals VN for odd N and WN for even
m , which means that {D VN}N ∈N is asymptotically approximable by
N . Then [ V•] = V in T 0,1
trace polynomials. Therefore,
N→∞Z τN (p) dµN = lim
N→∞Z τN (p) dνN = λ(p).
λV (p) = lim
N even
N odd
In fact, Lemma 8.4 implies that the non-commutative laws λ which occur as limits in Theorem
4.1 are precisely the free Gibbs laws for potentials V ∈ Em(c, C)0,1. In particular, Theorem 7.1
implies that χ = χ = χ∗ for every such law.
Remark 8.5. We have not proved that the law λV is uniquely characterized by the Schwinger-
Dyson equation λ[DV (X)f (X)] = λ⊗λ[Df (X)], although something like this is implied by [13].
One could hope to prove this by letting the semigroup T V
t act on an abstract space of Lipschitz
functions which is the completion of trace polynomials (where the metric now allows x to come
from any tracial von Neumann algebra rather than only the Rω-embeddable algebras). We
would want to show that if λ satisfies the Schwinger-Dyson equation, then λ(T V
t u) = λ(u), but
to justify the computation, we need to show more regularity of T V
t u than we have done in this
paper. In the SDE approach as well, the proof that λV is characterized by Schwinger-Dyson is
subtle when we do not assume more regularity for V (see [14], [13]).
70
DAVID JEKEL
8.3. Examples of Convex Potentials. A natural class of examples of functions in Em(c, C)0,1
are those of the form
where ǫ is a small positive parameter,
V (x) =
1
2kxk2
2 + ǫf (u)
u = (u1, . . . , um),
uj =
xj + 4i
xj − 4i
.
and f is a real-valued trace polynomial in u and u∗. Computations similar to those of §3.2
show that the normalized Hessian of Jac(Df (u(x))) with respect to x is bounded uniformly in
N . Therefore, V ∈ Em(1/2, 3/2)0,1 for sufficiently small ǫ. Similar examples are described in
the introduction of [13]. More generally, we can replace the trace polynomial f (u) by a power
series where the individual terms are trace monomials in u.
The class Em(c, C)0,1 does not include trace polynomials in x because if g is a trace polyno-
mial of degree ≥ 3, then we cannot have g(x) convex and g(x) − (C/2)kxk2
2 concave (globally).
However, if we consider a potential which is a small perturbation of a quadratic (as considered
in [20], [22]), we can fix this problem by introducing an operator-norm cut-off as follows.
Let f be a scalar-valued trace polynomial and let us denote
V (ǫ)(x) = kxk2
2 + ǫf (x).
function such that φ(t) = t for t ≤ R and φ(t) = 0 for t ≥ 2R. Let
sa be given by ΦN (x) = (φ(x1), . . . , φ(xm)).
V (ǫ)
N (x) = kxk2
2 + ǫfN (ΦN (x)).
(8.4)
Let φ : R → R be a C∞
Φ : MN (C)m
sa → MN (C)m
c
(8.5)
We will prove the following.
Proposition 8.6. Let V (ǫ)
have [ V (ǫ)
•
N be given as above. Then [ V (ǫ)
•
] ∈ T 0,1
m . Moreover, given δ > 0, we
] ∈ Em(1 − δ, 1 + δ)0,1 for sufficiently small ǫ (depending on f , R, and δ).
As a consequence, we will deduce the following result about measures defined by V (ǫ) re-
stricted to an operator-norm ball (without the smooth cut-off Φ).
Proposition 8.7. Let 2 < R′ < R, let f be a trace polynomial, and let V (ǫ) be as in (8.4). Let
dµ(ǫ)
N (x) =
1
ZN
exp(−N 2V (ǫ)
N (x))1kxk≤R dx.
For sufficiently small ǫ (depending on f , R, and R′), we have the following. The measure µ(ǫ)
N
exhibits exponential concentration around a non-commutative law λ(ǫ) ∈ Σm,R′. If X ∈ (M, τ )
is a non-commutative m-tuple realizing the law λ(ǫ), then the conjugate variable is given by
DV (ǫ)(X). Moreover, we have
χ(λ(ǫ)) = χ(λ(ǫ)) = χ∗(λ(ǫ)) = lim
N→∞(cid:18) 1
N 2 h(µ(ǫ)
N ) +
m
2
log N(cid:19) .
m or T 0,1
To fix notation for the remainder of this section, functions without a subscript, such as f ,
will denote elements of T 0
m , and Df will denote the "gradient" defined in the abstract
space T 0,1
m as the limit of the "gradients" of trace polynomials approximating f . However, fN
will denote fMN (C)m
, and DfN will denote the normalized gradient N∇fN defined in the usual
sense of calculus with respect to h·,·i2. Moreover, HfN = Jac(DfN ) will denote the Hessian of
fN with respect to h·,·i2.
In order to prove Proposition 8.6, we must understand D[fN ◦ ΦN ] and H[fN ◦ ΦN ]. To this
end, we recall some results of Peller [28] on non-commutative derivatives of φ(x), where φ is a
smooth function on the real line.
sa
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
71
For a polynomial φ in one variable, the non-commutative derivative Dφ ∈ ChXi ⊗ ChXi
defined by Definition 3.6 can be written as the difference quotient
Dφ(s, t) =
φ(s) − φ(t)
,
s − t
(8.6)
(8.7)
(8.8)
where we view ChXi⊗ ChXi as a subset of functions on R2 with the variables s and t. However,
the above difference quotient makes sense whenver φ : R → C is smooth. Thus, it defines an
extension of D to continuously differentiable functions φ of one variable.
Similarly, if φ is a polynomial, then the higher order non-commutative derivatives Dnφ can be
viewed as functions of n + 1 variables, which are obtained through iterated difference quotients
and thus their definition can be extended to smooth functions φ. (However, beware that we
have not defined Dn
j φ if φ is a non-polynomial function of multiple variables.)
by R, one seeks to control the norm of Dφ in the projective tensor product L∞[−R, R]b⊗L∞[−R, R].
If φ is a polynomial, then to estimate φ(X)−φ(Y ) for operators X and Y with norm bounded
Similarly, if φ is a smooth function and φ(X) and φ(Y ) are defined through functional calculus,
one can estimate the operator norm kφ(X) − φ(Y )k by representing φ as an integral of simpler
functions (e.g. by Fourier analysis) whose non-commutative derivatives are easier to analyze.
In this case, it is convenient to write Dφ as an integral rather than a sum of simple tensors.
We thus consider the integral projective tensor powers of the space of bounded Borel functions
B(R). The integral projective tensor product B(R) b⊗in consists of Borel functions G on Rn which
admit a representation
for some measure space (Ω, µ) such that
G1(x1, ω) . . . Gn(xn, ω) dµ(ω)
G(x1, . . . , xn) =ZΩ
ZΩkG1(·, ω)kB(R) . . .kG1(·, ω)kB(R) dµ(ω) < +∞
and we define kGkB(R) b⊗in to be the infimum of (8.7) over all representations (8.6).
Given G ∈ B(R)bωin, bounded self-adjoint operators x0, . . . , xn and bounded operators y1,
. . . , yn, we define
G(x0, . . . , xn)#(y1 ⊗ ··· ⊗ yn) =ZΩ
G0(x0, ω)y1G1(x1, ω) . . . ynGn(xn, ω) dµ(ω),
where G0, . . . , Gn satisfy (8.6). This is well-defined by [28, Lemma 3.1]. If the xj's and yj's are
elements of a tracial von Neumann algebra (M, τ ), we have by the non-commutative Holder's
inequality (see §2.3) that if 1/α = 1/α1 + ··· + 1/αn, then
kG(x0, . . . , xn)#(y1 ⊗ ··· ⊗ yn)kα ≤ kGkB(R)
(8.9)
b⊗i(n+1)ky1kα1
. . .kynkαn
,
Moreover, we have the following bounds on the non-commutative derivatives of φ as a corollary
of the results of [28].
Proposition 8.8. There exists a constant Kn such that for all φ ∈ C∞
(8.10)
c (R),
Proof. As in [28, §2], choose w ∈ C∞
for ξ > 0. Let Wk and W #
denotes the Fourier transform. It is shown in [28, Theorem 5.5] that
c
kDnφkB(R) b⊗i (n+1) ≤ KnZR bφ(ξ)ξn dξ.
such that 0 ≤ w ≤ χ[−1/2,2] and Pk∈Z w(2−kξ) = 1
k be given by cWk(ξ) = w(2−nξ) and cW #
k (ξ) = w(−2−kx) whereb·
k ∗ φkL∞(R)(cid:17) .
2nk(cid:16)kWk ∗ φkL∞(R) + kW #
kDnφkB(R) b⊗i(n+1) ≤ KnXk∈Z
72
DAVID JEKEL
This can be estimated by the right hand side of (8.10) (for a possibly different constant) by a
standard Fourier analysis computation.
(cid:3)
Proof of Proposition 8.6. Recall that V (ǫ)
2 + ǫfN ◦ ΦN . Thus, to show that the
sequence V (ǫ)
0 , it suffices to prove this for fN ◦ ΦN . To this end, it is
sufficient to show that for each r > 0, there is a sequence of trace polynomials {g(k)}k∈N such
that
N defines an element of T m
N (x) = 1
2kxk2
and
lim
k→∞
sup
N ∈N
sup
x∈MN (C)m
sa:kxk∞≤r g(k)(x) − fN ◦ ΦN (x) = 0
lim
k→∞
sup
N ∈N
sup
x∈MN (C)m
sa:kxk∞≤rkDg(k)(x) − D[fN ◦ ΦN (x)]k2
Fix r > 0. By standard approximation techniques, there exist Schwarz functions φ(k) : R → R
such that φ(k)[−r,r] is a polynomial and φ(k) → φ in the Schwarz space as k → ∞. By
Proposition 8.8, we have Dnφ(k) → Dnφ in B(R) b⊗i(n+1) as k → ∞ for every n.
N (x1, . . . , xm) = (φ(k)(x1), . . . , φ(k)(xm)). Then fN ◦ Φ(k)
mial g(k) on {kxk∞ ≤ r}. Because of the spectral mapping theorem,
N is given by a trace polyno-
Let Φ(k)
sup
kxk≤rkΦ(k)
N (x) − ΦN (x)k∞ ≤ m sup
t∈[−r,r]φ(k)(t) − φ(t)
which is independent of N and vanishes as k → ∞. Thus, our trace polynomials g(k) approxi-
mate fN ◦ ΦN uniformly on the operator norm ball {x : kxk∞ ≤ r}.
Next, we must show that Dg(k) approximates D[fN ◦ ΦN ] uniformly in k·k2 on the operator
norm ball {kxk∞ ≤ r}. By the chain rule, we have
Dj[fN ◦ ΦN ] = Jacj(ΦN )t[DjfN ],
where Dj and Jacj are the normalized gradient and Jacobian with respect to the variable
xj ∈ MN (C)sa. Now
Jacj(ΦN )(x)y = Dφ(xj )#y.
Now Dφ viewed as an element of the tensor product C[X] ⊗ C[X] is is invariant under the flip
map that switches the order of the tensorands; this is because Dφ is represented as a difference
quotient for one-variable functions. Flip invariance implies that
τN [(Dφ(xj )#y)z] = τN [y(Dφ(xj )#z)],
which means that the operator Jacj(ΦN )(x) on MN (C)sa is self-adjoint. Hence,
Dj[fN ◦ ΦN ](x) = Jacj(ΦN (x))[Dj fN ](x) = Dφ(xj )#DjfN (ΦN (x)).
This function is given by a trace polynomial on {kxk∞ ≤ r}, and specifically it must equal
Djg(k) because Djg(k) is uniquely determined as a trace polynomial by the fact that it is the
gradient of g(k)MN (C)m
sa for every N . Moreover, for kxk∞ ≤ r, we have
Dφk(xj )#Djf (Φk(x)) = Dφk(xj)#Dj f (Φ(x)) + Dφk(xj )#[Djf (Φk(x)) − Djf (Φ(x))].
The first term converges to Dφ(xj )#Djf (Φ(x)) in k·k2 uniformly on {kxk∞ ≤ r} using (8.9)
with estimates independent of N . Similarly, because the images of Φk and Φ are contained in
an operator norm ball and Djf is K-Lipschitz in k·k2 on this ball for some K > 0, we have
Djf (Φk(x)) − Djf (Φ(x)) → 0 uniformly. This in turn implies that the second term goes to
sequence of trace polynomials g(k) such that gk → f ◦ Φ and Dg(k) → D(f ◦ Φ) uniformly on
{kxk∞ ≤ r}. This means that f ◦ Φ ∈ T 1,0
m .
zero because Dφk(xj) is uniformly bounded in B(R)b⊗iB(R). Thus, for every r > 0, there is a
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
73
It follows that the sequence V (ǫ)
N defines a function in T 0,1
m for every ǫ. It remains to show that
this function is in Em(1− δ, 1 + δ)0,1 for sufficiently small ǫ. To this end, it suffices to show that
fN ◦ ΦN defines a function in Em(−a, a)0,1 for some real a > 0. Thus, we only need to obtain
some uniform upper and lower bounds on the operator norm of H[fN ◦ΦN ] that are independent
of N . However, this is equivalent to showing that Dj(fN ◦ ΦN ) = DφN (xj )#DjfN (Φ+ N (x)) is
we see that
Lipschitz in k·k2 for each j (uniformly in N ). Because D2φ is bounded in B(R)b⊗iB(R)b⊗iB(R),
kDφ(xj )#y − Dφ(x′
j )#yk2 ≤ Kkxj − x′
jk2kyk2
for some constant K. Together with the fact that DjfN (ΦN (x)) is Lipschitz in k·k2, this implies
that Dj(fN ◦ ΦN ) is Lipschitz in k·k2 as desired.
sa given by the potential V . Let
Proof of Proposition 8.7. Let µN be the measure on MN (C)m
δ be a number in (0, 1) to be chosen later. By Proposition 8.6, we have that V (ǫ) ∈ Em(1 −
δ, 1 + δ)0,1 for sufficiently small ǫ. By Theorem 4.1, the laws µN concentrate around a non-
commutative law λ. Furthermore, in Theorem 4.1 (1), we can take M = 0 and c = 1 − δ and
C = 1 + δ, so that
(cid:3)
lim sup
N→∞
RN ≤
2
(1 − δ)1/2 + kD V (0)k2
1 − δ
δ
+
(1 − δ)3/2 .
Note that D V (0) = DV (0) = ǫDf (0) is a scalar multiple of the identity matrix since f is a
trace polynomial. Because R′ > 2, we may choose δ sufficiently small that
2
δ
(1 − δ)1/2 +
(1 − δ)3/2 < R′.
Then by choosing ǫ (and hence kD V (0)k2) sufficiently small, we can arrange that R∗ =
lim supN→∞ RN < R′. This implies that the measures µN concentrate on the ball {kxk∞ ≤ R′}.
For kxk∞ ≤ R, we have V (x) = V (x), and therefore µN is the (normalized) restriction of µN
to {kxk∞ ≤ R}. It follows that µN concentrates around the law λ as well.
If X ∈ (M, τ ) realizes the law λ, then kXk∞ ≤ R′ since λ ∈ Σm,R∗ ⊆ Σm,R′ by Theorem 4.1
(2). Moreover, by Proposition 5.9, the conjugate variables for λ are given by D V (X) = DV (X).
Moreover, by Theorem 7.1 applied to µN , we have
χ(λ) = χ(λ) = χ∗(λ) = lim
N→∞(cid:18) 1
N 2 h(µN ) +
m
2
log N(cid:19) .
In the last equality, we can replace µN by µN as in the proof of Proposition 5.4 because µN is
concentrated on {kxk∞ ≤ R′}.
(cid:3)
Remark 8.9. The approach given here probably does not give the optimal range of ǫ for Propo-
sition 8.7. To get the best result, one would want a more direct way to extend the potential
V : {kxk∞ ≤ R} → R to a potential V defined everywhere. This leads us to ask the following
question.
Suppose that V is a real-valued function in the closure of trace polynomials with respect
to the norm kfkT 0
m,R, and hence V defines a function {x : kxk∞ ≤ R} → R for
x ∈ MN (C)m
2 is concave on {kxk ≤ R},
then does V extend to a potential V ∈ Em(c, C)0,1? What if we allow V to have slightly worse
constants c and C?
The construction of extensions that preserve the convexity properties is not difficult, but it
is less obvious how to construct an extension that one can verify preserves the approximability
by trace polynomials.
2 is convex and V (x) − (C/2)kxk2
sa. If V (x) − (c/2)kxk2
m,R + kDfkT 1
74
DAVID JEKEL
References
[1] G. W. Anderson, A. Guionnet, and O. Zeitouni, An Introduction to Random Ma-
trices, Cambridge Studies in Advanced Mathematics, Cambridge University Press, 2009.
[2] C. Antharaman and S. Popa, An Introduction to II1 Factors, 2017. preprint available
at http://www.math.ucla.edu/ popa/Books/IIun-v10.pdf.
[3] V. Barbu, A semigroup approach to Hamilton-Jacobi equations in Hilbert space, in Semi-
groups, Theory and Applications, H. Brezis, M. G. Crandall, and F. Kappel, eds., Longman
Scientific and Technical, Essex, 1986.
[4] A. R. Barron, Entropy and the central limit theorem, Ann. Prob., 14 (1996), pp. 336 -- 342.
[5] P. Biane, Segal-Bargmann transform, functional calculus on matrix spaces and the theory
of semi-circular and circular systems, Journal of Functional Analysis, 144 (1997), pp. 232
-- 286.
[6] P. Biane, M. Capitaine, and A. Guionnet, Large deviation bounds for matrix Brow-
nian motion, Invent. Math., 152 (2003), pp. 433 -- 459.
[7] H. J. Brascamp and E. H. Lieb, On extensions of the Brunn-Minkowski and Pr´ekopa-
Leindler theorems, including inequalities for log concave functions, and with an application
to the diffusion equation, Journal of Functional Analysis, 22 (1976), pp. 366 -- 389.
[8] V. Capraro, A survey on Connes' embedding conjecture, arXiv e-prints, (2010).
[9] G. C´ebron, Free convolution operators and free Hall transform, Journal of Functional
Analysis, 265 (2013), pp. 2645 -- 2708.
[10] M. G. Crandall, H. Ishii, and P.-L. Lions, User's guide to viscosity solutions of
second order partial differential equations, Bull. Amer. Math. Soc., 27 (1992), pp. 1 -- 67.
[11] M. G. Crandall and P.-L. Lions, Viscosity solutions of hamilton-jacobi equations,
Trans. Am. Math. Soc., 277 (1983), pp. 1 -- 42.
[12] R. C. da Silva, Lecture notes on non-commutative Lp-spaces, arXiv e-prints, (2018).
[13] Y. Dabrowksi, A Laplace principle for Hermitian Brownian motion and free entropy I:
the convex functional case, arXiv e-prints, (2017).
[14] Y. Dabrowski, A non-commutative path space approach to stationary free stochastic dif-
ferential equations, arXiv e-prints, (2010).
[15] Y. Dabrowski, A. Guionnet, and D. Shlyakhtenko, Free transport for convex po-
tentials, arXiv e-prints, (2016).
[16] B. K. Driver, B. C. Hall, and T. Kemp, The large-N limit of the Segal-Bargmann
transform on UN , Journal of Functional Analysis, 265 (2013), pp. 2585 -- 2644.
[17] T. C. Duvillard and A. Guionnet, Large deviations upper bounds for the laws
of matrix-valued processes and non-communicative entropies, Ann. Probab., 29 (2001),
pp. 1205 -- 1261.
[18] I. Ekeland and J. M. Lasry, On the number of periodic trajectories for a hamiltonian
flow on a convex energy surface, Ann. Math., 112 (1980), pp. 283 -- 319.
[19] L. C. Evans, Partial Differential Equations, vol. 19 of Graduate Studies in Mathematics,
American Mathematical Society, Providence, Rhode Island, 2 ed., 2010.
[20] A. Guionnet and E. Maurel-Segala, Combinatorial aspects of random matrix models,
Alea, 1 (2006), pp. 241 -- 279.
[21] A. Guionnet and D. Shlyakhtenko, Free diffusions and matrix models with strictly
convex interaction, Geometric and Functional Analysis, 18 (2009), pp. 1875 -- 1916.
, Free monotone transport, Inventiones mathematicae, 197 (2014), pp. 613 -- 661.
[22]
[23] G. Harg´e, A convex/log-concave correlation inequality for gaussian measure and an ap-
plication to abstract wiener spaces, Probab. Theory Relat. Fields, 130 (2004), pp. 415 -- 440.
[24] T. Kato, Trotter's product formula for an arbitrary pair of self-adjoint contraction semi-
groups, in Topics in Functional Analysis, Essays Dedicated to M.G. Krein, I. Gohberg and
AN ELEMENTARY APPROACH TO FREE ENTROPY THEORY FOR CONVEX POTENTIALS
75
M. Kac, eds., vol. 3 of Advances in Mathematics, Supplementary Studies, Academic Press,
1978.
[25] J. M. Lasry and P. L. Lions, A remark on regularization in Hilbert spaces, Israel Journal
of Mathematics, 55 (1986), pp. 257 -- 266.
[26] M. Ledoux, A heat semigroup approach to concentration on the sphere and on a compact
Riemannian manifold, Geometric & Functional Analysis, 2 (1992), pp. 221 -- 224.
[27] A. Nica and R. Speicher, Lectures on the Combinatorics of Free Probability, vol. 335
of London Mathematical Society Lecture Note Series, Cambridge University Press, 2006.
[28] V. V. Peller, Multiple operator integrals and higher operator derivatives, Journal of
Functional Analysis, 233 (2006), pp. 515 -- 544.
[29] G. Pisier and Q. Xu, Non-commutative Lp-spaces, in Handbook of the geometry of
Banach spaces, W. B. Johhson and J. Lindenstrauss, eds., vol. 2, Elsevier Science B.V.,
2003, pp. 1459 -- 1517.
[30] E. M. Rains, Combinatorial properties of Brownian motion on the compact classical
groups, Journal of Theoretical Probability, 10 (1997), pp. 659 -- 679.
[31] Y. P. Razmyslov, Trace identities of full matrix algebras over a field of characteristic
zero, Mathematics of the USSR-Izvestiya, 8 (1974), p. 727.
[32]
, Trace identities and central polynomials in the matrix superalgebras Mn,k, Mathe-
matics of the USSR-Sbornik, 56 (1987), p. 187.
[33] B. Simon, Functional Integration and Quantum Physics, Academic Press, 1979.
[34]
, Trace Ideals and Their Applications, Mathematical Surveys and Monographs, Amer-
ican Mathematical Society, Providence, RI, 2 ed., 2005.
[35] H. F. Trotter, On the product of semi-groups of operators, Proc. Amer. Soc., 10 (1959),
pp. 545 -- 551.
[36] D. Voiculescu, Addition of certain non-commuting random variables, Journal of Func-
[37]
[38]
[39]
[40]
tional Analysis, 66 (1986), pp. 323 -- 346.
, Limit laws for random matrices and free products, Inventiones mathematicae, 104
(1991), pp. 201 -- 220.
, The analogues of entropy and of Fisher's information in free probability, I, Comm.
Math. Phys., 155 (1993), pp. 71 -- 92.
, The analogues of entropy and of Fisher's information in free probability, II, Invent.
Math., 118 (1994), pp. 411 -- 440.
, The analogues of entropy and of Fisher's information in free probability V, Invent.
Math., 132 (1998), pp. 189 -- 227.
, Free entropy, Bulletin of the London Mathematical Society, 34 (2002), pp. 257 -- 278.
[41]
[42] D. Voiculescu, K. J. Dykema, and A. Nica, Free Random Variables, vol. 1 of CRM
Monograph Series, American Mathematical Society, 1992.
Department of Mathematics, UCLA, Los Angeles, CA 90095
E-mail address: [email protected]
URL: www.math.ucla.edu/∼davidjekel/
|
1107.5205 | 1 | 1107 | 2011-07-26T13:06:46 | Arveson dichotomy and essential fractality | [
"math.OA"
] | The notions of fractal and essentially fractal algebras of approximation sequences and of the Arveson dichotomy have proved extremely useful for several spectral approximation problems. The purpose of this short note is threefold: to present a short new proof of the fractal restriction theorem, to relate essential fractality with Arveson dichotomy, and to derive a restriction theorem for essential fractality. | math.OA | math |
Arveson dichotomy and essential fractality
Steffen Roch
Dedicated to Vladimir S. Rabinovich on the occasion of his 70th
birthday
Abstract
The notions of fractal and essentially fractal algebras of approximation
sequences and of the Arveson dichotomy have proved extremely useful for
several spectral approximation problems. The purpose of this short note
is threefold: to present a short new proof of the fractal restriction theo-
rem, to relate essential fractality with Arveson dichotomy, and to derive a
restriction theorem for essential fractality.
Keywords: Arveson dichotomy, essential spectral approximation, essential frac-
tality, essential fractal restriction of approximation sequences
2010 AMS-MSC: 65J10, 46L99, 47N40
1 Preliminaries
Let H be an infinite dimensional separable Hilbert space. We denote by L(H)
the C ∗-algebra of the bounded linear operators and by K(H) the ideal of the
compact operators on H.
A sequence P = (Pn)n≥1 of orthogonal projections of finite rank which con-
verge strongly to the identity operator on H is called a filtration on H. Given
a filtration P, let F P stand for the set of all sequences A = (An) of opera-
tors An : im Pn → im Pn such that the sequence (AnPn) converges strongly to
an operator W P(A) ∈ L(H). Since every sequence in F P is bounded by the
Banach-Steinhaus theorem, one can introduce pointwise defined operations
(An) + (Bn) := (An + Bn),
(An)(Bn) := (AnBn),
(An)∗ := (A∗
n)
(1)
and the supremum norm k(An)kF := supn kAnk, which make F P to a unital C ∗-
algebra and W P : F P → L(H) to a unital ∗-homomorphism. This homomorphism
is also known as the consistency map associated with the filtration P.
Set δ(n) := rank Pn := dim im Pn < ∞ for every n and choose an orthonormal
basis in each of the spaces im Pn. Every operator An ∈ L(im Pn) can be identified
1
with its matrix representation with respect to the chosen basis and, thus, with
an element of the C ∗-algebra Cδ(n)×δ(n) of all δ(n) × δ(n) matrices with complex
entries. The choice of a basis in each space im Pn makes F P to a special instance
of an algebra of matrix sequences in the following sense. Given a sequence δ
of positive integers, we let F δ stand for the set of all bounded sequences (An)
of matrices An ∈ Cδ(n)×δ(n).
Introducing again pointwise operations and the
supremum norm, we make F δ to a C ∗-algebra with identity element (Iδ(n)), the
algebra of matrix sequences with dimension function δ. The set of all sequences
in F δ which tend to zero in the norm forms a closed ideal of F δ which we denote
by Gδ. For example, the algebra of matrix sequences with constant dimension
function δ = 1 is l∞(N), but in what follows we will be mainly interested in strictly
increasing dimension functions, as they occur in the context of filtrations.
When passing from F P to F δ with δ(n) := rank Pn, one loses the embedding
of the matrix algebras L(im Pn) ∼= Cδ(n)×δ(n) into a common Hilbert space. It
makes thus no sense to speak about strong convergence of a sequence in F δ. But
it will turn out that algebras of matrix sequences provide a suitable frame to
formulate and study stability problems as well as a lot of other problems which
do not depend upon an embedding into a Hilbert space. Moreover, some of the
notions and assertions discussed in this paper remain meaningful in the much
more general context, when F C is the direct product of a sequence C = (Cn)n≥1
of unital C ∗-algebras. The associated ideal of zero sequences in F C, which can
be identified with the direct sum of the family C in a natural way, will then be
denoted by GC.
The following will serve as a running example in this paper. We consider the
algebra of the finite sections discretization for Toeplitz operators with continuous
generation function. For a continuous function a on the complex unit circle T,
the associated Toeplitz operator is the operator T (a) on l2(Z+) which is given by
the infinite matrix (ai−j)∞
i,j=0, with ak denoting the kth Fourier coefficient of a.
Note that T (a) is a bounded operator and kT (a)k = kak∞. For n ∈ N, put
Pn : l2(Z+) → l2(Z+),
(xn)n≥0 7→ (x0, x1, . . . , xn−1, 0, 0, . . .).
Then P = (Pn) is a filtration on l2(Z+). We let S(T(C)) stand for the smallest
closed subalgebra of F P which contains all sequences (PnT (a)im Pn) of finite sec-
tions of Toeplitz operators T (a) with a ∈ C(T). Let Rn : im Pn → im Pn be the
reflection operator
(x0, x1, . . . , xn−1, 0, 0, . . .) 7→ (xn−1, . . . , x1, x0, 0, 0, . . .).
It is not hard to see that for each sequence A = (An) ∈ S(T(C)), the strong
limit fW (A) := s-limRnAnRnPn exists and that fW is a unital and fractal ∗-
homomorphism from S(T(C)) to L(l2Z+). The following is a by now classical
result by Bottcher and Silbermann [5], see also Chapter 2 in [6] and Sections 1.3,
1.4 and 1.6 in [7].
2
Theorem 1 (a) The algebra S(T(C)) consists of all sequences (PnT (a)Pn +
PnKPn + RnLRn + Gn) where a ∈ C(T), K, L ∈ K(l2(Z+)), and (Gn) ∈ GP.
(b) For every sequence A ∈ S(T(C)), the coset A + GP is invertible in the quo-
tient algebra S(T(C))/GP if and only if the operators W P(A) and fW (A) are
invertible.
Due to its transparent structure, the algebra S(T(C)) served as a basic example
for the development of algebraic methods in asymptotic numerical analysis. These
methods have found fruitful applications in the stability analysis of different
approximation methods for numerous classes of operators; see the monographs [7,
8, 17] for an overview. In particular, I would like to emphasize the finite sections
method for band-dominated operators, a topic which was mainly influenced and
shaped by Vladimir S. Rabinovich and the limit operator techniques developed
by him, see [9, 10, 11, 12, 13] and [15] for an overview. In fact, the algebra of the
finite sections method for band-dominated operators is the first real-life example
of an essentially fractal, but not fractal, algebra (these notions will be introduced
below).
2 Fractality
As it was observed in [14, 16], several natural approximation procedures lead
to C ∗-subalgebras A of the algebra F which are distinguished by the property
of self-similarity: Given a subsequence of a sequence in A, one can uniquely
reconstruct the full sequence up to a sequence which tends to zero in the norm.
These algebras were called fractal in [16]. The goals of this section is to recall
the basic definitions and some consequences of fractality, and to give a short
proof of the known fact that every separable subalgebra of F possesses a fractal
restriction.
In this section, we let F := F C be the product of a family C = (Cn)n∈N of
unital C ∗-algebras and G := GC the associated ideal of zero sequences.
2.1 Definition and first consequences
For each strictly increasing sequence η : N → N, let Fη stand for the product of
the family (Cη(n))n∈N of C ∗-algebras, and write Gη for the associated ideal of zero
sequences. The elements of Fη can be viewed of as subsequences of sequences
in F . The canonical restriction mapping Rη : F → Fη, (An) 7→ (Aη(n)) is a
∗-homomorphism from F onto Fη and maps G onto Gη. More generally, for each
C ∗-subalgebra A of F , we let Aη denote the image of A under Rη. Clearly, Aη is
a C ∗-subalgebra of Fη. We call algebras obtained in this way restrictions of A.
Definition 2 (a) Let A be a C ∗-subalgebra of F . A ∗-homomorphism W from
A into a C ∗-algebra B is called fractal if it factors through RηA for every strictly
3
increasing sequence η : N → N, i.e., if for each such η, there is a mapping
Wη : Aη → B such that W = WηRηA.
(b) A C ∗-subalgebra A of F is fractal if the canonical homomorphism
A → A/(A ∩ G), A 7→ A + (A ∩ G)
is fractal.
(c) A sequence A ∈ F is fractal if the smallest C ∗-subalgebra of F which contains
the sequence A and the identity sequence is fractal.
For example, if P is a filtration, then the associated consistency map W P is fractal
(since the strong limit of a sequence (An) ∈ F P can be determined from each
subsequence of (An)). For the same reason, the homomorphism fW appearing in
Theorem 1 is fractal.
The fractal subalgebras of F are distinguished by their property that every se-
quence in the algebra can be rediscovered from each of its (infinite) subsequences
up to a sequence tending to zero. Note that, by Definition 2, a fractal sequence
always lies in a unital fractal algebra, whereas a fractal algebra needs not to be
unital.
Assertion (a) of the following theorem provides an equivalent characterization
of the fractality of an algebra. Proofs of Theorems 3 and 4 are given in [16] and
in Section 1.6 of [7].
Theorem 3 (a) A C ∗-subalgebra A of F is fractal if and only if the implication
Rη(A) ∈ Gη ⇒ A ∈ G
(2)
holds for every sequence A ∈ A and every strictly increasing sequence η.
(b) If A is a fractal C ∗-subalgebra of F , then Aη ∩ Gη = (A ∩ G)η for each strictly
increasing sequence η.
(c) A unital C ∗-subalgebra of F is fractal if and only if each of its elements is
fractal.
The following criterion will prove to be useful in order to verify the fractality of
many specific algebras of approximation methods.
Theorem 4 A unital C ∗-subalgebra A of F is fractal if and only if there is a
family {Wt}t∈T of unital and fractal ∗-homomorphisms Wt from A into unital C ∗-
algebras Bt such that the following equivalence holds for every sequence A ∈ A:
The coset A + A ∩ G is invertible in A/(A ∩ G) if and only if Wt(A) is invertible
in Bt for every t ∈ T .
For example, since W P and fW are fractal homomorphisms, we conclude from
Theorem 1 (b) and from the previous theorem that the algebra S(T(C)) is frac-
tal.
4
The property of fractality has striking consequences for asymptotic spectral prop-
erties of a sequence A = (An), see [14, 16] and Chapter 3 in [7]. Here we only
mention a few of them which are relevant for what follows. For every element a
of a unital C ∗-algebra A, we let σ2(a) denote the set of all non-negative square
roots of points in the spectrum of a∗a. In case A = Cn×n, the numbers in σ2(a)
are known as the singular values of a.
Proposition 5 Let A be a fractal C ∗-subalgebra of F and A = (An) a sequence
in A. Then
(a) the sequence A is stable if and only if it possesses a stable subsequence;
(b) the limit limn→∞ kAnk exists and is equal to kA + Gk;
(c) the limit limn→∞ σ2(An) exists with respect to the Hausdorff distance on R
and is equal to σ2(A + G).
2.2 The fractal restriction theorem
The preceding proposition and related results from [7] indicate that it is a question
of vital importance in numerical analysis to single out fractal subsequences of a
given sequence in F . The following theorem states that such subsequences always
exist.
Theorem 6 Let A be a separable C ∗-subalgebra of F . Then there exists a strictly
increasing sequence η : N → N such that the restricted algebra Aη = RηA is a
fractal subalgebra of Fη.
Since finitely generated C ∗-algebras are separable, this result immediately im-
plies:
Corollary 7 Every sequence in F possesses a fractal subsequence.
Theorem 6 was first proved in [14]. We shall give a much shorter proof here, which
is based on the following converse of assertion (b) of Proposition 5 (whereas the
original proof used the converse of assertion (c) of this proposition).
Proposition 8 Let A be a C ∗-subalgebra of F and L a dense subset of A. If
the sequence of the norms kAnk converges for each sequence (An) ∈ L, then the
algebra A is fractal.
Proof. First we show that if the sequence of the norms converges for each
sequence in L, then it converges for each sequence in A. Let (An) ∈ A and ε > 0.
Choose (Ln) ∈ L such that k(An − Ln)k = sup kAn − Lnk < ε/3, and let n0 ∈ N
be such that kLnk − kLmk < ε for all m, n ≥ n0. Then, for m, n ≥ n0,
kAnk − kAmk ≤ kAnk − kLnk + kLnk − kLmk + kLmk − kAmk
≤ kAn − Lnk + kLnk − kLmk + kLm − Amk ≤ ε.
5
Thus, (kAnk) is a Cauchy sequence, hence convergent. But the convergence of the
norms for each sequence in A implies the fractality of A by Theorem 3. Indeed,
if a subsequence of a sequence (An) ∈ A tends to zero, then 0 = lim inf kAnk =
lim kAnk, whence (An) ∈ G.
Proof of Theorem 6. Let {Am}m∈N of Asa be a dense countable subset of A
which consists of sequences Am = (Am
n )n∈N. Let η1 : N → N be a strictly increas-
ing sequence such that the sequence of the norms kA1
η1(n)k converges. Next let η2
be a strictly increasing subsequence of η1 such that the sequence (kA2
η2(n)k)n∈N
converges. We proceed in this way and find, for each k ≥ 2, a strictly increasing
subsequence ηk of ηk−1 such that the sequence (kAk
ηk(n)k)n∈N converges. Define
the sequence η by η(n) := ηn(n). Then η is strictly increasing, and the sequence
(kAk
η(n)k)n∈N converges for every k ∈ N.
Since the sequences Rη(Am) with k ∈ N form a dense subset of the restricted
η(n))n∈N has the property
η(n)k converges, the assertion follows from
algebra Aη, and since each sequence Rη(Am) = (Ak
that the sequence of the norms kAk
Proposition 8.
3 Essential fractality
Recall that a C ∗-subalgebra A of F is fractal if each sequence (An) ∈ A can
be rediscovered from each of its (infinite) subsequences modulo a sequence in
the ideal G. There are plenty of subalgebras of F which arise from concrete
discretization methods and which are fractal (the finite sections algebra S(T(C))
for Toeplitz operators is one example). On the other hand, the algebra of the
finite sections method for band-dominated operators is an example of an algebra
which fails to be fractal. But the latter algebra enjoys a weaker form of fractality
which we called essential fractality in [15]. Basically, a C ∗-subalgebra A of F
is essentially fractal if each sequence (An) ∈ A can be rediscovered from each
of its (infinite) subsequences modulo a sequence in the ideal K of the compact
sequences. The role of this ideal in numerical analysis can be compared with the
role of the ideal of the compact operators in operator theory.
In this section, we first recall the definition of a compact sequence and state
some useful characterizations of compactness and the definitions of J -fractality
and essential fractality from [15]. The main goal of this section is to derive an
analogue of the fractal restriction theorem for essential fractality.
Unless otherwise stated, we let F = F δ be an algebra of matrix sequences
with dimension function δ and G := Gδ the associated ideal of zero sequences in
this section.
6
3.1 Compact sequences
Slightly abusing the notation, we call a sequence (Kn) ∈ F a sequence of rank
one matrices if the rank of every matrix Kn is less than or equal to one. The
product of a sequence of rank one matrices with a sequence in F is a sequence
of rank one matrices again. Hence, the set of all finite sums of sequences of rank
one matrices forms an (in general, non-closed) ideal of F . We let K denote the
closure of this ideal and refer to the elements of K as compact sequences. Thus, K
is the smallest closed ideal of F which contains all sequences of rank one matrices,
and a sequence (An) ∈ F is compact if, and only if, for every ε > 0, there is a
sequence (Kn) ∈ F such that
sup
kAn − Knk < ε and
sup
rank Kn < ∞.
(3)
n
n
Note that K contains the ideal G, and that the restriction of a compact sequence is
compact. More precisely, if K is a compact sequence in the algebra F δ of matrix
sequences with dimension function δ and if η is a strictly increasing sequence,
then the restriction RηK is a compact sequence in the algebra RηF δ ∼= F δ◦η of
matrix sequences with dimension function δ ◦ η.
An appropriate notion of the rank of a sequence in F can be introduced as
follows. A sequence A ∈ F has finite essential rank if it is the sum of a sequence
in G and a sequence (Kn) with supn rank Kn < ∞. If A is of finite essential rank,
then there is a smallest integer r ≥ 0 such that A can be written as (Gn) + (Kn)
with (Gn) ∈ G and supn rank Kn ≤ r. We call this integer the essential rankof
A and write ess rank A = r. Thus, the sequences of essential rank 0 are just
the sequences in G. If A is not of finite essential rank, we set ess rank A = ∞.
Clearly, the sequences of finite essential rank form an ideal of F which is dense
in K, and for arbitrary sequences A, B ∈ F one has
ess rank (A + B) ≤ ess rank A + ess rank B,
ess rank (AB) ≤ min {ess rank A, ess rank B}.
Given a filtration P = (Pn) on a Hilbert space H, we identify the algebra F P
with the algebra F of matrix sequences with dimension function δ(n) := rank Pn.
Note that this identification requires the choice of an orthogonal basis in each
space im Pn. We define the ideal KP of the compact sequences in F P in the
same way as before. It is clear that then the ideal KP can be identified with K,
independently of the choice of the bases.
For example, using the explicit description of the finite sections algebra of
Toeplitz operators in Theorem 1 (a), it is not hard to show that the intersection
S(T(C)) ∩ K consists of all sequences
(PnKPn + RnLRn + Gn) with K, L compact and (Gn) ∈ G
(4)
7
and that the essential rank of the sequence (4) is equal to rank K + rank L.
There are several equivalent characterizations of compact sequences, see [15]. In
what follows we shall need a characterization of a compact sequence (Kn) in terms
of the asymptotic behavior of the singular values of the entries Kn. To state this
criterion, we denote the decreasingly ordered singular values of an n × n matrix
A by
kAk = Σ1(A) ≥ Σ2(A) ≥ . . . ≥ Σn(A) ≥ 0
(5)
and recall from Linear Algebra that A∗A and AA∗ are unitarily equivalent,
whence Σk(A) = Σk(A∗), and that every matrix A has a singular value decom-
position (SVD)
A = E∗ diag (Σ1(A), . . . , Σn(A))F
(6)
with unitary matrices E and F .
The announced characterization of compact sequences in terms of singular
values reads as follows. See Sections 4.2 and 5.1 in [15] for the proof of this and
the following theorem.
Theorem 9 The following conditions are equivalent for a sequence (Kn) ∈ F :
(a) limk→∞ supn≥k Σk(Kn) = 0;
(b) limk→∞ lim supn→∞ Σk(Kn) = 0;
(c) the sequence (Kn) is compact.
A sequence in F is called a Fredholm sequence if it is invertible modulo K. As the
compact sequences, Fredholm sequences can be characterized in terms of singular
values. Let σ1(A) ≤ . . . ≤ σn(A) denote the increasingly ordered singular values
of an n × n-matrix A.
Theorem 10 The following conditions are equivalent for a sequence (An) ∈ F :
(a) (An) is a Fredholm sequence.
(b) There are sequences (Bn) ∈ F and (Jn) ∈ K with supn rank Jn < ∞ such that
BnAn = In + Jn for all n ∈ N.
(c) There is a k ∈ Z+ such that lim inf n→∞ σk+1(An) > 0.
3.2 J -fractal algebras
Our next goal is to introduce fractality of an algebra A with respect to an arbi-
trary ideal J in place of G. The results presented in this subsection hold in the
general case, when F is the product of a family (Cn)n∈N of unital C ∗-algebras. We
start with a criterion for the fractality of the canonical quotient map A → A/J .
8
Theorem 11 Let A be a C ∗-subalgebra of F and J a closed ideal of A. The
canonical homomorphism πJ : A → A/J is fractal if and only if the following
implication holds for every sequence A ∈ A and every strictly increasing sequence
η : N → N
Rη(A) ∈ Jη =⇒ A ∈ J .
(7)
Proof. Let πJ be fractal, i.e., for each η, there is a mapping πJ
η
πJ = πJ
J ∈ J such that Rη(A) = Rη(J). Applying the homomorphism πJ
of this equality we obtain πJ (A) = πJ (J) = 0, whence A ∈ J .
such that
η RηA. Let Rη(A) ∈ Jη for a sequence A ∈ A. We choose a sequence
η to both sides
For the reverse implication, let A and B be sequences in A with Rη(A) =
Rη(B). Then Rη(A − B) = 0 ∈ Jη, and (7) implies that A − B ∈ J . Thus, the
mapping
πJ
η : Aη → A/J , Rη(A) 7→ A + J
is correctly defined, and it satisfies πJ
η RηA = πJ .
Let now J be a closed ideal of F . Then A ∩ J is a closed ideal of A, and the
preceding theorem states that the canonical mapping πA∩J : A → A/(A ∩ J ) is
fractal if and only if the implication
Rη(A) ∈ (A ∩ J )η =⇒ A ∈ J
(8)
holds for every sequence A ∈ A and every strictly increasing sequence η. It would
be much easier to check this implication if one would have
(A ∩ J )η = Aη ∩ Jη
(9)
foe every η, in which case the implication (8) reduces to Rη(A) ∈ Jη ⇒ A ∈ J .
Recall from Theorem 3 (b) that (9) indeed holds if J = G and if the canonical
homomorphism πA∩G : A → A/(A∩G) is fractal. One cannot expect an analogous
result for arbitrary closed ideals J of F , as the following example shows.
Example 12 Let A := S(T(C)) the algebra of the finite sections method for
Toeplitz operators and K the ideal of the compact sequences in the corresponding
algebra F . Then
J := {(Kn) ∈ K : lim
n→∞
kK2nk = 0}
is a closed ideal of F . Employing again the explicit description of S(T(C)) in
Theorem 1 (a), it is not hard to see that S(T(C)) ∩ J = G. Consequently,
the canonical homomorphism πS(T(C))∩J coincides with πG and is, thus, fractal.
But Gη = (S(T(C)) ∩ J )η is a proper subset of S(T(C))η ∩ Jη for the sequence
η(n) := 2n + 1. Indeed, the sequence (P2n+1KP2n+1) belongs to S(T(C))η ∩ Jη
for each compact operator K.
The previous considerations suggest the following definitions. Note that both
definitions coincide if J is a closed ideal of A and F .
9
Definition 13 Let A be a C ∗-subalgebra of F .
(a) If J is a closed ideal of A then A is called J -fractal if the canonical homo-
morphism πJ : A → A/J is fractal.
(b) If J is a closed ideal of F then A is called J -fractal if A is (A ∩ J )-fractal
and if (A ∩ J )η = Aη ∩ Jη for every strictly increasing sequence η : N → N.
The following results show that J -fractality implies what one expects: A sequence
in a J -fractal algebra belongs to J or is invertible modulo J if and only if at
least one of its subsequences has this property.
Theorem 14 Let J be a closed ideal of F . A C ∗-subalgebra A of F is J -fractal
if and only if the following implication holds for every sequence A ∈ A and every
strictly increasing sequence η
Rη(A) ∈ Jη =⇒ A ∈ J .
(10)
Proof. Let A be J -fractal and A ∈ A a sequence with Rη(A) ∈ Jη. Then
Rη(A) ∈ Aη ∩ Jη = (A ∩ J )η, and the (A ∩ J )-fractality of A implies A ∈ J
via Theorem 11.
Conversely, let (10) hold for every strictly increasing sequence η. From The-
orem 11 we conclude that A is (A ∩ J )-fractal. Further, the inclusion ⊆ in
(A ∩ J )η = Aη ∩ Jη is obvious. For the reverse inclusion, let A be a sequence
in F with Rη(A) ∈ Aη ∩ Jη. Then there are sequences B ∈ A and J ∈ J such
that Rη(A) = Rη(B) = Rη(J). Since Rη(B) ∈ Jη, the implication (10) gives
B ∈ J . Hence, Rη(B) ∈ (A ∩ J )η, and since Rη(B) = Rη(A), one also has
Rη(A) ∈ (A ∩ J )η.
Theorem 15 Let J be a closed ideal of F and A a J -fractal and unital C ∗-
subalgebra of F . Then the following implication holds for every sequence A ∈ A
and every strictly increasing sequence η
Rη(A) + Jη is invertible in Fη/Jη =⇒ A + J is invertible in F /J .
(11)
Proof. Let A ∈ A be such that Rη(A) + Jη is invertible in Fη/Jη. Since C ∗-
algebras are inverse closed, this coset is also invertible in (Aη +Jη)/Jη. The latter
algebra is canonically ∗-isomorphic to Aη/(Aη ∩ Jη), hence, to Aη/(A ∩ J )η by
J -fractality of A. Thus, the coset Rη(A) + (A ∩J )η is invertible in Aη/(A ∩J )η.
Choose sequences B ∈ A and J ∈ A ∩ J such that
Rη(A) Rη(B) = Rη(I) + Rη(J)
where I denotes the identity element of F . Applying the homomorphism πA∩J
to both sides of this equality one gets
η
πA∩J (A) πA∩J (B) = πA∩J (I) + πA∩J (J)
which shows that AB − I ∈ J . Hence, A is invertible modulo J from the
right-hand side. The invertibility from the left-hand side follows analogously.
10
Corollary 16 Let J be a closed ideal of F and A a J -fractal and unital C ∗-
subalgebra of F . Then a sequence A ∈ A
(a) belongs to J if and only if there is a strictly increasing sequence η such that
Aη belongs to Jη.
(b) is invertible modulo J if and only there is a strictly increasing sequence η
such that Aη is invertible modulo Jη.
We still mention the following simple facts for later reference.
Proposition 17 Let J be a closed ideal of F and A a J -fractal C ∗-subalgebra
of F . Then
(a) every C ∗-subalgebra of A is J -fractal.
(b) if I is an ideal of F with J ⊆ I and if (A ∩ I)η = Aη ∩ Iη for each strictly
increasing sequence η : N → N, then A is I-fractal.
Proof. (a) Let B be a C ∗-subalgebra of A, and let B be a sequence in B with
Rη(B) ∈ Jη for a certain strictly increasing sequence η. Then Rη(B) ∈ Aη ∩ Jη.
Since A is J -fractal, Theorem 14 implies that B ∈ J . Hence B is J -fractal,
again by Theorem 14.
(b) Let Rη(A) ∈ Iη for a sequence A ∈ A and a strictly increasing sequence η. By
hypothesis, Rη(A) ∈ (A∩I)η. Choose a sequence J ∈ A∩I with Rη(A) = Rη(J).
The J -fractality of A implies that A − J ∈ J , whence A ∈ J + J ⊆ I. By
Theorem 14, A is I-fractal.
3.3 Essential fractality and Fredholm property
Let again F be the algebra of matrix sequences with dimension function δ and K
the associated ideal of compact sequences. We call the K-fractal C ∗-subalgebras
of F essentially fractal.
Note that each restriction Fη of F is again an algebra of matrix sequences
(with dimension function δ ◦ η); hence, the restriction Kη of K is just the ideal
of the compact sequences related with Fη. If we speak on compact subsequences
and Fredholm subsequences in what follows, we thus mean sequences RηA ∈ Kη
and sequences RηA which are invertible modulo Kη, respectively. In these terms,
Corollary 16 reads as follows.
Corollary 18 Let A be an essentially fractal and unital C ∗-subalgebra of F .
Then a sequence A ∈ A is compact (resp. Fredholm) and only if one of the
subsequences of A is compact (resp. Fredholm).
The following is a consequence of Proposition 17.
Corollary 19 Let A be a fractal C ∗-subalgebra of F . If (A ∩ K)η = Aη ∩ Kη for
each strictly increasing sequence η : N → N, then A is essentially fractal.
11
Essential fractality has striking consequences for the behavior of the smallest
singular values.
Theorem 20 Let A be an essentially fractal and unital C ∗-subalgebra of F . A
sequence (An) ∈ A is Fredholm if and only if there is a k ∈ N such that
lim sup
σk(An) > 0.
n→∞
(12)
If (An) is Fredholm then, by Theorem 10 (c), lim inf n→∞ σk (An) > 0
Proof.
for some k ∈ N, whence (12). Conversely, let (12) hold for some k. We choose
a strictly increasing sequence η such that limn→∞ σk(Aη(n)) > 0. Thus, the
restricted sequence (Aη(n))n≥1 is Fredholm by Theorem 10. Since A is essentially
fractal, Corollary 18 (b) implies the Fredholm property of the sequence (An) itself.
Consequently, if a sequence (An) in an essentially fractal and unital C ∗-subalgebra
of F is not Fredholm, then
lim
n→∞
σk(An) = 0
for each k ∈ N.
(13)
In analogy with operator theory, we call a sequence (An) with property (13) not
normally solvable.
Corollary 21 Let A be an essentially fractal and unital C ∗-subalgebra of F .
Then a sequence in A is either Fredholm or not normally solvable.
Example 22 Consider the finite sections algebra S(T(C)) for Toeplitz operators.
It is a simple consequence of Theorem 1 (a) that (S(T(C))∩K)η = S(T(C))η ∩Kη
for each strictly increasing sequence η. Since S(T(C)) is fractal and G ⊂ K, the
algebra S(T(C)) is essentially fractal by Corollary 19.
3.4 Essential fractal restriction
Our final goal is an analogue of Theorem 6 for essential fractality. Recall that
we based the proof of Theorem 6 on the fact that there is a sequence η such that
the norms kAη(n)k converge for each sequence (An). We start with showing that
η can be even chosen such that not only the sequences (kAη(n)k) = (Σ1(Aη(n)))
converge, but every sequence (Σk(Aη(n))) with k ∈ N. Here, Σ1(A) ≥ . . . ≥ Σn(A)
denote the decreasingly ordered singular values of the n × n-matrix A.
Proposition 23 Let A be a separable C ∗-subalgebra of F . Then there is a strictly
increasing sequence η : N → N such that the sequence (Σk(Aη(n)))n≥1 converges
for every sequence (An)n≥1 ∈ A and every k ∈ N.
12
Proof. First consider a single sequence (An) ∈ A. We choose a strictly increasing
sequence η1 : N → N such that the sequence (Σ1(Aη1(n)))n≥1 converges, then a
subsequence η2 of η1 such that the sequence (Σ2(Aη2(n)))n≥1 converges, and so on.
The sequence η(n) := ηn(n) has the property that the sequence (Σk(Aη(n)))n≥1
converges for every k ∈ N.
Now let (Am)m≥1 be a countable dense subset of A, consisting of sequences
Am = (Am
n )n≥1. We use the result of the previous step to find a strictly increasing
sequence η1 : N → N such that the sequences (Σk(A1
η1(n)))n≥1 converge for every
k ∈ N, then a subsequence η2 of η1 such that the sequences (Σk(A2
η2(n)))n≥1
converge for every k, and so on. Then the sequence η(n) := ηn(n) has the
property that the sequences (Σk(Am
η(n)))n≥1 converge for every pair k, m ∈ N.
Let η be as in the previous step, i.e., the sequences (Σk(Am
η(n)))n≥1 converge
for every k ∈ N and for every sequence Am = (Am
n )n≥1 in a countable dense
subset of A. We show that then the sequences (Σk(Aη(n)))n≥1 converge for every
k ∈ N and every sequence A = (An) in A. Fix k ∈ N and let ε > 0. Using the
well known inequality Σk(A) − Σk(B) ≤ kA − Bk we obtain
Σk(Aη(n)) − Σk(Aη(l))
≤ Σk(Aη(n)) − Σk(Am
η(n)) + Σk(Am
η(n)) − Σk(Am
η(l))
+ Σk(Am
η(l)) − Σk(Aη(l))
≤ kAη(n) − Am
η(n)k + Σk(Am
≤ 2 kA − AmkF + Σk(Am
η(n)) − Σk(Am
η(n)) − Σk(Am
η(l)) + kAm
η(l)).
η(l) − Aη(l)k
Now choose m ∈ N such that kA − AmkF < ε/3 and then N ∈ N such that
Σk(Am
η(l)) < ε/3 for all n, l ≥ N. Then Σk(Aη(n)) − Σk(Aη(l)) < ε
for all n, l ≥ N. Thus, (Σk(Aη(n)))n≥1 is a Cauchy sequence, hence convergent.
η(n)) − Σk(Am
Proposition 24 Let A be a C ∗-subalgebra of F with the property that the se-
quences (Σk(An))n≥1 converge for every sequence (An) ∈ A and every k ∈ N.
Then A is essentially fractal.
Proof. Let K = (Kn) ∈ A and let η : N → N be a strictly increasing sequence
such that Kη ∈ Kη. Then, by Theorem 9 (b),
lim
k→∞
lim sup
n→∞
Σk(Kη(n)) = 0.
(14)
By hypothesis, lim supn→∞ Σk(Kη(n)) = limn→∞ Σk(Kn). Hence, (14) implies
limk→∞ limn→∞ Σk(Kn) = 0, whence K ∈ K by assertion (a) of Theorem 9.
Thus, every sequence in A which has a compact subsequence is compact itself.
Thus A is essentially fractal by Theorem 14.
Theorem 25 Let A be a separable C ∗-subalgebra of F . Then there is a strictly
increasing sequence η : N → N such that the restricted algebra Aη = RηA is
essentially fractal.
13
Indeed, if η is as in Proposition 23, then the restriction Aη is essentially fractal
by Proposition 24.
We know from Theorems 6 and 25 that every separable C ∗-subalgebra of F has
both a fractal and an essentially fractal restriction. If is an open question whether
this fact holds for arbitrary closed ideals J of F in place of G or K, i.e., whether
one can always force J -fractality by a suitable restriction.
4 Essential spectral approximation
In a series of papers [1, 2, 3], Arveson studied the question of whether one can
discover the essential spectrum of a self-adjoint operator A from the behavior
of the eigenvalues of the finite sections PnAPn of A. More generally, one might
ask whether one can discover the essential spectrum of a self-adjoint sequence
A = (An) ∈ F (i.e., the spectrum of the coset A + K, considered as an element of
the quotient algebra F /K) from the behavior of the eigenvalues of the matrices
An? To answer this question, Arveson introduced the notions of essential and
transient points, and he discovered (under an additional condition) a certain
dichotomy:
if A is a self-adjoint band-dominated operator, then every point in
R is either transient or essential; see Subsection 4.2. The goal of this section
is to relate the essential spectral approximation with the property of essential
fractality. In particular, we will see that a subalgebra A of F is essentially fractal
if and only if every self-adjoint sequence in A has Arveson dichotomy.
4.1 Essential spectra of self-adjoint sequences
Given a self-adjoint matrix A and a subset M of R, let N(A, M) denote the num-
ber of eigenvalues of A which lie in M, counted with respect to their multiplicity.
If M = {λ} is a singleton, we write N(A, λ) in place of N(A, {λ}). Thus, if λ is
an eigenvalue of A, then N(A, λ) is its multiplicity.
Let A = (An) ∈ F be a self-adjoint sequence. Following Arveson [1, 2, 3],
a point λ ∈ R is called essential for this sequence if, for every open interval U
containing λ,
lim
n→∞
N(An, U) = ∞,
and λ ∈ R is called transient for A if there is an open interval U which contains
λ such that
N(An, U) < ∞.
sup
n∈N
Thus, λ ∈ R is not essential for A if and only if λ is transient for a subsequence
of A, and λ is not transient for A if and only if λ is essential for a subsequence
of A. Moreover, if a point λ is transient (resp. essential) for A, then is is also
transient (resp. essential) for every subsequence of A.
14
Theorem 26 Let A ∈ F be a self-adjoint sequence. A point λ ∈ R belongs to
the essential spectrum of A if and only if it is not transient for the sequence A.
Proof. Let A = (An) be a bounded sequence of self-adjoint matrices. First let
λ ∈ R \ σ(A + K). We set Bn := An − λIn and have to show that 0 is transient
for the sequence (Bn). Since λ ∈ R \ σ(A + K), the sequence (Bn) is Fredholm.
By Theorem 10 (c), there is a k ∈ Z+ such that
lim inf
n→∞
σk+1(Bn) =: C > 0 and
lim inf
n→∞
σk(Bn) = 0.
Let U := (−C/2, C/2). Since the singular values of a self-adjoint matrix are
just the absolute values of the eigenvalues of that matrix, we conclude that
N(Bn, U) ≤ k for all sufficiently large n. Thus, 0 is transient.
Conversely, let λ ∈ R be transient for (An). We claim that (An − λIn) is a
Fredholm sequence. By transiency, there is an interval U = (λ−ε, λ+ε) with ε >
0 such that supn∈N N(An, U) =: k < ∞. Let Tn denote the orthogonal projection
from Cδ(n) onto the U-spectral subspace of An. Then rank Tn is not greater than
k. It is moreover obvious that the matrices Bn := (An − λPn)(I − Tn) + Tn are
invertible for all n ∈ N and that their inverses are uniformly bounded by the
maximum of 1/ε and 1. Hence, (B−1
n ) ∈ F and
(An − λPn)(I − Tn)B−1
n = I − TnB−1
n .
(15)
Since (Tn) is a compact sequence (of essential rank not greater than k), this
identity shows that the coset (An − λIn) + K is invertible from the right-hand
side. Since this coset is self-adjoint, it is then invertible from both sides. Thus,
(An − λIn) is a Fredholm sequence.
Proposition 27 The set of the non-transient points and the set of the essential
points of a self-adjoint sequence A ∈ F are compact.
Proof. The first assertion is an immediate consequence of Theorem 26. The
second assertion will follow once we have shown that the set of the essential
points of A is closed.
Let (λk) be a sequence of essential points for A = (An) with limit λ. Assume
that λ is not essential for A. Then there is a strictly increasing sequence η :
N → N such that λ is transient for Aη. Let U be an open neighborhood of λ
with supn∈N N(Aη(n), U) =: c < ∞. Since λk → λ and U is open, there are a
k ∈ N and an open neighborhood Uk of λk with Uk ⊆ U. Clearly, N(Aη(n), Uk) ≤
N(Aη(n), U) ≤ c. On the other hand, since λk is also essential for the restricted
sequence Aη, one has N(Aη(n), Uk) → ∞ as n → ∞, a contradiction.
Note that the set of the non-transient points of a self-adjoint sequence is non-
empty by Theorem 26, whereas it is easy to construct self-adjoint sequences
without any essential point: take a sequence which alternates between the zero
15
and the identity matrix.
In contrast to this observation, the following result
shows that sequences which arise by discretization of a self-adjoint operator,
always possess essential points. Let H be an infinite dimensional separable Hilbert
space with filtration P := (Pn), and define the algebra F P as in Section 1. One
can think of F P as a C ∗-subalgebra of the algebra F δ with dimension function
δ(n) := rank Pn.
Theorem 28 Let A := (An) ∈ F P be a self-adjoint sequence with strong limit
A. Then every point in the essential spectrum of A is an essential point for A.
Proof. We show that A − λI is a Fredholm operator if λ ∈ R is not essential for
A. Then λ is transient for a subsequence of A, i.e., there are an infinite subset
M of N and an interval U = (λ − ε, λ + ε) with ε > 0 such that
N(An, U) =: k < ∞.
sup
n∈M
(16)
Let Tn denote the orthogonal projection from H onto the U-spectral subspace of
AnPn. By (16), the rank of the projection Tn is not greater than k if n ∈ M. So
we conclude that the operators Bn := (An − λPn)(I − Tn) + Tn are invertible for
all n ∈ M and that their inverses are uniformly bounded by the maximum of 1/ε
and 1. Hence,
(An − λPn)(I − Tn)B−1
n = I − TnB−1
n
(17)
nr )r≥1 of (TnB−1
nr )r≥1 of ((I − Tn)B−1
for all n ∈ M. By the weak sequential compactness of the unit ball of L(H), one
finds weakly convergent subsequences ((I − Tnr )B−1
n )n∈M and
(Tnr B−1
n )n∈M with limits B and T , respectively. The product of a
weakly convergent sequence with limit C and a ∗-strongly convergent sequence
with limit D is weakly convergent with limit CD. Thus, passing to subsequences
and taking the weak limit in (17) yields (A − λI)B = I − T . Further, the rank of
T is not greater than k by Lemma 5.7 in [4]. Thus, (A − λI)B − I is a compact
operator. The compactness of B(A − λI) − I follows similarly. Hence, A is a
Fredholm operator.
Arveson gave a first example where the inclusion in Theorem 28 is proper. Specif-
ically, he constructed a self-adjoint unitary operator A ∈ L(l2(N)) with
σ(A) = σess(A) = {−1, 1}
(18)
such that 0 is an essential point of the sequence (PnAPn).
4.2 Arveson dichotomy and essential fractality
We say that a self-adjoint sequence A ∈ F enjoys Arveson's dichotomy if every
real number is either essential or transient for this sequence. Note that Arveson
dichotomy is preserved when passing to subsequences. Arveson introduced and
16
studied this property in [1, 2, 3].
In particular, he proved the dichotomy of
the finite sections sequence (PnAPn) when A is a self-adjoint band-dominated
operator which satisfies a Wiener and a Besov space condition. A generalization
to arbitrary band-dominated operators was obtained in [15].
Theorem 29 The set of all self-adjoint sequences in F with Arveson dichotomy
is closed in F .
Proof. Let (An)n∈N be a sequence of self-adjoint sequences in F with Arveson
dichotomy which converges to a (necessarily self-adjoint) sequence A in the norm
of F . Then An + K → A + K in the norm of F /K. Since An + K and A + K are
self-adjoint elements of F /K, this implies that the spectra of An + K converge to
the spectrum of A + K in the Hausdorff metric. Thus, by Theorem 26, the sets
of the non-transient points of An converge to the set of the non-transient points
of A. Since the An have Arveson dichotomy by hypothesis, this finally implies
that the sets of the essential points of An converge to the set of the non-transient
points of A in the Hausdorff metric.
Let now λ be a non-transient point for A and assume that λ is not essential for
A. Then there is a strictly increasing sequence η : N → N such that λ is transient
for the restricted sequence Aη. As we have seen above, there is a sequence (λn),
where λn is an essential point for An, with λn → λ. Since the property of being
an essential is preserved under passage to a subsequence, λn is also essential for
the restricted sequence (An)η.
Since the sequences (An)η also have Arveson dichotomy and since (An)η → Aη
in the norm of Fη, we can repeat the above arguments to conclude that the sets
Mn of the essential points for (An)η converge to the set M of the non-transient
points for Aη in the Hausdorff metric. Since λn ∈ Mn by construction, this
implies that λ ∈ M. This means that λ in not transient for Aη, a contradiction.
Here is the announced result which relates Arveson dichotomy with essential
fractality.
Theorem 30 Let A be a unital C ∗-subalgebra of F . Then A is essentially fractal
if and only if every self-adjoint sequence in A has Arveson dichotomy.
Proof. First let A be essentially fractal. Let A be a self-adjoint sequence
in A and λ ∈ R a point which is not essential for A. The λ is transient for
a subsequence of A, thus, 0 is transient for a subsequence of A − λI. From
Theorem 26 we conclude that this subsequence has the Fredholm property. Then,
by Corollary 18 (b) and since A is essentially fractal, the sequence A − λI itself
is a Fredholm sequence. Thus, 0 is transient for A − λI by Theorem 26 again,
whence finally follows that λ is transient for A. Hence, A has Arveson dichotomy.
Now assume that A is not essentially fractal. Then, by Theorem 14, there are
a sequence A = (An) ∈ A and a strictly increasing sequence η : N → N such that
17
the restricted sequence Aη belongs to Kη but A 6∈ K. The self-adjoint sequence
A∗A has the same properties, i.e., (A∗A)η = A∗
ηAη ∈ Kη, but A∗A 6∈ K.
Since A∗
ηAη ∈ Kη, the essential spectrum of A∗
ηAη (i.e., the spectrum of the
ηAη + Kη in Fη/Kη) consists of the point 0 only. Thus, by Theorem 26,
coset A∗
0 is the only non-transient point for the restricted sequence A∗
ηAη.
µAµ 6∈ Kη. Hence, the essential spectrum of A∗
Since A∗A 6∈ K, there is a strictly increasing sequence µ : N → N such
that µ(N) ∩ η(N) = ∅ and A∗
µAµ
contains at least one point λ 6= 0, and this point is non-transient for A∗
µAµ by
Theorem 26 again. Hence, there is a subsequence ν of µ such that λ is essential
for A∗
ηAη as we have seen above. Thus, λ is
neither transient nor essential for A∗A. Hence, the sequence A∗A does not have
Arveson dichotomy.
νAν, but λ 6= 0 is transient for A∗
Corollary 31 Every self-adjoint sequence in F possesses a subsequence with
Arveson dichotomy.
Proof. Let A be a self-adjoint sequence in F . The smallest closed subalgebra A
of F which contains A is separable. By Theorem 25, there is an essentially fractal
restriction Aη of A. Then Aη is a subsequence of A with Arveson dichotomy by
the previous theorem.
References
[1] W. Arveson, Improper filtrations for C ∗-algebras: Spectra of unilateral
tridiagonal operators. -- Acta Sci. Math. (Szeged) 57(1993), 11 -- 24.
[2] W. Arveson, C ∗-algebras and numerical linear algebra. -- J. Funct. Anal.
122(1994), 333 -- 360.
[3] W. Arveson, The role of C ∗-algebras in infinite dimensional numerical
linear algebra. -- Contemp. Math. 167(1994), 115 -- 129.
[4] A. Bottcher, S. M. Grudsky, Toeplitz matrices, Asymptotic Linear
Algebra and Functional Analysis. -- Hindustan Book Agency, New Delhi
2000.
[5] A. Bottcher, B. Silbermann, The finite section method for Toeplitz
operators on the quarter-plane with piecewise continuous symbols. -- Math.
Nachr. 110(1983), 279 -- 291.
[6] A. Bottcher, B. Silbermann, Introduction to Large Truncated Toeplitz
Matrices. -- Springer-Verlag, Berlin, Heidelberg 1999.
[7] R. Hagen, S. Roch, B. Silbermann, C ∗-Algebras and Numerical Anal-
ysis. -- Marcel Dekker, Inc., New York, Basel 2001.
18
[8] S. Prossdorf, B. Silbermann, Numerical Analysis for Integral and Re-
lated Operator Equations. -- Akademie-Verlag, Berlin, 1991, and Birkhauser
Verlag, Basel, Boston, Stuttgart 1991.
[9] V. S. Rabinovich, S. Roch, Fredholm properties of band-dominated oper-
ators on periodic discrete structures. -- Complex Anal. Oper. Theory 2(2008),
4, 637 -- 681.
[10] V. S. Rabinovich, S. Roch, B. Silbermann, Fredholm theory and finite
section method for band-dominated operators. -- Integral Equations Oper.
Theory 30(1998), 452 -- 495.
[11] V. S. Rabinovich, S. Roch, B. Silbermann, Band-dominated opera-
tors with operator-valued coefficients, their Fredholm properties and finite
sections. -- Integral Eq. Oper. Theory 40(2001), 3, 342 -- 381.
[12] V. S. Rabinovich, S. Roch, B. Silbermann, Limit Operators and
their Applications in Operator Theory. -- Oper. Theory: Adv. Appl. 150,
Birkhauser, Basel 2004.
[13] V. S. Rabinovich, S. Roch, B. Silbermann, On finite sections of band-
dominated operators. -- In: Oper. Theory: Adv. Appl. 181, Birkhauser 2008,
385 -- 391.
[14] S. Roch, Algebras of approximation sequences: Fractality. -- In: Oper.
Theory: Adv. Appl. 121, Birkhauser, Basel 2001, 471 -- 497.
[15] S. Roch, Finite sections of band-dominated operators. -- Memoirs AMS Vol.
191, 895, Providence, R.I., 2008.
[16] S. Roch, B. Silbermann, C ∗-algebra techniques in numerical analysis. --
J. Oper. Theory 35(1996), 2, 241 -- 280.
[17] S. Roch, P. A. Santos, B. Silbermann, Non-commutative Gelfand
Theories. A Tool-kit for Operator Theorists and Numerical Analysts. -- Uni-
versitext, Springer, London 2011.
Author's address:
Steffen Roch, Technische Universitat Darmstadt, Fachbereich Mathematik, Schloss-
gartenstrasse 7, 64289 Darmstadt, Germany.
E-mail: [email protected]
19
|
1205.4649 | 1 | 1205 | 2012-05-21T16:23:48 | New C*-completions of discrete groups and related spaces | [
"math.OA",
"math.GR"
] | Let $\Gamma$ be a discrete group. To every ideal in $\ell^{\infty}(\G)$ we associate a C$^*$-algebra completion of the group ring that encapsulates the unitary representations with matrix coefficients belonging to the ideal. The general framework we develop unifies some classical results and leads to new insights. For example, we give the first C$^*$-algebraic characterization of a-T-menability; a new characterization of property (T); new examples of "exotic" quantum groups; and, after extending our construction to transformation groupoids, we improve and simplify a recent result of Douglas and Nowak. | math.OA | math |
NEW C∗-COMPLETIONS OF DISCRETE GROUPS
AND RELATED SPACES
NATHANIAL P. BROWN AND ERIK GUENTNER
Abstract. Let Γ be a discrete group. To every ideal in ℓ∞(Γ) we associate a C∗-algebra
completion of the group ring that encapsulates the unitary representations with matrix co-
efficients belonging to the ideal. The general framework we develop unifies some classical
results and leads to new insights. For example, we give the first C∗-algebraic characteriza-
tion of a-T-menability; a new characterization of property (T); new examples of "exotic"
quantum groups; and, after extending our construction to transformation groupoids, we
improve and simplify a recent result of Douglas and Nowak [8].
1. Introduction
Since their introduction by von Neumann, amenable groups have played an important role
in many areas of mathematics. They have been studied from a variety of perspectives and
in many different contexts, and a vast literature is now devoted to them. More recently, the
concept of an amenable action of a (non-amenable) group was introduced by Zimmer, and
subsequently developed by many authors. An elementary connection between these theories
is the fact that every action of an amenable group is an amenable action. Less obvious, but
equally well-known, is that if a group acts amenably on a compact space fixing a probability
measure then the group itself is amenable.
This last fact is the launching point of a recent paper by Douglas and Nowak [8], in which,
among other things, they introduce conditions on an amenable action sufficient to guarantee
that the group acting is a-T-menable – in other words, that it admits a metrically proper,
affine isometric action on a Hilbert space. An amenable group is a-T-menable, so that
one may imagine hypotheses involving existence of a quasi-invariant measure together with
conditions on the associated Radon-Nikodym cocycle. Precisely, suppose a discrete group Γ
acts amenably on the compact Hausdorff topological space X, and that µ is a probability
measure on X which is quasi-invariant for the action. Define upper and lower envelopes of
the Radon-Nikodym cocycle by
ρ(x) = sup
s∈G
ds∗µ
dµ
(x),
and ρ(x) = inf
s∈G
ds∗µ
dµ
(x);
here, s∗µ is the translate of the measure µ by the group element s, and ds∗µ/dµ is the Radon-
Nikodym derivative. Douglas and Nowak show that if ρ is integrable, or if ρ is nonzero, then
The first named author was partially supported by DMS-0856197. The second named author was partially
supported by DMS-0349367.
1
2
NATHANIAL P. BROWN AND ERIK GUENTNER
the group Γ is a-T-menable. They ask whether amenability of Γ follows from either of these
conditions. In this note, we shall prove that this is indeed the case. See Corollary 5.11 and
surrounding discussion.
Our initial result lead us to the following question: if one wishes to conclude a-T-menability
of Γ, what are the appropriate hypotheses? To answer this question, we introduce appropriate
completions of the group ring of Γ, and of the convolution algebra Cc(X ⋊ G) in the case of
an action. Precisely, for every algebraic ideal in ℓ∞(Γ) we associate a completion – for ℓ∞(Γ)
we recover the full C∗-algebra, for cc(Γ) we recover the reduced C∗-algebra, and for c0(Γ)
we obtain new C∗-algebras well-adapted to the study of a-T-menability and a-T-menable
actions. Our results in this context are summarized:
Theorem. Let Γ be a discrete group acting on a compact Hausdorff topological space X. Let
C ∗
c0(X ⋊ Γ) be the completions with respect to the ideal c0(Γ). We have:
c0(Γ) and C ∗
(1) Γ is a-T-menable if and only if C ∗(Γ) = C ∗
(2) if the action of Γ on X is a-T-menable then C ∗(X ⋊ Γ) = C ∗
c0(Γ);
c0(X ⋊ Γ).
Further, under the hypotheses of Douglas and Nowak, Γ is a-T-menable if and only if its
action on X is a-T-menable.
Apart from this theorem, and ancillary related results, we develop some general aspects
of our ideal completions. We study when an ideal completion recovers the full or reduced
group C∗-algebra and give examples when it is neither – this gives rise to 'exotic' compact
quantum groups. We recover a standard characterization of amenability – equality of the
full and reduced group C∗-algebras – we obtain the characterization of a-T-menability stated
above, and we characterize Property (T) in terms of ideal completions.
Acknowledgement. The first author thanks the math department at the University of Hawai'i
for embodying the aloha spirit during the sabbatical year when this work was carried out.
He also thanks Yehuda Shalom and Rufus Willett for helpful remarks and suggestions, re-
spectively. Both authors thank Jesse Peterson for sharing his insights.
2. Ideals and C∗-completions
Throughout, Γ will denote a (countable) discrete group and D ⊳ ℓ∞(Γ) will be an algebraic
(not necessarily norm-closed) two-sided ideal. If π : Γ → B(H) is a unitary representation
and vectors ξ, η ∈ H are given, the ℓ∞-function
is a matrix coefficient (function) of π. The map associating to a pair of vectors their matrix
coefficient function is sesquilinear; concretely, given finitely many vectors vi, wj ∈ H, if we
πξ,η(s) := hπs(ξ), ηi
set ξ =P αivi and η =Pj βjwj then we have
πξ,η =Xi,j
αi ¯βjπvi,wj .
In particular, if a linear subspace of ℓ∞(Γ) contains the πvi,wj then it contains πξ,η as well.
NEW C∗-COMPLETIONS OF DISCRETE GROUPS
3
Definition 2.1. Let D ⊳ ℓ∞(Γ) be an ideal. A unitary representation π : Γ → B(H) is a
D-representation if there exists a dense linear subspace H0 ⊂ H such that πξ,η ∈ D for all ξ,
η ∈ H0.
Definition 2.2. Let D ⊳ ℓ∞(Γ) be an ideal. Define a C∗-norm on the group ring C[Γ] by
kxkD := sup{ kπ(x)k : π is a D-representation};
let C ∗
D(Γ) denote the completion of C[Γ] with respect to k · kD.
We shall refer to the C∗-algebra C ∗
D(Γ), and its generalizations defined below, as ideal
completions; these will be our primary objects of study.
Evidently, C ∗
uniquely to C ∗
D(Γ) has the universal property that every D-representation of Γ extends
D(Γ). We shall refer to such representations as D-representations of C ∗
D(Γ).
By virtue of its universal property, the full group C∗-algebra of Γ surjects onto every
ideal completion. For some D the ideal completion C ∗
D(Γ) does not contain the group
ring – it may even be the zero C ∗-algebra! – so is not strictly speaking a 'completion'.
However, if D contains the ideal cc(Γ) of finitely supported functions then C ∗
D(Γ) is indeed
a completion of the group ring – this follows because the regular representation of Γ, being
a cc-representation, extends to C ∗
D(Γ).
Remark 2.3. If D is a closed ideal, then every matrix coefficient of a D-representation belongs
to D (that is, not just those associated to the dense subspace H0). For example, this is the
case for the ideal c0(Γ) of functions vanishing at infinity.
Remark 2.4 (Tensor products). The tensor product of a D-representation and an arbitrary
representation is again a D-representation. Suppose π : Γ → B(H) is a D-representation
and σ : Γ → B(K) is arbitrary. For vi ∈ H0 and wi ∈ K we have
(π ⊗ σ)v1⊗w1,v2⊗w2 = πv1,v2σw1,w2 ∈ D,
since D is an ideal; further, such simple tensors have dense span in H ⊗ K.
Remark 2.5 (Direct sums). An arbitrary direct sum of D-representations is again a D-
representation. This follows since in the definition we only require a dense subspace.
As a consequence, C ∗
D(Γ) has a faithful D-representation. Indeed, for each element x ∈
D(Γ) there is a D-representation π such that π(x) 6= 0. Taking direct sums one easily
C ∗
constructs a faithful D-representation of C ∗
D(Γ).
Definition 2.6. An ideal D ⊳ ℓ∞(Γ) is translation invariant if it is invariant under both the
left and right translation actions of Γ on ℓ∞(Γ).
Every nonzero, translation invariant ideal in ℓ∞(Γ) contains the ideal cc(Γ). It follows that
the ideal completion with respect to a translation invariant ideal surjects onto the reduced
group C∗-algebra of Γ.
4
NATHANIAL P. BROWN AND ERIK GUENTNER
Remark 2.7 (Cyclic representations). Let D be a translation invariant ideal. If v is a cyclic
vector for a representation π : Γ → B(H) and πv,v ∈ D, then π is a D-representation. Indeed,
a computation confirms that if ξ = πg1(v) and η = πg2(v), then
πξ,η(s) = πv,v(g−1
2 sg1),
so that also πξ,η ∈ D. Setting H0 = span{πs(v) : s ∈ Γ} we see that H0 is dense in H and
that the matrix coefficients coming from vectors in H0 belong to D.
Proposition 2.8. Let φ : Γ1 → Γ2 be a group homomorphism and let Di ⊳ ℓ∞(Γi) be ideals
if f ∈ D2 then f ◦ φ ∈ D1. Then φ extends to a C∗-
satisfying the following condition:
homomorphism C ∗
D1(Γ1) → C ∗
D2(Γ2).
Proof. Apply the following simple observation to a faithful D2-representation of Γ2: under
the stated hypotheses, if π is a D2-representation of Γ2, then π ◦ φ is a D1-representation of
Γ1; in particular it extends to C ∗
(cid:3)
D1(Γ1).
Corollary 2.9. The following assertions hold.
(1) Suppose Di⊳ℓ∞(Γ) are ideals and D2 ⊂ D1; there is a quotient map C ∗
D1(Γ) → C ∗
D2(Γ)
(extending the identity map on the group ring).
(2) Suppose Λ ⊂ Γ is a normal subgroup, D2 ⊳ ℓ∞(Γ/Λ) is an ideal and D1 ⊳ ℓ∞(Γ) is an
ideal containing the image of D2 under the inclusion ℓ∞(Γ/Λ) ⊂ ℓ∞(Γ); there is a
surjection (extending the homomorphism φ)
C ∗
D1(Γ) → C ∗
D2(Γ/Λ).
Proof. Both statements are immediate from the proposition. The first is also equivalent to
the inequality, k · kD1 ≤ k · kD2, which is immediate from the definitions.
(cid:3)
We close this introductory section by looking at several basic examples. Our first example
is trivial, since both algebras in question satisfy the same universal property.
Proposition 2.10. For every discrete group Γ, the completion with respect to the ideal ℓ∞(Γ)
is the universal (or full ) group C∗-algebra: C ∗
(cid:3)
ℓ∞(Γ) = C ∗(Γ).
Proposition 2.11. For every discrete group Γ and every p ∈ [1, 2] the completion with
respect to the ideal ℓp(Γ) is the reduced group C∗-algebra: C ∗
ℓp(Γ) = C ∗
r (Γ).
if π : Γ → B(H)
Proof. This follows from the Cowling-Haagerup-Howe Theorem (cf. [7]):
has a cyclic vector v ∈ H and πv,v ∈ ℓ2(Γ), then π is weakly contained in the regular
representation.1 Indeed, fix a nonzero x ∈ C ∗
ℓp(Γ). We can find a cyclic ℓp(Γ)-representation
π such that π(x) 6= 0 – simply restrict a faithful ℓp(Γ)-representation to an appropriate
cyclic subspace. Since π is weakly contained in the regular representation, x cannot be in
the kernel of the map C ∗
(cid:3)
ℓp(Γ) → C ∗
r (Γ).
1Actually, for the Cowling-Haagerup-Howe Theorem it suffices to have πv,v ∈ ℓ2+ε(Γ) for all ε > 0. Thus,
the proposition generalizes to the ideal D := ∩ε>0ℓ2+ε(Γ), with exactly the same proof.
NEW C∗-COMPLETIONS OF DISCRETE GROUPS
5
It follows from functoriality (part (1) of Corollary 2.9) that if a translation invariant ideal
D is contained in ℓp(Γ) for some p ∈ [1, 2] then C ∗
r (Γ). This applies, in particular,
to the ideal of finitely supported functions. In contrast, the ideals ℓp(Γ) for finite p give rise
to the universal group C∗-algebra only if Γ is amenable.
D(Γ) = C ∗
Proposition 2.12. If there exists p ∈ [1, ∞) for which C ∗(Γ) = C ∗
ℓp(Γ), then Γ is amenable.
In the proof, and at a number of places below, we shall use the notion of a positive
definite function: recall that h : Γ → C is positive definite if for every s1, . . . , sn ∈ Γ, the
matrix [h(sis−1
j )]i,j ∈ Mn(C) is positive (semidefinite).
ℓp(Γ), then C ∗(Γ) admits a faithful ℓp(Γ)-representation π and, taking
Proof. If C ∗(Γ) = C ∗
an infinite direct sum if necessary, we may assume π(C ∗(Γ)) contains no compact operators.
In this case, Glimm's lemma implies that π weakly contains the trivial representation. Thus,
let vn be unit vectors such that kπs(vn) − vnk → 0 for all s ∈ Γ. Approximating the vn's
with vectors having associated matrix coefficients in ℓp(Γ), we may assume πvn,vn ∈ ℓp(Γ) for
all n ∈ N. Since πvn,vn are positive definite functions tending pointwise to one, we conclude
that Γ is amenable.
(cid:3)
Remark 2.13. In the previous proof we have used the following elementary fact:
if there
exist positive definite functions hn ∈ ℓp(Γ) for which hn → 1 pointwise, then Γ is amenable.
Lacking a reference, we provide the following argument. For k larger than p the functions
n are positive definite, converge pointwise to one, and belong to ℓ1(Γ) ⊂ C ∗
hk
r (Γ). To get
finitely supported functions with similar properties, consider fn ∈ C(Γ) which approximate
the square roots of the hk
n is approximated by the finitely
supported positive definite function f ∗
n in the norm of C ∗
n ∗ fn.
r (Γ), so that hk
3. Positive definite functions and the Haagerup property
Though very simple, the proof of Proposition 2.12 suggests a general result. We begin
with a lemma isolating the role of translation invariance.
Lemma 3.1. Suppose D ⊳ ℓ∞(Γ) is a translation invariant ideal and h ∈ D is positive
definite. The GNS representation corresponding to h is a D-representation.
Proof. Immediate from Remark 2.7:
v ∈ H is the canonical cyclic vector, then πv,v = h ∈ D.
if π is the GNS representation associated to h, and
(cid:3)
Theorem 3.2. Let D ⊳ ℓ∞(Γ) be a translation invariant ideal. We have that C ∗(Γ) = C ∗
D(Γ)
if and only if there exist positive definite functions hn ∈ D converging pointwise to the
constant function 1.
Proof. First assume that the canonical map C ∗(Γ) → C ∗
D(Γ) is an isomorphism. Replacing
ℓp(Γ) with D in the proof of Proposition 2.12, we see how to construct the desired positive
definite functions.
6
NATHANIAL P. BROWN AND ERIK GUENTNER
For the converse, suppose hn ∈ D are positive definite functions such that hn(s) → 1.
To prove that C ∗(Γ) = C ∗
D(Γ), it suffices to observe that vector states coming from D-
representations are weak-∗ dense in the state space of C ∗(Γ), since this implies that the map
C ∗(Γ) → C ∗
D(Γ) has a trivial kernel. So let ϕ be a state on C ∗(Γ). Then the formula
ϕn Xs∈Γ
αss! :=Xs∈Γ
αshn(s)ϕ(s)
determines a state on C ∗(Γ) – it is the composition of ϕ and the completely positive Schur
multiplier C ∗(Γ) → C ∗(Γ) associated to hn, cf. [5]. Since the norms of the ϕn are uniformly
bounded, we have that ϕn → ϕ in the weak-∗ topology. Also, it's clear that ϕnΓ ∈ D since
it is the product of hn and ϕΓ. So the previous lemma implies the GNS representations
associated to the ϕn are D-representations, concluding the proof.
(cid:3)
It has been open for some time whether the Haagerup property (≡ a-T-menability, see
[6]) admits a C∗-algebraic characterization. The previous theorem easily implies such a
characterization, which is perfectly analogous to a well-known fact about amenable groups.
To see the parallel, we isolate two more canonical ideal completions.
Definition 3.3. Let C ∗
finitely supported functions; let C ∗
functions vanishing at infinitey.
cc(Γ) denote the ideal completion associated to the ideal cc(Γ) of
c0(Γ) denote the ideal completion associated to c0(Γ), the
Recall that Γ is amenable if there exist positive definite functions in cc(Γ) converging point-
wise to one; similarly Γ has the Haagerup property if there exist positive definite functions
in c0(Γ) converging pointwise to one. Since both cc(Γ) and c0(Γ) are translation invariant
ideals, our next result follows immediately from Theorem 3.2. Having already observed that
C ∗
r (Γ), the first statement is classical. The second statement is closely related to
the following fact: Γ has the Haagerup property if and only if it admits a c0-representation
weakly containing the trivial representation [6].
cc(Γ) = C ∗
Corollary 3.4. For a discrete group Γ we have: Γ is amenable if and only if C ∗(Γ) = C ∗
Γ has the Haagerup property if and only if C ∗(Γ) = C ∗
cc(Γ);
(cid:3)
c0(Γ).
Since C ∗
c0(Γ) admits a faithful c0-representation, we have an analogue of the fact that every
representation of an amenable group is weakly contained in the left regular representation.
Corollary 3.5. If Γ has the Haagerup property, then every unitary representation is weakly
contained in a c0-representation.2
(cid:3)
Jesse Peterson asked if Property (T) can be characterized in this context, and suggested
the following proposition. For the definition of Property (T) we refer to [2].
Proposition 3.6. A discrete group Γ has Property (T) precisely when the following condition
holds: ℓ∞(Γ) is the only translation invariant ideal D for which C ∗(Γ) = C ∗
D(Γ).
2This can also be deduced from the existence of a c0-representation weakly containing the trivial one.
NEW C∗-COMPLETIONS OF DISCRETE GROUPS
7
Proof. First, suppose that Γ has Property (T) and that D is a translation invariant ideal
for which C ∗
D(Γ) = C ∗(Γ). We must show that D = ℓ∞(Γ). But, by Theorem 3.2 there
exist positive definite functions hn ∈ D converging pointwise to one. Since Γ has Property
(T) they converge uniformly to one. Thus, some hn is bounded away from zero, and so is
invertible in ℓ∞(D).
Conversely, suppose that Γ does not have Property (T). Let
D = { f ∈ ℓ∞(Γ) : inf
s /∈F
f (s) = 0 for every finite F ⊂ Γ }.
One readily checks that D is a proper, translation invariant ideal in ℓ∞(Γ). We shall show
that C ∗
D(Γ) = C ∗(Γ). By Theorem 3.2 we must exhibit positive definite functions in D
converging pointwise to one. Since Γ does not have Property (T) there exists an unbounded,
conditionally negative type function ψ on Γ; the desired functions are hn = e−ψ/n.
(cid:3)
4. Quantum groups
Recall that a compact quantum group is a pair (A, ∆) where A is a unital C∗-algebra and
∆ : A → A ⊗ A is a unital ∗-homomorphism satisfying the following two conditions:
(1) (∆ ⊗ idA)∆ = (idA ⊗ ∆)∆, and
(2) ∆(A)(A ⊗ 1) and ∆(A)(1 ⊗ A) are dense subspaces of A ⊗ A.3
The map ∆ is the co-multiplication and the first property is called co-associativity.
Discrete groups provided an early source of examples of quantum groups.
Indeed, the
assignment ∆(s) = s ⊗ s on group elements determines a co-associative map
(1)
∆ : C[Γ] → C[Γ] ⊗ C[Γ],
and one can check that ∆(C[Γ])(C[Γ] ⊗ 1) = ∆(C[Γ])(1 ⊗ C[Γ]) = C[Γ] ⊗ C[Γ]. Thus, if
A is a C∗-algebra containing C[Γ] as a dense subalgebra and for which ∆ can be extended
continuously to a map A → A ⊗ A, then A is a compact quantum group. Every discrete
group gives rise to two canonical compact quantum groups: the universal property ensures
that A = C ∗(Γ) is a compact quantum group whereas Fell's absorption principle implies that
A = C ∗
r (Γ) is a compact quantum group. Anything between these extremes is considered
'exotic' (cf. [11]). The purpose of this section is to provide examples of such exotic compact
quantum groups.
Proposition 4.1. For every ideal D ⊳ ℓ∞(Γ), the ideal completion C ∗
quantum group.
D(Γ) is a compact
Proof. We shall show that the map ∆ in (1) extends continuously to C ∗
To this end, fix a faithful D-representation C ∗
the composite
D(Γ).
D(Γ) ⊂ B(H). By Remark 2.9, we can regard
D(Γ) → C ∗
D(Γ)⊗C ∗
∆ : C[Γ] → C ∗
D(Γ) ⊗ C ∗
D(Γ) ⊂ B(H ⊗ H)
as a D-representation and the universal property ensures that ∆ extends to C ∗
D(Γ).
(cid:3)
3All tensor products in this definition are spatial (cf. [5]).
8
NATHANIAL P. BROWN AND ERIK GUENTNER
Thus our task is to provide examples of groups Γ and ideals D ⊳ ℓ∞(Γ) for which C ∗(Γ) 6=
D(Γ) 6= C ∗
r (Γ). Though a bit ad hoc, our first examples are easy to handle.
C ∗
Proposition 4.2. Suppose that Γ has Property (T ) and that Λ⊳Γ is a non-amenable normal
subgroup of infinite index. Suppose D1 ⊳ ℓ∞(Γ) and D2 ⊳ ℓ∞(Γ/Λ) are ideals satisfying the
following conditions:
(1) D1 is proper and translation invariant;
(2) D2 is translation invariant;
(3) D1 contains the image of D2 under the inclusion ℓ∞(Γ/Λ) ⊂ ℓ∞(Γ).
Then C ∗
D1(Γ) is an exotic compact quantum group.
D1(Γ) 6= C ∗
Proof. We must show that C ∗(Γ) 6= C ∗
invariant, and Γ has Property (T) the first inequality follows from Proposition 3.6.
r (Γ). Since D1 is proper and translation
To prove the second inequality, suppose to the contrary that C ∗
D1(Γ) = C ∗
r (Γ). Applying
(2) of Corollary 2.9 we obtain a ∗-homomorphism
C ∗
r (Γ) = C ∗
D1(Γ) → C ∗
D2(Γ/Λ) → C ∗
r (Γ/Λ)
extending the homomorphism Γ → Γ/Λ.
C ∗
r (Γ/Λ) defines a character of C ∗
r (Γ) → C ∗
r (Λ) ⊂ C ∗
It follows that Λ is amenable – the composite
(cid:3)
r (Λ).
Remark 4.3. While the hypotheses of the previous proposition may seem a bit contrived,
examples are plentiful. An extension of Property (T) groups will again have Property (T)
[2, Proposition 1.7.6]. As for the ideals, taking D1 to be the ideal generated by the image of
D2 in ℓ∞(Γ) we have: if D2 is translation invariant, then so is D1; if D2 ⊂ c0(Γ/Λ), then D1
is proper.
Free groups also provide natural examples. We thank Rufus Willett for suggesting the
proof of the following result.
Proposition 4.4. Let F be a free group on two or more generators. There exists a p ∈ (2, ∞)
such that C ∗(F) 6= C ∗
ℓp(F) 6= C ∗
r (F).
Proof. Since F is not amenable, Proposition 2.12 implies that C ∗(F) 6= C ∗
p. We must find some p such that C ∗
ℓp(F) 6= C ∗
r (F).
ℓp(F) for all finite
Let S ⊂ F be the standard symmetric generating set and let · denote the corresponding
word length. A seminal result, first proved by Haagerup [10], states that for every n ∈ N,
hn(s) := e−s/n
is positive definite. Clearly hn → 1 pointwise. Fixing n, we have hn ∈ ℓpn(Γ) for sufficiently
large pn; indeed if pn is chosen so that S < epn/n, or equivalently Se−pn/n < 1, then
(e−s/n)pn =
Xs∈Γ
∞
Xk=1(cid:18)Xs=k
e−kpn/n(cid:19) ≤
∞
Xr=1(cid:0)Ske−kpn/n(cid:1) =
∞
Xr=1(cid:0)Se−pn/n(cid:1)k < ∞.
NEW C∗-COMPLETIONS OF DISCRETE GROUPS
9
Let πn : C ∗
ℓpn (Γ) → B(Hn) be the GNS representations corresponding to hn, and let vn ∈
Hn be the canonical cyclic vector. Since since hn(s) → 1 we see that kπn(s)vn − vnk → 0, for
all s ∈ Γ. Hence the direct sum representation ⊕πn weakly contains the trivial representation.
It follows that we cannot have C ∗
r (Γ) for all n – otherwise ⊕πn would be defined on
C ∗
r (Γ) and nonamenability prevents the trivial representation from being weakly contained
in any representation of C ∗
(cid:3)
ℓpn (Γ) = C ∗
r (Γ).
Remark 4.5. The previous proposition is not optimal; Higson, Ozawa and Okayasu [13] have
independently shown that the C ∗-algebras C ∗
ℓp(Fn) are mutually non-isomorphic. On the
other hand, extracting the crucial ingredients from the proof, we see that the phenomenon
presented there is very general.
Indeed, suppose that Γ is a non-amenable, a-T-menable
group admitting an N-valued conditionally negative type function ψ satisfying an estimate
of the following form: there exists C > 0 such that for every k we have
(2)
#{ s ∈ Γ : ψ(s) = k } ≤ C k.
Taking hn(s) = e−ψ(s)/n the above proof applies verbatim to show that C ∗
r (Γ) for
some p. This applies, for example, to infinite Coxeter groups – the word length function
corresponding to the standard Coxeter generators satisfies the hypothesis for ψ [3].
ℓp(Γ) 6= C ∗
Remark 4.6. Continuing the previous remark, suppose a non-amenable group Γ acts (cellu-
larly) on a CAT(0) cube complex X. The combinatorial distance d in the one skeleton of X
defines an N-valued conditionally negative type function on Γ by
ψ(x) = d(x0, s · x0),
where x0 is any arbitrarily chosen vertex in X [12].4 If Γ is finitely generated and the orbit
map s 7→ s · x0 : Γ → X is a quasi-isometric embedding then the inequality (2) is satisfied.
These two conditions hold in many common situations: by the Svarc-Milnor Lemma, they
are automatic if the action is proper and cocompact [4] and the complex is finite dimensional;
they also hold for the action of Thompson's group F or, more generally, a finitely generated
diagram group with Property B , on its Farley complex [1].
5. Topological Dynamical Systems
Let Γ be a discrete group, and let X be a compact Hausdorff space on which Γ acts by
homeomorphisms. Thinking of transformation groupoids, let Cc(X ⋊ Γ) denote the convo-
lution algebra of compactly supported functions on X × Γ. We shall represent elements of
this algebra as finite formal sums P fss, where each fs ∈ C(X); we shall view Γ as a subset
of the convolution algebra in the obvious manner.
Definition 5.1. Let D ⊳ ℓ∞(Γ) be an ideal. A ∗-representation π : Cc(X ⋊ Γ) → B(H) is a
D-representation if πΓ is a D-representation in the sense of Definition 2.1.
4While stated only for finite dimensional complexes, the proof given is valid in greater generality.
10
NATHANIAL P. BROWN AND ERIK GUENTNER
Definition 5.2. Let D ⊳ ℓ∞(Γ) be an ideal. Define a C∗-norm k · kD on Cc(X ⋊ Γ) by
(cid:13)(cid:13)(cid:13)X fss(cid:13)(cid:13)(cid:13)D
:= supn(cid:13)(cid:13)(cid:13)
π(cid:16)X fss(cid:17)(cid:13)(cid:13)(cid:13)
: π is a D-representationo ,
and let C ∗
D(X ⋊ Γ) denote the completion of Cc(X ⋊ Γ) with respect to k · kD.
As before, every D-representation extends uniquely to C ∗
D(X ⋊ Γ)
admits a faithful D-representation. Considering the universal properties, one sees that
ℓ∞(X ⋊ Γ) is the universal (or full) crossed product C∗-algebra, denoted C ∗(X ⋊ Γ). It is
C ∗
not clear whether the analogue of Proposition 2.11 holds in the present context.
D(X ⋊ Γ); further, C ∗
Recall that a function h : X × Γ → C is positive definite if for each finite set s1, . . . , sn ∈ Γ
and point x ∈ X, the matrix
[h(si.x, sis−1
j )]i,j ∈ Mn(C)
is positive (semi-definite); here x 7→ s.x denotes the action of s on X. Recall also that to
each positive definite h we can associate a completely positive Schur multiplier
on finite sums mh is given by the formula
mh : C ∗(X ⋊ Γ) → C ∗(X ⋊ Γ);
where, slightly abusing notation, we have written h(s) ∈ C(X) for the function x 7→ h(x, s)
[5, Proposition 5.6.16].
mh(cid:16)X fss(cid:17) =X fsh(s)s
Lemma 5.3. Suppose that D⊳ℓ∞(Γ) is a translation invariant ideal; suppose that h : X ×Γ →
C is positive definite, and that the function H(s) := kh(s)k belongs to D. Then, for every
state ϕ on C ∗(X ⋊ Γ), the GNS representation associated to ϕ ◦ mh is a D-representation.
Proof. Note that for f ∈ ℓ∞(Γ) we have, f ∈ D ⇐⇒ f ∈ D – this follows from the polar
decomposition f = uf , in which u ∈ ℓ∞(Γ) is the unitary of 'pointwise rotation'. Also, D
is hereditary in the sense that if 0 ≤ g ≤ f and f ∈ D, then g ∈ D. To see this, define
h ∈ ℓ∞(Γ) to be zero wherever f is, and h(s) = g(s)
f (s) otherwise; evidently g = hf ∈ D.
Now, fix two functions f , g ∈ C(X) and two group elements s, t ∈ Γ. Let v denote the
canonical image of f s in the GNS Hilbert space; similarly let w denote the image of gt. Since
the linear span of such elements is dense, it suffices to show πv,w ∈ D, where π denotes the
GNS representation. By the first paragraph, and our assumptions on D and H, it suffices
to show πv,w is bounded above by a constant times some translate of H. This, however, is
a straightforward calculation.
(cid:3)
With the previous lemma in hand, the proof of the following result is very similar to its
analog in the group case, Theorem 3.2 – use Schur multipliers to approximate arbitrary
states by vector states. Whereas Theorem 3.2 was an 'if-and-only-if' statement, we do not
know if the converse holds.
NEW C∗-COMPLETIONS OF DISCRETE GROUPS
11
Theorem 5.4. Let D ⊳ ℓ∞(Γ) be a translation invariant ideal. Assume there exist positive
definite functions hn : X ×Γ → C satisfying hn → 1 uniformly on compact sets and for which
each Hn(s) := khn(s)k belongs to D. Then C ∗(X ⋊ Γ) = C ∗
(cid:3)
D(X ⋊ Γ).
Definition 5.5. An action of Γ on X is amenable if there exist positive definite functions
hn ∈ Cc(X ⋊ Γ) such that hn → 1 uniformly on compact sets; it is a-T-menable if there exist
positive definite functions hn ∈ C0(X ⋊ Γ) such that hn → 1 uniformly on compact sets.
Remark 5.6. Just as every action of an amenable group is amenable, every action of an a-T-
menable group is a-T-menable. Indeed, if h ∈ c0(Γ) is positive definite, then a computation
confirms that the function h ∈ C0(X × Γ) defined by h(x, s) = h(s) is as well. The assertion
now follows easily from the definitions.
Definition 5.7. Let C ∗
supported functions on Γ; let C ∗
of functions vanishing at infinity.
cc(X ⋊Γ) denote the ideal completion associated to the ideal of finitely
c0(X ⋊ Γ) denote the ideal completion associated to the ideal
We draw several corollaries, the analogs of of Corollaries 3.4 and 3.5 in the group case.
Again, whereas Corollary 3.4 was an equivalence, the converse of the first corollary is open.
Corollary 5.8. Let Γ be a discrete group, acting on X.
C ∗(X ⋊ Γ) = C ∗
cc(X ⋊ Γ); if the action is a-T-menable, then C ∗(X ⋊ Γ) = C ∗
If the action is amenable then
c0(X ⋊ Γ). (cid:3)
Corollary 5.9. Let Γ be a discrete group, acting on X. If the action is amenable then every
covariant representation (that is, ∗-homomorphism C ∗(X ⋊ Γ) → B(H)) is weakly contained
in a cc-representation; if the action is a-T-menable then every covariant representation is
weakly contained in a c0-representation.
Proof. In the case of an amenable action, the hypothesis implies that C ∗(X ⋊Γ) = C ∗
and C ∗
replace cc(Γ) by c0(Γ) throughout.
cc(X ⋊Γ),
cc(X ⋊ Γ) has a faithful cc-representation. For an a-T-menable action systematically
(cid:3)
Corollary 5.10. Let Γ be a discrete group acting on X. Let π : C ∗(X ⋊ Γ) → B(H) be a
covariant representation for which πΓ weakly contains the trivial representation of Γ. We
have:
(1) the action is amenable if and only if Γ is amenable;
(2) the action is a-T-menable if and only if Γ is a-T-menable.
Proof. The 'if' statements are trivial (see Remark 5.6). For the 'only if' statements, let D
stand for the appropriate ideal, either cc(Γ) or c0(Γ). Corollary 5.8 implies that C ∗
D(X ⋊Γ) =
C ∗(X ⋊ Γ), so that we can form the composition
C ∗
D(Γ) → C ∗
D(X ⋊ Γ) = C ∗(X ⋊ Γ) → B(H).
The proof is completed by recalling that Γ is amenable (respectively, a-T-menable) if and
only if there is a cc(Γ) (respectively, c0(Γ))-representation of Γ weakly containing the trivial
representation.
(cid:3)
12
NATHANIAL P. BROWN AND ERIK GUENTNER
To close, we return to the result of Douglas and Nowak described in the introduction.
Suppose a discrete group Γ acts on X, and that µ is a quasi-invariant measure on X (in
other words, elements of Γ map µ-null sets to µ-null sets). For each element s ∈ Γ the
Radon-Nikodym derivative is the non-negative, measurable function ρs satisfying
ZX
f dµ =ZX
s.f ρsdµ,
for every measurable function f ; here, f 7→ s.f denotes the action of s on f . The ρs allow
one to construct a covariant representation of X ⋊Γ on the Hilbert space L2(X, µ): functions
in C(X) act by multiplication and elements s ∈ Γ act by the unitaries
Us(f ) := (s.f )ρ1/2.
Corollary 5.11. Let Γ be a discrete group acting on X, with quasi-invariant measure µ.
Suppose the representation of Γ on L2(X, µ) weakly contains the trivial representation. We
have:
(1) the action is amenable if and only if Γ is amenable;
(2) the action is a-T-menable if and only if Γ is a-T-menable.
In particular, such an action of a non-a-T-menable group can never be a-T-menable.
(cid:3)
This result, which follows immediately from the previous corollary, generalizes Theorem 3
and Corollary 4 of Douglas and Nowak [8] – the hypotheses of their results imply the existence
of a non-zero fixed vector in L2(X, µ), namely the square root of ρ or ρ, as appropriate. See
[8, Lemma 6].
For invariant probability measures we get an analogue of a well-known amenability result.
Corollary 5.12. Let Γ be a discrete group acting on X, with invariant probability measure
µ. The action is a-T-menable if and only if Γ is a-T-menable. In particular, no measure-
preserving action of a non-a-T-menable group can be a-T-menable.
Proof. The constant functions in L2(X, µ) are invariant for Γ.
(cid:3)
References
[1] G. Arzhantseva, V. Guba and M. Sapir, Metrics on diagram groups and uniform embeddings in a
Hilbert space, Comment. Math. Helv. 81 (2006), 911–929.
[2] B. Bekka, P. de la Harpe and A. Valette, Kazhdan's property (T ), New Mathematical Monographs,
11. Cambridge University Press, Cambridge, 2008. xiv+472 pp.
[3] M. Bozejko, T. Januszkiewicz and R. Spatzier, Infinite Coxeter Groups do not have Kazhdan's Property,
J. Operator Theory 19 (1988), 63–67.
[4] M. Bridson and A. Haefliger, Metric spaces of non-positive curvature. Grundlehren der mathematische
Wissenschaften, 319. Springer, Berlin, 1999.
[5] N.P. Brown and N. Ozawa, C ∗-algebras and finite-dimensional approximations. Graduate Studies in
Mathematics, 88. American Mathematical Society, Providence, RI, 2008. xvi+509 pp.
[6] P. Cherix, M. Cowling, P. Jolissaint, P. Julg and A. Valette, Groups with the Haagerup property.
Gromov's a-T-menability, Progress in Mathematics, 197. Birkhauser Verlag, Basel, 2001. viii+126 pp.
NEW C∗-COMPLETIONS OF DISCRETE GROUPS
13
[7] M. Cowling, U. Haagerup and R. Howe, Almost L2 matrix coefficients, J. Reine Angew. Math. 387
(1988), 97110.
[8] R. G. Douglas and P. W. Nowak, Hilbert C∗-modules and amenable actions, Studia Math. (to appear).
[9] E. Guentner and J. Kaminker, Exactness and uniform embeddability of discrete groups, J. London
Math. Soc. 70 (2004), 703718.
[10] U. Haagerup, An Example of a Non-Nuclear C ∗-algebra with the Metric Approximation Property,
Inven. Math. 50 (1979), 279–293.
[11] D. Kyed and P.M. Soltan, Property (T) and exotic quantum group norms, preprint, arXiv:1006.4044.
[12] G. Niblo and L. Reeves, Groups acting on CAT(0) cube complexes, Geom. Topol. 1 (1997), 1–7.
[13] R. Okayasu, Free group C∗-algebras associated with ℓp, preprint 2012.
Department of Mathematics, Penn State University, State College, PA 16802, USA
E-mail address: [email protected]
Department of Mathematics, University of Hawai'i at M¯anoa, Honolulu, HI 96822
E-mail address: [email protected]
|
1703.06732 | 1 | 1703 | 2017-03-20T13:32:25 | Banach-Alaoglu theorem for Hilbert $H^*$-module | [
"math.OA"
] | We provided an analogue Banach-Alaoglu theorem for Hilbert $H^*$-module. We construct a $\Lambda$-weak$^*$ topology on a Hilbert $H^*$-module over a proper $H^*$-algebra $\Lambda$, such that the unit ball is compact with respect to $\Lambda$-weak$^*$ topology. | math.OA | math |
BANACH-ALAOGLU THEOREM FOR HILBERT
H ∗-MODULE
ZLATKO LAZOVI ´C
Abstract. We provided an analogue Banach-Alaoglu theorem for Hilbert
H ∗-module. We construct a Λ-weak∗ topology on a Hilbert H ∗-module
over a proper H ∗-algebra Λ, such that the unit ball is compact with
respect to Λ-weak∗ topology.
Keywords: H ∗-algebra, H ∗-module, compact set.
MSC(2010): Primary: 46H25; Secondary: 57N17, 46A50.
1. Introduction
A Hilbert H ∗-module L over an H ∗-algebra Λ is a right Λ-module which
possesses a τ (Λ)-valued product, where τ (Λ) = {ab a, b ∈ Λ} is the trace-
class. At the same time, H is a Hilbert space with the inner product given
by the action of the trace on the τ (Λ)-valued product.
The notion of H ∗-module is introduced by Saworotnow in [8] under the
name of generalized Hilbert space. It has been studied by Smith [10], Giellis
[4] Molnar [6], Cabrera [3], Martinez et al.
[3], Baki´c and Guljas [2] and
others. Saworotnow has proved that the trace-class in a proper H ∗-algebra
Λ has pre-dual. For Hilbert H ∗-modules, a generalization of Reisz theorem
holds, i.e. for each bounded Λ-linear functional on L, there is xf ∈ L such
that f (x) = [xf , x] for all x ∈ L.
Paschke [7] showed that self-dual Hilbert W ∗-modules are dual Banach
spaces and found topology on module such that the unit ball is compact.
In the present paper we find a topology on Hilbert H ∗-module L over a
proper H ∗-algebra Λ such that the unit ball in L is compact with respect to
this topology.
2. Basic notations and preliminary results
We recall that an H ∗-algebra is a complex associative Banach algebra with
an inner product h·, ·i such that ha, ai = a2 for all a ∈ Λ, and for each
a ∈ Λ there exists some a∗ ∈ Λ such that hab, ci = hb, a∗ci and hba, ci =
hb, ca∗i for all b, c ∈ Λ. The adjoint a∗ of a need not be unique (see [1]).
Proper H ∗-algebra Λ is an H ∗-algebra which satisfies aΛ = 0 ⇒ a = 0 (or
Λa = 0 ⇒ a = 0). An H ∗-algebra Λ is proper if and only if each a ∈ Λ has
Date: Received: , Accepted: .
∗Corresponding author.
1
2
F. AUTHOR
a unique adjoint a∗ ∈ Λ (see [1, Theorem 2.1]). An H ∗-algebra is simple
algebra if it has no nontrivial closed two-sided ideals.
The trace-class in a proper H ∗-algebra Λ is defined as the set τ (Λ) =
{ab a, b ∈ Λ}. The trace-class is selfadjoint ideal of Λ and it is dense in
Λ, with norm τ (·). The norm τ is related to the given norm · on Λ by
τ (a∗a) = a2 for all a ∈ Λ. There exists a continuous linear form sp on
τ (Λ) (trace) satisfying sp(ab) = sp(ba) = ha∗, bi . In particular, sp(a∗a) =
sp(aa∗) = ha, ai = a2 = τ (a∗a).
A Hilbert Λ-module is a right module L over a proper H ∗-algebra Λ
provided with a mapping [·, ·] : L × L → τ (Λ), which satisfies the following
conditions:
[x, αy] = α[x, y]; [x, y + z] = [x, y] + [x, z]; [x, ya] = [x, y] a;
[x, y]∗ = [y, x]; L is a Hilbert space with the inner product hx, yi = sp ([x, y])
for all α ∈ C, x, y, z ∈ L, a ∈ Λ and for all x ∈ L, x 6= 0 there is a ∈ Λ,
a 6= 0 such that [x, x] = a∗a.
A Λ-linear functional on L is a mapping w′ : L → τ (Λ) such that w′(xa +
yb) = w′(x)a + w′(y)b for all x, y ∈ L, a, b ∈ Λ. It is bounded if there exists
c > 0 such that τ (w′(x)) ≤ cx for all x ∈ L.
In this case we define
w′ = supp{c τ (w′(x)) ≤ cx for all x ∈ L}. The norm space of all
bounded Λ-linear functional f on L we denoted by L′.
Let R(Λ) be the set of all bounded linear operators S on Λ such that
S(xy) = (Sx)y for all x, y ∈ Λ and let C(Λ) be the closed subspace of R(Λ)
generated by the operators of the form La : x → ax, a ∈ Λ.
We now state some theorems which will be necessary for the proof of main
results.
Theorem 2.1. [8, Theorem 3] Each bounded Λ-linear functional f on L is
of the form [xf , ·] for some xf ∈ L.
Theorem 2.2. [9, Lemma 1] If a ∈ Λ then the mapping fa : S → sp(Sa),
defined on C(Λ), is a bounded linear functional and fa = τ (a).
Theorem 2.3. [9, Theorem 1] Each bounded linear functional on C(Λ) is of
the form fa for some a ∈ τ (Λ). The correspondence a ↔ fa is an isometric
isomorphism between τ (Λ) and C(Λ)∗. Also, τ (Λ) is a Banach algebra.
For more details, we refer to [8, 9, 1, 2, 5].
3. Results
Let L be a Hilbert H ∗-module over a proper H ∗-algebra Λ and let B1(L) is
the unit ball in L. We construct a topology on W ′ such that the unit ball
in L′ is compact with respect to this topology. Define Λ′-weak∗ topology on
L′ with the base
Lw′
0,x1,..,xn,S1,..,Sn,δ = (cid:8)w′ ∈ L′ (cid:12)
(cid:12)sp(Sj(w′(xj) − w′
0(xj)))(cid:12)
(cid:12) < δ, j = 1, .., n(cid:9) ,
for w′
0 ∈ L′, xj ∈ L, Sj ∈ C(Λ), and δ > 0.
SHORT TITLE
3
The main result of this paper, the compactness of the unit ball, will be
proven in Theorem 3.2 and its corollary. Before that, we state and prove a
useful lemma.
Lemma 3.1. Let Λ be a H ∗-algebra and let S ∈ C(Λ). Then the operator
Sa : Λ → Λ, Sa(x) = S(ax) belongs to C(Λ).
Proof. From
Sa(λx) = S(aλx) = λS(ax) = λSa(x), Sa(xy) = S(axy) = S(ax)y = Sa(x)y
and
Sa(x) = S(ax) ≤ S · ax ≤ S · a · x,
it follows that Sa belongs to R(Λ). If Lan converges to S, then from
(Lana − Sa)(x) = (Lana(x) − Sa(x) = anax − S(ax)
= Lan (ax) − S(ax) ≤ Lan − S · ax
≤ Lan − S · a · x,
it follows that Lana converges to Sa. Thus Sa ∈ C(Λ).
(cid:3)
Theorem 3.2. The set
K = (cid:8)w′ ∈ L′ w′(x) ≤ 1 for all x ∈ B1(L)(cid:9)
is compact in Λ′-weak∗ topology.
Proof. The neighborhood B1(L) is absorbing because for each x ∈ L, x 6= 0,
there exists a number γ(x) = x > 0 such that x ∈ γ(x)B1(L). For all
w′ ∈ K, it holds w′(x) ≤ 1, x ∈ B1(L), hence w′(x) ≤ x, x ∈ L.
According to Banach-Alaoglu theorem and Theorem 2.3, the set Dx =
{a ∈ Λa ≤ γ(x)} is compact in weak∗-topology on τ (Λ) given by semi-
norms pS(·) = sp(S(·)), S ∈ C(Λ). Let τ1 be the product weak∗-topology
on P = Qx Dx, the cartesian product of all Dx, one for each x ∈ L. Since
each Dx is weak∗-compact, it follows, from Tychonoff's theorem, that P is
τ1-compact.
From definition of P we have
P = {f : L → τ (Λ) f (x) ≤ x for all x ∈ L}.
The set P can contain Λ-nonlinear functionals.
It is clear that K ⊂ L′ ∩ P . It follows that K inherits two topologies:
one from L′ (its Λ′-weak∗ topology, to which the conclusion of the theorem
refers) and the other, τ1, from P . We will see that these two topologies
coincide on K, and that K is a closed subset of P .
We now prove that topologies τ1 and Λ′-weak∗ coincide on K. Fix some
0 ∈ K. Then
w′
W1 = { w′ ∈ L′ (cid:12)(cid:12)sp(Sj(w′(xj))) − sp(Sj(w′
0(xj)))(cid:12)(cid:12) < δ,
x1, x2, ..., xn ∈ L, S1, ..., Sn ∈ C(Λ)}
4
and
F. AUTHOR
W2 = {f ∈ P sp(Sj(f (xj))) − sp(Sj(Λ0(xj))) < δ,
x1, x2, ..., xn ∈ L, S1, ..., Sn ∈ C(Λ)}
are local bases in (L′, Λ-weak∗) and (P, τ1), respectively. From K ⊂ L′ ∩ P
we have K ∩ W1 = K ∩ W2, so topologies coincide on K.
Suppose w′
0 is in the τ1-closure of K.
If S is from C(Λ), then, from
Lemma 3.1, operator S(a·) : Λ → Λ belongs to C(Λ) for all a ∈ Λ. For
any a, b ∈ Λ, S ∈ C(Λ), x, y ∈ L, ε > 0, there is Λ-linear w′ ∈ K from
0). Therefore,
Vxa+yb,w′
it holds
0,S(b·),ε (τ1-neighborhood of w′
0,S(a·),ε ∩ Vy,w′
0,S,ε ∩ Vx,w′
(cid:12)(cid:12)sp(S[a(w′(x) − w′
0(x))])(cid:12)(cid:12) < ε,
(cid:12)sp(S[(w′(xa + yb) − w′
(cid:12)
(cid:12)(cid:12)sp(S[b(w′(y) − w′
0(xa + yb))])(cid:12)
(cid:12) < ε.
0(y))])(cid:12)(cid:12) < ε,
We have
sp(S(w′
= sp (S[(w′
0(xa + yb) − w′
0(x)a − w′
0(y)b))
0(xa + yb) − w′(xa + yb)) − (w′
0(x)a − w′(x)a)
− (w′
0(y)b − w′(y)b)] )
0(xa + yb) − w′(xa + yb)) − S[(w′
0(y) − w′(y))b] )
0(xa + yb) − w′(xa + yb)))(cid:12)
= sp( S(w′
− S[(w′
(cid:12)sp(S(w′
+ (cid:12)
≤ (cid:12)
(cid:12)sp(S[(w′
0(y) − w′(y))b])(cid:12)
(cid:12)
0(x) − w′(x))a]
(cid:12) + (cid:12)
(cid:12)sp(S[(w′
0(x) − w′(x))a])(cid:12)
(cid:12)
= (cid:12)(cid:12)sp(S(w′
0(xa + yb) − w′(xa + yb)))(cid:12)(cid:12) + (cid:12)(cid:12)sp(aS[(w′
0(x) − w′(x))])(cid:12)(cid:12)
+ (cid:12)(cid:12)sp(bS[(w′
≤ ε + ε + ε = 3ε.
0(y) − w′(y))])(cid:12)(cid:12)
Since ε > 0 was arbitrary, we see that
sp(S(w′
0(xa + yb))) = sp(S(w′
0(x)a) + S(w′
0(y)b))
for all S ∈ C(Λ), i.e.
sp(S(w′
0(xa + yb) − w′
0(x)a − w′
0(y)b)) = 0
for all S ∈ C(Λ). For S = Lw′
0(x)a − w′
0(xa + yb) − w′
sp((w′
0(xa+yb)−w′
0(y)b))∗(w′
0(x)a−w′
0(xa + yb) − w′
0(y)b) we have
0(x)a − w′
0(y)b)) = 0.
Hence w′
Λ-linear.
0(xa + yb) = w′
0(x)a + w′
0(y)b for all x, y ∈ L, a, b ∈ Λ, so w′
0 is
Let x ∈ B1(L). For arbitrary ε > 0 and S ∈ CΛ there is w′ ∈ B1(L′) such
that sp(S(w′
sp(S(w′
0(x))) − sp(S(w′(x))) < ε. Hence
0(x))) < sp(S(w′(x))) + ε ≤ S · w′(x) + ε ≤ S + ε.
Next, from Theorem 2.2 we have
SHORT TITLE
5
w′
0(x) = τ (w′
0(x)) = fw′
0(x) =
sup
S∈C(Λ), S=1
sp(S(w′
0(x)))
≤
sup
S + ε ≤ 1 + ε
S∈C(Λ), S=1
for arbitrary ε > 0. Thus w′
0(x) ≤ 1.
We have proven that w′
0 ∈ B1(L′), and that B1(L′) is a closed subset of
P .
Since P is τ1-compact, B1(L′) is a closed subset of P , and τ1 and Λ′-weak∗
topology coincide on B1(L′), we have that B1(L′) is Λ′-weak∗ compact. (cid:3)
Corollary 3.3. The unit ball B1(L) in L is compact in Λ-weak∗ topology
with the base
Wx0,y1,...,yn,S1,...,Sn,δ = {x ∈ L sp(Sj([x, yj] − [x0, yj])) < δ, j = 1, ..., n} ,
for x0, yj ∈ L, Sj ∈ C(Λ), δ > 0.
Proof. For each w ∈ L′ there is y ∈ L such that w′(x) = [y, x] for all x ∈ L
(Theorem 2.1), so from Theorem 3.2 it follows that the unit ball B1(L) in
L is compact in given topology.
(cid:3)
Acknowledgement. The author was supported in part by the Ministry of
education and science, Republic of Serbia, Grant 174034.
References
[1] W. Ambrose, Structure theorems for a special class of Banach algebras, Trans. Amer.
Math. Soc. 57 (1945), 364 -- 386.
[2] D. Baki´c and B. Guljas, Operators on Hilbert H ∗-modules, J. Operator theory 46
(2001), 123 -- 137.
[3] M. Cabrera, J. Martinez and A. Rodr´ıguez, Hilbert modules revisited: Orthonormal
bases and Hilbert-Schmidt operators, Glasgow Math. J 37 (1995), 45 -- 54.
[4] G. R. Giellis, A characterization of Hilbert modules, Proc. Amer. Math. Soc. 36
(1972), 440 -- 442.
[5] D. Ilisevi´c, On redundance of one of the axioms of generalized normed space, Glasnik
Matematicki 37 (2002), 135 -- 141.
[6] L. Moln´ar, Modular bases in a Hilbert A-module, Czechoslovak Math. J. 42 (1992),
649 -- 656.
[7] W. L. Paschke, Inner product modules over B ∗-algebras, Trans. Amer. Math. Soc.
182 (1973), 443 -- 468.
[8] P. P. Saworotnow, A generalized Hilbert space, Duke Math. J. 35 (1968), 191 -- 197.
[9] P. P. Saworotnow, Trace class and centralizers of an H ∗-algebra, Proceedings of the
American Mathematical Society 26, no. 1 (Sep., 1970), 101 -- 104.
[10] J. F. Smith, The structure of Hilbert modules, J. London Math. Soc. 8 (1975), 741 --
749.
(Zlatko Lazovi´c) Faculty of Mathematics, University of Belgrade, 11 000
Belgrade, Serbia
E-mail address: [email protected]
|
1303.1424 | 3 | 1303 | 2013-10-21T00:24:34 | A II$_1$ factor approach to the Kadison-Singer problem | [
"math.OA"
] | We show that the Kadison-Singer problem, asking whether the pure states of the diagonal subalgebra $\ell^\infty\Bbb N\subset \Cal B(\ell^2\Bbb N)$ have unique state extensions to $\Cal B(\ell^2\Bbb N)$, is equivalent to a similar statement in II$_1$ factor framework, concerning the ultrapower inclusion $D^\omega \subset R^\omega$, where $D$ is the Cartan subalgebra of the hyperfinite II$_1$ factor $R$, and $\omega$ is a free ultraflter. While we do not settle the problem in this latter form, we prove that if $A$ is any singular maximal abelian subalgebra of $R$, then the inclusion $A^\omega \subset R^\omega$ does satisfy the Kadison-Singer property. | math.OA | math | A II1 FACTOR APPROACH
TO THE KADISON-SINGER PROBLEM
SORIN POPA
Dedicated to R.V. Kadison and I.M. Singer
Abstract. We show that the Kadison-Singer problem, asking whether the pure
states of the diagonal subalgebra ℓ∞N ⊂ B(ℓ2N) have unique state extensions to
B(ℓ2N), is equivalent to a similar statement in II1 factor framework, concerning the
ultrapower inclusion Dω ⊂ Rω, where D is the Cartan subalgebra of the hyperfinite
II1 factor R (i.e. a maximal abelian ∗-subalgebra of R whose normalizer generates R,
e.g. D = L∞([0, 1]Z) ⊂ L∞([0, 1]Z ⋊ Z = R), and ω is a free ultraflter. Instead, we
prove here that if A is any singular maximal abelian ∗-subalgebra of R (i.e., whose
normalizer consists of the unitary group of A, e.g. A = L(Z) ⊂ L∞([0, 1]Z) ⋊ Z = R),
then the inclusion Aω ⊂ Rω does satisfy the Kadison-Singer property.
3
1
0
2
t
c
O
1
2
]
.
A
O
h
t
a
m
[
3
v
4
2
4
1
.
3
0
3
1
:
v
i
X
r
a
0. Introduction
A famous problem posed by Kadison and Singer in the late 1950s ([KS]) asks
whether any pure state on the diagonal ℓ∞N of the algebra B(ℓ2N), of all linear
bounded operators on the Hilbert space ℓ2N, has unique state extension to B(ℓ2N).
We will refer to this property of the inclusion of algebras ℓ∞N ⊂ B(ℓ∞N) as the
Kadison-Singer property. As already pointed out in [KS], it is equivalent to the
following property for operators on the Hilbert space, known as the paving property:
if x ∈ B(ℓ2N) has only 0 on the diagonal, then for any ε > 0, there exists a
finite partition of N into subsets Y1, ..., Yn, such that if pi ∈ ℓ∞N denotes the
characteristic function of Yi, viewed as a diagonal operator operator on ℓ2N, then
kΣn
i=1pixpik ≤ εkxk. It was later shown in [An1, An2] that this is in fact equivalent
to the following finite dimensional version of the property, known as the uniform
paving property: for any ε > 0, there exists n = n(ε) such that for any m and any
Supported in part by NSF Grant DMS-1101718 and a Simons Fellowship
1
Typeset by AMS-TEX
2
SORIN POPA
x ∈ B(ℓ2
m) with 0 on the diagonal, there exists a partition of {1, 2, ..., m} into n sets
Yi, such that the corresponding diagonal operators pi satisfy kΣipixpik ≤ εkxk.
The Kadison-Singer problem has attracted much interest over the years, proving
to have deep connections to a large number of fields of mathematics, with interesting
equivalent re-formulations in harmonic analysis, frame theory, discrepancy theory,
etc. Several partial results have been obtained so far (see e.g.
[A1], [A2], [AkA],
[BT], [BeHKW], [We], etc), showing for instance that certain classes of operators
in B(ℓ2N) can indeed be paved. We refer the reader to [CaFTW] for a beautiful,
comprehensive account on this problem, and on its interdisciplinary aspects.
In this paper, we attempt a new approach to the problem, based on a reformula-
tion in II1 factor framework. Recall that a II1 factor is a von Neumann algebra M
that is infinite dimensional, has trivial center and a completely additive trace state
τ . Any maximal abelian ∗-subalgebra (MASA) A of a II1 factor M is diffuse (i.e.
has no atoms) and if A is also countably generated, then A ≃ L∞([0, 1]) ([vN]).
algebra L∞([0, 1]), with the role of the Lebesgue integral R · dµ played by the
positive functional τ : M → C which satisfies τ (1) = 1 (it is a state) and τ (xy) =
τ (yx), ∀x, y ∈ M (it is a trace). A specific type of (non-commutative) analysis
has been developed in this framework, often exploiting the interplay between the
operator norm and the Hilbert-norm implemented by the trace, as well as ergodicity
properties of the Ad-action of the unitary group of M . One should note that the
algebra Mm×m(C), of m by m matrices with complex entries (≃ B(ℓ2
m)), has both
a trace state (given by the normalized trace tr) and is a factor, but it is finite
dimensional. However, inductive limits and ultraproducts of these algebras give
rise to II1 factors.
II1 factors can in fact be viewed as non-commutative versions of the function
Thus, the most "basic" example of a II1 factor is the hyperfinite II1 factor
R of Murray and von Neumann, defined as the infinite tensor product (R, τ ) =
⊗k(M2×2(C), tr)k. By [MvN2], R is in fact the unique approximately finite di-
mensional II1 factor, and by [C1] it is even the unique amenable II1 factor. So
R can be represented in many different ways, for instance as the group measure
space II1 factor L∞(X) ⋊ Γ, associated with a free ergodic measure preserving ac-
tion of a countable amenable group Γ on a probability space (X, µ). In particular,
R = L∞([0, 1]Z) ⋊ Z, where Z y X = [0, 1]Z is the Bernoulli action. When viewed
this way, R has D = L∞(X) as a natural Cartan subalgebra, i.e. a MASA D ⊂ R
whose normalizer generates R. By [CFW], [OW] the Cartan subalgebra of R is in
fact unique, up to conjugacy by an automorphism of R. We may thus represent
D ⊂ R as the infinite tensor product ⊗k(D2)k ⊂ ⊗k(M2×2(C))k, where D2 is the
diagonal subalgebra in M2×2(C).
But the hyperfinite II1 factor R also has MASAs A ⊂ R whose normalizer is
KADISON-SINGER PROBLEM
3
trivial, i.e. the only unitary elements u ∈ R normalizing A, uAu∗ = A, are the
unitaries in A. Such MASAs are called singular and their existence was discovered
in [D1]. A typical example of singular MASA is given by the subalgebra L(Z) ⊂ R,
generated by the canonical unitary implementing the Bernoulli action Z y [0, 1]Z,
in the above representation of the hyperfinite factor R = L∞([0, 1]Z) ⋊ Z.
There is an ultraproduct procedure of constructing II1 factors from a free ultrafil-
ter ω on N and a sequence of factors (Mm, τ ), with Mm either II1, or finite dimen-
sional with dimMm ր ∞ (see [W], [F]). The initial motivation for our work has
been the observation that the Kadison-Singer property for ℓ∞N ⊂ B(ℓ2N), as well
as its paving version, are equivalent to the analogue statements for the ultrapower
inclusions Dω ⊂ Rω, respectively ΠωDm ⊂ ΠωMm×m(C). Paving here means that
if x ∈ Rω (resp. x ∈ ΠωMm×m(C)) has trace preserving expectation onto Dω (resp.
ΠωDm) equal to 0, then for any ε > 0, there exists a partition of 1 with finitely
many projections p1, ..., pn in Dω (resp. ΠωDm), such that kΣn
The operator norm of an element y in a II1 factor can be calculated by the
formula kyk = limn τ ((y∗y)2n)1/2n. So in order to pave x one needs to control the
"higher moments" τ ((y∗y)n) for y = Σipixpi. Our idea here is to approach such
calculations by using a technique developed in [P6], which consists of building the
paving pi by patching together small, "infinitesimal" pieces of projections, with
"incremental" control of the moments.
Ideally, one wants to build the partition
pi so that to be "free independent" with respect to the given x, because then the
paving diminishes the operator norm by √ε if the mesh of the partition is less than
ε, due to norm calculations in [V2].
i=1pixpik ≤ εkxk.
In the case of the Cartan subalgebra D ⊂ R, the independence "breaks" after
the 3rd moment, more precisely we show that given any x ∈ Rω ⊖ Dω, Dω contains
finite partitions with projections pi that are 3-independent to x, but if x normalizes
D then xux∗u∗ = u∗xux∗, for any u ∈ Dω, so 4-independence may fail in general.
Nevertheless, our approach does provide "free paving" for any ultrapower Aω ⊂
Rω of a singular MASA A ⊂ R, in fact for any ultraproduct of singular MASAs in
II1 factors (N.B.: by [P3] any II1 factor contains singular MASAs). Note that this
result provides the first case when the Kadison-Singer property is established for a
MASA in an infinite dimensional von Neumann factor.
0.1. Theorem (Kadison-Singer for ultrapowers of singular MASAs). Let
Am ⊂ Mm, m ≥ 1, be a sequence of singular MASAs in II1 factors and denote
A = ΠωAm ⊂ ΠωMm = M, their ultraproduct, over a free ultrafilter ω on N. Then
A ⊂ M satisfies the Kadison-Singer property, i.e. any pure state on A has a unique
state extension to M. Moreover, A ⊂ M has the uniform paving property: if x ∈ M
has 0-expectation on A, then ∀ε > 0, ∃ p1, ..., pn partition of 1 with projections in
4
SORIN POPA
A, with n ≤ Cε−6 for some universal constant C, such that kΣn
i=1pixpik ≤ εkxk.
As we mentioned before, the way we prove the above result is by showing that
given any x ⊥ A, there exists a diffuse abelian subalgebra B0 ⊂ A which is free in-
dependent to {x, x∗}, i.e. any alternating word Πk
i=1uixi, with letters xi ∈ {x, x∗},
ui ∈ B0 ⊖ C, has 0-trace. The presence of "asymptotic freeness" in a MASA
A ⊂ M characterizes in fact singularity, and for it to be satisfied, asymptotic 4-
independence is actually sufficient (B0 ⊂ Aω is n-independent to X ⊂ M ω ⊖ Aω if
τ (Πk
i=1uixi) = 0, ∀k ≤ n, ui ∈ B0 ⊖ C, xi ∈ X). In turn, we will show in Section 3
and 5.3.1 that existence of asymptotic 3-independence holds in any MASA.
0.2. Theorem (Characterizations of singularity for MASAs). Let A be a
MASA in a II1 factor M . The following are equivalent:
1◦ A is singular in M ;
2◦ Aω is maximal amenable in M ω;
3◦ Aω is maximal among the ∗-subalgebras P ⊂ M ω that contain Aω and are
countably generated both as a left and right Aω-modules (i.e., ∃X ⊂ P countable
such that spAωX and spXAω are dense in P ).
4◦ Given any countable set X ⊂ M ω ⊖ Aω, there exists B0 ⊂ Aω diffuse such
that B0, X are free independent relative to Aω.
5◦ Given any self-adjoint element x ⊂ M ⊖ A, there exists B0 ⊂ Aω diffuse such
that B0, {x} are 4-independent.
The paper is organized as follows: In Section 1 we recall a classic result from
[KS], on the equivalence between the unique state extension of pure states from
ℓ∞N to B(ℓ2N) and the paving property, as well as other basic facts. In Section 2
we prove the equivalence between the Kadison-Singer problem and several similar
statements in II1 factor framework. In Section 3 we show that given any MASA
A in a II1 factor M , one can pave any finite set X ⊂ M ⊖ A with respect to the
L2-norm given by the trace. This result has already been shown in ([P1]; see also
A.1 in [P5]), but we give it here a different proof (which gives better estimates of
the paving size), by showing that A contains finite partitions of arbitrary small
mesh that are approximately 2-independent to X. In Section 4 we prove Theorem
0.1 (as Corollary 4.3). We do this by utilizing the L2-paving from Section 3 and
the "incremental patching method" from [P6].
In Section 5 we derive Theorem
0.2 (as Theorem 5.2.1) and obtain several related results, including existence of
approximate 3-independence in arbitrary MASAs (see Theorem 5.3.1). We also
formulate a conjecture generalizing Kadison-Singer (see 5.5.1).
KADISON-SINGER PROBLEM
5
While we made an effort to make this paper as self-contained as possible, for the
most basic facts on von Neumann algebras and II1 factors, we refer the reader to
the classic books [D2], [KR].
This work was completed during my stay at the Jussieu Math Institute in Paris,
during the year 2012-2013. I want to gratefully acknowledge A. Connes, G. Pisier,
G. Skandalis and S. Vassout, for their kind hospitality and support.
Note added in the proof. The present paper has been posted on the arXiv on
March 22nd, as arXiv:13031424. Since then, A. Marcus, D. Spielman and N. Srivas-
tava have posted the paper Interlacing Families II: Mixed Characteristic Polynomi-
als and the Kadison-Singer Problem (arXiv:1306.3969), where they solve the classic
Kadison-Singer problem in the affirmative. They do this by settling a finite dimen-
sional version of the paving property emphasized in [AkA] and [We]. Note that, due
to the equivalent re-formulation of the problem established in Theorem 2.2 below,
their result also implies that the inclusions Dω ⊂ Rω and ΠωDm ⊂ ΠωMm×m(C)
have the Kadison-Singer property.
1. Preliminaries
We recall in this section a result from [KS], showing that pure states on a maximal
abelian von Neumann subalgebra A of a von Neumann algebra M have unique
state extensions to M if and only if all elements in M have a certain "paving
property" relative to A. For the reader's convenience, we have included a proof. It
is essentially the original one from [KS], but explained in more modern terms, and
adding the reformulation of paving in terms of "relative Dixmier property". We
also introduce some necessary terminology and prove some basic related results.
1.1. Notation. Let M be a von Neumann algebra and A ⊂ M a maximal abelian
∗-subalgebra (hereafter abbreviated MASA) in M. If x ∈ M then we denote by
CA(x) the norm closure of the convex hull of the set {uxu∗ u ∈ U(A)}. Also,
given a finite n-tuple of unitaries V = (v1, ..., vn) in A and y ∈ M, we denote
TV (y) = n−1Σn
i ∈ CA(y). Note right away that the commutativity of A
implies TU (TV (y)) = TV (TU (y)) for any two such tuples U, V . Also, kTU (y)k ≤ kyk
and TU (a1ya2) = a1TU (y)a2, ∀a1, a2 ∈ A (i.e., the maps TU are A-bimodular).
1.2. Theorem (Kadison-Singer [KS]). If A ⊂ M is a MASA in a von Neu-
mann algebra M, then the following conditions are equivalent:
(1.2.1) Any pure state on A has a unique pure state extension to M.
(1.2.2) CA(x) ∩ A 6= ∅, ∀x ∈ M.
(1.2.3) CA(x) ∩ A is a single point set {EA(x)}, ∀x ∈ M.
i=1viyv∗
6
SORIN POPA
(1.2.4) For all x ∈ M and all ε > 0 there exists a finite partition of 1 with projec-
tions qk ∈ A such that d(Σkqkxqk, A) ≤ ε.
(1.2.5) For all x ∈ M, there exists a unique element E(x) ∈ A with the property
that ∀ε > 0, ∃qk ∈ P(A) a finite partition of 1 such that kΣkqkxqk − E(x)k ≤ ε.
Moreover, if these conditions are satisfied then E(x) = EA(x) and the map EA
satisfies the following additional properties:
w
∩ A = {EA(x)}, ∀x ∈ M.
(i) CA(x)
(ii) EA is the unique conditional expectation of M onto A.
(iii) Given any pure state ψ on A, ψ ◦ EA is the unique state extension of ψ to M,
and it is a pure state.
Proof. The implication (1.2.3) =⇒ (1.2.2) is trivial. If (1.2.2) is satisfied and a, b ∈
CA(x) are distinct, then there exist tuples U , V such that kTU (x)− ak ≤ ka− bk/4
and kTV (y)−bk ≤ ka−bk/4. But kTV (TU (x))−ak = kTV (TU (x)−a)k ≤ kTU (x)−ak
and kTU (TV (x)) − bk = kTU (TV (x) − b)k ≤ kTV (x) − bk. Thus we have
ka − bk ≤ kTV (TU (x)) − ak + kTV (TU (x)) − TU (TV (x))k + kTU (TV (x)) − bk
= kTV (TU (x)) − ak + kTU (TV (x)) − bk ≤ ka − bk/4 + ka − bk/4 = ka − bk/2,
a contradiction. This proves (1.2.2) =⇒ (1.2.3).
A similar argument shows that (1.2.4) and (1.2.5) are equivalent.
Assuming now (1.2.3), we prove (1.2.5) as well as the properties (i) − (iii). Let
x ∈ M, kxk ≤ 1, and ε > 0. Let u1, ..., un ∈ A be so that kTU (x) − EA(x)k ≤ ε/2,
where U = (u1, ..., un). For each i = 1, ..., n let {eij}j ∈ A be spectral projections
of ui such that if we denote vi = Σjλijeij , then kui − vik ≤ ε/4. Thus, if we denote
V = (v1, ..., vn), then kTU (x) − TV (x)k ≤ ε/2 and hence kTV (x) − EA(x)k ≤ ε.
By taking into account that if {qk}k denotes a relabeling of the set of projections
{eij}i,j, then ΣkqkTv(x)qk = Σkqkxqk, we thus get
kΣkqkxqk − EA(x)k = kΣkqk(Tv(x) − EA(x))qkk ≤ kTv(x) − EA(x)k ≤ ε,
j=1, w = Σn
proving the existence part of (1.2.5). Since any element of the form Σjpjxpj, with
p1, ..., pn a partition with projections in A, is of the form TW (x) ∈ CA(x), where
W = (wj−1)n
i=1λ(i−1)pi, where λ = exp(2πi/n), by the uniqueness in
(1.2.3) we get the uniqueness part in (1.2.5) and that E(x) = EA(x). We have also
implicitly shown that (1.2.5) =⇒ (1.2.2).
Since A is abelian (thus amenable), there exists a conditional expectation E :
M → A (obtained by taking a Banach limit of appropriate averages TU ). By
KADISON-SINGER PROBLEM
7
(1.2.3), for any fixed x ∈ M and any ε > 0, there exists a tuple V = (v1, ..., vn) in
A such that fixed kTV (x) − EA(x)k ≤ ε. Thus
kE(x) − EA(x)k = kTV (E(x)) − EA(x)k
= kE(TV (x) − EA(x)) ≤ kTV (x) − EA(x)k ≤ ε.
w
Fix now x0 ∈ M and let y0 ∈ CA(x)
Since ε > 0 was arbitrary, this shows that E(x) = EA(x), ∀x ∈ M, proving (ii).
∩ A and {Uι}ι∈I be a net of tuples of
unitaries in A such that the weak limit of {TUι (x0)}ι is equal to y0. Let Limι
be a Banach limit over ι and for each x ∈ M denote Φ(x) = LimιTUι (x). Then
Φ : M → M is linear, positive, A-bimodular, Φ(a) = a, ∀a ∈ A, and Φ(x0) = y0.
But then E(x) = EA(Φ(x)) is a conditional expectation of M onto A satisfying
E(x0) = y0. By (ii), this forces y0 = EA(x0), proving (i).
Let now ψ be a pure state on A. By Gelfand-Naimark, ψ is given by the eval-
uation at some point in the spectrum Ω of A (thus Ω is a hyperstonian compact
space and A = C(Ω)). In particular, ψ is multiplicative and takes only the val-
ues 0, 1 on the set of projections P(A), with ψ(1) = 1. This implies that any
state extension ϕ of ψ to M has A in its centralizer Mϕ.
Indeed, because if
ψ(p) = 0 for some p ∈ A, then by the Cauchy-Schwartz inequality for ϕ we have
ϕ(px) ≤ ϕ(p)1/2ϕ(x∗x)1/2 = 0, ϕ(xp) ≤ ϕ(xx∗)1/2ϕ(p)1/2 = 0, ∀x ∈ M. Since
for any projection p ∈ A we either have ψ(p) = 0 or ψ(p) = 1 and 1 ∈ Mϕ, this
shows that P(A) ⊂ Mϕ, thus all A is contained in Mϕ. Hence, ϕ is constant on
CA(x), which contains EA(x) by (1.2.3), implying that ψ(x) = ψ(EA(x)). This
proves (1.2.3) =⇒ (1.2.1) and (1.2.3) =⇒ (iii).
We have shown so far that (1.2.2) − (1.2.5) are equivalent and that they imply
(1.2.1) and (i) − (iii). To prove the remaining implication (1.2.1) =⇒ (1.2.2),
let b ∈ Mh and fix a point t ∈ Ω in the spectrum of A. Letting c0 = inf{a(t)
a ∈ Ah, a ≥ b}, c1 = sup{a(t) a ∈ Ah, a ≤ b}, we first show that condition
(1.2.1) implies c0 = c1. For if not, then the maps ψi : A + Cb → C defined by
ψi(y + αb) = y(t) + αci, i = 0, 1, y ∈ A, α ∈ C, are well defined, linear and positive;
thus kψik = 1 and by Hahn-Banach each ψi can be extended to a norm-1 linear
functional ϕi : M → C; we have thus obtained two states ϕ0, ϕ1 on M, which
extend the pure state t and are distinct (because ϕ0(b) 6= ϕ1(b)), contradicting
(1.2.1). Let now ε > 0 and for each t ∈ Ω denote ct = inf{a(t) a ∈ Ah, a ≥
b} = sup{a(t) a ∈ Ah, a ≤ b}. Let a±
t and
ct + ε/2 > a+
t (t) > ct − ε/2. By the continuity of a±
t ∈ A = C(Ω) as a
function on Ω, there exists an open-closed neighborhood Ωt of t in Ω such that
ct + ε/2 > a+
t (t′) > ct − ε/2, ∀t′ ∈ Ωt. Thus, if we denote by pt ∈ C(Ω) the
t (t′), a−
t (t), a−
t ∈ Ah be such that a+
t ≥ b ≥ a−
8
SORIN POPA
characteristic function of Ωt, then pt is a projection in A satisfying
t pt ≥ (ct − ε/2)p2.
(ct + ε/2)pt ≥ a+
t pt ≥ ptbpt ≥ a−
In particular, kptbpt − ctptk ≤ ε. Since Ω is compact, there exist t1, ..., tn ∈ Ω such
that ∪iΩti = Ω.
If we now take q1 to be the characteristic function of Ωt1 and
for each j ≥ 2, pj to be the characteristic function of Ωj \ ∪j−1
i=1 Ωi, viewed as a
projection in A, it follows that kΣjqjbqj − Σjctj qjk ≤ ε with Σjctj qj ∈ CA(b).
(cid:3)
1.3. Remark. The above proof actually shows that properties (1.2.1) − (1.2.4)
are equivalent for any given element x ∈ M. More precisely, we have proved the
following "local" statement: Let A be a MASA in a von Neumann algebra M and
let x ∈ M. The following properties are equivalent:
(1.3.1) Any two state extensions on M of a pure state on A coincide at x.
(1.3.2) CA(x) ∩ A 6= ∅.
(1.3.3) CA(x) ∩ A is a single point set {EA(x)}.
(1.3.4) For all ε > 0, there exists a finite partition of 1 with projections qk ∈ A
such that d(Σkqkxqk, A) ≤ ε.
(1.3.5) There exists a unique element E(x) ∈ A such that for all ε > 0, there exists
a finite partition of 1 with projections qk ∈ A such that kΣkqkxqk − E(x)k ≤ ε.
Moreover, if these conditions are satisfied for x, then E(x) = EA(x) and the
following additional properties hold true:
w
∩ A = {EA(x)}.
(i) CA(x)
(ii) Any conditional expectation E of M onto A (which always exist because A is
abelian) satisfies E(x) = EA(x).
(iii) Any extension of a pure state ψ on A to a state ϕ on M, satisfies ϕ(x) =
ψ(EA(x)).
1.4. Definitions. Let M be a von Neumann algebra and A ⊂ M a MASA in M.
We will use the following terminology:
(1.4.1) A ⊂ M satisfies the Kadison-Singer (abbreviated KS) property if (1.2.1) is
satisfied. Condition (1.2.4) is referred to as the paving property for A ⊂ M (the
term was coined in [A2]). Also, condition (1.2.3) is called the relative Dixmier
property for A ⊂ M, because of its relation to a phenomenon first emphasized in
[D2] (the "Dixmier averaging by unitaries"). Note that by Theorem 1.2 these three
properties for A ⊂ M are actually equivalent, and they imply 1.2(i)− (iii) as well.
KADISON-SINGER PROBLEM
9
(1.4.2) An element x ∈ M can be paved (over A) if condition (1.3.5) is satisfied.
A set X ⊂ M can be paved if each x ∈ X can be paved. If x ∈ M can be paved
and ε > 0, then we denote by n(A ⊂ M; x, ε) (or simply n(x, ε) if no confusion
is possible) the smallest number n for which there exists a partition of 1 with n
projections p1, ..., pn ∈ A, such that kΣn
i=1pixpi − EA(x)k ≤ εkx − EA(x)k, where
EA(x) is given by Remark 1.3.
More generally, if E : M → A is a conditional expectation, x ∈ M and ε > 0,
then we say that x can be ε-paved with respect to E, if there exists a finite partition
with projections p1, ..., pn ∈ A such that kΣipixpi − E(x)k ≤ εkx − E(x)k and we
denote by n(E; x, ε) the smallest number of such projections. If there exists no such
finite partition, then we let n(E; x, ε) = ∞. One should note that if E ′ : M → A is
another expectation, then kE ′(x) − E(x)k ≤ εkx − E(x)k and kΣipixpi − E ′(x)k ≤
2εkx − E ′(x)k, so we have n(E; x, ε) ≥ n(E ′; x, 2ε). Thus, taking one expectation
or another doesn't really change the nature of the function n(·; x, ε) and they are
all "comparable" to n(d; x, ε), which is by definition the smallest n for which there
exists a partition of 1 with projections p1, ...., pn ∈ A such that d(Σipixpi, A) ≤
εd(x, A). In case it is clear from the context what expectations we take, then E
will not be mentioned, and we just use the notation n(A ⊂ M; x, ε). Also, in case
there exists a normal conditional expectation of M onto A (e.g. when M = M is
a finite von Neumann algebra), then ε-pavings are always considered with respect
to this expectation.
(1.4.3) A set X ⊂ M has the uniform paving property (over A) if it can be paved
and if n(A ⊂ M; X, ε) def= sup{n(x, ε) x ∈ X} is finite, ∀ε > 0. If this holds true
for X = M, we say that A ⊂ M has the uniform paving property and use the
notation n(A ⊂ M; ε) for n(A ⊂ M; M, ε). We call this function the paving size
of A ⊂ M. We will be interested in the order of magnitude of the (decreasing)
functions n(A ⊂ M; x, ε), i.e. up to the equivalence relation f (ε) ∼ g(ε) for
functions f, g requiring the existence of positive constants 0 < c < C < ∞ such
that c ≤ f (ε)/g(ε) ≤ C, ∀ε > 0. As we will see below, the uniform paving
property appears naturally in this context, being often equivalent to the usual
paving property (notably in the case D ⊂ B), a fact first pointed out by Anderson
in [A1], [A2].
(1.4.4) Let D = ℓ∞N be the diagonal MASA in the algebra B = B(ℓ2N) of all linear
bounded operators on the Hilbert space ℓ2N. It is easy to see that the conditional
expectation B ∋ (αjk)j,k∈N 7→ (αkk)k ∈ D is the unique conditional expectation of
B onto D and that it is normal. We use the terminology "the classic Kadison-Singer
problem" for the question of whether D ⊂ B has the KS property. By Theorem
1.2, this property is equivalent to the paving property for D ⊂ B. The terminology
10
SORIN POPA
"Kadison-Singer conjecture" is sometimes used for the statement predicting that
the KS property does hold true for this inclusion, despite the fact that, in their
paper, Kadison and Singer expressed the belief that the property doesn't actually
hold true for D ⊂ B....
The next result summarizes some well known paving properties, notably J. An-
derson's observations that uniform paving for D ⊂ B is equivalent to paving and
that the classic Kadison-Singer problem is equivalent to Dk ⊂ Mk×k(C) having
uniformly bounded paving size (see [A1], [A2]).
1.5. Proposition. 0◦ Let A be a MASA in the von Neumann algebra M, p ∈
P(A) a projection and A ⊂ N ⊂ M an intermediate von Neumann algebra. If
x ∈ pMp (respectively x ∈ N ) then n(Ap ⊂ pMp; x, ε) = n(A ⊂ M; x, ε) (resp.
n(A ⊂ N ; x, ε) = n(A ⊂ M; x, ε)).
1◦ Let {Ai ⊂ Mi}i be a family of MASAs in von Neumann algebras and for each
i let xi ∈ Mi, kxik ≤ 1. Denote A = ⊕iAi, M = ⊕iMi, x = ⊕ixi ∈ M. Then
n(A ⊂ M; x, ε) = supi n(Ai ⊂ Mi; xi, ε) and n(A ⊂ M; ε) = supi n(Ai ⊂ Mi; ε),
∀ε > 0.
2◦ If a MASA A in a von Neumann algebra M has the property that there exists
a sequence of mutually orthogonal projections pn ∈ A with embeddings θn : M ֒→
pnMpn such that θn(A) = Apn, ∀n, then A ⊂ M has the paving property iff it has
the uniform paving property.
3◦ The diagonal MASA D = ℓ∞N in the algebra of all linear bounded operators
B = B(ℓ2N) on the Hilbert space ℓ2N, has the paving property iff it has the uniform
paving property. Moreover, n(D ⊂ B; ε) = supk n(Dk ⊂ Mk×k(C); ε), ∀ε > 0.
4◦ If A is a MASA in a von Neumann algebra M, with E : M → A an expec-
tation, and 1 > ε > 0, then we have sup{n(E; x, ε2) x ∈ M} ≤ (sup{n(E; y, ε)
y ∈ M})2. Thus, in order for A ⊂ M to have the uniform paving property, it is
sufficient that for some ε < 1 we have sup{n(E; y, ε) y ∈ M} < ∞.
Proof. Parts 0◦ and 1◦ are trivial and 2◦ is an immediate consequence of 1◦. Then
2◦ implies the equivalence in the first part of 3◦.
To establish the formula in 3◦, note that if a sequence of projections qn in ℓ∞N
is convergent in the weak operator topology to some element q ∈ ℓ∞N, then q is
itself a projection and qn converges to q in the strong operator topology as well.
The inequality n(D ⊂ B; ε) ≥ supk n(Dk ⊂ Mk×k(C); ε) is trivial because
the right hand side is equal to n(⊕kDk ⊂ ⊕kMk×k(C); ε) and one can embed
⊕kMk×k(C) into B in a way that takes ⊕kDk onto D.
For the inequality ≤ let T ∈ B be so that kTk ≤ 1 and T has 0 on the diagonal.
Let Pk ∈ D = ℓ∞N be the projection onto the first k coordinates. Let {pk,j}j
KADISON-SINGER PROBLEM
11
be a partition of Pk into n = supk n(Dk ⊂ Mk×k(C); ε) projections such that
kΣjpk,jT pk,jk ≤ ε. Let k1 < k2 < .... be a subsequence such that {pkm,j}m is
weakly convergent for each j = 1, 2, ..., n and denote by qj the corresponding weak
limit. By the above observation, for each j, {pkm,j}m converges in fact in the
strong operator topology and qj is a projection. Also, since Σjpkm,j = Pkm and
Pkm so-converges to 1B, it follows that Σjqj = 1 as well. Thus, {qj}j is a partition
of 1 by n projections and since {Σjpkm,jT pkm,j}m is so-convergent to ΣjqjT qj and
the operator norm is inferior semicontinuous with respect to the so-convergence, it
follow that kΣjqj T qjk ≤ lim supm kΣjpkm,jT pkm,jk ≤ ε.
Finally, part 4◦ is immediate from the definitions.
(cid:3)
1.6. Remark. While the classic Kadison-Singer problem is still open, one should
point out that a large number of beautiful paving results have been obtained over
the years, showing that the equivalent conditions 1.3 are satisfied for many classes
of operators x ∈ B(ℓ2Z). Thus, it is shown in [A2] that if x is in the C∗-algebra for
the reduced C∗-algebra of the group Z, C∗
in the operator norm-closure
of the span of the range of the left regular representation λ of the group Z, then
x can be paved. In [BeHKW] it is shown that matrices with non-negative entries
in Mn×n(C) can be paved, while in [BT] it is shown that if an element x in the
weak closure L(Z) ≃ L∞(T) of C∗
r (Z) in B(ℓ2Z) has Fourier coefficients satisfying
certain growth properties, then x can be paved. Also, a number of results have been
obtained in [AkA], [A1], [A2], [CaFTW], [We], etc, showing that in order to solve
the paving conjecture, it is sufficient to be able to pave certain particular classes of
elements (e.g. projections with small diagonal entries in [AkA]).
r (Z), i.e.
2. Kadison-Singer in II1 factor framework
We prove in this section that the KS property for the inclusion of the diagonal
MASA D = ℓ∞N into the algebra B = B(ℓ2N), of all linear bounded operators on
the Hilbert space ℓ2N, is equivalent to the KS property of MASAs in II1 factors
obtained as ultraproducts of certain Cartan inclusions. Also, we use a dilation trick
to prove that in order to pave arbitrary elements in an ultraproduct of inclusions
of MASAs, it is sufficient to pave projections that expect on scalars.
From now on, we fix once for all an (arbitrary) free ultrafilter ω on N. All
finite von Neumann algebras that we consider are assumed equipped with a faithful
normal trace state, generically denoted by τ (unless otherwise specified).
If Mn, n ≥ 1, is a sequence of finite von Neumann algebras then, we denote
by ΠωMn their ω-ultraproduct, i.e., the finite von Neumann algebra obtained as
the quotient of ⊕nMn by its ideal Iω = {(xn) limω τ (x∗
nxn) = 0}, endowed with
the trace τ (y) = limω τ (yn), where (yn)n ∈ ⊕nMn is in the class y ∈ ⊕nMn/Iω
12
SORIN POPA
([W]). Recall that if Mn are factors and dim Mn → ∞, then ΠωMn is a II1 factor
([W], [F]) and that if An ⊂ Mn are MASAs, n ≥ 1, then ΠωAn is a MASA in
[P1]). If A ⊂ M is a MASA in a finite von Neumann algebra,
ΠωMn (see e.g.
then Aω ⊂ M ω denotes its ω-ultrapower, i.e. the ultraproduct of infinitely many
copies of A ⊂ M . Note that M naturally embeds into M ω, as the von Neumann
subalgebra of constant sequences.
Recall that a Cartan subalgebra A in a finite von Neumann algebra M is a
MASA in M whose normalizer NM (A) = {u ∈ U(M ) uAu∗ = A} generates M ,
i.e. N (A)′′ = M (see [FM]).
2.1. Notations (a) We denote the inclusion ΠωDn ⊂ ΠωMn×n(C) by D(ω) ⊂
M(ω), or simply D ⊂ M. Note that given any sequence of MASAs An ⊂ Mn×n(C),
the von Neumann algebra ΠnAn ∈ M is unitary conjugate to D in M. One should
point out that D is not a Cartan subalgebra in M, in fact M has no Cartan
subalgebras (cf. [P2]; see 2.3.2◦ below). Also, D, M are non-separable (cf. [F]) and
M is non-amenable (because it contains L(F2), as first noticed in [Wa]).
(b) We represent the hyperfinite II1 factor R as the infinite tensor product
⊗n(M2×2(C), tr)n, where tr is the normalized trace on M2×2(C). Also, we de-
note by D ⊂ R the Cartan subalgebra obtained as the infinite tensor product of
the diagonals D2 ⊂ M2×2(C). Recall that any other Cartan subalgebra A ⊂ R is
conjugate to D by an automorphism of R (cf [CFW]). Thus, if Dω ⊂ Rω is the
ω-ultrapower of D ⊂ R, then any ultraproduct ΠωAn ⊂ Rω, with An ⊂ R Car-
tan subalgebras, is conjugate to Dω by an automorphism θ = (θn)n of Rω, where
θn ∈ Aut(R) is so that θn(An) = D. We denote by R ⊂ Rω the von Neumann
algebra Dω ∨ R, generated by Dω and R, or equivalently by Dω and NR(D).
If Γ ⊂ NR(D) is any countable subgroup generating the hyperfinite equivalence
relation R associated with D ⊂ R (cf [FM]), then R is generated by Dω and Γ.
Moreover, if Γ acts freely on D, then Γ acts freely on Dω as well and so we can
view R as the crossed product Dω ⋊ Γ. Finally, note that R is an amenable II1
von Neumann algebra, but not a factor, in fact any sequence (an)n ∈ Dω with
an ∈ 1 ⊗j≥kn (D2)j for some kn → ∞, lies in R′ ∩ Dω = R′ ∩ Dω = Z(R) (the
center of R).
Note that, while Dω is Cartan in R, Dω is not Cartan in Rω, in fact Rω has no
Cartan subalgebras (by [P2]; see 2.3.2◦ below).
2.2. Theorem. D ⊂ B has the KS property (equivalently, the paving property) if
and only if D ⊂ M (resp. Dω ⊂ Rω, resp. Dω ⊂ R) has this property. Moreover,
all these inclusions have the same paving size (whether finite or infinite):
(2.2.1)
n(D ⊂ B; ε) = n(D ⊂ M; ε) = n(Dω ⊂ Rω; ε) = n(Dω ⊂ R; ε).
KADISON-SINGER PROBLEM
13
They also have the same paving size as the Cartan subalgebra inclusions D ⊂
NM(D)′′ and Dω ⊂ NRω (Dω)′′.
Proof. Consider the inclusion A0 = ⊕∞
by 1.5.1◦ and 1.5.3◦ we have
n=1Dn ⊂ ⊕∞
n=1Mn×n(C) = M0 and note that
n(D ⊂ B; ε) = sup
n
n(Dn ⊂ Mn×n(C); ε) = n(A0 ⊂ M0; ε).
Embed now A0 into D and then extend this to an embedding of M0 into R
so that the matrix units of each direct summand Mn×n(C) are in the normalizing
groupoid of D ⊂ R (this is possible because D is Cartan in R; in fact, semiregularity
is sufficient). Note that this implies M0 and D make a commuting square, i.e.
EM0ED = EDEM0 = EA0. Also, we trivially have
n(Dω ⊂ Rω; ε) ≥ n(Dω ⊂ R; ε) ≥ n(Dω ⊂ R; R, ε) ≥ n(Dω ⊂ R; M0, ε).
Let x = (xn)n ∈ M0 ⊂ R be so that EDω (x) = EA0 (x) = 0 and note that
n(Dω ⊂ R; x, ε) = supn n(Dω ⊂ R; xn, ε). Let sn ∈ D ⊂ Dω denote the support
projection of Mn×n(C) in A0 ⊂ D. Each Dωsn ⊂ (M0 ∨ Dω)sn is of the form
C(Ω) ⊂ C(Ω) ⊗ Mn×n(C). Thus, if p1, ..., pm ∈ P(Dω) is a partition of 1 such that
kΣipixpik ≤ ε, then the evaluation at any point t ∈ Ω of pisn, 1 ≤ i ≤ m, gives
a partition of 1Mn×n(C) with m projections qi in Dn such that kΣm
i=1qixnqik ≤ ε.
This shows that n(Dω ⊂ R; M0, ε) ≥ n(A0 ⊂ M0; ε). Since the latter is equal to
n(D ⊂ B; ε) and to supn n(Dn ⊂ Mn×n(C); ε), in order to end the proof of the fact
that D ⊂ B, Dω ⊂ Rω, Dω ⊂ R (as well as A0 ⊂ M0) have the same paving size,
it is sufficient to show that supn n(Dn ⊂ Mn×n(C); ε) ≥ n(Dω ⊂ Rω; ε).
To this end, let x = (xn)n ∈ Rω ⊖ Dω and note that one can take each xn to
belong to R ⊖ D and such that kxnk ≤ kxk + cn, ∀n, for some cn → 0. Moreover,
since there exists an increasing sequence of 2k × 2k matrix subalgebras M2k ⊂ R
with diagonal subalgebra D2k ⊂ D making a commuting square with D ⊂ R such
that ∪kM2k
= D, we may replace each xn by EM2kn (xn), and
thus assume xn ∈ M2kn ⊖ D2kn , ∀n. Let {qn
j }j ⊂ D2kn be a partition of 1 with
K = supn n(Dn ⊂ Mn×n(C); ε) projections such that kΣK
j k ≤ εkxnk. If
we denote by qj = (qn
j=1qjxqjk ≤ εkxk. This proves
the desired inequality.
Let now ε > 0 and assume m = n(D ⊂ B; ε) = supn n(Dn ⊂ Mn×n(C); ε) is
finite. Any x ∈ M with kxk ≤ 1 and ED(x) = 0 can be represented by a sequence
x = (xn)n with xn ∈ Mn×n(C) such that kxnk ≤ 1 + cn, EDn(xn) = 0, for some
j )n ∈ Dω, it follows that kΣK
= R and ∪kD2k
j=1qn
j xnqn
w
w
14
SORIN POPA
j k ≤ ε(1 + cn). But then pj = (pn
cn → 0. For each n there exists a partition of 1 with projections pn
j , 1 ≤ j ≤ m,
j xnpn
j=1pn
such that kΣm
j )n ∈ D gives a partition
of 1 satisfying kΣm
j=1pjxpjk ≤ ε. Thus, supn n(Dn ⊂ Mn×n(C); ε) ≥ n(D ⊂ M; ε).
Conversely, assume m = n(D ⊂ M; ε) is finite. Let x be an element in Mk×k(C),
for some k ≥ 1, with kxk ≤ 1, EDk (x) = 0. For each n larger than k, embed
Mk×k(C) into Mn×n(C) by first letting n = kdn + rn, with dn, rn ∈ N, rn < k,
then letting sn
j ∈ Dn be mutually orthogonal projections of trace k/n, and then
j ⊂ sn
identifying Dk ⊂ Mk×k(C) with Dnsn
j via some isomorphism θn
j ,
for each j = 1, ..., dn, and mapping diagonally
j Mn×n(C)sn
Mk×k(C) ∋ y 7→ θn(y) def= Σjθn
j (y) ∈ Σdn
j=1sn
j Mn×n(C)sn
j ⊂ Mn×n(C).
i θn(x)pn
i sn
j θj(x)pn
j sn
j θj(x)pn
j sn
i )n, with pn
i = Σdn
j k < ε + δ.
Then consider the embedding θ : Mk×k(C) → M, by θ(y) = (θn(y))n. Let
p1, ..., pm ∈ P(D) be a partition of 1 such that kΣipiθ(x)pik ≤ ε. One can then
choose representing sequences pi = (pn
i ∈ P(Dn), 1 ≤ i ≤ m, a partition
of 1 for each n. We claim that for any δ > 0 there exist n and j ∈ {1, ..., dn}, such
that kΣipn
Indeed, for if not then for every n and j = 1, ..., dn the spectral projection of
Σipn
i sn
j corresponding to the interval [ε + δ, 1] is non-zero, thus having
i sn
trace at least 1/n. Since Σipn
j , the spectral
projection corresponding to [ε + δ, 1] of Σipn
i has trace ≥ dn/n. But this
implies that the spectral projection corresponding to [ε + δ/2, 1] of Σipiθ(x)pi =
(Σipn
i )n ∈ M has trace ≥ limn dn/n = 1/k. Thus kΣipiθ(x)pik ≥ ε + δ/2,
a contradiction.
j (x)pn
j θn
i sn
If we now choose some n and j ∈ {1, ..., dn} satisfying kΣipn
j k < ε+δ,
and let qi, 1 ≤ i ≤ m, be the pre-image in Dk of the partition {pn
j }i via θn, then
i sn
kΣm
i=1qixqik < ε+δ. Letting δ = 1/n, n = 1, 2, ..., we obtain a sequence of partitions
of 1 by projections {q1(n), ..., qm(n)}n in Dk satisfying kΣm
i=1qi(n)xqi(n)k < ε+1/n.
But the unit ball of Dk is compact in the operator norm, so by taking the limit
over some subsequence, we get a partition of 1 with projections q1, ..., qm ∈ Dk
with kΣiqixqik ≤ ε. Thus, m ≥ n(Dk ⊂ Mk; ε), and since k was arbitrary, m ≥
supk n(Dk ⊂ Mk; ε).
j=1Σipn
i θn(x)pn
j θn
j (x)pn
i sn
i θn(x)pn
j sn
Finally, since Dω ⊂ R ⊂ N (Dω)′′ ⊂ Rω, we have
n(Dω ⊂ R; ε) ≤ n(Dω ⊂ NM(D)′′; ε) ≤ n(Dω ⊂ Rω; ε),
and since the first and last terms are equal, they must all be equal. Similarly,
D ⊂ N (D)′′ ⊂ M implies n(D ⊂ N (D)′′; ε) ≤ n(D ⊂ M; ε), while arguments
KADISON-SINGER PROBLEM
15
above show that supn n(Dn ⊂ Mn×n(C); ε) ≤ n(D ⊂ N (D)′′; ε), with the first of
these terms equal to n(D ⊂ M; ε).
(cid:3)
If M is a finite von Neumann algebra with its faithful normal trace state τ ,
then we denote by kxk2 = τ (x∗x)1/2, x ∈ M , the L2 (or Hilbert) norm given
by the trace. We denote by L2M the Hilbert space obtained by completing M
in this L2 norm and view M in its standard representation, as left multiplication
representation on L2M . We also use the notation L1M for the completion of M
in the norm kxk1 = τ (x). We view the elements in L2M (resp. L1M ) as square
summable (resp. summable) operators affiliated with M ⊂ B(L2M ), in the usual
way. All self-adjoint elements affiliated with M (in particular elements in L2M ,
L1M ) have spectral decomposition belonging to M and they can be multiplied. In
particular, we have L2M · L2M = L1M .
A finite von Neumann algebra M with a normal faithful trace is separable if it is
separable with respect to the k·k2-norm given by the trace. This condition is easily
seen to be equivalent to M being countably generated. A von Neumann algebra
is diffuse if it has no minimal (non zero) projection. Any abelian von Neumann
algebra A which is diffuse and separable is isomorphic to L∞([0, 1]) (or to L∞(T)).
Moreover, if A is endowed with a faithful normal state τ , then the isomorphism
A ≃ L∞([0, 1]) can be taken so that to carry τ onto the integral R · dµ, where µ is
the Lebesgue measure on [0, 1].
It is well known that all separable diffuse abelian von Neumann subalgebras in
an utraproduct II1 factor are unitary conjugate (see e.g. [P2]). We will show below
that any II1 factor M that has this property will automatically have several other
properties, like absence of Cartan subalgebras (already noticed in [P2]) and the
fact that in order to pave arbitrary elements over a MASA in M , it is sufficient
to pave projections that expect on scalar multiples of 1. (Note that in Anderson's
formulation of the KS problem as the uniform paving property in Mk×k(C), k ր ∞,
the reduction of the problem to paving special elements, such as projections with
constant diagonal, has been subject of much study, see [AkA], [A2], [CaFTW], etc.)
2.3. Proposition. 1◦. Assume a II1 factor M has the property that given any
projection p ∈ M , any two separable diffuse abelian von Neumann subalgebras of
pM p are unitary conjugate. Then M satisfies the following properties
(a) Given any MASA A in M , there exists a diffuse abelian von Neumann sub-
algebra B0 ⊂ M perpendicular to A.
(Recall from [P2] that two von Neumann
subalgebras B1, B2 of a finite von Neumann algebra M are said to be perpendicular
if τ (b1b2) = 0, ∀bi ∈ Bi with τ (bi) = 0, i = 1, 2.)
(b) M has no separable MASAs and no Cartan subalgebras.
16
SORIN POPA
(c) If A is a MASA in M , then A ⊂ M has the paving property iff any projection
in M that expects on a scalar multiple of a projection in A can be paved. Moreover,
if P0 denotes the set of such projections, then the paving size n(ε) of A ⊂ M satisfies
n(ε) ≤ n(A ⊂ M ; P0, ε/50)2(ε−1 + 1)2.
2◦ If {Mn}n is a sequence of finite von Neumann factors with dimMn → ∞,
then the ultraproduct II1 factor M = ΠωMn satisfies the assumption in part 1◦,
i.e., given any projection p ∈ M , any two separable diffuse abelian von Neumann
subalgebras of pM p are unitary conjugate. Thus, M = ΠωMn satisfies properties
(a), (b), (c) as well.
Proof. 2◦ is well known (see e.g. Lemma 7.1 in [P2]).
1◦ Part (a) is shown in the proof of Theorem 7.3 in [P2], but let us recall the
argument here for completeness. Let D ⊂ A be a separable diffuse von Neumann
subalgebra. Since any two separable diffue abelian subalgebras in M are unitary
conjugate and since M contains copies of the hyperfinite II1 factor, we may assume
D is the Cartan subalgebra of such a subfactor R ⊂ M , represented as D = D⊗∞
2 ⊂
M2×2(C)⊗∞ = R. Let D0
2 ⊂ M2×2(C) be a maximal abelian subalgebra of M2×2(C)
⊗∞ ⊂ R. Then D ⊥ D0 and since
that is perpendicular to D2 and denote D0 = D0
2
both D, D0 are MASAs in R, we have ED′∩M (D0) = ED′∩R(D0) = ED(D0) = C,
i.e. D0 ⊥ D′ ∩ M ⊃ A, proving (a).
To prove (b), let A be a MASA in M . If A is separable, then it has a diffuse proper
subalgebra, A0 ⊂ A, which cannot be unitary conjugate to A because it is not a
MASA. Moreover, by part (a), there exist separable diffuse abelian subalgebras
D, D0 in M such that D ⊂ A and D0 ⊥ A. Let u ∈ U(M ) be so that uDu∗ = D0.
Indeed, because for any
Then u is perpendicular to the normalizer of A in M .
v ∈ NM (A) and any partition pi ∈ D of mesh ≤ ε, we have
τ (uv)2 = τ (Σipiuvpi)2 ≤ kΣipiuvpik2
2 = Σiτ (u∗piuvpiv∗) = Σiτ (pi)2 ≤ ε.
Since ε > 0 was arbitrary, τ (uv) = 0. Thus u ⊥ NM (A)′′.
To prove (c), note first that for any MASA in a II1 factor M and any x ∈ M ,
the paving size over A of any element x ∈ M behaves well with respect to scalar
translations and multiplications:
(1) n(x, ε) = n(x + α1, ε), n(αx, ε) = n(x, ε/α).
Now note that if y ∈ M is a δ-perturbation of x ∈ M , then any ε-paving of y
gives a ε + δ paving of x, more precisely:
(2) If kx−yk ≤ 2−1δ(1+ε)−1kx−EA(x)k and kΣipi(y−EA(y))pik ≤ εky−EA(y)k,
then
KADISON-SINGER PROBLEM
17
kΣipi(x − EA(x))pik
≤ k(x − y) − EA(x − y)k + kΣipi(y − EA(y))pik
≤ k(x − y) − EA(x − y)k + εky − EA(y)k
≤ (1 + ε)k(x − y) − EA(x − y)k + εkx − EA(x)k
≤ (ε + δ)kx − EA(x)k
(3) If x1, x2 ∈ M can be paved, then x1 + x2 can be paved. More specifically,
if x = y1 + iy2, yi = y∗
i , is the decomposition of x into its real and imaginary
parts, then kyik ≤ kxk and so if pi, qj ∈ P(A) are partitions of 1 such that
kΣjpi
j}i,j
satisfies
j−EA(yi)k ≤ εkyi−EA(yi)k, i = 1, 2, then the partition {pk}k = {p1
jyipi
i p2
kΣkpkxpk − EA(x)k ≤ kΣkpky1pk − EA(y1)k + kΣkpky2pk − EA(y2)k
≤ kΣjp1
j y1p1
j − EA(y1)k + kΣjp2
j y2p2
j − EA(y2)k
≤ εky1 − EA(y1)k + εky1 − EA(y1)k ≤ 2εkx − EA(x)k.
Thus, n(x, 2ε) ≤ n(ℜx, ε)n(ℑx, ε).
Let now x = x∗ ∈ M with EA(x) = 0 and kxk = 1. Note that y0 = 12−1(x +
5kxk) satisfies 1/3 ≤ y0 ≤ 1/2, thus 1/3 ≤ EA(y0) ≤ 1/2. Denote by ek the spectral
projection of EA(y0) corresponding to the interval [1/3 + (k − 1)ε/6, 1/3 + kε/6)
and note that there are ≤ ε−1 + 1 many such non-zero projections.
If we let
tk = 1/3 + (k − 1)ε/6, ak = EA(y0)ek, then tkek ≥ ak ≥ tk−1ek. Thus, if we
denote b = Σkt−1
k−1ak then 1 ≤ b ≤ 1 + ε/2 and (1 + ε/2)−1/2 ≤ b−1/2 ≤ 1. Notice
now that y = b−1/2y0b−1/2 satisfies EA(y) = Σktkek, 1/3 − ε/12 ≤ y ≤ 1/2 and
ky0 − yk ≤ 2k1 − b−1/2kky0k ≤ k1 − b−1/2k ≤ ε/4.
k ∈ Aek, of equal trace
τ (ej
k −
ykjk ≤ ε/4, with (1/3 − ε/12)ej
k ∈
A(1 − ej
k)/tk. To see that this is possible, we
need to have τ (ej
k), which is easily seen to hold true due to our
choices. Note also that τ (pj
k). Take B0 to be a separable diffuse abelian
von Neumann subalgebra of pj
k which is perpendicular to Apj
k be a
projection in B0 such that τ (f j
k). Let v ∈ M be a partial isometry such
We now split each ek into the sum of four projections ej
k) = τ (ek)/4, ek = Σ4
k)/tk ≤ τ (1− ej
k) ≥ τ (ej
kM pj
k) = τ (ej
k) be a projection of trace τ (ej
k. We still have kej
k ≤ ykj ≤ 1/2ej
j=1ej
k, and denote ykj = ej
kyej
k and EA(ykj) = tkej
ky0ej
k. Let pj
k. Let f j
k − yj
k − yj
18
SORIN POPA
k and vv∗ = f j
that v∗v = ej
diffuse abelian von Neumann subalgebras of f j
chosen so that in addition we have v(ej
k, which due to the assumption that any two separable
k are unitary conjugate, can be
kM f j
k − ykj)v∗ ∈ B0.
Denote gkj = ykj + v(ej
is easy to check that g2
k − ykj)v∗ + v(ykj(ej
k − ykj))1/2 + (ykj(ej
kj = gkj, i.e., gkj ∈ P(M ). Moreover
k − ykj))1/2v∗. It
k − ykj)v∗) + EA(ykj)
EA(gkj) = EA(pj
= τ (v(ej
k + ej
kgkjpj
kgkjej
k − ykj)v∗)/τ (pj
k) = EA(v(ej
k)pj
k + tkej
k + ej
k).
k + ej
Thus, by our assumption, each one of the projections gkj ∈ (pj
k + ej
k)
can be paved over A(pj
k can
be paved, implying that y can be paved. By (2), it follows that y0 can be paved,
and by (1), x can be paved as well. Thus, any selfadjoint element can be paved, so
by (3) any element in M can be paved.
k can be paved, so Σk,jej
k). Thus ykj = ej
k)M (pj
kyej
k = tk(pj
kgkjej
k + ej
(cid:3)
Moreover, if we keep track of the number of projections necessary in the above
pavings we see that in order to ε/2-pave a selfadjoint element, n(P0, ε/50)(ε−1 + 1)
many projections are sufficient. By (3), we get that n(ε) ≤ n(P0, ε/50)2(ε−1 + 1)2.
2.4. Remarks. 1◦ Let A ⊂ M be a MASA in a II1 factor and x ∈ M ⊖A, kxk ≤ 1.
If we view x as an element in M ω, then its ε-paving number over Aω is ≤ n iff
for any δ > 0, there exists a partition of 1 with projections p1, ..., pn ∈ P(A) with
the property that the spectral projection of Σipixpi corresponding to the interval
(ε, ∞) has trace ≤ δ.
2◦ We already mentioned in 2.1 (b) several properties of the algebra R: given any
representation of R as crossed product D ⋊ Γ, for some free action of a (necessarily
countable amenable) group Γ on D, Γ acts freely on Dω as well and we have
R = Dω ⋊ Γ; thus, R is amenable and has Dω as a Cartan subalgebra, but it has
large, non-separable center. In addition, note that due to Rohlin's theorem and
Følner's condition for amenable groups, any two free actions of the same amenable
group Γ y Dω are conjugate by a unitary element in NRω (Dω). Moreover, the 1-
cohomology for Dω ⊂ R vanishes, so if Γ, Λ ⊂ NR(D) are two countable amenable
groups of unitaries that implement free actions on D and ∆ : Γ ≃ Λ is a group
isomorphism, then there exists u ∈ NRω (Dω) such that uugu∗ = ∆(ug), ∀ug ∈ Γ.
In particular, if u1, u2 ∈ NR(D) act freely on D, then there exists u ∈ NRω (Dω)
such that uu1u∗ = u2. Note that in fact all these properties hold also for countable
amenable subgroups Γ ⊂ NM(D) acting freely on D.
3◦ Recall that in (4.1.(iii) and 4.3.3◦ of [P7]) it was shown that if B is a separable
amenable von Neumann subalgebra in a II1 von Neumann algebra M such that the
KADISON-SINGER PROBLEM
19
Pimsner-Popa index [PiP] of the inclusions pBp ⊂ pM p is infinite for any projection
p ∈ B, p 6= 0, then there exist non-normal conditional expectations of M onto B,
and thus EB is not the unique conditional expectation of M onto B. In particular,
if A is a separable MASA in a II1 von Neumann algebra M , then EA is not the
unique conditional expectation of M onto A, and thus A ⊂ M cannot have the KS
property, nor the paving property. We recall the argument in [P7], emphasizing a
simplification that occurs in the case of a MASA.
First one constructs a singular state ϕ on M with ϕB = τ as follows: Like
in [P7], the hypothesis implies there exists bn ∈ L1M+ satisfying EB(bn) = 1
and τ (s(bn)) ≤ 2−n, where for a positive element b ≥ 0 in M , s(b) denotes the
(In the case B = A is a MASA in M , the argument
support projection of b.
becomes much simpler, as one can take bn = 2nqn ∈ M , with qn the following
projection: let ekk ∈ A be a partition of 1 with 2n mutually equivalent projections
and complement it to a set ejk ∈ M of matrix units, then define qn = 2−nΣj,kejk,
which is a projection in M with EA(qn) = ΣkekkEA(qn)ekk = EA(Σkekkqnekk) =
2−n.) Then ϕn = τ (· bn) defines a normal state on M which, since EB(bn) = 1,
satisfies τ (ybn) = τ (y), ∀y ∈ B. Take ϕ to be a state on M obtained as a weak-limit
of ϕn. Then we still have ϕB = τB while ϕ is singular on M (because for each
fixed n we have ϕ(1 − ∨m≥ns(bn)) = 0 and limn(1 − ∨m≥ns(bn)) = 1).
Next, since B is amenable, by Connes' Theorem we can find a countable amena-
ble subgroup U0 ⊂ U(B) such that U ′′
0 = B. For each x ∈ M , put ψ(x) =
R ϕ(uxu∗)du, where the integral is in the Banach limit sense, over an invariant
mean on the countable amenable group U0. Then ψ defines a state on M which
is in the σ(M ∗, M )-closure of a countable set of singular states on M . By [Ak], it
follows that ψ is singular as well. Also, by its definition, ψ has the span of U0 in
its centralizer and ψB = τB. If now a ∈ B is arbitrary and an ∈ spU0, kank ≤ kxk
are so that kx− bnk2 → 0, then by Cauchy-Schwartz inequality for ψ, for all x ∈ M
we have
ψ(ax) − ψ(anx) ≤ ψ((a − an)(a − an))1/2ψ(x∗x)1/2
= τ ((a − an)(a − an)∗)1/2ψ(x∗x)1/2 = ka − ank2ψ(x∗x)1/2 → 0.
Similarly, ψ(xa) − ψ(xan) → 0. Since ψ(anx) = ψ(xan), ∀n, this shows that
ψ(ax) = ψ(xa), i.e. B is in the centralizer of ψ. Taking E : M → B to be the
unique conditional expectation satisfying ψ(E(x)a) = ψ(xa), ∀x ∈ M , a ∈ B, we
have constructed this way a singular (thus non-normal) conditional expectation of
M onto B.
3. Paving in the L2-norm
Given a MASA A in a finite von Neumann algebra M and x an element in M
20
SORIN POPA
with EA(x) = 0, our strategy for estimating the norm of its paving y = Σipixpi,
for pi ∈ P(A) a partition of 1, will be to calculate the moments τ ((y∗y)n) and use
the well known formula kyk2 = limn τ ((y∗y)n)1/n. More generally, in order to have
kyk ≤ c, we need to prove that τ ((y∗y)n) ≤ c2n, for large enough n. One way of
controlling these moments is to construct the partitions {pi}i ⊂ A so that to have
"high order of independence" with respect to x.
We will use the following terminology in this respect: Two sets V, W ⊂ M⊖C are
n-independent if any alternating word x1y1....xkyk, with k ≤ n and x1 ∈ V ∪ {1},
x2, .., xk ∈ V , y1, ..., yk−1 ∈ W , yk ∈ W ∪ {1}, has trace 0 (unless k = 1 and
x1 = y1 = 1). An algebra B0 ⊂ M is n-independent to V if V and B0 ⊖ C
are n-independent. Note that 1-independence amounts to what one usually calls
τ -independence.
More generally, if P ⊂ M is a von Neumann subalgebra, then two sets V ⊂
i=1xiyi) = 0, for all
M ⊖ P , W ⊂ M ⊖ C1 are n-independent relative to P if EP (Πk
1 ≤ k ≤ n, all x1 ∈ V ∪ {1}, xi ∈ V , yk ∈ W ∪ {1}, yi ∈ W .
In this section we prove a general fact about independence, showing that given
any MASA A in a finite von Neumann algebra M , we can find partitions of 1 in A
that are "asymptotically 2-independent" with respect to any given countable set of
elements in M⊖A. In other words, given any countable set X ⊂ M ⊖A, there exists
a diffuse abelian subalgebra B0 ⊂ Aω such that any word with at most 4 alternating
letters in X and B0 ⊖ C has trace 0. In particular, kpixpjk2 = kpik2kpjk2kxk2,
for any pi, pj ∈ P(A), x ∈ X, so any partition pi ∈ B0 with small mesh gives
L2-pavings of x ∈ X, simultaneously for all x ∈ X: kΣipixpik2
2Σiτ (pi)2 ≤
maxi{τ (pi)}ikxk2
2. We construct such partitions recursively, but another method,
where moments are controlled through incremental patching, can be used instead
(see Section 5.3).
2 = kxk2
3.1. Lemma. 1◦ If ξ ∈ L2M , u ∈ U(M ) are so that u2 = 1, τ (ξ∗uξu∗) ≤ ckξk2
2,
for some 1 ≥ c > 0, then p1 = (1 + u)/2, p2 = (1 − u)/2 is a partition of 1 with
projections satisfying kp1ξp1 +p2ξp2k2 ≤ (1+c)/√2kξk2. If in addition τ (uξ∗ξ) ≤
2and τ (u) ≤ c/2, then kpiξpik2 ≤ (1 + 2c)/√2kξk2kpik2, i = 1, 2.
ckξk2
2◦ If ξ ∈ L2M and u ∈ U(M ) satisfy τ (ξ∗uξu∗) ≤ ckξk2
2, for some c ≤
2−7, and n ≥ 27, then the spectral projections {ek}1≤k≤n of u defined by ek =
e[e2πi(k−1)/n,e2πik/n)(u), give a partition of 1 and satisfy kΣkekξekk2 ≤ 3/4kξk2.
2 + 2ℜτ (ξ∗uξu∗) ≤ (2 + 2c)kξk2
Proof. 1◦ We have kξ + uξu∗k2
2. Since
2 ≤ (1 + c)/2kξk2
p1ξp1 + p2ξp2 = 2−1(ξ + uξu∗), we get kp1ξp1 + p2ξp2k2
2 and the
first part of the statement follows. If τ (uξ∗ξ) ≤ ckξk2
2 and τ (u) ≤ c/2 as well,
then τ (p1) − 1/2 = τ (u)/2 ≤ c/4. Thus τ (p1) ≥ 1/2 − c/4 ≥ 1/4 and also
2 = 2kξk2
KADISON-SINGER PROBLEM
21
1/2 ≤ τ (p1) + c/4 ≤ τ (p1)(1 + c)/4, so we get:
kp1ξp1k2
≤ (1/4 + 3c/4)kξk2
2 = τ ((1 + u)ξ∗(1 + u)ξ)/4 = τ (ξ∗ξ)/4 + τ (uξ∗ξ)/2 + τ (uξ∗uξ)/4
2kp1k2
2.
2(1 + c)τ (p1) ≤ (1 + 3c)2/2kξk2
2 ≤ (1 + 3c)/2kξk2
Similarly, by using that p2 = (1 − u)/2, we obtain
kp2ξp2k2
2 ≤ (1 + 3c)/2kξk2
2(1 + c)τ (p2) ≤ (1 + 2c)2/2kξk2
2kp2k2
2.
2◦ We may clearly assume kξk2 = 1. Note that if we denote λk = e2πik/n,
then ku − Σkλkekk ≤ 2πi/n − 1 < 2π/n. Since the elements {ejξek}1≤j,k≤n are
mutually orthogonal in the Hilbert space L2M , by using first Pythagora's theorem
and then the inequality λ∗
4 − 4kΣkekξekk2
≥ kΣj6=k(λ∗
j λk − 1 ≤ 2, ∀j, k, we get:
2 − 4kΣkekξekk2
2 = 4kξk2
2 = 4kΣj6=kej ξekk2
j ej)ξ(Σkλkek) − ξk2
2
2
j λk − 1)ejξekk2
≥ ku∗ξu − ξk2
2 = k(Σjλ∗
2 − 4ku − Σkλkekk
= 2 − 2ℜτ (ξ∗u∗ξu) − 4ku − Σkλkekk ≥ 2 − 2c − 8π/n.
If we now choose c < 2−7 and n ≥ 27, then from the first and last term of the
above estimates we get
kΣkekξekk2
2 ≤ 1/2 + c/2 + 2π/n ≤ 9/16.
(cid:3)
3.2. Remark. For the following lemmas, it will be useful to recall that a unitary
representation π of a group G on a Hilbert space H is weak mixing if any of the
following equivalent conditions is satisfied:
(3.2.1) Given any finite subset F ⊂ H and any ε > 0, there exists g ∈ G such that
hπ(g)(ξ), ηi ≤ ε, ∀ξ, η ∈ F ;
(3.2.2) π has no non-zero finite dimensional invariant subspaces H0 ⊂ H;
(3.2.3) The representation π ⊗ π of G on H ⊗ H is ergodic, i.e.
non-zero vectors.
(3.2.4) For any finite dimensional subspace H0 ⊂ H and any ε > 0, there exists
g ∈ G such that, if p0 denotes the orthogonal projection of H onto H0 and we
still denote by π the representation of G on the space of Hilbert-Schmidt operators
HS(H) ≃ H ⊗ H, then T r(π(g)(p0)p0) ≤ εT r(p0).
it has no fixed
22
SORIN POPA
3.3. Lemma. If B ⊂ M is a diffuse von Neumann subalgebra, then the action Ad
of its unitary group U(B) on L2(M ⊖ (B′ ∩ M )) is weak mixing. Moreover, if B is
abelian, then the restriction of this Ad-action to the subgroup of period 2 elements,
U 0(B) def= {u ∈ U(B) u2 = 1}, is still weak mixing.
Proof. If the action is not weak mixing, then there exists a non-zero finite dimen-
sional subspace H0 ⊂ L2(M ⊖ (B′ ∩ M )) satisfying uH0u∗ = H0, ∀u ∈ U(B). In
particular, if A ⊂ B is a diffuse abelian subalgebra of B and U 0 = U 0(A) denotes the
group of unitaries of period two in A as in part 2◦, then H0 is invariant to the repre-
sentation ξ 7→ Ad(u)(ξ) = uξu∗ of U 0 on H0. Since the image of this representation
is an abelian subgroup V 0 of U(H0), it can be diagonalized. Thus, H0 = Σj Cξj,
with ξ1, ξ2, ...., ξn an orthonormal basis of H0 such that Ad(U 0)(ξj) ⊂ Tξj, ∀j, and
since any element in V 0 has period 2, we actually have Ad(U 0)(ξj) ⊂ {±ξj}, ∀j.
But the group (U 0, k k2) is Polish and contractible. This can be seen by taking a
k · k2-continuous path {pt 0 ≤ t ≤ 1} ⊂ P(A) with τ (pt) = t, pt ≤ pt′ iff t ≤ t′,
then defining the continuous path of group morphisms Φt : U 0(A) → U 0(A), by
Φt(u) = pt + u(1− pt), which satisfies Φ0(u) = u, Φ1(u) = 1, ∀u ∈ U 0(A). Since U 0
is contractible and the representation Ad is continuous, and since the one dimen-
sional representation u 7→ Ad(u) lies in {±1}, this representation must be trivial,
i.e. uξju∗ = ξj, ∀u ∈ U 0, ∀j. Hence, uξ = ξu for all ξ ∈ H0 and for all u ∈ U(A)
(because U 0 generates A as a von Neumann algebra). Since any u ∈ U(B) lies in
a diffuse abelian von Neumann subalgebra A ⊂ B, it follows that uξ = ξu, for all
u ∈ U(B) and all ξ ∈ H0. This means that H0 ⊂ L2(B′ ∩ M ), while at the same
time H0 ⊥ B′ ∩ M , implying that H0 = 0.
(cid:3)
We have seen in the previous lemma that if A is a diffuse abelian von Neumann
subalgebra in M , then the Ad-action of the group U 0(A) (of period 2 unitaries in A)
on L2(M ⊖ A′ ∩ M ) is weak mixing. We will show in the next lemma that one can
choose the corresponding mixing elements in U 0(A) so that to be approximately
τ -independent with respect to any given finite set in M and to have approximately
0-trace. Proving this in the abelian case is quite straightforward. But due to its
possible independent interest, we will actually prove this type of result for arbitrary
diffuse von Neumann subalgebras B ⊂ M , a fact that will make the argument a bit
more lengthy.
3.4. Lemma. Let M be a finite von Neumann algebra and B ⊂ M a diffuse von
Neumann subalgebra. Given any finite dimensional subspaces X ⊂ L2(M ⊖ (B ∨
(B′ ∩ M ))), Y ⊂ L1M , and any δ > 0, there exists a period 2 unitary element
u ∈ B such that τ (uξ∗
Proof. We first prove that given any finite dimensional subspace H0 ⊂ L2(M ⊖ B ∨
1 u∗ξ2) ≤ δkξ1k2kξ2k2, τ (uη) ≤ δkηk1, ∀ξ ∈ X, η ∈ Y .
KADISON-SINGER PROBLEM
23
(B′ ∩ M )), there exists a diffuse abelian von Neumann subalgebra A ⊂ B such that
EA′∩M (H0) = 0. Since H0 is perpendicular to B∨(B′∩M )), it is also perpendicular
to B′ ∩ M , so by Lemma 3.3 there exists u ∈ U(B) such that τ (ξ∗uξu∗) ≤ ckξk2
2,
∀ξ ∈ H0. By Lemma 3.1, if c = 2−7 and n1 = 27, then there exists a partition of
1 with n1 projections {e1
jk2 ≤ 3/4kξk2, ξ ∈ H0. Since
H0 ⊥ B ∨ B′ ∩ M , we also have Σje1
1 ∩ M ),
where B1 = Σje1
j Be1
j (cf. Lemma 2.1 in [P1]). We can thus continue recursively,
replacing H0 by Σje1
jH0e1
j and B by B1, to get a partition of 1 with n2 projections
{e2
jk2. Thus,
j}j in B such that kΣje1
j ⊥ Σje1
j ξe1
j (B ∨ B′ ∩ M )e1
k}k ⊂ B1, which refines {e1
j}j and satisfies kΣje2
j = B1 ∨ (B′
jH0e1
kξe2
j ξe1
kk2 ≤ 3/4kΣke1
kξe2
kk2, ∀ξ ∈ H0.
kΣke2
j ξe1
jk2 ≤ (3/4)2kΣke2
kξe2
kk2 ≤ 3/4kΣje1
j ξem
j }nm
j=1 ⊂ B, where nm = 27m, such that kΣiem
If we define A to be the von Neumann algebra generated by {em
By iterating this procedure we get a sequence of finer and finer partitions of 1,
j k2 ≤ (3/4)mkξk2 ≤ εkξk2,
{em
∀ξ ∈ H0 (Note that the number n = nm of projections necessary to get (3/4)m ≤ ε
is majorized by 27 ln(1/ε)/ ln(4/3)+1, thus n ≤ 27(1/ε)7 ln 2/ ln(4/3) ≈ 27(1/ε)17.02, so
the order of magnitude of the size of the partition is ε−17.02).
1 ≤ j ≤
nm, m ≥ 1}, it follows that EA′∩M (ξ) = 0, ∀ξ ∈ H0. Consider now the group
U 0 = U 0(A) and note that it is Polish with respect to the topology implemented
by k k2. Also, (U 0, k k2) is connected, in fact even contractible (due to the same
construction as in the above proof of Lemma 3.3). Consider the Hilbert space K =
HS(L2(sp(AF A)))⊕ HS(L2M ) and the unitary representation π of U 0 on K given
by u 7→ Ad(LuRu)⊕ Ad(Lu), ∀u ∈ U 0, where HS(H) denotes the space of Hilbert-
Schmidt operators on the Hilbert space H (i.e. HS(H) = {x ∈ B(H) T r(x∗x) <
∞}), and Lu (resp. Ru) are the operators of left (resp. right) multiplication by
u ∈ U 0 ⊂ A. Thus, if we identify in the usual way HS(H) ≃ H⊗H∗, then for each
ξ1, ξ2 ∈ sp(AH0A), η1, η2 ∈ L2M we have
j
π(u)(ξ1 ⊗ ξ∗
2 ⊕ η1 ⊗ η∗
2 ) = (uξ1u∗ ⊗ uξ∗
2 u∗ ⊕ uη1 ⊗ η∗
2 u∗).
Let p = p1 ⊕ p2 ∈ K, where p1 is the orthogonal projection of sp(AH0A) onto
H0 and p2 is the orthogonal projection of L2M onto this same space. Let Kp =
cow{π(u)(p) u ∈ U 0}. Note that π(u)(Kp) = Kp and kπ(u)(x)kK = kbkK, ∀x ∈ K.
Note also that all elements x in Kp are of the form x = x1 ⊕ x2, with x1 ∈
HS(sp(AF A)) ⊂ B(sp(AF A)), x2 ∈ HS(L2M ) ⊂ B(L2M ) positive operators when
viewed as acting on the corresponding Hilbert space.
Since Kp is convex, weakly closed and bounded in K, there exists a unique
element b = b1⊕b2 ∈ Kp of minimal Hilbert norm k kK. Since π(u)(b) ∈ Kp and has
24
SORIN POPA
the same norm as b, it follows that π(u)(b) = b, ∀u ∈ U0. Thus, Ad(LuRu)(b1) =
b1, Ad(Lu)(b2) = b2, ∀u ∈ U 0. This means that, as (positive) operators on the
corresponding Hilbert space, b1 commutes with LuRu and b2 with Lu, ∀u ∈ U 0.
Thus, the spectral decomposition of b1 (resp b2) commutes with these unitaries.
Since b1, b2 are Hilbert-Schimdt, they are in particular compact, so any spectral
projection corresponding to (t, ∞) for t > 0 is finite dimensional. It follows that
if b1 6= 0 (resp. b2 6= 0) then there exists a non-zero finite dimensional subspace
H1 ⊂ sp(AF A) such that uH1u∗ = H1 (resp. H2 ⊂ L2M such that uH2 = H2),
∀u ∈ H2.
Let us first notice that this implies H2 = 0. Indeed, because U 0H2 = H2 implies
AH2 = H2, which contradicts the finite dimensionality of H2, unless H2 = 0.
To see that H1 = 0 as well, note that H1 ∋ ξ 7→ Ad(u)(ξ) ∈ H1, ∀u ∈ U 0
defines a continuous unitary representation of the abelian Polish group U 0 on H1.
Since the image of this representation is an abelian subgroup of U(H0), it can
be diagonalized. Thus, H0 = Σj Cξj, with ξ1, ξ2, ...., ξn an orthonormal basis of
H0 such that Ad(U 0)(ξj) ⊂ Tξj, ∀j. Since any element in U 0 has period 2, we
actually have Ad(U 0)(ξj) ⊂ {±ξj}, ∀j. But as we have noticed above, the Polish
group (U 0, k k2) is connected (even contractible), implying that uξju∗ = ξj, or
equivalently uξ = ξu, ∀u ∈ U 0, ∀j. Hence, aξ = ξa for all ξ ∈ H1 and all a ∈ A
(because U 0 generates A as a von Neumann algebra), i.e. H1 ⊂ L2(A′ ∩ M ). But
EA′∩M (F ) = 0 implies EA′∩M (AF A) = 0 and thus EA′∩M (sp(AF A)) = 0, so in
particular EA′∩M (H1) = 0. We have thus proved that H1 ⊂ L2(A′ ∩ M ) and
H1 ⊥ L2(A′ ∩ M ), showing that H1 = 0.
This implies b = 0 and thus 0 ∈ Kp. Hence, for any δ > 0 there exists u ∈
U 0 such that T r(π(u)(p)p)) < δ.
Indeed, for if there exists δ0 > 0 such that
T r(π(u)(p)p)) ≥ δ0, ∀u ∈ U 0, then T r(xp) ≥ δ0 for all x ∈ Kp, in particular for
x = 0 ∈ Kp, giving 0 ≥ δ0, a contradiction.
But T r(π(u)(p)p) < δ implies that we have both uH0 ⊥δ H0 and uH0u∗ ⊥δ H0,
in particular τ (uξ∗
1 u∗ξ2) < δkξ1k2kξ2k2, τ (uη) < δkηk1, for all non-zero elements
ξ1, ξ2 ∈ X, η ∈ Y ′Y ′∗ + C1. Thus, if we embed Y (⊂ L1M ) in some Y ′Y ′∗ for some
appropriate finite subspace Y ′ ⊂ L2M , then all required conditions are satisfied.
(cid:3)
3.5. Lemma. Let M be a finite von Neumann algebra and H1 ⊂ L2M , H2 ⊂ L1M
finite dimensional spaces. Given any δ > 0, there exists δ′ > 0 such that if x ∈ M
satisfies kxk ≤ 1 and kxk2 ≤ δ′, then kxξk2 ≤ δkξk2, kxηk1 ≤ δkηk1, ∀ξ ∈ H1,
η ∈ H2.
Proof. The first part follows from the fact that norm k k2 implements the so-
topology on the unit ball of M while the product with a compact operator (such as
KADISON-SINGER PROBLEM
25
the orthogonal projection of L2M onto the finite dimensional space H1) turns so-
convergence into operator norm convergence. To prove the second part, note that
it is sufficient to show that given any δ > 0 and any finite set {ηi}i ⊂ H2 which
is "δ/2-dense" in the L1-unit ball of H2, there exists δ′ > 0 such that if x ∈ M
satisfies kxk ≤ 1, kxk2 ≤ δ′, then kxηik1 ≤ δ/2. In turn, this fact is an immediate
consequence of the first part, the Cauchy-Schwartz inequality and the fact that
any η ∈ L1M , kηk1 = 1 can be decomposed as a product ξ1ξ2 with ξi ∈ L2M ,
kξ1k2 = kξ2k2 = 1.
(cid:3)
3.6. Theorem. Let n ≥ 1 be an integer. Given any finite von Neumann algebra
M , any diffuse von Neumann subalgebra B ⊂ M , any finite sets X ⊂ L2(M ⊖
B ∨ (B′ ∩ M )), Y ⊂ L1M and any α > 0, there exists a finite dimensional von
Neumann subalgebra C ⊂ B generated by 2n minimal projections of trace 2−n such
that
(a) τ (a1ξ1a2ξ2) ≤ αka1k2ka2k2, ∀ξ1, ξ2 ∈ X, ∀a1, a2 ∈ C ⊖ C.
(b) τ (ηa) ≤ αkak, ∀a ∈ C ⊖ C, ∀η ∈ Y ∪ XX ∗.
In particular, if q1, ..., q2n ∈ C are the minimal projections in C, then
2 − kξk2
2kτ (qi)τ (qj) ≤ 3 · 2−nα, ∀i, j, ∀ξ ∈ X;
(a′) kqiξqjk2
(b′) τ (ηqi) − τ (η)τ (qi)) ≤ α, ∀i, ∀η ∈ Y ∪ XX ∗.
(c′) kqiξqik2 ≤ (2−n/2kξk2 + 2α1/2)kqik2; kΣiqiξqik2
(d′) kqiξqik1 ≤ (2−n/2kξk2 + 2α1/2)τ (qi), ∀i, ∀ξ ∈ X.
Proof. Note that, without any loss of generality, we may assume X = X ∗, Y = Y ∗,
kξk2 = 1, kηk1 = 1, ∀ξ ∈ X, ∀η ∈ Y .
We prove the statement by induction over n ≥ 0. If n = 0 then C = C1 and the
conditions are trivially satisfied. Assume we have proved the statement up to some
n. Thus, there exists an abelian 2n-dimensional ∗-subalgebra C0 ⊂ B generated by
minimal projections q0
2n ∈ B of trace 2−n such that for all a, a1, a2 ∈ C0 ⊖ C,
ξ1, ξ2 ∈ X, η ∈ Y ∪ XX ∗ ∪ {1} we have
2 + 3α, ∀i, ∀ξ ∈ X.
2 ≤ 2−nkξk2
1, ..., q0
(1)
τ (a1ξ1a2ξ2) ≤ α′ka1k2ka2k2, τ (ηa) ≤ α′kak,
where α′ = 2−n−2α.
i Xq0
Denote B0 = ΣiqiBqi and let X0, respectively Y0, be the linear span of the finite
i 1 ≤ i ≤ 2n}. Note that the condition
0 ∩ M . Indeed, by Lemma 2.1 in [P1], we
i and so if x ∈ X, y ∈ B ∨ B′ ∩ M , then
0 ∪ {q0
set Σi,jq0
X ⊥ B ∨ B′ ∩ M implies X0 ⊥ B0 ∨ B′
i (B ∨ B′ ∩ M )q0
have B0 ∨ B′
j , respectively Y ∪ X0X ∗
0 ∩ M = Σiq0
26
SORIN POPA
j yq0
the latter equality being due to the fact that q0
i ) = δjkτ (q0
kΣiq0
j xq0
j xq0
i yq0
τ (q0
j yq0
j ) = 0,
j ) = δjkτ (xq0
j yq0
j ∈ B ∨ B′ ∩ M .
Let δ = 2−n−2α. By Lemma 3.5, there exists 1 ≥ δ′ > 0 such that if x ∈ M
satisfies kxk ≤ 1 and kxk2 ≤ δ′, then kxξk2 ≤ 3−1δkξk2, kxηk1 ≤ 3−1δkηk1,
∀ξ ∈ X0, η ∈ Y0. By Lemma 3.4, there exists v ∈ U 0(B0) such that
(2)
τ (vξ∗
1v∗ξ2) ≤ 3−1δkξ1k2kξ2k2, τ (ηv) ≤ δ′2kηk1, ∀ξ1, ξ2 ∈ X0, η ∈ Y0.
i ) ≤ 2−nδ′2 (by last
Since v is a period 2 unitary commuting with all q0
part of (2) and the fact that q0
i ∈ Y0), using the fact that B0 is diffuse we can split
i = q2i−1 + q2i of trace 2−n−1
each projection q0
such that the period 2 unitary element u = Σi(q2i−1 − q2i) satisfies ku − vk2 ≤ δ′.
Thus, u satisfies k(v − u)ξk2 ≤ 3−1δkξk2, k(v − u)ηk1 ≤ 3−1δkηk1, ∀ξ ∈ X0, η ∈ Y0.
Combining with (2) and using the Cauchy-Schwartz inequality, we get:
i into the sum of two projections q0
i , and τ (vq0
(3)
τ (uξ∗
1u∗ξ2) ≤ τ (vξ∗
1 u∗ξ2) + k(u − v)ξ∗
1k2ku∗ξ2k2
≤ τ (vξ∗
1v∗ξ2) + kvξ∗
≤ τ (vξ∗
1k2k(u − v)ξ2k2 + k(u − v)ξ∗
1k2ku∗ξ2k2
1v∗ξ2) + 2δkξ1k2kξ2k2/3 ≤ δkξ1k2kξ2k2.
Moreover, since δ′ ≤ 1/3 we have δ′2 ≤ δ/3′ and thus
(3')
τ (uη) ≤ τ ((u − v)η) + τ (vη) ≤ δkηk1, ∀η ∈ Y0.
Denote C the linear span of {qj 1 ≤ j ≤ 2n+1} and note that C = C0 + uC0,
2 + kbik2
2,
C0 ⊥ uC0. Let xi = ai + ubi ∈ C, with ai, bi ∈ C0. Thus, kxik2
i = 1, 2. Take ξ1, ξ2 ∈ X. Then we have
(4)
2 = kaik2
τ (x1ξ1x2ξ2)
= τ (a1ξ1a2ξ2) + τ (b1uξ1a2ξ2) + τ (a1ξ1ub2ξ2) + τ (b1uξ1b2uξ2)
= τ (a1ξ1a2ξ2) + τ (u(ξ1a2ξ2b1)) + τ (u(b2ξ2a1ξ1)) + τ (u(ξ1b2)u(ξ2b1))
By (1), for the first term on the last line in (4), we have the estimate
(5)
τ (a1ξ1a2ξ2) ≤ α′ka1k2ka2k2 ≤ α′kx1k2kx2k2.
KADISON-SINGER PROBLEM
27
Since ξ1a2ξ2b1 and b2ξ2a1ξ1 belong to Y0, by (3′) it follows that for the second
term on the last line of (4) we have the estimate:
(6)
τ (u(ξ1a2ξ2b1)) ≤ α′kξ1a2ξ2b1k1 ≤ α′kξ1a2k2kξ2b1k2
≤ α′ka2kkb1k ≤ 2nα′kb1k2ka2k2 ≤ 2nα′kx1k2kx2k2,
where for the last row we have used the fact that for a ∈ C0 we have kak ≤ 2n/2kak2.
Similarly, for the third term of the last line in (4) we have
τ (u(b2ξ2a1ξ1)) ≤ 2nα′kx1k2kx2k2.
(7)
Finally, for the fourth term of the last line in (4), by (3) and the fact that
ξ1b1, ξ2b2 ∈ X0, we get
(8)
τ (u(ξ1b2)u(ξ2b1)) ≤ α′kξ1b2k2kξ2b2k2
≤ α′kb1kkb2k ≤ 2nα′kb1k2kb2k2 ≤ 2nα′kx1k2kx2k2.
By combining (4) − (8), we thus obtain for all ξ1, ξ2 ∈ X and x1, x2 ∈ C
τ (x1ξ1x2ξ2) ≤ 4 · 2nα′kx1k2kx2k2 ≤ αkx1k2kx2k2
Similarly, if η ∈ Y and x = a + bu ∈ C, then ηb ∈ Y0, kηbk1 ≤ 2nkbk ≤ 2nkxk,
and by the second part of (1) and (3′) we get
(9)
(10)
τ (xη) ≤ τ (aη) + τ (u(ηb))
≤ α′kak + δkηbk1 ≤ α′kxk + 2nδkxk ≤ αkxk,
showing that C satisfies conditions (a) and (b) of the statement.
If we now assume (a) and (b) are satisfied and combine them with the identity
2 = τ (qiξ∗qj ξ) = τ ((qi − τ (qi))ξ∗qjξ) + τ (qi)τ (qjξξ∗)
kqjξqik2
= τ ((qi − τ (qi))ξ∗(qj − τ (qj))ξ) + τ (qj)τ ((qi − τ (qi))ξ∗ξ)
+τ (qi)τ ((qj − τ (qj))ξξ∗) + τ (qi)τ (qj)τ (ξ∗ξ)
then we get:
kqjξqik2
2 − τ (qi)τ (qj)kξk2
2 ≤ 2−nα + 2 · 2−nα = 3 · 2−nα.
This proves that (a) and (b) imply (a′), while (b′) is trivial from (b) and (c′) from
(a′). Finally, (d′) follows from the first part of (c′), via the Cauchy-Schwartz in-
equality:
kqiξqik1 = sup{τ (qiξqix) x ∈ M, kxk ≤ 1}
≤ kqiξqik2 sup{kqixk2 x ∈ M, kxk ≤ 1} = kqiξqik2kqik2.
(cid:3)
28
SORIN POPA
3.7. Corollary. Let M be a II1 von Neumann algebra and A ⊂ M a MASA in
M . Let X ⊂ M ⊖ A, Y ⊂ M be finite sets, n ≥ 1 an integer and α > 0. There
exists a partition of 1 with projections q1, ..., q2n ∈ A of trace 2−n such that if C
denotes the algebra generated by {qi}i then for all x ∈ X, y ∈ Y and i = 1, 2, ..., 2n
we have:
(a) τ (a1x1a2x2) ≤ αka1k2ka2k2, ∀x1, x2 ∈ X, ∀a1, a2 ∈ C ⊖ C
(b) τ (yqi) − τ (y)τ (qi)) ≤ α, ∀i, ∀y ∈ Y ∪ XX ∗.
Moreover, we have for all x ∈ X and 1 ≤ i, j ≤ 2n:
2 − kxk2
(c) kqixqjk2
2τ (qi)τ (qj) ≤ 3 · 2−nα.
(d) kqixqik1 ≤ (2−n/2kxk2 + 2α1/2)τ (qi).
Proof. Since A is a MASA we have A′ ∩ M = A and since M is II1, A must be
diffuse. Thus, Theorem 3.6 applies.
(cid:3)
3.8. Lemma. Let M be a finite von Neumann algebra and A ⊂ M a MASA.
Given any separable von Neumann subalgebra Q0 ⊂ M , there exists a separable von
Neumann algebra Q ⊂ M that contains Q0, such that EA(Q) = A ∩ Q (i.e. Q
and A ⊂ M make a commuting square, as in Sec. 1.2 in [P4]) and A0 = A ∩ Q is
maximal abelian in Q.
Proof. First note that by Theorem 3.6, given any countable set X = {xn}n ⊂
M there exists a countably generated abelian von Neumann subalgebra B1 ⊂ A
Indeed, this is obtained by taking B1 to be
such that EB′
generated by the set {EA(xn)}n and partitions {pn
j,m}j ⊂ A, n, m ≥ 1, satisfying
j,mk2 ≤ 2−m (which exist by Theorem 3.6). If we then take
kΣjpn
X to be a k k2-dense subset in the unit ball of Q0, it follows that the von Neumann
algebra Q1 generated by B1 and Q0 satisfies:
(1) B1 ⊂ A is separable and satisfies EB′
and is separable;
1∩M (Q0) ⊂ B1; Q1 is generated by Q0, B1
j,m(xn − EA(xn))pn
1∩M (xn) ⊂ B1, ∀n.
Using this first part, it follows that we can construct recursively an increasing
sequence of inclusions of separable von Neumann algebra Bn ⊂ Qn, n ≥ 1, satisfying
the properties:
(2) Bn ⊂ A, EB′
by Bn and Qn−1.
n∩M (Qn−1) ⊂ Bn and Qn is the von Neumann algebra generated
If we now define A0 = ∪nBn
clearly satisfied.
w
and Q = ∪nQn
w
, then all required conditions are
(cid:3)
KADISON-SINGER PROBLEM
29
3.9. Theorem. Let Mn be a sequence of finite factors with dimMn → ∞ and for
each n, let An ⊂ Mn be a MASA. Denote by A = ΠωAn ⊂ ΠωMn. Let Q ⊂ ΠωMn
be an arbitrary separable von Neumann subalgebra such that EA(Q) = A ∩ Q, i.e.
Q and A ⊂ ΠωMn make a commuting square, and denote B1 = A ∩ Q. There
exists a diffuse von Neumann subalgebra B0 ⊂ A such that B0 is 2-independent to
Q ⊖ B1 and τ -independent to B1, more precisely:
(a) τ (x1a1x2a2) = 0, ∀x1 ∈ Q, x2 ∈ Q ⊖ B1, a1 ∈ B0 ⊖ C, a2 ∈ B0.
(b) τ (xa) = τ (x)τ (a), ∀x ∈ Q, a ∈ B0, i.e. (B0 ∨ B1, τ ) ≃ (B0, τ )⊗(B1, τ ).
(c) kexfk2
Proof. Let y0 = 1, y1, ... ∈ Q, be k k2-dense in the unit ball of Q and denote
xn = yn − EA(yn). By Corollary 3.7, for each n ≥ 1, there exists a 2n-dimensional
∗-subalgebra Cn ⊂ A generated by minimal projections of trace 2−n such that if
we denote by un ∈ Cn a unitary element with the properties u2n
n) = 0,
1 ≤ k ≤ 2n − 1, then the following inequalities hold true for all 1 ≤ k, l ≤ n,
1 ≤ p, q ≤ 2n − 1:
(1)
2, ∀x ∈ Q ⊖ B1, e, f ∈ P(B0).
2 = τ (e)τ (f )kxk2
n = 1, τ (up
τ (up
nxkuq
nxl) < 1/n; τ (ykup
n) < 1/n.
This implies that if un = (un,m)m, xn = (xn,m)m, yn = (yn,m)m are so that
un,m ∈ U(Am), kyn,mk ≤ kynk and xn,m = yn,m − EAm(yn,m), then for each n
there exists a neighborhood Vn of ω such that if m ∈ Vn then for all 1 ≤ k, l ≤ n,
1 ≤ p, q ≤ 2n − 1 we have
(2)
n,mxl,m) < 1/n; τ (yk,mup
n,m) < 1/n.
n,mxk,muq
τ (up
Define u = (um)m ∈ A by letting um := un,m, ∀m ∈ Vn\Vn−1. Conditions (2) then
imply that u is a Haar unitary element in A and that B0 := {un}′′
n ⊂ A satisfies
the independence conditions (a), (b) (and thus (c) as well).
(cid:3)
4. Asymptotic freeness and paving over singular MASAs
Recall from [D1] that a MASA A in a finite von Neumann algebra M is singular if
the only unitaries in M that normalize A are the unitaries in A, i.e. NM (A) = U(A).
It is easy to see that the existence of such a MASA in a finite von Neumann algebra
M implies M is necessarily of type II1 (unless M = A). For concrete examples of
singular MASAs in II1 factors, see [D1], [P2] and Section 5.1 below. Note that by
[P3], any separable II1 factor has singular MASAs. The prototype singular MASA
in the hyperfinite II1 factor R is the abelian von Neumann algebra L(Z) generated
30
SORIN POPA
by the canonical unitary implementing the Bernoulli action Z y X = [0, 1]Z, in
the representation of R given by the Murray-von Neumann group measure space
construction [MvN1], R = L∞(X) ⋊ Z.
The main result of this section shows that if A ⊂ M is a singular MASA in
a finite von Neumann algebra, then the associated ultrapower MASA inclusion
Aω ⊂ M ω satisfies the paving property, and thus the KS property as well. In fact,
we prove that given any countable set X = X ∗ ⊂ M ω perpendicular to Aω, there
exists a diffuse subalgebra B0 of Aω which is free independent to X, relative to Aω,
i.e., any alternating word in X, B0 ⊖ C has 0-expectation onto Aω (in other words,
X, B0 are n-independent relative to Aω, ∀n). In particular, due to calculations of
norms in [V2], this implies that any x ∈ X which is a selfadjoint element with two
point spectrum, has the property that any partition of small mesh with projections
in A0, provides a paving of x. As we saw in 2.3.2◦, this is sufficient to ensure that
ANY x ∈ M ω with EAω (x) = 0 can be paved with finite partitions in Aω, and thus
Aω ⊂ M ω satisfies the Kadison-Singer property.
More precisely, we prove the following:
4.1. Theorem. Let S = {An ⊂ Mn}n be a sequence of singular MASAs in finite
von Neumann algebras and denote M = M(S, ω) = ΠωMn, A = A(S, ω) = ΠωAn.
Then we have:
(a) If X ⊂ M ⊖ A, Y ⊂ A are countable sets, then there exists a diffuse von Neu-
mann subalgebra B0 ⊂ A such that B0 is τ -independent to Y and free independent
yixi) = 0, for all k ≥ 1 and all x0 ∈ X ∪ {1},
to X relative to A, i.e., EA(x0
xi ∈ X, yi ∈ B0 ⊖ C, 1 ≤ i ≤ k.
(b) Let B ⊂ M be a countably generated von Neumann subalgebra such that EA(B)
= A ∩ B, i.e. B and A make a commuting square. Then there exists a diffuse von
Neumann subalgebra B0 ⊂ A such that if we denote by B1 = A ∩ B, then B0 and
B1 are in tensor product situation and B ∨ B0 = B ∗B1 (B0 ⊗ B1). In particular, if
B ⊥ A then B ∨ B0 = B ∗ B0.
k
Π
i=1
One should note that by Lemma 3.8, any separable subalgebra Q ⊂ M is con-
tained in a larger von Neumann subalgebra B ⊂ M satisfying the commuting square
condition in part (b) of 4.1.
The above theorem implies that given any countable set X ⊥ A, one can find
partitions of arbitrarily small mesh in A that are free with respect to X. For special
type of elements x ∈ M with 0-expectation on A, such as unitaries or selfadjoint
elements with 2-points spectrum, any such "free paving" diminishes the operator
norm, due to Kesten-type phenomena [Ke] and Voiculescu's calculations of spectra
for products of free-independent variables [V2]:
KADISON-SINGER PROBLEM
31
4.2. Corollary. Let An ⊂ Mn, A, M be as above. Then we have:
1◦ If u ∈ M is a unitary element such that EA(u) = 0, then for any n ≥ 1, there
exists a partition of 1 with n projections q1, ..., qn ∈ A such that kΣn
i=1qiuqik ≤
(√n − 1 + 1)/n.
2◦. If e is a projection in M such that EA(e) = τ (e)1 and τ (e) ≤ 1/2, then for any
n ≥ τ (e)−1 there exists a partition of 1 with n projections q1, ..., qn ∈ A such that
i=1qieqi − τ (e)1k ≤ 2/√n. Also, there exists p ∈ A of trace τ (p) = 1/2 such that
kΣn
kpep + (1 − p)e(1 − p)k ≤ (τ (e)(1 − τ (e))1/2 + 1/2.
As we saw in Proposition 2.3, the paving of projections that expect on scalars
on ultrapowers of MASAs, is in fact sufficient to ensure paving of any element, so
from 4.2 above we deduce:
4.3. Corollary (Kadison-Singer for ultraproduct of singular MASAs). Let
An ⊂ Mn, A, M be as above. Then the inclusion A ⊂ M satisfies the KS property.
Thus, any pure state on A has a unique state extension to M and EA is the unique
conditional expectation of M onto A. Moreover, A ⊂ M has the uniform paving
property, with paving size n(ε) majorized by a scalar multiple of ε−6.
v
v
k
Π
i=1
def= {x0
The proof of Theorem 4.1 will follow quite closely the type of arguments that we
have developed in [P6]. We will thus use extensively the notations, terminology and
technical lemmas from that paper, which we recall here in details, for the reader's
convenience.
4.4. Notation. Let M be a von Neumann algebra. If v ∈ M is a partial isometry
with v∗v = vv∗, X ⊂ M is a subset and k ≤ n are nonnegative integers then denote
def= X and X k,n
X 0,n
vixi xi ∈ X, 1 ≤ i ≤ k − 1, x0, xk ∈ X ∪ {1}, vi ∈
{vj 1 ≤ j ≤ n}}.
4.5. Lemma. Let A be a singular MASA in the finite von Neumann algebra M .
Let ε > 0, n ≥ 1 an integer, f ∈ A a non-zero projection, and F = F ∗ ⊂ M ⊖ A,
Y ⊂ A finite sets. There exists a partial isometry v ∈ Af such that τ (vv∗) > τ (f )/2
and kEA(x)k1 ≤ ε, ∀x ∈
Proof. It is clearly sufficient to prove the statement in case kxk, kyk ≤ 1, ∀x ∈ F ,
y ∈ Y . Let δ > 0. Denote ε0 = δ, εk = 2kεk−1, k ≥ 1. Let W = {v ∈ Af vv∗ ∈
P(A), kEA(x)k1 ≤ εkτ (v∗v), τ (yvk) ≤ ετ (vv∗), ∀1 ≤ k ≤ n, x ∈ F k,n
, y ∈ Y }.
Endow W with the order ≤ in which w1 ≤ w2 iff w1 = w2w∗
1w1. (W , ≤) is then
clearly inductively ordered. Let v be a maximal element in W . Assume τ (v∗v) ≤
τ (f )/2 and denote p = f − v∗v. If w is a partial isometry in Ap and u = v + w,
, and τ (yvk) ≤ ετ (vv∗), ∀y ∈ Y , 1 ≤ k ≤ n.
v
n
∪k=1
F k,n
v
32
SORIN POPA
then for x = x0
k
Π
i=1
uixi ∈ F k,n
u we have
(1)
x = x0Πk
i=1vixi + ΣℓΣiz0,iΠℓ
j=1wij zj,i,
where the second sum is taken over all ℓ = 1, 2, . . . , k and all i = (i1, . . . , iℓ),
with 1 ≤ i1 < · · · < iℓ ≤ k, and where wij = ws whenever vj = vs, z0,i =
x0v1x1 · · · xi1−1p, zj,i = pxij vij +1 · · · vij+1 xij+1 p, for 1 ≤ j < ℓ, and zℓ,i = pxiℓviℓ+1
· · · vkxk .
By applying part (d) of Corollary 3.7 to the finite set X of all elements of the form
pzp − EAp(pzp) ∈ pM p ⊖ Ap, where z is of the form zj,i, for some i = (i1, . . . , iℓ),
1 ≤ j ≤ ℓ − 1, ℓ ≥ 2, as well as to the set Y of elements EAp(pzp) for such z, it
follows that ∀α > 0, ∃q ∈ P(Ap) such that
(2)
kqzq − EAp(pzp)qk1,pM p < ατpM p(q),
kEAp(pzp)qk1,pM p ≤ (1 + α)kEAp(pzp)k1,pM p τ (q)
Since for y1, y2, y ∈ M with ky1k ≤ 1, ky2k ≤ 1 we have kEA(y1yy2)k1 ≤
ky1yy2k1 ≤ kyk1, it follows that for any l ≥ 2 we have:
(3)
kEA(z0,iwi1 z1,iwi2 z2,i . . . wiℓ zℓ,i)k1
≤ kwi1 z1,iwi2k1 = kqz1,iqk1 = kqz1,iqk1,pM pτ (p),
which by applying consecutively to z = z1,i the two inequalities in (2), is further
majorised by
(4)
≤ (kEAp(z1,i)qk1,pM p + ατpM p(q))τ (p)
≤ (1 + α)kEAp(z1,i)k1,pM pτpM p(q)τ (p) + ατpM p(q))τ (p)
= (1 + α)(kEA(z1,i)k1τ (p)−1)(τ (q)τ (p)−1)τ (p) + ατ (q)
= (1 + α)kEA(pxi1 vi1+1 · · · vi2 xi2 p)k1τ (p)−1τ (q) + ατ (q).
But since p lies in A, we have kEA(pyp)k1 = kpEA(y)pk1 ≤ kEA(y)k1 for any y ∈
M . Also, since 1 ≤ i1 < i2 ≤ ℓ and i2−i1 ≤ k−1, the element y = xi1 vi1+1 · · · vi2 xi2
belongs to F m,n
with m = i2 − i1 − 1 ≤ k − 2. Altogether, it follows that the last
term in (4) is majorized by
v
KADISON-SINGER PROBLEM
33
(5)
(1 + α)kEA(xi1vi1+1 · · · vi2 xi2 )k1τ (p)−1τ (q) + ατ (q)
≤ (1 + α)εk−2τ (vv∗)τ (p)−1τ (q) + ατ (q)
≤ (1 + α)εk−2τ (q) + ατ (q),
where the last inequality is due to the fact that τ (vv∗) ≤ 1/2 implies τ (vv∗)/τ (p)
≤ 1. If we now take α ≤ ε0/4, from the first term of (3) and last term of (5), we
get that for all i = (i1, ..., iℓ) with ℓ ≥ 2 we have
(6)
kEA(z0,iwi1 z1,iwi2 z2,i . . . wiℓ zℓ,i)k1 ≤ 2εk−2τ (q).
Since 2εk−2 ≤ εk−1 and since there are at most
in the sum in (1) for which ℓ ≥ 2, from (6) we get
k
Σ
i=2(cid:18) k
i (cid:19) = 2k − k − 1 elements
ℓ
Yj=1
wij zj,i
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)1
≤ (2k − k − 1)εk−1τ (q)
(7)
Xℓ≥2Xi
EA
z0,i
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Finally, from the sum on the right hand side of (1) we will now estimate the
terms with ℓ = 1. These are terms which are obtained from x0v1x1v2x2 . . . vkxk
by replacing exactly one vi by wi, so they are of the form z = z0,iwiz1,i, where
i = 1, 2, ..., k, z0,i = x0v1x1...vi−1xi−1p, z1,i = pxivi+1...vkxk and wi = ws if
vi = vs. Note that there are k of them.
One should notice at this point that in the above estimates we only used the
fact that w∗w = ww∗ = q and that A is a MASA, not the actual form of w, nor
the fact that A is singular. We will make the appropriate choice for w now, to get
the necessary estimates for these last terms (in the process, we will also deal with
the condition τ (yvk) ≤ ετ (vv∗), ∀y ∈ Y , which is required for w to belong to W ).
The singularity assumption on A will play a crucial role, due to the following:
4.6. Lemma. Let A ⊂ M be a singular MASA. Let Y1 = Y ∗
1 ⊂ M ⊖ A, Y2 ⊂ M
be finite sets, q ∈ A a nonzero projection. Given any β > 0 and n ≥ 1 there exists
a unitary element w ∈ Aq such that kEA(y1wiy2)k1 < β, kEA(y2wiy1)k1 < β,
τ (y2wj) ≤ β, for all y1 ∈ Y1, y2 ∈ Y2, 0 < i ≤ n.
Proof. We may clearly assume kyik ≤ 1, ∀yi ∈ Yi, i = 1, 2. Let hM, eAi be
the Jones basic construction von Neumann algebra of the inclusion A ⊂eA M ,
34
SORIN POPA
endowed with its canonical (semifinite) trace T rhM,eAi. Consider the semifinite von
Neumann algebra M = hM, eAi⊕2n ⊕ B(L2M )⊕2n and denote by T r the trace on
M defined by T r(x1, x2, ..., x2n, y1, y2, ..., y2n) = ΣjT rhM,eAi(xj) +ΣiT rB(L2M )(yi).
Let K0 ⊂ M denote the convex hull of the set
1eAy1)w−j, wj(Σy2∈Y2y2eCy∗
2)w−j)1≤j≤n w ∈ U(Aq)} ⊂ M.
{(wj(Σy1∈Y1 y∗
One should notice right away that each j'th entry zj ∈ qhM, eAiq, of the first 2n
coordinates of an element z = (zj)j ∈ K0 satisfies zj = qzjq and eAzj = 0 = zjeA
(the latter because eAwjy1eA = wjeAy1eA = wjEA(y1)eA = 0, ∀y1 ∈ Y1; similarly
eAy1wjeA = 0).
Note further that K0 is bounded both in the operator norm on M (by Σy1ky1k2+
Σy2ky2k2 ≤ Y1+Y2) and in the Hilbert-norm k k2,T r on M (by Σy1kEA(y∗
1y1)k2
2+
Σy2 τ (y2y∗
is a weakly compact
bounded subset in both M and L2(M, T r). In particular, K contains a unique
element b ∈ K with kbk2,T r = min{kzk2,T r z ∈ K}.
2 ) ≤ Y1 + Y2). Thus, its weak closure K = K0
w
Note also that the group U(Aq) acts on K0 by
j)1≤j≤n) = (wjxjw−j, wjx′
σw((xj, x′
jw−j)j, ∀w ∈ U(Aq),
j = b′
j = b′
Hence, if b = (bj, b′
and that this action preserves the Hilbert norm k k2,T r. Thus, σ extends to an
action of U(Aq) on K, still denoted by σ. Since kσw(b)k2,T r = kbk2,T r, by the
uniqueness of b as the element of minimal norm in K, it follows that σw(b) = b,
∀w ∈ U(Aq).
j)j are the 4n components of b, then for each j with 1 ≤ j ≤
n, we have wjbj = bjwj, wjb′
jwj for all w ∈ U(Aq). Since any unitary element
in Aq can be expressed as a j'th power of a unitary in Aq, it follows that ubj = bju,
ub′
ju, ∀u ∈ U(Aq). But since any element in Aq is a linear combination of
unitary elements in Aq, this implies bj ∈ Aq′ ∩ qhM, eAiq = A′ ∩ qhM, eAiq and
b′
j ∈ Aq′ ∩ qB(L2M )q. But by (1.4 in [P8]), the supremum of finite projections in
A′∩qhM, eAiq is equal to the supremum of the projections qveAv∗q with v ∈ N (A).
Since A is singular in M , this implies bj = eAbjeA. But eAbj = 0, and so bj = 0,
∀j. On the other hand, since b′
j ∈ B(L2M ) are Hilbert-Schmidt (thus compact)
and commute with the diffuse algebra Aq, it follows that b′
We have thus proved that 0M = (0, ..., 0) ∈ K. This implies that for any β > 0
j = 0, ∀j, as well.
there exists w ∈ U(Aq) such that
(1)
T rhM,eAi(wj(y1eAy∗
1)w−j(Σy2 y2eAy∗
2))
KADISON-SINGER PROBLEM
35
+T rB(L2M )(wj(Σy2 y2eCy∗
2)w−jeC) < β2,
for all y1 ∈ Y1, where the sums are taken over y2 ∈ Y2. Indeed, for if not then
T rhM,eAi(wj(Σy1 y1eAy1
+T rB(L2M )(wj(Σy2 y2eCy∗
∗)w−j(Σy2 y2eAy∗
2))
2)w−jeC) ≥ β2,
for all w ∈ U(Aq). By taking convex combinations over w ∈ U(Aq) and then weak
limits, this would imply T r((bjΣy2 y2eAy∗
j)j ∈ K, in
particular for b = 0, thus 0 ≥ β2 > 0, a contradiction.
jeC)j) ≥ β2, ∀b = (bj, b′
In particular, any w ∈ U(Aq) that satisfies (1), will also satisfy
2, b′
(2)
T rhM,eAi(wj(y1eAy1
∗)w−jy2eAy∗
2) < β2
T rB(L2M )(wjy2eCy∗
2w−j eC) < β2,
for all y1 ∈ Y1, y2 ∈ Y2 and all j with 1 ≤ j ≤ n. The second set of these
inequalities, translates into τ (wjy2)2 ≤ β2, ∀y2 ∈ Y2, 1 ≤ j ≤ n. At the same
time, by taking into account the definitions of k k1 and of T rhM,eAi, and by using
the Cauchy-Schwartz inequality in (hM, eAi, T r), the first set of the inequalities
entails the estimates
(3)
kEA(y1wjy2)k1 = sup{τ (y1wjy2a) a ∈ A, kak ≤ 1}
= sup{T r(eAy1wjy2eAa) a ∈ A, kak ≤ 1}
≤ T r(eAy∗
1eAy1wjy2eA)1/2T r(eA)1/2
2w−jy∗
1eAy1wjy2eAy∗
and similarly kEA(y2wjy1)k1 ≤ β, ∀y1 ∈ Y1 = Y ∗
= T r(w−jy∗
2)1/2 ≤ β,
1 , y2 ∈ Y2, j = ±1, ±2, ..., ±n.
(cid:3)
End of proof of 4.5: Denote by Z the set of elements of the form x0v1x1...vi−1xi−1p,
or pxivi+1...vkxk, for all possible choices arising from elements in
. By
applying Lemma 4.6 to β = εk−1τ (q)/2k, n ≥ 1 and Y2 = Y ∪ Z ∪ Z ∗ ∪ {EA(z)
z ∈ Z ∪ Z ∗}, Y1 = {y2 − EA(y2) y2 ∈ Y2}, it follows that there exists w ∈ U(Aq)
such that
∪k=1
n
F k,n
v
36
SORIN POPA
(8)
kEA(((x0v1x1 . . . vj−1xj−1 − EA(x0v1x1 . . . vj−1xj−1p)wjxjvj+1 . . . vkxk)k1
≤ εk−1τ (q)/2k,
(8')
kEA(x0v1x1 . . . vj−1xj−1wj (xjvj+1 . . . vkxk − EA(pxjvj+1 . . . vkxk)))k1
(8")
≤ εk−1τ (q)/2k.
τ (wjy2) ≤ ετ (q).
From (8) and (8′), it follows that for each element with ℓ = 1 in the summation
j=1wij zj,i in (1), i.e., of the form x0v1x1 . . . vj−1xj−1wjxjvj+1 . . . vkxk,
ΣℓΣiz0,iΠℓ
we have the estimate:
(9)
kEA(x0v1x1 . . . vj−1xj−1wjxjvj+1 . . . vkxk)k1
≤ 2εk−1τ (q)/2k + kEA(x0v1x1 . . . vj−1xj−1)wjEA(xjvj+1 . . . vkxk)k1
≤ εk−1τ (q)/k + γ,
where γ is the minimum of kEA(x0v1x1 . . . vj−1xj−1)qk1, kqEA(xjvj+1 . . . vkxk)k1,
which by the second inequality in (2) is majorized by the minimum between k(1 +
α)EA(px0v1x1 . . . vj−1xj−1p)k1τ (q) and (1+α)kEA(pxjvj+1 . . . vkxkp)k1τ (q). Since
kEA(pyp)k1 = kpEA(y)pk1 ≤ kEA(y)k1, the latter is majorized by the minimum
between (1+α)kEA(x0v1x1 . . . vj−1xj−1)k1τ (q), (1+α)kEA(xjvj+1 . . . vkxk)k1τ (q).
Both elements x0v1x1 . . . vj−1xj−1, xjvj+1 . . . vkxk belong to some F j,n
v with j ≤
k − 1, and at least one of them with j 6= 0. Thus, by the properties of v we have
γ ≤ (1 + α)εk−1τ (vv∗)τ (q). Since α was taken ≤ ε0/4 ≤ 1/4, one gets γ ≤ εk−1.
Hence, the last term in (9) is majorized by εk−1τ (q)/k + εk−1τ (q). Since there
are k terms with ℓ = 1, obtained by taking j = 1, ..., k, by summing up over j in
(9) and combining with (7), we deduce by applying EA to (1) the following final
estimate:
(10)
kEA(x)k1 ≤ kEA(x0Πk
i=1vixi)k1 + ΣℓΣikEA(z0,iΠℓ
j=1wij zj,i)k1
≤ εkτ (vv∗) + (2k − k − 1)εk−1τ (q) + (k + 1)εk−1τ (q)
KADISON-SINGER PROBLEM
37
= εkτ (vv∗) + εkτ (ww∗) = εkτ ((v + w)(v + w)∗).
At the same time, from (8′′), we get
(10')
τ (ujy2) = τ (vj + wj)y2) ≤ τ (vjy2) + τ (wjy2)
≤ ετ (vv∗) + ετ (q) = ετ (uu∗).
Altogether, this shows that u = v + w ∈ W . Since u ≥ v and u 6= v, this
We conclude that τ (v∗v) > τ (f )/2.
, then εn =
contradicts the maximality of v ∈ W .
21+2+...+nδ < 2n2
If we now take δ ≤ ε/2n2
δ ≤ ε and the statement follows.
(cid:3).
Lemma 4.7. Let An ⊂ Mn, A ⊂ M be as in 4.1. Let X ⊂ M ⊖ A, Y ⊂ A
be countable sets and f a non-zero projection in A. Then there exists a partial
isometry w in Af such that τ (ww∗) ≥ τ (f )/2, EA(x) = 0 and τ (wjy) = 0, for all
n ≥ k ≥ 1, all x ∈ X k,n
Proof. Let X = {xk}k, Y = {yk}k be enumerations of the sets and let xk = (xk
n)n,
n)n be representations of xk ∈ ΠωMn, yk ∈ ΠωAn, which we can take so
yk = (yk
that xk
n ∈ Mn satisfy EAn(xk
n) = 0, for all k. Let also fn ∈ P(An) be so that
f = (fn)n. By applying Lemma 4.5 for the inclusion An ⊂ Mn, the projection
fn ∈ An, the positive element ε = 2−n, the integer n and the finite sets Xn =
{xk
n k ≤ n}, we get a partial isometry wn in Anfn such that
τ (w∗
w and all y ∈ Y .
n k ≤ n}, Yn = {yj
nwn) ≥ τ (fn)/2 and
kEAn(x)k1 ≤ 2−n, ∀x ∈ ∪k≤n
τ (wj
ny) ≤ 2−n, ∀y ∈ Yn, 1 ≤ j ≤ n.
(Xn)k,n
wn ,
But then w = (wn) clearly satisfies the required conditions.
(cid:3)
Proof of 4.1. Since 4.1(b) is an immediate consequence of part 4.1.(a), we only
need to prove the latter. To do this, we construct recursively a sequence of partial
isometries v1, v2, .... ∈ A such that
(i) vj+1v∗
(ii) EA(x) = 0, ∀n ≥ k ≥ 1, ∀x ∈ X k,n
(iii) τ (vi
j ) ≥ 1 − 1/2j, ∀j ≥ 1.
jy), ∀n ≥ 1, y ∈ Y .
j vj = vj and τ (vjv∗
vj
38
SORIN POPA
Assume we have constructed vj for j = 1, ..., m.
then we let vj = vm for all j ≥ m.
f = 1 − v∗
mvm ∈ A. Note that EA(x′) = 0, for all x′ ∈ X ′ def= ∪k≤n
X k,n
vm .
If vm is a unitary element,
If vm is not a unitary element, then let
If we apply now Lemma 4.7 to A ⊂ M, the projection f ∈ A, and to the
countable set X ′ ⊂ M, then we get a partial isometry u ∈ Af such that τ (uu∗) ≥
τ (f )/2 and EAf (x) = 0 for all x ∈ ∪k≤n
u . But then vm+1 = vm + u will satisfy
both (i) and (ii) for j = m + 1.
(X ′)k,n
It follows now from (i) that the sequence vj converges in the norm k k2 to
a unitary element v ∈ A, which due to (ii) and (iii) will satisfy the conditions
required in part (a) of 4.1.
(cid:3)
To deduce 4.2.1◦ from 4.1, we'll need the following:
Lemma 4.8. Let u, v be unitary elements in a finite von Neumann algebra M
such that τ (u) = 0 and τ (vj) = 0, for any non-zero integer j with j ≤ n − 1, for
some n ≥ 2. Assume τ (x0y1x1y2....ykxk) = 0, for any k ≥ 1 and any choice of
yi ∈ {u, u∗}, x1, ..., xk−1 ∈ {vj 1 ≤ j ≤ k − 1} and x0, xk ∈ {vj 1 ≤ j ≤
k − 1} ∪ {1}. Then we have:
(a) {u∗vjuv−j j = 1, 2, ..., n − 1} are freely independent Haar unitaries in M .
(b) kΣn−1
Proof. Part (a) is easy to check and we leave it as an exercise (see e.g.
similar calculations).
j=0 vjuv−jk ≤ √n − 1 + 1.
[AO] for
To deduce part (b), recall that by a well known result of Kesten ([Ke]),
if
w1, ..., wm are freely independent Haar unitaries in a finite von Neumann algebra
M , then kΣm
i=1wik = √m. Thus, by (a) we get:
j=0 vjuv−jk = ku(1 + Σn−1
kΣn−1
j=1 v∗vjuv−jk ≤ 1 + kΣn−1
= k1 + Σn−1
j=1 u∗vjuv−j)k
j=1 u∗vjuv−jk = 1 + √n − 1.
Proof of 4.2. By Theorem 4.1 there exists a diffuse von Neumann subalgebra B0 ⊂
A = Aω such that any word with alternating letters from {u, u∗}, B0 ⊖ C1, has
trace 0. Let v ∈ B0 be a unitary element such that τ (vj) = for j = 1, 2, .., n− 1 and
vn = 1. Thus, if λ ∈ C is a primitive n'th root of 1 then v = Σn−1
k=0 λkek+1, where
ek ∈ B0 are spectral projections of v with τ (ek) = 1/n, ∀k. An easy calculation
shows that n−1Σn−1
k=1ekuek. But then 4.2.1◦ follows from 4.8 (b).
j=0 vjuv−j = Σn
(cid:3)
KADISON-SINGER PROBLEM
39
To prove 4.2.2◦ let B0 ⊂ Aω be free with respect to the von Neumann algebra
Ce+C(1−e). Then the calculation of norms in [V2] shows that if q is any projection
of trace 1/n in B0 with 1/n ≤ τ (e), then kqeq − τ (e)qk ≤ 2/√n. By applying again
[V2] for p ∈ B0 with τ (p) = 1/2, we get ke(p − τ (p)1)ek = (τ (e)(1 − τ (e)))1/2
and thus kpepk ≤ τ (p) + (τ (e)(1 − τ (e)))1/2 = 1/2 + (τ (e)(1 − τ (e)))1/2. Since
τ (1− p) = 1/2 as well, we get similarly k(1− p)e(1− p)k ≤ 1/2 + (τ (e)(1− τ (e)))1/2
(cid:3)
Proof of 4.3. By Proposition 2.2, in order to prove Corollary 4.3, it is sufficient
to prove that any projection e ∈ M whose expectation on A is a scalar multiple
of some projection f ∈ A, can be paved. But this is indeed the case, because
Af ⊂ f Mf is itself an ultraproduct of singular inclusions, for which 4.2 applies.
(cid:3)
5. Final remarks
5.1. Examples of singular MASAs. Dixmier's first examples of singular MASAs
A in II1 factors M ([D1]), were constructed from group-subgroup situations, H ⊂ G,
as A = L(H) ⊂ L(G) = M , with G infinite conjugacy class (ICC) and H ⊂ G
an abelian subgroup satisfying certain conditions. These conditions are met for
instance by wreath product inclusions groups H ⊂ G = K ≀ H, with H infinite
abelian and K non-trivial and by the the inclusions L(Z) ⊂ L(Z ∗ Γ0), for any
non-trivial group Γ0. Another criterion for singularity of MASAs in factors was
found in [P2]. It can be used to recover the previous examples, as well as others.
It shows for instance that A = L∞([0, 1]) is singular in A ∗ N for any finite von
Neumann algebra N . It also shows that the group algebra A = L(H) is singular
in any crossed product II1 factor M = B⊗H ⋊ H, arising from a Bernoulli action
H y B⊗H, for any non-trivial finite von Neumann "base"-algebra B. In fact, by
(3.1 in [P11]), all these MASAs A are singular in the following stronger sense: If
u ∈ U(M ) is so that uAu∗ ∩ A is diffuse, then u ∈ A. This absorption phenomenon
from ([P11]) is actually valid for any inclusion L(H) ⊂ M = N ⋊ H, arising from
a mixing action of H on a finite von Neumann algebra N .
Another strengthening of the notion of singularity for a MASA A ⊂ M was
emphasized in [P3] and it requires that the only automorphisms of M that normalize
A are the inner automorphisms Ad(u) with u ∈ U(A). Such MASAs were called
ultrasingular in [P3], but we will call them supersingular from now on, because they
have the property that any two automorphisms of M that coincide on A must differ
by some Ad(u), with u ∈ U(A). Equivalently, embeddings with same range of M
into another algebra are uniquely determined by their values on A. It was shown in
[P3] that any II1 factor M whose outer automorphim group Out(M ) is countable
(e.g. if M has property T, by [C2]), do have supersingular MASAs.
40
SORIN POPA
We note here that results from (Section 4 and 5 of [P12]) show in particular
that if one reduces the singular M ASA, A = L(Z) ⊂ L([0, 1]Z) ⋊ Z = R, by a
projection p ∈ A which is not fixed by any "rotation" by a character γ ∈ Z, then
Ap ⊂ pRp ≃ R is supersingular. Moreover, if p, q ∈ A are not conjugate by such a
rotation, then Ap, Aq are distinct singular MASAs in R. More precisely, we have:
5.1.1. Theorem [P12]. Let H be a torsion free abelian group (such as H = Z) and
H y X = X H
0 a Bernoulli H-action. Let R = L∞(X) ⋊ H and denote A = L(H).
If p, q ∈ A are non-zero projections and θ : pRp ≃ qRq is an isomorphism carrying
Ap onto Aq, then there exists a character γ ∈ H such that θ is the restriction of
θγ ∈ Aut(R) to pRp. Moreover, the only automorphisms of pRp ≃ R that normalize
Ap are the restrictions of the automorphisms θγ that satisfy γ(Y ) = Y (a.e.), where
Y ⊂ T is the subset with characteristic function χY = p. In particular, if {pt t ∈
(0, 1]} is a family of projections in L(H) with τ (pt) = t, then Apt ⊂ ptRpt ≃ R
provide a family of distinct singular MASAs in the hyperfinite II1 factor, which are
supersingular for t 6∈ Q.
5.2. Characterizations of singularity for MASAs. Another strengthening of
singularity for MASAs was discovered in ([P4]), where it is shown that if A is a
diffuse abelian von Neumann algebra and N is any finite von Neumann algebra,
then A is maximal amenable (equivalently, maximal injective) in A ∗ N .
We notice in 5.2.1 below an immediate consequence of Theorem 4.1, showing that
for any singular MASA A ⊂ M , the ultrapower Aω is maximal amenable in M ω,
i.e., if Aω ⊂ P ⊂ M ω for some amenable von Neumann algebra P , then P = Aω.
Moreover, any P ⊂ M ω that contains Aω and has countable dimension both as a
left and right Hilbert module, must coincide with Aω.
We also provide an alternative characterization of singularity for MASAs in terms
of moments, as those MASAs that contain Haar unitaries which are asymptotically
free with respect to sets perpendicular to it. For this to happen, asymptotic 4-
independence is in fact sufficient. This should be compared to Theorem 3.9 where
it was shown that asymptotic 2-independence occurs for any MASA, and to 5.3.1
below, which shows that in fact in arbitrary MASAs asymptotic 3-independence
occurs as well.
5.2.1. Theorem. Let An ⊂ Mn be a sequence of MASAs in finite von Neumann
algebras Mn and denote A = ΠωAn, M = ΠωMn. The following are equivalent:
1◦ There exists a sequence of projections pn ∈ An such that lim
n→ω
Anpn is singular in pnMnpn, ∀n.
2◦ A is singular in M;
τ (pn) = 1 and
KADISON-SINGER PROBLEM
41
3◦ A is maximal amenable in M;
4◦ If H ⊂ L2M is a Hilbert A-bimodule with dimHA, dimA H ≤ ℵ0 (i.e., ∃X ⊂
H countable such that spAX and spXA are dense in H), then H is of the form
L2Ap, for some p ∈ P(A). In particular, A is maximal among subalgebras P ⊂ M
that contain A and have the property that L2P is countably generated both as a left
and right Hilbert A-module.
5◦ Given any countable set X ⊂ M ⊖ A, there exists B0 ⊂ A diffuse such that
B0, X are free independent relative to A.
6◦ Given any countable set X ⊂ M ⊖ A, there exists B0 ⊂ A diffuse such that
B0, X are 4-independent.
7◦ Given any selfadjoint element x ∈ M ⊖ A, there exists B0 ⊂ A diffuse such
that B0, {x} are 4-independent.
n→ω
τ (pn) = 1, we may assume qn ≤ pn and unqnu∗
Moreover, if An ⊂ Mn are all equal to the same MASA A ⊂ M then the above
are also equivalent to A ⊂ M being singular and in 6◦, 7◦ it is sufficient to take X,
resp. {x} inside M ⊖ A.
Proof. If un ∈ NMn (An), then u = (un)n normalizes A as well, acting non-trivially
iff u 6∈ A. Also, if en is the maximal projection in An with the property that
unen ∈ An, then u acts nontrivially on A iff lim
τ (en) = 0. This shows that
2◦ =⇒ 1◦. The converse is implicit in [P10] (due to Remark 5.2 in [P3]). Indeed,
for if u ∈ NM(A) is not in A, then there exists a non-zero projection q ∈ A such
that uqu∗q = 0 and u, q can be represented by sequences u = (un)n, q = (qn)n,
with un unitaries in Mn, qn projections in An, such that unqnu∗
nqn = 0. Moreover,
n ≤ pn. But by ([P10]),
since lim
n→ω
by the singularity of Anpn ⊂ pnMnpn, for each n there exists a unitary element
vn ∈ qnMnqn such that kEAn(unvnu∗
n)k2 ≤ kqnk2/n. Thus, v = (vn)n ∈ A satisfies
uvu∗ ⊥ A, a contradiction. Thus 1◦, 2◦ are equivalent.
The implication 1◦ =⇒ 5◦ is shown in Theorem 4.1.(a), and 5◦ =⇒ 6◦ =⇒ 7◦
are trivial. To see that 6◦ =⇒ 1◦, assume there exist vn ∈ Mn partial isometries
such that vnv∗
n =
Anvnv∗
n. If we denote v = (vn)n and u ∈ A would be a Haar unitary that's 4-
independent with respect to X = {v, v∗}, then the equality vuv∗u∗ = u∗vuv∗ (due
to abelianess of Aω) implies 0 6= τ (vv∗) = τ (vu∗v∗uvuv∗u∗) = 0, a contradiction.
Taking X = {v + v∗}, this actually proves 7◦ =⇒ 1◦ as well.
The implication 3◦ =⇒ 2◦ is trivial. To prove the converse, note that if
N is any von Neumann algebra that strictly contains A, then there exist two
orthogonal projections p1, p2 ∈ A that are equivalent via some partial isometry v
n, v∗
nvn are mutually orthogonal projections in An and vnAnv∗
42
SORIN POPA
in N (exercise!). If q = 2−1(p1 + p2 + v + v∗), then q is a projection in N such
that EA(q) = 2−1p, where p = p1 + p2. By Theorem 4.1, there exists a diffuse
von Neumann subalgebra B0 ⊂ A such that any alternating word in B0 ⊖ C and
q − 2−1p has trace 0. Thus, the algebras B0p and Cq + Cp are free independent,
implying that if u ∈ B0p is a Haar unitary then u and (v + v∗)u(v + v∗) generate
a copy of the free group factor L(F2) inside pN p. Since amenability is a hereditary
property (by [S]), this shows that N cannot be amenable.
4◦ =⇒ 2◦ is trivial (because if A is not singular and u ∈ NM(A) \ U(A),
then the von Neumann algebra P generated by u and A has the countable set
X = {un n ∈ Z} satisfying spXA, spAX dense in P 6= A. Finally, to prove
2◦ =⇒ 4◦, assume X ⊂ H ⊖ A is a separable subspace such that the span of
both XA and AX are k k2-dense in H ⊖ A. By 4.1.(a), there exists B0 ⊂ A
diffuse such that B0 is free independent to X relative to A. In particular, given
any Haar unitary u ∈ B0, we have EA(x∗
1ux2) = 0, for all x1, x2 ∈ X. Thus,
uX ⊥ spXA = H ⊖ A, a contradiction.
The last part of the statement, when all An ⊂ Mn are assumed to be equal, is
now trivial.
(cid:3)
5.3. Controlling moments through incremental patching. In Theorem 3.9,
we have proved that if A ⊂ M is an arbitrary MASA, then for any countable
X ⊂ M ω ⊖ Aω, there exists a diffuse von Neumann subalgebra B0 ⊂ Aω such that
B0 is 2-independent with respect to X. We chose to prove this through a "global"
construction of finite dimensional approximations of such a 2-independent B0. But
we can also prove this result differently, through the method used in the proofs of
the previous section, and which consists in controlling the moments incrementally,
by patching "infinitesimal pieces" of an appropriate Haar unitary. This method
does use a technical result from Section 3, namely property (d′) of Theorem 3.6,
but which was already known since [P1] (see also A.1 in [P5]): If M is a finite von
Neumann algebra, A ⊂ M a MASA and X ⊂ M ⊖ A a finite set, then given any
ε > 0, there exists a non-zero projection q ∈ A such that kqxqk1 ≤ ετ (q), ∀x ∈ X.
In fact, as shown in 5.3.1 below, the "incremental patching" method can be used
to obtain a slightly stronger result for arbitrary MASAs A ⊂ M , showing that
one can construct separable, diffuse von Neumann subalgebras B0 ⊂ Aω that are
3-independent with respect to any given countable set X ⊂ M ω ⊖ Aω. As we saw
in Theorem 5.2.1, this is the best one can do for an arbitrary MASA, as existence
of a B0 that's 4-independent with respect to any given countable set X ⊂ M ⊖ A
forces A to be singular (in which case B0 can even be chosen free independent with
respect to the given X).
5.3.1. Theorem. Let Mn be a sequence of finite factors with dimMn → ∞ and for
KADISON-SINGER PROBLEM
43
each n, let An ⊂ Mn be a MASA. Denote by A = ΠωAn ⊂ ΠωMn. Let Q ⊂ ΠωMn
be an arbitrary separable von Neumann subalgebra such that EA(Q) = A ∩ Q, i.e.
Q and A ⊂ ΠωMn make a commuting square, and denote B1 = A ∩ Q. There
exists a diffuse von Neumann subalgebra B0 ⊂ A such that B0 is 3-independent
to Q ⊖ B1, more preciseley: τ (xa) = 0, ∀x ∈ Q, a ∈ B0 ⊖ C; τ (x1a1x2a2) = 0,
τ (x1a1x2a2x3a3) = 0, ∀xi ∈ Q⊖ B1, ai ∈ B0 ⊖ C (N.B.: the odd level independence
relations follow from the even ones).
Proof. We proceed along the lines of the proofs of Lemmas 4.5, 4.7 and Theorem
4.1, from the previous section. If F is a subset in a von Neumann algebra and v a
partial isometry with vv∗ = v∗v, then we denote
F k
v,n = {Πk
j=1vij xj xj ∈ F, 1 ≤ ij ≤ n, 1 ≤ j ≤ k}}.
We first prove the following:
Fact. Let M be a finite von Neumann algebra and A ⊂ M a MASA. Given any
finite set F ⊂ M ⊖ A, with kxk ≤ 1, ∀x ∈ F , any n ≥ 1 and any δ > 0, there exists
a Haar unitary v ∈ A such that τ (x) ≤ δ, ∀x ∈
∪k=1
v,n.
3
F k
To prove this, denote by W = {v ∈ A vv∗ ∈ P(A), τ (x) ≤ δτ (v∗v), ∀x ∈
v,n, τ (vm) = 0, ∀m 6= 0}. Endow W with the order ≤ in which w1 ≤ w2 iff
k=1F k
∪3
w1 = w2w∗
1w1. (W , ≤) is then clearly inductively ordered. Let v be a maximal
element in W . Assume τ (v∗v) < 1 and denote p = 1 − v∗v.
If w is a partial
isometry in Ap and u = v + w, then by using that uij = vij + wij and expanding
x = ui1 x1ui2 x2...uik xk ∈ F k
u,n, k = 1, 2, 3, as a binomial product, we have τ (x) =
τ (Πk
j=1vij xj) + Στ (...xj−1wij xj....), and thus
(1)
τ (x) ≤ τ (Πk
j=1vij xj) + Στ (...xj−1wij xj....),
where the sum is taken over all terms that have at least one occurrence of wij . Since
j=1vij xj) ≤ δτ (vv∗). We will prove that we can choose w 6= 0
v ∈ W , we have τ (Πk
so that the summation on the right hand side of (1) is majorized by δτ (ww∗), giving
τ (x) ≤ δτ (vv∗) + δτ (ww∗) = δτ (uu∗). This will contradict the maximality of v,
thus showing that vv∗ = 1, i.e v is a Haar unitary in A. We construct w by first
making an appropriate choice for its support projection q = ww∗, then choosing w
as an appropriate Haar unitary in Aq.
In order to estimate the summation Στ (...xj−1wij xj....) in (1), note the follow-
in case k = 1 the sum has just one member, being of the form τ (wjy), for
ing:
some 1 ≤ j ≤ n, y ∈ X; in case k = 2, the sum has three terms, being of the form
(2)
τ (wj1x1vj2x2) + τ (vj1x1wj2 x2) + τ (wj1x1wj2 x2);
44
SORIN POPA
in case k = 3, the sum has seven terms, being of the form
(3)
τ (wj1x1vj2 x2vj3 x3) + τ (vj1x1wj2 x2vj3 x3) + τ (vj1x1vj2 x2wj3 x3)
+τ (wj1 x1wj2x2vj3 x3) + τ (vj1x1wj2 x2wj3x3) + τ (wj1x1vj2 x2wj3 x3)
+τ (wj1 x1wj2 x2wj3 x3).
Now note that for each summand for which we have 2 or 3 appearances of
non-zero powers of w in the above sums (one term for k = 2 and four terms
for k = 3), such appearances must be consecutive, i.e. they will be of the form
τ (....wiywj...), for some i, j 6= 0, y ∈ F ⊂ M ⊖ A (for one of the terms, one uses
the equality τ (wj1 x1vj2 x2wj3x3) = τ (x1vj2 x2wj3 x3wj1 )).
If q = ww∗, then for
each one of these terms we have τ (....wiywj...) ≤ kqyqk1. By (2.1 in [P1]), or (A.1
in [P5]), or by using 3.6(d′) in this paper, applied to the MASA Ap ⊂ pM p and
the set pF p ⊂ pM p⊖ Ap, one can choose the non-zero projection q ∈ Ap such that
kqyqk1 ≤ 2−3δτ (q), ∀y ∈ pF p. It thus follows that the sum of terms having two or
more appearances of powers of w are majorized by 2−1δτ (q) (because there is one
such term when k = 2 and four when k = 3).
All remaining terms and the case k = 1 have just one occurrence of wj, j 6= 0, i.e
are of the form τ (y1wjy2) = τ (wjEA(qy2y1q)), for some y1, y2 ∈ M , 1 ≤ j ≤ n.
There are k many such terms for each k = 1, 2, 3. Let's denote by Y0 the set of
all y1, y2 which appear this way, and note that this is a finite set in M . Thus
Y = EA(qY0 · Y0q) is finite as well. It is sufficient to find now a Haar unitary in Aq
such that τ (wjy) ≤ 2−8δτ (q), ∀y ∈ Y , 1 ≤ j ≤ n, because then the sum of the
k terms will be majorized by 2−1δτ (q) which added up to the quantity 2−1δτ (q)
that majorizes the terms with at least 2 occurrences of powers of w gives that
for all x ∈ ∪3
u,n, we have τ (x) ≤ δτ (uu∗). Since Aq is diffuse, it contains
a separable diffuse subalgebra A0 ⊂ Aq, which is isomorphic to L∞(T) with the
Lebesgue measure corresponding to τ (q)−1τA0. Let then w0 ∈ A0 be a Haar unitary
generating A0. Since {wm
0 }m tends to 0 in the weak operator topology and Y ⊂ A
0 y) ≤ 2−4δτ (q), for all y ∈ Y and
is a finite set, there exists n0 ≥ n such that τ (wm
m ≥ n0. But then w = wn0
is still a Haar unitary and it satisfies all the required
conditions.
k=1F k
0
This ends the proof of the Fact.
With this in hand, we proceed as follows: Let X0 ⊂ Q be k k2-dense countable
subset and denote by X = {y/kyk y = x − EA(x), x ∈ X0 \ A}. Note that X is a
countable subset of M ⊖ A and each element in X has operator norm equal to 1.
Write X as a sequence {xn}n. For each n we now apply the above Fact to the set
Fn = {x1, ..., xn} and δ = 1/n, to get a Haar unitary vn ∈ A such that
(4)
n xtj ) < 1/n, ∀ij, tj ∈ {1, ..., n}, k = 1, 2, 3.
τ (Πk
j=1vij
KADISON-SINGER PROBLEM
45
Let xn = (xn,m)m, vn = (vn,m)m be representations of the xn's and vn's with
xn,m ∈ Mm, kxn,mk ≤ 1, vn,m ∈ U(Am). Thus, (4) and the fact that vn are Haar
unitaries, translates into
(5)
lim
m→ω
τ (vk
n,m) = 0, ∀k 6= 0; lim
m→ωτ (Πk
j=1vij
n,mxtj ,m) < 1/n, ∀ij, tj ∈ {1, ..., n}.
Let Vn of ω denote the set of all m ∈ N with the property that
(6)
τ (vk
n,m) < 1/n, 1 ≤ k ≤ n; τ (Πk
j=1vij
n,mxtj ,m) < 1/n, ∀ij, tj ∈ {1, ..., n}.
From (5), it follows that Vn corresponds to a closed-open neighborhood of ω in
Ω, under the identification ℓ∞N = C(Ω). With this in mind, define recursively
W0 = N, Wn+1 = Wn ∩ Vn+1 ∩ {n ∈ N n > min Wn} and note that {Wn}n
this way defined is a strictly decreasing sequence of neighborhoods of ω satisfying
Wn ⊂ ∩j≤nVj.
Finally, let u = (um)m ∈ A be defined by um = vn,m, for m ∈ Wn \ Wn−1,
n ≥ 1. It is then immediate to check that u is a Haar unitary element in A and
that the von Neumann algebra B0 it generates satisfies the required 3-independence
conditions.
(cid:3)
5.4. Exact paving size for ultraproducts of singular MASAs. The order
of magnitude of the paving size in Corollary 4.3 should be ε−2, for any x, not only
for x = v unitary element with EA(v) = 0 and projections that expect on scalars
(i.e., the cases covered by Cor 4.2). We pose here two questions which, if answered
in the affirmative, would imply this fact:
(a) Can any x with EA(x) = 0 and norm ≤ 1/2 (or of norm ≤ c for an even
smaller universal constant c > 0) be written as a convex combination of unitaries
having 0-expectation on A? If so, then 4.2.1◦ would imply that n(x, ε) is majorized
by a constant multiple of ε−2 for any x ∈ M ω.
(b) Is it true that if M is a II1 factor and x = x∗ ∈ M ⊖ C, u ∈ U(M ) a
Haar unitary, such that τ (ui1xui2 x...) = 0, for any alternating word with ij 6= 0,
i=1uixu−ik has order of magnitude √n ? Again, if this would hold true in
then kΣn
this generality, then we would not need Proposition 2.3 at the end of the proof of
Theorem 4.3, the result following directly from 4.1(a), with the estimate ε−2 for
the order of magnitude of the paving size.
5.5. A conjecture generalizing Kadison-Singer. While we have not been
able to settle the classic Kadison-Singer problem in its equivalent formulations
of Theorem 2.2, i.e., by proving that one can pave all elements in Rω (resp.
in
46
SORIN POPA
M = ΠωMn×n(C)) over its MASA Dω (resp. D = ΠωDn), we believe this is true
and that in fact the following more general conjecture holds true:
5.5.1. Conjecture: Given any sequence of MASAs in finite factors, An ⊂ Mn, the
ultraproduct inclusion ΠωAn ⊂ ΠωMn has the Kadison-Singer (equivalently, the
paving) property.
References
[Ak] C. Akemann: The dual space of an operator algebra, Trans. Amer. Math. Soc.
126 (1967), 286-302.
[AkA] C. Akemann, J. Anderson: "Lyapunov theorems for operator algebras", Mem.
AMS 94 (1991).
[AkO] C. Akemann, Ostrand: Computing norms in group C∗-algebras, Amer. J. Math.
98 (1976), 1015-1047.
[A1] J. Anderson: Extensions, restrictions and representations of states on C∗-alge-
bras, Trans. Amer. Math. Soc. 249 (1979), 303-329.
Functional Analysis 31 (1979), 195-217.
[A2] J. Anderson: Extreme points in sets of positive linear maps on B(H), Jour.
[BeHKW] K. Berman, H. Halpern, V. Kaftal, G. Weiss: Matrix norm inequalities and the
relative Dixmier property Integral Equations and Op. Theory 11 (1988), 28-49.
[BT] J. Bourgain, L. Tzafriri: On a problem of Kadison and Singer, J. Reine Angew.
Math. 420 (1991), 1-43.
[CaFTW] P. Casazza, M. Fickus, J. Tremain, E. Weber: The Kadison-Singer Problem in
Mathematics and Engineering: a detailed account, in "Operator theory, operator
algebras, and applications" 299-355, Contemp. Math., Vol. 414, Amer. Math.
Soc., Providence, RI, 2006.
[C1] A. Connes: Classification of injective factors, Ann. of Math., 104 (1976), 73-115.
[C2] A. Connes: A type II1 factor with countable fundamental group, J. Operator
Theory 4 (1980), 151-153.
[CFW] A. Connes, J. Feldman, B. Weiss: An amenable equivalence relation is generated
by a single transformation, Erg. Theory Dyn. Sys. 1 (1981), 431-450.
[D1] J. Dixmier: Sous-anneaux ab´eliens maximaux dans les facteurs de type fini, Ann.
of Math. 59 (1954), 279-286.
[D2] J. Dixmier: "Les alg´ebres d'op´erateurs dans l'espace hilbertien", Gauthier-Vill-
ars, Paris 1957, 1969.
[F] J. Feldman: Nonseparability of certain finite factors, Proc. Amer. Math. Soc.
7 (1956), 23-26.
[KS] R.V. Kadison, I.M. Singer: Extensions of pure states, Amer. J. Math. 81 (1959),
383-400.
KADISON-SINGER PROBLEM
47
[KR] R.V. Kadison, J. Ringrose: "Fundamentals of the theory of operator algebras",
Pure and Applied Mathematics, 100, Academic Press, Inc. New York, 1983
[Ke] H. Kesten: Symmetric random walks on groups, Trans. Amer. Math. Soc. 92
(1959), 336-354.
[MvN1] F. Murray, J. von Neumann: On rings of operators, Ann. Math. 37 (1936),
116-229.
[MvN2] F. Murray, J. von Neumann: Rings of operators IV, Ann. Math. 44 (1943),
716-808.
[vN] J. von Neumann: Einige satze uber messbare abbildungen, Ann. of Math. 33
(1932), 574-586.
[OW] D. Ornstein, B. Weiss: Ergodic theory of amenable group actions I. The Rohlin
Lemma Bull. A.M.S. (1) 2 (1980), 161-164.
[PiP] M. Pimsner, S. Popa: Entropy and index for subfactors, Annales Scient. Ecole
Norm. Sup. 19 (1986), 57-106.
[P1] S. Popa: On a problem of R.V. Kadison on maximal abelian *-subalgebras in
factors, Invent. Math., 65 (1981), 269-281.
[P2] S. Popa: Orthogonal pairs of *-subalgebras in finite von Neumann algebras, J.
Operator Theory, 9 (1983), 253-268.
[P3] S. Popa: Singular maximal abelian *-subalgebras in continuous von Neumann
algebras, J. Funct. Analysis, 50 (1983), 151-166.
[P4] S. Popa: Maximal injective subalgebras in factors associated with free groups,
Advances in Math., 50 (1983), 27-48.
[P5] S. Popa: Classification of amenable subfactors of type II, Acta Mathematica,
172 (1994), 163-255.
[P6] S. Popa: Free independent sequences in type II1 factors and related problems,
Asterisque, 232 (1995), 187-202.
[P7] S. Popa: The relative Dixmier property for inclusions of von Neumann algebras
of finite index, Ann. Sci. Ec. Norm. Sup. 32 (1999), 743-767.
[P8] S. Popa: On a class of type II1 factors with Betti numbers invariants, Ann. of
Math 163 (2006), 809-899 (math.OA/0209310).
[P9] S. Popa: On the distance between MASA's in type II1 factors, Fields Institute
Communications, 30 (2001), 321-324.
[P10] S. Popa: Strong Rigidity of II1 Factors Arising from Malleable Actions of w-Rigid
Groups I, Invent. Math., 165 (2006), 369-408. (math.OA/0305306).
[P11] S. Popa: Strong Rigidity of II1 Factors Arising from Malleable Actions of w-Rigid
Groups II, Invent. Math., 165 (2006), 409-453. (math.OA/0407137).
[S] J. Schwartz: Two finite, non-hyperfinite, non-isomorphic factors, Comm. Pure
Appl. Math. 16 (1963), 19-26.
48
SORIN POPA
[V1] D. Voiculescu: Symmetries of some reduced free product C∗-algebras, In: "Op-
erator algebras and their connections with topology and ergodic theory", Lect.
Notes in Math. Vol. 1132, 566-588 (1985).
[V2] D. Voiculescu: Multiplication of certain noncommuting random variables., J.
Operator Theory 18 (1987), 223-235.
[Wa] S. Wasserman: On tensor product of certain group C∗-algebras, J. Funct. Anal.
23 (1976), 239-254.
[We] N. Weaver: The Kadison-Singer problem in discrepancy theory, Discrete Math.
278 (2004), 227-239.
[W] F. B. Wright: A reduction for algebras of finite type, Ann. of Math. 60 (1954),
560 - 570.
Math.Dept., UCLA, Los Angeles, CA 90095-1555
E-mail address: [email protected]
|
1611.07830 | 3 | 1611 | 2017-03-13T10:51:12 | On the definition of spacetimes in Noncommutative Geometry, Part I | [
"math.OA",
"math.DG"
] | In this two-part paper we propose an extension of Connes' notion of even spectral triple to the Lorentzian setting. This extension, which we call a spectral spacetime, is discussed in part II where several natural examples are given which are not covered by the previous approaches to the problem. Part I only deals with the commutative and continuous case of a manifold. It contains all the necessary material for the generalization to come in part II, namely the characterization of the signature of the metric in terms of a time-orientation 1-form and a natural Krein product on spinor fields. It turns out that all the data available in Noncommutative Geometry (the algebra of functions, the Krein space of spinor fields, the representation of the algebra on it, the Dirac operator, charge conjugation and chirality), but nothing more, play a role in this characterization. Thus, only space and time oriented spin manifolds of even dimension are considered for a noncommutative generalization in this approach. We observe that these are precisely the kind of manifolds on which the modern theories of spacetime and matter are defined. | math.OA | math |
On the definition of spacetimes in
Noncommutative Geometry: Part I
Fabien Besnard, Nadir Bizi
March 14, 2017
Abstract
In this two-part paper we propose an extension of Connes' notion of
even spectral triple to the Lorentzian setting. This extension, which we
call a spectral spacetime, is discussed in part II where several natural ex-
amples are given which are not covered by the previous approaches to the
problem. Part I only deals with the commutative and continuous case of
a manifold. It contains all the necessary material for the generalization to
come in part II, namely the characterization of the signature of the met-
ric in terms of a time-orientation 1-form and a natural Krein product on
spinor fields. It turns out that all the data available in Noncommutative
Geometry (the algebra of functions, the Krein space of spinor fields, the
representation of the algebra on it, the Dirac operator, charge conjuga-
tion and chirality), but nothing more, play a role in this characterization.
Thus, only space and time oriented spin manifolds of even dimension are
considered for a noncommutative generalization in this approach. We ob-
serve that these are precisely the kind of manifolds on which the modern
theories of spacetime and matter are defined.
Contents
1 Introduction
2 Local constructions
2.1 General definitions and conventions . . . . . . . . . . . . . . . . .
2.2 Real structures and local Wick rotations . . . . . . . . . . . . . .
2.3 Hermitian forms on the complex Clifford algebra . . . . . . . . .
2.4 Krein products on spinor spaces . . . . . . . . . . . . . . . . . . .
2.5 Characterization of the Lorentz and anti-Lorentz signatures . . .
2.6 Some Krein positive operators on spinor modules . . . . . . . . .
2.7 Real structure and charge conjugation. KO-dimension tables
. .
2.8 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
5
5
7
9
13
16
18
19
23
1
3 Global constructions
3.1 Global real structures
. . . . . . . . . . . . . . . . . . . . . . . .
3.2 Hermitian forms on the spinor bundle . . . . . . . . . . . . . . .
3.3 The Dirac operator and c-compatibility . . . . . . . . . . . . . .
3.4 The canonical spectral "triple" of a semi-Riemannian manifold .
3.4.1 The Krein space of spinor fields . . . . . . . . . . . . . . .
3.4.2 The antilorentzian and Lorentzian cases . . . . . . . . . .
3.4.3 Causality conditions . . . . . . . . . . . . . . . . . . . . .
. . . .
3.5 Wick rotation of the gamma matrices and Dirac operator
A Krein products on algebraic spinors
B C∗-algebra structures on Cl(V )
C Proofs of the claims in subsection 3.3
1 Introduction
23
24
26
27
30
30
31
31
32
35
36
37
Noncommutative geometry, as initiated by Alain Connes, is an operator alge-
braic framework which generalizes Riemannian manifolds in a way which harmo-
niously gathers continuous and discrete spaces, as well as truly noncommutative
examples. This generalization happens to be just enough to allow the recasting
of the Standard Model of particle physics as a noncommutative Kaluza-Klein
theory. This last application, thus far limited to the compact Riemannian case,
has been a strong motivation for the search of a semi-Riemannian extension of
Connes' noncommutative geometry.
There have been several important attempts in that direction already, but
this endeavour is plagued with difficulties. Some are analytical: among them
we find the noncompactness issue and the definition of the spectral action. This
last problem is probably the most elusive for the time being. A different kind
of problem is the characterization of the physical Lorentzian signature among
the general semi-Riemannian ones. This has been investigated in particular in
[Bes 15a], [Fra 11]. The present paper exclusively deals with this last issue. We
will present a solution which, as we will see, is different from existing ones on
some essential points. However we will argue that this solution encompasses
both discrete and continuous spaces, as it should, whereas its predecessors did
not.
However, it should be stressed that the solution we will propose is incom-
plete. What we will do is discuss the commutative and continuous case, and
use it to motivate the general definition of a noncommutative structure which
should correspond to Lorentzian spacetimes in the same way spectral triples cor-
respond to Riemannian manifolds. But in this definition we will totally ignore
the analytical aspects. Hence, in full rigor, it is only applicable to the finite-
dimensional (hence discrete) case. There are two reasons why we are being so
careless. The first and most important is to keep this work within reasonable
2
bounds. The second is that we think it is more urgent to present the reader
with a careful motivation of the general idea, followed by several examples, in
order to let him/her ponder the relevance of our approach, than elaborate the
subtler aspects of the question. These would only set us out of focus, as well
as, maybe, repell readers with a physicist background, who generally tend to
shun discussions which they think are too formal1. That said, we believe that it
should be possible to use the work [D-P-R 13] by van den Dungen, Paschke and
Rennie to complete the work presented here. In fact our results are incomplete
in yet another way: we only deal here with the even (KO-)dimensional case.
Let us now be more specific about what we are going to do. We will argue
that in order to describe Lorentzian manifolds, we must replace a spectral triple
(A,H, D) with a structure (A,K, D) which we call a spectral spacetime. Though
we will not describe in this introduction the details of the definition, let us just
say first that the Hilbert space H appearing in the data of the spectral triple,
must be replaced with a Krein space K. There is no novelty here. We will point
out that this requires us to restrict to the class of space-oriented manifolds. Since
we are also given a total (space+time) orientation, the manifolds which can be
described in a spectral way are space and time oriented ones. More importantly,
we will propose a definition of time orientation in the noncommutative case: such
an orientation will be defined to be a noncommutative 1-form β which must be
real in some sense and turn the Krein product into a scalar product. Much more
importantly perhaps, we will argue that we must not ask A to be a C∗-algebra,
in fact not even a ∗-algebra, just an algebra of operators acting on K. This step is
not as radical as one might think: it is the natural consequence of the downplay
of the Hilbert structure in front of the Krein structure. However, without a
C∗-structure we take the risk of not being able to reconstruct a space out of the
algebra, and this is why we will propose a natural axiom of reconstructibility,
which is that the algebra is closed under the Hilbert adjunction defined by at
least one orientation form β. In such a case the algebra A may posses many
different, yet isomorphic, C∗-structures: one for each orientation form satisfying
the reconstructibility condition.
As we have alluded to above, the main arguments we will give in favour of our
notion of spectral spacetime are the continuous case on one hand, and discrete
examples on the other. These discrete examples are constructed over a finite
graph. The first, which we call the canonical antilorentzian spectral spacetime
over a positively weighted graph is just an appropriate Wick rotation of the usual
spectral triple one can build on such a graph to recover the geodesic distance
thanks to Connes' distance formula. It is easily seen to be a spectral spacetime
but none of the structures proposed before by various authors. The second
example is an elaboration of the first, which we decorate with noncommutative
algebras on the vertices, and discrete parallel transport operators on the edges.
We call it the split Dirac structure over a graph. The Dirac operator in this
structure turns out to have an interesting relation with the discretization of the
1Actually, we are also mainly interested in the potential physical applications of noncom-
mutative geometry, and we believe that in the quantum gravitational regime the analytical
questions will essentially go away.
3
Dirac operator proposed by Marcolli and van Suijlekom in [M-vS 14]: the latter
is obtained by acting with the former on special "graph states" and projecting
out the result on the space of these states. Moreover, we will see that for
the Dirac structure to be a spectral spacetime it is necessary that the discrete
parallel transport operators satisfy relations which are exactly2 those required
of a spin connection on a manifold. Finally, this example will also show us that
the reconstructibility condition, however natural it might be, is quite restrictive
in the noncommutative case since it is equivalent to the existence of a parallel
(i.e. covariantly constant) timelike vector field.
We have approached our subject in a very slow and (we hope) careful way
keeping these general principles in mind while writing the text: stack definitions,
lemmas and propositions until the time is ripe to state a theorem, stay as
geometric as possible, stay as general as possible unless it spoils clarity, always
be clear on what convention is used, be as self-contained as possible, display
(almost) every computation. The resulting pace is admittedly very slow, and
it would certainly have been possible to take some short cuts. This is why we
will give below a little guide for readers who wish to go directly to the main
points. We hope nonetheless that the material which may be skipped at first
could prove useful on a second read for those who wish to deepen some side
question.
The two parts of the paper can essentially be read independently. In the first
part, which is devoted to the classical (i.e. continuous and commutative) case,
we do not claim much originality. It can be seen mainly as a motivation for the
definition of spectral spacetimes coming in the second part. However, we have
tried to give a unified and, we think, new presentation of several subjects (real
structures on Clifford algebras, Wick rotations, canonical hermitian forms on
Clifford algebras and on spinors) which are otherwise treated in isolation from
one another in the literature. Moreover we have proven some results which are
probably known as "folk theorems" by some, but that we could not locate in
the literature. Finally the first part contains some technical lemmas used in the
second part. The second part provides the definition of spectral spacetimes as
well as the study of several examples, as explained above.
The first part of the paper is organized as follows: after this introduction,
we make a somewhat long excursion into Clifford algebras in section 2. The
goal there will be to give a characterization of the Lorentzian signature of the
metric with tools available in noncommutative geometry. To this end, we first
give a description of Wick rotations in terms of commuting real structures,
then we recall the existence of Krein structures on spinor modules which are
compatible with a given real structure, and use them to characterize the Wick
rotations to Euclidean signature (Robinson's alternative, proposition 8). Finally
our goal is achieved in theorem 1, which characterizes the Lorentzian and anti-
Lorentzian signatures by the possibility of using a single vector to turn the Krein
structure on a given spinor module into a Hilbert structure. This section also
2There is actually a small difference, which does not show up in the most natural cases,
and is not worth explaining in this introduction.
4
makes the connection between the real structures on the Clifford algebra and
the charge conjugation operator which is used in noncommutative geometry,
and also contains some extra material on Clifford algebras, e.g. the σ-product,
which extends a given metric to the whole algebra.
In section 3 we globalize the notions introduced in section 2 to semi-rieman-
nian manifolds: real structures on Clifford bundles, Krein structures on spinor
bundle. We point out that Krein structures compatible with a given real struc-
ture exist if and only if the manifold is space or time orientable, according to
the signature (theorem 2). We also give a detailed explanation of the way in
which the geometric properties of the spin connection are encoded in the alge-
braic properties of the Dirac operator and charge conjugation. Finally we recall
how the Dirac operator can be Wick rotated, and we translate our local char-
acterization of Lorentzian signature from the previous section into global terms
(theorem 4), namely in terms of an orientation 1-form. We point out that the
exactness of this form amounts to a causality condition which is called stable
causality. This closes part I.
We advise the hurried reader to use the following strategy: read quickly
subsection 2.4 for the definition of Krein product on spinor space compatible
with a given real structure, then read subsection 2.5 to have theorem 1 in mind,
and skip to section 3.
In the whole paper some notions are defined in the bulk of the text. When
it happens, the name of the notion is always italicized.
2 Local constructions
2.1 General definitions and conventions
In this section we will be interested in the complex Clifford algebra generated by
a real vector space V of even dimension n = 2k equipped with a non-degenerate
quadratic form Q of signature (p, q). The bilinear form associated with Q will
be denoted by B. Our notations and conventions are summarized below:
• As is traditional in physics, what we call the signature of a real bi-
linear symmetric (or complex sesquilinear hermitian) form is the triple
(n+, n−, n0), where n0 is the nullity, n− the negative and n+ the positive
indices of inertia. When the form is non-degenerate, we leave out n0 = 0
and write (n+, n−) instead.
• We give names to the following cases : Lorentz signature is (n− 1, 1), anti-
Lorentz signature is (1, n − 1), Euclidean signature is (n, 0) and neutral
signature is (n, n).
• The real Clifford algebra Cl(V, Q) is defined using the following conven-
5
tion3
v2 = +Q(v)
for all v ∈ V . In later sections, in the context of manifolds, we will put
the emphasis on the bilinear form and write Cl(V, B) instead of Cl(V, Q).
We let Cl(V ) = Cl(V, Q) ⊗ C be its complexification, and V C be the
complexification of V . We consider that V ⊂ Cl(V, Q) ⊂ Cl(V ) in the
natural way, i.e. we do not write down explicitly the embedding i : V →
Cl(V, Q). We still denote by Q the natural extension of Q such that
v2 = Q(v) holds for all v ∈ V C.
• We let c : Cl(V ) → Cl(V ) be the real structure defined by c(a⊗λ) = a⊗ ¯λ.
• We let T : Cl(V ) → Cl(V ) be the unique linear antiautomorphism which
It is called the principal anti-involution.
restricts to the identity on V .
We often write aT instead of T (a).
• We note that c ◦ T = T ◦ c is the unique antilinear antiautomorphism of
Cl(V ) which extends IdV . We will write a× = c(aT ).
• We let γ be the principal involution, that is, the unique automorphism of
Cl(Q) which extends −IdV . This a grading operator which decomposes
the Clifford algebra into the sum Cl(V ) = Cl0(V ) ⊕ Cl1(V ) of its even
and odd parts.
• For any a and invertible g we write Lg(a) = ga, Rg(a) = ag, Adg(a) =
gag−1. The (complex) Clifford group is defined by ΓC = {g ∈ Cl(V )
Adg(V C) ⊂ V C}, the Pin group is P in(Q) = {g ∈ ΓCc(g) = g and ggT =
±1}, and the Spin group is Spin(Q) = P in(Q) ∩ Cl(V, Q)0. Remember
that the Clifford group is generated by non-isotropic vectors, and that its
elements satisfy gg× ∈ R.
• Let e1, . . . , en be a pseudo-orthonormal basis of V . Then we denote by
ω = e1 . . . en the volume element. It depends on the pseudo-orthonormal
basis chosen only up to a sign, which we can fix by chosing an orientation
of V . The volume element anticommutes with every odd element of Cl(V ).
It has the following properties :
ωT = (−1)
2 ω, ω2 = (−1)
• If ρ is a spinor representation, we set χ := (−i)
2 +qρ(ω). It is called the
chirality operator and always satisfies χ2 = 1. When the representation is
c-admissible (to be defined later), it will also satisfy χ× = (−1)qχ.
n
2 +q
n
n
Our general reference on Clifford algebras is [Cru 90]. For future use we note
the following fact : let φ : Cl(V ) → Cl(V ) be an automorphism or antiautomor-
phism which stabilizes V C. Then φ ◦ T and T ◦ φ coincide on V C, hence on the
whole algebra. Thus T and φ commute.
3We suggest the name "anti-Clifford" for the algebra defined using the convention v2 =
−Q(v). We think that making the sign conventions explicit by using the words anti-Clifford
and antilorentzian would be greatly beneficial to the mathematical physics community.
6
2.2 Real structures and local Wick rotations
What we wish to do in this section is to understand Wick rotations algebraically.
We are given at the start the vector space V equipped with a quadratic form
Q. The piece of data (V, Q) is equivalent to Cl(V, Q). We stress that the
Clifford algebra is not to be seen only as a real algebra, but as a real algebra
equipped with a particular set of generators, namely V . Since we want to
change the signature of the quadratic form, it is natural to embed Cl(V, Q) in its
complexification Cl(V ) which will remain constant when Q is varied. Recovering
Q from Cl(V ) amounts to fix a particular real form for the complex algebra
Cl(V ), by way of a real structure, i.e. an involutive antilinear automorphism
which stabilizes V C. The real structure which has Cl(V, Q) as its set of fixed
points will be denoted by c throughout the text. Clearly, the data (V, Q) and
(Cl(V ), c) are equivalent. Given a general real structure σ we define:
• The real subspace of σ-real vectors Vσ := {v ∈ V Cσ(v) = v}.
• The map uσ : v 7→ v+σ(v)
2 + i v−σ(v)
phism of real vector spaces from V onto Vσ.
, which is easily seen to be an isomor-
2
• The bilinear form Bσ(v, v′) := 1
associated quadratic form Qσ(v) := B(σ(v), v).
2 (B(σ(v), v′) + B(σ(v′), v)) on V , and its
Since σ(v) does not in general belong to V it is not obvious at first sight that
Bσ(v, v′) is real. However we can observe that Bσ(v, v′) = 1
2 (σ(v)v′ + v′σ(v) +
σ(v′)v + vσ(v′)) and on this form it is clear that σ(Bσ(v, v′)) = Bσ(v, v′), hence
this complex number is in fact real.
We also note that the restriction of the quadratic form Q to Vσ is real, since
σ(w2) = σ(w)2 = w2 for every w ∈ Vσ. Furthermore, since Vσ + iVσ = V C,
the complex algebra generated by Vσ is Cl(V ), which we can then identify with
Cl(Vσ, QVσ ) ⊗ C. In particular we note that QVσ is non-degenerate.
Finally the calculation Q(uσ(v)) = uσ(v)2 = 1
shows that uσ is an isometry from (V, Qσ) to (Vσ, Q).
2 (vσ(v) + σ(v)v) = Qσ(v)
Hence we have defined from σ a non-degenerate quadratic form Qσ on V
and a real subspace Vσ of V C such that Vσ ⊕ iVσ = V C and QVσ is real.
This construction can be inversed. Consider a n-dimensional real vector
space W of V C such that Q is real on W . Then W ⊕ iW = V C (indeed, if
there exists a nonzero w ∈ W such that iw ∈ W then Q(w + iw) is not real).
Define σ : V C → V C to be antilinear and satisfy σW = IdW . It is easy to check
that σ(v1v2 + v2v1) = σ(v1)σ(v2) + σ(v2)σ(v1), hence σ extends as an antilinear
algebra automorphism.
Hence we have the following one-to-one correspondence:
{ real structures on Cl(V )} ≃ {n − dimensional real subspaces
W ⊂ V C such that QW is real}
7
Since c is a fixed "background" structure, it is natural to be particularly
interested in real structures which commute with it, and we will call them
admissible real structures. They turn out to correspond to Wick rotations of
the quadratic form Q.
Lemma 1 The following are equivalent.
1. The real structure σ is admissible.
2. The subspace Vσ is stable by c.
3. The subspace V is stable by σ.
4. The real structure σ restricts to a Q-orthogonal symmetry of V .
5. The subspaces V+ := V ∩ Vσ and V− := V ∩ iVσ form a Q-orthogonal
decomposition of V .
Proof: It is immediate that (1)⇒(3), and for the converse it suffices to observe
that σ ◦ c and c ◦ σ are two algebra automorphisms of Cl(V ) which coincide on
V . The equivalence between (1) and (2) is obtained by symmetry.
Of course (4)⇒(3) is trivial, and to see that (3)⇒(4) we observe that Q(σ(v)) =
σ(v)2 = σ(v2) = σ(Q(v)) = Q(v), hence σ restricts to a Q-orthogonal transfor-
mation of V which is moreover involutive.
the R-linear operator σV , hence (5) and (4) are equivalent.
Finally we see that V+ is the +1-eigenspace and V− is the −1-eigenspace of
¶
Let us consider an admissible real structure σ. Using the decomposition
V = V+ ⊕ V− given by point 5 above, we see that if v = v+ + v−, with v± ∈ V±,
then the isometry
takes the simple form
uσ : (V, Qσ) −→ (Vσ, QVσ )
v 7−→ v+ + iv−
Hence uσ "puts an i" in front of the elements of V−: this is what is called
a Wick rotation. Clearly Qσ is positive definite iff Q is positive definite on V+
and negative definite on V−. In this latter case we say that u is a rotation "to
Euclidean signature".
As the proposition below shows, admissible real structures can be expressed
in terms of particular elements of the Clifford group.
Proposition 1
1. The real structures are of the form σ = Adb◦c with b ∈ ΓC
such that bc(b) = λ ∈ R. We also have bT = αb, α = ±1, b× = αc(b).
2. The real structure σ = Adb◦c with b ∈ ΓC is admissible iff c(b) = eiθb, with
θ ∈ R. In this case we can choose b to satisfy c(b) = b and b2 = λ = ±1,
in which case we say that it is real and normalized. Then b× = αb, b
belongs to P in(Q), and is unique up to a sign.
8
Proof: It is obvious that σ is an antilinear automorphism which preserves V C
when it is of the form Adb ◦ c with b in the Clifford group. If bc(b) = λ ∈ R it
is moreover an involution since σ2(a) = bc(b)ac(b)−1b−1 = a for all a ∈ Cl(V ).
Conversely, if σ is a real structure, then σ ◦ c is an automorphism of the
Clifford algebra which preserves V C, and it is then of the form Adb with b ∈ ΓC.
Hence σ = Adb ◦ c. Since σ is an involution we have bc(b) = λ ∈ C by the
calculation above. Since c(b) is then equal to b−1 up to a constant, it commutes
with b, from which we obtain that λ is real.
The other properties follow from the ones just proved: since b ∈ ΓC, bb× is
a constant, hence c(b)bT is a constant, and from c(b)b = λ ∈ R we get that b
is proportional to bT . The involutory property of T forces the proportionality
constant to be a sign α. Then b× = c(bT ) = αc(b).
Now it is easy to check that σ commtutes with c iff c(b) = eiθb for some
θ ∈ R. Then b′ = eiθ/2√λ
b is normalized and real and one has Adb = Adb′ . If b′′ is
another normalized and real element such that Adb′′ = Adb then b′′ = µb′ with
µ ∈ C. From reality one has µ ∈ R, and from normalization one has µ2 = 1.
Since b′ satisfies b′(b′)× = ±1 and is real it is in the Pin group.
¶
At this point one might like to have an example of a non-admissible real structure.
For this, consider R2 with an Euclidean metric and (e1, e2) an orthonormal basis.
Then for t ∈ R, let bt be the Clifford group element bt = cosh t + i(sinh t)e1e2. Then
c(bt) = b−t = b−1
. Hence bt satisfies the hypotheses of the first part of proposition 1,
but not the second (except if t = 0). If we denote by V the vector of components of
v ∈ V C in the chosen basis, then the real structure σt = Adbt ◦ c is given matricially
by V 7→ Ot ¯V where Ot is the matrix
t
Ot = (cid:18) cosh 2t
−i sinh 2t
i sinh 2t
cosh 2t (cid:19)
One then easily shows that the metric Bσt associated to the real structure σt =
Adbt ◦ c is cosh 2t times the original metric. More generally, one can show in (even)
dimension n and in the Euclidean case that there always exists an orthonormal basis
of V in which a general real structure is given by a matrix which is a direct sum of Ip,
−Iq, and 2 × 2 blocks ±Otk with Otk as above.
2.3 Hermitian forms on the complex Clifford algebra
Wick rotations to Euclidean signature are very special. This can be best seen
by extending the quadratic form Qσ to the whole Clifford algebra. The auto-
morphism σ◦ T will then appear naturally as an adjunction, and its exceptional
character when Qσ is positive definite will be made manifest.
Let (ei)1≤i≤n be a pseudo-orthonormal basis of V C. For any subset I ⊂
{1; . . . , n} we write eI = ei1 . . . eik where i1 < . . . < ik are the elements of I.
We know that (eI )I⊂{1;...,n} is a basis of Cl(V ). Hence there is a projection map
τn : Cl(V ) → C which sends an element of Cl(V ) to its coordinate on 1 = e∅.
The map τn is called the normalized trace, a name justified by the following
proposition (we refer to [Gar 11] p. 100 for the proof).
9
Proposition 2 The projection τn does not depend on the chosen pseudo-ortho-
normal basis. It is the unique linear form τn : Cl(V ) → R such that τn(ab) =
τn(ba) for all a, b ∈ Cl(V ) and τn(1) = 1. It also satisfies τn(aT ) = τn(a).
real structure, then τn(σ(a)) = τn(a).
It is also easy to see that if a ∈ Cl(V ), then τn(γ(a)) = τn(a), and if σ is a
Remark: We can see the above proposition as the reason behind the fact that a
product of distinct gamma-matrices is always traceless.
Given the normalized trace and a real structure, it is very natural to define
(a, b)σ := τn(σ(aT )b)
(1)
for all a, b ∈ Cl(V ). We call it the σ-product. It has remarkable properties.
Proposition 3 The σ-product is a non-degenerate hermitian form on Cl(V ).
The associated quadratic form restricts to Qσ on V and to Q on Vσ. It satisfies
(w1 . . . wk, w1 . . . wk)σ = Q(w1) . . . Q(wk)
(2)
for any vectors w1, . . . , wk ∈ Vσ. Moreover if (ei)1≤i≤n is a pseudo-orthonormal
basis of Vσ for Q then (eI )I⊂{1;...;n} is a pseudo-orthonormal basis of Cl(V ) for
the σ-product.
Proof: It is obvious that (., .)σ is sesquilinear. Moreover we have (b, a)σ =
τn(σ(bT )a) = τn(σ(b)T a) = τn(aT σ(b)) = τn(σ(aT )b) = (a, b)σ.
If v ∈ V , then (v, v)σ = τn(σ(v)v) = 1
Let us prove (2). We have :
2 τn(σ(v)v + vσ(v)) = Qσ(v).
(w1 . . . wk, w1 . . . wk)σ = τn(σ(wT
k ) . . . σ(wT
1 )w1 . . . wk)
= τn(wk . . . w1w1 . . . wk)
= w2
= Q(w1) . . . Q(wk)
1 . . . w2
k
Let (ei)1≤i≤n be a pseudo-orthonormal basis of Vσ. Then (eI , eI )σ = ±1 by
property (2). Moreover if I 6= J, (eI , eJ )σ = τn(eT
I eJ ) = ±τn(eI∆J ) = 0. Hence
(eI )I⊂{1;...;n} is a pseudo-orthonormal basis of Cl(V ), which shows that (., .)σ
¶
is not degenerate.
Note that the multiplicative property (2) of the σ-product does not generalize
to elements of V . The correct symmetrical statement which apply to elements
of V uses the c-product instead of the σ-product.
Here is an elegant (and useful) property of the σ-product.
Lemma 2 Let φ : (V, B) → (V ′, B′) be an isometry between two real vector
spaces equipped with nondegenerate bilinear forms. Let c, c′ be the canonical
real structures on Cl(V, B) and Cl(V ′, B′) respectively. Then the isomorphism
φ : Cl(V, B) → Cl(V ′, B′) which canonically extends φ transforms (., .)c into
(., .)c′ .
10
Proof: For any a, b ∈ Cl(V, B) we have
( φ(a), φ(b))c′ = τ ′
= τ ′
= τ ′
= τn(c(aT )b) = (a, b)c
n(c′( φ(a)T ) φ(b))
n(c′( φ(aT )) φ(b)), since φ(V ) = V ′
n ◦ φ(c(aT )b)
where the last step follows from the uniqueness of the normalized trace.
¶
Remark: Care must be taken in applying this lemma. For instance if we use
it on uσ we obtain (uσ(a), uσ(b))σ = τn(c(aT ) ∗σ b), where ∗σ is the Clifford prod-
uct corresponding to the quadratic form Qσ. In particular if a = b = v we obtain
(uσ(v), uσ(v))σ = τn(v ∗σ v) = Qσ(v) = Q(uσ(v)) which is correct.
Taking a Q-pseudo-orthonormal basis (ei)1≤i≤n in V , then the decomposi-
tion
n
Mk=0
V k = Cl(V, Q)
(3)
where V k = Vect{eI, I = k} is immediately seen to be orthogonal for the c-
product. The following proposition shows in particular that this decomposition
is independent of the chosen basis.
Proposition 4 Let φ : (V, B) → (V, B) be an isometry. Then φ(V k) = V k for
all k.
Proof: Clearly φ(V k) = V k for k = 0, 1. Let us suppose that φ(V j) = V j for
j ≤ k. The sum V 0 ⊕ . . .⊕ V k ⊕ V k+1 is orthogonal for the c-product, hence by
the lemma the sum φ(V 0) ⊕ . . . ⊕ φ(V k) ⊕ φ(V k+1) = V0 ⊕ . . . ⊕ V k ⊕ φ(V k+1)
also is. Since we obviously have φ(V k+1) ⊂ V 0 ⊕ . . . ⊕ V k+1 we obtain that
φ(V k+1) = V k+1.
¶
Here is another way to understand why the decomposition (3) is independent
of the chosen basis. There is a well-known vector space isomorphism Θ from
the exterior algebra ΛV to the Clifford algebra which is defined by
v1 ∧ . . . ∧ vr
ǫ(σ)vσ(1) . . . vσ(r)
(4)
ΛV −→ Cl(V, Q)
r! Xσ
7−→
1
Since Θ(ei1 ∧ . . . ∧ eik ) = ei1 . . . eik for distinct elements ei1 , . . . , eik of the
pseudo-orthonormal basis, we see that V k = Θ(ΛkV ). Moreover, there is a
well-known way to extend the bilinear form B to the exterior algebra, which is
to decree that ΛjV and ΛkV are orthogonal for j 6= k and to define
(u1 ∧ . . . ∧ uk, v1 ∧ . . . ∧ vk)B := det((B(ui, vj))1≤i,j≤k)
11
It is left to reader to check that Θ is an isometry from (ΛV, (., .)B) to
(Cl(V, Q), (., .)c). Of course what we have described using c can be extended
to a general real structure σ. However, when passing to the complexification,
the reference to the real structure, or quadratic form, vanishes, so that the
decomposition
n
Mk=0
V k ⊗ C = Cl(V )
(5)
is orthogonal for all σ-products. The vector space V k ⊗ C is described in
a basis-independent way as V k ⊗ C = Θ(ΛkV C), where we still write Θ the
natural extension of the isomorphism (4) to complex scalars.
Remark: When g is an element of the Clifford group, (g, g)c = τn(g×g) = g×g
because g×g is real. It is then customary to write N (g) = g×g and call it the spinor
norm.
If U : Cl(V ) → Cl(V ) is a linear map, we will write U ×σ for its adjoint
relative to the σ-product. We will also write
a×σ := σ(aT )
for an element a ∈ Cl(V ). The two notations are consistent thanks to the
property
L×σ
a = La×σ
(6)
which is easily checked. We will now see a simple yet important result which
tells us how the signature of (., .)σ depends on that of Qσ. We call it Garling's
alternative, since the only place where we could locate it is [Gar 11] (p 101),
where a direct combinatorial proof is given. We give below a slightly different
proof based on the following lemma.
Lemma 3 Let (K, (., .)) be a finite dimensional space equipped with a non-
degenerate hermitian form. If there exists U ∈ End(K) such that U ×U = −IdK,
then (., .) is neutral.
Proof: We have (U ψ, U η) = (ψ, U ×U η) = −(ψ, η) for all ψ, η ∈ K. Thus
if K = K+ ⊕ K− is an orthogonal decomposition of K into subspaces where
(., .) is positive definite and negative definite respectively, we see that K =
U K+ ⊕ U K− is an orthogonal decomposition where the signs are swapped.
¶
Since dim(U K±) = dim K± we conclude by Sylvester's law of inertia.
Proposition 5 (Garling's alternative) The σ-product on Cl(V ) is positive def-
inite whenever Qσ is. It is neutral in every other case.
Proof: Let (ei)1≤i≤n be a pseudo-orthonormal basis of Vσ. Using property (2)
of proposition 1 on the basis (eI ) we see that (., .)σ is positive definite if QVσ
is, and we know that QVσ and Qσ have the same signature.
12
that e2
and we conclude by lemma 3.
i = −1. Let U = Lei. Since L×σ
If Qσ (or equivalently QVσ ) is not positive definite, then there exists i such
= Lei , hence U U × = −IdCl(V )
¶
We can now characterize Wick rotations to Euclidean signature in terms of
ei = Le×σ
i
the corresponding real structure. The notations are the same as in lemma 1.
Proposition 6 Let σ be an admissible real structure. The following properties
are equivalent.
1. The couple (V+, V−) of Q-orthogonal supplementary subspaces of V is such
that Q is positive definite on V+ and negative definite on V−.
2. The quadratic form Qσ on V is positive definite.
3. The restriction of the quadratic form Q to Vσ is positive definite.
4. The σ-product (., .)σ is positive definite.
When these properties are satisfied we say that σ is an Euclidean real struc-
ture. Notice that when σ = Adb ◦ c is an Euclidean real structure, and b is
normalized, the signs such that b× = ±b and b2 = ±1 are no longer indepen-
dent. Indeed, b×σ = b−1b×b = b×, hence bb× is a positive operator and can only
be equal to 1.
Remark: Another specific feature of Euclidean real structures is that the involu-
tion a 7→ a×σ and the norm kak∞,σ = supkbkσ =1 kabkσ turn Cl(V ) into a C ∗-algebra.
For more on this, see appendix B.
2.4 Krein products on spinor spaces
In the rest of this section we let ρ : Cl(V ) → End(K) be a representation of
Cl(V ) on a finite dimensional complex vector space K equipped with a non-
degenerate hermitian form (., .)K. We call such a form a Krein product. Re-
member that a Krein space (see [Bog 74]) is a vector space equipped with a a
non-degenerate hermitian form (., .) and at least one operator β, called a fun-
damental symmetry such that β2 = 1 and (., β.) is positive definite and induces
a complete topology. In finite dimension the existence of β is obvious, thus K
equipped with a Krein product is a Krein space.
We will say that ρ is σ-compatible iff
∀a ∈ Cl(V ),∀ψ, φ ∈ K, (ρ(a)ψ, φ) = (ψ, ρ(a×σ )φ)
This is clearly equivalent to ask ρ(v) to be Krein-self adjoint for all v ∈ V .
When the Krein product is σ-compatible we will denote by A×σ the Krein
adjoint of A ∈ End(K).
We can always build a σ-compatible Krein product on K. First we can
suppose without loss of generality that ρ is irreducible. Then we can use ρ
to tranfer ×σ from Cl(V ) to an antilinear antiautomormphism A 7→ A×σ of
13
End(K). Such an antiautomorphism is necessarily the adjoint operation for
some non-degenerate hermitian form (., .)K
σ . To see this, use the simplicity of
End(K) to pick a β such that A×σ = βA†β−1 where A† is the adjoint for some
σ := h., β−1.i
scalar product h., .i. Then it is a simple matter to verify that (., .)K
has the required property.
Since this procedure is essentially in [Rob 88], we will call it Robinson's
transfert principle.
Proposition 7 There exists a non-degenerate hermitian form (., .)K
σ on K such
that ρ is σ-compatible, and when ρ is irreducible it is unique up to multiplication
by a non-zero real number.
We haven't proven the uniqueness part yet. It is a simple consequence of
Riesz' representation theorem, but we state it here as a separate lemma for
future reference. The proof is left to the reader.
Lemma 4 Let (K, (., .)) be a space equipped with a non-degenerate hermitian
form, and let (., .)′ be a hermitian form such that for all ψ, φ ∈ K, and for all
A ∈ B(K), one has (Aψ, φ)′ = (ψ, A×φ)′, where A× is the adjoint of A for
(., .). Then there exists λ ∈ R such that (., .)′ = λ(., .).
Every minimal left ideal of Cl(V ) is an irreducible module for the representa-
tion of Cl(V ) by left multiplication. Such a module is of the form Se := Cl(V )e
where e is a primitive idempotent of Cl(V ), that is an idempotent which cannot
be decomposed as a sum of two others. The module Se is sometimes called
an algebraic spinor module, and its elements are algebraic spinors. It is then
natural to restrict (., .)σ to Se and ask if it defines a σ-compatible Krein prod-
uct. In fact σ-compatibility is immediate, but the restriction of (., .)σ can be
degenerate. Of course when Qσ is positive definite everything works fine : (., .)σ
is positive definite and its restriction to every algebraic spinor module is a σ-
compatible scalar product. We will use this fact in the following proposition,
which is implicit in [Rob 88].
Remark: The positive definite case is all we really need, but we can wonder what
happens in general. It turns out that we just need to compute ee×σ : either it is zero,
in which case (., .)σ vanishes on Se, or it is not, in which case (., .)σ restricts to a
σ-compatible Krein product on Se. Moreover, is this latter case, there exists a unique
primitive idempotent f ∈ Cl(V ) such that f ×σ = f and Se = Sf . The proof of this
claim is given in appendix A for the interested reader.
Proposition 8 (Robinson's alternative) Let K be an irreducible spinor module
and (., .)K be a σ-compatible Krein spinor product.
If Qσ is positive definite, then (., .)K is definite, if Qσ is not positive definite,
then (., .)K is neutral.
Proof: Let us call ρ : Cl(V ) → End(K) the representation homomorphism.
and use lemma 3.
If Qσ is not positive definite, let ei ∈ Vσ be such that e2
i = −1, set U = ρ(ei)
14
If Qσ is positive definite, then let U : K → Se, where Se is a minimal left
ideal in Cl(V ), be a Clifford-module isomorphism, which we know exists by
irreducibility of K. By Garling's alternative, (., .)σ is positive definite, hence
its restriction to Se also is. We now transport (., .)σ to K thanks to U by the
formula (ψ, η)U := (U ψ, U η)σ, for all ψ, η ∈ K. Let us check that (., .)U is
σ-compatible:
∀a ∈ Cl(V ), (ρ(a)ψ, η)U = (U ρ(a)ψ, U η)σ
= (LaU ψ, U η)σ by the intertwining property of U
= (U ψ, La×σ U η)σ
= (U ψ, U ρ(a×σ )η)σ
= (ψ, ρ(a×σ )η)U
definite, therefore (., .)K is by lemma 4.
Now (ψ, ψ)U = 0 ⇒ (U ψ, U ψ)σ = 0 ⇒ U ψ = 0 ⇒ ψ = 0, hence (., .)U is
¶
Of course multiplication by −1 turns σ-compatible Krein products into σ-
compatible Krein products. Putting together Robinson's alternative and propo-
sition 7, we can therefore characterize the Euclidean signature of Qσ in the
following way.
Corollary 1 If the form Qσ is positive definite then there exists a σ-compatible
scalar product on every irreducible spinor module. If there exists at least one
σ-compatible scalar product on at least one irreducible spinor module then Qσ
is positive definite.
We now turn our attention to the case where σ is admissible.
Lemma 5 Let ρ be a c-compatible representation of Cl(V ) on a Krein space
(K, (., .)). Let σ = Adb ◦ c be an admissible real structure. Let x ∈ Cl(V ) and let
(., .)x = (., ρ(x).). Then (., .)x is a σ-compatible Krein product on K iff x = x×
and x is proportional to b−1.
Proof: The two properties are easily seen to be sufficient. Let us prove that
they are necessary. First we note that since Cl(V ) is simple and ρ 6= 0, then ρ
is injective. For (., .)x to be sesquilinear x must satisfy ρ(x) = ρ(x)× which is
equivalent to x = x× since ρ is injective and c-compatible. Now we note that
a×σ = σ(aT ) = σ ◦ c ◦ c(aT ) = Adb(a×) = ba×b−1.
is a scalar by the injectivity of ρ.
We must have ρ(a×σ ) = ρ(a)×σ = ρ(x)−1ρ(a)×ρ(x) and we obtain that xb
¶
Now if b is real then b× = ±b according to proposition 1, and we can take
x = b−1 or x = ib−1 in the above lemma. Let us store the result we obtain as a
proposition.
Proposition 9 Let ρ be a c-compatible representation of Cl(V ) on a Krein
space (K, (., .)). Let σ = Adb ◦ c be an admissible real structure with b real.
1. If b× = b, (., .)b := (., ρ(b)−1.) is a σ-compatible Krein product on K.
2. If b× = −b, (., .)b := (., iρ(b)−1.) is a σ-compatible Krein product on K.
15
2.5 Characterization of the Lorentz and anti-Lorentz sig-
natures
First we fix some terminology. If Q has the Lorentz signature, the open light
cone C of V is the non-convex cone of those v such that Q(v) < 0. It consists
of two connected components which we arbitrarily call C+ and C−. If Q has
the anti-Lorentz signature, the open light cone C is defined by Q(v) > 0 and
C+/C− are defined accordingly.
Let L be a subspace in V and sL be the orthogonal symmetry with respect
to L. That is, if V = L ⊕ W is an orthogonal decomposition, then sL is the
identity on L and minus the identity on W .
Our starting point will be the following evident observation:
• Q has the Lorentz signature iff there exists a line L such that B(−sL(.), .)
has Euclidean signature.
• Q has the anti-Lorentz signature iff there exists a line L such that B(sL(.), .)
has Euclidean signature.
Note that in both cases the line will belong to the open light cone.
Now all we have to do is to translate this observation in terms of admissible
real structures, and use the characterization of Euclidean signature we arrived
at in the previous subsection. In order to do this we first review some basic facts
about the implementation of orthogonal symmetries with the adjoint action of
vectors in Clifford algebras. Special care is needed about the signs.
Let v be a non-zero vector in V . Then v is invertible and v−1 = 1
Q(v) v. The
adjoint action of v on a vector u ∈ V satisfies Adv(v) = vvv−1 = v and if w ⊥ v,
Adv(w) = vwv−1 = −wvv−1 = −w. Thus Adv = sL where L is the line Rv.
Now notice that for any u ∈ V , Adω(u) = ωuω−1 = −uωω−1 = −u since u
is odd. Thus Adωv = Adω ◦ Adv = −sL.
Moreover, since c(ω) = ω and c(v) = v, the real structures σv := Adv ◦ c and
σωv := Adωv ◦ c are admissible. We have thus obtained the following:
Lemma 6 Let L be a line in V . Then
• The unique admissible real structure which restricts to sL is σv := Adv ◦ c,
where v ∈ L is a non-zero vector.
• The unique admissible real structure which restricts to −sL is σωv :=
Adωv ◦ c, where v ∈ L is a non-zero vector.
Putting together what we have learned so far, we obtain:
Theorem 1 Let V be a finite dimensional real vector space of even dimension
with non-degenerate quadratic form Q. Let ρ be a c-compatible irreducible rep-
resentation on the Krein space (K, (., .)). Then:
1. The signature of Q is anti-Lorentzian iff there exists v ∈ V such that
(., .)v := (., ρ(v)−1.) is definite.
16
2. If n = 2 or n = 6 modulo 8, then the signature of Q is Lorentzian iff there
exists v ∈ V such that (., .)v := (., ρ(ωv)−1.) is definite.
3. If n = 0 or n = 4 modulo 8, then the signature is Lorentzian iff there
exists v ∈ V such that (., .)v := (., iρ(ωv)−1.) is definite.
In every case, (., .)v is definite ⇔ v ∈ C, and there exists λ = ±1 such that
λ(., .)v positive definite ⇔ v ∈ C+ and λ(., .)v negative definite ⇔ v ∈ C−.
Proof:
1. We know that Q has anti-Lorentz signature iff there exists a line
L such that B(sL(.), .) is positive definite. Observe that B(sL(.), .) = Qσ,
where σ = Adv ◦ c, for any non-zero v ∈ L. We also know that (., .)v is a
σ-compatible Krein product on K by proposition 9, and we conclude by
Robinson's alternative.
2. Case 2 and 3 are similar to case 1 with the extra complication that b = ωv
satisfies b× = b for n = 2, 6 mod 8 and b× = −b for n = 0, 4 mod 8.
It is obvious that a line L is such that B(sL(.), .) is positive definite iff
L \ {0} lies inside the light cone. Moreover multiplication by −1 exchanges C+
with C− and (., .)v with −(., .)v, so that it only remains to show that (ψ, ψ)v
keeps a constant sign for a fixed non-zero ψ and v varying continuously in one
component of C, which is immediate by continuity of ρ.
¶
Cases 2 and 3 can be put together by saying that the signature is Lorentzian
iff there exists v ∈ V such that (., .)v := (χρ(v)., .) is definite, where χ is chirality
operator of noncommutative geometry.
Remark: Even if we could simplify the above theorem by taking ρ(v) instead
of ρ(v)−1 and so on, since they differ only by a non-zero constant, it could prove
misleading in the noncommutative situation. We will come back to this issue later.
We will not delve much into such matters, but the above characterization
of the Lorentzian/antilorentzian signature can be generalized easily. If Q is of
signature (p, q) and p is odd, then we can find p vectors v1, . . . , vp such that
σ = Adg ◦ c, with g = v1 . . . vq, is an Euclidean real structure. The converse is
also true, and these vectors will then automatically satisfy v2
i > 0. If p is even
we have to take g = v1 . . . vq and we will have v2
i < 0. The problem now is
to characterize the elements of the Clifford group which are of this form. For
this we can use the canonical isomorphism Θ : ΛV → Cl(V ). Thus we need to
consider two different classes of quadratic forms: even and odd ones, meaning
that p and q are both even or both odd, according to the case. We can then say
that:
• If Q is odd, then p is the only integer such that ∃g ∈ Θ(ΛpV ), Adg ◦ c is
an Euclidean real structure.
• If Q is even then q is the only integer such that ∃g ∈ Θ(ΛqV ), Adg ◦ c is
an Euclidean real structure.
17
We could then translate these properties in terms of the Krein structures on
spinor modules. However, in order to generalize to noncommutative geometry,
we would have to use noncommutative p-forms or q-forms instead of elements
of the exterior algebra over V , and this raises the issue of junk forms when p
(resp. q) is > 1. We leave this question open for the moment, but we note that
in the Lorentz case we have used the trick of composing with the chirality in
order to stick with 1-forms. It remains to see, using a more general approach,
if this trick is legal. This is the main reason, besides simplicity, why we will
mostly deal with the antilorentzian signature in this paper.
2.6 Some Krein positive operators on spinor modules
In this subsection the quadratic form Q is supposed to be antilorentzian. We
consider a c-compatible irreducible representation ρ of Cl(V ) into a Krein space
(K, (., .)), which we decompose as K = KL⊕KR where KL (resp. KR) is the −1-
eigenspace (resp. +1-eigenspace) of the chirality operator χ. It is immediately
seen by computing (ψ, χ2φ) and using χ× = −χ that KL and KR are self-
orthogonal for the Krein product (., .).
An operator A on K will be said to be Krein-positive iff (ψ, Aψ) > 0 for all
non-zero ψ ∈ K. We gather here some observations which will be useful later.
For brevity, we do not write explicitly the representation ρ.
We already know that if e is a timelike vector, (., .)e is definite, and we will
suppose here that the definition of the future light-cone and the Krein product
are compatible in the sense that e is future-directed iff it is Krein-positive. The
next lemma indicates what happens if we take a spacelike vector instead.
Lemma 7 Let w be a spacelike non-zero element of V and let (ψ, ψ)w :=
(ψ, wψ). Then KL and KR are orhogonal to each other for (., .)w. If n > 2, the
form (., .)w is neutral on each half-spinor module KL and KR.
Proof: First, since multiplication by w changes chirality, it is obvious that KL
and KR are orthogonal to each other for (., .)w since they are self-orthogonal
If n > 2 we can find a spacelike vector e such that e2 = −1 and
for (., .).
B(e, w) = 0. Let Ue : KL → KR be the map ψ 7→ eψ. Since (eψ, eψ)w =
(eψ, weψ) = −(eψ, ewψ) = (ψ, ψ)w, Ue is an isometry for (., .)w. Thus (., .)w,
which is globally neutral, must have the same signature on KR and KL, hence
the result.
¶
If n = 2 the form (., .)w has opposite signature on the 1-dimensional half-
spinor modules.
There are other Krein-positive operators besides future-directed vectors. An
important example is given in the following lemma.
Lemma 8 Let u, v ∈ V . If n > 2, u + χv is Krein positive iff u + v and u − v
are both timelike and future-directed.
18
Proof: First, using the fact that KL and KR are self-orthogonal we have
(ψ, (u + χv)ψ) = (ψL, (u − v)ψL) + (ψR, (u + v)ψR)
with ψ = ψL + ψR, ψL/R ∈ KL/R. If u± v are timelike and future-directed then
(., (u ± v).) are both scalar products, hence the two summands are positive.
Conversely, if u + v is timelike and past-directed, we can take ψL = 0 and
ψR 6= 0 to obtain a negative result. Similarly u − v cannot be past-directed. If
u + v is spacelike we can use lemma 7 to see that there exists a ψR such that
¶
(ψR, (u + v)ψR) < 0. A similar conclusion holds if u − v is spacelike.
If n > 2, a general ×-self-adjoint odd element of End(K) will have the form
u + χv + r, where u, v ∈ V and the remainder r is a linear combination of eI ,
I ⊂ {1; . . . ; n}, 1 < I < n− 1, with (ei)1≤i≤n is some pseudo-orthogonal basis.
Let us suppose that u + χv + r is Krein-positive. Let us observe that if a, b are
Krein-positive then τn(ab) > 0. Indeed, for all non-zero ψ ∈ K, (abψ, bψ) > 0
by the Krein-positivity of a and the invertibility of b. Hence ab is a positive
operator for the scalar product (., .)b, thus it must have a positive normalized
trace. Using this principle with a = u + χv + r and b a future-directed timelike
vector, we see that τn(ub + χvb + rb) > 0. However τn(χvb) = τn(rb) = 0
since χv and r are both orthogonal to b with respect to the c-product. Hence
we obtain that τn(ub) = B(u, b) > 0 for all future-directed timelike vectors b,
which proves that u is itself future-directed timelike. We state this result as a
lemma.
Lemma 9 If u + χv + r is Krein-positive, then u is timelike and future-directed.
As an example of such a Krein-positive operator with a non-zero r, one can
consider a pseudo-orthonormal basis (e1, . . . , e4) of R1,3 and let p = e1 + ite2e3
for t ∈] − 1; 1[.
2.7 Real structure and charge conjugation. KO-dimension
tables
In noncommutative geometry, the real structure is defined as an antilinear oper-
ator on the Krein space which is the abstract substitute of the space of sections
of the spinor bundle. When it is viewed in this way, it is sometimes called the
charge conjugation, and we will adopt this terminology in order to distinguish
it from its Clifford counterpart. To begin this section let us see how the two
notions relate locally.
Proposition 10 Let ρ : Cl(V ) → End(K) be an irreducible c-compatible rep-
resentation. Then there exists an antilinear operator C : K → K implementing
c, i.e. such that
ρ(c(a)) = Cρ(a)C−1
for all a ∈ Cl(V ). This operator can be chosen to satisfy C2 = ǫ and C×C = κ,
with ǫ and κ some signs, and in that case is unique up to multiplication by eiθ,
θ ∈ R.
19
We have C2
Let σ = Adb ◦ c be an admissible real structure, with b in the Clifford group
such that b2 = λ = ±1, b× = λ′b, λ′ = ±1, and c(b) = b. Let B = ρ(b). Then σ
is implemented on K by Cσ = BC = CB.
σ = λǫ, C×
σ Cσ = λλ′ κ, and if (., .)σ is a σ-compatible Krein product
on K, then (Cσ)×σ = λC×
σ and (Cσ)×σCσ = λ′C×C.
Proof: Fix any basis of K and denote by c.c. : ψ 7→ ¯ψ the complex conjugation
of coordinates with respect to this basis. We also denote by c.c. : A 7→ ¯A the
antilinear involution on operators induced by complex conjugation of coordi-
nates, that is : ¯Aψ = A ¯ψ. Let us write c = ρ ◦ c ◦ ρ−1. Since c is an antilinear
involution of End(K), c ◦ c.c. is a linear (involutive) automorphism and is thus
of the form AdC for some C ∈ End(K). We let C = C ◦ c.c. and we obtain that
ρ(c(a)) = Cρ(a)C−1.
The fact that c is an involution translates as C2 = ǫ, with ǫ a constant. Let
ψ ∈ K be such that Cψ 6= 0. Then C3ψ = ǫCψ = C(ǫψ), which shows that ǫ is
real. From the fact that ρ is c-compatible we obtain that ρ(c(a))× = ρ(c(a)×) =
ρ(c(a×)) and it boils down to CC× = κ, with κ a constant, which must be real.
Calculating C2(C×)2 we find that κ2 = ǫ2. Dividing C by pκ we can suppose
that κ = ±1 and ǫ = ±1. It is now easy to prove the uniqueness up to a phase.
σ Cσ =
C×B×BC = λλ′C×C = λλ′ κ.
σ (see
proposition 9), and if λ = −1, then (Cσ)×σ = (iB)−1C×
If λ = 1 then (Cσ)×σ = B−1C×
σ Cσ = λC×B×BC = λλ′C×B2C = λ′C×C. ¶
Hence we have (Cσ)×σCσ = λC×
If the representation is not irreducible, we can break it up into irreducible
parts and use the proposition on each one of them. The signs ǫ and κ can still be
defined globally since they do not depend on the representation, but the phase
θ can vary from block to block.
The second part is an obvious consequence of proposition 1. We have C×
σ B = C×
σ (iB) = −C×
σ .
Let us use proposition 10 to see how the KO table of signs change when we
perform a Wick rotation.
If we start from an antilorentzian quadratic form Q, a Wick rotation to
Euclidean signature corresponds to b = v with v2 = 1, v× = v, hence λ′ = λ = 1.
2 +1,
2 +1 in
In the Lorentzian case b = ωv with v2 = −1, hence b2 = −ω2v2 = (−1)
2 +1b. Hence one has λ = λ′ = (−1)
2 vω = (−1)
and b× = v×ω× = (−1)
this case.
n
n
n
n
We introduce the hopefully obvious notations CL, χL, CE, χE, CAL, χAL.
Then CE = CALρ(v) = CLρ(ωv). According to the above proposition we then
have:
when Wick rotating from antilorentzian to Euclidean signature, but
C2
E = C2
AL; C∗
ECE = C×
ALCAL
C2
E = (−1)
n
2 +1C2
L; C∗
ECE = (−1)
n
2 +1C×
L CL
when Wick rotating from Lorentzian to Euclidean signature. Here we have
used the notation ∗ instead of ×σ since we are dealing with a Hilbert adjoint.
20
When going from antilorentzian/Lorentzian to Euclidean signature, the vol-
ume element changes in the following way : ωAL/L = e1 . . . en 7→ ωE :=
uσ(ωAL/L) = iqωAL/L. But the Euclidean chirality element is χE = (−i)
2 ρ(ωE)
2 +qρ(ωAL/L). Hence the two expressions only differ by
whereas χAL/L = (−i)
a minus sign:
n
n
χE = −χL, χE = −χAL
Finally we note that in both cases CE has an additional −1 factor in its
Summarizing we have:
commutation relation with χE with respect to the one of CAL/L.
• Antilorentzian → Euclidean : ǫ 7→ ǫ, κ 7→ κ, ǫ′′ 7→ −ǫ′′
• Lorentzian → Euclidean : ǫ 7→ (−1)
These rules permit to fill the Lorentzian and antilorentzian KO table of signs
from the Euclidean one, using the fact that the metric dimension modulo 8 is
preserved by a Wick rotation (the KO dimension, of course, is not).
2 +1κ, ǫ′′ 7→ −ǫ′′
2 +1ǫ, κ 7→ (−1)
n
n
The real Clifford algebra Cl(V, Q) can be directly recovered as the algebra of
all a ∈ Cl(V ) such that [C, ρ(a)] = 0. In noncommutative geometry, one gener-
ally uses the operator J which selects Cl(V,−Q) (as originally in [Connes 95]).
To pass from one to the other is easy:
J = χC
We call C the ungraded charge conjugation and J the graded charge conju-
gation operator. In this work it is generally more convenient to use C, but we
stick to the traditional J in definitions and statements of theorem.
The signs ǫ, ǫ′, ǫ′′ which are given in the definition of spectral triples corre-
spond to the commutation rules of J with D and χ and to the square of J. In
the Lorentzian/antilorentzian context there is also a sign arising from J ×J.
When we pass from the C convention (signs with tildes) to the J convention
(signs without tildes) and vice versa we must use the following rules:
• ǫ′′ = ǫ′′ (J and C have the same commutation rules with χ).
• ǫ′ = −ǫ′ (J and D have a commutation sign opposite to the one of C and
D). In even dimension ǫ′ = 1 and ǫ′ = −1 always. Hence we do not let
these signs appear in the table.
• ǫ = ǫ′′ǫ (because J 2 = χCχC = ǫ′′C2).
• In the Lorentzian/antilorentzian case : κ = −κ (since J ×J = (χC)×χC =
−C×C), in the Euclidean case, κ = κ.
21
KO dim = n = p − q [8]
Jχ = ǫ′′χJ, Cχ = ǫ′′χC
J 2 = ǫ
C2 = ǫ
J ×J = κ
C×C = κ
0
1
1
1
1
1
2
-1
-1
1
1
1
4
-1
1
-1
1
1
6
1
-1
-1
1
1
Table 1: Euclidean table of KO signs
n [8]
KO dim = p − q [8]
ǫ
ǫ′′ = ǫ′′
ǫ = ǫ′′ǫ
κ
κ
0
2
-1
-1
1
-1
1
2
0
1
1
1
-1
1
4
6
1
-1
-1
-1
1
6
4
-1
1
-1
-1
1
Table 2: Antilorentzian table of KO signs.
n [8]
KO dim = p − q [8]
ǫ
ǫ′′ = ǫ′′
ǫ = ǫ′′ǫ
κ
κ
0
6
1
-1
-1
1
-1
2
0
1
1
1
-1
1
4
2
-1
-1
1
1
-1
6
4
-1
1
-1
-1
1
Table 3: Lorentzian table of KO signs.
22
2.8 Examples
In this section, which is optional, we derive the compatible Krein products on
spinors in some usual representations. We illustrate how the light cone appears
in each of the three different cases of theorem 1 in concrete computations with
In each case we take V = Rn,
gamma matrices. Let us fix the notations.
with a pseudo-orthonormal basis (eµ)µ=0,...,n−1. We have raised the indices to
be consistent with physicists conventions. We will have B(e0, e0) = −1 in the
Lorentz case and B(e0, e0) = 1 in the anti-Lorentz case, and e0 will belong to
C+ in both cases.
We will use a representation (for instance Dirac or Majorana4) in which the
gamma matrices γµ := ρ(eµ) satisfy (γµ)† = ±γµ where the sign is the same as
in (γµ)2 = ±1, and † is the adjoint relatively to the canonical scalar product on
K = CN , N = 2n/2.
We will consider a c-compatible Krein product on K denoted by (., .). Its
adjunction operation is denoted A 7→ A×. We denote by β the matrix of the
form (., .) in the canonical basis, that is, (., .) = h., β.i where h., .i is the canonical
scalar product. We know that (., .) is determined up to a constant, which we
reduce to a sign by requiring that β2 = 1.
The c-compatibility of ρ is equivalent to the requirement that (γµ)× = γµ
for all µ, which is in turn equivalent to β(γµ)†β−1 = γµ. We see then that :
• In the anti-Lorentz case, β is determined up to a sign by :
βγ0β−1 = γ0,
βγkβ−1 = −γk, k = 1, . . . , n − 1,
β† = β, β2 = 1 (7)
• In the Lorentz case, β is determined up to a sign by :
βγ0β−1 = −γ0,
βγkβ−1 = γk, k = 1, . . . , n − 1,
β† = β, β2 = 1 (8)
The obvious solution to (7) is β = ±γ0. The more or less obvious solution
to (8) is β = ±γ0ρ(ω) if n = 2, 6 [8] and β = ±iγ0ρ(ω) if n = 0, 4 [8].
In
other words in either case β = ±γ0χ where χ is the chirality operator (the "γ5
matrix" in this context).
Now the part about the light cone in theorem 1 tells us in the anti-Lorentz
case that v ∈ C if and only if βρ(v) is a definite hermitian matrix.
In the
Lorentz case it is βχρ(v) which must be definite, but since the two χ cancel, we
obtain in both cases that v ∈ C ⇔ γ0ρ(v) is a definite hermitian matrix.
3 Global constructions
Let (M, g) be a connected orientable semi-Riemannian manifold of dimension n
and signature (p, q). In this section the notion of spacelike (resp. timelike) will
be associated with the positive (resp. negative) index of inertia. Note that this
4In fact the Dirac representation is more natural, since the Dirac basis is at the same time
orthonormal for the scalar product h., .ie0 and pseudo-orthonormal for the Krein-product.
23
clashes with the relativistic convention in the antilorentzian case ! We see no
way of avoiding this problem. We will warn the reader when a confusion could
arise. We will revert to the usual convention in the following sections when we
are exclusively concerned with the antilorentzian/Lorentzian cases.
Given the Clifford bundle Cl(M, g) and its complexification Cl(M ) we now
seek to make the constructions of the preceding section global. We start with
real structures.
3.1 Global real structures
A (global) real structure σ on Cl(M ) is an involutive bundle map which is
antilinear and respect products over each fibre. Since Cl(M ) is given as the
complexification of Cl(M, g), we start with a given real structure c. We will
be interested in real structures σ commuting with c and such that the metric
g(σ(.), .) is Euclidean, i.e. Euclidean real structures. In view of proposition 6
such a real structure σ defines an orthogonal splitting T M = Es ⊕ Et of the
tangent bundle of M into spacelike and timelike subbundles, and the restriction
of σ to T M is the orthogonal symmetry with respect to Es. Conversely, given
such a splitting (which always exists, see [F-N 12], section 3), we can define the
spatial orthogonal symmetry globally, and extend it by the universal property
of the Clifford bundle. If we compose the result with c we obtain an Euclidean
real structure. Hence there is no obstruction to the existence of such objects.
However, σ is given at each point x ∈ M by σx = Adbx ◦ cx where bx in
an element of the Clifford group at x. Hence globally we will have σ = Adb ◦ c
where b is a section of the Clifford group such that Adb is smooth. But it is
clear that there will generally be an obstruction to the existence of a smooth b.
To get a feel of what this obstruction might be let us consider the antilorentzian
case. We know that σ defines an orthogonal splitting T M = Es ⊕ Et of the
tangent bundle, and in view of lemma 6, b itself is a section of Es. Asking this
section to be smooth is asking for the existence of a time-orientation5, that is a
nonvanishing timelike (in the sense of relativity) vector field ([Beem 81], p 5).
To deal with the general case we need a formal definition of space and time
orientations in the semi-Riemannian case6. Recall that the kernel of a p-form
ωx is the subspace of TxM consisting of those vectors v such that ω(v, ., . . . , .)
is the null p − 1-form.
Lemma 10 The following claims are equivalent.
1. There exists a p-form ω such that Ker(ωx) is a timelike q-dimensional
subspace of TxM for all x ∈ M .
2. There exists a p-form ω such that for every linearly independent family
(v1, . . . , vp) of spacelike tangent vectors at x, ωx(v1, . . . , vp) 6= 0.
5Reader beware ! What we have called spacelike in this semi-Riemannian context is what
is called timelike in Relativity when the mostly plus convention is adopted.
6Surprisingly we could not locate any in the literature.
24
3. Given any decomposition T M = Es ⊕ Et of the tangent bundle into the
sum of a spacelike and timelike subbundle, there exists a non-vanishing
top form ωs of Es.
Proof: To see that (1) entails (2) consider any family F = (v1, . . . , vp) of space-
like tangent vectors at x and let Sx ⊂ TxM be the linear subspace they span.
Since Kerωx is timelike and q-dimensional, one has Kerωx ⊕ Sx = TxM . Call
πx the projection onto Sx defined by this decomposition. It is immediate that
ωx(u1, . . . , up) = λ detF (πx(u1), . . . , πx(up)) for any vectors u1, . . . , up, where
detF is the determinant in the basis F of Sx and λ ∈ R. Clearly λ 6= 0 since ωx
does not vanish identically. Hence ωx(v1, . . . , vp) 6= 0.
The proof that (2) entails (3) is immediate. To obtain (1) from (3), consider
the projection π on Es defined by T M = Es⊕ Et, and extend ωs by the formula
ωx(u1, . . . , up) := ωs
x(π(u1), . . . , π(up))
Then ω is a p-form on T M and its kernel at x is clearly (Et)x.
¶
A similar lemma clearly holds with timelike/spacelike reversed. When the
properties stated in the lemma hold, we will say (M, g) is space (respectively)
time orientable. A form on T M with the two first properties will be called a
space (resp. time) orientation.
If ω′ is a space orientation p-form on M , a linearly independent family
(v1, . . . , vp) of spacelike vectors at x will be said to be positively oriented iff
ω′
x(v1, . . . , vp) > 0. A similar definition can be given with time-orientations.
Note that space/time orientation forms actually provide more information than
the notion of orientation on families of spacelike/timelike vectors just defined.
For example, suppose a Lorentz manifold is time-orientable and consider a 1-
form ω′′ with spacelike kernel. Then a timelike vector v at x will be said
to be positively oriented, or future-directed iff ω′′
x(v) > 0. Using the musical
isomorphism ♯ provided by g one can define a vector field ξ = (ω′′)♯, which is
timelike, since it is orthogonal to the distribution of spacelike subspaces Ker(ω′′),
and non-vanishing. This is indeed the usual definition of time-orientation for
Lorentz manifolds, even if the notion of future/past directed timelike vectors
would just require the continuous choice of a half light-cone at each point of the
manifold.
In the same vein we say that M is (totally) orientable if there exists a n-form
ω such that ωx has zero kernel (that is, ω is non-vanishing) for all x.
Now the canonical isomorphism Θx : ΛTxM ≃ Cl(TxM, gx) gives rises to a
bundle isomorphism (see for instance [L-M 89], prop. 3.5, chap. 2). We can
compose it with the musical isomorphism and obtain a canonical isomorphism
of vector bundles Θ : ΛT ∗M ≃ Cl(M, g).
If ω is a space orientation form,
then at each x we can decompose TxM into an orthogonal sum Ker(ωx) ⊕ Sx,
If B = (v1, . . . , vp) is an orthonormal basis of Sx, we
where Sx is spacelike.
obtain immediately that ωx(u1, . . . , up) = λx detB(π(u1), . . . , π(up)) with π the
orthogonal projection on Sx. Hence bx := Θx(ωx) = λxv1 . . . vp is seen to be an
element of the Clifford group at x. Thus b = θ(ω) is a smooth section of the
25
Clifford group bundle. Conversely if b is smooth section of the Clifford group
bundle which is locally of the form v1 . . . vp, with vi spacelike, we easily see that
θ−1(b) is a spatial orientation. Similar consideration apply to time orientations.
We can use these observations to globalize the discussion at the end of subsec-
tion 2.5. Moreover since we have assumed M to be orientable, space orientations
can be converted into time orientations, saving us the trouble distinguishing the
even and odd cases, and of translating the words timelike/spacelike when we
revert to the traditional convention. We can thus answer our question about
the obstruction to the definition of real structure through smooth sections of
the Clifford group bundle in the following way:
• An orientable semi-Riemannian manifold admits Euclidean real structures
defined through smooth sections of the Clifford group bundle iff it is time
and space orientable.
Now recall that a time-oriented Lorentzian/antilorentzian manifold is called
a spacetime. Hence an oriented Lorentzian/antilorentzian manifold admits Eu-
clidean real structures defined through smooth sections of the Clifford bundle
iff it is a spacetime.
3.2 Hermitian forms on the spinor bundle
In this subsection we consider a spin-c manifold M with metric g of signature
(p, q) and spinor bundle S. We call ρ the representation ρ : Cl(M ) → End(S).
We will often write a.Ψ instead of ρ(a)Ψ in order to simplify notations. Let
H : x 7→ Hx be a smooth field of nowhere degenerate hermitian forms on S.
We call this object a spinor metric. For any A ∈ End(S) we denote by A×
the map x 7→ A(x)×, where A(x)× is the adjoint relatively to H(x). As in the
local case, we say that H is c-compatible if ρ(a)× = ρ(a×) for all sections a of
Cl(M ). This is equivalent to ask that ρ(ξ) be self-adjoint for all vector fields
ξ ∈ Γ(T M ). What we want now is to globalize Robinson's transfert principle.
The following theorem extends results in [Baum 80].
Theorem 2 There exists a c-compatible spinor metric iff (M, g) is time ori-
entable when p, q are even, and iff (M, g) is space orientable when p, q are odd.
Proof: We deal with the odd case only, the even case being completely similar.
Let us suppose that there exists a c-compatible spinor metric H. Consider
a covering of M by open sets (Uα)α∈A such that T M and S are trivial over Uα,
and a subordinate partition of unity (fα)α∈A. For each x ∈ M we let Ix be the
finite set of indices α such that fα(x) 6= 0. On each Uα let us choose a section ψα
of the spinor bundle which is constant in some trivialization SUα ≃ Uα × C2n/2
and non-vanishing.
Let us define the p-form
ω(X1, . . . , Xp) = Xα
fα(x)Hx(ψα, irΘ(X1 ∧ . . . ∧ Xs)ψα)
(9)
26
The integer r is [ p
2 ]. It ensures that ω is a real p-form. Indeed:
Θ(X1 ∧ . . . ∧ Xp)× =
ǫ(σ)Xσ(p) . . . Xσ(1)
1
p! Xσ
(−1)r
p! Xτ
ǫ(τ )Xτ (1) . . . Xτ (p)
=
= (−1)rΘ(X1 ∧ . . . ∧ Xp)
where (−1)r is the signature of the reversal permutation (1, . . . , p) 7→ (p, . . . , 1).
Let (e1, . . . , ep, ep+1, . . . , en) be a pseudo-orthonormal basis of TxM such that
e1, . . . , ep are spacelike. Then we know that Hx(., ir(e1 . . . ep)−1.) is definite.
Since (e1 . . . ep)−1 = ep . . . e1 = (−1)re1 . . . ep, we have either:
or
∀α ∈ Ix, Hx(ψα, ire1 . . . epψα) > 0
∀α ∈ Ix, Hx(ψα, ire1 . . . epψα) < 0
In both cases we have ωx(e1, . . . , ep) 6= 0. Now, since ω is multilinear alter-
nate, we have for any vectors u1 . . . , up ∈ Span(e1, . . . , ep):
ωx(u1, . . . , up) = det(uj
i )1≤i,j≤pωx(e1, . . . , ep)
(10)
where uj
i is the j-th component of ui in the basis (e1, . . . , ep). Since (e1, . . . , ep) is
any orthonormal family of spatial vectors, this shows that ωx(u1, . . . , up) 6= 0 for
any linearly independent family of spatial vectors. Thus ω is a space orientation.
Conversely, since M is space-orientable, there exists a Euclidean real struc-
ture σ = Adb◦c where b is a smooth section of the Clifford group bundle. On the
Riemannian manifold (M, gσ) it is well-known that there exists a σ-compatible
spinor metric. Let us call H such a spinor metric. Then H(., .) = H(., irb.)
¶
defines a c-compatible spinor metric for M .
Of course a c-compatible spinor metric is unique only up to multiplication
by a nonvanishing real function.
In view of the above theorem we come to the conclusion that the framework
we will have a chance to generalize to the noncommutative world is that of
space and time-orientable spin manifolds. We think it noteworthy that these
purely mathematical considerations have led us to manifolds which are "physics
friendly" in the sense that 1) we can put matter fields on them, 2) we can define
an arrow of time, 3) particles will not change their chirality if transported around
the universe.
3.3 The Dirac operator and c-compatibility
In the case of a manifold we have seen that we can express neatly the c-
compatibility of a spinor metric by saying that vector fields are self-adjoint.
27
Unfortunately this cannot be directly generalized in the noncommutative set-
ting. It is therefore important to have an alternative formulation which is better
suited to this generalization. This formulation is quite simple: the Dirac oper-
ator has to be self-adjoint. We could close this section there, but we prefer to
develop a little bit in order to be completely precise and also because it will be
useful later on.
We will need to consider a spin manifold. Our definition of a spin structure
on a spin-c manifold (M, g, S) is the algebraic one, which is directly applicable in
noncommutative geometry: it is antilinear bundle map J : S → S which satisfies
Jρ(a)J −1 = −ρ(c(a)) and J 2 = ±IdS. If S is equipped with a c-compatible
spinor metric H, we will also have JJ × = ±IdS (see proposition 10). We refer
to [Sch 00] for the equivalence of this definition and the traditional one involving
the possibility of lifting the frame bundle to a spin bundle.
By progressively enriching the structure we can define various kinds of con-
nections on S. We will use the notation
(Ψ, Ψ′)H = ZM
Hx(Ψ(x), Ψ′(x))p det(g)dx
(11)
We call this the Krein product on spinor fields, a name which will be justified
later on.
Definition 1 Let (M, g, S) be a spin-c manifold and H be a spinor metric on
S. Let ∇ be a connection on S .
1. If ∇X (a · Ψ) = (∇LC
X a) · Ψ + a · ∇X Ψ, for all sections a of the Clifford
bundle, vector fields X and spinor fields Ψ, Ψ′ with compact support, then
∇ is said to be a Clifford connection.
2. If X · (Ψ, Ψ′)H = (∇X Ψ, Ψ′)H + (Ψ,∇X Ψ′)H , for all X, Ψ, Ψ′ as above,
∇ is said to be metric.
3. If in addition M is a spin manifold and J is a spin structure on it, then
∇ is said to preserves spin if ∇X J = J∇X for all vector field X.
4. If M is a spin manifold and is space and time oriented, J is a spin struc-
ture on it, and H is a c-compatible metric on S, then ∇ is said to be a
spin connection if it is a Clifford connection which is metric and preserves
spin.
Let x, y ∈ M , let λ be a curve joining these points. Let hλ : Sx → Sy
be the parallel transport operator of ∇ along λ. Similarly, we denote by hLC
the parallel transport of the Levi-Civita connection along λ. Since this is an
isometry from (TxM, gx) onto (TyM, gy) we can consider its canonical extension
hLC
λ which is an isomorphism between Cl(TxM ) and Cl(TyM ). The above
infinitesimal definitions can be given integrated forms. More precisely, with the
same hypotheses as in definition 1, we can say that
λ
28
1. ∇ is a Clifford connection iff for all x, y, λ, hλ intertwines the action of the
Clifford algebras at x and at y on the spinors. More precisely, this means
that for all a ∈ Cl(TxM ) the following diagram commutes
Sx
a
Sx
hλ
Sy
hλ
hLC
λ (a)
/ Sy
2. ∇ is metric iff the parallel transport hλ is an isometry from (Sx, Hx) onto
(Sy, Hy).
3. ∇ preserves spin iff the following diagram commutes for all x, y, λ:
Sx
hλ
Sy
Jx
Jy
Sx
hλ
/ Sy
Going back and forth from the infinitesimal to the integrated properies is
not difficult using the formula
(∇X Ψ)(x) = lim
t→0
h−1
λ (Ψ(λ(t)) − Ψ(x)
t
where λ : [0; 1] → M is a curve such that λ(0) = x and ( d
dt λ)(0) = X, and
the similar formula for ∇LC. The integrated versions continue to make sense in
the discrete context, as we will see.
Remark: It is immediate that the holonomy operators of a Clifford (resp. spin)
connection belong to the Clifford (resp. spin) group. A Clifford connection which
is metric for a c-compatible spinor metric would have its holonomies in the spin-c
group and would deserve to be called a spin-c connection. Such objects play a role
in the theory of Seiberg-Witten invariants and are indeed called that way. Clifford
connections were defined in [B-G-V 92].
Two connections on S differ by an End(S)-valued 1-form, and it is easy to
see that this 1-form must be scalar-valued in the case of Clifford connections. If
the connections are also metric, the 1-form has values in iR, and if they commute
with J it has values in R, so we obtain the uniqueness of the spin connection.
Moreover the spin connection always exists (see theorem 9.8 in [GB-V-F 01] for
the Riemannian case, [Bizi XX] for the general case).
Every Clifford connection gives rise to a Dirac operator which will have the
local form D = −iPµ γ(dxµ)∇∂µ , where γ : T ∗M → End(S) is the composi-
tion of the musical isomorphism defined by g, and the representation ρ (hence
γ(dxµ)ψ = ρ((dxµ)♯)ψ).
29
/
/
/
/
/
/
In noncommutative geometry we will be left with (an abstract version of) the
sections of the spinor bundle, Krein product, Dirac operator and spin structure,
with no direct hold on the connection, Clifford algebra and vector fields. The
following facts are then crucial:
1. The the Dirac operator commutes with the spin structure iff the latter
commutes with the Clifford connection and anti-commutes with vector
fields.
2. The Dirac operator is essentially self-adjoint with respect to the Krein
product (., .)H iff the spinor metric H is c-compatible and the Clifford
connection is metric with respect to it.
We prove these claims in appendix C.
We can conclude from this that the Dirac operator of a Clifford connection
∇ is essentially self-adjoint and commutes with J iff ∇ is the spin connection.
We will recover a discrete analog of this result in section 3.4 of [Bes 16].
3.4 The canonical spectral "triple" of a semi-Riemannian
manifold
In this section we consider a space and time orientable spin manifold M , with
a given spinor bundle S equipped with c-compatible spinor metric H and rep-
resentation ρ, charge conjugation J, and the canonical Dirac operator D corre-
sponding to the spin connection ∇.
3.4.1 The Krein space of spinor fields
We let Γ∞
c (S) denotes the space of smooth sections of S with compact support.
On this space we have already defined the hermitian form (., .)H . Since (M, g)
is space orientable we can consider a space-orientation7 form β, and the cor-
responding smooth field of Clifford group elements b = θ(β).
In fact we can
suppose without loss of generality that bx lies in the Pin group for all x, since
here c(bx) = bx and we have b2
x ∈ R. Hence we can replace b with b/pb2 (see
proposition 1 for details). We set B = ρ ◦ b, and we recall that we can define a
definite spinor metric (., .)b by premultiplying spinors by B−1 or iB−1 according
to the negativity index of the metric g (see proposition 9). For instance in case
B = B×
(Ψ, Φ)b = ZM
Hx(Ψ(x), B−1
x Φ(x))√−gdx
(12)
is a definite hermitian form on Γ∞
c (S). Since M is connected we can suppose
that it is positive definite up to an overall change of sign when choosing β. We
call K the Hilbert space obtained by completing Γ∞
c (S) with respect to this
scalar product. It turns out that this completion does not depend on the choice
7Once again, in the antilorentzian convention this is a time-orientation !
30
x = 1 if b×
x = bx and b2
x = −1 if b×
of β (see [Baum 80] section 3.3.1 or [Bizi XX]). Moreover, as we have already
noticed at the end of section 2.3, since Adb ◦ c is an Euclidean real structure,
we have b2
x = −bx. In either case B2 = 1. We
then obtain canonically a Krein space K with fundamental symmetry B.
The canonical spectral "triple" of a Riemannian spin manifold of even di-
mension is the following bunch of objects : the pre-C∗-algebra A = C∞(M ),
Hilbert space of L2-spinor fields H, representation π of A on H by pointwise
multiplication, canonical Dirac operator D, charge conjugation operator J and
chirality operator χ. We see that in the semi-Riemannian space and time ori-
entable context we must replace H with the Krein space K described above, but
without fixing a particular fundamental symmetry: doing so would be equiva-
lent to fix a particular space orientation. We will see later that the C∗-structure
on A has to be dropped too. But for the moment we turn to the question of the
characterization of the signature in the Lorentzian and antilorentzian cases.
3.4.2 The antilorentzian and Lorentzian cases
Given what we have done just above, characterizing the Lorentzian/antilorentzian
signature of a manifold metric purely in terms of the data available in non-
commutative geometry is easy. Beware that we now, and for the rest of the
paper, revert to the usual convetion of calling "timelike" the vectors such that
g(v, v) > 0 in the antilorentzian case. Recall that if ω is a 1-form we denote by
γ(ω) the Clifford multiplication by ω, that is γ(ω) = ρ(γ♯).
Theorem 3 Let (M, g) be a semi-Riemannian space and time orientable spin
manifold of even dimension, with given c-compatible spinor metric H, spin
structure J and chirality operator χ. Let K be the Krein space of spinor fields
equipped with the Krein product (., .)H . Then (M, g) is
1. antilorentzian iff there exists a never vanishing 1-form β such that JβJ −1 =
−β and (., γ(β)−1.)H is positive definite.
2. Lorentzian iff there exists a never vanishing 1-form β such that JβJ −1 =
−β and (., γ(β)−1χ.)H is positive definite.
3.4.3 Causality conditions
A very specific feature of the Lorentzian/antilorentzian signature is the existence
of a local causal structure. Two approaches which aim at using precisely this
feature in noncommutative geometry are reviewed in [Bes 15b]. Though we will
not follow this strategy, it is natural to interpret one of the many global causality
conditions one can put on a spacetime in the terms of theorem 3, namely stable
causality.
Stable causality is the fourth stronger causality condition distinguished in
[M-S 08], with global hyperbolicity on top.
Intuitively a spacetime (M, g) is
stably causal if (M, gǫ) is causal for every small enough perturbation gǫ of the
31
metric. For the rigorous definition see [M-S 08]. Any simply connected 2-
dimensional spacetime satisfies this property. The following theorem gives a
caracterization of stable causality in terms of some classes of functions.
Theorem 4 ([S´an 05]) Let (M, g) be a spacetime. Then the following are equiv-
alent
1. There exists a time function.
2. There exists a temporal function.
3. (M, g) is stably causal.
Recall that a time function is a continuous strictly increasing function, and
a temporal function is a smooth function on a spacetime with a past-directed
timelike gradient ([M-S 08], def 3.48).
Remark: The consequences of this theorem for both noncommutative geometry
and quantum gravity might prove far-reaching. First, since causality is defined through
(noncommutative) time functions in noncommutative geometry ([Bes 15b]), the equiv-
alence of the first two points is important since it shows that one looses nothing when
using the pre-C ∗-algebra C∞(M ) instead of its C ∗-completion C(M ). Moreover, the
resilience of a stably causal metric under small fluctuations could mean that this con-
dition alone stands a chance of defining a notion a causality in the quantum regime.
Putting together theorems 3 and 4, one sees immediately that stably causal
spacetimes are those on which the 1-form β can be chosen to be exact.
3.5 Wick rotation of the gamma matrices and Dirac op-
erator
We take the same hypotheses and notations as in the previous subsection, ex-
cept that it will be more convenient here to work with the ungraded charge
conjugation operator C. Also it will be important to consider a normalized
space-orientation form β, so that BB× = 1 and B2 = ±1.
We know then that σ = Adb ◦ c is a Euclidean real structure on the Clifford
bundle Cl(M ), and we have seen how σ can induce a Wick rotation of the
metric. Namely, we have a new metric gσ(., .) = g(σ(.), .), and an isometry
uσ = 1+i
2 σ from (T M, gσ), to the subbundle Vσ of Cl(M ) equipped
with the quadratic form Q(w) = w2. This isometry extends to an isomorphism
uσ : Cl(M, gσ) → Clσ = {(x, a) ∈ Cl(M )σ(a) = a} ⊂ Cl(M ). What we would
like to do is to use this isometry to "Wick rotate" the Dirac operator directly,
short-circuiting the metric, since we do not have a direct access to the metric
in noncommutative geometry. The strategy, which, as far as we know, was first
used in section 3.4 of [Dun 15], is to Wick rotate the gamma matrices appearing
in the Dirac operator.
2 Id + 1−i
We note first that ρ ◦ uσ is a representation of Cl(M, gσ). Moreover Cσ :=
BC = CB is an ungraded charge conjugation operator corresponding to σ (see
proposition 10).
32
Let us use σ to decompose the tangent bundle T M into a g-orthogonal sum
T M+⊕ T M−, corresponding at each point to the decomposition given in lemma
1 (hence v is a section of T M+ iff σ(v) = v). Then if (eµ) is a local g-pseudo-
orthonormal frame such that eµ(x) lies either in TxM+ or in TxM−, then (eµ)
will also be a gσ-pseudo-orthonormal frame.
We can then define gamma matrices with respect to ρ and ρ ◦ uσ :
• γµ := ρ(eµ), which satisfy {γµ, γν} = 2g(eµ, eν),
• γσ
µ := ρ(uσ(eµ)) which satisfy {γσ
Since ρ(σ(a)) = Cσρ(a)C−1
σ , we have
1 − i
2 CσγµC−1
ν } = 2gσ(eµ, eν)
1 − i
2
BγµB−1
γσ
µ =
µ , γσ
σ =
γµ +
γµ +
1 + i
1 + i
2
2
Let ×σ be the adjunction in End(S) for the σ-compatible Krein product
given by proposition 9. We use the same symbol to denote the adjunction in
End(K) with respect to (., .)b. Then for any operator A on S or End(K) we
have A×σ = BA×B−1. Hence we can also write
γσ
µ =
1 + i
2
γµ +
1 − i
2
γ×σ
µ
We must also consider the gamma matrices γµ = ρ((e∗
µ)♯) = g(eµ, eµ)γµ,
and similarly γµ
σ = gσ(eµ, eµ)γσ
µ . We have
Observe that gσ(eµ, eµ) = g(σ(eµ), eµ)) = κ(µ)g(eµ, eµ), where κ(µ) = ±1
µ , we have
1 + i
1 − i
2
γµ
2
γµ +
σ = gσ(eµ, eµ)(cid:0)
µ (cid:1)
according to whether eµ ∈ TxM ±. Since ρ(σ(eµ)) = γ×σ
γµ(cid:1)
σ = g(eµ, eµ)(cid:0)
1 − i
2
γ×σ
µ +
1 + i
γµ
2
γ×σ
hence
γµ
σ =
1 + i
2
(γµ)×σ +
1 − i
2
γµ
(13)
Now we want to replace the matrices γµ appearing in D by γµ
σ . In view of
(13), we define
Dσ =
1 + i
2
D×σ +
1 − i
2
D
which can also be written
Dσ =
1 + i
2
BDB−1 +
1 − i
2
D
and from CD = −DC and Cσ = BC we also obtain
33
(14)
(15)
Dσ = −
1 + i
2 CσDC−1
σ +
1 − i
2
D
It is clear on expression (14) that D×σ
σ = Dσ. From (16) we get
{Dσ,Cσ} = −
= −
= 0
1 + i
2 CσDC−1
1 + i
2 CσD +
σ Cσ +
1 − i
2
1 − i
2
DCσ −
DCσ − Cσ
1 − i
2 C2
1 + i
2 CσDC−1
σ + Cσ
1 + i
2 CσD
σ +
σDC−1
(16)
1 − i
2
D
(17)
(Hence [Dσ, Jσ] = 0.)
Note that in this computation we used the fact that B2 is constant.
Finally, if we compute the action of Dσ on a spinor field from (15), we obtain
(cid:18) 1 + i
2
BDB−1 +
1 − i
2
D(cid:19) Ψ =
1 + i
γµ∇eµ Ψ
2
eµ B−1)Ψ
1 + i
iBγµ∇eµ (B−1Ψ) + i
iBγµ(∇LC
1 + i
2
= iγµ
σ∇eµ Ψ +
= iγµ
σ ∂µΨ + . . .
2
(18)
In the last line of this expression we have stressed that even though the
Wick rotated Dirac operator Dσ is not, except for very special metrics, the
Dirac operator of the Wick rotated metric gσ, it has the same principal symbol.
In particular, the metric on M defined by Connes' distance formula from Dσ is
gσ.
The conclusion of this discussion is that Dσ has all the algebraic properties
required to form a spectral triple, together with the Hilbert space (K, (., .)b),
charge conjugation Jσ, chirality −χ, and the representation π of C∞(M ) by
(We have been very uncaring about the analytical
pointwise multiplication.
properties. For more on this question see [D-P-R 13]).
Hence if we set S = (C(M ), K, (., .), π, D, J, χ) and Sβ = (C(M ), K, (., .)b, π,
Dσ, Jσ,−χ) the map Wβ : S 7→ Sβ is defined without reference to the Clifford
In the Lorentzian and
bundle as long as we know how to let β act on K.
antilorentzian cases this action can indeed be defined thanks to π and D only.
The target of Wβ is a spectral triple, and the nature of the source is to be
defined in [Bes 16].
If we want to go in the other direction, we must provide a distinguished form
β (or equivalently b, c = Adb ◦ σ or J = B−1Jσ). From
u−1
σ =
1 − i
2
Id +
1 + i
2
c
we guess that
D =
1 − i
2
D×
σ +
1 + i
2
Dσ =
1 − i
2
B−1DσB +
1 + i
2
Dσ
(19)
34
This is is readily checked (with the first formula we need to use the fact
that (D×σ )× = (D×)×σ which is true because BB× is constant, and with the
second formula we need to use BDB−1 = B−1DB which is true because B2 is
constant).
Example : We start from M = R2k with the constant Euclidean metric. The
spinor bundle is trivial and identified with M × C2k
. We let eµ = ∂µ. To Wick
rotate to antilorentzian signature we take g = e0 (for instance). Then Vσ is the
vector subspace generated by e0 and iea, a = 1..2k, and G = γ0 is constant.
Then (14) gives
i(γ0∂0 −Xa
γa∂a) +
1 − i
2
i(γ0∂0 +Xa
γa∂a)
(iγa)∂a)
(γa∂a)), with γµ antilorentzian gamma matrices
Dσ =
1 + i
2
= i(γ0∂0 −Xa
= i(γ0∂0 −Xa
= iXa
= Dantilor
γa∂a, raising the indices with the antilorentzian metric
(20)
In this simple example Dσ is the canonical Dirac operator of gσ, since the
Christoffel symbols vanish and B is constant.
A Krein products on algebraic spinors
Proposition 11 Let Se be an algebraic spinor module.
1. If ee×σ = 0 then (., .)σ is zero on Se, hence is not a Krein product.
2. If ee×σ 6= 0 then :
(a) There exists a unique primitive idempotent f ∈
Cl(V ) such that f ×σ = f and Se = Sf .
(b) The
left
representation L of Cl(V )
on
(Se, (., .)σ) is σ-compatible.
on Se, which exists by proposition (7).
Proof: Suppose ee×σ = 0. Then for all a, b ∈ Cl(V ), (ae, be)σ = τn(e×σ a×σ be) =
τn(a×σ bee×σ) = 0.
Now suppose ee×σ 6= 0. Let (., .)a be a σ-compatible Krein spinor product
Since e is a primitive idempotent, p := Le ∈ End(Se) is a rank one projec-
tion. Note that p(e) = e2 = e, hence e ∈ Im(p), so that Im(p) = Ce. Observe
also that (Kerp)⊥ = Im(p×) (the orthogonal and adjoint are relative to (, )a) is
one-dimensional8. Let us call v a non-zero vector in this space.
8Proof of these relations : see appendix.
35
If v were isotropic, then for all ψ, η ∈ Se one would have (p×(ψ), p×(η))a =
(ψ, pp×(η))a = 0. By the non-degeneracy of (., .)a this would imply pp× = 0
which is excluded since ee×σ 6= 0.
Then we can suppose that (v, v) = ǫ = ±1. Let us call q ∈ Se the orthogonal
projection on Cv = (Kerp)⊥. Explicitly one has q(ψ) = ǫ(v, ψ)v, and it is easy
to check that q2 = q = q×. Moreover, Kerq = v⊥ = Kerp. Since v is not
isotropic, v /∈ Ker(q), hence Se = Ker(q) ⊕ Cv. If ψ ∈ Ker(q), then ψ ∈ Kerp,
hence pq(ψ) = p(ψ) = 0. Moreover pq(v) = p(v). Thus p = pq. Similarly it is
easy to check that qp = q using Se = Kerp ⊕ Ce.
e = ef , f = f e. This shows a). The uniqueness is easy to prove.
Let us call f = L−1(q) ∈ Cl(V ). Using L one sees that f 2 = f , f ×σ = f ,
Let us show b). Since we know that (aψ, φ)σ = (ψ, a×σ φ)σ for all a ∈ Cl(V )
and φ, ψ ∈ Se by (6), then by lemma 4 there exists λ ∈ R such that (., .)σ =
λ(., .)a on Se. There thus suffices to show that λ 6= 0.
Let us consider f as an element of Se. Then (f, f )σ = τn(f ×σ f ) = τn(f ).
Now f being a non zero projection, its normalized trace is not zero. On the
other hand we have (f, f )σ = λ(f, f )a. This shows that λ 6= 0, and that (., .)σ
¶
is σ-compatible.
Let Sf be an algebraic spinor module with f a primitive idempotent such
that f ×σ = f . Then for each x = af and y = bf one has x×σ y = f a×σ bf ∈
f Cl(V )f . But remember that in any finite-dimensional algebra A over k, if f is
a primitive idempotent then f Af is a division algebra over k. Hence we see that
the range of the map (x, y) 7→ x×σ y defined on Se × Se is a division algebra over
R, hence is isomorphic to R, C or H. Such a map is called9 a spinor product
in [Lou 81]. Thus we see that the σ-compatible Krein product on algebraic
spinor modules are just the composition of Lounesto's spinor product with the
normalized trace. It should certainly be possible to relate the properties of these
two kinds of products.
Remark: Given a σ-compatible representation ρ in a Krein space K, we see that
ρ(Se) is the set of operators a such that Im(a×) ⊂ L where L is a line. The set of
minimal left ideals Se has thus the structure of P (C2n/2
). Moreover, the ideals on
which (., .)σ vanishes are exactly those whose corresponding line is isotropic. They
form a submanifold of the projective space which is the image by the quotient map of
the isotropic cone of the Krein product.
B C∗-algebra structures on Cl(V )
Let σ be an Euclidean real structure. Let K be an irreducible spinor module
and h., .iσ a σ-compatible scalar product on K. Then let
kakρ
σ := sup
ψ6=0
haψ, aψiσ
hψ, ψiσ
9We thank Christian Brouder for drawing our attention to this reference.
36
σ) is a C∗-algebra isomorphic to M2n/2(C) with the
Obviously, (Cl(V ),×σ,k.kρ
usual involution ∗ and operator norm k.k associated with the canonical scalar
product. Hence if we equip Cl(V ) with σ we obtain an algebra with two com-
muting involutions, × and ×σ, one of them being a C∗-involution. This kind of
structure is called a Krein C∗-algebra in [Kawa 06] (see also [B-R 13]). Thus,
defining a Wick rotation of the metric to Euclidean signature exactly corre-
sponds to giving Cl(V ) a Krein C∗-algebra structure. Note that we used a
particular representation to make this remark (in example 4.10 of [Kawa 06],
where this point is also emphasized, the natural representation of the Clifford
algebra on the exterior algebra is used). However if we use the σ-product there
is no need to leave the Clifford algebra.
Indeed, notice that the σ-product (a, b)σ = τn(a×σ b) is transported by ρ to
the Hilbert-Schmidt scalar product on M2n/2(C). In particular one has kak2
σ =
Tr(ρ(a)∗ρ(a)). Now write kAkHS := [Tr(A∗A)]1/2 for the Hilbert-Schmidt norm
of a matrix A and define kAk∞,HS = sup{kABkHS kBkHS = 1}. Then it is
easy to see that kAk∞,HS is in fact equal to kAk. Hence the C∗-norm on
Cl(V ) corresponding to the involution ×σ can be defined without using the
representation ρ by
kak∞,σ := sup{kabkσ kbkσ = 1}
(21)
Conversely, if ∗ is a C∗-involution which commutes with × and stabilizes
V C then it is of the form ×σ for some Euclidean real structure σ. More-
over if σ′ is another Euclidean real structure, then (Cl(V ),×σ,k.k∞,σ) and
(Cl(V ),×σ′,k.k∞,σ′ ) are isomorphic. In fact this isomorphism is canonical and
is just Adbb′ −1.
C Proofs of the claims in subsection 3.3
.
∂µ
Let (M, g) be a semi-riemannian spin manifold, with spinor bundle S and spin
structure J. Let ∇ be a Clifford connection and D be the Dirac operator D
associated to it. We write ∇µ := ∇ ∂
Proposition 12 The Dirac operator and spin structure commute if and only if
{J, X} = [J,∇X ] = 0 for all X ∈ Γ(T M ).
Proof: The fact that {J, X} = [J,∇X ] = 0 are sufficient conditions is a
straightforward calculation.
Let us prove that they are necessary. Let f be a smooth real-valued function.
Since [J, D] = [J, f ] = 0, we deduce that J commutes with [D, f ] = −iγ(df ).
We infer easily that J anti-commutes with any real vector field or differential
form. Next we consider the differential operator ∇′ = J −1∇J. This operator is
a Clifford connection. Indeed, for any X, Y ∈ Γ(T M ), ψ ∈ Γ(S):
∇′
X (Y · ψ) = J −1∇X J(Y · ψ)
37
= −J −1∇X (Y · Jψ)
= −J −1(∇LC
= (∇LC
X Y ) · ψ + Y · ∇′
X Y ) · Jψ + Y · ∇X (Jψ)
X ψ
(22)
There thus exists a complex-valued one-form ω such that: ∇′
This can also be written: ω(X) = J −1[∇X , J]. We have:
X−∇X = ω(X).
0 = [J, D]
= [J,−iXµ
= −iXµ
= −iJ Xµ
γ(dxµ)∇µ]
γ(dxµ)[J,∇µ]
γ(dxµ)ωµ
¶
This implies that ω = 0, and thus that [∇X , J] = 0.
Let H be a spinor metric and (., .)H be the associated product on spinor
fields defined by (11).
Proposition 13 The Dirac operator is essentially self-adjoint if and only if
X × = X for all X ∈ Γ(T M ), and the Clifford connection ∇ is metric for H.
Remark: Our spinor bundle S thus has to be a Dirac bundle (see [L-M 89], p
114).
Proof: Let us assume that the Dirac operator is self-adjoint. Let f be a real-
valued smooth function. Then f is self-adjoint. This implies that [D, f ] =
−iγ(df ) is anti-self-adjoint. We easily conclude from this that all real vector
fields and all real-valued 1-forms are self-adjoint. In particular, the γ(dxµ) are
self-adjoint. Let ψ, φ be spinor fields with compact support. We have:
0 = (ψ, Dφ)H − (Dψ, φ)H
= Z pgXµ
0 = iZ pgXµ
[H(−iγ(dxµ)∇µψ, φ) − H(ψ,−iγ(dxµ)∇µφ)] dx
[H(∇µψ, γ(dxµ)φ) + H(ψ, γ(dxµ)∇µφ)] dx
We also have that:
Hence
[∇µ, γ(dxµ)] = γ(∇LC
= −Xα
µαγ(dxα)
Γµ
µ (dxµ)),
Xµ
[∇µ, γ(dxµ)] = −Xµ,α
Γµ
µαγ(dxα)
38
= −Xα
∂αpg
pg
γ(dxα)
Substituting in the integral above gives us:
from which we infer thatPµ γ(dxµ)∇µ = Pµ ∇µγ(dxµ)+Pµ γ(dxµ)(∂µpg)/pg.
Z dxXµ hpgH(∇µψ, γ(dxµ)φ) +pgH(ψ,∇µ(γ(dxµ)φ)) + (∂µpg)H(ψ, γ(dxµ)φ)i = 0
Finally, an integration by part yields:
[H(∇µψ, γ(dxµ)φ) + H(ψ,∇µ(γ(dxµ)φ)) − ∂µH(ψ, γ(dxµ)φ)] dx = 0
Z dxpgXµ
for all ψ, φ ∈ Γc(S). Now, the expression between brackets can be proven to be
C∞(M, C)-linear in φ (and anti-linear in ψ as well). Indeed, H(∇µψ, γ(dxµ)φ) is
clearly linear in φ. let f ∈ C∞(M, C). We replace φ by f φ in the two remaining
terms:
H(ψ,∇µ(γ(dxµ)f φ)) − ∂µH(ψ, γ(dxµ)f φ) = H(ψ,∇µ[f (γ(dxµ)φ)]) − ∂µ[f H(ψ, γ(dxµ)φ)]
= H(ψ, (∂µf )γ(dxµ)φ) + f∇µ(γ(dxµ)φ))
−(∂µf )H(ψ, γ(dxµ)φ) − f ∂µH(ψ, γ(dxµ)φ)
= f [H(ψ,∇µ(γ(dxµ)φ)) − ∂µH(ψ, γ(dxµ)φ)]
Thus, for all f ∈ C∞(M, C) and ψ, φ ∈ Γc(S):
Z dxpgf Xµ
[H(∇µψ, γ(dxµ)φ) + H(ψ,∇µ(γ(dxµ)φ)) − ∂µH(ψ, γ(dxµ)φ)] = 0
which implies that:
[H(∇µψ, γ(dxµ)φ) + H(ψ,∇µ(γ(dxµ)φ)) − ∂µH(ψ, γ(dxµ)φ)] = 0
Xµ
at every x. Now suppose that ∇0 is a local spin connection defined around
x. We know that ∇ = ∇0 + A where A is a scalar 1-form, and that ∇ is metric
if and only if A has pure imaginary values. Using (23) we find that
(23)
(Aµ + ¯Aµ)H(ψ, γ(dxµ)φ) = 0
(24)
Xµ
n
for all spinor fields ψ, φ with small enough compact support containing x.
Since H is non-degenerate the orthogonal of {γ(dx2)φ, . . . , γ(dxn)φ} for H is
a subbundle of dimension 2
2 − n + 1 ≥ 1 of the spinor bundle. Since the
orthogonal of {γ(dx1)φ, . . . , γ(dxn)φ} has codimension 1 in it, we can, at least
locally, consider a spinor field ψ such that H(ψ, γ(dxµ)φ) = 0 for µ = 2, . . . , n
and H(ψ, γ(dx1)φ) 6= 0. Hence A1 + ¯A1 = 0, and of course the same can be
done for the other indices. The Clifford connection ∇ is thus metric.
The converse can be proven easily following the same steps. It is in fact a
standard result of spin geometry. See for example [L-M 89].
¶
39
References
[Baum 80] H. Baum, Spin-Strukturen und Dirac-Operatoren uber pseudorie-
mannschen Mannigfaltigkeiten, Sekt. Mathematik d. Humboldt-Univ., 1980.
[Beem 81] J. Beem, P. Ehrlich, Global Lorentzian Geometry, Marcel Dekker:
New York, 1981
[B-G-V 92] N. Berline, E. Getzler, M. Vergne, Heat Kernels and Dirac Opera-
tors, Springer (1992)
[Bizi XX] N. Bizi, Ph. D. Thesis (in progress)
[B-R 13] P. Bertozzini, K. Rutamorn, Krein C∗-modules, Chamchuri Journal of
Mathematics (2013) 5:23-44, arXiv:1409.1343
[Bes 15a] F. Besnard, Noncommutative ordered spaces, examples and counterex-
amples, Class. Quantum Grav. 32 (2015) 135024 arXiv:1312.2442
[Bes 15b] F. Besnard, Two roads to noncommutative causality, J. Phys.: Conf.
Ser. 634 (2015)
[Bes 16] F. Besnard, On the definition of spacetimes in Noncommutative Ge-
ometry, part II
[Bog 74] J. Bogn´ar, Indefinite inner product spaces, Springer-Verlag, Berlin-
Heidelberg-New York, 1974
[Connes 94] A. Connes, Noncommutative Geometry, Academic Press, 1994
[Connes 95] A. Connes, Noncommutative Geometry and Reality, J. Math. Phys.
36, 6194 (1995)
[Cru 90] A. Crumeyrolle, Orthogonal and Symplectic Clifford Algebras, Spinor
Structures, Springer, 1990
[F-N 12] F. Finster, M. Nardmann, Some curvature problems in semi-
Riemannian geometry, in Global Differential Geometry, Spinger Proceed-
ings in Mathematics, C. Bar, J. Lohkamp, M. Schwarz Eds, vol 17 (2012)
arXiv:1010.2311
[Fra 11] N. Franco, Lorentzian approach to noncommutative geometry, PhD the-
sis (2011)
[GB-V-F 01] J.M. Gracia-Bond´ıa, J.C. V´arilly, H. Figueroa, Elements of Non-
commutative Geometry, Birkhauser, 2001
[Gar 11] D.J.H. Garling, Clifford Algebras, an introduction, London Mathemat-
ical Society Student Texts 78, Cambridge University Press (2011)
[Kawa 06] K. Kawamura, Algebra with indefinite involution and its representa-
tion in Krein space, arXiv:math/0610059
40
[L-M 89] H.B. Lawson, M-L Michelsohn, Spin Geometry, Princeton University
Press, 1989
[M-vS 14] M. Marcolli, W. D. van Suijlekom, Gauge Networks in noncommuta-
tive geometry, J. Geom. Phys. 75, 71–91 (2014), arXiv:1301.3480
[Lou 81] P. Lounesto, Scalar products on spinors and an extension of the
Brauer-Wall groups, Found. Phys. 11, 9/10 (1981)
[M-S 08] E. Minguzzi, M. S´anchez, The causal hierarchy of spacetimes, Recent
developments in pseudo-Riemannian geometry, ESI Lect. Math. Phys, 299–
358 (2008)
[Moo 01] J. D. Moore, Lecture Notes on Seiberg-Witten invariants, Lecture
Notes in Mathematics, 1629 (2nd ed.), Springer (2001)
[Rob 88] P.L. Robinson, Spinors and Canonical Hermitian Forms, Glasgow
Mathematical Journal, 30, 3, pp 263–270 (1988)
[S´an 05] M. S´anchez, Causal hierarchy of spacetimes, temporal functions and
smoothness of the Geroch's splitting. A revision. Contemporanea Matemat-
ica 29, 127–155 (2005)
[Sch 00] H. Schroeder, On the definition of geometric Dirac operators,
arXiv:math/0005239
[Dun 15] K. van den Dungen, Lorentzian geometry and physics in Kasparov's
theory, PhD Thesis, 2015
[D-P-R 13] K. van den Dungen, M. Paschke, A. Rennie, Pseudo-Riemannian
spectral triples and the harmonic oscillator, J. Geom. Phys. 73, 37–55 (2013)
arXiv:1207.2112
41
|
1712.09499 | 3 | 1712 | 2019-05-20T05:17:59 | The Picard groups for unital inclusions of unital $C^*$-algebras | [
"math.OA"
] | We shall introduce the notion of the Picard group for an inclusion of $C^*$-algebras. We shall also study its basic properties and the relation between the Picard group for an inclusion of $C^*$-algebras and the ordinary Picard group. Furthermore, we shall give some examples of the Picard groups for unital inclusions of unital $C^*$-algebras. | math.OA | math |
THE PICARD GROUPS FOR UNITAL INCLUSIONS
OF UNITAL C ∗-ALGEBRAS
KAZUNORI KODAKA
Abstract. We shall introduce the notion of the Picard group for
an inclusion of C ∗-algebras. We shall also study its basic properties
and the relation between the Picard group for an inclusion of C ∗-
algebras and the ordinary Picard group. Furthermore, we shall
give some examples of the Picard groups for unital inclusions of
unital C ∗-algebras.
1. Introduction
In the previous paper [13] we introduced the notion of the strong
Morita equivalence for inclusions of C ∗-algebras. Then in [13, Proposi-
tion 2.3], we showed that the strong Morita equivalence for inclusions
of C ∗-algebras is an equivalence relation. Thus in the same way as in
Brown, Green and Rieffel [2], we can define the Picard group for an
inclusion of C ∗-algebras. Also, in this paper we shall study its basic
properties as in [2] and [10]. In the last section, we shall give some ex-
amples of the Picard groups for unital inclusions of unital C ∗-algebras.
For a unital C ∗-algebra A, let Mn(A) be the n × n-matrix algebra
over A and In denotes the unit element in Mn(A). We identify Mn(A)
with A ⊗ Mn(C).
Let A and B be C ∗-algebras and X an A − B-bimodule. We denote
its left A-action and right B-action on X by a · x and x · b for any
by x ∈ X.
a ∈ A, b ∈ B, x ∈ X, respectively. Also, we denote by eX the dual
B − A-bimodule of X and we denote by ex the element in eX induced
For each C ∗-algebra A, let M(A) be its multiplier C ∗-algebra and
for any automorphism α of A, let α be the automorphism of M(A)
induced by α.
2. Preliminaries
In this section, we give several notations and basic facts used in the
paper.
Let α be an automorphism of a C ∗-algebra A. Following [2] but in a
slightly different way, we construct an A − A-equivalence bimodule Xα
2010 Mathematics Subject Classification. 46L05.
Key words and phrases. equivalence bimodules, inclusions of C ∗-algebras, the
Picard group.
1
induced by α as following: Let Xα = A as C-vector spaces and the left
A-action on Xα and the left A-valued inner product are defined in the
usual way, but we define the right A-action on Xα by x · a = xα(a) and
the right A-valued inner product by hx, yiA = α−1(x∗y) for any a ∈ A,
x, y ∈ Xα. We call Xα the A − A-equivalence bimodule induced by α.
Let A ⊂ C be a unital inclsion of unital C ∗-algebras and EA a
conditional expectation from C onto A.
Definition 2.1. ([17, Definition 1.2.2 and Lemma 2.1.6]) A finite set
{(ui, u∗
i=1 ⊂ C × C is called a quasi-basis for EA if
i )}n
nXi=1
uiEA(u∗
i c) = c =
nXi=1
EA(cui)u∗
i
for any c ∈ C. Also, we say that EA is of Watatani index-finite type if
there exists a quasi-basis for EA and in this case we define IndW (EA),
Watatani index of EA by
IndW (EA) =
nXi=1
uiu∗
i ∈ C.
By [17, Proposition 1.2.8], IndW (EA) is an invertible element in C ′ ∩
C and it does not depend on the choice of quasi-bases.
Proposition 2.1. ([17, Proposition 1.4.1]) With the above notation, we
suppose that EA is of Watatani index-finite type. If there is another
conditional expectation F A from C onto A, there is the unique element
h ∈ A′ ∩ C with EA(h) = 1 such that
for any c ∈ C.
F A(c) = EA(hc)
Under the same situation as above, we regard C as a right Hilbert
A-module as follows: We define the right A-action on C by x · a = xa
and the right A-valued inner product by hx, yiA = EA(x∗y) for any
a ∈ A, x, y ∈ C. Let BA(C) be the C ∗-algebra of all adjointable
right A-module operators on C and KA(C) the C ∗-algebra of all ad-
jointable right A-module "compact" operators on C. We regard EA as
an element in BA(C). We denote it by eA and we call eA the Jones
projection for EA. Also, we regard c ∈ C as an element in BA(C) by
the left multiplication in C.
Definition 2.2. ([17, Definition 2.1.2]) Let C1 be the closure of the
linear span of {ceAd ∈ BA(C) c, d ∈ C}. We call the C ∗-algebra C1
the C ∗-basic construction for EA.
By the definition of C1, C1 is strongly Morita equivalent to A with
respect to the equivalence bimodule C. By [17, Lemmas 2.1.3 and
2.1.6], C1 = KA(C) = BA(C) and C1 is the linear span of {ceAd ∈
BA(C) c, d ∈ C}.
2
Definition 2.3. ([17, Definition 2.3.2]) Let EC be the linear map from
C1 onto C defined by
EC(ceAd) = IndW (EA)−1cd
for any c, d ∈ C. Then EC is a conditional expectation from C1 onto
C and we call EC the dual conditional expectation of EA.
i )}n
1
2
By [17, Proposition 2.3.4], EC is of Watatani index-finite type. Let
i=1 be a quasi-basis for EA and we set wi = uieA(IndW (EA))
i=1 is a quasi-basis for
{(ui, u∗
for i = 1, 2, . . . , n. Then the finite set {(wi, w∗
EC.
i )}n
Definition 2.4. ([13, Definition 2.1]) Inclusions of C ∗-algebras A ⊂ C
and B ⊂ D with AC = C and BD = D are strongly Morita equivalent
if there are a C − D-equivalence bimodule Y and its closed subspace
X satisfying the following conditions:
(1) a · x ∈ X, Chx, yi ∈ A for any a ∈ A, x, y ∈ X and ChX, Xi = A,
ChY, Xi = C,
(2) x · b ∈ X, hx, yiB ∈ B for any b ∈ B, x, y ∈ X and hX, XiD = B,
hY, XiD = D.
Then we say that A ⊂ C is strongly Morita equivalent to B ⊂ D with
respect to the C − D-equivalence bimodule Y and its closed subspace
X. We note that X is regarded as an A − B-equivalence bimodule and
that if strong Morita equivalence inclusions A ⊂ C and B ⊂ D are
unital, we do not need to take the closure in Definition 2.4.
Let A ⊂ C and B ⊂ D be as above. Let EA and EB be conditional
expectations from C and D onto A and B, respectively. Let EX be a
linear map from Y onto X.
Definition 2.5. ([13, Definition 2.4]) We call EX a conditional expec-
tation from Y onto X with respect to EA and EB if EX satisfies the
following:
(1) EX (c · x) = EA(c) · x for any c ∈ C, x ∈ X,
(2) EX (a · y) = a · EX(y) for any a ∈ A, y ∈ Y ,
(3) EA(Chy, xi) = ChEX(y), xi for any x ∈ X, y ∈ Y ,
(4) EX (x · d) = x · EB(d) for any d ∈ D, x ∈ X,
(5) EX (y · b) = EX (y) · b for any b ∈ B, y ∈ Y ,
(6) EB(hy, xiD) = hEX(y), xiD for any x ∈ X, y ∈ Y .
With the above notation, we have the following by [13, Theorem 2.9].
Theorem 2.2. We suppose that inclusions A ⊂ C and B ⊂ D are
strongly Morita equivalent with respect to a C − D-equivalence bimod-
ule Y and its closed subspace X. If there is a conditional expectation
EA of Watatani index-finite type from C onto A, then there are a con-
ditional expectation of Watatani index-finite type from D onto B and a
conditional expectation EX from Y onto X with respect to EA and EB.
3
Also, if there is a conditional expectation EB of Watatani index-finite
type from D onto B, then we have the same result as above.
Next, we mention on the upward basic construction of equivalence
bimodules. Let A ⊂ C, B ⊂ D and X ⊂ Y be as above. We suppose
that there are conditional expectation EA and EB from C and D onto
A and B, respectively and that they are of Watatani index-finite type.
We also suppose that there is a conditional expectation EX from Y
onto X with respect to EA and EB. Let C1 and D1 be the C ∗-basic
constructions for EA and EB, respectively. And let EC and ED be the
dual conditional expectations of EA and EB, respectively. We regard
C and D as a C1 − A-equivalence bimodule and a D1 − B-equivalence
bimodule, respectively. Let
Also, let EY be the linear map from Y1 onto Y defined by
Y1 = C ⊗A X ⊗B eD.
for any c ∈ C, d ∈ D, x ∈ X. Furthermore, let φ be the linear map
from Y to Y1 defined by
EY (c ⊗ x ⊗ ed) = IndW (EA)−1c · x · d∗
φ(y) =Xi,j
ui ⊗ EX (u∗
i · y · vj) ⊗ evj
j )} are quasi- bases for EA
for any y ∈ Y , where {(ui, u∗
and EB, respectively. By [13, Lemma 6.1 and Corollary 6.3], we obtain
the following:
i )} and {(vj, v∗
Theorem 2.3. ([13, Lemma 6.1 and Corollary 6.3]) We can regard Y
as a closed subspace of a C1 − D1-equivalence bimodule Y1 using the
map φ and the inclusions C ⊂ C1 and D ⊂ D1 are strongly Morita
equivalent with Y1 and its closed subspace Y .
By [13, Lemma 6.4 and Remark 6.6], we can see that EY is a condi-
tional expectation from Y1 onto Y with respect to EC and ED and it
is independent of the choice of quasi-bases for EA and EB.
Definition 2.6. ([13, Definition 6.5]) We call Y1 the upward basic con-
struction of Y for EX, and EY is called the dual conditional expectation
of EX.
3. Definitions and basic properties
Let A ⊂ C, B ⊂ D and K ⊂ L be inclusions of C ∗-algebras with
AC = C, BD = D and KL = L, respectively. Let Y and W be a
C − D-equivalence bimodule and a D − L-equivalence bimodule and X
and Z their closed subspaces satisfying Conditions (1), (2) in Definition
2.4, respectively. That is, the inclusions A ⊂ C and B ⊂ D are strongly
Morita equivalent with respect to the C − D-equivalence bimodule Y
4
and its closed subspace X and the inclusion B ⊂ D and K ⊂ L are
strongly Morita equivalent with respect to the D − L-equivalence bi-
module and its closed subspace Z. Let X ⊗D Z be the closure of the
linear span of the set
{x ⊗ z ∈ Y ⊗D W x ∈ X, z ∈ Z}.
Clearly X ⊗D Z is a closed linear subspace of Y ⊗D W and we can
regard X ⊗D Z as an A − K-equivalence bimodule.
Lemma 3.1. With the above notation, X ⊗DZ is isomorphic to X ⊗B Z
as A−K-equivalence bimodules, where X ⊗D Z is regarded as an A−K-
equivalence bimodule.
Proof. Let (X ⊗D Z)0 be the linear span of the set
{x ⊗ z ∈ Y ⊗D W x ∈ X, z ∈ Z}
and let (X ⊗B Z)0 be the algebraic relative tensor product of the A−B-
equivalence bimodule X and the B − K-equivalence bimodule Z. We
note that X ⊗B Z is the completion of (X ⊗B Z)0. Let π0 be the
map from (X ⊗B Z)0 to (X ⊗D Z)0 defined by π0(x ⊗B z) = x ⊗D z
for any x ∈ X, z ∈ Z.
It is well-defined and surjective. By easy
computations, π0 preserves the left A-valued inner product and the
right K-valued inner product. Hence we obtain an A − K-equivalence
bimodule isomorphism π of X ⊗B Z onto X ⊗D Z. Therefore, we obtain
the conclusion.
(cid:3)
By Lemma 3.1, we identify X ⊗B Z with X ⊗D Z, the closed subspace
of Y ⊗D W under the above situations.
Let A ⊂ C be an inclusion of C ∗-algebras with AC = C. Let Y
be a C − C-equivalence bimodule and X its closed subspace satisfy-
ing Conditions (1), (2) in Definition 2.4. Let Equi(A, C) be the set
of all such pairs (X, Y ) as above. We define an equivalence relation
" ∼ " as follows: For (X, Y ), (Z, W ) ∈ Equi(A, C), (X, Y ) ∼ (Z, W )
in Equi(A, C) if and only if there is a C − C-equivalence bimodule iso-
morphism Φ of Y onto W such that the restriction of Φ to X, ΦX is
an A − A-equivalence bimodule isomorphism of X onto Z. We denote
by [X, Y ], the equivalence class of (X, Y ) in Equi(A, C). We remark
here that we have the following lemma:
Lemma 3.2. We suppose that inclusions of C ∗-algebras A ⊂ C and
B ⊂ D are strongly Morita equivalent with respect to C −D-equivalence
bimodules Y and W and their closed subspaces X and Z, respectively.
If there is a C − D-equivalence bimodule isomorphism Φ of Y onto W
such that ΦX is a bijection from X onto Z, then ΦX is an A − B-
equivalence bimodule isomorphism of X onto Z.
Proof. Let a ∈ A, b ∈ B and x, y ∈ X. By Definition 2.4, a · x ∈ X
and since ΦX is a bijection from X onto Z, Φ(a · x) ∈ Z. Furthermore,
5
since Φ is a C − D-equivalence bimodule isomorphism of Y onto W ,
Φ(a · x) = a · Φ(x). Similarly we obtain that Φ(x · b) = Φ(x) · b. Also
by Definition 2.4,
AhΦ(x), Φ(y)i = ChΦ(x) , Φ(y)i = Chx, yi = Ahx, yi
hΦ(x), Φ(y)iB = hΦ(x) , Φ(y)iD = hx, yiD = hx , yiB.
Hence ΦX is an A − B-equivalence bimodule isomorphism of X onto
Z.
(cid:3)
By Lemma 3.2, we can see that for (X, Y ), (Z, W ) ∈ Equi(A, C),
(X, Y ) ∼ (Z, W ) in Equi(A, C) if and only if there is a C−C-equivalence
bimodule isomorphism Φ of Y onto W such that Φ(X) = Z.
Let Pic(A, C) = Equi(A, C)/ ∼. We define the product in Pic(A, C)
as follows: For (X, Y ), (Z, W ) ∈ Equi(A, C)
[X, Y ][Z, W ] = [X ⊗A Z , Y ⊗C W ],
where the A − A-equivalence bimodule X ⊗A Z is identified with the
closed subspace X ⊗C Z of Y ⊗C W defined in the above. We note
that Y ⊗C W and its closed subspace X ⊗A Z satisfy Conditions (1),
(2) in Definition 2.4 by [13, Proposition 2.3]. By Lemma 3.1 and easy
computations, we can see that Pic(A, C) is a group. We regard (A, C)
as an element in Equi(A, C) in the evident way. Then [A, C] is the unit
element in Pic(A, C). For any element (X, Y ) ∈ Equi(A, C), (eX,eY ) ∈
Equi(A, C) and [eX,eY ] is the inverse element of [X, Y ] in Pic(A, C),
where eX and eY are the dual A − A-equivalence bimodule of X and the
dual C − C-equivalence bimodule of Y , respectively. We note that eX
can be a closed subspace of eY . We call the group Pic(A, C) defined
in the above, the Picard group of the inclusion of C ∗-algebras A ⊂ C
with AC = C.
Let A ⊂ C and B ⊂ D be inclusions of C ∗-algebras with AC = C
and BD = D, respectively. We suppose that A ⊂ C and B ⊂ D are
strongly Morita equivalent with respect to a C − D-equivalence W and
its closed subspace Z. Let g be the map from Pic(A, C) to Pic(B, D)
defined by
g([X, Y ]) = [eZ ⊗A X ⊗A Z , fW ⊗C Y ⊗C W ]
for any (X, Y ) ∈ Equi(A, C), where eZ ⊗A X ⊗A Z is regarded as a
closed subspace of fW ⊗C Y ⊗C W in the same way as in Lemma 3.1.
Lemma 3.3. With the same notation as above, g is an isomorphism
of Pic(A, C) onto Pic(B, D),
Proof. The C − C-equivalence bimodule W ⊗DfW is isomorphic to the
C − C-equivalence bimodule C by the isomorphism
W ⊗D fW −→ C : z ⊗ ew 7→ Chz, wi
6
for any z, w ∈ W . The restriction of the above isomorphism to Z ⊗B eZ
is an isomorphism of Z ⊗B eZ onto A by Definition 2.4. Also, the C −C-
equivalence bimodule C ⊗C Y is isomorphic to the C − C-equivalence
bimodule Y by the isomorphism
C ⊗C Y −→ Y : c ⊗ y 7→ c · y
for any c ∈ C, y ∈ Y . The restriction of the above isomorphism to
A ⊗A X is an isomorphism of A ⊗A X onto X by Definition 2.4. By the
above discussions, we can see that g is an isomorphism of Pic(A, C)
onto Pic(B, D).
(cid:3)
Let α be an automorphism of C such that the restriction of α to
A, αA is an automorphism of A. Let Aut(A, C) be the group of all
such automorphisms. We construct an element in Equi(A, C) from
an element in Aut(A, C) as follows: Let α ∈ Aut(A, C). Let Yα be
the C − C-equivalence bimodule induced by α in the same way as in
Preliminaries. Let Xα be the A − A-equivalence bimodule induced by
αA in the same way as above. Then clearly (Xα, Yα) ∈ Equi(A, C) and
for any α, β ∈ Aut(A, C),
[Xα◦β , Yα◦β] = [Xα , Yα][Xβ , Yβ]
in Pic(A, C). Let π be the map from Aut(A, C) to Pic(A, C) defined
by
π(α) = [Xα , Yα]
for any α ∈ Aut(A, C). By the above discussions, π is a homomor-
phism of Aut(A, C) to Pic(A, C). Let u be a unitary element in M(A).
Then Ad(u) is a generalized inner automorphism of A. Since AC = C,
by Izumi [7] u ∈ M(C). Thus Ad(u) is also a generalized inner auto-
morphism of C. Let Int(A, C) be the set of all such automorphisms in
Aut(A, C). We note that Int(A, C) = Int(A). Let ı be the inclusion
map of Int(A, C) to Aut(A, C).
Lemma 3.4. With the above notation, the sequence
1 −→ Int(A, C)
ı−→ Aut(A, C)
π−→ Pic(A, C)
is exact.
Proof. Let u ∈ M(A). Then u ∈ M(C). We show that C ∼= YAd(u)
as C − C-equivalence bimodules. Let Φ be the map from C to YAd(u)
defined by Φ(x) = xu∗ for any x ∈ C. Then for any a ∈ C, x, y ∈ C,
Φ(a · x) = Φ(ax) = axu∗ = a · Φ(x)
Φ(x · a) = xu∗uau∗ = Φ(x) · a
ChΦ(x), Φ(y)i = Chxu∗ , yu∗i = xy∗ = Chx, yi
hΦ(x), Φ(y)iC = hxu∗, yu∗iC = Ad(u)−1(ux∗yu∗) = x∗y = hx, yiC.
7
Hence Φ is a C −C-equivalence bimodule isomorphism of C onto YAd(u).
Furthermore, since u ∈ M(A), xu∗ ∈ XAd(u) for any x ∈ A. In the same
way as above, we can see that ΦAd(u) is an A−A-equivalence bimodule
isomorphism of A onto XAd(u). Thus
[XAd(u) , YAd(u)] = [A, C]
in Pic(A, C). Let α ∈ Aut(A, C) with [Xα, Yα] = [A, C] in Pic(A, C).
Then there is a C − C-equivalence bimodule isomorphism Φ of C onto
Yα such that ΦA is an A − A-equivalence bimodule isomorphism of A
onto Xα. In the same way as the proof of [2, Proposition 3.1], we can
obtain unitary elements u1 ∈ M(C) and u ∈ M(A) such that
u1 = (Φ ◦ α−1 , Φ) , u = ((Φ ◦ α−1)A , ΦA)
α = Ad(u∗
1) , αA = Ad(u∗),
where (Φ ◦ α−1, Φ) and ((Φ ◦ α−1)A, ΦA) are double centralizers of
C and A, respectively. Then for any a ∈ A, u1a = (Φ ◦ α−1)(a) =
ua. Since AC = C, u1 = u. Hence π([XAd(u∗) , YAd(u∗)]) = [A, C].
Therefore, we obtain the conclusion.
(cid:3)
Let A ⊂ C be an inclusion of C ∗-algebras such that A is σ-unital
and AC = C. Let K be the C ∗-algebra of all compact operators on a
countably infinite dimensional Hilbert space and let As = A ⊗ K and
C s = C ⊗ K, respectively. Let (X, Y ) ∈ Equi(As, C s). Let LX and LY
be the linking C ∗-algebras induced by X and Y , respectively. Let
p =(cid:20)1A ⊗ 1M (K) 0
0(cid:21) ,
0
q =(cid:20)0
0 1A ⊗ 1M (K)(cid:21)
0
in M(LX ). Then p and q are full projections in M(LX ). By easy
computations, we can see that LX LY = LY . Hence M(LX ) ⊂ M(LY )
by Izumi [7]. Since p and q are full projections in M(LX ), by Brown [1,
Lemma 2.5], there is a partial isometry w ∈ M(LX ) such that w∗w = p,
ww∗ = q. Then we note that w ∈ M(LY ). Let θ be the map from pLY p
to qLY q defined by
θ((cid:20)x 0
0 0(cid:21)) = w(cid:20)x 0
0 0(cid:21) w∗
for any x ∈ C s. By easy computations, we can see that θ is an iso-
morphism of pLY p onto qLY q. Identifying pLY p and qLY q with C s,
we can regard θ as an automorphism of C s. We also denote it by the
same symbol θ. Since w ∈ M(LX ), θAs is an automorphism of As. Let
(Xθ, Yθ) be the element in Equi(A, C) induced by θ. Then we can see
that [Xθ, Yθ] = [X, Y ] in Pic(A, C) in the same way as in the proof of
[2, Theorem 3.4]. By the above discussions and Lemma 3.4. we obtain
the following proposition:
8
Proposition 3.5. With the above notation and assumptions, the se-
quence
1 −→ Int(As, C s)
ı−→ Aut(As, C s)
π−→ Pic(As, C s) −→ 1
is exact.
4. Some lemmas
In this section, we shall prepare some lemmas for the next sections.
Let A ⊂ C, B ⊂ D and K ⊂ L be unital inclusions of unital C ∗-
algebras. Let EA, EB and EK be conditional expectations from C, D
and L onto A, B and K, respectively. We suppose that they are of
Watatani index-finite type. Also, we suppose that A ⊂ C and B ⊂ D
are strongly Morita equivalent with respect to a C − D-equivalence
bimodule Y and its closed subspace X and suppose that B ⊂ D
and K ⊂ L are strongly Morita equivalent with respect to a D − L-
equivalence bimodule W and its closed subspace Z. Also, we suppose
that A′ ∩ C = C1. Then B ′ ∩ D = C1 and K ′ ∩ L = C1 by [13, Lemma
10.3].
Lemma 4.1. With the above notation and assumptions, there is the
unique conditional expectation EX from Y onto X with respect to EA
and EB.
Proof. By Theorem 2.2, we can see that there are a conditional expec-
tation F B of Watatani index-finite type from D onto B and a condi-
tional expectation EX from Y onto X with respect to EA and F B.
But F B = EB by Proposition 2.1 since A′ ∩ C = C1. Hence EX is a
conditional expectation from Y onto X with respect to EA and EB.
Next we show the uniqueness of EX . Let F X be another conditional
expectation from Y onto X with respect to EA and EB. Then by the
definitions of EX and F X, for any x ∈ X, y ∈ Y ,
hx , EX (y)iB = EB(hx, yiD) = hx , F X(y)iB.
Hence EX(y) = F X (y) for any y ∈ Y .
(cid:3)
eX from
eX (ey) =
By Lemma 4.1, there is the unique conditional expectation E
eY onto eX with respect to EB and EA.
Lemma 4.2. With the above notation and assumptions, E
^EX(y) for any y ∈ Y .
Proof. This is immediate by Definition 2.5 and routine computations.
(cid:3)
Also, by Lemma 4.1, there are the unique conditional expectations
EX and EZ from Y and W onto X and Z with respect to EA, EB and
EB, EK, respectively. Also, there is the unique conditional expectation
EX⊗B Z from Y ⊗D W onto X ⊗B Z with respect to EA and EK. Let
9
C1, D1 and L1 be the C ∗-basic constructions for EA, EB and EK,
respectively. Let Y1 and W1 be the upward basic constructions of Y
and W for EX and EZ, respectively and let (Y ⊗D W )1 be the upward
basic construction of Y ⊗D W for EX⊗B Z.
Lemma 4.3. With the above notation and asumptions, (Y ⊗D W )1
Y1 ⊗D1 W1 as C1 − D1-equivalence bimodules.
∼=
Proof. By the definitions of Y1, W1 and (Y ⊗D W )1,
Thus
Y1 = C⊗AX ⊗B eD , W1 = D ⊗B Z ⊗K eL,
(Y ⊗D W )1 = C ⊗A X ⊗B Z ⊗K eL
Y1 ⊗D1 W1 = C ⊗A X ⊗B eD ⊗D1 D ⊗B Z ⊗K eL.
Φ(c ⊗ x ⊗ ed ⊗ d′ ⊗ z ⊗el) = c ⊗ x ⊗ EB(d∗d′) · z ⊗el
Let Φ be the map from Y1 ⊗D1 W1 to (Y ⊗D W )1 defined by
for any c ∈ C, x ∈ X, d, d′ ∈ D, z ∈ Z, l ∈ L. By routine computa-
tions, Φ is a C1 − D1-equivalence bimodule isomorphism of Y1 ⊗D1 W1
onto (Y ⊗D W )1. Therefore, we obtain the conclusion.
(cid:3)
Let {(ui, u∗
i )}, {(vi, v∗
i )} and {(si, s∗
i )} be quasi-bases for conditional
expectations EA, EB and EK, respectively. We recall that Y and W
are regarded as closed subspaces of Y1 and W1 by the linear maps φY
and φW defined by
φY (y) =Xi,j
φW (w) =Xk,l
ui ⊗ EX (u∗
vk ⊗ EZ(v∗
i · y · vj) ⊗ evj,
k · w · sl) ⊗esl
for any y ∈ Y , w ∈ W , respectively.
Lemma 4.4. With the above notation and asssumptions,
EX⊗BZ(y ⊗ w) =Xj
EX(y · vj) ⊗ EZ(v∗
j · w)
for any y ∈ Y , w ∈ W .
Proof. By Lemma 4.1, we have only to show the right hand side in the
above equation defines a linear map satisfying Conditions (1)-(6) in
Definition 2.5. They are proved in the routine computations.
(cid:3)
Also, we recall that Y ⊗D W is regarded as a closed subspace of
(Y ⊗D W )1 by the linear map φY ⊗DW defined by
φY ⊗DW (y ⊗ w) =Xi,l
for any y ∈ Y , w ∈ W .
ui ⊗ EX⊗B Z(ui · y ⊗ w · sl) ⊗esl
10
Lemma 4.5. With the above notation, let Φ be the C1 −L1-equivalence
bimodule isomorphism of Y1 ⊗D1 W1 onto (Y ⊗D W )1 defined in Lemma
4.3. Then
φY ⊗DW = Φ ◦ (φY ⊗ φW ).
Proof. For any y ∈ Y , w ∈ W ,
φY (y) ⊗ φW (w) = Xi,j,t,l
By the definition of Φ,
ui ⊗ EX (u∗
i · y · vj) ⊗ evj ⊗ vt ⊗ EZ(v∗
t · w · sl) ⊗esl.
Φ(φY (y) ⊗ φW (w))
ui ⊗ EX(u∗
ui ⊗ EX(u∗
ui ⊗ EX(u∗
= Xi,j,t,l
= Xi,j,t,l
=Xi,j,l
φY ⊗DW (y ⊗ w) =Xi,l
=Xi,j,l
i · y · vj) ⊗ EA(v∗
j vt) · EZ(v∗
i · y · vj) ⊗ EZ(EA(v∗
j vt)v∗
t · w · sl) ⊗esl
t · w · sl) ⊗esl
i · y · vj) ⊗ EZ(v∗
ui ⊗ EX⊗B Z(u∗
j · w · sl) ⊗esl.
i · y ⊗ w · sl) ⊗esl
On the other hand, by Lemma 4.4
Therefore we obtain the conclsuion.
ui ⊗ EX(u∗
i · y · vj) ⊗ EZ(v∗
j · w · sl) ⊗esl.
(cid:3)
5. The C ∗-basic construction
Let A ⊂ C be a unital inclusion of unital C ∗-algebras. We suppose
that A′ ∩ C = C1 and that there is a conditional expectation EA of
Watatani index-finite type from C onto A. We denote its Watatani
index by IndW (EA). Then IndW (EA) ∈ C1. Let C1 be the C ∗-basic
construction for EA and eA the Jones projection for EA. Let (X, Y ) ∈
Equi(A, C). Let Y1 be the upward basic construction of Y for EX .
Then Y1 is uniquely determined by Lemma 4.1. We recall that Y is
regarded as a closed subspace of Y1 by the map, which is denoted by
φY , from Y to Y1 defined by
φY (y) =Xi,j
ui ⊗ EX(u∗
i · y · uj) ⊗ euj,
i )}i is a quasi-basis for EA, which is independent of the
where {(ui, u∗
choice of a quasi-basis for EA (See Preliminaries). Let f be the map
from Pic(A, C) to Pic(C, C1) defined by
f ([X, Y ]) = [Y, Y1]
11
for any (X, Y ) ∈ Equi(A, C). Then by Theorem 2.3, (Y, Y1) ∈ Equi(C, C1).
Hence f is well-defined. In this section, we shall show that f is an iso-
morphism of Pic(A, C) onto Pic(C, C1). First we show that f is a
homomorphism of Pic(A, C) to Pic(C, C1).
Lemma 5.1. With the same notation as above, f is a homomorphism
of Pic(A, C) to Pic(C, C1).
Proof. Let (X, Y ), (Z, W ) ∈ Equi(A, C). Then
f ([X, Y ][Z, W ]) = f ([X ⊗A Z , Y ⊗C W ]) = [Y ⊗C W , (Y ⊗C W )1],
where (Y ⊗C W )1 is the upward basic construction of Y ⊗C W for EX⊗AZ
where EX⊗AZ is the conditional expectation from Y ⊗C W onto X ⊗A Z
with respect to EA and EA. By Lemmas 4.3, 4.5, we can see that there
is a C1 −C1-equivalence bimodule isomorphism preserving the elements
in Y ⊗C W . Therefore, we obtain that
f ([X, Y ][Z, W ]) = f ([X, Y ])f ([Z, W ]).
(cid:3)
Let EC be the dual conditional expectation from C1 onto C. Let
eC and C2 be the Jones projection and the C ∗-basic construction for
EC, respectively. Then the unital inclusion C1 ⊂ C2 is strongly Morita
equivalent to the unital inclusion A ⊂ C with respect to the C2 − C-
equivalence bimodule C1 and its closed subspace C by [13, Lemma 4.2],
where C is regarded as a closed subspace of C1 by the linear map θC
defined by
θC(a) = IndW (EA)
1
2 aeA
for any a ∈ C. Let g be the map from Pic(A, C) to Pic(C1, C2) defined
by
g([X, Y ]) = [C ⊗A X ⊗A eC , C1 ⊗C Y ⊗C fC1]
for any (X, Y ) ∈ Equi(A, C). Then g is an isomorphism of Pic(A, C)
onto Pic(C1, C2) by Lemma 3.3. Let f1 be the homomorphism of
Pic(C, C1) to Pic(C1, C2) defined by
f1([Y, Y1]) = [Y1, Y2]
for any (Y, Y1) ∈ Equi(C, C1), where Y2 is the upward basic construc-
tion of Y1 for EY and EY is the conditional expectation from Y1 onto
Y with respect to EC and EC.
Lemma 5.2. With the above notation, f1 ◦ f = g on Pic(A, C).
Proof. Let (X, Y ) ∈ Equi(A, C). By the definitions of f and f1,
(f1 ◦ f )([X, Y ]) = [Y1, Y2],
where Y1 = C ⊗A X ⊗A eC and Y2 = C1 ⊗C Y ⊗C fC1. Also,
g([X, Y ]) = [C ⊗A X ⊗A eC , C1 ⊗C Y ⊗C fC1]
12
by the definition of g. We note that Y1 is regarded as a closed subspace
Y2 by the linear map φY1 from Y1 to Y2 defined by
φY1(c ⊗ x ⊗ ed) =Xi,j
wi ⊗ EY (w∗
i · c ⊗ x ⊗ ed · wj) ⊗fwj
1
θC⊗AX⊗A
for any c, d ∈ C, x ∈ X, where {(wi, w∗
defined by wi = IndW (EA)
2 uieA. We also note that C ⊗A X ⊗A eC
is regarded as a closed subspace of C1 ⊗C Y ⊗C fC1 by the linear map
i )} is a quasi-basis for EC
eC from C ⊗A X ⊗A eC to C1 ⊗C Y ⊗C fC1 defined by
eC(c ⊗ x ⊗ ed) = IndW (EA)ceA ⊗ x ⊗gdeA
for any c, d ∈ C, x ∈ X. In order to show that f1 ◦ f = g, we need to
prove that
θC⊗AX⊗A
φY1(c ⊗ x ⊗ ed) = θC⊗AX⊗A
eC(c ⊗ x ⊗ ed)
for any c, d ∈ C, x ∈ X. For any c, d ∈ C, x ∈ X,
uieA ⊗ EY (eAu∗
uieA ⊗ EY (EA(u∗
i · c ⊗ x ⊗ ed · ujeA) ⊗ gujeA
j d)) ⊗ gujeA
i c) ⊗ x ⊗ ^EA(u∗
uieA ⊗ EA(u∗
i c) · x · EA(d∗uj) ⊗ gujeA
i c)eA ⊗ x ⊗ [ujEA(u∗
j d)eA]e
φY1(c ⊗ x ⊗ ed)
= IndW (EA)2Xi,j
= IndW (EA)2Xi,j
= IndW (EA)Xi,j
= IndW (EA)Xi,j
= IndW (EA)ceA ⊗ x ⊗gdeA
eC(c ⊗ x ⊗ ed)
= θC⊗AX⊗A
uiEA(u∗
by the definition of EY (See Preliminaries). Therefore, we obtain the
conclusion.
(cid:3)
By Lemmas 3.3, 5.2, we can see that (g−1 ◦ f1) ◦ f = id on Pic(A, C).
Next, we shall show that
f ◦ (g−1 ◦ f1) = id
on Pic(C, C1). Let (Y, Y1) ∈ Equi(C, C1). Then
(g−1 ◦ f1)([Y, Y1]) = g−1([Y1, Y2]) = [eC ⊗C1 Y1 ⊗C1 C , fC1 ⊗C2 Y2 ⊗C2 C1],
13
where Y2 is the upward basic construction of Y1 for EY . Thus Y2 =
(f ◦ g−1 ◦ f1)([Y, Y1])
C1 ⊗C Y ⊗C fC1. Hence
= [fC1 ⊗C2 Y2 ⊗C2 C1 , C ⊗A eC ⊗C1 Y1 ⊗C1 C ⊗A eC]
= [fC1 ⊗C2 C1 ⊗C Y ⊗C fC1 ⊗C2 C1 , C ⊗A eC ⊗C1 Y1 ⊗C1 C ⊗A eC].
By easy computations, fC1 ⊗C2 C1 ⊗C Y ⊗C fC1 ⊗C2 C1 is isomorphic to
Y as C − C-equivalence bimodules by the linear map ΦY from fC1 ⊗C2
C1 ⊗C Y ⊗C fC1 ⊗C2 C1 to Y defined by
ΦY (ec1 ⊗ y2 ⊗ d1) = EC(c∗
for any a1, b1, c1, d1 ∈ C1, y ∈ Y and y2 = a1 ⊗ y ⊗eb1. Also, C ⊗A eC ⊗C1
Y1 ⊗C1 C ⊗A eC is isomorphic to Y1 as C1 − C1-equivalence bimodules
by the linear map ΦY1 from C ⊗A eC ⊗C1 Y1 ⊗C1 C ⊗A eC to Y1 defined
for any a, b, c, d ∈ C, y1 ∈ Y1. We note thatfC1⊗C2 C1⊗C Y ⊗CfC1⊗C2 C1
is regarded as a closed subspace of C ⊗A eC ⊗C1 Y1 ⊗C1 C ⊗A eC by the
fC1⊗C2 C1 from fC1 ⊗C2 C1 ⊗C Y ⊗C fC1 ⊗C2 C1
to C ⊗A eC ⊗C1 Y1 ⊗C1 C ⊗A eC defined by
(ec1 ⊗ a1 ⊗ y ⊗ eb1 ⊗ d1)
ΦY1(c ⊗ea ⊗ y1 ⊗ b ⊗ ed) = ceAa∗ · y1 · beAd∗
fC1⊗C2 C1
eC ⊗C1 Y1⊗C1 C(u∗
linear map φ fC1⊗C2 C1⊗C Y ⊗C
φ fC1⊗C2 C1⊗C Y ⊗C
ui ⊗ E
1a1) · y · EC(b∗
1d1)
by
=Xi,j
i · ec1 ⊗ a1 ⊗ y ⊗ eb1 ⊗ d1 · uj) ⊗ euj
for any a1, b1, c1, d1 ∈ C1, y ∈ Y , where E
eC⊗C1 Y1⊗C1 C is the unique
conditional expectation from fC1 ⊗C2 C1 ⊗C Y ⊗CfC1 ⊗C2 C1 onto eC ⊗C1
Y1 ⊗C1 C with respect to EA and EA, which satisfies Conditions (1)-(6)
in Definition 2.5 by Lemma 4.1.
Lemma 5.3. With the above notation,
E
eC⊗C1 Y1⊗C1 C(ec1 ⊗ a1 ⊗ y ⊗ eb1 ⊗ d1) = ^EC(a∗
for any a1, b1, c1, d1 ∈ C1, y ∈ Y .
1c1) ⊗ y ⊗ EC(b∗
1d1)
Proof. Let y2 = a1 ⊗y ⊗eb1. Let {(ri, r∗
i )} be the quasi-basis for EC1, the
dual conditional expectation of EC from C2 onto C1, which is defined
by ri = IndW (EA)
2 wieC, where eC is the Jones projection for EC. Let
F C be the conditional expectation from C1, the C2 − C-equivalence
bimodule onto C, the C1 − A-equivalence bimodule with respect to
EC1 and EA, which is defined by
1
F C(ceCd) = cEC(d)eC
14
j · d1)
E
F
i · y2 · rj) ⊗ F C(r∗
for any c, d ∈ C. Also, let F
of EY1 (See Preliminaries),
eC be the conditional expectation from fC1
onto eC induced by F C. Then by Lemmas 4.2, 4.4, and the definition
eC⊗C1 Y1⊗C1 C(ec1 ⊗ y2 ⊗ d1)
=Xi,j
eC(ec1 · ri) ⊗ EY1(r∗
= IndW (EA)2Xi,j
= IndW (EA)Xi,j
= IndW (EA)Xi,j
i a1) ⊗ y ⊗ ^EC(w∗
i c1)) ⊗ EY1(EC(w∗
i a1) · y · EC(w∗
j b1)∗
1wiEC(w∗
i c1)))]e⊗ y
i c1))]e⊗ EC(w∗
⊗ F C(EC(w∗
⊗ F C(EC(w∗
eC( ^EC(w∗
[F C(EC(w∗
j d1))
j d1))
[F C(EC(a∗
F
j b1))
⊗ F C(EC(EC(b∗
1wj)w∗
= IndW (EA)[F C(EC(a∗
j d1))
1c1))]e⊗ y ⊗ F C(EC(b∗
1d1)).
Here since we regard EC(a∗
bimodule C1,
F C(EC(a∗
1c1)) =Xi
=Xi
1c1) as an element in the C2 − C-equivalence
F C(EC(a∗
1c1)wieCw∗
i )
EC(a∗
1c1)wiEC(w∗
i )eC = EC(a∗
1c1)eC.
Since we regard the element EC(a∗
bimodule C1 as the element IndW (EA)−
equivalence bimodule by [13, Section 4],
1c1)eC in the C2 − C-equivalence
1c1) in C, the C1 − A-
2 EC(a∗
1
F C(EC(a∗
1c1)) = IndW (EA)−
1
2 EC(a∗
1c1).
Similarly, F C(EC(b∗
the conclusion.
1d1)) = IndW (A)−
1
2 EC(b∗
1d1). Therefore, we obtain
(cid:3)
Lemma 5.4. With the above notation, f ◦ g−1 ◦ f1 = id on Pic(C, C1).
in the above. Also, let φ fC1⊗C2 C1⊗C Y ⊗C
Proof. Let ΦY be the isomorphism offC1⊗C2C1⊗CY ⊗CfC1⊗C2C1 onto Y
and ΦY1 the isomorphism of C ⊗A eC ⊗C1 Y1 ⊗C1 C ⊗A eC onto Y1 defined
map from fC1 ⊗C2 C1 ⊗C Y ⊗CfC1 ⊗C2 C1 into C ⊗A eC ⊗C1 Y1 ⊗C1 C ⊗A eC
defined in the above. It suffices to show that
fC1⊗C2 C1 be the injective linear
ΦY = ΦY1 ◦ φ fC1⊗C2 C1⊗C Y ⊗C
fC1⊗C2 C1
15
in order to prove that f ◦ g−1 ◦ f1 = id on Pic(C, C1). Let a1, b1, c1, d1 ∈
C1, y ∈ Y and let y2 = a1 ⊗ y ⊗ eb1. Then by Lemma 5.3,
φ fC1⊗C2 C1⊗C Y ⊗C
ui ⊗ E
ui ⊗ E
eC ⊗C1 Y1⊗C1 C(u∗
fC1⊗C2 C1(ec1 ⊗ y2 ⊗ d1)
i · ec1 ⊗ a1 ⊗ y ⊗ eb1 ⊗ d1 · uj) ⊗ euj
eC ⊗C1 Y1⊗C1 C(gc1ui ⊗ a1 ⊗ y ⊗ eb1 ⊗ d1uj) ⊗ euj
ui ⊗ [EC(a∗
1c1ui)]e⊗ y ⊗ EC(b∗
=Xi,j
=Xi,j
=Xi,j
1d1uj) ⊗ euj.
Hence
(ΦY1 ◦ φ fC1⊗C2 C1⊗C Y ⊗C
fC1⊗C2 C1)(ec1 ⊗ y2 ⊗ d1)
1d1uj)eAu∗
1a1) · y · EC(b∗
j
uieAEC(u∗
i c∗
uieAu∗
i EC(c∗
1a1) · y · EC(b∗
1d1)ujeAu∗
j
=Xi,j
=Xi,j
= EC(c∗
1a1) · y · EC(b∗
1d1).
On the other hand,
Therefore, we obtain the conclusion.
ΦY (ec1 ⊗ y2 ⊗ d1) = EC(c∗
1a1) · y · EC(b∗
1d1).
(cid:3)
Theorem 5.5. Let A ⊂ C be a unital inclusion of unital C ∗-algebras.
We suppose that A′∩C = C1 and that there is a conditional expectation
EA of Watatani index-finite type from C onto A. Let C1 be the C ∗-basic
construction for EA. Then Pic(A, C) ∼= Pic(C, C1).
Proof. This is immediate by Lemmas 5.2, 5.4.
(cid:3)
6. The Picard groups for inclusions of C ∗-algebras and
the ordinary Picard groups
In this section, we shall investigate the relation between the Picard
groups for inclusion of C ∗-algebras and the ordinary Picard groups.
Let A ⊂ C be an inclusion of C ∗-algebras with AC = C and let fC be
the homomorphism from Pic(A, C) to Pic(C) defined by
fC : Pic(A, C) → Pic(C) : [X, Y ] 7→ [Y ],
where Pic(C) is the ordinary Picard group of C. In this section, we
suppose that AC = C for any inclusions of C ∗-algebras A ⊂ C.
Let u be a unitary element in M(C) satisfying that uau∗, u∗au ∈ A
for any a ∈ A. We regard Au as an A − A-equivalence bimodule as
16
follows: In the usual way, we regard Au as a vector space over C. We
define the left A-action and the right action by
a · xu = axu,
xu · a = xua = x(uau∗)u
for any a, x ∈ A. We also define the left A-valued inner product and
the right A-valued inner product by
Ahxu, yui = xuu∗y∗ = xy∗,
hxu, yuiA = u∗x∗yu
for any x, y ∈ A. Furthermore, Au is a closed subspace of C, the trivial
C − C-equivalence bimodule and we can see that [Au, C] is an element
in Pic(A, C) by easy computations. Let Aut(A, C) be the group of
all automorphisms of C such that αA is an automorphism of A. Let
α ∈ Aut(A, C) and let [Xα, Yα] be the element in Pic(A, C) induced by
α, which is defined in Section 3
Lemma 6.1. With the above notation, let α ∈ Aut(A, C). Then the
following conditions are equivalent:
(1) [Xα, Yα] ∈ KerfC,
(2) There is a unitary element u ∈ M(C) satisfying that uau∗, u∗au ∈
A for any a ∈ A and that [Xα, Yα] = [Au, C] in Pic(A, C).
Proof. (1) ⇒ (2): We suppose that [Xα, Yα] ∈ KerfC. Then [Yα] = [C]
in Pic(C). Hence there is a unitary element u ∈ M(C) such that α =
Ad(u) on C. Since αA is also an automorphism of A, we can see that
uau∗, u∗au ∈ A for any a ∈ A. Let Φ be the linear map from C to YAd(u)
defined by Φ(x) = xu∗ for any x ∈ C. Then by the proof of Lemma 3.4,
Φ is a C −C-equivalence bimodule isomorphism of C onto YAd(u). Also,
we can see that ΦAu is an A − A-equivalence bimodule isomorphism
of Au onto Xα. Indeed, for any x ∈ A, Φ(xu) = xuu∗ = x ∈ Xα. By
Lemma 3.2 ΦAu is an A − A-equivalence bimodule isomorphism of Au
onto Xα.
(2) ⇒ (1): It is clear by the definition of fC.
(cid:3)
Let K(M(C)) be the set of all unitary elements u in M(C) such that
uau∗, u∗au ∈ A for any a ∈ A. Then K(M(C)) is a subgroup of the
group of all unitary elements in M(C).
Lemma 6.2. With the above notation, for any u, v ∈ K(M(C)),
[Au, C][Av, C] = [Auv, C]
in KerfC.
Proof. By the definition of the product in Pic(A, C),
[Au, C][Av, C] = [Au ⊗A Av, C ⊗C C]
in Pic(A, C). Hence we have only to show that
[Au ⊗A Av, C ⊗C C] = [Auv, C]
17
in Pic(A, C). Let Φ be the linear map from C ⊗C C to C defined by
Φ(x ⊗ y) = xy
for any x, y ∈ C. Then clearly Φ is a C − C-equivalence bimodule
isomorphism of C ⊗C C onto C. Also, for any x, y ∈ A,
Φ(xu ⊗ yv) = xuyv = xuyu∗uv ∈ Auv.
Hence by Lemma 3.2, we can see that ΦAu⊗AAv is an A−A-equivalence
bimodule isomorphism of Au ⊗A Av onto Auv. Therefore, we obtain
the conclusion.
(cid:3)
Lemma 6.3. With the above notation, let u ∈ K(M(C)). Then the
following conditions are equivalent:
(1) [Au, C] = [A, C] in Pic(A, C),
(2) There is a unitary element w ∈ M(A) such that w∗u ∈ C ′ ∩ M(C).
Proof. (1) ⇒ (2): By the proof of Lemma 6.1, [Au, C] = [XAd(u), YAd(u)]
in Pic(A, C). Hence since [XAd(u), YAd(u)] = [A, C] in Pic(A, C), by the
proof of Lemma 3.4, there is a unitary element w ∈ M(A) such that
Ad(u) = Ad(w) on C. Hence w∗u ∈ C ′ ∩ M(C) since M(A) ⊂ M(C).
(2) ⇒ (1): Since w∗u ∈ C ′ ∩ M(C), Ad(u) = Ad(w) on C. Thus
[Au, C] = [XAd(u), YAd(u)] = [XAd(w), YAd(w)] = [A, C]
in Pic(A, C) by the proofs of Lemmas 3.4 and 6.1.
(cid:3)
Let U(M(A)) be the group of all unitary elements in M(A). Then
U(M(A)) is a subgroup of K(M(C)) since M(A) ⊂ M(C) is a unital
inclusion.
Lemma 6.4. With the above notation, U(M(A)) is a normal subgroup
of K(M(C)).
Proof. Let u ∈ K(M(C)) and w ∈ U(M(A)). Let {wλ}λ∈Λ be a net in
A such that {wλ}λ∈Λ is strictly convergent to w ∈ M(A). Then since
uwλu∗ ∈ A for any λ ∈ Λ, uwu∗ ∈ U(M(A)). Therefore we obtain the
conclusion.
(cid:3)
Let K be the C ∗-algebra of all compact operators on a countably
inifinite dimensional Hilbert space. Let A ⊂ C be an inclusion of
C ∗-algebras with AC = C and As ⊂ C s the inclusion of C ∗-algebras
induced by A ⊂ C, where As = A ⊗ K and C s = C ⊗ K. Let S(As, C s)
be the subgroup of Pic(C s) defined by
S(As, C s) = {[Yα] ∈ Pic(C s) α ∈ Aut(As, C s)}.
Then by Proposition 3.5, S(As, C s) = ImfC s, where fC s is the homo-
morphism of Pic(As, C s) to Pic(C s) defined in the same way as in the
above. By Lemma 6.1
KerfC s = {[Asu , C s] ∈ Pic(As, C s) u ∈ K(M(C s))}.
18
Furthermore, we suppose that (C s)′ ∩ M(C s) = C1. Then by Lemmas
6.2, 6.3, KerfC s ∼= K(M(C s))/U(M(As)) as groups. Thus, we obtain
the following theorem:
Theorem 6.5. Let A ⊂ C be an inclusion of C ∗-algebras such that
A is σ-unital and AC = C. Let As ⊂ C s be the inclusion of C ∗-
algebras induced by A ⊂ C, where As = A ⊗ K and C s = C ⊗ K. We
suppose that (C s)′ ∩ M(C s) = C1M (C s). Then we have the following
exact sequence:
1 −→ K(M(C s))/U(M(As)) −→ Pic(As, C s)
f s
C−→ S(As, C s) −→ 1.
Again, we consider an inclusion of C ∗-algebras, A ⊂ C with AC = C.
Let fA be the homomorphism of Pic(A, C) to Pic(A) defined by
fA : Pic(A, C) → Pic(A) : [X, Y ] 7→ [X],
where Pic(A) is the ordinary Picard group of A. Let Aut0(A, C) be the
group of all automorphisms α of C with α = id on A. Then by easy
computations, Aut0(A, C) is a normal subgroup of Aut(A, C).
Lemma 6.6. With the above notation, let α ∈ Aut(A, C). Then the
following conditions are equivalent:
(1) [Xα, Yα] ∈ KerfA,
(2) There is a β ∈ Aut0(A, C) such that [Xα, Yα] = [Xβ, Yβ] in Pic(A, C).
Proof. (1) ⇒ (2): Since [Xα, Yα] ∈ KerfA, [Xα] = [A] in Pic(A). Hence
there is a unitary element u ∈ M(A) such that α = Ad(u) on A. Since
[XAd(u∗), YAd(u∗)] = [A, C] in Pic(A, C),
[Xα , Yα] = [XAd(u∗) , YAd(u∗)][Xα , Yα] = [XAd(u∗)◦α , YAd(u∗)◦α]
in Pic(A). Let β = Ad(u∗) ◦ α. Then β ∈ Aut0(A, C).
(2) ⇒ (1): Since Xβ = A, [Xβ , Yβ] ∈ KerfA. Hence [Xα, Yα] ∈ KerfA.
(cid:3)
Let π be the homomorphism of Aut(A, C) to Pic(A, C) defined by
π(α) = [Xα, Yα]
for any α ∈ Aut(A, C). Let AutI(A, C) be the subset of Aut(A, C)
defined by
AutI(A, C) = {α ∈ Aut(A, C) αA ∈ Int(A)}.
Then clearly AutI(A, C) is a subgroup of Aut(A, C). Also, AutI(A, C)
is a normal subgroup of Aut(A, C). Indeed, let α ∈ AutI(A, C). Then
there is a unitary element u ∈ M(A) such that α(a) = uau∗ for any
a ∈ A. Hence for any β ∈ Aut(A, C) and a ∈ A,
(β ◦ α ◦ β −1)(a) = β(uβ −1(a)u∗) = β(u)aβ(u∗)
since βA ∈ Aut(A), where β is an automorphism of M(C) induced by
β, whose restriction βM (A) is also an automorphism of M(A). Thus
Aut(A, C) is a normal subgroup of Aut(A, C).
19
Lemma 6.7. With the above notation, let α ∈ Aut(A, C). Then the
following conditions are equivalent:
(1) [Xα , Yα] ∈ KerfA,
(2) α ∈ AutI(A, C).
Proof. (1) ⇒ (2): By Lemma 6.6, there is a β ∈ Aut0(A, C) such
that [Xα , Yα] = [Xβ , Yβ] in Pic(A, C). Hence by Lemma 3.4, α ◦
β −1 ∈ Int(A, C). Thus there is a unitary element u ∈ M(A) such
that α ◦ β −1 = Ad(u), that is, α = Ad(u) ◦ β. Then for any a ∈ A,
α(a) = uβ(a)u∗ = uau∗. Hence α ∈ AutI(A, C).
(2) ⇒ (1): Since α ∈ AutI(A, C), there is a unitary element u ∈ M(A)
such that α(a) = uau∗ for any a ∈ A. Let β = Ad(u∗) ◦ α. Then β ∈
Aut(A, C) since M(A) ⊂ M(C) is a unital inclusion. Also, β(a) = a
for any a ∈ A. Hence β ∈ Aut0(A, C). Thus [Xα, Yα] = [Xβ, Yβ] in
Pic(A, C) by Lemma 3.4. Hence by Lemma 6.6, [Xα, Yα] ∈ KerfA. (cid:3)
Lemma 6.8. With the above notations, let α ∈ Aut(A, C). Then the
following conditions are equivalent:
(1) [Xα] = [A] in Pic(A),
(2) α ∈ AutI(A, C).
Proof. (1) ⇒ (2): By [2, Proposition 3.1], there is a unitary element
u ∈ M(A) such that αA = Ad(u). Hence α ∈ AutI(A, C).
(2) ⇒ (1): Since αA ∈ Int(A), we can see that [Xα] = [A] in Pic(A).
(cid:3)
Let As ⊂ C s be the inclusion of C ∗-algebras induced by the inclu-
sion of C ∗-algebras A ⊂ C with AC = C. We suppose that A is σ-
unital. Then by Proposition 3.5, the homomorphism π of Aut(As, C s)to
Pic(As, C s) is surjective. Hence
ImfAs = {[Xα] ∈ Pic(As) α ∈ Aut(As, C s)}
and by Lemma 6.6,
KerfAs = {[Xα, Yα] ∈ Pic(As, C s) α ∈ Aut0(As, C s)}.
Also, we have the exact sequence
1 −→ KerfAs −→ Pic(As, C s) −→ ImfAs −→ 1.
Since π is a surjective homomorphism of AutI(As, C s) onto KerfAs by
Lemma 6.7, we can see that
KerfAs ∼= AutI(As, C s)/Int(As, C s)
by Proposition 3.5. Also, fAs ◦ π is a surjective homomorphism of
Aut(As, C s) onto ImfAs, we can see that
ImfAs ∼= Aut(As, C s)/AutI(As, C s)
by Lemma 6.8. By the above discussions, we obtain the following
theorem:
20
Theorem 6.9. Let A ⊂ C be an inclusion of C ∗-algebras such that A
is σ-unital and AC = A. Let As ⊂ C s be the inclusion of C ∗-algebras
induced by A ⊂ C, where As = A ⊗ K and C s = C ⊗ K. Then we have
the following:
1 −→KerfAs −→ Pic(As, C s) −→ ImfAs −→ 1,
KerfAs ∼= AutI(As, C s)/Int(As, C s),
ImfAs ∼= Aut(As, C s)/AutI(As, C s).
7. Examples
In this section, we shall give some examples of the Picard groups for
unital inclusions of unital C ∗-algebras obtained by routine computa-
tions. Some other examples can be found in [11, Corollary 6.7]. We
begin this section with a trivial example.
Example 7.1. Let A be a C ∗-algebra. We regard A ⊂ A as an inclu-
sion of C ∗-algebras by the identity map. Then Pic(A, A) ∼= Pic(A).
Proof. Let be the map from Pic(A) to Pic(A, A) defined by
([X]) = [X, X]
for any [X] ∈ Pic(A). Then clearly is a monomorphism of Pic(A) to
Pic(A, A). Let [X, Y ] ∈ Pic(A, A). Since X is an A − A-equivalence
bimodule, A · X = X = X · A by [3, Proposition1.7]. Also, since Y is
an A − A-equivalence bimodule, by [3, Proposition 1.7],
Y = A · Y = AhX , Xi · Y = X · hX , Y iA = X · A = X.
Thus is surjective. Therefore, is an isomorphism of Pic(A) onto
Pic(A, A).
(cid:3)
Let A be a unital C ∗-algebra and we consider the unital inclusion
of unital C ∗-algebras C1 ⊂ A. Before we give the next example, we
prepare a lemma.
Lemma 7.2. Let A be a unital C ∗-algebra satisfying the following se-
quence
1 −→ Int(A) −→ Aut(A) −→ Pic(A) −→ 1
is exact. Then we have the following exact sequence:
1 −→ KerfA −→ Pic(C1, A)
fA−→ Pic(A) −→ 1.
Furthermore,
KerfA = { [Cu, A] ∈ Pic(C1, A) u ∈ U(A)},
where U(A) is the group of all unitary elements in A.
21
Proof. By the assumptions, for any [X] ∈ Pic(A), there is an automor-
phism α of A such that [X] = [Xα] in Pic(A), where Xα is the A − A-
equivalence bimodule induced by α. Then C1 is a closed subspace of
Xα and by easy computations, we can see that [C1, Xα] ∈ Pic(C1, A)
and that fA([C1, Xα]) = [Xα]. Thus fA is surjective. Next, we show
that
KerfA = { [Cu, A] ∈ Pic(C1, A) u ∈ U(A)}.
Let [X, Y ] ∈ KerfA. Then [Y ] = [A] in Pic(A). Hence there is an
A−A-equivalence bimodule isomorphism θ of Y onto A. Let Z = θ(X).
Then [Z, A] ∈ Pic(C1, A) and [Z, A] = [X, Y ] in Pic(C1, A) by easy
computations. Also, since Pic(C) = {[C]}, Z is a vector space over C
of dimension 1. Thus there is a unitary element u in A with u = 1
such that Z = Cu. We note that
Chu, ui = Ahu, ui = uu∗ ∈ C,
hu, uiC = hu, uiA = u∗u ∈ C.
Since uu∗, u∗u are positive numbers and u = 1, u is a unitary element
in A. Thus
KerfA ⊂ { [Cu, A] ∈ Pic(C1, A) u ∈ U(A)}.
Furthermore, by Lemma 6.1 we can see that [Cu, A] ∈ KerfA for any
u ∈ U(A) . Hence
KerfA = { [Cu, A] ∈ Pic(C1, A) u ∈ U(A)}.
(cid:3)
Let be the map from Pic(A) to Pic(C1, A) defined by
([Xα]) = [C1, Xα]
for any α ∈ Aut(A). Then by the proof of Lemma 7.2, is a homomor-
phism of Pic(A) to Pic(C1, A) with fA ◦ = id on Pic(A). Also, let κ
be the map from U(A) to KerfA defined by
κ(u) = [Cu, A]
for any u ∈ U(A). Then by Lemma 6.3, Kerκ = U(A ∩ A′) and κ is
surjective. Hence we obtain the following example.
Example 7.3. Let A be a unital C ∗-algebra satisfying the following
sequence
1 −→ Int(A) −→ Aut(A) −→ Pic(A) −→ 1
is exact. Then Pic(C1, A) is isomorphic to a semidirect product group
of U(A)/U(A ∩ A′) by Pic(A), that is,
Pic(C1, A) ∼= U(A)/U(A ∩ A′) ⋊s Pic(A).
We go to the next example. Let A be a unital C ∗-algebra satisfying
the following sequence
1 −→ Int(A) −→ Aut(A) −→ Pic(A) −→ 1
22
is exact. Let n be any positive integer with n ≥ 2. We consider the
unital inclusion of unital C ∗-algebras
where In is the unit element in Mn(A).
a ∈ A 7→ a ⊗ In ∈ Mn(A),
Lemma 7.4. With the above notation, the homomorphism fA of
Pic(A, Mn(A)) to Pic(A) is surjective.
Proof. Let X be any A − A-equivalence bimodule. Then by the as-
sumptions, there is an automorphism α of A such that [X] = [Xα] in
Pic(A), where Xα is the A−A-equivalence bimodule induced by α. Let
Xα⊗id be the Mn(A) − Mn(A)-equivalence bimodule induced by α ⊗ id.
Then [Xα, Xα⊗id] ∈ Pic(A, Mn(A)) and
fA([Xα, Xα⊗id]) = [Xα] = [X]
in Pic(A). Thus fA is surjective.
(cid:3)
Let be the map from Pic(A) to Pic(A, Mn(A)) defined by
([Xα]) = [Xα, Xα⊗id]
for any α ∈ Aut(A). Then is clearly a homomorphism of Pic(A) to
Pic(A, Mn(A)) with fA ◦ = id on Pic(A). Thus Pic(A, Mn(A)) is a
semidirect group of KerfA by Pic(A) by Lemma 7.4.
We compute KerfA. Let [X, Y ] ∈ KerfA. Since fA([X, Y ]) = [X] in
Pic(A), there is an A − A-equivalence bimodule isomorphism θ of A
onto X. We prepare the following lemma.
Lemma 7.5. Let A ⊂ C be a unital inclusion of unital C ∗-algebras.
Let (X, Y ) ∈ Equi(A, C) satisfying that there is an A − A-equivalence
bimodule isomorphism θ of A onto X. Then there is a β ∈ Aut0(A, C)
such that
[X, Y ] = [A, Yβ]
in Pic(A, C), where Yβ is the C − C-equivalence bimodule induced by
β.
Proof. By the discussions in [13, Section 2] and Rieffel [16, Proposition
2.1], there is an automorphism β of C defined by
for any c ∈ C since
β(c) = Chθ(1) · c , θ(1)i
Chθ(1) , θ(1)i = Ahθ(1) , θ(1)i = Ah1 , 1i = 1
and also we have hθ(1) , θ(1)iC = 1, where 1 is the unit element of A.
Then for any a ∈ A
β(a) = Chθ(1) · a , θ(1)i = Chθ(a) , θ(1)i = Ahθ(a), θ(1)i = Aha, 1i = a.
Hence β ∈ Aut0(A, C). Let ρ be the linear map from Y to Yβ defined
by
ρ(y) = Chy , θ(1)i
23
for any y ∈ Y . For any c ∈ C, c · θ(1) ∈ Y and
ρ(c · θ(1)) = Chc · θ(1) , θ(1)i = c Chθ(1) , θ(1)i = c.
Hence ρ is surjective. Also, for any y, z ∈ Y ,
Chρ(y) , ρ(z)i = Ch Chy , θ(1)i , Chz , θ(1)ii = Chy , θ(1)i Chθ(1) , zi
= Ch Chy , θ(1)i · θ(1) , zi = Chy · hθ(1) , θ(1)iC , zi
= Chy · h1 , 1iA , zi = Chy , zi,
hρ(y) , ρ(z)iC = h Chy , θ(1)i , Chz , θ(1)iiC
= β −1( Chθ(1) , yi Chz , θ(1)i)
= β −1( Ch Chθ(1) , yi · z , θ(1)i)
= β −1( Chθ(1) · hy , ziC , θ(1)i)
= (β −1 ◦ β)(hy , ziC) = hy , ziC.
Hence ρ is a C − C-equivalence bimodule isomorphism of Y onto Yβ.
Furthermore, for any a ∈ A,
ρ(θ(a)) = Chθ(a) , θ(1)i = Ahθ(a) , θ(1)i = Aha , 1i = a.
Therefore [X, Y ] = [A, Yβ] in Pic(A, C).
(cid:3)
We return to the unital inclusion of unital C ∗-algebras a ∈ A 7→
a ⊗ In ∈ Mn(A). By Lemma 7.5,
KerfA = {[A, Yβ] ∈ Pic(A, Mn(A)) β ∈ Aut0(A, Mn(A))}.
By Lemma 3.4
Ker π ∩ Aut0(A, Mn(A))
= {Ad(u) ∈ Aut0(A, Mn(A)) u is a unitary element in A ∩ A′},
where π is the homomorphism of Aut(A, Mn(A)) to Pic(A, Mn(A))
defined in Section 3. But Ker π ∩ Aut0(A, Mn(A)) = {1} since (u ⊗
In)a = a(u ⊗ In) for any unitary element u ∈ A ∩ A′, a ∈ Mn(A). Thus
we obtain the following example.
Example 7.6. Let A be a unital C ∗-algebra satisfying the following
sequence
1 −→ Int(A) −→ Aut(A) −→ Pic(A) −→ 1
is exact. Let n be a positive number with n ≥ 2. Then Pic(A, Mn(A)) is
isomorphic to a semidirect product group of Aut0(A, Mn(A)) by Pic(A),
that is,
Pic(A, Mn(A)) ∼= Aut0(A, Mn(A)) ⋊s Pic(A).
24
References
[1] L. G. Brown, Stable isomorphism of hereditary subalgebra of C ∗-algebras, Pa-
cific J. Math., 71 (1977), 335 -- 348.
[2] L. G. Brown, P. Green and M. A. Rieffel, Stable isomorphism and strong Morita
equivalence of C ∗-algebras, Pacific J. Math., 71 (1977), 349 -- 363.
[3] L.G. Brown, J. Mingo and N-T. Shen, Quasi-multipliers and embeddings of
Hilbert C ∗-bimodules, Canad. J. Math., 46 (1994), 1150-1174.
[4] F. Combes, Crossed products and Morita equivalence, Proc. London Math.
Soc., 49 (1984), 289 -- 306.
[5] R. E. Curto, P. S. Muhly and D. P. Williams, Cross products of strong Morita
equivalent C ∗-algebras, Proc. Amer. Math. Soc., 90 (1984), 528 -- 530.
[6] K. K. Jensen and K. Thomsen, Elements of KK-theory, Birkhauser, 1991.
[7] M. Izumi, Inclusions of simple C ∗-algebras, J. reine angew. Math., 547 (2002),
97 -- 138.
[8] T. Kajiwara and Y. Watatani, Jones index theory by Hilbert C ∗-bimodules and
K-Thorey, Trans. Amer. Math. Soc., 352 (2000), 3429 -- 3472.
[9] T. Kajiwara and Y. Watatani, Crossed products of Hilbert C ∗-bimodules by
countable discrete groups, Proc. Amer. Math. Soc., 126 (1998), 841 -- 851.
[10] K. Kodaka, Equivariant Picard groups of C ∗-algebras with finite dimensional
C ∗-Hopf algebra coactions, Rocky Mountain J. Math., 47 (2017), 1565-1615.
[11] K. Kodaka, The generalized Picard groups for finite dimensional C ∗-Hopf al-
gebra coactions on unital C ∗-algebras, preprint, arXiv:1805.08358.
[12] K. Kodaka and T. Teruya, The strong Morita equivalence for coactions of a
finite dimensional C ∗-Hopf algebra on unital C ∗-algebras, Studia Math., 228
(2015), 259 -- 294.
[13] K. Kodaka and T. Teruya, The strong Morita equivalence for inclusions of
C ∗-algebras and conditional expectations for equivalence bimodules, J. Aust.
Math. Soc., 105 (2018), 103 -- 144.
[14] H. Osaka, K. Kodaka and T. Teruya, The Rohlin property for inclusions of
C ∗-algebras with a finite Watatani index, Operator structures and dynamical
systems, 177 -- 195, Contemp Math., 503 Amer Math. Soc., Providence, RI,
2009.
[15] I. Raeburn and D. P. Williams, Morita equivalence and continuous -trace C ∗-
algebras, Mathematical Surveys and Monographs, 60, Amer. Math. Soc., 1998.
[16] M. A. Rieffel, C ∗-algebras associated with irrational rotations, Pacific J. Math.,
93 (1981), 415 -- 429.
[17] Y. Watatani, Index for C ∗-subalgebras, Mem. Amer. Math. Soc., 424, Amer.
Math. Soc., 1990.
Department of Mathematical Sciences, Faculty of Science, Ryukyu
University, Nishihara-cho, Okinawa, 903-0213, Japan
E-mail address: [email protected]
25
|
0906.2765 | 2 | 0906 | 2010-05-27T08:17:42 | Group measure space decomposition of II_1 factors and W*-superrigidity | [
"math.OA",
"math.DS",
"math.GR"
] | We prove a "unique crossed product decomposition" result for group measure space II_1 factors arising from arbitrary free ergodic probability measure preserving (p.m.p.) actions of groups \Gamma in a fairly large family G, which contains all free products of a Kazhdan group and a non-trivial group, as well as certain amalgamated free products over an amenable subgroup. We deduce that if T_n denotes the group of upper triangular matrices in PSL(n,Z), then any free, mixing p.m.p. action of the amalgamated free product of PSL(n,Z) with itself over T_n, is W*-superrigid, i.e. any isomorphism between L^\infty(X) \rtimes \Gamma and an arbitrary group measure space factor L^\infty(Y) \rtimes \Lambda, comes from a conjugacy of the actions. We also prove that for many groups \Gamma in the family G, the Bernoulli actions of \Gamma are W*-superrigid. | math.OA | math |
Group measure space decomposition of II1 factors
and W*-superrigidity
by Sorin Popa(1)(2) and Stefaan Vaes(3)(4)
Abstract
We prove a "unique crossed product decomposition" result for group measure space II1 factors
L∞(X) ⋊ Γ arising from arbitrary free ergodic probability measure preserving (p.m.p.) actions
of groups Γ in a fairly large family G, which contains all free products of a Kazhdan group and a
non-trivial group, as well as certain amalgamated free products over an amenable subgroup. We
deduce that if Tn denotes the group of upper triangular matrices in PSL(n, Z), then any free,
mixing p.m.p. action of Γ = PSL(n, Z) ∗Tn PSL(n, Z) is W∗-superrigid, i.e. any isomorphism
between L∞(X) ⋊ Γ and an arbitrary group measure space factor L∞(Y ) ⋊ Λ, comes from a
conjugacy of the actions. We also prove that for many groups Γ in the family G, the Bernoulli
actions of Γ are W∗-superrigid.
1
Introduction
Rigidity results have by now appeared in many areas of mathematics, and in several forms. The
most frequently encountered is when two mathematical objects with rich structure that are known to
be equivalent in some "weak sense", which ignores part of the structure, are shown to be isomorphic
as objects with the full structure. In the best of cases, such a result will also show that morphisms
which are equivalences in the weak sense are equivalent to isomorphisms in the stronger category,
thus leading to complete classification results and calculation of invariants.
Von Neumann algebras (also called W∗-algebras) provide a most natural framework for rigidity. In
fact, such phenomena are at the very core of this subject, relating it at the outset with group theory
and ergodic theory. This is due to the Murray and von Neumann classical group measure space
construction, which associates to a free ergodic measure preserving action Γ y X, of a countable
group Γ on a probability space (X, µ), a von Neumann algebra (called II1 factor) L∞(X) ⋊ Γ,
through a crossed product type construction [MvN36]. The study of these objects in terms of their
"initial data" Γ y X has been a central theme of the subject since the early 1940s. It soon led
to a new area in ergodic theory, studying group actions up to orbit equivalence (OE), i.e. up to
isomorphism of probability spaces carrying the orbits of actions onto each other, since an OE of
actions Γ y X, Λ y Y has been shown in [Si55] to be "the same as" an algebra isomorphism
L∞(X) ⋊ Γ ≃ L∞(Y ) ⋊ Λ taking the group measure space Cartan subalgebras L∞(X), L∞(Y ) onto
each other.
Thus, W∗-equivalence (or von Neumann equivalence) of group actions, requiring isomorphism of
their group measure space algebras, is weaker than OE. Since there are examples of non-OE actions
whose group measure space factors are isomorphic [CJ82, OP08], it is in general strictly weaker. In
(1)Partially supported by NSF Grant DMS-0601082
(2)Mathematics Department; University of California at Los Angeles, CA 90095-1555 (United States).
E-mail: [email protected]
(3)Partially supported by ERC Starting Grant VNALG-200749, Research Programme G.0231.07 of the Research
Foundation -- Flanders (FWO) and K.U.Leuven BOF research grant OT/08/032.
(4)Department of Mathematics; K.U.Leuven; Celestijnenlaan 200B; B -- 3001 Leuven (Belgium).
E-mail: [email protected]
1
turn, it has been known since [MvN43, Dy59] that OE is much weaker than classical conjugacy (or
isomorphism), which for free actions Γ y X, Λ y Y requires isomorphism of probability spaces
∆ : (X, µ) ≃ (Y, ν) satisfying ∆Γ∆−1 = Λ (so in particular Γ ≃ Λ). Rigidity in this context occurs
whenever one can establish that W∗- or OE-equivalence of certain group actions Γ y X, Λ y Y ,
forces the groups, or the actions, to share some common properties. The ideal such result, labeled
W∗- (respectively OE-) superrigidity, recovers the isomorphism class of Γ y X, from its W∗-class
(resp. OE-class).
W∗- and OE-rigidity can only occur for non-amenable groups, since by classical results of Connes
[Co76], all II1 factors L∞(X)⋊Γ with Γ amenable are mutually isomorphic and by [OW80, CFW81]
they are undistinguishable under OE as well. But the non-amenable case is extremely complex and
although signs of rigidity where detected early on [MvN43, Dy63, McDu70, Co75], for many years
progress has been slow, despite several breakthrough discoveries in the 1980s [Co80a, Zi80, CJ85,
CH89]. This changed dramatically over the last decade, with the advent of a variety of striking
rigidity results, in both group measure space II1 factors and OE ergodic theory:
[Fu99a, Fu99b,
Ga00, Ga01, Po01, MS02, Hj02, Oz03, Po03, Po04, Po05, HK05, IPP05, Po06a, PV06, Pe06, Ki06,
Va07, OP07, Io08, OP08, PV08a, PV08b, PV08c, CH08, Ki09].
Our purpose in this paper is to investigate the most "extreme" of the W∗-rigidity phenomena
mentioned above, i.e. W∗-superrigidity. Thus, we seek to find classes of group actions Γ y X with
the property that any isomorphism between L∞(X) ⋊ Γ and any other group measure space factor
L∞(Y ) ⋊ Λ, arising from an arbitrary free ergodic p.m.p. action Λ y Y , comes from a conjugacy
of the actions Γ y X, Λ y Y .
Note that W∗-superrigidity for an action Γ y X is equivalent to the "sum" between its OE-
superrigidity and the uniqueness, up to unitary conjugacy, of L∞(X) as a group measure space
Cartan subalgebra in L∞(X)⋊Γ. This makes W∗-superrigidity results extremely difficult to obtain,
since each one of these problems is notoriously hard. But while several large families of OE-
superrigid actions have been discovered over the last ten years [Fu99b, Po05, Po06a, Ki06, Io08,
Ki09], unique Cartan decomposition proved to be much more challenging to establish, and the
only existing results cover very particular group actions. Thus, a first such result, obtained by
Ozawa and the first named author in [OP07], shows that given any profinite action Γ y X, of a
product of free groups Γ = Fn1 × · · · × Fnk , with k ≥ 1, 2 ≤ ni ≤ ∞, any Cartan subalgebra of
M = L∞(X) ⋊ Γ (i.e. any maximal abelian subalgebra whose normalizer generates M ), is unitary
conjugate to L∞(X). A similar result, covering a more general class of groups Γ, was then proved
in [OP08]. More recently, Peterson showed in [Pe09] that factors arising from profinite actions of
non-trivial free products Γ = Γ1 ∗ Γ2, with at least one of the Γi not having the Haagerup property,
have unique group measure space Cartan subalgebra, up to unitary conjugacy. But so far, none of
these group actions could be shown to be OE-superrigid. Nevertheless, an intricate combination
of results in [Io08, OP08, Pe09] were used to prove the existence of virtually W∗-superrigid group
actions Γ y X in [Pe09], by a Baire category argument (following [Fu99a], virtual means that the
ensuing conjugacy of Γ y X and the target actions Λ y Y is up to finite index subgroups of Γ, Λ).
In this paper, we establish a very general unique Cartan decomposition result, which allows us to
obtain a wide range of W∗-superrigid group actions. Thus, we first prove the uniqueness, up to
unitary conjugacy, of the group measure space Cartan subalgebra in the II1 factor given by an
arbitrary free ergodic p.m.p. action of any group Γ belonging to a large family G of amalgamated
free product groups. By combining this with Kida's OE-superrigidity in [Ki09], we deduce that
any free, mixing p.m.p. action of Γ = PSL(n, Z) ∗Tn PSL(n, Z) is W∗-superrigid. In combination
with [Po05, Po06a], we prove that for many groups Γ in the family G, the Bernoulli actions of Γ
2
are W∗-superrigid. In combination with Gaboriau's work [Ga00] on cost, we find new groups Γ for
which all group measure space II1 factors L∞(X) ⋊ Γ have trivial fundamental group.
1.1 Statements of main results
More precisely, our family G contains all non-trivial free products Γ = Γ1 ∗ Γ2 with Γ1 satisfying
one of the following rigidity properties: Γ1 contains a non-amenable subgroup with the relative
property (T) or Γ1 contains two non-amenable commuting subgroups. The family G also contains
certain amalgamated free products Γ1 ∗Σ Γ2 over amenable subgroups Σ, see Definition 5.1.
Our results can be summarized as follows.
Theorem 1.1 (See Theorem 5.2). Let Γ be a group in the family G and Γ y (X, µ) an arbitrary
free ergodic p.m.p. action. Denote M = L∞(X) ⋊ Γ. Whenever Λ y (Y, η) is a free ergodic p.m.p.
action such that M = L∞(Y ) ⋊ Λ, there exists a unitary u ∈ M such that L∞(Y ) = u L∞(X)u∗.
We mention that the most general version of the above theorem (see Theorem 5.2) allows to handle
amplifications of the group measure space factors M as well.
Theorem 1.2 (See Theorem 6.2). Let n ≥ 3 and denote by Tn the subgroup of upper triangular
matrices in PSL(n, Z). Put Γ = PSL(n, Z) ∗Tn PSL(n, Z). Then, every free p.m.p. mixing action of
Γ is W∗-superrigid.
Whenever Γ is an infinite group and (X0, µ0) a non-trivial probability space, denote by Γ y
(X0, µ0)Γ the Bernoulli action of Γ with base space (X0, µ0), given by (g · x)h = xg−1h for all
g, h ∈ Γ and x ∈ X Γ
0 .
Theorem 1.3 (See Theorem 6.7). The Bernoulli action Γ y (X0, µ0)Γ of all of the following
groups is W∗-superrigid.
• Γ = Γ1 ∗Σ Γ2 with the following assumptions: Γ1 has property (T), Σ is an infinite, amenable,
proper normal subgroup of Γ2 and there exist g1, . . . , gk ∈ Γ1 such that Tk
For instance, we can take Γ = PSL(n, Z) ∗Σ (Σ × Λ), where Σ < Tn is an infinite subgroup of
the upper triangular matrices and Λ is an arbitrary non-trivial group.
i=1 giΣg−1
i = {e}.
• Γ = (H × H) ∗Σ Γ2 where H is a finitely generated non-amenable group with trivial center, Σ
is an infinite amenable subgroup of H that we embed diagonally in H × H and Σ is a proper
normal subgroup of Γ2.
We in fact obtain a more general version of this result as Theorem 6.7, which covers generalized
Bernoulli actions, Gaussian actions and certain co-induced actions (see Examples 6.8, 6.9 and 6.10).
Our methods also provide the following new examples of II1 factors which cannot be written as
group measure space factors.
Theorem 1.4. Let Γ ∈ G and assume that Γ is ICC. Let Γ y (X, µ) be an arbitrary ergodic p.m.p.
action and put M = L∞(X) ⋊ Γ. Then, M is a II1 factor. If, for some t > 0, the II1 factor M t
admits a group measure space decomposition, then the action Γ y (X, µ) must be free. Thus, L(Γ)
and all the factors of the form L∞(X) ⋊ Γ corresponding to non-free actions Γ y X, do not admit
a group measure space decomposition.
3
1.2 Rigidity of bimodules
Another typical rigidity paradigm encountered in mathematics is when certain invariants of math-
ematical objects which are supposed to take values in a certain range, are shown to take values in a
much smaller subset. Group measure space II1 factors provide a natural framework for this type of
rigidity as well, due to Murray and von Neumann's continuous dimension and a related invariant
for II1 factors M : the fundamental group F(M ). This is defined as the set of ratios τ (p)/τ (q) ∈ R+,
over all projections p, q ∈ M with pM p ≃ qM q, where τ denotes the (unique) normalized trace
(=dimension function) on M . Equivalently, F(M ) = {t > 0 M t ≃ M }. Thus, since the range of
the dimension function is all [0, 1], the group F(M ) seems to always be equal to R+. Supporting
evidence comes from the case M = L∞(X) ⋊ Γ with Γ amenable, when this is indeed the case (cf.
[MvN43]). So it came as a striking surprise when Connes showed that all factors L∞(X) ⋊ Γ with
Γ an ICC Kazhdan group and Γ y X ergodic, have countable fundamental group ([Co80a]).
The important progress in W∗-rigidity in recent years, led to the first actual computations of
fundamental groups F(M ) of group measure space factors M = L∞(X) ⋊ Γ: from the first such ex-
amples in [Po01], where F(M ) = 1, to examples of factors M with F(M ) any prescribed countable
subgroup of R+ [Po03] (see also [IPP05, Ho07]), and most recently examples with F(M ) uncount-
able, yet different from R+ [PV08a, PV08c]. We mention in this respect that all Bernoulli actions
Γ y (X, µ) appearing in Theorem 1.3 give rise to II1 factors L∞(X) ⋊ Γ with trivial fundamental
group (see Remark 6.6). In the same spirit, Corollary 5.3 provides new examples of groups Γ such
that L∞(X) ⋊ Γ has trivial fundamental group for all free ergodic p.m.p. actions Γ y (X, µ).
Another occurrence of the same type of rigidity paradigm is related to Jones' index for subfactors,
a numerical invariant for inclusions of II1 factors N ⊂ M which, like the fundamental group, is
defined with the help of the Murray-von Neumann continuous dimension. A priori, the range of
the index could well be all R+, but in his seminal work [Jo83], Jones proved that it is subject to
very surprising restrictions.
A unique feature of the II1 factor framework is that it allows a unifying approach to the two types of
rigidity phenomena (W∗-rigidity and restrictions on invariants), by considering finite index bimod-
ules between factors (as a generalization of isomorphism between factors and their amplifications).
Thus, by explicitly calculating all bimodules between factors in a certain class, one also obtains the
fundamental group of the corresponding factors, as well as all possible indices of its subfactors.
In the last section 7, we combine a generalization of Theorem 1.1 with the cocycle superrigidity
theorems of [Po05, Po06a] and techniques from [Va07], to give examples of group actions Γ y (X, µ)
such that the mere existence of a finite index bimodule between L∞(X) ⋊ Γ and L∞(Y ) ⋊ Λ for an
arbitrary free ergodic p.m.p. action Λ y (Y, η), implies that the groups Γ, Λ are virtually isomorphic
and their actions Γ y X, Λ y Y are virtually conjugate in a very precise sense (see Theorem 7.1
and Example 7.2). In particular, the fundamental group of any of these II1 factors L∞(X) ⋊ Γ is
trivial and the index of all their subfactors is an integer.
1.3 Comments on the proofs
As we mentioned before, the main difficulty in obtaining W∗-superrigidity lies in proving the
uniqueness of the group measure space Cartan decomposition.
Indeed, because once such a re-
sult is established, W∗-superrigidity can be derived from existing OE-superrigidity results. In our
case, 1.2 and 1.3 will follow from our uniqueness of the group measure space Cartan subalgebra
in Theorem 1.1 and the OE superrigidity theorems in [Po05, Po06a, Ki09]. By using intertwining
4
subalgebras techniques ([Po01, Po03]), in order to prove the uniqueness, up to unitary conjugacy,
of the Cartan subalgebra A = L∞(X) of a group measure space factor M = A ⋊ Γ, it is suffi-
cient to prove that given any other group measure space decomposition M = B ⋊ Λ, B = L∞(Y ),
there exists a B-A-bimodule H ⊂ L2(M ), which is finitely generated over A, a property that we
denote by B ≺M A. To get such bimodules, we use the deformation-rigidity theory introduced in
[Po01, Po03, Po04] (see [Po06b] for a survey), and in fact the whole array of subsequent develop-
ments in [IPP05, PV06, Po06a, Va07, Ho07, CH08], etc. But in order to "locate" the position of
the target Cartan subalgebra B, with respect to the initial (source) Cartan subalgebra A, through
these techniques, one needs some amount of rigidity for one of the group actions and a deformation
property for the other (like for example in 6.2 of [Po01], 7.1 of [Po04], 7.7 of [IPP05], 1.5 of [Po06a],
etc).
Since in W∗-superrigidity statements all assumptions must be on the side of the source Cartan
subalgebra A = L∞(X), it is thus crucial to show that either the deformation or the rigidity
properties of Γ y X automatically transfer to Λ y Y .
It is precisely the lack of satisfactory
"transfer" results that so far prevented from obtaining W∗-superrigidity results. We solve this
problem here by proving in Section 2 some very general "transfer of rigidity", from the source
to the target side. While the proofs of these results are quite subtle, let us give here a heuristic
explanation.
Assume that A ⋊ Γ = M = B ⋊ Λ and denote by (ug)g∈Γ, resp. (vs)s∈Λ, the canonical unitaries in
A ⋊ Γ, resp. B ⋊ Λ. Every element x ∈ M , has a Fourier expansion x =Pg∈Γ xgug with xg ∈ A
and we call the xg the Fourier coefficients of x w.r.t. {ug}. We similarly define Fourier coefficients
w.r.t. {vs}. We assume that Γ1 < Γ is a non-amenable subgroup with the relative property (T) and
try to transfer this rigidity property to some rigidity for Λ. More precisely, we show that for any
deformation φn of M (i.e. a sequence of c.p. maps on M tending pointwise in the Hilbert norm to
idM ), there exist a large n and an infinite subset {sk}k ⊂ Λ, such that φn(vsk ) ≈ vsk , ∀k, and such
that every Fourier coefficient of vsk w.r.t. {ug} tends to zero in k · k2 as k → ∞. We construct
as follows the set {sk}k. The functions ψn(s) = τ (φn(vs)v∗
s ) are positive definite and hence, define
the c.p. maps Ψn by Ψn(Ps bsvs) =Ps ψn(s)bsvs. Since Ψn → idM , the relative property (T) of
Γ1 < Γ ensures that Ψn(ug) ≈ ug uniformly in g ∈ Γ1. This forces ψn(s) ≈ 1 for many of the s ∈ Λ
in the support of the Fourier expansion of ug, g ∈ Γ1 w.r.t. {vs}. Among the s ∈ Λ with ψn(s) ≈ 1,
we can find a sequence sk such that the Fourier coefficients of vsk tend to zero as k → ∞, because
otherwise, it will follow that the ug, g ∈ Γ1 can roughly be intertwined into A, contradicting the
non-amenability of Γ1.
As it turns out, if we assume that Γ1 is freely complemented in Γ, i.e. Γ = Γ1 ∗ Γ2, then the
"tiny" initial information about the group Λ provided by the transfer of rigidity, is enough to prove
that B ≺M A. To see this, we first notice that if we apply the above transfer result to the word
length deformation mρ(Pg agug) =Pg ρgagug, as ρ → 1, then, for ρ close enough to 1, we have
mρ(vsk ) ≈ vsk uniformly in k. This implies that in the Fourier decomposition with respect to {ug},
all vsk are almost supported by words g ∈ Γ = Γ1 ∗ Γ2 of length uniformly bounded by some K.
On the other hand, since the Fourier coefficients of vsk w.r.t. {ug} tend to 0 in k · k2, the support
of the Fourier expansion of the vsk lies, as k → ∞, essentially outside any given finite subset of
Γ. If, by contradiction B 6≺M A, results from [Po03, IPP05] provide a unitary w ∈ B such that
the Fourier expansion of w is essentially supported by a set of words g ∈ Γ of length much larger
than K (g ≥ 2K will do). As we will explain now, this implies that vsk wv∗
sk are
almost orthogonal, contradicting the abelianess of B.
sk w∗ and w∗vsk wv∗
Indeed, first assume for simplicity that all vsk lie in the span of Aug, g ≤ K, with the support of
5
the Fourier expansion of vsk w.r.t. {ug} tending to infinity in Γ. Similarly, assume that w exactly
lies in the span of Aug, g ≥ 2K. Then, one concludes that all g ∈ Γ in the support of w∗vsk wv∗
sk
eventually have their first K letters in a fixed finite set independent of k, while all g ∈ Γ in the
sk w∗, have their first K letters eventually (as k → ∞) outside any fixed finite set.
support of vsk wv∗
In reality, we can only approximate in k · k2 and uniformly in k, the unitaries vsk by elements v′
sk
with such good properties. We similarly approximate w by w′. But since our reasoning involves
products of 4 elements, the Hilbert norm estimates cannot be handled unless one can control the
uniform norms of w − w′, vsk − v′
sk . We handle this problem through repeated "trimming" of
elements, via Herz-Schur multiplier techniques.
2 Preliminaries
If Γ y (A, τ ) is a trace preserving action of a countable group Γ, we denote by A ⋊ Γ the crossed
product von Neumann algebra, which is the unique tracial von Neumann algebra generated by A
and the group of unitaries (ug)g∈Γ satisfying
ugau∗
g = σg(a)
for all g ∈ Γ, a ∈ A and τ (aug) =(τ (a)
0
if g = e ,
if g 6= e .
When Γ y (X, µ) is a free ergodic p.m.p. action, the crossed product L∞(X) ⋊ Γ is called the
group measure space II1 factor associated with Γ y (X, µ). Then, L∞(X) is a Cartan subalgebra
of L∞(X) ⋊ Γ, called a group measure space Cartan subalgebra.
We denote by L(Γ) the group von Neumann algebra of a countable group Γ.
Recall that two free ergodic p.m.p. actions Γ y (X, µ) and Λ y (Y, η) are called
• conjugate, if there exists an isomorphism ∆ : X → Y of probability spaces and an isomorphism
δ : Γ → Λ of groups such that ∆(g · x) = δ(g) · ∆(x) almost everywhere,
• orbit equivalent, if there exists an isomorphism ∆ : X → Y of probability spaces such that
∆(Γ · x) = Λ · ∆(x) for almost all x ∈ X,
• W∗-equivalent (or von Neumann equivalent), if L∞(X) ⋊ Γ ∼= L∞(Y ) ⋊ Λ.
If two actions are conjugate, they are obviously orbit equivalent. On the other hand, two actions
are orbit equivalent if and only if there exists an isomorphism L∞(X) ⋊ Γ ∼= L∞(Y ) ⋊ Λ sending
L∞(X) onto L∞(Y ), see [Si55, FM77].
2.1 Bimodules and weak containment
Let M, N be tracial von Neumann algebras. An M -N -bimodule MHN is a Hilbert space H equipped
with a normal representation π of M and a normal anti-representation π′ of N such that π(M ) and
π′(N ) commute.
Given the bimodules MHN and NKP , one can define the Connes tensor product H ⊗N K which is
an M -P -bimodule, see [Co94, V.Appendix B].
An M -N -bimodule can be seen as well as a representation of the C∗-algebra M ⊗binor N op. If MH1
N
N are M -N -bimodules, we say that H1 is weakly contained in H2 if the corresponding
and MH2
6
representations π1, π2 of M ⊗binor N op satisfy Ker π1 ⊃ Ker π2. Weak containment behaves well
with respect to the Connes tensor product: if MH1
N, then K ⊗M H1
is weakly contained in K ⊗M H2 for every P -M -bimodule K (see e.g. [An95, Lemma 1.7]). A similar
statement holds for tensor products on the right.
We call ML2(M )M the trivial M -M -bimodule and define the coarse M -M -bimodule as the Hilbert
space L2(M ) ⊗ L2(M ) equipped with the bimodule structure a · ξ · b = (a ⊗ 1)ξ(1 ⊗ b). A finite
von Neumann algebra M is injective if the trivial M -M -bimodule is weakly contained in the coarse
M -M -bimodule.
N is weakly contained in MH2
The first type of bimodule that we encounter in this article, is the following. Let Γ y (Q, τ ) be a
trace preserving action and put M = Q ⋊ Γ. Whenever π : Γ → U (K) is a unitary representation,
define the Hilbert space Hπ = L2(M ) ⊗ K, with bimodule action given by
(aug) · ξ · (buh) = (aug ⊗ π(g))ξbuh
for all a, b ∈ Q, g, h ∈ Γ .
If the unitary representation π is weakly contained in the unitary representation ρ, then Hπ is weakly
contained in Hρ. If π = λ is the regular representation of Γ on ℓ2(Γ), then Hλ ∼= L2(M ) ⊗Q L2(M ).
When Q is injective, QL2(Q)Q is weakly contained in Q(L2(Q) ⊗ L2(Q))Q and hence,
L2(M ) ⊗Q L2(M ) ∼= L2(M ) ⊗Q L2(Q) ⊗Q L2(M )
is weakly contained in L2(M ) ⊗Q (L2(Q) ⊗ L2(Q)) ⊗Q L2(M ) = L2(M ) ⊗ L2(M ). So, for Q injective,
Hλ is weakly contained in the coarse M -M -bimodule. Finally, if Q is injective and if the unitary
representation π is weakly contained in the regular representation, then Hπ is weakly contained in
the coarse M -M -bimodule.
A next type of bimodule arises from Jones' basic construction [Jo83]. Let M be a tracial von
Neumann algebra with von Neumann subalgebra P . Denote by hM, eP i the von Neumann algebra
acting on L2(M ) generated by M and by the orthogonal projection eP of L2(M ) onto L2(P ).
Equivalently, hM, eP i is the commutant of the right P -action on L2(M ). Given a tracial state
τ on M , the von Neumann algebra hM, eP i carries a natural normal semi-finite faithful trace Tr
characterized by
Tr(aeP b) = τ (ab)
for all a, b ∈ M .
In particular, we can write the Hilbert space L2(hM, eP i) and consider it as an M -M -bimodule.
Then,
ML2(hM, eP i)M
∼= M(L2(M ) ⊗P L2(M ))M .
Again, if P is injective, it follows that ML2(hM, eP i)M is weakly contained in the coarse M -M -
bimodule.
2.2 Relative property (T) for an inclusion of finite von Neumann algebras
We recall from [Po01, Proposition 4.1] the following definition of relative property (T) for an
inclusion of tracial von Neumann algebras.
Definition 2.1. Let (M, τ ) be a tracial von Neumann algebra and let P ⊂ M be a von Neumann
subalgebra. The inclusion P ⊂ M is said to have the relative property (T) if the following property
holds: for every ε > 0, there exists a finite subset J ⊂ M and a δ > 0 such that whenever MHM is
an M -M -bimodule admitting a unit vector ξ with the properties
7
• ka · ξ − ξ · ak < δ for all a ∈ J ,
• hξ, a · ξi − τ (a) < δ and hξ, ξ · ai − τ (a) < δ for all a in the unit ball of M ,
there exists a vector ξ0 ∈ H satisfying kξ − ξ0k < ε and a · ξ0 = ξ0 · a for all a ∈ P .
Note that if Γ0 is a subgroup of the countable group Γ, then the inclusion L(Γ0) ⊂ L(Γ) has
relative property (T) if and only if Γ0 < Γ has the relative property (T) of Kazhdan-Margulis
[Po01, Proposition 5.1].
A normal completely positive map ϕ : M → M is said to be subunital if ϕ(1) ≤ 1 and subtracial if
τ ◦ ϕ ≤ ϕ. Let P ⊂ M be an inclusion of tracial von Neumann algebras. Relative property (T) then
has the following equivalent characterization: whenever ϕn : M → M is a sequence of subunital
and subtracial normal completely positive maps satisfying kx − ϕn(x)k2 → 0 for all x ∈ M , we have
that kx − ϕn(x)k2 → 0 uniformly on the unit ball of P .
2.3
Intertwining by bimodules
To fix notations, we briefly recall the intertwining-by-bimodules technique from [Po03, Section 2]
(see also [Va06, Appendix C]).
Let (M, τ ) be a tracial von Neumann algebra and assume that A, B ⊂ Mn(C) ⊗ M are possibly
non-unital von Neumann subalgebras. Denote their respective units by 1A and 1B. Then, the
following two conditions are equivalent.
• 1A(Mn(C) ⊗ L2(M ))1B admits an A-B-subbimodule that is finitely generated as a right B-
module.
• There is no sequence of unitaries un ∈ U (A) satisfying kEB(xuny∗)k2 → 0 for all x, y ∈
1B(Mn(C) ⊗ M )1A.
If one of these equivalent conditions hold, we write A ≺M B. Otherwise, we write A 6≺M B.
When M is a II1 factor and A, B ⊂ M are Cartan subalgebras, then A ≺M B if and only if there
exists a unitary u ∈ U (M ) such that A = uBu∗, see [Po01, Theorem A.1] (see also [Va06, Theorem
C.3]).
2.4 Cocycle superrigidity
Let Γ y (X, µ) be a p.m.p. action. If L is a Polish group, an L-valued 1-cocycle is a measurable
map ω : Γ× X → L satisfying ω(gh, x) = ω(g, h·x)ω(h, x) for all g, h ∈ Γ and almost all x ∈ X. We
say that Γ y (X, µ) is L-cocycle superrigid if every 1-cocycle with values in L is cohomologous to a
group morphism Γ → L. More precisely, this means that there exists a measurable map ϕ : X → L
and a group morphism δ : Γ → L satisfying ω(g, x) = ϕ(g · x)δ(g)ϕ(x)−1 for all g ∈ Γ and almost
all x ∈ X.
We say that Γ y (X, µ) is Ufin-cocycle superrigid if it is L-cocycle superrigid for all L in the class
Ufin of Polish groups that can be realized as the closed subgroup of the unitary group of a II1
factor with separable predual. Note that Ufin contains all countable groups and all compact second
countable groups. Using the distance given by the k · k2-norm every group in Ufin admits a separable
complete bi-invariant metric implementing the topology.
8
3 Two transfer lemmas
We want to establish W∗-superrigidity for certain known group actions Γ y (X, µ). Starting
from a "mysterious" alternative group measure space decomposition L∞(Y ) ⋊ Λ of the given factor
L∞(X) ⋊ Γ, we need to transfer certain properties of Γ y X to properties of the unknown group
Λ or of its unknown action Λ y Y . For our purposes, this transfer of properties is achieved in
Lemmas 3.1 and 3.2.
Lemma 3.1. Let M be a II1 factor with the following deformation property. We are given a
sequence ϕn of subunital, subtracial, normal, completely positive maps from M to M such that
kx − ϕn(x)k2 → 0 for all x ∈ M .
Let P ⊂ M be a von Neumann subalgebra and assume that P is injective.
Assume that the countable group Λ acts trace preservingly on the injective von Neumann algebra
Q and suppose that we have identified Q ⋊ Λ with pM p for some projection p ∈ M . We denote by
(vs)s∈Λ the canonical unitaries in Q ⋊ Λ.
Finally, assume that M0 ⊂ M is a von Neumann subalgebra with the relative property (T) and such
that M0 has no injective direct summand.
Then, for every ε > 0, there exists n and a sequence (sk)k in Λ such that
1. kϕn(vsk ) − vskk2 ≤ ε for all k,
2. kEP (xvsk y)k2 → 0 for all x, y ∈ M .
Proof. We first argue that we may actually assume that p = 1. Put N = pM p and define ψn : N →
N : ψn(a) = pϕn(a)p. Then, ψn is a sequence of subunital, subtracial, normal, completely positive
maps and ka − ψn(a)k2 → 0 for all a ∈ N . Let ε > 0. Since kp − ϕn(p)k2 → 0, we can take n0 such
that kψn(a) − ϕn(a)k2 ≤ ε
2 kak for all a ∈ N and all n ≥ n0. We now replace M by N and ϕn by
ψn. Since M0 is diffuse and M is a factor, we may assume that p ∈ M0 and finally replace M0 by
pM0p ⊂ N .
So, for the rest of the proof, we assume that p = 1. We have M = Q ⋊ Λ.
Define the positive-definite functions eϕn : Λ → C : eϕn(s) := τ (v∗
pointwise. We have to prove the following statement: for every ε > 0, there exists n and a sequence
(sk)k in Λ with the properties
s ϕn(vs)). Note that eϕn → 1
eϕn(sk) − 1 ≤ ε for all k and kEP (xvsk y)k2 → 0 for all x, y ∈ M .
Suppose that this statement is false. Fix ε > 0 such that for every n, it is impossible to find a
sequence (sk)k in Λ with properties (3.1).
(3.1)
Define for every n, the normal completely positive map θn : M → M satisfying θn(bvs) = eϕn(s)bvs
for all b ∈ Q, s ∈ Λ. Note that kθn(x) − xk2 → 0 for all x ∈ M . Since M0 ⊂ M has the relative
property (T), fix n such that kθn(w) − wk2
By assumption, it is impossible to find a sequence (sk)k in Λ with properties (3.1). Define
2 ≤ ε2/2 for all w ∈ U (M0).
It follows that we can take a finite subset F ⊂ M and a δ > 0 such that for all s ∈ V, we have
V = {s ∈ Λ eϕn(s) − 1 ≤ ε} .
Xx,y∈F
kEP (xvsy∗)k2
2 ≥ 2δ .
9
Consider the Hilbert space K = L2(hM, eP i) and the unitary representation
π : Λ → U (K) : π(s)ξ = vsξv∗
s .
ψ : Λ → [0, +∞) : ψ(s) = hξ0, π(s)ξ0i .
Since P is injective, the unitary representation π is weakly contained in the regular representation
of Λ. Define ξ0 ∈ K by the formula ξ0 =Px∈F x∗eP x. Define the positive definite function
Note that ψ(s) =Px,y∈F kEP (xvsy)k2
Define the normal completely positive map ρ : M → M such that ρ(bvs) = ψ(s)bvs for all b ∈ Q,
s ∈ Λ. Since Q is injective and since π is weakly contained in the regular representation of Λ,
it follows that the M -M -bimodule MKρ
M defined by ρ is weakly contained in the coarse M -M -
bimodule M(L2(M ) ⊗ L2(M ))M.
We claim that τ (w∗ρ(w)) ∈ [δ, 1] for all w ∈ U (M0). Since kθn(w)−wk2
suffices to prove that τ (w∗ρ(w)) ∈ [δ, 1] for every unitary w ∈ U (M ) satisfying kθn(w)−wk2
2 ≤ ε2/2 for all w ∈ U (M0), it
2 ≤ ε2/2.
2 and hence, ψ(s) ≥ 2δ for all s ∈ V.
Take such a unitary w and write w =Ps∈Λ bsvs. It follows that
eϕn(s) − 12 kbsk2
2
ε2
2
≥ kθn(w) − wk2
2 =Xs∈Λ
eϕn(s) − 12 kbsk2
kbsk2
2 .
2
≥ Xs∈Λ−V
≥ ε2 Xs∈Λ−V
We conclude that
Since w is unitary, it follows that
It now follows that
kbsk2
2 ≤
1
2
.
kbsk2
2 ≥
1
2
.
Xs∈Λ−V
Xs∈V
τ (w∗ρ(w)) =Xs∈Λ
≥ 2δXs∈V
ψ(s)kbsk2
ψ(s)kbsk2
2
2 ≥Xs∈V
kbsk2
2 ≥ δ .
Hence, our claim is proven.
Since τ (w∗ρ(w)) ∈ [δ, 1] for all w ∈ U (M0), the M -M -bimodule MKρ
M contains a non-zero M0-
central vector. On the other hand, the M -M -bimodule MKρ
M is weakly contained in the coarse M -
M -bimodule M(L2(M ) ⊗ L2(M ))M. It follows that the coarse M0-M0-bimodule admits a sequence
of almost M0-central unit vectors. This is a contradiction with the assumption that M0 has no
injective direct summand.
Lemma 3.2. Let M be a II1 factor with the following deformation property. We have an in-
clusion M ⊂ fM of M into the finite von Neumann algebra fM such that the M -M -bimodule
10
M(cid:0)L2(fM ) ⊖ L2(M )(cid:1)M is weakly contained in the coarse M -M -bimodule M(L2(M ) ⊗ L2(M ))M. We
are given a sequence αn ∈ Aut(fM ) such that ka − αn(a)k2 → 0 for all a ∈ fM .
Let P ⊂ M be a von Neumann subalgebra and assume that P is injective.
Assume that the countable group Λ acts trace preservingly on the injective von Neumann algebra
Q and suppose that we have identified Q ⋊ Λ = pM p for some projection p ∈ M . We denote by
(vs)s∈Λ the canonical unitaries in Q ⋊ Λ.
Finally, assume that M1, M2 ⊂ M are von Neumann subalgebras such that M1 and M2 commute
and M1, M2 have no injective direct summand.
Then, for every ε > 0, there exists n and a sequence sk ∈ Λ such that
1. kαn(vsk ) − EM (αn(vsk ))k2 ≤ ε for all k,
2. kEP (xvsk y)k2 → 0 for all x, y ∈ M .
Proof. We use throughout the trace τ normalized in such a way that τ (p) = 1. We use the notation
k · k2 accordingly.
Assume that the statement is false. Take ε > 0 such that for all n ∈ N, it is impossible to find a
sequence sk ∈ Λ satisfying properties 1 and 2 above.
Put N := Q ⋊ Λ, which we identified with pM p. Define the normal, unital ∗-homomorphism
β : N → N ⊗N given by β(avs) = avs ⊗ vs for all a ∈ Q, s ∈ Λ. Put eN := pfMp. Define
the N -N -bimodule NHN given by H := L2(N ) ⊗ L2(eN ) with x · ξ · y := β(x)ξβ(y). Put H0 :=
L2(N ) ⊗ (L2(eN ) ⊖ L2(N )). Note that H0 ⊂ H is an N -N -subbimodule. By our assumptions and
because Q is injective, the N -N -bimodule NH0
N is weakly contained in the coarse N -N -bimodule.
Since M1 is diffuse and M is a factor, we may assume that p ∈ M1. We replace M1 by pM1p and
M2 by M2p so that they become commuting subalgebras of N .
Put ε1 = 1
M1 is weakly contained in
the coarse M1-M1-bimodule, we can take a finite subset F ⊂ M1 and a ρ > 0 such that whenever
ξ ∈ H0 and kx · ξ − ξ · xk2 ≤ ρ for all x ∈ F, then kξk2 ≤ ε1.
Define the linear map
10 ε. Since M1 has no injective direct summand and since M1H0
θn : M2 → H : θn(a) = (1 ⊗ p)(id ⊗ αn)β(a)(1 ⊗ p) .
Take n large enough such that
kαn(p) − pk2
and k(id ⊗ α−1
n )β(x) − β(x)k2 , x ∈ F ,
are all sufficiently small in order to ensure that
kx · θn(a) − θn(a) · xk2 ≤ ρ for all x ∈ F , a ∈ U (M2) .
It follows that
kθn(a) − (id ⊗ EM )θn(a)k2 ≤ ε1
for all a ∈ U (M2) .
When choosing n, we can make sure that kαn(p) − pk2 ≤ ε1, yielding
kθn(a) − (id ⊗ αn)β(a)k2 ≤ 2ε1kak
for all a ∈ M2 .
11
As a conclusion, we get
k(id ⊗ αn)β(a) − (id ⊗ EM ◦ αn)β(a)k2 ≤ 5ε1
for all a ∈ U (M2) .
(3.2)
Define V = {s ∈ Λ kαn(vs) − EM (αn(vs))k2 ≤ ε}. By our assumption ex absurdo, it is impossible
to find a sequence sk ∈ Λ satisfying conditions 1 and 2 in the formulation of the lemma. Hence, we
find a finite subset F1 ⊂ M and a δ > 0 such that for all s ∈ V, we have
Xx,y∈F1
kEP (xvsy∗)k2
2 ≥ δ .
Define the N -N -bimodule NKN where
K = L2(N ) ⊗ p L2(hM, eP i)p and x · ξ · y = β(x)ξβ(y) .
Since P is injective, N(cid:0)p L2(hM, eP i)p(cid:1)N is contained in the coarse N -N -bimodule. Since Q is
injective, also NKN is weakly contained in the coarse N -N -bimodule. We now prove that K admits
a non-zero M2-central vector. Since M2 has no injective direct summand, this yields the required
contradiction.
Define the vector ξ ∈ K by the formula
ξ := Xx∈F1
1 ⊗ px∗eP xp .
We will prove that
ha · ξ · a∗, ξi ≥
3
4
Take a ∈ U (M2) and write a =Ps∈Λ asvs with as ∈ Q. Then,
(id ⊗ αn)β(a) − (id ⊗ EM ◦ αn)β(a) =Xs∈Λ
Using (3.2) and the definition of V ⊂ Λ, it follows that
asvs ⊗(cid:0)αn(vs) − EM (αn(vs))(cid:1) .
δ
for all a ∈ U (M2) .
(3.3)
(5ε1)2 ≥ k(id ⊗ αn)β(a) − (id ⊗ EM ◦ αn)β(a)k2
2
kask2
2 kαn(vs) − EM (αn(vs))k2
2
=Xs∈Λ
≥ ε2 Xs∈Λ\V
kask2
2 .
Using our definition of ε1, we conclude that
It follows that
1
4
2 ≤
kask2
Xs∈Λ\V
ha · ξ · a∗, ξi =Xs∈Λ
≥ δXs∈V
and hence Xs∈V
2(cid:16) Xx,y∈F1
kask2
kask2
2 ≥
3
4
δ .
kask2
2 ≥
3
4
.
kEP (xvsy∗)k2
2(cid:17)
So, we have shown (3.3). It follows that the unique vector of minimal norm in the closed convex
hull of {a · ξ0 · a∗ a ∈ U (M2)} is non-zero and M2-central. This ends the proof of the lemma.
12
4 Some Herz-Schur multipliers on amalgamated free products
Let Γ be a countable group. A function ϕ : Γ → C is called a Herz-Schur multiplier if ug 7→
ϕ(g)ug extends to an ultraweakly continuous, completely bounded linear map mϕ : L(Γ) → L(Γ).
The linear space of Herz-Schur multipliers on Γ is denoted by B2(Γ). Whenever ϕ ∈ B2(Γ), put
kϕkcb := kmϕkcb.
Let Γ = Γ1 ∗Σ Γ2 be an amalgamated free product. All elements g ∈ Γ have a natural length g,
arising by writing g as an alternating product of elements Γ1 − Σ and elements in Γ2 − Σ. By
convention, g = 0 if and only if g ∈ Σ.
The following result is probably well known. It is an immediate consequence of [BP93, Proposition
3.2] and we include a proof for the convenience of the reader.
Lemma 4.1. 1. For every K ∈ N, there exists K ′ > K and ψ ∈ B2(Γ) with the following properties.
• kψkcb ≤ 2.
• ψ(g) = 1 if g ≤ K and ψ(g) = 0 if g ≥ K ′.
• 0 ≤ ψ(g) ≤ 1 for all g ∈ Γ.
2. For every K ∈ N, there exists K ′ > K and ϕ ∈ B2(Γ) with the following properties.
• kϕkcb ≤ 3.
• ϕ(g) = 0 if g ≤ K and ϕ(g) = 1 if g ≥ K ′.
• 0 ≤ ϕ(g) ≤ 1 for all g ∈ Γ.
Proof. 1. Whenever 0 < ρ < 1, define θρ(g) = ρg. For all n ∈ N, put γn(g) = 1 if g = n
and γn(g) = 0 if g 6= n. By [BP93, Proposition 3.2], all θρ and γn belong to B2(Γ) and satisfy
kθρkcb = 1 and kγnkcb ≤ 4n + 1.
Choose K ∈ N. Take 0 < ρ < 1 close enough to 1 such that
Next, take K ′ > K large enough such that
(1 − ρn)(4n + 1) ≤
1
2
.
ρn(4n + 1) ≤
1
2
.
if g ≤ K ,
if K < g < K ′ ,
if g ≥ K ′ .
KXn=0
∞Xn=K ′
ψ(g) =
KXn=0
1
ρk
0
Define
Since
ψ = θρ +
(1 − ρn)γn −
ρnγn ,
∞Xn=K ′
we conclude that ψ ∈ B2(Γ) and kψkcb ≤ 2.
2. Take ψ as in 1 and put ϕ(g) = 1 − ψ(g).
13
5 Factors with unique group measure space Cartan subalgebra
Definition 5.1. We define the family G of groups Γ of the form Γ = Γ1 ∗Σ Γ2 with the following
properties.
1. Γ1 contains a non-amenable subgroup with the relative property (T) or Γ1 contains two
non-amenable commuting subgroups,
2. Σ is amenable and Γ2 6= Σ,
3. There exist g1, . . . , gk ∈ Γ such that
giΣg−1
i
is finite .
k\i=1
Our main result says that all group measure space II1 factors with groups Γ ∈ G have a unique
group measure space Cartan subalgebra. We also deal with amplifications. Therefore, denote by
Dn(C) the subalgebra of diagonal matrices in Mn(C). The following is a more general version of
Theorem 1.1.
Theorem 5.2. Let Γ be a group in the family G and Γ y (X, µ) a free ergodic p.m.p. action.
Denote M = L∞(X) ⋊ Γ. Whenever Λ y (Y, η) is a free ergodic p.m.p. action, p ∈ Mn(C) ⊗ M is
a projection and
π : L∞(Y ) ⋊ Λ → p(Mn(C) ⊗ M )p
is an isomorphism, there exists a projection q ∈ Dn(C) ⊗ L∞(X) and a unitary u ∈ q(Mn(C) ⊗ M )p
such that
π(L∞(Y )) = u∗(Dn(C) ⊗ L∞(X))u .
In the rest of this section, we prove Theorems 5.2 and 1.4. We already deduce the following result,
similar to [PV08c, Theorem 1.2] which dealt with certain free product groups Γ = Γ1 ∗ Γ2.
Corollary 5.3. Let Γ = Γ1 ∗Σ Γ2 be a group in the family G. Assume that Γ1 and Γ2 are finitely
generated and assume that at least one of the Γi has fixed prize with cost strictly larger than 1 (e.g.
Γ2 = Fn for 2 ≤ n < ∞).
For any free ergodic p.m.p. action Γ y (X, µ), the II1 factor L∞(X) ⋊ Γ has trivial fundamental
group. In other words, using the notation of [PV08c], we have Sfactor(Γ) = {{1}}.
Proof. Theorem 5.2 implies that the fundamental group of L∞(X)⋊Γ equals the fundamental group
of the orbit equivalence relation R = R(Γ y X). By [Ga00, Th´eor`eme IV.15], R has cost strictly
between 1 and ∞. It then follows from [Ga00, Proposition II.6] that R has trivial fundamental
group.
5.1 Deformation of amalgamated free product factors
Let M1, M2 be von Neumann algebras equipped with a faithful normal tracial state τ . Assume that
P is a common von Neumann subalgebra of M1 and M2 and that the traces of M1, M2 coincide
on P . Denote by M = M1 ∗P M2 the amalgamated free product with respect to the unique trace
preserving conditional expectations (see [Po93] and [VDN92]) and still denote by τ the canonical
14
tracial state on M . The Hilbert space L2(M ) can be explicitly realized as follows, where we denote
L2(M ◦
i ) := L2(Mi) ⊖ L2(P ).
L2(M ) = L2(P ) ⊕
i16=i2,i26=i3,··· ,in−16=in(cid:0)L2(M ◦
M
i1) ⊗P L2(M ◦
i2) ⊗P · · · ⊗P L2(M ◦
in )(cid:1) .
For all 0 < ρ < 1, denote by mρ ∈ B(L2(M )) the operator given by multiplication with the positive
scalar ρn on L2(M ◦
in). Actually (and this will incidentally be a
consequence from the following discussion), there is a unique normal unital completely positive
map mρ : M → M whose extension to L2(M ) is the mρ that we have just defined.
Following [IPP05, Section 2.2], we can define the following deformation of M . Define, for i = 1, 2,
i2) ⊗P · · · ⊗P L2(M ◦
i1 ) ⊗P L2(M ◦
fMi := Mi ∗P (P ⊗ L(Z)). Define fM := fM1 ∗P fM2 and observe that we have a canonical identification
fM = M ∗P (P ⊗ L(F2)). We now define a one-parameter group of automorphisms αt of fM .
Whenever α ∈ Aut(fM1) and β ∈ Aut(fM2) are both the identity when restricted to P , we have
a unique α ∗ β ∈ Aut(fM ) simultaneously extending α and β. Denote by u the canonical unitary
generator of L(Z) viewed as a subalgebra of fM1 and denote by v the canonical unitary generator of
L(Z) viewed as a subalgebra offM2. Let f : T → (−π, π] be the unique map satisfying z = exp(if (z))
for all z ∈ T. Define the self-adjoint elements h ∈ fM1 and k ∈ fM2 as h = f (u) and k = f (v). Put,
for all t ∈ R, ut = exp(ith) and vt = exp(itk). Then, Ad(ut) ∈ Aut(fM1) and Ad(vt) ∈ Aut(fM2)
are both the identity on P , so that we can define αt ∈ Aut(fM ) as αt := Ad(ut) ∗ Ad(vt).
Define ρt = τ (ut)2 = τ (vt)2. It happens to be that ρt = sin2(πt)
0, then ρt increases from 0 to 1. By using the definition of the trace on M , it is easy to see that
(πt)2 and hence, if t decreases from 1 to
EM (αt(x)) = mρt (x) .
This also shows that mρ is a normal unital completely positive map on M for all 0 < ρ < 1.
Observe that, when P is injective, the M -M -bimodule M(cid:0)L2(fM ) ⊖ L2(M )(cid:1)M is weakly contained
in the coarse M -M -bimodule M(L2(M ) ⊗ L2(M ))M (see [CH08, Proposition 3.1] for a detailed
argument). So, we are then in a situation where Lemma 3.2 can potentially be applied.
We now present one of the main technical results from [IPP05]. Since the statement as we need it,
is not exactly formulated in [IPP05], we give a sketch of proof, indicating the different steps from
[IPP05] that are needed to prove the result. Recall that the normalizer NN (Q)′′ of a von Neumann
subalgebra Q ⊂ N is the von Neumann algebra generated by the group of unitaries u ∈ N satisfying
uQu∗ = Q.
Theorem 5.4. Let M1, M2 be tracial von Neumann algebras with a common von Neumann sub-
algebra P on which the traces coincide. Denote by M = M1 ∗P M2 the amalgamated free product
w.r.t. the trace preserving conditional expectations. Denote, for 0 < ρ < 1, by mρ the completely
positive map on M introduced above. Let p ∈ M be a projection and Q ⊂ pM p a von Neumann
subalgebra.
If there exists 0 < ρ < 1 and δ > 0 such that τ (v∗mρ(v)) ≥ δ for all v ∈ U (Q), then Q ≺M P or
NpM p(Q)′′ ≺M Mi for some i ∈ {1, 2}. As explained above, NpM p(Q)′′ denotes the normalizer of
Q inside pM p.
Proof. Clearly, τ (v∗mρ(v)) increases when ρ increases. So, with the notation introduced before the
theorem, we can take 0 < t0 < 1 such that
τ (v∗αt(v)) = τ (v∗EM (αt(v))) = τ (v∗mρt(v)) ≥ δ
for all v ∈ U (Q) , 0 < t ≤ t0 .
15
In particular, we can take t of the form t = 2−n such that τ (v∗αt(v)) ≥ δ for all v ∈ U (Q). Denote by
It follows that y 6= 0 because τ (y) ≥ δ and that xy = yαt(x) for all x ∈ Q.
Assume that Q 6≺M P . We have to prove that NpM p(Q)′′ ≺M Mi for some i ∈ {1, 2}. Repeating
y ∈ pfM αt(p) the unique element of minimal k · k2 in the closed convex hull of {v∗αt(v) v ∈ U (Q)}.
[IPP05, Proof of 3.3] (see also [Ho07, Step (2) of 5.6]), we find a non-zero z ∈ pfMα1(p) satisfying
xz = zα1(x) for all x ∈ Q. As in [IPP05, Proof of 4.3] (actually, literally repeating [Ho07, Step
(3) of 5.6]), we can conclude that for i = 1 or i = 2, we have Q ≺M Mi. Since we assumed
that Q 6≺M P , [IPP05, Theorem 1.1] (see also [Ho07, Theorem 4.6]) implies that the normalizer
NpM p(Q)′′ can be embedded into Mi inside M .
For later use, we also record the following non-optimal inequality. As positive operators on L2(M ),
we have (1 − mρ)2 ≤ 1 − mρ ≤ 1 − m2
2. On the other
hand, for all t ∈ R and x ∈ M , we have
ρ. Hence, kx − mρ(x)k2
2 − kmρ(x)k2
2 ≤ kxk2
kαt(x) − EM (αt(x))k2
2 − kmρt(x)k2
2 .
2 = kxk2
2 − kEM (αt(x))k2
2 = kxk2
It follows that
kx − mρt (x)k2 ≤ kαt(x) − EM (αt(x))k2
for all x ∈ M , 0 < t < 1 .
(5.1)
Finally, let Γ = Γ1 ∗Σ Γ2 be an amalgamated free product of groups. Assume that Γ acts in a trace
preserving way on the tracial von Neumann algebra (A, τ ). Put M = A ⋊ Γ. Putting Mi = A ⋊ Γi
for i = 1, 2 and P = A ⋊ Σ, we have M = M1 ∗P M2 in a canonical way. Moreover, for all 0 < ρ < 1,
a ∈ A, g ∈ Γ, we have
mρ(aug) = ρgaug ,
where g denotes the length of g in the sense explained at the beginning of Section 4.
5.2 A combinatorial lemma
Fix an amalgamated free product Γ = Γ1 ∗Σ Γ2 and a trace preserving action Γ y (A, τ ) of Γ on
the tracial von Neumann algebra (A, τ ). Put M = A ⋊ Γ and P = A ⋊ Σ.
For every K ∈ N, denote by PK the orthogonal projection of L2(M ) onto the closed linear span of
{aug g ≤ K, a ∈ A}.
Whenever g0, h0 ∈ (Γ1 − Σ) ∪ (Γ2 − Σ), denote by Wg0,h0 the subset of Γ consisting of those g ∈ Γ
with g ≥ 2 admitting a reduced expression starting with g0 and ending with h0 (this means that
any reduced expression for g starts with g0σ and ends with σ′h0 for some σ, σ′ ∈ Σ). Denote by
Pg0,h0 the orthogonal projection of L2(M ) onto the closed linear span of {aug g ∈ Wg0,h0, a ∈ A}.
In general, whenever W ⊂ Γ, denote by PW the orthogonal projection of L2(M ) onto the closed
linear span of {aug g ∈ W, a ∈ A}.
Lemma 5.5. Let K ∈ N and assume that (yk) is a bounded sequence in M with the following
properties.
• yk = PK(yk) for all k.
• kEP (xykz)k2 → 0 for all x, z ∈ M .
16
Let g, h ∈ Γ with g, h ≥ K and write g, h as reduced expressions. Denote by g0 the first letter of
g and by h0 the last letter of h. Then, we can write
where ak, bk are bounded sequences in M satisfying the following properties.
ugykuh = ak + bk
• ak = Pg0,h0(ak) for all k.
• kbkk2 → 0.
Proof. We fix once and for all words with letters alternatingly from Γ1 − Σ and Γ2 − Σ representing
the elements g and h.
Let (g1, h1), · · · , (gN , hN ) be an enumeration of all pairs of words (g′, h′) satisfying
• g′ + h′ ≤ K,
• the fixed word representing g ends with the subword g′,
• the fixed word representing h starts with the subword h′.
Define Wi = gg−1
i Σh−1
i=1 Wi. Observe that
i h and W =SN
PWi(x) = ugu∗
giEP (ugiu∗
gxu∗
huhi)u∗
hiuh
for all x ∈ M . Hence, PWi is completely bounded as a map from M to M . The orthogonal
projections PWi commute and hence
1 − PW = (1 − PW1) · · · (1 − PWN ) .
So, PW is completely bounded on M . We put bk = PW (ugykuh) and ak = ugykuh − bk. So, ak and
bk are bounded sequences in M with ugykuh = ak + bk.
First observe that
kbkk2
2 ≤
NXi=1
kPWi(ugykuh)k2
2 =
NXi=1
kEP (ugiykuhi)k2
2 → 0 .
It remains to prove that ak = Pg0,h0(ak) for all k. But, if r ∈ Γ with r ≤ K and if grh admits no
reduced expression that starts with g0 and ends with h0, there must exist i ∈ {1, . . . , N } such that
girhi ∈ Σ and hence grh ∈ W . As a consequence, whenever y ∈ M with y = PK (y), we have
ugyuh − PW (ugyuh) ∈ Pg0,h0(L2(M )) .
This concludes the proof of the lemma.
17
5.3 Group measure space Cartan subalgebras can be intertwined into A ⋊ Σ
Fix Γ = Γ1 ∗Σ Γ2 satisfying the conditions 1 and 2 of Definition 5.1. Let Γ y (A, τ ) be a trace
preserving action of Γ on the tracial injective von Neumann algebra (A, τ ). Put M = A ⋊ Γ and
P = A ⋊ Σ. Note that P is injective, because A is injective and Σ is amenable.
Theorem 5.6. Whenever p ∈ M is a non-zero projection and pM p = B ⋊ Λ is a crossed product
decomposition where Λ y (B, τ ) is a trace preserving action on the abelian von Neumann algebra
B, then B ≺M P .
We will prove Theorem 5.6 by combining the transfer of rigidity lemmas 3.1, 3.2 with the following
result saying that any abelian algebra that is normalized by 'many' unitaries of short word length,
is itself uniformly of short length.
As in Section 5.1, define for every 0 < ρ < 1, the unital completely positive map mρ on M by
mρ(aug) = ρgaug for all a ∈ A, g ∈ Γ.
Lemma 5.7. Let p ∈ M be a projection and B ⊂ pM p an abelian von Neumann subalgebra. Put
ε = τ (p)/2072 and assume that we are given 0 < ρ < 1 and a sequence of unitaries vk ∈ pM p that
normalize B and satisfy
• kvk − mρ(vk)k2 ≤ ε/2 for all k,
• kEP (xvky)k2 → 0 for all x, y ∈ M .
Then, there exists a 0 < ρ0 < 1 and a δ > 0 such that τ (w∗mρ0(w)) ≥ δ for all w ∈ U (B).
Proof. Throughout the proof we make use of the Herz-Schur multipliers provided by Lemma 4.1.
Whenever ϕ ∈ B2(Γ), we can extend mϕ to A ⋊ Γ, without increasing the cb-norm, by putting
mϕ(aug) = ϕ(g)aug for all a ∈ A and g ∈ Γ.
Assume that the lemma is false. Take K ∈ N such that ρK ≤ 1/2. Denote as before by PK the
orthogonal projection of L2(M ) onto the closed linear span of {aug a ∈ A, g ≤ K}. Note that
kv − PK (v)k2 ≤ 2kmρ(v) − vk2
for all v ∈ M .
By Lemma 4.1.1, take K1 > K and ψ ∈ B2(Γ) satisfying kψkcb ≤ 2, ψ(g) = 1 if g ≤ K, ψ(g) = 0
if g ≥ K1 and 0 ≤ ψ(g) ≤ 1 for all g ∈ Γ. Define yk := mψ(vk). Since
kvk − PK (vk)k2 ≤ 2kmρ(vk) − vkk2 ≤ ε ,
we get kvk − ykk2 ≤ ε. So, the sequence (yk) satisfies
• kykk ≤ 2 for all k,
• yk = PK1(yk) for all k,
• kvk − ykk2 ≤ ε.
18
We claim that also kEP (xykz)k2 → 0 for all x, z ∈ M . Since (yk) is a bounded sequence and since
A ⊂ P , we may assume that x = ug, z = uh for some g, h ∈ Γ. Denote by P0 the orthogonal
projection of L2(M ) onto the closed linear span of {aur a ∈ A, r ∈ g−1Σh−1}. Then,
kEP (ugykuh)k2 = kP0(yk)k2 = kP0(mψ(vk))k2
= kmψ(P0(vk))k2 ≤ kP0(vk)k2
= kEP (ugvkuh)k2 → 0 .
This proves the claim.
By Lemma 4.1.2, take K2 > 2K1 and ϕ ∈ B2(Γ) satisfying kϕkcb ≤ 3, ϕ(g) = 0 if g ≤ 2K1,
ϕ(g) = 1 if g ≥ K2 and 0 ≤ ϕ(g) ≤ 1 for all g ∈ Γ.
Take 0 < ρ0 < 1 close enough to 1 such that ρK2
a unitary w ∈ U (B) such that
0 ≥ 1/2. By our assumption by contradiction, take
τ (w∗mρ0(w)) ≤
ε2
8
.
It follows that kPK2(w)k2 ≤ ε/2. By the Kaplansky density theorem, take w0 in the dense ∗-
subalgebra
with kw0k ≤ 1 and kw − w0k2 ≤ ε/2. Put x = mϕ(w0). Then,
M0 = span{aug a ∈ A, g ∈ Γ}
kmϕ(w) − xk2 = kmϕ(w − w0)k2 ≤ kw − w0k2 ≤
ε
2
.
Since kPK2(w)k2 ≤ ε/2, also kw − mϕ(w)k2 ≤ ε/2. So, our element x ∈ M has the following
properties.
• x ∈ span{aug a ∈ A, g > 2K1},
• kxk ≤ 3,
• kw − xk2 ≤ ε.
Since B is abelian, w ∈ U (B) and vk normalizes B, we have
p = vkwv∗
k w vkw∗v∗
k w∗
for all k. We now replace vk by yk and w by x. We use the estimates for kxk, kykk, kvk − ykk2 and
kw − xk2, to conclude that
kp−yk x y∗
k x yk x∗ y∗
k x∗k2
≤ ε(1 + 3 + 2 · 3 + 3 · 2 · 3 + 2 · 3 · 2 · 3 + 3 · 2 · 3 · 2 · 3 + 2 · 3 · 2 · 3 · 2 · 3 + 3 · 2 · 3 · 2 · 3 · 2 · 3)
= 1036ε =
1
2
τ (p) .
It follows that for all k,
τ (yk x y∗
k x yk x∗ y∗
k x∗) ≥
1
2
τ (p) .
We claim however that the left-hand side tends to 0 when k → ∞.
19
Since x ∈ span{aug a ∈ A, g > 2K1}, it suffices to prove that
τ (yk ugaug′ y∗
k uhbuh′ yk urcur′ y∗
k utdut′ ) → 0
(5.2)
for all a, b, c, d ∈ A and for all group elements g, g′, h, h′, k, k′, r, r′, t, t′ having length at least K1
and being chosen such that the concatenations gg′, hh′, rr′ and tt′ are reduced.
We now study the expressions
ut′ykug , ug′y∗
kuh , uh′ykur and ur′y∗
kut .
According to Lemma 5.5, every of these four expressions can be written as a sum a(i)
i = 1, 2, 3, 4, of bounded sequences satisfying the following properties: kb(i)
a(i)
k = Pg0,h0(a(i)
g, h, r, t.
k + b(i)
k ,
k k2 → 0 and we have
k ) where g0 is the first letter of resp. t′, g′, h′, r′ and h0 is the last letter of resp.
Then, the left hand side of (5.2) equals
k + b(1)
k ) a (a(2)
k + b(2)
k ) b (a(3)
k + b(3)
k ) c (a(4)
k + b(4)
τ(cid:0) (a(1)
k ) d(cid:1) .
Developing all the sums, the one term that only involves a(i)
equals zero because there is no
k
simplification between consecutive factors of the product and all the other terms tend to zero
because kb(i)
k k2 → 0 and the (a(i)
k ) are bounded.
We have reached a contradiction and hence we have proven the lemma.
It is now easy to prove Theorem 5.6.
Proof of Theorem 5.6. Denote by (vs)s∈Λ the canonical unitaries in B ⋊ Λ. Put ε = τ (p)/2072.
In the situation where Γ1 has a non-amenable subgroup H with the relative property (T), we
apply Lemma 3.1 to the completely positive maps mρ, ρ → 1 and the von Neumann subalgebra
M0 = L(H). In the situation where Γ1 has two non-amenable commuting subgroups H1, H2, we
apply Lemma 3.2 to the deformation αt, t → 0, introduced in Section 5.1 and the commuting
subalgebras L(H1), L(H2). Combined with (5.1), we always find 0 < ρ < 1 and a sequence sk ∈ Λ
such that
• kvsk − mρ(vsk )k2 ≤ ε/2 for all k.
• kEP (xvsk y)k2 → 0 for all x, y ∈ M .
By Lemma 5.7 we get a 0 < ρ < 1 and a δ > 0 such that τ (w∗mρ(w)) ≥ δ for all w ∈ U (B).
By Theorem 5.4 we have that B ≺M P or that NpM p(B)′′ ≺M A ⋊ Γi for some i = 1, 2. Since
NpM p(B)′′ = pM p and since Γi < Γ has infinite index, the second option is impossible, concluding
the proof of the theorem.
20
5.4 Proof of Theorems 5.2 and 1.4
Proof of Theorem 5.2. Put A = Mn(C) ⊗ L∞(X) with Γ y A acting trivially on Mn(C). In the
proof, we do not write the isomorphism π and we put B = L∞(Y ). So, A ⋊ Γ = Mn(C) ⊗ M and
p is a projection in A ⋊ Γ such that p(A ⋊ Γ)p = B ⋊ Λ. By Theorem 5.6, we get an intertwining
bimodule between B and A ⋊ Σ. So, we also have
B ≺M L∞(X) ⋊ Σ .
By condition 3 in Definition 5.1 and [PV06, Theorem 6.16], it follows that B ≺M L∞(X). Then,
[Po01, Theorem A.1] (see also [Va06, Theorem C.3]) provides the conclusion of the theorem.
Proof of Theorem 1.4. Since Γ acts ergodically on (X, µ) and Γ is ICC, it follows that M is a factor.
Assume that B is a group measure space Cartan subalgebra in pM np. Repeating the previous proof,
it follows that B ≺M L∞(X). By [OP07, Theorem 4.11], it follows that Γ y (X, µ) is free.
5.5
II1 factors with at least two group measure space Cartan subalgebras
rigidity condition (5.1.1), where Σ is amenable and different from Γ2 (5.1.2) and whereTk
By definition, the family G consists of amalgamated free products Γ = Γ1 ∗Σ Γ2, where Γ1 satisfies a
i=1 giΣg−1
is finite for some g1, . . . , gk ∈ Γ (5.1.3). Without this last condition, it is possible to give examples
of group actions such that M = L∞(X) ⋊ Γ admits at least two Cartan subalgebras that are
non-conjugate by an automorphism of M .
i
Indeed, Connes and Jones [CJ82] provide examples of free ergodic p.m.p. actions Γ × Σ y (X, µ)
such that the II1 factor L∞(X) ⋊ (Γ × Σ) has at least two group measure space Cartan subalgebras
that are non-conjugate by an automorphism. In their construction, Γ can be any non-amenable
group and Σ is a specific infinite amenable group. In particular, one can consider
(Γ1 ∗ Γ2) × Σ = (Γ1 × Σ) ∗Σ (Γ2 × Σ)
and provide examples where conditions 5.1.1 and 5.1.2 are satisfied, but condition 5.1.3 is not.
Mimicking [OP08, Section 7], we give other examples of II1 factors with at least two group measure
space decompositions. The following general non-uniqueness statement is a consequence of Example
5.8. Whenever Γ = H ⋊ G is a semi-direct product group with H being infinite abelian, then Γ
admits free ergodic p.m.p. actions such that the corresponding group measure space II1 factor
has at least two non unitarily conjugate group measure space Cartan subalgebras. Whenever
G = G1 ∗Σ G2, also Γ = (H ⋊ G1) ∗H⋊Σ (H ⋊ G2). As such, we get again examples where conditions
5.1.1 and 5.1.2 are satisfied, but condition 5.1.3 is not.
Example 5.8. Let H be an infinite abelian group and G yα H an action by automorphisms. Let
H ֒→ K be a dense embedding of H into the compact abelian group K. Assume that G y H
extends to an action by homeomorphisms of K that we still denote by α. Whenever G y (X, µ) is
a free ergodic p.m.p. action, consider the free ergodic p.m.p. action
H ⋊ G y K × X given by (h · (k, x) = (h + k, x)
g · (k, x) = (αg(k), g · x)
for all h ∈ H, g ∈ G, k ∈ K, x ∈ X .
21
L∞(K × X).
automorphism. This is, for instance, always the case when the action G y H is trivial. Then, the
L∞(K × X) ⋊ (H ⋊ G) = L∞(bH × X) ⋊ (bK ⋊ G) .
unitarily conjugate. Indeed, if hn ∈ H is a sequence tending to infinity in H, the unitaries uhn ∈
L(H) satisfy kEL∞(K×X)(auhnb)k2 → 0 for all a, b ∈ L∞(K × X) ⋊ (H ⋊ G).
It follows that
Dualizing the embedding H ֒→ K, we get the embedding bK ֒→ bH and the action of G by automor-
phisms of bK and bH. We canonically have
First of all, the group measure space Cartan subalgebras L∞(K × X) and L∞(bH × X) are never
L∞(bH) = L(H) 6≺ L∞(K × X). A fortiori, L∞(bH × X) cannot be unitarily conjugated onto
In certain examples, the Cartan subalgebras L∞(K × X) and L∞(bH × X) are conjugate by an
crossed product II1 factor is the tensor product of L∞(X) ⋊ G and L∞(K) × H = L∞(bH) ⋊ bK.
and L∞(bH) are conjugate by an automorphism [OW80, CFW81].
In other examples, the Cartan subalgebras L∞(K × X) and L∞(bH × X) are non conjugate by an
of p-adic integers for some prime number p. It follows that bK ⋊ G is the direct limit of a sequence
with the relative property (T)), while bK ⋊ G has the Haagerup property (as the direct limit
• For n = 3, H ⋊ G has property (T), while bK ⋊ G does not have property (T) (as the direct
The second tensor factor is the hyperfinite II1 factor and hence, the Cartan subalgebras L∞(K)
automorphism. Consider G = SL(n, Z) y H = Zn and Zn ֒→ K = Zn
p , where Zp denotes the ring
of groups with the Haagerup property).
limit of a strictly increasing sequence of groups).
of groups that are virtually isomorphic with SL(n, Z).
• For n = 2, H ⋊ G does not have the Haagerup property (because H is an infinite subgroup
Since both property (T) [Fu99a, Corollary 1.4] and the Haagerup property [Po01, Remark 3.5.6◦]
are measure equivalence invariants, it follows that for n = 2, 3, the group actions (H ⋊G) y (K ×X)
and (bK ⋊ G) y (bH × X) are not stably orbit equivalent. Hence, the corresponding group measure
space Cartan subalgebras are not conjugate by an automorphism either.
5.6 Amalgamated free products over groups with the Haagerup property
We mention that the result in Theorem 5.2 also holds for certain amalgamated free products
Γ = Γ1 ∗Σ Γ2 over non-amenable groups Σ with the Haagerup property, once the free ergodic p.m.p.
action Γ y (X, µ) is such that L∞(X) ⋊ Σ still has the Haagerup property. This latter condition
is not automatic, but holds for plain Bernoulli actions by [CSV09, Theorem 1.1]. More precisely,
apart from the Haagerup property of L∞(X) ⋊ Σ, one has to assume that Γ = Γ1 ∗Σ Γ2 is such that
Γ1 admits an infinite subgroup with property (T), that Σ has the Haagerup property, that Σ 6= Γ2
and that the 'malnormality' condition 5.1.3 holds.
In order to prove such a statement, it suffices to observe that Lemma 3.1 still holds when M0 has
property (T) and P has the Haagerup property. In the proof of Theorem 5.2, Lemma 3.1 is applied
to P = L∞(X) ⋊ Σ.
22
6 Stable W∗-superrigidity theorems
6.1 W∗-superrigidity
Let Γ y (X, µ) be a p.m.p. action. Denote A = L∞(X) and denote by (σg)g∈Γ the corresponding
group of automorphisms of A. We denote by Z1(Γ y X) the abelian group of scalar 1-cocycles for
the action Γ y X, i.e. the group of functions ω : Γ → U (A) : g 7→ ωg satisfying ωgh = ωgσg(ωh) for
all g, h ∈ Γ.
Let ∆ : X → Y be a conjugacy between the free ergodic p.m.p. actions Γ y X and Λ y Y , with
corresponding group isomorphism δ : Γ → Λ. Define the isomorphism ∆∗ : L∞(X) → L∞(Y ) :
∆(a) = a ◦ ∆−1. Whenever ω ∈ Z1(Γ y A), we get an isomorphism
θ : L∞(X) ⋊ Γ → L∞(Y ) ⋊ Λ : θ(aug) = ∆∗(aωg) uδ(g)
for all a ∈ L∞(X) , g ∈ Γ .
(6.1)
Definition 6.1. We call a free ergodic p.m.p. action Γ y (X, µ) W∗-superrigid if the following
property holds. If Λ y (Y, η) is a free ergodic p.m.p. action and π : L∞(X) ⋊ Γ → L∞(Y ) ⋊ Λ is an
isomorphism, then the groups Γ and Λ are isomorphic, their actions Γ y X, Λ y Y are conjugate
and, up to a unitary conjugacy, π is of the form (6.1).
Actually, the notion of stable W∗-superrigidity is more natural (see Definition 6.4 below), allowing
in the correct way for amplifications. Indeed, in order for Γ y (X, µ) to be W∗-superrigid in the
above sense, Γ should not have finite normal subgroups, which is a somewhat restrictive assumption.
We start with the following more precise version of Theorem 1.2.
Theorem 6.2. Let n ≥ 3 and denote by Tn the subgroup of upper triangular matrices in PSL(n, Z).
Put Γ = PSL(n, Z) ∗Tn PSL(n, Z). Then, every free ergodic p.m.p. action Γ y (X, µ) with the prop-
erty that all finite index subgroups of Tn act ergodically on (X, µ), is W∗-superrigid. In particular,
all free p.m.p. mixing actions of Γ are W∗-superrigid.
Proof. Take a free p.m.p. action Γ y (X, µ) such that the restriction to every finite index subgroup
of Tn is ergodic. Assume that L∞(X) ⋊ Γ = L∞(Y ) ⋊ Λ for some free ergodic p.m.p. action
Λ y (Y, η). By Theorem 5.2, we can unitarily conjugate L∞(Y ) onto L∞(X). Hence, Γ y (X, µ)
and Λ y (Y, η) are orbit equivalent. But then, [Ki09, Theorem 1.4] yields the conclusion of the
theorem.
6.2 Stable W∗-superrigidity
In order to introduce the more natural notion of stable W∗-superrigidity, we first discuss stable
If M is a II1 factor and t > 0, the amplification M t is defined as
isomorphism of II1 factors.
M t := p(Mn(C) ⊗ M )p, where p ∈ Mn(C) ⊗ M is a projection satisfying (Tr ⊗τ )(p) = t. Note that
M t is uniquely defined up to unitary conjugacy.
Definition 6.3. Let M and N be II1 factors. A stable isomorphism between M and N is an
isomorphism π : N → M t for some t > 0.
Given the isomorphism π : N → M t and a projection p ∈ Mn(C) ⊗ M with (Tr ⊗τ )(p) = t, define
the Hilbert space Hπ := (cid:0)M1,n(C) ⊗ L2(M )(cid:1)p. The formula a · ξ · b := aξπ(b) turns Hπ into an
M -N -bimodule such that the right N -action equals the commutant of the left M -action and vice
versa.
23
We equivalently define a stable isomorphism between M and N as being an M -N -bimodule MHN
such that the right N -action equals the commutant of the left M -action (and, equivalently, vice
versa). Every stable isomorphism MHN is unitarily equivalent with MHπ
N for an isomorphism
π : N → M t that is uniquely determined up to unitary conjugacy.
We call the number t > 0 the compression constant of the stable isomorphism MHπ
N.
Stable isomorphisms can be composed using the Connes tensor product. Of course, we have
Hπ ⊗N Hη ∼= H(id⊗π)η .
Fix free ergodic p.m.p. actions Γ y (X, µ) and Λ y (Y, η). Stabilizing the notions of W∗-
equivalence, orbit equivalence and conjugacy (see the beginning of Section 2), we introduce the
following terminology.
First recall that a free p.m.p. action Γ y (X, µ) is said to be induced from Γ0 y X0 if Γ0 < Γ is a
finite index subgroup, X0 ⊂ X is a non-negligible Γ0-invariant subset and, up to measure zero, the
sets g · X0, g ∈ Γ/Γ0, form a partition of X.
• We call stable W∗-equivalence between the actions Γ y (X, µ) and Λ y (Y, η), any stable
isomorphism between L∞(X) ⋊ Γ and L∞(Y ) ⋊ Λ.
• We call stable orbit equivalence between the actions Γ y (X, µ) and Λ y (Y, η), any stable
isomorphism between the orbit equivalence relations R(Γ y X) and R(Λ y Y ), i.e. any
isomorphism ∆ : X1 → Y1 between non-negligible subsets X1 ⊂ X, Y1 ⊂ Y satisfying
for a.e. x ∈ X1. We call µ(X1)/η(Y1) the compression constant of the stable orbit equivalence.
∆(X1 ∩ Γ · x) = Y1 ∩ Λ · ∆(x)
• We call stable conjugacy between the actions Γ y (X, µ) and Λ y (Y, η), any conjugacy
between the actions Γ0
H where Γ y X, Λ y Y are induced from
G
Γ0 y X0, Λ0 y Y0 and where G ⊳ Γ0, H ⊳ Λ0 are finite normal subgroups. The number
H [Λ:Λ0]
G [Γ:Γ0] is called the compression constant of the stable conjugacy.
G and Λ0
y X0
y Y0
H
y X0
From stable conjugacy to stable orbit equivalence. If Γ y (X, µ) is induced from Γ0 y X0 and
if G ⊳ Γ0 is a finite normal subgroup, the canonical stable orbit equivalence between Γ y (X, µ)
and Γ0
G is defined by taking a fundamental domain X1 ⊂ X0 for the action G y X0 and
G
restricting the quotient map X0 → X0
G to X1. The compression constant is (G [Γ : Γ0])−1. The
stable orbit equivalence associated with a stable conjugacy between Γ y (X, µ) and Λ y (Y, η) is
y Y0
defined as the composition of the canonical stable orbit equivalences with Γ0
H ,
G
together with the conjugacy between both actions.
From stable orbit equivalence to stable W∗-equivalence. Let ∆ : X1 → Y1 be a stable orbit equiva-
lence and denote by p ∈ L∞(X), q ∈ L∞(Y ) the projections with support X1, Y1. Then ∆ gives rise
to a canonical isomorphism π∆ : q(L∞(Y ) ⋊ Λ)q → p(L∞(X) ⋊ Γ)p satisfying π∆(b) = b ◦ ∆ for all
b ∈ L∞(Y1). The isomorphism π∆ amplifies to a stable W∗-equivalence L∞(Y )⋊Λ → (L∞(X)⋊Γ)t,
with t = µ(X1)/η(Y1), that we still denote by π∆.
Definition 6.4. A free ergodic p.m.p. action Γ y (X, µ) is said to be stably W∗-superrigid if the
following holds. Whenever π is a stable W∗-equivalence between Γ y (X, µ) and an arbitrary free
ergodic p.m.p. action Λ y (Y, η), it follows that the actions are stably conjugate and that π equals
the composition of
G , resp. Λ0
y X0
H
24
• the canonical stable W∗-equivalence given by the stable conjugacy,
• the automorphism of L∞(X) ⋊ Γ given by an element of Z1(Γ y X),
• an inner automorphism of L∞(X) ⋊ Γ.
Let Γ y (X, µ) be stably W∗-superrigid. If moreover Γ has no finite normal subgroups and if finite
index subgroups of Γ still act ergodically on (X, µ), then Γ y (X, µ) is W∗-superrigid in the sense
of Section 1.
For certain families of group actions Γ y (X, µ), all 1-cocycles with values in S1 are known to be
cohomologous to a group morphism and then we may assume that the corresponding automorphism
of L∞(X) ⋊ Γ is implemented by a character Γ → S1.
From stable W∗-equivalence to stable orbit equivalence. Let Γ y (X, µ) be a free ergodic p.m.p.
action satisfying the conclusion of Theorem 5.2. Whenever π is a stable W∗-equivalence between
Γ y (X, µ) and an arbitrary free ergodic p.m.p. action Λ y (Y, η), we then find a stable orbit equiv-
alence ∆ between Γ y X and Λ y Y such that π equals the composition of π∆, the automorphism
of L∞(X) ⋊ Γ given by an element of Z1(Γ y X) and an inner automorphism.
From stable orbit equivalence to stable conjugacy. Let Γ y (X, µ) be a free ergodic p.m.p. action
that is cocycle superrigid with arbitrary countable target groups (see paragraph 2.4 for terminology).
Whenever ∆ is a stable orbit equivalence between Γ y (X, µ) and an arbitrary free ergodic p.m.p.
action Λ y (Y, η), it follows (see [Po05, Proposition 5.11], [Va06, Lemma 4.7]) that the actions are
stably conjugate and that ∆ equals the composition of the canonical stable orbit equivalence given
by the stable conjugacy and an inner automorphism, i.e. an automorphism ∆0 of (X, µ) satisfying
∆0(x) ∈ Γ · x for a.e. x ∈ X. Even more so, it actually follows that there exists a finite normal
subgroup G ⊳ Γ and that Λ y Y is induced from Λ0 y Y0 such that the actions Γ
G and
G
Λ0 y Y0 are conjugate.
y X
Summarizing the previous two paragraphs, we have proven the following lemma.
Lemma 6.5. Let Γ y (X, µ) be a free ergodic p.m.p. action that satisfies the conclusion of Theorem
5.2 and that is cocycle superrigid with arbitrary countable target groups, as well as with target group
S1. Then, Γ y (X, µ) is stably W∗-superrigid. Even more precisely, we have the following.
Let Λ y (Y, η) be an arbitrary free ergodic p.m.p. action and π : L∞(X) ⋊ Γ → (L∞(Y ) ⋊ Λ)t a
∗-isomorphism for some t > 0. Then, Λ y Y is induced from Λ0 y Y0 and there exist
• a finite normal subgroup G ⊳ Γ and a group isomorphism δ : Γ
G → Λ0,
• a measure space isomorphism ∆ : X
G → Y0 conjugating the actions, i.e. ∆(g · x) = δ(g) · ∆(x)
for all g ∈ Γ
G and a.e. x ∈ X,
• a character ω : Γ → S1,
such that t = G
[Λ:Λ0] and such that, after a unitary conjugacy, π equals the composition of
• the automorphism πω of L∞(X) ⋊ Γ given by πω(aug) = ω(g)aug,
• the canonical isomorphism L∞(X) ⋊ Γ →(cid:0)L∞(cid:0) X
• the isomorphism π∆ : L∞(cid:0) X
G(cid:1) ⋊ Γ
G(cid:1) ⋊ Γ
G(cid:1)n where n = G,
G → L∞(Y0) ⋊ Λ0 given by π∆(aug) = (a ◦ ∆−1)uδ(g),
25
• the canonical isomorphism L∞(Y0) ⋊ Λ0 → (L∞(Y ) ⋊ Λ)
1
m where m = [Λ : Λ0].
Remark 6.6. Let Γ y (X, µ) be a stably W∗-superrigid, free ergodic p.m.p. action.
If Γ has
no finite normal subgroups and if the finite index subgroups of Γ act ergodically on (X, µ), then
L∞(X) ⋊ Γ has trivial fundamental group.
Recall that given any action Γ y I of a countable group Γ on a countable set I, the generalized
Bernoulli action with base probability space (X0, µ0), is defined as Γ y (X0, µ0)I where (g · x)i =
xg−1·i for all g ∈ Γ, x ∈ X I
Recall that given any orthogonal representation π : Γ → O(KR) of Γ on the real Hilbert space KR,
the Gaussian functor allows to define the Gaussian p.m.p. action of Γ on the Gaussian probability
space defined by KR.
0 and i ∈ I.
Theorem 6.7. Let Γ1, Γ2 be countable groups with a common infinite amenable subgroup Σ. As-
sume that Σ is a proper, normal subgroup of Γ2 and that there exist g1, . . . , gk ∈ Γ1 such that
Put Γ = Γ1 ∗Σ Γ2.
giΣg−1
i
is finite .
k\i=1
1. If Γ1 admits a non-amenable normal subgroup H with the relative property (T), then all of
the following actions are stably W∗-superrigid.
• Every free p.m.p. action Γ y (X, µ) whose restriction to Γ1 is a generalized Bernoulli
action Γ1 y (X0, µ0)I with the property that both H · i and Σ · i are infinite for all i ∈ I.
• Every free p.m.p. action Γ y (X, µ) whose restriction to Γ1 is a Gaussian action defined
by an orthogonal representation π : Γ1 → O(KR) with the property that both restrictions
πH and πΣ have no non-zero finite dimensional subrepresentations.
2. Suppose that Γ1 admits non-amenable commuting subgroups H and H ′ such that H is normal
in Γ1. If Γ y (X, µ) is a free p.m.p. action whose restriction to Γ1 is a generalized Bernoulli
action Γ1 y (X0, µ0)I with the properties that Stab i ∩ H ′ is amenable for all i ∈ I and that
both H · i and Σ · i are infinite for all i ∈ I, then Γ y (X, µ) is stably W∗-superrigid.
In particular, for all groups Γ mentioned in this theorem, the plain Bernoulli action Γ y (X0, µ0)Γ
is stably W∗-superrigid.
More precisely, all actions Γ y X appearing in the theorem satisfy the conclusions of Lemma 6.5.
Before proving Theorem 6.7, we provide the following concrete examples of stably W∗-superrigid
group actions.
Example 6.8.
1. Denote by Tn < PSL(n, Z) the subgroup of upper triangular matrices and
assume n ≥ 3. For an arbitrary non-trivial group Λ and an arbitrary infinite subgroup
Σ < Tn, put Γ = PSL(n, Z) ∗Σ (Σ × Λ). Whenever Γ y I is such that Σ · i is infinite for all
i ∈ I, the generalized Bernoulli actions Γ y (X0, µ0)I are stably W∗-superrigid.
2. Let Σ be an infinite amenable subgroup of the non-amenable group H. Assume that H is
finitely generated and that Σ ∩ Z(H) = {e}. Let Λ be an arbitrary non-trivial group. Define
Γ1 = H × H, view Σ as a subgroup of Γ1 diagonally and put Γ = Γ1 ∗Σ Γ2. Whenever Γ y I
is such that Σ · i and (H × {e}) · i is infinite for all i ∈ I, the generalized Bernoulli actions
Γ y (X0, µ0)I are stably W∗-superrigid.
26
Actually, exploiting the full strength of Theorem 6.7, many more examples of stably W∗-superrigid
actions can be given. Given an arbitrary faithful(5) p.m.p. action Λ y (X0, µ0), we construct a
stably W∗-superrigid, free ergodic p.m.p. action Γ y (X, µ) such that Λ is a subgroup of Γ and
such that Λ y (X0, µ0) is a quotient of the restriction of Γ y (X, µ) to Λ.
Recall first the construction of a co-induced action. Let Λ < Γ be a subgroup and Λ y (X0, µ0) a
p.m.p. action. Choose a map π : Γ → Λ satisfying π(gh) = π(g)h for all g ∈ Γ, h ∈ Λ. Define the
1-cocycle ω for the action Γ y Γ/Λ with values in Λ by the formula
ω(g, hΛ) = π(gh)π(h)−1 .
Different choices of π lead to cohomologous 1-cocycles ω. Define the probability space (X, µ) :=
(X0, µ0)Γ/Λ and the p.m.p. action Γ y (X, µ) given by
(g−1 · x)hΛ := ω(g, hΛ)−1 · xghΛ for all g, h ∈ Λ, x ∈ X .
We call Γ y (X, µ) the co-induced action of Λ y (X0, µ0) to Γ. Different choices of π lead to
conjugate actions.
Note that we can choose π such that π(h) = h for all h ∈ Λ. Then, the quotient map θ : X → X0 :
x 7→ xeΛ satisfies θ(h · x) = h · θ(x) for all h ∈ Λ, x ∈ X. Hence, Λ y X0 arises as a quotient of the
restriction of Γ y X to Λ.
Example 6.9. Take Γ = PSL(n, Z) ∗Σ (Σ × Λ) as in Example 6.8.1 or take Γ = (H × H) ∗Σ (Σ × Λ)
as in Example 6.8.2. Let Λ y (X0, µ0) be an arbitrary faithful p.m.p. action. Then, the co-induced
action Γ y (X, µ) is stably W∗-superrigid.
In the first example, put Γ1 = PSL(n, Z) and in the second example, put Γ1 = H × H. In order to
apply Theorem 6.7, it suffices to observe that the restriction of the co-induced action Γ y (X, µ)
to Γ1 is a Γ1-Bernoulli action. Indeed, the action Γ1 × Λ y Γ by left-right multiplication is free.
Therefore, we can choose π : Γ → Λ satisfying π(ghk) = π(h)k for all g ∈ Γ1, h ∈ Γ, k ∈ Λ.
Associated with π is the 1-cocycle ω for Γ y Γ/Λ, which now satisfies ω(g, hΛ) = e for all g ∈
Γ1, h ∈ Γ. Hence, the restriction Γ1 y (X, µ) is precisely the Bernoulli action Γ1 y (X0, µ0)Γ/Λ.
The latter can be seen as the plain Bernoulli action Γ1 y Y Γ1
0 , where Y0 := X Γ1\Γ/Λ
0
.
Example 6.10. Let Γ1 be an infinite group with property (T), Σ < Γ1 an infinite subgroup with
the Haagerup property and Λ an arbitrary non-trivial group. Put Γ = Γ1 ∗Σ (Σ × Λ). Using
paragraph 5.6 instead of Theorem 5.2, it follows that the plain Bernoulli action Γ y (X0, µ0)Γ is
stably W∗-superrigid. Similarly (cf. Example 6.9), the co-induced action Γ y (X, µ) of an arbitrary
faithful p.m.p. action Λ y (X0, µ0), follows stably W∗-superrigid.
6.3 Cocycle superrigidity for actions of amalgamated free products
Recall from paragraph 2.4 the notion of cocycle superrigidity. Theorem 6.7 will be proven as a
consequence of Theorem 5.2 and a number of cocycle superrigidity theorems from [Po05, Po06a]
that we recall here.
We first prove the following permanence lemma for L-cocycle superrigidity. It is a direct conse-
quence of techniques in [Po05, Section 3], but we give a short and full proof for the convenience of
the reader. We also deal with arbitrary finite index subgroups, which will be useful in Section 7.
(5)This means that non-trivial group elements act non-trivially.
27
Lemma 6.11. Let Γ = Γ1 ∗Σ Γ2 be an amalgamated free product where Σ ⊳Γ2 is a normal subgroup.
Let Γ y (X, µ) be a p.m.p. action and assume that its restriction Σ y (X, µ) is weakly mixing.
Let L be a Polish group whose topology is induced by a separable complete bi-invariant metric (for
instance, L ∈ Ufin).
If for every finite index subgroup Γ′
Γ′ y (X, µ) is L-cocycle superrigid for all finite index subgroups Γ′ < Γ.
y (X, µ) is L-cocycle superrigid, then
1 < Γ1 the action Γ′
1
Proof. From [Po05, Section 3] we deduce the following two principles. Let L be as in the lemma,
Λ y (X, µ) a p.m.p. action and Λ0 < Λ a subgroup such that Λ0 y (X, µ) is weakly mixing. Let
ω : Λ × X → L be a 1-cocycle such that ω(h, x) = δ(h) for all h ∈ Λ0, where δ : Λ0 → L is a group
morphism.
If Λ0 is normal in Λ, then ω is a group morphism on the whole of Λ.
Principle 1.
Indeed,
choose g ∈ Λ and put ψ(x) = ω(g, x). We have to prove that ψ is essentially constant. But
ψ(h · x) = δ(ghg−1) ψ(x) δ(h)−1 for all h ∈ Λ0 and a.e. x ∈ X. By [PV08b, Lemma 5.4] the function
ψ is essentially constant.
Principle 2. If ω is cohomologous to a group morphism on Λ, then ω is already a group morphism on
Λ. Indeed, by assumption we find a measurable map ψ : X → L and a group morphism ρ : Λ → L
such that ω(g, x) = ψ(g · x)−1ρ(g)ψ(x) for all g ∈ Λ and a.e. x ∈ X. It suffices to prove that ψ
is essentially constant. But ψ(h · x) = ρ(h)ψ(x)δ(h)−1 for all h ∈ Λ0 and a.e. x ∈ X. Again by
[PV08b, Lemma 5.4] the function ψ is essentially constant.
Take a finite index subgroup Γ′ < Γ and a 1-cocycle ω : Γ′ × X → L. We have to prove that ω is
cohomologous to a group morphism. Since Γ1 ∩ Γ′ is a finite index subgroup of Γ1, we may assume
that ω is already a group morphism on Γ1 ∩ Γ′. We prove that ω is then a group morphism on the
whole of Γ′.
Take a finite index subgroup Γ′′ < Γ′ such that Γ′′ is normal in Γ. Define Γ′
Σ′ = Σ ∩ Γ′′. Note that Σ′ < Σ has finite index and that Γ′
Also Σ′ y (X, µ) is still weakly mixing.
ig−1 for all g ∈ Γ and i = 1, 2. For
We prove by induction on g that ω is a group morphism on gΓ′
g = 0 we already know that ω is a group morphism on Γ′
1. In particular, ω is a group morphism
on Σ′ and by principle 1, also on Γ′
2. Assume that the statement is true for all elements of length
n − 1. Take g ∈ Γ with g = n. Write g = g0h where g0 = n − 1 and h ∈ Γ1 ∪ Γ2. If h ∈ Γ1, we
have gΓ′
1g−1 and, in
2g−1. Next, if h ∈ Γ2, we
particular, on gΣ′g−1. By principle 1, ω is also a group morphism on gΓ′
have gΓ′
2g−1 and, in
particular, on gΣ′g−1. Consider the 1-cocycle µ : Γ′
1 × X → L given by µ(h, x) = ω(ghg−1, g · x).
Since Γ′
y (X, µ) is L-cocycle superrigid, it follows that µ is cohomologous to a group morphism.
1
But µ is already a group morphism on Σ′. By principle 2, µ is a group morphism on Γ′
1 and hence,
ω is a group morphism on gΓ′
Define the subgroup Γ′′′ < Γ′ generated by gΓ′
ig−1 for all g ∈ Γ and i = 1, 2. We have already
proven that ω is a group morphism on Γ′′′. By construction, Γ′′′ is normal in Γ′. So, by principle 1,
ω is a group morphism on Γ′.
0 . By the induction hypothesis, ω is a group morphism on gΓ′
0 . By the induction hypothesis, ω is a group morphism on gΓ′
i = Γi ∩ Γ′′ and
2 are normal subgroups.
1g−1 = g0Γ′
1g−1
i ⊳ Γi, Σ′ ⊳ Γ′
2g−1 = g0Γ′
2g−1
1g−1.
Theorem 6.12. Let Γ = Γ1 ∗Σ Γ2 be an amalgamated free product over an infinite subgroup Σ that
is normal in Γ2. Let Γ y (X, µ) be a free ergodic p.m.p. action. Assume that the group Γ and its
action Γ y (X, µ) satisfy condition 1 or condition 2 in Theorem 6.7. Then, for every finite index
subgroup Γ′ < Γ, the action Γ′ y (X, µ) is Ufin-cocycle superrigid.
28
Proof. Both conditions 1 and 2 imply that Σ y (X, µ) is weakly mixing. By Lemma 6.11, it suffices
to prove that Γ′
1 < Γ1. Under
1
conditions 1, this is a consequence of [Po05, Theorem 0.1]. Under conditions 2, this follows from
[Po06a, Theorem 1.1].
y (X, µ) is Ufin-cocycle superrigid for every finite index subgroup Γ′
Finally, for later use in Section 7 we record the following lemma about weak 1-cocycles (cf. [PS03,
Theorem 4.1]).
Lemma 6.13. Let Γ y (X, µ) be a p.m.p. action that is Ufin-cocycle superrigid. Assume that
Ω : Γ × Γ → S1 is a scalar 2-cocycle on Γ and that ω : Γ × X → S1 is a measurable map satisfying
ω(gh, x) = Ω(g, h)ω(g, h · x)ω(h, x)
for all g, h ∈ Γ and a.e. x ∈ X. Then, there exist a measurable function ϕ : X → S1 and a map
δ : Γ → S1 such that
ω(g, x) = ϕ(g · x)δ(g)ϕ(x)
for all g ∈ Γ and a.e. x ∈ X. In particular, Ω is a coboundary: Ω(g, h) = δ(gh)δ(g)δ(h) for all
g, h ∈ Γ.
Proof. Form the twisted group von Neumann algebra N := LΩ(Γ) generated by unitaries ug, g ∈ Γ
satisfying uguh = Ω(g, h)ugh and equipped with a trace τ satisfying τ (ug) = 0 if g 6= e. Denote
by L the closed subgroup of U (N ) consisting of the unitaries λug with λ ∈ S1, g ∈ Γ. Writing
µ(g, x) = ω(g, x)ug, it follows that µ is a 1-cocycle for Γ y (X, µ) with values in L.
Since Γ y (X, µ) is Ufin-cocycle superrigid, we find a measurable map θ : X → L and a group
morphism η : Γ → L such that µ(g, x) = θ(g · x)η(g)θ(x)−1 for all g ∈ Γ and a.e. x ∈ X. Denoting
by ϕ(x), resp. δ(g), the S1-part of θ(x), resp. η(g), the lemma is proven.
6.4 Proof of Theorem 6.7
Proof of Theorem 6.7. Since Γ belongs to G the conclusion of Theorem 5.2 holds. By Theorem 6.12,
Γ y (X, µ) is Ufin-cocycle superrigid. In particular, Γ y (X, µ) is cocycle superrigid with arbitrary
countable target groups and with target group S1. So, the conclusions follow from Lemma 6.5.
6.5 Strong rigidity for group von Neumann algebras
Following up Connes' rigidity conjecture [Co80b], Jones asked in [Jo00] whether every isomorphism
between property (T) group von Neumann algebras L(G1) and L(G2) essentially comes from an
isomorphism between the groups G1 and G2 (cf. [Po06b, Statement 3.2']).
In [PV06, Theorem
7.13], we provided a family of (generalized) wreath product groups satisfying such a strong rigidity
result. The following corollary enlarges this family of groups.
i, j ∈ I with i 6= j. Denote by M the family of associated wreath product groups G =(cid:0)Z/2Z(cid:1)(I)
Corollary 6.14. Consider all group actions Γ y I on countable sets covered by Theorem 6.7.
Assume moreover that Γ has no finite normal subgroups and that (Stab i) · j is infinite for all
⋊ Γ.
If G1, G2 ∈ M, t > 0 and if π : L(G1) → L(G2)t is a stable isomorphism between L(G1) and L(G2),
then the groups G1, G2 are isomorphic through an isomorphism δ : G1 → G2, the amplification t
equals 1 and there exist a unitary w ∈ L(G2) and a group morphism ω : G1 → S1 such that
π(ug) = ω(g) w vδ(g)w∗
for all g ∈ G1 .
29
Here, (ug)g∈G1 and (vs)s∈G2 denote the natural unitaries generating L(G1) and L(G2).
In particular, the automorphism group of L(G) for G ∈ M, is generated by Aut(G), the group of
characters Char G and the inner automorphisms.
To prove Corollary 6.14, we use the following lemma, whose proof is contained in [PV06, Theorem
5.4]. Observe that Lemma 6.15 implies in particular that, whenever Γ y I is such that (Stab j) · k
is infinite for all j 6= k, the generalized Bernoulli actions Γ y (X0, µ0)I and Γ y (Y0, ρ0)I are
conjugate if and only if the base probability spaces are isomorphic. Such a statement is false in
general for plain Bernoulli actions Γ y (X0, µ0)Γ, see [St75].
Lemma 6.15. Let Γi y Ii, i = 1, 2. Assume that for both actions, (Stab j) · k is infinite for all
j 6= k. Consider the generalized Bernoulli actions
Γ1 y (X, µ) = (X0, µ0)I1
and Γ2 y (Y, ρ) := (Y0, ρ0)I2 .
If ∆ : X → Y is an isomorphism of probability spaces and δ : Γ1 → Γ2 an isomorphism of groups
such that ∆(g · x) = δ(g) · ∆(x) for all g ∈ Γ1 and almost all x ∈ X, there exists a bijection
η : I1 → I2 and there exist isomorphisms ∆i : X0 → Y0 of the base probability spaces such that
• η(g · i) = δ(g) · η(i) and ∆g·i = ∆i for all i ∈ I1, g ∈ Γ1,
• (cid:0)∆(x)(cid:1)η(i) = ∆i(xi) for all i ∈ I1 and almost all x ∈ X.
Proof of Corollary 6.14. Write Gi = (cid:0)Z/2Z(cid:1)(Ii)
L(cid:16)(cid:0)Z/2Z(cid:1)(Ii)(cid:17). Let π : L(G1) → L(G2)t be a ∗-isomorphism. By Theorem 6.7, we get that t = 1
and we get the existence of an isomorphism δ : G1 → G2, a character ω : Λ1 → S1 and a unitary
w ∈ L(G2) such that
⋊ Λi and identify L(Gi) = Ai ⋊ Λi, where Ai =
π(aug) = ω(g) w α(a) vδ(g) w∗ ,
where α : A1 → A2 is a ∗-isomorphism satisfying α(σg(a)) = σδ(g)(a) for all g ∈ Λ1 and a ∈ A1.
Lemma 6.15 describes the form of πA1. Since L(Z/2Z) has precisely two automorphisms, the
identity and the multiplication with the non-trivial character on Z/2Z, we are done.
6.6 Counterexamples to W∗-superrigidity
There are many free ergodic p.m.p. actions Γ y (X, µ) with Γ ∈ G and such that Γ y (X, µ)
is orbit equivalent with a large family of non-conjugate group actions. So, in these cases, the II1
factor L∞(X) ⋊ Γ has many group measure space decompositions (up to conjugacy of the actions),
but all of them have the same Cartan subalgebra (up to unitary conjugacy).
First of all, by [Ga05, PME6], if Γi and Γ′
and Γ′
Next, let Γ1 be an arbitrary group and Γ2, Γ′
Γ′ = Γ1∗Γ′
for all non-trivial base probability spaces (X0, µ0) and (Y0, η0).
2 infinite amenable groups. Denote Γ = Γ1 ∗ Γ2 and
are orbit equivalent
2. By [Bo09], the Bernoulli actions Γ y (X0, µ0)Γ and Γ′ y (Y0, η0)Γ′
i admit orbit equivalent actions (i = 1, 2), then Γ1 ∗ Γ2
1 ∗ Γ′
2 also admit orbit equivalent actions.
30
7 W∗-superrigidity for finite index bimodules
By definition, the action Γ y (X, µ) is stably W∗-superrigid if, for all free ergodic p.m.p. actions
Λ y (Y, η), stable isomorphism of L∞(X)⋊Γ and L∞(Y )⋊Λ implies, in a sense, stable isomorphism
of the groups Γ, Λ and their respective actions.
For certain of the actions listed in Theorem 6.7, one can go even further and prove that, whenever
H is a finite index bimodule between L∞(X) ⋊ Γ and L∞(Y ) ⋊ Λ, the groups Γ, Λ and their actions
are, in the following precise sense, virtually isomorphic, with H being implemented by this virtual
conjugacy and a finite dimensional unitary representation.
i=1 giΣg−1
i finite, does not provide sufficient absence
of normality of Σ < Γ. Therefore, we make in the following theorem, a much stronger (and certainly
non-optimal) assumption.
However, the existence of g1, . . . , gn ∈ Γ withTn
Theorem 7.1. Let Γ y (X, µ) be any of the actions listed in Theorem 6.7 and assume that Γ
admits an infinite index subgroup G such that gΣg−1 ∩ Σ is finite for all g ∈ Γ − G. If Λ y (Y, η) is
an arbitrary free ergodic p.m.p. action and H is a finite index(cid:0)L∞(X)⋊Γ(cid:1) -- (cid:0)L∞(Y )⋊Λ(cid:1) -- bimodule,
there exist
• a finite index subgroup Γ′ < Γ,
• a finite subgroup H < Aut(X, µ) such that gHg−1 = H for all g ∈ Γ′,
• finite index subgroups Λ′′ < Λ′ < Λ with Λ y Y being induced from Λ′ y Y ′ for some Y ′ ⊂ Y ,
such that Λ′′ ∼= Γ′
isomorphism ∆ : X
Γ′∩H and such that the action Λ′′ y Y ′ is conjugate with
H → Y ′ of probability spaces and a group isomorphism δ :
y X
Γ′
Γ′∩H
Γ′
Γ′∩H → Λ′′.
H , through an
Moreover, if H is irreducible, there exists an irreducible finite dimensional unitary representation
π : Γ′ → U (Ck) such that H is the composition of
• the (cid:0)L∞(X) ⋊ Γ(cid:1) -- (cid:0)L∞(X) ⋊ Γ′(cid:1) -- bimodule on the Hilbert space M1,k(C) ⊗ L2(M ), where
M = L∞(X) ⋊ Γ and where
z · ξ · (aug) = zξ(π(g) ⊗ aug)
for all z ∈ M, a ∈ L∞(X), g ∈ Γ′ ,
(7.1)
• the natural(cid:0)L∞(X)⋊Γ′(cid:1) -- (cid:0)L∞( X
H )⋊ Γ′
Γ′∩H(cid:1) -- bimodule on the Hilbert space L2(M0)pΓ′∩H, where
M0 = L∞(X) ⋊ Γ′
and
pΓ′∩H = Γ′ ∩ H−1 Xg∈Γ′∩H
ug ,
• the(cid:0)L∞( X
Λ and where
H )⋊ Γ′
Γ′∩H(cid:1) -- (cid:0)L∞(Y )⋊Λ(cid:1) -- bimodule on the Hilbert space L2(N ), where N = L∞(Y )⋊
(aug) · ξ · z = (a ◦ ∆−1)uδ(g)ξz
for all a ∈ L∞( X
H ), g ∈ Γ′
Γ′∩H , z ∈ N .
Example 7.2. Let Σ be an amenable group and Γ = Γ1 ∗Σ Γ2. Assume that Γ y I with Σ · i
being infinite for every i ∈ I. If one of the following conditions is satisfied, the generalized Bernoulli
action Γ y (X0, µ0)I satisfies the conclusions of Theorem 7.1.
31
1. Let Σ = Z and embed Σ into Γ1 = SL(n, Z), n ≥ 3, in the upper right corner. Assume that
Σ ⊳ Γ2 is a proper normal subgroup. Then, Theorem 7.1 applies by putting G = G0 ∗Σ Γ2,
where G0 consists of those matrices A with Ai1 = 0 for all 2 ≤ i ≤ n.
2. Let Σ be a common subgroup of Λ0, Λ1, Γ2 and assume that Σ is normal in Γ2, Σ 6= Γ2 and
Λ0 is non-amenable. Let Λ2 be an arbitrary non-trivial group. Put Γ1 = Λ0 × (Λ1 ∗ Λ2) and
view Σ < Γ1 diagonally. Theorem 7.1 applies by putting G = (Λ0 × Λ1) ∗Σ Γ2. Assume that
Λ0 or Λ1 ∗ Λ2 acts with infinite orbits on I.
In the proof of Theorem 7.1 we make use of the following lemma having some independent interest.
Lemma 7.3. Let Γ y (X, µ) and Λ y (Y, η) be free p.m.p. actions. Assume that Γ y (X, µ) is
cocycle superrigid with finite target groups and that all finite index subgroups of Γ act ergodically
on (X, µ). Let ∆ : X → Y be an m-to-1 measure preserving map satisfying ∆(g · x) = δ(g) · x for
some surjective group morphism δ : Γ → Λ.
Then there exists a group H with m elements and a free p.m.p. action H y (X, µ) such that,
viewing both Γ and H as subgroups of Aut(X, µ), we have
• gHg−1 = H for all g ∈ Γ,
• Ker δ = Γ ∩ H,
• ∆ is the composition of the canonical quotient map X → X
H with a conjugacy of the actions
Γ
Γ∩H
y X
H and Λ y Y .
Proof. Partition X into disjoint measurable subsets X1, . . . , Xm such that the restrictions ∆i :=
∆Xi are isomorphisms of Xi onto Y scaling the measure by the factor 1/m.
Put J := {1, . . . , m}. For g ∈ Γ and almost every x ∈ X, define the map ω(g, x) : J → J such
that
ω(g, x)i = j
if and only if
g · ∆−1
i (∆(x)) ∈ Xj .
So, by construction
g · ∆−1
i (∆(x)) = ∆−1
ω(g,x)i(∆(g · x))
almost everywhere. It follows that ω(gh, x) = ω(g, h · x) ◦ ω(h, x) almost everywhere. Hence, ω(g, x)
is a permutation of J and ω is a 1-cocycle for the action Γ y X with values in the permutation
group Sm. Since Γ y (X, µ) is assumed to be cocycle superrigid with finite target groups, we find
a measurable map ϕ : X → Sm and a group morphism η : Γ → Sm such that, writing
Ti : X → X : Ti(x) = ∆−1
ϕ(x)i(∆(x)) ,
we have g · Ti(x) = Tη(g)i(g · x) almost everywhere. By construction, every Ti is locally a m.p.
isomorphism. Moreover, the range of Ti is globally invariant under the finite index subgroup
Ker η < Γ. Hence, all Ti are m.p. automorphisms of (X, µ).
Define the equivalence relation R on X as R := {(x, y) ∈ X × X ∆(x) = ∆(y)}. By construction,
every equivalence class has m elements and the graph of every Ti belongs to R. Finally, if Ti(x) =
Tj(x) for all x in a non-negligible subset U , we can make U smaller and assume moreover that
ϕ(x) = σ for all x ∈ U . Hence, ∆−1
σj (∆(x)) for all x ∈ U . So, σi = σj and hence,
i = j. We have shown that R is the disjoint union of the graphs of the Ti, i = 1, . . . , m.
σi (∆(x)) = ∆−1
32
Define H < Aut(X, µ) as the group generated by T1, . . . , Tm. By construction, all h ∈ H commute
with Ker η. Since Ker η acts ergodically on (X, µ), it follows that H acts freely on (X, µ). Since
the orbit equivalence relation of H y X is contained in R, we get that H ≤ m. But H contains
the distinct elements T1, . . . , Tm, so that H = {T1, . . . , Tm}.
It remains to prove that Ker δ = Γ ∩ H. Since ∆(g · x) = δ(g) · ∆(x) and Λ acts freely on Y , it
follows that an element g ∈ Γ belongs to Ker δ if and only if the graph of g belongs to R. It follows
that Ker δ ≤ m and that Γ ∩ H ⊂ Ker δ. Conversely, if g ∈ Ker δ, take i such that the graphs
of g and Ti intersect non-negligibly. Since Ker δ is a finite normal subgroup of Γ, it follows that g
commutes with a finite index subgroup of Γ. We have already seen that also Ti commutes with a
finite index subgroup of Γ. It follows that g and Ti coincide almost everywhere, i.e. g ∈ Γ ∩ H.
Proof of Theorem 7.1. As a preliminary step, note that Lemmas 3.1 and 3.2 remain valid when we
identify Q ⋊ Λ with a finite index subfactor of pM p. This generalization almost has the same proof
and we explain this briefly for the case of Lemma 3.1. The deformation ϕn of M still allows to
define the completely positive maps eϕn on Λ, which in turn lead to a new deformation θn of Q ⋊ Λ.
By Jones' tunnel construction and for the appropriate value of s > 0, we can view the amplification
M s as a finite index subfactor of Q ⋊ Λ. Hence, we find inside Q ⋊ Λ a von Neumann subalgebra
M0 with the relative property (T) and without injective direct summand. So, we can finish the
proof in the same way as for the original Lemma 3.1.
Write A = L∞(X) and M = A ⋊ Γ. Put B = L∞(Y ) and N = B ⋊ Λ. We may assume that MHN
is an irreducible finite index bimodule. Let γ : N → pM mp be a finite index inclusion such that
MHN
∼= M(cid:0)(M1,m(C) ⊗ L2(M ))p(cid:1)γ(N ) .
View M m as the crossed product (Mm(C) ⊗ A) ⋊ Γ where Γ acts trivially on Mm(C). As in Section
5.1, define for every 0 < ρ < 1, the unital completely positive map mρ on M m by mρ(aug) = ρgaug
for all a ∈ Mm(C) ⊗ A, g ∈ Γ.
In the situation where Γ1 has a non-amenable subgroup with the relative property (T), apply Lemma
3.1 -- as generalized in the first paragraph of the proof -- to the completely positive maps mρ, ρ → 1
and the von Neumann subalgebra P = (A ⋊ Σ)m. In the situation where Γ1 has two non-amenable
commuting subgroups H1, H2, we apply the generalized Lemma 3.2. Put ε = (tr ⊗τ )(p)/2072.
Combined with (5.1), we always find 0 < ρ < 1 and a sequence sk ∈ Λ such that
• kγ(vsk ) − mρ(γ(vsk ))k2 ≤ ε/2 for all k.
• kEP (xγ(vsk )y)k2 → 0 for all x, y ∈ M .
By Lemma 5.7 we get a 0 < ρ < 1 and a δ > 0 such that τ (γ(w)∗mρ(γ(w))) ≥ δ for all w ∈ U (B). By
Theorem 5.4 we have that γ(B) ≺M A ⋊ Σ or that NpM mp(γ(B))′′ ≺M A ⋊ Γi for some i = 1, 2. But
NpM mp(γ(B)) contains γ(N ) which has finite index in pM mp. Since Γi < Γ has infinite index, this
rules out the possibility that NpM mp(γ(B))′′ ≺M A ⋊ Γi. So, we have shown that γ(B) ≺M A ⋊ Σ.
We claim that γ(B) ≺M A. Assume the contrary. Our assumption on G, together with the
regularity of B ⊂ N , implies that γ(N ) ≺M A ⋊ G (cf. [Va07, Lemma 4.2]). Since γ(N ) ⊂ pM mp
is of finite index, while G < Γ has infinite index, this is a contradiction. So, we have shown that
γ(B) ≺M A.
In bimodule language, this means that H admits a non-zero A-B-subbimodule K which is finitely
generated as a left A-module. Take a finite index inclusion ψ : M → qN kq such that
MHN
∼= ψ(M )q(Ck ⊗ L2(N ))N .
33
The existence of K means that B ≺N ψ(A). Taking relative commutants (see [Va07, Lemma 3.5]),
it follows that ψ(A) ≺N B.
By Theorem 6.12 the action Γ′ y X is Ufin-cocycle superrigid for every finite index subgroup
Γ′ < Γ. Therefore, we are in a situation where we can apply [Va07, Lemma 6.5] as well as the
first part of the proof of [Va07, Theorem 6.4]. Denote by Dk(C) ⊂ Mk(C) the algebra of diagonal
matrices. As a result, we get
• a finite index subgroup Γ′ < Γ and an irreducible projective representation π : Γ′ → U (Ck)
with obstruction 2-cocycle Ω : Γ′ × Γ′ → S1 given by π(g)π(h) = Ω(g, h)π(gh),
• a projection q′ ∈ Dk(C) ⊗ B and a finite index inclusion ψ′ : A ⋊Ω Γ′ → q′N kq′ satisfying
ψ′(A) = (Dk(C) ⊗ B)q′,
N -- bimodule given by
ψ′(cid:0)A ⋊Ω Γ′(cid:1)q′(Ck ⊗ L2(N ))N .
kZ × Λ and eY = Z
such that q′ = χZ . Denote by T : X → Z the isomorphism of measure spaces determined by
ψ′(a) = a ◦ T −1 for all a ∈ A. Normalize the measure η on Y in such a way that Z has measure 1
and hence, T is measure preserving.
such that H is the composition of the(cid:0)A⋊Γ(cid:1) -- (cid:0)A⋊Ω Γ′(cid:1) -- bimodule given by (7.1) and the(cid:0)A⋊Ω Γ′(cid:1) --
kZ × Y . Consider the natural action eΛ y eY and define Z ⊂ eY
Denote eΛ = Z
By construction, T (Γ′ · x) ⊂eΛ · T (x) for almost all x ∈ X. So, the formula
defines a 1-cocycle for the action Γ′ y X with values in eΛ. By Theorem 6.12, Γ′ y X is Ufin-
cocycle superrigid. So, we find a measurable map ϕ : X → eΛ and a group morphism δ : Γ′ → eΛ
such that, writing ∆(x) := ϕ(x) · T (x), we have ∆(g · x) = δ(g) · ∆(x) almost everywhere. Since
Γ′ y X is weakly mixing, we find i ∈ Z
kZ such that ∆(x) ∈ {i} × Y for almost all x ∈ X. Hence,
δ(Γ′) ⊂ {1} × Λ. From now on, we view ∆ as a map from X to Y and δ as a group morphism from
Γ′ to Λ. Put Λ′′ := δ(Γ′).
T (g · x) = ω(g, x) · T (x)
By construction, ∆ is locally a m.p. isomorphism, meaning that we can partition X into a sequence
of measurable subsets Xn such that the restriction of ∆ to each of the Xn is a m.p. isomorphism
of Xn onto a subset of Y . Since ∆(g · x) = δ(g) · ∆(x) and since (X, µ) is a probability space, it
follows that Ker δ is finite and that ∆ is an m-to-1 quotient map onto Y ′ ⊂ Y (cf. [Fu06, Theorem
1.8]). Observe that Y ′ is globally Λ′′-invariant. Partition X into subsets X1, . . . , Xm of measure
1/m such that ∆i := ∆Xi is an isomorphism of Xi onto Y ′.
Define the projection pKer δ in A ⋊ Γ′ by the formula
Put B1 := L∞(Y ′) = L∞(Y )q1, where q1 := χY ′. The pair ∆, δ yields a natural(cid:0)A⋊Γ′(cid:1) -- (cid:0)B1 ⋊Λ′′(cid:1) --
bimodule structure on the Hilbert space K := L2(A ⋊ Γ′)pKer δ. We can also write this bimodule
as
where γ : A ⋊ Γ′ → Mm(C) ⊗ (B1 ⋊ Λ′′) is a finite index inclusion satisfying
pKer δ := Ker δ−1 Xg∈Ker δ
ug .
γ(A ⋊ Γ′)(cid:0)Cm ⊗ L2(B1 ⋊ Λ′′)(cid:1)B1 ⋊ Λ′′
mXi=1
eii ⊗ (aXi ◦ ∆−1
i )
γ(a) =
for all a ∈ A .
34
Since T (x) ∈eΛ·∆(x) for almost all x ∈ X, we can take W ∈ Mm,k(C)⊗N satisfying W W ∗ = 1⊗q1,
W ∗W = q′ and γ(a) = W ψ′(a)W ∗ for all a ∈ A. Replace ψ′ by W ψ′( · )W .
It follows that, for all g ∈ Γ′, ψ′(ug)γ(ug)∗ commutes with γ(A) = ψ′(A) = Dm(C) ⊗ B1, which is
maximal abelian in Mm(C) ⊗ q1N q1. So, ψ′(ug) = γ(ωgug) where ωg ∈ U (A) satisfies
ωgσg(ωh) = Ω(g, h)ωgh .
By Lemma 6.13 we find w ∈ U (A) and a map µ : Γ′ → S1 such that ωg = µ(g)wσg(w∗) for all
g ∈ Γ′. Replacing g 7→ π(g) by g 7→ µ(g)π(g) and replacing ψ′ by (Ad ψ′(w)∗) ◦ ψ′, we may assume
that π is an ordinary unitary representation (i.e. Ω = 1) and that ψ′(z) = γ(z) for all z ∈ A ⋊ Γ′.
Since ψ′(A ⋊ Γ′) has finite index in Mm(C) ⊗ q1N q1 and since γ(A ⋊ Γ′) is contained in (B ⋊ Λ′′)m,
it follows that Λ′′ < Λ is a finite index subgroup.
If s ∈ Λ, then ∆−1(s·Y ′ ∩Y ′) is globally invariant under the finite index subgroup δ−1(Λ′′ ∩sΛ′′s−1)
of Γ′ and hence, must have measure 0 or 1. So, either s · Y ′ ∩ Y ′ has measure zero, or s · Y ′ equals
Y ′ up to measure zero. Define the subgroup Λ′ < Λ consisting of those s ∈ Λ for which s · Y ′ equals
Y ′ up to measure zero. By construction, Λ′′ < Λ′ < Λ and Λ y Y is induced from Λ′ y Y ′.
The theorem now follows by applying Lemma 7.3 to the actions Γ′ y (X, µ) and Λ′′ y Y ′.
References
[An95]
C. Anantharaman-Delaroche, Amenable correspondences and approximation properties for
von Neumann algebras. Pacific J. Math. 171 (1995), 309 -- 341.
L. Bowen, Orbit equivalence, coinduced actions and free products. Preprint. arXiv:0906.4573
[Bo09]
[BP93] M. Bo zejko and M.A. Picardello, Weakly amenable groups and amalgamated products.
[CH08]
Proc. Amer. Math. Soc. 117 (1993), 1039 -- 1046.
I. Chifan and C. Houdayer, Bass-Serre rigidity results in von Neumann algebras. To appear
in Duke Math J. arXiv:0805.1566
A. Connes, Sur la classification des facteurs de type II. C. R. Acad. Sci. Paris 281 (1975), 13-15.
A. Connes, Classification of injective factors. Ann. of Math. (2) 104 (1976), 73-115.
[Co75]
[Co76]
[Co80a] A. Connes, A factor of type II1 with countable fundamental group. J. Operator Theory 4 (1980),
151-153.
[Co80b] A. Connes, Classification des facteurs. In Operator algebras and applications, Part 2 (Kingston,
[Co94]
[CJ82]
[CJ85]
1980). Proc. Sympos. Pure Math. 38, Amer. Math. Soc., Providence, 1982, p. 43109.
A. Connes, Noncommutative Geometry. Academic Press, 1994.
A. Connes and V.F.R. Jones, A II1 factor with two non-conjugate Cartan subalgebras, Bull.
Amer. Math. Soc. 6 (1982), 211-212.
A. Connes and V.F.R. Jones, Property (T) for von Neumann algebras, Bull. London Math.
Soc. 17 (1985), 57-62.
[CFW81] A. Connes, J. Feldman and B. Weiss, An amenable equivalence relation is generated by a
single transformation. Ergodic Theory Dynam. Systems 1 (1981), 431-450.
[CSV09] Y. Cornulier, Y. Stalder and A. Valette, Proper actions of wreath products and gener-
alizations. Preprint. arXiv:0905.3960
[CH89] M. Cowling and U. Haagerup, Completely bounded multipliers of the Fourier algebra of a
[Dy59]
[Dy63]
simple Lie group of real rank one. Invent. Math. 96 (1989), 507-549.
H.A. Dye, On groups of measure preserving transformations I. Amer. J. Math. 81 (1959), 119-
159.
H.A. Dye, On groups of measure preserving transformations. II. Amer. J. Math. 85 (1963),
551-576.
35
[FM77]
[Fu99a]
[Fu99b]
[Fu06]
[Ga00]
[Ga01]
[Ga05]
[Hj02]
[HK05]
[Ho07]
[Io08]
J. Feldman and C.C. Moore, Ergodic equivalence relations, cohomology, and von Neumann
algebras, II. Trans. Amer. Math. Soc. 234 (1977), 325 -- 359.
A. Furman, Gromov's measure equivalence and rigidity of higher rank lattices. Ann. of Math.
150 (1999), 1059-1081.
A. Furman, Orbit equivalence rigidity. Ann. of Math. 150 (1999), 1083-1108.
A. Furman, On Popa's cocycle superrigidity theorem. Int. Math. Res. Not. IMRN (2007), Art.
ID rnm073, 46 pp.
D. Gaboriau, Cout des relations d'´equivalence et des groupes. Invent. Math. 139 (2000), 41 -- 98.
D. Gaboriau, Invariants ℓ2 de relations d'´equivalence et de groupes. Publ. Math. Inst. Hautes
´Etudes Sci. 95 (2002), 93 -- 150.
D. Gaboriau, Examples of groups that are measure equivalent to the free group. Ergodic Theory
Dynam. Systems 25 (2005), 1809-1827.
G. Hjorth, A converse to Dye's Theorem. Trans Amer. Math. Soc. 357 (2004), 3083-3103.
G. Hjorth and A. Kechris, Rigidity theorems for actions of product groups and countable
Borel equivalence relations. Mem. Amer. Math. Soc. 177 (2005), no. 833.
C. Houdayer, Construction of type II1 factors with prescribed countable fundamental group.
J. Reine Angew Math. 634 (2009), 169-207.
A. Ioana, Cocycle superrigidity for profinite actions of property (T) groups. Preprint.
arXiv:0805.2998
[IPP05] A. Ioana, J. Peterson and S. Popa, Amalgamated free products of weakly rigid factors and
[Jo83]
[Jo00]
[Ki06]
[Ki09]
calculation of their symmetry groups. Acta Math. 200 (2008), 85 -- 153.
V.F.R. Jones, Index for subfactors. Invent. Math. 72 (1983), 1-25.
V.F.R. Jones, Ten problems. In Mathematics:
Providence, 2000, p. 7991.
Y. Kida, Measure equivalence rigidity of the mapping class group. To appear in Ann. of Math.
arXiv:math/0607600
frontiers and perspectives, Amer. Math. Soc.,
Y. Kida, Rigidity of amalgamated free products in measure equivalence theory. Preprint.
arXiv:0902.2888
[McDu70] D. McDuff, Central sequences and the hyperfinite factor. Proc. London Math. Soc. (3) 21
[MS02]
(1970), 443-461.
N. Monod and Y. Shalom, Orbit equivalence rigidity and bounded cohomology. Ann. Math.
164 (2006), 825-878.
[MvN36] F.J. Murray and J. von Neumann, On rings of operators. Ann. Math. 37 (1936), 116-229.
[MvN43] F.J. Murray and J. von Neumann, Rings of operators IV, Ann. Math. 44 (1943), 716-808.
[OW80] D.S. Ornstein and B. Weiss, Ergodic theory of amenable group actions. Bull. Amer. Math.
[Oz03]
[OP07]
[OP08]
[Pe06]
[Pe09]
[Po93]
[Po01]
[Po03]
[Po04]
Soc. (N.S.) 2 (1980), 161-164.
N. Ozawa, Solid von Neumann algebras. Acta Math. 192 (2004), 111-117.
N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra. To appear
in Ann. Math. arXiv:0706.3623
N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra, II. To
appear in Amer. J. Math. arXiv:0807.4270
J. Peterson, L2-rigidity in von Neumann algebras. Invent. Math. 175 (2009), 417-433.
J. Peterson, Examples of group actions which are virtually W∗-superrigid, in preparation.
S. Popa, Markov traces on universal Jones algebras and subfactors of finite index. Invent. Math.
111 (1993), 375-405.
S. Popa, On a class of type II1 factors with Betti numbers invariants. Ann. of Math. 163 (2006),
809 -- 899.
S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, Part I.
Invent. Math. 165 (2006), 369 -- 408.
S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, II. Invent.
Math. 165 (2006), 409-452.
36
[Po05]
[Po06a]
[Po06b]
[PS03]
[PV06]
[PV08a]
[PV08b]
[PV08c]
[Si55]
[St75]
[Va06]
[Va07]
S. Popa, Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups.
Invent. Math. 170 (2007), 243 -- 295.
S. Popa, On the superrigidity of malleable actions with spectral gap. J. Amer. Math. Soc. 21
(2008), 981 -- 1000.
S. Popa, Deformation and rigidity for group actions and von Neumann algebras. In Proceedings
of the International Congress of Mathematicians (Madrid, 2006), Vol. I, European Mathematical
Society Publishing House, 2007, p. 445-477.
S. Popa and R. Sasyk, On the cohomology of Bernoulli actions. Ergodic Theory Dynam. Sys-
tems 27 (2007), 241-251.
S. Popa and S. Vaes, Strong rigidity of generalized Bernoulli actions and computations of their
symmetry groups. Adv. Math. 217 (2008), 833-872.
S. Popa and S. Vaes, Actions of F∞ whose II1 factors and orbit equivalence relations have
prescribed fundamental group. J. Amer. Math. Soc. 23 (2010), 383-403.
S. Popa and S. Vaes, Cocycle and orbit superrigidity for lattices in SL(n, R) acting on homo-
geneous spaces. In Geometry, rigidity and group actions, Proceedings of the Conference in honor
of R.J.Zimmer's 60th birthday, to appear. arXiv:0810.3630
S. Popa and S. Vaes, On the fundamental group of II1 factors and equivalence relations arising
from group actions. In Noncommutative geometry, Proceedings of the Conference in honor of
A.Connes' 60th birthday, to appear. arXiv:0810.0706
I.M. Singer, Automorphisms of finite factors. Amer. J. Math. 77 (1955), 117-133.
A.M. Stepin, Bernoulli shifts on groups and decreasing sequences of partitions. In Proceedings of
the third Japan-USSR symposium on probability theory (Tashkent, 1975), Lecture Notes in Math.
550, Springer, Berlin, 1976. p. 592-603.
S. Vaes, Rigidity results for Bernoulli actions and their von Neumann algebras (after Sorin Popa).
S´eminaire Bourbaki, exp. no. 961. Ast´erisque 311 (2007), 237 -- 294.
S. Vaes, Explicit computations of all finite index bimodules for a family of II1 factors. Ann. Sci.
´Ecole Norm. Sup. 41 (2008), 743-788.
[VDN92] D.V. Voiculescu, K.J. Dykema and A. Nica, Free random variables. CRM Monograph Series
[Zi80]
1, American Mathematical Society, Providence, RI, 1992.
R.J. Zimmer, Strong rigidity for ergodic actions of semisimple Lie groups. Ann. Math. 112
(1980), 511-529.
37
|
1609.09617 | 1 | 1609 | 2016-09-30T07:33:04 | Maximal amenable MASAs of the free group factor of two generators arising from the free products of hyperfinite factors | [
"math.OA"
] | In this paper, we give examples of maximal amenable subalgebras of the free group factor of two generators. More precisely, we consider two copies of the hyperfinite factor $R_i$ of type $\mathrm{II}_1$. From each $R_i$, we take a Haar unitary $u_i$ which generates a Cartan subalgebra of it. We show that the von Neumann subalgebra generated by the self-adjoint operator $u_1+u_1^{-1}+u_2+u_2^{-1}$ is maximal amenable in the free product. This provides infinitely many non-unitary conjugate maximal amenable MASAs. | math.OA | math |
MAXIMAL AMENABLE MASAS OF THE FREE GROUP
FACTOR OF TWO GENERATORS ARISING FROM THE FREE
PRODUCTS OF HYPERFINITE FACTORS
KOICHI SHIMADA
Abstract. In this paper, we give examples of maximal amenable subalgebras
of the free group factor of two generators. More precisely, we consider two
copies of the hyperfinite factor Ri of type II1. From each Ri, we take a Haar
unitary ui which generates a Cartan subalgebra of it. We show that the von
Neumann subalgebra generated by the self-adjoint operator u1+u−1
1 +u2+u−1
is maximal amenable in the free product. This provides infinitely many non-
unitary conjugate maximal amenable MASAs.
2
1. Introduction
In operator algebra theory, maximal amenable subalgebras have fascinated many
researchers because they give a good insight into ambient (non-amenable) factors.
The history of the research dates back to the middle of 1960's. In 1967, Kadison
conjectured that any maximal amenable von Neumann subalgebra of any factor of
type II1 was isomorphic to the hyperfinite factor of type II1. In the early 1980's,
Popa [12] solved this conjecture negatively. He showed that the generator subal-
gebra of any free group factor is maximal amenable (Even such a simple example
had not been shown to be maximal amenable until his work, which tells us the dif-
ficulty of the problem). In order to show that the subalgebra is maximal amenable,
he used the ultraproduct technique and looked at a property, which is called the
asymptotic orthogonality property. Even now, his method has a strong influence
upon the research on maximal amenability.
After Popa's work, his result has been generalized by many researchers. Popa's
In this
example can be seen as an amenable free component of a free product.
direction, Boutonnet -- Houdayer [2] reached a general structural theorem of maximal
amenable subalgebras of (amalgamated) free products which may be of type III (See
also Houdayer [8], Houdayer -- Ueda [9] and Ozawa [11]). Popa's result can also be
seen as a subalgebra arising from a subgroup of a group factor. In this direction,
Boutonnet -- Carderi [1] gave a sufficient condition for a subalgebra of any group
factor coming from a subgroup to be maximal amenable. It is also remarkable that
their proof is quite concise, which relies on the study of non-normal states.
However, in order to understand the inner structure of the free group factors, it is
also important to investigate subalgebras which come neither from free components
nor subgroups. The reason is that there are some non-trivial presentations of the
free group factors (See Dykema [5], Guionnet -- Shlyakhtenko [7] for example). This
means that the free group factors have certain flexibility, which makes them inter-
esting. Hence it is important to investigate what kind of non-trivial automorphisms
the free group factors admit; whether a given MASA is conjugate to the generator
1
2
KOICHI SHIMADA
subalgebras by automorphisms or not. For this purpose, it is helpful to consider
whether a given MASA has similar properties to those of the generator subalgebas.
In that sense, the radial MASA of the free group factors is an interesting example
(For the definition, see Cameron -- Fang -- Ravichandran -- White [4] for example). It is
not known whether the radial MASA is conjugate to one of the generator MASAs
by automorphisms or not; although it is not contained in any free component or
subgroup in an obvious way, there is no known property which distinguishes the
radial MASA from the generator MASAs. In particular, it was shown to be maxi-
mal amenable by [4] (See also Wen [16] for a simplified proof). Their strategy for
proving the maximal amenability was to determine the form of a sequence which
asymptotically commutes with the radial MASA by using a basis constructed by
Radulescu [14]. Although their strategy itself may be simple, it could not have
been carried out without their tough computing power.
Motivated by this example, we present new examples of maximal amenable von
Neumann subalgebras of a free group factor. More precisely, we consider two copies
of the hyperfinite factor Ri (i = 1, 2) of type II1. By Dykema [5], the free product
R1 ∗ R2 is isomorphic to the free group factor of two generators. From each Ri, we
take a Haar unitary ui which generates a Cartan subalgebra of it. We show that the
von Neumann subalgebra generated by the self-adjoint operator u1 + u−1
1 + u2 + u−1
2
is maximal amenable in the free product.
Although our construction is similar to that of the radial MASA, it has a differ-
ent aspect. Unlike the radial MASA, our construction provides "many" examples.
Although we do not know whether they are really mutually non-conjugate by au-
tomorphisms or not, they are neither mutually unitary conjugate nor conjugate by
automorphisms arising from the free components.
In order to show that our subalgebras are maximal amenable, we would like to
develop a similar strategy to that of Cameron -- Fang -- Ravichandran -- White [4], in
which they show Popa's asymptotic orthogonality property for the radial MASA.
However, there are two problems. First, although their proof depends on combi-
natorics of the free groups, we cannot expect to find out such a good group-like
structure in our setting. Therefore, we first consider the special case, namely, the
case when the Haar unitaries come from generators of the irrational rotation C∗-
algebras. After that, we reduce the general case to the special case. The idea of
reducing the problem to that in a case when the factor comes from the irrational
rotation C∗-algebras is conceived by Ge [6]. He embedded a system of the general
case into that of the special case. However, just an imitation of Ge [6] does not
work well in our setting. Our key idea for overcoming this difficulty is to embed
"asymptotically" a subalgebra into another one.
Even after the problem is reduced to the special case, there arises another prob-
lem; combinatorics of the irrational rotation C∗-algebras is complicated. The irra-
tional rotation C∗-algebras can be seen as deformations of Z2. Hence it is possible
to use their algebraic structures. However, since Z2 ∗ Z2 is "less free" than F2 is,
its combinatorics is more complicated, which requires further computations.
This paper is organized as follows. In Section 2, we explain the main theorem of
this paper. Sections 3 to 7 are devoted to showing the main theorem. In Section
3, we investigate the combinatorics for the special case and construct a good basis
of L2-space. In Sections 4 and 5, we do some analytical computations necessary to
show the asymptotic orthogonality property. In Section 6, we reduce the problem to
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS3
the special case and show that the subalgebras have the asymptotic orthogonality
property. In order to show the maximal amenability of subalgebras, besides the
asymptotic orthogonality property, we need to show the singularity of them.
In
Section 7, we show the singularity of the subalgebras and conclude that they are
maximal amenable.
Acknowledgment. The author would like to thank Chenxu Wen for explaining
the main idea of his recent work [16] and bringing the author's attention to this
topic. He also appreciates Professor Cyril Houdayer's many pieces of useful advice,
particularly introducing him Ge [6].
The main theorem of this paper is the following.
2. Main theorem
Theorem 1. For each i = 1, 2, let Ri be the hyperfinite factor of type II1 and wi ∈
Ri be a Haar unitary of Ri which generates a Cartan subalgebra of Ri. Then the von
Neumann subalgebra B generated by the self-adjoint operator w1 + w−1
1 + w2 + w−1
2
is maximal amenable in the free product R1 ∗ R2 with respect to the traces.
The following is an example of unitaries which satisfy the assumption of Theorem
1.
Example 2. Let θ be an irrational number and Aθ be the universal C∗-algebra gen-
erated by two unitaries u, v with uv = e2πiθvu (an irrational rotation C∗-algebra).
Then the C∗-algebra Aθ has a tracial state τ defined by τ (ukvl) = δk,0δl,0 for
k, l ∈ Z. Take the GNS representation of Aθ with respect to the trace τ . Then the
weak closure R of Aθ in the GNS representation is isomorphic to the hyperfinite
factor of type II1. For any k, l ∈ Z \ {0} without any non-trivial common divisor,
the unitary ukvl is Haar and generates a Cartan subalgebra of R.
Proof. It is possible to choose k′, l′ ∈ Z \ {0} with kl′ − k′l = 1 because k and l
do not have any non-trivial common divisor. Then the map u 7→ ukvl, v 7→ uk′
vl′
extends to an automorphism of R.
(cid:3)
We will see that Theorem 1 has a possibility of producing many examples of
maximal amenable MASAs. We see that the theorem provides many subalgebras
which are mutually non-unitary conjugate. Let ui, vi be generators of Ri explained
in the above example. Let wk,l
be a Haar unitary generating the von Neumann
i}′′. Set Bk,l := {wk,l
subalgebra {uk
i vl
2 )−1}′′ ⊂ R1 ∗ R2. We
show that Bk,l 6(cid:22) Bk′,l′ for any (k, l) 6= (k′, l′). Let an be unitaries of Bk,l which
converges weakly to 0. Then for any large M > 0, any small ǫ > 0, there exists
N > 0 such that for any n ≥ N , there exists a linear combination bn of the words
w with the following conditions.
1 )−1 + wk,l
i
1 + (wk,l
2 + (wk,l
(1) We have kan − bnk2 < ǫ.
(1) For any linear component w of bn, the length of w is not smaller than M .
(2) For any linear component w of bn, the ratio of the number of ui's (i = 1, 2)
in w and the number of vi's in w is k/l.
Take two words x, y of R1 ∗ R2. Then if we take M > 0 large enough compared
to the lengths of x and y, for any linear component w′ of xbny, the ratio of the
4
KOICHI SHIMADA
number of ui's in w′ and vi's in w′ is almost k/l. Thus we have EBk′ ,l′ (xbny) = 0,
which implies that Bk,l 6(cid:22) Bk′,l′.
In this way, it is possible to construct many non-unitary conjugate subalgebras.
However, our central interest is whether they are conjugate by automorphisms or
not. Thus we close this section with the following problem.
Problem 3. Are the maximal amenable subalgebras constructed in Theorem 1 mu-
tually conjugate by automorphisms? Are they conjugate to the generator MASA?
Although any two Cartan subalgebras of the hyperfinite factor of type II1 are
mutually conjugate by an automorphism, it is impossible to control the position of
Haar unitaries by the automorphism. Hence it is not clear whether it is possible to
conjugate two of them by automorphisms arising from free components.
3. Radulescu type basis
In order to show the main theorem, we first consider the case when the Haar
unitaries come from the irrational rotation C∗-algebras and later we reduce the
general case to the special case. Hence until the end of Section 4, we always assume
the following condition.
Condition. For each i = 1, 2, there exist Haar unitaries ui, vi and a complex
number d of absolute value 1 satisfying the following conditions.
(0) For any n 6= 0, we have dn 6= 1.
(1) The factor Ri is generated by ui and vi for each i.
(2) We have uivi = dviui for each i.
Set M := R1 ∗ R2, A := u1 + u−1
The purpose of this section is to construct a good basis of L2-space under this
1 + u2 + u−1
2 .
condition (Corollary 18), which is motivated by Radulesucu [14].
2 , v±1
1 , v±1
1 , u±1
3.1. Decomposing L2M into three pieces. Let w = w1 ··· wn be a word consists
of the letters {u±1
2 }. The number n of letters contained in the word
w is said to be the word length of w and denoted by w. We say that the word
w is reduced if the word length does not decrease by finitely many times of the
transformations u±1
i (i = 1, 2, s, t =
±1). On the set of reduced words, we introduce an equivalence relation defined by
us
i vt
i ←→ vt
i . Then any non-trivial word w is equivalent to a word of the following
form.
i vt
i ←→ 1 and us
i ←→ vt
i ←→ 1, v±1
i u∓1
i v∓1
i us
i us
uk1
i1 vl1
i1 uk2
i2 vl2
i2 ··· ukn
in
vln
in
,
where n ∈ Z≥1, i1,··· , in ∈ {1, 2} with i1 6= i2 6= ··· 6= in (This notation means
that i1 6= i2, i2 6= i3,··· , in−1 6= in), and kt + lt ≥ 1 for t = 1,··· , n. A word
of this form is said to be a completely reduced word. For each l ≥ 0, let W 0
l be
the set of all completely reduced words with length l consisting only of letters u±1
(i = 1, 2). Let W 1
l be the set of all completely reduced words with length l such
that if we write them as uk1
, exactly one of lt's is non-zero. Let
W 2
l be the set of all completely reduced words with length l such that if we write
them as uk1
, at least two of lt's are non-zero. We often regard
i2 ··· ukn
i1 uk2
i1 vl1
i2 vl2
vln
in
i1 vl1
i1 uk2
i2 vl2
vln
in
in
i
i2 ··· ukn
in
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS5
a reduced word as a unitary element of M . In the following, we identify the set
l as a subset of L2M . Then it is an orthonormal basis of L2M . Set
W i
Sl≥0, i=0,1,2
for each i = 0, 1, 2 and l ≥ 0. Set
W i
l := span W i
l
Wl := W 0
l ⊕ W 1
l ⊕ W 2
l
for each l ≥ 0. Let ql be the orthogonal projection onto Wl.
For each l ≥ 0, set
χl := Xw∈W 0
l
w,
which is a self-adjoint operator of M . As mentioned in p.299 of Radulescu [14], the
self-adjoint operators {χl}l≥1 satisfy
for l ≥ 2 and
Notice that A = {χ1}′′. Set
χlχ1 = χ1χl = χl+1 + 3χl−1
χ1χ1 = χ2 + 4.
Sl := S0
for i = 0, 1, 2,
For γ ∈ W i
k, k ≤ l − 1}
l and w ∈ W i
l := {ql(χ1w), ql(wχ1) w ∈ W i
Si
l ⊕ S2
l .
k with k ≤ l − 1, we have
hγ, ql(χ1w)i = hγ, χ1wi = hχ1γ, wi.
Similarly, we have hγ, ql(wχ1)i = hγχ1, wi. Hence γ ∈ W i
l means that when
we expand χ1γ and γχ1 by the orthonormal basis consisting of completely reduced
words, no term with length less than l appears. Hence we have the following
decomposition.
l ⊕ S1
l ⊖ Si
Lemma 4. We have
l ⊖ S0
l ) ⊕ (W 1
Wl ⊖ Sl = (W 0
l ⊖ S1
l ⊖ S2
l ).
l ⊕ W 2
Proof. For ξ ∈ Wl ⊖ Sl, let ξ = ξ0 + ξ1 + ξ2 ∈ W 0
l be the orthogonal
decomposition. Since any term of χ1ξ0 is not canceled by any term of χ1(ξ1+ξ2) and
since ξ is orthogonal to Sl, χ1ξ0 has no term with length less than l. Similarly, the
vector ξ0χ1 has no term with length less than l. Hence the vector ξ0 is orthogonal
to S0
(cid:3)
l ) ⊕ (W 2
l ⊕ W 1
l )⊥ and ξ2 ∈ (S2
l )⊥.
l . Similarly, we have ξ1 ∈ (S1
For r, s ∈ Z≥0 and ξ ∈ Wl, set
ξr,s := ql+r+s(χrξχs).
When r < 0 or s < 0, we define ξr,s as 0. We would like to see what relations there
are among these vectors. The following lemma is one of them.
Lemma 5. For ξ ∈ Wl with l ≥ 1, we have
χ1ξr,s = ξr+1,s + 3ξr−1,s
for r ≥ 1, s ≥ 0,
for r ≥ 0, s ≥ 1.
ξr,sχ1 = ξr,s+1 + 3ξr,s−1
6
KOICHI SHIMADA
Proof. This is shown by the same argument as that of the proof of Lemma 1 (a) of
Radulescu [14].
(cid:3)
l (cid:19) ⊕ (cid:18)Ll W 0
3.2. Structures of the space (cid:18)Ll W 2
present some formulas on vectors of the space (cid:18)Ll W 2
Lemma 6. For ξ ∈ (W 1
i = 1, 2} ∪ {v±l
i }, we have
l (cid:19). In this subsection, we
l (cid:19).
l (cid:19) ⊕(cid:18)Ll W 0
l ) ⊖ Sl which is orthogonal to the set {u±1
i v±(l−1)
l ⊕ W 2
i
χ1ξ0,s = ξ1,s
for s ≥ 0,
for r ≥ 0.
Proof. Let
ξr,0χ1 = ξr,1
λww
ξ = Xw∈ Wl
be the decomposition along the orthonormal basis Wl. We would like to show that
hχ1ξ0,s, w′′i = 0 for any w′′ ∈ Wk′ with k′ ≤ l + s − 1. In order to achieve this, we
decompose ξ into some pieces and compute the inner product of each component.
The word w ∈ Wl may begin with u±1
. We may consider these two kinds of
terms separately. In the following, we consider the case when λw = 0 if w begins
with v±1
implies that ql(χ1w0,s) = χ1w0,s).
(the other case is obvious; w 6= v±1
or v±1
j
j
j
i
Claim 1. For any word w ∈ Wl−1 and w′′ ∈ Wk′ with k′ ≤ l − 1, we have
h Xx∈ W 0
1 with x w∈ Wl
λx wx w, ql(χ1w′′)i = 0.
Proof of Claim 1. Notice that for different w's, the vectors χ1x w's are mutually
orthogonal. Thus no cancellation occurs among terms of χ1x w coming from differ-
ent w's. On the other hand, by assumption, we have hξ, ql(χ1w′′)i = 0. Thus we
get the conclusion of Claim 1.
(cid:3)
Next, we show another necessary claim.
Claim 2. Let w ∈ Wl−1 and x ∈ W 0
1 with x w ∈ Wl. Let w′ be a word of W 0
such that x w does not cancel with w′, that is, xww′ = x + w + w′. Then for
any j = 1, 2, p = ±1, if we multiply up
j from the left, it cancel with x ww′ if and
only if it cancel with x w.
s
Proof of Claim 2. Case 1. When x w 6= us
commute with x w. Hence for the word up
cancel with x w. Conversely, if the word up
should be strictly less than 1 + l + s. Thus the word up
j cannot
j , in order to cancel with x ww′, it should
j cancel with x w, the length of up
j x ww′
j cancels with x ww′.
, then the word up
i v±(l−s)
i
Case 2. When x w = us
i v±(l−s)
i
s 6= 0.
(s 6= 0), then this is directly checked by using
(cid:3)
We also need to show the following two claims.
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS7
Claim 3. Let w ∈ Wl−1 be a word orthogonal to v±(l−1)
(j = 1, 2) and x ∈ W 0
with x w ∈ Wl. Then for any reduced word w′, we have x ww′ = x + w + w′ if
and only if ww′ = w + w′. (In this claim, the assumption that the vector ξ is
orthogonal to u±1
Proof of Claim 3. This is shown by the same way as that of the proof of Claim
(cid:3)
i v±(l−1)
is used).
1
, v±l
i
j
i
2.
Claim 3'. Let x be u±1
3−j and w be v±(l−1)
j
of Claim 3 holds.
. Then the same conclusion as that
Proof of Claim 3'. This is trivial.
By Claims 1, 2, 3 and 3', for any w′ ∈ W 0
s with ww′ = l + s − 1, we have
(cid:3)
h
x∈ W 0
X
1 with x ww′∈ Wl+s
λx wx ww′, ql+s(χ1w′′)i = 0
for any w′′ ∈ Wk′ with k′ ≤ l + s − 1. When w and w′ run over all possible values,
the sum of the above equality is
0 = h Xw∈ Wl(cid:18) Xw′∈ W 0
= hξ0,s, ql+s(χ1w′′)i
= hχ1ξ0,s, w′′i
s , ww′=l+s
λwww′(cid:19), ql+s(χ1w′′)i
for all w′′ ∈ W 2
k′ with k′ ≤ l + s − 1. Hence χ1ξ0,s has no term with length less
than l + s − 1. Thus we have ξ1,s = ql+s+1(χ1ξ0,s) = χ1ξ0,s. The other equality is
shown in the same way.
(cid:3)
Lemma 7. For ξ ∈ W 0
l ⊖ S0
(1) When l ≥ 2, we have
l , we have the following statements.
for s ≥ 0,
χ1ξ0,s = ξ1,s
ξr,0χ1 = ξr,1
for r ≥ 0.
c1, c2 ∈ C, we have
(2) When l = 1 and ξ is c1(u1 + ǫu−1
1 ) + c2(u2 + ǫu−1
2 ) for some ǫ ∈ {±1} and
for s ≥ 0,
χ1ξ0,s = ξ1,s − ǫξ0,s−1
ξr,0χ1 = ξr,1 − ǫξr−1,0
for r ≥ 0.
Proof. This is shown by the same argument as that of the proof of Lemma 1 (b)(c)
of Radulescu [14].
(cid:3)
8
KOICHI SHIMADA
3.3. Structures of the space (cid:18)Ll W 1
structure of the space Ll W 1
l (cid:19). In this subsection, we investigate the
l . Fix l ∈ Z \ {0}. When l 6= ±1, set
1,± := vl−sgn(l)
γi,l
i
u±1
i
,
i
2 := vl−sgn(l)
γi,l
γi,l
3 := (u3−i + u−1
2
(u3−i + u−1
3−i),
3−i)vl−sgn(l)
,
1,+ − γi,l
1,−) − γi,l
3
i
γi
l :=
d−(l−sgn(l)) − dl−sgn(l) (γi,l
dl−sgn(l) − d−(l−sgn(l)) (dl−sgn(l)γi,l
2
1,+ − d−(l−sgn(l))γi,l
1,−) − γi,l
2 .
and
γi
l :=
For l > 0, we also set
W 1,α
l
:= span{u±1
j
i v±(l−1)
:= W 1
, v±(l−1)
j
u±1
i
l ⊕ Cv±l
±l, γi
i = 1, 2, j = 1, 2},
1 ⊕ Cv±l
2 ),
±l i = 1, 2}
l ⊖ (W 1,α
:= span{γi
W 1,β
l
W 1,α,1
l
is generated by the completely reduced words of the form xvk′
i
, v±(l−1)
= W 1,α,2
l
(u3−i − u−1
3−i) i = 1, 2}
= 0. Obviously, the space
l with
i x′ ∈ W 1
when l 6= 1 and
l
i
W 1,α,2
:= span{(u3−i − u−1
3−i)v±(l−1)
when l 6= 1. When l = 1, we set W 1,α,1
W 1,β
x, x′ ∈Sk≥0
Lemma 8. For any l ∈ Z>0, we have
k ,x + x′ ≥ 2, k′ 6= 0, i = 1, 2.
W 0
l
l
l = (W 1,β
W 1
l ⊖ S1
l ⊖ S1
Proof. First, note that we have W 1
l ). Thus in
order to get the conclusion, it is enough to determine the structure of the space
W 1,α ⊖ S1
l .
1 ⊕ Cv±l
2 .
l ⊖ S1
⊕ Cv±l
l ) ⊕ (W 1,β
l ) ⊕ W 1,α,1
l ⊖ S1
⊕ W 1,α,2
l ⊖ S1
l = (W 1,α
l
l
1
1
Claim For l ∈ Z \ {0,±1}, let ξ ∈ W 1
be a linear combination of
i, ut′
i vl−sgn(l)
vl−sgn(l)
(i, i′ = 1, 2, t, t′ = ±1). Then ξ is a linear combination
ut
1
l , (u2 − u−1
l , γ1
of γ1
Proof of Claim. For simplicity, we assume that l > 0 (When l < 0, Claim is
2 )vl−sgn(l)
and vl−sgn(l)
l ⊖ S1
l
(u2 − u−1
2 ).
shown in the same way). It is possible to write ξ as
c1,+vl−1
Then since ξ is orthogonal to χ1w with w ≤ l − 1, we have
d−(l−1)c1,+ + dl−1c1,− + c3,+ + c3,− = 0.
1 u1 + c1,−vl−1
1 u2 + c2,−vl−1
1 + c2,+vl−1
1 u−1
1 u−1
1 + c3,−u−1
2 + c3,+u2vl−1
2 vl−1
1
1
.
Since ξ is orthogonal to wχ1 with w ≤ l − 1, we have
c1,+ + c1,− + c2,+ + c2,− = 0.
l , (u2 − u−1
and vl−1
(u2 − u−1
Obviously, the vectors γ1
2 ) satisfy the above
two equalities. On the other hand, the vector space defined by these two equalities
2 )vl−1
is at most four dimensional. Thus ξ is a linear combination of γ1
and vl−1
l , (u2−u−1
2 )vl−1
l , γ1
l , γ1
(cid:3)
1
1
1
1
(u2 − u−1
2 ).
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS9
Thus ξ ∈ W 1
l ⊖ S1
l which is a linear combination of v±l
j , v±(l−1)
j′
(j, j′, j′′, i′, i′′ = 1, 2, t′, t′′ = ±1) is in fact contained in W 1,α,1
Cv±l
l
ut′
i′ , ut′′
⊕ W 1,α,2
i′′ v±(l−1)
⊕ Cv±l
1 ⊕
j′′
l
2 . Hence we have
W 1,α
l ⊖ S1
l
which implies the conclusion of the lemma.
Lemma 9. For any n, m ≥ 0, i = 1, 2, we have
l
l = W 1,α,1
⊕ W 1,α,2
⊕ Cv±1
1 ⊕ Cv±1
2 ,
χnγi,l
j χm ∈ span{(γi,l
j )r,s j = (1,±), 2, 3, r, s ≥ 0} ∨ Cvl−1
i
(cid:3)
.
Proof. Set
Notice that we have (vl−sgn(l)
i
χ1(uivl−sgn(l)
i
)0,s = χ1uivl−sgn(l)
Vi := span{(γi,l
)0,s, (vl−sgn(l)
i
j )r,s j = (1,±), 2, 3, r, s ≥ 0}.
)r,0 ∈ Vi∨ Cvl−sgn(l)
(cid:18) Xw6=u−1
i
i
i
w(cid:19)
χs − (vl−sgn(l)
i
, w=s
)1,s + vl−sgn(l)
i
= (uivl−sgn(l)
i
u−1
i
)0,s−1,
. For s ≥ 1, we have
χ1((u3−i + u−1
3−i)vl−sgn(l)
i
)0,s = χ1(u3−i + u−1
= ((u3−i + u−1
3−i)vl−sgn(l)
3−i)vl−sgn(l)
i
i
χs
)1,s + 2vl−sgn(l)
χs
i
and
i
i
χ1(vl−sgn(l)
(u3−i + u−1
j )0,s ∈ Vi ∨ Avl
3−i))0,s = (vl−sgn(l)
(u3−i + u−1
Thus we have χ1(γi,l
iA. Similarly, we have (γi,l
Hence by using Lemma 5, it is possible to conclude that χnγi
3.4. Constructing a Riesz basis of L2M ⊖ A.
Lemma 10. For ξ ∈ (W 0
the following statements.
l ) ⊕ ((W 1,α,2
⊕ W 1,β
l ⊖ S0
) ⊖ S1
l
l
3−i))1,s.
j )r,0χ1 ∈ Vi ∨ Avl
jχm ∈ Vi ∨ Avl
iA.
iA. (cid:3)
l ) ⊕ (W 2
l ⊖ S2
l ), we have
(1) (i) For any n, m ≥ 0 and l ≥ 2, we have
χnξχm = ξn,m − (ξn,m−2 + ξn−2,m) + ξn−2,m−2.
(ii) For any n, m ≥ 0 and l ≥ 2, we have
X
ξn,m =
r≤n,s≤m, (r,s) has the same parity as that of (n,m)
1 )+c2(u2 +ǫu−1
(2) When l = 1 and ξ is of the form c1(u1 +ǫu−1
and c1, c2 ∈ C, for any n, m ≥ 0, we have the following two statements.
2 ) for some ǫ ∈ {±1}
(i) We have
χrξχs.
χnξχm =ξn,m − (ξn,m−2 + ξn−2,m) + ξn−2,m−2
(−ǫ)k(ǫξn−k−1,m−k+1 + ǫξn−k+1,m−k−1 + 2ξn−k,m−k).
+Xk≥2
(ii) We have
ξn,m =
r≤n,s≤m, r−s has the same parity as that of n−m
X
ǫn−rχrξχs.
10
KOICHI SHIMADA
l
l ⊖ S0
l ) ⊕ ((W 1,α,2
Proof. By using lemmas 5, 6 and 7 , this is shown in the same way as that of
Lemma 2 of Radulescu [14].
(cid:3)
⊕ W 1,β
Lemma 11. For ξ ∈ (W 0
l ), l ≥ 1,
the projection pξ commutes with the projection qn and the range of pξ ∧ qn is the
subspace of L2(M ) spanned by {ξr,s r + s = n − l}.
Proof. This is shown in the same way as the fact mentioned in the paragraph
preceding to Lemma 2 of Radulescu [14]. However, for readers convenience, we
present a proof. Let rn be a projection onto the subspace span{ξr,s r + s = n− l}.
By Lemma 10 (1) (ii) (2) (i), we have ξr,s ∈ AξA. Hence we have rn ≤ pξ ∧ qn.
On the other hand, by Lemma 10 (1) (i) (2) (ii), we have pξ ≤ Pn rn. Hence the
range of qnpξ is contained in that of rn. Thus we have
l ) ⊕ (W 2
l ⊖ S2
) ⊖ S1
l
pξ ∧ qn ≤ qnpξqn ≤ rn
(The first equality holds without any assumption). Hence two projections pξ and
qn commute and we have pξ ∧ qn = rn.
(cid:3)
Lemma 12. For ξ, ξ′ ∈ (W 0
the following statements.
l )⊕ ((W 1,α,2
(1) When l ≥ 2, n, m, n′, m′ ≥ 0, we have
l )⊕ (W 2
⊕ W 1,β
l ), we have
l ⊖ S0
l ⊖ S2
)⊖ S1
l
l
hξn,m, ξ′n′,m′i = δn,n′δm,m′ 3n+mhξ, ξ′i.
(2) When l = 1 and ξ is of the form c1(u1 + ǫu−1
ǫ ∈ {±1} and c1, c2 ∈ C, for any n, m, n′, m′, we have
1 ) + c2(u2 + ǫu−1
2 ) for some
hξn,m, ξ′n′,m′i = δǫ,ǫ′δn+m,n′+m′3n+m(−3)−n−n′hξ, ξ′i.
Proof. By using lemmas 5, 6 and 7 , this is shown in the same way as Lemma 3 of
Radulescu [14].
(cid:3)
For l ≥ 1, let Pl be the projection onto the subspace of L2M spanned by {AwA
w ∈ Wk, k ≤ l − 1}.
Lemma 13. (1) The projection Pl−1 commutes with ql and the range of Pl−1ql is
exactly Sl.
l ) ⊖ Sl) with hξ, ξ′i = 0, we have
(2) For ξ, ξ′ ∈Sl≥1((W 0
AξA ⊥ Aξ′A.
1 ⊕ Cv±l
(3) For ξ ∈Sl≥1((W 0
2 ,
we have AξA ⊥ Aξ′A.
Proof. Although this is shown by the same argument as that of the proof of Lemma
4 of Radulescu [14], we present a proof .
⊕ W 1,β
l ⊕ W 2
l )⊖ Sl), ξ′ ∈ W 1,α,1
l ⊕ W 1,α,2
l ⊕ W 1,α,2
l ⊕ W 2
⊕ W 1,β
⊕ Cv±l
l
l
l
(1) We show this by induction on l. When l = 0, then statement (1) is obvious
because we have Pl−1 = 0. Assume that statement (1) holds for any k = 0,··· , l.
We first show the following claim.
Claim. We have ql+1(χpwχq) ∈ Si
Proof of Claim. Let w = w1 + w2 ∈ Si
l+1 for any p, q ≥ 0, w ∈ W i
k with k ≤ l.
k) be the orthogonal decom-
position. Then by the induction hypothesis, we have w1 ∈ ranP i
k−1. Hence it is a
finite sum of elements of the form χp′ w′χq′ for some p′, q′ ≥ 0, w′ ∈ W i
k′ (k′ < k).
Hence by induction, it is possible to show that w is a sum of elements of the form
χrγχs for some γ ∈ W i
k with k ≤ l, r, s ≥ 0. Hence we may assume that
k ⊕ (W i
k ⊖ Si
k ⊖ Si
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS11
k ⊖ Si
w ∈ W i
k with k ≤ l. However, by Lemma 10, when i = 0, 2, (1, α, 2) and (1, β),
it is enough to show that wp,q ∈ Si
l+1 for any p, q ≥ 0 with p + q = l − k. This
follows from Lemmas 5, 6 and 7. When i = (1, α, 1) or w = vk
i , by Lemma 9, it
is enough to show that (γi,k
j )p,q ∈ Si
l+1 for any p, q ≥ 0 with p + q = l − k. This
follows from Lemma 5 and the equalities of the proof of Lemma 9. Thus we have
ql+1(χpγχq) ∈ Sl+1. Thus Claim holds.
(cid:3)
By Claim, we have
l ∧ ql+1.
The second inclusion of the above is obvious. Hence P i
the range of P i
l+1 ⊂ ranP i
ranql+1P i
l ⊂ Si
l ∧ ql+1 is Si
l .
(2) When l 6= l′, then by statement (1), we have AξA ⊥ Aξ′A. When l = l′,
(3) When ξ ∈ W 1,α,2
, this is shown by the direct computation. In
i, we have AξA ⊥ ξ′. (cid:3)
then by Lemma 12, we have AξA ⊥ ξ′.
, ξ′ ∈ W 1,α,1
the other cases, this is shown by counting the number of vl
l
l
l commutes with ql+1 and
For each non-zero integer l, an integer k and r, s ≥ 0, set
ξi,l,k
r,s
:=
3
1
2 −1
r+s
X
w∈W 0
r , ending with u±1
3−i
wvl
iuk
1
X
w∈W 0
s , beginning with u±1
3−i
w′.
Set χ1 := χ1/√3. Then we have the following.
Lemma 14. We have
χ1ξi,l,k
r,s =(ξi,l,k
r+1,s + ξi,l,k
r−1,s
(dlξi,l,k+1
ξi,l,k
1,s + 1√3
0,s
+ d−lξi,l,k−1
0,s
)
(r ≥ 1)
(r = 0).
Similarly, we have
ξi,l,k
r,s χ1 =(ξi,l,k
r,s+1 + ξi,l,k
r,s−1
ξi,l,k
(ξi,l,k+1
r,1 + 1√3
r,0
+ ξi,l,k−1
r,0
)
(s ≥ 1)
(s = 0).
Proof. When r ≥ 1, the first equality follows from Lemma 5. We show the first
equality when r = 0. We have
χ1ξi,l,k
0,s =
1
3s/2−1 χ1vl
= √3ξi,l,k
1,s +
w
s , w1=u±1
3−i
iuk
i Xw∈W 0
3s/2−1 (dlvl
1
iuk+1
i + d−lvl
iuk−1
i
= √3ξi,l,k
1,s + dlξi,l,k+1
0,s
+ d−lξi,l,k−1
0,s
.
) Xw∈W 0
s , w1=u±1
3−i
w
Thus we have
χ1ξi,l,k
0,s = ξi,l,k
1,s +
1
√3
(dlξi,l,k+1
0,s
+ d−lξi,l,k−1
0,s
).
The second equality is shown in the same way.
(cid:3)
We have the following.
12
KOICHI SHIMADA
Lemma 15. (1) For any (i, l, k), we have
2
√6
3
(r, s ≥ 1)
(r = 0 or s = 0, (r, s) 6= (0, 0))
(r = s = 0).
kξi,l,k
r,s k2 =
(2) When (i, l, k, r, s) 6= (i′, l′, k′, r′, s′), then the vector ξi,l,k
r,s
is orthogonal to
ξi′,l′,k′
r′,s′
.
Proof. (1) The number of the words with length r ending with either u2 or u−1
is
2
exactly 2 · 3r−1 when r 6= 0 and 1 when r = 0. Thus we have the desired equality.
(cid:3)
(2) This is obvious.
Consider a finite sum
ξ :=
Let
i=1,2, j=(1,±),2,3, r,s≥0, l∈Z
X
r,s (γi,l
ai,l,j
j )r,s.
ξ = Xr,s,k,i=1,2
βi,l,k
r,s ξi,l,k
r,s
be the expansion along the orthonormal system {ξi,l,k
Lemma 16. Let ξ and {βk
r,s} be as above. Then we have the following.
r,s } (this is always possible).
(1) We have
βi,l,k
r,s =
1
3k/2 (√3
for any i, l, k, r, s.
k−1
Xj=0
dlsgn(k)jβi,l,sgn(k)
r+j,s+k−j−1 +
k−1
Xj=1
dlsgn(k)j βi,l,0
r+j,s+k−j)
(2) There exists a constant C such that if we have kξk2 ≤ 1, for any k0 ∈ N, we
βi,l,k
r,s 2 ≤
C
3 · 2k0
.
have
1
βi,l,k
r,s k2 ≤ Xk≥k0
The constant C does not depend on ξ and k0.
j ⊥ γi′,l′
r,s by βk
3k Xk≥k0
r,s ⊥ ξi′,l′,k′
r,s by αj
r,s, βi,l,k
(1) The components contributing to ξk
, γi,l
Proof. Notice that ξi,l,k
for any (i, l) 6= (i′, l′). Thus we may
assume that only one (i, l) appears in the sum of the definition of ξ. In the lest of
the proof, we denote αi,l,j
r,s.
r′,s′
j′
1,sgn(k) )r+j,s+k−j−1 (j = 0,··· ,k − 1). The coefficient coming from
r,s are (γi,l+sgn(l)
2
)r+k,s−1,(γi,l+sgn(l)
3
)r−1,s+k
and (γi,l+sgn(l)
(γi,l+sgn(l)
2
)r+k,s−1 is
The coefficient coming from (γi,l+sgn(l)
3
r+s
2 dlka2
3
r+k,s−1.
)r−1,s+k is
.
r−1,s+k
r+s
2 a3
3
The coefficient coming from (γi,l+sgn(l)
1,sgn(k) )r+j,s+k−j−1 (j = 0,··· ,k − 1) is
2 dlsgn(k)j a1,sgn(k)
r+s
3
r+j,s+k−j−1.
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS13
Hence the coefficient of ξk
r,s is
r+s
2 dlka2
r+k,s−1 + a3
r−1,s+k +
k−1
Xj=0
dlsgn(k)ja1,sgn(k)
r+j,s+k−j−1!.
βk
r,s = 3
We also have
dlka2
r+k,s−1 + a3
r−1,s+k
dlsgn(k)j a1,sgn(k)
r+j,s+k−j−1
k−1
+
Xj=0
r+k,s−1 + a3
r+k−2,s+1 + a1,sgn(k)
r+k−1,s
r+k−1,s + a3
r+k−3,s+2 + a1,sgn(k)
r+k−2,s+1)
r+k−3,s+2 + a2
r+k−2,s+1)
r+k−2,s+1 + a2
r+k−1,s)
= dlsgn(k)(k−1)(dlsgn(k)a2
− dlsgn(k)(k−1)(a3
+ dlsgn(k)(k−2)(dlsgn(k)a2
− dlsgn(k)(k−2)(a3
+ ··· − ···
+ dlsgn(k)(dlsgn(k)a2
− dlsgn(k)(a3
+ (dlsgn(k)a2
r+2,s+k−3 + a3
r+1,s+k−2)
r−1,s+k
r,s+k−1 + a2
r+1,s+k−2 + a3
r,s+k−1 + a1,sgn(k)
r,s+k−2)
+ a1,sgn(k)
r,s+k−1).
Thus we gate the conclusion.
(2) For each k, j = 0,··· , k − 1, set
ck
j :=
√3
3k/2 βsgn(k)
r+j,s+k−j−1 = 3
r+s
2 (dlsgn(k)a2
r+j+1,s+k−j−2+a3
r+j−1,s+k−j+a1,sgn(k)
r+j,s+k−j−1),
dk
j :=
1
3k/2 β0
r+j,s+k−j = 3
r+s
2 (a3
r+j−1,s+k−j + a2
r+j,s+k−j−1).
Then we have
k−1
Xj=0
j2 +
ck
k−1
Xj=1
k−1
j2 = 31−k
dk
3
3
k−1
r+j,s+k−j−12kξ1
β1
Xj=0
r+j,s+k−j−12 + 3−k
β1
4 · 3k k−1
Xj=0
Xj=0
4 · 3kk
4 · 3kkXr,s,k
4 · 3k
β1
r+j,s+k−j−1ξ1
r,sk2
βk
r,sξk
3
3
2
.
≤
=
≤
≤
k−1
Xj=1
r+j,s+k−j2
β0
r+j,s+k−j−1k2
2 +
r+j,s+k−j2kξ0
β0
2!
r+j,s+k−jk2
k−1
Xj=1
β0
r+j,s+k−j ξ0
r+j,s+k−jk2
2
r+j,s+k−j−1 +
k−1
Xj=1
14
KOICHI SHIMADA
Hence we have
βk
r,s2
k−1 + dk
Xk≥k0
≤ Xk≥k0 (cid:12)(cid:12)ck
≤ Xk≥k0 (2k + 1)(cid:12)(cid:12)ck
≤ Xk≥k0
≤
3(2k + 1)
4 · 3k
C
2k0
.
k−1 + ··· + ck
1 + dk
1 + ck
k−12 + dk
k−12 + ··· + ck
0(cid:12)(cid:12)!2
12(cid:12)(cid:12)!
(cid:3)
Set
Summarizing the above results, we have the following.
L := span{AW 1,α,1
l
A, Av±l
i A i = 1, 2, l > 0}.
Lemma 17. (See Lemma 3.2 of Cameron -- Fang -- Ravichandran -- White [4]) There
exists a sequence of orthonormal vectors {ξn}∞n=1 satisfying the following conditions.
l(n) for some l(n) ≥ 1, i(n) = 0, 2, (1, α, 2), (1, β).
(1) Each vector ξn lies in W i(n)
(2) The subspaces spanAξnA (n ∈ N ) are pairwise orthogonal in L2M .
(3) We have
L2M ⊖ (L2A ⊕ L) =M spanAξnA.
nA to spanAξ0
mA defined by (ξ0
basis of the subspace spanAξnA.
(4) For any n with l(n) > 1, the sequence {(ξn)r,s/k(ξn)r,sk2} is an orthonormal
(5) For each n, m > 0, there exists a bounded invertible operator Tn,m from the
m)r,s. Furthermore, there
subspace spanAξ0
exists a constant C0 satisfying Tn,m,T −1
(6) The subspace L is contained in the subspace Lr,s≥0, k,l∈Z, i=1,2 Cξi,l,k
Proof. This is shown by the same argument as that of the proof of Lemma 3.2
of Cameron -- Fang -- Ravichandran -- White [4]. However, for readers' convenience, we
present a proof. For each i = 0, 2, (1, α, 2), (1, β), l ≥ 1, choose an orthonormal
basis {ηl,i
i,j ≤ C0 for any n, m.
n)r,s 7→ (ξ0
k=1 of W i
k }Kl
r,s .
l ⊖ Si
{(ηl,i
l . Let {ξn} be a rearrangement of
k )}l≥1, i=0,2,(1,α,2),(1,β) k=1,··· ,Kl.
Then by construction, the sequence {ξn} satisfies condition (1). By Lemma 13 (2),
the sequence {ξn} satisfies condition (2).
We show that {ξn} satisfies condition (4). By Lemma 11, the set {(ξn)r,s} spans
AξnA. By condition (2) and Lemma 12, the vectors {(ξn)r,s}n,r,s are mutually
orthogonal if l(n) > 1. Thus we have condition (4).
Next, we show that the sequence {ξn} satisfies condition (3). By Lemma 13
(3), L is orthogonal to W spanAξnA. By Lemma 13 (2), the subspaces AξnA's are
mutually orthogonal. Thus it is enough to show that the set {AξnA}n really spans
L2M ⊖ (L2A⊕ L). Take an element ξ ∈ Wl ⊖ (L2A⊕ L) which is orthogonal to any
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS15
AξnA. Then ξ is orthogonal to the space Wl ⊖ Sl, which means that ξ ∈ Sl. On
the other hand, by the same argument as in the proof of Lemma 13 (1), any vector
w ∈ Wk is written as a linear combination of the form χnγχm for some n, m ≥ 0,
γ ∈ Wk′ ⊖ Sk′ (k′ ≤ k). Thus the vector ξ is orthogonal to AwA for any w ∈ Wk,
k ≤ l − 1. Hence by Lemma 13 (1), the vector ξ is orthogonal to Sl. Thus the
vector ξ is zero.
Condition (5) follows in the same way as the proof of Lemma 3.2 of Cameron --
(cid:3)
Fang -- Ravichandran -- White [4]. Condition (6) is trivial.
By an immediate consequence of Lemma 17, we have the following.
Corollary 18. We have the following.
(1) The family {(ξm)r,s m ∈ N, r, s ≥ 0} is a Riesz basis of L2M ⊖ (A ⊕ L).
(2) We have
L ⊂ Mi,l,k,r,s
ξi,l,k
r,s .
3) We have L ⊕ (L2M ⊖ (A ⊕ L)) = L2M ⊖ A.
4. Locating the support of the sequences of (M ω ⊖ Aω) ∩ A′
Now, we would like to explain how to show the maximal amenability of the
subalgebra. In order to show the maximal amenability, we look at the following
notion.
Definition 19. (Lemma 2.1 of Popa [12]) Let M be a factor of type II1 and A be
a von Neumann subalgebra of M . We say that the subalgebra A has the asymptotic
orthogonality property if for any x1 = (x1
n) ∈ (M ω ⊖ Aω) ∩ A′, any
y1, y2 ∈ M ⊖ A, we have τ ω(y∗1x1∗y2x2) = 0.
n), x2 = (x2
Although the definition of the asymptotic orthogonality property is rather tech-
nical, this property is crucial because of the following proposition.
Proposition 20. (Corollary 2.3 of Cameron -- Fang -- Ravichandran -- White [4], See
also Popa [12]) Let A be a singular maximal abelian subalgebra of a factor M of
type II1 with the asymptotic orthogonality property. Then it is maximal amenable.
This is why we would like to show that the subalgebra has the asymptotic or-
thogonality property. In order to achieve this, we first show that any sequence of
(M ω⊖Aω)∩A′ eventually get out of the space span{ξi,l,k
r,s , (ξn)r,s r ≤ M or s ≤ M}
for any M > 0. Then we show that any vectors η1, η2 orthogonal to the space and
any vector a, b,∈ M ⊖ A, the value τ (a ∗ η∗1bη2) is small if M is large enough. In
this section, we show the first part.
As in Section 3, set
1
√3
Then we have the following lemma.
χ1 :=
(u1 + u−1
1 + u2 + u−1
2 ).
Lemma 21. (See Lemma 4.3 of Cameron -- Fang -- Ravichandran -- White [4], See also
Lemma 11 of Wen [16]) Let D ⊂ A be a diffuse von Neumann subalgebra. Let
x = (xn) be an ω-centralizing sequence of M commuting with D, kxnk = 1 and
16
KOICHI SHIMADA
EA(xn) = 0 for all n. Assume that each xn is written as xn =Pm,r,s αn,m
for some αn,m
r,s (ξm)r,s
r,s ∈ C. Then for each M ∈ N, we have
kαn,m
r,s k2
Xm≥1,r≤M or s≤M
lim
n→ω
2 = 0.
Proof. This is shown by the same way as that of Lemma 4.3 of Cameron -- Fang --
Ravichandran -- White [4].
(cid:3)
By Lemma 14, we have
χ1ξk
=
r,s − ξk
r,s χ1
ξk
r+1,s + ξk
1,s + 1√3
ξk
ξk
r+1,0 + ξk
1,s + 1√3
ξk
Hence for x =Pi,l,r,s,k αi,l,k
r−1,s − ξk
(dlξk+1
r−1,0 − ξk
(dlξk+1
r,s+1 − ξk
0,s + d−lξk−1
r,1 − 1√3
0,s + d−lξk−1
r,s−1
0,s ) − ξk
(ξk+1
0,s ) − ξk
(r, s ≥ 1)
(r = 0, s ≥ 1)
(r ≥ 1, s = 0)
(r = s = 0).
0,s−1
0,s+1 − ξk
r,0 + ξk−1
r,0 )
r,1 − 1√3
r,s ∈ C for each r, s, i, l, k, write
r,0 + ξk−1
r,0 )
(ξk+1
r,s ξi,l,k
r,s , where αi,l,k
χ1x − x χ1 = Xr,s,i,l,k
r,s ξi,l,k
βi,l,k
r,s .
Then the complex number βk
r,s is the following.
0,s
+ d−lαi,l,k+1
r,s+1 − αi,l,k
r,s−1
(dlαi,l,k−1
0,s+1 + 1√3
(αi,l,k−1
r,1 − 1√3
(dl − 1)(αi,l,k−1
r,s with kxk2 = 1, s′ ≥ s ≥ 1, we have the following
(r, s ≥ 1)
(r = 0, s ≥ 1)
(r ≥ 1, s = 0)
(r = s = 0).
0,s
+ αi,l,k+1
)
− d−lαi,l,k+1
0,0
0,0
r,0
r,0
)
)
βi,l,k
r,s =
αi,l,k
r+1,s + αi,l,k
αi,l,k
1,s − αi,l,k
αi,l,k
r−1,0 + αi,l,k
αi,l,k
1,0 − αi,l,k
r−1,s − αi,l,k
0,s−1 − αi,l,k
r+1,0 − αi,l,k
0,1 + 1√3
Lemma 22. For x =P αi,l,k
Xr≥s′,i,l,k
r−s,0 + αi,l,k
αi,l,k
inequalities.
r,s ξi,l,k
r−s+2,0 + ··· + αi,l,k
1
r+s,0 −
√3(cid:16)αi,l,k−1
r+s−1,0(cid:17)
r−s+1,0 + ··· + αi,l,k−1
r,s 2!1/2
− Xr≥s′,i,l,k
r+s−1,0(cid:17)2(cid:19)1/2
αi,l,k
r+s,0(cid:17)
r−s,0 + ··· + αi,l,k
√3(cid:0)αi,l,k+1
r−s+1,0 + ··· + αi,l,k+1
r+s−1,0(cid:17) +
1
r+s−1,0(cid:1)2!1/2
r−s+1,0 + ··· + αi,l,k+1
1
−
√3(cid:16)αi,l,k+1
≤ Xr≥s,i,l,k
r,s −(cid:16)αi,l,k
αi,l,k
√3(cid:16)αi,l,k−1
≤ 3s−1C0k[x, χ1]k2.
+
1
r−s+1,0 + ··· + αi,l,k−1
Proof. This is shown in a similar way to Lemma 4.1 of Cameron -- Fang -- Ravichandran --
White [4].
(cid:3)
Similarly, we have the following.
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS17
ξi,l,k
r,s with kxk2 = 1, r′ ≥ r ≥ 1, we have the following
,s
inequalities.
Lemma 23. For x =P αi,l,k
Xs≥r′,i,l,k
αi,l,k
0,s−r + αi,l,k
0,s−r+2 + ··· + αi,l,k
0,s+r −
dl
√3(cid:16)αi,l,k−1
0,s+r−1(cid:17)
0,s−r+1 + ··· + αi,l,k−1
r,s 2!1/2
− Xs≥r′,i,l,k
αi,l,k
0,s−r+1 + ··· + αi,l,k+1
d−l
−
√3 (cid:16)αi,l,k+1
≤ Xs≥r,i,l,k
r,s −(cid:16)αi,l,k
αi,l,k
√3(cid:16)αi,l,k−1
≤ 3r−1C0k[x, χ1]k2.
dl
+
0,s+r−1(cid:17) 2(cid:19)1/2
0,s+r(cid:17)
0,s−r + ··· + αi,l,k
√3(cid:0)αi,l,k+1
0,s+r−1(cid:17) +
d−l
0,s−r+1 + ··· + αi,l,k−1
0,s−r+1 + ··· + αi,l,k+1
0,s+r−1(cid:1)2!1/2
Lemma 24. For xn = Pr,s,i,l,k αn,i,l,k
as n → ω, we have
1
√3
lim
αn,i,l,k
r,0 −
r,s
ξi,l,k
r,s with kxnk2 = 1 and k[xn, χ1]k2 → 0
(αn,i,l,k−1
r+1,0 + αn,i,l,k+1
r+1,0
)2 = 0,
n→ω Xr≥0, i,l,k
n→ω Xs≥0, i,l,k
lim
αn,i,l,k
0,s −
1
√3
(dlαn,i,l,k−1
0,s+1 + d−lαn,i,l,k+1
0,s+1
)2 = 0.
Proof. This is shown in a similar way to that in Lemma 4.2 of Cameron -- Fang --
Ravichandran -- White [4].
(cid:3)
r,s
Lemma 25. For xn =Pr,s,i,l,k αn,i,l,k
0 as n → ω, we have
n→ω Xi,l,k,s
n→ω Xi,l,k,r
lim
lim
r,s ∈ L with kxnk2 = 1 and k[xn, χ1]k2 →
ξi,l,k
αn,i,l,k
0,s
2 = 0,
αn,i,l,k
r,0
2 = 0.
Proof. By the middle ≤ the right of Lemma 22 and Lemma 24, for any ǫ > 0 and
s ∈ N, there exists a natural number Ns such that
(*)
Xr≥s,i,l,k
r,s − αn,i,l,k
αn,i,l,k
r+s,0 2!1/2
< ǫ
for any n ≥ Ns. Similarly, by Lemmas 23 and 24, for any ǫ > 0 and r ∈ N, there
exists a natural number Nr such that
(**)
Xs≥r,i,l,k
r,s − αn,i,l,k
αn,i,l,k
0,s+r 2!1/2
< ǫ
18
KOICHI SHIMADA
for any n ≥ Nr. Fix a natural number s0. By picking up terms over r ≥ 2s0 − s of
the first inequality for s = 1,··· s0, we have
r+s,0 2!1/2
for n ≥ N1,··· , Ns0. By the triangle inequality, we have
Xr≥2s0−s, i,l,k
r,s − αn,i,l,k
αn,i,l,k
< ǫ
αn,i,l,k
r,s
2!1/2
− Xr≥2s0,i,l,k
αn,i,l,k
r,0
< ǫ.
2!1/2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2!1/2
Xr≥2s0−s, i,l,k
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xr≥2s0, i,l,k
Here, we re-enumerated the index of the second term of the left hand side of the
above inequality. Hence we have
αn,i,l,k
r,0
2!1/2
< Xr≥2s0−s, i,l,k
αn,i,l,k
r,s
+ ǫ.
Taking the square of the above inequality, we have
Xr≥2s0, i,l,k
αn,i,l,k
r,0
2 < Xr≥2s0−s, i,l,k
≤ Xr≥2s0−s, i,l,k
αn,i,l,k
r,s
αn,i,l,k
r,s
2 + 2 Xr≥2s0−s, i,l,k
2 + 2C0ǫ + ǫ2.
αn,i,l,k
r,s
2!1/2
ǫ + ǫ2
Taking the average of the above inequality over s = 1,··· , s0, we have
Xr≥2s0, i,l,k
Hence we have
αn,i,l,k
r,0
2 <
1
s0
X
r≥2s0−s, s=1,··· ,s0, i,l,k
0 + 2C0ǫ + ǫ2.
C2
<
1
s0
αn,i,l,k
r,s
2! + 2C0ǫ + ǫ2
Similarly, we have
lim
n→ω Xr≥2s0, i,l,k
n→ω Xs≥2r0, i,l,k
lim
αn,i,l,k
r,0
2!1/2
C0√s0
.
≤
αn,i,l,k
0,s
2!1/2
C0√s0
.
≤
On the other hand, by Lemma 16 (2), for any k0, we have
lim
n→ω Xr<2s0, k≥k0,i,l
αn,i,l,k
r,0
2!1/2
<
C
2k0
.
Next, by looking at the partial sum over r = s(≥ 1) of inequality (*), for n ≥ Nr,
we have
r,r − αn,i,l,k
αn,i,l,k
2r,0
2!1/2
< ǫ.
Xi,l,k
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS19
By looking at the partial sum over r = s of inequality (**), there exists a natural
number Mr such that for any n ≥ Mr, we have
r,r − αn,i,l,k
αn,i,l,k
0,2r
2!1/2
< ǫ.
Xi,l,k
Hence for n ≥ Nr, Mr, we have
(***)
αn,i,l,k
0,2r − αn,i,l,k
2r,0
2!1/2
< 2ǫ.
Xi,l,k
By using this and Lemma 24, and by taking ǫ smaller, we have
(****)
αn,i,l,k
0,2r−1 − αn,i,l,k
2r−1,02!1/2
< 2ǫ.
Xi,l,k
Suppose that there exited r ≥ 2 with limn→ωPi,l αn,i,l,0
2 = c > 0. Take a small
positive number δ > 0, which depends on c and is determined later. Since xn ∈ L,
by Lemma 16 (1), we have
r,0
αn,i,l,k
r,s =
1
3k/2 (√3
k−1
Xj=0
dlsgn(k)jαn,i,l,sgn(k)
r+j,s+k−j−1 −
k−1
Xj=1
dlsgn(k)j αn,i,l,0
r+j,s+k−j)
for any i, l, k, r, s. On the other hand, by Lemma 24, we have
αn,i,l,k
r,0 −
1
√3
Xr≥0, i,l,k
(αn,i,l,k−1
r+1,0 + αn,i,l,k+1
r+1,0
)2 < δ
for any sufficiently large n. By inequalities (*), (**), (***) and (****), for any fixed
(r, s) ∈ N2, we have
r,s − αn,i,l,k
αn,i,l,k
r+s,0 2)1/2 < δ
(Xi,l,k
for any sufficiently large n. From now, for any fixed (r, k), we regard {αn,i,l,k
a vector of ℓ2({1, 2} × Z). Then we have
r,0
}i,l as
kαn,i,l,k
r,0 −
1
3k/2 (√3
dlsgn(k)jαn,i,l,sgn(k)
r+k−1,0 −
k−1
Xj=1
dlsgn(k)jαn,i,l,0
r+k,0)k2 < k
2
k−1
3
· 2δ,
k−1
Xj=0
(√3
kαn,i,l,k+sgn(k)
r+1,0
−
3
1
k+1
2
k
Xj=0
dlsgn(k)j αn,i,l,sgn(k)
r+k+1,0 −
k
Xj=1
dlsgn(k)j αn,i,l,0
r+k+2,0)k2 < k + 1
k
2
3
kαn,i,l,k−sgn(k)
r+1,0
−
(√3
1
k−1
2
3
k−2
Xj=0
dlsgn(k)jαn,i,l,sgn(k)
r+k−1,0 −
k−2
Xj=1
dlsgn(k)jαn,i,l,0
r+k,0)k2 < k − 1
k+1
2
3
·2δ,
·2δ.
20
KOICHI SHIMADA
Thus we have
δ > kαn,i,l,k
r,0 −
> −k + 1
k−1
2
3
−
−
>
k
2 (√3
k−2
r+k+1,0 −
dlsgn(k)j αn,i,l,sgn(k)
dlsgn(k)j αn,i,l,sgn(k)
Xj=0
Xj=0
r+k−1,0 − αn,i,l,0
r+k,0)
1
√3
√3
(√3
√3
2 k(√3αn,i,l,sgn(k)
1
k
(√3
dlsgn(k)jαn,i,l,sgn(k)
r+k−1,0 −
3
1
1
√3
(αn,i,l,k+sgn(k)
r+1,0
+ αn,i,l,k−k
r+1,0
)k2
(1 +
2
√3
) · 2δ +
3
1
k
2 k(√3
dlsgn(k)jαn,i,l,sgn(k)
r+k−1,0 −
k−1
Xj=1
dlsgn(k)j αn,i,l,0
r+k,0)
k−1
k
Xj=0
Xj=1
Xj=1
k−2
dlsgn(k)jαn,i,l,0
r+k+2,0)
dlsgn(k)j αn,i,l,0
r+k,0)k2
1
1
3
−
dlsgn(k)j αn,i,l,0
r+k+2,0)k2 − 6k + 1
r+k+1,0 −
Xj=−k+2
Xj=−k+1
> (A)r,k − 999δ
for any sufficiently large n (How large we should take n depends on r, k and δ).
Thus we have (A)r,k < 1000δ. Now, we have r +k = (r + 1)− (k− 1). Hence for
a fixed r ≥ 3, we have
(√3αn,i,l,sgn(k)
r+1,0 )k2 ≤ k(A)r−k−1,k − (A)r−k,k−1k2
− dlsgn(k)αn,i,l,0
1
3
k−1
k
r,0
3
δ
2
Thus we have
Hence we have
Thus we have
< 2000δ.
√3αn,i,l,sgn(k)
r,0
k
− dlsgn(k)αn,i,l,0
r+1,0k2 ≤ 10000δ.
r,0 + αn,i,l,−1
r,0
kαn,i,l,1
1
√3
−
(dl + d−l)αn,i,l,0
r+1,0k < 19990δ.
√3αn,i,l,0
r−1,0 −
k
1
√3
(dl + d−l)αn,i,l,0
r+1,0k < 20000δ
for any sufficiently large n (depending on r and δ). Hence if we take δ so large that
it satisfies δ < c/100000, we have
kαn,i,l,0
r+1 k2 ≥
4
3
c
if we take a large n. Take a large T ∈ N so large that it satisfies T > 1/c. Then,
by induction, for any T ≥ t > 0, we have
kαn,i,l,0
r+2t−1,0k2 ≥
4t
3t c
for any large n (depending on r, T and δ), which would contradict the fact that
k2 → 0 for any r ≥ 2. By using
2 ≤ 1. Hence we have kαn,i,l,0
Pi,l,k,r,s αn,i,l,k
other inequalities, it is possible to show that kαn,i,l,k
k2 → 0 for any r, k.
r,0
r,0
r,s
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS21
Hence we have
r,0
lim
αn,i,l,k
n→ωXi,l,k
2
n→ω Xr≥2s0, i,l,k
= lim
C2
0
s0
≤
+
C
2k0
+ 0.
+ Xr<2s0, k≥k0
+ Xr<2s0, k<k0!αn,i,l,k
r,0
2
0 /s0 < ǫ/2 and then we choose
For any ǫ > 0, we choose s0 so large that we have C2
k0 so huge that we have C/2k0 < ǫ/2. Then we have
2 < ǫ.
αn,i,l,k
0,r
Xr≥0, i,l,k
Thus we have limn→ωPr≥0, i,l,k αn,i,l,k
n→ωXi,l,k
we have
lim
0,r
Thus we are done.
2 = 0. By the same argument as above,
αn,i,l,k
2 = 0.
0,s
(cid:3)
5. Counting the number of words which contribute to the inner
product
In this section, we show that for any vectors η1, η2 ∈ span{ξi,l,k
r,s , (ξm)r,s r ≥
M or s ≥ M}, any vectors a, b ∈ M ⊖ A, the inner product τ (a∗η∗1bη2) is small if
M is large enough (Lemmas 27 and 26). This section corresponds to Section 5 and
Lemma 6.1 of Cameron -- Fang -- Ravichandran -- White [4].
Lemma 26. Let g, h be elements of M ⊖ A satisfying the following conditions.
in
i1 wi1 ··· ukn
(1) The vector g is of the form uk1
win for some n ≥ 1, i1 6= ··· 6= in,
k1,··· , kn ∈ Z, where for any s = 1,··· , n, the operator uks
wis satisfies either (a)
ks 6= 0, wis = 1 or (b) wis is a normalizing unitary of {uis}′′ which is orthogonal
to {uis}′′.
for some n′ ≥ 1, i′1 6= ··· 6= i′n′,
satisfies either (c)
t}′′ which is orthogonal to
k′1,··· , k′n ∈ Z, where for any t = 1,··· , n′, the operator uk′
k′t 6= 0, xi′
{ui′
(2) The vector h is of the form uk′
1
i′
1
is a normalizing unitary of {ui′
1 ··· uk′
= 1 or (d) xi′
xi′
n′
n
i′
n′
xi′
xi′
t
i′
t
is
t
t
t
t}′′.
(3) At least one of the operators uks
is
satisfies the above condition (b) or (d).
wis
(s = 1,··· , n), uk′
t
i′
t
xi′
t
(t = 1,··· , n′)
Then there exists a positive constant C > 0, which depends neither on wis 's nor
's, such that for any M > 4k := 4(k1 + k2 + ··· +kn + k′1 + ··· +k′n′), any
xi′
vectors η1, η2 of the space
t
span{(ξm)2M+r,2M+s m ≥ 0, r, s ≥ 0}
∨ span{ξi,l,k
2M+r,2M+s i = 1, 2, l ∈ Z, k ∈ Z, r, s ≥ 0} ∩ L!,
22
we have
KOICHI SHIMADA
hη1g, hη2i ≤ C
M 4
3M/2kη1k2kη2k2.
Before proving the lemma, we have to notice that the above ks's, k′t's, wis 's and
's are not completely determined by h and g. This lemma means that the above
xi′
equation holds for any vectors h and g which admit the above presentation.
t
Proof. Roughly speaking, the strategy is the following. Decompose the vectors
η∗1 and η2 into linear combinations of wyw′, w′′zw′′′, where w, w′, w′′, w′′′, y, z are
words with w,w′,w′′,w′′′ ≫ g,h. Then we cancel each two neighboring
words as possible. The vital point is that at least one of h∗ and g is not completely
canceled by condition (3). By using this fact, we show that for most (w, w′, w′′, w′′′),
we have τ (h∗wyw′gw′′zw′′′) = 0 (Claim 3).
Since the two spaces
and
span{(ξi
m)2M+r,2M+s m ≥ 0, r, s ≥ 0}
span{ξi,l,k
2M+r,2M+s i = 1, 2, l ∈ Z, k ∈ Z, r, s ≥ 0} ∩ L
are mutually orthogonal, it is enough to show the following claim.
Claim 1. Assume that vectors η1 and η2 belong to one of the above two spaces.
Then we have
hη1g, hη2i ≤ C
M 4
3M/2kη1k2kη2k2.
In the rest of this proof, we devote our attention to proving this claim. For
simplicity, we assume that both η1 and η2 belong to the former subspace (We can
handle the other three cases in the same way). Write η1 and η∗2 in the following
way.
3− 4M +r+s
2
λm,2M+r,2M+s(ξm)2M+r,2M+s,
3− 4M +r′+s′
2
µm′,2M+r′,2M+s′ (ξm′ )2M+r′,2M+s′ .
η1 = Xr,s≥0, m∈N
η∗2 = Xr′,s′≥0, m′∈N
Then by Lemma 17 (5), we have
kη1k2 ≥ C−1
kη2k2 ≥ C−1
λm,2M+r,2M+s2!1/2
0 Xr,s,m
0 Xr′,s′,m′ µm′,2M+r′,2M+s′2!1/2
,
.
Set two vectors hj′′′,j and gj′,j′′ in the following way. When win = 1, set
h∗j′′′,j := uk1−j′′′
i1
wi1 uk2
i2 ··· win−1 ukn−j
in
.
Here, j′′′ and j run over the following ranges. When n ≥ 2, (j′′′, j) run over
all j′′′ = 0,··· , k1, j = 0,··· , kn. When n = 1, k1 ≥ 0, (j′′′, j) run over all
j′′′ = 0,··· , k1, j = 0,··· , k1, 0 ≤ j + j′′′ ≤ k1. When n = 1, k1 < 0, (j′′′, j) run
over all j′′′ = 0,··· , k1, j = 0,··· , k1, 0 ≥ j + j′′′ ≥ k1.
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS23
When win 6= 1, set
for j′′′ = 0,··· , k1, j = 0.
When xi′
n′ = 1, set
h∗j′′′,j := uk1−j′′′
i1
wi1 ··· ukn
in
win
1−j′
gj′,j′′ := uk′
i′
1
xi′
1
uk′
2
i′
2 ··· u
k′
n′−j′′
i′
n′
.
Here, j′ and j′′ run over the following ranges. When n′ ≥ 2, (j′, j′′) run over
all j′ = 0,··· , k′1, j′′ = 0,··· , k′n. When n′ = 1, k′1 ≥ 0, (j′, j′′) run over all
j′ = 0,··· , k′1, j′′ = 0,··· , k′1, 0 ≤ j′ + j′′ ≤ k′1. When n′ = 1, k′1 < 0, (j′, j′′) run
over all j′ = 0,··· , k′1, j′′ = 0,··· , k′1, 0 ≥ j′ + j′′ ≥ k′1.
When xi′
n′
6= 1, set
1−j′
gj′,j′′ := uk′
i′
1
xi′
1
uk′
2
i′
2 ··· u
k′
n′
i′
n′
xi′
n′
for j′ = 0,··· , k′1, j′′ = 0. Here, there is an important notice:
by condition (3), either h∗j′′′,j or gj′,j′′ is not 1.
Let
3− 2M +r+s−j−j′
2
3− 2M +r+s−j′′ −j′′′
2
y,
(ξm)M+r−j,M+s−j′ =Xy
(ξm′ )M+r′−j′′,M+s−j′′′ =Xz
z
be decompositions, where {y}, {z} are sets of mutually orthogonal non-zero scalar
multiples of complete reduced words, respectively.
We also define a subset V1(m, r, s, j, y) of W 0
Case 1. When win = 1, V1(m, r, s, j, y) is the set of all words w satisfying the
M in the following way.
following conditions.
the inverse of the j-th letter of h from the right.
(1) The first letter of w is neither the (j − 1)-st letter of h∗ from the right nor
(2) The last letter of w does not cancel with the first letter of y.
Case 2. When win 6= 1, V1(m, r, s, j, y) is the set of all words whose last letters
do not cancel with the first letter of y.
We define a subset V2(m, r, s, j′, y) of W 0
The set V2(m, r, s, j′, y) consists of all words w′ satisfying the following condi-
M in the following way.
tions.
(1) The first letter of w′ does not cancel with the last letter of y.
(2) The last letter of w′ is neither the (j′ − 1)-st letter of g from the left nor the
inverse of the j′-th letter of g from the left.
We define a subset V3(m′, r′, s′, j′′, z) of W 0
Case 1. When xi′
n′ = 1, V3(m′, r′, s′, j′′, z) is the set of all words w′′ satisfying
M in the following way.
the following conditions.
the inverse of j′′-th letter of g from the right.
(1) The first letter of w′′ is neither the (j′′ − 1)-st letter of g from the right nor
(2) The last letter of w′′ does not cancel with the first letter of z.
24
KOICHI SHIMADA
Case 2. When xi′
n′
6= 1, the set V3(m′, r′, s′, j′′, z) consists of all words whose
last letters do not cancel with the first letter of z.
We define a subset V4(m′, r′, s′, j′′′, z) of W 0
The set V4(m′, r′, s′, j′′′, z) consists of all words w′′′ satisfying the following con-
M in the following way.
ditions.
(1) The first letter of w′′′ does not cancel with z.
(2) The last letter of w′′′ is neither the (j′′′ − 1)-st letter of h∗ from the left nor
the inverse of the j′′′-th letter of h from the left.
Hereafter, if there is no danger of confusion, we sometimes abbreviate V1(m, r, s, j, y),
V2(m, r, s, j′, y), V3(m′, r′, s′, j′′, z) and V4(m′, r′, s′, j′′′, z) to V1, V2, V3 and V4, re-
spectively. In this setting, we have the following claim.
Claim 2. We have
3
h∗η1gη∗2 = 3−2M Xj,j′,j′′,j′′′ Xr,s,m,r′,s′,m′
w′(cid:19)gj′,j′′(cid:18)
w′′′(cid:19)u−j′′′
(cid:18) Xw′∈V2(m,r,s,j′,y)
(cid:18)
Xw′′′∈V4(m′,r′,s′,j′′′,z)
i1
,
j+j′ +j′′ +j′′′
2
uj′′′
i1 hj′′′,jXy,z(cid:18) Xw∈V1(m,r,s,j,y)
w′′(cid:19)µm′,2M+r′,2M+s′ z
Xw′′∈V3(m′,r′,s′,j′′,z)
w(cid:19)λm,2M+r,2M+sy
where j, j′, j′′, j′′′ run over the following ranges. When win = 1 and n ≥ 2, j, j′′′
run over all j = 0,··· , kn, j′′′ = 0,··· , k1. When win = 1, n = 1 and k1 ≥ 0, j, j′′′
run over all j = 0,··· , k1, j′′′ = 0,··· , k1, 0 ≤ j + j′′′ ≤ k1. When win = 1, n = 1
and k1 < 0, j, j′′′ run over all j = 0,··· , k1, j′′′ = 0,··· , k1, 0 ≥ j + j′′′ ≥ k1.
When win 6= 1, j, j′′′ run over j = 0, j′′′ = 0,··· , k1.
n′ = 1 and n′ ≥ 2, j′, j′′ run over all j′ = 0,··· , k1, j′′ = 0,··· , k′n.
n′ = 1, n′ = 1 and k1 ≥ 0, j′, j′′ run over all j′ = 0,··· , k′1, j′′0,··· , k′1,
When wi′
n′ = 1, n′ = 1 and k′1 < 0, j′, j′′ run over all j′ = 0,··· , k′1,
j′ + j′′ ≤ k1. When wi′
6= 1, j′, j′′ run over all j′′ = 0, j′ =
j′′ = 0,··· , k′1, j′ + j′′ ≥ k′1. When wi′
0,··· , k′1.
When wi′
n′
Proof of Claim 2. Obviously, we have
h∗η1gη∗2 = 3− 8M +r+s+r′+s′
2
λm,2M+r,2M+s(ξm)2M+r,2M+s
h∗ Xm,r,s
g Xm′,r′,s′
µm′,2M+r′,2M+s′ (ξm′ )2M+r′,2M+s′ .
We would like to look at words appearing in h∗(ξm)2M+r,2M+sg(ξm′ )2M+r′,2M+s′
and to reduce each neighboring two words as possible. Note that in this reduction,
we do temporary think that the letters wi and xi are free from ui. Consider the
component h∗y′gz′, where y′ is a linear component of (ξm)2M+r,2M+s, z′ is a linear
component of (ξm′ )2M+r′,2M+s′ . When we reduce each two neighboring two blocks
in this word as possible, then the word becomes a linear sum of words of the form
uj′′′
i1 hj′′′,jwyw′gj′,j′′ w′′zw′′′,
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS25
where y is a component of (ξm)M+r−j,M+s−j′ , z is a component of (ξm′ )M+r′−j′′,M+s′−j′′′
and w, w′, w′′, w′′′ are words of W 0
M satisfying the following conditions.
(1) The word w does not cancel with h∗j′′′,j or y.
(2) The word w′ does not cancel with y or gj′,j′′ .
(3) The word w′′ does not cancel with gj′,j′′ or z.
(4) The word w′′′ does not cancel with z or hj′′′,j.
We have to show that the above conditions are satisfied if and only if (w, w′, w′′, w′′′) ∈
n′ = 1. Other cases
V1 × V2 × V3 × V4. For simplicity, we consider when win = 1, xi′
are shown in the same way (and the argument is much easier). We show that if
w 6∈ V1, then the word w does not satisfy condition (1). Since a word w does not
cancel with hj′′′,j, the first letter of w cannot be the inverse of the last letter of
hj′′′,j. In order to cancel the (j − 1)-st letter of h from the right, there should be
the inverse of that letter ahead of w. Hence the first letter of w cannot be the
(j − 1)-st letter of h. Of course, if the last letter of w cancel with y , then the
word w cannot satisfy condition (1). Hence any word w 6∈ V1 cannot contribute to
the second summation. On the other hand, if w ∈ V1, then the word w satisfies
condition (1). Similar statements hold for w′, w′′ and w′′′. Thus we get the desired
expression.
(cid:3)
Claim 3. Let y be a linear component of (ξm)M+r−j,M+s−j′ , z be a linear
component of (ξm′ )M+r′−j′,M+s′−j′′ . Consider a word h∗j′′′,jyxw′gj′,j′′ w′′zw′′′. If
it satisfies τ (h∗j′′′ ,jwyw′gj′,j′′ w′′zw′′′) 6= 0, then at least one of the following state-
ments holds.
i
(1) We have w = u±M
.
(2) We have w′ = u±M
.
(3) We have w′′ = u±M
.
(4) We have w′′′ = u±M
.
i
i
i
Proof of Claim 3. Assume that none of statements (1) -- (4) holds. We would like
to show that τ (h∗j′′′,jwyw′gj′,j′′ w′′zw′′′) = 0. Let
i1 wi1 ··· ukn
j1 vq1
j1 ··· upm
xi′
1 ··· u
vq′
1 ··· u
1
j′
be the completely reduced word expressions.
h∗j′′′,j = uk1
wyw′ = up1
gj′,j′′ = uk′
1
i′
1
w′′zw′′′ = up′
1
j′
1
jm
k′
n′
i′
n′
p′
m′
j′
m′
in
win ,
vqm
jm
,
xi′
n′ ,
q′
m′
v
j′
m′
Case 1: neither h∗j′′′,j nor gj′,j′′ is 1. In order to get the conclusion, it is enough
to show that the following.
l′′
vl′′
(i) The vector h∗j′′′,jwyw′ is a linear sum of words of the form uk′′
n′′
1 ··· u
v
.
1
i′′
i′′
n′′
l′′′
k′′′
(ii) The vector gj′,j′′ w′′zw′′′ is a linear sum of words of the form uk′′′
vl′′′
n′′′
n′′′
v
1 ··· u
1
1
i′′′
i′′′
i′′′
i′′′
n′′′
n′′′
1
(iii) Any pair of linear components of h∗j′′′,jwyw′ and gj′,j′′ w′′zw′′′ does not
k′′
n′′
i′′
n′′
1
i′′
1
.
cancel at all.
26
KOICHI SHIMADA
Case 1-1: both win and xi′
n′ are 1. In this case, we have statements (i) -- (iii)
without any assumption.
Case 1-2: neither win nor xi′
n′ is 1. In Claim 2, we temporary thought that the
letters wi and xi were free from ui. However, in this claim, we do not. This may
cause h∗j′′′,j and w to cancel. Nonetheless, the assumption that w 6= u±M
ensures
that if we reduce h∗j′′′,jwyw′ as possible, it is either of the form
vqm
jm
i1 w∗in )win up2
(win up1
j2 vq2
i
i1 wi1 ··· ukn
uk1
in
j2 ··· upm
jm
(in 6= j2) or
i1 wi1 ··· ukn
uk1
in
win up1
j1 vq1
j1 ··· upm
jm
vqm
jm
(in 6= j1). Since win is orthogonal to the subalgebra {uin}′′ and normalizes it, by
freeness, the vector h∗j′′′,jwyw′ satisfies condition (i). Similarly, since w′′ 6= u±M
,
the vector gj′,j′′ w′′zw′′′ satisfies condition (ii). Since w′ 6= u±M
, if k′1 = 0, then
condition (iii) is shown by the same argument. If k′1 6= 0, then w′ and gj′,j′′ does
not cancel at all. Hence condition (iii) is trivial.
i
i
Case 1-3: exactly onr of win and xi′
n′ is 1. This case is treated by the combi-
nation of the arguments in Cases 1-1 and 1-2.
Case 2: exactly one of h∗j′′′,j and gj′,j′′ is 1. For simplicity, we assume gj′,j′′ = 1
(the other case is handled in the same way). We have
τ (h∗j′′′ ,jwyw′gj′,j′′ w′′zw′′′) = τ (w′ · (w′′zw′′′h∗j′′′,jwy)).
Since any component of h∗j′′′,j contains v±1
, for τ (w′·(w′′zw′′′h∗j′′′,jwy)), in order to
be non-zero, at least one of z and y contains v±1
. Since neither w′′′ nor w is u±M
,
by the same argument as that of Case 1, any component of w′′zw′′′h∗j′′′,jwy contains
at least one v±1
if we reduce it as possible. Thus we have τ (w′·(w′′zw′′′h∗j′′′,jwy)) =
0.
(cid:3)
j
j
j
i
Set
η1,j,j′ := 3−M+ j+j′
usM
i ∈V1
i=1,2, s∈{±1}, usM
2 Xr,s,mXy Xw∈V1
2 Xr,s,mXy
X
2 Xr,s,mXy Xw∈V1, w6=usM
2 Xr′,s′,m′Xz Xw′′∈V3
2 Xz
X
w!λm,2M+r,2M+sy Xw′∈V2
w′!.
i !λm,2M+r,2M+sy Xw′∈V2
w!λm,2M+r,2M+sy
X
w′′!µm′,2M+r′,2M+s′ z Xw′′′∈V4
w′′′!,
i !µm′,2M+r′,2M+s′ z Xw′′′∈V4
i=1,2, s∈{±1}, usM
i=1,2, s∈{±1},usM
i ∈V3
usM
i
w′!,
usM
i !.
i ∈V2
w′′′!,
η′1,j,j′ := 3−M+ j+j′
η′′1,j,j′ := 3−M+ j+j′
η∗2,j′,j′′ := 3−M+ j′′ +j′′′
η′∗2,j′,j′′ := 3−M+ j′′ +j′′′
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS27
η′′∗2,j′,j′′ := 3−M+ j′′ +j′′′
2 Xz Xw′′∈V3, w′′6=usM
i
w′′!µm′,2M+r′,2M+s′ z
X
i=1,2, s∈{±1}, usM
i ∈V4
usM
i !.
Then we have
kh∗j′′′,jη′1,j,j′k2 = kη′1,j,j′k2
λm,2M+r,2M+sy Xw′∈V2
λm,2M+r,2M+sy
w′!k2
Xw′∈W 0
M , yw′=y+M
w′!k2
≤ 4 · 3−M+ j+j′
≤ 4 · 3−M+ j+j′
= 4 · 3−M+ j+j′
3M/2 Xr,s,m
≤ 4
C0
2 k Xr,s,mXy
2 k Xr,s,mXy
2 k Xr,s,m
λm,2M+r,2M+s2!1/2
λm,2M+r,2M+s(ξm)M+r−j,2M+s−j′k2
C2
0 · 3k
3M/2 kη1k2.
≤ 4
Similar statements holds for other three vectors. We also have
kh∗j′′′,j(η1,j,j′ − (η′1,j,j′ + η′′1,j,j′ ))k2
≤ kh∗j′′′,jη1,j,j′k2
λm,2M+r,2M+s Xw∈V1
≤ 3−Mk Xr,s,mXy
w!y Xw′∈V2
w′!k2.
By the same argument as above, the above left hand side is not greater than
3−M · 3M · C0 Xr,s,m
λm,2M+r,2M+s2!1/2
≤ C2
0kη1k2.
Hence we have
hη1g, hη2i = τ (h∗η1gη∗2)
≤ Xj,j′,j′′,j′′′ τ (h∗j′′′ ,j(η′1,j,j′ + η′′1,j,j′ )gj′,j′′ η∗2,j,j′ )
+ τ (h∗j′′′ ,j(η1,j,j′ − (η′1,j,j′ + η′′1,j,j′ ))gj′,j′′ (η′∗2,j′,j′′ + η′′∗2,j′,j′′ )
+ τ (h∗j′′′ ,j(η1,j,j′ − (η′1,j,j′ + η′′1,j,j′ ))gj′,j′′ (η∗2,j′,j′′ − (η′∗2,j′,j′′ + η′′∗2,j′,j′′ ))!.
By Claim 3, the third term in the above sum is zero. Hence the right hand side is
not greater than
M 4 4 · 4 ·
C4
0
3M/2kη1k2kη2k2 + 0! ≤ 16 · C4
0 ·
M 4
3M/2kη1k2kη2k2.
Thus we are done.
We also have the following.
(cid:3)
28
KOICHI SHIMADA
Lemma 27. Let g1, g2 be words of W 0
l
be a vector satisfying the following condition.
for some non-negative integer l ≥ 0. Let h
The vector h is of the form uk′
1
i′
1
xi′
1 ··· uk′
n
i′
n′
xi′
n′
k′1,··· , k′n ∈ Z, where for any t = 1,··· , n′, the operator uk′
(c) k′t 6= 0, xi′
is a normalizing unitary of {ui′
orthogonal to {ui′
= 1 or (d) xi′
t}′′.
t
t
for some n′ ≥ 1, i′1 6= ··· 6= i′n′,
satisfies either
which is
xi′
t}′′ in Ri′
t
i′
t
t
t
Set k := k′1 +···k′n (Actually, this depends on the presentation of h. However,
we fix the presentation or take the minimum). Then there exists a constant C > 0,
which does not depend either on none of xit 's, such that for any M > 2max{l, k},
any vectors η1, η2 of
span{(ξm)M+r,M+s m ≥ 0, r, s ≥ 0}∨ span{ξi,l,k
M+r,M+s i = 1, 2, l ∈ Z\{0}, k ∈ Z, r, s ≥ 0}∩L!,
we have
hη1(g1 − g2), hη2i ≤ CM 43−M/2kη1k2kη2k2.
Proof. When h ∈ W 0
l , then the lemma is shown by the same argument as that of
the proof of Lemmas 5.1, 5.2, 5.3 and 6.1 of Cameron -- Fang -- Ravichandran -- White
[4]. When h 6∈ W 0
(cid:3)
l , then the lemma is shown by Lemma 27.
6. Asymptotic orthogonality property of the subalgebra
In this section, by using the results of Sections 4 and 5, we show that the subal-
gebra has the asymptotic orthogonality property. In order to achieve this, we have
to reduce the general cases to the special case, that is, the case when the Haar
unitaries come from generators of the irrational rotation C∗-algebras.
Lemma 28. Let α be a free action of Z/mZ on a diffuse separable abelian von
Neumann algebra C and β be an ergodic action of Z on a diffuse separable abelian
von Neumann algebra D. Let u, v be generating Haar unitaries of C and D,
respectively. Then for any positive number ǫ > 0, there exists a normal injec-
tive *-homomorphism θ from C ⋊α Z/mZ into D ⋊β Z satisfying θ(C) = D,
kθ(u) − vk2 < ǫ.
Proof. Since both u and v are generating Haar unitary, the map uk 7→ vk extends
to a *-isomorphism θ0 from C onto D satisfying θ0(u) = v. Since θ0 ◦ α ◦ θ−1
is a
free action of Z/(mZ) on D, it is possible to find a partition {pi}m
i=1 of unitary by
projections in D with
0
θ0 ◦ α ◦ θ−1
0 (pi) = pi+1
for any i = 1,··· , m, where pm+1 = p1. Then there exists a partition {ql}L
l=1 of p1
by projections in D such that there exist complex numbers {λ(i, l)}i=1,··· ,m,l=1,··· ,L ⊂
C with
Set
m
L
k
Xi=1
Xl=1
λ(i, l)θ0 ◦ αi−1 ◦ θ−1
0 (ql) − vk2 <
ǫ
2
.
v0 :=
m
L
Xi=1
Xl=1
λ(i, l)θ0 ◦ αi−1 ◦ θ−1
0 (ql).
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS29
Since the action β of Z on D is ergodic, there exists a unitary w of D ⋊β Z satisfying
the following two conditions.
(1) The unitary w normalizes D.
(2) We have wi−1qlw−i+1 = θ0◦αi−1◦θ−1
(3) We have wm = 1.
0 (ql) for any l = 1,··· , L, i = 1,··· , m+
1.
For x ∈ C, set
m
θ(x) :=
θ0 ◦ αi−1 ◦ θ−1
0 (p1)wi−1(θ0 ◦ α−i+1(x))w−i+1.
Xi=1
Claim 1. The map θ is a *-isomorphism from C onto D.
Proof of Claim 1. Notice that the map θ maps each αi−1 ◦ θ−1
θ−1
0 (ql) because we have
0 (ql) to θ0 ◦ αi−1 ◦
0 (q′l))
θ(αi′−1 ◦ θ−1
=Xi
=Xi
=Xi
=Xi
= θ0 ◦ αi′−1 ◦ θ−1
θ0 ◦ αi−1 ◦ θ−1
θ0 ◦ αi−1 ◦ θ−1
θ0 ◦ αi−1 ◦ θ−1
θ0 ◦ αi−1 ◦ θ−1
0 (ql′ ).
0 (ql′ )))w−i+1
0 (p1)wi−1(θ0 ◦ α−i+1(αi′−1 ◦ θ−1
0 (p1)wi−1(θ0 ◦ α−i+i′
0 (p1)wi−1(wi′−iql′ w−(i′−i))w−i+1
◦ θ−1
0 (ql′ ))w−i+1
0 (p1)θ0 ◦ αi′−1 ◦ θ−1
0 (ql′ )
Note that the third equality of the above computation follows from conditions (2)
and (3). On the other hand, the map Adwi−1 ◦ θ0 ◦ α−i+1 is a *-isomorphism from
C onto D. Hence so is its restriction to Cαi−1◦θ−1
0 (ql). Thus θ is a *-isomorphism
from C onto D.
(cid:3)
Claim 2. We have θ ◦ α = (AdwD) ◦ θ.
Proof of Claim 2. For x ∈ C, we have
wθ(x)w−1
0 (p1)wi−1(cid:18)θ0 ◦ α−i+1(x)(cid:19)w−i+1!w−1
=
θ0 ◦ αi−1 ◦ θ−1
= w m
Xi=1
Xi=1
(wip1w−i)(wiθ0 ◦ α−i+1(x)w−i)
Xi=1
θ0 ◦ αi ◦ θ−1
=
m
m
= θ(α(x)).
0 (p1)wi(θ0 ◦ α−i(α(x)))w−i
Claim 3. We have θ(θ−1
0 (v0)) = v0.
(cid:3)
30
KOICHI SHIMADA
Proof of Claim 3. As we have see in the proof of Claim 1, we have θ(αi−1 ◦
0 (ql)) = θ0 ◦ αi−1 ◦ θ−1
θ−1
0 (ql),
we have θ(θ−1
0 (v0)) = v0.
(cid:3)
0 (ql). Since v0 is a linear combination of θ0 ◦ αi−1 ◦ θ−1
By Claim 2, the *-isomorphism θ extends to a *-isomorphism form C ⋊α(Z/(mZ)
into D ⋊β Z satisfying θ(C) = D, θ(λα
1 ) = w. By Claim 3, we also have
kθ(u) − v)k2 ≤ kθ(u − θ−1
0 (v0))k + kv0 − vk
= ku − θ−1
0 (v0)k2 + kv0 − vk2
= kθ0(u) − v0k2 + kv0 − vk2
= 2kv0 − vk2
< ǫ.
(cid:3)
Let N be the free product of two hyperfinite factor Ri (i = 1, 2) of type II1 with
respect to their traces. For each i = 1, 2, choose a Haar unitary wi of Ri which
generates a Cartan subalgeba of Ri. Set
B := {w1 + w−1
1 + w2 + w−1
2 }′′ ⊂ N.
For each non-negative integer l ≥ 0, let χB
l be the sum of all reduced words of
{w±1
i }i=1,2 with length l. For each i = 1, 2, think of Ri as an increasing union
{Bi ⋊αi Gk}∞k=1 of von Neumann algebras of type I, where Bi := {wi}′′, {Gk}∞k=1 is
an increasing sequence of finite abelian groups and αi be a free ergodic probability
measure preserving action of Sk Gk on Bi. Set
Nk := (B1 ⋊α1 Gk) ∗ (B2 ⋊α2 Gk),
which is a von Neumann subalgebra of N .
Proposition 29. Let B ⊂ N be as above. For j = 1,··· , J , choose (xj
n) ∈
(N ω ⊖ Bω)∩ B′. Assume that for each n, there exists a positive integer kn > 0 with
xj
n ∈ Nkn for all j = 1,··· , J .
Then there exists a family of weakly continuous injective *-homomorphisms {θn :
Nkn → M}∞n=1 satisfying the following conditions.
(1) For each i = 1, 2, we have θn(wi) → ui as n → ω.
(2) For any n, any normalizing unitary x ∈ Bi ⋊αi Gkn of Bi which is orthogonal
(3) We have (θn(xj
to Bi, the unitary θn(x) normalizes {ui}′′ and is orthogonal to {ui}′′.
n)) ∈ (M ω ⊖ Aω) ∩ A′.
Proof. By the previous lemma, there exists a *-homomorphism θ
into {ui, vi}′′ satisfying θ
which is a *-homomorphism from Nk into M .
k(Bi) = {ui}′′, kθ
k(wi) − uik2 < 2−k. Set θk := θ
i
k from Bi ⋊αi Gk
2
k ∗ θ
k,
1
i
i
Claim. There exists an increasing sequence {k′n} of natural numbers with the
following conditions.
(1) For any n, we have k′n > kn.
(2) For any n, we have kEA((θk′
(3) For any n, we have kθk′
n (xj
n))n)k2 ≤ kxj
n (wi) − uik2 < 1/2n.
nk2/2n.
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS31
Proof of Claim. For each n and j, since xj
n is orthogonal to B, xj
n can be
approximated by a linear sum of vectors of the following forms.
(1) The vectors b1 − b2, where b1 and b2 are words of w±1
(2) The words of λi
i
g, w±1
j
of the same length.
for some i, j = 1, 2, g ∈ (Sk Gk) \ {0} with at least
one λi
g.
If gt 6= 0 for some t, then we have
EA(θk(wl1
i1 λi1
g1 ··· wlm
im
λim
gm )) = 0,
which implies that the image of any vector of the second form by EA ◦ θk converges
to 0 as k → ∞. We also have θk(wl1
, which implies that
any vector of the first form converges to a vector orthogonal to A. Thus we have
EA(θk(xj
n)) → 0 as k → ∞. We also have θk(wi) → ui as k → ∞. Thus Claim
holds.
i1 ··· wkm
i1 ··· ulm
) → ul1
im
im
(cid:3)
Set θn := θk′
n .
Then the family {θn : Nk′
n))k2 < 1/2n−1 → 0 as n → ω, we have (θn(xj
n → M} is a family of weakly continuous injective *-
homomorphisms satisfying conditions (1) and (2). We show condition (3). Since we
have kEA(θn(xj
n)) ∈ M ω ⊖ Aω. Next,
we show that θ(x) commutes with A. Take y ∈ A. Since we have kθ−1
n (ui)− wik2 =
kui − θn(wi)k2 → 0 as n → ω, the sequence {θ−1
n (y)} converges to an operator z of
B in the strong operator topology. Thus we have xnθ−1
n (y)xn converges
to 0 as n tends to ω. Hence θ(x) commutes with A.
(cid:3)
Remark 30. For each x ∈ N , set θ(x) := (θn(Ekn (x)))n ∈ M ω. Then the map θ
is a weakly continuous injective *-homomorphism from N into M ω.
n (y) − θ−1
Theorem 31. The subalgebra B has the asymptotic orthogonality property.
n), x2 = (x2
n) ∈ (N ω ⊖ Bω) ∩ B′ and y1, y2 ∈ N ⊖ B.
Proof. Choose x1 = (x1
Let θn : Nkn → M , θ : N → M ω be injective *-homomorphisms chosen in the
n). Then we have θ(x1), θ(x2) ∈
previous proposition corresponding to (x1
(M ω ⊖ Aω) ∩ A′. Hence by Lemmas 27 and 25, we may assume that the vectors
θn(x1
n), (x2
n) and θn(x2
n) lie in
span{(ξm)2Mn+r,2Mn+s m ≥ 0, r, s ≥ 0}
∨ span{ξk
2Mn+r,2Mn+s k ∈ Z, r, s ≥ 0} ∩ L!,
where Mn → ω as n → ω. On the other hand, when we regard N as Sk Nk
the Hilbert subspace N ⊖ B is linearly spanned by the following things.
(1) The vectors b1 − b2, where b1 and b2 are words of w±1
(2) The words of λi
of the same length.
i
g, w±1
j
for some i, j = 1, 2, g ∈ (Sk Gk) \ {0} with at least
one λi
g.
weak
,
Hence we may assume that y1 and y2 are of the above forms. Then by Proposition
29 (1) (2), for each n, the vectors θn(y1) and θn(y2) satisfy assumptions of Lemmas
27 and 26. Hence by Lemmas 27 and 26, we have τ (y∗1 (x1
n) → 0 as n → ω. (cid:3)
n)∗y2x2
32
KOICHI SHIMADA
7. The subalgebra is maximal amenable.
In this section, we show that the subalgebra B is maximal amenable. In order
to achieve this, we use Proposition 20. In the previous section, we have already
shown that the subalgebra B has the asymptotic orthogonality property. Hence in
order to show that the subalgebra is maximal amenable, it is enough to show that
it is singular. In order to achieve this, we show the mixing property.
Definition 32. (Definition 3.1 of Cameron -- Fang -- Mukherjee [3]) Let A be a diffuse
abelian von Neumann subalgebra of a factor M of type II1. The subalgebra A is said
to be mixing if for any a, b ∈ M ⊖ A, any sequence {un} of unitaries of A which
converges to 0 weakly, we have kEA(aunb)k2 → 0.
It is known that if a subalgebra is mixing, then it is singular (Proposition 1.1 of
Jolissaint -- Stalder [10]).
Lemma 33. Let w be a reduced word of {w±1
words of the forms
g1 wk1
i1 ··· λin
λj1
h1 ··· wlm
g = λi1
h = wl1
j1
,
gn wkn
in
λjm
hm
jm
,
1 , w±1
2 } with w = M > 0 and g, h be
respectively. Suppose that at least one of gs's is non-zero and that one of ht's is
non-zero. Assume that they satisfy the following conditions.
(1) The word w is not w±M
(2) The words wkn
in
(3) The words w and wl1
j1
i
.
and w do not cancel, that is, we have wkn
in
w = wkn
in w.
do not cancel.
Then we have EB(gwh) = 0.
Proof. This is shown by the same argument as that of Claim 3 of the proof of
Lemma 26.
(cid:3)
i with length p. Then there exists a positive constant
Lemma 34. Let v =Pp≥M λpwp/kwpk2 be a vector of L2B with kvk2 = 1, where
wp is the sum of all words of w±1
Cg,h > 0 with kEB(gvh)k2 ≤ Cg,h3M/2.
Proof. If either g or h is in W 0
l , then there is nothing to show (Here, we use the
mixing property of the radial MASA implicitly. See Sinclair -- Smith [15] Theorems
3.1 and 5.1. Although the statements are slightly different, they essentially show
the mixing property of the radial MASA). Write g and h as the following way.
g = λi1
h = wl1
g1 wk1
j1 λj1
i1 ··· λin
h1 ··· wlm
gn wkn
in
λjm
hm
jm
,
.
Set g := kn, h := l1. By the previous lemma, we have kEB(gwph)k2 ≤ 4gh.
Thus we have
kEB(gvh)k2 ≤ 4gh(cid:18)Xp≥M
≤ 4gh(cid:18)Xp≥M
1
λp
3p/2(cid:19)
λp2(cid:19)1/2(cid:18)Xp≥M
1
3p(cid:19)1/2
≤ Cg,h
1
3M/2 .
(cid:3)
MAXIMAL AMENABLE MASAS ARISING FROM FREE PRODUCTS OF HYPERFINITE FACTORS33
Proposition 35. The subalgebra B is mixing in N . In particular, it is singular in
N .
Proof. Notice that if uk → 0 weakly, then uk is approximated by operators v of the
form Pp≥Mk
λpwp/kwpk2 in the strong operator topology, where Mk → ∞. Hence
this is obvious by the previous lemma.
(cid:3)
Theorem 36. The subalgebra B is maximal amenable in N .
Proof. By Theorem 31, the subalgebra B has the asymptotic orthogonality prop-
erty. By Lemma 35, the subalgebra B is a singular abelian von Neumann subalgebra
of N . Thus by Proposition 20, the subalgebra B is maximal amenable in N .
(cid:3)
References
[1] R. Boutonnet and A. Carderi, Maximal amenable von Neumann subalgebras arising from
maximal amenable subgroups, Geom. Funct. Anal. 25 (2015) 1688 -- 1705.
[2] R. Boutonnet and C. Houdayer, Amenable absorption in amalgamated free product von
Neumann algebras, preprint, arXiv:1606.00808.
[3] J. Cameron, J. Fang and K. Mukherjee, Mixing subalgebras of finite von Neumann algebras,
New York J. Math. 19 (2013), 343 -- 366.
[4] J. Cameron, J. Fang, M. Ravichandran and S. White, The radial masa in a free group factor
is maximal injective, J. London Math. Soc. (2) 82 (2010), 787 -- 809.
[5] K. Dykema, Free products of hyperfinite von Neumann algebras and free dimension, Duke
Math. J. 69 (1993), 97 -- 119.
[6] L.Ge, On maximal injective subalgebras of factors, Adv. Math. 118 no. 1 (1996), 34 -- 70.
[7] A. Guionnet and D. Shlyakhtenko, Free monotone transport, Invent. Math. 197 no. 3 (2014),
613 -- 661.
[8] C. Houdayer, Gamma stability in free product von Neumann algebras, Comm. Math. Phys.
336 (2015), 831 -- 851.
[9] C. Houdayer and Y. Ueda, Asymptotic structure of free product von Neumann algebras,
preprint arXiv:1503.02460, to appear in Math. Proc. Cambridge Philos. Soc.
[10] P. Jolissaint, Y. Stalder, Strongly singular MASAs and mixing actions in finite von Neumann
algebras, Ergodic Theory Dynam. Systems 28 no. 6 (2008), 1861 -- 1878.
[11] N. Ozawa, A remark on amenable von Neumann subalgebras in a tracial free product, Proc.
Japan Acad. Ser. A Math. Sci. 91 no. 7 (2015), 104.
[12] S. Popa, Maximal injective subalgebras in factors associated with free groups, Adv. Math.
50 no. 1 (1983), 27 -- 48.
[13] S. Popa, Orthogonal pairs of *subalgebras in finite von Neumann algebras, J. Oper. Theo. 9
no. 2 (1983), 253 -- 268.
[14] R. Radulescu, Singularity of the radial subalgebra of L(FN ) and the Pukanszky invariant.
Pacific. J. Math. 151 no. 2 (1991), 297 -- 306.
[15] A. Sinclair and R. Smith, The Laplacian MASA in a free group factor, Trans. Amer. Math.
Soc. 355 no. 2 (2003), 465 -- 475.
[16] C. Wen, Maximal amenability and disjointness for the radial masa, J. Funct. Anal. 270 no.
2 (2016), 787 -- 801.
E-mail address: [email protected]
Department of Mathematics, Kyoto University, Kita-shirakawa Oiwake-cho, Sakyo-
ku, Kyoto, 606-8502, Japan
|
1301.2618 | 1 | 1301 | 2013-01-11T22:02:28 | Type II$_1$ factors with arbitrary countable endomorphism group | [
"math.OA",
"math.DS",
"math.GR"
] | In \cite{Ioana:vNsuperrigidity}, Ioana introduced three new invariants of type II$_1$ factors: the one-sided fundamental group, the endomorphism semigroup and the set of right-finite bimodules. In \cite{Ioana:vNsuperrigidity}, he does not provide many computations of these invariants. In particular, the question whether these invariants can be trivial is left open. We give an explicit example of a type II$_1$ factor for which all three invariants are trivial. More generally, for any countable left-cancellative semigroup $G$, we construct a type II$_1$ factor $M$ whose endomorphism semigroup is precisely $G$. | math.OA | math |
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE
ENDOMORPHISM GROUP
STEVEN DEPREZ
Abstract. In [I], Ioana introduced three new invariants of type II1 factors: the one-sided
fundamental group, the endomorphism semigroup and the set of right-finite bimodules. In
[I], he does not provide many computations of these invariants. In particular, the question
whether these invariants can be trivial is left open. We give an explicit example of a type
II1 factor for which all three invariants are trivial. More generally, for any countable left-
cancellative semigroup G, we construct a type II1 factor M whose endomorphism semigroup
is precisely G.
Introduction and Overview of the paper
In [I, section 10.(II)], Ioana introduced three new invariants of type II1 factors, but he provided
few concrete computations. Here we provide a large class of type II1 factors where we can
compute these invariants. The invariants in question are "one-sided versions" of three classical
invariants. Let M be a type II1 factor. Then the one-sided fundamental group F s(M ) is
defined to be
F s(M ) = {t ∈ R∗
+ there is a normal injective ∗ -homomorphism ϕ : M → M t}.
Observe that this set contains 1, and is closed under multiplication and under taking sums. In
+ whenever (0, 1)∩
fact, it is closed under taking infinite sums. This implies that F s(M ) = R∗
F s(M ) (cid:54)= ∅. In particular, whenever the fundamental group of M is non-trivial, it follows
+. Similarly, it follows that F s(L(Fn)) = R∗
that the one-sided fundamental group is all of R∗
+
for all n ∈ N, because L(Fn) ⊂ L(Fn+1) = L(Fn)t where t =
n < 1. Our examples are all
on the other side of the spectrum: they satisfy F s(M ) = N.
(cid:113) n−1
The second invariant that Ioana introduced is the one-sided version on the outer automor-
phism group. This is called the endomorphism semigroup End(M ). It is the set of all normal
injective ∗-homomorphisms ϕ : M → M , and two such ∗-homomorphisms ϕ1, ϕ2 are identified
if there is a unitary u ∈ M such that ϕ1 = Adu ◦ϕ2. This way, it is clear that End(M ) is a
unital semigroup. But it does not have to be left nor right cancellative. For example, End(R),
where R is the hyperfinite II1 factor, is neither left nor right cancellative. This is easy to see
Postdoc at the university of Copenhagen (from September 2011).
Partially supported by ERC Advanced Grant no. OAFPG 247321.
Supported by the Danish National Research Foundation (DNRF) through the Centre for Symmetry and
Deformation.
Department of mathematics, Copenhagen university, Universitetsparken 5, 2500 O, Copenhagen).
email: [email protected].
1
2
STEVEN DEPREZ
explicitly: remember that R ∼= R⊗ R ∼= (R⊗ R)(cid:111) (Z /2), where Z /2 acts outerly on R⊗ R by
swapping the components of the tensor product. Write ϕ1, ϕ2 : R → R⊗R for the embeddings
that are given by ϕ1(x) = x ⊗ 1 and ϕ2(x) = 1 ⊗ x. Denote by ψ : R ⊗ R → (R ⊗ R) (cid:111) (Z /2)
the obvious embedding. Then we see that
(id⊗ϕ1) ◦ ϕ1 = (id⊗ϕ2) ◦ ϕ1 : R → R ⊗ R ⊗ R
ψ ◦ ϕ1 = ψ ◦ ϕ2 in End(R)
but id⊗ϕ1 (cid:54)= id⊗ϕ2 in End(R)
but ϕ1 (cid:54)= ϕ2 in End(R).
Even though in general End(M ) does not have to be left cancellative, our examples will be.
We show that every countable left-cancellative unital semigroup appears as the endomor-
phism semigroup of some type II1 factor. In particular, we find a type II1 factor with trivial
endomorphism semigroup. This solves Ioanas question for an example of such a type II1
factor.
In fact, we show even more. Ioana introduced a third invariant that contains both the one-
sided fundamental group and the endomorphism semigroup. This is the set RFBimod(M ) of
all M -M bimodules H that have finite dimension as a right M -module, up to isomorphism of
M -M bimodules. This set is closed under the Connes tensor product and under finite direct
sums. It is even closed under infinite direct sums, provided that the dimensions (i.e. the right
dimensions over M ) form a convergent series.
This invariant contains the previous two invariants. The set of all right dimensions of M -M
bimodules is precisely the one-sided fundamental group. Moreover, the sum and product
in R∗
+ correspond to the direct sum and the Connes tensor product. The endomorphism
semigroup corresponds precisely to the set of all M -M bimodules with right dimension equal
to 1, and the product in End(M ) corresponds to the Connes tensor product in RFBimod(M ).
We give an example of a type II1 factor for which RFBimod(M ) is as small as possible, i.e.
all right-finite M -M bimodules are trivial bimodules (direct sums of L2(M )).
The results in this paper are based on Popas deformation/rigidity theory. More precisely, we
combine techniques and results from [P2, P3], [IPP], [PV4, PV6], [IPV] and [PV1, PV2] in
order to reduce the computation of End(M ) to a problem in ergodic theory.
The ergodic-theoretic problem is the following. Let Λ (cid:121)(Y, ν) be an ergodic probability
measure preserving (p.m.p.) action of a not necessarily countable group. Another probability
measure preserving action Λ (cid:121)(Z, η) of the same group is said to be a factor of Λ (cid:121)(Y, ν) if
there is a p.m.p. quotient map ∆ : Y → Z such that ∆(λy) = λ∆(y) for almost all y ∈ Y and
this for all λ ∈ Λ. The map ∆ is called a factor map. We denote by Factor(Λ (cid:121)(Y, ν)) the
set of all factor maps from (Y, ν) to itself. Composition of factor maps defines a semigroup
operation, and the identity map is the identity element for this operation. This way, we see
that Factor(Λ (cid:121)(Y, ν)) is a right-cancellative semigroup.
Given an ergodic action Λ (cid:121)(Y, ν), we construct a type II1 factor MY such that End(MY ) =
Factor(Λ (cid:121)(Y, ν))op. It is easy to see that every compact group G is G = Factor(G (cid:121)(G, h))
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
3
where h denotes the Haar measure on G. Hence G appears also as the endomorphism semi-
group of some type II1 factor M . In these cases, all endomorphisms of M are in fact isomor-
phisms. Every compact right-cancellative unital semigroup is automatically a group, so this
observation covers whole compact right-cancellative case.
The discrete case is more interesting. Here we have examples of semigroups that are not
groups. Already the semigroup of natural numbers with addition form such a semigroup. In
section 6, we show that every countable right-cancellative semigroup appears as the factor
semigroup of an ergodic p.m.p. action Λ (cid:121)(Y, ν). Hence every countable left-cancellative
semigroup appears as End(M ) for some type II1 factor M .
Given Λ (cid:121)(Y, ν), we construct MY as follows. We can consider Λ (in fact, a quotient of Λ)
as a subgroup of Autν(Y ). Observe that Factor(Λ (cid:121)(Y, ν)) only depends on the closure of
Λ in the usual Polish topology on Autν(Y ). Hence we can replace Λ by a countable dense
subgroup without changing Factor(Λ (cid:121)(Y, ν)). From now on we assume that Λ is countable.
Let Γ1 be a hyperbolic property (T) group with trivial endomorphism semigroup. Let Σ ⊂ Γ1
be an amenable subgroup. Consider Γ = Γ1 ∗Σ(Σ × Λ). Let Γ (cid:121) I be an action of Γ on a
countable set.
Let (X0, µ0) be an atomic probability space with unequal weights, and consider the generalized
Bernoulli action Γ (cid:121)(X, µ) = (X0, µ0)I . Consider the obvious quotient map π : Γ → Λ.
Following [PV4, PV6], we define an action of Γ on X×Y by the formula g(x, y) = (gx, π(g)y).
Then we set MY = L∞(X × Y ) (cid:111) Γ.
Let us give an idea why End(MY ) = Factor(Λ (cid:121)(Y, ν))op. One inclusion is easy. Given
∆ ∈ Factor(Λ (cid:121)(Y, ν)), we define an embedding ϕ∆ : MY → MY by the formula ϕ((a ⊗
b)ug) = (a ⊗ ∆∗(b))ug, where ∆∗(b) = b ◦ ∆ for every function b ∈ L∞(Y ). The application
∆ (cid:55)→ ϕ∆ embeds Factor(Λ (cid:121) Y )op into End(MY ).
Now, let ϕ : MY → MY be an endomorphism of MY . We want to show that ϕ = ϕ∆ up
unitary conjugacy. Denote A = L∞(X) and B = L∞(Y ). Techiques from [P2, P3] show that
ϕ(B(cid:111)Γ) ⊂ B(cid:111)Γ, up to a unitary. This result depends crucially on the fact the Γ1 has property
(T) while the action Γ (cid:121) X is a generalized Bernoulli action. Similarly, techniques from [IPP]
show that ϕ((A ⊗ B) (cid:111) Γ1) ⊂ (A ⊗ B) (cid:111) Γ1 up to unitary conjugacy. In fact, we can assume
that both unitaries are the same. This result uses the facts that Γ1 has property (T) while Γ
is an amalgamated free product. Because Γ1 is hyperbolic, [PV1, PV2] shows that ϕ(A) can
not be in B (cid:111) Γ1 = B ⊗ L(Γ1). Then [IPV, theorem 5.1] shows that C = ϕ(A)(cid:48) ∩ (A⊗ B) (cid:111) Γ1
embeds into A ⊗ B. Now we apply [IPV, theorem 6.1] to conclude that ϕ(A ⊗ B) ⊂ A ⊗ B.
Moreover, the endomorphism ϕ : B ⊗ L(Γ1) → B ⊗ L(Γ1) is described in the following way.
We can consider every element in B ⊗ L(Γ1) as a map from Y to L(Γ1). There is a field of
group endomorphisms δy : Γ1 → Γ1 (y ∈ Y ) such that ϕ(ug)(y) = uδy(g). All of these are
inner, so we can assume that they are all trivial. Now we know that ϕ(A ⊗ B) ⊂ A ⊗ B and
ϕ(ug) = ug for all g ∈ Γ1. For a good choice for the action Γ (cid:121) I, a direct computation shows
that in fact ϕ = ϕ∆ for some factor map ∆ ∈ Factor(Λ (cid:121) Y ).
4
STEVEN DEPREZ
Of course, in the above idea of the proof, we have been ignoring a lot of technical conditions.
A more precise statement and proof are given in 4. In section 1, we remind the reader of
some well-known results that are crucial for this paper. In section 2, we extend [IPV, theorem
5.1 and 6.1] to our setting. Section 3 introduces two properties of groups that are crucial in
the next section. There we show our main result, theorem 4.1. In order to apply that main
theorem, we need to give an example of a group Γ and an action Γ (cid:121) I that satisfies the
conditions of theorem 4.1. This is not very hard, but it is technical. Section 5 is devoted
to such an example. Finally, in section 6, we show that all countable right-cancellative
semigroups appear as End(M ) for some type II1 factor M .
1. Preliminaries and Notations
1.1. Relatively weakly mixing actions. Relative weak mixing plays a crucial role in the
proof of theorem 2.2. This property was introduced by Furstenberg in [F] and Zimmer in
[Z1, Z2], in the case of actions on probability spaces. In [P4], Popa generalized this to actions
on von Neumann algebras.
Definition 1.1 (see [P4, lemma 2.10]). Let D ⊂ (B, τ ) be an inclusion of finite von Neumann
algebras. Assume that a countable group Γ acts trace-preservingly on B and leaves D globally
invariant. Denote the action by α. We say that Γ acts weakly mixingly on B relative to D if
one of the following equivalent conditions holds.
(1) There exists a sequence of group elements (gn)n in Γ such that
for all x ∈ B, y ∈ B (cid:9) D.
(cid:107)ED(xαgn(y))(cid:107)2 → 0
(2) Every Γ-invariant positive element a with finite trace in the basic construction (cid:104)B, eD(cid:105),
(3) The action of Γ on L2(B) ⊗D L2(B) is ergodic relative to L2(D). I.e. all Γ-invariant
must be a ∈ eDDeD.
vectors ξ ∈ L2(B) ⊗D L2(B) are ξ ∈ L2(D).
(4) The only right D-submodules of L2(B) that have finite dimension over D and are
Γ-invariant, are contained in L2(B).
proof of equivalence of these conditions. For a proof that conditions (1) and (2) are equiva-
lent, we refer to [P4, lemma 2.10]. Equivalence of conditions (2) and (3) follows from the
fact the L2((cid:104)B, eD(cid:105)) = L2(B) ⊗D L2(B). Condition (2) and (4) are equivalent because the
D-submodules of L2(B) correspond 1-to-1 with the projections in (cid:104)B, eD(cid:105). The dimension of
the submodule is precisely the trace of the corresponding projection, and the submodule is
(cid:3)
Γ-invariant if and only if its corresponding projection is.
In the case where B is abelian, we recover the classical definition of relative weak mixing:
Observation 1.2. Let Γ (cid:121)(X, µ) be a p.m.p. action and suppose that p : (X, µ) → (Y, ν) is
a factor of this action. Then L∞(Y, ν) ⊂ L∞(X, µ) is a Γ-invariant von Neumann subalgebra.
Then the following are equivalent:
(1) Γ (cid:121) L∞(X, µ) is weakly mixing relative to L∞(Y, ν), in the sense defined above.
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
(2) Γ (cid:121)(X, µ) is weakly mixing relative to p : X → Y , in the classical sense. I.e.
diagonal action of Γ on X ×Y X is ergodic.
5
the
1.2. Generalized co-induced actions. We introduced generalized co-induced actions in
[D]. Here we repeat the construction and generalize some properties.
Definition 1.3. Let Λ (cid:121) I be an action of a countable group on a countable set. Let ω :
Λ × I → Λ0 be a cocycle. Suppose that Λ0 acts probability measure preservingly on (Y0, ν0).
Define an action of Λ on (Y, ν) = (Y0, ν0)I by the formula (gy)i = ω(g, g−1i)yg−1i. This
action is called the generalized co-induced action of Λ0 (cid:121)(Y0, ν0), with respect to ω.
Lemma 1.4. Let Λ (cid:121) I be an action of a countable group on a countable set, and let
ω : Λ × I → Λ0 be a cocycle. Suppose that Λ0 (cid:121)(Y0, ν0) is a probability measure preserving
action. Consider the generalized co-induced action Λ (cid:121)(Y, ν) = (Y0, ν0)I .
• If all orbits of Λ (cid:121) I are infinite, then the generalized co-induced action Λ (cid:121)(Y, ν) is
• Suppose that Λ (cid:121) I and ω satisfy the following three conditions.
weakly mixing.
Stab{i} × {i} surjectively onto Λ0
Si = {g ∈ Λ gi = i and ω(g, i) = e} acts with infinite orbits on I \ {i}.
-- Λ acts transitively on I.
-- There exists an i ∈ I such that (or equivalently, for all i ∈ I) ω maps the set
-- There exists an i ∈ I such that (or equivalently, for all i ∈ I) the subgroup
Then every measurable Λ-invariant map f : Y → Y is of the form f (x)i = f0(xi),
where f0 : Y0 → Y0 is a measurable Λ0-invariant map.
Proof. The first point follows in the same way as for generalized Bernoulli actions. See for
example [PV3, proposition 2.3].
For the second point, it is clear that every map of the given form is measurable and Λ-
invariant. On the other hand, let f : Y → Y be a Λ-invariant map. Fix i ∈ I and consider
the composition fi : Y → Y0 of f with the quotient onto the i-th component of Y . Observe
that fi is Si-invariant, because Si does not act on the i-th component. But by the first point,
Si acts ergodically on Y I -- {i}
. So we see that f (x)i = fi(x) = f0(xi) for some measurable
map f0 : Y0 → Y0, but only for the one i ∈ I we fixed.
Moreover, observe that f0(ω(g, i)x0) = ω(g, i)f0(x0) for all g ∈ Stab{i}. Since ω maps
Stab i × {i} onto Λ0, we find that f0 is Λ0-equivariant. Let j ∈ I be another element. We
find g ∈ Λ such that gj = i. Then we compute that
0
f (x)j = ω(g−1, j)(gf (x))i = ω(g, i)−1f0((gx)i) = f0(xj).
(cid:3)
2. Generalizing two results from [IPV]
Our main theorem depends on the results from sections 5 and 6 from [IPV]. But in fact, we
need a slightly more general statement of these two results. The proof is mainly an application
6
STEVEN DEPREZ
of the direct integral decomposition of a von Neumann algebra. Nevertheless, we think it is
worthwhile to give a careful statement and a short proof.
0 ) and consider the von Neumann algebra M = (A (cid:111) Γ) ⊗ B.
Theorem 2.1 (a version of [IPV, section 5]). Let Γ act on a countable set I in such a way
that there is a number κ ∈ N such that StabF is finite whenever F ≥ κ. Choose a standard
probability space (X0, µ0). Suppose that (B, τ ) is a finite type I von Neumann algebra. Write
A = L∞(X I
Let p ∈ L(Γ) ⊗ B be a projection. Let D ⊂ pM p be an abelian subalgebra. Write G for the
normalizer of D inside pM p. Denote the intersection G0 = G ∩U(p(L(Γ) ⊗ B)p). Assume
that
• D does not embed into B inside M .
• G(cid:48)(cid:48) does not embed into (A (cid:111) Stab{i}) ⊗ B inside M , for any i ∈ I.
• G(cid:48)(cid:48) does not embed into L(Γ) ⊗ B inside M .
• G(cid:48)(cid:48)
0 does not embed into (L Stab{i}) ⊗ B inside L(Γ) ⊗ B for any i ∈ I.
Then we get that C = D(cid:48) ∩ pM p≺f
M A ⊗ B.
Proof. Set (cid:101)D = Z(C), and observe that (cid:101)D is still an abelian subalgebra of pM p that is still
normalized by G, and D ⊂ (cid:101)D. Hence (cid:101)D still satisfies the four conditions above. But we also
get that pZ(B) ⊂ (cid:101)D.
direct integral decomposition B =(cid:82) ⊕
We decompose (cid:101)D and C into (cid:101)D = (cid:82) ⊕
We write Z(B) = L∞(Y, ν) for some standard measure space (Y, ν). Then we can take the
Y pydν(y),
where each py ∈ L(Γ) ⊗ By.
Y Bydν(y). Likewise we can decompose p =(cid:82) ⊕
Y Dydν(y) and C = (cid:82) ⊕
Y Cydν(y) respectively. Denote
My = (A(cid:111)Γ)⊗By, and observe that Cy is the relative commutant of Dy inside pyMypy. Each
unitary u in G decomposes intro a direct integral of unitaries uy, each of which normalizes
Dy.
All in all, we see that the inclusion Dy ⊂ pyMypy satisfies the conditions of [IPV, theorem 5.1].
M A⊗ B. (cid:3)
We obtain that Cy ≺f
My A⊗ By for almost all y ∈ Y . Hence it follows that C ≺f
We also want to give a similar variant to [IPV, theorem 6.1] But for our main theorem,
we need a slightly more general version: using the notations from [IPV, section 6], we can
not assume that (Adγ(s))s∈Λ acts weakly mixingly on Z(C). Instead, we can only assume
that the action is weakly mixing relative to a discrete subalgebra D ⊂ Z(C). This is not a
hard generalization, but we have to adapt the statement of the theorem slightly. In order to
simplify the statement of theorem 2.2, we incorporate [IPV, corollary 6.2.1].
Theorem 2.2 (our variant of [IPV, theorem 6.1 and corollary 6.2.1]). Let Γ (cid:121)(X, µ) be a
free, ergodic and p.m.p. action. Let (B, τ ) be a finite type I von neumann algebra. Write
A = L∞(X) and consider the von Neumann algebra M = (A (cid:111) Γ) ⊗ B. Let p ∈ L(Γ) ⊗ B be
a projection with finite trace.
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
7
Let C ⊂ pM p be a von Neumann subalgebra and suppose that γ : Λ → U(p(L(Γ) ⊗ B)p) ∩
N pM p(C) is a group morphism satisfying the following conditions.
• Λ does not have any non-trivial finite dimensional unitary representations.
• γ(Λ)(cid:48)(cid:48) does not embed into any L(Centr{g}) ⊗ B for any e (cid:54)= g ∈ Γ.
• Consider the action of Λ on Z(C) by conjugating with γ(Λ). We assume that this
• Z(C)(cid:48) ∩ pM p = C and C ≺f A ⊗ B.
action is weakly mixing relative to D = Z(C) ∩ p(L(Γ) ⊗ B)p.
Then we have the following
• a partial isometry v ∈ L(Γ) ⊗ B ⊗ B(C, (cid:96)2(N)) ⊗ B(C, (cid:96)2(N)) with left support equal to
p and with right support q = v∗v inside (cid:101)B = B ⊗ B((cid:96)2(N)) ⊗ (cid:96)∞(N), and
• a group morphism δ : Λ → G where G ⊂ U(L(Γ) ⊗ (cid:101)B) is the group
,
(cid:88)
(cid:88)
v∗Cv = q(A ⊗ (cid:101)B)q and v∗γ(s)v = δ(s) for all s ∈ Λ.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) pg ∈ Z(q(cid:101)Bq) are projections with
pg = q
G =
pgug
g∈Γ
g
such that
Proof. Write Z(B) = L∞(Y, ν) for some measure space (Y, ν). Decompose B as the direct
integral of (By)y∈Y . Observe that pZ(B) ⊂ D ⊂ Z(C). Hence we can decompose both D and
C as a direct integral over Y of von Neumann algebras Dy and Cy. Write My = (A (cid:111) Γ)⊗ By
and observe that M is the direct integral of the My. We decompose p as the direct integral
of the projections py.
Observe that Cy ≺f
My A⊗ By and that Z(Cy)(cid:48) ∩ pyMypy = Cy almost everywhere. Moreover,
we can consider the γ(s) as measurable maps y (cid:55)→ γ(s, y) from Y into the group of unitary
elements of py(L(Γ) ⊗ By)py. We remark that the γ(s, y) normalize Cy. Consider the action
of Λ on Z(Cy) that is given by conjugation with γ(s, y). This action is still weakly mixing
relative to Dy.
But Dy embeds fully into A ⊗ By, while Dy is contained in L(Γ) ⊗ By. Hence it fully embeds
into By (see lemma 2.3 below). Since By is a finite type I factor, it follows that Dy is an
abelian discrete von Neumann algebra. Since Λ does not have non-trivial finite dimensional
unitary representations, it can not act trace-preservingly on such a von Neumann algebra,
unless the action is trivial. So we find a countable number of projections py,n ∈ Dy such that
(cid:80)
n pn,y = py and Dypn,y = C pn,y. In other words, Λ acts weakly mixingly on Z(Cy)pn,y.
We can now apply [IPV, theorem 6.1 and corollary 6.2.1] to the inclusion Cypn,y ⊂ pn,yMypn,y.
We find a partial isometry vn,y ∈ L(Γ) ⊗ By ⊗ B(C, (cid:96)2(N)) with left support equal to pn,y and
n,yvn,y ∈ By ⊗ B((cid:96)2(N)), and such that
with right support qn,y = v∗
n,yCyvn,y = qn,y(A ⊗ By ⊗ B((cid:96)2(N)))qn,y.
v∗
Moreover, we find a group morphism δy,n : Λ → Γ and a finite-dimensional unitary represen-
tation πn,y : Λ → U(pn,y(By ⊗ B((cid:96)2(N)))pn,y) such that
n,yγ(s, y)vn,y = πn,y(s) ⊗ uδ(s) for all s ∈ Λ.
v∗
8
STEVEN DEPREZ
But we assumed that Λ does not have any non-trivial finite dimensional unitary representa-
tions, so we see that πn,y(s) = pn,y.
(cid:19)
(cid:88)
n
(cid:18)(cid:90) ⊕
In fact, reading the proof of [IPV, theorem 6.1] carefully, we see that we can do all this in
such a way that the vn,y depend measurably on y. Hence we can consider the partial isometry
⊗ e1,n ∈ (cid:101)N L(Γ) ⊗ B ⊗ B(C, (cid:96)2(N)) ⊗ B(C, (cid:96)2(N)).
This partial isometry has left support p and its right support is given by q =(cid:82) ⊕
(cid:80)
en,ndν(y) ∈ (cid:101)B. A direct computation shows that
n qn,y ⊗
vn,ydν(y)
Y
v =
Y
The measureable field δ(s, y, n) = δn,y(s) of group morphisms satisfies the condition
v∗Cv = q(A ⊗ (cid:101)B)q.
(cid:88)
v∗γ(s)v = q
χ{(y,n)δ(s,y,n)=g} ⊗ ug,
so we see that v∗γ(s)v ∈ G for all s ∈ Λ.
g
(cid:3)
Lemma 2.3. Let A, B be finite von Neumann algebras with trace-preserving actions of a
countable group Γ. Consider M = (A ⊗ B) (cid:111) Γ. Let D ⊂ p(B (cid:111) Γ)p be a von Neumann
subalgebra of some corner of B (cid:111) Γ. If D embeds into A⊗ B inside M , then D already embeds
into B inside B (cid:111) Γ.
Proof. Suppose that D did not embed into B inside B (cid:111) Γ. By definition, we find a sequence
(vn)n of unitaries in D such that
(cid:107)EB(xvny)(cid:107)2 → 0 for all x, y ∈ B (cid:111) Γ.
We want to show that D does not embed into A ⊗ B, inside M . So we want to show that
(cid:107)EA⊗B(xvny)(cid:107)2 → 0 for all x, y ∈ M.
By Kaplansky's density theorem, we can assume that x = (a1 ⊗ b1)ug and y = uh(a2 ⊗ b2)
for some a1, a2 ∈ A and b1, b2 ∈ B and g, h ∈ Γ. We compute that
(cid:107)EA⊗B(xvny)(cid:107)2 = (cid:107)(a1 ⊗ b1) EA⊗B(ugvnuh)(a2 ⊗ b2)(cid:107)2
≤ (cid:107)a1(cid:107)(cid:107)a2(cid:107)(cid:107)b1(cid:107)(cid:107)b2(cid:107)(cid:107)EB(ugvnuh)(cid:107)2 → 0,
as required.
(cid:3)
3. Anti-(T) groups and groups with small normalizers
It is well-known that a group Γ that has the Haagerup property does not contain an infinite
property (T) subgroup. Slighly more generally, for any p.m.p. action Γ (cid:121)(X, µ), we know that
the corresponding group-measure space von Neumann algebra L∞(X) (cid:111) Γ does not contain
a diffuse von Neumann subalgebra with property (T), see [P1].
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
9
Definition 3.1 (anti-(T) group). We say that a group Γ is anti-(T) if for every trace pre-
serving action Γ (cid:121)(A, τ ) on an amenable von Neumann algebra A and for every projection
p ∈ A (cid:111) Γ, the von Neumann algebra p(A (cid:111) Γ)p does not contain a diffuse von Neumann
subalgebra with property (T).
This definition differs from the notion of an anti-(T) group that was introduced in [HPV].
Every group that has the anti-(T) property in the [HPV]-sense is anti-(T) in our sense, but
not the other way around. The advantage of our notion is that it is stable under arbitrary
amalgamated free products.
Lemma 3.2. If Γ1, Γ2 are anti-(T) groups, then Γ = Γ1 ∗Σ Γ2 is still anti-(T).
Proof. Let Γ (cid:121)(A, τ ) be a trace-preserving action on an amenable von Neumann algebra. Let
p ∈ A (cid:111) Γ be a projection and suppose that Q ⊂ p(A (cid:111) Γ)p is a property (T) subalgebra.
By [IPP, theorem 5.1], we know that Q≺ A (cid:111) Γi for i = 1 or 2. In particular, there is a
∗-homomorphism θ : Q (cid:111) q(A (cid:111) Γi)nq for some n ∈ N and q ∈ (A (cid:111) Γi)n. Since Γi was
(cid:3)
anti-(T), it follows that Q is not diffuse.
In [OP], Ozawa and Popa show that the free groups have the following property: the normal-
izer of every diffuse abelian subalgebra B ⊂ L(Fn) is amenable. This result was generalized
later in [PV1]. There Popa and Vaes show that for every trace-preserving action Fn (cid:121)(A, τ ) on
a finite amenable von Neumann algebra and every diffuse abelian subalgebra B ⊂ p(A (cid:111) Fn)p
of a corner of the crosses product, we have the following dichotomy. Either B embeds into
A or the normalizer of B is amenable. In [PV2], they show that this property holds for all
hyperbolic groups.
In this section, we introduce a similar but weaker property of groups. The advantage of our
property is that it is implied by the Haagerup property and that it is stable under taking
amalgamated free products.
We first introduce what we mean when we say that an abelian subalgebra B ⊂ M has a large
normalizer. In the following, we denote by Dn(C) the algebra of diagonal n× n matrices with
complex entries.
Definition 3.3. Let B ⊂ M be an abelian subalgebra of a finite von Neumann algebra M .
We say that B has a large normalizer if N M (B) contains a group G that generates a diffuse
property (T) subalgebra of M .
Definition 3.4 (groups with small normalizers). We say that a group Γ has small normalizers
if, for every trace-preserving action Γ (cid:121)(A, Tr) on a finite amenable von Neumann algebra,
and every projection p ∈ M = A (cid:111) Γ, every diffuse abelian subalgebra B ⊂ pM p with large
normalizer, embeds into A inside M .
It follows immediately from [PV2, theorem 3.1 and lemma 2.4] that the following groups have
small normalizers.
• hyperbolic groups
10
STEVEN DEPREZ
• lattices in rank one simple Lie groups with finite center
• Sela's limit groups
Any anti-(T) group Γ has small normalizers, simply because no amplification of the crossed
product A (cid:111) Γ can have a von Neumann subalgebra with property (T).
Theorem 3.5. The amalgamated free product of groups with small normalizers over an amal-
gam that has the anti-(T) property still has small normalizers.
Proof. Let Γ1, Γ2 be two groups with small normalizers, and let Σ be a common subgroup
with the Haagerup property. Consider the amalgamated free product Γ = Γ1 ∗Σ Γ2. Let Γ
act trace preservingly on an amenable finite von Neumann algebra (A, τ ). Denote M = A(cid:111) Γ
and take a projection p ∈ M . Let B ⊂ pM p be an abelian subalgebra with a large normalizer.
Consider a subgroup G ⊂ N pM p(B) that generates a property (T) subalgebra in pM p. Denote
by N = B ∨ G ⊂ pM p the von Neumann subalgebra that is generated by B and G.
Now we use techniques from [IPP]. But we use the versions as stated in [PV5], because
these versions are more convenient. Consider the word-length deformation as defined in [IPP,
section 2.3] (see also [PV5, section 5.1]): we define completely positive maps mρ : M → M
by mρ(aug) = ρgaug where g denotes the word-length of g and ρ is a real number between
0 and 1. These completely positive maps converge to the identity pointwise, as ρ → 1.
Because G(cid:48)(cid:48) has property (T), we know that mρ converges to the identity uniformly in (cid:107)·(cid:107)2
on G. Observe that G(cid:48)(cid:48) does not embed into A (cid:111) Σ because Σ is anti-(T). So [PV5, lemma
5.7] yields a real number 0 < ρ0 < 1 and a δ > 0 such that
τ (w∗mρ0(w)) > δ for all unitaries w ∈ B.
8 δ2 for all v ∈ G. Because
2 δ for all v ∈ G and w ∈ U(B).
By property (T) we find ρ ≥ ρ0 such that (cid:107)v − mρ(v)(cid:107)2 < 1
τ (w∗mρ0(w)) increases with ρ0, we can assume that ρ = ρ0. Hence we can conclude that
τ (v∗w∗mρ0(wv)) > 1
Theorem [IPP, theorem 4.3] shows that N embeds into either A (cid:111) Γ1 or A (cid:111) Γ2. Without
loss of generality, we can assume that N embeds into A (cid:111) Γ1. Hence we find a non-zero
partial isometry v ∈ M1,n(C) ⊗ M and a ∗-homomorphism θ : N → q(Mn(C) ⊗ A (cid:111) Γ1)q
such that xv = vθ(x) for all x ∈ N . We can assume that q is the support projection of
EMn(C)⊗A(v∗v). The subalgebra θ(B) ⊂ q((Mn(C) ⊗ A) (cid:111) Γ1)q still has large normalizer.
Because Γ1 was assumed to have small normalizers, we see that θ(B) embeds into Mn(C)⊗ A
inside (Mn(C) ⊗ A) (cid:111) Γ1. It follows that B embeds into A.
(cid:3)
4. Proof of the main result
We want to show that all semigroups of the form Factor(Λ (cid:121) Y )op appear is End(M ) for some
type II1 factor M . We can always replace Λ by a countable dense subgroup of Λ ⊂ Aut(Y, ν)
without changing Factor(Λ (cid:121) Y ). From now on we assume that Λ is countable. Lemma 6.3
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
11
in section 6 below shows that we can also assume that Λ is anti-(T) and that all cocycles
ω : Λ × Y → K with values in a compact group are trivial.
Recall the construction of M from the introduction. Let Γ = Γ1 ∗Σ(Σ×Λ) be an amalgamated
free product group. Take an action Γ (cid:121) I of Γ on a countable set I. Choose a purely atomic
base space (X0, µ0) with unequal weights. Set (X, µ) = (X0, µ0)I and consider the generalized
Bernoulli action Γ (cid:121) X. Consider that canonical quotient morphism π : Γ → Λ and define
a new action Γ (cid:121) X × Y by the formula g(x, y) = (gx, π(g)y) for all g ∈ Γ and almost all
(x, y) ∈ X × Y .
Define M = L∞(X×Y )(cid:111)Γ. Observe that every ∆ ∈ Factor(Λ (cid:121) Y ) defines an endomorphism
ϕ∆ : M → M by the formula ϕ∆((a ⊗ b)ug) = (a ⊗ ∆∗(b))ug, where ∆ast is defined by
∆∗(b)(y) = b(∆(y)) for all b ∈ L∞(Y ) and almost all i ∈ Y . It is easy to see that two such
endomorphisms ϕ∆, ϕη are unitarily equivalent if an only if ∆ = η. The map ∆ (cid:55)→ ϕ∆ embeds
Factor(Λ (cid:121) Y )op into End(M ).
We give a set of conditions on the group Γ1 and the action Γ (cid:121) I that ensures that all
endomorphisms ϕ : M → M are of the form ϕ∆ for some ∆ ∈ Factor(Λ (cid:121) Y ).
In fact,
under the same conditions, we also find the all right-finite M -M bimodules are direct sums
of bimodules of the form H∆ = ϕ∆(M ) L2(M )M . In particular, we see that F s(M ) = N.
In the next section, we give an explicit example of a group Γ1 and an action Γ (cid:121) I that satisfy
the conditions of theorem 4.1.
Theorem 4.1. Let Λ (cid:121)(Y, ν) be an ergodic probability measure preserving action. Assume
that Λ is anti-(T) as in definition 3.1 and that all 1-cocycles ω : Λ × Y → K with values in
compact groups are trivial.
Let Γ1 ba a countable group that satisfies the conditions G1, . . . ,G5 below.
G1 Γ1 has the small normalizers property introduced above.
G2 Γ1 does not have non-trivial finite dimensional unitary representations.
G3 all centralizers CentrΓ1{g} ⊂ Γ1 of finite-order elements e (cid:54)= g ∈ Γ1 have the Haagerup
property.
G4 Γ1 contains a property (T) subgroup G.
G5 G contains a non-amenable subgroup G0 such that the commensurator of G0 and G
generate all of Γ1.
Let Σ ⊂ Γ1 be an amenable subgroup that is not virtually abelian. Consider the amalgamated
free product Γ = Γ1 ∗Σ(Σ × Λ). Let Γ (cid:121) I be an action of Γ on a countable set, satisfying the
conditions A1, . . . ,A5 below.
A1 Γ acts transitively on I
A2 All stabilizers Stab{i} have the Haagerup property and Stab{i} ∩ Γ1 is abelian.
A3 the stabilizers Stab{i, j} of two-point sets are trivial.
A4 The only injective group morphisms θ : Γ1 → Γ1 that map each StabΓ1{i} into some
A5 there is i0 ∈ I such that Stab{i0} ∩ sΓ1s−1 is infinite for all s ∈ Λ.
StabΓ1{j} up to finite index are inner.
12
STEVEN DEPREZ
Let X0 be an atomic probability space with unequal weights. Write X = X I
0 and let Γ act on X
by generalized Bernoulli action. Consider the natural quitient morphism π : Γ → Λ and define
a probability measure preserving action Γ (cid:121) X × Y by by the formula g(x, y) = (gx, π(g)y).
Denote M = L∞(X × Y ) (cid:111) Γ.
Then for every normal ∗-homomorphism ϕ : M → M from M into itself is of the form
ϕ((a ⊗ b)ug) = (a ⊗ ∆∗(b))ug for some factor map ∆ ∈ Factor(Λ (cid:121) Y ),
up to unitary conjugacy.
Moreover, if H is a right-finite dimensional M -M bimodule, them H is (isomorphic to) a finite
direct sum of M -M bimodules of the form ϕ∆(M ) L2(M )M for factor maps ∆ ∈ Factor(Λ (cid:121) Y ).
Proof. Observe that the statement about endomorphisms follows immediately from the state-
ment about right-finite bimodules. So let H be a right-finite bimodule of M . Then we
know that H is of the form H = ϕ(M )p(Cn ⊗ L2(M ))M for some ∗-homomorphism ϕ : M →
p(M ⊗ Mn(C))p from M into some finite amplification pM np of M .
We prove in 5 steps that Tr(p) = k for some natural number k and ϕ is of the form
ϕ((a ⊗ b)ug) =
ei,i ⊗ (a ⊗ (∆i)∗(b))ug,
up to unitary conjugacy, for some factor maps ∆1, . . . , ∆k ∈ Factor(Λ (cid:121) Y ).
i=1
k(cid:88)
Throughout the proof we will use the following notations for various subalgebras of M .
• A = L∞(X) and B = L∞(Y ).
• M1 = (A ⊗ B) (cid:111) Γ1 and M2 = (A ⊗ B) (cid:111) (Σ × Λ).
• M(i) = (A ⊗ B) (cid:111) Stab{i} for i ∈ I.
• N = B (cid:111) Γ.
We will also combine the notations above. That way, we write M1,(i) for M1∩ M(i). Similarly,
we write N1 = N ∩ M1 and so on.
We use the notation M n = M ⊗ Mn(C) for the amplification of M by integer numbers.
Similarly, we write Bn = B ⊗ Mn(C) and N n = N ⊗ Mn(C). We denote the action of Γ on
A by σ.
step 1: We can assume that p ∈ N n
1 and
ϕ(M1) ⊂ pM n
1 p and ϕ(N ) ⊂ pN np.
Since Γ2 has the Haagerup property while G has property (T), we see that ϕ(L(G))(cid:54)≺M M2.
By [IPP, theorem 5.1], we find a partial isometry v ∈ M ⊗ Mn,m(C) with left support vv∗ = p
and with right support q = v∗v ∈ M m
1 q. We conjugate ϕ
by v and assume already that p ∈ M n
1 , and such that v∗ϕ(L(G))v ⊂ qM m
1 and ϕ(L(G)) ⊂ pM n
1 p.
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
13
Similarly, we see that ϕ(L(G)) does not embed into M(i) for any i ∈ I. Applying [IPV,
Corollary 4.3] (which is a version of [P2, theorem 4.1]) to the rigid inclusion ϕ(L(G)) ⊂ pM n
1 p,
we find a partial isometry w ∈ M1⊗Mn,k(C) with left support p = ww∗ and with right support
r = w∗w ∈ N k
1 r. We conjugate ϕ by w and we assume that
p ∈ N n
1 , satisfying wϕ(L(G))w∗ ⊂ rN k
1 and that ϕ(L(G)) ⊂ pN n
1 p.
Observe that ϕ(L(G0)) is contained in pN n
1 p, but it does not embed into P nor into N(i),1 for
any i ∈ I. So by [IPP, theorem 1.2.1] and [V2, lemma 4.2.1] (which is based on [P2, section
3]), it follows that its quasi-normalizer is still contained in pN n
1 p. But this quasi-normalizer
contains ϕ(B ⊗ L(CommΓ1(G0))), and together with ϕ(L(G)), this algebra generates ϕ(N1).
We conclude that ϕ(N1) ⊂ pN n
1 p.
1 p with g ∈ G normalize
We know that ϕ(A) is an abelian subalgebra and all ϕ(ug) ∈ pM n
ϕ(A). Moreover, ϕ(L(G)) does not embed into P . It is shown in [IPP, theorem 1.4.1] that
then ϕ(A) itself is contained in pM n
We know that ϕ(L(Λ)) commutes with ϕ(L(Σ)) ⊂ pN np. Once we show that ϕ(L(Σ)) does
not embed into any N(i), inside N , then we can apply [V2, lemma 4.2.1] and conclude that
ϕ(L(Λ)) is contained in pN np. In that case, we find that ϕ(N ) ⊂ pN np. It remains to show
that ϕ(L(Σ)) does not embed into N(i) for any i ∈ I. Observe that N(i),1 = B ⊗ L(StabΓ1{i})
is abelian, while Σ is not virtually abelian. It follows that ϕ(L(Σ)) does not embed into N(i),1.
Hence (see [V2, remark 3.3]) we find a sequence (vm)m of unitaries in ϕ(L(Σ)) such that
1 p. Hence all of ϕ(M1) is contained in pM n
1 p.
→ 0 for all x, y ∈ N1 and for all i ∈ I.
→ 0 for all x, y ∈ N and for all i ∈ I.
(cid:13)(cid:13)(cid:13)EN(i),1(xvmy)
(cid:13)(cid:13)(cid:13)2
We want to show that(cid:13)(cid:13)(cid:13)EN(i)(xvmy)
(cid:13)(cid:13)(cid:13)2
fourier expansion of vm as vm =(cid:80)
(cid:13)(cid:13)(cid:13)EN(i)(xvmy)
By Kaplanski's density theorem, we can assume that x = ug, y = uh with g, h ∈ Γ. Write the
vk,muk. Then we compute that
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:88)
k∈Γ1
k∈Γ1
(cid:13)(cid:13)(cid:13)2
2
=
=
EN(i)(ugvk,mukh)
(cid:88)
2
(cid:107)vk,m(cid:107)2
2 .
k∈Γ1∩g−1 Stab{i}h−1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
If this last sum is non-empty, then there is a k0 ∈ Γ1 ∩ g−1 Stab{i}h−1. Then it follows that
Γ1 ∩ g−1 Stab{i}h−1 = StabΓ1{g−1i}k0. So we see that
(cid:88)
(cid:13)(cid:13)(cid:13)EN(g−1i),1
(cid:13)(cid:13)(cid:13)EN(i)(xvmy)
(cid:13)(cid:13)(cid:13)2
k∈StabΓ1{g−1i}k0
(vmu∗
k0)
(cid:13)(cid:13)(cid:13) → 0.
(cid:107)vk,m(cid:107)2
=
2
=
2
14
STEVEN DEPREZ
We have shown that ϕ(L(Λ)) does not embed into N(i) for any i. It follows that ϕ(N ) ⊂ pN np.
This finishes the proof of our first step.
step 2: We write A0 = L∞(X0) and for all subsets J ⊂ I, we denote by AJ
0 ) the
subalgebra of A that consists of functions that depend only on the components indexed by J.
Let J ⊂ I be an infinite subset that is invariant under an infinite group H ⊂ Γ1. Then we
show that ϕ(AJ
0 ) does not embed into B inside M1, for any i ∈ I.
0 = L∞(X J
Observe that Γ1 acts trivially on B, so B is contained in the center of M1, in fact, it is the
center of M1. Remark that any given element e (cid:54)= g ∈ Γ1 can fix at most one j ∈ I. So H
acts freely on AJ
0 , because J is infinite.
0 . But Bm is of finite type I and θ(AJ
Now suppose that ϕ(AJ
0 ) embeds into B inside M1. Then we find a non-zero partial isometry
v ∈ M1 ⊗ Mn,m(C) and a ∗-homomorphism θ : AJ
0 → qBmq, for some projection q ∈ Bm,
such that ϕ(x)v = vθ(x) for all x ∈ AJ
0 ) is a (non-
unital) abelian subalgebra. Up to a unitary in Bm, we can assume that q ∈ B ⊗ Dm(C)
0 ) ⊂ q(B ⊗ Dm(C))q (see for example [V1, lemma C.2] for an argument). Taking a
and θ(AJ
non-zero component of v, we can assume that m = 1 and hence that v ∈ M1 ⊗ Mn,1(C). Set
r = vv∗ ∈ ϕ(AJ
Set wg = ϕ(ug) for all g ∈ H. We claim that r is orthogonal to wgrw∗
g ∈ H. Since H acts freely on AJ
Then we compute that
g for all e (cid:54)= g ∈ H. Fix
0 such that a− σg(a) is invertible.
1 p and observe that ϕ(x)r = r(1 ⊗ θ(x)) for all x ∈ AJ
0 .
0 , we find an element a ∈ AJ
0 )(cid:48) ∩ pM n
ϕ(a)rwgrw∗
g = r(1 ⊗ θ(a))wgrw∗
g = rwg(1 ⊗ θ(a))rw∗
g = rwgrϕ(a)w∗
g = rwgrw∗
gϕ(σg(a)).
But we know that rwgrw∗
is invertible, it follows that r is orthogonal to wgrw∗
g.
g commutes with ϕ(AJ
0 ), so rwgrw∗
gϕ(a− σg(a)) = 0. Since a− σg(a)
0 ) can not embed into B inside M1.
Since H is an infinite group, we have found an infinite sequence of pairwise orthogonal pro-
jections with the same trace in the finite von Neumann algebra pM n
1 p. This contradiction
shows that ϕ(AJ
step 3: From now on, we allow n = ∞. In that case, we denote M∞(C) for B((cid:96)2(N)), and
we write Mk,∞(C) for B(Ck, (cid:96)2(C)) and finally, we write D∞(C) = (cid:96)∞(N).
With these notations we can assume that p ∈ Bn is a projection with finite trace and that
ϕ(A ⊗ B) is contained in p(A ⊗ Bn)p. Moreover, we can assume that ϕ(ug) = ug for all
g ∈ Γ1.
Denote C = ϕ(A)(cid:48) ∩ pM n
1 p, and observe that M1 = (A (cid:111) Γ1)⊗ B, because Γ1 acts trivially on
Y . We want to apply theorem 2.1 to conclude that C embeds into A⊗ B inside M1. We check
its four conditions. The first condition is satisfied by step 2 above. Observe that ϕ(L(G))
has property (T) and hence can not embed into any of the amenable algebras M(i),1 for any
i ∈ I. This shows that the second and fourth condition are also satisfied.
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
15
If the third condition were not satisfied, then we had that ϕ(A (cid:111) G)≺M1 N1. So we find a
partial isometry 0 (cid:54)= v ∈ M1 ⊗ Mn,m(C) and a ∗-homomorphism θ : A (cid:111) G → qN m
1 q, for some
q, such that ϕ(x)v = vθ(x) for all x ∈ A (cid:111) G. We can assume that q is the support projection
of EN1(v∗v). Observe that θ(A) is an abelian subalgebra of qN m
1 q with large normalizer.
Since Γ1 has small normalizers, we see that θ(A) embeds into B inside N1. So there exists
0 (cid:54)= w ∈ N1 ⊗ Mm,k(C) and a ∗-homomorphism ρ : A → rBkr such that θ(x)w = wρ(x).
Since vw (cid:54)= 0, it follows that ϕ(A) embeds into B inside M1. But that is impossible by step
2 above. Hence also the third condition of theorem 2.1 is satisfied.
M1 A ⊗ B. We want to apply theorem 2.2 to the inclusion C ⊂ pM n
We conclude that C ≺f
1 p.
Denote by γ : Γ1 → U(pN n
1 p) the group morphism that is defined by γ(g) = ϕ(ug). Observe
that Γ1 does not have any non-trivial finite-dimensional representation and that G ⊂ Γ1 is
a property (T) subgroup. We show that γ(Γ1)(cid:48)(cid:48) does not embed into L(CentrΓ1{g}) ⊗ B for
any e (cid:54)= g ∈ Γ1. If g is a finite-order element, then the centralizer has the Haagerup property,
) ⊗ B is a
by assumption. If on the other hand g has infinite order, then we know that L(g
diffuse abelian subalgebra of L(Γ1) ⊗ B. If γ(Γ1)(cid:48)(cid:48) embeds into the centralizer of g, then it
) ⊗ B has large normalizer. But Γ1 is a group with small normalizers, so
follows that L(g
) ⊗ B embeds into B, or still, that g has finite order. This contradicts
we conclude that L(g
our assumption. We conclude that γ(Γ1)(cid:48)(cid:48) does not embed into L(CentrΓ1{g}) ⊗ B for any
e (cid:54)= g ∈ Γ1.
We show that the action by conjugation on Z(C) is weakly mixing relative to D = Z(C) ∩
1 p. Let H ⊂ L2(Z(C)) be a finite dimensional, γ(Λ1)-invariant right D-submodule. Ob-
pN n
serve that then HpN n
1 p), and it has finite dimen-
sion on the right. Remark that γ(Γ1)(cid:48)(cid:48) does not embed into N(i),1 for any i ∈ I, so [V2, lemma
4.2.1] shows that any γ(Λ1)(cid:48)(cid:48)-pN n
1 p) that has finite dimension on
the right, must be contained in L2(pN n
We have just shown that C ≺f
Z(C)(cid:48) ∩ pM n
B ⊗ (cid:96)∞(N) ⊗ B((cid:96)2(N)), and such that
M1 A ⊗ B and it is obvious from the definition of C that
B(Cn, (cid:96)2(N) ⊗ (cid:96)2(N)) with left support p = vv∗ and with right support q = v∗v ∈ (cid:101)B =
1 p = C, so we can apply theorem 2.2. This yields a partial isometry v ∈ N1 ⊗
1 p). It follows that H is contained in L2(D).
1 p subbimodule of L2(pM n
1 p subbimodule of L2(pM n
1 p is a γ(Λ1)(cid:48)(cid:48)-pN n
Z
Z
Z
v∗Cv = A ⊗ (cid:101)B.
Of course we still have that v∗ϕ(N )v ⊂ q(N ⊗ B((cid:96)2(N)) ⊗ B((cid:96)2(N)))q.
Moreover, we find a group morphism δ from Γ1 to the group
(cid:88)
g∈Γ1
G =
pgug
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) pg ∈ B ⊗ (cid:96)∞(N) are projections with
(cid:88)
g
pg = q
.
This group morphism satisfies v∗ϕ(ug)v = δ(g) for all g ∈ Γ1.
Denote by Z ⊂ Y × N the support of q. Then we can view that group morphism ∆ : Γ1 → G
as a measurable field (δz)z∈Z of group morphisms δz : Γ1 → Γ1. We consider δ(g) as a map
from Z to L(Γ1), for all g ∈ Γ1. As such, this map is given by δ(g)(z) = uδz(g).
16
STEVEN DEPREZ
factor and that ϕ(A (cid:111) Γ1) is contained in (A (cid:111) Γ1) ⊗ q(cid:101)Bq. If not all δz were injective, then
Denote by r = χU q the central projection in (A (cid:111) Γ1)⊗ q(cid:101)Bq that corresponds to U . But then
We first show that almost all the δz with z ∈ Z are injective. Observe that A (cid:111) Γ1 is a
there is an element g ∈ Γ1 and a non-null set U ⊂ Z such that δz(g) = e for all z ∈ U .
the ∗-homomorphism that maps x ∈ A (cid:111) Γ1 to rϕ(x) is not injective. This contradicts the
factoriality of A (cid:111) Γ1.
Fix i ∈ I and suppose that there is a non-null set V ⊂ Z such that δz(StabΓ1{i}) ∩ Stab{j}
does not have finite index in δz(StabΓ1{i}), for any j ∈ I and for all z ∈ V . In other words,
0)r2 ⊂ r2(cid:101)Br2.
projection in (A (cid:111) Γ1)⊗ q(cid:101)Bq that corresponds to V ⊂ Z. Then we see that ϕ(Ai
δz(StabΓ1{i}) acts with infinite orbits on I, for all z ∈ V . Denote by r2 = χV q the central
0ug)r2 is contained in r2(cid:101)Br2 for all g ∈ Γ1. Hence we
) ⊂ r2(cid:101)Br2, but this contradicts step 2. So we can conclude that for almost all
But we also get that ϕ(Agi
get that ϕ(AΓ1i
z ∈ U and for every i ∈ I, a finite index subgroup of δz(StabΓ1{i}) is contained in StabΓ1{j}
for some j ∈ I.
0 )r2 = ϕ(ugAi
0
We know that all such injective group morphisms from Γ1 to itself are inner. So each δz is
inner. This is the same this as saying that δ itself is conjugate to g (cid:55)→ 1 ⊗ ug ∈ G, inside G.
So we find an element u ∈ G such that uϕ(ug)u∗ = ug for all g ∈ Γ1. We conjugate ϕ by u
and assume that ϕ(ug) = ug. Remark that u normalizes (cid:101)B, so we still have that
ϕ(A ⊗ B) ⊂ A ⊗ q(cid:101)Bq ⊂ A ⊗ q(B ⊗ B((cid:96)2(N)) ⊗ B((cid:96)2(N)))q ∼= A ⊗ qB∞q.
step 4: We can assume that there is a cocycle (bs)s∈Λ with values in U(pBnp) such that
ϕ(ug) = bπ(g)ug for all g ∈ Γ. Moreover, we can assume that ϕ(a) = a for all a ∈ A.
Denote by I0 ⊂ I the set of all i ∈ I such that StabΓ1{i} is infinite. Observe that, for each
{i}
i ∈ I0, we have that ϕ(A
0 ) commutes with L(StabΓ1). But L(StabΓ1) does not embed into
B (cid:111) Stab{i, j} for any j (cid:54)= i, simply because Stab{i, j} is trivial. Now we can apply [V2,
lemma 4.2.1] and we obtain that ϕ(A
{i}
0 ) is contained in A
is in fact the identity morphism on A
{i}
0 ⊗ pBnp.
{i}
We show that ϕ
0 , for all i ∈ I0. Take an element
g ∈ Γ1 such that gi (cid:54)= i. Observe that gi is still in I0. Remember that A0 = L∞(X0, µ0)
and that (X0, µ0) is a purely atomic probability space with unequal weights. Take a minimal
projection q in A0. Take a maximal abelian subalgebra B0 ⊂ pBnp such that q1 = ϕ(q) and
∼= L∞(Z, η) for some probability
q2 = ϕ(σg(q)) = σg(q) are both in A⊗ B0. We know that B0
space (Z, η). We consider q1 and q2 as measurable maps from Z to A. Observe that q1(z) and
q2(z) are independent for almost all z ∈ Z, i.e. τ (q1(z)q2(z)) = τ (q1(z))τ (q2(z)) = τ (q1(z))2.
(cid:18)(cid:90)
Write f (z) = τ (q1(z)) for all z ∈ (cid:101)Z. Then we know that
(cid:19)2
(cid:90)
{i}
0
A
f (z)2dz = τ (q1q2) = τ (qσg(q)) = τ (q)2 =
f (z)dz
Z
Z
The only positive functions satisfying this condition are the constant functions, so we see that
f (z) = τ (q) almost everywhere.
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
17
{i}
{x} in A
x qx, so we see that 1 =(cid:80)
0
x ∈ X0. Moreover, we know that 1 =(cid:80)
write the Fourier expansion of bs as bs =(cid:80)
{i}
0 . We find that ϕ(qx)(z)
{i}
0 with trace τ (ϕ(qx)(z)) = µ0({x}), for almost every z ∈ Z and for all
x ϕ(qx)(z) for almost all
We can do the same thing for all minimal projections qx = χ
is a projection in A
z ∈ Z. Since µ0 has unequal weights, it follows that ϕ(qx)(z) = qx a.e, for all x ∈ X0.
We prove that, for every s ∈ Λ, we get ϕ(us) = bsus for some unitary bs ∈ pBnp. Write
s ∈ p(Bn (cid:111) Γ)p. By condition A5, there is an element i0 ∈ I such that Λi0 ⊂ I0.
bs = ϕ(us)u∗
In particular, we see that Σi0 and sΣi0 are contained in I0. So bs commutes with AΣi0
. We
g∈Γ bs,gug, where bs,g ∈ pBnp. Since bs commutes
with Aσi0
. If bs,g is nonzero,
. It follows that g ∈ Stab(Σi0). Since Σ ⊂ Γ1
then we conclude that a = σg(a) for all a ∈ AΣi0
is not abelian, we know that Σ can not be contained in StabΓ1{i0}. So Σi0 contains at least
two elements and hence g ∈ Stab(Σi0) = {e}. We can conclude that bs ∈ pBnp, or still
ϕ(us) = bsus with bs ∈ U(pBnp).
Since Γ1 and Λ generate the group Γ, we see that ϕ(ug) = bπ(g)ug for all g ∈ Γ. It remains to
show that ϕ(a) = a for all a ∈ A. We know already that ϕ(a) = a for all a ∈ A
. Since Γ
acts transitively on I, the same holds for all a ∈ A.
0 , it follows that a⊗bs,g = σg(a)⊗bs,g for all g ∈ Γ and all a ∈ AΣi0
0
0
{i0}
0
step 5: Conclude that theorem 4.1 holds.
Consider p ∈ Bn as a map from Y to Mn(C), or B((cid:96)2(N)) if n = ∞. Then we see that Tr(p(y))
is Λ-invariant. So by ergodicity of the Λ-action, it is constant, and up to conjugation by a
unitary in Bn we can assume that p itself is constant, i.e. p ∈ 1⊗ Mn(C). Reducing n we can
assume that p = 1. Since p was a finite projection, it follows that n is now finite. Now we
can consider (bs)s as a cocycle for the action Λ (cid:121) Y with values in Un(C). We assumed that
all such cocycles are trivial, so up to conjugation with a unitary in Bn, we can assume that
bs = 1 for all s ∈ Λ.
Remark that ϕ(B) ⊂ (A ⊗ Bn) ∩ (Bn (cid:111) Γ) = Bn by steps 1 and 3. Since B is abelian, we
can assume that ϕ(B) ⊂ B ⊗ Dn(C). Any such ∗-homomorphism is given by a quotient map
∆ : Y ×{1, . . . , n} → Y . This quotient map is Λ-equivariant because ϕ(us) = us for all s ∈ Λ.
The image ∆(Y × {k}) is Λ-invariant and non-null, so it must be all of Y up to measure 0.
In other words, the formula ∆k(y) = ∆(y, k) defines a factor map for the action of Λ in Y .
This works for all k = 1, . . . , n, so we see that our original right-finite bimodule H is a direct
sum H = ⊕n
(cid:3)
k=1H∆, finishing the proof of the theorem.
5. An example of a group action satisfying the conditions of theorem 4.1
In order to apply theorem 4.1, we have to find an action Γ (cid:121) I that satisfies the long list
of conditions given there. Such examples are necessarily rather complicated. This section is
devoted to the description of one such example.
18
STEVEN DEPREZ
As prescribed by theorem 4.1, the group Γ is of the form Γ = Γ1 ∗Σ(Σ × Λ). We build
the group Γ1 from an arithmetic lattice in Sp(n, 1). We refer to [M] for an introduction
to arithmetic lattices in Lie groups. For this section, we only need to know that suitable
arithmetic subgroups are indeed lattices in the corresponding Lie groups.
Consider the set Hur of Hurwitz quaternions, i.e.
(cid:26)
Hur =
a + bi + cj + dk
(cid:12)(cid:12)(cid:12)(cid:12) either a, b, c, d ∈ Z or a, b, c, d ∈ Z +
1
2
(cid:27)
,
so the components are allowed to be either integers or half-integers, but mixtures are not
allowed. This is a ring under the usual addition and multiplication of quaternions, so i2 =
j2 = k2 = −1 and ij = k = −ji. We denote the element 1
2 (1 + i + j + k) by h. The skew field
of quaternions is denoted by H.
The quaternions come with a natural involution defined by a + bi + ci + dk = a − bi − ci −
dk. This involution reverses the order of the multiplication, i.e. x y = x y. Consider the
sesquilinear form B : Hur3 × Hur3 → Hur on Hur3 that is defined by B(ξ, η) = ξ0η0 −
i=1 ξiηi. Observe that this form is of signature (2, 1). Consider the group
G = PSp(B, Hur) = {A ∈ M3(Hur) B(Aξ, Aη) = B(ξ, η) for all ξ, η ∈ Hur3}/{±1}.
This group is an arithmetic lattice in the Lie group Sp(2, 1)/{±1}. As such it has property
(T), and by [PV2], G has the small normalizers property.
(cid:80)2
We remark that SL2Z embeds into G, in the following way. Observe that SL2Z is exactly the
set of matrices in SL2Z[i] that preserve the non-definite Hermitian form B0(ξ, η) = ξ2iη1 −
ξ1iη2. Consider the linear transformation A : Z[i]2 → Hur2 ⊂ Hur3 that is defined by the
matrix
A =
(cid:19)
hi
ih −ihi
(cid:18) h
(cid:18)ABA−1 0
(cid:19)
.
Observe that B(Aξ, Aη) = B0(ξ, η), and that the matrix A is invertible over Hur. So A
defines an embedding of SL2Z into G mapping a matrix B ∈ SL2Z to the block matrix
.
0
1
(cid:18)a b
(cid:19)
c d
Consider the subgroup G0 of all elements in SL2Z that are represented by matrices of the
form
where a, b, c, d are integers
and b = 0 mod 6, a = d = 1 mod 6 and c = 0 mod 7.
Define the group Γ1 = G∗G0 SL2Q, and consider its subgroup Σ = ST2Q ⊂ SL2Q of upper
triangular matrices in SL2Q. Now it follows from theorem 3.5 that Γ1 has the small normal-
izers property. The group G ⊂ Γ1 has property (T). We also see that G0 is almost normal in
SL2Q. But G and SL2Q together generate all of Γ1. It was shown in [vNW] that SL2Q does
not have any non-trivial finite-dimensional representations. Because the group G0 generates
all of G as a normal subgroup, the same is true for Γ1. Moreover, the group Σ is amenable,
but not virtually abelian.
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
19
We show that CentrΓ1{g} has the Haagerup property for every finite-order element e (cid:54)= g ∈ Γ1.
Observe that a finite-order element e (cid:54)= g in Γ1 is conjugate to an element in one of the
components G or SL2Q. We can assume without loss of generality that g ∈ G or g ∈ SL2Q.
Moreover, G0 is torsion-free, so g is not conjugate to an element in G0. Hence the centralizer
CentrΓ1{g} is still contained in the same component G respectively SL2Q. If g were in SL2Q,
then this already implies that the centralizer of g has the Haagerup property. On the other
hand, if g were in G, then the centralizer in Γ1 is just the centralizer in G. This last centralizer
is a discrete subgroup of the centralizer C of g in the Lie group Sp(2, 1)/{±1}. The centralizer
C in Sp(2, 1)/{±1} is a Lie group of strictly smaller dimension than Sp(2, 1)/{±1}, hence
its Lie algebra does not contain a copy of sp(2, 1) nor of sl2 R (cid:110) R2. Now it follows from
[CCJ+] or [dC] that C has the Haagerup property. The same is true for its discrete subgroup
CentrΓ1{g}.
Till now, we have checked all the conditions on the group Γ1 that do not depend on the action
Γ (cid:121) I. We define the action Γ (cid:121) I as follows. Choose a one-to-one map Λ (cid:51) λ (cid:55)→ nλ ∈ N.
For every Λ ∈ Λ, we set xλ = nλ + i. Consider the matrix
xλ 0 nλ
0
0
nλ 0 xλ
1
∈ Hur3×3 .
Bλ =
This matrix defines an element in G. Consider likewise the element
(cid:18) 1
(cid:19)
Cλ =
nλ
1
−1 0
∈ SL2 Q .
Finally, we define the element hλ = BλCλ ∈ Γ1 and we consider the subgroup H ⊂ Γ
generated by the λhλλ−1. Finally, we define I to be the set of left cosets of H, i.e I = Γ/H,
with the natural action of Γ by left translation.
Observe that all stabilizers of this action are conjugate to H, which in turn is isomorphic to
F∞. In particular, the stabilizers all have the Haagerup property. Moreover, the intersection
of any conjugate of H with Γ1 is either trivial or a copy of Z. So the stabilizers StabΓ1{i}
are abelian. Remark also that H = StabΓ{e} has an infinite intersection with all λΓ1λ−1, so
StabΓ1{λH} is infinite for all λ ∈ Λ.
It remains to check that all stabilizers of two-point sets are trivial, and that injective group
morphisms δ : Γ1 → Γ1 that map stabilizers into stabilizers up to finite index are inner. We
prove a lemma that implies both facts.
Lemma 5.1. Let e (cid:54)= a, b ∈ H be elements of H, let g be an element of Γ and consider an
injective group morphism δ : Γ1 → Γ1. Observe that δ defines a group morphism (we abuse
the notation and keep using the letter δ) δ : H → Γ by the formula δ(λhλλ−1) = λδ(hλ)λ−1.
If we have a = gδ(b)g−1, then it follows that δ is inner, say δ = Adh and moreover gh ∈ H.
20
STEVEN DEPREZ
Proof. We begin by studying the injective group morphisms δ : Γ1 → Γ1.
step 1: Every injective group morphism δ : Γ1 → Γ1 is bijective and moreover it is given by
(cid:19)
δ = Adh ◦(AdE ∗G0 AdF ) where h ∈ Γ1 and either
(cid:19)
0
0 u
with V = A
and E =
A−1
(cid:19)
(cid:19)
(cid:19)
(cid:19)
(cid:18)v 0
(cid:18)vk
0 v
F =
and E =
with V = A
(cid:18)1 0
(cid:18)1
0 1
0
0 −1
0
0 −vk
A−1
(cid:18)V
(cid:18)V
(cid:12)(cid:12)(cid:12) α, β, γ = 1, i, j, k
0
0 u
(cid:111)
or F =
(cid:110)±α,
where u ∈ U =
±β±γ√
2
,
±1±i±j±k
2
and v ∈ U ∩ C.
Let δ : Γ1 → Γ1 be an injective group morphism. Since δ(G) has property (T), it must be
contained in a conjugate of one of the two components G, SL2Q of Γ1. It can not be contained
in a conjugate of SL2Q because that group has the Haagerup property. So we find y0 ∈ Γ1
such that y0δ(G)y0 ⊂ G.
Consider Zariski the closure of G1 of y0δ(G)y−1
inside the Lie group Sp(2, 1)/{±1}. Then we
know that G1 is a Lie group. If G1 were not equal to Sp(2, 1)/{±1}, then we know that its
Lie algebra is also strictly smaller that sp(2, 1). So it does not contain a Lie subalgebra of the
form sp(2, 1) nor of the form sl2 R (cid:110) R2. By [CCJ+, dC], it follows that G1 has the Haagerup
property, which is absurd because it contains the discrete property (T) group y0δ(G)y−1
0 . So
we conclude that G1 = Sp(2, 1)/{±1}. Now the Margulis superrigidity theorem (or better
said, its version for Sp(n, m), see [C]) shows that Ady0 ◦δG extends to an isomorphism of
Sp(2, 1)/{±1}. All such isomorphisms are inner, so we find an element y in the product
(Sp(2, 1)/{±1}) Γ1 such that δ(x) = y−1xy for all x ∈ G.
0
On the other hand, we know that the element
(cid:18)1 1
0 1
(cid:19)
∈ SL2 Q
1 + e1,2 =
(cid:18)v
(cid:19)
has roots of all order. The only elements in Γ1 that have roots of all orders are conjugate
to 1 + re1,2 for some r ∈ Q. Hence we find an element z ∈ GL2Q Γ1 such that δ(1 + e1,2) =
z−1(1 + e1,2)z. Write
xv =
∈ SL2 Q with v ∈ R\0.
0
0 v−1
Then we know that xv(1 + e1,2)x−1
v = 1 + v2e1,2. It follows that δ(xv) = z−1xvz and then it
also follows that δ(1 + e2,1) = z−1(1 + e2,1)z. Since the 1 + re1,2 and 1 + re2,1 generate SL2Q,
it follows that δ(x) = z−1xz for all x ∈ SL2Q.
For every element x ∈ G0, we must have δ(x) = y−1xy and δ(x) = z−1xz. It follows that
zy−1 commutes with G0. The only elements in GL2Q Γ1(Sp(1, 2)/{±1}) that commute with
G0 are of the form EF −1 where E and F are of the form described earlier. It follows that δ
is indeed of the required form.
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
21
Form now on, we assume that δ is of the form AdE ∗ AdF where E and F are as in step 1.
step 2: Replacing g by an element in Hgδ(H) and adapting a, b accordingly, we can assume
λλ−1 and b = µδ(hµ)tµ−1 for some g1 ∈ Γ1, some
that g is of the form λg1µ−1, while a = λhs
λ, µ ∈ Λ and s, t ∈ Z.
Among the elements in Hgδ(H), we can assume that g has minimal length. Write g =
λ0g1λ1 . . . gnλn where λ0, λ1 ∈ Λ, λ1, . . . , λn−1 ∈ Λ -- {e} and g1, . . . , gn ∈ Γ1 \ Σ. Write b as
a reduced word b = µ1b1µ−1
m where bi ∈ h
Z
µi for all i. Then we study 5
cases.
case 1: n ≥ 2, m ≥ 2.
In that case, we know that g1 (cid:54)∈ h
Z
λ0
Σ and that gn (cid:54)∈ Σδ(hλ−1
. . . µmbmµ−1
1 µ2b2µ−1
2
gδ(b)g−1 = λ0g1λ1 . . . gnλnµ1δ(b1)µ−1
1
Z
)
. The word
. . . µmδ(bm)µ−1
m λ−1
n g−1
n
n . . . g−1
1 λ−1
0
n
Z
λ0
Z
Σδ(hλ1)
1 or λn = µ−1
m . But even then, we know that gnδ(b1)
can not be contained in Σ. In any case, we get a reduced word for gδ(b)g−1
Σ. Such an element can never be
is reduced, except that maybe λn = µ−1
and δ(bm)g−1
that starts with something of the form λ0g1 where g1 (cid:54)∈ h
contained in H.
case 2: n = 1, m ≥ 2 and g1 (cid:54)∈ h
Z
λ0
In this case, the argument is the same as for the first case: the obvious word for gδ(b)g−1 is
almost reduced, and it can never be contained in H.
case 3: n = 1 and g1 ∈ h
Z
λ0
We can assume that g1 ∈ Σ, so g is of the form λg1. For gδ(b)g−1 to be in H, we need
at least that δ(bm)g−1
. A direct computation shows that this is only possible if
for some k1 ∈ Σ and s ∈ Z. But then it follows that
g1 = e. So we have that δ(bm) = k1hs
λµ1
δ(bm−1)k−1
. As before, it follows that k1 = e. By induction, we see
that δ(bi) ∈ h
Z
, we can assume that b is of the required
λµi
form. The conjugating element g is already of the required form, and a is then automatically
of the right form.
case 4: n ≥ 2, m = 1
In this case, we can assume that g1 (cid:54)∈ h
Z
λ0
Σ. We get the following word for gδ(b)g−1:
for all i. Replacing b by any µibiµ−1
is contained in Σh
1 ∈ Σh
Z
λµm−1
Σδ(hλ1)
Z
λµm
Z
1
.
i
gδ(b)g−1λ0g1 . . . λn−1gnλnµ1δ(b1)µ−1λ−1
n g−1
n λ−1
n−1 . . . g−1
1 λ−1
0 .
If λn is not µ−1
1 , this word is already reduced. On the other hand, form the definition of H
and Σ, it is clear that no element of the form δ(hλ)s can be conjugated into Σ, inside Γ1. So
if λnµ1 were e, then we still had that
gδ(b)g−1λ0g1 . . . λn−1(gnδ(b1)g−1
n )λ−1
n−1 . . . g−1
1 λ−1
0
is a reduced word. In both cases, it is clear that gδ(b)g−1 can not be contained in H.
22
STEVEN DEPREZ
case 5: n = 1, m = 1 and g1 (cid:54)∈ h
Z
λ0
Now we see that
Z
Σδ(hλ1)
a = gδ(b)g−1 = λ0g1λ1µ1δ(b1)µ−1
1 λ−1
1 g−1
1 λ−1
0 .
If λ1µ1 (cid:54)= e, then this is a reduced expression for an element that is not in H. Otherwise, a,
b and g are of the required form. This finishes the proof of step 2.
It follows that δ = id, s = t, µ = λ and g = hr
λ for some r ∈ Z.
λλ−1 and b = µht
µµ−1, and we also have that a = gδ(b)g−1.
step 3:
We know that g = λg1µ−1, a = λhs
It follows that g1δ(ht
µ)g−1
1 = hs
λ
λ. Moreover, hs
λ , Bλ}G0{e, δ(Cµ), δ(Bµ)−1}.
λ = BλCλ . . . BλCλ is a reduced word for hs
λ ∈ Γ1 has minimal
Observe that hs
length among its conjugates, in the amalgamated free product decomposition of Γ1. The
same is true for δ(hµ). It follows that s = ±t. We can assume that g1 has minimal length
λ g1δ(hµ)r2 with r1, r2 ∈ Z. With this assumption, it also follows that
among the elements hr1
g1 ∈ {e, C−1
In any case, it follows that there are g2, g3 ∈ G0 such that g2δ(Bµ)σ2 = Bσ3
±1. Moreover, we get that g1 = bg2c where b ∈ {e, C−1
also see that σ2σ3 = −1 if and only if either b or c is e, but not both.
Using the fact that δ = AdE ∗ AdF with E, F as in step 1, we see that g2EBσ2
We write g2 and g3 as block matrices
λ g3 where σ2, σ3 =
λ , Bλ} and c ∈ {e, δ(Cλ), δ(Bλ)−1}. We
µ = Bσ3
λ g3E.
(cid:18)AXA−1 0
(cid:19)
0
1
g2 =
and g3 =
(cid:18)AY A−1 0
(cid:19)
0
,
1
where X and Y are 2 × 2 matrices with integer coefficients that are upper triangular with
eigenvalues 1, modulo 6, and that are lowe triangular with arbitrary eigenvalues modulo 7.
Denote τ σ2(xµ) for xµ if σ2 = 1 and xµ if σ2 = −1. We use also the similar notations with σ3
and xλ. Then we find that
(1)
(2)
(3)
AXV A−1
τ σ2(xµ) 0
= σ1
(cid:32)nµ
(cid:33)
(cid:33)
(cid:32)nµ
AXV A−1
0
0
0
0
AY V A−1
τ σ3(xλ) 0
(cid:32)nλ
(cid:33)
(cid:33)
(cid:32)nλ
AY V A−1
u
0
0
= σ1σ2σ3
u
= σ1σ2σ3
0
0
and uτ σ2(xµ) = σ1τ σ3(xλ)u
(4)
where V and u are as in the definition of E in step 1, and σ1 = ±1. The equation (4) implies
that xµ = xλ, so it follows that µ = λ. Moreover, comparing the real parts in (4), we see
that the sign σ1 = 1.
From equation (2), we conclude that
ξ = A−1
(cid:33)
(cid:32)1
0
=
h
−ih
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
23
is an eigenvector of XV with eigenvalue x = σ2σ3u.
We have to take a little care with what we mean by eigenvalues and eigenvectors over non-
abelian rings, but the definition is exactely as in (2): the matrix is on the left of the vector,
while the scalar is on the right. In a skew field, like H, we say that two elements x, y are
equivalent if they are conjugates of each other. Observe that in H two elements are equivalent
if and only if they have the same modulus and the same real part. For the vector spaces Hn,
we always consider scalar multiplication on the right and matrix multiplication on the left.
If Z is a square matrix over H and η is an eigenvector of Z with eigenvalue y ∈ H. Then we
see that a scalar multiple ξz is still an eigenvector of Z but with eigenvalue z−1yz.
Now we see that the vector (1, ih) is an eigenvector of XV with eigenvalue u = σ2σ3huh.
Observe that u is still in U. Write X and V as matrices
(cid:18)a b
(cid:19)
c d
X =
(cid:18)v 0
(cid:19)
0 v
and V =
or V =
0
0 −vk
,
(cid:18)vk
(cid:19)
.
2
where v ∈ U ∩ C and a, b, c, d are integers. Because (1, ih) is an eigenvector, we see that
cv + dvih = ih(av + bvih) or cvk − dvkih = ih(avk − bvkih). Observe that v, vih, ihv and
ihvih are linearly independent over R unless v = ±1. We know that X (cid:54)= 0, so we conclude
that v = ±1. Similarly, vk, vkih, ihvk and ihvkih are linearly independent over R unless
v = ± 1−i√
We check both cases separately. Suppose we are in the first case, with v = ±1.One easily
computes that c = −b and a = d + b. Since X ∈ G0, we see that b is a multiple of 6 and d
is congruent to 1 modulo 6. Remark that u = av + bvih, while b is a multiple of 6 and a is
congruent to 1 modulo 6. The only element in U that can be written in this form is v itself,
so we conclude that u = v, i.e u = σ2σ3v = ±1. But if σ2σ3 = −1, then equation (4) tells us
that xλ = xλ, which is simply not the case. So u = v and hence δ = id. Moreover, we see
that X = 1 and so g1 = 1. So in the first case, we are done.
Assume now that we are in the second case and v = ± 1−i√
. A direct computation shows that
a = d and c = −b− a. Moreover, a,−b, d are congruent to 1 modulo 6 while c is a multiple of
6, and b is a multiple of 7. Since avk − bvkih = u ∈ U, we see that this is not possible. This
contradiction shows that we must have been in the first case, and the lemma is proven. (cid:3)
2
6. Examples of endomorphism semigroups
In this section, we use the results of sections 4 and 5 in order to show that many semigroups
appear as End(M ) for some type II1 factor M . The starting point is the following.
Theorem 6.1. Let Λ (cid:121)(Y, ν) be any probability measure preserving action of a not necessarily
discrete group. Then there is a type II1 factor M such that
End(M ) ∼= Factor(Λ (cid:121)(Y, ν))op
and RFBimod(M ) ∼= {formal finite direct sums of elements of Factor(Λ (cid:121)(Y, ν))op}
24
STEVEN DEPREZ
Proof. First of all, we can assume that Λ ⊂ Autν(Y ). Take any countable dense subgroup
Λ0 of Λ. We observe that Factor(Λ (cid:121) Y ) = Factor(Λ0 (cid:121) Y ). So we can assume that Λ is a
countable group.
Then this theorem is a direct consequence of our main result 4.1, the flexible class of examples
(cid:3)
in section 5, and of lemma 6.3 below.
Before we prove lemma 6.3, we show that many left cancellative semigroups appear this way.
First, observe that any compact left cancellative semigroup with unit is in fact a group. But
then we get that G = Factor(G (cid:121)(G, h)) where h denotes the Haar measure on G. So the
compact case is easy to handle, but nor very interesting.
A more interesting class of left cancellative semigroups is the class of discrete left cancellative
semigroups. This class contains proper semigroups, like N, and even semigroups that can not
be embedded into groups (for example they are not right cancellative). We show that any
discrete left cancellative semigroup appears as the (opposite of the) semigroup of factors of
some probability measure preserving action.
Let G be a left cancellative semigroup. Let (Y0, ν0) be a standard nonatomic probability
space and consider (Y, ν) = (Y0, ν0)G. Every element g ∈ G defines a measure preserving
quotient map fg : Y → Y by the formula fg(x)h = xgh. Then we see that fgfh = fhg. So we
see that Gop ⊂ Factor(Y, ν). If we take a non-atomic base space (Y0, ν0), and we choose an
appropriate subgroup Λ ⊂ Autν(Y ), then we get that Gop = Factor(Y, ν):
Lemma 6.2. Let G be a left cancellative semigroup with unit e, and let (Y0, ν0) be a non-
atomic probability space. Set (Y, ν) = (Y0, ν0)G as before. Then there is a subgroup Λ ⊂
Autν(Y ) such that
Factor(Y, ν) = Gop.
Proof. Denote by G ⊂ Autν(Y ) the closed subgroup of Autν(Y ) that is generated by the
following transformations:
ψ∆ : Y → Y ψ∆(x)h = ∆(xh)
∆(xh)
xh
ϕU,g,∆ : Y → Y ψ∆(x)h =
(cid:40)
for all ∆ ∈ Autν0(Y0)
for all ∆ ∈ Autν0(Y0) with ∆(U ) = U
and for all g ∈ G, U ⊂ Y0
if xhg ∈ U
if xhg (cid:54)∈ U
These automorphisms commute with all the fg with g ∈ G.
In other words, we see that
Gop ⊂ Factor(G (cid:121) Y ). We show that this is actually an equality. Let f : Y → Y be a
measure preserving quotient map that commutes with the action of G.
step 1: Because f commutes with all the ψ∆, there is an injective map α : G → G such
that f (x)g = xα(g) for almost all x ∈ Y , and for all g ∈ G.
Fix a set U ⊂ Y0 with 0 < ν0(U ) < 1, and consider the group H = {∆ ∈ Autν0(Y0) ∆(U ) =
U}. Then we know that H acts weakly mixingly on both U and Y0 -- U . In particular, for
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
25
.
0
0
0
0
i0
y ∈ Y I
0
denote
every finite set I, we know that the H-invariant subsets of (Y0, ν0)I are precisely the disjoint
unions of sets of the form U I1 × (Y0 -- U )I -- I1 for subsets I1 ⊂ I.
i0-component. Write (cid:101)U = f−1
Fix i0 ∈ G and denote by fi0 : Y → Y0 the composition of f with the quotient map onto the
(U ). For a fixed finite subset I ⊂ G and an element y ∈ Y I
(cid:12)(cid:12) f (y, x)i0 ∈ U(cid:9)
(cid:101)UI,y =(cid:8) x ∈ Y G -- I
0 ,
(cid:110)
(cid:111)
(cid:12)(cid:12) νG -- I
(cid:101)UI =
((cid:101)UI,y) > 0
Now it is clear that (cid:101)U ⊂ (cid:101)UI × Y G -- I
U I1 × (Y0 -- U )I -- I1 for subsets I1 ⊂ I.
We claim that there is an i ∈ I such that U{i} × Y I\{i}
and observe that the function p (cid:55)→ (1 − p)n−1 + np(1 − p)n−2 tends to 1 as p → 0, but the
derivative in 0 is positive. So, taking p small enough, we can assume that (1 − p)p + np(1 −
(V ) and
. Moreover, (cid:101)UI is a disjoint union of sets of the form
⊂ (cid:101)UI . To prove this claim, put n = I
p)n−1 > (1 − p). Take a subset V ⊂ U with measure ν0(V ) = p. Write (cid:101)V = f−1
define (cid:101)VI in the same way as we defined (cid:101)UI . Then it is clear that (cid:101)VI ⊂ (cid:101)UI , but also that
0 ((cid:101)VI ) ≥ p. Observe that (cid:101)VI is a disjoint union of sets of the form V I1 × (Y0 -- V )I -- I1 for
It follows that (cid:101)VI must contain at least one of the sets (Y0 -- V )I or
0 = (cid:101)UI , while in the
-- (cid:101)U ) < ε.
-- (cid:101)U ) < ε.
× U{i} ⊂ (cid:101)UI0. For every finite set I0 ⊂ I ⊂ G,
× U{i} ⊂ (cid:101)UI , with the same i ∈ I0 ⊂ I. As a consequence we find
× U{i} ⊂ (cid:101)U . Comparing the measures, we see that this inclusion is actually an
Observe that, for any ε > 0, there is a finite set I ⊂ G such that ν((cid:101)UI × Y G -- I
Taking ε = ν0(U )(1− ν0(U )), we find a finite subset I0 ⊂ G such that ν((cid:101)UI × Y G -- I
Then there is a unique i ∈ I0 with Y I0 -- {i}
we see that Y I -- {i}
that Y G -- {i}
equality. In other words, f−1
Every subset V ⊂ U is of the form V = U ∩ ∆(U ) for some automorphism ∆ : Y0 → Y0.
Hence we compute that
νI
some subset I1 ⊂ I.
(Y0 -- V )I -- {i} × V {i} for some i ∈ I. In the first case, it follows that Y I
second case we find that Y I -- {i}
× U{i} ⊂ (cid:101)UI . In both cases, we have proven our claim.
(U ) = U{i}.
0
0
i0
i0
0
0
0
0
f−1
i0
(V ) = f−1(U ) ∩ ψ∆(f−1
(U ))
i0
= U{i} ∩ ∆(U ){i}
= V {i}
i0
(V ) equals V {i}. In other words, f (x)i0 = xi almost everywhere.
The same proof works for subsets of Y0 -- U . Hence we get that for every V ⊂ Y0, the inverse
image f−1
We can do this for every i0 ∈ G and find a map α : G → G such that f (x)g = xα(g). The map
α is injective because otherwise f would not be measure preserving. This finishes the proof
of step 1.
step 2: Because f also commutes with the ϕU,g,∆, it follows that f = fk for some k ∈ G.
26
STEVEN DEPREZ
In fact, we only need that f commutes with ϕU,g,∆ for one non-trivial set U ⊂ Y0 and one
non-trivial automorphism ∆ : Y0 → Y0, but for all g ∈ G. Remember that f is of the form
f (x)h = xα(h) almost everywhere and for all h ∈ G. The fact that f commutes with ϕU,g,∆
implies that α(hg) = α(h)g. Set k = α(e) where e is the unit of g. Then it follows that
α(g) = α(e)g = kg. In other words, we see that f = fk. This concludes the proof of step 2,
(cid:3)
and hence the proof of lemma 6.2.
Finally we prove the technical lemma we needed in the proof of theorem 6.1.
Lemma 6.3. Let Λ (cid:121)(Y, ν) be an action of a countable group on a probability space (Y, ν),
then there is a probability measure preserving action (cid:101)Λ (cid:121)((cid:101)Y , ν) of an anti-(T) group (cid:101)Λ (see
definition 3.1) such that
Factor((cid:101)Λ (cid:121)((cid:101)Y , ν)) = Factor(Λ (cid:121)(Y, ν))
and such that this new action does not have any non-trivial cocycles to compact groups.
Moreover, the new action is ergodic.
Proof. We use the generalized co-induced actions that were introduced in [D]. For the con-
venience of the reader, we repeated the construction and basic properties in preliminary 1.2.
Without loss of generality, we can assume that Λ = F∞. Denote by an the n-th canonical
generator. Consider now the following group:
∗ Z2
(cid:107)
F1
∗ Z2
(cid:107)
F2
∗ . . .)))
,
(cid:125)
G
where Σ =
0 1
H
(cid:107)
G0
× (Z2
(cid:107)
B
(cid:123)(cid:122)
(cid:110) (Z2
(cid:107)
A
∗Σ (SL2 Z
(cid:101)Λ = SL2 Q (cid:110) Q2
(cid:125)
(cid:123)(cid:122)
(cid:124)
(cid:124)
(cid:27)
(cid:19)
(cid:26)(cid:18)1 n
(cid:19)n n ∈ Z
(cid:26)(cid:18)1 −1
(cid:27)
(cid:101)Λ1 = H ∗Σ(G0 (cid:110) (A × B)),
n ∈ Z
⊂ H.
⊂ G
∼=
1
0
Observe that (cid:101)Λ is an anti-(T) group because it is an amalgamated free product of poly-
Haagerup groups. There is an obvious quotient from (cid:101)Λ onto the group
where the Fn are mapped to the identity element in (cid:101)Λ1. Consider the group (cid:101)Λ0 = ST2Q of
upper triangular matrices in SL2Q, and let (cid:101)Λ act on I =(cid:101)Λ1/(cid:101)Λ0 by left translation. Define a
cocyle ω :(cid:101)Λ × I → F∞ by the following relations.
ω(s, i) = e
ω(s, i) = e
ω(s, i) = adet(sb)
if s ∈ Fn and a reduced word for i starts with a letter from H
if s ∈ Fn and
(a, b)g ∈ G0 (cid:110) (A × B) is the first letter of a reduced word for i.
Above, we denoted det(sb) for the determinant of the matrix whose columns are s and b.
n
if s ∈(cid:101)Λ0
27
TYPE II1 FACTORS WITH ARBITRARY COUNTABLE ENDOMORPHISM GROUP
cocycle clearly satisfies the conditions of lemma 1.4. So we see already that
Factor((cid:101)Λ (cid:121)((cid:101)Y , ν)) = Factor(Λ (cid:121)(Y, ν)),
Consider the generalized co-induced action(cid:101)Λ (cid:121)(cid:101)Y of Λ (cid:121) Y , associated to the cocycle ω. This
and moreover that (cid:101)Λ acts ergodically on (cid:101)Y .
It remains to show that every cocycle α :(cid:101)Λ×(cid:101)Y → G, to a compact group G, is in fact trivial.
When restricted to(cid:101)Λ0, the action is just a generalized Bernoulli action. We can apply Popa's
measurable map ϕ : (cid:101)Y → G such that ϕ(gx)α(g, x)ϕ(x)−1 is independent of the x-variable
for every g ∈ Z2. Since Z2 is almost normal in H and acts weakly mixingly on (cid:101)Y , the same
Analogously, using the rigid inclusion A ⊂ G, we find a measurable function ϕ2 : (cid:101)Y → G such
cocycle superrigidity theorem [P4] to the relatively rigid inclusion Z2 ⊂ Q2 ⊂ H. We find a
that ϕ2(gx)α(g, x)ϕ2(x)−1 is independent of the x-variable for all g ∈ G. For g ∈ Σ, both
is in fact true for all g ∈ H.
θ(g) = ϕ(gx)α(g, x)ϕ(x)−1 and θ2(g) = ϕ2(gx)α(g, x)ϕ2(x)−1
are independent of x. So we find that
(ϕϕ−1
for all g ∈ Σ. Since Σ acts weakly mixingly on (cid:101)Y , it follows that ϕ = ϕ2 up to a constant.
θ :(cid:101)Λ → G.
So we can assume that ϕ actually equals ϕ2, and α is cohomologuous to a group morphism
2 )(gx) = θ(g)(ϕϕ−1
2 )(x)θ2(g)−1
Since the only group morphism H → G is the trivial morphism, we see that at least H ⊂ ker θ.
In particular, Σ ⊂ ker θ. The smallest normal subgroup of G that contains Σ is G itself, so it
follows that G ⊂ ker θ. But then θ is the trivial group morphism.
(cid:3)
[CCJ+] P.-A. Cherix, M. Cowling, P. Jolissaint, P. Julg, and A. Valette. Groups with the Haagerup property,
References
[C]
[D]
[dC]
volume 197 of Progress in Mathematics. Birkhauser Verlag, Basel, 2001. Gromov's a-T-menability.
K. Corlette. Archimedean superrigidity and hyperbolic geometry. Ann. of Math. (2), 135(1):165 -- 182,
1992.
Y. de Cornulier. Haagerup property for subgroups of SL2 and residually free groups. Bull. Belg. Math.
Soc. Simon Stevin, 13(2):341 -- 343, 2006.
S. Deprez. Explicit examples of equivalence relations and II1 factors with prescribed uncountable
fundamental group. preprint. ArXiV:1010:3612.
[F]
H. Furstenberg. Ergodic behavior of diagonal measures and a theorem of Szemer´edi on arithmetic
progressions. J. Analyse Math., 31:204 -- 256, 1977.
[HPV] C. Houdayer, S. Popa, and S. Vaes. A class of groups for which every action is W∗ superrigid. To
appear in: Groups, Geometry, and Dynamics. ArXiV:1010.5077.
A. Ioana. W ∗-superrigidity for Bernoulli actions of property (T) groups. J. Amer. Math. Soc.,
24(4):1175 -- 1226, 2011.
[I]
[IPP] A. Ioana, J. Peterson, and S. Popa. Amalgamated free products of w-rigid factors and calculation of
their symmetry groups. Acta Math., 200(1):85 -- 153, 2008.
[IPV] A. Ioana, S. Popa, and S. Vaes. A class of superrigid group von Neumann algebras. To appear in:
[M]
Annals of mathematics. ArXiV:1007.1412.
D. W. Morris. Introduction to Arithmetic groups. work in progress. ArXiV:math/0106063.
28
[OP]
[P1]
[P2]
[P3]
[P4]
[PV1]
[PV2]
[PV3]
[PV4]
[PV5]
[PV6]
[V1]
[V2]
STEVEN DEPREZ
N. Ozawa and S. Popa. On a class of II1 factors with at most one Cartan subalgebra. Ann. of Math.
(2), 172:713 -- 749, 2010.
S. Popa. On a class of II1 factors with Betti numbers invariants. Ann. of Math., 163:809 -- 899, 2006.
S. Popa. Strong rigidity of II1 factors arising from malleable actions of w-rigid groups. I. Invent. Math.,
165:369 -- 408, 2006.
S. Popa. Strong rigidity of II1 factors arising from malleable actions of w-rigid groups. II. Invent.
Math., 165:409 -- 452, 2006.
S. Popa. Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups. Invent.
Math., 170:243 -- 295, 2007.
S. Popa and S. Vaes. Unique Cartan decomposition for II1 factors arising from arbitrary actions of
free groups. preprint. ArXiV:1111.6951.
S. Popa and S. Vaes. Unique Cartan decomposition for II1 factors arising from arbitrary actions of
hyperbolic groups. To appear in: Journal fur die reine und angewandte Mathematik (Crelle's Journal).
ArXiV:1201.2824.
S. Popa and S. Vaes. Strong rigidity of generalized Bernoulli actions and computations of their sym-
metry groups. Adv. Math., 217:833 -- 872, 2008.
S. Popa and S. Vaes. Actions of F∞ whose II1 factors and orbit equivalence relations have prescribed
fundamental group. J. Amer. Math. Soc., 23:383 -- 403, 2010.
S. Popa and S. Vaes. Group measure space decomposition of II1 factors and W∗-superrigidity. Invent.
Math., 182:371 -- 417, 2010.
S. Popa and S. Vaes. On the fundamental group of II1 factors and equivalence relations arising from
group actions. In Quanta of Maths, Proc. of the Conference in honor of A. Connes' 60th birthday,
Clay math institute proceedings, volume 11, pages 519 -- 541, 2011.
S. Vaes. Rigidity results for Bernoulli actions and their von Neumann algebras (after Sorin Popa).
S´eminaire Bourbaki, exp. no. 961. Ast´erisque, 311:237 -- 294, 2007.
S. Vaes. Explicit computations of all finite index bimodules for a family of II1 factors. Annales Scien-
tifiques de l'Ecole Normale Sup´erieure, 41:743 -- 788, 2008.
[vNW] J. von Neumann and E. Wigner. Minimally almost periodic groups. Ann. Math., 41:746 -- 750, 1940.
[Z1]
R. J. Zimmer. Ergodic actions with generalized discrete spectrum. Illinois J. Math., 20(4):555 -- 588,
1976.
R. J. Zimmer. Extensions of ergodic group actions. Illinois J. Math., 20(3):373 -- 409, 1976.
[Z2]
|
1512.08026 | 4 | 1512 | 2016-10-30T20:34:09 | An inverse semigroup approach to the C*-algebras and crossed products of cancellative semigroups | [
"math.OA"
] | We give a new definition of the semigroup C*-algebra of a left cancellative semigroup, which resolves problems of the construction by X. Li. Namely, the new construction is functorial, and the independence of ideals in the semigroup does not influence the independence of the generators. It has a group C*-algebra as a natural quotient. The C*-algebra of the old construction is a quotient of the new one. All this applies both to the full and reduced C*-algebras. The construction is based on the universal inverse semigroup generated by a left cancellative semigroup. We apply this approach to connect amenability of a semigroup to nuclearity of its C*-algebra. Large classes of actions of these semigroups are in one-to-one correspondence, and the crossed products are isomorphic. A crossed product of a left Ore semigroup is isomorphic to the partial crossed product of the generated group. | math.OA | math |
AN INVERSE SEMIGROUP APPROACH TO THE
C*-ALGEBRAS AND CROSSED PRODUCTS OF
CANCELLATIVE SEMIGROUPS
MARAT AUKHADIEV
Abstract. We give a new definition of the semigroup C*-algebra of a
left cancellative semigroup, which resolves problems of the construction
by X. Li. Namely, the new construction is functorial, and the indepen-
dence of ideals in the semigroup does not influence the independence of
the generators. It has a group C*-algebra as a natural quotient. The
C*-algebra of the old construction is a quotient of the new one. All this
applies both to the full and reduced C*-algebras. The construction is
based on the universal inverse semigroup generated by a left cancellative
semigroup. We apply this approach to connect amenability of a semi-
group to nuclearity of its C*-algebra. Large classes of actions of these
semigroups are in one-to-one correspondence, and the crossed products
are isomorphic. A crossed product of a left Ore semigroup is isomorphic
to the partial crossed product of the generated group.
1. Introduction
Working with semigroup C*-algebras of cancellative semigroups we face
several significant problems. The full semigroup C*-algebra C ∗(S) of a left
cancellative semigroup S was introduced by X. Li in [14], and it was im-
mediately noticed by the author in Section 2.5 that this construction is
not functorial, i.e. not every semigroup morphism of two cancellative semi-
groups extends to a *-homomorphism of their C*-algebras. It fails already
for a morphism with domain the free monoid, as we show in Example 5.14.
The reason for such behavior is that the generators of C ∗(S) imitate the left
regular representation of S, while for functoriality one needs to consider a
larger class of isometric representations.
Another problem concerns the reduced semigroup C*-algebra. Many re-
sults in [14], [15] and in other papers on the subject assume independence
of the constructible right ideals in S. But this fails even in the simplest
example of an abelian semigroup Z+ \ {1} with usual addition operation,
see Paragraph 5.15.
We solve these problems by constructing the universal inverse semigroup
of S, and associating to S the full and reduced C*-algebras of this inverse
1991 Mathematics Subject Classification. 46L05; 20Mxx; 47L65.
Key words and phrases. semigroup; amenability; C*-algebra; crossed product; partial
action; nuclear C*-algebra.
1
2
MARAT AUKHADIEV
semigroup. These algebras have quotients isomorphic to C ∗(S) and C ∗
r (S)
correspondingly, and a certain quotient is a group C*-algebra for some group
associated with S.
An inverse semigroup naturally arises in the left regular representation
of a left cancellative semigroup S, and relations between the C*-algebras of
these two semigroups were studied by several authors. The latest results in
this direction were obtained by M. Norling in [21], where C ∗
r (S) is described
as a quotient of the C*-algebra of the left inverse hull Il(S) of S. The author
also gives a surjective homomorphism between the full C*-algebras C ∗(S) →
C ∗(Il(S)). One can see that the full semigroup C*-algebra C ∗(S) is by
definition a C*-algebra of an inverse semigroup W generated by the elements
of S as isometries, where idempotents correspond to the constructible right
ideals in S. These ideals are domains and images of operators in Il(S).
The semigroup Il(S) represents only the canonical action of S on itself,
and cannot capture all possible actions of S, neither can W . As we show, a
more efficient way is first to embed S in a universal inverse semigroup, and
then obtain through it the C*-algebra of S, and actions and crossed products
by S. For this purpose we use the notion of a free inverse semigroup and
the results on the problem of embedding cancellative semigroups in inverse
semigroups.
Any injective action of a cancellative semigroup generates an inverse semi-
group of partial bijections. This leads to a construction of the universal
inverse semigroup S∗ generated by S, as we explain in Section 2. We give
a description of S∗ and its relation to the left inverse hull. The existence
follows from the work [28] of B. Shain.
In Section 3 we answer the question when the universal inverse semigroup
S∗ is E-unitary. This question is important, because so far the class of E-
unitary inverse semigroups is the most well-studied. The answer is that S∗
is E-unitary precisely when S is embeddable in a group.
In this case we
give a concrete model for S∗, describing it as a semigroup generated by an
action on the group G generated by S. This action is the one which gives the
isomorphism C ∗(S∗) ∼= C ∗(E) ⋊ G due to a result by Milan and Steinberg
in [16].
Section 4 contains a study of the reduced semigroup C*-algebra of S∗.
We prove that the left regular representation of S is a subrepresentation of
the left regular representation V of S∗. It follows that C ∗
r (S) is a quotient of
C ∗
r (S∗). When S embeds in a group, the model for S∗ described in Section
3 allows to compute V . We prove that V is a direct sum of representations,
each of them can be realized as some restriction of the left regular represen-
tation of G onto a subspace of ℓ2(G), where G is a group generated by S.
And the left regular representation of S is one of these summands.
We compare the full semigroup C*-algebras of S and S∗ in Section 5. By
virtue of general theory of inverse semigroups, a natural quotient of C ∗(S∗)
is a group C*-algebra C ∗(G), where G is the maximal group homomorphic
image of S∗, and the same holds for their reduced versions.
In the case
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
3
∗
when S embeds in a group, G is isomorphic to the group generated by S.
We point out that the assignment S 7→ C ∗(S∗) is functorial, unlike the
assignment S 7→ C ∗(S). Example 5.14 of a free monoid F+
n illustrates this
difference. Unlike C ∗(F+
n ), which almost never admits a homomorphism
onto C ∗(S) for an n-generated left cancellative semigroup S, the quotient
map C ∗(F+
) is the
n
Cuntz algebra On.
) 7→ C ∗(S∗) always exists. A natural quotient of C ∗(F+
n
In Paragraph 5.15 we note that the idempotent generators of S∗ under
the left regular representation are linearly independent, due to the general
theory of inverse semigroups. Therefore, an important question of indepen-
r (S∗). We note
dence of constructible right ideals has no importance for C ∗
that C ∗(S∗) and C ∗
r (S∗) are nuclear for an abelian semigroup S.
∗
We apply our constructions to the connection between amenability of a
cancellative semigroup S and nuclearity of its reduced C*-algebra in Section
6. In particular, we prove that if S embeds in an amenable group, then C ∗
r (S)
is nuclear. Thus we generalize results of [15] on this question. Moreover,
in the above mentioned case, S∗ has the weak containment property and
C ∗
r (S∗) is nuclear. We also show that amenability of S implies amenability
of S∗ in any case.
It is a known fact that C ∗
n is not
∗) is not nuclear,
amenable. Using our constructions, we obtain that C ∗
which makes it a more natural candidate for the C*-algebra associated with
the free monoid.
n ) is nuclear despite the fact that F+
r (F+
n
r (F+
In Section 7 we give connections between actions of S and S∗ on spaces,
C*-algebras, and prove isomorphisms of crossed products. We consider ac-
tions of cancellative semigroups by endomorphisms; the case of automor-
phisms was studied in [14]. According to the definition by [30], an inverse
semigroup acts on a C*-algebra by *-isomorphisms between closed two-sided
*-ideals of the C*-algebra. Therefore, with a view to connect actions by
cancellative semigroups with actions by inverse semigroups, we are forced to
restrict ourselves to the case when the images of endomorphisms are ideals.
First we prove that injective actions of a cancellative semigroup S are in
one-to-one correspondence with unital actions of S∗, so that an action of
one induces an action of the other. Then we prove an isomorphism between
the crossed products A ⋊α S and A ⋊ α S∗, where α is induced by α or
the other way round. The case of a unital C*-algebra A is more common
for the crossed products by cancellative semigroups, hence we consider the
unital and the non-unital case separately. We use the definition of a crossed
product with non-unital C*-algebra of N. Larsen [13]. This result allows us
to describe C ∗(S∗) as a crossed product of a commutative C*-algebra by S,
and C ∗(S) as a crossed product by S∗. If S is a left Ore semigroup, then any
unital action of S∗ can be dilated to an action of a group, and the crossed
product is Morita equivalent to the group crossed product.
In Section 8 we study connections of actions and crossed products of
semigroups with partial actions and partial crossed products of groups. If
4
MARAT AUKHADIEV
S is embeddable in a group, the model for S∗ gives a partial action of the
group G (the group generated by S), such that C ∗(S∗) is isomorphic to a
partial crossed product by G. This isomorphism is precisely the one given by
Milan and Steinberg in [16]. A stronger result holds in the case of a left Ore
semigroup, that is, a semigroup S such that G = S−1S is a group. Using
the semigroup S∗ and the work of Exel and Vieira [10], we prove that any
injective action of S generates a partial action of G, and the corresponding
crossed products are isomorphic.
1.1. Let us recall the main definitions and facts used in this paper. Let P
be a semigroup. Elements x and x∗ in P are called inverse to each other if
xx∗x = x, and x∗xx∗ = x∗.
The semigroup P is called an inverse semigroup if for any x ∈ P there exists
a unique inverse element x∗ ∈ P . Further, P always stands for an inverse
semigroup. We proceed to recall basic facts on inverse semigroups.
Theorem 1.2. (V. V. Vagner [31]). For a semigroup S in which every ele-
ment has an inverse, uniqueness of inverses is equivalent to the requirement
that all idempotents in S commute.
The set of idempotents of an inverse semigroup P forms a commutative
semigroup denoted E(P ). In fact,
E(P ) = {xx∗x ∈ P } = {x∗xx ∈ P }.
Every inverse semigroup P admits a universal morphism onto a group
G(P ), which is the quotient by the congruence: s ∼ t if se = te for some
e ∈ E(P ). The group G(P ) is called the maximal group homomorphic image
of P . Note that G(P ) is always trivial if P contains a zero, i.e. an element
denoted 0, satisfying 0 · a = a · 0 = 0 for any a ∈ P . Let σ : P → G(P )
denote the quotient homomorphism onto the maximal group homomorphic
image of P . The semigroup P is called E-unitary if σ−1(1) = E(P ).
1.3. A semigroup S is called left (right) cancellative if for any a, b, c ∈ S
the equation ab = ac (ba = ca) implies b = c. A unit in a semigroup is an
element denoted by 1, satisfying a · 1 = 1 · a = a for any a in the semigroup.
1.4. Let us compare inverse and left cancellative semigroups. These two
classes of semigroups have radical differences, which follow directly from the
definitions. The notion of an inverse semigroup is a natural generalization of
the notion of a group, where a group inverse element (ss−1 = 1 and s−1s = 1)
is substituted by a "generalized inverse" (ss∗s = s and s∗ss∗ = s∗). This
is the reason why at the early stage the inverse semigroups were called
"generalized groups" ([31]).
Inverse semigroups have many idempotents
and may have a zero, while a left cancellative semigroup may have only
one idempotent, namely the unit element, and no zero element. A left
cancellative semigroup is very often embedded in a group, while an inverse
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
semigroup is a subsemigroup in a group only if it is a group itself.
particular, the intersection of these classes is the class of groups.
5
In
One meets the consequence of these differences in the theory of semigroup
C*-algebras, starting with the left (right) regular representation. An inverse
semigroup is represented on itself by partial bijections. A left cancellative
semigroup is represented on itself by injective maps, where the domain is
the whole semigroup. An inverse semigroup has an involution, which is a
map assigning to every element of S its inverse element. And the pres-
ence of an involution makes it very natural to consider *-representations in
B(H). Despite the different nature, soon after the establishment of inverse
semigroups, it was noticed that these two classes are closely related. In the
following section we construct a universal inverse semigroup generated by a
left cancellative semigroup.
2. Universal inverse semigroup
2.1. Recall the basic example of an inverse semigroup. Let X be a set, and
let Y ⊂ X. A one-to-one map α : Y → X is called a partial bijection of X.
In particular, any injective map X → X is a partial bijection of X. Suppose
that α and β are partial bijections of X with domains Y and Z respectively.
Then the product αβ is defined to be a composition of α and β with domain
β−1(β(Z) ∩ Y ). The set I(X) of partial bijections with this product forms
an inverse semigroup called the symmetric inverse semigroup of X. Note
that this semigroup contains a zero and a unit.
In what follows we always assume that every semigroup contains a unit
element, denoted by 1.
2.2. The first inverse semigroup associated with a left cancellative semigroup
S arises from the left regular representation of S. For any a ∈ S, define an
operator of left ranslation λa : S → S by λa(b) = ab for all b ∈ S. Since S
is left cancellative, each λa is injective. Then {λaa ∈ S} forms a semigroup
of partial bijections on S, and it is a subsemigroup of I(S). The inverse of
λa is a partial bijection with a domain {abb ∈ S}. Taking inverses of all
such partial bijections and products of them, one obtains a subsemigroup
of I(S). This is an inverse semigroup called the left inverse hull of S ([5]),
denoted Il(S).
2.3. More generally, suppose we have an injective action α of a left can-
It means that for every s ∈ S, the
cellative semigroup S on a space X.
map αs : X → X is injective and αs ◦ αt = αst for all t ∈ S. Denote the
image of αs by Ds ⊂ X. Then αs is a bijection between X and its image
Ds, and there exists an inverse map, which we denote by α∗
s : Ds → X.
For convenience set Ds∗ = X for every s ∈ S. One can easily verify that
Dst = αs(Dt).
Clearly, α∗
s ◦ αs is the identity on X and αs ◦ α∗
follows that αs ◦ α∗
s ◦ αs = αs and α∗
s ◦ αs ◦ α∗
s = α∗
s is the identity on Ds. It
s. But the composition
6
MARAT AUKHADIEV
t α∗
s ◦ αt is defined only on a subset of X, namely on α∗
α∗
t (Ds ∩ Dt). Thus we
put D(s∗t)∗ = α∗
t (Ds ∩ Dt) and define the product α∗
sαt = (α∗
s ◦ αt)Dt∗ s.
One should check that this definition is compatible with multiplication in S,
namely α∗
sαv. Continuing this way we define all finite products
w of the maps from the collection F = {αs, α∗
t for all t, s ∈ S} with domain
Dw. We put α∗∗
s = αs for all s ∈ S.
stαv = α∗
2a∗
It is easy to see that for a1, ..., an ∈ F , the element (a1a2...an)∗ =
a∗
n...a∗
1 is the inverse (in a semigroup sense) of w = a1a2...an, and that
ww∗ and w∗w are idempotents. Obviously, w∗w and v∗v commute for any
words v, w and any idempotent has the form w∗w. Thus, we get an inverse
semigroup, which is a subsemigroup in a set of all partial bijections on a
space X. Note that in the case that α is an action of S by injective maps,
we have α∗
sαs = id, so αs is an isometry. We have verified the following
statement.
Lemma 2.4. An injective action α of a left cancellative semigroup S on a
space X generates an inverse semigroup S∗
α ⊂ I(X).
This motivates a notion of a universal inverse semigroup generated by a
left (right) cancellative semigroup. A problem of embedding a semigroup
in an inverse semigroup is analogous to the well-known and widely studied
problem of embedding it in a group. Recall a famous result of O. Ore and
P. Dubreil on this problem, which we will use later.
Theorem 2.5. ([6]) A semigroup S can be embedded into a group G such
that G = S−1S if and only if it is left and right cancellative and for any
p, q ∈ S we have Sp ∩ Sq 6= ∅. The group G is called the group of left
quotients of S.
2.6. The question of embedding a semigroup in an inverse semigroup is more
general and is approached in the following way. For any set X there exists
a free inverse semigroup F (X), generated by X (see [27] for the proof).
Thus, if S is a semigroup, we can consider the quotient of F (S) by all
the relations in S. Namely, if xy = z in S we put xy ∼ z in F (S), and
the same for inverses. The resulting semigroup is called the free inverse
semigroup of S, and we denote it SF . So, the question becomes whether
the natural map S → SF is an embedding. In fact, there were found many
sufficient conditions for this to hold. Among these results we mention the
most important for our research.
Theorem 2.7. (B. Shain [28]). If a semigroup S is left (or right) cancella-
tive, then S can be embedded into an inverse semigroup.
2.8. Therefore, working with a left cancellative semigroup S we always have
an embedding S ֒→ SF . B. Shain also gave an explicit form of SF . Namely,
it is the semigroup generated by the set
{vp, v∗
p = v∗
p}
p : p ∈ S, vpv∗
pvp = vp, v∗
pvpv∗
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
7
with the additional requirement that all idempotents in SF commute. Since
we want S to be represented by isometries, we take the congruence on SF
generated by the equivalence relation v∗
pvp ∼ 1 for all p ∈ S. The quotient
inverse semigroup, denoted by S∗, is then generated by isometries vp for
p ∈ S. The semigroup S∗ is in some sense the largest inverse semigroup
generated by S as a semigroup of isometries.
2.9. For the moment we have mentioned two inverse semigroups associated
with a given left cancellative semigroup S, constructed in different ways: the
universal inverse semigroup S∗, and the left inverse hull Il(S). In order to
see the relation between them, we make a short review of [5] and the notion
of the left inverse hull. Note that we adapt all the relations and notations,
because, unlike the left cancellative case selected here, the semigroup S in [5]
is right cancellative. Also, a semigroup with a unit is always left and right
reductive, so the semigroup S which we consider fits into the requirements
of [5]. For the generators of F (S), we use symbols vp for p ∈ S as above.
The left inverse hull Il(S) is proved in [5] to be isomorphic to the quotient
of F (S) by four collections of relations. In the notation of [5], the above
mentioned relations v∗
pvp ∼ 1 for all p ∈ S are denoted R1, and R4 is a
collection of all relations in S. Therefore, passing to the quotient of F (S)
by the congruence generated by R1 ∪ R4 we obtain exactly S∗ (introduced
above). The relations R2 and R3 ensure that the elements are equivalent if
the corresponding operators in Il(S) have the same domain and act in the
same way (see below).
An explicit form for the domains of the maps in Il(S) is given using the
notation in [14]. For any subset A ⊂ S and any a ∈ S set
(2.1)
(2.2)
aA = {ax x ∈ A}
a−1A = {x ∈ Sxa ∈ A}
a1 λa2...λ−1
an λan−1...λ−1
an−1 λan ∈ Il(S) is the set a−1
In [5] these sets are denoted aA and a : A respectively. Then the domain of
φ = λ−1
2 (a1S)), which is the
a2 λa1 = φ∗ ∈ Il(S). Such domains are right ideals in
image of λ−1
S. In fact, these sets were called the constructible right ideals of S by X. Li
in [14] and used there for the definition of C ∗(S) and for the study of C ∗
r (S).
The set of such ideals with an empty set is denoted by J , so J is the set of
all domains (and images) of all maps in Il(S).
n (an−1...a−1
We can now formulate the relations R2 and R3 on F (S). The first of
them introduces the zero element.
v∗
a1va2 ...v∗
an−1 van ∼R2 uv∗
an−1 vanw for any u, w ∈ F (S)
a1 va2...v∗
n (an−1...a−1
iff a−1
2 (a1S)) = ∅;
The relations R3 ask for elements to be equivalent if when represented on
S they act in the same way.
v∗
a1 va2...v∗
an−1 van ∼R3 v∗
b1vb2...v∗
bk−1vbk
8
MARAT AUKHADIEV
iff a−1
n (an−1...a−1
and for any x ∈ a−1
2 (a1S)) = b−1
k (bk−1...b−1
2 (a1S))
n (an−1...a−1
2 (b1S))
there exist x1, ...xn, y1, ...yk ∈ S, such that x1 = y1 and
anx = an−1xn, an−2xn = an−3xn−1, ..., a2x2 = a1x1,
bkx = bk−1yk, bk−2yk = bk−3yk−1, ..., b2y2 = b1y1.
The main result of [5] is that Il(S) is isomorphic to the quotient F (S)/R∗,
where R∗ is the congruence generated by R = R1 ∪ R2 ∪ R3 ∪ R4. Comparing
this fact with our definition of S∗, we obtain the following.
Lemma 2.10. There are surjective homomorphisms of inverse semigroups
(2.3)
F (S)
α→ SF β
→ S∗ γ
→ Il(S),
where α is the quotient by R4, β by R1, and γ by R2 ∪ R3.
3. E-unitarity and a model for the universal inverse semigroup
As explained in Paragraph 1.1, for any inverse semigroup P there exists
a maximal group homomorphic image G(P ), which is not trivial if the semi-
group does not contain a zero. This is the case for S∗ due to its definition.
Lemma 3.1. For a left cancellative semigroup S, its universal inverse semi-
group S∗ (as well as SF ) is E-unitary if S is embeddable in a group. In this
case G(S∗) = G(SF ) is isomorphic to the group generated by S.
Proof. Firstly, SF → G(SF ) factors through β : SF → S∗ given by v∗
ava ∼
1. Therefore, by maximality SF and S∗ share the same maximal group
homomorphic image, i.e. G(S∗) = G(SF ).
Let S∗ be E-unitary and assume σ(va) = σ(vb) in G(S∗). Then vbv∗
a is a
self-adjoint idempotent, and we obtain
va = vav∗
b vb = vbv∗
avb = vbv∗
avbv∗
ava = vbv∗
ava = vb
Therefore, σ is an embedding of the semigroup {va : a ∈ S} in G(S∗), which
can be identified with S. Since va are generators of S∗, their image under σ
generates the group G(S∗) by maximality.
If SF is E-unitary and σβ(va) = σβ(vb), then vbv∗
potent in SF and therefore so is vbv∗
group in this case as well.
a is a self-adjoint idem-
a ∈ S∗. Hence, S is embeddable in a
(cid:3)
3.2. In the case when S is embeddable in a group, there exists a model
realizing the universal inverse semigroups S∗ and SF . First, let us give a
model for SF , the predecessor of S∗.
For any set X there exists a free inverse semigroup F (X). The proof
in [27] and [3], uses essentially the free group G(X) on the set X, and
embedding of X in the free group. The idea of our model is based on the
model given in [3] for F (X).
As shown in Lemma 2.10, SF is a quotient of the free inverse semigroup on
F (S) by congruence generated by relations in the semigroup S. Therefore,
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
9
the quotient of the model for F (S) gives a correct model for SF if S is
embeddable in a group. And in this case, a group generated by S is used
instead of the free group on the set. We denote by G a group generated by
S, and let exp(G) be the set of all finite subsets of G containing 1. Due to
maximality of the maximal group homomorphic image of SF , and the fact
that SF is generated by elements va, we have G(SF ) = G. Denote by σ the
quotient map SF → G.
For any A ∈ exp(G), g ∈ G, a, b ∈ S define a relation
(3.1)
A ∪ {1, g, ga, gab} ∼ A ∪ {1, g, gab}
One can easily verify that this is an equivalence relation and generates a
congruence on exp(G). It follows in particular that
{1, g, ga−1, ga−1b−1} ∼ {1, g, ga−1b−1},
{1, g, ga−1, gb} ∼ {1, ga−1, gb}.
We will formally write g1 ≤ g ≤ g2 if g = g1a, g2 = gb for some a, b ∈ S.
Then the relation ∼ on exp(G) can be formulated as
A ∪ {g} ∼ A if and only if there exist g1, g2 ∈ A such that g1 ≤ g ≤ g2.
For any A ∈ exp(G) denote its equivalence class by [A] and the set of all
equivalence classes by exp(G)/ ∼. Clearly, ∼ is stable under taking union,
i.e.
A ∼ A′, B ∼ B′ ⇒ A ∪ B ∼ A′ ∪ B′.
This equivalence is also stable under multiplication by elements of G from
the left:
gA = {1} ∪ {ga : a ∈ A},
so that the class g[A] = [gA] is well-defined. Note that this is not an
action of G on exp(G), because in general (gh)A 6= g(hA). And we have
g−1(gA) = {g−1} ∪ A for any A ∈ exp(G), g ∈ G.
Define a partial order on exp(G)/ ∼, for any A, B ∈ exp(G):
[A] ≤ [B] if and only if there exists B′ ∼ B such that A ⊂ B′
This means in particular, that for any A, B ∈ exp(G) and g ∈ G, a, b ∈ S
A ⊂ B ⇒ [A] ≤ [B],
[{1, ga−1, ga−1b−1}] ≤ [A ∪ {1, g, ga−1b−1}],
[{1, g, ga}] ≤ [A ∪ {1, g, gab}],
[{1, g, ga−1}] ≤ [A ∪ {1, g, ga−1b−1}]
[{1, g, ga−1}] ≤ [A ∪ {1, gb, ga−1}],
The partial order ≤ on exp(G)/ ∼ is stable under multiplication from the
left by elements of G due to stability of ∼.
We say that an element g = a1a2...an ∈ G is written in a reduced form if
ai's are alternating elements from S and S−1 with ai 6= a−1
i+1 for all 1 ≤ i ≤
n − 1. Define a set Ig = {1, a1, a1a2, ..., a1a2...an} ∈ exp(G) corresponding
to a fixed reduced form of g. A set A ∈ exp(G) is called full if for any g ∈ A
it contains the subset Ig for some reduced form of g. This means that any
full set equals Ig1 ∪ Ig2 ∪ ... ∪ Ign for some gi ∈ G. Denote by E the set of
10
MARAT AUKHADIEV
equivalence classes with respect to ∼ of full sets in exp(G). Note that a full
set may be equivalent to a non-full set.
Define the set
G = {(g, [A]) ∈ G × E : [Ig] ≤ [A] for some reduced form of g}.
In fact, it is sufficient to require that g ∈ A′ for some full set A′ ∼ A. We
define product and inverse operation on G:
(3.2)
(g, [A])(h, [B]) = (gh, [A ∪ gB]),
(g, [A])∗ = (g−1, g−1[A]).
(3.3)
Since [Ig] ≤ [A], we have that [Ig−1] = g−1[Ig] ≤ g−1[A]. Hence (g, [A])∗ ∈ G.
Theorem 3.3. The set G with product and inverse operation defined by
(3.2) and (3.3) forms an inverse semigroup isomorphic to SF .
Proof. To see that the product is associative, take g, h, f ∈ G and full sets
A, B, C in exp(G) and compute
A ∪ gB ∪ (gh)C = A ∪ (gB ∪ {g}) ∪ (ghC) = A ∪ g(B ∪ hC).
For any (g, [A]) ∈ G we may assume Ig ⊂ A, which implies g(g−1A) = A.
Consequently,
(g, [A])(g−1, g−1[A])(g, [A]) = (g, [A ∪ g(g−1A) ∪ A]) = (g, [A]).
Checking similarly the corresponding equation for (g−1, g−1[A]), we obtain
that (3.3) defines an inverse element for (g, [A]).
The idempotents in G correspond to elements of E:
(g, [A])(g, [A])∗ = (1, [A]) ↔ [A] ∈ E.
The product [A][B] = [A ∪ B] on E is commutative. Moreover, [A] ≤ [B] if
and only if idempotents (1, [B]) ≤ (1, [A]).
Thus, G is an inverse semigroup.
As mentioned above, the semigroup SF is a quotient of the free inverse
semigroup F (S) on S as a set by equivalence relation generated by relations
among elements in S. Therefore it is sufficient to show that the model
described is a quotient of the model for F (S) by relations in S. Let T be
the free monoid generated by the set S. For a, b ∈ S we denote their product
in T by ab and their product in S by a · b. Now T is a cancellative monoid
and clearly T F = F (S) by definition, with a maximal group homomorphic
image G(T ) equal to the free group on the set S and to G(F (S)). One can
easily see that the model for T F coincides with the model in [27] for F (S),
which we describe further.
Since G(T ) is a free group, every its element has a reduced form. For any
g ∈ G(T ) with a reduced form r(g) = a1a2...an, where ai ∈ S ∪ S−1 ⊂ G(T ),
we define
g = {1, a1, a1a2, ..., a1a2...an}
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
11
The set A ∈ exp(G(T )) is called saturated if g ∈ A implies g ⊂ A. Then
F (S) = {(a, A) : A is saturated and r(a) ∈ A} ⊂ G(T ) × exp(G(T ))
The product and inverse operation on F (S) are given by
(a, A)(b, B) = (ab, A ∪ bB),
(a, A)∗ = (a−1, a−1A).
The idempotent semilattice in F (S) consists of elements (1, A) for all satu-
rated sets A.
The generators of F (S) are
ta = (a, {1, a}), t∗
a = (a−1, {1, a−1})
for all a ∈ S, where a−1 is the inverse of a in G(T ). Note that for a, b ∈ S
we have
tatb = (ab, {1, a, ab}) and t(a·b) = (a · b, {1, a · b}),
and the same for their inverses. The homomorphism F (S) → SF is a quo-
tient map by equivalence tatb ∼ t(a·b). This consists of equivalences on G(T )
and exp(G(T )):
ab ∼ a · b, {1, a, ab} ∼ {1, a · b}.
It is easy to see that the quotient of G(T ) by the first equivalence equals
G. The second reduces subsets of G(T ) to subsets of G and induces a new
equivalence on it, which is given by (3.1). Under this equivalence a saturated
subset of G(T ) turns into a full subset of G. Thus we obtain the model G.
For the sake of completeness, we give an isomorphism between G and
SF . For any element in SF we define a corresponding element in G × E.
Obviously, any element of SF can be written as a product of alternating
symbols of the type va and v∗
b , for a, b ∈ S. We call such a form of an
element a reduced form. Let s ∈ SF be an element with a reduced form
v∗
a1va2...v∗
an , where a1, ..an ∈ S. Define
Is = {1, σ(v∗
a1 va2), ..., σ(s)} ∈ exp(G),
a1 ), σ(v∗
s = (σ(s), [Is]) ∈ G × E.
Then because of the equivalence relation we put on exp(G), the map s 7→ s
does not depend on the choice of the reduced form for s. Clearly, this map
is a *-homomorphism.
Let us define the reverse map G → SF . Consider a full set
Ig = {1, a−1
2 ...a−1
corresponding to g ∈ G and its reduced form g = a−1
2 , ..., a−1
1 b1a−1
1 , a−1
1 b1, a−1
1 b1a−1
n bn} ∈ exp(G)
1 b1a−1
n bn.
2 ...a−1
First assume that Ig contains precisely one element of S−1 (or S), de-
note it a−1
(or a1). Then assume that Ig contains precisely one element in
1
a−1
1 S (or a1S−1), etc. Under all these assumptions we get a unique chain
a−1
1 , b1, ...a−1
n , bn and define
sg = v∗
a1 vb1...v∗
an vbn,
12
MARAT AUKHADIEV
as an element corresponding to (g, [Ig]).
If on some step this is not true and we have, for instance
a−1
1 b1...a−1
k bka−1
k+1bk+1...a−1
i bi = a−1
1 b1...a−1
k bkc−1
for some c ∈ S, then we split Ig into a union of sets
Ig1 = {1, a−1
1 b1...bk, a−1
1 ...bkc−1,
1 , ..., a−1
i+1, ..., a−1
1 , ..., a−1
, ..., a−1
1 ...bkc−1a−1
1 ...bk, a−1
1 ...bkc−1b−1
i+1...bn}
1 ...bkc−1,
...b−1
k+1},
i
a−1
1 ...bkc−1a−1
and Ig2 = {1, a−1
a−1
1 ...bkc−1b−1
i
and then consider these sets separately and check them for the assumptions.
Repeating this process at the end we get Ig = ∪l
j=1Igl, where each of the
sets is full and satisfies the required property, and at least one of them ends
with g. For each of these sets Igj = {1, c−1
m dm} define sgj
as above.
1 d1, ...c−1
1 d1c−1
1 , c−1
j=1Igl]) is (sg1s∗
Now the element corresponding to (g, [∪l
)sgr , where
sgr is the element corresponding to the set containing g. The element corre-
sponding to the idempotent (1, [∪l
l ). This map is well-
defined on equivalence classes, because it depends on a unique representative
of the class. Since any full set is a union of Ig, we obtain a well-defined map
G → SF . One can easily verify that this map is also a *-homomorphism.
Clearly, the maps SF → G and G → SF are inverse to one another.
(cid:3)
j=1Igl]) is (s1s∗
g1...sgls∗
gl
1...sls∗
For the model of S∗ we need to formulate the homomorphism β : SF → S∗
in terms of the model of SF . Recall that β is given on SF by equivalence
v∗
ava ∼′ 1 for all a ∈ S and generates the following equivalence for the model
of SF .
A ∪ {1, g, ga−1} ∼′ A ∪ {1, g}
for any A ∈ exp(G), g ∈ G, a ∈ A. It follows in particular that
{1, a−1} ∼′ {1}, {1, g, ga} ∼′ {1, ga}
Similarly to ∼, this new equivalence is stable under taking the union and
multiplying by elements of G from the left as defined above. Moreover, ∼′
substitutes ∼:
{1, g, ga, gab} ∼′ {1, ga, gab} ∼′ {1, gab}
Denote by E′ the quotient of exp(G) by ∼′, and by [A] the equivalence
class of A ∈ exp(G). The product, inverse operation and partial order on
E′ are defined similarly to E. Thus Theorem 3.3 implies the following.
Corollary 3.4. The inverse semigroup S∗ is isomorphic to the inverse semi-
group
{(g, [A]) ∈ G × E′ : [Ig] ≤ [A] for some reduced form of g},
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
13
with the product and inverse operation defined above, idempotent semilattice
equal to E′, generated by elements
va = (a, [{1, a}]), v∗
a = (a−1, [{1}]).
Corollary 3.5. The inverse semigroups SF and S∗ are E-unitary if and
only if S is embeddable in a group.
Proof. Due to Lemma 3.1, it is sufficient to show the "if" part. Suppose S
generates a group G. Then looking at the model for SF given by Theorem 3.3,
any s ∈ SF equals (σ(s), [A]) for some set A ∈ exp(G). Moreover, s is an
idempotent if and only if it equals (1, [A]). Therefore, σ−1(1) = E(SF ). A
similar proof works for S∗.
(cid:3)
4. The reduced C*-algebras of the universal inverse
semigroups
4.1. As a consequence of the facts described in the previous section, there
is a connection between the representation theories of S and S∗.
Let P be an inverse semigroup. A *-representation of P is a homomor-
phism π of P into B(H) such that π(s∗) = π(s)∗ for any s ∈ P . It is clear
that each π(s) is a partial isometry. We say that π is unital if it sends the
unit in P to the identity operator. We want to avoid further subtle details
concerning the zero element. For this we ask that a *-representation of an
inverse semigroup should assign the zero operator to the zero element in P
if the latter exists.
4.2. Let S be a left cancellative semigroup. Similarly to the definition given
in [11], we say that a representation π of S is an inverse representation if the
set π(S) ∪ π(S)∗ generates a semigroup of partial isometries, i.e. generates
an inverse semigroup.
It is known that the left regular representation of
S is inverse. Note that the well-known requirement of commuting range
projections is not sufficient for a representation to be inverse. An example
of an abelian semigroup with isometric but non-inverse representation and a
condition for admitting such a representation for general abelian semigroup
were given in [2].
Lemma 4.3. There are one-to-one correspondences between inverse repre-
sentations of S and *-representations of SF , and between isometric inverse
representations of S and unital *-representations of S∗.
Proof. Given an inverse representation π of S and p ∈ S, set π(vp) = π(p),
∗) = π(p)∗. Then extend π to SF multiplicatively. Uniqueness
and π(vp
of an inverse then follows from Vagner's theorem (1.2) and the fact that
a product of two partial isometries is a partial isometry if and only if the
source projection of the first one commutes with the range projection of
the second (see also Proposition 2.3 in [26]). Given a *-representation π
of SF just set π(s) = π(s), and the image of an inverse semigroup under
*-homomorphism is an inverse semigroup.
14
MARAT AUKHADIEV
The second statement is verified similarly. If π is a unital *-representation
(cid:3)
pvp = 1 we get that π(p) = π(vp) is an isometry.
of S∗, since v∗
4.4. Recall the definition of the reduced C*-algebra of an inverse semigroup
(see [23] for details). Consider the Hilbert space ℓ2(P ) with standard basis
δs, s ∈ P . Define the left regular representation V : P → B(ℓ2(P )) by
(4.1)
Vsδt = (cid:26) st
0
if s∗st = t,
otherwise
Then V is a *-representation. The reduced C*-algebra of P is C ∗
r (P ) =
C ∗(Vs s ∈ P ) ⊂ B(ℓ2(P )).
4.5. Recall the construction of the C*-algebra of a left cancellative semi-
group (see [14]). Let S be a left cancellative semigroup. Consider the Hilbert
space ℓ2(S) with standard basis δp, p ∈ S. Define Vp ∈ B(ℓ2(S)) by
Vpδq = δpq
for all p, q ∈ S. Then one can check that
(4.2)
p δq = (cid:26) δr
V ∗
0
if q = pr,
otherwise
This is a faithful representation of S by isometries called the left regular
representation of S. The C*-algebra C ∗
r (S) = C ∗(Vp p ∈ S) ⊂ B(ℓ2(S)) is
the reduced semigroup C*-algebra of S.
Lemma 4.6. The left regular representation of S induces a non-degenerate
unital *-representation V ′ of S∗ on ℓ2(S).
Proof. As noticed before, the left regular representation of S is inverse, i.e.
the semigroup V (S) generated by the set {Vp p ∈ S} ∪ {V ∗
p p ∈ S} is
an inverse semigroup. The reason is that any element of V (S) is a shift
operator on the standard basis {δp}. Then due to Lemma 4.3, V induces a
*-representation of S∗ on ℓ2(S), given on the generators by
V ′(vp)δq = δpq
(cid:3)
Note that V ′ is not in general faithful, see Example 5.14. But the image
V ′(S∗) can be identified with the left inverse hull Il(S) represented on ℓ2(S).
Lemma 4.7. The left regular representation V of the inverse semigroup S∗
restricts to a *-representation on ℓ2(S) ⊂ ℓ2(S∗). This subrepresentation
coincides with the *-representation V ′ and its image is Il(S).
Proof. The Hilbert space ℓ2(S) is naturally embedded in ℓ2(S∗) by the map
given by δs 7→ δvs for all s ∈ S. It is sufficient to show invariance of ℓ2(S)
under the generating operators. For s, t ∈ S due to (4.1) and equality
v∗
s vs = 1, we have
V (vs)δvt = δvsvt = δvst
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
15
s , let us show that vsv∗
Before checking the same for v∗
t = sr for some r ∈ S. The implication "⇐" is obvious.
s vt = vt if and only if
Now suppose t 6= sr for any r ∈ S. Then using Lemma 4.6, we have
V ′(vt)δ1 6= δsr for any r ∈ S. Since by relation (4.2) operator V ′(vs)V ′(v∗
s )
is a projection onto a closed linear span of {δsr r ∈ S}, we have
s )V ′(vt)δ1 = 0 6= V ′(vt)δ1
s vt)δ1 = V ′(vs)V ′(v∗
V ′(vsv∗
We conclude that vsv∗
s vt 6= vt. Hence, using (4.1) we obtain
V (v∗
s )δvt = (cid:26) δvr
0
if t = sr,
otherwise
We see that ℓ2(S) as a subspace in ℓ2(S∗) is invariant under all V (vs), V (v∗
s )
an therefore under the whole image V (S∗). Moreover, using identification
δvt ↔ δt we get V ℓ2(S) = V ′.
(cid:3)
Lemma 4.8. The C*-algebra C ∗
lowing short exact sequence holds.
0 −→ Jr −→ C ∗
(4.3)
where Jr is the kernel of restriction of the left regular representation of S∗
onto ℓ2(S).
V ′(S∗) is isomorphic to C ∗
r (S) and the fol-
r (S∗) −→ C ∗
r (S) −→ 0
Proof. Due to Lemma 4.7 and Lemma 4.6, V ′ can be viewed as a restriction
of the left regular representation of S∗ onto ℓ2(S), and at the same time as
a *-representation of S∗ induced by the left regular representation VS of S.
So, viewing C ∗
r (S) as a C*-subalgebra in B(ℓ2(S∗)) generated by V ′(S∗) we
obtain C ∗
(cid:3)
r (S∗)/Jr ∼= C ∗
r (S).
4.9. In the case of a semigroup S embeddable in a group we can compute
the left regular representations of SF and S∗. Namely, both representations
are decomposable into a direct sum of representations, each of which can be
realized by a representation on the group G, generated by S. For this we
use the models for SF and S∗ given in Theorem 3.3 and Corollary 3.4 and
notations therein.
Recall that in this case the maximal group homomorphic image of SF
equals G, σ denotes the homomorphism SF → G. For any element s =
v∗
a1va2...v∗
an−1 van ∈ SF ,
Is = {1, σ(v∗
a1 ), σ(v∗
a1 va2 ), ..., σ(s)} ⊂ G.
This set depends on the form in which s is written. But all forms of s give
equivalent sets in the class [Is]. In the model for SF , s = (σ(s), [Is]) with
idempotent ss∗ corresponding to [Is].
Define for s ∈ SF , g ∈ G
Gs = {g ∈ G : there exist g1, g2 ∈ Is such that g1 ≤ g ≤ g2} ⊂ G,
Ls = {t ∈ SF : It∗ ⊂ Gs} = {t ∈ SF : It ⊂ σ(t)Gs} ⊂ SF ,
Ls,g = {t ∈ SF : It ⊂ σ(t)gσ(s)−1Gs}.
16
MARAT AUKHADIEV
Suppose g ∈ G, g1, g2 ∈ A, g1 ≤ g ≤ g2 and A ∼ Is for some fixed form
of s. Then by definition of ∼, there exist g′
1 , g′′
2 ∈ Is such that
g′
1 ≤ g1 ≤ g′
2, g′′
1, g′
2, g′′
1 ≤ g2 ≤ g′′
2
1 ≤ g ≤ g′′
Therefore, g′
2 and g ∈ Gs. This shows why Gs does not depend
on the form of s and is well-defined. Similarly the sets Ls and Ls,g are
well-defined.
For s1, s2 ∈ SF define
s1 ∼′ s2 if Is∗
1 ∼ Is∗
2 .
This is an equivalence relation on SF which does not generate a semigroup
congruence. We denote by [s] an equivalence class of s ∈ SF and by Q the
set of equivalence classes.
Theorem 4.10. Let S be embeddable in a group, denote by G the group
generated by S. The left regular representation V of SF is isomorphic to
s σ(s)) ⊂
L[s]∈Q πs, where πs are unital *-representations of SF on ℓ2(G−1
ℓ2(G) defined by
(4.4)
πs(t)δg = (cid:26) δσ(t)g
0
if t ∈ Ls,g,
otherwise
Proof. For any inverse semigroup P and any s1, s2 ∈ P we have s∗
if and only if there exists t ∈ P such that
1s1 = s∗
2s2
ts1 = s2 and t∗ts1 = s1.
Therefore, for any s ∈ P , the space
ℓ2({p ∈ P : p∗p = s∗s})
is invariant under the left regular representation of P .
Now fix an element s ∈ SF . Using the model for SF , s∗s = (1, [Is∗ ]).
Hence, p ∼′ s if and only if there exists t ∈ SF such that ts = p and
t∗ts = s. It follows that ℓ2([s]) is an invariant subspace of ℓ2(SF ) under the
representation V .
If t∗ts = s, then [It∗] = [σ(t)−1It] ≤ [Is]. It follows that for any fixed
form of t and s we have It∗ ⊂ Gs, i.e. t ∈ Ls. This implies σ(t)−1 ∈ Gs and
ts = (σ(t), [It])(σ(s), [Is]) = (σ(ts), [It ∪ σ(t)Is]) = (σ(ts), σ(t)[Is].
Hence, if t ∈ Ls, the product ts depends only on σ(t).
And due to Theorem 3.3 for any g ∈ G−1
s we have [Is] = [Is ∪ {g−1}] and
there exists an element t ∈ SF such that [It∗ ] ≤ [Is] and σ(t) = g.
Therefore, σ gives a bijection between [s] ⊂ SF and G−1
s σ(s) ⊂ G. Denote
by αs the corresponding isomorphism between ℓ2([s]) and ℓ2(G−1
s σ(s)).
Repeating calculations as above, for any r ∈ SF and ts ∈ [s], g = σ(t) we
have for any Ir∗ corresponding to a fixed form of r that
r∗r(ts) = ts ⇐⇒ Ir∗ ⊂ gσ(s)−1Gs ⇐⇒ r ∈ Ls,g.
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
17
Consequently, defining the representation πs by formula (4.4) we obtain
for any t, x ∈ SF :
αs(Vtδx) = πs(t)δσ(x).
Thus, αs is an isomorphism between the restriction of V onto ℓ2([s]) and
πs.
(cid:3)
4.11. A similar result holds for the semigroup S∗. We use σ to denote the
homomorphism S∗ → G. Recall the equivalence on SF defining S∗. For any
g ∈ G, a ∈ S:
v∗
ava ∼ 1, {1, g, ga} ∼′ {1, ga}.
For any s ∈ S∗ define
Ps = {t ∈ S∗ : It∗ ∼′ Is∗}.
Denote by R the set of all sets Ps ⊂ S∗. Also define
Ds = Is · S−1 = {ga−1 : g ∈ Is, a ∈ S} ⊂ G
Theorem 4.12. Let G be a group generated by a semigroup S. The left
regular representation V of S∗ is isomorphic to LPs∈R πs, where πs are
unital *-representations of S∗ on ℓ2(D−1
(4.5)
πs(t)δg = (cid:26) δσ(t)g
0
s σ(s)) ⊂ ℓ2(G) defined by
if It∗ ⊂ gσ(s)−1Ds,
otherwise
Proof. The proof is similar to Theorem 4.10, the difference is in the quasi
order ≤ on exp(G) under the equivalence ∼′. For A, B ∈ exp(G), [A] ≤ [B]
if and only if for any g ∈ A there exists a ∈ S such that ga ∈ B.
Fix an element s ∈ S∗ and Is for one of its forms. Then for any t ∈ S∗
we have [It∗] ≤ [Is] if and only if It∗ ⊂ Ds for any form of t. This implies
that σ(t) ∈ D−1
s σ(s). As in Theorem 4.10, when t satisfies
t∗ts = s, the product ts depends only on σ(t). Consequently,
and σ(ts) ∈ D−1
s
Ps = {ts : t∗ts = s} ⊂ S∗
is bijective to D−1
s σ(s) ⊂ G, with bijection implemented by σ.
Let g ∈ D−1
s σ(s). Then we can find t ∈ S∗ such that t∗ts = s and
σ(t) = gσ(s)−1. For any r ∈ S∗ we have r∗rts = ts if and only if
Ir∗ ⊂ Dts = ItsS−1 = σ(t)Ds = gσ(s)−1Ds.
Hence, the statement of Theorem follows.
(cid:3)
5. The full C*-algebras of the universal inverse semigroups
5.1. Let P be an inverse semigroup. Consider the space ℓ1(P ), define prod-
uct and involution:
btδt) = ( Xs,t∈P
asbtδst)
(Xs∈P
asδs)(Xt∈P
(Xs∈P
asδs)∗ = Xs∈P
asδs∗
18
MARAT AUKHADIEV
Then ℓ1(P ) is a Banach *-algebra. Any *-representation of P extends to a
*-representation of ℓ1(P ) and the converse is true. The full semigroup C*-
algebra C ∗(P ) of P is the completion of ℓ1(P ) under the supremum norm
over all *-representations of P ([23]).
5.2. For a left cancellative semigroup S let J be the set of all constructible
right ideals in S (see Paragraph 2.9). Define an inverse semigroup W gen-
erated by isometries ws, s ∈ S and projections eX , X ∈ J satisfying the
following relations for all s, t ∈ S, X, Y ∈ J :
wst = wswt, wseX w∗
s = esX,
eS = 1, e∅ = 0, eX∩Y = eXeY
The universal C*-algebra generated by W is the full semigroup C*-algebra
of S, denoted C ∗(S) ([14]).
Lemma 5.3. There exists a surjective *-homomorphism C ∗(S∗) → C ∗(S).
Proof. Clearly, W is an inverse semigroup. By Corollary 2.10 in [14], W is
generated by isometries {ws s ∈ S}, i.e. the relations defining W can be
formulated in terms of ws. Namely, for X = a−1
2 (a1S)) we have
n (an−1...a−1
a2 wa1)∗
an wan−1 ...w∗
eX = w∗
anwan−1 ...w∗
a2 wa1(w∗
Then W is defined by requirements that idempotent monomials are equal
when corresponding ideals in S are equal as sets, and that the product of
ideals is respected. Imposing the same relations on S∗ denoted R5, we obtain
that the quotient S∗/R5 equals W . Then this semigroup *-homomorphism
extends to a surjective *-homomorphism C ∗(S∗) → C ∗(S).
(cid:3)
Remark 5.4. In terms of Paragraph 2.9 the relations R5 consist of R2
and the first half of R3. Therefore, the *-homomorphism γ : S∗ → Il(S) in
Lemma 2.10 factors through S∗ → W defined above.
5.5. As pointed out in the Section 2.5 of [14], the construction of the full
semigroup C*-algebra is not functorial, i.e. not every semigroup morphism
of two cancellative semigroups extends to *-homomorphism of their C*-
algebras. This is demonstrated further in Example 5.14. The reason lies in
the construction of the inverse semigroup W in the definition of C ∗(S). As
we will see now, the map S → C ∗(S∗), on the contrary, is functorial.
Proposition 5.6. Let φ : S → T be a semigroup morphism between two left
cancellative semigroups. Then φ extends to a *-homomorphism φ : C ∗(S∗) →
C ∗(T ∗).
Proof. Due to Theorem 4.8 in [29], the imbedding S ֒→ S∗ is functorial, so φ
induces a *-homomorphism S∗ → T ∗. By the universal property of inverse
semigroup C*-algebras we obtain the required *-homomorphism between the
full C*-algebras.
(cid:3)
y
y
y
y
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
19
Theorem 5.7. For any left cancellative semigroup S we have the follow-
ing commutative diagram in which every line is a short exact sequence and
vertical maps are surjective *-homomorphism.
0 −−−−→ J −−−−→ C ∗(S∗) −−−−→ C ∗(S) −−−−→ 0
(5.1)
0 −−−−→ Jr −−−−→ C ∗
r (S∗) −−−−→ C ∗
r (S) −−−−→ 0
Here J is a closed ideal in C ∗(S∗) generated by {x − y x, y ∈ S∗, x ≈R5 y},
and Jr is defined in Lemma 4.8.
Proof. The upper short exact sequence is given by Lemma 5.3, the lower
one was proved in Lemma 4.8. The vertical maps are the left regular rep-
resentations of S∗ and S respectively. The commutativity of the diagram
then follows from the fact that all these *-homomorphisms are induced by
*-homomorphisms of the corresponding inverse semigroups.
(cid:3)
Remark 5.8. Due to Remark 5.4, C ∗
r (S) is a C*-algebra generated by a
*-representation of W (in the same way as S∗ and Il(S)) on ℓ2(S), which
may be not the same as the left regular representation of W and may be
not faithful. Therefore in general C ∗
r (S) is not isomorphic to C ∗
r (W ).
Corollary 5.9. Using the results of [21], one has the following commutative
diagram where each map is a surjective *-homomorphism:
C ∗(S∗) −−−−→ C ∗(S) −−−−→ C ∗(Il(S))
y
C ∗
r (S∗) −−−−→ C ∗
r (S) ←−−−− C ∗
r (Il(S))
Remark 5.10. Note that the last arrow is reversed due to the fact that
C ∗
r (S) is generated by a subrepresentation of the left regular representation
of S∗ (or Il(S)). Note also that for inverse semigroups we consider only
0-preserving representations, so that in our notation C ∗(Il(S)) is the same
as C ∗
0 (Il(S)) used in [21].
5.11. We denote G = G(S∗) the maximal group homomorphic image of S∗
defined in Paragraph 1.1. Then the semigroup homomorphism σ : S∗ → G
extends to a surjective *-homomorphism C ∗(S∗) → C ∗(G). This surjection
is a quotient homomorphism by a closed ideal I in C ∗(S∗) generated by
the set {1 − p p ∈ S∗, p is an idempotent }. And using the definition of S∗
one can see that this ideal is generated by a smaller set {1 − vsv∗
s s ∈ S}.
By Proposition 1.4 of [7] with a proof in Proposition 4.1 of [24], the *-
homomorphism P → G(P ) extends to a surjective *-homomorphism of the
reduced C*-algebras. Therefore we obtain the following statement.
Proposition 5.12. Let S be a left cancellative semigroup and G the maximal
group homomorphic image of S∗. Then the following diagram is commuta-
tive, every line is a short exact sequence and vertical maps are surjective
20
MARAT AUKHADIEV
*-homomorphism.
0 −−−−→ I −−−−→ C ∗(S∗) −−−−→ C ∗(G) −−−−→ 0
(5.2)
y
y
y
0 −−−−→ Ir −−−−→ C ∗
r (G) −−−−→ 0
Here I is a closed ideal in C ∗(S∗) generated by the set {1 − vsv∗
Ir is its image under the left regular representation of S∗.
r (S∗) −−−−→ C ∗
s s ∈ S} and
Remark 5.13. There is one more ideal in C ∗(S∗) worth to be mentioned,
namely a closed ideal C generated by all commutators. Clearly, C contains
the ideal I, because v∗
s vs = 1. In the case of an abelian semigroup S, these
ideals coincide. In the case S = Z+, one can easily verify C ∗(S) = C ∗(S∗)
is the Toeplitz algebra and then the ideal C is isomorphic to the ideal of
compact operators. In the general case C ∗(S∗)/C ∼= C ∗(Gab), where Gab is
an abelianization of G.
Example 5.14. For any natural number n let F+
n denote the free monoid
on n generators a1, ..., an with the empty word e. Let us compare C ∗(F+
n )
(see [15] for a detailed description) with C ∗(F+
n = X1 ⊔ X2 ⊔
n
...Xn ⊔ {e}, where Xi = aiF+
n for i = 1, ..., n are constructible right ideals in
Sn (see Paragraph 2.9). Therefore in C ∗(F+
). We have F+
n ) for i 6= j we have
∗
(5.3)
vaiv∗
aivaj v∗
aj = eXieXj = e∅ = 0,
n ), V ∗
ai vaj = 0. The same holds in C ∗
which implies v∗
the maximal group homomorphic image of Il(F+
there is no canonical surjective *-homomorhism from C ∗(F+
r (F+
aiVaj = 0. Hence,
n ) is trivial. It follows that
n ) onto C ∗(G).
Moreover, knowing that every semigroup is a homomorphic image of a free
monoid, one would expect the same to hold for the corresponding semigroup
C*-algebras, but this is not true. To see this consider the semigroup S =
Z+ × Z+, which is an abelian semigroup generated by t1 = (1, 0) and t2 =
(0, 1). Then there is a canonical homomorphism φ : F+
2 → S sending a1 7→ t1,
a2 7→ t2. And we have (1, 0)S ∩ (0, 1)S = (1, 1)S, hence vt1v∗
t2 =
v(1,1)v∗
(1,1) 6= 0. Consequently, due to (5.3) homomorphism φ does not extend
to a *-homomorphism C ∗(F+
t1vt2 v∗
2 ) → C ∗(S).
∗
In the semigroup F+
n
(and its C*-algebra) the product va1 v∗
a2 is
a non-zero idempotent which allows the desired homomorphism to exist.
And in general, the semigroup F+
has no zero, and its maximal group
n
homomorphic image is the free group Fn. So, from the diagram (5.2) we get
the following short exact sequence:
a1va2v∗
∗
0 → I → C ∗(F+
n
∗
) → C ∗(Fn).
The same holds for the reduced C*-algebras.
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
21
∗
∗
Since every semigroup is a homomorphic image of a free monoid using
Proposition 5.6 we deduce that for any n-generated cancellative semigroup
S the C*-algebra C ∗(S∗) is a homomorphic image of C ∗(F+
n
There is a natural quotient for C ∗(F+
n
ai. Multiplying p by idempotents vaj v∗
aj and so contains v∗
) in the case n = 2 and n = 3.
Consider a closed two-sided ideal In generated by the element p = 1 −
aj one by one we obtain that In
aivaj v∗
aivaj for all i 6= j. Consequently,
∗)/In is canonically isomorphic to the Cuntz algebra On.
) →
n ) → On, where the last map is given in Paragraph 8.2 of [15]. In order
aivaj to the ideal
Pi vaiv∗
contains every vaiv∗
the quotient C ∗(F+
n
The quotient homomorphism can also be seen as the composition C ∗(F+
n
C ∗(F+
to get the same in the case n ≥ 4 one has to add elements v∗
In.
).
∗
5.15. Many of the results in [14] and [15] hold only under the assumption
of independence of the constructible right ideals. According to [14], for a
cancellative semigroup S the set of constructible right ideals J is said to be
independent, if X = ∪n
i=1Xi for X, X1, ..., Xn ∈ J implies X = Xi for some
i.
If some of the constructible right ideals are not independent, then the
images of corresponding idempotents under the left regular representation
of S are not linearly independent.
In such case the diagonal subalgebra
D = C ∗(eX : X ∈ J ) in C ∗(S) is not isomorphic to its image under the left
regular representation.
Let us illustrate such situation on a semigroup S = Z+ \{1} = {0, 2, 3, ...}
with the usual addition operation. In the notation (2.1), (2.2) we have:
2 + S = {2, 4, 5, 6, ...},
3 + S = {3, 5, 6, 7, ...},
(−3) + 2 + S = {2, 3, 4, ...}.
It follows that (2 + S) ∪ (3 + S) = (−3) + 2 + S. Therefore, in C ∗
have
r (S) we
2 V3V ∗
3 = V ∗
V2V ∗
2 + V3V ∗
3 − V2V ∗
3 V2V ∗
But this problem does not exist for the algebra C ∗
2 V3.
r (S∗). Due to [32] the
left regular representation of an inverse semigroup P extends to a faithful
representation of ℓ1(P ) on ℓ2(P ). Hence, images of elements of S∗ under
the left regular representation are all linearly independent. The reason is
that there are enough vectors in the basis of ℓ2(S∗) to differ elements of
S∗, unlike the subspace ℓ2(S). In the same way the semigroup elements of
S∗ are independent in C ∗(S∗) and in C ∗(S), since these are full semigroup
C*-algebras of inverse semigroups (S∗ and W respectively).
We end this section with a note on nuclearity of semigroup C*-algebras
in the case when S is abelian. This is a direct consequence of Murphy's [19].
Proposition 5.16. For an abelian cancellative semigroup S the algebras
C ∗(S∗), C ∗
r (S), C ∗(Il(S)), C ∗
r (Il(S)) are nuclear.
r (S∗), C ∗(S), C ∗
22
MARAT AUKHADIEV
Proof. All of the mentioned C*-algebras in this case are generated by com-
muting isometries with commuting range projections, so we can apply The-
orem 4.8 of [19].
(cid:3)
6. Amenability and nuclearity
The classical definition of amenability is common for all semigroups [25],
we recall it further. Let P be a semigroup. Right action of P on ℓ∞(P ) is
given by
φt(x) = φ(tx),
where φ ∈ ℓ∞(P ), t, x ∈ P . A mean m on ℓ∞(P ) is called left invariant if
m(φt) = m(φ) for all t ∈ P , φ ∈ ℓ∞(P ). The semigroup P is left amenable if
there exists a left invariant mean. For an inverse semigroup left amenability,
right amenability and (two-sided) amenability are equivalent. A result of
Duncan and Namioka [8] states that an inverse semigroup is amenable
if and only if its maximal group homomorphic image is amenable. The
following result of [17] connects amenability of inverse semigroup with the
weak containment property.
Theorem 6.1. (D. Milan [17]) Let P be an E-unitary inverse semigroup.
Then the following statements are equivalent
(1) P is amenable,
(2) P satisfies the weak containment property, i.e. C ∗(P ) = C ∗
(3) the maximal group homomorphic image G(P ) is amenable.
r (P ),
Corollary 6.2. Let S be embeddable in a group. Then the following condi-
tions are equivalent and imply that C ∗
r (S) and C ∗(S) are nuclear:
(1) the group G generated by S is amenable,
(2) S∗ is amenable,
(3) SF is amenable,
(4) S∗ has the weak containment property, i.e. C ∗(S∗) = C ∗
(5) SF has the weak containment property,
(6) C ∗
(7) C ∗
r (S∗) is nuclear,
r (SF ) is nuclear.
r (S∗),
Proof. By Corollary 3.5, inverse semigroups SF and S∗ are E-unitary and
G(SF ) = G(S∗) coincides with the group generated by S. Applying Theorem 6.1,
we get equivalence of all the conditions 1)-5) above. By Proposition 6.6. in
[18], the maximal group homomorphic image G(P ) of an inverse semigroup
P is amenable if and only if the universal groupoid of P is amenable, which
is equivalent to nuclearity of its reduced C*-algebra by Theorem 5.6.18 in
[4]. Hence, we obtain equivalence of 6) and 7) to 1). Finally, conditions 6)
and 7) and the diagram (5.1) imply that C ∗
r (S) and C ∗(S) are nuclear. (cid:3)
Corollary 6.3. Let S be a left amenable cancellative semigroup. Then S∗
and SF are amenable inverse semigroups.
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
23
Proof. By Proposition 1.27 in [25], every left amenable cancellative semi-
group embeds in a group G, such that G = SS−1 and G is amenable.
Hence, condition 1) in Corollary 6.2 is satisfied.
(cid:3)
r (S∗) (or C ∗
r (S) does not imply nuclearity
r (SF )). The counterexample is given by the free monoid
n , the subsemigroup generating the free group Fn. By Example 5.14,
∗) →
) is not nuclear. This is natural considering the fact
n is not amenable. Nevertheless, as shown by Nica in [20], the C*-
Remark 6.4. Nuclearity of the C*-algebra C ∗
of C ∗
F+
there exist surjective *-homomorphisms C ∗(F+
n
C ∗
that F+
algebra C ∗(F+
∗) → C ∗(Fn) and C ∗
r (Fn). Hence, C ∗
r (F+
n
∗
r (F+
n
n ) is nuclear.
We finish this section with a result connecting directly amenability of S∗
to amenability of S, analogous to Proposition 1.27 in [25].
Proposition 6.5. Let S be a left cancellative left amenable semigroup. Then
S∗ is an amenable inverse semigroup.
Proof. We consider S embedded in S∗ by a map a ∈ S 7→ va ∈ S∗ and
ℓ∞(S) ⊂ ℓ∞(S∗). Let m0 be a left invariant mean on ℓ∞(S) and define a
mean on ℓ∞(S∗) by m(φ) = m0(φS), where for any x ∈ S we put
φS(x) = φ(vx).
Since S∗ is generated by S, it is sufficient to check left invariance of m on
b for all a, b ∈ S. Take φ ∈ ℓ∞(S∗) and a ∈ S and calculate
generators va, v∗
using left invariance of m0:
m(φva) = m0((φva)S ) = m0(φSa) = m0(φS ) = m(φ),
a) = m0((φv∗
m(φv∗
Hence, m is a left invariant mean on ℓ∞(S∗).
a)S ) = m0((φv∗
a)Sa) = m0((φv∗
ava)S) = m0(φS) = m(φ)
(cid:3)
7. Crossed products of universal inverse semigroups
Definition 7.1. Let P be an inverse semigroup. An action α of P on a
space X is a *-homomorphism P → I(X), αs : Ds∗ → Ds, such that the
union of all Ds coincides with X. We call it unital if the image of the unit
element in P is the identity map on X. If X is a locally compact Hausdorff
topological space, we require every αs to be continuous and Ds to be open
in X.
Lemma 7.2. There is a one-to-one correspondence between actions of S on
a space X by injective maps and unital actions of S∗ on X.
Proof. Let α be an action of S on X, i.e. αsαt = αst for any s, t ∈ S, such
that each αs is injective. By Lemma 2.4, denoting Ds ⊂ X the image of αs
and defining α∗
s : Ds → X as the inverse of αs, we get a set generating an
inverse subsemigroup S∗
sαs is
an identity on X for any s ∈ S. Hence, there is a surjective *-homomorphism
α in I(X). Clearly, in this semigroup the map α∗
24
MARAT AUKHADIEV
α : S∗ → S∗
we see that α is a unital action of S∗ on X.
α, which gives an action of S∗ by partial bijections on X. And
Now suppose α is a unital action of S∗ on X. Then define α(s) = α(vs)
for all s ∈ S. Then multiplicativity follows immediately. Unitality of α
implies that
α(s)∗ α(s) = α(v∗
s vs) = id.
Hence, for every s ∈ S α(s) is a bijection with the domain equal to X. (cid:3)
Remark 7.3. The previous lemma holds also in topological setting, i.e.
when X is a locally compact Hausdorff topological space and each αs is
continuous.
Definition 7.4. By an injective action with ideal images α of a left can-
cellative semigroup S on a C*-algebra A we mean a set of injective *-
homomorphisms αs on A such that for every s, t ∈ S, αst = αsαt and
αs(A) is a closed two-sided *-ideal in A. In this case we say that (α, S, A)
is an injective C*-dynamical system.
A partial automorphism φ on a C*-algebra A is a *-isomorphism φ : J1 →
J2, where J1, J2 are closed two-sided *-ideals in A. For a C*-algebra A
denote by I(A) the inverse semigroup of partial automorphisms on A, with
a product and an inverse map defined similarly to I(X) (see Section 2).
An action α of an inverse semigroup P on a C*-algebra A is a *-homo-
morphism P → I(A), αs : Es∗ → Es, such that the union of all Es coincides
with A. In this case we say that (α, P, A) is a C*-dynamical system. If P
has a unit 1 and α1 = idA, we say that the action α is unital.
Lemma 7.5. There is a one-to-one correspondence between injective actions
of S with ideal images on a C*-algebra A and unital actions of S∗ on A.
Proof. For an injective action α of S on A, we define for any s ∈ S the
s = A and the range Evs = αs(A) of α(vs) = αs. For the inverse
domain Ev∗
map we put α(v∗
s . Following the proof of Lemma 7.2,
we obtain an action of S∗ on the underlying space of A. Since αs is a *-
homomorphism, the same is true for α(vs), α(v∗
s ) and the products of such
maps. Hence, α given by Lemma 7.2 is an action of S∗ on the C*-algebra
A. The reverse statement follows similarly from Lemma 7.2.
(cid:3)
s : Evs → Ev∗
s ) = α∗
Remark 7.6. Notice that if A is unital and α is an induced action of S∗
on A as in the Lemma, we may extend α(v∗
s ) to a *-endomorphism on A by
setting α(v∗
s(αs(1)a). But one should remember that this extension
is injective only on Evs. Then α is an action of S∗ on A by *-endomorphisms.
s )(a) = α∗
Definition 7.7. Let S be a left cancellative semigroup with an action α on
a C*-algebra A. A covariant representation (see [12]) of the C*-dynamical
system (α, S, A) is a pair (π, T ) in which
(1) π is a non-degenerate *-representation of A on H,
(2) T : S → B(H) is a unital isometric representation of S,
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
25
(3) the covariance condition π(αs(a)) = Tsπ(a)T ∗
s holds for every a ∈
A, s ∈ S.
Lemma 7.8. If (π, T ) is a covariant representation of an injective C*-
dynamical system with ideal images (α, S, A) and A is unital, then T is an
inverse representation of S.
Proof. It is sufficient to show that all idempotents in the semigroup ST gen-
erated by T (S) ∪ T (S)∗ commute, i.e. xx∗yy∗ = yy∗xx∗ for any monomials
x and y. Obviously, in the notation of Lemma 7.5, TsT ∗
s = π(αs(1)) ∈
π(Es). Due to the Remark 7.6 and the covariance condition, π(α(v∗
s )(a)) =
T ∗
s π(a)Ts. Generally, for any monomial x ∈ ST and a ∈ A we have
(7.1)
xx∗ = π(α(x)(1)),
(7.2)
xπ(a)x∗ = π(α(x)(a))
For x = T ∗
s , s ∈ S using the fact that images of α are ideals we deduce:
T ∗
s π(a) = T ∗
s π(aα(s)(1)) = T ∗
s π(α(s)(1)a) = T ∗
s π(a)TsT ∗
s ,
(7.3)
(7.4)
π(a)Ts = π(aα(s)(1))Ts = π(α(s)(1)a)Ts = TsT ∗
s π(a)Ts
For general type of monomial these formulas can be shown by induction on
the length of monomial. Suppose the formulas hold for the length equal n
and take x = yz, where y has length n and y is a generator. First assume
z = Ts and for a ∈ A calculate
π(a)x = π(a)yTs = yy∗π(a)yTs =
By (7.2), we have y∗π(a)y ∈ π(A), and due to (7.4)
If z = T ∗
s , by T ∗
y(y∗π(a)y)Ts = yTsT ∗
s Ts = 1 we similarly have
s y∗π(a)yTs = xx∗π(a)x.
π(a)x = π(a)yT ∗
s = yy∗π(a)yT ∗
s = yT ∗
s Tsy∗π(a)yT ∗
s = xx∗π(a)x.
Thus, for any monomial x we have
(7.5)
π(a)x = xx∗π(a)x.
In the same way, splitting x into Tsy or T ∗
s y and using (7.3), we obtain
(7.6)
Then for any monomials x, y using that yy∗ ∈ π(A) by (7.1), we get
xπ(a) = xπ(a)x∗x.
xx∗yy∗ (7.6)
= xx∗yy∗xx∗ (7.5)
= yy∗xx∗
Definition 7.9. Let P be an inverse semigroup with an action α on a C*-
algebra A. A covariant representation (see [30]) of the C*-dynamical system
(α, P, A) is a pair (π, T ) in which
(1) π is a non-degenerate *-representation of A on H,
(cid:3)
26
MARAT AUKHADIEV
(2) T : P → B(H) is a unital *-representation of P , such that for every
s ∈ P , T ∗
s TsH = π(Es∗)H and TsT ∗
s H = π(Es)H
(3) the covariance condition π(αs(a)) = Tsπ(a)T ∗
s holds for every a ∈
Es∗, s ∈ P .
Lemma 7.10. Let α be an injective action of a left cancellative semigroup
S on a C*-algebra A and α the induced action of S∗ on A. Then there is a
one-to-one correspondence between the covariant representations of (α, S, A)
and (α, S∗, A).
Proof. Let (π, T ) be a covariant representation of (α, S, A) on H. By Lemma
4.3 and Lemma 7.8, T induces a *-representation T of S∗ on H given by
Tvs = Ts, Tv∗
s = T ∗
s .
The condition (7.2) gives the condition (3) of the Definition 7.9.
Now we prove condition (2) of Definition (7.9). Let v ∈ S∗, a ∈ Ev∗ and
b = αv(a). Due to covariance condition, we have
Tvπ(a) T ∗
π(αv(a)) = Tvπ(a) T ∗
v = Tv T ∗
v
(7.7)
Hence, π(Ev)H ⊂ Tv T ∗
v H.
v = Tv T ∗
v π(αv(a)).
We prove the reverse inclusion by induction on the length of v. First
suppose v = vsw, where s ∈ S, w ∈ S∗, and assume that the inclusion is
proved for w. It implies that for x ∈ H, the vector Tv T ∗
Tv∗
s x
Tvsπ(ai)yi for some ai ∈ Ew and yi ∈ H. Hence,
can be approximated by Pi
Tvsπ(ai) T ∗
vs
π(αs(ai)) Tvsyi ∈ π(Evsw)H
we obtain
Tv T ∗
v x = Tvs
Tw T ∗
w
v x ≈ Xi
Tvsyi = Xi
v x = T ∗
vs
Now suppose v = v∗
proved for w. Similarly the vector Tv T ∗
Pi
s w for s ∈ S, w ∈ S∗, and assume that the inclusion is
Tvsx is approximated by
T ∗
vsπ(ai)yi for some ai ∈ Ew and yi ∈ H. Denote by uλ the approximate
unit of A. Due to the fact that π is a non-degenerate representation of
A, π(uλ) converges to the identity operator on H in the strong operator
topology, i.e. y ≈ π(uλ)y for any y ∈ H. Then we obtain
Tvsπ(uλ) T ∗
Tw T ∗
w
vs π(ai)yi =
Tv T ∗
T ∗
vs
Tvs
T ∗
vs
T ∗
v x = Xi
vsπ(ai)yi ≈ Xi
T ∗
Xi
vsπ(αs(uλ))π(ai)yi = Xi
T ∗
vsπ(αs(bi,λ))yi,
where bi,λ = αs(uλ)ai ∈ Evs ∩ Ew. Since bi,λ ∈ Evs, using (7.7) we get
Xi
T ∗
vsπ(bi,λ)yi = Xi
vsπ(bi,λ) Tvs
T ∗
T ∗
vsyi = Xi
π(α∗
vs(bi,λ)) T ∗
vs yi,
which belongs to π(Ev∗
covariant representation of (α, S∗, A).
s w)H due to definition of Ev∗
s w. Thus, (π, T ) is a
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
27
If (π, T ) is any covariant representation of (α, S∗, A), then Ts = Tvs gives a
unital inverse representation of S by Lemma 4.3. Since α is just a restriction
of α, the covariance condition also holds.
(cid:3)
Remark 7.11. The reverse statement to the previous Lemma also holds.
Let S be a left cancellative semigroup and α be an action of the universal
inverse semigroup S∗ on a C*-algebra A, α the induced action of S on A.
Then there is a correspondence between the covariant representations of
(α, S∗, A) and (α, S, A). The proof is the same as above.
7.12. We now recall the definitions of crossed products by partial automor-
phisms of inverse semigroups and crossed products by injective actions of
cancellative semigroups. For the case of inverse semigroups acting on (in
general non-unital) C*-algebras by partial automorphisms there is a well-
established definition, given in [30], [9], [10].
For cancellative semigroups there exist several constructions of the crossed
product, by automorphisms ([14]) and by endomorphisms ([12]), both of
them for unital C*-algebras. Though an automorphism is a particular case
of an endomorphism, the construction of the crossed product by an auto-
morphism is not a particular case of the one by an endomorphism.
The main difference of the constructions is that in the case of an au-
tomorphism the crossed product contains the whole semigroup C*-algebra,
while the crossed product by an endomorphism contains only isometries cor-
responding to the elements of the semigroup. A connection between crossed
products by automorphisms and crossed products by inverse semigroups is
given in Proposition 5.7 of [15]. It describes the isomorphism between A⋊a
αS
and (A⊗D)⋊β S in the case when S is a subsemigroup of a group. Here D is
the canonical commutative C*-subalgebra in C ∗(S), and S is some specific
*-homomorphic image of S∗. Thus, the subject of the present paper is the
notion of crossed product by an endomorphism of S.
Definition 7.13. Let α be an action of a left cancellative semigroup S on
a unital C*-algebra A. The crossed product associated to the C*-dynamical
system (α, S, A) is a C*-algebra A ⋊α S with a unital *-homomorphism
iA : A → A ⋊α S and an isometric representation iS : S → A ⋊α S such that
(1) (iA, iS ) is a covariant representation for (α, S, A),
(2) for any other covariant representation (π, T ) there is a representation
π × T of A ⋊α S such that π = (π × T ) ◦ iA and T = (π × T ) ◦ iS,
(3) A ⋊α S is generated by iA(A) and iS(S) as a C*-algebra.
As noticed in [12], the crossed product is non-trivial if S acts by injec-
tive endomorphisms. A ⋊α S can be defined as a C*-algebra generated by
monomials aws and (aws)∗ = w∗
s a∗ where a ∈ iA(A), and ws = iS(s). The
completion is taken with respect to the norm given by the supremum of
(π × T )(x) over all covariant representations (π, T ). Obviously, for any
covariant representations (π, T ), the representation π × T of A ⋊α S is given
28
by
MARAT AUKHADIEV
(π × T )(w∗
t aws) = T ∗
t π(a)Ts.
Note that since A and S are unital, the elements ws generate a semigroup
isomorphic to S∗ and we can change the notation ws to vs.
Definition 7.14. Let α be an action of an inverse semigroup P on a C*-
algebra A. Denote by L the linear space of finite sums Ps∈P asδs where
as ∈ Es and δs is a formal symbol. Multiplication and involution are defined
in the following way.
(aδs)(bδt) = αs(αs∗ (a)b)δts, (aδs)∗ = αs∗(a∗)δs∗.
For any covariant representation (π, T ) define a non-degenerate *-represen-
tation of L:
(π × T )(Xs∈P
asδs) = Xs∈P
π(as)Ts
The crossed product A ⋊α P is the Hausdorff completion of L in the norm
x = supΠΠ(x),
where the supremum is taken over all representations of L of the form Π =
π × T for all covariant representations (π, T ).
Theorem 7.15. Let A be a unital C*-algebra and let α be an injective action
with ideal images of S on A, α be an action of S∗ on A, where one of them
is induced by another. Then the crossed product C*-algebras A ⋊α S and
A ⋊ α S∗ are isomorphic.
Proof. Due to Lemma 7.10 and Remark 7.11, it is sufficient to give a *-
isomorphism of the underlying *-algebras of monomials, K for A ⋊α S and
L for A ⋊ α S∗ respectively.
First note that any monomial in K can be written as av∗
s1vs2...vsn, where
a ∈ Ev∗
vs2 ...vsn in the notation of Lemma 7.5. Indeed, using the covariance
s1
condition and the assumption that the images of the endomorphisms are
ideals, we obtain for any a ∈ A, s ∈ S:
s a = v∗
v∗
s αs(1)a = v∗
s(αs(1)a)v∗
s ,
avs = avsv∗
s αs(α∗
vsa = αs(a)vs.
s vs = aαs(1)vs,
s(αs(1)a)) = α∗
As noticed in the proof of Lemma 7.5, we may assume that α(v∗
s(αs(1)a) for any a ∈ A, remembering that α(v∗
α∗
on Evs. Therefore, for any monomial x = v∗
xa = α(x)(a)x.
s )(a) =
s ) is an isomorphism only
s1vs2...vsn and a ∈ A we obtain
Define φ : K → L on generators by a → aδ1, vs → α(vs)(1)δvs and on an
arbitrary monomial x and a ∈ Ex by
φ(ax) = aδx.
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
29
Extend φ linearly to K. Then clearly, φ(K) = L. To show that φ is
multiplicative, calculate the product of arbitrary monomials av∗
t1 vt2...vtn and
bv∗
s1vs2...vsk , where a ∈ Ex, b ∈ Ey, x = v∗
t1vt2...vtn , y = v∗
axby = aα(x)(b)xy = α(x)(α(x∗)(a)b)xy.
s1vs2...vsk :
Therefore, due to the Definition 7.14
φ(axby) = φ(α(x)(α(x∗)(a)b)xy) = α(x)(α(x∗)(a)b)δxy = (aδx)(bδy).
In the same way we verify that φ preserves involution:
φ(ax)∗ = (aδx)∗ = α(x∗)(a∗)δx∗ = φ(α(x∗)(a∗)x∗) = φ(x∗a∗).
Thus, φ is a *-isomorphism onto L.
Corollary 7.16. There exist an injective action β of S on C ∗(E), where
E is the semigroup of idempotents in S∗, and an action α of S∗ on EJ =
{eX X ∈ J }. With respect to thess actions,
(cid:3)
C ∗(S∗) ∼= C ∗(E) ⋊ β S, C ∗(S) ∼= C ∗(EJ ) ⋊ α S∗.
Proof. For any inverse semigroup P there exists an action β by partial bijec-
tions on its subsemigroup E of idempotents. Namely, for x ∈ P , the domain
of βx is Dx∗ = {f : f = x∗xf }, and βx(f ) = xf x∗. This action extends to
an action β of P on the commutative C*-algebra C ∗(E) and by Proposition
4.11 of [30], C ∗(P ) is isomorphic to the crossed product C ∗(E) ⋊β P . Take
P = S∗ and see that β is unital and βvs is injective since vs are isometries for
s ∈ S, and the images of the extension of β to C ∗(E) are closed ideals. Then
by Lemma 7.5 we get an action β of S on C ∗(E) by injective endomorphisms
with ideal images. Applying Theorem 7.15 we obtain
C ∗(S∗) ∼= C ∗(E) ⋊β S∗ ∼= C ∗(E) ⋊ β S.
By Lemma 2.14 in [14], C ∗(S) ∼= C ∗(EJ ) ⋊α S, where α is an injective
action given on generators by αs(eX ) = esX for s ∈ S, X ∈ J . Therefore,
by Lemma 7.5 α generates an action α of S∗ on C ∗(EJ ). Using again
Theorem 7.15, we obtain the required isomorphism.
(cid:3)
Remark 7.17. The isomorphism C ∗(S) ∼= C ∗(EJ ) ⋊α S of [14] mentioned
above could be deduced directly from the fact that C ∗(S) is a C*-algebra of
an inverse semigroup W , which is a quotient of S∗, using Theorem 7.15.
Corollary 7.18. Let S be a left Ore semigroup generating a group G =
S−1S and α be a unital action of S∗ on a C*-algebra A. Then there exists
a unique up to isomorphism C*-dynamical system (B, G, β), where β is an
action of G by automorphisms of a C*-algebra B and an embedding i : A ֒→
B such that
(1) β dilates α, i.e. βs ◦ i = i ◦ αvs for all s ∈ S,
(2) ∪s∈Sβ−1
(3) A ⋊α S∗ is isomorphic to i(1)(B ⋊β G)i(1), which is a full corner.
s (i(A)) is dense in B,
Proof. Use Theorem 7.15 and apply Theorem 2.4 in [12].
(cid:3)
30
MARAT AUKHADIEV
7.19. The crossed product with a non-unital C*-algebra is defined using
the multiplier algebra (see [13] for details), but the notion of the covari-
ant representation is the same as in Definition 7.7. For an extendible *-
homomorphism between C*-algebras φ : A → M (B), its unique strictly con-
tinuous extension is denoted φ : M (A) → M (B).
Definition 7.20. Let (α, S, A) be a C*-dynamical system, where A is non-
unital and S is a semigroup. A crossed product for (α, S, A) is a C*-algebra
B denoted A⋊αS with a proper homomorphism iA : A → B and a semigroup
homomorphism iS : S → Isom(M (B)) such that
(1) (iA, iS ) is a covariant representation for (α, S, A),
(2) for any other covariant representation (π, T ) there is a non-degene-
rate representation π × T of A ⋊α S such that
π = (π × T ) ◦ iA and T = (π × T ) ◦ iS,
(3) A ⋊α S is generated by {iA(a)iS (s) a ∈ A, s ∈ S} as a C*-algebra.
Theorem 7.21. Let A be a non-unital C*-algebra and let α be an extendible
injective action of S on A, β be an extendible action of S∗ on A, where one
of them is induced by another. Then the crossed product C*-algebras A ⋊α S
and A ⋊β S∗ are isomorphic.
Proof. The action α of S on A extends to an action α on M (A). Then es =
α(s)(1M (A)) is a strict limit of α(s)(uλ) = αs(uλ) and thus it is the projection
onto Evs = αs(A). Moreover, δ is a unit in M (Evs) and the map m → esmes
implements an embedding M (Evs) ֒→ M (A). On the other hand, one can
easily verify that α(s)(M (A)) ⊂ M (Evs). Hence, α(s)(M (A)) = M (Evs)
and we also have that α(s)(M (A)) is a closed two-sided *-ideal in M (A).
By Lemma 7.5 we also obtain an action β of S∗ on M (A), which clearly
extends the action β on A.
By definition and the fact that non-degenerate representations of A are
extendible, the crossed product A ⋊α S is a closed ideal in M (A) ⋊α S
generated as a C*-algebra by {iM (A)(a)iS (s) a ∈ A, s ∈ S}. Take a, b ∈ A,
s, t ∈ S and calculate the following products inside M (A) ⋊α S.
iS(s)∗iM (A)(a) = iS (s)∗iS(s)iS (s)∗iM (A)(a) = iS(s)∗iM (A)(esa) =
iM (A)(α(s)∗(esa))iS (s)∗,
iM (A)(a)iS (s)iM (A)(b)iS (t) = iM (A)(a)iS(s)iM (A)(b)iS (s)∗iS(s)iS (t) =
iM (A)(aαs(b))iS (s)iS(t).
It follows that A⋊αS is the closure of the linear span of the set {iM (A)(a)x a ∈
A, x = iS(s1)∗iS(s2)...iS (sn), si ∈ S, n ∈ N}. As usual, we call elements of
the form iS(s1)∗iS(s2)∗...iS (sn)∗ monomials. Repeating the same reasoning
as in Theorem 7.15, for any product iM (A)(a)iS (s1)∗iS(s2)...iS (sn) we may
assume that a ∈ Es∗
1s2...sn.
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
31
In the same way the crossed product A ⋊β S∗ can be viewed as a closed
ideal in the unital crossed product M (A) ⋊β S∗. Denote by jA,x the embed-
ding of Ex in M (A) ⋊β S∗. Then A ⋊β S∗ is the closure of the linear span of
the set {jA,x(a)δx x ∈ S∗} with the algebraic structure given by Definition
7.14.
Theorem 7.15 gives an isomorphism φ between the C*-algebras M (A)⋊αS
β S∗. For any a ∈ A and a monomial x the map φ sends
and M (A) ⋊
iM (A)(a)x to jA,x(a)δx. By the Lemma 7.10, the covariant representations
of the systems (α, S, A) and (β, S∗, A) are the same. Therefore, iA(a) = 0
iff jA,1(a) = 0. Hence φ restricts to an isomorphism between A ⋊α S and
A ⋊β S∗.
(cid:3)
Corollary 7.22. Let α be an action of S on a locally compact Hausdorff
space X by continuous injective maps and β the induced action of S∗. Then
the crossed product C*-algebras C0(X)⋊αS and C0(X)⋊β S∗ are isomorphic.
Proof. The action α induces an action α on C0(X) by the formula
αs(f )(x) = f (α∗
s(x)),
where f ∈ C0(X), x ∈ Ds in the notation of Lemma 7.2. Then clearly
Es = C0(Ds) ⊂ C0(X), so αs : C0(X) → C0(Ds) and every αs is injective.
The action of S∗ given by α by Lemma 7.5 coincides with the action β
induced by β on C0(X). Therefore by Theorem 7.21 the crossed products
are isomorphic.
(cid:3)
8. Partial crossed products, Ore semigroups
The results of [16] show that a C*-algebra of an E-unitary inverse semi-
group P is isomorphic to a partial crossed product of the commutative sub-
algebra generated by idempotents in P by a partial action of G, where G is
the maximal group homomorphic image of P .
Definition 8.1. A partial action α of a group G on a set X is a pair
({Dg}g∈G, {αg}g∈G), where Dg are subsets of X and αg : Dg−1 → Dg are
bijections, satisfying for any g, h ∈ G:
(1) D1 = X, α1 = idX ,
(2) αg(Dg−1 ∩ Dh) = Dg ∩ Dgh,
(3) (αgαh)(x) = αgh(x) for all x ∈ Dh−1 ∩ Dh−1g−1.
Theorem 8.2. (D. Milan, B. Steinberg [16].) Let P be an E-unitary inverse
semigroup with idempotent set E and maximal group image G. Then there
exists a partial action of G on E such that
r (P ) ∼= C ∗(E) ⋊r G.
In a view of Corollary 3.5, we get an immediate Corollary.
C ∗(P ) ∼= C ∗(E) ⋊ G, C ∗
Corollary 8.3. If a semigroup S is embeddable in a group, C ∗(SF ) ∼=
C ∗(EF ) ⋊ G, C ∗(S∗) ∼= C ∗(E) ⋊ G, where EF and E are subsemigroups
32
MARAT AUKHADIEV
of idempotents in SF and S∗ correspondingly, and G is the group gener-
ated by S. The same holds for the reduced C*-algebras and reduced crossed
products.
The models for SF and S∗ described in Theorem 3.3 and Corollary 3.4
give us concrete formulas for the partial actions in Corollary above.
For every g ∈ G define using the notation of Theorem 3.3:
Dg−1 = {[A] ∈ EF : there exist g1, g2 ∈ A such that g1 ≤ g−1 ≤ g2},
or equivalently Dg−1 = {[A] : [{1, g−1}] ≤ [A]}. For [A] ∈ Dg−1 define
αg([A]) = g[A] = [{1} ∪ {gh : h ∈ A}]
One can easily see that αg is a bijection between Dg−1 and Dg−1 and the
formula 3) from the Definition is satisfied as well. Hence, α is a partial
action of G on EF , and it coincides with the partial action in Theorem 8.2.
Consider S∗ and its idempotent subsemigroup E. Now [A] denotes an
equivalence class defined in Corollary 3.4. For every g ∈ G we define
Dg−1 = {[A] ∈ E : g ∈ A · S−1},
where A · S−1 denotes pointwise product of sets and S−1 = {a−1 : a ∈ S} ⊂
G. For [A] ∈ Dg−1 define
αg([A]) = g[A] = [{1} ∪ {gh : h ∈ A}]
Again, α is a partial action of G on S∗, which gives an isomorphism C ∗(S∗) ∼=
C ∗(E) ⋊ G.
8.4. For a particular class of semigroups we can say more about the con-
nection between S∗ and partial crossed products of groups.
Definition 8.5. A partial action α of a group G on a C*-algebra A is a
pair ({Eg}g∈G, {αg}g∈G), where Eg are closed two-sided *-ideals in A and
αg : Eg−1 → Eg are *-isomorphisms, satisfying for any g, h ∈ G:
(1) E1 = A,
(2) αg(Eg−1 ∩ Eh) = Eg ∩ Egh,
(3) (αgαh)(x) = αgh(x) for all x ∈ Eh−1 ∩ Eh−1g−1.
Then (A, G, α) is a C*-partial dynamical system.
In [10] it was shown, that the partial actions and partial representations of
a group G are in one-to-one correspondence with actions and representations
of a special inverse semigroup S(G). Moreover, an isomorphism between a
partial crossed product by G and a crossed product by S(G) was proved in
[10] and earlier in [30]. We recall these results.
Following [10] S(G) is a semigroup generated by elements tg for all g ∈ G
satisfying the following relations:
(8.1)
tg−1tgth = tg−1tgh
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
33
(8.2)
(8.3)
tgthth−1 = tghth−1
tgt1 = tg
Then S(G) is an inverse semigroup with unit t1 and involution t∗
g = tg−1.
A partial representation of G on a Hilbert space H is a map T : G → B(H),
sending g → Tg, where Tg satisfy relations (8.1) -- (8.3).
Lemma 8.6. ([10]) Partial actions (partial representations) of G are in
one-to-one correspondence with actions (resp. *-representations) of S(G).
Theorem 8.7. ([10]). Let α : G → I(A) be a partial action of a group G
on a C*-algebra A, and β the action of S(G) induced by α. Then A ⋊α G
and A ⋊β S(G) are isomorphic.
Let us consider a particular case of a group. Namely, let S be a left Ore
semigroup, so that by Theorem 2.5 of Ore and Dubreil there exists a group
G such that G = S−1S. Now we study the connections between S(G) and
the inverse semigroup S∗ generated by S as defined in Section 2.
In S(G) all elements are partial isometries, including ts for all s ∈ S.
Moreover, S(G) is generated not only by elements corresponding to S. It
follows that S(G) is not isomorphic to SF or to S∗. We implement the choice
of generators by setting which elements should be represented by isometries.
Namely, define on S(G) a relation t∗
sts ∼ 1 for all s ∈ S and denote by R
the generated congruence.
Lemma 8.8. The quotient semigroup of S(G) by the congruence R is iso-
morphic to S∗.
Proof. For all g ∈ G we denote by tg the image of tg under the quotient
map S(G) → S(G)/R. Then S(G) is characterised by R and the conditions
(8.1) -- (8.3). For any s, p ∈ S set g = s−1, h = p and consider equation (8.1):
stp = tsts−1p
tst∗
s we obtain ts−1p = t∗
Multiplying from the left by t∗
In the same way
setting g = s, h = p in (8.2) we get tsp = tstp. Since a quotient of an
inverse semigroup is an inverse semigroup, we deduce that S(G)/R satisfies
the definition of S∗ having ts at the place of vs. The converse is also true, let
us show for instance (8.1) for the generators of S∗. Take arbitrary g = p−1q,
h = s−1r where p, q, r, s ∈ S. Let qs−1 = a−1b for some a, b ∈ S, then
gh = p−1qs−1r = p−1a−1br and we have
stp.
vavq = vbvs =⇒ v∗
avb
s = v∗
q v∗
Then the left hand side of (8.1) equals
q vp)(v∗
(v∗
= v∗
q vqv∗
s vr) = v∗
pvq)(v∗
q vpv∗
pv∗
avbvr = (v∗
q vpv∗
Therefore, the map ts−1p → v∗
s vp is an isomorphism between S(G)/R and
S∗.
(cid:3)
pvqv∗
q vp)(v∗
avbvr =
avbvr)
q v∗
pv∗
34
MARAT AUKHADIEV
Combining this result with Proposition 6.3 and Theorem 4.2 (presented
here as Lemma 8.6) in [9] and Lemmas 4.3, 7.2, 7.5, we immediately get the
following.
Corollary 8.9. Any isometric inverse representation of S induces a partial
representation of G. Any injective action of S on a space X induces a partial
action of G on X. Any injective action of S on a C*-algebra A with ideal
images induces a partial action of G on A.
Remark 8.10. The reverse statement is true under some conditions. If a
*-representation T of S(G) on some Hilbert space H factors through the
quotient map S(G) → S(G)/R, then clearly T induces a *-representation of
S∗, which we denote by the same symbol. It follows that any partial repre-
sentation T of G which satisfies the property that Ts are isometries for all
s ∈ S, gives a *-representation of S∗ (and an isometric inverse representation
of S).
Theorem 8.11. Let S be a left Ore semigroup, and let α be an extendible
injective action of S on a C*-algebra A, and α the induced partial action
of G on A. Then the crossed product A ⋊α S is isomorphic to the partial
crossed product A ⋊ α G.
Let (π, T ) be a covariant representation of (β, S(G), A). We have Et∗
Proof. By Theorems 7.15 and 7.21, the crossed product A⋊α S is isomorphic
to the inverse semigroup crossed product A ⋊α S∗. On the other hand, by
the Lemma 8.6 the partial action α induces an action β of S(G) on A, and
by Theorem 8.7 the crossed products A ⋊α G and A ⋊β S(G) are isomorphic.
So, it remains to prove the isomorphism between A ⋊α S∗ and A ⋊β S(G).
s =
= A due to the fact that β is induced by α. By condition (2) of
Et∗
Definition 7.9, for any s ∈ S the operator T ∗
s ts is a partial isometry
with initial and final spaces equal to H, hence a unitary; and at the same
time it is a projection. It implies that Tts is an isometry on H. By Remark
8.10, the representation T : S(G) → B(H) factors through the quotient map
S(G) → S∗ defined in Lemma 8.8 and gives a representation T of S∗ on H.
So, we obtain a covariant representation (π, T ) of (α, S∗, A).
s
tsTts = Tt∗
If (π, T ) is a covariant representation of (α, S∗, A), then Corollary 8.9
gives a covariant representation of (β, S(G), A). Therefore these dynamical
systems have the same set of covariant representations. The underlying
algebra L for the two crossed products are different, but the completions
under the supremum norm over all covariant representations are isomorphic.
Indeed, if s, t ∈ S(G) are such that s ∼R t, then for any a ∈ Es = Et and
for any covariant representation (π, T ) we have
π × T (aδs) = π(a)Ts = π(a)Tt = π × T (aδt).
Thus, A ⋊α S∗ and A ⋊β S(G) are isomorphic.
(cid:3)
C*-THEORY OF THE UNIVERSAL INVERSE SEMIGROUPS
35
9. Conclusion
One can see that the semigroup C*-algebras (both universal and reduced)
of a cancellative semigroup S are in fact C*-algebras (the full C*-algebra
and a C*-algebra generated by some special representation) of some inverse
semigroup. All phenomena of these algebras, discussed in the Introduction,
can be explained by this fact, and indicate that the concept of these algebras
is imperfect. We have shown that to every left cancellative semigroup S one
can associate a universal inverse semigroup S∗. Then the full and reduced
C*-algebras of S∗ do not have the mentioned problems and can be regarded
as "new" C*-algebras of S. The universal inverse semigroup captures many
properties of S, of its "old" C*-algebras and also of actions and crossed
products by S. All together this convinces us that S∗ serves the purpose of
describing the C*-theory of S.
Acknowledgements. The author is thankful to Prof. S. Echterhoff for
many useful discussions and to Dr. H. Thiel for pointing at mistakes. The
research was supported by the Alexander von Humboldt Foundation. A part
of this research was supported by the Russian Foundation of Basic Research
(RFBR), grant 14-01-31358.
References
[1] M. Aukhadiev, Yu. Kuznetsova, Quantum semigroups generated by locally compact
semigroups, arXiv:1504.00407v3, 2015.
[2] M. Aukhadiev, V. Tepoyan, Isometric representations of totally ordered semigroups.
Lobachevskii J. Math. 33 (2012), no. 3, 239 -- 243.
[3] T. O. Banakh, I. Yo. Guran, O. V. Gutik, Free topological inverse semigroups, Matem-
atychni Studii, 15 (2001), 23 -- 43.
[4] N. P. Brown and N. Ozawa, C*-algebras and Finite-dimensional approximations, Grad-
uate Studies in Mathematics, vol. 88, American Mathematical Society, Providence, RI,
2008.
[5] A. Cherubini, M. Petrich, The inverse hull of right cancellative semigroups, Journal of
Algebra 111 (1987), 74 -- 113.
[6] P. Dubreil, Sur les probl`emes d'immersion et la th´eorie des modules, C. R. Acad. Sci.
Paris 216 (1943), 625 -- 627.
[7] J. Duncan, A. L. T. Paterson, C*-algebras of inverse semigroups. Proc. Edinburgh
Math. Soc. (2) 28 (1985), no. 1, 4158.
[8] J. Duncan, I. Namioka, Amenability of inverse semigroups and their semigroup alge-
bras. Proc. Roy. Soc. Edinburgh Sect. A 80 (1978), no. 3 -- 4, 309 -- 321.
[9] R. Exel, Partial actions of groups and actions of inverse semigroups, Proc.AMS, 126
(1998), no. 12, 3481 -- 3494.
[10] R. Exel, F. Vieira, Actions of inverse semigroups arising from partial actions of
groups, J. Math. Anal. Appl., 363 (2010), 86 -- 96.
[11] S. A. Grigoryan, V. H. Tepoyan. On isometric representations of the perforated semi-
group, Lobachevskii Journal of Mathematics, January 2013, Volume 34, Issue 1, pp
85 -- 88
[12] M. Laca. From endomorphisms to automorphisms and back: dilations and full corners,
J. London Math. Soc. (2) 61 (2000), 893 -- 904
[13] N. Larsen, Non-unital semigroup crossed products, Math. Proc. R. Ir. Acad. 100A
(2000), no. 2, 205 -- 218
36
MARAT AUKHADIEV
[14] X. Li, Semigroup C*-algebras and amenability of semigroups, J. Funct. Anal. 262
(2012), no. 10, 4302 -- 4340.
[15] X. Li, Nuclearity of cemigroup C*-algebras and the connection to amenability,
arXiv:1203.0021
[16] D. Milan, B. Steinberg, On inverse semigroup C*-algebras and crossed products,
Groups, Geometry and Dynamics, Vol.8 (2014), no. 2, 485 -- 512.
[17] D. Milan, C*-algebras of inverse semigroups: amenability and weak containment. J.
Operator Theory, 63(2):317 -- 332, 2010.
[18] S. M. Lalonde, D. Milan, Amenability and Uniqueness for Groupoids Associated with
Inverse Semigroups, arXiv:1511.01517.
[19] Murphy, Gerard J. C*-algebras generated by commuting isometries. Rocky Mountain
J. Math. 26 (1996), no. 1, 237267.
[20] Alexandru Nica. C*-algebras generated by isometries and Wiener-Hopf operators. J.
Operator Theory, 27(1):1752, 1992.
[21] Norling, Magnus Dahler, Inverse semigroup C*-algebras associated with left cancella-
tive semigroups. Proc. Edinb. Math. Soc. (2) 57 (2014), no. 2, 533564.
[22] O. Ore, Linear equations in non-commutative fields, Ann. Math. (2) 32 (1931), 463 --
477.
[23] Alan L.T. Paterson. Groupoids, Inverse Semigroups, and their Operator Algebras,
Birkhauser, 1998.
[24] Paterson, Alan L. T., Weak containment and Clifford semigroups. Proc. Roy. Soc.
Edinburgh Sect. A 81 (1978), no. 1-2, 2330.
[25] Alan L.T. Paterson. Amenability, Mathematical Surveys and Monographs, vol. 29,
American Mathematical Society, Providence, RI, 1988.
[26] Popov, Alexey I.; Radjavi, Heydar Semigroups of partial isometries. Semigroup Forum
87 (2013), no. 3, 663678
[27] G. B. Preston, Free inverse semigroups. Collection of articles dedicated to the memory
of Hanna Neumann, IV. J. Austral. Math. Soc. 16 (1973), 443-453.
[28] B. Schein, Subsemigroups of inverse semigroups, Le Matematiche, vol. LI (1996)-
Supplemento, pp. 205 -- 227
[29] B. Shain, Transformative semigroups of transformations (Russian), Mat. Sb. (N.S.),
71(113):1 (1966), 6582
[30] N. Sieben, C*-crossed products by partial actions and actions of inverse semigroups,
J. Austral. Math. Soc. (Series A) 63 (1997), 32 -- 46
[31] Vagner, V. V. Generalized groups. (Russian) Doklady Akad. Nauk SSSR (N.S.) 84,
(1952). 11191122.
[32] Wordingham, J. R. The left regular *-representation of an inverse semigroup. Proc.
Amer. Math. Soc. 86 (1982), no. 1, 5558.
WWU Munster, Einsteinstr. 62, 48149 Munster, Germany.
Kazan State Power Engineering University, Krasnoselskaya 51, 420066 Kazan,
Russia.
[email protected]
|
1812.08123 | 2 | 1812 | 2019-11-05T15:14:37 | Roots of Completely Positive Maps | [
"math.OA",
"math.PR",
"quant-ph"
] | We introduce the concept of completely positive roots of completely positive maps on operator algebras. We do this in different forms: as asymptotic roots, proper discrete roots and as continuous one-parameter semigroups of roots. We present structural and general existence and non-existence results, some special examples in settings where we understand the situation better, and several challenging open problems. Our study is closely related to Elfving's embedding problem in classical probability and the divisibility problem of quantum channels. | math.OA | math |
ROOTS OF COMPLETELY POSITIVE MAPS
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
Abstract. We introduce the concept of completely positive roots of completely positive
maps on operator algebras. We do this in different forms: as asymptotic roots, proper
discrete roots and as continuous one-parameter semigroups of roots. We present struc-
tural and general existence and non-existence results, some special examples in settings
where we understand the situation better, and several challenging open problems. Our
study is closely related to Elfving's embedding problem in classical probability and the
divisibility problem of quantum channels.
1. Introduction
In many mathematical settings, the concept of a square-root or higher order roots is
familiar, e.g. in the context of real numbers, in the context of matrices, in the context
of real-valued functions, or measures. What all of these settings have in common is the
underlying structure of a semigroup. To be slightly more formal, given a semigroup (A, ⋆)
and given a ∈ A and n ∈ N, we can ask whether there exists some x ∈ A such that
a = x ⋆ x ⋆ · · · ⋆ x (x appearing n times). Then we may call x an n-th root of a. If such
a root exists for all n, we would call a infinitely divisible. We may also ask whether there
is a one-parameter semigroup (xt)t∈R+ in A (namely xs+t = xs ⋆ xt, for all s, t ∈ R+),
such that xt0 = a for some t0 > 0. If this is the case then a may be called embeddable
(into a continuous semigroup). Finally, if there is a topology on A, we may also look for
asymptotic roots or asymptotic embeddability, that is, whether there is a one parameter
semigroup (xt)t∈R+ , with limt→∞ xt = a.
Yuan [30] deals with some of these questions in the full generality of topological semi-
In the
groups, but without further structure it seems that one cannot say too much.
present paper we would like to add such structure and look at unital normal completely
positive (UNCP) maps on von Neumann algebras. They arise in many ways in operator
algebras and quantum physics and are natural objects to study (cf. [14, 23, 27]). Since
'composition' is an associative operation on UNCP maps, it makes sense to study the
question of roots also in this setting, namely: given a von Neumann algebra A, a number
n ∈ N, and a UNCP map φ : A → A, is there another UNCP map ψ : A → A such that
φ = ψn? It turns out that currently surprisingly little is known in general.
However, there are a number of connection points with results in related areas. For
example, we could specialize to commutative algebras. Then UNCP maps become sto-
chastic maps of Markov chains in classical probability theory. A notable special case is
that of (discrete or continuous) convolution semigroups of probability measures. Exis-
tence and uniqueness criteria for roots of stochastic maps have been studied earlier (see
e.g. [17, 16]). Suppose a given Markov chain on a countable state space converges to an
invariant distribution (an absorbing state). Then typically such a convergence happens
exponentially (i.e., asymptotically) over time. In discrete time there are some instances
when the convergence takes place in finite time. An analysis of transition probability ma-
trices of such Markov chains can be seen in [5] and [24]. This does not work in continuous
time if the semigroup generator is bounded, [15]. But the condition of boundedness of the
generator might be too strong, and it is widely believed that under some very minimal
Date: 5 November 2019.
1991 Mathematics Subject Classification. primary: 46L57; secondary: 60J10, 81P45.
Key words and phrases. complete positivity; divisibility; Markov chains; matrix algebras; operator
algebras; quantum information.
1
2
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
continuity assumptions on the transition semigroup, convergence in finite time should be
impossible. However, we are not aware of any proof. Surprisingly the noncommutative
counterpart is more involved and convergence in finite continuous time to a given pure
state is indeed possible as has been shown in [3]; more precisely, for a given normal pure
state ϕ on B(H), identified with the completely positive map φ = ϕ(·)1, it is possible
to construct a quantum Markov semigroup (τt)t∈R+ (a strongly continuous one-parameter
semigroup of UNCP maps) on B(H) which coincides with φ at all times t ≥ 1; in other
words, for all states ψ, we get ψ ◦ τt = ϕ, for t ≥ 1. Hence convergence in the continuous
setting is possible in finite time, for all pure states on B(H). A trivial consequence is that
φ has an n-th root, for every n ∈ N. It is natural to ask what happens in the case of
φ = ϕ(·)1 where ϕ is a mixed state. What can be said about n-th roots or semigroups of
roots of φ? And in light of our first observation, what can be said about n-th roots and
semigroups of roots of more general completely positive maps φ, not only those arising
from states?
if S is infinitely divisible, i.e.
If we drop the assumption of convergence to an invariant distribution, many things
can happen. E.g. the question of continuous roots of a given stochastic map makes sense
here and is known under the name of Elfving's embedding problem, dating back to 1937
[13]: given a stochastic map S, when can we find a map L such that eL = S and all etL,
for t ≥ 0, are stochastic maps? A number of necessary and sufficient criteria have been
found over the years, see e.g. [8, 10, 18, 26] for a non-exhaustive list. One particularly
interesting condition is:
it has n-th roots for all n, then
it is embeddable into a semigroup [18]. The noncommutative analogue of this question
is not so easy but we may restrict ourselves to finite dimensions to start with. Indeed it
should be noted that completely positive (trace-preserving) maps in finite dimensions form
the basis of quantum information theory (there termed "quantum channels" [27]). The
question of asymptotic behavior of sequences of compositions of quantum channels appears
relevant in quantum information problems, e.g. in the context of entanglement breaking
maps [21], and the question of "divisibility" of quantum channels has also been studied in
a few places. A number of divisibility criteria can be found in [11, 28, 29, 2] and some of
the questions we pose here have also been discussed in [28, 2] but with slightly different
terminology and complementary answers. Most notably though, it has been shown that
the complexity of the problems of deciding whether a given quantum channel (or stochastic
map) has a square-root or whether it is embeddable into a continuous semigroup are NP-
hard [6, 7, 2]. This means that the set of such quantum channels has no simple expression
other than explicit enumeration of its elements, and it is impossible to find a simply
verifiable criterion for the existence of such roots as in the case of positive numbers or
matrices, for example. However, this should not discourage us from looking for interesting
new relations or characterizations, at least for some special classes of UNCP maps, and
that is what we would like to do here.
Our outline is as follows. In Section 2, we start by describing the quantum analogue and
generalization of the "exponential" convergence to a given invariant distribution: given a
UNCP map, is there a continuous one-parameter semigroup that converges to φ as t → ∞?
We completely clarify this question. Such semigroups we will call asymptotic continuous
roots. As a byproduct we obtain an affirmative answer to a question of Arveson (Problem
3 in [1, p.387]) through very elementary methods.
We then move on to the question of proper roots in the finite-time setting, where
Section 3 deals with the n-th root case while Section 4 deals with the continuous semigroup
case. We are able to provide several existence and non-existence results under different
additional assumptions, e.g. regarding the dimension or structure of the algebra or the
range of the CP map. In particular, for the case of states on Md or B(H) or Cd we have a
complete characterization of existence of n-th roots. However, we are still far from a full
understanding and have to leave some questions in the form of conjectures and difficult
open problems. We hope some readers will feel stimulated to think about them and come
up with nice solutions. Our diversity of results together with the few related results in
ROOTS OF COMPLETELY POSITIVE MAPS
3
literature, e.g. [16, 6, 7, 2], indicate that a "complete and elegant" characterization is
unlikely to be found though.
In this paper we are mainly concerned with the existence or non-existence of roots
of UNCP maps. Whenever such roots exist, they are typically far from unique, and a
subsequent natural question would be to find a useful characterization of all such roots
for a given UNCP map. We deal with UNCP maps (in finite and infinite dimension) and
to some minor extent with the commutative special case of stochastic maps though we do
not look at the related question of nonnegative roots of (entry-wise) nonnegative matrices
as can be found in other places, e.g. [19].
Throughout this paper, all our C*-algebras or von Neumann algebras are supposed
to live on some Hilbert space, and all Hilbert spaces here are taken as complex and
separable, sometimes even finite-dimensional, with scalar products linear in the second
variable. B(H) will denote the bounded linear operators on the Hilbert space H and Md
the d × d matrices with complex entries; N denotes the natural numbers without 0. For
an introduction to completely positive maps we refer the reader to any good textbook on
the topic, e.g. [14, 23, 25].
2. Asymptotic roots
In the present section we work in the C*-algebraic setting because it appeared more
natural to us; however, everything can be adjusted and translated in a straight-forward
way to the von Neumann algebraic setting, cf. also Remark 2.3 below, which would also
bring it more in line with the subsequent sections.
Definition 2.1. Given a unital C*-algebra A and a bounded unital completely positive
(UCP) map φ : A → A,
(ad) an asymptotic discrete root of φ is a UCP map τ : A → A such that τ n → φ
(pointwise in norm), as n → ∞;
(ac) an asymptotic continuous root of φ is a uniformly continuous one-parameter semi-
group (τt)t≥0 of UCP maps on A such that τt → φ (pointwise in norm), as t → ∞.
We then have:
Theorem 2.2. Let A be a unital C*-algebra and φ a UCP map of A. Then the following
three statements are equivalent:
(i) φ is idempotent, i.e., φ2 = φ;
(ii) φ has an asymptotic continuous root;
(iii) φ has an asymptotic discrete root.
Proof. (i) ⇒ (ii). Suppose φ2 = φ. Then define the map
L := φ − id = −(id −φ) : A → A,
which is bounded and conditionally completely positive [14, Sec.4.5] and therefore gener-
ates a uniformly continuous UCP semigroup τ . We find
Ln(x) = (−1)n(id −φ)n(x) = (−1)n(id −φ)(x),
x ∈ A,
and therefore
∞
τt(x) =
We see
(−t)n(id −φ)n
n!
x =
(−t)n
n!
∞
Xn=0
(id −φ)(x) + φ(x) = e−t(id −φ)(x) + φ(x).
Xn=0
x ∈ A,
so τt → φ uniformly, as t → ∞, so τ is an asymptotic continuous root for φ.
kτt(x) − φ(x)k = e−t kx − φ(x)k ≤ e−t(1 + kφk)kxk,
(1)
(ii) ⇒ (iii) is obvious.
(iii) ⇒ (i). Suppose now that there is an asymptotic discrete root τ of φ. Then the fact
that τ m → φ as m → ∞ allows us to make the following manipulations:
φ ◦ τ n(x) = lim
m→∞
τ m ◦ τ n(x) = lim
n+m→∞
τ n+m(x) = φ(x),
x ∈ A,
4
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
and therefore
so φ2 = φ.
φ2(x) = lim
m→∞
φ ◦ τ m(x) = lim
m→∞
φ(x) = φ(x),
x ∈ A,
(cid:3)
Remark 2.3. Let A be a unital C*-algebra and φ a UCP map of A.
(i) An asymptotic root of φ is in general not unique.
(ii) We did not specify the dimension of A and the Hilbert space H on which it acts.
In fact, the statements are interesting in both finite and infinite dimensions.
(iii) If φ has an asymptotic (discrete/continuous) root with respect to the strong op-
erator topology then the above proof shows that φ also has an asymptotic (dis-
crete/continuous) root with respect to the uniform topology.
(iv) The definition, theorem and proof continue to hold true upon replacing C*-algebras
by von Neumann algebras, UCP by UNCP, and the uniform by the strong operator
topology.
Remark 2.4. As a byproduct, the theorem answers Problem 3 in [1, p.387] affirmatively,
namely given an eigenvalue list (λ1, λ2, . . .) with 0 ≤ λi ≤ 1 and Pi λi = 1 as in [1,
Sec.12.4], consider the density matrix
D := diag(λ1, λ2, . . .) ∈ B(H),
the normal state ϕ = tr(D·) on B(H), and the UNCP map
φ = ϕ(·)1 : B(H) → B(H).
Then the asymptotic root in the above proof is a UNCP semigroup with bounded generator
that has ϕ as absorbing state: for every normal state ψ on B(H) and every x ∈ B(H), we
get
ψ ◦ τt(x) − ϕ(x) = ψ(τt(x) − ϕ(x)1) ≤ kτt(x) − ϕ(x)1k ≤ 2 e−t kxk
where the last inequality follows from (1); thus,
meaning that ϕ is an absorbing state for (τt)t≥0, which answers Problem 3 in [1, p.387].
kψ ◦ τt − ϕk → 0,
t → ∞,
3. Proper discrete roots
In this and the following section, we work exclusively with von Neumann algebras.
3.1. General statements. Our fundamental definition is the following:
Definition 3.1. Given a von Neumann algebra A, a UNCP map φ : A → A and an
integer n ∈ N \ {1}, a proper n-th discrete root of φ is a UNCP map τ : A → A such that
τ n = φ and τ k 6= φ for all k < n. We call n the order of τ .
We need a notational convenience which turns out to be crucial in many proofs and
characterizations:
Definition 3.2. For every UNCP map φ on a von Neumann algebra A, we define the
support projection as the smallest projection pφ ∈ A such that φ(pφ) = 1. We write
p′
φ := 1 − pφ ∈ A.
The existence and the uniqueness of pφ follow from [12, Prop.I.4.3, p.63], roughly as
follows: one first realizes that the set of x ∈ A such that φ(x∗x) = 0 forms a σ-weakly
closed left ideal in A. For such ideals there exists a maximal projection p such that the
ideal consists of all x ∈ A with x = xp. This is exactly the projection p = p′
φ = 1 − pφ
from the preceding definition.
We use the following block matrix decomposition of x ∈ A:
x =(cid:18)x11 x12
x21 x22(cid:19) =(cid:18)pφxpφ pφxp′
φ(cid:19) .
p′
φxpφ p′
φ
φxp′
(2)
ROOTS OF COMPLETELY POSITIVE MAPS
5
A first useful fact is the following variation of [4, Th.4.2] about the relation with nilpo-
tent NCP maps:
Lemma 3.3. Let A be a von Neumann algebra, φ a UNCP map of A and n ∈ N. Suppose
there exists a proper n-th discrete root τ of φ. Then
(i) τ (pφ) ≥ pφ;
(ii) there also exists a nilpotent NCP map α : p′
φAp′
φ → p′
φAp′
φ of order at most n such
that
(iii) for every (cid:18)0 x
y
A such that
0
0 α(x)(cid:19) ,
τ(cid:18)0 0
0 x(cid:19) =(cid:18)0
z(cid:19) ∈ A w.r.t. to the above block decomposition, there is (cid:18) 0 x′
φAp′
φ;
x ∈ p′
y′
z′(cid:19) ∈
y z(cid:19) =(cid:18) 0 x′
τ(cid:18)0 x
z′(cid:19) ;
y′
(iv) pφτ (·)pφ restricts to a proper discrete root of pφφ(·)pφ on pφApφ of order at most
n.
Proof. (i). We first notice that τ (pφ) ≤ 1 since τ was assumed to be UNCP. Therefore
0 ≤ pφτ (pφ)pφ ≤ pφ. Let us write b := pφ − pφτ (pφ)pφ ≥ 0. We would like to show that
b = 0.
To start with,
This implies
τ ◦ φ = τ ◦ τ n = τ n+1 = τ n ◦ τ = φ ◦ τ.
φ(pφ) = 1 = τ (1) = τ (φ(pφ)) = φ(τ (pφ)) = φ(pφτ (pφ)pφ),
thus
φ(b) = φ(pφ − pφτ (pφ)pφ) = 0.
Let eb ∈ A be the support projection of b, which can be defined through Borel functional
calculus. Notice that eb ≤ pφ because b ≤ pφ, so pφ − eb is a subprojection of pφ. Then
it follows from the construction of the spectral theorem (in projection-valued measures
form [22, Sec.7.3]) that φ(b) = 0 if and only if φ(eb) = 0. Since we have already proved
φ(b) = 0, we find
φ(pφ − eb) = φ(pφ) − φ(eb) = 1 − 0 = 1.
Thus pφ − eb fulfills the properties of a support projection of φ and therefore must be equal
to pφ due to its uniqueness, so eb = 0, hence b = 0.
(ii). Unitality of τ together with part (i) implies τ (p′
φ) ≤ p′
φ, hence giving rise to an NCP map
map with image in p′
φAp′
φ. Thus, τ (p′
φ · p′
φ) is a NCP
Since τ is an n-th root of φ, we have
α := τ ↾p′
φAp′
φ
: p′
φAp′
φ → p′
φAp′
φ.
0 = φ(cid:18)0 0
0 x(cid:19) = τ n(cid:18)0 0
0 x(cid:19) =(cid:18)0
0 αn(x)(cid:19) ,
0
implying that α is nilpotent of order at most n.
x ∈ p′
φAp′
φ,
(iii) Part (i) shows that τ (p′
φ) ≤ p′
φ. Using the block decomposition in (2), we can write
(cid:18)0 0
0 1(cid:19) ≥ τ(cid:18)0 0
for every x ∈ pφAp′
φ with x∗x ≤ p′
0
x∗ 0(cid:19)(cid:18)0 x
0 0(cid:19)∗
0 1(cid:19) ≥ τh(cid:18) 0
0 0(cid:19)i ≥ τ(cid:18)0 x
0 0(cid:19) =(cid:18)0 x′
τ(cid:18)0 x
0 z′(cid:19)
φ. This means that
τ(cid:18)0 x
0 0(cid:19) ,
6
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
with certain x′ ∈ pφAp′
τ , we have, for any x ∈ pφAp′
φ, z′ ∈ p′
φAp′
φ, y ∈ p′
φ. Together with part (ii) and the self-adjointness of
φApφ, z ∈ p′
φAp′
φ:
τ(cid:18)0 x
z(cid:19) =(cid:18) 0 x′
z′(cid:19)
y′
y
φ, y′ ∈ p′
φApφ, z′ ∈ p′
φAp′
φ.
with certain x′ ∈ pφAp′
(iv) Since pφ ∈ A and pφφ(pφ)pφ = pφ = pφτ (pφ)pφ by part (i), it is clear that both
pφφ(·)pφ and pφτ (·)pφ restrict to UNCP maps on pφApφ. Moreover, it follows from part
(iii) that
pφτ (pφτ (·)pφ)pφ = pφτ 2(·)pφ,
and by induction, since τ n = φ, we get that pφτ (·)pφ is a proper discrete root of pφφ(·)pφ
of order at most n.
(cid:3)
If φ is idempotent then there is generally more hope to say something about roots. A
particularly nice case of idempotency is that where φ has rank one, namely φ = ϕ(·)1 for
some normal state ϕ on A. In that case, we get the following easy correspondence:
Lemma 3.4. Given a von Neumann algebra A, a normal state ϕ on A and n ∈ N, let
φ = ϕ(·)1, which is UNCP. Then a map τ on A is a proper n-th discrete root of φ if and
only if τ = φ + α with α some normal nilpotent map of order n such that α ◦ φ = 0 = φ ◦ α
and φ + α is CP.
Proof. (⇒) Consider α := τ − φ. Clearly α is normal. Since φ ◦ τ = τ ◦ φ = τ n+1 = φ
we have, for all k ≥ 1 αk = τ k − φ. Now since τ k 6= φ for k < n and τ n = φ we have
αk 6= 0 for k < n and αn = 0. i.e., α is nilpotent of order n. Also φ ◦ α = φ ◦ (τ − φ) = 0 =
(τ − φ) ◦ φ = α ◦ φ. The converse part (⇐) is trivial.
(cid:3)
When p := dim A < ∞, Lemma 3.4 shows that any UNCP map arising from a state on
A cannot have proper discrete roots of order higher than p. Indeed the following lemma
shows that the order of such a root must be strictly less than p:
Lemma 3.5. Let A be a finite dimensional von Neumann algebra of dimension p. Let τ
be a UNCP map on A. Then the following are equivalent:
(i) τ n = φ := ϕ(·)1 for some state ϕ on A and for some n ∈ N.
(ii) τ = φ + α for some nilpotent map α and φ := ϕ(·)1 for some state ϕ with α ◦ φ =
0 = φ ◦ α.
(iii) 0 is an eigenvalue of τ with algebraic multiplicity p − 1.
(iv) tr τ k = 1 for all k ≥ 1.
In any of these equivalent cases, τ is a root of order at most p − 1.
Proof. The idea of the proof is to treat φ and τ as linear maps on Cp.
(i) ⇔ (ii) follows from Lemma 3.4.
(i) ⇒ (iii). As τ n = φ has rank 1, 0 is an eigenvalue of τ n of multiplicity p − 1, hence 0
is an eigenvalue of τ of multiplicity p − 1.
(iii) ⇒ (i). Looking at the Jordan normal form of τ it is clear that τ n has rank 1 for
some n ∈ N. Since τ n is unital, there is a state ϕ on A such that τ n(x) = ϕ(x)1 for all
x ∈ A.
(iii) ⇒ (iv) is obvious as τ (1) = 1.
(iv) ⇒ (iii). Let λ1 = 1, λ2, λ3, ..., λm be the distinct eigenvalues of τ with algebraic mul-
tiplicity a1, a2, ..., am, respectively. From (iv) we have (a1−1)λk
m = 0 for
all k ≥ 1. Consider the Vandermonde matrix V := (λj−1
)1≤i,j≤n ∈ Mm . Then as the λi's
are mutually distinct, we have det V 6= 0. Also note that V ((a1−1)λ1, a2λ2, ..., amλm)′ = 0.
This implies that ((a1 − 1)λ1, a2λ2, ..., amλm) = 0. Hence a1 = 1, m = 2 and λ2 = 0. That
means 0 is an eigenvalue of τ with algebraic multiplicity p − 1.
2 +· · ·+amλk
1 +a2λk
i
Now regarding our final statement, let τ be a proper n-th discrete root of φ = ϕ(·)1 on
A. It is clear from (iii) ⇔ (i) that n is the maximal possible size of all Jordan blocks of τ .
Hence n ≤ p − 1.
(cid:3)
ROOTS OF COMPLETELY POSITIVE MAPS
7
Remark 3.6. It is worth pointing out that a proper n-th discrete root τ for a state ϕ
is "absorbing", namely ψ ◦ τ k = ϕ, for all k ≥ n and all other states ψ. So in this case
τ is also an asymptotic discrete root. The same is true for proper versus asymptotic
continuous roots, as shall become clear from the following section, cf. Proposition 4.5. In
general though, there is no clear relationship between proper and asymptotic roots.
Here are some examples regarding existence and non-existence of roots of UNCP maps
in finite dimensions. We start with a map which has no nontrivial proper discrete roots
at all.
Example 3.7. Let φ : M2 → M2 be the UNCP map defined by φ(cid:18)a b
c d(cid:19) =(cid:18)d 0
0 a(cid:19). We
claim that φ has no proper discrete root. Suppose for contradiction there exists a proper
n-th discrete root τ for φ, then τ n = φ and τ ◦ φ = τ ◦ τ n = τ n+1 = τ n ◦ τ = φ ◦ τ . Let
τ(cid:18)1 0
0 0(cid:19) =(cid:18)a11 a12
a21 a22(cid:19) ,
τ(cid:18)0 0
1 0(cid:19) =(cid:18)c11
c22(cid:19) ,
c21
c12
b12
b21
τ(cid:18)0 1
0 0(cid:19) =(cid:18)b11
b22(cid:19) ,
τ(cid:18)0 0
0 1(cid:19) =(cid:18)d11 d12
d21 d22(cid:19) .
Since τ ◦ φ = φ ◦ τ and τ (1) = 1, we have a12 = a21 = d12 = d21 = 0 and a11 = d22 and
d11 = a22 6= 0. It follows that
(cid:18)0 0
0 1(cid:19) = τ n(cid:18)1 0
(cid:18)1 0
0 0(cid:19) = τ n(cid:18)0 0
0 0(cid:19) =(cid:18)an
0 1(cid:19) =(cid:18)∗
11 + ∗ 0
0
∗(cid:19)
22 + ∗(cid:19) ,
0
0 dn
where all ∗'s are nonnegative terms depending on a11 and a22 only. In particular we see
from these equalities that a11 = d22 = 0 and the only possible solution is
τ(cid:18)1 0
0 0(cid:19) =(cid:18)0 0
0 1(cid:19) ,
τ(cid:18)0 0
0 1(cid:19) =(cid:18)1 0
0 0(cid:19) ,
i.e., τ = φ. Thus φ has no proper n-th discrete root.
The following map has only a proper square root.
Example 3.8. Let φ : M2 → M2 be the idempotent UNCP map defined by φ(cid:18)a b
(cid:18)a 0
0 d(cid:19). Then φ has a proper square root τ(cid:18)a b
c d(cid:19) =
0 a(cid:19) but φ has no other proper
c d(cid:19) =(cid:18)d 0
discrete roots, which can be proved in the same style as Example 3.7.
Finally, a map with proper discrete roots of all orders:
Example 3.9. Let φ : M2 → M2 be the UNCP map defined by
For every n ∈ N, define
τ1/n : M2 → M2,
φ(cid:18)a b
c d(cid:19) =(cid:18)a
c
2
b
2
d(cid:19) .
c d(cid:19) = a
c
1
2
n
τ1/n(cid:18)a b
b
1
2
n
d ! .
Then τ1/n is a UNCP map and τ n
1/n = φ, so τ1/n is a proper n-th discrete root of φ.
Example 3.10. Let φ : M2 → M2 be the UNCP map defined by
φ(cid:18)a b
c d(cid:19) =(cid:18)d
b
2
c
2
a(cid:19) .
Then φ has an n-th root for every odd n ∈ N \ {1} but not for even n. This is again proved
in the same way as Example 3.7.
8
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
So we are led to the following problem:
Problem 3.11. Suppose A = Md or B(H) and φ is a UNCP map on A. Then for which
n ∈ N is there a proper n-th discrete root of φ?
Though we have got some illustrative examples here, a general characterization of ex-
istence and non-existence of proper discrete roots is expected to be complicated and does
involve more details about the map φ, as the following subsection indicates. Similar facts
have been pointed out in [2] and it matches the findings in [16, Sec.4].
3.2. Proper discrete roots for states on Md and B(H). We can say much more by
specializing the results of the preceding subsection to the setting of normal states on B(H)
or Md, which we are going to do now.
Theorem 3.12. Suppose d < ∞ and ϕ is a state on Md of support dimension r :=
dim(pφCd). Then φ = ϕ(·)1 has a proper n-th discrete root on Md if and only if 1 < n ≤
d + r2 − r − 1.
λ1 ≥ . . . ≥ λr > λr+1 = 0 = . . . = λd.
Proof. We split the proof into two steps, depending on r. First of all, we may choose
k=1 λkhek, ·eki and
and fix a basis (ek)k=1,...,d such that ϕ is in diagonal form, so ϕ = Pd
(Step 1) Suppose r = d, so ϕ is faithful. We have to prove that φ has a proper n-th
discrete root if and only if 1 < n ≤ d2 − 1. First we see from Lemma 3.5 that if τ is a
proper n-th discrete root of φ then n ≤ d2 − 1. We write
α := τ − φ,
which is nilpotent of order n with α(1) = 0 = φ ◦ α owing to Lemma 3.4.
Let us introduce the scalar product
h·, ·iϕ : (x, y) ∈ Md × Md 7→ ϕ(x∗y).
Then α restricts to a linear nilpotent map from Md ⊖C1 into itself, and this subspace
has dimension d2 − 1. An upper bound on the order of nilpotency is therefore d2 − 1, so
n ≤ d2 − 1.
Next we would like show that we can actually attain this upper bound. To this end,
consider an orthonormal basis (1, Y1, . . . Yd2−1) of Md with respect to h·, ·iϕ such that
Y ∗
i = Yi and φ(Yi) = 0, for all i, which can always be achieved. Then define
α(1) = α(Yd2−1) = 0, α(Yi) = εYi+1,
i = 1, . . . , d2 − 2,
with suitable ε > 0 still to be determined, and
τ := φ + α.
Then it is clear that α is nilpotent of order d2 − 1 and so τ d2−1 = φ because φ ◦ α = α ◦ φ
but τ k 6= φ for k < d2 − 1. Moreover, α is self-adjoint, namely α(x∗) = α(x)∗ for all
x ∈ Md, thus is τ . In order to show that τ is a proper discrete root, it remains to show
that τ is CP. To this end, we compute the Choi matrix Cτ ∈ Md(Md) of τ , cf. [23], and
find
Cτ =
. . .
τ (e11)
...
τ (ed1)
. . .
τ (e1d)
...
τ (edd)
τ (e∗
11)
...
τ (e∗
1d)
. . .
. . .
τ (e∗
d1)
...
τ (e∗
dd)
τ (e11)∗
...
. . .
τ (ed1)∗
...
τ (e1d)∗
. . .
τ (edd)∗
= C ∗
τ
=
=
.
λ11
...
0
0
. . .
. . .
0
0
...
λd1
so Cτ is self-adjoint for all ε. We notice that Cτ depends continuously on ε and that for
ε = 0, we get
This matrix lies in the interior of the convex cone of positive matrices because all λi > 0.
Choosing ε > 0 small enough, we therefore find that Cτ must still be inside this cone. By
We therefore have to find the minimal number n′ such that
for all x, y, z, and we claim that it can be at most d − r.
τ n′(cid:18)1 x
z(cid:19) =(cid:18)1 0
0 1(cid:19) ,
y
To this end, let us write
N
Since
L∗
τ =
Xi=1
0 1(cid:19) =
Xi=1
N
i (·)Li, Li =(cid:18)Ai Bi
Ci Di(cid:19) .
ϕ(cid:18)C ∗
i Di(cid:19) =
i Ci C ∗
i Ci D∗
i Di
D∗
N
Xi=1
0 = ϕ ◦ τ(cid:18)0 0
(3)
ϕ ↾Mr (C ∗
i Ci)
ROOTS OF COMPLETELY POSITIVE MAPS
9
Choi's theorem, cf. [23], this implies that τ is CP, hence it is a proper discrete root of
order d2 − 1.
In order to get a proper discrete root of order n < d2 − 1, all we have to do is change
the map α accordingly, e.g
α(1) = α(Yn) = . . . = α(Yd2−1) = 0, α(Yi) = εYi+1,
i = 1, . . . , n − 1,
and proceed in the same way as above.
(Step 2) Next we examine the case r < d and write Mr for pφ Md pφ. Suppose τ is a
root of φ. Then by Lemma 3.3(iv),
τ ′ := pφτ (·)pφ : Mr → Mr
defines a proper discrete root of the faithful state ϕ ↾Mr on Mr, hence its maximal order
is r2 − 1 according to (Step 1) above. As shown in Lemma 3.3(iii), we have the following
action in block decomposition:
in particular
y
∗
τ(cid:18)w x
τ r2−1(cid:18)w x
z(cid:19) =(cid:18)τ ′(w) ∗
∗(cid:19) ,
z(cid:19) =(cid:18)ϕ ↾Mr (w)1 ∗
∗(cid:19) .
∗
y
and ϕ ↾Mr is faithful, we obtain Ci = 0 for all i. Moreover, it follows from Lemma 3.3 that
is nilpotent and CP, and it follows from [4, Cor.2.5] that the order of nilpotency
τ ↾p′
is at most dim(p′
φH) = d − r =: r′. Therefore
φ Md p′
φ
0 = τ r′(cid:18)0 0
0 1(cid:19) =
so
N
Xi1,...,ir′ =1(cid:18)0
0 D∗
ir′ · · · D∗
0
i1 Di1 · · · Dir′(cid:19) ,
for all i1, . . . , ir′ ∈ {1, . . . , N }. Moreover, unitality of τ ′ implies that
Di1 · · · Dir′ = 0,
N
A∗
i Ai = 1.
Xi=1
We have Li1 · · · Lik−1Lik =(cid:18)Ai1 · · · Aik Mi1,i2,...,ik
Di1Di2 · · · Dik(cid:19) for every k ∈ N, where
0
Mi1,i2,...,ik =Ai1 · · · Aik−1Bik + Ai1 · · · Aik−2Bik−1Dik + . . .
Now it follows from (4) and (5) that
+ Ai1Bi2Di3 · · · Dik + Bi1Di2 · · · Dik .
τ r′(cid:18)1 x
z(cid:19) =
y
N
Xi1,...,ir′ =1(cid:18)A∗
ir′ · · · A∗
i1Ai1 · · · Air′ A∗
M ∗
i1,i2,...,ir′ Ai1 · · · Air′ M ∗
ir′ · · · A∗
i1Mi1,i2,...,ir′
i1,i2,...,ir′ Mi1,i2,...,ir′(cid:19) .
(4)
(5)
10
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
Furthermore,
N
Xi0,i1,...,ir′ =1
M ∗
i0,i1,i2,...,ir′ Mi0,i1,i2,...,ir′ =
=
Similarly
N
Xi0,i1,...,ir′ =1
M ∗
i0,i1,i2,...,ir′ Ai0Ai1 · · · Air′ =
=
By induction we find that
N
Xi0,i1,...,ir′ =1
Xi1,...,ir′ =1
N
N
Xi0,i1,...,ir′ =1
Xi1,...,ir′ =1
N
M ∗
i1,i2,...,ir′ A∗
i0Ai0Mi1,i2,...,ir′
M ∗
i1,i2,...,ir′ Mi1,i2,...,ir′ .
M ∗
i1,i2,...,ir′ A∗
i0Ai0Ai1 · · · Air′
M ∗
i1,i2,...,ir′ Ai1 · · · Air′ .
τ r′+k(cid:18)1 x
z(cid:19) = τ r′(cid:18)1 x
z(cid:19) =(cid:18)1 0
0 1(cid:19) ,
y
y
∀k ∈ N,
and together with (3) we see that n′ can be at most r′, so the order of τ on Md can be at
most r2 − 1 + r′ = r2 − 1 + d − r.
It remains to show that all orders n = 2, . . . , r2 − 1 + d − r can be attained. First of all,
following the ideas in (Step 1) and given a root of order n = 1, . . . , r2 − 1 on Mr, there is
l = 1, . . . , r and w ∈ Mr such that
N
Xi1,...,in−2=1
(cid:16)
A∗
in−2 · · · A∗
i1wAi1 · · · Ain−2(cid:17)ll
6=(cid:16)
N
Xi1,...,in−1=1
A∗
in−1 · · · A∗
i1wAi1 · · · Ain−1(cid:17)ll
.
Then setting all Di = 0 and Bi = el,i for i = 1, . . . r′, we can obtain roots of orders
n + 1 = 2, . . . , r2 on Md. In order to get order n = r2 + n′, we keep Bi = el,i for i = 1, . . . r′
and choose for D1 any contractive nilpotent matrix of order n′ + 1 and all other Di = 0.
This way we achieve
N
Xi1,...,in′+1=1
M ∗
i1,i2,...,in′+1
Mi1,i2,...,in′+1 =
6=
N
Xi1,...,in′ =1
Xi1,...,in′
N
−1=1
M ∗
i1,i2,...,in′ Mi1,i2,...,in′
M ∗
i1,i2,...,in′
Mi1,i2,...,in′
−1,
−1
so in total we have a root τ of order r2 + n′, completing the proof of the theorem.
(cid:3)
We can adapt the construction in the preceding proof to obtain the corresponding
statement in B(H) as follows:
Theorem 3.13. Suppose H is infinite-dimensional separable and ϕ is a normal state on
B(H). Then φ = ϕ(·)1 has a proper n-th discrete root on B(H), for every n ∈ N.
Proof. Let r = dim(pφH) and r′ = dim(p′
φH). We distinguish two cases.
Case r′ = ∞. Here we choose α as a contractive nilpotent CP map of order n on
φH). We define
B(p′
τ(cid:18)w x
z(cid:19) :=(cid:18)ϕ ↾B(pφH) (w)1
0
y
α(z) + ϕ ↾B(pφH) (w)(1 − α(1))(cid:19)
0
Then τ is a proper n-th discrete root.
ROOTS OF COMPLETELY POSITIVE MAPS
11
Case r′ < ∞. Then r = ∞ and we may assume as in the proof of Theorem 3.12 that
the density matrix is in diagonal form with respect to a fixed orthonormal basis (ei) of H
and with entries λ1 ≥ λ2 ≥ . . .. Consider the projection pn onto span{e1, . . . , en}. Then
ϕn :=
1
ϕ(pn)
ϕ ↾B(pnH)
defines a faithful state on B(pnH). We may then proceed as in (Step 1) of the proof
of Theorem 3.12 to find a nilpotent map αn : B(pnH) → B(pnH) of order n such that
αn(pn) = 0 = ϕn ◦ αn. We rescale αn by ϕ(pn) and extend it trivially to B(H) ⊖ B(pnH)
and denote the resulting normal nilpotent map by α. Then
τ := φ + α
is a proper n-th discrete root of φ.
(cid:3)
3.3. Classical probability theory -- proper discrete roots of states on finite-
dimensional commutative von Neumann algebras. We would like to briefly spe-
cialize our general findings to the case of finite classical probability spaces because also
here we get some interesting results. Note that a map τ : Cd → Cd is UCP if and only if τ
is a stochastic matrix and a map ϕ : Cd → C is a state if and only if there is a probability
vector p = (p1, p2, ..., pd)′ ∈ Cd such that ϕ(x) = hp, xi, for all x ∈ Cd.
In this subsection, we will use the following special notation. For x ∈ Cn, y ∈ Cm
we define xihy := xy∗ ∈ Mn,m . For any x = (x1, ..., xd)′ ∈ Cd and m < d, we write
x(m) := (x1, ..., xm)′ ∈ Cm. We write 1 for the unit matrix but also for the unit vector
(1, . . . , 1)′ ∈ Cd. Sometimes we will add subscripts or superscripts to 0 and 1 in order to
indicate the space on which it is acting but we try to avoid this when it is obvious from
the context.
As according to Lemma 3.5, a state on Cd can have proper discrete roots only up to
order d − 1, the states on C and C2 will not have any proper discrete roots. The following
example is a construction of proper n-th discrete roots of states on Cd, for all 2 ≤ n ≤ d−1
and d > 2.
Example 3.14. Let d > 2 and ϕ be a state on Cd given by a probability vector p =
(p1, p2, ..., pd)′. Let φ := ϕ(·)1. Then φ is the stochastic matrix φ = 1ihp.
First let us consider the case when ϕ is faithful, i.e., p = (p1, p2, ..., pd)′ with pi > 0, for
all i. Let 2 ≤ n ≤ d − 1. Note that φ = 1ihp is diagonalizable and of rank one, so we can
write φ = S(cid:18)1 0
α0 ∈ Md−1 of order n and let α = εS(cid:18)0
0 0(cid:19) S−1, with a suitable invertible matrix S. Consider a nilpotent matrix
0 α0(cid:19) S−1. If ε > 0 is small enough then all
entries of φ + α are non-negative because ϕ was assumed to be faithful. By construction
we have got φ ◦ α = 0 = α ◦ φ and hence by Lemma 3.4, τ = φ + α is a proper n-th discrete
root of φ.
0
Now let us assume that ϕ is not faithful. Without loss of generality we can assume that
p = (p1, p2, ..., pr, 0, ..., 0)′, pi > 0 for all i = 1, 2, ..., r < d. Let us consider two separate
cases, namely r ≤ 2 and 2 < r < d, because our construction of n-th roots works differently
in these two cases.
Case r ≤ 2. Given r ≤ n ≤ d − 1, let
τ = 1(r)ihp(r)
y(d − n)ihe(r)
0
1 Sn!
where y(d− n) =(cid:18)1(d−n)
0(n−r)(cid:19) ∈ Cd−r and Sn ∈ Md−r is the operator defined by Sn(e(d−r)
i
) =
for i = d − n, d − n + 1, ..., d − r − 1,
0 for i = d − r, 1, 2, , ..., d − n − 1 and Sn(e(d−r)
i
) = e(d−r)
i+1
12
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
and e(d−r)
of φ. (Note that when n = r, we have y(d − n) = 1(d−r) and Sn = 0.)
is the i-th canonical basis vector in Cd−r. Then τ is a proper n-th discrete root
i
Case r > 2. Given 2 ≤ n ≤ d − 1, decompose n = n1 + n2, with suitable 1 ≤ n1 ≤ r − 1
and 1 ≤ n2 ≤ d − r. Let τ[r,n1] be an n1-th root of 1(r)ihp(r) as in the case of faithful ϕ
above. Then we define
τ =
τ[r,n1]
y(d − n2)ihe(r)
j
0
Sn2!
where y(d − n2) and Sn2 are as in the previous case and j is chosen as follows: if n1 ≥ 2
then choose j such that the j-th row of τ n1−1
, while for n1 = 1 we
choose j = 1. Then τ is a proper n-th discrete root of φ.
[r,n1] is different from p(r)′
We summarize the result of the preceding example as follows:
Theorem 3.15. A state ϕ on Cd has a proper n-th discrete root if and only if 2 ≤ n ≤ d−1.
Or in more probabilistic terms: given a probability distribution p on a probability space with
d elements, there is a stochastic map S that leaves p invariant and such that Sn = 1ihp
and Sk 6= 1ihp for k < n if and only if 2 ≤ n ≤ d − 1.
Here ϕ may be regarded as a stochastic matrix of rank 1. For stochastic matrices of
rank > 1, we have no complete and simple characterization though some partial charac-
terizations with necessary or sufficient conditions are known, e.g. in [16]. The case of rank
d is closely related to Elfving's embedding problem [13, 8].
We continue to use the notation from Section 3.
4. Proper continuous roots
Definition 4.1. Given a von Neumann algebra A and a UNCP map φ : A → A, a proper
continuous root of φ is a strongly-continuous one-parameter semigroup (τt)t≥0 of UNCP
maps on A such that τ1 = φ and τt 6= φ, for all 0 < t < 1.
In this definition one might also consider seemingly more general semigroups with τt0 =
φ for some t0 > 0. However, since we can always reduce the situation to the case t0 = 1
by rescaling, we decided to keep things simple and consider only the case t0 = 1. For
more information on strongly continuous one-parameter semigroups in general, we refer
the reader to [1, 9].
Proposition 4.2. Let A be a finite-dimensional von Neumann algebra and φ : A → A a
UNCP map. Then the following are equivalent:
(i) φ has a proper continuous root;
(ii) φ is bijective and has a proper n-th discrete root, for every n ∈ N \ {1}.
Proof. (i) ⇒ (ii). If (τt)t≥0 is a proper continuous root, then it must be a uniformly con-
tinuous UNCP semigroup, hence of the form τt = etL with some (bounded) conditionally
completely positive generator L, cf. [14, Sec.4.5], so e−L is an inverse of φ (in the sense of
linear maps on A) and τ1/n is a proper n-th discrete root of φ, for every n ∈ N.
(ii) ⇒ (i). If φ has a proper n-th discrete root for every n ∈ N (this is called infinitely
divisible in [11]) then according to [11, Cor.4] there are a conditional expectation E :
A → A and a conditionally completely positive generator L such that φ = eL E. Since φ
and eL are invertible, so is E and hence E must be the identity map because A is finite-
dimensional. Thus we may choose τt = etL, for all t ≥ 0, to obtain a proper continuous
root of φ.
(cid:3)
Remark 4.3. In the classical case, namely if A is commutative, φ is automatically bijective
if it has a proper n-th discrete root for every n ∈ N. This is one of the characterizations of
ROOTS OF COMPLETELY POSITIVE MAPS
13
Markovianity in the context of Elfving's embedding problem due to Kingman [18, Prop.7].
On the other hand, in the noncommutative case, bijectivity is not automatic. E.g. consider
This has proper n-th roots for all n but is clearly not bijective. Similarly, we see that the
UNCP map
0
0
x22 x23
x32 x33
x11
0
0
φ : M3 → M3, φ(x) =
φ : M2 → M2, φ(cid:18)a b
c d(cid:19) =(cid:18) d
b/2
.
a (cid:19)
c/2
from Example 3.10 is bijective but has proper n-th roots only for odd n ∈ N \ {1}, hence
it has no proper continuous root.
The following example provides a bijective UNCP map in finite dimensions where the
In fact, it is a simple "interpolation" of the
conditions in the proposition are verified.
construction in Example 3.9:
Example 4.4. Let φ : M2 → M2 be the UNCP map defined by
For every t ∈ [0, ∞), define
φ(cid:18)a b
b
2
c
2
c d(cid:19) =(cid:18)a
τt(cid:18)a b
d(cid:19) .
c d(cid:19) =(cid:18) a
c
2t
b
2t
d(cid:19) .
τt : M2 → M2,
Then (τt)t≥0 is a proper continuous root of φ, namely τ1 = φ and the semigroup property
and continuity are a straight-forward verification.
Embedding this example into a higher (possibly infinite) dimensional space, we can get
continuous roots for certain UNCP maps in higher dimensions as well. A more complete
criterion as to when such continuous roots exist seems out of reach. Notice that this might
be even more difficult than Problem 3.11.
Yet if φ arises from a state, we can say a little bit more:
Proposition 4.5. Let A be a von Neumann algebra, ϕ a state on A and φ = ϕ(·)1. If
(τt)t≥0 is a proper continuous root of φ then
(i) ϕ ◦ τt = ϕ, for every t ≥ 0, i.e., ϕ is τ -invariant;
(ii) ψ ◦ τt = φ, for every t ≥ 1 and every UNCP map ψ, i.e., all UNCP maps composed
with τt converge to φ in finite time; in particular, τt = φ, for all t ≥ 1.
Proof. (i) Since φ = τ1, for all t ≥ 1, we get from the linearity and the semigroup properties
of τ :
ϕ ◦ τt(x)1 = τ1 ◦ τt(x) = τt+1(x) = τt ◦ τ1(x) = τt(ϕ(x)1) = ϕ(x)τt(1) = ϕ(x)1,
x ∈ A.
(ii) For all t ≥ 1 and x ∈ A, we have, using the unitality and the semigroup property
of τ :
ψ ◦ τt(x) = ψ ◦ τt−1(τ1(x)) = ψ ◦ τt−1(ϕ(x)1) = ϕ(x)ψ ◦ τt−1(1) = ϕ(x)ψ(1) = φ(x).
(cid:3)
The property that (τt)t≥0 stabilizes after time t = 1 is very particular to states, cf.
Example 4.4 for a counter-example. In the special case where φ arises from a state and
moreover A = B(H), we can provide a partial classification of proper continuous roots:
Theorem 4.6. Let A = B(H) with H infinite-dimensional, ϕ a normal state on A and
φ = ϕ(·)1.
(i) If dim(pφH) = 1, i.e. ϕ is a pure state, then φ has a proper continuous root.
(ii) If 1 < dim(pφH) < ∞, i.e. ϕ is a finite convex combination of (at least two) pure
states, then φ has no proper continuous root.
14
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
(iii) If dim(pφH) = ∞, i.e. ϕ is an infinite convex combination of pure states, and
moreover 0 < dim(p′
φH) < ∞ then φ has no proper continuous root.
Proof. (i). This is taken from [3, Ex.1.3] and included here for the sake of completeness.
Since ϕ is pure, we can write ϕ = hξ, ·ξi, where ξ is a suitable vector in H. We decompose
H = Cξ ⊕ L2[0, 1], so pφ is the projection onto the first, p′
φ the projection onto the second
component. Let (St)t≥0 be the standard nilpotent right-shift semigroup on L2[0, 1] defined
as follows: for f ∈ L2[0, 1], t ∈ [0, ∞) and s ∈ [0, 1], set
St(f )(s) =(cid:26) f (s − t) : s − t ∈ [0, 1]
: otherwise.
0
(6)
Then with respect to the decomposition H = Cξ ⊕ L2[0, 1], define
x12S∗
t
x21 x22(cid:19) 7→(cid:18) x11
τt : B(H) → B(H), (cid:18)x11 x12
t )(cid:19) .
τt(cid:18)x11 x12
x21 x22(cid:19) = τt(x) = (1 ⊕ St)x(1 ⊕ St)∗ + ϕ(x)(cid:0)1 − (1 ⊕ St)(1 ⊕ St)∗(cid:1),
t + x11(1 − StS∗
Stx21 Stx22S∗
This can be written as
and it is straight-forward to verify that (τt)t≥0 is a strongly continuous semigroup, every
τt is UNCP and τ1(x) = ϕ(x)1 = φ(x). Thus (τt)t≥0 forms a proper continuous root of φ.
(ii) Suppose a proper continuous root (τt)t≥0 of φ existed. As in Lemma 3.3(4) we see
that (pφτt(·)pφ)t≥0 restricts to a continuous root of pφφ(·)pφ on pφApφ. However, we know
from Proposition 4.2 that such a continuous root cannot exist because pφφ(·)pφ is not
bijective on pφApφ, so we reach a contradiction. Thus φ cannot have a proper continuous
root.
φ according to Lemma 3.3(i), we see that (τt ↾p′
(iii) Suppose for contradiction a proper continuous root (τt)t≥0 of φ existed. Since
φ) ≤ p′
τt(p′
)t≥0 forms an NCP semi-
φAp′
φ
φ) <
group, and according to Lemma 3.3(ii), it is nilpotent with τ1 ↾p′
∞, a CP semigroup must be of the form (etL)t≥0 with bounded conditionally CP map L.
= eL 6= 0, which
Then e−L is the inverse of eL (as a linear map), so we get 0 = τ1 ↾p′
is a contradiction, so φ cannot have a proper continuous root.
= 0. If 0 < dim(p′
Ap′
φ
Ap′
φ
φ
φ
Problem 4.7. In the setting of Theorem 4.6, does φ have a proper continuous root in the
following two missing cases
(cid:3)
(iv) dim pφ = ∞ with dim p′
(v) dim pφ = ∞ with dim p′
φ = 0;
φ = ∞?
We wish to point out that the two cases are equivalent, so it suffices to study (iv).
Remark 4.8. In [3], the roots in case (i) of Theorem 4.6 have been completely classified
in terms of E0-semigroups in standard form, cf. [1] and [20, Def.2.12].
Remark 4.9. A similar construction can be used in order to get a proper continuous
root (Tt)t≥0 of a pure state on an uncountable classical probability space C[0, 1], namely
consider
Tt : C[0, 1] → C[0, 1],
Ttf (s) =(cid:26) f (s − t) : s − t ≥ 0
: otherwise.
f (0)
A pure state on C[0, 1] corresponds to an evaluation functional evx, with some x ∈ [0, 1].
Then evx ◦Tt equals a pure state at all times t ∈ [0, 1], in particular evx ◦T1 = ev0. In
contrast, in the noncommutative case of A = B(H) as in Theorem 4.6(i) suppose ψ 6= ϕ
is another pure state. Then ψ ◦ τt equals the pure states ψ at time t = 0 and ϕ at t = 1
but in between it is a convex combination of two pure states depending on t. Moreover,
for countable classical states space, we expect that no proper continuous root exists at
all. This indicates a stark difference between the commutative and the noncommutative
setting.
ROOTS OF COMPLETELY POSITIVE MAPS
15
Acknowledgments. BVRB thanks S. Kirkland for a mathematical idea which helped
us to construct Example 3.14. We also thank T. Cubitt and M. Skeide for helpful com-
ments on a former version of this manuscript. BVRB furthermore acknowledges financial
support from J.C. Bose Fellowships. RH thanks the Indian Statistical Institute for the
hospitality he received during research visits. NM thanks the Department of Atomic En-
ergy, Government of India, for financial support and the IMSc Chennai for providing the
necessary facilities. VU thanks the National Board for Higher Mathematics, India, for his
PhD fellowship.
References
[1] W. Arveson. Noncommutative dynamics and E-semigroups. Springer (2003).
[2] J. Bausch and T. Cubitt. The complexity of divisibility. Lin. Alg. Appl. 504, 64-107 (2016).
[3] B.V.R. Bhat. Roots of states. Commun. Stoch. Anal. 6, 85-93 (2012).
[4] B.V.R. Bhat and N. Mallick. Nilpotent completely positive maps. Positivity 18(3), 567 -- 577 (2014).
[5] I. Brosh and Y. Gerchak. Markov chains with finite convergence time. Stochastic Process. Appl. 7(3),
247-253 (1978).
[6] T. Cubitt, J. Eisert and M.M. Wolf. The complexity of relating quantum channels to master equations.
Commun. Math. Phys. 310, 383 -- 417 (2012).
[7] T. Cubitt, J. Eisert and M.M. Wolf. Extracting dynamical equations from experimental data is NP-
hard. Phys. Rev. Lett. 108, 120503 (2012).
[8] S.G. Dani. Convolution roots and embeddings of probability measures on locally compact groups.
Indian J. Pure Appl. Math. 41(1), 241-250 (2010).
[9] E.B. Davies. One-parameter semigroups. Academic Press (1980).
[10] E.B. Davies. Embeddable Markov matrices. Electron. J. Probab. 15 (47) 1474 -- 1486 (2010).
[11] L.V. Denisov. Infinitely divisible Markov mappings in quantum probability. Theory Probab. Appl.,
33(2), 392 -- 395 (1988).
[12] J. Dixmier. Von Neumann algebras. Elsevier North Holland Publishing (1981).
[13] G. Elfving. Zur Theorie der Markoffschen Ketten. Acta Societatis scientiarum Fennicae, Nova series
A 2(8) (1937).
[14] D.E. Evans and Y. Kawahigashi. Quantum symmetries on operator algebras. Oxford University Press
(1998).
[15] P.W. Glynn and D.L. Iglehart. Conditions under which a Markov chain converges to its steady state
in finite time. Probability in the engineering and informational sciences 2, 377-382 (1988).
[16] N.J. Higham and L. Lin. On pth roots of stochastic matrices. Lin. Alg. Appl. 435(3), 448-463 (2011).
[17] S.-G. Hwang and S.-S. Pyo. Doubly stochastic matrices whose powers eventually stop. Lin. Alg. Appl.
330, 25-30 (2001).
[18] J.F.C. Kingman. The imbedding problem for finite Markov chains. Z. Wahrscheinlichkeitstheorie 1,
14-24 (1962).
[19] H. Minc. Nonnegative matrices. Wiley (1988).
[20] R.T. Powers. New examples of continuous spatial semigroups of *-endomorphisms of B(H). Internat.
J. Math. 10(2), 215-288 (1999).
[21] M. Rahaman, S. Jaques and V.I. Paulsen. Eventually entanglement breaking maps. J. Math. Phys.
59, 062201 (2018).
[22] M. Reed and B. Simon. Methods of modern mathematical physics I. Functional analysis. Academic
Press (1980).
[23] E. Størmer. Positive linear maps of operator algebras. Springer (2013).
[24] E. J. Subelman. On the class of Markov chains with finite convergence time. Stoch. Processes Appl.
4(3), 253-259 (1976).
[25] M. Takesaki. Theory of operator algebras I. Springer (2002).
[26] A. Van-Brunt. Infinitely divisible nonnegative matrices, M -matrices, and the embedding problem for
finite state stationary Markov chains. Lin. Alg. Appl. 541, 163 -- 176 (2018).
[27] M.M. Wolf. Quantum channels and operations guided tour (2011 lecture notes on: www-m5.ma.
tum.de/Allgemeines/MichaelWolfEn).
[28] M.M. Wolf and J.I. Cirac. Dividing quantum channels. Commun. Math. Phys. 279,147-168 (2008).
[29] M.M. Wolf, J. Eisert, T. Cubitt and J.I. Cirac. Assessing non-Markovian dynamics. Phys. Rev. Lett.
101, 150402 (2008).
[30] J. Yuan. On the construction of one parameter semigroups in topological semigroups. Pacific J. Math.
65, 285-292 (1976).
16
B.V. RAJARAMA BHAT, ROBIN HILLIER, NIRUPAMA MALLICK, AND VIJAYA KUMAR U.
Indian Statistical Institute, Stat-Math. Unit, R V College Post, Bengaluru 560059, India
Lancaster University, Department of Mathematics and Statistics, Lancaster LA1 4YF,
United Kingdom
The Institute of Mathematical Sciences, IV Cross Road, CIT Campus, Taramani, Chennai
600113, India
Indian Statistical Institute, Stat-Math Unit, R V College Post, Bengaluru 560059, India
|
1111.7240 | 1 | 1111 | 2011-11-30T17:13:56 | Structure of associative subalgebras of Jordan operator algebras | [
"math.OA"
] | We show that any order isomorphism between ordered structures of associative unital JB-subalgebras of JBW algebras is implemented naturally by a Jordan isomorphism. Consequently, JBW algebras are determined by the structure of their associative unital JB subalgebras. Further we show that in a similar way it is possible to reconstruct Jordan structure from the order structure of associative subalgebras endowed with an orthogonality relation. In case of abelian subalgebras of von Neuman algebra it is we shown that order isomorphisms of the structure of abelian C*-subalgebras that are well behaved with respect to the structure of two by two matrices over original algebra are implemented by *-isomorphisms. | math.OA | math |
Structure of associative subalgebras of
Jordan operator algebras
Jan Hamhalter and Ekaterina Turilova
Czech Technical University in Prague - El. Eng.
Department of Mathematics, Technicka 2,
166 27 Prague 6, Czech Republic
[email protected]
Kazan Federal University,
Faculty of Computational Mathematics and Cybernetics,
Kremlevskaya 18, Kazan, Russia,
e-mail: [email protected]
Abstract: We show that any order isomorphism between ordered structures of associative unital
JB-subalgebras of JBW algebras is implemented naturally by a Jordan isomorphism. Consequently,
JBW algebras are determined by the structure of their associative unital JB subalgebras. Further
we show that in a similar way it is possible to reconstruct Jordan structure from the order structure
of associative subalgebras endowed with an orthogonality relation. In case of abelian subalgebras of
von Neumann algebras we show that order isomorphisms of the structure of abelian C ∗-subalgebras
that are well behaved with respect to the structure of two by two matrices over original algebra are
implemented by ∗-isomorphisms.
keywords: Jordan algebras, structure of associative subalgebras
1 Introduction and Preliminaries
JB algebra is a far reaching generalization of the self-adjoint part (A, ◦) of a
C ∗-algebra (A, ·) endowed with the product
x ◦ y =
1
2
(xy + yx)
x, y ∈ A .
Jordan product is not associative in general. An associative Jordan algebras
are quite special. More precisely, basic result of spectral theory says that JB
algebra (A, ◦) is associative if, and only if, it is isomorphic to the algebra C0(X)
of all continuous real-valued functions on a locally compact Hausdorf space
1
X vanishing at infinity.
In a similar way, JBW algebras, i.e. JB algebras
with preduals, are generalisation of von Neumann algebras. The associative
subalgebras of a given JB algebra are mutually overlapping and, when ordered by
set-theoretic inclusion, they form a partially ordered set (poset for short). This
poset can be classified as a semi-lattice: any two elements admit an infimum,
which is a set theoretic intersection. If two JBW algebras are isomorphic, then
their corresponding posets of associative subalgebras are order isomorphic, and
so we obtain an order theoretic invariant of JBW algebras. An interesting
question arises whether, on the contrary, the poset of associative subalgebras
determine fully structure of JBW algebras. More precisely, in this note we
shall be mainly concentrated on on the following problem: Let A1 and A2
be JBW algebras and L(A1) and L(A2) be some structures of associative JB
subalgebras of A1 and A2, respectively, ordered by set theoretic inclusion. Let
L(A1) → L(A2) be an order isomorphism. Is there a unique Jordan isomorphism
ψ : A1 → A2 such that
ϕ(C) = ψ(C)
for all C ∈ L(A1)?
We say in this case that the Jordan isomorphism ψ implements ϕ. The pos-
itive answer to this question means that, perhaps surprisingly, complicated
functional-analytic structure of JBW algebras can be encoded into "discrete"
algebraic ordered structure. We show that the solution to the problem stated
above depends on the type of associative subalgebras chosen. First we show
that the problem has positive answer if we consider associative unital JB sub-
algebras. (Unital subalgebra is a subalgebra containing the unit of the whole
algebra.) In the sequel let ASU (A) denote the poset of all unital associative JB
subalgebras of a JB algebra A. The following result shows that unital in case
when A is associative then the structure ASU (A) determines A. This result has
been proved for abelian subagebras in [8]. Its translation to associative Jordan
algebras is straightforward.
1.1. Theorem. Let A and B be unital associative JB algebras and ϕ : ASU (A) →
ASU (B) be an order isomorphism. Suppose that dim A 6= 2. Then there is a
unique Jordan isomorphism ψ : A → B that implements ϕ.
In the first part of the paper we extend this result in the following way.
Suppose that A is a JBW algebra without Type I2 direct summand that is
not isomorphic to R ⊕ R. Then, for each order isomorphism ϕ : ASU (A) →
ASU (B), where B is another J BW algebra, there is a unique Jordan isomor-
phism ψ : A → B, implementing ϕ. In the light of counterexamples we show
that the assumptions are not superfluous. Our result generalizes the correspond-
ing results for abelian C ∗-subalgebras of von Neumann algebras obtained in [8]
by J.Hamhalter.
An important types of associative subalgebras are singly generated subalge-
bras. These algebras may not have a unit in general. It is therefore natural to
consider the problem of determining whole Jordan structure by structure of all
2
associative subalgebras. Let us denote by AS(A) the system of all associative
JB subalgebras of a JBW algera A, again ordered by set theoretic inclusion. In
this case the situation is different. Since any Jordan isomorphism preserves the
unit, it can implement only order isomorphism that preserves unital associative
subalgebras. Therefore, in order to recover the structure of the algebra, we have
to consider some additional structure on the set AS(A). One of the natural pos-
sibilities seems to be an orthogonality relation on AS(A) introduced as follows.
Let C1, C2 ∈ AS(A). We say that C1 and C2 are orthogonal (in the symbols
C1 ⊥ C2) if C1 and C2 operator commute, and x ◦ y = 0 whenever x ∈ C1 and
y ∈ C2. We shall show that more or less any transformation of the structure
of associative subalgebras that preserves the order and orthogonality relation in
both directions, is implemented by a unique Jordan isomorphism. This result
is new also in the context of von Neumann algebras.
In case of von Neumann algebras it is known that the order structure of abelian
subalgebras does not suffice to recover the whole structure. The reason is a
deep result of Connes [4], who showed that there is a von Neumann factor M
that is not anti-isomorphic to itself. It means that M and its opposite algebra
M o have identical structures of all abelian subalgebras, although they are not
∗-isomorphic. (Let us recall that opposite algebra has identical linear structure
and norm but the multiplication is reversed.) In this light, the following open
problem has been formulated by A.Doring and J.Harding [6]: What additional
structure on the poset of abelian subalgebras suffices for recovering whole von
Neumann algebra structure? In the last part of this paper we show one possi-
bility in this direction using the structure of two by two matrices over original
algebra.
We briefly discuss the relevance of the results obtained to foundations of
quantum theory.
Le us now recall a few concepts and fix the notation. (Our standard refer-
ence for Jordan algebras is [9].) JB algebra is a real Banach algebra with the
commutative product, ◦, that satisfies the following properties for all a, b ∈ A:
(i) a ◦ (b ◦ a2) = (a ◦ b) ◦ a2 (ii): ka2k ≤ ka2 + b2k, kak2 = ka2k. JBW algera
is a JB algebra that is a dual Banach space. Throughout the paper A shall
denote the JBW algebra. We write A1 = {a ∈ A kak ≤ 1}, A+ = {a2 a ∈ A}.
For a ∈ A, mappings Ta, Ua : A → A are defined by letting Ta(b) = a ◦ b,
Ua(b) = 2a ◦ (a ◦ b) − a2 ◦ b. It is well known that Ua is positive in the sense
that Ua(A+) ⊂ A+. By a projection we mean an idempotent of A. The symbol
P (A) shall be reserved for the set of all projections in A. We shall often use
the fact that, for a projection p ∈ A, Up(A) is a JBW subalgebra of A with the
unit p. For x1, . . . , xn ∈ A we shall write J B(x1, . . . , xn) for a JB subalgebra of
A generated by x1, . . . , xn. Two elements a, b ∈ A are said to be operator com-
muting if TaTb = TbTa. Two subsets C, D ⊂ A are called operator commuting
if each element of C operator commutes with each element of D. The center
Z(A) of A is the set of all elements that operator commute with each element
of A. If the center is trivial, then A is called a factor. Positive functional ϕ on
3
A is a a linear functional such that ϕ(a) ≥ 0 for all a ∈ A+. A state is a norm
one positive functional. We shall need the concept of the spin factor. Let Hn
be a real Hilbert space of dimension n ≥ 2. Let Vn = Hn ⊕ R have the norm
ka + l1k = kak + klk, a ∈ Hn, l ∈ R. Define a product in Vn by
(a + l1) ◦ (b + µ1) = (µa + lb) + (< a, b > +lµ)1 ,
where a, b ∈ Hn, l, µ ∈ R. Then Vn is a JBW algebra that is called the spin
factor. It is known that spin factors are exactly Type I2 factors in structural
theory of JBW algebras.
Le us now recall a few concepts of theory of ordered structures. Suppose
that (X, ≤) is a poset with a smallest element 0. We say that element x ∈ X
covers element y ∈ X (in the symbol x ⊲ y), if x ≥ y, x 6= y, and the following
implication holds: if x ≥ z ≥ y, then either x = z or y = z. The elements that
cover 0 are called atoms. Let x ∈ X and suppose that there is a finite sequence
x1, . . . , xn such that x = x1 ⊲ x2 ⊲ · · · ⊲ xn ⊲ 0. The minimal n with this property
is then called the height of an element x.
2 Unital associative subalgebras
Let 1 denote the unit of the JBW algebra A.
In this section we shall deal
with order isomorphisms of the structure ASU (A) of all unital associative JB
subalgebras of A containing 1, ordered by set theoretic inclusion. Let B be
another JBW algebra. The bijection ϕ : ASU (A) → ASU (B) is called an order
isomorphism if it preserves the order in both directions, i.e. if
C1 ⊂ C2 ⇐⇒ ϕ(C1) ⊂ ϕ(C2)
for all C1, C2 ∈ ASU (A) .
We say that order isomorphism ϕ is implemented by a Jordan isomorphism
ψ : A → B if
ϕ(C) = ψ(C)
for all C ∈ ASU (A) .
The main result of this section is the following theorem:
2.1. Theorem. Let A be a JBW algebra without Type I2 direct summand that
is not isomorphic to R ⊕ R. Let B be another JBW algebra. For any order
isomorphism
ϕ : ASU (A) → ASU (B) ,
there is a unique Jordan isomorphism ψ : A → B, implementing ϕ.
We shall prove this theorem in series of statements.
2.2. Lemma. Let X be a locally compact Hausdorff space. If card X ≥ k ≥ 2,
then C0(X) contains a proper unital subalgebra of dimension at least k − 1.
4
Proof: By the assumption, there are distinct points x1, . . . , xk in X. Let us
consider a subalgebra C = {f ∈ C0(X) f (x1) = f (xk)}. It is a proper unital
subalgebra. Using Uryson lemma we argue that C contains k − 1 functions gi,
(i ≤ k − 1), such that gi(xj ) = δij (i, j = 1, . . . , k − 1). Therefore, dim C ≥ k − 1.
(cid:3)
Let us remark that the algebra ASU (A) has a smallest element -- the space
generated by 1. Next lemma characterises the atoms in the poset ASU (A).
2.3. Lemma. C ∈ ASU (A) is an atom if, and only if, there is a projection p
in A, different from 0 and 1, such that
C = sp{p, 1 − p} .
Proof: Any two-dimensional unital associative algebra contains only one-
dimensional subalgebra generated by the unit 1 as its proper subalgebra and
so it is an atom in the poset ASU (A). Let us prove the reverse implication.
Suppose that C is an atom in ASU (A). Then C is isomorphic to C(X), where
X is a compact Hausdorff space. If C has three distinct points, then, according
to Lemma 2.2, C contains a proper two-dimensional subalgebra. This is a con-
tradiction. So X = {x, y} and, in turn, C = sp{p, 1 − p}, where p is a projection
corresponding to the characteristic function of the set {x}.
(cid:3)
2.4. Lemma. Let C be a maximal unital associative subalgebra of A and p ∈ C
be a projection. Then dim Up(C) = 1 if, and only if, dim Up(A) = 1 .
Proof: The reverse implication is trivial and so we show that dim Up(C) = 1
implies dim Up(A) = 1. Suppose, for a contradiction, that dim Up(C) = 1 and
dim Up(A) ≥ 2. Then there is a positive a, with 0 ≤ a ≤ p, that is not a multiple
of p. We can write
C = p R ⊕ (1 − p) ◦ C .
(1)
As a ∈ Up(A), a operator commutes with all elements of U1−p(A) (see [9, Lemma
2.6.3., p. 48]). But U1−p(A) contains U1−p(C) = (1 − p) ◦ C, and so, by (1), a
operator commutes with C. In view of maximality of C, we obtain that a ∈ C.
But then a ∈ R p, which is a contradiction.
(cid:3)
The following proposition characterises the JBW algebras that admit asso-
ciative unital subalgebra that is simultaneously minimal and maximal.
2.5. Proposition. A contains a two-dimensional maximal associative unital
subalgebra if, and only if, A is a Type I2 factor or R ⊕ R.
Proof: Suppose that C is a two-dimensional maximal associative unital sub-
algebra of A. Then C = sp{p, 1 − p}, where p is a projection. By the previous
lemma
dim Up(A) = dim U1−p(A) = 1 .
5
It means that p and 1 − p are atomic projections in A. Let us denote by c(p)
and c(1 − p) central supports of p and q, respectively. As p is an atom in A, its
central cover, c(p), is an atom in the center Z(A). (Let us recall that central
cover c(p) is the smallest central projection z such that z ≥ p. This fact is
probably known, but since we have not been able to find appropriate quotation,
we give full argument here. Suppose that z is a central projection such that
z ≤ c(p). By [9, Lemma 4.3.5, p. 103], we can compute
c(z ◦ p) = z ◦ c(p) = z .
By atomicity of p we conclude that either z ◦ p = 0 or z ◦ p = p. The former
case gives that z = 0, while the latter case means that z = c(p). Therefore c(p)
is an atom. By the same reason c(1 − p) is an atomic central projection. Now
we have two possibilities:
c(p)c(1 − p) = 0
or
c(p) = c(1 − p) .
Consider first the case, when c(p) and c(1 − p) are orthogonal. As
1 ≥ c(p) + c(1 − p) ≥ p + 1 − p = 1 ,
we obtain that c(p) = p and c(1 − p) = 1 − p. Whence, A is isometric to R ⊕ R.
If c(p) = c(1 − p), then
c(p) ≥ p + 1 − p = 1 .
and so c(p) = c(1 − p) = 1. Atomicity of c(p) in the center yields that A is
a factor. As A contains two orthogonal projections with unit central supports
that sums to 1, A has to be a factor of Type I2.
Let us now prove the reverse implication. Any Type I2 factor is a spin factor
Vn = R ⊕ Hn, where Hn is a Hilbert space. Each nontrivial projection in Vn is
of the form
1
2
(1 + µ) ,
where µ is a unit vector in Hn. It follows that each nontrivial projection in Vn
is atomic. It can be easily verified that, for a unit vector µ ∈ Hn ,
1
2
(1 + µ) ,
1
2
(1 − µ)
is a pair of orthogonal atomic projections, whose linear span is a two-dimensional
maximal unital associative subalgebra of A.
(cid:3)
Finally, we have passed to the proof of the main theorem.
Proof of Theorem 2.1: Suppose that ϕ : ASU (A) → ASU (B) is an order
isomorphism. Let X be a maximal associative unital subalgebra of A. By
Proposition 2.5, we have that dim X ≥ 3. By Theorem 1.1 we see that there is
6
a unique Jordan isomorphism ψX : X → ϕ(X) implementing ϕ. We will show
the interplay between various ψC , where C runs through ASU (A). Let E and
D be two different maximal elements of ASU (A). We show that
ψE = ψD
on C = E ∩ D .
If dim C ≥ 3, the equality above follows from uniqueness of implementing Jordan
isomorphism established in Theorem 1.1. Therefore, we concentrate on the case
when dim C = 2, which means that
C = sp{p, 1 − p} , where p ∈ P (A) .
It can be verified easily that if dim Up(A) = dim U1−p(A) = 1, then C is a
maximal associative subalgebra. Indeed, if x ∈ A operator commutes with C,
then p ◦ x = Up(x) = l1p for some l1 ∈ R. Similarly, (1 − p) ◦ x = U1−p(x) = l2p
for some l2 ∈ R. It implies that x = p ◦ x + (1 − p) ◦ x = l1p + l2p ∈ C.
However, it would mean, by Proposition 2.5, that A is either Type I2 or R ⊕ R,
which is excluded by the assumption. So it cannot happen that p and 1 − p are
simultaneously one-dimensional projections. Consider first the case when
dim Up(A) = 1
and
dim U1−p(A) ≥ 2 .
By Lemma 2.3
ϕ(C) = sp{q, 1 − q}
for some q ∈ P (B). Obviously,
ψD(C) = ψE(C) = sp{q, 1 − q} .
As dim Up(D) = 1, we infer that
dim UψD(p)(ψD(D)) = 1 .
In virtue of Lemma 2.4, we have that
We shall prove further that
dim UψD (p)(B) = 1 .
dim U1−ψD(p)(B) ≥ 2 .
Assume the opposite and try to reach a contradiction. If dim U1−ψD(p)(B) = 1 ,
then ψD(C) (= ϕ(C)) is a maximal associative subalgebra of B (see the argu-
ments before). As ϕ is isomorphism, the same must be true of C.
It would
mean that C = D = E, contradicting E 6= D. Suppose, without loss of gen-
erality, that q = ψD(p). By applying the same arguments to ψE we have that
ψE(p) = q, and so ψD = ψE on C.
7
As the second case, suppose that dim Up(A) ≥ 2 and dim U1−p(A) ≥ 2.
Employing Lemma 2.4, there exist elements x ∈ D, y ∈ E such that x ≤ p, y ≤
1 − p and such that the sets {p, x} and {1 − p, y} are linearly independent. Set
V = J B(1, p, x) , W = J B(1, p, y) , U = J B(1, p, x, y) .
These algebras are associative. It holds that dim V = dim W ≥ 3. By Theo-
rem 1.1 we have
ψV (p) = ψD(p)
ψW (p) = ψE(p) .
(2)
(3)
But the inclusions V, W ⊂ U entail
ψV (p) = ψW (p) = ψU (p) .
We can conclude again that
ψD = ψE
on C .
Let us now define the map ψ. If x ∈ A, then x is contained in some maximal
associative unital JB subalgebra X. Put
ψ(x) = ψX (x) .
By the previous reasoning this definition does not depend on the choice of X.
As ψX is a Jordan isomorphism, we see that ψ is an isometry. The restriction of
ψ to the projection lattice P (A) gives a finitely additive measure, which means
that
ψ(p + q) = ψ(p) + ψ(q) ,
whenever p and q are orthogonal projections. Let us consider a state f ∈ B ∗.
Then f ◦ ψ is a nonnegative scalar measure. By deep Gleason type theorem for
JBW algebras proved by Bunce and Wright in [1, 2], f ◦ ψ extends uniquely
to a state ψf on A. (For this argument the assumption on Type I2 factor is
necessary.) Take x, y ∈ A. Then
f (ψ(x + y)) = ψf (x + y) = ψf (x) + ψf (y) = f (ψ(x) + ψ(y)) .
As the set of states is separating, we see that
ψ(x + y) = ψ(x) + ψ(y) ,
establishing the linearity of ψ.
It follows immediately from the definition of
ψ, that ψ is a Jordan homomorphism that implements ϕ. Applying the same
argument to ϕ−1, we see that ψ is a Jordan surjective isomorphism.
Finally, let us establish the uniqueness of ψ. Suppose that ϕ is implemented
is a Jordan
by two Jordan isomorphisms ψ1, ψ2 : A → B. Then = ψ1 ◦ ψ−1
2
8
isomorphism that implements the identity on ASU (A). For any x ∈ A take
a maximal associative unital subalgebra C of A containing x. By assumption,
dim C ≥ 3, and so (x) = x, for all x ∈ C by Theorem 1.1. In other words, is
an identity and so ψ1 = ψ2.
(cid:3)
Let us remark that the assumption on the absence of Type I2 factor is
not superfluous for establishing the uniqueness of implementing Jordan isomor-
phism. Indeed, consider spin factor Vn = R ⊕ Hn. Then, it can be verified easily
that all associative unital subalgebras (except for whole algebra and the algebra
generated by the unit) are of the form
Aξ = sp(cid:26) 1
2
(1 + ξ),
1
2
(1 − ξ)(cid:27) ,
where ξ is a unit vector. System (Aξ)ξ, where ξ runs through the unit ball of Hn
is a family of subalgebras covering whole V and intersecting each other in the
subalgebra sp{1}. Let ψξ be a Jordan automorphism of Aξ. The union of the
graphs of ψ′
ξs gives a Jordan automorphism of A. By this way we can get many
Jordan automorphisms of V implementing the identical order isomorphism of
ASU (Vn).
3 Non unital associative subalgebras
This section is devoted to reconstruction of JBW algebra from its structure of all
nonunital associative subalgebras. Let us denote by AS(A) the structure of all
associative JB subalgebras of A. As we remarked already in the introduction,
the order isomorphism between posets of associative subalgebras may not be
implemented by any Jordan isomorphism. For example, let us take an algebra
R ⊕ R. Then AS(A) consists of the whole algebra (the largest element), the zero
algebra (the smallest element), and then of three atoms. Any permutation of the
atoms that leave other algebras unchanged is an order isomorphism. One of the
atoms is the algebra generated by the unit. This atom is invariant with respect
to any isomorphism implemented by a Jordan isomorphism. So there are order
automorphisms of AS(A) that are not implemented by Jordan automorphisms.
In the light of it, we shall introduce another structure on AS(A) that will
complete the information contained in the order. We say that two algebras C
and D in AS(A) are orthogonal (in symbols C ⊥ D) if x ◦ y = 0 for all x ∈ C
and y ∈ D. Let us note that orthogonality implies that C and D operator
commute. Indeed, let us take positive elements x ≤ 1 and y ≤ 1 from C and D,
respectively. Since for any polynomial p with p(0) = 0 we have that p(x) ∈ C,
we infer that p(x) ◦ y = 0. Consequently, using function calculus we can see
that for any continuous function f , with f (0) = 0, we have f (x) ◦ y = 0. As the
range projection r(x) of x is a strong operator limit of elements of the form f (x)
above, we conclude that r(x) ◦ y = 0. By symmetry we see that r(x) and r(y)
are orthogonal projections. By the well known result Ur(x)(A) and Ur(y)(A)
operator commute. As x ∈ Ur(x)(A) and y ∈ Ur(y)(A), we obtain that x and
9
y operator commute. Since any element in a JB algebra is a difference of two
positive elements, we have shown that C and D operator commute.
Next theorem is the main result of this section.
3.1. Theorem. Let A and B be JBW-algebras such that A has no Type I2
direct summand. Let ϕ : (AS(A), ⊂, ⊥) → (AS(B), ⊂, ⊥) be an order isomor-
phism preserving the orthogonality, ⊥, in both directions. Then there is a unique
Jordan isomorphism
ψ : A → B
such that
ϕ(C) = ψ(C)
for all C ∈ AS(A) .
Proof will be divided into a few next steps.
3.2. Proposition. Let A be a JBW algebra. The following conditions are
equivalent:
(i) X ∈ AS(A) has dimension k
(ii) X = sp{p1, . . . , pk}, where p1, . . . , pk are orthogonal nonzero projections.
(iii) There is a chain in AS(A) such that
X ⊲ Xk−1 ⊲ · · · ⊲ X1 ⊲ {0} .
Proof: (i) ⇒ (ii) Each associative JB algebra is isomorphic to the algebra
C0(X) of real valued continuous functions on a locally compact Hausdorff topo-
logical space K that vanish at infinity. If dim C0(X) = k, then X must contain
precisely k points, i.e. K = {x1, . . . , xk}. Characteristic functions, pi, of single-
tons {xi} give the desired system of projections.
(ii) ⇒ (iii) It is immediate by setting
Xk−1 = sp{p2, · · · , pk}, . . . , X1 = sp{p1} .
(iii) ⇒ (ii) We shall proceed by induction on k. Suppose that X = C0(K)
is an atom in AS(A). We shall prove that K is singleton. Indeed, assume K
has two distinct points, say x and y. They can be separated by a continuous
function f in C0(K) in the sense that f (x) = 1, f (y) = 0. The algebra J B(f )
is an associative nonzero subalgebra consisting of functions vanishing at y. It
means that this algebra is a proper nonzero subalgebra of X, which contradicts
the fact that X is an atom.
Suppose now that X ⊲ Xk−1, where dim Xk = k − 1. We shall prove that
dim Xk = k. Suppose, for a contradiction, that dim X ≥ k + 1, which forces
that card K ≥ k + 1. Let us represent X as a space C0(K) again. Then
Xk−1 is by induction hypothesis a linear span of projections p1, . . . , pk−1, each
corresponding to a clopen subset Z1, . . . , Zk−1 of K. Then Z1, . . . , Zk−1, Zk =
10
K \ Sk−1
i=1 Zi is a covering of K by disjoint clopen sets. As card K ≥ k + 1, at
least one of these sets, say Z1, contains two distinct points, say x and y. Using
basic properties of locally compact spaces, there are function f, g in C0(X) such
that f (x) = 1, f (y) = 0, g(x) = 0, g(y) = 1, and such that f and g vanish
outside Z1. It is then clear that the algebra J B(f, g, p2, . . . , pk), where pk is a
projection corresponding to Zk, is a proper subalgebra of X that contains Xk−1
as its proper subalgebra. This is a contradiction.
(cid:3)
3.3. Corollary. There is a one-to-one correspondence between atoms in AS(A),
and nonzero projections in A such that each atom C ∈ AS(A) is of the form
C = sp{p}, where p ∈ P (A) \ {0}.
3.4. Lemma. Let p and q be nonzero orthogonal projections and z be a nonzero
projection different from p and q. Then
z = p + q ,
if, and only if, the set {p, q, z} is a 2-dimensional asociative subalgebra.
Proof: If {p, q, z} is a two-dimensional associative algebra, then z must be
equal to p + q. Indeed, if z 6= p + q, then for one of the remaining projections,
say p, we have z ⊥ p. But then q = z + p, which is in contradiction with q ⊥ p.
The reverse implication is obvious.
(cid:3)
Proof of the Theorem 3.1: Suppose that ϕ : AS(A) → AS(B) is an order
isomorphism preserving the orthogonality in both directions. Since ϕ preserves
faithfully one-dimensional subalgebras (i.e. atoms), we can define a map
by equation
ψ : P (A) \ {0} → P (B) \ {0}
sp{ψ(p)} = ϕ(sp{p}) .
Then, according to Corollary 3.3, ψ is a bijective map of P (A) onto P (B). As
two nonzero projections p, q are orthogonal if, and only if, sp{p} and sp{q}
are orthogonal algebras, we see that ψ preserves orthogonality of projections.
Suppose now that
z = p + q ,
where z, p, q are nonzero projections. Thanks to the fact that ϕ preserves two-
dimensional subalgebras (Propositon 3.2), we have
ϕ(sp{p, q}) = sp{ψ(p), ψ(q)} ,
and by Lemma 3.4 we see that
ψ(z) = ψ(p + q) = ψ(p) + ψ(q) .
In other words, defining ψ(0) = 0, ψ : P (A) → P (B) is a bounded finitely addi-
tive vector measure. Using now Gleason type theorem and the same arguments
11
as in the proof of Theorem 2.1, we infer that ψ extends to a unique bounded
linear map between A and B. We shall denote this map by the same symbol.
As ψ preserves the projections, it is Jordan homomorphism. Applying the same
argument to the inverse, ϕ−1, we deduce that ψ is a Jordan isomorphism. Let
us now take an arbitrary JB subalgebra C ∈ AS(A). Fix an element 0 ≤ x ≤ 1
in C. Then
x = Xn
1
2n
pn ,
where (pn) is a sequence of projections in C. As ψ(pn) ∈ ϕ(C) for all n, we
conclude that
ψ(x) = Xn
1
2n
ψ(pn) ∈ ϕ(C) .
Hence, ψ(C) ⊂ ϕ(C). Applying the same argument to the inverse map ϕ−1 we
see that
ψ(C) = ϕ(C) .
Uniqueness of ψ is a consequence of the fact that Jordan map is uniquely
(cid:3)
determined by its values on the projections.
Let us notice that, by the arguments in the proof, it is, in fact, possible to
recover Jordan structure from the structure of elements of height two in AS(A).
4 Recovering full C ∗-structure
In this section we shall make an initial step to finding structure on abelian
subalgebras that contain all information on whole algebra already and not only
on the Jordan part of the associative product. One of the natural possibilities
is to consider the structure of two by two matrices over original algebra. As
it is well known, for example, Jordan maps that are well behaved with respect
to this structure, are already ∗-homomorphisms. This motivates the following
definition that can be viewed as order theoretic version of the amplification for
maps. First, some notation will be useful. In the sequel let A be a von Neumann
algebra with self-adjoint part A. M2(A) will denote the algebra of two by two
matrices over A. We will identify it with the tensor product M2 ⊗ A, where M2
is the algebra of two by two complex matrices. The algebra A will be embedded
into M2(A) diagonally as 1 ⊗ A. By Abel(A) we shall mean the structure of
abelian subalgebras of A ordered by set inclusion. The orthogonality relation, ⊥,
on Abel(A) will be naturally extended from the corresponding structure AS(A):
We say that C, D ∈ Abel(A) are orthogonal (in symbols C ⊥ D) if xy = 0 for
all x ∈ C and y ∈ D.
4.1. Definition. Let A and B be von Neumann algebras. Let ϕ : Abel(A) →
Abel(B) be an order isomorphism preserving the orthogonality in both direc-
tions. We say that ω:Abel(M2(A)) → Abel(M2(B)) is an amplification of ϕ if
the following conditions are satisfied.
12
(i) ω is an order isomorphism preserving orthogonality in both directions.
(ii) For all abelian subalgebras C ∈ Abel(A) and D ∈ Abel(B) we have
ω(C ⊗ D) = C ⊗ ϕ(D) .
4.2. Proposition. Suppose that A has no direct summand of Type I2 and B is
another von Neumann algebra. Let ϕ : Abel(A) → Abel(B) be an order isomor-
phism preserving the orthogonality in both directions that has an amplification
ω. Then ϕ is implemented by a ∗-isomorphism between A and B.
Proof: By our results from the previous section (rewritten easily into the
language of von Neumann algebras), ω is induced by a Jordan ∗-isomorphism
Ψ : M2(A) → M2(B), and ϕ is induced by a Jordan ∗-isomorphism ψ : A → B.
Let p and q be projections in A and B, respectively. Then
sp{p} ⊗ sp{q} = sp{p ⊗ q}
is an abelian subalgebra which is mapped by ω onto abelian subalgebra
sp{p} ⊗ sp{q′} ,
where q′ ∈ P (B) .
Therefore Ψ(p ⊗ q) ∈ sp{p ⊗ q′} and taking into account that Ψ is a Jordan map
(and so preserving projections), we infer that
Ψ(p ⊗ q) = p ⊗ q′ .
(4)
By the spectral theorem, we obtain from (4) that for all x ∈ A, y ∈ B we have
Ψ(x ⊗ y) = x ⊗ z ,
where z is some element of A. In other words,
Ψ = id ⊗ ,
where : A → B is a linear map. Because preserves projections, it is a Jordan
map. As 1 ⊗ A is identified with A, we have that = ψ by uniqueness of
a Jordan map implementing ϕ. So ψ is a Jordan map whose amplification is
again Jordan. Consequently, ψ is 2-positive. By the result of Choi [3, Cor. 3.2],
any two-positive unital Jordan map is a ∗-homomorphism. This concludes the
proof.
(cid:3)
Let us remark that in mathematical foundations of quantum theory, the
system of observables is given by a JB algebra or a JBW algebra. This model
extends the classical system in which observables are given by elements of an
associative JB algebra, i.e. by the algebra of continuous functions on some
phase space. From this point of view the quantum system is a union of mu-
tually overlapping and interacting classical systems. It seems to be natural to
13
capture the structure of classical subsystems by the structure of associative sub-
algebras considered in this paper. In this context, our results may be viewed as
possibility to reconstruct global quantum system and its symmetries (given by
Jordan isomorphisms) from classical subsystems. That is why, apart from its
own mathematical interest, our study has been motivated by extensive research
in mathematical foundations of quantum theory (including topos theory) [5, ?].
References
1. L.J.Bunce, J.D.M.Wright: Continuity and linear extensions of quantum
measures on Jordan operator algebras. Math. Scand. 64 (1989), 300-306.
2. L.J.Bunce, J.D.M.Wright, Quantum measures and states on Jordan alge-
bras, Commun. Math. Phys. 98, 187-202 (1985).
3. M.D.Choi: A Schwarz inequality for positive linear maps on C ∗-algebras,
Illionis J. Math. (1974), 565-577.
4. A.Connes, A factor not isomorphic to itself, Annals of Mathematics 101
no. 3 (1962) 536-554.
5. A.Doring, C.J.Isham, A topos foundation for theories of physics: IV. Cat-
egories of systems, J. Math. Phys 49 953518 (2008).
6. A.Doring and J.Harding, Abelian subalgebras and the Jordan structure of
von Neuman algebras, arXiv: 1009.4945v1.
7. J.Hamhalter: Quantum Measure Theory, Kluwer Academic Publishers,
Dordrecht, Boston, London, 2003.
8. J.Hamhalter: Isomorphisms of ordered structures of abelian C ∗-subalgebras
of C ∗-algebras, Journal of Mathematical Analysis and Applications, 383,
(2011), 391-399.
9. H.Hanche-Olsen and E.Stormer: Jordan Operator Algebras, Pitman Ad-
vanced Publishing, Boston, London, Melbourne, (1984).
14
|
1703.10999 | 2 | 1703 | 2018-03-03T12:21:37 | Rokhlin dimension for compact quantum group actions | [
"math.OA",
"math.LO"
] | We show that, for a given compact or discrete quantum group $G$, the class of actions of $G$ on C*-algebras is first-order axiomatizable in the logic for metric structures. As an application, we extend the notion of Rokhlin property for $G$-C*-algebra, introduced by Barlak, Szab\'{o}, and Voigt in the case when $G$ is second countable and coexact, to an arbitrary compact quantum group $G$. All the the preservations and rigidity results for Rokhlin actions of second countable coexact compact quantum groups obtained by Barlak, Szab\'{o}, and Voigt are shown to hold in this general context. As a further application, we extend the notion of equivariant order zero dimension for equivariant *-homomorphisms, introduced in the classical setting by the first and third authors, to actions of compact quantum groups. This allows us to define the Rokhlin dimension of an action of a compact quantum group on a C*-algebra, recovering the Rokhlin property as Rokhlin dimension zero. We conclude by establishing a preservation result for finite nuclear dimension and finite decomposition rank when passing to fixed point algebras and crossed products by compact quantum group actions with finite Rokhlin dimension. | math.OA | math | MODEL THEORY AND ROKHLIN DIMENSION
FOR COMPACT QUANTUM GROUP ACTIONS
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
Abstract. We show that, for a given compact or discrete quantum group G, the class of actions of G on
C*-algebras is first-order axiomatizable in the logic for metric structures. As an application, we extend the
notion of Rokhlin property for G-C*-algebra, introduced by Barlak, Szab´o, and Voigt in the case when G is
second countable and coexact, to an arbitrary compact quantum group G. All the the preservations and rigidity
results for Rokhlin actions of second countable coexact compact quantum groups obtained by Barlak, Szab´o, and
Voigt are shown to hold in this general context. As a further application, we extend the notion of equivariant
order zero dimension for equivariant *-homomorphisms, introduced in the classical setting by the first and third
authors, to actions of compact quantum groups. This allows us to define the Rokhlin dimension of an action
of a compact quantum group on a C*-algebra, recovering the Rokhlin property as Rokhlin dimension zero. We
conclude by establishing a preservation result for finite nuclear dimension and finite decomposition rank when
passing to fixed point algebras and crossed products by compact quantum group actions with finite Rokhlin
dimension.
8
1
0
2
r
a
M
3
]
.
A
O
h
t
a
m
[
2
v
9
9
9
0
1
.
3
0
7
1
:
v
i
X
r
a
1. Introduction
The Rokhlin property is a freeness condition for actions of groups on C*-algebras. It has been intensively
studied in recent years due to, among other things, the strong implications it has on the structural properties
of fixed point algebras and crossed products. While generalizations to noncompact groups, such as Z [32], or R
[27], have been considered, the most common setting where the Rokhlin property has been studied is the one
of finite or, more generally, compact groups; see [17, 29]. In this case, it has recently been shown implicitly
[1], and explicitly in [22], that the Rokhlin property is of model-theoretic nature. Namely it corresponds to
the *-homomorphism defining the action being positively existential in the sense of first order logic for metric
structures.
Building on this work, the notion of Rokhlin property and the above-mentioned preservation results have
been generalized to the more general setting of actions of coexact compact quantum groups on C*-algebras in
[2]. This work has significantly expanded the scope of the preservation results for Rokhlin actions, paving the
way of finding several new examples of classifiable C*-algebras arising from compact quantum group actions.
Furthermore, approach from [2] has also contributed to a simplification and better understanding of the Rokhlin
property, even in the classical setting.
While very fruitful, the Rokhlin property is also quite restrictive. In order to circumvent this problem, the
notion of Rokhlin dimension has been recently introduced in the setting of actions of finite groups [28], compact
groups [19], or R [27]. In this setting, the Rokhlin property corresponds to having Rokhlin dimension equal to
zero. Furthermore, unlike Rokhlin actions, actions with finite Rohlin dimension are prevalent, even in the case
of C*-algebras that do not admit any Rokhlin action, such as the Jiang -- Su algebra.
The model-theoretic description of the Rokhlin property has been generalized to Rokhlin dimension in [22], in
terms of the notion, introduced therein, of equivariant order zero dimension for an equivariant *-homomorphism.
Such a notion subsumes the notion of positively existential equivariant *-homomorphism, which corresponds to
Date: September 6, 2018.
2000 Mathematics Subject Classification. Primary 20G42, 46L55, 54H05; Secondary 03E15, 37A55.
Key words and phrases. C*-algebra, compact quantum group, quantum group action, Rokhlin property, Rokhlin dimension,
logic for metric structures, axiomatizable class, positive existential embedding.
This work was initiated during a visit of E.G. and M.L. to the Mathematisches Forschungsinstitut Oberwolfach in August 2016,
supported by an Oberwolfach Leibnitz Fellowship of M.L. Parts of this work were carried out during a visit of E.G. and M.K. to
the California Institute of Technology in January 2017, and during a visit of E.G. and M.L. to the Centre de Recerca Matem`atica
in March 2017 in occasion of the Intensive Research Programme on Operator Algebras. The authors gratefully acknowledge the
hospitality and the financial support of all these institutions. E.G. was partially funded by SFB 878 Groups, Geometry and
Actions, and by a postdoctoral fellowship from the Humboldt Foundation. M.K. was partially supported by the NSF Grant DMS-
1700259. M.L. was partially supported by the NSF Grant DMS-1600186. This work is part of the project supported by the grant
H2020-MSCA-RISE-2015-691246-QUANTUM DYNAMICS..
1
2
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
having equivariant order zero dimension equal to zero. This new perspective has been used in [22] to recover and
extend various preservation results for fixed point algebras and crossed products of actions with finite Rokhlin
dimension.
The goal of this paper is to show how these notions and results naturally extend to the more general setting of
actions of compact quantum groups. To this purpose, we begin by showing that, for a fixed discrete or compact
quantum group G, the class of actions of G on C*-algebras (G-C*-algebras) is first-order axiomatizable in a
suitable language in the logic for metric structures. This provides a notion of positively existential G-equivariant
*-homomorphism between G-C*-algebras, as well as a notion of ultraproducts and reduced products for G-C*-
algebras consistent with the one considered in [2]. This perspective is then used to show how to remove all
coexactness and separability assumptions in the preservation and rigidity results from [2].
We then consider the natural generalization to the notion of equivariant order zero dimension for equivariant
*-homomorphisms, and define Rokhlin dimension for compact quantum group actions in terms of such a notion.
We use the perspective of order zero dimension to establish many relevant facts about actions with finite Rokhlin
dimension, some of which are new even in the well studied case of finite group actions. We conclude by showing
how the preservation results for nuclear dimension and decomposition rank under fixed point algebras and
crossed product from [19, 22, 28] admit natural extensions to this setting.
In order to make the present paper accessible to readers that are not necessarily familiar with quantum
groups, we recall all the definitions and results that we use. The interested reader can find more information in
the monographs [6, 47] and the surveys [34, 35, 39]. Both compact and discrete quantum groups are subsumed
by the more general class of locally compact quantum groups as defined and studied in [36 -- 38, 48]. While this
allows one to give a unified treatment, it is also technically more demanding. In order to make the paper more
accessible, and since all our results regard quantum groups that either compact or discrete, we will present all
the notions that we consider in these special cases. The interested reader is referred to [36 -- 38, 48] as well as
the monograph [47] for more information on locally compact quantum groups.
For convenience of the readers unfamiliar with first order logic for metric structures, we include an appendix
containing the notions and results from the logic for metric structures that are used in the present paper. We
will work in the framework of logic for metric structures with domains of quantification as considered in [15].
A good introduction to this topic is offered by the monograph [3]. A systematic study of C*-algebras from the
perspective of model theory has been undertaken [12]; see also [4, 7 -- 9, 11, 13 -- 16, 24, 25, 40, 41].
The present paper is divided into five sections, besides this introduction. In Section 2 we recall the notion of
discrete quantum group, discrete quantum group action, and show that the class of actions of a given discrete
quantum group on C*-algebras is first-order axiomatizable. The same is done in Section 3 for compact quantum
groups. Section 4 contains some results, to be used in the following sections, relating the notion of ultraproduct
of G-C*-algebras with crossed products and stabilization.
In Section 5 we introduce the notion of positive
existential embeddings and Rokhlin property for a G-C*-algebra, generalizing notions introduced in [2] when
G is coexact and second countable. The main results of [2] are then generalized to the case of an arbitrary
compact quantum group G. Finally, Section 6 contains the notion of G-equivariant order zero dimension for
morphisms between G-C*-algebras, which is used to define the Rokhlin dimension of a G-C*-algebra. Our
main preservation results for nuclear dimension and decomposition rank for G-C*-algebras with finite Rokhlin
dimension are presented here.
Given a subset X of a Banach space E, we denote by [X] the closure of the linear span of X inside E.
If A is a C*-algebra, then we let M (A) be its multiplier algebra, and by A its minimal unitization. We
canonically identify A with an essential ideal of M (A) and of A. We denote by ⊗ the injective tensor product
of Banach spaces, and the minimal tensor product of C*-algebras (which indeed coincides with the injective
tensor product as Banach spaces). The algebraic tensor product of complex algebras is denoted in this paper
by ⊙. A *-homomorphism π : A → B between C*-algebras is said to be nondegenerate if [π(A)B] = B. Given a
nondegenerate *-homomorphism π : A → M (B), we also denote by π its unique extension to a (unital, strictly
continuous) *-homomorphism π : M (A) → M (B). We say that a C*-subalgebra A of B is nondegenerate if the
inclusion map from A to B is nondegenerate. We will frequently use in the following that if A is a C*-algebra,
and I is a dense two-sided ideal in A, then I contains an increasing approximate unit for A.
Given a Hilbert space H, we let B(H) be the space of bounded linear operators on H, and K(H) be the
space of compact operators on H. Given vectors ξ, η ∈ H we denote by φξ,η the corresponding vector linear
functional φξ,η(T ) := hη, T ξi. We will frequently use the leg notation for elements in a tensor product [47,
Notation 7.1.1]. For a Hilbert space H, we let Σ ∈ B(H ⊗ H) denote the flip unitary. If T ∈ B(H ⊗ H), we let
T12, T23, T13 ∈ B(H ⊗ H ⊗ H) be defined by T12 = T ⊗ idH, T23 = idH ⊗ T , and T13 = Σ23T12Σ23 = Σ12T23Σ12.
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
3
More generally, one can similarly define the operators Ti1...ik ∈ B(H ⊗ · · · ⊗ H) for any sequence of indices
i1, . . . , ik.
2. An axiomatization of discrete quantum group actions
2.1. Discrete quantum groups. A discrete quantum group G is a C*-algebra c0(G) which is a direct sum of
full matrix algebras endowed with a nondegenerate *-homomorphism ∆ : c0(G) → M (c0(G) ⊗ c0(G)) (comulti-
plication) such that:
• (∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆;
• [(c0(G) ⊗ 1)∆(c0(G))] = [(1 ⊗ c0(G))∆(c0(G))] = c0(G) ⊗ c0(G).
The discrete quantum group G is said to be second countable if c0(G) is separable.
Let G be a discrete quantum group. Then there exist an index set Λ and finite dimensional Hilbert spaces
Hλ, for λ ∈ Λ, such that c0(G) is isomorphic to Lλ∈Λ K(Hλ). Set c0(G)λ = K(Hλ) for λ ∈ Λ, and denotes its
unit by 1λ. We identify c0(G)λ with a subalgebra of c0(G), M (c0(G)) withQλ∈Λ c0(G)λ, and M (c0(G) ⊗ c0(G))
with Qµ,ν∈Λ(c0(G)µ ⊗ c0(G)ν ).
For λ, µ, ν ∈ Λ, we fix a homomorphism ∆λ
µ,ν : c0(G)λ → c0(G)µ ⊗ c0(G)ν satisfying
for every x ∈ c0(G)λ.
∆(x) = (∆λ
µ,ν (x))µ,ν∈Λ ∈ Yµ,ν∈Λ
c0(G)µ ⊗ c0(G)ν
Remark 2.1. Let µ, ν ∈ Λ, and set Λµ,ν = {λ ∈ Λ : ∆λ
µ,ν 6= 0}. By [49, Proposition 2.2], the set Λµ,ν is finite.
Since ∆ is nondegenerate, we have
for every µ, ν ∈ Λ. Therefore the canonical extension of ∆ to a unital *-homomorphism
∆λ
µ,ν(1λ) = 1µ ⊗ 1ν
Xλ∈Λµ,ν
is defined by
c0(G)µ ⊗ c0(G)ν
∆ : Yλ∈λ
∆((xλ)λ∈Λ) =
c0(G)λ → Yµ,ν∈Λ
µ,ν(x)
∆λ
Xλ∈Λµ,ν
c0(G)µ ⊗ c0(G)ν .
∈ Yµ,ν∈Λ
µ,ν∈Λ
2.2. Discrete quantum group actions. Let G be a discrete group, and let A be a C*-algebra.
Definition 2.2. A (left) action of G on A is an injective nondegenerate *-homomorphism α : A → M (c0(G)⊗A)
such that:
(1) (∆ ⊗ id) ◦ α = (id ⊗ α) ◦ α (action condition);
(2) [(c0(G) ⊗ 1)α(A)] = c0(G) ⊗ A (density condition).
We will refer to a C*-algebra A endowed with a distinguished action of G as a G-C*-algebra.
If (A, α)
and (B, β) are G-C*-algebras, then a *-homomorphism φ : A → B is said to be G-equivariant if it satisfies
(id ⊗ φ) ◦ α = β ◦ φ. It follows from the density condition that, if (ui) is an approximate unit for A, then (α(ui))
is an approximate unit for c0(G) ⊗ A.
Let α : A → M (c0(G)⊗A) be a nondegenerate injective *-homomorphism. For λ ∈ Λ, let αλ : A → c0(G)λ ⊗A
be the coordinate function for α, so that
The conditions from Definition 2.2 can be restated as
α(a) = (αλ(a))λ∈Λ ∈ Yλ∈Λ
c0(G)λ ⊗ A.
(1') (idc0(G)µ ⊗ αν ) ◦ αµ =Pλ∈Λµ,ν
(2') [(c0(G)λ ⊗ 1)αλ(A)] = c0(G)λ ⊗ A for every λ ∈ Λ.
In turn, by Cohen's factorization theorem, (2') is equivalent to the assertion that for every λ ∈ Λ, for every
a ∈ c0(G)λ ⊗ A of norm at most 1, and every ε > 0, there exist x ∈ c0(G)λ and b ∈ A of norm at most 1 such
that k(x ⊗ 1)α(b) − ak < ε.
(∆λ
µ,ν ⊗ idA) ◦ αλ for every µ, ν ∈ Λ;
4
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
Example 2.3. If G is a classical discrete group, then one can regard G as a discrete quantum group by
considering the C*-algebra c0(G) of functions from G to C vanishing at infinity. In this case, one has that
Λ = G and c0(G)λ = C for every λ ∈ G. The comultiplication function ∆ : c0(G) → M (c0(G) ⊗ c0(G)) is
defined by setting ∆λ
µν (1) = δλ,µν for λ, µ, ν ∈ G.
Every discrete quantum group G such that c0(G) is a commutative C*-algebra arises from a classical discrete
group in this fashion.
2.3. Axiomatization. We continue to fix a discrete quantum group G. We now describe a natural (multi-
sorted) language LC*
G that has G-C*-algebras as structures. For comparison, one can refer to the language
considered in the axiomatization of operator systems from [24, Appendix B].
G has, for every n ≥ 1, sorts S(n) and C(n) to be interpreted as Mn(C)⊗ A
2.3.1. The language. The language LC*
and Mn(C), respectively. For each of these, the domains of quantifications should be interpreted as the balls
with respect to the norm centered at the origin.
The function and relation symbols consist of:
(1) function and relation symbols for the C*-algebra operations and the C*-algebra norm on Mn(C) ⊗ A
and Mn(C) for every n ∈ N;
(2) function symbols for the Mn(C)-bimodule structure on Mn(C) ⊗ A;
(3) constant symbols the elements of Mn(C);
(4) for any bounded linear map T : Mn(C) → Mm(C), function symbols C(n) → C(m) and S(n) → S(m) to
be interpreted as T and T ⊗ idA;
(5) function symbols for the canonical inclusions of each of the tensor factors in Mn(C) ⊗ A;
(6) function symbols S → S(dλ) to be interpreted as the injective *-homomorphisms αλ : A → c0(G)λ ⊗ A
that define the action.
It is clear that any G-C*-algebra can be seen as an LC*
G -structure in a canonical way.
2.3.2. The axioms. We now describe axioms for the class of G-C*-algebras in the language LC*
Such axioms are designed to guarantee the following:
G described above.
(1) the interpretations of the sort S(n) and C(n) are C*-algebras;
(2) the domains of quantifications are interpreted as the balls (see [12, Example 2.2.1]);
(3) the sorts S(n) and C(n), the symbols for the canonical inclusions of each of the tensor factors in Mn(C)⊗A,
the symbols for the maps T and T ⊗ idA for any bounded linear map T : Mn(C) → Mm(C), the symbols
for the Mn(C)-bimodule structure on Mn(C) ⊗ A, and the symbols for the elements of Mn(C), are
interpreted as they should be (see [24, Appendix C]);
(4) the interpretation of the symbols for the maps αλ that define the action are isometric *-homomorphisms;
(5) for every µ, ν ∈ Λ,
(idc0(G)µ ⊗ αν ) ◦ αµ = Xλ∈Λµ,ν
(∆λ
µ,ν ⊗ idA) ◦ αλ;
(6) for every a ∈ c0(G)λ ⊗ A of norm at most 1, and every ε > 0, there exist x ∈ c0(G)λ and b ∈ A, both
of norm at most 1, such that k(x ⊗ 1)α(b) − ak < ε.
It is clear from the discussion above that these axioms indeed axiomatize the class of G-C*-algebras. Fur-
thermore, it is easy to see that these axioms are all given by conditions of the form σ ≤ r where σ is a positive
primitive ∀∃-LC*
G -sentence and r ∈ R. Therefore, this shows that the class of G-C*-algebras is positively prim-
itively ∀∃-axiomatizable in the language LC*
G in the sense of Definition A.3. One can observe that, when G
is second-countable, the language LC*
G is separable for the class of G-C*-algebras; see Definition A.1. More
generally, the density character of LC*
G for the class of G-C*-algebras is equal to the size of Λ. It is not difficult
to verify that an LC*
G -morphism between G-C*-algebras is precisely a G-equivariant *-homomorphism, while an
LC*
G -embedding is an injective G-equivariant *-homomorphism.
2.3.3. Ultraproducts. Once G-C*-algebras are regarded as LC*
G -structures as described above, one can consider
ultraproducts of G-C*-algebras as a particular instance of ultraproducts in first-order logic for metric structures;
see [15]. It follows from Los' theorem that the ultraproduct of G-C*-algebras is again a G-C*-algebra.
More generally, one can consider reduced products with respect to an arbitrary filter F as defined in [23]. It
is easy to see that, in the particular case when F is the filter of cofinite subsets of N, then the reduced power
of a G-C*-algebra (A, α) with respect to F coincides with the G-C*-algebra (A∞, α∞) constructed in [2]. It
follows from the fact that the class of G-C*-algebras is positively primitively ∀∃-axiomatizable in the language
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
5
G together with Corollary A.6 that the reduced product of G-C*-algebras is again a G-C*-algebra. This
LC*
recovers [2, Lemma 2.6] as a particular case.
2.3.4. Other languages. Occasionally, it is useful to consider C*-algebras as structures in a language other
than the standard language for C*-algebras LC*. These other languages are useful to capture properties that
are preserved by more general classes of morphisms, rather than just *-homomorphisms, such as for example
completely positive contractive maps or completely positive contractive order zero maps. Several languages
are considered in [22, Section 3]. For each such language L, one can consider a corresponding G-equivariant
language LG. This can defined as in Subsubsection 2.3.4, by starting with the language L rather than LC*.
3. An axiomatization of compact quantum group actions
3.1. Compact quantum groups. A (reduced, C*-algebraic) compact quantum group G is given by a unital
C*-algebra C(G) endowed with a unital *-homomorphism (comultiplication) ∆ : C(G) → C(G) ⊗ C(G) and a
faithful state h : C(G) → C (the Haar state) satisfying
• (∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆,
• [∆(C(G))(1 ⊗ C(G))] = [∆(C(G))(C(G) ⊗ 1)] = C(G) ⊗ C(G), and
• (h ⊗ id) ◦ ∆ = (id ⊗ h) ◦ ∆ = h
where in the last equation we identify C with the space of scalar multiples for the identity in C(G). The first
two conditions assert that C(G) endowed with the comultiplication ∆ is a unital Hopf C*-algebra [43, Definition
2.2]. A compact quantum group is said to be second countable if C(G) is separable.
A unitary representation of G on a Hilbert space H is a unitary u ∈ M (C(G) ⊗ K(H)) satisfying (∆⊗id)(u) =
u13u23. In the following we will only consider the case when H is finite-dimensional, in which case one has that
u ∈ C (G) ⊗ B (H). We also identify the unitary representation u of G on H with the linear map H → C(G) ⊗ H
given by η 7→ u(1 ⊗ η) for all η ∈ H. A subspace K ⊆ H is invariant if u(1 ⊗ K) ⊆ C(G) ⊗ K. Direct sum
and tensor product of unitary representations is defined in the usual way. A unitary representation is called
irreducible if it has no non-trivial invariant closed subspace. Every irreducible unitary representation of G is
finite-dimensional, and every unitary representation of G is equivalent to a direct sum of irreducible unitary
representations.
We let Rep(G) be the set of unitary representations of G on finite-dimensional Hilbert spaces. For λ ∈ Rep(G)
we let uλ ∈ C(G) ⊗ K(Hλ) be the associated representation, and we denote by dλ the dimension of Hλ. We also
define Irr(G) to be the set of equivalence classes of irreducible unitary representations of G. Given λ, µ ∈ Irr(G),
we set δλ,µ = 1 if λ and µ are equivalent, and δλ,µ = 0 otherwise. For λ ∈ Rep(G) and ξ, η ∈ Hλ, set
uλ
ξ,η = (id ⊗ φξ,η)(uλ) ∈ C(G). These are called the matrix coefficients of the unitary representation λ.
If
λ ∈ Irr(G), then we fix an orthonormal basis {eλ
ej ,ek for 1 ≤ j, k ≤ dλ. For an
arbitrary λ ∈ Rep(G), we write λ as a direct sum of irreducible unitary representations λ1 ⊕ · · · ⊕ λn, and then
we consider the orthonormal basis of Hλ = Hλ1 ⊕ · · · ⊕ Hλn associated with the given orthonormal bases of
Hλ1 , . . . , Hλn .
k=1 of Hλ and set uλ
jk = uλ
k}dλ
Define O(G) to be the dense selfadjoint subalgebra of C(G) given by
O(G) =(cid:8)uλ
ξ,η : λ ∈ Rep(G), ξ, η ∈ Hλ(cid:9) = span(cid:8)uλ
ij : λ ∈ Irr(G), 1 ≤ i, j ≤ dλ(cid:9) ,
and observe that it is invariant under the comultiplication. The induced *-bialgebra structure on O(G) turns
it into a Hopf *-algebra [47, Definition 1.3.24]. The counit map ǫ on O(G) is defined by ǫ(uλ
ξ,η) = hξ, ηi, while
the antipode map S is given byS(uλ
η,ξ)∗, for ξ, η ∈ Hλ and λ ∈ Rep(G).
ξ,η) = (uλ
Remark 3.1. The C*-algebra C(G) together with its comultiplication can be recovered from O(G) as the
closure of the image of O(G) in the GNS representation induced by the restriction of the Haar state to O(G).
By [6, Theorem 1.8], for every λ ∈ Irr(G) there is a positive invertible operator Fλ ∈ K(Hλ) satisfying:
h(uλ
ξ,η(uλ
ζ,χ)∗) =
hξ, ζi hχ, Fληi
dimq(λ)
and h((uλ
ξ,η)∗uλ
ζ,χ) = (cid:10)ξ, (Fλ)−1ζ(cid:11) hχ, ηi
dimq(λ)
,
for all ξ, η ∈ Hλ, where dimq(λ) = Tr(Fλ) is the quantum dimension of λ. Furthermore, one has
h(uλ
ξ,η(uµ
ζ,χ)∗) = h((uλ
ξ,η)∗uµ
ζ,χ) = 0
whenever λ, µ are nonequivalent, for all ξ, η ∈ Hλ, and all ζ, χ ∈ Hµ. In particular, we deduce that
h(uλ
ij(uµ
kℓ)∗) = δλµδik
F λ
ℓj
dimq(λ)
.
(1)
6
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
for λ, µ ∈ Irr(G), 1 ≤ i, j ≤ dλ, and 1 ≤ k, ℓ ≤ dµ. Similarly to the classical case, to each λ, µ ∈ Irr(G) there
corresponds a contragradient representation λIrr(G) (cf. [39, Section 6]).
For λ ∈ Rep(G), set
Then C(G)λ is a finite-dimensional subspace of C(G) which is invariant under the comultiplication. If λ, µ ∈
Irr(G), then C(G)λ and C(G)µ are orthogonal with respect to the inner product defined by the Haar state h.
It follows that O(G) is equal to the algebraic direct sum of C(G)λ for λ ∈ Irr(G). For λ, µ ∈ Rep(G), we have
C(G)λ =(cid:8)uλ
ξ,η : ξ, η ∈ Hλ(cid:9) = span(cid:8)uλ
ij : 1 ≤ i, j ≤ dλ(cid:9) .
C(G)λ + C(G)µ = C(G)λ⊕µ, span(C(G)λC(G)µ) = C(G)λ⊗µ, and (C(G)λ)∗ = C(G)λ.
Define O( G) to be the space of linear functionals on C(G) of the form x 7→ h(xy) for some y ∈ O(G). For
λ ∈ Rep(G) and 1 ≤ s, m ≤ dλ, let ωλ
sm ∈ O( G) be given by
It follows from Equation 1 that
ωλ
sm(x) = dimq(λ)
dλ
Xk=1
(Fλ)−1
mkh(x(uλ
sk)∗).
ln) ◦ ∆ = δmlδµνωµ
for λ, µ ∈ Irr(G), 1 ≤ s, m ≤ dλ, and 1 ≤ l, n ≤ dλ. More generally, one can define
ln) = δµλδlsδmn and (ωλ
sm ⊗ ωµ
sm(uµ
ωλ
sn
dλ
ωλ
ξ,η = dimq(λ)
Xk=1(cid:10)(Fλ)−1ek, η(cid:11) h(x(uλ
ξ,ek )∗)
for ξ, η ∈ Hλ, and observe that(ωλ
ξ,η ⊗ ωλ
ζ,χ) ◦ ∆ = hη, ζi ωλ
ξ,χ for every ξ, η, ζ, χ ∈ Hλ.
Set P λ
ks = (ωλ
ks ⊗ id) ◦ ∆ for 1 ≤ k, s ≤ dλ. From the relations above, it follows that P λ
ln = δλµδmlP λ
sn for
λ, µ ∈ Irr(G), 1 ≤ s, m ≤ dλ, and 1 ≤ l, n ≤ dµ. Furthermore, P λ
smP µ
kk is a projection operator onto
k=1 P λ
kk is a projection onto C(G)λ.
C(G)λ,k := span(cid:8)uλ
ik : 1 ≤ i ≤ dλ(cid:9) ,
and P λ :=Pdλ
3.2. The fundamental unitaries. We now recall the important concept of fundamental unitaries associated
with a given compact quantum group G.
Definition 3.2. The fundamental unitaries associated with G are the unitary operators V, W ∈ B(L2(G) ⊗
L2(G)) defined as follows.
• The operator W is the adjoint of the unitary operator induced by the linear map with dense image
C(G) ⊗ C(G) → C(G) ⊗ C(G) given by x ⊗ y 7→ ∆(y)(x ⊗ 1).
• Similarly, V is the unitary operator induced by the linear map with dense image C(G) ⊗ C(G) →
C(G) ⊗ C(G) given by x ⊗ y 7→ ∆(x)(1 ⊗ y).
Abbreviate K(L2(G)) to KG, and identify the Banach-space dual K∗
G with the predual of B(L2(G)). The
fundamental unitaries satisfy the pentagon equations
The C*-algebra C(G) ⊆ B(L2(G)) can be recovered as the left leg of W :
V12V13V23 = V23V12
and W12W13W23 = W23W12.
C(G) = [{(id ⊗ φξ,ξ′ ) W : ξ, ξ′ ∈ L2(G)}] = [{(id ⊗ ω)(W ) : ω ∈ K∗
Similarly, the group C*-algebra C ∗(G) = c0( G) can be recovered as the right leg of W :
C ∗(G) = [{(φξ,ξ′ ⊗ id) (W ) : ξ, ξ′ ∈ L2(G)}] = [{(ω ⊗ id)(W ) : ω ∈ K∗
G}].
G}].
One can recover C(G) as the right leg of V , and c0( G) as the left leg of V . Furthermore, W belongs to
M (C(G) ⊗ c0( G)) and V belongs to M (c0( G) ⊗ C(G)).
Define normal *-homomorphisms ∆, ∆ : B(L2(G)) → B(L2(G) ⊗ L2(G)) by
∆(x) = W ∗(1 ⊗ x)W and
∆(x) = W ∗(1 ⊗ y) W ,
for all x, y ∈ B(L2(G)). The restriction of ∆ to C(G) gives the comultiplication of C(G), while the restriction
of ∆ to c0( G) gives the comultiplication of c0( G).
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
7
3.3. Compact quantum group actions. Let G be a (reduced, C*-algebraic) compact quantum group, and
let A be a C*-algebra.
Definition 3.3. A (left, continuous) action of G on A is an injective nondegenerate *-homomorphism α : A →
C(G) ⊗ A satisfying the following conditions:
(1) (∆ ⊗ id) ◦ α = (id ⊗ α) ◦ α (action condition);
(2) [(C(G) ⊗ 1)α(A)] = C(G) ⊗ A, where C(G) ⊗ 1, α(G), and C(G) ⊗ A are canonically regarded as
subalgebras of M (C(G) ⊗ A) (density condition).
If A is a C*-algebra endowed with a distinguished action α of G on A, then we refer to the pair (A, α) as a
G-C*-algebra. A canonical example of G-C*-algebra is C(G) endowed with the left translation action of G on
C(G) given by the comultiplication ∆.
Suppose now that λ ∈ Rep(G). An intertwiner between λ and α is a linear map v : Hλ → A such that
α(vξ) = (id ⊗ v)uλ(1 ⊗ ξ)
for every ξ ∈ Hλ. The space of intertwiners between λ and α is denoted by Int(λ, α). The λ-isotypical component
(or λ-spectral subspace) is the closed subspace
Aλ = {vξ : ξ ∈ Hλ, v ∈ Int(λ, α)} .
Define Eλ
ks : A → Aλ as Eλ
ks = (ωλ
relations (1) that Eλ
closed subspace of A. Indeed, one can alternatively describe Aλ as
smEµ
ks ⊗ id) ◦ α for 1 ≤ k, s ≤ dλ. It follows from the Schur orthogonality
ln = δλµδmlEsn for λ, µ ∈ Irr(G). The proof of [45, Theorem 1.5] shows that Aλ is a
Aλ = {a ∈ A : α(a) ∈ C(G)λ ⊗ A} = {a ∈ A : α(a) ∈ C(G)λ ⊗ Aλ} .
Furthermore Aλ + Aµ ⊆ Aλ⊕µ, AλAµ ⊆ Aλ⊗µ, and (Aλ)∗ = Aλ for every λ, µ ∈ Rep(G).
It is also shown in [45, Theorem 1.5] that Eλ = Pdλ
kk is a
projection onto a closed subspace Aλ,k of Aλ of dimension cα ∈ N ∪ {∞}. Such a dimension cα is called the
1,i))cλ
multiplicity of λ in the spectrum of α. Let (aλ
i=1 is a basis of Aλ,s. For
sk(aλ
s = 1, . . . , dλ, set aλ
s,i for every k = 1, 2, . . . , dλ.
From this it easily follows that (ǫ ⊗ id)α(c) = c for every c ∈ Aλ. Define O(A) = span {Aλ : λ ∈ Irr(A)}. It
is shown in [45, Theorem 1.5] that O(A) is a dense selfadjoint subalgebra of A, called Podle´s subalgebra or
algebraic core of (A, α).
i=1 be a basis of Aλ,1. Then (Eλ
s=1 uλ
kk is a projection onto Aλ. Moreover, Eλ
k,i) =Pdλ
1,i)cλ
k,i) = aλ
s1(aλ
ks⊗aλ
s,i, and α(aλ
s,i = Eλ
s1(aλ
1,i). Then Eλ
k=1 Eλ
The spectral subspace Aα := At associated with the trivial representation t of G is a nondegenerate C*-
subalgebra of A, called the fixed point algebra. The projection Et onto Aα is a faithful conditional expectation,
and Aλ is an Aα-bimodule for every λ ∈ Rep(G). The formula ha, bi = Et(a∗b) defines a (right) full Hilbert
Aα-module structure on A [6, Lemma 3.8 and Lemma 3.19], and two spectral subspaces Aλ, Aµ associated with
nonequivalent λ, µ ∈ Irr(G) are orthogonal with respect to this C*-bimodule structure.
Lemma 3.4. Let λ ∈ Rep(G). For a =Pij uλ
kXij
ij : 1 ≤ i, j ≤ dλ(cid:9) is a basis for C(G)λ.
kakC(G)⊗A = sup
that (cid:8)uλ
We close this subsection with the following lemma.
.
ijkC(G) ≤ 1
ij ⊗ aij ∈ C(G)λ ⊗ A, one has
µij aijkA : µij ∈ C, kXij
µijuλ
Proof. It is enough to observe that C(G)λ ⊗ A is endowed with the injective Banach space tensor product, and
Lemma 3.5. Let λ ∈ Irr(G).
(1) Let 1 ≤ i, j ≤ dλ. Then (P λ
ij ⊗ id) ◦ α = α ◦ Eλ
ij . Thus, (P λ ⊗ id) ◦ α = α ◦ Eλ.
(2) Set Eλ =Pdλ
j=1 Eλ
jj . For a ∈ A, one has
1 ⊗ Eλ(a) = X1≤s,t≤dλ
((uλ
st)∗ ⊗ 1)(α ◦ Eλ
st)(a).
Proof. Part (1) is straightforward, so we show (2). By linearity, we can assume that a has the form a = aλ
some 1 ≤ i ≤ cλ and 1 ≤ k ≤ dλ. Using the fact that O(G) is a Hopf *-algebra, we have
k,i for
(m ◦ (S ⊗ id) ◦ ∆)(x) = ε(x)1
8
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
for every x ∈ O(G). Therefore
1 ⊗ aλ
k,i = (ǫ ⊗ id)α(aλ
k,i) = (m ⊗ id)(S ⊗ id ⊗ id)(∆ ⊗ id)α(aλ
k,i)
(m ⊗ id)(S ⊗ id ⊗ id)(∆ ⊗ id)(ukj ⊗ aλ
j,i)
ks ⊗ uλ
sj ⊗ aλ
j,i)
uλ
sj ⊗ aλ
(m ⊗ id)(S ⊗ id ⊗ id)(uλ
= X1≤j≤dλ
= X1≤j,s≤dλ
j,i(cid:17)
= X1≤j,s≤dλ(cid:16)(cid:0)uλ
sk(cid:1)∗
⊗ 1(cid:17)
= X1≤s≤dλ(cid:16)(cid:0)uλ
sk(cid:1)∗
Xj=1
⊗ 1(cid:17) α(cid:0)aλ
= X1≤s≤dλ(cid:16)(cid:0)uλ
sk(cid:1)∗
s,i(cid:1)
= X1≤s≤dλ(cid:16)(cid:0)uλ
⊗ 1(cid:17)(cid:0)α ◦ Eλ
sk(cid:1)∗
sk(cid:1)(cid:0)aλ
k,i(cid:1)
= X1≤s,t≤dλ(cid:16)(cid:0)uλ
⊗ 1(cid:17)(cid:0)α ◦ Eλ
st(cid:1)∗
st(cid:1)(cid:0)aλ
k,i(cid:1) .
j,i
uλ
sj ⊗ aλ
dλ
This concludes the proof.
3.4. Axiomatization. Throughout this subsection, we fix a compact quantum group G. Our goal is to show
that there is a natural language LC*
G in the logic for metric structures that allows one to regard G-C*-algebra
as LC*
G -structures, in such a way that the class of G-C*-algebra is axiomatizable. We begin by describing the
language, and then present the axioms for the class of G-C*-algebras.
3.4.1. The language. The language LC*
G has the following sorts:
• a sort S, to be interpreted as the C*-algebra A where G acts;
• a sort S(0) to be interpreted as C(G);
• a sort S(1) to be interpreted as C(G) ⊗ A;
• a sort C to be interpreted as the algebra of complex numbers.
For each of the sorts above, the domains of quantifications are as follows:
n for λ ∈ Rep(G) and n ∈ N, to be interpreted as the ball of radius
• S has domains of quantification Dλ
n of Aλ ⊆ A centered at the origin;
• S(0) has domains of quantification D(0),λ
n
for λ ∈ Rep(G) and n ∈ N, to be interpreted as the ball of
radius n of C(G)λ ⊆ C(G) centered at the origin;
• S(1) has domains of quantification D(1),λ
n
for λ ∈ Rep(G) and n ∈ N, to be interpreted as the ball of
radius n of C(G)λ ⊗ Aλ ⊆ C(G) ⊗ A centered at the origin;
• C has domains of quantification Dn of n ∈ N, to be interpreted as the balls of radius n of C centered at
the origin.
The function and relation symbols consist of:
• function and relation symbols for the C*-algebra operations and C*-algebra norm of A, C(G), C(G)⊗A,
and C;
• constant symbols for the elements of C and for the elements of O(G);
• function symbols for the O(G)-bimodule structure of C(G) ⊗ A;
• for every λ ∈ Rep(G) and u ∈ C(G)λ, a function symbol S → S(1) for the canonical map A → C(G)⊗A;
• for every bounded linear map T : C(G) → C(G) that maps O(G) to itself, a function symbol S(1) → S(1)
to be interpreted as T ⊗idA : C(G)⊗A → C(G)⊗A and a function symbol S(0) → S(0) to be interpreted
as T : C(G) → C(G);
• for every ω ∈ O( G), a function symbol S(1) → S to be interpreted as the map ω ⊗ idA : C(G) ⊗ A → A
(slice maps) and a function symbol S(0) → C to be interpreted as ω : C(G) → C;
• a function symbol S → S(1) for the *-homomorphisms α : A → C(G) ⊗ A defining the action.
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
9
3.4.2. Axioms. We now describe axioms for the class of G-C*-algebras in the language LC*
will use the notation introduced in Subsection 3.1 and Subsection 3.3. Particularly, we set Eλ
for λ ∈ Rep(G) and 1 ≤ i, j ≤ dλ, and Eλ :=Pdλ
(P λ ⊗ id) ◦ Eλ. The axioms are designed to guarantee the following:
ii. We also set P λ ⊗ id :=Pdλ
i=1 Eλ
i=1 P λ
G described above. We
ij ⊗ id) ◦ α
ii ⊗ id and P λ ⊗ Eλ :=
ij := (ωλ
(1) the interpretations of the sort S, S(0), and S(1) are C*-algebras;
(2) the interpretation of the symbol for α is an isometric *-homomorphism;
(3) the sort C is interpreted as the complex numbers, and its domains are interpreted correctly;
(4) the interpretation of the domains Dλ,S
is the ball of radius n of the range of Eλ centered at the origin
n
(see also [12, Example 2.2.1]);
(5) the interpretation of the domain Dλ,S
n
is the ball of radius n of the range of P λ ⊗ Eλ centered at the
origin;
(6) the interpretation of the sort S(0) is isomorphic to C(G) as a C*-algebra;
(7) the interpretation of the sort S(1) is isomorphic to C(G) ⊗ A as a C(G)-bimodule;
(8) the norm on S(1) is the minimal (injective) tensor product norm (see Lemma 3.4);
(9) the function symbols for T : C(G) → C(G) and T ⊗ idA : C(G) ⊗ A → C(G) ⊗ A, where T is a bounded
linear map sending O(G) to itself, are interpreted correctly;
(10) the function symbols for the slice maps ωλ
(11) the interpretations of the symbols for Eλ
ij ⊗ idA : C(G) ⊗ A → A are interpreted correctly;
ij satisfy
ij Eµ
Eλ
sn = δλµδisδjn
for λ, µ ∈ Irr(G), 1 ≤ i, j ≤ dλ, and 1 ≤ s, n ≤ dµ;
(12) the interpretations of the symbols for Eλ satisfy Eλ = Eλ1 + · · · + Eλn whenever λ ∈ Rep(G) is the
direct sum of irreducible unitary representations λ1 ⊕ · · · ⊕ λn, and the fixed orthonormal basis on
Hλ = Hλ1 ⊕ · · · ⊕ Hλn is obtained from the fixed orthonormal bases of Hλ1 , . . . , Hλn ;
(13) the interpretations of the symbols for Eλ and P λ satisfy
for every λ ∈ Irr(G) and 1 ≤ i, j ≤ dλ (see part (1) of Lemma 3.5);
(14) for every λ ∈ Irr(G), we have
(P λ
ij ⊗ id) ◦ α = α ◦ Eλ
ij
1 ⊗ Eλ = X1≤s,t≤dλ
((uλ
st)∗ ⊗ 1)(α ◦ Eλ
st).
Next, we show that the axioms (1) -- (14) indeed axiomatize the class of G-C*-algebras.
Proposition 3.6. If G is a compact quantum group, then an LC*
if it is given by a G-C*-algebra.
G -structure satisfies (1) -- (14) above if and only
Proof. The discussion in Subsection 3.3 shows that any G-C*-algebra (A, α), when regarded as an LC*
G -structure,
satisfies the axioms (1) -- (14) above. We prove the converse, by showing that (13) implies that the action condition
from Definition 3.3 holds, and that (14) implies that the density condition from Definition 3.3 holds.
For every λ ∈ Irr(G) and 1 ≤ i, j ≤ dλ, in view of the axioms from (13) we have
(ωλ
ij ⊗ id ⊗ id) ◦ (id ⊗ α) ◦ α = (id ◦ α) ◦ (ωλ
= (id ◦ α) ◦ (P λ
= (ωλ
ij ⊗ id ⊗ id) ◦ α
ij ◦ id) ◦ α
ij ⊗ id ⊗ id) ◦ (∆ ⊗ id) ◦ α.
Since this is true for every λ ∈ Irr(G) and 1 ≤ i, j ≤ dλ, we conclude that (id ⊗ α) ◦ α = (∆ ⊗ id) ◦ α, which is
exactly the action condition from Definition 3.3.
For the density condition, let λ ∈ Irr(G) and a ∈ Aλ. Then
1 ⊗ a = X1≤s,t≤dλ
((uλ
st)∗ ⊗ 1)(α ◦ Eλ
st)(a) ∈ [(C(G) ⊗ 1)α(Aλ)].
Since this holds for every λ ∈ Irr(G), we have that [(C(G) ⊗ 1)α(A)] = C(G) ⊗ A, as desired.
Remark 3.7. One similarly checks that an LC*
*-homomorphism, while an LC*
G -embedding is an injective G-equivariant *-homomorphism.
G -morphism between G-C*-algebra is precisely a G-equivariant
10
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
It is easy to see that the axioms above are given by conditions of the form σ ≤ r, where σ is a positive primitive
G -sentence and r ∈ R. Therefore, the class of G-C*-algebras is positively primitively ∀∃-axiomatizable in
∀∃-LC*
the language LC*
G , in the sense of Definition A.3.
3.4.3. Freeness. We continue to fix a compact quantum group G. We recall here the notion of freeness for a
G-C*-algebra A from [10].
Definition 3.8. A G-C*-algebra (A, α) is free if [(1 ⊗ A)α(A)] = C(G) ⊗ A.
It is proved in [10, Theorem 2.9] that such a definition recovers the usual notion of freeness in topological
dynamics when G is a classical compact group and A is an abelian C*-algebra.
Example 3.9. The left translation action of G on C(G) is free. More generally, if A is any C*-algebra, then
the G-action on C(G) ⊗ A given by ∆ ⊗ idA is free as well.
The following equivalent reformulation of freeness is an easy consequence of the fact that O(A) is a dense
*-subalgebra of A, the fact that O(G) is a dense *-subalgebra of C(G), and part (1) of Lemma 3.5.
Lemma 3.10. A G-C*-algebra (A, α) is free if and only if [(1 ⊗ O(A))α(Aλ)] is equal to C(G)λ ⊗ A, for every
λ ∈ Irr(G). It follows that the class of free G-C*-algebras is definable by a uniform family of positive existential
LC*
G -formulas.
3.4.4. Ultraproducts. Regarding G-C*-algebras as structures in the language LC*
G described above provides a
natural notion of ultraproduct of G-C*-algebras, as a particular instance of the notion of ultraproduct in the
logic for metric structures. Since the class of G-C*-algebras is axiomatizable in the language LC*
G , it follows
that the ultraproduct of G-C*-algebras is a G-C*-algebra.
More generally, one can consider reduced products of G-C*-algebras with respect to an arbitrary filter F
as particular instances of reduced products of metric structures as defined in [23]. Concretely, suppose that
F is a filter on a set I, and let (Ai, αi) be an I-sequence of G-C*-algebras. For λ ∈ Rep(G), let Aλ be the
Multiplication and involution on A are induced by pointwise operations (it is important to notice here that
One can regard Aλ as a closed subspace of Aµ whenever λ, µ ∈ Rep(G) and λ is contained in µ. Therefore, the
Banach space obtained as the vector space of bounded sequences (ai) ∈Qi∈I (Ai)λ endowed with the seminorm
k(ai)k = lim supF kaik. (This is just the reduced product QF (Ai)λ of the I-sequence of Banach spaces (Ai)λ.)
union Sλ∈Rep(G) Aλ has a natural normed vector space structure. Let A denote its completion. We will write
[ai]F for the element of A corresponding to the bounded sequence (ai)i∈I in Qi∈I Ai.
(aibi) is automatically a bounded sequence in Qi∈I (Ai)λ⊗µ and (a∗
i ) is a bounded sequence in Qi∈I (Ai)λ).
with the reduced product of Banach spaces QF (C(G)λ ⊗ (Ai)λ); see for instance [26, Lemma 7.4], where
α : Aλ → QF (C(G)λ ⊗ (Ai)λ) = C(G)λ ⊗ Aλ for every λ ∈ Rep(G). The induced function α : A → C(G) ⊗ A
the case of ultraproducts is considered. Therefore, the assignment (ai)i∈I 7→ (α(ai))i∈I , induces a function
Finally, since C(G)λ is finite-dimensional for every λ ∈ Rep(G), one can isometrically identify C(G)λ ⊗ Aλ
is easily seen to be an action of G on A. In the following, we will denote such an action by αF .
Remark 3.11. In view of [2, Proposition 2.13], considering such an explicit construction of the reduced product
shows that it coincides with the continuous part of the sequence algebra as defined in [2, Definition 2.11] when
F is the filter of cofinite subsets of N and G is a second countable coexact compact quantum group (which is
the only case considered in [2, Definition 2.11]).
We now show that, when G is a classical compact group, the reduced product of G-C*-algebras as LC*
G -
structures agrees with the continuous part of the sequence algebra of a G-C*-algebra considered in the C*-
algebra literature; see, for example, [22].
Proposition 3.12. Let G be a compact group. Suppose that F is a filter on an index set I, and let {(Ai, α(i)) : i ∈
αF is continuous.
I} be G-C*-algebras. Denote by αF the (not necessarily continuous) action of G on QF Ai given by pointwise
application of the actions α(i). Then the reduced ultraproduct QG
F Ai agrees with the subalgebra of QF Ai where
Proof. For convenience, denote by B the subalgebra of QF Ai where G acts continuously. Regard QG
C*-subalgebra ofQF Ai, in such a way that the inclusionQG
G on QG
F Ai ⊆QF Ai is G-equivariant. Since the action of
F Ai is continuous, one has QG
F Ai ⊆ B. On the other hand, since the action of G on B is continuous,
F Ai as a
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
11
the *-subalgebra O(B) = Sλ∈Rep(G) Bλ consisting of the spectral subspaces for B, is dense in B. For every
λ ∈ Rep(G), and using the projections onto the spectral subspaces, one can see directly that
Bλ =YF
(Ai)λ ⊆YG
F
Ai.
Since this holds for every λ ∈ Rep(G), we have O(B) ⊆QG
F Ai and hence B ⊆QG
F Ai.
In particular, it follows from these observations that the (positive) existential theory of a G-C*-algebras
as defined in [22, Subsection 2.2] coincides with the (positive) existential theory of a G-C*-algebras as an
LC*
G -structure.
3.4.5. Other languages. As for the case of discrete quantum group actions, it is sometimes useful to consider
C*-algebras and G-C*-algebras as structures in a language that is different from the standard language for
C*-algebras. These other languages allow one to capture properties that are preserved under not necessarily
multiplicative maps. Several such languages are considered in [22, Section 3]. If L is such a language, one can
define its corresponding G-equivariant analogue LG as above, so that G-C*-algebras can also be regarded as
LG-structures.
4. Crossed products and reduced products
4.1. The dual of a compact quantum group. The dual G of a compact quantum group G is the discrete
quantum group defined as follows. By assumption, the GNS representation associated with the Haar state h on
C(G) defines a faithful representation of C(G). We denote by L2(G) the corresponding Hilbert space, and we
identify C(G) with a subalgebra of B(L2(G)). We let x 7→ xi be the canonical map from C(G) to L2(G), so that
x yi = xyi, and hx, yi = h(x∗y) for all x, y ∈ C(G). For λ ∈ Rep(G), set L2(G)λ =nuλ
Recall that O( G) denotes the space of linear functionals on C(G) of the form x 7→ h(xa) for some a ∈ O(G).
It coincides with the space of linear functionals on C(G) of the form x 7→ h(ax) for some a ∈ O(G), and it is also
equal to the linear span of ωλ
ij for λ ∈ Irr(G) and 1 ≤ i, j ≤ dλ. This is an algebra with respect to convolution.
The antipode map S of O(G) defines an involution on O( G) by setting φ∗ = φ(S(x)∗). For λ ∈ Rep(G), set
ξ,η : ξ, η ∈ Hλo ⊆ L2(G).
c0( G)λ = span(cid:8)ωλ
ξ,η : ξ, η ∈ Hλ(cid:9) ,
ξ,η 7→ Tξ,η is a *-homomorphism from c0( G)λ onto K(Hλ), where Tξ,η is the usual rank-one
and observe that ωλ
operator. Furthermore O( G) is isomorphic to the algebraic direct sum of c0( G)λ for λ ∈ Irr(G). One can
define an injective *-representation of O( G) on L2(G) by setting ω xi = (id ⊗ ω)∆(x)i for x ∈ C(G) and
ω ∈ O( G). We will identify O( G) with its image inside B(L2(G)), and let C ∗(G) = c0( G) be the closure of
O( G) inside B(L2(G)). For λ ∈ Rep(G), the subspace L2(G)λ is invariant under c0( G)λ, and the inclusion
c0( G)λ ⊆ B(L2(G)λ) is isometric. The C*-algebra C ∗(G) = c0( G) is also called the group C*-algebra of the
compact quantum group G.
The multiplication operation on C(G) defines a nondegenerate *-homomorphism ∆ : O( G) → M (O( G) ⊙
O( G)) such that (ω1 ⊗ 1) ∆(ω2) belongs to O( G) ⊙ O( G) for every ω1, ω2 ∈ O( G), and
(ω1 ⊗ 1) ∆(ω2)
∆(ω2)(1 ⊗ ω1)
:
x ⊗ y 7→ (ω1 ⊗ ω2)(∆(x)(1 ⊗ y))
: x ⊗ y 7→ (ω2 ⊗ ω1)((x ⊗ 1)∆(y)).
Such a nondegenerate *-homomorphism extends to c0( G) and defines a *-homomorphism ∆ : c0( G) → M (c0( G)⊗
c0( G)) which is also nondegenerate. This defines a discrete quantum group G, which is the dual of G.
For the purpose of defining the dual of a given action, it is convenient to consider the opposite discrete group
G of G.
Definition 4.1. Let G be a compact quantum group. Following the notation of [2], we define G to be the
discrete group such that c0( G) is equal to c0( G) as a C*-algebra, but endowed with the opposite comultiplication
∆(ω) = Σ ◦ ∆(ω) ◦ Σ.
The Hopf *-algebra O( G) is equal to O( G) but endowed with the opposite comultiplication.
12
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
4.2. The crossed product of a compact quantum group action. We fix a compact quantum group G
and a G-C*-algebra (A, α). We proceed to define the associated (reduced) crossed product.
Definition 4.2. (See [6, Definition 5.28] and [6, Proposition 5.32]). Fix a nondegenerate faithful *-representa-
tion A → B(H), under which we regard α(A) ⊆ C(G) ⊗ A as a subalgebra of B(L2(G) ⊗ H) and similarly for
c0( G) ⊗ 1. The reduced crossed product G ⋉α,r A is defined as
[(c0( G) ⊗ 1)α(A)] ⊆ B(L2(G) ⊗ H).
The crossed product is canonically endowed with an action α of G, called the dual action, which is defined by
setting
α((ω ⊗ 1)α(a)) = ( ∆(ω) ⊗ 1)(1 ⊗ α(a)) ∈ M (c0( G) ⊗ G ⋉α A) ⊆ B(L2(G) ⊗ L2(G) ⊗ H).
We will regard G ⋉α,r A as a G-C*-algebra endowed with such an action of G. Given ω ∈ c0( G) and a ∈ A,
we denote by ω ⋉ a the element (ω ⊗ 1)α(a) of G ⋉α,r A. It is shown in [6, Theorem 5.31] that the reduced
crossed product G ⋉α,r A as defined above coincides with the full crossed product in the sense of [6, Definition
5.27], and therefore can also be denoted by G ⋉α A. It follows from the universal property of the full crossed
products that, if (B, β) and (C, γ) are G-C*-algebras, then a G-equivariant *-homomorphism φ : B → C induces
a G-equivariant *-homomorphism G⋉φ : G⋉β,rB → G⋉γ,rC by setting (G⋉φ)(ω ⋉a) = ω ⋉φ(a). Furthermore,
if φ is nondegenerate, then G ⋉ φ is nondegenerate [47, Theorem 9.4.8]. Recall that a completely positive order
zero map φ : A → B between C*-algebra is order zero if, whenever a, b are positive elements of A satisfying
ab = ba = 0, then φ (a) φ (b) = φ (b) φ (a) = 0 [50, Definition 2.3]. One can then easily deduce from this and the
structure theorem for completely positive order zero maps the following lemma; see also [18, Proposition 2.3].
Lemma 4.3. Suppose that (A, α), (B, β), (C, γ) are G-C*-algebras. Assume that A ⊆ B and A ⊆ C are
G-invariant subalgebras. Let θ : B → C be a G-equivariant completely positive order zero A-bimodule map.
Then the assignment
defines a completely positive G-equivariant order zero (G ⋉α A)-bimodule map G ⋉ θ : G ⋉β B → G ⋉γ C. If
furthermore θ is a (nondegenerate) *-homomorphism, then G ⋉ θ is a (nondegenerate) *-homomorphism.
ω ⋉ b 7→ ω ⋉ θ(b)
Suppose now that F is a filter on some index set I. For every ℓ ∈ I, let (Aℓ, αℓ) be a G-C*-algebra. Fix also,
for every ℓ ∈ I, a nondegenerate faithful *-representation Aℓ → B(Hℓ). Consider the corresponding reduced
hand, one can also consider, for every ℓ ∈ I, the crossed product G ⋉αℓ Aℓ ⊆ B(L2(G) ⊗ H), and then the
reduced product of G-C*-algebras
F Aℓ ⊆QF B(Hℓ) ⊆ B(QF Hℓ), which is a G-C*-algebra endowed with the action
F Aℓ can be naturally represented on L2(G) ⊗QF Hℓ. On the other
product of G-C*-algebrasQG
αF . Then the crossed product G ⋉αF QG
(G ⋉αℓ Aℓ) ⊆YF
Y
G
F
These algebras do not coincide in general, but there is always a canonical map in one direction, as we show
next.
Proposition 4.4. Let the notation be as in the discussion above. Then the assignment θ ⋉ [a(ℓ)]F 7→ [θ ⋉ a(ℓ)]F ,
B(L2(G) ⊗ Hℓ) ⊆ B(YF
(L2(G) ⊗ Hℓ)).
F Aℓ)λ, determines an injective G-equivariant *-homomorphism
for λ ∈ Rep(G), for θ ∈ c0( G)λ, and a ∈ (QG
G ⋉αF YG
F
Aℓ →Y
G
F
(G ⋉αℓ Aℓ).
Proof. To simplify the notation, we drop the subscript F when denoting elements of a reduced product via their
representative sequences. Suppose that
for 1 ≤ i, j ≤ dλ, and set
Aℓ)λ
F
aij = [a(ℓ)
ij ] ∈ (YG
ij ] ∈ G ⋉αF YG
F
ij ⋉ [a(ℓ)
ωλ
z =Xij
Aℓ ⊆ B(L2(G) ⊗YF
H).
Let pλ ∈ B(L2(G)) be the orthogonal projection onto L2(G)λ. Then z(pλ ⊗ 1) = (pλ ⊗ 1)z(pλ ⊗ 1) and
kzk = k(pλ ⊗ 1)zpλk.
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
13
For ℓ ∈ I, set
and
ij ⋉ a(ℓ)
ωλ
ij ∈ G ⋉α A ⊆ B(L2(G) ⊗ H),
z(ℓ) =Xij
G
U
[z(ℓ)] ∈Y
(G ⋉αℓ Aℓ) ⊆YF
B(L2(G) ⊗ H) ⊆ B(YF
(L2(G) ⊗ H)).
Then [z(ℓ)]F (pλ ⊗ 1) = (pλ ⊗ 1)[z(ℓ)](pλ ⊗ 1) and k[z(ℓ)]k = k(pλ ⊗ 1)[z(ℓ)](pλ ⊗ 1)k. Since pλ is a finite-rank
projection with range L2(G)λ, we have
(pλ ⊗ 1)[z(ℓ)](pλ ⊗ 1) ∈ (pλ ⊗ 1)B(YF
We conclude that
(L2(G) ⊗ H))(pλ ⊗ 1) = B(L2(G)λ ⊗YF
H).
k[z(ℓ)]k = k(pλ ⊗ 1)[z(ℓ)](pλ ⊗ 1)k = lim sup
k(pλ ⊗ 1)z(ℓ)(pλ ⊗ 1)k
= k(pλ ⊗ 1)z(pλ ⊗ 1)k = kzk.
F
This shows that the assignment θ ⋉ [a(ℓ)] 7→ [θ ⋉ a(ℓ)] yields a well-defined isometric linear map
Φ : G ⋉αF YG
F
Aℓ →Y
G
F
(G ⋉α Aℓ).
The fact that Φ is a G-equivariant *-homomorphism can be verified directly by means of the expression for
the multiplication and involution in the crossed product, together with the definition of the dual action. This
finishes the proof.
Suppose now that α is a continuous action of G on a C*-algebra A. Fix a nondegenerate faithful *-
representation A → B(H). One defines the reduced crossed product G ⋉α,r A to be
G ⋉α,r A = [(C(G) ⊗ 1)α(A)] ⊆ B(L2(G) ⊗ H),
As before, we denote the element (x ⊗ 1)α(a) of G ⋉α,r A by x ⋉ a. The reduced crossed product is endowed
with a canonical action of G (dual action) defined by the *-homomorphism α : G ⋉α,r A → C(G) ⊗ ( G ⋉α,r A),
x ⋉ a 7→ (∆(x) ⊗ 1)(1 ⊗ a). The reduced crossed product construction is functorial, and a (nondegenerate)
G-equivariant *-homomorphism φ induces a (nondegenerate) G-equivariant *-homomorphism G ⋉ φ between
the reduced crossed products. This allows one to prove the analogue of Lemma 4.3 in this context. It is clear
that, for λ ∈ Rep(G), the corresponding spectral subspace of G ⋉α,r A is just
( G ⋉α,r A)λ = span{ω ⋉ a : ω ∈ C(G)λ, a ∈ A}.
Lemma 4.5. Suppose that G is a compact quantum group, F is a filter on a set I, and (Aℓ, αℓ) is a G-C*-algebra
for every ℓ ∈ I. The assignment of
u ⋉ [a(ℓ)]F 7→ [u ⋉ a(ℓ)]F
for λ ∈ Rep(G), θ ∈ C(G)λ, and a ∈Q
G
F Aℓ defines an injective G-equivariant *-homomorphism
G ⋉αF ,r Y
G
F
Aℓ →YG
F
( G ⋉αℓ,r Aℓ).
Proof. This is easy to see directly, using the fact that C(G)λ is finite-dimensional for every λ ∈ Rep(G).
Remark 4.6. Consider the particular case of Lemma 4.5 when (Aℓ, αℓ) is equal to a fixed G-C*-algebra (A, α)
F ( G ⋉α,r A) has the property
for every ℓ ∈ I. Then the G-equivariant *-isomorphism θ : G ⋉αF ,r Q
F A → QG
that
G
θ ◦ ( G ⋉ ∆A) = ∆ G⋉α,rA,
where ∆A : A →Q
F A and ∆ G⋉α,rA : G ⋉α,r A →QG
G
F ( G ⋉α,r A) are the diagonal embeddings.
14
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
4.3. Stabilizations. We continue to fix a compact quantum group G. Let (A, α) is a G-C*-algebra. Fix a
nondegenerate faithful representation A → B(H). Let V ∈ M (c0( G) ⊗ C(G)) ⊆ B(L2(G) ⊗ L2(G)) denote the
multiplicate unitary associated with G, and set X = Σ ◦ V ◦ Σ ∈ M (C(G) ⊗ c0( G)) ⊆ B(L2(G) ⊗ L2(G)).
Definition 4.7. ([43, Section 2]). Adopt the notation from the discussion above. The stabilization of (A, α) is
the G-C*-algebra (KG ⊗ A, αK), where αK : KG ⊗ A → C(G) ⊗ KG ⊗ A is the *-homomorphism given by
αK(T ⊗ a) = X ∗
12(1 ⊗ T ⊗ 1)α(a)13X12 ∈ C(G) ⊗ KG ⊗ A ⊆ B(L2(G) ⊗ L2(G) ⊗ H).
In the following, given a G-C*-algebra (A, α), we regard KG ⊗ A as a G-C*-algebra with respect to αK.
When G is a classical compact group, αK is just the diagonal action, where KG is endowed with the action
of G by conjugation induced by the left regular representation.
The assignment (A, α) 7→ (KG ⊗ A, αK) is functorial, in the sense that any *-homomorphism φ : (A, α) →
(B, β) induces a *-homomorphism id⊗φ : (KG ⊗ A, αK) → (KG ⊗ B, βK).
Proposition 4.8. Let F be a filter on a set I, and let (Aℓ, α(ℓ))ℓ∈I be G-C*-algebras. Then the assignment
T ⊗ [a(ℓ)]F 7→ [T ⊗ a(ℓ)]F defines a G-equivariant *-homomorphism
F
(KG ⊗ Aℓ).
Aℓ →YG
F Aℓ coincides with the maximal tensor product. Therefore the
F (KG ⊗ Aℓ) in view of
the universal property of the maximal tensor product. A straightforward computation shows that such a map
F Aℓ is endowed with the stabilization of the reduced product action αF , and
KG ⊗YG
Proof. Observe that tensor product KG ⊗QG
assignment in the statement gives a well-defined *-homomorphism ψ : KG ⊗QG
is G-equivariant when KG ⊗QG
QG
F (KG ⊗ Aℓ) is endowed with the reduced product of the stabilizations (α(ℓ)
Denote by 1 ∈ C(G) the unit, and let 1i ∈ L2(G) be the corresponding vector, and 1i h1 ∈ KG the
F Aℓ →QG
F )ℓ∈I.
F
corresponding rank one projection.
Lemma 4.9. The C*-subalgebra 1i h1 ⊗ A ⊆ KG ⊗ A is a G-invariant C*-subalgebra of (KG ⊗ A, αK), and
the injective *-homomorphism a 7→ 1i h1 ⊗ a is G-equivariant.
Proof. The invariance property of the Haar state h can be written as h(y)1 = Pi h(y0,i)y1,i = Pi h(y1,i)y0,i
where ∆(y) = Pi y0,i ⊗ y1,i. We claim that (1 ⊗ 1i h1)X = 1 ⊗ 1i h1. Indeed for every x, y ∈ C(G) we have
that
(1 ⊗ 1i h1)X x ⊗ yi = (1 ⊗ 1i h1)ΣV y ⊗ xi = (1 ⊗ 1i h1)Σ ∆(y)(1 ⊗ x)i
= (1 ⊗ 1i h1) ∆op(y)(x ⊗ 1)i =Xi
=Xi
h(y0,i) y1,ix ⊗ 1i = h(y) x ⊗ 1i = (1 ⊗ 1i h1) x ⊗ yi .
(1 ⊗ 1i h1) y1,ix ⊗ y0,ii
Henceforth
αK(1i h1 ⊗ a) = X12(1 ⊗ 1i h1 ⊗ 1)α(a)13X12 = X12(1 ⊗ 1i h1 ⊗ 1)α(a)13(1 ⊗ 1i h1 ⊗ 1)X12
= (1 ⊗ 1i h1 ⊗ 1)α(a)13(1 ⊗ 1i h1 ⊗ 1) = (1 ⊗ 1i h1 ⊗ 1)α(a)13.
This concludes the proof.
In the following lemma, we consider the ordered selfadjoint operator space language Losos as introduced in
[22, Subsection 3.1], as well as its A-bimodule version Losos,A-A as introduced in [22, Subsection 3.5]. Recall
that the language Losos,A-A does not have a distinguished relation symbol for the metric. Instead, for every
finite subset F of A, the language Losos,A-A contains a distinguished pseudometric symbol dF , to be interpreted
in KG ⊗ A as the pseudometric
dA
F (x, y) = sup {k(1 ⊗ a)(x − y)k, k(x − y)(1 ⊗ a)k : a ∈ F } .
G
G
and Losos,A-A
Given a compact or discrete quantum group G, one can add symbols for the G-action to obtain languages
Losos
If A is a G-C*-algebra, then KG ⊗ A can be regarded as structure in the language of
LC*,KG⊗A-KG⊗A
G
Fix an increasing approximate unit (uj)j∈J for A contained in Aα. Suppose now that B is a G-C*-algebra
containing A as a nondegenerate G-C*-subalgebra. In particular the approximate unit (uj)j∈J for A is also
.
.
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
15
an approximate unit for B. The notion of positively quantifier-free definable substructure in a language L is
recalled in Definition A.14.
Lemma 4.10. Adopt the notation and assumptions from the discussion above. The G-C*-subalgebra 1i h1⊗B
of KG ⊗ B is a positively quantifier-free Losos,KG⊗A-KG⊗A
-definable substructure relative to the class of G-C*-
algebras of the form KG ⊗ B.
G
Proof. We have already observed that 1i h1 ⊗ B is indeed a G-C*-subalgebra of KG ⊗ B. The fact that it is
positively quantifier-free Losos,KG⊗A-KG⊗A
-definable is witnessed by the formulas ϕj,F (x) defined by
G
where (uj)j∈J is the fixed approximate unit for Aα, and F ranges among the finite subsets of KG ⊗ A.
ka((1i h1 ⊗ uj)x(1i h1 ⊗ ui) − x)k
sup
a∈F
In the statement of Lemma 4.10, it is important that the G-C*-algebra KG ⊗ B is regarded as a structure in
the language Losos,KG⊗A-KG⊗A
G
, rather than a structure in the language Losos
G .
Remark 4.11. A similar discussion as above can be done when (A, α) is a G-C*-algebra. Fix a nondegenerate
faithful representation A → B(H). Set X = V ∗ ∈ M (c0( G) ⊗ C(G)). We define the stabilization of (A, α) as
the G-C*-algebra (KG ⊗ A, αK), where the action αK is defined by
αK(T ⊗ a) = X ∗
12(1 ⊗ T ⊗ 1)α(a)13 X12 ∈ c0( G) ∈ KG ⊗ A ⊆ B(L2(G) ⊗ L2(G) ⊗ H).
5. Existential embeddings and the Rokhlin property
G -structures gives the notion of positively LC*
5.1. Existential embeddings. Let G be either a compact or discrete quantum group. Considering G-
C*-algebras as LC*
G -existential *-homomorphism between G-C*-
algebras; see Definition A.10. When the algebras are separable, and the group is second countable and either
compact and coexact or discrete and exact (which is the only case considered in [2]), a G-equivariant homo-
morphism is positively LC*
G -existential if and only if it is sequentially split in the sense of [2, Definition 3.1]; see
A.11. In this section, we show that the results from [2], phrased in terms of positive existential embeddings,
can be obtained without any assumptions on the algebras or the group.
Positive existential *-homomorphisms are preserved by functors under general assumptions. If G is a com-
pact or discrete quantum group, we regard G-C*-algebras as the objects of a category with G-equivariant
U A the canonical
*-homomorphisms as morphisms. In the following proposition, we denote by ∆A : A → QG
G -embedding of a G-C*-algebra into the corresponding ultrapower.
diagonal LC*
Proposition 5.1. Let G0 and G1 be either compact or discrete quantum groups, let F be a functor from the
category of G0-C*-algebras to the category of G1-C*-algebras. Assume that for any index set I, for any countably
incomplete ultrafilter U over I, and for any G0-C*-algebra A, there exists a G1-equivariant *-homomorphism
G0 -existential *-homomorphisms to
U F (A) such that θ ◦ F (∆A) = ∆F (A). Then F maps LC*
θ : F (QG0
U A) →QG1
LC*
G1 -existential *-homomorphisms.
Proof. This is immediate using the semantic characterization of positively existential morphisms from Proposi-
tion A.11.
Let φ : (A, α) → (B, β) be an injective nondegenerate *-homomorphism. We identify A with a G-invariant
subalgebra of B via φ. Suppose that A0 is a G-C*-subalgebra of A. Both A and B have a natural A0-bimodule
structure, and hence can be regarded as structures in the language LC*,A0-A0
[22, Subsection 3.5]. Recall that
this is obtained from the language of G-C*-algebras by adding function symbols for the A0-bimodule structure,
and replacing the distinguished relation symbol for the metric with pseudometric function symbols dF where F
ranges among the finite subsets of A0. Then dF is interpreted in B as the pseudometric
G
dF (x, y) = sup {ka(x − y)k, k(x − y)ak : a ∈ F } .
Proposition 5.2. Let G be a compact or discrete quantum group, let (A, α) and (B, β) be G-C*-algebras, and
let φ : A → B be a nondegenerate injective G-equivariant *-homomorphism. Fix a G-C*-subalgebra A0 of A
containing an approximate unit for A. Then φ is LC*
G -existential if and only if it is LC*,A0-A0
-existential.
G
Proof. As remarked above, one can assume that φ : A → B is the inclusion map. The forward implication is
obvious. The converse implication is easily shown using an increasing approximate unit for A contained in A0,
which is also an approximate unit for B since the inclusion A ⊆ B is nondegenerate by assumption.
16
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
A reason to consider the notion of positive LC*
G -existential *-homomorphism is that it allows one to conclude
that several properties pass from the target algebra to the domain algebra; see also [1]. The following proposition
is just a special instance of Proposition A.12.
Proposition 5.3. Let G be a compact or discrete quantum group. Suppose that C is a class of G-C*-algebras
that is definable by a uniform family of positive existential LC*
G -formulas. Suppose that (A, α) and (B, β) are
G-C*-algebras, and φ : (A, α) → (B, β) is an LC*
G -existential *-homomorphism. If (B, β) belongs to C, then
(A, α) belongs to C.
The following result is established in [2, Proposition 3.3] when G is compact and coexact or discrete and
exact, and second countable. Here, we remove these assumptions.
Proposition 5.4. Let G be a compact or discrete quantum group, and let (A, α) and (B, β) be G-C*-algebras.
Suppose that φ : A → B is a nondegenerate G-equivariant *-homomorphism. If φ is positively LC*
G -existential,
then G ⋉ φ : (G ⋉α,r A, α) → (G ⋉β,r B, β) is positively LC*
G -existential.
Proof. When G is compact, this is a consequence of Proposition 5.1 and Proposition 4.4. When G is discrete, this
follows from Lemma 4.5 and Remark 4.6, together with the semantic characterization of positively existential
embeddings from Proposition A.11.
The following result is established in [2, Proposition 3.6, Proposition 3.7, Proposition 3.8] when G is coexact,
and second countable. Here, we remove these assumptions, and provide a simpler proof.
Proposition 5.5. Let G be a compact quantum group, let (A, α) and (B, β) be G-C*-algebras, and let φ : A → B
be a nondegenerate G-equivariant *-homomorphism.
G -existential, then φAα : Aα → Bβ is positively LC*-existential.
G -existential if and only if id ⊗ φ : (KG ⊗ A, αK) → (KG ⊗ B, βK) is
(1) If φ is positively LC*
(2) φ : (A, α) → (B, β) is positively LC*
positively LC*
G -existential.
existential.
(3) φ is positively LC*
G -existential if and only if G ⋉ φ : (G ⋉α,r A, α) → (G ⋉β,r B, β) is positively LC*
G -
Proof. (1): This is an immediate consequence of Proposition A.15, after observing that the fixed point algebra
of a G-C*-algebra is an LC*
G -definable G-C*-subalgebra.
(2): We can assume that A ⊆ B and φ : A → B is the inclusion map. The forward implication is a
consequence of Proposition 5.3 and Proposition 4.8. For the converse, observe that we can identify A with the
G-C*-subalgebra 1i h1 ⊗ A of KG ⊗ A. A similar observation applies to B, so we identify id ⊗ φ with the
inclusion map KG ⊗ A ⊆ KG ⊗ B, and φ with the restriction of id ⊗ φ to 1i h1 ⊗ A. We can regard KG ⊗ A
and KG ⊗ B as KG ⊗ A-bimodules, and hence as structures in the language LC*,KG⊗A-KG⊗A
. Observe that
id ⊗ φ is nondegenerate. It follows from Proposition 5.2 that id ⊗ φ is positively LC*,KG⊗A-KG⊗A
-existential.
Furthermore, by Lemma 4.10, 1i h1 ⊗ A ⊆ KG ⊗ A and 1i h1 ⊗ B ⊆ KG ⊗ B are positively quantifier-
free LC*,KG⊗A-KG⊗A
-existential by Proposition A.15. In
particular, φ is positively LC*,A-A
-existential. Since φ is nondegenerate, a further application of Proposition 5.2
shows that φ is positively LC*
-definable. Therefore φ is positively LC*,KG⊗A-KG⊗A
G
G -existential. This concludes the proof.
G
G
G
G
(3): The forward implication is a consequence of Proposition 5.4. The converse implication follows from the
other implication and Item (2) above, in view of the Baaj -- Skandalis -- Takesaki -- Takai duality for compact and
discrete quantum groups; see [2, Theorem 1.20], [47, Chapter 9], [6, Theorem 5.33].
5.2. The Rokhlin property. A generalization of the Rokhlin property for actions of classical compact groups
on separable C*-algebras has been considered in [2, Section 4] for coexact second countable compact quantum
groups. Here, we remove all separability and coexactness assumptions. We fix a compact quantum group G
throughout the rest of this section.
Definition 5.6. A G-C*-algebra (A, α) is said to have the (spatial) Rokhlin property if the map α : (A, α) →
(C(G) ⊗ A, ∆ ⊗ idA) is positively LC*
G -existential.
It follows from [2, Lemma 1.24] and Proposition A.11 that Definition 5.6 agrees with [2, Definition 4.1] in the
particular case when G is coexact and second countable, and A is separable (which is the only case considered
in [2, Definition 4.1]).
The following result generalizes [2, Proposition 4.5], with a simple and conceptual proof.
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
17
Proposition 5.7. Suppose that G is a compact quantum group, and (A, α) is a G-C*-algebra with the Rokhlin
property. Then (A, α) is free.
Proof. Observe that (C(G) ⊗ A, ∆ ⊗ idA) is a free G-C*-algebra. If (A, α) has the Rokhlin property, then (A, α)
is free in view of this observation and Proposition 5.3.
The main result of [2] asserts that, whenever G is a coexact second countable compact quantum groups and
(A, α) is a separable G-C*-algebra, then several properties of A are preserved under taking crossed products or
passing to the fixed point algebra. One can deduce the natural generalization of such a statement to arbitrary
compact quantum groups from the properties of positively existential embeddings established above.
Theorem 5.8. Let (A, α) be a G-C*-algebra with the Rokhlin property. Then the canonical nondegenerate
inclusion Aα ֒→ A is positively LC*-existential, and the canonical G-equivariant nondegenerate inclusion G ⋉α
G -existential. Here K G ⊗ A is regarded as a G-C*-algebra endowed with the
A ֒→ K G ⊗ A is positively LC*
stabilization of the trivial action of G on A.
Proof. If (A, α) has the Rokhlin property, then by definition α : (A, α) → (C(G) ⊗ A, ∆ ⊗ idA) is positively
LC*
G -existential. By part (1) of Proposition 5.5, the restriction of α to the fixed point algebras is positively
LC*-existential. Observing that the fixed point algebra of (C(G) ⊗ A, ∆ ⊗ idA) is equal to 1 ⊗ A ⊆ C(G) ⊗ A,
we deduce that the embedding Aα ֒→ 1 ⊗ A is positively LC*-existential. This concludes the proof of the first
assertion.
By Proposition 5.4, the positively LC*
G -existential *-homomorphism α : (A, α) → (C(G) ⊗ A, ∆ ⊗ idA) induces
a positively LC*
G -existential *-homomorphism G ⋉ α : G ⋉α A → G ⋉∆⊗idA (C(G) ⊗ A). A particular instance of
the Baaj -- Skandalis -- Takesaki -- Takai duality for compact quantum groups -- see [2, Theorem 1.20], [47, Chapter
9], [6, Theorem 5.33] -- gives that there exists a G-equivariant *-isomorphism ρ : G ⋉∆⊗idA (C(G)⊗ A) → KG ⊗ A
such that ρ ◦ (G ⋉ α) : G ⋉α A → KG ⊗ A is the canonical inclusion. Therefore we conclude that the latter
*-homomorphism is positively LC*
G -existential as well.
As an application, we extend several preservations results for crossed products by actions with the Rokhlin
property; see [17] and [1]. When G is coexact and second countable, this recovers the main result of [2], although
the assertions concerning real rank and stable rank are new even in this case.
Corollary 5.9. Let (A, α) be a G-C*-algebra with the Rokhlin property. If A satisfies any of the following
properties, then so do the fixed point algebra Aα and the crossed product G ⋉α A:
(1) being simple;
(2) being separable, nuclear, and satisfying the UCT;
(3) being separable and D-absorbing for a given strongly self-absorbing C*-algebra D or for D = K(H);
(4) being expressible as a direct limit of certain weakly semiprojective C*-algebras (see Theorem 3.10 in [17]
for the precise statement). This includes UHF-algebras (or matroid algebras), AF-algebras, AI-algebras,
AT-algebras, countable inductive limits of one-dimensional NCCW-complexes, and several other classes.
(5) having nuclear dimension at most n;
(6) having decomposition rank at most n;
(7) having real rank at most n;
(8) having stable rank at most n.
Proof. In view of Theorem 5.8, the canonical inclusions Aα ֒→ A and G ⋉α A ֒→ KG ⊗ A are positively
LC*
G -existential. Items (1),(3),(4),(5),(8),(9) can be obtained by observing that the corresponding properties
are definable by a uniform family of LC*
G -formulas as shown in [12, Theorem 2.5.1, Theorem 2.5.2] and [22,
Proposition 4.3]; see also [22, Remark 3.2]. The items (2), (6), (7) can alternatively be obtained by applying the
fact that the corresponding properties are definable by a uniform family of LC*,nuc
-formulas as shown in [12,
Section 5]; see also [22, Remark 3.5]. The language LC*,nuc
is the nuclear language for C*-algebras introduced
in [22, Subsection 3.3].
G
G
Finally, the K-theory formula for fixed point algebras of Rokhlin actions of finite groups from [29] generalizes
to the setting of Rokhlin actions of compact quantum groups. This has been shown in [2, Theorem 5.11] for
separable coexact compact quantum groups, but the proof applies equally well in general.
Theorem 5.10. Let G be a compact quantum group, and (A, α) be a G-C*-algebra. If (A, α) has the Rokhlin
property, then the canonical inclusion Aα ֒→ A induces an injective morphism K∗(Aα) ֒→ K∗(A) in K-theory,
18
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
whose range is
where ιA : A → C(G) ⊗ A, a 7→ 1 ⊗ a is the trivial action of G on A.
{x ∈ K∗(A) : K∗(α)(x) = K∗(ιA)(x)} ,
In fact, under the assumptions of Theorem 5.10, one can conclude that the inclusion of K0(Aα) into K0(A)
is positively existential. Here we regard K0-group as structures in the language of dimension groups (G, +, u)
endowed with domains of quantifications to be interpreted as the subsets {x ∈ G : − nu ≤ x ≤ nu} for n ∈ N.
This corresponds to the notion of ultrapower of dimension groups considered in [46].
5.3. Rigidity. Suppose that G is a compact quantum group, and (B, β) is a G-C*-algebra. Let B be the
minimal unitization of B. Then the G-action β has a unique extension to a G-action β on B. It is easy to
see that if a *-homomorphism φ : A → B is positively LC*
G -existential, then its unique extension to a unital
*-homomorphism φ : A → B is positively LC*,1
G by
adding a constant symbols for the multiplicative unit. The following definition is introduced in the setting of
compact quantum group actions in [2, Definition 5.1].
G -existential, where LC*,1
is the language obtained from LC*
G
Definition 5.11. Let G be a compact quantum group, (A, α), (B, β) be two G-C*-algebras, and φ1, φ2 : A →
B be G-equivariant *-homomorphisms. Then φ1, φ2 are approximately G-unitarily equivalent, in formulas
φ1 ≈u,G φ2, if there exists a net (vi) of unitary elements of the fixed point algebra Bβ such that φ2 is the limit
of Ad(vi) ◦ φ1 is the topology of pointwise norm convergence.
When B is separable, one can replace nets with sequences in Definition 5.11.
In the case when G is the
trivial group, Definition 5.11 recovers the usual notion of approximate unitary equivalence φ1 ≈u φ2 for the
*-homomorphisms φ1, φ2.
A rigidity result for Rokhlin actions of coexact second countable compact quantum groups, generalizing
results for finite and compact from [20, 22, 29, 42] and for finite quantum groups from [33], has been obtained
in [2, Theorem 5.10]. In the rest of this section, we observe here that such a result holds for arbitrary (not
necessarily coexact) second countable compact quantum groups.
Theorem 5.12. Let G be a second countable compact quantum group, let A be a separable C*-algebra, and let
α(0), α(1) be G-actions on A. If (A, α(0)) and (A, α(1)) have the Rokhlin property, then α(0) ≈u α(1) if and only
if α(0) ≈u,G α(1).
One can deduce from Theorem 5.12 the following corollary, similarly as Proposition [2, Proposition 6.4] is
deduced from [2, Theorem 5.10].
Corollary 5.13. Let G be a second countable compact quantum group, and D be a strongly self-absorbing
C*-algebra. Then there exists at most one conjugacy class of G-actions on D with the Rokhlin property.
The rest of this subsection is dedicated to the proof of Theorem 5.12.
We fix separable G-C*-algebras (A, α) and (B, β), and homomorphisms φ, φ1, φ2 : (A, α) → (B, β). We will
use tacitly the fact that the unitary group of a unital C*-algebras is positively quantifier-free LC*,1-definable
with respect to the class of unital C*-algebras. Indeed, it is the zeroset of the stable positive quantifier-free
LC*,1-formula max {kxx∗ − 1k, kx∗x − 1k}.
Lemma 5.14. Suppose that β ◦ φ ≈u (id ⊗ φ) ◦ α, and that β has the Rokhlin property. Then for every finite
subset F of A and every ε > 0, there exists a unitary v in the unitization of C(G) ⊗ B such that
and
k(β ◦ Ad(v) ◦ φ − (id ⊗ (Ad(v) ◦ φ)) ◦ α)(x)k < ε
sup
x∈F
(k[φ(x), v]k − k(β ◦ φ − (id ⊗ φ) ◦ α)(x)k) < ε.
sup
x∈F
Proof. Fix a finite subset F of A and ε > 0. By [2, Lemma 5.4], there exists a unitary v in the unitization of
C(G) ⊗ B such that
k((∆ ⊠ β) ◦ Ad(v) ◦ (1 ⊠ φ) − (id ⊗ (Ad(v) ◦ (1 ⊠ φ))) ◦ α)(x)k < ε
sup
x∈F
and
(k[(1 ⊠ φ)(x), v]k − k(β ◦ φ − (id ◦ φ) ◦ α)(x)k) < ε.
sup
x∈F
(2)
(3)
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
19
Since β has the Rokhlin property, the map
is positively LC*
G -existential. Therefore its unique unital extension
1 ⊠ idB : ( B, β) → (C(G) ⊠ B, ∆ ⊠ β)
1 ⊠ idB : (B, β) → (C(G) ⊠ B, ∆ ⊠ β)
is positively LC*,1
quantifier-free LC*,1
G -existential. Since the conditions from Equation 2 and Equation 3 can be expressed by positive
G -formulas, we conclude that there exists a unitary v ∈ B as wanted.
Lemma 5.15. Suppose that β ◦ φ ≈u (id ⊗ φ) ◦ α, and β has the Rokhlin property. Then there exists a
G-equivariant *-homomorphism ψ : (A, α) → (B, β) such that ψ ≈u φ.
Proof. One replaces [2, Lemma 5.5] with Lemma 5.14 in the proof [2, Proposition 5.6].
We prove the following lemma directly using the definition of positively LC*
G -existential G-equivariant *-
homomorphism.
Lemma 5.16. Suppose that (C, γ) is a G-C*-algebra, and ψ : (B, β) → (C, γ) is positively LC*
G-equivariant *-homomorphism. If ψ ◦ φ1 ≈u,G ψ ◦ φ2 then φ1 ≈u φ2.
G -existential
Proof. Observe that the unital extension ψ : ( B, β) → ( C, γ) is a positively LC*,1
G -existential G-equivariant *-
homomorphism. Fix ε > 0 and a finite subset F of A. Since by assumption ψ ◦ φ1 ≈u,G ψ ◦ φ2, there exists
v ∈ Cγ such that
kv(ψ ◦ φ1)(x)v∗ − (ψ ◦ φ2)(x)k < ε.
sup
x∈F
Since ψ is a positively LC*
supx∈F kvφ1(x)v∗ − φ2(x)k < ε. This concludes the proof.
G -existential G-equivariant *-homomorphism, there exists a unitary u ∈ Bβ such that
The following proposition in the case when G is coexact is [2, Corollary 5.9]. The proof is analogous, where
one replaces [2, Proposition 5.8] with Lemma 5.16.
Proposition 5.17. Suppose that G is a second countable compact quantum group, and (A, α) and (B, β) are
separable G-C*-algebras. Let φ1, φ2 : (A, α) → (B, β) be G-equivariant *-homomorphisms. If φ1 ≈u φ2 and
(B, β) has the Rokhlin property, then φ1 ≈u,G φ2.
Finally, using Proposition 5.17 instead of [2, Corollary 5.9], and Lemma 5.15 instead of [2, Corollary 5.6],
one can prove Theorem 5.12 reasoning as in the proof of [2, Theorem 5.10].
As fruitful as the Rokhlin property is, it is also very rare.
In fact, there are many very interesting C*-
algebras that do not admit any Rokhlin action of a nontrivial compact quantum group. For example, we have
the following result. For θ ∈ R \ Q, we denote by Aθ the irrational rotation algebra. Also, we write O∞ for the
Cuntz algebra on infinitely many generators.
Proposition 5.18. Let G be a nontrivial finite quantum group and let θ ∈ R \ Q. There do not exist any
actions of G on either Aθ or O∞ with the Rokhlin property.
Proof. For classical finite groups, this is well known; see, for example, Section 3 in [19]. Suppose that G is not
classical finite quantum group. Then C(G) is a finite dimensional C*-algebra which is not commutative. Find
n ∈ N, with n > 1, and orthogonal projections p1, . . . , pn ∈ C(G) which add up to 1C(G) and are unitarily
equivalent in C(G).
Let A be either Aθ or O∞. Suppose that α is an action of G on A, and assume by contradiction that α has
the Rokhlin property. Therefore α : (A, α) → (C (G) ⊗ A, ∆ ⊗ idA) is positively LC*
G -existential. By considering
the projections pj ⊗ 1A ∈ C(G) ⊗ A, we deduce that there exist orthogonal projections q1, . . . , qn ∈ A which add
up to 1A and are unitarily equivalent in A. In the case of O∞, this would imply that the class of unit of O∞
in its K0-group is divisible by n > 1, which is not true. For the case of Aθ, and denoting its unique trace by τ ,
j=1 τ (qj ) = nτ (q1), since unitarily equivalent projections have the same value on
traces. The range of τ on traces is known not to contain any rational which is not an integer, so this is again a
contradiction. This finishes the proof.
we would get 1 = τ (1Aθ ) =Pn
Nonexistence results like the one just explained are the main motivation for introducing a more flexible
notion, the Rokhlin dimension, which is the content of the next section. Despite not admitting any Rokhlin
action, the algebras O∞ and Aθ have many actions with finite Rokhlin dimension; see, for instance, [19].
20
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
6. Order zero dimension and Rokhlin dimension
6.1. Order zero dimension. The notion of positive LC*
G -existential G-equivariant *-homomorphism admits
a natural generalization, which has been introduced in the classical setting in [22, Section 5]. We consider
here its natural extension to compact quantum groups. In the following definition, given a *-homomorphism
θ : A → B, we consider B as an A-bimodule, with respect to the A-bimodule structure defined by a · b = θ(a) b
and b · a = b θ(a) for every a ∈ A and b ∈ B.
Definition 6.1. Let G be either a discrete or a compact quantum group, let (A, α) and (B, β) be G-C*-algebras.
Fix a cardinal number κ larger than the density characters of A, B, and L2(G), and a countably incomplete
κ-good filter F . We say that a *-homomorphism θ : (A, α) → (B, β) has G-equivariant order zero dimension at
most d, written dimG
oz(θ) ≤ d, if there exist G-equivariant completely positive contractive order zero A-bimodule
F A, αF ) such that the sum ψ = ψ0 + · · · + ψd is a contractive linear map such
maps ψ0, . . . , ψd : (B, β) → (QG
that the following diagram commutes
A
∆A
❂❂❂❂❂❂❂❂
θ
B
F A
/ QG
=③③③③③③③③
ψ
The G-equivariant order zero dimension dimG
oz(θ) ≤ d, if such a d
exists, and ∞ otherwise. The order zero dimension of a *-homomorphism between C*-algebras can be obtained
as the particular instance of Definition 6.1 when G is the trivial group.
oz(θ) of θ is the least d ∈ N such that dimG
Remark 6.2. Definition 6.1 does not depend on the choice of the countably incomplete κ-good ultrafilter F .
This can be seen, for instance, by considering the syntactic characterization presented in Remark 6.3 below.
Furthermore, when G is second countable, and A, B are separable, one can choose F to be any countably
incomplete filter, such as the filter of cofinite subsets of N.
Remark 6.3. One can give a syntactic reformulation of the notion of G-equivariant order zero dimension.
To this purpose, one can consider the ordered operator space language Losos introduced in [22, Subsection
3.1], and the order zero language Loz introduced in [22, Subsection 3.2]. Adding function symbols for the
A-bimodule operations give the A-bimodule ordered operator space language Losos,A-A and the A-bimodule
order zero language Loz,A-A. In both these languages, the distinguished symbol for the metric is replaced by
pseudometric symbols dF for F ranging among the finite subsets of A, to be interpreted as the pseudometric
dF (x, y) = max {ka(x − y)k, k(x − y)ak : a ∈ F } .
G
and Loz,A-A
One can consider the G-equivariant version Losos,A-A
for each of these languages, which are
defined starting from these languages by adding symbols for a G-action as in Subsection 3.4. The Losos,A-A
-
morphisms between G-C*-algebras are precisely the G-equivariant completely positive contractive A-bimodule
maps, while the Loz,A-A
-morphisms are precisely the G-equivariant completely positive contractive order zero
A-bimodule maps. One can then rephrase Definition 6.1 by asserting that the order zero dimension of a
nondegenerate *-homomorphism θ : (A, α) → (B, β) is at most d if and only if for every positive quantifier-free
Loz,A-A
-formula ψ(x, z, y), where the variables z have
finite-dimensional C*-algebras as sorts, for every tuples a in A, b in B, and w in finite-dimensional C*-algebras,
and for every ε > 0, there exist tuples c0, . . . , cd in A such that
-formula ϕ(z, y), for every positive quantifier-free Losos,A-A
G
G
G
G
G
ψ(a, w, c0 + · · · + cd) ≤ ψ(θ(a), w, b) + ε
and ϕ(w, cj) ≤ ϕ(w, b) + ε for j = 0, 1, . . . , d.
If A0 is a G-C*-subalgebra of A containing an approximate unit for A, then one can replace Losos,A-A
Loz,A-A
in the discussion above.
with Losos,A0-A0
and Loz,A0-A0
G
and
G
G
G
Remark 6.4. Suppose that θ : (A, α) → (B, β) is a nondegenerate *-homomorphism. Fix a countably incom-
plete κ-good ultrafilter U, where κ is larger than the density character of A, B, and L2(G). Let also A0 be a
G-C*-subalgebra of A containing an approximate unit for A. In the case when G is compact, assume furthermore
that A0 is contained in the fixed point algebra (which is a nondegenerate C*-subalgebra of A). Following [30,
U A. It is easy to see that
-structure. It follows
oz(θ) ≤ d, then there exist G-equivariant completely positive contractive
Section 1] and [1, Remark 1.3], we consider the G-C*-subalgebra A0 ·QG
A0 ·QG
U A · A0 can be identified with the ultrapower of (A, α) regarded as an LC*,A0-A0
from the observations above that if dimG
U A · A0 of QG
G
/
=
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
21
order zero A-bimodule maps ψ0, . . . , ψd : B → A0 ·QG
ψ ◦ θ is the diagonal embedding from A to A0 ·QG
U A · A0.
Recall that a unital completely positive order zero map is a *-homomorphism. Next, we prove a similar result
with 'unital' being replaced by 'nondegenerate'. If α is a G-action on A, then α admits a unique extension to
a G-action on A, which we still denote by α.
U A · A0 such that ψ := ψ0 + · · · + ψd is contractive and
Lemma 6.5. Let φ : A → B be a completely positive order zero map between C*-algebras. If φ is nondegenerate,
then it is a *-homomorphism.
Proof. If A is unital, then φ : A → B ⊂ M (B) is unital, and the conclusion follows from the structure theorem
for completely positive order zero maps from [50]. Suppose that A is not unital, and let A be its unitization.
We let B∗∗ be the second dual of B, which we identify with the enveloping von Neumann algebra of B. Fix
an increasing approximate unit (uj)j∈J for A, and set g = supj∈J φ(uj) ∈ B∗∗. By [50, Proposition 3.2], the
(unique) linear map φ : A → B∗∗ extending φ with φ(1) = g, is completely positive of order zero. Since φ is
nondegenerate, we must have g = 1 ∈ B∗∗. Therefore φ is a *-homomorphism, and hence so is φ.
Lemma 6.6. Suppose that G is a compact or discrete quantum group, and (A, α) is a G-C*-algebra. The
inclusion map (A, α) ֒→ ( A, α) is positively LC*,A-A
-existential.
G
Proof. Consider an approximate unit (uj)j∈J for A. Fix a quantifier-free LC*,A-A
a in A.
(bk + λkuj)n
-formula, ϕ (¯x, ¯y), and a tuple
k=1 is a tuple in A satisfying the condition ϕ(a, ¯y) < r, then, for a suitable j ∈ J,
k=1 is a tuple in A satisfying the same condition.
If (bk + λk1)n
G
Proposition 6.7. Let G be either a compact or discrete quantum group, and let θ : (A, α) → (B, β) be an
injective *-homomorphism between G-C*-algebras.
(1) If (C, γ) is a G-C*-algebra, and ψ : (B, β) → (C, γ) is a *-homomorphism, then
dimG
oz(ψ ◦ θ) + 1 ≤ (dimG
oz(ψ) + 1)(dimG
oz(θ) + 1).
(2) If D is any C*-algebra, then dimG
(3) Let I be a directed set, and let ((Ai, αi), θij)i,j∈I be a direct system of G-C*-algebras. For every
oz(θi∞) ≤
αi) be the canonical *-homomorphism. Then dimG
oz(θ ⊗ idD) ≤ dimG
oz(θ).
i ∈ I, let θi∞ : (Ai, α) → (lim
−→
lim supj dimG
oz(θij ).
Ai, lim
−→
(4) Let I be a directed set, and for k = 0, 1, let ((A(k)
i
, α(k)
i
), θ(k)
)) be a compatible family of *-homomorphism. Then dimG
ij )i,j∈I be a direct system of G-C*-algebras.
oz(lim−→ηi) ≤
Let (ηi : (A(0)
lim supi∈I dimG
, α(0)
oz(ηi).
i
i
) → (A(1)
i
, α(1)
i
(5) Suppose that S is a positively quantifier-free Losos,A-A
G
-definable G-C*-subalgebra relative the class C of
G-C*-algebras that contain (A, α). Then θ maps S(A,α) to S(B,β), and dimG
oz(θS(A,α) ) ≤ dimG
oz(θ).
Proof. (1) -- (4) can be proved similarly as for classical groups; see [22, Proposition 5.4]. (5) is an easy consequence
of the definition.
The following is one of our main results regarding order zero dimension. In the proof of part (5), we will
use the fact that if ψ : A → B is a completely positive contractive order zero map, then ψ(a)ψ(bc) = ψ(ab)ψ(c)
for all a, b, c ∈ A. This fact follows easily by considering the induced *-homomorphism C0((0, 1]) ⊗ A → B [50,
Corollary 4.1].
Theorem 6.8. Let G be either a compact or discrete quantum group, and let θ : (A, α) → (B, β) be a nonde-
generate injective *-homomorphism. Then:
oz(θ) = 0 if and only if θ is positively LC*
(1) dimG
(2) The G-equivariant order zero dimension of θ : ( A, α) → ( B, β) is equal to dimG
(3) The G-equivariant order zero dimension of G ⋉ θ : (G ⋉α,r A, α) → (G ⋉β,r B, β)) is less than or equal
G -existential.
oz(θ).
to dimG
oz(θ).
(4) If (B, β) is free, and dimoz
G (θ) < +∞, then (A, α) is free.
Proof. We prove the theorem for compact G, since the case of discrete G is analogous.
It is obvious that if θ is positively LC*
(1):
oz(θ) = 0. Conversely, suppose that
dimG
oz(θ) = 0. Fix a countably incomplete κ-good ultrafilter U, where κ is larger than the density character of A,
G -existential, then dimG
22
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
B, and L2(G). Then by Remark 6.4, there exists a completely positive contractive order zero A-bimodule map
U A · A is the diagonal inclusion. Since φ is nondegenerate and
ψ is an A-bimodule map, ψ is nondegenerate. Therefore ψ is a *-homomorphism, witnessing that θ is positively
LC*
U A · A such that ψ ◦ φ : A → A ·QG
ψ : B → A ·QG
G -existential.
(2): We can assume that θ : A → B is the inclusion map. Suppose that dimG
oz(θ) ≤ d. By Lemma 6.6 and
Remark 6.4, there exist G-equivariant completely positive order zero A-bimodule maps
ψ0, . . . , ψd : B → Aα ·YG
U
A · Aα ⊆ Aα ·YG
U
A · Aα
such that ψ = ψ0 + · · · + ψd is contractive and ψ ◦ θA : A → Aα ·QG
ψ0, . . . , ψd are A-bimodule maps, given a ∈ Aα we have
aψ(1) = ψ(a1) = ψ(a) = a,
U A · Aα is the diagonal embedding. Since
and similarly ψ(1)a = a. This shows that ψ(1) = 1 and that
is the diagonal embedding. We also have, for j = 0, 1, . . . , d.
ψ ◦ θ : A → Aα ·YG
U
A · Aα ⊆YG
U
A
(a + λ1)ψj(b + µ1) = aψj(b + µ1) + λψj(b + µ1) = ψj(ab + µa + λb + λµ1) = ψj ((a + λ1)(b + µ1)).
Therefore ψj is an A-bimodule map. Hence ψ0, . . . , ψd witness that dimG
oz(θ) ≤ d, so dimG
oz(θ) ≤ dimG
oz(θ).
Conversely, suppose that dimG
oz(θ) ≤ d. Observe that Aα is a closed two-sided ideal of the fixed point algebra
of ( A, α). Therefore, by Remark 6.4, there exist G-equivariant completely positive contractive order zero A-
U A · Aα. Therefore the
oz(θ), as desired.
bimodule maps ψ0, . . . , ψd : B → Aα ·QG
A · Aα. Note that Aα ·QG
restriction of the maps ψ0, . . . , ψd to B witness that dimG
A · Aα = Aα ·QG
oz(θ) = d. Let us first consider the case when A, B are unital C*-algebras, in which
case θ is a unital *-homomorphism. Fix a cardinal κ larger than the density characters of A, B, and L2(G),
U A be maps as in the definition of
and a countably incomplete κ-good ultrafilter U. Let ψ0, . . . , ψd : B → QG
oz(θ) ≤ d. Consider the following maps:
(3): Suppose that dimG
oz(θ) ≤ d, so dimG
oz(θ) ≤ dimG
dimG
U
U
for j = 0, . . . , d obtained as in Lemma 4.3;
• the G-equivariant *-homomorphism G ⋉ θ : G ⋉α A → G ⋉β B; and
• the G-equivariant completely positive contractive A-bimodule maps G ⋉ ψj : G ⋉β B → G ⋉α (QG
• the G-equivariant injective *-homomorphism Ψ : G ⋉α (QG
Then G ⋉ ψ =Pd
j=0 G ⋉ ψj is a G-equivariant completely positive A-bimodule map satisfying
G
U A from Proposition 4.4.
U A) →Q
G
U A)
Ψ ◦ (G ⋉ ψ) ◦ (G ⋉ θ) = ∆A : A →Y
U
A.
Since G ⋉ ψ is completely positive, we have kG ⋉ ψk = k(G ⋉ ψ)(1)k = 1. Therefore dim
Consider now the case when A and B are not necessarily unital. Then dimG
oz(θ) ≤ d by (2). By applying the
G
oz(G ⋉ θ) ≤ d. Observe now that G ⋉α A ⊆ G ⋉α A and G ⋉β B ⊆ G ⋉ B
result to the unital case, we have dim
are positively quantifier-free LAα-Aα
-definable with respect to the class of G-C*-algebras of the form G ⋉γ C
for some G-C*-algebra (C, γ) such that (A, α) embeds equivariantly into (C, γ). Therefore, by part (5) of
Proposition 6.7, the restriction G ⋉ θ : G ⋉α A → G ⋉β B also has G-equivariant order zero dimension at most d
in view of the semantic characterization of G-equivariant order zero dimension from Remark 6.3. We conclude
that dim
G
oz(G ⋉ θ) ≤ dimG
oz(θ).
G
(4): We can assume that A ⊆ B is a nondegenerate G-C*-subalgebra, and θ : A → B is the inclusion map.
Let x ∈ C(G) ⊗ A, and ε > 0. Using freeness for β, find n ∈ N, and tuples b1, . . . , bn, c1, . . . , cn ∈ B with
G
oz(G ⋉ θ) ≤ d.
Set M = maxk=1,...,n kbkk and d = dimG
oz(θ) < ∞. Since A ⊆ B is nondegenerate, choose u ∈ A+ satisfying
kx −
β(bk)(1 ⊗ ck)k < ε/2.
n
Xk=1
for all k = 1, . . . , n.
kx(1 ⊗ u) − xk < ε/3, and kcku − ckk < ε/(3(d + 1)nM )
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
23
Let κ be a cardinal larger than the density characters of A, B, and L2(G). Find G-equivariant completely
U A as in the definition of order zero dimension. For j = 0, . . . , d,
, using functional
It is easily checked that πj is again a G-equivariant
positive order zero maps ψ0, . . . , ψd : B →QG
let πj : B →QG
U A be the completely positive contractive order zero map given by πj = ψ1/2
calculus for order zero maps (see [50, Corollary 3.2]).
A-bimodule map.
j
For j = 0, . . . , d and k = 1, . . . , n, set bj,k = πj(bk) and cj,k = ψj(ck). For j = 0, . . . , d, denote by φj : C(G) ⊗
U A the G-equivariant completely positive order zero A-bimodule map φj = idC(G) ⊗ πj. In the
rest of this proof, for elements e and f in some C*-algebra and δ > 0, we write e ≈δ f to mean ke − f k < δ. By
the comments before this proposition, we have
B → C(G) ⊗QG
φj(β(bk)(1 ⊗ ck))φj (1 ⊗ u) = φj(β(bk))φj (1 ⊗ cku)
= αU (bj,k)(1 ⊗ cj,ku)
≈ε/(3(d+1)n) αU (bj,k)(1 ⊗ cj,k).
In the following computation, we use the observation above at the first step; the definition of the elements bk
and ck at the third; the fact that u has a square root at the fourth, together with the comments before this
proposition regarding order zero maps; the fact that φj is a (C(G)⊗ A)-bimodule map at the fifth; the definition
of φj at the seventh; and the properties of the maps ψj at the eighth:
d
n
Xj=0
Xk=1
αU (bj,k)(1 ⊗ cj,k) ≈ε/3
d
n
d
φj n
Xk=1
Xj=0
β(bk)(1 ⊗ ck)! φj(1 ⊗ u)
d
d
d
≈ε/3
Xk=1
d
Xj=0
φj(x)φj (1 ⊗ u) =
φj (β(bk)(1 ⊗ ck))φj (1 ⊗ u) =
φj (x(1 ⊗ u1/2))φj (1 ⊗ u1/2)
Xj=0
Xj=0
Xj=0
Xj=0
k=1 αU (bj,k)(1 ⊗ cj,k)k < ε. Since the diagonal embedding of A → QG
j=0Pn
(idC(G) ⊗ ψ)(1 ⊗ u) = x(1 ⊗ u) ≈ε/3 x.
φj (1 ⊗ u1/2)φj (1 ⊗ u1/2) = x
φ2
j (1 ⊗ u)
d
Xj=0
= x
= x
We conclude that kx −Pd
positively LC*
G -existential, we conclude that there exist bj,k, cj,k ∈ A satisfying
U A is
This shows that (A, α) is free.
kx −
n
d
Xk=1
Xj=0
α(bj,k)(1 ⊗ cj,k) < ε.
Theorem 6.9. Let G be a compact quantum group, and let θ : (A, α) → (B, β) be a nondegenerate injective
*-homomorphism. Then:
(1) The G-equivariant order zero dimension of idKG ⊗ θ : (KG ⊗ A, αK) → (KG ⊗ B, βK) is equal to dimG
(2) The G-equivariant order zero dimension of G ⋉ θ : (G ⋉α,r A, α) → (G ⋉β,r B, β) is equal to dimG
(3) The G-equivariant order zero dimension of θAα is less than or equal to dimG
oz(θ).
oz(θ).
oz(θ).
Proof. (1): The proof of this fact is identical to the proof of part (2) of Proposition 5.5.
(2): One inequality follows from Theorem 6.8. As in the proof of part (3) of Proposition 5.5, one can now
deduce that in fact equality holds using Item (1) above and the Baaj -- Skandalis -- Takesaki -- Takai duality for
compact quantum groups.
(3): This is a particular instance of part (5) of Proposition 6.7 in the case of the fixed point subalgebra,
G -definable G-C*-subalgebra relative to the class of G-C*-algebras when G is a compact quantum
which is an Losos
group.
The following preservation result for *-homomorphisms with finite order zero dimension has been established
in [22, Proposition 5.22]. Given a C*-algebra A, we let dimnuc(A) be the nuclear dimension of A [51], and dr(A)
be the decomposition rank of A [31, Definition 3.1].
24
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
Proposition 6.10. Let A, B be C*-algebras, and θ : A → B be a *-homomorphism. Then
and
dimnuc(A) + 1 ≤ (dimoz(θ) + 1)(dimnuc(B) + 1)
dr(A) + 1 ≤ (dimoz(θ) + 1)(dr(B) + 1).
More generally, Proposition 6.10 applies to any dimension function for (nuclear) C*-algebras that is (nucle-
arly) positively ∀∃-axiomatizable in the sense of [22, Definition 5.15, Definition 5.16.].
6.2. Rokhlin dimension. In this subsection we fix a compact quantum group G. We consider C(G) ⊗ A as a
G-C*-algebra with respect to the action given by ∆ ⊗ idA.
Definition 6.11. The Rokhlin dimension dimRok(A, α) of a G-C*-algebra (A, α), is the G-equivariant order
zero dimension of α : (A, α) → (C(G) ⊗ A, ∆ ⊗ idA).
It follows from [2, Lemma 1.24] and [22, Lemma 5.13] that when G is classical, Definition 6.11 recovers the
usual notion of Rokhlin dimension for G-C*-algebras from [21, Definition 3.2].
The following is the main technical fact for actions with finite Rokhlin dimension.
Theorem 6.12. Let (A, α) be a G-C*-algebra. Then
G
dimoz(Aα ֒→ A) ≤ dimRok(A, α) = dim
oz(G ⋉α A ֒→ K G ⊗ A).
Here K G ⊗ A is regarded as a G-C*-algebra endowed with the stabilization of the trivial action of G on A.
Proof. Let d be the Rokhlin dimension of (A, α). We consider α as a G-equivariant *-homomorphism α :
(A, α) → (C (G) ⊗ A, ∆ ⊗ idA). We thus have that d is equal to the G-equivariant order zero dimension of α.
By part (3) of Theorem 6.9, dimG
oz (α) = dimRok (A, α). The restriction αAα is the embedding
Aα ֒→ 1 ⊗ A ⊆ C(G) ⊗ A, so this proves the first equality.
oz (αAα ) ≤ dimG
By part (2) of Theorem 6.9, the G-equivariant *-homomorphism G⋉α : G⋉αA → G⋉∆⊗idA C(G)⊗A induced
by α is equal to d. A particular instance of the Baaj -- Skandalis -- Takesaki -- Takai duality for compact quantum
groups yields a G-equivariant *-isomorphism ρ : G⋉∆⊗idA (C(G)⊗A) → KG ⊗A such that ρ◦(G⋉α) : G⋉α A →
KG ⊗ A is the canonical inclusion. This shows the G-equivariant *-homomorphism G ⋉α A ֒→ KG ⊗ A also has
G-equivariant order zero dimension d.
Corollary 6.13. Suppose that G is a compact quantum group, and (A, α) is a G-C*-algebra. Then
dimnuc(Aα) + 1 ≤ (dimRok(A, α) + 1)(dimnuc(A) + 1)
and
Proof. This is an immediate consequence of Theorem 6.12 and Proposition 6.10.
dr(G ⋉α A) + 1 ≤ (dimRok(A, α) + 1)(dr(A) + 1).
Remark 6.14. Let θ : (A, α) → (B, β) be a G-equivariant *-homomorphism. Using part (1) of Proposition 6.7
at the fourth step, we get
dimRok(A, α) + 1 = dimG
≤ (dimG
oz(α) + 1 ≤ dimG
oz(β) + 1)(dimG
oz((id ⊗ θ) ◦ α) + 1 = dimG
oz(β ◦ θ) + 1
oz(θ) + 1) = (dimRok(B, β) + 1)(dimG
oz(θ) + 1).
The following dimensional inequalities follow from the remark above and [22, Theorem 5.40 and Theorem
4.41]. We denote by Z the Jiang-Su algebra, and by O2 and O∞ the Cuntz algebras on two and infinitely many
generators, respectively.
Theorem 6.15. Let (A, α) be a G-C*-algebra, and let U be UHF-algebra of infinite type. Then
dimRok(A ⊗ Z, α ⊗ idZ ) ≤ 2 dimRok(A ⊗ U, α ⊗ idU ) + 1
and
dimRok(A ⊗ O∞, α ⊗ idO∞) ≤ 2 dimRok(A ⊗ O2, α ⊗ idO2) + 1.
Finally, we show that actions with finite Rokhlin dimension are free. This result is new even for actions
of classical finite groups. Previous partial results were considerably more technical, and the quantum group
perspective makes the argument significantly more transparent.
Theorem 6.16. Let (A, α) be a G-C*-algebra. If dimRok(α) < ∞, then α is free.
Proof. By Example 3.9, the G-C*-algebra C(G) ⊗ A is free. Since α : A → C(G) ⊗ A has finite G-equivariant
order zero dimension, the result follows from part (4) of Theorem 6.8.
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
25
6.3. Duality. One can isolate a notion of dimension for actions of discrete quantum groups, which is dual to
the notion of Rokhlin dimension for actions of compact quantum groups. Suppose that G is a compact quantum
group, and let (A, α) be a G-C*-algebra. Consider the reduced crossed product G ⋉α,r A, and the canonical
inclusion ιC(G) : C(G) → M ( G ⋉α,r A). Denote by V the fundamental unitary of G. Consider now the unitary
Then the map Ad(V ∗
α ) turns M ( G ⋉α,r A) into a G-C*-algebra.
Vα := (idc0( G) ⊗ ιC(G))(V ) ∈ M (c0( G) ⊗ ( G ⋉α,r A)).
Definition 6.17. Let (A, α) be a G-C*-algebra. The representation dimension dimrep(A, α) of (A, α) is the
G-equivariant order zero dimension of the canonical embedding (A, α) → ( G ⋉α,r A, Ad(V ∗
α )).
In the case of dimension zero, and when G is coexact and A is separable, Definition 6.17 recovers the notion
of spatial approximate representability from [2, Definition 4.7]. The following result generalizes [2, Theorem
4.12], which is then the case of dimension zero.
Theorem 6.18. Let G be a compact quantum group and let (A, α) be a G-C*-algebra. Then
dimRok(A, α) = dimrep(G ⋊α,r A, α).
Proof. Using the identifications from [2, Proposition 4.10, Proposition 4.11], the conclusion follows from part (2)
of Theorem 6.9 applied to α : A → C(G) ⊗ A.
Appendix
We recall here the fundamental notions concerning first order logic for metric structure. A good introduction
to this subject can be found in [3]. The systematic study of C*-algebras from the perspective of model theory
has been undertaken in [12]. We will consider the framework of languages with domains of quantification as
introduced in [15], which is particularly suitable for dealing with structures from functional analysis. In fact,
we will consider a slightly more general setting, which is necessary for our purposes, where the interpretation of
the domains of quantifications of a given sort are only required to be dense in the interpretation of the sort. We
will also consider the situation where, rather than a single distinguished metric single, the language contains an
upward directed collection of pseudometric symbols. This is useful when considering structures such as modules
and bimodules. It is clear that all the results of [15] go through in this slightly more general setting.
A.1. Syntax. A language L is given by:
• a collection of sorts S,
• a collection of function symbols f ,
• a collection of relation symbols R,
• for each sort S, an ordered, upward directed collection of domains of quantification D for S,
• for each sort S, a collection of variables of sort S,
• for each sort S, an ordered, upward directed collection of pseudometric symbols dS of arity 2 with input
sorts all equal to S,
• for each function and relation symbol B, a natural number nB (its arity),
• for each function symbol f , a distinguished tuple Sf
• for each relation symbol R, a distinguished tuple SR
• for each function symbol f and for each tuple D1, . . . , Dnf of domains of quantifications for Sf
nf of input sorts and an output sort Sf ,
nR of input sorts,
nf ,
and for each tuple of pseudometric symbols d1, . . . , dnf of sort S1, . . . , Snf , distinguished domains of
quantification Df
for the output
sort Sf of f , and a distinguished continuity modulus f
and a distinguished pseudometric symbol df
1 , . . . , Sf
1 , . . . , SR
D1,...,Dnf ,d1,...,dnf
1 , . . . , Sf
D1,...,Dnf
D1,...,Dnf ,d1,...,dnf
;
• for each relation symbol R and for each tuple D1, . . . , DnR of domains of quantifications for SR
1 , . . . , SR
nf ,
and for each tuple of pseudometric symbols d1, . . . , dnf of sort S1, . . . , Snf , a distinguished compact
interval DR
in R and a distinguished continuity modulus R
D1,...,Dn .
D1,...,DnR
A function symbol f is allowed to have arity nf = 0, in which case it is called a constant symbol. Given a
language L, one can define the notion of L-term by recursion. Formally, one declares that variables are terms,
and if t1, . . . , tn are L-terms and f is an n-ary function symbol in L, then f (t1, . . . , tn) is an L-term. If t is an
L-term and x1, . . . , xk are the variables that appear in t, then one also writes t as t(x1, . . . , xk).
26
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
Formulas are defined starting from terms by recursion. A basic L-formula in the free variables x = (x1, . . . , xk)
is an expression ϕ(x) of the form R(t1, . . . , tn) where t1(x), . . . , tn(x) are L-terms and R is an n-ary relation
symbol in L. A quantifier-free L-formula is an expression ϕ(x) of the form
q(ψ1(x), . . . , ψn(x))
where ψ1, . . . , ψn are basic L-formulas in the free variables x and q : Rn → R is a continuous function. Such
a quantifier-free formula is positive if q : Rn → R has the property that si ≤ ti for i = 1, 2, . . . , n implies
that q(s1, . . . , sn) ≤ q(t1, . . . , tn).
If furthermore q is of the form q(z1, . . . , zn) = max {u1(z1), . . . , un(zn)}
where u1, . . . , un : R → R are continuous nondecreasing functions, then we say that ϕ(x) is a positive primitive
quantifier-free L-formula.
A (positive/positive primitive) existential L-formula in the free variables x1, . . . , xk is an expression of
the form ϕ(x1, . . . , xk) = inf y1 · · · inf yℓ ψ(x1, . . . , xk, y1, . . . , yℓ) where ψ(x, y) is a (positive/positive primitive)
quantifier-free L-formula. A (positive/positive primitive) ∀∃-L-formula in the free variables x1, . . . , xk is an ex-
pression ϕ(x1, . . . , xk) of the form supz1 · · · supzm inf y1 · · · inf yℓ ψ(x1, . . . , xk, y1, . . . , yℓ, z1, . . . , zm) where ψ(x, y)
is a (positive/positive primitive) quantifier-free L-formula. Finally, arbitrary (positive/positive primitive) L-
formulas are defined similarly as above, but allowing an arbitrary finite number of alternations between sup
and inf. A (positive/positive primitive) L-sentence is a (positive/positive primitive) L-formula with no free
variables.
For each L-formula ϕ(x1, . . . , xk), for each choice D1, . . . , Dk of domains of quantification for the sorts of
x1, . . . , xk, and for each choice of pseudometric symbols d1, . . . , dk for the sorts x1, . . . , xk one can define by
recursion on the complexity of the formula, a uniform continuity modulus ϕ
for ϕ, as well as
a compact interval Dϕ
in R where ϕ takes values. A collection F of formulas in the free variables
x1, . . . , xk is uniformly equicontinuous if for any choice D1, . . . , Dk of domains of quantification for the sorts of
D1,...,Dk,d1,...,dk
D1,...,Dk
x1, . . . , xk, the continuity modulinϕ
nDϕ
: ϕ ∈ Fo are all contains in a common compact interval in R.
D1,...,Dk,d1,...,dk
D1,...,Dk
: ϕ ∈ Fo are uniformly bounded, and the compact intervals
A.2. Semantics. An L-structure M if given by:
• for each sort S, a set SM ,
• for each sort S, and for each pseudometric symbol dS of sort S, dM
S is a pseudometric on S, such that
the uniformity defined by these pseudometrics [44, Definition 2.1] is complete [44, Definition 1.14], and
such that the assignment dS 7→ dM
S is order preserving (where the order for pseudometric is pointwise
comparison);
• for each function symbol f with input sorts Sf
• for each domain of quantification D for S, a closed subset DM of SM such that the assignment D 7→ DM
is order preserving (where the collection of closed subsets of SM is ordered with respect to inclusion),
and such that the union of DM when D ranges among the domains of quantification for S is dense in
SM with respect to the uniformity described above,
1 , . . . , Sf
1 )M ×· · ·×
nf )M → (Sf )M (the interpretation of f in M ) such that, for any choice of domains of quantification
nf , the restriction of f M to
with respect
(Sf
D1, . . . , Dnf and pseudometric symbols d1, . . . , dnf for the sorts Sf
DM
to the metrics max{d1, . . . , dnf } and df
nf is a uniformly continuous map with continuity modulus f
D1,...,Dnf ,d1,...,dnf
, and its range is contained in (Df
nf and output sort Sf , a function f M : (Sf
• for each relation symbol R with input sorts SR
)M ;
1 )M ×
nR )M → R (the interpretation of R in M ) such that, for any choice of domains of quantification
1 ×
and its range is
· · · × (SR
D1, . . . , Dnf and pseudometric symbols d1, . . . , dnf for SR
· · · × DM
contained in DR
D1,...,Dnf
nf and output sort SR, a function RM : (SR
nf is a uniformly continuous map with continuity modulus R
nR, the restriction of f M to DM
D1,...,Dnf ,d1,...,dnf
1 , . . . , SR
1 ×· · ·×DM
D1,...,Dnf ,d1,...,dnf
1 , . . . , SR
1 , . . . , Sf
.
D1,...,DnR
Suppose now that t(x1, . . . , xk) is an L-term in the variables x1, . . . , xk of sorts S1, . . . , Sk, and that M
is an L-structure. Then one can define by recursion on the complexity of t the output sort S of t and the
interpretation tM of t in M , which is a function tM : SM
k → SM . Similarly, for any L-formula
ϕ(x1, . . . , xk) in the free variables x1, . . . , xk of sorts S1, . . . , Sk, one can defined by recursion on the complexity
of ϕ its interpretation ϕM in M , which is a function ϕM : SM
k → R. For any choice of domains
of quantification D1, . . . , Dk for the sorts of x1, . . . , xk, the restriction of ϕM to DM
is uniformly
1 × · · · × SM
1 × · · · × SM
1 × · · · × DM
k
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
27
continuous with the continuity modulus ϕ
interval Dϕ
of R.
D1,...,Dk
D1,...,Dk
as in the previous section, and takes values in the compact
Suppose now that C is a class of L-structures. Then C defines a metric in the class of L-formulas by setting,
for every L-formulas ϕ(x1, . . . , xk), ψ(x1, . . . , xk),
dC(ϕ, ψ) = sup ϕ(a) − ψ(a)
where M ranges among all the L-structures in C, and a ranges among all the tuples in M of the correct sorts.
Definition A.1. The language L is separable for a class of L-structures C if the space of L-formulas endowed
with the metric associated with C as above is separable. More generally, we define the density character of L
for C to be the density character of the space of L-formulas with respect to the metric dC.
Definition A.2. A class C of L-structures is axiomatizable in the language L if there exists a collection A of
L-conditions of the form σ ≤ r (axioms) where σ is an L-sentence and r ∈ R, such that for any L-structure M ,
M belongs to C if and only if σM ≤ r for any condition in A.
Definition A.3. A class of L-structures is (positively/positively primitively) ∀∃-axiomatizable in the language
L if it is axiomatizable by a collection of L-conditions of the form σ ≤ r where σ is a (positive/positive primitive)
∀∃-L-sentence and r ∈ R.
The following notion has been essentially introduced in [12, Definition 5.7.1].
Definition A.4. A class C of L-structures is definable by a uniform collection of positive existential L-formulas
if for every choice of sorts S1, . . . , Sk in L and domains of quantification D1, . . . , Dk for S1, . . . , Sk, there exist
and a uniformly equicontinuous family F (D1, . . . , Dk) of positive existential L-formulas in the free variables
x1, . . . , xk of sorts S1, . . . , Sk such that, for every L-structure M , the following assertions are equivalent:
(1) M belongs to C;
(2) for every choice of sorts S1, . . . , Sk in L, of domains of quantification D1, . . . , Dk for S1, . . . , Sk, of
for i = 1, 2, . . . , k, and for every ε > 0, there exists a formula ϕ(x1, . . . , xk) in
elements ai ∈ DM
i
F (D1, . . . , Dk) such that M = ϕ(a1, . . . , ak) ≤ ε.
It is clear that, if a class of L-structures is positively ∀∃-axiomatizable in the language L, then in particular
it is definable by a uniform collection of positive existential L-formulas.
A.3. Reduced products and ultraproducts. Reduced products and ultraproducts are a fundamental con-
struction in model theory. We recall here these notions and their basic properties. Suppose that L is a language
as above, I is an index set, Mi for i ∈ I is an L-structure, and F is a filter over I. One can define the reduced
of quantification D for S, and for each pseudometric symbol d of sort S, consider the pseudometric dM on
i defined by d((xi), (yi)) = lim supF dMi (xi, yi). The collection of such pseudometrics, when d ranges
, and then one can define DM to be
the Hausdorff completion of such a uniform space, endowed with the canonical pseudometrics induces by the
. We will denote by [xi]F the element of DM associated with the sequence (xi)i∈I ,
L-product QF Mi of the I-sequence (Mi)i∈I with respect to F as follows. For each sort S in L and domain
Qi DM
among the pseudometric symbols of sort S, defines a uniformity on Qi DM
pseudometrics on Qi DM
and we will call (xi) a representative sequence for [xi]F .
If D1, D2 are domains of quantification for S with D1 ≤ D2, then one can canonically identify DM
1 as a
closed subspace of DM
2 . Since the collection of domains of quantification for S is directed, the union of DM ,
where D ranges among all the domains of quantifications for S, is a (possibly incomplete) metric space. Define
then SM to be the completion of such a metric space. One can regard DM as a closed subspace of SM . Clearly,
by definition of SM , the union of DM , where D ranges among all the quantifications for S, is dense in SM .
i
i
For any function symbol f in L with input sorts Sf
nf and output sort Sf , one can then define a function
nf )M → (Sf )M by setting, for every choice of domains of quantification D1, . . . , Dn for
1 , . . . , Sf
1 )M × · · · × (Sf
f M : (Sf
Sf
1 , . . . , Sf
nf , and elements [a1
i ]F of DM
1 , . . . , DM
n ,
i ]F , . . . , [an
f ([a1
i , . . . , an
Similarly, for any relation symbol R in L with input sorts SR
function RM : (SR
for SR
1 )M × · · · × (SR
nR, and elements [a1
i ]F ) = [f (a1
i ]F , . . . , [xn
1 , . . . , SR
nf and output sort SR, one can define a
nR)M → R by setting, for every choice of domains of quantification D1, . . . , Dn
i ]F , . . . , [an
R([a1
1 , . . . , DM
n ,
i ]F ) = lim sup
i ]F , . . . , [an
i ]F of DM
i , . . . , an
R(a1
i ).
D1,...,Dn
)M .
i )]F ∈ (Df
1 , . . . , SR
F
28
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
One obtains in this way an L-structure M , which is called the reduced L-product of the I-sequence of L-
structures (Mi)i∈I with respect to the filter F , and denoted by QF Mi. In the case when F is an ultrafilter,
then QF Mi is called L-ultraproduct of the I-sequence (Mi)i∈I . The function that maps an element a of M
to the element of QF Mi with representative sequence constantly equal to a defines the canonical diagonal
L-embedding of M intoQF Mi. The fundamental properly of reduced L-products is the following result, known
as Los' theorem; see [15, Proposition 4.3].
Proposition A.5. Suppose that F is a filter on a set I, (Mi)i∈I is an I-sequence of L-structures, and
ϕ(x1, . . . , xk) is a formula in the free variables x1, . . . , xn.
If either ϕ is positive primitive quantifier-free,
or F is an ultrafilter, then for any [a1
i ], . . . , [ak
i ] in QF Mi one has
i ]) = lim sup
i ], . . . , [ak
ϕQF Mi ([a1
ϕMi (a1
i , . . . , ak
i ).
F
Corollary A.6. Suppose that C is a class of L-structures that is axiomatizable in the language L. Then C is
closed under L-ultraproducts. If C is furthermore positively primitively ∀∃-axiomatizable in the language L, then
C is closed under reduced L-products.
A.4. Types and saturation. A (closed) L-condition is an expression of the form ϕ(x) ≤ r for some L-formula
ϕ in the free variables x = (x1, . . . , xk) and r ∈ R. An L-type p(x1, . . . , xk) in the free variables x1, . . . , xk
is a collection of L-conditions ϕ(x) ≤ r. Such a type p is quantifier-free if all the conditions that appear in
it involve quantifier-free L-formulas, and positive quantifier-free if all the conditions that appear in it involve
positive primitive quantifier-free L-formulas. A realization of the L-type p(x) in an L-structure M is a tuple
a = (a1, . . . , ak) in M such that each ai belongs to the interpretation of the sort of xi, and ϕM (a) ≤ r for every
condition ϕ(x) ≤ r in p(x). In this case, we also write M = ϕ(a) ≤ r and M = p(a). An L-type is realized in
an L-structure M if it admits a realization in M . It is approximately realized in M if for any finite subset p0(x)
of p(x) and for any ε > 0, the type pε
0(x1, . . . , xk) consisting of the conditions ϕ(x) ≤ r + ε for any condition
ϕ(x) ≤ r in p0(x1, . . . , xk), is realized in M .
Suppose that M is an L-structure, and A is a subset of M . Then one can consider the language L(A)
obtained by adding to L a constant symbol for each element of A. Then M or any L-structure containing M
can be canonically regarded as an L(A)-structure.
Definition A.7. Suppose that κ is an uncountable cardinal. An L-structure M is L-κ-saturated if for every
subset A of M of density character less than κ and for any L(A)-type p, if p is approximately realized in M , then
it is realized in M . Replacing arbitrary types with positive quantifier-free types gives the notion of positively
quantifier-free L-κ-saturated structure.
In the particular case when κ = ℵ1, L-ℵ1-saturation is also called countable L-saturation, and positive
quantifier-free L-κ-saturation is also called positive quantifier-free countable L-saturation. The fundamental
property of ultrapowers of structures with respect to nonprincipal ultrafilters over N is that they are countably
saturated; see [15, Proposition 4.11].
Proposition A.8. Suppose that F is a countably incomplete filter. Let C be a class of L-structures such that L is
separable for C. If M is an L-structure in C, then the reduced power QF M is countably positively quantifier-free
L-saturated. If F is a countably incomplete ultrafilter, then the ultrapower QU M is countably L-saturated.
Proposition A.8 admits a generalization to an arbitrary uncountable cardinal κ. In this more general setting,
one needs to consider (ultra)filters that are moreover κ-good. The definition of κ-good is given in [5, Section 6.1]
for ultrafilters, but it applies equally well to filters. Every countably incomplete filter is ℵ1-good. In particular,
every nonprincipal ultrafilter over N is ℵ1-good. The same proof as [5, Theorem 6.1.8] gives the following.
Proposition A.9. Let κ be an uncountable cardinal, and let F be a countably incomplete κ-good filter. Let C
be a class of L-structures such that L has density character less than κ for C. If M is an L-structure, then the
reduced power QF M is positively quantifier-free L-κ-saturated. If U is a countably incomplete κ-good ultrafilter,
then the ultrapower QU M is L-κ-saturated.
A.5. Existential embeddings. Suppose that L is a language in the logic for metric structures as above, M, N
are L-structures, and T : M → N is a function. Then T is an L-morphism if, for every sort S in L and for every
domain of quantification D for S, T maps SM to SN and DM to DN , and for any atomic formula ϕ(x1, . . . , xk)
one has that ϕN (T (a1), . . . , T (ak)) ≤ ϕM (a1, . . . , ak) for any a1, . . . , an of the same sorts as x1, . . . , xk. An
L-morphism is an L-embedding if for any atomic formula ϕ(x1, . . . , xk) one has that ϕN (T (a1), . . . , T (ak)) =
ϕM (a1, . . . , ak) for any a1, . . . , an of the same sorts as x1, . . . , xk. Suppose that T : M → N is an L-embedding,
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
29
and A is a subset of M . Recall that L(A) is the language obtained from L by adding a constant symbols ca for
any element a of A. Then one can regard both M and N as L(A)-structures, by interpreting ca as a in M and
as T (a) in N . We recall here the notion of (positively) L-existential L-embedding.
Definition A.10. Suppose that M, N are L-structures, and T : M → N is an L-embedding. Then T is a
(positively) L-existential if for every (positive) quantifier-free L(M )-condition ϕ(x1, . . . , xn) ≤ r satisfied in N
and for every ε > 0, then L(M )-condition ϕ(x1, . . . , xn) ≤ r + ε is satisfied in M .
The following characterization of L-existential L-embeddings is an immediate consequence of Los' theorem
and saturation of reduced powers. Recall that any countably incomplete filter is ℵ1-good.
Proposition A.11. Fix an uncountable cardinal κ, a class C of L-structures, and structures M, N in C of
density character less than κ. Suppose that F is a countably incomplete κ-good filter. Let T : M → N be an
L-morphism. The following assertions are equivalent:
(1) T is a positively L-existential L-embedding;
Furthermore, if F is an ultrafilter, then the following assertions are equivalent:
(1) T is an L-existential L-embedding;
(2) there exists an L-morphism S : N →QF M such that S ◦ T is the diagonal L-embedding M →QF M .
(2) there exists an L-embedding S : N →QU M such that S ◦ T is the diagonal L-embedding M →QU M .
The following preservation result for positively L-existential L-embedding can be easily verified directly.
Proposition A.12. Suppose that C is a collection of L-structures that is definable by a uniform collection
of positive existential L-formulas. Suppose that A and B are L-structures, and T : A → B is a positively
L-existential L-embedding. If B belongs to C, then A belongs to C.
A.6. Definability. Suppose that C is a class of L-structures. Fix domains of quantifications D1, . . . , Dn for L
of sorts S1, . . . , Sn. A uniform assignment S in D1 × · · · × Dn relative to the class C is an assignment M 7→ SM ,
where M ranges among the L-structures in C and SM is a closed subset DM
; see [12, Definition
3.2.1]. The following definition is equivalent to the one given in [12, Definition 3.2.1]; see [12, Theorem 3.2.2].
1 × · · · × DM
k
Definition A.13. Suppose that S is a uniform assignment in D1 × · · · × Dn relative to the class C. Then S is
an L-definable set in D1 × · · · × Dn relative to the class C if for every ε > 0, for every choice of pseudometric
symbols d1, . . . , dn of sorts S1, . . . , Sn there exists an L-formula ϕ(x1, . . . , xk) in the free variables x1, . . . , xn of
sorts S1, . . . , Sn such that, for every L-structure M in C and for every (a1, . . . , an) ∈ D1 × · · · × Dn,
We say that S is positively existentially L-definable in the L-formulas ϕ above can be chosen to be positive
existential.
ϕM (a1, . . . , an) −
inf
(b1,...,bn)∈SM
max
1≤i≤n
(cid:12)(cid:12)(cid:12)(cid:12)
< ε.
di(ai, bi)(cid:12)(cid:12)(cid:12)(cid:12)
Observe that, every choice of pseudometric symbols d1, . . . , dn of sort S1, . . . , Sn defines a pseudometric ρ on
the space of L-definable sets in D1 × · · · × Dn relative to the class C , obtained by setting
ρ(S, S′) = sup
M ∈C
ρM (SM , S′M ),
where ρM is the Hausdorff metric on the space of closed subsets of DM
n associated with the metric
d(a, b) = max1≤i≤n di(ai, bi) on DM
n . The collection of such pseudometrics ρ, obtaining by letting
d1, . . . , dn range among all the pseudometric symbols of sort S1, . . . , Sn, define a uniform structure on the space
of L-definable sets in D1 × · · · × Dn relative to the class C . It is clear from the definition that the space of
(positively existentially) L-definable sets is complete with respect to such a uniform structure.
1 × · · · × DM
1 × · · · × DM
Suppose now that M is an L-structure, and C is a class of L-structures as above. Then one can consider
the class C(M ) of L-structures in C that contain a distinguished copy of M . Then one can naturally regard the
structures in C(M ) as L(M )-structures. We consider a natural notion of L-definable substructure. Suppose as
above that C is a class of L-structures.
Definition A.14. A positively existentially L-definable substructure relative to the class C is an assignment
(M, D) 7→ SD,M where M ranges among the L-structures in C and D ranges among the domains of quantification
in L such that:
• SD,M is a closed subset of DM ,
30
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
• setting S(M )D := SD,M for every sort S and every domain of quantification D of sort S defines an
L-substructure S(M ) of M ,
• for every domain D, the assignment M 7→ SD,M is a positively existentially L-definable set in D relative
to the class C.
The following proposition is an easy consequence of the definition of positive existential L-embedding.
Proposition A.15. Suppose that M, N are L-structures, and T : M → N is a positively L-existential L-
embedding. Suppose that S is a positively existentially L(M )-definable substructure relative to a class C of
L(M )-structures containing M and N . Then T maps S(M ) to S(N ), and the restriction T S(M) : S(M ) → S(N )
is a positively L-existential L-embedding.
References
[1] Sel¸cuk Barlak and G´abor Szab´o, Sequentially split *-homomorphisms between C*-algebras, International
Journal of Mathematics 27 (2016), no. 13.
[2] Sel¸cuk Barlak, G´abor Szab´o, and Christian Voigt, The spatial Rokhlin property for actions of compact
quantum groups, Journal of Functional Analysis 272 (2017), no. 6, 2308 -- 2360.
[3] Itaı Ben Yaacov, Alexander Berenstein, C. Ward Henson, and Alexander Usvyatsov, Model theory for
metric structures, Model theory with applications to algebra and analysis. Vol. 2, London Mathematical
Society Lecture Note Series, vol. 350, Cambridge University Press, 2008, pp. 315 -- 427.
[4] Kevin Carlson, Enoch Cheung, Ilijas Farah, Alexander Gerhardt-Bourke, Bradd Hart, Leanne Mezuman,
Nigel Sequeira, and Alexander Sherman, Omitting types and AF algebras, Archive for Mathematical Logic
53 (2014), no. 1-2, 157 -- 169.
[5] Chen C. Chang and H. Jerome Keisler, Model theory, second ed., North-Holland Publishing Co.,
Amsterdam-New York-Oxford, 1977, Studies in Logic and the Foundations of Mathematics, 73.
[6] Kenny De Commer, Actions of compact quantum groups, arXiv:1604.00159 (2016).
[7] Christopher J. Eagle, Ilijas Farah, Bradd Hart, Boris Kadets, Vladyslav Kalashnyk, and Martino Lupini,
Fraıss´e limits of C*-algebras, Journal of Symbolic Logic 81 (2016), no. 2, 755 -- 773.
[8] Christopher J. Eagle, Isaac Goldbring, and Alessandro Vignati, The pseudoarc is a co-existentially closed
continuum, Topology and its Applications 207 (2016), 1 -- 9.
[9] Christopher J. Eagle and Alessandro Vignati, Saturation and elementary equivalence of C*-algebras, Journal
of Functional Analysis 269 (2015), no. 8, 2631 -- 2664.
[10] David Alexandre Ellwood, A new characterisation of principal actions, Journal of Functional Analysis 173
(2000), no. 1, 49 -- 60.
[11] Ilijas Farah and Bradd Hart, Countable saturation of corona algebras, Comptes Rendus Math´ematiques de
l'Acad´emie des Sciences 35 (2013), no. 2, 35 -- 56.
[12] Ilijas Farah, Bradd Hart, Martino Lupini, Leonel Robert, Aaron Tikuisis, Alessandro Vignati, and Winter
Winter, Model theory of C*-algebras, arXiv:1602.08072 (2016).
[13] Ilijas Farah, Bradd Hart, Mikael Rørdam, and Aaron Tikuisis, Relative commutants of strongly self-
absorbing C*-algebras, Selecta Mathematica 23 (2017), no. 1, 363 -- 387.
[14] Ilijas Farah, Bradd Hart, and David Sherman, Model theory of operator algebras I: stability, Bulletin of the
London Mathematical Society 45 (2013), no. 4, 825 -- 838.
[15]
[16]
, Model theory of operator algebras II: model theory, Israel Journal of Mathematics 201 (2014),
no. 1, 477 -- 505.
, Model theory of operator algebras III: elementary equivalence and II1 factors, Bulletin of the
London Mathematical Society 46 (2014), no. 3, 609 -- 628.
[17] Eusebio Gardella, Crossed products by compact group actions with the Rokhlin property, Journal of Non-
commutative Geometry, in press.
[18]
, Regularity properties and Rokhlin dimension for compact group actions, Houston Journal of Math-
ematics, in press.
[19]
[20]
[21] Eusebio Gardella, Ilan Hirshberg, and Luis Santiago, Rokhlin dimension: tracial properties and crossed
, Rokhlin dimension for compact group actions, Indiana Journal of Mathematics, in press.
, Classification theorems for circle actions on Kirchberg algebras, II, arXiv:1406.1208 (2014).
products, preprint.
[22] Eusebio Gardella and Martino Lupini, Equivariant logic and applications to C*-dynamics, arXiv:1608.05532
(2016).
MODEL THEORY AND ROKHLIN DIMENSION FOR COMPACT QUANTUM GROUP ACTIONS
31
[23] Saeed Ghasemi, Reduced products of metric structures: a metric Feferman-Vaught theorem, The Journal
of Symbolic Logic 81 (2016), no. 3, 856 -- 875.
[24] Isaac Goldbring and Thomas Sinclair, On Kirchberg's embedding problem, Journal of Functional Analysis
269 (2015), no. 1, 155 -- 198.
[25]
[26] Stefan Heinrich, Ultraproducts in Banach space theory, Journal fur die Reine und Angewandte Mathematik
, On the axiomatizability of C*-algebras as operator systems, arXiv:1603.05444 (2016).
313 (1980), 72 -- 104.
[27] Ilan Hirshberg, G´abor Szab´o, Wilhelm Winter, and Jianchao Wu, Rokhlin dimension for flows,
arXiv:1607.02222 (2016).
[28] Ilan Hirshberg, Wilhelm Winter, and Joachim Zacharias, Rokhlin dimension and C*-dynamics, Communi-
cations in Mathematical Physics 335 (2015), no. 2, 637 -- 670.
[29] Masaki Izumi, Finite group actions on C*-algebras with the Rohlin property, I, Duke Mathematical Journal
122 (2004), no. 2, 233 -- 280.
[30] Eberhard Kirchberg, Central sequences in C*-algebras and strongly purely infinite algebras, Operator Al-
gebras: The Abel Symposium 2004, Abel Symp., vol. 1, Springer, Berlin, 2006, pp. 175 -- 231.
[31] Eberhard Kirchberg and Wilhelm Winter, Covering dimension and quasidiagonality, International Journal
of Mathematics 15 (2004), no. 01, 63 -- 85.
[32] Akitaka Kishimoto, The Rohlin property for automorphisms of UHF algebras, Journal fur die reine und
angewandte Mathematik 1995 (1995), no. 465, 183 -- 196.
[33] Kazunori Kodaka and Tamotsu Teruya, The Rohlin property for coactions of finite dimensional C*-Hopf
algebras on unital C*-algebras, Journal of Operator Theory 74 (2015), no. 2, 329 -- 369.
[34] Johan Kustermans and Lars Tuset, A survey of C*-algebraic quantum groups. I, Irish Mathematical Society
Bulletin (1999), no. 43, 8 -- 63.
[35]
, A survey of C*-algebraic quantum groups. II, Irish Mathematical Society Bulletin (2000), no. 44,
6 -- 54.
[36] Johan Kustermans and Stefaan Vaes, Locally compact quantum groups, Annales Scientifiques de lcole Nor-
male Suprieure 33 (2000), no. 6, 837 -- 934.
[37]
[38]
, The operator algebra approach to quantum groups, Proceedings of the National Academy of Sciences
of the United States of America 97 (2000), no. 2, 547 -- 552 (electronic).
, Locally compact quantum groups in the von Neumann algebraic setting, Mathematica Scandinavica
92 (2003), no. 1, 68 -- 92.
[39] Ann Maes and Alfons Van Daele, Notes on compact quantum groups, Nieuw Archief voor Wiskunde. Vierde
Serie 16 (1998), no. 1-2, 73 -- 112.
[40] Shuhei Masumoto, A Fraıss´e theoretic approach to the Jiang -- Su algebra, arXiv:1612.00646 (2016).
[41]
[42] Norio Nawata, Finite group actions on certain stably projectionless C*-algebras with the Rohlin property,
, Jiang-Su Algebra as a Fraıss´e limit, arXiv:1602.00124 (2016).
Transactions of the American Mathematical Society 368 (2016), no. 1, 471 -- 493.
[43] Ryszard Nest and Christian Voigt, Equivariant Poincar´e duality for quantum group actions, Journal of
Functional Analysis 258 (2010), no. 5, 1466 -- 1503.
[44] Jan Pachl, Uniform spaces and measures, Fields Institute Monographs, vol. 30, Springer, New York, 2013.
[45] Piotr Podle´s, Symmetries of quantum spaces. Subgroups and quotient spaces of quantum SU2 and SO3
groups, Communications in Mathematical Physics 170 (1995), no. 1, 1 -- 20.
[46] Timothy Rainone and Christopher Schafhauser, Crossed products of nuclear C*-algebras by free groups and
their traces, arXiv:1601.06090 (2017).
[47] Thomas Timmermann, An invitation to quantum groups and duality, EMS Textbooks in Mathematics,
European Mathematical Society, 2008.
[48] Stefaan Vaes, Locally compact quantum groups, Ph.D. thesis, Katholieke Universiteit Leuven, 2001.
[49] Alfons Van Daele, Discrete quantum groups, Journal of Algebra 180 (1996), no. 2, 431 -- 444.
[50] Wilhelm Winter and Joachim Zacharias, Completely positive maps of order zero, Munster Journal of Math-
ematics 2 (2009), 311 -- 324.
[51]
, The nuclear dimension of C*-algebras, Advances in Mathematics 224 (2010), no. 2, 461 -- 498.
32
EUSEBIO GARDELLA, MEHRDAD KALANTAR, AND MARTINO LUPINI
Eusebio Gardella, Westfalische Wilhelms-Universitat Munster, Fachbereich Mathematik, Einsteinstrasse 62, 48149
Munster, Germany
E-mail address: [email protected]
URL: https://wwwmath.uni-muenster.de/u/gardella/
Mehrdad Kalantar, Department of Mathematics, University of Houston, Philip Guthrie Hoffman Hall, 3551
Cullen Blvd., Houston, TX 77204, USA
E-mail address: [email protected]
URL: https://www.math.uh.edu/~kalantar/
Martino Lupini, Mathematics Department, California Institute of Technology, 1200 E. California Blvd, MC 253-37,
Pasadena, CA 91125
E-mail address: [email protected]
URL: http://www.lupini.org/
|
1505.00302 | 2 | 1505 | 2015-08-22T15:05:32 | Maximal amenability and disjointness for the radial masa | [
"math.OA"
] | We prove that the radial masa C in the free group factor is disjoint from other maximal amenable subalgebras in the following sense: any distinct maximal amenable subalgebra cannot have diffuse intersection with C. | math.OA | math |
Maximal amenability and disjointness for the radial masa
Chenxu Wen
Abstract
We prove that the radial masa C in the free group factor is disjoint from other
maximal amenable subalgebras in the following sense: any distinct maximal amenable
subalgebra cannot have diffuse intersection with C.
Introduction
Amenability is one of the most central notions in the study of von Neumann algebras.
Amenable von Neumann algebras are very well understood: Connes' work on the charac-
terization of amenability [6] is a milestone for the entire theory. Thus, in order to study
non-amenable von Neumann algebras, it is natural to consider their amenable subalgebras.
Fuglede and Kadison [7] showed that for any II1 factor, there always exists a maximal
hyperfinite subfactor, thus answered a question of Murray and von Neumann about the
double relative commutant. Later on, during a conference at Baton Rouge in 1967, Kadison
asked a series of famous questions about von Neumann algebras (see for example [8]).
Among them is the following:
Question. Is every self-adjoint element in a II1 factor contained in a hyperfinite subfactor?
Popa answered this question in the negative, by showing that the generator masa in the
free group factor is maximal amenable, [22].
If (M, τ ) is a finite von Neumann algebra with a faithful normal tracial state τ and ω
is a free ultrafilter, we'll write M ω as the ultraproduct of (M, τ ). The key insight of Popa
[22] is that the inclusion A ⊂ M , where M = L(Fn) with n ≥ 2 and A the generator masa,
satisfies the asymptotic orthogonality property, which we define below:
Definition. Let A ⊂ M be an inclusion of finite von Neumann algebras. We say that
the inclusion satisfies the asymptotic orthogonality property (AOP for short), if for any free
ultrafilter ω on N, (xn)n ∈ A′∩M ω⊖Aω and y1, y2 ∈ M ⊖A, we have that y1(xn)n ⊥ (xn)ny2.
Since Popa, there are many results considering maximal amenable subalgebras. Ge
[9, Theorem 4.5] showed that any diffuse amenable finite von Neumann algebra can be
1
realized as a maximal amenable subalgebra of the free group factor. Shen [28] showed that
the Nn∈N A is maximal amenable inside Nn∈N M , where A is the generator masa in the
free group factor M , thus gave an example of a maximal masa in a McDuff-II1 factor.
Cameron, Fang, Ravichandran and White [4] proved that the radial masa in the free group
factor is maximal amenable. Brothier [3] gave an example in the setting of planar algebras.
Boutonnet and Carderi [1] showed that the subalgebra coming from a maximal amenable
subgroup in a hyperbolic group, is maximal amenable. Houdayer [11] showed that the
factors coming from free Bogoljubov actions contains concrete maximal amenable masa's,
see also [13]. All these results use Popa's AOP approach.
Very recently a new method via the study of centralizers of states, is developed by Bou-
tonnet and Carderi [2]. In particular, they are able to show that the subalgebra coming from
the upper-triangular matrix subgroup of SL(3, Z), is maximal amenable inside L(SL(3, Z)).
See Ozawa's remark [18] for an application of this new approach.
Another central theme in the theory is the study of the free group factors ([16], [30],
[10], [17], [19]). One of the motivating questions of this paper is a conjecture by J. Peterson
(see the end of [21]):
Conjecture. For the free group factor, any diffuse amenable subalgebra is contained in a
unique maximal amenable subalgebra.
Houdayer's result on Gamma stability of free products [13, Theorem 4.1] implies that
the generator masa satisfies Peterson's conjecture. The proof again is relying on the AOP.
See also Ozawa's proof [18] via the centralizer approach.
One subalgebra of the free group factor under intense study is the radial masa. So let
M = L(FN ) with 2 ≤ N < ∞ be the free group factor with finitely many generators and
denote by C the von Neumann subalgebra of M generated by ω1 := Pg∈FN ,g=1 ug. Note
that ω1 is only well-defined for free groups with finitely many generators. It was proved by
Pytlik [26, Theorem 4.3] that C is a masa in M , called the radial masa or the Laplacian
masa. Moreover, Radulescu [27, Theorem 7] showed that C is singular and Cameron, Fang,
Ravichandra and White [4, Corollary 6.3] proved that it is maximal amenable in M .
Recall that a result of Popa [23, Corollary 4.3] shows that generator masa's coming from
different generators cannot be unitarily conjugate inside M . This implies that the radial
masa C cannot be unitarily conjugate with the generator masa A inside M . However,
whether they are conjugate via some automorphism, is still unknown.
The main result of this paper is the following:
Theorem. Let M = L(FN ) with 2 ≤ N < ∞ and let C ⊂ M be the radial masa. Then
every amenable subalgebra of M having diffuse intersection with C, must be contained in
C.
This is the first example of such disjointness for an maximal amenable subalgebra which
is not known to be in a free position.
2
The approach taken in this paper is to show a stronger version of Popa's AOP. The main
idea is that, according to a computation by Radulescu [27], ℓ2(FN ) ⊖ L2(C) admits a very
nice decomposition as a direct sum of C-C-bimodules. Moreover, each bimodule contains a
Riesz basis whose interaction with the left and right actions of C is very similar to that of
the canonical basis of the free group. In other words, the radial masa C and L2(M ) ⊖ L2(C)
behave almost freely.
Acknowledgement
We would like to thank Jesse Peterson, for suggesting the question and for many valuable
discussions and consistent encouragement. We are very grateful to Cyril Houdayer, R´emi
Boutonnet and Stuart White for a careful reading and numerous helpful suggestions on an
early version of the paper. We also thank Sorin Popa and Allan Sinclair for their comments.
Last but not least, we would like to thank the anonymous referee for providing many helpful
comments.
Proof of the Theorem
Recall that a von neumann algebra M is said to be solid, if the relative commutant of any
diffuse amenable subalgebra is amenable [17]. M is called strongly solid, if for any diffuse
amenable subalgebra D of M , the normalizer of D generates an amenable subalgebra of M
[19]. Clearly strong solidity implies solidity. It is shown in [19] that free group factors are
strongly solid.
The following proposition is inspired by [22, Lemma 3.1, Theorem 3.2] and [4, Lemma
2.2, Corollary 2.3]:
Proposition 1. Let M be a strongly solid II1 factor and A ⊂ M a singular masa in M .
Assume in addition that for any diffuse von Neumann subalgebra B ⊂ A and any free
ulltrafilter ω, the following holds:
(xk)ky2.
for any (xk)k ∈ B′T M ω ⊖ Aω and for any y1, y2 ∈ M ⊖ A, we have that y1(xk)k ⊥
Then any amenable subalegbra of M containing B, must be contained in A.
Proof. As shown by [4, Lemma 2.2, Corollary 2.3], AOP and singularity imply that A is
maximal amenable in M .
Let B ⊂ Q ⊂ M be an amenable subalgebra. By solidity of M , A ⊂ B′T M is amenable.
Since A is maximal amenable, we conclude Q′T Q ⊂ B′T M ⊂ A.
Let z be the maximal central projection of Q such that Qz is type II1. Now suppose
that z 6= 0.
3
Since Qz is amenable and of type II1, Popa's intertwining theorem ([24, Theorem A.1])
see [4, Lemma 2.2].
easily implies that there is a unitary u ∈ (Qz)′T(Qz)ω, such that EAω (u) = 0. For a proof,
Now let C be a masa in Qz which contains Bz. Again by solidity and maximal injectivity,
C ⊂ Az. Since Qz is of type II1, there exists two non-zero projections p1, p2 ∈ C and a
partial isometry v ∈ Qz, such that vv∗ = p1, v∗v = p2, p1p2 = 0. Then we have EA(v) =
EA(p1vp2) = p1EA(v)p2 = 0 so that vu ⊥ uv. However we also know that vu = uv, hence
v = 0. This contradicts that p1, p2 6= 0.
Thus, Q has to be of type I. Let C be a masa in Q containing B. Again C ⊂ A. By
Kadison's result [15], C is regular in Q. Both A and Q lie in the normalizer of C, so they
together generate an amenable algebra containing A. By maximal amenability of A, it
follows that Q ⊂ A.
Thus, in order to confirm Peterson's conjecture for the radial masa, it suffices to prove
the strong-AOP as in Proposition 1 for the radial masa. This section is mainly devoted to
establish this fact.
Let Γ = FN +1, N ∈ N. Write K := 2N + 1 for later use. Denote by ωn =Pg∈G,g=n ug,
for n = 1, 2, 3, · · · and let ω0 = ue. Let M = L(Γ) be the free group factor and let
C = {ω1}′′ ⊂ M be the radial masa. {ωn}n≥0 forms an orthogonal basis for L2(C).
Let Ki be the finite-dimensional subspace of H := L2(M ) spanned by all words of
length i and we denote by Qi the orthogonal projection from H onto Ki. For ξ ∈ Ki and
n, m ∈ NS{0}, we define the following
ξn,m :=
Qi+m+n(ωnξωm)
K (n+m)/2
.
Radulescu [27] discovered that there is a nice decomposition of H ⊖ L2(C) = Li≥1 Hi
into a direct sum of C-C-bimodules, each Hi has a distinguished unit vector ξi, which is
from Kl(i), for some l(i) ∈ N, such that Hi is generated by ξi as a C-C-bimodule.
Moreover, by [27, Lemma 3, Lemma 6], for those i with l(i) ≥ 2, we have that
n,m}n,m≥0 forms an orthonormal basis for Hi. For those i with l(i) = 1 (there are
n,m}n,m≥0 is no longer an orthonormal basis for Hi, however for
{ξi
finitely many such i's), {ξi
any i, j ≥ 1, the linear mapping Ti,j : Hi → Hj, given by
Ti,j(ξi
n,m) = ξj
n,m,
extends uniquely to an invertible bounded operator. Furthermore, there is a universal
constant C1 > 0 such that
kT ±1
i,j k ≤ C1, ∀i, j ≥ 1.
Remark 2. Recall that in a separable Hilbert space, a sequence of vectors {vn} forms
a Riesz basis (for the basics of Riesz basis, see, e.g. [5]), if {vi} is the image of some
4
orthonormal basis under some bounded invertible operator. It is also equivalent to the fact
that there exists some A, B > 0 such that for any (cn) ∈ ℓ2, AP cn2 ≤ kP cnvnk ≤
BP cn2. In this case, every vector x in the Hilbert space has a unique decomposition
x = P cnvn, for some (cn) ∈ ℓ2. It follows that (cid:8)ξi
n,m(cid:9)i≥1,n,m≥0 forms a Riesz basis for
x =Pi≥1,n,m≥0 ai
L2(M )⊖ L2(C). Consequently, for any x ∈ L2(M )⊖ L2(C), there is a unique decomposition
n,m}i≥1,n,m≥0 the Radulescu
n,m for some (cid:0)ai
basis for L2(M ) ⊖ L2(C).
n,m(cid:1)n,m,i
∈ ℓ2. We call {ξi
Sometimes it will be convenient to use the following convention: we write ξi
n,mξi
n,m for all
n, m ∈ Z, where we define ξi
n,m = 0 whenever n < 0 or m < 0.
The key computation in [4] is that when considering the AOP in the case of the radial
masa, the Radulescu basis plays the same role as the canonical basis for the generator masa
case. However, in our approach, the Radulescu basis suffers from a lack of right modularity.
Instead, {ωnξiωm}, after proper normalization, is the more natural basis to work with.
We collect some relations between ωnξiωm and ξi
n,m's, due to Radulescu, in the following
lemma:
Lemma 3 (Lemma 2, 6 in [27]). The following statements hold for all n, m ≥ 0:
(1) If l(i) ≥ 2, then ωnξiωm = K
(2) If l(i) = 1, then there is some σ = σ(i) ∈ {−1, 1} such that
n,m−2 + ξi
n,m − K
2 ξi
n+m−2
n+m−4
n+m
2
2
ξi
n−2,m−2;
ωnξiωm = K
n+m
2 ξi
n,m − K
n+m−2
n,m−2 + ξi
n−2,m + σξi
n+m−2k
2
(−σ)kK
+Xk≥2
2
(cid:0)ξi
(cid:0)σξi
(cid:0)ξi
n−2,m(cid:1) + K
n−1,m−1(cid:1)
n−k−1,m−k+1 + σξi
n−k+1,m−k−1 + 2ξi
n−k,m−k(cid:1) ;
(3) For all i, j ≥ 1, the linear mapping Si,j : Hi → Hj given by
Si,j(cid:0)ωnξiωm(cid:1) = ωnξjωm, ∀n, m ≥ 0,
is well-defined and extends to an invertible bounded operator between the two subspaces, with
supi,j kS±1
i,j k ≤ C2, for some uniform constants 0 < C2 < ∞.
ωnξiωm
n,m :=
forms a Riesz basis for L2(M ) ⊖ L2(C).
K (n+m)/2(cid:27)i≥1,n,m≥0
Lemma 4. (cid:26)ηi
Therefore, for any x ∈ L2(M )⊖L2(C), there is a unique decomposition x =Pi≥1,n,m≥0 bi
where (cid:0)bi
Proof. By (3) of the previous lemma, it suffices to prove the conclusion for some fixed i ≥ 1
with l(i) ≥ 2.
n,m(cid:1)i≥1,n,m≥0
∈ ℓ2.
n,mηi
n,m
5
Fix i ≥ 1 with l(i) ≥ 2 and (an,m)n,m ∈ ℓ2. We will omit the superscript i, since no
confusion will appear. Using part (1) of the previous lemma, we have
Xn,m≥0
an,mηn,m = Xn,m≥0
an,m(cid:18)ξn,m −
ξn,m−2
K
−
ξn−2,m
K
+
ξn−2,m−2
=Xn,m(cid:16)an,m −
=Xn,m(cid:18)(cid:16)an,m −
an,m+2
K
−
an+2,m
K
+
an+2,m+2
K 2
an,m+2
K (cid:17) −
1
K (cid:16)an+2,m −
K 2 (cid:19)
(cid:17) ξn,m
K (cid:17)(cid:19) ξn,m,
an+2,m+2
hence by repeated use of the triangle inequality, we have
(cid:13)(cid:13)(cid:13)X an,mηn,m(cid:13)(cid:13)(cid:13)2
an+2,m+2
1/2
2
K (cid:17)(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
2
K (cid:12)(cid:12)(cid:12)
an+2,m+2
1/2
an+2,m −
an,m −
(cid:18)(cid:16)an,m −
=
Xn,m≥0(cid:12)(cid:12)(cid:12)(cid:12)
≥
Xn,m≥0(cid:12)(cid:12)(cid:12)
K(cid:19)
≥(cid:18)1 −
Xn,m≥0(cid:12)(cid:12)(cid:12)
K(cid:19)2
≥(cid:18)1 −
Xn,m≥0
1
1
an,m+2
1/2
an,m+2
K (cid:17) −
2
K (cid:12)(cid:12)(cid:12)
an,m −
1
1
−
K (cid:16)an+2,m −
K
Xn,m≥0(cid:12)(cid:12)(cid:12)
2
K (cid:12)(cid:12)(cid:12)
1/2
an,m+2
1/2
.
an,m2
The other side of the inequality is easy, since each an,m only appears at most four times.
Thus there is a B > 0, such that
i,j k are uniformly bounded, there is a C0 > 0 such
So we are done.
Remark 5. Because both kT ±1
that kT ±1
i,j k ≤ C0, kS±1
2
2
i,j k and kS±1
(cid:13)(cid:13)(cid:13)X an,mηn,m(cid:13)(cid:13)(cid:13)
i,j k ≤ C0, and for any (cid:0)ci
C0 Xi,n,m(cid:12)(cid:12)ci
≤ k Xn,m≥0,i≥1
n,m(cid:12)(cid:12)
C0 Xi,n,m(cid:12)(cid:12)ci
≤ k Xn,m≥0,i≥1
n,m(cid:12)(cid:12)
1
1
2
2
≤ BX an,m2 ,
n,m(cid:1) ∈ ℓ2,
ci
n,mξi
n,mk2
ci
n,mηi
n,mk2
6
,
2
2
2 ≤ C0 Xi,n,m(cid:12)(cid:12)ci
n,m(cid:12)(cid:12)
2 ≤ C0 Xi,n,m(cid:12)(cid:12)ci
n,m(cid:12)(cid:12)
For each k ∈ N, define Lk, L′
k : L2(M ) ⊖ L2(C) → L2(M ) ⊖ L2(C) as follows
ai
n,mξi
Lk
Xi≥1,n,m≥0
k
Xi≥1,n,m≥0
n,m
:= Xi≥1,n≤k,m≥0
n,m
:= Xi≥1,n≤k,m≥0
k) is the left "projection" onto the span of(cid:8)ξi
bi
n,mηi
L′
ai
n,mξi
n,m,
bi
n,mηi
n,m,
i.e. Lk (resp. L′
)
with the first subscript no larger than k. However one should be warned that they are merely
idempotents, instead of projections, due to the presence of those i's with l(i) = 1. We can
also define Rk, R′
k for the right "projections" in the similar fashion. All these idempotents
are bounded operators. Let Lk ∨ Rk := Lk + Rk − LkRk, L′
Lemma 6. L′
n,m(cid:9)i,n,m
n,m(cid:9)i,n,m
(resp.(cid:8)ηi
k is right C-modular, ∀k ≥ 0.
k := L′
k + R′
k − L′
k ∨ R′
kR′
k .
Proof. Since {ωn}n≥0 forms an orthogonal basis for L2(C) and {ηi
basis for L2(M ) ⊖ L2(C), it is sufficient to show that L′
n,m's clearly implies that ηi
n,mωl) = L′
n,mωl ∈ span{ηi
The definition of the ηi
k(ηi
tiplying ωl on the right does not change neither the upper nor left index of ηi
k(ηi
L′
n,m)ωl and the proof is complete.
n,mωl) = L′
k(ηi
n,m}i≥1,n,m≥0 is a Riesz
n,m)ωl.
k(ηi
n,k}k≥0, that is, mul-
n,m, thus
We will need the following result from [4]:
Lemma 7 (Lemma 4.3, Theorem 6.2 in [4]). Given (xk)k ∈ M ω ⊖ C ω, if for every k0 ∈
N, we have that limk→ω k(Lk0 ∨ Rk0) (xk)k2 = 0, then for any y1, y2 ∈ L2(M ) ⊖ L2(C),
y1(xk)k ⊥ (xk)ky2.
Now we state the key technical result of this paper.
Proposition 8. Let Γ = FN +1 be a non-abelian free group with finitely many generators
and M = L(Γ) the corresponding group von Neumann algebra. Denote by C the radial
masa of M and suppose that B ⊂ C is a diffuse von Neumann subalgebra. Then for any
y1(xk)k ⊥ (xk)ky2 in L2(M ω).
We will break the proof into several lemmas.
(xk)k ∈ B′T M ω ⊖ C ω and y1, y2 ∈ M ⊖ C, where ω is a free ultrafilter, we have that
Let (xk)k ∈ B′T M ω ⊖ C ω and y1, y2 ∈ M ⊖ C be given. For each k, we can assume
(cid:8)ξi
n,m(cid:9)i≥1,n,m≥0 and (cid:8)ηi
xk ∈ M ⊖ C ⊂ L2(M ) ⊖ L2(C), xk ≤ 1 and write its decompositions with respect to
n,m(cid:9)i≥1,n,m≥0, respectively:
xk = Xi≥1,n,m≥0
ai,k
n,mξi
n,m = Xi≥1,n,m≥0
bi,k
n,mηi
n,m,
7
are from ℓ2.
n,m(cid:17)i≥1,n,m≥0
where both (cid:16)ai,k
converges to 0 weakly. Recall that (cid:26) ωi
n,m(cid:17)i≥1,n,m≥0
ωi2(cid:27)i≥0
and (cid:16)bi,k
Since B is diffuse, we can choose a sequence {uk}k in the unitary group of B, which
is an orthonormal basis for L2(C). More-
over, for any fixed N0 ≥ 0, ωnωm will be supported on those ωi's with i > N0, provided
that m − n > N0. We first need to approximate each uk using finite linear combinations
of ωi's.
Lemma 9. For each fixed N0, there exists a sequence {Sk}k≥1 of non-empty, disjoint, finite
subsets of N ∪ {0} and a sequence of strictly increasing natural numbers {nk}k≥1, such that
in the decomposition with respect to {ωi}i≥0, the supports of elements from {ωmωn : m ∈
Si, n ≤ N0} and the supports of elements from {ωmωn : m ∈ Sj, n ≤ N0} are disjoint,
whenever i, j ≥ 1, i 6= j. Moreover, there exists a sequence {vk}k in C, with vk ∈ span{ωi :
i ∈ Sk} such that vk ≤ 2 and vk − unk 2 ≤
1
2k .
Moreover, one can construct {vk}, {Sk} such that there is a sequence {Fk} of strictly
increasing natural numbers such that L′
L2(M ) ⊖ L2(C).
N0 (vix) = L′
− L′
Fi+1
N0(cid:16)vi(cid:16)L′
Fi(cid:17) (x)(cid:17), for all x ∈
Proof. Throughout this lemma, for any x ∈ C, we always consider the Fourier expansion
of x with respect to {ωi}i≥0. Moreover, if x =Pi≥0 aiωi and F ⊂ N ∪ {0}, we will use the
notation PF (x) :=Pi∈F aiωi.
We construct {Sk}, {nk} and {vk} inductively. Since span{ωn}n≥0 is a weakly dense
*-subalgebra of C, Kaplansky Density Theorem implies that there exists a sequence {zk}k
of elements in C, whose Fourier expansions are finitely supported, such that zk ≤ 3/2 and
1
4k . For each k, suppose that zk is supported on {ωi}i∈Tk , where Tk ⊂ N∪{0} is
uk −zk2 ≤
some finite subset. Let n1 = 1, v1 = z1 and S1 = T1. Then v1 ≤ 2 and v1 − un12 ≤ 1/2
and v1 ∈ span{ωi : i ∈ S1}.
Now suppose that S1, · · · , Sk and n1, · · · , nk have already been chosen. Then there exists
a finite subset Fk+1 ⊂ N∪{0}, such thatS1≤i≤k Si ⊂ Fk+1 and for any S ⊂ N∪{0}\Fk+1, we
always have that in the decomposition with respect to {ωi}i≥0, the supports of elements from
{ωmωn : m ∈ ∪1≤i≤kSi, n ≤ N0} and the supports of elements from {ωmωn : m ∈ S, n ≤ N0}
are disjoint (for example, one can let Fk+1 = {0, 1, · · · , max ∪1≤i≤k Si + 3N0}). Now since
uk → 0 weakly, there is a natural number nk+1 > nk, such that with respect to the basis
{ωi}i≥0, the Fourier coefficient of znk+1 has absolute value less than
, for each
1
4kFk+1ωi
8
0 ≤ i ≤ Fk+1. Let Sk+1 := Tnk+1\Fk+1, vk+1 := PSk+1(znk+1). Then
= kP(Tn
kvk+1 − unk+1k2 ≤(cid:13)(cid:13)vk+1 − znk+1(cid:13)(cid:13)2 +(cid:13)(cid:13)znk+1 − unk+1(cid:13)(cid:13)2
k+1 \Sk+1)(znk+1)k2 +(cid:13)(cid:13)znk+1 − unk+1(cid:13)(cid:13)2
=(cid:13)(cid:13)PFk+1(znk+1)(cid:13)(cid:13)2 +(cid:13)(cid:13)znk+1 − unk+1(cid:13)(cid:13)2
22k Fk+1(cid:19)1/2
≤(cid:18) Fk+1
1
2k+1 ,
4k+1
+
≤
1
and an easy estimate of the ℓ1-norm gives us
kvk+1k ≤ kzk+1k +
Fk+1
2kFk+1
≤ 2.
The last statement can be achieved by letting the supports of {vk}k mutually far away.
For example, choose the gap between Si and Sj to be greater than 3N0 and let Fk be the
collection of elements of N ∪ {0} between minn∈Sk n − N0 and maxn∈Sk n + N0.
Thus by taking a subsequence if necessary, we may assume that {vk} is a sequence in C,
1
such that vk ∈ span{ωi : i ∈ Sk} for some finite subset Sk ⊂ N, vk ≤ 2, vk − uk2 ≤
2k
and viωk ⊥ vjωk, for all i, j ≥ 1, i 6= j and all 0 ≤ k ≤ N0 and there is a sequence {Fk} of
strictly increasing natural numbers such that L′
all x ∈ L2(M ) ⊖ L2(C).
Fi(cid:17) (x)(cid:17), for
N0(cid:16)vi(cid:16)L′
N0 (vix) = L′
− L′
Fi+1
Lemma 10. limk→ω(cid:13)(cid:13)L′
Proof. Fix a small ǫ > 0. First choose some large N1 < N2 such that 2PN2
N0(xk)(cid:13)(cid:13)2
0 + 1
= 0.
≤ ǫ. Then we have
and
i=N1 kvi − uik2
2 ≤ ǫ
2 C 2
N2 − N1
4(cid:13)(cid:13)L′
N0(cid:13)(cid:13)
N2
N2
lim
k→ω
Xi=N1(cid:10)L′
N0(vixk), L′
N0(vixk)(cid:11) ≥ lim
k→ω
= lim
k→ω
= lim
k→ω
N2
Xi=N1(cid:10)L′
Xi=N1(cid:10)L′
Xi=N1(cid:10)L′
N2
N0(uixk), L′
N0(xkui), L′
N0(xk)ui, L′
k→ω(cid:13)(cid:13)L′
N0(xk)(cid:13)(cid:13)
= (N2 − N1) lim
9
N0(uixk)(cid:11) − ǫ
N0(xkui)(cid:11) − ǫ
N0(xk)ui(cid:11) − ǫ
2
2 − ǫ.
The second line uses the assumption that (xk)k ∈ B′ ∩ M ω and the third line uses the fact
that L′
N0 is a right-C modular map, i.e.
L′
N0(xa) = L′
N0(x)a, ∀x ∈ L2(M ) ⊖ L2(C), ∀a ∈ C.
Meanwhile,
N2
N2
N2
≤
N2
2
Fi+1 − L′
Fi+1 − L′
N0 (vixk) , L′
Xi=N1(cid:10)L′
Fi(cid:17) (xk)(cid:17) , L′
N0(cid:16)vi(cid:16)L′
Xi=N1DL′
N0(vixk)(cid:11) =
Fi(cid:17) (xk), vi(cid:16)L′
Xi=N1Dvi(cid:16)L′
≤(cid:13)(cid:13)L′
N0(cid:13)(cid:13)
Fi(cid:17) (xk)(cid:13)(cid:13)(cid:13)
2 kvik2(cid:13)(cid:13)(cid:13)(cid:16)L′
Xi=N1(cid:13)(cid:13)L′
N0(cid:13)(cid:13)
j≥1,Fi+1≤n≤Fi+1,m≥0(cid:12)(cid:12)(cid:12)
n,m(cid:12)(cid:12)(cid:12)
X
Xi=N1
≤ 4(cid:13)(cid:13)L′
N0(cid:13)(cid:13)
n,m(cid:12)(cid:12)(cid:12)
Xj≥1,0≤n≤FN2 ,m≥0(cid:12)(cid:12)(cid:12)
≤ 4(cid:13)(cid:13)L′
N0(cid:13)(cid:13)
2 ≤ 4(cid:13)(cid:13)L′
≤ 4(cid:13)(cid:13)L′
N0(cid:13)(cid:13)
N0(cid:13)(cid:13)
Therefore, we conclude that limk→ω(cid:13)(cid:13)L′
N0(xk)(cid:13)(cid:13)
trarily small. Thus the proof for Lemma 10 is complete.
N0 2C 2
N2 − N1
2 C 2
0 kxkk2
Fi+1 − L′
2 C 2
0 .
0 + 1
N2
2
2
2
≤
2
bj,k
2 C0
4L′
C0
2
2
2
bj,k
Lemma 11. limk→ω kLN0(xk)k2 = 0.
Proof. We use the relations between ηi
Lemma 4.
n,m and ξi
n,m, as stated above in Lemma 3 and
First, since xk = Pi≥1,l(i)≥2,n,m≥0 ai,k
n,m, it suffices to
consider separately the part with i's such that l(i) ≥ 2 and the part with i's such that
l(i) = 1.
n,mξi
n,mξi
n,m ⊕Pi≥1,l(i)=1,n,m≥0 ai,k
n,m −K −1(cid:0)ξi
n,m−2 + ξi
. Therefore
bi
n+2,m+2
n−2,m(cid:1)+K −2ξi
n−2,m−2
N0(cid:16)vi(cid:16)L′
Fi+1 − L′
Fi(cid:17) (xk)(cid:17)E
Fi+1 − L′
Fi(cid:17) (xk)E
≤ ǫ can be made arbi-
For the i's with l(i) ≥ 2, recall that ηi
n,m = ξi
so that ai
n,m = bi
n,m −
bi
n,m+2
K
−
bi
n+2,m
K
+
K 2
10
LN0
Xi≥1,l(i)≥2,n,m≥0
ai,k
n,mξi,k
n,m
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2
2
2
=
=
ai,k
≤ 16
bi,k
n,m −
n,m(cid:12)(cid:12)(cid:12)
X
X
i≥1,l(i)≥2,N0≥n≥0,m≥0(cid:12)(cid:12)(cid:12)
i≥1,l(i)≥2,N0≥n≥0,m≥0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i≥1,l(i)≥2,N0+2≥n≥0,m≥0(cid:12)(cid:12)(cid:12)
n,m(cid:12)(cid:12)(cid:12)
X
≤ 16C0(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
N0+2
Xi≥1,l(i)≥2,n,m≥0
bi,k
L′
bi,k
n,mηi,k
n,m
,
2
2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
bi,k
n,m+2
K
2
bi,k
n+2,m
K
−
+
bi,k
n+2,m+2
K 2
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
and the last term goes to 0 as k → ω, by the previous lemma.
Now consider the i's with l(i) = 1. As there are only finitely many such i's, we may
restrict our attention to a single fixed i.
For some σ ∈ {1, −1}, we have that
bi,k
n,mηi,k
Xn,m≥0
n,m −
n,m ξi,k
K l (cid:16)σξi,k
(−σ)l
bi,k
n,m =Xn,m
+Xl≥2
=Xn,m bi,k
+Xl≥2
(−σ)l
ξi,k
n−2,m
K
ξi,k
n,m−2
K
−
+ σ
ξi,k
n−1,m−1
K
n−l−1,m−l+1 + σξi,k
n−l+1,m−l−1 + 2ξi,k
n−l,m−l(cid:17)!
n,m −
bi,k
n+2,m
K
bi,k
n,m+2
K
−
+
σbi,k
n+1,m+1
K
K l (cid:16)σbi,k
n+l+1,m+l−1 + σbi,k
n+l−1,m+l+1 + 2bi,k
n+l,m+l(cid:17)!ξi,k
n,m.
Therefore, for any fixed ǫ > 0, N0 ≥ 0, we find a large integer N1 ≫ N0, to be specified
11
later, and we let K0 = N1 − N0. By the triangle inequality,
1/2
Xn≤N0,m≥0
ai,k
n,m2
1
bi,k
n,m −
≤
Xn≤N0,m≥0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
K l
Xn≤N0,m≥0(cid:12)(cid:12)(cid:12)
+ X2≤l≤K0
K l
Xn≤N0,m≥0(cid:12)(cid:12)(cid:12)
+ Xl≥K0+1
1
bi,k
n+2,m
K
bi,k
n,m+2
K
−
+
σbi,k
n+1,m+1
K
1/2
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
σbi,k
n+l+1,m+l−1 + σbi,k
n+k−1,m+k+1 + 2bi,k
σbi,k
n+l+1,m+l−1 + σbi,k
n+l−1,m+l+1 + 2bi,k
1/2
2
n+k,m+k(cid:12)(cid:12)(cid:12)
2
n+l,m+l(cid:12)(cid:12)(cid:12)
1/2
.
We estimate the third term in the above inequality first:
σbi,k
n+l+1,m+l−1 + σbi,k
n+l−1,m+l+1 + 2bi,k
σbi,k
n+l+1,m+l−1 + σbi,k
n+l−1,m+l+1 + 2bi,k
1
1
K l
Xn≤N0,m≥0(cid:12)(cid:12)(cid:12)
K l
Xn,m≥0(cid:12)(cid:12)(cid:12)
K l 4
Xn,m≥0(cid:12)(cid:12)(cid:12)
1
Xl≥K0+1
≤ Xl≥K0+1
≤ Xl≥K0+1
≤ Xl≥K0+1
1/2
bi,k
2
n,m(cid:12)(cid:12)(cid:12)
1
K l 4C0 kxkk2 ≤
4C0
K K0(K − 1)
,
1/2
2
n+l,m+l(cid:12)(cid:12)(cid:12)
2
n+l,m+l(cid:12)(cid:12)(cid:12)
1/2
hence we can choose N1 large enough so that K0 is large, such that the third term is less
than ǫ/3, for any k.
Now we estimate the first and the second terms. To this end, we choose a large k0 =
k0(N1, ǫ), such that for any k ≥ k0, we have that 4C0K0(cid:18)Pm≥0,n≤N1+1(cid:12)(cid:12)(cid:12)
than ǫ/3. Thus both the first and the second term can be bounded above by ǫ/3. Combine
all these pieces together, we conclude that
2(cid:19)1/2
n,m(cid:12)(cid:12)(cid:12)
is less
bi,k
LN0
Xn,m≥0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ai,k
n,mξi,k
n,m
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
≤ C0
Xn≤N0,m≥0
ai,k
n,m2
1/2
≤ C0ǫ,
when k is close enough to ω. Since ǫ > 0 is arbitrary, we are done.
12
Proof of Proposition 8. The same proof for Lemma 11 shows that limk→ω kRN0(xk)k2 = 0.
So Lemma 7 applies.
Remark 12. In fact, the same conclusion as in Proposition 8 holds, if we replace the
assumption "B ⊂ C diffuse" by "B ⊂ C ω diffuse".
Theorem 13. The radial masa satisfies Peterson's conjecture.
Proof. It is shown in [19] that L(FN ), N ≥ 2 is strongly solid, and the fact that the radial
masa is singular is shown in [27](another proof can be found in [29]). Therefore, Proposi-
tion 1 and Proposition 8 imply the result.
Remark 14. One can also use [12, Theorem 8.1] and Proposition 8 to conclude Theorem 13.
We thank Boutonnet for pointing it out to us.
In fact, one can state a more general structural result for the inclusion C ⊂ L(FN +1).
Theorem 15. Let M = L(FN +1) be a free group factor with 1 ≤ N < ∞ and let C ⊂ M
be the radial masa. If Q ⊂ M is a von Neumann subalgebra that has a diffuse intersection
with C, then there exists a sequence of central projections en ∈ Z(Q), n ≥ 0 such that
• e0Q ⊂ C;
• For all n ≥ 1, enQ is a non-amenable II1 factor such that en(Q′ ∩ M ω) = en(Q′ ∩ M )
is discrete and abelian (even contained in C).
Proof. Let e0 ∈ Z(Q) be the maximal projection such that e0Q is amenable. Then Qe0 ⊕
C(1 − e0) is amenable and has a diffuse intersection with C so it is contained in C by
Theorem 13. Moreover, Q(1 − e0) has a discrete center, by solidity of M . This gives a
sequence of central projections {en}n≥1 such that for all n ≥ 1, enQ is a non-amenable II1
factor.
Now fix n ≥ 1. By [14, Lemma 2.7], one can find a central projection e ∈ Z((enQ)′ ∩enM en)
such that
• e((enQ)′ ∩ enM en) = e((enQ)′ ∩ (enM en)ω) is discrete;
• (en − e)((enQ)′ ∩ (enM en)ω) is diffuse.
By [20, Proof of Theorem 4.3], the fact that (en − e)((enQ)′ ∩ (enM en)ω) is diffuse implies
that (en − e)Q is amenable. Since enQ has no direct summand, this forces e = en.
Finally, (Q ∩ C)′ ∩ M is amenable, again by solidity. As it contains C, it has to be equal
to C. In particular Q′ ∩ M ⊂ (Q ∩ C)′ ∩ M ∩ C. So the last part of the theorem is true.
13
Remark 16. In [13, Theorem 3.1], Houdayer showed the general situation for free products
of σ-finite von Neumann algebras, which contains the strong-AOP for the generator masa
in a free group factor as a special case. Also, the strong-AOP as in Proposition 8 means
that for any diffuse subalgebra B of the radial masa C, the inclusion C ⊂ M has the AOP
relative to B, in the sense of [12, Definition 5.1]. The unique maximal injective extension
for any diffuse subalgebra of the generator masa is first shown by Houdayer [13, Theorem
4.1]. A proof via the study of centralizers is obtained by Ozawa [18].
Remark 17. Note that the disjointness result as in Theorem 13 is not true for arbitrary
maximal amenable masa of a II1 factor. For instance, if the inclusion A ⊂ M has some nice
decomposition, then A does not have the uniqueness property as the generator masa in the
above corollary. We give some such examples:
• Let M = A1 ∗A0 A2 be the amalgamated free product with Ai amenable, and A0
diffuse, A0 6= Ai, i = 1, 2, then A0 can be contained in different maximal amenable
subalgebras.
• Let M1, M2 both be the free group factor and Ai ⊂ Mi the corresponding gener-
ator masa, i = 1, 2. Then A = A1⊗A2 is a maximal injective subalgebra inside
M = M1⊗M2. However, many other injective subalgebras could contain the diffuse
subalgebra A1 ⊗ 1.
• Let Λ < Γ be a singular subgroup in the sense of Boutonnet and Carderi ([2, Definition
1.2]) and suppose Γ acts on a finite diffuse amenable von Neumann algebra Q. Then
Q ⋊ Λ is maximal injective inside Q ⋊ Γ, by [2, Theorem 1.3]. However again there are
lots of different injective subalgebras containing Q but are not contained in Q ⋊ Λ.
Remark 18. We would like to mention an example in the ultra-product setting. Let A ⊂ M
be a singular masa inside a separable II1 factor. Then for any free ultrafilter ω, A := Aω is a
maximal injective masa in M := M ω, a result due to Popa ([25, Theorem 5.2.1]). However,
it is well known that any two separable abelian subalegebras in a ultraproduct of II1 factors
are unitarily conjugate ([23, Lemma 7.1]). In particular, A is both contained in a maximal
injective masa and a maximal hyperfinite subfactor of M.
References
[1] R´emi Boutonnet and Alessandro Carderi. Maximal amenable subalgebras of von neumann
algebras associated with hyperbolic groups. arXiv, (1310.5864), 2013.
[2] R´emi Boutonnet and Alessandro Carderi. Maximal amenable von neumann subalgebras arising
from maximal amenable subgroups. arXiv, (1411.4093), 2014.
[3] Arnaud Brothier. The cup subalgebra of a II1 factor given by a subfactor planar algebra is
maximal amenable. Pacific J. Math., 269(1):19 -- 29, 2014.
14
[4] Jan Cameron, Junsheng Fang, Mohan Ravichandran, and Stuart White. The radial masa in a
free group factor is maximal injective. J. London Math. Soc., 82(2):787 -- 809, 2010.
[5] Ole Christensen. Frames, Riesz bases, and discrete Gabor/wavelet expansions. Bull. Amer.
Math. Soc. (N.S.), 38(3):273 -- 291 (electronic), 2001.
[6] Alain Connes. Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1. Ann. of Math.
(2), 104(1):73 -- 115, 1976.
[7] Bent Fuglede and Richard V. Kadison. On a conjecture of Murray and von Neumann. Proc.
Nat. Acad. Sci. U. S. A., 37:420 -- 425, 1951.
[8] Li Ming Ge. On "Problems on von Neumann algebras by R. Kadison, 1967". Acta Math.
Sin. (Engl. Ser.), 19(3):619 -- 624, 2003. With a previously unpublished manuscript by Kadison,
International Workshop on Operator Algebra and Operator Theory (Linfen, 2001).
[9] Liming Ge. On maximal injective subalgebras of factors. Adv. Math., 118(1):34 -- 70, 1996.
[10] Liming Ge. Applications of free entropy to finite von neumann algebras, II. Ann. of Math. (2),
147(1):143 -- 157, 1998.
[11] Cyril Houdayer. A class of II1 factors with an exotic abelian maximal amenable subalgebra.
Trans. Amer. Math. Soc., 366(7):3693 -- 3707, 2014.
[12] Cyril Houdayer. Structure of II1 factors arising from free bogoljubov actions of arbitrary groups.
Adv. Math., 260:414 -- 457, 2014.
[13] Cyril Houdayer. Gamma stability in free product von neumann algebras. Commun. Math.
Phys., 336(2):831 -- 851, 2015.
[14] Adrian Ioana. Cartan subalgebras of amalgamated free product II1 factors. Ann. Sci. ´Ecole
Norm. Sup. (4), 48(1):71 -- 130, 2015.
[15] Richard V. Kadison. Diagonalizing matrices. Amer. J. Math., 106(6):1451 -- 1468, 1984.
[16] F.J. Murray and J. von Neumann. On rings of operators. IV. Ann. of Math. (2), 44(4):716 -- 808,
1943.
[17] Narutaka Ozawa. Solid von Neumann algebras. Acta Math., 192(1):111 -- 117, 2004.
[18] Narutaka Ozawa. A remark on amenable von neumann subalgebras in a tracial free product.
arXiv, 1501.06373, 2015.
[19] Narutaka Ozawa and Sorin Popa. On a class of II1 factors with at most one Cartan subalgebra.
Ann. of Math. (2), 172(1):713 -- 749, 2010.
[20] Jesse Peterson. L2-rigidity in von neumann algebras. Invent. Math., 175:417 -- 433, 2009.
[21] Jesse Peterson and Andreas Thom. Group cocycles and the ring of affiliated operators. Invent.
Math., 185(3):561 -- 592, 2011.
[22] Sorin Popa. Maximal injective subalgebras in factors associated with free groups. Adv. Math.,
50:27 -- 48, 1983.
15
[23] Sorin Popa. Orthogonal pairs of ∗-subalgebras in finite von Neumann algebras. J. Operator
Theory, 9(2):253 -- 268, 1983.
[24] Sorin Popa. On a class of type II1 factors with Betti numbers invariants. Ann. of Math. (2),
163(3):809 -- 899, 2006.
[25] Sorin Popa. A II1 factor approach to the Kadison-Singer problem. Commun. Math. Phys.,
332(1):379 -- 414, 2014.
[26] T. Pytlik. Radial functions on free groups and a decomposition of the regular representation
into irreducible components. J. Reine Angew. Math., 326:124 -- 135, 1981.
[27] Florin Radulescu. Singularity of the radial subalgebra of L (FN ) and the Puk´anszky invariant.
Pacific J. Math., 151(2):297 -- 306, 1991.
[28] Junhao Shen. Maximal injective subalgebras of tensor products of free group factors. J. Funct.
Anal., 240(2):334 -- 348, 2006.
[29] Allan M. Sinclair and Roger R. Smith. The Laplacian MASA in a free group factor. Trans.
Amer. Math. Soc., 355(2):465 -- 475 (electronic), 2003.
[30] D. Voiculescu. The analogues of entropy and of fisher's information measure in free probability
theory III: the absence of cartan subalgebras. Geom. and Funct. Anal., 6(1):172 -- 199, 1996.
Department of Mathematics, Vanderbilt University, 1326 Stevenson Center, Nashville,
TN 37240, United States
E-mail address : [email protected]
16
|
1111.5753 | 1 | 1111 | 2011-11-24T13:29:20 | An equivalence theorem for reduced Fell bundle C*-algebras | [
"math.OA"
] | We show that if E is an equivalence of upper semicontinuous Fell bundles B and C over groupoids, then there is a linking bundle L(E) over the linking groupoid L such that the full cross-sectional algebra of L(E) contains those of B and C as complementary full corners, and likewise for reduced cross-sectional algebras. We show how our results generalise to groupoid crossed-products the fact, proved by Quigg and Spielberg, that Raeburn's symmetric imprimitivity theorem passes through the quotient map to reduced crossed products. | math.OA | math |
AN EQUIVALENCE THEOREM FOR REDUCED FELL BUNDLE
C ∗-ALGEBRAS
AIDAN SIMS AND DANA P. WILLIAMS
Abstract. We show that if E is an equivalence of upper semicontinuous Fell
bundles B and C over groupoids, then there is a linking bundle L(E ) over the
linking groupoid L such that the full cross-sectional algebra of L(E ) contains
those of B and C as complementary full corners, and likewise for reduced cross-
sectional algebras. We show how our results generalise to groupoid crossed-
products the fact, proved by Quigg and Spielberg, that Raeburn's symmetric
imprimitivity theorem passes through the quotient map to reduced crossed
products.
Contents
1.
Introduction
2. Background
3. Linking bundles
4. The reduced norm
4.1.
4.2. Regular Representations and the reduced C ∗-algebra
5. The equivalence theorem
6. The reduced symmetric imprimitivity theorem
References
Induced representations
1
3
5
7
8
9
11
14
16
1. Introduction
The purpose of this paper is to prove a reduced equivalence theorem for cross-
sectional algebras of Fell bundles over groupoids, and to prove that the imprimitivity
bimodule which implements the equivalence between the reduced C ∗-algebras is a
quotient of the Muhly-Williams equivalence bimodule between the full C ∗-algebras
[16].
An increasingly influential interpretation of Hilbert bimodules (or C ∗-correspon-
dences) is to regard them as generalized endomorphisms of C ∗-algebras. Imprim-
itivity bimodules represent isomorphisms, and a Fell bundle over a groupoid G is
then the counterpart of an action of G on a C0(G0)-algebra A. The cross-sectional
algebras of the bundle are analogues of groupoid crossed products. For example,
Date: 1 December 2011.
2000 Mathematics Subject Classification. 46L55.
Key words and phrases. Fell bundle, groupoid, groupoid equivalence, reduced C ∗-algebra,
equivalence theorem, Hilbert bimodule, C ∗-correspondence, Morita equivalence.
1
2
AIDAN SIMS AND DANA P. WILLIAMS
if G is a group and each imprimitivity module is of the form αA for an auto-
morphism α of A (see, for example [18]), then the cross-sectional algebras of such
bundles are precisely those arising from group crossed products; Fell and Doran
called these semidirect products in their magnum opus [5, §VIII.4.2]. In particu-
lar, if A = C0(G(0)) and each fibre of the Fell-bundle is 1-dimensional, then the
cross-sectional algebras are the usual groupoid C ∗-algebras.
r (G) ∼= C ∗
The classical result which motivates this paper is that if groups G and H act
freely, properly and transitively on the same locally compact Hausdorff space P
and the actions commute, then the groups are the same. To see why, fix x ∈ P .
Then for each g ∈ G, there is a unique h ∈ H such that g · x = x · h, and since the
actions commute, g 7→ h is an isomorphism of G with H. Hence C ∗(G) ∼= C ∗(H)
and C ∗
r (H). A particularly powerful viewpoint on this is the following. If
P op is a copy of the space P , but with the actions reversed so that G acts on the
right and H on the left, then L = G ⊔ P ⊔ P op ⊔ H is a groupoid, called the linking
groupoid, with two units. The isotropy at one unit is G and the isotropy at the other
is H, and conjugation in L by any element of P determines an isomorphism from
G to H. At the level of C ∗-algebras, we obtain the following very nice picture: the
actions of G and H on P induce convolution-like products Cc(G) × Cc(P ) → Cc(P )
and Cc(P )×Cc(H) → Cc(P ), and Cc(L) decomposes as a block 2×2 matrix algebra
Cc(L) ∼= (cid:18) Cc(G)
Cc(P op) Cc(H)(cid:19) .
Cc(P )
Moreover, the universal norm on Cc(L) restricts to the universal norm on each of
Cc(G) and Cc(H), and likewise for reduced norms. The characteristic function 1P
of P is a partial isometry in the multiplier algebra of each of C ∗(L) and C ∗
r (L) and
r (G) ∼=
conjugation by 1P implements the isomorphisms C ∗(G) ∼= C ∗(H) and C ∗
C ∗
r (H).
When G and H do not act transitively, the actions of G and H on P induce
actions of G on P/H and of H on G\P . The picture at the level of groups is
now somewhat more complicated, but the C ∗-algebraic picture carries over nicely:
replacing Cc(G) with Cc(G, Cc(P/H)) and Cc(H) with Cc(H, Cc(G\P )) in the
matrix above, we obtain a ∗-algebra L(P ). The C ∗-identity allows us to extend
the norm on C0(P/H) ⋊ G to a norm on L(P ). Moreover, this norm is consistent
with the norm on C0(G\P ) ⋊ H, and the completion of L(P ) in this norm contains
C0(P/H) ⋊ G and C0(G\P ) ⋊ H as complementary full corners. Further, this whole
apparatus descends under quotient maps to reduced crossed products.
To prove an analogue of this equivalence theorem in the context of Fell bundles,
one uses the notion of an equivalence of Fell bundles specified in [16]. The concept
is closely modeled on the situation of groups; but the natural objects on which Fell
bundles act are Banach bundles in which the fibres are equivalence bimodules. That
is, given Fell bundles B and C over groupoids G and H, an equivalence between
the two is, roughly speaking, an upper semicontinuous Banach bundle E over a
space Z such that Z admits actions of G and H making it into an equivalence of
groupoids in the sense of Renault, each fibre Ez of E is an imprimitivity bimodule
from the fibre Br(z) of B over r(z) to Cs(z), and there are fibred multiplication
operations B ∗ E → E and E ∗ C → E which are compatible with the bundle maps,
and which implement isomorphisms Bx ⊗Bu Ez ∼= Ex·z. Muhly and Williams show
in [16, Theorem 6.4] that given such an equivalence, the full cross-sectional algebras
FELL NOTES
3
C ∗(G, B) and C ∗(H, C ) are Morita equivalent (Kumjian proves the corresponding
statement for reduced C ∗-algebras in the r-discrete situation in [13]).
In this paper, we show that Muhly and Williams's Morita equivalence passes to
reduced algebras. We do so by constructing a linking bundle L(E ) = B⊔E ⊔E op⊔C
and showing that Γc(L; L(E )) has a matrix decomposition as above. We then prove
that the completion of L(E ) in the universal norm is a linking algebra for a Morita
equivalence between C ∗(G, B) and C ∗(H, C ), and likewise for reduced C ∗-algebras.
We conclude by showing how to recover a generalisation of Quigg and Spielberg's
theorem [19] which says that the symmetric imprimitivity bimodule arising in Rae-
burn's symmetric imprimitivity theorem [20] passes under the quotient map to an
imprimitivity bimodule for reduced crossed products.
Our reduced equivalence theorem itself is not new: late in the development of this
paper, we learned that Moutou and Tu also prove that equivalent Fell bundles have
Morita equivalent reduced cross-sectional algebras [14]. It appears that Moutou and
Tu deal only with Fell bundles whose underlying Banach bundles are required to be
continuous rather than just upper semicontinuous. (Upper semicontinuous bundles
turn out to be the more natural object in the context of C ∗-algebras -- see [16]
and especially [23, Appendix C]). Moreover Moutou and Tu restrict attention to
principle G-spaces for their groupoid equivalences. But these are minor points and
the arguments of [14] would surely go through unchanged to our setting. The main
new contribution in this article that we develop the linking bundle technology to
show explicitly that the full cross-sectional algebras of the linking bundle is a linking
algebra for the full cross sectional algebras of B and C , and that the quotient map
from the full to the reduced cross-sectional algebra of the linking bundle implements
the quotients C ∗(G, B) → C ∗
r (H, C ). In particular, if
I C
is the ideal of C ∗(H, C ) consisting of elements whose reduced norm is zero,
r
then the equivalence bimodule Xr which we obtain between reduced cross-sectional
algebras is the quotient of the equivalence bimodule X between full algebras by
X · I C
to the corresponding ideal I B
r
of C ∗(G, B).
r . Consequently, induction over X carries I C
r
r (G, B) and C ∗(H, C ) → C ∗
2. Background
Recall that for second-countable locally compact Hausdorff groupoids G and H,
a G -- H equivalence is a locally compact Hausdorff space Z which is simultaneously
a free and proper left G-space and a free and proper right H-space (with continuous
open fibre maps) such that the actions of G and H on Z commute, the map rZ
induces a homeomorphism of Z/H with G(0) and the map sZ induces a homeomor-
phism of G\Z with H (0). Then G acts on Z ∗r Z by g · (y, z) = (g · y, g · z), and
the formula h · [g, h]H = g defines a homeomorphism [·, ·]H : G\(Z ∗r Z) → H; and
G[·, ·] : Z ∗s Z → G is defined similarly (see [15, Definition 2.1] for details).
Recall that an upper semicontinuous Banach bundle over a locally compact Haus-
dorff space Z is a topological space B together with a continuous open surjection
q : B → Z such that each Bz := q−1(z) is a Banach space and: b 7→ kbk is upper
semicontinuous; addition is continuous from B ∗q B → B; scalar multiplication is
continuous on B; and kbik → 0 and q(bi) → z implies bi → 0z ∈ q−1(z). The con-
cept of an upper semicontinuous Banach bundle goes back to [3], where they were
called (H)-bundles, and the work of Hofmann [2, 6 -- 8]. Fell calls such bundles loose
in [4, Remark C.1]. Further details and comments concerning upper semicontinuous
4
AIDAN SIMS AND DANA P. WILLIAMS
Banach bundles are given in [16, Appendix A] and in the C ∗-case in [23, Appen-
dix C]. As in [16], a Fell bundle over a locally compact Hausdorff groupoid G is an
upper semicontinuous Banach bundle q : B → G endowed with a continuous bilin-
ear associative map (a, b) 7→ ab from B(2) := { (a, b) ∈ B × B : s(q(a)) = r(q(b)) }
to B such that
(a) q(ab) = q(a)q(b) for all (a, b) ∈ B(2);
(b) q(a∗) = q(a)−1 for all a ∈ B;
(c) (ab)∗ = b∗a∗ for all (a, b) ∈ B(2);
(d) for each u ∈ G(0), the fibre Au := q−1(u) is a C ∗-algebra under these
operations; and
(e) for each g ∈ G \ G(0), the fibre Bg := q−1(g) is an Ar(g) -- As(g)-imprimi-
tivity bimodule with actions determined by multiplication in B and inner
products ha , bi = a∗b and ha , bi = ab∗.
As a notational convenience, we define r, s : B → G(0) by r(a) := rG(q(a)) and
s(a) := sG(q(a)). See [16] for more details regarding Fell bundles over groupoids.
Remark 1. In the context of bundles over groups, the fibres in a Fell bundle are not
always assumed to be imprimitivity bimodules (they are not assumed to be full --
see [13, 2.4]). Bundles in which all the fibres are indeed imprimitivity bimodules
are then called saturated. We take this condition as part of our definition. It should
also be observed that the underlying Banach bundle of a Fell bundle over a group
is always continuous [1, Lemma 3.30].
Remark 2. In our notation the fibre of B over a unit u can be denoted either Au
or Bu. The dual notation allows us to emphasise its dual roles. We write Au to
emphasise its role as a C ∗-algebra, and Bu to emphasise its role as an imprimitiv-
ity bimodule. The C ∗-algebra A := Γ0(G(0); B) is called the C ∗-algebra of the Fell
bundle B over G(0).
We recall from [16] the definition of an equivalence of Fell bundles. First, fix a
second-countable locally compact Hausdorff groupoid G, a left G-space Z, a Fell
bundle qG : B → G, and a Hausdorff space E together with a continuous open
surjection q : E → Z. Again, as a notational convenience, we shall write r for the
composition rZ ◦ q : E → G(0). we say that B acts on the left of E if there is a
pairing (b, e) 7→ b · e from B ∗ E = { (b, e) ∈ B × E : s(b) = r(e) } to E such that
(a) q(b · e) = qG(b)q(e) for (b, e) ∈ B ∗ E ;
(b) a · (b · e) = (ab) · e whenever (a, b) ∈ B(2) and (b, e) ∈ B ∗ E ;
(c) kb · ek ≤ kbkkek for (b, e) ∈ B ∗ E .1
If E is a right H-space, and qH : C → H is a Fell bundle, then a right action of C
on E is defined similarly.
Now fix second-countable locally compact Hausdorff groupoids G and H and a
G -- H equivalence Z. Suppose that qG : B → G and qH : C → H are Fell bundles.
Fix a Banach bundle q : E → Z. We write E ∗s E for { (e, g) ∈ E × E : s(e) = s(g) }
and we define E ∗r E similarly. We call E a B -- C equivalence if:
(a) there are a left action of B on E and a right action of C on E which
commute;
1The equality appearing in the corresponding item in [16] is a typographical error.
FELL NOTES
5
(b) there are sesquilinear maps
such that the relations
B
h· , ·i : E ∗s E → B and h· , ·i
: E ∗r E → C
C
(i) qG(
(ii)
(iii) b
(iv)
B
B
B
B
he , f i) = G[q(e), q(f )] and qH (he , f i
,
= hf , ei
C
B
he , f i∗ =
he , f i =
he , f i · g = e · hf , gi
hf , ei and he , f i∗
hb · e , f i and he , f i
B
C
C
C
c = he , f · ci
C
and
C
) = [q(e), q(f )]H ,
are satisfied whenever they make sense; and
(c) under the actions described in (a) and the inner-products defined in (b),
each Ez := q−1(z) is an Ar(z) -- Ds(z)-imprimitivity bimodule.
As in [22], if G, H are second-countable locally compact Hausdorff groupoids with
Haar systems λ, β and Z is a G -- H equivalence, we write Z op for the "opposite
equivalence" Z op = { ¯z : z ∈ Z } with r(¯z) = s(z), s(¯z) = r(z), h · ¯z := z · h−1 and
¯z · g := g−1 · z. Then L := G ⊔ Z ⊔ Z op ⊔ H with L(0) := G(0) ⊔ H (0) ⊆ L is a
groupoid containing G and H as subgroupoids: we extend the inverse map to Z and
Z op by setting z−1 := ¯z; and multiplication between Z and G, H is implemented by
the left and right actions, while multiplication between Z and Z op is implemented
by G[·, ·] and [·, ·]H . See [22, Lemma 5] for details. There is a Haar system on L
determined by
κw(F ) :=
RG F (g) dλw(g) +RH F (z · h) dβs(z)(h)
RG F (¯y · g) dλs(¯y)(g) +RH F (h) dβw(h)
if w ∈ G(0)
if w ∈ H (0)
Z and σv
for F ∈ Cc(L) and w ∈ L(0) (see [22, Lemma 6]). For u ∈ G(0) and v ∈ H (0), we
Zop for the restrictions of κu to Z and of κv to Z op. The main results
write σu
of [22] say that C ∗(L, κ) contains C ∗(G, λ) and C ∗(H, β) as the complementary full
corners determined by the multiplier projections 1G(0) and 1H(0) , and that this
Morita equivalence passes under the quotient map C ∗(L, κ) → C ∗
r (L, κ) to reduced
groupoid C ∗-algebras. Our goal in this article is to establish the corresponding
statement for Fell bundles. As a first step, we show in the next section how to
construct from an equivalence of Fell bundles a linking bundle over the linking
groupoid.
3. Linking bundles
Let G and H be locally compact Hausdorff groupoids, let Z be a G -- H equiva-
lence, and let L be the linking groupoid as above. Suppose that pG : B → G and
pH : C → H are upper-semicontinuous Fell bundles, and that q : E → Z is a bundle
equivalence. We denote by A the C ∗-algebra Γ0(G(0); p−1
G (G(0))) of the bundle B,
and by D the C ∗-algebra Γ0(H (0); p−1
H (H (0))) of C ; so the fibre over u ∈ G(0) is
Au, the fibre over v ∈ H (0) is Dv, and each Ez is an Ar(z) -- Ds(z)-imprimitivity
bimodule.
Let E op = { ¯e : e ∈ E } be a copy of the topological space E endowed with
the conjugate Banach space structure α¯e + ¯f = (αe + f ) on each fibre. Then
qop : E op → Z op is an upper-semicontinuous Banach bundle with qop(¯e) = q(e).
We have s(¯e) = s(qop(¯e)) = r(e) and likewise r(¯e) = s(e), so we obtain a right
B-action and a left C -action on E op by
(1)
¯e · b = b∗ · e
and
c · ¯e = e · c∗.
6
AIDAN SIMS AND DANA P. WILLIAMS
h¯e , ¯f i = he , f i
The inner products on E op ∗r E op and E op ∗s E op are given by h¯e , ¯f i
he , f i
. Routine calculations show that each Eop(¯z) is the dual
and
imprimitivity bimodule E(z)∼ of E(z). Since s(z) = r(¯z) and r(z) = s(¯z), axioms
(a), (b) and (c) of [16, Definition 6.1] hold, so E op is a C -- B-equivalence.
=
B
B
C
C
Let L(E ) = B ⊔ E ⊔ E op ⊔ C and define L(q) : L(E ) → L by
L(q)B = pG, L(q)C = pH, L(q)E = q
and L(q)E op = qop.
Since e 7→ ¯e is a fiberwise-isometric homeomorphism from E to E op and since z 7→ ¯z
is a homeomorphism from Z to Z op, the bundle L(E ) is an upper semicontinuous
Banach bundle. Let
L(E )(2) = { (a, b) ∈ L(E ) × L(E ) : s(L(q)(a)) = r(L(q)(b)) }.
Define m : L(E )(2) → L(E ) to coincide with the given multiplications on B and C
and with the actions of B and C on E and E op, and to satisfy
m(e, ¯f ) =
he , f i for (e, f ) ∈ E ∗s E
and m(¯e, f ) = he , f i
for (e, f ) ∈ E ∗r E .
B
C
We define a 7→ a∗ on L(E ) to extend the given involutions on B and C by setting
e∗ = ¯e on E and ¯e∗ = e on E op.
Lemma 3. With notation as above, the bundle L(E ) is a Fell bundle over L.
Moreover, the C ∗-algebra Γ0(L(0); L(q)−1(L(0))) is isomorphic to A ⊕ D.
Proof. We know already that L(E ) is an upper-semicontinuous Banach bundle, that
each L(E )u is a C ∗-algebra and each L(E )x is a L(E )r(x) -- L(E )s(x)-imprimitivity
bimodule. The fibre map q preserves multiplication and involution by definition of
these operations. The operations are continuous because they are continuous on
each component of L(E ) and of L(E ) ∗ L(E ), and the components are topologically
disjoint. That (ab)∗ = b∗a∗ is clear on B ∗ B and C ∗ C , follows from the inner-
product axioms on E ∗ E op an E op ∗ E , and follows from (1) for the remaining
pairings. Associativity for triples from E ∗ E op ∗ E and E op ∗ E ∗ E op follows from
the imprimitivity bimodule axiom
, and is clear for all other
triples.
he , f ig = ehf , gi
⋆
⋆
The map f 7→ (f G(0) , f H(0) ) is a surjection Γ0(L(0); L(q)−1(L(0))) → A ⊕ D,
and the inverse makes sense because G(0) and H (0) are topologically disjoint. Hence
Γ0(L(0); L(q)−1(L(0))) ∼= A ⊕ D
(cid:3)
Resume the hypotheses of Lemma 3.
It is routine to check that (pGϕ)(g) :=
χG(0) (r(g))ϕ(g) determines a bounded self-adjoint map on Γc(G; B) under the
inner-product (ϕ, ψ) 7→ ϕ∗ψ, and hence extends to a multiplier projection, also
denoted pG, of C ∗(G; B). Taking adjoints, (ϕpG)(a) = χG(0) (s(a))ϕ(a). The cor-
responding projection pH for H is defined similarly.
Remark 4. As in [22], we think of ϕ ∈ Γc(L, L(E )) as a matrix
ϕZop ϕH(cid:19)
(cid:18) ϕG
ϕZ
where ϕG is the restriction of ϕ to G ⊆ L and similarly for the other terms. With
respect to this decomposition, we have
ϕψ = (cid:18)ϕGψG + ϕZ ψZop
ϕZop ψG + ϕH ψZ ϕZop ψZ + ϕH ψH(cid:19) ,
ϕGψZ + ϕZ ψH
FELL NOTES
7
where we have used juxtaposition for the convolution product restricted to the
various corners.2 Moreover ϕG = pGϕpG, ϕZ = pGϕpH , ϕZop = pH ϕpG, and
ϕH = pHϕpH .
Lemma 5. Resume the hypotheses of Lemma 3. Then pG and pH are full multiplier
projections of C ∗(L; L(E )).
Proof. We just show that pG is full; the corresponding statement for pH follows by
symmetry. Fix ϕ, ψ ∈ Γc(L; L(E )). Using the matrix notation established above,
we have
ϕpGψ = (cid:18) ϕGψG
ϕZop ψG ϕZop ψZ(cid:19) .
ϕGψZ
That elements of the form ϕGψG span a dense subalgebra of Γc(G; B) is clear. That
elements of the form ϕGψZ span a dense subspace of Γc(Z; E ) and likewise that
elements of the form ϕZop ψG span a dense subspace of Γc(Z op; E op) follows from
[16, Proposition 6.10]. That elements of the form ϕZop ψZ span a dense subspace of
Γc(H; C ) follows from the argument which establishes axiom (IB2) in [16, Section 7].
(cid:3)
Recall that the inductive-limit topology on Cc(X) for a locally compact Haus-
dorff space X is the unique finest locally convex topology such that for each compact
K ⊆ X, the inclusion of Cc(X)K = {f ∈ Cc(X) : supp(f ) ⊆ K} into Cc(X) is con-
tinuous (see for example [4, II.14.3] or [21, §D.2]). In particular, [21, Lemma D.10]
says that to check that a linear map L from Cc(X) into any locally convex space
M is continuous, it suffices to see that if fn → f uniformly and if all the supports
of the fn are contained in the same compact set K, then L(fn) → L(f ).
Remark 6. We are now in a situation analogous to that of [22, Remark 8]. By
the Disintegration Theorem for Fell bundles, [16, Theorem 4.13], any pre-C ∗-norm
k · kα on Γc(L; L(E )) which is continuous in the inductive-limit topology is dom-
inated by the universal norm. Hence the argument of [22, Remark 8] shows that
pGC ∗
α(H; C )-imprimitivity bimodule. So to prove
that C ∗(G; B) is Morita equivalent to C ∗(H, C ) we just need to show that for
α(L; L(E ))pH is a C ∗
α(G; B) -- C ∗
F ∈ pGΓc(L; L(E ))pG, the universal norms kF kC ∗(L,L(E )) and (cid:13)(cid:13)F G(cid:13)(cid:13)C ∗(G;B) coin-
cide, and similarly for the reduced algebras (the corresponding statements for H
hold by symmetry).
4. The reduced norm
In this section we recall the construction of the reduced cross-sectional algebra
of a Fell bundle. We first discuss how to induce representations from C ∗-algebra
of the restriction of a Fell bundle to a closed subgroupoid up to representations
of the C ∗-algebra of the whole bundle. We then apply this construction to the
closed subgroupoid G(0) of G to induce representations of the C ∗-algebra A =
Γ0(G(0); BG(0)) up to representations of C ∗(G; B). These are, by definition, the
regular representations whose supremum determines the reduced norm.
2In fact, the products in the matrix can be expressed in terms of the inner-products and module
actions from [16, Theorem 6.4].
8
AIDAN SIMS AND DANA P. WILLIAMS
4.1. Induced representations. Let G be a second countable locally compact
Hausdorff groupoid with Haar system {λu}u∈G(0). Let q : B → G be a separable
Fell bundle as described in [11, §1.3]. Assume that H is a closed subgroupoid of G
with Haar system {αu}u∈H(0) . We write qH : BH → H for the Fell bundle obtained
by restriction to H. We want to induce representations of C ∗(H, BH ) to C ∗(G; B)
using the Equivalence Theorem [16, Theorem 6.4] for Fell bundles. We will use the
set-up and notation from [10, §2]. In particular, we recall that GH(0) = s−1(H (0)) is
a (H G, H)-equivalence where H G is the imprimitivity groupoid (GH(0) ∗s GH(0) )/H.
Let σ : H G → G be the continuous map given by σ(cid:0)[x, y](cid:1) = xy−1. The pull-back
Fell bundle σ∗q : σ∗B → H G is the Fell bundle σ∗B = {([x, y], b) : [x, y] ∈ H G, b ∈
B, σ([x, y]) = q(b)} with bundle map σ∗([x, y], b) = [x, y] over H G.
Let E = q−1(GH(0) ); then q restricts to a map q : E → GH(0) . We wish to make
this Banach bundle into a σ∗B -- BH -equivalence (see [16, Definition 6.1]). It is
clear how BH acts on the right of E , and we get a left action of σ∗B via
(Since q(b) = xy−1, q(be) = xh as required.) The "inner products" on E ∗r E and
E ∗s E are given by
(cid:0)[x, y], b(cid:1) · e := be
for q(e) = yh.
he , f i
BH
= e∗f
and
σ∗B
he , f i = (cid:0)[q(e), q(f )], ef ∗(cid:1),
respectively. It now straightforward to check that E is a σ∗B -- BH -equivalence.
By [16, Theorem 6.4], Γc(GH(0) ; E ) is a pre-imprimitivity bimodule with actions
and inner products determined by
(2)
(3)
(4)
(5)
F · ϕ(z) = ZG
ϕ · g(z) = ZH
(h) = ZG
hϕ , ψi(cid:0)[x, y](cid:1) = ZG
hϕ , ψi
⋆
⋆
F(cid:0)[z, y](cid:1)ϕ(y) dλs(z)(y),
ϕ(zh)g(h−1) dαs(z)(h),
ϕ(y)∗ψ(yh) dλr(h)(y),
ϕ(xh)ψ(yh)∗ dαs(x)(h)
for F ∈ Γc(H G; σ∗B), ϕ, ψ ∈ Γc(GH(0) ; E ) and g ∈ Γc(H; BH ). The completion
X = XG
H is a C ∗(H G, σ∗B) -- C ∗(H, BH )-imprimitivity bimodule.
Remark 7. It is pleasing to note that the formalism of Fell bundles is such that
equations (2) -- (5) are virtually identical to those in the scalar case: see [10, Eq.
(1) -- (4)].3 The only difference is that complex conjugates in the scalar case are
replaced by adjoints.
To construct induced representations using the machinery of [21, Proposi-
tion 2.66], we need a nondegenerate homomorphism V : C ∗(G; B) → L(X) which
will make X into a right Hilbert C ∗(G; B) -- C ∗(H, BH )-bimodule (the data needed
to induce representations a la Rieffel.) Define V : Γc(G; B) → Lin(Γc(GH(0) ; E ))
by
(6)
V (f )(ϕ) := ZG
f (y)ϕ(y−1z) dλr(z)(y).
3Well, they would be if it weren't for the typos in equations (1) and (4) in [10].
FELL NOTES
9
By the Tietz Extension Theorem for upper semicontinuous Banach bundles [16,
Proposition A.5], each ϕ ∈ Γc(GH(0) ; E ) is the restriction of an element of Γc(G; B).
So the argument of [10, Remark 1] and the paragraph which follows yields
hV (f )ϕ , ψi
⋆
= hϕ , V (f ∗)ψi
⋆
.
The map f 7→ hV (f )ϕ , ψi
⋆
existence of approximate units in Γc(G; B) implies that
is continuous in the inductive-limit topology, and the
{ V (f )ϕ : f ∈ Γc(G; B) and ϕ ∈ Γc(GH(0) ; E ) }
spans a dense subspace of Γc(GH(0) ; E ). Then [11, Proposition 1.7] implies that V
is bounded and extends to a nondegenerate homomorphism as required.
Now if L is a representation of C ∗(H, BH ), then the induced representation
H L of C ∗(G; B) acts on the completion of X ⊙ HL with respect to
IndG
Fix ϕ ∈ Γc(GH(0) ; E ). Writing f · ϕ for V (f )ϕ, we have
(ϕ ⊗ h ψ ⊗ k) = (cid:0)L(hψ , ϕi
⋆
)h k(cid:1)HL
.
and, as in [10, Remark 1], f · ϕ = f ∗ ϕ.
(IndG
H L)(f )(ϕ ⊗ h) = f · ϕ ⊗ h,
4.2. Regular Representations and the reduced C ∗-algebra. Regular repre-
sentations are, by definition, those induced from A = Γ0(G(0); B). Thus, in the
notation of Section 4.1, H = G(0), GH(0) = G, E = B, and we write A in place of
BG(0) (see Remark 2); in particular each B(x) is a A(r(x)) -- A(s(x))-imprimitivity
= ϕ∗ψG(0) , with the product being computed in
bimodule. We also have hϕ , ψi
⋆
Γc(L; L(E )).
Let π be a representation of A on Hπ. Let π be the extension of π to M (A), and
let i : C0(G(0)) → M (A) be the map characterised by (cid:0)i(f )a(cid:1)(u) = f (u)a(u) for
u ∈ G(0). Then ϕ := π ◦ i is a representation of C0(G(0)) on Hπ which commutes
with π. Example F.25 of [23] shows that there is a Borel Hilbert bundle G(0) ∗ H
and a finite Radon measure µ on G(0) such that π is equivalent to a direct integral
R ⊕
G(0) πu dµ, and such that if L : f → Lf is the diagonal inclusion of C0(G(0))
in B(L2(G(0) ∗ H , µ)), then π(i(f )a) = Lf π(a) for a ∈ A. So each πu factors
through Au. We will usually write πu(cid:0)a(u)(cid:1) in place of πu(a) for a ∈ A. See
on the completion of Γc(G; B) ⊙ L2(G(0) ∗ H , µ) with respect to (cid:0)ϕ ⊗ h ψ ⊗ k(cid:1) =
(cid:0)π(cid:0)ψ∗ ∗ ϕ(cid:1)h k(cid:1), and a quick calculation yields
[17, p. 46] for more details. The regular representation Ind π = IndG
G(0) π then acts
(7)
(cid:0)ϕ ⊗ h ψ ⊗ k(cid:1) = ZG(0) ZG(cid:0)πu(cid:0)ψ(x)∗ϕ(x)(cid:1)h(u) k(u)(cid:1) dλu(x) dµ(u).
Then Ind π acts by:
(8)
(Ind π)(f )(ϕ ⊗ h) = V (f )(ϕ) ⊗ h = f · ϕ ⊗ h.
We next define the reduced algebra of a Fell bundle. We define the reduced norm
by analogy with the one-dimensional case as the supremum of the norms determined
by induced representations of A. We then show that this agrees, via V , with the
operator norm on L(X). This is equivalent to Definition 2.4 and Lemma 2.7 of [14],
though the roles of definition and lemma are interchanged.
10
AIDAN SIMS AND DANA P. WILLIAMS
Definition 8. We define the reduced norm on Γc(G; B) by
kf kr := sup{ k(Ind π)(f )k : π is a representation of A }.
Since the kernel of Ind π depends only on the kernel of π (see [21, Corollary 2.73]),
we have kf kr = k(Ind π)(f )k for any faithful representation π of A. We define the
reduced C ∗-algebra C ∗
r (G,B) :=
{ a ∈ C ∗(G, B) : kakr = 0 }.
r (G; B) of B to be the quotient of C ∗(G; B) by IC ∗
Lemma 9. Let X = XG
termined by (6). Then ker V = IC ∗
C ∗
r (G; B) into L(X). In particular, kV (f )k = kf kr.
G(0) and V : C ∗(G; B) → L(X) the homomorphism de-
r (G;B) and V factors through an injection of
Proof. Let π be a faithful representation of A. Then for any x ∈ X, h ∈ Hπ and
f ∈ C ∗(G; B), we have
(9)
(cid:13)(cid:13)(Ind π)(f )(x ⊗ h)(cid:13)(cid:13)
= (cid:13)(cid:13)V (f )(x) ⊗ h(cid:13)(cid:13)
2
2
= (cid:0)π(cid:0)hV (f )(x) , V (f )(x)i
⋆(cid:1)h h(cid:1).
Thus if V (f ) = 0, then (Ind π)(f ) = 0. On the other hand, given x and f , we can
find a unit vector h such that the right-hand side of (9) is at least
1
2(cid:13)(cid:13)π(cid:0)hV (f )(x) , V (f )(x)i
⋆(cid:1)(cid:13)(cid:13) =
2(cid:13)(cid:13)V (f )(x)(cid:13)(cid:13)
2
.
1
Therefore Ind π(f ) = 0 implies that V (f ) = 0. We have shown that ker V =
ker(Ind π), and hence V factors through an injection of C ∗
r (G; B) into L(X) as
claimed.
(cid:3)
We digress briefly to check that the definition of the reduced C ∗-algebra which
we have given is compatible with existing definitions on some special cases.
Example 10 (The Scalar Case: Groupoid C ∗-Algebras). Let B = G × C so
that C ∗(G; B) = C ∗(G). So A = C0(G(0)), and π defined by multiplication on
L2(G(0), µ) is a faithful representation of A. Then Ind π acts on the completion H
of Cc(G) ⊙ L2(G(0)) and, if we let ν = µ ◦ λ, then (7) becomes
(cid:0)ϕ ⊗ h ψ ⊗ k(cid:1) = ZG(cid:0)ϕ(x)h(s(x)) ψ(x)k(s(x))(cid:1) dν−1(x).
Hence there is a unitary U from H onto L2(G, ν−1) defined by U (ϕ ⊗ h)(x) =
ϕ(x)h(s(x)), and U intertwines Ind π with the representation (Ind µ)(f )ξ(x) =
RG f (y)ξ(y−1x) dλr(x)(y). Hence our definition of the reduced norm agrees with
the usual definition (see [22, §3], for example), and C ∗
r (G × C) = C ∗
r (G).
Example 11 (Groupoid Crossed Products). Suppose that (A , G, α) is a dynamical
system and form the associated semidirect product Fell bundle B = r∗A as in
[16, Example 2.1]. Working with the appropriate A -valued functions, as in [16,
Example 2.8], a quick calculation starting from (7) gives
(10)
(cid:0)f ⊗ h g ⊗ k(cid:1) =
ZG(0) ZG(cid:0)πu(cid:0)α−1
x (cid:0)f (x)(cid:1)(cid:1)h(u) πu(cid:0)α−1
x (cid:0)g(x)(cid:1)(cid:1)k(u)(cid:1) dλu(x) dµ(u).
(Since λu is supported on Gu, each α−1
x (g(x)) ∈ Au, so the integrand makes sense.)
Let G ∗ Hs be the pull back of G(0) ∗ H via s. Given a representation π of A, there
is a unitary U from the space of Ind π to L2(G ∗ Hs, ν−1) defined by U (f ⊗ h) =
FELL NOTES
11
πs(x)(cid:0)α−1
given by
x (cid:0)f (x)(cid:1)(cid:1)h(cid:0)s(x)(cid:1). This U intertwines Ind π with the representation Lπ
(11)
Lπ(f )ξ(x) = ZG
πs(x)(cid:0)α−1
x (cid:0)f (y)(cid:1)(cid:1)ξ(y−1x) dλr(x)(y).
Applying this with a faithful representation π of A, we deduce that A ⋊α,r G ∼=
C ∗
r (G, r∗A ).
Remark 12. In Examples 10 and 11, the essential step in finding a concrete real-
ization of the space of Ind π is to "distribute the πu" in the integrand in (7) to
both sides of the inner product. But for general Fell bundles, πu(ϕ(x)) makes no
sense for general x ∈ Gu. This often makes analyzing regular representations of
Fell bundle C ∗-algebras considerably more challenging.
Example 13. Any representation πu of Au determines a representation πu ◦ ǫu of
A by composition with evaluation at u (in the direct-integral picture, π = πu ◦ ǫu
is a direct integral with respect to the point-mass δu). We abuse notation slightly
and write Ind πu for Ind(πu ◦ ǫu) which acts on the completion of Γc(G; B) ⊙ Hπu
under
(12)
(ϕ ⊗ h ψ ⊗ k) = ZG(cid:0)πu(cid:0)ψ(x)∗ϕ(x)(cid:1)h(u) k(u)(cid:1) dλu(x).
Equation (12) depends only on ϕGu and ψGu; and conversely each element of
Γc(Gu; B) is the restriction of some ϕ ∈ Γc(G; B) by the Tietz Extension Theorem
for upper semicontinuous Fell bundles [16, Proposition A.5]. So we can view the
space of Ind πu as the completion of Γc(Gu; B) ⊙ Hπu with respect to (12).
5. The equivalence theorem
Fix for this section second-countable locally compact Hausdorff groupoids G and
H with Haar systems λ and β, a G -- H equivalence Z, Fell bundles pG : B → G and
pH : C → H and a B -- C equivalence q : E → Z. Let κ denote the Haar system on
L obtained from [22, Remark 11], and let L(q) : L(E ) → L be the linking bundle
of Section 3.
Theorem 14. Suppose that F ∈ Γc(L; L(E )) satisfies f (ζ) = 0 for all ζ ∈ L \ G.
Let f := F G ∈ Γc(G; B). Then kF kC ∗(L,L(E )) = kf kC ∗(G;B) and kF kC ∗
r (L;L(E )) =
r (G;B). Moreover, pGC ∗(L, L(E ))pH is a C ∗(G; B) -- C ∗(H; C )-imprimitivity
kf kC ∗
bimodule, and pGC ∗
r (H, C )-imprimitivity bimodule
which is the quotient module of pGC ∗(L, L(E ))pH by the kernel Ir of the canonical
homomorphism of C ∗(H, C ) onto C ∗
r (L, L(E ))pH is a C ∗
r (G; B) -- C ∗
r (H, C ).
Remark 15. Recall the set-up of Example 10. It is not difficult to see that if G and
H are groupoids and Z is a G -- H equivalence, then the trivial bundle Z × C is a
(G × C) -- (H × C) equivalence. Hence we recover Theorem 13, Proposition 15 and
Theorem 17 of [22] from Theorem 14.
To prove Theorem 14, we first establish some preliminary results. Our key tech-
nical result is a norm-estimate for the representations of Γc(G; B) ⊆ Γc(L; L(E ))
coming from elements of H 0.
Let {ρv
Z}v∈H(0) be the Radon measures on Z introduced in [22, Theorem 13].
For each v ∈ H (0), fix ζ ∈ Z with s(ζ) = v, and define a D(v)-valued form on
12
AIDAN SIMS AND DANA P. WILLIAMS
Y0 = Γc(Z; E ) by
(13)
hϕ , ψi
D
(v) = ZG
hϕ(x−1 · ζ) , ψ(x−1 · ζ)i
C
dλr(ζ)(x).
Left-invariance of ρ implies that this formula does not depend on the choice of
(v) is the restriction to L(0)
ζ ∈ Z such that s(ζ) = v. The map (ϕ, ψ) 7→ hϕ , ψi
of the product ϕ∗ψ computed in Γc(L; L(E )).
D
The following lemma constructs what is essentially an "integrated form" of the
modules used in [14, Proposition 4.3].
Lemma 16. With respect to the pre-inner product (13), Y0 is a pre-Hilbert D-
module whose completion, Y is a full right Hilbert D-module.
D
D
·i
= (ϕ∗ψ)L(0) .
,
Proof. That (13) takes values in D follows from the observation above that
hϕ , ψi
Since each Ez is an imprimitivity bimodule, the
(v) is all of Dv. Since D is a C0(H (0))-algebra, to see that
range of h·
X := span{ hϕ , ψi
: ϕ, ψ ∈ Y0 } = D, it therefore suffices to show that X is a
C0(H (0))-module (see, for example [23, Proposition C.24]), for which one uses the
right action of C0(H (0)) on Y0 to check that (hϕ , ψi
(v)
is bilinear from X × C0(H (0)) to X. An argument like that of page 6 shows that
for f ∈ Γc(L(0); L(q)−1(L(0))), Mf (ψ)(g) := f (r(g))ψ(g) determines a multiplier of
C ∗(L; L(E )), so Y0 is a pre-Hilbert D-module which is full since X = D.
(cid:3)
· f )(v) := f (v)hϕ , ψi
D
D
D
Remark 17. Since D is a C0(H (0))-algebra, to each v ∈ H (0) there corresponds a
quotient module
Y(v) := Y /Y · Iv,
where Iv = { d ∈ D : d(v) = 0 }. As in [21, Proposition 3.25], Y(v) is a right
if we denote by x(v) the image of x in Y(v), then we have
Hilbert D(v)-module:
= hx , yi
k, we obtain kyk =
Indeed, the Y(v) are isomorphic to the modules used in [14,
(cid:10)x(v) , y(v)(cid:11)D(v)
(v). Since kyk2 = khy , yi
D
D
supv∈H(0) ky(v)k.
Proposition 4.3].
Fix T ∈ L(Y).
for each v ∈ H (0) there is an ad-
jointable operator Tv on Y(v) satisfying Tv(x(v)) = (T x)(v) for all x ∈ Y. Since
Since T is D-linear,
(cid:10)Tv(x(v)) , y(v)(cid:11)D(v)
= hT x , yi
D
(v), we have kT k = supv∈H(0) kTvk.
Proposition 18 ([14, Proposition 4.3]). There is a homomorphism M from
C ∗(G; B) to L(Y) such that if f ∈ Γc(G; B) and ϕ ∈ Y0, then
(14)
M (f )ϕ(ζ) = ZG
f (x)ϕ(x−1 · ζ) dλr(ζ)(x).
We have IC ∗
kM (f )k ≤ kf kr.
r (G;B) ⊂ ker M , and M factors through C ∗
r (G; B).
In particular,
Proof. Direct computation shows that
hM (f )ϕ , ψi
D
= hϕ , M (f ∗)ψi
D
for f ∈ Γc(G; B) and ϕ, ψ ∈ Γc(Z; E ).
A calculation using Remark 17, the Cauchy-Schwartz inequality for Hilbert mod-
ules ([21, Lemma 2.5]) and the characterization of inductive-limit topology con-
tinuous maps out of Cc(G) in terms of eventually compactly supported uniform
is continuous in the inductive-limit
convergence shows that f 7→ hM (f )ϕ , ψi
D
FELL NOTES
13
topology. The existence of approximate identities as in [16, Proposition 6.10] then
implies that span{ M (f )ϕ : f ∈ Γc(G; B) and ϕ ∈ Y0 } is dense in Y in the induc-
tive limit topology, so [11, Proposition 1.7] implies that M is bounded and extends
to C ∗(G; B).
Since kM (f )k = supv∈H(0) kMv(f )k, it now suffices to show that
(15)
kMv(f )k ≤ kf kr
for all v ∈ H (0).
Fix v ∈ H (0) and choose ζ ∈ Z such that s(ζ) = v. Let u = r(ζ). For any
e ∈ q−1(ζ) we have
B(cid:10)ϕ(x · ζ) , e(cid:11) ∈ B(G[x · ζ, ζ]) = B(x). Thus we can define
Ue : Γc(Z; E ) → Γc(Gu; B) by U (ϕ)(x) =
B(cid:10)ϕ(x · ζ) , e(cid:11).
Just as in Remark 17, we can form the quotient module X(u), and the map
V : C ∗(G; B) → L(X) from Lemma 9 gives operators Vu(f ) ∈ L(X(u)) such that
(u) depends only on the ϕGu and ψGu,
kVu(f )k ≤ kf kr. The inner product hϕ , ψi
⋆
and every element of Γc(Gu; B) extends to an element of Γc(G; B) by the Tietz Ex-
tension Theorem for upper semicontinuous Banach bundles [16, Proposition A.5].
So we can view Ue as a map from Γc(Z; E ) to X(u).
Using that
one computes to see that
B(cid:10)ϕ(x−1ζ) , e(cid:11)∗
B(cid:10)ϕ(x−1 · ζ) , e(cid:11) =
B(cid:10)e · hϕ(x−1 · ζ) , ϕ(x−1 · ζ)i
C
, e(cid:11),
(16)
hUe(ϕ) , Ue(ϕ)i
⋆
(u) = (cid:10)e · hϕ , ϕ(cid:11)B
(v) , ei
D
.
So if kek ≤ 1, the Cauchy-Schwartz inequality for Hilbert modules ([21, Lemma 2.5])
implies that kUe(ϕ)kX(u) ≤ kϕkY(v).
For x ∈ Gu, that the pairing
Ue(M (f )ϕ)(x) = ZG
B(cid:10)· , ·(cid:11) is A-linear in the first variable gives
B(cid:10)f (y)ϕ(y−1x · ζ) , e(cid:11) dλr(x)(y) = Vu(f )U (ϕ)(x).
Fix f ∈ Γc(G; B) and ǫ > 0. Fix ϕ ∈ Γc(Z; E ) such that kϕkY(v) = 1 and such
that kM (f )ϕkY(v) > kMv(f )k − ǫ. By (16), there exists e ∈ q−1(ζ) with kek = 1
such that
Hence
kUe(M (f )ϕ)kX(u) > kM (f )ϕkY(v) − ǫ.
kMv(f )k − 2ǫ < kM (f )ϕkY(v) − ǫ < kUe(M (f )ϕ)kX(u) = kVu(f )Ue(ϕ)kX(u) ≤ kf kr.
Letting ε → 0 gives (15).
(cid:3)
Proof of Theorem 14. Since every representation of C ∗(L; L(E )) restricts to a rep-
resentation of C ∗(G; B), we have kF kC ∗(L;L(E )) ≤ kf kC ∗(G;B), so we just have to
establish the reverse inequality. The argument for this is nearly identical to that
of [22, Proposition 15]. The key differences are that: [16, Theorem 6.4] is used in
place of [17, Theorem 5.5]; and [16, Proposition 6.10] is used to obtain an approx-
imate identity for both Γc(G; B) and Γc(Z; E ) which can be used in place of the
approximate identity in C (L) to establish the analogue of [22, Equation (10)] and
to complete the norm approximation at the end of the proof.
We now turn to the proof that the reduced norms agree. Fix faithful represen-
tations πu of the A(u) and τv of the D(v). Then
kF kC ∗
r (L;L(E )) = maxn sup
u∈G(0)
k(IndL πu)(F )k, sup
v∈H(0)
k(Ind τv)(F )ko.
14
AIDAN SIMS AND DANA P. WILLIAMS
Fix u ∈ G(0). Let H1 be the space of IndG πu; that is, the completion of
Γc(Gu; B) ⊙ Hπu as in Example 13. The representation IndL πu acts on the
completion of Γc(Lu; L(E )) ⊙ Hπu which decomposes as H1 ⊕ H2 where H1 =
Γc(Gu; L(E )) ⊙ Hπu and H2 = Γc(Zu; L(E )) ⊙ Hπu . Moreover, the restriction of
IndL πu to H2 is the zero representation. Hence
kF kC ∗
r (L;L(E )) ≥ sup
u∈G(0)
k IndL πu(f )k = kf kC ∗
r (G;B),
and it suffices now to establish that k IndL τv(F )k ≤ supu k IndL πu(F )k for all
F ∈ Γc(L; L(E )) and v ∈ H (0).
Fix v ∈ H (0). Then IndL τv acts on the completion of Γc(Lv; L(E )) ⊙ Hτv which
again decomposes as a direct sum H3 ⊕ H4 (here H3 = Γc(Z op
v ; L(E )) ⊙ Hτv and
H4 = Γc(Hv; L(E )) ⊙ Hτv ). The restriction to H4 is zero, and H3 is the completion
of Γc(Z; E ) under
(ϕ ⊗ h ψ ⊗ k) = ZZ(cid:0)τv(cid:0)(cid:10)ψ(ζ) , ϕ(ζ)(cid:11)C(cid:1)h k(cid:1) dρv(ζ).
An inner-product computation shows that if Y is the Hilbert D-module of
Lemma 16, then H3 is isomorphic to the completion of Y ⊙ Hτv under
(ϕ ⊗ h ψ ⊗ k) = (cid:0)τv(cid:0)hψ , ϕi
D
(v)(cid:1)h k(cid:1),
and then the restriction of (IndL τv)(F ) to H3 is Y- Ind τv. Hence k IndL τv(F )k =
kY- Ind τv(f )k. Since Y- Ind τv[x ⊗ h] = [M (f )x ⊗ h] for all x, we have ker M ⊂
ker Y- Ind τv. Hence Proposition 18, implies that kY- Ind τv(f )k ≤ kf kC ∗
r (G;B) as
required.
The final statement follows from [21, Theorem 3.22].
(cid:3)
6. The reduced symmetric imprimitivity theorem
Suppose that K and H are locally compact groups acting freely and properly
on the left and right, respectively, of a locally compact space P . Suppose also that
we have commuting actions α and β of K and H, respectively, on a C ∗-algebra
D. Then we can form the induced algebras IndP
K(D, α) and get
dynamical systems
H (D, β) and IndP
σ : K → Aut(cid:0)IndP
H (D, β)(cid:1)
and τ : H → Aut(cid:0)IndP
K(D, α)(cid:1)
for the diagonal actions as in [23, Lemma 3.54]. Then Raeburn's Symmetric Im-
primitivity Theorem says that the crossed products
IndP
H (D, β) ⋊σ K and
IndP
K(D, α) ⋊τ H
are Morita equivalent. In [19], Quigg and Spielberg proved that Raeburn's Morita
equivalence passed to the reduced crossed products.
(Kasparov had a different
proof in [12, Theorem 3.15] and an Huef and Raeburn gave a different proof of the
Quigg and Spielberg result in [9, Corollary 3].)
We consider the corresponding statements for groupoid dynamical systems. Let
(A , G, α) be a groupoid dynamical system as in [17, §4]. Recall that the associated
crossed product A ⋊α G is a completion of Γc(G; r∗A ). If π is a representation
of A := Γ0(G(0); A ), then the associated regular representation of A ⋊α G is the
representation Lπ := IndA π acting on L2(G ∗ Hs, ν−1) as in (11). The reduced
crossed product, A ⋊α,r G is the quotient of A ⋊α G by the common kernel of the
FELL NOTES
15
Lπ with π faithful. Let B := r∗A with the semidirect product Fell bundle structure
so that C ∗(G, B) is isomorphic to A ⋊α G, then it follows from Example 11 that
C ∗
r (G, B) is isomorphic to the reduced crossed product A ⋊α,r G.
Now let H be a locally compact group and let X be a locally compact Hausdorff
H-space. Let G be the transformation groupoid G = H × X. As in [17, Exam-
ple 4.8], suppose that A = Γ0(X; A ) is a C0(X)-algebra, and define lt : H →
Aut(C0(X)) by lth(ϕ)(x) := ϕ(h−1 · x). Suppose that β : H → Aut A is a C ∗-
dynamical system such that
βh(ϕ · a) = lth(ϕ) · βh(a)
for h ∈ H, ϕ ∈ C0(X) and a ∈ A.
following [17, Example 4.8], we obtain a groupoid dynamical system
Then,
(A , G, α) where
Let ∆ be the modular function on H. Then the map Φ : Cc(H, A) → Γc(G; r∗A )
given by
α(h,x)(cid:0)a(h−1 · x)(cid:1) = βh(a)(x).
Φ(f )(h, x) = ∆(h)
1
2 f (h)(x)
extends to an isomorphism of A ⋊β H with A ⋊α G.
Fix a representation π of A. By decomposing π as a direct integral over X
one checks that (f ⊗π ξg ⊗π η) = (Φ(f ) ⊗π ξΦ(g) ⊗π η) for f, g ∈ Cc(G, A) and
ξ, η ∈ Hπ. We use this to show that U (f ⊗ h) = Φ(f ) ⊗ h determines a unitary from
the space of the regular representation IndA π of A ⋊β H to the space of the regular
representation IndA π of A ⋊α G which intertwines IndA π(f ) and IndA π(Φ(f ))
for all f . Therefore Φ factors through an isomorphism A ⋊β,r H ∼= A ⋊α,r G.
Now, back to the set-up of Raeburn's Symmetric Imprimitivity Theorem. Since
IndP
H (D, β) is a C0(P/H)-algebra, it is the section algebra of a bundle B over
P/H. It is shown in [17, Example 5.12] that there is a groupoid action σ of the
transformation groupoid K × P/H on B such that
Similarly,
IndP
H (D, β) ⋊σ K ∼= B ⋊σ (K × P/H).
IndP
K(D, α) ⋊τ H ∼= A ⋊τ (K\P ⋊ H)
for an appropriate bundle A over K\P and action τ . Furthermore, the trivial
bundle E := P ×A is an equivalence between (B, K ×P/H, σ) and (A , K\P ×H, τ )
in the sense of [17, Definition 5.1].
(Thus Raeburn's Symmetric Imprimitivity
Theorem is a special case of [17, Theorem 5.5].) Therefore the Quigg-Spielberg
result follows from following corollary of our main theorem.
Corollary 19. Suppose that q : E → Z is an equivalence between the groupoid
dynamical systems (B, H, β) and (A , G, α). Then the Morita equivalence of [17,
Theorem 5.5] factors through a Morita equivalence of the reduced crossed products
B ⋊β,r H and A ⋊α,r G.
Proof. Recall that r∗A := {(a, x) ∈ A × G : r(a) = r(x)} is a Fell bundle over
G with bundle map (a, x) 7→ x, multiplication (a, x)(b, y) = (aαx(b), xy) and in-
volution (a, x)∗ = (αx−1 (a), x−1) (see [16, Example 2.1]), and similarly for r ∗ B.
Define maps r∗A ∗ E → E and E ∗ r∗B → E by
(a, g) · a := a · αg(a)
and
a · (b, h) := βh(a) · b.
16
AIDAN SIMS AND DANA P. WILLIAMS
Define pairings
h· , ·i : E ∗s E → A and h· , ·i
: E ∗r E → B by
r∗B
r∗A
r∗A
ha , bi
ha , bi = (cid:0)ha , αG[q(a),q(b)](b)i
= (cid:0)ha , β[q(a),q(b)]H (b)i
r∗B
A
r(a)
A
r(a)
and
, G[q(a), q(b)](cid:1)
, [q(a), q(b)]H(cid:1).
It is routine though tedious to show that E is an r∗A -- r∗B equivalence.
The Morita equivalence X of [17, Theorem 5.5] and the Morita equivalence
pGC ∗(L, L(E ))pH of Theorem 14 are both completions of Γc(Z; E ). From the for-
mulae for the actions of r∗A on E , we see that the identity map on Γc(Z; E)
determines a left-module map from X to pGC ∗(L, L(E ))pH , and similarly on the
right. So it suffices to show that the norms on X and on pGC ∗(L, L(E ))pH coin-
cide. For this, observe that the formula [17, Equation (5.1)] for the A ×α G-valued
inner-product on Γc(Z; E ) is precisely the convolution formula for multiplication
of the corresponding elements of Γc(L; L(E )) with respect to the Haar system κ
described in [22].
(cid:3)
References
[1] A. Buss, R. Meyer, and C. Zhu, A higher category approach to twisted actions on C ∗-algebras,
preprint, 2009. (arXiv:math.OA.0908.0455v1).
[2] John Dauns and Karl H. Hofmann, Representation of rings by sections, Mem. Amer. Math.
Soc. 83 (1968), 1 -- 180.
[3] Maurice J. Dupr´e and Richard M. Gillette, Banach bundles, Banach modules and automor-
phisms of C ∗-algebras, Vol. 92, Pitman (Advanced Publishing Program), Boston, MA, 1983.
[4] James M. G. Fell and Robert S. Doran, Representations of ∗-algebras, locally compact groups,
and Banach ∗-algebraic bundles. Vol. 1, Pure and Applied Mathematics, vol. 125, Academic
Press Inc., Boston, MA, 1988. Basic representation theory of groups and algebras.
[5]
, Representations of ∗-algebras, locally compact groups, and Banach ∗-algebraic bun-
dles. Vol. 2, Pure and Applied Mathematics, vol. 126, Academic Press Inc., Boston, MA, 1988.
Banach ∗-algebraic bundles, induced representations, and the generalized Mackey analysis.
[6] Karl Heinrich Hofmann, Banach bundles, 1974. Darmstadt Notes.
[7]
, Bundles and sheaves are equivalent in the category of Banach spaces, in K-theory
and operator algebras (Proc. Conf., Univ. Georgia, Athens, Ga., 1975), Lecture Notes in
Math, vol. 575, Springer, Berlin, 1977, pp. 53 -- 69.
[8] Karl Heinrich Hofmann and Klaus Keimel, Sheaf-theoretical concepts in analysis: bundles
and sheaves of Banach spaces, Banach C(X)-modules, in Applications of sheaves (Proc. Res.
Sympos. Appl. Sheaf Theory to Logic, Algebra and Anal., Univ. Durham, Durham, 1977),
Lecture Notes in Math., vol. 753, Springer, Berlin, 1979, pp. 415 -- 441.
[9] Astrid an Huef and Iain Raeburn, Regularity of induced representations and a theorem of
Quigg and Spielberg, Math. Proc. Cambridge Philos. Soc. 133 (2002), 249 -- 259.
[10] Marius Ionescu and Dana P. Williams, Irreducible representations of groupoid C ∗-algebras,
Proc. Amer. Math. Soc. 137 (2009), 1323 -- 1332.
[11] Steven Kaliszewski, Paul S. Muhly, John Quigg, and Dana P. Williams, Coactions and Fell
bundles, New York J. Math. 16 (2010), 315 -- 359.
[12] Gennadi G. Kasparov, Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91
(1988), 147 -- 201.
[13] Alex Kumjian, Fell bundles over groupoids, Proc. Amer. Math. Soc. 126 (1998), 1115 -- 1125.
[14] El-Kaıoum M. Moutuou and Jean-Louis Tu, Equivalence of fell systems and their reduced
C ∗-algebras, 2011. (arXiv:math.OA.1101.1235v1).
[15] Paul S. Muhly, Jean N. Renault, and Dana P. Williams, Equivalence and isomorphism for
groupoid C ∗-algebras, J. Operator Theory 17 (1987), 3 -- 22.
[16] Paul S. Muhly and Dana P. Williams, Equivalence and disintegration theorems for Fell bun-
dles and their C ∗-algebras, Dissertationes Math. (Rozprawy Mat.) 456 (2008), 1 -- 57.
FELL NOTES
17
[17]
, Renault's equivalence theorem for groupoid crossed products, NYJM Monographs,
vol. 3, State University of New York University at Albany, Albany, NY, 2008. Available at
http://nyjm.albany.edu:8000/m/2008/3.htm.
[18] Michael V. Pimsner, A class of C ∗-algebras generalizing both Cuntz-Krieger algebras and
crossed products by Z, in Free probability theory (Waterloo, ON, 1995), Fields Inst. Commun.,
vol. 12, Amer. Math. Soc., Providence, RI, 1997, pp. 189 -- 212.
[19] John Quigg and Jack Spielberg, Regularity and hyporegularity in C ∗-dynamical systems,
Houston J. Math. 18 (1992), 139 -- 152.
[20] Iain Raeburn, Induced C ∗-algebras and a symmetric imprimitivity theorem, Math. Ann. 280
(1988), 369 -- 387.
[21] Iain Raeburn and Dana P. Williams, Morita equivalence and continuous-trace C ∗-algebras,
Mathematical Surveys and Monographs, vol. 60, American Mathematical Society, Providence,
RI, 1998.
[22] Aidan Sims and Dana P. Williams, Renault's equivalence theorem for reduced groupoid C ∗-
algebras, J. Operator Theory (2012), in press. (arXiv:math.OA.1002.3093).
[23] Dana P. Williams, Crossed products of C ∗-algebras, Mathematical Surveys and Monographs,
vol. 134, American Mathematical Society, Providence, RI, 2007.
School of Mathematics and Applied Statistics, University of Wollongong, NSW
2522, Australia
E-mail address: [email protected]
Department of Mathematics, Dartmouth College, Hanover, NH 03755-3551
E-mail address: [email protected]
|
1102.4413 | 1 | 1102 | 2011-02-22T05:33:06 | From graphs to free products | [
"math.OA",
"math.FA"
] | We investigate a construction which associates a finite von Neumann algebra $M(\Gamma,\mu)$ to a finite weighted graph $(\Gamma,\mu)$. Pleasantly, but not surprisingly, the von Neumann algebra associated to to a `flower with $n$ petals' is the group von Neumann algebra of the free group on $n$ generators. In general, the algebra $M(\Gamma,\mu)$ is a free product, with amalgamation over a finite-dimensional abelian subalgebra corresponding to the vertex set, of algebras associated to subgraphs `with one edge' (or actually a pair of dual edges). This also yields `natural' examples of (i) a Fock-type model of an operator with a free Poisson distribution; and (ii) $\C \oplus \C$-valued circular and semi-circular operators. | math.OA | math |
From graphs to free products
Madhushree Basu, Vijay Kodiyalam and V.S.Sunder
The Institute of Mathematical Sciences, Chennai, India
e-mail: [email protected], [email protected], [email protected]
November 21, 2018
Abstract
We investigate a construction which associates a finite von
Neumann algebra M (Γ, µ) to a finite weighted graph (Γ, µ).
Pleasantly, but not surprisingly, the von Neumann algebra as-
sociated to to a 'flower with n petals' is the group von Neu-
mann algebra of the free group on n generators. In general,
the algebra M (Γ, µ) is a free product, with amalgamation over
a finite-dimensional abelian subalgebra corresponding to the
vertex set, of algebras associated to subgraphs 'with one edge'
(or actually a pair of dual edges). This also yields 'natural'
examples of (i) a Fock-type model of an operator with a free
Poisson distribution; and (ii) C ⊕ C-valued circular and semi-
circular operators.
1 Preliminaries
There has been a serendipitous convergence of investigations being
carried out independently by us on the one hand, and by Guionnet,
Jones and Shlyakhtenko on the other - see [GJS1], [KS1], [KS2],
[GJS2]. As it has turned out, we have been providing independent
proofs, from slightly different viewpoints, of the same facts. Both
the papers [KS2] and [GJS2], establish that a certain von Neumann
algebra associated to a graph is a free product with amalgamation of
a family of von Neumann algebras corresponding to simpler graphs.
The amalgamated product involved subgraphs indexed by vertices
in [KS2], while the subgraphs are indexed by edges in [GJS2]. This
paper was motivated by trying to understand how the proof of our
result in [KS2] was also drastically simplfied by considering edges
rather than vertices. And, this third episode in our series seems to
have the following points in its favour:
1
• It does make certain cumulant computations and consequent
free independence assertions much more transparent.
• It brings to light a quite simple 'Fock-type model' of free Pois-
son variables.
• By allowing non-bipartite graphs, we get the aesthetically pleas-
ing fact mentioned in the abstract regarding the 'flower on n
petals'.
We investigate, in a little more detail, the construction in [KS2]
which associated a von Neumann probability space to a weighted
graph. We begin by recalling the set-up:
By a weighted graph we mean a tuple Γ = (V, E, µ), where:
• V is a (finite) set of vertices;
• E is a (finite) set of edges, equipped with 'source' and 'range'
maps s, r : E → V and '(orientation) reversal' invoution map
E ∋ e 7→ e ∈ E with (s(e), r(e)) = (r(e), s(e)); and
• µ : V → (0,∞) is a 'weight or spin function' so normalised that
Pu∈V µ2(v) = 1
We let Pn = Pn(Γ) denote the set of paths of length n in Γ
and let Pn(Γ) denote the vector space with basis {[ξ] : ξ ∈ Pn(Γ)}.
We think of ξ = ξ1ξ2 ··· ξn as the 'concatenation product' where
ξi denotes the i-th edge of ξ. We write F (Γ) = ⊕n≥0Pn(Γ) for
the indicated direct sum, and equip it with the following slightly
complicated multiplication: if ξ ∈ Pm(Γ), η ∈ Pn(Γ), then [ξ]#[η] =
Pmin(m,n)
ζk =
if ξm−j+1 = eηj∀1 ≤ j ≤ k
Here, and elsewhere, we adopt the convention that if ξ ∈ Pn,
then ξ = ξ1ξ2 ··· ξn denotes concatenation product, with ξi ∈ E and
we write s(ξi) = vξ
In particular, notice that P0(Γ) = {v : v ∈ V }, and that if v =
s(ξ), w = r(ξ) for some ξ ∈ Pn, and if u1, u2 ∈ V , then [u1][ξ][u2] =
δu1,vδu2,w[ξ]; and less trivially, if ξ ∈ P1 and η ∈ Pm, m ≥ 1, then
if r(ξ) 6= s(η)
if r(ξ) = s(η) but ξ 6= eη1
if ξ = eη1
[ζk], where ζk ∈ Pm+n−2k is defined by
µ(vξ
µ(vξ
0
[ξ]#[η] =
i−1 (so also r(ξi) = s(ξi+1) = vξ
i ).
m)
m−k )
[ξ1ξ2 ··· ξm−kηk+1ηk+2 ··· ηn]
0
[ξη1...ηm]
[ξη1...ηm] + µ(r(ξ))
µ(s(ξ)) [η2 ··· ηm]
2
k=0
otherwise
We define φ : F (Γ) → P0 by requiring that if ξ ∈ Pn, then
if n > 0
if n = 0
φ([ξ]) =(cid:26) 0
[ξ]
and finally define
τ = µ2 ◦ φ
where we simply write µ2 for the linear extension to P0(Γ) which
agrees with µ2 on the basis P0(Γ).
It was shown in [KS]1 that (F (Γ), τ ) is a tracial non-commutative
*-probability space, with e∗ = e, that the mapping y 7→ xy ex-
tends to a ∗-algebra representation F (Γ) → L(L2(F (Γ)), τ ) and that
M (Γ, µ) = λ(F Γ))′′ ⊂ L(L2(F (Γ)), τ ) is in standard form. Before
proceeding further, it is worth noting that for ξ, η ∈ ∪nPn(Γ), we
have
τ ([ξ]#[η]∗) = δξ,ηµ(r(ξ))µ(s(ξ)) ,
and hence, if we write {ξ} = (µ(s(ξ))µ(r(ξ)))− 1
∪n≥0Pn(Γ)} is an orthonormal basis for H(Γ) = L2(F (Γ), τ ).
2 [ξ], then {{ξ} : ξ ∈
2 The building blocks
Our interest here is the examination of just how M (Γ, µ) depends on
(Γ, µ). We begin by spelling out some simple examples, which will
turn out to be building blocks for the general case.
Example 2.1.
1. Suppose V = E = 1, say V = {v} and E =
{e}. Then we must have e = e, s(e) = r(e) = v, µ(v) = 1,Pn =
{en} and {ξ(n) = {en} : n ≥ 0} (where {e0} = {v}) is an
orthonormal basis for H(Γ); and the definitions show that x =
λ(e) satisfies xξn = ξ(n+1)+ξ(n−1). Thus x is a semi-circular
element and M (Γ) = {x}′′ ∼= LZ.
2. Suppose V = 1,E = 2, say V = {v} and E = {e1, e2} sup-
pose e2 = ee1. Then we must have s(ej) = r(ej) = v, µ(v) = 1.
Further {{e1},{e2}} is an orthonormal basis for H2 = P1(Γ),
and Pn(Γ) is isomorphic to ⊗nH2. Thus H(Γ) may be iden-
tified with the full Fock space F(H2) and the definitions show
that x1 = λ(e1) may be identifed as x1 = l1 + l∗2, where the lj
denote the standard creation operators. It follows that x1 is a
circular element and M (Γ) = {x1}′′ ∼= LF2.
1Actually, [KS] treated only the case of bipartite graphs, and sometimes re-
stricted attention to the case of the Perron-Frobenius weighting; but for the the
proof of statements made in this paragraph, none of those restrictions is necessary.
3
µ(v)
3. Suppose V = 2,E = 2, say V = {v, w} and E = {e,ee}
and suppose s(e) = v, r(e) = w and µ(w) ≤ µ(v). Write ρ =
µ(w) (≥ 1). If we let pv = λ([v]), pw = λ([w]), it follows that
Hv = ran pv (resp., Hw = ran pw) has an orthonormal basis
given by {{η(n)} : n ≥ 0} (resp., {{ξ(n)} : n ≥ 0} where η(n) ∈
Pn (resp., ξ(n) ∈ Pn) and η(n)k = e or e (resp., ξ(n)k = e or
e according as k is odd or even).
Writing x = λ(e), we see that with respect to the decomposition
H(Γ) = Hv ⊕ Hw, the operator x has a matrix decomposition
of the form
x =(cid:20) 0 t
0 0 (cid:21)
where t ∈ L(Hw,Hv) is seen to be given by
t[ξ(n)] = x[ξ(n)]
= [e]#[eeee··· (n terms)]
= [η(n + 1)] + ρ−1[η(n − 1)] ;
and hence,
t{ξ(n)} = (µ(s(ξ(n))µ(r(ξ(n)))− 1
2 t[ξ(n)]
= (µ(w)µ(r(ξ(n)))− 1
= (ρ−1µ(v)µ(r(η(n ± 1)))− 1
2 (cid:0)[η(n + 1)] + ρ−1[η(n − 1)](cid:1)
2{η(n − 1)}
2 (cid:0)[η(n + 1)] + ρ−1[η(n − 1)](cid:1)
= ρ
1
2{η(n + 1)} + ρ− 1
It is a fact - see Proposition 2.2 - that t∗t has has absolutely
continuous spectrum. This fact has two consequences:
(i) if t = ut is the polar decomposition of t, then u maps Hw
isometrically onto the subspace M = ran t of Hv, and if z is
the projection onto Hv ⊖ M then τ (z) = µ2(v) − µ2(w); and
(ii) W ∗(t) ∼= LZ.
Since pv + pw = 1 and z ≤ pv, the definitions are seen to
show that M (Γ, µ) is isomorphic to C⊕ M2(LZ) via the unique
isomorphism which maps pv, pw, z, u and t, respectively, to
(1,(cid:18) 1 0
0 0 (cid:19)), (0,(cid:18) 0 0
0 0 (cid:19)), and
0 a (cid:19)) for some positive a with absolutely continuous
(1,(cid:18) 0 0
0 0 (cid:19)), (0,(cid:18) 0 1
0 1 (cid:19)), (1,(cid:18) 0 0
spectrum which generates LZ as a von Neumann algebra. (This
4
must be compared with Lemma 17 of [GJS2], bearing in mind
that their µ is our µ2.)
Proposition 2.2. Let ℓ2(N) have its standard orthonormal basis
{δn : n ∈ N}. (For us, N = {0, 1, 2,··· }.) Let ℓδn = δn+1 denote
the creation operator (or unilateral shift), with ℓ∗δn = δn−1 (where
δ−1 = 0). Let ρ > 1 and t = ρ
2 ℓ∗. Then,
2 ℓ + ρ− 1
1
1. t∗t leaves the subspace ℓ2(2N) invariant;
2. δ0 is a cyclic vector for the restriction to ℓ2(2N) of t∗t, call it
aρ; and
3. the (scalar) spectral measure of aρ associated to δ0 is absolutely
continuous with respect to Lebesgue measure; in fact aρ has a
free Poisson distribution.
Proof. A little algebra shows that
t∗t = (ρ
1
1
2 ℓ∗ + ρ− 1
2 ℓ∗)
= ℓ2 + ℓ∗2 + (ρ + ρ−1) − ρ−1p0 ,
2 ℓ + ρ− 1
2 ℓ)(ρ
where p0 is the rank one projection onto Cδ0. It is seen that this
operator leaves both subspaces ℓ2(2N) and ℓ2(2N + 1) invariant, with
its restrictions to these subspaces being unitarily equivalent to ℓ +
ℓ∗ + (ρ + ρ−1)− ρ−1p0 and ℓ + ℓ∗ respectively. Since the spectral type
does not change under scalar translation, we may assume without
loss of generality that aρ = ℓ + ℓ∗ − ρ−1p0 and establish that a0 has
absolutely continuous scalar spectral measure corresponding to δ0.
Write a0 = ℓ + ℓ∗ so that aρ = a0 − ρ−1p0. Let the scalar spectral
measures of a0 and aρ be denoted by µ and µρ respectively, and
consider their Cauchy transforms given by
Fλ(z) = h(aλ − z)−1δ0, δ0i =ZR
dµλ(x)
x − z
for λ ∈ {0, ρ} and z ∈ C+ = {ζ ∈ C : Im(ζ) > 0}.
It follows from the resolvent equation that
Fρ(z) = h(aρ − z)−1δ0, δ0i
= h(a0 − z)−1δ0, δ0i + h(aρ − z)−1ρ−1p0(aλ − z)−1δ0, δ0i
= F0(z) + ρ−1Fρ(z)F0(z) ;
Hence
Fρ(z) =
F0(z)
1 − ρ−1F0(z)
=
ρF0(z)
ρ − F0(z)
(2.1)
5
It is seen from Lemma 2.21 of [NS] - after noting that the G of
that Lemma is the negative of the F0 here - that F0(z) = z−√z2−4
where √z2 − 4 is a branch of that square root such that √z2 − 4 =
√z + 2√z − 2 where the two individual factors are respectively de-
fined by using the branch-cuts {∓2− it : t ∈ (0,∞). (This choice en-
sures that limz→∞F0(z) = 0, which is clearly necessary.) It follows
that F0, which is holomorphic in C+, actually extends to a continu-
ous function on C+∪ R, and that if we write f0(a) = limb↓0 F0(a+ib),
then we have
2
2f0(t) =
−t + √t2 − 4
−t + i√4 − t2
−t − √t2 − 4
if t ≥ 2
if t ∈ [−2, 2]
if t ≤ −2
(2.2)
It is easy to check that f0 is strictly increasing in (−∞,−2), as
well as in in (2,∞), has non-zero imaginary part in (−2, 2), and
satisfies f (R\ (−2, 2)) = [−1, 0)∪ (0, 1]. Since ρ > 1, we may deduce
that F0(z) 6= ρ ∀z ∈ C+ ∪ R, and hence that also Fρ extends to a
continuous function on C+∪ R with equation (2.1) continuing to hold
for all z ∈ C+ ∪ R. Writing fλ(t) = Fλ(t + i0) for λ ∈ {0, ρ}, we find
that
fρ(t) =
and hence that
ρf0(t)
ρ − f0(t)
=
1
f0(t)−1 − ρ−1
,
Im(fρ(t)) = −
Im(f0(t)−1)
f0(t)−1 − ρ−12
Im(f0(t))
=
1 − f0(t)ρ−12
= ρ2 Im(f0(t))
f0(t) − ρ2
= 1[−2,2](t)
ρ2√4 − t2
2f0(t) − ρ2 .
Now, for t ∈ [−2, 2], we see that
f0(t) − ρ2 = −t + i√4 − t2
2
1
− ρ2
4(cid:0)(t + 2ρ)2 + 4 − t2(cid:1)
= ρ2 + ρt + 1 .
=
It follows from Stieltje's inversion formula that our aρ has ab-
solutely continuous scalar spectral measure µρ, with density given
6
by
gρ(t) =
1
π
Imfρ(t)
= 1[−2,2](t)
ρ2√4 − t2
2π(ρ2 + ρt + 1)
.
Hence the operator t∗t = aρ + (ρ + ρ−1)1 has has absolutely
continuous scalar spectral measure, with density given by
g(t) = gρ(t − (ρ + ρ−1))
= 1[(ρ+ρ−1)−2,(ρ+ρ−1)+2](t)
= 1[(ρ+ρ−1)−2,(ρ+ρ−1)+2](t)
ρ2p4 − (t − (ρ + ρ−1)
2πρ−2(ρ2 + ρ(t − ρ − ρ−1) + 1)
ρ2p4 − (t − (ρ + ρ−1)
2πρ−1t
2
2
If we write λ = ρ2 and α = ρ−1, we see that α(1 + λ) and recognise
the fact that not only does t∗t have absolutely continuous spectrum,
but - by comparing with equation (12.15) of [NS] -even that it ac-
tually has a free Poisson distribution, with rate ρ2 and jump size
ρ−1. However, we actually discovered this fact about t∗t having a
free Poisson distribution with the stated λ and α by a cute cumulant
computation which we present in the final section, both for giving
a combinatorial rather than analytic proof of this Proposition, and
because we came across that proof first.
(cid:3)
3 Some free cumulants
Before proceeding with the further study of a general (Γ, µ), we will
need an alternative description of M (Γ, τ ).
Let Gr(Γ) = ⊕n≥0Pn(Γ) be equipped with a ∗-algebra structure
wherein [ξ]◦ [η] = [ξη]) and [ξ]∗ = [ξ] = [eξn ···fξ1] for ξ ∈ Pn, η ∈ Pm.
It turns out - see [KS]2 - that Gr(Γ) and F (Γ) are isomorphic as
∗-algebras. While the multiplication is simpler in Gr(Γ), the trace τ
on F (Γ) turns out, when transported by the above isomorphism, to
be given by a slightly more complicated formula. (It is what has been
called the Voiculescu trace by Jones et al.) We shall write tr for this
transported trace on Gr(Γ), and φ for the tr-preserving conditional
expectation of M (Γ, µ)(= λ(Gr(Γ))′′) onto P0(Γ). We shall use the
2The remark made in an earlier footnote, concerning assumptions regarding
bipartiteness of Γ, applies here as well.
7
same letter φ to denote restrictions to subalgebras which contain
P0(Γ).
We wish to regard (Gr(Γ), φ) as an operator-valued non-commutative
probability space over P0(Γ), our first order of business being the de-
termination of the P0(Γ)-valued mixed cumulants in Gr(Γ).
Proposition 3.1. The P0(Γ)-valued mixed cumulants in Gr(Γ) are
given thus:
κn([e1], [e2],··· , [en]) = 0 unless n = 2 and e2 = ee1; and if e2 =
ee1 with s(e1) = v, r(e1) = w, then κ2([e1], [ee1]) = µ(w)
Proof. The proof depends on the 'moment-cumulant' relations which
guarantee that in order to prove this proposition, it will suffice to
establish the following, which is what we shall do:
µ(v) [v].
(a) Define κn : (Gr(Γ))n → P0(Γ) to be the unique multilin-
ear map which is defined when the arguments are tuples of paths as
asserted in the proposition; note that (i) it is 'balanced' over P0(Γ) in
the sense that κn(x1,··· , xi−1b, xi,··· , xn) = κn(x1,··· , xi−1, bxi,··· , xn)
for all xj ∈ Gr(Γ), b ∈ P0(Γ) and 1 < i ≤ n, and (ii) is P0(Γ)-bilinear
meaning κn(bx1, x2,··· , xn−1, xnb′) = bκn(x1, x2,··· , xn−1, xn)b′ for
all xj ∈ Gr(Γ), b, b′ ∈ P0(Γ);
(b) define the 'multiplicative extensions' κπ : (Gr(Γ))n → P0(Γ)
for π ∈ N C(n) by requiring, inductively, that if [k, l] is an interval
constituting a class of π, and if we write σ for the element of N C(n−
l + k − 1) given by the restriction of π to {1,··· , k − 1, l + 1,··· , n},
so that 'π = σW 1[k,l]' then
κπ(x1,··· , xn) = κσ(x1,··· , xk−1κl−k+1(xk,··· , xl), xl+1,··· , xn)
= κσ(x1,··· , xk−1, κl−k+1(xk,··· , xl)xl+1,··· , xn);
(c) and verify that for any e1,··· , en ∈ P1(Γ),
φ([e1]··· [en]) = Xπ∈N C(n)
κπ([e1], [e2],··· , [en]).
(3.3)
For this verification, we first assert that if e1, e2,··· , en ∈ E
and π ∈ N C(n), the quantity κπ([e1], [e2],··· , [en]) (yielded by the
unique 'multiplicative extension' of the κn's as in (b) above) can be
non-zero only if
(i) e1e2 ··· en is a meaningfully defined loop based at s(e1), mean-
ing f (ei) = s(ei+1) for 1 ≤ i ≤ n, with en+1 being interpreted as e1;
(ii) π ∈ N C2(n) is a pair partition of n (and in particular n is
even), such that {i, j} ∈ π ⇒ ej = eei;
8
and if that is the case, then,
κπ([e1], [e2],··· , [en]) =
Y{i,j}∈π
i<j
µ(r(ei)
µ(r(ej)
[s(e1)] .
(3.4)
We prove this assertion by induction on n. This is trivial for
n = 1 since κ1 ≡ 0. By the inductive definition of the multiplicative
extension, it is clear that if κπ([e1], [e2],··· , [en]) is to be non-zero,
π must contain an interval class of the form {k, k + 1} such that
ek+1 = eek; if σ denotes π{1,2,··· ,k−1,k+2,···n} we must have
κπ([e1],··· , [en]) =
κσ([e1],··· , [ek−1][s(ek)], [ek+2],··· , [en])
µ(r(ek))
µ(r(ek+1))
µ(r(ek))
µ(r(ek+1))
µ(r(ek))
µ(r(ek+1))
=
=
κσ([e1],··· , [ek−1], [s(ek)][ek+2],··· , [en])
κσ([e1],··· , [ek−1][r(ek+1)], [ek+2],··· , [en]) ;
and for this to be non-zero, we must have r(ek−1) = s(ek) = r(ek+1) =
s(ek+2), in which case we would have
µ(r(ek))
µ(r(ek+1))
κσ([e1],··· , [ek−1], [ek+2],··· , [en]) ,
κπ([e1],··· , [en]) =
and the requirement that κσ([e1],··· , [ek−1], [ek+2],··· , [en]) be non-
zero, along with the induction hypothesis, finally completes the proof
of the assertion.
Now, in order to verify equation 3.3, it suffices to check that for
any v ∈ V , we have
tr([e1][e2]··· [en][v]) = Xπ∈N C(n)
tr(κπ([e1], [e2],··· , [en])[v]).
(3.5)
First observe that both sides of equation 3.5 vanish unless e1 ··· en
is a meaningfully defined path with both source and range equal to v
(since tr is a trace and [v] is idempotent). In view of our description
above of the multiplicative extension κπ, we need, thus, to verify that
for such a loop, we have
tr([e1 ··· en]) = Xπ∈N C2(n)
δej , eei
µ(r(ei)
µ(r(ej)
µ2(s(e1)),
Y{i,j}∈π
i<j
9
but that is indeed the case (see equation (3) and the proof of Propo-
sition 5 in [KS1] ).
(cid:3)
In order to derive the true import of Proposition 3.1, we should
first introduce some notation:
For each dual pair e, e of edges - with, say, s(e) = v, r(e) = w - we
shall write Γe = (Ve, Ee, µe) where Ve = V, µe = µ and Ee = {e, e}
(with source, range and reversal in Ee as in E). If e = e, the above
definitions are to be suitably interpreted. Now for 'the true import
of Proposition 3.1':
Corollary 3.2. With the foregoing notation, we have:
Gr(Γ, µ) = ∗P0(Γ){Gr(Γe, µe) : {e, e} ⊂ E}
and hence, also
M (Γ, µ) = ∗P0(Γ){M (Γe, µe) : {e, e} ⊂ E} .
φ
→ B is a 'non-
Proof. Proposition 3.3.3 of [S1] shows that if A
commutative probability space over B', if {Ai : i ∈ I} is a family
: i ∈ I} gener-
of subalgebras of A containing B, such that {Ai
ates A, and if Gi is a set of generators of the algebra Ai, then A is
the free product with amalgamation over B of {Ai : i ∈ I} if and
only if the mixed B-valued cumulants κn(x1,··· , xn) vanish when-
ever x1,··· ,··· xn ∈ ∪iGi, unless all the xi belong to the same Gk
for some k. The desired assertion then follows from Proposition 3.1.
(cid:3)
The following assertion, advertised in the abstract, is an imme-
diate consequence of Corollary 3.2 and Examples 2.1 (1) and (2).
Corollary 3.3. If Γn denotes the 'flower with n petals' (thus V =
1,E = n), then M (Γ) ∼= LFn, independent of the reversal map on
E.
Remark 3.4. We may deduce from Proposition 3.1 that the x = λ(e)
of Example 2.1 (3) is a P0(Γ)-valued circular operator, in the sense
of [Dyk] (see Definition 4.1), with covariance (α, β) where α(b) =
φ(x∗bx) and β(b) = φ(xbx∗) for all b ∈ P0 are the completely positive
self-maps of P0(Γ)(= Cpv ⊕ Cpw) induced by the matrices
α =(cid:20) 0 ρ−1
0
0 (cid:21) and β =(cid:20) 0 0
ρ 0 (cid:21) .
10
If s = x + x∗, it follows then that s is a P0(Γ)-valued semi-
circular element (since κn(sb1, sb2,··· sbn−1, s) = 0 unless n = 2 and
κ2(sb, s) = η(b) where η is the (completely) positive self-map of C⊕ C
induced by the matrix
η =(cid:20) 0 ρ−1
0 (cid:21) .
ρ
4 Narayana numbers
Recall the Narayana numbers N (n, k) defined for all n, k ∈ N with
1 ≤ k ≤ n by
N (n, k) = {π ∈ N C(n) : π = k}.
Define the associated polynomials Nn by
Nn(T ) =
nXk=1
N (n, k)T k.
Recall also that a random variable in a non-commutative proba-
bility space (A, τ ) is said to be free Poisson with rate λ and jump size
α if its free cumulants are given by κn = λαn for all n ∈ N. An easy
application of the moment-cumulant relations shows that an equiv-
alent condition for a random variable to be free Poisson with rate λ
and jump size α is that its moments are given by µn = αnNn(λ) for
all n ∈ N.
We now illustrate an application of this characterisation of a free
Poisson variable in the situation of §2, Example 2.1 (3). There,
x = λ(e) has a matrix decomposition involving t ∈ L(Hw,Hv) where
t∗t was shown to have a free Poisson distribution. We will ver-
ify below by a cumulant computation that t∗t is free Poisson with
rate ρ2 and jump size ρ−1 in the non-commutative probability space
pwM (Γ, µ)pw.
Begin by observing that x∗x has a non-zero entry only in the w-
corner and that this entry is t∗t. Thus the trace in M (Γ, µ) of (x∗x)n
is µ2(w) times the trace - call it trw - in pwM (Γ, µ)pw of (t∗t)n. We
now compute tr((x∗x)n) = tr(([e]∗[e])n).
First apply the moment-cumulant relations and Proposition 3.1
to conclude that
φ(([e]∗[e])n) = Xπ∈N C(2n)
κπ([e]∗, [e],··· , [e]∗, [e]).
11
While this sum ranges over all π ∈ N C(2n), Proposition 3.1 enables
us to conclude that unless π is a non-crossing pair partition, its
contribution vanishes. Thus we have:
φ(([e]∗[e])n) = Xπ∈N C2(2n)
κπ([e]∗, [e],··· , [e]∗, [e]).
Now we use the well-known bijection between non-crossing pair
partitions (or equivalently, Temperley-Lieb diagrams) on 2n points
and all non-crossing partitions on n points. We will denote this bi-
jection as π ∈ N C2(2n) ↔ π ∈ N C(n). This is illustrated by exam-
ple in Figure 4 for π = {{1, 8},{2, 5},{3, 4},{6, 7}, {9, 12}, {10, 11}}
and may be summarised by saying that the black regions of the
Temperley-Lieb diagram for π ∈ N C2(2n) correspond to the classes
of π ∈ N C(n). Note that in Figure 4 the numbers above refer to the
1
2
1
3
4
2
5
6
3
7
4
8
9
11
10
5
12
6
{{1,3,4},{2},{5,6}}
Figure 1: π ∈ N C2(12) ↔ π ∈ N C(6)
vertices while those below refer to the black segments.
It follows from Proposition 3.1 that for any π ∈ N C2(2n), the
term κπ([e]∗, [e],··· , [e]∗, [e]) is a scalar multiple of pw where the
scalar is given by a product of n terms each of which is ρ = µ(v)
µ(w) or
ρ−1 = µ(w)
µ(v) . Classes of π for which the smaller element is odd give
ρ, while those for which the smaller element is even give ρ−1. Thus
κπ([e]∗, [e],··· , [e]∗, [e]) evaluates to ρ(πodd−πeven)pw = ρ(2πodd−n)pw,
where, of course, πodd (resp. πeven) denotes the number of classes
of π whose smaller element is odd (resp. even).
Our main combinatorial observation is contained in the following
simple lemma.
Lemma 4.1. For any π ∈ N C2(2n), πodd = π.
Proof. We induce on n with the basis case n = 1 having only one
π with πodd = π = 1. For larger n, consider a class of π of the
form {i, i + 1}, and remove it to get ρ ∈ N C2(2n − 2). A moment's
thought shows that if i is odd then πodd = ρodd + 1 = ρ + 1 = π,
while if i is even then πodd = ρodd = ρ = π.
(cid:3)
12
Thus:
ρ(2πodd−n)pw
ρ(2π−n)pw
φ(([e]∗[e])n) = Xπ∈N C2(2n)
= Xπ∈N C(n)
nXk=1 X{π∈N C(n):π=k}
nXk=1
N (n, k)ρ2k−npw
=
=
ρ2k−npw
Hence tr(([e]∗[e])n) =Pn
Pn
k=1 N (n, k)ρ2k−n. Now the characterisation of free Poisson ele-
ments in terms of their moments shows that t∗t is free Poisson with
rate ρ2 and jump size ρ−1.
k=1 N (n, k)ρ2k−nµ2(w) and thus trw((t∗t)n) =
Remark 4.2.
1
2 ℓ + ρ− 1
1. Thus, for t = ρ
2 ℓ∗, we have shown that
t∗t is a free Poisson element with rate ρ2 and jump size ρ−1.
By scaling by an appropriate constant, we can similarly obtain
such simple Fock-type models of free Poisson elements with
arbitrary jump size and rate.
2. Similar scaling, and the fact that eiθℓ is unitarily equivalent to
ℓ (by a unitary operator which fixes δ0) show that, in fact, if
t = aℓ + bℓ∗ for any a, b ∈ C, then t∗t is a free Poisson element.
Acknowledgement: We would like to thank M. Krishna for pa-
tiently leading us through the computation of Cauchy transforms of
rank-one perturbations as we struggled with an apparent contradic-
tion, which was finally resolved when we realised a problematic minus
sign stemming from a small mistake in choice of square roots. (We
claim no originality for this problem, for the same incorrect sign also
surfaces on page 33 of [NS] - cf. our formula (2.2) and the formula
there for g , when −∞ < t < −2.)
References
[1] [Dyk] Ken Dykema, Hyperinvariant subspaces for some B-
circular operators, With an appendix by Gabriel Tucci. Math.
Ann. 333 (2005), no. 3, 485-523.
13
[GJS1] A. Guionnet, V. F. R. Jones and D. Shlayakhtenko, Ran-
dom matrices, free probability, planar algebras and subfactors,
arXiv:0712.2904v2.
[GJS2] A. Guionnet, V.F.R. Jones, D. Shlyakhtenko, A semi-
finite algebra associated to a planar algebra, e-print arXiv
(math.OA)0911.4728.
[KS1] Vijay Kodiyalam and V.S. Sunder, Guionnet-Jones-
Shlyakhtenko subfactors associated to finite-dimensional Ka¸c
algebras, e-print arXiv (math. OA) 0901 3180.
[KS2] Vijay Kodiyalam and V.S. Sunder, On the Guionnet-Jones-
Shlyakhtenko construction for graphs, e-print arXiv (math. OA)
0911.2047
[NS] A. Nica and R. Speicher, Lectures on the Combinatorics of Free
Probability, LMS Lecture Note Series, vol. 335, Cambridge Uni-
versities Press, Cambridge, 2006.
[S1] R. Speicher, Combinatorial theory of the free product with amal-
gamation and operator-valued free probability theory, Memoirs of
the AMS 132 (1998).
[S2] R. Speicher, A New Example of 'Independence' and 'White
Noise', Probab. Th. Rel. Fields, 84, 141-159 (1990)
14
|
1203.4530 | 1 | 1203 | 2012-03-20T18:06:27 | De Finetti theorem on the CAR algebra | [
"math.OA",
"math-ph",
"math-ph"
] | The symmetric states on a quasi local C*-algebra on the infinite set of indices J are those invariant under the action of the group of the permutations moving only a finite, but arbitrary, number of elements of J. The celebrated De Finetti Theorem describes the structure of the symmetric states (i.e. exchangeable probability measures) in classical probability. In the present paper we extend De Finetti Theorem to the case of the CAR algebra, that is for physical systems describing Fermions. Namely, after showing that a symmetric state is automatically even under the natural action of the parity automorphism, we prove that the compact convex set of such states is a Choquet simplex, whose extremal (i.e. ergodic w.r.t. the action of the group of permutations previously described) are precisely the product states in the sense of Araki-Moriya. In order to do that, we also prove some ergodic properties naturally enjoyed by the symmetric states which have a self--containing interest. | math.OA | math |
DE FINETTI THEOREM ON THE CAR ALGEBRA
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
Abstract. The symmetric states on a quasi local C ∗ -- algebra on
the infinite set of indices J are those invariant under the action of
the group of the permutations moving only a finite, but arbitrary,
number of elements of J. The celebrated De Finetti Theorem de-
scribes the structure of the symmetric states (i.e. exchangeable
probability measures) in classical probability. In the present paper
we extend De Finetti Theorem to the case of the CAR algebra, that
is for physical systems describing Fermions. Namely, after show-
ing that a symmetric state is automatically even under the natural
action of the parity automorphism, we prove that the compact
convex set of such states is a Choquet simplex, whose extremal
(i.e. ergodic w.r.t. the action of the group of permutations pre-
viously described) are precisely the product states in the sense of
Araki -- Moriya.
In order to do that, we also prove some ergodic
properties naturally enjoyed by the symmetric states which have a
self -- containing interest.
Mathematics Subject Classification: 46L53, 46L05, 60G09,
46L30, 46N50.
Key words: Non commutative probability and statistics; C ∗ --
algebras, states; Exchangeability; Applications to quantum physics.
1. introduction
Exchangeable or, equivalently, symmetrically dependent sequences
of random variables and symmetric states have been investigated in a
wide way both in Probability Theory and Operator Algebras.
After De Finetti's pioneering work [13] for 2 -- point valued random
variables, it has been shown that more and more general sequences of
exchangeable random variables are mixtures of independent identical
distributed (i.i.d. for short) sequences. One of its most general version
in classical probability was obtained by Hewitt and Savage in [18] for
exchangeable random variables distributed on X = E × E × . . ., E
being a compact Hausdorff space.
A noncommutative extension of this result for infinite tensor product
A of a single C ∗-algebra B, was given by Størmer in [31], where it
Date: July 20, 2018.
1
2
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
is shown that the symmetric states on A form a simplex whose ∗ --
weakly closed set of extremal points is made exactly by the product
states. Indeed Størmer's theorem, when reduced to abelian C ∗-algebras
is nothing else than Hewitt -- Savage result.
Other similar characterizations can be found in [15, 19] for Boson
quantum systems, or in [3, 17] in the case of continuous index set.
More recently in [21, 22, 23], the authors got some De Finetti's type
theorems in Free Probability, whereas in [1] some results on the struc-
ture of symmetric (exchangeable) states on general C ∗-algebras have
been obtained.
Besides infinite sequences case, Diaconis and Freedman in [14] ob-
tained a De Finetti theorem for finite sequences of exchangeable ran-
dom variables. Namely they showed that the first k random variables
of a permutation invariant distribution of n random variables can be
approximated by a convex combination of k i.i.d. random variables,
the error being of order O(k2/n). This has been the starting point for
a recent intensive investigation of finite De Finetti results in Quantum
Information Theory and the problem of the Entanglement. Unfortu-
nately, when one generalizes to the tensor product case such a result,
it comes out that the approximating error, contrarily to the classical
situation, depends on the dimension of the state space (see, e.g.
[11]
for details). Hence, even if a general extension to infinite dimensional
quantum systems does not exist, some precise estimates, independent of
the dimension, have been evaluated for many concrete physical classes.
In particular, the dependence on the local dimension is removed, for
example, when there is a bound on the number of the ways in which
the system is measured, or the n -- particle reduced density matrix is
separable [12], when one treats an exponential version of the theorem
in the case of coherent states [20], or when one takes orthogonal invari-
ant states [24]. Moreover, an analogous result can be found in the so
called Quantum Key Distribution (QKD), when one deals with Gauss-
ian states against general attacks [28], or when one modifies the sym-
metric hypothesis in a fully compatible way with continuous -- variables
QKD protocols [25].
Although none of the above mentioned results concern Fermions,
nowadays there is a rapidly increasing investigation of properties and
models dealing with physical systems describing such particles. We
mention for example the following issues which is far to be complete.
The investigation of the ground states of lattice systems based on an-
ticommutation annihilators [26]. The introduction and the study of
the notion of the product state, and the application to general ther-
[4, 5] and
modynamical properties of Fermi lattice systems (see e.g.
DE FINETTI THEOREM
3
the references cited therein). The connection with the Markov Prop-
erty, Quantum Statistical Systems, Quantum Information Theory and
Entanglement, of chains of Fermi systems [2, 16]. Disordered systems
based on Fermions [6].
Thus, for the natural applications to Quantum Physics and also the
general implication in Quantum Probability, it is then natural to ad-
dress the study of the structure of the symmetric states on the Fermi
algebra, that is the CAR C ∗-algebra. Up to the knowledge of the
authors, no result concerning the systematic study of this subject is
present in literature. This is the object of the present paper. In more
details, our purpose consists in:
(1) characterizing the extremal points of the convex set of symmet-
ric states on the CAR algebra,
(2) showing that every symmetric state is the barycenter of a unique
maximal Radon measure which is pseudosupported on the ex-
tremal states,
(3) proving that the extremal states form a ∗ -- weakly closed sub-
set (thus the symmetric states are isomorphic to the regular
probability measures on a compact Hausdorff space [7]),
(4) determining the type of the von Neumann factors generated by
the extremal (i.e. product) states.
One of the main tools in De Finetti theorem for infinite tensor product
of C ∗ -- algebras [31] is the asymptotic Abelianess property with respect
to the permutation group.
In the CAR algebra this property is not
satisfied, because of the anticommutation relations between spatially
separated operators. As a consequence, the results relative to the struc-
ture of symmetric states in [31] can not be directly imported in our case.
Hence, after verifying that the group of permutations which fix all but
a finite number of the points in an arbitrary set J (denoted by PJ),
acts as a group of automorphisms on CAR(J) (the CAR algebra on
J), we establish a result which plays a crucial role in the sequel. Each
symmetric (i.e. invariant under the action of the group of permutations
above) state on CAR(J) is even. This property is exploited throughout
the paper in order to obtain the results listed above. Indeed, it is used
in Theorems 5.3 and 6.1 to prove the equivalence between an extremal
symmetric state on CAR(J) and a product state in the Araki -- Moriya
sense (the product being constructed starting by a single even state on
M2(C)), thus reaching point (1).
Further, the even property makes sure that the couple (CAR(J), PJ)
is PJ -- Abelian (see Theorem 4.2). This allows to prove that the ∗ --
weakly compact convex set of symmetric states is a Choquet simplex,
4
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
and, consequently, to obtain the ergodic decomposition presented in
(2). The unique decomposing measure is supported on the extremal
states when J is countable, and pseudosupported when J is uncount-
able (see Theorem 5.5 and the final discussion in Section 6).
As a consequence of the above discussion, one has that each sym-
metric state is (at least in the countable case) a mixture of product
states of Araki -- Moriya. This statement, which is our main result, can
be seen as the extension of De Finetti's theorem to the CAR algebra.
The ending parts of Sections 5 and 6 are aimed to reach points (3)
and (4).
In particular, some results due to Størmer in [31] and the
identification of every even state on M2(C) with a single point of a
closed segment (Lemma 2.2), give the extremal (i.e. product) states
are a ∗ -- weakly closed subset. Furthermore, the very special structure
of a product state, allows to achieve point (4), as Propositions 5.7, 5.8
and the relative following discussions.
As stressed above, the Fermi algebra is not asymptotically Abelian
with respect to PJ , but the even nature of a symmetric state yields a
"weak" asymptotic Abelianess property, see (ii) of Theorem 4.1. This
result, coupled with the property that PJ acts as a large group of
automorphisms (cf. Definition 3.4 and Theorem 4.2), provides some
ergodic properties (cf. Proposition 4.3 and Proposition 5.4) for averages
and orbits of symmetric states, which, even not used for establishing the
main results presented in paper, may have a self -- containing interest.
2. the CAR algebra
Denote by [a, b] := ab − ba, {a, b} := ab + ba, the commutator and
anticommutator between elements a, b, respectively.
We start by quickly reviewing the basic properties of the Fermion C ∗-
algebra, which, due to Pauli Exclusion Principle, is generated by the
annihilation and creation operators satisfying the Canonical Anticom-
mutation Relations. Indeed, let J be an arbitrary set. The Canonical
Anticommutation Relations (CAR for short) algebra over J is the C ∗ --
algebra CAR(J) with the identity 1I generated by the set {aj, a†
j j ∈ J}
(i.e. the Fermi annihilators and creators respectively), and the relations
(aj)∗ = a†
j , {a†
j, ak} = δjk1I , {aj, ak} = {a†
j, a†
k} = 0 , j, k ∈ J .
The parity automorphism Θ acts on the generators as
Θ(aj) = −aj , Θ(a†
j) = −a†
j ,
j ∈ J
DE FINETTI THEOREM
5
and induces on CAR(J) a Z2 -- grading. This grading yields CAR(J) =
CAR(J)+ ⊕ CAR(J)−, where
CAR(J)+ := {a ∈ CAR(J) Θ(a) = a} ,
CAR(J)− := {a ∈ CAR(J) Θ(a) = −a} .
Elements in CAR(J)+ and in CAR(J)− are called even and odd, re-
spectively.
Notice that, by definition,
CAR(J) = CAR 0(J) ,
where
is the (dense) subalgebra of the localized elements.
CAR 0(J) :=[{CAR(I) I ⊂ J finite }
A map T : A1 → A2 between C ∗ -- algebras with Z2 -- gradings Θ1 and
Θ2, is said to be even if it is grading -- equivariant:
T ◦ Θ1 = Θ2 ◦ T .
The previous definition, applied to states ϕ ∈ S(CAR(J)), leads to
ϕ ◦ Θ = ϕ, that is ϕ is even if and only if it is Θ -- invariant.
When the index set J is countable, the CAR algebra CAR(J) is
isomorphic to the C ∗ -- infinite tensor product of J -- copies of M2(C),
C∗
M2(C)
.
CAR(J) ∼OJ
(2.1)
(2.2)
Such an isomorphism is established by a Jordan -- Klein -- Wigner trans-
formation, as shown in [36], Exercise XIV. We briefly report it for the
convenience of the reader. Fix any enumeration j = 1, 2, . . . of the set
J. Let Uj := aja†
n=1 Un,
and denote
jaj, j = 1, 2, . . . . Put V0 := 1I, Vj :=Qj
j − a†
e11(j) := aja†
j ,
e21(j) := Vj−1a†
j ,
e12(j) := Vj−1aj ,
e22(j) := a†
jaj .
{ekl(j) k, l = 1, 2}j∈J provides a system of commuting matrix units in
CAR(N). In order to obtain
C∗
M2(C)
,
CAR(N) ∼ON
fix any segment [1, l] ⊂ N and consider the system of matrix units
localized in r ∈ N
{εirjr(r)}irjr=1,2 ⊂ON
C∗
M2(C)
,
6
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
together with the system of matrix units
{eirjr (r)}irjr=1,2 ⊂ CAR(N)
arising from the Jordan -- Klein -- Wigner construction (2.2). The above
isomorphism is simply described by
(2.3)
ei1j1(1) · · · eiljl(l) 7→ εi1j1(1) ⊗ · · · ⊗ εiljl(l)
for each ik, jk = 1, 2, k = 1, 2, . . . , l and l ∈ N.
Remark 2.1. Notice that, fixed r ∈ N, eirjr(r) is localized in the seg-
ment [1, r] and, moreover, eirir (r) is always localized in the site r. But
eirjr (r) is not necessarily localized in the site r if ir 6= jr.
Anyone of such isomorphisms depends on a predefined order of the
countable index set J. Thus, it cannot be directly used to investigate
the exchangeable properties of the states under consideration.
Thanks to (2.1), CAR(J) has a unique tracial state τ as the extension
of the unique tracial state on CAR(I), I < +∞. Let I ⊂ J be a
finite set and ϕ ∈ S(CAR(J)). Then there exists a unique positive
element T ∈ CAR(I) such that ϕ⌈CAR(I)= τ ⌈CAR(I)(T · ). The element
T is called the adjusted density matrix of ϕ⌈CAR(I). For the standard
applications to quantum statistical mechanics, one also uses the density
matrix w.r.t. the unnormalized trace.
We recall here the description of product state (cf.
[5]). We start
with the case of finite sets. Namely, let I1, I2 ⊂ J with I1, I2 < ∞
one among them is even, then according to Theorem 1 of [5], the prod-
and I1T I2 = ∅. Fix ϕ1 ∈ S(CAR(I1)), ϕ2 ∈ S(CAR(I2)). If at least
uct state extension (called product state for short) ϕ ∈ S(CAR(I1S I2))
is uniquely defined. We write, with an abuse of notation, ϕ = ϕ1ϕ2.
Let T1 ∈ CAR(I1), T2 ∈ CAR(I2) be the adjusted densities relative
to ϕ1 ∈ S(CAR(I1)), ϕ2 ∈ S(CAR(I2)), respectively. As at least one
among T1 and T2 is even, [T1, T2] = 0 and T := T1T2 is a well defined
positive element of CAR(I1S I2) which is precisely the density matrix
of ϕ = ϕ1ϕ2. The product state ϕ ∈ S(CAR(I1 ∪ I2)) is even if and
only if ϕ1 and ϕ2 are both even.
Now we pass to the description of the product state on CAR(J)
symbolically written as
ρ ,
ϕ :=Yj∈J
where ρ is a single even state on M2(C) ∼ CAR({j}). For j ∈ J denote
ιj : M2(C) → CAR(J) the corresponding embedding. For each finite
subset I := {j1, . . . , jI} ⊂ J, let ϕI ∈ S(CAR(I)) be the product state
DE FINETTI THEOREM
7
given, on the elementary generators, by
I
ϕI(ιj1(A1) · · · ιjI(AI)) =
ρ(Ak) ,
Yk=1
where A1, . . . , AI ∈ M2(C).
If I1 ⊂ I2, it is immediate to see that
ϕI2⌈CAR(I1)= ϕI1. So the direct limit lim−→ϕI, when I ↑ J is a well defined
state on the dense ∗ -- algebra of the localized elements CAR0(J), which
extends by continuity to a state ϕ which is the product state of a single
ρ is
even state ρ ∈ M2(C). A necessarily even product state ϕ =YJ
then uniquely determined by the even state ρ, and the next lemma
shows that each even state on M2(C) can be seen as a single point of
a closed segment.
Lemma 2.2. For every even state ρ on M2(C), there exists a unique
µ ∈ [0, 1] such that
ρ(cid:18) a b
c d (cid:19) = µa + (1 − µ) d
Proof. By the usual identification of CAR({j}), j ∈ J with M2(C),
one sees that for every state ρ on M2(C), there exists a unique positive
with µ(1 − µ) − (b2 + f 2) ≥ 0, hence µ ∈ [0, 1].
matrix T such that ρ = τ (T · ). In particular T =(cid:18) µ
0 = ρ(cid:18) 0 1
0 0 (cid:19) = b − ıf , hence b = f = 0. This ends the proof.
b − ıf 1 − µ (cid:19),
If ρ is even, then
b + ıf
(cid:3)
We refer to the above µ and 1 − µ as the eigenvalues of the even
state ρµ on M2(C), the latter inherited by the Z2 -- grading arising from
M2(C) ∼ CAR({j}).
Let ρµ be an even state as before, and denote ϕµ, ωµ the correspond-
ing product states on A := CAR(N) and B :=ON
tively. Denote γ : A → B the isomorphism described via (2.3).
M2(C)
, respec-
C∗
Lemma 2.3. Under the above notations, for each µ ∈ [0, 1] we get
ϕµ = ωµ ◦ γ.
Proof. By the definition of the product states on A and B, it is enough
to check the result for the system of the matrix units. Let {ekl(j)k, l =
1, 2}j∈N, {εkl(j)k, l = 1, 2}j∈N be the canonical systems of matrix units
described above. By taking into account
M2(C)
C∗
of CAR(N) and NN
8
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
that ϕµ is even and Remark 2.1, we get for the restrictions to any
segment [1, l],
ϕµ(ei1j1(1) · · · eiljl(l)) = ϕµ(ei1i1(1) · · · eilil(l))δi1j1 · · · δiljl
=ρµ(ei1i1(1)) · · · ρµ(eilil(l))
=ωµ(εi1i1(1) ⊗ · · · ⊗ εilil(l))δi1j1 · · · δiljl
=ωµ(εi1j1(1) ⊗ · · · ⊗ εiljl(l)) .
(cid:3)
3. the group of the permutations and its action on the
CAR algebra
We firstly present a result, probably known to the experts, crucial
in the sequel. Let
Gα ,
G = [α∈A
where A is a directed set, and Gα, α ∈ A are finite subgroups of the
group G such that α < β implies Gα ⊂ Gβ. Consider a unitary repre-
sentation {U(g) g ∈ G} of G acting on a Hilbert space H. Denote
Eα, E the selfadjoint projections onto the subspaces of H consisting
of the invariant vectors under the action of Gα and G, respectively.
Of course, the net {Eα α ∈ A} is decreasing. It is straightforward
to see (cf.
[33], Section 2.17) that it converges in the strong operator
topology to a projection
In general, P ≥ E. In addition, it is a standard fact to verify that
P := s − lim
α
Eα .
Eα =
1
Gα Xg∈Gα
U(g) .
Here, we give the analogue of the von Neumann Ergodic Theorem for
the case under consideration.
Proposition 3.1. Under the above notations, we get
s − lim
α
Eα = E .
Proof. We have only to prove that P ≤ E. Fix h ∈ G and ξ ∈ H such
that
ξ = lim
α
1
Gα Xg∈Gα
U(g)ξ .
DE FINETTI THEOREM
9
As A is a directed set and the sequence of the groups {Gα α ∈ A}
is increasing, there exists αh ∈ A such that α > αh implies h ∈ Gα.
Then, after a standard change of variables in the sum, we obtain
U(h)ξ =U(h) lim
α
U(g)ξ = U(h) lim
α>αh
1
Gα Xg∈Gα
= lim
α>αh
= lim
α
1
Gα Xg∈Gα
Gα Xg∈Gα
1
U(hg)ξ = lim
α>αh
1
Gα Xg∈Gα
U(g)ξ = ξ .
1
Gα Xg∈Gα
U(g)ξ
U(g)ξ
(cid:3)
We now introduce and recall some notations and definitions which
will be used throughout the paper. Let G be a group and A a C ∗ --
algebra which we suppose always to be unital. One says that G acts
as a group of automorphisms of A if there is a representation α : g ∈
G 7→ αg ∈ Aut(A). The state ϕ ∈ S(A) is called G -- invariant if
ϕ = ϕ ◦ αg for each g ∈ G. The subset SG(A) of the G -- invariant states
is ∗ -- weakly compact in S(A), and its extremal points are called ergodic
states (w.r.t.the action of G). For G acting as a group of automorphisms
of A and a state ϕ ∈ SG(A), (πϕ, Hϕ, Uϕ, Ωϕ) is the GNS covariant
quadruple canonically associated to ϕ (see, e.g. [8, 35]). If (πϕ, Hϕ, Ωϕ)
is the GNS triple associated to ϕ, the unitary representation Uϕ of G
on Hϕ is uniquely determined by
πϕ(αg(A)) = Uϕ(g)πϕ(A)Uϕ(g)−1 ,
Uϕ(g)Ωϕ = Ωϕ , A ∈ A , g ∈ G .
If ϕ ∈ SG(A), by BG(ϕ) := (Zϕ)α we denote the fixed point algebra of
the center
Zϕ := πϕ(A)′′\ πϕ(A)′
under the adjoint action ad(Uϕ) of G. We will refer to the set
HG
ϕ := {ξ ∈ Hϕ Uϕ(g)ξ = ξ , g ∈ G}
and Eϕ as the (closed) subspace of Hϕ of the invariant vectors w.r.t. the
action of G, and the relative selfadjoint projection onto it, respectively.
Let (A, G) be a C ∗ -- dynamical system as above, together with ϕ ∈
SG(A). The invariant state ϕ is said to be G -- abelian if all the operators
Eϕπϕ(A)Eϕ mutually commute. The C ∗ -- dynamical system (A, G) is
G -- abelian if ϕ is G -- abelian for each ϕ ∈ SG(A).
Let J be any set. By definition the group of the permutations PJ of
J is made by those permutations leaving fixed all the elements of J but
10
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
a finite number of them. Then it is the direct limit of the (sub)groups
of the permutations PI, I running on all the finite subsets of J, that is
PJ :=[{PI I ⊂ J finite } .
It is expected that the group of the permutations PJ acts, in a natural
way, as a group of automorphisms of CAR(J). However this is the case,
according to the following
Proposition 3.2. The map g ∈ PJ 7→ ag−1j ∈ CAR(J), j ∈ J, ex-
tends to an action g ∈ PJ 7→ αg ∈ Aut(CAR(J)) by automorphisms of
CAR(J).
Proof. By a standard argument, we have CAR(J) ≡ A(ℓ2(J)), that is
the CAR algebra on ℓ2(J) under the notations in Section 5.2.2 of [9].
Our action on the indices is nothing but a unitary action on ℓ2(J).
This means that PJ acts as a group of Bogoliubov automorphisms of
CAR(J).
(cid:3)
We now report a result crucial in the sequel.1 Denote n := {1, . . . , n}
the finite set made of exactly n elements. If m ≤ n, m can be consid-
ered, in a canonical way, as a subset of n.
Lemma 3.3. Let 1 ≤ m, n < N. Then, for some constant c(m, n)
depending only on m, n, we have
{g ∈ PN m ∩ gn 6= ∅}
(N − 1)!
≤ c(m, n) .
Proof. Set A := {g ∈ PN m ∩ gn 6= ∅}c, and
Γ =(N − m)[ln(N − m) − 1] + (N − n)[ln(N − n) − 1]
−(N − m − n)[ln(N − m − n) − 1] − N(ln N − 1) .
It is straightforwardly seen that
A
N!
=
(N − m)!(N − n)!
(N − m − n)!N!
≈s(N − m)(N − n)
N(N − m − n)
eΓ ,
after using the Stirling formula, for N → +∞. By retaining only the
leading terms up to the order 1/N, we get
A
N!
≈ e− mn
N ≈ 1 −
mn
N
1Compare with the connected estimation in Theorem 13 of [14], concerning the
classical case.
DE FINETTI THEOREM
which leads to
{g ∈ PN m ∩ gn 6= ∅}
N!
= 1 −
A
N!
≈
mn
N
.
11
(cid:3)
We end the section reporting the definition (cf. [30], Definition 3.3)
concerning the action of a group as a Large Group of Automorphisms.
Definition 3.4. Let g ∈ G 7→ αg ∈ Aut(A) be an action of a group G
on the C ∗ -- algebra A. We say that G is represented (or acts) as a large
group of automorphisms if, for each selfadjoint A and each ϕ ∈ SG(A)
conv ({πϕ(αg(A)) g ∈ G})\ πϕ(A)′ 6= ∅ .
In the next section we will establish that PJ acts on CAR(J) as
a large group of automorphisms (cf. Theorem 4.2). We underline
that this property is not directly used for the main result of the paper
concerning the structure of the symmetric (i.e.
invariant under the
action of PJ ) states. It comes out only for constructing a conditional
expectation from the GNS von Neumann algebra of CAR(J) onto the
invariant elements of the center, that allows to obtain some convergence
results which may have some interest in general (see Proposition 4.3
and Proposition 5.4).
4. symmetric states on the CAR algebra
A state ϕ ∈ S(CAR(J)) is called symmetric if it is invariant under
the action of the group PJ of all the finite permutations of the set
J. Following the notation introduced above, SPJ (CAR(J)) denotes
the ∗ -- weakly compact subset of all the symmetric states of CAR(J).
Furthermore we refer to E (SPJ (CAR(J))) as the set of all the extremal
symmetric states, that is the invariant states which are ergodic w.r.t.
the action of PJ .
In the section we investigate some of the basic ergodic properties
enjoyed by the symmetric states. To this aim, denote M the Cesaro
Mean w.r.t. PJ , given for a generic object f (g) by
M{f (g)} := lim
I↑J
f (g) ,
1
PI Xg∈PI
provided the l.h.s. exists in the appropriate sense. As usual, I ⊂ J
runs over all the finite parts of J.
Theorem 4.1. Let ϕ ∈ SPJ (CAR(J)). Then the following assertions
hold true.
12
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
(i) The state ϕ is even.
(ii) The state ϕ is asymptotically Abelian in average:
M{ϕ(C[αg(A), B]D)} = 0 , A, B, C, D ∈ CAR(J) .
(iii) ϕ ∈ E (SPJ (CAR(J))) if and only if it is weakly clustering:
M{ϕ(αg(A)B)} = ϕ(A)ϕ(B) , A, B ∈ CAR(J) .
Proof. (i) Let A be localized and odd. Proposition 3.1 gives
{Eϕπϕ(A)Eϕ, Eϕπϕ(A∗)Eϕ}
= M{Eϕπϕ(A)Uϕ(g)πϕ(A∗)Eϕ + Eϕπϕ(A∗)Uϕ(g)πϕ(A)Eϕ}
=M(Eϕπϕ({A, αg(A∗)})Eϕ)
By Lemma 3.3 and the CAR relations, one finds that the quantity above
is equal to zero, where the limit in the Cesaro mean is understood in
the strong operator topology.2 This implies that Eϕπϕ(A)Eϕ = 0, that
is
ϕ(A) = hEϕπϕ(A)EϕΩϕ, Ωϕi = 0 .
Hence ϕ vanishes on the localized odd elements. Since CAR(J) ∼
CAR+(J)L CAR−(J) as a Banach space, we can approximate a generic
odd element A with a sequence {An}n∈N made by odd and localized
elements. By the above result, one obtains
ϕ(A) = ϕ(lim
n
An) = lim
n
ϕ(An) = 0 .
This means that ϕ vanishes on all the odd elements, that is ϕ is even.
(ii) The same computations as before show that, if A or B is even,
then
(4.1)
[Eϕπϕ(A)Eϕ, Eϕπϕ(B)Eϕ] = 0 .
By a standard approximation argument, we can reduce the matter to
localized elements. Fix A even. Proposition 3.1, Lemma 3.3 and (4.1),
give
M{ϕ(Cαg(A)BD)} = M{ϕ(αg(A)CBD)}
=hπϕ(A)Eϕπϕ(CBD)Ωϕ, Ωϕi = hπϕ(CBD)Eϕπϕ(A)Ωϕ, Ωϕi
=M{ϕ(CBDαg(A))} = M{ϕ(CBαg(A)D)} .
By considering A odd and splitting C, D in their even and odd parts,
the result is reached by similar computations as above.
2The analogous case based on the spatial translations has been treated in [9],
Example 5.2.21, where the particular form of the action of Zd on CAR(Zd) as the
shift, allows to reach directly the result without using Lemma 3.3.
DE FINETTI THEOREM
13
(iii) The weak clustering condition is equivalent to dim(HPJ
ϕ ) = 1,
and it is immediate to show that implies ergodicity (see [29], Proposi-
tion 3.1.10). The converse assertion follows from Theorem 4.2, that is
CAR(J) is PJ -- abelian (see [29], Proposition 3.1.12).
(cid:3)
Theorem 4.2. The group PJ acts as a Large Groups of Automor-
phisms on CAR(J), and the C ∗ -- dynamical system (CAR(J), PJ , α) is
PJ -- abelian.
Proof. By taking into account (4.1) and (i) of Theorem 4.1, we conclude
that (CAR(J), PJ ) is PJ -- Abelian. Fix an arbitrary integer n ∈ N,
A, B1, . . . , Bn, C ∈ CAR(J) and ϕ ∈ SPJ (CAR(J)). Now, by (ii) of
Theorem 4.1, for each ε > 0 there exists a finite set I ⊂ J such that,
for any k = 1, . . . , n,
αg(A)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ϕ C ∗" Xg∈PI
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
PI Xg∈PI
PI ! , Bk# C!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ϕ(C ∗[αg(A), Bk]C)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
< ε .
1
This leads to the assertion by Theorem 3.5 of [30].
(cid:3)
As PJ is acting on CAR(J) as a large group of automorphisms, by
Theorem 3.1 of [30], for each state ϕ ∈ SPJ (CAR(J)) there exists a
conditional expectation
Φϕ : πϕ(CAR(J))′′ → BPJ (ϕ)
of the von Neumann algebra πϕ(CAR(J))′′ onto BPJ (ϕ). As the state
ϕ is asymptotically Abelian in average (cf. (ii) of Theorem 4.1) we
prove the following result which, even if is not used in the sequel, may
have an interest in itself.
Proposition 4.3. Let ϕ ∈ SPJ (CAR(J)) and A ∈ CAR(J). Then
(4.2)
w − lim
I↑J
1
PI Xg∈PI
Uϕ(g)πϕ(A)Uϕ(g)−1 = Φϕ(πϕ(A)) .
Proof. Let {Iβ} ⊂ {I} any subnet of the Cauchy net {I} of all the
Uϕ(g)πϕ(A)Uϕ(g)−1 converges
finite subsets I ⊂ J such that
1
PIβ Xg∈PIβ
in the weak operator topology, which exists by compactness. Fix an
arbitrary h ∈ PJ. Then there exists a βh such that I ⊃ Iβh implies
14
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
h ∈ PI. We get
w − lim
Iβ
1
Uϕ(h)(cid:18) 1
PIβ Xg∈PIβ
PIβ Xg∈PIβ
PIβ Xg∈PIβ
1
= w − lim
Iβ ⊃Iβh
= w − lim
Iβ ⊃Iβh
1
= w − lim
Iβ
PIβ Xg∈PIβ
Uϕ(g)πϕ(A)Uϕ(g)−1(cid:19)Uϕ(h)−1
Uϕ(hg)πϕ(A)Uϕ(hg)−1
Uϕ(g)πϕ(A)Uϕ(g)−1
Uϕ(g)πϕ(A)Uϕ(g)−1 .
Thus, each weak limit point of the Cesaro net on the l.h.s. of (4.2) is
invariant. By using the asymptotic Abelianess property (ii) of Theorem
4.1, and arguing as in the proof of Lemma 5.3 of [30], one shows that
the limit above is in Zϕ. Then it belongs to BPJ (ϕ). But BPJ (ϕ) can
contain at most one of such limit points since PJ acts on CAR(J) as a
large group of automorphisms (cf. proof of Theorem 3.1 in [30]). As a
consequence, the limit is precisely Φϕ(πϕ(A)).
(cid:3)
5. the structure of the symmetric states: de finetti
theorem
The present and the following sections are mainly concerned with the
characterization of the extremal symmetric states. In particular here
we consider the case in which J is countable, hence J ≡ N. We pre-
liminary report the definition of the sequence of permutations {gn}n∈N
in [31] given by
(5.1)
2n−1 + k if 1 ≤ k ≤ 2n−1 ,
if 2n−1 < k ≤ 2n ,
k − 2n−1
if 2n < k .
k
gn (k) :=
Definition 5.1. A state ϕ ∈ S(CAR(N)) is said to be strongly clus-
tering if, for every A, B ∈ CAR(N)
lim
n
ϕ(αgn(A)B) = ϕ(A)ϕ(B) .
Lemma 5.2. If ϕ ∈ SPN(CAR(N)) is extremal, then for each A ∈
CAR(N)
w − lim
n
[πϕ(αgn(A))Ωϕ] = ϕ(A)Ωϕ .
DE FINETTI THEOREM
15
Proof. By a standard approximation argument, we can reduce the mat-
ter to A ∈ CAR0(N). Let g ∈ PN. As shown in the proof of Lemma 2.6
of [31], there exists nA,g such that n > nA,g implies αggn(A) = αgn(A).
This means that any weak limit point (which exists by compactness) of
the sequence (cid:8)πϕ(αgn(A))Ωϕ(cid:9) is an invariant vector under the action
of PN, that is it belongs to HPN
ϕ . Let
ξ := w − lim
k
Uϕ(gnξ(k))πϕ(A)Uϕ(gnξ(k))−1Ωϕ ≡ w − lim
k
be one of such limit points. Since ξ is a vector in HPN
ϕ and ϕ is extremal,
by (iii) of Theorem 4.1, one has ξ = Γ(A, ξ)Ωϕ. By using (5.1), we
obtain
πϕ(αgnξ (k)(A))Ωϕ
Γ(A, ξ) =(cid:10) lim
k
= lim
Uϕ(gnξ(k))πϕ(A)Uϕ(gnξ(k))−1Ωϕ, Ωϕ(cid:11)
k (cid:10)πϕ(αgnξ (k)(A)Ωϕ, Ωϕ(cid:11) = ϕ(A) .
Namely, there is only one of such weak limit points in Hϕ, which is
ϕ(A)Ωϕ.
(cid:3)
Theorem 5.3. Let ϕ ∈ SPN(CAR(N)). Then the following are equiva-
lent.
(i) ϕ ∈ E(SPN(CAR(N))),
(ii) ϕ is strongly clustering,
ρ for some even state ρ ∈ M2(C).
(iii) ϕ =YN
Proof. (i) ⇒ (ii) Suppose ϕ is extremal and take A, B ∈ CAR(N).
Then by Lemma 5.2, we get
lim
n
ϕ(Aαgn(B)) = lim
n
hπϕ(αgn(B))Ωϕ, πϕ(A∗)Ωϕi = ϕ(A)ϕ(B) ,
that is ϕ is strongly clustering.
(ii) ⇒ (i) Choose a vector ξ ⊥ Ωϕ belonging to HPN
ϕ , and fix ε > 0.
Then there exists B ∈ CAR(N) such that kξ − πϕ(B)Ωϕk < ε/2. Let
A ∈ CAR(N) such that kπϕ(A)k ≤ 1. We get
hξ, πϕ(A)Ωϕi = hUϕ(gn)πϕ(A∗)Uϕ(gn)−1ξ, Ωϕi
≤hUϕ(gn)πϕ(A∗)Uϕ(gn)−1πϕ(B)Ωϕ, Ωϕi + ε/2 = ϕ(αgn(A∗)B) + ε/2 .
Suppose now ϕ is strongly clustering. By taking the limit for n → ∞
on both sides, we get
hξ, πϕ(A)Ωϕi ≤ ϕ(A∗)ϕ(B) + ε/2 < ε .
As ε > 0 is arbitrary and Ωϕ is cyclic for πϕ(CAR(N)), we get ξ = 0.
This means that HPN
ϕ is one dimensional, which implies (and it is indeed
equivalent by Proposition 3.1.12 of [29]) that ϕ is extremal.
16
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
Obviously, if ϕ is a product state of a single state as in (iii), then it
is strongly clustering. Suppose now that (ii) holds true. After fixing
j ∈ N, we start by identifying CAR({j}) with M2(C) and, as usual,
denoting ιj : M2(C) → CAR(N) the related embedding.
By Theorem 1 of [5], any product state is uniquely determined by
the product of the values of the state on the generators. Hence, it is
enough to check, for each n ∈ N and A1, . . . , An ∈ M2(C),
(5.2)
ϕ(ι1(A1) · · · ιn(An)) =
ρ(Aj) ,
n
Yj=1
The proof now proceeds as in Theorem 2.7 of [31]. We give the de-
tails with the appropriate modifications. Define for j ∈ N, ρj(A) :=
ϕ(ιj(A)). As ϕ is even (cf. Theorem 4.1.(i)), all the ρj are even. In
addition, for i, j ∈ N, ρi = ρj =: ρ as ϕ is symmetric. Equation (5.2)
can be achieved by an induction procedure. Indeed, for n = 1 it follows
immediately, so we suppose it holds true till n − 1. Fix ε > 0. By using
the inductive hypothesis and the strong clustering property , we get
(5.3)
ϕ(ι1(A1) · · · ιn(An−1)αgm(ιn(An))) −
= ϕ(ι1(A1) · · · ιn(An−1)αgm(ιn(An))) − ϕ(ι1(A1) · · · ιn(An−1))ρ(An) < ε
for some m > n. Choose now a permutation g ∈ PN such that g(j) = j
if 1 ≤ j < n and g(n) = gm(n). Then
ϕ(ι1(A1) · · · ιn(An)) =ϕ(αg(ι1(A1) · · · ιn(An)))
=ϕ(ι1(A1) · · · ιn(An−1)αgm(ιn(An)))
which, combined with (5.3), leads to the assertion as ε > 0 is arbitrary.
(cid:3)
When a state is extremal among the symmetric ones, the result of
Proposition 4.3 can be strengthened as we see in the following propo-
sition, inserted for the sake of completeness.
Proposition 5.4. If ϕ ∈ E(SPN(CAR(N))). Then for each A ∈ CAR(N),
(5.4)
w − lim
n
Uϕ(gn)πϕ(A)Uϕ(gn)−1 = ϕ(A)1I = Φϕ(πϕ(A)) .
Proof. By a standard approximation argument, it is enough to con-
sider ξ = πϕ(B)Ωϕ, η = πϕ(C ∗)Ωϕ, and reduce the matter to A, B, C ∈
CAR0(N). Let A = A+ + A−, B = B+ + B− the split of A, B into
it is enough to assume
the even and odd part. By Theorem 5.3,
(cid:12)(cid:12)(cid:12)(cid:12)
n
Yj=1
ρ(Aj)(cid:12)(cid:12)(cid:12)(cid:12)
DE FINETTI THEOREM
17
that ϕ is strongly clustering. As ϕ is even, by using the standard
(anti)commutation relations, we get
hUϕ(gn)πϕ(A)Uϕ(gn)−1ξ, ηi = lim
n
ϕ(Cαgn(A)B)
lim
n
= lim
n
ϕ(CBαgn(A+)) + lim
n
ϕ(CB+αgn(A−)) − lim
n
ϕ(CB−αgn(A−))
=ϕ(CB)ϕ(A+) + ϕ(CB+)ϕ(A−) − ϕ(CB−)ϕ(A−)
=ϕ(CB)ϕ(A+) + ϕ(CB+)ϕ(A−)
=ϕ(CB)ϕ(A+) + ϕ(CB+)ϕ(A−) + ϕ(CB−)ϕ(A−)
=ϕ(CB)ϕ(A) = h(ϕ(A)1I)ξ, ηi .
This means that the above weak limit is in BPN(ϕ). But, as observed
in the proof of (4.2), it is nothing else than Φϕ(πϕ(A)).
(cid:3)
Notice that, in the case of the C ∗ -- tensor product (i.e. when the
system is asymptotically Abelian in norm) (5.4) is satisfied for each
symmetric state, see [31], Lemma 2.6. In our situation, we have merely
the expected average property (4.2), and (5.4) is satisfied only for ex-
tremal symmetric states.
Having characterized the extremal symmetric states as the Araki --
Moriya product states, we are interested in the ergodic decomposi-
tion of an invariant state under the action of the permutation group.
Namely, we want to express every symmetric state as the barycenter
of a unique (maximal) Radon probability measure on SPN(CAR(N))
whose support is E(SPN(CAR(N))). The existence of such a measure
is granted, since SPN(CAR(N)) is metrizable (see [8], Proposition 4.1.3
It is unique if and only if SPN(CAR(N)) is a
and Theorem 4.1.11).
Choquet simplex (see [8], Theorem 4.1.15). This is our case, as we see
in the following result.
Theorem 5.5. SPN(CAR(N)) is a Choquet simplex. Then, for each
ϕ ∈ SPN(CAR(N)) there exists a unique maximal Radon probability
measure µ on SPN(CAR(N)) such that
ϕ(A) =Z ψ(A)dµ(ψ) , A ∈ CAR(N) ,
and µ(E(SPN(CAR(N))) = 1.
Proof. By Theorem 4.2, (CAR(N), PN) is PN -- Abelian, then, as a conse-
quence of Theorem 3.1.14 of [29], SPN(CAR(N)) is a Choquet simplex.
The last part is a rephrasing of the above discussion.
(cid:3)
The property for the set of extremal states to be ∗ -- weakly closed
In fact, in [7] it is shown that a simplex with
implies a nice result.
18
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
closed boundary is affinely isomorphic to the probability measures on
a compact Hausdorff space. These facts are stated in the following
proposition, whose proof, based on Lemma 2.2, Theorem 5.3 and ar-
guments analogous to those developed in Theorem 2.8 of [31], is left to
the reader.
Proposition 5.6. The Choquet simplex SPN(CAR(N)) has a ∗ -- weakly
closed boundary and is affinely isomorphic to the probability measures
on a closed interval.
Put A := CAR(N). Let ϕ ∈ S(A). Obviously, πϕ(A)′′ is a hyperfinite
von Neumann algebra. A state ϕ on a C ∗-algebra A is called factor
state if the double commutant πϕ(A)′′ of πϕ(A) is a factor. The state
ϕ is said of type X if πϕ(A)′′ is of type X, where X = I∞, II1, II∞, III
or IIIλ, λ ∈ [0, 1] according to the Connes' classification of the type III
factors [10].
Proposition 5.7. Let ρµ be a even state in M2(C) with eigenvalues µ
ρµ the corresponding product state on CAR(N).
and 1−µ, and ϕµ :=YN
Then
(1) ϕµ is a factor state of type I∞ if and only if µ = 0 or µ = 1.
(2) ϕµ is a factor state of type II1 if and only if µ = 1/2.
(3) πϕµ(A)′′ is of type IIIλ if and only if 0 < µ < 1/2 and λ = µ
1−µ ,
or 1/2 < µ < 1 and λ = 1−µ
µ .
Proof. By Lemma 2.3, we get (under the same notations) πϕµ(A)′′ ∼
πωµ(B)′′. But the πωµ(B)′′ are factors whose type is determined by the
ratio between the smallest and the largest eigenvalue of the trace op-
erator describing ρµ, according to the three possibilities listed above in
the statement. The reader is referred to A.17 of [32] and the references
cited therein.
(cid:3)
Notice that we do not have the type II∞, and furthermore, the type
II1 occurs only for ϕ1/2. Thus, the latter gives rise the trivial face made
of a singleton.3 The cases of the portions corresponding to I∞ and III
are covered by the next result. Recall that a face of a given simplex K
is a convex subset F of K such that, if χ ∈ F , ψ ∈ K and for µ > 0,
ψ ≤ µχ, implies ψ ∈ F . The proof of the forthcoming proposition
follows mutatis mutandis the lines of the analogous Lemma 2.9 in [31].
The details are left to the reader.
3The II∞ might appear for the more general situation investigated in [16], where
the regrouped CAR algebra with 2d generators is attacked to each site.
DE FINETTI THEOREM
19
Proposition 5.8. If X denotes I∞ or III, and
SPN(CAR(N))X := {ϕ ∈ SPN(CAR(N)) ϕ is of type X} ,
then SPN(CAR(N))X is a face of SPN(CAR(N)).
Recalling that the boundary of an arbitrary nontrivial compact con-
vex K is nonvoid by the Krein -- Milman Theorem, by Propositions 5.6
and 5.7, one has that E(SPN(CAR(N))III) consists of two open con-
nected components of E(SPN(CAR(N))), E(SPN(CAR(N))I∞) is given
by two states, that is the pure states, whereas E(SPN(CAR(N))II1) is
made by one state, that is the unique trace.
We end the section with a brief sketch on the extension of the situa-
tion in Proposition 5.8 to more general cases and leave further details
to the reader.
Let A be a separable C ∗ -- algebra and E ∈ (0, 1) be a Borel set. Fix
a state ϕ ∈ S(A) and put M := πϕ(A)′′. Let
M =Z ⊗
Γ
Mγ dν(g)
(5.5)
be its direct integral decomposition into von Neumann factors. Here,
(Γ, ν) is a standard probability measure space which can be chosen
as (spec(Zϕ), νω⌈Zϕ ), spec(Zϕ) and ω being the spectrum of Zϕ, and a
faithful normal state on M respectively, see e.g. [29, 35].
We say that the state ϕ is of type XE if M contains only type IIIλ
factors for some λ ∈ E in its direct integral decomposition (5.5). Let
ΓE ⊂ Γ be the subset made of the γ such that Mγ is a type IIIλ factor
for some λ ∈ E. By Theorem 2.2 of [34] (see also (ii) in Theorem 21.2
of [27]), ΓE is a ν -- measurable set. The fact that ϕ is of type XE simply
means that ν(ΓE) = 1. Put
SPN(A)XE := {ϕ ∈ SPN(A) ϕ is of type XE} .
Remark 5.9. By using the same lines as in Lemma 2.9 of [31], we can
show that SPN(CAR(N))XE is a face of SPN(CAR(N)).
In fact, let ϕ = λϕ1 + (1 − λ)ϕ2, λ ∈ [0, 1] be a convex combination
of states in SPN(CAR(N))XE . As in the proof of Lemma 2.9 of [31], we
find P1, P2 ∈ Zϕ such that
πϕk (A)′′ = Pkπϕ(A)′′ ,
k = 1, 2 .
Of course, P1πϕ(A)′′ contains only type IIIλ factors, λ ∈ E, in its
direct integral factor decomposition (5.5). The same happens to (P2 −
P1P2)πϕ(A)′′ as P2 − P1P2 ≤ P2. Since one can show that P1 + P2 −
P1P2 = I, it follows that ϕ is of type XE. So SPN(CAR(N))XE is
closed under convex combinations. Let now take ω ∈ SPN(CAR(N)),
20
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
ϕ ∈ SPN(CAR(N))XE with ω ≤ λϕ for some λ > 0. As before, there
exists a projection P ∈ Zϕ such that πω(A)′′ = P πϕ(A)′′, which implies
that ω ∈ SPN(CAR(N))XE .
6. symmetric states in the uncountable case
The starting point to obtain the characterization of the extremal
symmetric states is the clustering property given in Definition 5.1. It
is the bridge between the extremality and the property to being a
product state (cf. Theorem 5.3). Suppose that CAR(J) is generated
by infinitely uncountably many annihilators, that is J is uncountable.
Fix a state ϕ ∈ S(CAR(J)). We generalize the clustering property in
the following way.
We start with a pair I := (I, ν), where I ⊂ J is countable and
ν : I → N is a bijection defining an order i1, i2, . . . on I. Define the
sequence {gI
n := gν(in), where gk ∈ PN is given in (5.1).
Fix now ϕ ∈ S(CAR(J)).
It is said strongly clustering if, for each
A, B ∈ CAR(J) there exists I := (I, ν) with A, B ∈ CAR(I) such that
n}n∈N ⊂ PJ as gI
(6.1)
lim
n
ϕ(αgI
n (A)B) = ϕ(A)ϕ(B) .
n(A) = αgI
As a preliminary fact we note that, if I := (I, ν) with A ∈ CAR(I)
localized, and g ∈ PJ , then there exists nA,g such that n > nA,g implies
αggI
n (A). In fact, as g changes only a finite number of indices
in I, say up to m, and A is localized, say in the first s elements of I,
it is enough to choose n such that max{m, s} ≤ 2n−1. By using this
fact, Lemma 5.2 holds true also in the present situation. Namely, fix
A ∈ CAR(J). Then for each I = (I, ν) such that A ∈ CAR(I)
(6.2)
w − lim
n
[πϕ(αgI
n (A))Ωϕ] = ϕ(A)Ωϕ ,
provided ϕ is extremal.
Now we are ready to extend Theorem 5.3 to uncountable case.
Theorem 6.1. Let ϕ ∈ SPJ (CAR(J)). Then ϕ ∈ E(SPJ (CAR(J))) if
ρ for some even state ρ ∈ M2(C).
and only if ϕ =YJ
Proof. If ϕ is a product state, it is obviously strongly clustering. Sup-
pose now ϕ be strongly clustering and fix a sequence {Ai}i∈N ⊂ M2(C).
The proof proceeds by induction. Choose I := (I, ν) such that A :=
ιj1(A1) · · · ιjn−1(An−1), B := ιjn(An) belong to I and (6.1) holds true.
After noticing that
ϕ(gI
n(A)B) = ϕ⌈CAR(I)(gI
n(A)B) .
DE FINETTI THEOREM
21
we can reduce the matter to the countable situation, that is to Theorem
5.3, as CAR(I) ∼ CAR(N). Thus a symmetric state is a product one
if and only if it is strongly clustering. On the other hand, if ϕ is
extremal we conclude by (6.2) that ϕ is strongly clustering. For the
reverse implication we reason as in Theorem 5.3 as well. Indeed, choose
a vector ξ ⊥ Ωϕ belonging to HPJ
ϕ and fix ε > 0. Then there exists B ∈
CAR(J) such that kξ − πϕ(B)Ωk < ε/2. Let A ∈ CAR(J) such that
kπϕ(A)k ≤ 1. We get for each I = (I, ν) such that A, B ∈ CAR(I),
(6.3)
hξ, πϕ(A)Ωϕi ≤ ϕ(αgI
n(A∗)B) + ε/2 .
As ϕ is strongly clustering, there exists an I := (I, ν) such that A, B ∈
CAR(I) and (6.1) holds. Taking the limit on both sides in (6.3), one
obtains
hξ, πϕ(A)Ωϕi ≤ lim
n
ϕ(αgI
n(A∗)B) + ε/2 = ϕ(A∗)ϕ(B) + ε/2 < ε .
We conclude as before that ξ = 0 and then ϕ is extremal. Since we ver-
ified that both properties (i.e. being a product state and extremeness)
are equivalent to strong clustering, the proof is complete.
(cid:3)
Remark 6.2. Theorem 6.1 holds true mutatis mutandis for the case
of infinite tensor product of C ∗ -- algebras considered in [31], when the
index set is uncountable.
In the general (uncountable) case it is possible to achieve the ergodic
decomposition of a symmetric state ϕ as in Theorem 5.5. (countable
case). In fact, the convex of the symmetric states on CAR(J) is yet a
Choquet simplex (see [29], Theorem 3.1.14). The difference w.r.t. the
countable case consists in the fact that SPJ (CAR(J)) is not metriz-
able. Then the unique maximal decomposing measure µ of ϕ is only
pseudosupported on E(SPJ (CAR(J))). Namely, we still have
ϕ(A) =Z ψ(A)dµ(ψ) , A ∈ CAR(J) ,
but here E(SPJ (CAR(J)) is merely a Borel set which is not a Baire one.
Then the maximal measure µ satisfies µ(B) = 1 for each Baire set B
containing E(SPJ (CAR(J)) (see [8], Theorem 4.1.11).
Finally, one realizes that most of the results at the end of Section
5, except Remark 5.9, may be extended to the general (uncountable)
case. For example, as in Proposition 5.6, one sees that the simplex of
the symmetric states on CAR(J) has a ∗ -- weakly closed boundary and
is affinely isomorphic to the regular probability measures on the unit
interval.
22
VITONOFRIO CRISMALE AND FRANCESCO FIDALEO
Acknowledgements
The second -- named author would like to thank Simion Stoilow Insti-
tute of Mathematics of the Romanian Academy for the warm hospital-
ity, thanks to a BITDEFENDER invited professorship position, where
part of the present work has been done.
References
[1] Accardi L., Ben Ghorbal A., Crismale V., Lu Y. G. Singleton conditions
and quantum De Finetti's theorem, Infin. Dimens. Anal. Quantum Probab.
Relat. Top. 11 (2008), 639 -- 660.
[2] Accardi L., Fidaleo F., Mukhamedov F. Quantum Markov states and chains
on the CAR algebras, Infin. Dimens. Anal. Quantum Probab. Relat. Top.
10 (2007), 165 -- 183.
[3] Accardi L., Lu Y.G. A continuous version of De Finetti's theorem, Ann.
Probab. 21 (1993), 1478 -- 1493.
[4] Araki H., Moriya H. Equilibrium statistical mechanics of Fermion lattice
systems, Rev. Math. Phys. 15 (2003), 93 -- 198.
[5] Araki H., Moriya H. Joint extension of states of subsystems for a CAR
system, Commun. Math. Phys. 237 (2003), 105 -- 122.
[6] Barreto S. D., Fidaleo F. Disordered Fermions on lattices and their spectral
properties, J. Stat. Phys. 143 (2011), 657 -- 684.
[7] Bauer H. Schilowscher rand und dirichletsches problem, Ann. Inst. Fourier
11 (1961), 89 -- 136.
[8] Bratteli O., Robinson D. W. Operator algebras and quantum statistical me-
chanics I, Springer, Berlin -- Heidelberg -- New York, 1981.
[9] Bratteli O., Robinson D. W. Operator algebras and quantum statistical me-
chanics II, Springer, Berlin -- Heidelberg -- New york, 1981.
[10] Connes A. Une classification des facteurs de type III, Ann. Scient. ´Ec. Norm.
Sup. 6 (1973), 133 -- 252.
[11] Christandl M., Konig R., Mitchison G., Renner R. One -- and -- a -- half quantum
de Finetti theorems, Comm. Math. Phys. 273 (2007), 473 -- 498.
[12] Christandl M., Toner B. Finite de Finetti theorem for conditional probability
distributions describing physical theories, J. Math. Phys. 50 (2009), no. 4,
042104, 11 pp.
[13] De Finetti B. Funzione caratteristica di un fenomeno aleatorio, Atti Accad.
Naz. Lincei, VI Ser., Mem. Cl. Sci. Fis. Mat. Nat. 4 (1931), 251 -- 259.
[14] Diaconis P., Freedman D. A. Finite exchangeable sequences, Ann. Prob. 8
(1980), 745 -- 764.
[15] Fannes M., Lewis J. T., Verbeure A. Symmetric stats of composite systems,
Lett. Math. Phys. 15 (1988), 255 -- 260.
[16] Fidaleo F. Fermi Markov states, J. Operator Theory, 66 (2011), 385 -- 414.
[17] Freedman D. A. Invariance under mixing which generalize De Finetti's the-
orem: Continuous time parameter, Ann. Math. Stat. 34 (1963), 1194 -- 1216.
[18] Hewitt E., Savage L. F. Symmetric measures on Cartesian products, Trans.
Amer. Math. Soc. 80 (1955), 470 -- 501.
DE FINETTI THEOREM
23
[19] Hudson R. L., Moody G. R. Locally normal symmetric states and an ana-
logue of De Finetti's theorem, Z. Wahr. Verw. Gebiete 33 (1976), 343 -- 351.
[20] Konig R., Mitchison G. A most compendious and facile quantum de Finetti
theorem, J. Math. Phys. 50 (2009), no. 1, 012105, 20 pp.
[21] Kostler C. A noncommutative extended De Finetti theorem, J. Funct. Anal.
258 (2010), 1073 -- 120.
[22] Kostler C, Speicher R. A noncommutative De Finetti theorem: invariance
under quantum permutations is equivalent to freeness with amalgamation,
Commun. Math. Phys. 291 (2009), 473 -- 490.
[23] Lehner F. Cumulants in noncommutative probability theory IV. noncrossing
cumulants: De Finetti's theorem and Lp -- inequalities, J. Funct. Anal. 239
(2006), 214 -- 246.
[24] Leverrier A., Cherf N. J. A quantum De Finetti theorem in phase space
representation, Phys. Rev. A 80 010102 (2009).
[25] Leverrier A., Karpov E., Grangier P., Cherf N. J. Security of continuous-
variable QKD: exploiting symmetries in phase space, New Journal of Physics
11 115009 (2009)
[26] Matsui T. Ground states of Fermions on lattices, Commun. Math. Phys.
182 (1996), 723 -- 751.
[27] Nielsen O. A. Direct integral theory, Marcel Dekker, New York -- Basel, 1980.
[28] Renner R., Cirac I. J. A de Finetti representation theorem for infinite dimen-
sional quantum systems and applications to quantum cryptography, Phys.
Rev. Lett. 102 110504 (2009).
[29] Sakai S. C ∗ -- algebras and W ∗ -- algebras, Springer, Berlin -- Heidelberg -- New
York 1971.
[30] Størmer E. Large groups of automorphisms of C ∗ -- algebras, Commun. Math.
Phys. 5 (1967), 1 -- 22.
[31] Størmer E. Symmetric states of infinite tensor products of C ∗ -- algebras, J.
Funct. Anal. 3 (1969), 48 -- 68.
[32] Stratila S. Modular theory in operator algebras, Abacus press, Tunbridge
Wells, Kent (1981).
[33] Stratila S., Zsid´o, L. Lectures on von Neumann algebras, Abacus press, Tun-
bridge Wells, Kent, (1979).
[34] Sutherland C. Crossed products, direct integrals and Connes' classification
of type III factors, Math. Scand. 40 (1977), 209 -- 214.
[35] Takesaki M. Theory of operator algebras I, Springer, Berlin -- Heidelberg -- New
York 1979.
[36] Takesaki M. Theory of operator algebras III, Springer, Berlin -- Heidelberg --
New York 1979.
Vitonofrio Crismale, Dipartimento di Matematica, Universit`a degli
studi di Bari, Via E. Orabona, 4, 70125 Bari, Italy
E-mail address: [email protected]
Francesco Fidaleo, Dipartimento di Matematica, Universit`a di Roma
Tor Vergata, Via della Ricerca Scientifica 1, Roma 00133, Italy
E-mail address: [email protected]
|
1603.04678 | 1 | 1603 | 2016-03-14T09:40:27 | The $C^*$-algebras of quantum lens and weighted projective spaces | [
"math.OA",
"math.QA"
] | It is shown that the algebra of continuous functions on the quantum $2n+1$-dimensional lens space $C(L^{2n+1}_q(N; m_0,\ldots, m_n))$ is a graph $C^*$-algebra, for arbitrary positive weights $ m_0,\ldots, m_n$. The form of the corresponding graph is determined from the skew product of the graph which defines the algebra of continuous functions on the quantum sphere $S_q^{2n+1}$ and the cyclic group $\mathbb{Z}_N$, with the labelling induced by the weights. Based on this description, the K-groups of specific examples are computed. Furthermore, the K-groups of the algebras of continuous functions on quantum weighted projective spaces $C(\mathbb{WP}_q^n(m_0,\ldots, m_n))$, interpreted as fixed points under the circle action on $C(S_q^{2n+1})$, are computed under a mild assumption on the weights. | math.OA | math |
THE C ∗-ALGEBRAS OF QUANTUM LENS
AND WEIGHTED PROJECTIVE SPACES
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
Abstract. It is shown that the algebra of continuous functions on the quantum 2n+1-
dimensional lens space C(L2n+1
(N ; m0, . . . , mn)) is a graph C ∗-algebra, for arbitrary
positive weights m0, . . . , mn. The form of the corresponding graph is determined from
the skew product of the graph which defines the algebra of continuous functions on
the quantum sphere S 2n+1
and the cyclic group ZN , with the labelling induced by
the weights. Based on this description, the K-groups of specific examples are com-
puted. Furthermore, the K-groups of the algebras of continuous functions on quantum
weighted projective spaces C(WPn
q (m0, . . . , mn)), interpreted as fixed points under the
circle action on C(S 2n+1
), are computed under a mild assumption on the weights.
q
q
q
1. Introduction
The aim of the present paper is investigation of noncommutative C ∗-algebras of con-
tinuous functions on quantum deformations of two classes of (possibly singular) spaces,
namely the lens and weighted projective spaces. Central to these studies is identification
of the algebras in question as graph C ∗-algebras or subalgebras of graph C ∗-algebras
corresponding to certain actions of the circle group.
In classical geometry both (weighted) lens and weighted projective spaces are obtained
as quotients of groups acting (with possible different positive integer weights) on the
odd dimensional spheres. In the former case the acting group is a finite cyclic group,
in the latter it is the circle group. Depending on the choice of weights (relatively to
the order of the acting group) the action can be free or almost free, thus leading to
most easily accessible examples of orbifolds in the latter case, [31] or [3]. The fact
that these orbifolds are defined by explicit group actions on the spheres suggests an
accessible way of defining and studying their noncommutative counterparts through
the exploration of analogous actions on quantum odd dimensional spheres, [27]. On this
premise, quantum lens spaces were defined in [21], and more recently weighted projective
spaces were introduced in [9] with a specific aim of studying quantum or noncommutative
orbifolds, a task subsequently undertaken for example in [18], [15], [25], [19] and a series
of papers by the first author of the present paper. Surprising results of these initial
studies include observations that, upon deformation, classically non-free actions become
free (see e.g. [10], [6] or [15]) and deformations of non-smooth objects behave as if they
were deformations of smooth objects (see e.g. [8], [12]). Despite a significant progress
in understanding the structure of weighted projective spaces in special cases (see e.g.
Date: August 31, 2018.
2010 Mathematics Subject Classification. 46L87; 58B32; 58B34.
Key words and phrases. Quantum lens spaces; quantum weighted projective spaces; graph C*-
algebras.
1
2
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
[15], [10]) and deformations of classically singular lens spaces, the full picture is still not
complete and does not include some important classes of the objects in question. Guided
by the experience of working with quantum lens spaces that correspond to classically
non-singular case, [21], graph C ∗-algebras appear to offer an effective tool to fill in
this gap in our understanding of the structure of the relevant C ∗-algebras and their
K-theoretic invariants. We are exploiting this opportunity here.
A directed graph G = (G0, G1, , σ) consists of two sets G0 and G1 (the former the
set of vertices and the latter the set of edges) and two mappings , σ : G1 → G0, called
the range and source, respectively. Given a graph G with countably many vertices and
edges, C ∗(G) denotes the C ∗-algebra defined as follows, [17]. C ∗(G) is the universal
C ∗-algebra generated by a set {Pv v ∈ G0} of mutually orthogonal projections and a
set {Se e ∈ G1} of partial isometries which satisfy the following relations, for all edges
e 6= f ∈ G1 and all vertices v ∈ G0 emitting a finite number of edges,
(1.1a)
(1.1b)
S∗
e Sf = 0,
S∗
e Se = P(e),
SeS∗
e ≤ Pσ(e),
Pv = Xe∈G1 : σ(e)=v
SeS∗
e .
Graph C ∗-algebras include important classes of operator algebras, such as the Cuntz-
Krieger algebras [14] or AF algebras. Significant advantage of working with graph C ∗-
algebras stems from the ease with which one can calculate their K-theory and primitive
ideal spectrum. This feature of graph C ∗-algebras has been widely exploited in their
applications to the classification programme of general C ∗-algebras (for example, see
[30], [28], [16] or [4]). In addition, they have influenced recent developments in purely
algebraic ring theory, leading to the introduction of Leavitt path algebras [2] in an at-
tempt to explore their classification power beyond operator algebra theory; see [1] for
an illuminating review. More importantly from the point of view of the subject matter
of this text, algebras of continuous functions on quantum spheres and on deformations
of non-singular lens spaces can also be interpreted as graph C ∗-algebras [20], [21]. We
extend this interpretation to quantum deformations of all (weighted) lens spaces, in-
cluding those that are classically singular, and employ it to compute the K-theory of a
fairly general class of quantum weighted projective spaces and of particular examples of
(weighted, singular) quantum lens spaces.
The paper is organised as follows. In Section 2, we first recall the algebraic definition
of quantum lens spaces. The coordinate algebras of quantum lens spaces are defined as
fixed points of the weighted action of the cyclic group ZN on (the generators of) the
coordinate algebra of the quantum odd-dimensional sphere. We make no assumption on
the existence of common factors of weights and the order of the group. Next, using the
identification of the algebra of continuous functions on the quantum odd dimensional
sphere with the graph C ∗-algebra associated to a graph L2n+1 [20], we extend the cyclic
group action to the action on this algebra. The resulting fixed point algebra has been
shown in [21] to be a graph C ∗-algebra, provided all weights are coprime with the
order of the acting group. We extend this identification to all weights, thus relaxing
the coprimeness assumption. Similarly to [21] we construct a suitable graph LN ;m
2n+1 by
relating it to the skew product graph L2n+1 ×c ZN (where the labelling c is determined by
the weights), and using the result of Crisp [13] that fixed points of a finite group action
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
3
on a graph C ∗-algebra can be identified with specific corner of the algebra associated
to the skew product graph. The construction of LN ;m
2n+1 explores the values of weights
modulo the order of the cyclic group and thus heavily depends on them; we illustrate
this by a series of examples. The identification of the algebras of continuous functions on
quantum lens spaces as algebras associated to explicitly described graphs allows one for
more effective calculation of their K-groups, in particular the K0-groups; we illustrate
it be a series of examples too.
In Section 3, we study algebras of continuous functions on quantum weighted projec-
tive spaces WPn
q (m0, . . . , mn). On the algebraic level these are defined as fixed points of
weighted circle group actions on the quantum odd dimensional sphere, [9]. On the other
hand they can also be identified with a free or principal [11] action of the circle group
on quantum lens spaces such that all weights divide the order of the cyclic group, [10].
Both actions can be lifted to actions on continuous functions on quantum spheres and
lens spaces. The analysis of algebraic structure of quantum weighted projective spaces,
in particular of deriving generators and relations, is notoriously difficult, as one has to
deal not only with an increasing number of generators but also with relative divisibility
properties of the weights. Until now, even in the lowest dimensional case, n = 1, the
algebraic and operator algebraic structure of WP1
q(m0, m1) has been understood com-
In higher dimensions, the full list of
pletely only in the case of coprime weights [9].
(algebraic) generators of the coordinate algebra C(WPn
q (m0, . . . , mn)) is given in [15],
in the case the weights of the form mi = Qj6=i lj, where l0, . . . , ln are pairwise coprime
integers. Furthermore, such a list of generators is given in [10] provided all the weights
but the last one are equal to 1. We prove that the algebra of continuous functions on
WP1
q(m0, m1) is an AF graph C ∗-algebra, and compute its K-theory with no restric-
q(m0, m1)) does not
tions on the weights (Proposition 3.1).
depend on the actual values of the weights, but only on m1 divided by its greatest com-
mon divisor with m0, and hence, as a topological noncommutative space any quantum
projective line WP1
q(1, m) (with
m = m1/ gcd(m0, m1)). Finally we derive a short exact sequence which characterises
quantum weighted projective spaces with weights m0, . . . , mn−1 coprime with mn, and
use it to compute their K-theory in the case the weights have the property that for each
j ≥ 1 there is an i < j so that mi and mj are relatively prime.
q(m0, m1) is isomorphic to the quantum teardrop WP1
It turns out that C(WP1
2. Quantum weighted lens spaces as graph C ∗-algebras
In this section we prove that noncommutative algebras of continuous functions on all
quantum (weighted) lens spaces are graph C ∗-algebras.
The algebra of continuous functions on the quantum odd-dimensional sphere C(S2n+1
)
is defined as the universal C ∗-algebra with generators z0, z1, . . . , zn, subject to the fol-
lowing relations:
q
(2.1a)
(2.1b)
zizj = qzjzi
for i < j,
ziz∗
i = z∗
i zi + (q−2 − 1)
ziz∗
j = qz∗
zjz∗
j ,
n
Xj=i+1
j zi
n
Xj=0
for i 6= j,
zjz∗
j = 1,
4
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
where q is a real number, q ∈ (0, 1); see [27]. As explained in [20], the algebra C(S2n+1
)
can be interpreted as a C ∗-algebra associated to a graph L2n+1 defined as follows. L2n+1
has n + 1 vertices v0, v1, . . . , vn, and (n + 1)(n + 2)/2 edges eij, i = 0, . . . , n, j = i, . . . , n,
with vi the source and vj the range of eij.
q
Let us fix a sequence of positive integers m := m0, . . . , mn. For any natural number
N, C(S2n+1
q
) admits the action of the cyclic group ZN defined by
(2.2)
N
m : zi 7→ ζ mizi,
where ζ is a generator of ZN . Under the isomorphism C(S2n+1
takes the form
q
) ∼= C ∗(L2n+1), this action
(2.3)
N
m : Seij 7→ ζ miSeij , N
m : Pvi 7→ Pvi.
The fixed points of this action form the algebra of continuous functions on the quantum
lens space C(L2n+1
(N; m)) is a graph C ∗-
algebra provided all the mi are coprime with N. We extend this result to the general
case with arbitrary weight vector m, below.
It is shown in [21] that C(L2n+1
(N; m)).
q
q
q
As a matter of fact, C(L2n+1
(N; m)) is isomorphic to the full corner of the graph
C ∗-algebra associated to the skew product graph L2n+1 ×c ZN , where the labelling c
is induced from the ZN -action N
m , namely c : eij 7→ mi mod N. More explicitly, the
graph L2n+1 ×c ZN has vertices (vi, r), i = 0, . . . , n, r = 0, . . . , N − 1 and edges (eij, r),
i, j = 0, . . . n, i ≤ j, r = 0, . . . N − 1, with (vi, r − mi mod N) being the source and
(vj, r) being the range of (eij, r). For example, the skew product graph corresponding
to n = 1, N = 6, m0 = 1, m1 = 3, comes out as
(v0,0)
(v0,1)
(v0,2)
(v0,3)
(v0,4)
(v0,5)
(v1,0)
(v1,1)
(v1,2)
(v1,3)
(v1,4)
(v1,5)
(2.4)
i and the range vs
2n+1 has vertices vr
Based on the skew product graph L2n+1 ×c ZN we construct a graph LN ;m
2n+1 in the
following way. LN ;m
i , i = 0, . . . , n, r = 0, . . . , gcd(N, mi) − 1, and edges
j , and labelled additionally by a = 1, . . . , nrs
ers
ij;a with the source vr
ij ,
where nrs
ij is a number of paths in L2n+1 ×c ZN from (vi, r) to (vj, s) that do not pass
through (vk, t), with k = i + 1, . . . , j − 1 and t = 0, . . . , gcd(mk, N) − 1; such paths are
termed admissible. Since there are no edges (eij, r) in L2n+1 ×c ZN with i > j, one may
assume that i ≤ j in ers
ij;a. We refer to the index i as labelling the levels, and to index r
as labelling the loops in LN ;m
2n+1.
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
5
It is helpful to analyse the graphs L2n+1 ×c ZN and LN ;m
(2.5)
ci = gcd(N, mi),
di =
2n+1 more closely. Define
N
ci
,
and observe that di is coprime with mi/ci. The admissible paths from (vi, r) to (vj, s)
with r ∈ {0, 1, . . . , ci − 1}, s ∈ {0, 1, . . . , cj − 1} have the form
(eii1, r + mi)(ei1i2, r1) . . . (eikik+1, rk)(eik+1j, s),
t
rt = r + mi +
mia ≥ cit,
t = 1, . . . , k,
rk + mik+1 = s
(2.6)
where
(2.7)
Xa=1
(all sums are computed modulo N), and no vertex appears twice as the range (or,
equivalently, source) of any of the edges that compose into the path (2.6). There are no
admissible paths between (vi, r) and (vi, s) if r 6= s. Indeed if there were such a path,
then there would exist integers a, b such that
(2.8)
r − s = ami − bN.
The right hand side is divisible by ci, while the left hand side is not, since r − s < ci,
which gives the desired contradiction. On the other hand, there is exactly one path
connecting (vi, r) with itself:
(2.9)
(eii, r + mi)(eii, r + 2mi) . . . (eii, r + (di − 1)mi)(eii, r).
To see that (2.9) is admissible, first note that
r + dimi = r +
N
ci
mi = r +
mi
ci
N ≡ r mod N,
so that the condition in (2.7) is satisfied. Furthermore there are no edges in (2.9) with
the same source. Otherwise, there would need to exist integers a, b ∈ 0, . . . , di − 1 such
that r + ami = r + bmi modulo N. This would imply the existence of c ∈ Z such that
(a − b)mi = cN or, when both sides are divided by ci,
(2.10)
(a − b)
mi
ci
= cdi.
Since a − b < di and di is coprime with mi/ci, the left hand side is not divisible by di,
which gives the required contradiction. By the same token, (2.9) is the shortest path
connecting (vi, r) with itself. Any longer path would need to pass through the same
vertex at least twice, hence it would not be admissible. This proves the uniqueness.
Thus, If i = j, nrs
2n+1 and
there are no links between vertices vr
ii = δrs, i.e. there is a single loop attached to each vertex in LN ;m
If all the mi are coprime with N, the graph LN ;m
2n+1 coincides with the graph described
in [21, p. 257 -- 258]. At the other extreme, i.e. if all the mi divide N, then the graph
LN ;m
2n+1 consists of n + 1 levels of interconnected loops with mi mutually disconnected
loops at the i-th level.
Example 2.1. (1) Lkl;1,l
consists of one loop at level 0 and l-loops at level 1 with k links
connecting the loop in level 0 with each of the loops at level 1, so the corresponding
graph is:
3
i and vs
i if r 6= s.
6
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
v0
0
v0
1
(k)
(k)
(k)
...
v1
1
vl−1
1
,
(2.11)
where the labels in brackets over the straight arrows indicate their multiplicities.
(2) Lkl;1,1,l
5
consists of one loop at level 0, one loop at level 1 and l-loops at level 2
with the numbers of edges connecting different levels given by
2
hence the corresponding graph comes out as:
n00
01 = kl,
n0r
02 =
kl(k + 1)
− rk,
n0r
12 = k,
v0
0
(kl)
v0
1
(cid:16)kl(k+1)
2 (cid:17)
v0
2
(k)
(cid:16)k(lk+l−2)
(cid:17)
2
(k)
v1
2
(k)
vl−1
2
(cid:16)k(lk−l+2)
(cid:17)
...
2
(2.12)
(3) Lkl;1,l,l
5
consists of one loop at level 0, l loops at level 1 and l loops at level 2 with
the numbers of edges connecting different levels given by
n0r
01 = k,
n0r
02 =
k(k + 1)
2
,
n0r
12 = k,
i.e.
(2.13)
(k)
(k)
v0
1
v0
2
(k)
2 (cid:17)
(cid:16) k(k+1)
v1
1
(k)
v1
2
v0
0
...
(k)
2 (cid:17) (cid:16) k(k+1)
(cid:16) k(k+1)
2 (cid:17)
...
vl−1
1
(k)
vl−1
2
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
7
(4) Lkl;1,1,1,l
7
consists of one loop each at levels 0, 1 and 2, and l-loops at level 3 with
the numbers of edges connecting different levels given by
01 = n00
n00
12 = kl,
n0r
13 =
kl(k + 1)
2
− rk,
n0r
23 = k,
n00
02 =
kl(kl + 1)
2
,
n0r
03 =
kl(k + 1)
12
(2kl + l + 3) +
kr(r − 1)
2
−
kl(k + 1)
2
r.
The main result of this section is contained in the following
Theorem 2.2. As C ∗-algebras,
Proof.
Since C(L2n+1
q
C ∗(LN ;m
2n+1) ∼= C(L2n+1
q
(N; m)).
(N; m)) is obtained as fixed points of a finite abelian group
action on a graph C ∗-algebra, [13, Theorem 4.6] implies that it is isomorphic to
P(vi,0)! C ∗(L2n+1 ×c ZN ) n
Xi=0
P(vi,0)! .
Thus it suffices to prove that the following map
n
Xi=0
2n+1) → n
Xi=0
ψ : C ∗(LN ;m
P(vi,0)! C ∗(L2n+1 ×c ZN ) n
Xi=0
P(vi,0)! ,
given by
Pvr
i
7→ P(vi,r),
i = 0, . . . , n,
r = 0, 1, . . . , ci − 1,
and, for all admissible paths
α = (eii1, r + mi)(ei1i2, r1) . . . (eikik+1, rk)(eik+1j, s),
Sα 7→ S(eii1 ,r+mi)S(ei1 i2 ,r1) · · · S(eik ik+1 ,rk)S(eik+1 j,s),
extends to a C ∗-algebra isomorphism.
In view of the universal property of the graph C ∗-algebra, to prove that ψ extends
to a ∗-homomorphism it suffices to check that the images of the Pvr
i and Sα under ψ
satisfy relations (1.1) for LN ;m
2n+1. Conditions (1.1a) are obvious. To prove (1.1b), we fix
(vi, r) ∈ L2n+1 ×c ZN , and, for any ν ∈ N, split the set of all admissible paths from (vi, r)
into the subsets Aν of those of length less than n and Bν of those of length exactly ν.
We will prove by induction on ν that
(2.14)
P(vi,r) = Xα∈Aν
SαS∗
α + Xβ∈Bν
SβS∗
β.
The equation (2.14) obviously holds if ν = 1. Now, suppose that it also holds for some
(other) ν, and let
β = (eii1, r + mi)(ei1i2, r1) . . . (eikik+1, rk)(eik+1j, s) ∈ Bν.
Using (1.1b) at the range vertex (vj, s) of β one easily finds that
(2.15)
SβS∗
β =
n
Xl=s
SβS(ejl,s+mj)S∗
(ejl,s+mj)S∗
β.
8
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
All the paths
β′ = (eii1, r + mi)(ei1i2, r1) . . . (eikik+1, rk)(eik+1j, s)(ejl, s + mj)
that appear in (2.15) are admissible, which is obvious in case l 6= s. Otherwise, this
follows by the observation that no segment of an admissible path that is not a loop
can have length greater than di − 1 at any given level i (otherwise the path would pass
through the same vertex at least twice); if an edge with both the source and range at the
level i is added its range will be a vertex that is not a range for any edge yet, otherwise
one is led to contradiction as in (2.10). Therefore, β′ ∈ Aν+1 \ Aν or β′ ∈ Bν+1 and,
using the inductive hypothesis, we obtain
SαS∗
SαS∗
P(vi,r) = Xα∈Aν
= Xα∈Aν
= Xα∈Aν+1
SβS∗
β
α + Xβ∈Bν
α + Xβ ′∈Bν+1
α + Xβ∈Bν+1
SαS∗
SβS∗
β.
Sβ ′S∗
β ′ + Xβ ′∈Aν+1\Aν
Sβ ′S∗
β ′
By the principle of mathematical induction, (2.14) holds for all natural ν.
Since L2n+1 ×c ZN is a finite graph there exists ν such that Bν is an empty set, and
hence relation (1.1b) for P(vi,r) = ψ(Pvr
i ) is equivalent to (2.14) for this ν. This proves
that ψ extends to a ∗-homomorphism, which we still denote ψ. To prove that ψ is
injective, we apply the general Cuntz-Krieger uniqueness theorem, [29, Theorem 1.2].
Clearly, ψ(P(vi,r)) 6= 0 for all vertices vr
i . Also, the only loops without exits in graph
LN ;m
nn;1 attached to vertices at level n. For all r = 0, . . . cn − 1,
we have
2n+1 are the edges er
n := err
ψ(Ser
n) = (enn, r + mn)(enn, r + 2mn) . . . (enn, r + (dn − 1)mn)(enn, r).
It follows from the last part of the proof of Theorem 2.4 in [22] that this is a partial
unitary with full spectrum. Thus the hypothesis of [29, Theorem 1.2] holds and ψ is
injective.
Finally, we need to prove that ψ is surjective. First, let us consider a path α with
source and range both in the set {(vi, 0) i = 0, . . . , n}. Since every loop in the crossed-
product graph L2n+1 ×c ZN passes through one of the vertices (vi, ri), ri = 0, . . . ci − 1,
i = 0, . . . , n, the path α is a concatenation of admissible paths, i.e. α = α1α2 . . . αk,
with all the αk admissible. Therefore Sα = Sα1Sα2 · · · Sαk is in the image of C ∗(LN ;m
2n+1)
under ψ. ⊔⊓
Corollary 2.3. The following sequence of C ∗-algebras
(2.16)
0
/ (K ⊗ C(T))⊕ gcd(mn,N )
/ C(L2n+1
q
(N; m))
/ C(L2n−1
q
(N; m))
/ 0,
where K denotes compact operators on a separable Hilbert space, is exact.
Proof. Each level n vertex in graph LN ;m
2n+1 emits exactly one edge, to itself. On the
other hand, there exist infinitely many paths from vertices in other levels to each vertex
in level n. Thus the closed two-sided ideal of C ∗(LN ;m
2n+1) generated by projections Pvi
n,
/
/
/
/
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
9
i = 0, . . . , gcd(mn, N) − 1, is isomorphic to (K ⊗ C(T))⊕ gcd(mn,N ), [22] and [7], and the
corresponding quotient is isomorphic to C ∗(LN ;m
2n−1), [7]. Thus the exactness of (2.16)
follows from Theorem 2.2. ⊔⊓
The identification of the algebra of continuous functions on the quantum lens space
2n+1) allows one to design a method for computing
with the graph C ∗-algebra C ∗(LN ;m
K-groups of C(L2n+1
(N; m)). More precisely,
q
K0(C(L2n+1
q
(N; m))) = coker Φ,
K1(C(L2n+1
q
(N; m))) = ker Φ,
where Φ is the endomorphism of a free abelian group with generators vr
i given by
Φ(vr
i ) =Xj,s
(nrs
ij − δijδrs)vs
j ;
see [24, Theorem 3.2]. The complete computation of K1(C(L2n+1
(N; m))) is presented in
[10, Proposition 5.2], the more difficult computation of K0(C(L2n+1
(N; m))) boils down
to detailed analysis of numbers of links connecting various loops and then to derive the
Smith normal form of the matrix corresponding to the transformation Φ. Recall [26]
that every integer matrix A can be reduced (by row and column operations) to the
diagonal form with entries 0 or α1, . . . , αn, where
q
q
(2.17)
α1 = ∆1,
αi+1 =
∆i+1
∆i
,
where the ∆i are greatest common divisiors of all minors of A of size i. The torsion part
of the cokernel of the transformation defined by A is then
Zα1 ⊕ Zα2 ⊕ . . . ⊕ Zαn.
Below we give three examples of this.
Example 2.4.
K0(C(L5
q(kl; 1, 1, l))) = Zl ⊕(Zk ⊕ Zk
Z2k ⊕ Z k
2
if k is odd or l is even,
if k is even and l is odd.
Proof. The transformation Φ is determined by the integer (l + 2) × (l + 2)-matrix
with the first two columns
,
0
0
k
k
k
. . .
k
and all other entries 0. By simple row and column operations this matrix can be reduced
to the block diagonal form
0
kl
kl(k + 1)/2
kl(k + 1)/2 − k
kl(k + 1)/2 − 2k
. . .
kl(k + 1)/2 − k(l − 1)
k
0 0
kl(k − 1)/2 k 0
0
0 0
.
10
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
The right bottom corner gives the infinite part of K0(C(L5
q(kl; 1, 1, l))). The greatest
common divisor of the minor of size 2 is ∆2 = k2, while the greatest common divisor of
one-dimensional minors depends on the parity of k and l,
∆1 =(k
k
2
if k is odd or l is even,
if k is even and l is odd
.
In view of (2.17), this yields the stated finite part of K0(C(L5
q(kl; 1, 1, l))). ⊔⊓
Example 2.5.
K0(C(L5
q(kl; 1, l, l))) = Zl ⊕(Zl+1
k
Z2k ⊕ Z k
2
if k is odd,
if k is even.
⊕ Zl−1
k
Proof. The transformation Φ is determined by the integer (2l + 1) × (2l + 1)-matrix
with the first l + 1 columns
0
k
k
. . .
k
. . .
k(k + 1)/2
k(k + 1)/2
k(k + 1)/2
0
0
0
. . .
0
k
0
. . .
0
0
0
0
. . .
0
0
k
. . .
0
. . . 0
. . . 0
. . . 0
. . . 0
. . . 0
. . . 0
. . . 0
. . . 0
. . . k
,
and all other entries 0. By subtracting the l+1-st row from rows 2 to l, the first l rows can
be reduced to the zeros. This gives the infinite part of the group K0(C(L5
q(kl; 1, l, l))).
The greatest common divisors of the minors in the remaining matrix come out as
∆i =(ki
ki
2
if k is odd,
if k is even
,
i = 1, 2, . . . , l,
and ∆l+1 = kl+1. In view of (2.17), this yields the finite part of K0(C(L5
stated. ⊔⊓
q(kl; 1, l, l))) as
Example 2.6. Let
(2.18)
Then
α := n00
13 =
kl(k + 1)
2
,
β := n00
03 = α
2kl + l + 3
6
.
(2.19) K0(C(L7
q(kl; 1, 1, 1, l))) = Zl ⊕
Zk ⊕ Zk ⊕ Zk
⊕ Zk ⊕ Z6k
Z k
6
⊕ Zk ⊕ Z3k
Z k
3
⊕ Zk ⊕ Z2k
Z k
2
⊕ Z4k
⊕ Z k
Z k
2
⊕ Z12k
Z k
6
⊕ Z k
2
2
if kα & kβ,
if kα & β ≡ k
if kα & β ≡ k
if kα & β ≡ k
if k 6 α & k
2 β,
if k 6 α & k
2 6 β.
6 (mod k),
3 (mod k),
6 , 5k
3 , 2k
2 (mod k),
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
11
Proof. Let us first observe that
n0r
13 = α − kr,
n0r
03 = β − αr +
r(r − 1)
2
k.
Therefore, the matrix representing Φ is an (l + 3) × (l + 3)-matrix with the non-zero
entries contained in the first three columns
0
kl
kl(kl+1)
2
β
β − α
β − 2α + k
. . .
0
0
kl
α
α − k
α − 2k
. . .
k α − (l − 1)k
.
0
0
0
k
k
k
. . .
k
β − (l − 1)α + (l−1)(l−2)
2
By elementary row and column operations (starting with subtracting row 4 from all
subsequent rows) and using the divisibility of the entries by k, we arrive at the block
diagonal matrix
k 0 0 0
α k 0 0
β α k 0
0 0 0 0
.
The zero l × l-matrix in the bottom-right corner gives Zl as the infinite part of the group
K0(C(L7
q(kl; 1, 1, 1, l))), while the 3 × 3-matrix in the top-left corner gives the finite part
of K0(C(L7
q(kl; 1, 1, 1, l))). Its nature depends on the divisbiliity properties of α and β
and it splits into two parts. If α is divisible by k, then the matrix can be reduced to
The greatest common divisors of the minors thus read
k 0 0
0 k 0
β 0 k
.
∆1 = gcd(k, β),
∆2 = gcd(k2, kβ) = k∆1,
∆3 = k3,
thus yielding the diagonal entries:
α1 = gcd(k, β),
α2 = k,
α3 =
k2
gcd(k, β)
.
If k does not divide α, then it must be even, and α is divisible by k/2, thus leading to
the matrix
k 0 0
k
2 k 0
β k
2 k
,
and the corresponding Smith normal form entries
α1 = gcd(
k
2
, β),
α2 =
k
2
,
α3 =
2k2
gcd( k
2 , β)
.
12
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
Let us note that
β =
k(k + 1)(k + 2)l(l + 1)
12
+
(k − 1)k(k + 1)l(l − 1)
12
.
Hence β modulo k has to be a multiple of the sixth of k, and, in the case of even k, β
modulo k/2 has to be a multiple of the third of k/2. The analysis of all these possibilities
yields the stated form of K0(C(L7
q(kl; 1, 1, 1, l))). ⊔⊓
Remark 2.7. If l = 1, the results of Example 2.4 agree with that of [21, Proposition 2.3].
On the other hand, if l = 1, then numbers α and β defined in (2.18) come out as
α =
k(k + 1)
2
,
β = α
k + 2
3
,
and the second and fourth cases in (2.19) cannot occur. The remaining cases coincide
with the K-groups computed in [5, Example 6.6].
3. K-theory of quantum weighted projective spaces
The aim of this section is to calculate K-theory of a fairly general class of quan-
tum weighted projective spaces and to give a complete description of C ∗-algebras of
continuous functions on quantum weighted projective lines as graph AF-algebras.
As before, we fix a sequence of positive integers m := m0, . . . , mn. In addition to the
ZN -action (2.2), the algebra C(S2n+1
q
) admits the circle group action m,
(3.1)
m : zi 7→ ξmizi,
i = 0, . . . , n,
where ξ is the unitary generator of T (of infinite order). Fixed points C(WPn
q (m))
form the algebra of continuous functions on the quantum weighted projective space, [9].
) that transform
As explained in [10], for a fixed N, all the elements Pi xi of C(S2n+1
according to the rule
q
Xi
xi 7→Xi
ξriN xi,
ri ∈ Z,
q
) isomorphic to C(L2n+1
form a subalgebra of C(S2n+1
to the T-action m on C(L2n+1
element x ∈ C(L2n+1
x 7→ ξrN x under m. The actions N
via the short exact sequence of abelian groups
(N; m)). The action m gives rise
q (m)): an
(N; m)) transforms under m as x 7→ ξrx provided it transforms as
m and m can be uderstood as being derived from m
(N; m)) with fixed points being again C(WPn
q
q
q
1
/ T
/ T
/ ZN
/ 1,
where the (non-trivial) monomorphism is ξ 7→ ξN and the (non-trivial) epimorphism is
ξ 7→ ζ; see e.g. [23, Section A.1.1].
We describe C ∗-algebras of the quantum weighted projective spaces WP1
q(m), m =
(m0, m1), as AF graph algebras. Let g := gcd(m0, m1), m0 := m0/g and m1 := m1/g,
and define a graph W1(m) as follows. The graph has m1+1 vertices, denoted w0, . . . , w m1.
/
/
/
/
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
13
For each j ∈ {1, . . . , w m1} there are infinitely many edges from w0 to wj, denoted fjk,
k ∈ N, i.e.
w0
(∞) (∞)
(∞)
w1
w2
...
w m1
(3.2)
Proposition 3.1. For all values of m = (m0, m1), C(WP1
morphic to the graph C ∗-algebra C ∗(W1(m)). Consequently,
q(m)) is an AF-algebra iso-
K0(C(WP1
q(m))) = Z1+m1/ gcd(m0,m1),
K1(C(WP1
q(m))) = 0.
Proof. As explained above, C(WP1
q(m)) is isomorphic to the C ∗-algebra of fixed points
for the generalized gauge action of the circle group T on the graph algebra C ∗(L3), such
that
m(Seij ) = ξmiSeij
for i = 0, 1, j = i, 1.
We denote this fixed point algebra C ∗(L3)m and we construct a C ∗-algebra isomorphism
φ : C ∗(W1(m)) → C ∗(L3)m. At first we find targets for the generators of C ∗(W1(m))
inside C ∗(L3). Let
m1
Xj=2
φ(Pw0) := Pv0 −
e00 Se01S∗
Sj−2
e01(S∗
e00)j−2,
φ(Pw1) := Pv1,
φ(Pwj ) := Sj−2
e00 Se01S∗
φ(Sf1k) := S m1(k+1)−1
φ(Sfjk) := S m1(k+1)+j−2
e00
e00
e00)j−2,
e01(S∗
Se01(S∗
e11) m0(k+1),
Se01(S∗
e11) m0(k+1)S∗
for j = 2, . . . , m1,
for k ∈ N,
e01(S∗
e00)j−2,
for k ∈ N, j = 2, . . . , m1.
Clearly, these elements of C ∗(L3) are m-invariant and satisfy the defining relations
for the graph algebra C ∗(W1(m)). Thus, this assignment extends uniquely to a ∗-
homomorphism φ : C ∗(W1(m)) → C ∗(L3)m. Injectivity of φ follows from [29, Theorem
1.2], since there are no closed paths in graph W3(m) and φ(Pwj ) 6= 0 for all j = 0, . . . , m1.
It remains to verify that the map φ is surjective. First of all, C ∗(L3) is a closed span
of elements of the form SαS∗
β, where α and β are two paths with the common range.
The action m rescales each such an element by a suitable power of ξ. Applying the
conditional expectation from C ∗(L3) onto C ∗(L3)m (integration over the orbits), we see
that the fixed point algebra C ∗(L3)m is spanned by those elements SαS∗
β which are fixed
by m. Hence it suffices to show that all such elements are in the range of φ.
If both α and β end at v0, then we must have α = β and thus SαS∗
β = Pv0 =
j=2 φ(Pwj ). So suppose that α and β end at v1. If both α and β contain
β = Pv1 = φ(Pw1). So we
11 for some k, r ∈ N. Now, if β does not contain edge e01,
11 for some s ∈ N and we must have m0(k + 1) + m1r = m1s. This can only
β = Sα′S∗
β ′
only edges e11, then again we must have α = β and thus SαS∗
may assume that α = ek
then β = es
happen when k = t m1 − 1 and s = r + t m0 for some t ∈ N \ {0}. Since SαS∗
with α′ = ek
φ(Pw0) +P m1
00e01 and β′ = es−r
11 , we get SαS∗
β = φ(Sf1(t−1)) in this case.
00e01er
14
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
00e01ep
11 and β = el
It remains to consider the case α = ek
11 for some k, l, p, s ∈ N.
As above, from the start we may assume that p = 0. We must have m0(k + 1) = m0(l +
1)+m1s, and hence m0(k−l) = m1s. If k = l, then s = 0 and SαS∗
e00)k.
β = Sk
Write k = r m1 + t with r ∈ N and t ∈ {0, . . . , m1 − 1}. Then SαS∗
β equals: (i) φ(Pwt+2) if
r = 0 and t < m1 −1, (ii) φ(Sf(t+2)(r−1)S∗
f1r )
if t = m1 − 1. Note that this argument shows that for each path α in graph L3 there
exists an edge (or vertex) f in graph W1(m) such that φ(Sf ) = Pα.
) if r > 0 and t < m1 −1, (iii) φ(Sf1r S∗
e00Se01S∗
00e01es
e01(S∗
f(t+2)(r−1)
e11)t m0S∗
Finally, suppose that k 6= l. It suffices to consider the case k − l > 0, when also s > 0.
Then we must have s = t m0 and k − l = t m1 for some t ∈ N \ {0}, and thus SαSβ =
e01(S∗
e00)l. Let f, h be edges (or possibly vertices) in graph W1(m)
e00 Se01(S∗
Sl+t m1
such that φ(f ) = Sl+t m1
e00)l.
e00 Se01S∗
Since (l + t m1) − l is a multiple of m1, edges f and h have a common range. Then it is
e00)l,
a bit tedious but not difficult to verify that φ(f )φ(h)∗ = Sl+t m1
and this completes the proof of surjectivity of φ.
e00)l+t m1 and φ(h) = Sl
e00 Se01(S∗
e00Se01St m0
e11)t m0S∗
e11)t m0S∗
e11 (S∗
e01(S∗
e01(S∗
e01(S∗
As an immediate corollary of the isomorphism C(WP1
q(m)) ∼= C ∗(W1(m)), we obtain
the following exact sequence:
0
/ Km1/ gcd(m0,m1)
/ C(WP1
q(m))
/ C
/ 0.
It follows that C(WP1
q(m)) is an AF algebra and
q(m))) = Z1+m1/ gcd(m0,m1),
K0(C(WP1
K1(C(WP1
q(m))) = 0,
as required. ⊔⊓
Poposition 3.1 contains full classification of algebras of continuous functions on the
quantum weighted projective line: as a topological noncommutative space the quantum
projective line WP1
q(1, m), where
m = m1/ gcd(m0, m1).
q(m0, m1) is isomorphic to the quantum teardrop WP1
Proposition 3.2. Let m := m0, . . . , mn be positive integers such that there exists j ∈
{0, 1, . . . , n − 1} so that mj is relatively prime with mn. Then there exists an exact
sequence
(3.3)
0
/ Kmn
/ C(WPn
q (m))
/ C(WPn−1
q
(m))
/ 0.
Proof. We use the identification C(WPn
q (m)) ∼= C ∗(L2n+1)m. Let J be the closed
span of all SαS∗
β ∈ C ∗(L2n+1)m such that α, β are paths in L2n+1 with both α and β
ending at vertex mn. Then J is a closed, two-sided ideal of C ∗(L2n+1)m such that the
quotient C ∗(L2n+1)m/J is isomorphic to C ∗(L2n−1)m ∼= C(WPn−1
(m)). Thus it suffices
to show that J ∼= Kmn.
q
α and SβS∗
For each k ∈ {0, 1, . . . , mn − 1} let Jk be the closed, two-sided ideal of J generated
α such that α is a path in L2n+1 ending at vk and m(Sα) = ξlSα
by all projections SαS∗
∼= K and JkJr = {0} for k 6= r. Indeed, let
with l ≡ k (mod mn). We claim that Jk
SαS∗
β be two projections in Jk, as above. Then for a suitable integer t the
element SαSt
α. On
the other hand, if SαS∗
β ∈ Jr with k not congruent to r modulo mn,
then these two projections are not equivalent in J, since there is no t ∈ Z for which
SαSt
β is in the fixed point algebra C ∗(L2n+1)m. It remains to show that Jk 6= {0}
β is a partial isometry in Jk with domain SβS∗
β and range SαS∗
α ∈ Jk and SβS∗
mnS∗
kS∗
/
/
/
/
/
/
/
/
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
15
for each k. Let j < mn be such that mj and mn are relatively prime. Consider path
α = et
jjej(j+1)e(j+1)(j+2) . . . e(mn−1)mn. Since mj and mn are relatively prime, for each k
i=j mi is congruent to k modulo mn.
we can find a positive integer t such that tmj +Pmn−1
⊔⊓
Corollary 3.3. Let m := m0, . . . , mn be a sequence of positive integers such that for
each j ≥ 1 there is an i < j so that mi and mj are relatively prime. Then
K0(C(WPn
q (m))) = Z1+Pn
i=1 mi,
K1(C(WPn
q (m))) = 0.
Proof. We proceed by induction on n. Case n = 1 being contained in Proposition 3.1.
Applying the K-functor to (3.3) we obtain the six-term exact sequence
K0(Kmn)
/ K0(C(WPn
q (m)))
/ K0(C(WPn−1
q
(m)))
K1(C(WPn−1
q
(m)))
K1(C(WPn
q (m)))
K1(Kmn) .
Since outer terms in the bottom row vanish (the left one by inductive assumption) also
the middle term is 0, as required. Thus again using the inductive assumption and the
K-theory of compact operators we obtain a short exact sequence
0
/ Zmn
/ K0(C(WPn
q (m)))
/ Z1+Pn−1
i=1 mi
/ 0,
which splits as a sequence of abelian groups thus confirming the stated form of K-groups
of quantum weighted projective spaces. ⊔⊓
Acknowledgments
The first named author would like to extend his warmest thanks to the members
of IMADA, University of Southern Denmark, Odense, where the work on this paper
was partly carried out in May and December 2015. The research of the second named
author was partially supported by the FNU Project Grant 'Operator algebras, dynamical
systems and quantum information theory' (2013 -- 2015), the Villum Fonden Research
Grant 'Local and global structures of groups and their algebras' (2014 -- 2018), and by
the Mittag-Leffler Institute during his stay there in January-February, 2016.
References
[1] G. Abrams, Leavitt path algebras: the first decade, Bull. Math. Sci. 5 (2015), 59 -- 120.
[2] G. Abrams & G. Aranda Pino, The Leavitt path algebra of a graph, J. Algebra 293 (2005), 319 -- 334.
[3] A. Adem, J. Leida & Y. Ruan, Orbifolds and Stringy Topology. Cambridge University Press,
Cambridge (2007).
[4] P. Ara & R. Exel, Dynamical systems associated to separated graphs, graph algebras, and paradox-
ical decompositions, Adv. Math. 252 (2014), 748 -- 804.
[5] F. Arici, S. Brain & G. Landi, The Gysin sequence for quantum lens spaces, J. Noncommut. Geom.
9 (2015), 1077 -- 1111.
[6] F. Arici, F. D'Andrea & G. Landi, Pimsner algebras and noncommutative circle bundles,
arXiv:1506.03109 (2015), to appear in Noncommutative Analysis, Operator Theory and Appli-
cations (Birkhauser, Basel).
/
/
O
O
o
o
o
o
/
/
/
/
16
TOMASZ BRZEZI ´NSKI AND WOJCIECH SZYMA ´NSKI
[7] T. Bates, D. Pask, I. Raeburn & W. Szyma´nski, The C ∗-algebras of row-finite graphs, New York
J. Math. 6 (2000), 307 -- 324.
[8] T. Brzezi´nski, On the smoothness of the noncommutative pillow and quantum teardrops, SIGMA
10 (2014), 015.
[9] T. Brzezi´nski & S. A. Fairfax, Quantum teardrops, Commun. Math. Phys. 316 (2012), 151 -- 170.
[10] T. Brzezi´nski & S. A. Fairfax, Notes on quantum weighted projective spaces and multidimensional
teardrops, J. Geom. Phys. 93 (2015), 1 -- 10.
[11] T. Brzezi´nski & P. M. Hajac, The Chern-Galois character, C. R. Math. Acad. Sci. Paris 338
(2004), 113 -- 116.
[12] T. Brzezi´nski & A. Sitarz, Smooth geometry of the noncommutative pillow, cones and lens spaces,
arXiv:1410.6587 (2014), to appear in J. Noncomm. Geom.
[13] T. Crisp, Corners of graph algebras, J. Operator Theory, 60 (2008), 101 -- 119.
[14] J. Cuntz & W. Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56
(1980), 251 -- 268.
[15] F. D'Andrea & G. Landi, Quantum weighted projective and lens spaces, Commun. Math. Phys.
340 (2015), 325 -- 353.
[16] S. Eilers & M. Tomforde, On the classification of nonsimple graph C ∗-algebras, Math. Ann. 346
(2010), 393 -- 418.
[17] N. J. Fowler, M. Laca & I. Raeburn, The C ∗-algebras of infinite graphs, Proc. Amer. Math. Soc.,
128 (2000), 2391 -- 2327.
[18] A. J. Harju, Quantum orbifolds, arXiv:1412.4589 (2014).
[19] A. J. Harju, Dirac operators on quantum weighted projective spaces, Algebr. Represent. Theory 18
(2015), 1187 -- 1210.
[20] J. H. Hong & W. Szyma´nski, Quantum spheres and projective spaces as graph algebras, Commun.
Math. Phys. 232 (2002), 157 -- 188.
[21] J. H. Hong & W. Szyma´nski, Quantum lens spaces and graph algebras, Pacific J. Math. 211 (2003),
249 -- 263.
[22] A. Kumjian, D. Pask & I. Raeburn, Cuntz-Krieger algebras of directed graphs, Pacific J. Math.
184 (1998), 161 -- 174.
[23] C. Nastasescu & F. van Oystaeyen, Graded Ring Theory, North-Holland, Amsterdam-New York,
1982.
[24] I. Raeburn & W. Szyma´nski, Cuntz-Krieger algebras of infinite graphs and matrices, Trans. Amer.
Math. Soc. 338 (2004), 39 -- 59.
[25] A. Sitarz & J. J. Venselaar, The geometry of quantum lens spaces: real spectral triples and bundle
structure, Math. Phys. Anal. Geom. 18 (2015), no. 1, Art. 9, 19 pp.
[26] H. J. S. Smith, On systems of linear indeterminate equations and congruences, Phil. Trans. Royal
Soc. London, 151 (1861), 293 -- 326.
[27] Ya. S. Soibel'man & L. L. Vaksman, Algebra of functions on the quantum group SU(n+1), and
odd-dimensional quantum spheres. (Russian) Algebra i Analiz 2 (1990), 101 -- 120; translation in
Leningrad Math. J. 2 (1991), 1023 -- 1042.
[28] J. Spielberg, Semiprojectivity for certain purely infinite C ∗-algebras, Trans. Amer. Math. Soc. 361
(2009), 2805 -- 2830.
[29] W. Szyma´nski, General Cuntz-Krieger uniqueness theorem, Internat. J. Math. 13 (2002), 549 -- 555.
[30] W. Szyma´nski, The range of K-invariants for C ∗-algebras of infinite graphs, Indiana Univ. Math.
J. 51 (2002), 239 -- 249.
[31] W. P. Thurston, Three-dimensional geometry and topology. Vol. 1. Edited by Silvio Levy. Princeton
Mathematical Series, 35. Princeton University Press, Princeton, NJ, 1997.
QUANTUM LENS AND WEIGHTED PROJECTIVE SPACES
17
Department of Mathematics, Swansea University, Singleton Park, Swansea SA2 8PP,
U.K. & Department of Mathematics, University of Bia lystok, K. Cio lkowskiego 1M,
15 -- 245 Bia lystok, Poland
E-mail address: [email protected]
Department of Mathematics and Computer Science, University of Southern Den-
mark, Campusvej 55, 5230 Odense M, Denmark
E-mail address: [email protected]
|
1507.03296 | 4 | 1507 | 2016-03-14T22:12:04 | Amenability and covariant injectivity of locally compact quantum groups II | [
"math.OA",
"math.FA"
] | Building on our previous work, we study the non-relative homology of quantum group convolution algebras. Our main result establishes the equivalence of amenability of a locally compact quantum group $\mathbb{G}$ and 1-injectivity of $L^{\infty}(\widehat{\mathbb{G}})$ as an operator $L^1(\widehat{\mathbb{G}})$-module. In particular, a locally compact group $G$ is amenable if and only if its group von Neumann algebra $VN(G)$ is 1-injective as an operator module over the Fourier algebra $A(G)$. As an application, we provide a decomposability result for completely bounded $L^1(\widehat{\mathbb{G}})$-module maps on $L^{\infty}(\widehat{\mathbb{G}})$, and give a simplified proof that amenable discrete quantum groups have co-amenable compact duals which avoids the use of modular theory and the Powers--St{\o}rmer inequality, suggesting that our homological techniques may yield a new approach to the open problem of duality between amenability and co-amenability. | math.OA | math |
AMENABILITY AND COVARIANT INJECTIVITY OF LOCALLY
COMPACT QUANTUM GROUPS II
JASON CRANN
Abstract. Building on our previous work, we study the non-relative homol-
ogy of quantum group convolution algebras. Our main result establishes the
equivalence of amenability of a locally compact quantum group G and 1-
injectivity of L∞( bG) as an operator L1( bG)-module.
In particular, a locally
compact group G is amenable if and only if its group von Neumann algebra
V N (G) is 1-injective as an operator module over the Fourier algebra A(G). As
an application, we provide a decomposability result for completely bounded
L1( bG)-module maps on L∞( bG), and give a simplified proof that amenable dis-
crete quantum groups have co-amenable compact duals which avoids the use
of modular theory and the Powers -- Størmer inequality, suggesting that our ho-
mological techniques may yield a new approach to the open problem of duality
between amenability and co-amenability.
1. Introduction
The connection between amenability of a locally compact group G and injectivity
of its group von Neumann algebra V N (G) has been a topic of interest in abstract
harmonic analysis for decades. Amenability of G entails injectivity of V N (G) [14],
however, the converse is not true, e.g., if G = SL(n, R) for n ≥ 2; indeed, a result
of Connes' [4, Corollary 7], attributed to Dixmier, states that V N (G) is injective
for any separable connected locally compact group.
In [5], we clarified this connection by exploiting the T (L2(G))-module structure
of B(L2(G)), showing the equivalence of amenability of a locally compact group G
and covariant injectivity of V N (G), meaning the existence of a conditional expec-
tation E : B(L2(G)) → V N (G) commuting with the canonical T (L2(G))-action
[5, Theorem 4.2]. We also established a corresponding result at the level of locally
compact quantum groups G and studied the relationship between amenability of G
and relative 1-injectivity of its various operator modules over T (L2(G)).
In this paper we build on results from [5], focusing on the non-relative homology
of operator modules over L1(G) and T (L2(G)). Our main result states that a
group duality shows that in order to recover properties of G, one should not only
locally compact quantum group G is amenable if and only if L∞(bG) is 1-injective
as an operator L1(bG)-module. This new homological manifestation of quantum
consider the von Neumann algebraic structure of L∞(bG), but rather its operator
module structure. As an application, we provide a decomposability result in the
2010 Mathematics Subject Classification. Primary 22D35; Secondary 46M10, 46L89.
Key words and phrases. Locally compact quantum groups; Amenability; Injective Modules.
This work was completed as part of the author's doctoral thesis, and was supported by an
NSERC Canada Graduate Scholarship and a FCRF Joint PhD Scholarship. The author would
like to thank Matthias Neufang and Zhong-Jin Ruan for helpful discussions.
1
2
JASON CRANN
spirit of Haagerup [15] for completely bounded L1(bG)-module maps on L∞(bG).
Specifically, if G is amenable then
CBL1( bG)(L∞(bG)) = span CP L1( bG)(L∞(bG)).
Moreover, the proof of the above equality leads to a characterization of the pred-
ual Qr
cb(L1(G)) for an
arbitrary locally compact quantum group G as
cb(L1(G)) of the completely bounded (right) multipliers M r
which is new even in the group setting.
Qr
cb(L1(G)) ∼= C0(G)b⊗L1(G)L1(G),
Arguably the biggest open problem in abstract harmonic analysis on locally
compact quantum groups is the duality between amenability and co-amenability.
In the group setting, this is Leptin's theorem [29], which states that a locally
compact group G is amenable if and only if its Fourier algebra A(G) has a bounded
approximate identity. In the quantum group setting, many partial results have been
obtained over the years. Ruan showed that a compact Kac algebra G is co-amenable
if and only if its discrete dual bG is amenable [34, Theorem 4.5]. This equivalence
was later generalized by Tomatsu (and, independently by Blanchard and Vaes) to
arbitrary compact quantum groups [39, Theorem 3.8]. Tomatsu's argument relies
on the specific modular theory of discrete quantum groups in order to apply the
Powers -- Størmer inequality in a crucial step. As another application of our main
result, we give a considerably simplified proof of Tomatsu's theorem which avoids
the use of modular theory and the Powers -- Størmer inequality, suggesting that our
homological techniques may provide a new approach to the general duality problem
of amenability and co-amenability.
For regular quantum groups G, we obtain a version of our main result at the
predual level, showing the equivalence of discreteness of G and 1-projectivity of
L1(G) as an operator module over itself.
The paper is structured as follows. We begin in section 2 with some prelimi-
naries on the homology of operator modules, and include some new results on the
relationship between relative and non-relative homology. Section 3 is devoted to
a brief overview of the relevant machinery from locally compact quantum groups,
their associated operator modules, and completely bounded multipliers. Section 4
outlines the operator module structure of B(L2(G)) over T (L2(G)) and contains
new results which are used in the proof of the main theorem. Section 5 contains
the main result of the paper along with its aforementioned applications.
2. Preliminaries
Let A be a complete contractive Banach algebra. We say that an operator space
X is a right operator A-module if it is a right Banach A-module such that the
module map mX : Xb⊗A → X is completely contractive, where b⊗ denotes the
operator space projective tensor product. We say that X is faithful if for every
non-zero x ∈ X, there is a ∈ A such that x · a 6= 0, and we say that X is essential if
hX · Ai = X, where h·i denotes the closed linear span. We denote by mod − A the
category of right operator A-modules with morphisms given by completely bounded
module homomorphisms. Left operator A-modules and operator A-bimodules are
defined similarly, and we denote the respective categories by A − mod and A −
mod−A. If X ∈ mod−A is a dual operator space such that the action of a is weak*
continuous for all a ∈ A, then we say that X is a dual right operator A-module. We
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
3
let nmod−A denote the category of dual right operator A-modules with morphisms
given by weak*-weak* continuous completely bounded module homomorphisms,
and similarly for dual left operator A-modules.
Remark 2.1. Regarding terminology, in what follows we will often omit the term
"operator" when discussing homological properties of operator modules as we will
be working exclusively in the operator space category.
Let A be a completely contractive Banach algebra, X ∈ mod − A and Y ∈
A−mod. The A-module tensor product of X and Y is the quotient space Xb⊗AY :=
Xb⊗Y /N , where
N = hx · a ⊗ y − x ⊗ a · y x ∈ X, y ∈ Y, a ∈ Ai,
and, again, h·i denotes the closed linear span. It follows that
where CBA(X, Y ∗) denotes the space of completely bounded right A-module maps
CBA(X, Y ∗) ∼= N ⊥ ∼= (Xb⊗AY )∗,
isomorphism we say that X is an induced A-module. A similar definition applies
Φ : X → Y ∗. If Y = A, then clearly N ⊆ Ker(mX ) where mX : Xb⊗A → X is the
module map. If the induced mapping emX : Xb⊗AA → X is a completely isometric
for left modules. In particular, we say that A is self-induced if emA : Ab⊗AA ∼= A
Let A be a completely contractive Banach algebra and X ∈ mod − A. The
identification A+ = A ⊕1 C turns the unitization of A into a unital completely
contractive Banach algebra, and it follows that X becomes a right operator A+-
module via the extended action
completely isometrically.
x · (a + λe) = x · a + λx,
a ∈ A+, λ ∈ C, x ∈ X.
Let C ≥ 1. We say that X is relatively C-projective if there exists a morphism
module map m+
by the operator analogue of [7, Proposition 1.2]. We say that X is C-projective if
for every Y, Z ∈ mod − A, every complete quotient morphism Ψ : Y ։ Z, every
Φ+ : X → Xb⊗A+ satisfying kΦ+kcb ≤ C which is a right inverse to the extended
X : Xb⊗A+ → X. When X is essential, this is equivalent to the
existence of a morphism Φ : X → Xb⊗A satisfying kΦkcb ≤ C and mX ◦ Φ = idX
morphism Φ : X → Z, and every ε > 0, there exists a morphism eΦε : X → Y such
that keΦεkcb < CkΦkcb + ε and Ψ ◦eΦε = Φ.
When A = C, the definition of C-projectivity coincides with that of a C-
projective operator space [2, Definition 3.3]. The following relationship between
relative projectivity and projectivity appears to be new, and will be used to char-
acterize the 1-projectivity of quantum group convolution algebras.
Proposition 2.2. Let A be a completely contractive Banach algebra and let X ∈
mod − A.
If X is C1-projective in mod − C and is relatively C2-projective in
mod − A, then X is C1C2-projective in mod − A.
Proof. Let Y, Z ∈ mod − A, let Ψ : Y ։ Z be a complete quotient morphism and
Φ : X → Z be a morphism. By relative C2-projectivity, there exists a morphism
X ◦ α+ = idX and kα+kcb ≤ C2. Since X is a
C1-projective operator space, for every ε > 0, there exists a lifting Φε : X → Y
satisfying Ψ ◦ Φε = Φ and kΦεkcb < C1kΦkcb + ε/C2. The morphism (Φε ⊗ id) :
α+ : X → Xb⊗A+ satisfying m+
Xb⊗A+ → Y b⊗A+ then satisfies kΦε ⊗ idkcb < C1kΦkcb + ε/C2, and composing
4
JASON CRANN
with α+ together with the multiplication m+
Y : Y b⊗A+ → Y , we obtain a morphism
Y ◦ (Φε ⊗ id) ◦ α+ : X → Y satisfying keΦεkcb < C1C2kΦkcb + ε. Moreover,
using the module properties of the relevant morphisms we have
eΦε := m+
Y ◦ (Φε ⊗ id) ◦ α+
Z ◦ (Ψ ⊗ id) ◦ (Φε ⊗ id) ◦ α+
Z ◦ (Φ ⊗ id) ◦ α+
Ψ ◦eΦε = Ψ ◦ m+
= m+
= m+
= Φ ◦ m+
= Φ.
X ◦ α+
Hence, X is C1C2-projective.
(cid:3)
Note that the converse of Proposition 2.2 (when C1 = C2 = 1) is not true in
general as A is both 1-projective and relatively 1-projective in mod − A for any
unital C∗-algebra. However, the only C∗-algebra which is a 1-projective operator
space is C by [2, Theorem 3.4].
Given a completely contractive Banach algebra A and X ∈ mod − A, there is a
canonical completely contractive morphism ∆+ : X → CB(A+, X) given by
∆+(x)(a) = x · a,
x ∈ X, a ∈ A+,
where the right A-module structure on CB(A+, X) is defined by
(Ψ · a)(b) = Ψ(ab),
a ∈ A, Ψ ∈ CB(A+, X), b ∈ A+.
An analogous construction exists for objects in A − mod. Let C ≥ 1. We say that
X is relatively C-injective if there exists a morphism Φ+ : CB(A+, X) → X such
that Φ+ ◦ ∆+ = idX and kΦ+kcb ≤ C. When X is faithful, this is equivalent to the
existence a morphism Φ : CB(A, X) → X such that Φ ◦ ∆ = idX and kΦkcb ≤ C
by the operator analogue of [7, Proposition 1.7], where ∆(x)(a) := ∆+(x)(a) for
x ∈ X and a ∈ A. We say that X is C-injective if for every Y, Z ∈ mod − A, every
completely isometric morphism Ψ : Y ֒→ Z, and every morphism Φ : Y → X, there
exists a morphism eΦ : Z → X such that keΦkcb ≤ CkΦkcb and eΦ ◦ Ψ = Φ.
Clearly, when A = C, the definition of C-injectivity coincides with that of a C-
injective operator space [32, §24]. For general A, the dual X ∗ of any X ∈ mod − A
has a canonical left A-module structure, and it follows that X ∗ is C-injective in
A − mod whenever X is C-projective in mod − A by an operator module version
of [2, Theorem 3.5]. Moreover, X ∗ is relatively C-injective in A − nmod if and
only if X is relatively C-projective in mod − A. The next proposition also appears
to be new and will be used in the proof of our main result.
Proposition 2.3. Let A be a completely contractive Banach algebra and let X ∈
mod−A. If X is C1-injective in mod−C and is relatively C2-injective in mod−A,
then X is C1C2-injective in mod − A.
Proof. We first show that CB(A+, X) is C1-injective in mod−A using the standard
argument. To this end, let Y, Z ∈ mod−A, let κ : Y ֒→ Z be a completely isometric
morphism, and let α : Y → CB(A+, X) be a morphism. Define β : Y → X by
β(y) = α(y)(e), y ∈ Y , where e is the identity in A+. Then kβkcb ≤ kαkcb, and
by C1-injectivity of X in mod − C, there exists an extension eβ : Z → X satisfying
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
5
β = eβ ◦ κ and keβkcb ≤ C1kβkcb ≤ C1kαkcb. Define eα : Z → CB(A+, X) by
eα(z)(a) = eβ(z · a), for z ∈ Z, a ∈ A+. Then
for all z ∈ Z and a, b ∈ A. Thus, eα is a module map extending α such that
keαkcb ≤ keβkcb ≤ C1kαkcb.
Now, by relative C2-injectivity of X in mod − A there exists a morphism Φ+ :
CB(A+, X) → X satisfying Φ+ ◦ ∆+ = idX and kΦ+kcb ≤ C2. Thus, if Y, Z ∈
mod − A with κ : Y ֒→ Z a completely isometric morphism, and α : Y → X
is a morphism, then we may extend the morphism ∆+ ◦ α : Y → CB(A+, X) to
(eα(z) · a)(b) = eα(z)(ab) = eβ(z · ab) = eα(z · a)(b)
a morphism eα : Z → CB(A+, X) with keαkcb ≤ C1k∆+ ◦ αkcb ≤ C1kαkcb. The
morphism Φ+ ◦ eα : Z → X satisfies Φ+ ◦ eα ◦ κ = α and kΦ ◦ eαkcb ≤ C1C2kαkcb,
and is therefore the desired extension.
(cid:3)
The converse of Proposition 2.3 is not true in general (when C1 = C2 = 1).
Indeed, for any unital completely contractive Banach algebra A and any 1-injective
operator space X, it follows from the proof of Proposition 2.3 that CB(A, X) is
1-injective in mod − A. This clearly implies relative 1-injectivity in mod − A.
However, consider A = B(G) and X = C, where B(G) is the Fourier-Stieltjes
algebra of a non-amenable discrete group G. Since B(G) is the operator dual of
the full group C∗-algebra C∗(G), we have CB(A, X) = B(G)∗ = C∗(G)∗∗. If this
were a 1-injective operator space, the group C∗-algebra C∗(G) would be nuclear
[3], forcing G to be amenable by [28, Theorem 4.2].
Remark 2.4. Our notions of projectivity and injectivity are closer in spirit to the
approach taken in operator space theory [2] and the recent approach of Helemskii
[17] rather than Banach algebra homology, where the related notions are usually
studied solely from the relative perspective. In particular, we caution the reader
that "injectivity" as defined in our previous work [5] coincides with relative 1-
injectivity as defined above.
3. Locally Compact Quantum Groups
A locally compact quantum group is a quadruple G = (L∞(G), Γ, ϕ, ψ), where
L∞(G) is a Hopf-von Neumann algebra with co-multiplication Γ : L∞(G) →
L∞(G)⊗L∞(G), and ϕ and ψ are fixed left and right Haar weights on L∞(G),
respectively [27, 40]. For every locally compact quantum group G, there exists a
left fundamental unitary operator W on L2(G, ϕ) ⊗ L2(G, ϕ) and a right fundamen-
tal unitary operator V on L2(G, ψ) ⊗ L2(G, ψ) implementing the co-multiplication
Γ via
Γ(x) = W ∗(1 ⊗ x)W = V (x ⊗ 1)V ∗,
x ∈ L∞(G).
Both unitaries satisfy the pentagonal relation; that is,
(3.1)
W12W13W23 = W23W12
and
V12V13V23 = V23V12.
By [27, Proposition 2.11], we may identify L2(G, ϕ) and L2(G, ψ), so we will simply
use L2(G) for this Hilbert space throughout the paper. We denote by R the unitary
antipode of G.
Let L1(G) denote the predual of L∞(G). Then the pre-adjoint of Γ induces an
associative completely contractive multiplication on L1(G), defined by
⋆ : L1(G)b⊗L1(G) ∋ f ⊗ g 7→ f ⋆ g = Γ∗(f ⊗ g) ∈ L1(G).
6
JASON CRANN
The multiplication ⋆ is a complete quotient map from L1(G)b⊗L1(G) onto L1(G),
implying
hL1(G) ⋆ L1(G)i = L1(G).
Moreover, L1(G) is always self-induced. The proof is a simple application of [41,
Theorem 2.7], but we provide the details for the convenience of the reader.
Proposition 3.1. Let G be a locally compact quantum group. Then L1(G) is a
self-induced completely contractive Banach algebra.
Proof. Let em : L1(G)b⊗L1(G)L1(G) → L1(G) be the induced multiplication map.
Then em∗ : L∞(G) → (L1(G)b⊗L1(G)L1(G))∗ is nothing but the co-multiplication Γ.
Since (L1(G)b⊗L1(G)L1(G))∗ ∼= N ⊥, where N ⊆ L1(G)b⊗L1(G) is the closed linear
span of {f ⋆ g ⊗ h − f ⊗ g ⋆ h f, g, h ∈ L1(G)}, given X ∈ (L1(G)b⊗L1(G)L1(G))∗ ⊆
by [41, Theorem 2.7], and em∗ is surjective. Since em∗ = Γ is also a complete
L∞(G)⊗L∞(G), it follows that (Γ ⊗ id)(X) = (id ⊗ Γ)(X). Hence, X ∈ Γ(L∞(G))
isometry, the result follows.
For any locally compact quantum group G, the canonical L1(G)-bimodule struc-
(cid:3)
ture on L∞(G) is given by
f ⋆ x = (id ⊗ f )Γ(x)
and
x ⋆ f = (f ⊗ id)Γ(x)
for x ∈ L∞(G), and f ∈ L1(G). A left invariant mean on L∞(G) is a state
m ∈ L∞(G)∗ satisfying
(3.2)
hm, x ⋆ f i = hf, 1ihm, xi,
x ∈ L∞(G), f ∈ L1(G).
Right and two-sided invariant means are defined similarly. A locally compact quan-
tum group G is said to be amenable if there exists a left invariant mean on L∞(G).
It is known that G is amenable if and only if there exists a right (equivalently,
two-sided) invariant mean (cf. [11, Proposition 3]). We say that G is co-amenable
if L1(G) has a bounded left (equivalently, right or two-sided) approximate identity
(cf. [1, Theorem 3.1]).
The left regular representation λ : L1(G) → B(L2(G)) of G is defined by
λ(f ) = (f ⊗ id)(W ),
f ∈ L1(G),
and is an injective, completely contractive homomorphism from L1(G) into B(L2(G)).
Then L∞(bG) := {λ(f ) : f ∈ L1(G)}′′ is the von Neumann algebra associated with
the dual quantum group bG. Analogously, we have the right regular representation
ρ : L1(G) → B(L2(G)) defined by
ρ(f ) = (id ⊗ f )(V ),
f ∈ L1(G),
which is also an injective, completely contractive homomorphism from L1(G) into
B(L2(G)). Then L∞(bG′) := {ρ(f ) : f ∈ L1(G)}′′ is the von Neumann alge-
bra associated to the quantum group bG′.
It follows that L∞(bG′) = L∞(bG)′,
and the left and right fundamental unitaries satisfy W ∈ L∞(G)⊗L∞(bG) and
V ∈ L∞(bG′)⊗L∞(G) [27, Proposition 2.15]. Moreover, dual quantum groups al-
ways satisfy L∞(G) ∩ L∞(bG) = L∞(G) ∩ L∞(bG′) = C1 [43, Proposition 3.4].
If G is a locally compact group, we let Ga = (L∞(G), Γa, ϕa, ψa) denote the com-
mutative quantum group associated with the commutative von Neumann algebra
L∞(G), where the co-multiplication is given by Γa(f )(s, t) = f (st), and ϕa and ψa
are integration with respect to a left and right Haar measure, respectively. The dual
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
7
bGa of Ga is the co-commutative quantum group Gs = (V N (G), Γs, ϕs, ψs), where
V N (G) is the left group von Neumann algebra with co-multiplication Γs(λ(t)) =
λ(t) ⊗ λ(t), and ϕs = ψs is Haagerup's Plancherel weight (cf. [37, §VII.3]). Then
L1(Ga) is the usual group convolution algebra L1(G), and L1(Gs) is the Fourier
algebra A(G). It is known that every commutative locally compact quantum group
is of the form Ga [36, 42, Theorem 2; §2]. Therefore, every commutative locally
compact quantum group is co-amenable, and is amenable if and only if the under-
lying locally compact group is amenable. By duality, every co-commutative locally
compact quantum group is of the form Gs, which is always amenable [33, Theo-
rem 4], and is co-amenable if and only if the underlying locally compact group is
amenable, by Leptin's theorem [29].
k·k
denote
the reduced quantum group C∗-algebra of G. We say that G is compact if C0(G)
is a unital C∗-algebra, in which case we denote C0(G) by C(G). We say that G is
For a locally compact quantum group G, we let C0(G) := bλ(L1(bG))
discrete if L1(G) is unital. It is well-known that G is compact if and only if bG is
discrete, and in that case, L1(bG) ∼=L1{Tnα(C) α ∈ Irr(G)}, where Tnα(C) is the
space of nα × nα trace-class operators, and Irr(G) denotes the set of (equivalence
classes of) irreducible co-representations of the compact quantum group G [45].
For general G, the operator dual M (G) := C0(G)∗ is a completely contractive
Banach algebra containing L1(G) as a norm closed two-sided ideal via the map
L1(G) ∋ f 7→ f C0(G) ∈ M (G) [19].
We let Cu(G) be the universal quantum group C∗-algebra of G, and denote the
canonical surjective *-homomorphism by πu : Cu(G) → C0(G) [25]. The space
Cu(G)∗ then has the structure of a unital completely contractive Banach algebra
such that the map L1(G) → Cu(G)∗ given by the composition of the inclusion
L1(G) ⊆ M (G) and π∗
u : M (G) → Cu(G)∗ is a completely isometric homomor-
phism, and it follows that L1(G) is a norm closed two-sided ideal in Cu(G)∗ [25,
Proposition 8.3].
Let G be a locally compact quantum group. An element b′ ∈ L∞(bG)′ is said
to be a completely bounded right multiplier of L1(G) if ρ(f )b′ ∈ ρ(L1(G)) for all
f ∈ L1(G) and the induced map
mr
b′ : L1(G) ∋ f 7→ ρ−1(ρ(f )b′) ∈ L1(G)
cb(L1(G)) denote the space of all com-
is completely bounded on L1(G). We let M r
pletely bounded right multipliers of L1(G), which is a completely contractive Ba-
nach algebra with respect to the norm
k[b′
ij]kMn(M r
cb(L1(G))) = k[mr
bij
]kcb.
Completely bounded left multipliers are defined analogously and we denote by
M l
cb(L1(G)) the corresponding completely contractive Banach algebra. We now
review the relevant properties of completely bounded multipliers, adopting the no-
tation of [21].
Given b′ ∈ M r
cb(L1(G)), the adjoint Θr(b′) := (mr
b′ )∗ defines a normal com-
pletely bounded right L1(G)-module map on L∞(G).
In general, the restriction
Θr(b′)C0(G) leaves C0(G) invariant by [21, Proposition 4.1], and, together with [21,
Proposition 4.2], we have the completely isometric identifications
(3.3)
Θr : M r
cb(L1(G)) ∼= CBσ
L1(G)(L∞(G)) ∼= CBL1(G)(C0(G)).
8
JASON CRANN
It is known that M r
cb(L1(G)) is a dual operator space [19, Theorem 3.5], with
cb(L1(G)). When G = Gs is co-commutative, Haagerup and Kraus
predual Qr
gave a representation for elements of Qcb(L1(Gs)) = Qcb(A(G)) as ΩA,ρ for A ∈
C∗
λ(G) ⊗min K∞ and ρ ∈ A(G)b⊗T∞ [16, Proposition 1.5], where
hϕ, ΩA,ρi = h(Θr(ϕ) ⊗ idK∞ )(A), ρi, ϕ ∈ McbA(G);
C∗
λ(G) is the reduced C∗-algebra of G; the spaces K∞ and T∞ denote the com-
pact and trace-class operators on a countably infinite-dimensional Hilbert space,
respectively, and ⊗min denotes the minimum tensor product of C∗-algebras. This
was later generalized to the setting of Kac algbras by Kraus and Ruan [24, Theo-
rem 3.3]. Relying upon the general result [16, Lemma 1.6], their argument readily
extends to arbitrary locally compact quantum groups.
Proposition 3.2. Let G be a locally compact quantum group. Then
Qr
where hb′, ΩA,ρi = h(Θr(b′) ⊗ idK∞ )(A), ρi, b′ ∈ M r
cb(L1(G)) = {ΩA,ρ A ∈ C0(G) ⊗min K∞, ρ ∈ L1(G)b⊗T∞},
cb(L1(G)).
4. T (L2(G)) y B(L2(G))
Let G be a locally compact quantum group. The right fundamental unitary V
of G induces a co-associative co-multiplication
Γr : B(L2(G)) ∋ T 7→ V (T ⊗ 1)V ∗ ∈ B(L2(G))⊗B(L2(G)),
and the restriction of Γr to L∞(G) yields the original co-multiplication Γ on L∞(G).
The pre-adjoint of Γr induces an associative completely contractive multiplication
on the space of trace-class operators T (L2(G)), defined by
(cid:3) : T (L2(G))b⊗T (L2(G)) ∋ ω ⊗ τ 7→ ω (cid:3) τ = Γr
Since Γr is a complete isometry, it follows that Γr
we have
∗(ω ⊗ τ ) ∈ T (L2(G)).
∗ is a complete quotient map, so
(4.1)
T (L2(G)) = hT (L2(G)) (cid:3) T (L2(G))i.
Analogously, the left fundamental unitary W of G induces a co-associative co-
multiplication
Γl : B(L2(G)) ∋ T 7→ W ∗(1 ⊗ T )W ∈ B(L2(G))⊗B(L2(G)),
and the restriction of Γl to L∞(G) is also equal to Γ. The pre-adjoint of Γl induces
another associative completely contractive multiplication
(cid:1) : T (L2(G))b⊗T (L2(G)) ∋ ω ⊗ τ 7→ ω (cid:1) τ = Γl
It was shown in [19, Lemma 5.2] that the pre-annihilator L∞(G)⊥ of L∞(G) in
T (L2(G)) is a norm closed two sided ideal in (T (L2(G)), (cid:3)) and (T (L2(G)), (cid:1)),
respectively, and the complete quotient map
∗(ω ⊗ τ ) ∈ T (L2(G)).
(4.2)
π : T (L2(G)) ∋ ω 7→ f = ωL∞(G) ∈ L1(G)
is an algebra homomorphism from (T (L2(G)), (cid:3)), respectively, (T (L2(G)), (cid:1)), onto
L1(G).
x ∈ L∞(G).
By [27, Proposition 2.1] the unitary antipode R satisfies R(x) = bJx∗bJ, for
B(L2(G)) → B(L2(G)), via R(T ) = bJT ∗bJ, T ∈ B(L2(G)). The extended antipode
It therefore extends to a *-anti-automorphism (still denoted) R :
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
9
maps L∞(G) and L∞(bG) onto L∞(G) and L∞(bG′), respectively, and satisfies the
generalized antipode relations; that is,
(4.3)
(R ⊗ R) ◦ Γr = Σ ◦ Γl ◦ R and (R ⊗ R) ◦ Γl = Σ ◦ Γr ◦ R,
where Σ is the flip map on B(L2(G))⊗B(L2(G)). At the level of T (L2(G)), the
relations (4.3) mean
R∗(ω (cid:3) τ ) = R∗(τ ) (cid:1) R∗(ω) and R∗(ω (cid:1) τ ) = R∗(τ ) (cid:3) R∗(ω)
for all ω, τ ∈ T (L2(G)). We may therefore pass between the left and right products
using R, and as a result, we will often focus on the right product (cid:3) throughout the
article.
Since L2(G) ∼= L2(bG) for any locally compact quantum group G, applying the
above construction to the co-multiplication bΓ on L∞(bG) yields two dual products
b(cid:3) : T (L2(G))b⊗T (L2(G)) ∋ ω ⊗ τ 7→ ωb(cid:3)τ =bΓr
b(cid:1) : T (L2(G))b⊗T (L2(G)) ∋ ω ⊗ τ 7→ ωb(cid:1)τ =bΓl
∗(ω ⊗ τ ) ∈ T (L2(G)),
∗(ω ⊗ τ ) ∈ T (L2(G)).
This lifting of quantum group convolution to T (L2(G)) allows one to study prop-
erties of G and bG as well as their interactions a single space. One such interaction
was obtained in [23], and states that the dual products anti-commute.
Theorem 4.1. [23, Theorem 3.3] Let G be a locally compact quantum group. Then
for every ρ, ω, τ ∈ T (L2(G)) we have
(4.4)
(ρ (cid:3) ω)b(cid:3)τ = (ρb(cid:3)τ ) (cid:3) ω.
Equation (4.4) has an important consequence (Proposition 4.2) that will be used
in the proof of the main result.
For a locally compact quantum group G, the multiplications (cid:3) and (cid:1) define
operator T (L2(G))-bimodule structures on B(L2(G)) such that for x ∈ L∞(G) and
f = ωL∞(G) with ω ∈ T (L2(G)), we have
(4.5)
x (cid:3) ω = x (cid:1) ω = x ⋆ f
and ω (cid:1) x = ω (cid:3) x = f ⋆ x.
The bimodule actions of (T (L2(G)), (cid:3)) and (T (L2(G)), (cid:1)) on B(L2(G)) are there-
fore liftings of the usual bimodule action of L1(G) on L∞(G). For details on these
bimodules we refer the reader to [5, 19]. In what follows we denote the algebra of
completely bounded right (T (L2(G)), (cid:3))-module (respectively, left (T (L2(G)),b(cid:3))-
module) maps by CBT(cid:3) (B(L2(G))) (respectively, T b(cid:3) CB(B(L2(G)))).
In [19, Remark 7.4], the authors observe that for co-amenable G we have
CBT(cid:3) (B(L2(G))) ⊆ CBL∞(G)
L∞( bG)
(B(L2(G))),
where CBL∞(G)
L∞( bG)
maps on B(L2(G)) that leave L∞(G) globally invariant. As a corollary to the
commutation relation (4.4), we can remove the co-amenability hypothesis in the
above inclusion using the following "automatic" module property.
(B(L2(G))) is the algebra of completely bounded L∞(bG)-bimodule
Proposition 4.2. Let G be locally compact quantum group. Then
CBT(cid:3)(B(L2(G))) ⊆ T b(cid:3)CB(B(L2(G))).
10
JASON CRANN
Proof. Let Φ ∈ CBT(cid:3) (B(L2(G))), and fix ρ ∈ T (L2(G)) and T ∈ B(L2(G)). Then
for any ω, τ ∈ T (L2(G)), we have
Thus,
h(ρb(cid:3)T ) (cid:3) τ, ωi = hρb(cid:3)T, τ (cid:3) ωi = hT, (τ (cid:3) ω)b(cid:3)ρi
= hT, (τb(cid:3)ρ) (cid:3) ωi = hT (cid:3) (τb(cid:3)ρ), ωi.
hΦ(ρb(cid:3)T ), τ (cid:3) ωi = hΦ(ρb(cid:3)T ) (cid:3) τ, ωi = hΦ((ρb(cid:3)T ) (cid:3) τ ), ωi
= hΦ(T (cid:3) (τb(cid:3)ρ)), ωi = hΦ(T ) (cid:3) (τb(cid:3)ρ), ωi
= hΦ(T ), (τb(cid:3)ρ) (cid:3) ωi = hΦ(T ), (τ (cid:3) ω)b(cid:3)ρi
= hρb(cid:3)Φ(T ), τ (cid:3) ωi.
By (4.1) it follows that Φ(ρb(cid:3)T ) = ρb(cid:3)Φ(T ), as required.
Corollary 4.3. For any locally compact quantum group G, we have
(cid:3)
CBT(cid:3) (B(L2(G))) ⊆ CBL∞(G)
L∞( bG)
(B(L2(G))).
Proof. Let Φ ∈ CBT(cid:3)(B(L2(G))), and x, y ∈ L∞(bG). Then for any ρ ∈ T (L2(G))
and T ∈ B(L2(G)) we have
(xT y) (cid:3) ρ = (ρ ⊗ id)V (xT y ⊗ 1)V ∗ = (ρ ⊗ id)((x ⊗ 1)V (T ⊗ 1)V ∗(y ⊗ 1))
= (y · ρ · x ⊗ id)V (T ⊗ 1)V ∗ = T (cid:3) (y · ρ · x).
Thus, for any ω ∈ T (L2(G)) we obtain
hΦ(xT y), ρ (cid:3) ωi = hΦ(xT y) (cid:3) ρ, ωi = hΦ((xT y) (cid:3) ρ), ωi
= hΦ(T (cid:3) (y · ρ · x)), ωi = hΦ(T ) (cid:3) (y · ρ · x), ωi
= h(xΦ(T )y) (cid:3) ρ, ωi = hxΦ(T )y, ρ (cid:3) ωi.
any x ∈ L∞(G) and ρ ∈ T (L2(G)) we have
Again by (4.1), it follows that Φ is an L∞(bG)-bimodule map on B(L2(G)).
By Proposition 4.2 Φ ∈ T b(cid:3) CB(B(L2(G))), and since bV ∈ L∞(G)′⊗L∞(bG), for
(id ⊗ ρ)bV (Φ(x) ⊗ 1)bV ∗ = ρb(cid:3)Φ(x) = Φ(ρb(cid:3)x) = hρ, 1iΦ(x) = (id ⊗ ρ)(Φ(x) ⊗ 1).
It follows that bV (Φ(x) ⊗ 1)bV ∗ = Φ(x) ⊗ 1, which implies that bρ( f )Φ(x) = Φ(x)bρ( f )
for every f ∈ L1(bG). Since bρ(L1(bG)) is weak* dense in L∞(G)′, we have Φ(x) ∈
L∞(G)′′ = L∞(G). Thus, Φ leaves L∞(G) globally invariant, and the claim follows.
(cid:3)
In [5], we studied the existence of conditional expectations E : B(L2(G)) →
from G. We now complete this picture by studying the four remaining T (L2(G))-
L∞(bG) commuting with the four T (L2(G))-module structures on B(L2(G)) arising
module structures on B(L2(G)) arising from bG. We denote by T b(cid:1) CB(B(L2(G))) and
CBT b(cid:1) (B(L2(G))) the algebra of completely bounded left and right (T (L2(G)),b(cid:1))-
module maps on B(L2(G)), respectively, and similarly for (T (L2(G)),b(cid:3)).
ditional expectation E : B(L2(G)) → L∞(bG) in CBT b(cid:3) (B(L2(G))) if and only if
Proposition 4.4. Let G be a locally compact quantum group. There exists a con-
G = C1.
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
11
G = C1.
(cid:3)
(cid:3)
Proposition 4.5. Let G be a locally compact quantum group. There exists a con-
Proof. For any T ∈ B(L2(G)) and ρ ∈ T (L2(G)), we have Tb(cid:3)ρ ∈ L∞(bG), so if such
a conditional expectation E exists, then Tb(cid:3)ρ = E(Tb(cid:3)ρ) = E(T )b(cid:3)ρ. By density of
products (4.1), it follows that E(T ) = T . In particular B(L2(G)) ⊆ L∞(bG), which
entails that G = bG = C1. The converse is trivial.
ditional expectation E : B(L2(G)) → L∞(bG) in T b(cid:1)CB(B(L2(G))) if and only if
Proof. Using the extended unitary antipode of bG, denoted by bR, it follows that
bR ◦ E ◦ bR is a conditional expectation onto L∞(bG) in CBT b(cid:3)(B(L2(G))), so the
ditional expectation E : B(L2(G)) → L∞(bG) in T b(cid:3)CB(B(L2(G))) if and only if G
tation onto L∞(bG) in CBT(cid:3) (B(L2(G))), which, thanks to Proposition 4.2, lies in
Proposition 4.6. Let G be a locally compact quantum group. There exists a con-
Proof. If G is amenable, then by [5, Theorem 4.2] there exists a conditional expec-
result follows from Proposition 4.4.
On the other hand, if there exists such a conditional expectation E, then for any
T b(cid:3)CB(B(L2(G))).
is amenable.
x ∈ L∞(G) and ρ ∈ T (L2(G)) we have
ρb(cid:3)E(x) = E(ρb(cid:3)x) = hρ, 1iE(x).
G is amenable by [35, Theorem 3].
Proposition 4.7. Let G be a locally compact quantum group. There exists a con-
As in the proof of Corollary 4.3, this implies E(x) ∈ L∞(G) ∩ L∞(bG) = C1. Hence,
ditional expectation E : B(L2(G)) → L∞(bG) in CBT b(cid:1) (B(L2(G))) if and only if G
Proof. This follows from Proposition 4.6 using the extended unitary antipode bR.
is amenable.
(cid:3)
(cid:3)
We record the normal version of Proposition 4.7 for later use. The proof is left
to the reader to establish.
Proposition 4.8. Let G be a locally compact quantum group. There exists a normal
conditional expectation E : B(L2(G)) → L∞(bG) in CBT b(cid:1)(B(L2(G))) if and only if
G is compact.
5. Main Result & Applications
We are now in position to prove the main result of the paper -- the equivalence
theorem also reveals a duality between the left and right (T (L2(G)), (cid:3))-module
structures on B(L2(G)): amenability of G is captured by left injectivity of B(L2(G))
of amenability of bG and 1-injectivity of L∞(G) as an operator L1(G)-module. The
[5, Theorem 5.5], while amenability of bG is captured by right injectivity of B(L2(G)).
12
JASON CRANN
Theorem 5.1. Let G be a locally compact quantum group. The following conditions
are equivalent:
(1) bG is amenable;
(2) B(L2(G)) is 1-injective in mod − (T (L2(G)), (cid:3));
(3) L∞(G) is 1-injective in mod − L1(G).
Proof. (1) ⇒ (2): By [5, Proposition 5.8] amenability of bG implies that B(L2(G))
is relatively 1-injective in mod − (T (L2(G)), (cid:3)). Since B(L2(G)) is a 1-injective
operator space, the implication follows from Proposition 2.3.
(2) ⇒ (3): If B(L2(G)) is 1-injective in mod − (T (L2(G)), (cid:3)), there exists a
completely contractive morphism Φ : B(L2(G))⊗B(L2(G)) → B(L2(G)) which is
a left inverse to Γr, where the pertinent right (T (L2(G)), (cid:3))-module structure on
B(L2(G))⊗B(L2(G)) is defined by
X (cid:17) ρ = (ρ ⊗ id ⊗ id)(Γr ⊗ id)(X), X ∈ B(L2(G))⊗B(L2(G)), ρ ∈ T (L2(G)).
By the proof of Proposition 4.2 it follows that Φ(ρb(cid:17)X) = ρb(cid:3)Φ(X), where the left
module action b(cid:17) is given by
ρb(cid:17)X = (id ⊗ ρ ⊗ id)(bΓr ⊗ id)(X), X ∈ B(L2(G))⊗B(L2(G)), ρ ∈ T (L2(G)).
Furthermore, the proof of Corollary 4.3 entails the invariance
Φ(L∞(G)⊗B(L2(G))) ⊆ L∞(G).
Since Γl : B(L2(G)) → L∞(G)⊗B(L2(G)), the composition Φ ◦ Γl therefore maps
into L∞(G). Moreover, if x ∈ L∞(G) then Φ ◦ Γl(x) = Φ ◦ Γr(x) = x, so that
Φ ◦ Γl is a projection of norm one from B(L2(G)) onto L∞(G). Thus, L∞(G) is a
1-injective operator space.
Next, consider the map Ψ = ΦL∞(G)⊗L∞(G) : L∞(G)⊗L∞(G) → L∞(G). Since
the right (T (L2(G)), (cid:3))-module action on B(L2(G)) restricts to the canonical right
L1(G)-module action on L∞(G), it follows that Ψ is a completely contractive right
L1(G)-module map such that Ψ ◦ Γ = idL∞(G). Since L∞(G) is faithful in mod −
L1(G), this entails the relative 1-injectivity of L∞(G) in mod−L1(G), and therefore
the 1-injectivity of L∞(G) in mod − L1(G) by Proposition 2.3.
(3) ⇒ (1): Viewing B(L2(G)) as a right operator L1(G)-module via
T (cid:1) f = (f ⊗ id)Γl(T ) = (f ⊗ id)W ∗(1 ⊗ T )W,
f ∈ L1(G), T ∈ B(L2(G)),
1-injectivity of L∞(G) in mod−L1(G) gives an extension of idL∞(G) to a completely
contractive morphism E : B(L2(G)) → L∞(G). Proposition 4.7 then entails the
(cid:3)
amenability of bG.
Analogously, there is a left module version of Theorem 5.1 involving the left
product (cid:1), the proof of which follows similarly.
Theorem 5.2. Let G be a locally compact quantum group. The following conditions
are equivalent:
(1) bG is amenable;
(2) B(L2(G)) is 1-injective in (T (L2(G)), (cid:1)) − mod;
(3) L∞(G) is 1-injective in L1(G) − mod.
In the co-commutative setting, we obtain a new characterization of amenable
locally compact groups.
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
13
Corollary 5.3. Let G be a locally compact group. The following conditions are
equivalent:
(1) G is amenable;
(2) V N (G) is 1-injective in mod − A(G);
(3) V N (G) is 1-injective in A(G) − mod.
Remark 5.4. Corollary 5.3 highlights the significance of the non-relative homology
of V N (G) as an operator A(G)-module. In fact, relative 1-injectivity of V N (G)
in mod − A(G) is equivalent to inner amenability of G [6, Theorem 3.4]. Related
results at the level of quantum groups will appear in subsequent work.
In [15] Haagerup provided an elegant characterization of injective von Neumann
algebras via decomposability of completely bounded maps. More specifically, a von
Neumann algebra M is injective if and only if CB(M ) = span CP(M ), where CP(M )
is the set of completely positive maps Φ : M → M . The next result provides a
similar decomposition for L1(G)-module maps on L∞(G) when L∞(G) is 1-injective
in mod − L1(G).
Proposition 5.5. Let G be a locally compact quantum group.
If L∞(G) is 1-
injective in mod − L1(G) (equivalently, bG is amenable) then
CBL1(G)(L∞(G)) = span CP L1(G)(L∞(G)).
Proof. Viewing Mn(L∞(G)) as an operator L1(G)-module under the amplified ac-
tion:
[xij ] ⋆ f = [xij ⋆ f ],
[xij ] ∈ Mn(L∞(G)), f ∈ L1(G),
we claim that Mn(L∞(G)) is 1-injective in mod − L1(G) for any n ∈ N. Indeed,
the canonical morphism
∆n : Mn(L∞(G)) → CB(L1(G), Mn(L∞(G))) = Mn(CB(L1(G), L∞(G)))
is nothing but the nth amplification of ∆ : L∞(G) → CB(L1(G), L∞(G)), so the
nth amplification of a completely contractive module left inverse of ∆ (which exists
by 1-injectivity of L∞(G)) provides a completely contractive module left inverse
to ∆n. Since Mn(L∞(G)) is 1-injective in mod − C [38, Proposition XV.3.2], the
claim follows from Proposition 2.3.
Now, let Φ ∈ CBL1(G)(L∞(G)) be a complete contraction, and consider the
Paulsen system S ⊆ M2(L∞(G)) defined by
S =(cid:26)(cid:18)α1
y
x
β1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) x, y ∈ L∞(G), α, β ∈ C(cid:27).
β1(cid:19) ∈ S,
β1 (cid:19) , (cid:18)α1
Φ∗(y)
Φ(x)
x
y
Then S is an L1(G)-submodule of M2(L∞(G)) and Φ gives rise to a unital com-
pletely positive L1(G)-module map ΦS : S → M2(L∞(G)) via off-diagonalization
[31]:
ΦS(cid:18)(cid:18)α1
y
x
β1(cid:19)(cid:19) =(cid:18) α1
where Φ∗(y) = Φ(y∗)∗, y ∈ L∞(G). By 1-injectivity of M2(L∞(G)) in mod −
L1(G), the map ΦS extends to a completely contractive L1(G)-module map eΦ :
M2(L∞(G)) → M2(L∞(G)) such that
0 1(cid:19)(cid:19) =(cid:18)1
eΦ(cid:18)(cid:18)1 0
0
0
1(cid:19) .
14
JASON CRANN
Hence, eΦ is completely positive and is of the form
x21 x22(cid:19)(cid:19) =(cid:18)Ψ1(x11) Φ(x12)
Φ∗(x21) Ψ2(x22)(cid:19) ,
eΦ(cid:18)(cid:18)x11 x12
where Ψi ∈ CP L1(G)(L∞(G)) is associated to Pii ◦eΦ ◦ Pii, and
Pii ∈ CP L1(G)(M2(L∞(G)))
[xij ] ∈ M2(L∞(G)),
is the diagonal projection onto the (i, i)th entry for i = 1, 2. By [13, Proposition
5.4.2], it follows that the map
eΦL∞(G) : L∞(G) ∋ x 7→(cid:18)Ψ1(x) Φ(x)
Φ∗(x) Ψ2(x)(cid:19) ∈ M2(L∞(G))
is a completely positive L1(G)-module map. Thus, via polarization (as in [13,
Proposition 5.4.1]), it follows that Φ ∈ span CP L1(G)(L∞(G)).
(cid:3)
Remark 5.6. As the proof of Proposition 5.5 shows, when L∞(G) is 1-injective in
mod−L1(G) we can decompose any Φ ∈ CBL1(G)(L∞(G)) into a linear combination
of 4 completely positive L1(G)-module maps.
For a locally compact group G, it is well-known that B(G) = McbA(G) whenever
G is amenable [10, Corollary 1.8]. Using Proposition 5.5 together with [9, Theorem
5.2] we can now generalize this implication to arbitrary locally compact quantum
groups. We note that the same result was obtained under the a priori stronger
assumption that G is co-amenable [19, Theorem 4.2].
Corollary 5.7. Let G be a locally compact quantum group. If bG is amenable then
Cu(G)∗ ∼= M r
cb(L1(G)).
Proof. First, we claim that CBL1(G)(C0(G), L∞(G)) = CBL1(G)(C0(G)). One inclu-
sion is obvious, so let Φ ∈ CBL1(G)(C0(G), L∞(G)). Then the restriction of its ad-
joint Φ∗L1(G) ∈ L1(G)CB(L1(G), M (G)) = L1(G)CB(L1(G)), noting that L1(G) =
hL1(G)⋆L1(G)i is a closed ideal in M (G). Hence, (Φ∗L1(G))∗ ∈ CBσ
L1(G)(L∞(G)) =
CBL1(G)(C0(G)) by [21, Proposition 4.1]. But
h(Φ∗L1(G))∗(x), f i = hx, Φ∗L1(G)(f )i = hx, Φ∗(f )i = hΦ(x), f i
for all x ∈ C0(G) and f ∈ L1(G), so (Φ∗L1(G))∗ is an extension of Φ which leaves
C0(G) invariant, hence so too does Φ.
Letting I denote the complete isometry
CBL1(G)(C0(G)) ∋ Φ 7→ (Φ∗L1(G))∗ ∈ CBσ
L1(G)(L∞(G))
and R : CBL1(G)(L∞(G)) → CBL1(G)(C0(G), L∞(G)) the completely contractive
restriction map, it follows that P σ := I ◦ R : CBL1(G)(L∞(G)) → CBσ
L1(G)(L∞(G))
is a completely contractive projection onto CBσ
L1(G)(L∞(G)). Moreover, P σ maps
CP L1(G)(L∞(G)) onto CP σ
L1(G)(L∞(G)).
Since bG is amenable, CBL1(G)(L∞(G)) = span CP L1(G)(L∞(G)) by Proposition
L1(G)(L∞(G)) there exist Φi ∈ CP L1(G)(L∞(G)) i = 1, ..., 4
5.5, so given Φ ∈ CBσ
such that
Φ =
1
4
(Φ1 − Φ2 + i(Φ3 − Φ4)).
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
15
But then
Φ = P σ(Φ) =
1
4
(P σ(Φ1) − P σ(Φ2) + i(P σ(Φ3) − P σ(Φ4))),
and it follows that CBσ
span CP σ
L1(G)(L∞(G)) ∼= Cu(G)∗, so we have
L1(G)(L∞(G)) = span CP σ
L1(G)(L∞(G)). By [9, Theorem 5.2],
M r
cb(L1(G)) ∼= CBσ
L1(G)(L∞(G)) ∼= Cu(G)∗.
The observations in the proof of Corollary 5.7 lead to the following new charac-
terization of the predual of M r
cb(L1(G)).
Proposition 5.8. Let G be a locally compact quantum group. Then
(cid:3)
completely isometrically.
Qr
cb(L1(G)) ∼= C0(G)b⊗L1(G)L1(G)
Proof. As noted in the proof of Corollary 5.7, we have
CBL1(G)(C0(G)) = CBL1(G)(C0(G), L∞(G)).
cb(L1(G)) ∼= CBL1(G)(C0(G), L∞(G)) completely isometrically [21,
Thus, Θr : M r
Proposition 4.1]. We need to show that Θr is a weak*-weak* homeomorphism.
Since
CBL1(G)(C0(G), L∞(G)) ∼= (C0(G)b⊗L1(G)L1(G))∗
i)i∈I be a bounded net in M r
weak* homeomorphically, and Θr is a completely isometric isomorphism, it suffices
to show that Θr is weak* continuous on bounded sets (see [8, Lemma 10.1]). Let
cb(L1(G)) converging weak* to b′. By Proposition
(b′
cb(L1(G)),
3.2, for any A ∈ C0(G) ⊗min K∞ and ρ ∈ L1(G)b⊗T∞ we have ΩA,ρ ∈ Qr
ha′, ΩA,ρi = h(Θr(a′) ⊗ idK∞)(A), ρi,
i))i∈I converges point weak* to Θr(b′) in CB(C0(G), L∞(G)). Letting
cb(L1(G)).
Then (Θr(b′
a′ ∈ M r
where
q : C0(G)b⊗L1(G) ։ C0(G)b⊗L1(G)L1(G)
be the quotient map, and viewing Θr(b′
of the image q(C0(G) ⊗ L1(G)) of the algebraic tensor product C0(G) ⊗ L1(G)
i))i∈I imply that
(cid:3)
i) ∈ (C0(G)b⊗L1(G)L1(G))∗, the density
in C0(G)b⊗L1(G)L1(G), together with the boundedness of (Θr(b′
i))i∈I converges weak* to Θr(b′) in (C0(G)b⊗L1(G)L1(G))∗.
Remark 5.9. The identification of Qr
the co-commutative case, that is, for any locally compact group G we have
cb(L1(G)) in Proposition 5.8 is new even in
(Θr(b′
completely isometrically.
Qcb(G) ∼= C∗
λ(G)b⊗A(G)A(G)
For our final application, we now give a simplified proof of the fact that amenabil-
ity of a discrete quantum group implies co-amenability of its compact dual.
Theorem 5.10. A compact quantum group G is co-amenable if and only if bG is
amenable.
16
JASON CRANN
Proof. Co-amenability of G always implies amenability of bG [1, Theorem 3.2], so
assume that bG is amenable. By Theorem 5.1 we know that L∞(G) is 1-injective in
mod − L1(G). Let Φ : L∞(G)⊗L∞(G) → L∞(G) be a completely contractive left
inverse to Γ which is a right L1(G)-module map. As a unital complete contraction, Φ
is completely positive and ΦC(G)⊗minC(G) 6= 0 since C(G) is unital. By [1, Theorem
3.3] we also know that C(G) is nuclear, so let (Ψa)a∈A be a net of finite-rank,
unital completely positive maps converging to idC(G) in the point-norm topology.
For a ∈ A, consider the unital completely positive map Φa : C(G) → C(G) given
by
Φa = Φ ◦ (id ⊗ Ψa) ◦ ΓC(G).
The fact that Φa maps into C(G) follows from the density of Γ(C(G))(C(G) ⊗ 1) in
C(G)⊗minC(G) (see [26, Corollary 6.11]) together with the Γ(L∞(G))-module prop-
erty of Φ (see the proof of [5, Theorem 5.5]). However, the invariance Φa(C(G)) ⊆
C(G) will be a byproduct of the following argument.
Since Ψa is finite rank, there exist xa
1, ..., xa
na ∈ C(G) and µa
1, ..., µa
na ∈ M (G)
such that
Ψa(x) =
naXn=1
hµa
n, xixa
n,
x ∈ C(G), a ∈ A.
For each a ∈ A, and 1 ≤ n ≤ na, let Φ(a,n) : C(G) → C(G) be defined by
Φ(a,n)(x) = Φ(x ⊗ xa
n),
x ∈ C(G).
Then Φ(a,n) is completely bounded with kΦ(a,n)kcb ≤ kxa
L1(G)-module map. Hence, Φ(a,n) ∈ CBL1(G)(C(G)) = Θr(M r
M r
Θr(νa
cb(L1(G)) = Cu(G)∗ by Corollary 5.7, there exist νa
nkC(G), and is a right
cb(L1(G))). Since
n ∈ Cu(G)∗ such that Φ(a,n) =
n).
Let x ∈ C0(G). Then
Φa(x) = Φ((id ⊗ Ψa)(Γ(x)))
Φ(((id ⊗ µa
n)Γ(x)) ⊗ xa
n)
Φ(a,n)(((id ⊗ µa
n)Γ(x)))
=
=
=
naXn=1
naXn=1
naXn=1
naXn=1
= Θr(cid:18) naXn=1
=
Θr(νa
n)(Θr(µa
n)(x))
Θr(νa
n ⋆ µa
n)(x)
n ⋆ µa
νa
n(cid:19)(x).
Letting µa =Pna
as M (G) is a two-sided ideal in Cu(G)∗. Since Θr is an isometry,
n=1 νa
n ⋆µa
n, we obtain Φa = Θr(µa). Moreover, we have µa ∈ M (G)
kµakM(G) = kΘr(µa)kcb = kΦakcb = 1,
a ∈ A,
and since Φa converges to idC(G) in the point-norm topology it follows that
µa ⋆ x = Θr(µa)(x) → x,
x ∈ C(G).
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
17
Let µ be a weak* cluster point of (µa)a∈A in the unit ball of M (G) = C(G)∗. Then
µ is a right identity of M (G). The restricted unitary antipode R maps C(G) into
C(G) and satisfies R∗(µ ⋆ ν) = R∗(ν) ⋆ R∗(µ) for all ν ∈ M (G). Hence, R∗(µ) is a
left identity of M (G). It follows that ε := µ + R∗(µ) − µ ⋆ R∗(µ) is an identity for
M (G). Hence, G is co-amenable by [1, Theorem 3.1].
(cid:3)
Given a completely contractive Banach algebra A with a contractive approximate
identity, any essential module X ∈ mod − A is induced by [8, Proposition 6.4].
Since a locally compact quantum group G is co-amenable if and only if L1(G) has
a contractive approximate identity [18, Theorem 2], the next proposition supports
the idea that our methods may be applicable to the general duality problem of
amenability and co-amenability.
Proposition 5.11. Let G be a locally compact quantum group for which the dual
bG is amenable. Then for any closed right ideal I (cid:2) L1(G), the multiplication map
yields a completely isometric isomorphism
emI : Ib⊗L1(G)L1(G) ∼= hI ⋆ L1(G)i.
L1(G)-module.
In particular, if I is essential then Ib⊗L1(G)L1(G) ∼= I, that is, I is an induced
A and any closed right ideal J (cid:2) A, we have emA/J : (A/J)b⊗AA ∼= A/hJ · Ai
Proof. First, note that for any self-induced completely contractive Banach algebra
Indeed, identifying (A/hJ · Ai)∗ = hJ · Ai⊥ ⊆ A∗, it
completely isometrically.
follows that
(emA/J )∗ : hJ · Ai⊥ → ((A/J)b⊗AA)∗ = (NA/J )⊥
q : A ։ A/J be the complete quotient map, if X ∈ (NA/J )⊥ then (q∗ ⊗ id)(X) ∈
is equal to (emA)∗hJ·Ai⊥. In particular, (emA/J )∗ is a complete isometry. Letting
(NA)⊥, so there exists F ∈ A∗ such that (q∗ ⊗ id)(X) = (emA)∗(F ) as A is self-
induced. Clearly, F ∈ hJ · Ai⊥, so (emA/J )∗ is also surjective.
Since bG is amenable, by Theorem 5.1 L∞(G) is 1-injective in mod − L1(G).
Then for every 1-exact sequence of right A-modules
the induced sequence
0 → Y ֒→ Z ։ Z/Y → 0,
0 → CBL1(G)(Z/Y, L∞(G)) ֒→ CBL1(G)(Z, L∞(G)) ։ CBL1(G)(Y, L∞(G)) → 0
is 1-exact, where 1-exactness refers to an exact sequence of morphisms such that the
injection (֒→) is a complete isometry and the surjection (։) is a complete quotient
map. Taking the pre-adjoint of the above sequence we obtain the 1-exact sequence
In particular, take Y = I and Z = L1(G), and consider the commutative diagram:
0 → Y b⊗L1(G)L1(G) ֒→ Zb⊗L1(G)L1(G) ։ Z/Y b⊗L1(G)L1(G) → 0.
y
Ib⊗L1(G)L1(G) −−−−→ L1(G)b⊗L1(G)L1(G) −−−−→ (L1(G)/I)b⊗L1(G)L1(G)
−−−−→ L1(G)/hI ⋆ L1(G)i
hI ⋆ L1(G)i −−−−→
L1(G)
y
y
18
JASON CRANN
As L1(G) is self-induced, the last two columns are completely isometric isomor-
phisms, and since both rows are 1-exact, it follows that
completely isometrically.
emI : Ib⊗L1(G)L1(G) ∼= hI ⋆ L1(G)i
A locally compact quantum group G is said to be regular if
K(L2(G)) = h(id ⊗ ω)(σV ) ω ∈ T (L2(G))i,
(cid:3)
where K(L2(G)) denotes the ideal of compact operators on L2(G), σ denotes the
flip map on L2(G) ⊗ L2(G), and as usual, h·i denotes the closed linear span. For
example, Kac algebras are regular, as well as discrete and compact quantum groups
(see [20]). Under the assumption of regularity, we now obtain a version of Theorem
5.1 at the predual level.
Theorem 5.12. Let G be a locally compact quantum group. Consider the following
conditions:
(1) bG is compact (equivalently, G is discrete);
(2) T (L2(G)) is relatively 1-projective in (T (L2(G)), (cid:3)) − mod;
(3) L1(G) is 1-projective in L1(G) − mod.
Then (1) ⇔ (2) ⇒ (3), and when G is regular, the conditions are equivalent.
Proof. The implication (1) ⇒ (2) follows by an argument similar to the proof of
[5, Proposition 5.8] using a normal two-sided invariant mean bm′ on L∞(bG′), which
exists by compactness. The implication (2) ⇒ (1) follows similarly to Theorem
5.1, giving the relative 1-injectivity of L∞(G) in nmod − L1(G) and the existence
of a normal conditional expectation E : B(L2(G)) → L∞(G). Viewing B(L2(G))
as a right L1(G)-module under the (cid:1)-action, the relative 1-injectivity of L∞(G)
in nmod − L1(G) implies the existence of a normal condition expectation P :
B(L2(G)) → L∞(G) that is a right L1(G)-module map (see [44, Lemma 3.5] for
1-injective in nmod − L1(G), which implies that L1(G) is relatively 1-projective in
details). Thus, bG is compact by Proposition 4.8.
(2) ⇒ (3): By the above we know that bG is compact and L∞(G) is relatively
L1(G) − mod. By discreteness of G we have L1(G) ∼= L1{Tnα(C) α ∈ Irr(bG)},
where Tnα(C) is the space of nα × nα trace-class operators. Hence, L1(G) is 1-
projective in C − mod by [2, Proposition 3.6, Proposition 3.7]. The left version of
Proposition 2.2 then entails the 1-projectivity of L1(G) in L1(G) − mod.
Now, suppose that G is regular. Considering again the right L1(G)-module
structure on B(L2(G)) given by the (cid:1)-action, it follows from [20, Corollary 3.6]
that K(L2(G)) is an essential L1(G)-submodule of B(L2(G)), that is, K(L2(G)) =
hK(L2(G)) (cid:1) L1(G)i. We show (3) ⇒ (1).
morphism and L1(G) is 1-projective in L1(G) − mod, for every ε > 0 there exists
Since the multiplication mL1(G) : L1(G)b⊗L1(G) → L1(G) is a complete quotient
a morphism Φε : L1(G) → L1(G)b⊗L1(G) satisfying mL1(G) ◦ Φε = idL1(G) and
kΦεkcb < 1 + ε. Moreover, we know that L∞(G) is 1-injective in mod − L1(G)
cb(L1(G)) ∼= Cu(G)∗ by Corollary
as the dual a 1-projective module. Thus, M r
5.7, and L∞(G) is a 1-injective operator space. Hence, L∞(G) is semi-discrete
[3, 4, 12], so there exits a net (Ψi)i∈I of normal finite-rank complete contractions
Ψi : L∞(G) → L∞(G) converging to idL∞(G) in the point weak* topology. Using
ε : L∞(G)⊗L∞(G) → L∞(G) which is
the normal completely bounded morphism Φ∗
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
19
a left inverse of Γ, one can argue in a similar manner to Theorem 5.10 by averaging
the normal finite-rank maps Ψi into multipliers and use the fact that L1(G) is
a two-sided ideal in Cu(G)∗ to obtain a bounded net (fi)i∈I in L1(G) satisfying
f ⋆ fi − f → 0 weakly for all f ∈ L1(G). The standard convexity argument then
yields a bounded right approximate identity for L1(G), and G is necessarily co-
amenable.
Now, since π : (T (L2(G)), (cid:1)) → L1(G) is a complete quotient morphism, for
any ε > 0 it also has a right inverse morphism Ψε : L1(G) → T (L2(G)) with
kΨεkcb < 1 + ε. Then Ψ∗
ε : B(L2(G)) → L∞(G) is a normal completely bounded
right (T (L2(G)), (cid:1))-module projection onto L∞(G). Since L1(G) has a contractive
approximate identity and K(L2(G)) is an essential L1(G)-module, we know that
K(L2(G)) is induced, that is,
is a completely isometric isomorphism. Hence, so too is its dual
emK(L2(G)) : K(L2(G))b⊗L1(G)L1(G) → K(L2(G))
Then Ψ∗
(emK(L2(G)))∗ : T (L2(G)) ∼= CBL1(G)(K(L2(G)), L∞(G)).
K(L2(G)) and f ∈ L1(G) we have
exists ρ ∈ T (L2(G)) satisfying (emK(L2(G)))∗(ρ) = Ψ∗
εK(L2(G)) ∈ CBL1(G)(K(L2(G)), L∞(G)) = (emK(L2(G)))∗(T (L2(G))), so there
εK(L2(G))(y), f i = h(emK(L2(G)))∗(ρ)(y), f i = hρ, y (cid:1) f i = hρ (cid:1) y, f i.
By weak* density of K(L2(G)) in B(L2(G)), we obtain Ψ∗
T ∈ B(L2(G)). In particular,
εK(L2(G)). Then for all y ∈
ε(T ) = ρ (cid:1) T for all
hΨ∗
π(ρ) ⋆ x = ρ (cid:1) x = Ψ∗
ε(x) = x
for all x ∈ L∞(G) as Ψ∗
ε is a projection. Then π(ρ) is a right identity for L1(G),
and using the unitary antipode R as in Theorem 5.10 we may construct a two-sided
(cid:3)
identity for L1(G), that is, G is discrete, whence bG is compact.
Analogously, there is a right module version of Theorem 5.12.
Theorem 5.13. Let G be a locally compact quantum group. Consider the following
conditions:
(1) bG is compact (equivalently, G is discrete);
(2) T (L2(G)) is relatively 1-projective in mod − (T (L2(G)), (cid:1));
(3) L1(G) is 1-projective in mod − L1(G).
Then (1) ⇔ (2) ⇒ (3), and when G is regular, the conditions are equivalent.
Remark 5.14. It is not clear at this time whether we can replace relative 1-
projectivity of T (L2(G)) with 1-projectivity of T (L2(G)) in condition (2) of The-
orems 5.12 and 5.13. However, one cannot replace 1-projectivity of L1(G) with
relative 1-projectivity of L1(G) in condition (3) of Theorems 5.12 and 5.13, as, for
example, L1(G) is always relatively 1-projective for any locally compact group G
(see [7, Theorem 2.4]).
Remark 5.15. Combining [2, Theorem 3.12] with [22, Corollary 7], it follows that
a regular quantum group G is discrete if and only if L1(G) is a 1-projective operator
space. Theorem 5.12 therefore yields the following equivalence for regular quantum
groups: L1(G) is 1-projective in C − mod if and only if L1(G) is 1-projective in
L1(G) − mod.
20
JASON CRANN
References
[1] E. B´edos and L. Tuset, Amenability and co-amenability for locally compact quantum groups.
Internat. J. Math. 14(2003), no. 8, 865 -- 884.
[2] D. P. Blecher, The standard dual of an operator space. Pacific J. Math. 153(1992), no. 1,
15 -- 30.
[3] M.-D. Choi and E. G. Effros, Nuclear C ∗-algebras and injectivity: the general case. Indiana
Univ. Math. J. 26(1977), no. 3, 443 -- 446.
[4] A. Connes, Classification of injective factors. Cases I I1, I I∞, I I Iλ, λ 6= 1. Ann. of Math.
(2) 104(1976), no. 1, 73 -- 115.
[5] J. Crann and M. Neufang, Amenability and covariant injectivity of locally compact quantum
groups. Trans. Amer. Math. Soc. 368(2016), 495 -- 513.
[6] J. Crann and Z. Tanko, On the operator homology of the Fourier algebra. Preprint.
arXiv:1602.05259.
[7] H. G. Dales and M. E. Polyakov, Homological properties of modules over group algebras.
Proc. London Math. Soc. (3) 89(2004), no. 2, 390 -- 426.
[8] M. Daws, Multipliers,
self-induced and dual Banach algebras. Dissertationes Math.
470(2010), 62 pp.
[9] M. Daws, Completely positive multipliers of quantum groups. Internat. J. Math. 23(2012),
no. 12, 1250132.
[10] J. de Canni`ere, J and U. Haagerup, Multipliers of the Fourier algebras of some simple Lie
groups and their discrete subgroups. Amer. J. Math. 107(1985), no. 2, 455 -- 500.
[11] P. Desmedt, J. Quaegebeur and S. Vaes, Amenability and the bicrossed product construction.
Illinois J. Math. 46(2002), no. 4, 1259 -- 1277.
[12] E. G. Effros and C. E. Lance, Tensor products of operator algebras. Adv. Math. 25(1977),
no. 1, 1 -- 34.
[13] E. G. Effros and Z.-J. Ruan, Operator Spaces. London Mathematical Society Monographs,
New Series 23, Oxford University Press, New York, 2000.
[14] A. Guichardet, Tensor products of C ∗-algebras. Aarhus University Lecture Notes Series
no. 12, 1969.
[15] U. Haagerup, Injectivity and decomposition of completely bounded maps. Operator algebras
and their connections with topology and ergodic theory (Busteni, 1983), 170-222, Lecture
Notes in Math., 1132, Springer, Berlin, 1985.
[16] U. Haagerup and J. Kraus, Approximation properties for group C ∗-algebras and group von
Neumann algebras. Trans. Amer. Math. Soc. 44(1994), no. 2, 667 -- 699.
[17] A. Y. Helemskii, Metric version of flatness and Hahn -- Banach type theorems for normed
modules over sequence algebras. Studia Math. 206(2011), no. 2, 135 -- 160.
[18] Z. Hu, M. Neufang, and Z.-J. Ruan, Multipliers on a new class of Banach algebras, locally
compact quantum groups, and topological centres. Proc. Lond. Math. Soc. (3) 100(2010),
no. 2, 429 -- 458.
[19] Z. Hu, M. Neufang, and Z.-J. Ruan, Completely bounded multipliers over locally compact
quantum groups. Proc. Lond. Math. Soc. (3) 103(2011), no. 1, 1 -- 39.
[20] Z. Hu, M. Neufang, and Z.-J. Ruan, Convolution of trace class operators over locally compact
quantum groups. Canad. J. Math. 65(2013), no. 5, 1043 -- 1072.
[21] M. Junge, M. Neufang and Z.-J. Ruan, A representation theorem for locally compact quantum
groups. Internat. J. Math. 20(2009), no. 3, 377 -- 400.
[22] M. Kalantar, Compact operators in regular LCQ groups. Canad. Math. Bull. 57(2014), no. 3,
546 -- 550.
[23] M. Kalantar and M. Neufang, Duality, cohomology, and geometry of locally compact quantum
groups. J. Math. Anal. Appl. 406(2013), no. 1, 22 -- 33.
[24] J. Kraus, J and Z.-J. Ruan, Approximation properties for Kac algebras. Indiana Univ. Math.
J. 48(1999), no. 2, 469 -- 535.
[25] J. Kustermans, Locally compact quantum groups in the universal setting. Internat. J. Math.
12(2001), no. 3, 289 -- 338.
[26] J. Kustermans and S. Vaes, Locally compact quantum groups. Ann. Sci. ´Ecole Norm. Sup.
(4) 33(2000), no. 6, 837 -- 934.
[27] J. Kustermans and S. Vaes, Locally compact quantum groups in the von Neumann algebraic
setting. Math. Scand. 92(2003), no. 1, 68 -- 92.
AMENABILITY AND COVARIANT INJECTIVITY OF LCQG II
21
[28] C. E. Lance, On nuclear C ∗-algebras. J. Funct. Anal. 12(1973), 157 -- 176.
[29] H. Leptin, Sur l'alg`ebre de Fourier d'un groupe localement compact. C. R. Acad. Sci. Paris
S´er. A-B 266(1968), A1180 -- A1182.
[30] M. Neufang, Z.-J. Ruan and N. Spronk, Completely isometric representations of McbA(G)
and U CB( bG)∗. Trans. Amer. Math. Soc. 360(2008), no. 3, 1133 -- 1161.
[31] V. I. Paulsen, Every completely polynomially bounded operator is similar to a contraction. J.
Funct. Anal. 55(1984), no. 1, 1 -- 17.
[32] G. Pisier, Introduction to Operator Space Theory. London Mathematical Society Lecture
Note Series, 294. Cambridge University Press, Cambridge, 2003.
[33] P. F. Renaud, Invariant means on a class of von Neumann algebras. Trans. Amer. Math.
Soc. 170(1972), 285 -- 291.
[34] Z.-J. Ruan, Amenability of Hopf von Neumann algebras and Kac algebras. J. Funct. Anal.
139(1996), no. 2, 466 -- 499.
[35] P. So ltan and A. Viselter, A note on amenability of locally compact quantum groups. Canad.
Math. Bull. 57(2014), no. 2, 424 -- 430.
[36] M. Takesaki, A characterization of group algebras as a converse of Tannaka -- Stinespring --
Tatsuuma duality theorem. Amer. J. Math. 91(1969), 529 -- 564.
[37] M. Takesaki, Theory of Operator Algebras I. Encyclopedia of Mathematical Sciences 124,
Springer-Verlag Berlin -- Heidelberg -- New York, 2002.
[38] M. Takesaki, Theory of Operator Algebras III. Encyclopedia of Mathematical Sciences 127,
Springer-Verlag Berlin -- Heidelberg -- New York, 2003.
[39] R. Tomatsu, Amenable discrete quantum groups. J. Math. Soc. Japan 58(2006), no. 4, 949 --
964.
[40] S. Vaes, Locally Compact Quantum Groups. Ph. D. thesis, K.U. Leuven, 2000.
[41] S. Vaes, The unitary implementation of a locally compact quantum group action. J. Funct.
Anal. 180(2001), no. 2, 426 -- 480.
[42] S. Vaes and L. Vainerman, On low-dimensional locally compact quantum groups. Locally
compact quantum groups and groupoids (Strasbourg, 2002), 127 -- 187, IRMA Lect. Math.
Theor. Phys., 2, de Gruyter, Berlin, 2003.
[43] A. Van Daele, Locally compact quantum groups. A von Neumann algebra approach. SIGMA
10(2014), 082, 41 pp.
[44] M. C. White, Injective modules for uniform algebras. Proc. London Math. Soc. (3) 73(1996),
no. 1, 155 -- 184.
[45] S. L. Woronowicz, Compact quantum groups. Sym´etries quantiques (Les Houches, 1995),
845 -- 884, North-Holland, Amsterdam, 1998.
School of Mathematics and Statistics, Carleton University, Ottawa, ON, Canada
K1S 5B6
Department of Pure Mathematics, University of Waterloo, Waterloo, ON, Canada
N2L 3G1
Institute for Quantum Computing, University of Waterloo, Waterloo, ON, Canada
N2L 3G1
Department of Mathematics and Statistics, University of Guelph, Guelph, ON, Canada
N1G 2W1
E-mail address: [email protected]
|
1407.6929 | 4 | 1407 | 2015-04-02T20:23:27 | Similarity degree of type II$_1$ von Neumann algebras with Property $\Gamma$ | [
"math.OA"
] | In this paper, we discuss some equivalent definitions of Property $\Gamma$ for a type II$_1$ von Neumann algebra. Using these equivalent definitions, we prove that the Pisier's similarity degree of a type II$_1$ von Neumann algebra with Property $\Gamma$ is equal to $3$. | math.OA | math |
Similarity degree of type II1 von Neumann algebras with
Property Γ
Wenhua Qian, Don Hadwin, and Junhao Shen
Abstract. In this paper, we discuss some equivalent definitions of Property Γ for a type II1
von Neumann algebra. Using these equivalent definitions, we prove that the Pisier's similarity
degree of a type II1 von Neumann algebra with Property Γ is equal to 3.
1. Introduction
Kadison's Similarity Problem for a C∗-algebra is a longstanding open problem, which asks
whether every bounded representation ρ of a C∗-algebra A on a Hilbert space H is similar to a
∗-representation. i.e. whether there exists an invertible operator T in B(H), such that T ρ(·)T −1
is a ∗-representation of A. Significant progress toward this famous open problem was obtained
in [1] and [3]. We will refer to Pisier's book [10] for a wonderful introduction to the problem
and many of its recent developments.
Similarity degree for a unital C∗-algebra A, denoted by d(A), was defined by Pisier in [7].
Since its introduction, this new concept has greatly influenced the study of Kadison's Similarity
Problem for C∗-algebras.
In fact, it was shown in [7] that Kadison's Similarity Problem for
a unital C∗-algebra A has an affirmative answer if and only if d(A) < ∞. One of the most
surprising results on similarity degree was also obtained by Pisier in [11] when he proved that,
for an infinite dimensional unital C∗-algebra A, the similarity degree of A is equal to 2 if and
only if A is a nuclear C∗-algebra.
Several results on similarity degree for a unital C∗-algebra have now been known. For
example, if there is no tracial state on a unital C∗-algebra A, then d(A) = 3 ([3], [9]). The
similarity degree of a type II1 factor M with Property Γ is less than or equal to 5 ([9]). This
result was later improved in [2] to that the similarity degree of such M is equal to 3. When A
is a minimal tensor product of two C*-algebras, one of which is nuclear and contains matrices
of any order, it was proved in [12] that d(A) ≤ 5. Recently, it was shown in [4] that, if A is Z-
stable, then d(A) ≤ 5. In [14], it was shown that, if a separable C*-algebra A has Property c∗-Γ,
2000 Mathematics Subject Classification. Primary 46L10; Secondary 46L05.
Key words and phrases. Property Γ, Similarity problem, Similarity degree.
1
2
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
then d(A) = 3, which implies that a nonnuclear separable Z-stable C∗-algebra has similarity
degree 3.
In this paper, we will discuss properties of type II1 von Neumann algebras with Property Γ
and compute similarity degree for this class of von Neumann algebras. First result we obtained
in the paper is the following characterization of type II1 von Neumann algebras with Property
Γ.
Theorem 3.10. Suppose M is a type II1 von Neumann algebra and ZM is the center of M.
Let τ be the center-valued trace on M such that τ (a) = a for any a ∈ ZM. Then the following
statements are equivalent.
(i) M has Property Γ in the sense of Definition 3.1 in [13].
(ii) There exists a family of nonzero orthogonal central projections {qα : α ∈ Ω} in M with
sum I such that qαM is a countably decomposable type II1 von Neumann algbera with
Property Γ for each α ∈ Ω.
(iii) For any n ∈ N, any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exist n orthogonal equivalent
projections p1, p2, . . . , pn in M with sum I such that
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k.
(iv) There exists a positive integer n0 ≥ 2 satisfying for any ǫ > 0 and a1, a2, . . . , ak ∈
M, there exist n0 orthogonal equivalent projections p1, p2, . . . , pn0 in M with sum I
satisfying
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n0, 1 ≤ j ≤ k.
(v) For any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exists a unitary u in M such that
(a) τ (u) = 0;
(b) τ ((uaj − aju)∗(uaj − aju)) < ǫI,
∀ 1 ≤ j ≤ k.
(vi) For any n ∈ N, any normal tracial state ρ on M, any ǫ > 0 and a1, a2, . . . , ak ∈ M,
there exist n orthogonal equivalent projections p1, p2, . . . , pn in M with sum I such that
kpiaj − ajpik2,ρ < ǫ,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k,
where k · k2,ρ is the 2-norm on M induced by ρ.
(vii) There exists a positive integer n0 ≥ 2 satisfying for any normal tracial state ρ on M,
any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exist n0 orthogonal equivalent projections
p1, p2, . . . , pn0 in M with sum I satisfying
kpiaj − ajpik2,ρ < ǫ,
∀ 1 ≤ i ≤ n0, 1 ≤ j ≤ k,
where k · k2,ρ is the 2-norm on M induced by ρ.
(viii) For any normal tracial state ρ on M, any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exists a
unitary u in M such that (a) τ (u) = 0; and (b) kuaj − ajuk2,ρ < ǫ, for all 1 ≤ j ≤ k,
where k · k2,ρ is the 2-norm on M induced by ρ.
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
3
We should remark that, when a type II1 von Neumann algebra M is in fact a type II1 factor,
our definition of Property Γ coincides with Murray and von Neumann's original definition in
[6].
Combining preceding Theorem 3.10 and results in [14], we are able to calculate the exact
value of similarity degree for a type II1 von Neumann algebra with Property Γ and obtain the
next result as a generalization of Christensen's result in [2].
Theorem 4.2. If M is a type II1 von Neumann algebra with Property Γ, then the similarity
degree d(M) = 3.
Next we apply Theorem 4.2 to calculate values of similarity degrees for two classes of C∗-
algebras, which were also considered by Pisier in [9].
Suppose A is unital C∗-algebra. Let I be some index set and
l∞(I, A) = {(xi)i∈I :
for each i ∈ I, xi ∈ A and sup
i∈I
kxik < ∞}.
Corollary 4.4. If M is a type II1 factor with Property Γ, then d(l∞(I, M)) = 3 for any index
set I.
Corollary 4.5. Let C = M2(C) ⊗ M2(C) ⊗ . . . (infinite C∗-tensor product of 2 × 2 matrix
algebras). Then, for any infinite index set I, d(l∞(I, C)) = 3.
The organization of this paper is as follows.
In section 2, we give some preliminaries on
direct integrals of separable Hilbert spaces and von Neumann algebras acting on separable
Hilbert spaces. In section 3, we give a characterization of type II1 von Neumann algberas with
Property Γ and obtain some equivalent definitions. In section 4, by showing that every finite
subset F of of a type II1 von Neumann algebra M with Property Γ is contained in a separable
unital C∗-subalgebra with Property c∗-Γ, we obtain that d(M) = 3.
2. Preliminaries
2.1. Dixmier Approximation Theorem. We will need the following Dixmier Approxi-
mation Theorem in the paper.
Lemma 2.1. (Dixmier Approximation Theorem) Let M be a finite von Neumann algebra
with center Z. Let τ be the center-valued trace on M. If a ∈ M, then
{τ (a)} = Z ∩ (convM(a)=) ,
where convM(a)= is the norm closure of the convex hull of the set {uau∗ : u is a unitary in M}.
2.2. Direct integral theory. General knowledge about direct integrals of separable Hilbert
spaces and von Neumann algebras acting on separable Hilbert spaces can be found in [18] and
[5]. Here we list a few lemmas that will be needed in this paper.
Lemma 2.2. ([5]) Suppose M is a von Neumann algebra acting on a separable Hilbert space
H. Let Z be the center of M. Then there is a direct integral decomposition of M relative to Z,
i.e. there exists a locally compact complete separable metric measure space (X, µ) such that
4
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
(i) H is (unitarily equivalent to) the direct integral of {Hs : s ∈ X} over (X, µ), where
each Hs is a separable Hilbert space, s ∈ X.
(ii) M is (unitarily equivalent to) the direct integral of {Ms} over (X, µ), where Ms is a
factor in B(Hs) almost everywhere. Also, if M is of type In(n could be infinite), II1,
II∞ or III, then the components Ms are, almost everywhere, of type In, II1, II∞ or
III, respectively.
Moreover, the center Z is (unitarily equivalent to) the algebra of diagonalizable operators relative
to this decomposition.
The following lemma gives a decomposition of a normal sate on a direct integral of von
Neumann algebras.
Lemma 2.3. ([5]) If H is the direct integral of separable Hilbert spaces {Hs} over (X, µ),
M is a decomposable von Neumann algebra on H (i.e., every operator in M is decomposable
relative to the direct integral decomposition, see Definition 14.1.6 in [5]) and ρ is a normal
state on M. There is a positive normal linear functional ρs on Ms for every s ∈ X such that
ρ(a) =RX ρs(a(s))dµ for each a in M. If M contains the algebra C of diagonalizable operators
and ρEME is faithful or tracial, for some projection E in M, then ρsE(s)MsE(s) is, accordingly,
faithful or tracial almost everywhere.
Remark 2.4. From the proof of Lemma 14.1.19 in [5], we obtain that if ρ =
ωyn on
M, where {yn} is a sequence of vectors in H such that
kynk2 = 1 and ωy is defined on M
such that ωy(a) = hay, yi for any a ∈ M, y ∈ H, then ρs can be chosen to be
s ∈ X.
3. Some equivalent definitions of type II1 von Neumann algebras with Property Γ
In this section, we will give some equivalent definitions of Property Γ for type II1 von
Neumann algebras.
Let us recall the following definition of Property Γ for general type II1 von Neumann algebras
in [13]. Suppose M is a type II1 von Neumann algebra with a predual M♯. Suppose σ(M, M♯)
is the weak-∗ topology on M induced from M♯. We say that M has Property Γ if and only
if ∀ a1, a2, . . . , ak ∈ M and ∀ n ∈ N, there exist a partially ordered set Λ and a family of
projections
{piλ : 1 ≤ i ≤ n; λ ∈ Λ} ⊆ M
satisfying
(i) For each λ ∈ Λ, p1λ, p2λ, . . . , pnλ are mutually orthogonal equivalent projections in M
with sum I.
(ii) For each 1 ≤ i ≤ n and 1 ≤ j ≤ k,
(piλaj − ajpiλ)∗(piλaj − ajpiλ) = 0
lim
λ
in σ(M, M♯) topology.
∞
Pn=1
∞
Pn=1
ωyn(s) for each
∞
Pn=1
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
5
The following two lemmas are well-known. We include their proofs here for the purpose of
completeness.
Lemma 3.1. Suppose that M is a type II1 von Neumann algebra. Then the following are
true.
(a) For any nonzero element x ∈ M, there exists a normal tracial state ρ on M such that
ρ(x∗x) 6= 0.
(b) There exists a non-zero central projection q of M, such that qM is a countably decom-
posable type II1 von Neumann algebra.
Proof. Assume that M acts on a Hilbert space H.
(a). Let Z be the center of M and τ be the unique, normal, faithful, center-valued trace
on M such that τ (a) = a for all a ∈ Z (see Theorem 8.2.8 in [5]). Let x ∈ M be a non-zero
element. Then, from the fact that τ is faithful, we know that τ (x∗x) 6= 0. Let ρ be a normal
state on Z such that ρ(τ (x∗x)) 6= 0. Now the normal state ρ on Z can be extended to a normal
tracial state ρ on M by defining ρ(a) = ρ(τ (a)) for all a ∈ M. Therefore ρ is a normal tracial
state on M such that ρ(x∗x) 6= 0.
(b). Let ρ be a normal tracial state on M and I = {a ∈ M : ρ(a∗a) = 0}. Thus I is a
2-sided ideal in M. Let (Hρ, πρ, xρ) be the triple obtained from the GNS construction of ρ such
that Hρ is a Hilbert space, πρ : M → B(Hρ) is a ∗-representation, xρ ∈ B(Hρ) is a cyclic vector
for πρ satisfying that, for every a ∈ M, ρ(a) = hπρ(a)xρ, xρi. It follows Proposition III.3.12
in [17] that πρ : M → B(Hρ) is a normal representation. Combining with the fact that I is
a two-sided ideal, we can check that I = ker(πρ). Therefore I is closed in M in ultraweak
operator topology. By Proposition 1.10.5 in [16], there exists a central projection q in Z such
that I = (1 − q)M.
Now we claim that qM is countably decomposable. Suppose that there is a collection
I and the choice of the central projection q, we get that ρ(qα) > 0 for each α ∈ J. Now
of nonzero orthogonal projections {qα : α ∈ J} in qM such that q = Pα qα. Since ρ is a
normal tracial state on M, we know that ρ(q) = Pα ρ(qα). By the definition of the ideal
1 = ρ(q + (I − q)) = ρ(q) = Pα ρ(qα), where I is the identity of M. It follows that J is a
countable set and thus qM is countably decomposable.
(cid:3)
Lemma 3.2. Suppose M is a type II1 von Neumann algebra. Then there is a family of
orthogonal central projections {qα : α ∈ Ω} in M with sum I such that qαM is countably
decomposable for each α ∈ Ω.
Proof. By Lemma 3.1 and Zorn's lemma, there exists an orthogonal family {qα} of non-zero
central projections in M, which is maximal with respect to the property that qαM is countable
Assume, to the contrary, that Q 6= I. Then by Lemma 3.1, there is a nonzero central projection
q in (I − Q)M such that qM is countably decomposable. The existence of such q contradicts
decomposable for each α. Let Q = P qα. We claim that Q = I, where I is the identity of M.
with the maximality of the family {qλ}. Therefore I = P qα and the proof of the lemma is
completed.
(cid:3)
6
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
Remark 3.3. Suppose M is a type II1 von Neumann algebra with Property Γ. Let q be a
central projection of M. Then it follows directly from the definition of Property Γ that qM also
has Property Γ.
Lemma 3.4. Let M be a type II1 von Neumann algbera acting on a separable Hilbert space
H and ZM the center of M. Let τ be the center-valued trace on M such that τ (z) = z for any
z ∈ ZM. Let M = RXL Msdµ and H = RXL Hsdµ be the direct integral decompositions of
M and H relative to ZM as in Lemma 2.2. Assume that Ms is a type II1 factor with a trace
τs for each s ∈ X. Then for any a ∈ M,
for almost every s ∈ X.
τ (a)(s) = τs(a(s))Is
Proof. Fix a ∈ M. By the Dixmier Approximation Theorem, for each t ∈ N, there
: t ∈ N, 1 ≤ j ≤ kt} in M and scalars
exist a positive integer kt, a family of unitaries {v(t)
j
{λ(t)
j
: t ∈ N, 1 ≤ j ≤ kt} ⊆ [0, 1] such that
(i) for each t ∈ N, P1≤j≤kt
(ii) lim
t→∞
k P1≤j≤kt
λ(t)
j = 1;
λ(t)
j (v(t)
j )∗av(t)
j − τ (a)k = 0.
Since {v(t)
j
: t ∈ N, 1 ≤ j ≤ kt} is a countable set, we may assume that, for every s ∈ X,
v(t)
j (s) is a unitary in Ms for any t ∈ N and any 1 ≤ j ≤ kt. By Proposition 14.1.9 in [5], for
any t ∈ N, we have
k X1≤j≤kt
It follows that
λ(t)
j (v(t)
j )∗av(t)
j − τ (a)k = ess- sup
s∈X
k X1≤j≤kt
λ(t)
j (v(t)
j (s))∗a(s)v(t)
j (s) − τ (a)(s)k.
lim
t→∞
k X1≤j≤kt
λ(t)
j (v(t)
j (s))∗a(s)v(t)
j (s) − τ (a)(s)k = 0
(3.1)
for almost every s ∈ X. Again, by the Dixmier Approximation Theorem and the fact that each
Ms is a type II1 factor, (3.1) gives that
τ (a)(s) = τs(a(s))Is
for almost every s ∈ X.
(cid:3)
Lemma 3.5. Let M be a type II1 von Neumann algebra with center ZM. Let τ be the center-
valued trace on M such that τ (a) = a for any a ∈ ZM. Suppose ǫ > 0, x ∈ M and τ (x∗x) < ǫI.
Then for any tracial state ρ on M,
ρ(x∗x) < 2ǫ.
Proof. Note that τ (x∗x) < ǫI. It follows from the Dixmier Approximation Theorem that
there exist a positive integer n ∈ N, a family of unitaries {v1, v2, . . . , vn} in M and a family of
scalars {α1, α2, . . . , αn} ⊆ [0, 1] such that
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
7
αi = 1;
(a) P1≤i≤n
(b) kτ (x∗x) − P1≤i≤n
αiv∗
i x∗xvik < ǫ.
Since ρ is tracial, it follows from (a) and (b) that
ρ(x∗x) = ρ( X1≤i≤n
= ρ( X1≤i≤n
< 2ǫ.
αiv∗
i x∗xvi)
αiv∗
i x∗xvi) − τ (x∗x)) + ρ(τ (x∗x))
The proof is completed.
(cid:3)
Proposition 3.6. Suppose M is a type II1 von Neumann algebra acting on a separable
Hilbert space H. Let τ be the center-valued trace on M such that τ (a) = a for any a ∈ ZM,
where ZM is the center of M. Suppose that M has Property Γ. Then, for a1, a2, . . . , ak ∈ M,
any n ∈ N, any ǫ > 0, there exist n orthogonal equivalent projections p1, p2, . . . , pn in M with
sum I such that
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k.
Proof. Suppose M has Property Γ. Let M = RXL Msdµ and H = RXL Hsdµ be the
direct integral decompositions of M and H relative to the center ZM as in Lemma 2.2. We
might assume that Ms is a type II1 factor with a trace τs for each s ∈ X.
Fix a1, a2, . . . , ak ∈ M, n ∈ N, and ǫ > 0. By Corollary 4.2 in [13], there exist n orthogonal
equivalent projections p1, p2, . . . , pn in M with sum I such that
kpi(s)aj(s) − aj(s)pi(s)k2,s ≤ ǫ/2,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k,
(3.2)
for almost every s ∈ X, where k · k2,s is the trace norm induced by τs on Ms for each s ∈ X.
For any 1 ≤ i ≤ n, 1 ≤ j ≤ k, Lemma 3.4 gives
τ ((piaj − ajpi)∗(piaj − ajpi))(s) = τs((pi(s)aj(s) − aj(s)pi(s))∗(pi(s)aj(s) − aj(s)pi(s)))Is (3.3)
for almost every s ∈ X.
For any 1 ≤ i ≤ n, 1 ≤ j ≤ k, from (3.2), (3.3) and Proposition 14.1.9 in [5], it follows that
and, thus,
This finishes the proof.
kτ ((piaj − ajpi)∗(piaj − ajpi))k ≤ ǫ/2
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI.
(cid:3)
Lemma 3.7. Let M be a type II1 von Neumann algebra with a center ZM. Let M1 be a von
Neumann subalgebra of M and ZM1 be the center of M1. Suppose τM and τM1 are the center-
valued traces of M, and M1 respectively. For any x ∈ M1, we have kτM(x)k ≤ kτM1(x)k.
8
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
Proof. Let x be an element in M1. For any ǫ > 0, by the Dixmier Approximation Theorem,
there exist a positive integer k, a family of unitaries {vj : 1 ≤ j ≤ k} in M1 and scalars
{λj : 1 ≤ j ≤ k} ⊆ [0, 1] such that (i) P1≤j≤k
Hence,
j xvj − τM1(x)k ≤ ǫ.
λjv∗
λj = 1 and (ii) k P1≤j≤k
j xvj! k
j xvj! − τM(τM1(x))k + kτM(τM1(x))k
λjv∗
λjv∗
kτM(x)k = kτM X1≤j≤k
≤ kτM X1≤j≤k
≤ ǫ + kτM1(x)k
Since ǫ is arbitrary, we have kτM(x)k ≤ kτM1(x)k.
(cid:3)
Proposition 3.8. Suppose M is a countably decomposable type II1 von Neumann algebra.
Let τ be the center-valued trace on M such that τ (a) = a for any a ∈ ZM, where ZM is the
center of M. Suppose that M has Property Γ. Then, for a1, a2, . . . , ak ∈ M, any n ∈ N, any
ǫ > 0, there exist n orthogonal equivalent projections p1, p2, . . . , pn in M with sum I such that
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k.
Proof. Let a1, a2, . . . , ak be in M. By Lemma 3.6 in [14], there is a type II1 von Neumann
algebra M1 with separable predual and Property Γ such that {a1, . . . , ak} ⊆ M1 ⊆ M. From
Proposition 3.6, it follows that there exist n orthogonal equivalent projections p1, p2, . . . , pn in
M1 with sum I such that
τM1((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k,
where τM1 is the center-valued trace on M1. By Lemma 3.7, we obtain that
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k.
(cid:3)
Remark 3.9. Suppose M is a type II1 von Neumann algbera with center ZM. Let τ be the
center-valued trace on M such that τ (a) = a for any a ∈ ZM. Suppose {qα : α ∈ Ω} is a
family of nonzero orthogonal central projections in M with sum I. Therefore qαM is a type II1
von Neumann algebra with center qαZM. Let τα be the center-valued trace on qαM such that
τα(a) = a for any a ∈ qαZM. We have
τ (a) =Xα∈Ω
τα(qαa),
∀ a ∈ M.
Theorem 3.10. Suppose M is a type II1 von Neumann algebra and ZM is the center of M.
Let τ be the center-valued trace on M such that τ (a) = a for any a ∈ ZM. Then the following
statements are equivalent:
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
9
(i) M has Property Γ.
(ii) There exists a family of nonzero orthogonal central projections {qα : α ∈ Ω} in M with
sum I such that qαM is a countably decomposable type II1 von Neumann algbera with
Property Γ for each α ∈ Ω.
(iii) For any n ∈ N, any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exist n orthogonal equivalent
projections p1, p2, . . . , pn in M with sum I such that
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k.
(iv) There exists a positive integer n0 ≥ 2 satisfying for any ǫ > 0 and a1, a2, . . . , ak ∈
M, there exist n0 orthogonal equivalent projections p1, p2, . . . , pn0 in M with sum I
satisfying
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n0, 1 ≤ j ≤ k.
(v) For any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exists a unitary u in M such that
(a) τ (u) = 0;
(b) τ ((uaj − aju)∗(uaj − aju)) < ǫI,
∀ 1 ≤ j ≤ k.
(vi) For any n ∈ N, any normal tracial state ρ on M, any ǫ > 0 and a1, a2, . . . , ak ∈ M,
there exist n orthogonal equivalent projections p1, p2, . . . , pn in M with sum I such that
kpiaj − ajpik2,ρ < ǫ,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k,
where k · k2,ρ is the 2-norm on M induced by ρ.
(vii) There exists a positive integer n0 ≥ 2 satisfying for any normal tracial state ρ on M,
any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exist n0 orthogonal equivalent projections
p1, p2, . . . , pn0 in M with sum I satisfying
kpiaj − ajpik2,ρ < ǫ,
∀ 1 ≤ i ≤ n0, 1 ≤ j ≤ k,
where k · k2,ρ is the 2-norm on M induced by ρ.
(viii) For any normal tracial state ρ on M, any ǫ > 0 and a1, a2, . . . , ak ∈ M, there exists a
unitary u in M such that
(a) τ (u) = 0;
(b) kuaj − ajuk2,ρ < ǫ, for all 1 ≤ j ≤ k, where k · k2,ρ is the 2-norm on M induced by
ρ.
Proof. We will prove the result by showing that (i)⇒ (ii) ⇒ (iii) ⇒ (iv) ⇒ (v)⇒ (ii),
(iii)⇒ (i) and (iii) ⇒ (vi) ⇒ (vii) ⇒ (viii) ⇒ (ii).
(i) ⇒ (ii): It follows from Lemma 3.2 and Remark 3.3.
(ii) ⇒ (iii): Assume that there exists a family of nonzero orthogonal central projections
{qα : α ∈ Ω} with sum I such that qαM is a countably decomposable type II1 von Neumann
algbera with Property Γ for each α ∈ Ω. Fix n ∈ N, any ǫ > 0 and a1, a2, . . . , ak ∈ M. Then
qαaj,
∀ 1 ≤ j ≤ n.
aj =Xα
10
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
For each α ∈ Ω, by Proposition 3.8, there exist n orthogonal equivalent projections p(α)
in qαM with sum qα such that
n
1 , p(α)
2 , . . . , p(α)
τα((p(α)
i
(qαaj) − (qαaj)p(α)
i
)∗(p(α)
i
(qαaj) − (qαaj)p(α)
i
)) < ǫ · qα,
(3.4)
for all 1 ≤ i ≤ n, 1 ≤ j ≤ k, where τα is the center-valued trace on qαM. Let
p(α)
i
,
for all 1 ≤ i ≤ n.
pi =Xα
Then it is not hard to see that p1, . . . , pn are orthogonal equivalent projections in M with sum
I. By Remark 3.9 and inequality (3.4), we know
τ ((piaj − ajpi)∗(piaj − ajpi)) < ǫI,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k.
(iii) ⇒ (iv): It is obvious.
(iv) ⇒ (v): Assume that there exists a positive integer n0 ≥ 2 satisfying for any ǫ > 0 and
a1, a2, . . . , ak ∈ M, there exist n0 orthogonal equivalent projections p1, p2, . . . , pn0 in M with
sum I satisfying
τ ((piaj − ajpi)∗(piaj − ajpi)) <
ǫ
n2
0
I,
∀ 1 ≤ i ≤ n0, 1 ≤ j ≤ k.
Let λ = e2πi/n0 be the n0-th root of unit. Let
u = p1 + λp2 + · · · λn0−1pn0.
Since p1, . . . , pn0 are orthogonal equivalent projections in M, we know τ (u) = 0. A quick
computation shows that
τ ((uaj − aju)∗(uaj − aju)) < ǫI.
(v)⇒ (ii): Assume that (v) holds. From Lemma 3.2, there is a family of orthogonal central
projections {qα : α ∈ Ω} in M with sum I such that qαM is countably decomposable for each
α ∈ Ω.
Next we will show that qαM has Property Γ for each α in Ω. Let x1, . . . , xk be elements
in qαM. By the assumption (v), for any ǫ > 0, there exists a unitary u in M such that
(a) τ (u) = 0 and (b) τ ((uaj − aju)∗(uaj − aju)) < ǫI, for all 1 ≤ j ≤ k. Since qαM is
a countably decomposable type II1 von Neumann subalgebra, there exists a faithful normal
tracial state ρ on qαM. We can naturally extend ρ on qαM to a normal tracial state ρ on M
by defining ρ(x) = ρ(qαx) for all x in M. It is not hard to see that qαu is a unitary in qαM and
τα(qαu) = τ (qαu) = qατ (u) = 0, where τα is a center-valued trace on qαM. Moreover, by the
Dixmier Approximation Theorem, we have
ρ(x) = ρ(τ (x)),
∀ x ∈ M.
Hence
ρ(((qαu)aj − aj(qαu))∗((qαu)aj − aj(qαu))) = ρ ((uaj − aju)∗(uaj − aju))
= ρ (τ ((uaj − aju)∗(uaj − aju)))
≤ ǫ,
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
11
for all 1 ≤ i ≤ k. By Proposition 3.5 in [14], we conclude that qαM has Property Γ.
(iii)⇒ (i): Assume that (iii) is true. We assume that M acts on a Hilbert space H. Let
x1, . . . , xk be a family of elements in M. From (iii), for any positive integer n, there exists a
family of projections {pir : 1 ≤ i ≤ n, r ≥ 1} in M such that
1. For each r ≥ 1, p1,r, . . . , pn,r are orthogonal equivalent projections in M with sum I.
2. Moreover,
lim
r→∞
kτ ((pi,raj − ajpi,r)∗(pi,raj − ajpi,r))k = 0,
∀ 1 ≤ i ≤ n, 1 ≤ j ≤ k.
Thus, for any normal tracial state ρ on M, we have
kρ((pi,raj − ajpi,r)∗(pi,raj − ajpi,r))k = lim
r→∞
≤ lim
r→∞
lim
r→∞
kρ(τ ((pi,raj − ajpi,r)∗(pi,raj − ajpi,r)))k
kτ ((pi,raj − ajpi,r)∗(pi,raj − ajpi,r))k
(3.5)
Let {ρλ}λ∈Λ be the collection of all normal tracial states on M. For each λ ∈ Λ, let (πλ, Hλ, Iλ)
be the GNS representation, obtained from ρλ, of M on the Hilbert space Hλ = L2(M, ρλ) with
= 0.
a cyclic vector Iλ in Hλ. We also let K = Pλ∈Λ Hλ be the direct sum of Hilbert spaces {Hλ}
and π = Pλ∈Λ πλ : M → B(K) be the direct sum of {πλ}. Thus π is a ∗-representation of M
on K defined by
It is not hard to see that π is a normal ∗-representation and π(M) is also a von Neumann
algebra. By Lemma 3.1 (a), π is a ∗-isomorphism from M onto π(M).
We claim that, for all 1 ≤ i ≤ n, 1 ≤ j ≤ k,
(pi,raj − ajpi,r)∗(pi,raj − ajpi,r) → 0 in ultraweak operator topology (or in σ(M, M♯) topology).
Actually, the claim is equivalent to the statement that
π((pi,raj − ajpi,r)∗(pi,raj − ajpi,r)) → 0 in ultraweak topology.
Note that
{(pi,raj − ajpi,r)∗(pi,raj − ajpi,r) : 1 ≤ i ≤ n, 1 ≤ j ≤ k, r ∈ N}
is a bounded subset in M. It will be enough if we are able to show that
π((pi,raj − ajpi,r)∗(pi,raj − ajpi,r)) → 0 in weak operator topology,
or
π(pi,raj − ajpi,r) → 0 in strong operator topology.
(3.6)
By the construction of π, (3.6) follows directly from (3.5).
From the claim in the preceding paragraph, by the definition of Property Γ, we know that
M has property Γ.
(iii)⇒ (vi): From the Dixmier Approximation Theorem, for any normal tracial state ρ on
M, we have
ρ(x) = ρ(τ (x))
∀ x ∈ M.
π(x)((ξλ)) = (πλ(x)ξλ),
∀ (ξλ) ∈Xλ∈Λ
Hλ = K.
12
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
Now (vi) follows easily from (iii).
(vi) ⇒ (vii): It is obvious.
(vii) ⇒ (viii): It is similar to (iv)⇒ (v).
(viii) ⇒ (ii): Assume that (viii) holds. From Lemma 3.2, there is a family of orthogonal
central projections {qα : α ∈ Ω} in M with sum I such that qαM is countably decomposable
for each α ∈ Ω. We need to show that qαM has Property Γ for each α in Ω.
Since each qαM is a countably decomposable type II1 von Neumann algebra. There exists
a faithful normal tracial state ρα on qαM. Then the normal tracial state ρα on qαM can be
naturally extended to a normal tracial state ρ on M by defining ρ(x) = ρα(qαx) for all x ∈ M.
Let ǫ > 0 and a1, . . . , ak be elements in qαM. Since (viii) holds, there exists a unitary u in M
such that
(a) τ (u) = 0;
(b) kuaj − ajuk2,ρ < ǫ, for all 1 ≤ j ≤ k, where k · k2,ρ is the trace norm induced by ρ on
M.
Now it is not hard to verify that qαu is a unitary in qαM satisfying τα(qαu) = τ (qαu) = 0, where
τα is the unique center-valued trace on qαM. Moreover
k(qαu)aj − aj(qαu)k2,ρα = k(qαu)aj − aj(qαu)k2,ρ
= kuaj − ajuk2,ρ
< ǫ.
From Proposition 3.5 in [14], it follows that qαM has Property Γ for each α in Ω. This ends
the whole proof.
(cid:3)
4. Similarity degree of type II1 von Neumann algebras with Property Γ
Let us recall a definition of Property c∗-Γ for unital C∗-algebras given in [14]. Suppose A is
a unital C∗-algebra. We say A has Property c∗-Γ if it satisfies the following condition:
If π is a unital ∗-representation of A on a Hilbert space H such that π(A)′′ is a type
II1 factor, then π(A)′′ has Property Γ.
If A is a separable unital C∗-algebra with Property c∗-Γ, Theorem 5.3 in [14] gives that the
similarity degree of A is no more than 3. Indeed, it was shown in Theorem 5.3 in [14] that, for
any C∗-algebra B, if φ is a bounded unital homomorphism from A to B, then kφkcb ≤ kφk3.
Lemma 4.1. Suppose M is a type II1 von Neumann algebra with Property Γ. Let τ be the
center-valued trace on M such that τ (a) = a for any a ∈ ZM, where ZM is the center of M.
Suppose F is a finite subset of M. Then there exists a separable unital C∗-subalgebra A with
Property c∗-Γ satisfying F ⊆ A ⊆ M.
Proof. Let F1 = F = {x1, x2, . . . , xk} be a finite subset of M. Since M has Property Γ,
by Theorem 3.10, there exists a 2 × 2 system of matrix units {e(1)
11 , e(1)
12 , e(1)
21 , e(1)
22 } such that
11 + e(1)
(i1) e(1)
(ii1) τ ((e(1)
22 = I, where I is the identity of M.
ii x − xe(1)
ii )∗(e(1)
ii x − xe(1)
ii )) ≤ 1
2I, for each x ∈ F1.
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
13
From (ii1), by the Dixmier Approximation Theorem, there exist a positive integer n1, a family
of unitaries v(1)
n1 in M such that
1 , v(1)
2 , . . . , v(1)
(iii1) For each 1 ≤ i ≤ 2 and each x ∈ F1, there is an element y in the convex hull of
{(v(1)
t )∗(e(1)
Let F2 = F1 ∪ {e(1)
ii x − xe(1)
12 , e(1)
11 , e(1)
ii )∗(e(1)
21 , e(1)
22 } ∪ {v(1)
ii x − xe(1)
ii )v(1)
1 , . . . , v(1)
t
n1 }.
: 1 ≤ t ≤ n1} with kyk < 1.
Assume that F1 ⊆ F2 ⊆ · · · ⊆ Fm have been constructed for some m ≥ 2. Since M has Prop-
21 , e(m)
22 }
erty Γ, again by Theorem 3.10, there exists a 2 × 2 system of matrix units {e(m)
such that
11 , e(m)
12 , e(m)
11 + e(m)
(im) e(m)
(iim) τ ((e(m)
22 = I, where I is the identity of M.
ii x − xe(m)
ii )∗(e(m)
ii x − xe(m)
ii )) ≤ 1
m+1 I, for each x ∈ Fm.
From (iim), by the Dixmier Approximation Theorem, there exist a positive integer nm, a family
of unitaries v(m)
nm in M such that
, . . . , v(m)
, v(m)
1
2
(iiim) For each 1 ≤ i ≤ 2 and each x ∈ Fm, there is an element y in the convex hull of
{(v(m)
t
)∗(e(m)
ii x − xe(m)
Let Fm+1 = Fm ∪ {e(m)
11 , e(m)
12 , e(m)
ii )∗(e(m)
21 , e(m)
ii x − xe(m)
22 } ∪ {v(m)
ii )v(m)
, v(m)
1
2
t
: 1 ≤ t ≤ nm} with kyk < 1
m.
, . . . , v(m)
nm }.
Continuing this process, we are able to obtain a sequence {Fm}, a sequence of system of
units {e(m)
11 , e(m)
12 , e(m)
21 , e(m)
22 } such that
(0) {x1, . . . , xk} = F1 ⊆ F2 ⊆ · · · ⊆ Fm ⊆ · · · .
(1) For each m ≥ 1, {e(m)
(2) τ ((e(m)
ii )∗(e(m)
(3) For each i = 1, 2 and each x ∈ Fm, there is an element
12 , e(m)
11 , e(m)
ii x − xe(m)
21 , e(m)
ii )) ≤ 1
ii x − xe(m)
m+1 I, for each x ∈ Fm.
22 } is a system of units such that e(m)
11 + e(m)
22 = I.
y ∈ conv{(v∗(e(m)
ii x − xe(m)
ii )∗(e(m)
ii x − xe(m)
ii )v : v is a unitary in Fm+1}
satisfying
kyk <
1
m
.
Let A be the unital C∗-algebra generated by ∪m∈NFm. Then A is separable and it follows
from the preceding construction that,
(4) for i = 1, 2 and any x ∈ A, there exists a sequence of elements {ym}m≥1 in A such
ii )v :
that, for m ≥ 1, each ym is in the convex hull of {v∗(e(m)
v is a unitary in A} and lim
m→∞
ii x − xe(m)
ii x − xe(m)
ii )∗(e(m)
kymk = 0.
Now we are going to show this C∗-subalgebra A of M has Property c∗-Γ. Suppose π is a
unital ∗-representation of A on a Hilbert space H such that π(A)′′ is a type II1 factor. Notice
that for each m ∈ N, {e(m)
22 } is a 2 × 2 system of matrix units in A. It follows
that {π(e(m)
12 ), π(e(m)
11 ), π(e(m)
22 )
are orthogonal equivalent projections in π(A)′′ with sum I.
21 , e(m)
22 )} is also a system of matrix units. Hence π(e(m)
11 , e(m)
12 , e(m)
21 ), π(e(m)
11 ), π(e(m)
It follows from Condition (4) that
14
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
(4') for i = 1, 2 and any x ∈ π(A), there exists a sequence of elements {ym}m≥1 in π(A)
such that, for m ≥ 1, each ym is in the convex hull of
{v∗(π(e(m)
ii )x − xπ(e(m)
kymk = 0.
and lim
m→∞
ii ))∗(π(e(m)
ii )x − xπ(e(m)
ii ))v : v is a unitary in π(A)}
Let ρ be the unique trace on π(A)′′. Since ρ is tracial, Condition (4') implies that, for any
x ∈ π(A),
ρ((π(e(m)
ii )π(x) − π(x)π(e(m)
ii )∗(π(e(m)
ii )π(x) − π(x)π(e(m)
ii ))
lim
m→∞
= lim
m→∞
ρ(ym)
= 0.
(4.1)
By Kaplansky Density Theorem, it follows from (4.1) that
lim
m→∞
ρ((π(e(m)
ii )a − aπ(e(m)
ii )∗(π(e(m)
ii )a − aπ(e(m)
ii )) = 0
for any a ∈ π(A)′′. Note that a type II1 factor is always countably decomposable. By Proposition
3.5 in [14], π(A)′′ has Property Γ, whence we conclude that A has Property c∗-Γ.
The proof is completed.
(cid:3)
It was shown in [2] that the similarity degree of a type II1 factor with Property Γ is 3. The
following theorem gives a generalization.
Theorem 4.2. If M is a type II1 von Neumann algebra with Property Γ, then the similarity
degree d(M) = 3.
Proof. Since M is a von Neumann algebra of type II1, by Corollary 1.9 in [19], it is not
nuclear. It follows from Theorem 1 in [11] that d(M) ≥ 3. In the following we show that d(M)
is no more than 3.
Suppose φ : M → B(H) is a bounde unital homomorphism, where H is a Hilbert space.
We will show that kφkcb ≤ kφk3. In fact we are going to prove that, for any n ∈ N and any
x = (xij) ∈ Mn(M),
kφ(n)(x)k ≤ kφk3kxk.
(4.2)
Fix n ∈ N and x = (xij) ∈ Mn(M). We assume that kxk = 1. Notice that F = {xij : 1 ≤
i, j ≤ n} is a finite subset of M. By Lemma 4.1, there is a separable unital C∗-subalgebra A of
M with Property c∗-Γ such that F ⊆ A. Let φ be the restriction of φ on A. Then φ : A → B(H)
is a bounded unital homomorphism. It was shown in the proof of Theorem 5.3 in [14] that
k φkcb ≤ k φk3.
Since F ⊆ A, it follows from (4.3) that
kφ(n)(x)k = k φ(n)(x)k ≤ k φk3 ≤ kφk3.
Therefore d(M) = 3 and the proof is completed.
(4.3)
(cid:3)
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
15
Based on Theorem 4.2, a slight modification of the proof of Theorem 5.2 in [14] gives the
next corollary.
Corollary 4.3. Let M be a von Neumann algebra with the type decomposition
M = M1 ⊕ Mc1 ⊕ Mc∞ ⊕ M∞,
where M1 is a type I von Neumann algebra, Mc1 is a type II1 von Neumann algebra, Mc∞ is
a type II∞ von Neumann algebra and M∞ is a type III von Neumann algebra. Suppose Mc1 is
a type II1 von Neumann algebra with Property Γ. If φ is a bounded unital representation of M
on a Hilbert space H, which is continuous from M, with the topology σ(M, M♯), to B(H), with
the topology σ(B(H), B(H)♯), then φ is completely bounded and kφkcb ≤ kφk3.
Suppose A is a unital C∗-algebra. Let I be some index set and
l∞(I, A) = {(xi)i∈J :
for each i ∈ I, xi ∈ A and sup
i∈I
kxik < ∞}.
It was shown in [9] (Corollary 17) that if M is a type II1 factor with Property Γ, then
d(l∞(I, M)) ≤ 5 for any index set I. The next corollary gives an exact value of d(l∞(I, M)).
Corollary 4.4. If M is a type II1 factor with Property Γ, then d(l∞(I, M)) = 3 for any
index set I.
Proof. Assume that M is a type II1 factor with Property Γ. By Theorem 3.10, for any
index set I, l∞(I, M) is a type II1 von Neumann algebra with Property Γ. Therefore
d(l∞(I, M)) = 3.
(cid:3)
Let C be the CAR-algebra C = M2(C) ⊗ M2(C) ⊗ . . . (infinite C∗-tensor product of 2 × 2
matrix algebras). It was shown in [9] (Proposition 21) that, for any index set I, d(l∞(I, C)) ≤ 5.
The next corollary gives an exact value of d(l∞(I, C)).
Corollary 4.5. Let C = M2(C) ⊗ M2(C) ⊗ . . . (infinite C∗-tensor product of 2 × 2 matrix
algebras). Then, for any infinite index set I, d(l∞(I, C)) = 3.
⊕Ci, where Ci is a copy of C for each i ∈ I. Let R
and Ri be the canonical hyperfinite II1 factor generated by C and Ci respectively. Let τi be a
⊕Ri. We might assume that both M and A act naturally
Proof. Denote by A = l∞(I, C) = Pi∈I
trace on Ri. Let M = l∞(I, R) = Pi∈I
on the Hilbert space Pi∈I l2(Ri, τi). Denote by pi the projection in A such that piA = Ci. It
follows that Pi∈I
pi = I.
First we will prove the following two claims.
Claim 4.5.1. For any x1, . . . , xk in A and any ǫ > 0, there exists a system of matrix units
{Est : 1 ≤ s, t ≤ 2} in A such that E11 + E22 = I and
kxjEss − Essxjk < ǫ,
for all 1 ≤ s ≤ 2, 1 ≤ j ≤ k.
16
WENHUA QIAN, DON HADWIN, AND JUNHAO SHEN
Proof of Claim 4.5.1: For each i ∈ I, note that pix1, . . . , pixk are in a CAR algebra Ci. Hence
there exists a system of matrix units {e(i)
st : 1 ≤ s, t ≤ 2} in Ci such that e(i)
22 = pi and
11 + e(i)
kxje(i)
ss − e(i)
ss xjk < ǫ/2,
for all 1 ≤ s ≤ 2, 1 ≤ j ≤ k.
Let
Then {Est : 1 ≤ s, t ≤ 2} is a system of matrix units in A such that E11 + E22 = I and
e(i)
st ,
for all 1 ≤ s, t ≤ 2.
Est =Xi∈I
kxjEss − Essxjk = sup
i
kxje(i)
ss − e(i)
ss xjk < ǫ,
for all 1 ≤ s ≤ 2, 1 ≤ j ≤ k.
This finishes the proof of Claim 4.5.1.
Claim 4.5.2. For any x1, . . . , xk in A, there exists a separable C∗-subalgebra B of A such
that B is of Property c∗-Γ and all x1, . . . xk are in B.
Proof of Claim 4.5.2: Let F1 = {x1, . . . , xk}. By Claim 4.5.1, there exists a system of matrix
units {E(1)
st
: 1 ≤ s, t ≤ 2} in A such that E(1)
11 + E(1)
22 = I and
kxE(1)
ss − E(1)
ss xk < 1.
for all 1 ≤ s ≤ 2, 1 ≤ j ≤ k, and x ∈ F1.
Let F2 = F1 ∪ {E(1)
st
: 1 ≤ s, t ≤ 2}.
Assume that F1 ⊆ F2 ⊆ · · · ⊆ Fm have been constructed for some m ≥ 2. By Claim 4.5.1, we
22 = I
: 1 ≤ s, t ≤ 2} in A such that E(m)
11 + E(m)
know there exists a system of matrix units {E(m)
and
st
kxE(m)
ss − E(m)
ss xk <
.
for all 1 ≤ s ≤ 2, 1 ≤ j ≤ k, and x ∈ Fm.
1
m
Let Fm+1 = Fm ∪ {E(m)
st
: 1 ≤ s, t ≤ 2}.
Using similar arguments as in Lemma 4.1, we are able to obtain an increasing sequence of
subsets {Fm} of A such that (a) the C∗-subalgebra B generated by {Fm} in A is of Property
c∗-Γ; and (b) all x1, . . . xk are in B. This ends the proof of Claim 4.5.2.
(Continue the proof of Corollary:) From Claim 4.5.2, using similar arguments as in Theorem
4.2, we conclude that d(A) ≤ 3.
Next we will show that d(A) ≥ 3. Since I is an infinite set, let I0 be a countable infinite
subset of I. Then (Pi∈I\I0 pi)A is a closed two sided ideal of A. Moreover, (Pi∈I0 pi)A ∼=
A/(Pi∈I\I0 pi)A. By Remark 6 in [8], we know that d(A) ≥ d((Pi∈I0 pi)A). In order to show
that d(A) ≥ 3, it suffices to show that d((Pi∈I0 pi)A) ≥ 3. By replacing I by I0, we can assume
that I = N.
Let ω be a free ultra-filter of N and
J = {(xi) ∈ M(= l∞(N, R) =Xi∈I
⊕Ri) : lim
i→ω
τi(x∗
i xi) = 0}
SIMILARITY DEGREE OF TYPE II1 VON NEUMANN ALGEBRAS WITH PROPERTY Γ
17
be a closed two sided ideal of M. By Theorem 7.1 in [15], M/J is a type II1 factor. By Remark
12 in [9], d(M/J ) ≥ 3.
Let q : M → M/J be the quotient map. For any element (xi) ∈ M, by Kaplansky
Density Theorem, there exists an element (xi) ∈ A such that q((xi)) = q((xi)). In other words,
q(A) = M/J . By Remark 6 in [8], we get that d(A) ≥ d(M/J ). Combining with the result
from the preceding paragraph, we conclude that d(A) ≥ 3.
Therefore d(l∞(I, C)) = d(A) = 3, when I is an infinite set.
(cid:3)
References
[1] E. Christensen, On nonselfafjoint representations of C∗-algebras, Amer. J. Math. 103(1981), 817-833.
[2] E. Christensen, Finite von Neumann algebra factors with Property Γ, J. Funct. Anal. 186(2001), 366-380.
[3] U. Haagerup, Solution of the similarity problem for cyclic representations of C∗-algebras, Annals of Math.
118(1983), 215-240.
[4] M. Johanesov´a and W. Winter, The similarity degree for Z-stable C∗-algebras, Bull. Lond. Math. Soc.
44(6)(2012), 1215-1220.
[5] R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Algebras I,II, Academic Press,
Orlando, 1983, 1986.
[6] F.J. Murray and J. von Neumann, On rings of operators IV, Ann. of Math. (2) 44(1943), 716-808.
[7] G. Pisier, The similarity degree of an operator algebra, St. Petersburg Math. J. 10(1999), 103-146.
[8] G. Pisier, Similarity problems and length, Taiwanese Journal of Math. 5(2001), 1-17.
[9] G. Pisier, Remarks on the similarity degree of an operator algebra, Int. J. Math. 12, 403(2001), 403-414.
[10] G.Pisier, Similarity problems and completely bounded maps, Lecture notes in Mathematics, 1618. Springer-
Verlag, Berlin, 2001.
[11] G. Pisier, A similarity degree characterization of nuclear C∗-algebras, Publ. Res. Inst. Math. Sci. 42(2006),
No. 3, 691-704.
[12] F. Pop, The similarity problem for tensor products of certain C∗-algebras, Bulletin of the Australian Math-
ematical Society 70(03) 2004, 385 - 389.
[13] W. Qian and J. Shen, Hochschild cohomology of type II1 von Neumann algebras with Property Γ, Preprint.
[14] W. Qian and J. Shen, Similarity degree of a class of C∗-algebras, Preprint.
[15] S. Sakai, The theory of W∗-algebras, lecture notes, Yale University (1962).
[16] S. Sakai, C∗-algebras and W∗-algebras, Springer-Verlag 1971.
[17] M. Takesaki, Theory of Operator Algebras I, Springer 1979.
[18] J. von Neumann, On rings of operators. Reduction theory, Ann. of Math. 50(1949), 401-485.
[19] S. Wasserman, On tensor products of certain group C∗-algebras, Funct. Analysis, 23(1976), 239-254.
Wenhua Qian, Demartment of Mathematics and Statistics, University of New Hampshire,
Durham, NH 03824; Email: [email protected]
Don Hadwin, Demartment of Mathematics and Statistics, University of New Hampshire,
Durham, NH 03824; Email: [email protected]
Junhao Shen, Demartment of Mathematics and Statistics, University of New Hampshire,
Durham, NH 03824; Email: [email protected]
|
1202.1175 | 1 | 1202 | 2012-02-06T15:39:26 | Faithful compact quantum group actions on connected compact metrizable spaces | [
"math.OA",
"math.QA"
] | We construct faithful actions of quantum permutation groups on connected compact metrizable spaces. This disproves a conjecture of Goswami. | math.OA | math |
FAITHFUL COMPACT QUANTUM GROUP ACTIONS
ON CONNECTED COMPACT METRIZABLE SPACES
HUICHI HUANG
Abstract. We construct faithful actions of quantum permutation
groups on connected compact metrizable spaces. This disproves a
conjecture of Goswami.
1. Introduction
Compact quantum groups were introduced by Woronowicz in [13, 14].
They are noncommutative analogues of compact groups. Among all
literatures related to compact quantum groups, one particularly inter-
esting topic is the compact quantum group actions on commutative
or non-commutative unital C*-algebras (from the viewpoint of non-
commutative topology, that means actions on commutative or non-
commutative compact spaces). The actions of compact quantum groups
are the natural generalizations of actions of compact groups. It was
Podle´s who first formulated the concept of compact quantum group
actions, then established some basic properties [7]. Later, Wang intro-
duced the quantum permutation groups [11] and showed that they are
the universal compact quantum groups acting on finite spaces. After
that, many interesting actions are studied (see [1 -- 7] and the references
therein). But so far, all known (commutative) compact spaces admit-
ting genuine faithful compact quantum group actions are disconnected.
In [6], Goswami showed that there is no genuine faithful quantum iso-
metric action of compact quantum groups on the Riemannian man-
ifold G/T where G is a compact, semisimple, centre-less, connected
Lie group with a maximal torus T and conjectured that the quantum
permutations on (disconnected) finite sets are the only possible faithful
actions of genuine compact quantum groups on classical spaces. In this
paper, we construct faithful actions of quantum permutation groups on
Date: January 29, 2012.
2010 Mathematics Subject Classification. Primary 46L65, 16W22.
Key words and phrases. Compact quantum group, quantum permutation group,
connected compact metrizable space.
1
connected compact metrizable spaces and disprove Goswami's conjec-
ture.
The paper is organized as follows. In the next section we recall some
basic definitions and terminologies related to compact quantum groups
and their actions. Then in section 3, we construct faithful quantum
permutation group actions on connected compact metrizable spaces.
2. Preliminaries
In this section, we recall some definitions about compact quantum
groups. See [7, 10, 11, 13, 14] for more details. Throughout this paper,
the notation A ⊗ B for two unital C*-algebras A and B stands for
the minimal tensor product of A and B. For a ∗-homomorphism β :
B → B ⊗ A, use β(B)(1 ⊗ A) to denote the linear span of the set
{β(b)(1B ⊗ a)b ∈ B, a ∈ A} and β(B)(B ⊗ 1) to denote the linear
span of the set {β(b1)(b2 ⊗ 1A)b1, b2 ∈ B}. Denote by C the set of
complex numbers. For a compact Hausdorff space X and a unital C*-
algebra A, denote by C(X, A) the C*-algebra of continuous functions
mapping from X to A. Especially, when A = C, we write C(X, C) as
C(X). For x ∈ X, use evx to denote the evaluation functional on C(X)
at the point x.
Definition 2.1 (Definition 1.1 in [14]). A compact quantum group is a
unital C*-algebra A together with a unital ∗-homomorphism ∆ : A →
A ⊗ A such that
(1) (∆ ⊗ id)∆ = (id ⊗ ∆)∆;
(2) ∆(A)(1 ⊗ A) and ∆(A)(A ⊗ 1) are dense in A ⊗ A.
If A is non-commutative, we say that (A, ∆) is a genuine compact
quantum group. We say a unital C*-subalgebra Q of A is a compact
quantum quotient group of (A, ∆) if ∆(Q) ⊆ Q ⊗ Q, and ∆(Q)(1 ⊗ Q)
and ∆(Q)(Q ⊗ 1) are dense in Q ⊗ Q. That is, (Q, ∆Q) is a compact
If Q 6= A, we call Q a proper
quantum group [10, Definition 2.9].
compact quantum quotient group.
Definition 2.2 (Definition 1.4 in [7]). An action of a compact quantum
group (A, ∆) on a unital C*-algebra B is a unital ∗-homomorphism
α : B → B ⊗ A satisfying that
(1) (α ⊗ id)α = (id ⊗ ∆)α;
(2) α(B)(1 ⊗ A) is dense in B ⊗ A.
When A is non-commutative, we call α a genuine compact quantum
group action. We say that α is faithful if there is no proper compact
quantum quotient group Q of A such that α induces an action αq of
2
Q on B satisfying α(b) = αq(b) for all b in B [11, Definition 2.4]. An
action α is called ergodic if {b ∈ Bα(b) = b ⊗ 1A} = C1B. If A acts on
C(X) for a compact Hausdorff space X, we say that A acts on X.
i=1 aij = Pn
erated by aij for 1 ≤ i, j ≤ n under the relations a∗
and Pn
∗-homomorphism satisfying that ∆n(aij) = Pn
For any positive integer n, let An be the universal C*-algebra gen-
ij = aij = a2
ij
j=1 aij = 1. Let ∆n : An → An ⊗ An be the
k=1 aik ⊗ akj. Then
(An, ∆n) is a compact quantum group and is called a quantum permu-
tation group [11, Theorem 3.1]. Moreover, the quantum permutation
group An is a genuine quantum group when n ≥ 4 [12, The example
before Theorem 6.2].
Let Xn = {x1, x2, . . . , xn} be the finite space with n points. Define
ei for 1 ≤ i ≤ n to be the function on Xn such that ei(xj) = δij for
1 ≤ j ≤ n. There is a compact quantum group action αn : C(Xn) →
k=1 ek ⊗ aki for all 1 ≤ i ≤ n [11,
C(Xn) ⊗ An such that αn(ei) = Pn
Theorem 3.1].
3. Main Results
Let Y be a compact Hausdorff space.
Lemma 3.1. There exists an action α of the quantum permutation
k=1 ek ⊗ f ⊗ aki for all
group An on Xn × Y given by α(ei ⊗ f ) = Pn
1 ≤ i ≤ n and f ∈ C(Y ).
Proof. The quantum permutation group An acts on Xn by αn. Note
that C is also a compact quantum group. We denote the trivial action
of C on Y by σ. Then An ⊗ C is also a compact quantum group
and acts on Xn × Y [9, Theorem 2.1]. This action, denoted by α, is
defined by α := σ23(αn ⊗ σ), where σ23 : C(Xn) ⊗ An ⊗ C(Y ) ⊗ C →
C(Xn) ⊗ C(Y ) ⊗ An ⊗ C denotes the operator flipping the 2nd and 3rd
components. Note that An ⊗ C ∼= An. Therefore, we can say that An
acts on Xn × Y by α. Since αn(ei) = Pn
k=1 ek ⊗ aki and σ(f ) = f ⊗ 1,
we have that α(ei ⊗ f ) = Pn
k=1 ek ⊗ f ⊗ aki for all 1 ≤ i ≤ n and
(cid:3)
f ∈ C(Y ).
Let Y1 be a closed subset of Y . We define an equivalence relation
∼ on Xn × Y as follows. For y′, y′′ in Y and x′, x′′ in Xn, two points
(x′, y′) and (x′′, y′′) in Xn × Y are equivalent if one of the following is
true:
(1) y′ = y′′ ∈ Y1;
(2) y′ = y′′ and x′ = x′′.
Lemma 3.2. The quotient space Xn ×Y / ∼ is compact and Hausdorff.
3
Proof. For convenience, denote Xn ×Y by Z. The compactness of Z/ ∼
follows from the compactness of Z. To show that Z/ ∼ is Hausdorff,
it suffices to show that the subset R := {(z1, z2) ∈ Z 2z1 ∼ z2} of Z 2
is closed [8, Theorem 8.2]. Let (z1, z2) be in Z 2\R. Thus z1 ≁ z2. Use
(x′, y′) and (x′′, y′′) to denote z1 and z2 respectively.
Case 1. If y′ 6= y′′, then there exist two open subsets U and V of Y
such that y′ ∈ U, y′′ ∈ V and U ∩ V = ∅. Thus (Xn × U) × (Xn × V )
is an open neighborhood of (z1, z2) and is disjoint with R.
Case 2. If y′ = y′′, then x′ 6= x′′, and y′ /∈ Y1. Since Y1 is closed and
Y is compact Hausdorff, there exists an open subset U of Y containing
y′ and U is disjoint with Y1. Consequently ({x′} × U) × ({x′′} × U) is
an open neighborhood of (z1, z2) and is disjoint with R.
Combining Case 1 and Case 2, we obtain that Z 2\R is open. Hence
(cid:3)
R is closed and Z/ ∼ is Hausdorff.
Lemma 3.3. If an element F of C(Xn × Y ) satisfies that F (xi, y′) =
F (xj, y′) for some y′ in Y and all 1 ≤ i, j ≤ n, then α(F )(xk, y′) =
F (xj, y′)1An for all 1 ≤ j, k ≤ n.
Proof. Note that F can be written as Pn
i=1 ei ⊗ fi where f1, ..., fn are
in C(Y ). If F (xi, y′) = F (xj, y′), then fi(y′) = fj(y′). We obtain that
α(F )(xk, y) = (evk ⊗ evy ⊗ id)α(
nX
i=1
ei ⊗ fi)
=
=
=
nX
i=1
(evk ⊗ evy ⊗ id)α(ei ⊗ fi)
nX
nX
i=1
l=1
(evk ⊗ evy ⊗ id)(el ⊗ fi ⊗ ali)
nX
i=1
fi(y)aki
for any y ∈ Y and 1 ≤ k ≤ n. Since fi(y′) = fj(y′) for all 1 ≤ i, j ≤ n,
and Pn
i=1 aki = 1An for all 1 ≤ k ≤ n, we get
α(F )(xk, y′) =
nX
i=1
fi(y′)aki = fj(y′)
nX
i=1
aki = fj(y′)1An = F (xj, y′)1An
for all 1 ≤ j, k ≤ n. This completes the proof.
(cid:3)
Note that C(Xn × Y / ∼) is a C*-subalgebra of C(Xn × Y ).
Proposition 3.4. When the action α is restricted on C(Xn × Y / ∼),
it induces an action eα of An on Xn × Y / ∼.
4
Proof. We first prove the following:
(1)
α(C(Xn × Y / ∼)) ⊆ C(Xn × Y / ∼) ⊗ An.
Since C(Xn × Y / ∼) ⊗ An
∼= C(Xn × Y / ∼, An) and C(Xn × Y ) ⊗
∼= C(Xn × Y, An), an element c of C(Xn × Y ) ⊗ An belongs to
An
C(Xn ×Y / ∼)⊗An if and only if (evk ⊗evy ⊗id)(c) = (evl ⊗evy ⊗id)(c)
for all 1 ≤ k, l ≤ n and y ∈ Y1.
Therefore, to prove (1), it suffices to show that
(evk ⊗ evy ⊗ id)α(F ) = (evl ⊗ evy ⊗ id)α(F )
for all 1 ≤ k, l ≤ n, y ∈ Y1 and F in C(Xn × Y / ∼).
Let F be in C(Xn × Y / ∼). Then F can be written as Pn
i=1 ei ⊗ fi
for fi ∈ C(Y ) satisfying that fi(y) = fj(y) for all 1 ≤ i, j ≤ n and
y ∈ Y1. By Lemma 3.3, we have
(evk ⊗ evy ⊗ id)α(
nX
i=1
ei ⊗ fi) = fj(y)1An = (evl ⊗ evy ⊗ id)α(
nX
i=1
ei ⊗ fi)
for all y ∈ Y1 and 1 ≤ j, k, l ≤ n. This proves (1).
Next we show that α(C(Xn×Y / ∼))(1⊗An) is dense in C(Xn×Y / ∼
) ⊗ An.
α(F )(1 ⊗ a) by G. Note that F can be written as Pn
It is enough to show that F ⊗ a is in the closure of α(C(Xn × Y / ∼
))(1 ⊗ An) for all F in C(Xn × Y / ∼) and a in An. Denote F ⊗ a −
i=1 ei ⊗ fi for
fi ∈ C(Y ) satisfying that fi(y) = fj(y) for all 1 ≤ i, j ≤ n and y ∈ Y1.
By Lemma 3.3, we have α(F )(xi, y) = F (xj, y)1An for all xi, xj in Xn
and y in Y1. Thus GXn×Y1 = 0. For arbitrary ε > 0, let U be an
open subset of Y containing Y1 and satisfying that kG(xi, y)k < ε for
all (xi, y) ∈ Xn × U. By Urysohn's Lemma, there exists an f in C(Y ),
such that f Y1 = 0, f Y \U = 1 and 0 ≤ f ≤ 1. Denote 1⊗f ∈ C(Xn×Y )
by Hε. Then HεXn×Y1 = 0 and Hε(xi, y) = 1 for all xi in Xn and y in
Y \U. It follows from Lemma 3.3 that (α(Hε)G − G)(xi, y) = 0 for all
xi ∈ Xn and y ∈ Y \U. Since 0 ≤ Hε ≤ 1, for (xi, y) ∈ Xn × U, we
have
k(α(Hε)G − G)(xi, y)k ≤ kα(Hε) − 1kkG(xi, y)k < ε.
Hence kα(Hε)G − Gk < ε. Moreover, since α is an action of An on
Xn×Y , we have that α(C(Xn ×Y ))(1⊗An) is dense in C(Xn×Y )⊗An.
So there exist Fi ∈ C(Xn × Y ) and ai ∈ An for 1 ≤ i ≤ m where m is a
i=1 α(Fi)(1⊗ai)k < ε. It follows from
positive integer such that kG−Pm
5
0 ≤ Hε ≤ 1 that kα(Hε)G − α(Hε)Pm
i=1 α(Fi)(1 ⊗ ai)k < ε. Hence
kF ⊗ a − α(F )(1 ⊗ a) −
mX
i=1
α(HεFi)(1 ⊗ ai)k
= kG −
mX
i=1
α(HεFi)(1 ⊗ ai)k
≤ kG − α(Hε)Gk + kα(Hε)G −
mX
i=1
α(HεFi)(1 ⊗ ai)k < 2ε.
Note that HεFiXn×Y1 = 0 for all 1 ≤ i ≤ m. Thus HεFi is in C(Xn ×
Y / ∼) for all 1 ≤ i ≤ m. It follows that α(F )(1⊗a)+Pm
i=1 α(HεFi)(1⊗
ai) is in α(C(Xn×Y / ∼))(1⊗An). Since ε > 0 is arbitrary, we conclude
that F ⊗a is in the closure of α(C(Xn×Y / ∼))(1⊗An). This completes
the proof.
(cid:3)
Theorem 3.5. If Y1 6= Y , the action eα of An on Xn × Y / ∼ is faithful.
Proof. Suppose Y1 6= Y . Take a point y0 in Y but not in Y1. Since Y
is compact Hausdorff, there exists f ∈ C(Y ) such that f (y0) = 1 and
f Y1 = 0. Note that ei ⊗ f is in C(Xn × Y / ∼) for any 1 ≤ i ≤ n.
Suppose Q is a compact quantum quotient group of An such that eα is
an action of Q on Xn × Y / ∼. Then for any 1 ≤ k ≤ n,
(evk ⊗ evy0 ⊗ id)α(ei ⊗ f ) = (evk ⊗ evy0 ⊗ id)(
nX
l=1
el ⊗ f ⊗ ali)
= f (y0)aki = aki
is in Q. Since i is arbitrarily chosen, we get aki ∈ Q for any 1 ≤ k, i ≤ n.
(cid:3)
Thus Q = An. Therefore eα is faithful.
Proposition 3.6. If Y contains at least two points, the action eα is not
ergpdic.
Proof. Since Y consists of at least two points, there exist a non constant
function f ∈ C(Y ). Then 1 ⊗ f is in C(Xn × Y \ ∼) and not constant.
Also
This shows that eα is not ergodic.
eα(1 ⊗ f ) = 1 ⊗ f ⊗ 1.
(cid:3)
Proposition 3.7. If Y is connected and Y1 is nonempty, then Xn ×
Y / ∼ is connected.
Proof. As before, denote Xn × Y by Z. Take any nonempty closed and
open subset U of Z/ ∼. Denote by π the quotient map from Z onto
6
Z/ ∼. It follows that π−1(U) is a nonempty, closed and open subset of
Z. Since Xn is finite, we obtain that π−1(U) = Sxi∈X ′{xi} × Ai where
X ′ is a nonempty subset of Xn, and every Ai is a nonempty closed and
open subset of Y . Since Y is connected, we have Ai = Y for all xi ∈ X ′.
Take y ∈ Y1 and xi ∈ X ′. Let xj ∈ Xn. Then π(xj, y) = π(xi, y) ∈ U.
Thus xj is in X ′. Therefore X ′ = Xn and U = Z/ ∼. So Z/ ∼ is
connected.
(cid:3)
By Theorem 3.5 and Proposition 3.7, if we take a nonempty proper
closed subset Y1 of a connected compact Hausdorff space Y , then we
get a faithful action of An on a compact connected space Xn × Y / ∼.
To be more specific, we list some examples of Xn × Y / ∼.
Example 3.8.
(1) If Y = [0, 1] and Y1 = {0}, then Xn × Y / ∼ is a wedge sum
of n unit intervals by identifying (xi, 0) to a single point for all
1 ≤ i ≤ n. In this case Xn × Y / ∼ is a contractible compact
metrizable space.
(2) If Y = S 1 is a circle, and Y1 = {y0} for some point y0 in S 1, then
Xn × S 1/ ∼ will be the n circles touching at a point, which is a
connected compact metrizable space whose fundamental group
is the free group with n generators.
Remark 3.9. By Theorem 3.5, the quantum permutation group An
can act on the spaces in Example 3.8 faithfully. When n ≥ 4, this gives
us faithful genuine compact quantum group actions on connected com-
pact metrizable spaces. This disproves the conjecture of Goswami [6]
mentioned in the introduction. However, Proposition 3.6 tells us that
these faithful genuine compact quantum actions on compact connected
spaces are not ergodic. For this reason, we ask the following question:
Question 3.10. Are there any faithful ergodic genuine quantum group
actions on compact connected spaces?
Acknowledgement
The author is grateful to Professor Hanfeng Li for his long-term
support and encouragement. During the writing of this paper, the
author benefits a lot from many helpful discussions with him.
References
[1] T. Banica. Quantum automorphism groups of small metric spaces. Pa-
cific J. Math. 219(2005), no. 1, 27 -- 51.
[2] T. Banica. Quantum automorphism groups of homogeneous graphs. J.
Funct. Anal. 224 (2005), no. 2, 243 -- 280.
7
[3] J. Bhowmick and D. Goswami. Quantum group of orientation preserv-
ing Riemannian isometries. J. Funct. Anal. 257 (2009), 2530 -- 2572.
[4] J. Bhowmick, D. Goswami and A. Skalski. Quantum isometry groups
of 0-dimensional manifolds. Trans. Amer. Math. Soc. 363 (2011), no.
2, 901 -- 921.
[5] D. Goswami. Quantum Group of isometries in Classical and Non Com-
mutative Geometry. Comm. Math. Phys. 285 (2009), no. 1, 141C160.
[6] D. Goswami. Rigidity of action of compact quantum groups.
arXiv:1106.5107
[7] P. Podle´s. Symmetries of quantum spaces. Subgroups and quotient
spaces of quantum SU(2) and SO(3) groups. Comm. Math. Phys. 170
(1995), no.1, 1 -- 20.
[8] J. Rotman. An Introduction to Algebraic Topology. Graduate Texts in
Mathematics, 119. Springer-Verlag, New York, 1988.
[9] S. Wang. Tensor products and crossed products of compact quantum
groups. Proc. London Math. Soc. 71 (1995), no.3, 695 -- 720.
[10] S. Wang. Free products of compact quantum groups. Comm. Math.
Phys. 167 (1995), 671 -- 692.
[11] S. Wang. Quantum symmetry groups of finite spaces. Comm. Math.
Phys. 915 (1998), no. 1, 195 -- 211.
[12] S. Wang. Ergodic actions of universal quantum groups on operator
algebras. Comm. Math. Phys. 203 (1999), no. 2, 481 -- 498.
[13] S. L. Woronowicz. Compact matrix pseudogroups. Comm. Math. Phys.
111 (1987), no. 4, 613 -- 665.
[14] S. L. Woronowicz. Compact quantum groups. In:Sym´etries Quantiques,
Les Houches, 1995, North-Holland, Amsterdam, 1998, pp.845 -- 884.
Department of Mathematics, SUNY at Buffalo, Buffalo, NY 14260,
U.S.A.
E-mail address: [email protected]
8
|
1203.2176 | 4 | 1203 | 2012-08-18T00:26:54 | Some results on continuous deformed free group factors | [
"math.OA"
] | We construct a Fock space associated to a symmetric function $Q:U\times U \to (-1,1)$, where $U$ is a nonempty open subset of $\mathbb R^j$ for some $j$. Namely, we will have operator-valued distributions $a(x)$ and $a^+(y)$ satisfying $a(x)a^+(y)-Q(x,y)a^+(y)a(x)=\delta(x-y)$. Analogous to the $q_{ij}$-Fock space of Bozejko and Speicher, we have field operators arising as the sum of the creation and annihilation operators. These operators generate a von Neumann algebra analogous to the free group factors, and we will show that they are factors which do not have property $\Gamma$. It was pointed out to us by an anonymous referee that this is a special case of a theorem of Krolak. | math.OA | math |
SOME RESULTS ON CONTINUOUS DEFORMED FREE GROUP FACTORS
ADAM MERBERG
Abstract. We construct a Fock space associated to a symmetric function Q : U × U → (−1, 1), where U
is a nonempty open subset of Rj for some j. Namely, we will have operator-valued distributions a(x) and
a+(y) satisfying
a(x)a+(y) − Q(x, y)a+(y)a(x) = δ(x − y).
Analogous to the qij-Fock space of Bozejko and Speicher [3], we have field operators arising as the sum of
the creation and annihilation operators. These operators generate a von Neumann algebra analogous to the
free group factors, and are factors which do not have property Γ. It was pointed out to us by an anonymous
referee that this is a special case of a theorem of Krolak [6].
1. Introduction
In the study of operator algebras, much attention has been paid to the canonical commutation relations
(CCR) and the canonical anti-commutations (CAR). Bozejko and Speicher [2] considered an interpolation be-
tween these relations. Specifically, for q ∈ [−1, 1], they constructed creation operators c+(f ) and annihilation
operators c(f ) on a q-twisted Fock space Fq(H) satisfying the relations
c(f )c+(g) − qc+(g)c(f ) = hf, gi · 1.
In the q = 0 case, these are the creation and annihilation operators on the full Fock space.
It was shown by Voiculescu [10] that for a Hilbert space of dimension d ∈ {1, 2, . . . ,∞}, the Hermitian
parts of the creation operators on the free Fock space generate von Neumann algebras isomorphic to the
free group factor on d generators. Thus, we can view the algebras Γq(H) := {c(f ) + c+(f ) : f ∈ H}′′ as
q-deformations of the free group factors.
Various factoriality theorems have been proven for these algebras. First, Bozejko and Speicher [3] showed
that these are factors when dimH is infinite. ´Sniady [9] subsequently showed that Γq(H) is a factor for
dimH sufficiently large but finite. Ricard [8] showed that in fact Γq(H) is a factor for dimH ≥ 2.
More general deformations of the free group factors have also been considered. For H a Hilbert space
with basis {ei}i∈I , Bozejko and Speicher [3] constructed a solution to the qij -relations
c(ei)c+(ej) − qij c+(ej)c(ei) = δij,
for qij ∈ [−1, 1] as well as a further generalization of the relations arising from a contraction T ∈ B(H)
satisfying the braid relation (or Yang-Baxter relation) given by
(1 ⊗ T )(T ⊗ 1)(1 ⊗ T ) = (T ⊗ 1)(1 ⊗ T )(T ⊗ 1).
Kr olak [6] proved that if kTk < 1, which in the qij case corresponds to the condition sup{qij : i, j ∈ I} <
1, the resulting von Neumann algebra is a factor for dimH sufficiently large.
In another direction, Liguouri and Mintchev [7] and Bozejko, Lytvynov, and Wysoczanski [1] have con-
sidered creation and annihilation operators on a Fock space arising from a continuous commutation relation
associated with a Hermitian function Q from Rj × Rj (or some more general space) to the unit circle.
This construction also involves additional commutation relations on the creation operators, and includes the
anyons as a special case.
Here we will consider a continuous Q-commutation relation arising from a function taking values in (−1, 1).
Before we state the problem more explicitly, we introduce some notations which will be used throughout the
paper.
Notation 1. Let U be a nonempty open subset of Rj for some integer j ≥ 1. We also fix Q ∈ C(U × U ),
the space of continuous functions on U × U . Further assume that q := sup{Q(x, y) : x, y ∈ U} < 1 and
that Q is a symmetric function, that is Q(x, y) = Q(y, x). Also define H = L2(U ).
1
2
ADAM MERBERG
For points x, y ∈ U , we wish to consider, at least heuristically, infinitesimal creation and annihilation
operators on a Q-twisted Fock space satisfying the Q-commutation relation
a(x)a+(y) − Q(x, y)a+(y)a(x) = δ(x − y) · 1,
(1)
where δ is the usual Dirac δ, whence
Z Z δ(x − y)f (x, y) dx dy =Z f (y, y) dy.
Rigorously, this relation should be understood as a statement about operator-valued distributions, which
makes sense upon smearing with a test function and considering the resulting quadratic forms. The meaning
will be explained further in Section 2.
The operator-valued distributions a+(x) and a(x) will give rise to creation and annihilation operators
a(f ) and a+(f ) on a Q-deformed Fock space FQ(H). We will use these to define a Q-deformed field operator
w(f ) = a(f ) + a+(f ) and the von Neumann algebra ΓQ(H) generated by operators of this type.
This paper has four sections, not including this introduction. Section 2 will present the construction of
a deformed Q-Fock space with creation and annihilation operators realizing the Q-commutation relation.
In Section 3, we will discuss basic properties of the von Neumann algebra generated by the field operators
on this Fock space. In Section 4, we will show that the field operators arise as a limit in distribution of
operators on discrete qij-Fock spaces considered by Bozejko and Speicher in [3]. In Section 5, we will show
that the von Neumann algebra generated by these operators is a factor.
Acknowledgments. While working on this paper, the author was supported in part by a National Science
Foundation (NSF) Graduate Research Fellowship. He was also supported in part by funds from NSF grant
DMS-1001881. The author also benefited from attending the program "Bialgebras in Free Probability" at
the Erwin Schrodinger Institute in the Spring of 2011. His travel was supported by NSF grant DMS-1101630.
The author would like to thank Dan-Virgil Voiculescu for many enlightening conversations, Michael Hart-
glass for suggesting simplifications to the proof of Proposition 4, and an anonymous referee for offering
several corrections and for pointing out that several results are special cases of results of Krolak [6, 5].
We will construct our Q-Fock space by defining a deformed inner product on the algebraic Fock space.
2. The Q-Fock space
Fix n and define for 1 ≤ i ≤ n − 1 the operator T (n)
i
on H⊗n by
T (n)
i
f (x1, . . . , xn) = Q(xi, xi+1)f (x1, . . . , xi−1, xi+1, xi, xi+2, . . . , xn).
Evidently Ti is self-adjoint and bounded with norm at most q := supx,y Q(x, y) < 1. It is easily verified
that
(2)
i T (n)
T (n)
j = T (n)
j T (n)
i
for i − j ≥ 2
and T (n)
i T (n)
i+1T (n)
i = T (n)
i+1T (n)
i T (n)
i+1.
These relations are known as the Yang-Baxter relations, or sometimes the braid relations. Now let Sn denote
the symmetric group on n elements and for i = 1, . . . , n− 1 let πi be the permutation transposing i and i + 1
and fixing all other elements. We define the map φn first on the πi by φn(πi) = T (n)
and then on all of Sn
by quasi-multiplicative extension. This means that if π = πi1 ··· πik is a decomposition of π into a minimal
number of the πi then we define
i
φn(π) = φn(πi1 )··· φn(πik ) = T (n)
i1
··· T (n)
ik
.
That this definition does not depend on our choice of minimal length decompositions for π is a consequence
of the fact that the T (n)
satisfy (2). It follows from this definition that φn(σ1σ2) = φn(σ1)φn(σ2) whenever
σ1 + σ2 = σ1σ2. Here σk denotes the number of inversions of the permutation σk. That is,
i
Q ∈ B(H⊗n) by
P (n)
Q = Xσ∈Sn
φn(σ).
Equivalently, σk is the length of the shortest word for σk as a product of the fundamental transpositions.
We now define the operator P (n)
σk = {(i, j) : 1 ≤ i < j ≤ n, σk(i) > σk(j)} .
SOME RESULTS ON CONTINUOUS DEFORMED FREE GROUP FACTORS
3
By Theorem 2.3 of [3], the operator P (n)
Q is strictly positive.
Let Falg(H) be the algebraic Fock space on H,
∞
Falg(H) :=
H⊗n,
Mn=0
where the direct sum is the algebraic direct sum, so that only finite sums are permitted. Here H⊗0 is a
one-dimensional vector space generated by a distinguished unit vector Ω, which we call the vacuum vector.
The Q-inner products on the H⊗n naturally define a Q-inner product on Falg(H) by sesquilinear extension
of
Q gE0
hf, giQ =(Df, P (n)
0,
, m = n,
m 6= n,
for f ∈ H⊗n and g ∈ H⊗m. Here, h·,·i0 denotes the usual inner product on H⊗n. We now define the Q-Fock
space FQ(H) as the completion of Falg(H) with respect to the Q-inner product.
We are now almost ready to introduce the Q-creation and annihilation operators. We will define these in
terms of the free creation and annihilation operators. For f ∈ H, we define the free creation operator l+(h)
on Falg(H) by
l+(h)f = h ⊗ f
for f ∈ H⊗n, where we adopt the convention for the n = 0 case that h ⊗ Ω = h. We define l(h) to be the
free annihilation operator, given by
(l(h)f )(x1, . . . , xn−1) :=ZU
h(y)f (y, x1, . . . , xn−1) dy.
We now define for h ∈ H the Q-creation operator a+(h) and the Q-annihilation operator a(h) by
a+(h) := l+(h)
and a(h) := l(h)R(n)
Q
on H⊗n for n > 0, where
(3)
Q := 1 + T (n)
R(n)
1 + T (n)
1 T (n)
2 + ··· + T (n)
1
··· T (n)
n−2T (n)
n−1.
By writing each permutation σ ∈ Sn as the product of an element of S1 × Sn−1 and the minimal length
representative of the coset of σ, we can show that
(4)
P (n+1)
Q
= (1 ⊗ P (n)
Q )R(n+1)
Q
.
One can analogously define Q-deformed right creation and annihilation operators. In general, we will state
our results in terms of the left side versions, but analogous results hold on the right side with the same
proofs, and we will occasionally need to make use of these analogs.
It was pointed out to us by an anonymous referee that the following is actually a special case of Theorem
3.1 of [5].
Proposition 1. For h ∈ H, the operators a(h) and a+(h) are adjoints with respect to the Q-norm. Further-
more, for h ∈ H,
ka+(h)k ≤ khk
1
.
√1 − q
In particular, a+(h) and a(h) extend to bounded operators on FQ(H).
Proof. The proof of this theorem is very similar to that of Theorem 3.1 in [3]. We will first show that a(h)
and a+(h) are adjoints with respect to the Q inner product. The definitions imply that
whence it follows that
l+(h)T (n)
i = T (n+1)
i+1
l+(h),
l+(h)P (n)
Q = (1 ⊗ P (n)
Q )l+(h)
and P (n)
Q l(h) = l(h)(1 ⊗ P (n)
Q ).
4
ADAM MERBERG
By applying (4), for f ∈ H⊗n
Q
Q
gE0
(cid:10)a+(h)f, g(cid:11)Q =Da+(h)f, P (n+1)
gE0
=Df, l(h)P (n+1)
Q (cid:17) R(n+1)
=Df, l(h)(cid:16)1 ⊗ P (n)
=Df, P (n)
Q l(h)R(n+1)
Q a(h)gE0
=Df, P (n)
= hf, a(h)giQ .
gE0
Q
Q
gE0
This proves that a(h) and a+(h) are adjoints with respect to the Q-inner product.
We now prove the bound on ka+(h)k. Since kT (n)
i
k ≤ q for each i,
Thus,
(cid:13)(cid:13)(cid:13)
R(n)
Q (cid:13)(cid:13)(cid:13) ≤ 1 + q + q2 + ··· + qn−1 ≤
1
1 − q
.
P (n+1)
Q
P (n+1)
Q
Q
Q
= P (n+1)
(cid:17)∗
(cid:16)P (n+1)
(cid:17)∗(cid:16)1 ⊗ P (n)
(cid:16)R(n+1)
=(cid:16)1 ⊗ P (n)
Q (cid:17) R(n+1)
Q (cid:17)
Q (cid:17)(cid:16)1 ⊗ P (n)
(1 − q)2 (cid:16)1 ⊗ P (n)
Q (cid:17) .
≤
1
Q
Q
Since 1 ⊗ P (n)
Q and P (n+1)
Q
Therefore, for f ∈ H⊗n,
are positive operators, it follows that
P (n+1)
Q
≤
1
1 − q (cid:16)1 ⊗ P (n)
Q (cid:17) .
2
(cid:13)(cid:13)a+(h)f(cid:13)(cid:13)
Q
1
=(cid:10)a+(h)f, a+(h)f(cid:11)Q
= hh ⊗ f, h ⊗ fiQ
(h ⊗ f )E0
=Dh ⊗ f, P (n+1)
Q (h ⊗ f )E0
1 − q Dh ⊗ f, 1 ⊗ P (n)
≤
Q fE0
1 − q hh, hiQDf, P (n)
1 − q hh, hihf, fiQ
1 − qkhk2kfk2
≤
≤
≤
Q.
1
1
1
(cid:3)
We can represent an element f of the Fock space FQ(H) as a sequence of functions (f (0), f (1), . . .), with
f (n) ∈ H⊗n and
∞
Xn=0(cid:13)(cid:13)(cid:13)
2
Q
f (n)(cid:13)(cid:13)(cid:13)
< ∞.
SOME RESULTS ON CONTINUOUS DEFORMED FREE GROUP FACTORS
5
We are now ready to define the operator-valued distributions a(x) and a+(x). For f ∈ H⊗n, we define
these by
[a(x)f ](x1, . . . , xn−1) =(cid:16)R(n+1)
f (n+1)(cid:17) (x, x1, . . . , xn−1)
[a+(x)f ](x1, . . . , xn+1) = δ(x − x1)f (n−1)(x2, . . . , xn+1).
Q
These definitions, of course, makes no sense as functions, but should be interpreted as distributions on
C∞
c (U ). It is an immediate consequence of the definitions that
a(h) =ZU
h(x)a(x) dx and a+(h) =ZU
h(x)a+(x) dx,
for functions h ∈ C∞
forms. That is, for f ∈ H⊗n and g ∈ H⊗(n−1),
c (U ). These relations are understood rigorously in terms of the corresponding quadratic
(cid:10)f, a+(h)g(cid:11)Q =ZU
=ZU
h(x)(cid:10)f, a+(x)g(cid:11)Q dx
h(x)ZU n(cid:16)P (n)
Q f(cid:17) (x1, . . . , xn)δ(x − x1)g(x2, . . . , xn)dx1 . . . dxndx,
and similarly for a(h):
hg, a(h)fiQ =ZU
=ZU
h(x)hg, a(x)fiQ dx
h(x)ZU n−1(cid:16)P (n−1)
Q
It now follows from a simple computation that these operator-valued distributions satisfy the Q-commutation
relations (1).
g(cid:17) (x1, . . . , xn−1)R(n)
Q f (x, x1, . . . , xn−1)dx1 . . . dxn−1dx.
3. The Q-deformed free group von Neumann algebras
We now define the main operators of interest, the field operators w(h) by
This allows us to define the Q-deformed free group von Neumann algebra by
w(h) := a+(h) + a(h)
for h ∈ H.
ΓQ(H) := {w(h) : h ∈ H}′′ .
Before proving anything about these algebras, we will need some additional notation. We will sometimes
let a−(h) denote a(h) so that we can write av(h) for v ∈ {−, +} to denote either the annihilation or creation
operator.
Given a finite ordered set S, we will denote the set of pairings of S by P (S). That is, P (S) = ∅ if S has
odd cardinality, and if S = 2p then
P (S) = {{(a1, z1), . . . , (ap, zp)}a1 < z1, . . . , ap < zp,{a1, . . . , ap, z1, . . . , zp} = S}.
We will denote by I(V) the set of crossings of a pairing V, that is, for V = {(a1, z1), . . . , (ap, zp)},
I(V) = {(k, l) ∈ {1, . . . , r}2ak < al < zk < zl},
where the inequalities are in the ordering given on S.
For a pairing V ∈ P (S) for S ⊂ {1, . . . , n}, we define a function Qn
Q(xak , xal ).
Qn
V on U n by
V(x) = Y(k,l)∈I(V)
We will simplify notation by writing
Note that the δ on the right side is the Dirac delta.
δn
V(x) = Y(a,z)∈V
δ(xa − xz).
6
ADAM MERBERG
Proposition 2. Let f1, . . . , fn ∈ H and denote by S the set {1, . . . , n}. For v1, . . . , vn ∈ {−, +}
V (x)dx1 . . . dxn,
V (x)δn
Dv,VZ ···Z fn(xn)··· f1(x1)Qn
havn (fn)··· av1 (f1)Ω, Ωi = XV∈P (S)
where if n = 2p, Dv,V is defined by
p
Dv,V =
δvak ,− · δvzk ,+.
Yk=1
In particular, hw(fn)··· w(f1)Ω, Ωi = 0 when n is odd.
Proof. The proof of is by induction on N := {(j, k) : j < k, vj = +, vk = −}. The claim is easily seen to be
true in the case N = 0, so we proceed to assume that N > 0 and that the claim holds for N − 1. We will
assume that f1, . . . , fn lie in the dense subspace C∞
c (U ) of H and then use the Q-commutation relation (1).
Since N > 0, we can choose j minimal to satisfy vj = + and vj+1 = −. Now applying (1),
avn (xn)··· av1 (x1) = avn (xn)··· avj+2 (xj+2)a(xj+1)a+(xj )avj−1 (xj−1)··· av1 (x1)
= avn (xn)··· avj+2 (xj+2)(Q(xj , xj+1)a+(xj)a(xj+1) + δ(xj , xj+1))avj−1 (xj−1)··· av1 (x1)
= Q(xj , xj+1)avn (xn)··· avj+2 (xj+2)a+(xj )a(xj+1)avj−1 (xj−1)··· av1 (x1)
+ δ(xj − xj+1)avn (xn)··· avj+2 (xj+2)avj−1 (xj−1)··· av1 (x1)
(5)
We now consider the terms in the last line of (5) separately, denoting them by X1 and X2. For compactness
of notation, we define S′ = {1, . . . , j − 1, j + 1, j, j + 2, . . . , n} (as an ordered set) and S = {1, . . . , j − 1, j +
2, . . . , n} and also write f (x) for the product fn(xn)··· f1(x1).
For the first term we have by the inductive hypothesis,
Z ···Z f (x)hX1Ω, Ωi dx1 . . . dxn = XV∈P (S ′)
= XV∈P (S)
(j,j+1)6∈V
Dv,VZ ···Z f (x)Q(xj , xj+1)Qn
Dv,VZ ···Z f (x)Qn
V (x)δn
V (x) dx1 . . . dxn,
V (x)δn
V (x)dx1 . . . dxn
For the second term,
Z ···Z f (x)hX2Ω, Ωi dx1 . . . dxn = XV∈P ( S)
= XV∈P (S)
(j,j+1)∈V
Dv,VZ ···Z δ(xj − xj+1)f (x)Qn
Dv,VZ ···Z f (x)Qn
V (x)δn
V (x) dx1 . . . dxn.
V (x)δn
V (x) dx1 . . . dxn
The proposition now follows just by adding the results of the two computations just completed.
(cid:3)
Corollary 1. Let f1, . . . , fn and S be as in Proposition 2. Then
hw(fn)··· w(f1)Ω, Ωi = XV∈P (S)Z ···Z fn(xn)··· f1(x1)Qn
V (x)δn
V (x)dx1 . . . dxn,
Proof. Sum the formula of Proposition 2 over all choices of v1, . . . , vn.
Corollary 2. The vacuum state on ΓQ(H) is a trace.
Proof. The formula in Corollary 1 is invariant under cyclic permutations of the w(fi).
Proposition 3. The vacuum vector Ω ∈ FQ(H) is cyclic and separating for ΓQ(H).
(cid:3)
(cid:3)
SOME RESULTS ON CONTINUOUS DEFORMED FREE GROUP FACTORS
7
Proof. We first show that Ω is cyclic. It will suffice to show that an arbitrary f ∈ H⊗n is in the closure of
ΓQ(H)Ω. The proof is by induction on n. The cases of n = 0 and n = 1 are obvious, so we assume n > 1
and f ∈ L2(U n). If ǫ > 0, we can choose (fij ) ∈ H for i = 1, . . . n and j = 1, . . . , r such that
But then
f −
r
Xj=1
< ǫ/2.
r
f −
Xj=1
f1j ⊗ ··· ⊗ fnj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
w(f1j)··· w(fnj )Ω =
Xj=1
f −
r
f1j ⊗ ··· ⊗ fnj
+ g,
k=1 H⊗n. The claim now follows by applying the inductive hypothesis to g.
for some g ∈Ln−1
To show that Ω is separating for ΓQ(H), it will suffice to show that Ω is cyclic for ΓQ(H)′. We define the
anti-linear conjugation operator J : FQ(H) → FQ(H) by JXΩ = X ∗Ω for X ∈ ΓQ(H). This operator is
well-defined because by the tracial property kXΩk = kX ∗Ωk. Since JΓQ(H)J commutes with ΓQ(H), and
Ω is seen to be cyclic for JΓQ(H)J in the same way as for ΓQ(H), the claim follows.
(cid:3)
4. The discretization lemma
We will now show that the creation and annihilation operators a+(h) and a(h) can be realized as a
limit in distribution of operators on a discrete Fock space arising from the discrete commutation relation as
considered in [3]. Fix ǫ and let Uǫ := U ∩ ǫZj. We let Hǫ be a real Hilbert space with orthonormal basis
{ex : x ∈ Uǫ}. For x, y ∈ Uǫ, we define qxy = Q(x, y).
Bozejko and Speicher showed [3] that there is a qxy-Fock space on Hǫ with vacuum vector Ωǫ, creation
operators a+
ǫ (f ) and annihilation operators aǫ(e) for e ∈ H satisfying the discrete qxy-commutation relation
aǫ(ex)a+
ǫ (ey) − qij a+
ǫ (ey)cǫ(ex) = δxy · 1.
The creation operator a+
ǫ (ex) and the annihilation operator aǫ(ex) are adjoints with respect to the deformed
inner product on the Fock space. We will denote this Fock space by FQ,ǫ(Hǫ), its vaccum vector by Ωǫ, and
its inner product by h·,·iQ,ǫ.
Now define aǫ(f ) and a+
ǫ (f ) by
aǫ(f ) := ǫj/2 Xx∈Uǫ
f (x)aǫ(ex) and a+
ǫ (f ) := ǫj/2 Xx∈Uǫ
f (x)a+
ǫ (ex).
Evidently, (aǫ(f ))∗ = a+
ǫ (f ).
To simplify notation, we define for a pairing V,
where the δxa,xz on the right side is a Kronecker delta.
Dn
V (x) = Y(a,z)∈V
δxa,xz ,
Lemma 1. The family {aǫ(f ) : f ∈ C∞
{a(f ) : f ∈ C∞
c (U )} converges in joint ∗-distribution as ǫ → 0 to the family
c (U )} introduced in Section 2 where all of the distributions are with respect to the respective
vacuum states.
Proof. We will use the fact, as shown by Bozejko and Speicher in [3], that for v1, . . . , vn ∈ {+,−},
ǫ (xn)··· av1
havn
ǫ (x1)Ωǫ, Ωǫi = XV∈P (S)
Dv,VDn
V(x) Y(k,l)∈I(V)
qxak ,xal
,
8
ADAM MERBERG
where S = {1, . . . , n} and Dv,V is as in Proposition 2. Again writing f (x) for the product fn(xn)··· f1(x1),
we have that
ǫ→0havn
lim
ǫ (fn)··· av1
ǫ (f1)Ωǫ, Ωǫi = lim
ǫ→0
= lim
ǫ→0
ǫ (exn )··· av1
ǫ
ǫjn/2 Xx∈U n
ǫjn/2 Xx∈U n
ǫ
hf (x)avn
f (x) XV∈P (S)
ǫ (ex1)Ωǫ, Ωǫi
V (x) Y(k,l)∈I(V)
Dv,VDn
qxak ,xal
Dv,VDn
V(x)Qn
V (x) dx1 . . . dxn
=Z ···Z f (x) XV∈P (S)
= havn (fn)··· av1 (f1)Ω, Ωi .
(cid:3)
ǫ
We conclude this section by noting that the inner product on FQ,ǫ(Hǫ) is defined using positive operators
P (n)
Q,ǫ on H⊗n
such that
hξ, ηiQ,ǫ =Dξ, P (n)
for ξ, η ∈ H⊗n, where h·,·i0,ǫ denotes the inner product of the Free fock space on Hǫ. Since we have
assumed that supx,y Q(x, y) < 1, there is an operator R(n)
Q,ǫ =
(cid:16)1 ⊗ P (n)
Q,ǫ ηE0,ǫ
Q,ǫ of norm at most (1 − q)−1 such that P (n+1)
Q,ǫ. One can use this to show that P (n+1)
Q,ǫ ≤ (1 − q)−1(1 ⊗ P (n)
Q,ǫ(cid:17) R(n)
Q,ǫ ) for all ǫ.
,
5. The factoriality result
To state our main theorem, we will need to introduce the right field operator wr(f ) for f ∈ H. We define
wr(f ) = Jw(f )J,
where J : FQ(H) → FQ(H) is the canonical antilinear isometry defined by J(XΩ) = X ∗Ω. Equivalently,
wr(f ) = ar(f ) + a+
r (f ),
r (f ) are the right annihilation and right creation operators defined analogously to the left
where ar(f ) and a+
annihilation and left creation operators.
Theorem 1. Let g1, g2, . . . ∈ C∞
each d > 0, define
c (U ) be real-valued functions with gigj = 0 for i 6= j and kgik2 = 1. For
Then for d sufficiently large, ker Nd = CΩ and Nd > ǫ1 on FQ(H) ⊖ CΩ for some ǫ > 0.
Nd =
(w(gi) − wr(gi))2.
d
Xi=1
In view of a theorem of Connes [4], this theorem will have the following consequence. It was pointed out
to us by an anonymous referee that this follows immediately from the main theorem of Krolak in [6].
Corollary 3. The von Neumann algebra ΓQ(H) is a factor which does not have property Γ.
Proof. Choose Nd large enough that Nd > ǫ1 on the orthogonal complement of the vacuum subspace. If
X ∈ ΓQ(H) ∩ ΓQ(H)′ then (w(gi) − wr(gi)) X = 0 for i = 1, . . . , d. Thus XΩ ∈ ker Nd = CΩ. Since Ω is
separating, X ∈ C. Thus, ΓQ(H) is a factor of Type II1. By Theorem 2.1 of [4], ΓQ(H) does not have
property Γ.
(cid:3)
Our method of proof of Theorem 1 will be similar to that used by Krolak [6] and will require some
estimates.
Proposition 4. For each n, define operators
Ln : H ⊗ H⊗(n−1) → H⊗(n−2)
and Rn : H⊗(n−1) ⊗ H → H⊗(n−2)
by
Ln(h ⊗ f ) = l(h)f
and Rn(f ⊗ h) = r(h)f,
SOME RESULTS ON CONTINUOUS DEFORMED FREE GROUP FACTORS
9
where l(f ) and r(f ) are the free left and right annihilation operators, respectively acting on H⊗(n−1) as a
subspace of FQ(H). Suppose that g ∈ H with kgk = 1 and define D on H⊗n by D(f ) = g ⊗ f ⊗ g. Then
2
(cid:13)(cid:13)(cid:13)Ln+2(T (n+2)
··· T (n+2)
n+1
)D(cid:13)(cid:13)(cid:13)Q ≤ qn
and
n
(cid:13)(cid:13)(cid:13)Rn+2(T (n+2)
··· T (n+2)
1
)D(cid:13)(cid:13)(cid:13)Q ≤ qn.
Proof. We will prove the second statement, and the first can be proven analogously. Our approach is similar
to that of Lemma 7 in [6]. Namely, we will begin by showing that the operator Rn+2(T (n+2)
)D
commutes with P (n)
Q , the operator used to define the Q-inner product in Section 2. For this, it will suffice
to show that Rn+2(T (n+2)
)D commutes with φn(σ) for each σ ∈ Sn, where φn : Sn → H⊗n is
as in Section 2. By quasimultiplicativity of φn, we can further assume that σ is one of the fundamental
transpositions πk. Using the relation (2), we have
··· T (n+2)
··· T (n+2)
n
n
1
1
Rn+2(T (n+2)
n
··· T (n+2)
1
1
n
n
n
··· T (n+2)
··· T (n+2)
··· T (n+2)
··· T (n+2)
(T (n+2)
)Dφn(πk) = Rn+2(T (n+2)
= Rn+2(T (n+2)
= Rn+2(T (n+2)
= Rn+2(T (n+2)
= Rn+2T (n+2)
= T (n)
= φn(πk)Rn+2(T (n+2)
k Rn+2(T (n+2)
n
n
n
n
k
1
k
)DT (n)
)T (n+2)
k+1 D
k+1 T (n+2)
k+1 T (n+2)
k+2 T (n+2)
T (n+2)
k
k−1
k+2 T (n+2)
T (n+2)
k+1 T (n+2)
T (n+2)
k−1
··· T (n+2)
)D
k
k
1
1
··· T (n+2)
··· T (n+2)
1
)D
)D
)D.
··· T (n+2)
1
)D
1
··· T (n+2)
Q =Pσ∈Sn
=(cid:13)(cid:13)(cid:13)Rn+2(T (n+2)
≤ kRn+2k0(cid:13)(cid:13)(cid:13)
≤ 1 · qn · 1.
(T (n+2)
n
n
··· T (n+2)
1
··· T (n+2)
1
)Dφn(σ)(cid:13)(cid:13)(cid:13)0
)(cid:13)(cid:13)(cid:13)0 kDφn(σ)k0
n
(cid:13)(cid:13)(cid:13)Rn+2(T (n+2)
··· T (n+2)
1
)Dφn(σ)(cid:13)(cid:13)(cid:13)Q
Therefore, Rn+2(T (n+2)
n
··· T (n+2)
1
)D commutes with P (n)
φn(σ). In particular, this means that
In the last line, we have used the fact that D is an isometry in the 0-norm and Rn+2 is a contraction when
restricted to the subspace H⊗(n+1) ⊗ g.
(cid:3)
The next lemma provides an analog to parts of Lemma 8 of [6].
Lemma 2. There is a constant C, depending only on Q, such that all of the following estimates hold for
any orthonormal vectors h1, . . . , hd ∈ H:
(1) (cid:13)(cid:13)(cid:13)Pd
(2) (cid:13)(cid:13)(cid:13)Pd
(3) (cid:13)(cid:13)(cid:13)Pd
(4) (cid:13)(cid:13)(cid:13)Pd
i=1 a+(hi)a+
r (hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d and (cid:13)(cid:13)(cid:13)Pd
i=1 a+(hi)ar(hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d and (cid:13)(cid:13)(cid:13)Pd
i=1 a(hi)a(hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d and (cid:13)(cid:13)(cid:13)Pd
i=1 a+(hi)a(hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d and (cid:13)(cid:13)(cid:13)Pd
i=1 a(hi)ar(hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d
r (hi)a(hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d
i=1 ar(hi)ar(hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d
r (hi)ar(hi)(cid:13)(cid:13)(cid:13)Q ≤ C√d
i=1 a+
i=1 a+
Proof. We take C = 1
Q ) for all n is as established
in (3).
In general, to prove that an operator X has norm at most K, it will be sufficient to prove that
kXfk2 ≤ K 2kfk2 for all of the form f ∈ H⊗n where n ≥ 0 is arbitrary. To prove the first bound in part 1,
1−q , which is large enough so that P (n+1)
≤ C(1 ⊗ P (n)
Q
10
we have
d
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
a+(hi)a+
2
Q
r (hi)f(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ADAM MERBERG
2
d
d
Q
P (n+2)
Q
hi ⊗ f ⊗ hi,
Xj=1
Xj=1
hi ⊗ f ⊗ hi,
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
hi ⊗ f ⊗ hi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
=* d
Xi=1
≤ C2* d
Xi=1
hhi, hjihhi, hjiDf, P (n)
Xi,j=1
Xi=1Df, P (n)
Q fE0
= C2
= C2
d
d
d
= dC2kfk2
Q
(1 ⊗ P (n)
Q
hj ⊗ f ⊗ hj+
Q ⊗ 1)hj ⊗ f ⊗ hj+
Q fE0
0
For the second bound in part 1, we have that
where in the last line we have used the first bound in part 1.
The proof of the first bound in part 2 is similar:
d
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
a(hi)ar(hi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Q
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
a+(hi)ar(hi)f(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Q
2
d
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
≤ C√d,
r (hi)a+(hi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Q
2
d
d
d
d
Xi=1
Q
=
a+
a+
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
r (hi)a+(hi)!
∗(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Q
hi ⊗ ar(hi)f(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
Q (hi ⊗ ar(hi)f ) , hj ⊗ ar(hj)fEQ
Xi,j=1DP (n)
Xi,j=1D(1 ⊗ P (n−1)
≤ C
Xi=1DP (n−1)
kar(hi)fk2
Xi=1
≤ C
≤ dC2kfk2
Q.
ar(hi)f, ar(hi)fEQ
≤ C
Q
Q
Q
d
d
d
) (hi ⊗ ar(hi)f ) , hj ⊗ ar(hj)fEQ
The arguments used to prove the second inequality in part 2 and all the remaining estimates are similar to
those cases just completed.
(cid:3)
We will need one additional bound, which is analogous to the last part of Lemma 8 of [6].
Proposition 5. If h1, . . . , hd ∈ C∞
constant C, depending only on Q, such that
c (U ) are such that khik2 = 1 and hihj = 0 for i 6= 0 then there is a
d
Xi=1(cid:0)a(hi)a+(hi) − 1(cid:1)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤ Cq√d and
d
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1(cid:0)ar(hi)a+
r (hi) − 1(cid:1)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤ Cq√d.
SOME RESULTS ON CONTINUOUS DEFORMED FREE GROUP FACTORS
11
Proof. We will prove the first estimate; the proof of the second is analogous. It will suffice to show that for
To prove this result, we will make use of Lemma 1, which implies that in the notation of Section 4,
≤ q2C2dkfk2
Q .
We again choose C = 1
C(P (n−1)
by Vǫ. Applying the discrete commutation relations and rearranging terms,
1−q . For this choice of the constant, we have P (n)
ǫ (f1j)··· a+
2
Q
2
d
d
d
Q
= lim
c (U ),
f =Pj∈J f1j ⊗ ··· ⊗ fnj with f1j,··· , fnj ∈ C∞
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1(cid:0)a(hi)a+(hi) − 1(cid:1) f(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ǫ→0(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1(cid:0)a(hi)a+(hi) − 1(cid:1) f(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1(cid:0)aǫ(hi)a+
Q,ǫ ⊗ 1). We define fǫ = Pj∈J a+
kVǫfǫkQ,ǫ =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ǫj Xx1,x2∈Uǫ
Xi=1
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1 Xx1,x2∈Uǫ
≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1 Xx1∈Uǫ
ǫjhi(x1)hi(x2)(cid:0)Q(x1, x2)a+
ǫjhi(x1)hi(x2)Q(x1, x2)a+
hi(x1)hi(x2)aǫ(ex1)a+
x2∈Uǫ
d
d
d
ǫ (hi) − 1(cid:1)Xj∈J
a+
ǫ (f1j)··· a+
d
Q,ǫ ≥ C(1 ⊗ P (n−1)
ǫ (fnj)Ωǫ, fix ǫ > 0, and denote Pd
fǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Q,ǫ
ǫ (ex2)
− 1
− dfǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Q,ǫ
ǫ (ex2 )aǫ(ex1) + δx1,x2(cid:1) fǫ
ǫ (ex2)aǫ(ex1)fǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Q,ǫ
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1 −1 + ǫj Xx∈Uǫ
ǫ (ex2)aǫ(ex1)fǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ǫjhi(x1)hi(x2)Q(x1, x2)ex2 ⊗ aǫ(ex1)fǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ǫjhi(x1)hi(x2)Q(x1, x2)aǫ(ex1)fǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ǫjhi(x1)hi(x2)Q(x1, x2)a+
Q,ǫ
2
Q,ǫ
Q,ǫ
2
.
2
,
Q,ǫ
ǫ (fnj)Ωǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
) and also P (n)
Q,ǫ
i=1 (aǫ(hi)a+
Q,ǫ ≥
ǫ (hi) − 1)
hi(x)2! fǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Q,ǫ
.
Since khik2 = 1, the second term in the last line converges to 0 as ǫ → 0, whence we need only show that
the first term has the needed bound in the limit. Denoting this term by Sǫ, we have
2
Here we have used the fact that P (n+1)
the adjoint map is an isometry and then make use of our choice of C again:
Q
Q ). To further simplify this bound, we use the fact that
S2
ǫjhi(x1)hi(x2)Q(x1, x2)a+
Xi=1 Xx1∈Uǫ
≤ C(1 ⊗ P (n)
d
d
d
S2
ǫ =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1 Xx1,x2∈Uǫ
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1 Xx1,x2∈Uǫ
≤ C Xx2∈Uǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1 Xx1∈Uǫ
kgǫkQ,ǫ=1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1 Xx1∈Uǫ
kgǫkQ,ǫ=1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xi=1
Xi=1
ǫ ≤ C Xx2∈Uǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤ C Xx2∈Uǫ
≤ C2 Xx1,x2∈Uǫ
≤ C2 Xx1,x2∈Uǫ
sup
sup
d
d
d
d
2
Q,ǫ
2
Q,ǫ
kfǫk2
ǫ (ex1)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ǫjhi(x1)hi(x2)Q(x1, x2)ex1 ⊗ gǫ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ǫjhi(x1)hi(x2)Q(x1, x2)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
kgǫk2
2
Q,ǫkfǫk2
Q,ǫ
kfǫk2
Q,ǫ
Q,ǫ
ǫ2jhi(x1)2hi(x2)2 Q(x1, x2)2 kfǫk2
Q,ǫ.
12
ADAM MERBERG
In arriving at the last line we have made use of the fact that the hi are supported on disjoint sets. Since by
Lemma 1 we have kfǫk2
Q,ǫ → kfkQ as ǫ → 0, whence
S2
ǫ ≤
This gives the needed result.
lim sup
ǫ→0
d
Xi=1
C2Z Z Q(x1, x2)2hi(x1)2hi(x2)2 dx1dx2kfk2
Q ≤ C2q2d.
(cid:3)
Proof of Theorem 1. Expanding the definition of Nd we have,
d
Nd =
+
−
−
d
d
r (gi)a+
r (gi) + ar(gi)ar(gi) + a+
Xi=1(cid:0)a+(gi)a+(gi) + a(gi)a(gi) + a+(gi)a(gi) + a(gi)a+(gi)(cid:1)
Xi=1(cid:0)a+
Xi=1(cid:0)2a+(gi)a+
Xi=1(cid:0)a+
r (gi)a(gi) + ar(gi)a+(gi)(cid:1) .
r (gi) + 2a(gi)ar(gi) + a+(gi)ar(gi) + a(gi)a+
r (gi)ar(gi) + ar(gi)a+
d
r (gi)(cid:1)
r (gi)(cid:1)
Here we have used the fact that a+(gi)a+
operators.
r (gi) = a+
r (gi)a+(gi) and likewise for the left and right annihilation
For each i, we denote by Di the map on FQ(H) given by linear extension of f 7→ gi ⊗ f ⊗ gi for f ∈ H⊗n.
By the definition of the left and right annihilation operators,
a(gi)a+
r (gi)f = (a(gi)f ) ⊗ gi + Ln+2(T (n+2)
2
··· T (n+2)
n+1
)Di(f )
and
ar(gi)a+(gi)f = gi ⊗ (ar(gi)f ) + Rn+2(T (n+2)
n
··· T (n+2)
1
)Di(f ),
for f ∈ H⊗n, where Rn+2 and Ln+2 are as in Proposition 4. Now defining
r (gi)),
(a(gi)a+(gi) + ar(gi)a+
d
we have by Proposition 5 that kB1k ≤ 2Cq√d on FQ(H) ⊖ CΩ. Define also
B1 := −2d +
Xi=1
B2 :=
d
Xi=1(cid:16)Rn+2(T (n+2)
n
··· T (n+2)
1
)Dj(f ) + Ln+2(T (n+2)
2
··· T (n+2)
n+1
)Dj(f )(cid:17) .
By Proposition 4, we have kB2k ≤ 2qd. Finally letting
we have by Lemma 2 that kB3k ≤ 14C√d. This yields an inequality of operators,
B3 := Nd − 2d − B1 + B2,
NdFQ(H)⊖CΩ ≥ 2d(1 − q) − 2C√dq − 14C√d.
The expression on the right is positive for sufficiently large d.
Remark 1. We have assumed throughout that q := supx,y∈U Q(x, y) < 1. However, we can easily extend
the construction to the case of q = 1. Write U = Si∈I B(xi, ri) where B(xi, ri) denotes the open ball of
radius ri centered at xi ∈ Rj. For each N , define UN :=Si∈I B(cid:0)xi, N −1
Q(x, y) < 1,
so we can define HN := L2(UN ) and apply the construction to get a factor ΓQ(HN ). Moreover, we have
a natural inclusion ΓQ(HN ) ⊆ ΓQ(HN +1), so we can define SN ∈N ΓQ(HN ). The Fock space FQ(H) can
be constructed by the GNS construction. Finally, by choosing the functions g1, g2, . . . in Theorem 1 to be
x,y∈UN Q(x, y) ≤ sup
sup
N ri(cid:1). Then
x,y∈UN
(cid:3)
SOME RESULTS ON CONTINUOUS DEFORMED FREE GROUP FACTORS
13
supported in some UN , we see that we can construct an operator as in Theorem 1, so that Corollary 3 holds
as well.
References
[1] Marek Bozejko, Eugene Lytvynov, and Janusz Wysoczanski. Non-commutative L´evy processes for generalized (particularly
anyon) statistics. Arxiv preprint arXiv:1106.2933, 2011.
[2] Marek Bozejko and Roland Speicher. An example of a generalized Brownian motion. Comm. Math. Phys., 137(3):519 -- 531,
1991.
[3] Marek Bozejko and Roland Speicher. Completely positive maps on Coxeter groups, deformed commutation relations, and
operator spaces. Math. Ann., 300(1):97 -- 120, 1994.
[4] A. Connes. Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1. Ann. of Math. (2), 104(1):73 -- 115, 1976.
[5] Ilona Kr olak. Wick product for commutation relations connected with Yang-Baxter operators and new constructions of
factors. Comm. Math. Phys., 210(3):685 -- 701, 2000.
[6] Ilona Kr´olak. Factoriality of von Neumann algebras connected with general commutation relations -- finite dimensional
case. In Quantum probability, volume 73 of Banach Center Publ., pages 277 -- 284. Polish Acad. Sci. Inst. Math., Warsaw,
2006.
[7] A. Liguori and M. Mintchev. Fock representations of quantum fields with generalized statistics. Comm. Math. Phys.,
169(3):635 -- 652, 1995.
[8] ´Eric Ricard. Factoriality of q-Gaussian von Neumann algebras. Comm. Math. Phys., 257(3):659 -- 665, 2005.
[9] Piotr ´Sniady. Factoriality of Bozejko-Speicher von Neumann algebras. Comm. Math. Phys., 246(3):561 -- 567, 2004.
[10] Dan Voiculescu. Symmetries of some reduced free product C ∗-algebras. In Operator algebras and their connections with
topology and ergodic theory (Bu¸steni, 1983), volume 1132 of Lecture Notes in Math., pages 556 -- 588. Springer, Berlin, 1985.
Department of Mathematics, University of California, Berkeley, CA, USA 94720
E-mail address: [email protected]
|
1409.4056 | 2 | 1409 | 2015-05-14T15:42:16 | Finitely generated nilpotent group C*-algebras have finite nuclear dimension | [
"math.OA"
] | We show that group C*-algebras of finitely generated, nilpotent groups have finite nuclear dimension. It then follows, from a string of deep results, that the C*-algebra $A$ generated by an irreducible representation of such a group has decomposition rank at most 3. If, in addition, $A$ satisfies the universal coefficient theorem, another string of deep results shows it is classifiable by its Elliott invariant and is approximately subhomogeneous. We give a large class of irreducible representations of nilpotent groups (of arbitrarily large nilpotency class) that satisfy the universal coefficient theorem and therefore are classifiable and approximately subhomogeneous. | math.OA | math |
FINITELY GENERATED NILPOTENT GROUP C*-ALGEBRAS
HAVE FINITE NUCLEAR DIMENSION
CALEB ECKHARDT AND PAUL MCKENNEY
Abstract. We show that group C*-algebras of finitely generated, nilpotent groups
have finite nuclear dimension. It then follows, from a string of deep results, that
the C*-algebra A generated by an irreducible representation of such a group has
decomposition rank at most 3. If, in addition, A satisfies the universal coefficient
theorem, another string of deep results shows it is classifiable by its ordered K-
theory and is approximately subhomogeneous. We observe that all C*-algebras
generated by faithful irreducible representations of finitely generated, torsion free
nilpotent groups satisfy the universal coefficient theorem.
1. Introduction
The noncommutative dimension theories of Kirchberg and Winter (decomposition
rank) and of Winter and Zacharias (nuclear dimension) play a prominent role in the
theory of nuclear C*-algebras. This is especially apparent in Elliott's classification
program where finite noncommutative dimension is essential for a satisfying classifi-
cation theory. In [41], Winter and Zacharias express a hope that nuclear dimension
will "shed new light on the role of dimension type conditions in other areas of non-
commutative geometry." We share this hope and this work aims to use the theory of
nuclear dimension to shed new light on the representation theory of discrete nilpotent
groups.
A discrete group is Type I (and therefore has a "tractable" representation theory)
if and only if it has an abelian subgroup of finite index [38]. Therefore being Type
I is a highly restrictive condition for discrete groups and therefore for most discrete
groups, leaves many of the tools of classic representation theory out of reach. Recent
breakthroughs of several mathematicians (H. Lin, Z. Niu, H. Matui, Y. Sato and W.
Winter to name a few) gave birth to the possibility of classifying the C*-algebras
generated by the irreducible representations of nilpotent groups by their ordered K-
theory. A key missing ingredient was knowing whether or not the group C*-algebras
of finitely generated nilpotent groups have finite nuclear dimension. Our main result
(Theorem 4.4) supplies this ingredient. In particular we show that the nuclear di-
mension of C ∗(G) is bounded by 10h(G)−1 · h(G)! where h(G) is the Hirsch number of
G (see Section 2.1.1).
Each finitely generated nilpotent group has an algebraic "basis" of sorts and the
Hirsch number returns the size of this basis–it is therefore not surprising to see h(G)
appear in the nuclear dimension estimate.
C.E. was partially supported by NSF grant DMS-1262106.
1
2
CALEB ECKHARDT AND PAUL MCKENNEY
Fix a finitely generated nilpotent group G and an irreducible representation π of G.
The C*-algebra generated by π(G) is simple, nuclear, quasidiagonal with unique trace
and, by a combination of Theorem 4.4 with many deep results (see Theorem 2.10 for
a complete list), has finite decomposition rank. Therefore if C ∗(π(G)) satisfies the
universal coefficient theorem (see [36]), it is classified by its ordered K-theory and
isomorphic to an approximately subhomogeneous C*-algebra by [23, 26, 27] (see [27,
Corollary 6.2]). We observe in Section 4 that the work of Rosenberg and Schochet [36]
show that if G is torsion free and π is faithful (as a group homomorphism on G) then
C ∗(π(G)) satisfies the universal coefficient theorem.
If G is a two-step nilpotent group, it is well-known that the C*-algebras gener-
ated by irreducible representations of G are either finite dimensional or AT-algebras.
Indeed Phillips showed in [32] that all simple higher dimensional non commutative
tori (a class of C*-algebras that include C ∗(π(G)) when G is two-step and π is an
irreducible, infinite dimensional representation) are AT algebras. In some sense this
result forms the base case for our induction proof (see below for a more detailed de-
scription). Peeling back a couple layers, we mention that Phillips' work relies on that
of Elliott and Evans [12] and Kishimoto [21] (see also [3] and [24] for precursors to
Phillips' Theorem).
Since in general a discrete nilpotent group G is not Type I we are left with essentially
no possibility of reasonably classifying its irreducible representations up to unitary
equivalence. On the other hand if every primitive quotient of C ∗(G) satisfies the
UCT, then one could classify the C*-algebras generated by these representations
by their ordered K-theory. This provides a dual viewpoint to the prevailing one of
parametrizing irreducible representations by primitive ideals of C ∗(G) or by the space
of characters of G (see [19])–we thank Nate Brown for sharing this nice observation
with us.
Let us provide a broad outline of our proof. First we prefer to deal with torsion
free groups. Since every finitely generated nilpotent group has a finite index torsion
free subgroup we begin in Section 3 by showing that finite nuclear dimension is stable
under finite extensions. We then focus on the torsion free case.
We proceed by induction on the Hirsch number (see Section 2.1.1) of the nilpo-
tent group G. When dealing with representation theoretic objects, (like a group C*-
algebra) induction on the Hirsch number is sometimes more helpful than induction
on, say, the nilpotency class for the simple reason that non-trivial quotients of G have
Hirsch number strictly less than G while the nilpotency class of the quotient may be
unchanged.
By [31] (see also Theorem 2.7 below) we can view C ∗(G) as a continuous field over
the dual of its center, Z(G). Since Z(G) is a finitely generated abelian group, our task
is to bound the nuclear dimension of the fibers as the base space is already controlled.
Since we are proceeding by induction we can more or less focus on those fibers
homomorphisms) on Z(G). In the case that G is a two step nilpotent group, then
(we call them C ∗(G,eγ) ) induced by characters γ ∈ [Z(G) that are faithful (as group
C ∗(G,eγ) is a simple higher dimensional noncommutative torus. Phillips showed in
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS
3
[32] that any such C*-algebra is an AT-algebra and therefore has nuclear dimension
(decomposition rank in fact) bounded by 1. Phillips' result is crucial for us as it allows
us to reduce to the case that the nilpotency class of G is at least 3 and provides enough
"room" for the next step of the proof.
We then find a subgroup N of G, with strictly smaller Hirsch number, such that
G ∼= N ⋊ Z. (N is simply the subgroup generated by Zn−1(G) and all but one of the
generators of G/Zn−1(G); see section 2 below for definitions.) By induction we know
C ∗(N) has finite nuclear dimension and so we analyze the action of Z on fibers of
C ∗(N). It turns out that there are two cases: In the first case, the fiber is simple and
the action of Z on the fiber is strongly outer and hence the crossed product (which
is a fiber of C ∗(G)) absorbs the Jiang-Su algebra by [28]. This in turn implies finite
decomposition rank of the fiber by a string of deep results (see Theorem 2.10). If the
fiber is not simple, we can no longer employ the results of [28], but the non-simplicity
forces the action restricted to the center of the fiber to have finite Rokhlin dimension
and so the crossed product has finite nuclear dimension by [15].
We assume the reader is familiar with the basics of group C*-algebras and discrete
crossed products and refer them to Brown and Ozawa [5] for more information.
2. Preliminaries
2.1. Facts about Nilpotent Groups.
2.1.1. Group Theoretic Facts. We refer the reader to Segal's book [37] for more infor-
mation on polycyclic and nilpotent groups. Here we collect some facts about nilpotent
groups that we will use frequently.
Let G be a group and define Z1(G) = Z(G) to be the center of G. Recursively
define Zn(G) ≤ G to satisfy Zn(G)/Zn−1(G) := Z(G/Zn−1(G)). A group G is called
nilpotent if Zn(G) = G for some n and is called nilpotent of class n if n is the
least integer satisfying Zn(G) = G.
A group G is polycyclic if it has a normal series
{e} E G1 E · · · E Gn−1 E Gn
such that each quotient Gi+1/Gi is cyclic. The number of times that Gi+1/Gi is
infinite is called the Hirsch number of the group G and is denoted by h(G). The
Hirsch number is an invariant of the group. If G is polycyclic and N is a normal
subgroup then both N and G/N are polycyclic with
Every finitely generated nilpotent group is polycyclic.
h(G) = h(N) + h(G/N)
Let now G be finitely generated and nilpotent. Let Gf denote the subgroup con-
sisting of those elements with finite conjugacy class. By [1], the center Z(G) has
finite index in Gf . If in addition G is torsion free, then by [25] the quotient groups
4
CALEB ECKHARDT AND PAUL MCKENNEY
G/Zi(G) are also all torsion free. These facts combine to show that if G is torsion
free, then every non-central conjugacy class is infinite.
2.1.2. Representation Theoretic Facts.
Definition 2.1. Let G be a group and φ : G → C a positive definite function with
φ(e) = 1. If φ is constant on conjugacy classes, then we say φ is a trace on G. It is
clear that any such φ defines a tracial state on C ∗(G). Following the representation
theory literature we call an extreme trace a character. Every character gives rise to
a factor representation and is therefore multiplicative on Z(G).
Let G be a finitely generated nilpotent group. Let J be a primitive ideal of C ∗(G)
(i.e. the kernel of an irreducible representation). Moore and Rosenberg showed in [29]
that J is actually a maximal ideal. Shortly after this result, Howe showed in [16] (see
especially the introduction of [6]) that every primitive ideal is induced from a unique
character, i.e. there is a unique character φ on G such that
J = {x ∈ C ∗(G) : φ(x∗x) = 0}.
Let G be a finitely generated nilpotent group and φ a character on G. Set K(φ) =
{g ∈ G : φ(g) = 1}. By [16] and [6], G is centrally inductive, i.e. φ vanishes on the
complement of {g ∈ G : gK(φ) ∈ (G/K(φ))f }. In this light we make the following
definition.
Definition 2.2. Let N ≤ G and φ be a positive definite function on N. We denote
by eφ the trivial extension of φ to G, i.e. that extension of φ satisfying eφ(x) = 0
for all x 6∈ N.
We use the following well-known fact repeatedly,
Lemma 2.3. Let G be a finitely generated torsion free nilpotent group and γ a faithful
point argument there is a character ω on G that extends γ. Since G is nilpotent, every
character on Z(G). Then eγ is a character of G.
Proof. Since γ is a character and the extensioneγ is a trace on G, by a standard extreme
non-trivial normal subgroup intersects the center non-trivially. Since K(eγ) ∩ Z(G)
is trivial, we have K(eγ) is also trivial. By the preceding section, Gf = Z(G), i.e. ω
vanishes off of Z(G). But this means precisely that ω = eγ.
(cid:3)
It is clear from the definitions that every nilpotent group is amenable and therefore
by [22], group C*-algebras of nilpotent groups are nuclear. In summary, for every
primitive ideal J of C ∗(G), the quotient C ∗(G)/J is simple and nuclear with a unique
trace.
2.2. C*-facts.
Definition 2.4 ( [28]). Let G be a group acting on a C*-algebra A with unique trace
τ. Since τ is unique the action of G leaves τ invariant and therefore extends to an
action on πτ (A)′′, the von Neumann algebra generated by the GNS representation of
τ. If for each g ∈ G \ {e} the automorphism of πτ (A)′′ corresponding to g is an outer
automorphism, then we say that the action is strongly outer.
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS
5
The above definition can be modified to make sense even if A does not have a
unique trace (see [28]), but we are only concerned with the unique trace case. The
following theorem provides a key step in our main result.
Theorem 2.5 ( [28, Corollary 4.11]). Let G be a discrete elementary amenable group
acting on a unital, separable, simple, nuclear C*-algebra A with property (SI) and
finitely many extremal traces. If the action of G is strongly outer, then the crossed
product A ⋊ G is Z-absorbing.
The above theorem is actually given in greater generality in [28]. We will only
make use of it in the case where G is Abelian. Another key idea for us is the fact
that discrete amenable group C*-algebras decompose as continuous fields over their
centers. First a definition,
Definition 2.6. Let G be a group and φ a positive definite function on G. Then, we
write C ∗(G, φ) for the C*-algebra generated by the GNS representation of φ.
We recall the following special case of [31, Theorem 1.2]:
Theorem 2.7. Let G be a discrete amenable group. Then C ∗(G) is a continuous field
of C*-algebras over [Z(G), the dual group of Z(G). Moreover, for each multiplicative
character γ ∈ [Z(G) the fiber at γ is isomorphic to C ∗(G,eγ).
Definition 2.8. Let A be a C*-algebra. Denote by dimnuc(A) the nuclear dimen-
sion of A (see [41] for the definition of nuclear dimension).
It will be crucial for us to view C ∗(G) as a continuous field for our inductive step
to work in the proof of our main theorem. The following allows us to control the
nuclear dimension of continuous fields.
Theorem 2.9 ( [7, Lemma 3.1], [39, Lemma 5.1]). Let A be a continuous field of
C*-algebras over the finite dimensional compact space X. For each x ∈ X, let Ax
denote the fiber of A at x. Then
dimnuc(A) ≤ (dim(X) + 1)(sup
x∈X
dimnuc(Ax) + 1) − 1.
Throughout this paper we never explicitly work with decomposition rank, nuclear
dimension, property (SI) or the Jiang-Su algebra Z (for example, we never actually
build any approximating maps in proving finite nuclear dimension). For this reason we
do not recall the lengthy definitions of these properties but simply refer the reader to
[20, Definition 3.1] for decomposition rank, [41, Definition 2.1] for nuclear dimension,
[13, 18] for the Jiang-Su algebra and its role in the classification program, and [26,
Definition 4.1] for property (SI).
Finally, for easy reference we gather together several results into
Theorem 2.10 (Rørdam, Winter, Matui and Sato). Let A be a unital, separable,
simple, nuclear, quasidiagonal C*-algebra with a unique tracial state. If A has any of
the following properties then it has all of them.
(i) Finite nuclear dimension.
6
CALEB ECKHARDT AND PAUL MCKENNEY
(ii) Z-stability.
(iii) Strict comparison.
(iv) Property (SI) of Matui and Sato.
(v) Decomposition rank at most 3.
In particular, if A is a primitive quotient of a finitely generated nilpotent group C*-
algebra then it satisfies the hypotheses of this Theorem.
Proof. Winter showed (i) implies (ii) in [40]. Results of Matui and Sato [26] and
Rørdam [34] show that (ii), (iii) and (iv) are all equivalent. Strict comparison and
Matui and Sato's [27] shows (v). Finally (v) to (i) follows trivially from the definitions.
If G is a nilpotent group, by [10], any (primitive) quotient of C ∗(G) is quasidiagonal
and therefore satisfies the hypotheses of the theorem by the discussion in Section
2.1.2.
(cid:3)
3. Stability under finite extensions
In this section we show that if a nilpotent group G has a finite index normal
subgroup H such that dimnuc(C ∗(H)) < ∞, then dimnuc(C ∗(G)) < ∞. Perhaps sur-
prisingly, this portion of the proof is the most involved and relies on several deep
results of C*-algebra theory. Moreover we lean heavily on the assumption that G is
nilpotent.
This section exists because every finitely generated nilpotent group has a finite
index subgroup that is torsion free. Reducing to this case gives the reader a clear
idea of what is happening without getting bogged down in torsion. We begin with
the following special case that isolates most of the technical details.
Theorem 3.1. Let G be a finitely generated nilpotent group. Suppose H is a normal
subgroup of finite index such that every primitive quotient of C ∗(H) has finite nuclear
dimension. Then every primitive quotient of C ∗(G) has decomposition rank at most
3.
Proof. We proceed by induction on G/H. Since G is nilpotent, so is G/H. In particu-
lar G/H has a cyclic group of prime order as a quotient. By our induction hypothesis
we may therefore suppose that G/H is cyclic of prime order p.
Let e, x, x2, ..., xp−1 ∈ G be coset representatives of G/H. Let α denote the action
of G on ℓ∞(G/H) by left translation.
It is well-known, and easy to prove, that
ℓ∞(G/H) ⋊α G ∼= Mp ⊗ C ∗(H) and that under this inclusion C ∗(H) ⊆ C ∗(G) ⊆
Mp ⊗ C ∗(H) we may realize this copy of C ∗(H) as the C*-algebra generated by the
diagonal matrices
(λh, λxhx−1, ..., λxp−1hx−(p−1)), with h ∈ H.
(3.1)
Let (π, Hπ) be an irreducible representation of G. We show that C ∗(G)/ker(π) has
decomposition rank less than or equal to 3. Let τ be the unique character on G
that induces ker(π) (see Section 2.1.2) and set K(τ ) = {x ∈ G : τ (x) = 1}. By
assumption every primitive quotient of H/(H ∩ K(τ )) has finite nuclear dimension.
Since (cid:16)G/K(τ )(cid:17)/(cid:16)H/(H ∩ K(τ ))(cid:17) is a quotient of G/H it is either trivial, in which
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS
7
case we are done by assumption or it is isomorphic to G/H. We may therefore assume,
without loss of generality, that
(3.2)
the character τ that induces ker(π) is faithful on Z(G).
By a well-known application of Stinespring's Theorem (see [30, Theorem 5.5.1])
there is an irreducible representation idp ⊗ σ of Mp ⊗ C ∗(H), such that Hπ ⊆ Hidp⊗σ
and if P : Hidp⊗σ → Hπ is the orthogonal projection, then
P ((idp ⊗ σ)(x))P = π(x)
(3.3)
Let J = ker(σ) ⊆ C ∗(H) ⊆ C ∗(G) and JG ⊆ C ∗(G) be the ideal of C ∗(G) generated
by J. By (3.3) we have JG ⊆ ker(π).
for all
x ∈ C ∗(G).
We now consider two cases:
Case 1: JG is a maximal ideal of C ∗(G), i.e. JG = ker(π).
In this case we have
C ∗(H)/J ⊆ C ∗(G)/JG ⊆ Mp ⊗ (C ∗(H)/J).
We would like to reiterate that in general the copy of C ∗(H)/J, is not conjugate to
the diagonal copy 1p ⊗ (C ∗(H)/J), but rather the twisted copy of (3.1). If C ∗(H)/J
is conjugate to the diagonal copy, then the proof is quite short (see the last paragraph
of Case 1b), so most of the present proof consists of overcoming this difficulty.
∼= G/H denote the cyclic group of order p. Define an action β of Zp on
Let Zp
ℓ∞(G/H) ⋊α G by βt(f )(s) = f (s − t) for all f ∈ ℓ∞(G/H) (i.e. β acts by left
translation on ℓ∞(G/H)) and β(λg) = λg for all g ∈ G. Since the G-action α and
Zp-action β commute with each other it is easy to see that β defines an action of Zp
on ℓ∞(G/H) ⋊α G. Moreover
βt(x) = x
for all t ∈ Z/pZ if and only if x ∈ C ∗(G).
In particular β fixes JG ⊆ C ∗(G) pointwise. Notice that Mp ⊗ J is generated by
the β-invariant set {e1a1 + · · · + epap ai ∈ J}, where ei denotes the ith minimal
projection of Mp. Therefore, β leaves Mp ⊗ J invariant (although not pointwise), and
induces an automorphism of Mp ⊗ (C ∗(H)/J). Moreover the fixed-point subalgebra
of this induced automorphism is exactly C ∗(G)/JG. We now split further into two
subcases based on the behavior of β.
Case 1a: The action β y Mp ⊗ (C ∗(H)/J) is strongly outer (Definition 2.4).
By assumption, Mp ⊗ (C ∗(H)/J) has finite nuclear dimension. By Theorem 2.10,
Mp ⊗ (C ∗(H)/J) then has property (SI). Since β is strongly outer, by [28, Corollary
4.11], the crossed product Mp ⊗ (C ∗(H)/J) ⋊β Zp is Z-stable. Since Mp ⊗ (C ∗(H)/J)
∼= Mp, it follows that [Mp ⊗ (C ∗(H)/J)] ⋊β Zp
has a unique trace and ℓ∞(p) ⋊ Zp
has unique trace. It follows from [10] that [Mp ⊗ (C ∗(H)/J)] ⋊β Zp is quasidiagonal.
8
CALEB ECKHARDT AND PAUL MCKENNEY
Therefore by Theorem 2.10, Mp ⊗ (C ∗(H)/J) ⋊β Zp has decomposition rank at most
3.
By [35], the fixed point algebra of β,
i.e. C ∗(G)/JG is isomorphic to a cor-
ner of Mp ⊗ (C ∗(H)/J) ⋊β Zp, which by Brown's isomorphism theorem [4] im-
plies that C ∗(G)/JG is stably isomorphic to Mp ⊗ (C ∗(H)/J) ⋊β Zp and therefore
C ∗(G)/JG = C ∗(G)/ker(π) has decomposition rank at most 3 by [20, Corollary 3.9].
Case 1b: The action β y Mp ⊗ (C ∗(H)/J) is not stongly outer (Definition 2.4).
Choose a generator t of Zp and set β = βt (note that every βt is strongly outer
or none of them are). We will first show that β is actually an inner automorphism of
Mp ⊗ (C ∗(H)/J).
The unique trace on Mp ⊗ (C ∗(H)/J) restricts to the unique trace τ on C ∗(G)/JG.
We will use the common letter τ for both of these traces.
Let Gf ≤ G be the subgroup consisting of those elements with finite conjugacy
classes. By [1, Lemma 3], Gf /Z(G) is finite. Let (πτ , L2) be the GNS representation
of ℓ∞(G/H) ⋊α G associated with τ.
Since τ is multiplicative on Z(G) it easily follows that
for x, y ∈ Z(G).
By the Cauchy-Schwarz inequality it follows that in L2 we have
hλx, λyiτ = τ (y−1x) ∈ T,
(3.4)
λx =L2 τ (y−1x)λy for all x, y ∈ Z(G).
Let x1, ..., xn ∈ Gf be coset representatives of Gf /Z(G) and let C ⊆ G be a set of
coset representatives for G/Gf .
The preceding discussion shows that the following set spans a dense subset of L2 :
{f λtxi : f ∈ ℓ∞(G/H), t ∈ C, 1 ≤ i ≤ n}.
(3.5)
Since τ is β-invariant, for each minimal projection e ∈ ℓ∞(p) and t ∈ G we have
τ (eλt) = τ (β(e)λt). So for each f ∈ ℓ∞(G/H) and t ∈ G we have τ (f λt) = τ (f )τ (t).
Combining this with the fact that τ vanishes on infinite conjugacy classes, we have
(3.6)
Since β is not strongly outer there is a unitary W ∈ πτ (ℓ∞(G/H) ⋊α G)′′ such that
hf λt, gλsiτ = 0 for all f, g ∈ ℓ∞(G/H), and t, s ∈ C, t 6= s.
W πτ (x)W ∗ = πτ (β(x))
for all x ∈ ℓ∞(G/H) ⋊α G.
Since β leaves πτ (G) pointwise invariant, W must commute with πτ (λt) for all t ∈ G.
For t ∈ G, let Conj(t) = {sts−1 : s ∈ G} be the conjugacy class of t.
Let s ∈ G \ Gf . Suppose first that Conj(s) intersects infinitely many G/Gf cosets.
Let (sn)∞
n Gf are all distinct. Let
f ∈ ℓ∞(G/H). Since W commutes with the λt's, for each n ∈ N we have
n=1 be a sequence from G such that the cosets snss−1
hW, f λsiτ = hλs−1
W λsn, f λsiτ
= hW, αsn(f )λsnss−1
n
iτ .
n
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS
9
: n ∈ N} form an orthogonal family of vectors,
By (3.6), the vectors {αsn(f )λsnss−1
each with the same L2 norm. Since W has L2 norm equal to 1, this implies that
hW, f λsi = 0 for all f ∈ ℓ∞(G/H).
n
Suppose now that Conj(s) intersects only finitely many G/Gf cosets. Since Conj(s) is
infinite, and Z(G) has finite index in Gf , there is a y ∈ G such that Conj(s) ∩ yZ(G)
is infinite. Let sn be a sequence from G and tn a sequence of distinct elements from
Z(G) so snss−1
n = ytn for all n ∈ N.
Let f ∈ ℓ∞(G/H). Since the set {αg(f ) : g ∈ G} is finite we may, without loss
of generality, suppose that αsn(f ) = αsm(f ) for all n, m ∈ N. By (3.4), we have
λytn =L2 τ (t−1
hW, f λsiτ = hλs−1
m tn)λytm for all n, m ∈ N. We then have
W λsn, f λsiτ
= hW, αsn(f )λsnss−1
= hW, αs1(f )λytniτ
= τ (t−1
n
n
iτ
1 tn)hW, αs1(f )λyt1iτ
In particular, we have
τ (t−1
1 t1)hW, αs1(f )λyt1iτ = τ (t−1
Since τ is faithful on Z(G) (by (3.2)) we have 1 = τ (t−1
1. Hence hW, f λsiτ = 0.
1 t2)hW, αs1(f )λyt1iτ .
1 t2), and τ (t−1
1 t2) 6= τ (t−1
1 t1) =
We have shown that for all s ∈ G \ Gf and f ∈ ℓ∞(G/H) we have hW, f λsiτ = 0. By
(3.5) it follows that
W ∈ span{πτ (f λxi) : f ∈ ℓ∞(G/H), 1 ≤ i ≤ n} ⊆ πτ (ℓ∞(G/H)⋊G) ∼= Mp⊗(C ∗(H)/J)
By the way that β acts on ℓ∞(G/H) there are a1, ..., ap ∈ C ∗(H)/J so
W = e1p ⊗ a1 +
pX
i=2
ei,i−1 ⊗ ai.
2a∗
2, a∗
3, ..., a∗
Set U = diag(1, a∗
C ∗(H)/J, by (3.1) it follows that
p) ∈ Mp ⊗ (C ∗(H)/J). Since W commutes with
(3.7) U(cid:16)C ∗(H)/J(cid:17)U ∗ = 1Mp ⊗ C ∗(H)/J ⊆ U(C ∗(G)/JG)U ∗ ⊆ Mp ⊗ (C ∗(H)/J).
3 · · · a∗
2a∗
Let U be a free ultrafilter on N and for a C*-algebra A, let AU denote the ultrapower
of A. We think of A ⊂ AU via the diagonal embedding and write A′ ∩ AU for those
elements of the ultrapower that commute with this diagonal embedding.
By Theorem 2.10, C ∗(H)/J is Z-stable. By [33, Theorem 7.2.2], there is an em-
bedding φ of Z into (C ∗(H)/J)U ∩ (C ∗(H)/J)′. By both inclusions of (3.7) it is clear
that 1Mp ⊗ φ defines an embedding of Z into (C ∗(G)/JG)U ∩ (C ∗(G)/JG)′. Therefore,
again by [33, Theorem 7.2.2] it follows that C ∗(G)/JG is Z-stable. By Theorem 2.10,
C ∗(G)/JG has decomposition rank at most 3.
10
CALEB ECKHARDT AND PAUL MCKENNEY
Case 2: The ideal JG is not maximal. Recall the definition of σ from the begin-
ning of the proof.
For each i = 0, ..., p−1 define the representations of H, σi(h) = σ(xihx−i). By [10, Sec-
tion 3] either all of σi are unitarily equivalent to each other or none of them are. We
treat these cases separately.
Case 2a: All of the σi are unitarily equivalent to each other.
By the proof of Lemma 3.4 in [10] it follows that there is a unitary U ∈ Mp ⊗C ∗(H)/J
such that U(C ∗(H)/J)U ∗ = 1p ⊗ (C ∗(H)/J). Moreover from the same proof there
is a projection q ∈ Mp ⊗ 1C ∗(H)/J that commutes with U(idMp ⊗ σ(C ∗(G)))U ∗ so
q(U[idMp ⊗ σ(C ∗(G))]U ∗) ∼= C ∗(G)/ker(π). We therefore have a chain of inclusions
q ⊗ C ∗(H)/J ⊆ q(U[idMp ⊗ σ(C ∗(G))]U ∗) ⊆ qMpq ⊗ C ∗(H)/J.
We can now complete the proof as in the end of Case 1b (following nearly verbatim
everything that follows (3.7)).
Case 2b. None of the σi are unitarily equivalent to each other.
From the proof of Lemma 3.5 in [10] there is a projection q ∈ ℓ∞(p) ⊗ 1C ∗(H)/J
that commutes with (idMp ⊗ σ)(C ∗(G)) so q(idMp ⊗ σ(C ∗(G))) ∼= C ∗(G)/ker(π). But
since G acts transitively on G/H (and hence ergodically on ℓ∞(G/H)) we must have
q = 1. But this implies that ker(idMp ⊗ σC ∗(G)) = ker(π).
written uniquely as Pp−1
Recall the coset representatives e, x, x2, ..., xp−1 of G/H. Each x ∈ C ∗(G) can be
i=0 aiλxi for some ai ∈ C ∗(H). Fix an 0 ≤ i ≤ p − 1 and
consider λxi ∈ Mp ⊗ C ∗(H). If there is an index (k, ℓ) such that the (k, ℓ)-entry of
λxi is non-zero , then for any j 6= i the (k, ℓ)-entry of λxj must be 0. From this
observation it follows that
idp ⊗ σ(cid:16) p−1X
i=0
aiλxi(cid:17) = 0, if and only if
(idp ⊗ σ)(ai) = 0
for all 1 ≤ i ≤ p
In other words ker(π) = ker(idMp ⊗ σC ∗(G)) ⊆ JG and we are done by Case 1.
(cid:3)
Lemma 3.2. Let G be a finitely generated nilpotent group and N a finite index
subgroup of Z(G) and φ a faithful multiplicative character on N. Then there is a
finite set F of characters of G such that eφ (see Definition 2.6) is in the convex hull
of F .
Proof. This result follows from Thoma's work on characters [38] (see also the discus-
sion on page 355 of [19]). For the convenience of the reader we outline a proof and
keep the notation of [19]. Let Gf ≤ G denote the group consisting of elements with
finite conjugacy class. By [1] Z(G) (and hence N) has finite index in Gf .
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS 11
Let π be the GNS representation of Gf associated with eφGf . Since N has finite
index in Gf and π(N) ⊆ C it follows that π(Gf ) generates a finite dimensional C*-
algebra. Therefore there are finitely many characters ω1, ..., ωn of Gf that extend φ
(trivially as π(N) ⊆ C) and a sequence of positive scalars λi such that
(3.8)
eφGf =
nX
i=1
λiωi
The positive definite functions eωi on G need not be traces, but this is easily remedied
by the following averaging process (which we took from [19] and [38]).
Let x ∈ Gf . Then the centralizer of x in G, denoted CG(x), has finite index. Let
Ax be a complete set of coset representatives of G/CG(x). For each 1 ≤ i ≤ n and
x ∈ Gf define
(3.9)
i (x) =
eωG
1
[G : CG(x)] X
a∈Ax eωi(axa−1).
i
i
i are actually characters on G.
is extreme in the space of G-invariant traces on Gf (see [19, Page 355]).
i )
Then each eωG
is G-invariant, the trivial extension to G (which we still denote by eωG
Since each eωG
is a trace on G. We show that the eωG
By a standard convexity argument, there is a character ω on G which extends eωi
G.
Let K(ω) = {x ∈ G : ω(x) = 1}. Then K(ω) ⊆ Gf . Indeed if there is a g ∈ K(ω)\Gf ,
then g necessarily has infinite order (otherwise the torsion subgroup of G would be
infinite). Since every finitely generated nilpotent group has a finite index torsion free
subgroup and every nontrivial subgroup of a nilpotent group intersects the center
non-trivially, this would force K(ω) ∩ Z(G) to have non-zero Hirsch number, but by
assumption φ = ωN is faithful on a finite index subgroup of Z(G).
Since G is finitely generated and nilpotent, it is centrally inductive (see [6]). This
means that ω vanishes outside of Gf (ω) = {x ∈ G : xK(ω) ∈ (G/K(ω))f }. But since
K(ω) ⊆ Gf and ωN = φ, it follows that K(ω) is finite. From this it follows that
i .
Gf (ω) = Gf , i.e. ω must vanish outside of Gf . But this means precisely that ω = eωG
By (3.8) and (3.9) for x ∈ Gf we have
nX
i=1
λieωG
i (x) =
1
a∈Ax eωi(axa−1)(cid:17)
[G : CG(x)] X
(cid:16) nX
λieωi(axa−1)(cid:17)
a∈Ax eφ(x)
a∈Ax
i=1
1
=
i=1
λi(cid:16)
nX
[G : CG(x)] X
[G : CG(x)] X
= eφ(x).
i=1 λieωG
=
1
i (x) = 0 = eφ(x).
For x 6∈ Gf we clearly have Pn
(cid:3)
12
CALEB ECKHARDT AND PAUL MCKENNEY
Lemma 3.3. Set f (n) = 10n−1n!. Let G be a finitely generated nilpotent group. Let
H E G be normal of finite index. If dimnuc(C ∗(H/N)) ≤ f (h(H/N)) for every normal
subgroup of H, then dimnuc(C ∗(G/K)) ≤ f (h(G/K)) for every normal subgroup of
G.
Proof. We proceed by induction on h(G). If h(G) = 0, then G is finite and there
is nothing to prove. So assume that for every finitely generated nilpotent group A
with h(A) < h(G) that satisfies dimnuc(C ∗(A/N)) ≤ f (h(A/N)) for every normal
subgroup N of A, we have dimnuc(C ∗(A′/N ′)) ≤ f (h(A′/N ′)) where A′ is a finite
normal extension of A and N ′ is a normal subgroup of A′.
Let now H be a finite index normal subgroup of G that satisfies the hypotheses. If
G/Z(G) is finite, then dimnuc(C ∗(G)) = h(G) ≤ f (h(G)) by Theorem 2.9. Suppose
that G/Z(G) is infinite, i.e. h(Z(G)) < h(G). Since for any quotient G/K of G, the
group H/(H ∩ K) has finite index in G/K it suffices to show that dimnuc(C ∗(G)) ≤
f (G).
We use Theorem 2.7 to view C ∗(G) as a continuous field over [Z(G). We estimate the
nuclear dimension of the fibers. Let γ ∈ [Z(G). Suppose first that h(ker(γ)) > 0. The
fiber C ∗(G,eγ) is a quotient of the group C*-algebra C ∗(G/ker(γ)). By our induction
hypothesis,
dimnuc(C ∗(G/ker(γ))) ≤ f (h(G) − 1).
Now suppose on the other hand that F = ker(γ) is finite. If x ∈ G with [x, y] ∈
F for all y ∈ G, then yxF y−1 = xF for all y ∈ G, i.e. xF ∈ Z(G). Hence
Z(G/F )/(Z(G)/F ) is a finitely generated, nilpotent torsion group, hence is finite.
We therefore replace G with G/F and suppose that Z(G) contains a finite index
subgroup N such that γ is a faithful homomorphism on N (Notice we can not say
that γ is faithful on Z(G) since it may not be the case that Z(G/F ) = (Z(G)/F )).
By Lemma 3.2, there are finitely many distinct characters ω1, ..., ωn on G such that
eγ is a convex combination of the ωi. The GNS representation associated with eγ is
then the direct sum of the GNS representations associated with the ωi. Let Ji be the
maximal ideal of C ∗(G) induced by ωi (Section 2.1.2). Since all of the Ji are maximal
and distinct it follows (from purely algebraic considerations) that
C ∗(G,eγ) ∼=
nM
i=1
C ∗(G, ωi).
Since each ωi is a character on G, it follows by Theorem 3.1 that the decomposition
rank of C ∗(G, ωi) is bounded by 3 which also bounds the decomposition rank of
C ∗(G,eγ) by 3.
Since the nuclear dimension of all the fibers of C ∗(G) are bounded by f (h(G) − 1),
(cid:3)
by Theorem 2.9 we have dimnuc(C ∗(G)) ≤ 2h(G)f (h(G) − 1) ≤ f (h(G)).
The work of Matui and Sato on strongly outer actions (see Theorem 2.5) and of
Hirshberg, Winter and Zacharias [15] on Rokhlin dimension are both crucial to the
4. Main Result
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS 13
proof of our main result. In our case, their results turned extremely difficult problems
into ones with more or less straightforward solutions.
Lemma 4.1. Let α be an outer automorphism of a torsion free nilpotent group G.
Then for every a ∈ G, the following set is infinite.
{s−1aα(s) : s ∈ G}.
Proof. Suppose that for some a, the above set is finite. Then for any s ∈ G there are
0 ≤ m < n so that
s−maα(sm) = s−naα(sn)
or
a−1sn−ma = α(sn−m).
Therefore
sn−m = aα(sn−m)a−1 = (aα(s)a−1)n−m.
Since G is nilpotent and torsion free it has unique roots (see [25] or [2, Lemma 2.1]),
i.e. s = aα(s)a−1, or α is an inner automorphism.
(cid:3)
Lemma 4.2. Let G be a torsion free nilpotent group of class n ≥ 3. Suppose that
G = hN, xi where x ∈ G \ Zn−1(G), N ∩ hxi = {e} and Z(G) = Z(N). Let φ be a
trace on G that is multiplicative on Z(G) and that vanishes off of Z(G). Let α be the
automorphism of C ∗(N, φN ) induced by conjugation by x. Then α is strongly outer.
Proof. We first show that α induces an outer automorphism of N/Z(G). If not, then
there is a z ∈ N such that
xax−1Z(G) = z−1azZ(G)
for all a ∈ N.
In other words zx ∈ Z2(G). Since x 6∈ Zn−1(G) we also have z 6∈ Zn−1(G). Then
z−1Zn−1(G) = xZn−1(G), from which it follows that z 6∈ N.
Since G is torsion free so is N. Since φ is multiplicative on Z(N) and vanishes out-
side of Z(N) it follows that φ is a character (see Lemma 2.3) and therefore C ∗(N, φ)
has a unique trace (see Section 2.1). Let (πφ, L2(N, φ)) be the GNS representation
associated with φ. By way of contradiction suppose there is a W ∈ πφ(N)′′ such that
W πφ(g)W ∗ = πφ(α(g)) for all g ∈ N.
For each s ∈ N we let δs ∈ L2(N, φ) be its canonical image. Notice that if aZ(N) 6=
bZ(N), then δa and δb are orthogonal. In particular for any complete choice of coset
representatives C ⊆ N for N/Z(G), the set {δc : c ∈ C} is an orthonormal basis
for L2(N, φ). Since W is in the weak closure of the GNS representation πφ it is in
L2(N, φ) with norm 1. Therefore there is some a ∈ N such that hW, δai 6= 0.
We now have, for all s ∈ N,
hW, δai = hW πφ(s), δasi
= hπφ(α(s))W, δasi
= hW, δα(s)−1asi.
Since α is outer on N/Z(N), by Lemma 4.1 the set {α(s)−1asZ(N) : s ∈ N} is
infinite. Since distinct cosets provide orthogonal vectors of norm 1, this contradicts
the fact that W ∈ L2(N, φ).
14
CALEB ECKHARDT AND PAUL MCKENNEY
Essentially the same argument shows that any power of α is also not inner on
(cid:3)
πτ (N)′′, i.e. the action is strongly outer.
We refer the reader to the paper [15] for information on Rokhlin dimension of ac-
tions on C*-algebras. For our purposes we do not even need to know what it is,
simply that our actions have finite Rokhlin dimension. Therefore we omit the some-
what lengthy definition [15, Definition 2.3]. We do mention the following corollary
to the definition of Rokhlin dimension: If α is an automorphism of a C*-algebra A
and there is an α-invariant, unital subalgebra B ⊆ Z(A) such that the action of α on
B has Rokhlin dimension bounded by d, then the action of α on A also has Rokhlin
dimension bounded by d.
Lemma 4.3. Let G be a finitely generated, torsion free nilpotent group. Suppose
that G = hN, xi where N ⊳ G, N ∩ hxi = {e}, Z(G) ⊆ Z(N), and Z(G) 6= Z(N).
Let φ be a trace on G that is multiplicative on Z(G) and vanishes off Z(G). Let α
be the automorphism of C ∗(N, φN ) defined by conjugation by x. Then the Rokhlin
dimension of α is 1.
Proof. Consider the action of α restricted to Z(N). Since Z(N) 6= Z(G), α is not the
identity on Z(N). Since Z(N) is a free abelian group we have α ∈ GL(Z, d) where
d is the rank of Z(N). Since G is nilpotent, so is the group Z(N) ⋊α Z. Therefore
(1 − α)d = 0. In particular there is a y ∈ Z(N) such that (1 − α)(y) 6= 0 but
(1 − α)2(y) = 0. From this we deduce that α(y) = y + z for some z ∈ Z(G) \ {0}.
Therefore the action of α on C ∗(πφ(y)) ∼= C(T) is a rotation by φ(z). Since φ is
faithful we have φ(z) = e2πiθ for some irrational θ. By [15, Theorem 6.1] irrational
rotations of the circle have Rokhlin dimension 1. Since C ∗(πφ(y)) ⊆ Z(C ∗(N, φ)),
the remark preceding this lemma shows that the Rokhlin dimension of α acting on
C ∗(N, φ) is also equal to 1.
(cid:3)
Theorem 4.4. Define f : N → N by f (n) = 10n−1n!. Let G be a finitely generated
nilpotent group. Then dimnuc(C ∗(G)) ≤ f (h(G)).
Proof. We proceed by induction on the Hirsch number of G. If h(G) = 0, there is
nothing to prove. It is well-known that G contains a finite index torsion-free subgroup
N (see section 2.1.1). Therefore by Lemma 3.3 we may assume that G is torsion free.
We decompose C ∗(G) as a continuous field over [Z(G) as in Theorem 2.7. Let γ ∈
[Z(G). If γ is not faithful on Z(G), then since G is torsion free we have h(ker(γ)) > 0.
By our induction hypothesis we then have dimnuc(C ∗(G,eγ)) ≤ f (h(G) − 1).
C ∗(G,eγ) is a simple higher dimensional noncommutative torus and therefore an AT
algebra by [32]. AT algebras have nuclear dimension (decomposition rank in fact)
bounded by 1 [20].
Suppose now that γ is faithful on Z(G). If G is a 2 step nilpotent group, then
Suppose then that G is nilpotent of class n ≥ 3 and let Zi(G) denote its upper
central series. By [25] (see also [17, Theorem 1.2] ), the group G/Zn−1(G) is a free
abelian group. Let xZn−1(G), x1Zn−1(G), ..., xdZn−1(G) be a free basis for G/Zn−1(G).
Let N be the group generated by Zn−1(G) and {x1, ..., xd}. Then N is a normal
subgroup of G with h(N) = h(G) − 1 and G = N ⋊α Z where α is conjugation by x.
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS 15
Suppose first that Z(N) = Z(G). Since h(N) = h(G) − 1, the group C*-algebra
C ∗(N) has finite nuclear dimension by our induction hypothesis. Assuming Z(N) =
It then enjoys all of the properties of Theorem 2.10.
Z(G) means that eγ is a character on N, i.e. C ∗(N,eγ) is primitive quotient of C ∗(N).
By Lemma 4.2, the action of α on C ∗(N,eγ) is strongly outer. By [28, Corollary
4.11], the crossed product C ∗(N,eγ) ⋊α Z ∼= C ∗(G,eγ) is Z-stable, hence C ∗(G,eγ) has
decomposition rank bounded by 3 by Theorem 2.10.
Suppose now that Z(G) is strictly contained in Z(N). By [15, Theorem 4.1] and
Lemma 4.3 we have
dimnuc(C ∗(G,eγ)) = dimnuc(C ∗(N,eγ) ⋊α Z) ≤ 8(dimnuc(C ∗(N,eγ)) + 1) ≤ 9f (h(N)).
Therefore the nuclear dimension of every fiber of C ∗(G) is bounded by 9f (h(G) − 1)
and the dimension of its center is at most h(G) − 1. By Theorem 2.9, we have
dimnuc(C ∗(G)) ≤ 10h(G)f (h(G) − 1) = f (h(G)).
(cid:3)
4.1. Application to the Classification Program. Combining Theorem 4.4 with
results of Lin and Niu [23] and Matui and Sato [26, 27] (see especially Corollary 6.2
of [27]) we display the reach of Elliott's classification program:
Theorem 4.5. Let G be a finitely generated nilpotent group and J a primitive ideal of
C ∗(G). If C ∗(G)/J satisfies the universal coefficient theorem, then C ∗(G)/J is classi-
fiable by its ordered K-theory and is isomorphic to an approximately subhomogeneous
C*-algebra.
We do not know if every quotient of a nilpotent group C*-algebra satisfies the UCT,
but Rosenberg and Schochet show that satisfying the UCT is closed under Z-actions
which covers most cases of interest;
Theorem 4.6. Let G be a finitely generated, torsion free nilpotent group and π a
faithful irreducible representation of G. Then C ∗(π(G)) is classifiable by its ordered
K-theory and is an ASH algebra.
Proof. By Theorem 4.5 we only need to show that C ∗(π(G)) satisfies the UCT. Let
τ be the unique character inducing ker(π) (see Section 2.1.2). Since G is torsion free
so is G/Z(G) (Section 2.1.1). Therefore we have a normal series
Z(G) = G0 E G1 E · · · E Gn = G
∼= Z for all i = 1, ..., n. Let x1, ..., xn ∈ G be such that xiGi−1
such that Gi/Gi−1
generates Gi/Gi−1 for i = 1, ..., n. For i = 1, ..., n let αi be the automorphism of
C ∗(π(Gi−1)) defined by conjugation by xi. Since C ∗(π(G)) is an iterated crossed
product by the Z-actions αi, a repeated application of [36, Proposition 2.7] shows
that C ∗(π(G)) satisfies the UCT.
(cid:3)
16
CALEB ECKHARDT AND PAUL MCKENNEY
4.1.1. Unitriangular groups and K-theory. In light of Theorem 4.6 it is natural to
wonder about K-theory calculations for specifc groups and representations. Let us
announce a little progress in this direction. Let d ≥ 3 be an integer. Consider the
group of upper triangular matrices,
Ud = nA ∈ GLd(Z) : Aii = 1 and Aij = 0 for i > jo.
The center Z(Ud) ∼= Z is identified with those elements whose only non-zero, non-
diagonal entry can occur in the (1, d) matrix entry. Ud is a finitely generated d−1-step
nilpotent group. Fix an irrational θ and consider the character τθ on Ud induced from
the multiplicative character n 7→ e2πnθi. Then C ∗(Ud, τθ) is covered by Theorem 4.6.
In the case of d = 4, we show in [11], together with Craig Kleski that the Elliott
invariant of C ∗(U4, τθ) is given by K0 = K1 = Z10 with the order on K0 given by
those vectors x ∈ Z10 satisfying hx, (1, θ, θ2, 0, ..., 0)i > 0.
5. Questions and Comments
Very broadly this section contains one question: Is there a group theoretic char-
acterization of finitely generated groups whose group C*-algebras have finite nuclear
dimension? We parcel this into more manageable chunks.
Since finite decomposition rank implies strong quasidiagonality (see [20]) there are
many easy examples of finitely generated group C*-algebras with infinite decomposi-
tion rank [8]. There are also difficult examples of finitely generated amenable groups
with infinite nuclear dimension.
Theorem 5.1 (Giol and Kerr [14]). The nuclear dimension of C ∗(Z ≀ Z) is infinite.
Proof. Giol and Kerr construct several C*-algebras of the form C(X) ⋊Z with infinite
nuclear dimension. One notices that some of these algebras are actually quotients of
C ∗(Z ≀ Z), forcing the nuclear dimension of C ∗(Z ≀ Z) to be infinite by [41, Proposition
2.3].
(cid:3)
On the other hand if we restrict to polycyclic groups–given the finite-dimensional
feel of the Hirsch number and the role it played in the present work–it seems plau-
sible that all of these groups have finite nuclear dimension. Note that in [9] there
are numerous examples of polycyclic groups that are not strongly quasidiagonal and
therefore have infinite decomposition rank.
Question 5.2. Are there any polycyclic groups with infinite nuclear dimension?
Question 5.3. Are there any polycyclic, non virtually nilpotent groups with finite
decomposition rank or finite nuclear dimension?
The paper [9] also provides examples of non virtually nilpotent, polycyclic groups
whose group C*-algebras are strongly quasidiagonal. It seems that these groups could
be a good starting point for a general investigation into nuclear dimension of polycyclic
groups. The difficulty here is that these groups have trivial center and we therefore
do not have a useful continuous field characterization of their group C*-algebras as
in Theorem 2.7.
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS 17
Unfortunately we were unable to extend our results to the virtually nilpotent case,
and thus are left with the following question.
Question 5.4. Do virtually nilpotent group C*-algebras have finite nuclear dimen-
sion?
Finally we have
Question 5.5. If G is finitely generated and nilpotent, does C ∗(G) have finite de-
composition rank?
The careful reader will notice that the only part of our proof where we cannot
deduce finite decomposition rank is in the second case of the proof of Theorem 4.4.
There is definitely a need for both cases as there exist torsion free nilpotent groups
G such that whenever G/N ∼= Z with Z(G) ≤ N, then Z(N) is strictly bigger than
Z(G) (we thank the user YCor on mathoverflow.net for kindly pointing this out to
us). In general if a C*-algebra A has finite decomposition rank and an automorphism
has finite Rokhlin dimension, one cannot deduce that the crossed product has finite
decomposition rank (for example, consider α ⊗ β where α is a shift on a Cantor space
and β is an irrational rotation of T).
References
[1] Reinhold Baer. Finiteness properties of groups. Duke Math. J., 15:1021–1032, 1948.
[2] Gilbert Baumslag. Lecture notes on nilpotent groups. Regional Conference Series in Mathemat-
ics, No. 2. American Mathematical Society, Providence, R.I., 1971.
[3] Florin P. Boca. The structure of higher-dimensional noncommutative tori and metric Diophan-
tine approximation. J. Reine Angew. Math., 492:179–219, 1997.
[4] Lawrence G. Brown. Stable isomorphism of hereditary subalgebras of C ∗-algebras. Pacific J.
Math., 71(2):335–348, 1977.
[5] Nathanial P. Brown and Narutaka Ozawa. C ∗-algebras and finite-dimensional approximations,
volume 88 of Graduate Studies in Mathematics. American Mathematical Society, Providence,
RI, 2008.
[6] A. L. Carey and W. Moran. Characters of nilpotent groups. Math. Proc. Cambridge Philos.
Soc., 96(1):123–137, 1984.
[7] Jos´e R. Carri´on. Classification of a class of crossed product C ∗-algebras associated with resid-
ually finite groups. J. Funct. Anal., 260(9):2815–2825, 2011.
[8] Jos´e R. Carri´on, Marius Dadarlat, and Caleb Eckhardt. On groups with quasidiagonal C ∗-
algebras. J. Funct. Anal., 265(1):135–152, 2013.
[9] Caleb Eckhardt. A note on strongly quasidiagonal groups. arXiv:1309.2205 [math.OA], to appear
in Journal of Operator Theory, 2013.
[10] Caleb Eckhardt. Quasidiagonal representations of nilpotent groups. Adv. Math., 254:15–32,
2014.
[11] Caleb Eckhardt, Craig Kleski, and Paul McKenney. Classification of C*-algebras generated by
representations of the unitriangular group U4. In Preparation, 2014.
[12] George A. Elliott and David E. Evans. The structure of the irrational rotation C ∗-algebra. Ann.
of Math. (2), 138(3):477–501, 1993.
[13] George A. Elliott and Andrew S. Toms. Regularity properties in the classification program for
separable amenable C ∗-algebras. Bull. Amer. Math. Soc. (N.S.), 45(2):229–245, 2008.
[14] Julien Giol and David Kerr. Subshifts and perforation. J. Reine Angew. Math., 639:107–119,
2010.
18
CALEB ECKHARDT AND PAUL MCKENNEY
[15] Ilan Hirshberg, Wilhelm Winter, and Joachim Zacharias. Rokhlin dimension and C*-dynamics.
math.OA/1209.1618, 2012.
[16] Roger E. Howe. On representations of discrete, finitely generated, torsion-free, nilpotent groups.
Pacific J. Math., 73(2):281–305, 1977.
[17] S. A. Jennings. The group ring of a class of infinite nilpotent groups. Canad. J. Math., 7:169–
187, 1955.
[18] Xinhui Jiang and Hongbing Su. On a simple unital projectionless C ∗-algebra. Amer. J. Math.,
121(2):359–413, 1999.
[19] Eberhard Kaniuth. Induced characters, Mackey analysis and primitive ideal spaces of nilpotent
discrete groups. J. Funct. Anal., 240(2):349–372, 2006.
[20] Eberhard Kirchberg and Wilhelm Winter. Covering dimension and quasidiagonality. Internat.
J. Math., 15(1):63–85, 2004.
[21] A. Kishimoto. Unbounded derivations in AT algebras. J. Funct. Anal., 160(1):270–311, 1998.
[22] Christopher Lance. On nuclear C ∗-algebras. J. Functional Analysis, 12:157–176, 1973.
[23] Huaxin Lin and Zhuang Niu. Lifting KK-elements, asymptotic unitary equivalence and classi-
fication of simple C ∗-algebras. Adv. Math., 219(5):1729–1769, 2008.
[24] Qing Lin. Cut-down method in the inductive limit decomposition of non-commutative tori. III.
A complete answer in 3-dimension. Comm. Math. Phys., 179(3):555–575, 1996.
[25] A. I. Mal'cev. Nilpotent torsion-free groups. Izvestiya Akad. Nauk. SSSR. Ser. Mat., 13:201–212,
1949.
[26] Hiroki Matui and Yasuhiko Sato. Strict comparison and Z-absorption of nuclear C ∗-algebras.
Acta Math., 209(1):179–196, 2012.
[27] Hiroki Matui and Yasuhiko Sato. Decomposition rank of UHF-absorbing C*-algebras. Duke
Math. J., 163(14):2687–2708, 2014.
[28] Hiroki Matui and Yasuhiko Sato. Z-stability of crossed products by strongly outer actions II.
Amer. J. Math., 136(6):1441–1496, 2014.
[29] Calvin C. Moore and Jonathan Rosenberg. Groups with T1 primitive ideal spaces. J. Functional
Analysis, 22(3):204–224, 1976.
[30] Gerard J. Murphy. C ∗-algebras and operator theory. Academic Press Inc., Boston, MA, 1990.
[31] Judith A. Packer and Iain Raeburn. On the structure of twisted group C ∗-algebras. Trans.
Amer. Math. Soc., 334(2):685–718, 1992.
[32] N. Chrsitopher Phillips. Every simple higher dimensional noncommutative torus is an AT al-
gebra. arXiv:math/0609783 [math.OA], 2006.
[33] M. Rørdam. Classification of nuclear, simple C ∗-algebras. In Classification of nuclear C ∗-
algebras. Entropy in operator algebras, volume 126 of Encyclopaedia Math. Sci., pages 1–145.
Springer, Berlin, 2002.
[34] Mikael Rørdam. The stable and the real rank of Z-absorbing C ∗-algebras. Internat. J. Math.,
15(10):1065–1084, 2004.
[35] Jonathan Rosenberg. Appendix to: "Crossed products of UHF algebras by product type ac-
tions" [Duke Math. J. 46 (1979), no. 1, 1–23; MR 82a:46063 above] by O. Bratteli. Duke Math.
J., 46(1):25–26, 1979.
[36] Jonathan Rosenberg and Claude Schochet. The Kunneth theorem and the universal coefficient
theorem for Kasparov's generalized K-functor. Duke Math. J., 55(2):431–474, 1987.
[37] Daniel Segal. Polycyclic groups, volume 82 of Cambridge Tracts in Mathematics. Cambridge
University Press, Cambridge, 1983.
[38] Elmar Thoma. Uber unitare Darstellungen abzahlbarer, diskreter Gruppen. Math. Ann.,
153:111–138, 1964.
[39] Aaron Tikuisis and Wilhelm Winter. Decomposition rank of Z-stable C∗-algebras. Anal. PDE,
7(3):673–700, 2014.
[40] Wilhelm Winter. Nuclear dimension and Z-stability of pure C∗-algebras. Invent. Math.,
187(2):259–342, 2012.
NUCLEAR DIMENSION OF FINITELY-GENERATED NILPOTENT GROUP C*-ALGEBRAS 19
[41] Wilhelm Winter and Joachim Zacharias. The nuclear dimension of C ∗-algebras. Adv. Math.,
224(2):461–498, 2010.
Department of Mathematics, Miami University, Oxford, OH, 45056
E-mail address: [email protected]
E-mail address: [email protected]
|
1911.05500 | 1 | 1911 | 2019-11-13T14:24:44 | Functional Calculus for Elliptic Operators on Noncommutative Tori, I | [
"math.OA",
"math.FA"
] | In this paper, we introduce a parametric pseudodifferential calculus on noncommutative $n$-tori which is a natural nest for resolvents of elliptic pseudodifferential operators. Unlike in some previous approaches to parametric pseudodifferential calculi, our parametric pseudodifferential calculus contains resolvents of elliptic pseudodifferential operators that need not be differential operators. As an application we show that complex powers of positive elliptic pseudodifferential operators on noncommutative $n$-tori are pseudodifferential operators. This confirms a claim of Fathi-Ghorbanpour-Khalkhali. | math.OA | math |
FUNCTIONAL CALCULUS FOR ELLIPTIC OPERATORS ON
NONCOMMUTATIVE TORI, I
GIHYUN LEE AND RAPHAEL PONGE
Abstract. In this paper, we introduce a parametric pseudodifferential calculus on noncom-
mutative n-tori which is a natural nest for resolvents of elliptic pseudodifferential operators.
Unlike in some previous approaches to parametric pseudodifferential calculi, our parametric
pseudodifferential calculus contains resolvents of elliptic pseudodifferential operators that need
not be differential operators. As an application we show that complex powers of positive elliptic
pseudodifferential operators on noncommutative n-tori are pseudodifferential operators. This
confirms a claim of Fathi-Ghorbanpour-Khalkhali.
1. Introduction
UkUj " e2iπθjk UjUk,
j, k " 1, . . . , n.
Noncommutative tori are among the most well known examples of noncommutative spaces. For
instance, noncommutative 2-tori naturally arise from actions on circles by irrational rotations. In
general, the (smooth) noncommutative torus associated with a given anti-symmetric real n n-
matrix θ " pθjkq is a (Fr´echet) -algebra Aθ that is generated by unitaries U1, . . . , Un subject to
the relations,
When θ " 0 we recover the algebra C8pTnq of smooth functions on the ordinary torus Tn " pR{
Zqn. (We refer to Section 2 for a summary of the main definitions and properties regarding non-
commutative tori.) The recent work of Connes-Tretkoff [8] and Connes-Moscovici [7] is currently
spurring on an intensive research activity on the construction of a full differential geometric appa-
ratus on noncommutative tori (see, e.g., [6, 7, 8, 9, 13, 14, 15, 28, 29, 30, 38]). One specific issue
is the taking into account of the non-triviality of the Takesaki-Tomita modular automorphism
group. This is a purely noncommutative phenomenon.
An important tool in this line of research is the (local) short time asymptotic for heat semi-group
traces Trrae´tPs, where P is some Laplace-type operator and a P Aθ. In particular, the second
coefficient in this asymptotic is naturally interpreted as the so-called modular scalar curvature
(see [7, 8]). The derivation of this asymptotic heavily relies on the pseudodifferential calculus for
C-dynamical systems announced in the notes of Connes [3] and Baaj [2]. It was only recently
that detailed accounts on this ΨDO calculus became available (see [22, 23, 28]; see also [44]).
Nevertheless, there is a general understanding that the needed asymptotics can be obtained by
implementing in the setting of noncommutative tori the approach to the heat kernel asymptotic of
Gilkey [16] (see [7, 8, 28]). This actually requires a version with parameter of the pseudodifferential
calculus on noncommutative tori that is similar to the parametric ΨDO calculus of Gilkey [16]
and Shubin [41] (see [28]).
The parametric ΨDO calculus of [16, 41] is a natural nest for the resolvent pP ´ λq´1 when P
is an elliptic differential operator. However, when P is more generally a non-differential elliptic
ΨDO the resolvent pP ´ λq´1 falls out this parametric calculus. In fact, the calculus does not
contain non-differential ΨDOs, and so in this case even P and P ´ λ are not in this parametric
calculus. As a result, the approach of [16] does not allow us to derive short-time asymptotics
for traces TrrAe´tPs when P or A are not differential operators. Similarly, the approach of [41]
does not allow us to construct complex powers P s, s P C, when P is a "non-differential" elliptic
The
research for
this article was partially supported by NRF grants 2013R1A1A2008802 and
2016R1D1A1B01015971 (South Korea).
1
ΨDO (compare [40]). One approach to remedy is to use the weakly parametric pseudodifferential
calculus of Grubb-Seeley [19, 20].
The aim of this paper is to construct a parametric ΨDO calculus on noncommutative tori which
is a natural receptacle for the resolvent pP ´ λq´1, where P is an elliptic ΨDO that need not be
a differential operator. This parametric ΨDO calculus is inspired by the parametric Heisenberg
calculus of [32, 33]. Its construction avoids some of the technical difficulties of the construction
of the weakly parametric pseudodifferential calculus in [19, 20].
In addition, it contains as a
sub-calculus the version for noncommutative tori of the parametric pseudodifferential calculus
of [16, 41] (see [34]). This paper can considered as a first step toward a general functional calculus
for elliptic pseudodifferential operators on noncommutative tori. In forthcoming papers we are
planning on presenting applications of the parametric ΨDO calculus of this paper toward the
following objectives:
(i) Short-time asymptotics for traces Trrae´tPs, where P is an elliptic differential operator
and a P Aθ.
(ii) Construction of complex powers P s, s P C, of elliptic pseudodifferential operators P as
(iii) Precise study of the singularities of zeta functions ζpP ; sq " TrrP ´ss of elliptic pseudo-
differential operators P together with corresponding variation formulas when we let P
vary.
(iv) Extension of (i) to short-time asymptotics for traces TrrAe´tPs, where A and P are allowed
holomorphic families of pseudodifferential operators.
to be non-differential ΨDOs.
ÿ
jě0
In [16, 41] the parametric ΨDOs are associated with symbols with parameter,
(1.1)
ppx, ξ; λq „
pm´jpx, tξ; twλq " tm´jpm´jpx, ξ; λq,
pm´jpx, ξ; λq,
ř
ř
where w is a given positive number and the asymptotic expansion is meant in the sense that the
wq´Nq as ξ ` λ Ñ 8 with N arbitrarily large. Symbols of
remainder terms are Oppξ ` λ 1
differential operators ppx, ξq "
aαpxqξα fit into this framework, but the symbols of classical
ΨDOs ppx, ξq „
jě0 pm´jpx, ξq do not fit in as soon as there are infinitely many non-zero
homogeneous components pm´jpx, ξq.
The observation in [32, 33] is that we may afford to be somewhat loose with λ-decay. Namely,
it is enough to require the remainder terms in the asymptotic expansion (1.1) to be Opλdξ´Nq
where N is still arbitrarily large, but d is some fixed real number. This allows us to construct
complex powers P s, s P C, of a given elliptic ΨDO P as holomorphic families of ΨDOs, including
when P is not a differential operator. With some additional work this allows us to study the
singularities of the zeta functions ζpP ; sq " TrrP ´ss and ζpP, A; sq " TrrAP ´ss (where A is a
ΨDO) together with establishing uniform boundedness along vertical lines (at least when P is
strongly elliptic). A standard Mellin transform argument (see, e.g., [21]) then provides with short
time asymptotics for TrrAe´tPs. Therefore, this approach provides us with a sensible alternative
to using the weakly parametric pseudodifferential calculus of [19, 20].
The parametric pseudodifferential calculus presented in this paper implements this approach
on noncommutative tori. We refer to Section 3 for a review of the main definitions and properties
of the pseudodifferential calculus on noncommutative tori. Given any m P R, the space of m-th
order (standard) symbols SmpRn; Aθq consists of maps ρpξq P C8pRn; Aθq such that Bβ
ξ ρpξq is
Opξm´βq in Aθ. The ΨDO associated with a symbol ρpξq P SmpRn; Aθq is the (continuous)
ij
linear operator Pρ : Aθ Ñ Aθ given by
Pρu " p2πq´n
(1.2)
where the above integral is meant as an oscillating integral over Rn Rn and Rn Aθ Q ps, uq Ñ
αspuq is the -action of Rn on Aθ such that αspU mq " eismU m, m P Zn. We are more especially
interested in ΨDOs associated with classical symbols. They are symbols with an asymptotic
expansion ρpξq „
jě0 ρm´jpξq, where ρm´jpξq is homogeneous of degree m ´ j.
eisξρpξqα´spuqdsdξ,
u P Aθ,
ř
2
A first issue to address in the construction of the parametric pseudodifferential calculus is the
general shape of the domains of the parameters λ. For deriving short-time heat trace asymptotics
we may take these domains to be cones (where by a cone we shall always mean a cone with vertex at
the origin). However, the construction of complex powers of elliptic operators involves integrating
over contours in the complex plane that wind around the origin. Such contours cannot be inside
a cone with vertex at the origin. We take the domains to be (open) pseudo-cones, where by a
pseudo-cone we mean a connected set that agrees with a cone off some open disk about the origin
(see Section 4 for the precise definition). For instance, we obtain pseudo-cones by pasting/cutting
disks or annuli to/from cones.
Given a pseudo-cone Λ and d P R, a HoldpΛq-family in a locally convex space E is a holomophic
family that is OpΛdq in E as λ goes to 8 (the precise definition requires some uniformity condition;
see Definition 4.11). Intuitively speaking, the parametric ΨDO calculus of this paper is meant to
be a Hold-family version of the ΨDO calculus on noncommutative tori. The spaces of standard
symbols SmpRn; Aθq mentioned above are Fr´echet spaces. Given a pseudo-cone Λ and m, d P R,
the space of standard parametric symbols Sm,dpRn Λ; Aθq is just the space of HoldpΛq-families in
SmpRn; Aθq. The space of (classical) parametric symbols Sm,dpRn Λ; Aθq consists of parametric
symbols that have an asymptotic expansion,
ρpξ; λq „
ρm´jpξ; λq,
ρm´jptξ; twλq " tm´jρm´jpξ; λq,
t ą 0,
ÿ
jě0
where w is some positive number and the asymptotic expansion is meant in the sense that the
remainder terms can be made to be Opλdq Opξ´Nq in Aθ with N arbitrary large (see Section 4
for the precise meaning).
origin), then the homogeneous symbols ρm´jpξ; λq are defined on open sets of the form,
If we denote by Θ the conical part of Λ (i.e., Λ agrees with Θ off some open disk about the
(cid:32)
pξ, λq P pRnz0q C; λ P Θ or λ ă cξw
ΩcpΘq "
(
.
Suitably cut-off homogeneous symbols give rise to standard parametric symbols (Lemma 4.34).
This implies that classical parametric symbols are standard parametric symbols (see Proposi-
tion 4.39 for the precise relationship). There is also a version of Borel's lemma for parametric
symbols (see Proposition 4.43).
A parametric ΨDO is a family of ΨDOs Pρpλq associated via (1.2) with a parametric symbol
ρpξ; λq. We denote by Ψm,dpAθ; Λq the class of parametric ΨDOs that have a classical parametric
symbol in Sm,dpRn Λ; Aθq. The Hold-approach makes it simple to extend to parametric ΨDOs
the main properties of (non-parametric) ΨDOs via continuity arguments. This includes com-
position (Proposition 5.9), Sobolev regularity properties (Proposition 5.11), and Schatten-class
properties, including trace-class properties (see Proposition 5.14 and Proposition 5.15). For sake
of completeness we also introduce a class of parametric toroidal ΨDOs and show it agrees with
our original class of parametric ΨDOs (Proposition 5.25). This leads us to a characterization of
parametric smoothing operators (Proposition 5.31).
The heart of the matter of the paper is Section 6. Given an elliptic operator P of order w ą 0
with (classical) symbol ρpξq „
ř
jě0 ρw´jpξq we introduce its elliptic parameter cone,
ΘpPq "
rCz Sppρwpξqqs.
ď
We shall say that P is elliptic with parameter when ΘpPq ‰ H (which we assume thereon). We
may regard P ´ λ as an element of Ψw,1pAθ; Λq, even when P is not a differential operator. When
P is elliptic with parameter the parametric symbolic calculus of this paper allows us to get an
explicit construction of a parametrix for P ´ λ as an element of Ψ´w,´1pAθ; Λq, where Λ is any
pseudo-cone that is obtained by glueing to ΘpPq a disk about the origin (Theorem 6.8).
The existence of such a parametrix allows to give a somewhat precise localization of the
spectrum of P (see Theorem 6.10). More precisely, given any cone Θ such that Θz0 Ă ΘpPq,
there are at most finitely many eigenvalues of P that can be contained in Θ, and, furthermore,
}pP ´ λq´1} " Opλ´1q as λ Ñ 8 within Θ. This implies that any ray contained in ΘpPq that
ξPRnz0
3
does not cross the spectrum of P is automatically a ray of minimal growth (see Corollary 6.15).
The existence of such a ray is an important ingredient in the complex powers of elliptic operators
(see, e.g., [40]).
If we let ΘpPq be the cone obtained from ΘpPq by deleting rays through eigenvalues of P , then
ΘpPq is still a non-empty open cone (see Lemma 6.16). The main result of Section 6 states that the
resolvent pP ´λq´1 is an element of Ψ´w,´1pAθ; ΛpPqq, where ΛpPq is obtained by glueing to ΘpPq
any disk or marked disk about the origin that does not contain any eigenvalue of P (Theorem 6.19).
Combining this result with the properties of parametric ΨDOs provides us with sharp Sobolev
regularity properties and Schatten-class properties for the resolvent pP ´ λq´1. Namely, for any
s P R, the resolvent pP´λq´1 is an Hol´1`apΛpPqq-family in L pH psq
q for every a P r0, 1s
(Proposition 6.21) and, if we set p " nw´1, then this is also an Hol´1`pq´1pΛpPqq-family in L pq,8q
for every q ě p (Proposition 6.22). Furthermore, we have sharp asymptotic estimates for the traces
of operators of the form,
, H ps`awq
θ
θ
A0pP ´ λq´1A1 AN´1pP ´ λq´1AN ,
Aj P ΨajpAθq,
provided that ´N w ` a0 ` ` aN ă ´n (see Proposition 6.23). Such operators often appear in
resolvent expansions.
In [27] we shall apply the above results to construct complex powers P s of an elliptic ΨDO
as holomorphic families of ΨDOs. As a preview of this we show when P P ΨwpAθq is an elliptic
ΨDO with positive principal symbol and positive spectrum its complex powers P z, z P C, are
ΨDOs (Theorem 7.3). This implies that, for any elliptic ΨDO P P ΨwpAθq the absolute value
P " ?
It was
mentioned without proof in [12] that complex powers of positive elliptic ΨDOs on noncommutative
2-tori are ΨDOs. Thus, our results confirm the claim of [12]. In any case, we postpone to [27] a
more thorough account on complex powers of elliptic ΨDOs on noncommutative tori.
P P and its complex powers Pz, z P C, are ΨDOs as well (Theorem 7.5).
The paper is organized as follows. In Section 2, we give an overview the main facts on noncom-
mutative tori. In Section 3, we review the main definitions and properties regarding pseudodiffer-
ential operators on noncommutative tori. In Section 4, we introduce our classes of symbols with
parameter and derive some of their properties. In Section 5, we define our parametric pseudodif-
ferential operators and establish their main properties. In Section 6, given any ΨDO P that is
elliptic with parameter, we construct a parametrix for P ´ λ in our parametric ΨDO calculus and
show that the resolvent pP ´λq´1 is an element of this calculus. These results are used in Section 7
to show that the complex powers of positive elliptic ΨDOs are ΨDOs. In Appendix A, we review
basic facts on improper integrals with values in locally convex spaces. In Appendix B, we describe
the strong topology of L pA 1
bounded sets in A 1
θ .
θ , Aθq in terms of Sobolev norms. This includes a characterization of
Acknowlegements. G.L. acknowledges the support of BK21 PLUS SNU, Mathematical Sciences
Division (South Korea). R.P. wishes to thank University of New South Wales (Sydney, Australia)
and University of Qu´ebec at Montr´eal (Montr´eal, Canada) for their hospitality during the prepa-
ration of this manuscript.
2. Noncommutative Tori
In this section, we review the main definitions and properties of noncommutative n-tori, n ě 2.
We refer to [22], and the references therein (more especially [5, 35, 36]), for more comprehensive
accounts, including proofs of the results mentioned in this section.
Throughout this paper, we let θ " pθjkq be a real anti-symmetric n n-matrix (n ě 2). We
denote by θ1, . . . , θn its column vectors. In what follows we also let Tn " Rn{2πZn be the ordinary
n-torus, and we equip L2pTnq with the inner product,
ż
(2.1)
pξηq "
ξpxqηpxq ¯dx,
ξ, η P L2pTnq,
Tn
4
where we have set ¯dx " p2πq´ndx. For j " 1, . . . , n, let Uj : L2pTnq Ñ L2pTnq be the unitary
operator defined by
pUjξqpxq " eixj ξ px ` πθjq ,
ξ P L2pTnq.
We then have the relations,
UkUj " e2iπθjk UjUk,
j, k " 1, . . . , n.
(2.2)
The noncommutative torus Aθ is the C-algebra generated by the unitary operators U1, . . . , Un.
For θ " 0 we obtain the C-algebra C 0pTnq of continuous functions on the ordinary n-torus Tn.
Note that (2.2) implies that Aθ is the closure in L pL2pTnqq of the algebra A 0
θ is the
span of the unitary operators,
θ , where A 0
Let τ : L pL2pTnqq Ñ C be the state defined by the constant function 1, i.e.,
U k :" U k1
1 U kn
ż
n ,
τpTq " pT 11q "
`
pT 1qpxq ¯dx,
"
Tn
k " pk1, . . . , knq P Zn.
`
L2pTnq
.
T P L
In particular, we have
τ
1
0
U k
"
The GNS construction (see, e.g., [1]) allows us to associate with τ a -representation of Aθ as
if k " 0,
otherwise.
This induces a continuous linear trace on the C-algebra Aθ.
follows. Let pq be the sesquilinear form on Aθ defined by
(2.3)
The family tU k; k P Znu is orthonormal with respect to this sesquilinear form. In particular, we
have a pre-inner product on the dense subalgebra A 0
θ .
puvq " τ puvq ,
u, v P Aθ.
Definition 2.1. Hθ is the Hilbert space arising from the completion of A 0
pre-inner product (2.3).
θ with respect to the
When θ " 0 we recover the Hilbert space L2pTnq with the inner product (2.1). In what follows
we shall denote by } }0 the norm of Hθ. This notation allows us to distinguish it from the norm
of Aθ, which we denote by } }.
By construction pU kqkPZn is an orthonormal basis of Hθ. Thus, every u P Hθ can be uniquely
written as
ÿ
u "
`
uk "
uU k
ukU k,
,
(2.4)
where the series converges in Hθ. When θ " 0 we recover the Fourier series decomposition in
L2pTnq. By analogy with the case θ " 0 we shall call the series
kPZn ukU k in (2.4) the Fourier
series of u P Hθ.
Proposition 2.2. The following holds.
ř
kPZn
(1) The multiplication of A 0
This provides us with a unital -representation of Aθ. In particular, we have
θ uniquely extends to a continuous bilinear map Aθ Hθ Ñ Hθ.
}u} " sup
}v}0"1
@u P Aθ.
}uv}0
(2) The inclusion of A 0
θ into Hθ uniquely extends to a continuous embedding of Aθ into Hθ.
In particular, the 2nd part allows us to identify any u P Aθ with the sum of its Fourier series
The natural action of Rn on Tn by translation gives rise to an action on L pL2pTnqq which
in Hθ. In general this Fourier series need not converge in Aθ.
induces a -action ps, uq Ñ αspuq on Aθ such that
αspU kq " eiskU k,
for all k P Zn and s P Rn.
5
This action is strongly continuous, and so we obtain a C-dynamical system pAθ, Rn, αq. We are
especially interested in the subalgebra of the smooth elements of this C-dynamical system, i.e.,
the smooth noncommutative torus,
Aθ :" tu P Aθ; αspuq P C8pRn; Aθqu .
All the unitaries U k, k P Zn, are contained in Aθ, and so Aθ is a dense subalgebra of Aθ. When
θ " 0 we recover the algebra C8pTnq of smooth functions on the ordinary torus Tn.
iBsj . When θ " 0 the derivation δj is just the derivation Dxj " 1
i
B
Bxj
For j " 1, . . . , n, let δj : Aθ Ñ Aθ be the derivation defined by
u P Aθ,
δjpuq " Dsj αspuqs"0,
where we have set Dsj " 1
"
on C8pTnq. In general, for j, l " 1, . . . , n, we have
Uj
0
δjpUlq "
More generally, given u P Aθ and β P Nn
δβpuq " Dβ
if l " j,
if l ‰ j.
0 , β " N , we define
s αspuqs"0 " δβ1
1 δβn
n puq.
››δβpuq
›› ,
In what follows, we endow Aθ with the locally convex topology defined by the semi-norms,
Aθ Q u ÝÑ
β P Nn
0 .
(2.5)
With the involution inherited from Aθ this turns Aθ into a (unital) Fr´echet -algebra. It can be
further shown that Aθ is even a nuclear Fr´echet-Montel space (see, e.g., [22]).
(
Aθ). Given any u P Aθ we shall denote by Sppuq its spectrum, i.e.,
.
(cid:32)
λ P C; u ´ λ R A ´1
θ ) the group of invertible elements of Aθ (resp.,
In what follows, we denote by A ´1
(resp., A´1
Sppuq "
θ
θ
Proposition 2.3 (see [4, 22]). The following holds.
θ " A´1
(1) We have A ´1
(2) Aθ is stable under holomorphic functional calculus.
(3) For all u P Aθ, we have
θ X Aθ.
Sppuq " tλ P C; u ´ λ : Hθ Ñ Hθ is not a bijectionu .
Let A 1
θ be the topological dual of Aθ. We equip it with its strong topology, i.e., the locally
convex topology generated by the semi-norms,
xv, uy,
v ÝÑ sup
uPB
B Ă Aθ bounded.
It is tempting to think of elements of A 1
θ as distributions on Aθ. This is consistent with the
definition of distributions on Tn as continuous linear forms on C8pTnq. Any u P Aθ defines a
linear form on Aθ by
xu, vy " τpuvq
for all v P Aθ.
xu, vy " pvuq " puvq .
Note that, for all u, v P Aθ, we have
(2.6)
In particular, given any u P Aθ, the map v Ñ xu, vy is a continuous linear form on Aθ. This gives
rise to a continuous embedding of Aθ into A 1
θ . In view of (2.6) it uniquely extends to a continuous
embedding of Hθ into A 1
θ . Namely,
given any v P A 1
(2.7)
Here the unitaries U k, k P Zn, are regarded as elements of A 1
is the sum of its Fourier series in A 1
θ . This allows us to extend the definition of Fourier series to A 1
θ . It can be shown that every u P A 1
v,pU kqD
@
θ its Fourier series is the series,
where vk :"
ÿ
vkU k,
kPZn
.
θ
θ (see [22]).
6
3. Pseudodifferential Operators on Noncommutative Tori
In this section, we recall the main definitions and properties of the pseudodifferential calculus
on noncommutative tori [2, 3, 22, 23]. The exposition closely follows [22, 23] (see also [44]).
3.1. Symbols on noncommutative tori. There are various classes of symbols on noncommu-
tative tori. First, we have the standard symbols.
Definition 3.1 (Standard Symbols; see [2, 3]). SmpRn; Aθq, m P R, consists of maps ρpξq P
C8pRn; Aθq such that, for all multi-orders α and β, there exists Cαβ ą 0 such that
}δαBβ
ξ ρpξq} ď Cαβ p1 ` ξqm´β
@ξ P Rn.
Remark 3.2 (see [22]). SmpRn; Aθq is a Fr´echet space with respect to the topology generated by
the semi-norms,
Remark 3.3. Let pmjqjě0 Ă R be a decreasing sequence such that mj Ñ ´8. Given ρpξq in
C8pRn; Aθq and ρjpξq P SmjpRn; Aθq, j ě 0, we shall write ρpξq „
(3.1)
(3.2)
p1 ` ξq´m`β}δαBβ
sup
ξPRn
α`βďN
pmq
N pρq :" sup
p
ÿ
ρpξq ´
ρjpξq P SmNpRn; Aθq
jăN
N P N0.
ξ ρpξq},
ř
jě0 ρjpξq when
@N ě 0.
Note this implies that ρpξq P Sm0pRn; Aθq.
Definition 3.4. S pRn; Aθq consists of maps ρpξq P C8pRn; Aθq such that, for all N ě 0 and
multi-orders α, β, there exists CN αβ ą 0 such that
Ş
ξ ρpξq} ď CN αβ p1 ` ξq´N
}δαBβ
mPR SmpRn; Aθq.
@ξ P Rn.
Remark 3.5. S pRn; Aθq "
Definition 3.6 (Homogeneous Symbols). SqpRn; Aθq, q P C, consists of maps ρpξq P C8pRnz0; Aθq
that are homogeneous of degree q, i.e., ρptξq " tqρpξq for all ξ P Rnz0 and t ą 0.
Remark 3.7. If ρpξq P SqpRn; Aθq and χpξq P C8
p1 ´ χpξqqρpξq P S(cid:60)qpRn; Aθq.
Definition 3.8 (Classical Symbols; see [2]). SqpRn; Aθq, q P C, consists of maps ρpξq P C8pRn; Aθq
that admit an asymptotic expansion,
c pRnq is such that χpξq " 1 near ξ " 0, then
ÿ
jě0
ρpξq „
`
›››δαBβ
ξ
ρq´jpξq,
ÿ
ρ ´
ρq´j
pξq
ρq´j P Sq´jpRn; Aθq,
››› ď CN αβξ(cid:60)q´N´β.
where „ means that, for all N ě 0 and multi-orders α, β, there exists CN αβ ą 0 such that, for all
ξ P Rn with ξ ě 1, we have
(3.3)
jăN
Remark 3.9. SqpRn; Aθq Ă S(cid:60)qpRn; Aθq.
ř
Remark 3.10. Let χpξq P C8
the sense of (3.3) if and only if ρpξq „
Example 3.11. Any polynomial map ρpξq "
Example 3.12. Set xξy " p1 ` ξ2q 1
ř
jě0 ρq´jpξq in
c pRnq be such that χpξq " 1 near ξ " 0. Then ρpξq „
ř
jě0p1 ´ χpξqqρq´jpξq in the sense of (3.2).
αďm aαξα, aα P Aθ, m P N0, is in SmpRn; Aθq.
2 , ξ P Rn. Then xξyq P SqpRn; Aθq for all q P C (see, e.g., [22]).
7
3.2. Amplitudes and oscillating integrals. We briefly recall the construction of the Aθ-valued
oscillating integral in [22] (see also [37]). In this setting the oscillating integral is defined on Aθ-
valued amplitudes.
Definition 3.13 (Amplitudes; see [22]). AmpRn Rn; Aθq, m P R, consists of maps aps, ξq in
C8pRn Rn; Aθq such that, for all multi-orders α, β, γ, there is Cαβγ ą 0 such that
(3.4)
ξ aps, ξq} ď Cαβγ p1 ` s ` ξqm @ps, ξq P Rn Rn.
}δαBβ
Each space AmpRn Rn; Aθq, m P R, is a Fr´echet space with respect to the locally convex
s Bγ
s Bγ
ξ aps, ξq}, N P N0.
topology generated by the semi-norms,
(3.5)
q
pmq
N paq :"
sup
sup
α`β`γďN
ps,ξqPRnRn
p1 ` s ` ξq´m}δαBβ
ij
If aps, ξq P AmpRn Rn; Aθq with m ă ´2n, then the estimates (3.4) ensure us that the Aθ-
valued map ps, ξq Ñ eisξaps, ξq is absolutely integrable on Rn Rn (see [22]). Therefore, we may
define
(3.6)
where we have set ¯dξ " p2πq´ndξ. This defines a continuous linear map from AmpRn Rn; Aθq
to Aθ for every m ă ´2n (see [22]).
eisξaps, ξqds ¯dξ,
Ť
J0paq :"
In the following we set A`8pRn Rn; Aθq :"
mPR AmpRn Rn; Aθq.
Proposition 3.14 (see [22]). J0 uniquely extends to a linear map J : A`8pRn Rn; Aθq Ñ Aθ
that induces a continuous linear map on each amplitude space AmpRn Rn; Aθq, m P R.
More specifically, given χps, ξq P C8
be the differential operator on R2n defined by
c pRn Rnq such that χps, ξq " 1 near ps, ξq " p0, 0q, let L
L :" χps, ξq ` 1 ´ χps, ξq
s2 ` ξ2
pξjDsj ` sjDξjq,
ÿ
1ďjďn
where we have set Dxj " 1
iBxj , j " 1, . . . , n. Note that Lpeisξq " eisξ. Moreover, given any
m P R, it can be shown that the transpose Lt (in the sense of differential operators) gives rise
to a continuous linear map Lt : AmpRn Rn; Aθq Ñ Am´1pRn Rn; Aθq (see [22]). Successive
integrations by parts show that, given any aps, ξq P AmpRn Rn; Aθq, m P R, we have
Jpaq "
eisξpLtqNraps, ξqsds ¯dξ,
where N is any non-negative integer ą m ` 2n.
We call Jpaq the oscillating integral with amplitude aps, ξq.
ť
eisξaps, ξqds ¯dξ.
integral
3.3. ΨDOs on noncommutative tori. Let ρpξq P SmpRn; Aθq, m P R. Given any u P Aθ, it
can be shown that ρpξqα´spuq is an amplitude in Am`pRn Rn; Aθq, where m` " maxpm, 0q
(see [22]). Therefore, we can define
It is convenient write it as an
ij
ij
Pρu "
eisξρpξqα´spuqds ¯dξ,
where the above integral is meant as an oscillating integral. This defines a continuous linear
operator Pρ : Aθ Ñ Aθ (see [22]).
Definition 3.15. ΨqpAθq, q P C, consists of all linear operators Pρ : Aθ Ñ Aθ with ρpξq in
SqpRn; Aθq.
Remark 3.16. If P " Pρ with ρpξq in SqpRn; Aθq, ρpξq „
ρq´jpξq, then ρpξq is called a symbol
for P . This symbol is not unique, but its restriction to Zn and its class modulo S pRn; Aθq are
unique (see Remark 3.33). As a result, the homogeneous symbols ρq´jpξq are uniquely determined
by P . The leading symbol ρqpξq is called the principal symbol of P .
ř
8
ř
αďm aαδα, aα P Aθ (see [3, 5]).
ř
1 ` ` δ2
kPZn ukU k P Hθ;
ř
Example 3.17. A differential operator on Aθ is of the form P "
This is a ΨDO of order m with symbol ρpξq "
ř
aαξα (see [22]).
Example 3.18. The Laplacian ∆ :" δ2
n is a selfadjoint unbounded operator on Hθ
with domain Domp∆q :" tu "
kPZn k2uk2 ă 8u. It is isospectral to the
Laplacian on Tn. For all s P C set Λs " p1 ` ∆qs. Then Λs is the ΨDO with symbol xξys, and so
Λs P ΨspAθq (see [22]). In fact, the family pΛsqsPC form a 1-parameter group of ΨDOs.
3.4. Composition of ΨDOs. Suppose we are given symbols ρ1pξq P Sm1pRn; Aθq, m1 P R, and
ρ2pξq P Sm2pRn; Aθq, m2 P R. As Pρ1 and Pρ2 are linear operators on Aθ, the composition Pρ1Pρ2
makes sense as such an operator. In addition, we define the map ρ17ρ2 : Rn Ñ Aθ by
ρ17ρ2pξq "
(3.7)
Here, for every ξ P Rn, the map pt, ηq Ñ ρ1pξ`ηqα´trρ2pξqs is an amplitude, and so the right-hand
side makes sense as an oscillating integral (see [2, 23]).
Proposition 3.19 (see [2, 3, 23]). Let ρ1pξq P Sm1pRn; Aθq and ρ2pξq P Sm2pRn; Aθq, m1, m2 P R.
eitηρ1pξ ` ηqα´trρ2pξqsdt ¯dη,
ř
ξ P Rn.
ij
1
α!Bα
(1) ρ17ρ2pξq P Sm1`m2pRn; Aθq, and we have ρ17ρ2pξq „
(2) The operators Pρ1 Pρ2 and Pρ17ρ2 agree.
In the case of classical ΨDOs we further have the following result.
ř
ř
Proposition 3.20 (see [2, 3, 23]). Let P1 P Ψq1pAθq, q1 P C, have symbol ρ1pξq „
and let P2 P Ψq2pAθq, q2 P C, have symbol ρ2pξq „
jě0 ρ2,q2´jpξq.
(1) ρ17ρ2pξq P Sq1`q2pRn; Aθq with ρ17ρ2pξq „
pρ17ρ2qq1`q2´jpξq, where
Bα
ξ ρ1,q1´kpξqδαρ2,q2´lpξq,
ξ ρ1pξqδαρ2pξq.
ř
pρ17ρ2qq1`q2´jpξq "
j ě 0.
ÿ
(3.8)
1
α!
k`l`α"j
jě0 ρ1,q1´jpξq,
ř
jě0 ρq´jpξq.
(2) The composition P1P2 " Pρ17ρ2 is contained in Ψq1`q2pAθq.
3.5. Adjoints of ΨDOs and action on A 1
adjoint is any linear operator P : Aθ Ñ Aθ such that
θ . Given a linear operator P : Aθ Ñ Aθ, a formal
@u, v P Aθ,
pP uvq " puP vq
ij
where pq is the inner product (2.3). When it exists a formal adjoint must be unique.
Let ρpξq P SmpRn; Aθq, m P R, and set
ρ‹pξq "
eitηα´trρpξ ` ηqsdt ¯dη,
ξ P Rn.
(3.9)
For every ξ P Rn, the map pt, ηq Ñ α´trρpξ ` ηqs is an amplitude, and so the right-hand side
makes sense as an oscillating integral (see [23]).
Proposition 3.21 (see [2, 23]). Let ρpξq P SmpRn; Aθq, m P R.
ř
(1) ρ‹pξq P SmpRn; Aθq, and we have ρ‹pξq „
(2) Pρ‹ is the formal adjoint of Pρ.
1
α! δαBα
ξ rρpξqs.
(1) ρ‹pξq P S
(3.10)
We have the following version of this result for classical ΨDOs.
Proposition 3.22 (see [2, 23]). Let P P ΨqpAθq, q P C, have symbol ρpξq „
ř
sqpRn; Aθq, and we have ρ‹pξq „
ρ‹sq´jpξq, where
ξ rρq´kpξqs ,
δαBα
sqpAθq.
ρ‹sq´jpξq "
ÿ
k`α"j
1
α!
j ě 0.
(2) The formal adjoint P " Pρ‹ is contained in Ψ
An application of Proposition 3.21 is the following extension result.
Proposition 3.23 (see [23]). Let ρpξq P SmpRn; Aθq, m P R. Then Pρ uniquely extends to a
continuous linear operator Pρ : A 1
θ Ñ A 1
θ .
9
2 is a (classical) ΨDO of order s (cf. Example 3.18). Therefore, by Proposition 3.23
n be the Laplacian on Aθ. Given any s P R, the operator
, s P R, consists of all u P A 1
θ such
1``δ2
3.6. Sobolev spaces. Let ∆ " δ2
Λs " p1 ` ∆q s
it uniquely extends to a continuous linear operator Λs : A 1
Definition 3.24 (see [23, 43, 47]). The Sobolev space H psq
that Λsu P Hθ. It is equipped with the norm,
θ
θ Ñ A 1
θ .
u P H psq
Proposition 3.25. H psq
}u}s " }Λsu}0,
ř
ÿ
is a Hilbert space.
In terms of the Fourier decomposition u "
ukU k in A 1
θ we have
p1 ` k2qsuk2 ă 8,
u P H psq
ÿ
u P H psq
θ ðñ
s "
}u}2
p1 ` k2qsuk2,
(3.11)
kPZn
.
θ
θ
θ
.
Remark 3.31. S pZn; Aθq "
kPZn
č
ď
We have the following version of Sobolev's embedding theorem.
Proposition 3.26 (see [23, 47]). If t ą s, then the inclusion of H ptq
θ
into H psq
θ
is compact.
As the following result shows the Sobolev spaces provide us with a natural "topological" scale
of Hilbert spaces interpolating between Aθ and A 1
θ .
Proposition 3.27 (see [23, 31, 43]). The following holds.
(1) We have
Aθ "
H psq
θ
and
H psq
θ " A 1
θ .
(2) The topology of Aθ is generated by the Sobolev norms } }s, s P R.
(3) The strong topology of A 1
We have the following Sobolev mapping properties of ΨDOs on noncommutative tori.
sPR
θ coincides with the inductive limit of the H psq
sPR
θ
-topologies.
Proposition 3.28 (see [23]). Let ρpξq P SmpRn; Aθq, m P R.
(1) Pρ uniquely extends to continuous linear operator Pρ : H ps`mq
for every s P R.
(2) If m ď 0, then Pρ uniquely extends to a bounded operator Pρ : Hθ Ñ Hθ. This operator
Ñ H psq
θ
θ
is compact when m ă 0.
Let A Zn
θ denote the space of sequences with values in Aθ that are indexed by Zn. In addition, let
θ Ñ A Zn
3.7. Toroidal symbols and ΨDOs. In this section, we recall the relationship between standard
ΨDOs as defined above and the toroidal ΨDOs considered in [18, 29, 39].
pe1, . . . , enq be the canonical basis of Rn. For i " 1, . . . , n, the difference operator ∆i : A Zn
is defined by
.
The operators ∆1, . . . , ∆n pairwise commute. For α " pα1, . . . , αnq P Nn
We similarly define backward difference operators ∆ " ∆
defined by
1 ∆αn
n .
θ Ñ A Zn
is
∆iuk " uk`ei ´ uk,
0 , we set ∆α " ∆α1
n , where ∆i : A Zn
αn
θ
θ
θ
pukqkPZn P A Zn
1 ∆
pukqkPZn P A Zn
α1
θ
.
Definition 3.29 ([29, 39]). SmpZn; Aθq, m P R, consists of sequences pρkqkPZn Ă Aθ such that,
for all multi-orders α and β, there is Cαβ ą 0 such that
Definition 3.30. S pZn; Aθq consists of sequences pρkqkPZn Ă Aθ such that, for all N P N0 and
α P Nn
0 , there is CN α ą 0 such that
∆iuk " uk ´ uk´ei,
››δα∆βρk} ď Cαβp1 ` kqm´β
Ş
}δαρk} ď CN αp1 ` kq´N
mPR SmpZn; Aθq.
10
@k P Zn.
@k P Zn.
Following [29] (see also [18]) the ΨDO associated with a toroidal symbol pρkqkPZn P SmpZn; Aθq,
m P R, is the linear operator P : Aθ Ñ Aθ given by
ÿ
P u "
ukρkU k,
u "
ukU k P Aθ.
ÿ
kPZn
Any standard ΨDO is a toroidal ΨDO. More precisely, we have the following result.
Proposition 3.32 (see [8, 22, 39]). Let ρpξq P SmpRn; Aθq, m P R.
(1) The restriction of ρpξq to Zn is a toroidal symbol pρpkqqkPZn in SmpZn; Aθq.
(2) If pρpkqqkPZn P S pZn; Aθq, then ρpξq P S pRn; Aθq.
(3) For all u "
kPZn ukU k in Aθ, we have
ř
ÿ
kPZn
Pρu "
ukρpkqU k.
(3.12)
Remark 3.33. Using (3.12) we get
ρpkq " PρpU kqpU kq´1 " PρpU kqpU kq
(3.13)
Thus, the restriction to Zn of ρpξq is uniquely determined by Pρ. Combining this with the 2nd
part shows that the class of ρpξq modulo S pRn; Aθq is uniquely determined by Pρ.
for all k P Zn.
Conversely, toroidal symbols can be extended to standard symbols of the same order as follows.
φpξq " 1.
Lemma 3.34 ([39, Lemma 4.5.1]). There exists a function φpξq P S pRnq such that
ş
(i) φp0q " 1 and φpkq " 0 for all k P Znz0.
(ii) For every multi-order α, there is φαpξq P S pRnq such that Bα
(iii)
Let φpξq P S pRnq be a function satisfying the properties (i) -- (iii) of Lemma 3.34. Given any
ÿ
toroidal symbol pρkqkPZn in SmpZn; Aθq, m P R, define
ř
ř
φpξ ´ kqρk,
ρpξq "
(3.14)
kPZn
Given any k P Zn, we have ρpkq "
(cid:96)PZn φpk ´ (cid:96)qρ(cid:96) "
(cid:96)PZn δ(cid:96),kρ(cid:96) " ρk. Therefore, this provides
us with an extension map from SmpZn; Aθq to SmpRn; Aθq.
Proposition 3.35 (see [22, 29, 39]). Let pρkqkPZn P SmpZn; Aθq, m P R, and denote by P the
corresponding toroidal ΨDO.
ξ φpξq " ∆
ξ P Rn.
α
φαpξq.
(1) ρpξq is a standard symbol in SmpRn; Aθq.
(2) The operators P and Pρ agree.
Combining together Proposition 3.32 and Proposition 3.35 we arrive at the following result.
Proposition 3.36 (see [22]). For every m P R, the classes of toroidal and standard ΨDOs of
order m agree.
By construction the extension map pρkqkPZn Ñ ρpξq given by (3.14) is a right-inverse of the
restriction map ρpξq Ñ pρpkqqkPZn . It is also interesting to look at the reverse composition.
Proposition 3.37 (see [22]). Let ρpξq P SmpRn; Aθq, m P R, and denote by ρpξq the exten-
sion (3.14) of pρpkqqkPZn .
(1) ρpξq is a standard symbol in SmpRn; Aθq which agrees with ρpξq on Zn.
(2) P ρ " Pρ and ρpξq ´ ρpξq P S pRnq.
3.8. Smoothing operators. A linear operator R : Aθ Ñ A 1
to a continuous linear operator R : A 1
smoothing operator to A 1
θ is unique.
Proposition 3.38 ([22, 23]). Let R : Aθ Ñ Aθ be a linear map. The following are equivalent.
θ Ñ Aθ. Note that, as Aθ is dense in A 1
θ is called smoothing when it extends
θ , the extension of a
(i) R is a smoothing operator.
(ii) There is ρpξq P S pRn; Aθq such that R " Pρ, i.e., P P Ψ´8pAθq.
11
(iii) R extends to a continuous linear map R : H psq
θ Ñ H ptq
θ
for all s, t P R.
Remark 3.39. If R is in all the spaces ΨqpA q, q P C, then by Proposition 3.28 it extends to
a continuous linear map R : H psq
for all s, t P R, and so by using (ii) we see that
R P Ψ´8pAθq. Combining this with Remark 3.5 shows that Ψ´8pAθq "
θ Ñ H ptq
qPC ΨqpAθq.
θ
Ş
3.9. Ellipticity and parametrices. In this section, we recall the main facts regarding elliptic
operators.
Definition 3.40. An operator P P ΨqpAθq, q P C, is elliptic when its principal symbol ρqpξq is
invertible for all ξ P Rnz0.
Remark 3.41. The ellipticity condition implies that ρqpξq´1 P S´qpRn; Aθq (see [2, 23]).
Example 3.42. Suppose that the principal symbol ρqpξq of P is such that there is c ą 0 such that
(3.15)
pρqpξqηηq ě cξq}η}2
0
for all η P Hθ and ξ P Rnzt0u.
Then P is elliptic (see [23]). This condition is satisfied in the following examples:
- The (flat) Laplacian ∆ " δ2
- The conformal deformations k∆k, k P Aθ, k ą 0. These operators were introduced by
Connes-Tretkoff [8]. They were considered by various other authors as well.
1 ` ` δ2
n.
- The Laplace-Beltrami operators of [24].
Theorem 3.43 (see [2, 3, 23]). Let P P ΨqpAθq, q P C, be elliptic with symbol ρpξq „
ř
jě0 ρq´jpξq.
(1) P admits a parametrix Q P Ψ´qpAθq, i.e.,
ř
(2) Any parametrix Q P Ψ´qpAθq has symbol σpξq „
σ´qpξq " ρqpξq´1,
ρqpξq´1Bα
1
α!
P Q " QP " 1 mod Ψ´8pAθq.
ÿ
σ´q´jpξq " ´
(3.16)
(3.17)
k`l`α"j
lăj
jě0 σ´q´jpξq, where
ξ ρq´kpξqδασ´q´lpξq,
j ě 1.
An important application of Theorem 3.43 is the following version of the elliptic regularity
theorem.
Proposition 3.44 ([23]). Let P P ΨqpAθq, q P C, be elliptic, and set m " (cid:60)q.
(1) For any s P R and u P Aθ
1, we have
P u P H psq
θ ðñ u P H ps`mq
θ
.
(3) If m ą 0, then the operator P ´ λ is hypoelliptic in the above sense for every λ P C.
θ , we have
(2) The operator P is hypoelliptic, i.e., for any u P A 1
P u P Aθ ðñ u P Aθ.
(
(cid:32)
θ ; P u " 0
u P A 1
(cid:32)
If (cid:60)q ą 0, then, for every λ P C, we also have
u P A 1
Corollary 3.45 ([23]). Let P P ΨqpAθq, q P C, be elliptic. Then
Ă Aθ.
(
θ ; pP ´ λqu " 0
kerpP ´ λq :"
ker P :"
Ă Aθ.
12
3.10. Spectral theory of elliptic ΨDOs. Let P P ΨqpAθq be elliptic with m :" (cid:60)q ą 0. We
sqpAθq by Proposition 3.21.
shall regard P as an unbounded operator of Hθ with domain H pmq
. We also let P P Ψ¯qpAθq be
its formal adjoint. Recall that P P Ψ
Proposition 3.46 (see [2, 3, 23]). The following holds.
θ
(1) The operator P with domain H pmq
(2) The adjoint of P is the formal adjoint P with domain H pmq
(3) If P is formally selfadjoint (resp., P commutes with its formal adjoint), then P is selfdjoint
is closed, Fredholm and has closed range.
.
θ
θ
(resp., normal).
Definition 3.47. The resolvent set of P consists of all λ P C such that P ´ λ : H pmq
θ Ñ Hθ is a
bijection with bounded inverse. The spectrum of P , denoted by SppPq, is the complement of its
resolvent set.
Remark 3.48. If λ P Cz SppPq, then pP ´ λq´1 P L pHθ, H pmq
Remark 3.49. It can be shown that SppP q " tλ; λ P SppPqu (see [23, Corollary 12.8]).
Proposition 3.50 (see [23]). There are only two possibilities for the spectrum of P . Either
SppPq is all C, or this is an unbounded discrete set consisting of isolated eigenvalues with finite
multiplicity.
q.
θ
In the special case of normal operators we even obtain the following result.
Proposition 3.51 (see [23]). Suppose that P is normal and SppPq ‰ C (e.g., P is selfadjoint).
(1) We have an orthogonal decomposition,
(3.18)
Hθ "
kerpP ´ λq.
à
λPSppPq
(2) There is an orthonormal basis pe(cid:96)q(cid:96)ě0 of Hθ such that e(cid:96) P Aθ and P e(cid:96) " λ(cid:96)e(cid:96) with
λ(cid:96) Ñ 8 as (cid:96) Ñ 8.
3.11. Schatten-class properties of ΨDOs. We denote by K the closed two-sided ideal of
compact operators on Hθ. For p ě 1 we let L p and L pp,8q be the corresponding Schatten and
weak Schatten classes equipped with their usual norms (see, e.g., [17, 42]). In particular, we have
T T counted with multiplicity. Moreover, L p
p
ÿ
,
N 1´ 1
L pp,8q "
(
p ě 1
(
,
µkpTq " Oplog Nq
(cid:32)
(
(cid:32)
`
T P K ; TrTp ă 8
L p "
ÿ
(cid:32)
T P K ;
µkpTq " O
kăN
T P K ;
L p1,8q "
Here µkpTq is the pk ` 1q-th eigenvalue of T " ?
`
1
and L pp,8q are Banach ideals with respect to their natural norms,
ÿ
TrTp
p ,
"
µkpTq
ÿ
*
T P L p, p ě 1,
*
µkpTq
}T}L p1,8q " sup
Ně1
"
}T}L p "
kăN
1
N´1` 1
p
kăN
}T}L p1,8q " sup
Ně2
log N
kăN
Proposition 3.52 (see [23]). Let ρpξq P SmpRn; Aθq, m ă 0.
,
T P L pp,8q, p ą 1,
,
T P L p1,8q.
p ą 1,
.
(1) If m P r´n, 0q and p " nm´1, then Pρ P L pp,8q, and hence Pρ P L q for all q ą p.
(2) If m ă ´n, then Pρ P L 1.
13
4. Parametric Symbols
In this section, we introduce our classes of symbols with parameter. These classes are similar
Throughout the rest of this paper, we denote by C the complex plane with the origin deleted,
to the classes introduced in [32, 33].
i.e., C " Czt0u.
We refer to [22, Appendix C], and the references therein, for background on differentiable maps
In particular, if U Ă Rd is open, then a map
with values in a given locally convex space E .
f : U Ñ E is C N , N ě 0, when all the partial derivatives of order ď N exist and are continuous
on U . The map is C8 or smooth when it is C N for all N ě 0. Moreover, the space C8pU ; Eq of
"
smooth maps from U to E is a locally convex space with respect to the topology generated by the
semi-norms,
‰
u ÝÑ max
αďN
p
sup
xPK
Bα
x upxq
,
where N ranges over non-negative integers, K ranges over compact subsets of U and p ranges over
continuous semi-norms on E .
4.1. Pseudo-cones and HoldpΛq-families. In what follows, by a cone Θ Ă C we shall always
mean a cone with vertex at the origin, so that λ P Θ ñ tλ P Θ for all t ą 0.
Definition 4.1. A connected set Λ Ă C is called a pseudo-cone when there is a cone Θ Ă C and
a disk D about the origin such that ΛzD " ΘzD. The cone Θ is called the conical par t of Λ.
Remark 4.2. In the above definition the conical part Θ is unique.
Remark 4.3. If Λ is a pseudo-cone, then its interior IntpΛq and its closure Λ are pseudo-cones as
well.
Example 4.4. Any angular sector Θ " tφ ă arg λ ă φ1u is a cone, and hence is a pseudo-cone.
Chopping off a disk from Θ or glueing to it a disk or an annulus provides us with pseudo-cones
with conical component Θ.
Figure 1. Examples of Pseudo-Cones
Definition 4.5. Given pseudo-cones Λ and Λ1 we shall write Λ1 ĂĂ Λ to mean that Λ1 Ă IntpΛq.
Remark 4.6. If Θ and Θ1 are the respective conical parts of Λ and Λ1, then the relation Λ1 ĂĂ Λ
implies that Θ1 X S1 is a relatively compact subset of Θ X S1 (where S1 Ă C is the unit circle).
14
Remark 4.7. If Λ Ă C, then the relation Λ1 ĂĂ Λ implies there is an open disk D about the
Ť
origin such that Λ1 Ă CzD. Thus, there is r ą 0 such that λ ě r for all λ P Λ1.
Remark 4.8. Any open pseudo-cone Λ admits a pseudo-cone exhaustion Λ "
Λj are closed pseudo-cones such that Λj ĂĂ Λj`1.
jě0 Λj, where the
Throughout this section we let Λ be an open pseudo-cone.
Definition 4.9. HoldpΛq, d P R, consists of holomorphic functions f : Λ Ñ C such that, for every
pseudo-cone Λ1 ĂĂ Λ, there is CΛ1 ą 0 such that
fpλq ď CΛ1p1 ` λqd
@λ P Λ1.
We equip the vector space HoldpΛq, d P R, with the locally convex topology generated by the
semi-norms,
f ÝÑ sup
λPΛ1
p1 ` λq´dfpλq,
where Λ1 ranges over pseudo-cones Λ1 ĂĂ Λ. This turns HoldpΛq into a Fr´echet space. The space
Hol´8pΛq is also a Fr´echet space with respect to the topology generated by these semi-norms,
where Λ1 ranges over all pseudo-cones ĂĂ Λ and d ranges over R.
In what follows, given an open Ω Ă C and a locally convex space E , we shall say that a map
f : Ω Ñ E is holomorphic at a given point λ0 P Ω when
fpλq ´ fpλ0q
Bλfpλ0q :" lim
λÑλ0
λ ´ λ0
exists in E .
We say that f is holomorphic on Ω when it is holomorphic at every point of Ω.
Remark 4.10. Let f : Ω Ñ E be a holomorphic map. Then, for all λ P Ω, we have the Cauchy
formula,
(4.1)
fpλq " 1
2iπ
fpζqpζ ´ λq´1dζ,
ż
.
ż
Γ
"
‰
fpλq
Γ
ϕ
ϕ
where Γ is any direct-oriented Jordan curve contained in Ω whose interior contains λ. Indeed, for
all ϕ P E1, the composition ϕ f is a holomoprhic function on Ω, and so we have
fpζqpζ ´ λq´1dζ
pϕ fqpζqpζ ´ λq´1dζ " 1
2iπ
" 1
2iπ
As E 1 separates the points of E this gives (4.1). If we denote by D the interior of Γ, then the
integrand in the right-hand side of (4.1) lies in C8pDΓ; Eq, and so the integral actually converges
in C8pD; Eq (see, e.g., [22, Appendix C]). It then follows that f is a smooth map from Ω to E .
Moreover, by differentiating under the integral sign (see, e.g., [22, Appendix C]) we see that fpλq
satisfies the Cauchy-Riemann equation Bλf " 0. Conversely, any differentiable map satisfying this
equation is a holomorphic map.
Definition 4.11. Suppose that E is a locally convex space. Then HoldpΛ; Eq, d P R, consists
of holomorphic maps x : Λ Ñ E such that, for every continuous semi-norm p on E and every
pseudo-cone Λ1 ĂĂ Λ, there is CpΛ1 ą 0 such that
(4.2)
Remark 4.12. HoldpΛ; Eq, d P R, is a locally convex space with respect to the topology generated
by the semi-norms,
prxpλqs ď CpΛ1p1 ` λqd
@λ P Λ1.
xpλq ÝÑ sup
λPΛ1
p1 ` λq´dprxpλqs ,
(4.3)
where p ranges over continuous semi-norms on E and Λ1 ranges over pseudo-cones Λ1 ĂĂ Λ. In
addition, if E is a Fr´echet space, then HoldpΛ; Eq is a Fr´echet space as well.
Proposition 4.13. Suppose that E and F are locally convex spaces, and let d P R. Any continuous
linear map T : E Ñ F gives rise to a continuous linear map T : HoldpΛ; Eq Ñ HoldpΛ; Fq.
ż
Γ
15
Proof. Let xpλq P HoldpΛ; Eq, d P R. As T is a continuous linear map, the composition Trxpλqs
is a holomorphic map from Λ to F . Let q be a continuous semi-norm on F . Then q T is a
‰
continuous semi-norm on E , and so, for every pseudo-cone Λ1 ĂĂ Λ, we have
‰
`
"
"
xpλq
p1 ` λq´dq
xpλq
T
sup
λPΛ1
" sup
λPΛ1
p1 ` λq´dpq Tq
ă 8.
This shows that Trxpλqs is contained in HoldpΛ; Fq and depends continuously on xpλq. The proof
(cid:3)
is complete.
For bilinear bilinear maps we have the following statement.
Proposition 4.14. Suppose that Ei, i " 1, 2, and F are locally convex spaces, and let d1, d2 P R.
Any (jointly) continuous bilinear map Φ : E1 E2 Ñ F gives rise to a continuous bilinear map
Φ : Hold1pΛ; E1q Hold2pΛ; E2q Ñ Hold1`d2pΛ; Fq.
Proof. Let xjpλq P HoldjpΛ; Ejq, j " 1, 2. As x1pλq (resp., x2pλq) is a holomorphic family with
values in E1 (resp., E2) and the map Φ : E1 E2 Ñ F is continuous and bilinear, we see that
Φrx1pλq, x2pλqs, λ P Λ, is a holomorphic family in F . Moreover, the continuity of Φ means
that, given any continuous semi-norm q on F , there are a continuous semi-norm p1 on E1 and a
continuous semi-norm p2 on E2 such that
Φrx1, x2s
Φrx1pλq, x2pλqs
Thus, for every pseudo-cone Λ1 ĂĂ Λ, we have
ď sup
λPΛ1
ă 8.
p1 ` λq´d1p1rx1pλqs sup
λPΛ1
p1 ` λq´d2 p2rx2pλqs
p1 ` λq´pd1`d2qq
ď p1px1qp2px2q
@xi P Ei.
sup
λPΛ1
`
`
q
This shows that Φrx1pλq, x2pλqs is an Hold1`d2pΛq-family and depends continuously on x1pλq and
x2pλq. The proof is complete.
(cid:3)
In addition, it is also convenient to introduce the following class of maps.
Definition 4.15. Suppose that Ω is an open set of RN C for some N ě 1 and E is a locally
convex space. Then C8,ωpΩ; Eq consists of maps Ω Q pη, λq Ñ fpη; λq P E that are C8 with
respect to η and holomorphic with respect to λ.
Remark 4.16. By using the Cauchy formula (4.1) we can identify C8,ωpΩ; Eq with the (closed)
subspace of C8pΩ; Eq of solutions of the Cauchy-Riemann equation Bsλf " 0. In particular, when
Ω " V Λ, where V Ă RN is an open set, then under the natural identification C8pV Λ; Eq »
`
C8pΛ; C8pV ; Eqq we get a one-to-one correspondence,
Λ; C8pV ; Eq
.
C8,ωpΩ; Eq » Hol
››δαBβ
›› ď CΛ1αβp1 ` λqdp1 ` ξqm´β
4.2. Standard parametric symbols.
Definition 4.17. Sm,dpRn Λ; Aθq, m, d P R, consists of maps ρpξ; λq P C8,ωpRn Λ; Aθq such
that, for all pseudo-cones Λ1 ĂĂ Λ and multi-orders α and β, there is CΛ1αβ ą 0 such that
(4.4)
Example 4.18. Any symbol ρpξq P SmpRn; Aθq can be regarded as an element of Sm,0pRn Λ; Aθq
that does not depend on λ.
Remark 4.19. Sm,dpRnΛ; Aθq is a locally convex space with respect to the locally convex topology
generated by the semi-norms,
@pξ, λq P Rn Λ1.
ξ ρpξ; λq
(4.5)
where N ranges over non-negative integers and Λ1 ranges over pseudo-cones such that Λ1 ĂĂ Λ.
pξ,λqPRnΛ1
sup
p1 ` λq´dp1 ` ξqβ´m}δαBβ
ξ ρpξ; λq},
ρpξ; λq ÝÑ max
α`βďN
16
Remark 4.20. Let ρpξ; λq P Sm,dpRnΛ; Aθq. For every λ P Λ we get a symbol ρp; λq P SmpRn; Aθq,
and so we get a family in SmpRn; Aθq parametrized by Λ. This is actually an HoldpΛq-family. This
provides us with a one-to-one correspondence,
Sm,dpRn Λ; Aθq » HoldpΛ; SmpRn; Aθqq.
(4.6)
Under this correspondence the semi-norms (4.5) generate the topology of HoldpΛ; SmpRn; Aθqq,
and so we actually have a topological vector space isomorphism.
Definition 4.21. S´8,dpRn Λ; Aθq, d P R, consists of maps ρpξ; λq P C8,ωpRn Λ; Aθq such
that, given any N ě 0, for all pseudo-cones Λ1 ĂĂ Λ and multi-orders α and β, there is CΛ1N αβ ą 0
such that
›› ď CΛ1N αβp1 ` λqdp1 ` ξq´N
@pξ, λq P Rn Λ1.
››δαBβ
ξ ρpξ; λq
Remark 4.22. We have S´8,dpRn Λ; Aθq "
Remark 4.23. S´8,dpRn Λ; Aθq is a locally convex space with respect to the topology generated
by the semi-norms (4.5), where we let m ranges over all real numbers.
Ş
mPR Sm,dpRn Λ; Aθq.
Remark 4.24. The one-to-one correspondance (4.6) induces a topological vector space isomor-
phism,
(4.7)
S´8,dpRn Λ; Aθq » HoldpΛ; S pRn; Aθqq.
Combining the topological vector space identifications (4.6) and (4.7) with Proposition 4.13
yields the following result.
Proposition 4.25. Suppose that E is a locally convex space, and let d P R.
(1) Given m P R, any continuous linear map T : SmpRn; Aθq Ñ E gives rise to a continuous
(2) Any continuous linear map T : S pRn; Aθq Ñ E gives rise to a continuous linear map
linear map T : Sm,dpRn Λ; Aθq Ñ HoldpΛ; Eq.
T : S´8,dpRn Λ; Aθq Ñ HoldpΛ; Eq.
Likewise, by combining the identification (4.6) with Proposition 4.14 we arrive at the following
statement.
Proposition 4.26. Given m1, m2, m3 P R and d1, d2 P R, any (jointly) continuous bilinear
map Φ : Sm1pRn; Aθq Sm2pRn; Aθq Ñ Sm3pRn; Aθq gives rise to a continuous bilinear map
Φ : Sm1,d1pRn Λ; Aθq Sm2,d2pRn Λ; Aθq Ñ Sm3,d1`d2pRn Λ; Aθq.
4.3. Asymptotic expansions of parametric symbols. In what follows we let pmjqjě0 be a
(strictly) decreasing sequence of real numbers converging to ´8.
Definition 4.27. Given ρpξ; λq P C8,ωpRn Λ; Aθq and ρjpξ; λq P Smj ,dpRn Λ; Aθq, j ě 0, we
shall write ρpξ; λq „
ř
jě0 ρjpξ; λq when
ρpξ; λq ´
ÿ
jăN
(4.8)
ρjpξ; λq P SmN ,dpRn Λ; Aθq
for all N ě 0.
Remark 4.28. The condition (4.8) for N " 0 means that ρpξ; λq P Sm0,dpRn Λ; Aθq.
For our purpose, it will be convenient to have a qualitative version of the conditions (4.8).
ř
Lemma 4.29. Given ρpξ; λq P C8,ωpRn Λ; Aθq and ρjpξ; λq P Smj ,dpRn Λ; Aθq, j ě 0, the
following are equivalent:
(i) ρpξ; λq „
(ii) For all N ě 0, given any pseudo-cone Λ1 ĂĂ Λ and multi-orders α and β, as soon as J is
jě0 ρjpξ; λq in the sense of (4.8).
large enough there is a constant C ą 0 such that, for all pξ, λq P Rn Λ1, we have
›› ď Cp1 ` λqdp1 ` ξq´N .
(4.9)
`
››δαBβ
ξ
ÿ
ρ ´
ρj
jăJ
pξ; λq
17
`
jăJ
jăN
jăJ
NďjăJ
ÿ
ÿ
ρjpξ; λq.
ρpξ; λq ´
››ρpξ; λq ´
Proof. It is immediate that (i) implies (ii), so we only have to prove the converse. Suppose that
ÿ
(ii) holds. Let N ě 0. Then (ii) implies that, given any pseudo-cone Λ1 ĂĂ Λ, if J is large enough,
then there is C ą 0 such that
@pξ, λq P Rn Λ1.
ρjpξ; λq
(4.10)
ÿ
ÿ
If we take J large enough so as to have J ě N as well, then we have
›› ď Cp1 ` λqdp1 ` ξqmN
ρjpξ; λq " ρpξ; λq ´
Likewise, given any multi-orders α and β there is CN Λ1αβ ą 0 such that
ρjpξ; λq `
ř
If j ě N , then ρjpξ; λq satisfies the estimate (4.10), since it is contained in Smj ,dpRn Λ; Aθq Ă
jăN ρjpξ; λq satisfies the estimate (4.10).
SmN ,dpRn Λ; Aθq. Therefore, the difference ρpξ; λq ´
›› ď CN Λ1αβp1 ` λqdp1 ` ξqmN´β
››δαBβ
ř
jě0 ρjpξ; λq in the sense of (4.8). The proof is complete.
ř
pξ; λq
jăN ρjpξ; λq P SmN ,dpRn Λ; Aθq for all N ě 0. That is, we have
(cid:3)
ρj
This shows that ρpξ; λq ´
ρpξ; λq „
4.4. Homogeneous parametric symbols. Throughout the rest of this paper we let w be some
(cid:32)
fixed (strictly) positive real number. We also denote by Θ the conical component of Λ. Note this
is an open cone contained in C. In addition, given any c ě 0, we set
pξ, λq P pRnz0q C; λ P Θ or λ ă cξw
(4.11)
In particular, Ω0pΘq " pRnz0q Θ.
Definition 4.30. Sd
satisfy the following two properties:
mpΩcpΘq; Aθq, m, d P R, consists of maps ρpξ; λq P C8,ωpΩcpΘq; Aθq that
@pξ, λq P Rn Λ1.
(i) ρptξ; twλq " tmρpξ; λq for all pξ, λq P ΩcpΘq and t ą 0.
(ii) Given any cone Θ1 such that Θ1zt0u Ă Θ, for all compacts K Ă Rnz0 and multi-orders α,
ΩcpΘq "
(
.
ρ ´
ξ
jăN
β, there is CΘ1Kαβ ą 0 such that
`
1 ` λ
d
@pξ, λq P K Θ1.
}δαBβ
ξ ρpξ; λq} ď CΘ1Kαβ
(4.12)
Example 4.31. Any homogeneous symbol ρpξq P SmpRn; Aθq can be regarded as an element of
mpΩcpΘq; Aθq for every c ą 0.
S0
Remark 4.32. If ρpξ; λq P Sd
Remark 4.33. If ρ1pξ; λq P Sd1
contained in Sd1`d2
m´βpΩcpΘq; Aθq for all α, β P Nn
0 .
m2pΩcpΘq; Aθq, then ρ1pξ; λqρ2pξ; λq is
mpΩcpΘq; Aθq, then δαBβ
m1pΩcpΘq; Aθq and ρ2pξ; λq P Sd2
ξ ρpξ; λq P Sd
m1`m2pΩcpΘq; Aθq.
The next lemma shows how to cut off homogeneous parametric symbols in order to get standard
parametric symbols.
Lemma 4.34. Suppose that ρpξ; λq P Sd
χpξq " 1 for ξ ď pc´1Rq 1
mpΩcpΘq; Aθq, m, d P R. Let χpξq P C8
c pRnq be such that
w , where R ą 0 is such that Λ Ă Θ Y Dp0, Rq.
(i) If d ě 0, then p1 ´ χpξqqρpξ; λq P Sm,dpRn Λ; Aθq.
(ii) If d ă 0, then
(4.13)
p1 ´ χpξqqρpξ; λq P
Sm`wd1,d1pRn Λ; Aθq.
Proof. As the relation Λ1 ĂĂ Λ implies that Λ1 ĂĂ Θ Y Dp0, Rq, the restriction to Rn Λ of any
element in Sm,dpRn pΘ Y Dp0, Rqq; Aθq is contained in Sm,dpRn Λ; Aθq. Therefore, without
any loss of generality we may assume that Λ " Θ Y Dp0, Rq.
wu and U " RnzB " tξ P Rn; cξw ě Ru. As Λ " Θ Y Dp0, Rq,
we have U Λ " pU Θq Y pU Dp0, Rqq. It is immediate that U Θ Ă ΩcpΘq. Moreover,
Set B " tξ P Rn; ξ ă pc´1Rq 1
č
dďd1ď0
18
if pξ, λq P U Dp0, Rq, then cξw ě R ą λ, and hence pξ, λq P ΩcpΘq. Therefore, we see that
`
U Λ Ă ΩcpΘq, and hence we have
`
`
Rn Λ "
U Λ
Y
B Λ
Ă ΩcpΘq Y
B Λ
.
As ρpξ; λq P C8,ωpΩcpΘq; Aθq and χpξq " 1 on B, if we extend p1´χpξqqρpξ; λq to be zero on BΛ,
then we get a C8,ω-map on ΩcpΘqYpB Λq. By restriction this gives a map in C8,ωpRn Λ; Aθq.
It then remains to show that p1 ´ χpξqqρpξ; λq satisfies the estimates (4.4) on Rn Λ.
In what follows we set d´ " maxp0,´dq.
.
"
(4.16)
;ξw
`
`
›› ď C1p1 ` µqd
`
}δαBβ
Claim. Given any pseudo-cone Λ1 ĂĂ Λ and multi-orders α, β P Nn
(4.14)
Proof of the Claim. Without any loss of generality we may assume that Λ1 " Θ1Y Dp0, R1q, where
R1 ă R and Θ1 Ă C is a cone such that Θ1zt0u Ă Θ. Moreover, thanks to the homogeneity of
δαBβ
ξ ρpξ; λq} ď CΛ1αβp1 ` λqdξm`wd´´β
‰
0 , there is CΛ1αβ ą 0 such that
@pξ, λq P U Λ1.
ξ ρpξ; λq, for all pξ, λq P ΩcpΘq, we have
ξ´1ξ
ξ´1ξ;ξ´wλ
ξ ρpξ; λq " δαBβ
ξ ρ
" ξm´βδαBβ
ξ ρ
ξ´wλ
δαBβ
(4.15)
ξ
`
ξ ρpη; µq
››δαBβ
@pη, µq P Sn´1 Θ1.
we know by (4.12) there is C1 ą 0 such that
Set D1 " Dp0, R1q, so that Λ1 " Θ1 Y D1. As Sn´1 is a compact subset of Rnz0 and Θ1zt0u Ă Θ,
››δαBβ
If pξ, λq P U Θ1, then pξ´1ξ,ξ´wλq P Sn´1 Θ1. Therefore, by using (4.15) we see that, for
›› " ξm´β››δαBβ
all pξ, λq P U Θ1, we have
ξ ρpξ; λq
(cid:32)
`
`
pξ, λq P Sn´1 C; λ P Θ or λ ă cξw
Sn´1 Θ
If pξ, λq P U D1, then cξw ě R " RpR1q´1R1 ě RpR1q´1λ, and so ξ´wλ ď cR1R´1.
Thus, if we set K " Sn´1 Dp0, cR1R´1q, then K is compact and pξ´1ξ,ξ´wλq P K for all
pξ, λq P U D1. Note also that
›› ď C1ξm´β`
Sn´1 Dp0, cq
ξ´1ξ;ξ´wλ
1 ` ξ´wλ
Sn´1 C
ΩcpΘq X
"
"
Therefore, we see that K is a compact subset of ΩcpΘq. As ρpξ, λq P C8,ωpΩcpΘq; Aθq we deduce
there is C2 ą 0 such that, for all pη, µq P K, we have
(4.17)
2 " C2 suptp1`µq´d; µ P Dp0, cR1R´1qu. As mentioned above, if pξ, λq is in
where we have set C1
UD1, then pξ´1ξ,ξ´wλq P K. Therefore, by using (4.17) we deduce that, for all pξ, λq P UD1,
we have
››δαBβ
2p1 ` µqd,
ξ ρpη; µq
`
›› ď C2 ď C1
`
1 ` ξ´wλ
Note that if ξ P U , then cξw ě R, and so ξ´w ď cR´1 and 1 ě c´1Rξ´w. Thus,
ξ´1ξ;ξ´wλ
ξ ρpξ; λq
›› " ξm´β››δαBβ
2ξm´β`
›› ď C1
››δαBβ
(4.18)
(
Y
ξ ρ
ξ ρ
d
.
.
1 ` ξ´wλ ď 1 ` cR´1λ ď c1p1 ` λq,
d
.
(cid:3)
1 ` ξ´wλ ě c´1Rξ´w ` ξ´wλ ě c´1
`
where we have set c1 " maxp1, cR´1q. It then follows that
d
1 p1 ` λqdξwd´
d ď c
Combining this with (4.16) and (4.18) proves the claim.
Given α, β P Nn
1 ` ξ´wλ
ÿ
1 ξ´wp1 ` λq,
@pξ, λq P U Λ1.
0 , let Λ1 be a pseudo-cone such that Λ1 ĂĂ Λ. By Leibniz's rule we have
δαBβ
ξ rp1 ´ χpξqqρpξ; λqs "
ξ p1 ´ χpξqqδαBβ2
Bβ1
ξ ρpξ; λq,
(4.19)
β
β1
19
›››δαBβ
ξ rp1 ´ χpξqqρpξ; λqs
Suppose that d ă 0.
››› ď CΛ1αβp1 ` λqdp1 ` ξqm`wd´´β @pξ, λq P Rn Λ1.
where the sum ranges over all multi-orders β1 and β2 such that β1 ` β2 " β. Note that each
Bβ1
ξ p1 ´ χpξqq is uniformly bounded on Rn and vanishes on B Λ1. Combining this with (4.14)
allows us to show there is CΛ1αβ ą 0 such that
(4.20)
In particular, p1 ´ χpξqqρpξ; λq is
This shows that p1 ´ χpξqqρpξ; λq P Sm`wd´,dpRn Λ; Aθq.
contained in Sm,dpRn Λ; Aθq when d ě 0.
In this case d´ " ´d " d, and so p1 ´ χpξqqρpξ; λq is contained
in Sm`wd,dpRn Λ; Aθq. Note that if d1 P rd, 0s, then ρpξ; λq is contained in Sd1
q pΩcpΘq; Aθq,
qpΩcpΘq; Aθq. Therefore, the above arguments also show that
since Sd1
p1 ´ χpξqqρpξ; λq is in Sm`wd1,d1pRn Λ; Aθq for any d1 P rd, 0s. This gives (4.13). The proof is
(cid:3)
complete.
4.5. Classical parametric symbols. In what follows, given any r ą 0, we denote by Bprq
the ball of radius r about the origin in Rn. We observe that if c ą 0 and R ą 0 is such that
Λ Ă Θ Y Dp0, Rq, then
q pΩcpΘq; Aθq contains Sd
ΩcpΘq Ą
Indeed, pRnz0q Θ Ą pRnzBprqq Θ. Moreover, if λ ă R and ξ ě r ą pc´1Rq 1
"`
λ ă R ă cξw. Thus,
ΩcpΘq Ą
RnzBprq
RnzBprq
Λ.
Θ
Λ
w , then
`
RnzBprq
‰ď"`
Definition 4.35. Sm,dpRn Λ; Aθq, m, d P R, consists of maps ρpξ; λq P C8,dpRn Λ; Aθq for
which there are c ą 0 and ρm´jpξ; λq P Sd
m´jpΩcpΘq; Aθq, j ě 0, such that
‰
Dp0, Rq
for all r ą pc´1Rq 1
w .
`
RnzBprq
ÿ
Ą
ρpξ; λq „
ρm´jpξ; λq,
jě0
`
RnzBprq
`
ξ
jăJ
ρ ´
ř
pξ; λq
in the sense that, for all N ě 0 and multi-orders α, β, given any pseudo-cone Λ1 ĂĂ Λ and r ą 0
ÿ
Λ1, as soon as J is large enough there is CΛ1N Jrαβ ą 0 such that,
such that ΩcpΘq Ą
for all pξ, λq P pRnzBprqq Λ1, we have
ρm´j
(4.21)
››δαBβ
›› ď CΛ1N Jrαβξ´Np1 ` λqd.
ř
ρm´jpξq. As mentioned in Example 4.31 each
Example 4.36. Let ρpξq P SmpRn; Aθq, ρpξq „
m´jpΩcpΘq; Aθq for any c ą 0.
homogeneous symbol ρm´jpξq can be regarded as an element of S0
ρm´jpξq in the sense of (3.3) implies
Note also that having an asymptotic expansion ρpξq „
that we also have an asymptotic expansion in the sense of (4.21) with d " 0. It then follows that
ř
ρpξq P Sm,0pRn Λ; Aθq.
Remark 4.37. Let ρpξ; λq P Sm,dpRnΛ; Aθq. Thus, ρpξ; λq „
jě0 ρm´jpξ; λq, where ρm´jpξ; λq is
m´jpΩcpΘq; Aθq. Given any multi-orders α, β, we also have δαBβ
ξ ρm´jpξ; λq
in Sd
in the sense of (4.21), and hence δαBβ
ξ ρpξ; λq P Sm´β,dpRn Λ; Aθq.
c pRnq
Remark 4.38. Suppose we are given ρm´jpξ; λq P Sd
be as in Lemma 4.34. Recall that by Lemma 4.34 p1 ´ χpξqqρm´jpξ; λq P Sm´j`wd´,dpΩcpΘq; Aθq,
where d´ " maxp0,´dq. Then, the following are equivalent:
m´jpΩcpΘq; Aθq, j ě 0. Let χpξq P C8
ř
ř
jě0 ρm´jpξ; λq in the sense of (4.21).
jě0p1 ´ χpξqqρm´jpξ; λq in the sense of (4.8) or (4.9).
ř
jě0 δαBβ
(i) ρpξ; λq „
(ii) ρpξ; λq „
ξ ρpξ; λq „
Note that here (4.8) means that, for all N ě 0, as soon as J ě N ` wd´, we have
(4.22)
p1 ´ χpξqqρm´jpξ; λq P Sm´N,dpRn Λ; Aθq.
ρpξ; λq ´
ÿ
jăJ
This provides us with a quantitative version of the estimates (4.21).
20
4.6. Borel lemma for parametric symbols. In the following, given any m P R, we denote by
SmpRnq the space of scalar standard symbols on Rn. It consists of functions σpξq P C8pRnq such
that, for every multi-order α, there is Cα ą 0 such that
We equip SmpRnq with the locally convex topology generated by the semi-norms,
Bα
ξ σpξq ď Cαp1 ` ξqm´α
for all ξ P Rn.
σpξq ÝÑ sup
αďN
sup
ξPRn
p1 ` ξq´m`αBα
ξ σpξq,
N ě 0.
Combining the above remark with Lemma 4.34 we also get the following result.
Proposition 4.39. Let m, d P R.
(1) If d ě 0, then Sm,dpRn Λ; Aθq Ă Sm,dpRn Λ; Aθq.
(2) If d ă 0, then
Sm,dpRn Λ; Aθq Ă
Sm`wd1,d1pRn Λ; Aθq.
č
dďd1ď0
Lemma 4.40 ([26, Prop. 18.1.2]). Let χ P S pRnq be such that χp0q " 1. Then the family
pχpξqqą0 converges to 1 in SmpRnq as Ñ 0` for every m ą 0.
We have the following density result in Sm,dpRn Λ; Aθq.
Lemma 4.41. Let χpξq P S pRnq be such that χp0q " 1. Given any ρpξ; λq P Sm,dpRn Λ; Aθq,
m, d P R, as Ñ 0`, the family χpξqρpξ; λq converges to ρpξ; λq in Sm1,dpRn Λ; Aθq for all
m1 ą m.
Proof. Let m1 ą m. The product of Aθ gives rise to a continuous bilinear maps from Sm1pRn; Aθq
Sm2pRn; Aθq to Sm1`m2pRn; Aθq for all m1, m2 P R. Proposition 4.26 then ensures us that we get
a continuous bilinear map,
Sm1,d1pRn Λ; Aθq Sm2,d2pRn Λ; Aθq ÝÑ Sm1`m2,d1`d2pRn Λ; Aθq.
(4.23)
In particular, by using the continuity of the inclusion of Sm1´mpRnq into Sm1´m,0pRn Λ; Aθq we
obtain a continuous bilinear map,
Sm1´mpRnq Sm,dpRn Λ; Aθq ÝÑ Sm1,dpRn Λ; Aθq.
Combining Lemma 4.40 with the continuity of the above bilinear map (4.23) then gives the result.
(cid:3)
The proof is complete.
ř
The following is the version of Borel's lemma for standard parametric symbols.
sup
pξ,λqPRnΛj
jě0 ρjpξ; λq.
››δαBβ
pmq
j
pmq
j ď p
p1 ` λq´dp1 ` ξqβ´m
Ť
jě0 Λj be a pseudo-cone exhaustion of Λ with Λj ĂĂ Λj`1 (cf. Remark 4.8).
Lemma 4.42. For j " 0, 1, . . ., let ρjpξ; λq P Smj ,dpRn Λ; Aθq. Then there exists ρpξ; λq in
Sm0,dpRn Λ; Aθq such that ρpξ; λq „
Proof. Let Λ "
Given any m P R the topology of the Sm,dpRn Λ; Aθq is generated by the semi-norms,
j ě 0.
pρq " max
α`βďj
pmq
j`1. In addition, we let χpξq P S pRnq be such that χp0q " 1. For ą 0 we set
Note that p
χpξq " χpξq, ξ P Rn.
‰
We know by Lemma 4.41 that, for j " 1, 2, . . ., the family χpξqρjpξ; λq converges to ρjpξ; λq
in Smj´1,dpRn Λ; Aθq as Ñ 0` (since mj´1 ą mj). Thus, we can find j ą 0 such that
pmj´1q
p
j
"
ď 2´j. We also set 0 " 1. Given any N ě 0, for any (cid:96) ě 0, we have
pmNq
(cid:96)
ÿ
p1 ´ χjqρj
p
ř
p1 ´ χjqρj
jěNp1´χjpξqqρjpξ; λq converges normally with respect to each semi-
, (cid:96) ě N . As SmN ,dpRnΛ; Aθq is a Fr´echet space whose topology is generated by these
p1 ´ χjqρj
This implies that the series
2´j ă 8.
‰
ď
ξ ρpξ; λq
ÿ
‰
ď
jě(cid:96)`1
pmj´1q
p
j
ÿ
jě(cid:96)`1
pmNq
norm p
(cid:96)
››,
p
"
jě(cid:96)`1
"
21
semi-norms, we deduce that the series
for every N ě 0. Set
ř
ÿ
`
jěNp1 ´ χjpξqqρjpξ; λq converges in SmN ,dpRn Λ; Aθq
ρpξ; λq "
ÿ
1 ´ χjpξq
jě0
ρjpξ; λq.
ÿ
`
Then ρpξ; λq P Sm0,dpRn Λ; Aθq. Moreover, for all N ě 1, we have
ÿ
jăN
ρjpξ; λq "
χjpξqρjpξ; λq `
ř
ρpξ; λq ´
jăN ρjpξ; λq is in SmN ,dpRn Λ; Aθq for all N ě 0. That is, ρpξ; λq „
ř
jěNp1´χjpξqqρjpξ; λq converges in SmN ,dpRnΛ; Aθq and each map χjpξqρjpξ; λq
Here the series
is contained in S´8,dpRnΛ; Aθq Ă SmN ,dpRnΛ; Aθq. Therefore, we see that the remainder term
jě0 ρjpξ; λq.
ρpξ; λq ´
(cid:3)
The proof is complete.
1 ´ χjpξq
ρjpξ; λq.
ř
jăN
jěN
We are now in a position to get a version of Borel's lemma for classical parametric symbols.
Proposition 4.43. Given m P R and c ą 0, let ρm´jpξ; λq P Sd
there is ρpξ; λq P Sm,dpRn Λ; Aθq such that ρpξ; λq „
ř
Proof. Let χpξq P C8
c pRnq be as in Lemma 4.34. We know by Lemma 4.34 that p1´χpξqqρm´jpξ; λq
ř
lies in Sm`wd´´j,dpΩcpΘq; Aθq. Therefore, by Lemma 4.42 there is ρpξ; λq in Sm`wd´,dpRnΛ; Aθq
such that ρpξ; λq „
jě0p1 ´ χpξqqρm´jpξ; λq in the sense of (4.8). By Remark 4.38 this implies
jě0 ρm´jpξ; λq in the sense of (4.21). In particular, we see that ρpξ; λq P Sm,dpRn
that ρpξ; λq „
Λ; Aθq. The proof is complete.
(cid:3)
m´jpΩcpΘq; Aθq, j ě 0. Then
ř
jě0 ρm´jpξ; λq.
5. Parametric Pseudodifferential Operators
In this section, we introduce our classes of ΨDOs with parameter and derive some of their
properties.
5.1. Classes of ΨDOs with parameter. Let ρpξ; λq P Sm,dpRn Λ; Aθq, m, d P R. Given any
λ P Λ, we get a symbol in SmpRn; Aθq. We denote by Pρpλq the ΨDO associated with this symbol.
That is, Pρpλq is the continuous operator on Aθ defined by
eisξρpξ; λqα´spuqds ¯dξ,
Pρpλqu "
u P Aθ,
ij
In what follows we equip L pAθq with its strong topology (i.e., the topology of uniform conver-
θq with its strong topology. By [22,
where the above integral is meant as an oscillating integral (see Section 3).
gence on bounded subsets of Aθ). We similarly equip L pA 1
Proposition 5.4] and [23, Proposition 8.6] we have continuous linear maps,
SmpRn; Aθq Q ρpξq Ñ Pρ P L pAθq,
SmpRn; Aθq Q ρpξq Ñ Pρ P L pA 1
θq.
Combining this Proposition 4.25 gives the following result.
Proposition 5.1. For any ρpξ; λq P Sm,dpRn Λ; Aθq, m, d P R, the family Pρpλq is contained in
HoldpΛ; L pAθqq and uniquely extends to a family Pρpλq P HoldpΛ; L pA 1
Definition 5.2. Ψm,dpAθ; Λq, m, d P R, consists of families of operators Pρpλq : Aθ Ñ Aθ with
ρpξ; λq in Sm,dpRn Λ; Aθq.
Example 5.3. It follows from Example 4.36 that any operator P P ΨmpAθq can be regarded as an
element of Ψm,0pAθ; Λq. Combining this with the obvious inclusion Ψm,0pAθ; Λq Ă Ψm,1pAθ; Λq
allows us to regard P ´ λ as an element of Ψm,1pAθ; Λq.
θqq.
We also define ΨDOs with parameter of order ´8 as follows.
Definition 5.4. Ψ´8,dpAθ; Λq, d P R, consists of families of operators Pρpλq : Aθ Ñ Aθ with
ρpξ; λq in S´8,dpRn Λ; Aθq.
Remark 5.5. We have Ψ´8,dpAθ; Λq Ă
(see Corollary 5.32).
Ş
mPR Ψm,dpAθ; Λq. This inclusion is actually an equality
22
5.2. Composition of ΨDOs with parameter. Let ρ1pξ; λq P Sm1,d1pRn Λ; Aθq and ρ2pξ; λq P
Sm2,d2pRn Λ; Aθq. For each λ P Λ, we denote by ρ17ρ2pξ; λq the symbol given by (3.7). By
Proposition 3.19 the composition Pρ1pλqPρ2pλq is the ΨDO with symbol ρ17ρ2pξ; λq.
Proposition 5.6. Let ρ1pξ; λq P Sm1,d1pRnΛ; Aθq, m1, d1 P R, and ρ2pξ; λq P Sm2,d2pRnΛ; Aθq,
m2, d2 P R. Then ρ17ρ2pξ; λq P Sm1`m2,d1`d2pRn Λ; Aθq, and in the sense of (4.8) we have
ÿ 1
α!
ρ17ρ2pξ; λq „
ÿ
1
α!
αăN
Bα
ξ ρ1pξ; λqδαρ2pξ; λq.
Proof. For N ě 0, let 7N : Sm1pRn; Aθq Sm2pRn; Aθq Ñ Sm1`m2´NpRn; Aθq be the bilinear map
defined by
ρ17N ρ2pξq " ρ17ρ2pξq ´
Bα
ξ ρ1pξqδαρ2pξq,
ρjpξq P SmjpRn; Aθq.
By [23, Proposition 7.10] this is a continuous bilinear map. Therefore, by using Proposition 4.26
we see that if ρ1pξ; λq P Sm1,d1pRn Λ; Aθq and ρ2pξ; λq P Sm2,d2pRn Λ; Aθq, then ρ17N ρ2pξ; λq
is in Sm1`m2´N,d1`d2pRn Λ; Aθq. Thus,
ρ17ρ2pξ; λq ´
1
α!
ξ ρ1pξ; λqδαρ2pξ; λq P Sm1`m2´N,d1`d2pRn Λ; Aθq
Bα
ř
@N ě 0.
This means that ρ17ρ2pξ; λq „
that ρ17ρ2pξ; λq P Sm1`m2,d1`d2pRn Λ; Aθq. The proof is complete.
ξ ρ1pξ; λqδαρ2pξ; λq in the sense of (4.8). In particular, we see
(cid:3)
α!Bα
1
ÿ
αăN
In order to deal with the composition of ΨDOs associated with classical parametric symbols we
ř
ř
need the following two lemmas.
Lemma 5.7. Suppose that ρpξ; λq „
(cid:96)ě0 ρp(cid:96)qpξ; λq in the sense of (4.8), where ρp(cid:96)qpξ; λq is in
p(cid:96)q
Sm´(cid:96),dpRn Λ; Aθq and ρp(cid:96)qpξ; λq „
m´(cid:96)´kpξ; λq with ρ
m´(cid:96)´kpΩcpΘq; Aθq.
ÿ
kě0 ρ
Then ρpξ; λq P Sm,dpRn Λ; Aθq, and we have
p(cid:96)q
m´(cid:96)´kpξ; λq P Sd
ÿ
(5.1)
ρpξ; λq „
ρm´jpξ; λq, where ρm´jpξ; λq :"
p(cid:96)q
m´jpξ; λq P Sd
ρ
m´jpΩcpΘq; Λq.
jě0
0ď(cid:96)ďj
Proof. Let N ě 0 and J ě N ` wd´. By assumption,
ρp(cid:96)qpξ; λq P Sm´N,dpRn Λ; Aθq.
(cid:96)ăJ
ÿ
1 ´ χpξq
1 ´ χpξq
ÿ
`
ρpξ; λq ´
ÿ
`
ÿ
`
ÿ
ÿ
(cid:96)ăJ
`
1 ´ χpξq
(cid:96)ďjăJ
1 ´ χpξq
ř
jăJ
Let χpξq P C8
kăJ´(cid:96)
c pRnq be as in Lemma 4.34. In view of (4.22) for (cid:96) ă J we have
ρp(cid:96)qpξ; λq "
"
p(cid:96)q
m´(cid:96)´kpξ; λq mod Spm´(cid:96)q´pN´(cid:96)q,dpRn Λ; Aθq
ρ
p(cid:96)q
m´jpξ; λq mod Sm´N,dpRn Λ; Aθq.
ρ
(cid:96)ďjăJ
p(cid:96)q
m´jpξ; λq mod Sm´N,dpRn Λ; Aθq
ρ
Thus,
ρpξ; λq "
"
ρm´jpξ; λq mod Sm´N,dpRn Λ; Aθq,
ř
p(cid:96)q
m´jpξ; λq. Here ρm´jpξ; λq P Sd
where we have set ρm´jpξ; λq "
j
(cid:96)"0 ρ
using Remark 4.38 we see that ρpξ; λq „
shows that ρpξ; λq P Sm,dpRn Λ; Aθq. The result is thus proved.
m´jpΩcpΘq; Λq, and so by
jě0 ρm´jpξ; λq in the sense of (4.21). In particular, this
(cid:3)
23
ÿ
(5.2)
(5.3)
ř
Lemma 5.8. Let ρ1pξ; λq P Sm1,d1pRn Λ; Aθq and ρ2pξ; λq P Sm2,d2pRn Λ; Aθq, mi, di P R,
jě0 ρ2,m2´jpξ; λq, with ρi,mi´jpξ; λq in
be such that ρ1pξ; λq „
mi´jpΩcpΘq; Aθq, i " 1, 2. Then ρ1pξ; λqρ2pξ; λq P Sm1`m2,d1`d2pRn Λ; Aθq, and we have
Sdi
jě0 ρ1,m1´jpξ; λq and ρ2pξ; λq „
ř
ÿ
ρ1pξ; λqρ2pξ; λq „
pρ1ρ2qm1`m2´jpξ; λq, where
jě0
ρ1,m1´kpξ; λqρ2,m2´lpξ; λq P Sd1`d2
k`l"j
pρ1ρ2qm1`m2´jpξ; λq :"
Proof. For i " 1, 2 let ρpiqpξ; λq P Smi,dipRn Λ; Aθq be such that ρpiqpξ; λq „
with ρ
have
m1`m2´jpΩcpΘq; Aθq.
piq
mi´jpξ; λq
piq
mi´jpξ; λq P Smi´j,dipRn Λ; Aθq. Then (4.23) implies that, for all N ě 0 and l ă N , we
ρp1qpξ; λqρp2qpξ; λq "
p2q
m2´lpξ; λq "
ρp1qpξ; λqρ
p2q
p1q
m2´lpξ; λq mod Sm1`m2´N,d1`d2pRn Λ; Aθq.
m1´kpξ; λqρ
p2q
m2´lpξ; λq mod Sm1`m2´N,d1`d2pRn Λ; Aθq,
ρp1qpξ; λqρ
ř
jě0 ρ
lăN
ρ
ÿ
ÿ
ÿ
kăN´l
Thus,
ρp1qpξ; λqρp2qpξ; λq "
k`lăN
p1q
p2q
m2´lpξ; λq mod Sm1`m2´N,d1`d2pRn Λ; Aθq.
m1´kpξ; λqρ
ρ
ÿ
ÿ
This shows that, in the sense of (4.8), we have
(5.4)
ρp1qpξ; λqρp2qpξ; λq „
p2q
m2´lpξ; λq.
ρ
p1q
m1´kpξ; λqρ
ř
jě0
k`l"j
ř
jě0p1 ´ χpξqqρ1,m1´jpξ; λq and ρ2pξ; λq „
ρ1pξ; λq „
of (4.8). Therefore, in view of (5.4), in the sense of (4.8) once again, we have
w , i.e., rχpξq is as in Lemma 4.34. We know by Remark 4.38 that
c pRnq and
jě0p1 ´ χpξqqρ2,m2´jpξ; λq in the sense
R ą 0 is such that Λ Ă Θ Y Dp0, Rq. Set rχpξq " 1 ´ p1 ´ χpξqq2. Then rχpξq P C8
Bearing this in mind, let χpξq P C8
c pRnq be such that χpξq " 1 for ξ ď pc´1Rq 1
rχpξq " 1 for ξ ď pc´1Rq 1
ÿ
ÿ
ÿ
p1 ´ χpξqq2ρ1,m1´kpξ; λqρ2,m2´lpξ; λq
1 ´rχpξq
ρ1pξ; λqρ2pξ; λq „
„
pρ1ρ2qm1`m2´jpξ; λq.
w , where
k`l"j
jě0
`
ř
rχpξq is as in Lemma 4.34,
it follows from Remark 4.38 that in the sense of (4.21) we have
where pρ1ρ2qm`m2´jpξ; λq is given by (5.3). As pρ1ρ2qm`m2´jpξ; λq P Sd1`d2
ρ1pξ; λqρ2pξ; λq „
in Sm1`m2,d1`d2pRn Λ; Aθq. The proof is complete.
m1`m2´jpΩcpΘq; Aθq and
jě0pρ1ρ2qm1`m2´jpξ; λq. This immediately implies that ρ1pξ; λqρ2pξ; λq is
(cid:3)
ř
ř
Proposition 5.9. Let P1pλq P Ψm1,d1pAθ; Λq have symbol ρ1pξ; λq „
ρ1,m1´jpξ; λq, and let
P2pλq P Ψm2,d2pAθ; Λq have symbol ρ2pξ; λq „
pρ17ρ2qm1`m2´jpξ; λq, where
(1) ρ17ρ2pξ; λq P Sm1`m2,d1`d2pRn Λ; Aθq with ρ17ρ2pξ; λq „
We are now in a position to prove the following result.
ρ2,m2´jpξ; λq.
ř
jě0
(5.5)
pρ17ρ2qm1`m2´jpξ; λq "
Bα
ξ ρ1,m1´kpξ; λqδαρ2,m2´lpξ; λq,
(2) The composition P1pλqP2pλq " Pρ17ρ2pλq is in Ψm1`m2,d1`d2pAθ; Λq.
k`l`α"j
1
α!
j ě 0.
Proof. As the 2nd part is an immediate consequence of the first part and Proposition 3.20, we
only have to prove the first part. Given any α P Nn
0 , it follows from Remark 4.37 and Lemma 5.8
that, in the sense of (4.21), we have
Bα
ξ ρ1pξ; λqδαρ2pξ; λq „
Bα
ξ ρ1,m1´kpξ; λqδαρ2,m2´lpξ; λq,
ÿ
24
ÿ
pρ17ρ2qm1`m2´jpξ; λq,
where Bα
of α, k and l. Combining this with Lemma 5.7 and Proposition 5.6 then shows that, in the sense
of (4.21), we have
ξ ρ1,m1´kpξ; λqδαρ2,m2´lpξ; λq P Sd1`d2
m1`m2´α´k´lpΩcpΘq; Aθq for some c ą 0 independent
ÿ
ρ17ρ2pξ; λq „
m1`m2´jpΩcpΘq; Aθq is given by (5.5). In particular, we see that
where pρ17ρ2qm1`m2´jpξ; λq P Sd1`d2
ρ17ρ2pξ; λq is contained in Sm1`m2,d1`d2pRn Λ; Aθq. The proof is complete.
(cid:3)
5.3. Sobolev space mapping properties. Given any m, s P R, by [23, Proposition 10.4] we have
a continuous linear map SmpRn; Aθq Q ρpξq ÝÑ Pρ P L pH ps`mq
q. By combining this with
Proposition 4.25 and the version of Sobolev's embedding theorem provided by Proposition 3.26
we obtain the following result.
Proposition 5.10. Let ρpξ; λq P Sm,dpRn Λ; Aθq, m, d P R.
(1) The family Pρpλq is contained in HoldpΛ; L pH ps`mq
(2) If m ď 0, then Pρpλq P HoldpΛ; L pHθqq.
In the case of classical ΨDOs with parameter we get the following result.
qq for every s P R.
, H psq
, H psq
θ
θ
θ
θ
Proposition 5.11. Let Ppλq P Ψm,dpAθ; Λq, m, d P R.
, H psq
`
(1) If d ě 0, then Ppλq P HoldpΛ; L pH ps`mq
(2) If d ă 0, then, for every s P R, we have
Λ; L
Hold1`
Ppλq P
č
θ
θ
qq for every s P R.
H ps`m`wd1q
θ
, H psq
θ
.
dďd1ď0
(3) If m ď 0, then Ppλq P Hol
¯dpΛ; L pHθqq with ¯d :" maxpd, mw´1q.
Proof. The first two parts follow by combining the first part of Proposition 5.10 with Proposi-
tion 4.39. It remains to prove the last part.
Suppose that m ď 0.
If d ě 0, then by using the last two parts of Proposition 5.10 and
Proposition 4.39 we see that Ppλq is in HoldpΛ; L pHθqq and is even in HoldpΛ; K q when m ă 0.
Assume now that d ă 0. In the same way as above, given any d1 P rd, 0s, we see that Ppλq is in
Hold1pΛ; L pHθqq when m ` wd1 ď 0 and is in Hold1pΛ; K q when m ` wd1 ă 0. In particular,
we see that Ppλq is in Hol
(cid:3)
5.4. Schatten class properties. We refer to §§3.11 for the definitions of the Schatten classes
L p and weak Schatten classes L pp,8q on Hθ with p ě 1. Recall they are Banach ideals of L pHθq.
Lemma 5.12 ([23, Propositions 13.8 & 13.13, Corollary 13.11]). The following holds.
¯dpΛ; L pHθqq with ¯d :" maxpd, mw´1q. The proof is complete.
SmpRn; Aθq to L pp,8q and L q, q ą p.
(1) If ´n ď m ă 0 and we set p " nm´1, then ρ Ñ Pρ induces continuous linear maps from
(2) If m ă ´n, then ρ Ñ Pρ induces a continuous linear map from SmpRn; Aθq to L 1.
Combining this with Proposition 4.25 gives the following result.
Proposition 5.13. Let ρpξ; λq P Sm,dpRn Λ; Aθq with d P R and m ă 0. Set p " nm´1.
(1) If ´n ď m ă 0, then Pρpλq is contained in HoldpΛ; L pp,8qq and HoldpΛ; L qq for all q ą p.
(2) If m ă ´n, then Pρpλq P HoldpΛ; L 1q.
Let us now specialize the above result to classical ΨDOs with parameter.
Proposition 5.14. Let Ppλq P Ψm,dpAθ; Λq with d P R and ´n ď m ă 0. Set p " nm´1. In
addition, for q ě p set dpqq " pm ` nq´1qw´1.
č
(1) If d ě 0, then Ppλq P HoldpΛ; L pp,8qq.
(2) If mw´1 ă d ă 0, then
Holdpqq`
Λ; L pq,8q
,
where ¯p :" npm ` wdq´1.
Ppλq P
(5.6)
pďqď ¯p
25
(3) If d ď mw´1, then
(5.7)
Holdpqq`
Λ; L pq,8q
.
č
qěp
Ppλq P
`
Proof. Let ρpξ; λq P Sm,dpRn Λ; Aθq be such that Ppλq " Pρpλq. Suppose that d ě 0. Proposi-
tion 4.39 ensures us that ρpξ; λq P Sm,dpRn Λ; Aθq. As ´n ď m ă 0, it then follows from the 1st
part of Proposition 5.13 that Ppλq P HoldpΛ; L pp,8qq.
Suppose that d ă 0, and let q ě p. As dpqq " pm ` nq´1qw´1 and p " ´nm´1 we have
mw´1 ă dpqq ď
m ` np´1
w´1 " 0.
(5.8)
Thus, if d ď mw´1, then dpqq P rd, 0s for all q ě p. If d ą mw´1, then d ď dpqq if and only if
q ď ¯p, with ¯p :" npm ` wdq´1.
From now on we assume that either d ď mw´1 and q ě p, or d ą mw´1 and p ď q ď ¯p.
This ensures that dpqq P rd, 0s, and so by Proposition 4.39 this implies that ρpξ; λq is contained in
Sm´wdpqq,dpqqpRn Λ; Aθq. Note also that as m ě ´n, we have m ´ wdpqq ě m ě ´n. Moreover,
as by (5.8) we have dpqq ą mw´1, we also have m ´ wdpqq ă 0. It then follows from the 1st
part of Proposition 5.13 that Ppλq P HoldpqqpΛ; L pppqq,8qq, where ppqq :" npwdpqq ´ mq´1. As
dpqq " pm ` nq´1qw´1 it is immediate to check that ppqq " q. Therefore, we see that Ppλq is
contained in HoldpqqpΛ; L pq,8qq. This proves (5.6) and (5.7). The proof is complete.
(cid:3)
For the trace-class we have the following result.
Proposition 5.15. Let Ppλq P Ψm,dpAθ; Λq with d P R and m ă ´n.
(1) If d ą pm ` nqw´1, then Ppλq P HoldpΛ; L 1q and, for any pseudo-cone Λ1 ĂĂ Λ, there is
(5.9)
"
CΛ1 ą 0 such that
Tr
Tr
"
there is CΛ1d1 ą 0 such that
Ppλq
‰ ď CΛ1p1 ` λqd
‰ ď CΛ1d1p1 ` λqd1
Ppλq
@λ P Λ1.
@λ P Λ1.
(2) If d ď pm` nqw´1 ă d1 ď 0, then Ppλq P Hold1pΛ; L 1q and, for any pseudo-cone Λ1 ĂĂ Λ,
(5.10)
Proof. Let ρpξ; λq P Sm,dpRn Λ; Aθq be such that Ppλq " Pρpλq. We know by Proposition 4.39
that ρpξ; λq P Sm`wd´,dpRn Λ; Aθq, where d´ " maxp´d, 0q. Suppose that d ą pm ` nqw´1. If
d ě 0, then m ` wd´ " m ă ´n, and if d ă 0, then we also have m ` wd´ " m ´ wd ă ´n. It
then follows from the 2nd part of Proposition 5.13 that Ppλq P HoldpΛ; L 1q. This means that, for
every pseudo-cone Λ1 ĂĂ Λ, there is CΛ1 ą 0 such that
}Ppλq}L 1 ď CΛ1p1 ` λqd
@λ P Λ1.
Here }T}L 1 " TrpTq, T P L 1. Combining this with the inequality TrpTq ď }T}L 1 gives (5.9).
Suppose now that d ď pm ` nqw´1, and let d1 P ppm ` nqw´1, 0s. This implies that d ă d1 ď 0,
and so by using Proposition 4.39 we see that ρpξ; λq P Sm´wd1,d1pRn Λ; Aθq. As d1 ą pm` nqw´1,
we have m´ wd1 ă m´ wpm` nqw´1 " ´n. Therefore, by using the 2nd part of Proposition 5.13
once again we see that Ppλq P Hold1pΛ; L 1q. In the same way as above this gives (5.10). The
(cid:3)
proof is complete.
5.5. Toroidal ΨDOs with parameter. We know by Proposition 3.36 that the classes of standard
and toroidal ΨDOs on Aθ agree. We shall now explain how to obtain analogous results for ΨDOs
with parameter. We shall keep on using the notation of §§3.7.
As mentioned in [22] each toroidal symbol space SmpRn; Aθq, m P R, is a Fr´echet space with
respect to the locally convex topology generated by the semi-norms,
pρkqkPZn ÝÑ sup
α`βďN
sup
kPZn
p1 ` kq´m`β››δα∆βρk
››,
26
N ě 0.
pρkqkPZn ÝÑ sup
kPZn
p1 ` kq´N
››δαρk
››,
›› ď CΛ1αβp1 ` λqdp1 ` kqm´β
››δα∆βρkpλq
N ě 0, α P Nn
0 .
@pk, λq P Zn Λ1.
Likewise, S pZn; Aθq is a Fr´echet space with respect to the topology generated by the semi-norms,
Definition 5.16. Sm,dpZn Λ; Aθq, m, d P R, consists of sequences pρkpλqqkPZn Ă HolpΛq such
that, for all pseudo-cones Λ1 ĂĂ Λ and multi-orders α and β, there is CΛ1αβ ą 0 such that
Remark 5.17. In the same way as in Remark 4.20 the space Sm,dpZnΛ; Aθq is naturally identified
with HoldpΛ; SmpZn; Aθqq.
Definition 5.18. S´8,dpZn Λ; Aθq, d P R, consists of sequences pρkpλqqkPZn Ă HolpΛq such
that, given any N ě 0, for all pseudo-cones Λ1 ĂĂ Λ and multi-orders α, β, there is CΛ1N αβ ą 0
such that
›› ď CΛ1N αβp1 ` λqdp1 ` kq´N
››δα∆βρkpλq
@pk, λq P Zn Λ1.
Remark 5.19. S´8,dpZn Λ; Aθq "
Remark 5.20. In the same way as in Remark 4.20 and Remark 5.17 the space S´8,dpZn Λ; Aθq
is naturally identified with HoldpΛ; S pZn; Aθqq.
mPR Sm,dpZn Λ; Aθq.
In what follows, given m P R Y t´8u, by a toroidal ΨDO with parameter of order m we shall
mean a family of operators Ppλq : Aθ Ñ Aθ parametrized by λ P Λ for which there is a toroidal
symbol pρkpλqqkPZn P Sm,dpZn Λ; Aθq, such that
Ş
Ppλqu "
ukρkpλqU k
for all u "
ukU k P Aθ.
ÿ
kPZn
ÿ
kPZn
We will need the following versions of Proposition 3.32 and Proposition 3.35.
Lemma 5.21. The following linear maps are continuous:
(i) The restriction maps SmpRn; Aθq Q ρpξq Ñ pρpkqqkPZn P SmpZn; Aθq, m P R.
(ii) The extension maps SmpZn; Aθq Q pρkqkPZn Ñ ρpξq P SmpRn; Aθq, m P R, and S pZn; Aθq Q
pρkqkPZn Ñ ρpξq P S pRn; Aθq given by (3.14).
Proof. Recall that SmpRn; Aθq and SmpZn; Aθq are Fr´echet spaces. It is straightforward to check
that, for every m P R, the graph of the restriction map is a closed subspace of SmpRn; Aθq
SmpZn; Aθq. It then follows from the closed graph theorem we get a continuous linear map from
SmpRn; Aθq and SmpZn; Aθq. This proves (i). For a proof of (ii) see [22, Remark 6.23] and [22,
(cid:3)
Remark 6.26]. The proof is complete.
Combining the first part of Lemma 5.21 with Proposition 4.25 gives the following result.
Proposition 5.22. Let ρpξ; λq P Sm,dpRn Λ; Aθq, m, d P R. Then the restriction pρpk; λqqkPZn
is contained in Sm,dpZn Λ; Aθq.
Remark 5.23. It is immediate from the definitions of S´8,dpRn Λ; Aθq and S´8,dpZn Λ; Aθq
that, if ρpξ; λq P S´8,dpRn Λ; Aθq, then pρpk; λqqkPZn P S´8,dpZn Λ; Aθq.
Let pρkpλqqkPZn P Sm,dpZn Λ; Aθq. Given any λ P Λ we get a toroidal symbol pρkpλqqkPZn
in SmpZn; Aθq. We denote by ρpξ; λq P SmpRn; Aθq its extension. By using the 2nd part of
Lemma 5.21 and Proposition 4.13 we get the following result.
Proposition 5.24. Let pρkpλqqkPZn P Sm,dpZn Λ; Aθq, m P R Y t´8u, d P R. Then ρpξ; λq is
contained in Sm,dpRn Λ; Aθq.
We are now in a position to prove the following result.
Proposition 5.25. Given any m P R Y t´8u, the classes of toroidal and standard ΨDOs with
parameter of order m agree.
27
Proof. Let ρpξ; λq P Sm,dpRnΛ; Aθq. Given any λ P Λ, it follows from Proposition 3.32 that Pρpλq
is the toroidal ΨDO associated with the restriction pρpk; λqqkPZn . As pρpk; λqqkPZn P SmpZnΛ; Aθq
by Proposition 5.22 we see that Pρpλq is a toroidal ΨDO with parameter of order m.
Conversely, let pρkpλqqkPZn P SmpZn Λ; Aθq and denote by Ppλq the corresponding toroidal
ΨDO with parameter. It follows from Proposition 3.35 that Ppλq " P ρpλq. As we know from
Proposition 5.24 that the extension ρpξ; λq is contained in SmpRn Λ; Aθq, it then follows that
Ppλq is a standard ΨDO with parameter of order m. This completes the proof.
(cid:3)
Given any ρpξ; λq P Sm,dpRn Λ; Aθq and λ P Λ, we have a symbol ρpξ; λq P SmpRn; Aθq. We
then denote by ρpξ; λq the extension (3.14) of the toroidal symbol pρpk; λqqkPZn .
Lemma 5.26. The maps ρpξq Ñ ρpξq and ρpξq Ñ ρpξq ´ ρpξq are continuous linear maps from
SmpRn; Aθq to SmpRn; Aθq and S pRn; Aθq, respectively.
Proof. It follows from Lemma 5.21 that ρpξq Ñ ρpξq is a continuous linear map from SmpRn; Aθq to
itself, since this is the composition of the restriction and extension maps. This also implies that the
linear map ρpξq Ñ ρpξq´ ρpξq is continuous from SmpRn; Aθq to itself, and so its graph is a closed
subspace of SmpRn; Aθq SmpRn; Aθq. As this graph is contained in SmpRn; Aθq S pRn; Aθq
and the inclusion of S pRn; Aθq into SmpRn; Aθq is continuous, this graph is a closed subspace
of SmpRn; Aθq S pRn; Aθq as well. The closed graph theorem then ensures us that we have a
continuous linear map from SmpRn; Aθq to S pRn; Aθq. The proof is complete.
(cid:3)
This leads us to the following parameter version of Proposition 3.32 and Proposition 3.37.
Proposition 5.27. Let ρpξ; λq P Sm,dpRn Λ; Aθq, m, d P R.
(i) ρpξ; λq P Sm,dpRn Λ; Aθq and ρpξ; λq ´ ρpξ; λq P S´8,dpRn Λ; Aθq.
(ii) Suppose that pρpk; λqqkPZn P S´8,dpZn Λ; Aθq. Then ρpξ; λq P S´8,dpRn Λ; Aθq.
Proof. The first part follows from Proposition 4.25 and Lemma 5.26. It remains to prove part (ii).
Suppose that pρpk; λqqkPZn P S´8,dpZn Λ; Aθq. We then know by Proposition 5.24 that ρpξ; λq
is in S´8,dpRn Λ; Aθq. As ρpξ; λq´ ρpξ; λq P S´8,dpRn Λ; Aθq by part (i), we then deduce that
ρpξ; λq P S´8,dpRn Λ; Aθq. The proof is complete.
(cid:3)
Combining the 2nd part of Proposition 5.27 and (3.13) leads us to the following statement.
We equip L pA 1
Corollary 5.28. Let ρpξ; λq P Sm,dpRn Λ; Aθq be such that Pρpλq " 0. Then ρpξ; λq is contained
in S´8,dpRn Λ; Aθq.
5.6. Smoothing operators with parameter. We would like to have a HoldpΛq-version of the
characterization of smoothing operators provided by Proposition 3.38. Once again a smoothing
operator is a linear operator on Aθ that extends to a continuous linear map from A 1
θ to Aθ. Thus,
the space of smoothing operators is naturally identified with L pA 1
θ , Aθq with its strong topology, i.e., the topology of uniform convergence on
θ itself is equipped with its strong topology. Given s, t P R, we
bounded subsets of A 1
denote by } }s,t the operator norm of L pH psq
their inductive limit yields the topology of A 1
L pA 1
Proposition 5.29. The topology of L pA 1
Proof. See Appendix B.
By Proposition 3.27 the topology of Aθ is the projective limit of the H psq
θ , Aθq in terms of the norms } }s,t. Namely, we have the following result.
θ , Aθq is generated by the norms } }s,t, s, t P R.
-topologies while
θ . This enables us to understand the topology of
θ
θ , where A 1
θ , Aθq.
, H ptq
q.
θ
θ
(cid:3)
Proposition 5.30. The following holds.
(1) The linear map S pRn; Aθq Q ρpξq Ñ Pρ P L pA 1
(2) There is a continuous linear map L pA 1
for all R P L pA 1
θ , Aθq.
θ , Aθq is continuous.
θ , Aθq Q R Ñ ρRpξq P S pRn; Aθq such that R " PρR
28
θ
θ
θ
θ
, H ptq
, H ptq
q " L pH psq
θ , Aθq. This proves the first part.
It remains to prove the 2nd part. Given any R P L pA 1
Proof. Let s, t P R and set m " s ´ t. The quantization map ρpξq Ñ Pρ gives rise to a continuous
linear map from SmpRn; Aθq to L pH pt`mq
q. Combining this with the
continuity of the inclusion of S pRn; Aθq into SmpRn; Aθq shows that ρ Ñ }Pρ}s,t is a continuous
semi-norm on S pRn; Aθq. As by Proposition 5.29 the norms }}s,t, s, t P R, generate the topology
of L pA 1
θ , Aθq, it then follows that ρpξq Ñ Pρ is a continuous linear map from S pRn; Aθq to
L pA 1
θ , Aθq we know by Proposition 3.38
there is ρRpξq P S pRn; Aθq such that R is the ΨDO associated with ρRpξq. As the proof of [22,
Proposition 6.30] shows we may take ρRpξq to be the extension (3.14) of the toroidal symbol
pRrU kspU kqqkPZn (this can be also seen by using (3.13) and Proposition 3.37).
Claim. The linear map L pA 1
Proof of the Claim. Given N P N0 and α P Nn
0 , let s ă ´ n
ÿ
contained in H psq
ÿ
and it follows from (3.11) that }U k}s " p1`k2q s
2 ´ N . Any unitary U k, k P Zn, is
2 . As 2s` 2N ă ´n, we have
θ , Aθq Q R Ñ pRrU ksqkPZn P S pZn; Aθq is continuous.
θ
››p1 ` kqN U k}2
s "
kPZn
kPZn
p1 ` k2qsp1 ` kq2N ă 8.
2 it can be shown (see, e.g., [23, Appendix A])
››p1 ` kqN U k
››
.
θ
"
››δα
This shows that the family tp1 ` kqN U k; k P Znu is bounded in H psq
that there is Cs ą 0 independent of α such that
Bearing this in mind, set t " α ´ s. As ´s ą n
Therefore, for all k P Zn, we have
RpU kq
››δαu} ď Cs}u}t
››RpU kq
‰›› ď Csp1 ` kqN
››
‰›› ď CN α
››δα
››R
››
"
RpU kq
As }}s,t is a continuous semi-norm on L pA 1
linear map from L pA 1
As tp1 ` kqN U k; k P Znu is bounded in H psq
p1 ` kqN
p1 ` kqN
θ , Aθq to S pZn; Aθq. The claim is proved.
@u P Aθ.
t ď Cs
››R
sup
kPZn
››
s,t
s,t
θ
s .
, we deduce there is CN α ą 0 such that
@R P L pA 1
θ , Aθq.
θ , Aθq this shows that R Ñ pRrU ksqkPZn is continuous
(cid:3)
Bearing this in mind, it follows from the proof of [22, Lemma 6.29] that we have a continuous
linear map,
S pZn; Aθq Q pρkqkPZn ÝÑ pρkpU kqqkPZn P S pZn; Aθq.
Composing it with the extension map (3.14) we get a continuous linear map from S pZn; Aθq to
S pRn; Aθq. As ρRpξ; λq is precisely the image of pRrU ksqkPZn under that map, by using the claim
above we see that R Ñ ρRpξq is a continuous linear map from L pA 1
θ , Aθq to S pRn; Aθq. This
(cid:3)
proves the 2nd part. The proof is complete.
θ , Aθq
We are now in a position to prove the following HoldpΛq-version of Proposition 3.38.
Proposition 5.31. We have
č
`
`
`
Proof. It follows from Proposition 4.25 and Proposition 5.30 that we have quantization and symbol
maps,
Ψ´8,dpAθ; Λq " Hold
Λ; L pA 1
θ
θ
"
Hold
Λ; L
s,tPR
H psq
, H ptq
θ , Aθq
,
Q Rpλq ÝÑ ρRpξ; λq P S´8,dpRn Λ; Aθq.
Λ; L pA 1
`
S´8,dpRn Λ; Aθq Q ρpξ; λq ÝÑ Pρpλq P Hold
Hold
Λ; L pA 1
θ , Aθq
`
.
(5.11)
Furthermore, the symbol map is a right-inverse of the quantization map.
ensures us that Rpλq P HoldpΛ; L pA 1
If Rpλq P Ψ´8,dpAθ; Λq, then Rpλq " Pρpλq for some ρpξ; λq P S´8,dpRn Λ; Aθq, and so (5.11)
θ , Aθqq, then
θ , Aθqq. Conversely, if Rpλq P HoldpΛ; L pA 1
29
θ
θ
θ
θ
Hold
, H ptq
θ
, H ptq
Ş
ρRpξ; λq P S´8,dpRn Λ; Aθq and Rpλq " PρRpλq, and hence Rpλq P Ψ´8,dpAθ; Λq. This shows
that Ψ´8,dpAθ; Λq " HoldpΛ; L pA 1
θ , Aθqq.
s,tPR HoldpΛ; L pH psq
θ , Aθqq "
qq. The continuity
It remains to show that HoldpΛ; L pA 1
`
`
`
θ , Aθq into L pH psq
, H ptq
of the inclusions of L pA 1
q, s, t P R, and Proposition 4.13 ensure us that
Ş
θ
H psq
θ , Aθq
Ă Hold
Λ; L pA 1
Λ; L
(5.12)
s,tPR HoldpΛ; L pH psq
, H ptq
Conversely, let Rpλq P
qq. Then, for every λ P Λ, the operator Rpλq
is contained in all the spaces L pH psq
, H ptq
q, s, t P R, and so this is a smoothing operator by
Proposition 3.38. Moreover, given any s, t P R, for every pseudo-cone Λ1 ĂĂ Λ, there is CΛ1st ą 0
such that
for all s, t P R.
θ , Aθqq. Combining this with (5.12) shows that HoldpΛ; L pA 1
s,t ď CΛ1stp1 ` λqd
Ş
As by Proposition 5.29 the norms }}s,t, s, t P R, generate the topology of L pA 1
that Rpλq P HoldpΛ; L pA 1
Ş
s,tPR HoldpΛ; L pH psq
, H ptq
agrees with
Ş
mPR Ψm,dpAθ; Λq.
Corollary 5.32. Ψ´8,dpAθ; Λq "
mPR Ψm,dpAθ; Λq (cf. Remark 5.5). Conversely, if Rpλq is
Proof. We know that Ψ´8,dpAθ; Λq Ă
in Ψm,dpAθ; Λq for every m P R, then by Proposition 5.10 it is in HoldpΛ; L pH psq
, H ptq
qq for all
s, t P R. Proposition 5.31 then shows that Rpλq P Ψ´8,dpAθ; Λq. The proof is complete.
(cid:3)
θ , Aθq, this shows
θ , Aθqq
(cid:3)
qq. The proof is complete.
››
››Rpλq
for all s, t P Λ1.
θ
θ
θ
θ
θ
θ
θ
θ
6. The Resolvent of an Elliptic ΨDO
ř
In this section, we show that the pseudodifferential calculus with parameter from the previous
Throughout this section, we let P : Aθ Ñ Aθ be an elliptic ΨDO of order w ą 0 with symbol
two sections is a natural nest for the resolvent of elliptic operators.
ρpξq „
6.1. Parametrix construction. We start by constructing an explicit parametrix for P ´ λ. The
first task is single out the relevant set of parameters for which we have a parametrix.
Recall that by Proposition 2.3 ρwpξq is invertible if and only if it is invertible in L pHθq.
jě0 ρw´jpξq.
Remark 6.2. The fact that Sp ρwptξq " Sp twρwpξq " tw Sp ρwpξq for all t ą 0 implies that ΘpPq
is invariant under positive dilations. It then follows that ΘpPq is a cone in C.
Example 6.3. Suppose that ρwpξq is selfadjoint for all ξ P Rnzt0u (e.g., P is selfadjoint). Then
Sppρwpξqq Ă R for all ξ P Rnzt0u, and so ΘpPq Ą CzR.
Example 6.4. If ρwpξq is positive in the sense of (3.15), then Sppρwpξqq Ă p0,8q, and so ΘpPq
contains Czr0,8q. As ΘpPq is a cone and Sppρwpξqq cannot be empty, we see that ΘpPq " Czr0,8q.
Definition 6.5. We say that P is elliptic with parameter when ΘpPq ‰ H.
Throughout the rest of this section we assume that P is elliptic with parameter.
Lemma 6.6. ΘpPq is an open cone in C.
Proof. We know by Remark 6.2 that ΘpPq is a cone in C. Let λ0 P ΘpPq, and set φ0 " arg λ0 P
p´π, πs. By homogeneity the whole ray tteiφ0 ; t ą 0u is contained in ΘpPq. Together with the
ellipticity of P this implies that ρwpξq ´ teiφ0 is invertible for all ξ P Rnz0 and t ě 0. In addition,
set r0 " supξPSn´1 }ρwpξq}. If λ ą r0 and ξ P Sn´1, then λ ą }ρwpξq} for all ξ P Sn´1, and so
ρwpξq ´ λ is invertible.
30
„ ď
Definition 6.1. The elliptic parameter set of P is
Sppρwpξqq
(cid:32)
(
ΘpPq " Cz
λ P C; ρwpξq ´ λ is invertible for all ξ P Rnz0
ξPRnz0
"
.
for all ξ P Sn´1 and t P r0, rs. As A ´1
Let r ą r0. As mentioned above ρwpξq´teiφ0 P A ´1
is an
open set, we see that every pξ1, t1q P Sn´1 r0, rs admits an open neighborhood U1 Ă Sn´1 r0, rs
such that there is δ ą 0 for which ρwpξq ´ teiφ P A ´1
for all pξ, tq P U1 and φ P pφ0 ´ δ, φ0 ` δq.
The compactness of Sn´1 r0, rs allows us to cover Sn´1 r0, rs by finitely many such open sets.
Therefore, we can find δ ą 0 such that ρwpξq ´ teiφ P A ´1
for all pξ, tq P Sn´1 r0, rs and
φ P pφ0 ´ δ, φ0 ` δq. As r ą r0 and ρwpξq ´ λ P A ´1
θ when ξ P Sn´1 and λ ą r0, we deduce that
ρwpξq´ λ is invertible for all ξ P Sn´1 and all λ in the open angular sector Θ :" t arg λ´ φ0 ă δu.
By homogeneity this implies that ρwpξq ´ λ is invertible for all ξ P Rnz0 and λ P Θ. That is, Θ
is contained in ΘpPq. As Θ is an open set this implies that ΘpPq is a neighborhood of λ0. Thus,
ΘpPq is a neighborhood of each of its points, and so this is an open set. The proof is complete. (cid:3)
θ
θ
θ
θ
In what follows, setting c :" inft}ρwpξq´1}´1; ξ " 1u we define
ΩcpPq " tpξ, λq P pRnz0q C; λ P ΘpPq or λ ă cξwu.
(6.1)
In the notation of (4.11) this is just the open set ΩcpΘpPqq associated with the open cone ΘpPq.
Lemma 6.7. The following holds.
(1) ρwpξq ´ λ is invertible for all pξ, λq P ΩcpPq.
(2) pρwpξq ´ λq´1 P S´1´wpΩcpPq; Aθq.
Proof. Set Ω0pPq " pRnz0q ΘpPq. We have
ΩcpPq " Ω0pPq Y tpξ, λq P pRnz0q C; λ ă cξwu.
The very definition of ΘpPq means that ρwpξq ´ λ is invertible for all pξ, λq P Ω0pPq. Let pξ, λq P
pRnz0q C be such that λ ă cξw. The definition of c implies that }ρwpηq´1} ď c´1 for all
η P Sn´1. In particular, the inequality holds for η " ξ´1ξ. Thus,
`
´1} ď ξ´wc´1 ă λ´1.
}ρwpξq´1} " ξ´w}ρw
ξ´1ξ
This implies that λ´1 ´ ρwpξq´1 is invertible. As ρwpξq ´ λ " λρwpξqpλ´1 ´ ρwpξq´1q, we deduce
that ρwpξq ´ λ is invertible when 0 ă λ ă cξw. This proves the first part.
The homogeneity of ρwpξq implies that pρwptξq ´ twλq´1 " t´wpρwpξq ´ λq´1 for all pξ, λq P
ΩcpPq and t ą 0. Moreover, the differentiability of the inverse map u Ñ u´1 of Aθ ensures us
that pρwpξq ´ λq´1 P C8pΩcpPq; Aθq (see [23, Lemma 11.2]). Moreover, as Bλpρwpξq ´ λq´1 "
pρwpξq ´ λq´1Bλpλqpρwpξq ´ λq´1 " 0, we see that pρwpξq ´ λq´1 is holomorphic with respect to
λ, and hence pρwpξq ´ λq´1 P C8,ωpΩcpPq; Aθq.
To complete the proof it just remains to show that pρwpξq ´ λq´1 P C8,´1pΩ0pPq; Aθq. Note
that if λ ą }ρwpξq}, then
}pρwpξq ´ λq´1} " λ´1}p1 ´ λ´1ρwpξqq´1} ď
1 ´ λ´1}ρwpξq} .
(6.2)
2λ for all ξ P K,
Let K Ă Rnz0 be compact. If λ ą 2 supt}ρwpξq}; ξ P Ku, then }ρwpξq} ď 1
and so by using (6.2) we get }pρwpξq ´ λq´1} ď 2λ´1. It then follows that, for every pseudo-cone
Θ1 ĂĂ ΘpPq, there is CΘ1K ą 0 such that
(6.3)
Given any multi-orders α and β the multi-derivative δαBβ
products of the form,
ξ pρwpξq´ λq´1 is a linear combination of
}pρwpξq ´ λq´1} ď CΘ1K
@pξ, λq P K Θ1.
´1
1 ` λ
`
pρwpξq ´ λq´1rδα1Bβ1
ξ ρwpξqspρwpξq ´ λq´1 rδαkBβk
ξ ρwpξqspρwpξq ´ λq´1.
Combining this with (6.3) allows us to show that, for every pseudo-cone Θ1 ĂĂ ΘpPq, there is
CΘ1Kαβ ą 0 such that
}δαBβ
ξ pρwpξq ´ λq´1} ď CΘ1Kαβ
@pξ, λq P K Θ1.
´1
1 ` λ
`
This shows pρwpξq ´ λq´1 P C8,´1pΩ0pPq; Aθq. The proof is complete.
(cid:3)
λ´1
31
In what follows, given any R ą 0, we denote by ΛR the open pseudo-cone defined by
ΛR " ΘpPq Y Dp0, Rq,
where Dp0, Rq is the open disk of radius R centered at the origin.
Theorem 6.8. Suppose that P is elliptic with parameter. Then, for every R ą 0, the following
holds.
(1) P ´ λ admits a parametrix Qpλq P Ψ´w,´1pAθ; ΛRq in the sense that
pP ´ λqQpλq " QpλqpP ´ λq " 1 mod Ψ´8,0pAθ; ΛRq.
(2) Any parametrix Qpλq P Ψ´w,´1pAθ; ΛRq as above has symbol σpξ; λq „
´1
where σ´w´jpξ; λq P S´1´w´jpΩcpPq; Aθq is given by
ρwpξq ´ λ
σ´wpξ; λq "
,
`
ρwpξq ´ λ
`
´1Bα
σ´w´jpξ; λq " ´
(6.4)
(6.5)
(6.6)
ÿ
1
α!
k`l`α"j
lăj
ξ ρw´kpξqδασ´w´lpξ; λq,
j ě 1.
ř
jě0 σ´w´jpξ; λq,
Proof. We regard P ´ λ as an element of Ψw,1pAθ; Aθq (cf. Example 5.3). We claim that the
formulas (6.5) -- (6.6) define maps σ´w´jpξ; λq P S´1´w´jpΩcpPq; Aθq for j " 0, 1, . . .. We prove this
by induction on j. For j " 0 this is the contents of Lemma 6.7. Assume the result is true for l ă j
with j ě 1. Then σ´w´jpξ; λq is a linear combination of terms of the form,
k ` l ` α " j,
(6.7)
ξ ρw´kpξqδασ´w´lpξ; λq,
´1Bα
ρwpξq ´ λ
l ă j.
`
w´k´αpΩcpPq; Aθq Ă S1
as well. Therefore, by using Remark 4.33 we see that each product of the form (6.7) is contained
By Proposition 4.43 there is σpξ; λq P S´w,´1pRnΛR; Aθq such that σpξ; λq „
ξ ρw´kpξq as
w´k´αpΩcpPq; Aθq. Moreover, by assumption σ´w´lpξ; λq P
Here pρwpξq´ λq´1 P S´1´wpΩcpPq; Aθq. As mentioned in Example 4.31 we can regard Bα
an element of S0
S´1´w´lpΩcpPq; Aθq, and so it follows from Remark 4.32 that δασ´w´lpξ; λq is contained in S´1´w´lpΩcpPq; Aθq
ř
in S´1´w´jpΩcpPq; Aθq, and hence σ´w´jpξ; λq P S´1´w´jpΩcpPq; Aθq.
jě0 σ´w´jpξ; λq.
ř
Set Qpλq " Pσpλq. Then Qpλq P Ψ´w,´1pAθ; ΛRq. Moreover, by Proposition 5.9 the composite
pP ´ λqQpλq is contained in Ψ0,0pAθ; ΛRq and has symbol rpρ ´ λq7σspξ; λq. Moreover, it follows
`
from (5.5) and (6.5) -- (6.6) that rpρ ´ λq7σspξ; λq „
jě0rpρ ´ λq7σs´jpξ; λq, where
0pξ; λq "
ρwpξq ´ λ
ÿ
σ´wpξ; λq " 1,
and for j ě 1 the homogeneous component rpρ ´ λq7σs´jpξ; λq is equal to
"
pρ ´ λq7σ
`
‰
ρwpξq ´ λ
σ´w´jpξ; λq `
ξ ρw´kpξqδασ´w´lpξ; λq " 0.
Bα
1
α!
k`l`α"j
lăj
This means that pρ ´ λq7σpξ; λq „ 1, i.e., pρ ´ λq7σpξ; λq ´ 1 P S´8,0pRn ΛR; Aθq. Thus,
pP ´ λqQpλq " Ppρ´λq7σpλq " 1
mod Ψ´8,0pAθ; ΛRq.
The invertibility of ρwpξq ´ λ also enables us to construct σ´w´jpξ; λq P S´1´w´jpΩcpPq; Aθq,
"
j " 0, 1, . . ., such that in the notation of (5.5), for j " 0, we have
0pξ; λq " σ´wpξ; λq
ρwpξq ´ λ
(6.8)
and rσ7pρ ´ λqs´jpξ; λq, j ě 1, is equal to
`
ρwpξq ´ λ
‰
σ7pρ ´ λq
`
σ´w´jpξ; λq
ÿ
(6.9)
`
" 1,
ξ σ´w´kpξ; λqδαρw´lpξq " 0.
Bα
ř
By Proposition 4.43 there is σpξ; λq P S´w,´1pRn ΛR; Aθq such that σpξ; λq „
jě0 σ´w´jpξ; λq.
As above (6.8) -- (6.9) implies that σ7pρ ´ λq ´ 1 is in S´8,0pRn ΛR; Aθq. Thus, if we set Qpλq "
Pσpλq, then Qpλq P Ψ´w,´1pAθ; ΛRq and QpλqpP ´ λq " 1 modulo Ψ´8,0pAθ; ΛRq.
k`l`α"j
1
α!
kăj
32
Set Rpλq " pP ´ λqQpλq ´ 1 and Rpλq " QpλqpP ´ λq ´ 1. Then we have
" Qpλq ` QpλqRpλq,
Qpλq " Qpλq ` RpλqQpλq.
`
QpλqpP ´ λqQpλq " Qpλq
QpλqpP ´ λqQpλq "
`
1 ` Rpλq
1 ` Rpλq
As QpλqRpλq and RpλqQpλq are in Ψ´8,´1pAθ; ΛRq, we see that Qpλq ´ Qpλq is contained in
Ψ´8,´1pAθ; ΛRq. Thus,
QpλqpP ´ λq " QpλqpP ´ λq " 1 mod Ψ´8,0pAθ; ΛRq.
This shows that Qpλq is a parametrix for P ´ λ in the sense of (6.4).
The above considerations to compare Qpλq and Qpλq also shows that if Q1pλq P Ψ´w,´1pAθ; ΛRq
is any other parametrix in the sense of (6.4), then Q1pλq ´ Qpλq is in Ψ´8,´1pAθ; ΛRq. Let
σp1qpξ; λq P S´w,´1pRn ΛR; Aθq be the symbol of Q1pλq. Then Corollary 5.28 ensures us that
σp1qpξ; λq ´ σpξ; λq is contained in S´8,´1pRn ΛR; Aθq, and so σpξ; λq „
σ´w´jpξ; λq, where
the σ´w´jpξ; λq are given by (6.5) -- (6.6). This proves the 2nd part. The proof is complete.
(cid:3)
ř
Remark 6.9. The smoothing process of Proposition 4.43 involves cutoffs by means of functions
χpξq P C8
c pRnq. These are elements of S´8,0pRn ΛR; Aθq which are by no mean unique. This
creates an ambiguity with values in S´8,´1pRn ΛR; Aθq in the construction of the symbol of
the parametrix of P ´ λ. Note also that, in view of (6.6), the λ-decay of symbols σ´w´jpξ; λq
need not decrease as j becomes large, since we may have some non-zero contribution from terms
like pρwpξq ´ λq´1ρw´j`1pξqσ´w´1pξ; λq with j arbitrary large. Because of all this we can't really
expect obtaining better λ-decay for the remainder terms in (6.4). When we restrict the parameter
set to ΘpPq we can improve the decay by adding some meromorphic singularities near λ " 0
(see the proof of Theorem 6.10 below). In some forthcoming work, we will explain that when P
is a differential operator and we also take the parameter set to ΘpPq, then there is no need to
use cut-off functions anymore and we then can modify the parametrix construction so as to have
λ-decay of any order.
6.2. Localization of the spectrum and rays of minimal growth. We shall now use Theo-
rem 6.8 to look at the localization of the spectrum of P .
Theorem 6.10. Suppose that P is elliptic with parameter.
(1) The spectrum of P is an unbounded discrete subset of C consisting of eigenvalues with
(2) For any cone Θ1 such that Θ1zt0u Ă ΘpPq the following holds.
finite multiplicity.
(a) Θ1 contains at most finitely many eigenvalues of P .
(b) There are r0 and C ą 0 such that
››pP ´ λq´1
›› ď Cλ´1
@λ P Θ1zDp0, r0q.
(6.10)
Proof. By Theorem 6.8 there are Qpλq P Ψ´w,´1pAθ; ΘpPqq and Rpλq P Ψ´8,0pAθ; ΘpPqq such
that pP ´ λqQpλq ` Rpλq " 1. Set Q1pλq " Qpλq ´ λ´1Rpλq and R1pλq " λ´1P Rpλq. Then
Q1pλq P Ψ´w,´1pAθ; ΘpPqq and R1pλq P Ψ´8,´1pAθ; ΘpPqq. Moreover, we have
pP ´ λqQ1pλq " pP ´ λqQpλq ´ λ´1P Rpλq ` Rpλq " 1 ´ R1pλq.
(6.11)
Let Θ1 be a cone such that Θ1zt0u Ă ΘpPq. This implies that Θ1zDp0, rq ĂĂ ΘpPq for any r ą 0.
Moreover, as R1pλq is in Ψ´8,´1pAθ; ΘpPqq, Proposition 5.31 allows us to regard it as an element
of Hol´1pΘpPq; L pHθqq. Therefore, there is CΘ1r ą 0 such that }R1pλq} ď CΘ1rλ´1 for all
λ P Θ1zDp0, rq. It then follows that there is r0 ą 0 such that }R1pλq} ď 1
2 for all λ P Θ1zDp0, r0q.
This ensures that, for every λ P Θ1zDp0, r0q, the operator 1 ´ R1pλq is invertible in L pHθq and
}p1 ´ R1pλqq´1} ď p1 ´ }R1pλq}q´1 ď 2.
As Q1pλq P Ψ´w,´1pAθ; ΘpPqq, it follows from Proposition 5.11 that
Q1pλq P Hol´1
ΘpPq; L pHθq
X Hol0
ΘpPq; L pHθ, H pwq
θ
q
.
`
`
33
In particular, the equalities (6.11) hold on Hθ. Let λ P Θ1zDp0, r0q. Then Q1pλqp1 ´ R1pλqq´1 is
in L pHθ, H pwq
q, and by using (6.11) we see that on L pHθq we have
pP ´ λqQ1pλq
1 ´ R1pλq
´1 " p1 ´ R1pλqq
`
´1 " 1.
1 ´ R1pλq
`
θ
That is, Q1pλqp1 ´ R1pλqq´1 is a right inverse of P ´ λ.
We can similarly construct Q2pλq P Ψ´w,´1pAθ; ΘpPqq and R2pλq P Ψ´8,´1pAθ; ΘpPqq such
that Q2pλqpP ´ λq " 1 ´ R2pλq. We regard R2pλq as an element of Hol´1pΘpPq; L pH pwq
qq. In
the same way as above, by taking r0 large enough 1 ´ R2pλq becomes invertible in L pH pwq
q
for all λ P Θ1zDp0, r0q. It then can be shown that p1 ´ R2pλqq´1Q2pλq P L pHθ, H pwq
q is a left
inverse of P ´ λ on its domain H pwq
. Therefore, it must agree with Q1pλqp1 ´ R1pλqq´1, and so
Q1pλqp1 ´ R1pλqq´1 is a bounded two-sided inverse of P ´ λ.
All this shows that Θ1zDp0, r0q is contained in Cz SppPq. This implies that SppPq ‰ C, and
so by Proposition 3.50 the spectrum of P is an unbounded discrete subset of C consisting of
eigenvalues with finite multiplicity. In particular, there are at most finitely many eigenvalues of P
in the disk Dp0, r0q. As there are no eigenvalues in Θ1zDp0, r0q, we then see that Θ1 may contain
at most finitely many of them.
Finally, as Q1pλq P Hol´1pΘpPq; L pHθqq and Θ1zDp0, r0q ĂĂ ΘpPq there is CΘ1r0 ą 0 such
that }Q1pλq} ď CΘ1r0λ´1 for all λ P Θ1zDp0, r0q. Furthermore, as shown above, if λ P Θ1zDp0, r0q
then pP ´ λq´1 " Q1pλqp1´ R1pλqq´1 and }p1´ R1pλqq´1} ď 2. Therefore, for all λ P Θ1zDp0, r0q,
we have
θ
θ
θ
θ
}pP ´ λq´1} " }Q1pλqp1 ´ R1pλqq´1} ď }Q1pλq}}p1 ´ R1pλqq´1} ď 2CΘ1r0λ´1.
(cid:3)
This gives (6.10). The proof is complete.
Definition 6.11. A ray L Ă C is called a ray of minimal growth for P when the following two
conditions are satisfied:
(i) L does not contain any eigenvalue of P .
(ii) }pP ´ λq´1} " Opλ´1q as λ goes to 8 along L.
Remark 6.12. The existence of ray of minimal growths is a crucial ingredient in the construction
of complex powers and logarithms of elliptic operators (see, e.g., [40]).
Example 6.13. If P is selfadjoint, then every ray L Ă CzR is a ray of minimal growth.
As immediate consequences of Theorem 6.10 we get the following results.
Corollary 6.14. For any cone Θ1 such that Θ1zt0u Ă ΘpPq, all but finitely many rays contained
in Θ1 are rays of minimal growth for P .
Corollary 6.15. Any ray L Ă ΘpPq that does not contain any eigenvalue of P is a ray of miminal
growth.
6.3. Analysis of the resolvent. In what follows we set
ΘpPq " ΘpPqzttλ; t ą 0, λ P Sp Pu.
That is, ΘpPq is obtained from ΘpPq by deleting all the rays through eigenvalues of P .
In
particular, this is a cone. Moreover, in view of Corollary 6.15, every ray contained in ΘpPq is a
ray of minimal growth.
Lemma 6.16. ΘpPq is a non-empty open cone in C.
Proof. Let λ0 P ΘpPq. As ΘpPq is an open cone, there is an open cone Θ1 containing λ0 and such
that Θ1zt0u Ă ΘpPq. Such a cone can be obtained for instance as the cone generated by a compact
neighborhood of λ0 in ΘpPq. Thanks to Theorem 6.10 we know that Θ1 contains at most finitely
many eigenvalues of P . Let L1, . . . , LN be the rays contained in Θ1 that contain eigenvalues of P .
Set Θ1 :" Θ1zpL1 Y Y LNq. Then Θ1 is a non-empty open cone which is contained in ΘpPq. In
particular, this implies that ΘpPq ‰ H. Moreover, if λ0 P ΘpPq, then Θ1 is an open neighborhood
of λ0 contained in ΘpPq. This shows that ΘpPq is a neighborhood of each of its point, and hence
(cid:3)
this is an open set. The proof is complete.
34
Definition 6.17. Set R0 " distp0, SppPq X Cq. The pseudo-cone ΛpPq is defined by
#
ΛpPq "
"
‰
ΘpPq Y Dp0, R0q
Dp0, R0qzt0u
ΘpPq Y
if 0 R SppPq,
if 0 P SppPq.
θ
θ
q, i.e., pP ´ λq´1 P HolpΛpPq; L pHθqq.
Proposition 6.18. pP ´ λq´1 is contained in Hol´1pΛpPq; L pHθqq.
Proof. As ΛpPq Ă Cz SppPq, we have a holomorphic map λ Ñ P ´ λ from ΛpPq to the invertible
elements of L pH pwq
, Hθq. By taking inverses we then get a holomorphic map from ΛpPq to
L pHθ, H pwq
Let Λ1 be a pseudo-cone such that Λ1 ĂĂ ΛpPq. Denote by Θ1 its conical component. Then
Θ1zt0u is contained in ΘpPq Ă ΘpPq, and so by Theorem 6.10 the estimate (6.10) holds on
ΘpPqzDp0, r0q for r0 large enough. Note that Λ1zrΘpPqzDp0, r0qs is pre-compact in ΛpPq, since
this is a bounded set whose closure is contained in ΛpPq. It then follows there is CΛ1 ą 0 such
that
This shows that pP ´ λq´1 P Hol´1pΛpPq; L pHθqq. The proof is complete.
›› ď CΛ1p1 ` λq´1
››pP ´ λq´1
@λ P Λ1.
(cid:3)
We are now in a position to prove the main result of this paper.
ř
Theorem 6.19. Suppose that P is elliptic with parameter. Then the following holds.
QpλqpP ´ λq " 1 ´ R1pλq
(1) The resolvent pP ´ λq´1 is contained in Ψ´w,´1pAθ; ΛpPqq.
(2) pP ´ λq´1 has symbol σpξ; λq „
σ´w´jpξ; λq, where σ´w´jpξ; λq is given by (6.5) -- (6.6).
Proof. By construction ΛpPq Ă ΘpPq Y Dp0, R0q. Therefore, by Theorem 6.8 there is Qpλq in
Ψ´w,´1pAθ; ΛpPqq such that
(6.12)
where R1pλq and R2pλq are in Ψ´8,0pAθ; ΛpPqq. Multiplying the first equality by pP ´ λq´1 gives
Qpλq " p1 ´ R1pλqqpP ´ λq´1, i.e.,
(6.13)
Proposition 5.1 and Proposition 5.31 ensure us that Qpλq and R1pλq are in Hol´1pΛpPq; L pAθqq
and Hol0pΛpPq; L pA 1
θ , Aθqq, respectively. Moreover, thanks to Proposition 6.18 we also know
that pP ´ λq´1 is contained in Hol´1pΛpPq; L pHθqq, and so it can be regarded as an element of
θqq. Therefore, the composition R1pλqpP ´ λq´1 gives rise to an element of
Hol´1pΛpPq; L pAθ, A 1
Hol´1pΛpPq; L pAθqq. Combining all this with (6.13) we deduce that pP ´ λq´1 is contained in
Hol´1pΛpPq; L pAθqq.
Multiplying by pP ´ λq´1 the 2nd equality in (6.12) gives Qpλq " pP ´ λq´1p1 ´ R2pλqq. That
pP ´ λq´1 " Qpλq ` R1pλqpP ´ λq´1.
pP ´ λqQpλq " 1 ´ R2pλq,
and
is,
pP ´ λq´1 " Qpλq ` pP ´ λq´1R2pλq.
(6.14)
As pP ´ λq´1 P Hol´1pΛpPq; L pAθqq and R2pλq is in Hol0pΛpPq; L pA 1
θ , Aθqq by Proposition 5.31
the composition pP ´ λq´1R2pλq is in Hol´1pΛpPq; L pA 1
θ , Aθqq. Therefore, by Proposition 5.31
this is an element of Ψ´8,´1pAθ; ΛpPqq. Combining this with (6.14) then shows that pP ´ λq´1
is contained in Ψ´w,´1pAθ; ΛpPqq. This proves the 1st part. The 2nd part follows from the 2nd
part of Theorem 6.8 and the fact that pP ´ λq´1 is a parametrix. The proof is complete.
(cid:3)
Remark 6.20. Suppose that 0 P SppPq. In this case 0 R ΛpPq, and so Theorem 6.19 asserts that
pP ´ λq´1 is a Hol´1-family of ΨDOs off λ " 0. By using analytic Fredholm theory it can be
shown that at λ " 0 we actually have a meromorphic singularity in L pHθq with finite rank poles
(see, e.g., [10, Appendix C]). It can be actually shown that the poles are smoothing operators and
by removing the meromorphic singularity we actually get a Hol´1-family of ΨDOs near λ " 0
(see [27]).
We mention a few consequences of Theorem 6.19. First, we have the following refinement of
Proposition 5.11.
35
.
¯
¯
´
ΛpPq; L
`
Proposition 6.21. For every s P R, we have
Hol´1`a
pP ´ λq´1 P
č
0ďaď1
`
H psq
θ
, H ps`awq
θ
´
Proof. We know by Theorem 6.19 that pP ´ λq´1 P Ψ´w,´1pAθ; ΛpPqq. Therefore, by using
Proposition 5.11 we see that, for every s P R, we have
ΛpPq; L
, H psq
Substituting ´1 ` a for d1 and s ` aw for s with a P r0, 1s gives the result.
pP ´ λq´1 P
H ps´w`wd1q
´1ďd1ď0
č
Hold1
(cid:3)
.
θ
θ
In another direction we have the following Schatten class properties of the resolvent.
č
Proposition 6.22. Set p " nw´1. Then
pP ´ λq´1 P
Hol´1`pq´1
Proof. We know by Theorem 6.19 that pP ´ λq´1 P Ψ´w,´1pAθ; ΛpPqq. We thus are in the setting
of Proposition 5.14 with d " ´1 and m " ´w. Here mw´1 " ´1 " d, and so (5.7) gives
¯
.
´
ΛpPq; L pq,8q
¯
´
ΛpPq; L pq,8q
.
qěp
č
qěp
pP ´ λq´1 P
Holdpqq
where dpqq " p´w ` nq´1qw´1 " ´1 ` pnw´1qq´1 " ´1 ` pq´1. This proves the result.
(cid:3)
One approach to derive spectral asymptotics is to use resolvent-type expansions. This often
leads us to estimate traces of operators of the form,
QpA0, . . . , ANqpλq " A0pP ´ λq´1A1 AN´1pP ´ λq´1AN ,
Aj P ΨajpAθq.
For instance, when A1 " AN " 1 we get the operator A0pP ´ λq´N . Note that by Proposi-
tion 5.9 and Theorem 6.19 QpA0, . . . , ANqpλq is contained in Ψm,´NpAθ; ΛpPqq where m " a´N w
and we have set a " a0 `` aN . In particular, we obtain a family of trace-class operators when
a ´ N w ă ´n.
Proposition 6.23. Let Aj P ΨajpAθq, aj P R, j " 0, . . . , N . Set a " a0 ` ` aN , and assume
that ´N w ` a ă ´n.
(1) If a ă ´n, then QpA0, . . . , ANqpλq P Hol´NpΛpPq; L 1q and, for any pseudo-cone Λ ĂĂ
ΛpPq, there is CΛ ą 0 such that
"
QpA0, . . . , ANqpλq
(2) If a ě ´n and δ ą pa ` nqw´1, then QpA0, . . . , ANqpλq P Hol´N`δpΛpPq; L 1q and, for
any pseudo-cone Λ ĂĂ ΛpPq, there is CΛδ ą 0 such that
‰ ď CΛp1 ` λq´N
‰ ď CΛδp1 ` λq´N`δ
@λ P Λ.
@λ P Λ.
Tr
Tr
"
QpA0, . . . , ANqpλq
Proof. As mentioned above QpA0, . . . , ANqpλq P Ψm,´NpAθ; ΛpPqq with m " a ´ N w. We thus
are in the framework of Proposition 5.15 with d " ´N . Here pm ` nqw´1 " pa ´ N w ` nqw´1 "
´N ` pa ` nqw´1, and so the condition d ą pm ` nqw´1 becomes ´N ą ´N ` pa ` nqw´1, i.e.,
a ă ´n. The result then follows from Proposition 5.15.
(cid:3)
7. Complex Powers of Positive Elliptic Operators
In this section, we use the results of the previous section to show that the complex powers of
positive elliptic ΨDOs are ΨDOs. We refer to [27] for a more comprehensive account on complex
powers of elliptic ΨDOs on noncommutative tori.
In what follows, we let P P ΨwpAθq be an elliptic ΨDO with symbol ρpξq „
ř
jě0 ρw´jpξq. We
make the following positivity assumptions:
36
(P1) The principal symbol ρwpξq is positive, i.e., ρwpξq is selfadjoint and has positive spectrum
for all ξ ‰ 0.
(P2) P is formally selfadjoint and has non-negative spectrum.
Note that we do not assume P to be a differential operator. The ellipticity of P and (P2) ensure us
that P with domain H pwq
is a selfadjoint Fredholm operator on Hθ with non-negative (discrete)
spectrum. In addition, (P1) and (P2) imply that ΘpPq " ΘpPq " Czr0,8q, and so we have
θ
`
ΛpPq " Czrr0,8q,
where r0 :" dist
0, SppPqz0
" }P ´1}´1.
Given z P C and λ P Czp´8, 0s, we put λz " ez log λ, where λ Ñ log λ is the continuous
determination of the logarithm on Czp´8, 0s with values in the stripe t(cid:60)λ ă πu. In particular,
the power functions λ Ñ λz are holomorphic functions on Czp´8, 0s. Moreover, they agree with
the standard power functions on p0,8q.
For z P C we define the operator P z by standard Borel functional calculus, with the convention
that P z " 0 on ker P (i.e., λz " 0 for λ " 0). Alternatively, if pe(cid:96)q(cid:96)ě0 is any orthonormal eigenbasis
of P with P e(cid:96) " λ(cid:96)e(cid:96), then we have
P ze(cid:96) " λz
(cid:96) e(cid:96)
@(cid:96) ě 0.
For (cid:60)z ď 0 we obtain a bounded operator on Hθ. For (cid:60)z ą 0 we get a closed unbounded operator
with domain H pw(cid:60)zq
. In particular, the domain of P z always contains Aθ. We also have the
group property,
θ
(7.1)
In addition, for (cid:60)z ă 0 we have the integral formula,
P z1`z2u " P z1 P z2u
ż
@u P Aθ @zi P C.
P z " 1
2iπ
Γ
λzpP ´ λq´1dλ,
where the integral converges in L pHθq and Γ is a downward-oriented keyhole contour of the form
Γ " BΛprq, with
Λprq :" tλ P C; (cid:60)λ ď 0 or λ ď ru ,
0 ă r ă r0.
Note that Λprq is a pseudo-cone and Λprq ĂĂ ΛpPq.
Γ
r
Λprq
Figure 2. Pseudo-Cone Λprq and Contour Γ " BΛprq
The main aim of this section is to show that the powers P z are ΨDOs. To reach this end we
need a couple of lemmas.
Lemma 7.1. Suppose that Λ is an open pseudo-cone containing Γ. Let σpξ; λq P Sm,dpRn Λ; Aθq
with m P R and d ă ´1.
ş
Γ σpξ; λqdλ converges in SmpRn; Aθq.
(i) The integral ρpξq "
37
(ii) We have
ż
Pσpλqdλ " Pρ,
where the integral converges in L pAθq.
Γ
ż
ξwΓ
ρpξq :"
"
λ P C;
`
Θ "
λzσpξ; λqdλ,
ξ ‰ 0,
*
π
4
ă arg λ ă 3π
4
.
Proof. The first part follows from the 1st part of Proposition A.7 and the fact that SmpRn; Aθq is
a Fr´echet space. The 2nd part follows from the 2nd part of Proposition A.7 and the continuity of
the linear map SmpRn; Aθq Q ρ Ñ Pρ P L pAθq. This completes the proof.
(cid:3)
Define
c " inf
ξ"1
and c1 " sup
ξ"1
}ρwpξq} .
The ellipticity of P ensures us that c ą 0. Moreover, as ρwpξq has positive spectrum, its spectrum
is contained in the interval r}ρwpξq´1}´1,}ρwpξq}s for every ξ ‰ 0. As by homogeneity ρwpξq "
ξwρwpξ´1ξq, we deduce that
(7.2)
In what follows we define ΩcpPq as in (6.1). Namely, as ΘpPq " Czr0,8q, we have
Sppρwpξqq Ă
cξw, c1ξw
@ξ P Rnz0.
‰
ΩcpPq " tpξ, λq P pRnz0q C; λ P Czrcξw,8qu .
Lemma 7.2. Let z P C, (cid:60)z ă 0, and σpξ; λq P S´1
contour of the form Γ " BΛprq with 0 ă r ă c. Then the integral
m pΩcpPq; Aθq, m P R. In addition, let Γ be a
››´1
››ρwpξq´1
"
converges in Aθ and defines a homogeneous symbol in Sm`wpz`1qpRn; Aθq.
Proof. As σpξ; λq is in S´1
m pΩcpPq; Aθq this is a C8,ω-map from ΩcpPq to Aθ. Set
We get an open cone containing iRz0 such that Θz0 is contained in Czr0,8q " ΘpPq. Thus, by
m pΩcpPq; Aθq, for any compact K Ă Rnz0 and multi-orders α
the very definition of the space S´1
›› ď CKαβ
and β, there is CKαβ ą 0 such that
ξ σpξ; λq
"
(7.3)
λ P C; arg λ ď π
4
@pξ, λq P K Θ.
*
For s ą 0 we also set
››δαBβ
and λ ă csw
Λ1psq :" Θ Y
´1
1 ` λ
Thus, Λ1psq is an open pseudo-cone with conical part Θ. Note that λz P Hol
(cid:60)zpΛpsqq. We also
observe that the contour ξwΓ is contained in Λ1psq for s ě ξ. Furthermore, if s ď ξ and λ P C
is such that λ ă csw, then λ ă cξw, and so pξ, λq P ΩcpPq. Thus,
for all s P p0,ξs.
tξu Λ1psq Ă ΩcpPq
Let ξ P Rnz0. As tξu Λ1pξq Ă ΩcpPq the map λ Ñ σpξ; λq is holomorphic map from
Λ1pξq to Aθ. Combining this with the estimates (7.3) for K " tξu and β " 0 we see that this
ş
is a Hol´1pΛpξqq-map, and so the map λ Ñ λzσpξ; λq is in Hol´1`(cid:60)zpΛ1pξq; Aθq. Recall that
ξwΓ Ă Λ1pξq and Aθ is a Fr´echet space. Therefore, by the 1st part of Proposition A.7, for every
ξwΓ λzσpξ; λqdλ converges in Aθ. Furthermore, by the 3rd part of
ξ ‰ 0, the integral ρpz; ξq :"
Proposition A.7 we may replace the contour ξwΓ by swΓ with any s P p0,ξs.
Let R ą 0 and set UR " tξ P Rn; ξ ą Ru. If ξ P UR, then Rw ă ξw, and so we have
.
ż
ρpξq "
λzσpξ; λqdλ
@ξ P UR.
In addition, if 0 ă λ ă cRw, then λ ă cξw, and so pξ, λq P ΩcpPq. Therefore, we see that
UR Λ1pRq Ă ΩcpPq. Combining this with the estimates (7.3) shows that σpξ; λq P C8,´1pUR
RwΓ
38
Λ1psq
csw
Figure 3. Pseudo-Cone Λ1psq
ş
Λ1pRq; Aθq, and hence λzσpξ; λq P C8,´1`(cid:60)zpUR Λ1pRq; Aθq. As C8pUR; Aθq is a Fr´echet space
(see, e.g., [22, Proposition C.23]), we may apply 1st part of Proposition A.7 to see that the integral
RwΓ λzσpξ; λqdλ converges in C8pUR; Aθq. It then follows that ρpξq is a C8-map from UR to Aθ
ż
for every R ą 0, and hence it lies in C8pRnz0; Aθq.
Let us show that ρpξq is homogeneous. Let ξ P Rnz0 and t ą 0. By Proposition A.6 we have
ż
ρptξq "
λzσptξ; λqdλ "
twz`wλzσptξ; twλqdλ.
ξwΓ
twξwΓ
As σptξ; twλq " tmσpξ; λq we obtain
ρptξq " twz`w`m
λzσpξ; λqdλ " tm`wpz`1qρpξq.
ż
ξwΓ
This shows that ρpξq is homogeneous of degree m ` wpz ` 1q, and so ρpξq P Sm`wpz`1qpRn; Aθq.
(cid:3)
ř
The proof is complete.
It follows from Theorem 6.19 that pP ´ λq´1 is in Ψ´w,´1pAθ; ΛpPqq and has symbol σpξ; λq „
jě0 σ´w´jpξ; λq, where
(7.4)
(7.5)
σ´w´jpξ; λq " ´
ξ ρw´kpξqδασ´w´lpξ; λq,
j ě 1.
ÿ
`
ρwpξq ´ λ
ρwpξq ´ λ
´1Bα
`
´1
,
σ´wpξ; λq "
1
α!
k`l`α"j
lăj
ř
ż
jě0 ρwz´jpz; ξq, with
ρwz´jpz; ξq :" 1
2iπ
We are now in a position to state and prove the main result of this section.
For every j ě 0 the symbol σ´w´jpξ; λq is in S´1´w´jpΩcpPq; Aθq, where ΩcpPq is defined as above.
Theorem 7.3. Assume that (P1) and (P2) hold. For every z P C, the operator P z is in ΨwzpAθq
and has symbol ρpz; ξq „
λzσ´w´jpξ; λqdλ,
ξ ‰ 0,
γξ
(7.6)
where σ´w´jpξ; λq is given by (7.4) -- (7.5) and γξ is any clockwise-oriented Jordan curve in Czp´8, 0s
that encloses Sppρwpξqq. In particular, the principal symbol of P z is pρwpξqqz.
ř
Proof. Let z P C, (cid:60)z ă 0. As mentioned above pP ´ λq´1 " Pσpλq with σpξ; λq P S´w,´1pRn
σ´w´jpξ; λq, where σ´w´jpξ; λq is given by (7.4) -- (7.5). In par-
ΛpPq; Aθq such that σpξ; λq „
ticular, by using Proposition 4.39 we see that σpξ; λq P S0,´1pRn ΛpPq; Aθq. It then follows from
ż
ż
Γ λzσpξ; λqdλ converges in S0pRn; Aθq, and on Aθ
Lemma 7.1 that the integral ρpz; ξq " p2iπq´1
we have
ş
P z " 1
2iπ
λz dλ
P ´ λ
Γ
" 1
2iπ
39
Γ
λzPσpλqdλ " Pρpz;q,
where the 2nd integral converges in L pAθq.
Without any loss of generality we may assume that Γ " BΛprq with 0 ă r ă minpr0, cq. Given
any j ě 0, as σ´w´jpξ; λq is contained in S´1´w´jpΩcpPq; Aθq, it follows from Lemma 7.2 that we
ż
define a homogeneous symbol in SmpjqpRn; Aθq with mpjq " p´w ´ jq ` wpz ` 1q " wz ´ j by
letting
(7.7)
ρwz´jpz; ξq :" 1
2iπ
ξwΓ
λzσ´w´jpξ; λqdλ,
ξ ‰ 0.
Let ξ P Rnz0. Recall that σ´wpλ; ξq " pρwpξq ´ λq´1, and so λ Ñ σ´wpξ; λq is a holomorphic
map from Cz Sppρwpξqq to Aθ. Using (7.5) and arguing by induction then shows that, for every
j ě 0, the map λ Ñ σ´w´jpξ; λq is holomorphic on Cz Sppρwpξqq. We also know from (7.2) that
Sppρwpξqq Ă rcξw, c1ξws. As ξwΓ Ă Czprξw,8q Ă Czrcξw,8q, we see that the contour ξwΓ
is contained in Cz Sppρwpξqq. Therefore, by the 3rd part of Proposition A.7, in the formula (7.7)
we may replace the contour ξwΓ by any clockwise-oriented Jordan curve γξ in Czp´8, 0s whose
ż
interior contains Sppρwpξqq. That is, we have
ρwz´jpz; ξq " 1
ż
2iπ
(7.8)
In particular, for j " 0 we get
λzσ´w´jpξ; λqdλ,
`
Let us show that ρpz; ξq „
ř
ρwzpz; ξq " 1
2iπ
jě0 ρwz´jpz; ξq. By Remark 4.38 the asymptotic expansion
jě0 σ´w´jpξ; λq implies that, for all N ě 0 and J ě N ` w, there is rN Jpξ; λq in
σpξ; λq „
ÿ
`
S´w´N,´1pRn ΛpPq; Aθq such that
dλ " ρwpξqz.
ρwpξq ´ λ
´1
j ě 0.
ř
λz
γξ
γξ
1 ´ χpξq
σpξ; λq "
jăJ
σ´w´jpξ; λq ` rN Jpξ; λq,
c pRnq is as in Lemma 4.34. In our setting this means that χpξq " 1 for ξ ă
w for some R ě r0. Moreover, it follows from Lemma 4.34 that p1 ´ χpξqqσ´w´jpξq is
where χpξq P C8
pc´1Rq 1
contained in S´j,´1pRn ΛpPq; Aθq. In any case, for all ξ P Rn, we have
ż
jăJ
ż
`
ÿ
ρpz; ξq " 1
2iπ
"
ż
λzσpξ; λqdλ
`
Γ
1
2iπ
λz
Γ
1 ´ χpξq
ż
σ´w´jpξ; λqdλ ` 1
ş
2iπ
Γ λzrN Jpξ; λqdλ converges in S´w´NpRn; Aθq. In
λzrN Jpξ; λqdλ.
Γ
It follows from Lemma 7.1 that the integral
addition, let j ě 0 and ξ P Rn. We claim that
λz
1
2iπ
1 ´ χpξq
σ´w´jpξ; λqdλ " p1 ´ χpξqq ρwz´jpz; ξq.
Γ
(7.9)
If ξ ď pc´1Rq 1
w , then both sides of the equation are zero, and so the equation holds. Suppose
now that ξ ą pc´1Rq 1
w . This implies that cξw ą R ě r0 ą r. As Sppρwpξqq Ă rcξw, c1ξws
we see that Γ is contained in Cz Sppρwpξqq. Therefore, in the same way as above, in the integral
ż
`
in (7.9) we may replace the contour Γ by any Jordan curve γξ as in (7.8). Thus,
ż
ż
1 ´ χpξq
`
λz
`
σ´w´jpξ; λqdλ " 1
2iπ
"
1 ´ χpξq
" p1 ´ χpξqq ρwz´jpz; ξq.
1
2iπ
γξ
γξ
σ´w´jpξ; λqdλ
λzσ´w´jpξ; λqdλ
1
2iπ
Γ
1 ´ χpξq
λz
This confirms our claim.
40
jăJ
It follows from all this that, for all N ě 1 and J ě N ` w, we have
ř
p1 ´ χpξqq ρwz´jpz; ξq mod S´w´NpRn; Aθq.
ÿ
ρpz; ξq "
ř
ř
In the same way as in the proof of Lemma 4.29 this implies that ρpz; ξq „
jě0p1´χpξqqρwz´jpz; ξq
in the sense that ρpz; ξq ´
jăNp1 ´ χpξqqρwz´jpz; ξq P Sw(cid:60)z´NpRn; Aθq for all N ě 0. It then
jě0 ρwz´jpz; ξq in the sense of (3.3) (cf. [22, Remark 3.21]). In particular,
follows that ρpz; ξq „
this shows that ρpz; ξq P SwzpRn; Aθq, and so the power P z " Pρpz;q is an operator in ΨwzpAθq.
This proves the theorem when (cid:60)z ă 0.
Suppose now that (cid:60)z ě 0, and let m P N be such that (cid:60)z ´ m ă 0. The group property (7.1)
implies that P z " P mP z´m on Aθ. Here P m P ΨmwpAθq, and, as (cid:60)z´ m ă 0, the first part of the
proof shows that P z´m P Ψwpz´mqpAθq. It then follows from Proposition 3.20 that P z P ΨwzpAθq.
It remains to show that the homogeneous symbols ρwz´jpz; ξq of P z are given by the for-
mula (7.6). By the first part of the proof the formula holds true when (cid:60)z ă 0. Let ξ P Rnz0.
We observe that the integrals p2iπq´1
λzσ´w´jpξ; λqdλ give rise to holomorphic maps from C
to Aθ. In particular, this implies that z Ñ ρwz´jpz; ξq is a holomorphic map from the half-space
t(cid:60)z ă 0u to Aθ for every j ě 0. Furthermore, the equality P z " P mP z´m and Proposition 3.20
ÿ
show that, for any m P N and (cid:60)z ă m, we have
ş
γξ
ρwz´jpz; ξq "
Bα
ξ ρmw´kpm; ξqδαρwpz´mq´lpz ´ m; ξq.
1
α!
k`l`α"j
It then follows that the map z Ñ ρwz´jpz; ξq is holomorphic on every half-space t(cid:60)z ă mu, m P N,
and hence it is holomorphic on all C. Therefore, if we fix ξ P Rnz0, then both sides of (7.6) are
holomorphic maps on C. As they agree on the half-space t(cid:60)z ă 0u they agree everywhere. The
(cid:3)
proof is complete.
Remark 7.4. It was mentioned without proof in [12] that complex powers of positive elliptic ΨDOs
on noncommutative 2-tori are ΨDOs. Therefore, Theorem 7.3 confirms the claim of [12].
We end the section with the following consequence of Theorem 7.3.
Theorem 7.5. Let P P ΨwpAθq be elliptic and have principal symbol ρwpξq.
(1) The absolute value P :" ?
(2) For every z P C, the operator Pz is in ΨwzpAθq and has principal symbol ρwpξqz.
P P is in ΨwpAθq and has principal symbol ρwpξq.
Proof. By Proposition 3.20 and Proposition 3.21 the operator P P is in Ψ2wpAθq and has principal
symbol ρwpξqρwpξq. In particular, its principal symbol is positive and invertible, i.e., P P is
elliptic and it satisfies (P1). It also satisfies (P2), since P P is formally selfadjoint and pP P uuq "
a
pP uP uq ě 0 for all u P Aθ, which implies that the eigenvalues of P P must be non-negative.
Therefore, by Theorem 7.3 the absolute value P :" ?
P P is in ΨwpAθq and has principal symbol
ρwpξqρwpξq " ρwpξq. Note that P satisfies (P1) and (P2) as well, and so by Theorem 7.3
again, for every z P C, the operator Pz is in ΨwzpAθq and has principal symbol ρwpξqz. The
(cid:3)
proof is complete.
Appendix A. Improper Integrals with Values in Locally Convex Spaces
some locally convex space.
In this appendix we gather a few facts about convergence of improper integrals with values in
ş
In what follows we let E be a locally convex space. Given any continuous map f : ra, bs Ñ E its
ş8
a fpλqdλ is defined in the same way as the Riemann integral of scalar-valued
a fpλqdλ converges in E
Given a continuous map f : ra,8q Ñ E we say that the integral
Riemann integral
functions (see, e.g., [11, 25]; see also [22, Appendix B]).
b
b
ş
a fpλqdλ exists. In this case we set
fpλqdλ " lim
bÑ8
41
ż 8
a
ż
b
fpλqdλ.
a
when limbÑ8
ş8
Proposition A.1. Suppose that T : E Ñ F is a continuous linear map with values in some locally
a fpλqdλ converges in
convex space F . Let f : ra,8q Ñ E be a continuous map whose integral
E . Then the integral
"
‰
ż 8
ş8
¯
a Trfpλqsdλ converges in F , and we have
fpλqdλ
¯
´ż 8
dλ " T
´ż
fpλq
a
a
b
"
‰
fpλq
ş
ş
Proof. The composition Trfpλqs is a continuous map from ra,8q to F . Moreover, we have
ż
a fpλqdλq for 0 ă a ă b (see, e.g., [22, Proposition B.5]). Thus, in F we have
a Trfpλqsdλ " Tp
T
.
b
b
¯
´ż 8
dλ " lim
b
fpλqdλ
" T
fpλqdλ
.
a
a
a
T
ϕ
(A.1)
bÑ8 T
dλ " ϕ
ż 8
lim
bÑ8
ş8
(cid:3)
This proves the result.
Remark A.2. In the special case E " C the above results shows that, for every ϕ P E 1, the integral
"
‰
0 ϕrfpλqsdλ is convergent, and we have
fpλq
´ż 8
¯
ş8
fpλqdλ
ş8
a fpλqdλ.
ş8
Note that this property uniquely determines the value of
a fpλqdλ converges in E .
Proposition A.3. Let f : ra,8q Ñ E be a continuous map such that
ż 8
Let φ : ra1,8q Ñ ra,8q be a C 1-diffeomorphism. Then
a1 φ1pλqpf φqpλqdλ converges in E , and
we have
φ1pλqpf φqpλqdλ "
fpλqdλ.
ż 8
Proof. The assumption that φ is a diffeomorphism ensures us that φ is increasing, φpa1q " a, and
limλÑ8 φpλq " 8. Therefore, by using [22, Proposition B.7] we get
fpλqdλ
φ1pλqpf φqpλqdλ "
fpλqdλ ÝÑ
as b Ñ 8.
ż 8
φpbq
ż
ż
a1
a
a
a
.
b
a1
This proves the result.
a
a
(cid:3)
It is convenient to interpret the above notion of convergence in term of Lebesgue integrations
of maps with values in locally convex spaces. A natural setting for this notion of integration is the
setting of quasi-complete locally convex Suslin spaces (see [45]; see also [22, Appendix B]). This
encompasses a large class of locally convex spaces, including separable Fr´echet spaces such as Aθ.
However, in the precise setting of this paper, we can bypass this as follows.
In what follows, given any Borel set Y Ă C, we say that a measurable map f : Y Ñ E is
"
absolutely integrable when
fpλq
(A.2)
p
‰
dλ ă 8
ż
for every continuous semi-norm p on E .
Y
Recall that a locally convex space is quasi-complete when every bounded Cauchy net is con-
vergent. For instance, Fr´echet spaces and their weak duals, as well as inductive limits of Fr´echet
spaces are quasi-complete (see, e.g., [46]).
Proposition A.4. Assume that E is quasi-complete. Then, for every absolutely integrable con-
tinuous map f : ra,8q Ñ E , the integral
Proof. Let f : ra,8q Ñ E be an absolutely integrable continuous map. Let p be a continuous
ż
ı
semi-norm on E . Given any b ě a we have
fpλqdλ
"ż
ı
Similarly, given any c ě 0, we have
fpλqdλ
ş8
a fpλqdλ converges in E .
‰
"
fpλq
fpλq
"
fpλq
dλ ă 8.
‰
dλ ď
ż 8
ż
‰
fpλq
dλ ď
ı
"ż
ż
fpλqdλ ´
ż 8
fpλqdλ
"ż
" p
b`c
b`c
b`c
"
"
‰
ď
ď
dλ.
p
p
p
p
p
p
a
a
a
b
b
b
a
a
b
42
b
b
ş8
"
‰
fpλq
dλ Ñ 0 as b Ñ 8. Therefore,
b p
The absolute-integrability condition (A.2) ensures us that
we see that, for every continuous semi-norm p on E , we have
ÝÑ 0
fpλqdλ
b`c
b
ı
"ż
ż
a
p
sup
cě0
fpλqdλ ´
ş
ş8
ş
a fpλqdλ; b ě au is a bounded Cauchy net in E . As E is quasi-complete, it then
a fpλqdλ converges in E . The
(cid:3)
a fpλqdλ exists in E . That is, the integral
as b Ñ 8.
a
b
b
This shows that t
follows that limbÑ8
proof is complete.
As an immediate consequence we have the following dominated convergence result.
"
‰
fpλq
Proposition A.5. Suppose that E is quasi-complete. Let f : ra,8q Ñ E be a continuous map
such that, for every continuous semi-norm p on E , there is a function gppλq P L1ra,8q such that
ď gppλq
Then fpλq is absolutely integrable, and the integral
ş8
a fpλqdλ converges in E .
ş
ş
The above considerations extend verbatim to integrals along rays Lφpaq " eiφra,8q with a ě 0
Lφpaq fpλqdλ converges
a fpeiφλqdpeiφλq exists in E . In this case, and if Lφpaq is outward-oriented
and 0 ď φ ă 2π. If f : Lφpaq Ñ E is a continuous map, then the integral
in E when limbÑ8
(i.e., it is directed toward 8), then we set
@λ ě a.
p
b
ż
ż
a
There is a change of sign when Lφpaq is inward-oriented.
fpλqdλ " lim
bÑ8
Lφpaq
b
fpeiφλqdpeiφλq.
More generally, we can consider keyhole contours, by which we mean contours of the form,
Γ " Lφ1paq Y Cφ1,φ2paq Y Lφ2paq,
a ě 0, φ1 ą φ2 ě φ1 ´ 2π,
where Cφ1,φ2paq is the arc of circle taeiφ; φ1 ě φ ě φ2u. The orientation of Γ is chosen so that
it agrees with the clockwise-orientation on Cφ1,φ2paq. Thus, Lφ1paq is inward-oriented, whereas
Lφ2paq is outward-oriented. For instance when φ1 " π, φ2 " 0 and a " 0 the contour Γ is just the
real line p´8,8q.
when the integrals
Figure 4. Examples of Keyhole Contours
Given a continuous map f : Γ Ñ E we say that the contour integral
ż
ş
Lφ1paq fpλqdλ and
fpλqdλ "
ş
fpλqdλ `
ż
ż
Lφ2paq
where the middle integral in the r.h.s. is defined as a Riemann integral.
Cφ1,φ2paq
Lφ1paq
Γ
ż
fpλqdλ `
fpλqdλ,
Lφ2paq fpλqdλ both converge in E . In this case we set
ş
Γ fpλqdλ converges in E
43
At the exception of Proposition A.3 all the previous results of this section hold verbatim for
ş
this kind of improper integrals. We also have the following version of Proposition A.3.
Proposition A.6. Suppose that Γ is a keyhole contour, and let f : Γ Ñ E be a continuous
tΓ fpt´1λqdλ
map whose integral
converges in E , and we have
Γ fpλqdλ converges in E. Then, for every t ą 0, the integral
ş
ż
tΓ
ż
Γ
fpt´1λqdλ " t
fpλqdλ.
We are mostly interested in the case where fpλq is an HoldpΛq-map, where Λ is some open
pseudo-cone containing the contour Γ. Note that keyhole contours themselves are pseudo-cones.
Proposition A.7. Assume that E is quasi-complete. Let fpλq P HoldpΛ; Eq, d ă ´1, where Λ is
some open pseudo-cone. In addition, let Γ be a keyhole contour contained in Λ.
(1) The contour integral
(2) Let T : E Ñ F be a continuous linear map with values in some other locally convex space
ş
ş
Γ fpλqdλ converges in E .
´ż
ż
Γ Trfpλqsdλ converges in F , and we have
¯
F . Then, the integral
(3) Let Γ1 Ă Λ be either a keyhole contour or a clockwise-oriented Jordan curve, and assume
there is a domain Ω Ă Λ such that BΩ " Γ Y Γ1. Then, we have
‰
fpλq
T
Γ
"
ż
dλ " T
ż
fpλqdλ
.
Γ
fpλqdλ.
Γ1
Γ
fpλqdλ "
‰
fpλq
ď p1 ` λqd
"
p
@λ P Γ.
Proof. As Γ is a pseudo-cone ĂĂ Λ, for every continuous semi-norm p on E , there is CΓp ą 0 such
that
As d ă ´1 and E is quasi-complete, it follows from Proposition A.5 that fpλq is absolutely
Γ fpλqdλ converges in E . This proves the first part. Combining it
integrable and the integral
with Proposition A.1 gives the 2nd part.
It remains to prove the 3rd part. Let Γ1 Ă Λ be either a keyhole contour or a clockwise-oriented
Jordan curve, and assume there is a domain Ω Ă Λ such that BΩ " ΓY Γ1. In addition, let ϕ P E 1.
The composition ϕfpλq is contained in HoldpΛq, and so this is an integrable holomorphic function
on Ω. Therefore, if we orient BΩ suitably, then we get
0 "
ϕrfpλqs dλ "
ϕrfpλqs dλ.
Combining this with (A.1) we obtain
fpλqdλ
ϕ
" ϕ
As E1 separates the point of E by Hahn-Banach theorem, we see that
proves the 3rd part. The proof is complete.
ş
@ϕ P E 1.
Γ fpλqdλ "
ş
Γ1 fpλqdλ. This
(cid:3)
ż
Γ1
ż
´ż
Γ
ϕrfpλqs ´
¯
fpλqdλ
Γ1
¯
ş
ż
´ż
BΩ
Γ
In this Appendix, we include a proof of Proposition 5.29. Recall that the topology of L pA 1
θ , Aθq
Appendix B. Proof of Proposition 5.29
is generated by the semi-norms,
R ÝÑ sup
uPB
B Ă A 1
We have the following description of bounded sets in A 1
θ .
Lemma B.1. A subset B Ă A 1
contained and bounded in H psq
.
θ is bounded in A 1
}δαpRuq},
θ
44
θ bounded, α P Nn
0 .
θ if and only if there s P R such that B is
θ
θ
θ
θ
into A 1
Conversely, let B be a bounded set of A 1
is a neighborhood of the origin in H psq
Bearing this in mind, a subset B Ă A 1
θ agrees with the inductive limit of the H psq
θ is bounded when, given any neigborhood U of the
θ is
for every s P R.
θ , there is t ą 0 such that tB Ă U . As the inclusion of H psq
θ
is bounded in A 1
θ .
Proof. This a a special case of a general result for inductive limits of compact inclusions of Banach
spaces, but we shall give a proof for reader's convenience. By Proposition 3.27 the strong topology
of A 1
-topologies. Equivalently, this is the strongest
locally convex topology with respect to which the inclusion of H psq
θ is continuous for every
s P R. Thus, a basis of neighborhoods of the origin in A 1
θ consists of all convex balanced sets U
such that U X H psq
origin in A 1
continuous, it is immediate that any bounded set of H psq
θ . With a view toward contradiction assume that B
is not a bounded set in any H psq
. Given s P R and r ą 0 let us denote by Bpsqprq the open
ř
ball in H psq
of radius r about the origin. The above assumption means that, for every s P R, we
cannot find r ą 0 such that B Ă Bpsqprq. In particular, there is up0q P B such that up0q R Bp0qp1q,
θ
ř
ř
p0q
p0q
u
k 2 ą 1. Likewise, there is
k 2 ą 1, and so there is an integer N0 such that
kďN0 u
ř
i.e.,
p1q
p0q
k 2, and so there is also an integer N1 such that
up1q P B such that
k 2 ą 2
kďN0 u
p0q
p1q
kďN1xky´2u
k 2. Repeating this argument allows us to construct sequences
k 2 ą 2
ÿ
ÿ
pN(cid:96)q(cid:96)ě0 Ă N0 and pup(cid:96)qq(cid:96)ě0 Ă B such that
(B.1)
, s P R, into A 1
2 ą 2
xky´2(cid:96)
u
ř
2
θ
θ
p(cid:96)q
k
p(cid:96)q
k 2, and define
kďN(cid:96)
ÿ
for all (cid:96) ě 0.
*
ř
xky´2u
kďN0 u
xky´2p(cid:96)`1qu
ř
"
č
xky´2(cid:96)u
kďN(cid:96)
u P A 1
U "
θ ;
p(cid:96)`1q
k
kďN(cid:96)`1
For (cid:96) ě 0 set µ(cid:96) " 2´(cid:96)
.
(cid:96)ě0
kďN(cid:96)
kďN(cid:96)
2 , u P A 1
xky´2(cid:96)uk2q 1
xky´2(cid:96)uk2 ă µ(cid:96)
Note that (B.1) ensures us that pµ(cid:96)q(cid:96)ě0 is an increasing sequence.
ř
Claim. U is a neighborhood of the origin in A 1
θ .
Proof of the Claim. For (cid:96) ě 0 set p(cid:96)puq " p
θ . This defines semi-norms
on A 1
θ . In particular, we see that U is a convex balanced set. Therefore, in order to prove the
is a neighborhood of the origin in H psq
claim it is enough to show that U X H psq
for every s P R.
In fact, as the inclusion of H psq
into H ps1q
is continuous for s ą s1, it is enough to show this for
s " ´(cid:96)0 with (cid:96)0 P N.
ÿ
?
µ(cid:96)0q. We observe that if u P Bp´(cid:96)0qpq and (cid:96) ě (cid:96)0, then
xky´2(cid:96)uk2 ď
!
č
θ ; p(cid:96)puq ă ?
µ(cid:96)u for (cid:96) ě (cid:96)0. Thus,
u P Bp´(cid:96)0qpq; p(cid:96)puq ă ?
, it then follows that U X Bp´(cid:96)0qpq is an open set
for every (cid:96)0 P N. The
(cid:3)
As the semi-norms p(cid:96) are continuous on H p´(cid:96)0q
of H p´(cid:96)0q
proof is complete.
xky´2(cid:96)0uk2 ă 2 ă µ(cid:96)0 ď µ(cid:96).
)
ÿ
Given (cid:96)0 P N, let P p0,
is a neighborhood of the origin in H p´(cid:96)0q
This shows that Bp´(cid:96)0qpq Ă tu P A 1
, and so U X H p´(cid:96)0q
U X Bp´(cid:96)0qpq "
kďN(cid:96)
kPZn
(cid:96)ă(cid:96)0
µ(cid:96)
.
θ
θ
θ
θ
θ
θ
θ
θ
The above claim leads us to a contradiction as follows. As B is bounded in A 1
θ and U is a
θ , there is t ą 0 such that tB Ă U . In particular, tup(cid:96)q P U for
neighborhood of the origin in A 1
all (cid:96) ě 0. Now, choose (cid:96) so that 2´(cid:96) ď t2. Then we have
xky´2(cid:96)t2u
p(cid:96)q
k 2 ě
2´(cid:96)xky´2(cid:96)u
p(cid:96)q
k 2 " µ(cid:96).
ÿ
kďN(cid:96)
ÿ
kďN(cid:96)
45
Thus, tup(cid:96)q cannot be contained in U . This is a contradiction. Therefore, if B is a bounded set
of A 1
(cid:3)
for some s P R. This proves the result.
θ , then it must be a bounded set of H psq
θ
We are now in a position to prove Proposition 5.29.
θ
Proof of Proposition 5.29. Given s, t P R, the continuity of the inclusions H psq
H ptq
implies that the natural embedding of L pA 1
so } }s,t is a continuous semi-norm on L pA 1
θ , Aθq.
norms } }t, t P R. Therefore, the topology of L pA 1
B Ă A 1
θ , Aθq is generated by the semi-norms.
θ bounded,
Moreover, we know by Proposition 3.27 that the topology of Aθ is generated by the Sobolev
θ Ă A 1
θ and Aθ Ă
q is continuous, and
θ , Aθq into L pH psq
}Ru}t,
, H ptq
θ
θ
R ÝÑ sup
uPB
t P R.
θ . By Lemma B.1 there is s P R
; }u}s ď u.
Bearing this in mind, let t P R and let B be a bounded subset of A 1
such that B is a bounded subset of H psq
Thus, for all R P L pA 1
, i.e., there is ą 0 such that B Ă tu P H psq
θ
θ
θ , Aθq, we have
}Ru}t ď sup
sup
uPB
uPH psq
}u}sď
θ
}Ru}t ď
}Ru}t " }R}s,t.
sup
uPH psq
}u}sď1
θ
All this ensures us that the norms }}s,t, s, t P R, generate the topology of L pA 1
is complete.
θ , Aθq. The proof
(cid:3)
References
[1] Arveson, W.: An invitation to C*-algebras. Springer-Verlag, 1981.
[2] Baaj, S.: Calcul pseudo-diff´erentiel et produits crois´es de C-alg´ebres. I, II. C. R. Acad. Sc. Paris, s´er. I,
307 (1988), 581 -- 586, 663 -- 666.
[3] Connes, A.: C*-alg`ebres et g´eom´etrie differentielle. C. R. Acad. Sc. Paris, s´er. A, 290 (1980), 599 -- 604.
[4] Connes, A.: An analogue of the Thom isomorphism for crossed products of a C-algebra by an action of
R. Adv. Math. 39 (1981), 31 -- 55.
[5] Connes, A.: Noncommutative geometry. Academic Press, Inc., San Diego, CA, 1994.
[6] Connes, A.; Fathizadeh, F.: The term a4 in the heat kernel expansion of noncommutative tori. Munster
J. Math. 12 (2019), 239 -- 410.
[7] Connes, A.; Moscovici, H.: Modular curvature for noncommutative two-tori. J. Amer. Math. Soc. 27
(2014), no. 3, 639 -- 684.
[8] Connes, A.; Tretkoff, P.: The Gauss-Bonnet theorem for the noncommutative two torus. Noncommutative
geometry, arithmetic, and related topics, pp. 141 -- 158, Johns Hopkins Univ. Press, Baltimore, MD, 2011.
[9] Dabrowski, L.; Sitarz, A.: An asymmetric noncommutative torus. SIGMA Symmetry Integrability Geom.
Methods Appl. 11 (2015), Paper 075, 11pp..
[10] Dyatlov, S.; Zworski, M.: Mathematical theory of scattering resonances. Graduate Studies in Mathematics,
Vol. 200, AMS, 2019, 634 pp..
[11] Falb, P.L.; Jacobs, M.Q.: On differentials in locally convex spaces. J. Differential Equations 4 (1968),
444 -- 459.
[12] Fathi, A.; Ghorbanpour, M.; Khalkhali, M.: The curvature of the determinant line bundle on the non-
commutative two torus. J. Math. Phys. Anal. Geom. 20 (2017), 1 -- 20.
[13] Fathizadeh, F.; Khalkhali, M.: The Gauss-Bonnet theorem for noncommutative two tori with a general
conformal structure. J. Noncommut. Geom. 6 (2012), no. 3, 457 -- 480.
[14] Fathizadeh, F.; Khalkhali, M.: Scalar curvature for noncommutative four-tori. J. Noncommut. Geom. 9
(2015), no. 2, 473 -- 503.
[15] Floricel, R.; Ghorbanpour, A.; Khalkhali, M.: The Ricci curvature in noncommutative geometry. Preprint
arXiv:1612.06688, 26 pp.. To appear in J. Noncommut. Geom..
[16] Gilkey, P.B.: Invariance theory, the heat equation, and the Atiyah-Singer index theorem. 2nd edition.
Studies in Advanced Mathematics. CRC Press, Boca Raton, FL, 1995.
[17] Gohberg, I.C.; Krein, M.G.: Introduction to the theory of linear nonselfadjoint operators. Translations of
Mathematical Monographs, Vol. 18. American Mathematical Society, Providence, R.I., 1969.
[18] Gonz´alez-P´erez, A.M.; Junge, M.; Parcet, J.: Singular integrals in quantum Euclidean spaces. Preprint
arXiv:1705.01081, 87 pp.. To appear in Mem. Amer. Math. Soc..
[19] Grubb, G.: Functional calculus of pseudodifferential boundary problems. Second edition. Progress in Math-
ematics, 65. Birkhauser Boston, Inc., Boston, MA, 1996. x+522 pp.
[20] Grubb, G.; Seeley, R.: Weakly parametric pseudodifferential operators and Atiyah-Patodi-Singer boundary
problems. Invent. Math. 121 (1995), 481 -- 530.
46
[21] Grubb, G.; Seeley, R.: Zeta and eta functions for Atiyah-Patodi-Singer operators. J. Geom. Anal. 6
(1996), 31 -- 77.
[22] Ha, H.; Lee, G.; Ponge, R.: Pseudodifferential calculus on noncommutative tori, I. Oscillating integrals.
Int. J. Math. 30 (2019), 1950033 (74 pages).
[23] Ha, H.; Lee, G.; Ponge, R.: Pseudodifferential calculus on noncommutative tori, II. Main properties. Int.
J. Math. 30 (2019), 1950034 (73 pages).
[24] Ha, H.; Ponge, R.: Laplace-Beltrami operators on noncommutative tori. Preprint arXiv:1905.09048, 28
pages.
[25] Hamilton, R.S.: The inverse function theorem of Nash and Moser. Bull. Trans. Amer. Math. Soc. 7 (1982),
65 -- 222.
[26] Hormander, L.: The analysis of linear partial differential operators III. Springer-Verlag, Berlin-Heidelberg-
New York-Tokyo, 1985.
[27] Lee, G.; Ponge, R.: Complex powers of elliptic operators on noncommutative tori. In preparation.
[28] Lesch, M.; Moscovici, H.: Modular curvature and Morita equivalence. Geom. Funct. Anal. 26 (2016), no.
3, 818 -- 873.
[29] L´evy, C.; Neira-Jim´enez, C.; Paycha, S.: The canonical trace and the noncommutative residue on the
noncommutative torus. Trans. Amer. Math. Soc. 368 (2016), no. 2, 1051 -- 1095.
[30] Liu, Y.: Hypergeometric function and modular curvature II. Connes-Moscovici functional relation after
Lesch's work. Preprint, arXiv:1811.07967, 29 pp..
[31] Polishchuk, A.: Analogues of the exponential map associated with complex structures on noncommutative
two-tori. Pacific J. Math. 226 (2006), 153 -- 178.
[32] Ponge, R.: Calcul hypoelliptique sur les vari´et´es de Heisenberg, r´esidu non commutatif et g´eom´etrie
pseudo-hermitienne. PhD Thesis, Univ. Paris-Sud (Orsay), December 2000.
[33] Ponge, R.: Calcul fonctionnel sous elliptique et r´esidu non commutatif pour les vari´et´es de Heisenberg.
C. R. Acad. Sc. Paris. s´er. I, 332 (2001), 611-614.
[34] Ponge, R.: Heat kernel asymptotics for elliptic operators on noncommutative tori. In preparation.
[35] Rieffel, M.: C-algebras associated with irrational rotations. Pacific J. Math. 93 (1981), 415 -- 429.
[36] Rieffel, M.: Noncommutative tori-A case study of non-commutative differentiable manifolds. Contemp.
Math., 105, Amer. Math. Soc., Providence, RI, 1990, pp. 191 -- 211.
[37] Rieffel, M.: Deformation quantization for actions of Rd. Mem. Amer. Math. Soc. 106 (1993) no. 506,
x+98 pp..
[38] Rosenberg, J.: Levi-Civita's theorem for noncommutative tori. SIGMA 9 (2013), 071, 9 pages.
[39] Ruzhansky, M.: Turunen, V.: Pseudodifferential operators and symmetries. Background analysis and
advanced topics. Pseudo-differential operators. Theory and applications, 2, Birkhauser, Basel, 2010.
[40] Seeley, R.T.: Complex powers of an elliptic operator. Proc. Sympos. Pure Math., Vol. X, pp. 288 -- 307.
American Mathematical Society, Providence, RI, 1967.
[41] Shubin, M.A.: Pseudodifferential operators and spectral theory, 2nd edition. Springer, Berlin, 2001.
[42] Simon, B.: Trace ideals and their applications. Second edition. Mathematical Surveys and Monographs,
120. American Mathematical Society, Providence, RI, 2005.
[43] Spera, M.: Sobolev theory for noncommutative tori. Rend. Sem. Mat. Univ. Padova 86 (1992), 143 -- 156.
[44] Tao, J.: The theory of pseudo-differential operators on the noncommutative n-torus. J. Phys.: Conf. Ser.
965 012042 (2018), 1 -- 12.
[45] Thomas, G.E.F.: Integration of functions with values in locally convex Suslin spaces. Trans. Amer. Math.
Soc. 212 (1975), 61 -- 81.
[46] Tr`eves, F.: Topological vector spaces, distributions and kernels. Academic Press, New York-London, 1967.
[47] Xiong, X.; Xu, Q.; Yin, Z.: Sobolev, Besov and Triebel-Lizorkin spaces on quantum tori. Mem. Amer.
Math. Soc. 252 (2018) no. 1203, 86 pages.
Department of Mathematical Sciences, Seoul National University, Seoul, South Korea
E-mail address: [email protected]
School of Mathematics, Sichuan University, Chengdu, China
E-mail address: [email protected]
47
|
1307.4201 | 1 | 1307 | 2013-07-16T08:56:29 | Effect algebras with state operator | [
"math.OA"
] | State operators on convex effect algebras, in particular effect algebras of unital JC-algebras, JW-algebras and convex sigma-MV algebras are studied and their relations with conditional expectations in algebraic sense as well as in the sense of probability on MV-algebras are shown. | math.OA | math |
EFFECT ALGEBRAS WITH STATE OPERATOR
A. JEN COV ´A, S. PULMANNOV ´A
Abstract. State operators on convex effect algebras, in particular effect alge-
bras of unital JC-algebras, JW-algebras and convex σ-MV algebras are studied
and their relations with conditional expectations in algebraic sense as well as
in the sense of probability on MV-algebras are shown.
1. Introduction
Effect algebras have been introduced by Foulis and Bennett [18] (see also [19, 32]
for equivalent definitions) for modeling unsharp measurements in quantum mechan-
ical systems [6]. They are a generalization of many structures which arise in the
axiomatization of quantum mechanics (Hilbert space effects [34]), noncommutative
measure theory and probability (orthomodular lattices and posets, [1, 44]), fuzzy
measure theory and many-valued logic (MV-algebras [9, 8]).
A state, as an analogue of a probability measure, is a basic notion in algebraic
structures used in the quantum theories (see e.g., [13]), and properties of states
have been deeply studied by many authors.
In MV-algebras, states as averaging the truth value were first studied in [41].
In the last few years, the notion of a state has been studied by many experts in
MV-algebras, e.g, [46, 31].
Another approach to the state theory on MV-algebras has been presented re-
cently in [17]. Namely, a new unary operator was added to the MV-algebras struc-
ture as an internal state (or so-called state operator). MV-algebras with the added
state operator are called state MV-algebras. The idea is that an internal state
has some properties reminiscent of states, but, while a state is a map from an
MV-algebra into [0, 1], an internal state is an operator of the algebra. State MV-
algebras generalize, for example, H´ajek's approach [27] to fuzzy logic with modality
P r (interpreted as probably) with the following semantic interpretation: The prob-
ability of an event a is presented as the truth value of P r(a). For a more detailed
motivation of state MV-algebras and their relation to logic, see [17].
In [5], the notion of a state operator was extended from MV-algebras to the more
general frame of effect algebras. A state operator is there defined as an additive,
unital and idempotent operator on E. A state operator on E is called strong, if it
satisfies the additional condition
(1)
τ (τ (a) ∧ τ (b)) = τ (a) ∧ τ (b) whenever τ (a) ∧ τ (b) exists in E.
1991 Mathematics Subject Classification. Primary 81P10, 46L53, Secondary 17C50, 47C15.
Key words and phrases. effect algebra, convex effect algebra, MV algebra, state operator,
JC-effect algebra, conditional expectation.
This work was supported by grant VEGA 2/0059/12 and by Science and Technology Assistance
Agency under the contract no. APVV-0178-11.
1
2
A. JEN COV ´A, S. PULMANNOV ´A
Since MV-algebras form a special subclass of effect algebras, so-called MV-effect
algebras, it was shown that the definition of a state operator on effect algebras
coincides with the original definition on MV-algebras if and only if the state op-
erator is strong. Moreover, if τ is faithful, i.e., τ (a) = 0 implies a = 0, then it is
automatically strong.
In the present paper, we show that state operators on an effect algebra E are
related with states on E in the following way: (1) every state on E induces a state
operator on the tensor product [0, 1] ⊗ E. (2) If E admits an ordering set of states,
then every state operator on E induces a state on E. We study state operators
mainly on convex effect algebras. Convex effect algebras as effect algebras with
an additional convexity structure were introduced and studied in [24, 3, 4]. It was
proved in [24], that every convex effect algebra is isomorphic with the interval [0, u]
in an ordered real linear space (V, V +), where u is an order unit. Moreover, (V, u)
is an order unit space, i.e. u is an archimedean order unit, if and only if E admits
an ordering set of states [3]. Using the tensor product of an effect algebra with
the interval [0, 1] of reals, we show that every effect algebra E admitting at least
one state, can be embedded into a convex one. Moreover, a state operator on E
extends to a state operator on its convex envelope. It is therefore not too restrictive
to concentrate our interest on convex effect algebras with state operators. We show
that a state operator on a convex effect algebra is an affine mapping which extends
to a linear, positive and idempotent mapping of the corresponding ordered linear
space into itself. Moreover, the state operator is faithful if and only if its extension
is faithful.
A prototype of an effect algebra is the set of Hilbert space effects E(H), i.e., self-
adjoint operators between the zero and identity operator on a (complex) Hilbert
space H with respect to the usual ordering of self-adjoint operators. The partial
operation ⊕ is defined as the usual operator sum of two effects whenever this sum
is also an effect, and the effect algebra ordering coincides with the original one we
started with. This effect algebra is naturally convex. It plays an important role
in quantum measurement theory, because the most general quantum observables,
positive operator valued measures (POVMs), have their ranges in it [6]. On the
set B(H) of the bounded operators on H, we consider the von Neumann [43] and
Luders [38] conditional expectations, and we show that their restrictions to E(H)
are faithful, hence strong state operators. Motivated by these examples, we study
relations between state operators and conditional expectations on so-called JC-
effect algebras.
Recall that a JC-algebra J is a norm-closed real vector subspace of bounded self-
adjoint operators on a Hilbert space H, closed under the Jordan product a ◦ b =
1
2 (ab + ba). A JC-algebra is called a JW-algebra if it is closed in the weak topology
[49]. We study JC-algebras containing a unit element 1, and the interval [0, 1] is
then called a JC-effect algebra. Let τ : E(J ) → E(J ) be a state operator. Since
E(J ) is the [0, 1] interval in the ordered vector space (J , J +), τ extends to a linear,
positive idempotent and unital mapping τ : J → J . Such maps were studied in
several papers, we use mostly the results in [15].
We call a state operator τ : E(J ) → E(J ) a conditional expectation iff it has
the property τ (τ (a)bτ (a)) = τ (a)τ (b)τ (a) for all a, b ∈ E(J ). We show that a
state operator τ on E(J ) is a conditional expectation iff its range is a JC-sub-
effect algebra of E(J ), and any faithful state operator is a conditional expectation.
EFFECT ALGEBRAS WITH STATE OPERATOR
3
Moreover, a state operator can be expressed as a combination of a conditional
expectation and so-called Jordan state operator. We also show that if the JC-effect
algebra is the unit interval in C(X; R), the set of real-valued continuous functions on
a compact Hausdorff space X, then a state τ operator is a conditional expectation
iff τ is strong.
In the probability theory on MV-algebras, an MV-conditional expectation on a
σ-MV-algebra M with respect to a σ-MV-subalgebra N of M in a σ-additive state
m on M was introduced in [14]. We study relations between the MV-conditional
expectations and state operators on convex σ-MV-algebras. Since a convex σ-MV-
algebra M is isomorphic with the interval [0, 1] in C(X; R) ([13, Theorem 7.3.12],
a state operator τ on M is a conditional expectation iff τ is strong.
We show that an MV-conditional expectation induces a strong state operator on
the quotient of M with respect to the kernel of the state m.
On the other hand, a strong σ-additive state operator (hence a conditional ex-
pectation) τ on M induces an MV-conditional expectation on M with respect to
the sub-MV-algebra equal to the range of τ and in states of the form s ◦ τ , where
s is any σ-additive state.
2. Basic definitions
Definition 2.1. An effect algebra (EA, for short) is a structure (E; 0, 1, ⊕) con-
sisting of a set E; elements 0 and 1 in E called the zero and the unit; and a
partially defined binary operation ⊕ on E called the orthosummation, such that,
for all d, e, f ∈ E:
(EA1) If e ⊕ f is defined, then f ⊕ e is defined and e ⊕ f = f ⊕ e.
(EA2) If d ⊕ e and (d ⊕ e) ⊕ f are defined, then e ⊕ f and d ⊕ (e ⊕ f ) are defined
and (d ⊕ e) ⊕ f = d ⊕ (e ⊕ f ).
(EA3) For each e ∈ E there exists a unique element e⊥ ∈ E, called the orthosup-
plement of e, such that e ⊕ e⊥ = 1.1
(EA4) e ⊕ 1 is defined only if e = 0.
In an effect algebra E, a partial ordering is defined by a ≤ b iff there is c ∈ E such
that a ⊕ c = b. It turns out that the element c, if it exists, is uniquely defined. This
enables us to introduce the partial operation ⊖ by b ⊖ a := c 2 iff a ⊕ c = b. Clearly,
b ⊖ a is defined iff a ≤ b. With respect to this partial order we have 0 ≤ a ≤ 1 for
all a ∈ E.
Effect algebras were introduced in [18]. Notice that effect algebras are equivalent
to D-posets [32], which are partial algebraic structures based on the operation ⊖
[13].
The operation ⊕ can be extended to suitable sequences a1, a2, . . . , an of elements
(not necessarily all different) by recurrence. We say that the elements a1, a2, . . . , an
are orthogonal iff their orthosum a1⊕a2⊕· · ·⊕an =: ⊕i≤nai is defined. An arbitrary
family of elements {aλ : λ ∈ I} in E is called orthogonal, iff every its finite subfamily
is orthogonal. Let F (I) denote the set of all finite subsets of the index set I. If the
element a := WF ∈F (I) ⊕i∈F ai exists, then a is called the orthosum of the orthogonal
1If we write an equation involving an orthosum e ⊕ f without explicitly stating that e ⊕ f is
defined, we are tacitly assuming that it is defined.
2The notation := means 'equals by definition'.
4
A. JEN COV ´A, S. PULMANNOV ´A
family {aλ : λ ∈ I}. The effect algebra E is called σ-orthocomplete (orthocomplete)
iff every countable (arbitrary) orthogonal family has an orthosum.
Example 2.2. An orthomodular lattice (L; ≤,⊥ , 0, 1) is organized into an EA by
defining a ⊕ b for a, b ∈ L iff a ≤ b⊥, in which case a ⊕ b := a ∨ b.
Conversely, an effect algebra E is an OML iff it is lattice ordered and a ≤ b⊥ =⇒
a ∧ b = 0.
Example 2.3. Recall that an MV-algebra is a system (M ; ⊞,′ , 0), where (M ; ⊞, 0)
is a commutative monoid with neutral element 0, and for each x, y ∈ M the following
equations hold:
(i) (x′)′ = x,
(ii) x ⊞ 1 = 1, where 1 = 0′,
(iii) x ⊞ (x ⊞ y′)′ = y ⊞ (y ⊞ x′)′.
In every MV-algebra one can define further operations as follows:
x ⊡ y = (x′ ⊞ y′)′, x ∨ y = (x′ ⊞ y)′ ⊞ y
x ∧ y = (x′ ∨ y′)′, x ⊟ y = (x′ ⊞ y)′.
A prototype of an MV-algebra is the unit interval [0, 1] ⊆ R, with operations
x ⊞ y = min{1, x + y}, and x′ = 1 − x. An MV-algebra becomes a boolean algebra
iff the following identity holds: a ⊞ a = a.
An MV-algebra (M ; ⊞,′ , 0) is organized into an EA by defining a⊕ b for a, b ∈ M
iff a ⊡ b = 0, in which case a ⊕ b := a ⊞ b.
Conversely, an EA E can be made an MV-algebra iff E is a lattice and
a ∧ b = 0 =⇒ a ≤ b⊥,
in which case E is called an MV-effect algebra. The MV-algebra operations are
defined by a ⊞ b := a ⊕ (a⊥ ∧ b) and a′ := a⊥.
MV-algebras and MV-effect algebras are in one-to one correspondence, and we
identify them.
Notice that Boolean algebras coincide with the subclass of MV-algebras satisfy-
ing the additional identity a ⊞ a = a.
Example 2.4. Let H be a (complex) Hilbert space. A (self-adjoint) operator
a : H → H is called an effect iff 0 ≤ a ≤ I (with respect to the usual ordering
of self-adjoint operators). Let E(H) denote the set of all effects on H, then E(H)
can be organized into an effect algebra with constants 0 and I as zero and unit,
respectively, by putting a ⊕ b is defined iff a + b is an effect, and in this case,
a ⊕ b = a + b. The effect algebra ordering coincides with the original one that we
used in the definition of effects.
Example 2.5. Let (G; ≤, +, 0) be an abelian partially ordered group. Then for
any u ∈ G, u ≥ 0, the interval G[0, u] := {x ∈ G : 0 ≤ x ≤ u} can be organized
into an effect algebra by defining a ⊕ b = a + b, provided that a + b ≤ u. The
effect algebra ordering coincides with the restriction of the partial order in G to the
interval G[0, u]. Effect algebras of this type are called interval effect algebras.
Notice that E(H) is an interval effect algebra.
Indeed, E(H) = B(H)sa[0, I],
where B(H)sa is the self-adjoint part of B(H).
Moreover, by [39], MV-algebras are equivalent with unit intervals in abelian
ℓ-groups with a strong unit.
EFFECT ALGEBRAS WITH STATE OPERATOR
5
On the other hand, there are examples of OMLs that are not interval effect
algebras [42, 51].
A subset F of an effect algebra E is a subeffect algebra of E iff (i) a ∈ F implies
a⊥ ∈ F , (ii) a, b ∈ F and a ⊕ b exists in E implies a ⊕ b ∈ F .
Let E be an effect algebra and a ∈ E. The interval E[0, a] := {x ∈ E : x ≤ a}
with the operation ⊕ restricted to E[0, a] (i.e., x ⊕a y exists if x ⊕ y exists in E and
x ⊕ y ≤ a) is an effect algebra with unit element a.
Let E, F be two EAs. A mapping φ : E → F is said to be:
(i) a morphism iff φ(1) = 1, and p ⊥ q, p, q ∈ E implies φ(p) ⊥ φ(q), and
φ(p ⊕ q) = φ(p) ⊕ φ(q);
(ii) a monomorphism iff φ is a morphism and φ(p) ⊥ φ(q) iff p ⊥ q;
(iii) an isomorphism iff φ is a surjective monomorphism, and we say that E is
isomorphic to F ;
(iv) a state iff φ is a morphism and F is the MV effect algebra [0, 1].
If E and F are σ-orthocomplete (orthocomplete) effect algebras, then a mor-
phism φ : E → F is a σ-morphism (complete morphism) iff φ(⊕i∈Nai) = ⊕i∈Nφ(ai),
whenever ⊕i∈Nai exists in E (iff φ(⊕i∈I ai) = ⊕i∈I φ(ai) whenever ⊕i∈I ai exists for
an arbitrary index set I).
Let E, F, L be EAs. A mapping β : E × F → L is called a bimorphism iff:
(i) a, b ∈ E with a ⊥ b, q ∈ F imply β(a, q) ⊥ β(b, q) and β(a ⊕ b, q) =
β(a, q) ⊕ β(b, q);
(ii) c, d ∈ F with c ⊥ d, p ∈ E imply β(p, c) ⊥ β(p, d) and β(p, c ⊕ d) =
β(p, c) ⊕ β(p, d);
(iii) β(1, 1) = 1.
If, in addition, E and F are MV-effect algebras, then also the following properties
are required:
(iv) a1, a2 ∈ E, b ∈ F imply β(a1 ∨ a2, b) = β(a1, b) ∨ β(a2, b), β(a1 ∧ a2) =
β(a1b) ∧ β(a2, b);
(v) a ∈ E, b1, b2 ∈ F imply β(a, b1 ∨ b2) = β(a, b1) ∨ β(a, b2), β(a, b1 ∧ b2) =
β(a, b1) ∧ β(a, b2).
Recall that a bimorphism β : P × Q → L, where P, Q and L are σ-effect algebras
(σ-MV-effect algebras), is a σ-bimorphism iff whenever ai ∈ P, bi ∈ Q are increasing
then β(W ai, b) = W β(ai, b) for every b ∈ Q, and β(a,W bi) = W(a, bi) for every
a ∈ P .
The following is a generalization of the definition of tensor product of effect
algebras in [10].
Definition 2.6. Let K be a class of effect algebras and E, F ∈ K. We say that a pair
(T, ρ) consisting of an effect algebra T ∈ K and a K- bimporphism ρ : E × F → T is
a tensor product of E and F in the class K iff the following condition is satisfied:
(T) If L ∈ K and β : E × F → L is a K- bimorphism, there exists a unique
K-morphism φ : T → L such that β = φ ◦ ρ.
The following theorem was proved in [12, Theorem 5.3].
Theorem 2.7. Let both effect algebras P and Q posses at least one state. Then
the tensor product P ⊗ Q of P and Q exists in the category of effect algebras. In
addition, for any state µ on P and any state ν on Q there is a unique state µ ⊗ ν
on P ⊗ Q such that µ ⊗ ν(p ⊗ q) = µ(p)ν(q), p ∈ P , q ∈ Q.
6
A. JEN COV ´A, S. PULMANNOV ´A
Tensor products in the category of σ-effect algebras were studied in [25]. Tensor
products in the category of MV-algebras (equivalently, MV-effect algebras) were
studied in [40].
Notice that for the unit interval [0, 1] ⊆ R, in the category of σ-effect algebras,
[0, 1] ⊗σ [0, 1] = [0, 1], where the bimorphism is given by (α, β) 7→ αβ [25], while in
the category of effect algebras, [0, 1] ⊗ [0, 1] 6= [0, 1] [45].
3. State operator
Recently, in [17], the notion of an MV-algebra with internal state, state MV-
algebra (SMV-algebra for short), was introduced as follows. An SMV-algebra is a
structure (M, σ) = (M ; ⊞,′ , σ, 0), where (M ; ⊞,′ , 0) is an MV-algebra, and σ is a
unary operator on M satisfying, for each x, y ∈ M :
(σ1) σ(0) = 0.
(σ2) σ(x′) = (σ(x))′.
(σ3) σ(x ⊞ y) = σ(x) ⊞ σ(y ⊟ (x ⊡ y)).
(σ4) σ(σ(x) ⊞ σ(y)) = σ(x) ⊞ σ(y).
In [5],the notion of an internal state (called also a state operator) was extended
to the more general frame of effect algebras.
Definition 3.1. A state operator on an EA E is a mapping τ : E → E such that,
for all e, f ∈ E,
(i) τ (1) = 1;
(ii) τ (e ⊕ f ) = τ (e) ⊕ τ (f );
(iii) τ (τ (a)) = τ (a).
The couple (E; τ ), where E is an effect algebra and τ is a state operator, will be
called a state effect algebra (SEA).
Lemma 3.2. For every SEA the following properties hold:
(i) τ (0) = 0.
(ii) τ (a⊥) = (τ (a))⊥.
(iii) a ≤ b =⇒ τ (a) ≤ τ (b), and τ (b ⊖ a) = τ (b) ⊖ τ (a).
(iv) If a ∧ b exists, then τ (a ∧ b) ≤ τ (a), τ (b), and if a ∨ b exists, then τ (a), τ (b) ≤
τ (a ∨ b).
(v) τ (E) is a subeffect algebra of E.
Lemma 3.3. [5] A state operator on an MV-effect algebra (E; ⊕, 0, 1) is a state
operator on the MV-algebra (E; ⊞,′ , 0) in the sense of [17] if and only if the following
additional condition is satisfied:
(2)
∃τ (e) ∧ τ (f ) =⇒ τ (τ (e) ∧ τ (f )) = τ (e) ∧ τ (f ).
Proof. Let (E; ⊞,′ , 0, 1) be the MV-algebra corresponding to the MV-effect algebra
(E; ⊕, 0, 1). Then τ (0) = 0, τ (a′) = τ (a)′ by Lemma 3.2. Moreover, for any
a, b ∈ E, τ (a ⊞ b) = τ (a ⊕ a⊥ ∧ b) = τ (a) ⊕ τ (a⊥ ∧ b) = τ (a) ⊞ τ ((a ∨ b′)′) =
τ (a) ⊞τ ((a′ ⊞b′)′ ⊞b′)′) = τ (a) ⊞τ (b ⊟(a⊙b)), as required. Finally, τ (τ (a) ⊞τ (b)) =
τ (τ (a) ⊕ τ (a)⊥ ∧ τ (b)) = τ (a) ⊕ τ ((τ (a⊥) ∧ τ (b)) = τ (a) ⊕ τ (a)⊥ ∧ τ (b) = τ (a) ⊞ τ (b).
The converse statement follows from the fact that if σ is an internal state on the
(cid:3)
MV-algebra (E; ⊞,′ , 0, 1), then the range of σ is MV-subalgebra of E.
In [5], a state operator τ satisfying property (2), is called a strong state operator.
Notice that an effect algebra need not be a lattice, in general, and so condition (2)
EFFECT ALGEBRAS WITH STATE OPERATOR
7
might be difficult to check. Nevertheless, the following lemma shows an important
situation in which this condition always holds.
Definition 3.4. A state operator τ on E is faithful if for any a ∈ E, τ (a) = 0
implies a = 0.
Lemma 3.5. [5] If τ be a faithful state operator on an effect algebra E, then τ is
strong.
Proof. Let a, b ∈ E and assume that c := τ (a) ∧ τ (b) exists in E. From c ≤ τ (a) we
get τ (c) ≤ τ (τ (a)) = τ (a), and similarly, τ (c) ≤ τ (b), hence τ (c) ≤ c. This entails
τ (c ⊖ τ (c)) = 0, and as τ is faithful, we get c = τ (c), which is (2).
(cid:3)
In the next theorem, we show that every state operator on an MV-effect algebra
M induces a faithful state operator on a quotient of M .
Recall that a subset I of an effect algebra E is an ideal iff (1) a ∈ I and b ≤ a
imply b ∈ I, and (2) a, b ∈ I and a ⊥ b imply a ⊕ b ∈ I. By [13, Theorem 3.1.32],
a subset I of an MV-effect algebra is an MV-algebra ideal iff it is an effect algebra
ideal. Consequently, the quotient M I := {[a] : a ∈ M }, where [a] denote the
congruence class containing a, a ∈ M , is an MV-effect algebra.
Theorem 3.6. Let τ be a state operator on an MV-effect algebra M . Put Iτ :=
{a ∈ M : τ (a) = 0}, then Iτ is an MV-algebra ideal, and τ : M Iτ → M Iτ defined
by τ [a] := [τ (a)] is a faithful state operator on M Iτ .
Proof. It is easy to check that Iτ is an effect algebra ideal, hence it is also an MV-
algebra ideal, and M Iτ is an MV-effect algebra. Define τ [a] := [τ (a)]. We have
a ∼ b iff a∆b ∈ Iτ , where a∆b = (a ∨ b) ⊖ (a ∧ b) is the symmetric difference [13,
Theorem 2.2.21 (v)]. Hence a ∼ b iff τ (a∆b) = 0 and τ (a∨b) ≥ τ (a), τ (b) ≥ τ (a∧b)
yields τ (a) = τ (b). This proves that τ is well defined. Clearly, τ : M Iτ → M Iτ ,
and τ [1] = [τ (1)] = [1]. Moreover, [a] ⊥ [b] iff there are a1 ∼ a, b1 ∼ b with a1 ⊥ b1
and [a] ⊕ [b] = [a1 ⊕ b1]. Therefore, τ ([a] ⊕ [b]) = τ ([a1 ⊕ b1]) = [τ (a1 ⊕ b1]) =
[τ (a1)]⊕[τ (b1)] = τ ([a])⊕ τ ([b]), so that τ is additive. Finally, τ (τ [a]) = τ ([τ (a)]) =
[τ (τ (a))] = [τ (a)] = τ [a], hence τ is idempotent, and hence a state operator.
Moreover, if τ ([a]) = 0, then [τ (a)] = 0, hence τ (a) ∈ I. But then τ (τ (a)) =
(cid:3)
τ (a) = 0. It follows that a ∈ Iτ , whence [a] = 0. Consequently, τ is faithful.
In the rest of this section we show relations between states on effect algebras and
their state operators.
Theorem 3.7. Every state on an effect algebra E induces a state operator on the
tensor product [0, 1] ⊗ E.
Proof. Let s : E → [0, 1] be a state on E. Since the identity mapping i : [0, 1] →
[0, 1] is (the unique) state on [0, 1], by [12], the tensor product [0, 1] ⊗ E exists.
Moreover, there is a unique state τ := i ⊗ s on [0, 1] ⊗ E with τ (α ⊗ a) = αs(a). It
is easy to check that the mapping α 7→ α ⊗ 1 is an isomorphism between [0, 1] and
[0, 1] ⊗ 1 ⊆ [0, 1] ⊗ E. Thus τ (τ (α ⊗ a)) = τ (αs(a) ⊗ 1) = αs(a) ⊗ 1. It follows that
τ is a state operator on [0, 1] ⊗ E.
(cid:3)
Definition 3.8. A set Ω of states on an EA E is called ordering iff for all e, f ∈ E,
ω(e) ≤ ω(f ) ∀ω ∈ Ω implies e ≤ f .
Theorem 3.9. [22] An effect algebra E admits an ordering set Ω of states if and
only if E is isomorphic to an effect subalgebra of [0, 1]Ω.
8
A. JEN COV ´A, S. PULMANNOV ´A
Theorem 3.10. Every state operator on an effect algebra E with ordering set of
states induces a state on E.
Proof. Let Ω be the ordering set of states on E and let τ : E → E be a state
operator on E. Since τ (E) is a subeffect algebra of E, Ω is an ordering set of
states also on τ (E). By Theorem 3.9, there is a monomorphism φ : τ (E) → [0, 1]Ω.
Put M := [0, 1]Ω, then M is an MV-algebra, and for a maximal ideal I of M , the
quotient M/I is isomorphic to a subalgebra of [0, 1] (e.g., [13, Prop. 2.2.33]). Let
q : M → M/I be the quotient mapping. Define s(a) := q(φ(τ (a))), a ∈ E. From
the properties of the mappings τ, φ and q we easily derive that s is a state on E. (cid:3)
4. Convex effect algebras
An effect algebra E is convex [24] iff for every a ∈ E and λ ∈ [0, 1] ⊆ R there
exists an element λa ∈ E such that the following conditions hold:
(C1) If α, β ∈ [0, 1] and a ∈ E, then α(βa) = (αβ)a.
(C2) If α, β ∈ [0, 1] with α + β ≤ 1 and a ∈ E, then αa ⊥ βa and (α + β)a =
αa ⊕ βa.
(C3) If a, b ∈ E with a ⊥ b and λ ∈ [0, 1], then λa ⊥ λb and λ(a ⊕ b) = λa ⊕ λb.
(C4) If a ∈ E, then 1a = a.
A map (λ, a) → λa that satisfies (C1) -- (C4) is called a convex structure on E.
Notice that 0a = 0 for every a ∈ E. Observe that a convex structure on E is
bimorphism from [0, 1] × E into E.
Example 4.1. It can be shown that every effect algebra which has at least one
state can be embedded into a convex one. Indeed, let E be an effect algebra with
at least one state. Then the tensor product (T, ρ) of [0, 1] and E exists. Similarly,
the tensor product (Tα, ρα) of [0, α] and E exists for every 0 6= α ∈ [0, 1]. Define
the mapping iα : [0, α] × E → T by iα(λ, a) = ρ(λ, a). The image of iα is in
the interval [0, ρ(α, 1)] of T , and iα : [0, α] × E → [0, ρ(α, 1)] is a bimorphism.
Therefore there is a morphism ψα : Tα → [0, ρ(α, 1)] such that ψα ◦ ρα = iα. Then
ψα(ρα(λ, a)) = ρ(λ, a), and since ρα(λ, a) are generating elements of Tα, while
ρ(λ, a) are generating elements of the interval [0, ρ(α, 1)] in T (λ ∈ [0, α], a ∈ E),
it follows that ψα is an isomorphism of Tα onto the interval [0, ρ(α, 1)] in T .
Define, for every α ∈ [0, 1], the mapping jα : [0, 1]× E → Tα, jα(λ, a) = ρ(αλ, a).
This mapping is a bimorphism, and hence it extends to a unique morphism φα :
T → Tα. We claim that for α ∈ [0, 1], x ∈ T , (α, x) 7→ φα(x) defines a convex
structure on T .
To prove (C1), let α, β ∈ [0, 1]. Then φα◦φβ(ρ(λ, a)) = φαρ(βλ, a) = ρ(αβλ, a)) =
φαβ(ρ(λ, a)). By uniqueness of the extensions, φα ◦ φβ = φαβ, which proves (C1).
To prove (C2), let α, β ∈ [0, 1] be such that α + β ∈ [0, 1]. Then for every
a ∈ E, λ ∈ [0, 1], φα(ρ(λ, a)) + φβ(ρ(λ, a)) = ρ(αλ, a) + ρ(βλ, a) = ρ(αλ + βλ, a) =
φα+β(ρ(λ, a)), which yields φα+β = φα + φβ, which is (C2).
(C3) follows from the fact that φα is a morphism for every α, and (C4) follows
from φ1(ρ(λ, a)) = ρ(λ, a) for all λ ∈ [0, 1] and every a ∈ E.
In the same way we prove that there is a convex structure on the tensor product
of [0, 1] ⊗σ E in the category of σ-effect algebras.
Example 4.2. Let H be a complex Hilbert space and let E(H) be the effect algebra
of Hilbert space effects from Example 2.4. For λ ∈ [0, 1] and a ∈ E(H), (λ, a) 7→ λa,
EFFECT ALGEBRAS WITH STATE OPERATOR
9
where λa is the usual scalar multiplication for operators, gives a convex structure
on E(H) and so E(H) becomes a convex effect algebra.
Example 4.3. Let (Ω, A) be a measurable space and let E(Ω, A) be the set of
measurable functions on Ω with values in [0, 1]. If we define, for f, g ∈ E(Ω, A),
f ⊥ g iff f + g ≤ 1 (pointwise), and in this case f ⊕ g(ω) := f (ω) + g(ω), ω ∈ Ω,
we obtain an effect algebra, called effect algebra of fuzzy events. Moreover, if we
define for λ ∈ [0, 1], (λf )(ω) := λf (ω) as the usual multiplication, we can see
that E(Ω, A) is a convex effect algebra. Fuzzy events are basic concepts in fuzzy
probability theory [2, 23].
Example 4.4. The latter above two examples are special cases of the following
example. Let V be an ordered real linear space with zero θ and with a strict positive
cone K. Recall that the partial order is defined by x ≤K y iff y − x ∈ K. We say
that K is generating iff V = K − K. Let u ∈ K with u 6= θ and form the interval
[θ, u] := {x ∈ K : x ≤K u}. For x, y ∈ [θ, u] we define x ⊥ y iff x + y ≤K u and in
this case we define x ⊕ y = x + y. It is easy to check that ([θ, u]; ⊕, θ, u) is an effect
algebra with x⊥ = u − x for every x ∈ [θ, u]. A straightforward verification shows
that (λ, x) 7→ λx, λ ∈ [0, 1] is a convex structure on [θ, u] so that ([θ, u]; ⊕, θ, u) is a
convex effect algebra. We say that [θ, u] generates K iff K = R+.[θ, u] and we say
that [θ, u] generates V iff [θ, u] generates K and K generates V [24].
It has been shown in [24] that every convex effect algebra is equivalent to a
convex effect algebra described in Example 4.4. We have the following result. We
recall that two ordered linear spaces (V1; K1) and (V2; K2) are order isomorphic iff
there exists a linear bijection T : V1 → V2 such that T (K1) = K2.
Theorem 4.5. [24, Theorem 3.1] Every convex effect algebra (E; ⊕, 0, 1) is affinely
isomorphic to an effect algebra ([θ, u]; ⊕, θ, u), where [θ, u] is a generating interval
for an ordered linear space (V ; K); and the effect algebra order ≤ on [θ, u] coincides
with the linear space order ≤K restricted to [θ, u]. Moreover, (V ; K) is unique in the
sense that if E is affinely isomorphic to an interval [θ1, u1] that generates (V1; K1),
then (V1; K1) is order isomorphic to (V ; K).
The following lemma shows that if an effect algebra E is embedded into the
tensor product [0, 1] ⊗ E, then a state operator on E can be extended to a state
operator on [0, 1] ⊗ E.
Lemma 4.6. Let τ : E → E be a state operator on the effect algebra E. Let (T, ρ)
be the tensor product of [0, 1] and E. Then τ uniquely extends to a state operator
on T .
Proof. Define a mapping ν : [0, 1]×E → T by ν(λ, a) = ρ(λ, τ (a)). Since τ is a state
operator and ρ a bimorphism, we obtain that ν is a bimorphism. Therefore, by the
universal property of the tensor product, there is a unique (effect algebra) morphism
τ o : T → T such that ν = τ o ◦ ρ. Moreover, τ o(τ o(ρ(λ, a)) = τ o(ρ(λ, τ (a)) =
ρ(λ, τ (τ (a))) = ρ(λ, τ (a)) = τ o(ρ(λ, a)). Since T is generated by elements of the
form ρ(λ, a), we obtain that τ o is idempotent.
(cid:3)
Lemma 4.7. Let τ : E → E be a state operator on a convex effect algebra E. Then
for every α ∈ [0, 1] ⊆ R and every a ∈ E, τ (αa) = ατ (a).
10
A. JEN COV ´A, S. PULMANNOV ´A
Proof. Let n be a nonnegative integer and a ∈ E such that na exists in E. Then
τ (na) = τ (a ⊕ a ⊕ · · · ⊕ a) = τ (a) ⊕ · · · ⊕ τ (a) = nτ (a). Now assume that n > 0,
a ∈ E, then
nτ (
1
n
a) = τ (
= τ (
a) ⊕ τ (
1
n
a ⊕ · · · ⊕
1
n
1
n
a) ⊕ · · · ⊕ τ (
a)
1
n
1
n
1
n
a) = τ (n(
a)) = τ (a),
hence τ ( 1
n a) = 1
n τ (a). It follows that for every m ≤ n we have τ ( m
n a) = m
n τ (a).
For every α ∈ [0, 1] there are sequences of rational numbers (qn)n, (rn)n such that
qn, rn ∈ Q∩[0, 1] and qn ↑ α, rn ↓ α. For every a ∈ E, qna ↑ αa, rna ↓ αa. Therefore
qnτ (a) ≤ τ (αa) ≤ rnτ (a). Taking the limits for n → ∞, we get τ (αa) = ατ (a). (cid:3)
Theorem 4.8. Let E be a convex effect algebra, and let [θ, u] be the generating
interval in an ordered linear space (V ; K) induced by E by Theorem 4.5. Then
any state operator τ : E → E extends to a linear, idempotent endomorphism p :
V → V such that p(K) ⊆ K. In addition, if x, y ∈ V +, and τ is strong, then
p(p(x) ∧ p(y)) = p(x) ∧ p(y) whenever p(x) ∧ p(y) exists.
n ) = n(τ ( a
n ) ⊕ τ ( b
n a). Let a, b ∈ V +, we find n such that a+b
Proof. We can (and do) identify E with the interval [θ, u]. Let a ∈ V +, then
since u is an order unit, there is n ∈ N such that a ≤ nu, hence 1
n a ≤ u. Put
p(a) := nτ ( 1
n ≤ u, and so p(a + b) =
nτ ( a+b
n )) = p(a) + p(b). Now for every x ∈ V , x = x1 − x2,
x1, x2 ∈ V +, and we may put p(x) = p(x1) − p(x2). If also x = y1 − y2, y1, y2 ∈ V +,
then from x1 + y2 = y1 + x2, we get p(x1) + p(x2) = p(y1) + p(y2), so that p(x)
is well defined. For a ∈ V +, p(p(a)) = p(nτ ( a
n )) = p(a).
From this we derive that p2 = p.
n )) = nτ (τ ( a
n )) = np(τ ( a
To finish the proof, assume that τ is strong. Let p(x) ∧ p(y) exist for x, y ∈ V +.
Then there is n ∈ N such that θ ≤ x, y ≤ n1, hence θ ≤ x/n, y/n ≤ 1. Therefore
p(p(x) ∧ p(y)) = p(n(τ (x/n) ∧ τ (y/n))) = nτ (τ (x/n) ∧ τ (y/n)) = n(τ (x/n) ∧
τ (y/n)) = p(x) ∧ p(y).
(cid:3)
5. State operators and conditional expectations on JC-effect
algebras
Let us consider the following examples.
Example 5.1. Let (Ω, Σ, µ) be a (finite) measure space. Let E be the set of all
Σ-measurable functions f : Ω → [0, 1]. E can be organized into an MV-algebra
by putting f ⊞ g = min{f + g, 1} and f ′ = 1 − f , where 1(ω) = 1 for all ω ∈ Ω.
Moreover, we also have E = {f : 0 ≤ f ≤ 1} ⊆ L2(Ω, Σ, µ). It has been shown that
a linear transformation T : L2 → L2 is a conditional expectation if and only if (0)
T is self-adjoint and idempotent (i.e, T is a projection); (1) T (1) = 1; and (2) if
f ∈ L2, g ∈ L2 then T (max{T f, T g}) = max{T f, T g} [48]. It can be easily seen
that if we restrict T to E, we obtain a strong state operator on E.
Example 5.2. With the standard quantum-mechanical Hilbert space formalism,
i=1 pixpi) , x ∈ B(H) is the state after a Luders -- von Neumann
i=1 and
with the eigen-projections (pi)∞
i=1 pi = I, when the physical system was in
the initial state µ prior to the measurement. While von Neumann [43] originally
µB(x) := µ(P∞
measurement of an observable B = P∞
i=1,P∞
i=1 bipi with discrete spectrum (bi)∞
EFFECT ALGEBRAS WITH STATE OPERATOR
11
considered only the case when the pi are one-dimensional projections (i.e. atoms
in the projection lattice P(H) of the Hilbert space H), that is, only an observable
B with a non-degenerate spectrum, Luders [38] later extended von Neumann's
measurement process to the case when pi need not longer be one-dimensional, and
the observable B may be degenerated.
The operator p := P∞
i=1 pixpi is the von Neumann-Luders conditional expec-
tation on B(H) with respect to B. Let τ (a) := Pi∈N piapi, a ∈ E(H), be the
restriction of p to E(H).
It is easy to check that this τ satisfies axioms (i), (ii), (iii) of Definition 3.1.
Moreover, if a commutes with all pi, then τ (a) = a, and conversely, τ (a) = a
i=1 piapi = pjapj, whence the range of τ consists of all effects
implies pja = pj P∞
that commute with all of the projections pi.
Moreover, the state operator induced by the von Neumann-Luders conditional
expectation is faithful, hence it is strong. Indeed, let Pi piapi = 0, then piapi = 0,
hence pia = 0 for all i, and hence Pi pia = a = 0.
Since E(H) is far from being a lattice [21, 37], but it is closed under the triple
products of the form aba, a, b ∈ E(H), it may be reasonable to consider, instead of
property (2), the following property:
(3)
∀a ∈ E(H), τ (τ (a)bτ (a)) = τ (a)τ (b)τ (a).
It is easy to check that the state operator induced by the von Neumann-Luders
conditional expectation satisfies (3).
Motivated by the latter examples, we will study state operators on so-called JC-
effect algebras, which generalize the effect algebra E(H) and their relations with
conditional expectations.
Recall that a JC-algebra is a norm-closed real vector subspace of bounded self-
adjoint operators on a Hilbert space H, closed under the Jordan product a ◦ b =
1
2 (ab + ba), [16]. A JC-algebra is called a JW-algebra if it is closed in the weak
topology. It was shown that the lattice of projections in a JW-algebra J must be
complete, hence J contains a unit 1, [49]. In what follows, we will suppose that
a JC-algebra contains the unit as well. In particular, if J = Asa is the set of all
self-adjoint operators in a C*-subalgebra A ⊂ B(H), we will suppose that A is
unital.
Let J be a JC-algebra and let J + = J ∩B(H)+ be the cone of positive operators
in J . Then J + = {a ◦ a : a ∈ J }. The set of effects in J is called a JC-effect
algebra, it will be denoted by E(J ) = J ∩ E(H). A sub-effect algebra E ⊂ E(J ) is
called a sub-JC-effect algebra if a2 ∈ E for all a ∈ E.
Recall that the triple product on a Jordan algebra is defined by
{xyz} = (x ◦ y) ◦ z + (y ◦ z) ◦ x − (z ◦ x) ◦ y.
In a JC-algebra J , we have
in particular,
{abc} =
1
2
(abc + cba),
a, b, c ∈ J ,
{aba} = 2(a ◦ b) ◦ a − (a2) ◦ b = aba.
It is clear that if a, b ∈ E(J ), then aba ∈ E(J ) and if E ⊂ E(H) is a sub-effect
algebra, then E is a sub-JC-effect algebra if and only if aba ∈ E whenever a, b ∈ E.
12
A. JEN COV ´A, S. PULMANNOV ´A
Let τ : E(J ) → E(J ) be a state operator. Since E(J ) is the [0, I] interval in
the ordered vector space (J , J +), τ extends to a linear, positive idempotent and
unital mapping τ : J → J , by Theorem 4.8. Such maps were studied in [15]
and it was shown that the image τ (J ) is (has an isometric Jordan representation
as) a JC-algebra, with the product τ (a) ⋆ τ (b) = τ (τ (a) ◦ τ (b)). Moreover, if τ is
faithful, τ (J ) is a Jordan subalgebra in J . All of the results in this section are
easy consequences of the results of [15], nevertheless, we include some proofs for
the convenience of the reader.
A crucial result used in this paragraph is the Kadison-Schwarz inequality [30],
extended to the context of JC-algebras [15], which states that if p : J → B(H)
is a linear, positive and contractive map, then p(a2) ≥ p(a)2 for all a ∈ J .
In
particular, a state operator τ on E(J ) satisfies
τ (a)2 ≤ τ (τ (a)2) ≤ τ (a2),
a ∈ E(J ).
(4)
We will show below that unital additive maps satisfying equality in either of these
inequalities are special cases of state operators. Moreover, we will show that each
state operator is a composition of such maps.
Lemma 5.3. Let p : J → J be a positive, linear mapping such that p(1) = 1 and
let a ∈ J . The following are equivalent:
(a) p(a2) = p(a)2;
(b) p(b ◦ a) = p(b) ◦ p(a) for all b ∈ J ;
(c) p(aba) = p(a)p(b)p(a) for all b ∈ J .
Proof. The first part of the proof is completely analogical to a proof of a similar
statement for maps on C*-algebras satisfying Schwarz inequality, [26].
(a) =⇒ (b): Let p(a2) = p(a)2, b ∈ Asa, t ∈ R. Then by the Kadison-Schwarz
inequality,
(5)
2tp(a) ◦ p(b) =p(ta + b)2 − t2p(a)2 − p(b)2
≤p((ta + b)2) − t2p(a2) − p(b)2
=2tp(a ◦ b) + p(b2) − p(b)2
Dividing the inequality by t and letting t → ±∞, we obtain (b).
(b) =⇒ (c): Note that since an = a ◦ an−1, we obtain by induction from (b)
that p(an) = p(a)n, n ∈ N. In particular, p(a2)2 = p(a)4 = p((a2)2), so that both
a and a2 satisfy the condition (a). Further, we have
p((a ◦ b) ◦ a) = p(a ◦ b) ◦ p(a) = (p(a) ◦ p(b)) ◦ p(a)
and by the part (a) =⇒ (b) of the proof,
p(a2 ◦ b) = p(a2) ◦ p(b) = p(a)2 ◦ p(b).
Hence
p(aba) = p(2(a ◦ b) ◦ a − a2 ◦ b) = 2(p(a) ◦ p(b)) ◦ p(a) − p(a)2 ◦ p(b) = p(a)p(b)p(a).
(c) =⇒ (a) is obvious, by putting b = 1.
(cid:3)
Definition 5.4. A mapping τ : E(J ) → E(J ) will be called a conditional expec-
tation on E(J ) iff the following conditions are satisfied:
(i) τ (1) = 1;
EFFECT ALGEBRAS WITH STATE OPERATOR
13
(ii) if a ⊥ b, then τ (a + b) = τ (a) + τ (b);
(iii) for all a, b ∈ E(J ), τ (τ (a)bτ (a)) = τ (a)τ (b)τ (a).
Notice that putting b = 1 in (iii) we get
(iii') τ (τ (a)2) = τ (a)2,
that is, equality holds in the first inequality of (4). We show in Corollary 5.6 that
we may replace (iii) with (iii') in Definition 5.4.
We will next study the relation between state operators and conditional expec-
tations on E(J ).
Theorem 5.5. Let τ : E(J ) → E(J ) be a conditional expectation, then τ is a state
operator on E(J ). Its extension τ satisfies
τ ({τ (a)bτ (c)}) = {τ (a)τ (b)τ (c)},
a, b, c ∈ J
In particular, the image τ (J ) is a Jordan subalgebra of J .
Proof. To show that τ is a state operator, it is enough to check that τ 2 = τ .
By property (iii) of conditional expectations we obtain τ (τ (a)2) = τ (a)2 for all
a ∈ E(J ). Since τ (a) = kakτ ( a
kak ) for positive elements in J and kak1 − a ≥ 0 for
all a ∈ J , it is easily checked that we have
(6)
τ (τ (a)2) = τ (a)2,
a ∈ J
Applying this equality to 1 + a for a ∈ E(J ), we obtain
2τ (a) = τ (1 + a)2 − τ (a)2 − 1 = τ (τ (1 + a)2) − τ (τ (a)2) − 1 = 2τ (τ (a)),
hence τ is idempotent.
It follows that τ (a) satisfies the property (a) of Lemma 5.3 for all a ∈ J . The
rest of the proof now follows by Lemma 5.3 (c) by linearity of the triple product.
(cid:3)
Corollary 5.6. Let τ : E(J ) → E(J ) be a unital additive mapping. Then τ is a
conditional expectation if and only if its range τ (E(J )) is a sub-JC-effect algebra
in E(J ).
Proof. It is clear that the range is a sub-JC-effect algebra iff (iii') holds. Suppose
it is true, then by the proof of Theorem 5.5, τ is idempotent, so that the condition
(iii) of Definition 5.4 follows by Lemma 5.3. The converse statement is obvious.
(cid:3)
Theorem 5.7. Let τ : E(J ) → E(J ) be a faithful state operator. Then τ is a
conditional expectation on E(J ).
Proof. Let a ∈ E(J ), then by the Kadison-Schwaz inequality, e := τ (τ (a)2) − τ (a)2
is an element in E(J ) such that τ (e) = 0. Since τ is faithful, this implies that
τ (τ (a)2) = τ (a)2. The rest now follows by Lemma 5.3.
(cid:3)
We now consider equality in the second inequality of (4).
Lemma 5.8. Let p : J → J be a linear positive unital idempotent map. Let a ∈ J .
Then the following are equivalent.
(a) p(a2) = p(p(a)2));
(b) p(a ◦ b) = p(p(a) ◦ p(b)), for all b ∈ J ;
14
A. JEN COV ´A, S. PULMANNOV ´A
(c) p(aba) = p(p(a)p(b)p(a)), for all b ∈ J .
Proof. The proof is very similar to the proof of Lemma 5.3.
Note that if a = p(a) is in the range of p in the above lemma, then a trivially
satisfies condition (a), hence we have
(cid:3)
(7)
(8)
for all a, b ∈ J .
p(p(a) ◦ b) = p(p(a) ◦ p(b))
p(p(a)bp(a)) = p(p(a)p(b)p(a))
Theorem 5.9. Let τ : E(J ) → E(J ) be a unital additive map such that τ (a2) =
τ (τ (a)2) for all a ∈ E(J ). Then τ is a state operator. Moreover,
Iτ := {a ∈ J , τ (a2) = 0}
is a Jordan ideal in J and the map [a]Iτ 7→ τ (a) is an isometric Jordan isomor-
phism of J Iτ onto the range τ (J ) with Jordan structure given by a ⋆ b = τ (a ◦ b).
Proof. It is clear that τ (a2) = τ (τ (a)2) extends to all a ∈ J +, in particular to a+I,
a ∈ E(J ). It follows that τ is idempotent, hence a state operator. Using Lemma
5.8, it is easy to see that τ (a2) = τ (τ (a)2) holds for all a ∈ J .
We next show that Iτ = {c ∈ J , τ (c) = 0}. Indeed, let c ∈ Iτ , then by Kadison
inequality, τ (c)2 ≤ τ (c2) = 0. Conversely, if τ (c) = 0, then τ (c2) = τ (τ (c)2) = 0.
Hence Iτ is a Jordan ideal if and only if bab ∈ Iτ for all a ∈ Iτ and b ∈ J , [16]. By
Lemma 5.8,
τ (bab) = τ (τ (b)τ (a)τ (b)) = 0
By [16], J Iτ is a JC-algebra. It is clear that φ : [a] 7→ τ (a) is a well defined
linear unital map J Iτ onto τ (J ). Moreover,
φ([a]2) = φ([a2]) = τ (a2) = τ (τ (a)2),
[a] ∈ J Iτ
this implies that φ is a Jordan homomorphism with respect to the product a ⋆ b on
τ (J ). To show that φ is an isometry, note that a − τ (a) ∈ Iτ for all a and hence
k[a]k = inf{ka + ck, c ∈ Iτ } ≤ kτ (a)k = kτ (a + c)k ≤ ka + ck,
c ∈ Iτ .
(cid:3)
Definition 5.10. A map on E(J ) satisfying the conditions of Theorem 5.9 will be
called a Jordan state operator on E(J ).
We now turn to the case of a JW-algebra.
Definition 5.11. We say that an additive map τ : E(J ) → E(J ) is normal iff for
every ascending net (aα)α, aα ր a implies τ (aα) ր τ (a).
Notice that τ is normal iff for any summable family (ai)i∈I of effects, we have
τ (⊕i∈I ai) = ⊕i∈I τ (ai). Recall that a state ρ on an effect algebra is completely
additive iff ρ(⊕i∈I ai) = Pi∈I ρ(ai) whenever the orthosum ⊕i∈I ai exists.
Let J be a JW-algebra, then E(J ) is an orthocomplete effect algebra, and
completely additive states on E(J ) coincide with the restrictions of normal states
on J to E(J ). Hence a map τ on E(J ) is normal if and only if φ ◦ τ is normal for
all normal states φ.
EFFECT ALGEBRAS WITH STATE OPERATOR
15
Corollary 5.12. Let J be a JW-algebra and let τ : E(J ) → E(J ) be a unital
normal map. Then τ is a conditional expectation if and only if its range is a sub-
JW-effect algebra in E(J ).
The following definition is analogous to the definition of a support of a normal
state [35, Definition 7.2.4].
Definition 5.13. For a normal state operator τ on E(J ), the support of τ is the
complement e of the maximal projection f ∈ E(J ) with the property τ (f ) = 0.
By Lemma 1.2 in [15], we have that τ (a) = τ (eae) for every a ∈ E(J ) and
τ (a) = 0 implies eae = 0, moreover, eτ (a) = τ (a)e for every a ∈ E(J ).
Corollary 5.14. Let τ be a normal state operator on E(J ), and let e be the support
of τ . Then τe, defined by τe(a) = τ (a)e, is a normal faithful conditional expectation
on eE(J )e = E(J )[0, e].
Let us now turn to the case when the state operator τ on a JC-effect algebra E(J )
is not faithful. The next Theorem (together with Theorem 5.9) is a reformulation
of [15, Corollary 1.5]).
Theorem 5.15. Let J be a JW-algebra and let τ : E(J ) → E(J ) be a normal
state operator. Then
(i) Eτ := {a ∈ E(J ), τ (a2) = τ (τ (a)2)} is a sub-JW-effect algebra in E(J );
(ii) there is a faithful normal conditional expectation µ on E(J ) with range Eτ ;
(iii) there is a normal Jordan state operator φ on Eτ such that τ = φ ◦ µ.
Proof. Let e be the support of τ . Let us define
µ(a) := τ (a)e + (1 − e)a(1 − e) = τe(eae) + (1 − e)a(1 − e),
a ∈ E(J ).
It is easy to see that µ is a normal state operator on E(J ). Suppose that µ(a) = 0
for some a ∈ E(J ), then we must have τ (a)e = (1 − e)a(1 − e) = 0. It follows that
τ (a) = τ (τ (a)) = τ (τ (a)e) = 0, so that eae = ae = 0 and a = ae + a(1 − e) = 0,
hence µ is faithful. By Theorem 5.7, µ is a conditional expectation.
We will show that the range of µ is Eτ : let a = µ(a), then it is quite clear that
τ (a2) = τ (τ (a)2). Conversely, suppose a ∈ Eτ , then by using (7), τ ((a−τ (a))2) = 0,
hence (a − τ (a))e = e(a − τ (a)) = 0. It follows that ae = ea = τ (a)e and µ(a) =
ae + a(1 − e) = a. This proves (i) and (ii).
Let φ be the restriction of τ to Eτ , then φ is clearly a normal Jordan state
operator and we have
φ ◦ µ(a) = τ (τ (a)e + (1 − e)a(1 − e)) = τ (a),
a ∈ E(J ),
so that (iii) is true.
(cid:3)
Let J be a JC-algebra and let A be the C*-subalgebra in B(H) generated by
J . Then we may identify the second dual J ∗∗ with the strong operator closure of
J in the second dual A∗∗ of A, J ∗∗ is thus a JW-algebra, [16, 15]. Moreover, if
p is a unital positive projection on J , then p extends to a normal unital positive
projection p∗∗ : J ∗∗ → J ∗∗.
Let now a be any element in the JW-effect algebra E(J ∗∗), then since the Ka-
plansky density theorem holds for JC-algebras, there is a net {aα} of elements in
the unit ball of J converging to a1/2 in the strong operator topology. It follows
16
A. JEN COV ´A, S. PULMANNOV ´A
that {a2
α} is a net of element in E(J ), converging to a, so that E(J ∗∗) is the strong
operator closure of E(J ) in J ∗∗. It also folows that any state operator τ on E(J )
extends to a normal state operator τ ∗∗ on E(J ∗∗), given by the restriction of τ ∗∗.
Corollary 5.16. Let J be a JC algebra and let τ be a state operator on E(J ).
Then there is a faithful normal conditional expectation µ on E(J ∗∗) and a Jordan
state operator φ on the range of µ such that τ = φ ◦ µE(J ).
6. State operators and conditional expectations on effect algebras
of Abelian C*-algebras
Let A be an abelian C*-algebra. Then A is isomorphic with C(X), the set of
all continuous complex valued functions on a compact Hausdorff space X. Notice
that the unit interval E(A) is the MV-effect algebra consisting of all continuous
functions f : X → [0, 1] ⊆ R, this will be denoted by C1(X). In fact, C1(X) is the
unit interval in the real abelian C*-algebra C(X; R) of continuous real functions on
X, which is a JC-algebra, with Jordan product being the usual product of functions.
Hence C1(X) is also a JC-effect algebra. Hence we may compare the notion of a
strong state operator and a conditional expectation.
Theorem 6.1. Let X be a compact Hausdorff space and let τ be a state operator
on C1(X). Then τ is a conditional expectation if and only if τ is strong.
Proof. Let τ be a conditional expectation on M . By Corollary 5.6, the range of
τ is a JC-effect-subalgebra of M .
It follows that τ (M ) is the unit interval in a
C*-subalgebra of C(X), and is therefore lattice ordered. This implies that τ is
strong.
Let τ be a strong state operator on C1(X) and let τ be its extension to C(X; R).
The range τ (C(X; R)) is a closed linear subspace of C(X; R) with the positive
cone τ (C(X, R)) ∩ C(X; R)+ = τ (C(X; R)+). By Theorem 4.8, we obtain that
τ (C(X; R)+) is lattice ordered, and hence τ (C(X; R)) is a lattice. Let us define
a relation x ∼ y, x, y ∈ X, iff f (x) = f (y) for all f ∈ τ (C(X; R)). Then ∼ is
an equivalence, and let Y := {[x] : x ∈ X} denote the quotient space, then Y is
compact Hausdorff. For all f ∈ τ (C(X; R)), define f [x] = f (x), then f [x] is well
defined, and does not depend on the choice of x ∈ [x]. The mapping f 7→ f is an
isometric linear isomorphism C(X; R) → C(Y ; R), and (f ∧ g)∼[x] = f ∧ g(x) =
f (x) ∧ g(x) = f [x] ∧ g[x]. The set of functions B := { f : Y → R : f ∈ τ (C(X; R))}
satisfies the following conditions:
(i) B separates points.
(ii) B contains the constant function 1.
(iii) If f ∈ B then αf ∈ B for all constants α ∈ R.
(iv) B is a boolean ring; that is, if f, g ∈ B, then f + g ∈ B and max{f, g} ∈ B.
By the boolean ring version of the Stone - Weierstrass theorem [28, Theorem 7.29],
B is dense in C(Y, R) hence B = C(Y ; R) and the range τ (C(X; R)) of τ is a JC-
subalgebra of C(X; R). By Corollary 5.6, τ is a conditional expectation.
(cid:3)
Let τ be any state operator on C1(X) and let τ be its extension to C(X; R).
By decomposing any f ∈ C(X) into real and imaginary parts, τ can be extended
to C(X), we denote this extension again by τ . It is easy to see that τ is a linear
positive and idempotent map on C(X). By Theorems 6.1 and 5.5, τ is strong if and
EFFECT ALGEBRAS WITH STATE OPERATOR
17
only if τ is a conditional expectation in the sense of operator algebras, as introduced
in [50].
Let now τ be any state operator on C1(X) with extension τ on C(X). Then
Iτ := {a ∈ C1(X) : τ (a) = 0} is an ideal in C1(X) and τ induces a faithful state
operator on the quotient C1(X)Iτ , see Theorem 3.6. Put Iτ := {f ∈ C(X) :
τ (f ∗f ) = 0}.
Theorem 6.2. Let τ be a state operator on C1(X). Then Iτ is a closed ideal in
C(X) and C1(X)Iτ ≃ E(C(X)Iτ ).
Proof. We start with a simple observation. For all f ≥ 0, τ (f ) = 0 ⇔ τ (f 2) = 0.
Indeed, let f ∈ C(X), f ≥ 0. Assume τ (f ) = 0. Then f 2 ≤ f .f , hence
0 ≤ τ (f 2) ≤ f τ (f ) = 0. Conversely, let τ (f 2) = 0. By Kadison inequality,
0 = τ (f 2) ≥ τ (f )2, which entails τ (f ) = 0.
To prove that Iτ is an ideal, let f, g ∈ Iτ and α, β ∈ C. Then 0 ≤ αf +
βg ≤ αf + βg, whence τ (αf + βg) ≤ ατ (f ) + βτ (g) = 0. This entails
τ (αf + βg2) = 0, hence αf + βg ∈ Iτ . Now let f ∈ Iτ , g ∈ C(X). Then
(f g)∗(f g) = f ∗f g∗g ≤ g∗gf ∗f , whence τ ((f g)∗(f g)) = 0, and f g ∈ Iτ .
Let f ∈ C1(X) and let [f ] ∈ C1(X)Iτ and [f ]∼ ∈ C(X)Iτ be the equivalence
classes containing f , i.e. [f ] = {g ∈ C1(X), f ∆g ∈ Iτ } and [f ]∼ = {g ∈ C(X), f −
g ∈ Iτ }. For g ∈ C1(X), f ∆g = f − g, so that g ∈ [f ] iff g ∈ [f ]∼. Moreover, let
f 1/2 = h, then [f ]∼ = [h2]∼ = ([h]∼)2 ≥ 0 and similarly for [1]∼ − [f ]∼ = [1 − f ]∼,
so that [f ]∼ ∈ E(C(X)Iτ ). It follows that [f ] 7→ [f ]∼ is a well defined injective map
C1(X)Iτ → E(C(X)Iτ ) and it is easy to see that it is additive. It is now enough
to check that any equivalence class E(C(X)Iτ ) contains an element of C1(X). So
let h ∈ C(X), [h]∼ ∈ E(C(X)Iτ ). We may clearly suppose that h ≥ 0, so that
¯h = h ∧ 1 ∈ C1(X). We will show that ¯h ∈ [h]∼.
Since Iτ is a closed ideal in C(X), Iτ = {f ∈ C(X), f (x) = 0, x ∈ K}, where
K ⊂ X is a closed subset given by K = {x ∈ X, f (x) = 0, f ∈ Iτ } [35, Theorem
3.4.1]. Since [1 − h]∼ ≥ 0, there is a positive element g ∈ C(X) and some f ∈ Iτ
such that 1 − h = g + f . Hence h(x) ≤ 1 for x ∈ K, so that h(x) − ¯h(x) = 0 for
x ∈ K, that is, h − ¯h ∈ Iτ and ¯h ∈ [h]∼.
(cid:3)
We have C(X)Iτ ≃ C(K), where the isomorphism is given by [f ]∼ 7→ f K. We
will call the set K the support of τ .
Corollary 6.3. Let τ be a state operator on C1(X). Then there is a closed subset
K ⊂ X such that C1(X)Iτ ≃ C1(K). Moreover, τ induces a faithful conditional
expectation µ on C(K), such that µ(f K) = τ (f )K, f ∈ C(X).
Let now τ be a state operator on M and let τ be its extension to C(X). Then
for x ∈ X, the map f 7→ τ (f )(x) is a state on C(X). By the Riesz representation
theorem, there is some Radon probability measure λx on X such that τ (f )(x) =
R f (y)λx(dy) for all f ∈ C(X). Let K be the support of τ . Since for f ≥ 0, f ∈ Iτ
iff τ (f ) = 0, it is easy to see that for any x ∈ X, the support of λx must lie in K.
Moreover, since any function g ∈ C(K) has a continuous extension to X, we see
that g 7→ R g(y)λx(dy) defines a linear positive unital map ϕ : C(K) → C(X) and
τ (f ) = ϕ(f K), f ∈ C(X).
18
A. JEN COV ´A, S. PULMANNOV ´A
Theorem 6.4. A map τ on C1(X) is a Jordan state operator if and only if there
is a closed subset K ⊂ X and a linear positive unital extension ϕ : C(K) → C(X)
such that τ (f ) = ϕ(f K) for all f ∈ C1(X).
Proof. Let τ be a state operator. Then τ is a Jordan state operator if and only
if τ (f )(x) = f (x) for all x ∈ K, f ∈ C1(X) Indeed, if τ is Jordan, then by the
proof of Theorem 5.9, g ∈ Iτ iff τ (g) = 0, for any real g ∈ C(X). Hence for
f ≥ 0, f 2 − τ (f )2 ∈ Iτ , so that (f 2 − τ (f )2)(x) = 0 for all x ∈ K. It follows that
f (x) = τ (f )(x) on K for f ∈ C1(X). Conversely, for x ∈ K, τ (f )(x) = f (x) for all
f ∈ C1(X) implies τ (f )2(x) = f 2(x) = τ (f 2)(x), so that g := τ (f 2) − τ (f )2 ∈ Iτ .
Since g ≥ 0, this is equivalent with τ (g) = 0. Hence τ is Jordan.
Let now τ be a Jordan state operator and let K be its support. By the remarks
above, τ defines a linear positive unital map ϕ : C(K) → C(X) such that τ (f ) =
ϕ(f K). Moreover, by the previous paragraph, we clearly have τ (f )(x) = f (x) for
x ∈ K and for all f ∈ C(X). Let g ∈ C(K) and let f ∈ C(X) be a continuous
extension of g, then ϕ(g)(x) = τ (f )(x) = f (x) = g(x) for x ∈ K.
Conversely, if τ is of the given form, then τ is clearly a state operator and
τ (f )(x) = ϕ(f K)(x) = f (x) for x ∈ K, so that τ is Jordan.
(cid:3)
7. State operators and conditional expectations on convex
MV-algebras
In this section, we will need some facts from the theory of MV-algebras, see
[8, 13] for more details.
Let (M ; ⊞,∗ , 0) be an MV-algebra, then M is called a σ- MV-algebra if it is a
σ-lattice. The set B(M ) := {a ∈ M : a ⊞ a = a} of all idempotents in M , with
the operations inherited from M , is a boolean σ-algebra. It is the largest boolean
σ-subalgebra of M .
A special example of a σ-MV-algebra is the following. Let X be a nonempty set.
A tribe T over X is a collection of functions T ⊆ [0, 1]X such that the zero function
0(x) = 0 is in T and the following is satisfied:
(1) f ∈ T =⇒ 1 − f ∈ T ;
(2) f, g ∈ T =⇒ f ⊞ g := min(f + g, 1) ∈ T ;
(3) fn ∈ T , n ∈ N and fn ր f (pointwise) =⇒ f ∈ T .
Every tribe is an MV-algebra - so-called Bold algebra of fuzzy sets, and also a σ-
MV-algebra where the lattice operations ∨, ∧ coincide with the pointwise supremum
and infimum, respectively, of [0, 1]-valued functions defined on X. Idempotents in
T are elements of the form χB ∈ T and the sets B ⊆ X, χB ∈ T (so-called crisp
sets) form a σ-algebra of sets, which we also denote by B(T ). A σ-additive state m
on a tribe T over X determines a probability measure on B(T ), by P (A) := m(χA)
for any A ∈ B(T ). By [7], each σ-additive state on T has the following integral
representation: for any f ∈ T , m(f ) = RX f dP .
The following theorem is a generalization of the Loomis-Sikorski theorem to
σ-MV-algebras [40, 11].
Theorem 7.1. Let M be a σ-MV-algebra. Then there exists a tribe M ∗ over a
compact Hausdorff space X and a σ-homomorphism η of M ∗ onto M .
Elements of the set X can be identified with extremal states on M . For each
a ∈ M there exists a unique continuous function a∗ ∈ M ∗ such that η(a∗) = a, this
EFFECT ALGEBRAS WITH STATE OPERATOR
19
function is given by a∗(ω) := ω(a), ω ∈ X. A function f ∈ M ∗ has the same image
as a∗ iff the set {ω ∈ X : f (ω) 6= a∗(ω)} is a meager set (i.e., a countable union of
nowhere dense sets). Moreover, B(M ∗) is mapped onto B(M ).
Let m be a σ-additive state on M . Then m∗ = m ◦ η is a state on M ∗ and
P ∗ : A ∈ B(M ∗) 7→ m∗(χA) is a probability measure on B(M ∗). For any a ∈ M we
have
m(a) = m(η(a∗)) = m∗(a∗) = ZX
a∗dP ∗.
In what follows, by a state we will always mean a σ-additive state.
The Loomis-Sikorski theorem enables us to extend some results from probability
theory to σ-MV-algebras. Conditional expectations on σ-MV-algebras have been
studied in [14]. Similar results for the special case of MV-algebras with product
were obtained in [36]. Let M be a σ-MV-algebra, N a sub-σ-MV-algebra of M and
let N ∗ be the sub-tribe of M ∗ generated by {b∗ ∈ M ∗ : b ∈ N }. Let m be a state on
M . For b∗ ∈ M ∗ we denote P ∗
b∗ B(M ∗)
is a σ-additive measure on B(M ∗), and P ∗
b∗ (a∗) = P ∗(B ∩ A) if a∗ = χA, b∗ = χB.
The following definition was introduced in [14].
b∗ (a∗) := m∗(b∗ ∧ a∗), a∗ ∈ M ∗. Clearly, P ∗
Definition 7.2. [14, Definition 4.1] An MV-conditional expectation of a ∈ M given
N in the state m is a B(N ∗)-measurable function m(aB(N )) : X → R such that
for any b ∈ B(N ),
(9)
ZX
m(aB(N ))(ω)dP ∗
b∗ (ω) = m(a ∧ b).
Clearly, the function m(aB(N )) is integrable with respect to P ∗ and is deter-
[P ∗]. Definition 7.2 implies that m(aB(N )) coincides with
mined uniquely a.e.
the Kolmogorovian conditional expectation of a∗ given B(N ∗) with respect to P ∗.
Notice that the function m(aB(N )) need not belong to N ∗, in general.
In [14], the equality (9) was extended to
(10)
ZX
m(aB(N ))(ω)dP ∗
b∗ (ω) = ZX
a∗dP ∗
b∗ (ω)
for all b ∈ N . Moreover, the following was proved.
Theorem 7.3. [14, Theorem 4.2] The following properties are satisfied a.e. [P ∗]:
(1) m(0B(N )) = 0, m(1B(N )) = 1
(2) a ⊡ b = 0 implies m(a ⊞ bB(N )) = m(aB(N )) + m(bB(N ))
(3) ∀a ∈ M , 0 ≤ m(aB(N )) ≤ 1.
(4) For a given sequence (an)n, an ր a implies m(anB(N )) ր m(aB(N )).
From now on, we will assume that the σ-MV-algebra M is a convex MV-effect
algebra. It is then easy to see that M is weakly divisible, that is, for any integer
n ∈ N, there is an element v ∈ M such that nv is defined and nv = 1; we then
write v := 1
n . Let X be as in Theorem 7.1. Then we have the following.
Theorem 7.4. [13, Theorem 7.3.12] An MV-algebra M is weakly divisible and σ-
complete if and only if X is basically disconnected and a 7→ a∗ is an MV-algebra
isomorphism of M onto C1(X).
On the other hand, for every constant α, 0 ≤ α ≤ 1, α1 ∈ M . It follows that the
tribe M ∗ contains all constant functions with values in [0, 1], an therefore by [47,
Theorem 8.4.1], [13, Theorem 7.1.7], M ∗ contains all B(M ∗)-measurable functions.
20
A. JEN COV ´A, S. PULMANNOV ´A
In the next lemma we show that if N is a convex σ-sub-MV algebra of M 3 then
N ∗ consists of all B(N ∗)- measurable functions. The following setup is inspired by
[48].
Lemma 7.5. Let M be a convex σ-MV-algebra and let N be a convex σ-sub-MV-
algebra of M . Let N ∗ := {f ∈ M ∗ : η(f ) ∈ N }, and B(N ∗) be the set of all
characteristic functions belonging to N ∗. Then N ∗ consists exactly of all B(N ∗)-
measurable functions in M ∗.
Proof. From the properties of N it easily follows that B(N ∗) is an algebra of subsets
i=1 Ai ր
i=1 Ai ), and since N is a
i=1 Ai ∈ B(N ∗). This shows that B(N ∗)
of X. If Ai ∈ B(N ∗) for i = 1, 2, 3, . . ., then Sk
i=1 Ai η-almost everywhere, we have η(χSk
χS∞
σ-MV-algebra, η(χS∞
is a σ-algebra of subsets of X.
i=1 Ai ) ∈ N , hence S∞
i=1 Ai ∈ B(N ∗). Since χSk
i=1 Ai) ր η(χS∞
We show that every function f ∈ N ∗ is B(N ∗)-measurable. By definition, every
characteristic function from N ∗ is B(N ∗)-measurable. Let us consider an arbitrary
f ∈ N ∗. Let α ∈ [0, 1), and put A := {x ∈ X : f (x) > α}. Define g(x) :=
max(f (x), α), then g(x) = f (x) if x ∈ A, g(x) = α if x /∈ A and since N ∗ is convex
and closed under lattice operations, g is also an element of N ∗. Since g(x) ≥ α, we
have h(x) := g(x) − α ∈ N ∗, and h(x) > 0 for x ∈ A, while h(x) = 0 for x /∈ A.
We denote An := {x ∈ X : h(x) > 1
n }; then A1 ⊂ A2 ⊂ . . ., Sn An = A.
Furthermore for each n = 1, 2, . . . we define the functions gn(x) := n. min(h(x), 1
n ).
Clearly, 0 ≤ gn(x) ≤ 1 and gn ∈ N ∗.
Evidently, gn(x) = 1 for x ∈ An, gn(x) = n.h(x) for x /∈ An. Since for x /∈ A,
h(x) = 0, we see that gn(x) ր χA(x) for all x ∈ X. Since N is a σ-MV algebra, this
entails that χA ∈ N ∗. This proves that all functions f ∈ N ∗ are B(N ∗)-measurable.
On the other hand, going, as usual, from characteristic functions to simple func-
tions etc., we show that N ∗ contains all B(N ∗)-measurable functions from M ∗.
Hence N ∗ consists precisely of all B(N ∗)-measurable functions from M ∗. In par-
ticular, if f, g ∈ N ∗ then f.g ∈ N ∗.
(cid:3)
Corollary 7.6. Let M be a convex σ-MV-algebra, m a σ-additive state on M ,
and N a convex σ-MV-subalgebra of M . Then for every a ∈ M , the conditional
expectation m(aN ) ∈ N ∗, and η(m(aN )) belongs to N .
Let M be a σ-MV-algebra, N a σ-MV-subalgebra of M , and m a σ-additive
state on M . Then Im := {a ∈ M : m(a) = 0} is a σ-MV algebra ideal on M and
the quotient M := M Im is a σ-MV-algebra. Put N := {[a] : a ∈ N }, hence N is
the subclass of M consisting of all equivalence classes having a representative in N .
Theorem 7.7. Let M be a convex σ-MV algebra, N a convex σ-MV-subalgebra of
M , and m a σ-additive state on m. Define the mapping τ : M → M by τ [a] :=
[η(m(aN ))]. Then τ is a strong state operator on M , and the range of τ is τ ( M ) =
N .
Proof. (1) First we prove that τ is well defined. That is, we have to prove that
for any a1, a2 ∈ M , a1 ∼ a2 implies η(m(a1N )) ∼ η(m(a2N )). Now a1 ∼ a2 iff
m(a1∆a2) = 0. Let b ∈ B(N ) with b∗ = χB, B ∈ B(N ∗). Then (a1 ∧ b)∆(a2 ∧ b) =
(a1∆a2) ∧ b, so that a1 ∼ a2 entails that m((a1 ∧ b)∆(a2 ∧ b)) = 0, and hence
3N is a convex sub-MV-algebra of a convex MV-algebra M iff N is a sub-MV-algebra of M
and a ∈ N , α ∈ [0, 1] implies αa ∈ N .
EFFECT ALGEBRAS WITH STATE OPERATOR
21
m(a1 ∧ b) = m(a2 ∧ b) for all b ∈ B(N ). Write f1 := m(a1N ), f2 := m(a2N ).
Then f1, f2 are B(N ∗)-measurable, and we have RB(f1 − f2)dP ∗ = 0, ∀B ∈ B(N ∗).
Putting B1 = {ω : f1(ω) > f2(ω)}, B2 := {ω : f1(ω) < f2(ω)}, we obtain P ∗(B1) =
0 = P ∗(B2). From this we have P ∗{ω : f1 − f2 6= 0} = 0 =⇒ m ◦ η{ω :
f1∆f2 6= 0} = 0 =⇒ m ◦ η(f1∆f2) = 0 =⇒ m(η(f1)∆η(f2)) = 0, which yields
η(m(a1N )) ∼ η(m(a2N )).
(2) Next we prove that τ is a σ-additive state operator. By Theorem 7.3 (1),
τ [1] = [1]. By Theorem 7.3 (2) and (3), a ⊡ b = 0 implies m(a ⊕ bN ) = m(aN ) +
m(bN ) ≤ 1 a.e. [P ∗], whence τ [a ⊕ b] = τ [a] ⊕ τ [b]. Similarly, σ-additivity follows
by Theorem 7.3 (4). For every [c] ∈ N , there is c1 ∈ N with [c] = [c1]. Then by
Lemma 7.5, c∗
1 = m(c1N ), which implies
τ [c] = [η(c∗
1)] = [c1]. Moreover, τ (τ [c]) = τ [c1] = [c1] = τ [c]. Conversely, for every
a ∈ M , m(aN ) ∈ N ∗, hence η(m(aN )) ∈ N , which gives τ [a] ∈ N . It follows
that τ is a state operator and the range of τ is N . But N is an MV-algebra, hence
τ [a] ∧ τ [b] ∈ N , and therefore τ (τ [a] ∧ τ [b]) = τ [a] ∧ τ [b], and hence τ is strong. (cid:3)
1 ∈ N ∗ is B(N ∗)-measurable, therefore c∗
Let τ be a σ-additive state operator on a convex σ-MV-algebra M . Then (by
Theorem 7.4) we may define a map τ ∗ : C1(X) → C1(X) by
(11)
τ ∗(a∗) := τ (a)∗
We show that τ ∗ is a σ-additive state operator on C1(X). Indeed, τ ∗(1∗) = τ (1)∗ =
1∗; a∗ +b∗ ≤ 1∗ implies (a∗+b∗) = (a⊕b)∗, whence τ ∗(a∗+b∗) = τ (a⊕b)∗ = τ (a)∗ +
τ (b)∗ = τ ∗(a∗) + τ ∗(b∗); τ ∗(τ ∗(a∗)) = τ ∗(τ (a)∗) = (τ (τ (a))∗ = τ (a)∗ = τ ∗(a∗);
n ր a∗ implies an ր a, so that τ (an) ր τ (a), whence τ (an)∗ ր τ (a)∗,
moreover, a∗
n) ր τ ∗(a∗) (notice that by [20, Lemma 9.12], the isomorphism
which yields τ ∗(a∗
a 7→ a∗ preserves countable suprema and infima).
Conversely, if τ ∗ is a state operator on C1(X), then τ (a) := η(τ ∗(a∗)) defines a
state operator on M . In addition, from ω(a∧b) = inf{ω(a), ω(b)} for every extremal
state ω ∈ X, we obtain (a ∧ b)∗ = a∗ ∧ b∗, which yields τ ∗(τ ∗(a∗) ∧ τ ∗(b∗)) =
τ ∗(τ (a)∗ ∧ τ (b)∗) = τ ∗((τ (a) ∧ τ (b))∗) = (τ (τ (a) ∧ τ (b)))∗, and it shows that τ is
strong iff τ ∗ is strong. Therefore, studying state operators on M may be replaced
by studying state operators on C1(X) and Theorem 6.1 applies, hence τ is a strong
state operator on M iff τ ∗ is a conditional expectation on C1(X), in the sense of
Section 6.
In the next theorem we show that a strong state operator on M yields an MV-
conditional expectation on M .
Theorem 7.8. Let M be a convex σ-MV-algebra and τ a σ-additive strong state
operator on M . Then there is a convex σ-MV subalgebra N of M and a state m on
M such that for every a ∈ M , τ ∗(a∗) is an MV-conditional expectation of a with
respect to N in m.
Proof. Assume that M is a convex σ-MV-algebra. We may identify M with C1(X),
and define, for a, b ∈ M , a.b := η(a∗.b∗). As τ is strong, it is a conditional expec-
tation, hence for all a, b ∈ M , τ (a.τ (b)) = τ (a).τ (b).
Let s be a σ-additive state on M . Then m := s ◦ τ is also a σ-additive state
on M , and m ◦ η is a σ-additive state on M ∗. Let µ be the probability measure
µ := m∗B(M ∗). Put N := τ (M ), then it is clear that N is a convex σ-MV-
subalgebra of M . For all a ∈ M , b ∈ B(N ),
22
A. JEN COV ´A, S. PULMANNOV ´A
m(a ∧ b) = ◦τ ◦ η(a∗.b∗).
Taking into account that by (11) τ (a)) = τ (η(a∗)) = η(τ ∗(a∗)), and that b = τ (b)
implies b∗ = τ ∗(b∗), we obtain
s ◦ τ ◦ η(a∗.b∗) = s ◦ η(τ ∗(a∗.b∗))
= s ◦ η(τ ∗(τ ∗(a∗.τ ∗(b∗))))
= s ◦ τ ◦ η((τ ∗(a∗).τ ∗(b∗)))
= ZX
τ ∗(a∗)b∗dµ,
which shows that τ ∗(a∗) is an MV-conditional expectation of a.
(cid:3)
References
[1] E.G. Beltrametti, G. Cassinelli: The logic of quantum mechanics, Addison-Wesley, Reading
1981.
[2] E.G. Beltrametti, S. Bugajski: A classical axtension of quantum mechanics, J. Phys. A:
Math. Gen. 28 (1995), 3329 -- 3343.
[3] E. Beltrametti, S. Bugajski, S. Gudder, S. Pulmannov´a: Convex and linear effect algebras,
Rep. Math. Phys. 44 (1999), 359 -- 379.
[4] S. Bugajski, S. Gudder, S. Pulmannov´a: Convex effect algebras, state ordered effect alge-
bras, and ordered linear spaces, Rep. Math. Phys. 45 (2000), 371 -- 388.
[5] D. Buhagiar, E. Chetcuti, A. Dvurecenskij: Loomis-Sikorski theorem and Stone duality
for effect algebras with internal state, Fuzzy Sets and Systems 172 (2011), 71 -- 86.
[6] P. Busch, P.J. Lahti, P. Mittelstaedt: The Quantum Theory of Measurement, Lecture
Notes in Physics, Springer, Berlin, 1991.
[7] D. Butnariu, E. P. Klement: Triangular-norm based measures and their Markov kernel
representation, J. Math. Anal. Appl. 162(1991), 111 -- 143.
[8] R. Cignoli, I.M.L. D'Ottaviano, D. Mundici: Algebraic Foundations of Many-valued Rea-
soning, Kluwer, Dordrecht, 2000.
[9] C.C. Chang: Algebraic analysis of many valued logic, Trans. Amer. Math. Soc. 88 (1958),
467 -- 490.
[10] A. Dvurecenskij: Tensor product of difference posets, Trans. Amer. Math. Soc. 347 (1995),
1043 -- 1057.
[11] A. Dvurecenskij: Loomis-Sikorski theorem for σ-complete MV-algebras and ℓ-groups, J.
Austral. Math. Soc. Ser. A 68 (2000), 261-277.
[12] A. Dvurecenskij: Tensor product of difference posets or effect algebras, Int. J. Theor. Phys.
34 (1995), 1337 -- 1348.
[13] A. Dvurecenskij, S.Pulmannov´a: New Trends in Quantum Structures, Kluwer, Dordrecht,
2000.
[14] A. Dvurecenskij, S.Pulmannov´a:Conditional probability on σ-MV-algebras, Fuzzy Sets and
Systems 155 (2005), 102 -- 118.
[15] E.G. Effros, E. Størmer: Positive projections and Jordan structure in operator algebras,
Math. Scand. 45 (1979), 127 -- 138.
[16] E.G. Effros, E. Størmer: Jordan algebras of self-adjoint operators, Trans. Amer. Math.
Soc. 127 (1967), 313 -- 316.
[17] T. Flaminio, F. Montagna: MV-algebras with internal states and probabilistic fuzzy logics,
Int. J. Approx. Reasoning 50 (2009) 138 -- 152.
[18] D.J. Foulis, M.K. Bennett: Effect algebras and unsharp quantum ligics, Found. Phys. 24
(1994), 1325 -- 1346.
[19] R. Giuntini, H. Greuling: Toward a formal language for usharp properties, Found. Phys.
19 (1989), 931 -- 945.
EFFECT ALGEBRAS WITH STATE OPERATOR
23
[20] K.R. Goodearl: Partially Ordered Abelian Groups with Interpolation, Math. Surveys and
Monographs No. 20, Providence, Rhode Island, 1986.
[21] S. Gudder: Lattice properties of quantum effects, J. Math. Phys. 37 (1996), 2637 -- 2642.
[22] S. Gudder: Effect algebras are not adequate models for quantum mechanics, Found. Phys.
40 (2010), 1566 -- 1577.
[23] S. Gudder: Fuzzy probabilitry theory, Demonstratio Math. 31 (1998), 235 -- 254.
[24] S. Gudder, S. Pulmannov´a: Representation theorem for convex effect algebras, Comment.
Math. Univ. Carolinae 39 (1998), 645 -- 659.
[25] S. Gudder: Morphisms, tensor products and σ-effect algebras, Rep. Math. Phys. 42 (1998),
321 -- 346.
[26] M. Hamana: Injective envelopes of C*-algebras, J. Math. Soc. Japan 31 (1) (1979), 181 --
197.
[27] P. H´ajek: Metamathematics of Fuzzy Logic, Kluwer, Dordrecht, 1998.
[28] E. Hewitt, K.Stromberg: Real and Abstract Analysis, Springer, Berlin, 1965.
[29] A.S. Holevo: Probabilistic and Statistic Aspects of Quantum Thepry, North-Holland, Am-
sterdam, 1982.
[30] R.V. Kadison: A generalized Schwarz inequality and algebraic invariants for operator
algebras, Ann. Math. 56 (1952), 494 -- 503.
[31] J. Kuhr, D. Mundici: De Finetti theorem and Borel states in [0, 1]-valued algebraic logic,
Internat. J. Approx. Reason. 46 (2007), 605 -- 616.
[32] F. Kopka, F. Chovanec: D-posets, Math. Slovaca 44 (1994), 21 -- 34.
[33] P.J. Lahti, M.J. M¸aczy´nski: On the order structure of the set of effects in quantum me-
chanics, J. Math.Phys. 36 (1995), 1673 -- 1680.
[34] G. Ludwig: Foundations of Quantum Mechanics, Springer, New York, 1983.
[35] R.V. Kadison, J.R. Ringrose: Fundamentals of the Theory of Operator Algebras, vol. I,
Academic Press, New York, 1983.
[36] K. Kraus, States, Effects and Operations, Fundamental Notes of Quantum Theory,
Springer-Verlag, 1988.
[37] P.J. Lahti, M.J. M¸aczy´nski: On the order structure of the set of effects in quantum me-
chanics, Rep. Math. Phys. 39 (1977), 339 -- 351.
[38] G. Luders: Uber die Zustandsanderung durch den Messprocess, Ann. Phys. 8 (1951),
322 -- 328.
[39] D. Mundici: Interpretation of AF C*-algebras in Lukaszievicz sentential calculus, J. Funct.
Anal. 65 (1986), 15 -- 63.
[40] D. Mundici: Tensor product and the Loomis-Sikorski theorem for MV-algebras, Adv. Appl.
Math. 22 (1999), 227 -- 248.
[41] D. Mundici: Averaging the truth-value in Lukasziewicz logic, Stud. Log. 55 (1995)113 -- 127.
[42] M. Navara: An orthomodular lattice admitting no group-valed measure, Proc. Amer. Math.
Soc. 122 (1994), 7 -- 12.
[43] J. von Neumann: Mathematische Grundlagen der Quantenmechanik, Springer, Berlin,
1932.
[44] P. Pt´ak, S. Pulmannov´a: Orthomodular Structures as Quantum Logics, Kluwer, Dordrecht,
1991.
[45] S. Pulmannov´a: Tensor product of divisible effect algebras, Bull. Austral. Math. Soc. 68
(2003), 127 -- 140.
[46] B. Riecan, D. Mundici: Probability on MV-algebras, In: Handbook of Measure Theory,
Vol. II (E. Pap, ed.), Elsevier Science, Amsterdam 2002, pp. 869 -- 909.
[47] B. Riecan, T. Neubrunn: Integral Measure and Ordering, Kluwer, Dordrecht, 1997.
[48] Z. Sid´ak: On relations between strict sense and wide sense conditional expectations, Teor.
Verojatnost. i Primenen. 2 (1957), 283 -- 288.
[49] D.M. Topping: Jordan algebras of self-adjoint operators, A.M.S. Memoir No 53, Provi-
dence, Rhode Island, 1965.
[50] H. Umegaki: Conditional expectations in operator algebras, Tohoku Math. J. 6 (1954),177 --
181.
[51] H. Weber: There are orthomodular lattices without non-trivial group-valued states: A
computer based construction, J. Math. Anal. Appl. 183 (1994), 89 -- 93.
24
A. JEN COV ´A, S. PULMANNOV ´A
Mathematical Institute, Slovak Academy of Sciences, Stef´anikova 49, 814 73 Bratislava,
Slovakia
E-mail address: [email protected], [email protected]
|
1603.02643 | 2 | 1603 | 2016-08-13T03:50:24 | Semidirect products of C*-quantum groups: multiplicative unitaries approach | [
"math.OA"
] | C*-quantum groups with projection are the noncommutative analogues of semidirect products of groups. Radford's Theorem about Hopf algebras with projection suggests that any C*quantum group with projection decomposes uniquely into an ordinary C*-quantum group and a "braided" C*-quantum group. We establish this on the level of manageable multiplicative unitaries. | math.OA | math |
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS:
MULTIPLICATIVE UNITARIES APPROACH
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
Abstract. C∗-quantum groups with projection are the noncommutative ana-
logues of semidirect products of groups. Radford's Theorem about Hopf alge-
bras with projection suggests that any C∗-quantum group with projection
decomposes uniquely into an ordinary C∗-quantum group and a "braided"
C∗-quantum group. We establish this on the level of manageable multiplicative
unitaries.
1. Introduction
Many important Lie groups like the Poincaré group or the group of motions of
Euclidean space are defined as semidirect products of smaller building blocks. What
is the quantum group analogue of a semidirect product? Such a notion should be
useful to understand quantum deformations of semidirect products.
For a semidirect product of groups, we need two groups G and H and an action
of G on H by group automorphisms. Since non-commutative quantum groups
cannot act on other quantum groups by automorphisms, we need a different point
of view: semidirect product groups are the same as groups with a projection. A
semidirect product of groups G ⋉ H comes with a canonical group homomorphism
p : G ⋉ H → G ⋉ H,
(g, h) 7→ (g, 1H),
which is idempotent, that is, p2 = p. Its kernel and image are H ⊆ G ⋉ H and
G ⊆ G ⋉ H, respectively. The conjugation action of G on H needed for a semidirect
product is the restriction of the conjugation action of G ⋉ H on itself. Therefore,
an idempotent group homomorphism p : K → K on a group K is equivalent to a
semidirect product decomposition of K.
Now consider a quantum group with a projection, that is, with an idempotent
quantum group endomorphism. What corresponds to the building blocks G and H
in a semidirect product of groups? If "quantum group" means "Hopf algebra," then
a theorem by Radford [15] answers this question. Here we consider C∗-quantum
groups, meaning C∗-bialgebras coming from manageable multiplicative unitaries
(see [21, 28]). More precisely, we work on the level of the multiplicative unitaries
themselves to avoid analytical difficulties.
Let us first recall Radford's Theorem. It splits a Hopf algebra C with a projection
p : C → C into two pieces A and B. The "image" of the projection A is a Hopf
algebra as well. The "kernel" of the projection B is only a Hopf algebra in a
certain braided monoidal category, namely, the category of Yetter -- Drinfeld modules
2000 Mathematics Subject Classification. 46L89 (81R50 18D10 ).
Key words and phrases. quantum group, braided quantum group, semidirect product, bosoni-
sation, multiplicative unitary, braided multiplicative unitary, quantum E(2) group.
Supported by the German Research Foundation (Deutsche Forschungsgemeinschaft (DFG))
through the Research Training Group 1493. The second author was also supported by a Fields --
Ontario postdoctoral fellowship. The third author was partially supported by the Alexander von
Humboldt-Stiftung and the National Science Center (NCN), grant 2015/17/B/ST1/00085.
1
2
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
over A. The tensor product of two Yetter -- Drinfeld algebras is again a Yetter --
Drinfeld algebra, for the diagonal Yetter -- Drinfeld module structure and a certain
deformed multiplication. The comultiplication on B is a homomorphism to the
deformed tensor product B ⊠ B.
Radford's Theorem contains two constructions. One puts together A and B
into their "semidirect product" C and describes the projection p on C. The other
splits C into the two factors A and B, with the Hopf algebra structure on A and
the A-Yetter -- Drinfeld algebra and braided Hopf algebra structure on B. The first
construction is called "bosonisation" by Majid [8]. The analogue of this construction
for C∗-quantum groups is described in [12], except for the projection that we expect
on this semidirect product.
In particular, the appropriate analogues of Yetter --
Drinfeld algebras and their deformed tensor product ⊠ are described in [12] for
arbitrary C∗-quantum groups. For regular C∗-quantum groups with Haar weights,
this is already done by Nest and Voigt [13].
The "projections" on C∗-quantum groups that we use are morphisms as intro-
duced in [10, 14]. That is, a quantum group morphism from (C, ∆C ) to (A, ∆A) is
a bicharacter in UM( C ⊗ A). Several equivalent descriptions of such morphisms
are given in [10], including functors between the categories of C∗-algebra coac-
tions that preserve the underlying C∗-algebra, and Hopf ∗-homomorphisms be-
tween the associated universal quantum groups. These are more general than Hopf
∗-homomorphisms between the reduced quantum group C∗-algebras.
Thus a C∗-quantum group with projection consists of a C∗-quantum group (C, ∆C )
with a unitary multiplier P ∈ UM( C ⊗C) with certain properties. To express these,
we use a manageable multiplicative unitary W ∈ U(H ⊗ H) that generates C; in
particular, W satisfies the pentagon equation
W
W
23
(1.1)
Then C and C act faithfully on H. Write P for P viewed as an operator on H ⊗ H.
The condition that P is a bicharacter is equivalent to
in U(H ⊗ H ⊗ H).
12 = W
W
W
13
12
23
(1.2) P23W
12 = W
12
P13P23
and W
23
P12 = P12P13W
23
in U(H ⊗ H ⊗ H).
The condition that P is idempotent for the composition of quantum group homo-
morphisms is equivalent to the pentagon equation for P:
in U(H ⊗ H ⊗ H).
P23P12 = P12P13P23
(1.3)
Thus a C∗-quantum group with projection is determined by two unitaries W, P ∈
U(H ⊗ H) that satisfy (1.1) -- (1.3); in addition, W must be manageable. Equa-
tion (1.3) means that P is a multiplicative unitary in its own right. It is manageable
if W is. The C∗-quantum group (A, ∆A) it generates is the image of the projection.
It is much more difficult to describe the other factor B. As a C∗-algebra, it should
be the generalised fixed-point algebra for a canonical coaction of (A, ∆A) on (C, ∆C ).
In the group case, this says that C0(H) is the generalised fixed-point algebra for
the left or right translation action of G on C0(G ⋉ H). Unless G is compact, this
requires Rieffel's generalisation of fixed-point algebras to group actions that are
"proper" in a suitable sense (see [9, 16]). Buss [2, 3] has generalised this theory
to locally compact quantum groups. We only need the special case of quantum
homogeneous spaces, which is also treated by Vaes [24]. All these approaches need
some regularity assumptions on (A, ∆A) and are technically difficult.
We may avoid these difficulties by staying on the level of multiplicative unitaries.
We already described a C∗-quantum group with projection through two multiplica-
tive unitaries W, P ∈ U(H ⊗ H) on the same Hilbert space that are linked by the
conditions (1.2). We find that any such pair comes from a "braided multiplicative
unitary" over the C∗-quantum group (A, ∆A) generated by P.
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
3
A braided multiplicative unitary is a unitary F ∈ U(L ⊗ L) for a Hilbert space L
with a Yetter -- Drinfeld module structure over (A, ∆A). That is, L carries corepre-
sentations U ∈ U(K(L) ⊗ A) and V ∈ U(K(L) ⊗ A) that are linked by a Yetter --
Drinfeld commutation relation. In addition, F is equivariant for the tensor product
corepresentations U U and V V on L ⊗ L and satisfies the braided pentagon
equation:
(1.4)
F23F12 = F12(
LL
)23F12(
LL
)∗
23
F23
in U(L ⊗ L ⊗ L).
LL
Here
Hilbert space L with itself, see [12].
denotes the braiding operator on the tensor product of the Yetter -- Drinfeld
Since A and A are represented faithfully on H, the unitaries U and V are de-
termined by their images U and V in U(L ⊗ H). It is convenient to replace V by
V := ΣV∗Σ ∈ U(H ⊗ L). We also write W instead of P; the multiplicative unitary
for the semidirect product quantum group will be denoted by WC .
Thus a braided multiplicative unitary is a family of four unitaries W ∈ U(H⊗H),
U ∈ U(L ⊗ H), V ∈ U(H ⊗ L), and F ∈ U(L ⊗ L) for two Hilbert spaces H
and L; these unitaries satisfy seven conditions: the pentagon condition for W; one
corepresentation condition each for U and V, which link them to W; the Yetter --
Drinfeld condition linking U and V; the equivariance of F with respect to U U
and V V; and the braided pentagon equation for F. We show that given these
four unitaries subject to these seven conditions, the unitary
(1.5)
WC
1234 := W
13
U23 V∗
34
F24 V34
in U(H ⊗ L ⊗ H ⊗ L)
is multiplicative. Furthermore, the unitaries WC and P := W
U23 on H ⊗ L ⊗
H ⊗ L satisfy the conditions (1.1) -- (1.3) that characterise C∗-quantum groups with
projection. The only analytic issue is to prove that WC is manageable if the braided
multiplicative unitary is manageable in a suitable sense. Otherwise, the claim is
proved by a direct computation. This has to be lengthy, however, because all seven
conditions on our four unitaries must play their role.
13
Conversely, let WC , P ∈ U(H ⊗ H) be unitaries satisfying the conditions (1.1) --
(1.3), with WC manageable. Then we construct a braided multiplicative unitary
based on the unitary W = P ∈ U(H ⊗ H), that is, we construct a Hilbert space L
and unitaries U ∈ U(L ⊗ H), V ∈ U(H ⊗ L), and F ∈ U(L ⊗ L) satisfying the
conditions for a braided multiplicative unitary, and we check that this braided
multiplicative unitary is manageable. When we construct a pair (WC , P) out of
this data as in (1.5), then we do not get back the same data we started with
because the underlying Hilbert spaces have changed. We show, however, that the
resulting C∗-quantum groups with projection are the same. This isomorphism is
also implemented by a quantum group isomorphism in the category constructed
in [10].
When we start with a manageable braided multiplicative unitary, form the
crossed product as in (1.5) and go back, we also get a different braided multiplica-
tive unitary, which should be "equivalent" to the one we started with. Since we do
not discuss how a braided multiplicative unitary generates a braided C∗-quantum
bialgebra, we cannot yet express this equivalence.
We treat one example of a braided multiplicative unitary in detail, namely, the
one that defines the simplified quantum E(2) group, a variant of the quantum E(2)
group introduced by Woronowicz in [26]. We write down the braided multiplicative
unitary and check that it is manageable. Similar computations appear in [1, 26].
4
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
2. Projections on Quantum Groups
A C∗-quantum group is, by definition, a C∗-bialgebra that is generated by a
manageable multiplicative unitary, see [21, 28]. We do not assume a C∗-quantum
group to have Haar weights. We fix a C∗-quantum group H = (C, ∆C ) and let
W ∈ U(H ⊗ H) be a manageable multiplicative unitary on a Hilbert space H that
generates it. Let H = ( C, ∆C ) be the dual quantum group.
A bialgebra morphism (A, ∆A) → (C, ∆C ) between two C∗-bialgebras is a C∗-algebra
morphism f : A → C (that is, a nondegenerate ∗-homomorphism A → M(C)) mak-
ing the following diagram commute:
∆A
∆C
f
A
C
A ⊗ A
f ⊗ f
C ⊗ C
This notion of morphism is too restrictive, however, because a group homomor-
phism G → H need not induce a morphism C∗
r (H). When we speak of
morphisms of C∗-quantum groups, we will mean those introduced by Ng [14], and
we shall use the equivalent characterisations of these morphisms in [10].
r (G) → C∗
Definition 2.1. A C∗-quantum group with projection is a C∗-quantum group with
an idempotent quantum group endomorphism.
Before we make this definition explicit, we consider the commutative case. It al-
lows us to view C∗-quantum groups with projection as C∗-quantum group analogues
of semidirect products of groups.
Proposition 2.2. Let (C, ∆C ) be a commutative C∗-quantum group with projec-
tion. Then C ∼= C0(G ⋉ H) for a semidirect product group, with the corresponding
comultiplication, and the projection on C comes from the group homomorphism
G ⋉ H → G ⋉ H, (g, h) 7→ (g, 1H); here G and H are locally compact groups and G
acts continuously on H by automorphisms. Conversely, any semidirect product
group gives a commutative C∗-quantum group with projection in this way.
Proof. Since C is commutative, C ∼= C0(K) for a locally compact group K. A quan-
tum group homomorphism from C to itself is equivalent to a group homomorphism
K → K, and the composition of quantum group homomorphisms also corresponds
to the composition of group homomorphisms. Thus a projection on C corresponds
to a group homomorphism p : K → K with p ◦ p = p. Let G ⊆ K and H ⊆ K
be the image and kernel of p, respectively; these are locally compact groups as
well. Since H is a normal subgroup, conjugation in K lets G ⊆ K act continuously
on H by automorphisms. The continuous maps m : G × H → K, (g, h) 7→ g · h,
and n : K → G × H, k 7→ (p(k), p(k−1)k), are inverse to each other and hence
homeomorphisms. The multiplication is given by
m(g1, h1) · m(g2, h2) = g1h1g2h2 = g1g2(g−1
2 h1g2)h2 = m(g1g2, (g−1
2 h1g2)h2).
Thus the homeomorphism m is also a group isomorphism K ∼= G⋉H. The converse
(cid:3)
assertion is routine to check.
Now we make Definition 2.1 explicit in several different ways, corresponding to
some of the equivalent characterisations of quantum group morphisms in [10]. First
we use unitaries satisfying pentagon equations.
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
5
Proposition 2.3. A C∗-quantum group with projection is given by a Hilbert space H
and two unitaries P, W ∈ U(H ⊗ H) that satisfy
W
13
W
23,
12
12 = W
12 = W
12
W
23
W
P23W
W
P13P23,
23,
23
P12 = P12P13W
P23P12 = P12P13P23
in U(H ⊗ H ⊗ H).
In addition, W is manageable as a multiplicative unitary.
All four equations in Proposition 2.3 are variants of the pentagon equation.
Proof. [10, Lemma 3.2] describes a quantum group morphism from H to itself by
a unitary P ∈ U(H ⊗ H) on the same Hilbert space H on which the manageable
multiplicative unitary W lives, subject to the two conditions (1.2), which are the
second and third equation in our statement. The first equation is the pentagon
equation for W. The fourth equation says that the quantum group endomorphism
(cid:3)
associated to P is idempotent by [10, Definition 3.5].
Our first goal is to prove the following structural result:
Proposition 2.4. Any idempotent endomorphism p : H → H of a C∗-quantum
group H splits. That is, there are a C∗-quantum group G and quantum group
morphisms a : G → H, b : H → G with a ◦ b = p and b ◦ a = idG.
The C∗-quantum group G is called the image of the idempotent endomorphism p.
We first construct this image, then we describe a and b and then we prove a ◦ b = p
and b ◦ a = idG. The proof of Proposition 2.4 will be finished by Lemma 2.7.
The fourth equation in Proposition 2.3 says that P is a multiplicative unitary.
Proposition 2.5. The multiplicative unitary P ∈ U(H ⊗ H) is manageable.
Proof. The multiplicative unitary W is manageable by assumption. This requires
the existence of certain auxiliary operators Q and fW. We use the same operator Q
for P. [28, Theorem 1.6] gives a unitary eP ∈ U(H ⊗ H) with
(cid:0)x ⊗ u P z ⊗ y(cid:1) =(cid:0)z ⊗ Qu eP x ⊗ Q−1y(cid:1)
for all x, z ∈ H, u ∈ D(Q) and y ∈ D(Q−1). Lemma A.2 shows that P commutes
with Q ⊗ Q. So eP and Q witness the manageability of the multiplicative unitary P
(see [28, Definition 1.2]).
(cid:3)
Proposition 2.5 shows that P generates a C∗-quantum group G = (A, ∆A), which
is called the image of P. Let G = ( A, ∆A) be its dual.
The unitary P is the image of a unitary multiplier P ∈ U( C ⊗ C) by [10, Lemma
3.2]. Hence slices of P are multipliers of C ⊆ B(H) and C ⊆ B(H), respectively.
These slices generate A and A, respectively, so A ⊆ M(C) and A ⊆ M( C).
Lemma 2.6. The embeddings i : A → M(C) and j : A → M( C) are bialgebra
morphisms G → H and G → H.
Proof. First we claim that i and j are C∗-algebra morphisms, that is, i(A) · C = C
and j( A) · C = C. The third condition in Proposition 2.3 is equivalent to
P∗
12
W
23
P12 = P13W
23
in U(H ⊗ H ⊗ H).
When we slice the first two legs on both sides by ω1 ⊗ ω2 for ω1, ω2 ∈ B(H)∗ and
close in norm, we get C = A · C. The same argument works for j.
6
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
The conditions in Proposition 2.3 also imply
W
23
W∗
12
P12W∗
P23W
23 = P12P13 = P23P12P∗
12 = P13P23 = P∗
12
23,
P23P12.
23 for all x ∈ D⊗C and (idD⊗∆A)(x) = P23x12P∗
Since (idD ⊗∆C)(x) = W
23
for all x ∈ D ⊗ A, the first equation says that id A ⊗ ∆C and id A ⊗ ∆A agree on P.
Since slices of P generate A, this implies ∆CA = ∆A, that is, i is a bialgebra
(cid:3)
morphism. So is j by a similar argument.
23x12W∗
The bialgebra morphisms i and j give quantum group morphisms
Vi = (id ⊗ i)(WA) ∈ U( A ⊗ C)
Vj = (j ⊗ id)(WA) ∈ U( C ⊗ A)
from A to C,
from C to A.
The quantum groups G and H may be generated by the multiplicative unitaries
P and W on the same Hilbert space H. Then the unitaries Vi and Vj are both
represented by the same unitary P on H ⊗ H; the conditions in Proposition 2.3
allow us to view P as a quantum group homomorphism G → H, H → G, H → H,
or as the identity quantum group homomorphism on G.
Lemma 2.7. The composite quantum group homomorphism Vi ◦ Vj : H → G → H
is the given projection P ∈ U( C ⊗ C) on H. The other composite G → H → G is
the identity on G.
Proof. The composition of quantum group homomorphisms is described in [10] by
a pentagon-like equation. The two claims in the lemma are both equivalent to the
(cid:3)
pentagon equation for P.
The description of a projection on a C∗-quantum group by a pair of bialgebra
morphisms (i, j) is unwieldy because it mixes quantum groups and their duals and
because the composition G → H → G is computed only indirectly.
The quantum group morphism H → G is usually not representable by a bialgebra
morphism C → A. We may, however, also represent the quantum group morphism j
by a bialgebra morphism ju : Cu → Au between the universal quantum groups, see
[10, Theorem 4.8]. Similarly, i lifts to a bialgebra morphism iu : Au → Cu. A
C∗-quantum group with projection is equivalent to a C∗-quantum group H with a
bialgebra morphism p : Cu → Cu satisfying p ◦ p = p by [10, Theorem 4.8]. Our
analysis above shows that for any such p there are a C∗-quantum group (A, ∆A) and
bialgebra morphisms ju : Cu → Au and iu : Au → Cu with p = iu ◦ ju and ju ◦ iu =
idA. Thus a quantum group with projection is equivalent to two C∗-quantum groups
with bialgebra morphisms ju : Cu → Au and iu : Au → Cu with ju ◦ iu = idA.
Next we replace j by right and left quantum group morphisms:
Proposition 2.8. A C∗-quantum group with projection is equivalent to two C∗-quan-
tum groups H = (C, ∆C ) and G = (A, ∆A) with morphisms i : A → C and
∆R : C → C ⊗ A such that the following diagrams commute:
∆A
∆C
A
i
C
A ⊗ A
C
i ⊗ i
∆C
C ⊗ C
C ⊗ C
∆R
C ⊗ A
idC ⊗ ∆R
∆C ⊗ idA
C ⊗ C ⊗ A
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
7
C
∆R
C ⊗ A
∆R
C ⊗ A
∆R ⊗ idA
idC ⊗ ∆A
C ⊗ A ⊗ A
∆A
∆R
A
i
C
A ⊗ A
i ⊗ idA
C ⊗ A
Another equivalent set of data is a pair of morphisms i : A → C and ∆L : C → A⊗C
with commutative diagrams
∆A
∆C
A
i
C
A ⊗ A
C
i ⊗ i
∆C
C ⊗ C
C ⊗ C
∆L
A ⊗ C
∆L ⊗ idC
idA ⊗ ∆C
A ⊗ C ⊗ C
C
∆L
A ⊗ C
∆L
A ⊗ C
idA ⊗ ∆L
∆A ⊗ idC
A ⊗ A ⊗ C
∆A
∆L
A
i
C
A ⊗ A
idA ⊗ i
A ⊗ C
Finally, the quantum group with projection is equivalent to a triple of morphisms
i : A → C, ∆R : C → C ⊗ A and ∆L : C → A ⊗ C satisfying all the above conditions
and, in addition,
C
∆C
C ⊗ C
∆C
C ⊗ C
∆R ⊗ idC
idC ⊗ ∆L
C ⊗ A ⊗ C
Then the following diagram also commutes:
C
∆R
C ⊗ A
∆L
A ⊗ C
∆L ⊗ idA
idA ⊗ ∆R
A ⊗ C ⊗ A
Proof. We have already seen that any projection on a C∗-quantum group H has an
image G and that there are a bialgebra morphism i : A → C and a quantum group
morphism j : H → G with j ◦ i = idG and i ◦ j = p, where p denotes the given
projection on H. Now we describe j by a right quantum group morphism ∆R as in
[10, Definition 5.1].
The first diagram above says that i is a bialgebra morphism. The second and
third diagram together say that ∆R is a right quantum group homomorphism
from C to A. The fourth diagram says that the composite A → C → A of these
quantum group morphisms is the identity map. Therefore, the other composite
C → A → C is idempotent, hence a projection. Thus i and ∆R give a projection
on H with image G. Conversely, any projection on a C∗-quantum group H has an
image by Proposition 2.4, which gives i and ∆R as above.
Replacing right by left quantum group morphisms shows that pairs (i, ∆L) as
above are also equivalent to C∗-quantum groups with projection. Of the two dia-
grams that relate ∆R and ∆L, the first one characterises when the right and left
quantum group homomorphisms ∆R and ∆L describe the same quantum group
morphism, and the second one commutes automatically, see [10, Lemma 5.7]. (cid:3)
8
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
Let A and B be C∗-algebras and T ∈ U(A ⊗ B). Then B is generated by T
in the sense of [27, Definition 4.1] if, for any representation ξ : B → B(H) and
any C∗-algebra C ⊂ B(H), the condition (idA ⊗ ξ)T ∈ M(A ⊗ C) implies that
ξ ∈ Mor(B, C).
Definition 2.9 ([4, Definition 3.2]). Let I = (C, ∆C ) and G = (A, ∆A) be quantum
groups. We call G a closed quantum subgroup of I in the sense of Woronowicz if
there is a bicharacter V ∈ U( C ⊗ A) that generates G.
In the situation of Proposition 2.8, (A, ∆A) is indeed a closed quantum subgroup
of (C, ∆C ) because the bicharacter (j ⊗ idA)(WA) ∈ U( C ⊗ A) generates A. This
is to be expected because (A, ∆A) is even a retract of (C, ∆C ) in the category of
quantum group morphisms.
2.1. Semidirect products. In this section, we are going to show that the semidi-
rect product construction in [12, Section 6] gives examples of C∗-quantum groups
with projection. Since we do not use this construction in the rest of the article,
we do not recall the notation and setup from [12]. Readers unfamiliar with the
semidirect product construction in [12] may skip this section.
Let G = (A, ∆A) be a C∗-quantum group. Let (B, β, β) be an A-Yetter -- Drinfeld
algebra, that is, β : B → B ⊗ A and β : B → B ⊗ A are continuous coactions of A
and A that satisfy the compatibility condition in [12, Definition 5.11]. The twisted
tensor product B ⊠ B = B ⊠W B is defined in [12]. We also require a coassociative
comultiplication ∆B : B → B ⊠B. Then [12, Theorem 6.8] describes a coassociative
comultiplication ∆C on C := A⊠B and shows that the C∗-bialgebra H = (C, ∆C ) is
bisimplifiable if (B, ∆B) is bisimplifiable. Furthermore, ∆C is injective if and only
if ∆B is injective. It is not studied in [12] when (C, ∆C ) is a C∗-quantum group: by
our definition, this would require a multiplicative unitary that generates it. If C is
unital, then this automatically exists and we are dealing with a compact quantum
group. In the non-compact case, we need some sort of multiplicative unitary for B
to get one for C.
For now, we disregard this issue. We want to describe a projection on H with
image G, and the description of projections in Proposition 2.8 makes sense in our
situation. Thus we are going to define morphisms
i : A → C,
∆R : C → C ⊗ A,
∆L : C → A ⊗ C
with the properties listed in Proposition 2.8. If we know for some reason that H is a
C∗-quantum group, that is, comes from a manageable multiplicative unitary, then
(i, ∆L, ∆R) as in Proposition 2.8 give a projection on H with image G. Actually, we
only need either ∆L or ∆R for this purpose. We provide both, however, and check
all conditions in Proposition 2.8.
The morphism i : A → A ⊠ B = C is the canonical embedding from the twisted
tensor product, which is denoted j1 or ιA in [12]. The right coaction ∆R : C → C⊗A
is the one constructed in [12, Lemma 6.5].
It is the unique one for which the
embeddings i = ιA : A → C and ιB : B → C are equivariant; that is,
∆R(ιA(a) · ιB(b)) = (ιA ⊗ idA)(∆A(a)) · (ιB ⊗ idA)(β(b)).
To construct ∆L, we equip A ⊗ A with the right A-coaction idA ⊗ ∆A on the
second tensor factor; this is a continuous A-coaction, and ∆A : A → A ⊗ A is an
A-equivariant morphism. Therefore, there is an A-equivariant morphism ∆A ⊠
idB : A ⊠ B → (A ⊗ A) ⊠ B. We let ∆L be the composite of ∆A ⊠ idB with the
isomorphism (A ⊗ A) ⊠ B ∼= A ⊗ (A ⊠ B) = A ⊗ C from [12, Lemma 3.14]. We may
also rewrite
A ⊠W B ∼= B ⊠bW
A
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
9
by [12, Proposition 5.1]. This is exactly the reduced crossed product for the
A-coaction on B by [12, Section 6.3]. After this identification, ∆L becomes the
dual coaction on the reduced crossed product as described in [12, Section 6.3].
Proposition 2.10. The morphisms i, ∆R and ∆L constructed above make all the
diagrams in Proposition 2.8 commute.
Proof. Of the ten diagrams in Proposition 2.8, the last one commutes automatically
if the others do, and the first and fifth one are the same. So we have to check eight
commuting diagrams. The maps ∆C , ∆R and ∆L are defined to have certain
composites with ιA and ιB:
∆C ◦ ιA = (ιA ⊗ ιA)∆A,
∆R ◦ ιA = (ιA ⊗ idA)∆A,
∆L ◦ ιA = (idA ⊗ ιA)∆A,
∆C ◦ ιB = Ψ23 ◦ ∆B,
∆R ◦ ιB = (ιB ⊗ idA)β,
∆L ◦ ιB = 1A ⊗ ιB,
where Ψ23 : B ⊠ B → C ⊗ C is the restriction of the map Ψ in [12, Proposition 6.6]
to the second two legs; that is, Ψ23j1(b) = (ιB ⊗ ιA)β(b) and Ψ23j2(b) = (1 ⊗ ιB)(b)
for all b ∈ B.
In particular, (ιA ⊗ ιA) ◦ ∆A = ∆C ◦ ιA says that the first and fifth diagram
commute, ∆R ◦ ιA = (ιA ⊗ idA)∆A says that the fourth diagram commutes, and
∆L ◦ ιA = (idA ⊗ ιA)∆A says that the eighth diagram commutes.
The remaining diagrams in Proposition 2.8 involve equalities of two maps defined
on C. Two maps f, f ′ defined on C are equal if and only if f ◦ ιA = f ′ ◦ ιA and
f ◦ ιB = f ′ ◦ ιB. For all remaining diagrams, it is trivial to check that they
commute after composing with ιA because of the explicit formulas above. The
third and seventh diagram do not involve ∆C , so the composites with ιB are also
given explicitly, which makes them trivial to check; in fact, they say simply that
∆R and ∆L are a right and a left coaction, respectively, which is already checked
in [12].
The condition on B for the sixth diagram is also trivial because ∆L only does
something complicated on ιA(A) and ∆C maps ιB(B) into ιB(B) ⊗ C.
For the second diagram, we must check (∆C ⊗ idA)∆RιB = (idC ⊗ ∆R)∆C ιB.
Since ∆B is A-equivariant, (∆B ⊗ idA) ◦ β = (β ⊲⊳ β) ◦ ∆B. Using the definition
of ∆C , we may rewrite our goal as (Ψ23 ⊗ idA)(β ⊲⊳ β)∆B = (idC ⊗ ∆R)Ψ23∆B.
From this, we may cancel the factor ∆B, so it suffices to check that
(Ψ23 ⊗ idA)(β ⊲⊳ β) = (idC ⊗ ∆R)Ψ23.
This is an equality of maps B ⊠ B → C ⊗ C ⊗ A, which we may check on both legs
separately. On the first leg, this reduces to the condition (idB ⊗ ∆A)β = (β ⊗ idA)β
that says that β is a coaction, and on the second leg this is trivial. This finishes
the proof that the second diagram commutes
In the condition from the ninth diagram on B, we may cancel the factor ∆B
from ∆C , so it suffices to check that (idC ⊗ ∆L)Ψ23 = (∆R ⊗ idC )Ψ23 as maps
B ⊠ B → C ⊗ A ⊗ C. This is once again checked separately on the two factors B.
So we must check that the maps idC ⊗ ∆L and ∆R ⊗ idC take the same values
both on (ιB ⊗ ιA)β(b) and on 1 ⊗ ιB(b) for all b ∈ B. This reduces to the coaction
condition for β on (ιB ⊗ ιA)β(b) and is trivial on 1 ⊗ ιB(b).
(cid:3)
3. Braided Multiplicative Unitaries
The definition of a braided multiplicative unitary is as complicated as the def-
inition of a braided C∗-quantum group. Recall that the latter is relative to a
C∗-quantum group G = (A, ∆A) which generates the braiding. The underlying
C∗-algebra B of a braided C∗-quantum group carries continuous coactions β and β
10
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
of G and G, respectively, which satisfy the Yetter -- Drinfeld compatibility condition
which characterises coactions of the quantum codouble of G. Finally, there is the
comultiplication ∆B : B → B ⊠ B, which is equivariant with respect to β and β
and coassociative. Thus a braided C∗-quantum group contains four coactions or
comultiplications ∆A, β, β, ∆B, which must satisfy seven algebraic conditions:
(1) ∆A is coassociative;
(2) β is a coaction of (A, ∆A);
(3) β is a coaction of ( A, ∆A);
(4) β and β satisfy the Drinfeld commutation relation, so that they give a
coaction of the quantum codouble;
(5) ∆B is equivariant with respect to the coaction β;
(6) ∆B is equivariant with respect to the coaction β;
(7) ∆B is coassociative.
The tensor product ⊠ is not symmetric unless G is trivial. Thus X ⊠′ Y := Y ⊠ X
gives another equally reasonable tensor product. We may also consider braided
quantum groups where the comultiplication takes values in B ⊠′ B instead of B ⊠ B.
Actually, these C∗-algebras are canonically isomorphic through the flip map, which
interchanges the two factors B. Thus there are two kinds of braided C∗-quantum
group, and taking the "coopposite," that is, composing ∆B with the flip map Σ
and leaving everything else the same, gives a bijection between the two types.
Remark 3.1. The definition above simplifies somewhat if G is quasitriangular. Then
a corepresentation β determines a corepresentation β so as to form a coaction of
the quantum codouble. Since β is a coaction constructed naturally from β, the
conditions (3), (4) and (6) above are redundant. A similar simplification occurs for
braided multiplicative unitaries. Since we are concerned with the general theory
here, we do not explore this situation any further.
When we turn to multiplicative unitaries, we replace C∗-algebras by Hilbert
spaces on which they act faithfully; comultiplications and coactions are replaced
by unitaries on appropriate tensor product Hilbert spaces that implement the coac-
tions through conjugation. So to specify a braided multiplicative unitary, we need
two Hilbert spaces and four unitaries that satisfy seven conditions, which corre-
spond to the seven conditions for the comultiplications and coactions listed above.
Moreover, there are two slightly different kinds of braided multiplicative unitaries,
depending on whether we use the "standard" braiding or its opposite; which braid-
ing is standard and which is opposite is, of course, a mere convention. The following
definition contains the details:
Definition 3.2. Let H and L be Hilbert spaces and let W ∈ U(H ⊗ H) be a
manageable multiplicative unitary; in particular, W satisfies the pentagon equation
(3.3)
W
23
W
12 = W
12
W
13
W
23.
A top-braided multiplicative unitary on L relative to W is given by unitaries
U ∈ U(L ⊗ H),
V ∈ U(H ⊗ L),
F ∈ U(L ⊗ L)
which satisfy the following conditions:
• U is a right corepresentation of W:
(3.4)
(3.5)
23
W
U12 = U12U13W
23
• V is a left corepresentation of W:
V13 V23
V23W
12 = W
12
in U(L ⊗ H ⊗ H);
in U(H ⊗ H ⊗ L);
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
11
• the corepresentations U and V are Drinfeld compatible:
(3.6)
U23W
13
V12 = V12W
13
U23
in U(H ⊗ L ⊗ H);
• F is invariant with respect to the right corepresentation U U := U13U23
of W on L ⊗ L:
(3.7)
U13U23F12 = F12U13U23
in U(L ⊗ L ⊗ H);
• F is invariant with respect to the left corepresentation V V := V13 V12
of W on L ⊗ L:
(3.8)
V13 V12F23 = F23 V13 V12
in U(H ⊗ L ⊗ L);
• F satisfies the top-braided pentagon equation
(3.9)
F23F12 = F12(
LL
)23F12(L L)23F23
in U(L ⊗ L ⊗ L);
LL
LL =
here the braiding
ZΣ for the flip Σ, x ⊗ y 7→ y ⊗ x, and the unique unitary Z ∈ U(L ⊗ L)
that satisfies
∈ U(L ⊗ L) and L L = (
)∗ are defined as
LL
(3.10)
Z13 = V23U∗
12
V∗
23
U12
in U(L ⊗ H ⊗ L).
A bottom-braided multiplicative unitary on L relative to W is given by the same
unitaries U, V, F satisfying (3.4) -- (3.8) and the bottom-braided pentagon equation
(3.11)
F23F12 = F12(L L)23F12(
LL
)23F23
in U(L ⊗ L ⊗ L).
Two corepresentations U and V on a Hilbert space L satisfying (3.6) are equiv-
alent to a corepresentation of the quantum codouble of the quantum group associ-
ated to W. It is shown in [12] that these corepresentations form a braided monoidal
category. Our conventions differ from those in [12] because we use a left corepre-
sentation V instead of the corresponding right corepresentation V := Σ V∗Σ. The
compatibility condition (3.6) and the definition of the braiding operator above are
equivalent to those in [12] up to this change of notation. The operator Z in (3.10)
L2
exists because W is manageable.
defined as above form a braiding on the tensor category of triples (L, U, V); the
operators L1 L2 give the opposite braiding.
It is shown in [12] that the operators
L1
In a braided monoidal category, the leg numbering notation should use the
)23F12(L L)23 or
)23 in the two braided pentagon equations (3.9) and (3.11). We
is
braiding operators. This explains why we replace F13 by (
(L L)23F12(
should also have replaced F23 by
defined as Z ′ΣL,L⊗L, where Z ′ is the unique operator on (L ⊗ L) ⊗ L with
L L⊗L; the braiding operator
F12
LL
LL
L⊗L
L
L
L⊗L
134 = ( V V)234U∗
Z ′
12( V V)∗
234
U12
in U(L ⊗ H ⊗ L ⊗ L).
Since we are dealing with a braided monoidal category, we also have
L
L⊗L
= LL
23
LL
12,
L⊗L = LL
L
12
LL
23.
Since F is invariant with respect to both corepresentations, it commutes with
any operator that is constructed in a natural way out of them, such as Z ′. This
implies
F23 = L⊗L
L
F12
L L⊗L = L L⊗LF12
L
L⊗L
,
so here the braiding has no effect. This also implies
LL
23F12
L L
23 = L L
12F23
LL
12,
L L
23F12
LL
23 = LL
12F23
L L
12.
12
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
Such equations are easier to digest as pictures:
F
=
F
F
=
F
LL
23F12
L L
23
L L
12F23
LL
12
L L
23F12
LL
23
LL
12F23
L L
12
The top-braided pentagon equation (3.9) uses the version of F13 where F acts on
the two top strands, whereas the bottom-braided pentagon equation (3.11) uses the
version of F13 where F acts on the two bottom strands; this explains our notation.
The braided pentagon equation is the usual pentagon equation if and only if F
commutes with ΣZΣ. Sufficient conditions for this are Z = 1, U = 1 or V = 1.
From now on, we restrict attention to top-braided multiplicative unitaries, so
braided multiplicative unitary means top-braided multiplicative unitary.
Definition 3.12. The dual of a braided multiplicative unitary (U, V, F) over W is
(V, U,bF) over cW, where cW := ΣW∗Σ, V := Σ V∗Σ, U := ΣU∗Σ, and
∈ U(L ⊗ L).
LL
bF := L LF∗
for (cW, V, U) is the opposite braiding L L for (W, U, V).
Therefore, the dual of the dual is the braided multiplicative unitary that we started
with, even if the braiding is not symmetric.
Proposition 3.13. Let (U, V, F) be a top-braided multiplicative unitary over W.
The braiding operator
LL
that V is a left corepresentation of W if and only if V is a right corepresentation
Its dual (V, U,bF) is a top-braided multiplicative unitary over cW := ΣW∗Σ.
Proof. It is well-known that the dualcW is again a multiplicative unitary, that U is
a right corepresentation of W if and only if U is a left corepresentation of cW, and
of cW. Routine computations show that the Drinfeld compatibility condition and
the invariance conditions are also preserved by the duality. The top-braided (or
bottom-braided) pentagon equation for the dual is equivalent to the top-braided (or
bottom-braided) pentagon equation for the original braided multiplicative unitary
(cid:3)
because the duality replaces the braiding by the opposite braiding.
Now we define when a braided multiplicative unitary (W, U, V, F) is manageable.
This requires W to be manageable, that is, there are a strictly positive operator Q
(3.14)
on H and a unitary fW ∈ U(H ⊗ H) with W∗(Q ⊗ Q)W = Q ⊗ Q and
(cid:0)x ⊗ u W z ⊗ y(cid:1) =(cid:0)z ⊗ Qu fW x ⊗ Q−1y(cid:1)
for all x, z ∈ H, u ∈ D(Q) and y ∈ D(Q−1) (see [28, Definition 1.2]). Here H is
the conjugate Hilbert space, and an operator is strictly positive if it is positive and
self-adjoint with trivial kernel. The condition W∗(Q ⊗ Q)W = Q ⊗ Q means that
the unitary W commutes with the unbounded operator Q ⊗ Q.
Definition 3.15. Let W ∈ U(H ⊗ H) be a manageable multiplicative unitary and
let Z and Q be as above. A braided multiplicative unitary (U, V, F) over W is
manageable if there are a strictly positive operator QL on L and a unitary eFeZ ∗ ∈
U(L ⊗ L) such that
(3.16)
(3.17)
(3.18)
(3.19)
U(QL ⊗ Q)U∗ = QL ⊗ Q,
V(Q ⊗ QL) V∗ = Q ⊗ QL,
F(QL ⊗ QL)F∗ = QL ⊗ QL,
(x ⊗ u Z ∗F y ⊗ v) = (y ⊗ QL(u) eFeZ ∗ x ⊗ Q−1
L (v))
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
13
for all x, y ∈ L, u ∈ D(QL) and v ∈ D(Q−1
L ).
and to clarify the manageability of the dual of a braided multiplicative unitary.
We have written eFeZ ∗ and not eF in (3.19) to make the formula more symmetric
We now describe the operator eZ that we want to use. The corepresentation U
of W on L induces a contragradient corepresentation on L. This is of the form eU∗,
where eU ∈ U(L ⊗ H) satisfies a variant of (3.14), see [28, Theorem 1.6] and [21,
Proposition 10]. Since eU∗ is a right corepresentation of W on L, there is a unique
unitary eZ ∈ U(L ⊗ L) that satisfies
eZ13 = V23eU12 V∗
We use this unitary in (3.19). Of course, it does not matter which unitary eZ we
use because we may absorb it in eF.
Proposition 3.21. The dual of a manageable braided multiplicative unitary is
again manageable.
in U(L ⊗ H ⊗ L).
23eU∗
(3.20)
12
Proof. Let (U, V, F) be a manageable top-braided multiplicative unitary over W,
let Z and eZ be as in (3.10) and (3.20). Let fW, Q witness the manageability of W
and let eF and QL witness the manageability of (U, V, F).
On L⊗H⊗L, both U12 and V23 commute with QL ⊗Q⊗QL by (3.16) and (3.17).
Hence so does Z by (3.10). Thus
(3.22)
Z(QL ⊗ QL)Z ∗ = QL ⊗ QL.
Together with (3.18), this implies that Z ∗F commutes with QL ⊗ QL. This together
L , compare the proof of
L ⊗ Q−1
Lemma A.2 or [28, Proposition 1.4.(1)].
with (3.19) implies that eFeZ ∗ commutes with QT
The unitaryeU commutes with QT
1.4.(1)]. This together with (3.17) and (3.20) implies
L⊗Q−1, compare Lemma A.2 or [28, Proposition
(3.23)
eZ(QT
L ⊗ Q−1
L )eZ ∗ = QT
L ⊗ Q−1
L ,
compare the proof of (3.22). Hence
(3.24)
eF(QT
because eFeZ ∗ commutes with QT
If y ∈ D(QL), x ∈ D(Q−1
L ⊗ Q−1
L )eF∗ = QT
L ⊗ Q−1
L
L ⊗ Q−1
L as well.
L ), and u, v ∈ L, then
(3.25)
(x ⊗ u Z ∗F y ⊗ v) = (QL(y) ⊗ u eFeZ ∗ Q−1
L (x) ⊗ v);
this is proved like [28, Proposition 1.4 (2)]. We rewrite this using the unitaries
eZ,ebF ∈ U(L ⊗ L) defined by
eZ :=(cid:0)ΣeZ ∗Σ(cid:1)T⊗T
,
ebF :=(cid:0)ΣeF∗Σ(cid:1)T⊗T
.
14
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
By definition, Z ∗bF = ΣF∗ZΣ and ebFeZ ∗ = (ΣeZeF∗Σ)T⊗T. Thus (3.25) gives
(x ⊗ u Z ∗bF y ⊗ v) = (x ⊗ u ΣF∗ZΣ y ⊗ v)
= (y ⊗ v ΣZ ∗FΣ x ⊗ u)
= (v ⊗ y Z ∗F u ⊗ x)
L (v) ⊗ x)
= (QL(u) ⊗ y eFeZ ∗ Q−1
= (Q−1
= (x ⊗ Q−1
L (v) ⊗ x eZeF∗ QL(u) ⊗ y)
L (v) ΣeZeF∗Σ y ⊗ QL(u))
= (y ⊗ QL(u) (ΣeZeF∗Σ)T⊗T x ⊗ Q−1
= (y ⊗ QL(u) ebFeZ ∗ x ⊗ Q−1
L (v)).
L (v))
Since the unitary Z for the dual braided multiplicative unitary becomes Z, the
(cid:3)
operators QL and ebF witness the manageability of bF.
3.1. Semidirect product multiplicative unitaries. In this section, we con-
struct a semidirect product multiplicative unitary WC and a projection P out of a
braided multiplicative unitary (U, V, F) over a multiplicative unitary W. We show
that the semidirect product multiplicative unitary WC is manageable if the braided
multiplicative unitary (U, V, F) is manageable.
The formulas and proofs below are explicit but lengthy because all four unitaries
W, U, V, F must enter in the definitions of WC and P and all seven conditions on
them must be used in the proofs.
Theorem 3.26. Let (U, V, F) be a braided multiplicative unitary over a multiplica-
tive unitary W. Define WC
(3.27)
(3.28)
1234, P ∈ U(H ⊗ L ⊗ H ⊗ L) by
F24 V34,
WC
1234 := W
P1234 := W
13
13
U23 V∗
34
U23.
Then WC and P satisfy the four pentagon-like equations in Proposition 2.3. Thus
they give a C∗-quantum group with projection when WC is manageable.
Proof. We first verify the pentagon equation (3.3) for WC
1234. Let
XXX = WC
3456
WC
1234(WC )∗
3456.
We will rewrite this in several steps using the conditions in Definition 3.2. We use
{. . .} to highlight which part of the formula we are modifying in the following step.
Definition (3.27) gives
XXX = W
35{U45 V∗
56
F46 V56 and W
Since U45 V∗
56
U23 commute,
13
F46 V56}{W
13
U23} V∗
34
F24 V34 V∗
56
F∗
46
V56U∗
45
W∗
35.
XXX = {W
35
W
13}U23U45 V∗
56
F46{ V56} V∗
34
F24 V34{ V∗
56}F∗
46
V56U∗
45
W∗
Now we use the pentagon equation (3.3) for W and commute V56 with V∗
34
W∗
34}F24{ V34F∗
46} V56U∗
56{F46 V∗
45
XXX = W
U23}U45 V∗
Equations (3.4) and (3.8) turn this into
15{W
W
35
13
35.
35.
F24 V34:
W
13
W
15
U23U25W
35
U45{ V∗
56
V∗
34} V∗
36
F46{ V36F24 V∗
36}F∗
46
V36{ V34 V56}U∗
45
W∗
35.
Commuting V56 with V34 and V36 with F24 gives
XXX = W
13
W
15
U23U25{W
35
U45 V∗
34} V∗
56
V∗
36
F46F24F∗
46
V36 V56{ V34U∗
45
W∗
35}.
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
15
Now (3.6) gives
XXX = W
13
W
15
U23U25 V∗
34
U45{W
35
V∗
56
We transform this using (3.5):
V∗
36}F46F24F∗
46{ V36 V56W∗
35}U∗
45
V34.
XXX = W
13
15
W
U23U25 V∗
34
35 with F46F24F∗
46:
We commute W
U45 V∗
56{W
35}F46F24F∗
46{W∗
35} V56U∗
45
V34.
XXX = W
13
W
15
U23U25 V∗
34
U45 V∗
56{F46F24F∗
46} V56U∗
45
V34.
Now we use the braided pentagon equation (3.9) and the definition of the braiding
through Z:
F24}Z46F26Z ∗
46
V56U∗
45
V34.
XXX = W
13
W
15
Now we commute U25 with V∗
U23{U25 V∗
34, V∗
56
34}U45{ V∗
56 with F24:
U25U45F24{ V∗
XXX = W
13
W
15
U23 V∗
34
56Z46}F26{Z ∗
46
V56U∗
45} V34.
Equation (3.10) implies U45 V∗
56Z46 = V∗
56
U45, so this becomes
XXX = W
13
W
15
U23 V∗
34{U25U45F24U∗
45} V∗
56{U45F26U∗
45} V56 V34.
Now we use (3.7) and commute F26 with U45:
XXX = W
15}{U23 V∗
34
13{W
15 with U23 V∗
U23 V∗
34
F24 V34}{W
34
F26 V56}{ V34}.
F24}{U25 V∗
56
U25 V∗
56
15
F26 V56} = WC
1234
WC
1256.
F26 V56 with V34 to get
Finally, we commute W
XXX = {W
13
This is the desired pentagon equation for WC
Next we show that P satisfies the pentagon equation:
P3456P1234P3456 = W
= W
= P1234P1256.
U45W
W
W
13
35
15
13
W∗
U23U∗
45
U23W∗
35 = W
35 = W
W
13
35
W
U23W∗
35
13
U23U25 = W
15
35
U23W
15
U25
13
F24 and W
U25 V∗
56
1234.
15
The first and last equalities are the definition of P; the second step commutes U45
with W
U23; the third step uses the pentagon equation (3.3) for W; the fourth
step uses (3.4); the fifth step commutes W
13
15 with U23.
Next we prove P3456WC
1234 = WC
1234
P1256P3456 or, equivalently, P3456WC
P∗
3456 =
1234
WC
1234
P1256:
P3456WC
1234
P∗
3456 = W
= W
U45W
35
W
13
35
13
U23 V∗
34
U23U45 V∗
34
= W
13
W
15
W
35
U23U45 V∗
34
F24 V34U∗
45
F24 V34U∗
45
W∗
35
W∗
35
F24 V34U∗
45
W∗
35
= W
13
W
15
= W
13
= W
13
= W
13
= W
13
U23W
U23 V∗
34
U23 V∗
34
U23 V∗
34
U23U25W
U25 V∗
34
15
35
U45 V∗
34
U45W
35
W
15
W
15
U25U45F24U∗
45
F24U25 V34
W∗
35
V34
F24 V34U∗
45
U∗
F24W∗
45
35
V34
F24 V34W
15
U25 = WC
1234
P1256.
The first and last equalities are the definitions of WC and P; the second, sixth,
and eighth steps commute unitaries in different legs; the third step uses the penta-
gon equation (3.3) for W; the fourth step uses (3.4); the fifth step uses (3.6) and
commutes unitaries in different legs; the seventh step uses (3.7).
16
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
Finally, we prove WC
P∗
1234
WC
3456
3456 by computing P∗
1234
WC
3456
P1234:
P1234 = P1234P1256WC
U45 V∗
56
U45 V∗
56
W∗
13
W∗
13
W
W
W
35
35
13
3456
P1234 = U∗
23
= U∗
23
= U∗
23
F46 V56W
13
U23
F46 V56U23
W
15
U∗
23
W
35
W
35
= W
15
= W
15
U25W
35
F46 V56U23
F46 V56
U45 V∗
56
U23U45 V∗
56
U45 V∗
56
F46 V56 = P1256WC
3456.
The first and last steps are the definitions of WC and P; the second and fourth steps
commute unitaries in different legs; the third step uses the pentagon equation (3.3)
(cid:3)
for W; the fifth step uses (3.4).
Theorem 3.29. Let W be a manageable multiplicative unitary and let (U, V, F)
be a manageable braided multiplicative unitary over W. Then the multiplicative
unitaries WC := W
F24 V34 and P := W
U23 are manageable.
U23 V∗
34
13
13
right corepresentation of W by the same argument; in particular, it works for U, so
Proof. Let fW and Q witness the manageability of W, and let eF and QL witness
the manageability of F. The construction of the unitary eU in (A.1) works for any
we get eU ∈ U(L ⊗ H) with
(3.30)
for all x, z ∈ L, u ∈ D(Q) and y ∈ D(Q−1).
Let QC := Q ⊗ QL ∈ U(H ⊗ L) and
(cid:10)x ⊗ u(cid:12)(cid:12)U(cid:12)(cid:12)z ⊗ y(cid:11) =(cid:10)z ⊗ Qu(cid:12)(cid:12)eU(cid:12)(cid:12)x ⊗ Q−1y(cid:11)
1234 :=fW13 V∗
fWC
34eF24eU23 V34 ∈ U(H ⊗ L ⊗ H ⊗ L).
(3.31)
We claim that these operators witness the manageability of WC. It is clear that QC
is strictly positive.
The operators W
13, V34, U23 and F24 all commute with QC ⊗ QC = Q ⊗ QL ⊗
Q ⊗ QL by the manageability assumptions. Hence WC commutes with QC ⊗ QC. It
remains to check (3.14) for WC , fWC and QC. We relegate this technical computa-
tion to Lemma A.5 in the appendix. This finishes the proof that WC is manageable.
(cid:3)
Now Proposition 2.5 shows that P is manageable as well.
3.2. Analysis of a quantum group with projection. In this section, we con-
struct a braided multiplicative unitary from a quantum group with projection. Our
starting point is a Hilbert space H with two unitaries WC , P ∈ U(H ⊗ H) satisfying
the conditions in Proposition 2.3. We must construct another Hilbert space L with
operators U ∈ U(L ⊗ H), V ∈ U(H ⊗ L) and F ∈ U(L ⊗ L) as in Definition 3.2.
In particular, the corepresentations U and V form a Drinfeld pair for the multi-
plicative unitary P. The simplest general construction of such a Drinfeld pair lives
on the tensor product Hilbert space L := H ⊗ H, where H denotes the conjugate
Hilbert space of H. Therefore, we will use this rather large Hilbert space.
Let G = (A, ∆A) be the C∗-quantum group generated by P, which is manageable
by Proposition 2.5. Let H = (C, ∆C ) be the C∗-quantum group generated by the
manageable multiplicative unitary WC . By construction, we have inclusion maps
ι : C → B(H) and ι : C → B(H), which are non-degenerate ∗-homomorphisms. The
reduced bicharacter is the unique unitary WC ∈ U( C⊗C) with WC = (ι⊗ι)(WC ) or,
ιι. By construction, A ⊆ M(C) and A ⊆ M( C) as C∗-subalgebras
briefly, WC = WC
of B(H).
The representations (ι, ι) form a Heisenberg pair for the quantum group (C, ∆C )
in the notation of [11]. This Heisenberg pair generates an anti-Heisenberg pair
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
17
α : C → B(H), α : C → B(H) by [11, Lemma 3.6]. Thus
(3.32)
WC
1αWC
α3 = WC
α3WC
13WC
1α
in U( C ⊗ K(H) ⊗ C).
The restriction of a Heisenberg or anti-Heisenberg pair for H to G remains a Heisen-
berg or anti-Heisenberg pair, respectively. Thus
in U( A ⊗ K(H) ⊗ A).
P1αP α3 = P α3P13P1α
(3.33)
To make computations shorter, we shall use leg numbering notation such as Pij , WC
U(H ⊗ H ⊗ H ⊗ H) for 1 ≤ i < j ≤ 4. This means the unitary acting on the ith
and jth tensor factor by applying the appropriate representations of C or C to the
two legs of P or WC , respectively. For instance, P12 = (α ⊗ ι)(P) ⊗ 1H⊗H. This
notation is not ambiguous if we also specify the Hilbert space on which the operator
acts because we have given one representation of C and C on H and H each. We
let
ij ∈
(3.34)
(3.35)
(3.36)
U := P23P13 := (ι ⊗ ι)P23 · (α ⊗ ι)P13
V := P12P13 := (ι ⊗ α)P12 · (ι ⊗ ι)P13
F := P∗
24WC
24WC
14
14P∗
in U(H ⊗ H ⊗ H),
in U(H ⊗ H ⊗ H),
in U(H ⊗ H ⊗ H ⊗ H).
Theorem 3.37. The unitaries P ∈ U(H ⊗ H), U ∈ U(L ⊗ H), V ∈ U(H ⊗ L), and
F ∈ U(L ⊗ L) form a braided multiplicative unitary.
The proof of Theorem 3.37 will take some work. The precise formulas for α
and α will only matter in the end when we check the manageability of our braided
multiplicative unitary. Thus our construction really uses that WC is manageable
(or at least modular). The pentagon equation (3.3) for P holds by assumption.
Equations (3.4) and (3.5), which say that U and V are corepresentations, amount
to
P34P23P13 = P23P13P24P14P34
in U(H ⊗ H ⊗ H ⊗ H),
P23P24P12 = P12P13P14P23P24
in U(H ⊗ H ⊗ H ⊗ H).
We get both equations using the pentagon equation for P twice, in legs where the
representations of H and hence of G form a Heisenberg pair. Thus U and V are
corepresentations of P. The Drinfeld compatibility condition (3.6) becomes
P34P24P14P12P13 = P12P13P14P34P24
in U(H ⊗ H ⊗ H ⊗ H ⊗ H).
The anti-Heisenberg and Heisenberg properties of our representations on the second
and third leg give P24P14P12 = P12P24 and P13P14P34 = P34P13. Hence
P34P24P14P12P13 = P34P12P24P13 = P12P34P13P24 = P12P13P14P34P24
as needed. Thus U and V are Drinfeld compatible. The equivariance of F with
respect to U in (3.7) amounts to
14P∗
P25P15P45P35P∗
14P25P15P45P35
14 = P∗
24WC
24WC
24WC
24WC
14P∗
(3.38)
in U(H ⊗ H ⊗ H ⊗ H ⊗ H). Since we have a Heisenberg pair on the fourth leg, we
may use the conditions in Proposition 2.3 to simplify
14P15P45) = WC
14P25P15P45 = WC
24P25P45WC
14 = P45WC
24WC
14.
WC
24WC
24P25(WC
Hence the right side in (3.38) becomes
P∗
14P∗
24WC
24WC
14P25P15P45P35 = P∗
14P∗
24P45WC
24WC
14P35.
Plugging this in and cancelling WC
24WC
14P35, we see that (3.38) is equivalent to
P25P15P45P∗
14P∗
24 = P∗
14P∗
24P45
or P24P25P14P15P45 = P45P24P14.
18
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
Since we have Heisenberg pairs on the second and fourth legs, this follows by two
applications of the pentagon equation for P.
The condition (3.8) about F being equivariant with respect to V becomes
(3.39)
P14P15P12P13P∗
25P∗
35WC
35WC
25 = P∗
25P∗
35WC
35WC
25P14P15P12P13
in U(H ⊗ H ⊗ H ⊗ H ⊗ H). Since we have an anti-Heisenberg pair on the second and
a Heisenberg pair on the third leg, we get WC
35 =
WC
35P13. Thus the right side of (3.39) becomes
25P15P12 = P12WC
25 and P13P15WC
P∗
25P∗
= P14P∗
= P14P∗
35WC
25P∗
25P∗
35WC
25P14P15P12P13
35(WC
35WC
35P12(WC
25P15P12)P13 = P14P∗
25P∗
35P13)WC
25 on the right to transform our
25P∗
35P12WC
35P12P13P15WC
25 = P14P∗
35WC
25P13
35WC
25.
35WC
Now we may cancel P14 on the left and WC
condition into P15P12P13P∗
35 = P∗
25P∗
25P∗
35P12P13P15 or, equivalently,
P35P25P15P12P13 = P12P13P15P35P25
in U(H ⊗ H ⊗ H ⊗ H ⊗ H). The anti-Heisenberg pair condition on the second leg
gives P25P15P12 = P12P25, the Heisenberg pair on the third leg gives P13P15P35 =
P35P13. Plugging this in, our condition becomes
P35P12P25P13 = P12P35P13P25,
which is manifestly true. Thus our operators satisfy (3.8) as well.
Checking the braided pentagon equation (3.9) is a long computation. We may
omit it because of the following trick. In the proof of Theorem 3.26, the braided
pentagon equation is used exactly once. Therefore, if all the other conditions in
Definition 3.2 hold, then the braided pentagon equation (3.9) is both sufficient
and necessary for the usual pentagon equation for the unitary WD constructed
in Theorem 3.26. Thus the proof of Theorem 3.37 is finished up to the braided
pentagon equation, and this follows from the proof of the following theorem.
Theorem 3.40. Let WC , P ∈ U(H⊗H) define a C∗-quantum group with projection
as in Proposition 2.3. Construct a braided multiplicative unitary (P, U, V, F) on the
Hilbert space L = H ⊗ H with the unitaries defined in (3.34) -- (3.36). From this,
construct a multiplicative unitary WD with a projection PD on the Hilbert space
H ⊗ L ⊗ H ⊗ L ∼= H ⊗ H ⊗ H ⊗ H ⊗ H ⊗ H
by Theorem 3.26.
The braided multiplicative unitary (P, U, V, F) and the multiplicative unitary WD
are manageable. And WD generates the same C∗-quantum group as WC. The
isomorphism between these C∗-quantum groups maps P to PD.
Roughly speaking, going from a C∗-quantum group with projection to a braided
C∗-quantum group and back gives an isomorphic C∗-quantum group with projec-
tion.
The definitions in Theorem 3.26 amount to
45P∗
26P∗
WD := P14P34P24P∗
= P14P34P24P∗
46P∗
46P∗
26P∗
36WC
36WC
36WC
36WC
26P46,
26P45P46
PD := P14P34P24.
Our first task is to construct representations π and π of C and C that form a
Heisenberg pair and that satisfy (π ⊗ π)WC = WD. This implies that WD satisfies
the pentagon equation. As we remarked above, this implies the braided pentagon
equation (3.9) for F, which still remained to be proven.
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
19
Lemma 3.41. There is a representation π′ : C → B(H ⊗ H ⊗ H) such that
(3.42)
(3.43)
(id C ⊗ π′)WC = P12WC
(id C ⊗ π′)P = P12P14
14
in U( C ⊗ K(H ⊗ H ⊗ H)),
in U( C ⊗ K(H ⊗ H ⊗ H)).
Proof. Let P = σ(P)∗ ∈ U(C ⊗ C) for the flip σ. Define ϕ : C → M(C ⊗ K(H)) by
ϕ(c) := P1 α(1 ⊗ α(c))P∗
1 α. Then
(id C ⊗ ϕ)WC = P2 αWC
1α
P∗
2 α = σ23(P∗
α3WC
1αP α3)
in M( C ⊗ C ⊗ K(H)).
1αP α3 = P α3P13WC
The second condition in Proposition 2.3 is equivalent to WC
1α in
U( C ⊗ K(H) ⊗ C) because we have an anti-Heisenberg pair on the second leg. Thus
(id C ⊗ ϕ)WC = σ23(P13WC
1α) = P12WC
1α
in M( C ⊗ C ⊗ K(H)).
We may define π′(c) := (cid:0)(ι ⊗ ι ◦ α−1)ϕ(c)(cid:1)13 because α is automatically injective
(see [18, Proposition 3.7]). This is the unique representation that satisfies (3.42).
Replacing WC by P in the above computations gives (3.43).
(cid:3)
Lemma 3.44. Let π′ be as in the previous lemma. The pair of representations
(π, π) of C and C on H ⊗ H ⊗ H defined by
π(c) := P∗
13π′(c)P13,
π(c) := P∗
13((α ⊗ ι) ∆C (c))23P13
is an H-Heisenberg pair.
Proof. Let π′(c) := ((α ⊗ ι) ∆C (c))23. The lemma is equivalent to (π′, π′) being
H-Heisenberg. Recall that ( ∆C ⊗ idC )WC = WC
13. Lemma 3.41 gives
23WC
WC
π′5WC
1π′ = WC
45WC
35P12WC
14 = P12WC
45WC
14WC
35
in U( C ⊗ K(H ⊗ H ⊗ H) ⊗ C). Since we have a Heisenberg pair on the fourth leg,
the pentagon equation (3.3) gives
45WC
P12WC
14WC
35 = P12WC
15WC
1π′ WC
15WC
45WC
35,
14WC
π′5 in U( C ⊗K(H⊗H⊗H)⊗C). (cid:3)
which is equivalent to WC
π′5WC
1π′ = WC
Lemma 3.45. WD = (π ⊗ π)WC and PD = (π ⊗ π)P in U(H ⊗ L ⊗ H ⊗ L).
(π ⊗ π)WC = P∗
Proof. Cancelling P45 gives WD = P14P34P24P∗
as in the proof of Lemma 3.44, we get
36P24WC
26P13P46 = P∗
Since we have a Heisenberg pair on the third leg, WC
13P34P24P13WC
(π ⊗ π)WC = P∗
46P34WC
13P∗
46P∗
46P∗
13P34P24WC
46P∗
36P13 = P13P16WC
36WC
26P46.
36P13WC
26P46.
36. Hence
26P∗
36WC
36WC
26P46. Computing
Since we have Heisenberg pairs on the third and fourth legs, P34P13 = P13P14P34
and P46Pi4 = Pi6Pi6P46 for all i = 1, 2, 3. Using these identities (for the expressions
within brackets in the computation below) we get
P∗
46P∗
13P34P24P13P16 = P∗
46(P∗
13P34P13)P24P16 = P∗
46P14P34P24P16
= (P∗
46P14P16)P34P24 = P14(P∗
46P34)P24 = P14P34P∗
= P14P34(P∗
46P24)P∗
46P∗
36P24
36 = P14P34P24P∗
46P∗
26P∗
36.
The last two computations together give
(π ⊗ π)WC = P14P34P24P∗
46P∗
26P∗
36WC
36WC
26P46 = WD.
Equation (3.43) allows a similar computation with P instead of WC . This gives
(π ⊗ π)P = P14P34P24P∗
46P∗
26P∗
36P36P26P46 = PD.
(cid:3)
20
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
The following remarks apply to any Heisenberg pair (π, π) for a C∗-quantum
group H = (C, ∆C ) on a Hilbert space H′. Being a Heisenberg pair means that
(π ⊗ π)WC is a multiplicative unitary. It is unclear, in general, whether this mul-
tiplicative unitary is manageable.
If it is manageable, then we claim that the
C∗-quantum group that it generates is isomorphic to the one we started with. The
representations in a Heisenberg pair are automatically faithful by [18, Proposition
3.7]. Hence we may view C and C as subalgebras of B(H′), and (π ⊗ π)WC is
a unitary multiplier of C ⊗ C ⊆ B(H′ ⊗ H′).
It makes no difference whether
we take slices on the first leg with elements of B(H′)∗ or C ∗: both generate the
same C∗-subalgebra of B(H′), namely, π(C). The comultiplication on the quantum
group generated by (π ⊗ π)WC is defined so that the isomorphism π is a Hopf
∗-homomorphism. Thus the C∗-quantum group generated by (π ⊗ π)WC is isomor-
phic to H for any Heisenberg pair for which (π ⊗π)WC is manageable. Furthermore,
Lemma 3.45 shows that this Hopf ∗-isomorphism maps P to PD, so we also get the
same projection on our C∗-quantum group.
Thus the proof of Theorem 3.40 will be finished once we show that WD and the
braided multiplicative unitary (U, V, F) are manageable. By Theorem 3.29, WD is
manageable once (U, V, F) is manageable. So it remains to prove this.
The braiding on L ⊗ L comes from the unique unitary Z that verifies (3.10). A
23 in U(H ⊗ H ⊗ H ⊗ H) does the
14P∗
24P∗
13P∗
simple computation shows that Z = P∗
job. This gives
Z ∗F = P23P13P24P14P∗
14P∗
24WC
24WC
14 = P23P13WC
24WC
14.
Now we use that (ι, ι) is the standard Heisenberg pair, generated by WC, and
that the anti-Heisenberg pair (α, α) is constructed as in [11, Lemma 3.6]; that is,
α(a) := aT◦ι◦ RC and α(a) := aT◦ι◦RC . Thus
Z ∗F = Pι⊗T◦ι◦RC
23
PT◦ι◦ RC ⊗T◦ι◦RC
13
(WC
24)ι⊗ι(WC
14)T◦ι◦ RC ⊗ι.
U(H ⊗ H), see Appendix A. Since P is manageable by Proposition 2.5, so is the
Let QC and gWC ∈ U(H ⊗ H) witness the manageability of WC = (ι ⊗ ι)WC ∈
dual bP = ΣP∗Σ. This is witnessed by a certain unitary ebP ∈ U(H ⊗ H). We have
(WC )TιR C ⊗ι = (WC )Tι⊗ιRC = (fWC )∗ and PTιR C ⊗TιRC = PT⊗T by [28, Theorem
1.6 (5)] and [10, (19)]. Similarly, (ΣPι⊗T◦ι◦RC Σ)∗ = PT◦ι◦RC ⊗ι =ebP. Thus
in U(H ⊗ H ⊗ H ⊗ H).
WC
This unitary and Q witness the manageability of the braided multiplicative unitary
(U, V, F). The rather technical proof of this fact is relegated to the appendix, see
Lemma A.8.
4. Examples of Quantum Groups with Projections
The simplest examples of semidirect products of connected Lie groups are E(2) =
R2 ⋊ T and the real and complex ax + b-groups R ⋊ R×
>0 and C ⋊ C×, where the
second, multiplicative factor acts by multiplication on the first, additive factor. The
group E(2) is the group of isometries of the plane. Another very important example
is the Poincaré group, the semidirect product of the Lorentz group with R4.
When quantising such groups, one may try to preserve the semidirect product
structure, that is, construct C∗-quantum groups with projection. For instance, the
Let Q := QT
C ⊗ QC. Then Q ⊗ Q commutes with F, Q ⊗ QC commutes with U, and
QC ⊗ Q commutes with V. Define eF ∈ U(L ⊗ L) by
14)∗(P∗)T⊗T
(3.46)
in U(H ⊗ H ⊗ H ⊗ H).
13
Z ∗F = Σ23ebP23Σ23PT⊗T
eF :=fWC
24(WC
14)∗
24(fWC
13
23 ePT⊗T
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
21
quantum E(2) groups by Woronowicz [26] have obvious morphisms to the circle
group T and back that compose to the identity on T. The quantum az + b groups
introduced by Woronowicz [29] and Sołtan [20] -- which deform C ⋊ C× -- have
q = qZ+iR ⊆ C× (with multiplication as group
obvious morphisms to the group C×
structure) and back, which compose to the identity on C×
q ; see also [6, Example
3.7]. The quantum ax + b group by Woronowicz and Zakrzewski [31] has an obvious
projection onto the group R×
>0
∼= R.
There are also quantum versions of semidirect product groups that appear to
have no such projection. This includes the az + b groups by Baaj and Skandalis,
see [25, Section 5.3], the ax+b groups by Stachura [23] and the κ-Poincaré groups by
Stachura [22]. These examples are all constructed using the formalism of quantum
group extensions of [25]. Quantum group extensions are compared with quantum
groups with projection in [7].
As an example of our theory, we are going to construct a braided multiplicative
unitary that generates "simplified quantum E(2)," a variant of quantum E(2) also
due to Woronowicz (unpublished); whereas the quantum E(2) groups in [26] deform
a double cover of E(2), the simplified variants deform E(2) itself. A common feature
of simplified quantum E(2) and the quantum groups with projection mentioned
above is that the image of the projection is a classical, Abelian group. This is to
be expected when deforming semidirect products by Abelian groups because these
cannot be deformed to quantum groups in interesting ways. We begin by observing
some common features of braided multiplicative unitaries in case W generates an
Abelian group G.
Let G be the dual group. The corepresentations U and V in a braided multiplica-
tive unitary are equivalent to representations of G and G on the Hilbert space L,
respectively. The compatibility condition (3.6) for U and V says here that the repre-
sentations of G and G commute. Thus we may combine them to one representation
of G × G on L.
We can further normalise this representation because the left regular representa-
tion of any quantum group absorbs every other representation. The operator F ⊗ 1
on L ⊗ L2( G × G) is a braided multiplicative unitary if and only if F is, and it
generates an equivalent semidirect product quantum group. Thus we may assume
without loss of generality that our representation is a multiple of the left regular
representation:
L = L2( G × G) ⊗ L0
for some separable Hilbert space L0, with C∗( G × G) acting only on the first tensor
factor, by the regular representation. We may identify C0(G × G) ∼= C∗( G × G) and
L2( G × G) ∼= L2(G × G) by the Fourier transform, and the regular representation
of C∗( G × G) on L2( G × G) ∼= L2(G × G) becomes the standard representation of
C0(G × G) on L2(G × G) by pointwise multiplication.
For some examples, a variant of the above is useful: if the representation of G×G
on L factors through an Abelian locally compact group H, then we may use H in-
stead of G× G in the above simplification. That is, we seek a braided multiplicative
unitary on the Hilbert space L2(H) ⊗ L0 with G × G acting only on the first ten-
sor factor, through the regular representation of H and the given homomorphism
G × G → H.
For instance, the compact quantum group Uq(2) is a semidirect product of the
braided quantum group SUq(2) by the circle T (see [5]), and the relevant represen-
tation of Z × T factors through a homomorphism Z × T → T, (n, z) 7→ λn · z for
some λ ∈ T. For the quantum az + b groups, G = Cq is self-dual, and the relevant
22
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
representation of G × G factors through the map G × G ∼= G × G → G, where the
second map is the multiplication map (x, y) 7→ x · y.
When we have simplified L to L2(G× G)⊗L0 with C0(G× G) acting by pointwise
is the operator of pointwise multiplication
LL
multiplication, the braiding operator
with the circle-valued function
(g1, χ1, g2, χ2) 7→ χ1(g2).
(4.1)
The conditions (3.7) and (3.8) for F mean that F ∈ U(L ⊗ L) is a G × G-equivariant
operator with respect to the tensor product representation of G × G on L ⊗ L. In
terms of the above spectral analysis, f ∈ C0(G × G) acts on L ⊗ L by pointwise
multiplication on each fibre with the function
∆∗f (g1, χ1, g2, χ2) = f (g1g2, χ1χ2)
for all (g1, χ1, g2, χ2) ∈ G × G × G × G. An operator on L ⊗ L is G × G-equivariant
if and only if it commutes with the operators of pointwise multiplication by ∆∗f
for f ∈ C0(G × G).
Summing up,
it suffices to look for braided multiplicative unitaries over an
Abelian group G on the Hilbert space L = L2(G × G, ℓ2(N)). Such a braided
multiplicative unitary F ∈ U(L ⊗ L) must commute with the operators of pointwise
multiplication by functions in ∆∗Cb(G × G) ⊆ Cb(G × G × G × G), and it must sat-
isfy the braided pentagon equation (3.9), where the braiding is given by pointwise
multiplication with the function in (4.1).
4.1. Simplified quantum E(2) groups. Now we specialise to simplified quan-
tum E(2) groups. They were already treated in [17] except for the manageability
of the braided multiplicative unitary in question. A C∗-algebraic version of this
construction appears in [19].
Here the image of the projection is the circle group G = T, so G = Z. The anal-
ysis above suggests to construct a braided multiplicative unitary for quantum E(2)
on a Hilbert space of the form L2(T × Z) ⊗ L0. Actually, we shall not need L0 and
work on L2(T × Z) itself. First we describe the standard multiplicative unitary W
generating T.
Let H := ℓ2(Z) and let {ep}p∈Z be an orthonormal basis of H. Define
uep := ep+1
and
N ep := pep.
The shift u is unitary and generates the regular representation of Z on H. The
operator N is self-adjoint with spectrum Z, and the resulting representation of
C0(Z) ∼= C∗(T) is the Fourier transform of the regular representation of T on L2(T).
These operators generate a representation of the crossed product C0(Z) ⋊ Z ∼=
K(ℓ2Z). That is, they satisfy the commutation relation
(4.2)
u∗ N u = N + 1.
The multiplicative unitary generating T is
N ⊗1 =ZZ×T
W := (1 ⊗ u)
zs dE N (s) ⊗ dEu(z),
ek ⊗ el 7→ ek ⊗ el+k,
where dE N and dEu are the spectral measures of N and u, respectively. The
commutation relation (4.2) implies the pentagon equation
(1 ⊗ 1 ⊗ u)1⊗ N ⊗1(1 ⊗ u ⊗ 1)
N ⊗1⊗1(1 ⊗ 1 ⊗ u∗)1⊗ N ⊗1 = (1 ⊗ u ⊗ u)
N ⊗1⊗1.
for W, that is, W is a multiplicative unitary. Simple computations show that
Q = 1 and fW := (1 ⊗ u∗) N ⊗1 witness the manageability of W. Slices of W
in the first and second leg clearly generate the C∗-algebras C∗( N ) ∼= C0(Z) and
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
23
C∗(u) ∼= C0(T). The comultiplications defined by W satisfy ∆C(T)(u) := u ⊗ u and
∆C0(Z)( N ) := N ⊗ 1 ∔ 1 ⊗ N , where N ⊗ 1 ∔ 1 ⊗ N means the unbounded affiliated
element of C0(Z × Z) given by the closure of the essentially self-adjoint operator
N ⊗ 1 + 1 ⊗ N .
Let 0 < q < 1. We identify Z× T ∼= C×
q = qZ+iR ⊆ C× by mapping (n, z) 7→ qn ·z.
As suggested above, we are going to construct a multiplicative unitary on the Hilbert
space L2(Z × T) = L2(C×
q ) with the regular representation of Z × T. We choose
the orthonormal basis ek,l := δk ⊗ zl for k, l ∈ Z in L2(C×
q ) and thus identify
L ∼= H ⊗ H.
In our chosen basis, Z acts by αn(ek,l) = ek+n,l for n, k, l ∈ Z
and T acts by αζ(ek,l) = ζ l · ek,l for k, l ∈ Z, ζ ∈ T. Thus the right and left
corepresentations U ∈ U(L ⊗ H) and V ∈ U(H ⊗ L) and the resulting braiding
operator
LL
∈ U(L ⊗ L) are
U = W
V = W
23,
12,
ek,l,m 7→ ek,l,m+l
em,k,l 7→ em,k+m,l
LL = ZΣ = W∗
23Σ,
ek,l ⊗ en,p 7→ en,p ⊗ ek−p,l.
We also describe the representations of C(T) ∼= C∗(Z) and C0(Z) ∼= C∗(T) on L
through a unitary operator U and a self-adjoint operator N with spectrum Z and
commuting with U:
U(ek,l) := ek+1,l
N (ek,l) := lek,l.
We define a closed operator Υ = ΦΥΥ on L by
Υek,l := q2k+lek,l,
ΦΥek,l := ek,l+1,
Υek,l := q2k+lek,l+1.
The operator ΦΥ is unitary and Υ is strictly positive with spectrum qZ ∪ {0}, and
ΦΥ and Υ satisfy the following commutation relations:
(4.3)
ΦΥΥΦ∗
Υ = q−1Υ,
UΦΥ = ΦΥU,
Υ = N − 1,
ΦΥ N Φ∗
UΥU ∗ = q−2Υ,
Υ and N strongly commute.
Thus Υ−1(ek,l) = q−2k−l+1ek,l−1. The closed operator
X := Υq−2 N ⊗ Υ−1,
ek,l ⊗ en,p 7→ q2(k−n)−(l+p)+1ek,l+1 ⊗ en,p−1,
on L ⊗ L is normal because X : ek,l ⊗ en,p 7→ q2(k−n)−(l+p)+1ek,l ⊗ en,p commutes
with its phase ΦX : ek,l ⊗ en,p 7→ ek,l+1 ⊗ en,p−1 in the polar decomposition. The
spectrum of X is Cq := C×
q ∪ {0}. Both X and ΦX strongly commute with U ⊗ U
and N ⊗ 1 ∔ 1 ⊗ N . Thus X is equivariant for the tensor product representations
U U and V V. Hence any circle-valued function F : Cq → T gives a unitary F (X)
on L ⊗ L that is equivariant with respect to U U and V V.
We want to choose F so that F (X) satisfies the braided pentagon equation. Since
the functional calculus is compatible with conjugation by unitaries, the top-braided
pentagon equation for F (X) says
(4.4) F (F (X23)X12F (X23)∗) = F (X12)
LL
23F (X12)L L
23
= F (X12)F (Z23X13Z ∗
23).
We compute
Z23X13Z ∗
23(ek,l ⊗ en,p ⊗ er,s) = q2(k−p−r)−(l+s)+1(ek,l+1 ⊗ en,p ⊗ ep,s−1).
Thus
Z23X13Z ∗
23 = Υq−2 N ⊗ q−2 N ⊗ Υ−1 = X12 · X23.
24
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
Hence (4.4) becomes
(4.5)
F (F (X23)X12F (X23)) = F (X12)F (X12X23).
The quantum exponential function is defined in [30] by
Fq(z) :=
1 + q2kz
1 + q2kz
∞Yk=1
−1
z ∈ Cq \ {−q2Z},
otherwise.
(4.6)
This product converges absolutely outside −q2Z, and (cid:12)(cid:12) 1+q2k z
Thus Fq is a unitary multiplier of C0(Cq), and
1+q2k z(cid:12)(cid:12) = 1 for all z 6= −q2k.
F := Fq(Υ−1q−2 N ⊗ Υ)
is a unitary operator on L ⊗ L.
Theorem 4.7. The triple (U, V, F) is a manageable braided multiplicative unitary
on L relative to W.
The proof will occupy the rest of this section. For (U, V, F) to be a braided
multiplicative unitary, it only remains to verify the braided pentagon equation,
which is equivalent to (4.4). We shall use the properties of the quantum exponential
function established in [30].
The operators
R := X12 = Υq−2 N ⊗ Υ−1 ⊗ 1,
S := X12X23 = Υq−2 N ⊗ q−2 N ⊗ Υ−1
are normal and satisfy the commutation relations in [30, (0.1)], that is, their phases
and their absolute values strongly commute and
Φ∗
RSΦR = qS,
ΦSRΦ∗
S = qR.
Since R−1S = X23 is also normal with spectrum Cq, [30, Theorems 2.1 -- 2] apply
and show that R ∔ S is normal with spectrum Cq and
Fq(X23) · X12 · Fq(X23)∗ = Fq(R−1S) · R · Fq(R−1S)∗ = R ∔ S.
Moreover, [30, Theorem 3.1] gives
Fq(R)Fq(S) = Fq(R ∔ S).
Both results of [30] together give (4.5) for F = Fq; this is equivalent to the braided
pentagon equation for F.
Now we turn to braided manageability. First we compute the unitary eZ. It is
the unique unitary on L ⊗ L that satisfies (3.20). The contragradient eU∗ of U is
given in the standard basis ek,l ⊗ em of L ⊗ H by eU
eZ ∈ U(L ⊗ L) acts on the standard basis by
Equivalently, eZ = (1 ⊗ U) N T⊗1.
eZ(ek,l ⊗ en,p) = ek,l ⊗ en−l,p.
Next we define the operator QL required by Definition 3.15:
(ek,l ⊗ em) = ek,l ⊗ em−l. Hence
∗
QLek,l := q−lek,l.
This is a strictly positive operator on L with spectrum qZ ∪ {0}. It commutes with
U and N and therefore satisfies (3.16) and (3.17). The operator QL ⊗ QL, mapping
ek,l ⊗ en,p 7→ q−(l+p)ek,l ⊗ en,p, commutes with X = Υq−2 N ⊗ Υ−1 and therefore
with F = Fq(X). Thus (3.18) holds as well.
It suffices to
check this if the vectors x, y, u, v involved are standard basis vectors x = ek,l,
Finally, we need a unitary eF ∈ U(L ⊗ L) that satisfies (3.19).
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
25
y = en,p, u = ea,b, v = ec,d for L. Using our explicit formulas for Z and eZ, we may
rewrite (3.19) as
(4.8)
Substituting γ = c + l and F = Fq(X), this becomes
(ek,l ⊗ ea−l,b F en,p ⊗ ec,d) = (en,p ⊗ q−bea,b eF ek,l ⊗ qdec+l,d).
(en,p ⊗ ea,b eF ek,l ⊗ eγ,d) = qb−d(ek,l ⊗ ea−l,b Fq(X) en,p ⊗ eγ−l,d)
for a, b, γ, d, k, l, n, p ∈ Z. So the issue is whether the bilinear form eF defined by
To compute the right hand side in (4.8), we Fourier transform the restrictions
this equation is unitary.
of Fq to the circles z = qn, n ∈ Z, and write
Fq(z) = Xm∈Z
Fm(z)Φm
z ,
see [1] or [29, Appendix A]. The scalars Fm(qn) for m, n ∈ Z are real and satisfy
Fm(qn) = (−q)mF−m(qn−m).
The vectors en,p ⊗ eγ−l,d and ek,l ⊗ ea−l,b are eigenvectors of X with eigenvalues
q2(n−γ+l)−(p+d)+1 and q2(k−a+l)−(b+l)+1, respectively. And Φm
X acts on these vectors
by ek,l ⊗ ea,b 7→ ek,l+m ⊗ ea,b−m. Thus
qb−d(ek,l ⊗ ea−l,b Fq(X) en,p ⊗ eγ−l,d)
qb−d(ek,l ⊗ ea−l,b Φm
X Fm(X) en,p ⊗ eγ−l,d)
qb−d · Fm(q2(n−γ+l)−(p+d)+1)δk,nδl,p+mδa−l,γ−lδb,d−m
= Xm∈Z
= Xm∈Z
= δk,nδa,γδp,l+b−d · qb−d · Fd−b(q2k−2a+l−b+1)
= δk,nδa,γδp,l+b−d · (−1)b−d · Fb−d(q2k−2a+l−d+1).
Now we define an unbounded normal operator eX on L ⊗ H with spectrum Cq by
eX(ek,l ⊗ en,p) = q2(k−n)+l−p+1ek,l ⊗ en,p, ΦeX (ek,l ⊗ en,p) = −ek,l+1 ⊗ en,p+1,
so eX(ek,l ⊗ en,p) = −1 · q2(k−n)+l−p+1ek,l+1 ⊗ en,p+1. We claim that the unitary
eF := Fq(eX)∗ will do:
(en,p ⊗ ea,b eF ek,l ⊗ eγ,d) = Xm∈Z
= Xm∈Z
Fm(eX) en,p ⊗ ea,b)
(−1)mFm(q2(n−a)+p−b+1)δk,nδl,p+mδγ,aδd,b+m
(ek,l ⊗ eγ,d Φm
eX
= δk,nδa,γδp,l+b−d · (−1)d−b · Fd−b(q2k−2a+l−d+1).
This is equal to the result of the computation above, so (4.8) holds. Thus our
braided multiplicative unitary is manageable, and Theorem 4.7 is proved.
Appendix A. Some Manageability Techniques
Let WA ∈ U(HA ⊗ HA) and WB ∈ U(HB ⊗ HB) be manageable multiplicative
unitaries as in [28, Definition 1.2], which generate C∗-quantum groups G = (A, ∆A)
and H = (B, ∆B).
Let V ∈ U( A ⊗ B) be a bicharacter from G to H. Let V ∈ U(HA ⊗ HB) be the
concrete realisation of V. Then V ∈ U(HA ⊗ HB) is adapted to WB in the sense of
26
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
[28, Definition 1.3] by [10, Lemma 3.2]. Thus V ∈ U(HA ⊗ HB) is manageable by
(A.1)
[28, Theorem 1.6]; that is, there is a unitary eV ∈ U(HA ⊗ HB) with
(cid:0)x ⊗ u(cid:12)(cid:12)V(cid:12)(cid:12)z ⊗ y(cid:1) =(cid:0)z ⊗ QBu(cid:12)(cid:12)eV(cid:12)(cid:12)x ⊗ Q−1
B y(cid:1)
for all x, z ∈ HA, u ∈ D(QB) and y ∈ D(Q−1
the manageability condition for WB.
B ); here QB is one of the operators in
Lemma A.2. V(QA ⊗ QB)V∗ = QA ⊗ QB and eV(QT
A ⊗ Q−1
B .
t : R → Aut(A) of A acts through conjugation by Qit
Proof. The scaling group τ A
[28, Theorem 1.5.5], and similarly for B. [10, (20)] says that V is fixed by τ A
This means that V commutes with QA ⊗ QB, as asserted.
B )eV∗ = QT
A ⊗ Q−1
A by
t ⊗ τ B
t .
Hence the left-hand side of (A.1) does not change if we replace x, u, z and y by
A(x), Qit
B(y), respectively, for any t ∈ R. Thus the right-hand
A(z) and Qit
Qit
sides also remain the same, that is,
B(u), Qit
(cid:0)z ⊗ QBu(cid:12)(cid:12)eV(cid:12)(cid:12)x ⊗ Q−1
Hence eV = (cid:16)(cid:2)QT
A(cid:3)it
commutes with QT
B y(cid:1) =(cid:16)(cid:2)QT
A(cid:3)−it
B (cid:17)eV(cid:16)(cid:2)QT
A(cid:3)it
A ⊗ Q−1
B .
⊗ Q−it
z ⊗ Qit
⊗ Q−it
x ⊗ Qit
BQBu(cid:12)(cid:12)(cid:12)eV(cid:12)(cid:12)(cid:12)(cid:2)QT
B u(cid:17).
A(cid:3)−it
B (cid:17) for all t ∈ R. This says that eV
BQ−1
(cid:3)
Lemma A.3. Let Q be a self-adjoint, strictly positive operator on a Hilbert space H.
There is an orthonormal basis (ei)i∈N in H with ei ∈ D(Q) ∩ D(Q−1) and strong
convergence
(A.4)
Xi∈N
Q−1eiihQei = idH.
Proof. For n ∈ Z and λ ∈ [22n−1, 22n+1), let f (λ) = 2−2nλ and g(λ) = λ/f (λ).
The function
R>0 ∋ λ 7→ f (λ) ∈ [2−1, 2)
is piecewise linear and bounded with bounded inverse. Hence the Borel func-
tional calculus for self-adjoint operators gives Q′ := f (Q), which is self-adjoint and
bounded with a bounded inverse. We also get the self-adjoint operator Q′′ = g(Q),
which has countable spectrum {22n n ∈ Z}. Thus H is the orthogonal direct
sum of the 22n-eigenspaces of H. We choose orthonormal bases for all these
eigenspaces and put them together to an orthonormal basis (ei)i∈N of H. We
have Q = Q′Q′′ and Q−1 = (Q′)−1(Q′′)−1 by functional calculus. Since the op-
erators (Q′)±1 are bounded and self-adjoint, Q and Q′′ have the same domain.
Thus ei ∈ D(Q) ∩ D(Q−1) because it is an eigenvector of Q′′ with some positive
eigenvalue 22n for some n ∈ Z depending on i. Since (Q′)±1 are bounded and
Q′′ei = 22nei, we may rewrite
Q−1eiihQei = (Q′)−1(Q′′)−1eiihQ′Q′′ei = (Q′)−12−2neiihQ′22nei
The sumPi∈Neiihei converges strongly to the identity on H because (ei)i∈N is an
orthonormal basis for H. Since (Q′)±1 are bounded, the sum over (Q′)−1eiiheiQ′
converges strongly to (Q′)−1 · 1 · Q′ = 1.
(cid:3)
= (Q′)−1eiiheiQ′.
The following lemma completes the proof of Theorem 3.29.
Lemma A.5. In the situation of the proof of Theorem 3.29, WC , Q and fWC
verify (3.14).
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
27
It suffices to
Proof. We continue in the notation of the proof of Theorem 3.29.
check (3.14) when x, z, u, y are tensor monomials: x = x1 ⊗ x2, z = z1 ⊗ z2,
u = u1 ⊗ u2, y = y1 ⊗ y2 with x1, z1 ∈ H, x2, z2 ∈ L, u1 ∈ D(Q), u2 ∈ D(QL),
y1 ∈ D(Q−1), y2 ∈ D(Q−1
L ). This implies the assertion for all x, y, z, u.
Equation (3.10) is equivalent to U23 V∗
U23Z ∗
13, which gives
(A.6)
WC
1234 = W
13
V∗
34
F24) V34.
34 = V∗
34
U23(Z ∗
24
We first concentrate on the part U23(Z ∗
24
F24) in (A.6). Let (ei)i∈N be an or-
thonormal basis of L. Then (3.30) and (3.19) give
hx2 ⊗ u1 ⊗ u2U12Z ∗
13
F13z2 ⊗ y1 ⊗ y2i
13
=X(cid:10)x2 ⊗ u1 ⊗ u2(cid:12)(cid:12)U12 ·(cid:0)eiihei ⊗ 1 ⊗ 1(cid:1) · Z ∗
=X hx2 ⊗ u1Uei ⊗ y1i · hei ⊗ u2Z ∗Fz2 ⊗ y2i
=XDz2 ⊗ QL(u2)(cid:12)(cid:12)(cid:12)eF(cid:12)(cid:12)(cid:12)ei ⊗ Q−1
=Dz2 ⊗ Q(u1) ⊗ QL(u2)(cid:12)(cid:12)(cid:12)eF13eU12(cid:12)(cid:12)(cid:12)x2 ⊗ Q−1(y1) ⊗ Q−1
F13(cid:12)(cid:12)z2 ⊗ y1 ⊗ y2(cid:11)
L (y2)EDei ⊗ Q(u1)(cid:12)(cid:12)(cid:12)eU(cid:12)(cid:12)(cid:12)x2 ⊗ Q−1(y1)E
L (y2)E
Since V commutes with Q ⊗ QL, it also preserves the domains of (Q ⊗ QL)−1,
and we get an equivalent statement if we replace u1 ⊗ u2 and y1 ⊗ y2 above by
V(u1 ⊗ u2) and V(y1 ⊗ y2), respectively. This gives
(A.7) Dx2 ⊗ u1 ⊗ u2(cid:12)(cid:12)(cid:12) V∗
U12Z ∗
13
23
F13 V23(cid:12)(cid:12)(cid:12)z2 ⊗ y1 ⊗ y2E
=Dz2 ⊗ Q(u1) ⊗ QL(u2)(cid:12)(cid:12)(cid:12) V∗
23eF13eU12 V23(cid:12)(cid:12)(cid:12)x2 ⊗ Q−1(y1) ⊗ Q−1
L (y2)E .
Now let (ǫj)j∈N be a basis of H as in Lemma A.3, that is,
Xj
Q−1(ǫj)ihQ(ǫj ) = idH.
We compute
23
34
13
V∗
34
U23Z ∗
24
U23Z ∗
24
13 ·(cid:0)1 ⊗ 1 ⊗(cid:12)(cid:12)ǫjihǫj ⊗ 1(cid:1) · V∗
F24 V34(cid:12)(cid:12)(cid:12)z1 ⊗ z2 ⊗ y1 ⊗ y2E
hx1 ⊗ u1Wz1 ⊗ ǫii ·Dx2 ⊗ ǫi ⊗ u2(cid:12)(cid:12)(cid:12) V∗
·Dz2 ⊗ Q(ǫi) ⊗ QL(u2)(cid:12)(cid:12)(cid:12) V∗
Dx1 ⊗ x2 ⊗ u1 ⊗ u2(cid:12)(cid:12)(cid:12)W
= Xj (cid:10)x1 ⊗ x2 ⊗ u1 ⊗ u2(cid:12)(cid:12)W
=Xj
= Xj Dz1 ⊗ Q(u1)(cid:12)(cid:12)(cid:12)fW(cid:12)(cid:12)(cid:12)x1 ⊗ Q−1(ǫi)E
= Dz1 ⊗ z2 ⊗ Q(u1) ⊗ QL(u2)(cid:12)(cid:12)(cid:12)fW13 V∗
Thus (3.14) holds for WC , Q and fWC .
Lemma A.8. The unitary eF ∈ U(L ⊗ L) defined by (3.46) satisfies the manage-
F24 V34(cid:12)(cid:12)
(cid:12)(cid:12)z1 ⊗ z2 ⊗ y1 ⊗ y2(cid:11)
U12(Z ∗F)13 V23(cid:12)(cid:12)(cid:12)z2 ⊗ y1 ⊗ y2E
23eF13eU12 V23(cid:12)(cid:12)(cid:12)x2 ⊗ Q−1(y1) ⊗ Q−1
34eF24eU23 V34(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)x1 ⊗ x2 ⊗ Q−1(y1) ⊗ Q−1
L (y2)E
L (y2)E.
ability condition (3.19) for F ∈ U(L ⊗ L) in Theorem 3.37.
(cid:3)
28
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
Proof. It suffices to check (3.19) when x, u, y, v are tensor monomials: x = ¯x1 ⊗ x2,
u = ¯u1 ⊗ u2, y = ¯y1 ⊗ y2, v = ¯v1 ⊗ v2, with x1, x2, y1, y2 ∈ H, v1, u2 ∈ D(QC ) and
u1, v2 ∈ D(Q−1
C ).
First we focus on the part Σ23ebP23Σ23PT⊗T
13
of H in Lemma A.3. Then (¯ei)i∈N is an orthonormal basis of H with
. Let (ei)i∈N be an orthonormal basis
(A.9)
Xi∈N
QeiihQ−1ei = idH.
13
(cid:12)(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ ¯v1(cid:29)
Equation (3.14) for P and bP gives
(cid:28)¯x1 ⊗ x2 ⊗ ¯u1(cid:12)(cid:12)(cid:12)(cid:12)Σ23ebP23Σ23PT⊗T
=X(cid:28)¯x1 ⊗ x2 ⊗ ¯u1(cid:12)(cid:12)(cid:12)(cid:12)Σ23ebP23Σ23 ·(cid:0)1 ⊗ 1 ⊗ eiihei(cid:1) · PT⊗T
=X(cid:28)x2 ⊗ ¯u1(cid:12)(cid:12)(cid:12)(cid:12)ΣebPΣ(cid:12)(cid:12)(cid:12)(cid:12)y2 ⊗ ¯ei(cid:29)(cid:10)¯x1 ⊗ ¯ei(cid:12)(cid:12)PT⊗T(cid:12)(cid:12)¯y1 ⊗ ¯v1(cid:11)
=X(cid:28)¯u1 ⊗ x2(cid:12)(cid:12)(cid:12)(cid:12)ebP(cid:12)(cid:12)(cid:12)(cid:12)¯ei ⊗ y2(cid:29) hy1 ⊗ v1Px1 ⊗ eii
=XDei ⊗ Q−1
C (x2)(cid:12)(cid:12)(cid:12)bP(cid:12)(cid:12)(cid:12)u1 ⊗ QC(y2)ED¯x1 ⊗ QC (v1)(cid:12)(cid:12)(cid:12)eP(cid:12)(cid:12)(cid:12)¯y1 ⊗ Q−1
13
(cid:12)(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ ¯v1(cid:29)
C (ei)E .
13
Lemma A.2 shows that QC ⊗ CC commutes with bP. Hence
(cid:28)¯x1 ⊗ x2 ⊗ ¯u1(cid:12)(cid:12)(cid:12)(cid:12)Σ23ebP23Σ23PT⊗T
=XDQC (ei) ⊗ x2(cid:12)(cid:12)(cid:12)bP(cid:12)(cid:12)(cid:12)Q−1
=X(cid:10)x2 ⊗ QC(ei)(cid:12)(cid:12)P∗(cid:12)(cid:12)y2 ⊗ Q−1
=XD¯y2 ⊗ Q−1
(cid:12)(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ ¯v1(cid:29)
C (u1)(cid:11)Dy1 ⊗ Q−1
C (u1)(cid:12)(cid:12)(cid:12)(P∗)T⊗T(cid:12)(cid:12)(cid:12)¯x2 ⊗ QC (ei)EDy1 ⊗ Q−1
C (u1) ⊗ y2ED¯x1 ⊗ QC(v1)(cid:12)(cid:12)(cid:12)eP(cid:12)(cid:12)(cid:12)¯y1 ⊗ Q−1
C (ei)E
C (ei)(cid:12)(cid:12)(cid:12)ePT⊗T(cid:12)(cid:12)(cid:12)x1 ⊗ QC (v1)E
C (ei)(cid:12)(cid:12)(cid:12)ePT⊗T(cid:12)(cid:12)(cid:12)x1 ⊗ QC (v1)E .
Now (A.9) gives
(A.10) (cid:28)¯x1 ⊗ x2 ⊗ ¯u1(cid:12)(cid:12)(cid:12)(cid:12)Σ23ebP23Σ23PT⊗T
=Dy1 ⊗ ¯y2 ⊗ Q−1
13
(cid:12)(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ ¯v1(cid:29)
C (u1)(cid:12)(cid:12)(cid:12)(P∗)T⊗T
23 ePT⊗T
13
A similar computation gives
(cid:12)(cid:12)(cid:12)x1 ⊗ ¯x2 ⊗ QC(v1)E .
(A.11) D¯x1 ⊗ x2 ⊗ u2(cid:12)(cid:12)(cid:12)WC
13)∗(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ v2E
23(fWC
=Dy1 ⊗ ¯y2 ⊗ QC (u2)(cid:12)(cid:12)(cid:12)fWC
23(WC
13)∗(cid:12)(cid:12)(cid:12)x1 ⊗ ¯x2 ⊗ Q−1
C (v2)E .
SEMIDIRECT PRODUCTS OF C*-QUANTUM GROUPS
29
Let (ej)j∈N be an orthonormal basis of H. Equations (A.10) and (A.11) imply
14)∗(cid:12)(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ ¯v1 ⊗ v2(cid:29)
24(fWC
·(cid:0)¯ejih¯ej ⊗ ekihek ⊗ 1 ⊗ 1(cid:1)
13
13
WC
(cid:28)¯x1 ⊗ x2 ⊗ ¯u1 ⊗ u2(cid:12)(cid:12)(cid:12)(cid:12)Σ23ebP23Σ23PT⊗T
= Xj,kD¯x1 ⊗ x2 ⊗ ¯u1 ⊗ u2(cid:12)(cid:12)(cid:12)Σ23ebP23Σ23PT⊗T
= Xj,kD¯x1 ⊗ x2 ⊗ ¯u1(cid:12)(cid:12)(cid:12)Σ23ebP23Σ23PT⊗T
= Xj,k Dy1 ⊗ ¯y2 ⊗ QC(u2)(cid:12)(cid:12)(cid:12)fWC
= Dy1 ⊗ ¯y2 ⊗ Q−1
C (u1) ⊗ QC(u2)(cid:12)(cid:12)(cid:12)fWC
23(WC
Dej ⊗ ¯ek ⊗ Q−1
13
· WC
24(fWC
(cid:12)(cid:12)(cid:12)¯ej ⊗ ek ⊗ ¯v1E
D¯ej ⊗ ek ⊗ u2(cid:12)(cid:12)(cid:12)WC
13)∗(cid:12)(cid:12)(cid:12)ej ⊗ ¯ek ⊗ Q−1
C (v2)E
C (u1)(cid:12)(cid:12)(cid:12)(P∗)T⊗T
23 ePT⊗T
(cid:12)(cid:12)(cid:12)
23 ePT⊗T
14)∗(P∗)T⊗T
14)∗(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ ¯v1 ⊗ v2E
13)∗(cid:12)(cid:12)(cid:12)¯y1 ⊗ y2 ⊗ v2E
23(fWC
(cid:12)(cid:12)(cid:12)x1 ⊗ ¯x2 ⊗ QC (v1)E
C (v2)E
x1 ⊗ ¯x2 ⊗ QC (v1) ⊗ Q−1
24(WC
13
13
This is the equation we have to check.
(cid:3)
References
[1] Saad Baaj, Représentation régulière du groupe quantique E
R. Acad. Sci. Paris Sér.
http://gallica.bnf.fr/ark:/12148/bpt6k58688425/f1025.item. MR 1168528
µ(2) de Woronowicz, C.
I Math. 314 (1992), no. 13, 1021 -- 1026, available at
[2] Alcides Buss, Generalized fixed point algebras for coactions of locally compact quantum
groups, Doctoral Thesis, Univ. Münster, 2007.
[3]
, Generalized
quantum groups, Münster
fixed
point
algebras
J. Math.
pact
295 -- 341,
http://nbn-resolving.de/urn:nbn:de:hbz:6-55309458343. MR 3148214
(2013),
for
6
coactions
of
locally
available
com-
at
[4] Matthew Daws, Paweł Kasprzak, Adam G. Skalski, and Piotr Mikołaj Sołtan, Closed quantum
subgroups of locally compact quantum groups, Adv. Math. 231 (2012), no. 6, 3473 -- 3501,
doi: 10.1016/j.aim.2012.09.002. MR 2980506
[5] Paweł Kasprzak, Ralf Meyer, Sutanu Roy, and Stanisław Lech Woronowicz, Braided quantum
SU(2) groups, J. Noncommut. Geom. (2016), accepted. arXiv: 1411.3218.
[6] Paweł Kasprzak and Piotr Mikołaj Sołtan, Quantum groups with projection on
von Neumann algebra level, J. Math. Anal. Appl. 427 (2015), no. 1, 289 -- 306,
doi: 10.1016/j.jmaa.2015.02.047. MR 3318200
[7]
, Quantum groups with projection and extensions of locally compact quantum groups
(2014), eprint. arXiv: 1412.0821.
[8] Shahn Majid, Algebras and Hopf algebras in braided categories, Advances in Hopf algebras
(Chicago, IL, 1992), 1994, pp. 55 -- 105. MR 1289422
[9] Ralf Meyer, Generalized fixed point algebras and square-integrable groups actions, J. Funct.
Anal. 186 (2001), no. 1, 167 -- 195, doi: 10.1006/jfan.2001.3795. MR 1863296
Lech Woronowicz,
Stanisław
[10] Ralf Meyer,
Sutanu
Roy,
and
phisms
http://nbn-resolving.de/urn:nbn:de:hbz:6-88399662599. MR 3047623
quantum groups, Münster
J. Math. 5 (2012),
of
1 -- 24,
Homomor-
at
available
[11]
[12]
, Quantum group-twisted tensor products of C∗-algebras, Internat. J. Math. 25 (2014),
no. 2, 1450019, 37, doi: 10.1142/S0129167X14500190. MR 3189775
, Quantum group-twisted tensor products of C∗-algebras II, J. Noncommut. Geom.
(2016), accepted. arXiv: 1501.04432.
[13] Ryszard Nest and Christian Voigt, Equivariant Poincaré duality for quantum group actions,
J. Funct. Anal. 258 (2010), no. 5, 1466 -- 1503, doi: 10.1016/j.jfa.2009.10.015. MR 2566309
[14] Chi-Keung
Op-
erator
at
http://www.theta.ro/jot/archive/1997-038-002/1997-038-002-001.html. MR 1606928
Morphisms
38
Ng,
Theory
multiplicative
unitaries,
available
203 -- 224,
(1997),
no.
of
J.
2,
30
RALF MEYER, SUTANU ROY, AND STANISŁAW LECH WORONOWICZ
[15] David E. Radford, The structure of Hopf algebras with a projection, J. Algebra 92 (1985),
no. 2, 322 -- 347, doi: 10.1016/0021-8693(85)90124-3. MR 778452
[16] Marc A. Rieffel, Integrable and proper actions on C ∗-algebras, and square-integrable represen-
tations of groups, Expo. Math. 22 (2004), no. 1, 1 -- 53, doi: 10.1016/S0723-0869(04)80002-1.
MR 2166968
[17] Sutanu Roy, C∗-Quantum groups with projection, Ph.D. Thesis, Georg-August Universität
Göttingen, 2013, http://hdl.handle.net/11858/00-1735-0000-0022-5EF9-0.
[18]
, The Drinfeld double for C ∗-algebraic quantum groups, J. Operator Theory 74 (2015),
no. 2, 485 -- 515, doi: 10.7900/jot.2014sep04.2053. MR 3431941
[19]
[20] Piotr Mikołaj Sołtan, New quantum "az + b" groups, Rev. Math. Phys. 17 (2005), no. 3,
, Braided C ∗-quantum groups (2016), eprint. arXiv: 1601.00169.
313 -- 364, doi: 10.1142/S0129055X05002339. MR 2144675
[21] Piotr Mikołaj Sołtan and Stanisław Lech Woronowicz, From multiplicative unitaries to quan-
tum groups. II, J. Funct. Anal. 252 (2007), no. 1, 42 -- 67, doi: 10.1016/j.jfa.2007.07.006.
MR 2357350
[22] Piotr Stachura, Towards a topological (dual of) quantum κ-Poincaré group, Rep. Math. Phys.
57 (2006), no. 2, 233 -- 256, doi: 10.1016/S0034-4877(06)80019-4. MR 2227008
[23]
, On the quantum 'ax + b' group, J. Geom. Phys. 73 (2013), 125 -- 149,
doi: 10.1016/j.geomphys.2013.05.006. MR 3090106
[24] Stefaan Vaes, A new approach to induction and imprimitivity results, J. Funct. Anal. 229
(2005), no. 2, 317 -- 374, doi: 10.1016/j.jfa.2004.11.016. MR 2182592
[25] Stefaan Vaes and Leonid Vainerman, Extensions of
locally compact quantum groups
construction, Adv. Math. 175 (2003), no. 1, 1 -- 101,
and the bicrossed product
doi: 10.1016/S0001-8708(02)00040-3. MR 1970242
[26] Stanisław Lech Woronowicz, Quantum E(2) group and its Pontryagin dual, Lett. Math. Phys.
23 (1991), no. 4, 251 -- 263, doi: 10.1007/BF00398822. MR 1152695
[27]
[28]
[29]
[30]
, C ∗-algebras generated by unbounded elements, Rev. Math. Phys. 7 (1995), no. 3,
481 -- 521, doi: 10.1142/S0129055X95000207. MR 1326143
, From multiplicative unitaries to quantum groups, Internat. J. Math. 7 (1996), no. 1,
127 -- 149, doi: 10.1142/S0129167X96000086. MR 1369908
, Quantum 'az + b' group on complex plane, Internat. J. Math. 12 (2001), no. 4,
461 -- 503, doi: 10.1142/S0129167X01000836. MR 1841400
, Operator equalities related to the quantum E(2) group, Comm. Math. Phys. 144
(1992), no. 2, 417 -- 428, available at http://projecteuclid.org/euclid.cmp/1104249324.
MR 1152380
[31] Stanisław Lech Woronowicz and Stanisław Zakrzewski, Quantum 'ax + b' group, Rev. Math.
Phys. 14 (2002), no. 7-8, 797 -- 828, doi: 10.1142/S0129055X02001405. MR 1932667
E-mail address: [email protected]
Mathematisches Institut, Georg-August Universität Göttingen, Bunsenstrasse 3 -- 5,
37073 Göttingen, Germany
E-mail address: [email protected]
School of Mathematics and Statistics, Carleton University, 1125 Colonel By Drive,
K1S 5B6 Ottawa, Canada.
E-mail address: [email protected]
Instytut Matematyczny Polskiej Akademii Nauk, ul. Śniadeckich 8, 00-656 Warszawa,
Poland, and, Katedra Metod Matematycznych Fizyki, Wydział Fizyki, Uniwersytet
Warszawski, ul. Pasteura 5, 02-093 Warszawa, Poland
|
1709.00118 | 2 | 1709 | 2019-01-27T19:32:24 | Completely bounded maps and invariant subspaces | [
"math.OA",
"math.FA"
] | We provide a description of certain invariance properties of completely bounded bimodule maps in terms of their symbols. If $\mathbb{G}$ is a locally compact quantum group, we characterise the completely bounded $L^{\infty}(\mathbb{G})'$-bimodule maps that send $C_0(\hat{\mathbb{G}})$ into $L^{\infty}(\hat{\mathbb{G}})$ in terms of the properties of the corresponding elements of the normal Haagerup tensor product $L^{\infty}(\mathbb{G}) \otimes_{\sigma{\rm h}} L^{\infty}(\mathbb{G})$. As a consequence, we obtain an intrinsic characterisation of the normal completely bounded $L^{\infty}(\mathbb{G})'$-bimodule maps that leave $L^{\infty}(\hat{\mathbb{G}})$ invariant, extending and unifying results, formulated in the current literature separately for the commutative and the co-commutative cases. | math.OA | math |
COMPLETELY BOUNDED MAPS AND INVARIANT
SUBSPACES
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
Abstract. We provide a description of certain invariance properties of
completely bounded bimodule maps in terms of their symbols. If G is a
locally compact quantum group, we characterise the completely bounded
L∞(G)′-bimodule maps that send C0( G) into L∞( G) in terms of the
properties of the corresponding elements of the normal Haagerup tensor
product L∞(G) ⊗σ h L∞(G). As a consequence, we obtain an intrinsic
characterisation of the normal completely bounded L∞(G)′-bimodule
maps that leave L∞( G) invariant, extending and unifying results, for-
mulated in the current literature separately for the commutative and
the co-commutative cases.
1. Introduction
Let M be a von Neumann algebra, acting on a Hilbert space H, with com-
mutant M′. The normal Haagerup tensor product M⊗σ hM was introduced
by E. G. Effros and A. Kishimoto in [9], where it was shown that to each of
its element χ there corresponds, in a canonical way, a completely bounded
M′-bimodule map Φχ on the space B(H) of all bounded linear operators
on H. Let us call χ the symbol of the map Φχ. The transformation Φχ is
in addition weak* continuous precisely when χ belongs to the, smaller, ex-
tended Haagerup tensor product M ⊗eh M, introduced by U. Haagerup [12].
The present paper is concerned with the question of how specific invariant
subspace properties of the map Φχ are reflected in its symbol χ.
As a first motivation, we single out Herz-Schur multipliers of the Fourier
algebra A(G) of a locally compact group G, introduced by J. de Canniere
and U. Haagerup [7]. These objects have played a prominent role in opera-
tor algebra theory, in particular in the study of approximation properties of
the von Neumann algebra VN(G) and the reduced group C*-algebra C ∗
r (G)
of G [6, 13, 17]. The numerous applications they have found were facili-
tated to a great extent by the description, due to J. E. Gilbert and to M.
Bozejko and G. Fendler [5], which identifies them with a part of the Schur
multipliers on the direct product G × G. Recall that the space of Schur
multipliers [12, 21, 22, 26] coincides with the extended Haagerup tensor
product L∞(G) ⊗eh L∞(G), where L∞(G) is identified with the algebra of
operators of multiplication by essentially bounded functions on G, acting
Date: 27 January 2019.
1
2
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
on the Hilbert space L2(G). Herz-Schur multipliers can be identified with
those completely bounded weak* continuous L∞(G)-bimodule maps that
leave VN(G) invariant. As L∞(G) ⊗eh L∞(G) can be naturally embedded
in L∞(G × G), a natural problem, addressed in [24] and [20], is that of
characterising the functions ϕ : G × G → C that are symbols of Herz-Schur
multipliers.
Exchanging the roles of L∞(G) and VN(G), one may consider the space of
completely bounded weak* continuous VN(G)′-bimodule maps on B(L2(G)).
Since VN(G) ⊗eh VN(G) naturally embeds in VN(G × G), to each such map
there corresponds canonically an element of the Neumann algebra VN(G ×
G). Those such maps that leave L∞(G) invariant were characterised by M.
Neufang [19] (see also [20]) as arising in a canonical fashion from bounded
complex measures on G.
In [1], under the restriction that G be weakly
amenable, we provided an intrinsic characterisation of the symbols from
VN(G) ⊗eh VN(G) whose maps leave L∞(G) invariant, showing that, when
viewed as elements of VN(G × G), they are precisely those supported, in the
sense of P. Eymard [11], on the anti-diagonal of G.
The invariance properties in the two cases described above were placed
in the same context in [15], where the authors showed that, if G is a locally
compact quantum group with underlying von Neumann algebra L∞(G) and
a dual quantum group G with underlying von Neumann algebra L∞( G), the
completely bounded weak* continuous L∞(G)′-bimodule maps that leave
L∞( G) invariant correspond in a canonical fashion to the completely bounded
left multipliers of the predual L1( G) of L∞( G) (we direct the reader to [15]
for the definitions and further details). However, the problem of charac-
terising intrinsically the elements of L∞(G) ⊗eh L∞(G) which are symbols
of maps leaving the von Neumann algebra L∞( G) invariant, has not been
addressed in the literature.
One of the main aims of the present paper is to exhibit such a characteri-
sation. We unify results in the literature that are currently stated in the two
extreme cases -- for commutative and co-commutative locally compact quan-
tum groups -- arriving at a condition that captures both simultaneously. In
fact, our results go beyond this aim, as we are able to drop the requirement
of weak* continuity, and characterise intrinsically those elements of the nor-
mal Haagerup tensor product L∞(G)⊗σ hL∞(G) that give rise to completely
bounded (but not necessarily weak* continuous) maps on B(L2(G)) sending
the reduced C*-algebra C0( G) of G into L∞( G).
Specialising to the case of co-commutative quantum groups, we obtain a
generalisation of our previous results from [1], removing the assumption of
weak amenability imposed therein. On the other hand, specialising to the
case of commutative quantum groups, we obtain the first, to the best of our
knowledge, rigorous characterisation of the Schur multipliers ϕ : G × G → C
whose corresponding map on B(L2(G)) leaves VN(G) invariant, in terms of
the function ϕ alone; the result states that this happens if and only if, for
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
3
every r ∈ G, the equality ϕ(sr, tr) = ϕ(s, t) holds for marginally almost all
(s, t) ∈ G × G.
Our main result in the setting of quantum groups, Theorem 3.1, is ob-
tained as a consequence of a more general statement (Theorem 2.3) that
expresses invariance properties of completely bounded bimodule maps in
terms of their symbol. This general viewpoint is presented in Section 2,
after obtaining some preparatory results. It is applied, in Section 3, to the
case of locally compact quantum groups, while Sections 4 and 5 contain the
further applications to co-commutative and commutative quantum groups,
respectively.
We finish this section by setting basic notation. We denote by B(H)
the algebra of all bounded linear operators acing on a Hilbert space H.
If M ⊆ B(H) is a von Neumann algebra, we let M∗ denote its predual,
consisting of all normal functionals on M, and equip it with the operator
space structure arising from its natural embedding into M∗. We denote
by CBM(B(H)) (resp. CBσ
M(B(H))) the operator space of all completely
bounded (resp. weak* continuous, or normal, completely bounded) M-
bimodule maps on B(H). The algebraic tensor product of vector spaces X
and Y is denoted by X ⊙ Y, the space of all n by n matrices -- by Mn, and
the tensor product X ⊙ Mn is written as Mn(X ) and identified with the
space of all n by n matrices with entries in X . If X and Y are operator
spaces, we denote by X ⊗h Y the Haagerup tensor product of X and Y, and
by X ⊗Y their operator projective tensor product. If X and Y are moreover
dual operator spaces, X ¯⊗Y stands for their weak* spatial tensor product.
If ϕ : X × Y → Z (where Z is another operator space) and n, m ∈ N, we
let ϕ(m,n) : Mn(X ) × Mm(Y) → Mnm(Z) be the ampliation of ϕ, given by
ϕ(m,n)((xi,j), (yp,q)) = (ϕ(xi,j, yp,q)i,p,j,q. The identity operator is denoted
by 1, and we let M ⊗ 1 = {a ⊗ 1 : a ∈ M}; it will be clear from the
context on which Hilbert space 1 acts. We will use throughout the paper
basic results from operator space theory, and we refer the reader to the
monographs [3, 10, 21, 22] for the necessary background.
2. A characterisation of invariance
Let X and Y be dual operator spaces and Bilσ(X , Y) be the operator
space of all normal (i.e. separately weak* continuous) completely bounded
bilinear forms on X × Y. The normal Haagerup tensor product [3, 9] of X
and Y is the operator space dual of Bilσ(X , Y):
X ⊗σ h Y
def
= Bilσ(X , Y)∗.
It is characterised by the following universal property: for every dual op-
erator space Z and every normal completely bounded bilinear map φ :
X × Y → Z, there exists a (unique) normal completely bounded lineari-
sation φ : X ⊗σ h Y → Z. It is easy to see that the normal Haagerup tensor
product is functorial: if X1 and Y1 are dual operator spaces and φ : X → X1
4
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
and ψ : Y → Y1 are normal completely bounded maps then there exists a
(unique) normal completely bounded map φ ⊗ ψ : X ⊗σ h Y → X1 ⊗σ h Y1
such that
(1)
(φ ⊗ ψ)(a ⊗ b) = φ(a) ⊗ ψ(b),
a ∈ X , b ∈ Y.
Let M be a von Neumann algebra acting on a Hilbert space H. There
exists a canonical complete isometry, that is also a weak* homeomorphism,
from M ⊗σ h M onto the space CBM′(B(H)) of all completely bounded M′-
bimodule maps on B(H) [9]. For χ ∈ M ⊗σ h M, we let Φχ : B(H) → B(H)
be its corresponding map; we have that
(2)
Φa⊗b(x) = axb,
x ∈ B(H), a, b ∈ M.
We call χ the symbol of Φχ. If ξ, η ∈ H, x ∈ B(H) and Fx,ξ,η : M × M → C
is the (normal completely bounded) bilinear form given by Fx,ξ,η(a, b) =
haxbξ, ηi, then [9, (2.5)]
hχ, Fx,ξ,ηi = hΦχ(x)ξ, ηi, χ ∈ M ⊗σ h M.
It follows that the map χ → Φχ is continuous when M ⊗σ h M is equipped
with its weak* topology and CBM′(B(H)) -- with the point-weak operator
topology.
Recall that the extended Haagerup tensor product M ⊗eh M [10] coin-
cides with the weak* Haagerup tensor product M ⊗w∗ h M [4], and can
be canonically identified with the operator space CBσ
M′(B(H)) of all nor-
mal completely bounded M′-bimodule maps on B(H):
for each element
χ ∈ M ⊗eh M, there exist families (ai)i∈I, (bi)i∈I ⊆ M such that the series
i bi are weak* convergent and the corresponding map
Pi∈I aia∗
i and Pi∈I b∗
Φχ : B(H) → B(H) is given by
where the series converges in the weak* topology of B(H). We write χ ∼
Pi∈I ai ⊗ bi.
Note that the canonical inclusion M ⊗eh M ⊆ M ⊗σ h M is completely
isometric. If N ⊆ B(H) is a(nother) von Neumann algebra, we let
CBσ,N
M′ (B(H)) = {Φ ∈ CBσ
M′(B(H)) : Φ(N ) ⊆ N }.
Let H be a Hilbert space, M ⊆ B(H) be a von Neumann algebra and
def
=
f, g ∈ M∗. By the functoriality of the weak* spatial tensor product, Lf
f ⊗ id and Rg
from M ¯⊗M into M; note that
def
= id ⊗g are well-defined normal completely bounded maps
Lf (a ⊗ b) = f (a)b and Rg(a ⊗ b) = g(b)a,
a, b ∈ M,
kLf kcb = kf k and kRgkcb = kgk (see [27]). By the functoriality of the
normal Haagerup tensor product,
Lf ⊗ Rg : (M ¯⊗M) ⊗σh (M ¯⊗M) → M ⊗σ h M
Φχ(x) =Xi∈I
aixbi,
x ∈ B(H),
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
5
is a weak* continuous completely bounded map (see [10]).
Let m : M×M → M denote operator multiplication. The map m extends
uniquely to a weak* continuous completely contractive map (denoted in the
same way) m : M ⊗σh M → M [10]. We let m∗ : M∗ → M∗ ⊗eh M∗ be its
predual.
Lemma 2.1. For every ψ ∈ (M ¯⊗M) ⊗σh (M ¯⊗M) there exists (a unique)
T (ψ) ∈ M ¯⊗M ¯⊗M such that
hT (ψ), f ⊗ ω ⊗ gi = hm((Lf ⊗ Rg)(ψ)), ωi,
f, g, ω ∈ M∗.
Moreover, the map
T : (M ¯⊗M) ⊗σh (M ¯⊗M) → M ¯⊗M ¯⊗M
is linear, contractive and weak*-continuous, and
(3)
T (χ1 ⊗ χ2) = (χ1 ⊗ 1)(1 ⊗ χ2)
for all χ1, χ2 ∈ M ¯⊗M.
Proof. Let ψ ∈ (M ¯⊗M) ⊗σh (M ¯⊗M) and Fψ : M∗ × M∗ × M∗ → C be
the trilinear map given by
(4)
Fψ(f, ω, g) = hm((Lf ⊗ Rg)(ψ)), ωi,
f, g, ω ∈ M∗.
Let X = (fi,j)i,j ∈ Mn(M∗), Y = (gk,l)k,l ∈ Mm(M∗), and
LX : M ¯⊗M → Mn(M), RY : M ¯⊗M → Mm(M)
be the maps given by
LX (χ) = (Lfi,j (χ))i,j, RY (χ) = (Rgk,l(χ))k,l, χ ∈ M ¯⊗M.
Standard arguments show that LX and RY are completely bounded and
kLXkcb ≤ kXkMn(M∗), kRY kcb ≤ kY kMm(M∗).
By [10, p. 149], the map
LX ⊗ RY : (M ¯⊗M) ⊗σh (M ¯⊗M) → Mn(M) ⊗σh Mm(M)
is completely bounded with kLX ⊗ RY kcb ≤ kXkMn(M∗)kY kMm(M∗).
By [10, Theorem 6.1], the canonical shuffle map
Sσ : Mn(M) ⊗σh Mm(M) = (M ¯⊗Mn) ⊗σh (M ¯⊗Mm)
→ (M ⊗σh M) ¯⊗(Mn ⊗σh Mm)
is a complete contraction. Since the minimal norm is dominated by the
Haagerup norm, we have, on the other hand, that the map
θ : Mn ⊗σh Mm = Mn ⊗h Mm → Mnm
is a complete contraction. It follows that the canonical map
τ : Mn(M) ⊗σh Mm(M) → Mnm(M ⊗σh M),
6
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
induced by Sσ and θ, is a complete contraction. Therefore, for each p ∈ N
and Z = (ωs,t)s,t ∈ Mp(M∗), we have
kF (m,n,p)
ψ
(X, Y, Z)kMmnp
(p)
∗ (ωs,t)ikMmnp
= kh((Lfi,j ⊗ Rgk,l)(ψ), m
≤ k((Lfi,j ⊗ Rgk,l)(ψ))i,j,k,lkMnm(M⊗σhM)k(m∗(ωs,t))s,tkMp(M∗⊗ehM∗)
≤ k(LX ⊗ RY )(ψ)kMn(M)⊗σhMm(M)kZk ≤ kψkkXkkY kkZk.
Thus, the map Fψ is a completely bounded trilinear form on M∗×M∗×M∗,
with
(5)
kFψkcb ≤ kψk.
It therefore linearises to a bounded linear functional on M∗ ⊗M∗ ⊗M∗,
which we denote again by Fψ. Since (M∗ ⊗M∗ ⊗M∗)∗ ≡ M ¯⊗M ¯⊗M, there
exists T (ψ) ∈ M ¯⊗M ¯⊗M such that
(6)
By (5),
Fψ(w) = hT (ψ), wi, w ∈ M∗ ⊗M∗ ⊗M∗.
kT (ψ)k = sup{hT (ψ), wi : kwk ≤ 1} = sup{kFψ(w)k : kwk ≤ 1}
= kFψk ≤ kψk.
The uniqueness of T (ψ) follows from the density of M∗ ⊙ M∗ ⊙ M∗ in
M∗ ⊗M∗ ⊗M∗.
We show that T is weak* continuous. To this end, for each element w of
M∗ ⊗M∗ ⊗M∗, define a linear functional Ew on (M ¯⊗M) ⊗σh (M ¯⊗M) by
Ew(ψ) = hT (ψ), wi, ψ ∈ (M ¯⊗M) ⊗σh (M ¯⊗M).
It clearly suffices to prove that Ew is weak* continuous, for each w. By [23, p
75], it suffices to show that the kernel of Ew is weak* closed which, by virtue
of the Krein-Shmulian Theorem [23, p 152], is equivalent to the fact that the
intersection of ker(Ew) with every norm closed ball of (M ¯⊗M)⊗σh (M ¯⊗M)
is weak* closed.
Fix w ∈ M∗ ⊗M∗ ⊗M∗. Let C > 0, ψ ∈ (M ¯⊗M) ⊗σh (M ¯⊗M) and
(ψα)α∈A be a net in ker(Ew) that converges to ψ in the weak* topology,
such that kψαk ≤ C, α ∈ A. Fix ǫ > 0. Let w0 ∈ M∗ ⊙ M∗ ⊙ M∗ be such
that kw − w0k < ǫ/3C. By the weak* continuity of Lf ⊗ Rg and m, (4) and
(6) imply that there exists α0 ∈ A such that
Ew0(ψα) − Ew0(ψ) = Fψα(w0) − Fψ(w0) < ǫ/3, α ≥ α0.
An ǫ/3-argument now implies that
Ew(ψ) = Ew(ψα) − Ew(ψ) = Fψα(w) − Fψ(w) < ǫ, α ≥ α0.
It follows that Ew(ψ) = 0 and the weak* continuity of T is established.
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
7
Finally, we show (3). Assume first that χ1 = a ⊗ b and χ2 = c ⊗ d where
a, b, c, d ∈ M. Then
hm ((Lf ⊗ Rg)(χ1 ⊗ χ2)) , ωi = hm(Lf (χ1) ⊗ Rg(χ2)), ωi
= hLf (χ1)Rg(χ2), ωi
= f (a)g(d) hbc, ωi
= ha ⊗ bc ⊗ d, f ⊗ ω ⊗ gi
= h(χ1 ⊗ 1)(1 ⊗ χ2), f ⊗ ω ⊗ gi,
giving T (χ1 ⊗ χ2) = (χ1 ⊗ 1)(1⊗ χ2). By linearity, the equality holds for any
χ1, χ2 in the algebraic tensor product M ⊙ M. As T is weak* continuous
and the multiplication in M ¯⊗M ¯⊗M is separately weak* continuous, we
obtain the equality (3) for any χ1, χ2 ∈ M ¯⊗M.
(cid:3)
Intuitively, the map T from Lemma 2.1 "multiplies the two
Remark.
middle variables" in the four-term tensor product, leaving the outer variables
intact, thus producing a three-variable element.
Let H be a Hilbert space, and M and N be von Neumann algebras acting
on H. We fix for the rest of the section a unitary operator W ∈ B(H) ¯⊗N
(resp. V ∈ N ¯⊗B(H)), and define maps
ΓW : M → B(H ⊗ H) and Γ′
V : M → B(H ⊗ H)
by letting
ΓW (a) = W ∗(1 ⊗ a)W and Γ′
V (b) = V (b ⊗ 1)V ∗.
(7)
Clearly, ΓW and Γ′
a normal completely bounded map
V are normal completely bounded maps; thus, there exists
ΓW ⊗ Γ′
V : M ⊗σ h M → (B(H) ¯⊗B(H)) ⊗σ h (B(H) ¯⊗B(H))
such that
(ΓW ⊗ Γ′
V )(a ⊗ b) = ΓW (a) ⊗ Γ′
V (b),
a, b ∈ M.
For an operator T ∈ B(H ⊗ H), let T1,2 = T ⊗ 1 and T2,3 = 1 ⊗ T . Recall
that, for χ ∈ M ⊗σ h M, we denote by Φχ the corresponding completely
bounded mapping on B(H).
Lemma 2.2. Let H be a Hilbert space, M and N be von Neumann algebras
acting on H, χ ∈ M ⊗σ h M, W ∈ B(H) ¯⊗N , V ∈ N ¯⊗B(H) be unitary
operators, and f, g ∈ B(H)∗. Then
(8) (f ⊗ id ⊗g)(W1,2(T ◦(ΓW ⊗ Γ′
V )(χ))V2,3) = Φχ((f ⊗ id)(W )(id ⊗g)(V )).
Proof. Assume first that χ = a ⊗ b, where a, b ∈ M. Using Lemma 2.1, we
have
T ◦ (ΓW ⊗ Γ′
V )(χ) = T (ΓW (a) ⊗ Γ′
V (b))
= T ((W ∗(1 ⊗ a)W ) ⊗ (V (b ⊗ 1)V ∗))
= W ∗
1,2(1 ⊗ a ⊗ 1)W1,2V2,3(1 ⊗ b ⊗ 1)V ∗
2,3.
8
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
Hence
W1,2(T ◦ (ΓW ⊗ Γ′
V )(χ))V2,3 = (1 ⊗ a ⊗ 1)W1,2V2,3(1 ⊗ b ⊗ 1).
Thus, in order to establish (8), it suffices to show that, whenever T ∈
B(H) ¯⊗N and S ∈ N ¯⊗B(H), we have
(9) (f ⊗ id ⊗g)((1 ⊗ a ⊗ 1)T1,2S2,3(1 ⊗ b ⊗ 1)) = Φχ((f ⊗ id)(T )(id ⊗g)(S)).
To this end, assume first that T = T1 ⊗ T2 and S = S1 ⊗ S2, where T1, S2 ∈
B(H) and T2, S1 ∈ N . Then
(f ⊗ id ⊗g)((1 ⊗ a ⊗ 1)T1,2S2,3(1 ⊗ b ⊗ 1))
= (f ⊗ id ⊗g)((1 ⊗ a ⊗ 1)(T1 ⊗ T2S1 ⊗ S2)(1 ⊗ b ⊗ 1))
= (f ⊗ id ⊗g)(T1 ⊗ Φχ(T2S1) ⊗ S2)
= f (T1)g(S2)Φχ(T2S1) = Φχ((f ⊗ id)(T )(id ⊗g)(S)).
By linearity, (9) holds true if T (resp. S) belongs to the algebraic tensor
product B(H) ⊙ N (resp. N ⊙ B(H)). Equation (9) now follows from the
weak* continuity of Φχ.
By linearity, (8) holds whenever χ is an element of the algebraic tensor
product M ⊙ M. Now assume that χ ∈ M ⊗σ h M is arbitrary. Let
(χα)α∈A ⊆ M ⊙ M be a net converging to χ in the weak* topology of
M ⊗σ h M. Then
Φχα((f ⊗ id)(W )(id ⊗g)(V )) →WOT
α∈A Φχ((f ⊗ id)(W )(id ⊗g)(V )).
On the other hand, the weak* continuity of one-sided operator multiplication
and that of the maps ΓW ⊗ Γ′
V , T and f ⊗ id ⊗g imply that
(f ⊗ id ⊗g)(W1,2(T ◦ (ΓW ⊗ Γ′
(f ⊗ id ⊗g)(W1,2(T ◦ (ΓW ⊗ Γ′
V )(χα))V2,3) →w∗
α∈A
V )(χ))V2,3).
Identity (8) now follows.
Let
(cid:3)
A(W, V ) = [Lf (W )Rg(V ) : f, g ∈ B(H)∗]
k·k
;
we call A(W, V ) the reduced operator space of the pair (W, V ). The assump-
tions W ∈ B(H) ¯⊗N and V ∈ N ¯⊗B(H) imply that A(W, V ) is a (norm
closed) subspace of N .
Theorem 2.3. Let H be a Hilbert space, M and N be von Neumann alge-
bras acting on H, χ ∈ M ⊗σ h M, and W ∈ B(H) ¯⊗N , V ∈ N ¯⊗B(H) be
unitary operators. The following are equivalent:
(i) Φχ(A(W, V )) ⊆ N ;
(ii) T ◦ (ΓW ⊗ Γ′
V )(χ) ∈ B(H) ¯⊗N ¯⊗B(H).
Proof. (i)⇒(ii) By Lemma 2.2,
(f ⊗ id ⊗g)(W1,2(T ◦ (ΓW ⊗ Γ′
V )(χ))V2,3) ∈ N
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
9
for all f, g ∈ B(H)∗. It follows from [27] that
W1,2(T ◦ (ΓW ⊗ Γ′
V )(χ))V2,3 ∈ B(H) ¯⊗N ¯⊗B(H).
Thus,
T ◦ (ΓW ⊗ Γ′
V )(χ) ∈ B(H) ¯⊗N ¯⊗B(H).
(ii)⇒(i) is immediate from Lemma 2.2.
(cid:3)
3. Applications to locally compact quantum groups
We refer the reader to [18, 28] (see also [15]) for background on (von Neu-
mann algebraic) locally compact quantum groups. In the sequel, we recall
only those elements of the theory that will be essential for the statements
and the proofs of our results. A locally compact quantum group is a quadru-
ple G = (L∞(G), Γ, ϕ, ψ), where L∞(G) is a von Neumann algebra equipped
with a co-associative co-multiplication Γ : L∞(G) → L∞(G) ¯⊗L∞(G), and ϕ
and ψ are (normal faithful semifinite) left and right Haar weights on L∞(G),
respectively. The left Haar weight induces an inner product
(10)
on the subspace Rϕ = {x ∈ L∞(G) : ϕ(x∗x) < ∞}. We let L2(G, ϕ) denote
the completion of Rϕ with respect to the norm induced by (10). We define
the Hilbert space L2(G, ψ) in a similar way.
hx, yiϕ = ϕ(y∗x)
Let W ∈ B(L2(G, ϕ) ⊗ L2(G, ϕ)) (resp. V ∈ B(L2(G, ψ) ⊗ L2(G, ψ))) be
the left (resp. right) fundamental unitary associated with G and note (see
(7)) that
Γ = ΓW = Γ′
V .
Let L1(G) be the predual of L∞(G). The pre-adjoint of Γ induces an asso-
ciative completely contractive multiplication ∗ on L1(G) by letting
f1 ∗ f2 = (f1 ⊗ f2) ◦ Γ.
The left regular representation λ : L1(G) → B(L2(G, ϕ)) is defined by
(11)
λ(f ) = Lf (W );
note that λ is an injective completely contractive algebra homomorphism.
We have that L∞( G)
is the von Neumann algebra
associated to the dual quantum group G of G. If Σ is the flip operator on
L2(G, ϕ) ⊗ L2(G, ϕ) and W = ΣW ∗Σ, its co-multiplication Γ is given by
def
= {λ(f ) : f ∈ L1(G)}′′
Γ(x) = W ∗(1 ⊗ x) W ,
x ∈ L∞( G).
The norm closure
C0( G)
def
= {λ(f ) : f ∈ L1(G)}
is referred to as the reduced C*-algebra associated with G.
Analogously, the right regular representation ρ : L1(G) → B(L2(G, ψ)) is
given by
ρ(g) = Rg(V ),
g ∈ L1(G);
10
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
note that ρ is an injective completely contractive algebra homomorphism.
We have that L∞( G′) = {ρ(g) : g ∈ L1(G)}
is the von Neumann algebra
associated with a quantum group denoted G′. We note that L∞( G′) =
L∞( G)′, W ∈ L∞(G) ¯⊗L∞( G) and V ∈ L∞( G′) ¯⊗L∞(G).
′′
Let
Γop : L∞(G) → L∞(G) ¯⊗L∞(G)
be the map given by Γop(x) = (σ ◦ Γ)(x), where
σ : L∞(G) ¯⊗L∞(G) → L∞(G) ¯⊗L∞(G),
(12)
is the flip. Note that Γop = Γ′
. It is known that L2(G, ϕ) and L2(G, ψ)
W
can be canonically identified; we use L2(G) for this Hilbert space in the rest
of this paper.
Theorem 3.1. Let G be a locally compact quantum group and χ ∈ L∞(G)⊗σ h
L∞(G). The following are equivalent:
σ(T ) = ΣT Σ,
(i) Φχ(C0( G)) ⊆ L∞( G);
(ii) T ◦ (Γ ⊗ Γop)(χ) ∈ L∞(G) ¯⊗1 ¯⊗L∞(G).
Proof. Note that, if f, g ∈ L1(G) then (f ⊗id)(W ) = λ(f ) and (id ⊗g)( W ) =
λ(g). Thus,
A(W, W ) = [λ(f )λ(g) : f, g ∈ L1(G)]
k·k
and hence, by equation [14, 2.5], AW, W = C0( G).
By Theorem 2.3, condition (i) is equivalent to the condition
T ◦ (Γ ⊗ Γop)(χ) ∈ B(L2(G)) ¯⊗L∞( G) ¯⊗B(L2(G)).
On the other hand, by Lemma 2.1,
T ◦ (Γ ⊗ Γop)(χ) ∈ L∞(G) ¯⊗L∞(G) ¯⊗L∞(G).
Since L∞( G) ∩ L∞(G) = C1, we conclude by [27] and Theorem 2.3 that (i)
and (ii) are equivalent.
(cid:3)
Corollary 3.2. Let G be a locally compact quantum group and suppose that
χ ∈ L∞(G) ⊗eh L∞(G). The following are equivalent:
(i) Φχ(L∞( G)) ⊆ L∞( G);
(ii) T ◦ (Γ ⊗ Γop)(χ) ∈ L∞(G) ¯⊗1 ¯⊗L∞(G).
Proof. The statement follows from Theorem 3.1, the weak* continuity of Φχ
and the fact that C0( G) is weak* dense in L∞( G).
(cid:3)
4. The co-commutative case
In this section, we specialise the results from Section 3 to co-commutative
locally compact quantum groups G [18]. Let G be a locally compact group
equipped with left Haar measure. We thus consider the case where L∞(G)
coincides with the von Neumann algebra VN(G) of G, acting on the Hilbert
space L2(G). We have that C0(G) is equal to the reduced groups C*-algebra
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
11
r (G) of G. We denote by λ (resp. ρ) the left (resp. right) regular repre-
C ∗
sentation of G on L2(G). Using the canonical identification
(13)
VN(G) ¯⊗ VN(G) ≡ VN(G × G),
we have that the co-multiplication Γ : VN(G) → VN(G × G) is given by
Γ(λ(s)) = λ(s) ⊗ λ(s),
s ∈ G.
In this case, L1(G) coincides with the Fourier algebra A(G) of G. We refer
the reader to [11] for the basic concepts and results from Abstract Harmonic
Analysis that will be used in this section, but note here that A(G) consists
of (complex valued) continuous functions on G vanishing at infinity, the
operation is point-wise, and its Gelfand spectrum can be homeomorphically
identified with G.
Note that, in the case under consideration, L∞( G) coincides with L∞(G),
which we hereafter view as the von Neumann algebra of operators on L2(G)
of multiplication by the corresponding essentially bounded functions. Un-
der this identification, C0( G) coincides with the subalgebra C0(G) of all
continuous functions vanishing at infinity.
Equation (13) allows us to identify the predual A(G × G) of VN(G × G)
with the operator projective tensor product A(G) ⊗A(G). Note that the pre-
adjoint of Γ is the associative multiplication map mA : A(G) ⊗A(G) → A(G)
given by mA(f ⊗ g) = f g.
We set
Ah(G) = A(G) ⊗h A(G) and Aeh(G) = A(G) ⊗eh A(G);
note that there is a completely isometric inclusion Ah(G) ⊆ Aeh(G) [10].
We have completely isometric identifications
Ah(G)∗ = VN(G) ⊗eh VN(G) and Aeh(G)∗ = VN(G) ⊗σ h VN(G).
In [1], Aeh(G) (resp. Ah(G)) was identified with a completely contractive
Banach algebra of separately (resp. jointly) continuous complex-valued func-
tions on G × G, equipped with pointwise multiplication; specifically, the
function, corresponding to an element v ∈ Aeh(G) is (denoted in the same
fashion and) given by
v(s, t) = hλs ⊗ λt, vi,
s, t ∈ G.
Note that Ah(G) is a regular (semi-simple) Banach algebra whose Gelfand
spectrum can be homeomorphically identified with G × G.
Recall that the multiplication map m : VN(G) × VN(G) → VN(G) (given
by m(S, T ) = ST ) extends uniquely to a weak* continuous completely con-
tractive map (denoted in the same fashion) m : VN(G)⊗σhVN(G) → VN(G).
If m∗ : A(G) → A(G) ⊗eh A(G) is the pre-adjoint of m then
m∗(f )(s, t) = f (st),
f ∈ A(G), s, t ∈ G.
12
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
M. Daws has shown [8] that, if f ∈ A(G) then m∗(f ) belongs to the algebra
M cbAh(G) of completely bounded multipliers [1] of Ah(G); in particular,
m∗(f )u ∈ Ah(G), whenever u ∈ Ah(G).
Moreover,
(14)
m∗ : A(G) → M cbAh(G)
is completely contractive (we refer the reader to [8, Theorem 9.2] and the
remark after its proof).
Theorem 3.1 specialises in the case under consideration to the follow-
ing result, providing a characterisation of the completely bounded maps on
B(L2(G)) that are not necessarily normal and which map C0(G) into L∞(G).
Corollary 4.1. Let G be a locally compact group and χ ∈ VN(G) ⊗σ h
VN(G). The following are equivalent:
(i) Φχ(C0(G)) ⊆ L∞(G);
(ii) T ◦ (Γ ⊗ Γop)(χ) ∈ VN(G) ¯⊗1 ¯⊗ VN(G).
If χ ∈ VN(G) ⊗σ h VN(G) and v ∈ Aeh(G), we let v · χ be the (unique)
element of VN(G) ⊗σ h VN(G) such that
hv · χ, wi = hχ, vwi, w ∈ Aeh(G).
If χ ∈ VN(G) ⊗eh VN(G) and v ∈ Ah(G) then v · χ denotes the similarly
defined element of VN(G) ⊗eh VN(G). If χ ∈ VN(G) ⊗eh VN(G), we denote
by suppeh(χ) the support of χ when the latter is viewed as a functional on
Ah(G); thus,
supp eh(χ) = {z ∈ G × G : v ∈ Ah(G), v(z) 6= 0 =⇒ v · χ 6= 0}
(see [11] and [16]). Let p : VN(G) ⊗σ h VN(G) → VN(G) ⊗eh VN(G) be the
natural projection given by p(χ) = χAh(G) (or, equivalently, the dual of the
inclusion map Ah(G) → Aeh(G)), and define suppσ(χ) = suppeh(p(χ)).
If E ⊆ G × G is a closed set, let
Jh(E) = {u ∈ Ah(G) : u has compact support, disjoint from E}
k·k
,
and write Jh(E)⊥,eh (resp. Jh(E)⊥,σ) for its annihilator in VN(G)⊗ehVN(G)
(resp. VN(G) ⊗σ h VN(G)). It is well-known that
Jh(E)⊥,eh = {χ ∈ VN(G) ⊗eh VN(G) : supp eh(χ) ⊆ E};
(15)
the following, similar, description of Jh(E)⊥,σ is straightforward from the
definitions.
Remark 4.2. Let E be a closed subset of G × G. Then
Jh(E)⊥,σ = {χ ∈ VN(G) ⊗σ h VN(G) : supp σ(χ) ⊆ E}.
Let
∇ = {(s, s−1) : s ∈ G}
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
13
be the anti-diagonal of G × G. Recall the flip σ : VN(G × G) → VN(G × G)
defined in (12) and note that
(id ⊗σ)(cid:0)a ⊗ b ⊗ c(cid:1) = a ⊗ c ⊗ b,
a, b, c ∈ VN(G).
In condition (iii) of the next theorem, we identify VN(G) ⊗eh VN(G) with a
subspace of VN(G × G) (see [4, Corollary 3.8]).
Theorem 4.3. Let χ ∈ VN(G) ⊗σ h VN(G). The following are equivalent:
(i) T ◦ (Γ ⊗ Γ)(χ) ∈ VN(G) ¯⊗1 ¯⊗ VN(G);
(ii) suppσ(χ) ⊆ ∇.
If χ ∈ VN(G) ⊗eh VN(G) then these conditions are also equivalent to
(iii) T ◦ (Γ ⊗ Γ)(χ) = (id ⊗σ)(cid:0)χ ⊗ 1(cid:1).
Proof. Assume that χ = λs ⊗ λt, where s, t ∈ G. For f, g, ω ∈ A(G), using
Lemma 2.1, we obtain
hT ◦ (Γ ⊗ Γ)(χ), f ⊗ ω ⊗ gi = hT (λs ⊗ λs ⊗ λt ⊗ λt), f ⊗ ω ⊗ gi
= hλs ⊗ λst ⊗ λt, f ⊗ ω ⊗ gi
= f (s)ω(st)g(t) = f (s)m∗(ω)(s, t)g(t)
= hλs ⊗ λt, m∗(ω)(f ⊗ g)i
= hχ, m∗(ω)(f ⊗ g)i
= hm∗(ω) · χ, f ⊗ gi.
It is easy to see that the linear span of {λs ⊗ λt : s, t ∈ G} is weak*
dense in VN(G) ⊗σ h VN(G). Further, since the map ψ → m∗(ω) · ψ on
VN(G) ⊗σ h VN(G) is weak* continuous, we have that
(16)
hT ◦ (Γ ⊗ Γ)(χ), f ⊗ ω ⊗ gi = hm∗(ω) · χ, f ⊗ gi,
for all χ ∈ VN(G) ⊗σ h VN(G) and all f, g, ω ∈ A(G).
(i)⇒(ii) By assumption, there exists ϕ ∈ VN(G) ¯⊗ VN(G) such that
T ◦ (Γ ⊗ Γ)(χ) = (id ⊗σ)(ϕ ⊗ 1).
Thus, if f, g, ω ∈ A(G) then (16) implies
(17)
hm∗(ω) · χ, f ⊗ gi = h(id ⊗σ)(ϕ ⊗ 1), f ⊗ ω ⊗ gi = hω(e)ϕ, f ⊗ gi.
Fix (s, t) ∈ G × G such that st 6= e. Let f, g ∈ A(G) be such that
f (s)g(t) 6= 0. Choose ω1, ω2 ∈ A(G) such that ω1(e) = ω2(e) but ω1(st) 6=
ω2(st). Thus,
(18)
(m∗(ω1) − m∗(ω2))(f ⊗ g)(s, t) 6= 0.
By (17), for all f0, g0 ∈ A(G), we have
h(m∗(ω1) − m∗(ω2))(f ⊗ g) · χ, f0 ⊗ g0i
= h(m∗(ω1) − m∗(ω2)) · χ, (f f0) ⊗ (gg0)i
= h(ω1(e) − ω2(e))ϕ, (f f0) ⊗ (gg0)i = 0.
14
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
We have that (m∗(ω1) − m∗(ω2))(f ⊗ g) ∈ Ah(G) and hence
h(m∗(ω1) − m∗(ω2))(f ⊗ g) · p(χ), f0 ⊗ g0i = 0.
Since A(G) ⊙ A(G) is dense in Ah(G), we conclude that
(m∗(ω1) − m∗(ω2))(f ⊗ g) · p(χ) = 0.
In view of (18), this shows that (s, t) 6∈ suppeh(p(χ)), and thus suppσ(χ) ⊆
∇.
(ii)⇒(i) Assume that suppσ(χ) ⊆ ∇. By Remark 4.2,
hp(χ), ui = 0 for any u ∈ Ah(G) ∩ Cc(G × G) with supp(u) ∩ ∇ = ∅.
Let u ∈ Ah(G) ∩ Cc(G × G). Choose h ∈ A(G) that takes the value 1 on the
(compact) set {st : (s, t) ∈ supp u} ∪ {e}. Since {e} is a set of synthesis for
A(G), for every ω ∈ A(G) with ω(e) = 1, there exists hn ∈ A(G) ∩ Cc(G),
n ∈ N, vanishing on a neighbourhood of {e}, such that
kω − h − hnkA(G) →n→∞ 0.
Since m∗ is a complete contraction from A(G) into M cbAh(G) [8], we have
that m∗(ω − h − hn) ∈ M cb Ah(G) and
km∗(ω − h − hn)ukh ≤ kω − h − hnkA(G)kukh →n→∞ 0.
Since supp(m∗(hn)u) ∩ ∇ = ∅, we have that hp(χ), m∗(hn)ui = 0. On the
other hand, m∗(h)u = u and hence
hp(χ), (m∗(ω) − 1)ui = hp(χ), (m∗(ω) − m∗(h))ui
= hp(χ), (m∗(ω − h − hn)ui →n→∞ 0,
i.e. hp(χ), (m∗(ω) − 1)ui = 0. In particular, since m∗(ω)(f ⊗ g) ∈ Ah(G), we
obtain
hχ, m∗(ω)(f ⊗ g)i = hp(χ), m∗(ω)(f ⊗ g)i = hp(χ), f ⊗ gi = hχ, f ⊗ gi
for all f , g ∈ A(G) ∩ Cc(G). Using (16), we obtain that, if f, g, ω ∈ A(G)
then
h(f ⊗ id ⊗g)(T ◦ (Γ ⊗ Γ)(χ)), ωi = hT ◦ (Γ ⊗ Γ)(χ), f ⊗ ω ⊗ gi
= ω(e)hχ, f ⊗ gi = hhχ, f ⊗ gi1, ωi.
Thus,
(f ⊗ id ⊗g)(T ◦ (Γ ⊗ Γ)(χ)) = hχ, f ⊗ gi1,
(19)
It follows that T ◦ (Γ ⊗ Γ)(χ)) ∈ VN(G) ¯⊗1 ¯⊗ VN(G).
f, g ∈ A(G).
Now suppose that χ ∈ VN(G) ⊗eh VN(G).
In this case, (19) can be
rewritten as
hT ◦ (Γ ⊗ Γ)(χ), f ⊗ ω ⊗ gi = h(id ⊗σ)(cid:0)χ ⊗ 1(cid:1), f ⊗ ω ⊗ gi,
completing the proof.
(cid:3)
In the next corollary, M (G) denotes the Banach algebra of all complex
Borel measures on G. The equivalence (i)⇔(ii) extends [1, Theorem 6.10].
The integral in (iii) is understood in the weak sense.
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
15
Corollary 4.4. Let χ ∈ VN(G) ⊗eh VN(G). The following are equivalent
(i) Φχ ∈ CBσ,L∞(G)
(ii) supp χ ⊆ ∇;
VN(G)′ (B(L2(G)));
(iii) there exists µ ∈ M (G) such that Φχ(x) = RG λsxλ∗
B(L2(G)).
sdµ(s), x ∈
Proof. The equivalence (i)⇔(ii) follows from Corollary 3.2 and Theorem 4.3,
while the equivalence (i)⇔(iii) was proved in [19].
(cid:3)
5. The commutative case
In this section, we specialise the results from Section 3 to commutative
locally compact quantum groups G. Thus, fixing a second countable lo-
cally compact group G, we have that L∞(G) = L∞(G), C0(G) = C0(G),
L∞( G) = VN(G) and C0( G) = C ∗
r (G). We identify the tensor product
L∞(G) ¯⊗L∞(G) with the von Neumann algebra L∞(G × G) in the canonical
way. It is readily verified that the flip σ : L∞(G × G) → L∞(G × G) is given
by
(20) σ(h)(s, t) = h(t, s), for almost all (s, t) ∈ G × G, h ∈ L∞(G × G),
while the co-multiplication Γ : L∞(G) → L∞(G × G) -- by
Γ(a)(s, t) = a(st), for almost all (s, t) ∈ G × G, a ∈ L∞(G).
(21)
It follows that Γop : L∞(G) → L∞(G × G) is given by
Γop(a)(s, t) = a(ts), for almost all (s, t) ∈ G × G, a ∈ L∞(G).
(22)
Note that the predual of L∞(G) is (completely) isometric to L1(G) and the
pre-adjoint of Γ is the usual convolution product:
(f ∗ g)(t) =Z f (s)g(s−1t)dt,
t ∈ G, f, g ∈ L1(G).
Theorem 3.1 specialises in the case under consideration to the follow-
ing result, providing a characterisation of the completely bounded maps on
B(L2(G)) that are not necessarily normal and which map C ∗
r (G) into VN(G).
Corollary 5.1. Let χ ∈ L∞(G) ⊗σ h L∞(G). The following are equivalent:
(i) Φχ(C ∗
(ii) T ◦ (Γ ⊗ Γop)(χ) ∈ L∞(G) ¯⊗1 ¯⊗L∞(G).
r (G)) ⊆ VN(G);
In the rest of this subsection, we restrict our attention to the case where
the elements χ from Corollary 5.1 belong to the extended Haagerup tensor
product L∞(G)⊗ehL∞(G). Those are precisely the elements χ ∈ L∞(G)⊗σ h
L∞(G), for which the corresponding map Φχ is (L∞(G)-bimodular and)
weak* continuous, and are widely known in the literature as Schur multipli-
ers.
Set V ∞(G) = L∞(G) ⊗eh L∞(G) and note that V ∞(G) = (L1(G) ⊗h
L1(G))∗, up to a complete isometry. Let χ ∈ V ∞(G) and assume that
χ ∼ P∞
that
16
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
i=1 ai ⊗ bi, where (ai)i∈N and (bi)i∈N in L∞(G) are sequences such
∞
∞
ai(s)2 < ∞ and esssup
t∈G
esssup
s∈G
Xi=1
bi(t)2 < ∞
Xi=1
[4]. We identify χ with the essentially bounded function (denoted in the
same way) χ : G × G → C, given by
(23)
χ(s, t) =
ai(s)bi(t),
s, t ∈ G,
∞
Xi=1
and thus consider V ∞(G) as a subspace of L∞(G × G).
Given a function ϕ : G × G → C and r ∈ G, let ϕr : G × G → C be the
function given by ϕr(s, t) = ϕ(sr, tr).
Lemma 5.2. Let χ ∈ V ∞(G) and r ∈ G. Then χr ∈ V ∞(G) and
Φχr (T ) = ρrΦχ(ρ∗
rT ρr)ρ∗
r,
T ∈ B(L2(G)).
Proof. For a ∈ L∞(G) let ar ∈ L∞(G) be given by ar(s) = a(sr), s ∈ G.
A direct verification shows that ρrMaρ∗
i=1 ai ⊗
i=1(ai)r(s)(bi)r(t), for almost all (s, t). Now, if T ∈
r = Mar . Clearly, if χ ∼ P∞
bi then χr(s, t) = P∞
B(L2(G)) then
ρrΦχ(ρ∗
rT ρr)ρ∗
r =
=
∞
∞
Xi=1
Xi=1
(ρrMaiρ∗
r)T (ρrMbiρ∗
r)
M(ai)r T M(bi)r = Φχr (T ).
(cid:3)
Recall that, if X is a measure space and X × X is equipped with the
product measure, a measurable subset E ⊆ X × X is called marginally null
[2] if there exists a null set M ⊆ X such that E ⊆ (M × X) ∪ (X × M ). If
w1, w2 : X × X → C are measurable functions, we say that w1 and w2 are
equal marginally almost everywhere (m.a.e.) if the set {(x, y) : w1(x, y) 6=
w2(x, y)} is marginally null. We note that if χ ∈ V ∞(G) then the function χ
is well-defined up to a marginally null subset of G×G and that it completely
determines the corresponding map Φχ (see [25]).
In the proof of the following lemma, given an element of L∞(G × G), we
choose any representative of its equivalence class, and denote it still by the
same symbol.
Lemma 5.3. Let χ ∈ V ∞(G × G). The function χ : G × G × G → C, given
by
χ(s, r, t) = χ(s, r, r, t),
s, r, t ∈ G,
is a well-defined element of L∞(G × G × G).
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
17
Proof. For a measurable function c : Z → R, where Z is a measure space,
and a positive real number δ, let Ec(δ) = {z ∈ Z : c(z) > δ}.
Suppose that χ = a ⊗ b, where a, b ∈ L∞(G × G). In order to show that χ
is measurable in this case, it suffices to assume that a ≥ 0 and b ≥ 0. Then
E χ(δ) = ∪λ∈Q+{(s, r, t) : (s, r) ∈ Ea(λ) and (r, t) ∈ Eb(δ/λ)}.
If F ⊆ G × G is a set, setting F r = {s ∈ G : (s, r) ∈ F } and Fr = {t ∈ G :
(r, t) ∈ F }, we have
{(s, t) : (s, r, t) ∈ E χ(δ)} = ∪λ∈Q+(Ea(λ)r × G) ∩ (G × Eb(δ/λ)r),
for each r ∈ G. It follows that χ is a measurable function, when G × G × G
is equipped with product measure.
For a general χ ∈ V ∞(G × G), write χ ∼P∞
subset of G × G such that
∞
i=1 ai ⊗ bi and let M be a null
∞
Xi=1
Set χn =Pn
ai(s, t)2 < C and
bi(s, t)2 < C whenever (s, t) 6∈ M.
Xi=1
i=1 ai ⊗ bi, n ∈ N. It is clear that
χ(s, r, t) = lim
n→∞
χn(s, r, t),
(s, r, t) 6∈ (M × G) ∪ (G × M ).
It follows from the previous paragraph that χ is measurable.
Let P∞
i=1 ai ⊗ bi and P∞
j=1 cj ⊗ dj be two w*-representations of the same
element χ ∈ V ∞(G × G). Then the corresponding functions on G × G ×
G × G, let us call them temporarily χ1 and χ2, are equal marginally almost
everywhere on (G × G) × (G × G) (see [25] for details). Let M ⊆ G × G
be a null set such that χ1(s, r, p, t) = χ2(s, r, p, t) whenever ((s, r), (p, t)) 6∈
(M × (G × G)) ∪ ((G × G) × M ). Then χ1(s, r, r, t) = χ2(s, r, r, t) whenever
(s, r) 6∈ M and (r, t) 6∈ M , that is, whenever (s, r, t) 6∈ (M × G)∪ (G× M ). It
follows that the function χ is a well-defined element of L∞(G × G × G). (cid:3)
Lemma 5.4. Let G be a second countable locally compact group and χ ∈
V ∞(G × G). Then T (χ) = χ.
Proof. For ω, f , g ∈ L1(G), we have
hT (χ), f ⊗ ω ⊗ hi = hm((Lf ⊗ Rg)(χ)), ωi.
Suppose that (ai)i∈N, (bi)i∈N ∈ L∞(G × G) and C > 0 are such that
ai(s, t)2 < C and
bi(s, t)2 < C for almost all (s, t)
∞
Xi=1
∞
i=1 ai ⊗ bi.
Xi=1
and χ ∼P∞
Let χn =Pn
Lf (ai)(r) =ZG
i=1 ai ⊗ bi, n ∈ N. It is easy to see that
ai(s, r)f (s)ds and Rg(bi)(r) =ZG
bi(r, t)g(t)dt
18
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
and hence
(24)
hm((Lf ⊗ Rg)(χn)), ωi =ZGZG×G
n
Xi=1
ai(s, r)bi(r, t)f (s)ω(r)g(t)ds dt dr.
By [10, p. 147], χn →n→∞ χ in the weak* topology of L∞(G × G) ⊗σ h
L∞(G×G); hence the left hand side of (24) converges to hm((Lf ⊗Rg)(χ)), ωi.
On the other hand, χn →n→∞ χ marginally almost everywhere on (G ×
G) × (G × G). Similarly to the proof of Lemma 5.3, we see that χn →n→∞ χ
almost everywhere on G × G × G. Moreover,
χn(s, r, t) ≤ n
Xi=1
ai(s, r)2!1/2 n
Xi=1
bi(r, t)2!1/2
< C,
for almost all (s, r, t) ∈ G × G × G, and hence Lebesgue's Dominated Con-
vergence Theorem implies that the right hand side of (24) converges to
Thus,
ZGZG×G
χ(s, r, t)f (s)ω(r)g(t)dsdtdr.
hT (χ), f ⊗ ω ⊗ gi = h χ, f ⊗ ω ⊗ gi,
and the proof is complete.
(cid:3)
Lemma 5.5. Let χ ∈ V ∞(G) and χ : G × G × G × G → C be the function
given by χ(s, r, p, t) = χ(sr, tp). Then χ ∈ V ∞(G × G) and
(25)
(Γ ⊗ Γop)(χ) = χ, m.a.e. on (G × G) × (G × G).
Proof. Suppose that χ ∼ P∞
i=1 ai ⊗ bi is a w*-representation of χ. By the
functoriality of the weak* Haagerup tensor product, Γ⊗Γop is a well-defined
map from V ∞(G) into V ∞(G × G) and
∞
(Γ ⊗ Γop)(χ) ∼
(see [4, p 136]). However,
n
Γ(ai) ⊗ Γop(bi)
Xi=1
Γ(ai) ⊗ Γop(bi) →n→∞ χ, m.a.e. on (G × G) × (G × G).
Xi=1
In fact, there is a null set M ⊆ G such that
∞
n
Xi=1
ai(sr)bi(tp) →n→∞
ai(sr)bi(tp)
Xi=1
whenever (sr, tp) /∈ (M × G) ∪ (G × M ). Letting M ♯ = {(s, r) : sr ∈ M }
and M ♭ = {(p, t) : tp ∈ M }, we have the convergence whenever (s, r, p, t) /∈
(M ♯ × (G × G)) ∪ (G × G) × M ♭. We finally note that M ♯ and M ♭ have
measure zero. The conclusion follows.
(cid:3)
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
19
Set
V ∞
inv(G) = {χ ∈ V ∞(G) : χr = χ a.e., for all r ∈ G}
(see [25] and [24]).
Theorem 5.6. Let G be a second countable locally compact group and χ ∈
V ∞(G). The following are equivalent:
(i) Φχ(VN(G)) ⊆ VN(G);
(ii) T ◦ (Γ ⊗ Γop)(χ) ∈ L∞(G) ¯⊗1 ¯⊗L∞(G);
(iii) T ◦ (Γ ⊗ Γop(χ)) = (id ⊗σ)(cid:0)χ ⊗ 1(cid:1);
(iv) χ ∈ V ∞
inv(G).
Proof. (i)⇔(ii) follows from Corollary 5.1 and the weak* continuity of Φχ.
(ii)⇒(iv) By Lemmas 5.4 and 5.5,
(26)
T ◦ (Γ ⊗ Γop(χ))(s, r, t) = (Γ ⊗ Γop)(χ)(s, r, r, t) = χr(s, t),
for almost all (s, r, t) ∈ G × G × G. Thus, there exist a null set M ⊆ G and
a function ϕ ∈ L∞(G × G) such that
χr = ϕ a.e.,
if r 6∈ M.
Let ξ, ξ0, η, η0 ∈ L2(G) and write ξη∗ for the rank one operator on L2(G)
given by (ξη∗)(ζ) = hζ, ηiξ. By Lemma 5.2, ϕ ∈ V ∞(G) and, whenever
r 6∈ M , we have
hΦϕ(ξη∗)ξ0, η0i = hρrΦχ(ρ∗
= hΦχ((ρ∗
r(ξη∗)ρr)ρ∗
rη)∗)ρ∗
rξ)(ρ∗
rξ0, η0i
rξ0, ρ∗
rη0i.
Since the right regular representation is strongly continuous, the map from
G into B(L2(G)), sending r to (ρ∗
rη)∗, is norm continuous. Since the
family {(ρ∗
rη)∗ : r ∈ G} is uniformly bounded, we have that the map
r ξ)(ρ∗
rξ)(ρ∗
r → hΦχ((ρ∗
rη0i
r ξ)(ρ∗
rη)∗)ρ∗
rξ0, ρ∗
is continuous. Now Lemma 5.2 implies that
hΦϕ(ξη∗)ξ0, η0i = hΦχr (ξη∗)ξ0, η0i,
r ∈ G,
which shows that χr = ϕ almost everywhere for every r ∈ G; in particular,
ϕ = χ almost everywhere, and (iv) is established.
(iv)⇒(iii) By (20),
(27) (id ⊗σ)(χ ⊗ 1)(s, r, t) = χ(s, t),
for almost all (s, r, t) ∈ G × G × G,
and the claim now follows from (26).
(iii)⇒(ii) follows from (27).
(cid:3)
Acknowledgement. We are grateful to Jason Crann for a number of
fruitful conversations on the topic of this paper.
20
M. ALAGHMANDAN, I. G. TODOROV, AND L. TUROWSKA
References
[1] M. Alaghmandan, I. G. Todorov and L. Turowska, Completely bounded bi-
module maps and spectral synthesis, Internat. J. Math. 28 (2017), no. 10, 1750067, 40
pp.
[2] W. B. Arveson, Operator algebras and invariant subspaces, Ann. Math. (2) 100
(1974), 433-532.
[3] D. P. Blecher and C. Le Merdy, Operator algebras and their modules -- an oper-
ator space approach, Oxford University Press, 2004.
[4] D. Blecher and R. R. Smith, The dual of the Haagerup tensor product, J. London
Math. Soc. (2) 45 (1992), 126-144.
[5] M. Bozejko and G. Fendler, Herz-Schur multipliers and completely bounded mul-
tipliers of the Fourier algebra of a locally compact group, Boll. Un. Mat. Ital. A (6) 2
(1984), no. 2, 297-302.
[6] N. P. Brown and N. Ozawa, C*-algebras and finite dimensional approximations,
American Mathematical Society, 2008.
[7] J. de Canniere and U. Haagerup, Multipliers of the Fourier algebras of some
simple Lie groups and their discrete subgroups, Amer. J. Math. 107 (1985), no. 2,
455-500.
[8] M. Daws, Multipliers, self-induced and dual Banach algebras, Dissertationes Math.
470 (2010), 62 p.
[9] E. G. Effros and A. Kishimoto, Module maps and Hochschild-Johnson cohomol-
ogy, Indiana Univ. Math. J. 36 (1987), 257-276.
[10] E. G. Effros and Z.-J. Ruan, Operator space tensor products and Hopf convolution
algebras, J. Operator Theory 50 (2003), no. 1, 131-156.
[11] P. Eymard, L'alg`ebre de Fourier d'un groupe localement compact, Bull. Soc. Math.
France 92 (1964), 181-236.
[12] U. Haagerup, Decomposition of completely bounded maps on operator algebras, un-
published manuscript.
[13] U. Haagerup and J. Kraus, Approximation properties for group C*-algebras and
group von Neumann algebras, Trans. Amer. Math. Soc. 344 (1994), no. 2, 667-699.
[14] Z. Hu, M. Neufang and Z.-J. Ruan, Completely bounded multipliers over locally
compact quantum groups, Proc. London Math. Soc. (3) 103 (2011), 1-39.
[15] M. Junge, M. Neufang and Z.-J. Ruan, A representation theorem for locally com-
pact quantum groups, Internat. J. Math. 20 (2009) 377-400.
[16] Y. Katznelson, An introduction to harmonic analysis, Cambridge University Press,
2004.
[17] S. Knudby, The weak Haagerup property, preprint, arXiv 1401.7541.
[18] J. Kustermans and S. Vaes, Locally compact quantum groups, Ann. Scient. `Ecole
Normale Sup`erieure 33 (2000), no. 6, 837-934.
[19] M. Neufang, Abstrakte harmonische Analyse und Modulhomomorphismen uber von
Neumann-algebren, PhD thesis, Universitat des Saarlandes, 2000.
[20] M. Neufang, Z.-J. Ruan and N. Spronk, Completely isometric representations of
McbA(G) and UCB( G), Trans. Amer. Math. Soc. 360 (2008), 1133-1161.
[21] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge Univer-
sity Press, 2002.
[22] G. Pisier, An introduction to the theory of operator spaces, Cambridge University
Press, 2003.
[23] H. H. Shaefer, Topological vector spaces, Springer-Verlag, 1999.
[24] N. Spronk, Measurable Schur multipliers and completely bounded multipliers of the
Fourier algebras, Proc. London Math. Soc. (3) (1991), 156-175.
[25] I. G. Todorov, Herz-Schur multipliers, Lecture notes (2014), available at
https://www.fields.utoronto.ca/programs/scientific/13-14/harmonicanalysis/.
COMPLETELY BOUNDED MAPS AND INVARIANT SUBSPACES
21
[26] I. G. Todorov, Interactions between harmonic analysis and operator theory, Serdica
Math. J. 41 (2015), 13-34.
[27] J. Tomiyama, Tensor products and approximation problems of C ∗-algebras, Publ.
Res. Inst. Math. Sci. 11 (1975/76), 163-183.
[28] S. Vaes, Locally compact quantum groups, Ph.D. thesis, Katholieke Universitiet Leu-
ven (2001).
School of Mathematics and Statistics, Carleton University, Ottawa, ON,
Canada H1S 5B6
E-mail address: [email protected]
Mathematical Sciences Research Centre, Queen's University Belfast, Belfast
BT7 1NN, United Kingdom, and School of Mathematical Sciences, Nankai Uni-
versity, 300071 Tianjin, China
E-mail address: [email protected]
Department of Mathematical Sciences, Chalmers University of Technol-
ogy and the University of Gothenburg, Gothenburg SE-412 96, Sweden
E-mail address: [email protected]
|
1805.02037 | 1 | 1805 | 2018-05-05T10:27:54 | Nonlinear $\ast$-Jordan-Type Derivations on von Neumann Algebras | [
"math.OA"
] | Let $\mathcal{H}$ be a complex Hilbert space, $\mathcal{B(H)}$ be the algebra of all bounded linear operators on $\mathcal{H}$ and $\mathcal{A} \subseteq \mathcal{B(H)}$ be a von Neumann algebra without central summands of type $I_1$. For arbitrary elements $A, B\in \mathcal{A}$, one can define their $\ast$-Jordan product in the sense of $A\diamond B = AB+BA^\ast$. Let $p_n(x_1,x_2,\cdots,x_n)$ be the polynomial defined by $n$ indeterminates $x_1, \cdots, x_n$ and their $\ast$-Jordan products. In this article, it is shown that a mapping $\delta: \mathcal{A} \longrightarrow \mathcal{B(H)}$ satisfies the condition $$ \delta(p_n(A_1, A_2,\cdots, A_n))=\sum_{k=1}^n p_n(A_1,\cdots, A_{k-1}, \delta(A_k), A_{k+1},\cdots, A_n) $$ for all $A_1, A_2,\cdots, A_n \in \mathcal{A}$ if and only if $\delta$ is an additive $\ast$-derivation. | math.OA | math |
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON
NEUMANN ALGEBRAS
WENHUI LIN
Abstract. Let H be a complex Hilbert space, B(H) be the algebra of all
bounded linear operators on H and A ⊆ B(H) be a von Neumann algebra
without central summands of type I1. For arbitrary elements A, B ∈ A, one
can define their ∗-Jordan product in the sense of A ⋄ B = AB + BA∗. Let
pn(x1, x2, · · · , xn) be the polynomial defined by n indeterminates x1, · · · , xn
and their ∗-Jordan products. In this article, it is shown that a mapping δ :
A −→ B(H) satisfies the condition
δ(pn(A1, A2, · · · , An)) =
n
X
k=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A if and only if δ is an additive ∗-derivation.
Contents
1.
Introduction
2. Preliminaries
3. Main Theorem and Its Proof
4. Related Topics for Future Research
References
1
3
4
16
18
1. Introduction
Let A be an associative ∗-algebra over the complex field C. For any A, B ∈ A, one
can denote a "new product" of A and B by A⋄B = AB+BA∗, and this new product
⋄ is usually said to be ∗-Jordan product. Such kind of product based on Jordan
bracket naturally appears in relation with the so-called Jordan ∗-derivations and
plays an important role in the problem of representability of quadratic functionals
by sesqui-linear functionals on left-modules over ∗-algebras (see [17, 21, 22]). The
product is workable for us to characterize ideals, see [2, 16, 18]. Especial attention
has been paid to understanding mappings which preserve the product AB + BA∗
between ∗-algebras, see [4, 6, 8, 9, 27].
The question of to what extent the multiplicative structure of an algebra de-
termines its additive structure has been considered by many researchers over the
past decades. In particular, they have investigated under which conditions bijective
Date: August 8, 2018.
2010 Mathematics Subject Classification. 47B47, 46L10.
Key words and phrases. Nonlinear ∗-Jordan-type derivation, von Neumann algebra.
1
2
WENHUI LIN
mappings between algebras preserving the multiplicative structure necessarily pre-
serve the additive structure as well. The most fundamental result in this direction
is due to W. S. Martindale III [14] who proved that every bijective multiplicative
mapping from a prime ring containing a nontrivial idempotent onto an arbitrary
ring is necessarily additive. Later, a number of authors considered the Jordan-type
product or Lie-type product and proved that, on certain associative algebras or
rings, bijective mappings which preserve any of those products are automatically
additive, see [1 -- 4, 6, 8, 9, 19, 20, 27].
An additive mapping δ : A −→ A is called an additive derivation if δ(AB) =
δ(A)B + Aδ(B) for all A, B ∈ A. Furthermore, δ is said to be an additive ∗-
derivation provided that δ is an additive derivation and satisfies δ(A∗) = δ(A)∗ for
all A ∈ A. Let δ : A −→ A be a mapping (without the additivity assumption). We
say that δ is a nonlinear ∗-Jordan derivation if
δ(A ⋄ B) = δ(A) ⋄ B + A ⋄ δ(B),
holds true for all A, B ∈ A. Similarly, a mapping δ : A −→ A is called a nonlinear
∗-Jordan triple derivation if it satisfies the condition
δ(A ⋄ B ⋄ C) = δ(A) ⋄ B ⋄ C + A ⋄ δ(B) ⋄ C + A ⋄ B ⋄ δ(C)
for all A, B, C ∈ A, where A ⋄ B ⋄ C = (A ⋄ B) ⋄ C . We should be aware that ⋄ is
not necessarily associative.
Given the consideration of ∗-Jordan derivations and ∗-Jordan triple derivations,
we can further develop them in one natural way. Suppose that n ≥ 2 is a fixed
positive integer. Let us see a sequence of polynomials with ∗
p1(x1) = x1,
p2(x1, x2) = x1 ⋄ x2 = x1x2 + x2x∗
1,
p3(x1, x2, x3) = p2(x1, x2) ⋄ x3 = (x1 ⋄ x2) ⋄ x3,
p4(x1, x2, x3, x4) = p3(x1, x2, x3) ⋄ x4 = ((x1 ⋄ x2) ⋄ x3) ⋄ x4,
· · · · · · ,
pn(x1, x2, · · · , xn) = pn−1(x1, x2, · · · , xn−1) ⋄ xn
= (· · · ((
x1 ⋄ x2) ⋄ x3) ⋄ · · · ⋄ xn−1) ⋄ xn.
Accordingly, a nonlinear ∗-Jordan n-derivation is a mapping δ : A −→ A satisfying
the condition
n−2
{z }
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A. This notion makes the best use of the definition of
Lie-type derivations and that of ∗-Lie-type derivations, see [5, 12, 13]. By the
definition, it is clear that every ∗-Jordan derivation is a ∗-Jordan 2-derivation and
each ∗-Jordan triple derivation is a ∗-Jordan 3-derivation. One can easily check that
each nonlinear ∗-Jordan derivation on A is a nonlinear ∗-Jordan triple derivation.
But, we don't know whether the converse statement is true. ∗-Jordan 2-derivations,
∗-Jordan 3-derivations and ∗-Jordan n-derivations are collectively referred to as
∗-Jordan-type derivations. ∗-Jordan-type derivations on operator algebras have
been studied by several authors. Let H be a complex Hilbert space and B(H) be
the algebra of all bounded linear operators on H. Li et al in [10] showed that if
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
3
A ⊆ B(H) is a von Neumann algebra without central summands of type I1, then
δ : A −→ B(H) is a nonlinear ∗-Jordan derivation if and only if δ is an additive
∗-derivation. More recently, this result is extended to the case of nonlinear ∗-
Jordan triple derivations by Zhao and Li [26]. Taghavi et al [23] and Zhang [25]
independently investigate ∗-Jordan derivations on factor von Neumann algebras,
respectively. It turns out that each nonlinear ∗-Jordan derivation on a factor von
Neumann algebra is an additive ∗-derivation.
Inspired by the afore-mentioned works, we will concentrate on giving a descrip-
tion of nonlinear ∗-Jordan-type derivations on von Neumann algebras. The orga-
nization of this paper is as follows. We recall and collect some indispensable facts
with respect to von Neumann algebras in the second section 2. The third Section 3
is devoted to our main result Theorem 3.1 and its proof. The main theorem states
that every nonlinear ∗-Jordan-type derivation on a von Neumann algebra without
central summands of type I1 is an additive ∗-derivation. Similar statements are
also given for factor von Neumann algebras and standard operator algebras with-
out proofs. Some potential topics for the future research are presented in the last
Section 4.
2. Preliminaries
Throughout this paper, H denotes a complex Hilbert space and B(H) is the
algebra of all bounded linear operators on H. A von Neumann algebra A is weakly
closed, self-adjoint algebra of operators on H containing the identity operator I.
The set Z(A) = {S ∈ A : ST = T S for all T ∈ A} is called the centre of
A. A projection P is called a central abelian projection if P ∈ Z(A) and P AP is
abelian. Recall that the central carrier of A, denoted by A, is the smallest central
projection P satisfying the condition P A = A. It is straightforward to check that
the central carrier of A is the projection onto the closed subspace spanned by
{BA(x) : B ∈ A, x ∈ H}. If A is self-adjoint, then the core of A, denoted by A,
is sup{S ∈ Z(A) : S = S ∗, S ≤ A}. If P is a projection, it is clear that P is the
largest central projection Q with Q ≤ P . A projection P is said to be core-free if
P = 0. It is not difficult to see that P = 0 if and only if I − P = I.
To round off the proof of our main theorem, we need to give some necessary
lemmas.
Lemma 2.1. [15, Lemma 4] Let A be a von Neumann without central summands
of type I1. Then each nonzero central projection in A is the central carrier of a
core-free projection in A.
Lemma 2.2. [10, Lemma 2.2] Let A be a von Neumann algebra on a Hilbert space
H. Let A ∈ B(H) and P ∈ A is a projection with P = I. If ABP = 0 for all
B ∈ A, then A = 0.
Lemma 2.3. Let A be a von Neumann algebra without central summands of type
I1. For any A ∈ A and for any positive integer n ≥ 2, we have
and
pn(cid:18)A,
1
2
I, · · · ,
1
2
pn(cid:18) 1
2
I, · · · ,
1
2
(A + A∗)
1
2
I(cid:19) =
I, A(cid:19) = A.
4
WENHUI LIN
Proof. By a recursive calculation, we know that
pn(cid:18)A,
1
2
I, · · · ,
1
2
2
I(cid:19) = pn−1(cid:18) 1
= pn−2(cid:18) 1
= pn−3(cid:18) 1
2
2
(A + A∗),
I, · · · ,
(A + A∗) ,
I, · · · ,
(A + A∗) ,
I, · · · ,
1
2
1
2
1
2
I(cid:19)
I(cid:19)
I(cid:19)
1
2
1
2
1
2
Similarly, we also have
pn(cid:18) 1
2
I, · · · ,
1
2
= · · ·
=
1
2
(A + A∗) .
I, A(cid:19) = pn−1(cid:18) 1
= pn−2(cid:18) 1
2
2
I, · · · ,
I, · · · ,
I, A(cid:19)
I, A(cid:19)
1
2
1
2
= · · ·
= A.
(cid:3)
Let A be a von Neumann algebra without central summands of type I1. By
Lemma 2.1, we know that there exists a nonzero central projection P such that
P = 0 and P = I. For the convenience of discussion, let us set P1 = P , P2 = I − P .
i,j=1 Aij . We denote the imaginary
We write Aij = PiAPj. Thus one gets A = P2
unit by i.
Lemma 2.4. Let A be a von Neumann algebra without central summands of type
I1. For any A ∈ A with A = P2
(a) P1 ⋄ A = 0 implies that A12 = A21 = A11 = 0.
(b) P2 ⋄ A = 0 implies that A12 = A21 = A22 = 0.
(c) (P2 − P1) ⋄ A = 0 implies that A11 = A22 = 0.
i,j=1 Aij and Aij ∈ Aij , we have
Proof. Let us first prove the assertion (a). We have
0 = P1 ⋄ A = P1A + AP ∗
1 = P1A + AP1
= A12 + A21 + 2A11,
which leads to A12 = A21 = A11 = 0.
The other two assertions can be achieved by an analogous manner.
(cid:3)
3. Main Theorem and Its Proof
We are in a position to give the main theorem of this article which can be stated
as follows.
Theorem 3.1. Let A be a von Neumann algebra without central summands of type
I1. Then a mapping δ : A −→ B(H) satisfies the rule
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
5
for all A1, A2, · · · , An ∈ A if and only if δ is an additive ∗-derivation.
Proof. The proof of this theorem can be realized via a series of claims.
Claim 1. δ(0) = 0.
δ(0) = δ(pn(0, 0, · · · , 0)) =
Claim 2. For any A ∈ A, we have δ(cid:0) 1
Note that the fact 1
2 I, · · · , 1
n
k
pn(0, · · · , δ(0), · · · , 0) = 0.
Xk=1
2 I(cid:1) ⋄ A = 0.
2 I(cid:1). By Lemma 2.3 it follows that
2
2 I = pn(cid:0) 1
I(cid:19) =δ(cid:18)pn(cid:18) 1
pn(cid:18) 1
Xk=1
2 (cid:18)δ(cid:18) 1
n − 1
=
=
2
2
n
I, · · · ,
k
1
2
I(cid:19)(cid:19)
I, δ(cid:18) 1
I(cid:19) + δ(cid:18) 1
1
2
2
I, · · · ,
I, · · · ,
2
1
2
I(cid:19) ,
I(cid:19)∗(cid:19) + δ(cid:18) 1
2
1
2
I(cid:19)
I(cid:19) .
δ(cid:18) 1
2
This gives
That is,
δ(cid:18) 1
2
2
I(cid:19) + δ(cid:18) 1
δ(cid:18) 1
I(cid:19) ⋄
1
2
2
I(cid:19)∗
= 0.
I = 0.
(3.1)
Using the relation (3.1), we get
2
n−1
1
2
I, · · · ,
δ(A) =δ(cid:18)pn(cid:18) 1
pn(cid:18) 1
I(cid:19) A + Aδ(cid:18) 1
Xk=1
=δ(cid:18) 1
I, A(cid:19)(cid:19)
I, δ(cid:18) 1
I(cid:19)∗
I, · · · ,
1
2
=
2
2
2
2
k
+ δ(A).
I(cid:19) ,
1
2
I, · · · ,
I, A(cid:19) + pn(cid:18) 1
2
1
2
I, · · · ,
I, δ(A)(cid:19)
1
2
for all A ∈ A. Thus we obtain
That is,
δ(cid:18) 1
2
I(cid:19) A + Aδ(cid:18) 1
2
I(cid:19)∗
= 0.
δ(cid:18) 1
2
I(cid:19) ⋄ A = 0.
Claim 3. For any All ∈ All, Bij ∈ Aij (i, j, l = 1, 2, i 6= j), we have
δ(All + Bij ) = δ(All) + δ(Bij).
We only need to prove the case of i = l = 1, j = 2, and the proofs of the other cases
are rather similar and are omitted here. Let us write
M = δ(A11 + B12) − δ(A11) − δ(B12).
6
WENHUI LIN
It is sufficient for us to show that M = 0. Since
pn(cid:18) 1
2
I, · · · ,
I, P2, A11(cid:19) = 0
1
2
and
pn(cid:18) 1
2
I, · · · ,
I, P2, A11 + B12(cid:19) = pn(cid:18) 1
2
1
2
I, · · · ,
I, P2, B12(cid:19) ,
1
2
we by Claim 2 have
I, · · · ,
I, δ (P2) , A11 + B12(cid:19) + pn(cid:18) 1
2
1
2
I, · · · ,
I, P2, δ (A11 + B12)(cid:19)
1
2
2
2
pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, · · · ,
I, P2, A11 + B12(cid:19)(cid:19)
I, P2, A11(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
2
1
2
1
2
1
2
I, δ (P2) , A11(cid:19) + pn(cid:18) 1
I, δ (P2) , B12(cid:19) + pn(cid:18) 1
1
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
1
2
I, P2, B12(cid:19)(cid:19)
I, P2, δ (A11)(cid:19)
I, P2, δ (B12)(cid:19) .
1
2
1
2
We therefore get
pn(cid:18) 1
2
I, · · · ,
I, P2, M(cid:19) = 0.
1
2
In light of Lemma 2.3, we obtain
p2 (P2, M ) = P2 ⋄ M = 0.
It follows from Lemma 2.4 that
M12 = M21 = M22 = 0.
Notice that
and
pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, B12(cid:19) = (P2 − P1) ⋄ B12 = 0
1
2
pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, A11 + B12(cid:19) = pn(cid:18) 1
2
1
2
I, · · · ,
I, P2 − P1, A11(cid:19) .
1
2
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
7
By Claim 2, we observe that
I, · · · ,
1
2
I, δ(P2 − P1), A11 + B12(cid:19)
I, P2 − P1, δ (A11 + B12)(cid:19)
I, P2 − P1, A11 + B12(cid:19)(cid:19)
I, P2 − P1, B12(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
1
2
1
2
1
2
2
I, · · · ,
I, · · · ,
I, · · · ,
2
2
pn(cid:18) 1
+ pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
2
I, · · · ,
1
2
I, δ(P2 − P1), B12(cid:19) + pn(cid:18) 1
I, δ(P2 − P1), A11(cid:19) + pn(cid:18) 1
1
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, P2 − P1, A11(cid:19)(cid:19)
1
2
1
2
I, P2 − P1, δ (B12)(cid:19)
I, P2 − P1, δ (A11)(cid:19) .
1
2
I, · · · ,
Thus we arrive at
pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, M(cid:19) = 0.
1
2
Taking into account Lemma 2.3, we get
p2 (P2 − P1, M ) = (P2 − P1) ⋄ M = 0.
Applying Lemma 2.4 yields that
M11 = 0.
We therefore have M = 0. That is,
δ(A11 + B12) = δ(A11) + δ(B12).
The other cases can be verified by an analogous manner.
Claim 4. For any B12 ∈ A12, C21 ∈ A21, we have
δ(B12 + C21) = δ(B12) + δ(C21).
We only need to show that
M = δ(B12 + C21) − δ(B12) − δ(C21) = 0.
pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, B12(cid:19) = (P2 − P1) ⋄ B12 = 0
1
2
Since
and
pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, C21(cid:19) = pn(cid:18) 1
2
1
2
I, · · · ,
I, P2 − P1, B12 + C21(cid:19) = 0,
1
2
8
WENHUI LIN
we obtain
2
2
pn(cid:18) 1
+ pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
2
Hence, we have
I, · · · ,
1
2
I, δ(P2 − P1), B12 + C21(cid:19)
I, P2 − P1, δ (B12 + C21)(cid:19)
I, P2 − P1, B12 + C21(cid:19)(cid:19)
I, P2 − P1, B12(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
1
2
1
2
1
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, · · · ,
1
2
I, · · · ,
2
1
2
I, δ(P2 − P1), B12(cid:19) + pn(cid:18) 1
I, δ(P2 − P1), C21(cid:19) + pn(cid:18) 1
pn(cid:18) 1
I, · · · ,
1
2
2
2
I, P2 − P1, M(cid:19) = 0.
I, · · · ,
I, · · · ,
I, P2 − P1, C21(cid:19)(cid:19)
1
2
1
2
I, P2 − P1, δ (B12)(cid:19)
I, P2 − P1, δ (C21)(cid:19) .
1
2
I, · · · ,
Applying Lemma 2.3 gives
p2 (P2 − P1, M ) = (P2 − P1) ⋄ M = 0.
By Lemma 2.4, we know that
M11 = M22 = 0.
Note that the facts
and
pn(cid:18) 1
2
I, · · · ,
1
2
2
1
2
I, · · · ,
I, B12, P1(cid:19) = 0
pn(cid:18) 1
I, B12 + C21, P1(cid:19) = pn(cid:18) 1
2
I, · · · ,
I, C21, P1(cid:19) .
1
2
Using similar computations as the above, we get
pn(cid:18) 1
2
I, · · · ,
I, M, P1(cid:19) = 0.
1
2
In view of Lemma 2.3 and the fact M11 = M22 = 0, one can see that
On the other hand, we should remark that
M21 = 0.
pn(cid:18) 1
2
I, · · · ,
1
2
and
I, B12 + C21, P2(cid:19) = pn(cid:18) 1
pn(cid:18) 1
I, C21, P2(cid:19) = 0.
I, · · · ,
1
2
2
2
I, · · · ,
I, B12, P2(cid:19)
1
2
Using similar arguments as the above, one can get M12 = 0.
Claim 5. For all A11 ∈ A11, D22 ∈ A22, we have
δ(A11 + D22) = δ(A11) + δ(D22).
It is sufficient to prove that
M = δ(A11 + D22) − δ(A11) − δ(D22) = 0.
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
9
Since
and
pn(cid:18) 1
2
we konw that
I, · · · ,
I, P2, A11(cid:19) = 0
2
1
2
I, · · · ,
pn(cid:18) 1
I, P2, A11 + D22(cid:19) = pn(cid:18) 1
I, δ (P2) , A11 + D22(cid:19) + pn(cid:18) 1
1
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
1
2
1
2
I, P2, D22(cid:19) ,
I, P2, δ (A11 + D22)(cid:19)
1
2
2
2
pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
I, · · · ,
I, · · · ,
I, P2, A11 + D22(cid:19)(cid:19)
I, P2, D22(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
2
1
2
1
2
I, · · · ,
I, · · · ,
1
2
I, δ (P2) , D22(cid:19) + pn(cid:18) 1
I, δ (P2) , A11(cid:19) + pn(cid:18) 1
1
2
2
2
I, · · · ,
I, · · · ,
1
2
I, P2, A11(cid:19)(cid:19)
I, P2, δ (D22)(cid:19)
I, P2, δ (A11)(cid:19) .
1
2
1
2
I, · · · ,
Thus we obtain
1
2
By invoking of Lemma 2.3, we arrive at
I, · · · ,
pn(cid:18) 1
2
I, P2, M(cid:19) = 0.
It follows from Lemma 2.4 that
p2 (P2, M ) = P2 ⋄ M = 0.
We should remark that
pn(cid:18) 1
2
M12 = M21 = M22 = 0.
I, · · · ,
I, P1, D22(cid:19) = P1 ⋄ D22 = 0
1
2
and that
pn(cid:18) 1
2
I, · · · ,
I, P1, A11 + D22(cid:19) = pn(cid:18) 1
I, · · · ,
I, P1, A11(cid:19) .
1
2
1
2
2
Using similar discussions as the above, one can get
Hence we conclude that M = 0. That is,
M11 = 0.
δ(A11 + D22) = δ(A11) + δ(D22).
Claim 6. For any A11 ∈ A11, B12 ∈ A12, C21 ∈ A21 and D22 ∈ A22, we have
(a) δ(A11 + B12 + C21) = δ(A11) + δ(B12) + δ(C21),
(b) δ(B12 + C21 + D22) = δ(B12) + δ(C21) + δ(D22).
Let us first prove the result (a). For convenience, let us set
M = δ(A11 + B12 + C21) − δ(A11) − δ(B12) − δ(C21).
We shall prove that M = 0. In view of the facts
pn(cid:18) 1
2
I, · · · ,
I, P2, A11(cid:19) = 0
1
2
10
and
WENHUI LIN
pn(cid:18) 1
2
I, · · · ,
I, P2, A11 + B12 + C21(cid:19) = pn(cid:18) 1
2
1
2
I, · · · ,
I, P2, B12 + C21(cid:19) ,
1
2
we by Claim 4 get
2
2
pn(cid:18) 1
+ pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
2
I, · · · ,
1
2
I, δ (P2) , A11 + B12 + C21(cid:19)
I, P2, δ (A11 + B12 + C21)(cid:19)
I, P2, A11 + B12 + C21(cid:19)(cid:19)
I, P2, B12 + C21(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
1
2
1
2
1
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, · · · ,
1
2
I, δ (P2) , B12 + C21(cid:19) + pn(cid:18) 1
I, δ (P2) , A11(cid:19) + pn(cid:18) 1
1
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, P2, A11(cid:19)(cid:19)
1
2
I, · · · ,
I, P2, δ (B12) + δ (C21)(cid:19)
1
2
I, P2, δ (A11)(cid:19)
1
2
By Lemma 2.3 we know that
pn(cid:18) 1
2
I, · · · ,
I, P2, M(cid:19) = P2 ⋄ M = 0.
1
2
It follows from Lemma 2.4 that
In order to show M11 = 0, we should note that
M12 = M21 = M22 = 0.
pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, C21(cid:19) = 0
1
2
and that
pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, A11 + B12 + C21(cid:19) = pn(cid:18) 1
2
1
2
I, · · · ,
I, P2 − P1, A11 + B12(cid:19) .
1
2
Using Claim 3, we see that
2
2
pn(cid:18) 1
+ pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
2
I, · · · ,
1
2
I, δ(P2 − P1), A11 + B12 + C21(cid:19)
I, P2 − P1, δ (A11 + B12 + C21)(cid:19)
I, P2 − P1, A11 + B12 + C21(cid:19)(cid:19)
I, P2 − P1, A11 + B12(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
1
2
1
2
1
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, · · · ,
1
2
I, δ(P2 − P1), A11 + B12(cid:19) + pn(cid:18) 1
I, δ(P2 − P1), C21(cid:19) + pn(cid:18) 1
1
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, P2 − P1, C21(cid:19)(cid:19)
1
2
I, · · · ,
I, P2 − P1, δ (A11) + δ (B12)(cid:19)
1
2
I, P2 − P1, δ (C21)(cid:19) .
1
2
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
11
This implies that
0 = pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, M(cid:19) = (P2 − P1) ⋄ M.
1
2
According to Lemma 2.4, we know that M11 = 0. Thus we arrive at
δ(A11 + B12 + C21) = δ(A11) + δ(B12) + δ(C21).
Considering the relations
and
δ(cid:18)pn(cid:18) 1
2
I, · · · ,
I, P1, B12 + C21 + D22(cid:19)(cid:19)
1
2
δ(cid:18)pn(cid:18) 1
2
I, · · · ,
I, P2 − P1, B12 + C21 + D22(cid:19)(cid:19) ,
1
2
together with the previous calculations, we assert that
δ(B12 + C21 + D22) = δ(B12) + δ(C21) + δ(D22).
Claim 7. For any A11 ∈ A11, B12 ∈ A12, C21 ∈ A21 and D22 ∈ A22, we have
δ(A11 + B12 + C21 + D22) = δ(A11) + δ(B12) + δ(C21) + δ(D22).
We only need to prove that
M = δ(A11 + B12 + C21 + D22) − δ(A11) − δ(B12) − δ(C21) − δ(D22) = 0.
Note the facts that
pn(cid:18) 1
2
I, · · · ,
I, P1, D22(cid:19) = 0
1
2
and
pn(cid:18) 1
2
I, · · · ,
I, P1, A11 + B12 + C21 + D22(cid:19) = pn(cid:18) 1
2
1
2
I, · · · ,
I, P1, A11 + B12 + C21(cid:19) .
1
2
Applying Claim 6 (a) yields that
I, · · · ,
1
2
I, · · · ,
I, · · · ,
I, · · · ,
I, δ (P1) , A11 + B12 + C21 + D22(cid:19)
I, P1, δ (A11 + B12 + C21 + D22)(cid:19)
I, P1, A11 + B12 + C21 + D22(cid:19)(cid:19)
I, P1, A11 + B12 + C21(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
1
2
1
2
1
2
2
I, · · · ,
1
2
I, P1, D22(cid:19)(cid:19)
I, δ (P1) , D22(cid:19)
I, P1, δ (D22)(cid:19)
1
2
2
2
2
2
pn(cid:18) 1
+ pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
1
2
1
2
2
1
2
1
2
I, · · · ,
I, δ (P1) , A11 + B12 + C21(cid:19) + pn(cid:18) 1
I, P1, δ (A11 + B12 + C21)(cid:19) + pn(cid:18) 1
I, δ (P1) , A11 + B12 + C21 + D22(cid:19)
I, P1, δ (A11) + δ (B12) + δ (C21) + (D22)(cid:19) .
I, · · · ,
1
2
2
I, · · · ,
12
WENHUI LIN
Thus we obtain
0 = pn(cid:18) 1
2
I, · · · ,
I, P1, M(cid:19) = P1 ⋄ M.
1
2
So M12 = M21 = M11 = 0 by Lemma 2.4.
Similarly, using the relations
pn(cid:18) 1
2
I, · · · ,
I, P2, A11(cid:19) = 0
1
2
and
pn(cid:18) 1
2
I, · · · ,
I, P2, A11 + B12 + C21 + D22(cid:19) = pn(cid:18) 1
2
1
2
I, · · · ,
I, P2, B12 + C21 + D22(cid:19) ,
1
2
one can get M22 = 0. The proof of this claim is completed.
Claim 8. For any Aij , Bij ∈ Aij (i, j = 1, 2), we have
δ(Aij + Bij ) = δ(Aij ) + δ(Bij ).
Case 1: i 6= j.
Note that
pn(cid:18) 1
2
I, · · · ,
I, Pi + Aij , Pj + Bij(cid:19) = (Pi + Aij ) ⋄ (Pj + Bij)
1
2
= Aij + Bij + A∗
ij + Bij A∗
ij.
In light of Claim 6, we know that
I, · · · ,
2
δ(cid:18)pn(cid:18) 1
=δ(Aij + Bij) + δ(cid:0)A∗
I, Pi + Aij, Pj + Bij(cid:19)(cid:19)
ij(cid:1) .
ij(cid:1) + δ(cid:0)BijA∗
1
2
(3.2)
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
13
On the other hand, we by Claim 3 and Claim 4 have
I, · · · ,
I, Pi + Aij, δ (Pj + Bij )(cid:19)
1
2
I, · · · ,
I, Pi + Aij , Pj + Bij(cid:19)(cid:19)
1
2
1
2
1
2
I, · · · ,
I, · · · ,
2
2
2
δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
+ pn(cid:18) 1
+ pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
+ δ(cid:18)pn(cid:18) 1
2
2
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, · · · ,
2
2
2
1
2
I, · · · ,
I, · · · ,
I, δ (Pi + Aij) , Pj + Bij(cid:19) + pn(cid:18) 1
I, δ (Pi) , Pj(cid:19) + pn(cid:18) 1
I, δ (Pi) , Bij(cid:19) + pn(cid:18) 1
I, Pi, δ (Pj)(cid:19) + pn(cid:18) 1
I, Aij, δ (Pj )(cid:19) + pn(cid:18) 1
I, Pi, Pj(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
I, Aij, Pj(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
1
2
1
2
1
2
1
2
I, · · · ,
I, · · · ,
I, · · · ,
1
2
2
2
2
2
I, δ (Aij ) , Pj(cid:19)
1
2
I, δ (Aij ) , Bij(cid:19)
I, Pi, δ (Bij)(cid:19)
I, Aij, δ (Bij )(cid:19)
I, Pi, Bij(cid:19)(cid:19)
1
2
1
2
1
2
I, · · · ,
I, · · · ,
=δ (Pi ⋄ Bij) + δ (Aij ⋄ Pj) + δ (Aij ⋄ Bij )
=δ(Bij ) + δ(cid:0)Aij + A∗
=δ (Bij ) + δ (Aij) + δ(cid:0)A∗
ij(cid:1)
ij(cid:1) + δ(cid:0)BijA∗
ij(cid:1) + δ(cid:0)Bij A∗
ij(cid:1) .
Compare (3.2) with (3.3) gives
I, Aij , Bij(cid:19)(cid:19)
1
2
(3.3)
Case 2: i = j.
δ (Bij + Aij ) = δ (Bij) + δ (Aij ) .
Let us set M = δ(Aii + Bii) − δ(Aii) − δ(Bii). Let us take l = 1, 2, but l 6= i.
Since
and
pn(cid:18) 1
2
I, · · · ,
1
2
we know that
I, Pl, Aii(cid:19) = 0
2
1
2
I, · · · ,
pn(cid:18) 1
I, Pl, Bii(cid:19) = pn(cid:18) 1
2
I, · · · ,
I, Pl, Aii + Bii(cid:19) = 0,
1
2
I, · · · ,
I, δ (Pl) , Aii + Bii(cid:19) + pn(cid:18) 1
2
1
2
I, · · · ,
I, Pl, δ (Aii + Bii)(cid:19)
1
2
2
2
pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
2
2
2
I, · · · ,
I, · · · ,
I, Pl, Aii + Bii(cid:19)(cid:19)
I, Pl, Aii(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
2
1
2
1
2
I, · · · ,
1
2
I, δ (Pl) , Aii(cid:19) + pn(cid:18) 1
I, Pl, δ (Aii)(cid:19) + pn(cid:18) 1
1
2
2
2
I, · · · ,
I, · · · ,
I, · · · ,
1
2
I, Pl, Bii(cid:19)(cid:19)
I, δ (Pl) , Bii(cid:19)
I, Pl, δ (Bii)(cid:19) .
1
2
1
2
I, · · · ,
14
WENHUI LIN
Then we have
pn(cid:18) 1
2
I, · · · ,
I, Pl, M(cid:19) = Pl ⋄ M = 0.
1
2
By invoking of Lemma 2.4, we arrive at Mli = Mil = Mll = 0.
The last step is to show that Mii = 0. Since
2
pn(cid:18) 1
pn(cid:18) 1
2
I, · · · ,
I, · · · ,
1
2
1
2
I, Pl, Cli, Aii(cid:19) =CliAii + AiiC ∗
I, Pl, Cli, Bii(cid:19) =CliBii + BiiC ∗
li,
li,
and by Case 1 of this claim and Claim 4, we have
I, · · · ,
2
pn(cid:18) 1
+ pn(cid:18) 1
=δ(cid:18)pn(cid:18) 1
2
2
1
2
I, δ(Pl), Cli, Aii + Bii(cid:19) + pn(cid:18) 1
I, Pl, Cli, δ(Aii + Bii)(cid:19)
I, Pl, Cli, Aii + Bii(cid:19)(cid:19)
1
2
1
2
2
I, · · · ,
I, · · · ,
I, · · · ,
I, Pl, δ(Cli), Aii + Bii(cid:19)
1
2
=δ (p3 (Pl, Cli, Aii + Bii))
=δ (Cli(Aii + Bii) + (Aii + Bii)C ∗
li)
=δ(CliAii + AiiC ∗
li) + δ(CliBii + BiiC ∗
li)
1
2
I, Pl, Cli, Aii(cid:19)(cid:19) + δ(cid:18)pn(cid:18) 1
2
I, · · · ,
2
2
=δ(cid:18)pn(cid:18) 1
=pn(cid:18) 1
+ pn(cid:18) 1
+ pn(cid:18) 1
2
2
I, · · · ,
I, Pl, Cli, Bii(cid:19)(cid:19)
1
2
I, · · · ,
I, · · · ,
1
2
2
I, δ(Pl), Cli, Aii(cid:19) + pn(cid:18) 1
I, Pl, δ(Cli), Aii(cid:19) + pn(cid:18) 1
I, Pl, Cli, δ(Aii)(cid:19) + pn(cid:18) 1
1
2
1
2
2
2
I, · · · ,
I, · · · ,
1
2
I, δ(Pl), Cli, Bii(cid:19)
I, Pl, δ(Cli), Bii(cid:19)
I, Pl, Cli, δ(Bii)(cid:19)
1
2
1
2
I, · · · ,
I, · · · ,
Thus we obtain
I, Pl, Cli, M(cid:19)
2
I, · · · ,
0 =pn(cid:18) 1
1
2
=p3 (Pl, Cli, M )
=Cli ⋄ M.
It follows that CliMii + MiiC ∗
that I − P = I. In light of Lemma 2.2, we conclude that Mii = 0.
li = 0. That is, MiiC ∗
li = 0 for all Cli ∈ Ali. Note
As an immediate consequence of the previous Claims, we have
Claim 9. δ is an additive mapping.
Let us next show that δ is a ∗-derivation.
Claim 10. For any A ∈ A, we have δ(A∗) = δ(A)∗.
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
15
In view of Lemma 2.3, we konw that
pn(cid:18)A,
1
2
I, · · · ,
I(cid:19) =
1
2
1
2
(A + A∗).
It follows that
1
2
(δ(A) + δ(A∗)) =δ(cid:18)pn(cid:18)A,
=pn(cid:18)δ (A) ,
1
2
1
2
I, · · · ,
I, · · · ,
I(cid:19)(cid:19)
I(cid:19)
1
2
1
2
=
1
2
(δ(A) + δ(A)∗).
Thus we get δ(A∗) = δ(A)∗.
We next prove that δ is actually a derivation.
Claim 11. For any A, B ∈ A, we have δ(AB) = δ(A)B + Aδ(B).
Since
we obtain
pn(cid:18) 1
2
I, · · · ,
I, A, B(cid:19) = A ⋄ B = AB + BA∗,
1
2
δ(AB + BA∗) =δ(cid:18)pn(cid:18) 1
2
I, · · · ,
I, A, B(cid:19)(cid:19)
1
2
=pn(cid:18) 1
2
I, · · · ,
I, δ (A) , B(cid:19) + pn(cid:18) 1
2
1
2
I, · · · ,
I, A, δ (B)(cid:19)
1
2
=δ(A) ⋄ B + A ⋄ δ(B)
=δ(A)B + Bδ(A)∗ + Aδ(B) + δ(B)A∗.
It follows that
δ(AB) + δ(BA∗) = δ(A)B + Bδ(A)∗ + Aδ(B) + δ(B)A∗.
(3.4)
Replacing A (resp. B) by iA (resp. iB) in (3.4) and using Claim 9, we arrive at
δ(AB) − δ(BA∗) = δ(A)B − Bδ(A)∗ + Aδ(B) − δ(B)A∗.
(3.5)
Combining (3.4) with (3.5) gives
δ(AB) = δ(A)B + Aδ(B).
(cid:3)
By an analogous manner, we can prove
Theorem 3.2. Let B(H) be the algebra of all bounded linear operators on a complex
Hilbert space H and A ⊆ B(H) be a factor von Neumann algebra. Then a mapping
δ : A −→ A satisfies the rule
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A if and only if δ is an additive ∗-derivation.
16
WENHUI LIN
B(H) denotes the algebra of all bounded linear operators on a complex Hilbert
space H. Let us denote the subalgebra of all bounded finite rank operators by
F (H) ⊆ B(H). We call a subalgebra A of B(H) a standard operator algebra if
it contains F (H). It should be remarked that a standard operator algebra is not
necessarily closed in the sense of weak operator topology. This is quite different
from von Neumann algebras which are always weakly closed.
From ring theoretic prespective, standard operator algebras and factor von Neu-
mann algebras are both prime, whereas von Neumann algebras are usually semiprime.
Recall that an algebra A is prime if AAB = {0} impliess either A = 0 or B = 0.
An algebra is semiprime if AAA = {0} impliess A = 0. Every standard operator
algebra has the center CI, which is also the center of arbitrary factor von Neumann
algebra. An operator P ∈ B(H) is said to be a projection provided P ∗ = P and
P 2 = P . Any operator A ∈ B(H) can be expressed as A = RA + iIA, where i is
the imaginary unit, RA = A+A∗
. Note that both RA and IA are
self-adjoint.
and IA = A−A∗
2
2i
Combining our current methods with the techniques of [12], one can get
Theorem 3.3. Let H be an infinite dimensional complex Hilbert space and A be a
standard operator algebra on H containing the identity operator I. Suppose that A
is closed under the adjoint operation. Then a mapping δ : A −→ B(H) satisfies the
rule
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A if and only if δ is an additive ∗-derivation.
We must point out that the technical routes and proving methods of Theorems
3.2 and 3.3 are fairly similar to those of Theorem 3.1, and hence its proofs are
omitted here for saving space.
4. Related Topics for Future Research
The main purpose of this article is to concentrate on studying nonlinear ∗-Jordan-
type derivations on operator algebras. The involved operator algebras are based on
the algebra B(H) of all bounded linear operators on a complex Hilbert space H,
such as standard operator algebras, factor von Neumann algebras, von Neumann
algebras without central summands of type I1. Note that, unlike von Neumann
algebras which are always weakly closed, a standard operator algebra is not neces-
sarily closed. The current work together with [7, 10 -- 13, 23 -- 26] indicates that it is
feasible to investigate ∗-Jordan-type derivations and ∗-Lie-type derivations on op-
erator algebras under a unified framework -- η-∗-Jordan-type derivations. We have
good reasons to believe that characterizing η-∗-Jordan-type derivations on operator
algebras is also of great interest. In the light of the motivation and contents of this
article, we would like to end this article by proposing several open questions.
Let A be an associative ∗-algebra over the complex field C and η be a non-
zero scalar. For any A, B ∈ A, we can denote a "new product" of A and B by
A⋄η B = AB +ηBA∗. This new product ⋄η is usually said to be η-∗-Jordan product.
Clearly, 1-∗-Jordan product ⋄1 is the so-called ∗-Jordan product, and (−1)-∗-Jordan
product ⋄−1 is the so-called ∗-Lie product. Therefore, it is reasonable to say that η-
∗-Jordan products organically unify ∗-Jordan products with ∗-Lie products. There
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
17
are considerable works which are devoted to the study of mappings preserving the η-
∗-Jordan product between ∗-algebras, see [3, 4, 6, 8, 9, 19, 20, 27] and the references
therein.
Let δ : A −→ A be a mapping (without the additivity assumption). We say that
δ is a nonlinear η-∗-Jordan derivation if
δ(A ⋄η B) = δ(A) ⋄η B + A ⋄η δ(B),
holds true for all A, B ∈ A. Similarly, a mapping δ : A −→ A is called a nonlinear
η-∗-Jordan triple derivation if it satisfies the condition
δ(A ⋄η B ⋄η C) = δ(A) ⋄η B ⋄η C + A ⋄η δ(B) ⋄η C + A ⋄η B ⋄η δ(C)
for all A, B, C ∈ A, where A ⋄η B ⋄η C = (A ⋄η B) ⋄η C . We should note that ⋄ is
not necessarily associative.
Taking into account the definitions of η-∗-Jordan derivations and η-∗-Jordan
triple derivations, one can propose one much more common notion. Suppose that
n ≥ 2 is a fixed positive integer. Let us see a sequence of polynomials with scalar
η and ∗
p1(x1) = x1,
p2(x1, x2) = x1 ⋄η x2 = x1x2 + ηx2x∗
1,
p3(x1, x2, x3) = p2(x1, x2) ⋄η x3 = (x1 ⋄η x2) ⋄η x3,
p4(x1, x2, x3, x4) = p3(x1, x2, x3) ⋄η x4 = ((x1 ⋄η x2) ⋄η x3) ⋄η x4,
· · · · · · ,
pn(x1, x2, · · · , xn) = pn−1(x1, x2, · · · , xn−1) ⋄η xn
= (· · · ((
x1 ⋄η x2) ⋄η x3) ⋄η · · · ⋄η xn−1) ⋄η xn.
Accordingly, a nonlinear η-∗-Jordan n-derivation is a mapping δ : A −→ A satis-
fying the condition
n−2
{z }
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A. This notion is motivated by the definition of ∗-Jordan-
type derivations and that of ∗-Lie-type derivations. Then each ∗-Jordan derivation
is a 1-∗-Jordan 2-derivation and every ∗-Jordan triple derivation is a 1-∗-Jordan 3-
derivation. Likewise, each ∗-Lie derivation is a (−1)-∗-Jordan 2-derivation and ev-
ery ∗-Lie triple derivation is a (−1)-∗-Jordan 3-derivation. η-∗-Jordan 2-derivations,
η-∗-Jordan 3-derivations and η-∗-Jordan n-derivations are collectively referred to
as η-∗-Jordan-type derivations. η-∗-Jordan-type derivations on operator algebras
are intensively studied by several authors, [7, 10 -- 13, 23 -- 26] . A basic question in
this line is to investigate whether each nonlinear η-∗-Jordan-type derivation on an
operator algebra A with ∗ is an additive ∗-derivation. In view of the current work
and existing results in this direction, we propose several open questions.
Question 4.1. Let H be an infinite dimensional complex Hilbert space and A be
a standard operator algebra on H containing the identity operator I. Let η be a
non-zero scalar. Suppose that A is closed under the adjoint operation. A mapping
18
WENHUI LIN
δ : A −→ B(H) satisfies the following condition:
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A.
δ(ηA) = ηδ(A) hold for any A ∈ A ?
Is δ an additive ∗-derivation ? Does the relation
Question 4.2. Let B(H) be the algebra of all bounded linear operators on a complex
Hilbert space H and A ⊆ B(H) be a factor von Neumann algebra. Suppose that η
is a non-zero scaler. Let δ : A −→ B(H) be a mapping such that
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A. Is δ an additive ∗-derivation ? Do we have the relation
δ(ηA) = ηδ(A) for any A ∈ A ?
Question 4.3. Let B(H) be the algebra of all bounded linear operators on a complex
Hilbert space H and A ⊆ B(H) be a von Neumann algebra without central summands
of type I1. Let η be a non-zereo scalar. A mapping δ : A −→ B(H) satisfies the
following conditions:
δ(pn(A1, A2, · · · , An)) =
n
Xk=1
pn(A1, · · · , Ak−1, δ(Ak), Ak+1, · · · , An)
for all A1, A2, · · · , An ∈ A. Is δ an additive ∗-derivation ? Can we get the relation
δ(ηA) = ηδ(A) for any A ∈ A ?
References
[1] Z.-F. Bai and S.-P. Du, Maps preserving products XY −Y X ∗ on von Neumann
algebras, J. Math. Anal. Appl., 386 (2012), 103-109. 2
[2] M. Bresar and M. Fosner, On rings with involution equipped with some new
product, Publ. Math. Debrecen, 57 (2000), 121-134. 1, 2
[3] J.-L. Cui and C.-K. Li, Maps preserving product XY − Y X ∗ on factor von
Neumann algebras, Linear Algebra Appl., 431 (2009), 833-842. 2, 17
[4] L.-Q. Dai and F.-Y. Lu, Nonlinear maps preserving Jordan ∗-products, J. Math.
Anal. Appl., 409 (2014), 180-188. 1, 2, 17
[5] A. Fosner, F. Wei and Z.-K. Xiao, Nonlinear Lie-type derivations of von Neu-
mann algebras and related topics, Colloq. Math., 132 (2013), 53-71. 2
[6] D.-H. Huo, B.-D. Zheng and H.-Y. Liu, Nonlinear maps preserving Jordan
triple η-∗-products, J. Math. Anal. Appl., 430 (2015), 830-844. 1, 2, 17
[7] W. Jing, Nonlinear ∗-Lie derivations of standard operator algebras, Quaest.
Math., 39 (2016), 1037-1046. 16, 17
[8] C-.J. Li and F.-Y. Lu, Nonlinear maps preserving the Jordan triple 1-∗-product
on von Neumann algebras, Complex Anal. Oper. Theory, 11 (2017), 109-117.
1, 2, 17
[9] C.-J. Li, F.-Y. Lu and X.-C. Fang, Nonlinear mappings preserving product
XY +Y X ∗ on factor von Neumann algebras, Linear Algebra Appl., 438 (2013),
2339-2345. 1, 2, 17
NONLINEAR ∗-JORDAN-TYPE DERIVATIONS ON VON NEUMANN ALGEBRAS
19
[10] C.-J. Li, F.-Y. Lu and X.-C. Fang, Non-linear ξ-Jordan ∗-derivations on von
Neumann algebras, Linear Multilinear Algebra, 62 (2014), 466-473. 2, 3, 16,
17
[11] C.-J. Li, F.-F. Zhao and Q.-Y. Chen, Nonlinear skew Lie triple derivations
between factors, Acta Math. Sinica (English Series), 32 (2016), 821-830. 16,
17
[12] W.-H. Lin, Nonlinear ∗-Lie-type derivations on standard operator algebras,
Acta Math. Hungar., 154 (2018), 480-500. 2, 16, 17
[13] W.-H. Lin, Nonlinear ∗-Lie-type derivations on von Neumann algebras, Acta
Math. Hungar., https://doi.org/10.1007/s10474-018-0803-1. 2, 16, 17
[14] W. S. Martindale III, When are multiplicative mappings additive ? , Proc.
Amer. Math. Soc., 21 (1969), 695-698. 2
[15] C. R. Miers, Lie homomorphisms of operator algebras, Pacific J. Math., 38
(1971), 717-735. 3
[16] L. Moln´ar, A condition for a subspace of B(H) to be an ideal, Linear Algebra
Appl., 235 (1996), 229-234. 1
[17] L. Moln´ar, Jordan ∗-derivation pairs on a complex ∗-algebra, Aequationes
Math., 54 (1997), 44-55. 1
[18] L. Moln´ar, Jordan maps on standard operator algebras, in Z. Dar´oczy and
Zs. P´ales (Edt.), Functional Equations -- Results and Advances, pp. 305-320,
Kluwer Academic Publishers, 2001. 1
[19] L. Moln´ar, Non-linear Jordan triple automorphisms of sets of self-adjoint ma-
trices and operators, Studia Math., 173 (2006), 39-48. 2, 17
[20] L. Moln´ar, Multiplicative Jordan triple isomorphisms on the self-adjoint ele-
ments of von Neumann algebras, Linear Algebra Appl., 419 (2006), 586-600.
2, 17
[21] L. Moln´ar and P. Semrl Local Jordan ∗-derivations of standard operator alge-
bras, Proc. Amer. Math. Soc., 125 (1997), 447-454. 1
[22] P. Semrl, Jordan ∗-derivations of standard operator algebras, Proc. Amer.
Math. Soc., 120 (1994), 515-518. 1
[23] A. Taghavi, H. Rohi and V. Darvish, Non-linear ∗-Jordan derivations on von
Neumann algebras, Linear Multilinear Algebra, 64 (2016), 426-439. 3, 16, 17
[24] Y.-W. Yu and J.-H. Zhang, Nonlinear ∗-Lie derivations on factor von Neu-
mann algebras, Linear Algebra Appl., 437 (2012), 1979-1991. 16, 17
[25] F.-J. Zhang, Nonlinear skew Jordan derivable maps on factor von Neumann
algebras, Linear Multilinear Algebra, 64 (2016), 2090-2103. 3, 16, 17
[26] F.-F. Zhao and C.-J. Li, Nonlinear ∗-Jordan triple derivations on von Neu-
mann algebras, Math. Slovaca, 68 (2018), 163-170. 3, 16, 17
[27] F.-F. Zhao and C.-J. Li, Nonlinear maps preserving the Jordan triple ∗-product
between factors, Indag. Math., 29 (2018), 619-627. 1, 2, 17
Lin: College of Science, China Agricultural University, Beijing 100083, P. R. China
E-mail address: [email protected]
E-mail address: [email protected]
|
1808.01511 | 3 | 1808 | 2018-09-02T14:05:31 | Large irredundant sets in operator algebras | [
"math.OA",
"math.GN",
"math.LO"
] | A subset $\mathcal X$ of a C*-algebra $\mathcal A$ is called irredundant if no $A\in \mathcal X$ belongs to the C*-subalgebra of $\mathcal A$ generated by $\mathcal X\setminus \{A\}$. Separable C*-algebras cannot have uncountable irredundant sets and all members of many classes of nonseparable C*-algebras, e.g., infinite dimensional von Neumann algebras have irredundant sets of cardinality continuum.
There exists a considerable literature showing that the question whether every AF commutative nonseparable C*-algebra has an uncountable irredundant set is sensitive to additional set-theoretic axioms and we investigate here the noncommutative case.
Assuming $\diamondsuit$ (an additional axiom stronger than the continuum hypothesis) we prove that there is an AF C*-subalgebra of $\mathcal B(\ell_2)$ of density $2^\omega=\omega_1$ with no nonseparable commutative C*-subalgebra and with no uncountable irredundant set. On the other hand we also prove that it is consistent that every discrete collection of operators in $\mathcal B(\ell_2)$ of cardinality continuum contains an irredundant subcollection of cardinality continuum.
Other partial results and more open problems are presented. | math.OA | math |
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Abstract. A subset X of a C*-algebra A is called irredundant if no A ∈ X
belongs to the C*-subalgebra of A generated by X \ {A}. Separable C*-
algebras cannot have uncountable irredundant sets and all members of many
classes of nonseparable C*-algebras, e.g., infinite dimensional von Neumann
algebras have irredundant sets of cardinality continuum.
There exists a considerable literature showing that the question whether
every AF commutative nonseparable C*-algebra has an uncountable irredun-
dant set is sensitive to additional set-theoretic axioms and we investigate here
the noncommutative case.
Assuming ♦ (an additional axiom stronger than the continuum hypothesis)
we prove that there is an AF C*-subalgebra of B(ℓ2) of density 2ω = ω1
with no nonseparable commutative C*-subalgebra and with no uncountable
irredundant set. On the other hand we also prove that it is consistent that
every discrete collection of operators in B(ℓ2) of cardinality continuum contains
an irredundant subcollection of cardinality continuum.
Other partial results and more open problems are presented.
1. Introduction
Definition 1.1. Let A be a C*-algebra. A subset X ⊆ A is called irredundant if
and only if for every A ∈ X , the C*-subalgebra of A generated by X \ {A} does not
contain A. We define
irr(A) := sup{X : X is an irredundant set in A}.
Recall that the density of a C*-algebra A, denoted d(A) is the least cardinality
of a norm dense subset of A, i.e., A is separable if and only if d(A) is countable.
It is easy to see that irr(A) ≤ d(A) for every C*-algebra, as irredundant sets
must be norm discrete. When A is an infinite dimensional C*-algebra, then irr(A)
is infinite, because then A contains an infinite dimensional abelian C*-subalgebra
([41]) and locally compact infinite Hausdorff spaces contain pairwise disjoint infinite
collections of open sets which yield infinite irredundant sets (Proposition 3.12). In
this article, we are interested in uncountable irredundant sets in (C*-subalgebras
of) the algebra B(ℓ2) of all linear bounded operators on a separable Hilbert space.
Irredundant sets have been considered in the context of other structures. For
example, a subset of a Boolean algebra is called irredundant if none of its ele-
ments belongs to the Boolean subalgebra generated by the remaining elements.
We call such sets Boolean irredundant (Definition 3.8). In Banach spaces irredun-
dant sets, i.e., where no element belongs to the closed subspace spanned by the
2010 Mathematics Subject Classification. 03E35, 46L05, 46L85, 46J10, 54G12.
The research of the first named author was partially supported by doctoral scholarships
CAPES: 1427540 and CNPq: 167761/2017-0 and 201213/2016-8.
The research of the second named author was partially supported by grant PVE Ciencia sem
Fronteiras - CNPq (406239/2013-4).
1
2
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
remaining elements correspond exactly to biorthogonal systems ([20], see [30] for
some comparisons between this type of notions). Examples of Boolean irredundant
sets include independent families, ideal independent families or (almost) disjoint
families, but there are Boolean algebras of uncountable irredundance with no un-
countable families of the above-mentioned classes (see Remark 3.7). A collection
(xα, x∗α)α<κ ⊆ B × B∗ of a Banach space B is biorthogonal if x∗α(xα) = 1 and
x∗α(xβ ) = 0 for all α < β < κ. As usually linear functionals on a C*-algebra A
are not multiplicative there are many more biorthogonal systems than irredundant
sets in A, one can even consistently have a commutative C*-algebra C(K) with
countable irredundance but with uncountable biorthogonal systems ([9]).
One of our main motivations are consistent constructions of uncountable Boolean
algebras with no uncountable irredundant sets. They were first obtained by Rubin
([46]) under the assumption of ♦1 and then by Kunen ([40]) under the continuum
hypothesis CH (improved further by Todorcevic to a b = ω1 construction from 2.4
of [51]). Also some versions of the classical Ostaszewski's construction assuming ♦
from [43] have these properties as further constructions assuming ♣ from [20] as
well as forcing constructions from [6], [9], [29].
Some of the above constructions are of Boolean algebras and other of (locally)
compact Hausdorff totally disconnected spaces. Using the Stone duality one trans-
lates one language to the other easily. The fact that the Kunen or Ostaszewski
types of constructions mentioned above correspond to superatomic Boolean alge-
bras or equivalently their Stone spaces are scattered spaces (every subset has a
relative isolated point) yields the equality between the Boolean irredundance of the
Boolean algebra and the irredundance of the commutative C*-algebra of contin-
uous functions (Corollary 3.10). In particular the corresponding C(K)s have no
uncountable irredundant sets. In fact the scatteredness can be exploited further to
prove that the Banach spaces C(K) have no ucountable biorthogonal systems ([40],
[20]).
The first question we considered was whether such phenomena can take place if
the C*-algebra is made considerably noncommutative. One of our main results is:
Theorem 1.2. Assume ♦. There is a fully noncommutative nonseparable scattered
C*-algebra (of operators in B(ℓ2)) with no nonseparable commutative subalgebra and
with no uncountable irredundant set.
Proof. Apply Theorems 2.12 and 6.2.
(cid:3)
Here scattered C*-algebras are the noncommutative analogues of the scattered
locally compact spaces. The condition of being fully noncommutative means that
these algebras are "maximally noncommutative" among scattered algebras. These
notions are reviewed in Section 2.1.
1♦ is an additional axiom (introduced by R. Jensen) which is true in the universe of con-
structible sets. It says that there is a sequence (Sα)α<ω1 which "predicts" all subsets of ω1 in
the sense that for any X ⊆ ω1 the set {α < ω1 : X ∩ α = Sα} meets every closed an unbounded
subset of ω1, for details see [25] or [33]. ♦ has been recently successfully applied in the context of
nonseparable C*-algebras by Akemann, Farah, Hirshberg and Weaver ([3], [4], [17]). We will not
use the ♦ directly but will apply its consequence from Theorem 2.12 which was developed by S.
Todorcevic in [54].
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
3
Another motivation for our project was the result of Todorcevic ([52],
[53])
that assuming Martin's axiom MA and the negation of the CH every uncount-
able Boolean algebra has an uncountable irredundant set. Here the main question
remains open:
Question 1.3. Is it consistent that every nonseparable (AF, scattered) C*-algebra
(of operators in B(ℓ2)) contains an uncountable irredundant set?
It should be added that even the commutative general case is open, since the
result of Todorcevic provides uncountable irredundant sets in C(K)s only for Ks
totally disconnected and there can be nonmetrizable compact spaces with no totally
disconnected nonmetrizable compact subspace and similar examples (see [31]). So
it is natural to restrict initially the attention in the noncommutative problem to
C*-algebras corresponding to totally disconnected spaces, namely to approximately
finite dimensional C*-algebras (AF), i.e., where there is a dense subset which is the
union of a directed family of finite dimensional C*-subalgebras (see [16] for diverse
notions of approximate finite-dimensionality in the nonseparable context). Another
natural narrowing of the question is to consider only the scattered C*-algebras since
one of the conditions equivalent to being scattered for a C*-algebra of density ω1
is that each of its C*-subalgebras is AF. Attempting to answer Question 1.3 we
obtained several results which shed some light on it. Let us discuss them below.
If A is AF C*-algebra of density equal to the first uncountable cardinal ω1, then
it can be written as A = Sα<ω1 Aξ where Aξ ⊆ Aξ′ for all ξ < ξ′ < ω1 and
each Aξ is separable and AF. It follows from the result of Thiel in [48] (cf.
[42],
[49]) that each Aξ is singly generated by one element Aξ ∈ Aξ. Hence in the set
{Aξ : ξ < ω1} irredundant subsets are at most singletons. So there is no chance to
extract (possibly using some additional forcing axioms) an uncountable irredundant
set from an arbitrary norm discrete set of cardinality ω1 of operators in B(ℓ2).
The AF hypothesis allows nevertheless to avoid sets of operators as above.
whenever D ≤ D′ for D ∈ D and (D, ≤) is directed, then given any norm discrete
Namely, if A = SD∈D AD, where all ADs are finite-dimensional and AD ⊆ AD′
{Aξ : ξ < ω1} ⊆ SD∈D AD, which exists by the nonseparability of A, for every
finite F ⊆ ω1 the set
XF = {ξ < ω1 : Aξ ∈ AF }
is a finite superset of F , where AF is the C*-subalgebra generated by {Aη : η ∈ F }.
So, the search for an uncountable irredundant set among {Aξ : ξ < ω1} is equivalent
to the search for an uncountable X ⊆ ω1 such that XF ∩ X = F for every F ⊆ X.
However this combinatorial problem for a general function from finite subsets
of ω1 to themselves has the negative solution2. Nevertheless passing to the second
uncountable cardinal ω2 allows for a very general consistency result:
Theorem 1.4. It is consistent that 2ω = ω2 and for every norm discrete collection
of operators (Aξ : ξ < ω2) in B(ℓ2) there is a subset X ⊆ ω2 of cardinality ω2 such
that (Aξ : ξ ∈ X) is irredundant.
2It is enough to take XF to be of the form Y ∩ [(max F ) + 1] where Y ∈ µ is of minimal rank
which contains F and where µ is an (ω, ω1)-cardinal as in [32]. µ is originally due to Velleman
([55]). A positive result for general functions is that given n ∈ N and a function φ from finite
subsets of the n-th uncountable cardinal ωn into countable subsets of ωn there is an n-element set
X ⊆ ωn such that ξ 6∈ φ(X \ {ξ}) for any ξ ∈ X. In particular, this gives that any norm discrete
subset of cardinality ωn in any C*-algebra has an irredundant subset of cardinality n.
4
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
This is not a mere consequence of B(ℓ2) having density ω2 because by a result
of Brech and Koszmider ([8]) it is consistent that there exists a commutative C*-
subalgebra of ℓ∞ of density 2ω = ω2 with no uncountable irredundant set. The
cardinal ω2 in Theorem 1.4 can be replaced by any regular cardinal bigger than ω1
but it is not known if the result of [8] can be generalized to bigger cardinals than
ω2. Combining 1.4, 1.2 and knowing that ♦ implies CH we obtain:
Corollary 1.5. It is independent from ZFC whether there is a norm discrete col-
lection of operators (projections) (Aξ : ξ < 2ω) in B(ℓ2) with no uncountable (of
cardinality 2ω) irredundant subcollection of size 2ω.
It is independent from ZFC whether there is C*-subalgebra of B(ℓ2) of density
2ω with no uncountable (of size 2ω) irredundant set.
The commutative results mentioned above, in fact, are most often of topological
nature, where the compact Hausdorff space under the consideration is the Stone
space KA of a Boolean algebra A. For example, the reason the above-mentioned
Boolean algebras have countable irredundance is that the spread3 of KA × KA is
countable as the finite powers of the mentioned KAs are hereditarily separable.
Namely, in general we have irr(A) ≤ s(KA × KA) which was first noted in [23] and
easily follows from the characterization of irredundant sets in the commutative case
(Lemma 3.4). Also the Urysohn Lemma gives the inequality s(K) ≤ irr(C(K)) for
any locally compact Hausdorff K. This argument cannot be transferred to the
noncommutative setting since we do not have so general noncommutative Urysohn
Lemma (for noncommutative Uryshon Lemma see [2]). That is for constructing
an irredundant set of cardinality κ in a C*-algebra A it is enough to construct a
sequence of states (τα : α < κ) and a sequence of positive elements (Aα : α < κ) of A
such that τα(Aα) > 0 for all α < κ and τα(Aβ) = 0 for all distinct α, β < κ (Lemma
3.14), but a weak∗ discrete set of pure states does not produce the elements Aα as
above due to the lack of the Urysohn Lemma for nonorthogonal closed projections.
In fact assuming the Proper Forcing Axiom, PFA every nonseparable scattered C*-
algebra has an uncountable weak∗ discrete set of pure states (Corollary 3.17), but
this does not help us in constructing an uncountable irredundant set and answering
Question 1.3 in the positive in the scattered case.
A bolder approach to Question 1.3 would be to try to answer the following
question in the positive:
Question 1.6. Is it consistent (with MA and the negation of CH) that every non-
separable scattered (or even AF) C*-algebra has a nonseparable commutative sub-
algebra in one of its quotients?
Note that the class of scattered C*-algebras is closed under quotients and subal-
gebras and every locally compact scattered Hausdorff space is totally disconnected,
so the positive answer to the question above and the MA result of Todorcevic
mentioned above would give the positive answer to Question 1.3 in the scattered
case.
Known ZFC examples of nonseparable C*-algebras with no nonseparable com-
mutative subalgebras are the reduced group C*-algebra of an uncountable free
group as shown by Popa in [45] and the algebras of Akemann and Doner as shown
3The spread of a topological space K, denoted by s(K) is the supremum over the cardinalities
of discrete subspaces of K.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
5
in [7]. However the former is not AF (and has an uncountable irredundant set cor-
responding to the free generators of the group) and the latter has a nonseparable
commutative quotient c0(ω1) (which also has an obvious uncountable irredundant
set). Perhaps the algebra of [19] could provide the negative answer to Question 1.6.
The reason our algebra from Theorem 1.2 does not contain a nonseparable com-
mutative C*-subalgebra is that given any discrete sequence of projections in certain
dense subalgebra there are two of them which have maximal commutator equal to
1/2 (the fact that 1/2 is the maximal value is proved in [47]). However in such
an arbitrary sequence there are also two projections which almost commute (see
Theorem 6.2), so in this sense our algebra is quite random, that is no pattern re-
peats on any uncountable norm discrete subset of elements. In fact such behaviour
is already sensitive to infinitary combinatorics beyond ZFC determined by ♦ and
Open Coloring Axiom (OCA)4, namely we have:
Theorem 1.7. Assume OCA. For every 0 < ε < 1/2 among any sequence of
operators (Aξ : ξ < ω1) in B(ℓ2) there is an uncountable X ⊆ ω1 such that
• for every distinct ξ1, ξ2 ∈ X we have [Aξ1 , Aξ2 ] > 1/2 − ε, or
• for every ξ1, ξ2 ∈ X we have [Aξ1 , Aξ2 ] < ε.
However, assuming ♦ there is a scattered C*-algebra A ⊆ B(ℓ2) (it is in particular
AF) such that for every 0 < ε < 1/2 among any discrete sequence of projections
(Pξ : ξ < ω1) in A
• there are ξ1 < ξ2 < ω1 such that [Pξ1 , Pξ2 ] > 1/2 − ε,
• there are ξ1 < ξ2 < ω1 such that [Pξ1 , Pξ2 ] < ε.
Proof. Apply Corollary 4.6 and Theorem 6.2
(cid:3)
Another natural question related to uncountable irredundant sets in general C*-
algebras is the following:
Question 1.8.
(1) Is it true that d(A) ≤ 2irr(A) holds for every C*-algebra (every C*-algebra
of type I )?
(2) Can there be arbitrarily big C*-algebras with no uncountable irredundant
sets?
This is motivated by a Boolean result of McKenzie (see 4.2.3 of [28]) which says
that a Boolean algebra has a dense subalgebra not bigger than its irredundance.
This result has been generalized by Hida in [24] to all commutative C*-algebras
which implies that irr(A) ≤ 2d(A) holds for commutative A. We prove this in-
equality answering Question 1.8 for scattered C*-algebras in our Theorem 3.18.
In Section 2 we review scattered C*-algebras and constructions schemes which is
an elegant framework to deal with some constructions using ♦ recently introduced
by Todorcevic in [54]. It was already applied in several functional analytic, topo-
logical and combinatorial contexts in [54], [36], [37]. In Section 3 we prove basic
facts concerning irredundant sets in commutative and noncommutative setting. In
Section 4 we prove the OCA part of Theorem 1.7. Section 5 is devoted to defining
and investigating the partial order of finite dimensional approximations to our al-
gebra from Theorem 1.2. In the final Section 6 we use the appropriate construction
schemes described in Section 2 to construct the algebra from Theorem 1.2.
4For the statement of OCA see Definition 4.4.
6
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Notation and the terminology of this paper should be standard, however, it
draws from diverse parts of mathematics like Boolean algebras, operator theory,
set-theory, logic and general topology. When in doubt one could refer to textbooks
like [28], [39], [25], [33], [14].
In particular by an embedding (isomorphism onto
its image) we mean ∗-monomorphism (∗-isomorphism) of C∗-algebras which is not
necessarily unital, ℓ2(X) denotes the Hilbert space of square summable complex
functions defined on a set X, B(ℓ2(X)) denotes the C*-algebra of all bounded
operators on ℓ2(X), ℓ2 = ℓ2(N), h·, ·i denotes the scalar product, A+ denotes the
set of positive elements of a C*-algebra A, 1A denotes the unit of A and eA the
unitization of A, BA∗ denotes the dual ball of the algebra A, [A, B] = AB − BA
for A, B ∈ B(ℓ2), Mn denotes the C*-algebra of n × n matrices for n ∈ N, C(K)
denotes the C*-algebra of complex valued continuous functions on a compact K and
C0(X) the C*-algebra of complex valued continuous functions vanishing at infinity
on a locally compact X, χU denotes the characteristic function of a set U , Clop(K)
denotes the family of clopen subsets of a space K, ωn denotes the n-th uncountable
cardinal for n ∈ N, [X]n denotes the family of all n-element subsets of a set X,
[X]<ω denotes the family of all finite subsets of a set X; X < Y means that x < y
for all x ∈ X and y ∈ Y where X, Y are sets of ordinals.
We would like to thank Alessandro Vignati for his feedback on an earlier version
of this paper.
2. Preliminaries
2.1. Scattered C*-algebras. The reason why scattered C*-algebras will play an
important role in our investigation of irredundant sets is that in such algebras irre-
dundant sets can easily be replaced by irredundant sets of projections (Proposition
3.3), in particular the Boolean results pass to the C*-algebraic ones (Corollary
3.10). Moreover all commutative results culminate around the scattered case which
seems most basic.
Recall that a topological space is called scattered if it does not contain any perfect
subset or in other words if each (closed) nonempty subset has a relative isolated
point. The phenomena related to the scatteredness were already analysed by Cantor
which resulted in the notion of the Cantor-Bendixson derivative of a topological
space ([14]). The Boolean algebra manifestation of these phenomena was discovered
by Mostowski and Tarski in [38] as what is today known as superatomic Boolean
algebras. The importance of the class of Banach spaces of the form C(K), where
K is scattered, already implicitly known in the 30ties, was first systematically
revealed in [44].
Its generalization, Asplund Banach spaces, started to play an
important role in Banach space theory since the 60ties. It was Jensen in [26] who
first defined a scattered C∗-algebra but they were considered earlier by Tomiyama
[50] and Wojtaszczyk [56]. A recent survey [18] underlines the links of scattered
C*-algebras with its Boolean algebraic and commutative predecessors. Recall that
a projection p in a C*-algebra is called minimal if and only if pAp = Cp, i.e.,
minimal projections generalize isolated points. The ∗-subalgebra of A generated
by the minimal projections of A will be denoted I At(A). We have the following
observation from [18]:
Proposition 2.1. Suppose that A is a C∗-algebra.
(1) I At(A) is an ideal of A,
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
7
(2) I At(A) is isomorphic to a subalgebra of the algebra K(H) of all compact
operators on a Hilbert space H,
(3) I At(A) contains all ideals of A which are isomorphic to a subalgebra of
K(H) for some Hilbert space H,
(4) if an ideal I ⊆ A is essential and isomorphic to a subalgebra of K(H) for
some Hilbert space H, then I = I At(A).
A selected list of conditions equivalent to being scattered and which are relevant
to our paper is given below. Any of these conditions can be taken as the definition
of a scattered algebra.
Theorem 2.2 ([26, 27, 56, 35, 34, 50, 18]). Suppose that A is a C∗-algebra. The
following conditions are equivalent:
(1) Every non-zero ∗-homomorphic image of A has a minimal projection.
(2) There is an ordinal ht(A) and a continuous increasing sequence of closed
α (A))α≤ht(A) called the Cantor-Bendixson composition series for
ideals (I At
A such that I0 = {0}, Iht(A) = A and
I At(A/I At
α+1(A)},
α (A)) = {[a]I At
α (A) : a ∈ I At
for every α < ht(A).
(3) Every non-zero subalgebra of A has a minimal projection.
(4) Every non-zero subalgebra has a projection,
(5) Every subalgebra of A has real rank zero,
(6) A does not contain a copy of the C∗-algebra C0((0, 1]) = {f ∈ C((0, 1]) :
limx→0 f (x) = 0}.
(7) The spectrum of every self-adjoint element is countable.
Definition 2.3. [18] A scattered C*-algebra is called thin-tall if and only if ht(A)
from Theorem 2.2 (2) is equal ω1 and I At
α (A) is separable for each α < ω1.
α+1(A)/I At
In the nonseparable context we are especially interested in condition (2) which
was introduced in [18] which gives an essential composition series corresponding to
the Cantor-Bendixson derivative. A scattered C*-algebra is called fully noncom-
mutative if and only if for all α < ht(A) the algebra I At(A/Iα) is *-isomorphic to
the algebra of all compact operators on a Hilbert space. We have the following two
observations from [18]:
Proposition 2.4. Suppose that A is a scattered C∗-algebra. The following are
equivalent:
(1) A is fully noncommutative,
(2) the ideals of A form a chain,
(3) the centers of the multiplier algebras of any quotient of A are all trivial.
Proposition 2.5. Every scattered C∗-algebra A is atomic, i.e., the ideal I At(A)
is essential.
Recall that in a topological space a sequence of points {xξ : ξ < κ} is called
right-separated (left-separated) if and only if xξ 6∈ {xη : η > ξ} for all ξ < κ
(xξ 6∈ {xη : η < ξ} for all ξ < κ). Left and right separated sequences play an impor-
tant role in commutative set-theoretic topology because a regular space is hered-
itarily Lindelof (hereditarily separable) if it has no uncountable right-separated
8
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
(left-separated) sequences. Additional axioms like ♦, CH, MA, PFA5 have sub-
stantial impact on the existence of right or left separated sequences in regular
topological spaces, for example PFA implies that there are no regular S-spaces, i.e.,
every regular topological space which has an uncountable right-separated sequence
has an uncountable left-separated sequence as well (Theorem 8.9 of [51]).
Proposition 2.6. Suppose that A is a thin-tall C*-algebra. Then the dual ball BA∗
of A∗ contains an uncountable right-separated sequence of pure states in the weak∗
topology. In particular under the Proper Forcing Axiom (PFA) the dual ball BA∗
of A∗ contains an uncountable discrete set consisting of pure states.
Proof. Let (Iα)α<ω1 be the Cantor-Bendixson composition series of 2.2 (3). As
Iα+1/Iα is an essential ideal of A/Iα which is *-isomorphic with the the algebra
of all compact operators on ℓ2, we can embed A/Iα into B(ℓ2) with the range
containing all compact operators. Take τα to be a vector pure state on B(ℓ2)
composed with the quotient map and the embedding. So τα is a pure state on A
which is zero on Iα and there is Aα ∈ Iα+1 such that τα(Aα) = 1. Denote the set
of all pure states on A by P (A). Now consider
Uα = {τ ∈ P (A) : τ (Aα) > 0}.
Note that if τβ ∈ Uα then β ≤ α, so {τα : α < ω1} is right-separated in the weak∗
topology. So {τα : α < ω1} contains and uncountable left-separated sequence by
PFA (Theorem 8.9 or [51]). It is clear that a sequence which is both left and right
separated is discrete.
(cid:3)
2.2. Construction schemes. In this section we recall some definitions and results
from [54].
Definition 2.7. Let E, F ∈ [ω1]<ω.
(1) F < E whenever α < β for all α ∈ F and β ∈ E,
(2) F ⊑ E whenever there is α ∈ ω1 such that E ∩ α = F (we say that F is an
initial fragment of E or that E end-extends F ),
(3) F ❁ E whenever F ⊑ E and E \ F 6= ∅.
Definition 2.8. Let η be an ordinal and let (Fξ : ξ < η) = F ⊆ [ω1]<ω.
(1) F is cofinal if for all E ∈ [ω1]<ω there is F ∈ F such that E ⊆ F ,
(2) (Fξ : ξ < η) is a ∆-system of length η with root ∆ whenever Fξ ∩ Fξ′ = ∆
for all ξ < ξ′ < η,
(3) A ∆-system (Fξ : ξ < η) with root ∆ is increasing whenever Fξ \∆ < Fξ′ \∆
for all ξ < ξ′ < η,
(4) A subset of a ∆-system is called a subsystem,
(5) FF = {E ∈ F : E ( F } for F ⊆ ω1.
Definition 2.9. A pair of sequences (nk)k∈N ⊆ N and (rk)k∈N ⊆ N is called allowed
parameters if and only if
(1) r0 = r1 = n0 = 0
(2) nk ≥ 2 for all k ∈ N.
(3) each natural value appears in the sequence (rk)k∈N ⊆ N infinitely many
times
5For the statement of the Proper Forcing Axiom (PFA) or Martin's Axiom (MA) we refer the
reader, for example, to [25] or [51]. PFA implies among others MA, OCA and 2ω = ω2.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
9
(4) rk+1 < mk where m0 = 1, mk+1 = rk+1 + nk+1(mk − rk+1) for k > 0.
Definition 2.10. A construction scheme with a pair of allowed parameters (nk)k∈N ⊆
N and (rk)k∈N ⊆ N is a cofinal family F =Sn∈N Fn satisfying
(1) F0 = [ω1]1,
(2) If k > 0 and E, F ∈ Fk, then E = F and E ∩ F ⊑ E, F and
{φF,E[G] : G ∈ FE} = FF,
where φF,E : E → F is the order preserving bijection between E and F ,
(3) If k ≥ 0 and F ∈ Fk+1, then the maximal elements of FF are in Fk
and they form an increasing ∆-system of length nk+1 such that F is its
union. The family of all these maximal elements is called the canonical
decomposition of F .
Definition 2.11. Given a construction scheme F , we say that an F ∈ Fk for k > 0
captures a ∆-system (si : i < n) of finite subsets of ω1 with root s if the canonical
decomposition (Fi : i < nk) of F with root ∆ has the following properties:
(1) nk ≥ n, s ⊆ ∆, and si \ s ⊆ Fi \ ∆ for all i < n.
(2) φFi,Fj [si] = sj for all i < j < n.
When n = nk, we say that F fully captures the ∆-system.
Theorem 2.12 ([54]). Assume ♦. For any pair of allowed parameters (nk)k∈N
and (rk)k∈N there is a construction scheme F with these parameters and there is a
partition (Pn)n∈N of N into infinitely many infinite sets such that for every n ∈ N
and every uncountable ∆-system T of finite subsets of ω1 there exist arbitrarily large
k ∈ Pn and F ∈ Fk+1 which fully captures a subsystem of T .
3. Irredundant sets
3.1. Reducing irredundant sets to special ones. Because of Weierstrass-Stone
theorem for unital commutative C*-algebras sometimes it is useful to consider a
strengthening of being irredundant, where the subalgebras we generate are unital.
However this does not affect the cardinalities of irredundant sets much:
Lemma 3.1. Suppose that A is a unital C*-algebra and that X ⊆ A is its nonempty
irredundant set. Then there is x0 ∈ X such that no element x of X \ {x0} belongs
to the unital C*-subalgebra generated by X \ {x0, x}.
Proof. If no element x of X belongs to the unital C*-subalgebra generated by X \{x}
we are done by taking any element of X as x0.
Otherwise let x0 ∈ X belong to the unital C*-subalgebra generated by X \ {x0}
so x0 = λ1 + y where y is in the subalgebra generated by X \ {x0} and λ ∈ C \ {0}.
Suppose that there is x ∈ X \ {x0} in the unital C*-subalgebra generated by
X \ {x0, x}, i.e., x = λ′1 + z where z is in the subalgebra generated by X \ {x0, x}
and λ′ ∈ C \ {0}. So 1 is in the algebra generated by X \ {x0}, but this shows that
x0 = λ1 + y is in the subalgebra generated by X \ {x0}, a contradiction with the
fact that X is irredundant.
(cid:3)
Clearly any two orthogonal one-dimensional projections in M2 form an irredun-
dant set, however each of them is in the unital C*-algebra generated by the other
projection.
10
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Proposition 3.2. Suppose that A is a C*-algebra, κ is an infinite cardinal and
{Aξ : ξ < κ} is an irredundant set in A. Then there is an irredundant set {Bξ :
ξ < κ} consisting of positive elements of A.
Proof. Given X ⊆ κ let AX be the C*-subalgebra of A generated by {Aξ : ξ ∈ X}.
Clearly (Aη + A∗η)/2, (Aη − A∗η)/2i ∈ Aκ\{ξ} for every η 6= ξ and (Aξ + A∗ξ )/2 and
(Aξ − A∗ξ )/2i cannot both belong to Aκ\{ξ}. So {Bξ : ξ < κ} is irredundant set
consisting of self-adjoint elements, where Bξ ∈ {(Aξ + A∗ξ)/2, (Aξ − A∗ξ)/2i} is such
that Bξ 6∈ Aκ\{ξ}.
To prove that we can obtain the same cardinality irredundant set consisting of
all positive elements, by the above we may assume that the original Aξs are self-
adjoint. We have Aξ = Aξ + − Aξ−
} which
= (Aη − Aη)/2 for
does not belong to Aκ\{ξ}. But Aη + = (Aη + Aη)/2 and Aη−
. Note that there is Bξ ∈ {Aξ +, Aξ−
η belong to Aξ for all η ∈ κ \ {ξ}. So {Bξ : ξ < κ} is irredundant, as
(cid:3)
Aη =qA2
required.
The following proposition shows the role of being scattered while extracting
irredundant sets consisting of projections.
Proposition 3.3. Suppose that A is a scattered C*-algebra, κ is an infinite cardinal
and {Aξ : ξ < κ} is an irredundant set in A. Then there is an irredundant set
{Pξ : ξ < κ} consisting of projections.
Proof. By Lemma 3.2 we may assume that Aξs are self-adjoint. Let us adopt the
notation AX for X ⊆ κ from the proof of Lemma 3.2.
Since subalgebras of scattered algebras are scattered, A{ξ}s are scattered for each
ξ < κ and so of the form C0(K{ξ}) for some locally compact scattered K{ξ} which
must be totally disconnected. It follows that linear combinations of projections of
A{ξ}s are norm dense in A{ξ}s. Hence for each ξ < κ there is a projection Pξ ∈ A{ξ}
such that Pξ 6∈ Aκ\{ξ}. It follows that {Pξ : ξ < κ} is irredundant.
(cid:3)
3.2. Irredundant sets in commutative C*-algebras. The following two lem-
mas characterize irredundant sets in commutative C*-algebras.
Lemma 3.4. Suppose that K is compact Hausdorff space and X ⊆ C(K) is such
that no f ∈ X belongs to the unital C*-subalgebra of C(K) generated by X \ {f }.
Then for each f ∈ X there are xf , yf ∈ K such that f (xf ) 6= f (yf ) but g(xf ) =
g(yf ) for any g ∈ X \ {f }.
Consequently if X is a nonempty irredundant set in C(K), then there is h ∈ X
such that X \ {h} has the above property.
Proof. By the Gelfand representation we may assume that C(K) is the unital C*-
algebra generated by X . By the complex Stone-Weierstrass theorem the proper
C*-subalgebra generated by X \ {f } does not separate a pair of points of K, say
xf , yf . But they must be separated by f by the fact that X generated C(K).
The last part of the lemma follows from Lemma 3.1
(cid:3)
Lemma 3.5. Suppose that X is locally compact noncompact Hausdorff space and
X ⊆ C0(X) is irredundant then for every f ∈ X there are xf , yf ∈ X such that
either
• f (xf ) 6= 0 and g(xf ) = 0 for all g ∈ X \ {f }, or
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
11
• f (xf ) 6= f (yf ) but g(xf ) = g(yf ) for all g ∈ X \ {f }.
Points xf satisfying the first case form a discrete subspace of X
Proof. Let K = X ∪ {∞} be the one-point compactification of X. We will identify
C0(K) with a C*-subalgebra of C(K). Note that X satisfies the hypothesis of
Lemma 3.4, because if f = λ1 + g for f, g ∈ C0(X), the unit would be in C0(X)
which contradicts the hypothesis that X is noncompact. So we obtain the pairs of
points xf , yf ∈ K as in Lemma 3.4. The first case of the lemma corresponds to the
situation when one of the points xf , yf is ∞, say yf . But then h(yf ) = 0 for all
h ∈ C0(X) which implies f (xf ) 6= 0 and g(xf ) = 0 for all g ∈ X \ {f }.
Considering open sets Uf = {x ∈ X0 : f (x) 6= 0} we obtain neighbourhoods
(cid:3)
witnessing the discretness of the set of xf s satisfying the first case.
In fact discrete subsets of K provide strong irredundant subsets in C(K):
Remark 3.6. Suppose that K is compact Hausdorff space and D ⊆ K is discrete.
For each d ∈ D consider fd ∈ C(K) such that fd(d) 6= 0 and fd(d′) = 0 for all
d′ ∈ D \ {d}. Then fd does not belong to the ideal generated by {fd′ : d′ ∈ D \ {d}}.
In particular {fd : d ∈ D} is irredundant.
One should, however, note that there could be dramatic gap between the sizes
of discrete subsets and the sizes of irredundant sets:
Remark 3.7. Let K be the split interval, i.e., {0, 1}N×{0, 1} with the order topology
induced by the lexicographical order. Then K has no uncountable discrete subset
(in fact, K is hereditarily separable and hereditarily Lindelof ) but C(K) has an
irredundant set {χ[0N⌢0,x⌢0] : x ∈ {0, 1}N} of cardinality continuum.
Most of the literature concerning implicitly or explicitly irredundant sets are
related to Boolean algebras. As shown in the following lemma the relationship
between Boolean irredundance and irredundance for C*-algebras is very close in
the light of lemma 3.1.
Definition 3.8. A subset X of a Boolean algebra A is called Boolean irredundant if
for every x ∈ X the element x does not belong to the Boolean subalgebra generated
by X \ {x}.
Lemma 3.9. Suppose that A is a unital C*-algebra and B ⊆ A is a Boolean algebra
of projections in A and X ⊆ B is Boolean irredundant. Then X is irredundant in
A.
Suppose that K is a totally disconnected space and X ⊆ C(K) consists of projec-
tions and no element of x ∈ X belongs to the unital C*-algebra generated by X \{x}.
Then X is Boolean irredundant in the Boolean algebra {χU : U ∈ Clop(K)}.
Proof. Let C be the C*-subalgebra of A generated by B.
It is abelian, so it is
of the form C(K) where K is the Stone space of B.
It is enough to prove that
X is irredundant in C. But given a proper Boolean subalgebra, there are distinct
ultrafilters on the superalgebra which coincide on the subalgebra. These ultrafilters
are the points of K witnessing the irredundance of F as in Lemma 3.15.
(cid:3)
For commutative scattered C*-algebras the relationship between Boolean and
C*-algebraic irredundance is even closer:
12
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Corollary 3.10. Suppose that A is an infinite superatomic Boolean algebra. Then
the Boolean irredundance of A is the same as irr(C(KA)), where KA is the Stone
space of A.
Proof. As A is infinite, its Boolean irredundance is infinite (just take an infinite
pairwise disjoint collection). By the first part of Lemma 3.9 the Boolean irredun-
dance of A is not bigger than irr(C(KA)). On the other hand consider any infinite
irredundant subset X of C(KA). KA is scattered as A is superatomic and so C(KA)
is a scattered C*-algebra. By Proposition 3.3 there is an irredundant subset Y of
C(KA) of the same cardinality as X and consisting of projections in C(KA). By
removing at most one element of Y, by the second part of Lemma 3.9 and Lemma
3.1 it is a Boolean irredundant set in the Boolean algebra {χU : U ∈ Clop(KA)}
and this yields a Boolean irredundant set in A.
(cid:3)
The above positively answers Question 3.10 (3) of [12] in the case of a scattered
space.
Corollary 3.11. Suppose that K is an infinite Hausdorff compact space. Then
irr#(CR(K)) = irr(C(K)) where irr#(CR(K)) is the supremum over the cardinal-
ities of sets X of real-valued continuous functions on K such that no f ∈ X belongs
to the real unital Banach algebra generated by X \ {f }. In particular the π-weight
of K is bounded by irr(C(K)) and the density of C(K) is bounded by 2irr(C(K)).
Proof. Let X ⊆ C(K) be an infinite irredundant set. By Lemma 3.2 we may assume
that it consists of real-valued (non-negative) functions. As in the proof of Lemma
3.1, by removing at most one element we may assume that no f ∈ X belongs to
the real unital C*-algebra generated by X \ {f }. So irr#(CR(K)) ≥ irr(C(K)).
Now given a set X as in the lemma, by the real unital Weierstrass-Stone theorem
there are pairs of points xf , yf ∈ K such that f (xf ) 6= f (yf ) but g(xf ) = g(yf ) for
any g ∈ X \ {f } (cf. [30]). hence X is an irredundant set in the C*-algebra C(K)
by Lemma 3.4.
The last part of the corollary follows from Theorem 10 of [24] where π(K) ≤
irr#(CR(K)) is proved and from the fact that the weight of a regular space is
bounded by the exponent of its π-weight (Theorem 3.3. of [21]).
(cid:3)
3.3. Irredundant sets in general C*-algebras. Having developed the moti-
vations in the previous section now we move to the irredundant sets in general,
possibly noncommutative C*-algebras.
Proposition 3.12. Every infinite pairwise orthogonal collection of self-adjoint el-
ements in a C*-algebra is irredundant. In particular, every infinite dimensional
C*-algebra contains an infinite irredundant set.
Proof. This follows form the fact that given a self-adjoint element A of a C*-algebra
A the set {B ∈ A : AB = BA = 0} is a C*-subalgebra of A.
(cid:3)
Proposition 3.13. Suppose that an infinite dimensional C*-algebra A is a von
Neumann algebra. Then A has an irredundant set of cardinality continuum.
Proof. An infinite dimensional von Neumann algebra has an infinite pairwise or-
thogonal collection of projections and so it contains the commutative C*-algebra
ℓ∞ which is ∗-isomorphic to C(βN). The Boolean algebra ℘(N) is isomorphic to
{χU : U ∈ Clop(βN)} and so the Boolean irredundance of ℘(N) is equal to the
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
13
irredundance of ℓ∞ by Lemma 3.10. By considering an almost disjoint family (or
an independent family) of cardinality continuum of subsets of N we obtain an irre-
dundant set of cardinality continuum.
(cid:3)
The following Proposition corresponds to Remark 3.6.
Lemma 3.14. Suppose that A is a C*-algebra, κ a cardinal {Aξ : ξ < κ} ⊆ A+
and {τα : α < κ} a family of states such that
• τα(Aα) > 0
• τα(Aξ) = 0 for ξ 6= α.
Then {Aξ : ξ < κ} is irredundant.
Proof. As in the GNS construction one proves that Lα = {A ∈ A : τα(A∗A) = 0} is
a left-ideal in A, and in particular a C*-subalgebra. So Xξ ∈ Lα, where Aξ = X∗ξ Xξ
and so Aξ ∈ Lα for all ξ 6= α. However by Theorem 3.3.2. of [39] we have
0 < τα(Aα) ≤ kταkτα(A∗αAα),
so Aα 6∈ Lα.
(cid:3)
In the noncommutative case, for pure states {τα : α < κ} being discrete in
the weak* topology does not yield in general the existence of positive elements Aα
like in the lemma above, as the noncommutative Urysohn lemmas require extra
hypotheses ([2]).
The following proposition is a version of the commutative characterizations in
Lemmas 3.4 and 3.5. It could be interesting to remark that a version of the fol-
lowing proposition where "representations" are replaced by "irreducible represen-
tations" implies the noncommutative Stone-Weierstrass theorem, which remains a
well-known open problem. One should note that below one of the possibilities of
the representation is to be constantly zero.
Proposition 3.15 ([22]). Suppose that A is a C*-algebra and X ⊆ A is an irre-
dundant set. Then for all a ∈ X there are Hilbert spaces Ha and representations
π1
a, π2
a : A → B(Ha) such that π1
a(a) 6= π2
a(a) but
X \ {a} ⊆ {b ∈ A : π1
a(b) = π2
a(b)}.
3.4. Irredundance in scattered C*-algebras. The following proposition shows
that thin-tall algebras play a special role in the context of uncountable irredundant
sets.
Proposition 3.16. If there is a nonseparable scattered C*-algebra with no uncount-
able irredundant set, then it contains a thin-tall scattered C*-algebra.
Proof. First note that by the characterization of subalgebras of the algebra of com-
pact operators ([5]) a C*-algebra which is isomorphic to a subalgebra of the algebra
of all compact operators on a Hilbert space H but not isomorphic to a subalgebra
of the algebra of all compact operators on the separable Hilbert space H must con-
tain an uncountable pairwise orthogonal set which is irredundant by 3.12. So if a
scattered A has no uncountable irredundant set, then all the algebras I At(A/Iα)
are *-isomorphic to a subalgebra of the algebra of all compact operators on the
separable or finite dimensional Hilbert space, but as A is nonseparable ht(A) ≥ ω1
and so Iω1 is the required thin-tall subalgebra of A.
(cid:3)
14
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Corollary 3.17. Assume PFA. Suppose that A is a nonseparable scattered C*-
algebra. Then there is an uncountable weak∗ discrete set of pure states of A.
Proof. First suppose that A has a quotient that contains an uncountable orthogonal
set of projections. Then it is clear that we can find pure states which form a weak∗
discrete set. Otherwise using the argument as in the proof of Proposition 3.16 we
may assume that A is thin-tall. By Proposition 2.6 A has an uncountable weak∗
discrete set of pure states.
(cid:3)
Below we prove a simple noncommutative version of a Theorem of McKenzie
(see 4.2.3 of [28]):
Theorem 3.18. If A is a scattered C*-algebra, then
d(A) ≤ 2irr(A).
Proof. Let κ be the minimal cardinal such that I At(A) is a subalgebra of the
algebra of all compact operators on ℓ2(κ). By the characterization of subalgebras
of the algebra of compact operators ([5]) A must contain pairwise orthogonal set
of cardinality κ which is irredundant by 3.12. So κ ≤ irr(A). By the essentiality
of I At(A) which follows from Proposition 2.5 we can embed A into B(ℓ2(κ)), so
d(A) ≤ 2κ ≤ 2irr(A) as required.
(cid:3)
3.5. Extracting irredundant sets from a given collection of operators.
Proposition 3.19. There is a collection of operators (Aξ : ξ < ω1) in B(ℓ2) which
generates a nonseparable C*-subalgebra of B(ℓ2) with no two-element irredundant
subset. Any fully noncommutative thin-tall C*-algebra is generated by such a se-
quence.
Proof. Construct a fully noncommutative thin-tall C*-algebra A as in Theorem 7.6.
of [18], in particular with Cantor-Bendixson decomposition (I At
α (A))α<ω1 (see 2.2
(2)), where I At
α+1(A) is *-isomorphic to ^I At
α (A) ⊗ K(ℓ2).
By Theorem 8 of [42] any C*-algebra of the form B ⊕ K(ℓ2) is singly generated
if B is separable and unital. So for each α < ω1 pick Aα to be a single generator of
I At
α+1(A).
An alternative approach which gives the final statement of the Proposition is to
use the fact that scattered C*-algebras are locally finite dimensional (see [16] for
more on these notions in the nonseparable context) in the sense that each of its
finite subsets can be approximated from a finite dimensional C*-subalgebra ([34],
[35]). So I At
α (A) is locally finite dimensional and separable for each α < ω and so
AF. Thus the result of [48] implies that I At
α (A) is singly generated for each α < ω1.
So pick Aα+1 as before. This completes the proof of the theorem.
(cid:3)
Using the free set lemmas like in [13] one can prove that given a discrete set
of operators (Aα)α<ωn for n ∈ N there is an n-element irredundant set. However
there is much stronger consistent extraction principle:
Theorem 3.20. It is relatively consistent that whenever (Aξ : ξ < 2ω) is a collec-
tion of operators in B(ℓ2) which generates a C*-algebra of density continuum, then
there is a set I ⊆ 2ω of cardinality continuum such that (Aξ : ξ ∈ I) is irredundant.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
15
Proof. To obtain the relative consistency we will use the method of forcing (see
[33]). We start with the ground model V satisfying the generalized continuum
hypothesis (GCH) and we will consider the generic extension V [G] where G is a
generic set in the forcing P = F n(ω2, 2) for adding ω2 Cohen reals (see Chapter
VIII §2 of [33]).
Fix a ground model orthonormal basis (en : n ∈ N) for ℓ2 in V . In V [G] let
(Aξ : ξ < 2ω) be as in the theorem. By passing to a subset of cardinality 2ω = ω2
and using the hypothesis that (Aξ : ξ < 2ω) is a collection of operators in B(ℓ2)
which generates a C*-algebra of density continuum, we may assume that Aξ does
not belong to the C*-algebra generated by the operators (Aη : η < ξ) for each
ξ < ω1. Moreover by passing to a subsequence we may assume that there is a
rational ε > 0 such that kA − Aξk > ε for every A in the C*-algebra generated by
the operators (Aη : η < ξ) for each ξ < ω1.
Each Aξ can be identified with an N×N complex valued matrix (hAξ(en), emi)m,n∈N.
Let Aξ be P-names in V for these matrices. Using the standard argument of nice
names, the countable chain condition for P and passing to a subsequence using the
∆-system lemma for countable sets which follows from the GCH, we may assume
that there are permutations σξ,η : ω2 → ω2 which lift to the automorphisms of P
and the permutations σ′ξ,η of P names such that
and for every ξ, η ∈ ω2 we have that
σ′ξ,η( Aη) = Aξ,
P (cid:13) φ( x1, ..., xk) if and only if P (cid:13) φ(σ′ξ,η( x1), ..., σ′ξ,η( xk))
for any formula φ in k ∈ N free variables and any sequence x1, ..., xk of P-names
for k ∈ N (7.13 [33]). Using this for the formulas which say that the distance of Aξ
from any element of the C*-algebra generated by the operators (Aη : η ∈ F ) for
any finite F ⊆ ξ is bigger than ε, we conclude that P forces that no Aξ belongs the
C*-algebra generated by any countable collection from { Aη : η 6= ξ} (by considering
a permutation of ω2 which moves ξ above the countable set). This means that P
forces that no Aξ belongs the C*-algebra generated by the remaining operators
{ Aη : η 6= ξ}, i.e., that the collection is irredundant as required.
(cid:3)
The above is a version of applying a standard argument as in [52] in the context
of Boolean irredundance.
4. Commutators under OCA
The main consistent construction of this paper presented in the following sections
has a strong randomness properties. In this section we show that this randomness
does not take place for any uncountable collection of operators in B(ℓ2) under the
assumption of Open Coloring Axiom, OCA. We will follow the approach to the
strong operator topology from the book [11] of Davidson. Thus we have:
Definition 4.1. Let H be a Hilbert space. The strong operator topology (SOT) on
B(H) is defined as the weakest topology such that the sets
S(a, x) := {b ∈ B(H) : (b − a)(x) < 1}
16
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
are open for each a ∈ B(H) and x ∈ H . We denote by (B(H), τsot) and (B(H)1, τsot)
respectively the space B(H) and the unit ball of B(H) with the strong operator topol-
ogy.
Proposition 4.2. If H is a separable Hilbert space, then (B(H)1, τsot) is metrizable
and separable in the strong operator topology.
Proof. For metrizability see [11], Proposition I.6.3. For the separability fix some
orthonormal basis (en)n∈N and consider finite rank operators in the linear span of
one dimensional operators of the form v ⊗ w where v, w have finitely many nonzero
It is clear that such operators are
rational coordinates with respect to (en)n∈N.
SOT dense in B(l2)1, as required.
(cid:3)
By the remarks on page 16 and 17 of [11] we have the following:
Lemma 4.3. The multiplication on B(H)1 is jointly continuous in the SOT topol-
ogy and so every polynomial6 is SOT continuous on B(H)1.
We will follow the approach to the Open Coloring Axiom (OCA) from [15], page
55. Its weaker version was discovered by Abraham, Rubin and Shelah ([1]) and the
final form was introduced by Todorcevic in [51]. It is consistent with ZFC. In fact,
it is a consequence of the Proper Forcing Axiom (PFA). See Theorem 8 of [51].
Recall that
[X]2 = {{x, y} ⊆ X : x 6= y}.
It is well known that the original form of OCA from [51] for subsets of the reals is
equivalent to the version for separable metric spaces as in [15]:
Definition 4.4 (Todorcevic [51]). OCA denotes de following statement: If X is
a separable metric Hausdorff space and [X]2 = K0 ∪ K1 is a partition with K0
open7, then either there is an uncountable Y ⊆ X such that [Y ]2 ⊆ K0, or else
X =Sn∈N Xn where [Xn]2 ⊆ K1 for each n ∈ N.
Theorem 4.5 (OCA). Let (Aα)α<ω1 be an uncountable family in B(l2) and P (x, y)
be a polynomial satisfying kP (A, B)k = kP (B, A)k for all A, B ∈ B(l2). Then
given ε > 0, either there is an uncountable Γ0 ⊂ ω1 such that P (Aα, Aβ) ≤ ε
for every distinct α, β ∈ Γ0 or else there is an uncountable Γ1 ⊂ ω1 such that
P (Aα, Aβ) > ε for every distinct α, β ∈ Γ1.
Proof. As X is uncountable, by passing to an uncountable subset, we may assume
that there is M > 0 such that kAαk ≤ M for all α < ω1. Let X = {Aα : α <
ω1} ⊆ M B(ℓ2)1 and note that M B(ℓ2)1 is metric and separable by Proposition 4.2.
Define
K0 = {{A, B} ∈ [X]2 : P (A, B) > ε}
and K1 = [X]2 \ K0.
First note that the separability is hereditary for metric spaces, so X is metric
separable as a subspace of (M B(l2)1, τsot).
Now note that K0 is open.
Indeed if kP (A, B)k > ε, then there is x ∈ ℓ2
of norm one and δ > 0 such that kP (A, B)(x)k > ε + δ. Now if P (A′, B′) ∈
6By a polynomial P (x, y) we mean a expression in the form P (x, y) = Pi aixi + Pi biyi +
Pi,j ci,j xiyj + Pi,j di,j yixj + e0.
7We call K0 ⊆ [X]2 open if the symmetric set {(x, y) ∈ X × X : {x, y} ∈ K0} is open in
K × K \ ∆ in the product topology, where ∆ denotes the diagonal of X × X.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
17
S(P (A, B), x/δ), we have kP (A′, B′)(x) − P (A, B)(x)k < δ and so kP (A′, B′)k > ε;
Hence {{A′, B′} ∈ [X]2 : P (A′, B′) ∈ S(P (A, B), x/δ)} ⊆ K0. But (A, B) ∈
P −1[S(P (A, B), x/δ)] is open in X × X with the product SOT topology by the
continuity of P (Lemma 4.3).
So we are in the position of applying the OCA. From 4.4 we obtain the required
uncountable set Γ0 or Γ1.
(cid:3)
Corollary 4.6 (OCA). Let (Aα)α<ω1 be an uncountable family in B(l2). Then
given ε > 0, either there is an uncountable Γ0 ⊂ ω1 such that [Aα, Aβ] ≤ ε for
every α, β ∈ Γ0 or else there is an uncountable Γ1 ⊂ ω1 such that [Aα, Aβ] > ε
for every α, β ∈ Γ1.
Proof. Consider P (x, y) = xy − yx and apply Theorem 4.5.
(cid:3)
Remark 4.7. Let us remark on two trivial versions of the above results. First
let (An)n∈N be an infinite family in B(l2). Then given ε > 0, either there is an
infinite Γ0 ⊂ N such that [An, Am] ≤ ε for every n, m ∈ Γ0 or else there is an
infinite Γ1 ⊂ N such that [An, Am] > ε for every n, m ∈ Γ1. This follows from
the Ramsey theorem whose consistent generalization is the OCA.
Secondly note that if (Aα)α<ω1 is an uncountable family in a separable C*-
subalgebra of B(ℓ2), then by its second countability in the norm topology it follows
that for every δ > 0 there is an uncountable Γ0 ⊆ ω1 such that kAα − Aβk < δ
for every α, β ∈ Γ0 and so given any polynomial P satisfying P (x, x) = 0 and
ε > 0, by the norm continuity of P there is an uncountable Γ0 ⊆ ω1 such that
kP (Aα, Pβ)k < ε for every α, β ∈ Γ0.
In fact, in the nontrivial cases of Theorem 4.5 and Corollary 4.6 when (Aα)α<ω1
generates a nonseparable C*-subalgebra of B(ℓ2) we may assume that (Aα)α<ω1
forms a norm discrete set.
5. The partial order of finite dimensional approximations
5.1. Notation. The C*-algebras that we consider in the rest of this paper are sub-
algebras of B(ℓ2(ω1 ×N)). In fact, the subspaces ℓ2({ξ}×N) of ℓ2(ω1 ×N), which we
call columns will be invariant for all our algebras, so our algebras could be identified
with subalgebras of Πξ<ω1 B(ℓ2({ξ} × N)). Also the map πα : Πα≤ξ<ω1B(ℓ2({ξ} ×
N)) → B(ℓ2({α} × N) applied to the appropriate quotients, will be faithful (see
Lemma 5.24 (3)). Thus the purpose of this presentation of the algebras is re-
lated to the transparent structure of the Cantor-Bendixson composition series (see
Proposition 5.25 (3)).
For X ⊆ ω1 × N, we introduce the following notation:
• (eξ,n : ξ < ω1, n ∈ N) is the canonical orthonormal basis of ℓ2(ω1 × N),
• the family of all operators A in B(ℓ2(ω1 × N)) such that
-- ℓ2(X ∩ ({ξ} × N)) is A-invariant for all ξ < ω1,
-- A(eξ,n) = 0 whenever (ξ, n) 6∈ X,
will be denoted by BX ,
• the unit of the C*-algebra BX will be denoted by PX ,
• 1ξ,m,n is the operator in Bω1×N satisfying
1ξ,m,n(eη,k) =(eξ,m if k = n, ξ = η
otherwise,
0
18
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
• if A ∈ Bω1×N we define AX = APX ,
• if A ∈ Bω1×N and a ⊆ ω1 we define Aa as A(a × N),
• AX = {AX : A ∈ A} for A ⊆ Bω1×N and X ⊆ ω1 × N.
5.2. The definition of the partial order of finite-dimensional approxima-
tions.
Definition 5.1. We define a partial order P consisting of elements
p =(cid:0)ap, {np
where
ξ : ξ ∈ ap}, {Ap
ξ,m,n : ξ ∈ ap, n, m ∈ [0, np
ξ)}(cid:1),
(1) ap is a finite subset of ω1,
(2) np
(3) Ap
ξ ∈ N for each ξ ∈ ap,
ξ,m,n ∈ BXp for each ξ ∈ ap and n, m ∈ [0, np
Xp = {(ξ, n) : ξ ∈ ap; n ∈ [0, np
ξ), where
ξ)},
ξ,m,n = (Ap
(4) Ap
ξ,m,nξ) + 1ξ,m,n for each ξ ∈ ap and n, m ∈ [0, np
The order ≤P=≤ on P is defined by declaring p ≤ q if and only if:
ξ).
(a) ap ⊇ aq,
(b) np
ξ ≥ nq
(c) there is a (nonunital) *-embedding ipq : BX q → BX p such that ipq(Aq
ξ for ξ ∈ aq,
ξ,m,n) =
Ap
ξ,m,n for all ξ ∈ aq and m, n ∈ [0, nq
ξ),
(d) ip,q(A)Xq = A for all A ∈ BX q .
Definition 5.2. Suppose that p ∈ P and X ⊆ Xp. Then the C∗-subalgebra of BXp
generated by {Ap
ξ,m,n : (ξ, m), (ξ, n) ∈ X} is denoted by Ap
X .
Lemma 5.3. For every α ∈ ω1 and every p ∈ P we have
In particular Ap
Xp
= BXp .
Ap
Xp∩(α×N) = BXp∩(α×N).
Proof. We will prove it by induction on ap ∩ α. If ap ∩ α = ∅, then both of the
algebras are {0}. Suppose ap ∩ α = n + 1 and we have proved the Lemma for
every q ∈ P and α < ω1 such that aq ∩ α = n. Let ξ = max(ap ∩ α). By the
definition of BXp we have that BXp∩(α×N) is *-isomorphic to BXp∩(ξ×N) ⊕ B{ξ}×N.
By the inductive hypothesis, BXp∩(ξ×N) is generated by {Ap
η,m,n : η ∈ ap ∩ ξ; m, n ∈
[0, np
ξ,m,n − A for some
A ∈ BXp∩(ξ×N) and all m, n ∈ [0, np
ξ ) is included in the
algebra generated by {Ap
η)}. This together with the
inductive hypothesis completes the proof.
(cid:3)
ξ)}. But by (4) in Definition 5.1, we have that 1ξ,m,n = Ap
ξ). In particular, B{ξ}×[0,np
η,m,n : η ∈ ap ∩ α; m, n ∈ [0, np
Lemma 5.4. Suppose that α < ω1 and p, q ∈ P satisfy p ≤ q and A = ip,q(B),
where B ∈ Aq
Xq
. Then
kA[α, ω1)k = kB[α, ω1)k.
Proof. Since Bα and B[α, ω1) are in Aq
Xq
by Lemma 5.3, we have
kA[α, ω1)k = kip,q(B)[α, ω1)k = kip,q(Bα)[α, ω1) + ip,q(B[α, ω1))[α, ω1)k.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
19
But Bα ∈ Aq
the isomorphism ip,q, i.e., ip,q(Bα)[α, ω1) = 0 and so
Xq∩(α×N) by Lemma 5.3, and this generation must be preserved by
kA[α, ω1)k = kip,q(B[α, ω1)[α, ω1)k ≤ kip,q(B[α, ω1)k.
Since ip,q is an embedding (in particular an isometry) , we conclude that
kA[α, ω1)k ≤ kB[α, ω1)k.
The other inequality follows from Definition 5.1 (c-d).
(cid:3)
5.3. Density Lemmas. In the terminology related to partial orders occurring in
the theory of forcing a subset D of a partial order Q is said to be dense if for every
p ∈ P there is d ∈ D satisfying d ≤ p. In what follows we usually need stronger
information for Q = P namely that ad = ap.
Lemma 5.5. Suppose that ξ < ω1. Then
Dξ = {p ∈ P : ξ ∈ ap}
is a dense subset of P.
Proof. Let q ∈ P be such that ξ /∈ aq. Define p as follows:
• ap = aq ∪ {ξ},
• np
• Ap
η = nq
η,m,n = Aq
η for η ∈ aq and np
ξ = 1,
η,m,n for η ∈ aq and Ap
ξ,0,0 = 1ξ,0,0.
It is clear that p ∈ P. Also p ≤ q as IdBXq : BXq → BXp is a *-embedding good for
ip,q in Definition 5.1 (c).
(cid:3)
Lemma 5.6. Suppose that ξ < ω1; k ∈ N and q ∈ P is such that ξ ∈ aq. Then
there is
p ∈ Eξ,k = {p ∈ P : ξ ∈ ap, np
ξ ≥ k}
such that p ≤ q and ap = aq.
Proof.
Consider q ∈ P such that ξ ∈ aq but nq
ξ < k. Define p as follows:
η for η ∈ ap \ {ξ} and np
• ap = aq,
η = nq
• np
η,m,n = Aq
• Ap
ξ,m,n = Aq
• Ap
• Ap
ξ,m,n = 1ξ,m,n if n, m ∈ [0, k) and {n, m} ∩ [nq
η,m,n η ∈ aq \ {ξ},
ξ,m,n for n, m ∈ [0, nq
ξ = k,
ξ),
ξ, k) 6= ∅.
It is clear that p ∈ P ∩ Eξ,k. Also p ≤ q as IdBXq : BXq → BXp is a *-embedding
good for ip,q in Definition 5.1 (c).
(cid:3)
Lemma 5.7. Suppose that q ∈ P and X ⊆ Xq and that α ∈ aq. Then there is
p ≤ q such that p ∈ FX,α, where
FX,α = {p ∈ P : α ∈ ap, X ⊆ Xp, and ∀A ∈ Ap
X kA{α}k ≥ kA[α, ω1)k}.
Moreover, ap = aq and np
ξ = nq
ξ whenever ξ ∈ ap \ {α}.
20
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Proof. Let q ∈ P. We may assume that X = Xq. If α = max(ap), then there is
nothing to prove. So let aq \ (α + 1) = {ξ1, ..., ξk} for some k ∈ N and put
l =X{nq
ξi
: 1 ≤ i ≤ k}.
Consider Y = Xq ∩ ((α, ω1) × N). Let φ : Y → [nq
obtain a ∗-homomorphism i : BXq → B
where ir : BXq → B{α}×[nq
hir(A)(eα,nq
α+l) satisfies
α,nq
α+φ(ξi,k)), eα,nq
Xq∪(cid:0){α}×[nq
α,nq
α, nq
α + l) be any bijection. We
α+l)(cid:1) given by i(A) = A+ir(A)
α+φ(ξi′ ,k′)i = hA(eξi,k), eξi′ ,k′ i
for all (ξ, k), (ξ′, k′) ∈ Y and every A ∈ BXq . Define p in the following way
ξ if ξ ∈ ap \ {α} and np
• ap = aq,
ξ = nq
• np
• Ap
ξ,m,n = i(Aq
α,m,n = 1α,m,n if {m, n} ∩ [nq
• Ap
α + l,
ξ,m,n) for (ξ, m), (ξ, n) ∈ Xq,
α) 6= ∅.
α = nq
α, np
It is clear from the construction that p ∈ P as condition (4) of Definition 5.1 is
satisfied due to the fact that we change only Aq
ξ,m,n for ξ > α on {α} × N, and that
(a), (b) of Definition 5.1 are satisfied.
If we put ip,q = i, condition (c) follows from the fact that i is a ∗-embedding since
ξ,m,n for (ξ, m), (ξ, n) ∈
α + n) ∩ Xq = ∅. We also have ip,q(Aq
{α} × [nq
Xq. The construction yields (d) of Definition 5.1.
ξ,m,n) = Ap
α, nq
Finally to check the main assertion of the lemma note that by Lemma 5.4 for
any A ∈ BXq we have
kip,q(A){α}k = max(kip,q(A){α}k, kip,q(A){α} × [nq
α, nq
α + n)k) =
= max(kip,q(A){α}k, kA(α, ω1)k) = max(kip,q(A){α}k, kip,q(A)(α, ω1)k) =
for any A ∈ BXq as required since X ⊆ Xq.
= kip,q(A)[α, ω1)k
(cid:3)
Lemma 5.8. Let X ⊆ ω1 × N be finite and α ∈ X. If q ∈ FX,α and p ≤ q, then
p ∈ FX,α.
Proof. Let A ∈ Ap
First note that by Lemma 5.4
X . As X ⊆ Xq we have that A = ip,q(B) for some B ∈ Aq
X ⊂ Aq
Xq
.
kA[α, ω1)k = kB[α, ω1)k.
Now kB[α, ω1)k ≤ kB{α}k by the hypothesis that q ∈ FX,α. But kB{α}k ≤
kA{α}k by the fact that AX q = B by Definition 5.1 (d). So kA[α, ω1)k ≤ kA{α}k
as required.
(cid:3)
5.4. Basic amalgamations.
Definition 5.9. We say that two elements p, q ∈ P are in the convenient position
(as witnessed by σ : ap → aq) if and only if
and there is an order preserving bijection σ : ap → aq such that
∆ := ap ∩ aq < ap \ ∆ < aq \ ∆
• np
ξ = nq
σ(ξ) for ξ ∈ ap,
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
21
and the ∗-isomorphism of BXq onto BXp induced by σ, denoted by jσ, which is given
by
for every (ξ, k), (ξ, l) ∈ Xp and A ∈ BXq satisfies
hjσ(A)(eξ,k), eξ,li = hA(eσ(ξ),k), eσ(ξ),li
• jσ(Aq
σ(ξ),n,m) = Ap
ξ,n,m for every ξ ∈ ap, n, m ∈ [0, np
ξ).
Lemma 5.10. Suppose that two elements p, q ∈ P are in the convenient position
as witnessed by σ : ap → aq and that ξ ∈ ∆ = ap ∩ aq. Then Aq
ξ,n,m for
every n, m ∈ [0, np
ξ,n,m = Ap
ξ) = [0, nq
ξ).
Proof. Note that in Definition 5.9 the bijection σ must be the identity on ∆ because
it is order-preserving and ∆ is the initial fragment of both ap and aq and so any
ξ ∈ ∆ must have the same position in both ap and aq. So jσ(Aq
ξ,n,m and
it is enough to prove that jσ(Aq
ξ,n,m) = Ap
ξ,n,m) = Aq
ξ,n,m.
For η ∈ ap \ ∆ we have
hjσ(Aq
ξ,n,m)(eη,k), eη,li = hAq
ξ,n,m(eσ(η),k), eσ(η),li = 0
for every k, l ∈ N such that (η, k), (η, l) ∈ Xp as σ(η) ∈ aq \ ∆.
On the other hand for η ∈ ∆ we have σ(η) = η and so
hjσ(Aq
ξ,n,m)(eη,k), eη,li = hAq
ξ,n,m(eη,k), eη,li
for every k, l ∈ N such that (η, k), (η, l) ∈ Xp as σ(η) = η by Definition 5.9. Using
Definition 5.1 (4) this proves the required Aq
(cid:3)
ξ,n,m = jσ(Aq
ξ,n,m) = Ap
ξ,n,m.
Lemma 5.11. Suppose that p, q ∈ P are in the convenient position as witnessed by
σ : ap → aq. Then there is r ≤ p, q such that
ξ = np
• ar = ap ∪ aq,
• nr
• ir,p = IdBXp , ir,q = IdBXq .
ξ if ξ ∈ ap and nr
ξ = nq
ξ if ξ ∈ aq,
In particular,
• Ar
• Ar
ξ,m,n = Ap
ξ,m,n = Aq
ξ,m,n for each ξ ∈ ap and n, m ∈ [0, nr
ξ,m,n for each ξ ∈ aq and n, m ∈ [0, nr
ξ),
ξ).
The element r will be called the disjoint amalgamation of p and q.
Proof. Define r as in the lemma. As p, q ∈ P, it is easy to see that r ∈ P. To see
that r ≤ p, q note that IdBXp and IdBXq are ∗-embeddings into BXr .
(cid:3)
Lemma 5.12. Suppose that p, q are two elements of P in the convenient position
as witnessed by σq,p : ap → aq. Let U ∈ BXp∪Xq be a partial isometry satisfying
U U∗ = U∗U = PXp\Xq , where PXp\Xq is the projection on the space spanned by
{eξ,k : (ξ, k) ∈ Xp \ Xq}. Then there is rU = r ≤ p, q such that
ξ = np
• ar = ap ∪ aq,
• nr
• ir,p = IdBXp ,
• ir,q(A) = A + U jσq,p (A)U∗ for all A ∈ BXq ,
ξ if ξ ∈ ap, nr
ξ if ξ ∈ aq,
ξ = nq
in particular,
• Ar
• Ar
ξ,m,n = Ap
ξ,m,n = U Ap
σ−1
q,p(ξ),m,n
ξ,m,n for ξ ∈ ap and m, n ∈ [0, nr
ξ),
U∗ + Aq
ξ,m,n for ξ ∈ aq \ ap and m, n ∈ [0, nr
ξ).
22
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
ξ,m,nXq = Aq
The element rU will be called the U -including amalgamation of p and q; if U =
PXp\Xq , then rU is called the including amalgamation.
Proof. Define rU as in the lemma. It is clear by Definition 5.1 applied to p and q that
r ∈ P. rU ≤ p because IdBXp : BXp → BXr is a ∗-embedding. For rU ≤ q we note
that Ar
and that the formula ir,q(A) = A+U jσ(A)U∗ for all A ∈ BXq defines a *-embedding
from BXp to BXr . This is follows from the fact that sending A to U jσ(A)U∗ is a
*-homomorphism since BXp\Xq is Ap
-invariant, so ir,q is a *-homomorphism. But
its kernel is null since U jσ(A)U∗ = (U jσ(A)U∗)(Xp \ Xq) for all A ∈ BXq .
(cid:3)
Lemma 5.13. Suppose that v1, v2 are two orthogonal unit vectors of Cn for n > 1.
Then there is a unitary U ∈ Mn such that
U∗(cid:1)Xq = 0 since U U∗ = U∗U = PXp\Xq
ξ,m,n as(cid:0)U Ap
q,p (ξ),m,n
σ−1
Xp
k[U AU∗, A]k = 1/2
for every nonexpanding linear A ∈ Mn satisfying A(v1) = v1 and A(v2) = 0.
Proof. Choose an orthonormal basis v1, . . . vn of Cn starting with v1, v2 and consider
the orthogonal projection P ∈ Mn onto the line containing v1, so in particular we
have P (v1) = v1 and P (v2) = 0. Let U = V ⊕ In−2, U∗ = V ∗ ⊕ In−2, where
1√2
1√2! .
1√2
V = V ∗ = −1√2
0(cid:19) −1√2
1√2
1√2
0
So we obtain that
U P U∗ = −1√2
1√2
Hence
1√2
1√2!(cid:18)1
0
1√2! ⊕ 0n−2 =(cid:18) 1
2 − 1
− 1
2
2
1
2 (cid:19) ⊕ 0n−2.
[U P U∗, P ] = U P U∗P − P U P U∗ =
2 − 1
− 1
2
2
=(cid:18) 1
=(cid:18) 1
2
− 1
2
1
0 0(cid:19) ⊕ 0n−2 −(cid:18)1 0
2 (cid:19)(cid:18)1 0
0(cid:19) ⊕ 0n−2 −(cid:18) 1
2 − 1
− 1
2
0 0(cid:19)(cid:18) 1
0 (cid:19) ⊕ 0n−2 =(cid:18) 0
2 (cid:19) ⊕ 0n−2 =
0(cid:19) ⊕ 0.
2 − 1
0
− 1
2
1
2
0
2
1
2
And so k[U P U∗, P ]k = 1/2 and in particular
• [U P U∗, P ](v1) = (1/2)v2 and
• [U P U∗, P ](v2) = (−1/2)v1.
Since P equals A on the space spanned by v1 and v2, and U, U∗ leave this space
invariant, we have the same equalities for A instead of P , hence k[U AU∗, A]k ≥ 1/2.
The other inequality follows from the fact that k[B, C]k ≤ 1/2 for any two B, C
satisfying 0 ≤ B, C ≤ 1 by a result of Stampfli (Corollary 2 of [47]).
(cid:3)
Lemma 5.14. Suppose that p, q are two elements of P in the convenient position
as witnessed by σ : ap → aq such that ∆ < ap \∆ < aq \∆. Suppose that nq
ξ = n ≥ 1
for every ξ ∈ aq \ap and that v1 = (v1
n−1) are two orthogonal
unit vectors of Cn. Then there is r ≤ p, q such that
n−1), v2 = (v2
0, ...v1
0, ...v2
• ar = ap ∪ aq,
• nr
ξ = np
ξ if ξ ∈ ap, nr
ξ = nq
ξ if ξ ∈ aq,
and
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
23
• k[irq(A), irp(jσ(A))]k = 1/2
for every nonexpanding A ∈ BXq such that there is ξ ∈ aq\ap with A(Pk<n v1
Pk<n v1
keξ,k) =
keξ,k) = 0. We call r the (v1, v2)-anticommuting amal-
keξ,k and A(Pk<n v2
gamation of p and q.
Proof. By Lemma 5.13 for each ξ ∈ aq \ ap for ηξ = σ−1(ξ) there is a unitary
Uξ ∈ B{ηξ}×[0,n) such that
(∗)
k[Uξ(jσ(A){ηξ})U∗ξ , jσ(A){ηξ}]k = 1/2
whenever A ∈ BXq is nonexpanding such that
A(Xk<n
v1
keξ,k) =Xk<n
v1
keξ,k and A(Xk<n
v2
keξ,k) = 0.
Let U ∈ BX p∪X q be a partial isometry such that U ({ηξ} × [0, n)) = Uξ and
U is zero on the columns not in in X p \ X q, and U U∗ = PX p\X q . Consider the
U -including amalgamation rU ≤ p, q as in Lemma 5.12.
We claim that r = rU satisfies the lemma we are proving. Let A ∈ BXq be
keξ,k and
ξ)) for ξ ∈ aq \ ap are invariant for BX q
the operator A{ξ} is nonexpanding as well and so is jσ(A){ηξ}. By Lemma 5.12
we have ir,q(A) = A + U jσ(A)U∗ and ir,p(jσ(A)) = jσ(A), so for ηξ = σ−1(ξ) we
have
nonexpanding and ξ ∈ aq \ ap be such that A(Pk<n v1
A(Pk<n v2
keξ,k) = Pk<n v1
keξ,k) = 0. Since ℓ2({ξ} × [0, nq
[ir,q(A), ir,p(jσ(A))]({ηξ} × [0, n)) = [Uξ(jσ(A){ηξ})U∗ξ , jσ(A){ηξ}],
So by (*) we have k[ir,q(A), ir,p(jσ(A))]k ≥ 1/2. The other inequality follows from
the maximality of 1/2 (Corollary 2 of [47]).
(cid:3)
5.5. Types of 3-amalgamations.
Lemma 5.15. Suppose that p1, p2, p3 are distinct elements in P which are pairwise
in the convenient position as witnessed by σj,i : api → apj for 1 ≤ i < j ≤ 3 such
that ∆ < ap1 \ ∆ < ap2 \ ∆ < ap3 \ ∆. Then there is r ≤ p1, p2, p3 satisfying
• ar = ap1 ∪ ap2 ∪ ap3 ;
• there is n ∈ N such that for each ξ ∈ ar we have
n = nr
ξ > n′ = max{npi
ξ : ξ ∈ api , 1 ≤ i ≤ 3},
•
The element r is called the amalgamation of p1, p2, p3 of type 1.
r ∈\{FX,α : X ∈ {Xp1 , Xp2, Xp3}, α ∈ X},
Proof. Let ap1 = {α1, ..., αk} in the increasing order. Using Lemma 5.7 find p1 ≥
∈ FXp1 ,αj for 1 ≤ j ≤ k. Now using
p1
1 ≥ ... ≥ pk
ξ = n > n′ for
Lemma 5.6 several times find q1 ≤ pk
every ξ ∈ aq1.
1 such that aq1 = ap1 and nq1
1 such that ap1 = apk
and ap1
j
1
Now find q2, q3 ∈ P such that q2 ≤ p2 and q3 ≤ p3 and "isomorphic" with q1
i.e., with aq2 = ap2 , aq3 = ap3 and where q1, q2, q3 are pairwise in the convenient
position as witnessed by σj,i : aqi → aqj for 1 ≤ i < j ≤ 3. Note that by Lemma
5.8 we have
qi ∈\{FXpi ,α : α ∈ Xpi}.
24
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Now let s1 ≤ q1, q2 and s2 ≤ q1, q3 be the disjoint amalgamations as in Lemma
5.11. Note that s1 and s2 are in the convenient position as witnessed by Idap1 ∪σ3,2 :
ap1 ∪ ap2 → ap1 ∪ ap3 where as1 ∩ as2 = ap1. So now let r ≤ s1, s2 be the disjoint
amalgamation of s1 and s2 as in Lemma 5.11. Note that we have the final statement
of the lemma by Lemma 5.8.
(cid:3)
Lemma 5.16. Suppose that p1, p2, p3 are distinct elements in P which are pairwise
in the convenient position as witnessed by σj,i : api → apj for 1 ≤ i < j ≤ 3 such
that ∆ < ap1 \ ∆ < ap2 \ ∆ < ap3 \ ∆. Then there is r ≤ p1, p2, p3 satisfying
ξ
ξ = npi
if ξ ∈ api for 1 ≤ i ≤ 3,
• ar = ap1 ∪ ap2 ∪ ap3 ;
• nr
• ir,p1 = IdBXp1
• ir,p2(A) = A + jσ2,1 (A)(Xp1 \ Xp2 ) for all A ∈ BXp2 ,
• ir,p3(A) = A + jσ3,1 (A)(Xp1 \ Xp3 ) for all A ∈ BXp3 ,
In particular
ir,p3(A)ir,p2 (jσ3,2 (A)) = ir,p1(jσ3,1 (A))2
for every A ∈ BXp3 . The element r is called the amalgamation of p1, p2, p3 of type
2.
Proof. First consider s2 ≤ p1, p2 and s3 ≤ p1, p3 which are the including amalga-
mations of p1, p2 and p1, p3 as in Lemma 5.12. It is clear that s1 and s2 are in the
convenient position as witnessed by Idap1 ∪ σ3,2 : ap1 ∪ ap2 → ap1 ∪ ap3 Now let r
be the disjoint amalgamation of s1 and s2 as in Lemma 5.11. The properties of r
follow from Lemma 5.12 and Definition 5.1.
To prove the last statement of the lemma note that ir,p3(A)ir,p2 (jσ3,2 (A)) =(cid:0)A+
jσ3,1 (A)(Xp1 \Xp3)(cid:1)(cid:0)jσ3,2 (A)+jσ3,1 (A)(Xp1 \Xp3 )(cid:1) = (jσ3,1 (A))2 = ir,p1 (jσ3,1 (A))2.
(cid:3)
Lemma 5.17. Suppose that p1, p2, p3 are distinct elements in P which are pairwise
in the convenient position as witnessed by σj,i : api → apj for 1 ≤ i < j ≤ 3 such
that ∆ < ap1 \ ∆ < ap2 \ ∆ < ap3 \ ∆ and npi
ξ = n for some n > 1 and each
i ∈ {1, 2, 3} and that v1 = (v1
n−1), v2 = (v2
0, ...v2
n−1) are two orthogonal unit
vectors of Cn. Then there is r ≤ p1, p2, p3 satisfying
0, ...v1
• ar = ap1 ∪ ap2 ∪ ap3 ;
• nr
•
ξ = npi
ξ = n if ξ ∈ api for 1 ≤ i ≤ 3,
k[ir,pm(A), ir,p1 (jσm,1 (A))]k = 1/2,
for m = 2, 3 and for every nonexpanding A ∈ BXpm such that there is
keξ,k) = 0.
The element r is called the amalgamation of p1, p2, p3 of type 3 for vectors v1and
v2.
ξ ∈ am \ ap1 with A(Pk<n v1
keξ,k and A(Pk<n v2
keξ,k) =Pk<n v1
Proof. First consider s2 ≤ p1, p2 and s3 ≤ p1, p3 which are the (v1, v2)-anti-
commuting amalgamations of p1, p2 and p1, p3 as in Lemma 5.14. It is clear that
s1 and s2 are in the convenient position as witnessed by Idap1 ∪ σ3,2 : ap1 ∪ ap2 →
ap1 ∪ ap3 Now let r be the disjoint amalgamation of s1 and s2 as in Lemma 5.11.
The properties of s1 and s2 from the (v1, v2)-anti-commuting amalgamations s1
and s2 pass to r by Definition 5.1 (d).
(cid:3)
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
25
5.6. Inductive limits of directed families in P. In this section we adopt the
terminology where a directed set is a partial order (X, ≤) where for any two x, y ∈ X
there is z ∈ X such that z ≤ x, y. In this section we will consider inductive limits
AG of systems (Ap
: p ∈ G) where G ⊆ P is a directed subset of P with the order
Xp
≤=≤P. Here for p ≤ q the embeddings ipq : Aq
are given by Definition 5.1
Xq
(c), i.e., they satisfy ipq(Aq
ξ). Formally we
define AG differently in order to work with its convenient representation in Bω1×N
but then, in Lemma 5.22 we prove that the constructed algebra is the corresponding
inductive limit.
ξ,m,n for ξ ∈ aq and n, m ∈ [0, nq
ξ,m,n) = Ap
→ Ap
Xp
Definition 5.18. We say that G ⊆ P is covering if and only if
ω1 × N ⊆[{Xp : p ∈ G}.
Definition 5.19. Suppose that G ⊆ P is directed and covering. Then AG
Bω1×N is given by
ξ,n,m ∈
hAG
ξ,n,m(eη,k), eη,li = hAp
ξ,n,m(eη,k), eη,li
for any (all) p ∈ G such (η, k), (η, l), (ξ, n), (ξ, m) ∈ Xp.
Note that AG
ξ,n,m are well-defined if G is directed and covering. This is because
given two p, p′ ∈ G such that (η, k), (η, l), (ξ, n), (ξ, m) ∈ Xp, Xp′ there is q ≤ p, p′
which implies that Xp, Xp′ ⊆ Xq and so
hAp
ξ,n,m(eη,k), eη,li = hAq
ξ,n,m(eη,k), eη,li = hAp′
ξ,n,m(eη,k), eη,li
by Definition 5.1 (c-d). The following definition is parallel to Definition 5.2:
Definition 5.20. Suppose that G ⊆ P is directed and covering. AG is the subalgebra
of Bω1×N generated by the operators AG
ξ,m,n for all ξ ∈ ω1 and m, n ∈ N.
Let X be a subset of ω1 × N. We define AG
X to be the C∗-subalgebra of AG
ξ,m,n : (ξ, n), (ξ, m) ∈ X). In particular, for every α < ω1, by AG
α
generated by (AG
we mean the C∗-subalgebra of AG generated by {AG
ξ,m,n : ξ < α, m, n ∈ N}.
Lemma 5.21. Suppose that G ⊂ P is directed and covering and p ∈ G. There is a
∗-embedding iG,p : Ap
Xp
Xp such that
→ AG
(1) iG,p(Ap
(2) iG,p(Ap
ξ,m,n) = AG
ξ,m,n)Xp = Ap
ξ,m,n and
ξ,m,n
for every ξ, n, m such that (ξ, n), (ξ, m) ∈ X.
Proof. By Definitions 5.1 and 5.19, a map sending Ap
∗-homomorphism of Ap
Xp
for q ≤ p are null.
ξ,m,n extends to a
Xp . Its kernel must be null as the kernels of iq,p
ξ,m,n to AG
into AG
To prove the second part of the lemma, use the first part and Definition 5.19. (cid:3)
Lemma 5.22. Suppose that G ⊆ P is directed and covering. There is a ∗-isomorphism
j of AG and the inductive limit limp∈GAp
(ip,q : p ≤ q) such that
of the system (Ap
Xp
: p ∈ G) with maps
Xp
for each ξ ∈ ω1 and m, n ∈ N.
j(AG
ξ,n,m) = lim
p∈G
Ap
ξ,n,m
26
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Proof. As in Ex. 1. Chapter 6 of [39] it is enough to prove that for every p, q ∈ G
satisfying p ≤ q the diagram
iG,q
Aq
Xq
ip,q
iG,p
Ap
Xp
AG
Xq
⊆
AG
Xp
commutes. This follows from the fact that by Definition 5.19 we have iG,p(ip,q(Aq
AG
generate Aq
Xq
ξ,n,m) for ξ, m, n such that (ξ, m), (ξ, n) ∈ Xq. But these elements
(cid:3)
ξ,n,m = iG,q(Aq
.
ξ,n,m)) =
Definition 5.23. A family G ⊆ P is called F-rich if and only if G is directed,
covering and FX,α ∩ G 6= ∅ for every finite X ⊆ ω1 × N and α ∈ X, where FX,αs
are defined in Lemma 5.7.
Lemma 5.24. Let G ⊆ P be an F-rich family. Then for every α < ω1 the following
hold:
α is an ideal of AG equal to {A ∈ AG : A[α, ω1) = 0},
(1) AG
(2) there is a *-isomorphism jα : AG/AG
(3) the representation πα : AG[α, ω1) → AG{α} given by πα(A) = A{α} is
α → AG[α, ω1),
faithful.
Proof. As ℓ2({ξ} × N) are AG-invariant, it is clear that sending A ∈ AG to A[α, ω1)
is a ∗-homomorphism. So for (1) and (2) we are left with proving that its kernel is
equal to AG
α.
First note that the kernel contains every generator AG
ξ,n,m for ξ < α and m, n ∈ N
of AG
α and so includes AG
α. This is true by Definition 5.1 (4).
For the other inclusion let A ∈ AG satisfy A[α, ω1) = 0. Since AG is the
s for p ∈ G by Lemma 5.22, for every ε > 0 there is p ∈ G
such that kiG,p(B) − Ak < ε and so kiG,p(B)[α, ω1)k < ε. By Lemma
Xp∩(α×N) ⊆ Ap
, so we can apply iG,p to them.
inductive limit of Ap
Xp
and B ∈ Ap
Xp
5.3, Bα ∈ Ap
and B[α, ω1) ∈ Ap
Xp
By Lemma 5.4 and Definition 5.19 we have that
Xp
kiG,p(B[α, ω1))k = kiG,p(B)[α, ω1)k.
So we have
kA − iG,p(Bα)k = kA − iG,p(B) + iG,p(B[α, ω1))k ≤
≤ kA − iG,p(B)k + kiG,p(B)[α, ω1)k ≤ 2ε.
α since Bα ∈ Ap
α, we conclude that A ∈ AG
But iG,p(Bα) ∈ AG
iG,p(Bα) ∈ AG
Xp∩(α×N). As ε > 0 was arbitrary and
α, completing the proof of (1) and (2).
To prove (3) first note that since ℓ2({α} × ω1) is AG-invariant, it is clear that
Xq for q ∈ G. By
πα is a representation of AG[α, ω1). Now suppose that A ∈ AG
Lemma 5.21 there is B ∈ Aq
such that iG,q(B) = A. Since G is assumed to be
Xq
F-rich, by Lemmas 5.7 and 5.8 there is p ∈ FXq ,α such that p ≤ q. By Lemma
5.4 and Definition 5.19 we have kA[α, ω1)k = kip,q(B)[α, ω1)k. By the fact that
p ∈ FXq ,α we have that
kA{α}k ≥ kip,q(B){α}k ≥ kip,q(B)[α, ω1)k = kA[α, ω1)k.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
27
This shows that πα is an isometry when restricted to Sq∈G AG
dense in AG by Lemma 5.22, and so the representation is faithful.
Xq [α, ω1) which is
(cid:3)
Proposition 5.25. Suppose that G ⊆ P is an F-rich family. Then AG is a scattered
thin-tall fully noncommutative C*-algebra such that
α (AG) = AG
α,
(1) I At
(2) there is a ∗-isomorphism jα : AG/I At
α (AG) → AG[α, ω1) satisfying
jα([A]I At
α (A)) = A[α, ω1),
(3) the collection {[Aα,m,n]I At
α (A) : n, m ∈ N} satisfies the matrix units relations
and generates the essential ideal At(AG/I At
α (AG)).
Proof. By Theorem 1.4 of [18] it is enough to prove (1) - (3) to conclude that A is
a scattered thin-tall fully noncommutative C*-algebra.
The proof of (1) - (3) is by induction on α < ω1. For α = 0 we have that
Iα = {0} and so (1) and (2) are trivial. Also AG
0,n,m = 10,n,m by Definition 5.1, so
these elements satisfy the matrix unit relations. Moreover they generate the algebra
of all compact operators on ℓ2({0} × N) which is an essential ideal in B{0}×N. Since
π0 from Lemma 5.24 is faithfull, the collection {A0,m,n : n, m ∈ N} generates an
essential ideal isomorphic to an algebra of all compact operators on a Hilbert space,
so by Theorem 1.2 (4) of [18] this ideal is I At(AG) as required.
Now suppose we are done for β < α < ω1.
(1) If α is a limit ordinal, then by 1.4 of [18] and the inductive hypothesis we
have
I At
α (A) = [β<α
I At
β (A) = [β<α
Aβ = Aα.
If α = β + 1, then (3) of the inductive hypothesis implies (1).
(2) follows from Lemma 5.24.
(3) Is proved like in the case α = 0.
(cid:3)
6. An operator algebra along a construction scheme
In this section we adopt the terminology and the notation of Section 5. We will
use the constructions scheme of [54] described in Section 2.2 to build appropriate
F-rich families G in the partial order P of approximations whose inductive limit AG
will have interesting properties described in the introduction. To prove the main
theorem of this section we need one more general lemma:
Lemma 6.1. Suppose that A is an AF C*-algebra where {AD : D ∈ D} is a
directed family of finite-dimensional subalgebras with dense union. Let P ∈ A be
a projection. Then for every 0 < ε < 1 there is D ∈ D and a projection Q ∈ AD
such that kQ − P k < ε.
Proof. Let D ∈ D be such that there is A ∈ AD satisfying kA − P k < ε/6. By
considering (A + A∗)/2 instead of A we may assume that A is self-adjoint and
kA − P k < ε/6. As AD is finite dimensional, it is ∗-isomorphic to the direct sum of
full matrix algebras. Let π be the isomorphism. The matrix π(A) is self-adjoint, so
it can be diagonalized. As kA − P k < ε/6 we have that kA2 − Ak < ε/2 and so the
distance of each entry on the diagonal of the diagonalized π(A) from 0 or 1 cannot
28
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
be bigger than ε/2, so there is a projection Q ∈ AD such that kπ(Q) − π(A)k < ε/2
and hence kQ − Ak < ε/2 and kQ − P k < ε as required.
(cid:3)
Theorem 6.2. Suppose that there exists a construction scheme F with allowed
parameters (rk)k∈N and (nk)k∈N, where nk = 3 for each k ∈ N \ {0} and a partition
(Pm)m∈N of N into infinite sets such that for every m ∈ N and every uncountable
∆-system T of finite subsets of ω1 there exist F ∈ F of arbitrarily large rank in Pm
which fully captures a subsystem of T .
Then there is an F-rich family G of elements of P such that the scattered thin-tall
fully noncommutative C*-algebra AG has the following properties:
(1) There is a nondecreasing unbounded sequence (lk)k∈N ⊆ N and a directed
X : X = F × [0, lk), F ∈ Fk, k ∈ N}
family of finite dimensional algebras {AG
whose union B is dense in A such that whenever (Pξ : ξ < ω1) ⊆ B is a
family of projections which generate a nonseparable subalgebra of AG, then
for every ε > 0
(a) there are ξ1 < ξ2 < ξ3 < ω1 such that kPξ1 − Pξ2 Pξ3 k < ε,
(b) there are ξ1 < ξ2 < ω1 such that k[Pξ1 , Pξ2 k < ε,
(c) there are ξ1 < ξ2 < ω1 such that k[Pξ1 , Pξ2 ]k > 1/2 − ε.
(2) AG has no uncountable irredundant subset,
(3) AG has no nonseparable abelian subalgebra.
Proof. Fix an enumeration ((vm, wm) : m ≥ 3), with possible repetitions, of all
pairs of orthogonal complex vectors with finitely many coordinates, all of them
rational, such that vm, wm ∈ Cm for each m ≥ 3 (we abuse notation and identify
Cm′
with a subset of Cm for m′ ≤ m).
We construct the sequence (lk)k∈∈N ⊆ N and G = {pF : F ∈ F} ⊆ P by
induction with respect to k ∈ N such that F ∈ Fk. Moreover, for each k ∈ N we
require that whenever F, F ′ ∈ Fk are such that F \ F ′ < F ′ \ F (cf. Definition 2.10
(2)), then
(*) pF and pF ′ are in the convenient position as witnessed by φF ′,F .
(**) apF = F ,
(***) npF
ξ = lk for all ξ ∈ F and F ∈ Fk and k ∈ N.
For k = 0 we have that F1 = [ω1]1 by Definition 2.10 (1), so we define pF for
F = {ξ} to be the element of P such that
• apF = {ξ},
• nξ
• ApF
pF = l0 = 1,
ξ,0,0 = 1ξ,0,0.
Suppose that we have constructed pF s for all F ∈ Fk′ for k′ ≤ k satisfying (*) -
(***). Now we need to define the pF s for F ∈ Fk+1. Since nk+1 = 3, each F ∈ Fk+1
is the union of the maximal elements G1, G2, G3 of FF which form an increasing
∆-system by Definition 2.10 (3). If k ∈ P1, then we define pF as the amalgamation
of pG1, pG2, pG3 of type 1 from Lemma 5.15. If k ∈ P2, then we define pF as the
amalgamation of pG1, pG2, pG3 of type 2 from Lemma 5.16. If k ∈ Pm for m ≥ 3,
and lk < m, then we define pF as the amalgamation of pG1, pG2, pG3 of type 1
from Lemma 5.15. If k ∈ Pm for m ≥ 3, and lk ≥ m, then we define pF as the
amalgamation of pG1, pG2, pG3 of type 3 for vectors (vm, wm) from Lemma 5.17.
Observe that amalgamation of type 1 increases lk, so lk → ∞ when k → ∞.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
29
First let us note that our inductive hypothesis (*) - (***) is preserved when
we pass from k ∈ N to k + 1. Let F, F ′ ∈ Fk+1 be such that F \ F ′ < F ′ \
F . By Definition 2.10 (2) there is an order preserving bijection φF ′,F : F → F ′
and F ∩ F ′ < F \ F ′ < F ′ \ F .
In particular Definition 2.10 (2) implies that
the maximal elements of FF are sent by φF,′,F onto the maximal elements of
FF ′, on the other hand (3) of 2.10 implies that these maximal elements form the
canonical decomposition consisting of elements in Fk, which in fact are used in
the construction of pF or pF ′. Now to verify Definition 5.9 in order to check (*)
we note that the amalgamations described in Lemmas 5.15, 5.16, 5.17 consist of
constructions of operators which depend only on the place of the involved objects
in F , so Definition 5.9 and (*) are satisfied for pF and pF ′ . (**) and (***) follow
from the descriptions of the amalgamations from Lemmas 5.15, 5.16, 5.17. We have
lk+1 = lk if k ∈ N \ P1 and lk+1 > lk if k ∈ P1 and Definition 2.10 guarantees that
the amalgamations which follow Lemma 5.15 can be done in "the same way" up
pF ′ for any F, F ′ ∈ Fk+1
to the bijection φF ′,F and so obtaining nξ
and ξ ∈ F and ξ′ ∈ F ′. This completes the construction of G = {pF : F ∈ F} and
determines completely the C*-algebra AG as in Definition 5.20.
pF = lk+1 = nξ′
Now note that G is F-rich as in Definition 5.23. First note that pF ′ ≤ pF
whenever F ′ ⊆ F and F, F ′ ∈ F . This can be proved by induction on k ∈ N such
that F ∈ Fk. Note that it is true if F ′ is a maximal element of F F , because then
F ′ is in the canonical decomposition of F by Definition 2.10 (3) and we use pF ′ in
the construction of pF obtaining pF ′ ≤ pF by the Lemmas 5.15, 5.16, 5.17. Now we
proceed with the inductive argument, given F ′ ( F either F ′ is below a maximal
element G of FF or it is one of the maximal elements. The latter case is proved
above and the former follows from the inductive assumption for the pair F ′, G and
from the transitivity of the order in P.
To prove the directedness of G take F, F ′ ∈ F and use the cofinality of F in
[ω1]<ω (Definition 2.10) to find F ′′ ∈ F such taht F ∪ F ′ ⊆ F ′′. By the above
arguments we have pF , pF ′ ≤ pF ′′ .
Now let X = a × [0, l) ∈ [ω1 × N]<ω and α ∈ ω1 and aim at proving further parts
of the F-richness. Consider the ∆-system T = {a ∪ {α, ξ} : max(a ∪ {α}) < ξ < ω1}
of finite subsets of ω1. By the hypothesis there is k ∈ P1 with lk ≥ l and F ∈ F
such that F fully captures a subsystem of T . In particular F = G1 ∪ G2 ∪ G3 for
some G1, G2, G3 ∈ Fk and X ⊆ XpG1 and α ∈ apG1 . By the construction, we do
the amalgamation of type 1 like in Lemma 5.15 while constructing pF and so pF is
,α but this implies that it is in FX,α as required for F-richness in Definition
in FXpG1
5.23.
Proposition 5.25 implies that AG as in Definition 5.20 is a thin-tall fully non-
commutative scattered C*-algebra.
To prove (1) the directed family of finite dimensional subalgebras of AG is {AG
p ∈ G} as in Definition 5.20. By Lemma 5.21 the algebras AG
to the algebras Ap
Xp
Xp :
Xp are *-isomorphic
and they are finite dimensional since they are equal to BXp by
Xp : p ∈ G}.
Suppose that {Pξ : ξ < ω1} ⊆ B is a collection of projections which generate
a nonseparable subalgebra of AG. So, there must be distinct αξ ∈ ω1 such that
Lemma 5.3. Let B =S{AG
Pξ(cid:0){αξ} × N(cid:1) 6= 0. Since B{αξ}×N is invariant for AG it follows that Pξ(cid:0){αξ} × N(cid:1)
is a non-zero projection. Moreover it is not the unit of B{αξ}×N because such a unit
30
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
would produce a unit of AG/I At
impossible because AG is the union of proper ideals I At
αξ (AG) by Lemma 5.24 and Theorem 5.25, which is
α (AG) for α < ω1.
= AG
Fξ×[0,lξ) for each ξ ∈ ω1,
XpFξ
Let Fξ ∈ F be such that αξ ∈ Fξ, Pξ ∈ AG
Let Qξ ∈ A
not the unit of B{αξ}×[0,lξ). This can be obtained from the cofinality of F and the
fact that lk → ∞ when k → ∞.
where lξ = lk for Fξ ∈ Fk and Pξ(cid:0){αξ} × [0, lξ)(cid:1) is a nonzero projection which is
Qξs are projections and Qξ(cid:0){αξ} × [0, lξ)(cid:1) is a nonzero projection which is not the
pFξ
Fξ×[0,lξ) be such that iG,pFξ
unit of B{αξ}×[0,lξ) for each ξ < ω1.
forms an increasing ∆-system of elements of Fk′ for a fixed k′ ∈ N and that
By passing to an uncountable subset, we may assume that T = {Fξ : ξ < ω1}
(Qξ) = Pξ. Note that by Lemma 5.21
hQξ(eη,l), eη,l′ i − hQξ′(eφF
ξ′ ,Fξ (η),l), eφF
ξ′ ,Fξ (η),l′ i < ε/2lk′
for every (η, l), (η, l′) ∈ Fξ × [0, lk′) and every ξ < ξ′ < ω1. This guarantees that
(+)
kjφFξ ,F
ξ′ (Qξ) − Qξ′ k < ε/2
for every ξ < ξ′ < ω1. Now let us prove item (a) of (1). By the hypothesis on
F there is k ∈ P2 bigger than k′ and F ∈ Fk+1 which fully captures T , i.e. the
canonical decomposition of F is {G1, G2, G3} and there are ξ1 < ξ2 < ξ3 < ω1
such that Fξi ⊆ Gi and φGj ,Gi[Fξi ] = Fξj for all 1 ≤ i, j ≤ 3. As φGj ,Gi are order
preserving, they must agree with φFξj ,Fξi
on Fξi , so (+) implies that
kjφG3,Gi
(Qξ3) − Qξik < ε/2
holds for i = 1, 2. Since we use amalgamation of type 2 at the construction of pF
for k ∈ P2 by Lemma 5.16 we have
ipF ,pG3 (Qξ3 )ipF ,pG2 (jφG3 ,G2 (Qξ3 )) = ipF ,pG1 (jφG3 ,G1 (Qξ3 ))2,
and so
kipF ,pG3 (Qξ3 )ipF ,pG2 (Qξ2 ) − ipF ,pG1 (Qξ1 )2k < ε
and hence kPξ3 Pξ2 − Pξ1 k < ε since
(Qξi) = iG,pGξi
iG,pF ◦ ipF ,pGξi
by Definition 5.19 and Lemma 5.21. This completes the proof of (a) of (1). Item
(b) follows from (a) for ε/2 and by taking the adjoints.
(Qξi )) = iG,pFξi
(Qξi ) = iG,pGξi
(Qξi ) = Pξi
(ipGξi
,pFξi
Now let us prove item (c) of (1). For ξ < ω1 let Q′ξ ∈ BFξ×[0,lk′ ) be such
projections that kQξ − Q′ξk < ε/8 and there is on orthonormal basis in B{αξ}×[0,lk′ )
of eigenvectors for Q′ξ consisting only of vectors with all rational coordinates with
respect to our canonical basis (eαξ,l : 0 ≤ l < lk′ ). Note that by Lemma 5.3 we
have that Q′ξ ∈ A
not the unit of B{αξ}×[0,lk′ ) for each ξ < ω1, Q′ξ may be assumed to have the same
rank as Qξ and so there are orthogonal unit vectors vξ, wξ ∈ Clk′ with all rational
coordinates such that
. Since Qξ(cid:0){αξ} × [0, lk′ )(cid:1) is a nonzero projection which is
pFξ
XpFξ
Q′ξ(Xl<lk′
vξ
l eαξ,l) = Xl<lk′
vξ
l eαξ,l, Q′ξ(Xl<lk′
wξ
l eαξ,l) = 0.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
31
As there are only countably many such vectors we may assume that all of them are
equal to a pair (v, w) and moreover that
(++)
kjφFξ ,F
ξ′ (Q′ξ) − Q′ξ′ k < ε/4
for every ξ < ξ′ < ω1.
By the hypothesis on F there is k ∈ Pm bigger than k′ such that vm = v =
(v1, ...vlk′ ) and wm = w = (w1, ...wlk′ ) and there is F ∈ Fk+1 which fully captures
T , i.e., the canonical decomposition of F is {G1, G2, G3} and there are ξ1 < ξ2 <
ξ3 < ω1 such that Fξi ⊆ Gi and φGj ,Gi[Fξi ] = Fξj for all 1 ≤ i, j ≤ 3. Note that
αξis are not in the root of {G1, G2, G3} as they are not in the root of Fξs. As
φGj ,Gi are order preserving, they must agree with φFξj ,Fξi
on Fξi , so (++) implies
that
kjφG1,Gi
(Q′ξ1) − Q′ξik < ε/4
holds for i = 2, 3. Since we use amalgamation of type 3 at the construction of pF
for k ∈ Pm by Lemma 5.17 we have
k[ipF ,pG1 (Q′ξ1 ), ipF ,pG2 (jφG1 ,G2 (Q′ξ1 ))]k = 1/2
k[ipF ,pG1 (Q′ξ1 ), ipF ,pG2 (Q′ξ2 )]k ≥ 1/2 − ε/2
and so
and hence
as kQξ − Q′ξ′ k < ε/8 for each ξ < ξ′ < ω1, and finally
k[ipF ,pG1 (Qξ1 ), ipF ,pG2 (Qξ2)]k ≥ 1/2 − ε
k[Pξ1 , Pξ2 ]k ≥ 1/2 − ε
since
iG,pF ◦ ipF ,pGξi
by Definition 5.19 and Lemma 5.21. This completes the proof of (c) of (1).
(Qξi )) = iG,pFξi
(Qξi) = iG,pGξi
(Qξi ) = Pξi
(Qξi ) = iG,pGξi
(ipGξi
,pFξi
The proof of (2) will be based on (1) (a) and Lemma 6.1. Suppose that AG
contains an uncountable irredundant set {Qξ : ξ < ω1}. By Lemma 3.3 we may
assume that all Qξs are projections. For each ξ let Aω1\{ξ} be the C*-subalgebra of
AG generated by the set {Qη : η ∈ ω1 \ {ξ}}. By passing to an uncountable subset
we may assume that there is ε > 0 such that for each ξ < ω1 we have kA − Qξk ≥ ε
for each A ∈ Aω1\{ξ}. Let Pξ ∈ B be a projection satisfying kPξ − Qξk < ε/4 which
is obtained using Lemma 6.1. By (1) (a) there are ξ1 < ξ2 < ξ3 < ω1 such that
kPξ1 − Pξ2 Pξ3 k < ε/4. This implies that kQξ1 − Qξ2Qξ3 k < ε which contradicts the
defining property of ε and completes the proof of (2).
The proof of (3) will be based on (1) (c) and Lemma 6.1. Suppose that AG
contains a nonseparable abelian subalgebra. As subalgebras of scattered algebras
are scattered, and scattered locally compact spaces are totally disconnected, it fol-
lows that AG contains an uncountable Boolean algebra of (commuting) projections
{Qξ : ξ < ω1}. In particular kQξ − Qξ′k = 1 for all ξ < ξ′ < ω1.
Let Pξ ∈ B for ξ < ω1 be projections satisfying kPξ − Qξk < 1/10 for each
ξ < ω1 which is obtained using Lemma 6.1. In particular kPξ − Pξ′ k ≥ 8/10 for all
ξ < ξ′ < ω1 and so they generate a nonseparable C*-algebra.
We have kPξ1 Pξ2 − Qξ1Qξ2 k < 1/5 and kPξ2Pξ1 − Qξ2Qξ1k < 1/5 for each
ξ1 < ξ2 < ω1, so [Pξ1 , Pξ2 ] < 2/5 for each ξ1 < ξ2 < ω1. But by (1) (c) there are
ξ1 < ξ2 < ω1 such that k[Pξ1 , Pξ2]k ≥ 2/5, a contradiction.
32
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
(cid:3)
References
1. U. Abraham, M. Rubin, S. Shelah, On the consistency of some partition theorems for con-
tinuous colorings, and the structure of ℵ1-dense real order types. Ann. Pure Appl. Logic 29
(1985), no. 2, 123 -- 206.
2. C. Akemann, Left ideal structure of C*-algebras. J. Functional Analysis 6 1970, 305 -- 317.
3. C. Akemann, N. Weaver, Consistency of a counterexample to Naimark's problem. Proc. Natl.
Acad. Sci. USA 101 (2004), no. 20, 7522 -- 7525.
4. C. Akemann, N. Weaver, B(H) has a pure state that is not multiplicative on any masa. Proc.
Natl. Acad. Sci. USA 105 (2008), no. 14, 5313 -- 5314.
5. W. Arveson, An invitation to C ∗-algebras, Graduate Texts in Mathematics, No. 39. Springer-
Verlag, New York-Heidelberg, 1976.
6. M. Bell, J. Ginsburg, S. Todorcevic, Countable spread of expY and λY . Topology Appl. 14
(1982), no. 1, 1 -- 12.
7. T. Bice, P. Koszmider, A note on the Akemann-Doner and Farah-Wofsey constructions, Proc.
Amer. Math. Soc. 145 (2017), no. 2, 681 -- 687.
8. C. Brech, P. Koszmider, Thin-very tall compact scattered spaces which are hereditarily sepa-
rable. Trans. Amer. Math. Soc. 363 (2011), no. 1, 501 -- 519.
9. C. Brech, P. Koszmider, On biorthogonal systems whose functionals are finitely supported.
Fund. Math. 213 (2011), no. 1, 43 -- 66.
10. G. Carotenuto, An introduction to OCA - 2014, notes on lectures by Matteo Viale.
http://www.logicatorino.altervista.org/matteo viale/OCA.pdf
11. K. Davidson, C*-algebras by example. Fields Institute Monographs, 6. American Mathematical
Society, Providence, RI, 1996.
12. M. Dzamonja, I. Juhasz, CH, a problem of Rolewicz and bidiscrete systems, Topol. Appl. 158
(18) (2011) 2458 -- 2494.
13. P. Enflo, H. Rosenthal, Some results concerning Lp(µ)-spaces. J. Functional Analysis 14
(1973), 325 -- 348.
14. R. Engelking, General topology. Translated from the Polish by the author. Second edition.
Sigma Series in Pure Mathematics, 6. Heldermann Verlag, Berlin, 1989.
15. I. Farah, Analytic quotients: theory of liftings for quotients over analytic ideals on the inte-
gers. Mem. Amer. Math. Soc. 148 (2000), no. 702.
16. I. Farah, T. Katsura, Nonseparable UHF algebras I: Dixmier's problem. Adv. Math. 225
(2010), no. 3, 1399 -- 1430.
17. I. Farah, I. Hirshberg, Simple nuclear C*-algebras not isomorphic to their opposites. Proc.
Natl. Acad. Sci. USA 114 (2017), no. 24, 6244 -- 6249.
18. S. Ghasemi, P Koszmider, Noncommutative Cantor-Bendixson derivatives and scattered C*-
algebras, Topology Appl. 240 (2018), 183 -- 209.
19. S. Ghasemi, P. Koszmider, A non-stable C*-algebra with an elementary essential composition
series, arXiv:1712.02090
20. P. Hajek, V. Montesinos Santalucia, J. Vanderwerff, V. Zizler, Biorthogonal systems in Banach
spaces. CMS Books in Mathematics/Ouvrages de Mathematiques de la SMC, 26. Springer,
New York, 2008.
21. R. Hodel, Cardinal functions. I. Handbook of set-theoretic topology, 1 -- 61, North-Holland,
Amsterdam, 1984.
22. K. Hofmann, K. Neeb, Epimorphisms of C*-algebras are surjective. Arch. Math. (Basel) 65
(1995), no. 2, 134 -- 137.
23. L. Heindorf, A note on irredundant sets. Algebra Universalis 26 (1989), no. 2, 216 -- 221.
24. C. Hida, Two cardinal inequalities about bidiscrete systems. Topology Appl. 212 (2016), 71 -- 80.
25. T. Jech, Set theory. The third millennium edition, revised and expanded. Springer Monographs
in Mathematics. Springer-Verlag, Berlin, 2003.
26. H. Jensen, Scattered C*-algebras. Math. Scand. 41 (1977), no. 2, 308 -- 314.
27. H. Jensen, Scattered C*-algebras. II. Math. Scand. 43 (1978), no. 2, 308 -- 310 (1979).
28. S. Koppelberg, Handbook of Boolean algebras. Vol. 1. Edited by J. Donald Monk and Robert
Bonnet. North-Holland Publishing Co., Amsterdam, 1989.
LARGE IRREDUNDANT SETS IN OPERATOR ALGEBRAS
33
29. P. Koszmider, On a problem of Rolewicz about Banach spaces that admit support sets. J.
Funct. Anal. 257 (2009), no. 9, 2723 -- 2741.
30. P. Koszmider, Some topological invariants and biorthogonal systems in Banach spaces, Extr.
Math. 26 (2) (2011) 271 -- 294.
31. P. Koszmider, On the problem of compact totally disconnected reflection of nonmetrizability.
Topology Appl. 213 (2016), 154 -- 166.
32. P. Koszmider, On constructions with 2-cardinals. Arch. Math. Logic 56 (2017), no. 7-8, 849 --
876.
33. K. Kunen, An introduction to independence proofs. Studies in Logic and the Foundations of
Mathematics, 102. North-Holland Publishing Co., Amsterdam-New York, 1980.
34. M. Kusuda, C*-algebras in which every C*-subalgebra is AF. Q. J. Math. 63 (2012), no. 3,
675 -- 680.
35. H. X. Lin, The structure of quasimultipliers of C*-algebras, Trans. Amer. Math. Soc. 315
(1989), no. 1, 147 -- 172.
36. F. Lopez, S. Todorcevic, Trees and gaps from a construction scheme. Proc. Amer. Math. Soc.
145 (2017), no. 2, 871 -- 879.
37. F. Lopez, Banach spaces from a construction scheme. J. Math. Anal. Appl. 446 (2017), no.
1, 426 -- 435.
38. A. Mostowski, A. Tarski, Booleshe Ringe mit ordneter basis. Fund. Math. 32, 69 -- 86.
39. G. Murphy, C*-algebras and operator theory. Academic Press, Inc., Boston, MA, 1990.
40. S. Negrepontis, Banach spaces and topology. Handbook of set-theoretic topology, 1045 -- 1142,
North-Holland, Amsterdam, 1984.
41. T. Ogasawara, Finite-dimensionality of certain Banach algebras. J. Sci. Hiroshima Univ. Ser.
A. 17, (1954). 359 -- 364.
42. C. Olsen, W. Zame, Some C*-alegebras with a single generator. Trans. Amer. Math. Soc. 215
(1976), 205 -- 217.
43. A. Ostaszewski, On countably compact, perfectly normal spaces. J. London Math. Soc. (2) 14
(1976), no. 3, 505 -- 516.
44. A. Pe lczy´nski, Z. Semadeni, Spaces of continuous functions. III. Spaces C(Ω) for ω without
perfect subsets. Studia Math. 18 1959 211 -- 222.
45. S. Popa, Orthogonal pairs of ∗-subalgebras in finite von Neumann algebras. J. Operator The-
ory 9 (1983), no. 2, 253 -- 268.
46. M. Rubin, A Boolean algebra with few subalgebras, interval Boolean algebras and retractive-
ness. Trans. Amer. Math. Soc. 278 (1983), no. 1, 65 -- 89.
47. J.G. Stampfli, The norm of a derivation. Pacific J. Math. 33, 1970, 737 -- 747.
48. H. Thiel, The generator rank for C*-algebras. arXiv:1210.6608
49. H. Thiel, W. Winter, The generator problem for Z-stable C*-algebras. Trans. Amer. Math.
Soc. 366 (2014), no. 5, 2327-2343.
50. J. Tomiyama, A characterization of C*-algebras whose conjugate spaces are separable. Tohoku
Mathematical Journal, Second Series 15.1 (1963): 96-102.
51. S. Todorcevic, Partition problems in topology. Contemporary Mathematics, 84. American
Mathematical Society, Providence, RI, 1989.
52. S. Todorcevic, Irredundant sets in Boolean algebras, Trans. Am. Math. Soc. 339 (1) (1993)
35 -- 44.
53. S. Todorcevic, Biorthogonal systems and quotient spaces via Baire category methods, Math.
Ann. 335 (3) (2006), 687 -- 715.
54. S. Todorcevic, A construction scheme for non-separable structures. Adv. Math. 313 (2017),
564 -- 589.
55. D. Velleman, ω-morasses, and a weak form of Martin's axiom provable in ZFC. Trans. Am.
Math. Soc. 285 (2), 617 -- 627 (1984)
56. P. Wojtaszczyk, On linear properties of separable conjugate spaces of C ∗-algebras. Studia
Math. 52 (1974), 143 -- 147.
34
CLAYTON SUGUIO HIDA AND PIOTR KOSZMIDER
Departamento de Matem´atica, Instituto de Matem´atica e Estat´ıstica, Universidade
de Sao Paulo, Caixa Postal 66281, 05314-970, Sao Paulo, Brazil
E-mail address: [email protected]
Institute of Mathematics of the Polish Academy of Sciences, ul. ´Sniadeckich 8,
00-656 Warszawa, Poland
E-mail address: [email protected]
|
1709.04873 | 2 | 1709 | 2018-01-23T09:18:37 | Convolution semigroups on locally compact quantum groups and noncommutative Dirichlet forms | [
"math.OA",
"math.FA",
"math.PR",
"math.QA"
] | The subject of this paper is the study of convolution semigroups of states on a locally compact quantum group, generalising classical families of distributions of a L\'{e}vy process on a locally compact group. In particular a definitive one-to-one correspondence between symmetric convolution semigroups of states and noncommutative Dirichlet forms satisfying the natural translation invariance property is established, extending earlier partial results and providing a powerful tool to analyse such semigroups. This is then applied to provide new characterisations of the Haagerup Property and Property (T) for locally compact quantum groups, and some examples are presented. The proofs of the main theorems require developing certain general results concerning Haagerup's $L^{p}$-spaces. | math.OA | math | CONVOLUTION SEMIGROUPS ON LOCALLY COMPACT QUANTUM GROUPS AND
NONCOMMUTATIVE DIRICHLET FORMS
ADAM SKALSKI AND AMI VISELTER
ABSTRACT. The subject of this paper is the study of convolution semigroups of states on a locally
compact quantum group, generalising classical families of distributions of a Lévy process on a locally
compact group. In particular a definitive one-to-one correspondence between symmetric convolution
semigroups of states and noncommutative Dirichlet forms satisfying the natural translation invariance
property is established, extending earlier partial results and providing a powerful tool to analyse such
semigroups. This is then applied to provide new characterisations of the Haagerup Property and Prop-
erty (T) for locally compact quantum groups, and some examples are presented. The proofs of the
main theorems require developing certain general results concerning Haagerup's Lp-spaces.
8
1
0
2
n
a
J
3
2
]
.
A
O
h
t
a
m
[
2
v
3
7
8
4
0
.
9
0
7
1
:
v
i
X
r
a
INTRODUCTION
The connection between the convolution operation, defined on probability measures on R, and
independence of random variables is one of the key elementary features of classical probability
theory. In particular convolution semigroups of probability measures are precisely families of distri-
butions of R-valued stochastic processes with independent and identically distributed increments,
i.e. of Lévy processes. It is not too difficult to see that defining the convolution operation requires
only that the underlying space admits a semigroup structure; in fact the natural setup for studying
Lévy processes, on one hand sufficiently rich to allow plenty of important examples, and on the
other hand sufficiently specific to facilitate the application of strong functional-analytic methods, is
that of a locally compact group, denoted henceforth by G. Here convolution semigroups of measures
generate convolution semigroups of operators, acting either on the algebra of continuous functions
vanishing at infinity, C0(G), on the von Neumann algebra of essentially bounded measurable func-
tions, L∞(G), or on the scale of the Lp(G)-spaces. Among recent monographs describing (some of)
this vast area of study we recommend [App] and [Lia]. The convolution semigroups of operators
associated to a Lévy process form a specific subclass of Markov semigroups [BLSC]. The latter are
often studied via their generators; a related key tool, which will play a very important role in this
paper, is that of Dirichlet forms – quadratic forms on the appropriate L2-space satisfying the specific
conditions identified by Beurling and Deny (see [FOT] and references therein).
The development of mathematical approaches to quantum mechanics using the language of op-
erator algebras, dating back to von Neumann, has led to the study of quantum Markov semigroups,
understood as semigroups of completely positive unital maps on a von Neumann algebra equipped
with a reference state or weight and representing quantum stochastic evolutions of open systems
2010 Mathematics Subject Classification. Primary: 46L65, Secondary: 46L30, 46L53, 46L57, 47B38, 47D07.
Key words and phrases. locally compact quantum group; noncommutative Dirichlet form; convolution operator; convolu-
tion semigroup.
1
2
ADAM SKALSKI AND AMI VISELTER
(see for example [Dav1, Mey]). These, as their classical counterparts, are studied via their gener-
ators, and once again the Dirichlet forms become an indispensable tool. In the noncommutative
context Dirichlet forms were first studied by Albeverio and Høegh-Krohn, then by Davies and Lind-
say in the case where the reference state is tracial, and later by Goldstein and Lindsay [GL1] and
independently by Cipriani [Cip1] for general non-tracial states. Finally in [GL2] noncommutative
Dirichlet forms were investigated in the most general context, in which the role of the reference
'measure' was played by an arbitrary normal semi-finite faithful weight. For the history, motivations
behind the introduction of quantum Dirichlet forms and several examples we refer to the survey
[Cip2]. It has to be noted that the passage from the tracial to the non-tracial case vastly increases
the technical complexity of the problem, as for example one needs to consider Haagerup's Lp-spaces
[Ter1, Ter2] instead of the 'tracial' Lp-spaces of Nelson [PiX]. In the non-tracial context we may also
consider several different natural ways of passing from the maps at the von Neumann algebra level
to the maps on the corresponding L2-space, with the two most prominent ones being the so-called
GNS- and KMS-implementations, see for example [CaS].
In view of the first paragraph above it is natural to ask about the convolution structures which
might play an important role also in the quantum context. This was recognised relatively early in
the development of the theory of quantum Lévy process initiated by Schürmann (see [Sch, Fra]),
in which the underlying quantum probability space was represented by a Hopf ∗-algebra, or more
generally a ∗-bialgebra. The huge success of the theory of compact quantum groups due to Woro-
nowicz [Wor1, Wor2], and more generally locally compact quantum groups due to Kustermans and
Vaes [KV1], with the latter objects denoted below by G, has opened a possibility to study analogous
problems in the much richer analytic setting. Here, in parallel to the situation described in the
previous paragraph, the level of technical difficulty involved in the locally compact theory by far ex-
ceeds that of the compact case, where many questions can be still investigated via purely algebraic
means, exploiting the Hopf ∗-algebra Pol(G). Kustermans's and Vaes's theory admits perfect duality,
with G denoting the dual of G. Discrete quantum groups are duals of compact ones, which means
that also in the discrete case one can take advantage of certain algebraic techniques. Recall here
only that quantum groups are in fact studied indirectly, via associated algebras of functions, such as
C0(G), Cu
0(G) or L∞(G) (see [KV1, KV2, Kus2]).
Recent years brought a significant interest in convolution operators on arbitrary locally compact
quantum groups: these featured prominently in the work of Junge, Neufang, Ruan, Daws, Hu and
others (see e.g. [JNR] and [Daw2]). It should be noted that the quantum group context allows one
to treat the usual convolution operators and Herz–Schur multipliers on a locally compact group G
within the same framework, with the latter viewed as convolution operators on the dual quantum
group G. Perhaps slightly surprisingly, there has been less work on the convolution semigroups of
states on quantum groups beyond the compact case: some initial facts were established in [LS1]
(where in particular such semigroups were shown to be determined by densely defined generating
functionals) and then applied in [LS2] to study the analytic theory of quantum Lévy processes. On
the other hand, in an important recent paper [CFK], Cipriani, Franz and Kula continued the study
of convolution semigroups of states on compact quantum groups, earlier investigated in the context
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
3
of algebraic quantum Lévy processes, for the first time involving in the analysis the Dirichlet form
techniques.
In this work we undertake a deeper study of the convolution semigroups associated with a locally
compact quantum group G. As in the classical case these turn out to have several avatars: semig-
roups of states of Cu
0(G), or semigroups of completely positive operators on each of the algebras
0(G), C0(G), L∞(G), etc. Here we show further that every convolution operator on L∞(G) leads,
Cu
in two different ways (via the KMS- and GNS-implementations mentioned above) to a bounded
operator on L2(G). In presence of the natural symmetry (again, visible on all the levels at which
we can view the respective convolution operator or convolution semigroup) we use the theory of
noncommutative Dirichlet forms for weights to obtain the central result of the article (Theorem 3.4
in the main body of the paper).
Theorem 0.1. There exist 1 − 1 correspondences between the following objects:
• w∗-continuous convolution semigroups of states of Cu
0(G) invariant under the unitary antipode;
• C∗0 -semigroups of normal, unital, completely positive maps on L∞(G) that satisfy the intertwin-
ing relation with the co-product and are KMS-symmetric with respect to the left Haar weight of
G;
• C0-semigroups of selfadjoint completely Markov operators on L2(G) (with respect to the left
Haar weight of G) that belong to L∞( G);
• completely Dirichlet forms (with respect to the left Haar weight of G) on L2(G) that are invari-
ant under U (L∞( G)′), modulo multiplication of forms by a positive number.
All the notions featuring in the above result will be made precise in the remainder of the paper.
Here note only that the special property of the Dirichlet forms in the last point above can be naturally
interpreted as translational invariance, bringing us back to the classical context studied in [FOT,
Sections 1.4 and 1.5] and in [Den, Section 6].
The theorem above provides a definitive extension of the results of [CFK]. The technical diffi-
culties in the locally compact case are far more daunting: the generators of the respective semig-
roups admit no 'obvious' domain and one needs to deal with weights as opposed to states. The
key step of the proof uses ideas of Goldstein and Lindsay; however as the proof of one of the main
results in [GL2] appears to be incorrect, we have to provide a new argument to obtain the passage
from the Dirichlet form to the semigroup of operators acting on the von Neumann algebra, requir-
ing slightly stronger assumptions than those stated in [GL2]. Nevertheless, these are automatically
satisfied in our quantum group context. We believe this general result is of independent interest.
Returning to the quantum group framework, we furthermore present a very explicit connection
between the relevant noncommutative Dirichlet form and the respective generating functional as
studied in [LS1].
The above theorem and its technical extensions turn out to have significant applications to geo-
metric quantum group theory.
In particular we deduce that a second countable locally compact
quantum group G does not have Kazhdan's Property (T) (see [DSV] and references therein) if and
only if there exists a symmetric (in the sense of the first bullet point in Theorem 0.1) convolution
0( G) with an unbounded (equivalently, not everywhere defined) generating
semigroup of states of Cu
functional. On the other hand G has the Haagerup Property if and only if there exists a semigroup
4
ADAM SKALSKI AND AMI VISELTER
of Markov convolution operators on L2(G) (in the sense of the third bullet point of the Theorem
above) belonging to C0(G). Classically the passage between the representation theoretic aspects of
properties such as Property (T) and their geometric guises goes through the Schönberg correspond-
ence: suitable positive-definite functions are first assembled into a semigroup, whose generator is
a conditionally negative-definite function ψ, and then one identifies the desired affine action of the
group on a Hilbert space with the cocycle associated with ψ; the converse direction requires Schön-
berg's result. For discrete quantum groups one no longer has a direct notion of affine actions, and
works only with the respective cocycles – understood as certain derivations of the underlying Hopf
algebra (instances of such techniques can be found in [Kye] and [DFSW]). We emphasise that this
approach relies crucially on the algebraic nature of discrete quantum groups, which is absent from
the general locally compact case. Thus in this paper we provide a yet different perspective, encom-
passing all the results mentioned above: namely, we encode the representation theoretic properties
of a locally compact quantum group via the existence (or non-existence) of convolution semigroups
with particular properties.
Also the construction of examples in the locally compact setting is far more complicated than it
was in [CFK], once again mainly due to the lack of a canonical 'purely algebraic' domain for the
generators. We describe in detail the dual to the classical case, and for the classical case refer again
to [FOT, Sections 1.4 and 1.5] and [Den, Section 6]. We further show how to use cocycle twisting
to generate interesting examples acting on genuine locally compact quantum groups, e.g. on the
quantised Heisenberg group of Enock and Vainerman [EnV].
Our work opens several future directions of research. It is natural to ask about the possibility
of extending the key construction of derivations out of quantum Dirichlet forms due to Cipriani
and Sauvageot [CiS] beyond the tracial case, exploiting the additional quantum group structure.
We intend also to study further examples of non-trivial convolution semigroups using the Rieffel
deformation techniques of [Kas], in a sense dual to the procedure of cocycle twisting discussed in
this paper. Finally one might investigate how the properties of the noncommutative Dirichlet form
affect the long-term behaviour of the quantum Markov semigroup in question both in the quant-
itative and the qualitative senses, and exploit the Dirichlet form techniques to produce interesting
perturbations of the objects studied here (perhaps landing outside of the class of the convolution
semigroups).
The concrete plan of the paper is as follows: in Section 1 we discuss preliminaries concerning
von Neumann algebras, the associated Haagerup Lp-spaces, and noncommutative Dirichlet forms,
prove several technical results related to the Lp-embeddings introduced in [GL2] and introduce
notation and terminology concerning locally compact quantum groups. Section 2 treats convolu-
tion operators associated with quantum groups, first recalling known results, and then focusing on
the existence and properties of L2(G)-implementations and equivalences between various modes
of convergence. A short Section 3 contains the main general theorems of the paper, in particu-
lar Theorem 0.1, and connects the noncommutative Dirichlet form associated with a convolution
semigroup with the generating functional of [LS1]. Section 4 provides applications to geometric
quantum group theory, first concerning Property (T) and then the Haagerup Property, and Section
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
5
5 discusses several examples. Finally in the Appendix we provide a correct proof of [GL2, Theorem
4.7] under two different sets of mild additional assumptions.
Acknowledgements. The first author was partially supported by the National Science Centre (NCN)
grant no. 2014/14/E/ST1/00525. We thank Martin Lindsay and Stanisław Goldstein for encour-
agement, and Stuart White and the anonymous referee for valuable comments. During the last
stage of the preparation of this paper the second author was visiting Sutanu Roy at the School of
Mathematical Sciences of NISER, Bhubaneswar, India. He is grateful to him and to his colleagues
for their warm hospitality.
1. PRELIMINARIES
We begin with some basic notation and conventions. For a Banach space X, let B(X) stand for the
algebra of bounded operators on X. A (one-parameter) operator semigroup on X is a family (Tt)t≥0
in B(X) satisfying T (0) = I and the semigroup identity: T (t + s) = T (t)T (s) for all t, s ≥ 0. We
say that (Tt)t≥0 is a C0-semigroup if it is continuous at 0+ in the strong operator topology, namely
limt→0+ T (t)x = x for every x ∈ X [Dav2, EnN, Kan]. We will frequently use the fact that this
continuity condition is equivalent to continuity at 0+ in the weak operator topology, namely that
limt→0+ T (t)x = x weakly for every x ∈ X [Dav2, Proposition 1.23].
For an algebra A we denote by id : A → A the identity map and by 1 the unit of A, if it exists. For
subsets X, Y ⊆ A we write XY := span{xy : x ∈ X, y ∈ Y }.
Let A be a C∗-algebra. The multiplier algebra of A is denoted by M(A). For ω ∈ A∗ we define
ω ∈ A∗ by ω(x) := ω(x∗), x ∈ A. Representations of C∗-algebras are always assumed nondegenerate.
We use ⊗,⊗min,⊗ for the algebraic or Hilbert space tensor product, the minimal tensor product
of C∗-algebras, and the normal tensor product of von Neumann algebras, respectively. We write σ
for the flip map at the C∗- or the von Neumann algebra level.
For a Hilbert space H, write K(H) for the C∗-algebra of compact operators on H and U (H) for
the group of all unitary operators on H. Given u ∈ U (H), set Ad(u)(x) := uxu∗ for x ∈ B(H).
Inner products are linear in the left variable. For vectors ζ, η ∈ H, ωζ,η ∈ B(H)∗ is the functional
given by ωζ,η(x) := hxζ, ηi, x ∈ B(H). We let ωζ := ωζ,ζ. The ultraweak operator topology,
resp. ultrastrong operator topology, on a von Neumann algebra will be called simply the ultraweak
topology, resp. ultrastrong topology. An automorphism group of a C∗- or von Neumann algebra is
assumed by definition to be continuous with respect to the suitable topology, namely point–norm
or point–ultraweak, respectively.
Subsections 1.1–1.3 discuss preliminaries on von Neumann algebras and in particular on the
Haagerup Lp-spaces. Subsection 1.4 introduces locally compact quantum groups.
1.1. The Haagerup noncommutative Lp-spaces of von Neumann algebras and the maps j(q)
and i(p). In this subsection we introduce Haagerup's construction of Lp-spaces of von Neumann
algebras as announced in [Haa4] and described in full detail by Terp [Ter1], and the subsequent
work of Goldstein and Lindsay [GL2, Sections 1–2]. We assume that the reader is familiar with
modular theory [Str, Tak1, Tak2].
6
ADAM SKALSKI AND AMI VISELTER
We will rely heavily on the theory of unbounded operators on Hilbert spaces [DuS, Chapter XII].
For linear operators a, b on a Hilbert space H, not necessarily densely defined or closable, denote
by D(a) the domain of a, and write a + b and ab for the sum, respectively product, of a and b with
maximal domains. If a is densely defined, we denote its adjoint by a∗; if a is closable, we denote its
closure by a.
Fix a von Neumann algebra A acting, not necessarily standardly, on a Hilbert space H. Recall
that a (not necessarily densely defined or closable) linear operator a on H is affiliated with A if
y′a ⊆ ay′ for every y′ ∈ A′.
Fix a normal, semi-finite, faithful (n.s.f.) weight ϕ on A. Write σϕ = (σϕ
t )t∈R for the modular
automorphism group of ϕ, let
Nϕ := {a ∈ A : ϕ(a∗a) < ∞} ,
Mϕ := span{a ∈ A+ : ϕ(a) < ∞} = N ∗ϕNϕ,
let Tϕ denote the Tomita ∗-algebra of all elements a ∈ A that are entire analytic with respect to σϕ
and satisfy σϕ
z (a) ∈ Nϕ∩N ∗ϕ for every z ∈ C, and let Mϕ,∞ := T ∗ϕ Tϕ = TϕTϕ. The GNS construction
for the pair (A, ϕ) gives a Hilbert space L2(A, ϕ) and a map ηϕ : Nϕ → L2(A, ϕ). When viewing A
as acting on L2(A, ϕ) we do not use any additional symbol for the representation. Write ∇ϕ and Jϕ
for the modular operator and modular conjugation of ϕ, respectively, both acting on L2(A, ϕ). In
the sequel we usually omit ϕ from the notation.
We begin with an approximation lemma, providing a 'good' net of elements in the Tomita algebra
converging to 1 in the ∗-strong operator topology.
Lemma 1.1. For each δ > 0 there exists a net (ej)j∈J
in T such that for every z ∈ C we have:
(a) kσz(ej )k ≤ eδ(ℑz)2
(b) σz(ej) −−−→j∈J
for all j ∈ J ;
1 in the ∗-strong operator topology.
Proof. Combine [Ter2, Lemma 9] and [Str, Proposition 2.16], whose constructions agree: the former
proves the existence of a net satisfying (a), as well as (b) for z = 0, while the latter treats (a) and (b)
for δ = 1. Unfortunately, in proving (b), [Str] uses the Lebesgue dominated convergence theorem,
which is applicable only when one can choose in the proof sequences instead of nets (e.g., when A∗
is separable). This can be fixed via the following argument (see also the proof of [Kus1, Proposition
2.25]). Assume that A is in standard form in H and fix z ∈ C, ζ ∈ H. Write a := ℜz, b := ℑz. Given
σt(f ) dt (convergence in the ∗-strong operator topology).
f ∈ A with kfk ≤ 1 let e :=
Then we have √π√δ kσz(e)ζ − ζk ≤ eδb2RR e−δ(t−a)2 kσt(f )ζ − ζk dt. For ε > 0 one can approximate
For every t ∈ R we have kσt(f )ζ − ζk = (cid:13)(cid:13)f∇−itζ − ∇−itζ(cid:13)(cid:13) because σt is implemented by ∇it. By
uniform continuity of t 7→ ∇−itζ on I one can approximate the integral on I up to ε by Riemann
sums, again independently of f . This implies that if (fj)j∈J
in the closed unit ball of A converges
in the strong operator topology to 1 then the associated (ej)j∈J
the last integral up to ε by the integral on some (bounded) interval I in R independently of f .
√δ√πRR e−δt2
will satisfy σz(ej )ζ −−−→j∈J
The following material is taken from [Nel, Ter1] unless indicated otherwise. Let A be a semi-finite
von Neumann algebra acting, not necessarily standardly, on a Hilbert space K, and τ an n.s.f. trace
on A. A linear subspace E ⊆ K is called τ-dense if for every δ > 0 there exists a projection p ∈ A
ζ.
(cid:3)
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
7
such that Im p ⊆ E and τ (1− p) < δ. A closed, densely defined operator a affiliated with A is called
τ-measurable if D(a) is τ-dense. The set of all τ-measurable operators is denoted by τ A. Evidently,
A ⊆ τ A.
In fact, τ A is a ∗-algebra with respect to the sum a ∔ b := a + b, product a · b := ab
and scalar product λ · a := λa (the 'strong operations'), and the involution being the adjoint a∗
(a, b ∈ τ A, λ ∈ C).
Remark 1.2 (see [Ter1, proof of Proposition I.24] or [GL2, Corollary 1.2]). Every polynomial in
elements of τ A formed using sums and products of unbounded operators (without closures) has a
unique extension in τ A, namely the one given at each step by the operations in τ A (with closures).
It also equals the closure of the original polynomial (which is densely defined).
To introduce the Haagerup Lp-spaces, we recall that A and ϕ have been fixed above, and set
henceforth A := A ⋊σϕ R, the crossed product of A by σϕ acting on L2(R) ⊗ H ∼= L2(R, H). It is
generated as a von Neumann algebra by π(A) ∪ λ(R), where π is the canonical embedding of A
in A given by (π(a)f )(t) := σϕ
−t(a)(f (t)) (a ∈ A, f ∈ L2(R, H), t ∈ R) and λ is the left regular
representation of R tensored by 1H, namely (λ(s)f )(t) := f (t − s) (f ∈ L2(R, H), s, t ∈ R). Write h
for the injective, positive selfadjoint operator on L2(R, H) such that λ(s) = his for all s ∈ R, whose
existence and uniqueness follow from Stone's theorem. For simplicity, we usually suppress π. By
construction, σϕ
t = Ad(λ(t)) = Ad(hit) for every t ∈ R. The crossed product A is equipped with an
automorphism group θ = (θs)s∈R that is the action of R dual to σϕ, characterised by the equalities
θs(a) = a and θs(hit) = e−isthit for all s, t ∈ R and a ∈ A. For an n.s.f. weight ψ on A, we denote
by eψ the (n.s.f.) dual weight on A [Haa1, Haa2]. There is a unique n.s.f. trace τ on A that, using
the Pedersen–Takesaki Radon–Nikodym derivative notation, satisfies d eϕ
Denote by A+ the extended positive part of A. The definition of eψ in [Haa2] extends naturally
to every normal, not necessarily faithful or semi-finite, weight ψ on A, yielding a dual (normal,
not necessarily faithful or semi-finite) weight on A. For every such ψ, let kψ := d eψ
dτ ∈ A+ in the
sense of [Haa3, Theorem 1.12]; by definition, kϕ = h. Then kψ is a genuine (unbounded, positive,
selfadjoint) operator if and only if ψ is semi-finite, and in this case, kψ is τ-measurable if and only
if ψ is bounded, that is, belongs to A∗. Furthermore, the map (A∗)+ ∋ ψ 7→ kψ extends linearly to
an injection ψ 7→ kψ from A∗ into τ A.
For each t ∈ R the map θt extends naturally to a ∗-automorphism of τ A, also denoted by θt. We
dτ = h.
define, for p ∈ [1,∞],
Lp(A) :=nx ∈ τ A : θt(x) = e−t/px for all t ∈ Ro .
The set Lp(A) is a selfadjoint linear subspace of τ A, and it is spanned by Lp(A)+ := Lp(A)∩ (τ A)+.
Up to isomorphism, this construction does not depend on the n.s.f. weight ϕ. Several special cases
are of particular importance. The set L∞(A) coincides with A (that is, its image in A), while
L1(A) = {kψ : ψ ∈ A∗}. One can thus define a functional tr on L1(A) by the formula tr(kψ) := ψ(1),
ψ ∈ A∗. For p ∈ [1,∞) and a closed, densely defined linear operator a affiliated with A with polar
decomposition a = ua, we have a ∈ Lp(A) if and only if u ∈ A and ap ∈ L1(A). This means one
can define a norm on Lp(A), setting kakp := tr(ap)1/p, and thus making Lp(A) a Banach space.
The canonical linear map ψ 7→ kψ is an isometry from A∗ onto L1(A). The functional tr is 'trace-
like': for conjugate exponents p, q ∈ [1,∞], a ∈ Lp(A) and b ∈ Lq(A), we have a · b, b · a ∈ L1(A)
8
ADAM SKALSKI AND AMI VISELTER
and tr(a · b) = tr(b· a). The Banach space L2(A) is actually a Hilbert space with respect to the inner
product ha, bi := tr(a · b∗), a, b ∈ L2(A). Using the product of τ A, we see that A acts on L2(A) by
left multiplication. With respect to this representation of A, the quadruple (A, L2(A),∗, L2(A)+) is
a standard representation for A. More generally, for every p ∈ [1,∞), L∞(A) acts on Lp(A) by left
and right multiplication in τ A. These actions are contractive: kx · akp,ka · xkp ≤ kxk∞kakp for all
x ∈ L∞(A) and a ∈ Lp(A).
We proceed to describe some of [GL2, Sections 1–2]. For q ∈ [2,∞), define a left ideal of A by
N (q) :=(cid:8)a ∈ A : ah1/q is closable and ah1/q ∈ Lq(A)(cid:9).
We have N = N (2) ⊆ N (q) [GL2, p. 46]. Further define a map j(q) : N (q) → Lq(A) by N (q) ∋ a 7→
ah1/q. It is linear and injective [GL2, p. 48].
Remark 1.3. We have N (q) = (cid:8)a ∈ A : h1/qa∗ ∈ τ A(cid:9). Indeed, for a ∈ A, observe that 'ah1/q is
closable and ah1/q ∈ τ A' if and only if 'the adjoint of ah1/q, namely h1/qa∗, belongs to τ A'. This is
because if ah1/q is closable and ah1/q ∈ τ A, then also h1/qa∗ = (cid:0)ah1/q(cid:1)∗ ∈ τ A; on the other hand,
if (ah1/q)∗ = h1/qa∗ belongs to τ A, then in particular it is densely defined, so that ah1/q is closable,
hence ah1/q = (h1/qa∗)∗ ∈ τ A. Next, when this is the case, we have h1/qa∗ ∈ Lq(A) (equivalently:
ah1/q ∈ Lq(A)) automatically: this is showed in the second part of the proof of [GL2, Proposition
2.2] (after 'it remains only to show that').
Given p ∈ [1,∞), define the following ∗-subalgebra of A:
M(p) := spanna ∈ A+ : a1/2 ∈ N (2p)o = spannb∗c : b, c ∈ N (2p)o .
We have M = M(1) ⊆ M(p). Also, define a linear map i(p) : M(p) → Lp(A) as the unique linear
extension of the map that takes a ∈ A+ such that a1/2 ∈ N (2p) to j(2p)(a1/2)∗j(2p)(a1/2). Then i(p)
is injective and positivity preserving. For all this see [GL2, p. 49]. The notation i(∞) := idA will be
useful. Let us list some key properties of these maps.
Proposition 1.4.
a ∈ M(p) and b ∈ M(q), are equal [GL2, Proposition 2.10].
(a) For every p ∈ [1,∞) and a, b ∈ N (2p) we have i(p)(a∗b) = j(2p)(a)∗ · j(2p)(b) [GL2, Lemma 2.9].
(b) Let a, b ∈ A. All scalars tr(cid:0)i(p)(a) · i(q)(b)(cid:1), for conjugate exponents p, q ∈ [1,∞] such that
(c) For every p ∈ [1,∞), i(p)(M∞ ∩ A+) is dense in Lp(A)+ [GL2, Proposition 2.11 (b)].
(d) For every p ∈ [1,∞), the map i(p) : M(p) → Lp(A), viewed as a densely defined operator from
A with the ultraweak topology to Lp(A) with the weak topology, is closable. The pertinent
closure, which we denote by i(p), is injective and positivity preserving [GL2, Proposition 2.12].
(e) For every a ∈ N ∗ and b ∈ M∞ we have ϕ(aσ−i/2(b)) = tr(cid:0)a · i(1)(b)(cid:1) [GL2, Proposition 2.13
Remark that for p, q ∈ [1,∞] conjugate exponents, x ∈ Lp(A) is positive if and only if tr(x· y) ≥ 0
for all y ∈ Lq(A)+. Consequently, if p ∈ [1,∞) and a ∈ M(p), then a ≥ 0 if (and only if) i(p)(a) ≥ 0,
because for every b ∈ M+, 0 ≤ tr(cid:0)i(p)(a) · i(q)(b)(cid:1) = tr(cid:0)a · i(1)(b)(cid:1), thus ω(a) ≥ 0 for every ω ∈ A+
∗
(use (b), (c) above).
(c)].
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
9
Convention. From now on we will tacitly identify:
• the ∗-algebra L∞(A) with A (that is, with its canonical image in A);
• Lp(A)∗ with Lq(A) for p ∈ [1,∞), q ∈ (1,∞] conjugate exponents by the correspondence
sending b ∈ Lq(A) to the functional Lp(A) ∋ a 7→ tr(a · b);
• (in particular) the Banach space L1(A) with A∗ by the correspondence sending x ∈ L1(A)
to the functional L∞(A) ∋ y 7→ tr(xy);
• the GNS representation (L2(A, ϕ), the canonical left action of A, ηϕ) with the semi-cyclic
representation
(L2(A), the action of L∞(A) ∼= A on L2(A) by multiplication on the left, j(2)
ϕ );
ϕ (xa) for every x ∈ A ∼= L∞(A) and a ∈ N , j
(2)
ϕ has dense
indeed, we have x · j
range [GL2, Proposition 2.11 (a)], and it satisfies kj
ϕ (a)k2 = ϕ(a∗a)1/2 for all a ∈ N
by [GL2, Lemma 2.1 (c)] and the invariance of the L2(A)-norm under the adjoint map,
essentially following from the trace-like property of the functional tr.
(2)
ϕ (a) = j
(2)
(2)
1.2. Markov operators, KMS-symmetry and Dirichlet forms. In this subsection we discuss sev-
eral notions and results from [GL2, Sections 3–5] that will be fundamental for us. We continue to
use the same notation as in the previous subsection, so in particular, A and ϕ are fixed.
Let p ∈ [1,∞). Write (cid:2)0, h1/p(cid:3)Lp(A) := (cid:8)x ∈ Lp(A) : 0 ≤ x ≤ h1/p(cid:9), the right inequality being in
the sense of unbounded positive selfadjoint operators (recall that generally h1/p does not belong to
τ A!). By [GL2, Lemma 3.1, Proposition 3.2 and Corollary 3.4], we have
i(p)([0, 1]M(p)) =h0, h1/piLp(A)
,
(1.1)
and this convex set is norm-closed in Lp(A). Notice that the linear span of (cid:2)0, h1/p(cid:3)Lp(A) equals
i(p)(M(p)).
Definition 1.5 ([GL2, p. 57]). Let p ∈ [1,∞) and let S, T be linear operators on, respectively,
Lp(A),A, not necessarily everywhere defined. Then
• S is Markov with respect to ϕ if(cid:2)0, h1/p(cid:3)Lp(A) is contained in D(S) and is invariant under S;
• T is Markov with respect to ϕ if M ⊆ D(T ) and T ([0, 1]A ∩ D(T )) ⊆ [0, 1]A
• T is KMS-symmetric with respect to ϕ if M ⊆ D(T ) and
tr(cid:16)T a · i(1)(b)(cid:17) = tr(cid:16)i(1)(a) · T b(cid:17)
(∀a,b∈M);
;
• T is p-integrable with respect to ϕ if M ⊆ D(T ), T (M) ⊆ M(p) and the map i(p)(a) 7→
i(p)(T a), a ∈ M, is continuous (with respect to the Lp(A)-norm). When this is the case, we
denote by eT (p) the unique bounded extension of that map to Lp(A).
Remark that when T is everywhere defined, it is Markov if and only if it is positive (= positivity
preserving) and contractive. Also, for p = 2, if S is symmetric and Markov, then for every a, b ∈ M(2)
we have
tr(cid:16)Si(2)(a) · i(2)(b)(cid:17) =DSi(2)(a), i(2)(b)∗E =Di(2)(a), S(i(2)(b)∗)E
=Di(2)(a), (Si(2)(b))∗E = tr(cid:16)i(2)(a) · Si(2)(b)(cid:17) ,
(1.2)
10
ADAM SKALSKI AND AMI VISELTER
where we used the fact that S is positivity preserving, thus adjoint preserving, on i(2)(M(2)).
We make a short detour to discuss the general theory of quadratic forms [Dav2, Kat]. Fix a Hilbert
space H. A non-negative quadratic form on H arises from a semi-inner product Q : D(Q)× D(Q) →
C for a linear subspace D(Q) of H. Specifically, Q is assumed to be linear in the first variable,
hermitian: Q(η, ζ) = Q(ζ, η) for all ζ, η ∈ D(Q), and positive semi-definite: Q(ζ, ζ) ≥ 0 for all
ζ ∈ D(Q). All quadratic forms in this paper will be non-negative. By the polarisation identity, Q is
determined by the function Q′ : D(Q) → [0,∞) given by Q′ζ := Q(ζ, ζ), ζ ∈ D(Q) (usually it is Q′
which is called the quadratic form). It is also useful to consider the function Q′′ : H → [0,∞] given
by
Q′′ζ :=
Q′(ζ)
∞
ζ ∈ D(Q)
else
(ζ ∈ H).
The form Q is said to be densely defined if D(Q) is dense in H, and closed if for every sequence
(ζn)∞n=1 in D(Q) that converges to ζ ∈ H and satisfies that Q′(ζn− ζm) −−−−−→n,m→∞
0 we have ζ ∈ D(Q)
0. This is equivalent to Q′′ being lower semicontinuous.
and Q′(ζn − ζ) −−−→n→∞
Using the Q′′ 'face' of quadratic forms, one can add and compare them, and form limits of as-
cending sequences of them. For instance, if Q1, Q2 are quadratic forms on H, then the quadratic
form Q1 + Q2 on H is given by (Q1 + Q2)′′ = Q′′1 + Q′′2.
There are 1 − 1 correspondences between: (generally unbounded) positive selfadjoint operators
A on H; C0-semigroups (St)t≥0 of selfadjoint contractions on H; and closed, densely-defined (non-
negative) quadratic forms Q on H. They are given by St = e−tA (t ≥ 0) – with the equality
understood either via the functional calculus or in the standard semigroup theory sense, and Q =
QA where QA :=(cid:13)(cid:13)A1/2·(cid:13)(cid:13)2, that is, D(QA) = D(A1/2) and Q′A(ζ) =(cid:13)(cid:13)A1/2ζ(cid:13)(cid:13)2 for all ζ ∈ D(A1/2).
so again, A and ϕ are fixed. Let πI : L2(A) →(cid:2)0, h1/2(cid:3)L2(A) be the nearest-point projection onto the
closed convex set(cid:2)0, h1/2(cid:3)L2(A) (see (1.1)). A closed densely-defined quadratic form Q on L2(A) is
said to be Dirichlet with respect to ϕ if Q ◦ πI ≤ Q [GL2, p. 62].
In the sequel we will not distinguish between Q, Q′, Q′′, and denote these three maps by Q.
Finally, having completed the detour, we introduce Dirichlet forms using the previous notation;
At this point we refer the reader to the Appendix, where we point out a gap in the proof of
[GL2, Theorem 4.7] and propose two solutions under additional hypotheses. The first one is more
practical for our purposes. It is used to prove Corollary A.8, which plays an important role in our
paper.
1.3. New results concerning Lp-embeddings. We now present several new results pertaining to
the material of the previous two subsections.
The first lemma is not used later on, but it complements Lemma 1.8.
Lemma 1.6. Let b, c ∈ N , and consider a := b∗c ∈ M. Then i(1)(a) ∈ L1(A), viewed as an element of
A∗, equals ωJ η(b),J η(c).
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
11
Proof. Denote by ω the element of A∗ that corresponds to i(1)(a). Assume for the moment that
b, c ∈ T . Write S := J∇1/2 = ∇−1/2J. By Proposition 1.4 (e), for every x ∈ N ∩ N ∗,
ω(x) = tr(x · i(1)(a)) = ϕ(xσ−i/2(a)) =(cid:10)η(σ−i/2(a)), η(x∗)(cid:11) =D∇1/2
η(a),∇−1/2J η(x)E = hη(a), J η(x)i = hη(x), J η(a)i
=D∇1/2
= hJbJ η(x), J η(c)i = hxJ η(b), J η(c)i .
η(a), Sη(x)E
Since ω is normal and N ∩ N ∗ is ultraweakly dense in A, we get ω = ωJ η(b),J η(c).
b, cn −−−→n→∞
For the general case of b, c ∈ N , let (bn)∞n=1, (cn)∞n=1 be bounded sequences in T such that
bn −−−→n→∞
η(c)
([Tak2, Theorem VI.1.26] and [StZs, Section 10.21, Corollary 2]; or [Str, p. 29, (16)]). Then
b∗c = a ultraweakly and i(1)(an) = ωJ η(bn),J η(cn) → ωJ η(b),J η(c) in norm. Since i(1)
an := b∗ncn −−−→n→∞
is ultraweak–σ(A∗,A)-closable (Proposition 1.4 (d)), i(1)(a) = ωJ η(b),J η(c).
c in the strong operator topology and η(bn) −−−→n→∞
η(b), η(cn) −−−→n→∞
(cid:3)
The next lemma's assertion is precisely that of [GL2, Lemma 2.5] but it uses weaker assumptions.
Lemma 1.7. Let α ∈ C be such that s := ℜα ≥ 0, and let a ∈ D(σiα) = D(σis) ⊆ A be such that
a, σiα(a)∗ ∈ N (1/s). Then ahα = hασiα(a).
Proof. By [StZs, 9.24, Proposition], a ∈ D(σis) if and only if ahs ⊆ hsb for some b ∈ A, in which
case b = σis(a), thus ahα ⊆ hασiα(a). Now, follow the proof of [GL2, Lemma 2.5]: since a, σiα(a)∗ ∈
N (1/s), ahα is closable, ahα ⊆ hασiα(a), and both sides are in τ A, hence they are equal.
(cid:3)
The next lemma connects the map i(2) introduced above with the GNS map η.
Lemma 1.8.
(a) For every a = b∗c with b, c ∈ N (4) (so a ∈ M(2)) such that a ∈ D(σ−i/4) and σ−i/4(a) ∈ N we
(b) For every a ∈ M we have i(2)(a) = ∇1/4
η(a).
In particular, if a ∈ M ∩ T then i(2)(a) =
have i(2)(a) = η(σ−i/4(a)).
η(σ−i/4(a)).
Proof. (a) We have i(2)(a) ⊇ h1/4ah1/4 because Proposition 1.4 (a) implies the relations i(2)(b∗c) ⊇
j(4)(b)∗j(4)(c) = h1/4b∗ch1/4. On the other hand, since σ−i/4(a) ∈ N ⊆ N (4) and a∗ ∈ M(2) ⊆ N (4),
Lemma 1.7 implies that h1/4a = σ−i/4(a)h1/4, thus i(2)(a) ⊇ σ−i/4(a)h1/2. Hence Remark 1.2 entails
that i(2)(a) = σ−i/4(a)h1/2 = j(2)(σ−i/4(a)) = η(σ−i/4(a)).
(b) Let a ∈ M, and assume that a = b∗c for b, c ∈ N . As in the proof of Lemma 1.6,
find bounded sequences (bn)∞n=1, (cn)∞n=1 in T that satisfy bn −−−→n→∞
c in the strong
η(c). Then the sequence (b∗ncn)∞n=1 in
operator topology and η(bn) −−−→n→∞
b∗η(c) = η(a) weakly. Since
M∞ converges ultraweakly to a and η(b∗ncn) = b∗nη(cn) −−−→n→∞
n=1 is bounded,
J∇1/2
η(b∗ncn)(cid:1)∞
as was (η(b∗ncn))∞n=1, hence so is (cid:0)∇1/4
n=1 by the Phragmen–Lindelöf three lines theorem.
η(b∗ncn) = i(2)(b∗ncn) for all n ∈ N by (a),
Thus, ∇1/4
Proposition 1.4 (d) implies that i(2)(a) = ∇1/4
η(b∗ncn) = η(c∗nbn) = c∗nη(bn) for all n ∈ N, the sequence (cid:0)∇1/2
η(a) weakly. Since ∇1/4
η(b∗ncn) −−−→n→∞ ∇1/4
η(b), η(cn) −−−→n→∞
η(b∗ncn)(cid:1)∞
b, cn −−−→n→∞
η(a).
(cid:3)
12
ADAM SKALSKI AND AMI VISELTER
Recall the (generalised) operators kψ defined in Subsection 1.1. The last statement below can be
also deduced from a general argument concerning the Lp-extensions of automorphisms preserving
the fixed reference weight; for extending yet more general maps to Haagerup Lp-spaces in the state
context we refer to [HJX, Section 5] (the arguments there can be also adapted to the weight case).
Lemma 1.9.
(a) For every normal weight ψ on A and t ∈ R we have kψ◦σt = h−itkψhit in A+.
(b) tr is Ad(hit)-invariant.
(c) Let p ∈ [2,∞), q ∈ [1,∞) and t ∈ R. The sets N (p),M(q), D(i(q)) are invariant under σt,
and for every a ∈ N (p) and b ∈ D(i(q)) we have j(p)(σt(a)) = hitj(p)(a)h−it and i(q)(σt(b)) =
hiti(q)(b)h−it.
(d) Let p ∈ [1,∞). For t ∈ R define U (t) : Lp(A) → Lp(A) by x 7→ hitxh−it, x ∈ Lp(A). Then
(U (t))t∈R is a (well-defined) C0-group of surjective isometries.
Proof. (a) Let ψ be a normal weight on A and t ∈ R. We claim that ^ψ ◦ σt = eψ ◦ Ad(hit). Once this
is established, we get the following equalities in A+:
kψ◦σt =
d(^ψ ◦ σt)
dτ
=
d(eψ ◦ Ad(hit))
dτ
deψ
dτ
= Ad(h−it)(
) = Ad(h−it)(kψ).
To prove the claim, write π for the canonical normal injection of A into A. Recall that θ = (θs)s∈R
is the action of R on A dual to σ, and denote by T the canonical n.s.f. operator-valued weight from
A to π(A) given by T x = RR θs(x) ds, x ∈ A+. Then for every x ∈ A+, since θs(hit) = e−isthit, we
have
T (hitxh−it) =ZR
θs(hitxh−it) ds =ZR
(hitθs(x)h−it) ds = hit(T x)h−it
(1.3)
(the equalities being in A+, thus in [π(A)+). By definition, we have ^ψ ◦ σt = (cid:0)ψ ◦ σt ◦ π−1(cid:1)∧ ◦ T ,
where the hat symbol denotes the canonical extension to [π(A)+. But σt ◦ π−1 = π−1 ◦ Ad(hit)π(A),
so (1.3) implies that
^ψ ◦ σt = (ψ ◦ π−1 ◦ Ad(hit))∧ ◦ T = \(ψ ◦ π−1) ◦ T ◦ Ad(hit) = eψ ◦ Ad(hit).
, tr(Ad(h−it)(kψ)) = tr(kψ◦σt ) = (ψ ◦ σt)(1) = ψ(1) = tr(kψ).
(b) For each ψ ∈ A+
∗
(c) Let a ∈ N (p). We have σt(a) = hitah−it. Thus
1
1
1
1
1
h
p a∗h−it.
p σt(a)∗ = h
p hita∗h−it = hith
p a∗ is τ-measurable, so is hith
(1.4)
Since a ∈ N (p), namely h
p a∗h−it, proving that σt(a) ∈ N (p). Now
(1.4) means that j(p)(σt(a))∗ = hitj(p)(a)∗h−it, and taking adjoints gives j(p)(σt(a)) = hitj(p)(a)h−it.
Take b ∈ M(q). From the foregoing it is obvious that σt(b) ∈ M(q) and i(q)(σt(b)) = hiti(q)(b)h−it.
More generally, if b ∈ D(i(q)), let (bλ)λ∈I
be a net in M(q) such that bλ −−→λ∈I
b ultraweakly and
σt(b) ultraweakly and i(q)(σt(bλ)) = hiti(q)(bλ)h−it −−→λ∈I
i(q)(b) weakly. Then σt(bλ) −−→λ∈I
i(q)(bλ) −−→λ∈I
hiti(q)(b)h−it weakly by (b) (it is clear that hitLq(A)h−it = Lq(A)). Hence σt(b) ∈ D(i(q)) and
i(q)(σt(b)) = hiti(q)(b)h−it.
(d) Fix p ∈ [1,∞) and x ∈ Lp(A). For every t ∈ R we have (cid:12)(cid:12)hitxh−it(cid:12)(cid:12)p = hit xp h−it,
so (b) entails that (cid:13)(cid:13)hitxh−it(cid:13)(cid:13)p = kxkp. For every t ∈ R define U (t) : Lp(A) → Lp(A) by
x 7→ hitxh−it. By the foregoing, U (t) is an isometry from Lp(A) onto itself. Clearly, (U (t))t∈R
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
13
is a group. It is also point–weakly continuous. Indeed, it p = 1, then for every b ∈ A we have
tr(hitxh−itb) = tr(xh−itbhit) = tr(xσ−t(b)), and the function t 7→ ω(σ−t(b)) is continuous for
all ω ∈ A∗.
If p ∈ (1,∞), write q for the conjugate exponent of p. For every a ∈ M(p) and
b ∈ M ⊆ M(q), we have tr(cid:0)i(p)(σt(a)) · i(q)(b)(cid:1) = tr(cid:0)σt(a) · i(1)(b)(cid:1) by Proposition 1.4 (b), so that
R ∋ t 7→ tr(cid:0)i(p)(σt(a)) · i(q)(b)(cid:1) is continuous. Since i(p)(σt(a)) = U (t)(i(p)(a)) for all t ∈ R and
a ∈ M(p), i(p)(M(p)) is dense in Lp(A) and i(q)(M) is dense in Lq(A) (Proposition 1.4 (c)), the
group (U (t))t∈R is point–weakly continuous, thus point–norm continuous.
(cid:3)
We need one more approximation result, identifying a convenient core for the embedding map
operators as well.
a in the ∗-strong operator topology and i(p)(an) −−−→n→∞
is bounded by (cid:13)(cid:13)i(p)(b)(cid:13)(cid:13) + ε, bλ −−→λ∈I
i(p).
Proposition 1.10. For every p ∈ [1,∞), the subspace M∞ is an ultraweak–weak core for i(p). To
elaborate, for every b ∈ D(i(p)) and ε > 0 there exists a net (bλ)λ∈I in M∞ that is bounded by kbk
such that (cid:0)i(p)(bλ)(cid:1)λ∈I
b in the ∗-strong operator topology and
i(p)(bλ) −−→λ∈I
i(p)(b) in norm. If b is selfadjoint or positive, (bλ)λ∈I can be chosen to consist of such
Proof. By convexity, it suffices to prove the assertion for weak convergence of(cid:0)i(p)(bλ)(cid:1)λ∈I
Step 1: We claim that for every a ∈ D(i(p)) there is a sequence (an)∞n=1 in D(i(p)), all of whose
elements are entire analytic with respect to σ, such that kank ≤ kak and(cid:13)(cid:13)i(p)(an)(cid:13)(cid:13)p ≤(cid:13)(cid:13)i(p)(a)(cid:13)(cid:13)p for
every n ∈ N, and such that an −−−→n→∞
in norm, and if a selfadjoint or positive, so are (an)∞n=1.
σt(a) dt, where the integral con-
Indeed, fix a ∈ D(i(p)). Let n ∈ N and write an :=
verges in the ∗-strong operator topology. Then an is entire analytic with respect to σ. Furthermore,
by Lemma 1.9 (c) and (d), (cid:13)(cid:13)i(p)(σt(a))(cid:13)(cid:13)p = (cid:13)(cid:13)i(p)(a)(cid:13)(cid:13)p for all t ∈ R and t 7→ i(p)(σt(a)) ∈ Lp(A) is
norm continuous. ThusRR e−nt2
i(p)(σt(a)) dt converges in norm. Hence an ∈ D(i(p)) and i(p)(an) =
i(p)(σt(a)) dt. In particular, (cid:13)(cid:13)i(p)(an)(cid:13)(cid:13)p ≤ (cid:13)(cid:13)i(p)(a)(cid:13)(cid:13)p. The sequence (an)∞n=1 has the de-
√n√πRR e−nt2
Before proceeding we make the following simple observation. Let b ∈ D(i(p)) and e, f ∈ T .
Then f∗be ∈ M satisfies i(p)(f∗be) = σ i
(e). Indeed, for every d ∈ N (2p) we have
de ∈ N ⊆ N (2p) and j(2p)(de)∗ = h
(e∗)j(2p)(d)∗ by
2p d∗ = σ− i
[GL2, Lemma 2.5] (or the more general Lemma 1.7 above), hence j(2p)(de)∗ = σ− i
(e∗) · j(2p)(d)∗
by Remark 1.2. Using Proposition 1.4 (a), this entails that for all b ∈ M(p) = D(i(p)) we have
(e). The observation follows for all b ∈ D(i(p)) from the definition
i(p)(f∗be) = σ i
of i(p).
√n√πRR e−nt2
(f )∗ · i(p)(b) · σ i
(f )∗ · i(p)(b) · σ i
2p
2p e∗d∗ = σ− i
2p
sired properties.
1
2p d∗ ⊇ σ− i
2p
1
(e∗)h
2p
1
2p
2p
.
i(p)(a)
(e∗)h
2p
2p
Step 2: We claim that for every b ∈ D(i(p)) that is entire analytic with respect to σ and every
δ > 0 there exists a net (bλ)λ∈I
is bounded by
(1 + δ)4(cid:13)(cid:13)i(p)(b)(cid:13)(cid:13)p, bλ −−→λ∈I
i(p)(b) weakly,
and if b is selfadjoint or positive, so are (bλ)λ∈I
Fix b, δ as above. By Lemma 1.1 there is a net (eλ)λ∈I
converges to 1 in the ∗-strong operator topology such that the nets(cid:0)σ± i
in M∞ that is bounded by kbk such that(cid:0)i(p)(bλ)(cid:1)λ∈I
b in the ∗-strong operator topology and(cid:0)i(p)(bλ)(cid:1)λ∈I −−→λ∈I
(eλ)(cid:1)λ∈I
in T that is bounded by 1 and that
are bounded by
2p
.
14
ADAM SKALSKI AND AMI VISELTER
λbeλ(cid:1)λ∈I
would suffice.)
(cid:3)
is a net in M∞ that is bounded by kbk and that converges
is bounded by
1 + δ. Then (bλ)λ∈I := (e∗λeλbe∗λeλ)λ∈I
to b in the ∗-strong operator topology, and by the preceding paragraph,(cid:0)i(p)(bλ)(cid:1)λ∈I
(1 + δ)4(cid:13)(cid:13)i(p)(b)(cid:13)(cid:13). If p > 1 write q ∈ (1,∞) for the conjugate exponent of p. For every c ∈ M,
tr(cid:16)b · i(1)(c)(cid:17) = tr(cid:0)i(p)(b) · i(q)(c)(cid:1)
by Proposition 1.4 (b). As before, density (Proposition 1.4 (c)) implies that i(p)(bλ) −−→λ∈I
tr(cid:16)i(p)(bλ) · i(q)(c)(cid:17) = tr(cid:16)bλ · i(1)(c)(cid:17) −−→λ∈I
2
(eλ) −−→λ∈I
i(p)(b)
1 in the ∗-strong operator topology (Lemma
(e∗λeλ) in L1(A); equivalently,
i(1)(b) in
1 ultrastrongly. This completes the proof. (Remark: without needing
(e∗λeλ)) as elements of A∗. Consequently, i(1)(bλ) −−→λ∈I
(e∗λeλ) · i(1)(b) · σ i
2
2
2
If p = 1, assume also that σ± i
weakly.
1.1). For every λ ∈ I we have i(1)(bλ) = σ− i
i(1)(bλ) = (σ− i
norm because σ− i
(bλ)λ∈I
to be selfadjoint or positive if b is, taking(cid:0)e2
(e∗λeλ) −−−→λ→I
(e∗λeλ))(i(1)(b))(σ i
2
2
The last proposition gives a neat description of the 'interval' that plays a key role when Dirichlet
forms are introduced.
Corollary 1.11.
(a) For every p ∈ [1,∞) we have D(i(p)) ∩ A+ ⊆ D(i(p)) = M(p).
(b) The closed convex set(cid:2)0, h1/2(cid:3)L2(A) = i(2) ([0, 1]M(2)) is equal to
∇1/4η(cid:0)[0, 1]M∞(cid:1) = ∇1/4η ([0, 1]M).
Proof. (a) Let b ∈ D(i(p)) be positive and of norm 1. By Proposition 1.10, there is a net (bλ)λ∈I
of positive elements in M∞ ⊆ M(p) that are bounded by 1 such that bλ −−→λ∈I
b in the ∗-strong
operator topology and (cid:0)i(p)(bλ)(cid:1)λ∈I −−→λ∈I
i(p)(b) weakly. The convex set (1.1) is closed in Lp(A)
in norm, equivalently weakly. Therefore, i(p)(b) ∈ i(p)([0, 1]M(p)). From injectivity of i(p) we get
b ∈ [0, 1]M(p).
η(a) for every a ∈ M. Hence we obtain the inclusions
(b) By Lemma 1.8 (b), i(2)(a) = ∇1/4
∇1/4η(cid:0)[0, 1]M∞(cid:1) ⊆ ∇1/4η ([0, 1]M) ⊆ i(2) ([0, 1]M(2)). The reverse inclusion follows from Proposi-
tion 1.10 by convexity.
Remark 1.12. Let p ∈ [1,∞). If T : A → A is a normal operator that is p-integrable with respect to
ϕ, then T leaves D(i(p)) invariant and i(p)(T a) = eT (p)(i(p)(a)) for every a ∈ D(i(p)). Indeed, given
such a, find a net (aλ)λ∈I
T a ultraweakly and i(p)(T aλ) = eT (p)(i(p)(aλ)) −−→λ∈I eT (p)(i(p)(a))
Proposition 1.10. Then T aλ −−→λ∈I
weakly, so Proposition 1.4 (d) gives the assertion.
a ultraweakly and i(p)(aλ) −−→λ∈I
in M such that aλ −−→λ∈I
i(p)(a) weakly by
(cid:3)
1.4. Locally compact quantum groups. We introduce locally compact quantum groups in the von
Neumann algebraic setting ([KV2], see also [KV1, VD]), and describe some of their features.
Definition 1.13. A locally compact quantum group is a virtual object studied via a pair (M, ∆),
where:
• M is a von Neumann algebra;
• ∆ : M → M⊗M is a co-multiplication (or co-product), that is, a normal unital ∗-homomorphism
that is co-associative: (∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆;
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
15
• there exist n.s.f. weights ϕ, ψ on M, called the left and right Haar weights, respectively, that
satisfy
ϕ((ω ⊗ id)∆(x)) = ϕ(x)ω(1)
ψ((id ⊗ ω)∆(x)) = ψ(x)ω(1)
We set L∞(G) := M and L1(G) := L∞(G)∗.
for all ω ∈ M+
for all ω ∈ M+
∗ , x ∈ M+ with ϕ(x) < ∞,
∗ , x ∈ M+ with ψ(x) < ∞.
The left and right Haar weights are unique up to scaling. We write L2(G) for the GNS Hilbert
space of L∞(G) with respect to ϕ, and always view L∞(G) as acting on L2(G). Set ∇ := ∇ϕ,
J := Jϕ and η := ηϕ. There is an injective, positive, selfadjoint operator δ affiliated with L∞(G),
called the modular element, such that ψ = ϕδ in the sense of [Vae].
Every locally compact quantum group G has a dual locally compact quantum group, denoted
by G. The objects pertaining to it will be adorned with a 'hat', e.g. ∆, ϕ, ψ, ∇, J, η. We will not
elaborate on this construction, but mention that the GNS Hilbert space of L∞( G) with respect to ϕ
can and will be naturally identified with L2(G). So we shall always view L∞( G) too as acting on
L2(G). Importantly, the 'Pontryagin double dual property' holds: the dual of G is G.
There exists a distinguished (multiplicative) unitary W ∈ L∞(G)⊗ L∞( G), called the left regular
representation, which implements the co-multiplication by ∆(x) = W ∗(1 ⊗ x)W for all x ∈ L∞(G).
k·k is a C∗-algebra, which is ultraweakly dense
The subspace C0(G) := {(id ⊗ ω)(W ) : ω ∈ L1( G)}
in L∞(G). Furthermore, ∆(C0(G)) ⊆ M(C0(G) ⊗min C0(G)), W ∈ M(C0(G) ⊗min C0( G)) and
W = σ(W ∗). The right regular representation is a (multiplicative) unitary V ∈ L∞( G)′ ⊗ L∞(G)
with similar properties, and in particular ∆(x) = V (x⊗ 1)V ∗ for all x ∈ L∞(G). In fact L∞(G)′, the
commutant of L∞(G) in B(L2(G)), admits its own co-product, arising from the commutant quantum
group G′ (so that L∞(G′) := L∞(G)′) – see [KV2, Section 4] for details. All the objects associated
with G′, such as the co-product, Haar weights, etc., will be adorned with primes.
An object of fundamental importance is the antipode of G.
It is a generally unbounded, ul-
traweakly closed, linear map S on L∞(G) that is characterised by {(id ⊗ ω)(W ) : ω ∈ L1( G)} being
a ∗-ultrastrong core for S and the formula S ((id ⊗ ω)(W )) = (id ⊗ ω)(W ∗) for every ω ∈ L1( G). It
has a 'polar decomposition' as S = R ◦ τ−i/2, where the unitary antipode R is a ∗-anti-automorphism
of L∞(G) and τ−i is the 'analytic generator' of the scaling group τ = (τt)t∈R of automorphisms
of L∞(G). The relavant objects are given by the formulas R(x) = Jx∗ J and τt(x) = ∇itx ∇−it
(x ∈ L∞(G), t ∈ R). Additionally, R and τ reduce to a ∗-anti-automorphism and an automorphism
group, respectively, of the C∗-algebra C0(G).
Definition 1.14. A unitary representation of G on a Hilbert space H is a unitary U ∈ M(C0(G)⊗min
K(H)) that satisfies (∆ ⊗ id)(U ) = U13U23, where we use the standard leg numbering nota-
tion. A little more generally, for a C∗-algebra B, one defines similarly unitary representations
U ∈ M(C0(G) ⊗min B) of G.
For instance, the left regular representation W is a unitary representation of G on L2(G). The
trivial representation of G is 1 ∈ M(C0(G)) = M(C0(G) ⊗ C). The above-mentioned property of
the antipode extends as follows: for every unitary representation U of G on a Hilbert space H and
ω ∈ B(H)∗ we have (id ⊗ ω)(U ) ∈ D(S) and S ((id ⊗ ω)(U )) = (id ⊗ ω)(U∗).
16
ADAM SKALSKI AND AMI VISELTER
0(G) ⊗min Cu
0(G) → M(Cu
Locally compact quantum groups can be also studied via a universal C∗-algebra Cu
0(G) [Kus2],
0(G)). There is a unitary rep-
equipped with a co-multiplication ∆u : Cu
0( G)) of G, the right semi-universal version of W , satisfying
resentation W ∈ M(C0(G) ⊗min Cu
{(ω ⊗ id)(W) : ω ∈ L1(G)}k·k = Cu
0( G), determined by the following property: there is a 1 − 1 cor-
0( G) given as follows:
respondence between unitary representations of G and representations of Cu
a unitary representation U of G on a Hilbert space H is associated to the (unique) representation Φ
0( G) on H by U = (id⊗ Φ)(W), where we view Φ as taking values in M(K(H)) ∼= B(H). There
of Cu
0(G) ⊗min C0( G)), the left semi-universal version of W , with a similar
is also a unitary W ∈ M(Cu
universality property, and we have W = σ(W∗). There are also V, V for V .
Applying the (dual of the) above correspondence to the representation W and to the trivial
representation gives, respectively, the reducing morphism, which is a surjective ∗-homomorphism
0(G)∗
Λ : Cu
satisfying (ǫ ⊗ id)(W) = 1. We have (Λ ⊗ Λ) ◦ ∆u = ∆ ◦ Λ and
0(G) → C0(G) satisfying (Λ ⊗ id)(W) = W , and the co-unit, which is a character ǫ ∈ Cu
(id ⊗ Λ)(∆u(x)) = W∗(1 ⊗ Λ(x))W (∀x∈Cu
(1.5)
where the right-hand side makes sense when considering C0(G), C0( G) ⊆ B(L2(G)). In this point
of view we also have W ∈ M(Cu
The unitary antipode and scaling group of G, and thus also its antipode, lift to objects acting on
0(G) denoted by Su, Ru, τ u = (τ u
−i/2, we have Λ ◦ Ru = R ◦ Λ
Cu
t = τt ◦ Λ for all t ∈ R, and the fundamental property of the antipode has a universal
and Λ ◦ τ u
version:
t )t∈R, respectively. Again, Su = Ru ◦ τ u
0(G) ⊗min K(L2(G))).
0(G)),
(id ⊗ ω)(W) ∈ D(Su) and Su ((id ⊗ ω)(W)) = (id ⊗ ω)(W∗)
(∀ω∈L1( G)).
(1.6)
The co-multiplications ∆, ∆C0(G) and ∆u induce on L1(G), C0(G)∗ and Cu
0(G)∗, respectively,
convolution products, for instance by ω1 ⋆ ω2 := (ω1 ⊗ ω2)◦ ∆ (ω1, ω2 ∈ L1(G)), turning these spaces
0(G)∗. The restriction
into completely contractive Banach algebras. The co-unit ǫ is the unit of Cu
map and the map of composing with the reducing morphism allow us to embed L1(G) ֒→ C0(G)∗ ֒→
0(G)∗ as Banach algebras. What is more, each 'small' set is an ideal in every 'larger' one (see for
Cu
example [HNR] and references therein).
For every Banach algebra A, its dual A∗ becomes an A-bimodule as customary, namely, for θ ∈ A∗
and a ∈ A, one defines θa, aθ ∈ A∗ by (θa)(b) := θ(ab) and (aθ)(b) := θ(ba), b ∈ A. When
specialising to one of the Banach algebras L1(G), C0(G)∗, Cu
0(G)∗, we denote these bimodule actions
with a '·' rather than just by juxtaposition. For instance, by the definition of the convolution, the
actions of µ ∈ Cu
0(G)∗ on x ∈ Cu
0(G) ֒→ Cu
0(G)∗∗ are
µ · x = (id ⊗ µ)(∆u(x)), x · µ = (µ ⊗ id)(∆u(x)) ∈ M(Cu
0(G) is dense in Cu
0(G)∗ · Cu
0(G)).
0(G); the same holds for
0(G). These assertions follow from the quantum
0(G) and L1(G) · Cu
In fact, we have Cu
the other module map and for C0(G) in place of Cu
cancellation rules, e.g. span ∆u(Cu
0(G) ⊆ Cu
0(G))(1 ⊗ Cu
0(G)) = Cu
0(G) ⊗min Cu
0(G).
Example 1.15. Commutative locally compact quantum groups G, namely the ones whose L∞(G)
algebra is commutative, are in 1 − 1 correspondence with locally compact groups G. We have
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
17
0(G) = C0(G) = C0(G), (∆(f ))(t, s) = f (ts) for f ∈ L∞(G) and t, s ∈ G (after
L∞(G) = L∞(G), Cu
identifying L∞(G)⊗ L∞(G) with L∞(G× G)) and L2(G) = L2(G). The left and right Haar weights
ϕ, ψ are given by integration against the left and right Haar measures of G, respectively. As σϕ is
trivial, it follows immediately from Corollary 1.11 (b) that
h0, h1/2
ϕ iL2(G)
= i(2)
ϕ ([0, 1]M(2,ϕ)) =(cid:8)f ∈ L2(G) : 0 ≤ f ≤ 1 a.e.(cid:9) .
Example 1.16. Co-commutative locally compact quantum groups G, namely the ones whose L1(G)
algebra is commutative, are in 1 − 1 correspondence with duals G of locally compact groups G.
We have L∞( G) = VN(G), the left von Neumann algebra of G generated by the left translation
operators {λg : g ∈ G} on L2(G) = L2( G), C0( G) is the reduced group C∗-algebra C∗r (G) of G,
0( G) is the full group C∗-algebra C∗(G) of G, and ∆(λg) = λg ⊗ λg for all g ∈ G. Denote the
Cu
modular function of G by δ, and for f : G → C define f #, f∗ : G → C by f #(s) := δ(s)−1f (s−1)
and f∗(s) := δ(s)−1/2f (s−1), s ∈ G. Recall that the convolution of f ∈ L2(G) and g ∈ Cc(G) is
given by (f ∗ g)(s) := RG f (t)g(t−1s) dt (s ∈ G). The left and right Haar weights of G are both
equal to the Plancherel weight of G [Tak2, Section VII.3]. It is the n.s.f. weight associated to the
Tomita algebra Cc(G) ⊆ L2(G) with convolution as product and involution being Cc(G) ∋ f 7→ f #.
The associated modular operator ∆ is the operator of multiplication by δ with maximal domain
in L2(G), and the modular conjugation J is L2(G) ∋ f 7→ f∗. The left-bounded elements are all
f ∈ L2(G) such that the map given by Cc(G) ∋ g 7→ f ∗ g extends to a bounded operator πl(f ) on
L2(G). With respect to the Plancherel weight, from Corollary 1.11 (b) we deduce that
i(2) ([0, 1]M(2)) = ∇1/4η ([0, 1]M)
L2(G)
= ∇1/4η ({πl(f )∗πl(f ) : f ∈ L2(G) is left bounded and kπl(f )k ≤ 1})
L2(G)
.
Let f ∈ L2(G) be left bounded with kπl(f )k ≤ 1. We can use [Tak2, Theorem VI.1.26] to find a
sequence (fn)∞n=1 in Cc(G) that converges to f in L2(G) and such that kπl(fn)k ≤ 1 for all n ∈ N
and πl(fn) −−−→n→∞
πl(f ) in the strong operator topology. As in the proof of Lemma 1.8 (b), we then
have ∇1/4
η(πl(f )∗πl(f )) weakly. By convexity,
η(πl(fn)∗πl(fn)) −−−→n→∞ ∇1/4
i(2) ([0, 1]M(2)) = co(cid:8)δ1/4(f # ∗ f ) : f ∈ Cc(G),kπl(f )k ≤ 1(cid:9)L2(G)
.
We return to the general case. For µ ∈ C0( G)∗, the next lemma is just [NgV, Lemma 3.2 and
Remark 2]. Let us give full details (using a different argument) for completeness.
Lemma 1.17. For every µ ∈ Cu
0( G)∗ we have
(id ⊗ µ)(W) ∈ D(S) and S((id ⊗ µ)(W)) = (id ⊗ µ)(W∗).
Proof. We may suppose that µ is a state. Writing (H, π, ξ) for the GNS construction of µ, the operator
Wµ := (id⊗ π)(W) ∈ M(C0(G)⊗minK(H)) is a unitary representation of G on H and (id⊗ µ)(W) =
(id ⊗ ωξ)(W µ). Thus this operator belongs to D(S) and S((id ⊗ ωξ)(W µ)) = (id ⊗ ωξ)(W ∗µ) =
(id ⊗ µ)(W∗).
(cid:3)
18
ADAM SKALSKI AND AMI VISELTER
The following construction is useful. For µ ∈ Cu
µ ◦ τ u
√n
√πZR
e−nt2
µn :=
0(G)∗, let
t dt ∈ Cu
0(G)∗
(n ∈ N),
(1.7)
with the integrals converging in the w∗-topology. Then for all n ∈ N, µn is entire analytic with
respect to (the adjoint of) τ u and kµnk ≤ kµk. In addition, µn −−−→n→∞
µ in the w∗-topology, thus
kµnk −−−→n→∞ kµk.
The next definition and the proposition which follows it belong to the quantum group folklore.
Definition 1.18. A locally compact quantum group G is second countable if C0(G) is separable.
Proposition 1.19. For a locally compact quantum group G the following conditions are equivalent:
(a) G is second countable;
(b) G is second countable;
(c) Cu
0(G) is separable;
0( G) is separable;
(d) Cu
(e) L2(G) is separable.
Proof. It is obviously sufficient to prove that (a), (c) and (e) are equivalent.
(a) =⇒ (e) Since ϕ is the W ∗-lift of its restriction to C0(G), which is a KMS weight (in the sense
of [KV1]), it follows from modular theory that if A is a countable dense subset of C0(G) ∩ Nϕ ∩
N ∗ϕ, then JϕAJϕηϕ(A) is dense in L2(G) = ηϕ(C0(G) ∩ Nϕ ∩ N ∗ϕ) and the latter is also equal to
Jϕ(C0(G) ∩ Nϕ ∩ N ∗ϕ)Jϕηϕ(C0(G) ∩ Nϕ ∩ N ∗ϕ).
(e) =⇒ (c) This follows from the density of(cid:8)(id ⊗ ωζ,η)(W) : ζ, η ∈ L2(G)(cid:9) in Cu
(c) =⇒ (a) Simply use the canonical reducing surjection Cu
We will refer several times to the following particular classes of locally compact quantum groups:
compact (or dually, discrete), amenable, and co-amenable. Since their definitions are not necessary
for understanding the chief part of this paper, we will not give them here, but rather refer the reader
to [MVD, Wor1, Wor2] and to [BeT].
0(G) → C0(G).
0(G).
(cid:3)
2. THE LEFT AND RIGHT CONVOLUTION OPERATORS ASSOCIATED WITH LOCALLY COMPACT QUANTUM
GROUPS
In this section we discuss various incarnations of convolution operators associated to quantum
0(G)). After recall-
measures on locally compact quantum groups (i.e. continuous functionals on Cu
ing and sometimes rephrasing known results in Subsection 2.1, essentially following [JNR, Daw3,
DFSW, SaS], in Subsection 2.2 we introduce the conditions on the functional in question guarantee-
ing that the corresponding convolution operator possesses natural symmetry properties. Finally in
Subsection 2.3 we analyse the GNS- and KMS-implementations of convolution operators on L2(G).
0(G)∗. Since L1(G), when
2.1. Basic facts. Let G be a locally compact quantum group and µ ∈ Cu
viewed canonically as a subspace of the completely contractive Banach algebra Cu
0(G)∗, is an ideal,
the operators L1(G) → L1(G) given by L1(G) ∋ ω 7→ ω ⋆ µ and L1(G) ∋ ω 7→ µ ⋆ ω have cb-norms
dominated by kµk. Thus their adjoints, denoted by Lµ, Rµ : L∞(G) → L∞(G), respectively, are
normal and satisfy kLµkcb ,kRµkcb ≤ kµk. We also write µ · x := Lµ(x) and x · µ := Rµ(x). These
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
19
µ, Lu
µ : Cu
0(G) → Cu
0(G) given by the Cu
0(G)∗-bimodule structure of Cu
0(G)∗-bimodule, extending its canonical structure as an L1(G)-bimodule. We
make L∞(G) into a Cu
0(G),
also have the operators Ru
µ = (id ⊗ µ) ◦ ∆u. In fact in the literature
admitting also a simple direct description: for example Lu
one can find varying left/right conventions here (so that for example [SaS] calls the latter map a
right convolution operator). Finally note that we can recover µ by ǫ ◦ Ru
We gather some information from the literature about these operators in the subsequent results.
Notice that a completely bounded map T on L∞(G) satisfies ∆ ◦ T = (T ⊗ id) ◦ ∆ (resp. ∆ ◦ T =
(id ⊗ T ) ◦ ∆) if and only if it commutes with the operators Lω (resp. Rω), ω ∈ L1(G). We will see
below that a similar, and even stronger, statement holds for the maps Ru
µ. A bounded linear map
between two C∗-algebras is called strict if it is strictly continuous on bounded sets.
µ.
µ = µ = ǫ ◦ Lu
µ, Lu
Theorem 2.1. Let G be a locally compact quantum group.
(a) The maps Cu
(b) For every state µ ∈ Cu
0(G)∗ ∋ µ 7→ Rµ and Cu
0(G)∗ ∋ µ 7→ Lµ are injective.
0(G)∗ the left, respectively right, Haar weight is invariant under Rµ,
respectively Lµ.
(c) We have
0(G)) : T u ⊗ id is strict and ∆u ◦ T u = (T u ⊗ id) ◦ ∆u}
0(G)) : T u commutes with all maps Lu
(cid:8)Ru
0(G)∗(cid:9) = {T u ∈ CB(Cu
µ : µ ∈ Cu
= {T u ∈ B(Cu
=(cid:13)(cid:13)Ru
µ(cid:13)(cid:13)cb
Furthermore,(cid:13)(cid:13)Ru
(d) There exist 1 − 1 completely isometric correspondences between:
• completely bounded, normal maps T on L∞(G) that satisfy ∆ ◦ T = (T ⊗ id) ◦ ∆;
• completely bounded maps T ′ on C0(G) that commute with the operators Lω, ω ∈ L1(G);
• completely bounded right module maps M on L1(G) (that is, M (ω1 ⋆ ω2) = M (ω1) ⋆ ω2
µ(cid:13)(cid:13) = kµk for all µ ∈ Cu
ν, ν ∈ Cu
0(G)∗} .
0(G)∗.
for all ω1, ω2 ∈ L1(G)).
They are given by T ′ = TC0(G) and M = T∗.
(e) There exist 1 − 1 order-preserving correspondences between:
0(G)∗;
• positive functionals µ ∈ Cu
• completely positive, normal maps T on L∞(G) that satisfy ∆ ◦ T = (T ⊗ id) ◦ ∆;
• completely positive maps T ′ on C0(G) that commute with the operators Lω, ω ∈ L1(G);
0(G)∗;
• completely positive maps T u on Cu
• right module maps M on L1(G) with completely positive adjoints.
0(G) that commute with the operators Lu
ν, ν ∈ Cu
They are given by T = Rµ, T ′ = TC0(G), T u = Ru
kTkcb = kT ′k = kT ′kcb = kT uk = kT ukcb = kMk = kMkcb.
(f) Assume that G is amenable. There exist 1 − 1 correspondences between:
µ and M = T∗, and satisfy kµk = kTk =
0(G)∗;
• functionals µ ∈ Cu
• completely bounded, normal maps T on L∞(G) that satisfy ∆ ◦ T = (T ⊗ id) ◦ ∆;
• completely bounded maps T ′ on C0(G) that commute with the operators Lω, ω ∈ L1(G);
• completely bounded right module maps M on L1(G).
They are given by T = Rµ , T ′ = TC0(G) and M = T∗.
In each of (c)–(f) a similar statement holds for the left actions of Cu
0(G)∗.
20
ADAM SKALSKI AND AMI VISELTER
Proof. (a) See [Daw1, Proposition 8.3 and its proof] (or Lemma 2.3 below).
(b) This is [KNR, Lemma 3.4].
(c) The assertion is proved precisely as [LS1, Proposition 3.2].
(d) Follows from [JNR, Propositions 4.1 and 4.2], see also [SaS, Lemma 12] (and the preceding
remark, replacing complete positivity by complete boundedness and ignoring the non-degeneracy
condition).
(e) Follows from (c), (d) and [Daw3, Theorems 5.1 and 5.2].
(f) Follows from (d) and the left analogue of [Cra1, Proposition 5.10 and the discussion succeed-
(cid:3)
ing the definition of eρ in Section 3].
(a) By [Cra2, Theorem 7.2], the correspondence in (f) is isometric: kµk = kTkcb = kMkcb.
(b) The result of (f) was known before [Cra1] for co-amenable G [HNR, Proposition 3.1 or
Remark 2.2.
Theorem 4.2].
We now start discussing connections between various modes of convergence of different avatars
of convolution operators (see also [RuV, Theorem 4.6]).
Lemma 2.3. Let (µi)i∈I be a bounded net in Cu
conditions:
0(G)∗, and let µ ∈ Cu
0(G)∗. Consider the following
(a) µi −−→i∈I
(b) Ru
µi(x) −−→i∈I
(c) Rµi(x) −−→i∈I
(d) Rµi(x) −−→i∈I
(e) (Rµi)∗ −−→i∈I
µ in the w∗-topology;
0(G);
Ru
µ(x) weakly for every x ∈ Cu
Rµ(x) ultraweakly for every x ∈ C0(G);
Rµ(x) ultraweakly for every x ∈ L∞(G);
(Rµ)∗ in the point–norm topology.
Then (e) =⇒ (d) =⇒ (c) ⇐⇒ (b) ⇐⇒ (a), and if kµik = kµk for all i ∈ I, then all these conditions
are equivalent. Similar statements hold for operators of the form Lµ, µ ∈ Cu
Proof. The implications (e) =⇒ (d) =⇒ (c) are clear, and (a) implies (e) under the additional norm
assumption by [RuV, Theorem 4.6, (i) =⇒ (vii)].
0(G)∗.
0(G) → C0(G) be the reducing morphism. Then for every a ∈ Cu
Proving that (c) ⇐⇒ (a) can be done similarly to the proof of [RuV, Corollary 4.5] as follows.
0(G), ω ∈ L1(G) and ν ∈
0(G). This implies
0(G).
allows us to conclude that
Let Λ : Cu
0(G)∗ we have ν(ω · a) = ω(Rν(Λ(a))), where we used the action of L1(G) on Cu
Cu
that (a) =⇒ (c). Conversely, assuming (c), we see that µi(b) −−→i∈I
Since L1(G) · Cu
µi −−→i∈I
because Cu
The implication (a) =⇒ (b) is immediate, and the converse holds by the boundedness of (µi)i∈I
0(G), the boundedness of (µi)i∈I
µ(b) for all b ∈ L1(G) · Cu
0(G) is dense in Cu
µ in the w∗-topology.
0(G)∗ · Cu
0(G) ⊆ Cu
0(G) is dense in Cu
0(G).
The next result follows from, and is in fact equivalent to, [DFSW, Theorem 2.6]. For clarity, we
give its proof, whose second part is different from that of [DFSW]. In the language of [CaS] we
would say that we describe here the GNS-implementations of convolution operators.
(cid:3)
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
21
ϕ ) ⊆ M+
µ ∈ C0( G′).
µ and eLψ
µ = (µ⊗ id)(W∗) ∈ M(C0( G)) and eLψ
µ ∈ C0( G), and eLψ
0(G)∗, Nϕ,Mϕ are invariant under Rµ and Nψ,Mψ are invariant
Proposition 2.4. For every µ ∈ Cu
under Lµ. The maps ηϕ(x) 7→ ηϕ(Rµ(x)), x ∈ Nϕ, and ηψ(x) 7→ ηψ(Lµ(x)), x ∈ Nψ, are well
defined and extend to bounded operators on L2(G) denoted by eRϕ
µ , respectively. In fact, we
have eRϕ
µ = (id⊗ µ)(V) ∈ M(C0( G′)). Furthermore if µ ∈ L1(G),
then eRϕ
ϕ follows readily from Theorem 2.1 (b).
Proof. Suppose that µ is a state. That Rµ(M+
For every x ∈ Nϕ we have Rµ(x)∗Rµ(x) ≤ Rµ(x∗x) by Kadison's Cauchy–Schwarz inequality,
which applies because Rµ is a completely positive contraction. From Theorem 2.1 (b) we hence
get ϕ(Rµ(x)∗Rµ(x)) ≤ ϕ(Rµ(x∗x)) = ϕ(x∗x). For general µ ∈ Cu
0(G)∗, decompose µ as a linear
combination of states.
The equality eRϕ
alternative proof is the following. For every ω ∈ L1(G) we have eRϕ
definition of the operator W . Therefore, under the embedding L1(G) ֒→ Cu
(ω ⊗ id)(W∗) for all ω ∈ L1(G). Consider now the anti-homomorphisms Cu
by µ 7→ eRϕ
and that the algebra {eRϕ
Since W ∈ M(Cu
have eRϕ
µ ∈ C0( G) is clear.
µ = (µ ⊗ id)(W∗) can be easily deduced from [Kus2, Corollary 8.2, (1)]. An
ω = (ω ⊗ id)(W ∗) by the very
0(G)∗, we have eRϕ
ω =
0(G)∗ → B(L2(G)) given
µ and µ 7→ (µ ⊗ id)(W∗). Using the facts that they coincide on the ideal L1(G) of Cu
0(G)∗
ω : ω ∈ L1(G)} acts non-degenerately on L2(G), we infer that they coincide
µ ∈ M(C0( G)). Then the fact that for µ ∈ L1(G) we
0(G) ⊗ C0( G)), we have eRϕ
The proof for Lµ is similar.
0(G)∗.
on Cu
(cid:3)
We are now ready to start discussing GNS-implementations of convolution operators with respect
to the 'wrong-sided' weights (i.e. for example ψ for the right convolution operator).
Proposition 2.5.
(a) Let ζ, η be in D(δ− 1
2 ) (resp. D(δ
Rω (resp. Lω).
1
2 )) and ω := ωζ,η. Then Mψ (resp. Mϕ) is invariant under
(b) Let ζ be in D(δ− 1
1
2 ) (resp. D(δ
2 )), η ∈ L2(G) and ω := ωζ,η. Then Nψ (resp. Nϕ) is invariant
under Rω (resp. Lω). The map ηψ(x) 7→ ηψ(Rω(x)), x ∈ Nψ, (resp. ηϕ(x) 7→ ηϕ(Lω(x)),
x ∈ Nϕ) is well defined and extends to a bounded operator on L2(G) denoted by eRψ
ω (resp. eLϕ
ω),
which is equal to eRϕ
(resp. eLψ
1
2 ζ,η
2 ζ,η
− 1
).
ω
ω
δ
δ
Proof. We prove only the first assertions.
(a) It is enough to assume ζ = η. By [KV1, Result 7.6], in that case ψ (Rω(x)) = ψ(x)ω
(1)
δ− 1
2 ζ
for each x ∈ M+
ψ .
2 ∈ Nϕ}. Then
(b) Recall that ψ = ϕδ. Write Nψ,0 := {x ∈ L∞(G) : xδ
Nψ,0 ⊆ Nψ, and for every x ∈ Nψ,0 we have ηψ(x) = ηϕ(xδ
2 ) and η ∈ L2(G).
Let x ∈ Nψ,0. The identity ∆(δ) = δ ⊗ δ, which can be understood for example in the language of
functional calculus for operators affiliated with a C∗-algebra, or directly as W ∗(1 ⊗ δ)W = δ ⊗ δ,
yields (ωζ,η ⊗ id)(∆(x))δ
2 )W ⊇
2 is bounded and xδ
2 ). Fix ζ ∈ D(δ− 1
2 )). Indeed, we have ∆(xδ
2 ) = W ∗(1 ⊗ xδ
2 ζ,η ⊗ id)(∆(xδ
2 ⊆ (ω
δ− 1
1
1
1
1
1
1
1
22
ADAM SKALSKI AND AMI VISELTER
W ∗(1 ⊗ x)W W ∗(1 ⊗ δ
D(ωζ,η ⊗ id)(∆(x))δ
1
1
1
2 )W = ∆(x)(δ
2 ⊗ δ
2 α, βE =D∆(x)(ζ ⊗ δ
1
=D∆(xδ
1
2 ), and for every α ∈ D(δ
2 α), η ⊗ βE =D∆(x)(δ
2 ζ ⊗ α), η ⊗ βE =D(ω
2 ζ,η ⊗ id)(∆(xδ
δ− 1
1
1
2 )(δ− 1
1
2 ) = (ω
1
2 ) and β ∈ L2(G),
2 δ− 1
2 α), η ⊗ βE
2 ζ ⊗ δ
1
1
δ− 1
2 ζ,η ⊗ id)(∆(xδ
1
2 ))α, βE .
2 )) ∈ Nϕ, thus Rωζ,η (x) = (ωζ,η ⊗
By Proposition 2.4 we have Rω
id)(∆(x)) ∈ Nψ,0. We infer that
δ
(xδ
− 1
2 ζ,η
ηψ(Rωζ,η (x)) = ηϕ(Rω
(xδ
− 1
2 ζ,η
δ
ηϕ(xδ
− 1
2 ζ,η
δ
ηψ(x).
− 1
2 ζ,η
δ
1
2 )) = eRϕ
ω
1
2 ) = eRϕ
ω
δ
ω
− 1
2 ζ,η
ηψ(x). Proposition 2.4 ends the proof.
From the definition of ϕδ [Vae], [Tak2, Theorem VI.1.26], and the ultrastrong–norm (in fact, weak
operator–weak) closedness of ηψ, for each x ∈ Nψ we have Rωζ,η (x) ∈ Nψ and ηψ(Rωζ,η (x)) =
eRϕ
)t∈R be the universal modular automorphism groups of the universal left
and right Haar weights of G, respectively [Kus2, Proposition 8.6 and Definition 8.1]. Recall that
t )t∈R and Ru are the universal scaling group and unitary antipode of G. By [Kus2, Propositions
(τ u
7.2, 9.2, 9.3 and 10.4], for every t ∈ R,
)t∈R and (σψ,u
Let (σϕ,u
(cid:3)
t
t
∆u ◦ σϕ,u
∆u ◦ τ u
t = (τ u
t = (τ u
t ⊗ σϕ,u
t ⊗ τ u
t
) ◦ ∆u,
t ) ◦ ∆u = (σϕ,u
t ⊗ σψ,u
−t ) ◦ ∆u,
t = (σψ,u
∆u ◦ σψ,u
−t) ◦ ∆u,
∆u ◦ Ru = σ ◦ (Ru ⊗ Ru) ◦ ∆u.
t ⊗ τ u
(2.1)
From these identities we easily get the following lemma (arguing as in [SaS, proof of Proposition
4]).
Lemma 2.6. For every µ ∈ Cu
Lµ ◦ σψ
Rµ ◦ σϕ
t = σψ
t = σϕ
t ◦ Lµ◦τ u
−t ,
t ◦ Rµ◦τ u
t ,
0(G)∗ and t ∈ R we have
Lµ ◦ σϕ
t = τt ◦ Lµ◦σϕ,u
,
Rµ ◦ σψ
t = τ−t ◦ Rµ◦σψ,u
Lµ ◦ R = R ◦ Rµ◦Ru .
t
t
Lµ ◦ τt = τt ◦ Lµ◦τ u
Rµ ◦ τt = τt ◦ Rµ◦τ u
t = σϕ
t = σψ
,
−t
t ◦ Lµ◦σψ,u
−t ◦ Rµ◦σϕ,u
t
,
,
2.2. Symmetry. In this short subsection we associate the symmetry properties of the Hilbert space
incarnations of the convolution operators Lµ and Rµ with the behaviour of µ with respect to the
universal unitary antipode.
Proposition 2.7. Let µ ∈ Cu
0(G)∗. For all a, b ∈ Mψ we have
(i
(1)
ψ (a))(Lµ(b)) = (i
(1)
ψ (b))(Lµ◦Ru (a)).
(2.2)
(1)
ϕ (a))(Rµ(b)) = (i
(1)
ϕ (b))(Rµ◦Ru (a)).
Similarly, for all a, b ∈ Mϕ we have (i
Proof. We establish only the first assertion. Assume first that µ ∈ L1(G). Suppose that a, b ∈
Mψ,∞ ⊆ Tψ. By Proposition 1.4 (e) and Proposition 2.4,
ψ (a) ((id ⊗ µ)∆(b)) = ψ(cid:16)(id ⊗ µ)(∆(b))σψ
= µh(ψ ⊗ id)(cid:16)∆(b)(σψ
−i/2(a) ⊗ 1)(cid:17)i
ψ (a))(µ · b) = i
(i
−i/2(a)(cid:17)
(1)
(1)
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
23
(ψ ⊗ id is defined as an operator-valued weight). A similar computation and the von Neumann
algebraic analogue of [KV1, Proposition 5.20] give
i/2(a) ⊗ 1)∆(b)(cid:17)i
ψ (b))((µ ◦ R) · a) = (µ ◦ R)h(ψ ⊗ id)(cid:16)∆(a)(σψ
−i/2(b) ⊗ 1)(cid:17)i = µh(ψ ⊗ id)(cid:16)(σψ
(1)
(i
= µh(ψ ⊗ id)(cid:16)∆(b)(σψ
−i/2(a) ⊗ 1)(cid:17)i = (i
(1)
ψ (a))(µ · b)
as in (2.2). Suppose that a ∈ Mψ and b ∈ Mψ,∞. Let (an)∞n=1 be a sequence in Mψ,∞ that converges
ultraweakly to a. Then µ · b ∈ Mψ by Proposition 2.4, and from the foregoing and Proposition 1.4
(b),
(1)
ψ (a))(µ · b) = (i
(i
(1)
ψ (µ · b))(a) = lim
n→∞
ψ (µ · b))(an) = lim
(i
n→∞
(1)
(1)
ψ (an))(µ · b)
(i
(1)
= lim
n→∞
ψ (b))((µ ◦ R) · an) = (i
(i
ψ (b))((µ ◦ R) · a)
(1)
by normality of Lµ◦R. One can now prove (2.2) for all a, b ∈ Mψ in a similar fashion.
0(G)∗ and a, b ∈ Mψ. For every ω ∈ L1(G) we have ω · b ∈ Mψ,
and as L1(G) is an ideal in Cu
For the general case, let µ ∈ Cu
0(G)∗,
(1)
ψ (a))((µ ⋆ ω) · b) = (i
ψ (b))[((ω ◦ R) ⋆ (µ ◦ Ru)) · a] = (i
ψ (a))(µ · (ω · b)) = (i
(i
= (i
(1)
(1)
(1)
ψ (b))[(ω ◦ R) · ((µ ◦ Ru) · a)].
But for all c ∈ Mψ, by Proposition 1.4 (b) and by what we have already established above,
(1)
ψ (b))[((µ ⋆ ω) ◦ R) · a]
(2.3)
so that (i
(1)
(1)
(1)
ψ (b))((ω ◦ R) · c) = (i
(i
ψ (c))(ω · b) = (i
ψ (ω · b). Together with (2.3) we get
ψ (b)) ⋆ (ω ◦ R) = i
ψ (a))(µ · (ω · b)) = (i
(i
(1)
(1)
(1)
(2.4)
be a net in L1(G) · Mψ ⊆ Mψ that converges ultraweakly to b. Then (µ ◦ Ru) · a ∈ Mψ,
ψ (ω · b))((µ ◦ Ru) · a).
Let (bι)ι∈I
and from (2.4) and Proposition 1.4 (b) we deduce that
(1)
ψ (ω · b))(c),
(1)
ψ (a))(µ · b) = lim
(i
ι∈I
= lim
ι∈I
Recall Definition 1.5.
(1)
(1)
ψ (a))(µ · bι) = lim
(i
ι∈I
ψ ((µ ◦ Ru) · a) (bι) = i
ψ (bι))((µ ◦ Ru) · a)
(i
ψ ((µ ◦ Ru) · a) (b) = i
(1)
i
(1)
(1)
ψ (b) ((µ ◦ Ru) · a) .
(cid:3)
Corollary 2.8. Let µ ∈ Cu
(i
0(G)∗. Then Lµ is KMS-symmetric with respect to ψ, namely
(1)
ψ (b))(Lµ(a)) = (i
(1)
ψ (a))(Lµ(b))
(∀a,b∈Mψ ),
if and only if µ = µ ◦ Ru, if and only if Rµ is KMS-symmetric with respect to ϕ.
The terminology introduced below follows the analogy between GNS- and KMS-implementations.
Definition 2.9. For a von Neumann algebra A and an n.s.f. weight ϕ on A, an operator P ∈ B(A)
is called GNS-symmetric with respect to ϕ if P (Nϕ) ⊆ Nϕ and
hηϕ(P a), ηϕ(b)i = hηϕ(a), ηϕ(P b)i
(∀a,b∈Nϕ).
(2.5)
24
ADAM SKALSKI AND AMI VISELTER
Condition (2.5) is equivalent to symmetry of the respective GNS-implementation: the map over the
pre-Hilbert space ηϕ(Nϕ) given by ηϕ(a) 7→ ηϕ(P a), a ∈ Nϕ.
0(G)∗♯ is w∗-dense in Cu
0(G)∗♯ is a subalgebra of Cu
Analogously to the subalgebra L1
♯ (G) ⊆ L1(G) we define Cu
0(G)∗ such that there exists a (necessarily unique) µ♯ ∈ Cu
0(G)∗ to be the set of all
0(G)∗♯ ⊆ Cu
0(G)∗ satisfying µ♯(x) = µ(Su(x))
0(G)∗ with µ 7→ µ♯ being an involution.
0(G)∗ is entire
0(G)∗ by the construction (1.7), because if µ ∈ Cu
µ ∈ Cu
for every x ∈ D(Su). Then Cu
Note that Cu
analytic with respect to the adjoint of τ u then µ ◦ Ru ∈ Cu
Proposition 2.10. Let µ ∈ Cu
and only if µ ∈ Cu
Proof. By Proposition 2.4, Rµ, resp. Lµ, is GNS-symmetric with respect to ϕ, resp. ψ, if and only if
the operator (µ⊗id)(W), resp. (id⊗µ)(V), is selfadjoint. Using the (formal) identities (Su⊗id)(W) =
W∗ (see (1.6)) and (id ⊗ Su)(V) = V∗ and the fact that the subspaces {(id ⊗ ω)(W) : ω ∈ L1( G)}
and {(ρ ⊗ id)(V) : ρ ∈ L1( G′)} are cores for Su, one can mimic the proof of [Kus2, Proposition 3.1]
to derive the assertion.
0(G)∗. Then Rµ, resp. Lµ, is GNS-symmetric with respect to ϕ, resp. ψ, if
0(G)∗♯ and (µ)♯ = µ, resp. µ ∈ Cu
0(G)∗♯ .
0(G)∗♯ and µ♯ = µ.
(cid:3)
0(G)∗+, then Rµ is GNS-symmetric if and only if Lµ is. Furthermore, in this
Remark 2.11. If µ ∈ Cu
case, both operators are KMS-symmetric (with respect to ϕ, ψ, respectively).
Indeed, under the
GNS-symmetry assumption, for every x ∈ D(τ u
−i) we have µ(x) = µ(Su(x)) and Su(x) = (Ru ◦
−i(x)). This implies by a
−i/2)(x) = (τ u
τ u
0(G) that is entire analytic with respect
standard argument that µ is invariant under τ u: every x ∈ Cu
to τ u satisfies µ(τ u
z−i(x)) for
all z ∈ C by analyticity; since the function C ∋ z 7→ µ(τ u
z (x)) is bounded on {z ∈ C : 0 ≤ ℑz ≤ 1},
it is bounded on all of C, thus it is constant; and the set of all x as above is dense in Cu
0(G)
(see also [CFK, Proposition 4.8] or [FST, Proposition 3.5]). Therefore, every x ∈ D(Su) satisfies
µ(x) = µ((τ u
−i/2), thus µ(x) = µ((Su ◦ Su)(x)) = µ(τ u
t−i(x)) for all t ∈ R, thus µ(τ u
t (x))) = µ(τ u
−i/2 ◦ Ru)(x) ∈ D(τ u
−i(τ u
−i/2 ◦ Ru)(x)) = µ(Ru(x)), yielding that µ = µ ◦ Ru.
t (x)) = µ(τ u
z (x)) = µ(τ u
2.3. Lµ and Rµ at the Hilbert space level. In this subsection we characterise (Proposition 2.12)
the operators of the form Lµ, Rµ in terms of their GNS Hilbert space versions with respect to ψ, ϕ,
respectively, introduced in Proposition 2.4. We then prove that Lµ, Rµ admit KMS Hilbert space
versions with respect to ψ, ϕ (Lemma 2.14), and characterise the operators of these forms using the
KMS language (Proposition 2.16). This is done via the KMS Hilbert space versions of Lµ, Rµ with
respect to the 'wrong' weight for suitable functionals µ. The subsection ends with two convergence
results, Lemma 2.17 and Corollary 2.19.
Proposition 2.12. Let T : L∞(G) → L∞(G) be a completely positive, normal operator.
(a) Suppose that T (Nϕ) ⊆ Nϕ and that the map ηϕ(a) 7→ ηϕ(T a), a ∈ Nϕ, is bounded. Denote
0(G)∗+, if and only if
(b) Suppose that T (Nψ) ⊆ Nψ and that the map ηψ(a) 7→ ηψ(T a), a ∈ Nψ, is bounded. Denote
0(G)∗+, if and only if
its bounded extension to L2(G) by eT ϕ. Then T has the form Rµ, µ ∈ Cu
eT ϕ ∈ L∞( G).
its bounded extension to L2(G) by eT ψ. Then T has the form Lµ, µ ∈ Cu
eT ψ ∈ L∞( G)′.
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
25
If G is amenable then the same assertions are true for normal, completely bounded T and µ ∈ Cu
0(G)∗
(not necessarily positive).
1
1
1
1
ω
1
2 ζ,η
δ
1
2 ζ,η
δ
: ζ ∈ D(δ
)(V ) : ζ ∈ D(δ
2 ), η ∈ L2(G)}. Since δ
2 ), η ∈ L2(G)}′ = {(id ⊗ ω
Proof. We prove only (a), as the proof of (b) is very similar. The assertion follows easily from
Propositions 2.4 and 2.5. Let A := {ωζ,η : ζ ∈ D(δ
2 is densely defined, A
is dense in L1(G). By Theorem 2.1 (e) and (f), proving that T has the desired form amounts to
showing that it commutes with all operators Lω, ω ∈ L1(G). By density and continuity, this holds
if and only if for every a ∈ Nϕ and ω ∈ A, we have (T ◦ Lω)(a) = (Lω ◦ T )(a), which is the same
ω = eLϕ
as eT ϕeLϕ
ωeT ϕ on ηϕ(Nϕ), equivalently everywhere. So T has the desired form if and only if eT ϕ
belongs to {eLϕ
ω : ω ∈ A}′ = {eLψ
2 ), η ∈
L2(G)}′ = L∞( G)′′ = L∞( G).
Remark 2.13. The last result is an analogue of [CFK, Theorem 7.3, (1) ⇐⇒ (4)] for general, not
necessarily compact, locally compact quantum groups. Let us explain this fact, using the conven-
tions of [CFK] freely (in particular, sesquilinear forms will be linear in the right variable and we
use the Sweedler notation). Let G be a compact quantum group, h the Haar state on C(G), and
Pol(G) ⊆ C(G) the canonical Hopf ∗-algebra. Assume that the linear map L : Pol(G) → Pol(G)
from [CFK, Theorem 7.3] extends to a normal map on L∞(G), still denoted by L, and to a bounded
map on L2(G), denoted by eL. Consider condition (4) of that Theorem, saying that
(2.6)
where m∗,Q are the sesquilinear maps from Pol(G) to Pol(G), C, respectively, given by m∗(a, b) =
a∗b and Q(a, b) := −h(a∗L(b)) (a, b ∈ Pol(G)). For ζ, η ∈ L2(G), apply ωζ,η to both sides of (2.6).
Then
((ωζ,ηm∗) ⊗ Q)(∆(a), ∆(b)) = ((ωζ,ηm∗) ⊗ Q)(a(1) ⊗ a(2), b(1) ⊗ b(2))
Q(a, b)1 = (m∗ ⊗ Q)(∆(a), ∆(b))
(∀a,b∈Pol(G)),
(cid:3)
= −(cid:10)a(1)ζ, b(1)η(cid:11) h(a∗(2)L(b(2))) = −(cid:10)a(1)ζ, b(1)η(cid:11)Dηh(a(2)),eLηh(b(2))E
= −Da(1)ζ ⊗ ηh(a(2)), (1 ⊗eL)(b(1)η ⊗ ηh(b(2)))E .
Using the identity W ∗(ξ ⊗ ηh(c)) = c(1)ξ ⊗ ηh(c(2)) for c ∈ Pol(G) and ξ ∈ L2(G), we get
((ωζ,ηm∗) ⊗ Q)(∆(a), ∆(b)) = −DW ∗(ζ ⊗ ηh(a)), (1 ⊗eL)W ∗(η ⊗ ηh(b))E .
On the other hand, Q(a, b)ωζ,η(1) = −Dζ ⊗ ηh(a), (1 ⊗eL)(η ⊗ ηh(b))E. So it is evident that (2.6)
holds in and only if 1 ⊗ eL commutes with W , that is to say, eL belongs to L∞( G)′.
We now turn to 2-integrability of the maps Rµ and Lµ. From Corollary 2.8 and Theorem A.1
0(G) then these two Markov operators are KMS-
we know that if µ is an Ru-invariant state of Cu
symmetric, thus 2-integrable, with respect to ϕ and ψ, respectively. It turns out that these assump-
tions are not necessary. Recall the notation introduced in Definition 1.5 and in Proposition 2.4,
together with the notion of the commutant quantum group.
Lemma 2.14. Let µ ∈ Cu
0(G)∗.
26
ADAM SKALSKI AND AMI VISELTER
(b) The operator Rµ, resp. Lµ, is 2-integrable with respect to ϕ, resp. ψ, and the corresponding
µ ∈ D(τ′
µ(cid:13)(cid:13)(cid:1) ≤ kµk.
= τ−i/4(eRϕ
belong to M(C0( G)), M(C0( G′)), respectively; and if µ ∈ L1(G),
µ )k ≤ max(cid:0)(cid:13)(cid:13)eRϕ
−i/4(eLψ
µ(cid:13)(cid:13)(cid:1) ≤ kµk,
µ(cid:13)(cid:13),(cid:13)(cid:13)eRϕ
= τ′
µ ).
µ
µ
µ
µ
(d) The following conditions are equivalent:
they belong to C0( G), C0( G′), respectively.
−i/2) and kτ−i/4(eRϕ
µ) and eL(2,ψ)
(a) We have eRϕ
µ ∈ D(τ−i/2), eLψ
µ(cid:13)(cid:13),(cid:13)(cid:13)eLψ
µ )k ≤ max(cid:0)(cid:13)(cid:13)eLψ
−i/4(eLψ
kτ′
Hilbert space operators are eR(2,ϕ)
,eL(2,ψ)
(c) The operators eR(2,ϕ)
• µ is invariant under τ u;
• eR(2,ϕ)
• eL(2,ψ)
symmetric with respect to ψ – then eR(2,ϕ)
µ = (µ ⊗ id)(W∗) = (id ⊗ µ)( W) and eLψ
µ) = (id⊗ µ)( W∗) =(cid:0)eRϕ
µ) = R(cid:0)(cid:0)eRϕ
µ and eL(2,ψ)
= eRϕ
= eLψ
= eRϕ
= eLψ
µ;
µ .
µ .
µ
µ
µ
µ
Thus, if µ is positive, and if Rµ is GNS-symmetric with respect to ϕ – equivalently, Lµ is GNS-
4] (indeed, it is proved there that V = WdGop, and equality for the ambient semi-universal operators
follows from [Kus2, Proposition 6.6]).
Proof. (a) By Proposition 2.4, eRϕ
µ = (id ⊗ µ)(V). Hence,
µ(cid:1)∗ by Lemma 1.17. Using the equality S = R◦ τ−i/2, we
eRϕ
µ ∈ D(S) and S(eRϕ
µ(cid:1)∗(cid:1). Consider the strip D := (cid:8)z ∈ C : − 1
2 ≤ ℑz ≤ 0(cid:9),
have eRϕ
µ ∈ D(τ−i/2) and τ−i/2(eRϕ
and define f : D → L∞( G) by f (z) := τz(eRϕ
µ), z ∈ D. Then f is bounded and ultraweakly
2 )k = (cid:13)(cid:13)eRϕ
µ(cid:13)(cid:13) for all
µ(cid:13)(cid:13) and kf (t − i
continuous on D and analytic on D◦. Furthermore, kf (t)k = (cid:13)(cid:13)eRϕ
µ(cid:13)(cid:13),(cid:13)(cid:13)eRϕ
t ∈ R. The Phragmen–Lindelöf three lines theorem implies that kf (z)k ≤ max(cid:0)(cid:13)(cid:13)eRϕ
µ(cid:13)(cid:13)(cid:1) ≤ kµk
for all z ∈ D, and in particular for z = − i
4.
µ follows by using the equalities V = W dGop and dGop = G′ [KV2, Section
The assertion about eLψ
(b) Denote by τ u∗ the automorphism group (cid:0)(τ u
t )∗(cid:1)t∈R of Cu
respect to ϕ and the corresponding operator eR(2,ϕ)
∈ B(L2(G)) equals eRϕ
and Lµ is 2-integrable with respect to ψ and the corresponding operator eL(2,ψ)
−i/4(eLψ
eLψ
(τ u∗)−i/4(µ) = τ′
We verify only the claim about Lµ, assuming that µ is as in the last paragraph. Write µ(·) : C →
t for all t ∈ R. By Lemma
0(G)∗ for the unique entire analytic function that satisfies µ(t) = µ ◦ τ u
Cu
2.6, Lµ(t) ◦ σψ
t ◦ Lµ for every t ∈ R. Consequently, if a ∈ Tψ ⊆ L∞(G), then also Lµ(a) ∈ Tψ
and (Lµ(z) ◦ σψ
z ◦ Lµ)(a) for all z ∈ C (use Proposition 2.4). Thus, by Proposition 2.4
again and Lemma 1.8 (b), for every a ∈ Mψ ∩ Tψ we have Lµ(a) ∈ Mψ ∩ Tψ and
−i/4)(a))
0(G)∗. We first prove the follow-
0(G)∗ is entire analytic with respect to τ u∗, then Rµ is 2-integrable with
(τ u∗)i/4(µ) = τ−i/4(eRϕ
µ),
∈ B(L2(G)) equals
t = σψ
z )(a) = (σψ
(2)
ψ (Lµ(a)) = ηψ((σψ
i
if µ ∈ Cu
ing claims:
µ ).
µ
µ
(2.7)
= eLψ
Let a ∈ Mψ (or even a ∈ M(2)
ψ ). Pick a net (aλ)λ∈I
ultraweakly and i
ψ (a) weakly using Proposition 1.10. Then Lµ(aλ) −−→λ∈I
ψ (aλ) −−→λ∈I
in Mψ,∞ ⊆ Mψ ∩ Tψ such that aλ −−→λ∈I
−i/4(a))) = eLψ
µ(−i/4)(ηψ(σψ
µ(−i/4)(i
(2)
i
(2)
−i/4 ◦ Lµ)(a)) = ηψ((Lµ(−i/4) ◦ σψ
(2)
ψ (a)).
a
Lµ(a)
(2)
ψ (a)) weakly by (2.7).
ultraweakly, and since eLψ
µ(−i/4) is bounded, i
(2)
ψ (Lµ(aλ)) −−→λ∈I eLψ
µ(−i/4)(i
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
27
(2)
τ′
(2)
µn)(i
µn = τ′
µ
= τ′
µ(−i/4)(i
implies that i
−i/4(eLψ
(2)
ψ (Lµ(a)) = τ′
(2)
ψ (Lµn (a)) = τ′
−i/4((id ⊗ µ)(V)) = τ′
(2)
ψ (a)), so that
−t)(V) = V
µ ), and the claim follows.
0(G)∗. Consider the sequence (µn)∞n=1 in Cu
(2)
ψ ) and i
(τ u∗)−i/4(µ). Using the formula (τ′t ⊗ τ u
Hence, Proposition 1.4 (d) implies that Lµ(a) ∈ D(i
Lµ is 2-integrable with respect to ψ and eL(2,ψ)
= eLψ
for every t ∈ R we deduce that eL(2,ψ)
We now treat the general case. Let µ ∈ Cu
0(G)∗
defined by (1.7) and its properties listed in the succeeding text. By the foregoing, for every n ∈ N,
Lµn is 2-integrable with respect to ψ and eL(2,ψ)
µ ultraweakly,
µ ) ultraweakly.
and by (a), (τ′
Let a ∈ Mψ. Then Lµn(a) −−−→n→∞
Lµ(a) ultraweakly by Lemma 2.3 (because kµnk −−−→n→∞ kµk)
(2)
ψ (a)) weakly. Thus, Proposition 1.4 (d)
and i
(2)
−i/4(eLψ
ψ (a)). This proves that Lµ is 2-integrable with respect to ψ
−i/4(eLψ
µn ))n∈N is bounded by kµk, we have τ′
−i/4(eLψ
ψ (Lµ(a)) = eLψ
−i/4(eLψ
µn). Since eLψ
−i/4(eLψ
µn ) −−−→n→∞
µn −−−→n→∞ eLψ
−i/4(eLψ
(c) Recall that τ restricts to an automorphism group of the C∗-algebra C0( G), and also R(C0( G)) =
µ ∈ M(C0( G)), so also τ−i/2(eRϕ
µ) ∈ M(C0( G)). Similarly, if µ ∈ L1(G), then eRϕ
and eL(2,ψ)
µ) = R(cid:0)(cid:0)eRϕ
C0( G). By Proposition 2.4 we have eRϕ
Therefore, [Kus1, Proposition 2.44] implies that eRϕ
extension of τ from C0( G) to M(C0( G)), and the two meanings of τ−i/2(eRϕ
µ, R(cid:0)(cid:0)eRϕ
eR(2,ϕ)
(d) If µ is invariant under τ u, then the claim in the proof of (b) implies that eR(2,ϕ)
Conversely, by (b), the assumption eR(2,ϕ)
τ, and the identity (τ u
Remark 2.11, this is the case when µ ∈ Cu
µ(cid:1)∗(cid:1) ∈ M(C0( G)).
µ belongs to D(τ−i/2) in the sense of the strict
µ) coincide. As a result,
µ(cid:1)∗(cid:1) ∈ C0( G), and we
= eRϕ
µ.
µ = (µ ⊗ id)(W∗) is invariant under
t ⊗ τt)(W) = W for every t ∈ R yields that µ is invariant under τ u. Lastly, by
deduce from [Kus1, Proposition 1.24] the desired conclusion.
0(G)∗+ and Rµ is GNS-symmetric with respect to ϕ.
µ means that eRϕ
= τ−i/4(eRϕ
ψ (a)) −−−→n→∞
µ )(i
−i/4(eLψ
−i/4(eLψ
= eRϕ
µ
= τ′
µ ).
τ′
µ )(i
µ
(cid:3)
µ
µ
µ
For brevity we formulate the next result only for the left convolution operators; it also has a
natural right version. Recall that ν denotes the scaling constant of G. Define an injective, positive,
ηϕ(τt(a)) for all t ∈ R and a ∈ Nϕ. The
selfadjoint operator P on L2(G) by P it(ηϕ(a)) := νt/2
operators δ, P commute and (τt)t∈R = (Ad(P it))t∈R.
Lemma 2.15. Let e(·), f (·) be the spectral measures of P, δ, respectively. Define a subspace Z of L2(G)
by
Z :=
e ([1/n, n]) f ([1/m, m]) L2(G).
∞[n,m=1
(a) Z is dense in L2(G), and for all z, w ∈ C, Z ⊆ D(P zδw) and P zδwZ is dense in L2(G).
(b) For each ζ, η ∈ Z, Lωζ,η is 2-integrable with respect to ϕ and eL(2,ϕ)
.
Proof. (a) Injectivity of P and δ implies the first assertion. Since, in addition, P and δ also commute,
the second assertion follows. Notice that for every ξ ∈ Z, the function C2 ∋ (z, w) 7→ P zδwξ =
δwP zξ is (well defined and) entire analytic.
ωζ,η = eLψ
1
4 ζ,P
1
4 η
1
4 δ
− 1
4 δ
ω
P
(b) We first show that for every a ∈ Mϕ ∩ Tϕ we have
i(2)
ϕ (Lωζ,η (a)) = eLψ
ω
P
− 1
4 δ
1
4 ζ,P
1
4 δ
1
4 η
(i(2)
ϕ (a)).
(2.8)
28
ADAM SKALSKI AND AMI VISELTER
So let a ∈ Mϕ ∩ Tϕ. For every t ∈ R, using that σψ
σψ
s (δit) = νistδit for s ∈ R, and Lω ◦ σψ
t (·)δ−it, ∆(δit) = δit ⊗ δit, σϕ
t ◦ Lω◦τ−t for ω ∈ L1(G) (Lemma 2.6), we have
t (·) = δitσϕ
t = σψ
s (δit) =
σϕ
t (Lωζ,η (a)) = δ−itσψ
t (Lωζ,η (a))δit = σψ
t (δ−itLωζ,η (a)δit)
= σψ
t (Lωδ−itζ,δ−itη
(δ−itaδit)) = LωP −itδ−itζ,P −itδ−itη
(σψ
t (δ−itaδit))
= LωP −itδ−itζ,P −itδ−itη
(σϕ
t (a)).
Hence, Lωζ,η (a) is entire analytic with respect to σϕ, and for all z ∈ C
z (a)).
σϕ
z (Lωζ,η (a)) = LωP −iz δ−iz ζ,P −iz δ−iz η
(σϕ
Therefore, using Proposition 2.5 (a) and (b) and Lemma 1.8 (b) we obtain Lωζ,η (a) ∈ Mϕ ∩Tϕ and
ϕ (Lωζ,η (a)) = ηϕ(σϕ
i(2)
−i/4(Lωζ,η (a))) = ηϕ(Lω
− 1
4 δ
P
− 1
4 ζ,P
1
4 δ
1
4 η
(σϕ
−i/4(a)))
(i(2)
= eLψ
proving (2.8). Relying on this, on Proposition 1.10 and on Proposition 1.4 (d) again we conclude
that (2.8) holds for all a ∈ Mϕ.
(cid:3)
−i/4(a))) = eLψ
(ηϕ(σϕ
ϕ (a)),
1
4 ζ,P
1
4 ζ,P
1
4 η
1
4 η
1
4 δ
1
4 δ
− 1
− 1
4 δ
4 δ
ω
ω
P
P
The following result provides a key characterisation of the KMS-implementations of the convolu-
tion operators.
0(G)∗ (not necessarily positive).
If G is amenable then the same assertions are true for normal, completely
Proposition 2.16. Let T : L∞(G) → L∞(G) be a completely positive, normal operator that is 2-
integrable with respect to ϕ, resp. ψ, and let eT (2,ϕ), resp. eT (2,ψ), be the associated bounded operator
0(G)∗+, if and only if eT (2,ϕ) ∈ L∞( G),
on L2(G). Then T = Rµ, resp. T = Lµ, for some µ ∈ Cu
resp. eT (2,ψ) ∈ L∞( G)′.
bounded T and µ ∈ Cu
Proof. We prove the assertions concerning Rµ. As in the proof of Proposition 2.12, by Theorem 2.1,
T being of the desired form is equivalent to T commuting with the operators Lω, ω ∈ L1(G). We
use Lemma 2.15 and its notation. Put A := {ωζ,η : ζ, η ∈ Z}. The set A dense in L1(G) because Z
is dense in L2(G). Hence, the operator T has the desired form if and only if it commutes with the
operators Lω, ω ∈ A. Let ζ, η ∈ Z. Then T commutes with Lωζ,η if and only if T Lωζ,η = Lωζ,η T on
M(2)
ϕ by ultraweak density of M(2)
ϕ in L∞(G). By Remark 1.12, injectivity of i(2) (Proposition 1.4
(d)) and continuity, this is equivalent to eT (2,ϕ)eL(2,ϕ)
ϕ ), thus everywhere
on L2(G) by density of this subspace (Proposition 1.4 (c)). So T has the desired form if and only
if eT (2,ϕ) commutes with B := {eLψ
4 Z in
L2(G) and Proposition 2.4 imply that B is dense in L∞( G)′ in the weak operator topology, and the
assertion follows.
(cid:3)
ωζ,η = eL(2,ϕ)
: ζ, η ∈ Z}. Density of P − 1
ωζ,η eT (2,ϕ) on i(2)(M(2)
4 Z and of P
1
4 ζ,P
1
4 δ
4 δ
1
4 η
1
4 δ
− 1
4 δ
ω
P
1
1
We now supplement Lemma 2.3 by adding a few more conditions.
Lemma 2.17. Let (µi)i∈I be a bounded net in Cu
conditions:
0(G)∗, and let µ ∈ Cu
0(G)∗. Consider the following
(a) µi −−→i∈I
µ in the w∗-topology;
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
29
µ
0(G)∗
µ
µ
0(G).
i(2)
ϕ (x), i(2)
ν
ϕ (Rν (x)), i(2)
µi
(cid:1)i∈I
is bounded and i
(1)
ϕ (Rν(x)) · i(2)
ultraweakly in L∞( G);
strictly in M(C0( G)).
µ ultraweakly in L∞( G);
µ strictly in M(C0( G));
is bounded, so is (cid:0)eRϕ
is bounded by Lemma 2.14, and so (c') =⇒ (c).
µi(cid:1)i∈I
µi −−→i∈I eR(2,ϕ)
µi −−→i∈I eRϕ
µi −−→i∈I eRϕ
µi −−→i∈I eR(2,ϕ)
µi −−→i∈I eR(2,ϕ)
(b) eRϕ
(b') eRϕ
(c) eR(2,ϕ)
(c') eR(2,ϕ)
Then (a) ⇐⇒ (b) ⇐⇒ (c), and they are implied by either of (b'), (c'). If kµik = kµk for all i ∈ I, then
all these conditions are equivalent. Similar results hold for operators of the form Lµ.
Proof. The equivalence (a) ⇐⇒ (b) is trivial from Proposition 2.4 because (cid:8)(id ⊗ ω)(W) : ω ∈
L1( G)(cid:9) is dense in Cu
, and the implication (b') =⇒ (b) follows. Similarly,
Since (µi)i∈I
(cid:0)eR(2,ϕ)
(c) =⇒ (a) Assume that eR(2,ϕ)
and x, y ∈ Mϕ,
ϕ (y)(cid:11) =(cid:10)i(2)
(cid:10)eR(2,ϕ)
Since(cid:0)Rµi(cid:1)i∈I
(2.9) that Rµi(x) −−→i∈I
Rµ(x) ultraweakly for every x ∈ Mϕ. This implies that µi −−→i∈I
w∗-topology by Lemma 2.3 implication (c) =⇒ (a), because Mϕ ∩ C0(G) is dense in C0(G).
µ ultraweakly in L∞( G), then the boundedness of (cid:0)eR(2,ϕ)
(cid:1)i∈I
(b) =⇒ (c) If eRϕ
that eR(2,ϕ)
µ) = eR(2,ϕ)
τ−i/4(eRϕ
µi = τ−i/4(eRϕ
Assume that kµik = kµk for all i ∈ I and that µi −−→i∈I
µ, in the w∗-topology.
Since eRϕ
0(G)∗, it follows from [RuV, Theorem 4.6, (i) =⇒ (ii)] that
ν = ((ν ⊗ id)(W))∗ for all ν ∈ Cu
eRϕ
µi −−→i∈I eRϕ
µ strictly in M(C0( G)), showing that (a) =⇒ (b'). We now prove that (a) =⇒ (c'). For
every ω ∈ L1(G) we have(cid:13)(cid:13)(cid:0)eR(2,ϕ)
(cid:1)eR(2,ϕ)
µi −eR(2,ϕ)
0 by Lemma
2.14 and Lemma 2.3 implication (a) =⇒ (e) in its Lν-version, and similarly for the opposite product.
Writing τ−i/4 for the generator in the C∗-algebraic sense, the subspace (cid:8)eRϕ
ω = (ω ⊗ id)(W ∗) : ω ∈
L1(G)(cid:9) ⊆ D(τ−i/4), being norm-dense in C0( G) and invariant under the automorphism group τ of
ω) : ω ∈ L1(G)(cid:9) is norm-dense in C0( G).
C0( G), is a core for τ−i/4. As a result, (cid:8)eR(2,ϕ)
Boundedness of(cid:0)eR(2,ϕ)
ultraweakly. By Proposition 1.4 (b), for all ν ∈ Cu
ϕ (y∗)(cid:17) .
(2.9)
ϕ (Mϕ) is dense in L1(G) by Proposition 1.4 (c), we deduce from
µ in the
implies that it converges strictly to eR(2,ϕ)
ω⋆(µi−µ)(cid:13)(cid:13) ≤ kω ⋆ (µi − µ)k −−→i∈I
ultraweakly (Lemma 2.14).
µ, equivalently µi −−→i∈I
ϕ (y∗)(cid:17) = tr(cid:16)Rν (x) · i(1)
ϕ (y)(cid:11) = tr(cid:16)i(2)
µi −−→i∈I eRϕ
µi) −−→i∈I
(cid:13)(cid:13) =(cid:13)(cid:13)eR(2,ϕ)
= τ−i/4(eRϕ
'deformed representations'.
We now establish an estimate connecting the distance of the KMS-version of the convolution
operator from the identity to the distance between the underlying state and the co-unit. The proof
follows that of the main part of [DSV, Theorem 6.1, (c) =⇒ (a)], where a corresponding result
was shown for maps of the forms eRϕ
. Roughly speaking, we work with
(cid:13)(cid:13)eL(2,ψ)
µ − 1(cid:13)(cid:13) ≤ δ. Then kµ − ǫk ≤ 2
0(G) and 0 < δ < 1
Proposition 2.18. Let µ be a state of Cu
√2δ
.
1−√2δ
,eL(2,ψ)
2 be such that (cid:13)(cid:13)eR(2,ϕ)
µ in place of eR(2,ϕ)
− 1(cid:13)(cid:13) ≤ δ or
µ,eLψ
µ
µ
µ
µ
in M(C0( G)).
(cid:3)
µ
ω
µi
implies
µ
ω
µi
(cid:1)i∈I
30
ADAM SKALSKI AND AMI VISELTER
ν
k ≤ kνk. Let ω ∈ L1( G). The linear functional given by
Proof. Let (Hµ, πµ, ξµ) be the GNS construction for µ. Recall from Lemma 2.14 that for every ν ∈
0(G)∗ we have keR(2,ϕ)
Cu
is of norm at most kωk, and thus defines an element xω ∈ B(Hµ) of norm at most kωk. The map
L1( G) ∋ ω 7→ xω is evidently linear. Let ω1, ω2 ∈ L1( G). Assuming that ωj ∈ D((τ∗)−i/4) and
writing ω′j for the closure of ωj ◦ τ−i/4 (j = 1, 2), for all ρ ∈ B(Hµ)∗ we have
ρ◦πµ)) = (ω ◦ τ−i/4)[(id ⊗ (ρ ◦ πµ))( W)]
B(Hµ)∗ ∋ ρ 7→ ω(eR(2,ϕ)
ρ◦πµ ) = ω(τ−i/4(eRϕ
ρ(xω1xω2) = (ρ · xω1)(xω2) = (ω2 ◦ τ−i/4){(id ⊗ ρ)[(1 ⊗ xω1)(id ⊗ πµ)( W)]}
= ρ[xω1(ω′2 ⊗ πµ)( W)] = [(ω′2 ⊗ πµ)( W) · ρ](xω1)
= ω′1{(id ⊗ ρ)[(id ⊗ πµ)( W)(1 ⊗ (ω′2 ⊗ πµ)( W))]}
(unconventionally using '·' instead of juxtaposition). The identity ( ∆ ⊗ id)( W) = W13 W23 implies
that
(id ⊗ πµ)( W)(1 ⊗ (ω′2 ⊗ πµ)( W)) = (id ⊗ ω′2 ⊗ πµ)( W13 W23) = (id ⊗ ω′2 ⊗ πµ)( ∆ ⊗ id)( W).
From the last two formulas we get
ρ(xω1xω2) = ω′1[(id ⊗ ρ)(id ⊗ ω′2 ⊗ πµ)( ∆ ⊗ id)( W)]
= (ω′1 ⋆ ω′2)[(id ⊗ (ρ ◦ πµ))( W)].
Using the equality (τt ⊗ τt) ◦ ∆ = ∆ ◦ τt valid for all t ∈ R, we obtain
ρ(xω1xω2) = ((ω1 ⋆ ω2) ◦ τ−i/4)[(id ⊗ (ρ ◦ πµ))( W)] = ρ(xω1⋆ω2),
proving that xω1xω2 = xω1⋆ω2. Since L1( G) ∋ ω 7→ xω is continuous and D((τ∗)−i/4) is dense in
L1( G), we infer that xω1xω2 = xω1⋆ω2 for every ω1, ω2 ∈ L1( G).
By assumption, for every state ω ∈ L1( G),
kxωξµ − ξµk2 ≤ 2(1 − ℜhxωξµ, ξµi) = 2ℜω(1 − eR(2,ϕ)
µ
) ≤ 2δ.
As a result, the unique vector ξ ∈ Hµ of minimal norm in the (convex) closure C of {xωξµ : ω ∈
L1( G) is a state} satisfies kξ − ξµk ≤ √2δ. By the foregoing, C is globally invariant under each of
the contractions xω for a state ω ∈ L1( G), and so ξ is invariant under each of these operators. That
is, for every state ω ∈ L1( G),
ω(kξk2 1) = kξk2 = ωξ(xω) = (ω ◦ τ−i/4)[(id ⊗ (ωξ ◦ πµ))( W)].
Consequently, τ−i/4[(id ⊗ (ωξ ◦ πµ))( W)] = kξk2 1, that is, (id ⊗ (ωξ ◦ πµ))( W) = kξk2 1. Hence
kξk ξ ◦ πµ = ǫ, and as kµ − ǫk = k(ω 1
, the proof
ω 1
is complete.
(cid:3)
kξk ξ − ωξµ) ◦ πµk ≤ k 1
kξk
ξ + ξµkk 1
kξk
ξ − ξµk ≤ 2
√2δ
1−√2δ
The following result is the norm analogue of Lemmas 2.3 and 2.17.
Corollary 2.19. Let (µi)i∈I be a net of states of Cu
0(G). Then the following conditions are equivalent:
(a) µi −−→i∈I
(b) Ru
µi −−→i∈I
ǫ in Cu
0(G)∗-norm;
id in B(Cu
0(G))-norm;
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
31
1 in L∞( G)-norm;
(c) eRϕ
µi −−→i∈I
(d) eR(2,ϕ)
µi −−→i∈I
A similar statement holds true for maps of the form Lµ.
Proof. It is immediate that (a) =⇒ (b), (c). From Lemma 2.14 we have (a) =⇒ (d).
1 in L∞( G)-norm.
(b) =⇒ (a) Use that ǫ ◦ Ru
(c) =⇒ (a) This follows from the proof of [DSV, Theorem 6.1, (c) =⇒ (a)] (cf. the proof of
ν = ν for every ν ∈ Cu
0(G)∗.
Proposition 2.18 and ignore τ).
(d) =⇒ (a) Use Proposition 2.18.
(cid:3)
Question 2.20. While it is obvious that the equivalent conditions of Corollary 2.19 imply that Rµi −−→i∈I
id in B(L∞(G))-norm, we ask whether the converse is true. The answer is clearly affirmative when G
is co-amenable.
3. SEMIGROUPS OF CONVOLUTION OPERATORS
In this section we discuss the key notion for our paper, namely that of convolution semigroups,
and use the results of the last section and of the Appendix to obtain our central result, Theorem 3.4.
Definition 3.1 ([LS1]). Let G be a locally compact quantum group. A convolution semigroup of
0(G)∗+ such that µt ⋆ µs = µt+s for
positive functionals on Cu
ǫ in the
all t, s ≥ 0 and µ0 = ǫ. It is called w∗-continuous (or pointwise continuous) if µt −−−→t→0+
w∗-topology of Cu
0(G) (or on G) is a family (µt)t≥0 in Cu
ǫ in norm.
0(G)∗, and norm continuous if µt −−−→t→0+
The first result is essentially an immediate consequence of the facts established in the last section.
Theorem 3.2. Let G be a locally compact quantum group. There exist 1 − 1 correspondences between:
(a) w∗-continuous convolution semigroups (µt)t≥0 of states of Cu
(b) C0-semigroups (T u
ν, ν ∈ Cu
0(G);
t )t≥0 of completely positive maps of norm 1 on Cu
operators Lu
0(G)∗;
0(G) that commute with the
(c) semigroups (Tt)t≥0 of normal, unital, completely positive maps on L∞(G) that are point–
(d) C0-semigroups (Mt)t≥0 of norm 1 right module maps on L1(G) with completely positive ad-
ultraweakly continuous at 0+, and that satisfy ∆ ◦ Tt = (Tt ⊗ id) ◦ ∆ for every t ≥ 0;
joints.
t = Ru
They are given by Tt = Rµt, T u
µt and Mt = (Tt)∗. All semigroups are continuous in the respective
topologies on all of R+ (for instance, (Tt)t≥0 is a C∗0 -semigroup). Norm continuity of the semigroups
in (a) and (b) are equivalent, and the same is true for the semigroups in (c) and (d), and the former
implies the latter. Similar statements hold for maps of the form Lµ.
Proof. The correspondences exist by Theorem 2.1 (e) and Lemma 2.3. Since C0-semigroups (such
as (Mt)t≥0) are point–norm continuous on all of R+, all semigroups considered in the theorem are
continuous in the respective topologies on all of R+. By Corollary 2.19 equivalence (a) ⇐⇒ (b),
norm continuity of (a) and (b) are the equivalent, and this obviously implies that of (c), equivalently
(d).
(cid:3)
32
ADAM SKALSKI AND AMI VISELTER
Remark 3.3. The above Theorem 3.2 could be stated slightly more generally, namely, replace 'states'
by 'contractive positive functionals' in (a), 'unital' by 'contractive' in (c), and 'of norm 1' by 'con-
tractive' in (b) and (d) (where 'contractive' means 'of norm at most 1' throughout). This statement
easily reduces to the one we used, because by Lemma 2.3, all assumptions imply that (µt)t≥0 is
w∗-continuous; and since (µt)t≥0 consists of contractions, the semigroup (µt(1))t≥0 = (kµtk)t≥0 =
t k)t≥0 = (kTtk)t≥0 = (kMtk)t≥0 (having extended µt strictly to the multiplier algebra on the
(kT u
left) is continuous at 0+, hence it has the form (cid:0)etc(cid:1)t≥0, c ≤ 0. Thus, by normalising, we may
assume that (µt)t≥0 consists of states, so that Lemma 2.3 implication (a) =⇒ (e) applies.
We now prove Theorem 0.1 presented in the Introduction.
on Cu
0(G);
KMS-symmetric with respect to ϕ and satisfy ∆ ◦ Tt = (Tt ⊗ id) ◦ ∆ for every t ≥ 0;
Theorem 3.4. Let G be a locally compact quantum group. There exist 1 − 1 correspondences between:
(a) w∗-continuous convolution semigroups (µt)t≥0 of Ru-invariant contractive positive functionals
(b) C∗0 -semigroups (Tt)t≥0 of normal, contractive, completely positive maps on L∞(G) that are
(c) C0-semigroups (St)t≥0 of selfadjoint completely Markov operators on L2(G) with respect to ϕ
(d) completely Dirichlet forms Q with respect to ϕ that are invariant under U (L∞( G)′).
The correspondences are given by Tt = Rµt, St = eR(2,ϕ)
and Corollary A.8. Also, norm continuity of
(µt)t≥0, of (Tt)t≥0 and of (St)t≥0 are equivalent, and these are also equivalent to boundedness of Q.
Similar statements hold for maps of the form Lµ with invariance under U (L∞( G)) in (d).
Proof. The correspondence between (a) and (b) is given by Theorem 3.2, Remark 3.3 and Corollary
2.8.
that belong to L∞( G);
µt
0(G)∗+ (namely, ∆ ◦ Tt = (Tt ⊗ id) ◦ ∆) if and only if eT (2,ϕ)
The correspondence between (b), (c) and (d) relies on Corollary A.8: completely Dirichlet forms
Q with respect to ϕ correspond to semigroups (Tt)t≥0 of normal, completely positive contractions
on L∞(G) that are KMS-symmetric with respect to ϕ and satisfy that R+ ∋ t 7→ Tta is ultraweakly
continuous for all a ∈ Mϕ, and to C0-semigroups (St)t≥0 of selfadjoint completely Markov operators
(cid:1)t≥0 by −A; then A is a
on L2(G) with respect to ϕ. Denote the generator of (St)t≥0 = (cid:0)eT (2,ϕ)
positive selfadjoint operator on L2(G) and Q = (cid:13)(cid:13)A1/2·(cid:13)(cid:13)2. By Proposition 2.16, for each t ≥ 0, Tt
∈ L∞( G).
is of the form Rµt for µt ∈ Cu
Thus, the semigroup (Tt)t≥0 has the form(cid:0)Rµt(cid:1)t≥0 with (µt)t≥0 a convolution semigroup of positive
(cid:1)t≥0 commutes
functionals on Cu
with the unitaries of L∞( G)′, if and only if A (equivalently, A1/2) commutes with the unitaries of
L∞( G)′, that is, Q is invariant (as a quadratic form) under all unitaries of L∞( G)′. When this is
the case, for each t ≥ 0, Tt = Rµt is contractive, thus so is µt by Theorem 2.1 (e). By Lemma
2.3 implication (c) =⇒ (a), which applies because Mϕ ∩ C0(G) is dense in C0(G), the semigroup
(µt)t≥0 is w∗-continuous. So by Remark 3.3, one deduces that (Tt)t≥0 = (Rµt )t≥0 is a C∗0 -semigroup;
and furthermore, we can and shall assume below that (µt)t≥0 consists of states.
(cid:1)t≥0 is norm continuous
As for norm continuity, the C0-semigroup (St)t≥0 =(cid:0)eT (2,ϕ)
if and only if −A, equivalently Q, is bounded, and by Corollary 2.19, this is the same as (µt)t≥0
0(G) if and only if each of the elements of the C0-semigroup(cid:0)eT (2,ϕ)
(cid:1)t≥0 =(cid:0)eR(2,ϕ)
µt
t
t
t
t
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
33
µt
being norm continuous. In this case, (Rµt)t≥0 is clearly norm continuous. Conversely, for every
t ≥ 0, by Proposition 2.4 and the fact that Rµt is a (bounded) normal KMS-symmetric Markov
operator, Rµt is 1-integrable and eR(1,ϕ)∗
= Rµt (see also Theorem A.1). Hence, Rµt − id is adjoint
(1,ϕ)∗ = Rµt − id. As a result, [GL2, Proposition 4.1]
preserving and 1-integrable, and ^(Rµt − id)
yields that (cid:13)(cid:13)eR(2,ϕ)
µt − 1B(L2(G))(cid:13)(cid:13) = (cid:13)(cid:13) ^(Rµt − id)
(2,ϕ)(cid:13)(cid:13) ≤ (cid:13)(cid:13)Rµt − id(cid:13)(cid:13). So norm continuity of (Rµt )t≥0
implies that of (eR(2,ϕ)
Definition 3.5 ([LS1, p. 333, Definition]). Let (µt)t≥0 be a w∗-continuous convolution semigroup
0(G). Its generating functional is the (generally unbounded) functional
of positive functionals on Cu
γ over Cu
We recall the notion of the generating functional.
)t≥0.
(cid:3)
µt
0(G) given by
γ(x) := lim
t→0+
with maximal domain D(γ) consisting of all x ∈ Cu
µt(x) − ǫ(x)
0(G) for which this limit exists.
t
By [LS1, p. 333, Remarks], D(γ) is dense in Cu
0(G), and by [LS1, Theorem 3.7], (µt)t≥0 is uniquely
determined by γ. Observe that γ is hermitian: for every x ∈ D(γ) we have x∗ ∈ D(γ) and γ(x∗) =
γ(x), and it is conditionally positive: γ(x) ≥ 0 for all x ∈ D(γ) ∩ Cu
0(G)+ ∩ ker ǫ. If G is compact
and (µt)t≥0 consists of states, then additionally 1 ∈ D(γ) and γ(1) = 0.
Lemma 3.6. In the notation of Definition 3.5, γ is bounded ⇐⇒ γ is everywhere defined ⇐⇒
(µt)t≥0 is norm continuous.
Proof. This is mostly [LS1, Theorem 3.7], the only missing part is that γ being everywhere defined
implies the other conditions. But when this happens, the restriction of γ to ker ǫ is positive, thus
bounded. Therefore, γ is bounded on Cu
(cid:3)
0(G).
To describe the relation between the generators appearing in Theorem 3.4 we need the following
elementary lemma.
2
t
t
and
lim
t→0+
Lemma 3.7. Let A be a (generally unbounded) positive selfadjoint operator on a Hilbert space H.
Consider the C0-semigroup (e−tA)t≥0. Then
D(A1/2) =(cid:8)ζ ∈ H :
(cid:13)(cid:13)(cid:13)A1/2ζ(cid:13)(cid:13)(cid:13)
ωζ(cid:0) 1
(1 − e−tA)(cid:1) =Z[0,∞)
ωζ(cid:0) 1
(1 − e−tA)(cid:1) exists(cid:9)
(∀ζ∈D(A1/2)).
Proof. Write e(·) for the spectral measure of A. Let ζ ∈ H. For every t > 0,
(1 − e−tλ)he(dλ)ζ, ζi .
ωζ(cid:0) 1
(1 − e−tA)(cid:1)
= lim
t→0+
t (1 − e−tA)) exists, then by Fatou's lemma, R[0,∞) λhe(dλ)ζ, ζi is dominated by this
If limt→0+ ωζ( 1
limit, thus ζ ∈ D(A1/2). On the other hand, if ζ ∈ D(A1/2), namelyR[0,∞) λhe(dλ)ζ, ζi < ∞, then
relying on the inequality 1
t (1 − e−tλ) ≤ λ for t > 0 and λ ≥ 0, Lebesgue's dominated convergence
theorem yields that limt→0+ ωζ( 1
t (1 − e−tA)) =R[0,∞) λhe(dλ)ζ, ζi =(cid:13)(cid:13)A1/2ζ(cid:13)(cid:13)2.
1
t
(cid:3)
t
34
ADAM SKALSKI AND AMI VISELTER
We are now ready to provide the connection mentioned before the last lemma.
and
and
Proposition 3.8. In the context of Theorem 3.4, let γ denote the generating functional of (µt)t≥0. Then
i/4(cid:0)(ωζ ⊗ id)( W)(cid:1)(cid:3)
D(Q) =(cid:8)ζ ∈ L2(G) : τ u
Qζ = −γ(cid:2)τ u
i/4(cid:0)(ωζ ⊗ id)( W)(cid:1) ∈ D(γ)(cid:9)
(∀ζ∈D(Q)).
(cid:1)t≥0 by −A. By Lemma 3.7,
Proof. Denote again the generator of(cid:0)eR(2,ϕ)
(1−e−tA)(cid:1) exists(cid:9) =(cid:8)ζ ∈ L2(G) :
ωζ(cid:0) 1
D(Q) = D(A1/2) =(cid:8)ζ ∈ L2(G) :
t (ǫ−µt)(cid:1)
ωζ(cid:0) 1
ωζ(cid:0)eR(2,ϕ)
(1 − e−tA)(cid:1) = lim
t→0+
0(G)∗ and ω ∈ L1( G) we have ω(eR(2,ϕ)
By Definition 3.5, it suffices to observe that for each ν ∈ Cu
) =
i/4(cid:0)(ω ⊗ id)( W)(cid:1)(cid:3); and this is indeed true because eR(2,ϕ)
ν(cid:2)τ u
= τ−i/4(eRϕ
ν ) = τ−i/4((id ⊗ ν)( W)) by
Proposition 2.4 and Lemma 2.14 (b), so the desired statement is a consequence of the fact that
(ω ⊗ id)( W) ∈ D(τ u
t )( W) = W for every t ∈ R.
(cid:3)
Remark 3.9 (compare Remark 2.13). The last result generalises the analysis in [CFK, Theorem 7.1
and the preceding text], which treats the case when G is compact, and only gives a formula for the
values of Q at i(2) (Pol(G)), where Pol(G) is the canonical Hopf ∗-algebra that is dense in C(G).
Qζ =(cid:13)(cid:13)(cid:13)A1/2ζ(cid:13)(cid:13)(cid:13)
i/4) and the identity (τt ⊗ τ u
2
= lim
t→0+
(∀ζ∈D(Q)).
µt
t
t
1
ν
lim
t→0+
lim
t→0+
ωζ(cid:0)eR(2,ϕ)
t (ǫ−µt)(cid:1) exists(cid:9)
1
ν
4. PROPERTY (T) AND THE HAAGERUP PROPERTY
In this section we employ the results of the previous sections to characterise, for arbitrary second
countable locally compact quantum groups, Property (T) and the Haagerup Property in terms of
w∗-continuous convolution semigroups of states. For locally compact groups, these characterisa-
tions are well-known and of fundamental importance, but in the quantum group setting they were
hitherto established only for discrete quantum groups.
Lack of Property (T) is characterised in Subsection 4.1 by the existence of certain unbounded
generating functionals, that is: certain w∗-continuous, but not norm continuous, convolution semig-
roups of states (Theorem 4.6). This generalises [AkW, Theorem 3], [Jol, Theorem 1.2] and [BHV,
Theorems 2.10.4 and 2.12.4] for locally compact groups, and [Kye, Theorem 5.1] for unimodular
discrete quantum groups (see also [DSV, Theorem 6.8]). In order to achieve this 'algebra level' res-
ult, we first 'go down' to the Hilbert space level, where more techniques are available, and then 'go
back up' to the algebra level. These transitions are possible thanks to the Dirichlet forms machinery.
Our approach is very different from that of [Kye], and is closer to those of [AkW, Jol, DSV].
In the main result of Subsection 4.2, Theorem 4.12, we characterise the Haagerup Property by the
existence of certain w∗-continuous convolution semigroups whose corresponding KMS Hilbert space
level counterparts 'vanish at infinity' in positive time. This condition should be viewed as properness
of the associated generating functional. Our result extends [AkW, Theorem 10] for locally compact
groups and [DFSW, Theorem 7.18] for discrete quantum groups. The approach of our proof is
different from that of [AkW, DFSW], and instead relies on Proposition 4.10: roughly speaking,
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
35
instead of seeking a generator with the desirable properties, we construct the semigroup directly,
following the technique of [Sau], applied later for example in [CaS]. We note that Theorem 4.12
compares with the characterisation of the Haagerup Property of von Neumann algebras obtained
recently in [CaS, Theorem 6.7].
We need first to recall some terminology concerning quantum group representations.
Definition 4.1 ([DFSW, Definitions 3.3 and 4.1]). Let G be a locally compact quantum group and
U ∈ M(C0(G) ⊗min K(H)) be a unitary representation of G on a Hilbert space H.
(a) A vector ζ ∈ H is invariant under U if U (η ⊗ ζ) = η ⊗ ζ for all η ∈ L2(G).
(b) U has almost-invariant vectors if there is a net (ζi)i∈I
of unit vectors in H such that for all
η ∈ L2(G), kU (η ⊗ ζi) − η ⊗ ζik −−→i∈I
C0(G).
0.
(c) U is mixing if 'it has C0-coefficients', namely, if for every ζ, ξ ∈ H, we have (id ⊗ ωζ,ξ)(U ) ∈
4.1. Property (T). Property (T) for discrete quantum groups was first introduced in [Fim]; for
locally compact quantum groups it was formally defined in [DFSW]. For recent developments we
refer the reader to [DSV, BrK].
Definition 4.2. A locally compact quantum group has Property (T) if each of its unitary represent-
ations that has almost-invariant vectors admits a non-zero invariant vector.
The following lemma is elementary.
Lemma 4.3. Let (ai)i∈I and (bi)i∈I be nets of contractive operators on a uniformly convex Banach
space (e.g., a Hilbert space). If 1
1.
2 (ai + bi) −−→i∈I
1, then ai −−→i∈I
1 and bi −−→i∈I
We are ready to provide the following strengthening of [DSV, Theorem 6.1], involving the KMS-
implementations.
Theorem 4.4. For a locally compact quantum group G the following conditions are equivalent:
0(G), if (µi)i∈I converges to ǫ in the w∗-topology, then it
converges to 1 in the strict topology of
0(G);
in L∞( G) converges in the weak
converges to ǫ in norm;
(a) G has Property (T);
(b) for every net (µi)i∈I of states of Cu
(b') same as (b) but only for nets (µi)i∈I of Ru-invariant states of Cu
(c) for every net (µi)i∈I of states of Cu
(c') same as (c) but only for nets (µi)i∈I of Ru-invariant states of Cu
(d) for every net (µi)i∈I of states of Cu
(d') same as (d) but only for nets (µi)i∈I of Ru-invariant states of Cu
0(G), if (eRϕ
0(G), if the net (eR(2,ϕ)
M(C0( G)), then it converges to 1 in norm;
µi)i∈I
)i∈I
µi
0(G);
0(G).
(equivalently, strong) operator topology to 1, then it converges to 1 in norm;
Moreover, if G is second countable, one can replace 'net' by 'sequence' throughout. Similar assertions
hold for operators of the form Lµ, µ ∈ Cu
Proof. The equivalence of (a), (b) and (c) was proved in [DSV, Theorem 6.1]. The equivalences
(b) ⇐⇒ (c) ⇐⇒ (d) and (b') ⇐⇒ (c') ⇐⇒ (d') are consequences of Lemma 2.17 and Corollary
2.19. The implications (b) =⇒ (b'), (c) =⇒ (c') and (d) =⇒ (d') are trivial.
0(G)∗.
36
ADAM SKALSKI AND AMI VISELTER
Cu
1
ν◦Ru for every ν ∈ Cu
ν ) = eRϕ
be a net of states of
µi)i∈I converges to 1 strictly (i.e., in the strict topology of M(C0( G))). Then
2 (eRϕ
µi + R(eRϕ
= ( 1
converges to 1 strictly because R is unital and maps
µi)))i∈I
C0( G) into itself. Applying (c') to the net ( 1
0(G), we
2 (µi + µi ◦ Ru))i∈I
converges to 1 in norm. It follows from Lemma 4.3 that (eRϕ
µi + R(eRϕ
deduce that ( 1
µi)i∈I
µi)))i∈I
converges to 1 in norm, as desired.
(c') =⇒ (c) Notice that R(eRϕ
0(G) such that (eRϕ
(cid:16)eRϕ
2 (µi+µi◦Ru)(cid:17)i∈I
2 (eRϕ
0(G)∗. Let (µi)i∈I
of Ru-invariant states of Cu
The last assertion under the assumption that G is second countable follows from the easy obser-
vation that if U is a unitary representation of G on a Hilbert space H with (a net of) almost-invariant
vectors, then since L2(G) is separable (Proposition 1.19), one can find a sequence of almost-invariant
vectors.
(cid:3)
Definition 4.5. Let (µt)t≥0 be a w∗-continuous convolution semigroup of positive functionals on
0(G). We say that (µt)t≥0 has unbounded generator if its generating functional (Definition 3.5) is
Cu
unbounded.
By Lemma 3.6 and Theorem 3.4, when the convolution semigroup consists of Ru-invariant con-
tractive positive functionals, it has unbounded generator if and only if the associated completely
Dirichlet form is unbounded.
The following is the main result of this subsection.
Theorem 4.6. Let G be a second countable locally compact quantum group. Then G does not have
Property (T) ⇐⇒ there exists a w∗-continuous convolution semigroup of Ru-invariant states of Cu
0(G)
that has unbounded generator (equivalently, whose associated completely Dirichlet form is unbounded).
Proof. ( =⇒ ) Assume that G does not have Property (T). By Theorem 4.4 implication (d') =⇒ (a),
Corollary 2.8 and Theorem A.1, there is a sequence (µn)∞n=1 of Ru-invariant states of Cu
0(G) such
that the sequence (an)∞n=1 := (eR(2,ϕ)
)∞n=1 of selfadjoint contractions in L∞( G) converges to 1 in the
strong operator topology but not in norm. Clearly 1 − an ≥ 0 for all n ∈ N. Let (ζj)∞j=1 be a dense
sequence in L2(G). By passing to a subsequence, one may assume that there exists ε0 > 0 such that
for every n ∈ N,
µn
and
Whenever H is a (generally unbounded) positive selfadjoint operator on a Hilbert space, write
QH for the associated closed densely-defined (non-negative) quadratic form. For n ∈ N, consider
the positive operator ∆n := Pn
k=1 k(1 − ak) ∈ B(L2(G)). Then (∆n)∞n=1, equivalently (Q∆n)∞n=1,
is increasing. Let Q be the closed (non-negative) quadratic form that is the pointwise limit of
(Q∆n)∞n=1 (see [Dav2, Theorem 4.32]). It follows from (4.1) that the dense subset {ζj : j ∈ N} of
L2(G) is contained in the domain of Q. For every k ∈ N, ak being selfadjoint, contractive, and
Markov with respect to ϕ implies that Q1−ak is Dirichlet with respect to ϕ by [GL2, Lemma 5.2].
Hence, for all n ∈ N, Q∆n is Dirichlet with respect to ϕ, and therefore so is Q. More generally the
operators Rµk, k ∈ N, are (KMS-symmetric) completely positive contractions, thus the selfadjoint
k(1 − an)ζjk ≤
1
2n
(∀1≤j≤n)
k1 − ank ≥ ε0.
(4.1)
(4.2)
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
37
operators ak are completely Markov with respect to ϕ (see the terminology introduced in the Ap-
pendix and Theorem A.1), so as above, Q is completely Dirichlet with respect to ϕ. Additionally, for
all n ∈ N, Q∆n is invariant under U (L∞( G)′) for {ak : k ∈ N} ⊆ L∞( G), so Q also has this property.
Finally, it is clear from (4.2) that Q is unbounded. Theorem 3.4 (and Remark 3.3) produce the w∗-
continuous convolution semigroup of Ru-invariant states of Cu
0(G) that is associated with Q, which,
by the remark before the theorem, has unbounded generator.
( ⇐= ) Assume that G has Property (T), and let (µt)t≥0 be a w∗-continuous convolution semig-
0(G). By Theorem 4.4 implication (a) =⇒ (b),
roup of (even not necessarily Ru-invariant) states of Cu
µt −−−→t→0+
Remark 4.7. Comparing the above proof to those of the classical results that we extend, see e.g. [AkW,
Theorem 3], something seems to be 'missing', namely picking an increasing sequence (Kn)∞n=1 of
ǫ in norm; that is, the generator of (µt)t≥0 is bounded (Lemma 3.6).
compact neighbourhoods of the identity with G =S∞n=1 Kn and choosing the sequence of continu-
ous positive-definite functions in a way that makes the series converge uniformly on each of these
compact sets, so as to make the sum ψ continuous. We, however, chose the dense sequence (ζn)∞n=1
arbitrarily. The reason is that ψ is conditionally negative definite, and also measurable as the limit
of a sequence of continuous functions; therefore, it is equal in L∞(G) to a continuous conditionally
negative-definite function. This is because for every t ≥ 0, the measurable function e−tψ is positive
definite by Schönberg's theorem, thus it is equal in L∞(G) to a continuous positive-definite function
(see [Dev] for a nice survey on this).
(cid:3)
4.2. The Haagerup Property. The following definition was introduced in [DFSW].
Definition 4.8. A locally compact quantum group G has the Haagerup Property if it admits a unitary
representation that is mixing and has almost-invariant vectors.
In the following result, which extends [DFSW, Theorem 6.5 (i) ⇐⇒ (iii) ⇐⇒ (iv) and Proposition
6.12] by adding two more equivalent conditions, all approximate identities are understood in the
Banach algebraic sense.
Proposition 4.9. Let G be a locally compact quantum group. Then the following conditions are equi-
valent:
(a) G has the Haagerup Property;
(b) there is a net (µi)i∈I of states of Cu
C0( G);
C0( G);
(b') same as (b) for a net of Ru-invariant states;
(c) there is a net (µi)i∈I of states of Cu
(c') same as (c) for a net of Ru-invariant states.
0(G) such that (cid:0)eRϕ
µi(cid:1)i∈I
0(G) such that (cid:0)eR(2,ϕ)
(cid:1)i∈I
µi
is an approximate identity for
is an approximate identity for
Moreover, if G is second countable, one can replace 'net' by 'sequence' throughout. Similar assertions
hold for operators of the form Lµ, µ ∈ Cu
Proof. The equivalence (a) ⇐⇒ (b) ⇐⇒ (b') is [DFSW, Theorem 6.5 (i) ⇐⇒ (iii) and Proposition
6.12].
0(G)∗.
38
ADAM SKALSKI AND AMI VISELTER
of states of Cu
implication follows.
1 strictly in M(C0( G)) if and only
µi −−→i∈I
ǫ in the w∗-topology. Thus, to prove that
µ ∈ C0( G). Then as in the
µ ∈ D(τ−i/2)
µ ) ∈ C0( G). The desired
∈ C0( G). Then for n ∈ N,
) dt ∈ C0( G), where the integral converges in norm, is
0(G)∗ as in (1.7). If µ is Ru-
µ) and τt(eRϕ
µ) =
for every t ∈ R, we deduce from the ultraweak closedness of τ−i/4 that xn = τ−i/4(eRϕ
µn ),
µn = τi/4(xn) ∈ C0( G). Recall
0(G), eRϕ
By Lemma 2.17, for a net (µi)i∈I
if eR(2,ϕ)
1 strictly in M(C0( G)), if and only if µi −−→i∈I
µi −−→i∈I
(b) ⇐⇒ (c) and (b') ⇐⇒ (c'), we only need to take care of the issue of belonging to C0( G).
0(G)∗ be hermitian and satisfy eRϕ
(b) =⇒ (c) and (b') =⇒ (c'). Let µ ∈ Cu
µ) = R(cid:0)(cid:0)eRϕ
µ(cid:1)∗(cid:1) ∈ C0( G), thus eRϕ
µ(cid:1)∗(cid:1) = R(cid:0)(cid:0)eRϕ
proof of Lemma 2.14 (c), we have τ−i/2(eRϕ
in the C∗-algebraic sense [Kus1, Proposition 1.24]. Hence eR(2,ϕ)
= τ−i/4(eRϕ
0(G)∗ be such that eR(2,ϕ)
(c) =⇒ (b) and (c') =⇒ (b'). Let µ ∈ Cu
the operator xn :=
entire analytic with respect to τ in the C∗-algebraic sense. Define µn ∈ Cu
t and Ru. Since eR(2,ϕ)
invariant, so is µn because of the commutativity of τ u
eRϕ
now in the von Neumann algebraic sense (use Lemma 2.17). Hence eRϕ
that kµnk ≤ kµk for all n ∈ N and µn −−−→n→∞
Let (µi)i∈I
of(cid:0)
µi −−→i∈I
converges to ǫ in the w∗-topology. By the foregoing,(cid:0)eRϕ
= τ−i/4(eRϕ
µ in the w∗-topology, thus kµnk −−−→n→∞ kµk.
be a net of states as in the assumption. For i ∈ I and n ∈ N, write µi,n := (µi)n. Since
µi,n(cid:1)(i,n)∈I×N that also
ǫ in the w∗-topology (see above), there is a subnet (νj)j∈J
is an approximate identity for C0( G).
(cid:3)
The final assertion, related to second countability, is obvious.
√n√πRR e−nt2
µ
τt(eR(2,ϕ)
νj(cid:1)j∈J
µ
µ
µ
µ◦τ u
−t
1
kµi,nk
The next result is a generalisation of [CaS, Proposition 6.5], which in turn is based on [Sau] and
[JoM]. Essentially, for a C∗-algebra A, we replace B(H) and K(H) appearing in those results by
M(A) and A, respectively. We remark that the constructions in the proof are identical to those in
[CaS], that is, they do not depend on A. However, to verify that the operators belong to M(A) or
A we need to present all steps. A strictly continuous semigroup in M(A) is a family (St)t≥0 in M(A)
such that S0 = 1M(A), StSs = St+s for every t, s ≥ 0, and limt→0+ St = 1M(A) strictly.
Proposition 4.10. Let H be a Hilbert space and A be a separable C∗-subalgebra of B(H) acting non-
degenerately on H. Form M(A) in B(H). Let (Ci)i∈I be a family of closed convex subsets of H, and let
(Tn)∞n=1 be a sequence in M(A) with the following properties:
(a) for every n ∈ N, Tn is a selfadjoint contraction;
(b) for every n ∈ N and i ∈ I, Ci is invariant under Tn;
(c) Tn −−−→n→∞
1M(A) strictly.
Then there is a strictly continuous semigroup (St)t≥0 of selfadjoint contractions in M(A) leaving each
of the sets Ci invariant. If Tn ∈ A for every n ∈ N, then we can arrange that St ∈ A for every t > 0.
Proof. We follow closely the proof of [CaS, Proposition 6.5], explaining which changes should be
made and where. Write 1 := 1M(A).
Step 1: we first show that we may assume, without loss of generality, that the operators (Tn)∞n=1
commute.
Indeed, replacing (Tn)∞n=1 by a subsequence if necessary, we pick a dense sequence (al)∞l=1 in A
such that kTna − ak ≤ 2−n kak for all n ∈ N and a ∈ spannT k
j al : 1 ≤ j, l ≤ n − 1, 0 ≤ k ≤ n2o. For
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
39
n ∈ N, let
θn :=
1
n
(T1 + . . . + Tn) ∈ M(A) and ∆n := n1 − (T1 + . . . + Tn) = n(1 − θn) ∈ M(A)+,
and for α > 0, consider the positive contraction Rα,n := (1 + α∆n)−1 ∈ M(A). By [CaS, Lemma
6.4], Ci is invariant under Rα,n for every i ∈ I. Since kθnk ≤ 1, we have
1
nα
Rα,n = ((nα + 1)1 − nαθn)−1 =
1
nα + 1(cid:18)1 −
θn(cid:19)−1
=
nα + 1
nα + 1
(cid:18) nα
nα + 1(cid:19)k
θk
n,
∞Xk=0
where the sum converges in the norm of M(A). For every α > 0, repeating the calculations in [Sau,
pp. 88–89] or [JoM, pp. 43–44] yields that (Rα,na)∞n=1 converges in norm for all a ∈ {al}∞l=1, thus
for all a ∈ A. From selfadjointness of the operators Rα,n, n ∈ N, and strict completeness of M(A),
we deduce that (Rα,n)∞n=1 converges strictly in M(A) to some positive contraction ρα ∈ M(A), which
evidently preserves all sets Ci, i ∈ I. Moreover, by the further arguments in [Sau, p. 89] or [JoM,
p. 44], we have ρα −−−−→α→0+
1 strictly and αρα − βρβ = (α − β)ραρβ for α, β > 0, hence the operators
(ρα)α>0 commute. Therefore, we may replace (Tn)∞n=1 by(cid:0)ρ 1
n(cid:1)∞
n(cid:1)∞
It remains to show that if the operators (Tn)∞n=1 belong to A, so do (ρα)α>0, thus do (cid:0)ρ 1
n=1.
For n, m ∈ N, n < m, put ∆n,m := ∆m − ∆n. For α > 0 and n ∈ N, repeating the above
reasoning with (Tm)∞m=n+1 in lieu of (Tm)∞m=1 implies that the sequence(cid:0)(cid:0)1 + α
nα+1 ∆n,m(cid:1)−1(cid:1)∞
converges strictly to some positive contraction γα,n ∈ M(A). One shows as in [JoM, p. 45] that
nα+1(cid:0)ρα − (1 + α∆n)−1(cid:1)θnγα,n, and as in [JoM, p. 46] uses this to prove
ρα − (1 + α∆n)−1γα,n = nα
that the operator
n=1.
m=n+1
ψα,n := θn(1 + α∆n)−1γα,n +
nα
nα + 1(cid:0)ρα − (1 + α∆n)−1(cid:1)θnγα,n
satisfies kρα − ψα,nk ≤ 2
conclude that ρα ∈ A.
nα. Since (1 + α∆n)−1, ρα, γα,n ∈ M(A) and by assumption θn ∈ A, we
Step 2: we then argue that the assertion holds under the assumption that the operators (Tn)∞n=1
commute.
To this end we follow [JoM, proof of Lemme 2]. Let (al)∞l=1 be a dense sequence in A. Replacing
(Tn)∞n=1 by a subsequence if necessary, we may assume thatP∞n=1 kTnal − alk < ∞ for each l ∈ N.
Given n, m ∈ N, n < m, define θn, ∆n, ∆n,m as above, and given t ≥ 0, consider the positive
contraction St,n := e−t∆n ∈ M(A). These operators preserve the sets Ci, i ∈ I, by [CaS, Lemma 6.4].
From the commutativity assumption we get St,m = St,ne−t∆n,m. For every t ≥ 0, one checks as in
[JoM, p. 40] that (St,na)∞n=1 converges in norm for all a ∈ {al}∞l=1, thus for all a ∈ A. Consequently,
(St,n)∞n=1 converges strictly to a positive contraction St ∈ M(A), which leaves invariant each of the
sets Ci, i ∈ I. As in [JoM, p. 41], (St)t≥0 is a strictly continuous semigroup (equivalently, when
viewing these operators as elements of B(A), (St)t≥0 is a C0-semigroup).
It remains to show that if the operators (Tn)∞n=1 belong to A, so do (St)t>0. The argument
starting in [JoM, middle of p. 41] shows that there exists K > 0 such that for all t > 0 and n ∈ N,
k(1 − θn)Stk ≤ k(1 − θn)St,nk ≤ K√nt
. As a result, St = limn→∞ θnSt, and since θn ∈ A for all n ∈ N
by assumption, we conclude that also St ∈ A.
(cid:3)
40
ADAM SKALSKI AND AMI VISELTER
In the next corollary, which extends [CaS, Proposition 6.6], we use the following notation. For
k ∈ N, write L2(Mk) := L2(Mk, trk), where trk is the canonical (non-normalised) trace on Mk. For
a Hilbert space H and T ∈ B(H), we consider the amplification T (k) := 1 ⊗ T ∈ B(cid:0)L2(Mk) ⊗ H(cid:1).
Corollary 4.11. Let H be a Hilbert space and A be a separable C∗-subalgebra of B(H) acting non-
degenerately on H. View M(A) in B(H). Let (ki)i∈I be a family in N, and for each i ∈ I, let Ci be a
closed convex subset of L2(Mki)⊗ H. Let (Tn)∞n=1 be a sequence in M(A) with the following properties:
(a) for every n ∈ N, Tn is a selfadjoint contraction;
(b) for every n ∈ N and i ∈ I, Ci is invariant under T (ki)
(c) Tn −−−→n→∞
1M(A) strictly.
n
;
Then there is a strictly continuous semigroup (St)t≥0 of selfadjoint contractions in M(A) such that for
each t ≥ 0 and i ∈ I, Ci is invariant under S(ki)
. If Tn ∈ A for every n ∈ N, then we can arrange that
St ∈ A for every t > 0.
Proof. Observe that the constructions in the proof of Proposition 4.10 do not depend on the closed
convex sets in question, and they are 'amplification invariant'. The corollary thus follows.
(cid:3)
t
We are ready to present the main result of this subsection.
Theorem 4.12. Let G be a second countable locally compact quantum group. The following conditions
are equivalent:
(a) G has the Haagerup Property;
(b) there exists a w∗-continuous convolution semigroup (µt)t≥0 of Ru-invariant states of Cu
0(G) such
µt ∈ C0( G) for every t > 0.
that eR(2,ϕ)
µn
invariant states of Cu
µn (cid:1)∞
0(G) such that (cid:0)eR(2,ϕ)
Proof. The implication (b) =⇒ (a) follows from Proposition 4.9 implication (c') =⇒ (a) and Lemma
2.17 implication (a) =⇒ (c').
(a) =⇒ (b) By Proposition 4.9 implication (a) =⇒ (c') there exists a sequence (µn)∞n=1 of Ru-
n=1 is an approximate identity for C0( G). Each of the
operators eR(2,ϕ)
, n ∈ N, is selfadjoint and completely Markov (as defined in the Appendix) by Corol-
lary 2.8 and Theorem A.1. Applying Corollary 4.11 to this sequence with A := C0( G), H := L2(G),
tri⊗ϕ([0, 1]M(2,tri⊗ϕ)) ⊆ L2(Mi) ⊗ L2(G) (cf. Definition A.3),
and for all i ∈ N, ki := i and Ci := i
we obtain a strictly continuous semigroup (St)t≥0 in M(C0( G)) ⊆ L∞( G) ⊆ B(L2(G)) consisting of
selfadjoint (contractive) completely Markov operators on L2(G) with respect to ϕ with St ∈ C0( G)
for t > 0. Since (St)t≥0 is, in particular, a C0-semigroup on L2(G), Theorem 3.4 implies that there is
0(G) such that St = eR(2,ϕ)
a w∗-continuous convolution semigroup (µt)t≥0 of Ru-invariant states of Cu
for every t ≥ 0. This completes the proof.
(2)
(cid:3)
µt
5. EXAMPLES
This section will be devoted to discussing examples of convolution semigroups on a locally com-
pact quantum group G. As mentioned in the introduction, the classical theory (i.e. the case where
G is a classical locally compact group) is as rich as the general theory of Lévy processes, and we
refer to [FOT, Den] or [App, Lia] for further information. Here we focus on describing in detail
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
41
the dual case and then on providing a method of constructing genuinely quantum examples via
cocycle twisting. Finally note that several interesting compact quantum group examples are treated
in [CFK].
5.1. Co-commutative quantum groups. Let Γ be a locally compact group and consider G := Γ.
0(G) = C∗(Γ), the full C∗-algebra of Γ. We will tacitly use the natural identification between
Then Cu
C∗(Γ)∗ and the Fourier–Stieltjes algebra B(Γ) (see [Eym]) given as follows: when viewing B(Γ) as
contained algebraically in Cb(Γ), the pairing of µ ∈ B(Γ) and the image of f ∈ L1(Γ) in C∗(Γ) is
RΓ f (γ)µ(γ) dγ. In other words, letting (cid:0)λu
in M(C∗(Γ)) be the universal representation of Γ,
the identification is such that µ(λu
) = W Γ ∈ M(C0(Γ) ⊗min
C∗(Γ)), as a continuous function from Γ to C∗(Γ) with the strict topology, is the map γ 7→ λu
γ. Thus,
) = µ. The positive elements in C∗(Γ)∗ correspond to positive-
each µ ∈ B(Γ) satisfies (µ ⊗ id)(W∗Γ
definite functions in B(Γ). Since Ru(λu
γ−1 for each γ ∈ Γ, and since every positive-definite
function µ satisfies µ(γ−1) = µ(γ) for each γ ∈ Γ, we conclude that positive Ru-invariant elements
in C∗(Γ)∗ correspond to real-valued positive-definite functions on Γ.
γ) = µ(γ) for every γ ∈ Γ. Then σ(W∗Γ
γ(cid:1)γ∈Γ
γ) = λu
i,j=1 cicjθ(γ−1
A continuous function θ : Γ → R is conditionally negative definite if it satisfies the following
conditions: CND1. θ(e) = 0; CND2. θ(γ−1) = θ(γ) for all γ ∈ Γ; and CND3. for every n ∈ N,
γ1, . . . , γn ∈ Γ and c1, . . . , cn ∈ R with Pn
i γj) ≤ 0. Such
functions admit non-negative values, and Schönberg's theorem asserts that a continuous function
θ : Γ → R satisfying CND1 and CND2 is conditionally negative definite if and only if e−tθ is positive
definite for all t ≥ 0.
i=1 ci = 0, we have Pn
Recall that in the state space of C∗(Γ), that is, in the space of positive-definite functions µ ∈ B(Γ)
with µ(e) = 1, the σ(C∗(Γ)∗, C∗(Γ))-topology coincides with the topology of uniform convergence
on compact sets [Dix, Theorem 13.5.2]. We note further that w∗-continuous convolution semig-
roups (µt)t≥0 of Ru-invariant states of C∗(Γ) are in 1−1 correspondence with conditionally negative-
definite functions θ on Γ. Indeed, if θ is given, then the semigroup(cid:0)e−tθ(cid:1)t≥0 has the desired prop-
erties by Schönberg's theorem. Conversely, given a semigroup (µt)t≥0 as above, for every γ ∈ Γ,
(µt(γ))t≥0 is a continuous positive semigroup in (0, 1], hence of the form(cid:0)e−tθ(γ)(cid:1)t≥0 for some con-
tinuous function θ : Γ → R+, which clearly satisfies CND1 and CND2. By Schönberg's theorem, θ is
conditionally negative definite.
Since G = Γ has trivial scaling group, eR(2)
) for all µ ∈ B(Γ) by Lemma
2.14 (d) and Proposition 2.4 (since G is unimodular we omit the Haar weight from the notation).
As already explained, this means that eR(2)
µ = µ as elements of Cb(Γ) ⊆ B(L2(Γ)) (as multiplication
operators). So for a conditionally negative-definite function θ : Γ → R and the matching convolu-
µt (cid:1)t≥0 in B(L2(Γ)) is (cid:0)Me−tθ(cid:1)t≥0. It
tion semigroup (µt)t≥0 in B(Γ), the associated semigroup (cid:0)eR(2)
is now an exercise to check that the generator of this C0-semigroup is −A, where A is the positive
selfadjoint operator Mθ over L2(G) given by f 7→ θf with maximal domain (compare Lemma 3.7).
As a result, the associated Dirichlet form is Q : L2(G) → [0,∞] given by Qf := RΓ θ(γ)f (γ)2 dγ,
f ∈ L2(G).
5.2. Convolution semigroups via cocycle twistings. Let G be an arbitrary locally compact quantum
group.
µ = eRµ = (µ ⊗ id)(W∗Γ
42
ADAM SKALSKI AND AMI VISELTER
Definition 5.1. A unitary Ω ∈ L∞(G) ⊗ L∞(G) is said to be a 2-cocycle if it satisfies the following
equality:
We say that Ω is trivial if it is of the form (U ⊗ U )∆(U∗) for some unitary U ∈ L∞(G).
(1 ⊗ Ω)(id ⊗ ∆)(Ω) = (Ω ⊗ 1)(∆ ⊗ id)(Ω).
The following result is due to De Commer.
Theorem 5.2 ([DeC1, Theorem 9.1.4]). Let G be a locally compact quantum group and let Ω ∈
L∞(G) ⊗ L∞(G) be a 2-cocycle. Put M := L∞(G) and let for m ∈ M
∆Ω(m) := Ω∆(m)Ω∗.
Then the pair (M, ∆Ω) defines a locally compact quantum group GΩ (so in particular we have L∞(GΩ) =
L∞(G)).
It is easy to see that if Ω is trivial, then G ∼= GΩ.
Now the usefulness of this theorem in our context is in that it allows us to define interesting
convolution semigroups via cocycle twistings. Note that the cocycle twisting construction need not
preserve the Haar weight [FiV], nor does it have to preserve compactness [DeC2]! In particular we
need not have a straightforward equality L2(G) = L2(GΩ) or C0(G) = C0(GΩ).
Proposition 5.3. Consider a w∗-continuous convolution semigroup of states (µt)t≥0 on a locally com-
pact quantum group G. Assume that Ω ∈ L∞(G) is a 2-cocycle and that the associated (via Theorem
3.2 (c)) semigroup on L∞(G), (Tt)t≥0, satisfies the condition
(Tt ⊗ id)(Ω) = Ω.
Then (Tt)t≥0 arises also from a convolution semigroup of states (eµt)t≥0 on GΩ.
Proof. By Theorem 3.2 it suffices to show that the intertwining relation
∆Ω ◦ Tt = (Tt ⊗ id) ◦ ∆Ω
holds for each t ≥ 0. But the multiplicative domain arguments for the unital completely positive
map Tt ⊗ id show that for each m ∈ L∞(G) and t ≥ 0 we have
(Tt ⊗ id)(∆Ω(m)) = (Tt ⊗ id)(Ω∆(m)Ω∗) = (Tt ⊗ id)(Ω)((Tt ⊗ id)(∆(m)))(Tt ⊗ id)(Ω∗)
= Ω∆(Ttm)Ω∗ = ∆Ω(Ttm).
(cid:3)
Note that by the comments above we have no guarantee that this construction preserves KMS-
symmetry; it will however preserve that property if the extra assumptions guarantee that the Haar
weights of GΩ coincide with these of G.
As we want to produce examples related to non-classical quantum groups, we begin from an 'easy'
noncommutative L∞(G). Let then Γ be a locally compact group and consider G := Γ. Here on one
hand we know how all the convolution semigroups of states on G look like – they are associated to
conditionally negative-definite functions on Γ (Subsection 5.1), and on the other hand we have a
natural construction of the cocycle twists, as explained for example in [EnV].
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
43
Proposition 5.4. Suppose that Γ is a locally compact group with a closed abelian subgroup H (so that
via the Takesaki–Tatsuuma theorem from [TaT] we have a natural normal inclusion γ : VN(H) ֒→
VN(Γ) ∼= L∞(Γ), respecting the co-products). Let k : H × H → T be a measurable 2-cocycle on H,
i.e. a measurable T-valued function such that for all r, s, t ∈ H,
k(s, t)k(r, st) = k(r, s)k(rs, t).
Then if we first use the isomorphism VN(H) ∼= L∞( H) and then embed the element k into L∞(Γ) ⊗
L∞(Γ), it becomes a 2-cocycle on Γ, denoted henceforth by Ω. Further if ψ : Γ → C is a conditionally
negative-definite function vanishing on H, then the family (exp(−tψ))t≥0 of elements of B(Γ), viewed
as a convolution semigroup of states on Γ, satisfies the assumptions of Proposition 5.3. In particular
we obtain a new convolution semigroup on ΓΩ.
Proof. The first part is essentially contained in [EnV, Section 6] (the extra assumptions there are
not necessary for us thanks to [DeC1, Theorem 9.1.4]). The construction is as follows: we view k
as an element of VN(H) ⊗ VN(H) and then write Ω = (γ ⊗ γ)(k). It is easy to check (as in [EnV])
that then Ω is a 2-cocycle in the sense of Definition 5.1.
To show the second part it suffices to see that for t ≥ 0 the condition
takes the form
(Tt ⊗ id)(Ω) = Ω
(Rexp(−tψ) ⊗ id) ((γ ⊗ γ)(k)) = (γ ⊗ γ)(k),
(5.1)
so in particular if Rexp(−tψ)γ(VN(H)) = idγ(VN(H)), then the condition above is satisfied. Now the
latter is equivalent to the fact that exp(−tψ(h)) = 1 for all h ∈ H; in other words, to ψH = 0.
Remark 5.5. In fact the proof above shows that sometimes we can hope for the condition (5.1) to
hold even if ψ does not vanish everywhere on H. Suppose for simplicity that H is discrete (so that
H is compact). To understand what (5.1) says in terms of the relationship between k and ψ we
need to see that it holds if and only if the coefficients ch1,h2 and dh1,h2 (h1, h2 ∈ H), given by
(cid:3)
ch1,h2 :=(cid:10)(Rexp(tψ) ⊗ id) ((γ ⊗ γ)(k)) (δe ⊗ δe), δh1 ⊗ δh2(cid:11) ,
dh1,h2 := h(γ ⊗ γ)(k)(δe ⊗ δe), δh1 ⊗ δh2i ,
coincide. We have however
and
ch1,h2 = exp(tψ(h1))dh1,h2
dh1,h2 =Z HZ H
k(h1, h2)h1(h1)h2(h2) dh1 dh2,
where we use the explicit expressions for the isomorphism L∞( H) ∼= VN(H). Thus the sufficient
and necessary condition for (5.1) to hold is that ψ(h) = 0 for all h ∈ H such that
Z HZ H
k(h1, h2)h1(h1)h2(h2) dh1 dh2 6= 0.
Consider now the following example, based on [EnV]; we will again begin from a general setup,
and then pass to a more specific context.
44
ADAM SKALSKI AND AMI VISELTER
Example 5.6. Let H, G be abelian groups, with G acting on H by automorphisms, and let Γ :=
H ⋊ G. Suppose further that k : H × H → T is a measurable 2-cocycle and consider the quantum
group GΩ := bΓΩ constructed out of the pair (Γ, k) as in Proposition 5.4. Note that the quantised
Heisenberg group of [EnV] is a special case of that construction, if we consider Γ := Hn(R) =
Rn+1 ⋊ Rn (n ≥ 2), and k : [Rn+1 × [Rn+1 → T determined by a choice of j, k ∈ {1, . . . , n}, j 6= k,
and a parameter qjk ∈ R, given by the formula
k(u, v, u′, v′) := exp(iqjk uu′(vj v′k − vk v′j))
(u, u′ ∈ R, v, v′ ∈ Rn).
Return to the general case of H ⋊ G. Suppose that ψG : G → C is continuous, conditionally negative
definite. It follows directly from the definitions that ψ : Γ → C defined by
ψ(h, g) := ψG(g)
(h ∈ H, g ∈ G),
is a continuous, conditionally negative-definite function on Γ such that ψH = 0. Thus it satisfies
all the assumptions of Proposition 5.4. If we specify again to the context of Γ = Hn(R) we obtain
a natural construction associating to each Lévy process on Rn a convolution semigroup on the
quantised Heisenberg group of [EnV].
The discussion in the above example can be summarised in the following corollary. The second
statement follows from the remarks above and the fact that the (unitary) antipode does not change
in this quantisation of Hn(R) by [EnV, Theorem 5.2], because Hn(R) is unimodular, so in particular
its modular function is identically 1 on Rn+1.
Corollary 5.7. Let n ∈ N. Let (µt)t≥0 denote the family of distributions of a Lévy process on Rn.
Given a choice of j, k ∈ {1, . . . , n}, j 6= k, and a parameter qjk ∈ R, the construction described in the
above example yields a convolution semigroup of operators on Hq
n(R), the quantised Heisenberg group
of [EnV]. If the distributions µt are symmetric, the convolution semigroup obtained is KMS-symmetric.
Remark that it is proved in [EnV] that L∞(Hq
n(R)) ∼= L∞(R) ⊗ B(L2(Rn)).
APPENDIX A. [GL2, THEOREM 4.7, 'CONVERSELY']
The purpose of this Appendix is to point out a mistake in the 'conversely' statement of [GL2,
Theorem 4.7], and propose two solutions under additional hypotheses, Propositions A.4 and A.9.
Consider a von Neumann algebra A acting (not necessarily standardly) on a Hilbert space H and
an n.s.f. weight ϕ on A, and we denote A := A ⋊σϕ R ⊆ B(L2(R) ⊗ H). When indicating that an
operator is Markov, we suppress the n.s.f. weight in question when there is no confusion.
Theorem A.1 ([GL2, Theorem 4.7, first assertion]). Let T be a KMS-symmetric Markov operator on
A with domain M. Then T is ultraweakly continuous and p-integrable for p ∈ {1, 2}. Moreover,
(eT (1))∗ ⊇ T , and eT (2) is a symmetric Markov operator on L2(A).
The converse statement claimed that, starting with a symmetric Markov operator on L2(A) with
domain span(cid:2)0, h1/2(cid:3)L2(A) = span i(2)([0, 1]M(2)) = i(2)(M(2)), one can produce a corresponding
KMS-symmetric Markov operator on A. The problem with the proof is that M(2) might not be
closed under the absolute value map; this was used in [GL2] to establish that the operator T0 is
bounded.
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
45
A.1. First (better) approach. As in Subsection 1.1, let h be the canonical closed operator affiliated
with A, let θ be the action dual to σϕ, and let τ be the n.s.f. trace on A satisfying d eϕ
dτ = h. For n ∈ N
fixed, consider the von Neumann algebra Mn(A) ∼= Mn ⊗ A acting on the Hilbert space Cn ⊗ H
and the n.s.f. weight trn ⊗ ϕ on Mn(A), where trn is the canonical (non-normalised) trace on Mn.
Below we use tensor products of unbounded operators on Hilbert spaces [KaR, Section 11.2]. Then
σtrn⊗ϕ R, acting on L2(R) ⊗ Cn ⊗ H ∼= Cn ⊗ L2(R) ⊗ H,
σtrn⊗ϕ = idMn ⊗ σϕ, thus An := Mn(A) ⋊
equals Mn ⊗ A, the associated operator hn equals 1Mn ⊗ h, and the associated n.s.f. trace τn on An
satisfying d( ^trn⊗ϕ)
= hn equals trn ⊗ τ as ^trn ⊗ ϕ = trn ⊗ eϕ. Fixing a system (eij)1≤i,j≤n of matrix
units for Mn, we thus have a ∗-algebras isomorphism τn An ∼= Mn ⊗ τ A as follows: an element
a ∈ τn An is associated to the matrix (aij)1≤i,j≤n ∈ Mn⊗ τ A such that (eii⊗ 1)· a· (ejj ⊗ 1) = eij ⊗ aij
for all 1 ≤ i, j ≤ n, and conversely, a matrix (aij)1≤i,j≤n ∈ Mn ⊗ τ A is associated to Pn
i,j=1eij ⊗ aij,
where the upper dot stands for summation in the algebra τn An, namely, followed by taking the
closure.
dτn
Lemma A.2.
(a) For every p ∈ [1,∞), the subspace Lp(Mn(A)) ⊆ τn An is identified with the
(b) For every q ∈ [2,∞) we have N (q,trn⊗ϕ) ∼= Mn ⊗ N (q,ϕ), and for every a = (aij)1≤i,j≤n in this
subspace Mn ⊗ Lp(A) ⊆ Mn ⊗ τ A.
(c) For every p ∈ [1,∞) we have M(p,trn⊗ϕ) ∼= Mn ⊗ M(p,ϕ), and for every a = (aij)1≤i,j≤n in this
(q)
trn⊗ϕ(a) ∈ Lq(Mn(A)) is identified with(cid:0)j
trn⊗ϕ(a) ∈ Lp(Mn(A)) is identified with(cid:0)i
(p)
(q)
ϕ (aij)(cid:1)1≤i,j≤n.
ϕ (aij)(cid:1)1≤i,j≤n.
(p)
set, the element j
set, the element i
Proof. (a) This follows because the action on An ∼= Mn ⊗ A dual to σtrn⊗ϕ = idMn ⊗ σϕ is idMn ⊗ θ.
(b) Let a = (aij)1≤i,j≤n ∈ Mn(A). If a ∈ Mn ⊗ N (q,ϕ), namely h1/qa∗ji ∈ τ A for all 1 ≤ i, j ≤ n,
then
n a∗ = (1Mn ⊗ h1/q) X1≤i,j≤n
(1Mn ⊗ h1/q)(eji ⊗ a∗ij) = X1≤i,j≤n
eji ⊗ a∗ij ⊇ X1≤i,j≤n
eji ⊗ h1/qa∗ij
h1/q
P1≤i,j≤n
(the last equality checks easily). We have eji ⊗ h1/qa∗ij ∈ τn An for all 1 ≤ i, j ≤ n, hence h1/q
τn An, so that a ∈ N (q,trn⊗ϕ) from Remark 1.3. Also, by Remark 1.2, the operator h1/q
=(cid:0)j
equals
n a∗ ∈
trn⊗ϕ(a)∗
eji ⊗ h1/qa∗ij, i.e., it is associated to the matrix(cid:16)h1/qa∗ji(cid:17)1≤i,j≤n
1≤i,j≤n.
Conversely, if a ∈ N (q,trn⊗ϕ), then for fixed 1 ≤ i, j ≤ n, the element ai := (eii ⊗ 1)a also
belongs to N (q,trn⊗ϕ), for the latter is a left ideal in Mn(A). That is, D(h1/q
n a∗i ) is τn-dense. Now
(ejj ⊗ 1)h1/q
n a∗i ⊆ (ejj ⊗ h1/q)a∗i = eji ⊗ h1/qa∗ij,
so these operators also have τn-dense domains. Since τn = trn⊗ τ, standard trace calculations show
that D(h1/qa∗ij) is τ-dense, proving that aij ∈ N (q,ϕ) by Remark 1.3.
n = (ejj ⊗ 1)(1⊗ h1/q) ⊆ ejj ⊗ h1/q, thus (ejj ⊗ 1)h1/q
ϕ (aij)(cid:1)∗
n a∗ = j
(q)
(q)
(c) This follows readily from (b) and Proposition 1.4 (a).
(cid:3)
For simplicity, we keep writing M(2), i(2) for M(2,ϕ), i
(2)
ϕ , etc. Notice that the identification
of the linear spaces L2(Mn(A)) and Mn ⊗ L2(A) is actually an isomorphism of Hilbert spaces:
L2(Mn(A)) ∼= L2(Mn, trn) ⊗ L2(A).
46
ADAM SKALSKI AND AMI VISELTER
Definition A.3. An everywhere-defined operator T on A is called n-Markov if its amplification
idMn ⊗ T is Markov on Mn(A), namely T is n-positive and n-contractive. A (not necessarily every-
where defined) operator S on L2(A) will be called n-Markov with respect to ϕ if 1Mn ⊗ S (algebraic
tensor product) is Markov on L2(Mn(A)) ∼= L2(Mn, trn) ⊗ L2(A) with respect to trn ⊗ ϕ, namely
i(2)(M(2)) ⊆ D(S) and i
trn⊗ϕ([0, 1]M(2,trn⊗ϕ)) is preserved by 1Mn ⊗ S.
(2)
Clearly, an n-Markov operator is Markov.
Proposition A.4. Let S be a symmetric 2-Markov operator on L2(A) with domain i(2)(M(2)). Then
there is a unique everywhere-defined, normal, KMS-symmetric Markov operator T on A that satisfies
eT (2) ⊇ S.
Proof. As in the beginning of the proof of [GL2, Theorem 4.7, 'conversely'], since S is symmetric and
Markov and i(2) is injective, there is a linear map T0 : M(2) → M(2) given by i(2)(T0a) = Si(2)(a),
a ∈ M(2), which is KMS-symmetric and preserves the set [0, 1]M(2) (in particular, it is positivity
preserving). Analogously, using now the full strength of 2-Markovianity of S, the map 1M2 ⊗ S on
L2(M2, tr2) ⊗ i(2)(M(2)) induces the linear map idM2 ⊗ T0 on M2 ⊗ M(2) ∼= M(2,tr2⊗ϕ). This map
preserves the set [0, 1]M2⊗M(2), hence it is positivity preserving. To fill the gap in [GL2] we need to
show that T0 is bounded.
that converges to 1 in the strong
(eλ))λ∈I is bounded. As in the proof of Proposition 1.10,
λ) ∈
, and T0(eλa)∗ = T0(a∗eλ) because T0 is positivity, thus adjoint,
0 0 ) ∈ M(2,tr2⊗ϕ).
T0(a∗eλ) T0(a∗a)(cid:17) ≥ 0. By [Lan, Lemma 5.2 (ii)], for every
λ)η, η(cid:11) hT0(a∗a)ζ, ζi ≤ (kakkζkkηk)2.
Let a ∈ M(2). By Lemma 1.1 there is a net (eλ)λ∈I
operator topology and such that (σ− i
i(2)(eλa) −−→λ∈I
[0, 1]M(2), T0(a∗a) ∈h0,kak2 1iM(2)
Since idM2 ⊗ T0 is positivity preserving,(cid:16) T0(e2
ζ, η ∈ H,
i(2)(a) weakly. For every λ ∈ I, since T0 preserves [0, 1]M(2), we have T0(e2
0 0 ) ∈ M(2,tr2⊗ϕ), thus 0 ≤ (cid:16) e2
hT0(eλa)ζ, ηi2 ≤(cid:10)T0(e2
Consequently, (T0(eλa))λ∈I
converges ultraweakly to some c ∈ A with kck ≤ kak. For every b ∈ M, by Proposition 1.4 (b),
is bounded by kak, so by passing to a subnet, we may assume that it
preserving. In addition, ( eλ a
in [0, 1]M∞
λ eλa
a∗eλ a∗a(cid:17) = ( eλ a
0 0 )∗ ( eλ a
λ) T0(eλa)
4
tr(cid:16)i(2)(T0(eλa)) · i(2)(b)(cid:17) = tr(cid:16)T0(eλa) · i(1)(b)(cid:17) −−→λ∈I
tr(cid:16)c · i(1)(b)(cid:17) .
tr(cid:16)i(2)(a) · S(i(2)(b))(cid:17)
Since S is symmetric and adjoint preserving, the expression on the left-hand side is also equal to
tr(cid:16)S(i(2)(eλa)) · i(2)(b)(cid:17) = tr(cid:16)i(2)(eλa) · S(i(2)(b))(cid:17) −−→λ∈I
= tr(cid:16)S(i(2)(a)) · i(2)(b)(cid:17) = tr(cid:16)i(2)(T0a) · i(2)(b)(cid:17) = tr(cid:16)T0a · i(1)(b)(cid:17)
(see (1.2)). In conclusion, c = T0a, and we deduce that kT0ak ≤ kak, as desired.
Definition A.5. Recall that a Markov semigroup on A is a semigroup (Tt)t≥0 of everywhere-defined,
normal Markov operators (i.e., positive contractions) on A such that R+ ∋ t 7→ Tta is ultraweakly
continuous for all a ∈ M; and a Markov semigroup on L2(A) is a C0-semigroup of Markov operators
on L2(A) [GL2, p. 62] (see also Definition 1.5). For n ∈ N, we define n-Markov semigroups by
replacing 'Markov' by 'n-Markov'.
(cid:3)
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
47
Using Theorem A.1 and Proposition A.4 we obtain a corrected replacement of [GL2, Theorem
4.9].
Corollary A.6. If (Tt)t≥0 is a KMS-symmetric Markov semigroup on A, then(cid:0)eT (2)
t (cid:1)t≥0 is a symmetric
Markov semigroup on L2(A). Conversely, if (St)t≥0 is a symmetric 2-Markov semigroup on L2(A),
then there exists a KMS-symmetric Markov semigroup (Tt)t≥0 on A such that St = eT (2)
for all t ≥ 0
(in particular, (St)t≥0 consists of contractions).
For a quadratic form Q on a Hilbert space H and n ∈ N, define a quadratic form Q(n) on
t
L2(Mn, trn) ⊗ H by
Q(n)(ζ) :=
nXi,j=1
Q(ζij)
for (ζij)1≤i,j≤n = ζ ∈ L2(Mn, trn) ⊗ H.
Definition A.7. Let n ∈ N. A quadratic form Q on L2(A) is called n-Dirichlet with respect to ϕ if
Q(n) is Dirichlet on L2(Mn(A)) ∼= L2(Mn, trn) ⊗ L2(A) with respect to trn ⊗ ϕ.
Let (St)t≥0 be a C0-semigroup of selfadjoint contractions on a Hilbert space H and Q be the asso-
ciated (closed densely-defined) quadratic form. Let n ∈ N. Then the C0-semigroup (1Mn ⊗ St)t≥0
of selfadjoint contractions on L2(Mn, trn) ⊗ H has Q(n) as its associated quadratic form. When
H = L2(A), [GL2, Theorem 5.7] implies that Q is n-Dirichlet if and only if St is n-Markov for all
t ≥ 0.
We use the terminology completely Markov, resp. completely Dirichlet, to mean n-Markov, resp. n-
Dirichlet, for every n ∈ N. Finally, taking into account Corollary A.6 and the last paragraph, we
obtain the following.
Corollary A.8. There are 1 − 1 correspondences between the following classes:
• KMS-symmetric completely Markov semigroups (Tt)t≥0 on A;
• symmetric completely Markov semigroups (St)t≥0 on L2(A);
• completely Dirichlet forms Q on L2(A).
t
for all t ≥ 0, and (St)t≥0 is the C0-semigroup (of selfadjoint contractions)
They are given by: St = eT (2)
on L2(A) associated to Q.
A.2. Second approach. As an alternative to Proposition A.4 we give the next result. However, it
is less practical for our purposes because the new additional assumption cannot be expressed in
terms of quadratic forms using [GL2, Theorem 5.3] since convex sets like i(2)(M) and i(2)([0, 1]M)
are usually not closed in L2(A) by Proposition 1.10.
Proposition A.9. Let S be a symmetric Markov operator on L2(A) with domain i(2)(M(2)). Assume
that i(2)(M) is invariant under S. Then there is a unique everywhere-defined, normal, KMS-symmetric
Markov operator T on A that satisfies eT (2) ⊇ S.
Proof. Recall that since S is symmetric and Markov, there is a linear map T0 : M(2) → M(2) given
by i(2)(T0a) = S(cid:0)i(2)(a)(cid:1), a ∈ M(2), which is KMS-symmetric and preserves the set [0, 1]M(2).
By the additional assumption, M is invariant under T0. Fix a ∈ A+. Let (aλ)λ∈I
that converges ultraweakly to a. Since [0, 1]M(2) is invariant under T0, (T0aλ)λ∈I
[0,kak 1]M
be a net in
is
48
ADAM SKALSKI AND AMI VISELTER
a net in [0,kak 1]M(2), so one can assume, by passing to a subnet if necessary, that it converges
ultraweakly to some T a ∈ A+ of norm at most kak. For every b ∈ M, the KMS-symmetry of T0,
Proposition 1.4 (b), and the fact that T0(b) ∈ M imply that
tr(cid:16)i(1)(b) · T a(cid:17) = lim
tr(cid:16)i(1)(b) · T0(aλ)(cid:17) = lim
tr(cid:16)i(1)(T0(b)) · aλ(cid:17) = tr(cid:16)i(1)(T0(b)) · a(cid:17) .
tr(cid:16)T0(b) · i(1)(aλ)(cid:17)
λ∈I
λ∈I
= lim
λ∈I
(A.1)
By density of i(1)(M) in A∗ (Proposition 1.4 (c)), T a does not depend on the choice of (aλ)λ∈I
.
The function T : A+ → A+ is evidently additive and positively homogeneous, so it extends to a
well-defined positivity preserving linear map T : A → A, obviously satisfying kTk ≤ 4.
Clearly T0M ⊆ T , but in fact T0 ⊆ T .
converges ultraweakly to a such that (cid:0)i(2)(a′λ)(cid:1)λ∈I
then by (A.1), Proposition 1.4 (b), and symmetry of S (see (1.2)), for every b ∈ M,
tr(cid:16)i(1)(b) · T a(cid:17) = tr(cid:16)i(1)(T0(b)) · a(cid:17) = lim
is a net in M that
converges weakly to i(2)(a) (Proposition 1.10),
tr(cid:16)i(1)(T0(b)) · a′λ(cid:17) = lim
tr(cid:16)i(2)(T0(b)) · i(2)(a′λ)(cid:17)
λ∈I
Indeed, if a ∈ M(2)
+ and (a′λ)λ∈I
λ∈I
= tr(cid:16)i(2)(T0(b)) · i(2)(a)(cid:17) = tr(cid:16)i(2)(b) · i(2)(T0(a))(cid:17) = tr(cid:16)i(1)(b) · T0(a)(cid:17) .
This proves that T0 is bounded, and the proof is complete.
(cid:3)
REFERENCES
[AkW] C. A. Akemann and M. E. Walter, Unbounded negative definite functions, Canad. J. Math. 33
(1981), no. 4, 862–871.
[App] D. Applebaum, Lévy processes and stochastic calculus, second ed., Cambridge Studies in
Advanced Mathematics, vol. 116, Cambridge University Press, Cambridge, 2009.
[BeT] E. Bédos and L. Tuset, Amenability and co-amenability for locally compact quantum groups,
Internat. J. Math. 14 (2003), no. 8, 865–884.
[BHV] B. Bekka, P. de la Harpe, and A. Valette, Kazhdan's property (T), New Mathematical Mono-
graphs, vol. 11, Cambridge University Press, Cambridge, 2008.
[BLSC] D. Bakry, M. Ledoux, and L. Saloff-Coste, Markov semigroups at Saint-Flour, Probability at
Saint-Flour, Springer, Heidelberg, 2012.
[BrK] M. Brannan and D. Kerr, Quantum groups, property (T ), and weak mixing, Comm. Math.
Phys., to appear, doi: 10.1007/s00220-017-3037-0.
[CaS] M. Caspers and A. Skalski, The Haagerup approximation property for von Neumann algebras
via quantum Markov semigroups and Dirichlet forms, Comm. Math. Phys. 336 (2015), no. 3,
1637–1664.
[CFK] F. Cipriani, U. Franz, and A. Kula, Symmetries of Lévy processes on compact quantum groups,
their Markov semigroups and potential theory, J. Funct. Anal. 266 (2014), no. 5, 2789–2844.
[Cip1] F. Cipriani, Dirichlet forms and Markovian semigroups on standard forms of von Neumann
algebras, J. Funct. Anal. 147 (1997), no. 2, 259–300.
[Cip2]
, Dirichlet forms on noncommutative spaces, Quantum potential theory, Lecture Notes
in Math., vol. 1954, Springer, Berlin, 2008, pp. 161–276.
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
49
[CiS] F. Cipriani and J.-L. Sauvageot, Derivations as square roots of Dirichlet forms, J. Funct. Anal.
201 (2003), no. 1, 78–120.
[Cra1] J. Crann, Amenability and covariant injectivity of locally compact quantum groups II, Canad.
J. Math. 69 (2017), no. 5, 1064–1086.
[Cra2]
, Inner amenability and approximation properties of locally compact quantum groups,
preprint, arXiv:1709.01770, 2017.
[Dav1] E. B. Davies, Quantum theory of open systems, Academic Press [Harcourt Brace Jovanovich,
Publishers], London-New York, 1976.
[Dav2]
, One-parameter semigroups, London Mathematical Society Monographs, vol. 15,
Academic Press, Inc., London-New York, 1980.
[Daw1] M. Daws, Multipliers, self-induced and dual Banach algebras, Dissertationes Math. (Roz-
prawy Mat.) 470 (2010), 62 pp.
[Daw2]
, Multipliers of locally compact quantum groups via Hilbert C∗-modules, J. Lond.
Math. Soc. (2) 84 (2011), no. 2, 385–407.
[Daw3]
, Completely positive multipliers of quantum groups, Internat. J. Math. 23 (2012),
no. 12, 1250132, 23 pp.
[DeC1] K.
coactions
thesis,
algebraic
for
KU Leuven,
and
2009,
locally
Available
com-
at
Commer,
De
quantum groups,
Galois
pact
http://homepages.vub.ac.be/~kdecomme/PhD.html.
Ph.D.
[DeC2]
, On cocycle twisting of compact quantum groups, J. Funct. Anal. 258 (2010), no. 10,
3362–3375.
[Den] J. Deny, Méthodes hilbertiennes en théorie du potentiel, Potential Theory (C.I.M.E., I Ciclo,
Stresa, 1969), Edizioni Cremonese, Rome, 1970, pp. 121–201.
[Dev] A. Devinatz, On measurable positive definite operator functions, J. Lond. Math. Soc. 35
(1960), 417–424.
[DFSW] M. Daws, P. Fima, A. Skalski, and S. White, The Haagerup property for locally compact
quantum groups, J. Reine Angew. Math. 711 (2016), 189–229.
[Dix] J. Dixmier, C∗-algebras, North-Holland Mathematical Library, vol. 15, North-Holland Pub-
lishing Co., Amsterdam-New York-Oxford, 1977.
[DSV] M. Daws, A. Skalski, and A. Viselter, Around Property (T) for quantum groups, Comm. Math.
Phys. 353 (2017), no. 1, 69–118.
[DuS] N. Dunford and J. T. Schwartz, Linear operators. Part II. Spectral theory, Wiley Classics
Library, John Wiley & Sons Inc., New York, 1988.
[EnN] K.-J. Engel and R. Nagel, One-parameter semigroups for linear evolution equations, Graduate
Texts in Mathematics, vol. 194, Springer-Verlag, New York, 2000, With contributions by S.
Brendle, M. Campiti, T. Hahn, G. Metafune, G. Nickel, D. Pallara, C. Perazzoli, A. Rhandi,
S. Romanelli and R. Schnaubelt.
[EnV] M. Enock and L. Vaınerman, Deformation of a Kac algebra by an abelian subgroup, Comm.
Math. Phys. 178 (1996), no. 3, 571–596.
[Eym] P. Eymard, L'algèbre de Fourier d'un groupe localement compact, Bull. Soc. Math. France 92
(1964), 181–236.
50
ADAM SKALSKI AND AMI VISELTER
[Fim] P. Fima, Kazhdan's property T for discrete quantum groups, Internat. J. Math. 21 (2010),
no. 1, 47–65.
[FiV] P. Fima and L. Vainerman, Twisting and Rieffel's deformation of locally compact quantum
groups: deformation of the Haar measure, Comm. Math. Phys. 286 (2009), no. 3, 1011–
1050.
[FOT] M. Fukushima, Y. Oshima, and M. Takeda, Dirichlet forms and symmetric Markov processes,
extended ed., De Gruyter Studies in Mathematics, vol. 19, Walter de Gruyter & Co., Berlin,
2011.
[Fra] U. Franz, Lévy processes on quantum groups and dual groups, Quantum independent incre-
ment processes. II, Lecture Notes in Math., vol. 1866, Springer, Berlin, 2006, pp. 161–257.
[FST] U. Franz, A. Skalski, and R. Tomatsu, Idempotent states on compact quantum groups and
their classification on Uq(2), SUq(2), and SOq(3), J. Noncommut. Geom. 7 (2013), no. 1,
221–254.
[GL1] S. Goldstein and J. M. Lindsay, KMS-symmetric Markov semigroups, Math. Z. 219 (1995),
no. 4, 591–608.
[GL2]
, Markov semigroups KMS-symmetric for a weight, Math. Ann. 313 (1999), no. 1,
39–67.
[Haa1] U. Haagerup, On the dual weights for crossed products of von Neumann algebras. I. Removing
separability conditions, Math. Scand. 43 (1978/79), no. 1, 99–118.
[Haa2]
[Haa3]
[Haa4]
, On the dual weights for crossed products of von Neumann algebras. II. Application of
operator-valued weights, Math. Scand. 43 (1978/79), no. 1, 119–140.
, Operator-valued weights in von Neumann algebras. I, J. Funct. Anal. 32 (1979),
no. 2, 175–206.
, Lp-spaces associated with an arbitrary von Neumann algebra, Algèbres d'opérateurs
et leurs applications en physique mathématique (Proc. Colloq., Marseille, 1977), Colloq.
Internat. CNRS, vol. 274, CNRS, Paris, 1979, pp. 175–184.
[HJX] U. Haagerup, M. Junge, and Q. Xu, A reduction method for noncommutative Lp-spaces and
applications, Trans. Amer. Math. Soc. 362 (2010), no. 4, 2125–2165.
[HNR] Z. Hu, M. Neufang, and Z.-J. Ruan, Completely bounded multipliers over locally compact
quantum groups, Proc. Lond. Math. Soc. (3) 103 (2011), no. 1, 1–39.
[JNR] M. Junge, M. Neufang, and Z.-J. Ruan, A representation theorem for locally compact quantum
groups, Internat. J. Math. 20 (2009), no. 3, 377–400.
[Jol] P. Jolissaint, On property (T) for pairs of topological groups, Enseign. Math. (2) 51 (2005),
no. 1-2, 31–45.
[JoM] P. Jolissaint and F. Martin, Algèbres de von Neumann finies ayant la propriété de Haagerup et
semi-groupes L2-compacts, Bull. Belg. Math. Soc. Simon Stevin 11 (2004), no. 1, 35–48.
[Kan] S. Kantorovitz, Topics in operator semigroups, Progress in Mathematics, vol. 281, Birkhäuser
Boston, Inc., Boston, MA, 2010.
[KaR] R. V. Kadison and J. R. Ringrose, Fundamentals of the theory of operator algebras. Volume II:
Advanced theory, Graduate Studies in Mathematics, vol. 16, Amer. Math. Soc., Providence,
RI, 1997.
CONVOLUTION SEMIGROUPS ON QUANTUM GROUPS AND NONCOMMUTATIVE DIRICHLET FORMS
51
[Kas] P. Kasprzak, Rieffel deformation via crossed products, J. Funct. Anal. 257 (2009), no. 5,
1288–1332.
[Kat] T. Kato, Perturbation theory for linear operators, Classics in Mathematics, Springer-Verlag,
Berlin, 1995, Reprint of the 1980 edition.
[KNR] M. Kalantar, M. Neufang, and Z.-J. Ruan, Poisson boundaries over locally compact quantum
groups, Internat. J. Math. 24 (2013), no. 3, 1350023, 21 pp.
[Kus1] J. Kustermans, One-parameter representations on C∗-algebras, preprint, arXiv:funct-
an/9707009, 1997.
[Kus2]
, Locally compact quantum groups in the universal setting, Internat. J. Math. 12
(2001), no. 3, 289–338.
[KV1] J. Kustermans and S. Vaes, Locally compact quantum groups, Ann. Sci. École Norm. Sup. (4)
33 (2000), no. 6, 837–934.
[KV2]
, Locally compact quantum groups in the von Neumann algebraic setting, Math. Scand.
92 (2003), no. 1, 68–92.
[Kye] D. Kyed, A cohomological description of property (T) for quantum groups, J. Funct. Anal. 261
(2011), no. 6, 1469–1493.
[Lan] E. C. Lance, Hilbert C∗-modules. A toolkit for operator algebraists, London Mathematical
Society Lecture Note Series, vol. 210, Cambridge University Press, 1995.
[Lia] M. Liao, Lévy processes in Lie groups, Cambridge Tracts in Mathematics, vol. 162, Cambridge
University Press, Cambridge, 2004.
[LS1] J. M. Lindsay and A. G. Skalski, Convolution semigroups of states, Math. Z. 267 (2011),
no. 1-2, 325–339.
[LS2]
804.
, Quantum stochastic convolution cocycles III, Math. Ann. 352 (2012), no. 4, 779–
[Mey] P.-A. Meyer, Quantum probability for probabilists, Lecture Notes in Mathematics, vol. 1538,
Springer-Verlag, Berlin, 1993.
[MVD] A. Maes and A. Van Daele, Notes on compact quantum groups, Nieuw Arch. Wisk. (4) 16
(1998), no. 1-2, 73–112.
[Nel] E. Nelson, Notes on non-commutative integration, J. Funct. Anal. 15 (1974), 103–116.
[NgV] C.-K. Ng and A. Viselter, Amenability of locally compact quantum groups and their unitary
co-representations, Bull. Lond. Math. Soc. 49 (2017), no. 3, 491–498.
[PiX] G. Pisier and Q. Xu, Non-commutative Lp-spaces, Handbook of the geometry of Banach
spaces, Vol. 2, North-Holland, Amsterdam, 2003, pp. 1459–1517.
[RuV] V. Runde and A. Viselter, On positive definiteness over locally compact quantum groups,
Canad. J. Math. 68 (2016), no. 5, 1067–1095.
[SaS] P. Salmi and A. Skalski, Idempotent states on locally compact quantum groups II, Q. J. Math.
68 (2017), no. 2, 421–431.
[Sau] J.-L. Sauvageot, Strong Feller semigroups on C∗-algebras, J. Operator Theory 42 (1999),
no. 1, 83–102.
[Sch] M. Schürmann, White noise on bialgebras, Lecture Notes in Mathematics, vol. 1544,
Springer-Verlag, Berlin, 1993.
52
ADAM SKALSKI AND AMI VISELTER
[Str] ¸S. Stratila, Modular theory in operator algebras, Abacus Press, Tunbridge Wells, England,
1981.
[StZs] ¸S. Stratila and L. Zsidó, Lectures on von Neumann algebras, Abacus Press, Tunbridge Wells,
England, 1979.
[Tak1] M. Takesaki, Tomita's theory of modular Hilbert algebras and its applications, Lecture Notes
in Mathematics, Vol. 128, Springer-Verlag, Berlin, 1970.
[Tak2]
, Theory of operator algebras. II, Encyclopaedia of Mathematical Sciences, vol. 125,
Springer-Verlag, Berlin, 2003.
[TaT] M. Takesaki and N. Tatsuuma, Duality and subgroups, Ann. of Math. (2) 93 (1971), 344–
364.
[Ter1] M. Terp, Lp spaces associated with von Neumann algebras, Notes, Report No. 3a + 3b, Køben-
havns Universitets Matematiske Institut, June 1981.
[Ter2]
, Interpolation spaces between a von Neumann algebra and its predual, J. Operator
Theory 8 (1982), no. 2, 327–360.
[Vae] S. Vaes, A Radon-Nikodym theorem for von Neumann algebras, J. Operator Theory 46
(2001), no. 3, suppl., 477–489.
[VD] A. Van Daele, Locally compact quantum groups. A von Neumann algebra approach, SIGMA
Symmetry Integrability Geom. Methods Appl. 10 (2014), Paper 082, 41 pp.
[Wor1] S. L. Woronowicz, Compact matrix pseudogroups, Comm. Math. Phys. 111 (1987), no. 4,
613–665.
[Wor2]
, Compact quantum groups, Symétries quantiques (Les Houches, 1995), North-
Holland, Amsterdam, 1998, pp. 845–884.
INSTITUTE OF MATHEMATICS OF THE POLISH ACADEMY OF SCIENCES, UL. ´SNIADECKICH 8, 00-656 WARSZAWA, POLAND
E-mail address: [email protected]
DEPARTMENT OF MATHEMATICS, UNIVERSITY OF HAIFA, 31905 HAIFA, ISRAEL
E-mail address: [email protected]
|
1801.06974 | 3 | 1801 | 2018-08-30T07:05:17 | C*-superrigidity of 2-step nilpotent groups | [
"math.OA",
"math.GR"
] | We show that torsion-free finitely generated nilpotent groups are characterised by their group C*-algebras and we additionally recover their nilpotency class as well as the subquotients of the upper central series. We then use a C*-bundle decomposition and apply K-theoretic methods based on noncommutative tori to prove that every torsion-free finitely generated 2-step nilpotent group can be recovered from its group C*-algebra. | math.OA | math | C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
Abstract. We show that torsion-free nitely generated nilpotent groups are charac-
terisedbytheirgroupC*-algebrasandweadditionallyrecovertheirnilpotencyclassas
wellasthesubquotientsoftheuppercentralseries.WethenuseaC*-bundledecom-
positionandapplyK-theoreticmethodsbasedonnoncommutativetoritoprovethat
everytorsion-freenitelygenerated2-stepnilpotentgroupcanberecoveredfromits
groupC*-algebra.
∗-algebra
1
Introduction
It is a classical problemto recover a discrete group G from its various group algebras such as the integral
groupring ZGandthecomplexgroupring CG.NotablytheKadison-Kaplanskyunitconjecturepredictsthat
foratorsion-freegroup Geveryunitof CGisoftheform cug,for c∈ C× and g∈ G,makingitpossibleto
recover Gcanonicallyfrom CG.Intheoperatoralgebraicsetting,studyingcompletionsofthe*-algebra CG
indi erenttopologiesoftheboundedoperatorsB((cid:96)2G),suchasthereducedgroup C∗-algebra C∗
red(G)
orthegroupvonNeumannalgebra L(G),similarquestionshaveattractedstronginterestinthepastyears.
In the 80's, Connes conjectured in [10, page 45] that in analogy with Mostow and Margulis super-
rigidity every discrete group with innite conjugacy classes and Kazhdan's property (T) can be recov-
ered from its group von Neumann algebra; that is if G is such a group and H is any other group with
L(G) ≅ L(H) then G ≅ H . Derived from an old name for von Neumann algebras, this phenomenon
is termed W∗-superrigidity. In contrast to Connes' early vision of how the subject might develop, only in
[19]therstbreakthroughon W∗-superrigiditywasachieved, showingthatcertainiteratedwreathprod-
ucts are W∗-superrigid. In the introduction to their article, Ioana, Popa and Vaes draw attention to the
even more mysterious situation in C∗-superrigidity and in analogy with Kadison-Kaplansky's unit conjec-
red(G) always remembers a torsion-free group
ture point out that: "It seems not even known whether C∗
red(H) for some other group H, does it
G" -- in other words, if G is torsion-free group and C∗
follow that G≅ H? At the time of writing their article, only torsion-free abelian groups G were known
tobe C∗-superrigidbyaclassicalresultofScheinberg[33,Theorem1]. Moreprecisely, Gisisomorphicto
red(G)modulotheconnectedcomponentoftheidentity[17,Theorem8.58]. Note
theunitarygroupof C∗
that the restriction to torsion-free groups in Ioana-Popa-Vaes' remark is necessary and natural, since for
red(Z4) and the group C∗-algebra of every innite abelian torsion group is
example C∗
isomorphictothecontinuousfunctionsonthestandardCantorset.
Motivated by C∗-superrigidity of torsion-free abelian groups, certain torsion-free virtually abelian
groups were shown to be C∗-superrigid in [23], providing the rst examples of non-abelian torsion-free
C∗-superrigidgroups. Further C∗-superrigidityresultswerededucedin[8]bycombiningtheauthors're-
sulton W∗-superrigidityofamalgamatedfreeproductswiththeuniquetraceproperty[4]oftheirreduced
group C∗-algebra. In the present article we prove the following result, providing the rst natural class of
C∗-superrigidnon-abeliantorsion-freegroups.
lastmodiedonAugust31,2018
MSCclassication: 46L35;46L05,20C07,20F18
Keywords: C
∗-superrigidity,nilpotentgroup,noncommutativetorus,twistedgroup C
red(Z2× Z2) ≅ C∗
red(G)≅ C∗
8
1
0
2
g
u
A
0
3
]
.
A
O
h
t
a
m
[
3
v
4
7
9
6
0
.
1
0
8
1
:
v
i
X
r
a
byCalebEckhardtandSvenRaum
C*-superrigidityof2-stepnilpotentgroups
TheoremA. Everytorsion-freenitelygenerated2-stepnilpotentgroupis C∗-superrigid.
Wementionthatinanalogywithabeliangroups,atorsion-freenon-abeliannilpotentgroup Gisasfarfrom
being W∗-superrigidaspossible.Indeed,itfollowsfromConnes's[9]andthetheoryofdirectintegralsthat
L(G)≅ L∞([0, 1])⊗ R,where Rdenotesthehypernite II1 factor.
Ourglobalstrategytoprove C∗-superrigidityfollowsthesamelinesas[23]:werstcharacterisetorsion-
free nitely generated nilpotent groups in terms of their group C∗-algebras and recover their nilpotency
class. Thisisthesecondmainresultofthisarticle.
If
TheoremB. Let G be a torsion-free nitely generated nilpotent group and H some discrete group.
red(G), then H is a torsion-free nitely generated nilpotent group, too. Further the class
red(H) ≅ C∗
C∗
of H equalstheclassof Gandthesubquotientsoftheuppercentralseriesofbothgroupsagree.
Animportantingredientintheproofofthistheoremdescribesthecentreofthegroup C∗-algebraofanar-
bitrarytorsion-freevirtuallynilpotentgroup,whichisofindependentinterest.Givenatorsion-freevirtually
nilpotentgroup Gwithsomemaximalnormalniteindexnilpotentsubgroup N Gwewrite F = G~N
forthenitequotientgroup.Thentheouteractionof F on N inducesanactionof F onthecentreZ(N),
asisexplainedinSection3.
TheoremC. If G is a torsion-free virtually nilpotent group and N G is a maximal normal nite index
nilpotentsubgroupwithquotient G~N= F,thenZ(C∗G)= C∗(Z(N))F.
Letusnowexplainthestrategytorecoveratorsion-freenitelygenerated2-stepnilpotentgroup Gfrom
red(G)≅
itsgroup C∗-algebra.ArstapplicationofTheoremBmakesitclearthatanygroup H satisfying C∗
red(H) must be torsion-free nitely generated and nilpotent of class 2 and satisfyZ(H)≅Z(G). Ap-
C∗
red(G) as a C∗-bundle over their centre, we are able to recover
plying Theorem C to write C∗
an isomorphism of torsion-free abelian groups H~Z(H) ≅ G~Z(G). Further, analysing how K-theory
variesoverthedi erentbresofthis C∗-bundle,werecoverthe2-cocycledescribingthecentralextension
1→Z(G)→ G→ G~Z(G)→ 1. Infact,thebresof C∗
red(H)aresocallednoncommutative
tori, whose K-theory has been intensively studied by Elliott [14] and Rie el [30] and we make use of the
formerresults,summarisedinSection2.10.
Itisnaturaltoaskforpossibleextensionsofthepresentresult,eitherbeyondthescopeofnitelygenerated
groupsorbeyondtheclassof2-stepnilpotentgroups.Asfortherstextension,wearefacingtheproblem
tondanappropriatereplacementforTheorem2.3,whichplaysanimportantroleinSection4. Notethat
boththeoremsinquestionmakecrucialuseofniterankoftheabeliangroupsinvolved.Asforthesecond
extension, while the general strategy to prove C*-superrigidity still applies to nitely generated, torsion-
freenilpotentgroups,importantchallengeshavetobeovercome.UnderstandingofK-theoryandthetrace
pairing of twisted group C∗-algebras of nilpotent groups as well as the replacement for cup products as
usedintheproofofTheorem5.1arethetwomainsuchchallenges. Nevertheless,afterthisworkhasrst
appeared in form of a preprint some examples of C∗-superrigid nilpotent groups of arbitrary class have
beenprovided[28].
This article has 5 sections. After the introduction, thorough preliminaries provide all necessary infor-
mation to make this article readable by non-experts in operator algebras, group theory or cohomology
theory. InSection3wecalculatethecentreofthegroup C∗-algebraofavirtuallynilpotentgroup,proving
TheoremC.InSection4weprovidethecharacterisationoftorsion-freenitelygeneratednilpotentgroups
announcedinTheoremB.WethenproceedtoproveourmainTheoremAinSection5.
red(H)≅ C∗
red(G)≅ C∗
2
byCalebEckhardtandSvenRaum
and Zi+1={g∈ G ∀h∈ G∶[g, h]∈ Zi}
C*-superrigidityof2-stepnilpotentgroups
Acknowledgements
TheauthorswouldliketothankElizabethGillaspyforinformingusabouteachothersinterestinthepresent
article'ssubject. Further, wethankRémiBoutonnetforusefulremarksonapreviousversionofthiswork
andwethankAlainValetteforprovidinguswiththeideaofanelementaryprooffortorsion-freenilpotent
groupsofTheorem2.18aswellaspointingoutthereferences[31],[21]and[32].WethankYuheiSuzukifor
pointing out an error in [3, Proposition 3.3], which we used in our work, and providing us with a shorter
proofofTheorem2.3. Finally,wewouldliketothanktherefereeforhiscarefulreadingofthemanuscript
andpointingoutseveralinaccuracies.C.E.waspartiallysupportedbyagrantfromtheSimonsFoundation.
2 Preliminariesandvariousresults
All groups considered in this article are discrete unless explicitly stated di erently. A group G is called
nilpotentifitsuppercentralseriesdenedby
Z0={e}, Z1=Z(G)
terminatesinsome Zn= G. Theminimal nsuchthat Zn+1= Giscalledthenilpotencyclassof G,andwe
saythat Gis n-stepnilpotent.
Wereferthereadertotheexcellentexposition[6]ondiscretegroupsandtheiroperatoralgebras.Since
nilpotent groups are amenable, their maximal C∗-algebra C∗(G) and their reduced C∗-algebra C∗
red(G)
agree.Thelatterisdenedastheclosureofthecomplexgroupring CGundertheimageoftheleft-regular
representation λ ∶ G → U((cid:96)2G) ∶ λgδh = δgh. For amenable groups, we write C∗(G) for this group
C∗-algebra.
If B ⊂ A is a unital C∗-
Let us recall the notion of conditional expectations between C∗-algebras.
subalgebra,thenaconditionalexpectationfrom Aonto B isacontractiveprojection E∶ A→ B ofnorm
one.ByTomiyama'stheorem[6,Theorem1.5.10],everyconditionalexpectation E∶ A→ Bis B-bi-modular,
that is E(bab′)= bE(a)b′ for all b, b′ ∈ B and all a∈ A. If H ≤ G is an inclusion of groups, then there
is a natural faithful conditional expectation E ∶ C∗
red(H) uniquely dened by the property
E(ug)= 1H(g)ug.
2.1 Actionsontorsion-freeabeliangroups
InthissectionweproveTheorem2.3,whichmightbeknowntoexperts.Inordertoremainconsistentwith
thenotationoftheremainingarticle,wewriteabeliangroupsmultiplicatively.
Recall that an action of a discrete group G on a topological space X is called topologically free, if for
every g∈ Gࢨ{e}thexedpointset{x∈ X gx= x}ismeagerin X. Thenextlemmaisfolkloreanda
proofcanbefoundforexamplein[13,Lemma2.1].
Lemma2.1. Let H be a connected topological group and let G be a group acting faithfully by continuous
automorphismson G. Then G H istopologicallyfree.
Thenexttheorem'sproofmakesuseofthenotionofrank.If Aisanabeliangroupconsideredasa Z-module,
therankof Aisdenedas rank(A)= dimQ A⊗Z Q. Further,if Aisanabeliangroup,thenthePontryagin
dual A= Hom(A, S1)equippedwiththetopologyofpointwiseconvergenceisacompactabeliangroup.
Everyautomorphism αof Adenesacontinuousautomorphismof Abyprecomposition χ χ○α.Finally,
ifagroup Gactsona C∗-algebra Aby*-automorphisms,wedenoteby AG={x∈ A gx= xforall g∈ G}
red(G) → C∗
3
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
thexedpointalgebra.WethankYuheiSuzukiforindicatingashorterproofforthenexttheoremthanthe
oneoriginallypresented. Westartbyashortlemma.
Lemma2.2. Let F Tnbeanitegroupactingbycontinuousautomorphismssuchthatthequotientspace
satises Tn~F≅ Tn. Then F actstrivially.
Proof. The quotient map q ∶ Tn → Tn~F by a nite group has the path lifting property according to [7,
11.1.4]. So the induced map on fundamental groups π1(q) is a surjection, hence an isomorphism because
Zn≅ π1(Tn) is Hopan. Identifying Zn≅ π1(Tn) by the map that sends a∈ Zn to the image of the line
segment[0, a]⊂ Rn,weobservethat π1(q)factorsthroughthegroup Zn~a− f a a∈ Zn, f∈ F. Since
everyproperquotientof Znhasrankstrictlylessthan nand π1(q)isanisomorphism,itfollowsthat a= f a
forall a∈ Zn and f∈ F. Thisnishestheproofofthetheorem.
Theorem2.3. Let A be a torsion-free abelian group and G a group acting on A by automorphisms with
niteorbits. If C( A)G≅ C(Tn)forsome n∈ N,then A≅ Zn and Gactstriviallyon A.
Proof. Since G acts with nite orbits on A, we can write A=i∈I Ai as a directed union of subgroups,
each generated by a nite G-invariant set. Since every automorphism of Ai, which is trivial on its nite
generating set, is trivial on Ai, it follows that G Ai factors through a nite group, say G Fi. Note
that since A is torsion free, Ai is connected for each i, so that Lemma 2.1 says that Fi acts topologically
freelyon Ai. So Ai~Fi containsanopenanddensesubsetthatislocallyEuclideanofdimension rank Ai.
Inparticularthecoveringdimensionof Ai~Fi equalstherankof Ai. (Seee.g. [27,p.305]forthenotionof
coveringdimension).
Weput X= spec C∗(A)Gandobservethattheinclusions C∗(A)G⊃ C∗(Ai)Fi and C∗(A)⊃ C∗(A)G
provideuswithquotientmaps X Ai~Fi and A X. Thefollowingcommutativediagramshowsthat
theformerisanopenmap.
/ Ai~Fi
open
4
X
cont O
A
open /
/ Ai
rank(Ai)= dim Ai~Fi≤ dim X= n.
Indeed,bothmapsmarkedasopeninthediagramariseasquotientmapsfromgroupactions.Byhypothesis,
wehave X≅ Tn. Itfollowsthat
We showed that the rank of(Ai)i is uniformly bounded by n, so that rank A≤ n follows. So there is a
nite G-invariantsubset S⊂ Asuchthatwriting B=S,thedivisionclosuresatises A={g∈ A ∃k∈
N≥1∶ gk∈ B}.Theactionof Gon Bfactorsthroughanitegroup F,andsodoestheactionon A:indeed,
if a∈ A and k∈ N≥1 such that ak ∈ B, then(ga)k = gak = ak implies that((ga)⋅ a−1)k = e. Since A is
torsion-free,itfollowsthat ga= a.
Tonishtheproofnotethat F AistopologicallyfreebyLemma2.1,since Aisconnected. Further,
C( A~F)≅ C∗(A)G≅ C(Tn)showingthat Aislocallyeuclideanofdimension n,whichinturnimpliesthat
A≅ Tn bythestructuretheoremforabeliangroups[11,Theorem4.2.4]. So A≅ Zn andbyLemma2.2,we
ndthat F actstriviallyon A. Soalso Gactstriviallyon A.
/
/
/
O
O
O
/
/
O
O
O
O
byCalebEckhardtandSvenRaum
FC-groupsandthecentreofgroupC*-algebras
C*-superrigidityof2-stepnilpotentgroups
2.2
Inthissection,wegiveashortaccountofProposition2.5,providingadescriptionofthecentreofareduced
group C∗-algebra by means of certain abelian subgroups. To this end we will analyse its FC-centre. An
FC-group is a group G whose conjugacy classes are nite. More generally, the FC-centre of a group G is
{g∈ G g hasaniteconjugacyclass}. ThefollowingtheoremdescribesnitelygeneratedFC-groups.
Theorem2.4(SpecialcaseofTheorem2of[12]). AnitelygeneratedFC-groupisvirtuallyabelian.
Thefollowingpropositionisclassical. Weprovideashortprooffortheconvenienceofthereader.
Proposition2.5. Atorsion-freeFC-groupisabelian.
Proof. Itsucestoprovethateverynitelygeneratedtorsion-freeFC-group Gisabelian. ByTheorem2.4
we know that G is virtually abelian. So [2, Theorem 2] implies thatZ(G)≤ G has nite index. It follows
thatthecommutatorsubgroupof Gisnite,andhencetrivialsince Gistorsion-free.
The next proposition describes the centre of the group C∗-algebra of a torsion-free group in terms of an
abeliangroup.
Proposition2.6. Let Gbeatorsion-freegroup.ThentheFC-centreof Gisanormalabeliansubgroup A G
satisfyingZ(C∗
red(G))⊂ C∗(A). Inparticular,Z(C∗
Proof. The FC-centre A G is a torsion-free FC-group and hence abelian by Proposition 2.5. We have to
redG)⊂ C∗(A). If x∈Z(C∗
showthatZ(C∗
redG)wewrite x= xδe=∑g∈G xgδg∈ (cid:96)2(G). Since xiscentral,
wehave x= uhxu∗
forall h∈ Gandhence xg= xhgh−1 forall h∈ G.Since(xg)g is (cid:96)2-summable,itfollows
that xg= 0,unless g hasaniteconjugacyclassin G,meaning g∈ A. So x∈ C∗(A).
2.3 Centralextensionsandcocycles
Inthissectionweclarifytheexactrelationbetween n-stepnilpotentgroupsandextensiondatafor n− 1-
stepnilpotentgroups.Thebook[5]providesareferenceforthecohomologyofdiscretegroups.Recallthat
givenagroup H andanabeliangroup A,thereisaone-to-onecorrespondencebetweencentralextensions
andelementsof H2(H, A),whichsendsanextensionafterchoiceofasection s∶ H→ Gtothe2-cocycle
σ(h1, h2) = s(h1h2)−1s(h1)s(h2), which we call an extension cocycle [5, Theorem 3.12]. Two central
extensions giving rise to G1, G2 are called equivalent, if there is an isomorphism G1 → G2 making the
followingdiagramcommute
red(G))= C∗(A)G.
/ H
/ 1
1
/ A
/ G
h
1
G1
H
1
1
Notethattheautomorphismsof Aand H arexedtobetheidentity. Aimingataclassicationofcentral
extensions without this restriction, we introduce the action of Aut(H)× Aut(A) on H2(H, A) that is
inducedby ρ ψA○ ρ○(ψH× ψH)with ψH∈ Aut(H), ψA∈ Aut(A), ρ∈ Z2(H, A).
Before we proceed to spell out the relation between n-step nilpotent groups and extension data for
n− 1-stepnilpotentgroups,weneedtointroducethefollowingnotation.
/ A
id
/ A
id
/ H
/ 1.
/ G2
5
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
1
/ A
/ G
/ H
/ 1
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
Notation2.7. Let H beagroupand Aanabeliangroup. A2-cocycle σ∈ Z2(H, A)iscalledcentrallynon-
degenerateifforallnon-trivialelements g∈Z(H)thereis h∈ H suchthat σ(g, h)≠ σ(h, g).
Thefollowingtheoremmustbewell-known. Notbeingabletondareferenceweindicateaproofforthe
reader'sconvenience.
Theorem2.8. Let H beagroupand Aanabeliangroup. Associatingtoacentralextensionof Aby H the
extension2-cocyclein H2(H, A),weobtainabijectionbetween
• isomorphismclassesofgroups GsuchthatZ(G)≅ Aand G~Z(G)≅ H,and
• Aut(H)× Aut(A)orbitsofcentrallynon-degeneratecocyclesin H2(H, A).
Proof. Weprovidemapswhichareinversetoeachother. Given GsuchthatZ(G)≅ Aand G~Z(G)≅ H,
wechooseisomorphisms ϕA∶Z(G)→ Aand ϕH∶ G~Z(G)→ H. Thisprovidesuswithanextension
and after choice of a section s∶ H → G with a cocycle σ ∈ Z2(H, A), whose cohomology class is inde-
pendentofthechoiceofthesection. Notethat σ iscentrallynon-degenerate,since s(H)∩Z(G)={e}.
Further,theorbitof[σ]∈ H2(H, A)forthe Aut(H)× Aut(A)actionisindependentofthechoiceof ϕA
and ϕH. So we obtain a well-dened map assigning to the group G an Aut(H)× Aut(A) orbit of cen-
trallynon-degeneratecocyclesin H2(H, A). Sinceeveryisomorphism G≅ Grestrictstoanisomorphism
ofZ(G)≅Z( G) and induces an isomorphism G~Z(G)≅ G~Z( G), this assignment descends to a map
fromisomorphismclassesofgroups GsatisfyingthatZ(G)≅ Aand G~Z(G)≅ H to Aut(H)× Aut(A)
orbitsofcentrallynon-degeneratecocyclesin H2(H, A).
Let now σ ∈ Z2(H, A) and dene G = Aσ H, which as a set is the Cartesian product A× H and
becomesagroupwiththemultiplication
Theisomorphismclassoftheextension
doesonlydependontheclassof σin H2(H, A).Notethatif σiscentrallynon-degenerate,thenZ(G)= A.
Further,if ψA∈ Aut(A), ψH∈ Aut(H)andwedene σ= ψA○ σ○(ψ−1
H)aswellas G= Aσ H,
then
G∋(a, h)(ψA(a), ψH(h))∈ G
isagroupisomorphism. Soweobtainawell-denedmapfrom Aut(H)× Aut(A)orbitsin H2(H, A)to
isomorphismclassesofgroups GsuchthatZ(G)≅ Aand G~Z(G)≅ H.
WespecialisethestatementofTheorem2.8tothecaseofnilpotentgroups. ForlateruseinSection5,the
nextcorollarywillbereformulatedinSection2.4.
Corollary2.9. Associating to a nilpotent group G the triple(Z(G), G~Z(G), σG), where we denote by
σG∈ H2(G~Z(G),Z(G))theextensioncocycle,weobtainabijectionbetween
• isomorphismclassesof n-stepnilpotentgroups,and
• equivalence classes of triples(A, H, σ) with A abelian, H an n− 1-step nilpotent group and σ ∈
H2(H, A)acentrallynon-degeneratecocycle. Twotriples(A, H, σ)and( A, H, σ)areequivalentif
thereareisomorphisms ϕA∶ A→ Aand ϕH∶ H→ H suchthat σ○(ϕH× ϕH)= ϕA○ σ.
(a1, h1)(a2, h2)=(a1a2σ(h1, h2), h1h2).
H × ψ−1
/ G
/(a, e)
/ A
a
/ H
/ 1
1
6
/
/
/
/
/
/
/
/
/
byCalebEckhardtandSvenRaum
C*-superrigidityof2-stepnilpotentgroups
2.4 Cocyclesandskew-symmetricforms
In this section we recall the relationship between 2-cocycles on torsion-free abelian groups and skew-
symmetricforms.ItprovidesthereformulationofCorollary2.9asitwillbeusedinSection5.Thereferences
[22,Theorem7.1],[5,ChapterV.6Exercise5]or[18,ProofofTheorem2]showthisresultwithoutthestate-
mentof Aut(B)× Aut(A)-equivariance,whichcanberightawaycheckedbytheformulaprovidedbelow.
Wedenoteby B∧Bthewedgeproduct,whicharisesasthequotientgroup B⊗ZB~b⊗b′+b′⊗b b, b′∈ B.
Theorem2.10. Let Bbeafreeabeliangroupand Aanyabeliangroup.Thenthereisan Aut(B)×Aut(A)-
equivariantbijectionbetween H2(B, A)and A-valuedskew-symmetricformson B,givenby
H2(B, A)→ Hom(B∧ B, A)∶ σ(b1∧ b2 σ(b1, b2)− σ(b2, b1)).
Corollary2.11. Associatingtoa2-stepnilpotentgroup Gthetriple(Z(G), G~Z(G), ωG),wherewedenote
by ωG∈ Hom(G~Z(G)∧ G~Z(G),Z(G))theskew-symmetrisationoftheextensioncocycle,weobtaina
bijectionbetween
• isomorphismclassesoftorsion-freenitelygenerated2-stepnilpotentgroups,and
• equivalence classes of triples(A, B, ω) with A, B torsion-free nitely generated abelian and ω ∈
Hom(B∧B, A)a A-valuedskew-symmetricform.Twotriples(A, B, ω)and( A, B, ω)areequivalent
ifthereareisomorphisms ϕA∶ A→ Aand ϕB∶ B→ B suchthat ω○(ϕB∧ ϕB)= ϕA○ ω.
2.5 Homotopyclassesofevaluationmaps
ThefollowingstatementiseasilycheckedongeneratorsandwillbeusedinSection5.
Proposition2.12. Let h∶ S1→ Tn bealoop,let g∈ Zn≅ Tn andlet evg∶ Tn→ S1 betheevaluationmap.
The natural identication π1(Tn)≅ Zn induces a bilinear form , such that π1(evg)([h])=g,[h]
holdsunderthenaturalidentication π1(S1)≅ Z.
Proof. Notethatbothsidesoftheequation π1(evg)([h])=g,[h]onlydependonthehomotopyclassof
h. Theidentication π1(Tn)≅ Zn providesuswiththestandardloop h(e2πit)=(e2πih1t, . . . , e2πihnt),for
(h1, . . . , hn)∈ Zn.Further g=(g1, . . . , gn)∈ Zndenesthecharacter evg(e2πit1, . . . , e2πitn)= e2πi∑j gj tj.
Itfollowsthat(evg○h)(e2πit)= e2πi∑j hj gj andhence π1(evg)([h])=[evg○h]=∑j hjgj=g,[h]under
thenaturalidentication π1(S1)≅ Z.
2.6 C(X)-algebras
Inthissectionwerecallthenotionof C(X)-algebras,whichisforexampleexplainedin[35,AppendixC]in
themoregeneralcontextof C0(X)-algebras.
Denition2.13. Let Abeaunital C∗-algebraand X acompactspace.
• Aiscalleda C(X)-algebraifthereisaunitalembedding C(X)Z(A).
• If Aisa C(X)-algebraand x∈ X,thenthebreof Aat xisdenedasthequotientof Abytheideal
generatedby{f∈ C(X) f(x)= 0}. Itisdenoted Ax. Theimageof a∈ Ain Ax isdenoteby ax.
7
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
• Aiscalledacontinuous C(X)-algebra,ifitisa C(X)-algebraandthemap xaxiscontinuous
forall a∈ A.
If A is a continuous C(X)-algebra we will also say that A is a C∗-bundle over X. This terminology is
justiedby[35,TheoremC.26]inconnectionwith[35,PropositionC.10(b)],whichexplainstherelationof
C(X)-algebraswithFell's C∗-bundles[15].
In[3,Proposition3.3],itwasclaimedthataunital C(X)-algebra Awithfaithfulconditionalexpectation
E∶ A→ C(X)isacontinuous C(X)-algebra. Thisisnottrueasstated,butneedstheadditionalassump-
tionthateverystate Ex∶ Ax→ Cdenedby Ex(a)= E(a)(x)isfaithful. Adoptingthisextraassumption,
theprooffrom[3]applies.
Proposition2.14(See[3,Proposition3.3]). Let Abeaunital C(X)-algebra,withaconditionalexpectation
E∶ A→ C(X)suchthateach Ex∶ Ax→ C, x∈ X isfaithful. Then Aisacontinuous C(X)-algebra.
Notation2.15. If Aisaunital C∗-algebra,wedenotethespectrumofZ(A)by Glimm(A)andcallitthe
Glimmspaceof A. Notethat Aisa C(Glimm(A))-algebraanditisa C∗-bundleover Glimm(A)ifthere
isaconditionalexpectationfrom Aontoitscentre.
2.7
In this section we recall the notion of twisted group C∗-algebras, which is essential in the study of group
C∗-algebrasassociatedwithnilpotentgroups.Thefollowingtheoremisaspecialcaseof[29,Theorem1.2].
However,sinceitsprooffordiscretegroupsbecomeslesstechnical,weprovideitherefortheconvenience
ofthereader.
Theorem2.16(Specialcaseof[29,Theorem1.2]). Let Gbeanamenablediscretegroup.Write Z=Z(G),
H= G~Z anddenoteby σ∈ Z2(H, Z)anextensioncocycle. For χ∈ Z write σχ= χ○ σ∈ Z2(H, S1). Then
C∗(G)isacontinuous C( Z)-algebra,whosebreat χ∈ Z isisomorphicwith C∗(H, σχ). Moreprecisely,
for χ∈ Z wehaveacommutativediagram
TwistedgroupC*-algebras
C∗(G)
C∗(H, σχ)
C∗(G)χ.
Proof. Since C∗(Z)⊂Z(C∗(G))and C∗(Z)≅ C( Z)viatheFouriertransform,the C∗-algebra C∗(G)is
a C( Z)-algebra.
Recallthat σ∈ Z2(H, Z)denotesanextensioncocycle. Thismeansthat G≅ Zσ H istheuniversal
groupcontainingacentralcopyof Zandelements hfor h∈ Hsubjecttotherelation h1h2= σ(h1, h2)h1h2
forall h1, h2∈ H. Itfollowsthat C∗(G)istheuniversal C∗-algebrageneratedbyacentralcopyof C∗(Z)
and elements uh, h∈ H such that uh1uh2 = uσ(h1,h2)uh1h2
for all h1, h2∈ H. The bre of C∗(G) at the
element χ∈ Z isthequotientof C∗(G)bytherelations uz= χ(z)for z∈ Z. Itishenceisomorphicwith
theuniversal C∗-algebrageneratedbyelements vh for h∈ H thatsatisfy vh1vh2= σχ(h1, h2)vh1h2
forall
h1, h2∈ H. Thisispreciselythetwistedgroup C∗-algebra C∗(H, σχ).
Since Gisamenable,itsmaximalandreducedgroup C∗-algebrasagreesothatthenaturalconditional
expectationofreducedgroup C∗-algebrasprovidesafaithfulconditionalexpectation E∶ C∗(G)→ C∗(Z).
Since Eχidentieswiththenaturaltrace C∗(H, σχ)→ C,itisfaithful,sothatProposition2.14impliesthat
C∗(G)isacontinuous C( Z)-algebra.
8
/
/
&
&
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
The next proposition is possibly known to experts and it was already used in [23].
It provides a simple
algebraicmeantodetectwhichtwistedgroup C∗-algebrasareuntwisted.
Proposition2.17. Let Gbeadiscretegroupand σ∈ Z2(G, S1). Ifthereisaone-dimensional σ-projective
representationof G,then σ isinner. Sothefollowingstatementsareequivalent.
• σ isinner.
• C∗(G, σ)admitsacharacter,thatisamultiplicativepositivefunctional.
• C∗(G, σ)≅ C∗(G).
Proof. Let π∶ G→U(1)= S1 beaone-dimensional σ-projectiverepresentationof G. Forall g, h∈ Gwe
have π(g)π(h)= σ(g, h)π(gh) and hence σ(g, h)= π(g)π(h)π(gh). This proves that σ is inner. Since
charactersof C∗(G, σ)arein1-1correspondencewithone-dimensional σ-projectiverepresentationsof G,
we infer that C∗(G, σ) has a character if and only only if σ is inner. The latter implies that C∗(G, σ) ≅
C∗(G),whichinturnprovidesacharacterof C∗(G, σ).
2.8 GroupC*-algebrasoftorsion-freegroups
InthissectionwerecallacharacterisationofgroupC*-algebrasoftorsion-freenilpotentgroups.Infact,the
followingtheoremholdsforamuchlargerclassofgroupsbyHigsonandKasparov's[16](see[34,Chapter
6.3]fordetailedexplanations). Alreadyin[32, Proposition2.5]and[21, Corollary, p. 96]resultsappeared
thatapplytoshowthatgroup C∗-algebrasoftorsion-freenilpotentgroupsadmitnonon-trivialprojections.
Anelementaryproof,whoseideawascommunicatedtousbyAlainValette,wasprovidedin[31,Theorem
8]. Fortheconvenienceofthereaderweprovidetheshortargumenthere.
Theorem2.18. Let Gbeadiscretegroupsuchthat C∗(G)hasnonon-trivialprojections.Then Gistorsion-
free. Conversely,if Gistorsion-freeandnilpotentthen C∗(G)hasnonon-trivialprojections.
Proof. The rst statement is clear, since for any group G with a non-trivial torsion element g ∈ G, the
ord(g)∑ord(g)
formula
n=1 un
Let now G be a torsion-free nilpotent group of nilpotency class c. If c = 0, then G is trivial and the
conclusion is clear. Assume that c ≥ 1 and the statement is proven for c− 1. Write Z = Z(G) and let
σ∈ Z2(G~Z, Z)beanextensioncocyclefor Z G G~Z.Theorem2.16saysthat C∗(G)isa C∗-bundle
over Z withbre C∗(G~Z, χ○ σ)at χ∈ Z. Let p∈ C∗(G)beaprojection. Denoteby pχ theimageof pin
thebreat χ.Denotingby εthetrivialcharacteron Z,wendthat pε∈{0, 1}bytheinductionhypothesis.
Replacing pby 1− pifnecessary,wemayassumethat pε= 0.Since Z istorsion-free, Z isconnectedsothat
χpχisa{0, 1}-valuedfunctiononaconnectedspace,whichtakesthevalue 0. Henceitisconstantly
0anditfollowsthat p= 0. Thisnishestheproofofthetheorem.
2.9 Theprimitiveidealspaceofnilpotentgroups
InthissectionwerecallresultsofMoore-Rosenberg[26]ontheprimitiveidealspaceofnilpotentgroups.
Denition2.19. Let Aabea C∗-algebra. Thesetofprimitiveidealsof Ais
g∈ CG⊂ C∗(G)denesanon-trivialprojection.
1
Prim(A)={ker π π non-zeroirreducible*-representationof A}.
9
byCalebEckhardtandSvenRaum
X={I∈ Prim(A) I⊃
J∈X
J}
C*-superrigidityof2-stepnilpotentgroups
Thissetisendowedwiththeuniquetopologywhoseclosureoperationsatises
forall X⊂ Prim(A).
Itfollowsfromthedenitionthataprimitiveideal I Aisaclosedpointin Prim(A)ifandonlyifitisa
maximalideal.Recallinthiscontextthatatopologicalspaceiscalled T1ifallitspointsareclosed.Motivated
byanalogieswithLiegroups,thefollowingtheoremisstatedbyMoore-Rosenbergonlyfornitelygenerated
solvable groups G. Its proof however, is valid for arbitrary amenable groups as the authors mention [26,
p.220].
Theorem2.20(Moore-Rosenberg[26,Theorem5]). Let Gbeanitelygeneratedgroup.
• If Gisvirtuallynilpotent,then Prim(G)isa T1-space.
• If Gisamenableand Prim(G)isa T1-space,then Gisvirtuallynilpotent.
2.10 Elliott'sworkonnoncommutativetori
Inthearticle[14], ElliottundertookadetailedstudyoftheK-theoryofnoncommutativetori,inparticular
calculatingthetraceparingwiththe K0-group. InviewofTheorem2.16, theseresultswillbeessentialto
our investigations of group C∗-algebras of torsion-free nitely generated 2-step nilpotent groups. In this
sectionwerecalltheresultsofElliottthatwillbeusedandwexournotation.
Proposition2.21(See[14,Lemma2.3]). Let A be a torsion-free abelian group and ω∈ Hom(A∧ A, S1).
Thereisauniquemap K0(τ)∶ K0(C∗(A, ω))→ Rinducedbytracialstateson C∗(A, ω).
Theorem2.22(See[14,Theorem2.2anditsproof]). Let Abeatorsion-freeabeliangroupand γ∶[0, 1]→
Hom(A∧A, S1)apathwithcontractibleimagefrom ω0to ω1.Consider γasa C([0, 1])-valued2-cocycleon
Aanddenoteby C([0, 1])γ Athetwistedcrossedproduct. Thentheevaluationmaps ev0∶ C([0, 1])γ
A → C∗(A, ω0) and ev1 ∶ C([0, 1])γ A → C∗(A, ω1) induce isomorphisms in K-theory. Further, the
composition
doesnotdependonthechoiceof γ.
Notation2.23. If Aisatorsion-freeabeliangroupand ω∈ Hom(A∧ A, S1)isconnectedtothetrivialform,
wedenoteby
theuniquemaparisingformTheorem2.22.
BeforestatingthenexttheoremofElliott,recallthatK-theoryofanabelian C∗-algebraadmitsacup-product
(seee.g. [1]).
Theorem2.24(See[14,Theorem3.1]). Let Abeatorsion-freeabeliangroupand ω∈ Hom(A∧ A, S1)be
connected to the trivial form. Consider A as a subgroup of K1(C∗A) via the embedding AU(C∗A).
Then
πR~Z○ K0(τ)○ K0(ev1)○ K0(ev0)−1∶ K0(C∗(A, ω0)) -- → R~Z
τω∶ K0(C∗(A)) -- → R~Z
τω([a1]K1∪[a2]K1)= ω(a1∧ a2)
forall a1, a2∈ A.
10
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
3
ThecentreofgroupC*-algebrasofvirtuallynilpotentgroups
Inthissectionwedeterminethecentreofthegroup C∗-algebraofatorsion-freevirtuallynilpotentgroup.
It turns out to be immediately related to the centre of the group itself. Let G be a torsion-free virtually
nilpotentgroup,denoteby N Gsomemaximalniteindexnilpotentnormalsubgroupandlet F= G~N
be the quotient group. Choosing a section s∶ F → G for the quotient map, we dene the conjugation
action G∋ g s(f)gs(f)−1, which is up to inner automorphisms independent of the choice of s. This
provides a group homomorphism F → Out(N). SinceZ(N)≤ N is a characteristic subgroup, we can
compose the latter homomorphism with the restriction map Out(N)→ Out(Z(N))= Aut(Z(N)) to
obtain an action of F onZ(N). Having xed this notation we restate Theorem C from the introduction,
whichisthemaintheoremofthissection.
TheoremC. If G is a torsion-free virtually nilpotent group and N G is a maximal nite index nilpotent
normalsubgroupwithquotient G~N= F,thenZ(C∗G)= C∗(Z(N))F.
Thenextlemmawasalreadystatedandprovenin[24,Corollary2]fortorsion-freenitelygeneratednilpo-
tent groups. We need the more general version dropping the assumption of nite generation and hence
provideadi erentreference.
Lemma3.1([20,Corollary1.3]). Let Gbeatorsion-freenilpotentgroup. Then G~Z(G)istorsion-free.
Thenextlemmadescribesthesetofniteconjugacyclassesofatorsion-freevirtuallynilpotentgroup. Its
conclusionresemblesknownresultsfortorsion-freenilpotentgroups[25].
Lemma3.2. Let G be a torsion-free virtually nilpotent group with maximal normal nite index nilpotent
subgroup N G. Theneveryniteconjugacyclassof GliesinZ(N). SotheFC-centreof GequalsZ(N).
Proof. Let us start by showing that the FC-centre of G equalsZG(N). If g∈ZG(N), then g has a nite
conjugacyclassin G, since N ≤ Ghasniteindex. Viceversa, let g∈ Ghaveaniteconjugacyclassand
wewillshowthat g∈ZG(N). Then C={ghg−1h−1 h∈ N}isniteandwerstshowthat C={e},or
equivalently g∈ZG(N). Let{e}= Z0≤Z(N)= Z1≤ Z2≤≤ Zn= N betheuppercentralseriesof
N and assume that C ⊂ Zk+1 holds for some k≥ 0. Since Zk+1~Zk ≤ N~Zk is central, we obtain for all
h1, h2∈ N that
[gh1g−1h−1
showing that the image of C in N~Zk is a nite subgroup. Since N is torsion-free nilpotent, Lemma 3.1
saysthatalso N~Zk istorsion-free. Thisshowsthattheimageof C in N~Zk istrivialandhence C⊂ Zk.
Because k≥ 0wasarbitrary,weconcludethat C={e}orequivalently g∈ZG(N).
WeshowedthatZG(N)istheFC-centreof G. Sinceeverytorsion-freeFC-groupisabelianbyPropo-
sition 2.5, it follows that ZG(N) is abelian. We denote H = ZG(N), N ≤ G and observe that
Z(H)=ZG(N), sinceZG(N) is abelian. Because of the isomorphism H~Z(H)≅ N~Z(N) we con-
clude that H is nilpotent. Further, H is normal in G. Since N is a maximal normal nite index nilpotent
subgroup of G, it follows that H = N or equivalentlyZG(N)=Z(N), showing that the FC-centre of G
equalsZ(N).
Wearenowreadytoprovethemaintheoremofthissection.
ProofofTheoremC By Lemma 3.2 the FC-centre of G equalsZ(N). So Proposition 2.6 applies to show
thatZ(C∗(G)) = C∗(Z(N))G. Since the conjugation action of G onZ(N) factors through the nite
group F,weobtaintheidentication C∗(Z(N))G= C∗(Z(N))F.Thisnishestheproofofthetheorem.
1 ]Zk=[g(h1h2)g−1(h1h2)−1]Zk
1 ]Zk[gh2g−1h−1
2 ]Zk=[gh1g−1]Zk[gh2g−1h−1
2 ]Zk[h−1
11
byCalebEckhardtandSvenRaum
C*-superrigidityof2-stepnilpotentgroups
4 CharacterisinggroupC*-algebrasofnilpotentgroups
In this section we will prove Theorem B, which is not only one of our main results, but will furthermore
allow us to systematically invoke the structure of nilpotent groups when proving C∗-superrigidity results
in Section 5. Before proceeding to the proof of Theorem B, we prepare two lemmas providing important
ingredients.
Lemma4.1. Let Gbeatorsion-freenitelygeneratednilpotentgroupand H beavirtuallynilpotentgroup.
If C∗(G)≅ C∗(H),then H istorsion-freenitelygeneratedandnilpotent.
Proof. Weprovethelemmabyinductiononthenilpotencyclassof G. Iftheclassof Gisatmost 1,then
G is abelian. Since torsion-free abelian groups are C∗-superrigid by [33, Theorem 1], the isomorphism
C∗(G)≅ C∗(H) implies G≅ H. In particular, H is torsion-free nitely generated and nilpotent. Let us
assumethat Ghasclass candthattheconclusionofthelemmaisknownforallgroupsofnilpotencyclass
atmost c− 1.NotethatbyTheorem2.18,thegroup H istorsion-free.Let N H beamaximalniteindex
nilpotent normal subgroup. By Theorem C, we have the following identications of the centre of C∗(G)
and C∗(H).
Z(C∗G)= C∗(Z(G))
Z(C∗H)= C∗(Z(N))H.
Since N H has nite index, the action of H onZ(N) has nite orbits. Further, C∗(Z(N))H ≅
C∗(Z(G))≅ C(Tn) for some n∈ N follows from the identication C∗(H)≅ C∗(G). So Theorem 2.3
applies to the action H Z(N) by conjugation and we obtain the isomorphismZ(N)≅ Zn and know
that H actstriviallyonZ(N), implyingthatZ(N)⊂Z(H). SoZ(H)⊂Z(N)followsfrommaximality
of N H andinparticularZ(C∗H)= C∗(Z(N)). Considerthebreof C∗(G)≅ C∗(H)overthetrivial
characterofZ(N),whichbyTheorem2.16ontheonehandisisomorphicwith C∗(H~Z(N))andonthe
otherisatwistedgroup C∗-algebraof G~Z(G). Since C∗(H~Z(N))admitsacharacter,Proposition2.17
shows that C∗(G~Z(G)) ≅ C∗(H~Z(N)). The induction hypothesis applies to show that H~Z(N) is
torsion-freenitelygeneratedandnilpotent.SinceZ(N)≅ Znwasalreadyshown,weseethat H isacen-
tralextensionof Zn byatorsion-freenitelygeneratednilpotentgroup. So H itselfistorsion-freenitely
generatedandnilpotent.
Lemma4.2. Let G and H be two torsion-free nilpotent groups. If C∗(G)≅ C∗(H), then G and H have
thesameclassandthesubquotientsoftheiruppercentralseriesareisomorphic.
Proof. We prove the lemma by induction on the nilpotency class of G. If G is of nilpotency class at most
1,then Gisabelian. Sincetorsion-freeabeliangroupsare C∗-superrigidby[33,Theorem1],itfollowsthat
G≅ H. Assume now that the lemma is proven for all groups of nilpotency class at most c− 1 and let G
beofnilpotencyclass c. ByTheoremC,wehave C∗(Z(G))=Z(C∗(G))≅Z(C∗(H))= C∗(Z(H)). In
particular,Z(G)≅Z(H)by C∗-superrigidityoftorsion-freeabeliangroups. Thebreof C∗(G)overthe
trivialcharacterofZ(G)isisomorphicwith C∗(G~Z(G)),byTheorem2.16.Thesametheoremcombined
withProposition2.17showsthatthisbreisisomorphicwith C∗(H~Z(H)). Sotheinductionhypothesis
appliesandshowsthat G~Z(G)and H~Z(H)havethesamenilpotencyclass c− 1andthesubquotients
oftheiruppercentralseriesagree. Itfollowsthat Gand H havenilpotencyclass c. Moreover,wealready
showed thatZ(G)≅Z(H) completing the proof that the subquotients of the upper central series of G
and H areisomorphic.
12
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
Letusnowproceedtotheproofofthissection'smaintheorem.
ProofofTheoremB Weprovethistheorembyinductiononthenilpotencyclass cof G. If c= 0,then Gis
trivialandtheconclusionofthetheoremisobvious.If c= 1,then Gisatorsion-freeabeliangroupandthe
conclusion follows from C∗-superrigidity of such groups. Let us now assume that the theorem is proven
forgroupsofnilpotencyclassatmost c− 1andwewillshowitfor Gofnilpotencyclass c. Werstcollect
propertiesof C∗(G)usedinthesubsequentarguments.
• C∗(G)isamenablesince Gisamenable.
• Itadmitsnonon-trivialprojections,byTheorem2.18.
• Itscentreisisomorphicwith C(Tn)forsome n∈ N,byTheoremC.
• Itsprimitiveidealspaceis T1,byTheorem2.20.
red(H)≅ C∗(G)weconcludethat H isamenable. FurtherbyTheorem2.18, H
Fromtheisomorphism C∗
istorsion-free. Wenextshowthat H isnitelygenerated. ApplyingProposition2.6, wendthatthereis
a normal abelian subgroup A≤ H on which H acts with nite orbits such thatZ(C∗(H))= C∗(A)H.
CombinedwiththeisomorphismZ(C∗(H))≅Z(C∗G)≅ C(Tn)forsome n∈ N,Theorem2.3appliesto
show that H acts trivially on A. It follows thatZ(H)= A andZ(C∗(H))= C∗(A). Note that the bre
of C∗(H)atthetrivialcharacterof Ais C∗(H~A). So C∗(H~A)isisomorphicwithsomebreof C∗(G).
By Theorems C and 2.16, it follows that C∗(H~A) is isomorphic with a twisted group C∗-algebra of the
quotient G~Z(G). Now Proposition 2.17 implies that the twist is trivial and C∗(H~A)≅ C∗(G~Z(G)).
Sotheinductionhypothesisappliestoshowthat H~Aisatorsion-freenitelygeneratednilpotentgroup.
Because Aand H~Aarenitelygenerated, H isalsonitelygenerated. Since H isnitelygeneratedand
hasa T1 primitiveidealspace,Theorem2.20saysthat H isvirtuallynilpotent. Lemma4.1appliestoshow
that H isactuallytorsion-freenitelygeneratedandnilpotent.Lemma4.2saysthattheclassof H andthe
classof Gagreeandthatsubquotientsoftheiruppercentralseriesareisomorphic.
5 C*-superrigidityof2-stepnilpotentgroups
InthissectionweproveTheoremAshowingthatalltorsion-freenitelygenerated2-stepnilpotentgroups
are C∗-superrigid.TheoremAisderivedfromthetechnicalTheorem5.1,whosenotationwenowintroduce.
Let G be a torsion-free nitely generated 2-step nilpotent group with centre Z. Theorem C provides
a natural isomorphism ϕZ ∶ Z → Glimm(C∗G) induced by the Fourier transform. We denote the in-
duced isomorphism of fundamental groups by ϕZ = π1( ϕZ) ∶ Z → π1(Glimm(C∗G)). Further, if
ε∈ Glimm(C∗G) is an element such that the bre C∗(G)ε admits a character, we obtain a natural iso-
morphism C∗(G)ε ≅ C∗(G~Z) from Theorem 2.16 and Proposition 2.17.
In particular, the natural map
G~Z → K1(C∗(G)ε) is an isomorphism onto its image, which equals imU(C∗G)→ K1(C∗G)ε. We
denotethisisomorphismby ϕab∶ G~Z→ im(U(C∗G)→ K1(C∗G)ε).
RecallthenotationandElliott'sresultsonK-theoryofnoncommutativetorifromSection2.10. Inpar-
ticular,for χ∈ Glimm(C∗G)theisomorphism C∗(G)χ≅ C∗(G~Z, χ○ σ)fromTheorem2.16providesan
isomorphisminK-theory,sothatElliott'sworkgivesrisetoanisomorphism K∗(C∗(G)ε)→ K∗(C∗(G)χ),
whose restriction denes a unique map K0(C∗(G)ε)→ K0(C∗(G)χ)→ R~Z as described in Theorem
2.22. Wedenotethismapby τχ.
Wealsomakeuseofthecup-productinK-theoryoftheabelian C∗-algebra C∗(G)ε,forwhichwerefer
to[1].
13
.
Z
(5.1)
Z(χ)∈ Z and
fb1∧b2∶ Glimm(C∗G)→ R~Z∶ fb1∧b2(χ)= τχ(b1∪ b2).
Z(χ)○ ωG∈ Hom(G~Z∧ G~Z, S1). SoTheorem2.24appliesandsaysthat
A= π1(Glimm(C∗G)),
B= imU(C∗G)→ K1(C∗G)ε.
∀a∈ A
ω(b1∧ b2), a= π1(fb1∧b2)(a)
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
Theorem5.1. Let Gbeatorsion-freenitelygenerated2-stepnilpotentgroup. Let ε∈ Glimm(C∗G)such
that C∗(G)ε admitsacharacter. Put
Choosesomeisomorphism A≅ Zm inducingabilinearformon Aanddenean A-valuedskew-symmetric
form ω∈ Hom(B∧ B, A)bytheformula
where
Then(A, B, ω)∼(Z(G), G~Z(G), ωG)inthesenseofCorollary2.11.
Proof. Asintheintroductionofthissection,wewillwrite Z=Z(G). Westartbyshowingthat
fb1∧b2= evωG○(ϕ−1
ab)(b1∧b2)○ ϕ−1
ab∧ϕ−1
ab(bi)∈ G~Z for i∈{1, 2}. Let χ∈ Glimm(C∗G),consider ϕ−1
Let b1, b2∈ B andwrite gi= ϕ−1
notethat ϕ−1
Z(χ)○ ωG(g1∧ g2).
Z (χ)○ωG([g1]K1∪[g2]K1)= ϕ−1
τ ϕ−1
Theleft-handsidecanberewrittenas
Z (χ)○ωG([g1]K1∪[g2]K1)= τχ(b1∪ b2)
τ ϕ−1
thanks to the identication C∗(G)χ ≅ C∗(G~Z, χ○ ωG) from Theorem 2.16 and the denition of g1, g2.
Weobtain
Z(χ)ωG○(ϕ−1
whichisequivalenttoformula(5.1).
Weconsider Z equippedwiththebilinearforminducedfrom ϕZ∶ Z→ A. Let h∶ S1→ Glimm(C∗G)
bealoop. Then
ω○(ϕab∧ ϕab)(g1∧ g2),[h]= π1(fb1∧b2)([h])
(denitionof ω)
= π1(evωG(g1∧g2)○ ϕ−1
Z)([h])
(formula(5.1))
Z ○ h])
= π1(evωG(g1∧g2))([ ϕ−1
(elementarypropertyof π1)
Z ○ h]
=ωG(g1∧ g2),[ ϕ−1
(Proposition2.12)
Z([h])
=ωG(g1∧ g2), ϕ−1
(denitionof ϕZ)
=ϕZ○ ωG(g1∧ g2),[h].
(denitionofthebilinearformon Z)
Itfollowsthat ω○(ϕab∧ ϕab)= ϕZ○ ωG. Thisnishestheproofofthetheorem.
Wearenowreadytoprovethemainresultofthisarticle.
ProofofTheoremA Let Gbe atorsion-free nitely generated2-step nilpotentgroup and H some group
red(H). Theorem B says that H is also a torsion-free nitely generated nilpotent
such that C∗
group.Denoteby(A, B, ω)thetripleconstructedfrom C∗(G)bymeansofTheorem5.1.Then(A, B, ω)is
equivalentto(Z(G), G~Z(G), ωG)andto(Z(H), H~cZ(H), ωH)byTheorem5.1andtheisomorphism
C∗(G)≅ C∗(H). SoCorollary2.11impliesthat G≅ H,nishingtheproofofthetheorem.
red(G)≅ C∗
fb1∧b2(χ)= ϕ−1
ab∧ ϕ−1
ab)(b1∧ b2)
forall χ∈ Glimm(C∗G),
14
byCalebEckhardtandSvenRaum
C*-superrigidityof2-stepnilpotentgroups
References
[1] M.F.Atiyah. K-theory. NewYork-Amsterdam: W.A.Benjamin,Inc.,1967.
[2] R.Baer. Finitenesspropertiesofgroups. DukeMath.J.,15(4):1021 -- 1032,1948.
[3] E. Blanchard and I. Gogić. On unital C(X)-algebras and C(X)-valued conditional expectations of
niteindex. LinearMultilinearAlgebra,64(12):2406 -- 2418,2016.
[4] E.Breuillard,M.Kalantar,M.Kennedy,andN.Ozawa. C∗-simplicityandtheuniquetracepropertyfor
discretegroups. Publ.Math.Inst.HautesÉtud.Sci.,126(1):35 -- 71,2017.
[5] K.S.Brown.Cohomologyofgroups.,volume87ofGraduateTextsinMathematics.NewYork:Springer-
Verlag,1994. Correctedreprintofthe1982original.
[6] N.P.BrownandN.Ozawa.C*-algebrasandnite-dimensionalapproximations.,volume88ofGraduate
StudiesinMathematics. Providence,RI:AmericanMathematicalSociety,2008.
[7] R.Brown.Topologyandgroupoids.NorthCarleston,SC:Createspace,2006.Thirdeditionof"Elements
ofmoderntopology".
[8] I. Chifan and A. Ioana. Amalgamated free product rigidity for group von Neumann algebras. Adv.
Math.,329:819 -- 850,2018.
[9] A. Connes. Classication of injective factors. Cases II1, II∞, IIIλ, λ≠ 1. Ann. Math. (2), 74:73 -- 115,
1976.
[10] A.Connes.Classicationdesfacteurs.InOperatorAlgebrasandApplications,Part2(Kingston,Ontario,
1980),volume38ofProc.Sympos.PureMath.,pages43 -- 109.Providence,RI:AmericanMathematical
Society,1982.
[11] A. Deitmar and S. Echterho . Principles of harmonic analysis. Universitext. Cham-Heidelberg-New
York-Dordrecht-London: Springer-Verlag,2ndedition,2014.
[12] A.DuguidandD.McLain.FC-nilpotentandFC-solublegroups.Proc.Camb.Phil.Soc.,52:391 -- 398,1956.
[13] C.Eckhardt. Anoteonstronglyquasidiagonalgroups. J.Oper.Theory,73(2):417 -- 424,2015.
[14] G.A.Elliott. OntheK-theoryoftheC∗-algebrageneratedbyaprojectiverepresentationofatorsion-
In Operator algebras and group representations (Neptun, 1980), Vol I.,
free discrete abelian group.
volume17ofMonogr.Stud.Math.,pages157 -- 184.Boston,MA:Pitman,1984.
[15] J.M.G.Fell. Thestructureofalgebrasofoperatorelds. ActaMath.,106:233 -- 280,1961.
[16] N.HigsonandG.G.Kasparov. E-theoryand KK-theoryforgroupswhichactproperlyandisometrically
[17] K.H.HofmannandS.A.Morris. Thestructureofcompactgroups,volume25ofdeGruyterStudiesin
[18] N.J.S.Hughes. Theuseofbilinearmappingsintheclassicationofgroupsofclass2. Proc.Am.Math.
onHilbertspace. Invent.Math.,144(1):23 -- 74,2001.
Mathematics. Berlin: WalterdeGruyter&Co.,1998.
Soc.,2(5):742 -- 747,1951.
15
1989.
tion.
1956.
178:231 -- 286,2013.
C*-superrigidityof2-stepnilpotentgroups
byCalebEckhardtandSvenRaum
[19] A. Ioana, S. Popa, and S. Vaes. A class of superrigid group von Neumann algebras. Ann. Math. (2),
[20] S.A.Jennings. Thegroupringofaclassofinnitenilpotentgroups. Canad.J.Math.,7:169 -- 187,1955.
[21] E.KaniuthandK.F.Taylor. ProjectionsinC*-algebrasofnilpotentgroups. Manuscr.Math.,65:93 -- 111,
[22] A.Kleppner. Multipliersonabeliangroups. Math.Ann.,158:11 -- 34,1965.
[23] S.Knudby,S.Raum,H.Thiel,andS.White. OnC∗-superrigidityofvirtuallyabeliangroups. Inprepara-
[24] A.I.Mal'ev. Onaclassofhomogeneousspaces. Izv.Akad.NaukSSSR,Ser.Mat.,13:9 -- 32,1949.
[25] D. H. McLain. Remarks on the upper central series of a group. Proc. Glasg. Math. Assoc., 3:38 -- 44,
[26] C.C.MooreandJ.Rosenberg. Groupswith T1primitiveidealspace. J.Funct.Anal.,22:204 -- 224,1976.
[27] J.R.Munkres. Topology. UpperSaddleRiver,NJ:PrenticeHall,2ndedition,2000.
[28] T.Omland. FreenilpotentgroupsareC*-superrigid. arXiv:1803.10208,2018.
[29] J. A. Packer and I. Raeburn. On the structure of twisted group C∗-algeras. Trans. Am. Math. Soc.,
[30] M. A. Rie el. Projective modules over higher-dimensional non-commutative tori. Can. J. Math.,
[31] J.RonghuiandS.Pedersen. CertainfullgroupC*-algebraswithoutproperprojections. J.Oper.Theory,
[32] J. Rosenberg. C*-algebras, positive scalar curvature, and the Novikov conjecture. Publ. Math. Inst.
[33] S. Scheinberg. Homeomoerphism and isomphism of abelian groups. Can. J. Math., 26(6):1515 -- 1519,
[34] A.Valette. IntroductiontotheBaum-Connesconjecture. LecturesinMathematics: ETHZürich.Basel:
[35] D. P. Williams. Crossed products of C∗-algebras, volume 134 of Mathematical Surveys and Mono-
334(2):685 -- 718,1992.
40(2):257 -- 338,1988.
24:239 -- 252,1990.
HautesÉtud.Sci.,58:197 -- 212,1983.
1974.
BirkhäuserVerlag,2002.
graphs. Providence,RI:AmericanMathematicalSociety,2007.
SvenRaum
DepartmentofMathematics
StockholmUniversity
SE-10691Stockholm
Sweden
[email protected]
CalebEckhardt
DepartmentofMathematics,MiamiUniversity
Oxford,OH45056,USA
[email protected]
16
|
1209.6294 | 1 | 1209 | 2012-09-27T17:28:25 | The mathematical legacy of William Arveson | [
"math.OA",
"math.FA"
] | This is a retrospective of some of William Arveson's many contributions to operator theory and operator algebras. | math.OA | math |
THE MATHEMATICAL LEGACY OF WILLIAM ARVESON
KENNETH R. DAVIDSON
Dedicated to the memory of Bill Arveson
Abstract. This is a retrospective of some of William B. Arveson's
many contributions to operator theory and operator algebras.
1. Introduction
William Barnes Arveson, born in November 1934, completed his doctorate
in 1964 at UCLA under the supervision of Henry Dye. After an instructor-
ship at Harvard, Bill began a long career at the University of California,
Berkeley. He died in November 2011, shortly before his 77th birthday, of
complications resulting from surgery.
Arveson's work was deep and insightful, and occasionally completely rev-
olutionary. When he attacked a problem, he always set it in a general frame-
work, and built all of the infrastructure needed to understand the workings.
This is one of the reasons that his influence has been so pervasive in many
areas of operator theory and operator algebras.
He worked in both operator theory and operator algebras. I think it is
fair to say that he did not really distinguish between the two areas, and
proved over and over again that these areas are inextricably linked. In [9],
he writes in the introduction, and in fact demonstrated many times, that:
"In the study of families of operators on Hilbert space, the
self-adjoint algebras have occupied a preeminent position.
Nevertheless, many problems in operator theory lead obsti-
nately toward questions about algebras that are not neces-
sarily self-adjoint."
Bill's work moves across the landscape of operator theory and operator al-
gebras, and he developed many foundational ideas which wove the subject
together as an integrated whole.
It isn't possible in a short article to review all of Bill's contributions. I will
cover many of the highlights. The choices made reflect my personal taste,
and I extend my apologies to those who feel that their favourite Arveson
2010 Mathematics Subject Classification. 47A, 47D, 47L, 46L.
Key words and phrases.
dilation theory, completely positive maps, invariant sub-
spaces, maximal subdiagonal operator algebras, dynamical systems, commutative sub-
space lattices, nest algebras, automorphism groups, multivariable operator theory.
Author partially supported by an NSERC grant.
1
2
K. R. DAVIDSON
paper is omitted. There will be one glaring omission -- E0-semigroups and
CP semigroups. Beginning with [21], Arveson wrote 24 papers (more than
650 pages) and a book (450 pages) on this topic. It is too large a topic to
cover in this article, and Masaki Izumi has written a survey [69] of this area
to appear in this same volume.
Bill Arveson was well-known to be an outstanding expositor. His papers
were always meticulously written. In addition, he wrote three graduate level
texts. His An invitation to C*-algebras [17] is distinguished from other C*-
algebra texts in that it develops in detail the structure of type I C*-algebras
and their representation theory. A short course on spectral theory [24] is a
brief introduction to Banach algebras and operator theory. Noncommutative
dynamics and E-semigroups [26] is a full development of the theory of E0-
semigroups.
Bill was also a dedicated supervisor. According to the Mathematics Ge-
nealogy Project, he had 27 doctoral students. At the memorial ceremony in
Berkeley in February, many of his former students paid their respects and
told stories of Bill as a mentor, inspiration and friend. He will be greatly
missed. But his mathematics will live on.
2. Dilation theory
I will begin with Arveson's contributions to dilation theory. While this is
not his first work, it is perhaps his most influential. It has had ramifications
in many aspects of operator algebras and operator theory.
In 1953, Bela Sz.Nagy [102] showed that for every operator T on Hilbert
space of norm at most one, there is a dilation to a unitary operator U on a
larger Hilbert space K = K− ⊕ H ⊕ K+ of the form
where ∗ represents an unspecified operator.
In particular, the map from
polynomials in U to polynomials in T obtained by compression to H is easily
seen to be a contractive homomorphism. This recovers the von Neumann
inequality for contractions:
kp(T )k ≤ kpk∞ := sup
z≤1
p(z)
for all p ∈ C[z].
When T does not have a unitary summand, this map extends to yield a
map from the algebra H ∞ of bounded analytic functions on the unit disk into
the weakly closed algebra that T generates. This allows the transference of
powerful techniques from function theory to operator theory. See [103, 104]
for more in this direction.
These methods extended, in part, to pairs of commuting contractions. In
1964, Ando [5] showed that any commuting pair of contractions simultane-
ously dilate to a pair of commuting unitaries. In particular, they also satisfy
U =
∗ ,
0
∗
0
∗ T 0
∗
∗
ARVESON'S LEGACY
3
von Neumann's inequality for polynomials in two variables, where the norm
on polynomials is the supremum over the bidisk D2.
In the one variable
case, the minimal dilation is unique in a natural sense. However, difficulties
arise in the two variable case because this dilation is no longer unique.
For commuting triples, there is a serious road block. Briefly skipping
ahead in the history, Parrott [80] used Arveson's ideas to construct three
commuting contractions which do not dilate to commuting unitaries. Fur-
thermore, examples developed later from the work of Varopoulos [108] on
tensor product norms showed that even von Neumann's inequality can fail.
In two papers in Acta Math. [11, 13], Arveson revolutionized dilation
theory by reformulating it in the framework of an arbitrary operator algebra.
The algebra in Sz.Nagy's case is the disk algebra A(D), and for Ando is the
bidisk algebra A(D2).
There were a number of key ideas that are central to Arveson's approach.
First, an operator algebra A was a considered as a subalgebra of a C*-
algebra B. Second, the correct representations were completely contractive,
and yielded completely positive maps on A + A∗. Third, he claimed that
every operator algebra A lived inside a unique canonical smallest C*-algebra
called the C*-envelope. These ideas completely changed the way we look at
dilations.
To elaborate, the k ×k matrices over the C*-algebra B, Mk(B), is itself a
C*-algebra and thus carries a unique norm. This induces a family of norms
on the matrix algebras over A. Today we say that this gives A an operator
space structure. A map ϕ from A into B(H) (or into any C*-algebra) induces
maps ϕk from Mk(A) into Mk(B(H)) ≃ B(H(k)) by acting by ϕ on each
entry:
We say that ϕ is completely bounded when
ϕk((cid:2)aij(cid:3) =(cid:2)ϕ(aij)(cid:3).
kϕkcb := sup
k≥1
kϕkk < ∞.
It is completely contractive when kϕkcb ≤ 1. When A is an operator system
(a unital self-adjoint subspace), we say ϕ is completely positive when ϕk is
positive for all k ≥ 1.
Completely positive maps on C*-algebras were introduced by Stinespring
[100], where he proved a fundamental structure theorem for completely
positive maps. The notion was used by Størmer [101] in a study of positive
maps on C*-algebras. However, Arveson really recognized the power of this
idea, and turned completely positive maps into a major tool in operator
theory.
Every completely contractive unital map ρ of a unital operator algebra
A extends uniquely to a completely positive map ρ on the operator system
A + A∗. Arveson's Extension Theorem established that this map extends
to a completely positive map on the C*-algebra B.
π(a) =
∗ for all a ∈ A.
0
∗
0
∗ ρ(a) 0
∗
∗
4
K. R. DAVIDSON
Theorem 2.1 (Arveson's extension theorem 1969). Let S be an operator
system in a unital C*-algebra B. If ϕ is a completely positive map from S
into B(H), then there is a completely positive extension Φ of B into B(H).
In modern terminology, this says that B(H) is injective in the category of
operator systems with completely positive maps. This is a generalization of
Krein's theorem that positive linear functionals have a positive extension.
The norm is not increased because kΦk = kΦ(1)k = kϕ(1)k = kϕk.
Stinespring's Theorem shows that Φ is a corner of a ∗-representation π of
B. When restricted to A, we obtain a completely contractive representation
of the form
This is called a ∗-dilation of ρ. So we have
Theorem 2.2 (Arveson's dilation theorem 1969). Let A be a unital operator
algebra, and let ρ : A → B(H) be a representation. The following are
equivalent:
(i) ρ is completely contractive.
(ii) ρ is completely positive.
(iii) ρ has a ∗-dilation.
This result was highlighted in Paulsen's book [83] and the expanded revi-
sion [84]. In Arveson's paper, this fundamental result is a remark following
Theorem 1.2.9.
Arveson was focussed on the notion of a boundary representation, that
is, an irreducible ∗-representation π of B with the property that πA has a
unique completely positive extension to B. These representations are the
non-commutative analogue of points in the spectrum of a function algebra
which have a unique representing measure -- points in the Choquet boundary.
When there are sufficiently many boundary representations to represent A
completely isometrically, then there is a quotient of B which is the unique
minimal C*-algebra containing A completely isometrically. This is the C*-
envelope of A.
Arveson was only able to establish the existence of this C*-envelope in a
variety of special cases. For example, if B contains the compact operators
and the quotient by the compacts is not completely isometric on A + A∗,
then B itself is the C*-envelope. He analyzed a variety of singly generated
examples, always from the boundary representation viewpoint. In particu-
lar, Arveson introduced the notion of matrix range of an operator T :
Wn(T ) =(cid:8)ϕ(T ) : ϕ : C∗(T ) → Mn unital completely positive(cid:9) for n ≥ 1.
In [12], he proved that the matrix range of an irreducible compact operator
formed a complete unitary invariant. Finding unitary invariants for opera-
tors was an important open problem. This provides a new invariant even for
ARVESON'S LEGACY
5
matrices. This was strengthened in [13] to show that for any irreducible op-
erator T such that C∗(T ) contains the compact operators, the matrix range
is again a complete unitary invariant.
Theorem 2.3 (Arveson 1972). If T1 and T2 are irreducible operators such
C∗(Ti) contains the compact operators, then T1 and T2 are unitarily equiva-
lent if and only if
Wn(T1) = Wn(T2)
for all n ≥ 1.
A decade later, Hamana [66] established the existence of a unique mini-
mal injective operator system containing A, the injective envelope, and from
this deduced the existence of the C*-envelope in complete generality. How-
ever it shed little light on the existence of boundary representations.
In
2005, Dritschel and McCullough [55] produced a very different proof of the
existence of the C*-envelope based on ideas of Agler [1]. It was more direct,
and it showed that every representation had a maximal dilation, which is
a representation such that any further dilation (to a completely contractive
representation) can be obtained only by adjoining another representation by
a direct sum. This property in another guise was recognized by Muhly and
Solel [78], but their result relied on Hamana's theorem. This approach did
not deal with the question of irreducible representations either, so it still did
not establish the existence of boundary representations. Arveson revisited
this question in the light of this new proof, and established the existence of
sufficiently many boundary representations when A is separable [28].
In the past four decades, many tools have been developed for computing
C*-envelopes, and they can now be computed for many classes of examples.
They have become a basic tool for the study of nonself-adjoint operator
algebras.
Completely bounded maps took on a life of their own in the 1980s. Witt-
stock [110] proved the extension theorem for completely bounded maps.
Paulsen [82] showed that this follows directly from Arveson's extension the-
orem using a 2 × 2 matrix trick.
The notion of completely bounded and completely positive maps has had
a profound influence on self-adjoint operator algebras. While the goal of
Bill's paper was dilation theory, the theory of completely positive maps has
had a life of its own in the self-adjoint theory. The injectivity of B(H) led
to deep work on injective von Neumann algebras by Connes [46]. This was
followed by work on nuclear C*-algebras by Lance [73], and work by Choi,
Effros and Lance determining the tight connection between nuclearity of
a C*-algebra and the injectivity of the von Neumann algebra it generates
[57, 40, 42].
When Brown, Douglas and Fillmore [35, 36] did their groundbreaking
work on essentially normal operators and the Ext functor (a K-homology
theory for C*-algebras), Arveson pointed out how one obtains inverses in
the Ext group using completely positive maps. Later he wrote an important
paper [18] introducing quasicentral approximate units for C*-algebras, and
6
K. R. DAVIDSON
used this to provide a unified and transparent approach to Voiculescu's
celebrated generalized Weyl-von Neumann theorem [109], the Choi-Effros
lifting theorem [41], and the structure of the Ext groups for nuclear C*-
algebras.
Completely positive and completely bounded maps also play a central
role in Kadison's similarity problem, which asks whether every bounded
representation of a C*-algebra is similar to a ∗-representation. This was
confirmed for nuclear C*-algebras by Bunce [37] and Christensen [44]. Then
Haagerup [64] established the result for cyclic representations of arbitrary
C*-algebras. Along the way, he showed that a representation was similar
to a ∗-representation if and only if it was completely bounded. See Pisier's
monograph [88].
The theory of CP and CB maps led to a considerable revolution on
the nonself-adjoint side as well. One important application concerns the
Sz.Nagy-Halmos problem. An operator T is called polynomially bounded
operator if there is a constant C so that
kp(T )k ≤ Ckpk∞ for all p ∈ C[x].
It is completely polynomially bounded if this inequality holds for matrices
of polynomials. Halmos's refinement of Sz.Nagy's question asks whether
every polynomially bounded operator is similar to a contraction. Paulsen
[81] showed that a representation of a unital operator algebra is similar to a
completely contractive one if and only if it is completely bounded. He used
this to show that every completely polynomially bounded operator is similar
to a contraction [82] . Then Pisier [87] showed that there are polynomi-
ally bounded operators which are not similar to a contraction, solving the
Sz.Nagy-Halmos problem. See [84, 88] for good coverage of this material.
The notions of matrix norms and completely bounded maps led to an ab-
stract categorical approach to abstract operator systems, operator algebras
and operator spaces. This is significant because it frees the theory from
reliance on spatial representations. Choi and Effros [43] characterized ab-
stract operator systems. Ruan [96] established a GNS style representation
theorem for operator spaces; and Blecher, Ruan and Sinclair [34] did the
same for unital operator algebras. See [58, 89, 33] for a treatment of this
material.
3. Maximal subdiagonal algebras
In his early work, Arveson used nonself-adjoint operator algebras to find
non-commutative analogues of results in analytic functions theory, to study
dynamics on topological spaces, to study invariant subspaces, and even to
study the corona problem of Carleson. We will survey these results in the
next few sections.
K. Hoffman [67] defined a logmodular function algebra to be a function
algebra A on a compact Hausdorff space X such that the set {log h : h ∈
A−1} is norm dense in CR(X). He showed that this class shared many
ARVESON'S LEGACY
7
properties that the algebra H ∞ of bounded analytic functions on the disk
enjoys. In fact H ∞ satisfies the stronger property that every invertible real
function f in L∞ has the form log h for an invertible H ∞ function. Indeed
one can take h = exp(f + i f ) where f is the harmonic conjugate of f . We
say that H ∞ is strongly logmodular, or that it satisfies factorization.
Srinivasan and Wang [99] defined a weak-∗ closed subalgebra A of L∞(m)
to be a weak-∗ Dirichlet algebra if A + A∗ is weak-∗ dense in L∞(m) and the
functional ϕ(f ) = R f dm is multiplicative. Hoffman and Rossi [68] show
that these algebras are logmodular. A variety of other properties such as
factorization, variants of Szego's theorem and Beurling's invariant subspace
theorem are eventually shown to be equivalent to this property. See [32] for
a survey of this and the non-commutative generalizations.
In the paper [9] that came out of Arveson's thesis [7], he developed a non-
commutative setting for logmodularity. This setting is a finite von Neumann
algebra M with a faithful normal state τ . A weak-∗ closed subalgebra A
is tracial if there is a normal expectation Φ of M onto A ∩ A∗ which is a
homomorphism on A: Φ(ab) = Φ(a)Φ(b) for all a, b ∈ A, and preserves the
trace: τ Φ = τ . A (finite) maximal subdiagonal algebra is a tracial subalgebra
of M such that A + A∗ is weak-∗ dense in M .
It isn't hard to see that logmodularity of a function algebra is equivalent
to saying that {h2 : h ∈ A−1} is dense in CR(X), and that factorization is
the stronger property that every invertible real continuous function on X is
exactly h2 for some invertible h ∈ A. A subalgebra A of a C*-algebra B is
called logmodular if {a∗a : a ∈ A−1} is dense in Bsa, and has factorization if
this set coincides with the set of all invertible self-adjoint elements. By the
polar decomposition, this latter property is equivalent to saying that every
invertible element of B factors as b = ua for some a ∈ A and unitary u ∈ B.
Arveson [9] established the analogue of the Hoffman-Rossi result:
Theorem 3.1 (Arveson 1967). Every maximal subdiagonal algebra satisfies
factorization.
An interesting class of examples is obtained as follows. Start with a chain
L of projections (in the usual order) in a finite von Neumann algebra M
such that there is a normal expectation onto the commutant L′. Then
Alg L = {a ∈ M : ae = eae
for all e ∈ L}
is a maximal subdiagonal algebra. Arveson also provides interesting exam-
ples associated to ergodic group actions on measure spaces.
There are a variety of other important properties that H ∞ satisfies, and
surprisingly many of them have natural non-commutative analogues. To
describe them, we need to introduce the Fuglede-Kadison determinant [60]
which they define on II1 factors, and Arveson extends to all finite von
Neumann algebras. Define ∆(a) = eτ (log a) when a is invertible, and in
general
∆(a) = inf
ε>0
eτ (log(a+ε1))
8
K. R. DAVIDSON
This map is multiplicative, positive homogeneous, monotone, self-adjoint;
and it is norm continuous on the invertible elements. Arveson makes this
much more tractable by showing that
∆(a) = inf τ (ab) : b ∈ M −1, b > 0, τ (b) ≥ 1}.
From this it follows that
∆(a) ≤ ∆(Φ(a))
for all a ≥ 0.
Now if h ∈ H ∞, then Jensen's inequality says that
log h(0) ≤
log h(eit) dt.
1
2πZ 2π
0
Jensen's formula states that this is an equality when h is invertible. We also
mention Szego's theorem, which states that if f is a positive function in L1,
then
inf
h∈H∞, h(0)=1
1
2πZ 2π
0
h(eit)2f (t) dt = exp(cid:16) 1
2πZ 2π
0
log f (t) dt(cid:17).
Arveson defines the analogues in the non-commutative setting: a tracial
algebra A satisfies Jensen's inequality if
∆(Φ(a)) ≤ ∆(a)
for all a ∈ A
and satisfies Jensen's formula if
∆(Φ(a)) = ∆(a)
for all a ∈ A−1.
Szego's formula becomes:
∆(b) = inf{τ (ba∗a) : a ∈ A, ∆Φ(a) ≥ 1}.
Arveson [9] established that these three properties are equivalent, and that
they are satisfied in many classes of examples. Labuschagne [72] proved that
these properties hold for all maximal subdiagonal algebras. Then Blecher
and Labuschagne [31] showed that for the class of tracial algebras, these
properties are equivalent to logmodularity, and to factorization, and to the
weak-∗ Dirichlet property. See [32] for a survey of this material.
4. Dynamical systems
Arveson was interested in using operator algebras to obtain complete
invariants for a discrete dynamical system in both the topological and the
measure theoretic settings.
In the measure setting, consider an ergodic transformation τ of a measure
space (X, m) where m is a probability measure. Two transformations σ and
τ are conjugate if there is a measure preserving automorphism Φ of X so
that σ = Φτ Φ−1. An important problem is to determine when two ergodic
transformations are conjugate.
ARVESON'S LEGACY
9
This is easily transferred to a statement about operator algebras. Take
M = L∞(m) considered as acting on L2(m) by multiplication. The measure
preserving transformation τ induces an automorphism of M by
α(Mf ) = Mf ◦τ
so that α(χE) = χτ −1(E).
Ergodicity means that the only fixed points of α are the scalars. If σ induces
the automorphism β, then conjugacy of σ and τ converts to conjugacy of α
and β in the automorphism group of M .
Every automorphism α of M determines a unique unitary map U as
follows. Since L∞(m) is a dense subset of L2(m), we set
U (g) = α(Mg)1 for all g ∈ L∞(m).
Then one readily checks that this extends to a unitary operator such that
α(Mg) = U MgU ∗
for all g ∈ L∞(m).
The conjugacy of two transformations reduces to unitary equivalence of the
two automorphisms.
Arveson's idea was to construct an operator algebra A(α) as the norm
closure of
A0(α) =n kXi=1
MfiU i : fi ∈ L∞(m), k ≥ 0o.
His main result in [8] is that conjugacy of two automorphisms α and β is
equivalent to the unitary equivalence of A(α) and A(β).
His results actually got a lot more mileage in the topological setting,
which he and Josephson developed in [29]. Begin with a locally compact
Hausdorff space X and a proper map σ of X into itself. As we shall see,
construction of a nonself-adjoint operator algebra does not require the map σ
to be invertible. Two dynamical systems (X, σ) and (Y, τ ) are (topologically)
conjugate if there is a homeomorphism γ of Y onto X so that σ = γτ γ−1.
Again the issue is to decide when two systems are conjugate.
The map σ induces an endomorphism α of C(X) by composition. When
σ is a homeomorphism, this map is an automorphism. So we consider a
dynamical system as the pair (C(X), α). Arveson and Josephson construct
a nonself-adjoint operator algebra in an analogous way to the measure the-
oretic setting. Their construction depended on the existence of a quasi-
invariant measure and small fixed point set. Indeed they put rather stringent
conditions on the maps:
Suppose that (X, σ) and (Y, τ ) are two topological dynamical systems
satisfying the following conditions
(i) σ and τ are homeomorphisms.
(ii)
there is a regular Borel probability measure µ on X which is quasi-
invariant for σ, i.e. µ◦σ and µ are mutually absolutely continuous;
(iii) µ(O) > 0 for every non-empty open set O ⊂ X;
(iv)
the set P = {x ∈ X : σn(x) = x for some n ≥ 1} of periodic points
has measure zero;
10
K. R. DAVIDSON
(v)
there is a regular Borel probability measure ν on Y with these same
properties, and in addition τ is ergodic with respect to ν.
Arveson and Josephson use these properties to build algebras A(α) and A(β)
as in the measure theoretic setting, but using continuous functions rather
than L∞(µ), by constructing an explicit representation on L2(µ).
What I will describe instead is a more general construction due to Peters
[85] called the semicrossed product. This has the major advantages that
it does not depend on any measure, and we may consider non-invertible
systems (X, τ ) where τ is any proper continuous map from X into itself.
The key new notion is that of a covariant representation (π, V ) where
π is a ∗-representation of C(X) on a Hilbert space H and V ∈ B(H) is a
contraction such that
V π(f ) = π(α(f ))V
for all f ∈ C(X).
Peters defines the operator algebra C(X) ×α Z+ to be the universal operator
algebra for covariant representations. To form this algebra, let A0(α) be the
i=0 fiτ i for fi ∈ C(X) and k ≥ 0.
set of all formal polynomials of the formPk
Define a norm by
k
kXi=0
fiτ ik := supn(cid:13)(cid:13)
kXi=1
π(fi)V i(cid:13)(cid:13) : (π, V ) is covarianto.
The semicrossed product is the completion of A0(α) in this norm.
Peters shows that a sufficient family of covariant representations is given
by the orbit representations: fix x ∈ X and define a representation πx
πx(f ) = diag(cid:0)f (x), f (τ (x)), f (τ 2(x)), . . .(cid:1)
and set V to be the unilateral shift. The direct sum over a dense subset of X
yields a faithful completely isometric representation of A(α). In particular,
τ is represented by an isometry.
Peters shows that subject to the Arveson-Josephson hypotheses, their
algebra A(α) coincides with the semicrossed product C(X) ×α Z+.
I can now state the Arveson-Josephson result in this modern terminology.
Theorem 4.1 (Arveson-Josephson 1969). Suppose that (X, σ) and (Y, τ )
are two topological dynamical systems, and let α and β be the automorphisms
of C(X) and C(Y ) induced by σ and τ respectively. Furthermore assume
conditions (i) -- (v) listed above. Then the following are equivalent:
(X, σ) and (Y, τ ) are conjugate.
(1)
(2) C(X) ×α Z+ and C(Y ) ×β Z+ are completely isometrically isomor-
phic.
(3) C(X) ×α Z+ and C(Y ) ×β Z+ are algebraically isomorphic.
Peters improved on this result considerably by removing the conditions
(i), (ii), (iii) and (v) at the expense of a slight strengthening of (iv) to
(a) X is compact.
ARVESON'S LEGACY
11
(b) σ has no periodic points.
He obtained the same equivalence. Hadwin and Hoover [65] further im-
proved on this. Finally Davidson and Katsoulis [50] removed all of these
conditions, and showed that (1), (2) and (3) are equivalent for arbitrary
proper maps of locally compact Hausdorff spaces with no conditions on pe-
riodic points at all.
5. Invariant subspaces
The invariant subspace problem, which remains open today, was a popular
topic four decades ago. Arveson made some important contributions, which
as usual, stressed operator algebras rather than single operator theory.
If A is an algebra of operators on a Hilbert space H, let
Lat A = {M is a closed subspace of H : AM ⊂ M },
and if L is a collection of subspaces, then
Alg(L) = {T ∈ B(H) : T M ⊂ M for all M ∈ L}.
Halmos called an operator algebra reflexive if A = Alg Lat A. This algebra
is always unital and closed in the weak operator topology (wot-closed). An
operator algebra is transitive if it has no proper closed invariant subspace.
The transitive algebra problem asks whether there is a proper unital wot-
closed transitive operator algebra.
Arveson's first contribution to invariant subspace theory recognized the
advantage of having a masa (maximal abelian self-adjoint algebra) in A.
Note that he shows that the algebra is dense in the weak-∗ topology, rather
than weak operator topology. This is a subtle, and occasionally non-trivial,
strengthening.
Theorem 5.1 (Arveson's density theorem 1967). A transitive subalgebra of
B(H) which contains a masa is weak-∗ dense in B(H).
This result received a fair bit of interest, and a generalization was found
if A is an operator algebra containing a
by Radjavi and Rosenthal [93]:
masa, and Lat A is a chain (nest), then the wot-closure of A is reflexive.
Perhaps this result prompted Bill to revisit the issue. He decided to study
weak-∗ closed operator algebras containing a masa, to determine whether
all such operator algebras are reflexive. We assume that the Hilbert space
is separable. The masa D is spatially isomorphic to L∞(µ) acting on L2(µ)
for some sigma-finite Borel measure µ. If A ⊃ D, then Lat A ⊂ Lat D. If
M ∈ Lat D, then the orthogonal projection PM commutes with D; and by
maximality, PM belongs to D. So Lat A can be naturally identified with a
sublattice of the lattice of projections in D. In particular, these projections
all commute. In addition, the lattice is complete (as a lattice, and as a sot-
closed set of projections). Such a lattice is called a commutative subspace
lattice (CSL).
12
K. R. DAVIDSON
In a 100 page tour de force in the Annals of Mathematics, Arveson [14]
developed an extensive theory for these operator algebras. The first major
result is a spectral theorem for CSLs.
Theorem 5.2 (Arveson 1974). If L is a CSL on a separable Hilbert space,
there is a compact metric space X, a regular Borel measure µ on X, and a
Borel partial order (cid:22) on X so that L is unitarily equivalent to the lattice of
projections onto the (essentially) increasing Borel sets.
Then he develops a class of pseudo-integral operators and identifies when
one is supported on the graph of the partial order; and hence belongs to
Alg L. Curiously, while the pseudo-integral operators seem to depend on the
choice of coordinates, the weak-∗ closure of the pseudo-integral operators
supported on the graph of the partial order is independent of all choices.
He calls this algebra Amin(L). He shows that Lat Amin(L) = L and hence
Lat Alg L = L for all CSLs. One of the main results is:
Theorem 5.3 (Arveson 1974). If L is a CSL on a separable Hilbert space,
then there is a weak-∗ closed operator algebra Amin(L) containing a masa
with Lat(Amin(L)) = L so that: a weak-∗ closed operator algebra A contain-
ing a masa has Lat(A) = L if and only if Amin ⊆ A ⊆ Alg(L).
Surprisingly, it is possible for Amin(L) to be properly contained in Alg(L),
even if it is wot-closed. An example is constructed based on the failure of
spectral synthesis on S3 in harmonic analysis. He calls a lattice synthetic
if Amin(L) = Alg(L). Many examples of synthetic lattices are produced,
including lattices which are generated by finitely many commuting chains,
and lattices associated to certain ordered groups. This is a definitive answer
to the question.
The area of CSL algebras has received a lot of attention. See [20] for a
survey of some of this material. Also see [49] for further material.
6. Nest algebras and the corona theorem
Kadison and Singer [71] initiated a theory of triangular operator algebras.
While their algebras were not necessarily wot-closed or even norm-closed,
Ringrose [94, 95] developed a related class that are wot-closed and re-
flexive. He calls a complete chain of subspaces N a nest. A nest algebra
is T (N ) := Alg N , the algebra of all operators which are upper triangular
with respect to the nest N . In the first paper, he characterized the Jacobson
radical. In the second, he showed that any algebra isomorphism between
two nest algebras is given by a similarity.
Arveson's paper [16] was very influential for the subject of nest algebras.
In particular, he developed a distance formula from an arbitrary operator
to T (N ). Note that if N ∈ N and T ∈ T (N ), the invariance of N may
be expressed algebraically by T PN = PN T PN , or equivalently P ⊥
N T PN = 0.
ARVESON'S LEGACY
13
Therefore if A ∈ B(H) and T ∈ T (N ), we see that
kP ⊥
N APN k = kP ⊥
N (A − T )PN k ≤ inf
kT − Ak = dist(A, T (N )).
T ∈T (N )
This inequality remains valid if we take the supremum over all N ∈ N .
Surprisingly, this leads to an exact formula:
Theorem 6.1 (Arveson's distance formula 1975). Let N a nest on a Hilbert
space H; and let A ∈ B(H). Then
dist(A, T (N )) = sup
N ∈N
kP ⊥
N APN k.
This and other results about nest algebras proved important in the devel-
opment of the theory.
A separable continuous nest is a nest which is order isomorphic to I =
[0, 1]. The Volterra nest N consisting of subspaces
Nt = L2(0, t) ⊂ H = L2(0, 1)
for 0 ≤ t ≤ 1
is an example which is multiplicity free, in that the projections onto Nt
generate a masa. The infinite ampliation M given by Mt = L2((0, t) ×
I) ⊂ L2(I 2) generates an abelian von Neumann algebra with non-abelian
commutant. So T (N ) and T (M) are not unitarily equivalent. Ringrose
[95] asked whether they could be similar.
Three students of Arveson's eventually solved this problem completely.
Andersen [4] showed that any two continuous nests are approximately uni-
tarily equivalent. So in particular, there is a sequence of unitaries Un so
that
[0, 1] ∋ t → PMt − UnPNtU ∗
n
are continuous, compact operator valued functions converging uniformly to
0.
(Arveson [19] generalized this to certain CSL algebras.) Larson [74]
used this to solve Ringrose's problem by showing that any two continuous
nests are similar. Shortly afterwards, Davidson [48] showed that if two nests
are order isomorphic via a map which preserves dimensions of differences of
nested subspaces, then there is a similarity of the two nests implementing
the isomorphism; and the similarity may be taken to be a small compact
perturbation of a unitary operator. See [49] for more material on nest
algebras.
Arveson was also interested in finding an operator theoretic proof of the
famous corona theorem of Carleson [39]. This result is formulated as follows:
suppose that f1, . . . , fn are functions in H ∞ for which there is an ε > 0 so
that
(†)
inf
z∈D
fi(z)2 ≥ ε.
Then there are functions gi ∈ H ∞ so that Pn
i=1 figi = 1. A reformulation
of this result is that the unit disk D is dense in the maximal ideal space
of H ∞. Carleson's proof was very difficult, and people continue to look for
nXi=1
14
K. R. DAVIDSON
more accessible proofs. An easier (but not easy) proof, based on unpublished
ideas of Thomas Wolff, is now available. See Garnett's book [61].
Arveson had an operator theoretic approach to the corona theorem. H 2
is the Hardy space of square integrable analytic functions on the disk D
considered as a subspace of L2(T). Represent H ∞ on H 2 as multiplication
operators Th (these are analytic Toeplitz operators). In [16], he established
this variation of Carleson's theorem:
Theorem 6.2 (Arveson's Toeplitz corona theorem 1975). If f1, . . . , fn ∈
H ∞ and
(‡)
nXi=1
then there are gi ∈ H ∞ so that Pn
TfiT ∗
fi ≥ εI,
i=1 figi = 1.
This result is weaker than the corona theorem because (‡) is a stronger
hypothesis than (†). To see this, we use the fact that H 2 is a reproducing
kernel Hilbert space: the functions
kw(z) =
(1 − w2)1/2
1 − z ¯w
are unit vectors for w ∈ D with the property that T ∗
h kw = h(w)kw. Applying
the vector state ρw(A) = hAkw, kwi to (‡) for w ∈ D, one recovers Carleson's
condition (†).
It turns out that Toeplitz corona theorems are useful in establishing full
corona theorems. Ball, Trent and Vinnikov [30] established a Toeplitz
corona theorem for all complete Nevanlinna-Pick kernels as a consequence
of their commutant lifting theorem. This was a useful step towards a bona
fide corona theorem by Costea, Sawyer and Wick [47] for Drury-Arveson
space, among other spaces. For definitions of these concepts, see section 8.
7. Automorphism groups
In [15], Arveson developed a spectral analysis of automorphism groups
acting on C*-algebras and von Neumann algebras.
Here G is a locally compact abelian group with Haar measure dt. A
representation is a bounded representation t → Ut in B(X), where X is
a Banach space, and the representation is assumed to be continuous with
respect to a certain weak topology. In particular, sot-continuous unitary
representations on Hilbert space form an important special case. A standard
technique extends this to a bounded representation πU of L1(G) such that
πU (f )ξ =ZG
f (t)Utξ dt.
We may regard the Fourier transform as the Gelfand transform of the
commutative Banach algebra L1(G) into C( G), where G is the dual group
ARVESON'S LEGACY
15
consisting of continuous characters of G. When f ∈ L1(G), its zero set is
Z(f ) = {γ ∈ G : f (γ) = 0}.
We define the spectrum of the representation U to be
Moreover, if ξ ∈ H, then the spectrum of ξ is
sp U =\{Z(f ) : f ∈ L1(G), πU (f ) = 0}.
spU (ξ) =\{Z(f ) : f ∈ L1(G), πU (f )ξ = 0}.
Then if E is a closed subset of G, we define the spectral subspace to be
M U (E) = {ξ : spU (ξ) ⊂ E}.
When U is a strongly continuous unitary representation on Hilbert space,
Stone's theorem yields a projection valued measure dPγ on G so that Ut =
R G γ(t) dPγ ; and M U (E) = P (E)H. But in more general contexts, there is
often no version of Stone's theorem. Nevertheless, the spectral subspaces
form a substitute, and they determine the representation uniquely:
Theorem 7.1 (Arveson 1974). Let X be a Banach space, and let U and V
be two strongly continuous representations of G on X. If M U (E) = M V (E)
for every compact subset E of G, then U = V .
An important application of this theory is to a one parameter group of
automorphisms acting on a von Neumann algebra R ⊂ B(H). That is,
t → αt is a weak-∗ continuous representation of R into Aut(R). Arveson was
interested in finding conditions that determine when this group is unitarily
implemented, i.e. when αt = Ad Ut, where U is a strongly continuous unitary
representation with Ut ∈ R and sp U ⊂ [0, ∞). Let Rα(E) = {A ∈ R :
spα(A) ⊂ E}. The answer is given by:
Theorem 7.2 (Arveson 1974). Let (αt) be a weak-∗ continuous representa-
tion of R into Aut(R). Then there is a strongly continuous unitary repre-
sentation U with Ut ∈ R and sp U ⊂ [0, ∞) such that αt = Ad Ut for t ∈ R
if and only if
Arveson used this to provide a new proof of the famous theorem of Kadi-
son [70] and Sakai [97] that every derivation of a von Neumann algebra is
inner. Alain Connes [45] used the Arveson spectrum to define the more re-
fined Connes spectrum of an automorphism group that allows one to define
the type IIIλ factors.
One can show that
Rα[s, ∞)Rα[t, ∞) ⊂ Rα[s + t, ∞).
Moreover Ut =R +∞
−∞ eitx dP (x) where P is the spectral measure such that
Rα[t, ∞) = {0}.
\t∈R
P [t, ∞)H =\s<t
Rα[s, ∞)H.
16
K. R. DAVIDSON
It follows that H ∞
α (R) = Rα[0, ∞) is a weak-∗ closed subalgebra of R. This
is sometimes a maximal subdiagonal algebra, but not always. A very curious
result of Loebl and Muhly [75] exhibits an example where H ∞
α (R) is a proper
nonself-adjoint subalgebra of R which is wot-dense in R. A variant on
the transitive algebra problem deals with reductive algebras, meaning that
every invariant subspace is reducing (or that the orthogonal complement of
every invariant subspace is invariant).
It asks whether every wot-closed
reductive operator is a von Neumann algebra. Loebl and Muhly's construct
is a reductive weak-∗ closed algebra which is not self-adjoint (but the wot-
closure is a von Neumann algebra).
I will now skip ahead to some more recent results. As I mentioned in
the introduction, in the intervening period, Bill was very interested in E0-
semigroups and CP-semigroups. This work will be covered in the article
[69] by Izumi in this volume.
8. Multivariable operator theory
As discussed in section 2, the Sz.Nagy dilation theorem led to a devel-
opment of function-theoretic techniques in single operator theory. In par-
ticular, if T is a contraction with no unitary summand, then it has an H ∞
functional calculus. This has many ramifications. See [103] for an updated
look at this material.
The examples of Parrott and Varopolous, also mentioned in section 2,
explain why there are problems with a straightforward generalization of
this dilation theory to several variables by studying d-tuples of commuting
contractions.
It turns out that a different choice of norm condition leads to a better
theory. The more tractable norm condition on an d-tuple T1, . . . , Td is a row
contractive condition: consider T =(cid:2)T1, . . . , Td(cid:3) as an operator mapping
the direct sum H(d) of d copies of H into H and require that
kT k =(cid:13)(cid:13)
dXi=1
TiT ∗
i(cid:13)(cid:13)1/2 ≤ 1.
In the non-commutative case, the appropriate model for the dilation theo-
rem is best described as the left regular representation of the free semigroup
d , consisting of all words in an alphabet {1, . . . , d}, acting on ℓ2(F+
F+
d ) by
Liξw = ξiw for 1 ≤ i ≤ d, w ∈ F+
d .
The space ℓ2(F+
d ) can be realized as the full Fock space over E = Cd, namely
ℓ2(F+
d ) ≃
⊕Xk≥0
E⊗k = C ⊕ Cd ⊕ Cd2
⊕ . . .
ARVESON'S LEGACY
17
where E⊗k = Cdk
is identified with the span of words of length k. In this
view, the operators Li become the creation operators that tensor on the left
by ei, where e1, . . . , ed is the standard basis for Cd.
. . . , Vd(cid:3) where V ∗
isometry V = (cid:2)V1,
The dilation theorem is that every row contraction T dilates to a row
i Vj = δijI. Furthermore, there is
a decomposition Vi ≃ L(α)
i ⊕ Wi where α is a cardinal and W1, . . . , Wd are
generators of a representation of the Cuntz algebra Od. This was established
for d = 2 by Frazho [59], for finite d by Bunce [38] and for arbitrary d
including infinite values, and uniqueness, by Popescu [90]. Popescu has
written a long series of papers developing the Sz.Nagy-Foia¸s theory in this
context. The wot-closed algebras generated by row isometries have also
been extensively studied by Popescu beginning with [91] and Davidson and
Pitts beginning with [51].
The dilation theory for commuting row contractions came earlier, but
was exploited later. Drury [56] showed that every commuting row contrac-
tion with kT k < 1 dilates to a direct sum of copies of a certain d-tuple of
weighted shifts, S =(cid:2)S1, . . . , Sd(cid:3), now called the d-shift. This was refined
Ni are commuting normal operators such thatPd
by Muller and Vasilescu [79] to show that if kT k ≤ 1, then T dilates to a
row contraction V with commuting entries so that Vi ≃ S(α)
i ⊕ Ni where the
i=1 NiN ∗
i = I, a spherical
isometry. Perhaps the reason this did not go further at the time is that the
exact nature of the d-tuple S was not realized.
Arveson began a program of multivariable operator theory in [22]. He re-
proved the Drury-Muller-Vascilescu dilation theorem. Moreover, he showed
that the d-shift has a natural representation on symmetric Fock space over
E = Cd. Let Ek denote the symmetric tensor product of k copies of E,
which can be considered as the subspace of E⊗k which is fixed by the action
of the permutation group Sk which permutes the terms of a tensor product
of k vectors. Then
H 2
d =Xk≥0
⊕ Ek ⊂ ℓ2(F+
d ).
This space is coinvariant for the creation operators Li, and Si = PH 2
LiH 2
d
.
d
There is a natural basis
zk =
k!
ξw,
k! Xw=k, w(z)=zk
0, k =Pi ki, k! :=Qi ki! and zk := zk1
1 zk2
where k = (k1, . . . , kd) ∈ Nd
This is an orthogonal but not orthonormal basis since kzkk2 = k!
In this basis, Si is just multiplication by zi. The space H 2
considered as the space of analytic functions
2 · · · zkd
d .
k! =: ck.
d may now be
f (z) = Xk∈Nd
0
akzk
such that Xk∈Nd
0
ak2ck < ∞.
18
K. R. DAVIDSON
These series converge on the unit ball Bd of Cd, so are bona fide functions.
This is a reproducing kernel Hilbert space (RKHS), meaning that w ∈ Bd
separate points and the value of f (w) may be recovered by
f (w) = hf, kwi where kw(z) =
1
1 − hz, wi
, w ∈ Bd.
The space H 2
d is now often called Drury-Arveson space.
Every RKHS has an algebra of multipliers, those functions which multiply
the space into itself. The multiplier algebra Md = Mult(H 2
d consists
of bounded analytic functions on the ball Bd. It is the wot-closed unital
algebra generated by S1, . . . , Sd. However the norm is greater than the sup
norm, and is not comparable -- so Md is a proper subalgebra of H ∞(Bd).
d ) of H 2
The C*-algebra C∗(S) := C∗(S1, . . . , Sn) contains the compact opera-
tors K, and C∗(S)/K ≃ C(∂Bd), the boundary sphere of Bd -- which is the
spectrum of the generic spherical isometry. So one deduces that this is the
C*-envelope of the unital nonself-adjoint algebra Ad generated by S1, . . . ,
Sd.
The classical Nevanlinna-Pick theorem [86] for the unit disk states that
given z1, . . . , zn ∈ D and w1, . . . , wn ∈ C, there is a function h ∈ H ∞ with
khk∞ ≤ 1 so that h(zi) = wi if and only if
1 − zizj #n×n
" 1 − wiwj
is positive definite. A matrix version states that given z1, ..., zn in the disk,
and r × r matrices W1, . . . , Wn, there is a function F in the the unit ball of
Mr(H ∞) such that F (zi) = Wi if and only if the Pick matrix
" Ir − WiW ∗
1 − zizj #n×n
j
is positive semidefinite. Sarason [98] put this into an operator theoretic
context.
An RKHS H of functions on a set X is a natural place to study interpola-
tion. One asks the same question: given x1, . . . , xn ∈ X and r × r matrices
W1, . . . , Wn, is there a function F in the the unit ball of Mr(Mult(H))
so that F (zi) = Wi? It is easy to show that a necessary condition is the
positive semidefiniteness of the matrix
h(Ir − WiW ∗
j )hkxj , kxiiin×n
.
A space for which this is also a sufficient condition is called a complete
Nevanlinna-Pick kernel.
Complete Nevanlinna-Pick kernels were characterized by Quiggin [92]
and McCullough [76]. It follows from their characterization that H 2
d is a
complete Nevanlinna-Pick kernel. This was shown from a completely differ-
ent angle, by showing that the non-commutative wot-closed algebra Ld on
ARVESON'S LEGACY
19
Fock space generated by L1, . . . , Ld has a good distance formula for ideals
[52, 6], and hence has a Nevanlinna-Pick theory based on Sarason's ap-
It follows that Md is a complete quotient of Ld, and hence is a
proach.
complete Nevanlinna-Pick kernel. Finally, Agler and McCarthy [2] give
a new proof of the Quiggin-McCullough theorem, and show that Md for
d = ∞ is the universal complete Nevanlinna-Pick kernel in the sense that
every irreducible complete Nevanlinna-Pick RKHS imbeds in a natural way
into H 2
d . See [3] for an overview of this theory.
The confluence of these very different directions leading to the same space,
H 2
d , is remarkable: the good dilation theory for commuting row contractions,
the close relationship with the non-commutative model Ld, and the fact that
this is a complete Nevanlinna-Pick kernel. All play a role in making H 2
d a
very fertile environment in which to study multivariable operator theory.
Associated to any row contraction T is the completely positive contractive
map
ϕ(A) =
TiAT ∗
i .
dXi=1
Notice that T is spherical if and only if ϕ(I) = I. We call a row contraction
T pure if there is no spherical part. The decreasing sequence
I ≥ ϕ(I) ≥ ϕ2(I) ≥ . . .
always has a limit in the strong operator topology. This limit ϕ∞(I) is 0 if
and only if T is pure.
In [23], Arveson begins a much more algebraic approach to multivariable
operator theory. Douglas particularly has espoused the approach of con-
sidering a Hilbert space H with an operator algebra A acting on it via a
representation ρ : A → B(H) as a Hilbert module over A with the action
a · ξ = ρ(a)ξ. This approach was developed for function algebras in the
monograph [53] by Douglas and Paulsen.
In this context, one begins with a commuting row contractive d-tuple T
acting on H, and consider H as a module over C[z] := C[z1, . . . , zd] given
by p · ξ = p(T1, . . . , Td)ξ. In order to use results from algebra effectively,
assume that the defect operator
∆ := (I − T T ∗)1/2 =(cid:0)I −
dXi=1
TiT ∗
i(cid:1)1/2
has finite rank ; and define rank(T ) := rank(∆). We also say that H is a
Hilbert module of the same rank.
Form a finitely generated module over C[z] as follows
MH := {p · ξ : p ∈ C[z], ξ ∈ ∆H}.
20
K. R. DAVIDSON
This space is not closed; so it is a finitely generated C[z] module alge-
braically. From commutative algebra (Hilbert's syzygy theorem), one ob-
tains a free resolution
0 → Fn → Fn−1 → · · · → F1 → MH → 0.
Each Fi is a finitely generated free C[z]-module of rank βi < ∞. One can
define an invariant known as the Euler characteristic
χ(H) :=
(−1)i+1βi
nXi=1
which is independent of the choice of resolution.
Motivated by an analogy with the Gauss-Bonnet theorem relating the
curvature of a manifold to its topological invariants, Arveson also defines an
analytic invariant that he calls curvature. Set
T (z) =
ziTi = [T1 . . . Td][z1 . . . zd]∗.
dXi=1
This satisfies kT (z)k ≤ kzk2. Thus we may define a function on the ball by
F (z) = ∆(I − T (z)∗)−1(I − T (z))−1∆∆H ∈ B(∆H)
for z ∈ Bd.
The boundary values, appropriately weighted, exist in the sense that
K0(H) = lim
r→1−
(1 − r2) Tr F (rζ)
exists a.e. for ζ ∈ ∂Bd.
Define the curvature of H to be
K(H) =Z∂Bd
K0(ζ) dσ(ζ)
where σ is normalized Lebesgue measure on the sphere ∂Bd.
Theorem 8.1 (Arveson 2000). Let H be a finite rank Hilbert module. Then
and
K(H) = d! lim
n→∞
Tr(I − ϕn+1(I))
nd
χ(H) = d! lim
n→∞
rank(I − ϕn+1(I))
nd
.
It is immediate that 0 ≤ K(H) ≤ χ(H). Examples show that this in-
equality may be proper. However there is an important context in which
they are equal. Say that H is graded if there is a strongly continuous unitary
representation Γ of the circle T on H so that
Γ(t)TiΓ(t)∗ = tTi
for all 1 ≤ i ≤ d, t ∈ T.
This is called a gauge group on H. When H is graded, Fourier series al-
lows the decomposition of H into a direct sum of subspaces Hn = {ξ ∈
H : Γ(t)ξ = tnξ}. When H is finite rank, these are all finite dimensional.
Arveson calls the following result his analogue of the Gauss-Bonnet theorem:
ARVESON'S LEGACY
21
Theorem 8.2 (Arveson 2000). Let H be a graded finite rank Hilbert module.
Then
K(H) = χ(H).
In particular, the curvature is always an integer in the graded case. Arve-
son hypothesized that it was always an integer. This conjecture was verified
by Greene, Richter and Sundberg [62].
An important tool for multivariable operator theory is the Taylor spec-
trum and functional calculus [106, 107]. Arveson [25] builds a Dirac oper-
ator based on this theory. Let Z = Cd with basis e1, . . . , ed; and let ΛZ be
the exterior algebra over Z, namely
ΛZ = Λ0Z ⊕ Λ1Z ⊕ · · · ⊕ ΛdZ
where ΛkZ is spanned by vectors of the form x1 ∧ · · · ∧ xk. For z ∈ Z, there
is a Clifford operator given by
C(z)(x1 ∧ · · · ∧ xk) = z ∧ x1 ∧ · · · ∧ xk.
Now form H = H ⊗ ΛZ, and for λ ∈ Cd and z ∈ Z, define
Bλ =
(Ti−λiI)⊗C(ei), Dλ = Bλ+B∗
λ
and R(z) = I⊗(C(z)+C(z)∗).
dXi=1
Arveson observes that λ is in the Taylor spectrum of T if and only if Dλ is
invertible. The pair (D0, R) is called the Dirac operator for T .
⊕(H ⊗ ΛiZ) and
eH+ = Xi even
Now the graded space eH splits into its even and odd parts:
Note that D0 maps eH+ into eH− and vice versa. Let D+ = D0 eH+
as an operator in B(eH+, eH−). The relationship to curvature is the following
eH− = Xi odd
Theorem 8.3 (Arveson 2002). If T is a pure graded row contraction, then
both ker D+ and ker D∗
+ are finite dimensional, and
⊕(H ⊗ ΛiZ).
index theorem:
considered
(−1)dK(H) = ind D+ = dim ker D+ − dim ker D∗
+.
Stability of the Fredholm index leads to the corollary that two d-contractions
which are unitarily equivalent modulo the compact operators have the same
curvature.
Finally we discuss [27] in which Arveson considers pure graded finite
rank d-contractions. By the dilation theorem for commuting row contrac-
tions, such a row contraction is obtained from a graded submodule M of
d ⊗ Cn, where n is finite. Graded means that M is spanned by homo-
H 2
geneous polynomials, and thus is generated by a finite set of homogeneous
polynomials (Hilbert's basis theorem).
An easy calculation shows that the multipliers Si on H 2
d satisfy
for all 1 ≤ i, j ≤ d, p > d,
[Si, S∗
j ] = SiS∗
j − S∗
j Si ∈ Sp
22
K. R. DAVIDSON
where Sp is the Schatten p-class of compact operators with s-numbers lying
in ℓp. Examples suggest that this condition on the commutators [Ti, T ∗
j ]
should persist for graded submodules M , i.e. Ti = S(n)
M . Arveson estab-
lishes this for modules generated by monomials of the form zk1
d ⊗ ξ.
Moreover he makes the following conjecture:
1 . . . zkd
i
Conjecture 8.4 (Arveson 2005). If T = [T1, . . . , Td] is a pure graded finite
rank d-contraction, then [Ti, T ∗
j ] ∈ Sp for all 1 ≤ i, j ≤ d and p > d.
Douglas [54] further refines this conjecture first by enlarging the family of
Hilbert modules considered, and more significantly by considering the ideal
I of all polynomials annihilating the d-tuple T . Let Z be the zero set of this
ideal. Then Douglas conjectures that the commutators [Ti, T ∗
j ] should lie in
Sp for all p > dim(Z ∩ Bd).
There are no known counterexamples, so both of these conjectures remain
open. There has been a lot of recent interest. The best result so far, due to
Guo and Wang [63], establishes Arveson's conjecture when d = 2 or 3, and
when M is singly generated.
Acknowledgements.
I thank Matthew Kennedy, Vern Paulsen, Gilles
Pisier and David Pitts for reading the first draft of this paper and mak-
ing many helpful suggestions that improved the final version.
References
[1] J. Agler, An abstract approach to model theory, in Surveys of some recent results in
operator theory, Vol. II, pp. 1 -- 23, Longman Sci. Tech., Harlow, 1988.
[2] J. Agler and J. McCarthy, Complete Nevanlinna-Pick kernels, J. Funct. Anal. 175
(2000), 111 -- 124.
[3] J. Agler and J. McCarthy, Pick interpolation and Hilbert function spaces, Graduate
Studies in Mathematics 44, Amer. Math. Soc., Providence RI, 2002.
[4] N. Andersen, Compact perturbations of reflexive algebras, J. Funct. Anal. 38 (1980),
366 -- 400.
[5] T. Ando, On a pair of commuting contractions, Acta Sci. Math. (Szeged) 24 (1963),
88 -- 90.
[6] A. Arias and G. Popescu, Noncommutative interpolation and Poisson transforms,
Israel J. Math. 115 (2000), 205 -- 234.
[7] W. Arveson, Prediction theory and group representations, Ph.D. Thesis, University
of California, Los Angeles, 1964.
[8] W. Arveson, Operator algebras and measure preserving automorphisms, Acta Math.
118, (1967), 95 -- 109.
[9] W. Arveson, Analyticity in operator algebras, Amer. J. Math. 89 (1967), 578 -- 642.
[10] W. Arveson, A density theorem for operator algebras, Duke Math. J. 34 (1967), 635 --
647.
[11] W. Arveson, Subalgebras of C*-algebras, Acta Math. 123 (1969), 141 -- 224.
[12] W. Arveson, Unitary invariants for compact operators, Bull. Amer. Math. Soc. 76
(1970), 88 -- 91.
[13] W. Arveson, Subalgebras of C*-algebras II, Acta Math. 128 (1972), 271 -- 308.
[14] W. Arveson, Operator algebras and invariant subspaces, Ann. Math. 100 (1974), 433 --
532.
ARVESON'S LEGACY
23
[15] W. Arveson, On groups of automorphisms of operator algebras, J. Funct. Anal. 15
(1974), 217 -- 243.
[16] W. Arveson, Interpolation problems in nest algebras, J. Funct. Anal. 20 (1975), 208 --
233.
[17] W. Arveson, An invitation to C*-algebras, Graduate Texts in Mathematics 39,
Springer-Verlag, New York-Heidelberg, 1976.
[18] W. Arveson, Notes on extensions of C*-algebras, Duke Math. J. 44 (1977), 329 -- 355.
[19] W. Arveson, Perturbation theory for groups and lattices, J. Funct. Anal. 53 (1983),
22 -- 73.
[20] W. Arveson, Ten lectures on operator algebras, CBMS Regional Conference Series in
Mathematics, 55; Published for the Conference Board of the Mathematical Sciences,
Washington, DC by the American Mathematical Society, Providence, RI, 1984.
[21] W. Arveson, Continuous analogues of Fock space, Mem. Amer. Math. Soc. 80 (1989),
no. 409.
[22] W. Arveson, Subalgebras of C*-algebras III, Acta Math. 181 (1998), 159 -- 228.
[23] W. Arveson, The curvature invariant of a Hilbert module over C[z1, . . . , zd], J. Reine
Angew. Math. 522 (2000), 173 -- 236.
[24] W. Arveson, A short course on spectral theory, Graduate Texts in Mathematics 209,
Springer-Verlag, New York, 2002.
[25] W. Arveson, The Dirac operator of a commuting d-tuple, J. Funct. Anal. 189 (2002),
53 -- 79.
[26] W. Arveson, Noncommutative dynamics and E-semigroups, Springer Monographs in
Mathematics, Springer-Verlag, New York, 2003.
[27] W. Arveson, p-summable commutators in dimension d, J. Operator Theory 54 (2005),
101 -- 117.
[28] W. Arveson, The noncommutative Choquet boundary, J. Amer. Math. Soc. 21 (2008),
1065 -- 1084.
[29] W. Arveson and K. Josephson, Operator algebras and measure preserving automor-
phisms II, J. Funct. Anal. 4 (1969), 100 -- 134.
[30] J. Ball, T. Trent and V. Vinnikov, Interpolation and commutant lifting for multipli-
ers on reproducing kernel Hilbert spaces, Operator theory and analysis (Amsterdam,
1997), 89 -- 138, Oper. Theory Adv. Appl. 122, Birkhauser, Basel, 2001.
[31] D. Blecher and L. Labuschagne, Characterizations of noncommutative H ∞, Integral
Equations Operator Theory 56 (2006), 301 -- 321.
[32] D. Blecher and L. Labuschagne, Von Neumann algebraic H p theory,
in Function
spaces, pp. 89 -- 114, Contemp. Math. 435, Amer. Math. Soc., Providence, RI, 2007.
[33] D. Blecher, and C. Le Merdy, Operator algebras and their modules -- an operator space
approach, LMS Monographs, new series 30, Oxford University Press, Oxford, 2004.
[34] D. Blecher, Z.J. Ruan and A. Sinclair, A characterization of operator algebras, J.
Funct. Anal. 89 (1990), 188 -- 201.
[35] L. Brown, R. Douglas and P. Fillmore, Unitary equivalence modulo the compact op-
erators and extensions of C*-algebras, Proc. conference on Operator theory, Halifax,
NS, Lect. Notes Math. 345, Springer Verlag, Berlin, 1973.
[36] L. Brown, R. Douglas and P. Fillmore, Extensions of C*-algebras and K-homology,
Ann. Math. 105 (1977), 265 -- 324.
[37] J. Bunce, The similarity problem for representations of C*-algebras, Proc. Amer.
Math. Soc. 81 (1981), 409 -- 414.
[38] J. Bunce, Models for n-tuples of non-commuting operators, J. Funct. Anal. 57 (1984),
21 -- 30.
[39] L. Carleson, Interpolation by bounded analytic functions and the corona problem, Ann.
Math. (2) 76 (1962), 547 -- 559.
[40] M.D. Choi and E. Effros, Separable nuclear C*-algebras and injectivity., Duke Math.
J. 43 (1976), 309 -- 322.
24
K. R. DAVIDSON
[41] M.D. Choi and E. Effros, The completely positive lifting problem for C*-algebras, Ann.
Math. 104 (1976), 585 -- 609.
[42] M.D. Choi and E. Effros, Nuclear C*-algebras and injectivity: the general case, Indi-
ana Univ. Math. J. 26 (1977), 443 -- 446.
[43] M.D. Choi and E. Effros, Injectivity and operator spaces, J. Funct. Anal. 24 (1977),
156 -- 209.
[44] E. Christensen, On nonselfadjoint representations of C*-algebras, Amer. J. Math.
103 (1981), 817 -- 833.
[45] A. Connes, Une classification des facteurs de type III, Ann. Sci. ´Ecole Norm. Sup.
6 (1973), 133 -- 252.
[46] A. Connes, Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1, Ann.
Math. 104 (1976), 73 -- 115.
[47] S. Costea, E. Sawyer and B. Wick, The corona theorem for the Drury-Arveson Hardy
space and other holomorphic Besov -- Sobolev spaces on the unit ball in Cn, Anal. PDE
4 (2011), 499 -- 550.
[48] K. Davidson, Similarity and compact perturbations of nest algebras, J. Reine Angew.
Math. 348 (1984), 286 -- 294.
[49] K. Davidson, Nest Algebras, Pitman Research Notes in Mathematics Series 191,
Longman Scientific and Technical Pub. Co., London, New York, 1988.
[50] K. Davidson and E. Katsoulis, Isomorphisms between topological conjugacy algebras,
J. Reine Angew. Math. 621 (2008), 29 -- 51.
[51] K. Davidson and D. Pitts, Invariant subspaces and hyper-reflexivity for free semi-
group algebras, Proc. London Math. Soc. 78 (1999), 401 -- 430.
[52] K. Davidson and D. Pitts, Nevanlinna-Pick interpolation for non-commutative ana-
lytic Toeplitz algebras, Integral Eqtns. and Operator Theory 31 (1998), 321 -- 337.
[53] R. Douglas and V. Paulsen, Hilbert modules over function algebras, Pitman Research
Notes in Math. 217, Longman Scientific & Technical, Harlow, 1989.
[54] R. Douglas, A new kind of index theorem, Analysis, geometry and topology of elliptic
operators, pp. 369 -- 382, World Scientific Publications, Hackensack, NJ, 2006.
[55] M. Dritschel and S. McCullough, Boundary representations for families of represen-
tations of operator algebras and spaces, J. Operator Theory 53 (2005), 159 -- 167
[56] S. Drury, A generalization of von Neumann's inequality to the complex ball, Proc.
Amer. Math. Soc. 68 (1978), 300 -- 304.
[57] E. Effros and E. Lance, Tensor products of operator algebras, Adv. Math. 25 (1977),
1 -- 34.
[58] E. Effros and Z.J. Ruan, Operator spaces, LMS Monographs, new series 23, Oxford
University Press, New York, 2000.
[59] A. Frazho, Models for non-commuting operators, J. Funct. Anal. 48 (1982), 1 -- 11.
[60] B. Fuglede and R. Kadison, Determinant theory in finite factors, Ann. Math. 55
(1952), 520 -- 530.
[61] J. Garnett, Bounded analytic functions, revised first edition, Graduate Texts in Math.
236, Springer, New York, 2007.
[62] D. Greene, S. Richter and C. Sundberg, The structure of inner multipliers on spaces
with complete Nevanlinna-Pick kernels, J. Funct. Anal. 194 (2002), 311 -- 331.
[63] K. Guo and K. Wang, Essentially normal Hilbert modules and K-homology, Math.
Annalen 340 (2008), 907 -- 934.
[64] U. Haagerup, Solution of the similarity problem for cyclic representations of C*-
algebras, Ann. Math. 118 (1983), 215 -- 240.
[65] D. Hadwin and T. Hoover, Operator algebras and the conjugacy of transformations.,
J. Funct. Anal. 77 (1988), 112 -- 122.
[66] M. Hamana, Injective envelopes of operator systems, Publ. Res. Inst. Math. Sci. 15
(1979), 773-785.
ARVESON'S LEGACY
25
[67] K. Hoffman, Analytic functions and logmodular Banach algebras, Acta Math. 108
(1962), 271 -- 317.
[68] K. Hoffman and H. Rossi, Function theory and multiplicative linear functionals, Trans.
Amer. Math. Soc. 116 (1965), 536 -- 543.
[69] M. Izumi, E0-semigroups: around and beyond Arveson's work, J. Operator Theory
68 (2012), xxx -- xxx.
[70] R. Kadison, Derivations of operator algebras, Ann. Math. 83 (1966), 280 -- 293.
[71] R. Kadison and I. Singer, Triangular operator algebras, Amer. J. Math. 82 (1960),
227 -- 259.
[72] L. Labuschagne, A noncommutative Szego theorem for subdiagonal subalgebras of von
Neumann algebras, Proc. Amer. Math. Soc. 133 (2005), 3643 -- 3646.
[73] E. Lance, On nuclear C*-algebras, J. Funct. Anal. 12 (1973), 157 -- 176.
[74] D. Larson, Nest algebras and similarity transformations, Ann. Math. 121 (1985),
409 -- 427.
[75] R. Loebl and P. Muhly, Analyticity and flows in von Neumann algebras, J. Funct.
Anal. 29 (1978), 214 -- 252.
[76] S. McCullough, Carath´eodory interpolation kernels, Integral Equations Operator The-
ory 15 (1992), 43 -- 71.
[77] P. Muhly and B. Solel, Hilbert modules over operator algebras, Mem. Amer. Math.
Soc. 117 (1995), no. 559.
[78] P. Muhly and B. Solel, An algebraic characterization of boundary representations,
Nonselfadjoint operator algebras, operator theory, and related topics, 189 -- 196, Oper.
Theory Adv. Appl. 104, Birkhauser, Basel, 1998.
[79] V. Muller and F. Vasilescu, Standard models for some commuting multioperators,
Proc. Amer. Math. Soc. 117 (1993), 979 -- 989.
[80] S. Parrott, Unitary dilations for commuting contractions, Pacific J. Math. 34 (1970),
481 -- 490.
[81] V. Paulsen, Completely bounded homomorphisms of operator algebras, Proc. Amer.
Math. Soc. 92 (1984), 225 -- 228.
[82] V. Paulsen, Every completely polynomially bounded operator is similar to a contrac-
tion, J. Funct. Anal. 55 (1984), 1 -- 17.
[83] V. Paulsen, Completely bounded maps and dilations, Pitman Research Notes in Math-
ematics 217, Longman Scientific and Technical, Harlow, 1986.
[84] V. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in
Advanced Mathematics 78, Cambridge University Press, Cambridge, 2002.
[85] J. Peters, Semicrossed products of C*-algebras, J. Funct. Anal. 59 (1984), 498 -- 534.
[86] G. Pick, Uber die Beschrankungen analytischer Funktionen, welche durch vorgegebene
Funktionswerte bewirkt werden, Math. Ann. 77 (1916), 7 -- 23.
[87] G. Pisier, A polynomially bounded operator on Hilbert space which is not similar to a
contraction, J. Amer. Math. Soc. 10 (1997), 351 -- 369.
[88] G. Pisier, Similarity problems and completely bounded maps, 2nd, expanded edition,
Lecture Notes in Mathematics 1618, Springer-Verlag, Berlin, 2001.
[89] G. Pisier, Introduction to operator space theory, LMS Lecture Note Series 294, Cam-
bridge University Press, Cambridge, 2003.
[90] G. Popescu, Isometric dilations for infinite sequences of noncommuting operators,
Trans. Amer. Math. Soc. 316 (1989), 523 -- 536.
[91] G. Popescu, Multi-analytic operators on Fock spaces, Math. Ann. 303 (1995), 31 -- 46.
[92] P. Quiggin, For which reproducing kernel Hilbert spaces is Pick's theorem true?, In-
tegral Equations Operator Theory 16 (1993), 244 -- 266.
[93] H. Radjavi and P. Rosenthal, On invariant subspaces and reflexive algebras, Amer. J.
Math. 91 (1969), 683 -- 692.
[94] J. Ringrose, On some algebras of operators, Proc London Math. Soc. (3) 15 (1965),
61 -- 83.
26
K. R. DAVIDSON
[95] J. Ringrose, On some algebras of operators II, Proc. London Math. Soc. (3) 16 (1966),
385 -- 402.
[96] Z.J. Ruan, Subspaces of C*-algebras, J. Funct. Anal. 76 (1988), 217 -- 230.
[97] S. Sakai, Derivations of W*-algebras, Ann. Math. 83 (1966), 273 -- 279.
[98] D. Sarason, Generalized interpolation in H ∞, Trans. Amer. Math. Soc. 127 (1967),
179 -- 203.
[99] T. Srinivasan and J.K. Wang, Weak*-Dirichlet algebras, in Function algebras, F.
Birtel, ed., Scott Foresman and Co., 1966, pp. 216 -- 249.
[100] W. Stinespring, Positive functions on C*-algebras, Proc. Amer. Math. Soc. 6 (1955),
211 -- 216.
[101] E. Størmer, Positive linear maps of operator algebras, Acta Math. 110 (1963), 233 --
278.
[102] B. Sz. Nagy, Sur les contractions de l'espace de Hilbert, Acta Sci. Math. (Szeged)
15 (1953), 87 -- 92.
[103] B. Sz. Nagy and C. Foia¸s, Harmonic analysis of operators on Hilbert space, North
Holland Pub. Co., London, 1970.
[104] B. Sz. Nagy, C. Foia¸s, H. Bercovici and L. Kerchy, Harmonic analysis of operators
on Hilbert space, 2nd ed., Springer Verlag, New York, 2010.
[105] B. Sz. Nagy and C. Foia¸s, Dilatation des commutants d'operateurs, Comptes Rendus
Paris 266 (1968), A493 -- 495.
[106] J. Taylor, A joint spectrum for several commuting operators, J. Funct. Anal. 6 (1970),
172 -- 191.
[107] J. Taylor, The analytic functional calculus for several commuting operators, Acta
Math. 125 (1970), 1 -- 38.
[108] N. Varopoulos, On an inequality of von Neumann and an application of the metric
theory of tensor products to operators theory, J. Funct. Anal. 16 (1974), 83 -- 100.
[109] D. Voiculescu, A non-commutative Weyl -- von Neumann Theorem, Rev. Roum. Pures
Appl. 21 (1976), 97 -- 113.
[110] G. Wittstock, Ein operatorwertiger Hahn-Banach Satz, J. Funct. Anal. 40 (1981),
127 -- 150.
KENNETH R. DAVIDSON, Department of Pure Mathematics, University of
Waterloo, Waterloo, ON N2L 3G1, CANADA
E-mail address: [email protected]
|
1704.06990 | 1 | 1704 | 2017-04-23T22:08:33 | Random walks on Bratteli diagrams | [
"math.OA"
] | In the eighties, A. Connes and E. J. Woods made a connection between hyperfinite von Neumann algebras and Poisson boundaries of time dependent random walks. The present paper explains this connection and gives a detailed proof of two theorems quoted there: the construction of a large class of states on a hyperfinite von Neumann algebra (due to A. Connes) and the ergodic decomposition of a Markov measure via harmonic functions (a classical result in probability theory). The crux of the first theorem is a model for conditional expectations on finite dimensional C*-algebras. The proof of the second theorem hinges on the notion of cotransition probability. | math.OA | math |
RANDOM WALKS ON BRATTELI DIAGRAMS
JEAN RENAULT
Abstract. In the eighties, A. Connes and E. J. Woods made a connection between
hyperfinite von Neumann algebras and Poisson boundaries of time dependent random
walks.
I will explain this connection and will present two theorems given there: the
construction of a large class of states on a hyperfinite von Neumann algebra (due to
A. Connes) and the ergodic decomposition of a Markov measure via harmonic functions
(a classical result in probability theory). The crux of the first theorem is a model for
conditional expectations on finite dimensional C*-algebras. Our proof of the second
theorem hinges on the notion of cotransition probability.
1. Introduction.
The connection between operator algebras and ergodic theory goes back to the early
days of these subjects. More recently, A. Connes and E. J. Woods have uncovered in [4]
a new connection between hyperfinite von Neumann algebras and Markov chains. They
have identified the flow of weights of an ITPFI factor M as the Poisson boundary of
a group-invariant time-dependent Markov random walk on the real line R. They also
generalize this identification to arbitrary hyperfinite factors by introducing matrix-valued
random walks.
In the same article, they give a partial answer to a related question
in ergodic theory which goes back to G. Mackey. Given a locally compact group G,
characterize the measured G-spaces X which can be constructed as the Poisson boundary
of a random walk on G. A necessary condition is the amenability in the sense of Zimmer
of the G-space X (Zimmer had already shown the amenability of the Poisson boundary
of a time-independent random walk). Another necessary condition is its approximate
transitivity, a notion introduced earlier by Connes and Woods to characterize the flow of
weights of an ITPFI factor. They state that these conditions are also sufficient and show
that this is the case for transitive actions and when the group is R or Z. More generally,
one can ask what measured G-spaces can be constructed as the Poisson boundary of a
matrix-valued random walk on G. The complete answer is given by Adams, Elliott and
Giordano in [1] (see also [7]): these are exactly the amenable G-spaces.
My purpose here is not to explain these results, but to put into light some points which
are only implicit in [4]. I shall put on the front of the stage random walks on Bratteli
diagrams, which is in fact another name for time-dependent Markov chains (with dis-
crete time), and cotransition probabilities. While cotransition probabilities do not appear
explicitly in the case of UHF diagrams, they become crucial when studying arbitrary Brat-
teli diagrams. Their importance in the study of boundaries of random walks on Bratteli
diagrams is also emphasized in the recent work [18, 19] of A. Vershik. My scope will be
1991 Mathematics Subject Classification. Primary 22A22; Secondary 54H20, 43A65, 46L55.
Key words and phrases. Bratteli diagrams, hyperfinite von Neumann algebras.
1
2
Jean Renault
limited to the presentation of two known theorems which play a key role in [4]. The first
one, given here as Theorem 3.3, is Theorem 1 of [3], which makes a direct connection
between a large class of states on a hyperfinite von Neumann algebra and random walks
on a Bratteli diagram. The second one (Theorem 4.4) is a classical result in probability
theory [14, Proposition V-2-2] which gives the ergodic decomposition of a Markov mea-
sure via bounded harmonic functions. On the way, I will give the reduction of faithful
conditional expectations on a finite dimensional C*-algebra (Theorem 2.8), a definitely
well-known result which can be easily extracted from [3], because it provides the building
block of a random walk on a Bratteli diagram. The definitions of a matrix-valued random
walk on a group and of its Poisson boundary will be given in the last section only, where
they should appear more natural after the presentation of Theorem 3.3 and Theorem 4.4.
2. Conditional expectations on finite dimensional C*-algebras
Although the material of this section is not new, I have not found a reference for
Theorem 2.8 below which gives a complete invariant for a faithful conditional expectation
on a finite dimensional C*-algebra. The description of an inclusion of finite dimensional
C*-algebras in terms of matrix units, which is a part of the theorem, and its description by
a diagram, have been given by O. Bratteli in his fundamental paper [2] (see its proposition
1.7 and its section 1.8). The graphical description of a conditional expectation on a finite
dimensional C*-algebra appears in section 3 (iii) of [4] but without much detail. Our
proof will emphasize Cartan (or diagonal) subalgebras, which appear only implicitly in
the work of Bratteli. The corresponding groupoid models are also known as path models
or tail equivalence relations.
We first give the ingredients and the recipe to construct a faithful conditional expec-
tation Q of a finite dimensional C*-algebra M onto a sub-C*-algebra M. Then, we show
that every faithful conditional expectation is obtained by this recipe. Let us recall that we
can associate to an equivalence relation R on a finite set X a finite dimensional C*-algebra
M = C ∗(R): its elements are the functions (or matrices) f : R → C (here R ⊂ X × X
is the graph of the equivalence relation), the product is the matrix multiplication and
the involution is the usual complex conjugate of a matrix. It has a canonical matrix unit
(e(x, y))(x,y)∈R indexed by R. Consider now finite sets X, V, E, V equipped with surjec-
tions r : X → V , s : E → V , r : E → V . The triple (V, E, V ), which we call here a graph,
will be ubiquitous in this survey. Define
and the equivalence relations:
X = {(x, a) ∈ X × E : r(x) = s(a)}
R = {(x, y) ∈ X × X : r(x) = r(y)}
R = {(xa, yb) ∈ X × X : r(a) = r(b).
We can construct the C*-algebras C ∗(R) and C ∗(R). Moreover, the map j : C ∗(R) →
C ∗(R) given by
j(f )(xa, yb) = (cid:26) f (x, y)
0
if a = b
if a 6= b
Random walks on Bratteli diagrams
3
identifies C ∗(R) to a subalgebra of C ∗(R). We shall make this identification and view the
elements of C ∗(R) as functions on R. Then (V, E, V ) is the graph of the inclusion. We
leave as an exercise to the reader the proof of the following lemma:
Lemma 2.1. Let R′ be the following equivalence relation on E:
R′ = {(a, b) ∈ E × E : s(a) = s(b), r(a) = r(b)}
Then the map k : C ∗(R′) → C ∗(R) given by
k(g)(xa, yb) = (cid:26) g(a, b)
0
if x = y
if x 6= y
identifies C ∗(R′) to the commutant of C ∗(R) in C ∗(R).
Definition 2.1. A transition probability on the graph E is a function p : E → R∗
+ such
that for all v ∈ V , Ps(c)=v p(c) = 1.
Proposition 2.2. Let X, V, E, V be as above and let p be a transition probability on E.
The map Q : C ∗(R) → C ∗(R) defined by
Q(f )(x, y) = Xc
p(c)f (xc, yc),
where the sum is over all edges c ∈ E originating from the common range of x and y, is
a faithful conditional expectation onto C ∗(R).
Proof. This is a straighforward verification.
(cid:3)
We are going to prove a converse to Proposition 2.2: namely all faithful conditional
expectations Q : M → M, where M is a sub-C*-algebra of a finite dimensional C*-algebra
M , are of that form. We first recall the notion of Cartan subalgebra which will be our
main tool. It is an algebraic characterization of the canonical abelian subalgebra C(X)
of the C*-algebra C ∗(R) of an equivalence relation R on X as above.
Definition 2.2. An abelian subalgebra A of a von Neumann algebra M is called a Car-
tan subalgebra if it is maximal self-adjoint, regular and there exists a faithful normal
conditional expectation P : M → A. We then say that (M, A) is a Cartan pair.
Regularity means that the normalizer of A in M, which is defined here as
NM (A) = {v partial isometry of M : vAv∗ ⊂ A, v∗Av ⊂ A},
generates M as a von Neumann algebra. We recall the fact that the conditional expecta-
tion P is unique. The main result of [9] is that every Cartan pair (M, A) (if one assumes
that M acts on a separable Hilbert space) is of the form (W ∗(R, τ ), L∞(X, µ)) where R is
a countable Borel equivalence relation on a standard measured space (X, µ), where µ is a
quasi-invariant measure and τ ∈ Z 2(R, T) is a Borel twist. When M is finite dimensional,
the result of [9] is elementary: we let X be the spectrum of A and V be the spectrum of
the centre Z(M) of M. The inclusion Z(M) ⊂ A gives a surjective map r : X → V . We
let R be the equivalence relation admitting r as quotient map. Each x ∈ X corresponds to
a minimal projection e(x) in A; e(x) and e(y) are equivalent if and only if (x, y) ∈ R. We
choose a matrix unit (e(x, y))(x,y)∈R such that for all x ∈ X, e(x, x) = e(x). This matrix
4
Jean Renault
unit defines an isomorphism M → C ∗(R) sending A to C(X). Thus, when M is finite
dimensional, the twist is trivial. However, it does not admit a canonical trivialization.
Note also that in a finite dimensional C*-algebra, the notions of Cartan subalgebra and
of maximal abelian self-adjoint subalgebra agree. We shall need an easy lemma about
extension of matrix units.
Lemma 2.3. Let (M, A) be a finite dimensional Cartan pair and let (X, R) be the corre-
sponding equivalence relation. Then every partial matrix unit (e(x, y))(x,y)∈S in M, where
S is a subequivalence relation of R, can be extended to a full matrix unit (e(x, y))(x,y)∈R.
Proof. We fix an arbitrary full matrix unit (e(x, y), (x, y) ∈ R). There exists a function
c : S → T, where T is the group of complex numbers of module 1, such that e(x, y) =
c(x, y)e(x, y) for all (x, y) ∈ S.
It is a cocycle. Every cocycle on S is trivial: there
exists b : X → T such that c(x, y) = b(x)b(y) for all (x, y) ∈ S. Then, we define
e(x, y) = b(x)e(x, y)b(y) for all (x, y) ∈ R.
(cid:3)
The following lemma is a complement to Lemma III.1.14 of [15].
Lemma 2.4. Given an inclusion M ⊂ M of finite dimensional C*-algebras, a faithful
conditional expectation Q : M → M and a Cartan subalgebra A of M, there exists a
Cartan subalgebra A of M such that
(i) A ⊂ A,
(ii) NM (A) ⊂ NM (A), and
(iii) Q ◦ P = P ◦ Q0, where P is the conditional expectation from M onto A, P is the
conditional expectation from M to A and Q0 is the restriction of Q to A.
Proof. We let (X, R) be the equivalence relation defined by the pair (M, A): X is the
spectrum of A, V is the spectrum of the centre Z(M) of M and r : X → V is the quotient
map. We choose a matrix unit (e(x, y))(x,y) ∈ R of M with e(x, x) = e(x) minimal
projection corresponding to x. We choose a section σ for the map r : X → V . For each
v ∈ V , we set M v = e(σ(v))M e(σ(v)). There exists a unique state ϕ
of the algebra M v
such that Q(f ) = ϕ
(f )e(σ(v)) for all f ∈ M v. It is faithful because Q is faithful. Since
self-adjoint matrices are diagonalizable, there exists a Cartan subalgebra Av of M v such
◦ P v, where P v is the conditional expectation onto Av. For x ∈ X, we define
that ϕ
= ϕ
v
v
v
v
Ax = e(x, σ(r(x)))Ar(x)e(σ(r(x)), x)
Then A = ⊕x∈X Ax is a Cartan subalgebra of M . It contains A because for all x ∈ X, Ax
contains e(x) as its unit element. By construction e(x, y) belongs to the normalizer of A
in M , hence NM (A) ⊂ NM (A). Let v ∈ V and a ∈ M v. Then
On the other hand, since P v is the restriction of P to M v,
P ◦ Q(a) = P (ϕ
v
(a)e(σ(v))) = ϕ
(a)e(σ(v)).
v
Q ◦ P (a) = ϕ
(P v(a))e(σ(v)) = ϕ
v
v
(a)e(σ(v)).
Thus, P ◦ Q and Q ◦ P agree on M v Suppose now that a belongs to e(x)M e(y), where
(x, y) ∈ R. We write a = e(x, v)ave(v, y) with av ∈ M v. Since e(x, v) and e(v, x) belong
to M, Q(a) = e(x, v)Q(av)e(v, y) and since e(x, v) and e(v, x) belong to NM (A),
P ◦ Q(a) = e(x, v)P ◦ Q(av)e(v, y).
Random walks on Bratteli diagrams
5
On the other hand, since e(x, v) and e(v, x) belong also to NM (A), P (a) = e(x, v)P (av)e(v, y).
Hence
Q ◦ P (a) = e(x, v)Q ◦ P (av)e(v, y).
Therefore, P ◦ Q and Q ◦ P agree on e(x)M e(y). We deduce that they agree on M. This
implies that Q(A) = A and that we have the equality Q ◦ P = P ◦ Q0 where Q0 is the
restriction of Q to A.
(cid:3)
Definition 2.3. Let Q : M → M be a conditional expectation and let A, A be Cartan
subalgebras of M, M respectively. We say that we have a Cartan pairs inclusion if the
conditions (i) and (ii) of the lemma are satisfied; we then write (M, A) ⊂ (M, A). We
say that the inclusion (M, A) ⊂ (M, A) is compatible with Q if the condition (iii) of the
lemma is also satisfied.
Lemma 2.5. Let (M, A) ⊂ (M , A) be an inclusion of finite dimensional Cartan pairs.
Then the spectrum X of A is canonically identified to the fibered product X ×V E, where
X is the spectrum of A, E is the spectrum of M ′ ∩ A, V is the spectrum of Z(M) and the
fibered product is relative to the maps r : X → V and s : E → V given by the inclusions
Z(M) ⊂ M and Z(M) ⊂ M ′ ∩ A.
Proof. We let α be the action of NM (A) on X and α be the action of NM (A) on X. We
let π : X → X and π : X → E be the surjections corresponding to the inclusions A ⊂ A
and M ′ ∩ A ⊂ A. They satisfy r ◦ π = s ◦ q. Hence (π, q) maps X into the fibered product
X ×V E. This map is injective: let x, y ∈ X such that π(x) = π(y) and q(x) = q(y). The
elements of M ′ ∩ A are exactly the functions on X which are constant under the action α
of NM (A) on X. Therefore, the relation q(x) = q(y) implies the existence of u ∈ NM (A)
such that y = αu(x). This implies that π(x) = π(y) = αu(π(x)), hence ue(x) = e(x) and
y = x. The map is surjective. Let (x, c) ∈ X × E such that r(x) = s(c). Pick y ∈ X such
that q(y) = c. Since r(π(y)) = r(x), there exists u ∈ NM (A) such that x = αu(π(y)).
Then x = αu(y) does the job.
(cid:3)
An equivalent statement of the lemma is that A is canonically identified to A ⊗Z(M )
(M ′ ∩ A).
Lemma 2.6. Let (M, A) ⊂ (M , A) be an inclusion of finite dimensional Cartan pairs.
The commutant of M in M is denoted by M ′.
(i) If a belongs to M ′, then e(xa)ae(yb) = 0 if x 6= y.
(ii) M ′ ∩ A is a Cartan subalgebra of M ′.
(iii) the equivalence relation induced on the spectrum E of M ′ ∩ A by the normalizer
is
R′ = {(a, b) ∈ E × E : s(a) = s(b), r(a) = r(b)}
Proof. i) Assume that f commutes with M. If x 6= y,
e(xa)f e(yb) = e(xa)e(x)f e(yb) = e(xa)f e(x)e(yb) = 0.
ii) For c ∈ E, we denote by ǫ(c) the corresponding projection in M ′ ∩ A. According to
(i), ǫ(c) = Pr(x)=s(c) e(xc). Suppose that f ∈ M ′ commutes with the elements of M ′ ∩ A.
6
Jean Renault
Consider xa and xb with a 6= b. Then
e(xa)f e(xb) = e(xa)ǫ(a)f e(yb) = e(xa)f ǫ(a)e(yb) = 0.
Thus e(xa)f e(yb) = 0 if xa 6= yb, therefore f belongs to A.
(iii) Assume that ǫ(a)M ′ǫ(b) 6= 0. Then according to (i), there exists x ∈ X such that
(xa, xb) ∈ R. This implies that (a, b) ∈ R′. Conversely, if (a, b) ∈ R′, we pick x ∈ X such
that r(x) = s(a) = s(b). We choose a partial isometry u ∈ M such that uu∗ = e(xa),
u∗u = e(xb). Then P e(y, x)ue(x, y), where (e(x, y))(x,y)∈R is a matrix unit for R and the
sum is over the y's such that r(y) = r(x) is a partial isometry in M ′ with domain ǫ(b)
and range ǫ(a).
(cid:3)
Lemma 2.7. Let (M, A) ⊂ (M , A) be an inclusion of finite dimensional Cartan pairs
and let Q : M → M be a faithful conditional expectation which satisfies the condition (iii)
of Lemma 2.4. Then, with above notations, there exists a transition probability p on the
graph E such that for all c ∈ E,
Q(ǫ(c)) = p(c)e(s(c))
where ǫ(c) is the minimal projection in M ′ ∩ A corresponding to c ∈ E and e(v) is the
minimal projection in Z(M) corresponding to v ∈ V .
Proof. As earlier, we denote by Q0 the restriction of Q to A. We first check that Q0(ǫ(c))
belongs to Z(M): for a ∈ M,
aQ0(ǫ(c)) = Q0(aǫ(c)) = Q0(ǫ(c)a) = Q0(ǫ(c))a
Then, we observe that e(x)ǫ(c) = 0 if r(x) 6= s(c). Therefore e(v)Q0(ǫ(c)) = 0 for all v ∈ V
distinct from s(c): Q0(ǫ(c)) is proportional to e(s(c)). The constant of proportionality is
non zero because Q is supposed to be faithful. The equality e(v) = Ps(c)=v ǫ(c) gives the
equality Ps(c)=v p(c) = 1.
Theorem 2.8. Let Q : M → M be a faithful conditional expectation on a finite dimen-
sional C*-algebra and let A be Cartan subalgebra of M. We let (X, R) be the associated
equivalence relation. Then,
(cid:3)
(i) there exists A Cartan subalgebra of M such that (M, A) ⊂ (M , A) is a Cartan
pairs inclusion compatible with Q.
(ii) any isomorphism Φ : M → C ∗(R) carrying A onto C0(X) can be extended to
an an isomorphism Φ : M → C ∗(R) carrying Q into the model expectation
Qp : C ∗(R) → C ∗(R) constructed from the graph (V, E, V ) of the inclusion, the
spectrum X of A and the transition probability p of Lemma 2.7.
Proof. The first assertion is Lemma 2.4. We fix a Cartan subalgebra A satisfying (i). We
recall that the spectrum X of A can be identified with the fibered product X ×V E and
that the spectrum of M ′ ∩ A can be identified with E. We also recall from Lemma 2.6
that (M ′, M ′ ∩ A) is a Cartan pair defining the equivalence relation R′ on E. We pick a
matrix unit (ǫ(a, b))(a,b)∈R′ for the Cartan pair (M ′, M ′ ∩ A). Let Φ : M → C ∗(R) be an
isomorphism carrying A onto C0(X). There exists a unique matrix unit (e(x, y))(x,y)∈R for
the Cartan pair (M, A) which is sent by Φ onto the canonical matrix unit of C ∗(R). We
Random walks on Bratteli diagrams
7
define e(xa, yb) = e(x, y)ǫ(a, b) if r(x) = r(y) = s(a) = s(b) and r(a) = r(b). This defines
a partial matrix unit on a subequivalence relation of R. According to Lemma 2.3, it can
be completed into a full matrix unit (e(xa, yb))(xa,yb)∈R. The isomorphism Φ : M → C ∗(R)
defined by this matrix unit extends Φ and satisfies Φ ◦ Q = Qp ◦ Φ.
(cid:3)
In particular, the theorem shows that our path model of a conditional expectation gives
every faithful conditional expectation. We can recover from this theorem the main result
of [5], namely every conditional expectation on a finite dimensional C*-algebra can be
written as a pinching followed by slicing and averaging: one introduces an intermediate
level V1 in the inclusion graph (V, E, V ), whose vertices label the edges (thus V1 = E).
The graph (V, E, V ) is then written as the concatenation of two graphs (V, E1, V1) and
(V1, E2, V ).
In the second
graph, the vertices of V1 emit a single edge. With the ingredient X1 = X and (V1, E2, V ),
our recipe gives the inclusion C ∗(R1) ⊂ C ∗(R) where (xa, yb) ∈ R1 if and only if a = b.
The transition probability p2 ≡ 1 gives the restriction map Q2 : C ∗(R) → C ∗(R1) as its
associated conditional expectation. It is a pinching: in other words, it is of the form
In the first graph, the vertices of V1 receive a single edge.
Q2(f ) = Xc∈E
ǫ(c)f ǫ(c).
The conditional expectation Q1 : C ∗(R1) → C ∗(R) is an averaging: for every v ∈ V
Q1(f )(x, y) = Xs(c)=v
p(c)f (xc, yc)
for
r(x) = r(y) = v.
In [3], A. Connes uses a similar decomposition of an inclusion of type I von Neumann
algebras to construct inclusions of Cartan pairs.
3. Random walks on discrete Bratteli diagrams.
We first recall the classical definition of a Bratteli diagram.
V = `∞
Definition 3.1. A Bratteli diagram is a directed graph (V, E) where the set of vertices
n=1 E(n) are graded. For each n ≥ 1,
s(E(n)) = V (n − 1) and r(E(n)) = V (n), where s(e) and r(e) are respectively the source
and the range of the edge e.
n=0 V (n) and the set of edges E = `∞
We assume that each level of vertices V (n) is at most countable; we also assumes that
each vertex emits finitely many but at least one edge and that each vertex of a level n ≥ 1
receives finitely many but at least one edge.
Definition 3.2. Let (V, E) be a Bratteli diagram.
• A transition probability is a map p assigning to each vertex v ∈ V a probability
measure p(v) on the set of edges Ev = s−1(v) emanating from v. We shall view p
as a map p : E → R such that for all v ∈ V , Ps(e)=v p(e) = 1. We shall denote
by pn its restriction to E(n).
• An initial probability measure is a probability measure ν0 on the set of initial
vertices V (0).
• A random walk is a pair (p, ν0), where p is a transition probability and ν0 is an
initial probability measure.
8
Jean Renault
We shall assume in this section that p and ν0 have full support, in the sense that
p(e) > 0 for all e ∈ E and µ0(v) > 0 for all v ∈ V (0).
A Bratteli diagram (V, E) defines an ´etale equivalence relation (X, R) called the tail
equivalence relation of the diagram: X is the set of infinite paths x = e1e2 . . . where
en ∈ E(n) and r(en) = s(en+1). It is a locally compact Hausdorff totally disconnected
space admitting the cylinders
Z(a) = {aen+1en+2 . . .}
where a = a1a2 . . . an is a finite path (we assume implicitly that ai ∈ E(i)), as a base of
compact open subsets. Two infinite paths x = e1e2 . . . and y = f1f2 . . . are tail equivalent
if there exists n such that ei = fi for i > n. Its graph R is a locally compact Hausdorff
totally disconnected space admitting the cylinders
Z(a, b) = {(aen+1en+2 . . . , ben+1en+2 . . .)}
where (a, b) is a pair of equivalent finite paths: this means that they have same length n
and same range r(a) = r(b), where we define r(a1a2 . . . an) = r(an) ∈ V (n).
A random walk on a Bratteli diagram defines a measure on the path space X; it is a
particular case of the well-known construction of Markov measures.
Proposition 3.1. Given a random walk (p, ν0) on a Bratteli diagram (V, E), there is
a unique probability measure µ on X , called the Markov measure of the random walk
whose values on cylinder sets is given by µ(Z(a)) = ν0(s(a))p(a) where, for the finite
path a = a1a2 . . . an, p(a) = p1(a1)p2(a2) . . . pn(an)
and s(a) = s(a1).
As observed in [16, Section 3.2], Markov measures are quasi-invariant under the tail
equivalence relation. Let us recall that a measure µ on X is quasi-invariant under the
equivalence relation R if the measures r∗µ and s∗µ on R are equivalent, where R f d(r∗µ) =
R Py f (x, y)dµ(x) for f ∈ Cc(R) and s∗µ is similarly defined. Then its Radon-Nikodym
derivative Dµ = d(r∗µ)/d(s∗µ) is a cocycle, i.e. it satisfies Dµ(x, y)Dµ(y, z) = Dµ(x, z) for
a.e. (x, y, z) ∈ R(2). Here is a way to construct cocycles on the tail equivalence relation
R of a Bratteli diagram (V, E).
Definition 3.3. Let G be a group. A map D : R → G is called a quasi-product cocycle
if there exists a map q : E → G, called a potential, such that for all pairs of equivalent
finite paths (a, b) and all (az, bz) ∈ Z(a, b), D(az, bz) = q(a)q(b)−1 and where, as before,
q(a1a2 . . . an) = q(a1)q(a2) . . . q(an).
Since a quasi-product cocycle is locally constant, it takes at most countably many values
and it is continuous. The following result, which is a simple observation, is essential here.
Proposition 3.2. [16, Proposition 3.3] Let (p, ν0) be a random walk on a Bratteli diagram
(V, E).
(i) The associated Markov measure µ is quasi-invariant under the tail equivalence
relation R
(ii) Its Radon-Nikodym derivative Dµ is the quasi-product cocycle given by the poten-
tial q = (qn) defined by the relation
νn−1(s(e))pn(e) = qn(e)νn(r(e))
for
e ∈ E(n).
Random walks on Bratteli diagrams
9
where νn is the distribution of the random walk on V (n), defined inductively by
νn(w) = Pr(e)=w pn(e)νn−1(s(e)) for w ∈ V (n).
Note that the Radon-Nikodym derivative depends only on the potential q. This po-
tential q has a simple probabilitstic interpretation: it is the cotransition probability of the
random walk: let e be an edge in E(n) with range r(e) = w, then q(e) is the probability
that the random walk passes through e given that it is at w at time n. In his recent papers
[18, 19], A. Vershik also emphasizes the importance of cotransition probabilities in the
asymptotic study of random walks on Bratteli diagrams. I thank him for the reference
[6] on the subject. Cotransition probabilities are called "backward transition probabili-
ties" in [12]. Note that the cotransition probability q depends not only on the transition
probability p but also on the initial measure ν0. As shown by the next example, different
random walks may share the same cotransition probability.
Example 3.1. Random walks on the Pascal triangle. It is the time development of the
simple random walk on Z. Here, the Bratteli diagram is (V, E) where
V (n) = {(n, k) : k = 0, 1, . . . , n}; E(n) = {(n − 1, k, ǫ) : k = 0, 1, . . . , n − 1; ǫ ∈ {0, 1}}.
We consider the random walk defined by the transition probability
pn(n − 1, k) = (1 − t)δ(n−1,k,0) + tδ(n−1,k,1)
where 0 < t < 1. Since V (0) has a single vertex, the initial measure ν0 is the point mass
n=1{0, 1}.
n=1((1 − t)δ0 + tδ1). An
at this vertex. The infinite path can be written as the infinite product X = Q∞
Then, the Markov measure is the product measure µt = Q∞
elementary computation gives the cotransition probability
qn(n, k) = (1 −
k
n
) δ(n−1,k,0) +
k
n
δ(n−1,k−1,1)
It does not depend on t. For a finite path ǫ1ǫ2 . . . ǫn ending at (n, k = ǫ1 + . . . + ǫn), one
has
q(ǫ1ǫ2 . . . ǫn) = (cid:18)n
k(cid:19)−1
.
One deduces that the Radon-Nikodym of µt is D ≡ 1.
In other words, the measures
µt are invariant under the tail equivalence relation on (V, E). It is a well-known result.
It is also well-known (see for example [17, Example 4.2]) that these are the extremal
invariant probability measures (one has to add µ0 and µ1 which we have excluded from
our discussion).
We use now the construction given by Feldman and Moore in [9]: since (X, R, µ) is a
countable standard measured equivalence relation, one can construct its von Neumann
algebra M = W ∗(X, R, µ) and its state ϕ = µ ◦ P , where P is the expectation of M onto
A = L∞(X, µ), which is normal and faithful. By construction, M acts on the Hilbert
space L2(R, s∗µ). This representation is standard. It is known that the modular operator
∆ of ϕ is the operator of multiplication by Dµ and that the modular automorphism σϕ
is
implemented by the operator of multiplication by Dit
µ . A. Connes has given the following
characterization of the pairs (M, ϕ) arising from this construction.
t
10
Jean Renault
Theorem 3.3. [3, Theorem 1] Let ϕ be a faithful normal state on a von Neumann algebra
M. Then the following conditions are equivalent:
(i) there exists an increasing sequence (Mn) of finite dimensional subalgebras stable
under the automorphism group σ of ϕ whose union is weakly dense in M;
(ii) there exists a Bratteli diagram (V, E) and a random walk (p, ν0) on it such that the
pair (M, ϕ) is isomorphic to (W ∗(X, R, µ), µ ◦ P ), where R is the tail equivalence
relation on the infinite path space X of the diagram and µ is the Markov measure
of the random walk.
The original theorem of Connes is in terms of Krieger's factors. It is an intermediate
step to show that all hyperfinite type III0 factors are Krieger's factors. We consider here
von Neumann algebras arising from hyperfinite measured equivalence relations rather than
Krieger's factors. This makes the statement easier to prove.
Proof. (ii)⇒(i). We assume that M = W ∗(X, R, µ) as above. We let Mn be the sub-
algebra of M generated by the characteristic functions 1Z(a,b), where (a, b) is a pair of
joining paths of length n. Since the Radon-Nikodym derivative Dµ is constant on the
cylinder sets Z(a, b), An is stable under the automorphism group of ϕµ. Since Z(a, b) is
the disjoint union of Z(ae, be)'s where e ∈ E(n + 1) and s(e) = r(a) = r(b), we have
the inclusion Mn ⊂ Mn+1. The elements of the union M∞ of the Mn's are the locally
constant functions with compact support. Since M∞ is dense in Cc(R) with respect to the
inductive limit topology, it is dense in the weak topology. Since Cc(R) is weakly dense in
M, so is M∞.
(i)⇒(ii). Let (Mn)n∈N be as in (i). Without loss of generality, we may assume that
M0 = C1. Since for all n, Mn is stable under σ, the modular automorphism of the
restriction ϕn to Mn is the restriction σn of σ to Mn. Since for all n ≥ 1, Mn−1 is
invariant under σn, there exists a faithful expectation Qn−1,n : Mn → Mn−1 such that
ϕn = ϕn−1 ◦ Qn−1,n. We use inductively Theorem 2.8, to construct an increasing sequence
(An) of abelian subalgebras such that for all n ≥ 1, (Mn−1, An−1) ⊂ (Mn, An) is a Cartan
pair inclusion compatible with the conditional expectation Qn−1,n. The construction is
initialized by the only possible choice A0 = M0. Thus we obtain for each n ∈ N the
spectrum Xn of An and for each n ≥ 1 the graph (V (n − 1), E(n), V (n)) of the inclusion
Mn−1 ⊂ Mn and the transition probability pn : E(n) → R∗
+. From the same theorem, we
obtain for all n an isomorphism Φn : Mn → C ∗(Rn) sending An to C(Xn), where (Xn, Rn)
is the equivalence relation defined by (Mn, An), such that Φn extends Φn−1 and carrying
the conditional expectation Qn−1,n : Mn → Mn−1 into the model conditional expectation
Fpn : C ∗(Rn) → C ∗(Rn−1). Again, the construction is initialized by the only possible
isomorphism Φ0 : M0 → C(X0). Since the conditional expectations Pn : Mn → An satisfy
Qn−1,n ◦ Pn = Pn−1 ◦ Qn−1,n, we have ϕn = νn ◦ Pn, where νn is the restriction of ϕn to
An. The Bratteli diagram of the increasing sequence (Mn) of finite dimensional algebras
+ define
a transition probability on E. The initial measure ν0 is the point mass at the unique
point of X0. This defines the random walk. We let X be its infinite path space, R be the
tail equivalence and µ be the Markov measure of the random walk. The isomorphisms
n=1 E(n)). The transition probabilities pn : E(n) → R∗
is (V = `∞
n=0 V (n), E = `∞
Random walks on Bratteli diagrams
11
Φn : Mn → C ∗(Rn) extend to an isomorphism Φ∞ : M∞ → C00(R), where M∞ is the
union of the Mn's and C00(R) is the ∗-algebra of locally constant functions with compact
support. This isomorphism carries the restriction of the state ϕ to the restriction of
the state µ ◦ P , where P is the expectation of W ∗(R) onto L∞(X, µ). Since both von
Neumann algebras M and W ∗(X, R, µ) can be obtained from the GNS representation of
these states, Φ∞ extends to a normal ∗-isomorphism Φ : M → W ∗(X, R, µ) which sends
the weak closure A of the union of the An's to L∞(X, µ) and ϕ to µ ◦ P .
(cid:3)
Remark 3.1. The von Neumann algebra M in the theorem is a factor if and only if the
Markov measure µ is ergodic under the tail equivalence relation.
Its flow of weights
can be computed from the Radon-Nikodym derivative Dµ, or more concretely, from the
cotransition probability q. For example, the above transition probability pt on the Pascal
triangle gives the hyperfinite II1 factor. We shall return to the probabilistic identification
of the flow of weights in the last section.
Remark 3.2. States on AF-algebras constructed from a random walk on a Bratteli diagram
are called quasi-product states in [8]. Do we have a characterization (besides condition
(i) of the theorem) of the normal faithful states on a hyperfinite von Neumann algebra
which can be described as quasi-product states? Necessarily, theses states are almost
periodic (their modular operators are diagonalizable) and their centralizers contain a
Cartan subalgebra. In part II of [3], A. Connes shows that any faithful semifinite normal
weight on a hyperfinite factor of type III0 whose modular operator ∆ is diagonalizable
and such that 1 is isolated in its spectrum and its point spectrum contained in Q satisfies
condition (i) of the theorem (with Mn type I∞ rather than finite-dimensional).
4. Markov chains and Bratteli diagrams
In order to recover the general theory of time-dependent Markov chains (with discrete
time), it is necessary to generalize the notion of Bratteli diagram. Indeed, the original
definition is limited to Markov chains with at most countably many states. Generalized
Bratteli diagrams have been considered before, mostly in the topological setting, and are
part of the theory of topological graphs. Since we are considering objects of measure-
theoretical nature, we choose the Borel setting. We assume implicitly that the Borel
spaces are analytic.
Definition 4.1. We say that a directed graph (V, E) is a Borel graph if the sets of edges
E and the set of vertices V are endowed with a Borel structure and the source and range
maps are Borel. A Borel Bratteli diagram is a Bratteli diagram which is a Borel graph.
Before extending Definition 3.2 to Borel Bratteli diagrams , we need to make precise
our assumptions.
Definition 4.2. Let E, V be Borel spaces and let s : E → V be a Borel surjection. A
Borel s-system of probability measures p is a map assigning to each v ∈ V a probability
measure pv on s−1(v) and such that for all bounded Borel function f on E, the map
v → R f dpv is Borel.
12
Jean Renault
Given a Borel s-system p and a probability measure ν on V , we can form the probability
on E. The measure ν is the image s∗(νp) of νp and µ = νp is the disintegration of µ
along s. This construction requires in fact weaker assumptions on p: it suffices to have
measure µ = νp on E, defined byR f d(νp) = R (R f dpv)dν(v) for f bounded Borel function
pv defined for a.e. v and the ν-measurability of the function v → R f dpv for all bounded
Borel function f on E. However, the disintegration theorem of measures, as stated for
example in [11, Theorem 2.1], says that, conversely, given a probability measure µ on
E, there exists a Borel s-system of probability measures p, where ν = s∗(µ), such that
µ = νp, where ν = s∗(µ). It is unique in the sense that if νp = νp′, then pv = p′
v for
ν-a.e.v. With an abuse of language, we shall say that (ν, p) is the disintegration of µ along
s.
Notation. Consider the n-th floor V (n − 1) s←− E(n) r−→ V (n) of a Borel Bratteli diagram.
A probability measure µn on E(n) admits a disintegration µn = νn−1pn along s and a
disintegration µn = νnqn along r, where νn−1 = s∗µn and νn = r∗µn. This establishes
a bijection between pairs (νn−1, pn), where νn−1 is a probability measure on V (n − 1)
and pn is a Borel system of probability measures along s and pairs (νn, qn),where νn is a
probability measure on V (n) and qn is a Borel system of probability measures along r ,
given by the equation νn−1pn = νnqn.
Definition 4.3. Let (V, E) be a Borel Bratteli diagram.
• A transition probability p is a Borel system of probability measures for the source
map s : E → V . It will be usually viewed as a sequence p = (pn) of Borel systems
of probability measures for s : E(n) → V (n − 1).
• A cotransition probability q is a Borel system of probability measures for the range
map r : E → V . It will be usually viewed as a sequence q = (qn) of Borel systems
of probability measures for r : E(n) → V (n).
• A random walk on (V, E) is a sequence of probability measures µn on E(n) which
are compatible in the sense that for all n ≥ 1, r∗µn = s∗µn+1.
• The measures νn = r∗µn are called the one-dimensional distributions of the ran-
dom walk.
• The measure ν0 = s∗µ1 is called the initial distribution of the random walk.
Let (µn) be a random walk on the Bratteli diagram (V, E). The disintegration of µn
along s and r gives respectively a pair (νn−1, pn) and a pair (νn, qn) as above. Then
p = (pn) [resp. q = (qn)] is called the transition [cotransition] probability of the random
walk. The measures νn are called the one-dimensional distributions of the random walk.
They satisfy the relations νn−1 = s∗(νnqn) and νn = r∗(νn−1pn) for all n ≥. We say that
the sequence (νn) is q-compatible and p-compatible respectively.
Proposition 4.1. Let (V, E) be a Borel Bratteli diagram (V, E).
(i) Given a transition probability p and a probability measure ν0 on V (0), there exists
a unique random walk on (V, E) admitting p as its transition probability and ν0
as its initial distribution.
(ii) Given a cotransition probability q and a sequence of probability measures νn on
V (n) such that νn−1 = s∗(νnqn) for all n ≥ 1, there exists a unique random walk on
Random walks on Bratteli diagrams
13
(V, E) admitting q as its cotransition probability and (νn) as its one-dimensional
distribution.
Proof. This is clear. In the first case, we define inductively µn = νn−1pn. In the second
case, we define µn = νnqn.
(cid:3)
Corollary 4.2. Given a Borel Bratteli diagram (V, E), there is a one-to-one correspon-
dence between
(i) pairs (p, ν0), where p is a transition probability and ν0 is a probability measure on
V (0);
(ii) pairs (q, (νn)), where q is a cotransition probability and νn is a q-compatible se-
quence of probability measures on V (n)
given by the relation νn−1pn = νnqn.
From now on, a random walk on (V, E) will designate indifferently the measures µn on
E(n) as in Definition 4.3, the pair (p, ν0) or the pair (q, (νn)) as in the above corollary.
We recall the construction of the Markov measure of a random walk (see [14, V-1]). As
earlier, we introduce the infinite path space X. We let X n = E(1) ∗ . . . ∗ E(n) denote
the space of paths e1 . . . en of length n endowed with the product Borel structure. Then
X = lim←−X n is the projective limit with respect to the canonical projection X n ← X n+1.
Given a random walk (p, ν0), one first construct by induction a probability measure µn
on X n such that
Z f dµ1 = Z f (e1)dpv(e1)dν0(v),
Z f dµn = Z f (e1 . . . en)dpr(en−1)(en)dµn−1(e1 . . . en−1)
The sequence of measures (µn) is consistent. Therefore, there exists a unique probability
measure µ on X whose image in X n is µn. Note that the one-dimensional distribution νn
on V (n) is the image of µn by the range map r : X n → V (n). It is also the image of µ by
the map rn : X → V (n) such that rn(e1e2 . . .) = r(en).
It remains to characterize the Markov measure µ on X in terms of the cotransition
probability q. We have seen that in the framework of the previous section, the Markov
measure µ is quasi-invariant under the tail equivalence relation and its Radon-Nikodym
derivative D is the quasi-product cocycle defined by q. We then say that µ is a D-measure.
The notion of quasi-product cocycle does not admit a straightforward generalization in
the general framework. However, there exists (see [17, Proposition 3.7]) an equivalent
definition of a D-measure (known in statistical mechanics as the Dobrushin-Lanford-
Ruelle condition for Gibbs states) which can be easily extended. We let Xn be the space
of infinite paths en+1en+2 . . . starting at level n. The sequence of quotient maps
X π1−→ X1
π2−→ X2
π3−→ . . . πn−→ Xn
πn+1−−−→ . . .
defines the tail equivalence relation R on X: two infinite paths x and y are tail equivalent
if and only if there exist n such that πn ◦ . . . π2 ◦ π1(x) = πn ◦ . . . π2 ◦ π1(y). A cotransition
probability q defines an inductive system of expectations
B(X)
q1−→ B(X1)
q2−→ . . .
qn−→ B(Xn)
qn+1−−→
14
Jean Renault
where B(Y ) is the space of bounded complex-valued Borel functions on Y and
qn(f )(en+1en+2 . . .) = Z f (enen+1en+2 . . .)dqs(en+1)
n
(en).
Definition 4.4. Let q be a cotransition probability on the Borel Bratteli diagram (V, E).
A q-measure is a measure on the infinite path space X which factors through all expec-
tations qn . . . q2 q1.
Then we have the easy generalisation of Proposition 3.2:
Theorem 4.3. Let µ be a probability measure on the infinite path space of a Borel Bratteli
diagram (V, E). Then the following conditions are equivalent:
(i) µ is a q-measure;
(ii) µ is the Markov measure of a random walk admitting q as its cotransition proba-
bility.
m qm . . . q2 q1 where µn [resp. µn
Proof. Let µ be a Markov measure with transition probability p, cotransition probability
q and one-dimensional distribution (νn). Let us show that µ factors through qm . . . q2 q1
for all m. The measure µm on Xm, image of µ by πm ◦ . . . π2 ◦ π1, is the Markov measure
defined by the initial measure νm and the transition probability (pn), n > m. Since
measures on X are uniquely determined by their values on cylinder sets, it suffices to
show that for n > m, µn = µn
m] is the measure of the
random walk on X n = E(1) ∗ . . . ∗ E(n) [resp. X n
m = E(m + 1) ∗ . . . ∗ E(n)]. But this is
clear from the construction of the Markov measure.
Let µ be a q-measure. We define for all n the measure νn as the image of µ by the map rn :
X → V (n) such that rn(e1e2 . . .) = r(en). Because of the relation rn = sn◦πn◦. . . π2◦π1, it
is also the image of µn by the map sn : Xn → V (n) such that sn(en+1en+2 . . .) = s(en+1).
Since µn−1 = µn ◦ qn, νn−1 = s∗(νnqn): the sequence (νn) is q-compatible. Therefore it is
the one-dimensional distribution of a Markov chain with cotransition probability q. The
disintegration µ = µn(qn ◦ . . . ◦ q1) gives the disintegration (πn)∗µ = νn(qn ◦ . . . ◦ q1) where
πn : X → X n is the projection πn(e1ee . . .) = e1 . . . en. Therefore (πn)∗µ agrees with the
measure µn of the random walk. This suffices to conclude that µ is the Markov measure
of the random walk.
(cid:3)
Definition 4.5. Given a random walk with transition probability p and one-dimensional
distributions (νn) on a Bratteli diagram (V, E), a bounded harmonic sequence is a sequence
(hn) where hn belongs to L∞(V (n), νn), hn−1 = pn(hn ◦ r) and supn khnk∞ < ∞.
Notation. The bounded harmonic sequences, equipped with the norm supn khnk∞, form
a Banach space H(p, ν0) which is the projective limit of the sequence
(M)
L∞(V (0), ν0)
p1←− L∞(V (1), ν1)
p2←− . . .
pn←− L∞(V (n), νn)
pn+1←−− . . .
where the maps are the expectations defined by the transition probability p. More pre-
cisely, with an abuse of notation, we define pn(h) = pn(h ◦ r) for h ∈ L∞(V (n), νn). We
also note that H(p, ν0) depends only on the measure class [ν0] of ν0.
We can now give the ergodic decompostion of a Markov measure µ under the tail
equivalence R on the infinite path space X of a Bratteli diagram. Recall that R is defined
Random walks on Bratteli diagrams
15
by the maps πn ◦ . . . π2 ◦ π1 : X → Xn. We say that a function f on X is invariant if for
all n, there exists fn on Xn such that f = fn ◦ πn ◦ . . . π2 ◦ π1. We denote by L∞(X, µ)R
the subalgebra of invariant elements of L∞(X, µ).
Theorem 4.4. [14, Proposition V-2-2] Let µ be the Markov measure of a random walk on
a Borel Bratteli diagram (V, E) having a transition probability p and an initial measure
ν0. Then the ordered Banach spaces L∞(X, µ)R and H(p, ν0) are naturally isomorphic.
Proof. Let q be the cotransition probability of the random walk and (νn) be its one-
dimensional distribution. A positive element f of L∞(X, µ)R defines a finite measure µ′ =
f µ such that µ′ ≤ Mµ where M = kf k∞. Since f is invariant, µ′ is a q-measure. Therefore
n = (rn)∗µ′, where the map rn : X → V (n) is the same as in
the sequence of measures ν′
the proof of Theorem 4.3, is q-compatible. Moreover, we have the inequality ν′
n ≤ Mνn.
Hence there exists hn ∈ L∞(V (n), νn) such that ν′
n = hnνn. We also have khnk∞ ≤ M.
Since νn−1pn = νnqn, the condition ν′
n−1 = s∗(ν′
nqn) gives hn−1 = pn(hn ◦ r). Thus
h = (hn) is a positive bounded harmonic sequence of norm at least kf k∞. Conversely,
let h = (hn) be a positive bounded harmonic sequence. We define ν′
n = hnνn. From the
n−1 = s∗(ν′
relations hn−1 = pn(hn ◦ r) and νn−1 = s∗(νnqn), we deduce that ν′
nqn) and that
ν′
n−1(1). We set M = supn khnk∞. Since the sequence (ν′
n(1) = ν′
n) is q-compatible, there
exists a random walk (we no longer have probability measures but finite measures of the
same mass) and a Markov measure µ′ admitting q as cotransition probability and (ν′
n) as
n ≤ Mνn gives µ′n ≤ Mµn where µn and
one-dimensional distribution. The condition ν′
µ′n are the corresponding measures on X n. Therefore µ′ ≤ Mµ. There exists a unique
positive element f ∈ L∞(X, µ) such that µ′ = f µ and it satisfies kf k∞ ≤ M. The first
part shows that kf k∞ = M. Thus this correspondence gives an isomorphism between the
positive cones of L∞(X, µ)R and H(p, ν0) which extends to an isomorphism of the ordered
Banach spaces.
(cid:3)
Remark 4.1. The proof given here relies only on the Radon-Nikodym theorem and on the
disintegration theorem of probability measures. The classical proof of [14, Proposition
V-2-2] uses the martingale convergence theorem.
It gives explicit formulas relating an
element f of L∞(X, µ)R and a bounded harmonic sequence (hn) in H(p, ν0):
(1)
(2)
f (e1e2 . . .) = lim
n
hn(r(en))
a.e.
hn = Pn(fn)
where fn ∈ B(Xn) comes from the factorization f = fn ◦ πn ◦ . . . π2 ◦ π1 and µn = νnPn
is the disintegration of µn along the source map sn : Xn → V (n).
Definition 4.6. The point realization of the commutative von Neumann algebra L∞(X, µ)R
is called the tail boundary of the random walk. It is a standard Borel space P equipped
with a probability measure m. By definition, L∞(X, µ)R = L∞(P, m) and m is the re-
striction of the measure µ.
Remark 4.2. This terminology is not a standard one.
In [4], it is called the Poisson
boundary of the time dependent random walk. There is a dual construction of the tail
16
Jean Renault
boundary, given in [10], based on the inductive sequence of Banach spaces
(L)
L1(V (0), ν0)
q1−→ L1(V (1), ν1)
q2−→ . . .
qn−→ L1(V (n), νn)
qn+1−−→ . . .
where qn(f ) = qn(f ◦ s) for f ∈ L1(V (n − 1), νn−1). Its inductive limit can be written
L1(P, m) because it is an L-space. The adjoint of qn : L1(V (n − 1), νn−1) → L1(V (n), νn)
is the map pn : L∞(V (n), νn) → L∞(V (n − 1), νn−1) defined earlier. Therefore, the
projective limit H(p, ν0) of the sequence (M) is the dual of L1(P, m).
5. Matrix-valued random walks on groups
Definition 5.1. A matrix-valued random walk on a Borel group G is given by the following
data:
(i) a random walk (p, ν0) on a Borel Bratteli diagram (V, E),
(ii) a Borel map ρ : E → G.
Here is the construction of the Poisson boundary of a matrix-valued random walk. One
first construct a new Bratteli diagram, called the skew-product of (V, E, ρ). This is a
particular case of a construction given for graphs or higher rank graphs in [13].
Definition 5.2. Let Γ = (V, E) be a Borel Bratteli diagram and let ρ be a Borel map
from E to a Borel group G. The skew-product Γ(ρ) is the Bratteli diagram (V × G, E × G)
where s(e, g) = (s(e), g) and r(e, g) = (r(e), gρ(e)).
The skew-product Γ(ρ) carries a compatible Borel structure and an automorphic action
of G, given by h(v, g) = (v, hg) and h(e, g) = (e, hg). The infinite path space of Γ(ρ) can be
identified with X ×G, where X is the infinite path space of Γ: we associate to (e1e2 . . . , g)
the path (e1, g)(e2, gρ(e1)), . . .. The map ρ : E → G defines a G-valued quasi-product
cocycle c on the tail equivalence relation R on X of (V, E) according to Definition 3.3. The
tail equivalence relation of the skew-product Bratteli diagram Γ(ρ) is the skew-product
equivalence relation R(c) on X × G, as defined in [15, Definition I.1.6]. Explicitly
(e1e2 . . . , g) ∼ (f1f2 . . . , h) ⇔ ∃N : for n ≥ N, en = fn
gρ(e1) . . . ρ(en) = hρ(f1) . . . ρ(fn)
Let µ be the Markov measure defined by the random walk (p, ν0) on (V, E). Then the
Markov measure defined by the random walk (p, ν0 × λ), where p(v,g) = pv × δg and λ is
a finite measure equivalent to the Haar measure of G, is µ × λ.
Definition 5.3. Let (V, E, ρ : E → G, p, ν0) be a matrix-valued random walk on a locally
compact group G. Its Poisson boundary is the point realization (P, m) of
It is a measured G-space.
L∞(X × G, µ × λ)R(c) ≃ H(p, ν0 × λ).
Example 5.1. Time-dependent random walks on a group. This is the case when (V, E)
is a UHF diagram, i.e. there is only one vertex at each level. Let us assume that G is a
discrete group and that (pn) is a sequence of probability measures on G with finite support
Gn. The UHF diagram (V, E) is defined by E(n) = Gn. The map ρn : E(n) → G is the
inclusion map. The infinite path space of (V, E) is the product space X = Q Gn and its
Markov measure is the product measure µ = Q pn. The one-dimensional distributions
of the skew-product Bratteli diagram are the measures νn = λ ∗ p1 ∗ p2 . . . ∗ pn on G.
Random walks on Bratteli diagrams
17
For a concrete example, let G = Z with probability measures pn = (1 − t)δ0 + tδ1 for
all n. Then (V, E) is the UHF(2∞) diagram. Choosing δ0 as initial measure rather than
a finite measure equivalent to the counting measure on Z, one gets the random walk of
Example 3.1. This is the kernel diagram rather than the skew product diagram described
above. It can be checked directly that the bounded harmonic sequences are constant; in
other words, the Poisson boundary of the time dependent random walk is trivial.
Example 5.2. Flow of weights of hyperfinite von Neumann algebras. Let ϕ be a faithful
normal state on a von Neumann algebra M satisfying the equivalent conditions of Theo-
rem 3.3. Thus, there exists a Bratteli diagram (V, E) and a random walk (p, ν0) on it such
that the pair (M, ϕ) is isomorphic to (W ∗(X, R, µ), µ ◦ P ), where R is the tail equivalence
relation on the infinite path space X of the diagram and µ is the Markov measure of the
random walk. Since the Radon-Nikodym Dµ is the quasi-product cocycle defined by the
cotransition probability q : E → R∗
+ of the random walk, the flow of weights of M is the
Poisson boundary of the matrix-valued random walk on R∗
+ defined by q.
Acknowledgements. I thank T. Giordano and A. Vershik for fruitful discussions.
References
[1] S. Adams, G. Elliott and T. Giordano: Amenable actions of groups, Trans. Amer. Math. Soc., 344
(1994), no. 2, 803–322.
[2] O. Bratteli: Inductive limits of finite dimensional C*-algebras, Trans. Amer. Math. Soc., 171 (1971),
195–234.
[3] A.-Connes: On hyperfinite factors of type III0 and Krieger's factors, J. Funct. Anal., 18 (1975),
318–327.
[4] A. Connes and E. J. Woods: Hyperfinite von Neumann algebras and Poisson boundaries of time
dependent random walks, Pacific J. Math. 137 (1989), no 2, 225–243.
[5] C. Davis: Various averaging operations on subalgebras, Illinois J. Math., 3 (1959), 538–553.
[6] E. Dynkin: The exit space of a Markov process, Russian Math. Surveys 24, No. 4, (1969), 89–152.
[7] G. Elliott and T. Giordano: Amenable actions of discrete groups, Ergodic Theory Dyn. System, 13
(1993), no.2, 289–318.
[8] D. E. Evans: Quasi-product states on C∗-algebras, Lecture Notes in Mathematics, No. 1132, Springer-
Verlag, Berlin-New York, 1985, 129-151.
[9] J. Feldman and C. Moore: Ergodic equivalence relations, cohomologies, von Neumann algebras, I
and II, Trans. Amer. Math. Soc., 234 (1977), 289-359.
[10] T. Giordano and D. Handelman: Matrix-valued random walks and variations on property AT,
Munster J. of Math. 1(2008),15–72.
[11] P. Hahn: Haar measures for measure groupoids, Trans. Amer. Math. Soc., 242, (1978) , 1–33.
[12] J.G. Kemeny and J.L. Snell : Finite Markov chains, Van Nostrand, Princeton N.J., 1960.
[13] A. Kumjian and D. Pask: Higher rank graph C*-algebras, New York J. Math. 6 (2000), 1?20.
[14] J. Neveu: Bases math´ematiques du calcul des probabilit´es, Masson & Cie, Paris, 1964.
[15] J. Renault: A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, Vol. 793 Springer-
Verlag Berlin, Heidelberg, New York (1980).
[16] J. Renault: AF equivalence relations and their cocycles, Operator Algebras and Mathematical
Physics, Conference Proceedings, Constanza 2001, the Theta Foundation (2003), 365-377.
[17] J. Renault: The Radon-Nikodym problem for approximately proper equivalence relations, Ergodic
Theory Dyn. System, 25 (2005), 1643–1672.
[18] A. Vershik Equipped graded graphs, projective limits of simplices, and their boundaries, Zapiski
Nauchn. Semin. POMI 432, (2015), 83–104, English translation to appear in J. Math. Sci.
18
Jean Renault
[19] A. Vershik Asymptotic theory of path spaces of graded graphs and its applications, Japanese J. Math.
11, No. 2, (2016), 151–218 .
Universit´e d'Orl´eans et CNRS (UMR 7349 et FR2964), D´epartement de Math´ematiques,
F-45067 Orl´eans Cedex 2, France
E-mail address: [email protected]
|
1810.07049 | 1 | 1810 | 2018-10-16T14:48:31 | The module embedding theorem via towers of algebras | [
"math.OA",
"math.CT",
"math.QA"
] | Jones and Penneys showed that a finite depth subfactor planar algebra embeds in the bipartite graph planar algebra of its principal graph, via a Markov towers of algebras approach. We relate several equivalent perspectives on the notion of module over a subfactor planar algebra, and show that a Markov tower is equivalent to a module over the Temperley-Lieb-Jones planar algebra. As a corollary, we obtain a classification of semisimple pivotal C* modules over Temperley-Lieb-Jones in terms of pointed graphs with a Frobenius-Perron vertex weighting. We then generalize the Markov towers of algebras approach to show that a finite depth subfactor planar algebra embeds in the bipartite graph planar algebra of the fusion graph of any of its cyclic modules. | math.OA | math |
The module embedding theorem via towers of algebras
Desmond Coles, Peter Huston, David Penneys, and Srivatsa Srinivas
October 17, 2018
Abstract
Jones and Penneys showed that a finite depth subfactor planar algebra embeds in the bi-
partite graph planar algebra of its principal graph, via a Markov towers of algebras approach.
We relate several equivalent perspectives on the notion of module over a subfactor planar alge-
bra, and show that a Markov tower is equivalent to a module over the Temperley-Lieb-Jones
planar algebra. As a corollary, we obtain a classification of semisimple pivotal C∗ modules over
Temperley-Lieb-Jones in terms of pointed graphs with a Frobenius-Perron vertex weighting. We
then generalize the Markov towers of algebras approach to show that a finite depth subfactor
planar algebra embeds in the bipartite graph planar algebra of the fusion graph of any of its
cyclic modules.
1
Introduction
Jones' planar algebras [Jon99] are a powerful method to construct [BMPS12, GMP+18] and classify
[JMS14, Liu15, AMP15] finite index II1 subfactors. Many exotic examples have been constructed
via graph planar algebra embedding, i.e., by finding evaluable planar subalgebras of graph planar
algebras. By [JP11], any finite depth subfactor planar algebra embeds in the graph planar algebra of
its principal graph. This result also extends to infinite depth subfactor planar algebras by [MW10].
To date, graph planar algebra embedding has been used to construct:
• the E6 and E8 subfactor planar algebras [Jon01],
• group planar algebras [Gup08],
• Haagerup-Izumi quadratic subfactors [Pet10, Han10, MP15b, MP15a, PP15],
• quantum group planar algebras [LMP15], and
• the extended Haagerup fusion categories [BMPS12, GMP+18].
While none of the constructions above rely on the embedding theorem from [JP11], the embedding
theorem gives us the motivation to do the hard work of looking for the embedding. However, the
embedding theorem is necessary for Liu's important classification theorem for composites of A3
and A4 subfactor planar algebras [Liu15], in which he shows that higher quotients of A3 ∗ A4 do
not exist because the possible generator does not embed in the appropriate graph planar algebra.
As noted in [JP11], it was rather surprising that the dual principal graph made no appearance
in the embedding theorem. Adding to this mystery, certain examples above could be constructed
by embedding into planar algebras of bipartite graphs which are completely different from the
principal and dual principal graphs [Pet10, PP15, GMP+18]. The answer to why this occurs is the
following theorem:
1
Theorem 1.1 ([GMP+18]). Let P• be a finite depth subfactor planar algebra and C its unitary 2×2
multifusion category of projections with generator X ∈ P1,+, the unshaded-shaded strand. Endow
C with the canonical spherical structure inherited from P•. There is a bijective correspondence
between:
• planar †-algebra embeddings P• → GPA(Γ)•, where Γ is a finite connected bipartite graph,
and
• indecomposable finitely semisimple pivotal left C-module C∗ categories M whose fusion graph
with respect to X is Γ.
The proof in [GMP+18] is mostly in the language of tensor and module categories.
In this
article, we provide an independent proof in the original towers of algebras approach to subfactor
theory [Jon83, Wen88, GdlHJ89, Pop94] and the graph planar algebra embedding theorem [JP11].
Our starting point is the well-known correspondence between:
(1) unitary 2×2 multitensor categories C with orthogonal decomposition into simples 1C = 10⊕11
and generator X = 10 ⊗ X ⊗ 11 with its canonical spherical/balanced unitary dual functor
(see [Yam04, Pen18]),1 and
(2) Jones' subfactor planar algebras P• [Jon99].
In §2, we build on this correspondence by defining analogous notions of right modules for these
algebraic objects. We briefly describe these objects here, and we refer the reader to §2.2 for more
details.
A pivotal module category for C is a finitely semisimple C∗ category M which is an indecompos-
able right C-module category equipped with a faithful positive trace TrM
m on each endomorphism
C∗ algebra M(m → m) which is compatible with the right C-action [Sch13, GMP+18]. That is, for
all m ∈ M, c ∈ C, and f ∈ EndM(m (cid:67) c),
TrM
c) ◦ (f (cid:67) idc) ◦ (idm (cid:67) coevc)),
m(cid:67)c(f ) = TrM
m ((idm (cid:67) coev†
where (c, evc, coevc) is the canonical balanced dual of c ∈ C [LR97, Yam04, BDH14, Pen18].
A (connected) right planar module M• for a subfactor planar algebra P• is a sequence of finite
dimensional von Neumann algebras (Mk)k≥0 with dim(M0) = 1,2 together with an action of the
shaded planar module operad, which is a variation of Voronov's Swiss cheese operad [Vor99]. We
refer the reader to Definition 2.7 for the details, but we include a representative tangle below which
acts amongst the algebras Mk and the box spaces Pn,±: 3
2
1
(cid:63)
5
4
(cid:63)
(cid:63)
3
: (M3 ⊗ M1) ⊗ (P2,+ ⊗ P1,− ⊗ P3,+) → M4
product, direct sum, taking dual, and taking subobjects is equivalent to C; see Definition 2.4.
1 That X is a generator means that any proper full subcategory of C containing X which is closed under tensor
2 The adjective connected refers to the condition that dim(M0) = 1.
3 We use the convention that all Mk appear before Pn,± in the tensor product; this is not problematic, as the
tensor category of finite dimensional complex vector spaces is symmetric.
2
Here, one can glue shaded planar module tangles into the module input semidisks, and one can
glue shaded planar tangles into the circular input disks. In addition, the tower of algebras (Mk)k≥0
must satisfy that multiplication in the von Neumann algebra Mk is given by the tangle
k
2
k
1
k
: Mk ⊗ Mk → Mk,
and the ∗-structure on Mk is compatible with the reflection of tangles about a horizontal axis.
Notice this canonically identifies M0 = C as a von Neumann algebra. Under this identification, we
require that the linear functionals
trk := d−k ·
: Mk → M0 = C
k
The following theorem generalizes the correspondence between unitary 2 × 2 multitensor cate-
are faithful positive normalized traces, where d is the loop parameter of P•.
gories C with 1C = 10 ⊕ 11 and generator X = 10 ⊗ X ⊗ 11 and subfactor planar algebras P•.
Theorem A. Let P• be a subfactor planar algebra corresponding to (C, X) as above. There is an
equivalence between:
(1) pivotal right C-module C∗ categories (M, TrM) with choice of simple basepoint m = m (cid:67) 10,
and
(2) connected right planar modules M• for P•.
One passes from (2) to (1) in Theorem A by taking the category of projections, similar to the
correspondence between P• and (C, X) [MPS10, BHP12]. One passes from (1) to (2) using the
diagrammatic calculus for module categories, similar to how one gets a subfactor planar algebra
from (C, X) via the diagrammatic calculus for pivotal categories [Gho11, Pen18].
From a pivotal semisimple right C-module C∗ category (M, TrM) together with a choice of
simple basepoint m ∈ M with m = m (cid:67) 10, we build a tower of finite dimensional von Neumann
algebras by setting
Mn := EndM(m (cid:67) X ⊗ X ⊗ ··· ⊗ X ?
)
(cid:124)
(cid:123)(cid:122)
n tensorands
(cid:125)
where for our generator X ∈ C, we set X ? = X if n is even and X ? = X if n is odd. The trace
TrM endows each von Neumann algebra Mn with a faithful tracial state trn := TrM(idn)−1 TrM
together with canonical Jones projections en ∈ Mn+1 for all n ≥ 1. Based on the parity of n, the
en are defined for k ≥ 0 by
:= d−1(cid:0) idm (cid:67) id(X⊗X)⊗k ⊗(coevX ◦ coev
:= d−1(cid:0) idm (cid:67) idX ⊗ id(X⊗X)⊗k ⊗(ev
X )(cid:1) ∈ M2k+2
X ◦ evX )(cid:1) ∈ M2k+3.
†
†
e2k+1 =
2k
e2k+2 =
2k + 1
Here, (X, evX , coevX ) is the balanced dual of X, m is graphically represented by a red strand, and
the left hand side of m is shaded red to denote the absence of a left C-action.
We call M• = (Mn, trn, en+1)n≥0 a Markov tower as it satisfies the following axioms:
3
(M1) The projections (en) satisfy the Temperley-Lieb-Jones relations with modulus d > 0 (our
convention for d is eiei±1ei = d−2ei.)
(M2) For all x ∈ Mn, enxen = En(x)en, where En : Mn → Mn−1 is the canonical trace-preserving
conditional expectation.
(M3) For all n ≥ 1, En+1(en) = d−2.
(M4) For all n ≥ 1, we have the Pimsner-Popa pull down property [PP86]: Mn+1en = Mnen, which
is equivalent to MnenMn being a 2-sided ideal in Mn+1.
One should view a Markov tower as an analog of Popa's λ-lattices [Pop95] where we only have one
tower of algebras rather than a tower/lattice of commuting squares. Indeed, one should compare
(M1) and (M2) with (1.3.2) and (M3) and (M4) with (1.3.3') from [Pop95] respectively. We expect
that the notion of Markov tower with some compatibility axioms is the correct notion of a right
module for Popa's λ-lattices (see Remark 3.34). We leave this exploration to a future article as it
would take us too far afield.
Markov towers satisfy many nice properties exhibited by standard invariants of finite index II1
subfactors from [GdlHJ89, Ch. 4]; we mention a few here, and we refer the reader to §3 for more
details. The traces satisfy the Markov property trn+2(xen) = d−2 trn+1(x) for every x ∈ Mn+1, and
the Markov tower has a principal graph consisting of the non-reflected part of the Bratteli diagram
at each step. The tower is called finite depth if the principal graph is finite.
From a Markov tower, we can form a semisimple C∗ projection category M, whose simple
objects are in canonical bijection with the vertices of the principal graph. Moreover, the traces and
Jones projections canonically equip M with the structure of a pivotal right T LJ (d)-module C∗
category. Now any pointed bipartite graph (Γ, v) with a quantum dimension function on vertices
dim : V (Γ) → R>0 satisfying
gives us a Markov tower of modulus d, where we write w ∼ v to mean w is connected to v, and
the sum is taken with multiplicity. We thus get the following corollary, which should be compared
with the non-pivotal case in [DCY15].
Corollary B. Equivalence classes of pivotal T LJ (d)-module C∗ categories with simple basepoint
are in bijection with pointed connected bipartite graphs (Γ, v) with a quantum dimension function.
We now specialize to the hypotheses of the module embedding theorem, i.e., Q• is a finite
depth subfactor planar algebra, (C, X) is its corresponding spherical unitary multifusion category
of projections with generator X = 10 ⊗ X ⊗ 11 the unshaded-shaded strand, and (M, TrM, m) is a
pivotal right C-module C∗ category with simple basepoint m = m (cid:67) 10. In this case, the Markov
tower M• constructed above has finite depth, and its principal graph Γ is the fusion graph of (M, m)
with respect to X ∈ C. This means there is an r > 0 such that the inclusion M2r ⊂ (M2r+1, tr2r+1)
is strongly Markov, meaning that there is a finite Pimsner-Popa basis {b} for M2r+1 over M2r
b bb∗ [Wat90] is a scalar (see [Pop94,
1.1.4(c)]).
By [JP11, §2.3], the inclusion A0 := M2r ⊂ (M2r+1, tr2r+1) =: (A1, tr1) has a canonical associ-
ated planar †-algebra P•, which is built from the tower of higher relative commutants. Moreover,
by [JP11, Thm. 3.8], the planar algebra P• is non-canonically isomorphic to the bipartite graph
planar algebra G• of the Bratteli diagram of the inclusion A0 ⊂ A1, which is also the fusion graph
Γ. (This isomorphism depends on the loop algebra representation for A0 ⊂ A1 from [JP11, §3.1],
which amounts to choosing compatible bases for the algebras.)
b be2rb∗ = 1M2r+2, and the Watatani index (cid:80)
satisfying (cid:80)
d · dim(v) =
dim(w)
(cid:88)
w∼v
4
Theorem C (Module embedding). The unital †-algebra maps Φn,± := idm (cid:67) id(X⊗X)⊗r (cid:67) − :
Qn,± → Pn,±
Φ
x
n
x
2rm
n
give a planar †-algebra embedding Q• (cid:44)→ P•.
Choosing M = C00 ⊕ C10 and m = 10 corresponding to the unshaded empty diagram exactly
recovers the embedding into the graph planar algebra of the principal graph of Q• from [JP11].
Similarly, we get an embedding into the graph planar algebra of the dual principal graph by choosing
M = C10 ⊕ C11 and an arbitrary simple object m ∈ C10.
Notice we made three choices in our proof of the Module Embedding Theorem C; we picked
a simple object m ∈ M with m = m (cid:67) 10, an r ≥ 0 such that M2r ⊂ (M2r+1, tr2r+1) is strongly
Markov, and a planar †-algebra isomorphism Q• ∼= GPA(Γ)•. In our final Section 5.2, we explain
that different choices still produce an equivalent planar †-algebra embedding Q• → GPA(Γ)•.
Indeed, we show that the two corresponding strongly Markov inclusions are related by a shift and a
compression by a projection with central support 1, and these processes yield planar †-isomorphisms
on the associated canonical relative commutant planar algebras.
Acknowledgements. This article is the undergraduate research project of Desmond Coles and
Srivatsa Srinivas, which was supervised by Peter Huston and David Penneys during Summer 2018,
supported by Penneys' NSF DMS CAREER grant 1654159. The authors would like to thank
Corey Jones for many helpful conversations. Additionally, David Penneys would like to thank
Emily Peters and Noah Snyder for helpful conversations.
2 Modules for subfactor planar algebras
The standard invariant of a finite index II1 subfactor has many axiomatizations, including Popa's
λ-lattices [Pop95] and Jones' subfactor planar algebras [Jon99]. Here, we use the language of
subfactor planar algebras. We discuss the well-known correspondence between subfactor planar
algebras and their projection unitary multitensor categories. We then introduce the notion of a
planar module for a subfactor planar algebra, and we show it corresponds to a module category for
the projection unitary multitensor category.
2.1 Unitary multitensor categories and subfactor planar algebras
In this section, we rapidly recall the definitions of a subfactor planar algebra [Jon99] and its unitary
2 × 2 multitensor category of projections [Gho11, Pen18].
Definition 2.1. The shaded planar operad consists of shaded planar tangles with the operation
of composition. Shaded planar tangles have r ≥ 0 input disks each with 2ki boundary points, and
an output disk with 2k0 boundary points. Internal to the output disk are non-intersecting strings,
which either attach 2 distinct boundary points, or are closed loops. There is also a checkerboard
shading, and a distinguished interval marked (cid:63) for each input disk and the output disk.
If the
(cid:63) for the i-th disk is on an interval which meets an unshaded region, that disk has type (ki, +),
and if it meets a shaded region, the disk has type (ki,−). A tangle with r input disks has type
((k0,±0); (k1,±1), . . . , (kr,±r)) if the output disk has type (k0,±0) and the i-th input disk has type
5
(ki,±i). There is a natural definition of the composite tangle T ◦i S when the output disk of a
tangle T has the same type as the i-th input disk of a tangle T . We give a representative example
below, and we refer the reader to [Pet10, Jon12] for a more precise definition.
(cid:63)
(cid:63)
2
(cid:63)
3
(cid:63)
1
◦3 (cid:63)
(cid:63)
1
= (cid:63)
(cid:63)
2
(cid:63)
3
(cid:63)
1
1 ≤ i ≤ r and output disk of type (k0,±0) defines a multilinear map Z(T ) :(cid:81)r
The shaded planar operad also has a †-structure, with the tangle T † obtained by reflecting T about
a diameter.
A shaded planar algebra P• consists of a family Pn,± of C-vector spaces together with an action of
the shaded planar operad. That is, each shaded planar tangle T with input disks of type (ki,±i) for
i=1 Pki,±i → Pk0,±0,
and tangle composition corresponds to composition of multilinear maps. Each Pn,± should also
†
have a dagger structure so that for every tangle T and tuple η1 . . . r of inputs, Z(T †)(η
r) =
Z(T )(η1, η2, . . . ηn)†.
†
2 . . . η
†
1, η
Notation 2.2. We will try to shade our diagrams as much as possible for a shaded planar algebra.
However, sometimes shading our diagrams requires us to split into many cases. In order to avoid
this, we sometimes suppress the shading when it can be inferred from the indices. We also tend to
suppress the external boundary disk of a shaded planar tangle; when we do so, the (cid:63) is always on
the left. For explicit examples, compare (PA3) and (PA4) in the following definition.
Definition 2.3. A shaded planar algebra is called a subfactor planar algebra if moreover
(PA1) (finite dimensional) dim(Pn,±) < ∞ for all n ≥ 0.
(PA2) (evaluable/connected) dim(P0,±) = 1.
(PA3) (positive) (cid:104)x, y(cid:105)n,± :=
(cid:63) defines a positive-definite inner product on each Pn,±.
y†
n
(cid:63)
x
(PA4) (spherical) for all x ∈ P1,+,
x(cid:63)
=
x(cid:63)
.
In a subfactor planar algebra, closed contractible loops can be traded for a multiplicative scalar d >
0; we call this the loop parameter. By Jones' index rigidity theorem [Jon83], d ∈ {2 cos(π/k)k ≥ 3}∪
[2,∞).
Given a subfactor planar algebra P•, we get two towers of finite dimensional von Neumann
algebras P± = (Pn,±)n≥0 with Jones projections for k ≥ 0 given by
e2k+1,+ := 2k
e2k+2,+ := 2k + 1
e2k+1,− := 2k
e2k+2,− := 2k + 1
(1)
6
and traces for n ≥ 0 given by
trn,± := d−n
n .
(2)
We will see in §3.1 below that (Pn,±, trn,±, en+1,±)n≥0 has the structure of a Markov tower, which
comes with a principal graph. The principal graph of P+ is finite if and only if the principal graph
of P− is finite; in this case, P• is said to have finite depth.
Definition 2.4. A unitary 2 × 2 multitensor category C is an indecomposable rigid C∗ tensor
category which is Karoubi complete such that 1C has an orthogonal decomposition into simple
objects as 1C = 10 ⊕ 11. We write Cij := 1i ⊗ C ⊗ 1j for i, j ∈ {0, 1}. By [LR97], such a C is
automatically semisimple. When C is finitely semisimple, it is called a unitary 2 × 2 multifusion
category [EGNO15].
We say X ∈ C01 generates C if every object of C is isomorphic to a direct summand of an
alternating tensor power of X and X
X alt⊗n := X ⊗ X ⊗ ··· ⊗ X ?
alt⊗n
X
:= X ⊗ X ⊗ ··· ⊗ X
(cid:124)
(cid:123)(cid:122)
(cid:125)
(cid:124)
(cid:123)(cid:122)
?
(cid:125)
n tensorands
n tensorands
where X ? = X if n is odd and X when n is even, and X
= X when n is odd and X when n
is even. Here, (X, evX , coevX ) is the canonical balanced dual of X [BDH14, GL18, Pen18] which
satisfies the zig-zag axioms and the balancing equation
?
ψ(evX ◦(idX ⊗f ) ◦ ev
†
X ) = ψ(coev
†
X ◦(f ⊗ idX ) ◦ coevX )
∀f ∈ C(X → X)
where ψ : C(1C → 1C) → C is the linear functional such that ψ(id10) = ψ(id11) = 1.
The following theorem is well-known to experts.
Theorem 2.5. There is an equivalence of categories 4
(cid:26)
Subfactor planar algebras P•
(cid:27) ∼=
(cid:26) Pairs (C, X) with C a unitary 2×2 multitensor
category together with a generator X ∈ C01
(cid:27)
.
Starting with a subfactor planar algebra P•, one may form its unitary 2×2 multitensor category
of projections C [MPS10, BHP12, Pen18], which comes with a canonical generator corresponding
to the unshaded-shaded strand in P1,+, and the canonical spherical unitary dual functor [BDH14,
GL18, Pen18]. This unitary 2×2 multitensor category can also be thought of as a unitary 2-category
called the paragroup; we refer the reader to [BP14] for more details.
Starting with a pair (C, X), we get a subfactor planar algebra by defining
Pn,+ := EndC(X alt⊗n)
Pn,− := EndC(X
alt⊗n
),
and we define the action of the shaded planar operad via the diagrammatic calculus for pivotal
tensor categories. We refer the reader to [Gho11, Pen18] for more details.
4We suppress the subtlety about the right hand side of this equivalence being a contractible 2-category. We refer
the reader to [HPT16, Pen18] for more details.
7
2.2 Modules for unitary multitensor categories and subfactor planar algebras
We now define the various notions of module for
• a unitary 2 × 2 multitensor category C with its canonical unitary spherical structure and a
generator X ∈ C01, and
• a subfactor planar algebra Q•.
n
n (f ◦ g) for all f ∈ M(m → n) and g ∈ M(n → m).
m (g ◦ f ) = TrM
m (f† ◦ f ) ≥ 0 for all f ∈ M(m → n), and TrM
Definition 2.6. Let C be a unitary 2×2 multitensor category. A pivotal right C-module C∗ category
is a pair (M, TrM) where M is a semisimple right C-module C∗ category, and TrM is a family of
: M(n → n) → C on each endomorphism space for n ∈ M satisfying the
positive traces TrM
following axioms:
(Tr1) TrM
(Tr2) TrM
(Tr3) For all m ∈ M and c ∈ C, TrM
Notice that M = M0 ⊕ M1 where M0 = M (cid:67) 10 and M1 = M (cid:67) 11.
A pivotal right module category is called pointed if it is indecomposable, we have a chosen
simple object m ∈ M, and TrM is normalized so that TrM
m (idm) = 1C. Generally, we choose
m ∈ M0, but this choice is not essential.
When C is generated by a single X ∈ C01 and (M, TrM, m) is a pointed pivotal right module
category with m ∈ M0, we define the cyclic pivotal right module category Mm,X to be the (non
Karoubi complete!) full subcategory of M whose objects are of the form m (cid:67) X alt⊗k for k ≥
0, which is a pointed pivotal right module category over CX , the (non Karoubi complete!) full
subcategory of C whose objects are of the form X alt⊗k and X
m (f† ◦ f ) = 0 if and only if f = 0.
m(cid:67)c(f ) = TrM
m ((idm (cid:67) coev
†
c) ◦ (f (cid:67) idc) ◦ (idm (cid:67) coevc))
alt⊗k
for k ≥ 0.
We next give an appropriate definition of planar modules over a planar algebra as algebras over
another operad.5
Definition 2.7. The shaded planar module operad is a variant of the shaded planar operad, akin
to a shaded, stranded version of the Swiss-cheese operad introduced in [Vor99]. In this operad, the
starred region of the boundary of the output disk of a tangle is replaced by a vertical line on the
left side of a tangle, and the adjacent region inside the tangle must be unshaded. In addition to the
usual input disks, tangles may also have input semidisks, whose boundaries intersect the left wall.
Similar to the definition of type for an input disk, a semidisk (input or output) has type ki if it has
2ki boundary points which meet 2ki strings. A tangle with r input semidisks and s input disks has
type (k0; k1, . . . , kr; ((cid:96)1,±1), . . . , ((cid:96)s,±s)) if the output disk has type k0, the i-th input semidisk has
type ki, and the j-th input disk has type ((cid:96)j,±j). The operadic composition comes from plugging
tangles into semidisks when the types are compatible. A representative tangle appears below.
5 Our definition of a planar module over a planar algebra differs significantly from the annular planar modules
introduced in [Jon01]. Our planar modules will correspond to the above notion of module over a multitensor category,
whereas annular planar modules are more closely related with representations of the tube algebra/affine annular
category, which correspond to objects in the Drinfeld center of the multitensor category [GJ16].
8
2
1
(cid:63)
5
4
(cid:63)
(cid:63)
3
: (M3 ⊗ M1) ⊗ (P2,+ ⊗ P1,− ⊗ P3,+) → M4
Tangles of the shaded planar module operad can also be composed with shaded planar tangles, by
plugging a shaded planar tangle into an input disk. One should think of the box spaces for semidisks
as being endomorphisms of objects in a module category, while the involvement of full disks allows
a planar algebra to act on the module. The input disks and semidisks are also numbered, with the
numbering determining the order of the tensor factors in the domain of the action map as depicted
above. Like the shaded planar operad, the shaded planar module operad is a symmetric operad,
and vector spaces form a symmetric monoidal category, so we often suppress the numbering.
Definition 2.8. A right planar module M• for the subfactor planar algebra P• consists of a
sequence of finite dimensional C-vector spaces (Mk)k≥0 and a conjugate-linear map † : Mk → Mk
for all k ≥ 0, together with an action of the shaded planar module operad on the box spaces M•
and P• compatible with the composition of tangles and the shaded planar algebra structure on P•,
and the † operation. In other words, the box spaces M• and P• together must have the structure
of an algebra over the shaded planar module operad, which must extend the original shaded planar
operad algebra structure on P•.
Notice that each of the Mk has a †-algebra structure with multiplication given by the tangle
k
2
k
1
k
: Mk ⊗ Mk → Mk.
We require that each †-algebra Mk is a finite dimensional C∗/W∗ algebra. Moreover, we require
that for each k, the following map Mk → M0 is positive, faithful, and tracial:
Trk :=
: Mk → M0
k
(3)
We call M• connected if dim(M0) = 1. In this case, we can canonically identify M0 = C as a
C∗-algebra, and each Trk is a scalar-valued. We define trk := d−k Trk, where d is the loop parameter
of P•. Notice that the trk are faithful tracial states. We will see that, under the correspondence of
Theorem A, connected right planar modules correspond to cyclic pivotal right module C∗-categories.
Example 2.9. Given a subfactor planar algebra P•, P+ := (Pk,+)k≥0 is a cyclic right planar
module for P•, while P− := (Pk,−)k≥0 is a right planar module for the dual planar algebra of P•
obtained by reversing the shading.
Example 2.10. Suppose G• = GPA(Γ)• is the graph planar algebra of the bipartite graph Γ.
Then for any +/even vertex v of Γ, we get a cyclic right planar module M• = M(Γ, v)• by defining
9
⊗ x1 ⊗ ··· ⊗ xs
(cid:124)
(cid:123)(cid:122)
∈P•
m1 ⊗ ··· ⊗ mr
(cid:123)(cid:122)
(cid:125)
(cid:124)
M(Γ, v)k := pvGk,± and action of the planar module operad by
:= ZG•((cid:101)T )(pv ⊗ m1 ⊗ ··· ⊗ mr ⊗ Φ(x1) ⊗ ··· ⊗ Φ(xs))
(cid:125)
ZM•(T )
where (cid:101)T is obtained from T by first turning each half-open input semidisk into a closed input
disk in the interior of the output disk with its (cid:63) on the left, rounding out the 90◦ angles on the
left boundary into a smooth curve, putting the external (cid:63) on the left hand side, and inserting
one (0, +)-type input disk in the left-most region of the new tangle which is numbered first. We
illustrate this procedure on the tangle above:
∈M•
T =
2
1
(cid:63)
5
4
(cid:63)
(cid:63)
3
(cid:55)−→ (cid:101)T :=
(cid:63)
3
(cid:63)
6
(cid:63)
1(cid:63)
5
(cid:63)
(cid:63)
2
(cid:63)
4
2.3 Equivalence of modules
In this section, P• will denote a subfactor planar algebra, and (C, X) will denote its unitary 2 × 2
multitensor category of projections, where X = 10 ⊗ X ⊗ 11 is the generating unshaded-shaded
strand. We now sketch the proof of the following theorem.
Theorem (Theorem A). There is a canonical bijection between equivalence classes of
(1) indecomposable pivotal right C-module C∗ categories (M, TrM) with simple basepoint m =
m (cid:67) 10, and
(2) connected right planar modules M• for P•.
As an application, we get a classification of pivotal module C∗ categories for the 2-shaded
Temperley-Lieb-Jones category with parameter d in Corollary B, whose proof appears in §3.5.
Definition 2.11. Suppose (M, TrM, m) is an indecomposable pivotal right C-module C∗ category,
and m ∈ M is a distinguished simple object with m = m (cid:67) 10. We build a connected right planar
P•-module M• by defining Mk := EndC(m (cid:67) X alt⊗k) for k ≥ 0, and we define the action of the
shaded planar module operad via the diagrammatic calculus for M. The process is similar to that
in [JP11, Def. 3.2].
We first define a standard form for tangles of the shaded planar module operad, such that every
tangle is isotopic to one in standard form. We say a tangle is in standard form if
(SF1) Each disk and semidisk, including the output semidisk, is rectangular in shape, with an equal
number of strings emerging from the top and bottom,
(SF2) in the case of a disk, the starred boundary interval includes the left side, and
(SF3) a horizontal line through the tangle passes through a disk, semidisk, or extremum of a strand
at most once.
10
Given a shaded planar module tangle T with type (k0; k1, . . . , kr; ((cid:96)1,±1), . . . , ((cid:96)s,±s)) together
with appropriate inputs (f1, . . . , fr, x1, . . . , xs) with mi ∈ Mki and xj ∈ P(cid:96)j ,±j , we begin with
the identity morphism of m (cid:67) X altk0 in Mk0, and we move an imaginary horizontal line upwards
along the tangle. Each time the horizontal line passes a disk, semidisk, or local extrema (which
can happen only one at a time!), we compose with a morphism from M. In more detail, when the
horizontal line passes:
• the i-th input semidisk with vertical stands to the right, we compose with fj (cid:67) id, where id
is the appropriate identity morphism corresponding to the strands to the right of the input
semidisk
• the j-th input disk with vertical strands to the left and right, we compose with idm (cid:67)
idl ⊗x ⊗ idr where idl, idr correspond to the appropriate identity morphisms corresponding
to the strands to the left and right of the input disk
• a local extrema with vertical strands to the left and right, we compose with idm (cid:67) idl ⊗v⊗ idr
where v stands for the following (co)evaluation or its dagger depending on the shading:
†
X ,
(cid:32) coevX
(cid:32) coev
(cid:32) ev
†
X
(cid:32) evX
and idl, idr are the appropriate identity morphisms as above.
The output is the composite morphism in Mk0 = EndM(m (cid:67) X altk0). One then checks that the
resulting composite morphism is independent of the choice of standard form for the shaded planar
module tangle T .
Example 2.12. Here is an explicit example of a tangle in standard form, together with the cor-
responding multi-linear map obtained by composing the associated morphisms in M from bottom
to top:
†
idm (cid:67) idX ⊗ coev
X ⊗ idX
− (cid:67) idX⊗X
idm (cid:67) idX⊗X⊗X ⊗−
†
idm (cid:67) idX ⊗ coev
X ⊗ idX⊗X⊗X
idm (cid:67) idX⊗X ⊗ − ⊗ idX⊗X
†
idm (cid:67) idX⊗X⊗X ⊗ ev
X ⊗ idX
− (cid:67) idX⊗X⊗X
idm (cid:67) coevX ⊗ idX⊗X
2
1
(cid:63)
4
(cid:63)
3
: M1 ⊗ M2 ⊗ P2,+ ⊗ P1,− −→ M2
Definition 2.13. Given a connected right planar P•-module M•, we let M be its category of
projections. The objects of M are the orthogonal projections in Mk for k ≥ 0. The Hom-space
M(p → q) for p ∈ Mj and q ∈ Mk is only nonzero if j ≡ k mod 2; in this case, we define
M(p → q) :=
x ∈ M(j+k)/2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
11
k
x
j
=
q
x
p
k
k
j
j
.
Composition is given by a suitable version of the usual multiplication tangle, and the †-structure
is given by † in M•. Given a projection p ∈ Mk and projections q ∈ Pn,+ and r ∈ Pn,−, we define
depending on parity
p (cid:67) q :=
k=2j
p
2j
2j
q(cid:63)
n
n
p (cid:67) r
:=
k=2j+1
2j + 1
q(cid:63)
p
2j + 1
n
n
.
For morphisms f ∈ M(p → q) with p ∈ Mj and q ∈ Mk and g ∈ C(r → s) with r ∈ Pm,± and
s ∈ Pn,± such that p (cid:67) r and q (cid:67) s are well-defined, we define
f (cid:67) g := f
k
n
g(cid:63)
.
j
m
where shading depends on the parity of j and k. We leave the straightforward verification that
M is a right C-module C∗ category to the reader. Finally, we replace M with its unitary Karoubi
completion, which formally adds orthogonal direct sums and then takes the orthogonal projection
completion.
The distinguished simple basepoint of M is given by the identity projection 1M0 ∈ M0. The
: M(p → p) → C for p ∈ Mk is given by the non-normalized trace Trk from (3)
trace TrM
restricted to pMkp = M(p → p). By definition, we have assumed Trk to be tracial and positive on
endomorphisms. That the trace is also compatible with the action of C can be seen by composing
the trace tangle with the action tangles.
p
3 Markov towers and their projection categories
So far, we have presented two versions of the concept of a module over a subfactor planar algebra.
The algebraic data of each shares a common structure: that of a Markov tower of finite dimensional
tracial von Neumann algebras. Studying elementary properties of Markov towers will therefore
allow us to state many important results about planar modules in single, common language. The
definition of a Markov tower can obtained from the definition of Popa's λ-sequences of commuting
squares from [Pop95] by forgetting one of the towers, analogous to the way one defines a module for
an algebraic object by replacing one argument of the algebraic operation with an element from the
module. In short, Markov towers are the towers-of-algebras analog of a module category. In §3.5
below, we will see that Markov towers are exactly a λ-lattice approach to pivotal Temperley-Lieb-
Jones module categories; this motivates the view of subfactor planar modules as simply Markov
towers with an additional structure.
3.1 Markov towers and their elementary properties
Definition 3.1. A Markov tower M• = (Mn, trn, en+1)n≥0 consists of a sequence (Mn, trn)n≥0 of
finite dimensional von Neumann algebras, such that Mn is unitally included in Mn+1, each Mn has
a faithful normal tracial states such that trn+1 Mn = trn for all n ≥ 0, and there is a sequence of
Jones projections en ∈ Mn+1 for all n ≥ 1, such that:
(M1) The projections (en) satisfy the Temperley-Lieb-Jones relations:
i = ei = e∗
(TLJ1) e2
(TLJ2) eiej = ejei for i − j > 1, and
i for all i,
12
(TLJ3) there is a fixed constant d > 0 called the modulus such that eiei±1ei = d−2ei for all i.
(M2) For all x ∈ Mn, enxen = En(x)en, where En : Mn → Mn−1 is the canonical faithful trace-
preserving conditional expectation.
(M3) For all n ≥ 1, En+1(en) = d−2.
(M4) (pull down) For all n ≥ 1, Mn+1en = Mnen.
We call a Markov tower connected if dim(M0) = 1.
Remark 3.2. One should think of the preceding definition as obtained from Popa's definition of λ-
sequence [Pop95] and removing one of the two sequences of algebras, together with the commuting
square condition. Compare the existence of Jones projections, (M1), and (M2) with (1.3.2), and
(M3) and (M4) with (1.3.3') from [Pop95] respectively.
Remark 3.3. Observe that MnenMn is a 2-sided ideal in Mn+1 for all n ≥ 1 if and only if
the pull down condition holds. Indeed, if the pull down condition holds, then Mn+1MnenMn ⊆
Mn+1enMn = MnenMn; the same argument holds on the right by first taking adjoints. Conversely,
if MnenMn is a 2-sided ideal, then Mn+1en = (Mn+1en)en ⊆ (MnenMn)en = Mnen.
Proposition 3.4. A Markov tower satisfies the following elementary properties for n ≥ 1.
(EP1) The map Mn (cid:51) y (cid:55)→ yen ∈ Mn+1 is injective.
(EP2) For all x ∈ Mn+1, d2En+1(xen) is the unique element y ∈ Mn such that xen = yen [PP86,
Lem. 1.2].
(EP3) The traces trn+1 satisfy the following Markov property with respect to Mn and en: for all
x ∈ Mn, trn+1(xen) = d−2 trn(x).
(EP4) enMn+1en = Mn−1en.
(EP5) Xn+1 := MnenMn is a 2-sided ideal of Mn+1, and thus Mn+1 splits as a direct sum of von
Neumann algebras Xn+1 ⊕ Yn+1. (In [GdlHJ89, Thm. 4.1.4 and Thm. 4.6.3], Yn+1 is the
so-called 'new stuff '.) By convention, we define Y0 = M0 and Y1 = M1, so that X0 = (0) and
X1 = (0).
(EP6) The map aenb (cid:55)→ apnb gives a ∗-isomorphism from Xn+1 = MnenMn to (cid:104)Mn, pn(cid:105) = MnpnMn,
the Jones basic construction of Mn−1 ⊆ Mn acting on L2(Mn, trn).
(EP7) Under the isomorphism Xn+1
∼= MnpnMn, the canonical non-normalized trace Trn+1 on the
Jones basic construction algebra MnpnMn satisfying Trn+1(apnb) = trn(ab) for a, b ∈ Mn
equals d2 trn+1 Xn+1.
(EP8) If y ∈ Yn+1 and x ∈ Xn, then yx = 0 in Mn+1. Hence En+1(Yn+1) ⊆ Yn. ("The new stuff
comes only from the old new stuff" [GdlHJ89].)
(EP9) If Yn = (0), then Yk = (0) for all k ≥ n.
Proof.
(EP1) By (M3), d2En+1(yen) = y, so the proposed map has a left inverse.
(EP2) This follows directly from (M4) and (EP1).
13
(EP3) By (M3), for x ∈ Mn, we have trn+1(xen) = trn(En+1(xen)) = trn(xEn+1(en)) = d−2 trn(x).
(EP4) By (M4), enMn+1en = enMnen. By (M2), enMnen = Mn−1en.
(EP5) That MnenMn is a 2-sided ideal is equivalent to (M4) as in Remark 3.3.
(EP6) It suffices to show the map is injective, which also shows it is well-defined. Suppose(cid:80) aipnbi =
0. Then for all a, b ∈ Mn, we have 0 = pna ((cid:80) aipnbi) bpn = (cid:80) En(aai)En(bib)pn, and
therefore(cid:80) En(aai)En(bib) = 0 as Mn (cid:51) x (cid:55)→ xpn ∈ (cid:104)Mn, pn(cid:105) is injective by (EP1) applied
for all a, b ∈ Mn, and thus(cid:80) aienbi = 0, so the map is injective.
to the Jones tower for Mn−1 ⊂ (Mn, trn), which is a Markov tower. Hence
(cid:16)(cid:88)
(cid:88)
0 =
En(aai)En(bib)en = ena
aienbi
ben
(cid:17)
(EP7) For a, b ∈ Mn, by (EP3), Trn+1(apnb) = trn(ab) = trn(ba) = d2 trn+1(baen) = d2 trn+1(aenb).
(EP8) Since X0 = (0) and X1 = (0) by definition, we may assume n ≥ 2. As in the proof of [GdlHJ89,
Thm. 4.6.3.vi], we may assume y is a central projection in Mn+1 such that yen = 0. Then
for all aen−1b ∈ Xn, by (M1), yaen−1b = d2yaen−1enen−1b = d2aen−1yenen−1b = 0. The final
claim follows from znEn+1(y) = En+1(zny) = 0 where zn is the central support of en−1 in
Mn.
(EP9) This follows immediately from (EP8).
Remark 3.5. The foregoing observations all hold in the case where the Mn are arbitrary tracial von
Neumann algebras. In this paper, we restrict our attention to the finite dimensional case because
of the following, in which we obtain a principal graph for a Markov tower. To generalize to the
infinite case, a measure-theoretic replacement for the principal graph would need to be introduced.
Notice that by (EP6), the Bratteli diagram for the inclusion Mn ⊂ Mn+1 consists of the
reflection of the Bratteli diagram for the inclusion Mn−1 ⊂ Mn, together with possibly some new
edges and vertices corresponding to simple summands of Yn+1. By (EP8), the new vertices at level
n + 1 only connect to the vertices that were new at level n. This leads to the following definition.
Definition 3.6. The principal graph of the Markov tower M• consists of the new vertices at every
level n of the Bratteli diagram, together with all the edges connecting them. A Markov tower is
said to have finite depth if the principal graph is finite.
It follows that a Markov tower has finite depth if and only if there is n ∈ N such that Yn = (0),
as in (EP9). Let M• be a Markov tower with finite depth, and take the minimal integer n ∈ N such
that Yn = (0). Notice that for k < n, the Bratteli diagram of Mk ⊆ Mk+1 contains the reflection
of the Bratteli diagram of Mk−1 ⊆ Mk, along with additional vertices and edges which are part of
the principal graph. At the base, all of the Bratteli diagram for M0 ⊆ M1 is part of the principal
graph. We can therefore 'unravel' the Bratteli diagram for Mn ⊆ Mn+1 to obtain the principal
graph for the Markov tower M•.
Fact 3.7. If a Markov tower M• has finite depth and n ∈ N is such that Yn = (0), then for k ≥ n,
there is a canonical graph isomorphism between the principal graph of M• and the Bratteli diagram
for Mk ⊆ Mk+1.
14
Definition 3.8. The principal graph Γ of a Markov tower M• has a quantum dimension function
dim : V (Γ) → R>0 given as follows. Let v ∈ V (Γ), and let p ∈ Mk be a minimal projection with k
minimal corresponding to the vertex v. We define dim(v) := dk trk(p), and we note this dimension
is independent of the choice of p ∈ Mk representing v. Moreover, the quantum dimension function
dim satisfies the Frobenius-Perron property
d · dim(v) =
dim(w)
(4)
(cid:88)
w∼v
where we write w ∼ v to mean w is connected to v, and the above sum is taken with multiplicity.
3.2 Examples of Markov towers
We discuss various examples of Markov towers in great detail.
Example 3.9. The Temperley-Lieb-Jones algebras of modulus d ≥ 2 with the usual Jones projec-
tions and Markov traces form a Markov tower with principal graph A∞.
Example 3.10. Suppose (M, m, TrM) is a cyclic pivotal T LJ (d)-module C∗ category. Let
X ∈ T LJ (d) be the generating object corresponding to the unshaded-shaded strand. As de-
scribed in the introduction, we get a Markov tower by defining Mn := EndC(m (cid:67) X alt⊗n),
trn := TrM
m(cid:67)X alt⊗n, and Jones projections depending on parity by
m(cid:67)X alt⊗n(idm(cid:67)X alt⊗n)−1 TrM
:= d−1(cid:0) idm (cid:67) id(X⊗X)⊗k ⊗(coevX ◦ coev
:= d−1(cid:0) idm (cid:67) idX ⊗ id(X⊗X)⊗k ⊗(ev
X )(cid:1) ∈ M2k+2
X ◦ evX )(cid:1) ∈ M2k+3.
†
†
e2k+1 =
2k
e2k+2 =
2k + 1
The principal graph of M• is precisely the fusion graph of M with respect to X.
Example 3.11. We obtain the equivalent connected right planar module for the subfactor planar
algebra T LJ (d)• to Example 3.10 under Theorem A as follows. We define Mk := Mk with its
†-algebra structure and faithful tracial state trk from Definition 2.8. Jones projections are defined
depending on parity by
e2k+1 := 2k
e2k+2,+ := 2k + 1
.
Lemma 3.12. Suppose P• is a finite depth subfactor planar algebra and M• is a right planar mod-
ule for P•. Then the associated Markov tower M• has finite depth, with depth(M•) ≤ depth(P•).
Proof. Let r be minimal such that Pr+1,+ = Pr,+er,+Pr,+, and let {b} be a Pimsner-Popa basis for
Pr+1,+ over Pr,+ so that(cid:80)
b ber,+b∗ = 1Pr+1,+. Since
1Mr+1 = r + 1 =
(cid:63)
r + 1 = 1Pr+1,+,
we have that {b} is a Pimsner-Popa basis for Mr+1 over Mr. Hence M• has finite depth by (EP9).
The last claim follows immediately.
15
Definition 3.13. Recall from [Pop94] that an inclusion of finite von Neumann algebras A0 ⊂ A1
with a faithful normal tracial state tr1 on A1 is called a Markov inclusion if the canonical faithful
0J = (cid:104)A1, e1(cid:105) ⊂ B(L2(A1, tr1))
normal semifinite trace on the Jones basic construction A2 = JA(cid:48)
given by the extension of xe1y (cid:55)→ tr1(xy) is finite and Tr2(1)−1 Tr2 A1 = tr1.
basis for A1 over A0, which is a finite subset {b} ⊂ A1 such that 1A2 =(cid:80)
to x = (cid:80)
[A1 : A0] := Tr2(1) =(cid:80)
Following [JP11], we call such an inclusion strongly Markov if moreover there is a Pimsner-Popa
b be1b∗. This is equivalent
b bEA0(b∗x) for all x ∈ A1, and also to A2 = A1e1A1 by [Con80, Prop. 3(b)] (see also
Given a strongly Markov inclusion A0 ⊂ (A1, tr1), its Watatani index [Wat90] is the scalar
b bb∗. We refer the reader to [Pop94, 1.1.4(c)] for other equivalent properties
for the Watatani index in the presence of a Pimsner-Popa basis. We may iterate the Jones basic
construction to get a tower of von Neumann algebras (An, trn, en+1)n≥0 with faithful tracial states
such that each inclusion An ⊂ (An+1, trn+1) is strongly Markov with index [An+1 : An] = [A1 : A0]
i ∩ Aj for i ≤ j are always finite
[JP11]. By [Wat90, Prop. 2.7.3], the relative commutants A(cid:48)
dimensional von Neumann algebras.
[Wat90]).
Notation 3.14. From this point on, we reserve the notation A• = (An, trn, en+1)n≥0 for the Jones
tower of a strongly Markov inclusion of tracial von Neumann algebras and M• = (Mn, trn, en+1)n≥0
for a Markov tower.
Example 3.15. Given a strongly Markov inclusion of tracial von Neumann algebras A0 ⊂ (A1, tr1),
its Jones tower (An, trn, en+1)n≥0 is a (possibly infinite dimensional) Markov tower, as in Remark
3.5.
Neumann algebras (A(cid:48)
is a Markov tower of finite dimensional von Neumann algebras.
Taking the relative commutant with A0, we get a Markov tower of finite dimensional von
1∩An+1, en+2)n≥0
0∩An, en+1)n≥0. Similarly, (A(cid:48)
0 ∩ An, trn A(cid:48)
1 ∩ An+1, trn+1 A(cid:48)
We now classify all connected Markov towers in terms of pointed bipartite graphs and quantum
dimension functions.
Example 3.16. Suppose (Γ, v) is a locally finite pointed bipartite graph with countably many
vertices, and dim : V (Γ) → R>0 is a quantum dimension function satisfying (4). We construct a
connected Markov tower M• by defining M0 = C, and inductively constructing each Mk as dictated
by the principal graph Γ starting at v in the usual way [GdlHJ89, JS97]. We define the trace vector
for Mk by normalizing the vector obtained from dim applied to the minimal projections appearing
at level k.
It is straightforward to check that M• has principal graph Γ with basepoint v corresponding to
1M0. Moreover, by construction, the quantum dimension function of M• is exactly dim.
Indeed, the above example can be easily generalized to the following result.
Proposition 3.17. Connected Markov towers M• are classified up to ∗-isomorphism by pointed
bipartite graphs (Γ, v) with a quantum dimension function dim : V (Γ) → R>0 satisfying (4).
3.3 Operations on Markov towers to produce new Markov towers
In this section, we describe various operations on a Markov tower M• = (Mn, trn, en+1)n≥0 which
yield new Markov towers. We begin with shifting and compressing the tower. We then study the
multistep tower. For each of these operations, we discuss how the principal graph changes.
We omit the proof of the following straightforward proposition.
16
Proposition 3.18 (Shifting a Markov tower). Suppose M• = (Mn, trn, en+1)n≥0 is a Markov tower.
For any k ≥ 1, M•+k := (Mn+k, trn+k, en+k+1)n≥0 is also a Markov tower.
Remark 3.19. Notice that shifting a Markov tower simply truncates the Bratteli diagram, and
by Fact 3.7, the principal graph is unchanged.
Given a Markov tower M•, we obtain another Markov tower by compression by a non-zero
projection p ∈ P (M0). First, for all n ≥ 0, we define a faithful trace trp
n on pMnp by
n(x) := trn(p)−1 trn(pxp).
trp
(5)
It is straightforward to verify that the unique trace-preserving conditional expectation is given by
n : pMnp → pMn−1p
Ep
Ep
n(pxp) := En(pxp) = pEn(x)p
Notice that since [en, p] = 0 for all n ∈ N, for all pxp ∈ pMnp we have
enp(pxp)enp = penxenp = pEn(x)enp = Ep
n(pxp)enp,
(6)
(7)
so the conditional expectation is implemented by enp.
Proposition 3.20. Suppose M• = (Mn, trn, en+1)n≥0 is a Markov tower of finite dimensional von
Neumann algebras and p ∈ P (M0) is a nonzero projection. Then pM p• := (pMnp, trp
n, pen+1)n≥0
is a Markov tower, where trp
n is defined as in (5).
Proof. First, it is easy to see that the projections (pen)n≥1 satisfy the Temperley-Lieb-Jones rela-
tions (M1), since [en, p] = 0 for all n ≥ 0. That pen implements the trace-preserving conditional
expectation pMnp → pMn−1p as in (M2) was shown above in (7). Using (6), this immediately
n+1(pen) = pEn+1(en) = d−2p = d−21Mn, so (M3) holds. Finally, for all n ≥ 1,
implies that Ep
pMn+1p(pen) = pMn+1enp = pMnenp = pMnpen, so we have (M4).
Remark 3.21. We can determine the Bratteli diagram and principal graph for pM p•, as follows.
If p has central support 1, then the Bratteli diagram is unchanged. In general, any vertices on the
bottom row corresponding to simple summands of M0 where p does not have support disappear,
as well as those edges no longer supported from below. By proceeding up the tower and, at each
level, removing those vertices and edges no longer supported from below, we obtain the Bratteli
diagram for pM p•.
Notation 3.22. We will make heavy use of the string diagrammatic representation of Temperley-
Lieb-Jones diagrams. Ordinarily, for subfactors and planar algebras, Kauffman diagrams [Kau87]
are drawn with strings going from bottom to top. We put the number k above or next to a strand
to denote a bundle of k parallel strands, and the label is omitted for single strands. For example,
the generators Ei = dei are represented by
Ei = i
n−i−2 .
Of particular importance will be the cabled/multi-step Jones projections from [PP88] which were
of importance in [Bis97, JP11]:
f j+k
j
:= dk(k−1)(ej+kej+k−1 ··· ej+1)(ej+k+1ej+k ··· ej+2)··· (ej+2k−1ej+2k−2 ··· ej+k)
F j+k
j
:=
j
k
k
= dkf j+k
j
.
17
We record the following relation for later use:
j = dk(k−1)(ej+kej+k+1 ··· ej+2k−2ej+2k−1) ·
f j+k
k−1
j
k−1
(8)
Now suppose we fix j ≥ 0 and k ≥ 1. For n ∈ N, define the k-cabled Jones projections gn :=
f j+nk
j+(n−1)k. It is straightforward to verify using Kauffman's diagrammatic calculus for Temperley-
Lieb-Jones algebras that the projections (gn)n∈N satisfy the Temperley-Lieb-Jones relations (M1)
with d−2 replaced with d−2k.
We now show that taking every k-th algebra in a Markov tower gives us another Markov tower.
Proposition 3.23. Suppose M• = (Mn, trn, en+1) is a Markov tower, and let j ≥ 0 and k ≥ 1.
Define gn ∈ Mj+(n+1)k as in Notation 3.22. Then Mj+k• := (Mj+nk, trj+nk, gn+1)n≥0 is a Markov
tower.
Proof. We saw Condition (M1) holds from the diagrammatic calculus, and Conditions (M2) and
(M3) are straightforward induction arguments.
We prove (M4) by strong induction on k. The base case k = 1 is exactly (M4) for the original
Markov tower. Now suppose that (M4) holds for any multi-step towers with increment less than
k. Consider the multi-step tower of algebras (Mj+nk)n≥0, which has increment k. By Proposition
3.18, we may assume j = 0. Using (8) and (M4) for the original Markov tower, we have
k−1
M(n+1)kgn = M(n+1)kf (n−1)k+k
(n−1)k = M(n+1)k(enkenk+1 ··· e(n+1)k−2e(n+1)k−1) ·
(n−1)k
k−1
= M(n+1)k−1e(n+1)k−1 ·
k−1
k−1
(n−1)k
k−1
= M(n+1)k−1
.
(n−1)k
k
Since we may perform isotopy in the Temperley-Lieb-Jones subalgebra of M(n+1)k, we may decom-
pose the diagram on the right hand side as follows:
k−1
(n−1)k
k
= dk−2
k−2
k−2
k
(n−1)k
← d · f (n−1)k+1+(k−1)
(n−1)k+1
By the induction hypothesis, we have
k−1
M(n+1)k−1
= Mnk+(k−1)
(n−1)k
k
k−2
k−2
k
= Mnk
k−2
k−2
(n−1)k
k
= Mnkgn.
(n−1)k
18
This completes the proof.
Remark 3.24. If M• is a Markov tower, then we know from Proposition 3.23 that Mk• is also a
Markov tower. The Bratteli diagram for Mk• can be read off the original Bratteli diagram quite
easily: the vertices of the level of the Bratteli diagram corresponding to Mkn are the same in both
towers, while the number of edges between two vertices in the new diagram is the number of upward
paths between those vertices in the old diagram. Note that, since a vertex of the multistep principal
graph may belong to the 'old stuff' in the original Bratteli diagram, the number of edges between
adjacent vertices in the multistep principal graph is not simply the number of paths in the original
principal graph. In the case where k is odd, taking the k-step basic construction therefore collapses
the vertices of each k levels of the principal graph into one level; when k is even, we lose the odd
part of the principal graph entirely, but aside from this, the situation is the same.
3.4 The projection category of a Markov tower
We now define the category of projections of a Markov tower.
Definition 3.25. Let M• = (Mn, trn, en+1)n≥0 be a Markov tower. We define the category M
to be the unitary Karoubi completion (formally adding orthogonal direct sums, and then taking
the orthogonal projection completion) of the C∗ category M0 with finite dimensional Hom-spaces
defined as follows.
• The objects of M0 are the symbols [n] for n ≥ 0.
• Given n, k ≥ 0, we define M0([n] → [n + 2k]) := Mn+k and M0([n + 2k] → [n]) := Mn+k.
• The identity morphism in M0([n] → [n]) is 1Mn.
• For x ∈ M0([n] → [n + 2k]) or x ∈ M0([n + 2k] → [n]), we define x† := x∗ ∈ Mn+k.
• We define composition in three cases. In each, we make use of Temperley-Lieb diagrams. The
Jones projections of a Markov tower generate an image of the Markov tower of Temperley-Lieb
algebras, so each such diagram represents a well-defined element of M•.
(C1) If x ∈ M0([n] → [n + 2i]) and y ∈ M0([n + 2i] → [n + 2i + 2j]), we define
y · x ·
(cid:124)
∈ Mn+i+j = M0([n] → [n + 2i + 2j]).
(cid:125)
n i
j
(cid:123)(cid:122)
i
∈Mn+2i+j
y ◦ x := diEn+2i+j
n+i+j
We define the composite x† ◦ y† := (y ◦ x)†, which defines composition
M0([n + 2i + 2j] → [n + 2i]) ⊗ M0([n + 2i] → [n]) → M0([n + 2i + 2j] → [n]).
To show x† ◦ y† is well-defined, we check that when i = j = 0,
x† ◦ y† = x∗y∗ = (yx)∗ = (y ◦ x)†.
(9)
19
(C2) If x ∈ M0([n] → [n + 2i + 2j]) and y ∈ M0([n + 2i + 2j] → [n + 2i]) we define
y ◦ x := diEn+2i+j
n+i
y · x ·
(cid:124)
n j i
(cid:123)(cid:122)
i
∈Mn+2i+j
∈ Mn+i = M0([n] → [n + 2i]).
(cid:125)
As above, we define the composite x† ◦ y† := (y ◦ x)†, which defines composition
M0([m] → [m + 2k]) ⊗ M0([n + 2j] → [n]) → M0([m] → [n]).
To show that x† ◦ y† is well-defined, we check that when i = 0,
x† ◦ y† = En+j
n
(x∗y∗) = En+j
n
((yx)∗) = En+j
n
(yx)∗ = (y ◦ x)†.
(10)
(C3) If x ∈ M0([n + 2i] → [n]) and y ∈ M0([n] → [n + 2i + 2j]), we define
y ◦ x := y · d−i
n j i
i
· x ∈ Mn+2i+j = M0([n + 2i] → [n + 2i + 2j]).
As above, we define the composite x† ◦ y† := (y ◦ x)†, which defines composition
M0([n + 2i + 2j] → [n]) ⊗ M0([n] → [n + 2i]) → M0([n + 2i + 2j] → [n + 2i]).
To show that x† ◦ y† is well-defined, we check that when j = 0,
x† ◦ y† = x∗ · d−i
n i
i
· y∗ =
y · d−i
∗
n i
i
· x
= (y ◦ x)†.
(11)
M0([n] → [n])
Showing that composition is associative directly from the definitions above is a highly non-trivial
exercise using the axioms (M1) -- (M4) of a Markov tower. A better way to prove associativity is
to prove that each 4 × 4 (possibly non-associative) linking algebra [GLR85]
M0([n+2i] → [n])
M0([n+2i] → [n+2i])
M0([n+2(i+j)] → [n])
M0([n+2(i+j)] → [n+2i])
L :=
M0([n] → [n+2i])
M0([n] → [n+2(i+j)]) M0([n+2i] → [n+2(i+j)]) M0([n+2(i+j)] → [n+2(i+j)]) M0([n+2(i+j+k)] → [n+2(i+j)])
M0([n] → [n+2(i+j+k)]) M0([n+2i] → [n+2(i+j+k)]) M0([n+2(i+j)] → [n+2(i+j+k)]) M0([n+2(i+j+k)] → [n+2(i+j+k)])
is †/∗-isomorphic to a von Neumann algebra, which is necessarily associative! This technique also
offers the advantage that it simultaneously proves M0 is C∗.6
M0([n+2(i+j+k)] → [n])
M0([n+2(i+j+k)] → [n+2i])
Notice we have an equality of sets
Mn
L =
Mn+i+j+k
Mn+2i+j+k
Mn+i
Mn+i+j
Mn+2i+2j+k
Mn+i+j+k Mn+2i+j+k Mn+2i+2j+k Mn+2i+2j+2k
Mn+i+j
Mn+2i+j
Mn+2i+2j
Mn+i
Mn+2i
Mn+2i+j
(12)
.
6 Just as being a C∗ algebra is a property of a complex ∗-algebra, being a C∗ category is a property of a C-linear
dagger category.
20
We define the following map entry-wise; that is, for an element x ∈ L, we plug xab into the input
disk in the ab-th entry of the map π : L → p Mat4(Mn+2i+2j+2k)p given by
n i
i j
j k k
n i
i j
j k k
n i
i j
j k k
n i
i j
j k k
d−i−j−k
d−i−j−k
d−i−j−k
n
n
i
i
i j
i j
j k k
j k k
n
i
i j
j k k
n+2i j
j k k
n
i j
j i k k
n+2i j
j
k k
n
i j k k j
i
n+2i j
j k k
n i
i j
j k k
n
i j
j i k k
d−j−k
d−j−k
d−j−k
d−j−k
n+2i
n+2i
j
j
j k k
j k k
n+2i
j
j k k
n+2i+2j k k
n+2i
j k k j
n+2i+2j
k k
(13)
d−i−j−k
d−i−j−k
d−i−j−k
.
n i
i j
j k k
n
i j k k j
i
n+2i j
j
k k
n+2i
j k k j
d−j−k
d−i−j−k
d−k
n+2i+2j
k k
n+2i+2j
k k
n+2i+2j
k k
n+2i+2j+2k
d−k
d−k
n i
i j
j k k
n+2i j
j k k
n+2i+2j
k k
n+2i+2j+2k
where p ∈ Mat4(Mn+2i+2j+2k) is the following projection:
d−i−j−k
p := diag
n i
i j
j k k
, d−j−k
n+2i j
j k k
n+2i+2j k k
, d−k
n
i
i j
j k k
n+2i
j
j k k
n+2i+2j
k k
, 1n+2i+2j+2k
In Proposition 3.26 below, we verify the map π is an injective unital algebra map satisfying π(x†) =
π(x)∗, and is thus an isomorphism onto its image. Thus im(π) is a unital ∗-subalgebra of the finite
dimensional von Neumann algebra p Mat4(Mn+2i+2j+2k)p, which means im(π) is a von Neumann
algebra by the finite dimensional bicommutant theorem [Jon15, Thm. 3.2.1]. By looking at the
2 × 2 and 3 × 3 corners of the linking algebra L and (13), we immediately see:
• † is a dagger structure on M0,
• for every f ∈ M0([n] → [n + 2j]), there is a g ∈ M0([n] → [n]) and an h ∈ M0([n + 2j] →
[n + 2j]) such that f† ◦ f = g† ◦ g and f ◦ f† = h† ◦ h, and
• there are (pullback) norms on the (finite dimensional) Hom-spaces M0([n] → [n + 2i]) and
M0([n+2j] → [n]) which are sub-multiplicative with respect to composition and which satisfy
the C∗ axiom (cid:107)f† ◦ f(cid:107) = (cid:107)f 2(cid:107).7
Hence M0 is C∗, and thus so is its unitary Karoubi completion M.
Proposition 3.26. The map π : L → p Mat4(Mn+2i+2j+2k)p from Definition 3.25 is an injective
unital algebra map such that π(x†) = π(x)∗ for all x ∈ L.
We begin with the following lemma.
7 Notice that in a C∗ category, these norms can be recovered from spectral radii together with the positivity and
C∗ axioms. Thus these norms are not part of the data of the C∗ category.
21
Lemma 3.27. For all x ∈ Mn+2i+j, diEn+j+2i
n+j+i
Proof. We calculate
x ·
i
n i
j
j
·
j
j i
i
n
i
j
= x ·
i
n i
j
j
.
n i
j
j
j i
·
n
i
j
n i
x ·
· x · d−i
i
i
i
i
i
i
n i j
n i j
n i j
di En+j+2i
x ·
x ·
x ·
n+j+i
n i
n i
n i
i
i
· x ·
· x ·
n
i
i
i
j
j
j
n i
n i
n i
i ·
i ·
·
x ·
· En+j+3i
· d−i
·
n+j+2i
n j
j i
j
j
j
i
i
.
= diEn+j+3i
n+j+2i
= En+j+3i
n+j+2i
= En+j+3i
n+j+2i
= En+j+3i
n+j+2i
= diEn+j+2i
n+j+i
= diEn+j+2i
n+j+i
= diEn+j+2i
n+j+i
j i
i
j
j i
n
i
j
n j
i
i
i
n j
i
i
·
·
j i
n
i
j
j i
n
i
j
i j
·
i
j i
n
i
j
i
n
i
j
Proof of Proposition 3.26. By inspection of the definition of π from (13), it is clear that π is injec-
tive, unital, C-linear, and respects the †-structure. The difficulty is in seeing that π is an algebra
homomorphism. In the following, we suppress the rightmost 2k strings of entries of (13), as well
as the factor d−k, since they are essentially inert when only three objects are considered. Because
π respects †, as in Definition 3.25, we only need to consider 3 cases of composition.
22
Case 1: Let x : [n] → [n + 2i] and y : [n + 2i] → [n + 2i + 2j]. Then
π(y ◦ x) = d−jEn+2i+j
n+i+j
yx ·
n i
i
i j
i
=
(Lem. 3.27)
d−i−jyx ·
j
n
i
i
j
j
= π(y) · π(x).
Case 2: Let x : [n] → [n + 2i + 2j] and y : [n + 2i + 2j] → [n + 2i]. Then
j
n
·
yx ·
En+2i+j
n+2i
i
(cid:124)
n j i
·
yx ·
(cid:123)(cid:122)
yx ·
z
n
i
i
j
j
n j i
i
n j i
n
·
(cid:125)
·
n
i
i
j
j
·
i
i
n
i
i
j
j
π(y ◦ x) = d−jEn+2i+j
n+i
= d−i−jEn+2i
n+i
d−2i−jEn+2i+j
n+2i
=
(Lem. 3.27 for
z with j = 0)
=
(Prop. 3.23)
d−2i−2j
n i i
j
· yx ·
j
i
n i i
j
= d−i−2j
n
i j
· yx ·
j
i
j
= π(y) · π(x).
i
n j i
j
·
n
i
i
j
j
π(y ◦ x) = d−j
n j i
· x
n i i
j
j
= d−i−jy ·
n j i
j
n i i
j
·
· x
i
j
Case 3: Let x : [n + 2i] → [n] and y : [n] → [n + 2i + 2j]. Then
y · d−i
= d−i−jy ·
i
n
i j
i
j
·
· x = π(y) · π(x).
Corollary 3.28. Let M• be a Markov sequence and let M be its unitary Karoubi completion. Then
M is semisimple, and the isomorphism classes of simple objects are in canonical bijection with the
vertices of the principal graph.
Proof. All endomorphism algebras of M0 are finite dimensional C∗ algebras which are semisimple,
and thus M is semisimple. By (EP6), every minimal projection in Xn+2 for n ≥ 0 is equivalent
to a minimal projection in Mn via a partial isometry in M0([n] → [n + 2]). Explicitly, p ∈ Mn is
equivalent to pen+1 ∈ Mn+2 via the morphism p ∈ Mn+1 = M0([n] → [n + 2]), which is a partial
23
isometry using the definition of composition (C2) and (C3) in M0, and this exhausts all equivalence
classes of minimal projections in Xn+2. By recursion, we see that the equivalence classes of minimal
projections in M• are in canonical bijective correspondence with the minimal projections in the
(Yn)n≥0, which are exactly the vertices of the principal graph.
Remark 3.29. By Remark 3.19, the category of projections of a Markov tower is invariant (up to
equivalence) under applying shifts. By Remark 3.21, the category of projections of the compression
pM p• is the subcategory of the category of projections of M• generated by minimal projections
under p. In particular, if p (cid:54)= 0 and the Bratteli diagram of M• is connected, then the two categories
of projections are again equivalent. Finally, in the case of the multistep tower, we see by Remark
3.24 that the category of projections of Mk• is equivalent to the category of projections for M•
when k is odd, and to the subcategory generated by the even part when k is even.
3.5 Temperley-Lieb-Jones module categories
We now show that the category of projections of a connected Markov tower of modulus d can be
canonically endowed with the structure of a cyclic pivotal right Temperley-Lieb-Jones (T LJ (d))
module C∗ category. Moreover, all cyclic pivotal right T LJ (d)-module C∗ categories arise in this
way. Combined with the classification of connected Markov towers of modulus d from Proposition
3.17, we get the following result, which should be compared with [DCY15] in the non-pivotal setting.
Corollary (Corollary B). Cyclic pivotal right module C∗ categories for T LJ (d) are classified by
triples (Γ, dim, v0) where Γ = (V+, V−, E) is a bipartite graph, v0 is a distinguished vertex, and
dim : V+ (cid:113) V− → R>0 is a function satisfying dim(v0) = 1 and
(cid:88)
w∼v
dim(w) = d dim(v),
where we write w ∼ v to mean w is connected to v, and the above sum is taken with multiplicity.
Definition 3.30. If M• is a Markov tower with modulus d, the corresponding T LJ (d)-module is
just the category M of projections of M•, as in Definition 3.25. The T LJ action on M comes
from the fact that T LJ• is the initial Markov tower for a given parameter. By construction, to
define an action of T LJ on M, it suffices to define the action of the objects [m,±]T LJ on 1[n]M
for every n and m, and then define the action functorially on morphisms.
We set [m]M (cid:67) [n]T LJ := [m + n]M. For morphisms, we first consider the case where one
If g : [a]T LJ → [b]T LJ , we first add n strings to the left to obtain
morphism is the identity.
1n ⊗T LJ g ∈ T LJ ([a + n]T LJ → [b + n]T LJ ), and we define 1Mn
(cid:67) g ∈ M([a + n]M → [b + n]M)
to be the image of the element 1n ⊗T LJ g in Mn+(a+b)/2. The case of f (cid:67) 1 is more complicated.
If f ∈ M([n] → [n + 2k]) = Mn+k, then we define
f ·
f ·
n
k
k
k
n
k
j − k
if k ≥ j
k − j
if j ≤ k.
f (cid:67) 1j :=
Notice these formulas agree when j = k. For f ∈ M([n] → [m]) where m < n, we define
f (cid:67) 1 := (f† (cid:67) 1)†.
24
Next, we check that (f (cid:67) 1) ◦ (1 (cid:67) g) = (1 (cid:67) g) ◦ (f (cid:67) 1). We illustrate the case f : [n]M →
[n + 2k]M and g : [m]T LJ → [m + 2(cid:96)]T LJ , where both m ≥ k and (cid:96) ≥ k.
In other words,
g ∈ [k + j]T LJ → [k + j + 2(k + i)]T LJ for some non-negative i and j. The other cases are similar.
Let ι denote the inclusion in M•. In the following, since g ∈ T LJ , we represent g by a ticket
within Temperley-Lieb tangles, which we may freely move via isotopy. By definition, we have
(f (cid:67) 12i+j+3k) ◦ (1n (cid:67) g) =
(C1)
di+kEn+j+4k+2i
n+j+3k+i
k i
k k
i
nk
j k i k i
k
g
n j k
k i
k k
i
nk k j k i k
i
g
nk k j k i k i
k
g
n j k
n+j+4k+i
ι(f ) ·
ι(f ) · En+j+4k+2i
ι(f ) · En+j+4k+2i
nk k j k i k
nk k j k i k
n+j+4k+i
g
g
· ι(f ) · d−i
i
k
n j k
k i
k k
i
n j k
k i
k k
i
nk k j k k i
· ι(f ) · En+j+4k+2i
n+j+4k+i
nk k j k i k
n j k i k k k
= di+kEn+j+4k+i
n+j+3k+i
= di+kEn+j+4k+i
n+j+3k+i
= di+kEn+j+4k+i
n+j+3k+i
= di+kEn+j+4k+i
n+j+3k+i
(1 (cid:67) g) ◦ (f (cid:67) 1).
=
(C1)
To show that (cid:67) is a well-defined bifunctor, it remains to show that for morphisms f and g in M
and h and k in T LJ , we have (g (cid:67) 1) ◦ (f (cid:67) 1) = (g ◦ f ) (cid:67) 1 and (1 (cid:67) k) ◦ (1 (cid:67) h) = 1 (cid:67)
(k ◦ h). Functoriality in the right variable comes directly from the definition, but functoriality in
the left variable is more involved and done in cases. We illustrate a representative case. Suppose
25
f : [n]M → [n + 2i + 2j]M and g : [n + 2i + 2j]M → [n + 2i]M, and set m := k + i + j. Then
(g (cid:67) 1m) ◦ (f (cid:67) 1m) = di+jEn+3i+k+2j
n+2i+k+j
· ι(g) · ι(f ) ·
= d−iEn+3i+k+2j
n+2i+k+j
· ι(g) · ι(f ) ·
= d−i+jEn+3i+k+2j
· ι
n i i
k i j
j
j
j
k i j
j
j
n i i
i
i
j
j
j
j k i
k i j
n i i
n+2i
n j i
n i i j
n+2i+k+j
g · ι(f ) ·
En+2i+j
En+2i+j
g · ι(f ) ·
En+2i+j
g · ι(f ) ·
(cid:123)(cid:122)
En+2i+j
En+2i
g · ι(f ) ·
En+2i+j
d−2iι
(cid:124)
n j i
n j i
n j i
n+2i
n+2i
n+2i
n+i
n+i
z
i
i
i
i
i
i
i
i
i
j
j
j
k
n
i i
n j
k i j
n+2i
i j k i
n j i
nk i j i
g · ι(f ) ·
En+2i+j
· En+3i+k+2j
·
·
(cid:125)
g · ι(f ) ·
·
·
n+2i+k+j
n k j i
n j i
i k j
i k j
i k j
n i
n i
n i
i
i
i
i
i
i
i
i
i
i
= d−i+jι
= d−2iι
= d−3iι
=
(Lem. 3.27 for
z with j = 0)
= d−iι
i
j
k i j
i
i
j
n
·
i
n j i
i
i
n i
i k i j
j
j
i
i
k i j
i
i
j
n
i
·
n
i
i
i
n i
i k j
i
n k j i
i
= (g ◦ f ) (cid:67) 1m.
i
i
i
i
26
It is easier to check that the action is associative. Since we have already shown that (cid:67) is a
bifunctor, it suffices to check associativity of triples of morphisms when two are the identity. That
1 (cid:67) (1 ⊗ h) = (1 (cid:67) 1) (cid:67) h and that 1 (cid:67) (g ⊗ 1) = (1 (cid:67) g) (cid:67) 1 for all g, h ∈ T LJ follow directly
from the definitions of (cid:67) and ⊗. Finally, in case f ∈ M([n] → [n + 2i]), we have
f (cid:67) (1[j+i] ⊗ 1[k+i]) = f (cid:67) 1[j+k+2i] = ι(f )·
The other cases are similar.
n i
i j i k
= ι(f )·
i
i
n i
i j
i k
n j i
k i
i
= (f (cid:67) 1[j+i]) (cid:67) 1[k+i]
Remark 3.31. Notice that the principal graph of M• is exactly the fusion graph for the associated
T LJ -module category M with respect to the unshaded-shaded strand X ∈ T LJ with basepoint
the simple projection 1M0 ∈ M. By Remark 3.29, the operation of shifting the Markov tower
does not change M (up to equivalence), but corresponds to replacing the basepoint 1M0 with
[0] (cid:67) X alt⊗2n. Similarly, compressing by a minimal projection p ∈ Mn corresponds to moving
the basepoint to p. In contrast, the multistep basic construction, which may affect the principal
graph, is analagous to replacing X ∈ T LJ with X alt⊗2n, without changing the basepoint of M.
The T LJ -module structure on the category of projections M(n) of Mn• comes from combining
the action of the subcategory T LJ (n) of T LJ generated by subobjects of [kn]T LJ and the pivotal
T LJ -right module structure of T LJ (n).
Remark 3.32. Observe that we may identify the tensor category T LJ as the category of projec-
tions of the Markov tower T LJ•, where the tensor structure is given by the T LJ -module structure
from Definition 3.30. One should think of the definitions for composition and tensor product in the
category of projections as being obtained by isotoping the much simpler definitions for a planar
algebra into a form that can be written down in terms of the data of a Markov tower. Notice
that under this identification, ev[n] ∈ T LJ ([2n] → [0]) and coev[n] ∈ T LJ ([0] → [2n]) are both
identified with 1n ∈ T LJn under Definition 3.25; one then checks they satisfy the zig-zag axioms
using the definitions of (cid:67) from Definition 3.30 and composition from Definition 3.25.
Definition 3.33 (Pivotal module structure). The category M obtains a unitary trace TrM from
the Markov tower (M•, tr•), by renormalizing so that isomorphic projections have the same trace.
For [n] ∈ M, we define TrM
[n] satisfies
(Tr2) for all n ≥ 0. Now if x ∈ M([n] → [n + 2k]) = Mn+k and y ∈ M([n + 2k] → [n]) = Mn+k,
observe that
[n] : M([n] → [n]) = Mn → C by dn trn. It is clear that TrM
x · dk
n k
k
· y
TrM
[n](y ◦ x) :=
(C2)
dn trn(En+k
n
(yx)) = dn trn+k(yx) = dn+k · trn+k
TrM
[n+2k](x ◦ y),
=:
(C3)
and thus TrM satisfies (Tr1). To see that TrM satisfies (Tr3), we must show that for all n, k, and
f ∈ M([n] → [n]),
TrM
[n](cid:67)[k](f ) = TrM
[n]
(1 (cid:67) coev
(cid:16)
(cid:17)
†
[k]) ◦ (f (cid:67) 1[k]) ◦ (1 (cid:67) coev[k])
.
27
f · En+2k
n
dk
nk
k
= Tr
En+2k
n
ι(f ) · dk
.
nk
k
Now by Remark 3.32, it is straightforward to show the left and right sides of the above equation
are respectively equal to the left and right sides of the following equation, which trivially holds:
Tr(f ) = Tr
Proof of Corollary B. We saw in Definitions 3.30 and 3.33 how a connected Markov tower of modu-
lus d gives us a cyclic pivotal right T LJ (d)-module C∗ category. We saw in Example 3.10 that given
a cyclic pivotal right T LJ (d)-module C∗ category (M, m, TrM), defining Mn := End(m (cid:67) X alt⊗n)
and trn := TrM
m(cid:67)X alt⊗n defines a connected Markov tower. One now shows these two processes are
mutually inverse up to dagger equivalence.
Remark 3.34. While the process of defining a tensor structure on the category of projections
P of a Markov tower P• obtained from a planar algebra P• is fairly straightforward and similar
to Remark 3.32, it is far less obvious for a Markov tower coming from a standard λ-lattice as in
[Pop95]. Given a standard λ-lattice A•• with λ = d−2, we expect that a Markov tower M• of
modulus d such that Mn ⊃ A0n for all n ≥ 0 satisfying certain compatibility conditions is the
equivalent notion of a right module for A•• in the spirit of Theorems A and 2.5.
M0 ⊂ M1 ⊂ M2 ⊂ M3 ⊂ ···
∪
∪
∪
∪
A00⊂ A01⊂ A02⊂ A03⊂ ···
∪
∪
∪
A11⊂ A12⊂ A13⊂ ···
∪
∪
A22⊂ A23⊂ ···
∪
A33⊂ ···
. . .
We leave this exploration to a future joint article, as it would take us too far afield.
4 The canonical planar algebra from a strongly Markov inclusion
We begin this section by recalling the construction of the canonical planar †-algebra from a strongly
Markov inclusion. The reader is advised to review the definition of a strongly Markov inclusion
from Definition 3.13 before proceeding. We then discuss various operations on the inclusion, and
how such operations affect (or do not affect!) the planar algebra.
4.1 The canonical relative commutant planar algebra
There is a canonical planar algebra structure on the towers of relative commutants, called the
canonical planar †-algebra of a strongly Markov inclusion corresponding to A0 ⊆ (A1, tr1). Denote
by (An, trn)n≥0 the Jones tower for inclusion A0 ⊂ (A1, tr1). The box spaces are defined by the
relative commutants
Pn,+ := A(cid:48)
0 ∩ An
Pn,− := A(cid:48)
1 ∩ An+1,
28
which are finite dimensional by [Wat90, Prop. 2.7.3]. We refer the reader to [JP11, §2.3] for the
action of tangles. The †-structure is given by ∗ in the relative commutants. We remark that one
important feature of this construction is that it depends on the existence of a Pimsner-Popa basis
for A1 over A0, but not on a choice of basis.
The following theorem uniquely characterizes the canonical relative commutant planar †-algebra.
Theorem 4.1 ([JP11, Thm. 2.50]). Given a strongly Markov inclusion A0 ⊂ (A1, tr1), there is a
unique planar †-algebra P• of modulus d = [A1 : A0]1/2 whose box spaces are given by
Pn,+ := A(cid:48)
0 ∩ An
Pn,− := A(cid:48)
1 ∩ An+1,
such that
(PA1) The †-structure of Pn,± is given by x† = x∗ in the relative commutant, and stacking corre-
sponds to multiplication in the relative commutant:
= xy ∈ Pn,±.
n
x
n
y
n
(PA2) The Jones projection en ∈ A(cid:48)
0 ∩ An for the strongly Markov inclusion An−1 ⊂ An is given by
the following tangles depending on parity for k ≥ 0:
e2k+1,+ := 2k
e2k+2,+ := 2k + 1
.
(PA3) For x ∈ Pn,+ = A(cid:48)
0 ∩ An and {b} a Pimsner-Popa basis for A1 over A0,
• n
• n − 1
n
n − 1
x = x ∈ Pn+1,+ = A(cid:48)
0 ∩ An+1,
x = dEAn
An−1
(x) ∈ Pn−1,+ = A(cid:48)
0 ∩ An−1, and
•
n − 1
x
n − 1
= dEA(cid:48)
A(cid:48)
0
1
(x) = d−1(cid:80)
b bxb∗ ∈ Pn,− = A(cid:48)
1 ∩ An.
(PA4) For x ∈ Pn,− = A(cid:48)
1 ∩ An+1,
n
x = x ∈ Pn+1,+ = A(cid:48)
0 ∩ An+1.
n
We will proceed with the same general technique for defining the embedding of planar algebras
as in [JP11], in that we will also initially embed into the canonical †-planar algebra, and then make
use of the following theorem:
Theorem 4.2 ([JP11, Theorem 3.28]). The canonical planar †-algebra associated to the strongly
Markov inclusion of finite dimensional von Neumann algebras A0 ⊆ (A1, tr1) is isomorphic to the
bipartite graph planar †-algebra of the Bratteli diagram for the inclusion A0 ⊆ A1.
The following two subsections describe two isomorphisms between canonical †-planar algebras of
related strongly Markov inclusions. Both are well known to experts, and will motivate constructions
detailed in the later sections of this article.
29
4.2 The shift isomorphism
In the rest of this article, we will make extensive use of the following lemma adapted from [JP11,
Lem. 2.49], which provides sufficient conditions for a collection of maps to be a morphism of shaded
planar †-algebras.
Lemma 4.3 ([JP11, Variation of Lem. 2.49]). Suppose Φn,± : Pn,± → Qn,± is a collection of unital
†-algebra maps such that
(1) Φ maps Jones projections in Pn,+ to Jones projections in Qn,+,
(2) Φ commutes with the action of the following tangles:
n
n
n
n
n − 1
n − 1
right inclusion
Pn,+→Pn+1,+
left inclusion
Pn,−→Pn+1,+
right capping
Pn,+→Pn−1,+
n − 1
n − 1
left capping
Pn,+→Pn−1,−
.
Then Φ is a morphism of shaded planar †-algebras.
Suppose that A0 ⊂ A1 ⊂ (A2, tr2, e1) is a strongly Markov inclusion of von Neumann algebras
and (An, trn, en)n≥0 is the tower obtained by iterating the basic construction. We know from [JP11,
Cor. 2.18] that for any 0 ≤ k ≤ n, the inclusion Ak ⊂ An is strongly Markov. Thus, we can find a
Pimsner-Popa basis B for An over Ak.
By [JP11, Prop. 2.20], for 0 ≤ j ≤ 2n, we can represent Aj on L2(An, trn) via the multistep
j∩B(L2(An, trn)) where Jn is the modular conjugation. We
basic construction, and JnA2n−jJn = A(cid:48)
get a canonical trace on A(cid:48)
j(x) := tr2n−j(Jnx∗Jn) as discussed in Remark [JP11, Rem. 2.21].
Proposition 4.4 ([JP11, Prop. 2.24]). Let 0 ≤ j ≤ k ≤ n. Let {b} be a Pimsner-Popa basis for Ak
k ∩ B(L2(An, trn)), tr(cid:48)
k)
over Aj. The conditional expectation E
is given by:
j ∩ B(L2(An, trn)), tr(cid:48)
j) → (A(cid:48)
j by tr(cid:48)
A(cid:48)
A(cid:48)
j
k
: (A(cid:48)
(x) = d−2(k−j)(cid:88)
bxb∗
b
A(cid:48)
A(cid:48)
j
E
k
and is independent of the choice of basis.
From the definition of the canonical †-planar algebra P• of A•, one can work out that (in
the language of Lemma 4.3), right capping is simply the conditional expectation in the Markov
tower A•, while left capping is the expectation on relative commutants described in the previous
proposition. Thus, left capping is how the Pimsner-Popa basis shows itself graphically. The full
details can be found in [JP11, Prop. 2.47].
Corollary 4.5. Adding k strings to the left of x gives a unital ∗-algebra isomorphisms Pn,± →
Pn+k,±(cid:48) where ±(cid:48) = ± if k is even and ∓ if k is odd.
(cid:55)→ x
nk
x
n
30
Proof. We first prove the result for Pn,+ = A(cid:48)
0 ∩ An. The following implicitly uses a trick due to
Vaughan Jones that can be found in [JP11, Theorem 4.1] along with the pictorial description of
f n
n−k in the Multistep Basic Construction [JP11, Remark 2.44] in the equality marked (!). Let B
be a Pimsner-Popa basis for Ak over A0. For x ∈ A(cid:48)
0 ∩ An is uniquely
determined by
k ∩ An+k, the element y ∈ A(cid:48)
(x) = d−2k(cid:88)
x = EA(cid:48)
A(cid:48)
0
k
bxb∗ =
(!)
d−k
xkk
=
k
d−k
xk
=
y
.
n
nk
b∈B
The proof is similar for Pn,−.
Recall that the canonical planar †-algebra for the inclusion A0 ⊆ (A1, tr1) is denoted by P•.
We denote the canonical planar †-algebra for A2 ⊂ (A3, tr3) by Q•.
Theorem 4.6 (Shift Isomorphism). The map Φ : P• → Q•, obtained by adding two strings in front
of elements of Pn,±
(cid:55)→ x
n2
x
n
(cid:55)→ x
n2
x
n
defines a planar †-algebra isomorphism between P• and Q•.
Proof. The map is an isomorphism between box spaces due to Corollary 4.5. In order to show that
this map commutes with the action of tangles, we just have to show that it satisfies the requirements
of Lemma 4.3. We draw the string diagrams of Q• in blue in order to increase clarity.
(1) (Right Inclusion) Φ(x)
=
Φ(x)
=
x
n
2 n
2
n
(2) (Left Inclusion) Φ(x)
=
Φ(x)
=
x
n
n2
n2
(3) (Right Capping) Φ(x)
=
Φ(x)
=
x
n
2 n
2
n
(4) (Left Capping) We again apply the trick from [JP11, Theorem 4.1]. Let B be a Pimsner-Popa
basis of A3 over A2. Then
Φ(x)
n
= d−1(cid:88)
b∈B
bΦ(x)b∗ =
x
.
2
n
Remark 4.7. By Remark 3.19, the categories of projections of P• and Q• as in Theorem 4.6 are
equivalent.
31
4.3 The compression isomorphism
The following lemma is well known to experts. We provide a proof for convenience and completeness.
Lemma 4.8. Suppose N ⊂ M ⊂ B(H) is an inclusion of von Neumann algebras and p ∈ P (N ).
(1) p(N(cid:48) ∩ M ) = pN(cid:48) ∩ pM p.
(2) Suppose the central support of p in N is z ∈ Z(N ). The map x (cid:55)→ px is an isomorphism
z(N(cid:48) ∩ M ) → p(N(cid:48) ∩ M ).
Proof.
Proof of (1): The proof of (1) is similar to the proof of the standard fact that (pN p)(cid:48) = N(cid:48)p.
Clearly (N(cid:48) ∩ M )p ⊆ (N(cid:48)p) ∩ pM p. Suppose u is a unitary in (N(cid:48)p) ∩ pM p. Let K be the
closure of N pH. Let q ∈ B(H) be the projection onto K, which is clearly in N(cid:48) ∩ N = Z(N ).
Define u0 in B(K) by u0(npξ) := npuξ. One now verifies that u0 is an isometry and thus is well-
defined. Look at the operator u0q ∈ N(cid:48) ∩ B(H), and note that u = u0qp ∈ N(cid:48)p. We claim that
u0q ∈ M , so that u = u0qp ∈ (N(cid:48) ∩ M )p. First, for any m ∈ M(cid:48), n ∈ N , and ξ ∈ H, we have
mu0npξ = mnupξ = nupmξ = u0npmξ = u0mnpξ. Thus u0 ∈ qM q. Since q ∈ M , for all m ∈ M(cid:48),
we have u0qmξ = u0mqξ = mu0qξ. Hence u0q commutes with M(cid:48) on H, and u0q ∈ M .
Proof of (2): For x ∈ N(cid:48) ∩ M , we have p(zx) = px. Hence the map is surjective. We now show the
map is injective. Suppose x ∈ z(N(cid:48) ∩ M ) such that px = 0. By (1), z(N(cid:48) ∩ M ) = zN(cid:48) ∩ zM z.
Then for all unitary u ∈ U (N ), upz = z(up) ∈ zN , so 0 = upxu∗ = (upz)xu∗ = x(upzu∗) = xupu∗.
Taking sup over u ∈ U (N ) yields 0 = xz = x.
Similar to the discussion in §3.3, given an inclusion of tracial von Neumann algebras A0 ⊂
(A1, tr1), we obtain another inclusion of tracial von Neumann algebras by compression by a non-
zero projection p ∈ P (A0). We define a faithful trace trp
1 on pA1p by
1(x) := tr1(p)−1 tr1(pxp).
trp
(14)
It is straightforward to verify that the unique trace-preserving conditional expectation is given by
1 : pA1p → pA0p
Ep
Ep
1 (pxp) := E1(pxp) = pE1(x)p
Notice that since [e1, p] = 0, we have for all pxp ∈ pA1p, we have
e1p(pxp)e1p = pe1xe1p = pE1(x)e1p = Ep
1 (pxp)e1p,
(15)
(16)
so the conditional expectation is implemented by e1p.
Suppose now that A0 ⊂ (A1, tr1) is strongly Markov. We would like to show that pA2p with
2(x) := tr2(p)−1 tr2(pxp) and Jones projection pe1 is isomorphic to the basic construction
trace trp
of pA0p ⊂ (pA1p, trp
1), but we will need an extra assumption on p. (This extra assumption will
be automatic when A0 ⊂ A1 is a II1 subfactor; see also [Bis94, Lem. 2.4].) Toward this goal, we
recall the following recognition lemma based on [PP88, Prop. 1.2], [Jol90, Lem. 5.8], and [JS97,
Lem. 5.3.1].
Lemma 4.9 ([JP11, Lem. 2.15]). Suppose A0 ⊂ (A1, tr1) is a strongly Markov inclusion of tracial
von Neumann algebras, and (B, trB, p) is a tracial von Neumann algebra containing A1 together
with a projection p ∈ P (B) such that
(x)p for all x ∈ A1,
(R1) pxp = EA1
A0
32
(p) = [A1 : A0]−11A1, and
(R2) EB
A1
(R3) B is algebraically spanned by A1 and p, i.e., B = A1pA1 := span{apba, b ∈ A1}.
Then the map A2 → B given by ae1b (cid:55)→ apb is a (normal) unital ∗-isomorphism of von Neumann
algebras.
In this case, where (B, trB, p) is isomorphic to the basic construction A2 of A0 ⊆ (A1, tr1), we
call the inclusion A0 ⊆ (A1, tr1) ⊆ (B, trB, p) standard, after [JP11].
We now prove that compression by well-behaved projections of A0 preserves the strong Markov
structure. Here, 'well-behaved' is the condition A0 = A0pA0 := span{apba, b ∈ A0}, which implies
that the central support of p in A0 is 1.
Proposition 4.10. Suppose A0 ⊂ (A1, tr1) is a strongly Markov inclusion of tracial von Neumann
algebras, and p ∈ P (A0) is a projection such that A0pA0 = A0.
(1) Let {a} ⊂ A0 be a finite set such that(cid:80)
a apa∗ = 1A0.8 Then for any Pimsner-Popa basis {b}
for A1 over A0, {pbap} is a Pimsner-Popa basis for pA1p over pA0p.
(2) The inclusion pA0p ⊂ (pA1p, trp
(3) The inclusion pA0p ⊂ (pA1p, trp
Proof.
Proof of (1): For all pxp ∈ pA1p, we have
2, pe1)9 is standard.
1) is strongly Markov with index [pA1p : pA0p] = [A1 : A0].
1) ⊂ (pA2p, trp
(cid:88)
(cid:88)
(cid:88)
pbapa∗E1(b∗px)p =
pbE1(b∗px)p = pxp.
(cid:88)
pbapEp
1 (pa∗b∗p · pxp) =
pbap
b
a
b
Proof of (2): Given that there exists a Pimsner-Popa basis for pA1p over pA0p by part (1), the
inclusion is Markov if and only if the Watatani index [Wat90] is a scalar by [Pop94, 1.1.4(c)]. We
now calculate(cid:88)
(cid:88)
(cid:88)
(cid:88)
pbap(pbap)∗ =
pbapa∗b∗p =
pbb∗p = [A1 : A0]p = [A1 : A0] idpA1p .
pbap
b
a
b
Proof of (3): We show the hypotheses of Lemma 4.9 hold. We already saw (R1) holds in (16). To
see (R2) holds, note that Ep
2 (pxp) = pE2(x)p for all x ∈ A2 as in (15). Thus by part (2),
2 (pe1) = pE2(e1) = [A1 : A0]−1p = [pA1p : pA0p]−11pA1p.
Ep
Finally, (R3) follows immediately from the existence of a Pimsner-Popa basis for pA1p over pA0p
by part (1).
By iterating Lemma 4.9 and Proposition 4.10, we immediately obtain the following.
Corollary 4.11. Assume the hypotheses of Proposition 4.10, and let A• = (An, trn, en+1)n≥0 be
the Jones tower for A0 ⊂ (A1, tr1). Then pAp• := (pAnp, trp
n, pen+1)n≥0 is isomorphic to the Jones
tower of pA0p ⊂ (pA1p, trp
1).
8 Such a finite set necessarily exists by the same trick used in [Con80, Prop. 3(b)].
9 The definition of trp
2 is analogous to (14).
33
Suppose A0 ⊂ (A1, tr1) is a strongly Markov inclusion of tracial von Neumann algebras. Denote
by P• the canonical planar †-algebra whose box spaces are given by
Pn,+ := A(cid:48)
0 ∩ An
Pn,− := A(cid:48)
1 ∩ An+1.
Suppose p ∈ P (A0) is a projection such that A0pA0 = A0. By Corollary 4.11, the Jones tower
for pA0p ⊂ (pA1p, trp
1) is given by (pAnp, trp
n, pen+1)n≥0, and thus we get another canonical planar
†-algebra Q• whose box spaces are given by
0 ∩ pAnp
1 ∩ pAn+1p.
Qn,− := pA(cid:48)
Qn,+ := pA(cid:48)
By Lemma 4.8, the map Φn,± : x (cid:55)→ xp gives an isomorphism of von Neumann algebras Φn,± :
Pn,± → Qn,± for each n ≥ 0.
Theorem 4.12. The maps Φn,± : Pn,± → Qn,± constitute a planar †-algebra isomorphism.
Proof. We prove the unital ∗-algebra isomorphisms Φn,± satisfy the conditions of Lemma 4.3.
First, note that Φn,±(en) = pen, so Jones projections in P• map to Jones projections in Q• by
Corollary 4.11. Hence Condition (1) of Lemma 4.3 is satisfied.
The only interesting part in checking Condition (2) of Lemma 4.3 holds is verifying that left
capping commutes with Φn,±. First, by Proposition 4.4, if {b} is a Pimsner-Popa basis for A1 over
A0, then for all x ∈ Pn,+ = A(cid:48)
0 ∩ An,
= d−1(cid:88)
β
n − 1
x
n − 1
bxb∗.
This means that picking Pimsner-Popa bases {b} for A1 over A0 and {a} for pA1p over pA0p, we
must show that
n − 1
x
n − 1
= d−1p
(cid:88)
bxb∗ ?= d−1(cid:88)
b
a
Φn−1,−
apxa∗ =
n − 1
Φn,+(x)
n − 1
.
(17)
The trick is to carefully choose our Pimsner-Popa basis for A1 over A0. We take the Pimsner-Popa
basis {a} for pA1p over pA0p and we take the disjoint union with {(1 − p)b}, where {b} was our
Pimsner-Popa basis for A1 over A0. We now claim {c} = {a} ∪ {(1 − p)b} is a Pimsner-Popa basis
for A1 over A0. Indeed, since a = pap ∈ pA1p for all a ∈ {a}, we have
ce1c∗ =
ape1a∗ +
(cid:88)
b
(cid:88)
(cid:32)(cid:88)
a
(cid:88)
c
(cid:88)
p
(1 − p)be1b∗(1 − p) = p + (1 − p) = 1A2.
(cid:88)
(cid:88)
(1 − p)bxb∗(1 − p)
apxa∗.
(cid:33)
=
Thus for this special choice of Pimsner-Popa basis for A1 over A0, we immediately obtain
cxc∗ = p
a(px)a∗ +
c
a
b
a
Hence (17) holds, and the result follows.
34
5 The module embedding theorem via towers of algebras
We have finally developed the tools necessary to prove Theorem C, which turns a finite depth cyclic
right pivotal module (M, m) over the category of projections C of a subfactor planar algebra Q•
(or equivalently, a finite depth connected right planar module over Q•) into an embedding of Q•
into the graph planar algebra of the fusion graph of M with respect to the unshaded-shaded strand
X ∈ Q1,+. As a special case, we recover the embedding of a subfactor planar algebra into the graph
planar algebra of its own principal graph, described in [JP11], along with an embedding into the
graph planar algebra of the dual principal graph. We then verify that, up to an automorphism
of the graph planar algebra, the resulting embedding does not depend on the choice of generating
object for the module category.
5.1 The Embedding Theorem
Suppose M• is a a connected right planar module over Q•. Since Q• has finite depth, so does M•,
by Lemma 3.12. The tangle
n
(cid:63)
gives a map Q• → M•, which is injective since M• is non-zero and connected. Let M• be the
Markov tower obtained from M• from Example 3.11. Choose r ≥ 0 such that the inclusion M2r ⊆
(M2r+1, tr2r+1) ⊆ (M2r+2, tr2r+2, e2r+1) is standard. Then setting (An, trn) := (M2r+n, tr2r+n) for
n ≥ 0 as described in §4.2, A0 ⊆ (A1, tr1) is a strongly Markov inclusion, and the tower (An, tr2r+n)
is a strongly Markov tower. Let P• be the canonical relative commutant planar †-algebra of the
inclusion A0 ⊆ (A1, tr1) described in §4.1. Therefore, the tangle
2r n
(cid:63)
gives an embedding Φ : Qn → Pn on the level of Markov towers. In terms of the string calculus
of pivotal modules over tensor categories, Φ places 2r strings to the left of elements in Qn,+ and
2r + 1 strings to the left of elements in Qn,−:
Φ
x
n
x
2r
n
Here, the unshaded-shaded strand is a generator of the category of projections C of Q•, while the
shaded-unshaded red strand is a simple generator of a cyclic pivotal right module category over C.
Theorem 5.1. The map Φ is a †-planar algebra embedding.
Proof. For clarity, let us denote the conditional expectations and inclusions in Q• by E and ι,
reserving the plain symbols for their counterparts in P•. Similarly, let us denote the n-th Jones
projection in Q• by εn, reserving en for the Jones projection in P•.
We need only check the hypotheses of Lemma 4.3: that Φ commutes with the action of several
tangles.
35
• (Jones Projections) Φ(εn) = Φ
• (Conditional Expectation) Φ(En(x)) = Φ
• (Right Inclusion) ιn(x) = in
x
n − 1
=
x
=
2r
n − 1
2r
n − 1
=
=
2r + n − 1 = en
2r
x
=
x
= En(Φ(x))
n − 1
2r
n − 1
x
=
x
= in(Φ(x))
• (Left Capping) In §4.1, we discussed that the left-capping tangle in the canonical planar
n 1
n
1
n
1
∗-algebra is given for n ≥ 1 by
0∩An
EA(cid:48)
1∩An
A(cid:48)
(x) =
1
d2
bxb∗
(cid:88)
b
where {b} is a Pimnser Popa basis of A1 over A0. This means that, for any x ∈ Qn,+, we
have
2r + 1
2r + 1
2r + 1
0∩An
d2EA(cid:48)
1∩An
A(cid:48)
(Φ(x)) =
(cid:88)
b
b
2r
b∗
(cid:88)
b
b
2r
b∗
x
=
x
= d
1
x
= d2Φ
(cid:16)
(cid:17)
Qn,+Qn−1,−(x)
E
.
n − 1
n − 1
n − 1
• (Left Inclusion) The left inclusion ln : Pn,− → Pn+1,+ is just the inclusion A(cid:48)
1 ∩ An+1 →
0 ∩ An+1. Graphically, (cid:96)n : Qn,− → Qn+1,+ is equivalent to adding a string on the left.
A(cid:48)
Thus, for x ∈ Qn,−, we have that:
ln(Φ(x)) =
n
x
=
n
x
= Φ((cid:96)n(x)).
2r + 1
2r
1
We have checked that Φ : Q• → P• is a planar †-algebra inclusion. Let G• be the bipartite
graph planar algebra of the fusion graph of X acting on M. Then by [JP11, Th,. 3.33], we know
that P• is a planar †-algebra isomorphic to G•. Thus, we have an embedding of Q• into G•.
Corollary 5.2 (The Embedding Theorem). A finite depth subfactor planar algebra Q• can be
embedded into the bipartite graph planar algebra of the fusion graph of a connected right planar
Q•-module.
36
In particular, by considering (Q•, 10,+) as a connected right planar Q•-module, we recover the
main result of [JP11]. By instead considering (Q•, Y ) for Y a simple summand of the shaded-
unshaded strand X ∈ Q1,−, we obtain an embedding of Q• into the graph planar algebra for the
dual principal graph.
5.2 Invariance of the embedding
As the observant reader may have noted, we made three choices in defining the embedding map
from Q• (cid:44)→ GPA(Γ)•. First, we chose a simple object m ∈ M to get our Markov tower Mn :=
EndM(m (cid:67) X alt⊗n), and second, we chose r ≥ 0 such that the inclusion M2r ⊆ (M2r+1, tr2r+1) ⊆
(M2r+2, tr2r+2, e2r+1) is standard. Third, we chose a basis for the strongly Markov inclusion M2r ⊆
(M2r+1, tr2r+1) to obtain a planar †-algebra isomorphism from the canonical relative commutant
planar algebra P• of the inclusion to the graph planar algebra GPA(Γ)• of the fusion graph Γ. In
this section, we show that the embedding does not depend on these choices up to a †-automorphism
of GPA(Γ)•.
Definition 5.3. Suppose Q• is a subfactor planar algebra and P•,P(cid:48)• are two unitary shaded planar
algebras together with planar algebra embeddings Φ : Q• (cid:44)→ P• and Φ(cid:48) : Q• (cid:44)→ P(cid:48)•. We say the
embeddings Φ and Φ(cid:48) are equivalent if there is a planar †-algebra isomorphism Ψ : P• → P(cid:48)• such
that the following diagram commutes:
Q•
Φ
Φ(cid:48)
P•
Ψ
P (cid:48)•
We now treat our three choices for our embedding in reverse order. First, note that choosing
a different basis for the inclusion just alters the isomorphism P• ∼= GPA(Γ)• by a †-automorphism
of GPA(Γ)•, resulting in equivalent embeddings.
Second, suppose we chose a different r(cid:48) ≥ 0 such that the inclusion M2r(cid:48) ⊆ (M2r(cid:48)+1, tr2r(cid:48)+1) ⊆
(M2r(cid:48)+2, tr2r(cid:48)+2, e2r(cid:48)+1) is standard. Without loss of generality, we may assume r(cid:48) = r + k for
k ∈ N. Denoting the canonical relative commutant planar algebra for the strongly Markov inclusion
M2r+2k ⊆ (M2r+2k+1, tr2r+2k+1) by P(cid:48)•, we see that P(cid:48)• ∼= P• by iteratively applying the shift-by-2
planar algebra isomorphism from Theorem 4.6. Hence, replacing r with r(cid:48) results in an equivalent
embedding.
Third, suppose we chose the simple object n ∈ M0 ⊂ M instead of m, where M0 = M (cid:67) 10.
For i ≥ 0, define Ni := EndM(n (cid:67) X alt⊗i). Since M is indecomposable as a right C-module
category, there is a j > 0 such that n is a subobject of m (cid:67) X alt⊗2j. Fix an orthogonal projection p ∈
M2j = EndM(m (cid:67) X alt⊗2j) with image isomorphic to n. Notice that the compressed shifted Markov
2j+k, pe2j+k+1)k≥0 is ∗-isomorphic to the Markov tower (Nk, trk, fk+1)k≥0, where
tower (pM2j+kp, trp
we denote by ei the Jones projections in M• and by fi the Jones projections in N•. Again, since
M is indecomposable, we may fix k > 0 sufficiently large such that the following three conditions
hold:
(1) The projection p (cid:67) idX alt⊗2k has central support 1 in the finite dimensional von Neumann al-
gebra M2(j+k) = EndM(m (cid:67) X alt⊗2(j+k)), which is equivalent to M2(j+k) = M2(j+k)pM2(j+k)
by finite dimensionality.
(2) Setting r := 2(j+k), the inclusion M2r ⊆ (M2r+1, tr2r+1) ⊆ (M2r+2, tr2r+2, e2r+1) is standard.
37
(3) The inclusion N2k ⊆ (N2k+1, tr2k+1) ⊆ (N2k+2, tr2k+2, e2k+1) is standard.
Now, by Theorem 4.12, compressing M• by p (cid:67) idX alt⊗2k ∈ M2r gives a planar †-algebra isomor-
phism from the canonical relative commutant planar algebra P• of the strongly Markov inclusion
M2r ⊆ (M2r+1, tr2r+1) to the canonical relative commutant planar algebra P p• of the strongly
Markov inclusion pM2rp ⊆ (pM2r+1p, trp
2r+1 is defined analogously to (14), which, in
turn, is †-isomorphic to the canonical relative commutant planar algebra R• of the strongly Markov
inclusion N2k ⊆ (N2k+1, tr2k+1). Hence, replacing m with n results in an equivalent embedding.
2r+1), where trp
We have just proved the following.
Proposition 5.4. The embedding Q• (cid:44)→ GPA(Γ)• is invariant under our choices, up to equiva-
lence.
References
[AMP15] Narjess Afzaly, Scott Morrison, and David Penneys, The classification of subfactors with index at most
5 1
4 , 2015, arXiv:1509.00038, to appear Mem. Amer. Math. Soc.
[BDH14] Arthur Bartels, Christopher L. Douglas, and Andr´e Henriques, Dualizability and index of subfactors,
Quantum Topol. 5 (2014), no. 3, 289 -- 345, MR3342166 DOI:10.4171/QT/53 arXiv:1110.5671.
[BHP12] Arnaud Brothier, Michael Hartglass, and David Penneys, Rigid C∗-tensor categories of bimodules
over interpolated free group factors, J. Math. Phys. 53 (2012), no. 12, 123525, 43, MR3405915
DOI:10.1063/1.4769178 arXiv:1208.5505.
[Bis94]
[Bis97]
Dietmar Bisch, A note on intermediate subfactors, Pacific J. Math. 163 (1994), no. 2, 201 -- 216,
MR1262294.
Dietmar Bisch, Bimodules, higher relative commutants and the fusion algebra associated to a subfactor,
Operator algebras and their applications (Waterloo, ON, 1994/1995), 13-63, Fields Inst. Commun., 13,
Amer. Math. Soc., Providence, RI, 1997, MR1424954, (preview at google books).
[BMPS12] Stephen Bigelow, Scott Morrison, Emily Peters, and Noah Snyder, Constructing the extended
Haagerup planar algebra, Acta Math. 209 (2012), no. 1, 29 -- 82, MR2979509, arXiv:0909.4099,
DOI:10.1007/s11511-012-0081-7.
[BP14]
[Con80]
Stephen Bigelow and David Penneys, Principal graph stability and the jellyfish algorithm, Math. Ann.
358 (2014), no. 1-2, 1 -- 24, MR3157990, DOI:10.1007/s00208-013-0941-2, arXiv:1208.1564.
Alain Connes, On the spatial theory of von Neumann algebras, J. Funct. Anal. 35 (1980), no. 2, 153 -- 164,
MR561983.
[DCY15] Kenny De Commer and Makoto Yamashita, Tannaka-Kreın duality for compact quantum homogeneous
spaces II. Classification of quantum homogeneous spaces for quantum SU(2), J. Reine Angew. Math. 708
(2015), 143 -- 171, MR3420332 DOI:10.1515/crelle-2013-0074.
[EGNO15] Pavel Etingof, Shlomo Gelaki, Dmitri Nikshych, and Victor Ostrik, Tensor categories, Mathematical
Surveys and Monographs, vol. 205, American Mathematical Society, Providence, RI, 2015, MR3242743
DOI:10.1090/surv/205.
[GdlHJ89] Frederick M. Goodman, Pierre de la Harpe, and Vaughan F.R. Jones, Coxeter graphs and towers of
algebras, Mathematical Sciences Research Institute Publications, 14. Springer-Verlag, New York, 1989,
x+288 pp. ISBN: 0-387-96979-9, MR999799.
[Gho11]
[GJ16]
[GL18]
[GLR85]
Shamindra Kumar Ghosh, Planar algebras: a category theoretic point of view, J. Algebra 339 (2011),
27 -- 54, MR2811311, arXiv:0810.4186, DOI:10.1016/j.jalgebra.2011.04.017.
Shamindra Kumar Ghosh and Corey Jones, Annular representation theory for rigid C∗-tensor categories,
J. Funct. Anal. 270 (2016), no. 4, 1537 -- 1584, MR3447719 DOI:10.1016/j.jfa.2015.08.017 arXiv:1502.
06543.
Luca Giorgetti and Roberto Longo, Minimal index and dimension for 2-C∗-categories with finite-
dimensional centers, 2018, arXiv:1805.09234.
P. Ghez, R. Lima, and J. E. Roberts, W ∗-categories, Pacific J. Math. 120 (1985), no. 1, 79 -- 109, MR808930.
38
[GMP+18] Pinhas Grossman, Scott Morrison, David Penneys, Emily Peters, and Noah Snyder, The Extended
Haagerup fusion categories, 2018, arXiv:1810.06076.
[Gup08]
[Han10]
Ved Prakash Gupta, Planar algebra of the subgroup-subfactor, Proc. Indian Acad. Sci. Math. Sci. 118
(2008), no. 4, 583 -- 612.
Richard Han, A Construction of the "2221" Planar Algebra, ProQuest LLC, Ann Arbor, MI, 2010, Thesis
(Ph.D.) -- University of California, Riverside, MR2822034 arXiv:1102.2052.
[HPT16] Andr´e Henriques, David Penneys, and James E. Tener, Planar algebras in braided tensor categories, 2016,
arXiv:1607.06041.
[JMS14]
[Jol90]
[Jon83]
[Jon99]
[Jon01]
Vaughan F. R. Jones, Scott Morrison, and Noah Snyder, The classification of subfactors of index
at most 5, Bull. Amer. Math. Soc. (N.S.) 51 (2014), no. 2, 277 -- 327, MR3166042, arXiv:1304.6141,
DOI:10.1090/S0273-0979-2013-01442-3.
Paul Jolissaint, Index for pairs of finite von Neumann algebras, Pacific J. Math. 146 (1990), no. 1, 43 -- 70,
MR1073519, euclid.pjm/1102645309.
Vaughan F. R. Jones,
DOI:10.1007/BF01389127.
Index for subfactors,
Invent. Math. 72 (1983), no. 1, 1 -- 25, MR696688,
, Planar algebras I, 1999, arXiv:math.QA/9909027.
, The annular structure of subfactors, Essays on geometry and related topics, Vol. 1, 2, Monogr.
Enseign. Math., vol. 38, Enseignement Math., Geneva, 2001, MR1929335, pp. 401 -- 463.
[Jon12]
, Quadratic tangles in planar algebras, Duke Math. J. 161 (2012), no. 12, 2257 -- 2295, MR2972458,
arXiv:1007.1158, DOI:10.1215/00127094-1723608.
[Jon15]
,
Von
Neumann
algebras,
2015,
https://math.vanderbilt.edu/jonesvf/
VONNEUMANNALGEBRAS2015/VonNeumann2015.pdf.
Vaughan F. R. Jones and David Penneys, The embedding theorem for finite depth subfactor planar alge-
bras, Quantum Topol. 2 (2011), no. 3, 301 -- 337, arXiv:1007.3173, MR2812459, DOI:10.4171/QT/23.
Vaughan F.R. Jones and V.S. Sunder, Introduction to subfactors, London Mathematical Society Lec-
ture Note Series, 234. Cambridge University Press, Cambridge, 1997, xii+162 pp. ISBN: 0-521-58420-5,
MR1473221.
Louis H. Kauffman, State models and the Jones polynomial, Topology 26 (1987), no. 3, 395 -- 407, MR899057,
DOI:10.1016/0040-9383(87)90009-7.
Zhengwei Liu, Composed inclusions of A3 and A4 subfactors, Adv. Math. 279 (2015), 307 -- 371, MR3345186
DOI:10.1016/j.aim.2015.03.017 arXiv:1308.5691.
Zhengwei Liu, Scott Morrison, and David Penneys, 1-Supertransitive Subfactors with Index at
Most 6 1
889 -- 922, MR3306607, arXiv:1310.8566,
DOI:10.1007/s00220-014-2160-4.
5 , Comm. Math. Phys. 334 (2015), no. 2,
R. Longo and J. E. Roberts, A theory of dimension, K-Theory 11 (1997), no. 2, 103 -- 159, MR1444286
DOI:10.1023/A:1007714415067.
√
5, J. Funct. Anal. 269
[JP11]
[JS97]
[Kau87]
[Liu15]
[LMP15]
[LR97]
[MP15a]
Scott Morrison and David Penneys, 2-supertransitive subfactors at index 3 +
(2015), no. 9, 2845 -- 2870, MR3394622 DOI:10.1016/j.jfa.2015.06.023 arXiv:1406.3401.
[MP15b]
, Constructing spoke subfactors using the jellyfish algorithm, Trans. Amer. Math. Soc. 367 (2015),
no. 5, 3257 -- 3298, MR3314808 DOI:10.1090/S0002-9947-2014-06109-6 arXiv:1208.3637.
[MPS10]
[MW10]
Scott Morrison, Emily Peters, and Noah Snyder, Skein theory for the D2n planar algebras, J. Pure Appl.
Algebra 214 (2010), no. 2, 117 -- 139, arXiv:0808.0764 MR2559686 DOI:10.1016/j.jpaa.2009.04.010.
Scott Morrison and Kevin Walker, The graph planar algebra embedding theorem, 2010, preprint available
at tqft.net/gpa.
[Pen18]
David Penneys, Unitary dual functors for unitary multitensor categories, 2018, arXiv:1808.00323.
[Pet10]
[Pop94]
Emily Peters, A planar algebra construction of the Haagerup subfactor, Internat. J. Math. 21 (2010),
no. 8, 987 -- 1045, MR2679382, DOI:10.1142/S0129167X10006380, arXiv:0902.1294.
Sorin Popa, Classification of amenable subfactors of type II, Acta Math. 172 (1994), no. 2, 163 -- 255,
MR1278111, DOI:10.1007/BF02392646.
[Pop95]
, An axiomatization of the lattice of higher relative commutants of a subfactor, Invent. Math. 120
(1995), no. 3, 427 -- 445, MR1334479 DOI:10.1007/BF01241137.
39
[PP86]
[PP88]
[PP15]
[Sch13]
[Vor99]
[Wat90]
[Wen88]
[Yam04]
Mihai Pimsner and Sorin Popa, Entropy and index for subfactors, Ann. Sci. ´Ecole Norm. Sup. (4) 19
(1986), no. 1, 57 -- 106, MR860811.
, Iterating the basic construction, Trans. Amer. Math. Soc. 310 (1988), no. 1, 127 -- 133, MR965748.
David Penneys and Emily Peters, Calculating two-strand jellyfish relations, Pacific J. Math. 277 (2015),
no. 2, 463 -- 510, MR3402358 DOI:10.2140/pjm.2015.277-2 arXiv:1308.5197.
Gregor Schaumann, Traces on module categories over fusion categories, J. Algebra 379 (2013), 382 -- 425,
MR3019263 DOI:10.1016/j.jalgebra.2013.01.013 arXiv:1206.5716.
Alexander A. Voronov, The Swiss-cheese operad, Homotopy invariant algebraic structures (Baltimore,
MD, 1998), Contemp. Math., vol. 239, Amer. Math. Soc., Providence, RI, 1999, MR1718089 arXiv:math.
QA/9807037, pp. 365 -- 373.
Yasuo Watatani, Index for C∗-subalgebras, Mem. Amer. Math. Soc. 83 (1990), no. 424, vi+117.
Hans Wenzl, Hecke algebras of type An and subfactors, Invent. Math. 92 (1988), no. 2, 349 -- 383, MR936086
DOI:10.1007/BF01404457.
Shigeru Yamagami, Frobenius duality in C∗-tensor categories, J. Operator Theory 52 (2004), no. 1, 3 -- 20,
MR2091457.
40
|
1508.01530 | 2 | 1508 | 2016-01-11T23:14:11 | Completely contractive projections on operator algebras | [
"math.OA",
"math-ph",
"math.FA",
"math-ph"
] | The main goal of this paper is to find operator algebra variants of certain deep results of Stormer, Friedman and Russo, Choi and Effros, Effros and Stormer, Robertson and Youngson, Youngson, and others, concerning projections on C*-algebras and their ranges. (See papers of these authors referenced in the bibliography.) In particular we investigate the `bicontractive projection problem' and related questions in the category of operator algebras. To do this, we will add the ingredient of `real positivity' from recent papers of the first author with Read. | math.OA | math |
COMPLETELY CONTRACTIVE PROJECTIONS ON OPERATOR
ALGEBRAS
DAVID P. BLECHER AND MATTHEW NEAL
Abstract. The main goal of this paper is to find operator algebra variants of
certain deep results of Størmer, Friedman and Russo, Choi and Effros, Effros
and Størmer, Robertson and Youngson, Youngson, and others, concerning
projections on C ∗-algebras and their ranges.
(See papers of these authors
referenced in the bibliography.) In particular we investigate the 'bicontractive
projection problem' and related questions in the category of operator algebras.
To do this, we will add the ingredient of 'real positivity' from recent papers of
the first author with Read.
1. Introduction
In previous papers both authors separately studied projections (that is, idempo-
tent linear maps) and conditional expectations on unital operator algebras (that is,
closed algebras of operators on a Hilbert space that contain the identity operator)
and other classes of Banach spaces. (See papers of the authors referenced in the
bibliography below.) Results were proved such as: a completely contractive projec-
tion P on such an algebra A which is unital (that is, P (1) = 1) and whose range is
a subalgebra, is a 'conditional expectation' (that is P (P (a)bP (c)) = P (a)P (b)P (c)
for a, b, c ∈ A) [12, Corollary 4.2.9]. This is an analogue of the matching theorem
due to Tomiyama for C∗-algebras.
The main goal of our paper is to find variants, valid for operator algebras which
are unital or which have an approximate identity, of certain deeper results in the C∗-
algebra case, due to Størmer, Friedman and Russo, Effros and Størmer, Robertson
and Youngson, and others, concerning projections and their ranges. (See papers
of these authors referenced in the bibliography below.) In particular we wish to
investigate the 'bicontractive projection problem' and related problems (such as
the 'symmetric projection problem' and the 'contractive projection problem') in
the category of operator algebras. To do this, we will add the ingredient of 'real
positivity' from recent papers of the first author with Read in [18, 19, 20] (see also
[14, 15, 6, 17, 9]), A key idea in those papers is that 'real positivity' is often the
right replacement in general algebras for positivity in C∗-algebras. This will be our
guiding principle here too.
We now discuss the structure of our paper. In Section 2 we discuss completely
contractive projections on operator algebras. A well known and lovely result of
Choi and Effros [23, Theorem 3.1] (or rather, its proof), shows that the range
of a completely positive projection P : B → B on a C∗-algebra B, is again a
1991 Mathematics Subject Classification. Primary 46L05, 46L07, 47L07, 47L30; Secondary
46B40, 46H10, 46L30.
*Blecher was partially supported by a grant from the National Science Foundation. Neal was
supported by Denison University.
1
2
DAVID P. BLECHER AND MATTHEW NEAL
C∗-algebra with product P (xy). A quite deep theorem of Friedman and Russo,
and a simpler variant of this by Youngson, shows that something similar is true
if P is simply contractive, or if B is replaced by a ternary ring of operators ([30,
50]). The analogous result for unital completely contractive projections on unital
operator algebras is true too, and is implicit in the proof of [12, Corollary 4.2.9]
referred to above. However there seems to be no analogous result for (nonunital)
completely contractive projections on nonunital operator algebras without adding
extra hypotheses on P . The 'guiding principle' in the previous paragraph suggests
to add the condition that P is also 'real completely positive' (we define this below).
Then the question does make good sense, and we are able to prove the desired result.
Thus the range of a real completely positive completely contractive projection P :
A → A on an operator algebra with approximate identity, is again an operator
algebra with product P (xy). We also have a converse and several complements to
this result in Section 2, as well as some other facts about completely contractive
projections, such as how one is often able to reduce the problem to algebras which
have an identity. We also show that for algebras with no kind of approximate
identity, there is a biggest 'nice part' on which completely contractive projections
(and the other classes of projections discussed below) work well.
In Sections 3 -- 5 we turn from the 'contractive projection problem' to the 'bicon-
tractive projection problem' and related questions. A projection P is bicontractive
if both P and I − P are contractive. By the bicontractive projection problem for a
Banach space X, one usually means one or both of the following: 1) the characteri-
zation of all bicontractive projections P : X → X; or 2) the characterization of the
ranges of the bicontractive projections. On a unital C∗-algebra B it is known, by
work of some of the authors mentioned above, that the unital bicontractive projec-
tions are precisely 1
2 (I +θ), for a period 2 ∗-automorphism θ : B → B. The possibly
nonunital bicontractive projections P on B are of a similar form, and indeed if P is
also positive then q = P (1) is a central projection in the multiplier algebra M (B)
with respect to which P decomposes into a direct sum of 0 and a projection of the
above form 1
2 (I + θ), for a period 2 ∗-automorphism θ of qB. (See Theorem 3.2
for the idea of the proof of this.) Conversely, note that a map P of the latter form
is automatically 'completely bicontractive' (that is, is bicontractive at each matrix
level), indeed is 'completely symmetric' (that is, I − 2P is completely contractive),
and the range of P is a C∗-subalgebra, and P is a conditional expectation.
One may ask: what from the last paragraph is true for general operator alge-
bras? Again the guiding principle referred to earlier leads us to use in place of the
positivity in that result, the real positivity in the sense of [19, 20]. The next thing
to note is that now 'completely bicontractive' is no longer the same as 'completely
symmetric' for projections. The 'completely symmetric' case works beautifully, and
the solution to our 'symmetric projection problem' is presented in Theorem 3.7 in
Section 3. This result is one somewhat satisfactory generalization (we shall see oth-
ers later) to operator algebras of the C∗-algebraic theorem in the last paragraph.
For the more general class of completely bicontractive projections, as seems to be
often the case in generalizing C∗-algebraic theory to more general algebras, a first
look is disappointing. Indeed most of the last paragraph no longer works in general.
One does not always get an associated completely isometric automorphism θ with
P = 1
Indeed we have
solved here and elsewhere (see e.g. [11, 7, 13] and see also p. 92 -- 93 in the prequel to
2 (I + θ), and q = P (1) need not be a central projection.
PROJECTIONS ON OPERATOR ALGEBRAS
3
the last cited paper, as well as the paper [16] in preparation) many of the obvious
questions about contractive projections, completely contractive projections, and
conditional expectations, on operator algebras. Unfortunately many of the answers
are counterexamples. However, as also seems to be often the case, a closer look at
examples reveals an interesting question. Namely, given a real completely positive
projection P : A → A which is completely bicontractive, when is the range of P
an (approximately unital) subalgebra of A, so that P is a conditional expectation?
For operator algebras we consider this to be the correct version of the 'bicontractive
projection problem'. In Sections 3 and 4 we elucidate this question. We remark
that in a paper in preparation [16] we study the 'Jordan algebra' variants of many
of the results in Sections 2 -- 3 of the present paper. The Jordan variants of the
material in Section 4 is not going to be any better behaved than what we do there,
and so we do not plan discuss this much in [16].
In Section 4 we discuss the completely bicontractive projection problem, and
construct some interesting examples, and give some reasonable conditions under
which P (A) is a subalgebra and P is a conditional expectation. In particular we
solve in full generality our version of the 'bicontractive projection problem' for
uniform algebras (that is, closed subalgebras of C(K)), and indeed for any algebra
satisfying a condition related to semisimplicity. Theorem 4.9 is one of the main
results of the paper, giving a very general condition for P (A) being a subalgebra
in terms of certain support projections.
In fact, at the time of writing, for all
we know the condition in Theorem 4.9 is necessary and sufficient; at least we
have no examples to the contrary. In Section 5 we discuss another condition that
completely bicontractive projections may satisfy, and examine some consequences
of this.
In Section 6 we discuss Jordan homomorphisms on operator algebras,
and amongst other things, solve two natural 'completely isometric problems', for
Jordan subalgebras of operator algebras and for operator algebras, related to the
noncommutative Banach -- Stone theorem.
We now turn to precise definitions and notation. Any unexplained terms below
can probably be found in [12, 45, 18, 19], or any of the other books on operator
spaces. All vector spaces are over the complex field C. The letters K, H denote
Hilbert spaces. If X is an operator space we often write X+ for the elements in
X which are positive (i.e. ≥ 0) in the usual C∗-algebraic sense. We write Ball(X)
for the set {x ∈ X : kxk ≤ 1}. By an operator algebra we mean a not necessarily
selfadjoint closed subalgebra of B(H), the bounded operators on a Hilbert space
H. We write C∗(A) for the C∗-algebra generated by A, that is the smallest C∗-
subalgebra containing A. A unital operator space is a subspace X of B(H) or a
unital C∗-algebra containing the identity (operator). We often write this identity
as 1X. A map T : X → Y is unital if T (1X) = 1Y . We say that an algebra is
approximately unital if it has a contractive approximate identity (cai). For us a
projection in an operator algebra A is always an orthogonal projection lying in A,
whereas a projection on A is a linear idempotent map P : A → A. If A is a nonunital
operator algebra represented (completely) isometrically on a Hilbert space H then
one may identify the unitization A1 with A + C IH . The second dual A∗∗ is also an
operator algebra with its (unique) Arens product, this is also the product inherited
from the von Neumann algebra B∗∗ if A is a subalgebra of a C∗-algebra B. Note
that A has a cai iff A∗∗ has an identity 1A∗∗ of norm 1, and then A1 is sometimes
4
DAVID P. BLECHER AND MATTHEW NEAL
identified with A + C 1A∗∗. The multiplier algebra M (A) of such A may be taken
to be the 'idealizer' of A in A∗∗, that is {η ∈ A∗∗ : ηA + Aη ⊂ A}.
If A is an approximately unital operator algebra or unital operator space then
I(A) denotes the injective envelope, an injective unital C∗-algebra containing A. It
contains A as a subalgebra if A is approximately unital [12, Corollary 4.2.8]. For us
the most important properties of I(A) are, first, that it is injective in the category
of operator spaces, so that any from a subspace of an operator space Y into I(A)
extends to a complete contraction from Y to I(A). Second, I(A) is rigid, so that the
identity map on I(A) is the only complete contraction : I(A) → I(A) extending the
identity map on A. The C∗-envelope C∗e (A) is the C∗-subalgebra of I(A) generated
by A. If A is unital it has the property that given any unital complete isometry
T : A → B(K), there exists a unique ∗-homomorphism π : C∗(A) → C∗e (A) with
π ◦ T equal to the inclusion map of A in C∗e (A).
We recall that a contractive completely positive map on a C∗-algebra is com-
pletely contractive. A unital linear map between operator systems is positive and
∗-linear if it is contractive; and it is completely positive iff it is completely contrac-
tive. See e.g. [12, Section 1.3] for these.
A hereditary subalgebra, or HSA, in an operator algebra A is an approximately
unital subalgebra with DAD ⊂ D. See [10] for their basic theory. The support
projection of an HSA in A is the identity of its bidual, viewed within A∗∗.
We write rA = {x ∈ A : x + x∗ ≥ 0}, and call these the real positive elements.
This space may be defined purely internally without using the 'star', as the accretive
elements, which have several purely metric definitions (see e.g. [9, Lemma 2.4]).
Also rA is the closure of the positive real multiples of FA = {a ∈ A : k1 − ak ≤
1} (see [19]). Read's theorem states that any operator algebra with cai has a
real positive cai (see e.g. [8] for a proof of this). Since Mn(A)∗∗ ∼= Mn(A∗∗) (see
[12, Theorem 1.4.11]), we have that rMn(A∗∗) is the weak* closure of rMn(A), and
rMn(A) = Mn(A) ∩ rMn(A∗∗) for each n.
A linear map T : A → B between operator algebras or unital operator spaces is
real positive if T (rA) ⊂ rB. It is real completely positive, or RCP for short, if Tn is
real positive on Mn(A) for all n ∈ N. One may also define these maps in terms of
the set FA above [19], but the definitions are shown to be equivalent in [6, Section 2].
From [6]: a linear map T : A → B(H) on an approximately unital operator algebra
or unital operator space A is RCP iff T has a completely positive (in the usual
sense) extension T : C∗(A) → B(H). Here C∗(A) is a C∗-algebra generated by
A. We call this the 'generalized Arveson extension theorem'. Thus real completely
positivity on A is equivalent to P extending to a completely positive map on a
containing C∗-algebra. A unital completely contractive map on a unital operator
space is RCP, since it extends to a completely contractive map on a containing
unital C∗-algebra, and such maps are completely positive as we said above.
A TRO or ternary ring of operators is a subspace of B(K, H) such that ZZ∗Z ⊂
Z. A WTRO is a weak* closed TRO. The second dual of a TRO Z is a WTRO (see
[12, Chapter 8] for this and the next several facts). We write L(Z) for the linking
C∗-algebra of a TRO, this has 'four corners' ZZ∗, Z, Z∗, and Z∗Z. Here ZZ∗ is
the closure of the linear span of products zw∗ with z, w ∈ Z, and similarly for Z∗Z.
One gets a similar von Neumann algebra for WTRO's. A ternary morphism on a
TRO Z is a linear map T such that T (xy∗z) = T (x)T (y)∗T (z) for all x, y, z ∈ Z. A
tripotent is an element u ∈ Z such that uu∗u = u. We order tripotents by u ≤ v if
PROJECTIONS ON OPERATOR ALGEBRAS
5
and only if uv∗u = u. This turns out to be equivalent to u = vu∗u, or to u = uu∗v,
and implies that u∗u ≤ v∗v and uu∗ ≤ vv∗ [5]. If x ∈ Ball(Z), define u(x) to be
the weak* limit of the sequence (x(x∗x)n) in Z∗∗. This is the largest tripotent in
B∗∗ satisfying vv∗x = v (see [25]). If x ≥ 0, or if u(x) is a projection then u(x) is
also the weak* limit of powers xn as n → ∞ (see e.g. [15, 8]).
We will say that an idempotent linear P : X → X is a symmetric (resp. com-
pletely symmetric) projection, if kI−2Pk ≤ 1 (resp. kI−2Pkcb ≤ 1). This is related
to the notion of u-ideal [31], but we will not need anything from that theory. Such
are automatically bicontractive (resp. completely bicontractive). We say that P
is completely hermitian if P is hermitian in CB(X). Note that since exp(itP ) =
I − P + eitP it follows that P is completely hermitian iff kI − P + eitPkcb ≤ 1 for
all real t. This is essentially the notion of being (completely) 'bicircular'. Clearly
if P is completely hermitian then it is completely symmetric. We will not dis-
cuss (completely) hermitian projections much in this paper, these seem much less
interesting.
2. Completely contractive projections on approximately unital
operator algebras
Looking at examples it becomes clear that projections on operator algebras with
no kind of approximate identity can be very badly behaved. Hence we will say little
in our paper about such algebras. However it is worth mentioning that we can pick
out a 'good part' of such a projection. This is the content of our first result.
Proposition 2.1. Let P : A → A be a real completely positive completely con-
tractive map (resp. projection) on an operator algebra A (possibly with no kind of
approximate identity). There exists a largest approximately unital subalgebra D of
A, and it is a HSA (hereditary subalgebra) of A. Moreover, P (D) ⊂ D, and the
restriction P ′ of P to D is a real completely positive completely contractive map
(resp. projection) on D. In addition, P ′ is completely bicontractive (resp. completely
symmetric) if P has the same property.
Proof. By [20, Corollary 2.2], D = rA − rA is the largest approximately unital
subalgebra of A. This algebra is written as AH there, and was first introduced in
[19, Section 4]. Clearly P (D) ⊂ D. The rest is obvious.
(cid:3)
Remarks. 1) The last result is also true the word 'completely' removed through-
out, with the same proof.
2) Letting p be the support projection of the HSA D above, if P extends to
a completely positive complete contraction on a containing C∗-algebra (as in our
generalized Arveson extension theorem mentioned in the introduction, see also [6,
Theorem 2.6]) then one can show that P ∗∗(pa) = pP (a) and P ∗∗(ap) = P (a)p for
a ∈ A. It follows that P may be pictured as a 2 × 2 matrix with its 'good part'
above in the 1-1 corner. However in general it seems one can say little about the
other corners, they can be quite messy. This is why we focus on algebras with
approximate identities in our paper.
Proposition 2.2. A real completely positive completely contractive map (resp. pro-
jection) on an approximately unital operator algebra A, extends to a unital (real
completely positive) completely contractive map (resp. projection) on the unitiza-
tion A1. (If A is unital then we define A1 = A ⊕∞ C here.)
6
DAVID P. BLECHER AND MATTHEW NEAL
Proof. Suppose that P : A → A ⊂ B(H) is the projection. By [6, Theorem 2.6],
P extends uniquely to a completely positive completely contractive map C∗(A) →
B(H). By [22, Lemma 3.9] it extends further to a unital completely positive map
C∗(A) + C IH → B(H). The restriction of the latter map to A + C IH may be
viewed as a unital (real completely positive) completely contractive map on the
unitization A1 → A1, and it is evidently a projection if P was a projection.
(cid:3)
The previous result gives a way to 'reduce to the unital case'. However this
method does not seem to be helpful later in our paper when dealing with bicon-
tractive or symmetric projections, and we will need a different 'reduction to the
unital case'.
Lemma 2.3. Let P : A → A be a real positive contractive map on a unital operator
algebra. Then 0 ≤ P (1) ≤ 1.
Proof. The restriction of P to ∆(A) = A ∩ A∗ is real positive. Hence it is positive
by the proof of [6, Theorem 2.4]. So 0 ≤ P (1) ≤ 1.
Lemma 2.4. Suppose that E is a completely contractive completely positive projec-
tion on an operator system X. Then the range of E, with its usual matrix norms,
is an operator system with matrix cones En(Mn(X)+) = Mn(X)+ ∩ Ran(En), and
unit E(1).
(cid:3)
Proof. We will use the Choi-Effros characterization of operator systems [23]. Be-
cause Ran(En) is a ∗-subspace of Mn(X), with the inherited cone from Mn(X)+,
it is a partially ordered, matrix ordered, Archimidean ∗-vector space with proper
If x = x∗ there exists a positive scalar t with −t1 ≤ x ≤ t1, so that
cones.
−tE(1) ≤ x ≤ tE(1). So E(1) is an order unit. If x ∈ Ran(E) with kxkX ≤ 1 then
(cid:20) 1
x∗
x
1 (cid:21) ≥ 0
in M2(X). Applying E2, we deduce that
(cid:20) E(1)
x∗
x
E(1) (cid:21) ≥ 0,
so that the norm of x is ≤ 1 in the new 'order unit norm' (see [23, p. 179]). Con-
versely, if the last norm is ≤ 1, or equivalently if the last centered equation holds,
then it is a simple exercise in operator theory that kxkX ≤ 1, since kE(1)k ≤ 1
and E(1) ≥ 0. Thus the 'order unit norm' coincides with the old norm. Simi-
larly at each matrix level. By the Choi-Effros characterization of operator systems
(Ran(E), E(1)) is an operator system with the given matrix cones, and the order-
unit matrix norms are the usual norms.
(cid:3)
The following generalization of the Choi-Effros result referred to in the intro-
duction, solves the 'completely contractive projection problem' in the category of
approximately unital operator algebras and real completely positive projections.
We remark that the case when also P (1A) = 1A is implicit in the proof of [12,
Corollary 4.2.9].
Theorem 2.5. Let A be an approximately unital operator algebra, and P : A → A
a completely contractive projection which is also real completely positive. Then the
PROJECTIONS ON OPERATOR ALGEBRAS
7
range B = P (A) is an approximately unital operator algebra with product P (xy).
We have
P (P (a)b) = P (P (a)P (b)) = P (aP (b)),
a, b ∈ A.
In particular P (P (1)n)) = P (1) for all n ∈ N, if A is unital. With respect to the
'multiplication' P (xy), A is an bimodule over B , and P viewed as a map A → B is
a B-bimodule map (with B equipped with its new product). If A is unital then P (1)
is the identity for the latter product. Moreover (not assuming A unital), P extends
to a completely positive completely contractive projection on the injective envelope
I(A).
Proof. We give two proofs, since they both use techniques the reader will need to
be familiar with in the rest of our paper.
Set B = P (A). Extend P to a unital completely contractive projection P 1
on A1 by Proposition 2.2. We may then use the proof of [12, Corollary 4.2.9],
which proceeds by extending P to a unital (completely positive and) completely
contractive projection E on I(A1). It follows from the Choi-Effros result mentioned
early in our introduction, that Ran(E) is a unital C∗-algebra with product E(xy),
and B with product P (xy) is a unital subalgebra of this C∗-algebra. We also have
by the same Choi-Effros result (or its proof) that E(E(a)b) = E(E(a)E(b)) =
E(aE(b)) for all a, b ∈ A, giving the centered equation in the theorem statement.
This gives the first several assertions of our theorem. Note that
P (P (et)P (a)) = P (etP (a)) → P (P (a)) = P (a),
a ∈ A,
if (et) is a cai for A. Similarly on the right, so that (P (et)) is a cai for P (A) in its
new product. That A is actually a B-bimodule follows from the centered equation
in the theorem statement; for example because P (P (P (a)P (b))c) equals
P (P (P (a)P (b))P (c)) = P (P (a)P (b)P (c)) = P (P (a)P (P (b)P (c))) = P (P (a)P (P (b)c)).
The centered equation in the theorem statement is just saying that P is a B-
bimodule map for the given products. The final assertion about extending to I(A)
is easy from the above in the case that A is unital; the other case we will do below.
For the second proof, first suppose that A is unital. Let B = P (A), and set
X = A + A∗, Y = B + B∗ and v = P (1). By [6, Theorem 2.6], P extends uniquely
to a completely positive completely contractive map P ′ on X. Since X = A + A∗
this map is uniquely determined, it must be a1 + a∗2 7→ P (a1) + P (a2)∗, a projection
on X with range Y . By Lemma 2.4 (Y, v) is an operator system with positive cone
P (X+). Let j : Y → X be the inclusion map. Then extend P ′ to a completely
positive complete contraction P : I(A) → I(B) by [6, Theorem 2.6]. Extend j to a
completely positive complete contraction j : I(B) → I(A) by Arveson's extension
theorem [3]. Then P ◦ j equals the identity map on I(B) by rigidity of the injective
envelope, since P ◦ j = IB. Thus E = j ◦ P is a completely positive completely
contractive projection on I(A) extending P . We deduce just as in the last paragraph
that Ran(E) is a unital C∗-algebra with product E(xy), B with product P (xy) is
a unital subalgebra of this C∗-algebra, and P (P (a)b) = P (P (a)P (b)) = P (aP (b)).
Finally, if A is nonunital but approximately unital, then P ∗∗ is a completely
contractive projection on A∗∗, which is also real positive by the proof of the main
theorem in [6, Section 2]. By the unital case, P ∗∗(xy) is an operator algebra product
on P ∗∗(A∗∗), with unit v = P ∗∗(1). Hence by restriction P (xy) is an operator
algebra product on P (A), and the centered equation in the theorem statement holds
8
DAVID P. BLECHER AND MATTHEW NEAL
on A, as does the assertions about bimodules. Note that P ∗∗(A∗∗) = (P (A))⊥⊥,
so that P ∗∗(A∗∗) ∼= P (A)∗∗. So P (A)∗∗ is unital, and hence P (A) is approximately
unital (or this may be seen directly using the centered equation in the theorem
statement). Also, since v = P ∗∗(1) acts as an identity on P (A) in the new product,
we can identify P (A) + C v, as a unital operator space, with the unitization of P (A)
with its new operator algebra product. Then the restriction r of P ∗∗ to A + C 1A∗∗
can be viewed as a real completely positive completely contractive projection on
the unitization A1. By the last paragraph, r extends to a completely positive
completely contractive projection on I(A1). However, I(A1) = I(A) by e.g. [12,
Corollary 4.2.8].
(cid:3)
Remarks. 1) Thus the category of approximately unital operator algebras and
real completely positive projections forms a 'projectively stable category' in the
sense of Friedman and Russo (see Neal and Russo [44, p. 295 -- 296], and e.g. [30]).
Namely, if B is the category of Banach spaces with morphisms being the contractive
projections, a subcategory S of B is projectively stable if S is closed under images
of morphisms. That is, for an object E and morphism ϕ : E → E in S, ϕ(E) is
again an object in S (although not necessarily a 'subobject' with respect to the full
structure of objects in S). For example, the subcategory of unital C∗-algebras and
completely positive unital projections is projectively stable, by the theorem of Choi
and Effros used earlier. Other projectively stable categories are listed in the last
references; e.g. the subcategory of TRO's and completely contractive projections
is projectively stable by Youngson's theorem [50]. In the cited pages of [44] the
concept of a 'projectively rigid category' is discussed. The associated question for
us would be if the preduals of dual operator algebras are projectively (completely)
rigid in their sense. However the answer to this is in the negative, since the category
of Banach or operator spaces is not projectively (completely) rigid, and then one
can play the U(X) trick (described above Proposition 3.3 below) to answer the
operator algebra question.
2) If A is unital and C is the C∗-subalgebra of I(A) generated by P (A), then
the map P in the proof restricts to a ∗-homomorphism from C onto C∗e (B), the
latter viewed as a subalgebra of I(B) (or as a C∗-subalgebra of the space Ran(E)
in the proof, with its 'Choi-Effros product'). See e.g. [12, Theorem 1.3.14 (3)].
Lemma 2.6. Let A be a unital operator algebra, and P : A → A a contractive
projection, such that Ran(P ) contains an orthogonal projection q with P (A) =
qP (A)q. Then q = P (1A).
Proof. We have kq±(1−q)k ≤ 1, so that kq±P (1−q)k ≤ 1. Since P (A) = qP (A)q,
and q is an extreme point of the unit ball of qAq (the identity is an extreme point
of the unit ball of any unital Banach algebra), we have that P (1 − q) = 0. Thus
P (1) = q.
(cid:3)
The following 'reduction to the case of unital maps' works under a certain con-
dition which will be seen to be automatic in the setting found in the next Sections
of the paper.
Proposition 2.7. Let A be an approximately unital operator algebra, and P :
A → A a completely contractive projection. Then Ran(P ∗∗) contains an orthogonal
projection q such that P (A) = qP (A)q iff P ∗∗(1) is a projection.
In this case
PROJECTIONS ON OPERATOR ALGEBRAS
9
q = P ∗∗(1), and Ran(P ) is an approximately unital operator algebra with product
P (xy), and its bidual has identity q. Also, P is real completely positive, all the
conclusions of Theorem 2.5 hold, q is an open projection for A∗∗ in the sense of
[10], and
P (a) = qP (a)q = P ∗∗(qaq),
a ∈ A,
(and we can replace P ∗∗ by P here if A is unital). Hence P (A) = qP (A)q =
P ∗∗(qAq), and P 'splits' as the sum of the zero map on q⊥A + Aq⊥ + q⊥Aq⊥, and
a real completely positive completely contractive projection P ′ on qAq with range
equal to P (A). This projection P ′ on qAq is unital if A is unital.
Proof. Let Q = P ∗∗, a completely contractive projection on A∗∗. We can replace
Q by P below if A is unital. If P (A) = qP (A)q for a projection q then Q(A∗∗) =
qQ(A∗∗)q by standard weak* approximation arguments, so by the lemma, Q(1) = q.
Conversely, suppose that Q(1) = q is a projection. Then Q(q⊥) = 0. Note that
Ran(Q) is a unital operator space (in qA∗∗q). So Q, and hence also P , is real
completely positive by [6, Lemma 2.2], since it extends by e.g. [12, Lemma 1.3.6] to
a completely positive unital map from X +X∗ onto Y +Y ∗ where X = A∗∗ and Y =
Q(A∗∗). By extending Q further to a completely positive completely contractive
map on a containing C∗-algebra, and using the Kadison-Schwarz inequality, we
have
Q(aq⊥)∗Q(aq⊥) ≤ Q(q⊥a∗aq⊥) ≤ Q(q⊥) = 0,
a ∈ Ball(A∗∗).
Thus Q(a) = Q(aq) for all a ∈ A∗∗, and similarly Q(a) = Q(qa). Also Q(q)2 =
Q(q), and so P (A) = Q(qAq) = qP (A)q by Choi's multiplicative domain trick (the
latter is usually stated for unital maps, but the general case may be reduced to this
using [22, Lemma 3.9]).
The rest follows from Theorem 2.5 and its proof, with the exception of q being an
open projection for A∗∗. To see this, if (et) is a cai for P (A) with its new product
then using some of the facts here and in Theorem 2.5 we have et = P (etq) =
P (et)q → q weak*.
(cid:3)
Remarks. 1) Note that even a completely contractive completely positive
projection on a unital C∗-algebra need not have P (1) a projection. To see this,
choose a norm 1 element x 6= 1 in A+ and a state ϕ with ϕ(x) = 1, and consider
P = ϕ(·) x.
2) Unfortunately the projection q here need not be central, even if P is com-
pletely bicontractive. See the next example.
Example 2.8. Consider the canonical projection of the upper triangular matri-
ces A onto C E11. This is a completely bicontractive projection (which is also
completely bicontractive, completely hermitian, etc), but it does not extend to a
completely bicontractive projection on its C∗-envelope (or injective envelope) M2.
In this case note that A + A∗ = C∗(A) = C∗e (A) = I(A). On the positive side, the
range of this projection is a subalgebra of A.
Corollary 2.9. Let A be an approximately unital operator algebra, with an approx-
imately unital subalgebra B which is the range of a completely contractive projection
P on A. Then P is real completely positive, and all the conclusions of Theorem 2.5
hold. Hence P is a conditional expectation: P (a)b = P (ab) and bP (a) = P (ba) for
all b ∈ B = P (A) and a ∈ A. It follows that (P (et)) is a cai for B for any cai (et)
of A.
10
DAVID P. BLECHER AND MATTHEW NEAL
Proof. Consider P ∗∗, a completely contractive projection on A∗∗ with range B∗∗.
Of course B∗∗ has an identity of norm 1 as we said in the introduction. By Propo-
sition 2.7, P ∗∗ is real completely positive, and hence so is its restriction P . The
remaining assertions follow e.g. from Proposition 2.7, except for the last assertion
which is an easy consequence of the second last assertion.
(cid:3)
The last result, which may be seen as a converse to Theorem 2.5, generalizes the
fact from [12] mentioned at the start of the introduction. We showed in [7] that
this is all false with the word 'completely' removed, however see [37] for some later
variants valid for certain Banach algebras.
We will need the following results later. If P : M → M is a unital completely
contractive projection on a von Neumann algebra, there exists a support projection
e, the perp of the supremum of all projections in Ker(P ), as in [27, p. 129]. We
have e ∈ P (M )′, and P (x) = P (ex) = P (xe) for all x ∈ M , and if x ∈ M+ then
P (x) = 0 iff xe = 0 iff ex = 0 (see e.g. around Lemma 1.2 in [27]). Following the
idea in the proof of [27, Lemma 1.2 (3)] we have:
Proposition 2.10. Let P : M → M a weak* continuous unital completely con-
tractive projection on a von Neumann algebra M . Let e be the support projection of
P on M discussed above. Let N be the von Neumann algebra generated by P (M ).
Then P (x)e = eP (x)e = exe = xe for all x ∈ N .
Proof. For n = 0, 1, . . . , let An be the span of products of 2n elements from P (M ).
Then An is a unital ∗-subspace of M . Suppose that eP (x)e = exe for all x ∈ An.
Then for such x, set z = e(P (x∗x) − x∗ex)e. Following the steps in the proof
of [27, Lemma 1.2 (3)] with minor modifications we have P (z) = 0 and z ≥ 0,
so that by the facts above the present proposition we obtain z = eze = 0 and
eP (x∗x)e = ex∗xe. By the polarization identity eP (y∗x)e = ey∗xe for x, y ∈ An.
So eP (x)e = exe for all x ∈ An+1, and hence for all x ∈ N .
Corollary 2.11. Let P : A → A be a unital completely contractive projection on
an operator algebra. If P (A) generates A as an operator algebra, then (I − P )(A) =
Ker(P ) is an ideal in A. In any case, if D is the closed algebra generated by P (A)
then (I − P )(D) is an ideal in D.
Proof. We may assume that P (A) generates A. As above we extend P to a unital
completely contractive projection P on a C∗-algebra B (= I(A)). The second
adjoint of this is a weak* continuous unital completely contractive projection on
a von Neumann algebra M , and we continue to write this projection as P . Let
P also denote the restriction of the latter projection to the von Neumann algebra
If x ∈ (I − P )(A), then P (x) = 0, and so by
N generated by P (A) inside M .
Proposition 2.10 we have ex = xe = 0. Thus x ∈ e⊥M e⊥ (and is also in e⊥N e⊥).
So for y ∈ A we have P (xy) = P (exy) = 0. Similarly yx ∈ (I − P )(A), so the latter
is an ideal.
(cid:3)
(cid:3)
For later use we record that if P : A → A is a unital completely contractive
projection on an operator algebra, then in the language of the last proof,
D ∩ e⊥M e⊥ = D ∩ e⊥N e⊥ = Ker(PD) = (I − P )(D).
Indeed since we said above Proposition 2.10 that P (x) = P (exe) for any x ∈ M , we
have D∩e⊥M e⊥ ⊂ Ker(PD). Moreover, if d ∈ D with P (d) = 0 then the argument
PROJECTIONS ON OPERATOR ALGEBRAS
11
in the last proof shows that d ∈ D ∩ e⊥N e⊥ ⊂ D ∩ e⊥M e⊥. So D ∩ e⊥M e⊥ =
D ∩ e⊥N e⊥ = Ker(PD).
3. The symmetric projection problem and the bicontractive
projection problem
It turns out that the variant of the bicontractive projection problem for sym-
metric projections works out perfectly. This is the question of characterizing (com-
pletely) symmetric projections in the categories we are interested in, and their
ranges. Notice that if P : X → X is a projection on a normed space and if we let
θ = 2P − I, so that P = 1
2 (I + θ), then Ran (P ) is exactly the set of fixed points of
θ, and θ ◦ θ = I. Note too that θ is contractive if P is symmetric. From the latter
facts we deduce that θ is a bijective isometry whose inverse is itself. Also θ(1) = 1
if X is a unital algebra and P (1) = 1. Applying the same argument at each 'matrix
level' we see that:
Lemma 3.1. A projection P : X → X on an operator space is completely symmet-
ric (resp. symmetric) iff θ = 2P − I is a complete isometry (resp. an isometry),
and in this case θ is also a surjection. If X is also an algebra (resp. Jordan algebra)
and if θ is a homomorphism (resp. Jordan homomorphism) then the range of P is
a subalgebra (resp. Jordan subalgebra).
Proof. For the last part, Ran (P ) is exactly the fixed points of θ.
(cid:3)
Thus the (completely) 'symmetric projection problem' in some sense a special
case of the (complete) 'isometry problem': namely characterizing the linear (com-
plete) isometries between the objects in our category. That is, the key to solv-
ing the (completely) 'symmetric projection problem' is proving a 'Banach -- Stone
type theorem' in our category. The original Banach -- Stone theorem characterizes
unital isometries between C(K) spaces, and in particular shows that such are ∗-
isomorphisms. Putting this together with the last assertion of the last lemma, we
see that one of the hoped-for conclusions of the (completely) 'symmetric projection
problem', and by extension the (completely) 'bicontractive projection problem', is
that the range of the projection is a subalgebra. We will also show in the completely
symmetric case that if A is unital or approximately unital then so is P (A).
Let us examine what this all looks like in a C∗-algebra, where as predicted in
the last paragraph, much hinges on the known 'Banach -- Stone type theorem' for
C∗-algebras, due mainly to Kadison. The following is essentially well known (see
e.g. [29, 49]), but we do not know of a reference which has all of these assertions,
or is in the formulation we give:
Theorem 3.2. If P : A → A is a projection on a C∗-algebra A then P is bicon-
tractive iff P is symmetric. Then P is bicontractive and completely positive iff there
exists a central projection q ∈ M (A), such that P = 0 on q⊥A, and there exists a
period 2 ∗-automorphism θ of qA so that P = 1
Proof. Clearly symmetric projections are bicontractive. Conversely, if P is bicon-
tractive then by [29, Theorem 2] θ = 2P − IA is a linear surjective isometric Jordan
triple product preserving selfmap of A with θ ◦ θ = IA, and P = 1
2 (IA + θ). So P
is symmetric. If also P is positive then P , and hence also θ, is ∗-linear. Let Q be
the extension of P to the second dual.
2 (I + θ) on qA.
12
DAVID P. BLECHER AND MATTHEW NEAL
Suppose that θ : A → B is a linear isometric surjection between C∗-algebras. By
a result of Kadison [35], u = θ∗∗(1) is a unitary in B∗∗. Suppose now further that
θ is ∗-linear. Then u is selfadjoint, and uθ∗∗(·) is a unital isometry so selfadjoint.
Thus uθ(a) = θ(a∗)∗u = θ(a)u, for a ∈ A, so u is central. If a ∈ Asa then
uθ(a2) = (uθ(a))2 = θ(a)2 ∈ B,
Returning to our situation, let q = Q(1). This is a central projection in M (A),
since uθ(·) is a Jordan morphism. Thus uB = uθ(A) ⊂ B. So u ∈ M (B).
since u = 2q − 1. Since Q(q⊥) = 0, if a ∈ Ball(A)+ then
P (q⊥a) = P (q⊥aq⊥) ≤ Q(q⊥) = 0,
and so P = 0 on q⊥A. Also, since θ(q) = q and θ is Jordan triple product preserving,
it follows that P (qa) = 1
2 (qa + θ(qa)) = qP (a). Thus, P (qA) = qP (A), and the
restriction of P to qA is a unital bicontractive positive projection on a unital C∗-
algebra. Also θ(qa) = qθ(a) for a ∈ A, as we had above, so θ(qA) = qA. Hence
θ′ = θqA is a unital isometric isomorphism of qA.
(cid:3)
Remark. The example P (x) = 1
2 (x + xT ) on M2 shows the necessity of the
'completely positive' hypothesis in the part it pertains to. Note in this example P
is positive and contractive, and I − P is completely contractive.
We will 'generalize' the bulk of the last result and its proof in Theorem 3.7.
We next show that unlike in the C∗-algebra case, for projections on operator
algebras (completely) 'bicontractive' is not the same as (completely) 'symmetric'.
Also, these both also differ from the notion of (completely) 'hermitian'. We recall
that any operator space X ⊂ B(H) may be unitized to become an operator algebra
as follows. We define
U(X) = (cid:26)(cid:20) λ1IH
0
x
λ2IH (cid:21) : x ∈ X, λ1, λ2 ∈ C(cid:27) ⊂ B(H ⊕ H).
By definition, U(X) may be regarded as a subspace of the Paulsen system (see
1.3.14 and 2.2.10 in [12], or [45]). It follows from Paulsen's lemma (see [45, Lemma
8.1] or [12, Lemma 1.3.15]) that if v : X → X is a linear contraction (resp. complete
contraction), then the mapping θv on U(X) defined by
θv(cid:18)(cid:20) λ1
0
x
λ2 (cid:21)(cid:19) = (cid:20) λ1
0
v(x)
λ2 (cid:21) ,
x ∈ X, λ1, λ2 ∈ C,
is a contractive (resp. completely contractive) homomorphism.
Proposition 3.3. Suppose that X is an operator space, and that P : X → X
is a linear idempotent map. Then P is completely contractive (resp. completely
bicontractive, completely symmetric, completely hermitian) iff the induced map
P : U(X) → U(X) is a real completely positive and completely contractive (resp.
completely bicontractive, completely symmetric, completely hermitian) projection.
In particular these hold (and with the word 'completely' removed everywhere if one
wishes), if P : X → X is a linear idempotent map on a Banach space, when we
give X its minimal or maximal operator space structure (see e.g. [12, 45]).
Proof. If P is completely contractive then by Paulsen's lemma referred to above, the
unique ∗-linear unital extension of P to U(X)+U(X)∗ is completely contractive and
completely positive. Thus by [6, Section 2], P is real completely positive. Clearly
PROJECTIONS ON OPERATOR ALGEBRAS
13
P is a projection. Conversely, if P is completely contractive then so is P . We note
that I− P annihilates the diagonal, and acts as I−P in the 1-2-entry. Thus I−P is
completely contractive iff I− P is completely contractive. Also note that 2 P−I does
nothing to the diagonal, and acts as 2P −I in the 2-1-entry; that is 2 P −I = ^2P − I.
By Paulsen's lemma again, as above, 2 P − I is completely contractive iff 2P − I is
completely contractive. Finally, I − P + eit P multiplies each of the two diagonal
entries by eit, and acts as I − P + eitP in the 1-2-entry. Multiplying by e−it, we
see again by Paulsen's lemma that this is completely contractive iff P is completely
hermitian.
For the last assertion, if P is contractive (resp. bicontractive, symmetric, hermit-
ian) then it is completely contractive (resp. bicontractive, symmetric, hermitian) on
X with its minimal or maximal operator space structure by e.g. (1.10) and (1.12)
in [12]. We may then apply the case in the previous paragraph to obtain the result.
Conversely, if P is contractive (resp. bicontractive, etc)., then so is P since P is the
1-2 corner of P . The rest is clear.
(cid:3)
The previous result provides examples of real completely positive unital com-
pletely bicontractive (resp. completely contractive, completely symmetric) projec-
tions which are not completely symmetric (resp. completely bicontractive, com-
pletely hermitian). As we said, this is in contrast to the C∗-algebra case where
(complete) bicontractivity is equivalent to being (completely) symmetric [49, 29].
Example 3.4. A more specific example of a real completely positive unital com-
pletely bicontractive projection on an operator algebra which is not symmetric
arises by the last construction, from the following explicit completely bicontractive
projection which is not completely symmetric. Let Y be R2, the latter viewed as a
real Banach space whose unit ball is the unit ball of ℓ∞2
in the first quadrant and
the unit ball of ℓ1
2 in the second quadrant. Let X be the standard complexification
of Y . Take P : X → X to be the usual complexification of the 'projection onto the
first coordinate' on Y . We thank Asvald and Vegard Lima for this example, which
is a bicontractive projection on a Banach space which is not symmetric. Giving
X its minimal or maximal operator space structure makes X an operator space,
and makes P (by e.g. (1.10) or (1.12) in [12]) a completely bicontractive projection
which is not symmetric. Hence by Proposition 3.3 we get a real completely posi-
tive unital completely bicontractive projection on an operator algebra which is not
symmetric.
Example 2.8 shows that one cannot extend completely symmetric real com-
pletely positive projections to a bicontractive projection on a containing C∗-algebra.
Things are much better if P is a unital map:
Lemma 3.5. Let A be a unital operator algebra, and P : A → A a completely
symmetric unital projection. Then P is real completely positive, and the range of
P is a subalgebra of A. Moreover, P extends to a completely symmetric unital
projection on C∗e (A) (or on I(A)).
Proof. Any completely contractive unital map on A is real completely positive as we
said in the introduction. Let θ = 2P − I, so that P = 1
2 (I + θ). By Lemma 3.1 θ is
a unital complete isometry. So in fact θ is a homomorphism, by the Banach -- Stone
theorem for operator algebras [12]. This implies as we said in Lemma 3.1, that
Ran(P ) is a subalgebra of A. We can extend θ uniquely to a unital ∗-isomorphism
14
DAVID P. BLECHER AND MATTHEW NEAL
π : C∗e (A) → C∗e (A), with π◦π = I, and it follows that P = 1
symmetric extension of P . It is also completely positive.
2 (I +π) is a completely
Similarly, since one may extend π further to a unital ∗-automorphism of I(A),
there is a completely symmetric unital projection on I(A) extending P .
(cid:3)
Lemma 3.6. Let A be a unital operator algebra, and P : A → A a completely con-
tractive real positive projection which is bicontractive. Then P (1) = q is a projection
(not necessarily central), and all the conclusions of Theorem 2.5 and Proposition
2.7 hold. Also, there exists a unital completely bicontractive (real completely posi-
tive) projection P ′ : qAq → qAq such that P is the zero map on q⊥A + Aq⊥, and
P = P ′ on qAq. We have Ran(P ) = Ran(P ′)
Proof. Suppose that q = P (1). Then q ≥ 0 by Lemma 2.3, so that the closed
algebra B generated by q and 1 is a C∗-algebra. Note that P (qn) = q by Theorem
2.5, so that P (B) ⊂ B. Let Q = PB, this is a bicontractive projection on B,
and it is positive by the proof of Lemma 2.3. By [29, Theorem 2] and its proof,
q = P (1) is a partial isometry in B, hence is a projection. Hence all the conclusions
of Proposition 2.7 and Theorem 2.5 hold. So P (a) = qP (a)q = P (qaq) for a ∈ A,
and P (A) = qP (A)q = P (qAq), and P 'splits' as the sum of the zero map on
q⊥A + Aq⊥, and a unital projection E on qAq.
(cid:3)
Remark. A direct proof that P (1) is a projection in the case of the lemma
above: Let P (1) = q. As we saw in Theorem 2.5, P (qn) = q. Thus P (u(q)) = q,
(we defined u(·) in the introduction). Suppose that c is a positive scalar with
c(q − u(q)) of norm one. Then
(I − P )(c(q − u(q)) − u(q)) = c(q − u(q)) − u(q) − (c(q − q) − q) = (1 + c)(q − u(q))
which has norm > 1. By the contractivity of I − P we have a contradiction, unless
q = u(q). So q = P (1) is a projection.
The following is the solution to the symmetric projection problem in the category
of approximately unital operator algebras.
Theorem 3.7. Let A be an approximately unital operator algebra, and P : A → A
a completely symmetric real completely positive projection. Then the range of P is
an approximately unital subalgebra of A. Moreover, P ∗∗(1) = q is a projection in
the multiplier algebra M (A) (so is both open and closed).
Set D = qAq, a hereditary subalgebra of A containing P (A). There exists a period
2 surjective completely isometric homomorphism θ : A → A such that θ(q) = q, so
that θ restricts to a period 2 surjective completely isometric homomorphism D → D.
Also, P is the zero map on q⊥A + Aq⊥ + q⊥Aq⊥, and
In fact
P =
1
2
(I + θ) on D.
1
2
P (a) =
(a + θ(a)(2q − 1)) ,
a ∈ A.
The range of P is exactly the set of fixed points of θD in D.
Conversely, any map of the form in the last equation, for a period 2 surjective
completely isometric homomorphism θ : A → A and a projection q ∈ M (A) with
θ∗∗(q) = q, is a completely symmetric real completely positive projection.
PROJECTIONS ON OPERATOR ALGEBRAS
15
Proof. Applying Lemma 3.6 to P ∗∗ , we see that P ∗∗(1) = q is a projection, and all
the conclusions of Proposition 2.7 and Theorem 2.5 are true for us. We will silently
be using facts from these results below. In particular q is an open projection, so
supports an approximately unital subalgebra D of A with D⊥⊥ = qA∗∗q (see [10]).
Then θ = 2P − I is a linear completely isometric surjection on A by Lemma 3.1.
So by the Banach -- Stone theorem for operator algebras [12, Theorem 4.5.13], there
exists a completely isometric surjective homomorphism π : A → A and a unitary u
with u, u−1 ∈ M (A) with θ = π(·)u. We have
θ∗∗(1) = 2P ∗∗(1) − 1 = 2q − 1 = π∗∗(1)u = u,
so that u is a selfadjoint unitary (a symmetry), and q ∈ M (A). So qAq = D ⊂ A.
Since P (A) = qP (A)q, the range of P is contained in D, and the range of P is
exactly the set of fixed points of θ, which all lie in D. This implies that Ran(P ) is
a subalgebra of A. It is approximately unital and P is real completely positive by
Corollary 2.9.
We have θ∗∗(q) = q and π∗∗(q) = θ∗∗(q)u = qu = q. Then 2P (a) − a =
π(a)(2q − 1) and P (a) = 1
2 (a + π(a)(2q − 1)). Indeed
π(a) = (2P (a) − a)u = (2P (a) − a)(2q − 1) = 2P (a) − 2aq + a,
a ∈ A.
From this, or otherwise, one sees that π equals θ on D, and π(D) = (2P − I)(D) ⊂
D. However D = θ2(D) ⊂ θ(D) = π(D), so π(D) = D. This completes the main
part of the theorem.
2 (a + π(a)(2q− 1)) is clearly completely
For the converse, note that such P (a) = 1
symmetric on A, and
1
2
1
4
(a+π(a)(2q−1))) =
P (P (a)) = P (
since π is period 2, π∗∗(q) = q, and 2q − 1 is a symmetry. We have P ∗∗(1) =
2 (1 + π∗∗(1)(2q − 1)) = q, so that P is real completely positive by Proposition
2.7.
(a+2π(a)(2q−1)+π(π(a)(2q−1))(2q−1) = P (a),
(cid:3)
1
It follows easily from the last theorem that a completely symmetric real com-
pletely positive projection P on A extends to a completely symmetric projection
P on the C∗-envelope of A. Moreover, P (x) = 1
2 (a + π(a)(2q − 1)), for a ∗-
automorphism π. However, this extension will not in general be positive.
In the work [16] in progress, we prove the Jordan algebra variant of the last
result.
4. The bicontractive projection problem
The following could be compared with Corollary 2.11.
Lemma 4.1. Let A be a unital operator algebra, and P : A → A a unital projection
with I−P completely contractive. Let C = (I−P )(A) = Ker(P ). Then C2 ⊂ P (A).
We also have
(1) C is a subalgebra of A iff C2 = (0).
(2) If P is completely bicontractive (or more generally, if P (aP (b)) = P (P (a)P (b))
for a, b ∈ A) then P (A)C + CP (A) ⊂ C, and C3 ⊂ P (A).
(3) If the conditions in (1) hold, and if P is completely contractive (or more
generally, if P (aP (b)) = P (P (a)P (b)) for a, b ∈ A) then C is actually an
ideal in A.
16
DAVID P. BLECHER AND MATTHEW NEAL
(4) If P (A)C + CP (A) ⊂ C (see (2)), then θ = 2P − IA is a homomorphism
iff P (A) is a subalgebra of A, and then the range of P is the set of fixed
points of this automorphism θ.
Proof. Of course A = C +B, where B = P (A). By Youngson's result [50] applied to
an extension Q of I−P to a completely contractive projection on I(A) (which exists
by an easier variant of the proof in Theorem 2.5), we have (I − P )(wz) = Q(w(I −
P )(1)∗z) = 0 for z, w ∈ C. So P (wz) = wz ∈ B, and this is zero if the kernel is a
subalgebra. In any case, C2 ⊂ B. Assuming that P is completely contractive (or
that P (aP (b)) = P (P (a)P (b))), if z ∈ B and w ∈ C, then P (wz) = P (P (w)z) = 0,
so wz ∈ C. So CB ⊂ C and similarly BC ⊂ C. Thus C3 ⊂ BC ⊂ C. Hence if also
C is a subalgebra, then it is an ideal.
For (4), we may decompose A = C ⊕ B, where 1A ∈ B = P (A), and we have the
relations C2 ⊂ B, CB + BC ⊂ C. Using the latter it is a simple computation that
the period 2 map θ : x + y 7→ x − y for x ∈ B, y ∈ C is a homomorphism (indeed
an automorphism) on A iff P (A) is a subalgebra of A; and clearly P (A) is the fixed
points of θ.
(cid:3)
Remarks. 1) Replacing P by I − P we see that if P : A → A is a completely
contractive projection with P (1) = 0, then P (A) is a subalgebra iff P (A)2 = (0).
2) The kernel of a bicontractive projection need not be a subalgebra. For
example, consider P (f )(x) = 1
2 (f (x) + f (−x)) for f ∈ C([−1, 1]).
The following clarifies the unital version of the 'bicontractive projection problem'
in relation to the existence of an associated period two automorphism.
Corollary 4.2. If P : A → A is a unital idempotent on a unital operator algebra
let θ : A → A be the associated linear period 2 automorphism x + y 7→ x − y for
x ∈ Ran(P ), y ∈ Ker(P ). Then P is completely bicontractive iff kI ± θkcb ≤ 2. If
these hold then the range of P is a subalgebra iff θ is also a homomorphism, and
then the range of P is the set of fixed points of this automorphism θ. Also, P is
completely symmetric iff θ is completely contractive.
Proof. The first and last assertions are obvious, and the second follows by Lemma
4.1 (4).
(cid:3)
For us, the 'bicontractive projection problem' is whether the range of a com-
pletely bicontractive real completely positive projection on an approximately unital
operator algebra A, is an approximately unital subalgebra of A. This is not obvious,
although there are easy counterexamples if one drops some of the hypotheses. For
example consider the projection P on the upper triangular 2 × 2 matrices taking
the matrix with rows [a b] and [0 c] to the matrix with rows [(a − c)/2 0] and
[0 (c − a)/2] . This is completely contractive and extends to a completely contrac-
tive projection on the containing C∗-algebra, and can be shown to be bicontractive
with P (1) = 0, and its range is not a subalgebra.
We now show that the 'bicontractive projection problem' can be reduced to the
case that A is unital and P (1) = 1, by a sequence of three reductions. First, if
P : A → A is a completely bicontractive real completely positive projection on
an approximately unital operator algebra A, then P ∗∗ is a completely bicontractive
real completely positive projection on a unital operator algebra. Thus henceforth in
this section in the next we assume that A is unital. Second, Lemma 3.6 allows us to
PROJECTIONS ON OPERATOR ALGEBRAS
17
reduce further to the case that P (1) = 1: it asserted that P (1) = q is a projection
(not necessarily central), all the conclusions of Theorem 2.5 and Proposition 2.7
hold, and there exists a unital completely bicontractive (real completely positive)
projection P ′ : qAq → qAq such that P is the zero map on q⊥A + Aq⊥, and P = P ′
on qAq, and Ran(P ) = Ran(P ′).
Then the third of our reductions of the completely bicontractive projection prob-
lem puts the problem in a 'standard position'. Of course P (A) is a subalgebra if
and only if Q(D) is a subalgebra, where D is the closed subalgebra of A generated
by P (A), and Q = PD, which is a completely bicontractive unital projection on D.
That is, we may as well replace A by the closed subalgebra generated by P (A).
Thus by these three steps above, we have reduced the completely bicontractive
projection problem on approximately unital operator algebras to the 'standard
position' of a unital projection on a unital operator algebra, whose range generates
A. In this situation we obtain the following structural result.
Corollary 4.3. Let P be a completely bicontractive unital projection on a unital
operator algebra A. Let D be the algebra generated by P (A). Then (I − P )(D) =
Ker(PD) is an ideal in D and the product of any two elements in this ideal is zero.
Proof. We saw at the end of Section 2 that (I − P )(D) = Ker(PD) is an ideal in D,
and in the notation there it equals D ∩ e⊥M e⊥. Since D ∩ e⊥M e⊥ is a subalgebra
of D the result follows from Lemma 4.1.
(cid:3)
Remark. Note that (I − P )(D) ⊂ e⊥A, but P (D) is not a subset of eA.
The above shows that we can also solve the bicontractive projection problem
in the affirmative for real completely positive completely bicontractive projections
P on a unital operator algebra A such that the closed algebra generated by A is
semiprime (that is, it has no nontrivial square-zero ideals):
Corollary 4.4. Let P : A → A be a real completely positive completely bicontractive
projection on a unital operator algebra. If A is an operator algebra containing no
nonzero nilpotents, then P (A) is a subalgebra of A. Also if the closed algebra D
generated by P (A) is semiprime, then P (A) is a subalgebra of A.
Proof. By the second reduction above, we may assume that P is unital. By Corol-
lary 4.3, (I − P )(D) is an ideal in D with square zero, and so is (0) in these cases.
So P (A) = P (D) = D, a subalgebra.
(cid:3)
For the subcategory of uniform algebras (that is, closed subalgebras of C(K), for
compact K) which are unital or approximately unital, the bicontractive projection
problem coincides with the symmetric projection problem, and again there is a
complete solution:
Theorem 4.5. Let P : A → A be a real positive bicontractive projection on a
uniform algebra A, and suppose that A is unital or approximately unital). Then P
is (completely) symmetric, and so we have all the conclusions of Theorem 3.7. In
particular, P (A) is a subalgebra of A, and P is a conditional expectation.
Proof. Here bicontractive projections are the same as completely bicontractive pro-
jections (by e.g. (1.10) in [12]). By the obvious variant of the usual proof that
positive maps into a C(K) space are completely positive, we have that real pos-
itive maps into a uniform algebra are real completely positive. By the first two
18
DAVID P. BLECHER AND MATTHEW NEAL
reductions described above we can assume that A and P are unital. We also know
that B = P (A) is a subalgebra by Corollary 4.4, since e.g. nonzero nilpotents can-
not exist in a function algebra. Thus by Corollary 4.2 the map θ(x + y) = x − y
described there is an algebra automorphism of A, hence a (completely) isometric
isomorphism (since norm equals spectral radius). So P = 1
2 (I + θ) is (completely)
symmetric, and Theorem 3.7 applies.
(cid:3)
Remark. The idea in the last proof that θ is automatically isometric, since it
is an algebraic automorphism of a uniform algebra, and that this implies that P is
symmetric, was found together with Joel Feinstein after submission of our original
paper.
By e.g. Corollary 4.2, to find a counterexample to the conjecture that all (com-
pletely real positive) completely bicontractive unital projections have range which
is a subalgebra, we need a unital operator algebra D = C ⊕ E where C 6= (0), C2 =
0, CE + EC ⊂ C, and 1D ∈ E and E generates D, so in particular E2 is not
a subset of E, and with the projection maps onto C and E completely contrac-
tive. This is easy in a Banach algebra, one may equip ℓ1
3, with the standard basis
identified with symbols 1, a, b satisfying relations like b2 = 0, a2 = b, etc. (setting
C = {b}, E = Span{1, a}). To find an operator algebra example, we make a general
construction.
Example 4.6. Let V be a closed subspace of B(H)⊕ B(H), viewed as elements of
B(H (2)) supported on the 1-1 and 2-2 entries. Write v1 and v2 for the two 'parts'
of an element v ∈ V . Let C be the closed span of {v1w2 : v, w ∈ V }, and we will
assume that C 6= (0). Let B be the set of elements of B(H (3)) of form
λI
0
0
v1
λI
0
c
v2
λI
,
λ ∈ C, v ∈ V, c ∈ C.
Then the copy of C is an ideal in B with square zero, and it is generated by the
copy of V in B. If E is the sum of this copy of V and C IH(3) , then all the conditions
in the last paragraph needed for a counterexample hold, with the exception of the
projection onto E being completely contractive. We remark that one may also
describe B more abstractly as a set B = C 1 + V + C in B(K) where V and C are
closed subspaces of B(K) with the properties that (0) 6= C = V 2, V 3 = C2 = (0)
plus one more condition ensuring that p1V (1− p) = (0) where p1 is the left support
projection of C and p is the left support projection of C + V .
To fix the exception noted in the last paragraph, we consider the subalgebra
A = A(V ) of B(H (7)) consisting of all elements of form
λI
0
0
0
0
0
0
v1
λI
0
0
0
0
0
c
v2
λI
0
0
0
0
0
0
0
λI
0
0
0
0
0
0
0
λI
0
0
0
0
0
2v1
0
λI
0
0
0
0
0
2v2
0
λI
,
λ ∈ C, v ∈ V, c ∈ C.
The square-zero ideal in A consisting of the copy of C we will abusively write again
as C, and again let E be the sum of C IH(7) and the isomorphic copy of V in A(V ),
PROJECTIONS ON OPERATOR ALGEBRAS
19
so that A = C ⊕ E. Here C 6= (0), C2 = 0, CE + EC ⊂ C, and 1D ∈ E and E
generates A, as desired. The canonical projection map from A onto C is obviously
completely contractive.
A particularly simple case is when H is one dimensional, so that A ⊂ M7,
and where V = C I2. This algebra is obviously essentially just (i.e. is completely
isometrically isomorphic to) the subalgebra of M5 of matrices
λ ν
c
0 λ ν
0
0
0
0
0
0 λ 0
0
0
0
0
0
0 λ 2ν
0 λ
0
,
λ, ν, c ∈ C,
with the projection being 'replacing the 1-3 entry' by 0. Thus 5 × 5 matrices of
scalars will suffice to give an interesting example. However we will need the more
general construction later to produce a more specialized counterexample.
Consider the algebra U(V ) constructed from V as described in the paragraph
above Proposition 3.3. We define U0(V ) to be the subalgebra of U(V ) consisting of
elements of U(V ) with the two 'diagonal entries' identical. It is a subalgebra of of
B(H (4)).
Lemma 4.7. In the situation of Example 4.6, the map j : U0(V ) → B(H (3)) given
by
(cid:20) λI
0
v
λI (cid:21) 7→
λI
0
0
is completely contractive and unital.
1
2 v1
λI
0
0
1
2 v2
λI
(cid:3)
1√2
IH(3)
IH(3) ] and its transpose.
Proof. To prove that it is completely contractive simply notice that j may be viewed
as the composition of the canonical map U(V ) → U(W ) where W is the copy of V in
the algebra B above, and the map M2(B) → B given by pre- and postmultiplying
by [ 1√2
Corollary 4.8. If A = A(V ) is the unital operator algebra above in B(H (7)) then
the canonical projection P : A → A which replaces the 1-3 entry by 0, whose kernel
is a nontrivial square-zero ideal C generated by P (A), is a (real completely positive)
completely bicontractive and unital projection, but its range is not a subalgebra, and
it need not even be a Jordan subalgebra. A particularly simple 'case' is the algebra
(completely isometrically isomorphic to the algebra of ) of 5 × 5 scalar matrices
described above Lemma 4.7.
Proof. In the 5 × 5 matrix example it is easily checked that P (A) is not closed
under squares, hence is not a Jordan subalgebra.
It remains to prove that P
is completely contractive. However P is the composition of the canonical map
B(H (3) ⊕ H (4)) → B(H (4)) restricted to A, and the map x 7→ j(x) ⊕ x on U0(V ),
where j is as in Lemma 4.7.
(cid:3)
The following is another rather general condition under which the completely
bicontractive projection problem is soluble. Indeed as we said in the introduction,
all examples known to us of real completely positive completely bicontractive pro-
jections on unital operator algebras, whose range is a subalgebra, do satisfy the
criterion in Theorem 4.9.
20
DAVID P. BLECHER AND MATTHEW NEAL
Suppose further that either C P (A)∗ ⊂ M C
Theorem 4.9. Let A be a unital operator algebra in a von Neumann algebra M
(which could be taken to be B(H), or I(A)∗∗ as above) and let P : A → A be a
unital completely bicontractive projection. Let D be the closed algebra generated by
P (A), and let C = (I − P )(D).
w∗, the weak* closed left ideal in
M generated by C (this is equivalent to saying that the left support projection of
P (A)C∗) is dominated by the right support projection of C). Alternatively, assume
that P (A)∗C is contained in the weak* closed right ideal in M generated by C (or
equivalently that the right support projection of C∗P (A) is dominated by the left
support projection of C). Then P (A) is a subalgebra of A.
Proof. We assume the 'left ideal' condition, the other case is similar and left to the
reader (or can be seen by looking at the opposite algebra' Aop). By replacing A
by D we may assume that P (A) generates A. Suppose that M is a von Neumann
algebra on a Hilbert space H. We set p1 = ∨z∈C r(z)r(z)∗ to be the left support
projection of C, and set p2 = ∨z∈C r(z)∗r(z), the right support projection of C.
Note that zp1 = 0 for z ∈ C, which implies that p2p1 = 0. Let p = p1 + p2, a
projection. Our hypothesis is equivalent to saying that
We may write
C P (A)∗ = C P (A)∗p2.
M = (1 − p)M (1 − p) + (1 − p)M p + pM (1 − p) + pM p.
Thus M may be pictured as the direct sum of a von Neumann algebra with four
corners (thus having a 2 × 2 matrix form). Let y ∈ P (A). Then yC ⊂ C ⊂ pM , so
that (1−p)yp1 = 0. On the other hand, by hypothesis, p2y∗(1−p) = p2y∗p2(1−p) =
0. Thus (1 − p)yp2 = 0 and so (1 − p)yp = 0. Therefore
y = (1 − p)y(1 − p) + py(1 − p) + pyp = y(1 − p) + pyp.
Furthermore p2yC ⊂ p2C = p2p1C = (0), and so p2yp1 = 0. And by hypothesis,
p1yp2 = p1(p2y∗)∗ = p1(p2y∗p2)∗ = 0. So
Thus
p1P (A)p2 = (0).
If also x ∈ P (A) then x = x(1 − p) + p1xp1 + p2xp2, and so
y = y(1 − p) + p1yp1 + p2yp2.
xy = x(1 − p)y(1 − p) + p1xp1yp1 + p2xp2yp2,
and so again p1xyp2 = 0. Since C = (I − P )(A) we see that
(I − P )(xy) = p1(I − P )(xy)p2 = p1xyp2 − p1P (xy)p2 = 0,
so that xy = P (xy) ∈ P (A). So P (A) is a subalgebra.
(cid:3)
Remarks.
1) If C P (A)∗ ⊂ [BC] where B = C∗e (A) then the first hypothesis
in the previous result holds.
2) If A is the counterexample algebra of 5 × 5 scalar matrices described above
Lemma 4.7, it is very illustrative to compute the various associated objects of
interest in our paper. We leave the details to the reader as an exercise. Here
C∗(A) = C∗e (A) = I(A) = M3 ⊕ M2 ⊂ M5. If P is the projection in that example,
namely the map that replaces the 1-3 entry with 0, then C∗(P (A)) = I(P (A)) = 0⊕
PROJECTIONS ON OPERATOR ALGEBRAS
21
M2 ⊂ M5. A completely contractive completely positive projection P on C∗(A) =
I(A) that extends P is the map
x ⊕(cid:20) a b
d (cid:21) 7→
c
a+d
2
c
2
0
0
0
b
2
a+d
2
c
2
0
0
0
b
2
a+d
2
0
0
0
0
0
0
0
0
a b
c
d
,
x ∈ M3.
To see that this is completely contractive is a tiny modification of the proof of
Lemma 4.7. A completely contractive projection extending I−P is is the projection
onto the 1-3 coordinate. The support projection of P defined just before Proposition
2.10 is e = 0 ⊕ I2, which has complement r = I3 ⊕ 0. The projections p1, p2, p from
the proof of Theorem 4.9 are p1 = E11, p2 = E33, p = E11 + E33; and 1 − p =
E22 + E44 + E55. Also, r − p = E22. Note that C P (A)∗p2 6= C P (A)∗, of course.
Indeed this example is an excellent illustration of what is going on in the proof of
Theorem 4.9. Note that if we change the definition of A by replacing either the 2-3
entry or the 3-2 entry then the hypotheses of Theorem 4.9 are satisfied.
potheses of Theorem 4.9 is illustrative:
Examining why the general example described in 4.6 does not satisfy the hy-
it is not hard to see that if it did then
2 ∈ C then we obtain the
k=1 vk
1 wk
v1w2z∗2 = 0 for all v, w, z ∈ V . However if 0 6= Pn
contradiction 0 6= (Pn
2 )∗ = 0.
2 )(Pn
k=1 vk
k=1 vk
1 wk
1 wk
It would be interesting to investigate other conditions that might imply that
P (A) is a subalgebra, particularly when in the 'standard position' (namely P :
A → A is a unital completely bicontractive projection whose range generates A as
an operator algebra). Some which might be worth investigating are if the algebra
C in Theorem 4.9 is a maximal ideal in A, or if C contains the radical of A. Note
that any one of these conditions rules out our counterexamples above.
5. Another condition
We now look at another condition on a completely contractive projection P which
is automatic for bicontractive projections in the C∗-algebra case, namely that the
induced projection on Re(A) is bicontractive. We will not assume that the induced
projection on Re(A) has completely contractive complementary projection I − P .
We are not able to solve the problem yet, but have made some progress towards
the solution.
Lemma 5.1. Let A be a unital operator algebra, and let P : A → A a unital com-
pletely contractive projection such that the induced projection on Re(A) is bicon-
tractive. We also write P for an extension to a unital completely contractive weak*
continuous projection on the von Neumann algebra B∗∗, where B is a C∗-algebra
containing A as a unital-subalgebra (see the argument in the proof of Corollary
2.11). Let e be the support projection of P on B as in [27, p. 129]. If x ∈ A∩e⊥Be⊥
and x = a + ib with a = a∗, b = b∗, then ka+k = ka−k = kb+k = kb−k.
Proof. Suppose that kak ≤ 1. By the Kadison-Schwarz inequality
P (a)∗P (a) ≤ P (a∗a) = P (e⊥a∗ae⊥) ≤ P (e⊥) = 0.
22
DAVID P. BLECHER AND MATTHEW NEAL
So P (a) = 0, and similarly P (b) = 0. Suppose that ka+k > ka−k. Then by the
spectral theorem for a, there exists ǫ > 0 with
ka − ǫ1k = ka+ − a− − ǫ1k < ka+ − a−k = kak.
Thus ka − ǫ1k < k(I − P )(a − ǫ1)k, a contradiction. So ka+k ≤ ka−k. A similar
argument shows that ka−k ≤ ka+k, so that ka+k = ka−k. Similarly (or by replacing
x by ix), kb+k = kb−k.
Let y = Re(2x − x2). The last paragraph shows that kyk = ky+k = ky−k. Now
assume that kak = 1 ≥ kbk. So ka±k = 1. Let ψ be a state with ψ(a−) = 1. Then
ψ(a+) = 0 or else ψ(a) = ψ(a+) + ψ(a−) > 1 which is impossible. By standard
arguments these imply that ψ(a2
+) = 0. Since y = 2a+ − 2a− −
(a2
−), we have
−) = 1 and ψ(a2
−) + (b2
+ + a2
+ + b2
ψ(y) = −3 + ψ(b2
+ + b2
−).
It is well known that for a selfadjoint operator T = T+− T− = R− S with R, S ≥ 0,
we have kT+k ≤ kRk. Thus
kyk = ky+k ≤ k2a+ − a2
+ + (b2
+ + b2
−)k ≤ 2.
(cid:3)
+ +b2
+ +b2
−) we must have ψ(b2
−) = 1, so that kb2k = 1 = kbk.
Since ψ(y) = −3+ψ(b2
Replacing x by ix, we see that kak = kbk.
Lemma 5.2. Let A be a unital operator algebra, and let P : A → A a unital
completely contractive projection such that the induced projection on Re(A) is bi-
contractive. We also write P for an extension to a unital completely contractive
weak* continuous projection on the von Neumann algebra B∗∗, where B is a C∗-
algebra containing A as a unital-subalgebra (as in the last result). Let e be the
support projection of P on B as in [27, p. 129]. If x ∈ A ∩ e⊥Be⊥ and x = a + ib
with a = a∗, b = b∗, and kak = 1, then u(a)2 = u(b)2.
Proof. Since b = b∗, u(b) is a selfadjoint tripotent and u(b)2 is a projection. It is well
known that u(x)∗u(x) = u(x∗x) (to see this note that x(x∗x)n → xu(x∗x), so that
xu(x∗x) = u(x) from which the relation is easy). Hence u(b)2 = u(b2) = u(b2
−).
As we saw in the last proof, if ψ is a state with ψ(a−) = 1 then ψ(b2
−) = 1. Now
−) = 1 iff ψ(u(b)2) = 1 (see [25, Lemma 3.3
ψ(a−) = 1 iff ψ(u(a−)) = 1, and ψ(b2
(i)]). So {u(a−)}′ ∩ S(B) ⊂ {u(b)2}′ ∩ S(B), where S(B) is the state space and the
'prime' is as in [25]. From this, as is well known (and simple to prove), we have that
u(a−) ≤ u(b)2. Similarly, u(a+) ≤ u(b)2, so that u(a−) + u(a+) = u(a)2 ≤ u(b)2.
Similarly, u(b)2 ≤ u(a)2, so u(a)2 = u(b)2.
Lemma 5.3. Let A be a unital operator algebra, and let P : A → A a unital
completely contractive projection such that the induced projection on Re(A) is bi-
contractive. We also write P for an extension to a unital completely contractive
weak* continuous projection on the von Neumann algebra B∗∗, where B is a C∗-
algebra containing A as a unital-subalgebra (as in the last results). Let e be the
support projection of P on B as in [27, p. 129]. Suppose that x ∈ A ∩ e⊥Be⊥ has
norm 1. Then u(x)2 = 0 and x = u(x)+y for some y ∈ B∗∗ with u(x)y = u(x)y∗ =
yu(x) = y∗u(x) = 0. Finally, kRe xk = 1/2.
Proof. Suppose that x ∈ A ∩ e⊥Be⊥, and choose an angle θ so that kRe (eiθx)k
is maximized. By Lemma 5.1 this equals kIm (eiθx)k. Write z = eiθx = a + ib
with a = a∗, b = b∗. Scale z so that kak = 1 (so kbk = 1 by Lemma 5.1). Write
+ +b2
+ +b2
++b2
(cid:3)
PROJECTIONS ON OPERATOR ALGEBRAS
23
a = u(a) + a⊥ and b = u(b) + b⊥. Note that u(a)a⊥ = u(a)(a − u(a)) = 0,
since u(a)a = u(a)3a = u(a)2. Similarly a⊥u(a) = 0, and b⊥u(b) = u(b)b⊥ = 0.
Since u(b)2 = u(a)2 by Lemma 5.2, we have u(a)b⊥ = u(a)u(b)2b⊥ = 0, and
similarly b⊥u(a) = a⊥u(b) = u(b)a⊥ = 0. Hence a⊥ and b⊥ are contractions by the
orthogonality of u(a) and a⊥, and of u(b) and b⊥. Consider
1 + i
√2
(a + ib) =
a − b
√2
+ i
a + b
√2
.
and
√2
u(a) + u(b)
u(a) − u(b)
= u(a)2 a − b
√2
By the maximality property of θ we must have k a−b√2 k = k a+b√2 k ≤ 1. Now
,
= u(a)2 a + b
√2
so these are contractions. Squaring each of these we see that u(a)2 − 1
2 (u(a)u(b) +
u(b)u(a)) and u(a)2 + 1
2 (u(a)u(b) + u(b)u(a)) are contractions. Since u(a)2 is a
projection, hence an extreme point, we deduce that u(a)u(b) + u(b)u(a) = 0, or
u(a)u(b) = −u(b)u(a). Using this, if w1 = 1
2 (u(a) + iu(b)) then a simple compu-
tation shows that w1w∗1w1 = w1, so that w1 is a partial isometry. Let w = z/2.
Clearly kwk ≤ 1, but now we see that
√2
1 = kw1k = ku(a)2wk ≤ kwk.
Let w2 = 1
So kwk = 1 and kzk = kxk = 2. This proves the last assertion of the theorem, since
kRe(x)k = kIm(x)k ≤ 1 by the maximality property of θ, but they clearly cannot
be strict contractions since kxk = 2. So henceforth we may assume that θ = 0 and
z = x.
2 (a⊥ + i b⊥), so that w = w1 + w2, and w1w2 = w2w1 = w1w∗2 =
w∗2w1 = 0. Note that u(w) = w1+u(w2) 6= 0, and u(e−iθw) = e−iθu(w), so kxk = 2.
Also ww∗ = w1w∗1 + w2w∗2. Suppose ψ is a state with ψ(w2w∗2) = 1. Then since
kww∗k ≤ 1 we must have ψ(w1w∗1) = 0, which forces ψ(u(a)2) = ψ(u(b)2) = 0.
Thus ψ /∈ {u(a2)}′ = {a2}′ (see [25, Lemma 3.3 (i)]), so that ψ(a2
) 6= 1 since
⊥
a2 = u(a2) + a2
⊥
. On the other hand,
b2
⊥
4
1 = ψ(
a2
⊥
4
+
+ i(
b⊥a⊥
4 −
a⊥b⊥
4
)).
We deduce the contradiction that ψ(a2
) = 1. This contradiction shows that 1 −
⊥
w2w∗2 is strictly positive so that u(w2w∗2) = 0. Hence u(ww∗) = u(w1w∗1) = w1w∗1,
and so u(w) = limn (ww∗)n w = w1w∗1w = w1.
Finally, suppose that x ∈ A ∩ e⊥Be⊥ has norm 1 (so that x may be taken to
be our previous w). Then u(x)2 = w2
1 = 0. That x = u(x) + y, where u(x) is
orthogonal to y and y∗, follows because w = u(w) + w2 and u(w)w2 = u(w)w∗2 =
w2u(w) = w∗2u(w) = 0, the latter because u(w) is a linear combination of the
selfadjoint u(a), u(b), which are each orthogonal to a⊥ and b⊥.
(cid:3)
Corollary 5.4. If the conditions of the previous lemmas hold, and also A∩e⊥Be⊥ =
(0), then P (A) is a subalgebra of A.
Proof. For x, y ∈ P (A) we have exye = eP (xy)e by Proposition 2.10. So xy −
P (xy) ∈ e⊥B∗∗e⊥ ∩ A = (0). Thus P (A) is closed under products.
Corollary 5.5. If the conditions of the previous lemmas hold, and also B is com-
mutative, then A ∩ e⊥Be⊥ = (0), and P (A) is a subalgebra of A.
(cid:3)
24
DAVID P. BLECHER AND MATTHEW NEAL
Proof. By the proof of Lemma 5.3, if x ∈ A ∩ e⊥Be⊥, and eiθx = a + ib with
a = a∗, b = b∗ and kak = 1, we obtained u(a)u(b) = −u(b)u(a) = 0, and u(a) =
u(a)u(b)2 = 0. This is impossible since kak = 1, so A ∩ e⊥Be⊥ = (0). Then apply
Corollary 5.4.
(cid:3)
As in Section 4, to show P (A) is a subalgebra of A, we may replace A by D,
the closed algebra generated by P (A). After this is done, in the previous lemmas
A ∩ e⊥Be⊥ becomes (I − P )(D).
Theorem 5.6. Let P be a unital completely contractive projection on A such that
I − P is contractive on Re(A). Suppose that A is a subalgebra of MN for some
N ∈ N and let D be the closed algebra generated by P (A). Then every element of
(I − P )(D) is nilpotent. Furthermore, if D is semisimple then the range of P is a
subalgebra of A.
Proof. Note that (I − P )(D) is an ideal of D by Corollary 2.11. Suppose x ∈
(I − P )(D) has norm 1. Suppose that x is not nilpotent. Set y1 = x. By Lemma
5.3, x = u(x) + x′ where u(x)2 = 0 and x′ ⊥ u(x). Furthermore x2 = (x′ )2 lies in D
and x2 ⊥ u(x). Similarly, since x2 6= 0 we set y2 = x2/kx2k then y2 = u(y2) + (y2)′
where u(y2) and (y2)′ are perpendicular to each other, and u(y2) is perpendicular
to u(x) (this is because u(y2) is a limit of products beginning and ending with x2,
and e.g. x2u(x) = (x′ )2u(x) = 0. Continuing in this way we obtain an infinite
sequence of norm 1 elements yk such that that u(yn) ⊥ u(yk) for k ≤ n. It is well
known that u(y) 6= 0 if kyk = 1. This contradicts finite dimensionality. So x is
nilpotent. Since (I − P )(D) is an ideal of D consisting of nilpotents, it follows that
it lies in the Jacobson radical of D. Thus if D is semisimple then (I − P )(D) = (0),
so that P (A) = D as before.
(cid:3)
For the following we no longer assume A is finite dimensional but retain the
other assumptions of the above Theorem.
Lemma 5.7. Let P be a unital completely contractive projection on A such that
I − P is contractive on Re(A), and let D be the closed algebra generated by P (A).
If x ∈ (I − P )(D) and x = 1, then x2n
22n . Also, x is quasi-regular (that
is, quasi-invertible).
Proof. Let x = a+ib as in previous Lemmas. By Lemmas 5.1 and 5.2, a = b =
1/2. Hence Re(x2) = a2 − b2 ≤ 1/4. It follows again from Lemma 5.1 that
x2 ≤ 1/2. The first result now follows by considering normalizations of further
powers of 2, and using mathematical induction. It is easily seen that
≤ 2
∞
Xk=1
xk ≤ 1 +
∞
Xm=1
2m x2m
≤ 1 +
∞
Xm=1
2m 2
22m < ∞.
It follows that P∞k=1 xk converges, so that 1 − x is invertible.
Remark. It is still open whether the ideal (I − P )(D) above consists entirely
of quasi-regular elements. If this is the case, then the above Theorem 5.6 holds
for arbitrary unital operator algebras. Note too that the assertion about quasi-
regulars in Lemma 5.7 does follow from Lemma 5.1. That result shows that the
ideal (I − P )(D) in D has no nonzero real positive elements (for in the language of
that result, if a− = 0 then a+ = b+ = b− = 0). The ideas in the proof of Corollary
6.9 of [19] then also show that if x ∈ Ball((I − P )(D)) then x is quasi-regular.
(cid:3)
PROJECTIONS ON OPERATOR ALGEBRAS
25
6. Jordan morphisms and Jordan subalgebras of operator algebras
We recall that a Jordan homomorphism T : A → B is a linear map satisfying
T (ab + ba) = T (a)T (b) + T (b)T (a) for a, b ∈ A, or equivalently, that T (a2) = T (a)2
for all a ∈ A (the equivalence follows by applying T to (a + b)2). By a Jordan
operator algebra we shall simply mean a norm-closed Jordan subalgebra A of an
operator algebra, namely a norm-closed subspace closed under the 'Jordan product'
1
2 (ab + ba), or equivalently with a2 ∈ A for all a ∈ A (that is, A is closed under
squares).
It is natural to ask if the completely bicontractive algebra problem studied in
Section 4 becomes simpler if the range of the projection P : A → A is also a Jordan
subalgebra (that is P (a)2 ∈ P (A)) for all a ∈ A. We next dispense of this question:
Example 6.1. Let y = E21 ⊕ E12 ∈ M4, and let
0 0
1
0 0 −1 0
0
0 0
0 0
0
x =
.
0
0
0
Then xy = −yx, so that if F is the span of x and y then F is closed under squares.
However F is not an algebra since xy /∈ F . Let V = {z⊕ z ∈ M8 : z ∈ F}, and form
the algebra A = A(V ) described in Example 4.6. By Corollary 4.8 the canonical
projection P : A → A which replaces the 1-3 entry of a matrix in A(V ) by 0, is
a (real completely positive) completely bicontractive and unital projection, but its
range is not a subalgebra. However its range is a Jordan subalgebra; P (A) is closed
under squares since z2 = 0 for z ∈ F . Thus the completely bicontractive algebra
problem does not become simpler if the range of the projection P : A → A is also
a Jordan subalgebra.
The following variant of the 'Banach -- Stone theorem for C∗-algebras' will be
evident to 'JB-experts'.
Lemma 6.2. Let A be a unital C∗-algebra, and T : A → B(H) a unital complete
isometry such that T (A) is closed under taking squares (thus, T (A) is a Jordan
algebra). Then T (A) is a C∗-subalgebra of B(H), and T is a ∗-homomorphism.
Proof. Since such T is necessarily ∗-linear as we said in the introduction, T (A)
is a JB*-algebra, hence a selfadjoint JB*-triple (see e.g. [21]). By the theory of
JB*-triples T is a Jordan homomorphism.
look at
the selfadjoint part and use the fact that isometries in that category are Jordan
homomorphisms [34]; or it can be deduced using the C∗-envelope as in the next
In particular for each x ∈ Asa we have T (x2) = T (x)2, so by Choi's
proof).
multiplicative domain result (see e.g. [12, Proposition 1.3.11]) we have T (xy) =
T (x)T (y) for all y ∈ A. So T is a homomorphism and T (A) is a C∗-subalgebra. (cid:3)
It is natural to ask if the analogous result is true for operator algebras. That
is, if B is a closed unital Jordan subalgebra of an operator algebra A, and if B is
unitally and linearly completely isometric to another unital operator algebra, then
is B actually a subalgebra of A? If the algebra is also commutative this is true and
follows from the next result.
(Two other proofs of this:
26
DAVID P. BLECHER AND MATTHEW NEAL
Lemma 6.3. Let A be a unital operator algebra, and let T : A → B a unital
complete isometry onto a unital Jordan operator algebra. Then T is a Jordan
homomorphism, and T (an) = T (a)n for every n ∈ N and a ∈ A.
Proof. Note that T (a)3 is the Jordan product of T (a) and T (a)2, so T (A) is closed
under cubes. Similarly it is closed under every power. By the property of the
C∗-envelope mentioned in the introduction, there exists a ∗-homomorphism π :
C∗(T (A)) → C∗e (A) with π ◦ T = IA. So an = π(T (a)n) = π(T (an)). Since πT (A)
is one-to-one, the results follow.
(cid:3)
Remark. In the proof of the last result one could have instead used [2, Corollary
2.8].
We now answer the question above Lemma 6.3 in the negative:
Example 6.4. Let P : A → A be a completely contractive projection on an
operator algebra A on H whose kernel is an ideal I (see e.g. Corollary 2.11 or Lemma
4.1). Then it is known that B = A/I is an operator algebra (see [12, Proposition
2.3.4]), and the induced map P : B → P (A) is a completely isometric isomorphism,
and P will be unital if A and P are unital. If these hold, and in addition P (A) is
a Jordan subalgebra of A which is not a subalgebra, then T = P : B → A ⊂ B(H)
is a unital complete isometry such that T (B) is closed under taking squares (thus,
T (B) is a unital Jordan subalgebra), but T (B) is not a subalgebra, and T is not an
algebra homomorphism. In particular, we can take A to be the algebra in Example
6.1.
We finish our paper with another 'Banach -- Stone type theorem for operator al-
gebras':
Proposition 6.5. Suppose that T : A → B is a completely isometric surjection
between approximately unital operator algebras. Then T is real (completely) positive
if and only if T is an algebra homomorphism.
Proof. If T is an algebra homomorphism then by Meyer's theorem [12, Theorem
2.1.13] T extends to a unital completely isometric surjection between the unitiza-
tions, which then extends by Wittstock's extension theorem to a unital completely
contractive, hence completely positive, map on a generated C∗-algebra. So T is
real completely positive.
Conversely, suppose that T is real positive. We may assume that A and B
are unital, since T ∗∗ is a real positive completely isometric surjection between
unital operator algebras. By the Banach -- Stone theorem for operator algebras [12,
Theorem 4.5.13], there exists a completely isometric surjective homomorphism π :
A → B and a unitary u with u, u−1 ∈ B with T = π(·)u. The restriction of T
to C 1 is real positive, hence positive (see [6, Section 2]). Thus u ≥ 0, and so
u = (u2)
2 = 1. Hence T is an algebra homomorphism.
1
(cid:3)
Remark. One may also prove a limited version of this result for algebras with
no kind of approximate identity by using the ideas in the proof of Proposition 2.1.
There is also a Jordan variant of the last result, a simple adaption of the main
theorem in [2]. Here we just state the unital case (see [16] for more on this topic).
PROJECTIONS ON OPERATOR ALGEBRAS
27
Proposition 6.6. Suppose that T : A → B is an isometric surjection between
unital Jordan operator algebras. Then T is real positive if and only if T is a Jordan
algebra homomorphism.
Proof. If T is a Jordan algebra homomorphism and u = T (1A) and T (v) = 1B then
2u = u1B + 1Bu = 2T (1A · v) = 2T (v) = 2 · 1B. So u = 1B. However a unital
contractive map is real positive [6, Section 2].
The converse follows by the same proof as for Proposition 6.5, but using the
form of the Banach -- Stone theorem for operator algebras in [2, Corollary 2.10]. By
that result T (1) is a unitary u with u, u−1 ∈ B. Moreover T (·)u−1 is an isometric
surjection onto B, so by the same result it it is a Jordan homomorphism π. We
have T = π(·)u, and we finish as before.
(cid:3)
Acknowledgements. Some of this material was presented at the AMS Special
Session on Operator Algebras and Their Applications: A Tribute to Richard V.
Kadison, January 2015. In the first authors article [9] for the conference proceedings
for this AMS Special Session there is a very brief survey of some of the material in
the present paper. Corollary 4.2 and part of Theorem 4.5 were stated in [9] first,
and later added to the present paper. We thank Joel Feinstein for a conversation
during which we together arrived at an insight needed in these results. We are also
indebted to the referee for suggesting that we consider the Jordan algebra version
of the topics of this paper; this insight led to the work [16] in progress.
References
[1] C. A. Akemann, Left ideal structure of C ∗-algebras, J. Funct. Anal. 6 (1970), 305-317.
[2] J. Arazy and B. Solel, Isometries of nonselfadjoint operator algebras, J. Funct. Anal. 90
(1990), no. 2, 284 -- 305.
[3] W. B. Arveson, Subalgebras of C ∗
[4] T. Barton and R. Timoney, Weak*-continuity of Jordan triple products and its applications
−algebras, Acta Math. 123 (1969), 141 -- 224.
Math. Scand. 59 (1986), 177 -- 191.
[5] M. Battaglia, Order theoretic type decomposition of JBW*-triples, Quart. J. Math. Oxford,
42 (1991), 129 -- 147.
[6] C. A. Bearden, D. P. Blecher and S. Sharma, On positivity and roots in operator algebras,
Integral Equations Operator Theory 79 (2014), no. 4, 555 -- 566.
[7] D. P. Blecher, Are operator algebras Banach algebras? In: Banach algebras and their appli-
cations, p. 5358, Contemp. Math., 363, Amer. Math. Soc., Providence, RI, 2004.
[8] D. P. Blecher, Noncommutative peak interpolation revisited, Bull. London Math. Soc. 45
(2013), 1100 -- 1106.
[9] D. P. Blecher, Generalization of C*-algebra methods via real positivity for operator and
Banach algebras, Preprint 2015, to appear.
[10] D. P. Blecher, D. M. Hay, and M. Neal, Hereditary subalgebras of operator algebras, J.
Operator Theory 59 (2008), 333-357.
[11] D. P. Blecher and L. E. Labuschagne, Logmodularity and isometries of operator algebras,
Trans. Amer. Math. Soc. 355 (2002), 1621 -- 1646.
[12] D. P. Blecher and C. Le Merdy, Operator algebras and their modules -- an operator space
approach, Oxford Univ. Press, Oxford (2004).
[13] D. P. Blecher and B. Magajna, Duality and operator algebras II: Operator algebras as Banach
algebras, Journal of Functional Analysis 226 (2005), 485 -- 493.
[14] D. P. Blecher and M. Neal, Open projections in operator algebras I: Comparison Theory,
Studia Math. 208 (2012), 117 -- 150.
[15] D. P. Blecher and M. Neal, Open projections in operator algebras II: compact projections,
Studia Math. 209 (2012), 203 -- 224.
[16] D. P. Blecher and M. Neal, Contractive projections on Jordan operator algebras, in prepara-
tion.
28
DAVID P. BLECHER AND MATTHEW NEAL
[17] D. P. Blecher and N. Ozawa, Real positivity and approximate identities in Banach algebras,
Pacific J. Math. 277 (2015), 1 -- 59.
[18] D. P. Blecher and C. J. Read, Operator algebras with contractive approximate identities, J.
Functional Analysis 261 (2011), 188 -- 217.
[19] D. P. Blecher and C. J. Read, Operator algebras with contractive approximate identities II,
J. Functional Analysis 264 (2013), 1049 -- 1067.
[20] D. P. Blecher and C. J. Read, Order theory and interpolation in operator algebras, Studia
Math. 225 (2014), 61 -- 95.
[21] M. Cabrera Garcia and A. Rodriguez Palacios, Non-associative normed algebras. Vol. 1,
The Vidav-Palmer and Gelfand-Naimark theorems. Encyclopedia of Mathematics and its
Applications, 154. Cambridge University Press, Cambridge, 2014.
[22] M.-D. Choi and E. G. Effros, The completely positive lifting problem for C-algebras, Ann. of
Math. 104 (1976), 585 -- 609.
[23] M.-D. Choi and E. G. Effros, Injectivity and operator spaces, J. Funct. Anal. 24 (1977),
156 -- 209.
[24] C.H. Chu, M. Neal and B. Russo, Normal contractive projections preserve type, J. Operator
Theory 51 (2004), 281 -- 301.
[25] C. M. Edwards and G. T. Ruttimann, Compact tripotents in bi-dual JB*-triples, Math. Proc.
Camb. Philos. Soc. 120 (1996), 155 -- 173.
[26] E. G. Effros, Order ideals in a C ∗-algebra and its dual, Duke Math. J. 30 (1963), 391 -- 411.
[27] E. G. Effros and E. Størmer, Positive projections and Jordan structure in operator algebras,
Math. Scand. 45 (1979), 127 -- 138.
[28] Y. Friedman and B. Russo, Contractive projections on C0(K), Trans. AMS 273 (1982),
57 -- 73.
[29] Y. Friedman and B. Russo, Conditional expectation without order, Pacific J. Math. 115
(1984), 351 -- 360.
[30] Y. Friedman and B. Russo, Solution of the contractive projection problem, J. Funct. Anal.
60 (1985), 56 -- 79.
[31] G. Godefroy, N. J. Kalton and P. D. Saphar, Unconditional ideals in Banach spaces, Studia
Math. 104 (1993), 13 -- 59.
[32] M. Hamana, Triple envelopes and Silov boundaries of operator spaces, Math. J. Toyama
University 22 (1999), 77 -- 93.
[33] L. A. Harris, A generalization of C-algebras, Proc. London Math. Soc. (3) 42 (1981), no. 2,
331 -- 361.
[34] J. M. Isidro and A. Rodriguez-Palacios, Isometries of JB-algebras, Manuscripta Math. 86
(1995), no. 3, 337 -- 348.
[35] R. V. Kadison, Isometries of operator algebras, Ann. of Math. 54 (1951), 325 -- 338.
[36] R. V. Kadison, A generalized Schwarz inequality, Ann. of Math. 56 (1952), 494 -- 503.
[37] A. T. Lau and R. J. Loy, Contractive projections on Banach algebras, J. Funct. Anal. 254
(2008), 2513 -- 2533.
[38] M. Neal, Inner ideals and facial structure of the quasi-state space of a JB-algebra, J. Funct.
Anal. 173 (2000), 284 -- 307.
[39] M. Neal, E. Ricard and B. Russo, Classification of contractively complemented Hilbertian
operator spaces, J. Funct. Anal. 237 (2006), 589 -- 616.
[40] M. Neal and B. Russo, Projecteurs contractifs et espaces operateurs, C. R. Acad. Sci. Paris,
t. 331, Serie I (2000), 873 -- 878.
[41] M. Neal and B. Russo, Contractive projections and operator spaces, Trans. Amer. Math. Soc.
355 (2003), 2223 -- 2262.
[42] M. Neal and B. Russo, Operator space characterizations of C*-algebras and ternary rings,
Pacific J. Math. 209 (2003), 339 -- 364.
[43] M. Neal and B. Russo, Representation of contractively complemented Hilbertian operator
spaces on the Fock space, Proc. Amer. Math. Soc. 134, No. 2, (2005), 475-485
[44] M. Neal and B. Russo, Existence of contractive projections on preduals of JBW-triples, Israel
J. Math. 182 (2011), 293 -- 331.
[45] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Ad-
vanced Math., 78, Cambridge University Press, Cambridge, 2002.
[46] G. K. Pedersen, C*-algebras and their automorphism groups, Academic Press, London (1979).
PROJECTIONS ON OPERATOR ALGEBRAS
29
[47] A. G. Robertson and M.A. Youngson, Positive projections with contractive complements on
Jordan algebras, J. London Math. Soc. (2) 25 (1982), 365 -- 374.
[48] B. Russo, Structure of JB*-triples, In: Jordan Algebras, Proc. of Oberwolfach Conference
1992, Ed. W. Kaup, et al., de Gruyter (1994), 209 -- 280.
[49] E. Størmer, Positive projections with contractive complements on C ∗-algebras, J. London
Math. Soc. (2) 26 (1982), 132 -- 142.
[50] M. A. Youngson, Completely contractive projections on C-algebras, Quart. J. Math. Oxford
Ser. (2) 34 (1983), 507 -- 511.
Department of Mathematics, University of Houston, Houston, TX 77204-3008
E-mail address, David P. Blecher: [email protected]
Department of Mathematics, Denison University, Granville, OH 43023
E-mail address: [email protected]
|
1405.0431 | 4 | 1405 | 2016-03-15T11:06:34 | A noncommutative martingale convexity inequality | [
"math.OA",
"math.FA"
] | Let $\mathcal{M}$ be a von Neumann algebra equipped with a faithful semifinite normal weight $\phi$ and $\mathcal{N}$ be a von Neumann subalgebra of $\mathcal{M}$ such that the restriction of $\phi$ to $\mathcal{N}$ is semifinite and such that $\mathcal{N}$ is invariant by the modular group of $\phi$. Let $\mathcal{E}$ be the weight preserving conditional expectation from $\mathcal{M}$ onto $\mathcal{N}$. We prove the following inequality: \[\|x\|_p^2\ge\bigl \|\mathcal{E}(x)\bigr\|_p^2+(p-1)\bigl\|x-\mathcal{E}(x)\bigr\|_p^2, \qquad x\in L_p(\mathcal{M}),1<p\le2,\] which extends the celebrated Ball-Carlen-Lieb convexity inequality. As an application we show that there exists $\varepsilon_0>0$ such that for any free group $\mathbb{F}_n$ and any $q\ge4-\varepsilon_0$, \[\|P_t\|_{2\to q}\le1\quad\Leftrightarrow\quad t\ge\log{\sqrt{q-1}},\] where $(P_t)$ is the Poisson semigroup defined by the natural length function of $ \mathbb{F}_n$. | math.OA | math |
The Annals of Probability
2016, Vol. 44, No. 2, 867–882
DOI: 10.1214/14-AOP990
c(cid:13) Institute of Mathematical Statistics, 2016
A NONCOMMUTATIVE MARTINGALE CONVEXITY
INEQUALITY1
By ´Eric Ricard∗ and Quanhua Xu†,‡
Universit´e de Caen Basse-Normandie∗,
Wuhan University† and Universit´e de Franche-Comt´e‡
Let M be a von Neumann algebra equipped with a faithful semifi-
nite normal weight φ and N be a von Neumann subalgebra of M
such that the restriction of φ to N is semifinite and such that N is
invariant by the modular group of φ. Let E be the weight preserv-
ing conditional expectation from M onto N . We prove the following
inequality:
kxk2
p ≥ kE (x)k2
p + (p − 1)kx − E (x)k2
p,
x ∈ Lp(M), 1 < p ≤ 2,
which extends the celebrated Ball–Carlen–Lieb convexity inequality.
As an application we show that there exists ε0 > 0 such that for any
free group Fn and any q ≥ 4 − ε0,
kPtk2→q ≤ 1 ⇔ t ≥ log pq − 1,
where (Pt) is the Poisson semigroup defined by the natural length
function of Fn.
1. Introduction. Let M be a von Neumann algebra equipped with a
faithful semifinite normal weight φ. The associated noncommutative Lp-
spaces will be simply denoted by Lp(M). We refer to [11] for information on
noncommutative integration. Recall that if N is a von Neumann subalgebra
of M such that the restriction of φ to N is semifinite and such that N
is σφ-invariant [i.e., σφ
t (N ) = N for all t ∈ R], then there exists a unique
φ-preserving conditional expectation E from M onto N such that
E(axb) = aE(x)b,
a, b ∈ N , x ∈ M.
Here σφ denotes the modular group of φ. Moreover, E extends to a contrac-
tive projection from Lp(M) onto Lp(N ) for any 1 ≤ p < ∞. Below is our
main result.
Received May 2014; revised November 2014.
1Supported by ANR-2011-BS01-008-01 and NSFC Grant No. 11271292.
AMS 2000 subject classifications. Primary 46L51, 47A30; secondary 60G42, 81S25.
Key words and phrases. Noncommutative Lp-spaces, martingale convexity inequality,
hypercontractivity, free groups.
This is an electronic reprint of the original article published by the
Institute of Mathematical Statistics in The Annals of Probability,
2016, Vol. 44, No. 2, 867–882. This reprint differs from the original in pagination
and typographic detail.
1
2
E. RICARD AND Q. XU
Theorem 1. Let M, N and E be as above. If 1 < p ≤ 2, then
x ∈ Lp(M).
p + (p − 1)kx − E(x)k2
p,
p ≥ kE(x)k2
(1)
If 2 < p < ∞, the inequality is reversed.
kxk2
Inequality (1) is a martingale convexity inequality. It is closely related
to the celebrated convexity inequality of Ball, Carlen and Lieb [2] for the
Schatten classes Sp. Namely, for 1 < p ≤ 2, we have
p + 2(p − 1)kyk2
(2)
p,
p + kx − yk2
p ≥ 2kxk2
kx + yk2
x, y ∈ Sp.
In fact, it is easy to see that (2) is a special case of (1) by considering
M = B(ℓ2) ⊕ B(ℓ2). Conversely, the validity of (2) for any noncommutative
Lp-spaces implies (1). Indeed, we will deduce (1) from the following:
Theorem 2. Let M be any von Neumann algebra. If 1 < p ≤ 2, then
kx + yk2
p + kx − yk2
(3)
If 2 < p < ∞, the inequality is reversed.
p ≥ 2kxk2
p + 2(p − 1)kyk2
p,
x, y ∈ Lp(M).
What is new and remarkable in (2) or (3) is the fact that (p − 1) is the
best constant. In fact, if one allows a constant depending on p in place of
(p − 1), then (3) is equivalent to the well-known results on the 2-uniform
convexity of Lp(M). We refer to [2] for more discussion on this point. The
optimality of the constant (p − 1) has important applications to hypercon-
tractivity in the noncommutative case. It is the key to the solution of Gross's
longstanding open problem about the optimal hypercontractivity for Fermi
fields by Carlen and Lieb [4]. It plays the same role in [3] and [9]. Note that
the optimality of (p− 1) in (3) implies (p− 1) is also the best constant in (1).
It seems that (1) with this best constant is new even in the commutative
case.
Clearly, (2) implies (3) for injective M (or more generally, QWEP M)
since then Lp(M) is finitely representable in Sp. The proof of (2) in [2]
goes through a differentiation argument for the function t 7→ kx + tykp
p with
self-adjoint x and y. It seems difficult to directly extend their argument to
finite von Neumann algebras. The subtle point is the fact that to be able
to differentiate the above function, one needs the invertibility of x + ty for
all t ∈ [0, 1] except possibly countably many of them. This invertibility is
easily achieved in the matrix algebra case, that is, for M = Mn, the algebra
of n × n-matrices. Instead, we will use a pseudo-differentiation argument
which is much less rigid than that of [2]. The main novelty in our argument
can be simply explained as follows. We first cut the operator x + ty by its
spectral projections in order to reduce the general case to the invertible
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
3
one; to do so we need x + ty to be of full support for all t ∈ [0, 1]. We then
get this full support property for all t by adding to x + ty an independent
operator with diffuse spectral measure. Note that by standard perturbation
argument, it is easy to insure the full support (or even the invertibility) of
x + ty for one t.
An iteration of Theorem 1 immediately implies the following inequality
on noncommutative martingales.
Corollary 3. Let (Mn)n≥0 be an increasing sequence of von Neumann
subalgebras of M with w*-dense union in M. Assume that each Mn is
σφ-invariant, and φMn is semifinite. Let En be the conditional expectation
with respect to Mn. Then for 1 < p ≤ 2,
kxk2
p ≥ kE0(x)k2
p + (p − 1)Xn≥1
kEn(x) − En−1(x)k2
p,
x ∈ Lp(M).
For 2 < p < ∞, the inequality is reversed.
Another possible iteration is the following:
Corollary 4. Let (Mn)1≤n≤N be a family of von Neumann subalgebras
of M. Assume that each Mn is σφ-invariant and φMn is semifinite. Let E +
n
be the conditional expectation with respect to Mn and E −
n . Then
for 1 < p ≤ 2,
n = Id − E +
p ≥ X(εi)∈{+,−}N
kxk2
(p − 1){iεi=−}(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
NYi=1
E εi
i !(x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
For 2 < p < ∞, the inequality is reversed.
2
,
p
x ∈ Lp(M).
Applying it to the case where M = L∞({±1}N ) and Mn is the subal-
gebra of functions independent of the nth variable, we deduce the classical
hypercontractivity for the Walsh system (with operator valued coefficients).
Similarly, taking M to be the Clifford algebra with N generators, we obtain
the optimal hypercontractivity for Fermi fields as pointed out in [2, 4].
We end the paper with some applications to hypercontractivity for group
von Neumann algebras. In particular for the Poisson semigroup of a free
group, we obtain the optimal time for the hypercontractivity from L2 to Lq
for q ≥ 4.
2. The proofs. We will prove Theorems 1 and 2. Using the Haagerup
reduction theorem as in [7], one can reduce both theorems to the finite
case. Thus throughout this section M will denote a von Neumann algebra
4
E. RICARD AND Q. XU
equipped with a faithful tracial normal state τ . Lp(M) is then constructed
with respect to τ . We will first prove (3), then deduce (1) from it. 1 < p < 2
will be fixed in the sequel.
As explained before, the proof of (3) will be done by a pseudo-differentiation
argument. Recall that for a continuous function f from an interval I to R
its pseudo-derivative of second order at t is
D2f (t) = lim inf
h→0+
f (t + h) + f (t − h) − 2f (t)
h2
.
This pseudo-derivative shares many properties of the second derivative. For
instance, if D2f is nonnegative on I, then f is convex. Indeed, by adding εt2
to f (with ε > 0), we can assume that D2f (t) is positive for all t. If f was
not convex, there would exist t0 < t1 in I such that the function f − g takes
a positive value at some point of (t0, t1), where g is the straight line joining
the two points (t0, f (t0)) and (t1, f (t1)). So f − g achieves a local maximum
at a point s ∈ (t0, t1). Consequently, D2f (s) = D2(f − g)(s) ≤ 0, which is a
contradiction.
Our pseudo-differentiation argument consists in proving the following in-
equality for x, y ∈ Lp(M):
(D2
x,y)
D2kx + tyk2
p(0) ≥ 2(p − 1)kyk2
p.
Here the differentiation is, of course, taken with respect to the variable t.
The arguments from [2] can be adapted to give:
Lemma 5. Let a, b ∈ M be self-adjoint elements with a invertible. Then
a,b) holds.
(D2
Proof. As a is invertible in M, a + tb is also invertible for small t.
Introduce an auxiliary function ψ on R,
ψ(t) = ka + tbkp
p = τ ((a2 + t(ab + ba) + t2b2)p/2).
ψ is differentiable in a neighborhood of the origin and
ψ′(t) =
p
2
τ [(a2 + t(ab + ba) + t2b2)p/2−1((ab + ba) + 2tb2)].
As in [2] by functional calculus, the operator (a2 + t(ab + ba) + t2b2)p/2−1
admits the following integral representation:
(4)
(a2 + t(ab + ba) + t2b2)p/2−1
= cpZ ∞
0
sp/2−1
1
s + a2 + t(ab + ba) + t2b2 ds,
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
5
c−1
p =Z ∞
0
sp/2−1 1
s + 1
ds.
Thus ψ is twice differentiable at t = 0 and
where
(5)
ψ′′(0) = pτ (ap−2b2)
sp/2−1τ(cid:20)
− cpZ ∞
p − 1(cid:19)kak2−2p
p(cid:18) 2
2
p
0
ϕ′′(0) =
1
s + a2 (ab + ba)
1
s + a2 (ab + ba)(cid:21) ds.
ψ′(0)2 +
2
pkak2−p
p ψ′′(0) ≥
2
pkak2−p
p ψ′′(0).
It then follows that ϕ = ψ2/p is also twice differentiable at t = 0 and
Hence (D2
a,b) will be a consequence of
(6)
1
pkak2−p
p ψ′′(0) ≥ (p − 1)kbk2
p.
To prove the last inequality we claim that ψ′′(0) increases when a is replaced
by a. Indeed, the trace inside the integral in (5) is equal to twice the
following sum:
τ(cid:20)
a
s + a2 b
a
s + a2 b(cid:21) + τ(cid:20) a2
s + a2 b
1
s + a2 b(cid:21).
The second term above depends only on a (recalling that a is self-adjoint).
It remains to show that the first one increases when a is replaced by a.
By decomposing a into its positive and negative parts, we see that the first
term is equal to
τ(cid:20) a+
s + a2
+
b
a+
s + a2
+
b(cid:21) + τ(cid:20) a−
s + a2
−
b
a−
s + a2
−
b(cid:21) − 2τ(cid:20) a+
s + a2
+
b
a−
s + a2
−
b(cid:21).
All above traces are nonnegative. Therefore, the above quantity increases
when the subtraction is replaced by addition. Then tracing back the argu-
ment and noting that a = a+ + a−, we get the desired inequality
τ(cid:20)
a
s + a2 b
a
s + a2 b(cid:21) ≤ τ(cid:20) a
s + a2 b
a
s + a2 b(cid:21).
The positivity of a will facilitate the calculation of ψ′′(0) as explained in
Returning back to (5), we deduce the claim. Thus in the following we will
assume that a is a positive invertible element of M.
[2]. Since a + tb is positive for small t, we have
ψ(t) = τ ((a + tb)p).
6
E. RICARD AND Q. XU
Thus for t close to 0,
ψ′(t) = pτ ((a + tb)p−1b).
To calculate the second derivative we use the following integral representa-
tion:
Consequently,
s −
0
(a + tb)p−1 = dpZ ∞
ψ′′(0) = p dpZ ∞
sp−1(cid:20) 1
sp−1τ(cid:20) 1
0
s + a
1
s + a + tb(cid:21) ds.
b(cid:21) ds.
s + a
1
b
As shown in [2], the function
F : z 7→ τ(cid:20) 1
s + z
b
1
s + z
b(cid:21)
is convex on the positive cone of M.
Let u be the unitary operator in the polar decomposition of b (as M is
finite, the usual partial isometry in this decomposition can be chosen to be
a self-adjoint unitary). Then clearly
Now z′ = z+uzu
2
F (z) =
1
2
(F (z) + F (uzu)) ≥ F(cid:18) z + uzu
s + z′b(cid:21).
s + z′b
F (z) ≥ F (z′) = τ(cid:20)
2
1
1
commutes with u, so
(cid:19).
Let B be the Abelian von Neumann subalgebra of M generated by b, and
let Eb be the associated trace preserving conditional expectation. Then
F (z′) = τ(cid:20)Eb(cid:18) 1
s + z′b
1
s + z′(cid:19)b(cid:21).
However, the Kadison–Schwarz inequality implies
1
Eb(cid:18) 1
s + z′(cid:19) ≥ Eb(cid:18) 1
s + z′b1/2(cid:19)Eb(cid:18)b1/2
s + z′b
b.
Hence, by the positivity of the trace on products of positive elements, we
deduce
s + z′(cid:19) = Eb(cid:18) 1
s + z′(cid:19)2
1
F (z) ≥ τ(cid:20)Eb(cid:18) 1
s + z′(cid:19)2
b2(cid:21).
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
7
1
s + Eb(z′)
=
1
s + Eb(z)
.
Then by the operator convexity of 1
t , we have
Eb(cid:18) 1
s + z′(cid:19) ≥
ψ′′(0) ≥ p dpZ ∞
0
b
1
(7)
where
Lettingea = Eb(a), we have just shown
sp−1τ(cid:20) 1
s +ea
s +ea
eψ(t) = τ ((ea + tb)p).
eψ′′(0) = (p − 1)keak2−p
This finishes the proof of the lemma. (cid:3)
pkeak2−p
1
pkak2−p
p ψ′′(0) ≥
1
p
Finally, (6) immediately follows from (7). Indeed, by (7) and the Holder
inequality,
b(cid:21) ds = eψ′′(0),
p
τ (eap−2b2) ≥ (p − 1)kbk2
p.
For a ∈ M self-adjoint, we denote by s(a) = 1(0,∞)(a). s(a) is the support
of a, that is, the least projection e of M such that ea = a. We say that a
has full support if s(a) = 1.
Lemma 6. Let a, b ∈ M be self-adjoint with s(a) = 1. Then (D2
Proof. We will reduce this lemma to the previous one by cutting a + tb
with the spectral projections of a. Let e be a nonzero spectral projection of
a, and put ae = eae and be = ebe. Since a is of full support, ae is invertible
in the reduced von Neumann algebra Me = eMe. Thus Lemma 5 can be
applied to the couple (ae, be) in Me. Let ψe(t) = kae + tbek2
p as before. ψe is
twice differentiable at t = 0, and (6) holds with ψe in place of ψ.
a,b) holds.
Let e⊥ = 1 − e. Then for t in a neighborhood of the origin, we have [re-
calling that ϕ(t) = ka + tbk2
p]
ϕ(t) ≥ (ke(a + tb)ekp
p + ke⊥(a + tb)e⊥kp
p)2/pdef
= (ψe(t) + γe(t))2/p.
However,
Let
ψe(t) = kaekp
p + tψ′
e(0) +
t2
2
ψ′′
e (0) + o(t2)
as t → 0.
α(t) = kaekp
p + γe(t) = kae + e⊥(a + tb)e⊥kp
p = ka + te⊥be⊥kp
p.
8
Then
E. RICARD AND Q. XU
ϕ(t) ≥(cid:18)α(t) + tψ′
= α(t)2/p(cid:18)1 +
e(0) +
t2
2
ψ′′
2t
p
ψ′
e(0)
α(t)
e (0) + o(t2)(cid:19)2/p
p − 1(cid:19)t2 ψ′
p(cid:18) 2
ψ′′
e (0)
α(t)
t2
p
+
+
1
e(0)2
α(t)2 + o(t2)(cid:19)
= α(t)2/p +
ψ′
e(0)α(t)2/p−1 +
ψ′′
e (0)α(t)2/p−1
t2
p
2t
p
1
+
p − 1(cid:19)t2ψ′
p(cid:18) 2
≥ α(t)2/p +
2t
p
e(0)2α(t)2/p−2 + o(t2)
ψ′
e(0)α(t)2/p−1 +
t2
p
ψ′′
e (0)α(t)2/p−1 + o(t2).
By convexity of norms,
α(t)2/p + α(−t)2/p ≥ 2kak2
p = 2ϕ(0).
We then deduce that
ϕ(t) + ϕ(−t) − 2ϕ(0)
t2
2
p
≥
ψ′
e(0)
α(t)2/p−1 − α(−t)2/p−1
t
+
1
p
ψ′′
e (0)[α(t)2/p−1 + α(−t)2/p−1] + o(1).
The uniform smoothness of the norm kkp implies that the function α2/p−1
is differentiable at t = 0, and its derivative is equal to
(2 − p)kak1−p
p
τ (vap−1e⊥be⊥)
def
= δe,
where v is the unitary in the polar decomposition of a. It then follows that
Hence by (6),
D2ϕ(0) ≥
4
p
ψ′
e(0)δe +
2
p
ψ′′
e (0)kak2−p
p
.
D2ϕ(0) ≥
4
p
ψ′
e(0)δe + 2(p − 1)kbek2
p.
Thanks to the full support assumption of a, we can let e → 1 in the above
inequality. This limit procedure removes the first extra term, so we finally
get
D2ϕ(0) ≥ 2(p − 1)kbk2
p.
(cid:3)
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
9
Now we are ready to show (3).
Proof of Theorem 2. First by density, we need only to show (3) for
x, y ∈ M. Then notice that it suffices to do it for self-adjoint elements using
tensor trace. Given x, y ∈ M let
x
a classical 2 × 2-matrix trick. Indeed, let fM = M2 ⊗ M equipped with the
a =(cid:18) 0
x∗
0(cid:19) and b =(cid:18) 0
y∗
y
0(cid:19) .
Then a and b are self-adjoint. Moreover, by easy computations, (3) for x
and y is equivalent to the same inequality for a and b.
To use Lemma 6, we require that a + tb have full support for any t ∈ R.
This is achieved by a tensor product argument. Choose a positive element
c ∈ L∞([0, 1]) whose spectral measure with respect to Lebesgue measure is
diffuse (atomless), say c(t) = t for t ∈ [0, 1]. In other words, considered as a
random variable in the probability space [0, 1], the law of c is diffuse. On the
other hand, for t ∈ R, composing the spectral resolution of a + tb with the
trace τ , we can view a + tb as a random variable in another probability space
(Ω, P ). Now, consider the tensor von Neumann algebra L∞([0, 1])⊗M; it is
finite. M and L∞([0, 1]) are identified as subalgebras of L∞([0, 1])⊗M in
the usual way. Then for any ε > 0, the law of a + tb + εc is the convolution
of the laws of a + tb and εc. It is atomless since the law of c is atomless.
Consequently, the support of a + tb + εc is full in L∞([0, 1])⊗M.
Thus by Lemma 6 applied to the pair (a + εc, b), the function f (t) =
ka + tb + εck2
p satisfies D2(f )(t) ≥ 0, so it is convex. Hence
f (1) + f (−1) ≥ 2f (0); this is (3) for a + εc and b. Letting ε → 0 gives the
desired result. (cid:3)
p − (p − 1)t2kbk2
Finally, we deduce (1) from (3).
Proof of Theorem 1. Given x ∈ Lp(M) let a = E(x) and b = x−E(x).
Consider again the function f defined by
f (t) = ka + tbk2
p − (p − 1)t2kbk2
p.
Then (3) implies D2f ≥ 0, so f is convex. On the other hand, the function
g(t) = ka + tbk2
p is also convex and by the contractivity of E on Lp(M),
ka + tbkp ≥ kE(a + tb)kp = kakp.
Hence we conclude that the right derivative g′
r(0) ≥ 0,
too. Consequently, f is increasing on R+. In particular, f (1) ≥ f (0), which
is nothing but (1). (cid:3)
r(0) ≥ 0, so that f ′
10
E. RICARD AND Q. XU
3. Applications to hypercontractivity. We give in this section some ap-
plications to hypercontractivity inequalities on group von Neumann alge-
bras. Let G be a discrete group and vN (G) the associated group von Neu-
mann algebra. Recall that vN (G) is the von Neumann algebra generated by
the left regular representation λ: vN (G) = λ(G)′′ ⊂ B(ℓ2(G)). It is equipped
with a canonical trace τ , that is, τ (x) = hxe, ei, where e is the identity of
G. Given a function ψ : G → R+ with ψ(e) = 0, we consider the associated
Fourier–Schur multiplier initially defined on the family C[G] of polynomials
on G:
Pt :Xg∈G
x(g)λ(g) 7→Xg∈G
e−tψ(g)x(g)λ(g),
t > 0.
We will assume that Pt extends to a contraction on Lp(vN (G)) for every
1 ≤ p ≤ ∞. Schoenberg's classical theorem asserts that if ψ is symmetric and
conditionally negative, Pt is a completely positive map on vN (G). Since it is
trace preserving, Pt defines a contraction on Lp(vN (G)) for every 1 ≤ p ≤ ∞.
Thus in this case our assumption is satisfied.
The hypercontractivity problem for the semigroup (Pt)t>0 and for 1 < p <
q < ∞, consists in determining the optimal time tp,q > 0 such that
kPtkp→q ≤ 1
∀t ≥ tp,q.
We refer to [8, 9] for more information and historical references. It is easy
to check that if such a time tp,q exists, then ψ has a spectral gap, namely
inf g∈G\{e} ψ(g) > 0. After rescaling, we will assume that inf g∈G\{e} ψ(g) = 1.
In most-known cases the expected optimal time tp,q is attained, namely,
tp,q = logr q − 1
p − 1
.
It is a particularly interesting problem of determining the optimal time tp,q
when G = Fn is the free group on n generators with n ∈ N ∪ {∞}, and ψ is
its natural length function. Some partial results are obtained in [8, 9]. For
instance, by embedding vN (Fn) into a free product of Clifford algebras, it
is proved in [9] that for any q > 2,
1
q(cid:19) log√2.
2 −
t2,q ≤ logpq − 1 +(cid:18) 1
t2,q = logpq − 1.
On the other hand, Junge et al. [8] show that for any finite n there exists
q(n) such that if q ≥ q(n) is an even integer, then
The proof is combinatoric and based on lengthy calculations.
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
11
Here we provide an improvement. We will use Haagerup-type inequalities
[6]. Letting Sk be the set of words of length k in Fn and for any x ∈ vN (Fn)
supported on Sk, the original Haagerup inequality is
(8)
For q > 2 and k ∈ N, let Kk,q be the best constant in the following Khintchine
inequality for homogeneous polynomials x of degree k:
kxk∞ ≤ (k + 1)kxk2.
(cid:13)(cid:13)(cid:13)(cid:13)Xg∈Sk
x(g)λ(g)(cid:13)(cid:13)(cid:13)(cid:13)q ≤ Kk,q(cid:13)(cid:13)(cid:13)(cid:13)Xg∈Sk
x(g)λ(g)(cid:13)(cid:13)(cid:13)(cid:13)2
.
We will need the following:
Lemma 7. We have Kk,4 ≤ (k + 1)1/4.
Proof. Denote by gi the generators of Fn with the convention that
. For a multi-index i = (i1, . . . , id) with ij + ij+1 6= 0, we let gi =
g−i = g−1
gi1 ··· gid and i = d. So we may write
i
We compute x∗x according to simplifications that may occur
αiλ(gi).
x =Xi=k
x∗x = X0≤d≤k Xβ=d
i=j=k−d
αj,βαi,βλ(g−1
j gi),
ik−d6=jk−d6=−β1
where i, β denotes the multi-index obtained by superposing the two multi-
indices i and β. We then deduce that
kxk4
4 = kx∗xk2
ik−d6=jk−d
ik−d6=jk−d6=−β1
2 = X0≤d≤k Xi=j=k−d
≤ X0≤d≤k Xi=j=k−d
≤ X0≤d≤k(cid:18) Xi=k−d,β=d
(cid:18) Xβ=d
αj,βαi,β(cid:19)2
(cid:18) Xβ=d
αi,β2(cid:19) ·(cid:18) Xβ=d
αi,β2(cid:19) ·(cid:18) Xj=k−d,β=d
ik−d6=jk−d
ik−d6=−β1
ik−d6=−β1
jk−d6=−β1
jk−d6=−β1
αj,β2(cid:19)
αj,β2(cid:19)
= (k + 1)kxk4
2.
(cid:3)
12
E. RICARD AND Q. XU
Remark 8. Taking αi = 1 and by the free central limit theorem as n →
∞, one can see that the previous inequality is sharp. Thus Kk,4 = (k + 1)1/4.
This constant is the L4-norm of the kth Chebyshev polynomial for the semi-
circle law.
Using the Holder inequality we deduce from (8) and the previous lemma
that for any k ≥ 1,
(9)
(10)
Kk,q ≤ (k + 1)1−3/q ,
Kk,q ≤ (k + 1)1/2−1/q ,
q ≥ 4,
2 ≤ q ≤ 4.
We will also use the following elementary folklore:
Remark 9. Let T : Lp(M) → Lq(M) be a bounded linear map. Assume
that T is 2-positive in the sense that IdM2 ⊗ T maps the positive cone of
Lp(M2 ⊗ M) to that of Lq(M2 ⊗ M). Then
kT (x)kq ≤ kT (x)k1/2
q kT (x∗)k1/2
q
,
x ∈ Lp(M).
Consequently,
kTk = sup{kT (x)kq : x ∈ Lp(M)+,kxkp ≤ 1}.
Indeed, for any x ∈ Lp(M),
So the 2-positivity of T implies
x
x∗
(cid:18)x
x∗(cid:19) ≥ 0.
T (x∗) T (x∗)(cid:19) ≥ 0.
(cid:18) T (x)
T (x)
This yields a contraction c ∈ M such that T (x) = T (x)1/2cT (x∗)1/2. Then
the Holder inequality gives the assertion.
Theorem 10. There exists ε0 > 0 such that for any free Fn and any
q ≥ 4 − ε0,
kPtk2→q ≤ 1 ⇔ t ≥ logpq − 1.
Proof. The necessity is clear. The proof of the sufficiency will rely on
Remark 3.7 of [9]. Let σ be the automorphism of vN (Fn) given by σ(λ(gi)) =
λ(g−1
). Then Pt is hypercontractive from L2 to Lq with optimal time on
vN (Fn)σ, the fixed point algebra of σ. Let E be the conditional expectation
onto vN (Fn)σ. Note that E = Id+σ
and it commutes with Pt.
2
i
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
13
Fix q > 2. To prove kPtk2→q ≤ 1 for t ≥ log√q − 1, it suffices to show
kPt(x)kq ≤ kxk2 for any positive x ∈ C[Fn] by virtue of Remark 9. We need
one more reduction. Given complex numbers ζi of modulus 1, there exists
an automorphism πζ of vN (Fn) given by π(gi) = ζigi. It is an isometry on
all Lp's. Note that πζ and Pt commute. Thus to prove kPt(x)kq ≤ kxk2, we
may assume that x(gi) is real for every generator gi. We will fix a positive
x ∈ C[Fn] with the last property.
Then write x = y + z where y = E(x). Since x(gi) ∈ R, we have that z does
not have constant terms nor of degree 1. By Theorem 1 (or Theorem 3) and
Remark 3.7 of [9],
kPt(x)k2
q ≤ kPt(y)k2
q + (q − 1)kPt(z)k2
q ≤ kyk2
2 + (q − 1)kPt(z)k2
q.
Then for t = log√q − 1, decomposing z according to its homogeneous com-
ponents (zk) and using the Khintchine and the Cauchy–Schwarz inequalities,
we get
kPt(z)k2
≤Xk≥2
K 2
k,q
1
(q − 1)k kzk2
2.
We aim to find those q > 2 for which
e−tkkzkkq(cid:19)2
q ≤(cid:18)Xk≥2
Rq = (q − 1)Xk≥2
Rq ≤Xk≥2
K 2
k,q
1
(q − 1)k ≤ 1.
(k + 1)2(1−3/q)
1
(q − 1)k−1 .
For q ≥ 4, by (9) we have
The terms of the sum on the right-hand side are decreasing functions of q if
their derivatives are negative, that is, if
6(q − 1)
q2
≤
k − 1
log (k + 1)
.
Noting that the left-hand side of the above inequality is decreasing on q, one
easily checks that this inequality is true for q ≥ 4 and k ≥ 3. However, it is
true for k = 2 if and only if q ≥ q0, where
q0 =p3 log 3(p3 log 3 +p3 log 3 − 2) ≈ 5.36244.
We have the following numerical estimates:
R4 ≤ 0.92952
and
32(1−3/q0)
q0 − 1 −
31/2
3 ≤ 0.02613.
14
E. RICARD AND Q. XU
Hence if q ∈ [4, q0],
32(1−3/q0)
31/2
3
Rq ≤ R4 +
q0 − 1 −
We thus conclude that Rq < 1 for all q ≥ 4.
Since Rq is dominated by a continuous function of q, using (10) we get
a similar estimate for q ≥ 4 − ε0 for some ε0. A numerical estimate gives
ε0 ≈ 0.18. (cid:3)
< 1.
Remark 11.
Instead of Remark 3.7 of [9], we can equally use Theo-
rem A(iii) of [9] in the preceding proof. But the commutation of Pt and the
conditional expectation onto the symmetric subalgebra An
sym is less obvious.
It is likely that ε0 = 2, but other methods would have to be developed.
Gross's pioneering work [5] shows that hypercontractivity is equivalent
to the validity of log-Sobolev inequalities. In the present situation of free
groups, the validity of the hypercontractivity with optimal time in full gen-
erality (or equivalently, ε0 = 2) is equivalent to the following log-Sobolev
inequality in Lq for any q ≥ 2:
(SLq)
q
τ (xq−1L(x)) + kxkq
q logkxkq,
τ (xq log x) ≤
x ∈ D+.
2(q − 1)
Here L denotes the negative generator of (Pt), and D is a core for L where
the inequality makes sense. It is known that (SL2) implies (SLq) for all q;
see [10]. In the same spirit we can show that (SLp) implies (SLq) if q > p ≥ 2.
Let us record this explicitly here since it might be of interest. The semigroup
(Pt) can be any completely positive symmetric Markovian semigroup such
that D is rich enough.
Remark 12. Let q > p ≥ 2. Then (SLp) implies (SLq).
To check the remark we rewrite (SLq) in a symmetric form with respect
to q and its conjugate index q′ (provided that D is big enough):
(SLs
q)
L(x1/q)) + τ (x) log τ (x),
2 q′qτ (x1/q′
τ (x log x) ≤ 1
x ∈ D+.
r τ (z(1 − Pr)(y)). Let r >
Recall that for y ∈ Dom(L), τ (zL(y)) = limr→0
0 and x ∈ M+, and we will check that the function q 7→ q′qτ (x1/q′
(1 −
Pr)(x1/q)) is increasing for q ≥ 2; we put θ = 1
q . It is known from [1] that
there exists a positive symmetric Borel measure µr on σ(x)× σ(x) such that
1
τ (x1−θPr(xθ)) =Zσ(x)×σ(x)
s1−θtθdµr(s, t).
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
15
Hence, by symmetry, it suffices to show that
f : θ 7→
1 + u − uθ − u1−θ
θ(1 − θ)
f (θ) =Z 1
f ′′(θ) =Z 1
0
0
is convex on [0, 1] for u > 0 as f (θ) = f (1 − θ). One easily checks that
log(u)(uθ+(1−θ)(1−t) − uθt + u1−θ+θt − u(1−θ)(1−t)) dt,
log(u)3(t2(uθ+(1−θ)(1−t) − uθt)
+ (1 − t)2(u1−θ+θt − u(1−θ)(1−t))) dt ≥ 0.
We end this section with application to more general groups (G, ψ). If ψ
Passing to the limit in r gives the result if D is big enough.
is symmetric and satisfies the exponential order growth
∀R > 0
{g ∈ G : ψ(g) ≤ R} ≤ CρR
(11)
for some C > 0 and ρ > 1, then one of the main results of [8] shows that for
2 < q < ∞,
t2,q ≤ η logpq − 1
for any η > 2 when ρ is large compared to C. Their argument consists in first
considering the case q = 4 by combinatoric methods and then using Gross's
extrapolation. We will show that the martingale inequality in Theorem 1
easily implies a slight improvement. Note that our estimate on t2,q is as close
as to the expected optimal time as when q is sufficiently large, compared to
ρ and C.
q
Proposition 13. Assume (11) and 2 < q < ∞. Then
t2,q ≤(cid:18) q − 2
Proof. By (11), the range of ψ is countable. Let ψ(G) = {n0, n1, n2, . . .}
with n0 < n1 < n2 < ··· . Then n0 = 0 and n1 = 1. Let x ∈ vN (G) be a
logp2Cρ + logpq − 1(cid:19) ∨ log ρ.
polynomial, x =P x(g)λ(g), and let y = x − x(e). By Theorem 1
Let Bk = {g ∈ G : ψ(g) ≤ nk}, Sk = Bk \ Bk−1 and yk =Pg∈Sk
e−2tnkSk(cid:19) ·(cid:18)Xk≥1
kykk2
kPt(y)k2
q ≤ x(e)2 + (q − 1)kPt(y)k2
q .
e−tnkkykk∞(cid:19)2
∞ ≤(cid:18)Xk≥1
≤(cid:18)Xk≥1
∞
Sk (cid:19).
kPt(x)k2
x(g)λ(g).
Then
Actually exchanging the arguments, one has the following, slightly better
estimate that we will not use:
kPt(y)k2
e−2tnkSk2(q−2)/qkyk2
2.
By (11), for t > log ρ,
Xk≥1
e−2tnkSk ≤Xk≥1
e−(2t−log ρ)s ds
16
Obviously,
E. RICARD AND Q. XU
We get, using the Holder inequality,
kykk2
kPt(y)k2
kyk2
2.
x(g)2.
≤ SkXg∈Sk
e−2tnkSk(cid:19)(q−2)/q
x(g)(cid:19)2
∞ ≤(cid:18)Xg∈Sk
q ≤ e−4t/q(cid:18)Xk≥1
q ≤Xk≥1
(e−2tnk − e−2tnk+1)Bk ≤ 2CtZ ∞
e−(2t−log ρ) ≤ 2Ce−(2t−log ρ).
t
1
= 2C
2t − log ρ
Hence, if 2t ≥ q−2
q
the assertion. (cid:3)
log(2Cρ) + log(q − 1), we deduce kPt(x)kq ≤ kxk2, whence
REFERENCES
[1] Albeverio, S. and Høegh-Krohn, R. (1977). Dirichlet forms and Markov semi-
groups on C ∗-algebras. Comm. Math. Phys. 56 173–187. MR0461153
[2] Ball, K., Carlen, E. A. and Lieb, E. H. (1994). Sharp uniform convexity and
smoothness inequalities for trace norms. Invent. Math. 115 463–482. MR1262940
[3] Biane, P. (1997). Free hypercontractivity. Comm. Math. Phys. 184 457–474.
MR1462754
[4] Carlen, E. A. and Lieb, E. H. (1993). Optimal hypercontractivity for Fermi fields
and related noncommutative integration inequalities. Comm. Math. Phys. 155
27–46. MR1228524
[5] Gross, L. (1975). Logarithmic Sobolev inequalities. Amer. J. Math. 97 1061–1083.
MR0420249
[6] Haagerup, U. (1978/79). An example of a nonnuclear C ∗-algebra, which has the
metric approximation property. Invent. Math. 50 279–293. MR0520930
[7] Haagerup, U., Junge, M. and Xu, Q. (2010). A reduction method for noncom-
mutative Lp-spaces and applications. Trans. Amer. Math. Soc. 362 2125–2165.
MR2574890
[8] Junge, M., Palazuelos, C., Parcet, J. and Perrin, M. (2013). Hypercontractiv-
ity in group von Neumann algebras. Preprint.
[9] Junge, M., Palazuelos, C., Parcet, J., Perrin, M. and Ricard, ´E. (2015).
Hypercontractivity for free products. Ann. Sci. ´Ecole Norm. Sup. (4) 48 861–
889. MR3377067
A NONCOMMUTATIVE MARTINGALE CONVEXITY INEQUALITY
17
[10] Olkiewicz, R. and Zegarlinski, B. (1999). Hypercontractivity in noncommutative
Lp spaces. J. Funct. Anal. 161 246–285. MR1670230
[11] Pisier, G. and Xu, Q. (2003). Non-commutative Lp-spaces. In Handbook of the
Geometry of Banach Spaces, Vol. 2 (W. B. Johnson and J. Lindenstrauss,
eds.) 1459–1517. North-Holland, Amsterdam. MR1999201
Laboratoire de Math´ematiques Nicolas Oresme
Universit´e de Caen Basse-Normandie
14032 Caen Cedex
France
E-mail: [email protected]
School of Mathematics and Statistics
Wuhan University
Wuhan 430072
China
and
Laboratoire de Math´ematiques
Universit´e de Franche-Comt´e
25030 Besanc¸on Cedex
France
E-mail: [email protected]
|
1002.0605 | 5 | 1002 | 2011-02-27T13:36:37 | On Sofic Actions and Equivalence Relations | [
"math.OA",
"math.CO"
] | The notion of sofic equivalence relation was introduced by Gabor Elek and Gabor Lippner. Their technics employ some graph theory. Here we define this notion in a more operator algebraic context, starting from Connes' embedding problem, and prove the equivalence of this two definitions. We introduce a notion of sofic action for an arbitrary group and prove that amalgamated product of sofic actions over amenable groups is again sofic. We also prove that amalgamated product of sofic groups over an amenable subgroup is again sofic. | math.OA | math |
ON SOFIC ACTIONS AND EQUIVALENCE RELATIONS
L. P AUNESCU1
Abstract. The notion of sofic equivalence relation was introduced by Gabor Elek and
Gabor Lippner. Their technics employ some graph theory. Here we define this notion in a
more operator algebraic context, starting from Connes' embedding problem, and prove the
equivalence of this two definitions. We introduce a notion of sofic action for an arbitrary
group and prove that amalgamated product of sofic actions over amenable groups is again
sofic. We also prove that amalgamated product of sofic groups over an amenable subgroup
is again sofic.
Contents
1 Introduction
1.1 Definitions of hyperlinear and sofic actions . . . . . . . . . . . . . . . . . . .
1.2 The Feldman-Moore construction . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Definition of sofic equivalence relations . . . . . . . . . . . . . . . . . . . . .
1.4 Preliminaries
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Bernoulli shifts
3 Sofic actions
4 Sofic equivalence relations
2
2
4
6
7
10
13
16
1Work supported by the Marie Curie Research Training Network MRTN-CT-2006-031962 EU-NCG.
1
1
Introduction
1.1 Definitions of hyperlinear and sofic actions
We shall work with ultraproducts of matrix algebras instead of Rω. We denote such a
ultraproduct by Πk→ωMnk (C). It is well known that embedding in Rω is equivalent with
embedding in Πk→ωMnk (C). For the sake of completeness we include a proof. We shall
always denote by T r the normalized trace on a type I or type II finite factor.
Proposition 1.1. Rω embeds in some Πk→ωMnk (C). Thus any type II factor embedable
in Rω also embeds in a ultraproduct of matrices.
Proof. Approximating the hyperfinite factor by matrix algebras we can easily see that
R ⊂ M ω
nk
. Then Rω ⊂ (M ω
nk
nk
)ω ≃ M ω⊗ω
.
Some results in this article will have to do with soficity of certain objects. Because of
this, we shall work with permutations.
Notation 1.2. We shall denote by Dnk ⊂ Mnk the diagonal subalgebra and by Pnk ⊂ Mnk
the subgroup of permutation matrices. Given an ultraproduct Πk→ωMnk (C) denote by
Πk→ωDnk (C) and Πk→ωPnk (C) the corresponding subsets.
By a theorem of Popa (see [Po1], proposition 4.3) Πk→ωDnk (C) is a maximal abelian
nonseparable subalgebra of Πk→ωMnk (C). We now introduce a notion of hyperlinearity
and soficity for actions.
Definition 1.3. An action α of a countable group G on a standard Borelian space (X, B, µ)
is called hyperlinear if the crossed product L∞(X) ⋊α G embeds in Rω.
Definition 1.4. An action α of a countable group G on a standard Borelian space (X, B, µ)
is called sofic if the crossed product L∞(X) ⋊α G embeds in Πk→ωMnk (C) such that
L∞(X) ⊂ Πk→ωDnk (C) and G ⊂ Πk→ωPnk (C).
So for an action to be sofic instead of simply hyperlinear we want that the unitaries
implementing the action, to be permutation matrices. Note that also Elek and Lippner
defined a notion of soficity for action of F∞ as a preliminary step in defining the concept
of sofic equivalence relation. The two notions of sofic actions are not equivalent.
2
Next theorem shows that the property of sofic action is invariant under orbit
equivalence. While the proof is quite simple, this theorem hints very clearly that being
sofic is a property of the orbit equivalence relation, rather than of the action itself.
Theorem 1.5. Let α and β be two free orbit equivalent actions. If α is hyperlinear (sofic)
then also β is hyperlinear (sofic).
Proof. Let α be an action of G and β be an action of H such that L∞(X) ⋊α G ≃
L∞(X) ⋊β H (isomorphism that is identity on L∞(X)). By this we are done with the
hyperlinear part of the theorem. Consider now a sofic embedding of L∞(X) ⋊α G in
Πk→ωMnk . Let ug ∈ L(G), g ∈ G and vs ∈ L(H), s ∈ H denoting the corresponding
unitaries. For s ∈ H we can find projections {pgg ∈ G} in L∞(X) such that Pg pg = 1
and vs =Pg pgug. By next lemma, vs ∈ Πk→ωPnk and we are done.
Lemma 1.6. Let {eii ∈ N} be projections in Πk→ωDnk such that Pi ei = 1. Let
{uii ∈ N} be unitary elements in Πk→ωPnk such that Pi u∗
i eiui = 1. Then v =Pi eiui is
Proof. Using Pi ei = 1 we can easily construct projections ek
a unitary in Πk→ωPnk .
i ∈ Dnk such that:
1. ei = Πk→ωek
i ;
i = 1nk .
By hypothesis we have ui = Πk→ωuk
2. Pi ek
but vk are not necessary permutation matrices. Because Pi u∗
i then v = Πk→ωvk,
i eiui = 1 we have v∗v = 1,
so v is a unitary. Now, vk has only 0 and 1 entries and exactly one entry of 1 on each line.
Then vk∗vk is a diagonal matrix giving the number of 1 entries on each column.
i ∈ Pnk . If vk =Pi ek
i where uk
i uk
Denote by rk the number of columns in vk having only 0 entries. This number will
also represent the numbers of 0 entries on the diagonal of vk∗vk. Then:
vk∗vk − Id2
2 ≥
rk
nk
.
Because Πk→ωvk∗vk = 1 we have rk/nk →k→ω 0. We can "move" rk entries of 1 in vk on
the rk missing columns to get a permutation matrix wk ∈ Pnk . Then:
vk − wk2
2 =
2rk
nk
.
3
Combined with rk/nk →k→ω 0 we get v = Πk→ωwk. This will prove the lemma.
Definition 1.7. We shall call an element of Πk→ωPnk a permutation,
ultraproduct of permutations. A piece of permutation will be an element of the form
instead of an
e · p, where e is a projection in Πk→ωDnk and p is a permutation.
Lemma 1.6 now becomes a unitary that is a sum of pieces of permutations is a
permutation.
1.2 The Feldman-Moore construction
Based on theorem 1.5, a definition of sofic equivalence relation is possible for relations
generated by free actions. Unfortunately, not all equivalence relations have this property.
We adapt the definition using the Feldman-Moore construction. Let us recall some things
from [Fe-Mo].
Let (X, B, µ) be a standard space with a probability measure. Let E ⊂ X 2 an
equivalence relation on X such that E ∈ B × B. We shall work only with equivalence
relations that are countable, i.e. every equivalence class is countable, and µ-invariant.
Before we recall what this means we introduce some notation.
Denote by [E] set of all isomorphism with graph in E and by [[E]] set of all partial
isomorphism with graph in E:
[E] ={θ : X → X : θ bijection , graphθ ⊂ E};
[[E]] ={φ : A → B : A, B ⊂ X, φ bijection , graphφ ⊂ E}.
Definition 1.8. Let E an equivalence relation on (X, µ). Then E is called µ-invariant if
for any φ : A → B, φ ∈ [[E]] we have µ(A) = µ(B).
For a von Neumann algebra A, we shall denote by U (A) the group of unitaries in the
algebra A. The normalizer NM (A) and the normalizing pseudogroup GNM (A) are the
corresponding objects of [E] and [[E]] respectively. For A ⊂ M define:
NM (A) = {u ∈ U (M ) : uAu∗ = A};
GNM (A) = {v ∈ M partial isometry : vv∗, v∗v ∈ A, vAv∗ = Avv∗}.
4
Now we can construct the algebra M (E) associated to an equivalence relation.
Definition 1.9. A measurable function a : E → C is called finite if a is bounded and
there is a natural number n such that:
Card({x : a(x, y) 6= 0}) ≤ n ∀y ∈ X;
Card({y : a(x, y) 6= 0}) ≤ n ∀x ∈ X.
A finite function (matrix) is a bounded function with finite number of nonzero entries
on each line and column (having also a global margin). We shall multiply this functions
as general matrices and the definition of finite function guarantees we get a ∗-algebra.
Define:
M0(E) = {a : E → C : a finite function};
a(x, y)b(y, z);
a · b(x, z) =Xy
a∗(x, y) = a(y, x).
It is easy to check this is indeed a ∗-algebra. The trace is defined in a similar way as
in the case of matrices:
T r(a) =ZX
a(x, x)dµ.
The algebra M (E) will be the weak closure of M0(E) in the GNS representation of
(M0(E), T r). By general theory of von Neumann algebras, using the cyclic separating
vector of the GNS representation, we can still see elements of M (E) as measurable
functions a : E → C.
Let ∆ = {(x, x) : x ∈ X} the diagonal in E. Define the subalgebra of diagonal
matrices:
A = {a ∈ M (E) : supp(a) ⊂ ∆}.
We shall denote by δy
x the Kronecker delta function, i.e. δy
δy
x = 0. Notation χA stands for the characteristic function of A.
x = 1 iff x = y; otherwise
Definition 1.10. For θ ∈ [E] define uθ ∈ M (E) by: uθ(x, y) = δθ(y)
vφ(x, y) = χdom(φ)(y) · δφ(y)
x
.
x
. For φ ∈ [[E]], define
5
It is not hard to see that uθ ∈ N (A) for any θ ∈ [E] and vφ ∈ GN (A) for any
φ ∈ [[E]]. Moreover any u ∈ N (A) is of the form a · uθ, where a ∈ U (A) and θ ∈ [E]. Also
uθuψ = uθ◦ψ for θ, ψ ∈ [E]. This provides a group isomorphism between the Weyl group
N (A)/U (A) and [E].
Algebra A is maximal abelian in M (E). Also N (A)′′ = M (E). This properties make
A a Cartan subalgebra of M (E). We shall call A ⊂ M (E) a Cartan pair.
Motivation of Feldman-Moore construction was the invariance of crossed product up
to orbit equivalent actions. Next example shows this is indeed the right construction.
Exemple 1.11. Let α : G → Aut(X, µ) a free action. Denote by Eα the orbit equivalence
relation generated by α on X. Then:
L∞(X) ⋊α G ≃ M (Eα).
1.3 Definition of sofic equivalence relations
The notion of sofic equivalence relation was introduced by Gabor Elek and Gabor Lippner
(see [El-Li]). We shall provide a different definition here and prove in section 4 the
equivalence of the two definitions.
Definition 1.12. An equivalence relation E is called sofic if there is an embedding of
M (E) in some Πk→ωMnk such that A ⊂ Πk→ωDnk and N (A) ⊂ U (A) · Πk→ωPnk .
This has the advantage of being a compact definition, but in practice we shall need
the following type of embeddings.
Definition 1.13. Let E an equivalence relation and A ⊂ M (E) the Cartan pair associated
to E. We call an embedding Θ : M (E) → Πk→ωMnk sofic if Θ(A) ⊂ Πk→ωDnk and
Θ(uθ) ⊂ Πk→ωPnk for any θ ∈ [E].
Proposition 1.14. An equivalence relation E is a sofic if and only if its Cartan pair
A ⊂ M (E) admits a sofic embedding.
Proof. Let Θ : M (E) → Πk→ωMnk an embedding such that Θ(A) ⊂ Πk→ωDnk and
Θ(N (A)) ⊂ Θ(U (A)) · Πk→ωPnk .
6
For ϕ ∈ [E] we have a unique decomposition Θ(uϕ) = Θ(fϕ)vϕ, where fϕ ∈ U (A) and
vϕ ∈ Πk→ωPnk . Then:
Θ(fψ ◦ ϕ−1) = Θ(uϕ)Θ(fψ)Θ(u∗
ϕ) = Θ(fϕ)(vϕΘ(fψ)v∗
ϕ)Θ(f ∗
ϕ) = vϕΘ(fψ)v∗
ϕ.
Because of the uniqueness of the decomposition of Θ(uϕ) we have fϕψ = fϕ(fψ ◦ ϕ−1).
If χϕ denotes the projection with support {x ∈ X : ϕ(x) = x}, one has χϕuϕ = χϕ and
hence:
Θ(f ∗
ϕχϕ) = Θ(f ∗
ϕχϕuϕ) = Θ(χϕ)vϕ.
The conditional expectation of vϕ on Πk→ωDnk is a projection. Thus, taking the
ϕχϕ is positive and hence equal to χϕ.
conditional expectation on Θ(A) it follows that f ∗
So, for all ϕ ∈ [E] we have fϕχϕ = χϕ. Altogether, it follows that the formula:
α(uϕ) = f ∗
ϕuϕ for all ϕ ∈ [E],
α(a) = a for all a ∈ A
provides a well defined automorphism of M (E). The composition of Θ and α is the
required sofic embedding of M (E).
As a consequence of this proposition and lemma 1.6, we have the following result.
Proposition 1.15. Let α be a free action. Then Eα is a sofic equivalence relation if and
only if α is a sofic action.
1.4 Preliminaries
We include here some propositions that will be used in different situations. First an easy
observation.
Observation 1.16. Let Θ = Πk→ωΘk be a sofic embedding of some von Neumann
algebra M in Πk→ωMnk . Consider also {rk}k a sequence of natural numbers. Then
Θ ⊗ 1 = Πk→ωΘk ⊗ 1rk is again a sofic embedding of M in Πk→ωMnk ⊗ Mrk = Πk→ωMnkrk .
This trick will be used when we need to embed two algebras in the same Πk→ωMnk
(that is the same matrix dimension at each step).
7
Next sequence of lemmas will show that two sofic embeddings of the same hyperfinite
algebra are conjugate by a permutation.
Lemma 1.17. Let e, f two projections in Πk→ωDnk such that T r(e) = T r(f ). Then there
is a unitary u ∈ Πk→ωPnk such that f = ueu∗.
Proof. Let e = Πk→ωek and f = Πk→ωf k such that ek and f k are projections in Dnk .
Assume ek has tk entries of 1 and f k has sk entries of 1, so limk→ω tk/nk = T r(e) =
T r(f ) = limk→ω sk/nk. Choose pk
1 has the first tk entries of 1
on the diagonal. In the same way choose pk
2 has the first sk entries of
1 on the diagonal. Define pi = Πk→ωpk
i for i = 1, 2. Our constructions guarantee that
T r(p1ep∗
2) = limk→ω tk −sk/nk = 0. Then p1ep∗
1 ∈ Pnk such that pk
2 so define u = p∗
1 = p2f p∗
2 such that pk
2f kpk∗
1 −p2f p∗
1ekpk∗
2p1.
Lemma 1.18. Let {ei}m
that Pm
i=1 ei = 1 = Pm
unitary u ∈ Πk→ωPnk such that fi = ueiu∗ for all i = 1, . . . , m.
i=1 and {fi}m
i=1 two sequences of projections in Πk→ωDnk such
i=1 fi and T r(ei) = T r(fi) for each i = 1, . . . , m. Then there is a
Proof. Apply previous lemma for each i = 1, . . . , m to get elements ui ∈ Πk→ωPnk such
that uieiu∗
i=1 uiei. Then by lemma (1.6) we know u ∈ Πk→ωPnk .
Also ueiu∗ = uieiu∗
i = fi. Define u = Pm
i = fi.
Proposition 1.19. Let Θ1, Θ2 be two embeddings of L∞(X) in Πk→ωDnk . Then there
exists a unitary u ∈ Πk→ωPnk such that Θ2(a) = uΘ1(a)u∗ for every a ∈ L∞(X).
Proof. Let Am an increasing sequence of commutative finite dimensional subalgebras such
that L∞(X) = (∪mAm)′′. By previous lemma there exist a unitary um ∈ Πk→ωPnk such
that Θ2(a) = Adum ◦ Θ1(a) for a ∈ Am. We shall construct u ∈ Πk→ωPnk using a
diagonal argument. Let um = Πk→ωuk
m ∈ Pnk and Θi(a) = Πk→ωΘi(a)k with
Θi(a)k ∈ Dnk .
m with uk
Inductively choose smaller Fm ∈ ω, m ∈ N such that uk
mΘ1(a)kuk∗
m −Θ2(a)k2 < 1/m
for any a ∈ (Am)1, k ∈ Fm. Define uk = uk
m for k ∈ Fm \ Fm+1 and set u = Πk→ωuk.
Proposition 1.20. Let E be a hyperfinite equivalence relation and A ⊂ M (E) the Cartan
pair associated to E. Let Θ1, Θ2 two sofic embeddings of M (E) in Πk→ωMnk . Then there
exists a unitary u ∈ Πk→ωPnk such that Θ2(x) = uΘ1(x)u∗ for every x ∈ M (E).
8
Proof. Using the previous proposition we can assume Θ1 and Θ2 coincide on A. We shall
first prove this result in case of ergodic equivalence relation, i.e. M (E) is the hyperfinite
factor. By definition of hyperfinite equivalence relation and Feldman-Moore construction
(see also proof of 4.1 from [Po2]) there exists an increasing sequence of matrix algebras
{Nm}m≥1 of M (E) each of them with a set of matrix units {em
ij } such that:
1. M (E) is the weak closure of ∪mNm;
2. em
ii ∈ A and Pi em
ii = 1;
3. em
ij are of the form vθ with θ ∈ [[E]];
4. every ep
rs, for p ≤ m, is the sum of some em
ij .
Elements vθ are of the form e · uφ, where e is a projection in A and φ ∈ [E]. Combined
with Θl is sofic, we get Θl(em
ij ) is a piece of permutation. Define
pm =Xj
Then
Θ2(em
j1)Θ1(em
1j ).
pmp∗
m =Xi,j
=Xj
Θ2(em
i1)Θ1(em
1i)Θ1(em
j1)Θ2(em
1j)
Θ2(em
j1)Θ1(em
11)Θ2(em
1j ) =Xj
Θ2(em
jj) = 1,
so pm is a unitary. Using 1.6 we have pm ∈ Πk→ωPnk . Moreover:
pmΘ1(em
rs)p∗
Θ2(em
i1)Θ1(em
1i)Θ1(em
rs)Θ1(em
j1)Θ2(em
1j)
m =Xi,j
=Θ2(em
r1)Θ1(em
11)Θ2(em
1s) = Θ2(em
rs).
We obtained pmΘ1(x)p∗
m = Θ2(x) for x ∈ Nm. Employing another diagonal argument we
construct a permutation p ∈ Πk→ωPnk such that pΘ1(x)p∗ = Θ2(x) for x ∈ ∪mNm. Using
1 we are done.
The proof in general case works the same. The only difference is that {Nm}m≥1 are
finite dimensional algebras instead of matrix algebras, so we need to be more careful when
9
defining pm. Assume that Nm = N 1
Let {em
ij;v} a set of matrix units for N v
m ⊕ . . . ⊕ N t
m. Then define:
m ⊕ N 2
m, with N v
m factors for v = 1, . . . , t.
pm =
t
Xv=1Xj
Θ2(em
j1;v)Θ1(em
1j;v).
Computations that pm is a unitary and pmΘ1(em
rs)p∗
m = Θ2(em
rs) are the same.
2 Bernoulli shifts
In [El-Li] Elek and Lippner proved that equivalence relations generated by Bernoulli shifts
of sofic groups are sofic. We present here the nice proof of Narutaka Ozawa from [Oz].
Theorem 2.1. (Elek-Lippner) Equivalence relations generated by Bernoulli shifts of sofic
groups are sofic.
Proof. (Ozawa) Let G be a sofic group. Every Bernoulli shift is a free action. Using 1.15
we just need to prove that each Bernoulli shift of G is a sofic action.
Let X = {0, 1}G = {f : G → {0, 1}}. For distinct g1, g2 . . . gm, define the cylinder set:
ci1,i2,...,im
g1,g2,...,gm = {f ∈ X : f (gj) = ij ∀j = 1 . . . m},
and let Qi1,i2,...,im
that β(g)ci1,i2,...,im
g1,g2,...,gm be the projection onto this set. Then β is the action of G on X such
g1,g2,...,gm = ci1,i2,...,im
gg1,gg2,...,ggm.
Let Θ0 : G → Πk→ωPnk be a sofic embedding of G with T r(Θ0(g)) = 0 for each
g 6= e. Write Θ0(g) = Πk→ωpg;k such that pg;k ∈ Pnk . Define Θ : G → Πk→ωMnk ⊗ M2nk
by Θ = Θ0 ⊗ 1. Let Yk a set with nk elements and identify Dnk with L∞(Yk). Also let
Zk = {η : Yk → {0, 1}} and identify D2nk with L∞(Zk). Define now:
ci1,i2,...,im
g1,g2,...,gm;k = {(ξ, η) ∈ Ynk × Znk : η(p−1
gj ;k(ξ)) = ij, j = 1, . . . , m}.
Let Qi1,i2,...,im
g1,g2,...,gm;k ∈ Dnk ⊗ D2nk be the characteristic function of ci1,i2,...,im
g1,g2,...,gm;k. Define now
10
Θ(Qi1,i2,...,im
g1,g2,...,gm) = Πk→ωQi1,i2,...,im
g1,g2,...,gm;k. Then:
Θ(g)Θ(Qi1,i2,...,im
g1,g2,...,gm)Θ(g)∗ =Πk→ω(pg;k ⊗ 1)Qi1,i2,...,im
g1,g2,...,gm;k(p−1
g;k ⊗ 1)
=Πk→ωχ
{(ξ,η):(p−1
g;k⊗1)(ξ,η)∈c
i1 ,i2,...,im
g1 ,g2,...,gm ;k}
=Πk→ωχ{(ξ,η):η(p−1
gj ;kp−1
g;k(ξ))=ij , j=1,...,m}
=notΠk→ωχTk .
Θ(Qi1,i2,...,im
gg1,gg2,...,ggm) =Πk→ωχ{(ξ,η):η(p−1
ggj ;k(ξ))=ij , j=1,...,m}
=notΠk→ωχSk .
If (ξ, η) ∈ Tk∆Sk then for some j = 1, . . . , m we have p−1
fact that Θ0 is a sofic embedding it follows that Πk→ωχTk = Πk→ωχSk .
g;k(ξ) 6= p−1
gj;kp−1
ggj;k(ξ). Given the
The only thing left is to compute the trace of Θ(Qi1,i2,...,im
g1,g2,...,gm). For this, let Ak = {ξ ∈
gj;k(ξ) are different for j = 1, . . . , m}. Because T r(Θ0(g)) = 0 for g 6= e we have
Yk : p−1
limk→ω Card(Ak)/nk = 1. Then:
T r(Θ(Qi1,i2,...,im
g1,g2,...,gm)) = lim
k→ω
T r(Qi1,i2,...,im
g1,g2,...,gm;k) = lim
k→ω
1
nk2nk
(Xξ∈Ak
2nk−m + Xξ /∈Ak
vξ) =
1
2m .
This will prove that Θ is an embedding of L∞(X) ⋊β G, proving the soficity of the action
β.
The proof can be adapted to work for any Bernoulli shift. For a finite uniform Bernoulli
shift the proof works the same. A diagonal argument will prove the theorem in case
X = [0, 1]G (with product of Lebesgue measure). Any other Bernoulli shift will yield a
subalgebra of L∞([0, 1]G) ⋊ G.
Next easy proposition will be used in the proof of corollary 2.3.
Proposition 2.2. Let G act freely on a countable set I. Then the generalized Bernoulli
shift of G on {0, 1}I is sofic.
Proof. If G acts freely on I then I is of the form G × I ′ and the action is a shift on the
first component. The generalized Bernoulli shift on {0, 1}I is a classic Bernoulli shift on
X G where X = {0, 1}I ′
.
11
A formally weaker version of the following result was first obtain by Benoit Collins
and Ken Dykema ([Co-Dy]). Independently, Elek and Szabo proved this theorem using
different methods ([El-Sz2]).
Corollary 2.3. Amalgamated products of sofic groups over amenable groups are sofic.
Proof. Let G1, G2 be two sofic groups with a common amenable subgroup H. Let
X = {0, 1}G1∗H G2 equipped with product measure. Then G1 and G2 act on X as
generalized Bernoulli shifts and this actions coincide on H. Using the above proposition
(and 1.16) we can construct sofic embeddings Θi : L∞(X) ⋊ Gi → Πk→ωMnk for i = 1, 2.
By proposition (1.20) we can assume Θ1 = Θ2 on L∞(X) ⋊ H (here we use H amenable
and classic result from [CFW]). Note that now Θ1 acts on Θi(L∞(X)) by shifting with
G1 and Θ2 acts on the same space by shifting with G2. This will provide a representation
Θ of G1 ∗H G2 on Πk→ωPnk . Also, Θ acts on Θi(L∞(X)) as a classic Bernoulli shift. This
implies Θ is faithful, so G1 ∗H G2 is sofic.
Corollary 2.4. Let H be an abelian group and G a sofic group. Then H ≀ G (wreath
product) is sofic.
Proof. We shall work with the following presentation < SR > of the wreath product:
for every h ∈ H and g ∈ G};
S ={f h
R ={f e
g , ug :
g = e : ∀g ∈ G} ∪ {f h1
g1 f h2
g2 = f h2
g2 f h1
{f h1
g f h2
g = f h1h2
g
: ∀g ∈ G, ∀h1, h2 ∈ H}∪
g1 : ∀g1, g2 ∈ G, g1 6= g2∀h1, h2 ∈ H}∪
{ug1ug2 = ug1g2 : ∀g1, g2 ∈ G}∪
{ug1f h
g1 = f h
g2u−1
g1g2 : ∀g1, g2 ∈ G, h ∈ H}.
Consider first the case of Z2 ≀ G. Apply Elek-Lippner result to embed L(ZG
L(ZG
2
permutations. Instead, elements of the type f h
2 ) ⋊β G ≃
⋊ G) = L(Z2 ≀ G) in some Πk→ωMnk . Generators ug will be ultraproduct of
g are unitaries in Πk→ωDnk with ±1 entries.
Construct a sofic representation of Z2 ≀ G in Πk→ωP2nk by replacing a 1 entry with I2 and
a −1 entry with: 0 1
1 0 !.
12
Consider now the general case. Let Θ : L∞({0, 1}G) ⋊ G → Πk→ωMnk the sofic
embedding constructed in the last proof. Also let Λ : H → Pmk be a sofic embedding of
H. We shall construct Φ : H ≀ G → Πk→ωPnk ⊗ Pmk as follows:
Φ(ug) = Θ(g) ⊗ 1;
Φ(f h
g ⊗ 1 + c1
g ) = c0
g ⊗ Λ(h).
Relations in set R are easy to check (one needs H abelian for f h1
T r(Φ(ug)) = 0 and T r(Φ(f h
is injective.
g1 ). Also
g )) = 1/2. In order to finish the proof we need to see that Φ
g2 = f h2
g1 f h2
g2 f h1
The genereic element of H ≀ G is s = f h1
g1 f h2
g2 . . . f hn
gn ug with g1, g2 . . . gn distinct. Then:
Φ(s) =
X(i1,...,in)∈{0,1}n
g1,g2...gn ⊗ Λ(Πik=1hk)
ci1,i2...in
Assume Φ(s) = 1. Then for any (i1, . . . , in) ∈ {0, 1}n, Λ(Πik=1hk) = 1. This force hk = e
for any k. Then Φ(ug) = 1, so ug = e. It follows s = e.
(Θ(g) ⊗ 1).
3 Sofic actions
The goal would be to prove that every (free) action of a sofic group is sofic. While this
remains open we shall prove this fact for a family of groups. Let's first solve this ambiguity:
free or general actions.
Theorem 3.1. Let G be a group such that every free action is sofic. Then every action
of G is sofic.
Proof. Let α be an action of G on X. Let β : G → Aut(Y ) be a free action (eg. Bernoulli
shift). Define α ⊗ β : G → Aut(X × Y ) by g(x, y) = (gx, gy). With this definition
α ⊗ β is a free action of G, so it is sofic. We can embed L∞(X × Y ) ⋊α⊗β G in some
Πk→ωMnk satisfying the requirements of sofic action. The space L∞(X) can be embedded
in L∞(X ×Y ) by id⊗1. This embedding can be extended to an embedding of L∞(X)⋊α G
in L∞(X × Y ) ⋊α⊗β G. This will prove α is sofic.
Definition 3.2. Denote by S the class of groups for which every action is sofic.
13
While we can not prove that every sofic group is in S, we will provide some examples.
First goal is to deal with amenable groups.
Proposition 3.3. Each action of the integers admits a sofic embedding.
Proof. Let α : L∞(X) → L∞(X) the automorphism that generates the action. Choose
Θ : L∞(X) → Πk→ωDnk an embedding. Apply proposition 1.19 to Θ and Θ ◦ α to get a
unitary u ∈ Πk→ωPnk such that Adu ◦ Θ = Θ ◦ α.
As powers of permutation matrices are still permutation matrices, we have um ∈
Πk→ωPnk . Also umΘ(f )(um)∗ = Θ(αm(f )) for any f ∈ L∞(X). Now we have an
embedding Θ of the algebraic crossed product L∞(X)⋊α Z. In order to have an embedding
of the crossed product we need the relation T r(um) = 0 for any m ∈ Z∗.
Let Λ be an embedding of Z in some Πk→ωPrk using only elements of trace 0. Define
the embedding Θ ⊗ Λ of L∞(X) ⋊α Z in Πk→ωMnk·rk by:
Θ ⊗ Λ(T ) =Θ(T ) ⊗ 1
Θ ⊗ Λ(ug) =Θ(ug) ⊗ Λ(ug)
for T ∈ L∞(X)
for g ∈ Z.
This embedding Θ ⊗ Λ of the algebraic crossed product respects the trace of the von
Neumann crossed product. Using the unique feature of the type II case the closure of its
imagine will be the crossed product.
Proposition 3.4. Amenable groups are in S.
Proof. Let G be an amenable group and let α : G → Aut(X, µ) be a free action. Then
Eα is amenable. By [CFW] Eα is generated by an action β of Z. By previous proposition
beta is sofic. Because almost all equivalence classes of Eα are non-finite, β is free. Using
proposition 1.5 we deduce α is sofic. Combined with theorem 3.1, we get G ∈ S.
Next proposition will enlarge the class of groups for which such results hold.
Theorem 3.5. Let α1 and α2 be two sofic actions of G1 and G2 on the same space X.
Consider H, a common amenable subgroup of G1 and G2. Assume α1 and α2 coincide on
H, and this action of H is free. Then the action α1 ∗H α2 of G1 ∗H G2 is sofic.
14
Proof. Using 1.16 we can construct sofic embeddings of the two crossed products in the
same ultraproduct. So let Θi : L∞(X) ⋊ Gi → Πk→ωMnk , i = 1, 2. By 1.20 we can
assume Θ1 = Θ2 on L∞(X) ⋊ H (using the freeness of this action). Now we can construct
a representation Θ of the algebraic crossed product L∞(X) ⋊ (G1 ∗H G2) on Πk→ωMnk .
In order to embed the von Neumann crossed product the trace of each nontrivial ug,
g ∈ G1 ∗H G2 must be equal to 0. Because L(G) ⊂ L∞(X) ⋊ G we know that G1
and G2 are sofic. Then G1 ∗H G2 is sofic (see 2.3). There exist an embedding Λ of
G1 ∗H G2 in some Πk→ωPrk using only elements of trace 0. Define the embedding Θ ⊗ Λ
of L∞(X) ⋊α1∗H α2 G1 ∗H G2 like in 3.3.
Adapting the same methods we can prove this result for a countable family of actions.
Proposition 3.6. Let {αi}i∈N be a family of sofic actions of {Gi}i∈N on the same space.
Assume H is an amenable common subgroup of Gi and the actions αi coincide on H.
Then ∗H αi is sofic.
Corollary 3.7. Each action of a free group, including F∞ is sofic.
Proof. Corollary of 3.3 and 3.6.
We now recover with our methods the result of Elek-Lippner that any treeable
equivalence relation is sofic.
Proposition 3.8. Every treeable equivalence relation is sofic.
Proof. Well, treeable is some kind of freeness and freeness in general goes well with soficity.
Let E be a treeable equivalence relation on (X, µ). Fix a treeing of E, i.e. a countable
set of partial Borel isomorphism {φi}i∈N∗ ⊂ [[E]]. For each i we have φi = aiλi, where ai
is a projection in L∞(X) and λi ∈ [E].
Define an action α of F∞ on X such that α(γi) = λi (where {γi}i are the generators
of F∞). Being an action of F∞, α is sofic.
The von Neumann subalgebra of L∞(X) ⋊α F∞ generated by aiuγi
is naturally
isomorphic to M (E). Hence every sofic embedding of L∞(X) ⋊α F∞ can be restricted
to a sofic embedding of M (E) ⊂ L∞(X) ⋊α F∞.
15
We end this section with the following theorem.
Theorem 3.9. Class S is closed under amalgamated product over amenable groups. It is
strictly larger than the class of treeable groups.
Proof. First part of the theorem is 3.5 and 3.1. By 3.8 (and again 3.1) every treeable
group is in S.
Consider now the group G = Z ∗(2,3)Z Z. It is not treeable but G ∈ S. This example
is from [Ga].
Relation G ∈ S is just 3.3 and 3.5. By general theory of Gaboriau, the cost of G is
1 + 1 − 1 = 1. If amalgamation is done with good morphism (multiplication by 2 and 3)
then G is not amenable. This implies G is not treeable.
4 Sofic equivalence relations
Now we shall present from [El-Li] the original definition of Elek and Lippner of soficity for
actions and equivalence relations.
Definition 4.1. We call a basic sequence of projections for L∞(X) a collection
{ei,m}1≤i≤2m,m≥0 ⊂ L∞(X) with following properties:
1. spanw{ei,m}i,m = L∞(X);
2. µ(ei,m) = 2−m, 1 ≤ i ≤ 2m, m ≥ 0;
3. e2i−1,m + e2i,m = ei,m−1, m ≥ 1.
Let F∞ =< γ1, γ2, . . . >. For any r ∈ N denote by Wr the subset of reduced words
of length at most r containing only the first r generators and their inverses. We have
W0 ⊂ W1 ⊂ . . . and F∞ = ∪r≥0Wr.
Let α : F∞ y X a Borel action and fix {ei,r}1≤i≤2r a basic sequence of projections
for L∞(X). The following definition will allow us to keep track of the position of a point
x ∈ X relative to sets {ei,r}1≤i≤2r under the action of Wr ⊂ F∞.
Definition 4.2. Let r ∈ N. A r-labeled, r-neighborhood is a finite oriented multi-graph
containing:
16
1. a root vertex such that any vertex is connected to the root by a path of lenght at
most r;
2. ever vertex has a label from the set {1, . . . , 2r};
3. out-edges of every vertex have different colors from the set {γ1, γ−1
1 , . . . , γr, γ−1
r };
4. if edge xy is colored with γi then yx is colored by γ−1
i
.
Isomorphism classes of such objects form a finite set that we shall denote by U r,r.
For G ∈ U r,r, denote by RG the root vertex in G. For γ ∈ Wr let γRG be the vertex in
G obtained by starting from RG and following the path given by γ (if such a path exists).
Finally, let l(γRG) be the label of the vertex γRG in the set {1, 2, . . . , 2r}.
Let X be a space together with a basic sequence of projections. For an action α :
r (x) ∈ U r,r by taking the imagines of x under Wr and
r (x) ≡ G}.
F∞ y X and x ∈ X we can define Br
their labels with respect to {ei,r}1≤i≤2r . For G ∈ U r,r let T (α, G) = {x ∈ X : Br
Define also pG(α) = µ(T (α, G)).
If α is an action on a finite space Y (having the normalized cardinal measure) we
have the same definitions provided that we still have some subsets {ei,r}1≤i≤2r ,r≥0 of Y
satisfying the same summation relations. This are needed to give labels to our vertices.
Finite spaces with this kind of partitions are called X-sets. We are now ready to give the
definition.
Definition 4.3. An action α of F∞ is called sofic (in Elek-Lippner sense) if there exists
a sequence of actions αk of F∞ on X-sets such that for any r ≥ 1, for any G ∈ U r,r we
have limk→∞pG(αk) = pG(α).
Definition 4.4. An equivalence relation is called sofic if it is generated by a sofic action
of F∞ (all this in Elek-Lippner sense).
For actions of F∞ the two notions of soficity are different. With our definition every
action of F∞ is sofic (see 3.7). Instead for equivalence relations the two notions are the
same. This is what we shall prove now.
Proposition 4.5. Let E ⊂ X 2 be a sofic equivalence relation in sense of Elek-Lippner.
Then E is also sofic (M (E) admits a sofic embedding in some Πk→ωMnk ).
17
Proof. Let α : F∞ y (X, µ) be a sofic action in the sense of Elek-Lippner such that
E = Eα. Let αk a sequence of actions on X-sets Yk such that limk→∞pG(αk) = pG(α).
Finally let nk be the cardinal of Yk. We shall embed M (E) in Πk→ωMnk in a sofic way.
For this we need:
1. an embedding L∞(X) ⊂ Πk→ωDnk ;
2. a representation Θ of F∞ on Πk→ωPnk ;
3. the formula T r(f Θ(γ)) = RXγ
Xγ = {x ∈ X : γx = x}.
f dµ for every f ∈ L∞(X) and γ ∈ F∞, where
By hypothesis Yk are X-sets, so they come together with projections {ek
i,r}i,r. Construct
ei,r = Πk→ωek
i,r. We claim that {ei,r}1≤i≤2r ,r≥0 form a basic sequence of projections for
the algebra they generate. Relations ei,r = e2i−1,r+1 + e2i,r+1 are automatic, we just need
to prove that T r(ei,r) = 2−r.
Let {fi,r}1≤i≤2r,r≥0 ⊂ L∞(X) the basic sequence of projections used in the
construction of numbers pG(α). Fix i and r. Let U r,r
i = {G ∈ U r,r : l(RG) = i},
i.e. graphs such that the root has label i. Then T (α, G) ⊂ fi,r for each G ∈ U r,r
.
In the same way we have ek
Moreover: fi,r = ⊔G∈U r,r
T (αk, G).
Because limk→∞ pG(αk) = pG(α) we have T r(ei,r) = limk→ω T r(ek
i,r = ⊔G∈U r,r
i,r) = T r(fi,r) = 2−r.
T (α, G).
i
i
i
By identifying ei,r with fi,r we get an embedding of L∞(X). Now we construct the
representation Θ of F∞ in Πk→ωPnk . We identified set Yk with diagonal Dnk and we have
actions αk of F∞ that are defined on Yk. This will construct a representation. We need
to make sure Θ acts the same way as α.
Let γ be one of the generators of F∞. Fix i, j and r. Let U r,r
i,γj = {G ∈ U r,r :
l(RG) = i, l(γRG) = j)}, the set of graphs such that the root has label i and the vertex
connected with the root by the γ edge has label j (the existence of such an edge is a
requirement we ask now for G). It is easy to see that fi,r ∩ α(γ−1)(fj,r) = ⊔G∈U r,r
In the same way ek
T (αk, G). Using the hypothesis we get
T r(ei,r · Θ(γ−1)(ej,r)) = µ(fi,r ∩ α(γ−1)(fj,r)). This is enough to deduce that the action
that Θ induce on our embedding of L∞(X) is equal to α.
i,r ∩ αk(γ−1)(ek
j,r) = ⊔G∈U r,r
T (α, G).
i,γj
i,γj
For the third requirement let now γ ∈ F∞ an arbitrary element.
It is of course
sufficient to assume that f is one of the projections ei,r. We need to prove that
18
T r(ei,rΘ(γ)) = µ(Xγ ∩ ei,r). Lets say that in our construction we have Θ(γ) = Πk→ωγk.
Then T r(ei,rΘ(γ)) = limk→∞T r(ek
i,γ = {G ∈ U r,r : l(RG) = i, γRG = RG},
i.e. the set of G ∈ U r,r such that the root has label i and the path in G described by γ,
starting from the root, returns to the root. Then Xγ ∩ ei,r = ⊔G∈U r,r
formula with fixed points of γk and ek
T r(ei,rΘ(γ)) = µ(Xγ ∩ ei,r) and we are done.
i,r takes place. By limk→∞pG(αk) = pG(α) we get
i,rγk). Let U r,r
T (α, G). A similar
i,γ
Proposition 4.6. Let E be a sofic equivalence relation (M (E) embeds in some Πk→ωMnk ).
Then E is also sofic in the sense of Elek-Lippner.
Proof. By 1.14 we have a sofic embedding M (E) ⊂ Πk→ωMnk such that L∞(X) = A ⊂
Πk→ωDnk and uθ ⊂ Πk→ωPnk for any θ ∈ [E].
We shall denote by d the normalized Hamming distance on Pnk . In general γ, δ will
denote elements in F∞ and γi, δi will denote generators of F∞. Let α : F∞ y (X, µ) an
action that generates the equivalence relation E on X. For any element γ ∈ F∞, α(γ)
induce an element uγ ∈ N (A) and uγuδ = uγδ. We shall write uγ = Πk→ωuk
γ ∈ Πk→ωPnk .
Let Yk be a set with nk elements and identify algebra Dnk with L∞(Yk). For any
γi ∈ Pnk induce an automorphism of Yk. Denote it by αk(γi) and
generator γi of F∞, uk
extend αk by multiplicity to an action of F∞.
Let {ei,m}1≤i≤2m,m≥0 ⊂ L∞(X) a basic sequence of projections. Use it in order to
i,m}i,m respect the same
i,m}i,m so we have
construct sets T (α, G). Write ei,m = Πk→ωek
summation relations. Now elements in Yk are labelled by projections {ek
ingredients for constructing sets T (αk, G).
i,m such that {ek
We need to show that out of this actions we can find a subsequence satisfying the
definition of soficity, namely that limk→∞µnk(T (αk, G)) = µ(T (α, G)) for any r ∈ N and
any G ∈ U r,r (denote by µnk the normalized cardinal measure on a set with nk elements).
The subsequence is just to get rid of the ultrafilter and obtain classical limit for the
countable set of objects that we are working with.
Fix r ∈ N and ε > 0. Let us see that it is enough to find k ∈ N such that
µnk (T (αk, G))−µ(T (α, G)) < ε for any G ∈ U r,r. The phenomena here is that, if we fix a
G ∈ U r,r, when we pass from step r to r + 1 we have T (α, G) = ∪G′∈U r+1,r+1;G<G′T (α, G′)
(relation G < G′
is defined in an obvious way). So µ(T (αk′, G)) is a sum of other
19
µ(T (αk′, G′)), but a finite sum. When we choose our sequence {εr} we have to make
sure that it compensates this growth.
Sets {T (α, G) : G ∈ U r,r} form a partition of X. Let T (α, G) = Πk→ωT (α, G)k
such that {T (α, G)k : G ∈ U r,r} is a partition of Yk. We also have in Dnk projections
T (αk, G). We know that µnk (T (α, G)k) →k µ(T (α, G)) and we want to show that
µnk (T (αk, G)) →k µ(T (α, G)) .
Now fix G ∈ U r,r. We need to understand equations that describe points in T (α, G).
Remember that RG is the root vertex in G; for γ ∈ Wr, γRG is the vertex in G obtained
by starting from RG and following the path given by γ (if such a path exists). Finally,
l(γRG) is the label of the vertex γRG in the set {1, 2, . . . , 2r}. We can now state our
characterization of T (α, G).
A point x ∈ X is an element of the set T (α, G) iff:
1. α(γ)(x) ∈ el(γRG),r for any γ ∈ Wr for which γRG exists;
2. α(γ)(x) = α(δ)(x) ∀γ, δ ∈ Wr, γRG = δRG;
3. α(γ)(x) 6= α(δ)(x) ∀γ, δ ∈ Wr, γRG 6= δRG.
First condition gives the coloring of vertices. The other two give the structure of the
graph G. Let ε1 > 0 such that 2U r,r(Wr + 2Wr2)ε1 < ε. We want to find k ∈ N such
that for any G ∈ U r,r we have:
l(γRG),r(cid:17) < ε1
µnk(cid:16)αk(γ)(T (α, G)k ) \ ek
µnk(cid:16)T (α, G)k \ {y ∈ Yk : αk(γ)(y) = αk(δ)(y)}(cid:17) < ε1
µnk(cid:16)T (α, G)k \ {y ∈ Yk : αk(γ)(y) 6= αk(δ)(y)}(cid:17) < ε1
µnk (T (α, G)k) − µ(T (α, G)) < ε/2.
∀γ ∈ Wr;
∀γ, δ ∈ Wr, γRG = δRG
∀γ, δ ∈ Wr, γRG 6= δRG
(1)
(2)
(3)
(4)
First we shall prove that this four conditions are enough to guarantee µnk (T (αk, G))−
µ(T (α, G)) < ε for every G ∈ U r,r. Using (4) we just need to prove µnk (T (αk, G)) −
µnk (T (α, G)k) < ε/2.
Take x ∈ (T (α, G)k \ T (αk, G)) for some G ∈ U r,r. Following our characterization of
T (α, G), we have:
20
1. ∃γ ∈ Wr such that αk(γ)(x) does not have the right label, namely l(γRG);
2. or ∃γ, δ ∈ Wr such that γRG = δRG and αk(γ)(x) 6= αk(δ)(x);
3. or ∃γ, δ ∈ Wr such that γRG 6= δRG and αk(γ)(x) = αk(δ)(x).
Using (1), (2) and (3) we get:
µnk (T (α, G)k \ T (αk, G)) < Wrε1 + 2Wr2ε1.
Because both {T (αk, G)}G and {T (α, G)k)}G are partitions of Yk and the above formula
holds for any G ∈ U r,r we have:
µnk (T (αk, G)) − µnk (T (α, G)k) < U r,r(Wrε1 + 2Wr2ε1) < ε/2.
γζ2
i2
. . . γζs
is
Now back to the choice of k. Let γ = γi1γi2 . . . γis ∈ Wr. We should in fact
take γ = γζ1
, where ζj ∈ {±1}. The inverses will change nothing in our
i1
arguments and will only overload our notations. Due to Feldman-Moore construction
. . . uγis . Next, combine α(γ)(T (α, G)) ⊂ el(γR),r and
uγi2
we know that uγ = uγi1
α(γ)(T (α, G)) = uγT (α, G)u∗
γ to get uγT (α, G)u∗
γ ⊂ el(γR),r.
Consider now γ, δ ∈ Wr.
If γRG = δRG then α(γ)T (α,G) = α(δ)T (α,G) so
δ uγ) = µ(T (α, G)) (here we consider T (α, G) to be a projection of L∞(X) ⊂
δ uγ) = 0. Find k ∈ N such that (4) holds
T r(T (α, G)u∗
Πk→ωMnk ). If γRG 6= δRG then T r(T (α, G)u∗
and:
uk
uk
γi2
2 < ε1/4
γ − uk
. . . uk
γis
γi1
γ \ ek
γT (α, G)kuk∗
µnk(cid:16)uk
µnk (T (α, G)k) − T r(cid:16)T (α, G)kuk∗
T r(cid:16)T (α, G)kuk∗
γ(cid:17) < ε1/2
δ uk
l(γRG),r(cid:17) < ε1/2
δ uk
γ(cid:17) < ε1/2
∀γ = γi1γi2 . . . γis ∈ Wr
∀γ ∈ Wr, ∀G ∈ U r,r
∀G ∈ U r,r∀γ, δ ∈ Wr, γRG = δRG
∀G ∈ U r,r∀γ, δ ∈ Wr, γRG 6= δRG
(5)
(6)
(7)
(8)
By definition αk(γ) = αk(γi1)αk(γi2) . . . αk(γir ) and αk(γij )(P ) = uk
γij
P uk∗
γij
for any
projection P ∈ Dnk . Then:
αk(γ)(T (α, G)k) =(cid:16)uk
γi1
uk
γi2
. . . uk
γir(cid:17) T (α, G)k(cid:16)uk
γi1
uk
γi2
. . . uk
γir(cid:17)∗
.
(9)
21
Both uk
γ and uk
γi1
uk
γi2
. . . uk
γis
are elements in Pnk so by (5):
d(uk
γ, uk
γi1
uk
γi2
. . . uk
γis
) < ε1/4.
Combined with (9), we have µnk(cid:0)αk(γ)(T (α, G)k ) \ uk
get µnk(cid:16)αk(γ)(T (α, G)k) \ ek
l(γR),r(cid:17) < ε1/4 + ε1/2 < ε1, so we have (1).
γT (α, G)kuk∗
γ (cid:1) < ε1/4. Use (6) to
Let now γ = γi1γi2 . . . γis and δ = δj1δj2 . . . δjt such that γRG = δRG. Use (5) both
for γ and δ to get:
uk∗
δ uk
γ − (uk
δj1
uk
δj2
. . . uk
δjt
)∗(uk
γi1
uk
γi2
. . . uk
γir
)2 < ε1
As before, uk∗
δ uk
γ and (uk
δj1
. . . uk
δjt
)∗(uk
γi1
. . . uk
γir
) are elements in Pnk , so:
δ uk
γ, (uk
δj1
uk
δj2
. . . uk
δjt
)∗(uk
γi1
uk
γi2
. . . uk
γir
d(cid:16)uk∗
)(cid:17) < ε1/2,
Restricting this inequality just to fixed points in set T (α, G)k, we get:
T r(T (α, G)kuk∗
δ uk
γ) − T r(cid:16)T (α, G)k(uk
δj1
uk
δj2
. . . uk
δjt
)∗(uk
γi1
uk
γi2
. . . uk
γir
)(cid:17) < ε1/2.
(10)
Apply (9) for γ and δ to get:
µnk(cid:16)T (α, G)k \ {y ∈ Yk : αk(γ)(y) = αk(δ)(y)}(cid:17) = µnk(T (α, G)k)−
)(cid:17) .
T r(cid:16)T (α, G)k(uk
. . . uk
δjt
)∗(uk
γi1
. . . uk
γir
uk
γi2
uk
δj2
δj1
This combined with (10) and(7) yields (2).
Assume now γRG 6= δRG. Then:
µnk (T (α, G)k\{y ∈ Yk : αk(γ)(y) 6= αk(δ)(y)}) =
)∗(uk
γi1
. . . uk
δjt
uk
δj2
δj1
T r(cid:16)T (α, G)k(uk
uk
γi2
. . . uk
γir
)(cid:17) .
Inequalities (10) and (8) will imply (3).
Acknowledgements
It is my great pleasure to thanks Florin Radulescu for many discussions and ideas. Special
thanks to my friend and colleague Valerio Capraro for introducing me to the subject of
22
Connes' embedding problem and for his study and work on the subject from which I
benefited. I am very grateful to Stefaan Vaes for numerous remarks and corrections on
previous versions of the paper and also for considerably easier proofs for some results
including lemma 1.19 and lemma 3.3. Parts of this article were written during my stay
in Leuven in Spring 2010. Also, I want to thank Damien Gaboriau and Ken Dykema for
useful remarks.
References
[Ca-Pa] V. Capraro - L. Paunescu, Product Between Ultrafilters and Applications to the
Connes' Embedding Problem arXiv:0911.4978 (2009)
[CFW] A. Connes - J. Feldman - B. Weiss, An amenable equivalence relation is generated
by a single transformation, Erg. Theory Dyn. Sys. 1(1981), 431-450.
[Co-Dy] B. Collins - K. Dykema, Free products on sofic groups with amalgamation over
monotileably amenable groups arXiv:1003.1675 (2010).
[El-Li] G. Elek - G. Lippner, Sofic equivalence relations, arXiv:0906.3619 (2009).
[El-Sz1] G. Elek - E. Szabo, On sofic groups, J. Group Theory 9 (2006), Issue 2, 161-171.
[El-Sz2] G. Elek - E. Szabo, Sofic representations of amenable groups, arXiv:1010.3424
(2010).
[Fe-Mo] J. Feldman - C. C. Moore, Ergodic equivalence relations, cohomology, and von
Neumann algebras, II. Trans. Amer. Math. Soc. 234 (1977) no. 2, 325-359.
[Ga] D. Gaboriau, Cout des relations d'quivalence et des groupes, Invent. Math., 139
(2000), no. 1, 41-98.
[Ke-Mi] A. Kechris - B. Miller, Topics in orbit equivalence, Lecture Notes in Mathematics
1852 Springer Verlag.
[Oz] N. Ozawa, Unpublished lectures notes, available at http://people.math.jussieu.fr/
pisier/taka.talk.pdf
23
[Pe] V. Pestov, Hyperlinear and sofic groups: a brief guide, Bull. Symbolic Logic 14 (2008)
no. 4, 449480.
[Po1] S. Popa, On a Problem of R.V. Kadison on Maximal Abelian *-Subalgebras in
Factors, Inv. Math., 65 (1981), 269-281.
[Po2] S. Popa, Notes on Cartan subalgebras in type II1 factors, Math. Scand., 57 (1985),
171-188.
[Ra] F. Radulescu, The von Neumann algebras of the non-residually finite Baumslag group
< a, bab3a−1 = b2 > embeds into Rω, arXiv:math/0004172v3 (2000).
L.
P AUNESCU,
UNIVERSIT `A
DI
ROMA
TOR
VERGATA
and
INSTITUTE of MATHEMATICS "S. Stoilow" of the ROMANIAN ACADEMY email:
[email protected]
24
|
1505.01766 | 3 | 1505 | 2016-07-19T03:13:35 | Inverse semigroups associated to subshifts | [
"math.OA",
"math.CT",
"math.DS"
] | The dynamics of a one-sided subshift $\mathsf{X}$ can be modeled by a set of partially defined bijections. From this data we define an inverse semigroup $\mathcal{S}_{\mathsf{X}}$ and show that it has many interesting properties. We prove that the Carlsen-Matsumoto C*-algebra $\mathcal{O}_\mathsf{X}$ associated to $\mathsf{X}$ is canonically isomorphic to Exel's tight C*-algebra of $\mathcal{S}_{\mathsf{X}}$. As one consequence, we obtain that $\mathcal{O}_\mathsf{X}$ can be written as a partial crossed product of a commutative C*-algebra by a countable group. | math.OA | math |
Inverse semigroups associated to subshifts
Charles Starling∗
Abstract
The dynamics of a one-sided subshift X can be modeled by a set of partially defined
bijections. From this data we define an inverse semigroup SX and show that it has
many interesting properties. We prove that the Carlsen-Matsumoto C*-algebra OX
associated to X is canonically isomorphic to Exel's tight C*-algebra of SX. As one
consequence, we obtain that OX can be written as a partial crossed product of a
commutative C*-algebra by a countable group.
1
Introduction
An inverse semigroup is a semigroup S together with an involution ∗ : S → S such that for
all s ∈ S we have
ss∗s = s.
and such that s∗ is the only element for which this equation holds. On the other hand, a
partial isometry in a C*-algebra A is an element v such that
vv∗v = v.
The link between inverse semigroups and partial isometries in C*-algebras implied by the
above is hard to ignore, especially considering the mass of important examples of C*-
algebras which are generated by partial isometries. In fact, for many C*-algebras of interest
one can choose a countable generating set consisting of partial isometries which is also closed
under product and adjoint -- such a set is necessarily an inverse semigroup.
On the one hand, C*-algebras have provided interesting examples of inverse semigroups.
For examples of this, we can look to the graph inverse semigroups of [Pat02] and [LJ14],
the tiling inverse semigroup of [Kel97] and the AF inverse semigroups of [LS14]. As one
can imagine, each of these appears as a generating set of partial isometries in its namesake
C*-algebra.
On the other hand, if one knows a certain C*-algebra A is generated by an inverse
semigroup S, then (semigroup-theoretical) properties of S give rise to properties of A. The
catch is that S may not tell the whole story. For example, the Cuntz algebra O2 and
its Toeplitz extension T2 are both generated by inverse semigroups of partial isometries,
∗Supported by the NSERC grants of Benoıt Collins, Thierry Giordano, and Vladimir Pestov.
[email protected].
1
and both of these inverse semigroups are isomorphic (as semigroups) to the same inverse
semigroup P2 (called a polycyclic monoid in the literature). What is happening is that there
is a representation of P2 in both algebras, but the relation s1s∗
2 = 1 which holds in
the Cuntz algebra cannot be expressed using only the multiplication and involution inside
the inverse semigroup. Therefore, if one hopes to phrase simplicity (for example) of a C*-
algebra in terms of properties of a generating inverse semigroup, one has to care for how
the inverse semigroup is represented in the C*-algebra.1
1 + s2s∗
This led Exel in [Exe08] to define the notion of a tight representation of an inverse semi-
group. He showed that for an inverse semigroup S there always exists a C*-algebra, called
the tight C*-algebra of S and denoted C ∗
tight(S), which is universal for tight representations
of S. It turns out that O2 is universal for tight representations of P2, and T2 is universal for
all representations of P2 (a concept introduced by Paterson in [Pat99]). Many C*-algebras
of interest are isomorphic to the tight C*-algebra of their generating sets -- for example
see [EGS12] for tiling C*-algebras, [Exe08] for graph and higher rank graph C*-algebras,
[Sta15] for boundary quotients of certain Cuntz-Li algebras, and [?] for Katsura algebras
and self-similar group algebras.
tight(S) is simple, and in the case that C ∗
Hence, there has been interest in relating the properties of an inverse semigroup to
properties of its tight C*-algebra. The paper [EP14] of Exel and Pardo provides conditions
on S which guarantee that C ∗
tight(S) is nuclear these
conditions become necessary and sufficient (invoking the results of [BCFS14] regarding an
underlying groupoid). They also give a condition which further guarantees that C ∗
tight(S)
is purely infinite. Similar results are obtained in [Ste14]. Work of Milan and Steinberg
[MS14] gives conditions on S which imply that C ∗
tight(S) is isomorphic to the partial crossed
product of a commutative C*-algebra by a group, and further conditions which imply it is
Morita equivalent to a usual crossed product.
With that setup, we turn to the present paper. We are concerned with one-sided sub-
N which are invariant under the left
shifts over a finite alphabet a, ie closed subspaces of a
shift map. In [Mat97], Matsumoto associated a C*-algebra to such a space X which general-
ized the construction of Cuntz-Krieger C*-algebras [CK80] (which can be naturally viewed
as C*-algebras associated to shifts of finite type).
In a subsequent paper with Carlsen
[CM04] a slightly different construction was put forward, and again in [Car08]. We will
deal with the C*-algebra OX defined in [Car08], and call this a Carlsen-Matsumoto algebra.
This C*-algebra has been viewed under the lens of many different constructions: in [Car08]
it is presented as a Cuntz-Pimsner algebra, in [CS07] it is obtained as an Exel crossed
product by an endomorphism, in [Car04] it is obtained from a Renault-Deaconu groupoid,
and in [Tho10] Thomsen constructs it from a semi-´etale groupoid.
Here, we add another construction to the list, by constructing an inverse semigroup SX
from X and showing that OX is isomorphic to C ∗
tight(SX). Our motivations for doing so are to
add another example of a C*-algebra which can be seen as the tight C*-algebra of an inverse
semigroup and also to provide another interesting example of an inverse semigroup arising
from a C*-algebra. The other reason mentioned above that one might want to embark
on such an investigation -- that properties of the algebra can be gleaned from that of the
1It turns out that another way of approaching this problem is to specialize further to a class of inverse
semigroups called boolean inverse monoids. We do not pursue this here, and the interested reader is directed
to [LL13] for more details.
2
inverse semigroup -- is less pressing in this case, as Carlsen-Matsumoto algebras are already
quite well-studied. For instance, the papers [Tho10] and [CT12] combine to provide sharp
conditions under which OX is simple and purely infinite. We do however use the results
of [MS14] to show that OX can be seen as a partial crossed product of a commutative
C*-algebra by the free group over a.
This paper is organized in the following manner. After providing some background,
in Section 3 we define our inverse semigroup SX from X, and show that it satisfies some
nice properties. Section 4 is first devoted to establishing the isomorphism between OX and
C ∗
tight(SX), and then finishes by mentioning the partial crossed product result mentioned
above.
2 Preliminaries and notation
We will use the following general notation. If X is a set and U ⊂ X, let IdU denote the
map from U to U which fixes every point, and let 1U denote the characteristic function on
U, ie 1U : X → C defined by 1U (x) = 1 if x ∈ U and 1U (x) = 0 if x /∈ U. If F is a finite
subset of X, we write F ⊂fin X. We let N denote the set of natural numbers (starting at
1).
2.1
Inverse semigroups
An inverse semigroup is a semigroup S such that for every s ∈ S, there exists a unique
element s∗ ∈ S, with the property that
ss∗s = s,
s∗ss∗ = s∗.
The element s∗ is called the inverse of S. For s, t ∈ S, we have (s∗)∗ = s and (st)∗ = t∗s∗.
We typically assume that S has a neutral element 1 and a zero element 0 such that
1s = s1 = s for all s ∈ S
0s = s0 = 0 for all s ∈ S.
Even though we call s∗ the inverse of s, we need not have ss∗ = 1, although we always have
that (ss∗)2 = ss∗ss∗ = ss∗, which is to say that ss∗ (and indeed s∗s) is an idempotent. The
set of all idempotents in S is denoted
E(S) = {e ∈ S e2 = e}.
It is a nontrivial fact that if S is an inverse semigroup, then E(S) is closed under multipli-
cation and commutative. It is also clear that if e ∈ E(S), then e∗ = e.
Let X be a set, and let
I(X) = {f : U → V U, V ⊂ X, f bijective}.
Then I(X) is an inverse semigroup when given the operation of composition on the largest
possible domain and inverse given by function inverse; it is called the symmetric inverse
3
monoid on X. If e is an idempotent in I(X), then e = IdU for some U ⊂ X. The function
IdX is the neutral element for I(X), and the empty function is the 0 element for I(X). It
is an important fact (akin to the Cayley theorem for groups) that every inverse semigroup
is embeddable in I(X) for some set X -- this is the Wagner-Preston theorem.
Every inverse semigroup possesses a natural order structure. For an inverse semigroup
s, t ∈ S we say s 6 t if and only if ts∗s = s. On idempotents, this order has a nicer form -- if
e, f ∈ E(S) then e 6 f if and only if ef = e. This partial order is perhaps best understood
for elements of I(X), because if g, h ∈ I(X), then g 6 h if and only if h extends g as a
function.
2.2 Subshifts
As much as possible we use notation set in [Car08, CS07]. Let a be a finite set, called the
alphabet, and endow it with the discrete topology. The product space
a
N =Yn∈N
a
is called the one-sided full shift over a. If x = (xn)n∈N, we will write x in the shorter form
x = x1x2x3 · · · .
N → a
N given by σ(x1x2x3 · · · ) = x2x3 · · ·
is called the shift, and is a
The map σ : a
N is called a subshift if it is closed and σ(X) ⊂ X.
continuous surjection. A subspace X ⊂ a
If this is the case, we will also sometimes say that X is a one-sided subshift over a. Since
N is compact and metrizable, then so is any subshift over a.
a
For an integer k ≥ 1, we let a
k denote the set of words of length k in elements of a -- we
k as w1w2 · · · wk. We also let a0 = {ǫ}, and call ǫ the empty
k we write w = k and say that the length of w is k. We set a∗ = ∪k≥0a
k
again write an element w ∈ a
word. For w ∈ a
and say this is the set of words in a. Given v, w ∈ a∗, we may form their concatenation
vw = v1v2 · · · vvw1w2 · · · ww ∈ a
∗.
In addition, for all v ∈ a∗, we take vǫ = ǫv = v. Given v ∈ a∗ and x ∈ a
concatenate v and x:
N, we may also
vx = v1v2 · · · vvx1x2 · · · ∈ a
N.
Again, for all x ∈ a
is a prefix of y. For x ∈ X and k ∈ N we let
N we let ǫx = x. If v ∈ a∗, x ∈ a∗ ∪ a
N and y = vx, then we say that v
x[1,k] = x1x2 · · · xk,
x(k,∞) = xk+1xk+2 · · · .
In addition, if F ⊂ a∗ and w ∈ a∗ we let F w = {f w f ∈ F } and wF = {wf f ∈ F }.
For v ∈ a∗, we let C(v) = {vx ∈ a
N} and call sets of this form cylinder sets.
N. If X is a one-sided
These sets are closed and open in a
subshift over a and v ∈ a∗, then we set CX(v) = C(v) ∩ X, although the subscript will
frequently be dropped.
N, and generate the topology on a
N x ∈ a
4
3
Inverse semigroups associated to subshifts
Given a one-sided subshift X, the shift map σ : X → X is continuous, but in general it is
not a local homeomorphism. Still, it is locally a bijection, in that σC(a) is a bijection for
all a ∈ a. As mentioned in Section 2.1, inverse semigroups are a natural object with which
to study partially defined bijections, and so we use the partial bijections above to associate
an inverse semigroup SX to X.
3.1 Construction of SX
Following [Car08], for µ, ν ∈ a∗, we let
C(µ, ν) = {νx ∈ X µx ∈ X}
and notice that C(µ, µ) = C(µ). We note that
C(µ, ν) = C(ν) ∩ σ−ν(σµ(C(µ))).
Since the shift map is a closed map, C(µ, ν) is closed for every µ, ν ∈ a∗.
For each a ∈ a, let sa ∈ I(X) be defined by
sa : C(a, ǫ) → C(a, a)
sa(x) = ax.
For µ ∈ a∗ \ {ǫ}, we define sµ = sµ1sµ2 · · · sµµ so that
sµ : C(µ, ǫ) → C(µ)
sµ(x) = µx.
For the empty word ǫ, we take sǫ = IdX. It is clear that for each µ ∈ a∗, the map sµ is a
bijection between subsets of X.
Definition 3.1. Let X be a one-sided subshift over a. Then we let SX be the inverse
semigroup generated by {sǫ, sa a ∈ a} inside I(X).
We would like to find a convenient closed form for elements of SX. To this end, for each
F ⊂fin a∗ and ν ∈ a∗, let2
C(F ; ν) = {νx ∈ X f x ∈ X for all f ∈ F }
= \f ∈F
C(f, ν),
and let E(F ; ν) = IdC(F ;ν). We will also let E(µ, ν) = IdC(µ,ν). A short calculation shows
that
E(µ, ν) = sνs∗
µsµs∗
ν
2Late in preparation for this work, we discovered that sets of this form were already considered in
[Tho10], and were written C ′(ν; F ). We use our notation in solidarity with [Car08] and keep the "prefix"
data in the second entry and the "possible replacement prefixes" data in the first entry.
5
This product will be 0 unless v is a prefix of w or vice-versa. If w = vz, then
s∗
s∗
s∗
gsg! s∗
w
f sf! s∗
vsw Yg∈G
E(F ; v)E(G; w) = sv Yf ∈F
vsvsz Yg∈G
f sf! s∗
E(F ; v)E(G; w) = sv Yf ∈F
z Yf ∈F
f sf! szs∗
= sw Yf ∈F
f sf sz! s∗
f zsf z! Yg∈G
= sw Yf ∈F
gsg! s∗
vsvsz Yg∈G
wsw Yg∈G
gsg! s∗
gsg! s∗
= svszs∗
zs∗
s∗
zs∗
s∗
w
s∗
vz
s∗
s∗
s∗
gsg! s∗
s∗
vz
w
E(F ; ν) = sν Yf ∈F
f sf! s∗
ν.
s∗
The collection of all such elements will be important in the sequel -- we use the notation
EX = {E(F ; v) ∈ I(X) v ∈ a
∗, F ⊂fin a
∗} ∪ {∅}.
(1)
We note that the identity function on X is an element of E(X), taking E(F ; v) with v = ǫ
and F = {ǫ}. We also note that E(F ; v)E(G; w) 6= 0 if and only if C(F ; v) ∩ C(G; w) 6= ∅.
Lemma 3.2. If X is a one-sided subshift over a, then the set EX is closed under multipli-
cation. Furthermore, if w ∈ a∗ and e ∈ EX, then s∗
wesw ∈ EX.
Proof. Suppose that F, G ⊂fin a∗ and that v, w ∈ a∗. Then
(2)
= E(G ∪ F z; w).
A similar calculation shows that if v = wz, then
E(F ; v)E(G; w) = E(F ∪ Gz; v).
f sf! s∗
s∗
vsw.
Furthermore,
s∗
wE(F ; v)sw = s∗
wsv Yf ∈F
6
This is zero unless v is a prefix of w or vice-versa. If w = vz, then
s∗
wE(F ; v)sw = s∗
vsw
s∗
s∗
zs∗
= s∗
wsv Yf ∈F
f sf! s∗
vsv Yf ∈F
f sf! s∗
z Yf ∈F
f sf! s∗
= Yf ∈F
f sf sz! s∗
= s∗
vsvsz
zs∗
s∗
zs∗
s∗
= E({w} ∪ F z; ǫ)
vsvsz
vsvsz
If v = wz, then
s∗
wE(F ; v)sw = s∗
f sf! s∗
zs∗
wsw
zs∗
wsw
s∗
wswsz Yf ∈F
= sz Yf ∈F
f sf! s∗
= E(F ; z)E(w, ǫ)
= E(F ∪ {wz}; z)
s∗
where the last line is by our previous calculation.
Proposition 3.3. Let X be a one-sided subshift over a. Then
SX = {sαE(F ; v)s∗
β ∈ I(X) α, β, v ∈ a
∗, F ⊂fin a
∗} ∪ {0}.
(3)
Proof. We note that the containment "⊇" is trivial, because each element of the right hand
side is a finite product of elements from {sǫ, sa a ∈ a}. Hence, we will be done if we can
show that the set on the right hand side of (3) is itself an inverse semigroup, because the
right hand side contains {sǫ, sa a ∈ a}, and SX is the smallest inverse semigroup containing
these elements.
It is clear that (3) is closed under inverses, so we only need to show that it is closed
under multiplication. To this end, take α, β, δ, η, v, w ∈ a∗ and F, G ⊂fin a∗. The product
η) will again only be nonzero if β is a prefix of δ or vice-versa.
(cid:0)sαE(F ; v)s∗
If δ = βγ, then
β(cid:1) (sδE(G; w)s∗
(sαE(F ; v)s∗
β)(sδE(G, w)s∗
η) = sαE(F ; v)s∗
βsβsγE(G, w)s∗
η
= sαsγs∗
γE(F ∪ {βv}; v)sγE(G, w)s∗
η
and so by Lemma 3.2, this product is in SX. A similar argument applies to the case that
β = δγ. Hence SX is an inverse semigroup and we are done.
7
Remark 3.4. For sαE(F ; v)s∗
β ∈ SX, if it happens that αα = ββ = a ∈ a, then
sαE(F ; v)s∗
β = sα1···αα−1E(F ; av)s∗
β1···βα−1
.
For this reason, when working with elements of SX we will usually assume they are in
"lowest terms". To be more precise, we take the following form for SX:
SX = {sαE(F ; v)s∗
β α, β, v ∈ a
∗, F ⊂fin a
∗, αα 6= ββ} ∪ {0}.
(4)
If we take two such elements sαE(F ; v)s∗
then
β and sδE(G; w)s∗
η with αα 6= ββ and δδ 6= ηη,
sαE(F ; v)s∗
β = sδE(G; w)s∗
η ⇒ α = δ, β = η.
We caution that the above equality does not imply that E(F ; v) = E(G; w), Indeed, a short
calculation shows that
sαE(F ; v)s∗
β = sαE(F ∪ {αv, βv}; v)s∗
β.
We could put a condition on SX similar to (4) stating that we assume F contains αv and
βv when writing sαE(F ; v)s∗
β, but this will usually not be necessary.
In the proof of Proposition 3.3 we started computation of the product of two elements
of SX, but stopped when it became clear that the product was again back in SX. In the
following lemma, we record the details of the exact form of this product.
Lemma 3.5. Let X be a one-sided subshift over a, and take α, β, δ, η, v, w ∈ a∗ and F, G ⊂fin
a∗.
1. If δ = βγ and γ = vz for some γ, z ∈ a∗, then
(sαE(F ; v)s∗
β)(sδE(G; w)s∗
η) = sαγE(F zw ∪ G ∪ {γw} ∪ {δw}; w)s∗
η
2. If δ = βγ, v = γz, and z = wr for some γ, z, r ∈ a∗, then
(sαE(F ; v)s∗
β)(sδE(G; w)s∗
η) = sαγE(F ∪ Gr ∪ {βv}; z)s∗
η
3. If δ = βγ, v = γz, and w = zr for some γ, z, r ∈ a∗, then
(sαE(F ; v)s∗
β)(sδE(G; w)s∗
η) = sαγE(F r ∪ G ∪ {βvr}; w)s∗
η
4. If β = δγ and γ = wz for some γ, z ∈ a∗ then
(sαE(F ; v)s∗
β)(sδE(G; w)s∗
η) = sαE(F ∪ Gzv ∪ {γv} ∪ {βv}; v)s∗
ηγ
5. If β = δγ, w = γz, and z = vr for some γ, z, r ∈ a∗, then
(sαE(F ; v)s∗
β)(sδE(G; w)s∗
η) = sαE(F r ∪ G ∪ {δw}; z)s∗
ηγ
(5)
(6)
(7)
(8)
(9)
8
6. If β = δγ, w = γz, and v = zr for some γ, z, r ∈ a∗, then
(sαE(F ; v)s∗
β)(sδE(G; w)s∗
η) = sαE(F ∪ Gr ∪ {δwr}; v)s∗
ηγ
(10)
7. If none of the conditions in 1 -- 6 above hold, then (sαE(F ; v)s∗
β)(sδE(G; w)s∗
η) = 0.
Proof. This follows from Lemma 3.2, and is left to the enthusiastic reader.
Lemma 3.6. Let X be a one-sided subshift over a, let α, β, v ∈ a∗ and F ⊂fin a∗. Then
(sαE(F ; v)s∗
β)(sαE(F ; v)s∗
β)∗ = E(F ∪ {βv}; αv)
(sαE(F ; v)s∗
β)∗(sαE(F ; v)s∗
β) = E(F ∪ {αv}; βv).
Proof. This follows from Lemma 3.5.
Proposition 3.7. Let X be a one-sided subshift over a, let SX be as in (4), and let EX be
as in (1). Then E(SX) = EX.
Proof. This follows from Lemma 3.6 together with the fact that the set of idempotents of
an inverse semigroup S coincides with the set of elements of the form s∗s for s ∈ S.
Remark 3.8. A recent preprint of Boava, de Castro, and Mortari [BdM15] associates an
inverse semigroup to every labeled space. In [BCP12] Bates, Carlsen, and Pask associate
a labeled space to any one-sided subshift X, such that the C*-algebra associated to the
constructed labeled space is isomorphic to OX, see [BCP12, Example 4]. We caution that the
inverse semigroup that one obtains by combining [BdM15] and [BCP12] (say, SX) will not
be the same as our SX -- in fact E( SX) will be isomorphic to the Boolean algebra generated
by the C(v, w) as v and w range over a∗. Hence, their set of idempotents will contain
complements of the C(v, w) while ours (in general) will not. For the specific situation of
a subshift X our construction seems natural, as the only idempotents which appear in our
construction are those which arise directly from the partial bijections arising from the shift
map on X.
3.2 Properties of SX
We now discuss some useful properties which our newly-defined inverse semigroup SX may
possess.
Definition 3.9. Let S be an inverse semigroup with identity and zero (in other words, an
inverse monoid with zero).
1. We say that S is E∗-unitary if 0 6= e 6 s with e ∈ E(S) implies that s ∈ E(S).
2. If Γ is a group and φ : S \ {0} → Γ such that s, t ∈ S with st 6= 0 implies that
φ(st) = φ(s)φ(t), then we say that φ is a partial homomorphism from S to Γ. If, in
addition, φ−1(1Γ) = E(S), then φ is called an idempotent pure partial homomorphism.
9
3. We say that S is strongly E∗-unitary if there exists a group Γ and an idempotent-pure
partial homomorphism from S to Γ.
4. We say that S is F*-inverse if for each s ∈ S there exists a unique maximal element
above s.
5. We say that S is strongly F ∗-inverse if there exists a group Γ and an idempotent-pure
partial homomorphism φ from S to Γ such that for all g ∈ Γ, φ−1(g) has a maximal
element whenever it is nonempty.
Lemma 3.10. Let X be a one-sided subshift over a, and let SX be as in (4). Then SX is
E∗-unitary.
Proof. We note that for an idempotent e, e 6 s if and only if se = e. Suppose that
α, β, v, w ∈ a∗, that F, G ⊂fin a∗, that αα 6= ββ, and that E(G; w), sαE(F ; v)s∗
β 6= 0. We
have
(sαE(F ; v)s∗
β)E(G; w) = sαE(F ; v)(s∗
= sαE(F ′; v′)s∗
β
βE(G; w)sβ)s∗
β
for some F ′ ⊂fin a
∗, v′ ∈ a
∗
If this is equal to E(G; w), we must have that α = β. Since the last letters of α, β were
assumed to be unequal, this implies that α = β = ǫ, and hence sαE(F ; v)s∗
β is an idempotent
as required.
We now prove that SX is strongly E∗-unitary, which seems to make the above lemma
a waste because evidently being strongly E∗-unitary implies being E∗-unitary. Still, we
believe that the above lemma is instructive, so there it stays.
Lemma 3.11. Let X be a one-sided subshift over a, and let SX be as in (4). Then SX is
strongly E∗-unitary.
Proof. Let Fa denote the free group on the alphabet a. For α, β, v ∈ a∗, F ⊂fin a∗ such that
αα 6= ββ, we define a map φ : SX \ {0} → Fa by
φ(sαE(F ; v)s∗
β) = αβ−1.
We claim that this map is a partial homomorphism. To prove this, we take α, β, δ, η, v, w ∈
a∗ and F, G ⊂fin a∗ such that αα 6= ββ, δδ 6= ηη and suppose that (sαE(F ; v)s∗
0. This implies that either δ = βγ or β = δγ for some γ ∈ a∗.
β)(sδE(G; w)s∗
η) 6=
If δ = βγ for some γ ∈ a∗, then
φ(sαE(F ; v)s∗
β)φ(sδE(G; w)s∗
η) = αβ−1δη−1 = αβ−1βγη−1 = αγη−1.
On the other hand, in each of the first three cases of Lemma 3.5, the product of these two
elements is sαγAs∗
β for some A ∈ EX. Hence φ(sαE(F ; v)s∗
η) = αγη−1. The
case β = δγ is similar. Hence φ is a partial homomorphism.
βsδE(G; w)s∗
Furthermore, if φ(sαE(F ; v)s∗
β) = αβ−1 = 1Fa, then α = β, and as before this implies
that sαE(F ; v)s∗
β = E(F ; v), an idempotent. Thus, φ is idempotent pure.
10
Finally, we consider the last two properties from Definition 3.9.
Lemma 3.12. Suppose that α, β, v ∈ a∗, that F ⊂fin a∗, and that αα
sαE(F ; v)s∗
β. Furthermore, if s ∈ SX and sαE(F ; v)s∗
β 6 s, then s 6 sαs∗
β.
β 6 sαs∗
6= ββ. Then
Proof. Let t = sαE(F ; v)s∗
E(F ∪ {αv}; βv). We calculate
β. We first must show that sαs∗
βt∗t = t. By Lemma 3.6, t∗t =
sαs∗
βt∗t = sαs∗
= sαs∗
= sαE(F ∪ {αv}; v)s∗
= sαE(F ∪ {αv}; v)s∗
β
βE(F ∪ {αv}; βv)
βsβE(F ∪ {αv}; v)s∗
βsβs∗
β
β
= sαsvs∗
vs∗
= sαsv Yf ∈F
= sαE(F ; v)s∗
β
f sf! s∗
s∗
vs∗
β
αsαsv Yf ∈F
f sf! s∗
s∗
vs∗
β
Now, take δ, η, ∈ a∗ with δδ 6= ηη, A ∈ EX and let s = sδAs∗
η. Then
st∗t = sδAs∗
ηE(F ∪ {αv}; βv) = sδAs∗
ηE(F ∪ {αv}; βv)sηs∗
η = sδBs∗
η
for some B ∈ EX by Lemma 3.2. If this is equal to s, then δ = α and η = β. Thus by the
above calculation, we must have s 6 sαs∗
β as well.
We can now prove the following.
Proposition 3.13. Let X be a one-sided subshift over a and let SX be as in 4. Then SX is
strongly F ∗-inverse.
Proof. Let φ be as defined in the proof of Lemma 3.11. Then if φ−1(g) is not empty,
g = αβ−1 for some α, β ∈ a∗, and as in the proof of Lemma 3.12, sαs∗
β is maximal in
φ−1(αβ−1).
4 C*-algebras
We now turn our attention to the C*-algebras associated to the structures we have defined.
The main result of this section, Theorem 4.8, states that given a one-sided subshift X, a
certain C*-algebra OX associated to X is canonically isomorphic to a certain C*-algebra
C ∗
tight(SX) associated to the inverse semigroup SX. We first recall the construction of OX
due to Matsumoto and Carlsen, and then the construction of C ∗
tight(S) for a general inverse
semigroup S. Knowledge of C*-algebras is assumed -- one can find undefined terms in the
excellent reference [Dav96].
11
4.1 The Carlsen-Matsumoto algebra OX
Let X be a one-sided subshift over a, and consider ℓ∞(X), the C*-algebra of bounded
functions on X. Define eDX to be the C*-subalgebra of ℓ∞(X) generated by {1C(µ,ν) µ, ν ∈
a∗}. We can now define the algebra OX.
Definition 4.1. (See [CS07, Theorem 10]) Let X be a one-sided subshift over a. Then the
Carlsen-Matsumoto algebra OX is the universal C*-algebra generated by partial isometries
{Sµ}µ∈a∗ such that
1. SµSν = Sµν for all µ, ν ∈ a∗, and
2. The map 1C(µ,ν) 7→ SνS∗
µSµS∗
algebra generated by {Sµ µ ∈ a∗}.
ν extends to a ∗-homomorphism from eDX to the C*-
So OX is generated by a set of partial isometries {Sµ}µ∈a∗, and we view eDX as the
ν . One can show that OX is
= Sǫ. Furthermore, one can show that the elements {Sµ}µ∈a∗
subalgebra of OX generated by elements of the form SνS∗
unital, with unit IOX = I eDX
satisfy
µSµS∗
SaS∗
a = IOX,
Xa∈a
S∗
µSµSνS∗
S∗
µSµS∗
ν Sν = S∗
ν = SνS∗
ν S∗
ν SνS∗
µSµ,
µSµ.
(11)
(12)
(13)
(14)
In addition, if µ, ν ∈ a∗ with µ = ν, then
S∗
µSν 6= 0 ⇒ µ = ν.
Since eDX is a commutative C*-algebra, it is isomorphic to C(eX) for a certain compact
Hausdorff spaceeX. This space was presented as an inverse limit space in [Car04, Chapter 2],
and we reproduce this presentation here because we will use it to establish an isomorphism
between C ∗
tight(SX) and OX.
For x ∈ X and integer k ≥ 0, let
Pk(x) = {µ ∈ a
∗ µx ∈ X, µ = k}
For l ∈ N, we say that x, y ∈ X are l-past equivalent and write x ∼l y if Pk(x) = Pk(y) for
all k ≤ l. The l-past equivalence class of x ∈ X will be written as [x]l.
Let I = {(k, l) ∈ N2 k ≤ l}. For every (k, l) ∈ I we define another equivalence relation
k∼l on X by
xk∼ly ⇔ x[1,k] = y[1,k] and Pr(x(k,∞)) = Pr(y(k,∞)) for all r ≤ l.
We note that there is a typo in [Car04, Chapter 2] where the above is defined with r = l
rather than r ≤ l.3 The equivalence class of x ∈ X under k∼l will be written as k[x]l, and
3This was confirmed in private communication with Carlsen.
12
the set of all such equivalence classes will be written as kXl. It is clear that for all (k, l) ∈ I,
kXl is finite; we endow it with the discrete topology.
There is a partial order on I which respects this equivalence relation. For (k1, l1), (k2, l2) ∈
I we say
(k1, l1) ≤ (k2, l2) ⇔ k1 ≤ k2 and l1 − k1 ≤ l2 − k2.
We note that if (k, l), (r, s) ∈ I, then they have a common upper bound. Indeed, if k = r
then (k, max{l, s}) is an upper bound for (k, l) and (r, s), and if k < r then (r, max{l + r −
k, s}) is an upper bound for (k, l) and (r, s). If (k1, l1) ≤ (k2, l2) then it is straightforward
that
xk2∼l2y ⇒ xk1∼l1y.
Thus, for (k1, l1) ≤ (k2, l2), there is a map (k1,l1)π(k2,l2) : k2
Xl2 → k1
Xl1 such that
One can then form the inverse limit
(k1,l1)π(k2,l2)(k2[x]l2) = k1[x]l1
( kXl, π)
(k,j)∈I
eX = lim
=
(k[kxl]l)(k,l)∈I ∈ Y(k,l)∈I
kXl (k1, l1) ≤ (k2, l2) ⇒ k1[k2xl2]l1 = k1[k1xl1]l1
Which is a closed subspace of the space Q(k,l)∈I kXl when given the product topology of
the discrete topologies.
(15)
Let x ∈ X and take (k, l) ∈ I. The set
U(x, k, l) = {(r[rxs]s)(r,s)∈I ∈eX k[kxl]l = k[x]l}
is open and closed. Sets of this form generate the topology on eX.
We have the following lemma about the relation k∼l.
Lemma 4.2. Let X be a subshift, let v ∈ a∗, let F ⊂fin a∗, and let
k = v,
l = max{f , v : f ∈ F }.
Then for all x ∈ X, and all (r, s) ≥ (k, l), either r[x]s ⊂ C(F ; v) or r[x]s ∩ C(F ; v) = ∅.
Proof. Since (r, s) ≥ (k, l) implies that r[x]s ⊂ k[x]l, we need only prove the statement for
r = k and s = l. Suppose that we have k[x]l 6⊂ C(F ; v), and take y ∈ k[x]l \ C(F ; v). If v
is not a prefix of y, then this is true of all elements of k[x]l and we are done. So, suppose
that y = vy′. There must be an element f ∈ F such that f y′ /∈ X. If we have some other
element z = vz′ ∈ k[x]l, we must have that Pf (z′) = Pf (y′), and so f z′ /∈ X. This implies
that z /∈ C(F ; v) and we are done.
13
4.2 OX as the tight C*-algebra of SX
In this section we recall the definition of the tight C*-algebra of an inverse semigroup from
[Exe08]. We then show that the tight C*-algebra of SX is isomorphic to OX.
Let S be an inverse semigroup with 0, and let A be a C*-algebra. A map π : S → A is
called a representation of S if π(0) = 0, π(st) = π(s)π(t) and π(s)∗ = π(s∗) for all s, t ∈ S.
We are interested in a certain class of representations which we will now describe. For
F ⊂ Z ⊂ E(S), we say that F covers Z if for every z ∈ Z, there exists f ∈ F such that
f z 6= 0. If F covers {y ∈ E(S) y 6 x}, we say that F covers x.
Let X, Y ⊂fin E(S), and let
E(S)X,Y = {e ∈ E(S) e 6 x for all x ∈ X, ey = 0 for all y ∈ Y }.
A representation π : S → A with A unital is said to be tight if whenever X, Y, Z ⊂fin E(S)
such that Z is a cover of E(S)X,Y , then
_z∈Z
π(z) = Yx∈X
π(x)Yy∈Y
(1 − π(y)).
The tight C*-algebra of S, denoted C ∗
tight(S), is the universal C*-algebra generated by
one element for each element of S subject to the relations which say that the standard map
πt : S → C ∗
tight(S) is a tight representation.
At this point it is not clear that C ∗
tight(S) exists, but it was explicitly constructed in
associated to S. We do not go into specifics about inverse semigroup actions or groupoids
[Exe08] as a groupoid C*-algebra associated to an action of S on a certain space bEtight(S)
here, though we will define bEtight(S) as it is essential for establishing isomorphism we desire.
Recall that the natural partial order on S, when restricted to E(S), takes on a simpler
form: e 6 f ⇔ ef = e. A subset ξ ⊂ E(S) is called a filter if it does not contain the zero
element, is closed under products, and is "upwards directed", which is to say that if e ∈ ξ
and e 6 f then f ∈ ξ. A filter is called an ultrafilter if it is not properly contained in any
other filter. The set of filters is denoted bE0(S), and the set of ultrafilters is denoted bE∞(S).
The set bE0(S) may be viewed as a subset of the product space {0, 1}E(S). We let bE0(S)
inherit the subspace topology from the product topology (with {0, 1} given the discrete
topology). For e ∈ E(S), let
De = {ξ ∈ bE0(S) e ∈ ξ}.
Then sets of this form together with their complements form a subbasis for the topology on
It is a fact that C ∗
bE0(S). With this topology, bE0(S) is called the spectrum of E(S). We also let bEtight(S) =
bE∞(S), and call this the tight spectrum E(S). We will shorten De ∩ bEtight(S) to Dt
πt(E(S)) is ∗-isomorphic to C(bEtight) via the identification πt(e) 7→ 1Dt
Now, we take X to be a one-sided subshift over a, and describe bEtight(SX).
tight(S) exists and that the C*-subalgebra of C ∗
tight(S) generated by
e.
Lemma 4.3. If ξ ⊂ EX is a filter, then there exists x ∈ X ∪ a∗ such that if E(F ; v) ∈ ξ,
then v is a prefix of x.
e.
14
Proof. If ξ is a filter and E(G; v), E(H, w) ∈ ξ, then Lemma 3.2 shows that their product
is zero unless w is a prefix of v or vice-versa. The result follows.
Lemma 4.4. Let X be a one-sided subshift over a, and let
ηx = {E(F ; v) ∈ EX x ∈ C(F ; v)}.
Then bE∞ = {ηx x ∈ X}.
Proof. First, we show that ηx is a filter. If E(F ; v), E(G, w) ∈ ηx, then C(F ; v)∩C(G; w) =
C(H; z) 6= ∅ for some H ⊂fin a∗ and z ∈ a∗. Hence E(F ; v)E(G; w) = E(H; z), and so ηx is
closed under products. It is clear that ηx does not contain the zero element and is upwards
closed, so it is a filter.
Now, suppose that we have E(F ; v) such that E(F ; v)E(G; w) 6= 0 for all E(G; w) ∈ ηx.
Thus for each n ≥ 0, we can find yn ∈ C(F ; v) ∩ C(x1 · · · xn), and it is clear that the yn
converge to x in X. Since C(F ; v) is closed in X, we must have that x ∈ C(F ; v), and so
E(F ; v) ∈ ηx. This shows that ηx is an ultrafilter.
Now, suppose that ξ ⊂ EX is an ultrafilter. Then {C(F ; v) E(F ; v) ∈ ξ} is a collection
of closed subsets of the compact space X which has the finite intersection property, and so
the intersection
\E(F ;v)∈ξ
C(F ; v)
is nonempty. Take x in the above intersection. Then we must have that ξ ⊂ ηx, and since
ξ is assumed to be an ultrafilter, ξ = ηx.
We now return to the space eX from (15) which is the spectrum of the commutative
C*-algebra eDX. Our next proposition will establish a natural homeomorphism between eX
and bEtight(SX).
Proposition 4.5. The map θ :eX → bE0(SX) defined by
is continuous, injective, and θ(eX) = bEtight(SX). Hence, it is a homeomorphism from eX to
bEtight(SX).
Proof. First we show that θ is well-defined. Take (k[kxl]l)(k,l)∈I ∈eX and consider its image
θ(cid:0)(k[kxl]l)(k,l)∈I(cid:1) = {E(F ; v) ∈ EX k[kxl]l ⊂ C(F ; v) for some (k, l) ∈ I}
under θ -- it is clearly upwards closed and does not contain the zero element. To prove
closure under products, suppose we have (r, s), (t, u) ∈ I and E(F ; v), E(G; w) ∈ EX such
that r[rxs]s ⊂ E(F ; v) and t[txu]u ⊂ E(G; w). Let (k, l) be an upper bound for (r, s), (t, u)
in I. Then k[kxl]l is a subset of both r[rxs]s and t[txu]u, and so is contained in both C(F ; v)
and C(G; w). If E(F ; v)E(G; w) = E(H; z), then k[kxl]l ⊂ C(H; z), and so θ((k[kxl]l)(k,l)∈I)
is a filter.
must exist (k, l) ∈ I such that k[kxl]l 6=k [kyl]l. If (kxl)[1,k] 6= (kyl)[1,k], then
We now show that θ is injective. Suppose we have x, y ∈eX and that x 6= y. Then there
(16)
E [r≤l
Pr((kxl)(k,∞)); (kxl)[1,k]! ∈ θ(x)
15
E [r≤l
Pr((kyl)(k,∞)); (kyl)[1,k]! ∈ θ(y).
The product of these two elements is zero, so θ(x) 6= θ(y).
So, we instead suppose that there exists v ∈ a∗ with v = k and kxl = vx′, kyl = vy′, so
that
k[vx′]l 6= k[vy′]l.
Without loss of generality, there must exist w ∈ a∗ with w ≤ l such that wx′ ∈ X and
wy′ /∈ X. Hence, vx′ ∈ C(w, v), and vy′ /∈ C(w, v). Thus by Lemma 4.2, we must have that
k[vx′]l ⊂ C(w, v) and k[vy′]l∩C(w, v) = ∅. Similar to above, this implies that E(w, v) ∈ θ(x)
and E(w, v) /∈ θ(y). Hence θ(x) 6= θ(y), and θ is injective.
Now, we prove that θ is continuous. Take E(F ; v) ∈ EX, and as before take DE(F ;v) =
{ξ ∈ bE0(SX) E(F ; v) ∈ ξ}. Then
θ−1(DE(F ;v)) = {(k[kxl]l)(k,l)∈I ∈eX r[rxs]s ⊂ C(F ; v) for some (r, s) ∈ I}.
If (k[kxl]l)(k,l)∈I ∈ θ−1(DE(F ;v)), find (r, s) ∈ I such that r[rxs]s ⊂ C(F ; v). Then if
(k[kyl]l)(k,l)∈I ∈ U(rxs, r, s), r[rys]s = r[rxs]s ⊂ C(F ; v), and so U(rxs, r, s) ⊂ θ−1(DE(F ;v)).
On the other hand,
θ−1((DE(F ;v))c) = {(k[kxl]l)(k,l)∈I ∈eX r[rxs]s 6⊂ C(F ; v) for all (r, s) ∈ I}.
Take (k[kxl]l)(k,l)∈I ∈ θ−1((DE(F ;v))c), take k = v, and take l = max{f , v : f ∈ F }. Then
k[kxl]l ∩ C(F ; v) = ∅, and so U(kxl, k, l) ⊂ θ−1((DE(F ;v))c). The collection of all sets of the
form DE(F ;v) together with those of the form (DE(G;w))c form a subbasis for the topology
on bE0, so θ is continuous.
Finally, we must show that θ(eX) = bEtight(SX). For x ∈ X, let
x = (k[x]l)(k,l)∈I.
Because sets of the form U(x, k, l) for x ∈ X and (k, l) ∈ I form a basis for the topology on
We claim that θ(x) = ηx. If E(F ; v) ∈ ηx, then taking k = v and l = max{f , v : f ∈
F } gives us that k[x]l ⊂ C(F ; v), and so E(F ; v) ∈ θ(x). Conversely, if E(F ; v) ∈ θ(x), then
k[x]l ⊂ C(F ; v) for some (k, l) ∈ I. Hence x ∈ C(F ; v), E(F ; v) ∈ ηx, and so θ(x) = ηx.
eX, the set {x ∈eX x ∈ X} is dense in eX.
So θ :eX → bE0(SX) is continuous, injective, and maps a dense subspace of eX bijectively
onto a dense subspace of bEtight(SX). BotheX and bE0(SX) are second countable, and so θ(eX) ⊂
bEtight(SX). Since eX is compact, we must have that θ(eX) is a closed set in bE0(SX) which
contains bE∞(SX), and so it contains its closure bEtight(SX). Therefore θ(eX) = bEtight(SX), and
sinceeX is compact and bEtight(SX) is Hausdorff, θ :eX → bEtight(SX) is a homeomorphism.
Proposition 4.6. There exists a ∗-isomorphism Ψ : eDX → C(bEtight(SX)) such that Ψ(1C(w,v)) =
Now that we have the above homeomorphism, we can establish the conditions we need
for all w, v ∈ a∗. Furthermore, if E(F ; v) ∈ EX, then Ψ(1C(F ;v)) = 1Dt
to use the universal property of OX.
1Dt
E(w,v)
.
E(F ;v)
16
Proof. Take w, v ∈ a∗, and let k = v, l = max{w, v}. There are only finitely many
l-past equivalence classes, so pick a representative from each one, say xl
m(l). By
Lemma 4.2, C(w, v) is a finite disjoint union of k∼l equivalence classes, that is there exists
F ⊂ {1, . . . , m(l)} such that
2, . . . xl
1, xl
C(w, v) = [f ∈F
= [f ∈F
k[vxl
f ]l
C(v) ∩ σ−k([xl
f ]l).
(17)
From the proof of Proposition 4.5, if θ is as in (16), we must have that
θ−1(Dt
E(w,v)) = [f ∈F
U(vxl
f , k, l)
where again this is a disjoint union. Thus, if Θ is the ∗-isomorphism from C(bEtight(SX)) to
C(eX) induced by θ, we have
1U (vxl
f ,k,l).
Θ(1DE(w,v)) =Xf ∈F
By [Car04, Proposition 3 in Chapter 2], there exists a ∗-isomorphism ψ : eDX → C(eX) such
f ,v,l). By (17) we have
that ψ(1C(v)∩σ−v([xl
f ]l)) = 1U (vxl
Θ−1 ◦ ψ(cid:0)1C(w,v)(cid:1) = Θ−1 ◦ ψ(cid:16)1Sf ∈F C(v)∩σ−k ([xl
f ]l)(cid:17)
f ]l)(cid:17)!
ψ(cid:16)1C(v)∩σ−k ([xl
f ,k,l)!
= Θ−1 Xf ∈F
= Θ−1 Xf ∈F
1U (vxl
= Θ−1(Θ(1Dt
= 1Dt
.
E(w,v)
))
E(w,v)
Hence taking Ψ = Θ−1 ◦ ψ verifies the first statement. The second statement follows from
the fact that, for all F ⊂fin a∗ and v ∈ a∗, we have
1C(F ;v) = 1∩f ∈F C(f,v) =Yf ∈F
E(F ;v) = DtQf ∈F E(f,v) = \f ∈F
Dt
1C(f,v),
Dt
E(f,v).
We now establish what we need to use the universal property of C ∗
tight(SX)
17
Proposition 4.7. Let X be a one-sided subshift over a. Then the map π : SX → OX defined
by
π(sαE(F ; v)s∗
v S∗
β,
F ⊂fin a
∗; α, β, v ∈ a
∗
β) = SαSv Yf ∈F
f Sf! S∗
S∗
π(0) = 0
is a tight representation of SX.
Proof. Because Definition 4.1.1 and the relations (12), (13), (14) hold in OX, and each
Sµ is a partial isometry, the same computations from Lemma 3.2 hold in OX. Hence, the
products computed in Lemma 3.5 hold in OX, and so π is a representation of SX.
Now suppose we have X, Y, Z ⊂fin EX such that Z is a cover of EX,Y
X
. Then we know
that, for the universal tight representation πt, we have
By Proposition 4.6, πt(e) = Ψ ◦ π(e) for all e ∈ EX. Thus we have
(1 − πt(y)).
(1 − πt(y))
πt(z) = Yx∈X
_z∈Z
πt(x)Yy∈Y
_z∈Z
πt(z) = Yx∈X
πt(x)Yy∈Y
_z∈Z
Ψ ◦ π(z) = Yx∈X
Ψ ◦ π(x)Yy∈Y
Ψ _z∈Z
π(z)! = Ψ Yx∈X
π(x)Yy∈Y
_z∈Z
π(z) = Yx∈X
π(x)Yy∈Y
(Ψ ◦ π(1) − Ψ ◦ π(y))
(π(1) − π(y))!
(IOX − π(y))
and so, π is a tight representation.
Theorem 4.8. Let X be a one-sided subshift over a, let SX be as in (4), and let OX be as
in Definition 4.1. Then C ∗
tight(SX) and OX are ∗-isomorphic.
Proof. By Proposition 4.6 and the universal property of OX, there exists a ∗-homomorphism
tight(SX) such that κ(Sµ) = πt(sµ) for all µ ∈ a∗. By Proposition 4.7 and the fact
κ : OX → C ∗
that C ∗
tight(SX) is universal for tight representations for SX, there exists a ∗-homomorphism
τ : C ∗
tight(SX) → OX such that τ (πt(sµ)) = Sµ for all µ ∈ a∗. We therefore must have that
κ and τ are inverses of each other, and so C ∗
tight(SX) and OX are ∗-isomorphic.
4.3 OX as a partial crossed product
We close with a nice consequence of Theorem 4.8. Recall from Lemma 3.11 that SX is
strongly E∗-unitary. Any strongly E∗-unitary inverse semigroup S admits a universal group
U(S), that is there exists an idempotent-pure partial homomorphism ι : S \ {0} → U(S)
such that if every other idempotent-pure partial homomorphism from S factors through ι.
We have the following result about strongly E∗-unitary inverse semigroups from [MS14].
18
Theorem 4.9. (See [MS14, Theorem 5.3]) Let S be a countable strongly E∗-unitary inverse
semigroup. Then there is a natural partial action of U(S) on bEtight(S) such that the partial
crossed product C(bEtight(S)) ⋊ U(S) is isomorphic to C ∗
We do not define partial actions or partial crossed products here -- the interested reader
tight(S).
is directed to the excellent reference [Exe14].
In a preprint version of this work, we concluded the paper by using the above to deduce
that OX could be written as a partial crossed product by the universal group of SX. We
are grateful to the referee for pointing out that our results allow us to easily see what the
universal group is and to say even more about this partial crossed product. In what remains
of this paper, we implement the referee's suggestions.
The following Lemma is a consequence of our proof of Lemma 3.11.
Lemma 4.10. Let X be a one-sided subshift over a. Then U(SX) is isomorphic to Fa.
Proof. Let φ : SX :→ Fa be the partial homomorphism from the proof of Lemma 3.11. The
group U(SX) is generated by ι(sa) for a ∈ a, so there is a group homomorphism from Fa
to U(SX) which sends a ∈ a to ι(sa). From the definition of U(SX), there exists a group
homomorphism from U(SX) to Fa such that ι(sa) = a. Therefore, U(SX) is isomorphic to
Fa.
We now have the following from Theorem 4.9
Corollary 4.11. Let X be a one-sided subshift over a, and let OX be as in Definition 4.1.
Then there is partial action of Fa on X such that OX ∼= C(X) ⋊ Fa.
At this point we must direct the reader to the recent preprint [ED15] which constructs
by hand the partial action from Corollary 4.11, studies it in detail, and uses it to give
necessary and sufficient conditions on X to guarantee that OX is simple. The article [ED15]
appeared after the first preprint version of this work but before the final version was ac-
cepted. Therefore, the result in Corollary 4.11 is original to [ED15].
As the referee points out, one can say a little more about this partial crossed product.
Given a partial action θ of a group Γ on a space X, one can always construct a space
X ⊃ X and a global action θ of Γ on X such that the restriction of θ to X is the original
partial action -- this is called the enveloping action for θ, see [Aba03]. Unfortunately, even
if X is Hausdorff, X may not be. When X and X are both locally compact and Hausdorff,
then the partial crossed product C0(X) ⋊θ Γ is strongly Morita equivalent to the crossed
product C0( X) ⋊θ Γ, see [Aba03] for the details.
In our situation, [MS14, Corollary 6.17] says that because SX is F ∗-inverse, the space
for the enveloping action for the partial action in [ED15] and Corollary 4.11 is Hausdorff.
Therefore we have the following.
Corollary 4.12. Let X be a one-sided subshift over a, and let OX be as in Definition 4.1.
Then there exists a locally compact Hausdorff space Ω and an action of Fa on Ω such that
OX is strongly Morita equivalent to C0(Ω) ⋊ Fa.
Acknowledgment: I am grateful to the referee for an extremely careful reading, and
for pointing out that the results of this paper could be strengthened to include Lemma
4.10, Corollary 4.11, and Corollary 4.12.
19
References
[Aba03]
Fernando Abadie. Enveloping actions and takai duality for partial actions. Jour-
nal of Functional Analysis, 197(1):14 -- 67, 2003.
[BCFS14] Jonathan Brown, Lisa Orloff Clark, Cynthia Farthing, and Aidan Sims. Simplic-
ity of algebras associated to ´etale groupoids. Semigroup Forum, 88(2):433 -- 452,
2014.
[BCP12] Teresa Bates, Toke Meier Carlsen, and David Pask. C*-algebras of labelled
graphs III - K-theory computations. arXiv:1203.3072. To appear in Ergodic
Theory Dynam. Systems, 2012.
[BdM15] Giuliano Boava, Gilles G. de Castro, and Fernando de L. Mortari. Inverse semi-
groups associated with labelled spaces and their tight spectra. arXiv:1505.07123,
2015.
[Car04]
[Car08]
[CK80]
[CM04]
[CS07]
[CT12]
Toke Meier Carlsen. Operator algebraic applications in symbolic dynamics. PhD
thesis, University of Copenhagen, Copenhagen, Denmark, 2004.
Toke Meier Carlsen. Cuntz-Pimsner C*-algebras associated with subshifts. In-
ternational Journal of Mathematics, 19(01):47 -- 70, 2008.
Joachim Cuntz and Wolfgang Krieger. A class of C*-algebras and topological
Markov chains. Inventiones mathematicae, 56(3):251 -- 268, 1980.
Toke Meier Carlsen and Kengo Matsumoto. Some remarks on the C*-algebras
associated with subshifts. MATHEMATICA SCANDINAVICA, 95(1):145 -- 160,
2004.
Toke Meier Carlsen and Sergei Silvestrov. C*-crossed products and shift spaces.
Expositiones Mathematicae, 25(4):275 -- 307, 2007.
Toke Meier Carlsen and Klaus Thomsen. The structure of the C*-algebra of a
locally injective surjection. Ergodic Theory Dynam. Systems, 103(3):1226 -- 1248,
2012.
[Dav96] Kenneth R. Davidson. C ∗-algebras by example, volume 6 of Fields Institute
Monographs. American Mathematical Society, Providence, RI, 1996.
[ED15]
Ruy Exel and Mikhailo Dokuchaev.
arXiv:1511.00939, November 2015.
Partial actions and subshifts.
[EGS12] Ruy Exel, Daniel Gon¸calves, and Charles Starling. The tiling C*-algebra viewed
as a tight inverse semigroup algebra. Semigroup Forum, 84:229 -- 240, 2012.
[EP14]
Ruy Exel and Enrique Pardo. The tight groupoid of an inverse semigroup.
arXiv:1408.5278, August 2014.
20
[Exe08]
[Exe14]
[Kel97]
[LJ14]
[LL13]
[LS14]
Ruy Exel. Inverse semigroups and combinatorial C ∗-algebras. Bull. Braz. Math.
Soc. (N.S.), 39(2):191 -- 313, 2008.
Ruy Exel. Partial Dynamical Systems Fell Bundles and Applications. to appear
in NYJM Book Series, 2014.
Johannes Kellendonk. The local structure of tilings and their integer group of
coinvariants. Comm. Math. Phys., 187(1):115 -- 157, 1997.
Mark V. Lawson and David Jones. Graph inverse semigroups: their characteri-
zation and completion. J. Algebra, 409:444 -- 473, 2014.
Mark V. Lawson and Daniel H. Lenz. Pseudogroups and their ´etale groupoids.
Advances in Mathematics, 244(0):117 -- 170, 2013.
Mark V. Lawson and Phil Scott. AF inverse monoids and the structure of
countable MV-algebras. arXiv:1408.1231, 2014.
[Mat97] Kengo Matsumoto. On C*-algebras associated with subshifts.
International
Journal of Mathematics, 8:357 -- 374, 1997.
[MS14]
[Pat99]
[Pat02]
[Sta15]
[Ste14]
David Milan and Benjamin Steinberg. On inverse semigroup C*-algebras and
crossed products. Groups, Geomety, and Dynamics, 8(2):485 -- 512, 2014.
Alan Paterson. Groupoids, inverse semigroups, and their operator algebras.
Birkhauser, 1999.
Alan Paterson. Graph inverse semigroups, groupoids and their C*-algebras. J.
Operator Theory, 48:645 -- 662, 2002.
Charles Starling. Boundary quotients of C*-algebras of right LCM semigroups.
Journal of Functional Analysis, 268(11):3326 -- 3356, 2015.
Benjamin Steinberg. Simplicity, primitivity and semiprimitivity of ´etale groupoid
algebras with applications to inverse semigroup algebras. arXiv:1408.6014, Au-
gust 2014.
[Tho10] Klaus Thomsen. Semi-´etale groupoids and applications. Annales de linstitut
Fourier, 60(3):759 -- 800, 2010.
University of Ottawa, Department of Mathematics and Statistics. 585 King
Edward, Ottawa, ON, Canada, K1N 6N5 [email protected]
21
|
1204.5836 | 1 | 1204 | 2012-04-26T06:16:01 | Traces on cores of C*-algebras associated with self-similar maps | [
"math.OA",
"math.DS"
] | We completely classify the extreme tracial states onthe cores of the C*-algebras associated with self-similar maps on compact metric spaces. We present a complete list of them. The extreme tracial states are the union of the discrete type tracial states given by measures supported on the finite orbits of the branch points and a continuous type tracial state given by the Hutchinson measure on the original self-similar set. | math.OA | math | TRACES ON CORES OF C∗-ALGEBRAS ASSOCIATED WITH
SELF-SIMILAR MAPS
TSUYOSHI KAJIWARA AND YASUO WATATANI
Abstract. We completely classify the extreme tracial states on the cores of the C∗-algebras
associated with self-similar maps on compact metric spaces. We present a complete list of them.
The extreme tracial states are the union of the discrete type tracial states given by measures
supported on the finite orbits of the branch points and a continuous type tracial state given by
the Hutchinson measure on the original self-similar set.
KEYWORDS: traces, core, self-similar maps, C∗-correspondences
AMS SUBJECT CLASSIFICATION: 46L08, 46L55
1. Introduction
space K is a family of proper contractions γ = (γ1, . . . , γN ) on K such that K =SN
We investigate a relation between self-similar sets and C∗-algebras. Most of self-similar sets
are constructed from iterations of proper contractions. A self-similar map on a compact metric
i=1 γi(K). In
our former work Kajiwara-Watatani [14], we introduced C∗-algebras associated with self-similar
maps on compact metric spaces as Cuntz-Pimsner algebras for certain C∗-correspondences and
show that the associated C∗-algebras are simple and purely infinite. A related study on C∗-
algebras associated with iterated function systems is done by Castro [2]. A generalization to
Mauldin-Williams graphs is given by Ionescu-Watatani [9].
The gauge action of the associated C∗-algebra is an important tool. The fixed point algebra
under the gauge action is called the core. Recall that the dimension group of a topological
Markov shift is exactly the K-theory of the core of the corresponding Cuntz-Krieger algebra
[3]. This idea is extended to the subshifts by Matsumoto [18]. Therefore we expect that many
informations of a self-similar map as a dynamical system are contained in the structure of the
core of the corresponding C∗-algebra. In this note we study the trace structure of the fixed point
algebra under the gauge action. In fact, the trace structure is described by measures supported
on the orbits of the branched points of the self-similar map and the Hutchinson measure. We
present a complete list of the extreme traces on the core to classify them.
One of the key points is the structure of the cores of Cuntz-Pimsner algebras described by
Pimsner [20]. We need a result on the extension of traces on a subalgebra and an ideal to their
sum modified in [15] following after Exel and Laca [4]. We also recall the Rieffel correspondence
of traces between Morita equivalent C∗-algebras, which plays an important role in our study.
2
1
0
2
r
p
A
6
2
]
.
A
O
h
t
a
m
[
1
v
6
3
8
5
.
4
0
2
1
:
v
i
X
r
a
The extreme traces are described by measures supported on the finite orbits of branched
points of a self-similar map. For a branched point b and an integer r ≥ 0, we consider a
family (µ(b,r)
)r
i=0 of discrete measures on K satisfying a compatibility condition such that the
support of µ(b,r)
is on the (r − i)-th γ-orbit of the branch point b for 0 ≤ i ≤ r. By the Rieffel
correspondence of traces, we construct finite traces on the compact algebras of multiple tensor
i
i
1
products of the C∗-correspondence constructed from the original self-similar map. A sequence
of such compatible traces gives an extreme trace τ (b,r) on the core. On the other hand, the
Hutchinson measure on the self-similar set K also gives a continuous type trace τ∞ on the core.
We shall show that these traces τ (b,r) b ∈ Bγ, r = 0, 1, 2, . . . and τ∞ exhaust the extreme traces
on the core F (∞).
For the classification of extreme traces, the analysis of point masses is essential. A tracial
state on the core gives the traces on compact algebras K(X⊗n) by restriction. We get a certain
relation among the point masses of the measures on K corresponding to the traces on K(X⊗n)
and K(X⊗n+1) by Morita equivalence. The difficulty of the analysis comes from the fact that
K(X⊗n) is not included in K(X⊗n+1). Thus the necessary condition can not be described by
simple restriction. Here comes a technical use of a countable basis and a result of the extension
of traces on a subalgebra and an ideal to their sum modified in [15] following after Exel and
Laca [4].
For a tracial state of the core, we remove the discrete parts of it and get a finite trace which
has no point mass. By the argument of principle of proper contraction on tracial state space,
we can show that any tracial state on the core without point mass on A = C(K) is exactly the
trace associated with the Hutchinson measure. Combining these, we can present a complete list
of of extreme traces on the core. This gives a complete classification of the traces.
In [8] we completely classified the KMS states for the gauge action on the associated C∗-
algebra Oγ. We shall show that which trace on the core is extended to a KMS state on Oγ. We
recommend a paper [16] by Kumjian and J. Renalut for a general study of the KMS states on
C∗-algebras associated to expansive maps. In many cases, the inverse braches of an expansive
map gives a self-similar map in our study.
The content of the paper is as follows:
In section 2, we recall some basic notations and
typical examples of self-similar maps. In section 3, we give some preliminary results for C∗-
correspondences. In section 4, we describe Rieffel correspondence of finite traces between C∗-
algebras which are Morita equivalent by an equivalence bimodule of finite degree type using
countable bases. In section 5, we apply a result of trace extension to our situation. We get
a certain relation between point masses of the measures on K corresponding to the traces on
compact algebras of the tensor products of the C∗-correspondences. In section 6, we construct
model traces on the core . We also need a trace extension result for our construction. In section
7, we classify the extreme traces on the core. Since the set of branched points is assumed to be
finite, we can reduce the classificaion of the traces to the uniqueness of the invariant measure
without point mass. In fact, the final step of the classification is completed by the contraction
principle on the metric space of the probability measures using the Lipschitz norm and the fact
that the diameter of it is finite.
2. Self-similar maps
Let (Ω, d) be a (separable) complete metric space. A map f : Ω → Ω is called a proper
contraction if there exists a constant c and c′ with 0 < c′ ≤ c < 1 such that 0 < c′d(x, y) ≤
d(f (x), f (y)) ≤ cd(x, y) for any x, y ∈ Ω.
We consider a family γ = (γ1, . . . , γN ) of N proper contractions on Ω. We assume that N ≥ 2.
Then there exists a unique non-empty compact set K ⊂ Ω which is self-similar in the sense that
K =SN
i=1 γi(K). See Falconer [5] and Kigami [11] for more on fractal sets.
2
In this note we usually forget an ambient space Ω as in [14] and start with the following setting:
Let (K, d) be a compact metric set and γ = (γ1, . . . , γN ) is a family of N proper contractions
Definition 2.1. We say that γ satisfies the open set condition if there exists an open subset V
i=1 γi(V ) ⊂ V . Then V is an open dense
on K. We say that γ is a self-similar map on K if K =SN
of K such that γj(V ) ∩ γk(V ) = φ for j 6= k and SN
Let Σ = { 1, . . . , N}. For k ≥ 1, we put Σk = { 1, . . . , N}k. For a self-similar map γ on a
subset of K. See a book [5] by Falconer, for example.
compact metric space K, we introduce the following subsets of K:
i=1 γi(K).
Bγ ={ b ∈ K b = γi(a) = γj(a), for some a ∈ K and i 6= j, },
Cγ ={ a ∈ K γj(a) ∈ Bγ for some j },
Pγ ={ a ∈ K ∃k ≥ 1, ∃(j1, . . . , jk) ∈ Σk such that γj1 ◦ ··· ◦ γjk (a) ∈ Bγ},
Ob,k ={ γj1 ◦ ··· ◦ γjk (b) (j1, . . . , jk) ∈ Σk }
Orb = [b∈Bγ
(k ≥ 0), Ob =
[k=0
Ob.
∞
Ob,k, where Ob,0 = {b},
We call Bγ the branch set of γ, Cγ the branch value set of γ and Pγ the postcritical set of γ.
We define branch index at (γj(y), y) by eγ(γj(y), y) = #{i ∈ Σγj(y) = γi(y)}. We call Ob,k the
k-th γ orbit of b, and Ob the γ orbit of b.
Assumption A. Throughout the paper, we assume that a self-similar map (K, γ) satisfies
the following for simplicity:
(1) Their exists a continuous map h : K → K such that h(γj(y)) = y for y ∈ K and
j = 1, . . . , N , and h−1(y) =SN
(2) The sets Bγ and Pγ are finite sets.
j=1 γj(y).
Under the assumption A above, Bγ is the branch set of h and Cγ = h(Bγ) is the branch value
(3) The intersection of Bγ and Sb∈Bγ S∞k=1 Ob,k is empty.
of h. Moreover Pγ = S∞k=1 hk(Bγ), Ob,k = h−k(b) and Ob = S∞k=0 h−k(b) the backward orbit of
b under h. The condition (3) of the above assumption A means that for any branch point b,
the backward orbit of b under h does not intersect with the set of branch points except the 0-th
orbit b. The condition (3) is not trivial but satisfied in many cases including a tent map. The
condition (3) can be replaced by the following condition (3)':
(3)' Bγ ∩ Pγ is empty.
In fact, not(3) is equivalent to that there exists y ∈ Bγ ∩ (S∞k=1 h−k(b)). This means that
there exist y, b ∈ Bγ and an integer k ≥ 1 such that hk(y) = b. This is equivalent to not(3)'.
Moreover, since we assume that Pγ is finite, (K, γ) satisfiesthe open set condition. Let V =
K\Pγ. Then V is an open dense subset of K satisfying γj(V ) ∩ γk(V ) = ∅ for j 6= k.
In
fact, on the contrary suppose that γj(V ) ∩ γk(V ) 6= ∅. Then there exist u, v ∈ V such that
b := γj(u) = γk(v). Then h(b) = u = v and u ∈ Cγ ⊂ Pγ, which is a contradiction. We show
that V is invariant under γ. On the contrary suppose that x = γj(u) is not in V for some u ∈ V .
3
Then x is in Pγ and u = h(x) is not in Pγ. Hence x = hk(b) for some b ∈ Bγ and a natural
number k, so u = hk+1(b), which contradicts that u is not in Pγ.
Example 2.1. (tent map) Let K = [0, 1], γ1(y) = (1/2)y and γ2(y) = 1 − (1/2)y. Then a
family γ = (γ1, γ2) of proper contractions is a self-similar map. We note that Bγ = { 1/2},
Cγ = { 1} and Pγ = { 0, 1}. A continuous map h defined by
h(x) =(2x
−2x + 2
0 ≤ x ≤ 1/2
1/2 ≤ x ≤ 1
satisfies Assmption A (1). The map h is the ordinary tent map on [0, 1], and (γ1, γ2) is the
inverse branches of the tent map h. We note that Bγ = { 1/2}, Cγ = { 1} and Pγ = { 0, 1}.
We see that h(1/2) = 1, h(1) = 0, h(0) = 0. Hence a self-similar map γ = (γ1, γ2) satisfies the
assumption A above.
Example 2.2. Let K = [0, 1]. For n ≥ 2, we take n + 1 numbers ti (i = 0, . . . , n) such that
0 = t0 < t1 < ··· < tn−1 < tn = 1. Let h be a continuous map from K to K such that h
is linear on each subinterval [ti, ti+1] and takes value 0 or 1 at the endpoints of the subinterval
interchangingly. We denote by γi (i = 1, . . . , n) the inverse branches of h. Then γ = (γ1, . . . , γN )
satisfies the assumption A. We note that Bγ = { t1, . . . , tn−1}, and Pγ = { 0, 1}. If n = 2, then
Cγ = {0} or {1}. If n ≥ 3, then Cγ = { 0, 1}.
√ 3
Example 2.3. [14] (Koch curve) Let ω = 1
6 ∈ C. Consider the two contraction γ1,
γ2 on the triangle domain ∆ ⊂ C with vertices { 0, ω, 1} defined by γ1(z) = ωz and γ2(z) =
(1 − ω)(z − 1) + 1 for z ∈ C. Let τ be the reflection in the line x = 1/2. We put γ1 = γ1 and
γ2 = γ2 ◦ τ . Put K = T∞n=1T(j1,...,jn)∈Σn γj1 ◦ ··· ◦ γjn(∆). Then K is the Koch Curve and
(K, γ) is a self-similar map satisfying the assumption A. We note that Bγ = {ω}, Cγ = {1} and
Pγ = {0, 1}.
Example 2.4. [14] (Sierpinski gasket) Let P = (1/2,√ 3/2), Q = (0, 0), R = (1, 0), S =
(1/4,√ 3/4), T = (1/2, 0) and U = (3/4,√ 3/4). Let γ1, γ2 and γ3 be contractions on the
regular triangle T on R2 with three vertices P , Q and R such that
2 + i
γ1(x, y) =(cid:18) x
2
+
1
4
,
1
2
y(cid:19) ,
γ2(x, y) =(cid:16) x
2
,
y
2(cid:17) ,
γ3(x, y) =(cid:18)x
2
+
1
2
,
y
2(cid:19) .
We denote by αθ a rotation by angle θ. We put γ1 = γ1, γ2 = α−2π/3 ◦ γ2, γ3 = α2π/3 ◦ γ3. Then
γ1(∆P QR) = ∆P SU , γ2(∆P QR) = ∆T SQ and γ3(∆P QR) = ∆T RU , where ∆ABC denotes
the regular triangle whose vertices are A, B and C. Put K =T∞n=1T(j1,...,jn)∈Σn (γj1◦···◦γjn)(T ).
Then (K, γ) is a self similar map satisfying assumption A, and K is the Sierpinski gasket.
Bγ = { S, T, U }, Cγ = Pγ = { P, Q, R } and h is given by
h(x, y) =
γ−1
1 (x, y)
γ−1
2 (x, y)
γ−1
3 (x, y)
(x, y) ∈ ∆P SU ∩ K
(x, y) ∈ ∆T SQ ∩ K
(x, y) ∈ ∆T RU ∩ K,
3. Hilbert C∗-bimodules
Following [14], we introduce Hilbert C∗-bimodules (or C∗-correspondences) and their Cuntz-
Pimsner C∗-algebras for self-similar maps.
4
Let A be a C∗-algebra and X a Hilbert right A-module. Let L(X) the set of bounded linear
operators on X which have adjoints with respect to the A-inner product (xy)A. We denote by
θx,y the "rank one" operator given by θx,y(z) = x(yz)A for x, y, z ∈ X. We denote by K(X) the
C∗-algebra generated by rank one operators, which is sometimes called the compact ideal.
We say that X is a Hilbert bimodule over a C∗-algebras A if X is a Hilbert right A- module
with a *-homomorphism φ : A → L(X). We always assume that X is full and φ is injective.
We recall the construction of the Cuntz-Pimsner C∗-algebra OX associated with X. Let
F (X) =L∞n=0 X⊗n be the full Fock module of X with a convention X⊗0 = A. For ξ ∈ X, the
creation operator Tξ ∈ L(F (X)) is defined by
Tξ(a) = ξa
and Tξ(ξ1 ⊗ ··· ⊗ ξn) = ξ ⊗ ξ1 ⊗ ··· ⊗ ξn.
We define iF (X) : A → L(F (X)) by
iF (X)(a)(b) = ab
and iF (X)(a)(ξ1 ⊗ ··· ⊗ ξn) = φ(a)ξ1 ⊗ ··· ⊗ ξn
for a, b ∈ A. The Cuntz-Toeplitz algebra TX is the C∗-algebra acting on F (X) generated by
iF (X)(a) with a ∈ A and Tξ with ξ ∈ X.
Let jKK(X) → TX be the homomorphism defined by jK (θξ,η) = TξT ∗η . We consider the ideal
IX := φ−1(K(X)) of A. Let JX be the ideal of TX generated by {iF (X)(a)− (jK ◦φ)(a); a ∈ IX}.
Then the Cuntz-Pimsner algebra OX is defined as the quotient TX/JX . Let π : TX → OX be
the quotient map. We set Sξ = π(Tξ) and i(a) = π(iF (X)(a)). Let iK : K(X) → OX be the
homomorphism defined by iK(θξ,η) = SξS∗η. Then π((jK ◦ φ)(a)) = (iK ◦ φ)(a) for a ∈ IX.
The Cuntz-Pimsner algebra OX is the universal C∗-algebra generated by i(a) with a ∈ A
and Sξ with ξ ∈ X satisfying that i(a)Sξ = Sφ(a)ξ, Sξi(a) = Sξa, S∗ξ Sη = i((ξη)A) for a ∈ A,
ξ, η ∈ X and i(a) = (iK ◦φ)(a) for a ∈ IX . We usually identify i(a) with a in A. We also identify
Sξ with ξ ∈ X and simply write ξ instead of Sξ. There exists an action β : R → Aut OX defined
by βt(ξ) = eitξ for ξ ∈ X and βt(a) = a for a ∈ A, which is called the gauge action.
Let γ be a self-similar map on a compact metric space K satisfying Assumption A. Let
A = C(K), C =SN
i=1{ (γi(y), y) y ∈ K }, and X = C(C). For f , g ∈ X and a, b ∈ A, we define
left and right A module actions on X and an A-inner product by
(a · f · b)(γj(y), b) =a(γj(y))f (γj(y), y)b(y)
N
y ∈ K,
j = 1, . . . , N
(fg)A(y) =
Xj=1
f (γj(y), y)g(γj(y), y)
y ∈ K.
We define the representation φ of A in L(X) by φ(a)f = a · f . By [14], (X, φ) is proved to be
a C∗-correspondence over A. We denote by Oγ the Cuntz-Pimsner C∗-algebra associated with
X and call it the Cuntz-Pimsner algebra Oγ associated with a self-similar map γ.
We put JX = φ−1(K(X)). Then JX is an ideal of A.
We recall some basic facts on the Cuntz-Pimsner algebra Oγ associated with a self-similar
map γ.
Lemma 3.1. [14] Let γ be a self-similar map on a compact metric space K If (K, γ) satisfies the
open set condition, then the associated Cuntz-Pimsner algebra Oγ is simple and purely infinite.
Moreover JX remembers the branch set Bγ so that JX = { f ∈ A f (b) = 0
for each b ∈ Bγ }
5
Let X⊗n be the n-times inner tensor product of X and φn denotes the left module action of
A on X⊗n. Put
F (n) = A ⊗ I + K(X) ⊗ I + K(X⊗2) ⊗ I ··· + K(X⊗n) ⊂ L(X⊗n)
We embedd F (n) into F (n+1) by T 7→ T ⊗ I for T ∈ F (n). Put F (∞) =S∞n=0 F (n).
By Pimsner [20] we can identify F (n) with the C∗-subalgebra of Oγ generated by A and
SxS∗y for x, y ∈ X⊗k, k = 1, . . . , n under identifying SxS∗y with θx,y . Then the inductive
limit algebra F (∞) is exactly the fixed point subalgebra of Oγ under the gauge action. We
call F (∞) the core of OX . We note that F (n+1) = F (n) ⊗ I + K(X⊗n+1). Thus F (n) is a C∗-
subalgebra of F (n+1) containing unit and K(X⊗n+1) is an ideal of F (n+1). We sometimes write
F (n+1) = F (n) + K(X⊗n+1) for short.
4. The Rieffel correspondence of traces
In this section, we study the Rieffel correspondence of traces on Morita equivalent C∗-algebras
A and B. If C∗-algebras are not unital, then there does not exist a bijective correspondence
between the finite traces on Morita equivalent C∗-algebras A and B in general. For example,
consider A = C ⊕ C and B = C ⊕ K(H). We consider the case that a Hilbert module is
of finite degree type and show the following: If C∗-algebras A and B are Morita equivalent
by an equivalence bimodule of finite degree type for both sides, then there exists a bijective
correspondence between the finite traces on A and B. In general, if C∗-algebras A and B are
Morita equivalent, then there exists a bijective correspondence between the densely defined lower
semi-continuous traces on A and B written by countable bases, see, for example, N. Nawata
[19].
Since the correspondence can be explicitely written using basis, we recall the notion of basis
for a Hilbert C∗-module, see [12] and [10]. Let X = XA be a (right) Hilbert C∗-module X over
a C∗-algebra A. A family (ui)i∈I in X is called a (right) basis, (or a tight frame more precisely
as in [6]) of X if
x =Xi∈I
ui(uix)A for any x ∈ X,
where the sum is taken as unconditional norm convergence. Furthermore (ui)i∈I is called a finite
basis if (ui)i∈I is a finite set. If a Hilbert C∗-module is countably generated, then there exists a
countable basis (that is, finite or a countably infinite basis ) of X and written as {ui}∞i=1, where
some ui may be zero. Let {ui}∞i=1 be a countable basis of Hilbert right B-module X and {vj}∞j=1
a countable basis of Hilbert right C-module Y . Assume that there exists a ∗-homomorphisms
φ : B → L(Y ). Then {ui ⊗ vj}i=1,2,...,j=1,2,... constitutes a basis of X ⊗B Y for any ordering ([12]
and [15]).
Let A be a C∗-algebra. We denote by T (A) the set of tracial states on A. We assume
that T (A) is not empty. Let X be a countably generated Hilbert A-module and {ui}∞i=1 a
basis of X. For a tracial state τ on A, supnPn
i=1 τ ((uiui)A) ∈ [0,∞] does not depend on the
choice of basis {ui}∞i=1 as in [15], see also the proof of Proposition 4.1 of this paper. We put
dτ = supnPn
i=1 τ ((uiui)A). We call that X is of finite degree type if supτ∈T (A) dτ < ∞ ([12]
and [15]). For example, let X be the Hilbert bimodule associated with a self-similar map γ =
(γ1, . . . , γN ) of N proper contractions. Then XA is of finite degree type and supτ∈T (A) dτ = N .
In the papr we assume that a Hibert C∗-module is countably generated. Let A and B be C∗-
algebras. We say that an B − A equivalence bimodule X = BXA is of finite degree type for both
6
sides if T (A) and T (B) are not empty and XA and BX are of finite degree type, where XA and
BX are Hilbert C∗-modules considered as a right A-module and a left B-module respectively.
Proposition 4.1. If C∗-algebras A and B are Morita equivalent by an equivalence bimodule X
of finite degree type for both sides, then there exists a bijective correspondence between the finite
traces on A and B. Let {ui}∞i=1 and {wj}∞j=1 be countable bases of XA and BX respectively. The
correspondence is given explicitely as follows:
For a finite trace τ on A, we can define a finite trace π(τ ) on B by
For a finite trace σ on B, we can define a finite trace ψ(σ) on A by
π(τ )(b) =
τ ((uibui)A).
ψ(σ)(a) =
σ(B(wjawj)).
∞
Xi=1
∞
Xj=1
The correspondence does not depend on the choice of bases. The correspondence preserves the
order of traces. Moreover we have
π(τ )(B(xy)) = τ ((yx)A),
ψ(σ)((zw)A) = σ(B(wz)).
Proof. Firstly, we consider the case that B = K(XA). Since X is of finite degree type, for
T ∈ K(X)+, we have
∞
Xi=1
τ ((uiT ui)A) ≤ kTkdτ .
For general T ∈ K(X), we have P∞i=1 τ ((uiT ui)A) ≤ 4dτkTk < ∞. Hence π(τ ) is well-defined.
Let {vj}∞j=1 be another basis of XA. Then for T ∈ K(X), we have
vj(vjT ui)AT ui)A
∞
∞
∞
∞
∞
∞
∞
Xi=1
τ ((T uiT ui)A) =
τ
Xj=1
Xi=1
τ ((uiT ∗T ui)A) =
(
(vjT ui)∗A(vjT ui)A
τ
Xj=1
Xi=1
Xj=1
τ ((vjT ui)∗A(vjT ui)A)
=
Xi=1
Xj=1
Xi=1
τ ((vjT ui)A(vjT ui)∗A)
τ ((vjT ui)∗A(vjT ui)A) =
Xj=1
Xi=1
τ ((T ∗vj
τ ((T ∗vjT ∗vj)A) =
τ ((T ∗vjui(uiT ∗vj)A)) =
τ ((vjT T ∗vj)A).
ui(uiT ∗vj)A)A)
Xi=1
∞
∞
∞
∞
∞
∞
=
=
∞
∞
Xi=1
Xi=1
Xj=1
Xj=1
Xj=1
∞
∞
=
=
∞
Xj=1
This shows that for S ∈ K(X)+
∞
Xi=1
τ ((uiSui)A)
7
does not depend on the choice of the basis {ui}∞i=1 by taking T = S1/2. Moreover, the same
calculation shows that T 7→P∞i=1 τ ((uiT ui)A) gives a finite trace on K(XA) by the polarization
identity. Secondly we consider the general case. Since there exists an isomorphism of B onto
K(XA), we can define a finite trace π(τ ) on B by
Xi=1
τ ((uibui)A).
π(τ )(b) =
∞
Similary, for a finite trace σ on B, we can define a finite trace ψ(σ) on A by
ψ(σ)(a) =
∞
Xj=1
σ(B(wjawj)).
∞
Since B(xy)z = x(yz)A = θx,y(z),
∞
Xi=1
Xi=1
τ ((uiB(xy)ui)A) =
Xi=1
∞
=
∞
τ (uix(yui)A)A) =
Xi=1
τ ((yui)A(uix)A) = τ (y
τ ((uix)A(yui)A)
Xi=1
∞
ui(uix)A) = τ ((yx)A),
we have
Similarly we have
π(τ )(B(xy)) = τ ((yx)A).
Since the Hilbert bi-module is full on both sides, the correspondence is bijective. By definition,
the correspondence preserves the order of traces.
ψ(σ)((zw)A) = σ(B(wz)).
(cid:3)
5. Analysis for point mass
Let γ be a self-similar map on K which satisfies the assumption A. We denote by B(K) the
set of bounded Borel functions on K, and by M (K) the set of finite Borel measures on K. For
a Borel function a on K, we define a Borel function a by
a(y) =
1
eγ(γj(y), y)
N
Xj=1
a(γj(y)) = Xx∈h−1(y)
a(x).
We usually identify a finite Borel measure µ with the associated positive linear functional on
Even if a is continuous, a is not continuous in general, because we do not count the multiplicity
in {x ∈ h−1(y)}.
C(K). So, we may write as µ(a) := R adµ for a ∈ C(K) for short. For a Borel set B ⊂ K, we
may write µ(B) = µ(χB) if no confusion arize.
For a finite Borel measure µ on K, we define a finite Borel measure F (µ) on K by
F (µ)(a) = µ(a)
8
(a ∈ C(K)).
In particular, we have F (δy) = Px∈h−1(y) δx, where δy is the Dirac measure on y and the sum
on x should be taken without multiplicity. In fact
F (δy)(a) = δy(a) = a(y) =
Xj=1
1
eγ(γj(y), y)
a(γj(y)) = Xx∈h−1(y)
δx(a).
N
Lemma 5.1. Let X be a C∗-correspondence associated with a self-similar map γ on K. Let
{ui}∞i=1 be a countable basis of X. Then for a ∈ A and y ∈ K, we have
∞
Xi=1
(uiφ(a)ui)A(y) = a(y).
Proof. For f ∈ X = C(C), we have f (x, y) ≤ kfk2 and the identity P∞i=1 ui(uif )A = f
converges in norm k k2. Hence the left side converges also pointwisely. For each fixed y ∈ K,
k = 1, . . . , N , we consider the value of P∞i=1 ui(uif )A = f at (x, y) = (γk(y), y):
N
lim
n→∞ n
Xi=1
n→∞
ui(γk(y), y)(uif )A(y)! = lim
Xi=1
=f (γk(y), y).
n
ui(γk(y), y)
Xj=1
ui(γj(y), y)f (γj(y), y)
n
N
For j with γj(y) 6= γk(y), we take f ∈ X such that f (γk(y), y) = 1 and f (γj(y), y) = 0. Then
we have
n→∞
ui(γk(y), y)eγ(γk(y), y)ui(γk(y), y)!
ui(γj(y), y)f (γj(y), y)
ui(γk(y), y)
Xi=1
Xj=1
lim
n→∞ n
Xi=1
= lim
Xi=1
eγ(γk(y), y)ui(γk(y), y)2 = 1.
∞
=
Recall that the branch index at (γj(y), y) is given by eγ(γj(y), y) = #{i ∈ Σγj(y) = γi(y)} and
and that eγ(γj(y), y) express the multiplicity. For a ∈ A, we have
∞
∞
N
N
=
∞
Xi=1
Xj=1
a(γj(y))ui(γj(y), y)2
∞
Xi=1
(uiφn(a)ui)A(y) =
Xi=1
Xj=1
Xj=1
Xj=1
eγ(γj(y), y)
eγ(γj(y), y)
a(γj(y))
Xi=1
=
=
1
1
N
N
a(γj(y)) = a(y).
ui(γj(y), y)a(γj(y))ui(γj(y), y)
eγ(γj(y), y)ui(γj(y), y)2!
We shall rephrase the lemma above as follows:
9
(cid:3)
Corollary 5.2. Let X be the C∗-correspondence associated with a self-similar map γ on K. Let
{ui}∞i=1 be a countable basis of X. Then we have
∞
Xi=1
µ((uiφ(a)ui)A) = F (µ)(a).
Moreover let X⊗n be n-th tensor product of X and {vj}∞j=1 be a countable basis of X⊗n. Then
we have
∞
Xj=1
µ((vjφ(a)vj )A) = F n(µ)(a).
Proof. It is enough to recall that X⊗n
map γn = (γi1 ◦ ··· ◦ γin)(i1,...,in)∈Σn on K as in [14].
A is the Hilbert C∗-module associated with a self similar
(cid:3)
The identity above is proved for a certain concrete basis in Kajiwara [10] by a direct compu-
tation.
Let X be the C∗-correspondence associated with a self-similar map γ and n an integer with
n ≥ 1. We define yn
0 ∈ X⊗n by
yn
0 := (1/√ N )n1C ⊗ 1C ⊗ ··· ⊗ 1C ∈ X⊗n.
Since (yn
0yn
0 )A = I, for z ∈ X⊗n
K(XA)(zyn
0 )yn
0 = θz,yn
0 (yn
0 ) = z(yn
0yn
0 )A = z.
This shows that a singleton set {y0} is a finite basis for the left K(XA)-module K(XA)X⊗n.
Lemma 5.3. Let X be the C∗-correspondence associated with a self-similar map γ and n an
A and K(X ⊗n)X⊗n are of finite degree type
integer with n ≥ 1. Then Hilbert C∗-modules X⊗n
and full.
Proof. First, we prove that X⊗n
function on K, in Lemma 5.1. Then we have
A is of finite degree type for n = 1. Put a = 1K , a constant
∞
Xi=1
(uiui)A(y) = 1K (y) ≤ N · 1K ,
Hnece, for any tracial state τ on A, we have
∞
Xi=1
τ ((uiui)A) ≤ N.
Thus XA is of finite degree type. For general n, the conclusion follows from the fact that X⊗n
A
is the Hilbert C∗-module associated with a self similar map γn = (γi1 ◦ ··· ◦ γin)(i1,...,in)∈Σn on
0} constitutes a finite basis of K(X ⊗n)X⊗n, it is clear that K(X ⊗n)X⊗n is
K as in [14]. Since {yn
of finite degree type.
A is full. By definition, K(X ⊗n)X⊗n is full as a left K(X⊗n)-module.
0 )A = 1K , X⊗n
Since (yn
0 , yn
(cid:3)
10
Let {ξi}∞i=1 be a basis for X⊗n. By Proposition 4.1 and Lemma 5.3, for a finite Borel measure
µ, there exists a finte trace πn(µ) on K(X⊗n) such that
πn(µ)(T ) =
µ((ξiT ξi)A).
∞
Xi=1
and for ξ, η ∈ X⊗n,
πn(µ)(θξ,η) = µ((ηξ)A).
0 ∈ X⊗n constitutes a basis for the left Hilbert C∗-module K(X ⊗n)X⊗n, for a trace τ on
Since yn
K(X⊗n), there exists a finite Borel measure ψ(τ )n such that
0 ).
ψn(τ )(a) = τ (θyn
0 a,yn
for a ∈ A = C(K).
We shall study the traces on the core F (∞) =S∞n=0 F (n), where F (n) = A ⊗ I + K(X) ⊗ I +
K(X⊗2) ⊗ I ··· + K(X⊗n) ⊂ L(X⊗n). We usually identify F (n) with the C∗-subalgebra of Oγ
generated by a ∈ A and SxS∗y for x, y ∈ X⊗k, k = 1, . . . , n, by identifying SxS∗y with θx,y and a
with φ(a).
We note that F (n+1) = F (n) ⊗ I + K(X⊗n+1), where F (n) is a C∗-subalgebra of F (n+1)
containing unit and K(X⊗n+1) is an ideal of F (n+1). Any finite trace τ on F (∞) is determined
by a sequence (τ[n])n of its restrictions τ[n] to F (n). Each trace τ[n+1] is determmied by its
restriction τ[n] to F (n) and the restriction τn+1 to K(X⊗n+1) which is described by a finite Borel
measure µn+1 on K under the Riefell correspondence.
Conversely a compatible sequence (τ[n])n of traces on increasing subalgebras
A = F (0) ⊂ F (1) ⊂ F (2) ⊂ . . . F (n) ⊂ . . .
gives a trace on the inductive limit F (∞). First we find a sequence (µn)n of finite Borel measures
on K, which gives a sequence (τn)n of traces on K(X⊗n) by the Riefell correspondence. We shall
extend a trace τ[n] on F (n) and a trace τn+1 on K(X⊗n+1) to a trace τ[n+1] on F (n+1).
Therefore we need the following extension lemma of traces to construct traces on the core
and classify them.
Let A be a C∗-algebra and I be an ideal of A. For a positive linear functional ϕ on I, we
denote by ϕ the canonical extension of ϕ to A. Recall that for an approximate unit (uλ)λ of I
and x ∈ A, we have
ϕ(x) = lim
λ
ϕ(xuλ)
We refer [1] for the property of the canonical extension of states.
The following key lemma is proved in Proposition 12.5 of Exel and Laca [4] for state case,
and is modified in Kajiwara and Watatani [15] for trace case.
Lemma 5.4. Let A be a unital C∗-algebra. Let B a C∗-subalgebra of A containing the unit and
I an ideal of A such that A = B + I. Let τ be a finite trace on B, and ϕ a finite trace on I.
Suppose that the following conditions are satisfied :
(1) ϕ(x) = τ (x) for x ∈ B ∩ I.
(2) ϕ(x) ≤ τ (x) for positive x ∈ B.
11
Then there exists a unique finite trace on A which extends τ and ϕ. Conversely, if there exists
an finite trace on A, then its restrictions τ and ϕ on B and I must satisfy the above conditions
(1) and (2).
In order to describe an inductive construction of traces on the core smoothly, we need to
modify the Rieffel correspondence of finite traces τn on K(X⊗n) and finite measures µn on K
by a normalization as follows:
Lemma 5.5. Let γ be a self-similar map on K. Fix a natural number n. For a finite Borel
measure µ on K, there exist a unique finite trace τ = πn(µ) on K(X⊗n) such that
τ (θξ,η) =
1
N n µ((ηξ)A).
for ξ, η ∈ X⊗n. Conversely, for a finite trace τ on K(X⊗n), there exists a unique finite Borel
measure µ = ψn(τ ) on K such that πn(µn) = τn and
µ(a) = N nτ (θyn
0 a,yn
0 ),
for a ∈ A = C(K).
Proof. We can put πn(µ) =
1
N n πn(µ), ψn(τ ) = N n ψn(τ ).
(cid:3)
In the following we shall investigate compatibility conditions of finite traces τ[n] on F (n) in
terms of finite traces τn on K(X⊗n) and corresponding finite measures µn on K.
Proposition 5.6. Let γ be a self-similar map on K. Let µ be a probablity Borel measure on K,
which is identified with a normalized trace on C(K). For a positive integer r, let ϕ = πr(µ) be
the corresponding finite trace on K(X⊗r) under the modified Riefell correspondence in Lemma
5.5. Consider the canonical extension ϕ of ϕ to L(X⊗r). For an integer k with 0 ≤ k ≤ r, put
µk = 1
N r−k F r−k(µ). Then the restriction of ϕ to K(X⊗k) ⊗ I is exactly the corresponding finite
trace on K(X⊗k) for µk under the modified Riefell correspondence, that is, for x, y ∈ X⊗k, we
have
ϕ(θx,y ⊗ I) =
1
N k µk((yx)A).
12
Proof. Let (si)i be a basis of X⊗k and (tj)j be a basis of X⊗(r−k). Then (si ⊗ tj)i,j is a basis of
X⊗r = X⊗k ⊗ X⊗(r−k) ([12]). Using Lemma 5.1 for X⊗(r−k) and F r−k, for x, y ∈ X⊗k, we have
ϕ(θx,y ⊗ I) =
ϕ((θx,y ⊗ I)θsi⊗tj ,si⊗tj ) =
∞
∞
∞
∞
Xj=1
Xj=1
∞
1
N k
1
N k
1
∞
Xi=1
Xi=1
Xi=1
Xi=1
∞
∞
1
N k
1
∞
1
N k
Xj=1
ϕ(θθx,y(si)⊗tj ,si⊗tj )
∞
Xj=1
Xi=1
1
Xi=1
N r µ((si ⊗ tjθx,y(si) ⊗ tj)A) =
Xi=1
Xi=1
N r−k (F r−kµ)((six)A(ysi)A) =
1
si(six)A))A) =
N k µk((yx)A).
N r−k (F r−kµ)((siθx,y(si))A) =
1
N k
1
N k
∞
∞
∞
1
1
1
=
=
=
=
N r−k (F r−kµ)((y
Xi=1
1
N r−k µ((tj(siθx,y(si))Atj)A)
N r−k (F r−kµ)((six(ysi)A)A))
N r−k (F r−kµ)((ysi)A(six)A)
(cid:3)
Corollary 5.7. Let τ be a finite trace on F (n+1) and τn and τn+1 be its restriction to K(X⊗n)⊗I
and K(X⊗n+1). Let µn = ψn(τn) and µn+1 = ψn+1(τn+1) be the corresponding Borel measures
on K by modified Riefell correspondence. Then we have the following relations:
(1) For T ∈ K(X⊗n)+ ⊂ F (n), we have that
τn+1(T ⊗ I) = πn(cid:18)(cid:18) 1
N(cid:19) F (ψn+1(τn+1))(cid:19) (T ) ≤ τn(T ).
(2) For f ∈ B(K)+, we have that
ψn(πn+1(µn+1)K(X⊗n))(f ) =
1
N
F (µn+1)(f ) ≤ ψn(τn)(f ) = µn(f )
(3) For T ⊗ I ∈ K(X⊗n ⊗ I) ∩ K(X⊗n+1), we have that
τn(T ) = πn(cid:18)(cid:18) 1
N(cid:19) F (µn+1))(cid:19) (T ).
Proof. Apply Lemma 5.4 as A = F (n+1), B = F (n) ⊗ I and I = K(X⊗n+1).
(cid:3)
The Corollary above is used to show a relation between point masses of µn and µn+1 later.
The content of (2) in the Corolally above is expressed by the following commutative diagram:
T (K(X⊗n+1))
ψn+1y
M (K)
βn+1−−−−→ T (K(X⊗n))
ψny
1
N F
−−−−→ M (K)
where βn+1 is defined by βn+1(τn+1) = τn+1K(X⊗n)
obtained in [8].
In particular, if µn+1 is a Dirac measure δy, we have the following Corollary, which was
13
Corollary 5.8.
ψn((πn+1(δy)K(X⊗n)) =
1
N Xx∈h−1(y)
δx
where we do not count the multiplicity for the right hand side.
For f ∈ C(K), we define αn(f ) ∈ C(K) by αn(f )(x) = f (hn(x)). Then it holds that
ξf = φn(αn(f ))ξ, (ξ ∈ X⊗n).
For T ∈ K(X⊗n), we need to find f ∈ C(K) such that (T φn(αn(f ))) ⊗ I ∈ K(X⊗n) ⊗ I ∩
K(X⊗n+1). By Pimsner [20], it holds that
K(X⊗n) ⊗ I ∩ K(X⊗n+1) = K(X⊗nJX ) ⊗ I.
where JX := ϕ−1(K(X)) = { f ∈ C(K) f (b) = 0 for b ∈ Bγ}.
The right hand side is expressed as
K(X⊗nJX ) =C∗(Xi
=C∗(Xi
θξigi,ηifi ξi, ηi ∈ X⊗n, gi, fi ∈ JX)
θξi,ηifi ξi, ηi ∈ X⊗n, fi ∈ JX) .
For ξ, η ∈ X⊗n, we have that
θξ,ηf = θξ,φn(αn(f ))η = θξ,ηφn(αn(f )∗) ∈ L(X⊗n).
Since any T ∈ K(X⊗n) is a liimit of a sequence (Tk)k such that Tk =Pp=1 θxp,yp (xp, yp ∈ X⊗n)
is a finite sum of "rank one " operators, we have obtain the following lemma:
Lemma 5.9. Let T ∈ K(X⊗n) and f ∈ JX, then (T φn(αn(f )))⊗ I ∈ K(X⊗n)⊗ I ∩K(X⊗n+1).
The following Lemmas describe relations of point masses of the corresponding measures µn
and µn+1 on K for a finite trace on F (n+1).
Lemma 5.10. Let τ be a finite trace on F (n+1) and τn and τn+1 be its restriction to K(X⊗n)⊗ I
and K(X⊗n+1). Let µn = ψn(τn) and µn+1 = ψn+1(τn+1) be the corresponding Borel measures
on K by modified Riefell correspondence. Let b ∈ K and b = γi(a) for some i, that is, h(b) = a.
Then we have the following:
(1) If b is not in Bγ, then the point masses on a and b are related as µn+1({a}) = N µn({b}).
(2) If b is in Bγ, then µn+1({a}) ≤ N µn({b}).
Proof. (1)Suppose that b is not in Bγ. There exist a sequence of compact neighborhoods Fm of
b and open neighborhoods Vm of b such that Fm ⊂ Vm, Vm does not contain any element of Bγ
and Vm converges to b as m tends to ∞. We take fm ∈ C(K)+ such that 0 ≤ fm ≤ 1, fm = 1
on Fm and fm = 0 on Vm
. Since fm vanishes on Bγ, we have fm is in JX. Therefore
c
Then by (3) in Corollary 5.7, it holds that
θyn
0 ,yn
0 fm ⊗ I ∈ K(X⊗n) ⊗ I ∩ K(X⊗n+1).
τn(θyn
0 ,yn
0 fm) = πn(cid:18)(cid:18) 1
N(cid:19) F (µn+1))(cid:19) (θyn
14
0 ,yn
0 fm).
Then
1
1
N n µn((yn
0 fmyn
N n µn(fm) = 1
1
N n+1 F (µn+1)((yn
0 )A) =
N n+1 F (µn+1)(fm). Let m → ∞, then fm converges to the
0 fmyn
0 )A).
0yn
Since (yn
characterestic function χ{b} on the one point set {b}. Since χ{b} = χ{a},
0 )A = 1,
µn({b}) =
F (µn+1)({b}) =
1
N
1
N
µn+1({a}).
(2)Just apply (2) in Corollary 5.7. A similar calculation directly implies that
1
N
F (µn+1)(χ{b}) ≤ µn(χ{b}).
Therefore we have 1
N µn+1({a}) ≤ µn({b}).
6. Construction of model traces on the core
(cid:3)
For each b ∈ Bγ and r = 0, 1, 2, . . . , we shall construct a model trace τ (b,r) on F (∞).
We denote by δb the Dirac measure on b. For 0 ≤ i ≤ r, we define Borel measures µ(b,r)
i
on K
by
We note that
µ(b,r)
i
(f ) =
µ(b,r)
i
=
1
N r−i F r−i(δb),
1
N r−i X(j1,j2,...,jr)∈Σr−i
f (γj1 ◦ ··· ◦ γjr−i(b)),
(f ∈ A = C(K))
because h−k(b) does not contain the brancehd points for k = 1, 2, 3, . . . by the Assumption A.
For i > r we define Borel measures µ(b,r)
= 0 on K.
i
Proposition 6.1. Let γ be a self-similar map on K with assumption A. For each b ∈ Bγ and
r = 0, 1, 2, . . . , there exists a unique finite trace τ (b,r) on F (∞) such that τ (b,r)A = µ(b,r)
, the
restriction of τ (b,r) to K(X⊗i) is πi(µ(b,r)
) for 0 ≤ i ≤ r and the restriction of τ (b,r) to K(X⊗i)
is zero for i > r, that is,
0
i
τ (b,r)(f ) = (
1
N r F r(δb))(f ) =
1
N r Xx∈h−r(b)
f (x)
for
f ∈ A
and
τ (b,r)(θx,y) =
1
N i µ(b,r)
i
((yx)A) for x, y ∈ X⊗i,
i < r.
Proof. Recall that F (n) = A⊗ I +K(X)⊗ I +K(X⊗2)⊗ I ··· +K(X⊗n) ⊂ L(X⊗n). We identify
F (n) with the C∗-subalgebra of Oγ generated by A and SxS∗y for x, y ∈ X⊗k, k = 1, . . . , n by
identifying SxS∗y with θx,y. We note that F (n+1) = F (n) ⊗ I + K(X⊗n+1). We shall construct
finite traces τ (b,r)
on F (n) n = 0, 1, 2, . . . by induction on n. We choose a suitable sequence
(µ(b,r)
)n of compatible traces on
K(X⊗n) by the modified Riefell correspondence. We shall extend a trace τ (b,r)
on F (n) and a
trace τ (b,r)
)n of finite Borel measures on K, which gives a sequence (τ (b,r)
[n]
n+1 on K(X⊗n+1) to a trace τ (b,r)
[n+1] on F (n+1) by Lemma 5.4.
[n]
n
n
15
We put τ (b,r)
= π0(µ(b,r)
Let (ui)i be a basis for X⊗r. Then
[0] = τ (b,r)
0
0
) and τ (b,r)
1
= π1(µ(b,r)
1
). Then τ (b,r)
[0]
is a tracial state. In fact,
τ (b,r)
[0]
(1) = π0(µ(b,r)
0
) = (
1
N r F r(δb))(1) = 1
We show that τ (b,r)
[0]
extends to a trace τ (b,r)
1
on F (1). We note that for a ∈ A
because we have
1
N
F (µ(b,r)
1
) = µ(b,r)
0
τ (b,r)
1
(a) = π0(cid:18) 1
0
1
N
F (ψ1(τ (b,r)
))(a)(cid:19) = τ (b,r)
. For a ∈ A ∩ K(X), we have
(a) = τ (b,r)
(a) = π(b,r)
(a).
1
0
τ (b,r)
1
(a),
By Lemma 5.4, we can construct a finite trace τ (b,r)
[1]
on A +K(X) which extends τ (b,r)
Difine a trace τ (b,r)
i
:= πi(µ(b,r)
i
) on K(X⊗i) for 0 ≤ i ≤ r and a trace τ (b,r)
i
i > r.
By induction on n, assume that we constructed a finite trace τ (b,r)
[n]
1
[0]
and τ (b,r)
.
:= 0 on K(X⊗i)for
on F (n) satisfying that
r+1 = 0, clearly, τ (b,r)
r+1 = 0 on F (r+1). Hence τ (b,r)
r+1 = 0 ≤ τ (b,r)
[r]
on F (r).
Let T ∈ K(X⊗r) ⊗ I ∩ K(X⊗r+1) = K(X⊗rJX). We need to show that τ (b,r)
T ∈ K(X⊗rJX ) is written as T = P∞i=1 θξi,ηifi for fi ∈ JX . Since fi(b) = 0 for b ∈ Bγ, it holds
(T ) = 0. Any
that
r
τ (b,r)
r
(T ) =
µ(b,r)
r
((ηiξi)Afi) =
((ηiξi)Afi)(b) = 0.
∞
Xi=1
∞
µ(b,r)
r
Xi=1
[n+1] on F (n+1).
By Lemma 5.4, we can extend to a finite trace τ (b,r)
(iii)(Case that n > r):
n+1 = 0 and τ (b,r)
n = 0, we have that τ (b,r)
Since we put τ (b,r)
n+1 = 0 ≤ τ (b,r)
n+1 = 0 on F (n+1). Hence τ (b,r)
on F (n). Since F (n) ∩ K(X⊗n+1) = K(X⊗n+1) ∩ K(X⊗n) = K(X⊗nJX ) ⊗ I, τ (b,r)
n+1 are
equal to zero on the intersection F (n) ∩ K(X⊗n+1). Therefore by Lemma 5.4, we can extend to
a finite trace τ (b,r)
and τ (b,r)
[n]
[n]
[n+1] on F (n+1).
16
τ (b,r)(cid:12)(cid:12)K(X ⊗i) = πi(µ(b,r)
.
We consider the three cases that n < r, n = r and n > r.
(i)(Case that n < r):
) for 1 ≤ i ≤ n and τ (b,r)(cid:12)(cid:12)A = µ(b,r)
0
i
By Corollarly 5.7, for 0 ≤ i ≤ n and T ∈ K(X⊗i), we have that
n+1 ))(cid:19) (T ) = πi(cid:16)µ(b,r)
N n−i(cid:19) F (µ(b,r)
n+1 (T ⊗ I) = πi(cid:18)(cid:18) 1
n+1 = τ (b,r)
τ (b,r)
i
[n]
Thus τ (b,r)
F (n+1).
Since we put τ (b,r)
(ii)(Case that n = r):
on F (n). Therefore by Lemma 5.4, we can extend to a finite trace τ (b,r)
[n+1] on
)(cid:17) (T ) = τ (b,r)
[n] (T ⊗ I).
Hence we have constructed finite traces τ (b,r)
for 0 ≤ n′ ≤ n. The family τ (b,r)
[n]
τ (b,r)
[n′]
such τ (b,r)A = µ(b,r)
0
and τ (b,r)(cid:12)(cid:12)K(X ⊗i) = πi(µ(b,r)
i
[n]
(cid:12)(cid:12)(cid:12)F (n′)
on F (n) n = 0, 1, 2, . . . such that τ (b,r)
=
on F (n) n = 0, 1, 2, . . . gives a finite trace τ (b,r) on F (∞)
) for 0 ≤ i ≤ r and τ (b,r)(cid:12)(cid:12)K(X ⊗i) = 0 for i > r.
[n]
(cid:3)
We shall construct a continuous type trace τ (∞) which corresponds to the Hutchinson measure
µH on K.
For f ∈ C(K), we put
G(f )(y) =
1
N
f (γj(y)).
N
Xj=1
Then G : C(K) → C(K) is a positive linear map. We note that the dual map G∗ is a positive
linear map on C(K)∗. We denote by P (K) the space of probability measure on K. We note
that G∗ conserves P (K). We shall collect some folklore facts with simle direct proofs:
Lemma 6.2. Let γ be a self-similar map on K. Then we have the following:
(1) There exists a metric L on a compact space P (K) such that G∗ is a proper contraction
with respect to L.
(2) There exists a unique fixed point µH in P (K) under G∗ , which is called the Hutchinson
measure µH on K, that is, G∗(µH) = µH.
(3) ∩∞n=1G∗n(P (K)) = {µH}
(4) 1
N F (µH) = G∗(µH) = µH.
Proof. (1) We denote by Lip(K) the set of Lipschitz continuous functions on K. Recall that the
Lipschitz semi-norm k · kL on Lip(K) is defined by
a(x) − a(y)
,
d(x, y)
for a ∈ C(K).
kakL = sup
x6=y
We define a function L on P (K) × P (K) by
L(µ, ν) = sup
kakL≤1µ(a) − ν(a).
By Hutchinson [7], L is a metric on P (K) and the topology on P (K) given by L coincides with
the w*-topology.
We shall show
using d(γj(x), γj (y)) ≤ cd(x, y) for each j. We have
L(G∗(µ), G∗(ν)) ≤ cL(µ, ν)
G(a)(x) − G(a)(y)
d(x, y)
N
1
N
N
≤
Xj=1
Xj=1
≤
≤ckakL,
1
N
a(γj(x)) − a(γj(y))
d(x, y)
a(γj(x)) − a(γj(y))
d(γj(x), d(γj (y)))
d(γj(x)), γj (y))
d(x, y)
17
and it follows that kG(a)kL ≤ ckakL. Then we have
L(G∗(µ), G∗(ν)) = sup
kakL≤1µ(G(a)) − ν(G(a))
≤ sup
kakL≤1kG(a)kLL(µ, ν)
≤ sup
kakL≤1
ckakLL(µ, ν) ≤ cL(µ, ν).
It holds that L(G∗(µ), G∗(µ)) ≤ cL(µ, ν). This shows that G∗ is a proper contraction with
respect to L.
(2) is a consequence of (1).
(3)Since P (K) is compact, the diameter of P (K) with respect to L is finite. Therefore (1) and
(2) implies (3).
(4)Since µH does not have point mass, for a ∈ C(K)
µH(a) = µH(G(a)) = (G∗(µH))(a) = µH(a).
(cid:3)
(
1
N
F (µH ))(a) =
1
N
Proposition 6.3. Let γ be a self-similar map on K with assumption A. Let µH be the Hutchin-
son measure on K. Then there exists a unique finite trace τ∞ on F (∞) such that τ∞(a) =R aµH
for a ∈ A = C(K) and
τ∞(θx,y) =
1
N i µH((yx)A) for x, y ∈ X⊗i.
Proof. Difine a trace τ∞i
a ∈ A = C(K) and
:= πi(µH ) on K(X⊗i) for i = 0, 1, 2, . . . , that is, τ∞(a) = R aµH for
τ∞(θx,y) =
1
N i µH((yx)A)) for x, y ∈ X⊗i.
We shall construct finite traces τ∞[n] on F (n) n = 0, 1, 2, . . . by induction on n. By Corollarly 5.7,
for 0 ≤ i ≤ n and x, y ∈ X⊗i, we have that
τ∞n+1(θx,y⊗I) = ((cid:18) 1
N i(cid:19) (µH)((yx)A) = τ∞[n](θx,y⊗I)
Thus τ∞n+1 = τ∞[n] on F (n). Therefore by Lemma 5.4, we can extend to a finite trace τ∞[n+1] on
F (n+1).
N n+1−i(cid:19) (F n+1−i(µH))((yx)A) = ((cid:18) 1
N i(cid:19) ((cid:18)
(cid:3)
1
7. Classification of traces on the core
We investigate the place where point masses appear on K if a normalized trace τ on F (∞) is
restricted to A.
Proposition 7.1. Let γ be a self-similar map on K with assumption A. Then the restriction τ0 of
a finite trace τ on F (∞) to A does not have point mass except for points in Orb = ∪b∈Bγ∪∞r=0Ob,r,
where Ob,r = h−r(b).
Proof. Let µ0 = ψ0(τ0) be the corresponding Borel measures on K. Take a point a ∈ K which
is not in Orb. Then for any n = 0, 1, 2, 3, . . . , an := hn(a) /∈ Bγ, where a0 = a. Assume that
18
µ0({a}) > 0. We shall show that #(h−n(an)) ≥ N n−1. Firstly, consider the case that h−1(an)
contains an element bn−1 ∈ Bγ. Then
r=1 h−r(bn−1)) ∩ Bγ = ∅
(∪n−1
r=1 h−r(bn−1)) ∩ Bγ was not empty. Then there exists an
r=1 h−r(bn−1)) ∩ Bγ. Then there exists some r = 1, . . . , n − 1 with hr(c) =
r=1 h−r(bn−1)) ∩ Bγ = ∅. Then
Secondly, consider the case that h−1(an)∩Bγ = ∅ and h−2(an) contains an element bn−2 ∈ Bγ.
On the contrary suppose that (∪n−1
element c ∈ (∪n−1
bn−1. This contradicts to Assumption A (3). Therefore (∪n−1
#(h−(n−1)(bn−1)) = N n−1. Since #(h−1(an) ≥ 1, we have that #(h−n(an)) ≥ N n−1.
Then a similar argument shows that
(∪n−2
r=1 h−r(bn−2)) ∩ Bγ = ∅
Thus #(h−(n−2)(bn−2)) = N n−2 and #(h−1(an)) ≥ N . Therefore #(h−n(an)) ≥ N n−1. Con-
sider the case that h−i(an) ∩ Bγ = ∅ for i = 1, 2, . . . , r − 1 and h−r(an) contains an element
bn−r ∈ Bγ with r ≤ n. A similar argument shows that #(h−n(an)) ≥ N n−1. Finally consider
the case that h−i(an) ∩ Bγ = ∅ for i = 1, 2, . . . , n. Then #(h−n(an)) ≥ N n. In any cases, we
have #(h−n(an)) ≥ N n−1.
By Lemma 5.10, for any x ∈ h−n(an),
µ0({x}) ≥
= µ0({a0}) > 0.
µn({an})
N n
Therefore µ0(K) ≥ µ0(h−n(an)) ≥ N n−1µ0({a0}) > 0. Taking n → ∞, µ0(K) = ∞. This
contradicts to that µ0 is a finite measure. Therefore µ0({a}) = 0.
Proposition 7.2. Let γ be a self-similar map on K with assumption A. Let τ0 be the restriction
of a finite trace τ on F (∞) to A. Let µ0 = ψ0(τ0) be the corresponding Borel measures on K
by modified Riefell correspondence. For b ∈ Bγ, r ≥ 0, take any xb,r, x′b,r ∈ Ob,r = h−r(b), then
point masses µ0({xb,r}) = µ0({x′b,r})
Proof. This follows directly from Lemma 5.10.
(cid:3)
(cid:3)
Let τ be a finite trace on F (∞) and τi be the restriciton of τ to K(X⊗i). Let µi = ψi(τi) be the
corresponding Borel measures on K by modified Riefell correspondence. For b ∈ Bγ and r ≥ 0,
the proposition above helps us to define a constant cb,r ≥ 0 by the point mass cb,r = µ0({xb,r})
for any xb,r ∈ Ob,r = h−r(b). Let 0 ≤ i ≤ r. By Lemma 5.10, the point mass of µi at each
point of x ∈ Ob,r−i is determined by µi({x}) = cb,rN i and does not depend on the choice of such
x ∈ Ob,r−i.
Lemma 7.3. Let γ be a self-similar map on K with assumption A. For any b, b′ ∈ Bγ and
integers r, r′ ≥ 0, if (b, r) 6= (b′, r′), then Ob,r ∩ Ob′,r′ = ∅. Moreover for any z ∈ Orb =
∪b∈Bγ ∪∞r=0 Ob,r, there exist a unique b ∈ Bγ and a unique integer r ≥ 0 such that z ∈ Ob,r.
Proof. On the cotrary, assume that Ob,r ∩ Ob′,r′ 6= ∅. Then there exists x ∈ Ob,r ∩ Ob′,r′.
Suppose that b = b′, then r 6= r′. We may assume that r′ > r. Since hr(x) = b = hr′
have hr′−r(b) = b. This contradicts to Assumption A (3).
Suppose that b 6= b′. Since hr(x) = b and hr′
(x) = b′, r 6= r′. We may assume that r′ > r.
(x) = b′. This contradicts to Assumption A (3).
(cid:3)
Then we have hr′−r(b) = hr′−r(hr(x)) = hr′
Therefore Ob,r ∩ Ob′,r′ = ∅. The rest is clear.
(x), we
19
Recall that we difine an endomorphism α on C(K) by (α(a))(x) = a(h(x)) for a ∈ A = C(K)
and x ∈ K. Then (αi(a))(x) = a(hi(x)).
Proposition 7.4. Let γ be a self-similar map on K with assumption A. Let τ be a finite
trace τ on F (∞). For b ∈ Bγ and integer r ≥ 0, put a constant cb,r ≥ 0 by the point mass
cb,r = µ0({xb,r}) for any xb,r ∈ Ob,r = h−r(b). Then τ −Pb∈Bγ P∞r=0 N rcb,rτ (b,r) is a positive
finite trace on F (∞) whose restriction to A does not have positive point mass.
Proof. We shall show that τ ≥ Pb∈BγP∞r=0 N rcb,rτ (b,r). Since cb,r = µ0({xb,r}) for any xb,r ∈
Ob,r = h−r(b), we note that
∞
Xr=0
Xb∈Bγ
N rcb,r = µ0(∪b∈Bγ ∪∞r=0 Ob,r) ≤ 1.
It is enough to show that for any finite positive integer s.
Since τ and Pb∈BγPs
r=0 τ (b,r) are finite traces, it is sufficient to show that
s
τ ≥ Xb∈Bγ
Xr=0
N rcb,rτ (b,r)
s
τ (T ) ≥ Xb∈Bγ
Xr=0
N rcb,rτ (b,r)(T )
m (x) ≤ 1.
m (hi−r(b)) = 1.
for T ∈ F (p)+
for each positive integer p with s + 1 ≤ p. We fix a positive integer p and integers
r ≥ 0 such that r + 1 ≤ p. For b ∈ Bγ and integers i with r ≤ i ≤ p, we can choose a sequences
{g(b,r,i)
m }m=1,2,... of elements in A = C(K) satisfying the following:
(1) 0 ≤ g(b,r,i)
(2) g(b,r,i)
(3) For each i, the diameter of the support of g(b,r,i)
(4) αr(g(b,r,r)
(5) For any m, if (b, r) 6= (b′, r′),then the supports of αr(g(b,r,r)
) are disjoint.
In fact, for b ∈ Bγ and an integer r ≥ 0, Ob,r = h−r(b) is a finite set. Moreover if (b, r) 6= (b′, r′),
then Ob,r ∩ Ob′,r′ = ∅. Hence there exist open neibourhoods Ub,r for b ∈ Bγ and 0 ≤ r ≤ p such
that h−r(Ub,r) ∩ h−r′
(Ub′,r′) = ∅ if (b, r) 6= (b′, r′). For b ∈ Bγ and integers i with r < i ≤ p,
it is easy to find a sequences {g(b,r,i)
}m=1,2,... of elements in A = C(K)
satisfying the conditions (1) -(3) above and the support of g(b,r,r)
is included in Ub,r. Then
define
approaches zero as m tends to ∞.
m }m=1,2,... and {g′(b,r,r)
m )(x) for each x ∈ K.
)(x) ≤ αi(g(b,r,i)
) and αr′(g(b′,r′,r′)
m
m
m
m
m
m
We put f (b,r)
m = αr(g(b,r,r)
m
p
g(b,r,i)
m (hi−ry)
m
(y)
g(b,r,r)
m
(y) = g′(b,r,r)
Yi=r+1
r=0 f (b,r)
τ (T ) ≥ τ (T 1/2QT 1/2) = τ (T Q) = Xb∈Bγ
). Let Q =Pb∈BγPs
20
m . Since 0 ≤ Q ≤ 1, we have
s
τ (T f (b,r)
m )
Xr=0
for T ∈ F (p)+. We shall show that
lim
m→∞
τi(T f (b,r)
m ) = N rcb,rτ (b,r)
i
(T )
Since τi and τ (b,r)
for T ∈ K(X⊗i) with 0 ≤ i ≤ p.
subset of K(X⊗i). We may assume that T = θξ,η for ξ, η ∈ X⊗i.
are finite trace and 0 ≤ f (b,r)
i
Firstly, we consider the case that 0 ≤ i ≤ r. Let ξ, η ∈ X⊗i. Then we have
m ≤ 1, it suffices to prove Lemma for a dense
τi(θξ,ηφi(αr(g(b,r,r)
m
τi(θξ,ηαi(αr−i(g(b,r,r)
m
)))
lim
m→∞
Hence
))) = lim
m→∞
= lim
m→∞
= lim
m→∞
= lim
m→∞
= lim
m→∞
τi(cid:16)θ
ξ,φi(αi(αr−i(g(b,r,r)
m
N iψi(τi)(φi(αi(αr−i(g(b,r,r)
)))ηξ)A)
N iψi(τi)((ηαr−i(g(b,r,r)
m
N iψi(τi)((αr−i(g(b,r,r)
m
m
)))η(cid:17)
)ξ)A)
)(ηξ)A)
N icb,r(ηξ)A(γj1 ◦ ··· ◦ jr−i(b))
= X(j1,...,jr−i)∈Σr−i
=cb,rN r(τ (b,r)
i
)(θξ,η).
τi(T f (b,r)
m ) = cb,rN rτ (b,r)
i
(T )
lim
m→∞
for T ∈ K(X⊗i) with 0 ≤ i ≤ r.
Secondly, we consider the case that r + 1 ≤ i ≤ p. Let ξ, η ∈ X⊗i. We have
lim
m→∞
τi(cid:16)θξ,ηφi(αi(g(b,r,i)
m ))(cid:17) = lim
m→∞
= lim
m→∞
= lim
m→∞
= lim
m→∞
m
ξ,φi(αi(g(b,r,i)
τi(cid:16)θ
))η(cid:17)
m ))ηξ)A(cid:17)
ψi(τi)(cid:16)φi(αi(g(b,r,i)
m ξ)A(cid:17)
ψi(τi)(cid:16)(ηg(b,r,i)
m (ηξ)A(cid:17) = 0,
ψi(τi)(cid:16)g(b,r,i)
because ψi(τi) does not have a positive point mass at hi−r(b) ∈ Cγ ∪ Pγ, by Propsition 7.1. Then
we have
for T ∈ K(X⊗i). We assume T ∈ K(X⊗i) be positive. Then we have
τi(T αi(g(b,r,i)
m )) = 0
lim
m→∞
τi(T f (b,r)
m ) =τi(T 1/2αr(g(b,r,r)
≤τi(T 1/2αi(g(b,r,i)
=τi(T αi(g(b,r,i)
m )).
)T 1/2)
m
m )T 1/2)
Since τi(T αi(g(b,r,i)
m )) approaches 0 as m tends to ∞, we have
τi(T f (b,r)
m ) = 0.
lim
m→∞
21
for T ≥ 0. Since each T ∈ K(X⊗i) is expressed as T = T1 − T2 + i(T3 − T4) with Ti ≥ 0, we have
for T ∈ K(X⊗i).
lim
m→∞
τi(T f (b,r)
m ) = 0 = cb,rN rτ (b,r)
i
(T )
(cid:3)
Nextly we shall analyze traces without point mass.
Lemma 7.5. Let τ be a finite trace on F (∞). If the restriction τ0of τ to A does not have positive
point mass, then τn = ψn(τK(X ⊗n)) also does not have positive point mass.
Proof. If ψn(τK(X ⊗n))) has a positive point mass,then τ0 also has a positive point mass at some
point by Lemma 5.10.
(cid:3)
It is a key point that we can replace F by G∗.
Lemma 7.6. Let σ be a finite trace of F (∞) and σn = σK(X ⊗n) the restriction of σ to K(X⊗n).
If the corresponding measure ψn(σn) does not have positive point mass for any n, then it holds
that for any n
Moreover µn = ψn(σn) is a provability measure and µn = G∗(µn+1) for n.
1
N
F (ψn+1(σn+1)) = G∗(ψn+1(σn+1)).
Proof. Since ψn+1(σn+1) has no positive point mass, it holds that
1
N
F (ψn+1(σn+1))(a) =
1
N
ψn+1(σn+1)(a) = ψn+1(σn+1)(G(a)) = G∗(ψn+1(σn+1))(a).
It holds that σ0(I) = µ0(1K ) = 1, and µ0 is a probability measure. By
µ0(1K ) = G∗(µ1)(1K ) = µ1(G(1K )) = µ1(1K ),
µ1 is also a probability measure. Inductively, µn is a probability measure for each n.
(cid:3)
We denote by µH the Hutchinson measure on K. It is shown that there exists a unique tracial
n = µH for each n. The tracial state τ (∞) gives a continuous
state τ (∞) on F (∞) such that τ (∞)
type (infinite type) KMS state on Oγ.
Proposition 7.7. Let γ be a self-similar map on K with assumption A. Let σ be a tracial state
on F (∞). If the restriction µ0 of σ to A does not have positive point mass, then σ coincides with
τ (∞).
Proof. Let σn = σK(X ⊗n) be the restriction of σ to K(X⊗n) and µn = ψn(σn) be the corre-
sponding measure. Since 1
N F (ψn+1(σn+1)) = ψn(σn),
we have G∗(µn+1) = µn (n = 0, 1, . . . ). Hence
N F (ψn+1(σn+1)) = G∗(ψn+1(σn+1)) and 1
µn ∈
∞
\k=n+1
∞
\k=1
G∗k(P (K)) =
G∗k(P (K))
Therefore µn = µH for any n. This implies that σ = τ (∞)
(cid:3)
We state the following Theorem on the classification of extreme tracial states on the core of
C∗-algebras constructed from self-similar maps.
22
Theorem 7.8. Let γ be a self-similar map on K with assumption A and Oγ the C∗-algebra
associated with γ. Then for any finite trace τ on F (∞), there exist unique constants cb,r ≥ 0 and
c∞ ≥ such that
τ = Xb∈Bγ
∞
Xr=0
N rcb,rτ (b,r) + c∞τ∞
Moreover, the set of extreme tracial states of the core F (∞) of Oγ is exactly
{τ (∞)}[{τ (b,r) b ∈ Bγ, r = 0, 1, 2, . . . }.
Proof. Let τ be a finite trace τ on F (∞). Define cb,r = µ0({xb,r}) for any xb,r ∈ Ob,r = h−r(b).
Then σ := τ −Pb∈Bγ P∞r=0 N rcb,rτ (b,r) is a positive finite trace on F (∞) whose restriction to A
does not have positive point mass by Proposition 7.4.
If σ = 0, then nothing is to be proved.
If σ 6= 0, put c∞ = σ(1) 6= 0. Then 1
σ is a tracial state on F (∞) and its restriction to A does
σ coincides with τ (∞) by Proposition 7.7. Hence we
c∞
not have positive point mass. Therefore 1
c∞
have
τ = Xb∈Bγ
τ = Xb∈Bγ
∞
Xr=0
∞
Xr=0
N rcb,rτ (b,r) + c∞τ∞
N rdb,rτ (b,r) + d∞τ∞
We shall show the uniqueness of the expression. Assume that there exist constants db,r ≥ 0 and
d∞ ≥ such that
Then db,r = µ0({xb,r}) = cb,r, because we know the point mass of the restrictions of τ (b,r) to
A = C(K) and the point mass of the restriction of τ∞ to A = C(K) is zero. Then c∞τ∞ =
d∞τ∞. This implies c∞ = d∞. Hence the expression is unique.
The uniqueness of the expression implies that the set of extreme tracial states of the core
F (∞) of Oγ is exactly
{τ (∞)}[{τ (b,r) b ∈ Bγ, r = 0, 1, 2, . . . }.
(cid:3)
We explain the construction of extreme traces by two typical examples.
Example 7.1. We describe the extreme traces on the core of the C∗-algebras associated with
the tent map explicitly. We define discrete measures µ(r)
i on [0, 1] for 0 ≤ i ≤ r by
µ(r)
i (f ) =
1
2r−i
X(j1,...,jr−i)∈{ 1,2 }i
f (γj1 ◦ ··· ◦ γjr−i(δ1/2)).
We define traces τ (r) on F (∞) for r ≥ 0 by Morita equivalence and
i )
r + 1 ≤ i
for each i. We can also define a trace τ (∞) on F (∞) by
τ (∞)
i = µ∞,
i =( πi(µ(r)
τ (r)
0
23
0 ≤ i ≤ r
for each i where µ∞ = µH is the normalized Lebesgue measure on [0, 1] and coincides with the
Hutchinson measure. The the set of extreme traces on the core of C∗-algebras associated with
the tent map on [0, 1] is exactly { τ (0), τ (1), . . . , τ (∞) }.
Example 7.2. We write down extreme traces on the core of the C∗-algebras associated the
self-similar map generating the Sierpinski gasket K . We define measures on K by
µ(S,r)
i
=
µ(T,r)
i
=
µ(U,r)
i
=
1
3r−i
1
3r−i
1
3r−i
X
X
X
(j1,...,jr−i)∈{ 1,2,3 }r−i
(j1,...,jr−i)∈{ 1,2,3 }r−i
(j1,...,jr−i)∈{ 1,2,3 }r−i
f (γj1 ◦ ··· ◦ γjr−i(δS ))
f (γj1 ◦ ··· ◦ γjr−i(δT ))
f (γj1 ◦ ··· ◦ γjr−i(δU ))
Then we define traces τ (S,r), τ (T,r) and τ (U,r) for r = 0, 1, 2, . . . using µ(S,r)
by Morita equivalence. We can also define a trace τ (∞) on F (∞) by
i
τ (∞)
i = µH,
s, µ(T,r)
i
s and µ(U,r)
i
s
for each i where µH is the Hutchinson measure on the Sierpinski Gasket K.
Finally, we describe a connection of the extreme traces of the core and the KMS states of Oγ.
A general characterization of KMS states is given by Laca-Neshveyev [17]. Let β > log N and
we put
ρ(b,β) = Cb,β
where Cb,β is a normalizing constant.
∞
eβ(cid:19)j
Xj=0(cid:18) N
τ (b,j),
Proposition 7.9. Let γ be a self-similar map on K with assumption A. Then τ (∞) extends to
a log N -KMS state on Oγ and for β > log N , ρ(b,β) extends to a β-KMS state of Oγ.
Proof. This follows from the fact that the β-KMS state of Oγ is described concretely in section
6 of our paper [8] with Izumi.
(cid:3)
References
[1] B. Blackadar, Operator algebras, Encyclopedia of Math. Sci., 122, Springer, 2006
[2] G. Castro, C ∗-algebras associated with iterated function systems, Contemporary Math.503 (2010), 27-38,
Operator structures and dynamical systems.
[3] J. Cuntz and W. Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56 (1980),
251-268.
[4] R. Exel R. and M. Laca, Partial dynamical systems and the KMS condition, Commun. Math. Phys., 232
(2003), 223 -- 277.
[5] K. J. Falconer, Fractal Geometry, Wiley, 1997.
[6] M. Frank and D. Larson, Frames in Hilbert C ∗-modules and C ∗-algebras, J. Operator Theory 48 (2002),
273-314.
[7] J. E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30 (1981), 713 -- 747.
[8] M. Izumi, T. Kajiwara and Y. Watatani, KMS states and branched points, Ergodic Theory Dynam. Systems
27 (2007), 1887 -- 1918.
24
[9] M. Ionescu and Y. Watatani, C ∗-algebras associated with Mauldin-Williams graphs, Canad. Math. Bull. 51
(2008), 545-560.
[10] T. Kajiwara, Countable bases for Hilbert C ∗-modules and classification of KMS states, Contemporary
Math.503 (2010), 73-91, Operator structures and dynamical systems.
[11] J. Kigami, Analysis on Fractals , 2001, Cambridge University Press.
[12] T. Kajiwara, C. Pinzari and Y. Watatani, Jones index theory for Hilbert C∗ -- bimodules and its equivalence
with conjugation theory, J. Funct. Anal. 215 (2004), 1-49.
[13] T. Kajiwara, C. Pinzari and Y. Watatani, Ideal structure and simplicity of the C ∗ -- algebras generated by
Hilbert bimodules J. Funct. Anal. 159 (1998), 295-322
[14] T. Kajiwara and Y. Watatani, C*-algebras associated with self-similar sets, J. Operator Theory, 56(2006),
225-247.
[15] T. Kajiwara and Y. Watatani, KMS states on finite-graph C*-algebras, to appear in Kyushu Journal of
Mathematics.
[16] A. Kumjian and J. Renalut, KMS states on C∗-algebras associated to expansive maps, Proc. Amer. Math.
Soc. 134 (2006), 2067-2078.
[17] M. Laca and S. Neshveyev, KMS states of quasi-free dynamics on Pimsner algebras, J. Funct. Anal. 211
(2004) 457 -- 482.
[18] K. Matsumoto, K-theory for C ∗-algebras associated with subshifts, Math. Scand. 82 (1998), 237-255.
[19] N. Nawata, Fundamental group of simple C*-algebras with unique trace III, to appear in Canad. J. Math.
[20] M. Pimsner, A class of C ∗-algebras generating both Cuntz-Krieger algebras and crossed product by Z, Free
probability theory, AMS, (1997), 189 -- 212.
(Tsuyoshi Kajiwara) Department of Environmental and Mathematical Sciences, Okayama Univer-
sity, Tsushima, 700-8530, Japan
(Yasuo Watatani) Department of Mathematical Sciences, Kyushu University, Motooka, Fukuoka,
819-0395, Japan
25
|
1501.05944 | 2 | 1501 | 2015-06-01T16:07:09 | Martingale inequalities in noncommutative symmetric spaces | [
"math.OA",
"math.FA"
] | We provide generalizations of Burkholder's inequalities involving conditioned square functions of martingales to the general context of martingales in noncommutative symmetric spaces. More precisely, we prove that Burkholder's inequalities are valid for any martingale in noncommutative space constructed from a symmetric space defined on the interval $(0,\infty)$ with Fatou property and whose Boyd indices are strictly between 1 and 2. This answers positively a question raised by Jiao and may be viewed as a conditioned version of similar inequalities for square functions of noncommutative martingales. Using duality, we also recover the previously known case where the Boyd indices are finite and are strictly larger than 2. | math.OA | math |
MARTINGALE INEQUALITIES IN NONCOMMUTATIVE SYMMETRIC
SPACES
NARCISSE RANDRIANANTOANINA AND LIAN WU
Abstract. We provide generalizations of Burkholder's inequalities involving conditioned square
functions of martingales to the general context of martingales in noncommutative symmetric
spaces. More precisely, we prove that Burkholder's inequalities are valid for any martingale in
noncommutative space constructed from a symmetric space defined on the interval (0, ∞) with
Fatou property and whose Boyd indices are strictly between 1 and 2. This answers positively
a question raised by Jiao and may be viewed as a conditioned version of similar inequalities for
square functions of noncommutative martingales. Using duality, we also recover the previously
known case where the Boyd indices are finite and are strictly larger than 2.
1. Introduction
In classical martingale theory, a fundamental result due to Burkholder ([8, 9, 20]) can be
described as follows: given a probability space (Ω, Σ, P ), let {Σn}n≥1 be an increasing sequence
of σ-fields of Σ such that Σ =W Σn . If 2 ≤ p < ∞ and f = (fn)n≥1 is a Lp-bounded martingale
adapted to the filtration {Σn}n≥1, then (using the convention that Σ0 = Σ1),
where ≃p means equivalence of norms up to constants depending only on p. The random variable
E[dfn2Σn−1](cid:1)p/2i1/p
[Efnp]1/p ≃p hE(cid:0)Xn≥1
E[dfn2Σn−1](cid:1)1/2 is called the conditioned square function of the martingale f and
Edfnpi1/p
+hXn≥1
,
(1.1)
sup
n≥1
s(f ) =(cid:0)Pn≥1
the equivalence (1.1) is generally referred to as Burkholder's inequalities. The equivalence (1.1)
was established by Burkholder as the martingale difference sequence generalizations of Rosenthal's
inequalities [44] which state that if 2 ≤ p < ∞ and (gn)n≥1 is a sequence of independent mean-zero
random variables in Lp(Ω, Σ, P ) then
(1.2)
(cid:16)E(cid:12)(cid:12)Xn≥1
gn(cid:12)(cid:12)p(cid:17)1/p
≃p (cid:16)Xn≥1
Egn2(cid:17)1/2
+(cid:16)Xn≥1
Egnp(cid:17)1/p
.
Probabilistic inequalities involving independent random variables and martingales inequalities
play important roles in many different areas of mathematics. Burkholder/Rosenthal inequalities
in particular have many applications in probability theory and structures of symmetric spaces
in Banach space theory. On the other hand, a recent trend in the general study of martingale
inequalities is to find analogues of classical inequalities in the context of noncommutative Lp-
spaces. We refer to [39, 26, 28, 41] for additional information on noncommutative martingale
inequalities. Noncommutative analogues of (1.1) and (1.2) were extensively studied by Junge
2010 Mathematics Subject Classification. Primary: 46L53, 46L52. Secondary: 47L05, 60G42.
Key words and phrases. Noncommutative martingales, martingale inequalities, noncommutative symmetric
spaces, Boyd indices, interpolations.
Wu was partially supported by NSFC(No.11471337) and China Scholarship Council.
1
2
N. RANDRIANANTOANINA AND L. WU
and Xu in [29, 30]. They obtained that if 2 ≤ p < ∞ and x = (xn)n≥1 is a noncommutative
martingale that is Lp-bounded then
(cid:13)(cid:13)x(cid:13)(cid:13)p ≃p maxn(cid:13)(cid:13)sc(x)(cid:13)(cid:13)p,(cid:13)(cid:13)sr(x)(cid:13)(cid:13)p,(cid:0)Xn≥1(cid:13)(cid:13)dxn(cid:13)(cid:13)p
p(cid:1)1/po
(1.3)
(1.4)
(1.5)
(1.6)
where sc(x) and sr(x) denote the column version and the row version of conditioned square
functions which we refer to the next section for formal definitions. Moreover, they also treated
the corresponding inequalities for the range 1 < p < 2 which are dual versions of (1.3) and read
as follows: if x = (xn)n≥1 is a noncommutative martingale in L2(M) then
(cid:13)(cid:13)x(cid:13)(cid:13)p ≃p infn(cid:13)(cid:13)sc(y)(cid:13)(cid:13)p +(cid:13)(cid:13)sr(z)(cid:13)(cid:13)p +(cid:0)Xn≥1(cid:13)(cid:13)dwn(cid:13)(cid:13)p
p(cid:1)1/po
where the infimum is taken over all x = y + z + w with y, z, and w are martingales. The
differences between the two cases 1 < p < 2 and 2 ≤ p < ∞ are now well-understood in the field.
In [21], inequalities (1.3) and (1.4) were extended to the case of noncommutative Lorentz spaces
Lp,q(M) for 1 < p < ∞ and 1 ≤ q < ∞. Motivated by this extension, it is natural to ask if
some versions of noncommutative Burkholder's inequalities remain valid in the general context
of noncommutative symmetric spaces. This question was explicitly raised in [22, Problem 3.5].
Martingale inequalities in the general framework of rearrangement invariant spaces have long
been of interests. For the case of classical martingales, we refer reader to the work of Johnson
and Schechtman [24, 25] and the references therein. For the noncommutative settings, we recall
that generalizations of Burkholder-Gundy inequalities in noncommutative symmetric spaces were
recently established in [14, 22], extensions of Junge's noncommutative Doob maximal inequalities
in some symmetric spaces were treated in [13]. In a closely related topic, Le Merdy and Sukochev
studied Rademacher averages on noncommutative symmetric spaces ([33]). These Rademacher
averages turn out to provide one of the key ingredients in the solution of Burkholder-Gundy
inequalities in noncommutative symmetric spaces in [14, 22]. Naturally, the concept of Boyd
indices of symmetric spaces ([34]) and various interpolation techniques play significant roles in
all the results stated above.
The present paper solves the problem discussed above. Our main result can be summarized as
follows: assume that E is a rearrangement invariant function space on (0, ∞) that satisfies some
natural conditions and has nontrivial Boyd indices 1 < pE ≤ qE < ∞ and M is a semifinite von
Neumann algebra equipped with a faithful normal semifinite trace τ . We obtain generalizations
of (1.3) and (1.4) that read:
If 1 < pE ≤ qE < 2, then
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≃E infn(cid:13)(cid:13)sc(y)(cid:13)(cid:13)E(M) +(cid:13)(cid:13)sr(z)(cid:13)(cid:13)E(M) +(cid:13)(cid:13)(dwn)n≥1(cid:13)(cid:13)E(M⊗ℓ∞)o
where as in (1.4), the infimum is taken over all decompositions x = y + z + w with y, z, and w
are martingales in E(M, τ ).
If 2 < pE ≤ qE < ∞, then
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≃E maxn(cid:13)(cid:13)sc(x)(cid:13)(cid:13)E(M),(cid:13)(cid:13)sr(x)(cid:13)(cid:13)E(M),(cid:13)(cid:13)(dxn)n≥1(cid:13)(cid:13)E(M⊗ℓ∞)o.
We note that (1.6) was recently established by Dirksen in [12]. His approach follows closely
the original arguments used in [29] taking advantage of the fact mentioned earlier that the non-
commutative Burkholder-Gundy inequalities for square functions are valid for noncommutative
martingales in some general symmetric spaces. Thus, our main motivation is primarily to establish
the equivalence (1.5).
MARTINGALE INEQUALITIES
3
Our approach is based on another discovery made in the next section that, in some sense, one
inequality in the equivalence (1.4) can be achieved with a decomposition that works simultane-
ously for all 1 < p < 2. We refer to Theorem 2.10 below for more information. This simultaneous
decomposition allows us to efficiently apply results from interpolation theory. Namely, we use
concrete realization of noncommutative symmetric spaces as interpolations of noncommutative
Lp-spaces by means of K-functionals and J-functionals. The non-trivial inequality in the equiv-
alence (1.6) will be deduced from (1.5) using duality. Unlike the Lp-cases, this duality technique
does not seem to apply for the other direction. That is, at the time of this writing, we lack
necessary ingredients to deduce (1.5) from (1.6).
The paper is organized as follows: in Section 2, we provide some preliminary results concerning
noncommutative symmetric spaces, interpolation theory, and martingale inequalities. In partic-
ular, we establish a decomposition result for noncommutative martingales that sets up the use
of interpolations. Section 3 is devoted entirely to the statement and proof of our main result.
In the last section, we discuss some related results, provide examples, and point to related open
questions concerning Burkholder's inequalities.
Our notation and terminology are standard as may be found in the books [6, 34, 45].
2. Definitions and preliminary results
2.1. Noncommutative spaces. In this subsection, we review some basic facts on rearrangement
invariant spaces and their noncommutative counterparts that are relevant for our presentation.
For a semifinite von Neumann algebra M equipped with a faithful normal semifinite trace τ ,
let fM denote the topological ∗-algebra of all measurable operators with respect to (M, τ ) in the
sense of [36]. For x ∈ fM, define its generalized singular number by
µt(x) = inf{λ > 0; τ (ex(λ, ∞)) ≤ t},
t > 0
where ex is the spectral measure of x. The function t 7→ µt(x) from (0, ∞) into [0, ∞) is right-
continuous and nonincreasing ([19]). For the case where M is the abelian von Neumann algebra
L∞(0, ∞) with the trace given by integration with respect to the Lebesgue measure, fM becomes
the linear space of all measurable functions L0(0, ∞) and µ(f ) is the decreasing rearrangement
of the function f in the sense of [34].
We recall that a Banach function space (E, k · kE) on (0, ∞) is called symmetric if for any
g ∈ E and any measurable function f with µ(f ) ≤ µ(g), we have f ∈ E and kf kE ≤ kgkE . The
Kothe dual of a symmetric space E is the function space defined by setting:
E× =nf ∈ L0(0, ∞) :Z ∞
When equipped with the norm kf kE× := sup{R ∞
Banach function space.
0
f (t)g(t) dt < ∞, ∀g ∈ Eo.
f (t)g(t) dt : kgkE ≤ 1}, E× is a symmetric
0
The symmetric Banach function space E is said to have the Fatou property if, whenever 0 ≤
fα ↑α⊆ E is an upwards directed net with supα kfαkE < ∞, it follows that f = supα fα exists
in E and kf kE = supα kfαkE.
It is well-known that E has the Fatou property if and only if
the natural embedding of E into its Kothe bidual E×× is a surjective isometry. Examples of
symmetric spaces with the Fatou property are separable symmetric spaces and duals of separable
symmetric spaces.
Another concept that is central to the paper is the notion of Boyd indices which we now
introduce. Let E be a symmetric Banach space on (0, ∞). For s > 0, the dilation operator
4
N. RANDRIANANTOANINA AND L. WU
Ds : E → E is defined by setting
The lower and upper Boyd indices of E are defined by
Dsf (t) = f (t/s),
t > 0,
f ∈ E.
pE := lim
s→∞
log s
log kD1/sk
and qE := lim
s→0+
log s
log kD1/sk
,
It is well-known that 1 ≤ pE ≤ qE ≤ ∞ and if E = Lp for 1 ≤ p ≤ ∞ then
respectively.
pE = qE = p. We shall say that E has non-trivial Boyd indices whenever 1 < pE ≤ qE < ∞. We
refer to [6, 34] for any unexplained terminology from function space theory.
For a given symmetric Banach function space (E, k · kE) on the interval (0, ∞), we define the
corresponding noncommutative space by setting:
E(M, τ ) =(cid:8)x ∈ fM : µ(x) ∈ E(cid:9).
Equipped with the norm kxkE(M,τ ) := kµ(x)kE , the space E(M, τ ) is a complex Banach space
([32]) and is referred to as the noncommutative symmetric space associated with (M, τ ) corre-
sponding to the function space (E, k · kE). We remark that if 1 ≤ p < ∞ and E = Lp(0, ∞) then
E(M, τ ) = Lp(M, τ ) is the usual noncommutative Lp-space associated with (M, τ ).
Recall that a linear operator T : X → Y is called a semi-embedding if T is one to one and
T (BX) is a closed subset of Y where BX = {x ∈ X : kxk ≤ 1}. As in the commutative case, if
1 ≤ p < pE ≤ qE < q ≤ ∞ then the space E(M, τ ) is intermediate to the spaces Lp(M, τ ) and
Lq(M, τ ) in the sense that
Lp(M, τ ) ∩ Lq(M, τ ) ⊆ E(M, τ ) ⊆ Lp(M, τ ) + Lq(M, τ )
with the inclusion maps being continuous. Moreover,
if E satisfies the Fatou property, one
can readily verify that the second inclusion map E(M, τ ) ֒→ Lp(M, τ ) + Lq(M, τ ) is a semi-
embedding. These facts will be used in the sequel.
We end this subsection with the following elementary lemma. It will be used in the proof of
our main result. We include a proof for completeness.
Lemma 2.1. Assume that 1 < p < q < 2 and let u ∈ Lp(M) ∩ Lq(M). There exists a se-
quence (um)m≥1 in L1(M) ∩ L2(M) with limm→∞ kum − ukLp(M)∩Lq(M) = 0 and both sequences
(kumkp)m≥1 and (kumkq)m≥1 are increasing and converge to kukp and kukq, respectively.
Proof. Let eu denote the spectral measure of u. For each m ≥ 1, set em := eu([1/m, m]). Then
(em)m is an increasing sequence of projections in M that converges to 1 for the strong operator
topology. Clearly, for every m ≥ 1, τ (em) ≤ mpkukp
p < ∞. Set um := uem. It is easy to check
that um ∈ L1(M) ∩ M. A fortiori, um ∈ L1(M) ∩ L2(M). From the identities
max(cid:8)ku − umkp,ku − umkq(cid:9) = max(cid:8)ku(1 − em)kp, ku(1 − em)kq(cid:9)
= max(cid:8)k(1 − em)u2(1 − em)k1/2
q/2(cid:9),
p/2, k(1 − em)u2(1 − em)k1/2
we get that limn→∞ ku − umkLp(M)∩Lq (M) = 0. On the other hand, if s is equal to either p or q,
it follows from the identity kumks = kuemuk1/2
s/2 that (kumks)m forms an increasing sequence
that converges to kuks.
(cid:3)
We refer to [10, 16, 40, 46] for extensive discussions on various properties of noncommutative
symmetric spaces.
MARTINGALE INEQUALITIES
5
2.2. Function spaces and interpolations. In this subsection, we will discuss concrete de-
scription of certain classes of noncommutative symmetric spaces as interpolations of noncom-
mutative Lp-spaces that are relevant for our method of proof in the next section. We begin
by recalling that for a given compatible Banach couple (X0, X1), a Banach space Z is called
an interpolation space if X0 ∩ X1 ⊆ Z ⊆ X0 + X1 and whenever a bounded linear operator
T : X0 + X1 → X0 + X1 is such that T (X0) ⊆ X0 and T (X1) ⊆ X1 we have T (Z) ⊆ Z and
kT : Z → Zk ≤ C max{kT : X0 → X0k, kT : X1 → X1k} for some constant C. In this case, we
write Z ∈ Int(X0, X1). When C = 1, Z is called exact interpolation space. We refer to [6, 7, 31]
for more on interpolations.
In this paper we rely heavily on the notions of K-functionals and J-functionals which we now
review:
For a compatible Banach couple (X0, X1), we define the J-functional by setting for any x ∈
X0 ∩ X1 and t > 0,
As a dual notion, the K-functional is defined by setting for any x ∈ E0 + E1 and t > 0,
J(x, t; X0, X1) = max(cid:8)kxkX0, tkxkX1(cid:9).
K(x, t; X0, X1) = inf(cid:8)kx1kX0 + tkx2kX1; x = x1 + x2(cid:9).
If the compatible couple (X0, X1) is clear from the context, then we will simply write J(x, t) and
K(x, t) in place of J(x, t; X0, X1) and K(x, t; X0, X1), respectively. It is now quite well-known
that any symmetric Banach function space with the Fatou property that belongs to Int(Lp, Lq) is
given by a K-method. More precisely, we have the following result due to Brudnyi and Krugliak
(see for instance [31, Theorem 6.3]).
Theorem 2.2. Let E be a symmetric Banach function space on (0, ∞) with the Fatou property.
If E ∈ Int(Lp(0, ∞), Lq(0, ∞)) for 1 ≤ p < q ≤ ∞, then there exists a function space F on (0, ∞)
such that f ∈ E if and only if K(f, ·, Lp, Lq) ∈ F and there exists a constant C such that
C −1(cid:13)(cid:13)K(f, ·)(cid:13)(cid:13)F ≤ kf kE ≤ C(cid:13)(cid:13)K(f, ·)(cid:13)(cid:13)F .
We will use the corresponding J-method of the above theorem. This was studied in [4, 5].
We review the basic construction of this method and introduce a discrete version that is quite
essential in the next section.
Suppose that an element x ∈ X0 + X1 admits a representation
where u(·) is measurable function that takes its values in X0 ∩ X1 and the integral is convergent
in X0 + X1. For any given representation u(·), we set for s > 0,
Given a symmetric Banach function space F defined on (0, ∞), the interpolation space (X0, X1)F,j
consists of elements x ∈ X0 + X1 which admit a representation as in (2.1) and are such that
where the infimum is taken over all representation u of x as in (2.1). We refer to [4, 5] for a
comprehensive study of this interpolation method along with some other equivalent methods. As
noted above, we will need a discrete version of this method. This is standard but we could not
find any reference in the literature for this particular method so we provide the details.
(2.1)
(2.2)
(2.3)
x =Z ∞
0
u(t) dt/t,
j(u, s) =Z ∞
s
t−1J(u(t), t) dt/t.
(cid:13)(cid:13)x(cid:13)(cid:13)F,j = infn(cid:13)(cid:13)j(u, ·)(cid:13)(cid:13)Fo < ∞,
6
N. RANDRIANANTOANINA AND L. WU
We define the interpolation space (X0, X1)F,j to be the space of elements x ∈ X0 + X1 which
admit a representation
(2.4)
with uν ∈ X0 ∩ X1 and are such that
uν
(convergence in X0 + X1)
x =Xν∈Z
(cid:13)(cid:13)x(cid:13)(cid:13)F,j = infn(cid:13)(cid:13)(cid:13)j({uν }ν, ·)(cid:13)(cid:13)(cid:13)Fo < ∞,
where the decreasing function j({uν }ν , ·) is defined by
j({uν }ν, t) = Xγ≥ν+1
2−γJ(uγ, 2γ)
for t ∈ [2ν , 2ν+1),
and the infimum is taken over all representations of x as in (2.4). Clearly, the function j({uν }ν , t)
takes only countably many values. Thus, we may call (X0, X1)F,j as a discrete interpolation
method. As in the case of real interpolation methods, this discrete version is equivalent to the
continuous version described earlier. More precisely, we have:
Lemma 2.3. Let x ∈ X0 + X1. Then x ∈ (X0, X1)F,j if and only if x ∈ (X0, X1)F,j. More
precisely, the following inequalities hold:
The verification of Lemma 2.3 is a simple adaptation of standard arguments from interpolation
1
4(cid:13)(cid:13)x(cid:13)(cid:13)F,j ≤(cid:13)(cid:13)x(cid:13)(cid:13)F,j ≤ 4(cid:13)(cid:13)x(cid:13)(cid:13)F,j.
theory which we leave for the reader (see [7]).
Combining [4, Theorem 9.3], [5, Theorem 3.5], and Lemma 2.3, we may state the following
result which is one of the decisive tools we use in our proof.
Theorem 2.4. If E is a symmetric Banach function space on (0, ∞) with the Fatou property
then the following are equivalent:
(i) 1 < p < pE ≤ qE < q < ∞.
(ii) There exists a symmetric Banach function space F on (0, ∞) with nontrivial Boyd indices
such that:
E = (Lp(0, ∞), Lq(0, ∞))F,j
(with equivalent norms).
As is now well-known, the preceding interpolation result automatically lifts to the noncommu-
tative setting (see [40, Corollary 2.2]):
Corollary 2.5. Let E be a symmetric Banach function space on (0, ∞) with the Fatou property.
Then the following are equivalent:
(i) 1 < p < pE ≤ qE < q < ∞.
(ii) There exists a symmetric Banach function space F on (0, ∞) with nontrivial Boyd indices
such that for every semifinite von Neumann algebra (N , σ),
E(N , σ) = (Lp(N , σ), Lq(N , σ))F,j,
with equivalent norms depending only on E, p, and q.
We record a general fact about interpolations of linear operators between two noncommutative
spaces for further use.
MARTINGALE INEQUALITIES
7
Proposition 2.6 ([17]). Assume that E ∈ Int(Lp, Lq) and M and N are semifinite von Neumann
algebras. Let T : Lp(M) + Lq(M) → Lp(N ) + Lq(N ) be a linear operator such that T : Lp(M) →
Lp(N ) and T : Lq(M) → Lq(N ) are bounded. Then T maps E(M) into E(N ) and the resulting
operator T : E(M) → E(N ) is bounded. Moreover, we have the following estimate:
(cid:13)(cid:13)T : E(M) → E(N )(cid:13)(cid:13) ≤ C max(cid:8)(cid:13)(cid:13)T : Lp(M) → Lp(N )(cid:13)(cid:13),(cid:13)(cid:13)T : Lq(M) → Lq(N )(cid:13)(cid:13)(cid:9)
for some absolute constant C.
2.3. Noncommutative martingales. Let us now recall the general setup for noncommutative
In the sequel, we always assume that the von Neumann algebra M is such that
martingales.
M∗ is separable. Denote by (Mn)n≥1 an increasing sequence of von Neumann subalgebras of M
whose union is weak*-dense in M. For n ≥ 1, we assume that there exists a trace preserving
conditional expectation En from M onto Mn. It is well-known that if τn denotes the restriction
of τ on Mn, then En extends to a contractive projection from Lp(M, τ ) onto Lp(Mn, τn) for all
1 ≤ p ≤ ∞. More generally, if E is a symmetric Banach function space on (0, ∞) which is an
interpolation space of the couple (L1(0, ∞), L∞(0, ∞)) then En is bounded from E(M, τ ) onto
E(Mn, τn).
Definition 2.7. A sequence x = (xn)n≥1 in L1(M) is called a noncommutative martingale with
respect to (Mn)n≥1 if En(xn+1) = xn for every n ≥ 1.
If in addition, all xn's belong to E(M) then x is called an E(M)-martingale. In this case we
set
kxkE(M) = sup
n≥1
kxnkE(M).
If kxkE(M) < ∞, then x is called a bounded E(M)-martingale.
Let x = (xn)n≥1 be a noncommutative martingale with respect to (Mn)n≥1. Define dxn =
xn − xn−1 for n ≥ 1 with the usual convention that x0 = 0. The sequence dx = (dxn)n≥1 is
called the martingale difference sequence of x. A martingale x is called a finite martingale if
there exists N such that dxn = 0 for all n ≥ N. In the sequel, for any operator x ∈ E(M),
we denote xn = En(x) for n ≥ 1.
It is well-known that if 1 < p < ∞ and x = (xn)n≥1 is
a bounded Lp(M)-martingale then there exists x∞ ∈ Lp(M) such that xn = En(x∞) (for all
n ≥ 1) and kxkp = kx∞kp. For the general context of symmetric spaces, if E ∈ Int(Lp, Lq) for
1 < p ≤ q < ∞ and satisfies the Fatou property, then any bounded E(M)-martingale x = (xn)n≥1
is of the form (En(x∞))n≥1 where x∞ ∈ E(M) satisfying kxkE(M) ≃E kx∞kE(M), with equality
if E is an exact interpolation space. Indeed, by reflexivity, such statement can be readily verified
for Lp(M) + Lq(M) and for general such E, it follows from the fact that E(M) semi-embeds into
Lp(M) + Lq(M). Because of these facts, we often identify martingales with measurable operators
when appropriate.
Let us now review the definitions of Hardy spaces and conditioned Hardy spaces of noncommu-
tative martingales. Throughout, E is a symmetric function space on (0, ∞) with the additional
properties that it satisfies the Fatou property and E ∈ Int(L1, L∞).
We define the column space E(M; ℓc
2) to be the linear space of all sequences (an)n≥1 ⊂ E(M)
such that the infinite column matrix Pn≥1 an ⊗ en,1 ∈ E(M⊗B(ℓ2(N))) where B(ℓ2(N)) is the
the linear subspace {Pn≥1 an ⊗ en,1 : (an)n ∈ E(M; ℓc
algebra of all bounded operators on the Hilbert space ℓ2(N) equipped with its usual trace tr and
(en,m)n,m≥1 denotes the collection of all unit matrices in B(ℓ2(N)). It is not difficult to verify that
2)} is closed in E(M⊗B(ℓ2))and therefore
8
N. RANDRIANANTOANINA AND L. WU
we equip E(M; ℓc
2) with the norm
(cid:13)(cid:13)(cid:13)(an)n≥1(cid:13)(cid:13)(cid:13)E(M;ℓc
2)
:=(cid:13)(cid:13)(cid:13)Xn≥1
an ⊗ en,1(cid:13)(cid:13)(cid:13)E(M⊗B(ℓ2))
=(cid:13)(cid:13)(cid:13)(cid:0)Xn≥1
an2(cid:1)1/2(cid:13)(cid:13)(cid:13)E(M)
,
then it becomes a Banach space. We can also define the corresponding row space E(M; ℓr
be the space of all sequences (an)n≥1 ⊂ E(M) for which (a∗
norm
2) to
2) equipped with the
n)n≥1 ∈ E(M; ℓc
Following [39], we consider the column and row versions of square functions relative to a
martingale x = (xn)n≥1 as follows:
2)
(cid:13)(cid:13)(cid:13)(an)n≥1(cid:13)(cid:13)(cid:13)E(M;ℓr
dxk2(cid:17)1/2
n)n≥1(cid:13)(cid:13)(cid:13)E(M;ℓc
=(cid:13)(cid:13)(cid:13)(a∗
and Sr(x) =(cid:16)Xk≥1
.
2)
Sc(x) =(cid:16)Xk≥1
dx∗
k2(cid:17)1/2
,
E
= k(dxn)n≥1kE(M;ℓc
where convergences can be taken with respect to the measure topology. Define Hc
E(M) (respec-
tively, Hr
E(M)) to be the set of all martingales x = (xn)n≥1 in E(M) for which the martin-
gale difference sequence (dxn)n≥1 ∈ E(M; ℓc
2)) equipped with the norm
kxkHc
2)). It is easy to verify that
since conditional expectations are bounded in E(M), the normed spaces Hc
E(M)
embed isometrically into E(M⊗B(ℓ2(N))) with closed ranges and therefore they are Banach
spaces. Moreover, one can see from its definition that Hc
E(M) is simply the space of all mar-
= kSc(x)kE(M) < ∞. Similar statement is also
tingales x = (xn)n≥1 in E(M) for which kxkHc
valid for Hr
2) (respectively, E(M; ℓr
= k(dxn)n≥1kE(M;ℓr
2) (respectively, kxkHr
E(M) and Hr
E
E
E(M).
We now turn to the mixture Hardy spaces of noncommutative martingales. From the above
discussions, the spaces Hc
E(M) are compatible in the sense that they continuously
embed into the larger space E(M⊗B(ℓ2(N))). The Hardy space HE(M) is defined as follows.
For 1 ≤ pE ≤ qE < 2,
E(M) and Hr
HE(M) = Hc
E(M) + Hr
E(M)
equipped with the norm
where the infimum is taken over all y ∈ Hc
2 ≤ pE ≤ qE < ∞,
E(M) and z ∈ Hr
E(M) such that x = y + z. For
kxkHE = inf(cid:8)kykHc
E
+ kzkHr
E(cid:9),
HE(M) = Hc
E(M) ∩ Hr
E(M)
kxkHE = max(cid:8)kxkHc
E
, kxkHr
E(cid:9).
equipped with the norm
These definitions mirror the well-documented difference between the two cases 1 ≤ p < 2 and
2 ≤ p < ∞ for the special case where E = Lp(0, ∞). We refer to [14] and [22] for more information
and results related to space HE(M).
We now consider the conditioned versions of the above definitions. Our approach is based on
the conditioned spaces introduced by Junge in [26]. Since this is very crucial in the sequel, we
review the basic setup. Below, we use the convention that E0 = E1.
Let E : M → N be a normal faithful conditional expectation, where N is a von Neumann
p(M, E) to be the completion
subalgebra of M. For 0 < p ≤ ∞, we define the conditioned space Lc
of M ∩ Lp(M) with respect to the quasi-norm
(cid:13)(cid:13)x(cid:13)(cid:13)Lc
p(M,E) =(cid:13)(cid:13)E(x∗x)(cid:13)(cid:13)1/2
p/2.
MARTINGALE INEQUALITIES
9
It was shown in [26] that for every n and 0 < p ≤ ∞, there exists an isometric right Mn-module
map un,p : Lc
p(M, En) → Lp(Mn; ℓc
2) such that
un,p(x)∗un,q(y) = En(x∗y) ⊗ e1,1,
(2.5)
for all x ∈ Lc
q(M; En) with 1/p + 1/q ≤ 1. We now consider the increasing
sequence of expectations (En)n≥1. Denote by F the collection of all finite sequences (an)n≥1 in
L1(M) ∩ M. For 0 < p ≤ ∞, define the space Lcond
2) to be the completion of F with
respect to the norm:
p(M; En) and y ∈ Lc
(M; ℓc
p
(2.6)
(cid:13)(cid:13)(an)(cid:13)(cid:13)Lcond
p
(M;ℓc
2) =(cid:13)(cid:13)(cid:0)Xn≥1
En−1an2(cid:1)1/2(cid:13)(cid:13)p.
The space Lcond
von Neumann algebra by means of the following map:
(M; ℓc
p
2) can be isometrically embedded into an Lp-space associated to a semifinite
defined by setting
Up : Lcond
p
(M; ℓc
2) → Lp(M⊗B(ℓ2(N2)))
Up((an)n≥1) =Xn≥1
un−1,p(an) ⊗ en,1
where as above, (ei,j)i,j≥1 is the family of unit matrices in B(ℓ2(N)). From (2.5), it follows that
if (an)n≥1 ∈ Lcond
2) for 1/p + 1/q ≤ 1 then
2) and (bn)n≥1 ∈ Lcond
(M; ℓc
(M; ℓc
p
q
(2.7)
Up((an))∗Uq((bn)) =(cid:0)Xn≥1
En−1(a∗
nbn)(cid:1) ⊗ e1,1 ⊗ e1,1.
In particular, k(an)kLcond
2) = kUp((an))kp and hence Up is indeed an isometry. We note that
Up is independent of p in the sense of interpolation. Below, we will simply write U for Up. We
refer the reader to [26] and [28] for more details on the preceding construction.
(M;ℓc
p
Now, we generalize the notion of conditioned spaces to the setting of symmetric spaces. We
consider the algebraic linear map U restricted to the linear space F that takes its values in
L1(M⊗B(ℓ2(N2))) ∩ M⊗B(ℓ2(N2)). For a given sequence (an)n≥1 ∈ F, we set:
(cid:13)(cid:13)(an)(cid:13)(cid:13)Econd(M;ℓc
2) =(cid:13)(cid:13)(cid:0)Xn≥1
En−1an2(cid:1)1/2(cid:13)(cid:13)E(M) =(cid:13)(cid:13)U ((an))(cid:13)(cid:13)E(M⊗B(ℓ2(N2))).
This is well-defined and induces a norm on the linear space F. We define the Banach space
Econd(M; ℓc
2) to be the completion of F with respect to the above norm. Then U extends to an
isometry from Econd(M; ℓc
2) into E(M⊗B(ℓ2(N2))) which we will still denote by U .
Similarly, we may define the corresponding row version Econd(M; ℓr
2) which can also be viewed
as a subspace of E(M⊗B(ℓ2(N2))) as row vectors. This is done by simply considering adjoint
operators.
Now, let x = (xn)n≥1 be a finite martingale in L2(M) + M. We set
sc(x) =(cid:16)Xk≥1
Ek−1dxk2(cid:17)1/2
and sr(x) =(cid:16)Xk≥1
Ek−1dx∗
k2(cid:17)1/2
.
These are called the column and row conditioned square functions, respectively. Let FM denote
the set of all finite martingales in L1(M) ∩ M. Define hc
E(M)) as the
= ksr(x)kE(M)).
completion of FM under the norm kxkhc
E(M) may be
We observe that for every x ∈ FM , kxkhc
viewed as a subspace of Econd(M; ℓc
2). More precisely, we consider the map D : FM → F by
= ksc(x)kE(M) (respectively, kxkhr
E(M) (respectively, hr
2). Therefore, hc
= k(dxn)kEcond(M;ℓc
E
E
E
10
N. RANDRIANANTOANINA AND L. WU
setting D(x) = (dxn)n≥1. Then D extends to an isometry from hc
we will denote by Dc. In the sequel, we will make frequent use of the isometric embedding:
E(M) into Econd(M; ℓc
2) which
U Dc : hc
E(M) → E(M⊗B(ℓ2(N2))).
E
:= k(dxn)kE(M⊗ℓ∞). As above, we denote by Dd the isometric extension of D from hd
We can make similar assertions for the row case. That is, hr
E(M) embeds isometrically into
E(M⊗B(ℓ2(N2))). We also need the diagonal Hardy space hd
E(M) which is the space of all
martingales whose martingale difference sequences belong to E(M⊗ℓ∞) equipped with the norm
E(M)
kxkhd
into E(M⊗ℓ∞). We remark that since, under our assumptions on E, conditional expectations
are bounded on E(M), it follows that Dd(hd
E(M)) is a closed subspace of E(M⊗ℓ∞). This
shows in particular that hd
E(M) is a Banach space. Using the natural von Neumann algebra
embeddings, ℓ∞ ⊂ B(ℓ2(N)) ⊂ B(ℓ2(N2)), we can further state that hd
E(M) embeds isometrically
into E(M⊗B(ℓ2(N2))). Consequently, hd
E(M) are compatible as all three
isometrically embed into the larger Banach space E(M⊗B(ℓ2(N2))). We define the conditioned
version of martingale Hardy spaces as follows. If 1 ≤ pE ≤ qE < 2, then
E(M), and hr
E(M), hc
hE(M) = hd
E(M) + hc
E(M) + hr
E(M)
equipped with the norm
where the infimum is taken over all w ∈ hd
w + y + z. If 2 ≤ pE ≤ qE < ∞, then
E(M), y ∈ hc
E(M), and z ∈ hr
E(M) such that x =
kxkhE = inf(cid:8)kwkhd
E
+ kykhc
E
+ kzkhr
E(cid:9),
hE(M) = hd
E(M) ∩ hc
E(M) ∩ hr
E(M)
kxkhE = max(cid:8)kxkhd
E
, kxkhc
E
, kxkhr
E(cid:9).
equipped with the norm
For the case where E = Lp(0, ∞), we will simply write Hp(M), hp(M), ect.
in place of
HLp(M), hLp(M), ect. From the noncommutative Burkholder-Gundy inequalities and noncom-
mutative Burkholder inequalities proved in [29, 39], we have
Hp(M) = hp(M) = Lp(M)
with equivalent norms for all 1 < p < ∞. The latter equality constitutes the primary topic of
this paper.
We collect some basic properties of these various Hardy spaces for further use.
Proposition 2.8. Let 1 ≤ p < q < ∞ and assume that E ∈ Int(Lp, Lq).
p(M), hd
E(M) is complemented in E(M⊗ℓ∞) and hd
E(M) ∈ Int(hd
q (M));
(i) hd
(ii) If 1 < p < q < ∞, then hc
p(M), hs
E(M) ∈ Int(hs
then hs
E(M) is complemented in E(M⊗B(ℓ2(N2))) and for s ∈ {c, r},
q(M)).
Proof. For the first item, the complementation follows immediately from the simple fact that the
map Θ : Lr(M⊗ℓ∞) → Lr(M⊗ℓ∞) defined by:
Θ(cid:0)(an)n≥1(cid:1) = (En(an) − En−1(an))n≥1
is a bounded projection for all 1 ≤ r < ∞. By the interpolation result stated in Proposition 2.6,
Θ : E(M⊗ℓ∞) → E(M⊗ℓ∞) is a bounded projection and it is clear that its range is Dd(hd
E(M)).
The interpolation is an obvious consequence of the complementation result.
The second item is also a consequence of the known fact from [26] that if 1 < r < ∞, hc
r(M)
is complemented in Lr(M⊗B(ℓ2(N2))). Indeed, let Λ : Lr(M⊗B(ℓ2(N2))) → Lr(M⊗B(ℓ2(N2)))
MARTINGALE INEQUALITIES
11
r(M)) for all 1 < r < ∞.
be the bounded projection whose range is U Dc(hc
It is known
that Λ is independent of r in the sense of interpolation ([26, 29]). We deduce that it is also
a bounded projection from E(M⊗B(ℓ2(N2))) onto U Dc(hc
E(M)). The statement about the
range comes from the facts that on one hand, U Dc(FM ) ⊂ Λ(cid:2)E(M⊗B(ℓ2(N2)))(cid:3), and on the
other hand, Sn≥1(cid:2)L1(Mn⊗B(ℓ2(N2))) ∩ Mn⊗B(ℓ2(N2))] is a dense subset of E(M⊗B(ℓ2(N2)))
whose image Sn≥1 Λ(cid:2)L1(Mn⊗B(ℓ2(N2))) ∩ Mn⊗B(ℓ2(N2))] ⊆ U Dc(FM ). This shows that
Λ(cid:2)E(M⊗B(ℓ2(N2)))(cid:3) = U Dc(hc
E(M)). As in the first item, the statement on interpolation
(cid:3)
follows immediately from the complementation result.
Remark 2.9. Unlike the Lp-case, general descriptions of the duals of these more general condi-
tioned Hardy spaces appear to be unavailable . One of the difficulties that arises in trying to
develop such duality theory lies on the fact that, in some cases, E∗ may not be a function space.
For instance, if E = Lr,∞ for 1 ≤ p < r < q < ∞, then E ∈ Int(Lp, Lq), has the Fatou property,
but E∗ is highly nontrivial.
The next theorem is the main result of this section.
Its main feature is that it gives a de-
composition that provides norms estimates simultaneously for all p ∈ (1, 2). This fact is very
crucial in our approach in the next section. Before formally stating this result, we should clarify
that when 1 < p < 2, the noncommutative Burkholder inequalities ([29, Theorem 6.1]) imply
that for each w ∈ {d, c, r} there exists a bounded linear map ξw
p (M) → Lp(M). These
maps are one to one as they come from the isomorphism hp(M) ≈ Lp(M) and the definition of
hp(M) = hd
p (M) can be uniquely repre-
sented as measurable operator from Lp(M). This justifies the use of identification in item (i) in
the statement of the next theorem.
p(M). As a result, any element of hw
p(M) + hd
c (M) + hr
p : hw
Theorem 2.10. There exists a family {κp : 1 < p < 2} ⊂ R+ satisfying the following: if x ∈
hr
p(M)
L1(M) ∩ L2(M), then there exist a ∈T1<p<2
p(M), and c ∈T1<p<2
p(M), b ∈T1<p<2
such that:
hd
hc
(i) x = a + b + c;
(ii) for every 1 < p < 2, the following inequality holds:
(cid:13)(cid:13)a(cid:13)(cid:13)hd
p
+(cid:13)(cid:13)b(cid:13)(cid:13)hc
p
+(cid:13)(cid:13)c(cid:13)(cid:13)hr
p
≤ κp(cid:13)(cid:13)x(cid:13)(cid:13)p.
Proof. Case 1. Assume that M is finite and τ is a normalized trace. The proof uses a weak-type
decomposition from [43]. We consider the interpolation couple (L1(M), L2(M)).
Let x ∈ L2(M). According to [7, Lemma 3.3.2], there is a representation x = Pν∈Z uν
(convergent in L1(M)) of x satisfying, for every ν ∈ Z,
(2.8)
J(uν , 2ν ) ≤ 4K(x, 2ν ).
Since τ (1) = 1, we may apply [43, Theorem 3.1]. The statement of [43, Theorem 3.1] is only for
finite martingales but the construction used there can be applied verbatim to the case of infinite
L2(M)-bounded martingales. There exists an absolute constant κ > 0 such that, for each ν ∈ Z,
we can find three adapted sequences α(ν), β(ν), and γ(ν) in L2(M) such that:
(2.9)
(2.10)
dn(uν) = α(ν)
n + β(ν)
n + γ(ν)
n
for all n ≥ 1,
J(cid:0)Xn≥1
α(ν)
n ⊗ en, t(cid:1) ≤ κJ(uν , t), t > 0,
12
(2.11)
(2.12)
N. RANDRIANANTOANINA AND L. WU
J(cid:0)(cid:0)Xn≥1
J(cid:0)(cid:0)Xn≥1
En−1(β(ν)
En−1(γ(ν)
n
n 2)(cid:1)1/2, t(cid:1) ≤ κJ(uν , t), t > 0,
2)(cid:1)1/2, t(cid:1) ≤ κJ(uν , t), t > 0,
∗
where the J-functional in the left hand side of the inequality in (2.10) is taken relative to the
interpolation couple (L1,∞(M⊗ℓ∞), L2(M⊗ℓ∞)) and those from the left hand sides of (2.11) and
(2.12) are taken with respect to the interpolation couple (L1,∞(M), L2(M)). We set
αn =Xν∈Z
α(ν)
n , βn =Xν∈Z
β(ν)
n , and γn =Xν∈Z
γ(ν)
n .
Then we obtain three adapted sequence α = (αn)n, β = (βn)n, and γ = (γn)n. Define the
martingale difference sequences
dan = αn − En−1(αn), dbn = βn − En−1(βn), and dcn = γn − En−1(γn).
We claim that the resulting martingales a, b, and c satisfy the conclusion of the theorem. Indeed,
it is clear from the construction that x = a + b + c. For the second item, we will verify the
statement separately for a, b, and c. We begin with the martingale a. This will be deduced from
the next lemma. For 0 < θ < 1, 1 < p < 2, and an interpolation couple (X0, X1), (X0, X1)θ,p,J
and (X0, X1)θ,p,K denote the discrete real interpolation methods using the J-functionals and
K-functionals, respectively. We refer to [7] for definitions.
Lemma 2.11. For every 0 < θ < 1 and every 1 < p < 2,
(cid:13)(cid:13)(cid:13)(αn)n≥1(cid:13)(cid:13)(cid:13)[L1,∞(M⊗ℓ∞),L2(M⊗ℓ∞)]θ,p;J
≤ 4κkxk[L1(M),L2(M)]θ,p;K .
For ν ∈ Z, let [α](ν) = Pn α(ν)
n ⊗ en. The series Pν[α](ν) is a representation of Pn αn ⊗ en.
Then the lemma follows immediately from combining (2.8) and (2.10).
Fix 1 < p < 2 and 1/p = (1 − θ) + θ/2. We appeal to the known facts from [40] that for any
semifinite von Neumann algebra N , we have
Lp(N ) =(cid:2)L1,∞(N ), L2(N )(cid:3)θ,p,J and Lp(N ) =(cid:2)L1(N ), L2(N )(cid:3)θ,p,K.
The above lemma yields a constant cp such that
Applying the fact that conditional expectations are contractive projections in Lp(M) gives
(cid:16)Xn
kαnkp
p(cid:17)1/p
≤ cp(cid:13)(cid:13)x(cid:13)(cid:13)p.
=(cid:13)(cid:13)(cid:13)(αn)n≥1(cid:13)(cid:13)(cid:13)Lp(M⊗ℓ∞)
(cid:13)(cid:13)a(cid:13)(cid:13)hd
≤ 2cp(cid:13)(cid:13)x(cid:13)(cid:13)p.
p
Now we sketch the argument for b. We consider the conditioned spaces involved as subspaces of
Lr-spaces associated to M⊗B(ℓ2(N2)) for appropriate values of r. Then for every ν ∈ Z,
J(cid:0)(cid:0)Xn≥1
En−1(β(ν)
n 2)(cid:1)1/2, 2ν(cid:1) = J(cid:16)β(ν), 2ν ; L1,∞(M⊗B(ℓ2(N2))), L2(M⊗B(ℓ2(N2)))(cid:17) .
Therefore, (2.11) becomes,
(2.13)
J(cid:16)β(ν), 2ν ; L1,∞(M⊗B(ℓ2(N2))), L2(M⊗B(ℓ2(N2)))(cid:17) ≤ κJ(uν , 2ν ).
MARTINGALE INEQUALITIES
13
Using similar argument as in the estimate of norm of α with M⊗ℓ∞ replaced by M⊗B(ℓ2(N2)),
we get as in Lemma 2.11 that for every 0 < θ < 1 and 1 < p < 2,
(cid:13)(cid:13)β(cid:13)(cid:13)[L1,∞(M⊗B(ℓ2(N2))),L2(M⊗B(ℓ2(N2)))]θ,p;J
≤ 4κ(cid:13)(cid:13)x(cid:13)(cid:13)[L1(M),L2(M)]θ,p;K
.
As in the previous case, applying real interpolations with appropriate values of θ and p gives that
for every 1 < p < 2,
This is equivalent to
Using Kadison's inequality En−1(βn)∗En−1(βn) ≤ En−1βn2 for all n ≥ 1, we deduce that
Similar argument can be applied to the sequence γ to deduce the corresponding estimate
(cid:13)(cid:13)β(cid:13)(cid:13)Lp(M⊗B(ℓ2(N2))) ≤ cp(cid:13)(cid:13)x(cid:13)(cid:13)p.
En−1(βn2)(cid:1)1/2(cid:13)(cid:13)(cid:13)p
(cid:13)(cid:13)(cid:13)(cid:0)Xn≥1
≤ cp(cid:13)(cid:13)x(cid:13)(cid:13)p.
(cid:13)(cid:13)b(cid:13)(cid:13)hc
≤ 2cp(cid:13)(cid:13)x(cid:13)(cid:13)p.
(cid:13)(cid:13)c(cid:13)(cid:13)hr
≤ 2cp(cid:13)(cid:13)x(cid:13)(cid:13)p.
p
p
Combining the above three estimates clearly provides the second item in the statement of the
theorem. This completes the proof for the finite case.
Case 2. Assume now that M is infinite. Since M∗ is separable, the von Neumann algebra M
is σ-finite. We note first that Case 1 extends easily to any finite case with the trace τ being
not necessarily normalized (with the same constants as in the case of normalized trace). Since
there is a trace preserving conditional expectation E1 : M → M1, it is known that τ M1 remains
semifinite.
Fix an increasing sequence of projections (ek)k≥1 ⊂ M1 with τ (ek) < ∞ for all k ≥ 1 and such
that (ek)k≥1 converges to 1 for the strong operator topology. For each k, consider the finite von
Neumann algebra (ekMek, τ ekMek ) with the filtration (ekMnek)n≥1. If we denote by E (k)
the
trace preserving conditional expectation from ekMek onto ekMnek then E (k)
n is just the restriction
of En on ekMek. This is the case since the ek's were chosen from the smallest subalgebra M1.
Therefore, if y ∈ ekMek then one can easily verify that
n
p(ekMek) =(cid:13)(cid:13)y(cid:13)(cid:13)hr
p(M).
Let x ∈ L2(M) ∩ L1(M). For each k ≥ 1, ekxek ∈ L2(ekMek). From Case 1., there exists a
decomposition ekxek = a(k) + b(k) + c(k) with the property that for every 1 < p < 2,
(cid:13)(cid:13)y(cid:13)(cid:13)hd
p(ekMek) =(cid:13)(cid:13)y(cid:13)(cid:13)hd
p(M), (cid:13)(cid:13)y(cid:13)(cid:13)hc
(cid:13)(cid:13)a(k)(cid:13)(cid:13)hd
p(ekMek) =(cid:13)(cid:13)y(cid:13)(cid:13)hc
+(cid:13)(cid:13)b(k)(cid:13)(cid:13)hc
+(cid:13)(cid:13)c(k)(cid:13)(cid:13)hr
p(M), and (cid:13)(cid:13)y(cid:13)(cid:13)hr
≤ κp(cid:13)(cid:13)ekxek(cid:13)(cid:13)p.
p
p
p
Fix an ultrafilter U on N containing the Fr´echet filter. For any given 1 < p < 2, the weak-limit
along the ultrafilter U of the sequence (a(k))k≥1 exists in hd
p(M). It is crucial here to observe
that such weak-limits are independent of p (since they are automatically weak-limits of the same
sequence in L1(M⊗ℓ∞) + L2(M⊗ℓ∞)). Similar observations can be made with the sequences
(b(k))k≥1 and (c(k))k≥1. Set
a = w- lim
k,U
a(k), b = w- lim
k,U
b(k), and c = w- lim
k,U
c(k)
p(M), hc
p(M), and hr
in hd
p(M), respectively. We also observe that for every 1 < p < 2, it is easy
to verify that limk→∞ kekxek − xkp = 0. A fortiori, limk,U kekxek − xkp = 0. All these facts lead
to the decomposition:
x = a + b + c.
14
N. RANDRIANANTOANINA AND L. WU
Furthermore, for every 1 < p < 2, we have
(cid:13)(cid:13)a(cid:13)(cid:13)hd
p
+(cid:13)(cid:13)b(cid:13)(cid:13)hc
p
+(cid:13)(cid:13)c(cid:13)(cid:13)hr
p
≤ sup
k n(cid:13)(cid:13)a(k)(cid:13)(cid:13)hd
+(cid:13)(cid:13)b(k)(cid:13)(cid:13)hc
k (cid:13)(cid:13)ekxek(cid:13)(cid:13)p ≤ κp(cid:13)(cid:13)x(cid:13)(cid:13)p
p
p
≤ κp sup
po
+(cid:13)(cid:13)c(k)(cid:13)(cid:13)hr
where κp is the constant from Case 1. The proof is complete.
(cid:3)
Remarks 2.12. 1) Since the noncommutative Burkholder inequalities do not hold for p = 1, the
validity of our simultaneous decomposition can not include the left endpoint of the interval (1, 2).
On the other hand, using known estimates from real interpolation (θ, p, K) and (θ, p, J) methods
of classical Lebesgue spaces, we can derive that there is an absolute constant C such that for
1/p = (1 − θ) + θ/2, we have κp ≤ Cθ−2(1 − θ)−1/2−1/p. It follows that κp is of order (p − 1)−2
when p → 1 and of order (2 − p)−1 when p → 2. In particular, our method of proof does not
allow any extension of the decomposition to any of the endpoints of the interval [1, 2]. As we only
get that κp = O((p − 1)−2) when p → 1, our arguments do not yield the optimal order for the
constants for the noncommutative Burkholder inequalities from [43].
2) Junge and Perrin also considered simultaneous type decompositions for conditioned Hardy
spaces in [28]. Our Theorem 2.10 above should be compared with [28, Theorem 5.9]. See also
Corollary 4.4 below for similar type simultaneous decompositions for the case of martingale Hardy
space norms.
3. Burkholder's inequalities in symmetric spaces
The following is the principal result of this article. It provides extensions of noncommutative
Burkholder's inequalities for martingales in general noncommutative symmetric spaces.
Theorem 3.1. Let E be a symmetric Banach function space on (0, ∞) satisfying the Fatou
property. Assume that either 1 < pE ≤ qE < 2 or 2 < pE ≤ qE < ∞. Then
E(M) = hE(M).
That is, a martingale x = (xn)n≥1 is bounded in E(M) if and only if it belongs to hE(M) and
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≃E (cid:13)(cid:13)x(cid:13)(cid:13)hE
.
As noted in the introduction, the preceding theorem solves positively a question raised in [22].
The new result here is the case where 1 < pE ≤ qE < 2. The case 2 < pE ≤ qE < ∞ was
established by Dirksen in [12, Theorem 6.2] but we will also provide an alternative approach for
this range. We remark that under the assumptions of Theorem 3.1, the Banach function space E
is fully symmetric in the sense of [17] but this extra property will not be needed in the proof.
We divide the proof into four separate parts according to 1 < pE ≤ qE < 2 or 2 < pE ≤ qE < ∞,
each case involving two inequalities. The main difficulty in the proof is Part II below. Part III
will be deduced from Part II via duality. The other two parts will be derived from standard use
of interpolations of linear operators.
3.1. The case 1 < pE ≤ qE < 2. Let E be a symmetric Banach function space on (0, ∞) with
the Fatou property and satisfying 1 < pE ≤ qE < 2. Throughout the proof, we fix p and q so
that 1 < p < pE ≤ qE < q < 2. In this case, E ∈ Int(Lp, Lq).
Part I. We will verify that there exists a constant cE such that for every x ∈ hE(M), we have
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ cE(cid:13)(cid:13)x(cid:13)(cid:13)hE
.
MARTINGALE INEQUALITIES
15
This is a simple consequence of the noncommutative Burkholder inequalities and Proposition 2.8.
E (M).
. By interpolation, we deduce that(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ cE(cid:13)(cid:13)x(cid:13)(cid:13)hd
r
E (M).
Now, assume that x ∈ hE(M) is such that x = w + y + z where w ∈ hd
E(M), y ∈ hr
E(M), and
Indeed, for 1 < r < 2,(cid:13)(cid:13)x(cid:13)(cid:13)r ≤ cr(cid:13)(cid:13)x(cid:13)(cid:13)hd
Similar arguments also give (cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ cE(cid:13)(cid:13)x(cid:13)(cid:13)hc
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ cE(cid:0)(cid:13)(cid:13)w(cid:13)(cid:13)hd
E(M). Then we have
z ∈ hr
E (M) and (cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ cE(cid:13)(cid:13)x(cid:13)(cid:13)hr
+(cid:13)(cid:13)y(cid:13)(cid:13)hc
+(cid:13)(cid:13)z(cid:13)(cid:13)hr
E(cid:1).
E
E
Taking the infimum over all decompositions x = w + y + z provides the desired inequality.
(cid:3)
We should emphasize that the inequality from Part I may be interpreted as inclusion mappings
E : hw
in the following sense: if w ∈ {d, c, r}, then there is a natural bounded map ξw
E(M) → E(M).
In fact, ξw
) → E(M) defined by
(xn)n≥1 7→ limn→∞ xn (since (xn) is a finite sequence, the limit should be understood as the final
value of (xn)). We claim that these maps are one to one. To verify this claim, we consider the
following diagram:
E may be taken as the bounded extension of the map (FM , k · khw
E
hw
E(M)
ιy
ξw
E−−−−→
E(M)
yj
p (M) + hw
hw
q (M)
p,q−−−−→ Lp(M) + Lq(M),
ξw
p and ξw
q
p,q is the combination of ξw
where ι is the inclusion map from the interpolation in Proposition 2.8, j is the formal inclusion,
and ξw
introduced in the previous section. When z ∈ FM ,
we clearly have ξw
p,q are
one to one, so is ξw
E(M) is uniquely associated with the
operator ξw
E (a) ∈ E(M) which we will still denote by a. In Part II below, statement such as
x = a + b + c for x ∈ E(M), a ∈ hd
E(M) should be understood to
mean x = ξd
p,qι(z) = jξw
E. As a consequence, any element a ∈ hw
E (z). Therefore, the above diagram commutes. Since ι and ξw
E(M), and c ∈ hr
E(M), b ∈ hc
E(a) + ξc
E(b) + ξr
E(c).
Part II. We consider now the reverse inequalities. That is, there exists a constant βE such that
for every x ∈ E(M),
(cid:13)(cid:13)x(cid:13)(cid:13)hE
≤ βE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
The proof is much more involved and requires several steps. Our approach relies on two essential
facts. As stated in Corollary 2.5, noncommutative symmetric spaces have concrete representations
as interpolation spaces. The second fact is the simultaneous decomposition obtained in the
previous section.
According to Corollary 2.5, we may fix a symmetric Banach function space F on (0, ∞) with
nontrivial Boyd indices and such that for any semifinite von Neumann algebra N , we have:
E(N ) =(cid:2)Lp(N ), Lq(N )(cid:3)F,j,
where [·, ·]F,j is the interpolation method introduced in the previous section. We begin with the
following intermediate result.
Lemma 3.2. Let x ∈ Lp(M) ∩ Lq(M). For every ε > 0, there exist x(ε) ∈ Lp(M) ∩ Lq(M),
martingales a(ε) ∈ hd
E(M), and c(ε) ∈ hr
E(M), b(ε) ∈ hc
E(M) with:
(1) (cid:13)(cid:13)x − x(ε)(cid:13)(cid:13)Lp(M)∩Lq(M) < ε;
(2) x(ε) = a(ε) + b(ε) + c(ε);
16
N. RANDRIANANTOANINA AND L. WU
E
+(cid:13)(cid:13)b(ε)(cid:13)(cid:13)hc
(3) (cid:13)(cid:13)a(ε)(cid:13)(cid:13)hd
Proof. Let x ∈ Lp(M) ∩ Lq(M) and ε > 0. Using the interpolation couple (cid:0)Lp(M), Lq(M)(cid:1), fix
a representation x =Pν∈Z uν (convergent in Lp(M) + Lq(M)) such that
+(cid:13)(cid:13)c(ε)(cid:13)(cid:13)hr
E
E
(3.1)
≤ ηE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
(cid:13)(cid:13)(cid:13)j(cid:0){uν}ν , ·(cid:1)(cid:13)(cid:13)(cid:13)F
≤ 2(cid:13)(cid:13)x(cid:13)(cid:13)F,j.
Note that the uν's belong to Lp(M) ∩ Lq(M). Using Lemma 2.1, for each ν ∈ Z, we may choose
u(mν )
ν
∈ L1(M) ∩ L2(M) satisfying the following properties:
ν
(1) (cid:13)(cid:13)u(mν )
(2) (cid:13)(cid:13)u(mν )
(3) (cid:13)(cid:13)u(mν )
ν
ν
− uν(cid:13)(cid:13)Lp(M)∩Lq(M) ≤
(cid:13)(cid:13)q ≤(cid:13)(cid:13)uν(cid:13)(cid:13)q;
(cid:13)(cid:13)p ≤(cid:13)(cid:13)uν(cid:13)(cid:13)p.
ε
;
4ν+1
The last two conditions imply that for every ν ∈ Z and every t > 0,
which furthermore leads to the following inequality:
(3.2)
We define the operator x(ε) by setting:
ν
J(cid:0)u(mν )
j(cid:0){u(mν )
, t(cid:1) ≤ J(cid:0)uν, t(cid:1),
}ν , ·(cid:1) ≤ j(cid:0){uν}ν , ·(cid:1).
x(ε) =Xν∈Z
u(mν )
ν
ν
.
Then it satisfies the following norm estimates:
(cid:13)(cid:13)x(ε) − x(cid:13)(cid:13)Lp(M)∩Lq(M) ≤Xν∈Z(cid:13)(cid:13)u(mν )
ν
− uν(cid:13)(cid:13)Lp(M)∩Lq (M)
≤ ε.
≤Xν∈Z
ε
4ν+1
In particular, x(ε) ∈ Lp(M)∩ Lq(M) and the first item in the statement of Lemma 3.2 is satisfied.
The crucial fact here is that all u(mν )
's in the representation of x(ε) belong to L1(M) ∩ L2(M)
so that Theorem 2.10 can be applied to each of the u(mν )
's. That is, for every ν ∈ Z, there exist
aν ∈ ∩1<s<2hd
s(M), and cν ∈ ∩1<s<2hr
s(M) satisfying:
= aν + bν + cν
s(M), bν ∈ ∩1<s<2hc
(3.3)
ν
ν
u(mν )
ν
and if s ∈ {p, q}, then
(3.4)
where κ(p, q) = max{κp, κq}.
(cid:13)(cid:13)aν(cid:13)(cid:13)hd
s
+(cid:13)(cid:13)bν(cid:13)(cid:13)hc
s
+(cid:13)(cid:13)cν(cid:13)(cid:13)hr
s
≤ κ(p, q)(cid:13)(cid:13)u(mν )
ν
(cid:13)(cid:13)s
For each ν ∈ Z, we consider Dd(aν ) ∈ Lp(M⊗ℓ∞)∩Lq(M⊗ℓ∞), U Dc(bν) ∈ Lp(M⊗B(ℓ2(N2)))∩
Lq(M⊗B(ℓ2(N2))), and U Dc(c∗
ν) ∈ Lp(M⊗B(ℓ2(N2))) ∩ Lq(M⊗B(ℓ2(N2))).
First, we observe that (3.4) can be reinterpreted using the J-functionals as follows:
(3.5)
(3.6)
(3.7)
ν
J(cid:0)Dd(aν ), t) ≤ κ(p, q)J(u(mν )
J(cid:0)U Dc(bν ), t(cid:1) ≤ κ(p, q)J(u(mν )
J(cid:0)U Dc(c∗
ν ), t(cid:1) ≤ κ(p, q)J(u(mν )
ν
ν
, t), t > 0,
, t), t > 0,
, t), t > 0
MARTINGALE INEQUALITIES
17
where the J-functional on the left side of (3.5) is taken using the couple [Lp(M⊗ℓ∞), Lq(M⊗ℓ∞)]
and those on the left sides of inequalities (3.6) and (3.7) were computed using the couple
[Lp(M⊗B(ℓ2(N2))), Lq(M⊗B(ℓ2(N2)))].
We need the following properties of the sequences {Dd(aν )}ν∈Z, {U Dc(bν)}ν∈Z, and {U Dc(c∗
ν )}ν∈Z.
We refer to [11] for definition and criterion for unconditionally Cauchy series in Banach spaces.
Sublemma 3.3. (1) Pν∈Z Dd(aν) is a weakly unconditionally Cauchy series in E(M⊗ℓ∞).
(2) Pν∈Z U Dc(bν) andPν∈Z U Dc(c∗
kSN (c∗)kE(M⊗B(ℓ2(N2)))o
Moreover, there exists a constant κE such that:
kSN (a)kE(M⊗ℓ∞), sup
N ≥1
ν) are weakly unconditionally Cauchy series in E(M⊗B(ℓ2(N2))).
maxn sup
N ≥1
kSN (b)kE(M⊗B(ℓ2(N2))), sup
N ≥1
≤ κEkxkE(M),
where for each N ≥ 1, SN (a) = Pν≤N Dd(aν ), SN (b) = Pν≤N U Dc(bν), and SN (c∗) =
Pν≤N U Dc(c∗
To verify the first item in Sublemma 3.3, we note that if S if a finite subset of Z then it follows
ν ).
from (3.2) and (3.5) that
By the definition of [·, ·]F,j , for every finite sequence of scalars (θν)ν∈S with θν = 1, we have
}ν∈S, ·(cid:1) ≤ κ(p, q)j(cid:0){uν }ν, ·(cid:1).
j(cid:0){Dd(aν )}ν∈S , ·(cid:1) ≤ κ(p, q)j(cid:0){u(mν )
θνDd(aν )(cid:13)(cid:13)[Lp(M⊗ℓ∞),Lq(M⊗ℓ∞)]F,j
ν
where the last inequality is from (3.1). Now we use the facts that
(cid:13)(cid:13)Xν∈S
≤ κ(p, q)(cid:13)(cid:13)j(cid:0){uν }ν, ·(cid:1)kF ≤ 2κ(p, q)(cid:13)(cid:13)x(cid:13)(cid:13)F,j
E(M⊗ℓ∞) =(cid:2)Lp(M⊗ℓ∞), Lq(M⊗ℓ∞)(cid:3)F,j and E(M) =(cid:2)Lp(M), Lq(M)(cid:3)F,j
to deduce that there exists a constant κE such that:
(cid:13)(cid:13)Xν∈S
θνDd(aν )(cid:13)(cid:13)E(M⊗ℓ∞) ≤ κE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
Since this is the case for any arbitrary finite subset of Z, it proves that the series Pν∈Z Dd(aν ) is
weakly unconditionally Cauchy in E(M⊗ℓ∞).
The proof of the second item follows the same pattern. As above, if S is a finite subset of Z,
then it follows from (3.2) and (3.6) that:
j(cid:0){U Dc(bν)}ν∈S , ·(cid:1) ≤ κ(p, q)j(cid:0){u(mν )
ν
}ν∈S, ·(cid:1) ≤ κ(p, q)j(cid:0){uν}ν , ·(cid:1).
Using similar arguments as above, we may deduce that for every finite subset S of Z and for every
sequence of scalars (θν)ν∈S with θν = 1,
(3.8)
(3.9)
(cid:13)(cid:13)Xν∈S
θνU Dc(bν)(cid:13)(cid:13)E(M⊗B(ℓ2(N2))) ≤ κE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
This again shows that the seriesPν∈Z U Dc(bν) is weakly unconditionally Cauchy in E(M⊗B(ℓ2(N2))).
The proof for the series Pν∈Z U Dc(c∗
in Sublemma 3.3 follows from (3.8), (3.9), and the corresponding inequality for Pν∈Z U Dc(c∗
Next, we note that since Lp(M⊗ℓ∞) + Lq(M⊗ℓ∞) is a reflexive space, the series Pν∈Z Dd(aν )
is unconditionally convergent in Lp(M⊗ℓ∞) + Lq(M⊗ℓ∞). Similarly, both series Pν∈Z U Dc(bν )
ν ) is identical so we omit the details. The inequality stated
ν ).
Sublemma 3.3 is verified.
18
N. RANDRIANANTOANINA AND L. WU
ν ) are unconditionally convergent in Lp(M⊗B(ℓ2(N2))) + Lq(M⊗B(ℓ2(N2))).
and Pν∈Z U Dc(c∗
Now we set:
and
α(ε) :=Xν∈Z
Dd(aν ) ∈ Lp(M⊗ℓ∞) + Lq(M⊗ℓ∞),
U Dc(bν ) ∈ Lp(M⊗B(ℓ2(N2))) + Lq(M⊗B(ℓ2(N2))),
U Dc(c∗
ν ) ∈ Lp(M⊗B(ℓ2(N2))) + Lq(M⊗B(ℓ2(N2))).
β(ε) :=Xν∈Z
γ(ε) :=Xν∈Z
We claim that
(3.10)
maxn(cid:13)(cid:13)α(ε)(cid:13)(cid:13)E(M⊗ℓ∞),(cid:13)(cid:13)β(ε)(cid:13)(cid:13)E(M⊗B(ℓ2(N2))),(cid:13)(cid:13)γ(ε)(cid:13)(cid:13)E(M⊗B(ℓ2(N2)))o ≤ κE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
To verify this claim, we use the fact mentioned in the previous section that for every semifinite
von Neumann algebra N , the inclusion map from E(N ) into Lp(N )+Lq(N ) is a semi-embedding.
Indeed, if ρ = κEkxkE(M), then from Sublemma 3.3, (SN (a))N ≥1 is a sequence in the ρ-ball of
E(M⊗ℓ∞) that converges to α(ε) for the norm topology of Lp(M⊗ℓ∞) + Lq(M⊗ℓ∞). By semi-
embedding, we have α(ε) ∈ E(M⊗ℓ∞) with kα(ε)kE(M⊗ℓ∞) ≤ ρ.
Identical arguments can be
applied to β(ε) and γ(ε) to deduce that (cid:13)(cid:13)β(ε)(cid:13)(cid:13)E(M⊗B(ℓ2(N2))) ≤ ρ and (cid:13)(cid:13)γ(ε)(cid:13)(cid:13)E(M⊗B(ℓ2(N2))) ≤ ρ.
We have verified (3.10).
We observe that since the sequence (SN (a))N is from Dd(hd
that α(ε) ∈ Dd(hd
Lp+Lq (M)). That is, there exists a(ε) ∈ hd
But since (cid:13)(cid:13)α(ε)(cid:13)(cid:13)E(M⊗ℓ∞) ≤ κE(cid:13)(cid:13)x(cid:13)(cid:13)E(M), by the complementation stated in Proposition 2.8, we
E(M) with the norm estimate:
conclude that a(ε) ∈ hd
p(M)) ∩ Dd(hd
q (M)), it follows
Lp+Lq (M) such that α(ε) = Dd(a(ε)).
(cid:13)(cid:13)a(ε)(cid:13)(cid:13)hd
E
≤ κE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
Identical arguments can be applied to the sequences (SN (b))N and (SN (c))N to deduce that there
exist martingales b(ε) ∈ hc
E(M) such that β(ε) = U Dc(b(ε)), γ(ε) = U Dc((c(ε))∗),
and
E(M) and c(ε) ∈ hr
maxn(cid:13)(cid:13)b(ε)(cid:13)(cid:13)hc
E
,(cid:13)(cid:13)c(ε)(cid:13)(cid:13)hr
Eo ≤ κE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
It is clear from the construction that x(ε) = a(ε) + b(ε) + c(ε) and the last two inequalities clearly
implies the last item in Lemma 3.2. The proof is complete.
(cid:3)
The next step provides the desired decomposition for all x ∈ Lp(M) ∩ Lq(M).
Lemma 3.4. There exists a constant βE such that every x ∈ Lp(M) ∩ Lq(M) admits a decom-
position x = a + b + c where a ∈ hd
E(M), and c ∈ hr
E(M) satisfying:
E(M), b ∈ hc
(cid:13)(cid:13)a(cid:13)(cid:13)hd
E
+(cid:13)(cid:13)b(cid:13)(cid:13)hc
E
+(cid:13)(cid:13)c(cid:13)(cid:13)hr
E
≤ βE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
Proof. We use semi-embedding techniques. Using Lemma 3.2, we construct sequence of opera-
tors (x(m))m≥1 ⊂ Lp(M) ∩ Lq(M), sequences (a(m))m≥1 ⊂ hd
E(M), and
(c(m))m≥1 ⊂ hr
E(M), (b(m))m≥1 ⊂ hc
E(M) such that:
(i) limm→∞ kx(m) − xkLp(M)∩Lq(M) = 0;
(ii) x(m) = a(m) + b(m) + c(m) for all m ≥ 1:
(iii) (cid:13)(cid:13)a(m)(cid:13)(cid:13)hd
E
+(cid:13)(cid:13)b(m)(cid:13)(cid:13)hc
E
+(cid:13)(cid:13)c(m)(cid:13)(cid:13)hr
E
≤ ηE(cid:13)(cid:13)x(cid:13)(cid:13)E(M).
MARTINGALE INEQUALITIES
19
Let ρ = ηEkxkE(M). Then the sequence {Dd(a(m))}m≥1 belongs to the ρ-ball of E(M⊗ℓ∞).
Similarly, {U Dc(b(m))}m≥1 and {U Dc((c(m))∗)}m≥1 belong to the ρ-ball of E(M⊗B(ℓ2(N2))).
Since the spaces Lp(M⊗ℓ∞)+ Lq(M⊗ℓ∞) and Lp(M⊗B(ℓ2(N2)))+ Lq(M⊗B(ℓ2(N2))) are re-
for weak topology in Lp(M⊗ℓ∞)+Lq(M⊗ℓ∞) and both {U Dc(b(m))}m≥1 and {U Dc((c(m))∗)}m≥1
flexive, we may assume (after taking subsequences if necessary) that {Dd(a(m))}m≥1 converges toea
converge (for the weak topology of Lp(M⊗B(ℓ2(N2))) + Lq(M⊗B(ℓ2(N2)))) toeb and ec∗, respec-
As a consequence of the fact that the inclusion mappings are semi-embedings, it is clear that
tively.
these limits satisfy:
maxnkeakE(M⊗ℓ∞), kebkE(M⊗B(ℓ2(N2))), kec∗kE(M⊗B(ℓ2(N2)))o ≤ ρ.
E(M), b ∈ hc
Using similar arguments as in the proof of the previous lemma, we may deduce that there exist
a ∈ hd
E(M), and c ∈ hr
the previous inequality translates into
E(M) such that ea = Dd(a), eb = U Dc(b), ec∗ = U Dc(c∗), and
maxnkakhd
Eo ≤ ρ.
, kckhr
, kbkhc
E
E
Moreover, it is clear from (i) and (ii) that x = a + b + c. Thus, we have verified Lemma 3.4. (cid:3)
To conclude the proof of Part II, it is enough to note that since Lp(M) ∩ Lq(M) is dense in
E(M). The assertion that (cid:13)(cid:13)x(cid:13)(cid:13)hE
Lemma 3.4.
≤ βE(cid:13)(cid:13)x(cid:13)(cid:13)E(M) for all x ∈ E(M) then follows immediately from
3.2. The case 2 < pE ≤ qE < ∞. Assume now that E is a symmetric Banach function space on
(0, ∞) satisfying the Fatou property and 2 < pE ≤ qE < ∞.
Part III. We will verify that for every x ∈ hE(M),
(cid:13)(cid:13)xkE(M) .(cid:13)(cid:13)x(cid:13)(cid:13)hE
.
(1 − ε)(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ τ (xy∗).
+(cid:13)(cid:13)b(cid:13)(cid:13)hc
+(cid:13)(cid:13)c(cid:13)(cid:13)hr
E×
E×
(cid:13)(cid:13)a(cid:13)(cid:13)hd
E×
≤ κE× + ε.
(cid:13)(cid:13)a − a(cid:13)(cid:13)hd
(cid:13)(cid:13)b − b(cid:13)(cid:13)hc
(cid:13)(cid:13)c − c(cid:13)(cid:13)hr
E×
E×
E×
+(cid:13)(cid:13)a − a(cid:13)(cid:13)E×(M) < ε/3,
+(cid:13)(cid:13)b − b(cid:13)(cid:13)E×(M) < ε/3,
+(cid:13)(cid:13)c − c(cid:13)(cid:13)E×(M) < ε/3.
This will be deduced from Part II using duality. Let E× be the Kothe dual of E. The non-
commutative symmetric space E×(M) is the Kothe dual of E(M) in the sense of [18]. Since E
has the Fatou property, it follows that for every x ∈ E(M), we have kxkE(M) = kxkE××(M). In
particular, the closed unit ball of E×(M) is a norming set for E(M).
It is enough to verify the inequality for x ∈ L1(M) ∩ M. For ε > 0, choose y ∈ E×(M), with
kykE×(M) = 1, and such that
From [34, Proposition 2.b.2], the Boyd indices of E× satisfy 1 < pE× ≤ qE× < 2. Thus, using
Part II, it follows that y ∈ h
E×. We may choose a decomposition y = a + b + c satisfying:
From the discussion after the proof of Part I, a, b, and c can be represented by operators from
E×(M), which we will still denote by a, b, and c. By density, there exist N ≥ 1, a, b, and c in
L1(MN ) ∩ MN so that
20
N. RANDRIANANTOANINA AND L. WU
Now, τ (xy∗) = τ (xa∗)+τ (xb∗)+τ (xc∗) ≤ ε(cid:13)(cid:13)x(cid:13)(cid:13)E(M) +τ (xa∗)+τ (xb∗)+τ (xc∗) := ε(cid:13)(cid:13)x(cid:13)(cid:13)E(M) +
I + II + II. We estimate I, II, and III separately. Below, we denote by γ and tr the usual
traces on ℓ∞ and B(ℓ2(N2)), respectively. For I, we have the following estimates:
τ (dxnda∗
n)(cid:12)(cid:12)(cid:12)
I =(cid:12)(cid:12)(cid:12)
NXn=1
=(cid:12)(cid:12)(cid:12)τ ⊗ γ(cid:0)(dxn)1≤n≤N .(da∗
≤(cid:13)(cid:13)(dxn)1≤n≤N(cid:13)(cid:13)E(M⊗ℓ∞).(cid:13)(cid:13)(dan)1≤n≤N(cid:13)(cid:13)E×(M⊗ℓ∞)
.(cid:13)(cid:13)a(cid:13)(cid:13)hd
=(cid:13)(cid:13)x(cid:13)(cid:13)hd
≤(cid:0)ε/3 +(cid:13)(cid:13)a(cid:13)(cid:13)hd
E×(cid:1)(cid:13)(cid:13)x(cid:13)(cid:13)hd
n)1≤n≤N(cid:1)(cid:12)(cid:12)(cid:12)
E×
.
E
E
E(M) and hc
E×(M) as subspaces of E(M⊗B(ℓ2(N2))) and
n=1 τ (dxndb∗
n)(cid:12)(cid:12). Since the conditional
For II, we use the identifications of hc
expectations Ek's are trace invariant, we have:
E×(M⊗B(ℓ2(N2))), respectively. First, we write II = (cid:12)(cid:12)PN
ndxn))(cid:12)(cid:12)(cid:12)
ndxn)i(cid:12)(cid:12)(cid:12).
II =(cid:12)(cid:12)(cid:12)
NXn=1
=(cid:12)(cid:12)(cid:12)τh NXn=1
τ (En−1(db∗
En−1(db∗
We note that (dbn)1≤n≤N and (dxn)1≤n≤N are sequences from F and therefore (2.7) applies. We
may then write
II =(cid:12)(cid:12)(cid:12)τ ⊗ trhU ((dbn)1≤n≤N )∗U ((dxn)1≤n≤N )i(cid:12)(cid:12)(cid:12)
≤(cid:13)(cid:13)(cid:13)U ((dbn)n≥1)(cid:13)(cid:13)(cid:13)E×(M⊗B(ℓ2(N2)))(cid:13)(cid:13)(cid:13)U ((dxn)n≥1)(cid:13)(cid:13)(cid:13)E(M⊗B(ℓ2(N2)))
=(cid:13)(cid:13)b(cid:13)(cid:13)hc
.(cid:13)(cid:13)x(cid:13)(cid:13)hc
≤ (ε/3 +(cid:13)(cid:13)b(cid:13)(cid:13)hc
The proof that III ≤ (cid:0)ε/3 +(cid:13)(cid:13)c(cid:13)(cid:13)hr
)(cid:13)(cid:13)x(cid:13)(cid:13)hc
E×(cid:1)(cid:13)(cid:13)x(cid:13)(cid:13)hr
above estimates on I, II, and III, we derive that
E×
E×
.
E
E
E
Taking infimum over ε gives the desired inequality.
(1 − 2ε)(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ (κE× + 2ε)(cid:13)(cid:13)x(cid:13)(cid:13)hE
.
is identical so we omit the details. Combining the
Part IV. The remaining case is to show the reverse inequality (cid:13)(cid:13)x(cid:13)(cid:13)hE
easy application of the noncommutative Burkholder inequalities for Ls-bounded martingales when
2 < s < ∞ together with Proposition 2.8. Details are left to the reader.
(cid:3)
. (cid:13)(cid:13)x(cid:13)(cid:13)E(M). This is an
Remark 3.5. Our duality argument in Part III strongly relies on the theory of Kothe dualities for
noncommutative symmetric space from [18]. At the time of this writing, we do not know how to
incorporate this theory into martingales Hardy spaces.
MARTINGALE INEQUALITIES
21
For the case where M is a finite von Neumann algebra equipped with a normal tracial state τ , it
is more natural to consider symmetric Banach function spaces defined on the interval [0, 1]. How-
ever, the definition of the diagonal Hardy space uses the infinite von Neumann algebra M⊗ℓ∞.
In this case, we may consider an extension of symmetric Banach function space E on [0, 1] into a
symmetric Banach function space on (0, ∞) introduced in [23] (see also [1, 34] for more details).
E be the symmetric space on (0, ∞) of all measurable functions f for which µ(f )χ(0,1] ∈ E
Let Z 2
and µ(f )χ(1,∞) ∈ L2(0, ∞), endowed with the norm
(cid:13)(cid:13)f(cid:13)(cid:13)Z 2
E
= maxn(cid:13)(cid:13)µ(f )χ(0,1](cid:13)(cid:13)E,(cid:16) ∞Xn=0(cid:16)Z n+1
n
µu(f ) du(cid:17)2(cid:17)1/2o.
It was shown in [34, Theorem 2.f.1] that if E has nontrivial Boyd indices then Z 2
to E. Using the symmetric Banach function space Z 2
the following variant of Theorem 3.1:
E is isomorphic
E on the diagonal Hardy space, we may state
Theorem 3.6. Assume that (M, τ ) is a finite von Neumann algebra with τ being a normal tracial
state and E is a symmetric Banach function space on [0, 1] satisfying the Fatou property. Let
x = (xn)n≥1 be a bounded E(M)-martingale.
(1) If 1 < pE ≤ qE < 2, then
where the infimum runs over all decompositions x = w + y + z with w, y, and z are
martingales.
(2) If 2 < pE ≤ qE < ∞, then
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≃E infn(cid:13)(cid:13)w(cid:13)(cid:13)hd
Z2
E
+(cid:13)(cid:13)y(cid:13)(cid:13)hc
E
Eo
+(cid:13)(cid:13)z(cid:13)(cid:13)hr
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≃E maxn(cid:13)(cid:13)x(cid:13)(cid:13)hd
Z2
E
,(cid:13)(cid:13)x(cid:13)(cid:13)hc
E
Eo.
,(cid:13)(cid:13)x(cid:13)(cid:13)hr
The assumptions of Theorem 3.1 and Theorem 3.6 are equivalent to E ∈ Int(Lp, Lq) with
1 < p < q < 2 or 2 < p < q < ∞. Indeed, if E ∈ Int(Lp, Lq) then p ≤ pE ≤ qE ≤ q. We do not
know if our results extend to the case where E ∈ Int(Lp, L2) for 1 < p < 2 or E ∈ Int(L2, Lq) for
q > 2. On the other hand, since h1(M) ⊂ L1(M), the argument used in Part I of the proof of
Theorem 3.1 can be readily adjusted to provide the following:
If E ∈ Int(L1, Lq) where q ≤ 2, then there exists a constant cE such that for every martingale
x ∈ hE(M),
Similarly, if E ∈ Int(L2, Lq) where 2 < q < ∞, then there exists a constant cE such that for every
y ∈ E(M),
.
(cid:13)(cid:13)x(cid:13)(cid:13)E(M) ≤ cE(cid:13)(cid:13)x(cid:13)(cid:13)hE
(cid:13)(cid:13)y(cid:13)(cid:13)hE
≤ cE(cid:13)(cid:13)y(cid:13)(cid:13)E(M).
We conclude this section by pointing out that as with the case of noncommutative Burkholder-
Gundy inequalities, no equivalence of norms is known for the case where 1 < pE < 2 and
2 < qE < ∞.
4. Further remarks
We begin this section with a short discussion about the comparison between martingales Hardy
spaces and conditioned martingale Hardy spaces associated with general symmetric spaces. Com-
bining Theorem 3.1 with the main result of [22], we may state:
22
N. RANDRIANANTOANINA AND L. WU
Corollary 4.1. Let E be a symmetric Banach function space on (0, ∞) satisfying the Fatou
property. Assume that either 1 < pE ≤ qE < 2 or 2 < pE ≤ qE < ∞. Then
with equivalent norms.
HE(M) = hE(M) = E(M),
We recall the noncommutative Davis' decomposition established in [27, 38] which states that
H1(M) = h1(M). In view of this equivalence, it seems reasonable to expect that the assumption
pE > 1 is not needed in the first equality in Corollary 4.1. Unfortunately, our interpolation
techniques are not efficient enough to apply to the case pE = 1. We leave this as an open
problem.
Problem 4.2. Assume that pE = 1 and qE ≤ 2. Do we have HE(M) = hE(M)?
Our next result shows a connection between the diagonal Hardy space and the row/column
Hardy spaces.
Proposition 4.3. If qE < 2, then max(cid:8)(cid:13)(cid:13)a(cid:13)(cid:13)Hc
,(cid:13)(cid:13)a(cid:13)(cid:13)Hr
Proof. Let 1 ≤ s < 2. It is an immediate consequence of the space Ls/2(M) being a s/2-normed
space that if a ∈ hd
. The general case can be achieved
by interpolations. We appeal to a result from [2, Theorem 2 and Remark 4] which asserts that if
qE < q < 2, then E ∈ Int(L1, Lq).
s(M) then max(cid:8)(cid:13)(cid:13)a(cid:13)(cid:13)Hc
,(cid:13)(cid:13)a(cid:13)(cid:13)Hr
s(cid:9) ≤ (cid:13)(cid:13)a(cid:13)(cid:13)hd
E(cid:9) ≤ cE(cid:13)(cid:13)a(cid:13)(cid:13)hd
.
E
E
s
s
a contraction.
ι : hd
that ι : hd
When 1 ≤ s < 2, it follows from above that the formal identity ι : hd
If we denote by the natural isometry of Hc
s(M) is
s(M) into Ls(M⊗B(ℓ2(N))) then
s(M) → Ls(M⊗B(ℓ2(N))) is a contraction. We can now deduce from Proposition 2.8
E(M). This shows
1(M) is not necessarily
q(M)). This explains
(cid:3)
E(M) → E(M⊗B(ℓ2(N))) is bounded whose range sits in Hc
. We should point out here that in general, Hc
2) and we do not know if Hc
. (cid:13)(cid:13)a(cid:13)(cid:13)hd
E(M) ∈ Int(Hc
s(M) → Hc
1(M), Hc
E
s(M) ([29, Theorem 7.1]), it is
reasonable to assume that the corresponding statements to the row/column are also valid. That
for some constant cE. But we were unable to verify
for all a ∈ hc
s
s
≤ 21/s(cid:13)(cid:13)a(cid:13)(cid:13)hc
E
complemented in L1(M; ℓc
the role of in our argument.
that (cid:13)(cid:13)a(cid:13)(cid:13)Hc
Since for 1 ≤ s < 2, we have (cid:13)(cid:13)a(cid:13)(cid:13)Hc
is, (cid:13)(cid:13)a(cid:13)(cid:13)Hc
and (cid:13)(cid:13)a(cid:13)(cid:13)Hr
≤ cE(cid:13)(cid:13)a(cid:13)(cid:13)hc
these inequalities at this time.
E
E
E
≤ cE(cid:13)(cid:13)a(cid:13)(cid:13)hr
E
The following consequence of Theorem 2.10 now follows from Proposition 4.3 and the facts that
hc
p(M) ⊂ Hc
p(M) for 1 ≤ p < 2. It provides simultaneous decompositions
for martingale Hardy spaces norms that are related to the noncommutative Burkholder-Gundy
inequalities from [39].
p(M) and hr
p(M) ⊂ Hr
Corollary 4.4. There exists a family of constants {κ′
if x ∈ L1(M) ∩ L2(M), then there exist martingales y ∈ ∩1<p<2Hc
such that:
p : 1 < p < 2} ⊂ R+ satisfying the following:
p(M)
p(M) and z ∈ ∩1<p<2Hr
(i) x = y + z;
(ii) for every 1 < p < 2, the following inequality holds:
≤ κ′
(cid:13)(cid:13)y(cid:13)(cid:13)Hc
p
+(cid:13)(cid:13)z(cid:13)(cid:13)Hr
p
p(cid:13)(cid:13)x(cid:13)(cid:13)p.
A direct alternative approach to Corollary 4.4 is to use the weak-type inequality for square
functions from [42] (see also [37]) and then follow the same line of reasoning as in the proof of
MARTINGALE INEQUALITIES
23
Theorem 2.10. The above result is closely related to another type of simultaneous decompositions
considered by Junge and Perrin in [28, Theorem 3.3].
We note that Part I of the proof of Theorem 3.1 can also be deduced from Proposition 4.3 and
the noncommutative Burkholder-Gundy inequalities for symmetric spaces from [14] and [22].
As a final remark, we provide an example involving Orlicz functions. Let Φ be an Orlicz
function on [0, ∞) i.e, a continuous increasing convex function on [0, ∞) with Φ(0) = 0 and
limt→∞ Φ(t) = ∞. Two standard indices associated to the Orlicz function Φ are defined as
follows: let
MΦ(t) = sup
s>0
Φ(ts)
Φ(s)
,
t ∈ [0, ∞)
and
pΦ := lim
t→0+
log MΦ(t)
log t
,
qΦ := lim
t→∞
log MΦ(t)
log t
.
Then 1 ≤ pΦ ≤ qΦ ≤ ∞. The Orlicz space LΦ is the set of all Lebesgue measurable functions f
defined on (0, ∞) such that for some constant c > 0,
Z ∞
0
Φ(cid:16)f (t)/c(cid:17) dt < ∞.
If we equip LΦ with the Luxemburg norm:
= infnc > 0 :Z ∞
0
Φ(cid:16)f (t)/c(cid:17) dt ≤ 1o,
(cid:13)(cid:13)f(cid:13)(cid:13)LΦ
then LΦ is a symmetric Banach function space with the Fatou property. Moreover, the Boyd
indices of LΦ coincide with the indices pΦ and qΦ (see for instance [35]). Thus, Theorem 3.1
and Corollary 4.1 apply to martingales in the noncommutative space LΦ(M) whenever 1 < pΦ ≤
qΦ < 2 or 2 < pΦ ≤ qΦ < ∞. This example also motivates the consideration of the so-called
Φ-moment inequalities involving conditioned square functions. This direction will be explored
in a forthcoming article. We refer to [3, 13, 15] for recent progress on moment inequalities for
noncommutative martingales.
Acknowledgements. This work was carried out during the second-named author's visit to
Miami University. She would like to express her gratitude to the Department of Mathematics of
Miami University for its warm hospitality. The authors are indebted to the referee for helpful
suggestions that improved the presentation of the paper.
References
[1] S. Astashkin, F. Sukochev, and C. P. Wong, Disjointification of martingale differences and conditionally
independent random variables with some applications, Studia Math. 205 (2011), no. 2, 171 -- 200.
[2] S. V. Astashkin and L. Maligranda, Interpolation between L1 and Lp, 1 < p < ∞, Proc. Amer. Math. Soc.
132 (2004), no. 10, 2929 -- 2938 (electronic).
[3] T. Bekjan and Z. Chen, Interpolation and Φ-moment inequalities of noncommutative martingales, Probab.
Theory Related Fields 152 (2012), no. 1-2, 179 -- 206.
[4] C. Bennett, Banach function spaces and interpolation methods. I. The abstract theory, J. Funct. Anal. 17
(1974), 409 -- 440.
[5] C. Bennett, Banach function spaces and interpolation methods. II. Interpolation of weak-type operators, Linear
operators and approximation, II (Proc. Conf., Math. Res. Inst., Oberwolfach, 1974), Birkhauser, Basel, 1974,
pp. 129 -- 139. Internat. Ser. Numer. Math., Vol. 25.
[6] C. Bennett and R. Sharpley, Interpolation of operators, Academic Press Inc., Boston, MA, 1988.
[7] J. Bergh and J. Lofstrom, Interpolation spaces. An introduction, Springer-Verlag, Berlin, 1976, Grundlehren
der Mathematischen Wissenschaften, No. 223.
[8] D. L. Burkholder, Distribution function inequalities for martingales, Ann. Probab. 1 (1973), 19 -- 42.
24
N. RANDRIANANTOANINA AND L. WU
[9] D. L. Burkholder and R. F. Gundy, Extrapolation and interpolation of quasi-linear operators on martingales,
Acta Math. 124 (1970), 249 -- 304.
[10] V. I. Chilin and F. A. Sukochev, Symmetric spaces over semifinite von Neumann algebras, Dokl. Akad. Nauk
SSSR 313 (1990), 811 -- 815.
[11] J. Diestel, Sequences and series in Banach spaces, Graduate Text in Mathematics, 92, Springer-Verlag, New
York, 1984.
[12] S. Dirksen, Noncommutative Boyd interpolation theorems, Trans. Amer. Math. Soc, 367 (2015), no 6, 4079 --
4110.
[13] S. Dirksen, Weak-type interpolation for noncommutative maximal operators, arXiv:1212.5168v2 (2013).
[14] S. Dirksen, B. de Pagter, D. Potapov, and F. Sukochev, Rosenthal inequalities in noncommutative symmetric
spaces, J. Funct. Anal. 261 (2011), no. 10, 2890 -- 2925.
[15] S. Dirksen and E. Ricard, Some remarks on noncommutative Khintchine inequalities, Bull. Lond. Math. Soc.
45 (2013), no. 3, 618 -- 624.
[16] P. G. Dodds, T. K. Dodds, and B. de Pagter, Noncommutative Banach function spaces, Math. Z. 201 (1989),
583 -- 597.
[17] P. G. Dodds, T. K. Dodds, and B. de Pagter, Fully symmetric operator spaces, Integral Equations Operator
Theory 15 (1992), no. 6, 942 -- 972.
[18] P. G. Dodds, T. K. Dodds, and B. de Pagter, Noncommutative Kothe duality, Trans. Amer. Math. Soc. 339
(1993), 717 -- 750.
[19] T. Fack and H. Kosaki, Generalized s-numbers of τ -measurable operators, Pacific J. Math. 123 (1986), 269 -- 300.
[20] A. M. Garsia, Martingale inequalities: Seminar notes on recent progress, W. A. Benjamin, Inc., Reading,
Mass.-London-Amsterdam, 1973, Mathematics Lecture Notes Series.
[21] Y. Jiao, Burkholder's inequalities in noncommutative Lorentz spaces, Proc. Amer. Math. Soc. 138 (2010),
no. 7, 2431 -- 2441.
[22] Y. Jiao, Martingale inequalities in noncommutative symmetric spaces, Arch. Math. (Basel) 98 (2012), no. 1,
87 -- 97.
[23] W. B. Johnson, B. Maurey, G. Schechtman, and L. Tzafriri, Symmetric structures in Banach spaces, Mem.
Amer. Math. Soc. 19 (1979), no. 217, v+298.
[24] W. B. Johnson and G. Schechtman, Martingale inequalities in rearrangement invariant function spaces, Israel
J. Math. 64 (1988), no. 3, 267 -- 275 (1989).
[25] W. B. Johnson and G. Schechtman, Sums of independent random variables in rearrangement invariant function
spaces, Ann. Probab. 17 (1989), no. 2, 789 -- 808.
[26] M. Junge, Doob's inequality for non-commutative martingales, J. Reine Angew. Math. 549 (2002), 149 -- 190.
[27] M. Junge and T. Mei, Noncommutative Riesz transforms -- a probabilistic approach, Amer. J. Math. 132 (2010),
no. 3, 611 -- 680.
[28] M. Junge and M. Perrin, Theory of Hp-spaces for continuous filtrations in von Neumann algebras, Ast´erisque
(2014), no. 362, vi+134.
[29] M. Junge and Q. Xu, Noncommutative Burkholder/Rosenthal inequalities, Ann. Probab. 31 (2003), no. 2,
948 -- 995.
[30] M. Junge and Q. Xu, Noncommutative Burkholder/Rosenthal inequalities. II. Applications, Israel J. Math.
167 (2008), 227 -- 282.
[31] N. Kalton and S. Montgomery-Smith, Interpolation of Banach spaces, Handbook of the geometry of Banach
spaces, Vol. 2, North-Holland, Amsterdam, 2003, pp. 1131 -- 1175.
[32] N. J. Kalton and F. A. Sukochev, Symmetric norms and spaces of operators, J. Reine Angew. Math. 621
(2008), 81 -- 121.
[33] C. Le Merdy and F. Sukochev, Rademacher averages on noncommutative symmetric spaces, J. Funct. Anal.
255 (2008), no. 12, 3329 -- 3355.
[34] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces. II, Springer-Verlag, Berlin, 1979, Function spaces.
[35] L. Maligranda, Indices and interpolation, Dissertationes Math. (Rozprawy Mat.) 234 (1985), 49.
[36] E. Nelson, Notes on non-commutative integration, J. Funct. Anal. 15 (1974), 103 -- 116.
[37] J. Parcet and N. Randrianantoanina, Gundy's decomposition for non-commutative martingales and applica-
tions, Proc. London Math. Soc. (3) 93 (2006), no. 1, 227 -- 252.
[38] M. Perrin, A noncommutative Davis' decomposition for martingales, J. Lond. Math. Soc. (2) 80 (2009), no. 3,
627 -- 648.
[39] G. Pisier and Q. Xu, Non-commutative martingale inequalities, Comm. Math. Phys. 189 (1997), 667 -- 698.
MARTINGALE INEQUALITIES
25
[40] G. Pisier and Q. Xu, Non-commutative Lp-spaces, Handbook of the geometry of Banach spaces, Vol. 2, North-
Holland, Amsterdam, 2003, pp. 1459 -- 1517.
[41] N. Randrianantoanina, Non-commutative martingale transforms, J. Funct. Anal. 194 (2002), 181 -- 212.
[42] N. Randrianantoanina, A weak type inequality for non-commutative martingales and applications, Proc. London
Math. Soc. (3) 91 (2005), no. 2, 509 -- 542.
[43] N. Randrianantoanina, Conditioned square functions for noncommutative martingales, Ann. Probab. 35 (2007),
no. 3, 1039 -- 1070.
[44] H. P. Rosenthal, On the subspaces of Lp (p ≥ 2) spanned by sequences of independent random variables, Israel
J. Math. 8 (1970), 273 -- 303.
[45] M. Takesaki, Theory of operator algebras. I, Springer-Verlag, New York, 1979.
[46] Q. Xu, Analytic functions with values in lattices and symmetric spaces of measurable operators, Math. Proc.
Cambridge Philos. Soc. 109 (1991), 541 -- 563.
Department of Mathematics, Miami University, Oxford, Ohio 45056, USA
E-mail address: [email protected]
School of Probability and Statistics, Central South University, Changsha 410075, China and
Department of Mathematics, Miami University, Oxford, Ohio 45056, USA
E-mail address: [email protected]
|
1206.3466 | 1 | 1206 | 2012-06-15T13:55:14 | Ideals and hereditary subalgebras in operator algebras | [
"math.OA",
"math.FA"
] | This paper may be viewed as having two aims. First, we continue our study of algebras of operators on a Hilbert space which have a contractive approximate identity, this time from a more Banach algebraic point of view. Namely, we mainly investigate topics concerned with the ideal structure, and hereditary subalgebras (HSA's), which are in some sense generalization of ideals. Second, we study properties of operator algebras which are hereditary subalgebras in their bidual, or equivalently which are `weakly compact'. We also give several examples answering natural questions that arise in such an investigation. | math.OA | math |
IDEALS AND HEREDITARY SUBALGEBRAS IN OPERATOR
ALGEBRAS
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
Abstract. This paper may be viewed as having two aims. First, we continue
our study of algebras of operators on a Hilbert space which have a contrac-
tive approximate identity, this time from a more Banach algebraic point of
view. Namely, we mainly investigate topics concerned with the ideal structure,
and hereditary subalgebras (HSA's), which are in some sense generalization of
ideals. Second, we study properties of operator algebras which are hereditary
subalgebras in their bidual, or equivalently which are 'weakly compact'. We
also give several examples answering natural questions that arise in such an
investigation.
1. Introduction
For us, an operator algebra is a norm closed algebra of operators on a Hilbert
space. This paper may be viewed as having two aims. First, we continue our study
of the structure of operator algebras which have a contractive approximate identity
(cai), this time from a slightly more Banach algebraic point of view. Indeed this
paper may be seen as a collection of general results growing out of topics raised in [3]
concerning ideals and hereditary subalgebras (HSA's) of operator algebras. We recall
that HSA's are in some sense a generalization of ideals (a HSA D in A must satisfy
DAD ⊂ D) . Some of these general results are of a technical nature, and so this
paper should serve in part as a repository that will be useful in later development of
the themes of interest here. Second, we study properties of operator algebras which
are hereditary subalgebras in their bidual (this is equivalent to A being 'weakly
compact', see e.g. Lemma 5.1 below). We give several examples answering natural
questions that arise in such an investigation. For example, questions involving
semisimplicity or semiprimeness (we recall that an algebra A is semisimple if its
Jacobson radical J(A) = (0), and is semiprime if (0) is the only (closed) ideal with
square (0)).
In Section 2 (resp. Section 4) of the paper we present some general results about
ideals (resp. HSA's) in operator algebras. In Section 3 we discuss adjoining a square
root to an operator algebra, and use this to answer some natural questions. We
also discuss in Section 3 whether for an approximately unital operator algebra being
semisimple (resp. semiprime, radical) implies or is implied by A∗∗ being semisimple
(or semiprime, radical). (We do not think we know yet if A∗∗ semisimple implies
A semisimple if A is noncommutative.) In Section 5 we study operator algebras
A which are hereditary subalgebras of their bidual, which as we said above is
Date: Revision of June 13, 2012.
Key words and phrases. Operator algebras, one-sided ideals, hereditary subalgebra, approxi-
mate identity, semisimple algebra, semiprime algebra.
*Blecher was partially supported by a grant from the National Science Foundation.
1
2
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
equivalent to the multiplication x 7→ axb being weakly compact on A for all a, b ∈ A.
Some of the properties of such algebras are similar to operator algebras which are
one-sided ideals in their bidual, which were studied in [40, 3]. We also study the
more general class of algebras that we call nc-discrete, which means that all the open
projections are also closed (or equivalently lie in the multiplier algebra M (A)). Any
compact operator algebra is a HSA in its bidual; and we show that any operator
algebra which is a HSA in its bidual is nc-discrete. Neither of these two implications
are reversible though, as may be seen in Examples 6.1, 6.2, and Theorem 6.4. Indeed
in Section 6 we present examples of operator algebras exhibiting various properties
illustrating the topics of interest in this paper. In particular we give what is as far
as we can see the first explicit example in the literature of an interesting (i.e. not
reflexive in the Banach space sense) commutative algebra whose multiplication is
weakly compact but not compact. In Section 7 we discuss the diagonal of a quotient
algebra.
Some of the topics in the earlier parts of this paper and in Section 7 were inves-
tigated in the first author's Ph.D. thesis [2]. Some related results and topics may
be found there too. Others of our results related to HSA's have been presented at
various venues in 2010.
We now turn to notation and more precise definitions. The reader is referred for
example to [7, 6, 10] for more details on some of the topics below if needed. By
an ideal of an operator algebra A we shall always mean a closed two-sided ideal
in A. For us a projection is always an orthogonal projection, and an idempotent
merely satisfies x2 = x. If X, Y are sets, then XY denotes the closure of the span
of products of the form xy for x ∈ X, y ∈ Y . We recall that by a theorem due to
Ralf Meyer, every operator algebra A has a unique unitization A1 (see [7, Section
2.1]). Below 1 always refers to the identity of A1 if A has no identity.
If A is
a nonunital operator algebra represented (completely) isometrically on a Hilbert
space H then one may identify A1 with A + C IH . The second dual A∗∗ is also an
operator algebra with its (unique) Arens product, this is also the product inherited
from the von Neumann algebra B∗∗ if A is a subalgebra of a C∗-algebra B. Meets
and joins in B∗∗ of projections in A∗∗ remain in A∗∗, as can be readily seen for
example by inspecting some of the classical formulae for meets and joins of Hilbert
space projections, or by noting that these meets and joins may be computed in the
biggest von Neumann algebra contained inside A∗∗. Note that A has a cai iff A∗∗
has an identity 1A∗∗ of norm 1, and then A1 is sometimes identified with A+ C 1A∗∗.
In this case the multiplier algebra M (A) is identified with the idealizer of A in A∗∗
(that is, the set of elements α ∈ A∗∗ such that αA ⊂ A and Aα ⊂ A).
The set of compact operators on a Hilbert space is often called an elementary
C∗-algebra. We call a c0-direct sum of elementary C∗-algebras an annihilator C∗-
algebra.
The diagonal ∆(A) is defined to be A∩A∗, it is a C∗-algebra which is well defined
independently of the particular (completely isometric) representation of A. Most of
our algebras and ideals are approximately unital, i.e. have a contractive approximate
identity (cai), although for some results this is probably not necessary. We recall
that an r-ideal is a right ideal with a left cai, and an ℓ-ideal is a left ideal with a
right cai. We say that an operator algebra D with cai, which is a subalgebra of
another operator algebra A, is a HSA (hereditary subalgebra) in A, if DAD ⊂ D.
See [6] for the theory of HSA's (a few more results may be found in [3, 10]). HSA's
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
3
in A are in an order preserving, bijective correspondence with the r-ideals in A, and
also with the open projections p ∈ A∗∗, by which we mean that there is a net xt ∈ A
with xt = pxtp → p weak*. These are also the open projections p in the sense of
Akemann [1] in B∗∗, where B is a C∗-algebra containing A, such that p ∈ A⊥⊥.
The complement ('perp') of an open projection is called a closed projection. We
spell out some of the correspondences above: if D is a HSA in A, then DA (resp.
AD) is the matching r-ideal (resp. ℓ-ideal), and D = (DA)(AD) = (DA) ∩ (AD).
The weak* limit of a cai for D, or of a left cai for an r-ideal, is an open projection,
and is called the support projection. Conversely, if p is an open projection in A∗∗,
then pA∗∗ ∩ A and pA∗∗p ∩ A is the matching r-ideal and HSA pair in A.
It is a well-known fact that if J is an ideal of an operator algebra A, then the
quotient algebra A/J is isometrically isomorphic to an operator algebra [7, Propo-
sition 2.3.4]. Of course there is a 'factor theorem':
if u : A → B is a completely
bounded homomorphism between operator algebras, and if J is an ideal in A con-
tained in Ker(u), then the canonical map u : A/J → B is also completely bounded
with completely bounded norm kukcb = kukcb. If J = Ker(u), then u is a complete
quotient map if and only if u is a completely isometric isomorphism.
Let A be an operator algebra. The set FA = {x ∈ A : k1 − xk ≤ 1} equals
{x ∈ A : k1−xk = 1} if A is nonunital, whereas if A is unital then FA = 1+Ball(A).
Many properties of FA are developed in [10, 11]. If A is a closed subalgebra of an
operator algebra B then it is easy to see, using the uniqueness of the unitization,
that FA = A ∩ FB.
We write J(A) for the Jacobson radical (see e.g. [35]). It is a fact in pure algebra
that an algebra is semiprime (resp. semisimple) iff its unitization is semiprime (resp.
semisimple). Indeed J(A) = J(A1) (see [35, 4.3.3]). One trap to beware of is that
the C∗-algebra generated by a HSA D in an operator algebra A need not be a
HSA in a C∗-algebra generated by A. In particular C∗(D) need not be an HSA in
C∗
e (A). An example is U(M2), the subalgebra of M2(A) with 0 in
the 2-1-entry, and scalar multiples of the identity on the main diagonal, in the case
A = M2.
max(A) or in C∗
2. General results on ideals in operator algebras
The first two results that follow are obvious, and follow from the analogous
results for general operator spaces.
Theorem 2.1 (First Isomorphism Theorem). Let u : A → B be a complete quotient
map which is a homomorphism between operator algebras. Then, Ker(u) is an ideal
in A and A/Ker(u) ∼= B completely isometrically isomorphically. Conversely, every
ideal of A is of the form Ker(u) for a complete quotient map u : A → B, where A
and B are operator algebras.
Theorem 2.2 (Second Isomorphism Theorem). Let A be an approximately unital
operator algebra, let J be an ideal in A, and suppose that I is an ideal in J. Then,
(A/I)/(J/I) ∼= A/J completely isometrically isomorphically (as operator algebras).
Theorem 2.3 (Third Isomorphism Theorem). Let A be an approximately unital
operator algebra, and suppose that J and K are ideals in A, where J has a cai.
Then, J/(J ∩ K) ∼= (J + K)/K completely isometrically isomorphically. In partic-
ular, (J + K)/K is closed.
4
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
Proof. Note that by [19, Proposition 2.4], J + K is closed. Define a map u :
J/(J∩K) → (J +K)/K by u(j +J∩K) = j +K. This is a well-defined map and u is
one-to-one since Ker(u) = (0J/(J∩K)). Moreover, u is onto since x+K ∈ (J +K)/K
implies that x = j + k, where j ∈ J, k ∈ K and x + K = j + K = u(j + J ∩ K).
Since inf {kj + kk : k ∈ K} ≤ inf {kj + kk : k ∈ J ∩ K}, u is a contraction. Let
(et) be the cai for J and let k ∈ K. Then,
kj + kk ≥ ketj + etkk ≥ ketj + J ∩ Kk .
After taking the limit, we get kj + J ∩ Kk ≤ kj + kk, and so kj + J ∩ Kk ≤
kj + Kk. Hence, u is an isometry. Similarly, u is a complete isometry.
(cid:3)
For Banach algebras the 'Correspondence Theorem' states that for a Banach
algebra A and a closed ideal J in A, every closed subalgebra K of A/J is of the
form I/J, where I is a closed subalgebra of A with J ⊂ I ⊂ A. Also, every ideal
K of A/J is of the form I/J, where I is an ideal of A with J ⊂ I ⊂ A. Indeed
I = q−1(K) where q : A → A/J is the canonical map.
Theorem 2.4. Let A be an approximately unital operator algebra, let I be an ap-
proximately unital ideal in A and let J be an approximately unital ideal in I. Then,
I/J is an approximately unital ideal in A/J. Conversely, every approximately uni-
tal ideal of A/J is of the form I/J, where I is an approximately unital ideal in A
with J ⊂ I ⊂ A.
Proof. The first assertion is easy. The second assertion follows from [12, Proposition
3.1] (as in e.g. [10, Section 6], where the analogue of the above result is proved for
HSA's and certain one-sided ideals).
(cid:3)
Remark. Note that since [12, Proposition 3.1] is also valid for Arens regular Ba-
nach algebras, the previous result can be stated for such Banach algebras. Similarly
for Corollary 2.6 below.
Lemma 2.5. Suppose that A is an operator algebra such that A∗∗ is semiprime,
and that J is a closed ideal in A such that J 2 has a cai. Then J = J 2.
Proof. Let p be the support projection of J 2 in A∗∗. We have J 2(1 − p) = 0. If
ζ, η ∈ J ⊥⊥ and if at → η weak* and bs → ζ, then since atbs(1 − p) = 0 we have
atζ(1 − p) = 0 and 0 = ηζ(1 − p) = η(1 − p)ζ(1 − p). So (J ⊥⊥(1 − p))2 = (0). Since
A∗∗ is semiprime we have J ⊥⊥(1 − p) = (0), so that J ⊥⊥ is unital, so that J has a
cai.
(cid:3)
Corollary 2.6. Suppose that A is an approximately unital operator algebra, and
that J is an approximately unital ideal in A. Suppose that J has the property that
if I is an ideal with square J then I = J. Then A/J is semiprime. In particular,
if A∗∗ is semiprime then A/J is semiprime, for every approximately unital ideal J
in A.
Proof. Let K be an ideal in A/J such that K 2 = (0)A/J . There exists an ideal I
in A such that J ⊂ I ⊂ A and K = I/J (namely, the inverse image of K in A, see
Theorem 2.4). Since K 2 = (I/J)(I/J) = I 2/J = (0)A/J , we conclude that I 2 = J.
Under our hypotheses this forces I = J (the 'In particular' assertion uses Lemma
2.5 here). That is, K = I/J = (0)A/J , and so A/J is semiprime.
(cid:3)
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
5
Remarks. 1)
In view of the last results, and independently, it is of interest
to know whether every approximately unital ideal J in a semisimple or semiprime
algebra A has the property that if I is an ideal with square J then I = J. We will
see in the next section after Lemma 3.1 that this is false.
2) Algebras whose square (or nth power) is approximately unital are discussed
in [11, Section 3].
Proposition 2.7. Let A be a Banach algebra with no nonzero left annihilators,
and let {Iα} be an increasing family of ideals in A such that A = ∪Iα. If each Iα
is semiprime (resp. semisimple), then A is semiprime (resp. semisimple).
Proof. For each α, assume that Iα is semiprime. Let J be an ideal in A with
J 2 = (0). Then (J ∩ Iα)2 = (0) for each α. Since Iα is semiprime, J ∩ Iα = (0).
Hence JIα ⊂ J ∩ Iα = (0) for each α, so that JA = (0) and J = (0). Hence A is
semiprime.
Now suppose that each Iα is semisimple. Notice that J = Rad(A) is an ideal in
A and J ∩ Iα is an ideal in Iα for each α. Then J ∩ Iα = Rad(Iα) = (0) by [35,
Theorem 4.3.2]. Hence as in the first paragraph, J = (0) and A is semisimple. (cid:3)
Proposition 2.8. Suppose that A is an operator algebra with a bai, and that I and
J are ideals in A. If A/I ∼= A/J isomorphically as A-bimodules, then I = J.
Proof. If A is unital then this follows from the analogous result in pure algebra. If
A contains a bai, then its bidual A∗∗ is unital. If π : A/I → A/J is an A-bimodule
isomorphism, then π∗∗ maps (A/I)∗∗ ∼= A∗∗/I ⊥⊥ into (A/J)∗∗ ∼= A∗∗/J ⊥⊥. More-
over, π∗∗ is an A∗∗-bimodule isomorphism by the separate weak∗-continuity of the
Arens product. Since A∗∗ is unital, by the unital case we have I ⊥⊥ = J ⊥⊥. Thus
I = A ∩ I ⊥⊥ = A ∩ J ⊥⊥ = J.
(cid:3)
Remark. Note that the previous proposition is not true for general operator
algebras; the existence of a bai is needed. For example, let A = Span(x, y) where
If I = Span(x) and J = Span(y), then A/I ∼= A/J as
xy = x2 = y2 = 0.
A-bimodules, but I 6= J.
3. Example: adjoining a root to an algebra
In this section we show how to create examples of operator algebras by adjoining
a root. We then use this to answer several basic questions regarding operator
algebras with cai.
If A is an algebra, and S is in the center of A, we define an algebra
AS = (cid:26)(cid:20) x
Sy x (cid:21) : x ∈ A, y ∈ A1(cid:27) ⊂ M2(A1).
y
We identify A with the main diagonal of this algebra, and we set T to be the
matrix with rows 0, 1 and S, 0. Then any element of AS may be written as x + yT
for x ∈ A, y ∈ A1. In this notation, T 2 = S, and so now S has a square root even
if it did not have one before. A good example to bear in mind is the case that A
is the approximately unital ideal in the disk algebra A(D) of functions vanishing at
1, and S = z(1 − z) ∈ A, which has no root in A.
It is obvious that if A is an operator algebra then so is AS, and if A is com-
mutative then so is AS. If A has a cai but no identity then AS has no cai, but
A2
S = {x + yT : x, y ∈ A} does have a cai.
6
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
We say that an element a in an algebra A has no rational square root if there
exists no b, c ∈ A, with ac2 = b2 6= 0.
Lemma 3.1. Suppose that A is a commutative semisimple algebra, and S ∈ A, S 6=
0. If A is an integral domain then AS is semiprime. On the other hand, if A is
semisimple and S has no rational square root, and is not a divisor of zero, then AS
is semisimple.
Proof. If (x + yT )2 = x2 + y2S + 2xyT = 0, and A is an integral domain, then
x = 0 or y = 0. Since x2 + y2S = 0 we have x = y = 0.
In the semisimple case, suppose that all characters of AS vanish at x + yT ∈ AS.
If χ is a character of A1, define χ′(a + bT ) = χ(a) + αχ(b) for a ∈ A, b ∈ A1,
where α is a square root of χ(S). This defines a character on AS, and so we have
χ(x) + αχ(y) = 0. Thus x2 − Sy2 is in the kernel of every character of A1, so that
Sy2 = x2 . Thus x2 = 0, and since A is semiprime we have x = y = 0.
(cid:3)
In our disk algebra example mentioned above, the element z(1 − z) ∈ A has no
rational square root. To see this note that of course 1 − z does, and so we are
asking if zg2 = h2 is possible with g, h ∈ A(D). By Riemann's theorem in basic
complex analysis this equation implies that h/g has an analytic continuation k to
D such that k(z)2 = z on D, which is well known to be impossible (2k(z)k′(z) = 1
so k′ is unbounded at 0). We deduce from Lemma 3.1 that AS is semisimple if
S = z(1 − z). Here AS is a semisimple commutative operator algebra with no cai,
but A2
S has a cai. This solves the question posed in the Remark after Corollary 2.6
in the negative.
In [16] it is shown that semisimple B(ℓp) fails to have a semisimple second dual
if p 6= 2. This can also happen for operator algebras:
Proposition 3.2. Let A be an operator algebra.
(1) If A∗∗ is semiprime (resp. radical, semisimple and commutative) then A is
semiprime (resp. radical, semisimple).
(2) If A is semiprime (resp. approximately unital and radical, unital and semisim-
ple) then A∗∗ need not be semiprime (resp. radical, semiprime and hence
not semisimple).
If A is commutative and A∗∗ is semisimple, then A is semisimple by
Proof. (1)
e.g. [14, Proposition 2.6.25 (iv)]. If A∗∗ is semiprime and if J 2 = (0) in A then
(J ⊥⊥)2 = (0) in A∗∗, so that (0) = J ⊥⊥ = J. So A is semiprime.
It follows from [14, Proposition 2.6.25 (iii)] that if A∗∗ is radical then A is radical.
(2) If A is radical then A∗∗ is not radical (indeed A∗∗ is unital).
Suppose that the second dual of every unital semiprime operator algebra A was
semiprime. Then by Lemma 2.5, if J is an ideal in a unital semisimple operator
algebra such that J 2 has a cai, then J has a cai. However this is not true, as
may be seen from the disk algebra example two paragraphs above (one may take
J = AS, A = J 1 here). Hence the second dual of a commutative unital semisimple
operator algebra need not be semiprime.
(cid:3)
We shall see in Corollary 5.3 that the situation in the last result improves if A
is a HSA in its bidual.
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
7
4. General facts about HSA's
Proposition 4.1. If D is a HSA in an operator algebra A, then x ∈ D is quasi-
invertible in D iff x is quasi-invertible in A. Thus σD(x) \ {0} = σA(x) \ {0}. In
particular, D is a spectral subalgebra of A in the sense of e.g. [35, p. 245].
Proof. By [14, Proposition 2.6.25], x is quasi-invertible in D (resp. A) iff x is quasi-
invertible in D∗∗ (resp. A∗∗). Now D∗∗ = pA∗∗p, where p is the support projection
of D. It is a simple algebraic exercise that in an algebra A with an idempotent e,
an element of eAe is quasi-invertible in eAe iff it is quasi-invertible in A. So x is
quasi-invertible in D iff x is quasi-invertible in D∗∗ = pA∗∗p, iff x is quasi-invertible
in A∗∗, and hence iff x is quasi-invertible in A.
The last assertion follows from the usual formulation of the spectrum in terms
(cid:3)
of quasi-invertible elements.
The following answers a question posed in [3]. The first assertion is well known
with HSA's replaced by ideals [35].
Theorem 4.2. If D is a HSA in an operator algebra A, then J(D) = D ∩ J(A).
In particular, semisimplicity passes to HSA's.
Proof. Suppose that A is an operator algebra and D is a HSA in A. We recall that
J(A) may be characterized (see e.g. [35]) as the set of a ∈ A with r(ab) = 0 for all
b ∈ A1. Here r(·) denotes the spectral radius. Let x ∈ J(D) then since J(D) is
a nondegenerate D-module, by Cohen's factorization there exists d ∈ D, y ∈ J(D)
with x = dy. Now yft ad ∈ J(D) for all a ∈ A1, where (ft) is a cai for D (since D
is a HSA in A1). Since J(D) is closed we have yad ∈ J(D). Thus 0 = r(yad) =
r(dya) = r(xa) for all a ∈ A1. Hence x ∈ J(A). So J(D) ⊂ D ∩ J(A). The
if x ∈ D ∩ J(A), then A1x
converse follows from [35, Theorem 4.3.6 (c),(e)]:
consists of quasi-invertibles in A. Hence D1x consists of quasi-invertibles in A,
hence of quasi-invertibles in D by Proposition 4.1. So x ∈ J(D).
(cid:3)
We have a generalization of the last result:
Corollary 4.3. Suppose that D is a HSA in an operator algebra A, and that I is
an approximately unital ideal in A. Then
(1) D ∩ I = DID is a HSA in A, and J(D ∩ I) = J(D) ∩ J(I).
(2) (D ∩ I)⊥⊥ = D⊥⊥ ∩ I ⊥⊥.
(3) D + I is closed, and is a HSA in A.
Proof. (1) We have that (D ∩ I)A(D ∩ I) ⊂ (DAD) ∩ (IAI) ⊂ D ∩ I. Note
that DID ⊂ I ∩ D. Conversely, since D has a cai we have I ∩ D ⊂ DID. So
DID = I ∩ D. If (fs) is a cai for I and (eλ) is a cai for D, then (eλfseλ) is easily
seen to yield a cai for DID, by routine techniques. so I ∩ D = DID is a HSA in
A. By Theorem 4.2 we have
J(D ∩ I) = D ∩ I ∩ J(A) = D ∩ J(A) ∩ I ∩ J(A) = J(D) ∩ J(A)
as desired.
(3) Write (eλ) for the cai of D. If r ∈ I then eλreλ ∈ D ∩ I. Moreover if a ∈ D
then
ka − rk ≥ keλaeλ − eλreλk ≥ ka − eλreλk − ka − eλaeλk.
8
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
Also, ka − eλreλk ≥ ka + (I ∩ D)k. The above constitute the modifications of the
proof of [19, Proposition 2.4] that need to be made so that as in that proof we may
deduce that D + I is closed. Clearly (D + I)A(D + I) ⊂ DAD + I = D + I.
for closed subspaces E, F of any Banach space X,
(2) This follows from (3) and the fact from e.g. [13, Appendix A.3, A.5] that
(E ∩ F )⊥⊥ = (E⊥ + F ⊥)⊥ = E⊥⊥ ∩ F ⊥⊥,
if E + F is closed, or equivalently, if E⊥⊥ + F ⊥⊥ is closed.
(cid:3)
Remark. If I is any ideal in a HSA D of an operator algebra A, and if I ⊂ J(A),
then J(D/I) = J(D)/I. This follows from Theorem 4.2 and [35, Theorem 4.3.2
(b)].
5. Algebras that are HSA's in their bidual
That is, A has a cai, and AA∗∗A ⊂ A. We write Ma,b : A → A : x 7→ axb, where
a, b ∈ A. Recall that a Banach algebra is compact if the map Ma,a is compact for
all a ∈ A. We say that A is weakly compact if Ma,a is weakly compact for all a ∈ A.
We are concerned here mostly with operator algebras A that are HSA's in their
bidual. That is, A has a cai, and AA∗∗A ⊂ A. For algebras A that do not have a
cai, one could pass to the algebra AH described in [11], the biggest subalgebra with
a cai.
Lemma 5.1. An operator algebra A with cai is a HSA in its bidual iff the map
Ma,b : A → A : x 7→ axb is weakly compact for all a, b ∈ A, and iff A is weakly
compact in the sense just defined.
Similarly, Ma,b is compact on A for all a, b ∈ A iff A is compact.
If in addition A is commutative, then A is compact (resp. weakly compact) iff
multiplication by a is compact (resp. weakly compact) on A for all a ∈ A.
Proof. The first 'iff' follows by basic functional analysis (namely, the well known
fact that an operator T : X → Y is weakly compact iff T ∗∗(X ∗∗) ⊂ Y ). To see
the second and third 'iff' we use the fact that the compact (resp. weakly compact)
operators constitute a norm closed ideal. From this, first, if (et) is a cai for A
and Met,et is compact (resp. weakly compact), then so is Maet,etb for all a, b ∈ A.
Second, Ma,b is compact (resp. weakly compact) since Maet,etb → Ma,b in norm. (cid:3)
It is easy
to find Banach space reflexive examples showing that the converse is not true (see
Example 6.1). Note that the class of unital operator algebras which are HSA's in
their bidual, is the same as the class of unital operator algebras which are Banach
space reflexive. It is of interest to find nonreflexive weakly compact algebras which
are not compact, and we shall do this later in Subsection 6.4. In this connection we
remark that semisimple annihilator Banach algebras in the sense of [35, Chapter 8]
are compact, and are ideals in their bidual [35, Corollary 8.7.14].
Clearly then compact operator algebras are HSA's in their bidual.
Remark. In any commutative operator algebra A, two natural ideals to consider
are those constituting the elements a ∈ A with multiplication by a being compact
or weakly compact on A.
The property of being a HSA in the bidual passes to subalgebras and quotients:
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
9
Lemma 5.2. Let A be an operator algebra which is weakly compact. If B is a
closed subalgebra of A, then B is weakly compact. If I is a closed ideal in A, then
A/I is weakly compact.
Proof. We leave this as an exercise for the reader.
(cid:3)
Remark. Similarly, if A is an approximately unital ideal in its bidual then so
is any closed subalgebra, or quotient by a closed ideal (see [40]).
Corollary 5.3. Suppose that A is an operator algebra which is a HSA in its bidual.
Then A is semisimple (resp. semiprime) iff A∗∗ is semisimple (resp. semiprime).
Proof. The one direction follows from Theorem 4.2 (resp. [3, Proposition 2.5]). If
A is semisimple, let 0 6= η ∈ J(A∗∗). Then AηA ⊂ J(A∗∗) ∩ A ⊂ J(A) = (0) (using
Theorem 4.2). So η = 0, and A∗∗ is semisimple. If A is semiprime, and if J is
an ideal in A∗∗ with J 2 = (0), then (J ∩ A)2 = (0), so that J ∩ A = (0). Hence
AJA = (0) since AJA ⊂ J ∩ A. Since a cai of A converges weak* to the identity of
A∗∗ we deduce that J = (0). So A∗∗ is semiprime.
(cid:3)
Proposition 5.4. If A is an operator algebra which is a HSA in its bidual, and if
A has no ideals (resp. no closed ideals, no closed ideals with a cai), then every ideal
(resp. closed ideal, closed ideal with a cai) in A∗∗ contains A.
Proof. If J is a nontrivial ideal in A∗∗, then as in the proof of Corollary 5.3, AJA ⊂
J ∩ A = A or (0), and the latter is impossible. Similarly for the closed ideal case.
Similarly for the case of a closed ideal J with a cai, because by Corollary 4.3 the
ideal J ∩ A of A is also a HSA in A∗∗, so has a cai.
(cid:3)
An operator algebra A with cai is nc-discrete if every right ideal which has a
left cai, is of the form eA for a projection e ∈ M (A). Equivalently, all the open
projections are also closed (or equivalently are in M (A)). The first part of the
following was independently noticed recently in [34], and no doubt by others:
Proposition 5.5. A C∗-algebra which is a HSA in its bidual, or is nc-discrete, is
an annihilator C∗-algebra.
Proof. One well known characterization of annihilator C∗-algebras is that every
commutative C∗-subalgebra D has maximal ideal space which is topologically dis-
crete. Thus the HSA case of the proposition follows by Lemma 5.2 and the fact that
if a C0(K) space is an ideal in its bidual, then K is topologically discrete. The nc-
discrete case for C∗-algebras is another well known characterization of annihilator
C∗-algebras.
(cid:3)
Proposition 5.6. If an operator algebra A is a HSA in its bidual (resp. is compact),
then so is KI (A) for any cardinal I. Also, the c0-direct sum of operator algebras
which are HSA's in their bidual (resp. nc-discrete, ∆-dual), is a HSA in its bidual
(resp. is nc-discrete, ∆-dual).
Proof. We leave this as an exercise.
(cid:3)
Remark. We said earlier that compact approximately unital Banach algebras
are HSA's in their bidual. We recall that a semisimple Banach algebra A is a
modular annihilator algebra iff no element of A has a nonzero limit point in its
spectrum [35, Theorem 8.6.4]), and iff for every a ∈ A multiplication on A by
a is a Riesz operator (see [35, Chapter 8]). If A is also commutative then this is
10
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
equivalent to the Gelfand spectrum of A being discrete [31, p. 400]. By [35, Chapter
8], compact semisimple algebras are modular annihilator algebras. We note that
any radical semiprime algebra is a modular annihilator algebra by [35, Theorem
8.7.2]. There are some interesting commutative radical algebras in [18] which are
modulator annihilator algebras, but they are probably not approximately unital nor
are ideals in their bidual. One may ask if for algebras that are HSA's in their bidual
(or even ideals in their bidual), is the spectrum of every element finite or countable?
We have examples of algebras which are HSA's in their bidual with elements having
spectrum which does have nonzero limit points (see Example 6.1). Such algebras
are not modular annihilator algebras, but are Duncan modular annihilator algebras
in the sense of [35, Chapter 8] (see also [20]). Any semisimple operator algebra with
the spectrum of any element finite or countable, is a Duncan modular annihilator
algebra [35]. If A is a commutative approximately unital operator algebra which is
an ideal in its bidual, one may ask if the spectrum of A (eg. the set of characters
of A) scattered? In this case, and if A is not reflexive in the Banach space sense,
then the spectrum of A∗∗ equals the one point compactification of the spectrum of
A (see Theorem 5.10 (4)).
In the converse direction, Duncan modular annihilator algebras, or semisimple
operator algebras with the spectrum of every element finite or countable, need not
be nc-discrete. An example is the space c. We are not sure if every (approximately
unital) semisimple modular annihilator operator algebra is nc-discrete, or is a HSA
in its bidual, although this seems unlikely, even in the commutative case. .
Theorem 5.7. If an operator algebra A is a HSA in A∗∗, and if ∆(A) acts non-
degenerately on A, then ∆(A)∗∗ = ∆(A∗∗) = ∆(M (A)). In particular, every pro-
jection in A∗∗ is both open and closed.
Proof. That ∆(A) acts nondegenerately on A implies that A has a positive cai (et)
say. If A is a HSA in A∗∗ then etηet ∈ ∆(A) for all η ∈ ∆(A∗∗)+. If we repre-
sent A∗∗ as a weak* closed subalgebra of H containing IH , then A is represented
nondegenerately on H (via [7, Lemma 2.1.9] say), and so etζ → ζ for all ζ ∈ H.
It follows that etηet → η WOT, hence weak*. Thus ∆(A∗∗) ⊂ ∆(A)⊥⊥. Since
the converse inclusion is obvious we have ∆(A)∗∗ = ∆(A∗∗). This equals ∆(M (A))
by [3, Proposition 2.11]. For the last part, note that ∆(A) is a HSA in its bidual
by Lemma 5.13, and hence is an annihilator C∗-algebra by Proposition 5.5. Thus
any projection in ∆(A∗∗) = ∆(A)∗∗ is open and closed with respect to ∆(A)∗∗ and
hence also open with respect to A.
(cid:3)
Theorem 5.8. If an operator algebra A is a HSA in A∗∗ then A is nc-discrete.
Proof. We follow some ideas in the proof of [6, Proposition 5.1], which the reader
might follow. By Lemma 5.2, ∆(A) is a HSA in ∆(A)⊥⊥ = ∆(A)∗∗. Hence ∆(A)
is an annihilator C∗-algebra by Proposition 5.5.
Next, let p be an open projection in A∗∗. Suppose that A is a subalgebra of
a C∗-algebra B, generating B as a C∗-algebra. Then ApA ⊂ A, and by Cohen's
factorization BpB = BApAB ⊂ B. There is an increasing net xt ր p, with xt ∈ B
for all t. If b ∈ B+, then bxtb ր bpb ∈ B. Therefore, by Mazur's theorem, replacing
xt by convex combinations of the xt, we may assume that bxtb → bpb in norm, and
0 ≤ xt ≤ p. Then k√p − xtbk2 = kb(p − xt)bk → 0. Hence (p − xt)b → 0, so that
pb ∈ B. Therefore p is a left multiplier of B. However any projection which is a
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
11
left multiplier is a two-sided multiplier. Consequently, pA ∈ B ∩ A⊥⊥ = A, so p is
a left multiplier of A. Similarly, p is a right multiplier of A, hence p ∈ M (A). (cid:3)
In this case there are bijective correspondences between the right
ideals in A with left cai, HSA's in A, and orthogonal projections in the multiplier
algebra M (A). This will follow from the previous theorem and the basic facts about
HSA's (see [6, Section 2]).
Remark.
Lemma 5.9. Suppose that an approximately unital operator algebra A is a HSA in
its bidual, and that π : A → B(H) is a nondegenerate completely isometric repre-
sentation. Then A∗∗ ∼= π(A)
as dual operator algebras. Also, A∗∗ is an essential
extension of A (that is, every completely contractive linear map T : A∗∗ → B(H),
which restricts to a complete isometry on A, is a complete isometry), and A∗∗ em-
beds as a unital subalgebra of a C∗-algebra I(A) which is an injective envelope of
A.
w∗
w∗
Proof. Most of this is essentially in [29], and follows standard ideas (see [7, Section
2.6]), but for completeness we sketch a proof. Define QMπ(A) = {T ∈ B(H) :
π(A)T π(A) ⊂ π(A)}. It is easy to see using [7, Lemma 2.1.6] that π(A)T π(A) = (0)
implies that T = 0. The canonical weak* continuous representation π : A∗∗ →
π(A)
maps into QMπ(A), since π(a)π(η)π(b) = π(aηb) ∈ π(A). Clearly π is
one-to-one, since π(η) = 0 implies aηb = 0 for all a, b ∈ A, so that η = 0. In fact
π(A∗∗) = QMπ(A). To see this suppose that T ∈ QMπ(A), kTk ≤ 1. Suppose
that π(et)T π(es) = π(at,s) for each s, t. For fixed s the net (at,s) has a subnet
converging to ηs ∈ Ball(A∗∗). Suppose that η is a weak* limit point for (ηs) in
Ball(A∗∗). Then
π(a)π(η)π(b) = lim
µ
π(a)π(ηsµ )π(b),
a, b ∈ A,
where this limit and the ones below are weak* limits. However if s = sµ is fixed,
there is a net (tν) such that
π(a)π(ηs)π(b) = lim
ν
π(a)π(atν ,sπ(b) = π(a)π(etν )T π(es)π(b) = π(a)T π(es)π(b).
So
π(a)π(η)π(b) = lim
µ
π(a)π(ηsµ )π(b) = lim
µ
π(a)T π(esµ )π(b) = π(a)T π(b).
w∗
So π(η) = T . Thus π is isometric, hence its range is weak* closed, hence QMπ(A) =
. Once we that know π is isometric, applying this in the setting of Mn(A∗∗) ∼=
π(A)
Mn(A)∗∗ shows that π is completely isometric.
Suppose that z ∈ Ball(QM (A)). By the main theorem in [30] (see also the
quicker proof of [9, Theorem 5.2]), z corresponds to a unique element w ∈ Ball(I(A))
such that awb = azb for all a, b ∈ A. This defines a contractive one-to-one unital
If this w has
map ρ : QM (A) → I(A) which extends the identity map on A.
norm κ then ketzesk ≤ κ for each s, t, so that kzk ≤ κ. Hence ρ : A∗∗ → I(A) is
a unital isometry. By the usual trick (using the isometry applied on Mn(A∗∗) =
Mn(A)∗∗, and the fact that I(Mn(A)) = Mn(I(A)) by 4.2.10 in [7]), ρ is a complete
isometry. Since I(A) is an essential extension of A, we deduce that A∗∗ is an
essential extension of A. Now suppose that I(A∗∗) is an injective envelope of A∗∗
containing A∗∗ as a unital subalgebra. Any complete contraction on I(A∗∗) which
restricts to a complete isometry on A, must be a complete isometry on A∗∗ by the
12
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
last part, hence is a complete isometry on I(A∗∗) by rigidity. So I(A∗∗) is rigid for
A, hence is an injective envelope of A with the desired property.
(cid:3)
Remark. The bulk of the first paragraph of the last proof shows that π is
completely isometric. However this follows immediately from the second paragraph
(that fact there that A∗∗ is an essential extension). Nonetheless we felt it worthwhile
to include a more elementary argument.
Theorem 5.10. Let A be an operator algebra which is a HSA in its bidual.
(1) A is an Asplund space (that is, A∗ has the RNP). Also, A∗ has no proper
subspace that norms A.
(2) A∗∗ is a rigid extension of A in the Banach space category (that is, there
is only one contractive linear map from A∗∗ to itself extending IA).
(3) Any surjective linear complete isometry A∗∗ → A∗∗ is weak* continuous.
(4) There is a unique completely contractive extension π : A∗∗ → B(H) of any
nondegenerate completely contractive representation π : A → B(H), namely
the canonical weak* continuous extension. In particular, every character of
A∗∗ is weak* continuous.
Proof. (1) and (2) follow from [6, Theorem 2.10], which says that such A is 'Hahn-
Banach smooth', and well known properties of 'Hahn-Banach smooth' spaces due
to Godefroy and coauthors, and others (see e.g. [23]).
(3) This follows from the Remark at the end of Section 5 in [3].
(4) In fact this is true even if π is a linear complete contraction with π(et) → IH
weak*. Note that if π is a completely contractive extension of π, then for any unit
vector ζ ∈ H, hπ(·)ζ, ζi is the unique (and hence necessarily weak* continuous)
extension from [6, Theorem 2.10] of the state hπ(·)ζ, ζi. Thus hπ(1)ζ, ζi = 1.
Hence π(1) = I, and we may now appeal to [6, Proposition 2.11].
(cid:3)
Remark. Being an Asplund space is hereditary, so any closed subalgebra C of
an operator algebra A which is an Asplund space has ∆(C) an Asplund space. But
does not imply that ∆(C) is an annihilator C∗-algebra (a C∗-algebra which is an
Asplund space need not be annihilator, certainly C0(K)∗ may be a separable ℓ1
space without K being discrete (consider K the one point compactification of N).
As in [40, Proposition 3.14] we obtain:
Corollary 5.11. If an operator algebra A is a HSA in its bidual, and if A is
not reflexive, then it contains a copy of c0. Similarly, every approximately unital
subalgebra of A, and every quotient algebra of A, which is not reflexive, contains a
copy of c0.
The following is a variant of the 'Wedderburn theorem' for operator algebras
from [3]:
Corollary 5.12. A separable operator algebra A is σ-matricial in the sense of [3]
iff A is semiprime, a HSA in its bidual, and every HSA D in A with dim(D) > 1,
contains a nonzero projection which is not an identity for D.
Proof. Follows from Theorem 4.23 (vii) in [3] together with Theorem 5.8.
(cid:3)
The property of being nc-discrete also passes to subalgebras (and to quotients
by closed ideal having cai):
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
13
Lemma 5.13. Let A be a nc-discrete operator algebra with cai. If B is a closed
subalgebra of A with a cai, then B is nc-discrete. If I is a closed ideal in A, and if
I has a cai, then A/I is nc-discrete.
Proof. If A is nc-discrete, a subalgebra of a C∗-algebra B, and D is a closed ap-
proximately unital subalgebra of A, with p an open projection in B∗∗ which lies
in D⊥⊥, then p ∈ A⊥⊥, so p ∈ M (A). Then pd ∈ D⊥⊥ ∩ A = D for all d ∈ D.
Similarly dp ∈ D, so p ∈ M (D). Thus D is nc-discrete.
Next, suppose that A is nc-discrete, and that I is an approximately unital ideal
in A. If B is a C∗-algebra generated by A, then the C∗-algebra generated in B by
I, is an ideal J in B [12, Lemma 2.4]. Let q⊥ be the (central) support projection
of J, which equals the support projection of I. We make the identifications in the
proof of Lemma 5.2 above. Then
A/I ⊂ A∗∗/I ⊥⊥ ∼= A∗∗q ⊂ B∗∗q.
The map A/I → B/J is a completely isometric embedding, since its composition
with the 'canonical inclusion' B/J ⊂ B∗∗q is the complete isometry in the displayed
equation above. An open projection e in (A/I)∗∗ which is open with respect to
A/I, can thus be identified with a projection p ∈ B∗∗q such that there exists a net
(xt) ⊂ A with xtq → p weak*, and qxt = pxt for all t. By hypothesis, q is open
with respect to A, so that there is a net (ys) ⊂ A with ys → q weak* and qys = ys
for all s. Then xtys = xtqys → pq = p, and pxtys = qxtys = xtysq = xtys. It
follows that p is open in A∗∗. Thus p ∈ M (A). Since paq = ap ∈ A ∩ Aq for any
a ∈ A, it is clear that e ∈ M (A/I).
Proposition 5.14. If A is a nc-discrete approximately unital operator algebra, and
is an integral domain, then A has no nontrivial r-ideals.
(cid:3)
Proof. This is clear: The support projection p of any r-ideal is in M (A), as is p⊥,
and Ap⊥pA = (0).
(cid:3)
Proposition 5.15. An ideal with cai in a uniform algebra, which is nc-discrete, is
isometrically isomorphic to c0(I), for some set I.
Proof. If A is a nc-discrete ideal with cai in a uniform algebra, which we can take to
be A1, then by Lemma 5.13 and Proposition 5.5, ∆(A) is a commutative annihilator
C∗-algebra. Thus ∆(A) is densely spanned by its minimal projections. If f is the
sup of these minimal projections in ∆(A)∗∗, then obviously f is open with respect
to ∆(A), hence open with respect to A, and therefore is also closed and is in M (A)
(since A is nc-discrete). Then J = Af ⊥ is an ideal with cai in a uniform algebra
and J possesses no projections (for these would have to be in ∆(A)). On the
other hand, if f 6= 1 then J has proper closed 1-regular ideals by Theorem 3.3 in
[3] if necessary, and since J is nc-discrete (by Lemma 5.13) it contains nontrivial
projections by Proposition 3.5 in [3]. So f = 1.
If e is a minimal projection in ∆(A), then eA is a uniform algebra containing
no nontrivial projections, hence containing no proper nonzero closed ideals with
cai (or else the support projection f for such an ideal would satisfy f = f e ∈ A,
contradicting minimality of e). However every nontrivial uniform algebra contains
proper closed ideals with cai (for example those associated with Choquet boundary
points). Thus eA = C e. Hence 1A1 is the sum of a family {ei : i ∈ I} of mutually
orthogonal algebraically minimal projections in A, and so A ∼= c0(I).
(cid:3)
14
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
In [3] it is conjectured that C∗-algebras are exactly the operator algebras satis-
fying conditions of the type: every closed left ideal has a right cai. The following
is a complement to Theorem 5.1 of [3]. The hypotheses can be weakened further,
we just state a simple representative form of the result:
Proposition 5.16. A semisimple operator algebra A which is nc-discrete, such
that every right ideal in A has a left cai, is an annihilator C∗-algebra.
Proof. It is obvious that A is a left annihilator algebra. Thus A has dense socle
(see e.g. [35, Chapter 8]). The rest is as in Theorem 5.1 of [3].
(cid:3)
6. Examples of operator algebras that are ideals in their biduals
In this section we list several examples answering natural questions that arise
when investigating some of the topics of this paper.
6.1. A reflexive semisimple operator algebra. A unitization of some operator
algebra structure on ℓ2 with pointwise product, will be a unital reflexive commuta-
tive semisimple non-compact operator algebra. One such operator algebra structure
on ℓ2 may be explicitly represented as follows. Identify N with two disjoint copies
of N, and consider the span of E1k + Ekk, with k in the second copy of N, and 1
here from the first copy. This example is not a modular annihilator algebra (be-
cause the canonical maximal ideal has no nonzero annihilator), but it is a Duncan
modular annihilator algebra in the sense of [35, Chapter 8], which is more gen-
eral than a modular annihilator algebra. This example has no nontrivial r-ideals,
since it is reflexive and has no nontrivial projections. Its spectrum is the one point
compactification of N, which is a scattered topological space.
6.2. A nc-discrete semisimple operator algebra which is not a HSA in its
bidual. It is known that ℓ1 with pointwise product is isomorphic to an operator
algebra A say (actually this may be done in many ways, see e.g. [7, Chapter 5]), and
it is semisimple. Let A1 be the unitization of A, then A1 is commutative, unital,
semisimple, and it is not an ideal in its bidual, since it is unital but not reflexive.
We claim that the total number of orthogonal projections in A1 is finite. To see
this, let ej be the minimal idempotents in A coming from the canonical basis for
ℓ1. If p = λ1 + a is a projection in A1 then λ is either 1 or 0, so either p or 1 − p is
in A. Also, the only projections in A are finite sums of some of the ej. The sup of a
finite family of orthogonal projections is an orthogonal projection, so if there were
infinitely many distinct orthogonal projections in A then there would be arbitrarily
large finite sums of the ej represented in the operator algebra as norm 1 projections.
This is impossible, because in ℓ1 the norm of a sum of n of the ej is n. So we
have just finitely many orthogonal projections in A, and the orthogonal projections
in A1 are these projections and their complements. Depending on the choice of
representation, any finite ring of projections of ℓ1 can be the ring of orthogonal
projections in A (by basic similarity theory such a finite family of idempotents are
simultaneously similar to orthogonal projections); and the r-ideals of A1 include
the unital ideals pA and (1 − p)A1 for each projection p in the chosen ring.
We claim that the just mentioned ideals are the only r-ideals of A1, so that A1
is nc-discrete. To see this, suppose that J is an r-ideal. Case 1: x + 1 /∈ J for all
x ∈ A. In this case, J is an ideal in A. Setting E = {j ∈ N : ej ∈ J}, it is easy to
see that J = JE, where JE consists of the members of A with 'jth coordinate' zero
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
15
for all j /∈ E. This is isomorphic to ℓ1(E). If E is finite then J is finite dimensional,
hence J = eA1 for a projection e ∈ A1 as desired. However if E is infinite then
ℓ1(E) with pointwise product, or equivalently ℓ1 or A, cannot have a bai. Indeed if
A had a bai, then A ⊂ B(A) via the regular representation, and then the argument
at the start of [3, Section 4] gives the contradiction that (Pn
k=1 ej) is uniformly
bounded.
Case 2: x + 1 ∈ J for some x ∈ A. By the argument above, J ∩ A = JE for
some set E ⊂ N. If y + 1 ∈ J for some y ∈ A, then x − y ∈ JE. It follows that
J = C(1 + x) + JE. If j /∈ E then since ej + xej ∈ J we must have xej = −ej. This
can happen for at most a finite number of j; that is N\E is finite. Let q be the sum
of the ej for j ∈ N\E. Then x + q ∈ JE ⊂ J, so that 1 − q = 1 + x − (x + q) ∈ J.
Since q(1 + x) = 0 we see that J has an identity 1 − q, which necessarily has norm
1 since J is approximately unital.
Note that A1 is not a modular annihilator algebra since A has no annihilator,
but it is a Duncan modular annihilator algebra in the sense of [35, Section 8.6].
The remainder of our examples are commutative and radical operator algebras.
In this connection we remark that there are quite a number of papers on comm-
mutative radical operator algebras in the literature, but most of these algebras are
not approximately unital. See for example [18] and [37]. Indeed in [37] and several
related papers by Wogen, Larson, and others, one aim is to study an operator T in
terms of the norm closed algebra oa(T ) generated by T , particularly in cases where
the latter algebra is radical. The following is one of the best studied examples:
6.3. The Volterra operator and a subquotient of the disk algebra. Let V
be the Volterra operator on [0, 1]. Let AV be the norm closed algebra generated
by V . We may write AV = oa(T ) for an operator T with kI − Tk ≤ 1, indeed
let T = I − (I + V )−1 (it is well known that the norm of (I + V )−1 is 1 and its
spectrum is {1}). We have oa(T ) ⊂ AV , and the converse inclusion holds since
V = (I − T )−1T . This algebra has been studied extensively, for example in [37]
and [15, Corollary 5.11]. It is commutative, approximately unital, compact, radical,
and is an ideal in its bidual. Indeed, M (AV ) = V ′ ∼= A∗∗
V [15, Corollary 5.11]. As
we said in [10, Section 5], has no r-ideals; indeed all the closed ideals in this algebra
are known. The algebra AV is nc-discrete but not semiprime, in fact it has a dense
ideal consisting of nilpotent elements.
Jean Esterle suggested to the second author in 2009 to look at the example
D = B/[gB] where B is the ideal of functions in the disk algebra which vanish at
the point 1, and g(z) = exp((z + 1)/(z − 1)). Although g is not in the disk algebra,
it is well known from the theory of inner functions that gB ⊂ B, and that gB is a
closed proper ideal in B (see top of p. 84 in [28]). By [7, Proposition 2.3.4], D is a
commutative operator algebra, and it has a cai since B does. In fact it turns out
that D is completely isometrically isomorphic to the algebra AV above generated
by the Volterra operator. See e.g. [37], where it is pointed out that this leads to
mutual insight into both the operator theory in AV and its commutant, and the
function theory on the disk associated with an interesting class of ideals of the disk
algebra. For example we see from this that D is compact as a Banach algebra, a
fact that seems difficult to see by direct computations in D.
16
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
0
6.4. Weighted convolution algebras which are ideals in their bidual. In
this section we consider operator algebras formed from weighted convolution al-
gebras L1(R+, ω). By a weight we will mean a measurable function ω : R+ =
[0,∞) → (0,∞) with ω(0) = 1, which is submultiplicative in the sense that
ω(s + t) ≤ ω(s)ω(t) for all s, t ≥ 0. Then for 1 ≤ p < ∞, the set Lp(ω) =
Lp(R+, ω) of equivalence classes of measurable functions f : R+ → C such that
kfkp = (R ∞
f (t)p ω(t)p dt)1/p < ∞, is a Banach space with norm kfkp. As in
[14, Section 4.7], L1(ω) with convolution product is a Banach algebra, and it is
t = 0. Otherwise it is semisimple. We write Rxf for the
radical iff limt→∞ ω(t)
right translation of f by x; in other notation this is δx ∗ f . Similarly, we write
Lxf for the left translation of f by x. As in [10, Section 5], convolution induces
a contractive homomorphism f 7→ Mf from L1(ω) into B(L2(ω)), and we define
A = A(ω) to be the norm closure of the set of operators Mf for f ∈ L1(ω). This
is an operator algebra. We write k · kop for the operator norm on A(ω) or more
generally in B(L2(ω)). Whenever we refer below to 'the operator norm' it is this
one.
1
It is known that 1
ω is bounded on compact intervals, so the arguments in [10,
Corollary 5.3] work to show that A(ω) is an integral domain, and in particular is
semiprime, and is not an annihilator algebra. Being an integral domain, it has no
nontrivial idempotents, hence has zero socle. If it is radical then it is a modular
annihilator algebra in the sense of [35], by [35, Theorem 8.7.2]. If ω is right contin-
uous at 0 then there is a nonnegative cai for L1(ω), and hence for A(ω), consisting
of constant multiples of characteristic functions of a sequence of compact intervals
shrinking to 0 (by e.g. [14, 4.7.41]). In this case, since L1(ω) ∩ L2(ω) is dense in
L2(ω), it is clear that A(ω) acts nondegenerately on L2(ω). Hence if also A(ω) is an
ideal in its bidual then A(ω)∗∗ may be identified with the weak* closure of A(ω) in
B(L2(ω)) by Lemma 5.9, and A(ω) possesses no nontrivial r-ideals by Proposition
5.14 (and A(ω)∗∗ contains no nontrivial projections). We remark in passing that
[5, Theorem 2.2] states that if L1(ω) contains a nonzero compact element then ω is
radical.
Lemma 6.1. If a weight ω : [0,∞) → (0,∞) is right continuous at 0, and is not
regulated at some x > 0 (that is if ω(x+t)
ω(t)
9 0), then A(ω) is not compact.
ω(x+t)
ω(t) > 0, there exist an ǫ > 0 and an unbounded in-
Proof. Since lim supt→∞
creasing sequence of numbers (an) such that ω(x+an)
ω(an) > ǫ. Since ω is continuous at
0, it is bounded near 0. By submultiplicativity of ω it is bounded on any compact
interval.
If y ∈ [0, x] then ω(x + an) ≤ ω(y + an) ω(x − y), and it follows that
(changing ǫ if necessary)
(6.1)
1
ǫ ≤
ω(y + an)
ω(an) ≤ ǫ,
y ∈ [0, x].
If f is a nonzero nonnegative C∞ function supported on [0, x
multiplication by f is not compact on A(ω) . To see this define fn = 1
which is supported on [an, an + x
3 ], we claim that
ω(an) Ran f ,
3 ]. We have
kfnkL1(ω) = Z
0
x
3
f (t)
ω(t + an)
ω(an)
x
3
dt ≤ Z
0
f (t) ω(t) dt < ∞,
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
17
and kfnkL1(ω) ∈ [
, ǫkfkL1(R+)] by (6.1). Similarly
f (t)2 ω(t + an)2
dt ∈ [kfk2
L2(R+)
ǫ2
ω(an)2
, ǫ2kfk2
L2(R+)].
kf kL1(R+ )
ǫ
x
3
kfnk2
L2(ω) = Z
0
kf ∗f k2
L2(R+ )
ǫ
Similarly, kf ∗ fnkL2(ω) ≥
. Since f ∗ fn is supported on [an,∞), this
sequence converges weakly to zero in L2(ω). Thus (f ∗ fn) does not have a norm
convergent subsequence in L2(ω). Similarly, (f ∗ fn ∗ f ) does not have a norm
convergent subsequence in L2(ω), and so (f ∗ fn) does not have a norm convergent
subsequence in A(ω). Yet (fn) is a norm bounded sequence in L1(ω) and hence in
A(ω).
Corollary 6.2. If ω is any weight, and if A(ω) is compact then L1(ω) is compact.
Proof. By Lemma 6.1 we have that ω is regulated at all x > 0. Now apply [5,
Theorem 2.7].
(cid:3)
(cid:3)
Remark. The converse of Corollary 6.2 is false: L1(ω) may be compact without
A(ω) being compact. This follows from Theorem 6.4 below and [5, Theorem 2.9].
We say that the radical weight ω satisfies Domar's criterion if the function η(t) =
− log ω(t) is a convex function on (0,∞), and for some ǫ > 0 we have η(t)/t1+ǫ → ∞
as t → ∞. An obvious example of such a weight is ω(t) = e−t2
. In [10, Section 5]
we studied A(ω) if ω satisfies Domar's criterion.
Proposition 6.3. If a radical weight ω : [0,∞) → (0,∞) satisfies Domar's crite-
rion then A(ω) is compact.
Proof. The collection of compact operators on A is closed, hence it suffices to show
that g 7→ h ∗ g is compact on A for h ∈ L1(ω). Since the embedding of L1(R+) in
A is continuous, we may assume further that h is bounded (since simple functions
are dense in L1(ω)), and h has support contained in [ǫ, N ] say, for a fixed ǫ > 0.
Indeed since L1(ω) is an approximately unital Banach algebra, the convolution is
continuous and L1(ω) ∗ L1(ω) = L1(ω); hence we may assume that h = f ∗ g
for bounded f, g ∈ L1(ω) both with support contained in [ǫ, N ]. Clearly such
g ∈ L2(ω) since g and ω are bounded, and so a ∗ g ∈ L2(ω) for a ∈ Ball(A),
2 ∗ a ∗ g is a shift
and ka ∗ gkL2(ω) ≤ kakAkgkL2(ω). By [10, Corollary 5.6], if δ ǫ
of a ∗ g to the right by ǫ
2 , then this is in L1(ω), with norm there dominated by
ka ∗ gkL2(ω) ≤ CkgkL2(ω), for a constant C. On the other hand, if f2 is a shift of
f to the left by ǫ
2 , then on any compact subinterval of (0,∞) the function f2ω is
dominated by a constant times a left shift of fω, since ω is continuous. Since f
has compact support it follows that f2 ∈ L1(ω).
2 ∗(xn∗g)) is a bounded sequence
in L1(ω), by the inequality involving C in the last paragraph. Since the latter
2 ∗ (xn ∗ g)))
algebra is compact by [5], there is a convergent subsequence of (f2 ∗ (δ ǫ
in L1(ω), and hence in A. However f2 ∗ (δ ǫ
2 ∗ (xn ∗ g)) = f ∗ (xn ∗ g) = h∗ xn. Thus
multiplication by h is compact on A as desired.
Let (xn) ⊂ Ball(A). Then xn∗g ∈ L2(ω), and (δ ǫ
(cid:3)
Henceforth we take the weight ω to be a "staircase weight", namely on each inter-
val [n, n + 1) we assume that ω is a constant 1
, where (an) is a strictly and rapidly
an
increasing sequence of positive integers with a0 = 1. So long as an+m ≥ anam for
n, m ∈ N then ω is a weight function, and L1(ω) and A(ω) are commutative Banach
18
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
algebras with cai (see e.g. [5, 14]). It seems to be quite difficult to find examples of
commutative Banach algebras (which are not reflexive in the Banach space sense),
which are weakly compact (hence ideals in their biduals by Lemma 5.1) but not
compact. In fact this is mentioned as an open problem in [42]. The following gives
a commutative approximately unital operator algebra with this property.
Theorem 6.4. Let (ǫn) be a decreasing null sequence, and let (an) be a sequence
chosen with
(6.2)
a0 = 1 ,
an ≥ max{
2k
ǫn−k
ak an−k : k = 1,··· , n − 1}.
Then the operator algebra A(ω) for the associated staircase weight ω is not com-
pact, but is weakly compact, and hence is an ideal in its bidual. Also, A(ω) is a
commutative approximately unital radical operator algebra, which is not reflexive in
the Banach space sense, and which is topologically singly generated.
Clearly such a staircase weight ω is not regulated at numbers in (0, 1), and
hence A(ω) is not compact by Lemma 6.1. To show that it is weakly compact,
by the Eberlein-Smulian theorem it suffices to show that if F ∈ A and (Gn) is a
norm bounded sequence in A, then (F Gn) has a weakly convergent subsequence in
B(L2(ω)). Since the weakly compact operators are closed, it is enough to show this
in the case that F = π(f ) and Gn = π(gn) for f, gn continuous functions of compact
support on R+ (since such functions are dense in L1(ω) by the usual arguments,
and hence in A). Here (Gn) is uniformly bounded in the operator norm, and so has
a weak* convergent subsequence. By passing to this subsequence we may assume
that Gn → G weak*, and F Gn → F G weak*. We will show that a subsequence
F Gkn → F G weakly. Set Hn = Gn − G. We may, and will henceforth, assume that
kHnkop ≤ 1 for all n.
Let Pn denote the orthogonal projection onto the functions in L2(ω) supported
on [0, n], with P0 = 0, and set ∆n = Pn+1 − Pn. We have Pnπ(g) = Pnπ(g)Pn for
each n and g ∈ L1((ω), hence also Pnu = PnuPn for u in A or in ¯Aw∗.
Lemma 6.5. Assume the hypotheses and notation of Theorem 6.4 and the discus-
sion below it, and fix m ∈ N.
(1) Left multiplying by Pmπ(g) = Pmπ(g)Pm on A(ω) is a compact operator
(2) There exists k1 < k2 < ··· with kPm F Hknkop → 0 as n → ∞.
on A(ω) for each g ∈ L1(ω).
Proof. (1) Since the compact operators are closed it suffices to prove (1) for a
dense set of g ∈ L1(ω), as in Proposition 6.3, and as in that proof we may assume
that g = g1 ∗ g2 for bounded g1, g2 with support in [ǫ, N ] where 0 < ǫ < N . We
may also assume that g1, g2 are continuous, by the same idea.
For functions supported on [0, m] the L1(ω) norm is equivalent to the usual L1
norm. We view L1([0, m]) ⊂ L2([0, 1]) as a subspace of L1(ω) ∩ L2(ω) in this way.
Suppose that (an) is a bounded sequence in A. Then (g2∗an) is a bounded sequence
in L2(ω), and so if bn = Pm(g2 ∗ an) then (bn) is a bounded sequence in L1([0, m]).
We then note that left multiplying by Pmπ(g1) may be viewed as the operator
T : L1([0, m]) → C([0, m]) : h 7→ Pm(g1 ∗ h). Indeed T takes Ball(L1([0, m])) into
a uniformly bounded and equicontinuous subset of C([0, m]). By the Arzela-Ascoli
theorem, T is compact. Now
T (bn) = Pm(g1 ∗ Pm(g2 ∗ an)) = Pm(g1 ∗ g2 ∗ an) = Pm(g ∗ an).
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
19
Since T is compact and (bn) bounded, there will be a subsequence (Pm(g∗akn )) that
converges to a function in C([0, m]) in the uniform norm. Since the uniform norm
on C([0, m]) dominates the L1-norm, which is equivalent to the norm on L1(ω),
which dominates the operator norm, we deduce that (Pm(g ∗ akn )) converges in the
operator norm too.
(2) By (1), (PmF Gn) has a norm convergent subsequence, and the limit must
(cid:3)
be PmF G since F Gn → F G weak*. The last assertion is now obvious.
Lemma 6.6. Under the hypotheses of Theorem 6.4, if u ∈ A(ω)
and r ∈ N,
w∗
∞
ku (I − Pr) −
Xm=r
∆m u ∆mk ≤ ǫrkuk.
Further, ∆m u ∆m on Ran(∆m) is unitarily equivalent to P1 u P1 on Ran(P1), and
hence limr→∞ ku (I − Pr)k = kP1 u P1k. All norms here are the operator norm.
Proof. To establish the inequality, we recall that Pnu = PnuPn, and so
(6.3)
u (I − Pr) =
∞
Xm=r
u∆m =
∞
Xm=r
∆m u ∆m +
∞
Xm=r
∞
Xk=1
∆m+k u ∆m,
am
where these sums converge weak*. We next estimate k∆m+k u ∆mk.
Ran(∆m) then η is supported on [m, m + 1], and so kηkL2(ω) = a−1
∆m+k uη is supported on [m+k, m+k+1], and so k∆m+k uηkL2(ω) = a−1
Hence k∆m+k u ∆mk is
times the norm of ∆m+k u ∆m as an operator on
L2(R+). However the last norm equals the norm of ∆k u ∆0 as an operator on
L2(R+), since these operators are unitarily equivalent on their supports (since it
is easy to check that Lm∆m+k u ∆mRm = ∆k u ∆0, and the right shift Rm by m
is an isometry with adjoint Lm). By the norm identity we just established for
k∆m+k u ∆mk in the case m = 0, we deduce that
If η ∈
m kηkL2. Then
m+kk∆m+k uηkL2.
am+k
ak
a0 k∆k u ∆0k ≤
amak
am+k kuk.
k∆m+k u ∆mk =
am
am+k
For varying m, the operators ∆m+k u ∆m have mutually orthogonal left and right
supports, and so
∞
k
Xm=r
∆m+k u ∆mk = sup
m≥r k∆m+k u ∆mk ≤ kuk sup
m≥r
amak
am+k ≤ kuk
ǫr
2k ,
the last inequality following from Equation (6.2). Hence
∞
Xk=1
∞
k
Xm=r
∆m+k u ∆mk ≤ kuk ǫr.
By straightforward operator theory, by the observation above about mutually or-
thogonal left and right supports, we can interchange the double summation in
Equation (6.3) to obtain
∞
∞
∞
ku (I − Pr) −
Xm=r
∆m u ∆mk ≤
Xk=1
k
Xm=r
∆m+k u ∆mk ≤ kuk ǫr,
as desired.
Define Jg(t) = amg(t − m) if m ≤ t ≤ m + 1, and Jg(t) = 0 otherwise. Then
J is an isometry on Ran(P1), with kernel Ran(I − P1) = Ran(P1)⊥. Thus J is a
20
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
partial isometry, and its final space is clearly Ran(∆m). The reader can check that
JP1π(f )P1J ∗ = ∆mπ(f )∆m for f ∈ L1(ω), and the same will be true with π(f )
replaced by u, for u in A or ¯Aw∗. From this it is clear that ∆m u ∆m is unitarily
equivalent to P1 u P1 as stated. The final assertion of the Lemma is obvious since
we are adding elements with mutually orthogonal supports.
(cid:3)
Lemma 6.7. Under the hypotheses of Theorem 6.4, and in the notation above,
there is a sequence of integers 1 = b1 < c1 < b2 < c2 < ··· such that for each n,
k(I − Pbn+1)F Hcnk < 2−n ,
kPbnF Hcnk < 2−n.
The norms here are the operator norm.
Proof. Claim: left multiplication by F Pm is a compact operator on L2(ω). Since
F has compact support there exists an N such that F Pm = PN F Pm. Suppose
that (gn) is a bounded sequence in L2(ω). Then (Pmg) is a bounded sequence
in L1([0, m]), and by the method in the proof of Lemma 6.5 (1) we have that
(PN F Pmg) has a subsequence that converges to a function in C([0, N ]) in the
uniform norm. Since the uniform norm on C([0, N ]) dominates a constant multiple
of the L2(ω) norm, the subsequence converges in the latter norm. This proves the
Claim. It follows that F HnPm = HnF Pm is also compact on L2(ω) for any fixed
n. Thus (I − Ps)F HnPm → 0 in the operator norm as s → ∞ for any fixed n, m.
Assume that 1 = b1 < c1 < ··· < bk have already been chosen. By Lemma 6.5
(2), a subsequence of (Pbk F Hn) converges in norm to 0 by Lemma 6.5 (2). Thus we
may choose ck > bk with kPbk F Hckk < 2−k−1. Hence kP1F Hck P1k < 2−k−1. By
Lemma 6.6 we have limr→∞ kF Hck (I − Pr)k = kP1 F Hck P1k < 2−n−1. Choose a
particular r with kF Hck (I − Pr)k < 2−k−1. If s > r we have
k(I − Ps)F Hckk < 2−k−1 + k(I − Ps)F Hck Prk.
We know from the last line of the first paragraph of the proof that k(I−Ps)F Hck Prk →
0 as s → ∞, so we may choose bk+1 > ck with k(I − Pbk+1 )F Hckk < 2−k. This
completes the inductive step.
(cid:3)
Proof. (Completion of the proof of Theorem 6.4.) If we choose bn, cn as in Lemma
6.7 we have kF Hcn + (Pbn − Pbn+1)F Hcnk < 21−n for all n. Thus to show that
F Hcn → 0 weakly, which concludes our proof that A(ω) is weakly compact, it is
enough that Rn = (Pbn − Pbn+1)F Hcn → 0 weakly. But this is clear since any
bounded family (Rn) of operators on a Hilbert space H with mutually orthogonal
ranges converges weakly to 0. Indeed if P∞
n ≤ 1 then Rn → 0 weakly. To
see this note that for any state ϕ on B(H), we have P∞
n) ≤ 1. Thus by
the Cauchy-Schwarz inequality, ϕ(Rn)2 ≤ ϕ(RnR∗
n) → 0.
The assertion about the second dual follows from Lemma 5.1. To see that A(ω)
is radical, we first note that by mathematical induction on the case k = n − 1
in Equation (6.2), we have an ≥ an
n → ∞ and so
limt→∞ ω(t)
t = 0. Finally, A(ω) is not reflexive since if it were then it would have
an identity of norm 1, which cannot happen for radical algebras. It is topologically
singly generated by the constant function 1, by [14, Theorem 4.7.26].
(cid:3)
21+2+3+···+(n−1). Thus a
n=1 ϕ(RnR∗
n=1 RnR∗
1
ǫn−1
1
1
n
1
7. The diagonal of a quotient algebra
We remark that it is easy to see that if A and B are closed subalgebras of B(H)
then ∆(A ∩ B) = ∆(A) ∩ ∆(B).
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
21
Proposition 7.1. If J is an inner ideal in an operator algebra A (i.e. JAJ ⊂ J),
then J ∩ ∆(A) = ∆(J).
Proof. It is trivial that ∆(J) is a subalgebra of J ∩ ∆(A). Conversely, if JAJ ⊂ J,
then (J ∩∆(A))∆(A)(J ∩∆(A)) ⊂ J ∩∆(A). So J ∩∆(A) is a HSA in a C*-algebra,
hence it is selfadjoint. So x ∈ J ∩ ∆(A) implies that x∗ ∈ J ∩ ∆(A) ⊂ J, and so
x ∈ ∆(J).
(cid:3)
We will use the fact that the diagonal of an ideal J in A is an ideal in ∆(A) if it
is nonzero. Indeed, ∆(J) = ∆(A)∩ J and ∆(J)∆(A) ⊂ ∆(A)∩ (JA) ⊂ ∆(A)∩ J =
∆(J). Similarly, since J is a two-sided ideal, ∆(A)∆(J) ⊂ ∆(J). We sometimes
will silently use this fact. However it is not always true that ∆(A/J) ∼= ∆(A)/∆(J),
if J is an approximately unital ideal in an approximately unital operator algebra A.
A counterexample is given by the ideal of functions in the disk algebra vanishing at
two points on the circle, inside the ideal of functions vanishing at one point. Most
of the rest of this section is an attempt to understand this phenomenon, and to
give some conditions ensuring that ∆(A/J) ∼= ∆(A)/∆(J).
Proposition 7.2. Let A be an approximately unital operator algebra and let J be
an ideal in A. Then ∆(A)/∆(J) ⊂ ∆(A/J).
Proof. Let u : A → A/J be the canonical complete quotient map defined as u(x) =
x + J. The restriction of u to ∆(A) ⊂ A, u′, is a complete contraction. Since
u′ is a contractive homomorphism, it maps into ∆(A/J) by 2.1.2 of [7]. Hence,
we have a completely contractive map u′ = u∆(A) : ∆(A) → ∆(A/J), where
Ker(u′) = ∆(A)∩ J = ∆(J). By the fact mentioned above the proposition, ∆(J) is
an approximately unital ideal in ∆(A), and we deduce that ∆(A)/∆(J) ⊂ ∆(A/J)
completely isometrically.
(cid:3)
Lemma 7.3. Let J be an approximately unital ideal with positive cai in an ap-
proximately unital operator algebra A. Then ∆(A/J) ∼= ∆(A)/∆(J) canonically iff
every positive element of A/J lifts to an element b ∈ A such that b∆(J) ⊂ ∆(J).
Proof. One direction is obvious since as we said earlier, ∆(J) is an ideal in ∆(A).
For the other direction, let p be the support projection of J. The element b in
the statement satisfies b = bp + bp⊥. Now the canonical isometric homomorphism
A/J → A∗∗p⊥ ⊂ A∗∗ takes the diagonal of A/J into ∆(A∗∗). Thus bp⊥ ∈ ∆(A∗∗).
If (et) is a cai for ∆(J) then bet ∈ ∆(J), and in the limit we also have bp ∈
∆(J)⊥⊥ ⊂ ∆(A∗∗) (the latter since ∆(J) ⊂ ∆(A∗∗)). So b ∈ ∆(A∗∗) ∩ A = ∆(A).
From this the result is evident.
(cid:3)
For any operator algebra A, the diagonal ∆(A) acts nondegenerately on A if
and only if A has a positive cai, and if and only if 1∆(A)⊥⊥ = 1A∗∗. The latter
is equivalent to 1A∗∗ ∈ ∆(A)⊥⊥. Hence, we may use the statements '∆(A) acts
nondegenerately on A' and 'A has a positive cai' interchangeably.
Remark. If ∆(A) acts nondegenerately on A, then this does not imply that
∆(J) acts nondegenerately on J if J is an ideal of A. To see this, take any approx-
imately unital operator algebra J such that ∆(J) does not act nondegenerately
on J. Then, J is an ideal in A = M (J), and ∆(A) acts nondegenerately on A.
However, if ∆(A) acts nondegenerately on A, then ∆(A/J) acts nondegenerately
on A/J. Indeed, it is fairly evident by e.g. 2.1.2 in [7] that if J is an ideal in an
operator algebra A, and if A has a positive cai, then A/J has a positive cai.
22
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
Proposition 7.4. Let A be an operator algebra. If J is a left ideal in A with a
selfadjoint right cai, then ∆(J) = ∆(A) ∩ J is a left ideal in ∆(A), ∆(A) + J is
closed, and (J ∩ ∆(A))⊥⊥ = J ⊥⊥ ∩ ∆(A)⊥⊥.
Proof. The first statement is evident from Proposition 7.1. Write (eλ) for the
selfadjoint right cai of J.
If r ∈ ∆(A) then reλ ∈ ∆(A) ∩ J. This is what is
needed to make the idea in the proof of [19, Proposition 2.4] work, as in the proof
of Corollary 4.3, giving that ∆(A) + J is closed, By the proof of [13, Lemma 5.29
and Appendix A.1.5], (∆(A) ∩ J)⊥⊥ = (∆(A)⊥ + J ⊥)⊥ = ∆(A)⊥⊥ ∩ J ⊥⊥.
(cid:3)
Proposition 7.5. Let A be an approximately unital operator algebra such that
∆(A∗∗) = ∆(A)∗∗. If J is an ideal in A that contains a positive cai, then ∆(J ⊥⊥) =
∆(J)⊥⊥ and ∆((A/J)∗∗) = (∆(A)/∆(J))∗∗.
Proof. By Proposition 7.4, ∆(J)⊥⊥ = ∆(A)⊥⊥ ∩ J ⊥⊥. Clearly ∆(A∗∗) ∩ J ⊥⊥ =
∆(J ⊥⊥) by Proposition 7.1, and so ∆(J ⊥⊥) = ∆(J)⊥⊥. Let p ∈ A∗∗ be the support
projection of J. Then,
∆((A/J)∗∗) = ∆(A∗∗(1 − p)) = ∆(A∗∗)(1 − p) = ∆(A)∗∗(1 − p) = (∆(A)/∆(J))∗∗,
where we have used Remark 2.10 (ii) in [3].
(cid:3)
Corollary 7.6. Let A be an operator algebra with a positive cai that is a HSA
in its bidual. If J is an ideal in A that possesses a positive cai, then ∆(A/J) =
∆(A)/∆(J).
Proof. Since A is a HSA in its bidual and it contains a positive cai, we have
∆(A∗∗) = ∆(A)∗∗ by Theorem 5.7. Also, A/J is a HSA in its bidual (A/J)∗∗
by Lemma 5.2, and it can easily be seen by e.g. 2.1.2 in [7] to have a positive
cai. Hence ∆((A/J)∗∗) = ∆(A/J)∗∗ by Theorem 5.7. Moreover, since A/J is a
HSA in its bidual, ∆(A/J) is an annihilator C∗-algebra. We know by Proposition
7.2 that ∆(A)/∆(J) ⊂ ∆(A/J) completely isometrically. Hence ∆(A)/∆(J) is an
annihilator C∗-algebra as well. Using Proposition 7.5, we have (∆(A)/∆(J))∗∗ =
∆((A/J)∗∗) = ∆(A/J)∗∗. Hence, ∆(A)/∆(J) = ∆(A/J).
(cid:3)
We end with another result on the diagonal related to Corollary 4.3:
Proposition 7.7. If A is an approximately unital operator algebra, if D is a HSA
in A and if J is an approximately unital ideal in A, then ∆(D ∩ J)⊥⊥ = ∆(D)⊥⊥ ∩
∆(J)⊥⊥. If D and J have positive cais, then D ∩ J has a positive cai as well.
Proof. A modification of the proof of Proposition 7.4 or Corollary 4.3 shows that
∆(D ∩ J)⊥⊥ = ∆(D)⊥⊥ ∩ ∆(J)⊥⊥.
If D and J have positive cais, then
1(D∩J)⊥⊥ = 1D⊥⊥∩J ⊥⊥ = 1D⊥⊥1J ⊥⊥ ∈ ∆(D)⊥⊥ ∩ ∆(J)⊥⊥ = ∆(D ∩ J)⊥⊥.
That is, 1(D∩J)⊥⊥ ∈ ∆(D ∩ J)⊥⊥. Hence, D ∩ J has a positive cai.
(cid:3)
References
[1] C. A. Akemann, Left ideal structure of C ∗-algebras, J. Funct. Anal. 6 (1970), 305 -- 317.
[2] M. Almus, Structure in operator algebras, Ph.D. thesis, University of Houston, (2011).
[3] M. Almus, D. P. Blecher, and S. Sharma, Ideals and structure of operator algebras, J. Oper-
ator Theory 67 (2012), 397 -- 436.
[4] W. B. Arveson, Subalgebras of C ∗-algebras, Acta Math. 123 (1969), 141 -- 224.
IDEALS AND HSA'S IN OPERATOR ALGEBRAS
23
[5] W. G. Bade and H. G. Dales, Norms and ideals in radical convolution algebras, J. Funct.
Anal. 41 (1981), 77 -- 109.
[6] D. P. Blecher, D. M. Hay, and M. Neal, Hereditary subalgebras of operator algebras, J.
Operator Theory 59 (2008), 333-357.
[7] D. P. Blecher and C. Le Merdy, Operator algebras and their modules -- an operator space
approach, Oxford Univ. Press, Oxford (2004).
[8] D. P. Blecher and M. Neal, Open projections in operator algebras II: Compact projections,
Studia Math. 209 (2012), 203-224.
[9] D. P. Blecher and M. Neal, Metric characterizations II, To appear Illinois J. Math.
[10] D. P. Blecher and C. J. Read, Operator algebras with contractive approximate identities, J.
Functional Analysis 261 (2011), 188-217.
[11] D. P. Blecher and C. J. Read, Operator algebras with contractive approximate identities II,
Preprint, 2012.
[12] D. P. Blecher and M. Royce, Extensions of operator algebras I, J. Math. Anal. Appl. 339
(2008), 1451 -- 1467.
[13] D. P. Blecher and V. Zarikian, The calculus of one-sided M -ideals and multipliers in operator
spaces, Mem. Amer. Math. Soc. 179(2006), no. 842.
[14] H. G. Dales, Banach algebras and automatic continuity, London Mathematical Society Mono-
graphs. New Series, 24, Oxford Science Publications. The Clarendon Press, Oxford University
Press, New York, 2000.
[15] K. R. Davidson, Nest algebras. Triangular forms for operator algebras on Hilbert space,
Pitman Research Notes in Math., Vol. 191, Longman, Harlow, 1988.
[16] M. Daws and C. J. Read, Semisimplicity of B(ℓp)′′, J. Funct. Anal. 219 (2005), 177 -- 204.
[17] J. Dixmier, C ∗-algebras, North-Holland Publ. Co., Amsterdam, 1977.
[18] P. G. Dixon, Radical Q-algebras, Glasgow Math. J. 17 (1976), 119 -- 126.
[19] P. G. Dixon, Non-closed sums of closed ideals in Banach algebras, Proc. Amer. Math. Soc.
128 (2000), 3647 -- 3654.
[20] J. Duncan, B# modular annihilator algebras, Proc. Edinburgh Math. Soc. 15 (1966/1967),
89 -- 102.
[21] N. Dunford and J. T. Schwartz, Linear operators. Part II. Spectral theory. Selfadjoint opera-
tors in Hilbert space, Wiley-Interscience Publication, John Wiley and Sons, Inc., New York,
1988.
[22] F. Ghahramani, Weighted group algebra as an ideal in its second dual space, Proc. Amer.
Math. Soc. 90 (1984), 71 -- 76.
[23] G. Godefroy and P. Saphar, Duality in spaces of operators and smooth norms on Banach
spaces. Illinois J. Math. 32 (1988), 672 -- 695.
[24] P. Harmand, D. Werner, and W. Werner, M -ideals in Banach spaces and Banach algebras,
Lecture Notes in Math., 1547, Springer-Verlag, Berlin -- New York, 1993.
[25] D. Han, D. Larsen, Z. Pan, and W. Wogen, Separating vectors for operators, Proc. Amer.
Math. Soc. 135 (2007), 713 -- 723.
[26] D. M. Hay, Noncommutative topology and peak peak interpolation for operator algebras, Ph.D.
thesis, Houston 2006.
[27] D. M. Hay, Closed projections and peak interpolation for operator algebras, Integral Equations
Operator Theory 57 (2007), 491 -- 512.
[28] K. Hoffman, Banach spaces of analytic functions, Dover, 1988.
[29] M. Kaneda, Quasi-multipliers and algebrizations of an operator space, J. Funct. Anal. 251
(2007), 346 -- 359.
[30] M. Kaneda and V. I. Paulsen, Quasimultipliers of operator spaces, J. Funct. Anal. 217 (2004),
347 -- 365.
[31] K. B. Laursen and M. M. Neumann, An introduction to local spectral theory, London Mathe-
matical Society Monographs, New Series, 20, The Clarendon Press, Oxford University Press,
New York, 2000.
[32] R. J. Loy, C. J. Read, V. Runde, and G. A. Willis, Amenable and weakly amenable Banach
algebras with compact multiplication, J. Funct. Anal. 171 (2000), 78 -- 114.
[33] G. S. Mustafaev, On the discreteness of the structure space of weakly completely continuous
Banach algebras, Mat. Zh. 56 (2004), 411 -- 418; translation in Ukrainian Math. J. 56 (2004),
504 -- 511.
24
MELAHAT ALMUS, DAVID P. BLECHER, AND CHARLES JOHN READ
[34] T. Oikhberg and E. Spinu, Domination of operators in the noncommutative setting, Preprint
2012.
[35] T. W. Palmer, Banach algebras and the general theory of ∗-algebras, Vol. I. Algebras and
Banach algebras, Encyclopedia of Math. and its Appl., 49, Cambridge University Press,
Cambridge, 1994.
[36] G. K. Pedersen, C ∗-algebras and their automorphism groups, Academic Press, London (1979).
[37] J. Peters and W. Wogen, Commutative radical operator algebras, J. Operator Theory 42
(1999), 405 -- 424.
[38] C. J. Read, C ∗-algebras with dense nilpotent subalgebras, Bull. London Math. Soc. 35 (2003),
362-366.
[39] C. J. Read, On the quest for positivity in operator algebras, J. Math. Analysis and Applns.
381 (2011), 202 -- 214.
[40] S. Sharma, Operator spaces which are one-sided M-ideals in their bidual, Studia Math. 196
(2010), 121 -- 141.
[41] E. L. Stout, The theory of uniform algebras, Bogden & Quigley, Inc., Tarrytown-on-Hudson,
N. Y., 1971.
[42] A. Ulger, Some results about the spectrum of commutative Banach algebras under the weak
topology and applications, Monaths. Fur Math. 121 (1996), 353 -- 379.
Department of Mathematics, University of Houston, Houston, TX 77204-3008
E-mail address, Melahat Almus: [email protected]
Department of Mathematics, University of Houston, Houston, TX 77204-3008
E-mail address, David P. Blecher: [email protected]
Department of Pure Mathematics, University of Leeds, Leeds LS2 9JT, England
E-mail address, Charles John Read: [email protected]
|
1304.7453 | 1 | 1304 | 2013-04-28T10:54:52 | Which multiplier algebras are $W^*$-algebras? | [
"math.OA"
] | We consider the question of when the multiplier algebra $M(\mathcal{A})$ of a $C^*$-algebra $\mathcal{A}$ is a $ W^*$-algebra, and show that it holds for a stable $C^*$-algebra exactly when it is a $C^*$-algebra of compact operators. This implies that if for every Hilbert $C^*$-module $E$ over a $C^*$-algebra $\mathcal{A}$, the algebra $B(E)$ of adjointable operators on $E$ is a $ W^*$-algebra, then $\mathcal{A}$ is a $C^*$-algebra of compact operators.
Also we show that a unital $C^*$-algebra $\mathcal{A}$ which is Morita equivalent to a $ W^*$-algebra must be a $ W^*$-algebra. | math.OA | math |
WHICH MULTIPLIER ALGEBRAS ARE
W ∗-ALGEBRAS?
CHARLES A. AKEMANN, MASSOUD AMINI, AND MOHAMMAD B. ASADI
Abstract. We consider the question of when the multiplier al-
gebra M (A) of a C ∗-algebra A is a W ∗-algebra, and show that
it holds for a stable C ∗-algebra exactly when it is a C ∗-algebra
of compact operators. This implies that if for every Hilbert C ∗-
module E over a C ∗-algebra A, the algebra B(E) of adjointable
operators on E is a W ∗-algebra, then A is a C ∗-algebra of compact
operators.
Also we show that a unital C ∗-algebra A which is Morita equiv-
alent to a W ∗-algebra must be a W ∗-algebra.
1. Introduction
The main theme of this paper is around the question of when the
multiplier algebra M(A) of a C ∗-algebra A is a W ∗-algebra? For sepa-
rable C ∗-algebras, it holds exactly when A is a C ∗-algebra of compact
operators [2, Theorem 2.8]. For general C ∗-algebras, we get two partial
results in this direction. First we give an affirmative answer for stable
C ∗-algebras and deduce that if for every Hilbert C ∗-module E over A,
the algebra B(E) of adjointable operators on E is a W ∗-algebra, then
A is a C ∗-algebra of compact operators. This is related to our question
(with a much stronger assumption) as for E = A with its canonical
Hilbert A-module structure, B(E) = M(A). Second we show that if
M(A) is Morita equivalent to a W ∗-algebra, then it is a W ∗-algebra.
This is also related to our question, as if A is a C ∗-algebra of compact
operators, then M(A) is a W ∗-algebra.
The two partial answers take into account the notions of Hilbert
C ∗-algebras and Morita equivalence which are somewhat historically
related. In 1953, Kaplansky introduces Hilbert C ∗-modules to prove
that derivations of type I AW ∗-algebras are inner. Twenty years later,
Hilbert C ∗-modules appeared in the pioneering work of Rieffel [19],
2010 Mathematics Subject Classification. Primary 46L10, 47L30; Secondary
46L08.
Key words and phrases. Hilbert C ∗-modules, Morita equivalence, multiplier al-
gebras, operator algebras, W ∗-algebras.
The second and third authors were in part supported by a grant from IPM
(91430215 & 91470123).
1
2
C. AKEMANN, M. AMINI, AND M.B. ASADI
where he employed them to study (strong) Morita equivalence of C ∗-
algebras. Paschke studied Hilbert C ∗-modules as a generalization of
Hilbert spaces [16].
Hilbert C ∗-modules and Hilbert spaces differ in many aspects, such
as existence of orthogonal complements for submodules (subspaces),
self duality, existence of orthogonal basis, adjointability of bounded
operators, etc. However, when A is a C ∗-algebra of compact opera-
tors, then Hilbert A-modules behave like Hilbert spaces in having the
above properties. Indeed these properties characterize C ∗-algebras of
compact operators [5, 10, 14, 20].
2. C ∗-algebras of compact operators
In this section we give some characterizations of C ∗-algebras of com-
pact operators using properties of multiplier algebras. We also show
that these are characterized as C ∗-algebras A for which the algebra
B(E) of all adjointable operators is a W ∗-algebra, for any Hilbert A-
module E.
Definition 2.1. A C ∗-algebra A is called a C ∗-algebra of compact op-
erators if there exists a Hilbert space H and a (not necessarily surjec-
tive) ∗-isomorphism from A to K(H), where K(H) denotes the space
of compact operators on H.
This is exactly how Kaplansky characterized C*-algebras that were
dual rings [11, Theorem 2.1, p. 222] (see also [1]).
Theorem 2.2. For a C ∗-algebra A, the following are equivalent:
(i) A is a C ∗-algebra of compact operators.
(ii) The strict topology on the unit ball of M(A) is the same as the
strong∗-topology (viewing M(A) ⊆ A∗∗, the second dual of A).
Proof. Assume that (i) holds. Then A ∼= c0-P Lα K(Hα). Let aβ → 0
in the strict topology of the unit ball of M(A) ∼= ℓ∞-P Lα B(Hα).
Without loss of generality, we may assume that aβ ≥ 0, for all β. Let
η ∈ Lα Hα be a unit vector with ηα = 0 except for finitely many α.
Let pα be the rank one projection onto the non-zero ηα and pα = 0,
otherwise. Then p = P pα ∈ A, thus kaβpk → 0. Therefore kaβηk →
0, and the same holds for any η in the unit ball of Lα Hα, as {aβ} is
norm bounded. Hence aβ → 0 in the strong∗ topology.
Conversely if aβ ≥ 0 and aβ → 0 in the strong∗ topology. As above,
for any rank one projection p ∈ A, kaβpk = kpaβk → 0. Thus p can be
replaced by any finite linear combination of such minimal projections,
and this set is dense in A. Since {aβ} is norm bounded, aβ → 0 in the
strict topology. This shows that (i) implies (ii).
MULTIPLIER ALGEBRAS
3
Now assume that (ii) holds. By [2, Theorem 2.8], we need only
to prove that M(A) = A∗∗. For any positive element b in the unit
ball of A∗∗, there is a net {aβ} in the unit ball of A that converges
to b in strong∗ topology. Thus the net is strong∗ Cauchy, and hence
convergent in the strict topology to an element of M(A), as M(A) is
the completion of A in the strict topology [9, Theorem 3.6]. Therefore
b ∈ M(A), and we are done.
(cid:3)
Another characterization of C ∗-algebras of compact operators could
be obtained as a non unital version of the following result of J.A. Mingo
in [15], where he investigates the multipliers of stable C ∗-algebras.
Lemma 2.3. Suppose that H is a separable infinite dimensional Hilbert
space and A is a unital C ∗-algebra such that the multiplier algebra
M(A ⊗ K(H)) is a W ∗-algebra. Then A is a finite dimensional C ∗-
algebra.
We recall that a projection p in a C ∗-algebra A is called finite dimen-
sional if pAp is a finite dimensional C ∗-algebra. To prove a non unital
version of Mingo's result, we need some lemmas. The first lemma is
well-known, see for instance [4, Corollary 1.2.37].
Lemma 2.4. If A is a C ∗-algebra and p is a projection in the multiplier
algebra M(A), then M(pAp) ∼= pM(A)p, as C ∗-algebras.
Lemma 2.5. Let H be a Hilbert space and A be a C ∗-algebra.
A ⊗ K(H) is C ∗-algebra of compact operators, then so is A.
If
Proof. Suppose not. Then there is an element b ∈ A+ such that the
spectral projection ξ1(b) of b corresponding to {1} is not finite dimen-
sional in A. Let q be a one-dimensional projection in K(H). Then
(b ⊗ q)n is a decreasing sequence in the unit ball of the C*-algebra
A ⊗ K(H) of compact operators. By Theorem 2.2 it converges strictly,
hence (because it is decreasing) in norm to ξ1(b) ⊗ q ∈ A. Because
A⊗K(H) is a C ∗-algebra of compact operators, the projection ξ1(b)⊗q
must be finite rank, but
(ξ1(b) ⊗ q)(A ⊗ K(H))(ξ1(b) ⊗ q) = ξ1(b)Aξ1(b) ⊗ qK(H))q,
and the dimension of ξ1(b)Aξ1(b) is not finite by our assumption about
b.
(cid:3)
The next theorem is known for separable C ∗-algebras [2], here we
prove it with separability replaced by stability.
Theorem 2.6. If A is a stable C ∗-algebra such that the multiplier
algebra M(A) is a W ∗-algebra, then A is a C ∗-algebra of compact op-
erators.
4
C. AKEMANN, M. AMINI, AND M.B. ASADI
Proof. In order for the C ∗-algebra A to be a C ∗-algebra of compact
operators, it is necessary and sufficient that every positive element in A
can be approximated by a finite linear combination of finite dimensional
projections. Let a be a positive element in A and 0 ≤ a ≤ 1. Since the
multiplier algebra M(A) is a W ∗-algebra, we can define p ∈ M(A) as
the spectral projection of a, corresponding to an interval of the form
[s, t] where 0 < s < t. It suffices to show that pAp is finite dimensional.
Let g : [0, 1] → [0, 1] be a continuous function vanishing at 0, such that
g(r) = 1 for all r ∈ [s, t]. Then g(a) ∈ A and g(a)p = p. Hence p ∈ A.
Now let H be a separable infinite dimensional Hilbert space. Since
A is a stable C ∗-algebra, M(A) = M(A ⊗ K(H)) is a W ∗-algebra and
by Lemma 2.4,
M(pAp ⊗ K(H)) = M((p ⊗ 1)(A ⊗ K(H))(p ⊗ 1))
= (p ⊗ 1)M(A ⊗ K(H))(p ⊗ 1)
is a W ∗-algebra. Therefore by Lemma 2.3, p is finite rank.
(cid:3)
The non unital version of the Mingo's lemma follows.
Corollary 2.7. Suppose that H is a separable infinite dimensional
Hilbert space and A is a C ∗-algebra such that the multiplier algebra
M(A ⊗ K(H)) is a W ∗-algebra, then A is a C ∗-algebra of compact
operators.
Proof. Since A ⊗ K(H) is stable, it is a C ∗-algebra of compact opera-
tors, and so is A by Lemma 2.5.
(cid:3)
It is well known that if A is a W ∗-algebra and E is a selfdual Hilbert
A-module, then B(E) is a W ∗-algebra. The converse is not true, as
for E = A = c0, B(E) = ℓ∞ is a W ∗-algebra [19]. However, if A is
a C ∗-algebra of compact operators on some Hilbert space, then B(E)
is a W ∗-algebra, for every Hilbert A-module E [6]. Here we show the
converse.
Recall that the C ∗-algebra K(E) of compact operators on E is gen-
erated by rank one operators θξ,η(ζ) = ξhη, ζi, for ξ, η ∈ E, and the
multiplier algebra M(K(E)) is isomorphic to B(E). Also, if H is a sep-
arable infinite dimensional Hilbert space, then E = H ⊗ A is a Hilbert
C ∗-module over C ⊗ A = A, denoted by HA. It plays an important
role in the theory of Hilbert C ∗-modules.
Theorem 2.8. For any C ∗-algebra A, the following are equivalent:
(i) A is a C ∗-algebra of compact operators,
(ii) B(E) is a W ∗-algebra, for each Hilbert A-module E,
(iii) B(HA) is a W ∗-algebra.
MULTIPLIER ALGEBRAS
5
Proof. It is enough to show that (iii) implies (i). Since
K(HA) = K(H ⊗ A) ∼= K(H) ⊗ K(A) = K(H) ⊗ A
we have B(HA) ∼= M(K(H) ⊗ A). By assumption, B(HA) is a W ∗-
algebra and so A is a C ∗-algebra of compact operators by Corollary
2.7.
(cid:3)
J. Schweizer in [20] remarked that for a C ∗-algebra A, some problems
on Hilbert A-modules can be reformulated as problems on right ideals
of A, since submodules of a full Hilbert A-module are in a bijective
correspondence with the closed right ideals of A. Therefore, one may
wonder if the previous result could be reformulated in the language of
right ideals. Actually, if I is a (closed) right ideal of A, then I is a
right Hilbert A-module with inner product ha, bi = a∗b, for a, b ∈ I,
and in this case, K(E) equals to the hereditary C ∗-algebra I ∩ I ∗ and
so B(E) = M(I ∩ I ∗). Therefore, one may expect that C ∗-algebras A
of compact operators may be characterized by the property that for
every hereditary C ∗-subalgebra B of A, M(B) is a W ∗-algebra.
Unfortunately, this is not the case for non separable C ∗-algebras, as
the following counterexample shows. However, if A is separable and p
is a projection as in the proof of Theorem 2.6, then pAp is a separable
W ∗-algebra, hence finite dimensional (also see Theorem 2.8 in [2]).
Example 2.9. For the Stone-Cech compactification βN of the natural
numbers, the algebra of continuous functions C(βN) is a W ∗-algebra.
Let x be any point of βN that is not a natural number and let A be the
C ∗-subalgebra of C(βN) consisting of those functions vanishing at x.
Let B be a hereditary C*-subalgebra of A (which is an ideal, since A is
abelian). Then there is an open subset U of βN such that B consists of
functions in A that vanish outside U. Let V be the closure of U. Then
V is also open. For every c ∈ C(V ) we may extend c by zero outside V ,
and thereby view C(V ) as a W ∗-subalgebra of C(βN). Observe that
M(B) = C(V ): clearly B is an ideal in C(V ), so it suffices to note that
for any 0 6= c ∈ C(V ), cB 6= 0. To see this, we note that c is non-zero
on a nonvoid open subset W of V , hence W ∩ U \ x is a nonvoid open
set. Hence there exists a non-zero continuous function b with support
in W ∩ U \ x. Thus b ∈ B and cb 6= 0. Therefore M(B) = C(V ) is a
W ∗-algebra, but A cannot be a C ∗-algebra of compact operators.
3. Morita equivalence
The notion of (strong) Morita equivalence of C ∗-algebras was intro-
duced by M. Rieffel in [19]. Two C ∗-algebras A and B are (strongly)
Morita equivalent if there is an A-B-bimodule M, which is a left full
6
C. AKEMANN, M. AMINI, AND M.B. ASADI
Hilbert C ∗-module over A, and a right full Hilbert C ∗-module over B,
such that the inner products Ah·, ·i and h·, ·iB satisfy Ahx, yiz = xhy, ziB
for all x, y, z ∈ M. Such a module M is called an A-B-imprimitivity
bimodule.
It would be interesting to investigate those properties of C ∗-algebras
which are preserved under Morita equivalence. These include, among
other things nuclearity, being type I, and simplicity [3, 7, 12, 17, 18,
21, 22]. Now if one of the two Morita equivalent C ∗-algebras is a
W ∗-algebra, it is natural to ask if so is the other. The answer to
this question, as it posed is obviously negative, as Hilbert space H is
a K(H)-C-imprimitivity bimodule, and so C ∗-algebras K(H) and C
are Morita equivalent. However we may rephrase that question in the
following less trivial form.
Question 3.1. Suppose that C ∗-algebras A and B are Morita equiv-
alent and the C ∗-algebra M(A) is a W ∗-algebra, is it then true that
M(B) is a W ∗-algebra?
By Theorem 2.8, we can show that the above property holds for C ∗-
algebra A exactly when A is a C ∗-algebra of compact operators. In
fact, we have the following result.
Theorem 3.2. Let A be a C ∗-algebra such that M(B) is a W ∗-algebra,
for any C ∗-algebra B which is Morita equivalent to A. Then A is a
C ∗-algebra of compact operators.
Proof. Let B = K(HA). Since HA is a full Hilbert A-module, then B
is Morita equivalent to A. By assumption, B(HA) ∼= M(B) is a W ∗-
algebra, hence A is a C ∗-algebra of compact operators, by Theorem
2.8.
(cid:3)
However, we give an affirmative answer to the above question, when
both C ∗-algebras are unital.
Recall that a Hilbert C ∗-module E on a C ∗-algebra A is called self
dual if for every bounded linear A-module map ϕ : E → A there is an
element y ∈ E such that ϕ(·) = hy, ·i.
Lemma 3.3. Let E be a right Hilbert C ∗-module over a C ∗-algebra A
such that K(E) is unital. then
(i) E is self dual.
(ii) B(E) is a W ∗-algebra, whenever A is a W ∗-algebra.
Proof. By hypothesis there are elements x1, · · ·, xn and y1, · · ·, yn in
E such that Pn
i=1 θxi,yi = 1 ∈ K(E). Thus, for every bounded linear
MULTIPLIER ALGEBRAS
7
A-module map ϕ : E → A and x ∈ E we have
ϕ(x) = ϕ(
n
X
i=1
θxi,yix) = ϕ(
xihyi, xi) =
n
X
i=1
ϕ(xi)hyi, xi
n
X
i=1
n
=
n
X
i=1
hyiϕ(xi)∗, xi = h
yiϕ(xi)∗, xi.
X
i=1
Therefore ϕ(x) = hy, xi, where y = Pn
Now (ii) follows from (i) and [16, Proposition 3.10].
i=1 yiϕ(xi)∗. Hence E is selfdual.
(cid:3)
Now if E is an A-B-imprimitivity bimodule, then A ∼= KB(E) and
B ∼= KA(E). Therefore, the following partial answer to the above
question follows from the above lemma.
Theorem 3.4. Suppose that unital C ∗-algebras A and B are Morita
equivalent. Then A is a W ∗-algebra if and only if B is a W ∗-algebra.
A similar result can be proved for operator algebras. Let A and
B be operator algebras. We say that A and B are (strongly) Morita
equivalent if they are Morita equivalent in the sense of Blecher, Muhly,
In [8], it is proved that two C ∗-algebras are (strongly)
Paulsen [8].
Morita equivalent (as operator algebras) if and only if they are Morita
equivalent in the sense of Rieffel.
Theorem 3.5. Suppose that unital operator algebras A and B are
Morita equivalent. Then A is a dual operator algebra if and only if
B is a dual operator algebra.
Proof. Let π : A → B(H) be a completely isometric normal represen-
tation of A on some Hilbert space H. Then there exist a completely
isometric representation ρ : B → B(K) of B on a Hilbert spaces K and
subspaces X ⊆ B(K, H), Y ⊆ B(H, K) such that
π(A)Xρ(B) ⊆ X, ρ(B)Y π(A) ⊆ Y, π(A) = XY
k·k
, ρ(B) = Y X
k·k
Since π is normal, we have π(A) = π(A)
implies that Xρ(B)
Y ⊆ π(A). Therefore
w∗
w∗
. Now Xρ(B)Y ⊆ π(A)
w∗
Y Xρ(B)
Y X ⊆ Y π(A)X ⊆ ρ(B),
w∗
w∗
and so ρ(B)ρ(B)
ρ(B)
algebra.
⊆ ρ(B), hence ρ(B)
ρ(B) ⊆ ρ(B). Since ρ(B) is a unital algebra we have
= ρ(B). Therefore B is a dual operator
w∗
(cid:3)
8
C. AKEMANN, M. AMINI, AND M.B. ASADI
Acknowledgment. The authors would like to thank Professor G.
Eleftherakis, who suggested the proof of Theorem 3.5, and Professor J.
Tomiyama for helpful comments.
References
[1] C. A. Akemann, Left ideal structure of C*-algebras, Journal of Functional
Analysis 6 (1970), 305-317.
[2] C. A. Akemann, G.K. Pedersen, J. Tomiyama, Multipliers of C ∗-algebras, J.
Func. Anal. 13 (1973), 277-301.
[3] P. Ara, Morita equivalence and Pedersen ideals, Proc. Amer. Math. Soc. 219
(2001), 1041-1049.
[4] P. Ara, M. Mathieu, Local multipliers of C ∗-algebras, Springer, Berlin, 2003.
[5] L. Arambasic, Another characterization of Hilbert C ∗-modules over compact
operators, J. Math. Anal. Appl. 344 (2008), 735-740.
[6] D. Baki´c, B. Guljas, Hilbert C ∗-modules over C ∗-algebras of compact operators,
Acta Sci. Math. 68 (2002), 249-269.
[7] W. Beer, On Morita equivalence of nuclear C ∗-algebras, J. Pure Appl. Algebra
26 (1982), 49-267.
[8] D.P. Blecher, P. S. Muhly and V. I. Paulsen, Categories of operator modules;
Morita equivalence and projective modules, Memoirs Amer. Math. Soc., 681,
2000.
[9] R.C. Busby, Double centralizers and extensions of C ∗-algebras, Trans. Amer.
Math. Soc. 132 (1968), 79-99.
[10] M. Frank, Characterizing C ∗-algebras of compact operators by generic categor-
ical properties of Hilbert C ∗-modules, J. K-Theory 2 (2008), 453-462.
[11] I. Kaplansky, The structure of certain operator algebras, Trans. Amer. Math.
Soc. 70 (1951), 215-255.
[12] M. Kusuda, Morita equivalence for C ∗-algebras with the weak Banach-Saks
property, Quart. J. Math. 52 (2001), 455-461.
[13] E.C. Lance, Hilbert C ∗-modules, Camb. Univ. Press, Cambridge, 1995.
[14] B. Magajna, Hilbert C ∗-modules in which all closed submodules are comple-
mented, Proc. Amer. Math. Soc. 125 (1997), 849-852.
[15] J.A. Mingo, K-Theory and multipliers of stable C ∗-algebras, Tran. Amer. Math.
Soc. 299 (1987), 397-411.
[16] W. L. Paschke, Inner product modules over B*-algebras, Trans. Amer. Math.
Soc. 182 (1973), 443-468.
[17] I. Raeburn, D. P. Williams, Morita equivalence and continuous-trace C ∗-
algebras, Math. Surveys 60, Amer. Math. Soc., Providence, 1998.
[18] A. an Huef, I. Raeburn, and D. P. Williams, Properties preserved under Morita
equivalence of C ∗-algebras, Proc. Amer. Math. Soc. 135 (2007), 1495-1503.
[19] M. A. Rieffel, Morita equivalence for C ∗-algebras and W ∗-algebras, J. Pure
Appl. Alg. 5 (1974), 5196.
[20] J. Schweizer, A description of Hilbert C ∗-modules in which all closed submod-
ules are orthogonally closed, Proc. Amer. Math. Soc. 127 (1999), 2123-2125.
[21] H.H. Zettl, Strong Morita equivalence of C ∗-algebras preserves nuclearity,
Arch. Math. 38 (1982), 448452.
MULTIPLIER ALGEBRAS
9
[22] H.H. Zettl, Ideals in Hilbert modules and invariants under strong Morita equiv-
alence of C ∗- algebras, Arch. Math. 39 (1982), 6977.
Department of Mathematics, University of California, Santa Bar-
bara, CA 93106, USA
E-mail address: [email protected]
School of Mathematics, Tarbiat Modares University, Tehran 14115
134, Iran, and School of Mathematics, Institute for Research in Fun-
damental Sciences (IPM), Tehran 19395-5746, Iran
E-mail address: [email protected]
School of Mathematics, Statistics and Computer Science, College
of Science, University of Tehran, Enghelab Avenue, Tehran, Iran,
and School of Mathematics, Institute for Research in Fundamental
Sciences (IPM), Tehran 19395-5746, Iran
E-mail address: [email protected]
|
1504.08338 | 5 | 1504 | 2015-11-04T04:08:49 | Quantum $G_{2}$ categories have property (T) | [
"math.OA",
"math.QA"
] | We show that the rigid C*-tensor categories of finite dimensional type 1 unitary representations of the quantum groups $U_{q}(\mathfrak{g}_{2})$ corresponding to the exceptional Lie group $G_2$ for positive $q\ne 1$ have property (T). | math.OA | math |
QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
COREY JONES
Abstract. We show that the rigid C*-tensor categories of finite dimensional type 1 unitary representations
of the quantum groups Uq(g2) corresponding to the exceptional Lie group G2 for positive q (cid:54)= 1 have property
(T).
1. Introduction
Recently there has been a great deal of interest in approximation and rigidity properties for rigid C∗-
tensor categories and subfactors. Rigid C∗-tensor categories, introduced in their modern form by Longo and
Roberts [19], are structures which provide a unifying framework for symmetries appearing in a variety of
contexts. They make a prominent appearance in the theory of compact quantum groups as representation
categories [23], and arise as DHR super-selection sectors in algebraic quantum field theory [11], hence are
often described as encoding "quantum symmetries", generalizing the role of groups and their representations.
Rigid C∗-tensor categories are also realized as categories of bimodules appearing in the standard invariants
of finite index subfactors [13]. The standard invariant of the subfactor N ⊆ M is the 2-category of bimodules
that appear as tensor powers of N L2(M )M and its dual with respect to the relative tensor product. This
yields two tensor categories, the M -M bimodules and the N -N bimodules, related by N L2(M )M and its dual.
The standard invariant is a powerful invariant of a subfactor, and was first axiomatized in full generality
by Popa as standard λ-lattices [29]. Another useful realization is Jones' subfactors planar algebras [14].
Alternatively, standard invariants can be axiomatized as rigid C∗-tensor categories along with a tensor
generating Q-systems (see [21] for details), or via Ocneanu's paragroups in the finite depth case [25].
Popa introduced the concepts of approximation and rigidity properties for subfactors (see [27], [28], [30],
[31]). Popa's definitions can be formulated in terms of the symmetric enveloping inclusion T ⊆ S associated
to N ⊆ M (see [27], [31]), and he showed that the definitions only depend on the standard invariant of the
subfactor. Recently, approximation and rigidity properties were translated from Popa's original definitions
into the categorical setting by Popa and Vaes in [32]. This allows for the definitions of property (T), the
Haagerup property, and amenability to be defined in a conceptually uniform way for standard invariants
and rigid C∗-tensor categories without reference to an ambient subfactor. They introduce a representation
theory for standard invariants and rigid C∗-tensor categories generalizing the unitary representation theory
of groups, encoding in a natural way analytical properties. To this end, Popa and Vaes identify a class of
admissible representations of the fusion algebra of a category. This class admits a universal representation,
allowing for the construction of a univsersal algebra C∗(C), generalizing the universal C∗-algebra for groups.
Soon after the initial work of Popa and Vaes, admissible representations of the fusion algebra were in-
terpreted from different points of view. Neshveyev and Yamashita showed that admissible representations
can be understood as objects in the Drinfeld center of the ind-category [23].
In [10], the authors show
Key words and phrases. C∗ tensor category, Approximation/Rigidity Properties, Planar Algebras.
The author was partially supported by NSF grant DMS-1362138.
1
2
COREY JONES
that admissible representations have a natural interpretation in the annular representation theory for planar
algebras of Jones [16], [15] and the representation theory of the tube algebra of a category, introduced by
Ocneanu in [26]. The fusion algebra is a corner of the tube algebra, and in [10] it is shown that admissible
representations of the fusion algebra are precisely representations which are restrictions of representations
of the whole tube algebra. Since the tube algebra is computable in principle, this provides a method for
determining admissible representations.
While the notion of property (T) for subfactor standard invariants (due to Popa) has been in existence
for many years, examples have been somewhat elusive. Until recently, the only known examples of subfactor
standard invariants with property (T) came in some way from discrete (T) groups. In particular, the diagonal
subfactors and the Bisch-Haagerup subfactors, when constructed with property (T) groups, produce property
(T) standard invariants (see Popa [30], [31] and Bisch-Popa, [3] respectively). Arano showed in [2] that the
discrete dual of the compact quantum groups SUq(N ) have central property (T) for N ≥ 3 odd and positive
q (cid:54)= 1, which Popa and Vaes showed is equivalent to the corresponding representation category of SUq(N )
having property (T) (see [32]). From these categories, one can construct subfactors whose even bimodule
categories are equivalent to Rep(SUq(N )), providing the first examples of subfactors not coming in some
way from discrete groups whose standard invariants have (T). In light of this breakthrough, it is natural to
wonder if other quantum group categories may have property (T).
The categories Cq(g2) for positive q are the rigid C∗-tensor categories of finite dimensional type 1 unitary
representations of the Drinfeld-Jimbo quantum groups, corresponding to the exceptional Lie group G2 (see
[23]). They have been described diagrammatically by Kuperberg [17], [18]. These diagrammatic categories,
denoted (G2)q, have been further studied by Morrsion, Peters, and Snyder in [20], where they appear as
"small" examples of trivalent categories in their classification program, and a description of small idempotents
is obtained.
In this note, we show that the categories Cq(g2) have property (T) for positive q (cid:54)= 1 using their di-
agrammtic descriptions. This provides a new class of examples of subfactors with property (T) standard
invariant. We give a brief outline of the argument:
Since the fusion algebra of (G2)q is abelian (the category is braided), the universal C∗-algebra is isomorphic
to the algebra continuous functions on its spectrum, which corresponds to the space of irreducible admissible
representations. Having property (T ) in this setting simply translates to the trivial representation being
isolated in the spectrum. Identifying the fusion algebra (G2)q as polynomials in two self adjoint variables, the
possible irreducible representations correspond to points in the plane, defined by evaluation of polynomials.
Using some general restrictions, we reduce the possibilities for admissible representations to a rectangle,
with the trivial representation at a corner. Applying the description of small minimal idempotents provided
by Morrison, Peters, and Snyder [20], we define a function of the plane f (α, t) which is 0 at the trivial
representation, and has the property that this function must be non-negative at (α, t) for the representation
corresponding to that point to be admissible. Then using elementary calculus, we show in a neighborhood
of the trivial representation this function is strictly negative, implying that (G2)q has property (T). We note
our calculus arguments break down precisely when q = 1, which is to be expected since the classical G2
representation category is amenable.
Although the categories we study are quantum group category, we emphasis our proof mostly uses only
the basic skein theoretic description of the category. We also remark that a surprisingly small amount of
data from the actual category is used. We hope this note demonstrates the computational usefulness of
QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
3
diagrammatics for studying analytical properties of categories, which was also demonstrated in [4], where
the authors give a direct planar algebraic proof that the categories T LJ(δ) have the Haagerup property for
all δ ≥ 2 using Popa's original definitions.
The outline of the paper is as follows: We begin by describing rigid C∗-tensor categories and the tube
algebra. We then discuss admissible representations and property (T). Finally we give a short proof that
(G2)q has property (T). In the appendix, we discuss in greater detail the quantum group Uq(g2) and describe
how the ∗-structure from Cq(g2) transports to the planar algebra (G2)q.
Acknowledgements: The author would like to thank Noah Snyder and Yunxiang Ren for useful discus-
sions on (G2)q, and Vaughan Jones for his encouragement. He also thanks Marcel Bischoff, Arnaud Brothier,
and Michael Northington for helpful comments, and Shamindra Ghosh for many interesting discussions. The
author was supported by NSF grant DMS-1362138.
2. Preliminaries
2.1. Rigid C*-Tensor Categories. In this paper we will be concerned with semi-simple, C∗-categories
with strict tensor functor, simple unit and duals. We also assume that C has countably many isomorphism
classes of simple objects. Often in the literature, this is the definition of a rigid C∗-tensor category. We
briefly elaborate on the meaning of each of these words.
A C∗-category is a C-linear category C, with each morphism space M or(X, Y ) a Banach space, and
a conjugate linear, involutive, contravariant functor ∗ : C → C which fixes objects and satisfies for every
morphism f , f∗f = f f∗ = f2. We say the category is semi-simple if the category has direct sums,
sub-objects, and each M or(X, Y ) is finite dimensional.
A strict tensor functor is a bi-linear functor ⊗ : C × C → C, which is associative and has a distinguished
unit id ∈ Obj(C) such that X ⊗ id = X = id ⊗ X. The category is rigid if for each X ∈ Obj(C), there exists
X ∈ Obj(C) and morphism R ∈ M or(id, X ⊗ X) and R ∈ M or(id, X ⊗ X) satisfying the so-called conjugate
equations:
(1X ⊗ R
)(R ⊗ 1X ) = 1X and (1X ⊗ R∗)(R ⊗ 1X ) = 1X
∗
We say two objects are X, Y are isomorphic if there exists f ∈ M or(X, Y ) such that f∗f = 1X and
f f∗ = 1Y . We call an object X simple if M or(X, X) ∼= C. We note that for any simple objects X and Y ,
M or(X, Y ) is either isomorphic to C or 0. Two simple objects are isomorphic if and only if M or(X, Y ) ∼= C.
Isomorphism defines an equivalence relation on the collection of all objects and we denote the equivalence
class of an object by [X], and the set of isomorphism classes of simple objects Irr(C).
The semi-simplicity axiom implies that for any object X, M or(X, X) is a finite dimensional C∗-algebra
over C, hence a multi-matrix algebra. It is easy to see that each summand of the matrix algebra corresponds
to an equivalence class of simple objects, and the dimension of the matrix algebra corresponding to a simple
object Y is the square of the multiplicity with which Y occurs in X. In general for a simple object Y and
any object X, we denote by N Y
X the natural number describing the multiplicity with which [Y ] appears in
the simple object decomposition of X. If X is equivalent to a subobject of Y , we write X ≺ Y . We often
write X ⊗ Y simply as XY for objects X and Y .
For two simple objects X and Y , we have that [X ⊗ Y ] ∼= ⊕ZN Z
product of X and Y decomposes as a direct sum of simple objects of which N Z
object Z. The N Z
XY specify the fusion rules of the tensor category and are a critical piece of data.
XY [Z]. These means that the the tensor
XY are equivalent to the simple
4
COREY JONES
The fusion algebra is the complex linear span of isomorphism classes of simple objects C[Irr(C)], with
multiplication given by linear extension of the fusion rules. This algebra has a ∗-involution defined by
[X]∗ = [X] and extended conjugate-linearly. This algebra is a central object of study in approximation and
rigidity theory for rigid C∗-tensor categories.
For a more detailed discussion and analysis of the axioms of a rigid C∗-tensor category, see the paper of
Longo and Roberts [19], where these categories were first defined and studied with this axiomatization.
In a rigid C∗-tensor category, we can define the statistical dimension of an object d(X) = inf(R,R)RR,
where the infimum is taken over all solutions to the conjugate equations for an object X. The function
d( . ) : Obj(C) → R+ depends on objects only up to unitary isomorphism. It is multiplicative and additive
and satisfies d(X) = d(X) for any dual of X. We called solutions to the conjugate equations standard if
R = R = d(X) 1
2 , and such solutions are essentially unique. For standard solutions of the conjugate
equations, we have a well defined trace T rX on endomorphism spaces M or(X, X) given by
T rX (f ) = R∗(1X ⊗ f )R = R
∗
(f ⊗ 1X )R ∈ M or(id, id) ∼= C
This trace does not depend on the choice of dual for X or on the choice of standard solutions. We note
that T r(1X ) = d(X). See [19] for details.
2.2. The Tube Algebra. The tube algebra A of a semi-simple rigid tensor category C was introduced
by Ocneanu in [26]. This algebra has proved useful for computing the Drinfeld center Z(C), since finite
dimensional irreducible representations of A are in one-to-one correspondence with simple objects of Z(C)
(see [12] and [22]). Stefaan Vaes has pointed out that in general, arbitrary representations of A are in
one-to-one correspondence with objects in the category Z(ind-C) studied by Neshveyev and Yamashita in
[23].
For a rigid C∗-tensor category C, choose a representative Xk ∈ k for k ∈ Irr(C). We choose the strict
tensor identity id to represent its class and label it with index 0, so that X0 = id.
The tube algebra is defined as the algebraic direct sum (finiteley many non-zero terms)
A := ⊕i,j,k∈Irr(C)M or(Xk ⊗ Xi, Xj ⊗ Xk)
An element x ∈ A is given by a sequence xk
i,j ∈ M or(Xk ⊗ Xi, Xj ⊗ Xk) with only finitely many terms
non-zero. Recall for a simple object α and arbitrary β ∈ Obj(C), M or(α, β) has a Hilbert space structure
with inner product defined by η∗ξ = (cid:104)ξ, η(cid:105)1α. Note that this inner product differs by the tracial inner
product by a factor of d(α).
A carries the structure of an associative ∗-algebra, defined by
(cid:88)
(cid:88)
s,m,l∈Irr(C)
(x#)k
i,j = (R
(x · y)k
i,j =
(1j ⊗ V ∗)(xm
s,j ⊗ 1l)(1m ⊗ yl
i,s)(V ⊗ 1i)
V ∈onb(Xk, Xm⊗Xl)
∗
k ⊗ 1j ⊗ 1k)(1k ⊗ (xk
j,i)∗ ⊗ 1k)(1k ⊗ 1i ⊗ Rk)
where Rk ∈ M or(id, X k⊗Xk) and Rk ∈ M or(id, Xk⊗X k) are standard solutions to the conjugate equations
for Xk. We remark that we denote the ∗-involution by # to avoid confusion with the ∗-involution from C.
In the first sum above, onb denotes an orthonormal basis with respect to the inner product described above,
and we may have onb(Xk, Xm ⊗ Xl) = ∅ if Xk is not a sub-object of Xm ⊗ Xl. We mention the above
compact form for the definition of the tube algebra was borrowed from Stefaan Vaes.
QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
5
We define the subspaces Ak
i,j := M or(Xk ⊗ Xi, Xj ⊗ Xk) ⊂ A. For arbitrary objects α ∈ Obj(C), we have
a natural map Ψ : M or(α ⊗ Xi, Xj ⊗ α) → A given by
(cid:88)
(cid:88)
Ψ(f ) =
k≺α
V ∈onb(k,α)
(1j ⊗ V ∗)f (V ⊗ 1i).
We will use this map later, in our analysis of the categories (G2)q. For each k ∈ Irr(C), there is a projection
pm ∈ A0
m,m given by pm := 1m ∈ M or(id ⊗ Xm, Xm ⊗ id) ∈ A. In particular (pm)k
We see that A0,0 = p0Ap0 is a corner of the tube algebra. This corner is a unital ∗-algebra. Recall the
fusion algebra of C is the complex linear span of isomorphism classes of simple objects C[Irr(C)]. Multiplica-
tion is the linear extension of fusion rules and ∗ is given on basis elements by the duality. From the definition
of multiplication in A, one easily sees the following:
i,j = δk,0δi,jδj,m1m.
Proposition 2.1. The fusion algebra C[Irr(C)] is ∗-isomorphic to A0,0, via the map [Xk] → 1k ∈ (Xk ⊗
id, id ⊗ Xk) ∈ Ak
0,0.
2.3. Representations and property (T).
Definition 2.2. Rep(A) is the category whose objects are non-degenerate ∗-homomorphisms π : A →
B(H) for some Hilbert space H, and whose morphisms are bounded intertwiners.
This category can be given the structure of a braided C∗-tensor category, though in general it is not
rigid. In the case that C is fusion (Irr(C) < ∞), the category of finite dimensional representations is tensor
equivalent to the Drinfeld center Z(C) (see [7] for details).
Definition 2.3. A ∗-homomorphism π : C[Irr(C)] → B(H) of the fusion algebra is admissible if there
exists a π ∈ Rep(A) such that πA0,0 is unitarily equivalent to π.
We remark that this is not the original definition of admissible representation given by Popa and Vaes in
[32], but is equivalent by [10], Corollary 6.9.
Proposition 2.4. ([10], Corolloary 4.8) Let φ : A0,0 → C be a linear functional. The following are
equivalent:
(1) φ is a vector state in an admissible representation.
(2) φ(p0) = 1, and φ(x# · x) ≥ 0 for all x ∈ A0,k and k ∈ Λ.
We call the collection of functionals satisfying the equivalent conditions of the above proposition annular
states, denoted Φ0(A). The word annular comes from the correspondence between representations of the
tube algebra and representations of the affine annular category of a planar algebra introduced by Jones (see
[16], [15], [7], [10]). The idea is that annular states play a similar role in our representation theory to positive
definite functions in the representation theory of groups.
Every φ ∈ (cid:96)∞(Irr(C)) defines a linear functional φ on the fusion algebra C[Irr(C)] given by
(cid:88)
φ(
(cid:88)
ck[Xk]) =
ckd(Xk)φ([Xk])
Definition 2.5. φ ∈ (cid:96)∞(Irr(C)) is a cp-multiplier if φ is an annular state.
k
k
Again, this is not the original definition of cp-multiplier introduced by Popa and Vaes. For their original
definition see [32], Defintion 3.1. That the definition presented here is equivalent to theirs follows from [10],
Theorem 6.6. Without going into the details of the original definition, we describe its motivation as follows:
6
COREY JONES
Associated to a subfactor N ⊆ M , Popa defined a von Neumann algebra inclusion T ⊆ S, called the
symmetric enveloping inclusion [27]. Popa gave definitions for analytical properties of a subfactor in terms
of sequences of UCP maps φi : S → S which are T bimodular, satisfying certain convergence properties
(see [30]). It turns out that decomposing L2(S) as a T − T bimodule shows that such maps are determined
by functions on Irr(C), where C is the category of "even" bimodules of the subfactor. Their definition
of cp-multiplier is a necessary and sufficient condition for a function φ ∈ (cid:96)∞(Irr(C)) to produce a T -T
bimodular UCP map on S in the prescribed way. Furthermore, conditions for approximation and rigidity
properties on the sequence of UCP maps can be directly translated into conditions on the corresponding
functions in (cid:96)∞(Irr(C)) with out reference to the symmetric enveloping inclusion. This results in definitions
for approximation and rigidity properties for the category C without reference to a subfactor.
Definition 2.6. ([32], Definition 5.10) A rigid C∗-tensor category has property (T) if every sequence of
cp-multipliers φi which converges to 1 pointwise on Irr(C) converges uniformly.
There exists a C∗-closure, C∗(C), of the fusion algebra introduced in [32] such that continuous Hilbert
space representations of this C∗-algebra are in one-to-one correspondence with admissible representations.
It can be shown for X ∈ Irr(C), in any admissible representation π we have (cid:107)π(X)(cid:107) ≤ d(X) (see for example
[32], Proposition 4.2 or [10], Lemma 4.4).
This bound permits the definition of a universal admissible representation πu. C∗(C) is defined by taking
the C∗-completion of πu(C[Irr(C)]). The norm of an element in the fusion algebra can be written
We remark that in fact we can define such a universal representation and C∗-algebra for the entire tube
algebra A but we do not need this here (see [10], Definition 4.6).
(cid:107)f(cid:107)u = sup
φ∈Φ0(A)
φ(x# · x)
1
2 .
As with groups, there are many equivalent characterizations of property (T). We record the following
provided by Popa and Vaes:
Proposition 2.7. ([32], Proposition 5.5) C has property (T) if and only if there exists a projection
p ∈ C∗(C) such that αp = d(α)p for all α ∈ Irr(C).
The so-called trivial representation 1C is the one dimensional representation of the fusion algebra spanned
by v0 such that 1C([X])v0 = d(X)v0 for all X ∈ Irr(C). Viewing 1C as an annular state ([10], Lemma 4.11),
the corresponding cp-multiplier is the constant function 1 in (cid:96)∞(Irr(C)).
For categories with abelian fusion rules (for example, all braided categories), C∗(C) ∼= C(Z) for some
compact Hausdorff space Z. Points in Z correspond to one-dimensional representations of the fusion algebra,
so 1C ∈ Z. We have the following easy consequence of the above proposition:
Corollary 2.8. If C has abelian fusion rules so that C∗(C) ∼= C(Z) for some compact Hausdorff space
Z, then C has property (T) if and only if the trivial representation 1C is isolated in Z
Proof. If C has abelian fusion rules C∗(C) ∼= C(Z), where Z is the spectrum of C∗(C). If 1C is isolated in
the spectrum, then the characteristic function δ{1C} ∈ C(Z) ∼= C∗(C) is a projection satisfying the required
property. Conversely, if we had such a projection p, then it could be represented by the characteristic
function of some clopen set Y ⊆ Z. Since αp = d(α)p, this implies that when viewing an object α as a
function on Z, αY = d(α) = α(1C). Extending by linearity, we see that for an arbitrary element in the
QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
7
fusion algebra β, βY = 1C(β). This equality extends to the C∗-closure C∗(C) ∼= C(Z). Since the points of
Y are not separated by C(Z) from 1C, by the Stone-Weierstrass theorem we have Y = {1C}, hence {1C} is
(cid:3)
clopen, hence 1C is isolated in Z.
3. (G2)q categories
There are many ways to describe rigid C∗-tensor categories. One of the most useful is the planar algebra
approach introduced by Jones [14]. The idea is to use formal linear combinations of planar diagrams to
represent morphisms in your category. These diagrams satisfy some linear dependences called skein rela-
tions in modern parlance. The most famous skein relations are the ones defining the Jones and HOMFLY
polynomials.
The (G2)q categories we describe are a particularly nice type of planar algebra called a trivalent category.
These were introduced in their current form by Morrison, Peters, and Snyder [20]. Using dimension restric-
tions on morphism spaces as a notion of "small", they were able to classify the "smallest" examples. The
(G2)q categories appear in their classification list.
Definition 3.1. ( [20], Definition 2.4) A trivalent category C is a non-degenerate, evaluable, pivotal
category over C, with a tensor generating object X satisfying dim M or(id, X) = 0, dim M or(id, X ⊗X) = 1,
dim M or(id, X ⊗ X ⊗ X) = 1, generated (as a planar algebra) by a trivalent vertex for X .
We summarize the basic properties of trivalent categories:
(1) Objects in the category can be represented by N∪{0}, and correspond to tensor powers of a generating
object X.
(2) M or(k, m) is the complex linear span of isotopy classes of planar trivalent graphs embedded in a
rectangle, with m boundary points on the top of the rectangle, k boundary points on the bottom,
and no boundary points on the sides of the rectangle. These diagrams are subject to skein relations,
which are linear dependences among the trivalent graphs which make M or(k, m) finite dimensional.
(Note: We consider graphs with no vertices at all, namely line segments attached to the boundaries,
as trivalent graphs)
(3) M or(0, 0) ∼= C. In other words, our skein relations reduce every closed trivalent graph to a scalar
multiple of the empty trivalent graph. Identifying the empty graph with 1 ∈ C, this means we have
associated to every closed trivalent graph a complex number.
(4) Composition of morphisms is vertical stacking of rectangles.
(5) Tensor product on objects is addition of natural numbers, on morphisms it is horizontal stacking of
rectangles.
(6) Duality is given by rotation by π or −π (these manifestly agree in our setting).
(7) The linear functional T r : M or(k, k) → C given by connecting the top strings of the rectangle to the
bottom is non-degenerate.
Definition 3.2. A trivalent category is a C∗-trivalent category if the maps ∗ : M or(k, m) → M or(m, k)
given by reflecting graphs across a horizontal line and conjugating complex coefficients are well-defined
modulo the skein relations, and T r(x∗x) ≥ 0 for every x ∈ M or(k, m), and k, m ∈ N(cid:83){0}.
From a C∗-trivalent category, we can construct a rigid C∗-tensor category as follows: First, it can be shown
that a category satisfying all these conditions has a negligible category ideal, generated by diagrams with
T r(x∗x)=0. Quotienting by this produces a trivalent category with condition 8 replaced by T r(x∗x) > 0 .
8
COREY JONES
Next, we take the projection completion. Objects in this category will be projections living in some M or(k, k).
For two projections P ∈ M or(k, k), Q ∈ M or(m, m), M or(P, Q) = {f ∈ M or(k, m) : Qf P = f}. Now we
formally add direct sums to the category. The resulting category will have objects direct sums of projections,
and morphisms matrices of the morphisms between projections. The result is a rigid C∗-tensor category,
which we also call C.
Notice the duality map we have defined is automatically pivotal. Also the strict tensor identity id is
given by the empty diagram. Another consequence of the definitions is that the generating object X is
symmetrically self-dual (see [5], Definition 2.10).
The (G2)q trivalent categories were introduced by Kuperberg in [17] and [18]. Kuperberg showed that
these categories are equivalent to the category of (type 1) finite dimensional representations of the Drinfeld-
Jimbo quantum groups Uq(g2).
To define a trivalent category, it suffices to specify a set of skein relations.
In general it is a difficult
problem to determine whether an set of skein relations produces a trivalent category. In particular, one has
to verify that your relations are consistent and evaluable. Otherwise you may end up with M or(0, 0) being 0
dimensional, or with infinite dimensional morphism spaces. Kuperberg showed the following skein relations
are indeed consistent and evaluable, resulting in a trivalent category. The skein theory we present for (G2)q
can be found in [20], Definition 5.21. It differs from Kuperberg's description in two ways: The trivalent
vertex is normalized, and the q2 here is Kuperberg's q.
Definition 3.3. (G2)q for strictly positive q is the trivalent category defined by the following skein
relations:
= δ := q10 + q8 + q2 + 1 + q−2 + q−8 + q−10
= 0
=
= c
= a
+ b
+
+
QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
+ g
= f
Where
+ rotations
+ rotations
9
a =
b =
q2 + q−2
(q + 1 + q−1)(q − 1 + q−1)(q4 + q−4)
(q + 1 + q−1)(q − 1 + q−1)(q4 + q−4)2
1
c = − q2 − 1 + q−2
q4 + q−4
f = −
(q + 1 + q−1)(q − 1 + q−1)(q4 + q−4)
1
g = −
(q + 1 + q−1)2(q − 1 + q−1)2(q4 + q−4)2
1
[17], [18], and [20] shows this category is actually spherical. The duality maps ∪ and ∩ provide standard
solutions for the simple object (minimal projection) spanning M or(1, 1). This is the object X, which tensor
generates our category.
Kuperberg showed that this category is isomorphic (not just equivalent) to the spherical category generated
by the 7-dimensional fundamental representation (which we also call X) of Uq(g2) (see [18] Theorem 5.1). A
single string corresponds to the object X in Rep(Uq(g2)), hence the natural number k an an object in (G2)q
corresponds to the object X⊗k in Rep(Uq(g2)). Since X tensor generates Rep(Uq(g2)), we have the whole
category appearing.
In both Kuperberg's work and Morrison, Peters, and Snyder's no ∗-structure is considered. However,
Uq(g2) has a natural ∗-structure for positive q (cid:54)= 1 (along with all Drinfeld-Jimbo quantum groups), and
it is shown, for example, in [23], Chapter 2.4, that the category of finite dimensional ∗-representations is a
rigid C∗-tensor category. Every type 1 finite dimensional representation of Uq(g2) for q > 0 is unitarizable
([6], Chapter 10) and thus the rigid C∗-tensor category of finite dimensional unitary type 1 representations
Cq(g2) is monoidally equivalent to Rep(Uq(g2)). In the appendix, we show that the ∗-structure from quantum
groups transports to (G2)q as the trivalent ∗-structure defined by reflecting a diagram across a horizontal
line (see Proposition 5.1 in the Appendix). Thus we can consider (G2)q as a C∗-trivalent category.
To prove (G2)q has property (T), we need to know the structure of M or(2, 2). M or(2, 2) is a 4-dimensional
abelian C∗-algebra. To determine the minimal projections, we set
ξ :=(cid:112)δ2c4 + 2δ(c4 − 2c3 − c2 + 4c + 2) + (c2 − 2c − 1)2 =
This is manifestly non-zero for q (cid:54)= 1 and q > 0.
(1 + q2)2(1 − q2 + q6 − q8 + q10 − q14 + q16)
.
q6(1 + q8)
Proposition 3.4. ([20], Proposition 4.16) The minimal idempotents in the finite dimensional abelian
algebra M or(2, 2) are given by
and the two idempotents
1
δ
,
10
COREY JONES
y± =
−(δ+1)c2±ξ+1
±2ξ
+ δ(c2−2c−2)∓ξ+c2−2c−1
±2δξ
− δ(c+2)c±ξ+c2+1
±2ξ
+ δc+δ+c±ξ
.
In our setting, we see that the idempotents are in fact projections, since our basis is self-adjoint and all
coefficients are real numbers. These projections correspond to simple objects in the rigid C∗-tensor category
underlying (G2)q.
By Kuperberg's isomorphism, the fusion algebra of the underlying projection category of (G2)q is iso-
morphic to the fusion algebra of the category C q(g2) for positive q (cid:54)= 1, which in turn is isomorphic to the
complexification of the representation ring R(G2). It is well known that for compact, simply connected, sim-
ple Lie groups G, the representation ring R(G) is isomorphic to the ring of polynomials in the fundamental
representations. For a specific reference for G2 see [9], and in general, see [1], Theorem 6.41. This implies the
fusion algebra of (G2)q is the (commutative) complex polynomial algebra in 2 self-adjoint variables C[Z1, Z2],
where Z1 and Z2 correspond to the 14 and 7 dimensional fundamental representations of the quantum group
Uq(g2) respectively. (Note that self-adjointness of the variables follows from self-duality of the corresponding
representations).
The fusion graph with respect to X is given by
Here, the vertex at the bottom corresponds to the identity, the next highest vertex corresponds to X
itself, etc.
3.1. Property (T) for (G2)q. Now we consider the tube algebra A of the categories (G2)q for some positive
q (cid:54)= 1. In our analysis all such q will yield the same results, so we supress the dependence of the tube algebra
A on q for notational convenience. Recall simple objects in the category correspond to minimal projections
in some M or(k, k). Let us choose our set of representatives of projections so that it contains the empty
diagram id, the single string X, and the two projections y+ and y−, representing their equivalence classes.
For x ∈ M or(k, k), we let i : M or(k, k) → M or(k ⊗ id, id ⊗ k) be the canonical identification. Then define
∆(x) := Ψ(i(x)) ∈ A0,0.
where Ψ is defined in the discussion of the tube algebra. In our setting we see that the map in Proposition
2.1 is defined by applying ∆ to a projection.
Translating the fusion algebra description to our setting, the variable Z2 is represented by ∆(X), while
Z1 is represented by the projection ∆(y+) ∈ M or(2, 2). We see that
A0,0
∼= C[∆(y+), ∆(X)].
Xy+y- QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
11
Going back to our expression for y+, we see that δc+δ+c
= (1+q2+q4)(1+q8)
q4(1+q2)2
(cid:54)= 0. Thus
ξ
=
q4(1+q2)2
(1+q2+q4)(1+q8)
(y+) − −(δ+1)c2+ξ+1
2ξ
− δ(c2−2c−2)−ξ+c2−2c−1
2δξ
+ δ(c+2)c+ξ+c2+1
2ξ
(cid:32)
(cid:16)
(cid:33)
.
(cid:17)
∆(X)
.
Then since ∆(
) = ∆(X), we have
∆(
) =
q4(1+q2)2
(1+q2+q4)(1+q8)
∆(y+) − −(δ+1)c2+ξ+1
2ξ
∆(X)2 − δ(c2−2c−2)−ξ+c2−2c−1
2ξ
1 + δ(c+2)c+ξ+c2+1
2ξ
We denote H :=
. Since our polynomial expression for ∆(H) is linear in ∆(y+) and the terms
with powers of ∆(X) contain no ∆(y+) terms, we can perform an invertible transformation implementing a
change of basis, and write an arbitrary polynomial in ∆(y+) and ∆(X) as a polynomial in ∆(H) and ∆(X),
so that
A0,0
∼= C[∆(H), ∆(X)].
Therefore irreducible representations of A0,0 are 1-dimensional, and they are defined by assigning numbers
to ∆(H) and ∆(X). Let us denote by α the value assigned to ∆(H) and t the value assigned to ∆(X) in
our 1-dimensional representation. Let γα,t : A0,0 → C denote the 1-dimensional representation viewed as a
functional, given by evaluating polynomials in C[∆(H), ∆(X)] at the point (α, t).
The key point is that while arbitrary values of α and t determine a representation of A0,0, not all are
annular states (hence admissible representations). Recall that γα,t is admissible if and only if γα,t(x#· x) ≥ 0
for all x ∈ A0,k and for all k ∈ Irr((G2)q).
As a first restriction, for our representation to be admissible, t ∈ R since the object corresponding to a
single string is self-dual and our representation must be a ∗-representation. We also must have α ≥ 0, since
∆(H) = T # · T , where T ∈ M or(X ⊗ id, X ⊗ X) ⊆ AX
0,X is given by the trivalent vertex T :=
We know also that t ≤ δ by [32], Propsoition 4.2, or [10], Lemma 4.4. Here δ is the value of the closed
circle defined in terms of q in the description of the skein theory. This restricts the possible one dimensional
admissible representations to some subset Z ⊆ {(α, t) ⊆ R2 : α ≥ 0, t ∈ [−δ, δ]}.
Since the fusion algebra is isomorphic to the polynomial algebra in two self adjoint variables, and ir-
reducible representations correspond to evaluation at points Z ⊆ R2, the weak-∗ topology on Z as linear
functionals on C∗((G2)q) agrees with the (subspace) topology on the plane. The trivial representation cor-
responds to the point (0, δ). We will show that for positive q (cid:54)= 1, there is a neighborhood of the point (0, δ)
in the rectangle R+ × [−δ, δ] such that the functional γα,t is not an annular state.
To see this, let s :=
∈ M or(X ⊗ id, y− ⊗ X). We view s ∈ AX
0,y− ⊂ A.
For each pair (α, t), define the function f (α, t) := γα,t(s# · s) . This can be directly computed from the
representation of y− in terms of our planar algebra basis, and we obtain
f (α, t) = δ
−(δ + 1)c2 − ξ + 1
−2ξ
+ t2 δ(c2 − 2c − 2) + ξ + c2 − 2c − 1
−2δξ
− α
δ(c + 2)c − ξ + c2 + 1
−2ξ
+ t
δc + δ + c
−ξ
.
y- 12
COREY JONES
By construction, if the functional corresponding to (α, t) is an annular state, f (α, t) must be non-negative.
Proposition 3.5. For all positive q (cid:54)= 1, (G2)q has property (T ).
Proof. Since y− is a minimal projection in M or(2, 2) and is not equivalent to id in (G2)q, f (0, δ) = 0.
This can also be seen by direct computation. Let v := (x, y) ∈ R2 be a non-zero vector in the fourth quadrant
of the plane (including the axes), or in other words x ≥ 0 and y ≤ 0, but (x, y) (cid:54)= (0, 0). We wish to show
that f (x, δ + y) < 0 for sufficiently small v. This will demonstrate that in a neighborhood of (0, δ) in
R+ × [−δ, δ], the function f (α, t) will be strictly negative, hence the representation corresponding to γα,t is
not admissible.
We see that for positive q (cid:54)= 1,
(0,δ) =
∂f
∂α
δ(c + 2)c − ξ + c2 + 1
2ξ
= − 1 + q2 + 2q4 + q6 + q8
(q + q3)2
This is always strictly negative (for q (cid:54)= 0). We compute
(0,δ) =
∂f
∂t
(δ + 1)(c2 − c − 1)
−ξ
− 1 =
(−1 + q2)2(1 + q2 + q4)
q4
< 0.
> 0
∂f
∂v(0,δ) < 0 for v in the prescribed range. We remark that for q = 1, ∂f
proof breaks down as expected, since Rep(G2) is amenable.
This expression is strictly positive for all q (cid:54)= 1, q (cid:54)= 0. Therefore we have that the directional derivative
∂t (0,δ) = 0, hence this part of our
∂v(0,δ) is a continuous
Letting B denote the compact set of unit vectors in the fourth quadrant, since ∂f
function of v, there exists some M < 0 such that ∂f
∂v(0,δ) ≤ M < 0 for v ∈ B.
Now, it is straightforward to compute
∂2f
∂2t
(0,δ) =
2q6(1 + q2 + q4)
(1 + q2)2(1 + q2 + q4 + q6 + q8 + q10 + q12)
> 0,
and it is easy to see that all other second order partial derivatives are 0, and all higher order derivatives
with respect to both variables are 0. Then by Taylor's theorem, we have
f (x, δ + y) = x
(0,δ) + y
∂f
∂α
(0,δ) +
∂f
∂t
y2
2
∂2f
∂2t
(0,δ)
for arbitrary v = (x, y) in the fourth quadrant. Let v(cid:48) = 1(cid:107)v(cid:107) v. Since y2 ≤ (cid:107)v(cid:107)2, setting λ := 1
2
∂2f
∂2t (0,δ) gives
f (x, δ + y) = (cid:107)v(cid:107) ∂f
M
λ , then since M < 0, for 0 < (cid:107)v(cid:107) < , we see that f (x, y + δ) < 0. Therefore (G2)q has
(cid:3)
∂v(cid:48)(0,δ) + y2λ ≤ (cid:107)v(cid:107)M + (cid:107)v(cid:107)2λ = (cid:107)v(cid:107)(M + (cid:107)v(cid:107)λ).
If we set =
property (T) for positive q (cid:54)= 1.
4. Concluding Remarks
(1) Our result provides another class of examples of subfactors with property (T ) standard invariant. To
see this, we recall that Popa gave an axiomatization of standard invariants of subfactors as standard
λ-lattices [29]. These are towers unital inclusions of finite dimensional C∗- algebras, together with
some extra data. From a rigid C∗-tensor C and object X, one can construct a standard λ-lattice by
QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
13
C ⊂ End(X) ⊂ End(XX) ⊂ End(XXX) ⊂ · · ·
∪
∪
∪
C
⊂ End(X) ⊂ End(XX) ⊂ · · ·
Then this tower will be the standard invariant of some subfactor N ⊆ M by [29]. The category
of N -N bimodules will be isomorphic to the subcategory of C generated by XX. Applying this
construction to (G2)q, since X is self-dual and appears as a sub-object of X ⊗ X, the even bimodules
of this subfactor will be a category equivalent to Cq(g2) (as opposed to a proper sub-category).
Therefore, as in [32], Theorem 8.1, this subfactor will have property (T) standard invariant.
(2) We hope that methods similar to those presented here will be useful to deduce property (T) (or lack
thereof) for other quantum group categories which have a nice planar algebra description, particularly
the BM W planar algebras, which describe the quantum SO(n) and SP (2n) categories.
5. APPENDIX: ∗-structure for (G2)q
The point of this section is to show that the ∗-structure we described above for (G2)q (reflection of
diagrams about a horizontal line) gives a C∗-trivalent category. To do this from the purely planar algebra
perspective is quite a daunting task, since we would have to explicitly construct all idempotents, show they
are self adjoint, and that they have positive trace. Fortunately for us, due to Kuperberg's isomorphism,
(G2)q can be realized as the category generated by the fundamental 7-dimensional representation X of the
Hopf ∗-algebra Uq(g2), which carries with it a naturally occurring ∗-structure.
To elaborate, we know that quantum groups Uq(g) with positive q have a natural ∗-structure and that every
(type 1) finite dimensional representation is unitarizable, which means it is equivalent to a ∗-representation
of the Hopf ∗-algebra (see [6], Chapter 10). To obtain a rigid C∗-tensor category, we restrict our attention
to the finite dimensional type 1 unitary representations, and since every representation is unitarizable, this
category is monoidally equivalent to the whole category of finite dimensional type 1 representations.
Kuperberg's equivalence from Rep(Uq(g2)) to (G2)q used the fact that the fundamental 7-dimensional
representation was symmetrically self-dual. But being symmetrically self-dual depends on the specific choice
of duality maps and most importantly on the map implementing the equivalence from π to πc (the standard
dual, defined using the Hopf algebra antipode). For us to have a C∗-planar algebra, we need π to be unitarily
symmetrically self dual. Even if π is unitary, the canonical dual πc is not necessarily unitary, although we
know it is equivalent to a unitary representation. In [23], they explicitly identify the unitary dual, π for all
representations of Uq(g). Using this information, we explicitly compute the unitary intertwiner T ∈ (π, π) in
the matrix representations of π and π, and show that T along with the choices of (mutually adjoint) duality
maps implements a unitary symmetric self-duality on π.
Furthermore, these duality maps provide our rigid C∗ tensor category with a pivotal structure, such that
duality is compatible with the ∗-structure. Using Kuperberg's result, our category (forgetting the ∗-structure)
must be the trivalent category described by our skein theory. Then, compatibility of the ∗-structure with
the duality functor forces the ∗-structure described in Definition 3.2 for (G2)q, given by horizontal reflection
of diagrams.
We give a brief description of the Drinfeld-Jimbo Hopf ∗-algebra Uq(g2), following [23], Definition 2.4.1.
Let α1, α2 be the standard choice of simple roots of the G2 root system, pictured below:
14
COREY JONES
(cid:33)
(cid:32)
2 −3
−1
2
Let (aij) =
Aij := diai,j = (αi, αj) is the inner product matrix, given by
2 −3
−3
6
A =
(cid:32)
(cid:33)
be the Cartan matrix for the G2 root system. Let d1 := 1, d2 := 3, so that
For q > 0, q (cid:54)= 1, let qi = qdi , i = 1, 2. Then Uq(g2) is defined as the universal unital algebra generated
by elements Ei, FiKi, K−1
i
, i = 1, 2, satisfying the relations
KiK−1
KiEjK−1
i = K−1
i = qaij
i = q
i Ki = 1, KiKj = KjKi
−aij
i Ej, KiFjK−1
i
Ki − K−1
(cid:34)
qi − q−1
(cid:35)
i
i
1 − aij
Fj
[Ei, Fj] = δij
1−aij(cid:88)
1−aij(cid:88)
k=0
k=0
(−1)k
(cid:34)
(−1)k
For i (cid:54)= j
and
(cid:35)
k
1 − aij
k
i EjE1−aij−k
Ek
i
= 0
qi
i FjF 1−aij−k
F k
i
= 0
qi
(cid:34)
(cid:35)
m
k
[k]qi ![m−k]qi ! , [m]qi! = [m]qi[m − 1]qi
where
This algebra is a Hopf ∗-algebra with coproduct (cid:52)q define by
[m]qi !
=
qi
. . . [1]qi, and [n]qi = qn
i −q
qi−q
−n
i
−1
i
.
(cid:52)q(Ki) = Ki ⊗ Ki, (cid:52)q(Ei) = Ei ⊗ 1 + Ki ⊗ Ei, (cid:52)q(Fi) = Fi ⊗ K−1
i + 1 ⊗ Fi
and with involution given by K∗
The counit q and the antipode Sq are given by
i = Ki, E∗
i = FiKi, F ∗
i = K−1
i Ei.
q(Ki) = 1, q(Ei) = q(Fi) = 0,
Sq(Ki) = K−1
i
, Sq(Ei) = −K−1
i Ei, Sq(Fi) = −FiKi
A representation is type 1 if it decomposes as the direct sum of weight spaces (see [23], Definition 2.4.3).
The category of type 1 finite dimensional representations of this algebra has the structure of a rigid tensor
category, with tensor product defined using the comultiplication, and duality the vector space dual with
αα12 QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
15
action induced by the antipode. This is a standard result for Hopf algebras. A unitary representation (or
∗-representation) is simply a ∗-algebra homomorphism π : Uq(g) → B(H) for some Hilbert space H. We
consider here only finite dimensional type 1 unitary representations. This category is a rigid C∗-tensor
category. We refer the reader to [23], Chapter 2.4 for details. As we have mentioned above, every finite
dimensional representation of these Hopf algebras is unitarizable (see [6], Chapter 10 ) for positive q, so
that the category of unitary representations is monoidally equivalent to the category of all finite dimensional
representations. A key point in quantum group theory is that representations for positive q are in one-to-one
correspondence with classical representations of the Lie algebras, and have the same dimension (as vector
spaces) as the classical representations.
We will give an explicit unitary realization of the fundamental 7-dimensional representation X of Uq(g2).
Let v0 be the highest weight vector, normalized so that (cid:104)v0, v0(cid:105) = 1. Then define H as the Hilbert space
with orthonormal basis
v0, v1 := q
2 F1v0, v2 := q2F2F1v0, v3 := q3[2]
1
v5 := q
9
2 [2]−1F2F1F1F2F1v0, v6 := q5[2]−1
1
2
q F1F2F1v0, v4 := q3[2]−1
q F1F2F1F1F2F1v0.
q F1F1F2F1v0,
The action π of Uq(g2) on vectors can be worked out from the commutation relations. This yields a
∗-representation on which the actions of E1, E2, F1, F2 can be worked out explicitly. For example, we have
the matrix representations with respect to the above basis given by
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2
q 1
0
0
0
0
0
1
0 q[2]
2
q
0
0
0
0
0
0
0
0
0
0
0
1
[2]
2
q
0
0
0
0
0
0
0
0
0
0
0
0
0
0 q 1
0
0
2
0
0
0
0
0
0
0
2
0
q 3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0 q 3
0
0
0
0
2
0
0
0
0
0
0
0
π(E1) =
π(E2) =
π(K1) = diag(q, q−1, q2, 1, q−2, q, q−1) and π(K2) = diag(1, q3, q−3, 1, q3, q−3, 1)
If f ∈ B(H, K), then define j(f ) ∈ B(K, H) by j(f )ξ = f∗ξ for ξ ∈ K. Let ρ be the half-sum of
1 K 6
2 .
2ρ xK2ρ. Then we see that π(K2ρ) is a positive invertible operator (diagonal with
2 ∈ B(H). Then j(W ) ∈ B(H), and we define
positive roots and thus 2ρ the sum of positive roots. Then 2ρ = 10α1 + 6α2, so we define K2ρ := K 10
Then we have S2
respect to our chosen basis), and we define W := π(K−1
π( . ) = j(W )πc( . )j(W −1). This is manifestly equivalent to πc but has the advantage of being unitary.
q (x) = K−1
2ρ ) 1
Alternatively, the dual representation can be given using the unitary antipode Rq, defined by
Rq(Ki) = K−1
i
, Rq(Ei) = −qiK−1
i Ei, Rq(Fi) = −q−1
i FiKi
16
COREY JONES
Then if π : Uq(q2) → B(H), the unitary dual π : Uq(g2) → B(H) can be realized as π(x)ξ = π( Rq(x))∗ξ
for x ∈ Uq(g2).
(We note the standard dual πc is given by the same formula, with the standard antipode Sq in place of
Rq).
Since W is diagonal in the basis described above, the ith eigenvalue, Wi > 0, is clear. By inspection,
we see that WiW7−i+1 = 1. Consider the map T : H → H given by T (vi) = (−1)i+1v7−i+1. Using the
description provided above, it is straightforward to check that T ∈ Hom(π, π). Furthermore, T ∗ = T −1, and
is given by the same formula as T , with the bars reversed, and we have T ∗j(W ) = W −1T ∗
Consider the map r ∈ (C, H ⊗ H) given by r(1) =(cid:80)
i vi ⊗ vi (this map does not depend on basis). It
is easy to see that r ∈ (, π ⊗ πc). By definition j(W ) ∈ (πc, π). Then define R := (1 ⊗ (T ∗j(W )))r =
(W ⊗ T ∗)r ∈ (, π ⊗ π). We claim that the pair (R, R∗) provide a standard, symmetrically self-dual solution
to the conjugate equations for the object π. It is straightforward to check that
(1H ⊗ R∗)(R ⊗ 1H ) = (R∗ ⊗ 1H )(1H ⊗ R) = 1H .
(W −1 ⊗ j(W )) ◦ r = r. Similarly, if r(1) =(cid:80)
Now, to see symmetric self duality, we note that since W is self adjoint, r∗ ◦ (W −1 ⊗ j(W )) = r∗, and
j vj ⊗ vj ∈ B(C, H ⊗ H), then since T is a self adjoint unitary,
r∗ ◦ (T ∗ ⊗ T ) = r∗ and (T ⊗ T ∗) ◦ r = r. Using these relations and the fact that r, r themselves solve the
duality equations in the category of Hilbert spaces, we have
(1H ⊗ 1H ⊗ R∗) ◦ (1H ⊗ R ⊗ 1H ) ◦ R = R = (R∗ ⊗ 1H ⊗ 1H ) ◦ (1H ⊗ R ⊗ 1H ).
We let X = (π, H) described above. We define C to be the strict (up to Hilbert space associativity)
C∗-tensor category generated by the symmetrically self dual object X, with unit object id := q (the trivial
representation of the quantum group given by the counit), and duality maps compositions of R and R∗ in
the obvious fashion. Then objects are given by tensor powers X⊗n, where n ∈ Z+, and morphisms are
intertwiners of the corresponding quantum group representations. It is easy to check that the duality maps
we have defined induce a pivotal structure on this category (essentially following from the fact that (r, r)
induce a pivotal structure on finite dimensional Hilbert spaces), hence there is an unambiguously defined
dual f ∈ M or(X n, X m) for every f ∈ M or(X m, X n). Also we see that our ∗-structure is compatible with
duality in the sense that f∗ = f
, which follows since our solutions to the conjugate equations are mutually
adjoint, and from the corresponding property for Hilbert spaces with (r, r).
∗
Proposition 5.1. (G2)q is a C∗-trivalent category.
Proof. Neglecting the ∗- structure, C is precisely the G2 algebraic spider defined by Kuperberg, hence
must be isomorphic to (G2)q ( [17], 3.2). We now have a ∗-structure on the morphism spaces, which is positive
with respect to the trace (since C is manifestly a C∗-category). Note that by construction ∪∗ = R∗ = ∩.
Using the fusion rules, M or(X, X ⊗ X) is one dimensional, and is spanned in the planar algebra by
a rotationally invariant vertex t ∈ (X, X ⊗ X). We have t normalized so that t ◦ t = 1X , where t =
(1X ⊗ R∗) ◦ (1X ⊗ 1X ⊗ R∗ ⊗ 1X ) ◦ (1X ⊗ t ⊗ 1X ⊗ 1X ) ◦ (R ⊗ 1X ⊗ 1X ) ∈ (X ⊗ X, X) is the dual of t. We
know that t∗ = λt for some λ ∈ C. This is a C∗-category so t∗t = λ1X must be positive, hence λ > 0. We
= λ(cid:48)t. But t = t∗∗ = λλ(cid:48)t, hence λ(cid:48) = λ−1. Since ∗ is compatible duality as described in the
also have t
= t∗, hence λ−1t = λt, but since λ > 0, it must be that
paragraph before this proposition, we have that t
λ = 1.
∗
∗
QUANTUM G2 CATEGORIES HAVE PROPERTY (T)
17
The ∗ of t and ∩ determine the ∗ on an arbitrary diagram in M or(k, m), and it is simply the reflection
(cid:3)
about a horizontal line. Therefore (G2)q is a C∗-trivalent category.
References
[1] J. F. Adams 1969. Lectures on Lie Groups. University of Chicago Press.
[2] Y. Arano 2014. Unitary spherical representations of Drinfeld doubles. arXiv:1410.6238. To appear in Journal fr die reine
und angewandte Mathematik.
[3] D. Bisch and S. Popa 1998. Examples of subfactors with property T standard invariant. Geom. Funct. Anal. 9.2, pp. 215 -
225.
[4] A. Brothier and V.F.R. Jones 2015. Hilbert Modules over a Planar Algebra and the Haagerup Property. arXiv:1503.02708.
To appear in Journal of Functional Analysis.
[5] A. Brothier, D. Penneys and M. Hartglass 2013. Rigid C∗-tensor categories of bimodules over interpolated free group factors.
arXiv:1208.5505v2.
[6] V. Chari and A. Pressley 1994. A Guide to Quantum Groups. Cambridge University Press.
[7] P. Das, S. Ghosh and V. Gupta 2012. Drinfeld center of a planar algebra. arXiv:1203.3958.
[8] K. De Commer, A. Freslon, and M. Yamashita 2014. CCAP for universal discrete quantum groups. Comm. Math. Phys.
331, pp. 677 - 701.
[9] W. Fulton and J. Harris 2004. Representation Theory: A First Course. Springer, Graduate Texts in Mathematics 129.
[10] S.K. Ghosh and C. Jones 2015. Annular representation theory for rigid C∗-tensor categories. arXiv:1502.06543v4. To
appear in Journal of Functional Analysis.
[11] R. Haag 1996. Local Quantum Physics: Fields, Particles, Algebras. Springer, Texts and Monographs in Physics.
[12] M. Izumi 1999. The structure of sectors associated with the Longo-Rehren inclusion I. General Theory. Commun. Math.
Phys. 213, pp. 127 - 179
[13] V.F.R. Jones 1983. Index for Subfactors. Invent. Math. 73, pp. 1 - 25.
[14] V.F.R. Jones 2000. Planar Algebras I. arxiv:math.QA/9909027.
[15] V.F.R. Jones 2001. The annular structure of subfactors. Essays on geometry and related topics, Vol.1,2. pp.401 - 463.
Monogr. Enseign. Math., 38.
[16] V.F.R. Jones and S. Reznikoff 2006. Hilbert space representations of the annular Temperley-Lieb algebra. Pacific J. Math.
228.2 pp. 219 - 248.
[17] G. Kuperberg 1994. The quantum G2 link invariant. International J. Math. 5. No.1 pp. 61 - 85.
[18] G. Kuperberg 1996. Spiders for rank 2 Lie algebras. Comm. Math. Phys. 180, pp. 109 - 151.
[19] R. Longo and J.E. Roberts 1997. A theory of dimension. K-theory 11.2, pp. 103 - 159.
[20] S. Morrison, E. Peters, and N. Snyder 2015. Categories generated by a trivalent vertex. arXiv:1501.06869.
[21] M. Muger 2003. From subfactors to categories and topology I: Frobenius algebras and Morita equivalence of tensor cate-
gories. Journal of Pure and Applied Algebra 180.1, pp. 81 - 157.
[22] M. Muger 2003. From subfactors to categories and topology II: The quantum double of tensor categories and subfactors.
Journal of Pure and Applied Algebra 180.1, pp. 159 - 219.
[23] S. Neshveyev and L. Tuset 2013. Compact Quantum Groups and Their Representation Categories. Specialized Courses,
Vol. 20, SMF.
[24] S. Neshveyev and M. Yamashita 2015. Drinfeld center and representation theory for monoidal categories.
arXiv:1501.07390v1.
[25] A. Ocneanu 1988. Quantized groups, string algebras and Galois theory for algebras. Operator algebras and applications,
London Math. Soc. Lecture Note Ser., 136: pp. 119 172.
[26] A. Ocneanu 1994. Chirality for operator algebras . in Subfactors, ed. by H. Araki, et al., World Scientific, pp. 39 63.
[27] S. Popa 1994. Symmetric enveloping algebras, amenability and AFD properties for subfactors, Math. Res. Lett. 1, no. 4,
pp. 409 425.
[28] S. Popa 1994. Classification of amenable Subfactors of type II. Acta. Math. 172, pp. 163 - 225.
[29] S. Popa 1995. An axiomatization of the lattice of higher relative commutants. Invent. Math. 120, pp. 237 - 252.
[30] S. Popa 1996. Amenability in the theory of subfactors. Operator Algebras and Quantum Field Theory pp. 199 - 211.
18
COREY JONES
[31] S. Popa 1999. Some properties of the symmetric enveloping algebra of a subfactor, with applications to amenability and
property (T). Doc. Math. 4, pp. 665 - 744.
[32] S. Popa and S. Vaes 2014. Representation theory for subfactors, λ-lattices, and C∗-tensor categories. arXiv:1412.2732v2.
To appear in Communications in Mathematical Physics.
(Corey Jones) Vanderbilt University
Department of Mathematics
Nashville
USA
E-mail address: [email protected]
|
1005.4502 | 1 | 1005 | 2010-05-25T08:03:09 | Linear orthogonality preservers of Hilbert bundles | [
"math.OA",
"math.FA"
] | Due to the corresponding fact concerning Hilbert spaces, it is natural to ask if the linearity and the orthogonality structure of a Hilbert $C^*$-module determine its $C^*$-algebra-valued inner product. We verify this in the case when the $C^*$-algebra is commutative (or equivalently, we consider a Hilbert bundle over a locally compact Hausdorff space). More precisely, a $\mathbb{C}$-linear map $\theta$ (not assumed to be bounded) between two Hilbert $C^*$-modules is said to be "orthogonality preserving" if $\left<\theta(x),\theta(y)\right> =0$ whenever $\left<x,y\right> =0$. We prove that if $\theta$ is an orthogonality preserving map from a full Hilbert $C_0(\Omega)$-module $E$ into another Hilbert $C_0(\Omega)$-module $F$ that satisfies a weaker notion of $C_0(\Omega)$-linearity (known as "localness"), then $\theta$ is bounded and there exists $\phi\in C_b(\Omega)_+$ such that $$ \left<\theta(x),\theta(y)\right>\ =\ \phi\cdot\left<x,y\right>, \quad \forall x,y \in E. $$ On the other hand, if $F$ is a full Hilbert $C^*$-module over another commutative $C^*$-algebra $C_0(\Delta)$, we show that a "bi-orthogonality preserving" bijective map $\theta$ with some "local-type property" will be bounded and satisfy $$ \left<\theta(x),\theta(y)\right>\ =\ \phi\cdot\left<x,y\right>\circ\sigma, \quad \forall x,y \in E $$ where $\phi\in C_b(\Omega)_+$ and $\sigma: \Delta \rightarrow \Omega$ is a homeomorphism. | math.OA | math |
LINEAR ORTHOGONALITY PRESERVERS OF HILBERT BUNDLES
CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG
Abstract. Due to the corresponding fact concerning Hilbert spaces, it is natural to
ask if the linearity and the orthogonality structure of a Hilbert C ∗-module determine
its C ∗-algebra-valued inner product. We verify this in the case when the C ∗-algebra
is commutative (or equivalently, we consider a Hilbert bundle over a locally compact
Hausdorff space). More precisely, a C-linear map θ (not assumed to be bounded) between
two Hilbert C ∗-modules is said to be "orthogonality preserving" if hθ(x), θ(y)i = 0
whenever hx, yi = 0. We prove that if θ is an orthogonality preserving map from a full
Hilbert C0(Ω)-module E into another Hilbert C0(Ω)-module F that satisfies a weaker
notion of C0(Ω)-linearity (known as "localness"), then θ is bounded and there exists
φ ∈ Cb(Ω)+ such that
hθ(x), θ(y)i = φ · hx, yi ,
∀x, y ∈ E.
On the other hand, if F is a full Hilbert C ∗-module over another commutative C ∗-
algebra C0(∆), we show that a "bi-orthogonality preserving" bijective map θ with some
"local-type property" will be bounded and satisfy
where φ ∈ Cb(Ω)+ and σ : ∆ → Ω is a homeomorphism.
hθ(x), θ(y)i = φ · hx, yi ◦ σ,
∀x, y ∈ E
1. Introduction
It is a common knowledge that the inner product of a Hilbert space determines both the
norm and the orthogonality structures; and conversely, the norm structure determines
the inner product structure.
It might be a bit less well-known that the orthogonality
structure of a Hilbert space also determines its norm structure. Indeed, if θ is a linear
map between Hilbert spaces preserving orthogonality, then it is easy to see that θ is a
scalar multiple of an isometry (see [5, 6]).
We are interested in the corresponding relations for Hilbert C ∗-modules. Note that
in the case of a commutative C ∗-algebra C0(Ω), Hilbert C0(Ω)-modules are the same as
Hilbert bundles, or equivalently, continuous fields of Hilbert spaces over Ω. By modifying
the proof of [12, Theorem 6] (see also [13, 16, 9]), one can show that any surjective
isometry between two continuous fields of Hilbert spaces with non-zero fibres over each
2000 Mathematics Subject Classification. 46L08, 46M20, 46H40, 46E40.
Key words and phrases. automatic continuity, orthogonality preserving maps, module homomor-
phisms , local maps, Hilbert C*-modules, Hilbert bundles.
The authors are supported by Hong Kong RGC Research Grant (2160255), National Natural Science
Foundation of China (10771106), NCET-05-0219 and Taiwan NSC grant (NSC96-2115-M-110-004-MY3).
1
2
CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG
point is given by a homeomorphism and a field of unitaries. Thus, the norm structure
(and linearity) determines the unitary structure in this situation.
Our primary concern is the question of whether the orthogonality structure of a Hilbert
C ∗-module determines its unitary structure. More precisely, let A be a C ∗-algebra, and E
and F be two Hilbert A-modules. If θ : E → F is an A-module homomorphism , which
is not assumed to be bounded but preserves orthogonality (that is, hθ(x), θ(y)iA = 0
whenever hx, yiA = 0), we ask whether there is a central positive multiplier u in M(A)
such that
hθ(e), θ(f )iA = u he, f iA ,
∀e, f ∈ E.
When A = C, it reduces to the case of Hilbert spaces. Recently, D. Ilisevi´c and A.
Turnsek [10] gave a positive answer in the case when A is a standard C ∗-algebra (that
is, K(H) ⊆ A ⊆ L(H)).
In this article, we will give a positive answer when A is a commutative C ∗-algebra
(actually, we prove a slightly stronger result that replaces the A-linearity with the lo-
calness property; see Definition 2.1). On the other hand, we will also consider bijective
bi-orthogonality preserving maps between Hilbert C ∗-modules over different commuta-
tive C ∗-algebras. We show that if such a map also satisfies certain local-type property
(see Definition 3.10) but not assumed to be bounded, then it is given by a homeomor-
phism (between the base spaces) and a "continuous field of unitaries". We remark that
in this case of Hilbert C*-modules over different commutative C*-algebras, one cannot
defines "A-linearity" but have to consider localness property. This is one of the reasons
for considering local maps. We remark also that this case does not cover the case of
Hilbert C*-modules over the same commutative C*-algebra because we need to assume
that the map is both bijective and bi-orthogonality preserving.
Note that if Ω is a locally compact Hausdorff space and H is a Hilbert space, then
C0(Ω, H) is a Hilbert C0(Ω)-module. As far as we know, even in this case our results are
new, and the technique in the proofs are non-standard and non-trivial comparing with
those in the literatures [1, 4, 8, 11], concerning separating or zero-product preservers
(although some statements look similar). In a forthcoming paper of the authors, we will
study the case when the underlying C*-algebra is not commutative.
2. Terminologies and Notations
Recall that a (right) Hilbert C ∗-module E over a C ∗-algebra A is a right A-module
equipped with an A-valued inner product h·, ·i : E × E → A such that the following
conditions hold for all x, y ∈ E and all a ∈ A.
(1) hx, yai = hx, yia.
LINEAR ORTHOGONALITY PRESERVERS OF HILBERT BUNDLES
3
(2) hx, yi∗ = hy, xi.
(3) hx, xi ≥ 0, and hx, xi = 0 exactly when x = 0.
Moreover, E is a Banach space equipped with the norm kxk = khx, xik1/2. We also call
E a Hilbert A-module in this case. A complex linear map θ : E → F between two Hilbert
A-modules is called an A-module homomorphism if θ(xa) = θ(x)a for all a ∈ A and
x ∈ E. See, for example, [15] or [20], for a general introduction to the theory of Hilbert
C ∗-modules. In this paper, we are interested in the case when the underlying C ∗-algebra
A is abelian, that is, A = C0(Ω) consisting of all continuous complex-valued functions
defined on a locally compact Hausdorff space Ω vanishing at infinity.
Definition 2.1. Let A be a C ∗-algebra. Suppose that E and F are Hilbert A-modules.
A C-linear map θ : E → F is said to be local if θ(e)a = 0 whenever ea = 0 for any e ∈ E
and a ∈ A.
The idea of local linear maps is found in many researches in analysis. For example,
a theorem of Peetre [19] states that local linear maps of the space of smooth functions
defined on a manifold modeled on Rn are exactly linear differential operators (see [18]).
This is further extended to the case of vector-valued differentiable functions defined on a
finite dimensional manifold by Kantrowitz and Neumann [14] and Araujo [3], and in the
Banach C 1[0, 1]-module setting by Alaminos et. al. [2]. Note that every A-module homo-
morphism is local. Conversely, every bounded local map is an A-module homomorphism
([17, Proposition A.1]). See Remark 3.4 below for more information.
Notation 2.2. Throughout this article, Ω and ∆ are two locally compact Hausdorff spaces,
and Ω∞ is the one-point compactification of Ω. Moreover, E and F are respectively, a
(right) Hilbert C0(Ω)-module and a (right) Hilbert C0(∆)-module, while θ : E → F is
a C-linear map (not assumed to be bounded). We denote by BC0(Ω)(E, F ) the set of all
bounded C0(Ω)-module homomorphisms from E into F . For any ω ∈ Ω, we let NΩ(ω)
be the set of all compact neighborhoods of ω in Ω. If S ⊆ Ω, we denote by IntΩ(S) the
interior of S in Ω. Moreover, if U, V ⊆ Ω such that the closure of V is a compact subset
of IntΩ(U), we denote by UΩ(V, U) the collection of all λ ∈ C0(Ω) with 0 ≤ λ ≤ 1, λ ≡ 1
on V and λ vanishes outside U.
Note that any Hilbert C0(Ω)-module E can be regarded as a Hilbert C(Ω∞)-module,
and the results in [7] can be applied. In particular, E is the space of C0-sections (that is,
continuous sections that vanish at infinity) of an (F)-Hilbert bundle ΞE over Ω∞ (see [7,
p. 49]).
4
CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG
We define f (ω) := kf (ω)k for all f ∈ E and ω ∈ Ω. For any closed subset S ⊆ Ω∞
and ω ∈ Ω∞, we set
K E
S := {f ∈ E : f (ω) = 0, ∀ω ∈ S} and Iω := [V ∈NΩ∞ (ω)
K E
V
ω := K E
{ω}). Notice that K E
ω . Furthermore, K E
∞ = E and the fibre of ΞE
-module and
S is a Hilbert K C0(Ω)
S
(for simplicity, we also denote K E
at ω ∈ Ω∞ is given by ΞE
K E
S = E · K C0(Ω)
On the other hand, we denote
ω = E/K E
S
.
∆θ
:= (cid:8)ν ∈ ∆ : θ(E) * K F
ν (cid:9) = {ν ∈ ∆ : θ(e)(ν) 6= 0 for some e ∈ E} .
Then ∆θ is an open subset of ∆, and we put
Let Ω0 ⊆ Ω be an open set. As in [7, p. 10], we denote by ΞEΩ0 the restriction of ΞE
ΩE := (cid:8)ω ∈ Ω : ΞE
ω 6= (0)(cid:9) .
to Ω0 and by EΩ0 the set of C0-sections on ΞEΩ0. One can identify
C0(Ω0) = K C0(Ω)
Ω\Ω0
and EΩ0 = K E
Ω\Ω0.
3. Orthogonality preserving maps between Hilbert C0(Ω)-modules
Let us first recall the following two technical lemmas from [17, Lemmas 3.1 and 3.3,
and Theorem 3.7] (see also [17, Remark 3.4]), which summarize, unify, and generalize
techniques sporadically used in the literatures [4, 8, 11].
Lemma 3.1. If σ : ∆θ → Ω∞ is a map satisfying θ(cid:16)I E
σ is continuous.
σ(ν)(cid:17) ⊆ K F
ν (for any ν ∈ ∆θ), then
Lemma 3.2. Let σ : ∆ → Ω be a map (not assumed to be continuous) such that
ν for every ν ∈ ∆.
σ(ν)(cid:17) ⊆ K F
θ(cid:16)I E
(a) If Uθ :=(cid:8)ν ∈ ∆ : supkek≤1 kθ(e)(ν)k = ∞(cid:9), then σ(Uθ) is a finite set.
(b) If Nθ,σ := nν ∈ ∆ : θ(cid:16)K E
ν o, then Nθ,σ ⊆ Uθ and σ(Nθ,σ) consists of non-
σ(ν)(cid:17) * K F
isolated points in Ω.
(c) If σ is injective and sends isolated points in ∆ to isolated points in Ω, then Nθ,σ = ∅
and there exists a finite set T consisting of isolated points of ∆, a bounded linear map
θ0 : K E
for all ν ∈ T , such that
E = K E
T as well as linear maps θν : ΞE
σ(T ) → K F
σ(ν) → ΞF
ν
σ(T ) ⊕Lν∈T ΞE
σ(ν),
F = K F
T ⊕Mν∈T
ΞF
ν
and θ = θ0 ⊕Mν∈T
θν.
LINEAR ORTHOGONALITY PRESERVERS OF HILBERT BUNDLES
5
For any ν ∈ ∆ \ Nθ,σ, one can define θν : ΞE
σ(ν) → ΞF
ν by
(3.1)
or equivalently, θν(e(σ(ν))) = (θ(e))(ν) for all e ∈ E.
θν(cid:0)e + K E
σ(ν)(cid:1) = θ(e) + K F
ν ,
∀e ∈ E,
Lemma 3.3. Let σ and Uθ be the same as in Lemma 3.2. Suppose, in addition, that
σ is injective and θ is orthogonality preserving. Then there exists a bounded function
ψ : ∆ \ Uθ → R+ such that
(3.2)
hθ(e), θ(g)i(ν) = ψ(ν)2he, gi(σ(ν)),
∀e, g ∈ E, ∀ν ∈ ∆ \ Uθ.
Moreover, for each ν ∈ ∆θ, there is an isometry ιν : ΞE
σ(ν) → ΞF
ν such that
θ(e)(ν) = ψ(ν)ιν(e(σ(ν))),
∀e ∈ E, ∀ν ∈ ∆θ \ Uθ.
Proof. Fix any ν ∈ ∆θ \ Uθ. By Lemma 3.2(b), the map θν as in (3.1) is well-defined.
Suppose that η1 and η2 are orthogonal elements in ΞE
σ(ν) with η1 6= 0 (it is possible because
∆θ \ Nθ,σ ⊆ σ−1(ΩE)), and g1, g2 ∈ E with gi(σ(ν)) = ηi for i = 1, 2 . If V ∈ NΩ(σ(ν)) is
such that g1 is non-vanishing on V , then by replacing g2 with
(cid:18)g2 −
hg2, g1i
g12 g1(cid:19) λ
where λ ∈ UΩ({σ(ν)}, V ), we see that there are orthogonal elements e1, e2 ∈ E with
ei(σ(ν)) = ηi for i = 1, 2 . Hence, θν is non-zero (because ν ∈ ∆θ) and is an orthogonality
preserving C-linear map between Hilbert spaces. Consequently, there exist an isometry
ιν : ΞE
ν and a unique scalar ψ(ν) > 0 such that θν = ψ(ν)ιν. For any ν ∈ ∆ \ ∆θ,
we set ψ(ν) = 0. Then clearly (3.2) holds. Next, we show that ψ is a bounded function
on ∆ \ Uθ. Suppose that it is not the case. Then there exist distinct points νn ∈ ∆θ \ Uθ
such that ψ(νn) > n3. If en ∈ E with kenk = 1 and the modular function en(σ(νn)) =
σ(ν) → ΞF
phen, eni(σ(νn)) ≥ (n − 1)/n (note that νn ∈ σ−1(ΩE)), then because of (3.2),
θ(en)(νn) = ψ(νn)en(σ(νn)) > n2(n − 1).
As {σ(νn)} is a set of distinct points (note that σ is injective), by passing to a subsequence
if necessary, we can assume that there are Un ∈ NΩ(σ(νn)) such that Un ∩ Um = ∅ when
m 6= n. Now, pick any Vn ∈ NΩ(σ(νn)) with Vn ⊆ IntΩ(Un) and choose a function
∈ E. For any n ∈ N, as
k
λn ∈ UΩ(Vn, Un) for all n ∈ N. Define e := P∞
Un and en − enλ2
n2e − enλ2
n ∈ K E
k=1
n) ∈ K E
Vn, we have
ekλ2
k2
n = en(1 − λ2
kθ(enλ2
n)(νn)k
n2
=
kθ(en)(νn)k
n2
> n − 1
kθ(e)k ≥ kθ(e)(νn)k =
(by the relation between θ and σ) which is a contradiction.
(cid:3)
6
CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG
3.1. Hilbert bundles over the same base space.
Remark 3.4. For any e ∈ E, we denote
suppΩ e := {ω ∈ Ω : e(ω) 6= 0}.
It is not hard to check that the following statements are equivalent (which tells us that
local maps are the same as support shrinking maps [8]):
(i) θ is local (see Definition 2.1);
(ii) θ(cid:0)K E
V(cid:1) ⊆ K F
V for any non-empty open set V ;
(iii) suppΩ θ(e) ⊆ suppΩ e for every e ∈ E;
(iv) suppΩ θ(e)λ ⊆ suppΩ e for each e ∈ E and λ ∈ C0(Ω).
Theorem 3.5. Let Ω be a locally compact Hausdorff space, and let E and F be two
Hilbert C0(Ω)-modules. Suppose that θ : E → F is an orthogonality preserving local
C-linear map. The following assertions hold.
(a) θ ∈ BC0(Ω)(E, F ).
(b) There is a bounded non-negative function ϕ on Ω which is continuous on ΩE such
that
hθ(e), θ(g)i = ϕ · he, gi,
∀e, g ∈ E.
(c) There exist a strictly positive element ψ0 ∈ Cb(Ωθ)+ and J ∈ BC0(Ωθ)(EΩθ , FΩθ) such
that the fiber map Jω is an isometry for each ω ∈ Ωθ and
θ(e)(ω) = ψ0(ω)J(e)(ω),
∀e ∈ E, ∀ω ∈ Ωθ.
Proof. Note that the conclusions of Lemmas 3.2 and 3.3 hold for Ω = ∆ and σ = idΩ.
(a) By Remark 3.4 and Lemma 3.2(c), we see that θ is a C0(Ω)-module homomor-
phism. Furthermore, as θν (as in Lemma 3.2(c)) is an orthogonality preserving (and
hence bounded) linear map between Hilbert spaces for any ν ∈ T (where T is as in
Lemma 3.2(c), with σ = idΩ), we know from Lemma 3.2(c) that θ is bounded (note that
T is finite).
(b) By part (a), Uθ = ∅. Thus, Lemma 3.3 tells us that there exists a bounded non-
negative function ψ on Ω with hθ(e), θ(f )i = ψ2 · he, f i. Let ω ∈ ΩE and pick any e ∈ E
such that there is Uω ∈ NΩ(ω) with e(ν) 6= 0 for all ν ∈ Uω. Then ψ(ω) = θ(e)(ω)
for all
e(ω)
ω ∈ Uω. Hence ψ is continuous at ω, and ϕ(ω) = ψ(ω)2 is the required function.
(c) Note that Ωθ ⊆ ΩE because of part (a). Since ϕ(ω) > 0 (ω ∈ Ωθ), we know from part
(b) that ψ = ϕ1/2 gives a strictly positive element ψ0 in Cb(Ωθ)+. The equivalence in
[7, (2.2)] (consider E and F as Hilbert C(Ω∞)-bundles) tells us that the restriction of θ
LINEAR ORTHOGONALITY PRESERVERS OF HILBERT BUNDLES
7
induces a bounded Banach bundle map, again denoted by θ, from ΞEΩθ into ΞF Ωθ. For
each η ∈ ΞEΩθ, we define J(η) := ψ0(π(η))−1θ(η) (where π : ΞE → Ω is the canonical
projection). Then J : ΞEΩθ → ΞF Ωθ is a Banach bundle map (as η 7→ ψ0(π(η))−1 is
continuous) which is an isometry on each fibre (hence J is bounded) such that θ(η) =
ψ(π(η))J(η). This map J induces a map, again denoted by J, in BC0(Ωθ)(EΩθ , FΩθ) that
satisfies the requirement of part (c).
(cid:3)
It is natural to ask if one can find ϕ ∈ Cb(Ω) such that the conclusion of Theorem
3.5(b) holds. Unfortunately, the following example tells us that it is not the case in
general.
Example 3.6. Let Ω = R∞, the one-point compactification of the real line R. Consider
E = C0(R) = F as Hilbert C(Ω)-modules and θ(f )(t) = f (t) cos t for all f ∈ E and
t ∈ R. Then Ω \ ΩE = {∞} and ϕ(t) = cos t for any t ∈ R = ΩE. Thus, one cannot
extend ϕ to a continuous function on Ω.
Now, we can obtain the following commutative analogue of [10, 2.3]. This, together
with Corollary 3.9, asserts that the orthogonality structure of a Hilbert bundle determines
essentially its unitary structure, as we claimed in the Introduction. Note also that a large
portion of Lemma 3.2 were used to deal with the possibility of θ(K E
ν (such
situation does not exist for C0(Ω)-module homomorphism), and this corollary actually
has a much easier proof.
σ(ν)) * K F
Corollary 3.7. Let Ω be a locally compact Hausdorff space, and E and F be two Hilbert
C0(Ω)-modules. Suppose that θ : E → F is a C0(Ω)-module homomorphism which pre-
serves orthogonality. Then θ is bounded and there exists a bounded non-negative function
ϕ on Ω that is continuous on ΩE and satisfies hθ(e), θ(f )i = ϕ · he, f i for all e, f ∈ E.
Recall that a Hilbert C0(Ω)-module E is full if the C-linear span, hE, Ei, of
is dense in C0(Ω).
{he, f i : e, f ∈ E}
Remark 3.8. (a) E is full if and only if E * K E
ω for any ω ∈ Ω (or equivalently, ΩE = Ω).
In fact, if E ⊆ K E
ω , then f (ω) = 0 for any f ∈ hE, Ei and E is not full. Conversely, if E
is not full, then there exists ω ∈ Ω such that f (ω) = 0 for any f ∈ hE, Ei (because the
closure of hE, Ei is an ideal of C0(Ω)) and E ⊆ K E
ω .
(b) If E is full, then by part (a), the function ϕ in Theorem 3.5(b) (and Corollary 3.7) is
an element of Cb(Ω). However, there is no guarantee that this function is strictly positive.
8
CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG
(c) Suppose that F is full and θ is surjective orthogonality preserving local C-linear map.
If there exists ω ∈ Ω \ Ωθ, then F = θ(E) ⊆ K F
ω which contradicts the fullness of F (see
part (a)). Consequently, Ωθ = Ω. As θ ∈ BC0(Ω)(E, F ) (by Theorem 3.5(a)), we see that
Ω = Ωθ ⊆ ΩE and E is full (because of part (a)).
Corollary 3.9. Let Ω be a locally compact Hausdorff space, and let E and F be two
Hilbert C0(Ω)-modules. Suppose that F is full and θ : E → F is an orthogonality pre-
serving surjective local C-linear map. Then θ ∈ BC0(Ω)(E, F ). Moreover, there exist a
strictly positive element ψ ∈ Cb(Ω)+ and a unitary U ∈ BC0(Ω)(E, F ) such that θ = ψ · U.
Proof. Remark 3.8(c) tells us that Ωθ = Ω. By the surjectivity of θ, the bounded Banach
bundle map J in Theorem 3.5 is a unitary on each fibre. Therefore, the element U ∈
BC0(Ω)(E, F ) corresponding to J as given in [7, (2.2)] is a unitary.
(cid:3)
3.2. Hilbert bundles over different base spaces.
Definition 3.10. θ is said to be quasi-local
λ ∈ C0(∆), we have
if it is bijective and for any e ∈ E and
(3.3)
suppΩ θ−1(θ(e)λ) ⊆ suppΩ e.
Note that if ∆ = Ω, and if θ is both local and bijective (hence θ−1 is also local), then
θ is quasi-local by Remark 3.4.
Lemma 3.11. Suppose that θ is bijective and quasi-local and that both θ and θ−1 are or-
thogonality preserving. Then θ(e)θ(g) = 0 whenever e, g ∈ E with suppΩ e ∩ suppΩ g =
∅.
Proof. Suppose on the contrary that there exist e1, e2 ∈ E and ν ∈ ∆ such that suppΩ e1 ∩
suppΩ e2 = ∅ but kθ(e1)(ν)kkθ(e2)(ν)k 6= 0. As θ is orthogonality preserving, we may
assume that θ(e1)(ν) and θ(e2)(ν) are two orthogonal unit vectors in ΞF
ν . Let U, W ∈
N∆(ν) with W ⊆ Int∆(U) and kθ(ei)(µ)k > 1/2 for any µ ∈ U. Pick any λ ∈ U∆(W ; U).
Define hi ∈ F \ {0} for i = 1, 2 by
hi(µ) := (θ(ei)(µ)
0
λ(µ)
θ(ei)(µ) µ ∈ Int∆(U)
µ /∈ Int∆(U)
and set e′
i := θ−1(hi). The orthogonality of h1 and h2 (note that e1 and e2 are orthogonal),
together with that of h1 + h2 and h1 − h2 (as h1 = λ = h2), ensures the orthogonality
of e′
2 6= 0 which
contradicts the fact e′
(cid:3)
2, as well as that of e′
2. It follows that e′
2 = 0 (as θ is quasi-local).
1 and e′
1 + e′
2 and e′
1 − e′
1 = e′
1e′
LINEAR ORTHOGONALITY PRESERVERS OF HILBERT BUNDLES
9
Theorem 3.12. Let Ω and ∆ be locally compact Hausdorff spaces. Suppose that E is a
full Hilbert C0(Ω)-module and F is a full Hilbert C0(∆)-module. If θ : E → F is a bijective
C-linear map such that both θ and θ−1 are quasi-local and orthogonality preserving, then
θ is bounded and
(3.4)
θ(e)(ν) = ψ(ν)Jν(e(σ(ν))),
∀e ∈ E, ∀ν ∈ ∆,
where σ : ∆ → Ω is a homeomorphism, ψ is a strictly positive element in Cb(∆)+, and
Jν is a unitary operator from ΞE
ν such that for each fixed f ∈ E, the map
ν 7→ Jν(f (σ(ν))) is continuous.
σ(ν) onto ΞF
Proof. We consider E as a Hilbert C(Ω∞)-module. For each ν ∈ ∆, let
Sν := (cid:8)ω ∈ Ω∞ : θ(cid:0)K E
Ω∞\W(cid:1) * K F
We first show that Sν is a singleton set.
ω ∈ Ω∞, there is Wω ∈ NΩ∞(ω) such that θ(K E
with
ν for every W ∈ NΩ∞(ω)(cid:9) .
Indeed, assume that Sν = ∅. Then for any
ν . Consider ω1, ..., ωn ∈ Ω∞
Ω∞\Wω ) ⊆ K F
IntΩ∞(Wωk ) = Ω∞,
n
[k=1
Ω∞\Wω
k=1 to be a partition of unity subordinate to {IntΩ∞(Wωk )}n
and consider {ϕk}n
k=1. Then
for any e ∈ E, we have eϕk ∈ K E
ν . This shows that F = K F
ν
(as θ is surjective) which contradicts the fullness of F (see Remark 3.8(a)). Now, assume
that there are distinct elements ω1, ω2 ∈ Sν. Let V1 ∈ NΩ∞(ω1) and V2 ∈ NΩ∞(ω2) with
V1 ∩ V2 = ∅. By the definition of Sν, there exist e1, e2 ∈ E with suppΩ ei ⊆ Vi \ {∞} and
θ(ei)(ν) 6= 0 for i = 1, 2 which contradict Lemma 3.11. Thus, there is a unique element
σ(ν) ∈ Ω∞ with Sν = {σ(ν)}. Next, we claim that
and so θ(e) ∈ K F
k
(3.5)
θ(cid:0)I E
σ(ν)(cid:1) ⊆ I F
ν ,
∀ν ∈ ∆.
Consider any V ∈ NΩ∞(σ(ν)) and e ∈ K E
V . Pick any U ∈ NΩ∞(σ(ν)) with U ⊆ IntΩ∞(V ).
By the definition of σ, there exists g ∈ K E
Ω∞\U such that θ(g)(ν) 6= 0. Hence, there
is W ∈ N∆(ν) such that θ(g)(µ) 6= 0 for all µ ∈ W and Lemma 3.11 implies that
θ(e) ∈ K F
If there exists ν ∈ ∆ \ ∆θ, then for any f ∈ F , we have
f (ν) = 0 (because θ is surjective) which contradicts the fullness of F . Thus, ∆θ = ∆ and
σ : ∆ → Ω∞ is continuous (by Lemma 3.1). As θ−1 is also quasi-local and orthogonality
preserving, a similar argument as the above gives a continuous map τ : Ω → ∆∞ satisfying
ω for all ω ∈ Ω. Now, the argument of [17, Theorem 5.3] tells us that σ
W as claimed.
θ−1(cid:16)I F
τ (ω)(cid:17) ⊆ I E
is a homeomorphism from ∆ to Ω such that
θ(e · ϕ) = θ(e) · ϕ ◦ σ,
∀e ∈ E, ∀ϕ ∈ C0(Ω),
and by Lemma 3.2(c), there exists a finite set T consisting of isolated points of ∆ such that
θ restricts to a bounded map from K E
T . Since any ν ∈ T is an isolated point, θ
σ(T ) to K F
10
CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG
induces an orthogonality preserving (hence bounded) map θν from the Hilbert space ΞE
onto the Hilbert space ΞF
the fact that T is finite). By Lemma 3.3, there is a surjective isometry Jν : ΞE
such that
σ(ν)
ν . This shows that θ is bounded (because of Lemma 3.2(c) and
σ(ν) → ΞF
ν
θ(e)(ν) = ψ(ν)Jν(e(σ(ν))),
∀e ∈ E, ∀ν ∈ ∆.
Now the fullness of E implies that ψ(ν) > 0 (for every ν ∈ ∆) and clearly ν 7→ θ(e)(ν)
ψ(ν)
continuous.
is
(cid:3)
Note that the assumption of θ−1 being orthogonality preserving is necessary in Theorem
3.12 as can be seen from the following example.
Example 3.13. Let Ω be a (non-empty) locally compact Hausdorff space, and Ω2 be the
topological disjoint sum of two copies of Ω with j1, j2 : Ω → Ω2 being respectively the
embeddings into the first and the second copies of Ω in Ω2. Let H be a (non-zero) Hilbert
space, and let H2 be the Hilbert space direct sum of two copies of H. Then the map
θ : C0(Ω2, H) −→ C0(Ω, H2) defined by
θ(f )(ω) = (f (j1(ω)), f (j2(ω)))
is a bijective C-linear map preserving orthogonality satisfying Condition (3.3). However,
θ is not of the expected form. Note that θ−1 does not preserve orthogonality.
References
[1] Y. A. Abramovich, E. L. Arenson and A. K. Kitover, Banach C(K)-Modules and Operators Pre-
serving Disjointness, Pitman Research Notes in Mathematics 277, John Wiley & Sons, Inc., New
York, 1992.
[2] J. Alaminos, M. Bresar, M. Cerne, J. Extremera, A. R. Villena, 'Zero product preserving maps on
C 1[0, 1]', J. Math. Anal. Appl. 347 (2008), 472 -- 481.
[3] J. Araujo, 'Linear biseparating maps between spaces of vector-valued differentiable functions and
automatic continuity', Adv. Math. 187 (2004), no. 2, 488 -- 520.
[4] J. Araujo and K. Jarosz, 'Automatic continuity of biseparating maps', Studia Math. 155 (2003),
no. 3, 231 -- 239.
[5] A. Blanco and A. Turnsek, 'On maps that preserve orthogonality in normed spaces', Proc. Royal
Soc. Edingburgh 136A (2006), 709 -- 716.
[6] J. Chmieli´nski, 'Linear mappings approximately preserving orthogonality', J. Math. Anal. Appl.
304 (2005), 158 -- 169.
[7] M.J. Dupr´e and R.M. Gillette, Banach bundles, Banach Modules and Automorphisms of C ∗-algebras,
Pitman Research Notes in Mathematics 92, John Wiley & Sons, Inc., New York, 1983.
[8] H. L. Gau, J. S. Jeang and N. C. Wong, 'Biseparating linear maps between continuous vector-valued
function spaces', J. Australian Math. Soc. (Series A) 74 (2003), no. 1, 101 -- 111.
[9] M. H. Hsu and N. C. Wong, 'Isometric embeddings of Banach bundles', Taiwanese J. Math., to
appear.
[10] D. Ilisevi´c and A. Turnsek, 'Approximately orthogonality preserving mappings on C ∗-modules', J.
Math. Anal. Appl. 341 (2008), 298 -- 308.
[11] K. Jarosz, 'Automatic continuity of separating linear isomorphisms', Canad. Math. Bull. 33 (1990),
139 -- 144.
LINEAR ORTHOGONALITY PRESERVERS OF HILBERT BUNDLES
11
[12] J. S. Jeang and N. C. Wong, 'On the Banach-Stone problem', Studia Math. 155 (2003), no. 2,
95 -- 105.
[13] M. Jerison, 'The space of bounded maps into a Banach space', Ann. Math. 52 (1950), 309 -- 327.
[14] R. Kantrowitz and M.M. Neumann, 'Disjointness preserving and local operators on algebras of
differentiable functions', Glasg. Math. J. 43 (2001), 295 -- 309.
[15] E. C. Lance, Hilbert C ∗-modules. A toolkit for operator algebraists, London Mathematical Society
Lecture Note Series 210, Cambridge University Press, Cambridge, 1995.
[16] K. S. Lau, 'A representation theorem for isometries of C(X, E)', Pacific J. Math. 60 (1975), 229 -- 233.
[17] C. W. Leung, C. K. Ng and N. C. Wong, 'Automatic continuity and C0(Ω)-linearity of linear maps
between C0(Ω)-modules', preprint (2009).
[18] R. Narasimhan, Analysis on Real and Complex Manifolds, Advanced Studies in Pure Mathematics
1, North-Holland Publishing Co., Amsterdam, 1968.
[19] J. Peetre, 'R´ectification `a l'article "Une caract´erisation abstraite des op´erateurs diff´erentiels"', Math.
Scand. 8 (1960), 116 -- 120.
[20] I. Raeburn and D. P. Williams, Morita equivalence and continuous-trace C ∗-algebras, Mathematical
Surveys and Monographs 60, American Mathematical Society, Providence, RI, 1998.
(Chi-Wai Leung) Department of Mathematics, The Chinese University of Hong Kong,
Hong Kong.
E-mail address: [email protected]
(Chi-Keung Ng) Chern Institute of Mathematics and LPMC, Nankai University, Tianjin
300071, China.
E-mail address: [email protected]
(Ngai-Ching Wong) Department of Applied Mathematics, National Sun Yat-sen Univer-
sity, Kaohsiung, 80424, Taiwan, R.O.C.
E-mail address: [email protected]
|
1702.05611 | 1 | 1702 | 2017-02-18T13:29:01 | Stability of Fredholm property for regular operators on Hilbert $C^*$-modules | [
"math.OA"
] | We study the stability of Fredholm property for regular operators on Hilbert $C^*$-modules under some certain perturbations. We treat this problem when perturbing operators are (relatively) bounded or relatively compact. We also consider the perturbations of regular Fredholm operators in terms of the gap metric. In particular, we prove that the space of all regular Fredholm operators on a Hilbert $C^*$-module $E$ is open in the space of all regular operators on $E$ with respect to the gap metric. As an application, we construct some continuous paths of selfadjoint regular Fredholm operators with respect to the gap metric. | math.OA | math |
STABILITY OF FREDHOLM PROPERTY FOR REGULAR
OPERATORS ON HILBERT C ∗-MODULES
M. FOROUGH
Abstract. We study the stability of Fredholm property for regular operators
on Hilbert C ∗-modules under some certain perturbations. We treat this problem
when perturbing operators are (relatively) bounded or relatively compact. We
also consider the perturbations of regular Fredholm operators in terms of the gap
metric. In particular, we prove that the space of all regular Fredholm operators
on a Hilbert C ∗-module E is open in the space of all regular operators on E
with respect to the gap metric. As an application, we construct some continuous
paths of selfadjoint regular Fredholm operators with respect to the gap metric.
1. Introduction
In this paper, we investigate some forms of perturbations of regular Fredholm
operators on Hilbert C ∗-modules.
As in the Hilbert space case, one needs to consider unbounded adjointable operators
on Hilbert C ∗-modules, see [1], [17], [9], etc. The lack of the projection theorem
in Hilbert C ∗-modules caused the theory of unbounded operators on Hilbert C ∗-
modules to be more complicated than Hilbert space case. The regularity condition
was assumed for a closed densely defined operator on a Hilbert C ∗-module E with
densely defined adjoint, and this leads to have more reasonable theory like having a
continuous functional calculus.
Definition 1.1. Let t : D(t) ⊆ E → E be a closed densely defined operator with
densely defined adjoint on a Hilbert C ∗-module E. Then t is called regular if I + t∗t
has dense range.
In this paper, we denote by R(E) the set of all regular operators on a Hilbert
C ∗-module E. We say that a regular operator t is selfadjoint if t = t∗, and we denote
the set of all selfadjoint regular operators on E by SR(E).
Unbounded Fredholm operators on Hilbert C ∗-modules have been studied in several
papers, see [5], [15], among others. Joachim in [5] argued that it is natural to consider
unbounded Fredholm operators to describe some important invariants taking values
in K 0(X; A) and K 1(X; A), where X is a compact space and A is a unital C ∗-algebra.
In [15], the notion of noncommutative spectral flow for paths of selfadjoint regular
Fredholm operators is defined and studied.
2010 Mathematics Subject Classification. 46L08, 47A53, 47B25.
Key words and phrases. Perturbations, regular Fredholm operators, Cayley transform, the gap
metric, Hilbert C ∗-modules.
1
2
MARZIEH FOROUGH
Definition 1.2. (c.f [5]) Let t : D(t) ⊆ E → E be a regular operator. A bounded
adjointable operator G ∈ L(E) is called a pseudo left inverse of t if Gt is closable and
its closure Gt ∈ L(E) and Gt ≡ 1 modK(E).
Analogously, G ∈ L(E) is called a pseudo right inverse of t if tG is closable and its
closure tG ∈ L(E) and tG ≡ 1 modK(E).
A regular operator t on a Hilbert C ∗-module E is called Fredholm if t has a pseudo
right inverse as well as a pseudo left inverse.
The stability of Fredolm property for linear operators between Banach spaces or
Hilbert spaces are studied extensively, see [4], [8], and [13]. One basic problem here
is how to define a small perturbation for unbounded operators. One rather general
definition is based on the notion of relatively bounded perturbations, however, it
is still restrictive, see [7]. The most natural and general definition of smallness of a
perturbation is given by the gap metric in the space of closed linear operators between
Banach spaces. In [7], a systematic use of the gap metric was made. In this paper, we
aim to study the preservation of Fredholm property for regular operators on Hilbert
C ∗-modules. Following the Banach space case, we consider the small perturbations
by notions of (relatively) bounded perturbations, (relatively) compact perturbations
and the gap metric on the space of regular operators.
Theorem 1.3. (a) Let t : D(t) ⊆ E → E be a regular Fredholm operator on a Hilbert
C ∗-module E. Then there exists ε > 0 with the following property:
(i) D + t is a regular Fredholm operator for any selfadjoint bounded operator D on
E satisfying kDk ≤ ε.
(ii) If s : D(s) ⊆ E → E is a selfadjoint regular operator on E satisfying D(t) ⊆
D(s) and
ks(x)k ≤ ε(k(t + s)(x)k + kxk)
for all x ∈ D(t). Then t + s is a selfadjoint regular Fredholm operator.
(iii) If s is a regular operator on E with g(s, t) ≤ ε then s is Fredholm, where g is the
gap metric.
(b) Let t and s be selfadjoint regular operators on a countably generated Hilbert C ∗-
module E over unital C ∗-algebra A. Suppose that t is Fredholm and s is t-compact,
then s + t is a selfadjoint regular Fredholm operator.
Studying the perturbations of unbounded Fredholm operators on Banach spaces or
Hilbert spaces makes a heavy use of the fact that such operators have closed range,
finite dimensional kernel and cokernel. However, a regular Fredholm operator does
not have necessarily a closed range or finite dimensional kernel and cokernel. Hence
we can not employ the ideas and techniques applied in studying the perturbations
of unbounded closed densely defined Fredholm operators on Hilbert spaces in our
problem concerning the perturbations of regular Fredholm operators on Hilbert C ∗-
modules. To tackle our problem, we employ the Cayley transforms for selfadjoint
regular operators, and obtain a necessary and sufficient condition for a selfadjoint
regular operator to be Fredholm in terms of its Cayley transform.
STABILITY OF FREDHOLM PROPERTY
3
Theorem 1.4. Let t be a selfadjoint regular operator on a Hilbert C ∗-module E.
Then t is Fredholm if and only if bounded adjointable operator I + Ct is Fredholm.
Then we determine an explicit relation between the Cayley transform of a self-
adjoint regular operator and the Cayley transform of its perturbed operator. This
enables us to translate our problem in unbounded case to a problem in bounded ad-
jointable case and use the results available in the stability of Fredholm property for
bounded adjointable operators.
In the following proposition, we collect some results from [3] on the stability of
Fredholm property for bounded adjointable operators on Hilbert C ∗-modules.
Proposition 1.5. ([3]) Let A be a bounded adjointable Fredholm operator on a Hilbert
C ∗-module E, then the followings hold.
(i) There is ε > 0 such that any bounded adjointable operator D satisfying kA−Dk ≤ ε
is Fredholm.
(ii) If D is a bounded adjointable operator on E such that D −A is a compact operator
on E, then D is Fredholm.
(iii) If bounded adjointable operators U and V on E are invertible, then U AV is
Fredholm.
This paper is organized as follows. In section 2, we give a criterion for a selfad-
joint regular operator on a Hilbert C ∗-module to be Fredholm in terms of its Cayley
transform Ct. We also present a relation between the Cayley transform of a selfad-
joint regular operator and the Cayley transform of its perturbation by a selfadjoint
bounded operator or a selfadjoint regular operator.
In section 3, first we examine the stability of Fredholm property for regular operators
when perturbing operators are bounded adjointable operators. Then we turn to the
case of unbounded perturbing operators. We consider the relatively bounded pertur-
bations and relatively compact perturbations. Finally, we investigate the stability of
Fredholm property for regular operators in terms of the gap metric. In particular, we
show that the space of regular Fredholm operators is open in R(E) with respect to
the gap metric.
2. Cayley transforms of selfadjoint regular Fredholm operators
In this section we deal with the Cayley transform of a selfadjoint regular Fredholm
operator.
Let us start with recalling some notations regarding regular operators. Let t :
D(t) ⊆ E → E be a regular operator on a Hilbert C ∗-module E, then t∗ and tt∗ are
regular operators. Define Qt = (1 + t∗t)
t and Ft = tQt, then Qt and Ft
are bounded adjointable operators and D(t) = ran(Qt) (c.f [10]). The operator Ft
is called the bounded transform of t. We recall from [5] that a regular operator t is
Fredholm if and only if its bounded transform Ft is Fredholm.
2 , Rt = Q2
−1
Suppose that t is a selfadjoint regular operator then by Lemma 9.7 and Lemma 9.8
of [10], t + i and t − i are bijections. Define the Cayley transform of t, denoted by Ct,
by Ct := (t − i)(t + i)−1. Then Ct is a selfadjoint bounded operator. Furthermore,
by Theorem 10.4 of [10], the Cayley transform defines a bijection from SR(E) onto
{U ∈ L(E) : U is unitary; I − U has dense range}.
4
MARZIEH FOROUGH
In the next theorem, we give a criterion for a selfadjoint regular operator to be
Fredholm in terms of its Cayley transform. This characterization is essential to obtain
our main results.
Theorem 2.1. Let t be a selfadjoint regular operator on a Hilbert C ∗-module E.
Then t is Fredholm if and only if I + Ct is Fredholm.
Proof. Thanks to Lemma 2.2 of [5], it suffices to prove that Ft is Fredholm if and
only if I + Ct is Fredholm. To see this, first note that
(t − i)−1 = (t + i)(t2 + 1)−1 = tRt + iRt,
(t + i)−1 = (t − i)(t2 + 1)−1 = tRt − iRt.
These identities imply that
2tRt = (t − i)−1 + (t + i)−1.
Using above identity together with QttQt = tRt ((10.9) of [10]), we can get
I + Ct = I + (t − i)(t + i)−1 =
(t − i)((t − i)−1 + (t + i)−1) =
2(t − i)tRt = 2t2Rt − 2itRt
2tQttQt − 2itQtQt =
(2.1)
2Ft(Ft − iQt).
Observe that (10.3) and (10.9) of [10] shows that Ft − iQt is isometry. Moreover,
ran(Ft − iQt) = ran(t − i) since ran(Qt) = D(t). Now, Lemma 9.8 of [10] yields that
Ft − iQt is surjective, and so by Theorem 3.5 of [10], Ft − iQt is unitary. Therefore,
by Proposition (1.5), identity I + Ct = 2Ft(Ft − iQt) shows that Ft is Fredholm if
and only if I + Ct is Fredholm.
(cid:3)
We remark here that a similar criterion for a selfadjoint closed densely defined
operators on a Hilbert space to be Fredholm is obtained in [2]. However, the proof
given in [2] is different with our proof for the case of selfadjoint regular operators on
Hilbert C ∗-modules.
As an application of Theorem 2.1, we can obtain a sufficient condition for a selfad-
joint regular operator to be Fredholm. This condition is well-known among experts,
however we give other proof based on Theorem 2.1.
Corollary 2.2. Assume that t is a selfadjoint regular operator on a Hilbert C ∗-module
E such that (t + i)−1 is compact. Then t is Fredholm.
Proof. Observe that
I + Ct = 2I − (I − Ct) = 2I − (I − (t − i)(t + i)−1)
2I − ((t + i) − (t − i))(t + i)−1
= 2I − 2i(t + i)−1.
Thus I + Ct is Fredholm since it is a compact perturbation of Fredholm operator 2I.
Now, employ Theorem 2.1 to conclude that t is Fredholm.
(cid:3)
STABILITY OF FREDHOLM PROPERTY
5
Remark 2.3. Theorem 2.1 together with Theorem 10.4 of [10] imply that the Cayley
transform maps selfadjoint regular Fredholm operators on a Hilbert C ∗-module E
onto the set
{U ∈ L(E) : U is unitary ; I − U has a dense range; and I + U is F redholm}.
To study the bounded perturbations or relatively bounded perturbations of self-
adjoint regular Fredholm operators, we need to determine the relation between the
Cayley transform of the perturbed operator and the Cayley transform of the original
operator.
Proposition 2.4. Let t be a selfadjoint regular operator acting on a Hilbert C ∗-
module E and D is a selfadjoint bounded operator on E. Then we have
Ct+D − Ct = (I − Ct)D(D + t + i)−1.
Proof. The fact that operators t − i and t + i are bijections allows us to compute as
follows:
CD+t − Ct = (D + t − i)(D + t + i)−1 − (t − i)(t + i)−1 =
(t − i)((D + t + i)−1 − (t + i)−1) + D(D + t + i)−1 =
(t − i)(t + i)−1(−D)(t + D + i)−1 + D(D + t + i)−1 =
(I − Ct)D(D + t + i)−1.
(cid:3)
The following proposition can be deduced by a similar argument given in above
proposition.
Proposition 2.5. Let t be a selfadjoint regular operator on a Hilbert C ∗-module E
and s be a selfadjoint regular operator on E. Suppose that t + s is a selfadjoint regular
operator then
Ct+s − Ct = (I − Ct)s(s + t + i)−1
We recall here that there have been obtained some sufficient conditions which guar-
antee that sum of two selfadjoint regular operators remains regular and selfadjoint.
In [14], it is discussed that the perturbations of a regular selfadjoint operator by a
local bounded selfadjoint regular operator is again regular and selfadjoint. Moreover,
in Theorem 4.5 of [6], Kato-Rellich theorem on Hilbert spaces is extended to Hilbert
C ∗-modules, we use this theorem in the next section.
3. Perturbations of regular Fredholm operators
In this section, we study the stability of Fredholm property for regular operators
under certain perturbations.
6
MARZIEH FOROUGH
3.1. Bounded perturbations. We start this subsection with the following observa-
tion which can be useful in some of our proofs.
Remark 3.1. Let t be a selfadjoint regular operator acting on a Hilbert C ∗-module
E. Then we have
(t − i)−1 = (t + i)(t2 + 1)−1 = tRt + iRt = FtQt − iQ2
t .
Hence k(t − i)−1k ≤ 2.
Theorem 3.2. Let t be a selfadjoint regular Fredholm operator on a Hilbert C ∗-
module E. Then there exists ε > 0 such that D + t is a selfadjoint regular Fredholm
operator for any selfadjoint bounded operator D on E satisfying kDk ≤ ε.
Proof. By Theorem 2.1, I + Ct is Fredholm. Apply Proposition 1.5 to get ε > 0
with the following property: If S is a bounded adjointable operator on E satisfying
k(I + Ct) − Sk ≤ 4ε, then S is Fredholm. Assume that D is a selfadjoint bounded
operator on E with kDk ≤ ε. Then it follows by Proposition 2.4 and Remark 3.1 that
k(I + Ct+D) − (I + Ct)k ≤ kI − CtkkDkk(D + t + i)−1k ≤ 4ε.
Hence I + CD+t is Fredholm and so is D + t, as desired.
(cid:3)
Remark 3.3. We recall from [12] that one can correspond to any regular operator t
on Hilbert C ∗-module E a selfadjoint regular operator t = 0
t
t∗
0! on E ⊕ E.
Lemma 3.4. Let t be a regular Fredholm operator on a Hilbert C ∗-module E, then
0! is a selfadjoint regular Fredholm operator.
t
t = 0 t∗
Proof. Note that Ft = 0 Ft∗
Ft
0 ! = 0 F ∗
0 !, since Ft∗ = F ∗
Ft
t
t . Now, it is clear
that Ft is Fredholm if Ft is Fredholm.
(cid:3)
Remark 3.3 enables us to omit the assumption of being selfadjoint in Theorem 3.2.
Proposition 3.5. Let t be a regular operator on a Hilbert C ∗-module E. Then there
exists ε > 0 such that D + t is regular Fredholm operator on E for any bounded
adjointable operator D on E satisfying kDk ≤ ε.
Proof. If t is a regular operator on E and D is a bounded adjointable operator on
E then clearly we have (D + t)∗ = D∗ + t∗ and so ^(D + t) = D + t. Also, it is
easy to see that k Dk = kDk. Now apply Lemma 3.4 and Theorem 3.2 to conclude
proposition.
(cid:3)
Proposition 3.6. Suppose t is a regular Fredholm operator on a Hilbert C ∗-module
E and K is a compact operator on E. Then regular operator t + K is Fredholm.
Proof. By Remark 3.3, it suffices to prove that if t is a selfadjoint regular Fredholm
operator and K is a selfadjoint compact operator then selfadjoint regular operator
t+K is Fredholm. Indeed, compactness of operator K together with Proposition (2.4)
imply that I + Ct+K is a compact perturbation of Fredholm operator I + Ct. Hence
Proposition 1.5 implies that I + Ct+K is Fredholm and so t + K is Fredholm.
(cid:3)
STABILITY OF FREDHOLM PROPERTY
7
3.2. Unbounded perturbations.
Theorem 3.7. Let t : D(t) ⊆→ E be a selfadjoint regular Fredholm operator on
a Hilbert C ∗-module E. Then there exists ε > 0 with the following property: If
s : D(s) ⊆→ E is a selfadjoint regular operator satisfying D(t) ⊆ D(s) and
ks(x)k ≤ ε(k(t + s)(x)k + kxk)
for all x ∈ D(t). Then t + s is a selfadjoint regular Fredholm operator.
Proof. Note that Theorem 2.1 and Proposition 1.5 yield that there is a positive num-
ber ε′ such that any bounded adjointable operator D on E satisfying k(I +Ct)−Dk ≤
ε′ is also Fredholm. Put ε = min{ 1
10 }. Suppose that s is a selfadjoint regular op-
erator on E with D(t) ⊆ D(s) and satisfying
4 , ε′
(3.1)
ks(x)k ≤ ε(k(t + s)(x)k + kxk)
for all x ∈ D(t). We are going to show that t + s is a selfadjoint regular Fredholm
operator. For this, first note that it follows from ε ≤ 1
4 and 3.1 that
(3.2)
ks(x)k ≤
1
3
(kxk + kt(x)k)
for all x ∈ D(t). Hence Theorem 4.4 of [6] implies that t + s is a selfadjoint regular
operator on E. Observe that (t + s + i)−1(x) ∈ D(t) for all x ∈ E, so the inequality
(4.1) yields that
ks(s + t + i)−1(x)k ≤ ε(k(s + t)(s + t + i)−1xk + k(s + t + i)−1(x)k)
≤ εk(s + t + i)(s + t + i)−1xk + 2εk(s + t + i)−1(x)k)
for all x ∈ E. Employ Proposition 2.5 to conclude
≤ 5εkxk
(3.3) k(I + Ct+s)(x) − (I + Ct)(x)k ≤ kI − Ctkks(s + t + i)−1(x)k ≤ 10εkxk ≤ ε′kxk
for all x ∈ E. Thus I + Ct+s is Fredholm, and so t + s is Fredholm, as desired.
(cid:3)
For a closed operator on a Hilbert space, the domain equipped with the graph
scaler product is itself Hilbert space, see [7].
In [6], the analogous construction is
discussed for a (semi)regular operator. Let t be a regular operator and x, y ∈ D(t),
put
< x, y >t=< x, y > + < t(x), t(y) >
Then D(t) with this inner product becomes a Hilbert C ∗-module, denoted by E(t). It
follows from Proposition 2.4 of [6] that the inclusion map ι : D(t) → E is a bounded
adjointable operator.
Definition 3.8. Let t : D(t) ⊆→ E and s : D(s) ⊆→ E be regular operators on
a Hilbert C ∗-module E such that D(t) ⊆ D(s). We call s is relatively t-compact or
simply, a t-compact if s : E(t) → E is a compact operator.
Theorem 3.9. Let t and s be selfadjoint regular operators on a countably generated
Hilbert C ∗-module E over unital C ∗-algebra A. Suppose that t is Fredholm and s is a
t-compact operator, then s + t is a selfadjoint regular Fredholm operator.
8
MARZIEH FOROUGH
Proof. First note that Lemma A.4 of [15] yields that s + t is a selfadjoint regular op-
erator. It is straightforward to check that (t + i)−1 : E → E(t) is bounded adjointable
operator. Hence s(t + i)−1 is a compact operator on E since we assume that s is t-
compact. So it follows from the equation (t+s+i)−1 = (t+i)−1 −(t+s+i)−1s(t+i)−1
that s(t + s + i)−1 = s(t + i)−1 − s(t + s + i)−1s(t + i)−1 is a compact operator on
E. Hence, by Proposition 2.5, one can see that I + Ct+s is a compact perturba-
tion of Fredholm operator I + Ct. Therefore the desired result can be concluded by
Proposition 1.5 and Theorem 2.1.
(cid:3)
Corollary 3.10. Let t be a selfadjoint regular Fredholm operator on a Hilbert C ∗-
module E. Suppose that A is a selfadjoint bounded t-compact operator on E, then
A + t is a selfadjoint regular Fredholm operator.
3.3. Perturbations in terms of the gap metric. The most natural and general
definition of smallness of the perturbations of Fredholm operators on Hilbert spaces
or Banach spaces is given in terms of the gap metric. Indeed, Theorem IV4.30 of [7]
states that the set of Fredholm operators in the space of closed linear operators on a
Banach space is open with respect to the gap metric. In this subsection, we generalize
this result to regular Fredholm operators.
Let us first briefly recall the definition of the gap metric in the space of regular
operators on a Hilbert C ∗-module and some of its basic properties.
Let t be a regular operator on a Hilbert C ∗-module E and G(t) = {(x, t(x)) :
x ∈ D(t)} be its graph. Then Theorem 9.3 of [10] implies that G(t) is orthogonally
complemented in EL E, we denote by PG(t) the orthogonal projection onto G(t).
Assume that t and s are regular operators on E, define the gap metric between t and
s by g(t, s) = kPG(t) − PG(s)k.
The matrix representation of the orthogonal projection onto the graph of a regular
operator (see [10]) shows that gap metric is equivalent to the metric d given by
(3.4)
d(t, s) = sup{kRt − Rsk, kRt∗ − Rs∗ k, ktRt − sRsk}.
Suppose t and s are selfadjoint regular operators on a Hilbert C ∗-module E, then
Corollary 2.8 of [12] implies that
1
4
kCt − Csk ≤ d(t, s) ≤
1
2
kCt − Csk.
We recall from Proposition 3.7 of [12] that d(t, s) = d(t, s) for all regular operators t
and s on E, where t and s are defined in Remark (3.3).
Theorem 3.11. Let t be a selfadjoint regular Fredholm operator on a Hilbert C ∗-
module E. Then there exists r > 0 with the following property: If s is a selfadjoint
regular operator on E with g(s, t) ≤ r then s is Fredholm, where g is the gap metric.
Proof. Note that I +Ct is Fredholm by Theorem ??, hence we can employ Proposition
1.5 to get ε > 0 such that any bounded adjointable operator S on E satisfying
k(I + Ct) − Sk ≤ 4ε is Fredholm. Suppose that s is a selfadjoint regular operator on
E with d(s, t) ≤ ε, where d as given in equerrygap. Then Corollary 2.8 of [12] yields
that k(I + Ct) − (I + Cs)k ≤ 4ε. This shows that I + Cs is Fredholm and so s is
Fredholm. Finally, observe that the gap metric is equivalent to d, and this finishes
the proof.
(cid:3)
STABILITY OF FREDHOLM PROPERTY
9
Thanks to Remark 3.3, we can deduce the following theorem from Theorem 3.11.
Theorem 3.12. Let t be a regular Fredholm operator on a Hilbert C ∗-module E.
Then there exists r > 0 such that any regular operator s on E satisfying d(s, t) ≤ r is
also Fredholm.
Proof. Suppose that t is a regular Fredholm operator. By Lemma 3.4, t is a selfadjoint
Fredholm operator. Employ Theorem 3.11 to obtain r > 0 such that any selfadjoint
regular operator satisfying d(s, t) ≤ r is Fredholm. We show that this positive scalar
r satisfies in the claim of theorem. Let s be regular Fredholm operator satisfying
d(s, t) ≤ r. Then by Proposition 3.7 of [12], d(t, s) = d(t, s), hence s is Fredholm, and
so is s, as desired.
(cid:3)
Corollary 3.13. The space of all bounded adjointable Fredholm operators on a Hilbert
C ∗-module E is open in R(E) with respect to the gap metric.
Proof. Recall that Theorem 3.12 states that the space of regular Fredholm operators
is open in R(E) with respect to the gap metric. Moreover, by Theorem 3.4 of [12]
the space of all bounded adjointable operators on Hilbert C ∗-module E is open in
R(E) with respect to the gap metric. Therefore, the space of bounded adjointable
Fredholm operators is open in R(E) with respect to the gap metric, as desired. (cid:3)
We conclude this subsection by constructing some continuous paths of selfadjoint
regular Fredholm operators with respect to the gap metric.
Theorem 3.14. Let t be a selfadjoint regular Fredholm operator on a Hilbert C ∗-
module E and let λ 7−→ Aλ be a continuous path of selfadjoint bounded operators
on E with respect to the norm topology. Suppose that each Aλ is t-compact, then
λ 7−→ t + Aλ is a continuous path of selfadjoint regular Fredholm operators with
respect to the gap metric.
Proof. Proposition 3.10 implies that λ 7−→ t + Aλ is a path of selfadjoint regular
Fredholm operators on Hilbert C ∗-module E. By Proposition 2.4 we get that
kCt+Aλ − Ct+Aµ k ≤ kI − Ct+Aµ kkAλ − Aµkk(I + t + Aλ)k
≤ 4kAλ − Aµk.
So Corollary 2.8 of [12] shows that λ 7−→ t + Aλ is a continuous path of selfadjoint
regular Fredholm operators with respect to the gap metric.
(cid:3)
Theorem 3.15. Let t be a selfadjoint regular Fredholm operator on a Hilbert C ∗-
module E and let ε be the scaler as in Theorem 3.2. Suppose that λ 7−→ Aλ is
a continuous path of selfadjoint bounded operators on E with respect to the norm
topology with kAλk ≤ ε for each λ. Then λ 7−→ t + Aλ is a continuous path of
selfadjoint regular Fredholm operators with respect to the gap metric.
Proof. It follows from Theorem 3.2 that λ 7−→ Aλ + t is a path of selfadjoint regular
Fredholm operators. By a similar argument given in the proof of Theorem 3.14, one
can conclude the continuity of the path with respect to the gap metric
(cid:3)
10
MARZIEH FOROUGH
Acknowledgements
Part of this work was done during author's stay in International School for Ad-
vanced Studies (SISSA), Trieste, Italy, 2012. This work was supported by a grant
from IPM and by SISSA post-graduate fellowship.
References
1. S. Baaj, P. Julg. Th?eorie bivariante de Kasparov et operateurs non bornes dans les C ∗-modules
hilbertien, C. R. Acad. Sci. Paris Ser. I Math 269 (1983) 875-878 .
2. B. Booss-Bavenbek, M. Lesch, J. Phillips, Unbounded Fredholm operators and spectral flow,
Canad. J. Math. 57 (2005) 225–250.
3. R. Exel, A Ferdholm approach to Morita equivalence, K-Theory. 57 (1993) 285–308.
4. I. C. Gokhberg, M. G. Krein, Fundamental theorems on deficiency numbers, root numbers, and
indices of Linear operators, (Translated) Amer. math. Soc. Transl-Ser2, vol. 13.
5. M. Joachim, Unbounded Fredhom operators and K-theory, High-dimensional manifold topology,
177–199, World Sci. Publishing, river Edg, NJ, 2003.
6. J. Kaad, M. Lesch, A local global principle for regular operators in Hilbert C ∗-modules, J. Func.
Anal. 262 (2012) 4540–4569.
7. T. Kato, Perturbation of Linear operators (second edition), Springer-Verlage, Berlin, 2003 .
8. T. Kato, Perturbation theory for nulity, deficiency and other quantities of linear operators, J.
d'analyse Math. 6 (1958) 273–322.
9. D. Kucerovsky, The KK-product of unbounded modules, K-theory. 11 (1997) 17-34.
10. E. C. Lance, Hilbert C ∗-modules. A toolkit for operator algebrists, LMS Lecture Note Series
210, Cambridge University Press, 1995 .
11. J. Mingo, K-theory and multipliers of stable C ∗-algebras, Trans. Amer. Math. Soc. 299 (1987)
397-411 .
12. K. Sharifi, The gap between unbounded regular operators, J. Operator Theory. 65 (2011) 241–
253.
13. B. Sz.-Nagy, Perturbations des transformation lin`eares ferm`ees, Acta Sci. Math., Szeged. 14
(1951) 125–137.
14. K. van den Dungen, locally bounded perturbations and (odd) unbounded KK-theory, Preprint,
arXiv: 1608.02506.
15. C. Wahl, On noncommutative spectral flow, New York J. Math. 15 (2009) 319–351.
16. N. E. Wegge-Olsen, K-theory and C ∗-algebras. A friendly approach, Oxford University Press,
Oxford, England, 1993.
17. S. L. Woronowicz, Unbounded elements affiliated with C ∗-algebras and noncompact qantum
groups, Comm. Math. Phys. 136 (1991) 399-432.
School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O.
Box 19395-5746, Tehran, Iran
E-mail address: [email protected], [email protected]
|
1006.4263 | 2 | 1006 | 2010-07-14T21:34:34 | Rieffel Deformation of Homogeneous Spaces | [
"math.OA",
"math-ph",
"math-ph",
"math.QA"
] | Let H be a closed subgroup of a locally compact group G and let X=G/H be the quotient space of left cosets. Let C*X be the corresponding G-C*-algebra of continuous functions on X, vanishing at infinity. Suppose that L is a closed abelian subgroup of H and let f be a 2-cocycle on the dual group of L. Let G(f) be the Rieffel deformation of G. Using these data we may construct G(f)-C*-algebra C*X(f) - the Rieffel deformation of C*X. On the other hand we may perform the Rieffel deformation of the subgroup H obtaining the closed quantum subgroup H(f) of G(f) which in turn, by the results of Vaes, leads to the G(f)-C*-algebra G(f)/H(f). In this paper we show that G(f)/H(f) and C*X(f) are isomorphic G(f)-C*-algebras. We also consider the case where L is a subgroup of G but not of H, for which we cannot construct the subgroup H(f). Then C*X(f) cannot be identified with a quantum quotient. What may be shown is that it is a G(f)-simple object in the category of G(f)-C*-algebras. | math.OA | math | RIEFFEL DEFORMATION OF HOMOGENEOUS SPACES
P. KASPRZAK
Abstract. Let G1 ⊂ G be a closed subgroup of a locally compact group G and let X = G / G1
be the quotient space of left cosets. Let X = (C0(X), ∆X) be the corresponding G-C∗-algebra
where G = (C0(G), ∆). Suppose that Γ is a closed abelian subgroup of G1 and let Ψ be a
2-cocycle on the dual group Γ. Let GΨ be the Rieffel deformation of G. Using the results of
paper [8] we may construct GΨ-C∗-algebra XΨ - the Rieffel deformation of X. On the other
hand we may perform the Rieffel deformation of the subgroup G1 obtaining the closed quantum
1 ⊂ GΨ, which in turn, by the results of [16], leads to the GΨ-C∗-algebra GΨ/GΨ
subgroup GΨ
1 .
∼= XΨ. We also consider the case where Γ ⊂ G is not a
In this paper we show that GΨ/GΨ
subgroup of G1, for which we cannot construct the subgroup GΨ
1 . Then XΨ cannot be identified
with a quantum quotient. What may be shown is that it is a GΨ-simple object in the category
of GΨ-C∗-algebras.
1
0
1
0
2
l
u
J
4
1
]
.
A
O
h
t
a
m
[
2
v
3
6
2
4
.
6
0
0
1
:
v
i
X
r
a
Contents
1.
Introduction
2. Preliminaries
2.1. G-C∗-category
2.2. Quantum homogeneous spaces
3. Group algebra twist
4. Rieffel deformation of group coactions
5.
6. Rieffel deformation of homogeneous spaces - the quotient case
7. Rieffel deformation of G-simple C∗-algebras
References
Induction of the regular corepresentation
1
2
2
3
4
5
6
11
13
13
1. Introduction
The theory of locally compact quantum groups (LCQG) has already reached its maturity.
Almost ten years have passed since the appearance of the seminal paper of J. Kustermans and
S. Vaes [11], where the axiomatic theory was formulated. A locally compact quantum group is a
C∗-bialgebra (A, ∆) equipped with a left and a right Haar weight φ and ψ. Imposing some natural
conditions on ∆ (weak cancellation) and on the Haar weights φ and ψ (KMS-type conditions) the
authors were able to develop the theory, proving among other things the existence of the coinverse
which admits the polar decomposition κ = R ◦ τi/2 and showing that the theory is self dual. The
assumption of the existence of the Haar weights, which is a theorem for the classical groups and
for the compact quantum groups, may be perceived as a drawback. It seems that a Haar weights
free axiomatization is out of reach. There exists second formulation of the LCQG theory, due to
T. Masuda, Y. Nakagami and S.L. Woronowicz [12], in which the authors include the existence of
the coinverse κ in the axioms. But to develop the theory they still need to assume the existence
of a left Haar weight. In fact, it can be shown that both theories are equivalent.
One way of constructing examples of LCQGs is to start with a classical group G and search
for its deformations Gq. In general, the link between G and Gq is not rigorously described. There
2000 Mathematics Subject Classification. Primary 46L89, Secondary 20N99.
Supported by the Marie Curie Research Training Network Non-Commutative Geometry MRTN-CT-2006-031962.
Supported by Geometry and Symmetry of Quantum Spaces, PIRSES-GA-2008-230836.
1
2
P. KASPRZAK
is at least one mathematical procedure - the Rieffel deformation - where this correspondence is
clear. For the original approach of Rieffel we refer to [15]. In this paper we shall use our recent
approach to the Rieffel deformation which describes it in terms of crossed product construction
described in Section 3.
1 of the twisted group GΨ
(see [9]). The Rieffel deformation of G will be denoted by GΨ and its dual will be denoted by bGΨ.
The deformation procedure of a locally compact group in terms of the transition from bG to bGΨ is
Let G be a locally compact group and D a G - C∗-algebra. Using the results of [8] one may apply
the Rieffel deformation to D, obtaining the deformed GΨ - C∗-algebra DΨ. The concise account
of the deformation procedure of G - C∗-algebras is the subject of Section 4. In Section 2 beside
giving some preliminaries on G-C∗-algebras we also discuss the notion of a quantum homogeneous
space and the C∗-algebraic quotient space G/G1, a construction due to Vaes [16]. His construction
may be performed for any closed quantum subgroup G1 of a regular LCQG G - for regularity we
refer to [1]. The relation of G/G1 with the induction procedure of the regular corepresentation is
explained. The reader's awareness of this relation is crucial in the understanding of the proof that
XΨ ∼= GΨ/GΨ
1 . In Section 5 we perform the induction procedure of the regular corepresentation
W Ψ
1 and compare the resulting objects with their untwisted counterparts.
This enables us to prove that XΨ ∼= GΨ/GΨ
1 which is the subject of Section 6. Finally, in Section
7 we comment on the case where XΨ is not of the quotient type. In connection with it we show
that the Rieffel deformation of a G-simple C∗-algebra a GΨ-simple.
The particular case of the Rieffel deformation of a homogeneous space has been discussed in
[17]. In this paper J. Varilly treats the situation where there is given a sequence Γ ⊂ G1 ⊂ G of
closed subgroups with Γ being abelian and G compact. He shows that it is possible to perform
a covariant deformation of X = G / G1 obtaining a quantum homogeneous GΨ-space XΨ. In this
specific situation the difficulties that one encounters in general do not manifest themselves.
Throughout the paper we will freely use the language of C∗-algebras and the theory of locally
compact quantum groups. For the notion of multipliers and morphism of C∗-algebras we refer the
reader to [18]. For the theory of locally compact quantum groups we refer to [11] and [12].
I would like to express my gratitude to W. Szyma´nski for many stimulating discussions, which
greatly influenced the final form of this paper.
2. Preliminaries
2.1. G-C∗-category. Let us fix a notation related with a locally compact quantum group G. The
C∗-algebra and the comultiplication of G will be denoted by C0(G) and ∆G respectively. The von
Neumann algebra associated with G and the Hilbert space obtained by the GNS-representation
shall denote the reduced version of the dual of G. The modular conjugations related with the left
related with the left Haar weight will be denoted respectively by L∞(G) and L2(G). By bG we
Haar weight on G (on bG) will be denoted by J (by J).
The main subject of this paper is related with the quantum group actions on quantum spaces.
The following definition may be traced back to [13]. To formulate it we adopt the following
notation: a closed linear span of a subspace W ⊂ V of a Banach space V will be denoted by [W].
Definition 2.1. Let G be a LCQG. A G-C∗-algebra is a pair D = (D, ∆D) consisting of a C∗-
algebra D and a coaction ∆D ∈ Mor(D, C0(G) ⊗ D):
(ι ⊗ ∆D)∆D = (∆G ⊗ ι)∆D
such that [∆D(D)(C0(G) ⊗ 1)] = C0(G) ⊗ D. The C∗-algebra D will also be denoted by C0(D)
and the coaction ∆D will be denoted by ∆D.
Remark 2.2. Suppose that G corresponds to an ordinary locally compact group G. There is a
1-1 correspondence between G-C∗-algebras and G-C∗-algebras, i.e. C∗-algebras equipped with a
continuous action of G. In order to describe it we introduce the characters χg : C0(G) → C that
are associated with the points of G: χg(f ) = f (g), for any g ∈ G and f ∈ C0(G). Let D be a
continuous coaction of G. We define the corresponding continuous action α : G → Aut(C0(D)) by
the formula: αg(a) = (χg−1 ⊗ ι)∆D(a) for any g ∈ G and a ∈ C0(D).
3
The category of G-C∗-algebras was consider either explicitly or implicitly in many different
contexts - see for instance [6] where it appears explicitly in the context of the Landstad duality.
In order to specify the category of G-C∗-algebras we adopt the following notion of G-morphisms.
Definition 2.3. Let G be a LCQG and suppose that B and D are G-C∗-algebras. We say that a
morphism π : C0(B) → C0(D) is a G-morphism from B to D if ∆D ◦ π = (ι ⊗ π) ◦ ∆B. The set of
G-morphism from B to D will be denoted by MorG(B, D).
2.2. Quantum homogeneous spaces. Let G be a locally compact quantum group and X be
a G - C∗-algebra. The concept of a quantum homogeneous G-space exists so far only in the
case of G being compact and C0(X) being unital - see Definition 1.8 of [14]. In this case X is a
quantum homogeneous space if ∆X is ergodic. Let G be a compact quantum group corresponding
to a compact group G. It may be checked that there is a 1-1 correspondence between quantum
homogeneous G-spaces with the underlying commutative C∗-algebra C0(X) and the classical ho-
mogeneous G-spaces. Unfortunately the ergodicity assumption in the non-compact case cannot
serve as a replacement for homogeneity. The next approximation to the homogeneity is the notion
of G-simplicity:
Definition 2.4. Let D be G-C∗-algebra. We say that D is G-simple if for any G-C∗-algebra B
and any G-morphism π ∈ MorG(D, B) we have ker π = {0}.
In order to motivate the introduction the G-simplicity let us prove that in the case of the
compact quantum groups the homogeneity of D implies its G-simplicity.
Lemma 2.5. Let G = (A, ∆) be a compact quantum group with a faithful Haar measure h. Let
X be a quantum homogeneous space with ∆X being injective. Then X is G-simple.
Proof. Using the ergodicity of ∆X we may introduce the state on ρ : C0(X) → C such that
(h ⊗ ι)∆X(a) = ρ(a) for any a ∈ C0(X).
(1)
The faithfulness of h and the injectivity of ∆X imply the faithfulness of ρ. Let π ∈ MorG(X, B)
and suppose that a ∈ ker π. Note that (h ⊗ ι)∆X(a∗a) ∈ ker π:
π((h ⊗ ι)∆X(a∗a)) = (h ⊗ ι)∆X(π(a∗a)) = 0.
On the other hand, using (1) we may see that π((h ⊗ ι)∆X(a∗a)) = ρ(a∗a)1 which together with
the above computation shows that ρ(a∗a) = 0. The faithfulness of ρ implies that a = 0 hence
ker π = {0}.
(cid:3)
Let G be a locally compact quantum group and let G be the corresponding LCQG. It may
be checked that there is a 1-1 correspondence between G-simple C∗-algebras with the underlying
C∗-algebra being commutative and minimal G - spaces - the G-spaces with all orbit being dense
(see p.49, [3]).
As was already mentioned, a definition of a quantum homogeneous space appropriate for locally
compact quantum groups is not yet known. What has been generalized so far is the notion of the
quantum quotient space, due to Vaes [16]. But since the work of Podle´s [13] on quantum spheres
we know that a generic quantum homogeneous space is not of the quotient type. The Rieffel of
the homogeneous spaces deformation provides a new class of examples of non-compact quantum
spaces which in our opinion should also be considered as homogeneous and which generically are
not of the quotient type.
Let us move on to the discussion of the quantum quotient spaces. Let G be a regular LCQG
(for the notion of regularity we refer to [1]). Let G1 be the closed quantum subgroup in the sense
of Definition 2.5 [16]. As a part of the structure we have the injective normal ∗-homomorphism
π : L∞(bG1) → L∞(bG). Its commutant counterpart is denoted by π′ : L∞(bG1)′ → L∞(bG)′. The
definition of the measurable quotient space L∞(G/G1) goes as follows:
L∞(G/G1) = {a ∈ L∞(G) : aπ′(x) = π′(x)a, for any x ∈ L∞(bG1)′}.
A remarkable Theorem 6.1 [16]) provides the existence and the uniqueness of the C∗-algebraic
version C0(G/G1) of L∞(G/G1). For the needs of this paper we shall formulate it as a definition:
4
P. KASPRZAK
Definition 2.6. Let G be a LCQG which is regular and let G1 be its closed quantum subgroup.
The C∗-algebraic quotient G/G1 is the unique G-C∗-algebra defined by the following conditions:
• C0(G/G1) ⊂ L∞(G/G1) is a strongly dense C∗-subalgebra;
• ∆G/G1 is given by the restriction of ∆G to C0(G/G1);
• ∆G(L∞(G/G1)) ⊂ M(K(L2(G)) ⊗ C0(G/G1)) and the ∗-homomorphism
∆G/G1 : L∞(G/G1) → L(L2(G) ⊗ C0(G/G1)) is strict.
For the notion of strictness we refer to Definition 3.1 of [16]. Note that the symbol ∆G/G1
denotes the coaction on C0(G/G1) as well as the strict map defined on L∞(G/G1).
Remark 2.7. The difficulty of the proof of Theorem 6.1, [16] lies in the existence part. To
explain the idea of the proof (which will be also important to understand this paper) let us
consider the C∗-algebra C0(bG1) treated as a Hilbert C∗-module over itself. Performing the in-
duction procedure of the regular corepresentation W1 ∈ M(C0(G1) ⊗ C0(bG1)) we get the induced
Hilbert module Ind(C0(bG1)) equipped with the additional structure: the induced corepresenta-
tion Ind(W1) and the coaction γ : Ind(C0(bG1)) → M(Ind(C0(bG1)) ⊗ C0(bG)) of the dual quan-
tum group bG. The coaction γ is obtained by the induction procedure applied to the coaction
β : C0(bG1) → M(C0(bG1) ⊗ C0(bG)) given by β(a) = (ι ⊗ π)∆bG1
(a) for any a ∈ C0(bG1). The
pact operators K(Ind(C0(bG1)) on Ind(C0(bG1)) is canonically isomorphic with the C∗-algebra of a
crossed product. The coaction γ is identified with the dual coaction on the crossed product while
Ind(W1) is identified with the corepresentation implementing the coaction on the γ-invariants.
The C∗-algebra C0(G/G1) is defined as the Landstad-Vaes C∗-algebra of γ-invariants, by which
we mean the C∗-algebra satisfying the conditions of Theorem 6.7, [16].
additional structure consisting of γ and Ind(W1) enables one to prove that the C∗-algebra of com-
3. Group algebra twist
Let G be a locally compact group and let δ : G → R+ be the modular function. Suppose that
Γ is a closed abelian subgroup of G1, which in turn is a closed subgroup of G. For reasons which
will become clear later we shall assume that the modular function δ : G → R+ when restricted to
Γ is identically 1: δ(γ) = 1 for any γ ∈ Γ.
In order to perform the twisting procedure one has to fix a 2-cocycle Ψ on Γ, i.e. a continuous
function Ψ : Γ × Γ → T1 satisfying:
(i) Ψ(e, γ) = Ψ(γ, e) = 1 for all γ ∈ Γ;
(ii) Ψ(γ1, γ2 + γ3)Ψ(γ2, γ3) = Ψ(γ1 + γ2, γ3)Ψ(γ1, γ2) for all γ1, γ2, γ3 ∈ Γ.
For the theory of 2-cocycles we refer to [10].
In order to simplify the notation and make further computations less cumbersome we shall
assume that Ψ is a skew symmetric bicharacter on Γ:
Ψ(γ1 + γ2, γ3) =Ψ(γ1, γ3)Ψ(γ2, γ3),
Ψ(γ1, γ2) =Ψ(γ2, γ1).
(2)
(3)
We will only prove our results for such 2-cocycles although we have reasons to believe that they
hold for arbitrary 2-cocycles.
The role of the bicharacter Ψ ∈ M(C0(Γ) ⊗ C0(Γ)) will vary in the course of this paper. The
simplest variation is connected with the identification C0(Γ) ∼= C∗(Γ), which enables us to treat
Ψ as an element of M(C∗(Γ) ⊗ C∗(Γ)). Furthermore, the morphism ι ∈ Mor(C∗(Γ), C∗
l (G1))
which corresponds to the representation Γ ∋ γ 7→ Lγ ∈ M(C∗
l (G1)) enables us to treat Ψ as an
element of M(C∗
l (G1)). Finally, in some cases Ψ will be treated as an operator acting
on L2(G1) ⊗ L2(G1) or L2(G) ⊗ L2(G).
l (G1) ⊗ C∗
Let us describe the twisting procedure of (C∗
l (G), ∆). In what follows we shall use the notation
of Section 2.1, denoting the locally compact quantum group related to the pair (C∗
l (G), ∆) by bG.
In particular, C0(bG) = C∗
Ψ ∈ M(C0(bG) ⊗ C0(bG)): ∆bGΨ (a) = Ψ∗∆bG(a)Ψ. Using Corollary 5.3 of [4] we see that there exists
l (G) and ∆bG = ∆. We may twist the comultiplication on ∆bG by means of
5
is given by ∆bGΨ . (For more general results concerning the twist of a quantum group by a 2-cocycle
a locally compact quantum group bGΨ such that C0(bGΨ) = C0(bG), for which the comultiplication
we refer to [2].) Furthermore, with our modular assumption bGΨ is of the Kac type - the coinverse
κ bGΨ is an involutive anti-automorphism. The dual locally compact quantum group of bGΨ will be
denoted by GΨ. Its description in terms of the Rieffel deformation was given in [9]. It is also of the
Kac type - κ2
GΨ = id - which by Example 3.4 [1] implies that GΨ is regular. The only reason for
the assumption δΓ = 1 is to ensure the regularity of GΨ which is important in the construction
of the quotient of a locally compact quantum group by its closed quantum subgroup.
In what follows we shall give the formula for the multiplicative operators W Ψ ∈ L∞(GΨ) ⊗
L∞(bGΨ) and V Ψ ∈ L∞(GΨ)′ ⊗ L∞(bGΨ) where we adopted the notation of Section 2.1 of [16].
Using Theorem 1 of [5] we may see that the operator W Ψ acts on B(L2(G) ⊗ L2(G)) and it is of
the form W Ψ = ΨW Ψ′ where
Ψ′ = ( J ⊗ J)Ψ( J ⊗ J).
(4)
The assumption of the co-stability introduced in Section 2.7 of [5] is satisfied trivially in the
discussed case. The required group homomorphism t 7→ γt is the trivial homomorphism: γt = e ∈ Γ
(we replaced the γt introduced in [5] by γt which is more appropriate in our context).
(5)
we get
To give the formula for the multiplicative unitary V Ψ we only have to invoke the fact that it
cop.
cop is the flip of the comultiplication ∆bGΨ . By the skew-symmetry of Ψ
is the fundamental multiplicative unitary corresponding to the co-opposite quantum group bGΨ
The comultiplication of bGΨ
The twist construction that we applied to G can be also applied to G1 leading to bGΨ
1 and the
dual GΨ
1 . Note that we do not impose the modular condition on the embedding Γ ⊂ G1. Again
using Theorem 1 of [5] one may see that W Ψ
1 are related to W1 and V1 in the way analogous
to the case of G described above (see Eq. (5)). The aforementioned assumption of co-stability is
also satisfied but the homomorphism t 7→ γt may be non-trivial. In order to see that, we shall use
the modular function δ1 : G1 → R+. For any t ∈ R we may define the character γt ∈ Γ given by
hγt, γi = δit
1 (γ). The homomorphism t 7→ γt is the one that ensures the co-stability. It may be
checked that the quantum group GΨ
1 is a closed quantum subgroup of GΨ in the sense of Definition
V Ψ = Ψ∗ V Ψ′∗.
1 and V Ψ
2.5 of [16]. In particular, there exists the embedding π : L∞(bGΨ
acting between commutants will be denoted by π′ : L∞(bGΨ
1 )′ → L∞(bGΨ)′.
4. Rieffel deformation of group coactions
1 ) → L∞(bGΨ). Its counterpart
Suppose that G is a locally compact group containing an abelian closed subgroup Γ and let Ψ
be a skew-symmetric bicharacter on Γ (see Section 3). Let X = (C0(X), ∆X) be a G-C∗-algebra
In paper [9] we defined a GΨ-C∗-algebra XΨ - the Rieffel deformation of
(see Definition 2.1).
X. The deformation procedure may be viewed as a two steps procedure: first extend ∆X to the
morphism of the appropriate crossed products and then twist the extension by a unitary obtained
from Ψ. Let us be more precise.
Let β : G → Aut(C0(X)) be the continuous action corresponding to the coaction ∆X (see
Remark 2.2). Let α : Γ → Aut(C0(X)) be the restriction of β to the subgroup Γ. The C∗-
algebra C0(XΨ) is defined as the Landstad algebra of the Γ-product (Γ ⋉α C0(X), λ, ρ ¯Ψ) where
ρ ¯Ψ : Γ → Aut(Γ ⋉α C0(X)) is the ¯Ψ-deformed dual action (see Section 3, [9]). Let us note that
∆X ◦ αγ = (ργ,e ⊗ ι) ◦ ∆X, where ρ : Γ2 → Aut(C0(G)) is the action defined by ργ1,γ2(f )(g) =
f (γ−1
1 gγ2) for any γ1, γ2 ∈ Γ and g ∈ G. The universal property of the crossed product Γ ⋉α C0(X)
enables us to define the extension ∆Γ
X ∈ Mor(Γ ⋉α C0(X), Γ2 ⋉ρ C0(G) ⊗ Γ ⋉α C0(X)) of ∆X to
the level of crossed product.
In order to describe the aforementioned twisting step we define Υ ∈ M(Γ2⋉ρC0(G)⊗Γ⋉αC0(X))
as follows. Let Φ ∈ Mor(C∗(Γ2), Γ2 ⋉ρ C0(G) ⊗ Γ ⋉α C0(X)) be the morphism that corresponds to
the representation Γ ∋ (γ1, γ2) 7→ λe,γ1 ⊗ λγ2 ∈ M(Γ2 ⋉ρ C0(G) ⊗ Γ ⋉α C0(X)). Applying it to Ψ we
get Υ = Φ(Ψ). We may define ∆XΨ by the formula ∆XΨ (b) = Υ∆Γ
X(b)Υ∗ for any b ∈ Γ ⋉α C0(X).
6
P. KASPRZAK
By Theorems 4.3 and 4.4 of [8] ∆XΨ restricts to a morphism ∆XΨ : C0(XΨ) → C0(GΨ) ⊗ C0(XΨ)
which is a continuous coaction of GΨ on C0(XΨ). This defines GΨ-C∗-algebra XΨ.
5. Induction of the regular corepresentation
1 )) of GΨ
1 ), defined by
βΨ(a) = (ι ⊗ π)∆bGΨ
1 ) ⊗ C0(bGΨ
1 on C0(bGΨ
1 ), where C0(bGΨ
1 )) together with the induced corepresentation Ind(W Ψ
In this section we shall apply the induction procedure to the regular corepresentation W Ψ
1 ) ∈ L(C0(GΨ) ⊗ Ind(C0(bGΨ
1 ) → M(C0(bGΨ
1 ∈
1 ) is treated as a C∗-Hilbert module over
M(C0(GΨ
itself. For the induction procedure in the framework of LCQGs we refer to [16] (we shall also
1 )-Hilbert module
1 ))).
1 ) ⊗
adopt the notation of this paper). As a result, we obtain the induced C0(bGΨ
Ind(C0(bGΨ
We shall also apply the induction procedure to the right coaction βΨ : C0(bGΨ
C0(bGΨ)) of bGΨ on C0(bGΨ
where π ∈ Mor(C0(bGΨ
1 ), C0(bGΨ)) is the standard embedding. The result is the coaction
Ind(βΨ) : Ind(C0(bGΨ
1 )) → M(Ind(C0(bGΨ
Finally, the induced objects Ind(C0(bGΨ
untwisted counterparts Ind(C0(bG1)), Ind(W1), and Ind(β).
Let us first recall that the von Neumann algebra of the twisted quantum group bGΨ remains
unchanged L∞(bGΨ) = L∞(bG). This implies that the imprimitivity bimodule I Ψ, which is defined
stays undeformed: I Ψ = I. Nevertheless, the coaction αI Ψ : I Ψ → I Ψ ⊗ L∞(bGΨ): αI Ψ(v) =
I Ψ = {v ∈ B(L2(G1), L2(G)) vm = π′(m)v, for any m ∈ L∞(bGΨ
gets twisted. The relation with its untwisted counterpart is established by
1 ) and Ind(βΨ) will be compared with their
1 )) ⊗ C0(bGΨ)).
V Ψ(v ⊗ 1)(ι ⊗ π) V Ψ∗
the following formula:
1
1 )), Ind(W Ψ
1 )′}
(a),
1
(6)
by
αI Ψ (v) = Ψ∗αI(v)Ψ.
(7)
Let us analyze the strict ∗-homomorphism πbGΨ
[16] it was denoted by πl), given by:
1
: L∞(bGΨ
1 ) → L(L2(G)) ⊗ C0(bGΨ
1 )) (in Lemma 4.5,
πbGΨ
1
(m) = (π ⊗ ι)(W Ψ
1 (m ⊗ 1)W Ψ∗
1
)
1 ). It may be noted that the map πbGΨ
coincides with (π ⊗ id)∆bGΨ
1,cop
when
1
1 )) is expressed by
It is equipped with the strict ∗-homomorphism πΨ
l
homomorphism πΨ
r : L∞(bGΨ) → L(F Ψ), which are defined as follows:
h ⊗ a1) = (mi) ⊗
π bGΨ
1
πΨ
l (m)(i ⊗
π bGΨ
1
h ⊗ a1
: L∞(bGΨ) → L(F Ψ) and the strict ∗ - anti-
πΨ
r (m)(i ⊗
π bGΨ
1
h ⊗ a1) = i ⊗
π bGΨ
1
Jm∗ Jh ⊗ a1
for any m ∈ L∞(bGΨ), i ∈ I Ψ, h ∈ L2(G) and a1 ∈ C0(bGΨ
prove:
1 ). In order to compare F Ψ and F we
for any m ∈ L∞(bGΨ
appropriately interpreted. Its relation with πbG1
the twisting formula:
for any m ∈ L∞(bGΨ
1 ) = L∞(bG1). Using πbGΨ
1
F Ψ = I Ψ ⊗
π bGΨ
1
πbGΨ
1
(m) = ΨπbG1
(m)Ψ∗
: L∞(bG1) → L(L2(G)) ⊗ C0(bGΨ
we may introduce the C0(bGΨ
L2(G) ⊗ C0(bGΨ
1 ).
1 ) - module
Proposition 5.1. There exists a unitary transformation U ∈ L(F , F Ψ) such that
U (i ⊗
π bG1
h ⊗ a1) = i ⊗
π bGΨ
1
Ψ(h ⊗ a1)
7
for any i ∈ I, h ∈ L2(G) and a1 ∈ C0(bGΨ
Proof. For the existence of U it is enough to note that:
1 ). U intertwines πl with πΨ
l and πr with πΨ
r .
U (ia ⊗
π bG1
h ⊗ a1) = ia ⊗
π bGΨ
1
Ψ(h ⊗ a1)
= i ⊗
π bGΨ
1
ΨπbG1
(a)(h ⊗ a1)
= U (i ⊗
π bG1
πbG1
(a)(h ⊗ a1)).
The fact that U is unitary and that it possesses the required intertwining properties can be verified
by a straightforward computation.
(cid:3)
αL2(G)⊗C0( bGΨ
1 )(ξ) = V Ψ
1 ) → M(L2(G) ⊗
1 ) : L2(G) ⊗ C0(bGΨ
of the coaction αI Ψ given by (7) and the coaction αL2(G)⊗C0( bGΨ
Let us now introduce the coaction αF Ψ : F Ψ → M(F Ψ ⊗ C0(bGΨ)). It is defined as a product
1 ) ⊗ C0(bGΨ)) given by:
C0(bGΨ
for any ξ ∈ L2(G) ⊗ C0(bGΨ
The coaction αF Ψ corresponds to the corepresentation Y Ψ ∈ L(F Ψ ⊗ C0(bGΨ)) of bGΨ on F Ψ:
In order to compare Y Ψ with its undeformed counterpart Y ∈ L(F ⊗ C0(bG)) we compute:
1 ). The formula defining αF Ψ goes as follows:
1 )(ξ).
αF Ψ(i ⊗
π bGΨ
1
Y Ψ(f ⊗ a) = αF Ψ (f )(1 ⊗ a).
(U ∗ ⊗ 1)Y Ψ(U ⊗ 1)(i ⊗
π bG1
ξ) = αI Ψ (i) ⊗
π bGΨ
1
αL2(G)⊗C0( bGΨ
13(ξ ⊗ 1),
ξ ⊗ a)
⊗ι
= (U ∗ ⊗ 1)Y Ψ(i ⊗
π bGΨ
1
Ψξ ⊗ a)
= (U ∗ ⊗ 1)(αF Ψ (i ⊗
π bGΨ
1
Ψξ))(1 ⊗ a)
= (U ∗ ⊗ 1)(Ψ∗αI (i)Ψ ⊗
π bGΨ
1
⊗ι
Ψ∗
13
V13Ψ′∗
13Ψ12(ξ ⊗ 1))(1 ⊗ a)
⊗ι
⊗ι
= (Ψ∗αI (i)Ψ ⊗
π bG1
= (Ψ∗αI (i)Ψ ⊗
π bG1
= (Ψ∗αI (i) ⊗
π bG1
= (Ψ∗αI (i) ⊗
π bG1
⊗ι
⊗ι
Ψ∗
12Ψ∗
13
V13Ψ′∗
13Ψ12(ξ ⊗ 1))(1 ⊗ a)
Ψ∗
12Ψ∗
13Ψ12Ψ∗
23
V13Ψ′∗
13(ξ ⊗ 1))(1 ⊗ a)
Ψ13Ψ23Ψ∗
12Ψ∗
13Ψ12Ψ∗
23
V13Ψ′∗
13(ξ ⊗ 1))(1 ⊗ a)
V13Ψ′∗
13(ξ ⊗ 1))(1 ⊗ a)
= (πl ⊗ id)(Ψ∗)Y (πr ⊗ id)(Ψ′∗)(i ⊗
π bG1
In the fifth equality we used the easy to verify formula V13Ψ12 V ∗
equality we used: (πbG1
23 and in the sixth
⊗ ι)Ψ = Ψ13Ψ23. For the formula for Ψ′ we refer to (4). This computation
13 = Ψ12Ψ∗
ξ ⊗ a)
shows that the relation between the corepresentations Y ∈ L(F ⊗ C0(bG)) and Y Ψ ∈ L(F Ψ ⊗
C0(bGΨ)) is given by
(U ∗ ⊗ id)Y Ψ(U ⊗ id) = (πl ⊗ id)(Ψ∗)Y (πr ⊗ id)(Ψ′∗),
8
P. KASPRZAK
where we treat the anti-homomorphism πr : L∞(bG) → L(F ) as the homomorphism of the com-
mutant L∞(bG)′ (note that Ψ′ ∈ L∞(bG)′ ⊗ L∞(bG)).
We are now ready to compare the induced module Ind(C0(bGΨ
1 ) with their untwisted counterparts Ind(C0(bG1)) and Ind(W1). Let us first recall
tation Ind(W Ψ
the construction of the untwisted objects. By the results of [16] there exists a (unique) strict
∗-homomorphism Θ : B(L2(G)) → L(F ) such that
1 )) and the induced corepresen-
(Θ ⊗ ι) V = Y,
Θ( Jx∗ J) = πr(x),
such that
where by assumption v : L2(G) → F is a continuous map. It follows from the above definition
Ind(C0(bG1)) = {v : L2(G) → F : vx = Θ(x)v for all x ∈ B(L2(G))},
for any x ∈ L∞(bG). The induced C∗-module is defined as
that we may define the unitary transformation Φ : F → L2(G) ⊗ Ind(C0(bG1)) of C0(bG1)-modules,
for any h ⊗ v ∈ L2(G) ⊗ Ind(C0(bG1)). The induced corepresentation Ind(W1) ∈ L(C0(G) ⊗
Ind(C0(bG1))) of G is defined by the equation
in the sense of the identification F ∼= L2(G) ⊗ Ind(C0(bG1)). The isomorphism Φ defined above
intertwines the the strict ∗-homomorphism πl with the AdInd(W1) (see Proposition 3.7 and Section
4 of [16]):
where the leg numbering notation on the right hand side of this equation is to be understood
(ι ⊗ πl)W = W12 Ind(W1)13,
Φ∗(h ⊗ v) = v(h)
(8)
Φπl(x)Φ∗ = Ind(W1)(x ⊗ 1) Ind(W1)∗.
(9)
Let us write the corepresentation equation (∆G ⊗ ι) Ind(W1) = Ind(W1)13 Ind(W1)23 in terms of
W and Ind(W1):
Ind(W1)23W12 Ind(W1)∗
23 = W12 Ind(W1)13
(10)
Applying the character χg : C0(G) → C, g ∈ G to the first leg of the above equation we get the
following formula
Ind(W1)(Lg ⊗ 1) Ind(W1)∗ = Lg ⊗ Ind(W1)(g),
(11)
where we identified the induced corepresentation Ind(W1) ∈ L(C0(G) ⊗ Ind(C0(bG1))) with the
corresponding representation of the group G on K(Ind(C0(bG1))):
B(L2(G)) → L(F Ψ) we may define the induced C∗-module Ind(C0(bGΨ
ΦΨ : F Ψ → L2(G) ⊗ Ind(C0(bGΨ
The twisted objects are defined similarly. Starting with the strict ∗-homomorphism ΘΨ :
1 )) and the isomorphism
1 ) by the equation
12 Ind(W Ψ
Ind(W1)(g) = (χg ⊗ ι)(Ind(W1)).
1 )13
1 )). Defining Ind(W Ψ
l )W Ψ = W Ψ
1 ) by the formula:
(ι ⊗ πΨ
we may relate πΨ
l , ΦΨ and Ind(W Ψ
ΦΨπΨ
l (x)ΦΨ∗ = AdInd(W Ψ
1 )(x) = Ind(W Ψ
1 )(x ⊗ 1) Ind(W Ψ
1 )∗.
In order to compare Θ and ΘΨ let us define Ψ ∈ L(Ind(C0(bG1)) ⊗ L2(G)) by:
Ind(W1)12Ψ13 Ind(W1)∗
12 = Ψ13 Ψ23.
The existence of such Ψ follows from equation (11). Let us also introduce
where Σ is the flip operation on the tensor product.
Ψ′ = Φ∗Σ ΨΣΦ ∈ L(F )
(12)
(13)
Theorem 5.2. Let Θ : B(L2(G)) → L(F ) and ΘΨ : B(L2(G)) → L(F Ψ) be the strict ∗-
homomorphisms introduced above and let U ∈ L(F , F Ψ) be the unitary transformation introduced
in Proposition 5.1. Let Ψ′ ∈ L(F ) be the element introduced in Eq. (13). Then
ΘΨ(x) = U (Ad Ψ′ ◦Θ) (x)U ∗
(14)
9
1 )) given by:
for any x ∈ B(L2(G)).
1 )) are isomorphic, via the
In particular, Ind(C0(bG1)) and Ind(C0(bGΨ
isomorphism K : Ind(C0(bG1)) → Ind(C0(bGΨ
Ind(C0(bG1)) ∋ v 7→ K(v) = U Ψ′v ∈ Ind(C0(bGΨ
Proof. Since the algebras L∞(bGΨ)′ and L∞(GΨ)′ generate B(L2(G)), it is enough to check Eq.
(14) on them separately. It is easy to see that for any x ∈ L∞(bGΨ)′
which proves equality (14) on L∞(bGΨ)′. To prove it on L∞(GΨ)′ it is enough to see that (Ad Ψ′ ◦Θ⊗
Ind(C0(bG1)) by means of the isomorphism Φ (see (8)). Using (8) and (12) it is easy to show that
id) V Ψ = (U ∗ ⊗ id)Y Ψ(U ⊗ id). In the following computation we shall identify F with L2(G) ⊗
under the aforementioned identification we have:
(Ad Ψ′ ◦Θ)( Jx∗ J) = U ∗πr(x)U,
1 )).
(15)
(16)
(πl ⊗ id)Ψ = Ψ13 Ψ23.
(17)
We compute
(Ad Ψ′ ◦Θ ⊗ id)( V Ψ
13 ) = Ψ′
12Ψ∗
13
= Ψ′
12Ψ∗
13
= Ψ′
Ψ′∗
12
12
= (πl ⊗ id)(Ψ∗) V13(πr ⊗ id)(Ψ′∗) = Y Ψ,
Ψ′∗
V13Ψ′∗
12
13
V13 Ψ′∗
12Ψ′∗
13
V13Ψ′∗
Ψ∗
23Ψ∗
13
13
where in the third equality we used an easy to check formula:
V13 Ψ′∗
12
13 = Ψ′∗
V ∗
12
Ψ∗
23
and in the fourth equality we used (17).
(18)
(cid:3)
Let us now move on to the comparison of the induced corepresentations Ind(W Ψ
1 ) and Ind(W1).
In order to do that we introduce two elements Ψ, Ψ′ ∈ L(L2(G) ⊗ Ind(C0(bG1))) such that
(ι ⊗ πl)(Ψ) = Ψ12 Ψ13,
(ι ⊗ πl)(Ψ) = Ψ′
13.
12
Ψ′
(19)
(20)
Their existence follows from the bicharacter equation for Ψ and Eq. (11). Let us note that Ψ and
the element Ψ satisfying (17) are related by the formula
Ψ = Σ Ψ∗Σ.
(21)
We compute:
(ι ⊗ πΨ
l )(W Ψ) = Ψ23(ι ⊗ πl)(ΨW Ψ′) Ψ∗
23
Ψ′
13
Ψ∗
23
= Ψ23Ψ12 Ψ13W12 Ind(W1)13Ψ′
12
12 Ind(W1)13 Ψ′
= Ψ12W12 Ψ13Ψ′
13
Ψ13 Ind(W1)13 Ψ′
= Ψ12W12Ψ′
13
12
= W Ψ
12
Ψ13 Ind(W1)13 Ψ′
13,
where in the third equality we used
W12 Ψ13W ∗
12 Ind(W1)13 = Ψ′
12
12 = Ψ13 Ψ23,
Ψ∗
23.
Ind(W1)∗
13Ψ′
10
P. KASPRZAK
The first of these equalities follows from the bicharacter equation for Ψ whereas the second one
follows from the equality below:
Ind(W1)∗(Rg ⊗ 1) Ind(W1) = Rg ⊗ Ind(W1)(g).
This shows that with the identification Ind(C0(bG1)) ∼= Ind(C0(bGΨ
where the product on the right hand side is taken in the C∗-algebra L(L2(G) ⊗ Ind(C0(bG1)).
Finally, we shall compare the induced right coactions
1 )) of Theorem 5.2 we have
1 ) = Ψ Ind(W1) Ψ′,
Ind(W Ψ
Equation (6.2) of [16] defines the induced coaction Ind(β) as
Ind(β)Ψ : Ind(C0(bGΨ
1 )) ⊗ C0(bGΨ)),
Ind(β) : Ind(C0(bG1)) → M(Ind(C0(bG1)) ⊗ C0(bG)).
1 )) → M(Ind(C0(bGΨ
x) = W13(Φ ⊗ ι) (i ⊗ 1) ⊗
W ∗
π bG1
⊗ι
13(ι ⊗ β)(x)! .
(ι ⊗ Ind(β))Φ(i ⊗
π bG1
Similarly, Ind(βΨ) is defined by:
(ι ⊗ Ind(βΨ))ΦΨ(i ⊗
π bGΨ
1
x) = W Ψ
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
π bGΨ
1
⊗ι
13 (ι ⊗ βΨ)(x)! .
W Ψ∗
Combining (8), (15) and (21) one can easily check that
(ι ⊗ K) ΨΦ = ΦΨU.
(22)
Using this formula we compute:
(ι ⊗ Ind(βΨ))(ι ⊗ K) ΨΦ(i ⊗
π bG1
x)
= (ι ⊗ Ind(βΨ))ΦΨU (i ⊗
π bG1
x)
π bGΨ
1
⊗ι
π bGΨ
1
⊗ι
π bGΨ
1
⊗ι
W Ψ∗
W Ψ∗
13W ∗
= W Ψ
= W Ψ
= W Ψ
= W Ψ
13 (ι ⊗ βΨ)(Ψx)!
13 Ψ13Ψ12(ι ⊗ βΨ)(x)!
13Ψ12(ι ⊗ βΨ)(x)!
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι)(U ⊗ id) (i ⊗ 1) ⊗
= Ψ13W13(ΦΨ ⊗ ι)(U ⊗ id) (i ⊗ 1) ⊗
= Ψ13W13(ι ⊗ K ⊗ ι)( Ψ ⊗ 1)(Φ ⊗ id) (i ⊗ 1) ⊗
23(ι ⊗ β)(x)Ψ23!
13(ι ⊗ β)(x)Ψ23!
13(ι ⊗ β)(x)Ψ23!
13(ι ⊗ β)(x)Ψ23!
13Ψ12W ∗
13Ψ12Ψ∗
= W Ψ
= W Ψ
Ψ′∗
13W ∗
Ψ′∗
13W ∗
π bG1
⊗ι
π bG1
⊗ι
π bGΨ
1
⊗ι
π bGΨ
1
⊗ι
π bG1
⊗ι
W ∗
Ψ′∗
Ψ′∗
13(ι ⊗ β)(x)Ψ23!
W ∗
= (ι ⊗ K ⊗ ι)Ψ13 Ψ12 Ψ∗
23W13(Φ ⊗ id) (i ⊗ 1) ⊗
π bG1
⊗ι
11
13(ι ⊗ β)(x)! Ψ23
W ∗
= (ι ⊗ K ⊗ ι)Ψ13 Ψ12 Ψ∗
23(ι ⊗ Ind(β))(Φ(ι ⊗
π bG1
x))Ψ23.
Using the equality
(ι ⊗ Ind(β))( Ψ) = Ψ13 Ψ12
we get
(ι ⊗ Ind(βΨ))(ι ⊗ K) ΨΦ i ⊗
π bG1
x!
23 (ι ⊗ Ind(β)) ΨΦ ι ⊗
π bG1
x!! Ψ23.
= (ι ⊗ K ⊗ ι) Ψ∗
Hence we see that
(K ∗ ⊗ ι) Ind(βΨ)K(v) = Ψ∗ Ind(β)(v)Ψ
(23)
1 ))) we have
module over itself. Let W Ψ
In what follows we summarize the above considerations:
(K ∗ ⊗ ι) Ind(βΨ)(KxK ∗)(K ⊗ ι) = Ψ∗ Ind(β)(x) Ψ.
1 )). In particular, for any x ∈ K(Ind(C0(bGΨ
for any v ∈ Ind(C0(bGΨ
Theorem 5.3. Let C0(bGΨ) be the C∗-algebra of the quantum group bGΨ treated as the C∗-Hilbert
1 ∈ M(C0(GΨ) ⊗ C0(bGΨ)) be the left regular corepresentation of GΨ
1 )) → Ind(C0(bG1)) be
and βΨ be the right coaction of bGΨ defined in (6). Let K : Ind(C0(bGΨ
Ind(C0(bG1)) by means of K).
1 ) ∈ L(L2(G) ⊗ Ind(C0(bG1))) is given by
where Ψ, Ψ′ ∈ L(L2(G) ⊗ Ind(C0(bG1))) are the unitary elements defined by (19) and (20). The
(15) (in what follows we shall identify Ind(C0(bGΨ
the isomorphism introduced in Eq.
The induced coaction Ind(W Ψ
Ind(W Ψ
1 ) = Ψ Ind(W1) Ψ′
twisted induced coaction
1 )) with
is related with its untwisted counterpart Ind(β) by the following formula
Ind(β)Ψ : Ind(C0(bGΨ
1 )) → M(Ind(C0(bG1)) ⊗ C0(bGΨ))
Ind(βΨ)(v) = Ψ∗ Ind(β)(v)Ψ
(24)
(25)
for any v ∈ Ind(C0(bG1)), where Ψ ∈ L(Ind(C0(bG1)) ⊗ C0(bGΨ)) is the unitary element defined by
(17).
6. Rieffel deformation of homogeneous spaces - the quotient case
Let G be a locally compact group, G1 its closed subgroup and let X = G / G1 be the homo-
geneous space of left G1-cosets. The standard action of G on X gives rise to the G-C∗-algebra
X = (C0(X), ∆X). Suppose that Γ ⊂ G1 is a closed abelian subgroup and let Ψ be a 2-cocycle on
the dual group Γ. Using the results of paper [8] described in Section 4 we know that X may be
deformed to GΨ-C∗-algebra XΨ.
On the other hand, applying the Rieffel deformation to the subgroup G1 ⊂ G we get a closed
1 be the C∗-algebraic
quantum subgroup GΨ
quotient space of Definition 2.6. The aim of this section is to show that GΨ/GΨ
1
1 ⊂ GΨ in the sense of Definition 2.5 of [16]. Let GΨ/GΨ
∼= XΨ.
Let ι ∈ Mor(C0(X), C0(G)) be the standard embedding - it maps a function f ∈ C0(X) to
the same function on G which is constant on the right G1-cosets. Let α : Γ → Aut(C0(X)) and
ρ : Γ2 → Aut(C0(G)) be the actions introduced in Section 4. The action ρ of Γ2 on C0(G)
restricts to the action α on C0(X). To be more precise, the second copy of Γ in Γ2 acts trivially
on C0(X) and the action of the first copy coincides with α. By the results of Section 3.2 of [9] the
embedding ι may be twisted to the embedding ιΨ ∈ Mor(C0(XΨ), C0(GΨ)). The comultiplication
12
P. KASPRZAK
∆GΨ restricts to the coaction ∆XΨ on ιΨ(C0(XΨ)). The last statement follows from Theorem 4.11
of [9] and the description of ∆XΨ given in Section 4. In particular, ∆XΨ is implemented by the
multiplicative unitary:
∆XΨ (a) = W Ψ∗(1 ⊗ a)W Ψ,
(26)
for any a ∈ C0(XΨ).
As was explained in Remark 2.7, the crossed product of C0(GΨ/GΨ
1 ) by the coaction of GΨ
1 )))
coincides with K(Ind(C0(bGΨ
1 ))). Using Theorem 5.3 we see that we may identify K(Ind(C0(bGΨ
with K(Ind(C0(bG1))), which in turn may be identified with the C∗-algebra G⋉C0(G/G1). The for-
mer C∗-algebra may be identified with B = [C0(bG) C0(X)] ⊂ B(L2(G)) - the C∗-algebra generated
by C0(bG) and C0(X) inside B(L2(G)).
1 ))) = B the crossed product structure of K(Ind(C0(bGΨ
Under the identification K(Ind(C0(bGΨ
M(B ⊗ C0(bGΨ)) is implemented by the unitary bV Ψ introduced in (5). Eq.
corepresentation W Ψ ∈ M(C0(GΨ) ⊗ C0(bGΨ)) ⊂ M(C0(GΨ) ⊗ B). We are now well prepared to
1 )))
may be described as follows. Using Eq. (25) one can see that the dual coaction Ind(βΨ) : B →
(24) shows that
1 ) ∈ M(C0(GΨ) ⊗ B) can be identified with the regular
the induced corepresentation Ind(W Ψ
prove
Theorem 6.1. Let XΨ be the Rieffel deformation of the homogeneous space X and GΨ/GΨ
Vaes' quotient considered above. Then XΨ ∼= GΨ/GΨ
1 .
1 the
Proof. By the universal properties of the crossed product C∗-algebra Γ ⋉ρ C0(X) there exists a
unique morphism ιΓ ∈ Mor(Γ ⋉ρ C0(X), B) which is identity on C0(X) ⊂ M(Γ ⋉ρ C0(X)) and such
that it sends the unitary generator uγ ∈ M(C∗(Γ)) to the left shift Lγ ∈ M(B). Using Theorem 3.6,
[9] we may conclude that the restriction of ιΓ to C0(XΨ) gives rise to the injective morphism ιC0(XΨ)
which we shall denote by ιΨ ∈ Mor(C0(XΨ), B). Let ρΨ be the twisted dual action on Γ ⋉ρ C0(X).
In what follows we shall interpret it as a coaction ρΨ ∈ Mor(Γ ⋉ρ C0(X), Γ ⋉ρ C0(X) ⊗ C∗(Γ)).
Under this interpretation, the relation of ρΨ with its untwisted counterpart ρ is established by the
following formula:
for any x ∈ Γ ⋉ρ C0(X). Let us note that ιΓ intertwines the twisted dual coactions:
ρΨ(x) = Ψ∗ρ(x)Ψ
Ind(βΨ) ◦ ιΓ = (ιΓ ⊗ id) ◦ ρΨ.
In order to see it we have to observe that:
• ιΓ intertwines ρ and Ind(β):
Ind(β) ◦ ιΓ = (ιΓ ⊗ id) ◦ ρ.
(27)
(28)
(29)
Indeed, the equality Ind(β)(ιΓ(f )) = (ιΓ ⊗ id)(ρ(f )) for any f ∈ C0(X) is the consequence
of the the simultaneous invariance of f under ρ and Ind(β). Moreover Ind(β) ◦ ιΓ(uγ) =
Lγ ⊗ Lγ = (ιΓ ⊗ id) ◦ ρ(uγ) for any γ ∈ Γ. Using the fact C0(X) and C∗(Γ) generate
Γ ⋉ρ C0(X) we get a proof of (29).
• Let Ψ be the unitary element introduced in (12). Note that (ιΓ ⊗ id)Ψ = Ψ.
Using the above observations and Eqs. (25), (27) one easily gets (28) which in turn implies that
ιΨ(C0(XΨ)) ⊂ M(B)Ind(βΨ). The proof of the equality ιΨ(C0(XΨ)) = C0(GΨ/GΨ
1 ) is based on
Theorem 6.7 of [16]. Its application requires the check of the following two conditions:
• The map x 7→ W Ψ∗(1 ⊗ x)W Ψ defines a continuous coaction on ιΨ(C0(XΨ)). Indeed, this
follows from (26).
• We have [ιΨ(C0(XΨ)) C0(bG)] = B. In order to see that we compute
[ιΨ(C0(XΨ)) C0(bG)] = [ιΨ(C0(XΨ)) C∗(Γ) C0(bG)] = [ιΓ(Γ ⋉ρ C0(XΨ)) C0(bG)]
= [ιΓ(Γ ⋉ρ C0(X)) C0(bG)] = [ιΓ(C0(X)) C∗(Γ) C0(bG)]
= [C0(X) C0(bG)] = B
13
where in the first and the fourth equality we used the fact that [C∗(Γ) C0(bG)] = C0(bG)
and in the third equality we used the fact that Γ ⋉ρ C0(XΨ) = Γ ⋉ρ C0(X) (note that the
action of Γ on C0(X) and on C0(XΨ) is denoted by the same ρ).
This ends the proof of the fact that XΨ ∼= GΨ/GΨ
1 .
(cid:3)
7. Rieffel deformation of G-simple C∗-algebras
Let G be a locally compact quantum group corresponding to a locally compact group G. The
aim of this section is to prove that the Rieffel deformation XΨ of a G-simple C∗-algebra X is GΨ-
simple (see Definition 2.4). In particular, the Rieffel deformation of a quotient space X = G / G1
(also in the case when Γ is a subgroup of G but not of G1) is GΨ-simple. The idea of the following
proof is based in the functorial properties of the Rieffel deformation (see Section 3.2, [9]).
Theorem 7.1. Let X be a G-simple C∗-algebra. Let Γ be a closed abelian subgroup of G and Ψ a
2-cocycle on the dual group Γ. The Rieffel deformation XΨ of X is GΨ-simple.
Proof. Let B be GΨ-C∗-algebra and let π ∈ MorGΨ (C0(XΨ), C0(B)). The Rieffel deformation ΓΨ of
Γ is a quantum closed subgroup of GΨ. Actually, ΓΨ = Γ which easily follows from the abelianity
of Γ (see Appendix B of [7]). Let α : Γ → Aut(C0(XΨ)) be the action that corresponds to the Γ-
restriction of the coaction ∆XΨ . Similarly, we may introduce β : Γ → Aut(C0(B)). Obviously the
morphism π intertwines α and β: π◦αγ = βγ ◦π for any γ ∈ Γ. The deformation data (C0(B), β, ¯Ψ)
¯Ψ) and the deformed morphism
gives rise to the deformed C∗-algebra that we shall denote by C0(B
π ¯Ψ ∈ Mor(C0(X), C0(B
¯Ψ)). Note that (GΨ) ¯Ψ = G. Using the ideas presented in Section 4 we
may construct the twisted coaction ∆ ¯Ψ
¯Ψ and check that π ¯Ψ is a G-morphism. The
G-simplicity of X implies that ker π ¯Ψ = {0}. Using Proposition 3.8 [9] we see that ker π = {0},
which ends the proof of GΨ-simplicity of XΨ .
(cid:3)
B of G on B
References
[1] S. Baaj, G. Skandalis: Unitaires multiplicatifs et dualit´e pour les produits crois´es de C-algebres, Ann. Sci.
Ec. Nor. Sup. (4) 26 (1993), 425 -- 488.
[2] K. de Commer: Galois objects and cocycle twisting for locally compact quantum groups, to appear in
J. Op. Theory.
[3] M. Brin, G. Stuck: Introduction to Dynamical Systems, Camb. Univ. (2003)
[4] M. Enock, L. Vainerman: Deformation of a Kac algebra by an abelian subgroup, Commun. Math. Phys.
178 (1996), no. 3, p. 571 -- 596.
[5] P. Fima, L. Vainerman: Twisting and Rieffel's deformation of locally compact quantum groups. Deformation
of the Haar measure. Commun. Math. Phys. 286 (2009), no. 3, p. 1011 -- 1050.
[6] S. Kasliszewski, J. Quig: Categorical Landstad duality for actions, Indiana Univ. Math. J. 58 (2009), no. 1,
p. 415 -- 441.
[7] P. Kasprzak: Heisenberg-Lorentz Quantum Group, to appear in J. Noncom. Geom.
[8] P. Kasprzak: Rieffel deformation of groups coactions, to appear in Commun. Math. Phys.
[9] P. Kasprzak: Rieffel deformation via crossed products, J. Funct. Anal. 257 (2009), no. 5, p. 1288 -- 1332
[10] A. Kleppner: Multipliers on Abelian Groups, Math. Annalen 158, 11 -- 34 (1965).
[11] J. Kustermans, S. Vaes: Locally compact quantum groups. Ann. Sci. Ec. Norm. Sup. 33 (2000), no. 4, p.
837 -- 934.
[12] T. Masuda, Y. Nakagami, S.L. Woronowicz: A C∗-algebraic framework for quantum groups, Inter-
nat. J. Math. 14 (2003), no. 9, p. 903 -- 1001.
[13] P. Podle´s Quantum spheres, Lett. Math. Phys. 14 (1987), no. 3, p. 193 -- 202.
[14] P. Podle´s Symmetries of Quantum Spaces. Subgroups and Quotient Spaces of Quantum SU (2) and SO(3)
Groups. Commun. Math. Phys. 170 (1995), no. 1, p. 1 -- 20.
[15] M.A. Rieffel: Deformation quantization for action of Rd, Mem. Am. Math. Soc. 106 (1993), no. 506.
[16] S. Vaes: A new approach to induction and imprimitivity results, J. Funct. Anal. 229 (2005), no 2, p. 317 --
374.
[17] J. C. Varilly: Quantum symmetry groups of noncommutative spheres, Commun. Math. Phys. 221 (2001),
no. 3, p. 511 -- 523.
[18] S.L. Woronowicz: C∗-algebras generated by unbounded elements, Rev. Math. Phys. 7, No. 3, (1995), 481
-- 521.
Department of Mathematical Sciences, University of Copenhagen
14
P. KASPRZAK
On leave from Department of Mathematical Methods in Physics, Faculty of Physics, Warsaw Uni-
versity
E-mail address: [email protected]
RIEFFEL DEFORMATION OF HOMOGENEOUS SPACES
P. KASPRZAK
Abstract. Let G1 ⊂ G be a closed subgroup of a locally compact group G and let X = G / G1
be the quotient space of left cosets. Let X = (C0(X), ∆X) be the corresponding G-C∗-algebra
where G = (C0(G), ∆). Suppose that Γ is a closed abelian subgroup of G1 and let Ψ be a
2-cocycle on the dual group Γ. Let GΨ be the Rieffel deformation of G. Using the results of
paper [8] we may construct GΨ-C∗-algebra XΨ - the Rieffel deformation of X. On the other
hand we may perform the Rieffel deformation of the subgroup G1 obtaining the closed quantum
1 ⊂ GΨ, which in turn, by the results of [16], leads to the GΨ-C∗-algebra GΨ/GΨ
subgroup GΨ
1 .
∼= XΨ. We also consider the case where Γ ⊂ G is not a
In this paper we show that GΨ/GΨ
subgroup of G1, for which we cannot construct the subgroup GΨ
1 . Then XΨ cannot be identified
with a quantum quotient. What may be shown is that it is a GΨ-simple object in the category
of GΨ-C∗-algebras.
1
0
1
0
2
l
u
J
4
1
]
.
A
O
h
t
a
m
[
2
v
3
6
2
4
.
6
0
0
1
:
v
i
X
r
a
Contents
1.
Introduction
2. Preliminaries
2.1. G-C∗-category
2.2. Quantum homogeneous spaces
3. Group algebra twist
4. Rieffel deformation of group coactions
5.
6. Rieffel deformation of homogeneous spaces - the quotient case
7. Rieffel deformation of G-simple C∗-algebras
References
Induction of the regular corepresentation
1
2
2
3
4
5
6
12
13
14
1. Introduction
The theory of locally compact quantum groups (LCQG) has already reached its maturity.
Almost ten years have passed since the appearance of the seminal paper of J. Kustermans and
S. Vaes [11], where the axiomatic theory was formulated. A locally compact quantum group is a
C∗-bialgebra (A, ∆) equipped with a left and a right Haar weight φ and ψ. Imposing some natural
conditions on ∆ (weak cancellation) and on the Haar weights φ and ψ (KMS-type conditions) the
authors were able to develop the theory, proving among other things the existence of the coinverse
which admits the polar decomposition κ = R ◦ τi/2 and showing that the theory is self dual. The
assumption of the existence of the Haar weights, which is a theorem for the classical groups and
for the compact quantum groups, may be perceived as a drawback. It seems that a Haar weights
free axiomatization is out of reach. There exists second formulation of the LCQG theory, due to
T. Masuda, Y. Nakagami and S.L. Woronowicz [12], in which the authors include the existence of
the coinverse κ in the axioms. But to develop the theory they still need to assume the existence
of a left Haar weight. In fact, it can be shown that both theories are equivalent.
One way of constructing examples of LCQGs is to start with a classical group G and search
for its deformations Gq. In general, the link between G and Gq is not rigorously described. There
2000 Mathematics Subject Classification. Primary 46L89, Secondary 20N99.
Supported by the Marie Curie Research Training Network Non-Commutative Geometry MRTN-CT-2006-031962.
Supported by Geometry and Symmetry of Quantum Spaces, PIRSES-GA-2008-230836.
1
2
P. KASPRZAK
is at least one mathematical procedure - the Rieffel deformation - where this correspondence is
clear. For the original approach of Rieffel we refer to [15]. In this paper we shall use our recent
approach to the Rieffel deformation which describes it in terms of crossed product construction
described in Section 3.
1 of the twisted group GΨ
(see [9]). The Rieffel deformation of G will be denoted by GΨ and its dual will be denoted by bGΨ.
The deformation procedure of a locally compact group in terms of the transition from bG to bGΨ is
Let G be a locally compact group and D a G - C∗-algebra. Using the results of [8] one may apply
the Rieffel deformation to D, obtaining the deformed GΨ - C∗-algebra DΨ. The concise account
of the deformation procedure of G - C∗-algebras is the subject of Section 4. In Section 2 beside
giving some preliminaries on G-C∗-algebras we also discuss the notion of a quantum homogeneous
space and the C∗-algebraic quotient space G/G1, a construction due to Vaes [16]. His construction
may be performed for any closed quantum subgroup G1 of a regular LCQG G - for regularity we
refer to [1]. The relation of G/G1 with the induction procedure of the regular corepresentation is
explained. The reader's awareness of this relation is crucial in the understanding of the proof that
XΨ ∼= GΨ/GΨ
1 . In Section 5 we perform the induction procedure of the regular corepresentation
W Ψ
1 and compare the resulting objects with their untwisted counterparts.
This enables us to prove that XΨ ∼= GΨ/GΨ
1 which is the subject of Section 6. Finally, in Section
7 we comment on the case where XΨ is not of the quotient type. In connection with it we show
that the Rieffel deformation of a G-simple C∗-algebra is GΨ-simple.
The particular case of the Rieffel deformation of a homogeneous space has been discussed in
[17].
In this paper J. Varilly treats the situation where there is given a pair Γ ⊂ G1 ⊂ G of
closed subgroups with Γ being abelian and G compact. He shows that it is possible to perform
a covariant deformation of X = G / G1 obtaining a quantum homogeneous GΨ-space XΨ. In this
specific situation the difficulties that one encounters in general do not manifest themselves.
Throughout the paper we will freely use the language of C∗-algebras and the theory of locally
compact quantum groups. For the notion of multipliers and morphism of C∗-algebras we refer the
reader to [18]. For the theory of locally compact quantum groups we refer to [11] and [12].
I would like to express my gratitude to W. Szyma´nski for many stimulating discussions, which
greatly influenced the final form of this paper.
2. Preliminaries
2.1. G-C∗-category. Let us fix a notation related with a locally compact quantum group G. The
C∗-algebra and the comultiplication of G will be denoted by C0(G) and ∆G respectively. The von
Neumann algebra associated with G and the Hilbert space obtained by the GNS-representation
shall denote the reduced version of the dual of G. The modular conjugations related with the left
related with the left Haar weight will be denoted respectively by L∞(G) and L2(G). By bG we
Haar weight on G (on bG) will be denoted by J (by J).
The main subject of this paper is related with the quantum group actions on quantum spaces.
The following definition may be traced back to [13]. To formulate it we adopt the following
notation: a closed linear span of a subspace W ⊂ V of a Banach space V will be denoted by [W].
Definition 2.1. Let G be a LCQG. A G-C∗-algebra is a pair D = (D, ∆D) consisting of a C∗-
algebra D and a coaction ∆D ∈ Mor(D, C0(G) ⊗ D):
(ι ⊗ ∆D)∆D = (∆G ⊗ ι)∆D
such that [∆D(D)(C0(G) ⊗ 1)] = C0(G) ⊗ D. The C∗-algebra D will also be denoted by C0(D)
and the coaction ∆D will be denoted by ∆D.
Remark 2.2. Suppose that G corresponds to an ordinary locally compact group G. There is
a 1-1 correspondence between G-C∗-algebras and G-C∗-algebras, i.e. C∗-algebras equipped with
a continuous action of G. In order to describe it we introduce the characters χg : C0(G) → C
that are associated with the points of G: χg(f ) = f (g), for any g ∈ G and f ∈ C0(G). For
ang G-C∗-algebra D we define the corresponding continuous action α : G → Aut(C0(D)) by the
formula: αg(a) = (χg−1 ⊗ ι)∆D(a) for any g ∈ G and a ∈ C0(D).
3
The category of G-C∗-algebras was consider either explicitly or implicitly in many different
contexts - see for instance [6] where it appears explicitly in the context of the Landstad duality.
In order to specify the category of G-C∗-algebras we adopt the following notion of G-morphisms.
Definition 2.3. Let G be a LCQG and suppose that B and D are G-C∗-algebras. We say that a
morphism π : C0(B) → C0(D) is a G-morphism from B to D if ∆D ◦ π = (ι ⊗ π) ◦ ∆B. The set of
G-morphism from B to D will be denoted by MorG(B, D).
2.2. Quantum homogeneous spaces. Let G be a locally compact quantum group and X be
a G - C∗-algebra. The concept of a quantum homogeneous G-space exists so far only in the
case of G being compact and C0(X) being unital - see Definition 1.8 of [14]. In this case X is a
quantum homogeneous space if ∆X is ergodic. Let G be a compact quantum group corresponding
to a compact group G. It may be checked that there is a 1-1 correspondence between quantum
homogeneous G-spaces with the underlying commutative C∗-algebra C0(X) and the classical ho-
mogeneous G-spaces. Unfortunately the ergodicity assumption in the non-compact case cannot
serve as a replacement for homogeneity. The next approximation to the homogeneity is the notion
of G-simplicity:
Definition 2.4. Let D be G-C∗-algebra. We say that D is G-simple if for any G-C∗-algebra B
and any G-morphism π ∈ MorG(D, B) we have ker π = {0}.
In order to motivate the introduction the G-simplicity let us prove that in the case of the
compact quantum groups the homogeneity of D implies its G-simplicity.
Lemma 2.5. Let G = (A, ∆) be a compact quantum group with a faithful Haar measure h. Let
X be a quantum homogeneous space with ∆X being injective. Then X is G-simple.
Proof. Using the ergodicity of ∆X we may introduce the state on ρ : C0(X) → C such that
(h ⊗ ι)∆X(a) = ρ(a) for any a ∈ C0(X).
(1)
The faithfulness of h and the injectivity of ∆X imply the faithfulness of ρ. Let π ∈ MorG(X, B)
and suppose that a ∈ ker π. Note that (h ⊗ ι)∆X(a∗a) ∈ ker π:
π((h ⊗ ι)∆X(a∗a)) = (h ⊗ ι)∆X(π(a∗a)) = 0.
On the other hand, using (1) we may see that π((h ⊗ ι)∆X(a∗a)) = ρ(a∗a)1 which together with
the above computation shows that ρ(a∗a) = 0. The faithfulness of ρ implies that a = 0 hence
ker π = {0}.
(cid:3)
Let G be a locally compact quantum group and let G be the corresponding LCQG. It may
be checked that there is a 1-1 correspondence between G-simple C∗-algebras with the underlying
C∗-algebra being commutative and minimal G - spaces - the G-spaces with all orbit being dense
(see p.49, [3]).
As was already mentioned, a definition of a quantum homogeneous space appropriate for locally
compact quantum groups is not yet known. What has been generalized so far is the notion of the
quantum quotient space, due to Vaes [16]. But since the work of Podle´s [13] on quantum spheres
we know that a generic quantum homogeneous space is not of the quotient type. The Rieffel of
the homogeneous spaces deformation provides a new class of examples of non-compact quantum
spaces which in our opinion should also be considered as homogeneous and which generically are
not of the quotient type.
Let us move on to the discussion of the quantum quotient spaces. Let G be a regular LCQG
(for the notion of regularity we refer to [1]). Let G1 be a closed quantum subgroup in the sense
of Definition 2.5 [16]. As a part of the structure we have the injective normal ∗-homomorphism
π : L∞(bG1) → L∞(bG). Its commutant counterpart is denoted by π′ : L∞(bG1)′ → L∞(bG)′. The
definition of the measurable quotient space L∞(G/G1) goes as follows:
L∞(G/G1) = {a ∈ L∞(G) : aπ′(x) = π′(x)a, for any x ∈ L∞(bG1)′}.
A remarkable Theorem 6.1 [16]) provides the existence and the uniqueness of the C∗-algebraic
version C0(G/G1) of L∞(G/G1). For the needs of this paper we shall formulate it as a definition:
4
P. KASPRZAK
Definition 2.6. Let G be a LCQG which is regular and let G1 be its closed quantum subgroup.
The C∗-algebraic quotient G/G1 is the unique G-C∗-algebra defined by the following conditions:
• C0(G/G1) ⊂ L∞(G/G1) is a strongly dense C∗-subalgebra;
• ∆G/G1 is given by the restriction of ∆G to C0(G/G1);
• ∆G(L∞(G/G1)) ⊂ M(K(L2(G)) ⊗ C0(G/G1)) and the ∗-homomorphism
∆G/G1 : L∞(G/G1) → L(L2(G) ⊗ C0(G/G1)) is strict.
For the notion of strictness we refer to Definition 3.1 of [16]. Note that the symbol ∆G/G1
denotes the coaction on C0(G/G1) as well as the strict map defined on L∞(G/G1).
Remark 2.7. The difficulty of the proof of Theorem 6.1, [16] lies in the existence part. To explain
the idea of the proof (which will be also important to understand this paper) let us consider the C∗-
algebra C0(bG1) treated as a Hilbert C∗-module over itself. Performing the induction procedure
of the regular corepresentation W1 ∈ M(C0(G1) ⊗ C0(bG1)) we get the induced Hilbert module
Ind(C0(bG1)) equipped with the additional structure: the induced corepresentation Ind(W1) and
the coaction γ : Ind(C0(bG1)) → M(Ind(C0(bG1)) ⊗ C0(bG)) of the dual quantum group bG. The
coaction γ is obtained by the induction procedure applied to the right coaction β : C0(bG1) →
M(C0(bG1)⊗C0(bG)), β(a) = (ι⊗ π)∆bG1
(a) for any a ∈ C0(bG1). The additional structure consisting
of γ and Ind(W1) enables one to prove that the C∗-algebra of compact operators K(Ind(C0(bG1))
on Ind(C0(bG1)) is canonically isomorphic with the C∗-algebra of a crossed product. The coaction
γ is identified with the dual coaction on the crossed product while Ind(W1) is identified with
the corepresentation implementing the coaction on the γ-invariants. The C∗-algebra C0(G/G1)
is defined as the Landstad-Vaes C∗-algebra of γ-invariants, by which we mean the C∗-algebra
satisfying the conditions of Theorem 6.7, [16].
3. Group algebra twist
Let G be a locally compact group and let δ : G → R+ be the modular function. Suppose that
Γ is a closed abelian subgroup of G1, which in turn is a closed subgroup of G. For reasons which
will become clear later we shall assume that the modular function δ : G → R+ when restricted to
Γ is identically 1: δ(γ) = 1 for any γ ∈ Γ.
In order to perform the twisting procedure one has to fix a 2-cocycle Ψ on Γ, i.e. a continuous
function Ψ : Γ × Γ → T1 satisfying:
(i) Ψ(e, γ) = Ψ(γ, e) = 1 for all γ ∈ Γ;
(ii) Ψ(γ1, γ2 + γ3)Ψ(γ2, γ3) = Ψ(γ1 + γ2, γ3)Ψ(γ1, γ2) for all γ1, γ2, γ3 ∈ Γ.
For the theory of 2-cocycles we refer to [10].
In order to simplify the notation and make further computations less cumbersome we shall
assume that Ψ is a skew symmetric bicharacter on Γ:
Ψ(γ1 + γ2, γ3) =Ψ(γ1, γ3)Ψ(γ2, γ3),
Ψ(γ1, γ2) =Ψ(γ2, γ1).
(2)
(3)
We will only prove our results for such 2-cocycles although we have reasons to believe that they
hold for arbitrary 2-cocycles.
The role of the bicharacter Ψ ∈ M(C0(Γ) ⊗ C0(Γ)) will vary in the course of this paper. The
simplest variation is connected with the identification C0(Γ) ∼= C∗(Γ), which enables us to treat
Ψ as an element of M(C∗(Γ) ⊗ C∗(Γ)). Furthermore, the morphism ι ∈ Mor(C∗(Γ), C∗
l (G1))
which corresponds to the representation Γ ∋ γ 7→ Lγ ∈ M(C∗
l (G1)) enables us to treat Ψ as an
element of M(C∗
l (G1)). Finally, in some cases Ψ will be treated as an operator acting
on L2(G1) ⊗ L2(G1) or L2(G) ⊗ L2(G).
l (G1) ⊗ C∗
Let us describe the twisting procedure of (C∗
l (G), ∆). In what follows we shall use the notation
of Section 2.1, denoting the locally compact quantum group related to the pair (C∗
l (G) and ∆bG = ∆. We may twist the comultiplication on ∆bG by means of
In particular, C0(bG) = C∗
l (G), ∆) by bG.
5
is given by ∆bGΨ . (For more general results concerning the twist of a quantum group by a 2-cocycle
Ψ ∈ M(C0(bG) ⊗ C0(bG)): ∆bGΨ (a) = Ψ∗∆bG(a)Ψ. Using Corollary 5.3 of [4] we see that there exists
a locally compact quantum group bGΨ such that C0(bGΨ) = C0(bG), for which the comultiplication
we refer to [2].) Furthermore, with our modular assumption bGΨ is of the Kac type - the coinverse
κ bGΨ is an involutive anti-automorphism. The dual locally compact quantum group of bGΨ will be
denoted by GΨ. Its description in terms of the Rieffel deformation was given in [9]. It is also of the
Kac type - κ2
GΨ = id - which by Example 3.4 [1] implies that GΨ is regular. The only reason for
the assumption δΓ = 1 is to ensure the regularity of GΨ which is important in the construction
of the quotient of a locally compact quantum group by its closed quantum subgroup.
In what follows we shall give the formula for the multiplicative operators W Ψ ∈ L∞(GΨ) ⊗
L∞(bGΨ) and V Ψ ∈ L∞(GΨ)′ ⊗ L∞(bGΨ) where we adopted the notation of Section 2.1 of [16]. In
order to do it let us introduce an element F ∈ M(C0(Γ) ⊗ C0(Γ)): F (γ1, γ2) = Ψ(γ2, γ1 + γ2).
Using Theorem 1 of [5] and the the properties of Ψ we may see that the operator W Ψ acts on
B(L2(G) ⊗ L2(G)) and it is of the form W Ψ = ΨW Ψ′ where
Ψ′ = ( J ⊗ J)F ( J ⊗ J).
(4)
The assumption of the co-stability introduced in Section 2.7 of [5] is satisfied trivially in the
discussed case. The required group homomorphism t 7→ γt is the trivial homomorphism: γt = e ∈ Γ
(we replaced the γt introduced in [5] by γt which is more appropriate in our context).
we get
To give the formula for the multiplicative unitary V Ψ we only have to invoke the fact that it
cop.
cop is the flip of the comultiplication ∆bGΨ . By the skew-symmetry of Ψ
is the fundamental multiplicative unitary corresponding to the co-opposite quantum group bGΨ
The comultiplication of bGΨ
The twist construction that we applied to G can be also applied to G1 leading to bGΨ
1 and the
dual GΨ
1 . Note that we do not impose the modular condition on the embedding Γ ⊂ G1. Again
using Theorem 1 of [5] one may see that W Ψ
1 are related to W1 and V1 in the way analogous
to the case of G described above (see Eq. (5)). The aforementioned assumption of co-stability is
also satisfied but the homomorphism t 7→ γt may be non-trivial. In order to see that, we shall use
the modular function δ1 : G1 → R+. For any t ∈ R we may define the character γt ∈ Γ given by
hγt, γi = δit
1 (γ). The homomorphism t 7→ γt is the one that ensures the co-stability. It may be
1 is a closed quantum subgroup of GΨ in the sense of Definition
checked that the quantum group GΨ
V Ψ = Ψ∗ V Ψ′∗.
1 and V Ψ
(5)
2.5 of [16]. In particular, there exists the embedding π : L∞(bGΨ
acting between commutants will be denoted by π′ : L∞(bGΨ
1 )′ → L∞(bGΨ)′.
4. Rieffel deformation of group coactions
1 ) → L∞(bGΨ). Its counterpart
Suppose that G is a locally compact group containing an abelian closed subgroup Γ and let Ψ
be a skew-symmetric bicharacter on Γ (see Section 3). Let X = (C0(X), ∆X) be a G-C∗-algebra
In paper [9] we defined a GΨ-C∗-algebra XΨ - the Rieffel deformation of
(see Definition 2.1).
X. The deformation procedure may be viewed as a two steps procedure: first extend ∆X to the
morphism of the appropriate crossed products and then twist the extension by a unitary obtained
from Ψ. Let us be more precise.
Let β : G → Aut(C0(X)) be the continuous action corresponding to the coaction ∆X (see
Remark 2.2). Let α : Γ → Aut(C0(X)) be the restriction of β to the subgroup Γ. The C∗-
algebra C0(XΨ) is defined as the Landstad algebra of the Γ-product (Γ ⋉α C0(X), λ, ρΨ) where
ρΨ : Γ → Aut(Γ ⋉α C0(X)) is the Ψ-deformed dual action (see Section 3, [9]). Let us note that
∆X ◦ αγ = (ργ,e ⊗ ι) ◦ ∆X, where ρ : Γ2 → Aut(C0(G)) is the action defined by ργ1,γ2(f )(g) =
f (γ−1
1 gγ2) for any γ1, γ2 ∈ Γ and g ∈ G. The universal property of the crossed product Γ ⋉α C0(X)
enables us to define the extension ∆Γ
X ∈ Mor(Γ ⋉α C0(X), Γ2 ⋉ρ C0(G) ⊗ Γ ⋉α C0(X)) of ∆X to
the level of crossed product.
6
P. KASPRZAK
In order to describe the aforementioned twisting step we define Υ ∈ M(Γ2⋉ρC0(G)⊗Γ⋉αC0(X))
as follows. Let Φ ∈ Mor(C∗(Γ2), Γ2 ⋉ρ C0(G) ⊗ Γ ⋉α C0(X)) be the morphism that corresponds to
the representation Γ ∋ (γ1, γ2) 7→ λe,γ1 ⊗ λγ2 ∈ M(Γ2 ⋉ρ C0(G) ⊗ Γ ⋉α C0(X)). Applying it to ¯Ψ we
get Υ = Φ( ¯Ψ). We may define ∆XΨ by the formula ∆XΨ (b) = Υ∆Γ
X(b)Υ∗ for any b ∈ Γ ⋉α C0(X).
By Theorems 4.3 and 4.4 of [8] ∆XΨ restricts to a morphism ∆XΨ : C0(XΨ) → C0(GΨ) ⊗ C0(XΨ)
which is a continuous coaction of GΨ on C0(XΨ). This defines GΨ-C∗-algebra XΨ.
The reader may have noticed some differences between the above description of the deformation
procedure of D and the one given in [8]. They are due to the adopted definition of GΨ as the dual
of bGΨ and the fact D is a left G-C∗-algebra whereas in [8] we consider the right case.
5. Induction of the regular corepresentation
1 )) of GΨ
1 ), defined by
βΨ(a) = (ι ⊗ π)∆bGΨ
1 ) ⊗ C0(bGΨ
1 )) together with the induced corepresentation Ind(W Ψ
In this section we shall apply the induction procedure to the regular corepresentation W Ψ
1 on C0(bGΨ
1 ), where C0(bGΨ
1 ) ∈ L(C0(GΨ) ⊗ Ind(C0(bGΨ
1 ) → M(C0(bGΨ
1 ∈
1 ) is treated as a C∗-Hilbert module over
M(C0(GΨ
itself. For the induction procedure in the framework of LCQGs we refer to [16] (we shall also
1 )-Hilbert module
1 ))).
1 ) ⊗
adopt the notation of this paper). As a result, we obtain the induced C0(bGΨ
Ind(C0(bGΨ
We shall also apply the induction procedure to the right coaction βΨ : C0(bGΨ
C0(bGΨ)) of bGΨ on C0(bGΨ
where π ∈ Mor(C0(bGΨ
1 ), C0(bGΨ)) is the standard embedding. The result is the coaction
Ind(βΨ) : Ind(C0(bGΨ
1 )) → M(Ind(C0(bGΨ
Finally, the induced objects Ind(C0(bGΨ
untwisted counterparts Ind(C0(bG1)), Ind(W1), and Ind(β).
Let us first recall that the von Neumann algebra of the twisted quantum group bGΨ remains
unchanged L∞(bGΨ) = L∞(bG). This implies that the imprimitivity bimodule I Ψ, which is defined
stays undeformed: I Ψ = I. Nevertheless, the coaction αI Ψ : I Ψ → I Ψ ⊗ L∞(bGΨ): αI Ψ(v) =
I Ψ = {v ∈ B(L2(G1), L2(G)) vm = π′(m)v, for any m ∈ L∞(bGΨ
gets twisted. The relation with its untwisted counterpart is established by
1 )), Ind(W Ψ
1 ) and Ind(βΨ) will be compared with their
1 )) ⊗ C0(bGΨ)).
V Ψ(v ⊗ 1)(ι ⊗ π) V Ψ∗
the following formula:
1
1 )′}
(a),
1
(6)
by
αI Ψ (v) = Ψ∗αI(v)Ψ.
(7)
Let us analyze the strict ∗-homomorphism πbGΨ
[16] it was denoted by πl), given by:
1
: L∞(bGΨ
1 ) → L(L2(G)) ⊗ C0(bGΨ
1 )) (in Lemma 4.5,
πbGΨ
1
(m) = (π ⊗ ι)(W Ψ
1 (m ⊗ 1)W Ψ∗
1
)
1 ). It may be noted that the map πbGΨ
coincides with (π ⊗ id)∆bGΨ
1,cop
when
1
1 )) is expressed by
for any m ∈ L∞(bGΨ
appropriately interpreted. Its relation with πbG1
the twisting formula:
for any m ∈ L∞(bGΨ
1 ) = L∞(bG1). Using πbGΨ
1
F Ψ = I Ψ ⊗
π bGΨ
1
πbGΨ
1
(m) = ΨπbG1
(m)Ψ∗
: L∞(bG1) → L(L2(G)) ⊗ C0(bGΨ
we may introduce the C0(bGΨ
L2(G) ⊗ C0(bGΨ
1 ).
1 ) - module
It is equipped with the strict ∗-homomorphism πΨ
l
homomorphism πΨ
r : L∞(bGΨ) → L(F Ψ), which are defined as follows:
πΨ
l (m)(i ⊗
π bGΨ
1
h ⊗ a1) = (mi) ⊗
π bGΨ
1
h ⊗ a1
7
: L∞(bGΨ) → L(F Ψ) and the strict ∗ - anti-
πΨ
r (m)(i ⊗
π bGΨ
1
h ⊗ a1) = i ⊗
π bGΨ
1
Jm∗ Jh ⊗ a1
for any m ∈ L∞(bGΨ), i ∈ I Ψ, h ∈ L2(G) and a1 ∈ C0(bGΨ
prove:
1 ). In order to compare F Ψ and F we
Proposition 5.1. There exists a unitary transformation U ∈ L(F , F Ψ) such that
U (i ⊗
π bG1
h ⊗ a1) = i ⊗
π bGΨ
1
Ψ(h ⊗ a1)
for any i ∈ I, h ∈ L2(G) and a1 ∈ C0(bGΨ
Proof. For the existence of U it is enough to note that:
1 ). U intertwines πl with πΨ
l and πr with πΨ
r .
U (ia ⊗
π bG1
h ⊗ a1) = ia ⊗
π bGΨ
1
Ψ(h ⊗ a1)
= i ⊗
π bGΨ
1
ΨπbG1
(a)(h ⊗ a1)
= U (i ⊗
π bG1
πbG1
(a)(h ⊗ a1)).
The fact that U is unitary and that it possesses the required intertwining properties can be verified
by a straightforward computation.
(cid:3)
of the coaction αI Ψ given by (7) and the coaction αL2(G)⊗C0( bGΨ
Let us now introduce the coaction αF Ψ : F Ψ → M(F Ψ ⊗ C0(bGΨ)). It is defined as a product
C0(bGΨ
1 ) ⊗ C0(bGΨ)) given by:
1 ) : L2(G) ⊗ C0(bGΨ
1 ) → M(L2(G) ⊗
αL2(G)⊗C0( bGΨ
1 )(ξ) = V Ψ
13(ξ ⊗ 1),
for any ξ ∈ L2(G) ⊗ C0(bGΨ
1 ). The formula defining αF Ψ goes as follows:
αF Ψ(i ⊗
π bGΨ
1
ξ) = αI Ψ (i) ⊗
π bGΨ
1
⊗ι
αL2(G)⊗C0( bGΨ
1 )(ξ).
The coaction αF Ψ corresponds to the corepresentation Y Ψ ∈ L(F Ψ ⊗ C0(bGΨ)) of bGΨ on F Ψ:
Y Ψ(f ⊗ a) = αF Ψ (f )(1 ⊗ a).
8
P. KASPRZAK
In order to compare Y Ψ with its undeformed counterpart Y ∈ L(F ⊗ C0(bG)) we compute:
(U ∗ ⊗ 1)Y Ψ(U ⊗ 1)(i ⊗
π bG1
ξ ⊗ a)
= (U ∗ ⊗ 1)Y Ψ(i ⊗
π bGΨ
1
Ψξ ⊗ a)
= (U ∗ ⊗ 1)(αF Ψ (i ⊗
π bGΨ
1
Ψξ))(1 ⊗ a)
= (U ∗ ⊗ 1)(Ψ∗αI (i)Ψ ⊗
π bGΨ
1
⊗ι
Ψ∗
13
V13Ψ′∗
13Ψ12(ξ ⊗ 1))(1 ⊗ a)
⊗ι
⊗ι
= (Ψ∗αI (i)Ψ ⊗
π bG1
= (Ψ∗αI (i)Ψ ⊗
π bG1
= (Ψ∗αI (i) ⊗
π bG1
= (Ψ∗αI (i) ⊗
π bG1
⊗ι
⊗ι
Ψ∗
12Ψ∗
13
V13Ψ′∗
13Ψ12(ξ ⊗ 1))(1 ⊗ a)
Ψ∗
12Ψ∗
13Ψ12Ψ∗
23
V13Ψ′∗
13(ξ ⊗ 1))(1 ⊗ a)
Ψ13Ψ23Ψ∗
12Ψ∗
13Ψ12Ψ∗
23
V13Ψ′∗
13(ξ ⊗ 1))(1 ⊗ a)
V13Ψ′∗
13(ξ ⊗ 1))(1 ⊗ a)
= (πl ⊗ id)(Ψ∗)Y (πr ⊗ id)(Ψ′∗)(i ⊗
π bG1
In the fifth equality we used the easy to verify formula V13Ψ12 V ∗
equality we used: (πbG1
23 and in the sixth
⊗ ι)Ψ = Ψ13Ψ23. For the formula for Ψ′ we refer to (4). This computation
13 = Ψ12Ψ∗
ξ ⊗ a)
(U ∗ ⊗ id)Y Ψ(U ⊗ id) = (πl ⊗ id)(Ψ∗)Y (πr ⊗ id)(Ψ′∗),
shows that the relation between the corepresentations Y ∈ L(F ⊗ C0(bG)) and Y Ψ ∈ L(F Ψ ⊗
C0(bGΨ)) is given by
where we treat the anti-homomorphism πr : L∞(bG) → L(F ) as the homomorphism of the com-
mutant L∞(bG)′ (note that Ψ′ ∈ L∞(bG)′ ⊗ L∞(bG)).
We are now ready to compare the induced module Ind(C0(bGΨ
1 ) with their untwisted counterparts Ind(C0(bG1)) and Ind(W1). Let us first recall
tation Ind(W Ψ
the construction of the untwisted objects. By the results of [16] there exists a (unique) strict
∗-homomorphism Θ : B(L2(G)) → L(F ) such that
1 )) and the induced corepresen-
(Θ ⊗ ι) V = Y,
Θ( Jx∗ J) = πr(x),
such that
where by assumption v : L2(G) → F is a continuous map. It follows from the above definition
Ind(C0(bG1)) = {v : L2(G) → F : vx = Θ(x)v for all x ∈ B(L2(G))},
for any x ∈ L∞(bG). The induced C∗-module is defined as
that we may define the unitary transformation Φ : F → L2(G) ⊗ Ind(C0(bG1)) of C0(bG1)-modules,
for any h ⊗ v ∈ L2(G) ⊗ Ind(C0(bG1)). The induced corepresentation Ind(W1) ∈ L(C0(G) ⊗
Ind(C0(bG1))) of G is defined by the equation
in the sense of the identification F ∼= L2(G) ⊗ Ind(C0(bG1)). The isomorphism Φ defined above
intertwines the the strict ∗-homomorphism πl with the AdInd(W1) (see Proposition 3.7 and Section
4 of [16]):
where the leg numbering notation on the right hand side of this equation is to be understood
(ι ⊗ πl)W = W12 Ind(W1)13,
Φ∗(h ⊗ v) = v(h)
(8)
Φπl(x)Φ∗ = Ind(W1)(x ⊗ 1) Ind(W1)∗.
(9)
Let us write the corepresentation equation (∆G ⊗ ι) Ind(W1) = Ind(W1)13 Ind(W1)23 in terms of
W and Ind(W1):
(10)
Applying the character χg : C0(G) → C, g ∈ G to the first leg of the above equation we get the
following formula
23 = W12 Ind(W1)13
Ind(W1)23W12 Ind(W1)∗
Ind(W1)(Lg ⊗ 1) Ind(W1)∗ = Lg ⊗ Ind(W1)(g),
(11)
9
where we identified the induced corepresentation Ind(W1) ∈ L(C0(G) ⊗ Ind(C0(bG1))) with the
corresponding representation of the group G on K(Ind(C0(bG1))):
B(L2(G)) → L(F Ψ) we may define the induced C∗-module Ind(C0(bGΨ
ΦΨ : F Ψ → L2(G) ⊗ Ind(C0(bGΨ
The twisted objects are defined similarly. Starting with the strict ∗-homomorphism ΘΨ :
1 )) and the isomorphism
1 ) by the equation
12 Ind(W Ψ
Ind(W1)(g) = (χg ⊗ ι)(Ind(W1)).
1 )13
1 )). Defining Ind(W Ψ
l )W Ψ = W Ψ
1 ) by the formula:
(ι ⊗ πΨ
l , ΦΨ and Ind(W Ψ
we may relate πΨ
ΦΨπΨ
l (x)ΦΨ∗ = AdInd(W Ψ
1 )(x) = Ind(W Ψ
1 )(x ⊗ 1) Ind(W Ψ
1 )∗.
In order to compare Θ and ΘΨ let us define Ψ ∈ L(Ind(C0(bG1)) ⊗ L2(G)) by:
Ind(W1)12Ψ13 Ind(W1)∗
12 = Ψ13 Ψ23.
The existence of such Ψ follows from equation (11). Let us also introduce
Ψ′ = Φ∗Σ ΨΣΦ ∈ L(F )
(12)
(13)
where Σ is the flip operation on the tensor product.
Theorem 5.2. Let Θ : B(L2(G)) → L(F ) and ΘΨ : B(L2(G)) → L(F Ψ) be the strict ∗-
homomorphisms introduced above and let U ∈ L(F , F Ψ) be the unitary transformation introduced
in Proposition 5.1. Let Ψ′ ∈ L(F ) be the element introduced in Eq. (13). Then
ΘΨ(x) = U (Ad Ψ′ ◦Θ) (x)U ∗
(14)
1 )) given by:
for any x ∈ B(L2(G)).
1 )) are isomorphic, via the
In particular, Ind(C0(bG1)) and Ind(C0(bGΨ
isomorphism K : Ind(C0(bG1)) → Ind(C0(bGΨ
Ind(C0(bG1)) ∋ v 7→ K(v) = U Ψ′v ∈ Ind(C0(bGΨ
Proof. Since the algebras L∞(bGΨ)′ and L∞(GΨ)′ generate B(L2(G)), it is enough to check Eq.
(14) on them separately. It is easy to see that for any x ∈ L∞(bGΨ)′
which proves equality (14) on L∞(bGΨ)′. To prove it on L∞(GΨ)′ it is enough to see that (Ad Ψ′ ◦Θ⊗
Ind(C0(bG1)) by means of the isomorphism Φ (see (8)). Using (8) and (12) it is easy to show that
id) V Ψ = (U ∗ ⊗ id)Y Ψ(U ⊗ id). In the following computation we shall identify F with L2(G) ⊗
under the aforementioned identification we have:
(Ad Ψ′ ◦Θ)( Jx∗ J) = U ∗πr(x)U,
1 )).
(16)
(15)
(πl ⊗ id)Ψ = Ψ13 Ψ23.
(17)
We compute
(Ad Ψ′ ◦Θ ⊗ id)( V Ψ
13 ) = Ψ′
12Ψ∗
13
= Ψ′
12Ψ∗
13
= Ψ′
Ψ′∗
12
12
= (πl ⊗ id)(Ψ∗) V13(πr ⊗ id)(Ψ′∗) = Y Ψ,
V13Ψ′∗
Ψ′∗
13
12
V13 Ψ′∗
12Ψ′∗
13
Ψ∗
V13Ψ′∗
23Ψ∗
13
13
where in the third equality we used an easy to check formula:
V13 Ψ′∗
12
V ∗
13 = Ψ′∗
12
Ψ∗
23
and in the fourth equality we used (17).
(18)
(cid:3)
10
P. KASPRZAK
Let us now move on to the comparison of the induced corepresentations Ind(W Ψ
1 ) and Ind(W1).
In order to do that we introduce two elements Ψ, Ψ′ ∈ L(L2(G) ⊗ Ind(C0(bG1))) such that
(ι ⊗ πl)(Ψ) = Ψ12 Ψ13,
(ι ⊗ πl)(Ψ) = Ψ′
13.
12
Ψ′
(19)
(20)
Their existence follows from the bicharacter equation for Ψ and Eq. (11). Let us note that Ψ and
the element Ψ satisfying (17) are related by the formula
Ψ = Σ Ψ∗Σ.
(21)
We compute:
(ι ⊗ πΨ
l )(W Ψ) = Ψ23(ι ⊗ πl)(ΨW Ψ′) Ψ∗
23
Ψ′
13
Ψ∗
23
= Ψ23Ψ12 Ψ13W12 Ind(W1)13Ψ′
12
12 Ind(W1)13 Ψ′
= Ψ12W12 Ψ13Ψ′
13
Ψ13 Ind(W1)13 Ψ′
= Ψ12W12Ψ′
13
12
= W Ψ
12
Ψ13 Ind(W1)13 Ψ′
13,
where in the third equality we used
W12 Ψ13W ∗
12 Ind(W1)13 = Ψ′
12
12 = Ψ13 Ψ23,
23.
Ψ∗
Ind(W1)∗
13Ψ′
The first of these equalities follows from the bicharacter equation for Ψ whereas the second one
follows from the equality below:
Ind(W1)∗(Rg ⊗ 1) Ind(W1) = Rg ⊗ Ind(W1)(g).
This shows that with the identification Ind(C0(bG1)) ∼= Ind(C0(bGΨ
where the product on the right hand side is taken in the C∗-algebra L(L2(G) ⊗ Ind(C0(bG1)).
Finally, we shall compare the induced right coactions
1 )) of Theorem 5.2 we have
1 ) = Ψ Ind(W1) Ψ′,
Ind(W Ψ
Ind(β)Ψ : Ind(C0(bGΨ
1 )) ⊗ C0(bGΨ)),
Ind(β) : Ind(C0(bG1)) → M(Ind(C0(bG1)) ⊗ C0(bG)).
1 )) → M(Ind(C0(bGΨ
Equation (6.2) of [16] defines the induced coaction Ind(β) as
(ι ⊗ Ind(β))Φ(i ⊗
π bG1
Similarly, Ind(βΨ) is defined by:
(ι ⊗ Ind(βΨ))ΦΨ(i ⊗
π bGΨ
1
x) = W13(Φ ⊗ ι) (i ⊗ 1) ⊗
π bG1
⊗ι
13(ι ⊗ β)(x)! .
W ∗
x) = W Ψ
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
π bGΨ
1
⊗ι
13 (ι ⊗ βΨ)(x)! .
W Ψ∗
Combining (8), (15) and (21) one can easily check that
(ι ⊗ K) ΨΦ = ΦΨU.
(22)
Using this formula we compute:
(ι ⊗ Ind(βΨ))(ι ⊗ K) ΨΦ(i ⊗
π bG1
x)
= (ι ⊗ Ind(βΨ))ΦΨU (i ⊗
π bG1
x)
W Ψ∗
W Ψ∗
π bGΨ
1
⊗ι
π bGΨ
1
⊗ι
π bGΨ
1
⊗ι
π bGΨ
1
⊗ι
= W Ψ
= W Ψ
= W Ψ
Ψ′∗
13W ∗
Ψ′∗
13W ∗
13Ψ12Ψ∗
= W Ψ
13 (ι ⊗ βΨ)(Ψx)!
13 Ψ13Ψ12(ι ⊗ βΨ)(x)!
13Ψ12(ι ⊗ βΨ)(x)!
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι) (i ⊗ 1) ⊗
13(ΦΨ ⊗ ι)(U ⊗ id) (i ⊗ 1) ⊗
= Ψ13W13(ΦΨ ⊗ ι)(U ⊗ id) (i ⊗ 1) ⊗
= Ψ13W13(ι ⊗ K ⊗ ι)( Ψ ⊗ 1)(Φ ⊗ id) (i ⊗ 1) ⊗
23W13(Φ ⊗ id) (i ⊗ 1) ⊗
23(ι ⊗ β)(x)Ψ23!
13(ι ⊗ β)(x)Ψ23!
13(ι ⊗ β)(x)Ψ23!
13(ι ⊗ β)(x)Ψ23!
= (ι ⊗ K ⊗ ι)Ψ13 Ψ12 Ψ∗
13Ψ12W ∗
Ψ′∗
13W ∗
= W Ψ
= W Ψ
π bGΨ
1
⊗ι
π bG1
⊗ι
W ∗
π bG1
⊗ι
Ψ′∗
13(ι ⊗ β)(x)Ψ23!
13(ι ⊗ β)(x)! Ψ23
W ∗
W ∗
π bG1
⊗ι
π bG1
⊗ι
= (ι ⊗ K ⊗ ι)Ψ13 Ψ12 Ψ∗
23(ι ⊗ Ind(β))(Φ(ι ⊗
π bG1
x))Ψ23.
Using the equality
(ι ⊗ Ind(β))( Ψ) = Ψ13 Ψ12
we get
(ι ⊗ Ind(βΨ))(ι ⊗ K) ΨΦ i ⊗
π bG1
x!
23 (ι ⊗ Ind(β)) ΨΦ ι ⊗
π bG1
x!! Ψ23.
= (ι ⊗ K ⊗ ι) Ψ∗
Hence we see that
(K ∗ ⊗ ι) Ind(βΨ)K(v) = Ψ∗ Ind(β)(v)Ψ
for any v ∈ Ind(C0(bGΨ
1 )). In particular, for any x ∈ K(Ind(C0(bGΨ
(K ∗ ⊗ ι) Ind(βΨ)(KxK ∗)(K ⊗ ι) = Ψ∗ Ind(β)(x) Ψ.
1 ))) we have
11
(23)
In what follows we summarize the above considerations:
module over itself. Let W Ψ
Theorem 5.3. Let C0(bGΨ) be the C∗-algebra of the quantum group bGΨ treated as the C∗-Hilbert
1 ∈ M(C0(GΨ) ⊗ C0(bGΨ)) be the left regular corepresentation of GΨ
and βΨ be the right coaction of bGΨ defined in (6). Let K : Ind(C0(bGΨ
1 )) → Ind(C0(bG1)) be
12
P. KASPRZAK
The induced coaction Ind(W Ψ
the isomorphism introduced in Eq.
(15) (in what follows we shall identify Ind(C0(bGΨ
Ind(C0(bG1)) by means of K).
where Ψ, Ψ′ ∈ L(L2(G) ⊗ Ind(C0(bG1))) are the unitary elements defined by (19) and (20). The
1 ) ∈ L(L2(G) ⊗ Ind(C0(bG1))) is given by
1 ) = Ψ Ind(W1) Ψ′
twisted induced coaction
1 )) with
Ind(W Ψ
(24)
is related with its untwisted counterpart Ind(β) by the following formula
Ind(β)Ψ : Ind(C0(bGΨ
1 )) → M(Ind(C0(bG1)) ⊗ C0(bGΨ))
Ind(βΨ)(v) = Ψ∗ Ind(β)(v)Ψ
(25)
for any v ∈ Ind(C0(bG1)), where Ψ ∈ L(Ind(C0(bG1)) ⊗ C0(bGΨ)) is the unitary element defined by
(17).
6. Rieffel deformation of homogeneous spaces - the quotient case
Let G be a locally compact group, G1 its closed subgroup and let X = G / G1 be the homo-
geneous space of left G1-cosets. The standard action of G on X gives rise to the G-C∗-algebra
X = (C0(X), ∆X). Suppose that Γ ⊂ G1 is a closed abelian subgroup and let Ψ be a 2-cocycle on
the dual group Γ. Using the results of paper [8] described in Section 4 we know that X may be
deformed to GΨ-C∗-algebra XΨ.
On the other hand, applying the Rieffel deformation to the subgroup G1 ⊂ G we get a closed
1 be the C∗-algebraic
quantum subgroup GΨ
quotient space of Definition 2.6. The aim of this section is to show that GΨ/GΨ
1
1 ⊂ GΨ in the sense of Definition 2.5 of [16]. Let GΨ/GΨ
∼= XΨ.
Let ι ∈ Mor(C0(X), C0(G)) be the standard embedding - it maps a function f ∈ C0(X) to
the same function on G which is constant on the right G1-cosets. Let α : Γ → Aut(C0(X)) and
ρ : Γ2 → Aut(C0(G)) be the actions introduced in Section 4. The action ρ of Γ2 on C0(G)
restricts to the action α on C0(X). To be more precise, the second copy of Γ in Γ2 acts trivially
on C0(X) and the action of the first copy coincides with α. By the results of Section 3.2 of [9] the
embedding ι may be twisted to the embedding ιΨ ∈ Mor(C0(XΨ), C0(GΨ)). The comultiplication
∆GΨ restricts to the coaction ∆XΨ on ιΨ(C0(XΨ)). The last statement follows from Theorem 4.11
of [9] and the description of ∆XΨ given in Section 4. In particular, ∆XΨ is implemented by the
multiplicative unitary:
∆XΨ (a) = W Ψ∗(1 ⊗ a)W Ψ,
(26)
for any a ∈ C0(XΨ).
1 ) by the coaction of GΨ
1 )))
As was explained in Remark 2.7, the crossed product of C0(GΨ/GΨ
that the induced corepresentation Ind(W Ψ
1 )))
may be described as follows. Using Eq. (25) one can see that the dual coaction Ind(βΨ) : B →
1 ))). Using Theorem 5.3 we see that we may identify K(Ind(C0(bGΨ
coincides with K(Ind(C0(bGΨ
with K(Ind(C0(bG1))), which in turn may be identified with the C∗-algebra G⋉C0(G/G1). The for-
mer C∗-algebra may be identified with B = [C0(bG) C0(X)] ⊂ B(L2(G)) - the C∗-algebra generated
by C0(bG) and C0(X) inside B(L2(G)).
1 ))) = B the crossed product structure of K(Ind(C0(bGΨ
Under the identification K(Ind(C0(bGΨ
M(B ⊗ C0(bGΨ)) is implemented by the unitary bV Ψ introduced in (5). Eq. (24) shows in turn
corepresentation W Ψ ∈ M(C0(GΨ) ⊗ C0(bGΨ)) ⊂ M(C0(GΨ) ⊗ B). We are now well prepared to
1 ) ∈ M(C0(GΨ) ⊗ B) can be identified with the regular
Theorem 6.1. Let XΨ be the Rieffel deformation of the homogeneous space X and GΨ/GΨ
Vaes' quotient considered above. Then XΨ ∼= GΨ/GΨ
1 .
Proof. By the universal properties of the crossed product C∗-algebra Γ ⋉ρ C0(X) there exists a
unique morphism ιΓ ∈ Mor(Γ ⋉ρ C0(X), B) which is identity on C0(X) ⊂ M(Γ ⋉ρ C0(X)) and such
that it sends the unitary generator uγ ∈ M(C∗(Γ)) to the left shift Lγ ∈ M(B). Using Theorem 3.6,
[9] we may conclude that the restriction of ιΓ to C0(XΨ) gives rise to the injective morphism ιC0(XΨ)
1 the
prove
13
which we shall denote by ιΨ ∈ Mor(C0(XΨ), B). Let ρΨ be the twisted dual action on Γ ⋉ρ C0(X).
In what follows we shall interpret it as a coaction ρΨ ∈ Mor(Γ ⋉ρ C0(X), Γ ⋉ρ C0(X) ⊗ C∗(Γ)).
Under this interpretation, the relation of ρΨ with its untwisted counterpart ρ is established by the
following formula:
for any x ∈ Γ ⋉ρ C0(X). Let us note that ιΓ intertwines the twisted dual coactions:
Ind(βΨ) ◦ ιΓ = (ιΓ ⊗ id) ◦ ρΨ.
ρΨ(x) = Ψ∗ρ(x)Ψ
(27)
(28)
In order to see it we have to observe that:
• ιΓ intertwines ρ and Ind(β):
Ind(β) ◦ ιΓ = (ιΓ ⊗ id) ◦ ρ.
(29)
Indeed, the equality Ind(β)(ιΓ(f )) = (ιΓ ⊗ id)(ρ(f )) for any f ∈ C0(X) is the consequence
of the the simultaneous invariance of f under ρ and Ind(β). Moreover Ind(β) ◦ ιΓ(uγ) =
Lγ ⊗ Lγ = (ιΓ ⊗ id) ◦ ρ(uγ) for any γ ∈ Γ. Using the fact C0(X) and C∗(Γ) generate
Γ ⋉ρ C0(X) we get a proof of (29).
• Let Ψ be the unitary element introduced in (12). Note that (ιΓ ⊗ id)Ψ = Ψ.
Using the above observations and Eqs. (25), (27) one easily gets (28) which in turn implies that
ιΨ(C0(XΨ)) ⊂ M(B)Ind(βΨ). The proof of the equality ιΨ(C0(XΨ)) = C0(GΨ/GΨ
1 ) is based on
Theorem 6.7 of [16]. Its application requires the check of the following two conditions:
• The map x 7→ W Ψ∗(1 ⊗ x)W Ψ defines a continuous coaction on ιΨ(C0(XΨ)). Indeed, this
follows from (26).
• We have [ιΨ(C0(XΨ)) C0(bG)] = B. In order to see that we compute
[ιΨ(C0(XΨ)) C0(bG)] = [ιΨ(C0(XΨ)) C∗(Γ) C0(bG)] = [ιΓ(Γ ⋉ρ C0(XΨ)) C0(bG)]
= [ιΓ(Γ ⋉ρ C0(X)) C0(bG)] = [ιΓ(C0(X)) C∗(Γ) C0(bG)]
= [C0(X) C0(bG)] = B
where in the first and the fourth equality we used the fact that [C∗(Γ) C0(bG)] = C0(bG)
and in the third equality we used the fact that Γ ⋉ρ C0(XΨ) = Γ ⋉ρ C0(X) (note that the
action of Γ on C0(X) and on C0(XΨ) is denoted by the same ρ).
This ends the proof of the isomorphism XΨ ∼= GΨ/GΨ
1 .
(cid:3)
7. Rieffel deformation of G-simple C∗-algebras
Let G be a locally compact quantum group corresponding to a locally compact group G. The
aim of this section is to prove that the Rieffel deformation XΨ of a G-simple C∗-algebra X is GΨ-
simple (see Definition 2.4). In particular, the Rieffel deformation of a quotient space X = G / G1
(also in the case when Γ is a subgroup of G but not of G1) is GΨ-simple. The idea of the following
proof is based in the functorial properties of the Rieffel deformation (see Section 3.2, [9]).
Theorem 7.1. Let X be a G-C∗-algebra. Let Γ be a closed abelian subgroup of G and Ψ a 2-cocycle
on the dual group Γ. The Rieffel deformation XΨ of X is GΨ-simple.
Proof. Let B be a GΨ-C∗-algebra and let π ∈ MorGΨ (XΨ, B). The Rieffel deformation ΓΨ of Γ is
a quantum closed subgroup of GΨ. Actually, ΓΨ = Γ which easily follows from the abelianity of
Γ (see Appendix B of [7]). Let α : Γ → Aut(C0(XΨ)) be the action that corresponds to the Γ-
restriction of the coaction ∆XΨ . Similarly, we may introduce β : Γ → Aut(C0(B)). Obviously the
morphism π intertwines α and β: π◦αγ = βγ ◦π for any γ ∈ Γ. The deformation data (C0(B), β, ¯Ψ)
¯Ψ) and the deformed morphism
gives rise to the deformed C∗-algebra that we shall denote by C0(B
¯Ψ)). Note that (GΨ) ¯Ψ = G. Using the ideas presented in Section 4 we
π ¯Ψ ∈ Mor(C0(X), C0(B
may construct the twisted coaction ∆ ¯Ψ
¯Ψ and check that π ¯Ψ is a G-morphism. The
G-simplicity of X implies that ker π ¯Ψ = {0}. Using Proposition 3.8 [9] we see that ker π = {0},
which ends the proof of GΨ-simplicity of XΨ .
(cid:3)
B of G on B
14
P. KASPRZAK
References
[1] S. Baaj, G. Skandalis: Unitaires multiplicatifs et dualit´e pour les produits crois´es de C-algebres, Ann. Sci.
Ec. Nor. Sup. (4) 26 (1993), 425 -- 488.
[2] K. de Commer: Galois objects and cocycle twisting for locally compact quantum groups, to appear in
J. Op. Theory.
[3] M. Brin, G. Stuck: Introduction to Dynamical Systems, Camb. Univ. (2003)
[4] M. Enock, L. Vainerman: Deformation of a Kac algebra by an abelian subgroup, Commun. Math. Phys.
178 (1996), no. 3, p. 571 -- 596.
[5] P. Fima, L. Vainerman: Twisting and Rieffel's deformation of locally compact quantum groups. Deformation
of the Haar measure. Commun. Math. Phys. 286 (2009), no. 3, p. 1011 -- 1050.
[6] S. Kaliszewski, J. Quigg: Categorical Landstad duality for actions, Indiana Univ. Math. J. 58 (2009), no. 1,
p. 415 -- 441.
[7] P. Kasprzak: Heisenberg-Lorentz Quantum Group, to appear in J. Noncom. Geom.
[8] P. Kasprzak: Rieffel deformation of groups coactions, to appear in Commun. Math. Phys.
[9] P. Kasprzak: Rieffel deformation via crossed products, J. Funct. Anal. 257 (2009), no. 5, p. 1288 -- 1332
[10] A. Kleppner: Multipliers on Abelian Groups, Math. Annalen 158, 11 -- 34 (1965).
[11] J. Kustermans, S. Vaes: Locally compact quantum groups. Ann. Sci. Ec. Norm. Sup. 33 (2000), no. 4, p.
837 -- 934.
[12] T. Masuda, Y. Nakagami, S.L. Woronowicz: A C∗-algebraic framework for quantum groups, Inter-
nat. J. Math. 14 (2003), no. 9, p. 903 -- 1001.
[13] P. Podle´s Quantum spheres, Lett. Math. Phys. 14 (1987), no. 3, p. 193 -- 202.
[14] P. Podle´s Symmetries of Quantum Spaces. Subgroups and Quotient Spaces of Quantum SU (2) and SO(3)
Groups. Commun. Math. Phys. 170 (1995), no. 1, p. 1 -- 20.
[15] M.A. Rieffel: Deformation quantization for action of Rd, Mem. Am. Math. Soc. 106 (1993), no. 506.
[16] S. Vaes: A new approach to induction and imprimitivity results, J. Funct. Anal. 229 (2005), no 2, p. 317 --
374.
[17] J. C. Varilly: Quantum symmetry groups of noncommutative spheres, Commun. Math. Phys. 221 (2001),
no. 3, p. 511 -- 523.
[18] S.L. Woronowicz: C∗-algebras generated by unbounded elements, Rev. Math. Phys. 7, No. 3, (1995), 481
-- 521.
Department of Mathematical Sciences, University of Copenhagen
On leave from Department of Mathematical Methods in Physics, Faculty of Physics, Warsaw Uni-
versity
E-mail address: [email protected]
|
1512.06677 | 2 | 1512 | 2016-07-29T16:37:07 | A class of II_1 factors with exactly two group measure space decompositions | [
"math.OA"
] | We construct the first II_1 factors having exactly two group measure space decompositions up to unitary conjugacy. Also, for every positive integer $n$, we construct a II_1 factor $M$ that has exactly $n$ group measure space decompositions up to conjugacy by an automorphism. | math.OA | math |
A class of II1 factors with exactly two
group measure space decompositions
by Anna Sofie Krogager and Stefaan Vaes
Abstract
We construct the first II1 factors having exactly two group measure space decompositions
up to unitary conjugacy. Also, for every positive integer n, we construct a II1 factor M that
has exactly n group measure space decompositions up to conjugacy by an automorphism.
1
Introduction
The group measure space construction of Murray and von Neumann associates to every free
ergodic probability measure preserving (pmp) group action Γ y (X, µ) a crossed product von
Neumann algebra L∞(X) ⋊ Γ. The classification of these group measure space II1 factors is one
of the core problems in operator algebras. For Γ infinite amenable, they are all isomorphic to
the hyperfinite II1 factor R, because by Connes' celebrated theorem [Co75], even all amenable
II1 factors are isomorphic with R.
For nonamenable groups Γ, rigidity phenomena appear and far reaching classification theorems
for group measure space II1 factors were established in Popa's deformation/rigidity theory, see
e.g. [Po01, Po03, Po04]. In these results, the subalgebra A = L∞(X) of M = A ⋊ Γ plays a
special role. Indeed, if an isomorphism π : A ⋊ Γ → B ⋊ Λ of group measure space II1 factors
satisfies π(A) = B, by [Si55], π must come from an orbit equivalence between the group actions
Γ y A and Λ y B, so that methods from measured group theory can be used. The subalgebra
it is maximal abelian and the normalizer NM (A) = {u ∈ U (M ) uAu∗}
A ⊂ M is Cartan:
generates M . One of the key results of [Po01, Po03, Po04] is that for large classes of group
actions, any isomorphism π : A ⋊ Γ → B ⋊ Λ satisfies π(A) = B, up to unitary conjugacy.
In [Pe09, PV09], the most extreme form of rigidity, called W∗-superrigidity was discovered: in
certain cases, the crossed product II1 factor M = A ⋊ Γ entirely remembers Γ and its action
on A. These results were established by first proving that the II1 factor M has a unique group
measure space Cartan subalgebra up to unitary conjugacy and then proving that the group
action Γ y A is OE-superrigid. Since then, several uniqueness results for group measure
space Cartan subalgebras were obtained, in particular [Io10] proving that all Bernoulli actions
Γ y (X0, µ0)Γ of all infinite property (T) groups are W∗-superrigid.
Note that a general Cartan subalgebra A ⊂ M need not be of group measure space type, i.e.
there need not exist a group Γ complementing A in such a way that M = A ⋊ Γ. This is closely
related to the phenomenon that a countable pmp equivalence relation need not be the orbit
equivalence relation of a group action that is free. The first actual uniqueness theorems for
Cartan subalgebras up to unitary conjugacy were only obtained in [OP07], where it was proved
in particular that A is the unique Cartan subalgebra of A ⋊ Γ whenever Γ = Fn is a free group
and Fn y A is a free ergodic pmp action that is profinite. More recently, in [PV11], it was
shown that A is the unique Cartan subalgebra of A ⋊ Γ for arbitrary free ergodic pmp actions
of the free groups Γ = Fn.
KU Leuven, Department of Mathematics, Leuven (Belgium).
E-mails: [email protected] and [email protected].
Supported by European Research Council Consolidator Grant 614195, and by long term structural funding --
Methusalem grant of the Flemish Government.
1
It is known since [CJ82] that a II1 factor M may have two Cartan subalgebras A, B ⊂ M
that are non conjugate by an automorphism of M . Although several concrete examples of this
phenomenon were given in [OP08, PV09, SV11] and despite all the progress on uniqueness of
Cartan subalgebras, there are so far no results describing all Cartan subalgebras of a II1 factor
M once uniqueness fails. In this paper, we prove such a result and the following is our main
theorem.
Theorem A. (1) For every integer n ≥ 0, there exist II1 factors M that have exactly 2n group
measure space Cartan subalgebras up to unitary conjugacy.
(2) For every integer n ≥ 1, there exist II1 factors M that have exactly n group measure space
Cartan subalgebras up to conjugacy by an automorphism of M .
Two free ergodic pmp actions are called W∗-equivalent if they have isomorphic crossed product
von Neumann algebras. Thus, a free ergodic pmp action G y (X, µ) is W∗-superrigid if every
action that is W∗-equivalent to G y (X, µ) must be conjugate to G y (X, µ). Theorem A(2)
can then be rephrased in the following way: we construct free ergodic pmp actions G y (X, µ)
with the property that G y (X, µ) is W∗-equivalent to exactly n group actions, up to orbit
equivalence of the actions (and actually also up to conjugacy of the actions, see Theorem 6.3).
The II1 factors M in Theorem A are concretely constructed as follows. Let Γ be any torsion
free nonelementary hyperbolic group and let β : Γ y (A0, τ0) be any trace preserving action
on the amenable von Neumann algebra (A0, τ0) with A0 6= C1 and Ker β 6= {e}. We then
define (A, τ ) = (A0, τ0)Γ and consider the action σ : Γ × Γ y (A, τ ) given by σ(g,h)(πk(a)) =
πgkh−1(βh(a)) for all g, h, k ∈ Γ and a ∈ A0, where πk : A0 → A denotes the embedding as the
k'th tensor factor.
Our main result describes exactly all group measure space Cartan subalgebras of the crossed
product M = AΓ
0 ⋊ (Γ × Γ).
Theorem B. Let M = AΓ
0 ⋊ (Γ × Γ) be as above. Up to unitary conjugacy, all group measure
space Cartan subalgebras B ⊂ M are of the form B = BΓ
0 where B0 ⊂ A0 is a group measure
space Cartan subalgebra of A0 with the following two properties: βg(B0) = B0 for all g ∈ Γ and
A0 can be decomposed as A0 = B0 ⋊ Λ0 with βg(Λ0) = Λ0 for all g ∈ Γ.
In Section 5, we actually prove a more general and more precise result, see Theorem 5.1. In
Section 6, we give concrete examples and computations, thus proving Theorem A (see Theorem
6.3).
Note that in Theorems A and B, we can only describe the group measure space Cartan subal-
gebras of M . The reason for this is that our method entirely relies on a technique of [PV09],
using the so-called dual coaction that is associated to a group measure space decomposition
M = B ⋊ Λ, i.e. the normal ∗-homomorphism ∆ : M → M ⊗ M given by ∆(bvs) = bvs ⊗ vs
for all b ∈ B, s ∈ Λ. When B ⊂ M is an arbitrary Cartan subalgebra, we do not have such a
structural ∗-homomorphism.
Given a II1 factor M as in Theorem B and given the dual coaction ∆ : M → M ⊗ M associated
with an arbitrary group measure space decomposition M = B ⋊ Λ, Popa's key methods of
malleability [Po03] and spectral gap rigidity [Po06] for Bernoulli actions allow to prove that
∆(L(Γ×Γ)) can be unitarily conjugated into M ⊗L(Γ×Γ). An ultrapower technique of [Io11], in
combination with the transfer-of-rigidity principle of [PV09], then shows that the "mysterious"
group Λ must contain two commuting nonamenable subgroups Λ1, Λ2. Note here that the
same combination of [Io11] and [PV09] was recently used in [CdSS15] to prove that if Γ1, Γ2
are nonelementary hyperbolic groups and L(Γ1 × Γ2) ∼= L(Λ), then Λ must be a direct product
of two nonamenable groups. Once we know that Λ contains two commuting nonamenable
2
subgroups Λ1, Λ2, we use a combination of methods from [Io10] and [IPV10] to prove that
Λ1Λ2 ⊂ Γ × Γ. From that point on, it is not so hard any more to entirely unravel the structure
of B and Λ. Throughout these arguments, we repeatedly use the crucial dichotomy theorem
of [PV11, PV12] saying that hyperbolic groups Γ are relatively strongly solid:
in arbitrary
tracial crossed products M = P ⋊ Γ, if a von Neumann subalgebra Q ⊂ M is amenable relative
to P , then either Q embeds into P , or the normalizer of Q stays amenable relative to P .
2 Preliminaries
2.1 Popa's intertwining-by-bimodules
We recall from [Po03, Theorem 2.1 and Corollary 2.3] Popa's method of intertwining-by-
bimodules. When (M, τ ) is a tracial von Neumann algebra and P, Q ⊂ M are possibly non-
unital von Neumann subalgebras, we write P ≺M Q if there exists a nonzero P -Q-subbimodule
of 1P L2(M )1Q that has finite right Q-dimension. We refer to [Po03] for several equivalent for-
mulations of this intertwining property. If P p ≺M Q for all nonzero projections p ∈ P ′∩1P M 1P ,
we write P ≺f
M Q.
We are particularly interested in the case where M is a crossed product M = A ⋊ Γ by a
trace preserving action Γ y (A, τ ). Given a subset F ⊂ Γ, we denote by PF the orthogonal
projection of L2(M ) onto the closed linear span of {aug a ∈ A, g ∈ F }, where {ug}g∈Γ denote
the canonical unitaries in L(Γ). By [Va10, Lemma 2.5], a von Neumann subalgebra P ⊂ pM p
satisfies P ≺f
M A if and only if for every ε > 0, there exists a finite subset F ⊂ Γ such that
kx − PF (x)k2 ≤ kxkε
for all x ∈ P .
We also need the following elementary lemma.
Lemma 2.1. Let Γ y (A, τ ) be a trace preserving action and put M = A ⋊ Γ. If P ⊂ M is a
diffuse von Neumann subalgebra such that P ≺f A, then P ⊀ L(Γ).
Proof. Let ε > 0 be given. We have P ≺f A. So, as explained above, we can take a finite
set F ⊂ Γ such that ku − PF (u)k2 ≤ ε
2 for all u ∈ U (P ). Moreover, since P is diffuse, we
can choose a sequence of unitaries (wn) ⊂ U (P ) tending to 0 weakly. We will prove that
kEL(Γ)(xwny)k2 → 0 for all x, y ∈ M , meaning that P ⊀ L(Γ). Note that it suffices to consider
x, y ∈ (A)1.
Take x, y ∈ (A)1. Write wn = Pg∈Γ(wn)gug with (wn)g ∈ A. Then,
kEL(Γ)(PF (xwny))k2
τ (x(wn)gσg(y))2 → 0
2 = X
g∈F
since wn → 0 weakly. Take n0 large enough such that
kEL(Γ)(PF (xwny))k2 ≤
ε
2
for all n ≥ n0 .
Since PF is A-A-bimodular, we have that
kEL(Γ)(PF (xwny)) − EL(Γ)(xwny)k2 ≤ kPF (xwny) − xwnyk2 = kx(PF (wn) − wn)yk2 ≤
ε
2
for all n. We conclude that kEL(Γ)(xwny)k2 ≤ ε for all n ≥ n0.
3
2.2 Relative amenability
A tracial von Neumann algebra (M, τ ) is called amenable if there exists a state ϕ on B(L2(M ))
such that ϕM = τ and such that ϕ is M -central, meaning that ϕ(xT ) = ϕ(T x) for all x ∈ M ,
T ∈ B(L2(M )). In [OP07, Section 2.2], the concept of relative amenability was introduced.
The definition makes use of Jones' basic construction: given a tracial von Neumann algebra
(M, τ ) and a von Neumann subalgebra P ⊂ M , the basic construction hM, eP i is defined as
the commutant of the right P -action on B(L2(M )). Following [OP07, Definition 2.2], we say
that a von Neumann subalgebra Q ⊂ pM p is amenable relative to P inside M if there exists a
positive functional ϕ on phM, eP ip that is Q-central and satisfies ϕpM p = τ .
We say that Q is strongly nonamenable relative to P if Qq is nonamenable relative to P for every
nonzero projection q ∈ Q′ ∩ pM p. Note that in that case, also qQq is strongly nonamenable
relative to P for all nonzero projections q ∈ Q.
We need the following elementary relationship between relative amenability and intertwining-
by-bimodules.
Proposition 2.2. Let (M, τ ) be a tracial von Neumann algebra and Q, P1, P2 ⊂ M be von
Neumann subalgebras with P1 ⊂ P2. Assume that Q is strongly nonamenable relative to P1.
Then the following holds.
(1) If Q ≺ P2, there exist projections q ∈ Q, p ∈ P2, a nonzero partial isometry v ∈ qM p and a
normal unital ∗-homomorphism θ : qQq → pP2p such that xv = vθ(x) for all x ∈ qQq and
such that, inside P2, we have that θ(qQq) is nonamenable relative to P1.
(2) We have Q 6≺ P1.
Proof. (1) Assume that Q ≺ P2. By [Po03, Theorem 2.1], we can take projections q ∈ Q, p ∈ P2,
a nonzero partial isometry v ∈ qM p and a normal unital ∗-homomorphism θ : qQq → pP2p
such that xv = vθ(x) for all x ∈ qQq. Assume that θ(qQq) is amenable relative to P1 inside
P2. We can then take a positive functional ϕ on phP2, eP1ip that is θ(qQq)-central and satisfies
ϕpP2p = τ . Denote by eP2 the orthogonal projection of L2(M ) onto L2(P2). Observe that
eP2hM, eP1ieP2 = hP2, eP1i. We can then define the positive functional ω on qhM, eP1iq given
by
ω(T ) = ϕ(eP2 v∗T veP2)
for all T ∈ qhM, eP1 iq .
By construction, ω is qQq-central and ω(x) = τ (v∗xv) for all x ∈ qM q. Writing q0 = vv∗, we
have q0 ∈ (Q′ ∩ M )q and it follows that qQqq0 is amenable relative to P1. This contradicts the
strong nonamenability of Q relative to P1.
Finally, note that (2) follows from (1) by taking P1 = P2.
2.3 Properties of the dual coaction
Let M = B ⋊ Λ be any tracial crossed product von Neumann algebra. Denote by {vs}s∈Λ
the canonical unitaries and consider the dual coaction ∆ : M → M ⊗ M , i.e. the normal
∗-homomorphism defined by ∆(b) = b ⊗ 1 for all b ∈ B and ∆(vs) = vs ⊗ vs for all s ∈ Λ.
First, we show the following elementary lemma.
Lemma 2.3. A von Neumann subalgebra A ⊂ B ⋊ Λ satisfies ∆(A) ⊂ A ⊗ A if and only if
A = B0 ⋊ Λ0 for some von Neumann subalgebra B0 ⊂ B and some subgroup Λ0 < Λ that leaves
B0 globally invariant.
4
Proof. Let A ⊂ B ⋊ Λ be a von Neumann subalgebra satisfying ∆(A) ⊂ A ⊗ A. Let a ∈ A
and write a = Ps∈Λ asvs with as ∈ B. Let I = {s ∈ Λ as 6= 0}. Fix s ∈ I and define the
normal linear functional ω on B ⋊ Λ by ω(x) = τ (xv∗
2 vs. Since
∆(a) ∈ A ⊗ A, it follows that vs ∈ A. Similarly, we define a linear functional ρ on B ⋊ Λ by
ρ(x) = τ (xv∗
s ). Then (1 ⊗ ρ)∆(a) = asvs ∈ A and it follows that as ∈ A. Since this holds for
arbitrary s ∈ I, we conclude that A = B0⋊Λ0 where B0 = A∩B and Λ0 = {s ∈ Λ vs ∈ A}.
s a∗
s). Then (ω ⊗ 1)∆(a) = kask2
The proof of the next result is almost identical to the proof of [IPV10, Lemma 7.2(4)]. For the
convenience of the reader, we provide some details.
Proposition 2.4. Assume that (B, τ ) is amenable. If Q ⊂ M is a von Neumann subalgebra
without amenable direct summand, then ∆(Q) is strongly nonamenable relative to M ⊗ 1.
Proof. Using the bimodule characterization of relative amenability (see [OP07, Theorem 2.1]
and see also [PV11, Proposition 2.4]), it suffices to prove that the (M ⊗ M )-M -bimodule
M ⊗M(cid:0)L2(M ⊗ M ) ⊗
M ⊗1
L2(M ⊗ M )(cid:1)∆(M )
is weakly contained in the coarse (M ⊗ M )-M -bimodule. Denoting by σ : M ⊗ M → M ⊗ M
the flip automorphism, this bimodule is isomorphic with the (M ⊗ M )-M -bimodule
M ⊗M L2(M ⊗ M ⊗ M )(id⊗σ)(∆(M )⊗1) .
So, it suffices to prove that the M -M -bimodule M ⊗1L2(M ⊗ M )∆(M ) is weakly contained in
the coarse M -M -bimodule. Noting that this M -M -bimodule is isomorphic with a multiple of
the M -M -bimodule M(cid:0)L2(M ) ⊗B L2(M )(cid:1)M , the result follows from the amenability of B.
2.4 Spectral gap rigidity for co-induced actions
Given a tracial von Neumann algebra (A0, τ0) and a countable set I, we denote by (A0, τ0)I
(or just AI
by πi : A0 → AI
0) the von Neumann algebra tensor product NI (A0, τ0). For each i ∈ I, we denote
0 the embedding of A0 as the i'th tensor factor.
Definition 2.5. Let Γ y I be an action of a countable group Γ on a countable set I. We say
that a trace preserving action σ : Γ y (A0, τ0)I is built over Γ y I if it satisfies
σg(πi(A0)) = πg·i(A0)
for all g ∈ Γ, i ∈ I .
Assume that Γ y AI
0 is an action built over Γ y I. Choose a subset J ⊂ I that contains exactly
one point in every orbit of Γ y I. For every j ∈ J, the group Stab j globally preserves πj(A0).
This defines an action Stab j y A0 that can be co-induced to an action Γ y AΓ/ Stab i
. The
original action Γ y AI
0 is conjugate with the direct product of all these co-induced actions. In
particular, co-induced actions are exactly actions built over a transitive action Γ y I = Γ/Γ0.
0
Popa's malleability [Po03] and spectral gap rigidity [Po06] apply to actions built over Γ y I.
The generalization provided in [BV14, Theorems 3.1 and 3.3] carries over verbatim and this
gives the following result.
Theorem 2.6. Let Γ y I be an action of an icc group on a countable set. Assume that
Stab{i, j} is amenable for all i, j ∈ I with i 6= j. Let (A0, τ0) be a tracial von Neumann algebra
and (N, τ ) a II1 factor. Let Γ y (A0, τ0)I be an action built over Γ y I and put M = AI
0 ⋊ Γ.
If P ⊂ N ⊗ M is a von Neumann subalgebra that is strongly nonamenable relative to N ⊗ AI
0,
then the relative commutant Q := P ′ ∩ N ⊗ M satisfies at least one of the following properties:
5
(1) there exists an i ∈ I such that Q ≺ N ⊗ (AI
(2) there exists a unitary v ∈ N ⊗ M such that v∗Qv ⊂ N ⊗ L(Γ).
0 ⋊ Stab i) ;
Recall that for a von Neumann subalgebra P ⊂ M , we define QN M (P ) ⊂ M as the set of
elements x ∈ M for which there exist x1, . . . , xn, y1, . . . , ym ∈ M satisfying
xP ⊂
nX
i=1
P xi
and P x ⊂
mX
j=1
yjP .
Then QN M (P ) is a ∗-subalgebra of M containing P .
normalizer of P inside M .
Its weak closure is called the quasi-
The following lemma is proved in exactly the same way as [IPV10, Lemma 4.1] and goes back
to [Po03, Theorem 3.1].
Lemma 2.7. Let Γ y I be an action. Let A0 be a tracial von Neumann algebra and let Γ y AI
0
be an action built over Γ y I. Consider the crossed product M = AI
0 ⋊ Γ and let (N, τ ) be an
arbitrary tracial von Neumann algebra.
(1) If P ⊂ p(N ⊗ L(Γ))p is a von Neumann subalgebra such that P ⊀N ⊗L(Γ) N ⊗ L(Stab i) for
all i ∈ I, then the quasi-normalizer of P inside p(N ⊗ M )p is contained in p(N ⊗ L(Γ))p.
0 ⋊Stab i))q is a von Neumann subalgebra such that
0 ⋊ Stab{i, j}). Then the quasi-normalizer
(2) Fix i ∈ I and assume that Q ⊂ q(N ⊗(AI
0⋊Stab i) N ⊗ (AI
for all j 6= i, we have Q ⊀N ⊗(AI
of Q inside q(N ⊗ M )q is contained in q(N ⊗ (AI
0 ⋊ Stab i))q.
3 Transfer of rigidity
Fix a trace preserving action Λ y (B, τ ) and put M = B ⋊ Λ. We denote by {vs}s∈Λ the
canonical unitaries in L(Λ). Whenever G is a family of subgroups of Λ, we say that a subset
F ⊂ Λ is small relative to G if F is contained in a finite union of subsets of the form gΣh where
g, h ∈ Λ and Σ ∈ G.
Following the transfer of rigidity principle from [PV09, Section 3], we prove the following
theorem.
Theorem 3.1. Let Λ y (B, τ ) be a trace preserving action and put M = B ⋊ Λ. Let ∆ : M →
M ⊗ M be the dual coaction given by ∆(bvs) = bvs ⊗ vs for b ∈ B, s ∈ Λ. Let G be a family of
subgroups of Λ. Let P, Q ⊂ M be two von Neumann subalgebras satisfying
(1) ∆(P ) ≺M ⊗M M ⊗ Q,
(2) P ⊀M B ⋊ Σ for all Σ ∈ G.
Then there exists a finite set x1, . . . , xn ∈ M and a δ > 0 such that the following holds: whenever
F ⊂ Λ is small relative to G, there exists an element sF ∈ Λ − F such that
nX
i,k=1
kEQ (xivsF x∗
k) k2
2 ≥ δ .
Proof. Since ∆(P ) ≺M ⊗M M ⊗ Q, we can find a finite set x1, . . . , xn ∈ M and ρ > 0 such that
nX
i,k=1
kEM ⊗Q ((1 ⊗ xi)∆(w)(1 ⊗ x∗
k)) k2
2 ≥ ρ for all w ∈ U (P ) .
6
Given a subset F ⊂ Λ, we denote by PF the orthogonal projection of L2(M ) onto the closed
linear span of {bvs b ∈ B, s ∈ F }. Since P ⊀M B ⋊ Σ for all Σ ∈ G, it follows from [Va10,
Lemma 2.4] that there exists a net of unitaries {wj}j∈J ⊂ U (P ) such that kPF (wj)k2 → 0 for
s ∈ B
any set F ⊂ Λ that is small relative to G. For each j ∈ J, write wj = Ps∈Λ wj
svs with wj
and compute
ρ ≤
nX
i,k=1
kEM ⊗Q ((1 ⊗ xi)∆(wj)(1 ⊗ x∗
k)) k2
2 =
=
i,k=1
nX
nX
i,k=1
wj
svs ⊗ EQ(xivsx∗
k)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2
2
kwj
sk2
2 kEQ(xivsx∗
k)k2
2 .
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
X
X
s∈Λ
s∈Λ
We now claim that the conclusion of the theorem holds with δ = ρ
2 .
contradiction that there exists a subset F ⊂ Λ that is small relative to G such that
Indeed, assume for
nX
i,k=1
kEQ (xivsx∗
k) k2
2 < δ
for all s ∈ Λ − F .
Put K = max{kxik2kx∗
kk2
2 i, k = 1, . . . , n} and choose j0 ∈ J such that kPF (wj)k2
2 =
Ps∈F kwj
sk2
2 < ρ
4Kn2 for all j ≥ j0. We then get for j ≥ j0
nX
ρ ≤ X
2 ≤ n2KX
2 kEQ(xivsx∗
k)k2
kwj
kwj
sk2
s∈Λ
i,k=1
s∈F
2 + X
sk2
s∈Λ−F
kwj
sk2
2 δ <
ρ
4
+
ρ
2
,
which is a contradiction.
4 Embeddings of group von Neumann algebras
Following [Io10, Section 4] and [IPV10, Section 3], we define the height hΓ of an element in a
group von Neumann algebra L(Γ) as the absolute value of the largest Fourier coefficient, i.e.,
hΓ(x) = sup
g∈Γ
τ (xu∗
g)
for x ∈ L(Γ) .
Whenever G ⊂ L(Γ), we write
hΓ(G) = inf{hΓ(x) x ∈ G} .
When Γ is an icc group and Λ is a countable group such that L(Λ) = L(Γ) with hΓ(Λ) > 0, it
was proven in [IPV10, Theorem 3.1] that there exists a unitary u ∈ L(Γ) such that uTΛu∗ = TΓ.
We need the following generalization. For this, recall that a unitary representation is said to
be weakly mixing if {0} is the only finite dimensional subrepresentation.
Theorem 4.1. Let Γ be a countable group and p ∈ L(Γ) a projection. Assume that G ⊂
U (pL(Γ)p) is a subgroup satisfying the following properties.
(1) The unitary representation {Ad v}v∈G on L2(pL(Γ)p ⊖ Cp) is weakly mixing.
(2) If g ∈ Γ and g 6= e, then G′′ 6≺ L(CΓ(g)).
(3) We have hΓ(G) > 0.
Then p = 1 and there exists a unitary u ∈ L(Γ) such that uGu∗ ⊂ TΓ.
7
Proof. Write M = L(Γ) and denote by ∆ : M → M ⊗M : ∆(ug) = ug ⊗ug the comultiplication
on L(Γ). We first prove that the unitary representation on L2(∆(p)(M ⊗M )∆(p)⊖∆(Cp)) given
by {Ad ∆(v)}v∈G is weakly mixing. To prove this, assume that H ⊂ L2(∆(p)(M ⊗ M )∆(p))
is a finite dimensional subspace satisfying ∆(v)H∆(v∗) = H for all v ∈ G. Writing P =
G′′, it follows that the closed linear span of H∆(pM p) is a ∆(P )-∆(pM p)-subbimodule of
L2(∆(p)(M ⊗ M )∆(p)) that has finite right dimension. By [IPV10, Proposition 7.2] (using
that P 6≺ L(CΓ(g)) for g 6= e), we get that H ⊂ ∆(L2(pM p)). Since the unitary representation
{Ad v}v∈G on L2(pM p ⊖ Cp) is weakly mixing, we conclude that H ⊂ C∆(p).
Using the Fourier decomposition v = Pg∈Γ(v)gug, we get for every v ∈ G that
(v)g4 ≥ hΓ(v)4 ≥ hΓ(G)4 .
τ ((v ⊗ ∆(v))(∆(v)∗ ⊗ v∗)) = X
g∈Γ
Defining X ∈ M ⊗ M ⊗ M as the element of minimal k · k2 in the weakly closed convex hull of
{(v ⊗ ∆(v))(∆(v)∗ ⊗ v∗) v ∈ G}, we get that τ (X) ≥ hΓ(G)4, so that X is nonzero, and that
(v ⊗ ∆(v))X = X(∆(v) ⊗ v) for all v ∈ G. Also note that (p ⊗ ∆(p))X = X = X(∆(p) ⊗ p).
By the weak mixing of both Ad v and Ad ∆(v), it follows that XX ∗ is multiple of p ⊗ ∆(p) and
that X ∗X is a multiple of ∆(p) ⊗ p. We may thus assume that
XX ∗ = p ⊗ ∆(p)
and X ∗X = ∆(p) ⊗ p .
Define Y = (1 ⊗ X)(X ⊗ 1). Note that Y ∈ M ⊗ M ⊗ M ⊗ M is a partial isometry with
Y Y ∗ = p ⊗ p ⊗ ∆(p) and Y ∗Y = ∆(p) ⊗ p ⊗ p. Also,
Y = (v ⊗ v ⊗ ∆(v))Y (∆(v)∗ ⊗ v∗ ⊗ v∗)
for all v ∈ G .
Since Y is nonzero, it follows that the unitary representation ξ 7→ (v ⊗ v)ξ∆(v∗) of G on the
Hilbert space (p ⊗ p)L2(M ⊗ M )∆(p) is not weakly mixing. We thus find a finite dimensional
irreducible representation ω : G → U (Cn) and a nonzero Z ∈ Mn,1(C) ⊗ (p ⊗ p)L2(M ⊗ M )∆(p)
satisfying
(ω(v) ⊗ v ⊗ v)Z = Z∆(v)
for all v ∈ G .
By the weak mixing of Ad v and Ad ∆(v) and the irreducibility of ω, it follows that ZZ ∗ is a
multiple of 1⊗p⊗p and that Z ∗Z is a multiple of ∆(p). So, we may assume that ZZ ∗ = 1⊗p⊗p
It follows that Z ∗(Mn(C) ⊗ p ⊗ p)Z is an n2-dimensional globally
and that Z ∗Z = ∆(p).
{Ad ∆(v)}v∈G -invariant subspace of ∆(p)(M ⊗ M )∆(p). Again by weak mixing, this implies
that n = 1. But then, since τ (ZZ ∗) = τ (Z ∗Z), we also get that p = 1. So, Z ∈ M ⊗ M is a
unitary operator and ω : G → T is a character satisfying ω(v)(v ⊗ v)Z = Z∆(v) for all v ∈ G.
Denoting by σ : M ⊗ M → M ⊗ M the flip map and using that σ ◦ ∆ = ∆, it follows that
Zσ(Z)∗ commutes with all v ⊗ v, v ∈ G. By weak mixing, Zσ(Z)∗ is a multiple of 1. Using
that (∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆, we similarly find that (Z ⊗ 1)(∆ ⊗ id)(Z) is a multiple
of (1 ⊗ Z)(id ⊗ ∆)(Z). By [IPV10, Theorem 3.3], there exists a unitary u ∈ M such that
Z = (u∗ ⊗ u∗)∆(u). But then,
∆(uvu∗) = ω(v) uvu∗ ⊗ uvu∗
for all v ∈ G .
This means that uvu∗ ∈ TΓ for every v ∈ G.
5 Proof of Theorem B
As in [CIK13, Definition 2.7], we consider the class Crss of relatively strongly solid groups
consisting of all nonamenable countable groups Γ such that for any tracial crossed product
8
M = P ⋊ Γ and any von Neumann subalgebra Q ⊂ pM p that is amenable relative to P , we
have that either Q ≺ P or the normalizer NpM p(Q)′′ stays amenable relative to P .
The class Crss is quite large. Indeed, by [PV11, Theorem 1.6], all weakly amenable groups that
admit a proper 1-cocycle into an orthogonal representation weakly contained in the regular
representation belong to Crss. In particular, the free groups Fn with 2 ≤ n ≤ ∞ belong to
Crss and more generally, all free products Λ1 ∗ Λ2 of amenable groups Λ1, Λ2 with Λ1 ≥ 2 and
Λ2 ≥ 3 belong to Crss. By [PV12, Theorem 1.4], all weakly amenable, nonamenable, bi-exact
groups belong to Crss and thus Crss contains all nonelementary hyperbolic groups.
Theorem B is an immediate consequence of the more general Theorem 5.1 that we prove in
this section. In order to make our statements entirely explicit, we call a group measure space
(gms) decomposition of a tracial von Neumann algebra (M, τ ) any pair (B, Λ) where B ⊂ M
is a maximal abelian von Neumann subalgebra and Λ ⊂ U (M ) is a subgroup normalizing B
such that M = (B ∪ Λ)′′ and EB(v) = 0 for all v ∈ Λ \ {1}. This of course amounts to writing
M = B ⋊ Λ for some free and trace preserving action Λ y (B, τ ).
We then say that two gms decompositions (Bi, Λi), i = 0, 1, of M are
• identical if B0 = B1 and TΛ0 = TΛ1 ;
• unitarily conjugate if there exists a unitary u ∈ U (M ) such that uB0u∗ = B1 and uTΛ0u∗ =
TΛ1;
• conjugate by an automorphism if there exists an automorphism θ ∈ Aut(M ) such that
θ(B0) = B1 and θ(TΛ0) = TΛ1.
Theorem 5.1. Let Γ be a torsion free group in the class Crss. Let (A0, τ0) be any amenable
tracial von Neumann algebra with A0 6= C1 and β : Γ y (A0, τ0) any trace preserving action
such that Ker β is a nontrivial subgroup of Γ. Define (A, τ ) = (A0, τ0)Γ and denote by πk :
A0 → A the embedding as the k'th tensor factor. Define the action σ : Γ × Γ y (A, τ ) given by
σ(g,h)(πk(a)) = πgkh−1(βh(a)) for all g, k, h ∈ Γ and a ∈ A0. Denote M = A ⋊ (Γ × Γ).
Up to unitary conjugacy, all gms decompositions of M are given as M = B ⋊ Λ with B = BΓ
0
and Λ = Λ(Γ)
0 ⋊(Γ×Γ) where A0 = B0 ⋊Λ0 is a gms decomposition of A0 satisfying βg(B0) = B0
and βg(Λ0) = Λ0 for all g ∈ Γ.
Moreover, the gms decompositions of M associated with (B0, Λ0) and (B1, Λ1) are
(1) unitarily conjugate iff (B0, Λ0) is identical to (B1, Λ1),
(2) conjugate by an automorphism of M iff there exists a trace preserving automorphism θ0 :
A0 → A0 and an automorphism ϕ ∈ Aut(Γ) such that θ0(B0) = B1, θ0(TΛ0) = TΛ1 and
θ0 ◦ βg = βϕ(g) ◦ θ0 for all g ∈ Γ.
Note that in Proposition 5.14 at the end of this section, we discuss when the Cartan subalgebras
B = BΓ
0 are unitarily conjugate, resp. conjugate by an automorphism of M .
Lemma 5.2. Let Γ be a group in Crss and M = P ⋊ Γ any tracial crossed product. If Q1, Q2 ⊂
pM p are commuting von Neumann subalgebras, then either Q1 ≺M P or Q2 is amenable relative
to P .
Proof. Assume that Q1 6≺M P . By [BO08, Corollary F.14], there exists a diffuse abelian von
Neumann subalgebra A ⊂ Q1 such that A ⊀M P . Because Γ ∈ Crss, we get that NpM p(A)′′ is
amenable relative to P . Since Q2 ⊂ NpM p(A)′′, also Q2 is amenable relative to P .
It also follows that for groups Γ in Crss, the centralizer CΓ(L) of an infinite subgroup L < Γ
is amenable. So, torsion free groups Γ in Crss have the property that CΓ(g) is amenable for
9
every g 6= e. As a consequence, torsion free groups Γ in Crss are icc and even have the property
that every nonamenable subgroup L < Γ is relatively icc in the sense that {hgh−1 h ∈ L} is
an infinite set for every g ∈ Γ, g 6= e. Finally note that torsion free groups Γ in Crss have no
nontrivial amenable normal subgroups. In particular, every nontrivial normal subgroup of Γ is
relatively icc.
In the rest of this section, we prove Theorem 5.1. So, we fix a group Γ and an action Γ × Γ y A
as in the formulation of the theorem. We put M = A ⋊ (Γ × Γ).
Lemma 5.3. Let (N, τ ) be a tracial factor and let Q1, Q2 ⊂ N ⊗M be commuting von Neumann
subalgebras that are strongly nonamenable relative to N ⊗ 1. Then Q1 ∨ Q2 can be unitarily
conjugated into N ⊗ L(Γ × Γ).
Proof. Since A is amenable, we get that Q1 and Q2 are strongly nonamenable relative to
N ⊗ (A ⋊ L) whenever L < Γ × Γ is an amenable subgroup. For every g ∈ Γ, we denote by
Stab g ⊂ Γ × Γ the stabilizer of g under the left-right translation action Γ × Γ y Γ. We also
write Stab{g, h} = Stab g ∩ Stab h.
We start by proving that Q2 6≺ N ⊗ (A ⋊ Stab g) for all g ∈ Γ. Assume the contrary. Whenever
h 6= g, the group Stab{g, h} is amenable so that by Proposition 2.2, Q2 6≺ N ⊗ (A ⋊ Stab{g, h}).
Also by Proposition 2.2, we can take projections q ∈ Q2 and p ∈ N ⊗ (A ⋊ Stab g), a nonzero
partial isometry v ∈ q(N ⊗ M )p and a normal unital ∗-homomorphism
θ : qQ2q → p(N ⊗ (A ⋊ Stab g))p
such that xv = vθ(x) for all x ∈ qQ2q and such that, inside N ⊗ (A ⋊ Stab g), we have that
θ(qQ2q) is nonamenable relative to N ⊗ A and we have that θ(qQ2q) 6≺ N ⊗ (A ⋊ Stab{g, h})
whenever h 6= g.
Write P := θ(qQ2q)′ ∩ p(N ⊗ M )p. By Lemma 2.7, P ⊂ p(N ⊗ (A ⋊ Stab g))p. In particular,
v∗v ∈ p(N ⊗ (A ⋊ Stab g))p and we may assume that v∗v = p. Since Stab g ∼= Γ, we have
Stab g ∈ Crss and Lemma 5.2 implies that P ≺ N ⊗ A. Conjugating with v and writing
e = vv∗ ∈ (Q′
2 ∩ (N ⊗ M ),
it follows that Q1 ≺ N ⊗ A. By Proposition 2.2, this contradicts the strong nonamenability of
Q1 relative to N ⊗ A. So, we have proved that Q2 6≺ N ⊗ (A ⋊ Stab g) for all g ∈ Γ.
2 ∩ (N ⊗ M ))e ≺ N ⊗ A. Since Q1 ⊂ Q′
2 ∩ (N ⊗ M ))q, we find that e(Q′
Since Q1 is strongly nonamenable relative to N ⊗A and since Q2 6≺ N ⊗(A⋊Stab g) for all g ∈ Γ,
it follows from Theorem 2.6 that v∗Q2v ⊂ N ⊗ L(Γ × Γ) for some unitary v ∈ N ⊗ M . Since
v∗Q2v 6≺ N ⊗L(Stab g) for all g ∈ Γ, it follows from Lemma 2.7 that also v∗Q1v ⊂ N ⊗L(Γ×Γ).
This concludes the proof of the lemma.
We now also fix a gms decomposition M = B ⋊ Λ. We view Λ as a subgroup of U (M ). We
denote by ∆ : M → M ⊗ M the associated dual coaction given by ∆(b) = b ⊗ 1 for all b ∈ B
and ∆(v) = v ⊗ v for all v ∈ Λ.
Lemma 5.4. Writing Q1 = L(Γ × {e}) and Q2 = L({e} × Γ), we have ∆(Q2) ≺M ⊗M M ⊗ Qi
for either i = 1 or i = 2.
Proof. By Proposition 2.4, ∆(Q1) and ∆(Q2) are strongly nonamenable relative to M ⊗ 1. So
by Lemma 5.3, we can take a unitary v ∈ M ⊗ M such that
v∗∆(Q1 ∨ Q2)v ⊂ M ⊗ L(Γ × Γ) .
We therefore have the two commuting subalgebras v∗∆(Q1)v and v∗∆(Q2)v inside M ⊗L(Γ×Γ).
If v∗∆(Q1)v was amenable relative to both M ⊗ Q1 and M ⊗ Q2, then it would be amenable
10
relative to M ⊗ 1 by [PV11, Proposition 2.7], which is not the case. Hence v∗∆(Q1)v is
nonamenable relative to either M ⊗ Q1 or M ⊗ Q2. Assuming that v∗∆(Q1)v is nonamenable
relative to M ⊗ Q1, Lemma 5.2 implies that ∆(Q2) ≺ M ⊗ Q1.
In the following three lemmas, we prove that Λ contains two commuting nonamenable subgroups
Λ1, Λ2 < Λ. The method to produce such commuting subgroups is taken from [Io11] and our
proofs of Lemmas 5.5, 5.6 and 5.7 are very similar to the proof of [Io11, Theorem 3.1]. The
same method was also used in [CdSS15, Theorem 3.3]. For completeness, we provide all details.
Combining Theorem 3.1 and Lemma 5.4, we get the following.
Lemma 5.5. Denote by G the family of all amenable subgroups of Λ. For either i = 1 or i = 2,
there exists a finite set x1, . . . , xn ∈ M and a δ > 0 such that the following holds: whenever
F ⊂ Λ is small relative to G, we can find an element vF ∈ Λ − F such that
nX
k,j=1
kEQi(xkvF x∗
j )k2
2 ≥ δ .
Lemma 5.6. There exists a decreasing sequence of nonamenable subgroups Λn < Λ such that
Qi ≺M B ⋊ (Sn CΛ(Λn)) for either i = 1 or i = 2, where CΛ(Λn) denotes the centralizer of Λn
inside Λ.
Proof. As in Lemma 5.5, we let G denote the family of all amenable subgroups of Λ. We
denote by I the set of subsets of Λ that are small relative to G. We order I by inclusion and
choose a cofinal ultrafilter U on I. Consider the ultrapower von Neumann algebra M U and the
ultrapower group ΛU . Every v = (vF )F ∈I ∈ ΛU can be viewed as a unitary in M U .
Assume without loss of generality that i = 1 in Lemma 5.5 and denote by v = (vF )F ∈I the
element of ΛU that we found in Lemma 5.5. Denote by K ⊂ L2(M U ) the closed linear span of
M vM and by PK the orthogonal projection from L2(M U ) onto K. Put Σ = Λ ∩ vΛv−1. We
claim that Q2 ≺M B ⋊ Σ.
Assume the contrary. This means that we can find a sequence of unitaries an ∈ U (Q2) such
that kEB⋊Σ(xany)k2 → 0 for any x, y ∈ M . We prove that hanξa∗
n, ηi → 0 as n → ∞ for all
ξ, η ∈ K. For this, it suffices to prove that hanxvx′a∗
n, yvy′i → 0 for all x, x′, y, y′ ∈ M . First,
note that for all z ∈ M , we have EM (v∗zv) = EM (v∗EB⋊Σ(z)v) by definition of the subgroup
Σ. Hence
hanxvx′a∗
n, yvy′i = τ (EM (v∗y∗anxv)x′a∗
ny′∗)
= τ (EM (v∗EB⋊Σ(y∗anx)v)x′a∗
≤ kx′kky′kkEB⋊Σ(y∗anx)k2 → 0
ny′∗)
as wanted.
1
1
1
Next, Lemma 5.5 provides a finite set L ⊂ M such that Px,y∈L kEQU
(xvy∗) 6= 0. Put ξ = PK (EQU
(xvy∗) 6= 0, we get that kxvy∗ − EQU
ular, we can take x, y ∈ L such that EQU
ξ 6= 0. Since EQU
follows that kxvy∗ − ξk2 = kPK (xvy∗ − EQU
Since K is an M -M -bimodule and since Q1 commutes with Q2, we have that aξ = ξa for all
a ∈ Q2. In particular, hanξa∗
n, ξi → 0.
This proves that Q2 ≺M B ⋊ Σ.
2 > 0 in contradiction with the fact that hanξa∗
(xvy∗))k2 < kxvy∗k2. Hence ξ 6= 0.
n, ξi = kξk2
(xvy∗)k2
2 6= 0. In partic-
(xvy∗)). We claim that
(xvy∗)k2 < kxvy∗k2. Since xvy∗ ∈ K, it
1
1
1
It remains to show that there exists a decreasing sequence of subgroups Λn < Λ such that for
all n we have Λn /∈ G, and such that Σ = Sn CΛ(Λn). For every T ⊂ I, we denote by ΛT the
11
subgroup of Λ generated by {vF v−1
F, F ′ ∈ T }. An element w ∈ Λ belongs to Σ if and only
F ′
if there exists a T ∈ U such that w commutes with ΛT . Enumerating Σ = {w1, w2, . . .}, choose
Sn ∈ U such that wn commutes with ΛSn. Then put Tn := S1 ∩ . . . ∩ Sn. Note that Tn ∈ U and
by construction, Σ = Sn CΛ(ΛTn). It remains to prove that ΛT /∈ G for all T ∈ U .
Fix T ∈ U and assume that ΛT ∈ G. Fix an element F ′ ∈ T . Then {vF F ∈ T } ⊂ ΛT vF ′.
So, F1 := {vF F ∈ T } is small relative to G. Define T ′ ⊂ I by T ′ = {F ∈ I F1 ⊂ F }. Since
U is a cofinal ultrafilter and T ∈ U , we get T ∩ T ′ 6= ∅. So we can take F ∈ T with F1 ⊂ F .
Then, vF ∈ Λ − F ⊂ Λ − F1 but also vF ∈ F1. This being absurd, we have shown that ΛT /∈ G
for all T ∈ U .
Lemma 5.7. There exist two commuting nonamenable subgroups Λ1 and Λ2 inside Λ. More-
over, whenever Λ1, Λ2 < Λ are commuting nonamenable subgroups, L(Λ1Λ2) can be unitarily
conjugated into L(Γ × Γ).
Proof. From Lemma 5.6, we find a decreasing sequence of nonamenable subgroups Λn < Λ
such that Qi ≺M B ⋊ (Sn CΛ(Λn)) for either i = 1 or i = 2. Since Qi has no amenable
direct summand, we get that the group Sn CΛ(Λn) is nonamenable. It follows that CΛ(Λn) is
nonamenable for some n ∈ N. Denote Λ1 := Λn and Λ2 := CΛ(Λn).
When Λ1, Λ2 < Λ are commuting nonamenable subgroups, it follows from Lemma 5.3 applied
to N = C1 that L(Λ1) ∨ L(Λ2) can be unitarily conjugated into L(Γ × Γ).
From now on, we fix commuting nonamenable subgroups Λ1, Λ2 < Λ. By Lemma 5.7, after a
unitary conjugacy, we may assume that L(Λ1Λ2) ⊂ L(Γ × Γ).
Lemma 5.8. If N ⊂ L(Γ×Γ) is an amenable von Neumann subalgebra such that the normalizer
NL(Γ×Γ)(N )′′ contains L(Λ1Λ2), then N is atomic. Also, L(Λ1Λ2)′ ∩ L(Γ × Γ) is atomic.
Proof. Using [Va10, Proposition 2.6], we find a projection q in the center of the normalizer of
N such that N q ≺f L(Γ) ⊗ 1 and N (1 − q) ⊀ L(Γ) ⊗ 1.
Assume for contradiction that q 6= 1. Since N (1 − q) ⊀ L(Γ) ⊗ 1 and since Γ ∈ Crss, it
follows that L(Λi)(1 − q) is amenable relative to L(Γ) ⊗ 1 for both i = 1, 2. It then follows
from [PV11, Proposition 2.7] that L(Λ1)(1 − q) is nonamenable relative to 1 ⊗ L(Γ), hence
L(Λ2)(1 − q) ≺ 1 ⊗ L(Γ) by Lemma 5.2. By Proposition 2.2, we get a nonzero projection
q0 ≤ 1 − q that commutes with L(Λ2) such that L(Λ2)q0 is amenable relative to 1 ⊗ L(Γ). But
since L(Λ2)q0 is also amenable relative to L(Γ) ⊗ 1, it follows from [PV11, Proposition 2.7] that
L(Λ2)q0 is amenable relative to C1, hence a contradiction.
We conclude that q = 1 so that N ≺f L(Γ) ⊗ 1. By symmetry, we also get that N ≺f 1 ⊗ L(Γ).
Hence N ≺f C1, so that N is atomic.
To prove that L(Λ1Λ2)′ ∩L(Γ×Γ) is atomic, it suffices to prove that every abelian von Neumann
subalgebra D ⊂ L(Λ1Λ2)′ ∩ L(Γ × Γ) is atomic. But then D is amenable and its normalizer
contains L(Λ1Λ2), so that D is indeed atomic.
The proof of the following lemma is essentially contained in the proof of [OP03, Proposition
12].
Lemma 5.9. For every minimal projection e ∈ L(Λ1Λ2)′ ∩ L(Γ × Γ), there exist projections
p ∈ Mn(C)⊗L(Γ), q ∈ L(Γ)⊗Mm(C) and a partial isometry u ∈ Mn,1(C)⊗L(Γ×Γ)⊗Mm,1(C)
such that u∗u = e, uu∗ = p ⊗ q and such that
either uL(Λ1)u∗ ⊂ p(Mn(C) ⊗ L(Γ))p ⊗ q
or uL(Λ1)u∗ ⊂ p ⊗ q(L(Γ) ⊗ Mm(C))q
and uL(Λ2)u∗ ⊂ p ⊗ q(L(Γ) ⊗ Mm(C))q ,
and uL(Λ2)u∗ ⊂ p(Mn(C) ⊗ L(Γ))p ⊗ q .
12
Proof. By [PV11, Proposition 2.7], L(Λ2)e is nonamenable relative to either L(Γ)⊗1 or 1⊗L(Γ).
Assume that L(Λ2)e is nonamenable relative to L(Γ) ⊗ 1. By Lemma 5.2, L(Λ1)e ≺ L(Γ) ⊗ 1.
Take a projection p ∈ Mn(C)⊗L(Γ), a nonzero partial isometry v ∈ (p⊗1)(Mn,1(C)⊗L(Γ×Γ))e
and a unital normal ∗-homomorphism θ : L(Λ1) → p(Mn(C) ⊗ L(Γ))p such that
(θ(x) ⊗ 1)v = vx for all x ∈ L(Λ1) .
Since Γ ∈ Crss and Λ1 is nonamenable, the relative commutant θ(L(Λ1))′ ∩ p(Mn(C) ⊗ L(Γ))p
is atomic. Cutting with a minimal projection, we may assume that this relative commutant
equals Cp.
Write P := L(Λ1)′ ∩ L(Γ × Γ) and note that v∗v, e ∈ P with v∗v ≤ e. Since L(Λ2) ⊂ P , we
have that Z(P ) ⊂ L(Λ1Λ2)′ ∩ L(Γ × Γ). It follows that Z(P )e = Ce. So, eP e is a II1 factor and
we can take partial isometries v1, . . . , vm ∈ eP e with viv∗
i vi = e.
i ≤ v∗v for all i and Pm
Define u ∈ Mn,1(C) ⊗ L(Γ × Γ) ⊗ Mm,1(C) given by u = Pm
Since vP v∗ commutes with θ(L(Λ1)) ⊗ 1, we have vP v∗ ⊂ p ⊗ L(Γ) and we can define the
normal ∗-homomorphism η : v∗vP v∗v → L(Γ) such that vyv∗ = p ⊗ η(y) for all y ∈ v∗vP v∗v.
By construction, u∗u = e and uu∗ = p ⊗ q where q ∈ L(Γ) ⊗ Mm(C) is the projection given by
i=1 vvi ⊗ ei1.
i=1 v∗
q = Pm
i=1 η(viv∗
i ) ⊗ eii. Defining the ∗-homomorphism
eη : eP e → q(L(Γ) ⊗ Mm(C))q : eη(y) =
mX
i,j=1
η(viyv∗
j ) ⊗ eij
and using that L(Λ2)e ⊂ eP e, we get that
uL(Λ1)u∗ = θ(L(Λ1)) ⊗ q
This concludes the proof of the lemma.
and uL(Λ2)u∗ = p ⊗ eη(L(Λ2)e) .
Recall from Section 4 the notion of height of an element in a group von Neumann algebra (here,
L(Γ × Γ)), as well as the height of a subgroup of U (L(Γ × Γ)). The proof of the following lemma
is very similar to the proof of [Io10, Theorem 4.1].
Lemma 5.10. For every projection p ∈ L(Λ1Λ2)′ ∩ L(Γ × Γ), we have that hΓ×Γ(Λ1Λ2p) > 0.
Proof. Fix a minimal projection p ∈ L(Λ1Λ2)′∩L(Γ×Γ). It suffices to prove that hΓ×Γ(Λ1Λ2p) >
0. Using the conjugacy of Lemma 5.9, we see that the heights of Λ1p and Λ2p do not interact,
so that it suffices to prove that hΓ×Γ(Λip) > 0 for i = 1, 2. By symmetry, it is enough to prove
this for i = 1.
Assume for contradiction that hΓ×Γ(Λ1p) = 0. Take a sequence vn ∈ Λ1 such that hΓ×Γ(vnp) →
0. For every finite subset S ⊂ Γ × Γ, we denote by PS the orthogonal projection of L2(M ) onto
the linear span of L2(A)ug, g ∈ S. We claim that for every sequence of unitaries wn ∈ L(Γ× Γ),
every a ∈ M ⊖ L(Γ × Γ) and every finite subset S ⊂ Γ × Γ, we have that
lim
n
kPS(pvnawn)k2 = 0 .
Since PS(x) = Pg∈S EA(xu∗
g)ug, it suffices to prove that kEA(pvnawn)k2 → 0 for all a ∈
M ⊖ L(Γ × Γ). Every such a can be approximated by a linear combination of elements of the
form a0ug with a0 ∈ A ⊖ C1 and g ∈ Γ × Γ. So, we may assume that a ∈ A ⊖ C1. Such an
element a can be approximated by a linear combination of elementary tensors, so that we may
assume that a = Ni∈G ai for some finite nonempty subset G ⊂ Γ and elements ai ∈ A0 ⊖ C1
with kak ≤ 1. Note that σg(a) ⊥ σh(a) whenever g, h ∈ Γ × Γ and g · G 6= h · G (where we use
the left right action of Γ × Γ on Γ).
13
Choose ε > 0. By Lemma 5.9, we can take a finite subset F0 ⊂ Γ such that, writing F =
Γ × F0 ∪ F0 × Γ, we have kpv − PF (pv)k2 ≤ ε for all v ∈ Λ1. Then,
kEA(pvnawn) − EA(PF (pvn)awn)k2 ≤ ε
for all n, so that in order to prove the claim, it suffices to prove that kEA(PF (pvn)awn)k2 → 0.
Put κ = 2F0G2. Note that for every h ∈ Γ × Γ, the set {g ∈ F g · G = h · G} contains at
most κ elements. Using the Fourier decomposition for elements in L(Γ × Γ), we have
EA(PF (pvn)awn) = X
g∈F
(pvn)g (wn)g−1 σg(a) .
Thus, for all h ∈ Γ × Γ, we have
hEA(PF (pvn)awn), σh(a)i ≤ κ hΓ×Γ(pvn) .
But then, using the Cauchy-Schwarz inequality, we get that
kEA(PF (pvn)awn)k2
hEA(PF (pvn)awn) , (pvn)h (wn)h−1 σh(a)i
2 ≤ X
≤ κ hΓ×Γ(pvn) X
h∈F
h∈Γ×Γ
(pvn)h (wn)h−1
≤ κ hΓ×Γ(pvn) kpvnk2 kwnk2 ≤ κ hΓ×Γ(pvn) → 0 .
So, the claim is proved.
Put δ = kpk2/4. Because Γ ∈ Crss and B ⊂ M is a Cartan subalgebra, we have that B ≺f A.
By Lemma 2.1, we have B 6≺ L(Γ × Γ) and we can take a unitary b ∈ U (B) such that
kEL(Γ×Γ)(b)k2 ≤ δ. Since B ≺f A, we can take a finite subset S ⊂ Γ × Γ such that kpd −
PS(pd)k2 ≤ δ for all d ∈ U (B). For every n, we have that vnbv∗
n ∈ U (B). Therefore,
for all n. Since kEL(Γ×Γ)(b)k2 ≤ δ, writing b1 = b − EL(Γ×Γ)(b), we get
kpvnbv∗
n − PS (pvnbv∗
n)k2 ≤ δ
kPS (pvnbv∗
n) − PS(pvnb1v∗
n)k2 ≤ δ .
By the claim above, we can fix n large enough such that kPS (pvnb1v∗
proved that
n)k2 ≤ δ. So, we have
kpk2 = kpvnbv∗
nk2 ≤ 3δ < kpk2 ,
which is absurd. So, we have shown that hΓ×Γ(pΛ1) > 0 and the lemma is proved.
Lemma 5.11. There exists a unitary u ∈ L(Γ × Γ) such that uΛ1Λ2u∗ ⊂ T(Γ × Γ). Also, the
unitary representation {Ad v}v∈Λ1Λ2 is weakly mixing on L2(M ) ⊖ C1.
Proof. Write Λ0 = Λ1Λ2. Denote the action of Λ on B by γv(b) = vbv∗ for all v ∈ Λ, b ∈ B.
Define K < Λ as the virtual centralizer of Λ0 inside Λ, i.e. K consists of all v ∈ Λ such that the
set {wvw−1 w ∈ Λ0} is finite. Define B0 ⊂ B as the von Neumann algebra generated by the
unital ∗-algebra consisting of all b ∈ B such that {γv(b) v ∈ Λ0} spans a finite dimensional
subspace of B. Note that B0 is globally invariant under γv, v ∈ Λ0. Viewing M as the crossed
product M = B ⋊ Λ, we have by construction that the unitary representation {Ad v}v∈Λ0 is
weakly mixing on L2(M ) ⊖ L2(B0 ⋊ K).
For every g ∈ Γ, define Stab g as in the beginning of the proof of Lemma 5.3. We have L(Λ0) ⊂
L(Γ × Γ) and L(Λ0) 6≺ L(Stab g) for all g ∈ Γ. By Lemma 2.7, we have B0 ⋊ K ⊂ L(Γ × Γ).
14
Since we can take decreasing sequences of finite index subgroups Λi,n < Λi, i = 1, 2, such that
K = Sn CΛ(Λ1,nΛ2,n), it follows from Lemma 5.8 that L(K) is contained in a hyperfinite von
Neumann algebra. So, K is amenable and thus also B0 ⋊ K is amenable. Since B0 ⋊ K is
normalized by Λ0, it follows from Lemma 5.8 that B0 ⋊ K is atomic. So, K is a finite group and
B0 is atomic. We can then take a minimal projection p ∈ B0 ⋊ K and finite index subgroups
Λ3 < Λ1 and Λ4 < Λ2 such that p commutes with Λ3Λ4.
Lemmas 5.8, 5.9 and 5.10 apply to the commuting nonamenable subgroups Λ3, Λ4 < Λ. So,
by Lemma 5.10, we get that hΓ×Γ(pΛ3Λ4) > 0. By construction, the unitary representation
{Ad v}v∈Λ3Λ4 is weakly mixing on pL(Γ × Γ)p ⊖ Cp. For every g ∈ Γ × Γ with g 6= e, the
centralizer CΓ×Γ(g) is either amenable or of the form Γ × L or L × Γ with L < Γ amenable.
Therefore, L(Λ3Λ4) 6≺ L(CΓ×Γ(g)) for all g 6= e. It then first follows from Theorem 4.1 that
p = 1, so that we could have taken Λ3 = Λ1 and Λ4 = Λ2, and then also that there exists a
unitary u ∈ L(Γ × Γ) such that uΛ1Λ2u∗ ⊂ T(Γ × Γ).
Since we also proved that B0 ⋊ K = C1, it follows as well that the unitary representation
{Ad v}v∈Λ1Λ2 is weakly mixing on L2(M ) ⊖ C1.
Lemma 5.12. Whenever Λ2 ⊂ T({e} × Γ) is a nonamenable subgroup, we have M ∩ Λ′
L(Γ) ⊗ 1.
2 =
Proof. Define Γ2 < Γ such that TΛ2 = T({e} × Γ2). Then Γ2 is nonamenable and M ∩ Λ′
2 =
M ∩ L({e} × Γ2)′. Since Γ2 < Γ is relatively icc, we have M ∩ L({e} × Γ2)′ ⊂ A ⋊ (Γ × {e}).
Since the action {e} × Γ y A is mixing, it follows that M ∩ L({e} × Γ2)′ ⊂ L(Γ × {e}). So,
M ∩ Λ′
2 ⊂ L(Γ) ⊗ 1 and the converse inclusion is obvious.
Lemma 5.13. There exist commuting subgroups H1, H2 < Λ and a unitary u ∈ M such that
Λi < Hi for i = 1, 2 and uTH1H2u∗ = T(Γ × Γ).
Proof. By Lemma 5.11, after a unitary conjugacy, we may assume that Λ1, Λ2 < Λ are com-
muting nonamenable subgroups with Λ1Λ2 ⊂ T(Γ × Γ). Since Γ is torsion free and belongs to
Crss, we have that CΓ(g) is amenable for every g 6= e. Therefore, after exchanging Λ1 and Λ2 if
needed, we have Λ1 ⊂ T(Γ × {e}) and Λ2 ⊂ T({e} × Γ).
Denote by {γv}v∈Λ the action of Λ on B. Define H1 < Λ as the virtual centralizer of Λ2 inside
Λ. So, H1 consists of all v ∈ Λ that commute with a finite index subgroup of Λ2. Similarly,
define B1 as the von Neumann algebra generated by the ∗-algebra consisting of all b ∈ B such
that γv(b) = b for all v in a finite index subgroup of Λ2. Since finite index subgroups of Λ2 are
nonamenable, it follows from Lemma 5.12 that B1 ⋊ H1 ⊂ L(Γ) ⊗ 1. We also find that
L(Γ) ⊗ 1 ⊂ L(Λ2)′ ∩ (B ⋊ Λ) ⊂ B1 ⋊ H1 .
So, B1 ⋊ H1 = L(Γ) ⊗ 1. In particular, the subgroups H1, Λ2 < Λ commute. Because Γ ∈ Crss
and B1 ⊂ L(Γ) ⊗ 1 is normalized by H1, it follows that B1 is atomic. Since Λ1 < H1, the
unitaries v ∈ Λ1Λ2 normalize B1. By Lemma 5.11, they induce a weakly mixing action on B1.
Since B1 is atomic, this forces B1 = C1. We conclude that L(H1) = L(Γ) ⊗ 1.
We now apply Lemmas 5.8, 5.9 and 5.10 to the commuting nonamenable subgroups H1, Λ2 < Λ.
We conclude that hΓ(H1) > 0. Since L(H1) = L(Γ) ⊗ 1, the group H1 is icc. So, the action
{Ad v}v∈H1 on L(Γ) ⊖ C1 is weakly mixing. Since for g 6= e, the group CΓ(g) is amenable,
also L(H1) 6≺ L(CΓ(g)). So, by Theorem 4.1, there exists a unitary u1 ∈ L(Γ) such that
(u1 ⊗ 1)H1(u∗
1 ⊗ 1) = T(Γ × {e}).
15
Applying the same reasoning as above to the virtual centralizer of H1 inside Λ, we find a
subgroup H2 < Λ, containing Λ2 and commuting with H1, and we find a unitary u2 ∈ L(Γ)
such that (1 ⊗ u2)H2(1 ⊗ u∗
2) ⊂ T({e} ⋊ Γ). So, we get that
(u1 ⊗ u2)H1H2(u∗
1 ⊗ u∗
2) = T(Γ × Γ) .
Finally, we are ready to prove Theorem 5.1. As mentioned above, Theorem B is a direct
consequence of Theorem 5.1.
Proof of Theorem 5.1. First assume that (B0, Λ0) is a gms decomposition of A0 satisfying
βg(B0) = B0 and βg(Λ0) = Λ0 for all g ∈ Γ. Then, {βg}g∈Γ defines an action of Γ by au-
tomorphisms of the group Λ0. We can co-induce this to the action of Γ × Γ by automorphisms
of the direct sum group Λ(Γ)
given by (g, h) · πk(v) = πgkh−1(βh(v)) for all g, h, k ∈ Γ, v ∈ Λ0,
where πk : Λ0 → Λ(Γ)
denotes the embedding as the k'th direct summand. Putting B = BΛ
0
0
and Λ := Λ(Γ)
0 ⋊ (Γ × Γ), we have found the crossed product decomposition M = B ⋊ Λ. It is
easy to check that B ⊂ M is maximal abelian, so that (B, Λ) is indeed a gms decomposition
of M .
0
Conversely, assume that (B, Λ) is an arbitrary gms decomposition of M . By Lemma 5.13 and
after a unitary conjugacy, we have Γ × Γ ⊂ TΛ. Denoting by ∆ : M → M ⊗ M the dual
coaction associated with (B, Λ) and given by ∆(b) = b ⊗ 1 for all b ∈ B and ∆(v) = v ⊗ v for
all v ∈ Λ, this means that ∆(u(g,h)) is a multiple of u(g,h) ⊗ u(g,h) for all (g, h) ∈ Γ × Γ.
Denote A0,e := πe(A0) ⊂ A and observe that A0,e commutes with all u(g,g), g ∈ Ker β. Then,
∆(A0,e) commutes with all u(g,g) ⊗ u(g,g), g ∈ Ker β. Since Γ is a torsion free group in Crss, the
nontrivial normal subgroup Ker β < Γ must be nonamenable and thus relatively icc. It follows
that the unitary representation {Ad u(g,g)}g∈Ker β is weakly mixing on L2(M ) ⊖ L2(A0,e). This
implies that ∆(A0,e) ⊂ A0,e ⊗ A0,e.
By Lemma 2.3, we get a crossed product decomposition A0 = B0 ⋊ Λ0 such that πe(B0) =
B ∩ A0,e and πe(Λ0) = Λ ∩ A0,e. For every g ∈ Γ, we have that u(g,g) ∈ TΛ. So, u(g,g)
normalizes both B and A0,e, so that βg(B0) = B0. Also, u(g,g) normalizes both Λ and A0,e, so
that βg(Λ0) = Λ0. For every g ∈ Γ, we have that u(g,e) ∈ TΛ so that u(g,e) normalizes B and
Λ. It follows that πg(B0) ⊂ B and πg(Λ0) ⊂ Λ for all g ∈ Γ. We conclude that
BΓ
0 ⊂ B and Λ(Γ)
0 ⋊ (Γ × Γ) ⊂ TΛ .
(5.1)
0 and Λ(Γ)
Since A0 is generated by B0 and Λ0, we get that M is generated by BΓ
0 ⋊ (Γ × Γ).
Since M is also the crossed product of B and Λ, it follows from (5.1) that BΓ
0 = B and
TΛ(Γ)
0 ⋊ (Γ × Γ) = TΛ. In particular, B0 ⊂ A0 must be maximal abelian. So, (B0, Λ0) is a
gms decomposition of A0 that is {βg}g∈Γ-invariant, while the gms decomposition (B, Λ) of M
is unitarily conjugate to the gms decomposition associated with (B0, Λ0).
It remains to prove statements (1) and (2). Take {βg}g∈Γ-invariant gms decompositions (B0, Λ0)
and (B1, Λ1) of A0. Denote by (B, Λ) and (B′, Λ′) the associated gms decompositions of M .
To prove (1), assume that u ∈ M is a unitary satisfying uBu∗ = B′ and uTΛu∗ = TΛ′. It
follows that for all g ∈ Γ × Γ, we have uugu∗ ∈ U (A)uϕ(g) where ϕ ∈ Aut(Γ × Γ). Write
u = Ph∈Γ×Γ ahuh with ah ∈ A for the Fourier decomposition of u. It follows that {ϕ(g)−1hg
g ∈ Γ × Γ} is a finite set whenever ah 6= 0. Since Γ × Γ is icc, it follows that ah can only be
nonzero for one h ∈ Γ × Γ. So u is of the form u = ahuh. Since uh normalizes both B and Λ,
we may replace u with uu∗
h so that u ∈ U (A).
16
For each g ∈ Γ, we define Eg : A → A0 by Eg(x) = π−1
g (Eπg(A0)(x)), x ∈ A. Let (gn)n∈N be
a sequence in Γ that tends to infinity, and let b ∈ B0. Since (πgn(b))n∈N is an asymptotically
central sequence in A, we get that
B1 ∋ Egn(uπgn(b)u∗) → b ,
hence B0 ⊂ B1. By symmetry, it follows that B0 = B1. Similarly, we see that TΛ0 = TΛ1 so
we conclude that (B0, Λ0) and (B1, Λ1) are identical gms decompositions of A0.
To prove (2), assume that θ ∈ Aut(M ) is an automorphism satisfying θ(B) = B′ and θ(TΛ) =
TΛ′. Define the commuting subgroups Λ1, Λ2 < Λ′ such that θ(T(Γ × {e})) = TΛ1 and
θ(T({e} × Γ)) = TΛ2. Applying Lemma 5.13 to the gms decomposition (B′, Λ′) of M and these
commuting subgroups Λ1, Λ2 < Λ′, we find commuting subgroups H1, H2 < Λ′ and a unitary
u ∈ M such that Λi < Hi for i = 1, 2 and uTH1H2u∗ = T(Γ × Γ). Since Γ × {e} and {e} × Γ
are each other's centralizer inside Λ and since θ(TΛ) = TΛ′, we must have that Λi = Hi for
i = 1, 2.
Writing θ1 = Ad u ◦ θ, we have proved that θ1(T(Γ × Γ)) = T(Γ × Γ). This equality induces
an automorphism of Γ × Γ. Since Γ is a torsion free group in Crss, all automorphisms of Γ × Γ
are either of the form (g, h) 7→ (ϕ1(g), ϕ2(h)) or of the form (g, h) 7→ (ϕ1(h), ϕ2(g)) for some
automorphisms ϕi ∈ Aut(Γ). The formulas ζ(u(g,h)) = u(h,g) and ζ(πk(a)) = πk−1(βk(a)) for
all g, h, k ∈ Γ and a ∈ A0 define an automorphism ζ ∈ Aut(M ) satisfying ζ(B′) = B′ and
ζ(Λ′) = Λ′. So composing θ with ζ if necessary, we may assume that we have ϕ1, ϕ2 ∈ Aut(Γ)
such that θ1(u(g,h)) ∈ Tu(ϕ1(g),ϕ2(h)) for all g, h ∈ Γ. We still have that the gms decompositions
(θ1(B), θ1(Λ)) and (B′, Λ′) of M are unitarily conjugate.
Because u(g,g) commutes with πe(A0) for all g ∈ Ker β, the unitary representation on L2(M ) ⊖
C1 given by {Ad u(ϕ1(g),ϕ2(g))}g∈Ker β is not weakly mixing. There thus exists a k ∈ Γ such
that ϕ1(g)k = kϕ2(g) for all g in a finite index subgroup of Ker β. So after replacing θ by
(Ad u(e,k)) ◦ θ, which globally preserves B′ and Λ′, we may assume that ϕ1(g) = ϕ2(g) for
all g in a finite index subgroup of Ker β. Since Ker β < Γ is relatively icc, this implies that
ϕ1(g) = ϕ2(g) for all g ∈ Ker β. Since Ker β is a normal subgroup of Γ, it follows that
ϕ1(k)ϕ2(k)−1 commutes with ϕ1(g) for all k ∈ Γ and g ∈ Ker β. So, ϕ1 = ϕ2 and we denote
this automorphism as ϕ.
Taking the commutant of the unitaries u(g,g), g ∈ Ker β, it follows that θ1(πe(A0)) = πe(A0).
We define the automorphism θ0 ∈ Aut(A0) such that θ1 ◦ πe = πe ◦ θ0. Since θ1(u(g,g)) ∈
Tu(ϕ(g),ϕ(g)), we get that θ0 ◦ βg = βϕ(g) ◦ θ0 for all g ∈ Γ.
It follows that (θ0(B0), θ0(Λ0))
is a {βg}g∈Γ-invariant gms decomposition of A0. The associated gms decomposition of M is
(θ1(B), θ1(Λ)). This gms decomposition of M is unitarily conjugate with the gms decomposition
(B′, Λ′). It then follows from (1) that θ0(B0) = B1 and θ0(TΛ0) = TΛ1.
Proposition 5.14. Under the same hypotheses and with the same notations as in Theorem 5.1,
if (B0, Λ0) and (B1, Λ1) are {βg}g∈Γ-invariant gms decompositions of A0, then the associated
Cartan subalgebras of M given by BΓ
0 and BΓ
1 are
(1) unitarily conjugate iff B0 = B1 ;
(2) conjugate by an automorphism of M iff there exists a trace preserving automorphism θ0 :
A0 → A0 and an automorphism ϕ ∈ Aut(Γ) such that θ0(B0) = B1 and θ0 ◦ βg = βϕ(g) ◦ θ0
for all g ∈ Γ.
Proof. To prove (1), it suffices to prove that BΓ
1 if B0 6= B1. Take a unitary u ∈
U (B0) such that u 6∈ B1. Then kEB1 (u)k2 < 1. Let {g1, g2, . . .} be an enumeration of Γ and
define the sequence of unitaries (wn) ⊂ U (BΓ
0 ) by wn = πgn+1(u) πgn+2(u) · · · πg2n(u). Then
0 ⊀ BΓ
(xwny)k2 → 0 for all x, y ∈ M so that BΓ
kEBΓ
1 .
0 6≺ BΓ
1
17
To prove (2), denote by (B, Λ) and (B′, Λ′) the gms decompositions of M associated with
(B0, Λ0) and (B1, Λ1). Assume that θ ∈ Aut(M ) satisfies θ(B) = B′. Then, (B′, θ(Λ)) is
a gms decomposition of M . By Theorem 5.1, (B′, θ(Λ)) is unitarily conjugate with the gms
decomposition associated with a {βg}g∈Γ-invariant gms decomposition (B2, Λ2) of A0. By (1),
we must have B2 = B1. So the gms decompositions associated with (B0, Λ0) and (B1, Λ2)
are conjugate by an automorphism of M . By Theorem 5.1(2) there exists an automorphism
θ0 ∈ Aut(A0) as in (2).
6 Examples of II1 factors with a prescribed number of group
measure space decompositions
For every amenable tracial von Neumann algebra (A0, τ0) and for every trace preserving action
of Γ = F∞ on (A0, τ0) with nontrivial kernel, Theorem 5.1 gives a complete description of
all gms decompositions of the II1 factor M = AΓ
0 ⋊ (Γ × Γ) in terms of the Γ-invariant gms
decompositions of A0.
In this section, we construct a family of examples where these Γ-invariant gms decompositions
of A0 can be explicitly determined. In particular, this gives a proof of Theorem A. We will
construct A0 of the form A0 = L∞(K)⋊H1 where H1 is a countable abelian group and H1 ֒→ K
is an embedding of H1 as a dense subgroup of the compact second countable group K. Note
Ω(g, h) = 1 for all h ∈ H2, then g = e ; if h ∈ H2 and Ω(g, h) = 1 for all g ∈ H1, then h = e.
that we can equally view K as cH2 where H2 is a countable abelian group and the embedding
H1 ֒→ cH2 is given by a bicharacter Ω : H1 × H2 → T that is nondegenerate:
We can then view L∞(cH2) ⋊ H1 as being generated by the group von Neumann algebras L(H1)
and L(H2), with canonical unitaries {ug}g∈H1 and {uh}h∈H2, satisfying uguh = Ω(g, h)uhug for
all g ∈ H1, h ∈ H2.
if g ∈ H1 and
We call a direct sum decomposition H1 = S1 ⊕ T1 admissible if the closures of S1, T1 in cH2
give a direct sum decomposition of cH2. This is equivalent to saying that H2 = S2 ⊕ T2 in such
2 , where the "orthogonal complement" is
1 and T1 = T ⊥
1 , T2 = T ⊥
2 , S2 = S⊥
a way that S1 = S⊥
defined w.r.t. Ω.
Put Γ0 = SL(3, Z) and Hi = L3
Proposition 6.1. Let L1, L2 be torsion free abelian groups and L1 ֒→ cL2 a dense embedding.
H1 ֒→ cH2, defining the trace preserving action {βg}g∈Γ0 of Γ0 on A0 = L∞(cH2) ⋊ H1.
i . Consider the natural action of Γ0 on the direct sum embedding
Whenever L1 = P1 ⊕ Q1 is an admissible direct sum decomposition with corresponding L2 =
P2 ⊕ Q2, put Si = P 3
Then (B0, Λ0) is a {βg}g∈Γ0 -invariant gms decomposition of A0. Every {βg}g∈Γ0 -invariant
gms decomposition of A0 is of this form for a unique admissible direct sum decomposition
L1 = P1 ⊕ Q1.
i and define B0 = L(S1) ∨ L(S2), Λ0 = T1T2.
i , Ti = Q3
Proof. Let (B0, Λ0) be a {βg}g∈Γ0-invariant gms decomposition of A0. Define the subgroup
Γ1 < Γ0 as
Γ1 = Γ0 ∩
1 ∗ ∗
0 ∗ ∗
0 ∗ ∗
We also put H (1)
• a · g = g for all a ∈ Γ1 and g ∈ H (1)
1 .
1 = L1 ⊕ 0 ⊕ 0. Because L1 is torsion free, the following holds.
18
1 denotes the transpose of Γ1.
1 · h is infinite for all h ∈ H2 \ {0}, where ΓT
• Γ1 · g is infinite for all g ∈ H1 \ H (1)
1 .
• ΓT
From these observations, it follows that L(H (1)
1 ) is equal to the algebra of Γ1-invariant elements
in A0 and that L(H (1)
1 ) is also equal to the algebra of elements in A0 that are fixed by some
finite index subgroup of Γ1. Since both B0 and Λ0 are globally Γ0-invariant, it follows that
L(H (1)
We similarly consider H (2)
1 ) =
B(i)
0 ⋊ Λ(i)
generate H1 and H1 is abelian.
So, "everything" commutes and we conclude that L(H1) = B1 ⋊ Λ1 for some von Neumann
subalgebra B1 ⊂ B0 and subgroup Λ1 < Λ0.
for some von Neumann subalgebra B(1)
0
1 = 0 ⊕ L1 ⊕ 0 and H (3)
1 = 0 ⊕ 0 ⊕ L1. We conclude that L(H (i)
for all i = 1, 2, 3. The subgroups H (1)
and subgroup Λ(1)
1 , H (2)
and H (3)
1 ) = B(1)
0 ⋊ Λ(1)
0
0 < Λ0.
0
1
1
A similar reasoning applies to L∞(cH2) = L(H2) and we get that L(H2) = B2 ⋊ Λ2 for B2 ⊂ B
and Λ2 < Λ0.
Since L(H1)L(H2) is k · k2-dense in A0 and L(H1)∩L(H2) = C1, we get that Λ1Λ2 = Λ0 = Λ2Λ1
and Λ1 ∩ Λ2 = {e}. It then follows that for all bi ∈ Bi and si ∈ Λi, i = 1, 2, we have that
EB0(b1vs1b2vs2) equals zero unless s1 = e and s2 = e, in which case, we get b1b2. We conclude
that B1B2 is k · k2-dense in B0.
For every x ∈ L(Hi) and g ∈ Hi, we denote by (x)g = τ (xu∗
Comparing Fourier decompositions, we get for all xi ∈ L(Hi) that
g) the g-th Fourier coefficient of x.
x1x2 = x2x1 iff Ω(g, h) = 1 whenever g ∈ H1, h ∈ H2, (x1)g 6= 0, (x2)h 6= 0 .
(6.1)
Since Bi ⊂ L(Hi) and since B1, B2 commute, we obtain from (6.1) subgroups Si ⊂ Hi such
that Ω(g, h) = 1 for all g ∈ S1, h ∈ S2 and such that Bi ⊂ L(Si). Since B1B2 is dense in B0, it
follows that B0 ⊂ L(S1) ∨ L(S2). Since L(S1) ∨ L(S2) is abelian and B0 is maximal abelian, we
conclude that B0 = L(S1) ∨ L(S2). Thus, Bi = L(Si) for i = 1, 2. When g ∈ S⊥
2 , the unitary ug
commutes with L(S2), but also with L(S1) because L(S1) = B1 ⊂ L(H1) and L(H1) is abelian.
Since B0 is maximal abelian, we get that g ∈ S1. So, S1 = S⊥
The next step of the proof is to show that Λ0 is abelian, i.e. that Λ1 and Λ2 are commuting
subgroups of Λ0. Put Ti = Hi/Si. Since S1 = S⊥
1 , we have the canonical dense
2 and similarly S2 = S⊥
1 .
2 and S2 = S⊥
embeddings T1 ֒→ cS2 and T2 ֒→ cS1. Viewing L∞(cS1 × cS2) = L(S1) ∨ L(S2) as a Cartan
subalgebra of A0, the associated equivalence relation is given by the orbits of the action
T1 × T2 y cS1 × cS2 : (g, h) · (y, z) = (h · y, g · z) ,
viewing B0 = L(S1) ∨ L(S2), the same equivalence relation is given by the orbits of the action
where the actions on the right, namely T1 y cS2 and T2 y cS1, are given by translation. But
Λ0 y cS1 × cS2. We denote by ω : (T1 × T2) × (cS1 × cS2) → Λ0 the associated 1-cocycle. By
in L∞(cS1 ×cS2) given by the set
construction, for all g ∈ H1, h ∈ H2 and s ∈ Λ0, the support of EB(v∗
s uguh) is the projection
{(y, z) ∈ cS1 ×cS2 ω((gS1, hS2), (y, z)) = s} .
Since L(H1) = B1 ⋊ Λ1, for all g ∈ T1, the map (y, z) 7→ ω((g, e), (y, z)) only depends on the
first variable and takes values in Λ1 a.e. Reasoning similarly for h ∈ T2, we find ωi : Ti×bSi → Λi
such that
ω((g, e), (y, z)) = ω1(g, y)
and ω((e, h), (y, z)) = ω2(h, z)
a.e.
Writing (g, h) = (g, e)(e, h) and (g, h) = (e, h)(g, e), the 1-cocycle relation implies that
ω1(g, h · y) ω2(h, z) = ω2(h, g · z) ω1(g, y)
(6.2)
19
for all g ∈ T1, h ∈ T2 and a.e. y ∈ cS1, z ∈ cS2.
Define the subgroup G1 < Λ1 given by
G1 = {s ∈ Λ1 ∀t ∈ Λ2, tst−1 ∈ Λ1} .
Similarly, define G2 < Λ2. Note that G1 and G2 are normal subgroups of Λ0. Since Λ0 = Λ1Λ2
and Λ1 ∩ Λ2 = {e}, we also have that G1 and G2 commute. Rewriting (6.2) as
ω1(g, y) ω2(h, z) = ω2(h, g · z) ω1(g, h−1 · y) ,
we find that for all g ∈ T1, h ∈ T2 and a.e. y, y′ ∈ cS1, z ∈ cS2
ω2(h, z)−1 ω1(g, y′)−1 ω1(g, y) ω2(h, z) ∈ Λ1 .
Since L(H2) = B2 ⋊ Λ2, the essential range of ω2 equals Λ2. It thus follows that
ω1(g, y′)−1 ω1(g, y) ∈ G1
for all g ∈ T1 and a.e. y, y′ ∈ cS1. For every g ∈ T1, we choose δ1(g) ∈ Λ1 such that ω1(g, y) =
δ1(g) on a non-negligible set of y ∈ cS1. We conclude that ω1(g, y) = δ1(g) µ1(g, y) with
µ1(g, y) ∈ G1 a.e. We similarly decompose ω2(h, z) = δ2(h) µ2(h, z).
With these decompositions of ω1 and ω2 and using that G1, G2 are commuting normal subgroups
of Λ0, it follows from (6.2) that for all g ∈ T1, h ∈ T2, the commutator δ2(h)−1δ1(g)−1δ2(h)δ1(g)
belongs to G1G2, so that it can be uniquely written as η1(g, h)η2(g, h)−1 with ηi(g, h) ∈ Gi. It
then follows from (6.2) that
µ1(g, h · y) = δ2(h) η1(g, h) µ1(g, y) δ2(h)−1 ,
µ2(h, g · z) = δ1(g) η2(g, h) µ2(h, z) δ1(g)−1 ,
(6.3)
closed finite index subgroups. It follows that finite index subgroups of T2 act ergodically on
almost everywhere. Since S1 < H1 is torsion free, cS1 has no finite quotients and thus no proper
cS1. We claim that for every g ∈ T1, the map y 7→ µ1(g, y) is essentially constant. To prove
this claim, fix g ∈ T1 and denote ξ : cS1 → G1 : ξ(y) = µ1(g, y). For every h ∈ T2, define the
permutation
ρh : G1 → G1 : ρh(s) = δ2(h) η1(g, h) s δ2(h)−1 .
the essential range of ξ, it follows that {ρh}h∈T2 is an action of T2 on V1. The push forward
So, (6.3) says that ξ(h · y) = ρh(ξ(y)) for all h ∈ T2 and a.e. y ∈ cS1. Defining V1 ⊂ G1 as
via ξ of the Haar measure on cS1 is a {ρh}h∈T2-invariant probability measure on the countable
Choosing s ∈ V1, the set ξ−1({s}) ⊂ cS1 is non-negligible and globally invariant under a finite
index subgroup of T2. It follows that ξ(y) = s for a.e. y ∈ cS1, thus proving the claim.
set V1 and has full support. It follows that all orbits of the action {ρh}h∈T2 on V1 are finite.
Similarly, for every h ∈ T2, the map z 7→ µ2(h, z) is essentially constant. So we have proved
that ω1(g, y) = δ1(g) and ω2(h, z) = δ2(h) a.e. But then, (6.2) implies that Λ1 and Λ2 commute,
so that Λ0 is an abelian group.
Since A0 is a factor, BΛ0
0 = C1 and thus L(Λ0) ⊂ A0 is maximal abelian. Since L(Λ0) =
L(Λ1) ∨ L(Λ2) with L(Λi) ⊂ L(Hi), the same reasoning as with Bi ⊂ L(Hi), using (6.1), gives
us subgroups Ti ⊂ Hi such that L(Λi) = L(Ti) and T1 = T ⊥
1 . Since L(Hi) = Bi ⋊ Λi
with Bi = L(Si) and L(Λi) = L(Ti), we get that Hi = Si ⊕ Ti.
So far, we have proved that B0 = L(S1) ∨ L(S2) and L(Λ0) = L(T1) ∨ L(T2). In any crossed
product B0 ⋊ Λ0 by a faithful action, the only unitaries in L(Λ0) that normalize B0 are the
2 , T2 = T ⊥
20
multiples of the canonical unitaries {vs}s∈Λ. Therefore, TT1T2 = TΛ0. We have thus proved
that the gms decomposition (B0, Λ0) is identical to the gms decomposition (L(S1)∨L(S2), T1T2).
Since Λ0, H1 and H2 are globally {βg}g∈Γ0-invariant, it follows that Ti is a globally SL(3, Z)-
invariant subgroup of Hi. Thus, Ti = Q3
i for some subgroup Qi < Li. Since B0, H1 and H2
are globally Γ0-invariant, it follows in the same way that Si = P 3
for some subgroups Pi < Li.
i
Then, Li = Pi ⊕ Qi and P1, P2, as well as Q1, Q2, are each other's orthogonal complement
under Ω. So, L1 = P1 ⊕ Q1 is an admissible direct sum decomposition.
We now combine Proposition 6.1 with Theorem 5.1 and Proposition 5.14. We fix once and for all
Γ = F∞, Γ0 = SL(3, Z) and a surjective homomorphism β : Γ ։ Γ0 so that the automorphism
g 7→ (g−1)T of Γ0 lifts to an automorphism of Γ. An obvious way to do this is by enumerating
Γ0 = {g0, g1, . . .} and defining β : Γ → Γ0 by β(si) = gi for i ≥ 0, where (si)i∈N are free
generators of Γ. Note that Ker β is automatically nontrivial.
We also fix countable abelian torsion free groups L1, L2 and a dense embedding L1 ֒→ cL2. Put
i and let Γ act on H1 ֒→ cH2 through β. Then define A0 = L∞(cH2) ⋊ H1 together with
Hi = L3
the natural action β : Γ y (A0, τ0). Put (A, τ ) = (A0, τ0)Γ with the action Γ × Γ y (A, τ )
given by (g, h) · πk(a) = πgkh−1(βh(a)) for all g, h, k ∈ Γ, a ∈ A0. Write M = A ⋊ (Γ × Γ).
We call an isomorphism θ : L1 → L2 admissible if it extends to a continuous isomorphism
We call an automorphism of L1 admissible if it extends to a continuous automorphism of cL2.
cL2 → cL1.
Theorem 6.2. Whenever L1 = P1 ⊕ Q1 is an admissible direct sum decomposition with
corresponding L2 = P2 ⊕ Q2, we define B(P1, Q1) := (L(P 3
2 ))Γ and Λ(P1, Q1) =
(Q3
1 ) ∨ L(P 3
1 ⊕ Q3
2)(Γ) ⋊ (Γ × Γ).
• Every B(P1, Q1), Λ(P1, Q1) gives a gms decomposition of M .
• Every gms decomposition of M is unitarily conjugate with a B(P1, Q1), Λ(P1, Q1) for a
unique admissible direct sum decomposition L1 = P1 ⊕ Q1.
• Let L1 = P1 ⊕ Q1 and L1 = P ′
1 ⊕ Q′
1 be two admissible direct sum decompositions with
associated gms decompositions (B, Λ) and (B′, Λ′).
-- (B, Λ) and (B′, Λ′) are conjugate by an automorphism of M if and only if there exists an
1, or an admissible
admissible automorphism θ : L1 → L1 with θ(P1) = P ′
1, θ(Q1) = Q′
isomorphism θ : L1 → L2 with θ(P1) = P ′
2, θ(Q1) = Q′
2.
-- The Cartan subalgebras B and B′ are unitarily conjugate if and only if P1 = P ′
1.
-- The Cartan subalgebras B and B′ are conjugate by an automorphism of M if and only if
1 or an admissible
there exists an admissible automorphism θ : L1 → L1 with θ(P1) = P ′
isomorphism θ : L1 → L2 with θ(P1) = P ′
2.
Proof. Because of Proposition 6.1, Theorem 5.1 and Proposition 5.14, it only remains to de-
scribe all automorphisms ψ : A0 → A0 that normalize the action β : Γ y A0. This action β is
defined through the quotient homomorphism Γ ։ Γ0. Every automorphism of Γ0 = SL(3, Z)
is, up to an inner automorphism, either the identity or g 7→ (g−1)T . So, we only need to
describe all automorphisms ψ : A0 → A0 satisfying either ψ ◦ βg = βg ◦ ψ for all g ∈ Γ0, or
ψ ◦ βg = β(g−1)T ◦ ψ.
In the first case, reasoning as in the first paragraphs of the proof of Proposition 6.1, we get
that ψ(L(Hi)) = L(Hi) for i = 1, 2. So, for every g ∈ H1, ψ(ug) is a unitary in L(H1) that
normalizes L(H2). This forces ψ(ug) ∈ TH1 and we conclude that ψ(TH1) = TH1. Similarly,
In the second case, we obtain in the same way that ψ(TH1) = TH2 and
ψ(TH2) = TH2.
21
ψ(TH2) = TH1. The further analysis is analogous in both cases and we only give the details
of the first case.
We find automorphisms θi : Hi → Hi such that ψ(ug) ∈ Tuθi(g) for all i = 1, 2 and g ∈ Hi.
Since θi commutes with the action of SL(3, Z) on Hi, we get that θ1 = θ3, where θ : L1 → L1
is an admissible automorphism. It follows that ψ maps the gms decomposition associated with
L1 = P1 ⊕ Q1 to the gms decomposition associated with L1 = θ(P1) ⊕ θ(Q1). This concludes
the proof of the theorem.
The following concrete examples provide a proof for Theorem A.
1 , αk1
2 , . . . , αkn
Theorem 6.3. For all n ≥ 1, consider the following two embeddings πi : Zn ֒→ T2n.
• π1(k) = (αk1
2n) for rationally independent irrational angles αj ∈ T.
• π2(k) = (αk1, βk1, . . . , αkn, βkn) for rationally independent irrational angles α, β ∈ T.
Applying Theorem 6.2 to the embeddings π1 and π2, we respectively obtain
• a II1 factor M that has exactly 2n gms decompositions up to unitary conjugacy, and with
2n−1, αkn
the associated 2n Cartan subalgebras non conjugate by an automorphism of M ;
• a II1 factor M that has exactly n+1 gms decompositions up to conjugacy by an automorphism
of M , and with the associated n + 1 Cartan subalgebras non conjugate by an automorphism
of M .
Proof. Whenever F ⊂ {1, . . . , n}, we have the direct sum decomposition Zn = P (F) ⊕ P (F c)
where P (F) = {x ∈ Zn ∀i 6∈ F, xi = 0}.
In the case of π1, these are exactly all the admissible direct sum decompositions of Zn. Also, the
only admissible automorphisms of Zn are the ones of the form (x1, . . . , xn) 7→ (ε1x1, . . . , εnxn)
with εi = ±1. Since Zn 6∼= Z2n, there are no isomorphisms "exchanging L1 and L2".
In the case of π2, all direct sum decompositions and all automorphisms of Zn are admissible.
For every direct sum decomposition Zn = P1 ⊕ Q1, there exists a unique k ∈ {0, . . . , n}
and an automorphism θ ∈ GL(n, Z) such that θ(P1) = P ({1, . . . , k}) and θ(Q1) = P ({k +
1, . . . , n}). Again, there are no isomorphisms exchanging L1 and L2. So, the n + 1 direct sum
decompositions Zn = P ({1, . . . , k}) ⊕ P ({k + 1, . . . , n}), 0 ≤ k ≤ n, exactly give the possible
gms decompositions of M up to conjugacy by an automorphism of M . When k 6= k′, there
is no isomorphism θ ∈ GL(n, Z) with θ(P ({1, . . . , k})) = P ({1, . . . , k′}). Therefore, the n + 1
associated Cartan subalgebras are nonconjugate by an automorphism either.
Remark 6.4. When L1, L2 are torsion free abelian groups and L1 ֒→ cL2 is a dense embedding,
then the set of admissible homomorphisms L1 → L1 is a ring R that is torsion free as an additive
group. The admissible direct sum decompositions of L1 are in bijective correspondence with
the idempotents of R. As a torsion free ring, R either has infinitely many idempotents, or
finitely many that are all central, in which case their number is a power of 2. So, the number
of gms decompositions (up to unitary conjugacy) of the II1 factors produced by Theorem 6.2
is always either infinite or a power of 2.
Remark 6.5. Still in the context of Theorem 6.2, we call a subgroup P1 < L1 admissible if
L1 ∩ P1 = P1, where P1 denotes the closure of P1 inside cL2. Note that P1 < L1 is admissible
if and only if there exists a subgroup P2 < L2 such that P2 = P ⊥
2 . Whenever
2 ))Γ. It is easy to check
P1 < L1 is an admissible subgroup, we define B(P1) := (L(P 3
that all B(P1) are Cartan subalgebras of M and that B(P1) is unitarily conjugate with B(P ′
1)
if and only if P1 = P ′
1.
1 and P1 = P ⊥
1 ) ∨ L(P 3
22
It is highly plausible that these B(P1) describe all Cartan subalgebras of M up to unitary
conjugacy. We could however not prove this because all our techniques make use of the dual
coaction associated with a gms decomposition of M .
Also note that an admissible subgroup P1 < L1 cannot necessarily be complemented into an
admissible direct sum decomposition L1 = P1⊕Q1. In such a case, B(P1) is a Cartan subalgebra
of M that is not of group measure space type. However, M can be written as a cocycle crossed
⋊ (Γ × Γ), but the cocycle
product of B(P1) by an action of Λ(P1) = (cid:0)(L1/P1)3 ⊕ (L2/P2)3(cid:1)(Γ)
is nontrivial.
References
[BO08]
[BV14]
[Co75]
[CJ82]
[CIK13]
[CdSS15]
[Io10]
[Io11]
[IPV10]
[OP03]
[OP07]
[OP08]
[Pe09]
[Po01]
[Po03]
[Po04]
[Po06]
[PV09]
[PV11]
[PV12]
[Si55]
N. P. Brown and N. Ozawa, C∗-algebras and finite-dimensional approximations. Graduate
Studies in Mathematics 88. American Mathematical Society, Providence, 2008.
M. Berbec and S. Vaes, W∗-superrigidity for group von Neumann algebras of left-right wreath
products. Proc. Lond. Math. Soc. 108 (2014), 1113-1152.
A. Connes, Classification of injective factors. Ann. of Math. 104 (1976), 73-115.
A. Connes and V.F.R. Jones, A II1 factor with two non-conjugate Cartan subalgebras, Bull.
Amer. Math. Soc. 6 (1982), 211-212.
I. Chifan, A. Ioana and Y. Kida, W∗-superrigidity for arbitrary actions of central quotients
of braid groups. Math. Ann. 361 (2015), 563-582.
I. Chifan, R. de Santiago and T. Sinclair, W∗-rigidity for the von Neumann algebras of
products of hyperbolic groups. Geom. Funct. Anal. 26 (2016), 136-159.
A. Ioana, W∗-superrigidity for Bernoulli actions of property (T) groups. J. Amer. Math. Soc.
24 (2011), 1175-1226.
A. Ioana, Uniqueness of the group measure space decomposition for Popa's HT factors.
Geom. Funct. Anal. 22 (2012), 699-732.
A. Ioana, S. Popa and S. Vaes, A class of superrigid group von Neumann algebras. Ann.
Math. 178 (2013), 231-286.
N. Ozawa and S. Popa, Some prime factorization results for type II1 factors. Invent. Math.
156 (2004), 223-234.
N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra I. Ann.
Math. 172 (2010), 713-749.
N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra II.
Amer. J. Math. 132 (2010), 841-866.
J. Peterson, Examples of group actions which are virtually W∗-superrigid. Preprint.
arXiv:1002.1745
S. Popa, On a class of type II1 factors with Betti numbers invariants. Ann. of Math. 163
(2006), 809-899.
S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, I.
Invent. Math. 165 (2006), 369-408.
S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, II.
Invent. Math. 165 (2006), 409-452.
S. Popa, On the superrigidity of malleable actions with spectral gap. J. Amer. Math. Soc.
21 (2008), 981-1000.
S. Popa and S. Vaes, Group measure space decomposition of II1 factors and W∗-superrigidity.
Invent. Math. 182 (2010), 371-417.
S. Popa and S. Vaes, Unique Cartan decomposition for II1 factors arising from arbitrary
actions of free groups. Acta Math. 212 (2014), 141-198.
S. Popa and S. Vaes, Unique Cartan decomposition for II1 factors arising from arbitrary
actions of hyperbolic groups. J. Reine Angew. Math. 694 (2014), 215-239.
I. M. Singer, Automorphisms of finite factors. Amer. J. Math. 77 (1955), 117-133.
23
[SV11]
[Va10]
A. Speelman and S. Vaes, A class of II1 factors with many non conjugate Cartan subalgebras.
Adv. Math. 231 (2012), 2224-2251.
S. Vaes, One-cohomology and the uniqueness of the group measure space decomposition of
a II1 factor. Math. Ann. 355 (2013), 611-696.
24
|
1111.6183 | 1 | 1111 | 2011-11-26T18:46:28 | Some explicit computations and models of free products | [
"math.OA"
] | In this note, we first work out some `bare hands' computations of the most elementary possible free products involving $\mathbb{C}^2 ~(=\mathbb{C} \oplus \mathbb{C} $) and $M_2 ~(= M_2(\mathbb{C}))$. Using these, we identify all free products $C \ast D$, where $C,D$ are of the form $A_1 \oplus A_2$ or $M_2(B)$; $A_1,A_2,B$ are finite von Neumann algebras, as is $A_1 \oplus A_2$ with the 'uniform trace' given by $tr(a_1, a_2) = 1/2 (tr(a_1) + tr(a_2))\}$ and $M_2(B)$ with the normalized trace given by $tr((b_{i,j}))=1/2(tr(b_{1,1}) + tr(b_{2,2}))$. Those results are then used to compute various possible free products involving certain finite dimensional von-Neumann algebras, the free-group von-Neumann algebras and the hyperfinite $II_1$ factor. In the process, we reprove Dykema's result `$R \ast R \cong LF_2$'. | math.OA | math | Some explicit computations and models of free products
Madhushree Basu
November 7, 2018
Abstract
In this note, we first work out some 'bare hands' computations of the most elemen-
tary possible free products involving C2 (= C ⊕ C) and M2 (= M2(C)). Using these,
we identify all free products C ∗ D, where C, D are of the form A1 ⊕ A2 or M2(B);
A1, A2, B are finite von Neumann algebras, as is A1 ⊕ A2 with the 'uniform trace'
given by tr(a1, a2) = 1
2 (tr(a1) + tr(a2))} and M2(B) with the normalized trace given
by tr((bi,j)) = 1
2 (tr(b1,1) + tr(b2,2)). Those results are then used to compute various
possible free products involving certain finite dimensional von-Neumann algebras, the
free-group von-Neumann algebras and the hyperfinite II1 factor. In the process, we
reprove Dykema's result 'R ∗ R ∼= LF2'.
1
Introduction
The motivation behind the computations done in this paper comes from trying to understand
[D1] and [D2]. Most of the results here have been proved in those two papers. Further, our
proofs seem to constrain us to only regard direct sums and matrix algebras equipped with
the 'uniform trace', and consequently to results where powers of 2 keep cropping up. The
flip side of the coin is that our proofs involve nothing more than basic trigonometry and
simple cumulant calculations and require little prior background to follow.
We shall work with the model (L∞([0, π/2], 2
0 ·dt) ∗ LZ ∗ LZ) of LF3. Let u and v be
Haar unitaries generating the two LZs, and let c, s ∈ L∞([0, π/2]) be the functions defined
by c(θ) = cos θ and s(θ) = sin θ respectively.
2
πR π
Let the trace and cumulant on M2(LF3), resp., LF3, be denoted by T r and κ, resp. tr
1
1
0
2
v
o
N
6
2
]
.
A
O
h
t
a
m
[
1
v
3
8
1
6
.
1
1
1
1
:
v
i
X
r
a
and k respectively.
Then U and X are partial isometries satisfying U ∗U + UU ∗ = 1 and X ∗X + XX ∗ = 1;
0 0(cid:19), V =(cid:18)0 v
Let U =(cid:18)0 u
Let P = UU ∗ = (cid:18)1 0
0 0(cid:19), W =(cid:18)c −s
0 0(cid:19) and Q = XX ∗ = (cid:18)c2
−svs svc(cid:19).
c (cid:19), and let X = W V W ∗ =(cid:18)−cvs cvc
cs s2(cid:19) = W P W ∗. Then P, Q are
so each of them generates a copy of M2.
cs
s
projections of trace 1
2, each generating a copy of C2.
1
Finally, for finite von-Neumann algebra A, B and k ∈ N, we often use the notations Ak
and Mk(B) for A ⊕ A ⊕ · · · ⊕ A(k times) and Mk(C) ⊗ B respectively, with the trace always
being taken as the 'uniform trace'.
2 C2 ∗ C2 ∼= M2(LZ)
Lemma 2.1.
W ∗({P, Q}) = M2(LZ)
0(cid:19). Since cs is positive (in L∞([0, π
Proof. Note that P Q(1−P ) =(cid:18)0 cs
we see that the polar decomposition of P Q(1 − P ) is given by(cid:18)0 cs
0
0
2 ])) and has no kernel,
0(cid:19) =(cid:18)0 1
0 0(cid:19)(cid:18)0
0 cs(cid:19) .
0
Thus W ∗({P, Q}) contains all the matrix units, viz., P = e11, 1 − P = e22, e12 = polar part of
P Q(1 − P ) and e21 = e∗
12. On the other hand W ∗({P, Q}) clearly contains P QP = e11 ⊗ c2,
and hence also e11 ⊗W ∗({c2}); therefore W ∗({P, Q}) ⊃ e11 ⊗L∞([0, π
2 ]). Finally, by pre- and
post-multiplying by appropriate matrix units, we see that W ∗({P, Q}) ⊃ eij ⊗ L∞([0, π
2 ]) for
all i, j, and the proof of the lemma is complete.
Lemma 2.2. P and Q are free in M2(LF3).
Proof. Let P0 = P − 1/2 =(cid:18)1/2
0 −1/2(cid:19) and Q0 = Q − 1/2 =(cid:18)c2 − 1/2
cs
0
cs
s2 − 1/2(cid:19) be
sn, for the elements of L∞([0, π
the trace-less versions (i.e., translates with trace 0) of P and Q. We shall find it convenient
to write cn, resp.
2 ]) defined by cn(θ) = cos nθ, resp.,
sn(θ) = sin nθ.
We are going to verify that the trace of any alternating product in 2P0 =(cid:18)1
2Q0 =(cid:18)c2
0 −1(cid:19) and
s2 −c2(cid:19)is zero. Since we are working with a trace here, it is enough to prove that
T r((2P0.2Q0)r) = 0 and T r((2Q0.2P0)r2Q0) = 0.
s2
0
However,
whereas
0
s2
s2
−s2
(2P0.2Q0)r = (cid:18)(cid:18)1
= (cid:18)(cid:18) c2
= (cid:18) c2r
(2Q0.2P0)r2Q0 = (cid:18) c2r
= (cid:18)c2r+2
0 −1(cid:19)(cid:18)c2
s2 −c2(cid:19)(cid:19)r
c2(cid:19)(cid:19)r
c2r(cid:19)
0 −1(cid:19)
c2r(cid:19)(cid:18)1
s2r+2 −c2r+2(cid:19) ,
s2r+2
−s2r
−s2r
s2r
s2r
0
2
and the desired assertions follows from
2
2
πZ π
0
c2kdθ =
2
π
1
2k
(s2k(π) − s2k(0))
= 0 ∀k ∈ N
Now the validity of Proposition 2.3 follows immediately from Lemmas 2.1 and 2.2.
Proposition 2.3.
C2 ∗ C2 ∼= M2(LZ)
3 M2 ∗ M2
∼= M2(LF3)
En route to proving the main result, viz., Proposition 3.2, we shall also prove Proposition
3.1.
Proposition 3.1.
Proposition 3.2.
M2 ∗ C2 ∼= M2(LF2)
M2 ∗ M2
∼= M2(LF3)
We start with the following crucial lemma.
Lemma 3.3. U and X are ∗- free in M2(LF3).
Proof. Observe the following simple facts:
1. U 2 = 0 = (U ∗)2, V 2 = 0 = (V ∗)2
2. P 2
0 , Q2
0 ∈ C
3. P U = U, QX = X, UP = 0, XQ = 0
4. P, Q are free in M2(L∞)
In view of the the above relations, the sets of trace zero words in W ∗({U}) and W ∗({X})
are linearly generated by A = {U, U ∗, 2P0} and B = {X, X ∗, 2Q0} repectively. So we need
to check that every alternating product of elements from A and B has trace zero.
We may dispose of the 'trivial case' when Π is an alternating word in only P0 and Q0,
since that is covered by Lemma 2.2.
Consider a typical such product, say Π; we will prove that tr(Πij) = 0 for all i, j (for the
non-trivial case). Thus in particular we will have T r(Π) = tr(Π1,1)+tr(Π2,2)
2
= 0
3
One can see that each Πij is a sum of elements of the form
ω = f0(c, s)w0f1(c, s)w1f2(c, s) · · · wn−1fn(c, s) · · ·
(3.1)
where the fis are (possibly constant) polynomials in {c, s} and wi ∈ {u, u∗, v, v∗}. The
assumption that we are not dealing with the trivial case implies that there must exist at
least one wi in the string in the right hand side of equation 3.1.
Now, fix any such word ω is (of total length m, say) as in equation 3.1, which occurs as
a summand of Π; then
tr(ω) = Xσ∈N C(m)
kσ[f0(c, s), w0, f1(c, s), · · · , wn−1, fn(c, s), · · · ] .
We shall show that kσ(= kσ[f0(c, s), w0, f(c, s), · · · , wn−1, fn(c, s), · · · ]) = 0 for each σ ∈
NC(m), for each such ω.
But u and v are free Haar unitaries and hence R-diagonal. So following Proposition 15.1
in [NS], in order for kσ to be possibly non zero, it must be the case that the blocks of σ must
consist of either only {u, u∗} in alternate positions, or only {v, v∗} in alternate positions (in
each block the number of u = number of u∗ and same for v, v∗), or only {fi(c, s) : i}, all
occurring in a non-crossing fashion. Note that in effect ω must have the same number of u
and u∗ as well as the same number of v and v∗.
Example 3.4. ω0 = u(cs)v (cs(c2 − 1/2)s) v∗(c2s)u∗ = uf1(c, s)vf2(c, s)v∗f3(c, s)u∗ - with
m = 7 - can possibly give non zero cumulant only corresponding to two elements of NC(7),
namely
((1, 7), (3, 5), (2), (4), (6)) and ((1, 7), (3, 5), (2, 6), (4)) .
(Here, f1(c, s) = cs, f2(c, s) = cs(c2 − 1/2)s, f3(c, s) = c2s)
As at least one wi and necessarily also w∗
i must occur in the string defining ω, and σ is
non-crossing, we see that kσ can hope to be non-zero only if ω has a substring of the form
wf (c, s)w∗, with w ∈ {u, u∗, v, v∗} and f (c, s) as above, and with w and w∗ in the same
block of σ; thus ω must contain one of the one of the following four substrings:
1. uf (c, s)u∗
2. u∗f (c, s)u
3. vf (c, s)v∗
4. v∗f (c, s)v
Now consider the above four cases in the following way:
1. uf (c, s)u∗ can occur as a string in some summand ω of Πij only if USU ∗ occurs as
a substring in the alternating product expression of Π, where S = (2Q02P0)r2Q0.
Observe in this case that USU ∗ =(cid:18)−uc2r+2u∗ 0
0(cid:19) and that f (c, s) = −c2r+2.
0
4
2. Similarly u∗f (c, s)u can occur as a string in some summand ω of Πij only if U ∗SU occurs
as a substring in the alternating product expression of Π, where S = (2Q02P0)r2Q0.
Observe in this case that U ∗SU =(cid:18)0
0 u∗c2r+2u(cid:19) and that f (c, s) = c2r+2.
0
3. Similarly vf (c, s)v∗ can occur as a string in some summand ω of Πij only if XSX ∗ oc-
curs as a substring in the alternating product expression of Π, where S = (2P02Q0)r2P0.
Observe in this case that XSX ∗ = (cid:18)−cvc2r+2v∗c
−svc2r+2v∗c −svc2r+2v∗s(cid:19) and that again,
cvc2r+2v∗s
f (c, s) = ±c2r+2.
4. Similarly v∗f (c, s)v can occur as a string in some summand ω of Πij only if X ∗SX oc-
curs as a substring in the alternating product expression of Π, where S = (2P02Q0)r2P0.
Observe in this case that X ∗SX = (cid:18) sv∗c2r+2vs −sv∗c2r+2vc
cv∗c2r+2vc (cid:19) and that again,
−cv∗c2r+2vs
f (c, s) = ±c2r+2.
Thus, in any case, we find that for any σ ∈ NC(m) for which kσ can possibly be non-
zero, it must be the case that ω must contain a substring of the form wc2r+2w∗ with w ∈
{u, u∗, v, v∗} and with w and w∗ in the same block of σ. Since W ∗({c}) and W ∗({w}) are
free, we see that the only way kσ can be non-zero is for {c2r+2} to be a block of σ; but then
kσ has k(c2r+2) = tr(c2r+2) = 0 as a factor. Hence, inded, every kσ = 0, as asserted, and the
proof is complete.
Corollary 3.5. P and X are free in M2(LF3).
Proof. This follows from P ∈ W ∗({U}).
Lemma 3.6.
W ∗({P, X}) = M2(LF2)
2 ]) ⊂ N .
Proof. Since Q = XX ∗, it follows from Lemma 2.1 that W ∗({P, X}) contains each matrix
unit eij. It follows that W ∗({P, X}) contains M2(N ) where N is the von Neumann algebra
generated by the entries of X. By the last line in the proof of Lemma 2.1 shows that
L∞([0, π
Consider the bounded Borel functions fn defined on [0, 1] by fn(t) = (cid:26) 1
observe that fn(c)c converges strongly to 1 (since c is injective). Since cvc ∈ N , deduce
that v = limn(fn(c)cvcfn(c)) ∈ N . Since c and v are ∗-free, and c, v ∈ N , it follows
that N ⊃ LF2. Since the entries of P and X all lie in LF2, the proof of the Lemma is
complete.
t ≥ 1
t
n
0 t < 1
n
;
Remark 3.7. We can prove the above lemma also by proving W ∗({U, Q}) = M2(LF2) (where
U and Q are free), in an exactly similar way.
Lemma 3.8.
W ∗({U, X}) = M2(LF3)
5
Proof. In view of Lemma 3.6, we only need to observe that {u, v, c}′′ = LF3.
Finally, Proposition 3.1 follows from Lemma 3.6 and Corollary 3.5, while Proposition 3.2
follows from Lemma 3.8 and Lemma 3.3.
4
(A1 ⊕ A2) ∗ (B1 ⊕ B2) ∼= M2(A1 ∗ A2 ∗ B1 ∗ B2 ∗ LZ)
Let A1, A2 and B1, B2 be finite von-Neumann algebras. We denote the trace and cumulant
on matrix algebras over the finite von-Neumann algebras as T r and κ and those on the finite
von-Neumann algebras themselves as tr and k respectively.
We start by proving a simple but useful lemma on a certain property of the Kreweras
compliment of a non-crossing partition in NC(n) for any n ∈ N.
Lemma 4.1. Let π ∈ NC(n) and 1 ∼π n. Let V = (k′, · · · , (k + l)′) be an interval in its
Kreweras compliment K(π) for 1 ≤ k ≤ n, 0 ≤ l ≤ n − k. Then k ∼π (k + l + 1), where all
positive integers are taken modulo n.
Proof. Note that if (k) is a singloton block then (k − 1)′ ∼K(π) k′, a contradiction since V
is an interval. Similarly (k + l + 1) cannot be a singleton block. More generally suppose
rk ∈ {1, · · · , k} and sk ∈ {k, (k + l + 1), · · · , n} are minimum positive integers such that
k ∼π rk and k ∼π sk. Suppose rk+l+1 ∈ {1, · · · , k, (k+l+1)} and sk+l+1 ∈ {(k+l+1), · · · , n}
are the same for (k + l + 1). We already saw that we cannot have rk = sk = k or rk+l+1 =
sk+l+1 = (k + l + 1). For simplicity's sake we assume 1 < k, l < (n − k). The cases k = 1 or
l = n − k will follow similarly.
Case sk (cid:13) k: In this case sk being minimum in {(k + l + 1), · · · , n} such that k ∼π sk we
must have (sk − 1)′ ∼K(π) (k + l)′, a contradiction unless k + l + 1 = sk ∼π k.
Case sk = k: In this case rk being minimum in {1, · · · , k − 1} such that k ∼π rk, unless
rk = 1 we must have (rk − 1)′ ∼K(π) k′, a contradiction. On the other hand if rk = 1, then
rk+l+1 is forced to be (k + l + 1). Thus similarly we must have sk+l+1 = n, otherwise leading
into a contradiction. But 1 ∼π n. Hence k ∼π rk = 1 ∼π n = sk+l+1 ∼π (k + l + 1)
Proposition 4.2.
(A1 ⊕ A2) ∗ (B1 ⊕ B2) ∼= M2(A1 ∗ A2 ∗ B1 ∗ B2 ∗ LZ)
Proof. Consider the two matrix subalgebras(cid:18)A1
0
B1 ⊕ B2) of A1 ∗ A2 ∗ B1 ∗ B2 ∗ LZ, where W is the unitary matrix of the previous sections.
Following the method used in Lemma 3.6 we know that these two subalgebras indeed generate
the matrix algebra on the right side. We need to show that these are free.
0 A2(cid:19) (∼= A1 ⊕A2) and W(cid:18)B1
0 B2(cid:19) W ∗(∼=
0
6
0
Note that W(cid:18)b1
Suppose Π is an alternating product of matrices of the form(cid:18)a1
sb1c − cb2s sb1s + cb2c(cid:19)
b2(cid:19) W ∗ =(cid:18)cb1c + sb2s cb1s − sb2c
0
0
with aj ∈ Aj, bj ∈ Bj, tr(a1 + a2) = 0 = tr(b1 + b2).
0
a2(cid:19) and W(cid:18)b1
0
0
b2(cid:19) W ∗,
Instead of directly proving T r(Π) = 0, we will prove a stronger statement: tr(Πi1,i1) =
0∀i1 ∈ {1, 2}.
We will need to look at alternating products of the form
1
0
Π =(cid:18)a1
(cid:18)cb2r
sb2r
0
a1
2(cid:19)(cid:18)cb2
1c + sb2
1c − cb2
sb2
2 s cb2r
2 s sb2r
1s − sb2
2s cb2
2c
1s + cb2
2s sb2
1 s − sb2r
2 c
1 s + cb2r
2 c(cid:19)(cid:18)a2r+1
2c(cid:19) · · ·(cid:18)a2r−1
2 (cid:19) ,
a2r+1
1
0
1
0
0
1 c + sb2r
1 c − cb2r
0
2 (cid:19)
a2r−1
where ai
products (depending on which sort of matrix the product starts or ends with).
2) = 0 = tr(bi
2)∀i - as well as three other kinds of
j ∈ Bj, tr(ai
j ∈ Aj, bi
1 + ai
1 + bi
Note that by taking(cid:18)a1
1
0
0
a1
2(cid:19) or both(cid:18)a1
1
0
0
a1
2(cid:19) and(cid:18)a2r+1
1
0
0
2 (cid:19) as(cid:18)1
0 −1(cid:19), and
a2r+1
0
using the fact that we are working with traces here, we find that it is sufficient to consider
the one special case listed above.
For i1 ∈ {1, 2}, the (i1, i1)th diagonal entry Πi1,i1 of Π is a sum of words of the form
ω = a1
i1ti1,i2(ω)b2
i2(ω)t′
i2(ω),i3(ω)a3
i3(ω)ti3(ω),i4(ω)b4
i4(ω)t′
i4(ω),i5(ω) · · · a2r−1
i2r−1(ω)ti2r−1(ω),i2r(ω)b2r
i2r(ω)t′
i2r(ω),i1a2r+1
i1
,
for i2(ω), · · · , i2r(ω) ∈ {1, 2} and tij(ω),ij+1(ω), t′
ij(ω),ij+1(ω) ∈ {c, s} (the reason behind using
the cumbersome notation ij(ω) is to emphasize the dependence of the indices ij on the
particular summand ω).
Now since the a's and b's are free from the t's and t′'s, we see from Theorem 14.4, [NS],
that
tr(ω) = Xπ∈N C(2r+1)
kπ(a1
i1, b2
i2(ω), · · · , a2r+1
i1
)trK(π)(ti1,i2(ω), t′
i2(ω),i3(ω), · · · , ti2r−1(ω),i2r (ω), t′
i2r(ω),i1, 1).
Thus, in order to prove that tr(Πi1,i1) = 0, it is enough to prove that for any π ∈
NC(2r + 1), i1 ∈ {1, 2},
Xω a summand of Πi1 ,i1
i2(ω),··· ,i2r(ω)∈{1,2}
kπ(a1
i1, b2
i2(ω), · · · , a2r+1
i1
)trK(π)(ti1,i2(ω), t′
i2(ω),i3(ω) · · · , ti2r−1(ω),i2r (ω), t′
i2r(ω),i1, 1) = 0
The crux of the proof lies in the following key lemma.
(4.1)
7
Lemma 4.3. Let p ∈ N, i1 ∈ {1, 2}. For aj
write aj =(cid:18)a2j−1
1
0
0
2 (cid:19) and bj =(cid:18)cb2j
sb2j
a2j−1
i ∈ Ai, bj
1 c + sb2j
1 c − cb2j
2 s cb2j
2 s sb2j
1 s − sb2j
2 c
1 s + cb2j
2 c(cid:19); and define
i ∈ Bk, where i = 1, 2, j = 1, 2, · · · , p+1,
Π′ = a1b1a2 · · · apbpap+1 and Π′′ = b1a2 · · · apbp.
Then
Xω a summand of Π′
i1,i1
i2(ω),··· ,i2p(ω)∈{1,2}
k02p+1(a1
i1, b2
i2(ω), · · · , b2p
i2p(ω), a2p+1
i1
)tr12p+1(ti1,i2(ω), t′
i2(ω),i3(ω),
· · · , ti2p−1(ω),i2p(ω), t′
i2p(ω),i1, 1) = 0
In particular, when a1 = ap+1 =(cid:18)1
0 −1(cid:19),
0
Xω a summand of Π′′
i1 ,i1
i2(ω),··· ,i2p(ω)∈{1,2}
k02p−1(b2
i2(ω), a3
i3(ω), · · · , b2p
i2p(ω))tr12p(ti1,i2(ω), t′
i2(ω),i3(ω),
(4.2)
(4.3)
· · · , ti2p−1(ω),i2p(ω), t′
i2p(ω),i1) = 0
Proof. In this proof we write ij(ω) as ij for the sake of simplicity.
Xω is a summand of Π′
Xω is a summand of Π′
i2,··· ,i2p∈{1,2}
i1,i1
=
i2,··· ,i2p∈{1,2}
i1 ,i1
k02p+1(a1
i1, b2
i2, · · · , a2p+1
i1
)tr12p+1(ti1,i2, t′
i2,i3, · · · , ti2p−1,i2p, t′
i2p,i1, 1)
tr(a1
i1)tr(b2
i2) · · · tr(a2p+1
i1
)tr(ti1,i2t′
i2,i3 · · · ti2p−1,i2pt′
i2p,i1),
which is the (i1, i1) entry of
0
0
1)
(cid:18)tr(a1
(cid:18)ctr(b2p
where λ(= 0
str(b2p
1)c + str(b2
1)c − ctr(b2
2)(cid:19)(cid:18)ctr(b2
tr(a1
1 )c + str(b2p
1 )c − ctr(b2p
str(b2
2 )s ctr(b2p
2 )s str(b2p
2)s ctr(b2
2)s str(b2
1 )s − str(b2p
2 )c
1 )s + ctr(b2p
1)s − str(b2
1)s + ctr(b2
2 )c(cid:19)(cid:18)tr(a2p+1
1
0
)
2)c
2)c(cid:19) · · ·(cid:18)tr(a2p−1
0
tr(a2p+1
2
)
0
1
0
tr(a2p−1
)(cid:19)
)(cid:19) = λ(2P02Qo)p2P0
2
∈ C \ {0}
if ∃i : tr(ai
if ∀i, tr(ai
1) = tr(ai
1) = −tr(ai
2) = 0 or tr(bi
2) 6= 0, tr(bi
1) = tr(bi
1) = −tr(bi
2) = 0
2) 6= 0
Now the proof follows since by the proof of Lemma 2.2, each diagonal entry of an alter-
nating product in 2P0 and 2Q0 has trace zero.
8
Now let us fix an arbitrary π ∈ NC(2r + 1).
If π = 02r+1, then equation 4.1 follows from equation 4.2 for p = r.
If π 6= 02r+1 then ∃m < n ∈ {1, · · · , 2r + 1} such that m ∼π n.
Consider an interval, say V = (k, · · · , l − 1) in K(π) for 1 ≤ m ≤ k ≤ l − 1 ≤ n ≤ (r + 1).
Then by Lemma 4.1, k ∼π l (since m ∼π n).
Case 1: Suppose V is odd. Then either ak
il are joined through π
(depending on whether k or l is odd), which leads to corresponding kπ being zero, since
Ai, Bj are free ∀i, j (Theorem 11.20 [NS]).
ik and al
ik and bl
il or bk
Case 2: Suppose V is even, both k and l are odd. Then ak
Thus l − 1 (cid:13) k.
ik and al
il are joined through π.
Also note that if ik 6= il then due to Aik and Ail being free the corresponding kπ would
be zero. So we now assume that ik = il.
Let π0 = π \ {(k + 1) ∨ · · · ∨ (l − 1)} and π1 = K(π) \ {(k, · · · , l − 1)}.
Then
kπ(a1
i1, b2
i2(ω), · · · , a2r+1
i1
)trK(π)(ti1,i2(ω), · · · , t′
i2r(ω),i1)
Xω is a summand of Πi1 ,i1
i2(ω),··· ,i2r(ω)
=
=
Xω is a summand of Πi1 ,i1
i2(ω),··· ,ik(ω)(=il(ω)),il+1(ω),··· ,i2r(ω)
ik−1,ik, til,il+1, · · · , t′
trπ1(ti1,i2, · · · , t′
Xω is a summand of Πi1 ,i1
i2(ω),··· ,ik(ω)(=il(ω)),il+1(ω),··· ,i2r(ω)
ik−1,ik, til,il+1, · · · , t′
trπ1(ti1,i2, · · · , t′
kπ0(a1
i1, b2
i2, · · · , ak
il, al
il, bl+1
il+1, · · · , b2r
i2r , a2r+1
i1
)
i2r ,i1)(cid:16) Xik+1(ω),··· ,il−1(ω)
k0V (bk+1
ik+1
, · · · , bl−1
il−1
)tr1V (tik,ik+1, · · · , t′
il−1,il)(cid:17)
kπ0(a1
i1, b2
i2, · · · , ak
il, al
il, bl+1
il+1, · · · , b2r
i2r , a2r+1
i1
)
i2r ,i1)(cid:16) Xik+1(ω),··· ,il−1(ω)
k0l−k−1(bk+1
ik+1
, · · · , bl−1
il−1
)tr1l−k (tik,ik+1, · · · , t′
il−1,ik)(cid:17)
Write Π = Π(1,k)Π(k+1,l−1)Π(l,2r+1),
where Π(1,k) = a1b1 · · · a
2 , Π(k+1,l−1) = b
k+1
k+1
2 a
k+3
2
l−1
· · · b
2 , Π(l,2r+1) = a
l+1
2
· · · b2rar+1.
cases.
Further let eΠ = Π(1,k)Π(l,2r+1). Note that Π(1,k) or Π(l,2r+1) can be trivial for the extreme
Then using the fact that ik(ω) = il(ω) the above sum may be re-written as
9
Xeω is a summand of eΠi1,i1
i2(eω),··· ,ik(eω)(=il(eω)),il+1(eω),··· ,i2r(eω)
trπ1(ti1,i2, · · · , t′
ik−1,il, til,il+1, · · · , t′
tr1l−k (tik,ik+1, · · · , t′
il−1,ik)(cid:17)
Now we put p = l−k
2 , Π′′ = Π(k+1,l−1)
ik,ik
il(eω)) ∈ {1, 2},
X
eeω is a summand of Π
(k+1,l−1)
ik ,ik
ik+1(eeω),··· ,il−1(eeω)
kπ0(a1
i1, b2
i2, · · · , ak
il, al
il, bl+1
il+1, · · · , b2r
i2r , a2r+1
i1
)
i2r ,i1)(cid:16)
X
eeω is a summand of Π
(k+1,l−1)
ik ,ik
ik+1(eeω),··· ,il−1(eeω)
k0l−k−1(bk+1
ik+1
, · · · , bl−1
il−1
)
in equation 4.3 to get for any ik(eω) = ik(= il =
k0l−k−1(bk+1
ik+1(eeω)
, · · · , bl−1
il−1(eeω)
)tr1l−k(tik ,ik+1(eeω), · · · , t′
il−1(eeω),ik
) = 0,
thus proving equation 4.1, as desired.
Case 3: The case when V is even and both k and l are even is proved exactly as in the
previous case.
Corollary 4.4.
(A1 ⊕ A2) ∗ LZ ∼= M2(A1 ∗ A2 ∗ LF3)
Proof. Follows from Proposition 4.2 as well as by noting the fact that LZ ∼= LZ ⊕ LZ, both
being singly generated von Neumann algebras with non atomic distributions
5 M2(A) ∗ M2(B) ∼= M2(A ∗ B ∗ LF3)
Let A, B be finite von-Neumann algebras. The notations for trace and cumulant remain
same as above.
Proposition 5.1.
M2(A) ∗ M2(B) ∼= M2(A ∗ B ∗ LF3)
Proof. Set Y =(cid:18)1 0
0 u(cid:19) and Z =(cid:18)1 0
0 v(cid:19) in M2(LF3), where u, v are Haar unitaries as in
the introduction, such that A, B, {u}, {v}, {c} are free.
We want to show that Y ∗M2(A)Y (∼= M2(A)) and W Z ∗M2(B)ZW ∗(∼= M2(B)) are free
in M2(A ∗ B ∗ LF3). As before, following the method used in Lemma 3.6, we can conclude
that these two subalgebras generate the algebra on the right in the proposition.
10
For ai,j ∈ A, bi,j ∈ B,
a2,1 a2,2(cid:19)(cid:18)1 0
0 u(cid:19)
0
a1,2u
0 u∗(cid:19)(cid:18)a1,1 a1,2
Y ∗(ai,j)Y =(cid:18)1
=(cid:18) a1,1
u∗a2,1 u∗a2,2u(cid:19)
0 v(cid:19)(cid:18) c
b2,2(cid:19)(cid:18)1 0
0 v∗(cid:19)(cid:18)b1,1
−s c(cid:19)
v∗b2,1 v∗b2,2v(cid:19)(cid:18) c
b2,1
b1,2v
s
s
0
b1,2
c (cid:19)(cid:18)1
W Z ∗(bi,j)ZW ∗ =(cid:18)c −s
=(cid:18)c −s
c (cid:19)(cid:18) b1,1
=(cid:18)cb1,1c − sv∗b2,1c − cb1,2vs + sv∗b2,2vs cb1,1s − sv∗b2,1s + cb1,2vc − sv∗b2,2vc
sb1,1c + cv∗b2,1c − sb1,2vs − cv∗b2,2vs sb1,1s + cv∗b2,1s + sb1,2vc + cv∗b2,2vc(cid:19)
−s c(cid:19)
s
s
As above, to prove freeness, it is enough to check on an alternating product of the form
a1
1,2u
1,1
2,1 u∗a1
u∗a1
Π =(cid:18) a1
· · ·(cid:16)(cid:18)c −s
2,2u(cid:19)(cid:16)(cid:18)c −s
c (cid:19)(cid:18) b2r
b2r
1,2v
1,1
v∗b2r
2,1 v∗b2r
c (cid:19)(cid:18) b2
2,2v(cid:19)(cid:18) c
b2
1,2v
1,1
2,1 v∗b2
v∗b2
2,2v(cid:19)(cid:18) c
−s c(cid:19)(cid:17)(cid:18) a2r+1
1,1
2,1
s
s
s
u∗a2r+1
s
−s c(cid:19)(cid:17)
a2r+1
1,2 u
u∗a2r+1
2,2 u(cid:19) ,
for tr(ai
1,1) + tr(ai
2,2) = 0 = tr(bj
1,1) + tr(bj
2,2).
Here too we will prove that each diagonal entry of Π has trace zero.
Let ω be a summand of (i1, i1)th diagonal entry of the above product. Then ω is an
alternating product of the form
a1
i1,i2w2
i2,i3b3
i3,i4w4
i4,i5 · · · w2r−2
i2r−2,i2r−1b2r−1
i2r−1,i2r w2r
i2r ,i2r+1a2r+1
i2r+1,i1
where wk
ik,ik+1 ∈ ±{c, sv∗, us, ucv∗, vs, su∗, vcu∗} depending on ω.
As before from [NS] we can say that
tr(ω) = Xπ∈N C(r+1)
kπ(a1
i1,i2, b3
i3,i4, · · · , b2r−1
i2r−1,i2r , a2r+1
i2r+1,i1)trK(π)(w2
i2,i3, w4
i4,i5, · · · , w2r
i2r ,i2r+1, 1).
Using the fact that u∗u = v∗v = 1, as in Lemma 4.3, here also
Xω is a summand of Πi1 ,i1
k0r+1(a1
i1,i2, b3
i3,i4, · · · , b2r−1
i2r−1,i2r , a2r+1
i2r+1,i1)tr1r+1(w2
i2,i3, w4
i4,i5, · · · , w2r
i2r,i2r+1, 1)
is the (i1, i1)th entry of the matrix
11
(cid:16)tr(a1
=(cid:16)tr(a1
1,2)U + tr(a1
2,1)U ∗ + tr(a1
· · ·(cid:16)W (tr(b2r
· · ·(cid:16)tr(b2r
1,2)V + tr(b2r
1,2 )U + tr(a2r+1
2,1 )U ∗ + tr(a2r+1
1,1)2P0(cid:17)(cid:16)W (tr(b2
1,1)2P0(cid:17)(cid:16)tr(b2
2,1)V ∗ + tr(b2r
2,1)X ∗ + tr(b2r
1,2)X + tr(b2
1,1)2P0)W ∗(cid:17)(cid:16)tr(a2r+1
1,1)2Q0(cid:17)(cid:16)tr(a2r+1
2,1)X ∗ + tr(b2
1,1)2P0)W ∗(cid:17)
1,1)2Q0(cid:17)
1,1 )2P0(cid:17)
1,1 )2P0(cid:17),
1,2)U + tr(a1
2,1)U ∗ + tr(a1
1,2)V + tr(b2
2,1)V ∗ + tr(b2
1,2)X + tr(b2r
1,2 )U + tr(a2r+1
2,1 )U ∗ + tr(a2r+1
where U, V, X are trace zero partial isometries in M2(LF3) as defined in the introduction,
But by proof of Lemma 3.3, each diagonal entry of an alternating product in {U, U ∗, 2P0}
and {X, X ∗, 2Q0} has trace zero.
Thus exactly as Lemma 4.3, for any i1 ∈ {1, 2},
Xω is a summand of Πi1,i1
k0r+1(a1
i1,i2, b3
i3,i4, · · · , b2r−1
i2r−1,i2r , a2r+1
i2r+1,i1)tr1r+1(w2
i2,i3, w4
i4,i5, · · · , w2r
i2r,i2r+1, 1) = 0.
Now rest of the proof follows similarly as Proposition 4.2.
Proposition 5.2.
(A1 ⊕ A2) ∗ M2(B) ∼= M2(A1 ∗ A2 ∗ B ∗ LF2)
Proof. Here one needs to prove that(cid:18)A1
0 A2(cid:19) (∼= A1 ⊕A2) and W Z ∗M2(B)ZW ∗(∼= M2(B))
are free and they generate RHS. The proof is exactly similar to that of Proposition 4.2 or
Proposition 5.1, using Corollary 3.5.
0
Remark 5.3. As in Remark 3.7 here too we can have an exactly similar alternate proof
using W(cid:18)A1
0 A2(cid:19) W ∗ and Y M2(B)Y ∗ as model.
0
Corollary 5.4.
M2(A) ∗ LZ ∼= M2(A ∗ LF4)
Corollary 5.5. For k1, k2, l1, l2 ∈ N ∪ {0},
1. (LFk1 ⊕ LFk2) ∗ (LFl1 ⊕ LFl2) ∼= M2(LFk1+k2+l1+l2+1)
2. (LFk1 ⊕ LFk2) ∗ LZ ∼= M2(LFk1+k2+3)
3. M2(LFk) ∗ M2(LFl) ∼= M2(LFk+l+3)
12
4. (LFk1 ⊕ LFk2) ∗ M2(LFl) ∼= M2(LFk1+k2+l+2)
5. LZ ∗ M2(LFl) ∼= M2(LFl+4)
Proof. (1), (3) and (4) are direct consequences of Proposition 4.2, Proposition 5.1 and Propo-
sition 5.2 respectively. (2) and (5) follow from those as well as Corollary 4.4 and 5.4 (In fact
(3) follows directly from Theorem 5.4.1 [VDN]).
6 Applications
In this section, we use the results proved in the previous sections and compute various
possible free products involving the hyperfinite II1 factor R, (LFl)2m and M2n(LFk), where
m, n ∈ N, k, l ∈ N ∪ {0} (LF0 is considered as C). These results were proved in section 1-3
of [D2] in a much more general context, but with a different approach.
For the computation we frequently need to use Theorem 5.4.1 of [VDN]. For our purpose
it will be enough to stick to the following 2-dimensional version of the theorem: For k ≥ 2,
LFk
∼= M2(LF(k−1)4+1).
(6.1)
We now state an example of computing a free product of certain finite dimensional
algebras, using the above sections:
Example 6.1. C2n ∗ C2m ∼= M2(LF5−2(
M2(LF5− 4
2n−1
), n ≥ 1.
1
2n−1 + 1
2m−1 )), n ≥ 1.
In particular, C2n ∗ C2n ∼=
Proof. We first prove inductively that C2n ∗ C2n ∼= M2(LFan) for some an ∈ [1, ∞), for
all n ≥ 1. Then the basic step follows from Proposition 2.3. Also a1 = 1 by the same
Proposition.
C2n+1
∗ C2n+1 ∼= (C2n
⊕ C2n
) ∗ (C2n
∗ C2n
⊕ C2n
∗ C2n
)
∗ C2n
∗ LZ), by Proposition 2.3
∼= M2(C2n
∼= M2(M2(LFan) ∗ M2(LFan) ∗ LZ), by induction hypothesis
∼= M2(M2(LF2an+3) ∗ LZ), by Corollary 5.5
∼= M2(LF 2an+3−1
∼= M2(LF an
)
+2 ∗ LZ), by equation 6.1
4
2 + 5
2
Thus the induction is complete. Moreover we have the recurrence relation an+1 = an
2 + 5
2.
13
Now,
an+1 =
=
=
an−1
2 + 5
2
+
an
5
2
2
an−1
22 +
5
1
(
2n +
2
=
+
2
5
5
2.2
2
1
2n−1 +
+
5
2
a1
2n +
(
=
5
2
1
2n−1 +
1
2n−2 + · · · + 1) =
1
2n−2 + · · · + 1)
5(2n − 1)
1
2n +
2n
= 5 −
4
2n .
From the above calculations and equation 6.1,
C2n ∗ C2n ∼= M2(LF5− 4
), as required
∼= LF2− 1
2n−1
2n−1
, n ≥ 1.
Without loss of generality we can assume that n ≥ m. For m = n = 1 the proof follows
from Proposition 2.3.
Let n ≥ 2. Then,
C2n
∗ C2m ∼= M2(C2n−1
∗ C2n−1
∗ C2m−1
∼= M2(LF2− 1
∼= M2(LF5−2(
2n−2
∗ LF2− 1
2m−1 )).
1
2n−1 + 1
2m−2
∗ LZ)
∗ C2m−1
∗ LZ)
We are now ready to state the following proposition that summarizes the promised com-
putations of the free products involving certain finite dimensional von-Neumann algebras and
the free-group von-Neumann algebras. We will omit the proof since it is a simple exercise of
induction using the previous sections, similar to the above example.
Proposition 6.2. For m, n ∈ N, k, l ∈ N ∪ {0} and LF0 = C,
1. (LFk)2n ∗ (LFl)2m ∼= M2(LF5+ 2(k−1)
2n−1 + 2(l−1)
2m−1
). In particular (LZ)2n ∗ (LZ)2m ∼= M2(LF5).
2. M2n(LFk)∗(LFl)2m ∼= M2(LF5+ k−1
3. M2n(LFk) ∗ M2m(LFl) ∼= M2(LF5+ k−1
4n−1 + 2(l−1)
2m−1
4n−1 + l−1
4m−1
M2(LF5).
). In particular M2n(LZ)∗(LZ)2m ∼= M2(LF5).
). In particular M2n(LZ) ∗ M2m(LZ) ∼=
Remark 6.3. There exist explicitly computable functions f, g : [1, ∞) → [1, ∞), such that
whenever finite von-Neumann algebras A, B satisfy A ∗ B ∼= LFt for some t ∈ [1, ∞), then,
for all n ∈ N, we have
14
• A2n ∗ B2n ∼= M2(LFf (t)),
• M2n(A) ∗ M2n(B) ∼= M2(LFg(t)).
Remark 6.4. One can obviously extend the above proposition by taking LFk1 ⊕ · · · ⊕ LFk2n
for ki ≥ 0, instead of (LFk)2n.
We know that the hyperfinite II1 factor R can be constructed as an infinite tensor
scalar matrix algebras of dimension 2n. Again LZ ∼=
product of type I2n factors, i.e.
L∞([0, π
2 ]) can be thought of as an infinite tensor product of C2n.
Proposition (6.2) suggests - on 'taking the limit as m, n → ∞' , with k = l = 0 - that
1. LZ ∗ LZ ∼= M2(LF5) (trivially true);
2. R ∗ LZ ∼= M2(LF5) (see Theorem 5.4.3 [VDN]);
3. R ∗ R ∼= M2(LF5).
We could not come up with a matrix model to prove the above statements, approximating
LZ and R as by finite dimensional algebras as n → ∞ in Proposition 6.2. But we shall indeed
give a rigorous proof for the assertion about R ∗ R, as against the 'limiting' heuristics:
In view of the uniqueness of the hyperfinite II1 factor (see [MvN]), we know that
R ∼= M2(R)
(6.2)
Now using Theorem 5.4.3 of [VDN], i.e.
LFk ∗ R ∼= LFk+1
we may deduce the following:
Proposition 6.5. For finite von-Neumann algebra A1, A2, B,
1. R ∗ (A1 ⊕ A2) ∼= M2(A1 ∗ A2 ∗ LF3)
2. R ∗ M2(B) ∼= M2(B ∗ LF4)
Proof. By above equations and Proposition 5.2
R ∗ (A1 ⊕ A2) ∼= M2(R) ∗ (A1 ⊕ A2)
∼= M2(A1 ∗ A2 ∗ R ∗ LZ)
∼= M2(A1 ∗ A2 ∗ LF3)
The other statement follows similarly using Proposition 5.1.
15
Corollary 6.6. R ∗ R ∼= M2(LF5)
Proof. It follows from equation 6.2 and the above proposition.
We now wish to observe that our proof can also be led to the strengthened version
emphasize the distinction between the two free copies of the hyperfinite II1 factor. Consider
Proposition 6.7 of Corollary 6.6). Following [D1], let us write the left side as R ∗ eR to
R ∼= M2(R),eR ∼= M2(eR). Then we notice that by proof of Proposition 5.1, M2(R) on the
left hand side gets mapped into M2(R) on the right hand side (which is M2(R ∗ eR ∗ LF3) ∼=
0 u(cid:19).
M2(R ∗ LF4)) as conjugated by the unitary matrix Y ∗ where Y =(cid:18)1 0
On the other hand, by Proposition 5.2 and Theorem 5.4.3 of [VDN], we have
M2(R) ∗ LZ ∼= M2(R) ∗ (LZ ⊕ LZ)
∼= M2(R ∗ LF4),
where by Remark 5.3, M2(R) on the left hand side is mapped into M2(R) on the right hand
side in exactly the same manner as above, i.e. as conjugated by the same unitary matrix Y ∗.
In fact we note here that for projection P =(cid:18)1 0
P (M2(R) ∗ M2(eR))P ∼= P M2(R)P ∗ LF4
0 0(cid:19) ∈ M2(R),
∼= P (M2(R) ∗ LZ)P,
where the isomorphisms restricted to P M2(R)P (which is naturally isomorphic to R), in all
three cases are the identity maps.
Thus similarly as in Corollary 3.6 of [D1], we can conclude that
Proposition 6.7.
where the first isomorphism restricted to R on the left hand side is the identity map to R on
the right hand side.
R ∗ eR ∼= R ∗ LZ ∼= M2(LF5) ∼= LF2,
7 Acknowledgments
I wish to thank V. S. Sunder and Vijay Kodiyalam for helpful discussions.
8 References
[D1] K. Dykema: Interpolated factor Factors, Pacific J. Math. Volume 163, Number 1 (1994),
123-135
[D2] K. Dykema: Free products of hyperfinite von Neumann algebras and free dimension,
Duke Math. J. Volume 69, Number 1 (1993), 97-119.
16
[NS] A. Nica, R. Speicher: Lectures on the Combinatorics of Free Probability. Cambridge
University Press, 2006
[VDN] D. V. Voiculescu, K. Dykema, A. Nica: Free Random Variables, CRM Monograph
Series, Volume 1, American Mathematical Society, Providence, RI, 1992.
[MvN] F.J. Murray, J. von Neumann: On rings of operators IV Ann. of Math. (2) , 44
(1943) pp. 716808.
. INSTITUTE OF MATHEMATICAL SCIENCES, CIT Campus, Taramani, Chennai 600113,
India.
. E-mail address: [email protected]
17
|
1906.07487 | 1 | 1906 | 2019-06-18T10:35:14 | The tight groupoid of the inverse semigroups of left cancellative small categories | [
"math.OA",
"math.RA"
] | We fix a path model for the space of filters of the inverse semigroup $\mathcal{S}_\Lambda$ associated to a left cancellative small category $\Lambda$. Then, we compute its tight groupoid, thus giving a representation of its $C^*$-algebra as a (full) groupoid algebra. Using it, we characterize when these algebras are simple. Also, we determine amenability of the tight groupoid under mild, reasonable hypotheses. | math.OA | math |
THE TIGHT GROUPOID OF THE INVERSE SEMIGROUPS OF
LEFT CANCELLATIVE SMALL CATEGORIES.
EDUARD ORTEGA AND ENRIQUE PARDO
Abstract. We fix a path model for the space of filters of the inverse semigroup SΛ
associated to a left cancellative small category Λ. Then, we compute its tight groupoid,
thus giving a representation of its C ∗-algebra as a (full) groupoid algebra. Using it,
we characterize when these algebras are simple. Also, we determine amenability of the
tight groupoid under mild, reasonable hypotheses.
Introduction
In [15], Spielberg described a new method of defining C ∗-algebras from oriented combi-
natorial data, generalizing the construction of algebras from directed graphs, higher-rank
graphs, and (quasi-)ordered groups. To this end, he introduced categories of paths -- i.e.
cancellative small categories with no (nontrivial) inverses -- as a generalization of higher
rank graphs, as well as ordered groups.The idea is to start with a suitable combinato-
rial object and define a C ∗-algebra directly from what might be termed the generalized
symbolic dynamics that it induces. Associated to the underlying symbolic dynamics, he
present a natural groupoid derived from this structure. The construction also gives rise
to a presentation by generators and relations, tightly related to the groupoid presenta-
tion. In [16] he showed that most of the results hold when relaxing the conditions, so
that right cancellation or having no (nontrivial) inverses are taken out of the picture.
In [2], B´edos, Kaliszewski, Quigg and Spielberg use Spielberg's construction to ex-
tend the notion of self-similar graph introduced in [10] -- they termed it as "Exel-Pardo
systems" -- to the context of actions of group (potentially, of groupoids) on left cancella-
tive small categories. To this end, they use a Zappa-Sz´ep product construction, and
studied the representation theory for the Spielberg algebras of the new left cancellative
small category associated to this Zappa-Sz´ep product.
In the present paper, we study Spielberg construction, using a groupoid approach
based in the Exel's tight groupoid construction [7]. To this end, we study various inverse
semigroups associated to a left cancellative small category (see e.g. [6]), we compute a
"path-like" model for their tight groupoids, and we study the basic properties of its tight
groupoid. Also, we show that the tight groupoid for these inverse semigroups coincide
with Spielberg's groupoid [14]. With this tools at hand, we are able to characterize
simplicity for the algebras associated to finitely aligned left cancellative small categories,
2010 Mathematics Subject Classification. Primary: 46L05; Secondary: 46L80, 46L55, 20L05.
Key words and phrases. Left cancellative small category, inverse semigroup, tight representation,
tight groupoid, groupoid C ∗-algebra.
The second-named author was partially supported by PAI III grant FQM-298 of the Junta de An-
daluc´ıa, and by the DGI-MINECO and European Regional Development Fund, jointly, through grant
MTM2017-83487-P.
1
TIGHT GROUPOIDS OF LCSC
2
and in particular in the case of Exel-Pardo systems. Finally, we give, under mild and
necessary hypotheses, a characterization of amenability for such groupoid.
The contents of this paper can be summarized as follows: In Section 1 we recall some
known facts about small categories and inverse semigroups. In Section 2 we study basic
properties of the inverse semigroups SΛ and TΛ associated to a left cancellative small
category Λ. Section 3 is devoted to study filters on a left cancellative small category
and their path models. Section 4 deals with defining actions of SΛ on filter spaces,
and we picture their tight groupoids. In Section 5 we show that the tight groupoid of
SΛ is isomorphic (as topological groupoid) to the Spielberg's groupoid on Λ. Groupoid
properties characterizing simplicity on the associated algebras are stated in Section 6.
Section 7 is centered in analyzing Zappa-Sz´ep products, introduced in [2] to generalize
self-similar graphs of [10], from our particular perspective. We close the paper studying,
in Section 8, the amenability of the tight groupoid of Zappa-Sz´ep products.
1. Basic facts.
In this section we collect all the background we need for the rest of the paper.
1.1. Small categories. Given a small category Λ, we will denote by Λ◦ the class of its
objects, and we will identify Λ◦ with the identity morphisms, so that Λ◦ ⊆ Λ. Given
α ∈ Λ, we will denote by s(α) := dom(α) ∈ Λ◦ and r(α) := ran(α) ∈ Λ◦. The invertible
elements of Λ are
Λ−1 := {α ∈ Λ : ∃β ∈ Λ such that αβ = s(β)} .
Definition 1.1. Given a small category Λ, and let α, β, γ ∈ Λ:
(1) Λ is left cancellative if αβ = αγ then β = γ,
(2) Λ is right cancellative if βα = γα then β = γ,
(3) Λ has no inverses if αβ = s(β) then α = β = s(β).
A category of paths is a small category that is right and left cancellative and has no
inverses.
Notice that if Λ is either left or right cancellative, then the only idempotents in Λ are
Λ◦. Indeed, if αα = α, since α = r(α)α = αs(α) we have that α = s(α) or α = r(α).
Definition 1.2. Let Λ be a small category. Given α, β ∈ Λ, we say that β extends α
(equivalently α is an initial segments of β) if there exists γ ∈ Λ such that β = αγ. We
denote by [β] = {α ∈ Λ : α is an initial segment of β}. We write α ≤ β if α ∈ [β].
Lemma 1.3. Let Λ be a small category. Then
(1) the relation ≤ is reflexive and transitive,
(2) if Λ is left cancellative with no inverses, then ≤ is a partial order.
Proof. (1) Clearly α = αs(α), so α extends itself. If β = αα′ (α ≤ β) and γ = ββ′
(β ≤ γ), then γ = αα′β′ (α ≤ γ).
(2) Suppose that α = ββ′ (β ≤ α) and β = αα′ (α ≤ β). Then,
αs(α) = α = ββ′ = αα′β′ .
Thus, by left cancellation we have that s(α) = α′β′, whence since Λ has no inverses it
follows that α′, β′ ∈ Λ◦.
(cid:3)
TIGHT GROUPOIDS OF LCSC
3
Lemma 1.4 ([16, Lemma 2.3]). Let Λ be a LCSC (Left cancellative small category),
and let α, β ∈ Λ. Then, α ≤ β and β ≤ α if and only if β ∈ αΛ−1 = {αγ : γ ∈
Λ−1 with r(γ) = s(α)}.
We denote by α ≈ β if β ∈ αΛ−1. This is an equivalence relation.
Lemma 1.5 ([16, Lemma 2.5(ii)]). Let Λ be a LCSC, and let α, β ∈ Λ. Then the
following are equivalent:
(1) α ≈ β,
(2) αΛ = βΛ,
(3) [α] = [β].
Notation 1.6. Let Λ be a LCSC. Given α, β ∈ Λ, we say :
(1) α ⋓ β if and only if αΛ ∩ βΛ 6= ∅,
(2) α ⊥ β if and only if αΛ ∩ βΛ = ∅.
Definition 1.7. Let Λ be a LCSC, and let F ⊂ Λ. The elements of Tγ∈F γΛ are the
common extensions of F . A common extension ε of F is minimal if for any common
extension γ with ε ∈ γΛ we have that γ ≈ ε.
When Λ has no inverses, given F ⊆ Λ and given any minimal common extension ε of
F , if γ is common extension of F with ε ∈ γΛ then γ = ε. We will denote by
α ∨ β := {the minimal extensions of α and β} .
Notice that if α ∨ β 6= ∅ then α ⋓ β, but the converse fails in general.
Definition 1.8. A LCSC Λ is finitely aligned if for every α, β ∈ Λ there exists a finite
subset Γ ⊂ Λ such that αΛ ∩ βΛ =Sγ∈Γ γΛ.
When Λ is a finitely aligned LCSC, we can always assume that α ∨ β = Γ where Γ is
a finite set of minimal common extensions of α and β.
1.2. Inverse semigroups.
Definition 1.9. A semigroup S is an inverse semigroup if for every s ∈ S there exists
a unique s∗ ∈ S such that s = ss∗s and s∗ = s∗ss∗.
Equivalently, S is an inverse semigroup if and only if the subsemigroup E(S) := {e ∈
S : e2 = e} of idempotents of S is commutative [12, Theorem 1.1.3].
A monoid is a semigroup with unit. We say that a semigroup S has zero if there
exists 0 ∈ S such that 0s = s0 = 0 for every s ∈ S.
Definition 1.10. Given a set X, we define the (symmetric) inverse semigroup on X as
I(X) := {f : Y → Z : Y, Z ⊆ X and f is a bijection } ,
endowed with operation
g ◦ f : f −1(ran(f ) ∩ dom(g)) −→ g(ran(f ) ∩ dom(g)) ,
and involution
f ∗ := f −1 : ran(f ) −→ dom(f ) .
Notice that I(X) has unit IdX : X → X and zero being the empty map 0 : ∅ → ∅.
TIGHT GROUPOIDS OF LCSC
4
The Wagner-Preston Theorem [12, Theorem 1.5.1] guarantees that every inverse semi-
group is a ∗-subsemigroup of I(X) for some suitable set X.
Definition 1.11. Let S be an inverse semigroup, and let E(S) be its subsemigroup of
idempotents. Given e, f ∈ E(S), we say that e ≤ f if and only if e = ef . We extend
this relation to a partial order as follows: given s, t ∈ S, we say that s ≤ t if and only if
s = ss∗t = ts∗s.
Definition 1.12. We say that s, t ∈ S are compatible, denoted by s ∼ t, if both s∗t
and st∗ belong to E(S).
This concept will be essential to understand various properties.
are pairwise compatible.
We will say that S is (finitely) complete if for every (finite) subset Σ ⊆ S of pairwise
We say that S is (finitely) distributive if whenever Σ is a (finite) subset of S and
Lemma 1.13 ([12, Lemma 1.4.16]). Let Σ ⊆ S. IfWα∈Σ α ∈ S, then the elements of Σ
s ∈ S, ifWα∈Σ α ∈ S thenWα∈Σ sα ∈ S and s(cid:0)Wα∈Σ α(cid:1) =Wα∈Σ sα.
compatible elements we have thatWα∈Σ α ∈ S.
(1) Not necessarilyWα∈Σ α ∈ S,
(2) even ifWα∈Σ α ∈ S, it can happen thatWα∈Σ sα /∈ S.
The symmetric inverse monoid I(X) is complete and distributive [12, Proposition
1.2.1(i-ii)]. But this property is not necessarily inherited by its inverse subsemigroups.
Indeed, the point is that given Σ ⊆ S a set of pairwise compatible elements, and s ∈ S:
To understand when f, g ∈ I(X) are compatible elements, and describe who is f ∨g ∈
I(X), we address the reader to Lawson's monograph [12, Proposition 1.2.1].
2. The semigroups SΛ and TΛ
Given a LCSC Λ, we will define some inverse semigroups associated to Λ.
Definition 2.1. Let Λ be a LCSC. For any α ∈ Λ, we define two elements of I(Λ):
(1) σα : αΛ → s(α)Λ given by αβ 7→ β ,
(2) τ α : s(α)Λ → αΛ given by β 7→ αβ .
Clearly σα is injective, and since Λ is left cancellative so is τ α. Moreover,
σα = σατ ασα
and
τ α = τ ασατ α ,
for every α ∈ Λ.
Definition 2.2. Given a LCSC Λ, we define the semigroup
SΛ := hσα, τ α : α ∈ Λi .
Lemma 2.3. Let Λ be a LCSC. Then SΛ is an inverse semigroup.
Proof. It is clear, since I(Λ) is an inverse semigroup, and SΛ ⊆ I(Λ) is closed under
composition and inverses.
(cid:3)
In order to better understand its structure, we will need to consider finite aligned
LCSC. First, we introduce a definition.
TIGHT GROUPOIDS OF LCSC
5
Definition 2.4. Let Λ be a finitely aligned LCSC, and let s ∈ SΛ. We say that a
i=1 τ αiσβi is irredundant if for all 1 ≤ i 6= j ≤ n we have αi 6∈ [αj]
and βi 6∈ [βj].
presentation s =Wn
Remark 2.5. Let s =Wn
But since s :Sn
i=1 τ αiσβi ∈ SΛ, and suppose that for all 1 ≤ i 6= j ≤ n we have
βi 6∈ [βj]. Now suppose that there exists 1 ≤ i 6= j ≤ n with αi ≤ αj, so there exists
γ ∈ Λ such that αiγ = αj. Then we have that
s(βiγ) = αiγ = αj = s(βj) .
i=1 ran(τ αi) is a bijection, we have that βiγ = βj, so
βi ≤ βj, a contradiction. Thus, s is irredundant. Similarly it can be proved that s is
irredundant if and only if for all 1 ≤ i 6= j ≤ n we have αi 6∈ [αj].
i=1 dom(σβi) →Sn
Lemma 2.6 ([15, Lemma 3.3 & Theorem 6.3]). If Λ is a finite aligned LCSC, then every
f ∈ SΛ is the supremum of a finite family of elements of the form τ ασβ with α, β ∈ Λ
and s(α) = s(β). Moreover, if such a decomposition is irredundant, then is unique (up
to permutation).
Notice that, given any finite family {α1, . . . , αn} ⊂ Λ, the elements {τ αiσαi}n
i=1 ⊂ SΛ
i=1 τ αiσαi ∈ I(Λ), but not necessarily to SΛ. Thus,
in order to do some essential arguments we need to consider a new object.
are pairwise compatible, so that Wn
Definition 2.7. Let Λ be a finitely aligned LCSC. We define
TΛ =( n_i=1
τ αiσβi : {τ αiσβi}n
i=1 ⊂ SΛ are pairwise compatible) .
Clearly, by Lemma 2.6 SΛ ⊆ TΛ ⊂ I(Λ). Moreover, by [12, Proposition 1.4.20 &
Proposition 1.4.17], TΛ is closed by composition and inverses, and moreover, is finitely
distributive. Thus,
Lemma 2.8. Let Λ be a finitely aligned LCSC. Then, TΛ is an inverse semigroup con-
taining SΛ. Moreover, TΛ is the smallest finitely complete, finitely distributive, inverse
semigroup containing SΛ.
Now, we will proceed to understand who are the elements in E(TΛ) and the order
relation.
Lemma 2.9. Let Λ be a finitely aligned LCSC. Then e ∈ E(TΛ) if and only if e =
(cid:3)
i=1 τ αiσβi. By [12, Proposition 1.4.17], e∗ = e =
i=1 τ βiσαi, and
i=1 τ αiσαi for some α1, . . . , αn ∈ Λ.
Wn
Proof. Let e ∈ E(TΛ), then e = Wn
Wn
n_i=1
Thus,Wn
τ βiσβi! (βi) = n_i=1
i=1 τ βiσβi =Wn
βi = n_i=1
e = e∗e =
as desired.
τ βiσαiτ αiσβi =
τ βiσβi .
n_i=1
τ αiσβi! (βi) = αi ,
i=1 τ αiσβi. But then given 1 ≤ i ≤ n, we have that
Since TΛ is finitely distributive, we have that
τ βj σβj! .
τ αk σαk = τ αk σαk m_j=1
τ αkσαk (ε)σβj σβj (ε) .
_ε∈αk∨βj
m_j=1
τ αk σαkτ βj σβj =
τ αk σαk =
m_j=1
TIGHT GROUPOIDS OF LCSC
6
i=1 τ αiσαi, f =
j=1 τ βj σβj be idempotents of either SΛ or TΛ (written in irredundant form). Then, the
Proposition 2.10. Let Λ be a finitely aligned LCSC, and let e = Wn
Wm
following are equivalent:
(1) e ≤ f ,
(2) for each 1 ≤ k ≤ n, there exists 1 ≤ l ≤ m such that βl ≤ αk.
Proof. For (1) implies (2), let e, f ∈ E(TΛ) with e ≤ f . Then, e = Wn
f =Wm
j=1 τ βj σβj . Fix any 1 ≤ k ≤ n. Then, τ αk σαk ≤ e ≤ f if and only if
i=1 τ αiσαi and
Without loss of generality, we can assume that the decomposition is irredundant (by
using the reduction argument in the proof of [15, Theorem 6.3]. By Lemma 2.6, there
exist 1 ≤ l ≤ m and ε ∈ αk ∨ βl such that τ αk σαk = τ αkσαk (ε)σβlσβl (ε), whence αk =
αkσαk(ε) = βlσβl(ε). Thus, βl is an initial segment of αk if and only if βl ≤ αk if and
only if βl ∈ [αk].
For (2) implies (1), if βl ≤ αk, then τ αk σαk ≤ τ βlσβl ≤ f . Since this is true for all
1 ≤ k ≤ n, we have that e ≤ f , as desired.
Notice that, even we need TΛ to argue, the conclusion works for SΛ too.
(cid:3)
By an analog argument, we have the following result, extending Proposition 2.10 to
any couple of elements of SΛ.
i=1 τ αiσβi, t =
j=1 τ γj σδj be elements of either SΛ or TΛ (written in irredundant form). Then, the
Proposition 2.11. Let Λ be a finitely aligned LCSC, and let s = Wn
Wm
(1) s ≤ t,
(2) for each 1 ≤ k ≤ n, there exists 1 ≤ l ≤ m such that αk = γlǫ and βk = δlǫ for
following are equivalent:
some ǫ ∈ s(γl)Λ.
Now, we will connect SΛ with the semigroups appearing in [6, 15, 16].
Definition 2.12. Let Λ be a small category. A zigzag is an even tuple of the form
ξ = (α1, β1, α2, β2, . . . , αn, βn)
with αi, βi ∈ Λ, r(αi) = r(βi) for every 1 ≤ i ≤ n and s(αi+1) = s(βi) of every 1 ≤ i < n.
We will denote by ZΛ the set of zigzags of Λ. Given ξ ∈ ZΛ, we define s(ξ) = s(βn),
r(ξ) = s(α1) and ξ = (βn, αn, . . . , β1, α1).
Every ξ ∈ ZΛ defines a zigzag map ϕξ ∈ I(Λ) by
ϕξ = σα1τ β
1 · · · σαnτ βn .
We will denote Z(Λ) = {ϕξ : ξ ∈ ZΛ}.
TIGHT GROUPOIDS OF LCSC
7
Remark 2.13.
(1) For every α ∈ Λ we can define ξα := (r(α), α). Notice that ϕξα = τ α and
ϕξα
= σα.
(2) Z(Λ) is closed by concatenation, and ϕξ1 ◦ ϕξ2 = ϕξ1ξ2.
(3) For every ξ ∈ ZΛ, then ϕξ = ϕ−1
Thus,
.
ξ
Lemma 2.14 ([2, Section 7.2]). If Λ is a LCSC, then Z(Λ) is an inverse semigroup.
Moreover, Z(Λ) = SΛ.
Proof. First part is a consequence of Remark 2.13(2-3). For the second part, Remark
2.13(1-3) implies that SΛ ⊆ Z(Λ). On the other side, for every ξ ∈ ZΛ we have that
ϕξ ∈ SΛ, so that Z(Λ) ⊆ SΛ.
(cid:3)
Hence, when working with Z(Λ), we benefit of results in previous sections.
3. Filters on LCSC
Let Λ be a LCSC. We denote by E := E(SΛ) the semilattice of idempotents of the
inverse semigroup SΛ.
Definition 3.1. A nonempty subset η of E is a filter if:
(1) e ∈ η, f ∈ E and e ≤ f , then f ∈ η,
(2) e, f ∈ η then ef ∈ η.
The set of filters of E is denoted by E0. We can endow E0 with a topology, as follows.
Definition 3.2. For any X, Y ⊂ E finite subsets, define
U(X, Y ) := {η ∈ E0 : X ⊆ η and Y ∩ η = ∅} .
Then
TE = {U(X, Y ) : X, Y ⊆ E finite} ,
is a basis for a topology of E0, under which E0 is Hausdorff and locally compact space
(See e.g.[9]).
Definition 3.3. A filter η ∈ E0 is an ultrafilter if it is not properly contained in another
filter. Equivalently, η is maximal among the filters, partially ordered by inclusion.
A useful characterization is the following.
Lemma 3.4 ([7, Lemma 12.3]). A filter η ∈ E0 is an ultrafilter if and only if e ∈ E and
ef 6= 0 for every f ∈ η implies e ∈ η.
We denote by E∞ the subspace of ultrafilters of E0. Usually, E∞ is not closed in E0.
Definition 3.5. We define Etight as the closure of E∞ in E0. A filter in Etight is called
tight filter.
In order to characterize tight filters,we need to introduce some known concepts.
Definition 3.6. Given X, Y ⊂ E finite sets, we define
E X,Y = {e ∈ E : e ≤ x for every x ∈ X and ey = 0 for every y ∈ Y } .
TIGHT GROUPOIDS OF LCSC
8
Definition 3.7. Given a subset F of E, a outer cover for F is a subset Z ⊂ E such that
for every f ∈ F there exists z ∈ Z such that zf 6= 0. Moreover, Z is a cover for F if Z
is an outer cover for F with Z ⊆ F .
Given an idempotent e ∈ E, we say that Z ⊆ E is a cover for e if Z is a cover for the
set {f ∈ E : f ≤ e}.
Lemma 3.8 ([7, Theorem 12.9]). A filter η ∈ E0 is tight if and only if for every X, Y ⊂ E
finite sets and for every finite cover Z of E X,Y , η ∈ U(X, Y ) implies Z ∩ η 6= ∅.
3.1. Path models. Viewing some examples of LCSC, as graphs and k-graphs, we are
interested in obtain practical models of ultrafilters and tight filters. These models should
behave, somehow, as paths in a graph.
To guarantee that every filter has a such a path model, we introduce a restriction on
Λ.
Definition 3.9. Let Λ be a finitely aligned LCSC. We say that a filter η ∈ bE0 enjoys
condition (∗) if givenWn
i=1 τ αiσαi ∈ η, then there exists 1 ≤ j ≤ n such that τ αj σαj ∈ η.
[5,
Notice that, if Λ is singly aligned, then every filter enjoy condition (∗) (see e.g.
Proposition 3.5]).
Before showing how to construct the path model of η, let us show that there exist
filters where this property always holds.
Lemma 3.10. Let Λ be a finitely aligned LCSC. Then, any η ∈ Etight satisfies condition
(∗).
i=1 τ αiσαi. Define X = {e}, Y = ∅ and
i=1 τ αiσαi ≥ τ αj σαj for every 1 ≤ j ≤ n, it is clear that
i=1 τ αiσαi, then f τ αj σαj = 0 for every 1 ≤ j ≤ n will
τ αiσαi! = f e = f ,
Z = {τ αiσαi}n
Proof. Let e ∈ η. By Lemmas 2.6 and 2.9, e =Wn
i=1. Since e =Wn
Z ⊂ E X,Y . Also, if 0 6= f ≤ e =Wn
f τ αiσαi = f n_i=1
n_i=1
imply that
0 =
a contradiction. Hence, Z is a finite cover of E X,Y . Clearly, η ∈ U(X, Y ).
Thus, η ∈ Etight implies η ∩ Z 6= ∅ by Lemma 3.8, i.e. there exists 1 ≤ j ≤ n such that
(cid:3)
τ αj σαj ∈ η.
Corollary 3.11. If Λ is a finitely aligned LCSC, then every η ∈ E∞ satisfies condition
(∗).
Now, we proceed to introduce a set of paths.
Definition 3.12. Let Λ be a finitely aligned LCSC. A nonempty subset F of Λ is:
(1) Hereditary, if α ∈ Λ, β ∈ F and α ≤ β implies α ∈ F ,
(2) (upwards) directed, α, β ∈ F implies that there exists γ ∈ F with α, β ≤ γ.
We denote Λ∗ the set of nonempty, hereditary, directed subsets of Λ.
Notice that, if F ∈ Λ∗, there exists a unique v ∈ Λ0 such that F ⊂ vΛ. Indeed, given
any α, β ∈ F , there exists γ ∈ F with γ ≥ α, β, i.e., γ = α α = β β for some α, β ∈ Λ.
Thus, v = r(α) for any α ∈ F is the desired element.
TIGHT GROUPOIDS OF LCSC
9
Definition 3.13. Given η ∈ E0, we define
∆η := {α ∈ Λ : τ ασα ∈ η} .
Lemma 3.14. Let Λ be a finitely aligned LCSC. For every η ∈ E0 satisfying condition
(∗) we have that ∆η ∈ Λ∗.
Proof. By condition (∗), ∆η 6= ∅. Set α ∈ Λ, β ∈ ∆η such that α ≤ β. Then τ βσβ ≤
τ ασα. Since τ βσβ ∈ η and η is a filter, we have that τ ασα ∈ η, whence α ∈ ∆η.
Finally, suppose α, β ∈ ∆η. Then τ ασα, τ βσβ ∈ η, so that τ ασατ βσβ =Wε∈α∨β τ εσε ∈
η. By condition (∗) there exists δ ∈ (α ∨ β) ∩ ∆η, and since τ δσδ ≤ τ ασα, τ βσβ we have
that α, β ≤ δ ∈ ∆η, as desired.
(cid:3)
Definition 3.15. If Λ is a finitely aligned LCSC, we define
Thus,
bE∗ = {η ∈ E0 : η satisfies condition (∗)}.
Corollary 3.16. If Λ is a finitely aligned LCSC, then
is a well-defined map.
Φ : bE∗ −→ Λ∗
η 7−→ ∆η ,
Now, we will construct an inverse for this map.
Definition 3.17. Given F ∈ Λ∗, we define
ηF := {f ∈ E : f ≥ τ ασα for some α ∈ F } .
Lemma 3.18. If Λ is a finitely aligned LCSC, then for every F ∈ Λ∗ we have that
ηF ∈ bE∗.
Proof. Since F 6= ∅, the set {τ ασα : α ∈ F } ⊆ ηF , whence ηF 6= ∅. Set e ∈ ηF , f ∈ E
such that e ≤ f . By hypothesis there exists α ∈ F such that τ ασα ≤ e ≤ f then f ∈ ηF .
Now set e, f ∈ ηF . Then, there exists α, β ∈ F such that τ ασα ≤ e, τ βσβ ≤ f . Since
F is directed, there exists γ ∈ F with α, β ≤ γ. Thus, τ γσγ ≤ τ ασατ βσβ ≤ ef , whence
ef ∈ ηF .
Finally, if f ∈ ηF , then there exists α ∈ F such that f ≥ τ ασα. If f =Wm
j=1 τ βj σβj
(written in irredundant form), by Proposition 2.10 there exists 1 ≤ i ≤ m such that
τ βiσβi ≥ τ ασα. Since ηF ∈ E0, we have that τ βiσβi ∈ ηF . Hence, ηF satisfies condition
(∗), so we are done.
(cid:3)
Corollary 3.19. If Λ is a finitely aligned LCSC, then
is a well-defined map.
Ψ : Λ∗ −→ bE∗
F 7−→ ηF ,
,
Lemma 3.20. If Λ is a finitely aligned LCSC, then Φ and Ψ are naturally inverse
bijections.
TIGHT GROUPOIDS OF LCSC
10
Proof. Let η ∈ E0, and compute
Ψ ◦ Φ(η) = Ψ(Φ(η)) = Ψ(∆η) =
= {e ∈ E : e ≥ τ ασα for some α ∈ ∆η}
= {e ∈ E : e ≥ τ ασα ∈ η} .
Thus, Ψ ◦ Φ(η) ⊆ η. On the reverse sense, if e ∈ η and e =Wn
i=1 τ αiσαi, by condition
(∗) there exists 1 ≤ j ≤ n such that τ αj σαj ∈ η. Thus, αj ∈ ∆η and e ≥ τ αj σαj , whence
e ∈ Ψ ◦ Φ(η), and so Ψ ◦ Φ(η) ⊇ η.
Conversely, given F ∈ Λ∗, compute
Φ ◦ Ψ(F ) = Φ(Ψ(F )) = Ψ(ηF ) = ∆ηF
= {α ∈ Λ : τ ασα ≥ τ βσβ for some β ∈ F }
= {α ∈ Λ : α ≤ β for some β ∈ F } .
Clearly, Φ ◦ Ψ(F ) ⊆ F . On the reverse sense, if α ∈ F then τ ασα ∈ ηF , and thus α ∈ Λ
and τ ασα ≥ τ βσβ for some β ∈ F , whence α ∈ Φ ◦ Ψ(F ). Thus, F = Φ ◦ Ψ(F ).
(cid:3)
3.2. Topology of Λ∗. Before tracking E∞ and Etight through Ψ, we need to consider a
suitable topology defined on Λ∗.
Definition 3.21. Let Λ be a finitely aligned LCSC. Then given X, Y ⊂ Λ finite sets,
we define
MX,Y = {F ∈ Λ∗ : X ⊆ F and Y ∩ F = ∅} .
We will endow a topology on Λ∗, with a basis of open sets
{MX,Y : X, Y ⊆ Λ finite sets} .
We will show that Φ and Ψ are continuous (and thus homeomorphism) with these
simplify, we also use U(X, Y ) (see Definition 3.2) to denote the basic open sets of the
On the other side, since bE∗ ⊆ E0, we can equip bE∗ with the induced topology. To
topology for bE∗. Since both bE∞ and bEtight are subspaces of bE∗, in particular the closure
of bE∞ in E0 coincides with the closure of bE∞ in bE∗.
topologies on bE∗ and Λ∗.
is a basis for the topology of bE∗.
i=1. Set e = Wn
Proof. Let e, f1, . . . , fn ∈ E, and consider the basic open set U(X, Y ) where X = {e}
and Y = {fi}n
j=1 τ βi,j σβi,j for 1 ≤ i ≤ m. Define
Σ := {τ βi,j σβi,j }i=1,...,m j=1,...,mi. Now, given η ∈ U(X, Y ), we have that e ∈ η and
η ∩ Y = ∅.
i=1 τ αiσαi, fi = Wmi
Lemma 3.22. Let Λ be a finitely aligned LCSC. Then,
{U(X, Y ) : X = {τ ασα} , Y = {τ βiσβi}n
i=1} ,
Since η enjoys condition (∗), there exists 1 ≤ j ≤ n such that τ αj σαj ∈ η. If η ∩Σ 6= ∅,
then there exists τ βi,j σβi,j ∈ η, whence fi ≥ τ βi,j σβi,j ∈ η, and thus η ∩ Y 6= ∅, a
contradiction. Hence, there exists 1 ≤ j ≤ n such that η ∈ U({τ αj σαj }, Σ).
Conversely, if there exists 1 ≤ j ≤ n such that η ∈ U({τ αj σαj }, Σ), then τ αj σαj ≤ e,
and since τ αj σαj ∈ η we have that e ∈ η. Also, if η ∩ Y 6= ∅, then there exists
TIGHT GROUPOIDS OF LCSC
11
1 ≤ k ≤ m such that fk ∈ η. By condition (∗), there exists 1 ≤ l ≤ mk such that
τ βk,lσβk,l ∈ η, whence η ∩ Σ 6= ∅, a contradiction. Thus, η ∈ U(X, Y ), and then
(cid:3)
i=1 U({τ αiσαi}, Σ), so we are done.
U(X, Y ) =Sn
As a consequence, we have
Lemma 3.23. Let Λ be a finitely aligned LCSC. Then Φ and Ψ are homeomorphisms.
Proof. By Lemma 3.20, both are injections. Since they are mutually inverses, it is enough
to show that they are open maps. We will show that for Φ (the proof for Ψ is analog).
We denote by BΛ = {τ ασα : α ∈ Λ} ⊂ E. By Lemma 3.22, TE = {U(X, Y ) : X, Y ⊂
BΛ finite sets} is a basis for the topology of bE∗. Now, given a finite set E ⊂ BΛ, we
define E = {α ∈ Λ : τ ασα ∈ E} ⊂ Λ. Fix U(X, Y ) ∈ TE for some finite sets X, Y ⊂ BΛ,
and compute
Since Φ is a bijection, η ∩ Y = ∅ if and only if ∆η ∩ Y = ∅, and τ ασα ∈ η if and only if
α ∈ ∆η, whence
Φ(U(X, Y )) = Φ({η ∈ bE∗ : X ⊆ η and Y ∩ η = ∅}) .
Φ(U(X, Y )) = {∆η ∈ Λ : X ⊆ ∆η and Y ∩ ∆η = ∅} .
Since Φ is a bijection
Φ(U(X, Y )) = {C ∈ Λ∗ : X ⊆ C and Y ∩ C = ∅} = M X, Y .
Thus, Φ is open, as desired.
(cid:3)
Now, we will identify both Φ( E∞) and Φ( Etight) in an intrinsic way. To this end, we
will use a key result from [15].
Lemma 3.24 ([15, Lemma 7.3]). Let Λ be a countable finitely aligned LCSC. If C ⊂ Λ
is a directed subset and β ∈ Λ is such that β ⋓ α for every α ∈ C, then there exists
C ⊂ Λ directed subset such that {β} ∪ C ⊆ C. Moreover, if C is hereditary, then so is
C.
Definition 3.25. Given Λ a LCSC, we say that C ∈ Λ∗ is maximal if whenever C ⊂ D
with D ∈ Λ∗ we have that D = Λ. We will denote Λ∗∗ := {C ∈ Λ∗ : C is maximal}.
Lemma 3.26. Let Λ be a countable, finitely aligned LCSC. Then given η ∈ E0 the
following statements are equivalent:
(1) η ∈ E∞,
(2) ∆η ∈ Λ∗∗.
Proof. (1) ⇒ (2). First, if β ∈ Λ and β ⋓ α for every α ∈ ∆η, we will see that β ∈ ∆η.
Notice that β ⋓ α for every α ∈ ∆η if and only if τ ασατ βσβ 6= 0 for every α ∈ ∆η. Now,
i=1 τ αiσαi ∈ η. By Corollary 3.11, there exists 1 ≤ j ≤ n such that τ αj σαj ∈ η,
whence αj ∈ ∆η. Thus, τ βσβe ≥ τ βσβτ αj σαj 6= 0. Since η ∈ E∞, Lemma 3.4 implies
that τ βσβ ∈ η, and thus β ∈ ∆η.
let e =Wn
Now, suppose C ∈ Λ∗ and ∆η ⊂ C. If β ∈ C \ ∆η, since C is directed we have that
β ⋓ α for every α ∈ ∆η, whence by Lemma 3.4 β ∈ ∆η. Thus, ∆η is maximal.
for every e ∈ ηF . In particular, τ βσβτ ασα 6= 0 for every α ∈ F , whence β ⋓ α for every
(2) ⇒ (1). Set F ∈ Λ∗∗, and take ηF ∈ bE∗. First, pick β ∈ Λ such that τ βσβe 6= 0
0 6= τ γσγf ≤ gf = (τ α1σα1 · · · τ αnσαn) n_i=1
τ βiσβi!
=
(τ α1σα1 · · · \τ αiσαi · · · τ αnσαn)τ αiσαiτ βiσβi = 0 ,
n_i=1
TIGHT GROUPOIDS OF LCSC
12
α ∈ F . By Lemma 3.24 there exists F ∈ Λ∗ such that F ∪ {β} ⊂ F . Since F is maximal,
β ∈ F , and thus τ βσβ ∈ ηF . Now, let f ∈ E with f e 6= 0 for every e ∈ ηF . By Lemmas
2.6 and 2.9 we have f =Wn
0 6= f τ ασα = n_i=1
τ βiσβi! τ ασα =
n_i=1
i=1 τ βiσβi, whence for every α ∈ F
τ βiσβiτ ασα =
τ εσε .
n_i=1 _ε∈βi∨α
Suppose that for each 1 ≤ i ≤ n there exists αi ∈ F such that αi ∨ βi = ∅. Define g :=
τ α1σα1 · · · τ αnσαn ∈ ηF . Since F is directed, there exists γ ∈ F with {α1, . . . , αn} ≤ γ,
whence τ γσγ ≤ g. Thus,
a contradiction. Thus, there exists 1 ≤ j ≤ n such that α ∨ βj 6= ∅ for every α ∈ F .
Thus, βj ∈ F by the previous argument, and hence ηF ∋ τ βj σβj ≤ f , so that f ∈ ηF .
Then, ηF ∈ E∞, as desired.
(cid:3)
Notice that this means that Φ( E∞) = Λ∗∗. Since Φ is continuous and Etight = E∞
k·k bE∗
we have that
Φ(cid:16) Etight(cid:17) = Φ(cid:18) E∞
k·k bE∗(cid:19) = Φ( E∞)
k·kΛ∗
= Λ∗∗k·kΛ∗
=
= {C ∈ Λ : for every finite X, Y ⊂ Λ with C ∈ MX,Y , there exists D ∈ Λ∗∗ ∩ MX,Y }.
Now, we will introduce a couple of definitions for tight hereditary directed subsets of Λ,
and we will show that they are equivalent.
First definition is just the translation of Lemma 3.8 to the context of Λ∗, we need to
recover, and extend, the concept from [15].
Definition 3.27. Let Λ be a LCSC and α ∈ Λ. A subset F ⊂ r(α)Λ is exhaustive with
respect to α if for every γ ∈ αΛ there exists a β ∈ F with β ⋓ γ. We denote FE(α) the
collection of finite sets of r(α)Λ that are exhaustive with respect to α.
Notice that exhaustive sets corresponds to covers.
Definition 3.28. Let Λ be a LCSC and v ∈ Λ0. Then, C ∈ vΛ∗ is tight if for every
i=1 βiΛ we
α ∈ C, every {β1, . . . , βn} ∩ C = ∅ and any finite exhaustive set Z of αΛ \Sn
have that C ∩ Z 6= ∅. We denote by Λtight the set of tight hereditary directed sets.
Now we introduce the new definition.
Definition 3.29. Let Λ be a LCSC and v ∈ Λ0. We say that C ∈ vΛ∗ is E-tight if for
every α ∈ C and every finite set F of Λ with C ∩ F = ∅, there exists D ∈ Λ∗∗ with
α ∈ D and D ∩ F = ∅. We denote by ΛE−t the set of E-tight hereditary directed sets.
We have that ΛE−t = Φ( Etight). Now,
Lemma 3.30. Let Λ be a LCSC. Then ΛE−t = Λtight.
TIGHT GROUPOIDS OF LCSC
13
there exists zk ∈ Z such that w ⋓ zk. But then xzk
⋓ zk, a contradiction.
{z1, . . . , zm} ⊂ αΛ \Sn
Proof. First we prove ΛE−t ⊆ Λtight. Let C ∈ ΛE−t and let Y = {y1, . . . , yn} ⊂ Λ
be a finite set with C ∩ Y = ∅. Let α ∈ C and take any finite exhaustive set Z =
i=1 yiΛ. Suppose that Z ∩ C = ∅. By assumption there exists
D ∈ Λ∗∗ with α ∈ D and D ∩ {Y ∪ Z} = ∅. Since Y ∩ D = ∅, Z ∩ D = ∅ and D is
maximal, by Lemma 3.24 there exist xy1, . . . , xyn ∈ D with xyi⊥yi for every 1 ≤ i ≤ n,
and xz1, . . . , xzm ∈ D with xzj ⊥zj for every 1 ≤ j ≤ m. Therefore, since D is a directed
set, there exists w ∈ D with xyi ≤ w for 1 ≤ i ≤ n, xzj ≤ w for 1 ≤ j ≤ m and
α ≤ w. Observe that w /∈ yiΛ, because otherwise yiΛ ∩ xyiΛ 6= ∅ for every 1 ≤ i ≤ n, a
i=1 yiΛ,
i=1 yiΛ, and since Z is an exhaustive set of αΛ\Sn
contradiction. Thus, w ∈ αΛ\Sn
Now we will prove ΛE−t ⊇ Λtight. Let C ∈ vΛ∗ be tight, α ∈ C, and let F =
{β1, . . . , βn} ⊂ Λ such that F ∩C = ∅. Let Ψ(C) = ηC = {e ∈ E : τ γσγ ≤ e for some γ ∈
C}. Then, τ ασα ∈ ηC and τ βiσβi /∈ ηC for every 1 ≤ i ≤ n, so ηC ∈ U(X, Y ) where
X = {τ ασα} and Y = {τ βiσβi}n
j=1 be any finite cover of E X,Y ,
whence {γj}m
i=1 βiΛ is a finite exhaustive set. Therefore, by hypothesis,
there exists 1 ≤ k ≤ m such that γk ∈ C, and hence τ γk σγk ∈ ηC, so Z ∩ ηC 6= ∅. But
then, by Lemma 3.8, it follows that ηC is a tight filter. Since Etight is the closure of E∞,
there exists ξ ∈ E∞ such that ξ ∈ U(X, Y ). But then Φ(ξ) = ∆ξ ∈ Λ∗∗ by Lemma 3.26,
with α ∈ ∆ξ and ∆ξ ∩ F = ∅, as desired.
(cid:3)
j=1 ⊆ αΛ \Sn
i=1. Let Z = {τ γj σγj }m
E0, and we denote by E∗ = S∞
Example 3.31. Let E = (E0, E1, r, s) be a directed graph. A finite path α of E of length
n ≥ 1 is a sequence α1 · · · αn where αi ∈ E1 for 1 ≤ i ≤ n such that s(αi) = r(αi+1)
for 1 ≤ i ≤ n − 1. Given a path of length n we define s(α) = s(αn) and r(α) = r(α1).
We denote by En be the sets of paths of length n. If we define the paths of length 0 by
i=0 Ei the set of all finite paths of E. An infinite path
α = α1α2 · · · of E is an infinite sequence of edges αi ∈ E1 such that r(αi+1) = s(αi) for
every i ≥ 1. We denote by E∞ the set of infinite paths. A singular vertex of E is a vertex
v ∈ E0 such that r−1(v) ∈ {0, ∞}. We denote by E0
sing the set of singular vertices of E,
we denote by E0
inf = {v ∈ E0 : r−1(v) = ∞}.
sing = E0
Thus, E0
source = {v ∈ E0 : r−1(v) = ∅} and by E0
source ∪ E0
Then we define Λ to be the singly aligned LCSC given by the set of finite paths E∗.
Given a path α ∈ E∗ of length n we define Eα = {α1 · · · αi : 1 ≤ i ≤ n}, where Ev = {v}
for v ∈ E0. Moreover given an infinite path α ∈ E∞ we define Eα = {α1 · · · αi : i ≥ 1}.
It is straightforward to prove that
inf .
Λ∗ = [α∈E∞
Eα ∪ [α∈E ∗
Eα ,
and
Λ∗∗ = [α∈E∞
Eα ∪
[α∈E ∗, r(α)∈E 0
sink
Eα .
Now, given α ∈ E∗ of length n with r(α) ∈ E0
source, then Eα trivially belongs to ΛE−t. Suppose that s(α) ∈ E0
sing, let k ≤ n and β1, . . . , βm ⊆ E∗ \ Eα.
If s(α) ∈ E0
inf . Since
s(α) is an infinite emitter, we have that there exists e ∈ r−1(s(α)) such that e is not
contained in any path of β1, . . . , βm. Now, let γ be a path containing αe that is either
infinite or s(γ) ∈ E0
source. Then, α1 · · · αk ∈ Eγ and Eγ ∩ {β1, . . . , βm, } = ∅. Thus,
Eα ∈ ΛE−t. Conversely, given α ∈ E∗ of length n with s(α) ∈ E0 \ E0
sing, then Y = {αe :
e ∈ r−1(s(α))} is a finite set, and given any Eγ containing α must contain αe for some
e ∈ r−1(s(α)). Thus, Eα /∈ ΛE−t, and consequently Λ∗∗ ( Λtight.
TIGHT GROUPOIDS OF LCSC
14
4. Actions of SΛ
4.1. Basic definitions. We first recall the basic elements about partial actions of in-
verse semigroups on E0 and some subspaces of it. For further references see [9].
Definition 4.1. Let S be an inverse semigroup, let E := E(S) be its semilattice of
idempotents and let E0 be the locally compact Hausdorff space of filters on E. Given
any s ∈ S and any η ∈ E0 with s∗s ∈ η, we define
s · η := {f ∈ E : ses∗ ≤ f for some e ∈ η} ,
which is a filter containing ss∗. This defines a partial action of S on E0. For each s ∈ S,
the domain of s· is
Ds∗s := {η ∈ E0 : s∗s ∈ η} = U({s∗s}, ∅) ,
and the range of s· is Dss∗ := U({ss∗}, ∅). Thus, s· acts by local homeomorphisms. In
particular s· is continuous.
Since s · η ∈ E∞ for every η ∈ E∞ [9, Proposition 3.5], we have that s · η ∈ Etight for
every η ∈ Etight [7, Proposition 12.11].
the reason being that we are interested in define an action of SΛ on Λ∗, using the
homeomorphisms defined in the previous section. The essential step to be covered is to
Now, we specialize to the case of SΛ, when Λ is a finite aligned LCSC and η ∈ bE∗,
show that the action defined on E0 restricts to bE∗.
Lemma 4.2. Let Λ be a finitely aligned LCSC. Let s =Wn
i=1 τ αiσβi ∈ SΛ be irredundant.
Then, for any 1 ≤ i 6= j ≤ n such that βiΛ ∩ βjΛ 6= ∅ and for any η ∈ βiΛ ∩ βjΛ we
have that τ αiσβi(η) = τ αj σβj (η).
First, we need to prove a couple of results.
Proof. Since s ∈ SΛ, τ αiσβi and τ αj σβj are compatible. Thus,
τ αiσβi (ε)σαj σβj (ε),
τ αiσβiτ βj σαj = _ε∈βi∨βj
is an idempotent, whence by Lemma 2.9 we have that αiσβi(ε) = αjσβj (ε) for every
ε ∈ βi ∨ βj. Since η ∈ βiΛ ∩ βjΛ, there exists ε ∈ βi ∨ βj such that
Hence,
τ αiσβi(η) = αiσβi(ε)bη = αjσβj (ε)bη = τ αj σβj (η) .
η = εbη = βibβibη = βjbβjbη.
Lemma 4.3. Let Λ be a finitely aligned LCSC. Let s =Wn
If e =Wk
i=1 τ βiσβi ∈ SΛ for any 1 ≤ k ≤ n, then se =Wk
Proof. By hypothesis, βi (cid:2) βj whenever i 6= j. Fix 1 ≤ k ≤ n, and set t =Wk
k[i=1
TΛ. Now, se ∈ SΛ, and we have that
dom(se) = dom(t) =
βiΛ .
(cid:3)
i=1 τ αiσβi ∈ SΛ be irredundant.
i=1 τ αiσβi in SΛ.
i=1 τ αiσβi ∈
TIGHT GROUPOIDS OF LCSC
15
Let us prove that se = t as function in I(Λ); if so, then we conclude se = t ∈ SΛ. We
will compute the image of any element in dom(se). Pick any 1 ≤ i ≤ k, 1 ≤ j ≤ n.
Then, we have two options:
(1) βiΛ ∩ βjΛ = ∅: in this case, τ βiσβj = 0, and thus τ αiσβiτ βj σβj = τ αiσβi(βj) = 0.
(2) βiΛ ∩ βjΛ =Sε∈βi∨βj
εΛ: in this case
τ αiσβiτ βj σβj = _ε∈βi∨βj
τ αiσβi (ε)σε.
Thus, given any ε ∈ βi ∨ βj and any δ ∈ s(ε)Λ, we have that τ αiσβiτ βj σβj (εδ) =
τ αiσβi(εδ).
Applying Lemma 4.2, we conclude that se = t, as desired.
(cid:3)
Proof. As noticed before, s · η ∈ E0. Since s acts on η, we have s∗s ∈ η and ss∗ ∈ s · η.
If e ∈ η, then s∗se ∈ η. Hence, without lost of generality, we can assume that e = s∗se,
Lemma 4.4. Let Λ be a finite aligned LCSC, let η ∈ bE∗ and s ∈ SΛ. Then, s · η ∈ bE∗.
whence s∗(ses∗)s = e. Moreover, since η ∈ bE∗, we can assume that e = τ γσγ for some
Now, take s =Wn
i=1 τ αiσβi. By [12, Proposition 1.4.17(1)],
γ ∈ Λ.
s∗s =
τ βiσβi.
n_i=1
k=1 τ δk σδk, by Proposition 2.11 we have that
By Lemma 4.3, se = τ αjcβj σγ, whence ses∗ = se(se)∗ = τ αjcβj σαjcβj .
δk ∈ ∆s·ξ, and thus τ δk σδk ∈ s · ξ. Hence, s · ξ satisfies condition (∗), as desired.
Since e ≤ s∗s, by Proposition 2.11 there exists 1 ≤ j ≤ n such that βj ≤ γ, i.e. γ = βjbβj.
Now, given any idempotent f = Wm
sτ γσγs∗ ≤ f if and only if there exists 1 ≤ k ≤ m such that δk ≤ αjbβj ∈ ∆s·ξ. Then,
By restricting our attention to bE∗, we will use the notation Ds∗s and Dss∗ to refer to
the domain and range of the action of an element s ∈ SΛ on bE∗. Then, given s ∈ SΛ,
we can write (in a unique way up to irredundacy) s = Wn
that s∗s = Wn
i=1 τ βiσβi by Lemma 2.9, and thus Ds∗s ⊆ Sn
i=1 Dτ βi σβi = Ds∗s. Analogously,Sn
i=1 Dτ βi σβi ⊆ Ds∗s. Thus,Sn
Sn
4.2. The partial action on Λ∗. Since we have an homeomorphism Ψ : bE∗ → Λ∗ with
inverse Ψ (Lemma 3.23) we can transfer the action of SΛ on bE∗ to Λ∗. First we will fix
i=1 τ αiσβi by Lemma 2.6, so
i=1 Dτ βi σβi . On the other
side, if τ βiσβi ∈ η for some 1 ≤ i ≤ n, then τ βiσβi ≤ s∗s implies that s∗s ∈ η, whence
the domain and range.
i=1 Dτ αi σαi = Dss∗.
Definition 4.5. Let s = τ ασβ ∈ SΛ. Then, we define
(cid:3)
Eα := Ess∗ = Φ(Dss∗) = Φ(U({α}, ∅)) = {C ∈ Λ∗ : α ∈ C}
and
Eβ := Es∗s = Φ(Ds∗s) = Φ(U({β}, ∅)) = {C ∈ Λ∗ : β ∈ C} .
TIGHT GROUPOIDS OF LCSC
16
Given s =Wn
i=1 τ αiσβi we can define
Es∗s =
Eβi =
Φ(Dτ βi σβi ) = Φ(Ds∗s) .
n[i=1
n[i=1
The sets Es∗s and Ess∗ are the natural candidates for being the domain and range of
the partial action of SΛ on Λ∗.
Next step is to define the action. We will start by defining the action in the particular
case of s = τ ασβ. In this case, F ∈ Eβ if and only if β ∈ F . Then we define
τ ασβ · F = [β≤γ, γ∈F
[ασβ(γ)] ,
where [δ] is the set of initial segments of δ ∈ Λ, who clearly belong to Λ∗. Indeed:
(1) τ ασβ(β) = α, thus, τ ασβ · F 6= ∅.
(2) Set η1, η2 ∈ τ ασβ · F , this means that there exist γ1, γ2 extensions of β with
γ1, γ2 ∈ F such that ηi ≤ ασβ(γi) for i = 1, 2. Since F is directed, β ≤ γ1, γ2 ≤ δ
for some δ ∈ F . Thus,
(3) If δ ∈ τ ασβ · F and η ≤ δ, then there exists γ ≥ β and γ ∈ F such that
τ ασβ(γ1), τ ασβ(γ2) ≤ τ ασβ(δ) .
η ≤ δ ≤ ασβ(γ), so that η ∈ τ ασβ · F .
Moreover, τ ασβ · F ∈ Eα. Thus, in this case we have that
τ ασβ· : Eβ → Eα ,
F 7→ τ ασβ · F ,
is a well-defined map.
Now, set s =Wn
i=1 τ αiσβi ∈ SΛ, with
n[i=1
Es∗s =
Eβi
and
Ess∗ =
Eαi .
n[i=1
By Lemma 1.13, τ αiσβi and τ αj σβj are compatible for 1 ≤ i, j ≤ n, and then τ αiσβiτ βj σαj
and τ βiσαiτ αj σβj are idempotents in SΛ.
For every 1 ≤ i ≤ n and every F ∈ Eβi, we can define
s · F = τ αiσβi · F .
The only point to be checked is that, if F ∈ Eβi ∩ Eβj for 1 ≤ i 6= j ≤ n, then
τ αiσβi · F = τ αj σβj · F . To check this observe that, if F ∈ Eβi ∩ Eβj , then we have that
βi, βj ∈ F ∈ Λ∗. Thus, there exists γ ∈ F with βi, βj ≤ γ. Without lost of generality we
can assume that γ = βi ∨ βj. Then, γ = βiσβi(γ) = βjσβj (γ). But
τ αiσβi (ε)σαj σβj (ε) ∈ E(SΛ) ,
τ αiσβiτ βj σαj = _ε∈βi∨βj
so that for every ε ∈ βi ∨ βj we have that αiσβi(ε) = αjσβj (ε). In particular, for the γ
above, we have that αiσβi(γ) = αjσβj (γ). Hence,
[βi,βj≤γ, γ∈F
[αiσβi(γ)] = [βi,βj≤γ, γ∈F
[αjσβj (γ)] ,
that is τ αiσβi · F = τ αj σβj · F , as desired.
TIGHT GROUPOIDS OF LCSC
17
Because of this fact, we can define the map
as follows: given F ∈ Es∗s =Sn
F and βk+1, . . . , βn /∈ F . Thus,
s· : Es∗s → Ess∗ ,
i=1 Eβi we can assume, after re-indexing, that β1, . . . , βk ∈
s · F =
[Wk
i=1 βi≤γ, γ∈F
[αiσβi(γ)] .
4.3. The Key Lemma. Now, we will prove a result, essential to fix the dictionary.
Lemma 4.6. Let Λ be a finitely aligned LCSC. Then, for every s ∈ SΛ and any η ∈
i=1 τ αiσβi, so that s∗s =Wi=1 τ βiσβi and
Ds∗s ∩ bE∗ we have that s · ∆η = ∆s·η.
Proof. Given any s ∈ SΛ, we have that s =Wn
Ds∗s =Sn
i=1 Dτ βi σβi .
By the previous arguments, we can restrict the action of s on Ds∗s to an action of
si := τ αiσβi on Dτ βi σβi for each particular filter η ∈ Dτ βi σβi . Hence, we can reduce the
question to the case s = τ ασβ, s∗s = τ βσβ and η ∈ Dτ β σβ .
First observe that
s · η = {f ∈ E(SΛ) : f ≥ τ ασβeτ βσα for some e ∈ η} .
τ γj σγj ∈ η, whence
i=1 τ γiσγi, then there exists 1 ≤ j ≤ n such that
Since η satisfies condition (∗), if e =Wn
Now, τ ασβτ γσγτ βσα = Wǫ∈β∨γ τ ασβ (ǫ)σασβ (ǫ). Since τ βσβ ∈ η, we have that β ∈ ∆η.
s · η = {f ∈ E(SΛ) : f ≥ τ ασβτ γσγτ βσα for some γ ∈ ∆η} .
Hence, there exists ǫ ∈ (β ∨ γ) ∩ ∆η. Thus,
τ ασβτ ǫσǫτ βσα = τ ασβ (ǫ)σασβ (ǫ) ≤ τ ασβτ γσγτ βσα
for some ǫ ∈ (β ∨ γ) ∩ ∆η, and thus
s · η = {f ∈ E(SΛ) : f ≥ τ ασβ (γ)σασβ (γ) for some γ ∈ ∆η with β ≤ γ} .
that f ≥ τ ασβ (γ)σασβ (γ) if and only if there exists 1 ≤ k ≤ n such that δk ≤ ασβ(γ) for
some γ ∈ ∆η with β ≤ γ. Hence, we have that
i=1 τ δiσδi, since s · η ∈ bE∗ by Lemma 4.4, then by Proposition 2.10 we have
τ δiσδi ∈ E(SΛ) : there is 1 ≤ i ≤ n such that δi ≤ ασβ(γ) for some γ ∈ ∆η) .
Given f =Wn
s·η =( n_i=1
∆s·η = {δ ∈ Λ : δ ≤ ασβ(γ) for some γ ∈ ∆η with β ≤ γ} = [β≤γ, γ∈∆η
Thus,
[ασβ(γ)] = s · ∆η ,
(cid:3)
as desired.
Corollary 4.7. Let Λ be a countable, finite aligned LCSC. Then given s ∈ SΛ:
(1) s· restricts to an action on Λ∗∗,
(2) s· restricts to an action on Λtight.
TIGHT GROUPOIDS OF LCSC
18
Proof. Lemma 4.6 shows that for every s ∈ SΛ and for every η ∈ Ds∗s we have that
s · Φ(η) = Φ(s · η).
For (1) since Φ( E∞) = Λ∗∗, for any F ∈ Λ∗∗ we have that ηF = Ψ(F ) ∈ E∞, so that
s · F = s · ∆ηF = s · Φ(ηF ) = Φ(s · ηF ) ∈ Φ( E∞) = Λ∗∗ .
For (2) since s· is continuous and Λtight = Λ∗∗k·kΛ∗
4.4. The tight groupoid. Given an action of an inverse semigroup S on a locally
compact Hausdorff space X, we can associate to it a groupoid as follows: Consider
S × X := {(s, x) : x ∈ Ds∗s} with
the result derives from (1).
(cid:3)
(1) d(s, x) = x and r(s, x) = s · x,
(2) (s, x) · (t, y) is defined if t · y = x, and then (s, x) · (t, y) = (st, y),
(3) (s, x)−1 = (s∗, s · x) .
We say that (s, x) ∼ (t, y) if and only if x = y and there exists e ∈ E(S) with x ∈ De and
se = te. This is an equivalence relation, compatible with the groupoid structure. Thus,
we define S ⋊ X := S × X/ ∼, with the induced operations defined above. Moreover,
(S ⋊ X)(0) = X. Now to define the topology on S ⋊ X, given s ∈ S and U ⊆ Ds∗s an
open set, the subset
Θ(s, U) = {[s, x] : x ∈ U} ,
gives us a basis for S ⋊ Xunder which it is a locally compact ´etale groupoid.
When X = Etight, S ⋊ X is the tight groupoid of the inverse semigroup, denote by
Gtight(S). For extra information see for example [9].
We will show a nice description of Gtight(SΛ).
i=1 τ αiσβi ∈ SΛ and let ξ ∈
Ds∗s. Suppose that βk ∈ ∆ξ for some 1 ≤ k ≤ n. Then [τ αk σβk, ξ] = [s, ξ] ∈ Gtight(SΛ).
In particular
Gtight(SΛ) = {[τ ασβ, ξ] : s(α) = β, β ∈ ∆ξ} .
Lemma 4.8. Let Λ be a finitely aligned LCSC. Let s =Wn
Proof. Let s =Wn
Finally, given any [s, ξ] ∈ Gtight(SΛ) with s =Wn
i=1 τ αiσβi ∈ SΛ and let [s, ξ] ∈ Gtight(SΛ). Then by definition [s, ξ] =
[se, ξ] for any e ∈ ξ. Let 1 ≤ k ≤ n such that βk ∈ ∆ξ, whence e = τ αk σβk ∈ ξ. Then by
Lemma 4.3 we have that se = τ αk σβk. Thus, [s, ξ] = [τ αk σβk].
βk ∈ ∆ξ, since ξ satisfies condition (∗). So by the above [s, ξ] = [τ αk σβk, ξ].
i=1 τ αiσβi and ξ ∈ Etight, there exists
(cid:3)
Now we are ready to prove the following result.
Lemma 4.9. Let Λ be a finite aligned LCSC. Then, Gtight(SΛ) is topologically isomorphic
to SΛ ⋊ Λtight.
Proof. Since Φ : Etight → Λtight is a homeomorphism and for every s ∈ SΛ and η ∈ Etight,
we have that s · Φ(η) = Φ(s · η), we conclude that the map
ρ : SΛ × Etight → SΛ × Λtight
given by
(s, η) 7→ (s, Φ(η)) ,
is a groupoid isomorphism.
Now set s, t ∈ SΛ, η ∈ Ds∗s ∩ Dt∗t and e ∈ E(SΛ) with η ∈ De and se = te. Then,
Φ(η) ∈ Es∗s ∩ Et∗t and ∆η ∈ Ee, so that
(s, η) ∼ (t, η)
if and only if
(s, ∆η) ∼ (t, ∆η) .
TIGHT GROUPOIDS OF LCSC
19
Consequently, the isomorphism ρ induces an isomorphism
ρ : SΛ ⋊ Etight → SΛ ⋊ Λtight
given by
[s, η] 7→ [s, Φ(η)] .
Finally, given any s ∈ SΛ and U ⊆ Ds∗s open subset, we have that Φ(U) ⊆ Es∗s is an
open set and ρ(Θ(s, U)) = Θ(s, Φ(U)). Thus, ρ is a homeomorphism.
(cid:3)
We are ready to show when this groupoid is Hausdorff. First, we need to recall some
known facts.
Definition 4.10. A poset is a weak semilattice if the intersection of principal downsets
is finitely generated as a downset.
In the case of Λ being right cancellative, Λ can be seen as a subsemigroup of SΛ via
the natural map α 7→ τ α. Hence, when Λ is a left and right cancellative small
category, we have the following result.
Proposition 4.11 ([6, Proposition 3.6]). Let Λ be a left and right cancellative small
category. Then the following are equivalent:
(1) Λ is finitely aligned,
(2) Z(Λ) is a weak semilattice.
An important point is that, when Λ fails to be right cancellative, then Donsig & Millan
argument, fails.
Remark 4.12. Steinberg [18, page 1037] says that an inverse semigroup S is Hausdorff
if it is a weak semilattice
The following follows from [18, Section 5].
Corollary 4.13. Let S be a countable inverse semigroup. If S is Hausdorff, then so is
Gtight(S).
Thus,
Corollary 4.14. If Λ is a countable, finite aligned left and right cancellative small
category, then Gtight(SΛ) ∼= SΛ ⋊ Λtight is Hausdorff.
Proof. The conclusion follows by Lemma 4.9 and Corollary 4.13.
(cid:3)
If Λ fails to be right cancellative, Corollary 4.14 would fail in general.
4.5. Universal tight representations. In this subsection we quickly revisit the results
proved in [6], just fixing the essential hypotheses required to guarantee that these results
hold.
First, notice that the results of [6, Sections 1.1 and 2] do not require Λ to be other
than LCSC. In particular, the key result is [6, Proposition 3.4], that works correctly for
SΛ.
Second, to apply the results of [6, Section 3], we only need to fix the following facts:
(1) As noticed in [16, Remark before Theorem 10.10], the result required to prove
[6, Theorem 3.7] (namely, [15, Theorem 8.2]) do not depend on the amenability
of Spielberg's groupoid G∂Λ. Thus,
(2) Given Λ a (countable) finitely aligned LCSC, we define:
TIGHT GROUPOIDS OF LCSC
20
(a) C ∗(Λ) is the universal C ∗-algebra generated by a family {Tα : α ∈ Λ}
satisfying:
α for every v ∈ Λ0 and for all F ⊂ vΛ finite exhaustive
(b) Given any unital commutative ring R, we define RΛ the R-algebra generated
αTα = Ts(α).
(i) T ∗
(ii) TαTβ = Tαβ if s(α) = r(β).
(iii) TαT ∗
γ .
αTβT ∗
β =Wγ∈α∨β TγT ∗
(iv) Tv =Wα∈F TαT ∗
set.
by a family {Tα : α ∈ Λ} satisfying:
αTα = Ts(α).
(i) T ∗
(ii) TαTβ = Tαβ if s(α) = r(β).
(iii) TαT ∗
γ .
αTβT ∗
β =Wγ∈α∨β TγT ∗
(iv) Tv =Wα∈F TαT ∗
set.
α for every v ∈ Λ0 and for all F ⊂ vΛ finite exhaustive
In order to relate these algebras with the associated tight groupoid, we need to show
that the natural representations π : SΛ → C ∗(Λ) and π : SΛ → RΛ are universal tight.
With respect to its tighness, Donsig and Millan [6, Theorem 3.7] showed that these
representations are cover-to-joint, and the concluded that they are tight. As recently
observed by Exel [8], that could fail, so that there is an slight imprecision in the proof of
[6, Theorem 2.2]. Fortunately, Exel solved this problem [8, Corollary 5.2, Theorem 6.1],
so that the conclusion remains true. Hence, by [16, Theorem 10.15] and [6, Theorem
3.7], we have the following result.
Proposition 4.15. Let Λ be a (countable) finitely aligned LCSC. Then:
(1) The natural semigroup homomorphism
π :
SΛ → C ∗(Λ)
7→ TαT ∗
τ ασβ
β
is a universal tight representation of SΛ in the category of C ∗-algebras.
(2) For any unital commutative ring R, the natural semigroup homomorphism
π :
SΛ → RΛ
7→ TαT ∗
τ ασβ
β
is a universal tight representation of SΛ in the category of R-algebras.
Hence, because of [7, Theorem 13.3], Proposition 4.11 and [18, Corollary 5.3], we have
Theorem 4.16. Let Λ be a (countable) finitely aligned LCSC. Then:
(1) C ∗(Λ) ∼= C ∗(Gtight(SΛ)).
(2) For any unital commutative ring R, RΛ ∼= AR(Gtight(SΛ)).
In [14] Spielberg defines a groupoid G∂Λ for a category of paths Λ. We will show that
5. Spielberg's groupoid
this groupoid is topologically isomorphic to Gtight(SΛ).
First, Spielberg defines a topology in Λ∗ that coincides with the topology we introduce
Indeed, for α ∈ Λ and β1, . . . , βn ∈ αΛ \ {α}, and setting
in the definition 3.21.
TIGHT GROUPOIDS OF LCSC
21
E = αΛ \Sn
i=1 βiΛ we have that
E = {C ∈ Λ∗ : C ∩ γΛ ⊆ E for some γ ∈ C} = M{α},{β1,...,βn} .
Therefore, we have that ∂Λ = Λ∗∗ = Λtight.
Next result is a refined version of [15, Lemma 4.12].
Lemma 5.1. Let Λ be a finitely aligned LCSC. Let F, G ∈ Λ∗ and α, β ∈ Λ such that
τ α · F = τ β · G. Then there exists δ ∈ F and γ ∈ G such that αδ = βγ.
Proof. Let δ′ ∈ F , then αδ′ ∈ τ β · G. By definition, there exists γ′ ∈ G such that
αδ′ ≤ βγ′. Then, there is η ∈ Λ such that αδ′η = βγ′. Now, there exists ξ ∈ F such
that βγ′ ≤ αξ, and hence there is η′ ∈ Λ such that βγ′η′ = αξ, whence αδ′ηη′ = αξ.
Now, by left cancellation, we have that δ′ηη′ = ξ, so δ′η ≤ ξ ∈ F . Then, since F is
hereditary, it follows that δ′η ∈ F too. If we define δ := δ′η and γ := γ′, we are done. (cid:3)
Now, we recall the definition of Spielberg's groupoid associated to a small category
(see e.g. [14, pp. 729-730]). We start defining an equivalence relation on Λ × Λ × Λ∗ by
saying that (α, β, F ) ∼ (α′, β′, F ′) if there exist G ∈ Λ∗, γ, γ′ ∈ Λ such that F = τ γ · G,
F ′ = τ γ ′
· G, αγ = α′γ′ and βγ = β′γ′. Denote G = Λ × Λ × Λ∗/ ∼. Now, we define a
partial operation on G. To this end, fix the set of composable pairs
G(2) := {([α, β, F ], [γ, δ, G]) : τ β · F = τ γ · G)} ,
and define [α, β, F ]−1 = [β, γ, F ]. Given a pair ([α, β, F ], [γ, δ, G]) ∈ G(2) we define the
multiplication by
[α, β, F ][γ, δ, G] = [αξ, δη, H] ,
where ξ ∈ F and η ∈ G are the elements given in Lemma 5.1 such that βξ = γη, and
H = σξ · F = ση · G. Finally, the sets [α, β, U] := {[α, β, F ] : F ∈ U} for U an open
subset of Λ∗ forms a basis for the topology of G, under which G is an ´etale groupoid. By
Corollary 4.7, we have that G∂Λ = {[α, β, F ] ∈ G : F ∈ Λtight}.
Proposition 5.2. Let Λ be a countable, finitely aligned LCSC. Then the map
Φ : G∂Λ → SΛ ⋊ Λtight ,
[α, β, F ] 7→ [τ ασβ, τ β · F ] ,
is an isomorphism of topological groupoids.
Proof. First, let (α, β, F ) ∼ (α′, β′, F ′), that this, there exist G ∈ Λ∗, γ, γ′ ∈ Λ such that
F = τ γ · G, F ′ = τ γ ′
τ β · F = τ β · (τ γ · G) = τ βγ · G = τ β′γ ′
· G, αγ = α′γ′ and βγ = β′γ′. Then
· G = τ β′
· (τ γ ′
· G) = τ β′
· F ′.
Now, βγ ∈ τ β · F = τ β′ · F ′, and
(τ ασβ)(τ βγσβγ) = τ αγσβγ = τ α′γ ′
σβ′γ ′
= (τ α′
σβ′
)(τ β′γ ′
σβ′γ ′
) = (τ α′
σβ′
)(τ βγσβγ).
Hence, (τ ασβ, τ β · F ) ∼ (τ α′
σβ′
, τ β′
· F ′), and thus, Φ is a well-defined map.
Suppose that ([α, β, X], [γ, δ, Y ]) is a composable pair in G∂Λ. Since τ β · X = τ γ · Y ,
by [15, Lemma 4.12] there exist ξ, η ∈ Λ, and Z ∈ Λtight such that X = τ ξ · Z, Y = τ η · Z
and βξ = γη. Then, Φ([α, β, X]) = [τ ασβ, τ β · X], Φ([γ, δ, Y ]) = [τ γσδ, τ δ · Y ], and
Φ([α, β, X][γ, δ, Y ]) = Φ([αξ, δη, Z]) = [τ αξσδη, τ δη ·Z]. Notice that, since τ γσδ ·(τ δ ·Y ) =
τ γ · Y = τ γη · Z = τ βξ · Z = τ β · X, we can compute
[τ ασβ, τ β · X][τ γσδ, τ δ · Y ] = [τ ασβτ γσδ, τ δ · Y ].
TIGHT GROUPOIDS OF LCSC
22
On one side, τ δ · Y = τ δη · Z. On the other side, since βξ = γη, we have [τ ασβ, τ β · X] =
[τ αξσβξ, τ βξ · Z], [τ γσδ, τ δ · Y ] = [τ γησδη, τ δη · Z], and thus
[τ ασβ, τ β · X][τ γσδ, τ δ · Y ] = [τ αξσβξτ γησδη, τ δη · Z] =(1)
Since τ αξσβξτ γησδη = τ αξσγητ γησδη = τ αξσδη, we have
(1) = [τ αξσδη, τ δη · Z].
So, Φ is a groupoid homomorphism.
Suppose that [α, β, X], [γ, δ, Y ] in G∂Λ such that
Φ([α, β, X]) = [τ ασβ, τ β · X] = [τ γσδ, τ δ · Y ] = Φ([γ, δ, Y ]).
Then, τ β · X = τ δ · Y . By Lemma 5.1, there exist ξ ∈ X, η ∈ Y such that βξ = δη. Then,
the idempotent e = τ βξσβξ = τ δησδη lies in the right domain, and since [τ ασβ, τ β · X] =
[τ γσδ, τ δ · Y ], left cancellation give us τ αξσβξ = τ ασβ · e = τ γσδ · e = τ γησγη. Thus,
τ αξ = τ γη, whence αξ = γη, and hence [α, β, X] = [γ, δ, Y ]. So, Φ is injective.
Finally, let [τ ασβ, F ] ∈ SΛ ⋊ Λtight. Then Φ([α, β, σβ · F ]) = [τ ασβ, τ β · (σβ · F )] =
[α, β, τ βσβ · F ]. But since β ∈ F it follows that τ βσβ · F = F , so Φ is exhaustive, and
hence Φ is a topological groupoid isomorphism.
(cid:3)
6. Simplicity
In [15, Section 10] are given conditions in a category of paths Λ for G∂Λ being topolog-
ically free, minimal and locally contractive, but right-cancellation of Λ is crucial in the
proofs therein. We are going to use the isomorphism in Proposition 5.2 and the charac-
terization of these properties given in [9], to extend Spielberg results in [15, Section 10]
to finitely aligned LCSC.
Definition 6.1. Let S be an inverse semigroup, and let s ∈ S. Given an idempotent
e ∈ E such that e ≤ s∗s, we will say that:
(1) e is fixed under s, if se = e,
(2) e is weakly-fixed under s, if (sf s∗)f 6= 0, for every non-zero idempotent f ≤ e.
Definition 6.2. Given an action α : S y X, let s ∈ S, and let x ∈ Ds∗s.
(1) αs(x) = x, we will say that x is a fixed point for s. We denote by Fs the set of
fix points for s.
(2) If there exists e ∈ E, such that e ≤ s, and x ∈ De, we will say that x is a trivially
fixed point for s.
(3) We say that α is a topologically free action, if for every s in S, the interior of the
set of fixed points for s consists of trivial fixed points.
Given an action α : S y X, the groupoid S ⋊ X is effective if and only if the action
α is topologically free [9, Theorem 4.7].
Remark 6.3. Let Λ be a finitely aligned LCSC, and let SΛ y Etight be the associated
i=1 τ βiσβi ∈ ξ.
Since ξ satisfy condition (∗), there exists 1 ≤ j ≤ n such that τ βj σβj ∈ ξ. Let C ∈ Λtight
such that ξ = ηC. Then we have that βj ∈ C. By the definition of the action SΛ y Λtight
action. Let s ∈ SΛ, and ξ ∈ Ds∗s ∩ Etight. If s =Wn
i=1 τ αiσβi, then s∗s =Wn
TIGHT GROUPOIDS OF LCSC
23
we have that s·C = τ αj σβj ·C, and hence s·ξ = τ αj σβj ·ξ. Thus, without lost of generality,
we can assume that s = τ αj σβj .
Theorem 6.4. Let Λ be a countable, finitely aligned LCSC. If either Gtight(SΛ) is Haus-
dorff or E∞ = Etight, then the following are equivalent:
(1) Gtight(SΛ) is effective.
(2) For every s ∈ SΛ, and for every e ∈ EΛ which is weakly-fixed under s, there exists
a finite cover for e consisting of fixed idempotents.
(3) Given α, β ∈ Λ with r(α) = r(β) and s(α) = s(β), if αδ ⋓ βδ for every δ ∈ s(α)Λ
then there exists F ∈ FE(s(α)) such that αγ = βγ for every γ ∈ F .
Proof. The equivalence of (1) and (2) follows from Lemma 4.9 and Corollary 4.14 and
[9, Theorem 4.10, Theorem 3.16 and Theorem 4.7].
Let s = Wn
s∗s = Wn
2.5] there exists e = Wm
We assume (3), and we will prove that condition (iii) of [9, Theorem 4.10] holds.
i=1 τ αiσβi ∈ SΛ, and let ξ ∈ E∞ ∩ Ds∗s with s · ξ = ξ and ξ ∈ (Fs)◦. Let
i=1 τ βiσβi. From Remark 6.3, there exist C ∈ Λ∗∗ such that ξ = ηC, and
1 ≤ j ≤ n such that ξ ∈ Dτ βj σβj and s · ξ = τ βj σβj · ξ. Now, by [9, Proposition
k=1 τ γk σγk ∈ ξ with e ≤ s∗s such that ξ ∈ De ∩ E∞ ⊆ (Fs)◦.
Since ξ satisfies condition (∗) there exists 1 ≤ k ≤ m such that τ γk σγk ∈ ξ, whence
ξ ∈ Dτ γk σγk ∩ E∞ ⊆ (Fs)◦. Then, without lost of generality, we can assume that s = τ ασβ,
s∗s = τ βσβ, ξ = ηC with C ∈ Λ∗∗, β ∈ C, and there exists γ ∈ C with τ γσγ ≤ τ βσβ
such that ξ ∈ Dτ γ σγ ⊆ (Fs)◦. Then γ = βγ for some γ ∈ Λ, and since by hypothesis
τ ασβ · C = C, we have that τ ασβ(βγ) = αγ ∈ C . Now, by [9, Lemma 4.9], Dτ γ σγ ⊆ Fs
is equivalent to τ γσγ being weakly fixed under τ ασβ. But this means for every δ ∈ s(γ)Λ
we have that
αγδ ⋓ βγδ .
By hypothesis, there exists F ∈ FE(s(γ)) such that αγδ = βγδ for every δ ∈ F . We
claim that there exists δ ∈ F such that αγδ = βγ δ ∈ C. Indeed, since αγ ∈ C, we have
that E := σαγ · C ∈ Λ∗∗ ⊆ Λtight, ηE ∈ U(τ s(γ)σs(γ), ∅) and {τ δσδ : δ ∈ F } is a cover of
U(τ s(γ)σs(γ), ∅). Since ηE is a tight filter (because E∞ ⊆ Etight) there exists δ ∈ F with
τ δσδ ∈ ηE. Then, δ ∈ E = σαγ · C and hence αγδ ∈ C, as desired.
Now, we define g := τ αγ δσαγ δ ∈ EΛ. Then, 0 6= τ αγ δσαγ δ ≤ τ ασβ and ξ = ηC ∈
Dτ αγ δσαγ δ . Therefore, ξ is trivially fixed by s. Thus, condition (iii) of [9, Theorem 4.10]
is satisfied, and since either Gtight(SΛ) is Hausdorff or E∞ = Etight, then condition (2) is
satisfied by [9, Theorem 4.10], as desired.
Finally, let us assume (2). Let α, β ∈ Λ with r(α) = r(β) and s(α) = s(β) and
satisfying that αδ ⋓ βδ for every δ ∈ s(α), then the idempotent e = τ βσβ is weakly fixed
under s := τ ασβ, so by hypothesis there exists a finite cover Z of e consisting of fixed
idempotents under s. Then there exists a finite set F ⊆ s(β)Λ such that Z = {τ βγσβγ}.
Since the idempotents of Z are fixed under s, we have that
τ ασβ · τ βγσβγ = τ αγσβγ = τ βγσβγ ,
for every γ ∈ F . Thus, αγ = βγ for every γ ∈ F . But Z is a cover of τ βσβ, and hence
for every δ ∈ s(β)Λ there exists γ ∈ F such that τ βδσβδ · τ βγσβγ 6= 0, but this means
that βδ ⋓ βγ, and hence δ ⋓ γ by left-cancellation. Thus, F ∈ FE(s(β)), as desired. (cid:3)
TIGHT GROUPOIDS OF LCSC
24
Remark 6.5. Observe that if Λ has right cancellation, condition (3) in Theorem 6.4
reduces to aperiodicity as defined in [15, Definition 10.8]
Theorem 6.6 ([15, Theorem 10.14] & [9, Theorem 5.5]). If Λ is a countable, finitely
aligned LCSC, then the following statements are equivalent:
(1) Gtight(SΛ) is minimal.
(2) For every nonzero e, f ∈ EΛ, there are s1, . . . , sn ∈ SΛ, such that {sif s∗
i }n
i=1 is
an outer cover for e.
(3) For every α, β ∈ Λ there exists F ∈ FE(α) such that for each γ ∈ F , s(β)Λs(γ) 6=
∅.
Proof. By [9, Theorem 5.5] it is enough to prove the equivalence of conditions (2) and
(3). First, we will prove (3) ⇒ (2). Without lost of generality, we can assume that
e = τ ασα and f = τ βσβ. By hypothesis there exists F = {γ1, . . . , γn} ∈ FE(α) such that
for each i there exists δi ∈ s(β)Λs(γi). If we define si := τ γiσβδi for every 1 ≤ i ≤ n, we
have that {sif s∗
i=1, that is a cover for e.
i=1 = {τ γiσγi}n
i }n
(2) ⇒ (3). Let e = τ ασα and f = τ βσβ. By assumption there exist s1, . . . , sn such
i=1 is an outer cover of e. Without lost of generality we can assume that
that {sif s∗
i }n
i=1 τ γiσδi, and hence sf s∗ is an outer cover of e. But
n = 1, so s := s1 =Wm
sf s∗ =
τ γiσδi (εi)σγiσδi (εi) .
m_i=1 _εi∈β∨δi
Since Λ is finitely aligned β ∨ δi is finite for every 1 ≤ i ≤ m, and so the set {γiσδi(εi) :
1 ≤ i ≤ m , εi ∈ β ∨ δi} ∈ FE(α). Finally, observe that since εi ∈ β ∨ δi it follows that
s(β)Λs(εi) 6= ∅, but s(εi) = s(γiσδi(εi)).
(cid:3)
Then, we have the following result
Theorem 6.7. Let Λ be a countable, finitely aligned LCSC. If either Gtight(SΛ) is Haus-
dorff or E∞ = Etight, then the following statements are equivalent:
(1) C ∗(Λ) is simple.
(2) For any field K, KΛ is simple.
(3) The following properties hold:
(a) Given α, β ∈ Λ with r(α) = r(β) and s(α) = s(β), if αδ ⋓ βδ for every
δ ∈ s(α)Λ then there exists F ∈ FE(s(α)) such that αγ = βγ for every
γ ∈ F .
(b) For every α, β ∈ Λ there exists F ∈ FE(α) such that for each γ ∈ F ,
s(β)Λs(γ) 6= ∅.
Proof. By Theorem 4.16, C ∗(Λ) ∼= C ∗(Gtight(SΛ)), and for any field K, KΛ ∼= AK(Gtight(SΛ)).
By Theorem 6.4, condition (2(a)) is equivalent to Gtight(SΛ) being effective, and by The-
orem 6.6, condition (2(b)) is equivalent to Gtight(SΛ) being minimal. Then, (1) ⇔ (3) by
[3, Theorem 5.1], while (2) ⇔ (3) by [18, Theorem 3.5].
(cid:3)
7. Zappa-Sz´ep products of LCSC categories
In this section we will analyze the notion of Zappa-Sz´ep products of LCSC categories,
introduced in [2], inspired in the construction of self-similar graphs defined in [9].
TIGHT GROUPOIDS OF LCSC
25
Let Λ be a finitely aligned LCSC and let G be a discrete group (with unit 1G). We
will use multiplicative notation for the group operation.
We say that the group G acts on Λ by permutations when
r(g · α) = g · r(α)
and
s(g · α) = g · s(α)
for every α ∈ Λ, g ∈ G .
For the rest of the section we will assume that G acts by permutations on Λ.
A cocyle for the action of G on Λ is a function ϕ : G × Λ → G satisfying the cocyle
identity
ϕ(gh, α) = ϕ(g, h · α)ϕ(h, α)
for all g, h ∈ G, α ∈ Λ .
In particular the cocycle identity says that ϕ(1G, α) = 1G for every α ∈ Λ.
Definition 7.1. A map ϕ : G × Λ → Λ is a category cocycle if for all g ∈ G, v ∈ Λ0,
and α, β ∈ Λ with s(α) = r(β) we have
(1) ϕ(g, v) = g,
(2) ϕ(g, α) · r(α) = g · r(α),
(3) g · (αβ) = (g · α)(ϕ(g, α) · β),
(4) ϕ(g, αβ) = ϕ(ϕ(g, α), β).
We call (Λ, G, ϕ) a category system.
Definition 7.2. Let (Λ, G, ϕ) be a category system. We will denote by Λ ⋊ϕ G the
small category with
Λ ⋊ϕ G = Λ × G
and
(Λ ⋊ϕ G)0 = Λ × {e} ,
and r, s : Λ ⋊ϕ G → (Λ ⋊ϕ G)0 defined by
r(α, g) = (r(α), 1G)
and
s(α, g) = (g−1 · s(α), 1G) .
Moreover for (α, g), (β, h) with s(α, g) = r(β, h) we have that
We will call Λ ⋊ϕ G the Zappa-Sz´ep product of (Λ, G, ϕ).
(α, g)(β, h) = (α(g · β), ϕ(g, β)h) .
It was proved that Λ ⋊ϕ G is left cancellative whenever Λ is left cancellative [2,
Proposition 3.5], and as observe in [2, Remnark 3.9] the elements of the form (v, g)
where v ∈ Λ0 and g ∈ G are units of Λ ⋊ϕ G. Then given (α, g) ∈ Λ ⋊ϕ G and h ∈ G we
have that
(α, g)(g−1 · s(α), g−1h) = (α, h) ,
so (α, g) ≈ (α, h). Moreover, Λ ⋊ϕ G is finitely aligned (singly aligned) whenever Λ is
finitely aligned (singly aligned) [2, Proposition 3.12]. In particular,
(α, g) ∨ (β, h) = (α ∨ β) × {1G} .
Definition 7.3. A category system (Λ, G, ϕ) is called pseudo free if, whenever g · α = α
and ϕ(g, α) = 1G, then g = 1G.
Proposition 7.4 ([10, Proposition 5.6]). Let (Λ, G, ϕ) be a pseudo free category system.
Then, for all g1, g2 ∈ G, and α ∈ Λ, one has that
g1 · α = g2 · α and ϕ(g1, α) = ϕ(g2, α)
⇒
g1 = g2 .
TIGHT GROUPOIDS OF LCSC
26
Remark 7.5. Given a (Λ, G, ϕ) where Λ is a right cancellative category, it may happen
that Λ ⋊ϕ G fails to satisfy right cancellation. Given (α, a), (β, b) and (γ, g) in Λ ⋊ϕ G we
have that (α, a)(γ, g) = (β, b)(γ, g) if and only if α(a · γ) = β(b · γ) and ϕ(a, γ) = ϕ(b, γ).
In particular, the system is pseudo free if and only if Λ ⋊ϕ G is right cancellative.
Remark 7.6. Let (Λ, G, ϕ), and let F = {(γ1, h1), . . . , (γn, hn)} ⊆ Λ ⋊ϕ G. Then given
(α, g) ∈ Λ ⋊ϕ G. we have that F ∈ FE(α, g) of Λ ⋊ϕ G if and only if {γ1, . . . , γn} ∈ FE(α)
of Λ.
By the above remark the following results are a direct translation of Theorem 6.4 and
6.6.
Proposition 7.7. Let (Λ, G, ϕ) be a category system. If either Gtight(SΛ⋊ϕG) is Hausdorff
or E∞ = Etight, then the following is equivalent:
(1) Gtight(SΛ⋊ϕG) is effective,
(2) Given (α, a), (β, b) ∈ Λ ⋊ϕ G with r(α, a) = r(β, b) and s(α, a) = s(β, b), if
(α, a)(δ, d) ⋓ (β, b)(δ, d) for every (δ, d) ∈ s((α, a))(Λ ⋊ϕ G) then there exists
F ∈ FE(s(α, a)) such that (α, a)(γ, d) = (β, b)(γ, d) for every (γ, d) ∈ F .
(3) Given α, β ∈ Λ, a, b ∈ G with r(α) = r(b) and a−1 · s(α) = b−1 · s(β), if
α(a · δ) ⋓ β(b · δ) for every δ ∈ (a−1 · s(α))Λ then there exists F ∈ FE(a−1 · s(α))
such that α(a · γ) = β(b · γ) and ϕ(a, γ) = ϕ(b, γ) for every γ ∈ F .
Proposition 7.8. If (Λ, G, ϕ) is a category system, then the following statements are
equivalent:
(1) Gtight(SΛ⋊ϕG) is minimal.
(2) For every (α, a), (β, b) ∈ Λ ⋊ϕ G there exists F ∈ FE((α, a)) such that for each
(γ, g) ∈ F , s(β, b)(Λ ⋊ϕ G)s(γ, g) 6= ∅.
(3) For every α, β ∈ Λ there exists F ∈ FE(α) such that for each γ ∈ F , there exist
g ∈ G with s(β)Λ(g · s(γ)) 6= ∅.
Then, by an analog argument to that of Theorem 6.7, we have the following result
Theorem 7.9. Let (Λ, G, ϕ) be a category system such that Λ and G are countable.
If either Gtight(SΛ⋊ϕG) is Hausdorff or E∞ = Etight, then the following statements are
equivalent:
(1) C ∗(SΛ⋊ϕG) is simple.
(2) For any field K, KSΛ⋊ϕG is simple.
(3) The following properties hold:
(a) Given α, β ∈ Λ, a, b ∈ G with r(α) = r(b) and a−1 · s(α) = b−1 · s(β), if
α(a·δ)⋓β(b·δ) for every δ ∈ (a−1·s(α))Λ then there exists F ∈ FE(a−1·s(α))
such that α(a · γ) = β(b · γ) and ϕ(a, γ) = ϕ(b, γ) for every γ ∈ F .
(b) For every α, β ∈ Λ there exists F ∈ FE(α) such that for each γ ∈ F , there
exist g ∈ G with s(β)Λ(g · s(γ)) 6= ∅.
To end this section, we will have a look on the case of Λ = E∗, where E is a countable
graph. When G is a countable discrete group and E is a countable graph, there is a
definition of self-similar graph extending that of [10] (see [11, Definition 2.2]). In fact,
as shown in [11, Theorem 3.2], the case of arbitrary graphs can be reduced to the case
of row-finite graphs with no sources or sinks up to Morita equivalence (of both algebras
TIGHT GROUPOIDS OF LCSC
27
and groupoids); in this case, most of the properties enjoyed by the system are analog to
these found in the finite case.
In order to fix the relation between Gtight(SG,E) and Gtight(SE ∗⋊ϕG), we first need to
state the relation between SG,E and SE ∗⋊ϕG. On one side, we have
SG,E = {(α, g, β) : α, β ∈ E∗, g ∈ G, s(α) = g · s(β)}.
On the other side,
SE ∗⋊ϕG = hτ (α,g)σ(β,h) : α, β ∈ E∗, g, h ∈ G, g−1 · s(α) = h−1 · s(β)i.
Since (x, g) ∈ (E∗ ⋊ϕ G)−1 for all x ∈ E0, g ∈ G, we have that
τ (α,g)σ(β,h) = τ (α,g)σ(h−1·s(β),h−1)τ (s(β),h)σ(β,1G) = τ (α,gh−1)σ(β,1G).
Moreover, since E∗ is singly aligned, then so is E∗ ⋊ϕ G by [2, Proposition 3.12(ii)].
Thus, by [5, Theorem 3.2],
SE ∗⋊ϕG = {τ (α,g)σ(β,1G) : α, β ∈ E∗, g ∈ G, s(α) = g · s(β)} .
Hence, the map
π :
SG,E → SE ∗⋊ϕG
(α, g, β)
7→ τ (α,g)σ(β,1G)
is a well-defined, onto ∗-semigroup homomorphism. Let us characterize when π is injec-
tive. To this end, take (α, g, β), (γ, h, δ) ∈ SG,E such that
τ (α,g)σ(β,1G) = π(α, g, β) = π(γ, h, δ) = τ (γ,h)σ(δ,1G).
Being both equal functions, they must have the same domain, i.e. (β, 1G)(E∗ ⋊ϕ G) =
(δ, 1G)(E∗ ⋊ϕ G). Since (E∗ ⋊ϕ G)−1 = E0 × G, we conclude that β = δ. Moreover,
τ (α,g) = τ (γ,h) on their common domain, so that for every λ ∈ (g−1 · s(α))E∗ and for
every ℓ ∈ G we have
(α(g · λ), ϕ(g, λ)ℓ) = τ (α,g)(λ, ℓ) = τ (γ,h)(λ, ℓ) = (γ(h · λ), ϕ(h, λ)ℓ).
Since the self-similar action of G on E∗ preserves lengths of paths, we conclude that
α = γ, and that for every λ ∈ (g−1 · s(α))E∗ we have g · λ = h · λ and ϕ(g, λ) = ϕ(h, λ).
Thus, the existence of nontrivial kernel for π is equivalent to the existence of g ∈ G,
α ∈ E∗ such that for all λ ∈ s(α)E∗ satisfies g · λ = λ and ϕ(g, λ) = 1G; in other
words, injectivity of π is equivalent to the fact that the self-similar action of G on E∗ is
faithful on vertex-based trees of E. Notice that if (E, G, ϕ) is pseudo free, then the above
condition is trivially fulfilled, so that π will be an isomorphism in this case. Moreover,
being E∗ ⋊ϕ G singly aligned, we have that it is right cancellative exactly when (E, G, ϕ)
∼= SE ∗⋊ϕG, but also they are weak semilattices
is pseudo free. In this case, not only SG,E
by Proposition 4.11, so that their associated tight groupoids are Hausdorff by Corollary
4.14.
Now, we proceed to look at the relation between the corresponding tight groupoids
Gtight(SG,E) and Gtight(SE ∗⋊ϕG). First, notice that the idempotent semilattices of SE,
SG,E and SE ∗⋊ϕG coincide, so that the spaces of filters, ultrafilters and tight filters are
the same (up to natural isomorphism). Additionally, the partial actions SG,E y E0 and
TIGHT GROUPOIDS OF LCSC
28
SE ∗⋊ϕG y E0 are π-equivariant; also, the germ relation is compatible with π. Thus, π
induces a continuous, open, onto groupoid homomorphism
Φ : Gtight(SG,E) → Gtight(SE ∗⋊ϕG)
7→ [τ (α,g)σ(β,1G); η].
[α, g, β; η]
Using the Morita equivalence reduction [11, Theorem 3.2], we can assume that E is
row-finite with no sources or sinks, whence E∞ = Etight = E∞; let us reduce to this case,
in order to simplify the computations. We now will show that Φ is injective. To this end,
let [α, g, β; η] ∈ ker Φ. This means that τ (α,g)σ(β,1G) is an idempotent. According to the
computations done before, this happens exactly when α = β and for every λ ∈ s(α)E∗
we have that g · λ = λ and ϕ(g, λ) = 1G. Pick λ any initial segment in η ∈ E∞. Notice
that λ ∈ s(α)E∗, and thus (αλ, 1G, αλ) ∈ αη (seen as a filter), while
(α, g, α) · (αλ, 1G, αλ) = (α(gλ), ϕ(g, λ), αλ) = (αλ, 1G, αλ).
Hence, by the germ relation, if η = λη, then
[α, g, α; αη)] = [αλ, 1G, αλ; αλη)] ∈ Gtight(SG,E)(0).
Thus, Φ is a homeomorphism and an isomorphism of groupoids. This guarantees that,
independently of the choice for representing the self-similar graph system (G, E, ϕ), their
associated tight groupoids -and hence their algebras- are the same.
8. Amenability
Now we are going to study a case where we can deduce amenability of Gtight(SΛ⋊ϕG)
assuming that Gtight(SΛ) and G are amenable. Let Λ be a finitely aligned LCSC, and let
Γ be a subsemigroup of a group Q.
Definition 8.1 ([13, Definition 6.1]). Let Γ be a semigroup with unit element 1Q. A
Γ-graph is a LCSC Λ together with a map, called the degree map, d : Λ → Γ, such that:
(1) d(αβ) = d(α)d(β) for every α, β ∈ Λ with s(α) = r(β),
(2) for every α ∈ Λ and γ1, γ2 ∈ Γ with d(α) = γ1γ2, there are unique α1, α2 ∈ Λ with
s(α1) = r(α2), d(αi) = γi for i = 1, 2, such that α = α1α2 (unique factorization
property).
Observe that if Λ is a Γ-graph, the unique factorization property implies that Λ is
right and left cancellative category, and does not have inverses.
Given two elements γ1, γ2 ∈ Γ
γ1 ≤ γ2
if and only if
γ−1
1 γ2 ∈ Γ .
Lemma 8.2. Let Λ be a Γ-graph. Let α, β ∈ Λ with α ⋓ β. Then α ≤ β if and only if
d(α) ≤ d(β). In particular α = β whenever d(α) = d(β).
Proof. Let α, β ∈ Λ, and let ε ∈ α ∨ β, so there are δ, η ∈ Λ such that ε = αδ = βη.
Assume that d(α) ≤ d(β). So by the unique factorization property there exists γ, γ′ ∈ Λ
such that γγ′ = β and d(γ) = d(α). But then
d(α)d(δ) = d(αδ) = d(γγ′η) = d(γ)d(γ′η) .
Thus, by the unique factorization property α = γ, and hence α ≤ β, as desired.
Finally, if d(α) = d(β) then α ≤ β and β ≤ α. But since Λ has no inverses, it follows
(cid:3)
that α = β.
TIGHT GROUPOIDS OF LCSC
29
Definition 8.3. Let (Λ, G, ϕ) be a category system, where Λ is a Γ-graph. We say that
Γ is compatible with respect to (Λ, G, ϕ), if d(g · α) = d(α) for every g ∈ G and α ∈ Λ
(G-invariant).
Let Λ be a Γ-graph compatible with respect to (Λ, G, ϕ). Observe that since τ (α,g)σ(α,g) =
τ (α,1G)σ(α,1G) for every (α, g) ∈ Λ ⋊ϕ G, the set of idempotents EΛ = EΛ⋊ϕG coincide,
and hence so does their spaces of tight filters. We will denote by Etight the space of tight
filters of EΛ and EΛ⋊ϕG. By Lemma 4.3,
Gtight(SΛ) = {[τ ασβ, ξ] : ξ ∈ Etight, α, β ∈ Λ, s(α) = s(β), β ∈ ∆ξ}
and
Gtight(SΛ⋊ϕG)
= {[τ (α,a)σ(β,b), ξ] : ξ ∈ Etight, α, β ∈ Λ, a, b ∈ G, a−1 · s(α) = b−1 · s(β), β ∈ ∆ξ} .
Then, we will think of Gtight(SΛ) as an open subgroupoid of Gtight(SΛ⋊ϕG) via the map
[τ ασβ, ξ] 7→ [τ (α,1G)σ(β,1G), ξ].
The following remark is going to be used repeatedly during the rest of the paper
sometimes without mention it.
Remark 8.4. Let (α, a), (β, b) ∈ Λ ⋊ϕ G, and suppose that (α, a) ≤ (β, b), then there
exists (δ, d) ∈ Λ ⋊ϕ G with r(δ, d) = s(α, a), that is, r(δ) = a−1 · s(α) such that
(β, b) = (α, a)(δ, d) = (α(a · δ), ϕ(a, δ)d) .
Whence α(a · δ) = β and b = ϕ(a, δ)d. Hence, a · δ = σα(β) by left cancellation, so
δ = a−1 · σα(β) and d = ϕ(a, a−1 · σα(β))−1b = ϕ(a−1, σα(β))b because of the cocycle
identity. Therefore, (α, a) ≤ (β, b) if and only if α ≤ β, and then we have that
σ(α,a)(β, b) = (a−1 · σα(β), ϕ(a−1, σα(β))b) .
Lemma 8.5. Let Λ be a Γ-graph compatible with respect to (Λ, G, ϕ), then the map
d : Gtight(SΛ⋊ϕG) → Q ,
[τ (α,a)σ(β,b), ξ] 7→ d(α)d(β)−1 ,
is a well defined continuous groupoid homomorphism.
Gtight(SΛ).
In particular, d restricts to
Proof. First we will prove that d is well defined. Let [s, ξ] = [t, ξ] in Gtight(SΛ⋊ϕG), that
is, there exists f ∈ ξ such that sf = tf . Without lost of generality we can assume
that s = τ (α,a)σ(β,b), t = τ (δ,d)σ(η,c), and f = τ (γ,1G)σ(γ,1G) with α, β, γ ∈ ∆ξ such that
α, β ≤ γ. Then, by Remark 8.4, we have that
and
sf = τ (α,a)σ(β,b) · τ (γ,1G)σ(γ,1G)
= τ (α,a)(b−1·σβ (γ),ϕ(b−1,σβ (γ)))σ(γ,1G)
= τ (α(ab−1·σβ (γ)),ϕ(ab−1 ,σβ (γ)))σ(γ,1G) ,
tf = τ (δ,d)σ(η,c) · τ (γ,1G)σ(γ,1G)
= τ (δ,d)(c−1·ση(γ),ϕ(c−1,ση(γ)))σ(γ,1G)
= τ (δ(dc−1·ση(γ)),ϕ(dc−1,ση(γ)))σ(γ,1G) ,
TIGHT GROUPOIDS OF LCSC
30
But sf = tf , so then
(1)
α(ab−1 · σβ(γ)) = δ(dc−1 · ση(γ))
and
ϕ(ab−1, σβ(γ)) = ϕ(dc−1, ση(γ)) .
Therefore, by the G-invariance of d we have that
d(α(ab−1 · σβ(γ)))d(γ)−1 = d(α)d(ab−1 · σβ(γ))d(γ)−1
= d(ασβ(γ))d(γ)−1
= d(α)d(β)−1d(γ)d(γ)−1
= d(α)d(β)−1 ,
is equal to
d(δ(dc−1 · ση(ε′)))d(γσγ(ε′)θ)−1 = d(δ)(.dc−1 · ση(γ)θ)d(γ)−1
= d(δ)d(ση(γ))d(γ)−1
= d(δ)d(η)−1d(γ)d(γ)−1
= d(δ)d(η)−1 .
Hence, d([τ ασβ, ξ]) = d(α)d(β)−1 = d(δ)d(η)−1 = d([τ δση, ξ]), so d is well-defined.
Now, let [s, ξ], [t, ξ′] ∈ Gtight(SΛ⋊ϕG) such that ξ = t · ξ′.
If s = τ (α,a)σ(β,b) and
t = τ (δ,d)σ(η,c), then [s, ξ] · [t, ξ′] = [st, ξ′]. Observe that, since β, δ ∈ ∆ξ, by Lemma 8.2
there exists only one element ε ∈ (β ∨ δ) ∩ ∆ξ. We define f = τ (ε,1G)σ(ε,1G). We have
that ξ = t · ξ′, that means
∆ξ = {γ ∈ Λ : γ ≤ t(ν), ν ∈ ∆ξ′} = {γ ∈ Λ : γ ≤ τ (δ,d)σ(η,c)(ην), ην ∈ ∆ξ′}
= {γ ∈ Λ : γ ≤ δ(dc−1 · ν), ην ∈ ∆ξ′} .
But ε ∈ ∆ξ, that is ε ≤ δ(dc−1 · ν) for some ην ∈ ∆ξ′, and hence
d(ε) ≤ d(δ(dc−1 · ν)) = d(δ)d(dc−1 · ν) = d(δ)d(ν) .
As d(δ) ≤ d(ε), we have d(δ) ≤ d(ε) ≤ d(δ)d(ν). Therefore, there exists g, h ∈ Γ such
that d(ε) = d(δ)g and gh = d(ν). Now, by the unique factorization, there exist unique
elements ν1, ν2 ∈ Λ such that ν1ν2 = ν and d(ν1) = g and d(ν2) = h. But ην1 ∈ ∆ξ′, so
t(ην1) = δ(dc−1 · ν1), and d(δ(dc−1 · ν1)) = d(δ)g = d(ε). Hence by Lemma 8.2 we have
that ε = δ(dc−1 · ν1).
Then f t = t · τ (ην1,1G)σ(ην1,1G), so [s, ξ] = [sf, ξ] and [t, ξ′] = [f t, ξ′]. So, we can assume
that β = δ, and hence d([s, ξ]) = d(α)d(β)−1 and d([t, ξ′]) = d(β)d(η)−1. Thus,
Therefore,
τ (α,a)σ(β,b) · τ (β,d)σ(η,c) = τ (α,a)(b−1·s(β),b−1d)σ(η,c)
= τ (α,ab−1d)σ(η,c) .
d([st, ξ]) = d(α)d(η)−1
= d(α)d(β)−1d(β)d(η)−1
= d([s, ξ])d([t, ξ′]) .
Thus, d is a morphism of groupoids.
TIGHT GROUPOIDS OF LCSC
Finally, given g ∈ Γ, we have that
(d)−1(g) =
[
α,β∈Λ, d(α)d(β)−1=g
that is an open. Thus, d is continuous.
Θ(τ (α,a)σ(β,b), Dτ β σβ ) ,
31
(cid:3)
Let d : Gtight(SΛ⋊ϕG) → Γ be the cocycle defined in Lemma 8.5, and let us define
HΛ⋊ϕG := (d)−1(1G) = {[τ (α,a)σ(β,b), ξ] ∈ Gtight(SΛ⋊ϕG) : d(α)d(β)−1 = 1G} .
It is an open subgroupoid of Gtight(SΛ⋊ϕG).
Now in order to be able to decompose the groupoid HΛ⋊ϕG as a union of more treatable
groupoids, we need to impose some conditions on the semigroup Γ.
Definition 8.6. Let Γ ⊆ Q be a subsemigroup of a group Q with Γ ∩ Γ−1 = 1Q. We
say that Γ is a join-semilattice if given g1, g2 ∈ Γ
exists an it is unique. We will denote it by g1 ∨ g2.
inf{g ∈ Γ : g1, g2 ≤ g}
We now assume that Γ is a join-semilattice. Then, given g ∈ Γ, we define
H(g)
Λ⋊ϕG := {[τ (α,a)σ(β,b), ξ] : d(α) = d(β) ≤ g} .
Λ⋊ϕG is an open subgroupoid of HΛ⋊ϕG. Let [τ (α,a)σ(β,b), ξ], [τ (δ,d)σ(η,c), ξ′] ∈
We claim that H(g)
H(g)
Λ⋊ϕG two composable elements with g1 := d(β) = d(α) ≤ g and g2 := d(δ) = d(η) ≤ g.
Since β, δ ∈ ∆ξ we have that there exists ε ∈ (β ∨ δ) ∩ ∆ξ, and g1, g2 ≤ d(ε). Since Γ is a
join-semilattice we have that g1 ∨ g2 ≤ d(ε). Then by the unique factorization property
there exists ε1, ε2 ∈ Λ with d(ε1) = g1 ∨ g2 and ε = ε1ε2. Then ε1 ∈ ∆ξ and hence by
Lemma 8.2 we have that β, δ ≤ ε1. Then as shown in the proof of Lemma 8.5 we can find
elements [τ (α′,a′)σ(ε,b′), ξ], [τ (ε,d′)σ(η′,c′), ξ′] ∈ HΛ⋊ϕG with [τ (α,a)σ(β,b), ξ] = [τ (α′,a′)σ(ε,b′), ξ]
and [τ (δ,d)σ(η,c), ξ′] = [τ (ε,d′)σ(η′,c′), ξ′], and the product
[τ (α′,a′)σ(ε,b′), ξ] · [τ (ε,d′)σ(η′,c′), ξ′] = [τ (α′,a′(b′)−1)σ(η′,c′), ξ′] ∈ H(g1∨g2)
Λ⋊ϕG .
Therefore, H(g1∨g2)
Λ⋊ϕG ⊆ H(g)
Λ⋊ϕG, as desired.
Moreover, as a consequence of the above computation, given g1, g2 ≤ g we have that
H(g1)
Λ⋊ϕGH(g2)
Λ⋊ϕG. Then, if Γ is countable, there exists an ascending sequence
of elements g1, g2, . . . ∈ Γ such that for every g ∈ Γ there exists n ∈ N with g ≤ gn.
Λ⋊ϕG ⊆ H(g)
Whence, HΛ⋊ϕG =S∞
i=1 H(gi)
Λ⋊ϕG.
The next step will be to define a cocycle of the groupoids H(g)
Λ⋊ϕG onto G. In order to
do that we will need to make the following assumption in the Γ-graph Λ.
Definition 8.7. Let Λ be a Γ-graph. Then Λ satisfies property (⋆) if given F ∈ ∆tight
and g ∈ Γ, then there exists a unique β ∈ F with d(β) ≤ g such that whenever α ∈ F
satisfies d(α) ≤ g, we have that α ≤ β.
We can give some condition on Γ to guarantee that every Γ-graph satisfies condition
(⋆).
Proposition 8.8. Let Λ be a Γ-graph, and assume every bounded ascending sequence of
elements of Γ stabilizes. Then Λ satisfies property (⋆).
TIGHT GROUPOIDS OF LCSC
32
Proof. We define Fg := {β ∈ F : d(β) ≤ g}. Observe that given α, β ∈ Fg with
d(β) ≤ d(α), then d(α ∨ β) = d(α) ∨ d(β) = d(α) ≤ g, hence α ∨ β ∈ Fg, and the
unique factorization property says that α = α ∨ β, and hence β ≤ α. So it is enough to
prove that there exists α ∈ Fg such that d(β) ≤ d(α) for every β ∈ Fg.
Let α0 ∈ F , and let β ∈ Fd with d(β) (cid:2) d(α0). If such β does not exists, then we
are done. Otherwise, d(α0 ∨ β) = d(α0) ∨ d(β) ≤ g, and hence α0 ∨ β ∈ Fg, with
d(α0) < d(α0 ∨ β), because if d(α0) = d(α0 ∨ β) then α0 = α0 ∨ β by Lemma 8.2. Let
us define α1 := α0 ∨ β, so d(α0) < d(α1). Now if in this way we could construct an
infinite sequence α0, α1, α2, . . . ∈ Fg such that (.αi) < d(αi+1), then this will contradict
the hypothesis. Then will be n such that d(β) ≤ d(αn) for every β ∈ Fg, and so α := αn,
and we are done.
(cid:3)
Example 8.9. Every Nk-graph Λ satisfies property (⋆).
Now we are ready to define the promised cocyle.
Proposition 8.10. Let Λ be a Γ-graph compatible with respect to a pseudo free system
(Λ, G, ϕ), and suppose that Λ satisfies property (⋆). Then for every g ∈ Γ there exists
a continuous groupoid homomorphism
t(g) : H(g)
Λ⋊ϕG → G .
Proof. Let [s, ξ] ∈ H(g)
Λ⋊ϕG. By property (⋆) there exists β ∈ ∆ξ such that δ ≤ β for
every δ ∈ ∆ξ with d(δ) ≤ g. If we define f = τ (β,e)σ(β,e), then we have that [s, ξ] = [sf, ξ]
whenever s = τ (α,a)σ(δ,b) with d(δ) ≤ g, and
sf = τ (α,a)σ(δ,b)τ (β,e)σ(β,e) = τ (α,a)(b−1·σδ (β)),ϕ(b−1,σδ(β)))σ(β,e)
= τ (α(ab−1·σδ (β))),ϕ(ab−1 ,σδ(β)))σ(β,e) .
Thus, without lost of generality, any element [s, ξ] ∈ H(g)
Λ⋊ϕG has a representative of the
form [τ (α,a)σ(β,b), ξ], where β is the unique maximal element in ∆ξ satisfying d(β) ≤ g
given by property (⋆).
Under this choice of representative, we define t(g) : H(g)
t(g)([τ (α,a)σ(β,b), ξ]) = ab−1.
Λ⋊ϕG → G by the rule
Let us check that t(g) is well defined. To this end, let [s, ξ] and [s′, ξ] in H(g)
Λ⋊ϕG
with [s, ξ] = [s′, ξ]. By the above argument, we can assume that s = τ (α,a)σ(β,b) and
s′ = τ (α′,a′)σ(β,b′). Let h = τ (ββ′,1G)σ(ββ′,1G) for some β′ such that ββ′ ∈ ∆ξ and sh = s′h.
Then,
sh = τ (α(ab−1·β′),ϕ(ab−1,β′))σ(ββ′,1G) and s′h = τ (α(a′ b′−1·β′),ϕ(a′b′−1,β′))σ(ββ′,1G) .
Therefore we have that
α(ab−1 · β′) = α(a′b′−1 · β′)
and by left cancellation we have that
and
ϕ(ab−1, β′) = ϕ(a′b′−1, β′) ,
ab−1 · β′ = a′b′−1 · β′
and
ϕ(ab−1, β′) = ϕ(a′b′−1, β′) .
Hence, by Proposition 7.4, ab−1 = a′b′−1. Thus, t(g) is well-defined.
TIGHT GROUPOIDS OF LCSC
33
Now, given a composable pair [s, ξ], [t, ξ′], we can choose representatives s = [τ (α,a)σ(β,b), ξ]
and t = [τ (γ,c)σ(β′,b′), ξ′] with β, β′ unique maximal elements in ∆ξ, ∆ξ′ (respectively) sat-
isfying d(β), d(β′) ≤ g given by property (⋆). Since ξ = t · ξ′, we have that γ ∈ ∆ξ,
whence γ = β ∨ γ by property (⋆). Thus, the computation performed in Lemma 8.5
do not require replace β by any element δ with d(δ) > g, and thus this argument shows
that t(g) is a continuous groupoid homomorphism.
(cid:3)
Proposition 8.11. Let Λ be a Γ-graph compatible with respect to a pseudo free system
(Λ, G, ϕ), and suppose that Λ satisfies property (⋆), with G and Q countable amenable
groups. Moreover, assume that Γ is a join-semilattice. If the kernel of the map d :
Gtight(SΛ) → Q is amenable, then Gtight(SΛ⋊ϕG) is amenable.
Proof. By [13, Corollary 4.5] it is enough to prove that HΛ⋊ϕG is an amenable groupoid.
Λ⋊ϕG where H(gn)
Λ⋊ϕG are open subgroupoids with
Λ⋊ϕG )(0). Then by [1, Section 5.2(c)] it is enough
As observe above HΛ⋊ϕG = S∞
Λ⋊ϕG)(0) = (H(gn+1)
n=1 H(gn)
Λ⋊ϕG ⊆ H(gn+1)
H(gn)
to prove that the groupoids H(gn)
Λ⋊ϕG and (H(gn)
Λ⋊ϕG are amenable for every n. But now
(t(gn))−1(1G) = {[τ (α,a)σ(β,a), ξ] ∈ H(gn)
Λ⋊ϕG}
= {[τ (α,1G)σ(β,1G), ξ] ∈ H(gn)
= {[τ (α,1G)σ(β,1G), ξ] ∈ GΛ⋊ϕG : d(α) = d(β) ≤ gn}
= {[τ ασβ, ξ] ∈ GΛ⋊ϕG : d(α) = d(β) ≤ gn} ⊆ (d)−1(1Q) .
Λ⋊ϕG}
Therefore since (d)−1(1Q) is amenable by assumption, then (t(gn))−1(1G) is amenable.
So using again [13, Corollary 4.5] we have that H(gn)
(cid:3)
Λ⋊ϕG is amenable, as desired.
Next step will prove to prove that the kernel of the map d : Gtight(SΛ) → Q is
amenable. In order to do that we will prove that the groupoid Gtight(SΛ) is isomorphic
to the semigroup action groupoid of the Γ-graph Λ defined in [13, Section 5]. This
semigroup action groupoid has also a canonical cocyle c onto Q which kernel is amenable.
Now we will prove that the kernel of c is isomorphic to the kernel of d. First we introduce
the semigroup action groupoid.
Definition 8.12. Let X be a set and Γ ⊆ Q be a semigroup of a group Q containing
the identity 1Q. A left action of Γ on X consists of a subset Γ ⋆ X of Γ × X and a map
T : Γ ⋆ X → X sending (g, x) 7→ g · x, such that:
(1) for all x ∈ X, (e, x) ∈ Γ ⋆ X and e · x = x;
(2) for all (g, h, x) ∈ Γ × Γ × X, (gh, x) ∈ Γ ⋆ X if and only if (h, x) ∈ Γ ⋆ X and
(g, h · x) ∈ Γ ⋆ X, if this holds, g · (h · x) = (gh) · x.
For all g ∈ Γ, we define U(g) := {x : (g, x) ∈ Γ ⋆ X} and V (g) = {g · x : (g, x) ∈ Γ ⋆ X}
and Tg : U(g) → V (g) the map such that Tg(x) = g · x. The triple (X, Γ, T ) is called a
semigroup action.
Definition 8.13. A semigroup action (X, Γ, T ) is called directed if for all g, h ∈ Γ such
that U(g) ∩ U(h) 6= ∅ there exists r ∈ Γ with g, h ≤ r such that U(g) ∩ U(h) = U(r).
TIGHT GROUPOIDS OF LCSC
34
When (X, Γ, T ) is a directed semigroup action it is defined the groupoid
G(X, Γ, T ) = {(x, gh−1, y) ∈ X × Γ × X : (g, x), (h, y) ∈ Γ ⋆ X, g · x = h · y} .
Let X be a locally compact Hausdorff space such that U(g) and V (g) are open subsets
of X for every g ∈ Γ, and Tg : U(g) → V (g) is a local homeomorphism.
Given g, h ∈ Γ, A, B subsets of X, we define
Z(A, g, h, B) := {(x, gh−1, y) ∈ G(X, Γ, T ) : x ∈ A, y ∈ B and g · x = h · y} .
The family B of subsets Z(A, g, h, B), with A ⊆ U(g) and B ⊆ U(h) open subsets, such
that (Tg)A and (Th)B are injective, and Tg(A) = Th(B), forms a basis for the topology
of G(X, Γ, T ). With this topology G(X, Γ, T ) is a locally compact ´etale groupoid.
Now recall by Lemma 8.2, that given g ∈ Γ and F ∈ ∆∗, F ∩ d−1(g) is either empty
or contains just one element. Then we can define the following.
Definition 8.14. Let Λ be a Γ-graph. Then
Γ ⋆ Etight = {(g, ξ) ∈ Γ × Etight : ∃α ∈ ∆ξ such that d(α) = g} ,
and given (g, ξ) ∈ Γ ⋆ Etight, we define Tg(ξ) = σα · ξ, where α ∈ ∆ξ with d(α) = g.
U(g) := [α∈d−1(g)
Dτ ασα
and
V (g) = [α∈d−1(g)
[β∈Λs(α)
Dτ β σβ .
[
Let Λ be a Γ-graph with Γ a join-semilattice, because of the factorization property we
have that d(α ∨ β) = d(α) ∨ d(β), and hence
U(g) ∩ U(h) =
Dτ εσε = U(g ∨ h) .
α∈d−1(g), β∈d−1(h), ε∈α∨β
Thus, the semigroup action ( Etight, Γ, T ) is directed.
Proposition 8.15. Let Λ be a Γ-graph with Γ being join-semilattice. Then the map
Φ : Gtight(SΛ) → G( Etight, Γ, T ) ,
[τ ασβ, ξ] 7→ (τ ασβ · ξ, d(α)d(β)−1, ξ)
is an isomorphism of topological groupoids.
Proof. First observe that Φ is well-defined because of Lemma 8.5, and it is then clearly a
groupoid homomorphism. That Φ is a bijection follows from the definition of the inverse
Φ−1 : G( Etight, Γ, T ) → Φ : Gtight(SΛ) by Φ−1(ξ, gh−1, ξ′) = [τ ασβ, ξ], where β is the
unique element in ∆ξ with d(β) = h and α is the unique element in ∆ξ′ with d(α) = g
given by Lemma 8.2.
Now observe that the sets of the form Z(τ ασβ(Dτ β σβ ), g, h, Dτ βσβ ) for g, h ∈ Γ, β ∈
d−1(h) and α ∈ d−1(g) ∩ Λs(β) forms a basis for the topology of G( Etight, Γ, T ). But
Φ−1(Z(τ ασβ(Dτ β σβ ), g, h, Dτ βσβ )) = [τ ασβ, Dτ β σβ ], thus Φ is continuous and also open.
(cid:3)
Given a directed semigroup action (X, Γ, T ), there exists a natural groupoid homo-
morphism c : G( Etight, Γ, T ) → Γ defined by c(x, gh−1, yu) = gh−1 [13, Proposition 5.12],
and it is clear that Φ intertwines c and d, that is, c ◦ Φ = d. Therefore c−1(1Q) and
(1Q) are isomorphic as topological groupoids. But in the proof of [13, Theorem 5.13]
d
it is proved that c−1(1Q) is amenable, whence d
(1Q) is amenable too.
−1
−1
TIGHT GROUPOIDS OF LCSC
35
Theorem 8.16. Let Λ be a Γ-graph compatible with respect to a pseudo free system
(Λ, G, ϕ), and suppose that Λ satisfies property (⋆), with G and Q countable amenable
groups. Moreover, assume that Γ is a join-semilattice. Then Gtight(SΛ⋊ϕG) is amenable.
Acknowledgments
Part of this work was done during visits of the second author to the Institutt for
Matematiske Fag, Norges Teknisk-Naturvitenskapelige Universitet (Trondheim, Nor-
way), and during an stage of both authors to the ICMAT (Universidad Aut´onoma de
Madrid, Spain) as part of the Thematic Research Program: Operator Algebras, Groups
and Applications to Quantum Information in 2019, and the work was significantly sup-
ported by the research environment and facilities provided there. Both authors thank
the centers for their kind hospitality.
References
[1] C. Anantharaman-Delaroche, J. Renault, Amenable groupoids., Monographies de L'Enseignement
Math´ematique [Monographs of L'Enseignement Math´ematique], 36. L'Enseignement Math´ematique,
Geneva, 2000. 196 pp. ISBN: 2-940264-01-5
[2] E. B´edos, S. Kaliszewski, J. Quigg and J. Spielberg, On finitely aligned left cancellative small cate-
gories, Zappa-Sz´ep products and Exel-Pardo algebras, Theor. Appl. Categ. 33 (2018), Paper no. 42,
1346 -- 1406.
[3] J.H. Brown, L.O. Clark, C. Farthing and A. Sims, Simplicity of algebras associated to ´etale groupoids,
Semigroup Forum 88 (2014), 433 -- 452.
[4] N. Brownlowe, D. Pask, J. Ramagge, D. Robertson, M.F. Whittaker, Zappa-Sz´ep product groupoids
and C ∗-blends, Semigroup Forum 94 (2017), 500 -- 519.
[5] A.P. Donsig, J. Gensler, H. King, D. Milan and R. Wdowinski, On zigzag maps and the path category
of an inverse semigroup. arXiv:1811.04124v1.
[6] A.P. Donsig and D. Milan, Joins and covers in inverse semigroups and tight C ∗-algebras, Bull. Aust.
Math. Soc. 90 (2014), no. 1, 121 -- 133.
[7] R. Exel, Inverse semigroup and combinatorial C ∗-algebras, Bull. Braz. Math. Soc. 39(2) (2008),
191 -- 313.
[8] R. Exel, Tight and cover-to-joint representations of semilattices and inverse semigroups,
arXiv:1903.02911v1.
[9] R. Exel and E. Pardo, The tight groupoid of an inverse semigroup, Semigroup Forum 92 (2016),
274 -- 303.
[10] R. Exel and E. Pardo, Self-similar graphs, a unified treatment of Katsura and Nekrashevych C ∗-
algebras, Adv. Math. 306 (2017), 1046 -- 1129.
[11] R. Exel, E. Pardo and C. Starling, C ∗-algebras of self-similar graphs over arbitrary graphs,
arXiv:1807.01686v1.
[12] M.V. Lawson, Inverse semigroups. The theory of partial symmetries. World Scientific Publishing
Co., Inc., River Edge, NJ, 1998.
[13] J. Renault, D. Williams, Amenability of groupoids arising from partial semigroup actions and
topological higher rank graphs, Trans. Amer. Math. Soc. 369 (2017), no. 4, 2255 -- 2283.
[14] J. Spielberg, C ∗-algebras for categories of paths associated to the Baumslag-Solitar groups. J. Lon-
don Math. Soc. (2) 86 (2012), 728 -- 754.
[15] J. Spielberg, Groupoids and C ∗-algebras for categories of paths. Trans Amer. Math. Soc. 366 (2014),
no. 11, 5771 -- 5819.
[16] J. Spielberg, Groupoids and C ∗-algebras for left cancellative small categories. Indiana Univ. Math.
J. (to appear), arXiv:1712.07720v1.
[17] B. Steinberg, A groupoid approach to discrete inverse semigroup algebras. Adv. Math. 223 (2010),
689 -- 727.
TIGHT GROUPOIDS OF LCSC
36
[18] B. Steinberg, Simplicity, primitivity and semiprimitivity of ´etale groupoid algebras with applications
to inverse semigroup algebras.. J. Pure Appl. Algebra 220 (2016), no. 3, 1035 -- 1054.
Department of Mathematical Sciences, NTNU, NO-7491 Trondheim, Norway
E-mail address: [email protected]
Departamento de Matem´aticas, Facultad de Ciencias, Universidad de C´adiz, Campus
de Puerto Real, 11510 Puerto Real (C´adiz), Spain.
E-mail address: [email protected]
URL: https://sites.google.com/a/gm.uca.es/enrique-pardo-s-home-page/
|
1202.2103 | 1 | 1202 | 2012-02-09T20:30:03 | Analytic Free Semigroup Algebras and Hopf Algebras | [
"math.OA"
] | Let $\fS$ be an analytic free semigroup algebra. In this paper, we explore richer structures of $\fS$ and its predual $\fS_*$. We prove that $\fS$ and $\fS_*$ both are Hopf algebras. Moreover, the structures of $\fS$ and $\fS_*$ are closely connected with each other: There is a bijection between the set of completely bounded representations of $\fS_*$ and the set of corepresentations of $\fS$ on one hand, and $\fS$ can be recovered from the coefficient operators of completely bounded representations of $\fS_*$ on the other hand. As an amusing application of our results, the (Gelfand) spectrum of $\fS_*$ is identified. Surprisingly, the main results of this paper seem new even in the classical case. | math.OA | math |
ANALYTIC FREE SEMIGROUP ALGEBRAS
AND HOPF ALGEBRAS
DILIAN YANG
Abstract. Let S be an analytic free semigroup algebra. In this
paper, we explore richer structures of S and its predual S∗. We
prove that S and S∗ both are Hopf algebras. Moreover, the struc-
tures of S and S∗ are closely connected with each other: There is
a bijection between the set of completely bounded representations
of S∗ and the set of corepresentations of S on one hand, and S can
be recovered from the coefficient operators of completely bounded
representations of S∗ on the other hand. As an amusing appli-
cation of our results, the (Gelfand) spectrum of S∗ is identified.
Surprisingly, the main results of this paper seem new even in the
classical case.
1. Introduction
Let Si (i = 1, ..., n) be operators acting on a Hilbert space H. The
n-tuple S = [S1 · · · Sn] is said to be isometric if it is an isometry from
Hn to H, equivalently,
S∗
i Sj = δi,jI
(i, j = 1, ..., n).
metric n-tuple S is isomorphic to the Cuntz algebra On if Pi SiS∗
and isomorphic to the Cuntz-Toeplitz algebra En if Pi SiS∗
On one hand, it is well known that the C*-algebra generated by an iso-
i = I,
i < I. On
the other hand, the unital norm closed (non-selfadjoint) operator alge-
bra generated by S is completely isometrically isomorphic to Popescu's
non-commutative disk algebra An (cf. [27]). Because of those rigidi-
ties, Davidson and Pitts initiated the study of free semigroup algebras
in [13] in order to study the fine structure of isometric n-tuples. Since
then, free semigroup algebras have attracted a great deal of attention.
See, for example, [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 20, 21]. It should
be mentioned that the structure of isometric tuples has been completely
2000 Mathematics Subject Classification. 47L80, 16Txx.
Key words and phrases. free semigroup algebra, Hopf algebra, completely
bounded representation, corepresentation.
The author was partially supported by an NSERC Discovery grant.
1
2
D. YANG
described by Kennedy in [21] very recently. It turns out that an iso-
metric tuple has a higher-dimensional Lebesgue-von Neumann-Wold
decomposition.
A free semigroup algebra S is the unital wot-closed (non-selfadjoint)
operator algebra generated by an isometric tuple. The prototypical
example of free semigroup algebras is the non-commutative analytic
Toeplitz algebra Ln generated by the left regular representation of the
free semigroup F+
n with n generators. The algebra Ln was introduced in
[26] by Popescu in connection with a non-commutative von Neumann
inequality, and plays a prominent role in the theory of free semigroup
algebras. It is singled out to name a class of free semigroup algebras
analytic: A free semigroup algebra S generated by S is said to be
analytic if it is algebraically isomorphic to Ln. Analyticity plays a
fundamental role in the existence of wandering vectors. Recall that a
wandering vector for an isometric tuple S acting on H is a unit vec-
tor ξ in H such that the set {Swξ : w ∈ F+
n } is orthonormal. Here
Sw = Si1 · · · Sik for a word w = i1 · · · ik in F+
n . It was conjectured in
[8] that every analytic free semigroup algebra has wandering vectors.
This conjecture has been settled in [20] recently.
It is shown there
that a free semigroup algebra either has a wandering vector, or is a von
Neumann algebra.
The non-commutative disk algebra An is the unital norm closed al-
gebra generated by the left regular representation of F+
n . It was also
introduced by Popescu in [26]. He showed in [27] that the unital norm
closed algebra An generated by an arbitrary isometric n-tuple is com-
pletely isometrically isomorphic to An. The algebra An also plays an
important role in studying free semigroup algebras. The free semigroup
algebra S generated by an isometric n-tuple S is said to be absolutely
continuous if the representation of An induced by S can be extended to
a weak*-continuous representation of Ln. This notion was introduced
in [10] in order to obtain a natural analogue of the measure-theoretic
notion of absolute continuity, and played a central role there.
From the definitions and [8, Theorem 1.1] (cf. Theorem 2.2 below),
one sees that analyticity implies absolute continuity. The converse fails
in the classical case (i.e., when n = 1). Refer to, for example, [21, 29].
However it holds true in higher-dimensional cases [21].
Let S be an analytic free semigroup algebra generated by an isomet-
ric n-tuple. The main purpose of this paper is to explore richer struc-
tures of S and its predual S∗ (recall from [14] that S has a unique
predual), so that one could study those objects from other points of
view. This could lead us to understand them better and deeper. More
precisely, we show that both S and S∗ are Hopf algebras (Theorems
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
3
4.9 and 5.1). Moreover, the structures of S and S∗ are closely con-
nected with each other. We prove that there is a bijection between the
set of completely bounded representations of S∗ and the set of corep-
resentations of S (Theorem 6.3), and that the following duality result
holds true: S can be recovered from the coefficient operators of com-
pletely bounded representations of S∗ (Theorem 6.7). Furthermore, as
an amusing application of Theorem 6.3, we show that the (Gelfand)
spectrum of S∗ is precisely F+
n (Theorem 6.6). Rather surprisingly, ac-
cording to the author's best knowledge, the main results of this paper
seem new even in the classical case, that is, when S is the analytic
Toeplitz algebra.
It should be pointed out that this work connects with harmonic
analysis. It is shown in [13, 14] that Ln has a unique predual
Ln∗ = {[ξη∗] : ξ, η ∈ ℓ2(F+
n )},
where [ξη∗] is the rank one linear functional given by [ξη∗](A) = (Aξ, η)
for all A ∈ Ln. If we regard Ln as a non-selfadjoint analogue of the
group von Neumann algebra vN(Fn) of the free group Fn with n gener-
ators, then Ln∗ is analogous to the Fourier algebra A(Fn), the predual
of vN(Fn). See, for example, [19] for more information about Fourier
algebras. So one could think of Ln∗ as a sort of "Fourier algebra of the
free semigroup F+
n ". To seek an algebra structure for Ln∗ is actually
the starting motivation of this work. This point of view, in my opinion,
could be fruitful in further studying free semigroup algebras. In this
paper, we shall see certain analogy between Ln∗ and A(Fn).
The rest of the paper is organized as follows.
In Section 2, some
preliminaries on free semigroup algebras, dual algebras and operator
theory are given. Since the non-commutative analytic algebras play a
very crucial role in our paper, we take a closer look at them and prove
two main results in Section 3. Firstly, we prove that (Lm⊗Ln)∗ and
· · · L⊗k
n ] induced from the left regular representation of F+
Lm∗b⊗Ln∗ are completely isomorphically isomorphic. This isomorphism
will be used to prove that Ln∗ is a completely contractive Banach al-
gebra. Secondly, it is shown that for any k ≥ 1, the isometric n-tuple
[L⊗k
n is ana-
1
lytic (actually, more can be said). We will apply this result to construct
a Hopf algebra structure for an arbitrary analytic free semigroup S.
In Section 4, we show that every analytic free semigroup algebra S is
a Hopf dual algebra, while in Section 5, we prove that its predual S∗
is a Hopf convolution dual algebra. Those structures between S and
S∗ are tied by completely bounded representations of S∗. In Section
6, we show that, on one hand, there is a one-to-one correspondence be-
tween the set of completely bounded representations of S∗ and the set
4
D. YANG
of corepresentations of S; and on the other hand, S can be recovered
from the coefficient operators of completely bounded representations
of S∗. Moreover, it is shown that the (Gelfand) spectrum of S∗ is F+
n
if S is generated by an analytic isometric n-tuple.
2. Preliminaries
In this section, we will give some background which will be used
later. Also we take this opportunity to fix our notation.
Free semigroup algebras. The material in this subsection is mainly
from [6] and the references therein.
A free semigroup algebra S is the unital wot-closed operator algebra
generated by an isometric tuple. Let us start with a distinguished
representative of the class of those algebras -- the non-commutative
analytic Toeplitz algebra Ln. Let F+
n be the unital free semigroup
with the unit ∅ (the empty word), which is generated by the non-
commutative symbols 1, ..., n. We form the Fock space Hn := ℓ2(F+
n )
with the orthonormal basis {ξw : w ∈ F+
n }. Consider the left regular
representation L of F+
n : For i = 1, ..., n,
Li(ξw) = ξiw
(w ∈ F+
n ).
It is easy to see that L = [L1 · · · Ln] is isometric. Then the free
semigroup algebra generated by L is denoted as Ln, which is known
as the non-commutative analytic Toeplitz algebra. The case of n = 1
yields the wot-closed algebra generated by the unilateral shift, and so
reproduces the (classical) analytic Toeplitz algebra.
Similarly, one can define the right regular representation R of F+
n :
For i = 1, ..., n,
Ri(ξw) = ξwi
(w ∈ F+
n ).
Then by Rn we mean the free semigroup algebra generated by the
isometric n-tuple R = [R1 · · · Rn]. It is easy to see that Ln and Rn
are unitarily equivalent. Moreover, it turns out that Ln and Rn are
the commutants of each other: Rn = L′
n and Ln = R′
n.
As shown in [11], every element A ∈ Ln is uniquely determined by
its "Fourier expansion"
A ∼ X
w∈F+
n
awLw
in the sense that
Aξw = X
w∈F+
n
awξw
(w ∈ F+
n ),
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
5
where the "w-th Fourier coefficient" is given by aw = (Aξ∅, ξw). It is
often useful to heuristically view A as its Fourier expansion. For k ≥ 1,
the k-th Ces´aro sum of the Fourier series of A is defined by
Σk(A) = X
(1 −
w<k
w
k
)awLw,
where w is the length of the word w in F+
converges to A in the strong operator topology.
n . The sequence {Σk(A)}k
Another important operator algebra naturally associated with the
left regular representation L of F+
n is the non-commutative disk algebra
An, which, by definition, is the unital norm closed algebra generated
by L. Clearly, An is weak*-dense in Ln. The case of n = 1 yields the
(classical) disk algebra. The algebra An first appeared in [25], which is
related to multivariable non-commutative dilation theory. It was shown
in [27] that the unital norm closed operator algebra generated by an
arbitrary isometric n-tuple is completely isometrically isomorphic to
An.
Two classes of free semigroup algebras are particularly important
and they are defined in terms of Ln and An:
Definition 2.1. ([8, 10, 21]) Let S be a free semigroup algebra gen-
erated by an isometric n-tuple S.
(i) S is said to be analytic if it is algebraically isomorphic to Ln.
(ii) S is said to be absolutely continuous if the representation
of An induced by S can be extended to a weak*-continuous
representation of Ln.
Sometimes, we also call the isometric tuple S analytic in (i) and
absolute continuous in (ii), respectively.
In [8], analytic free semigroup algebras are called type L. The no-
tion used here follows [21]. Analyticity plays a fundamental role in
the existence of wandering vectors. In [8] it was conjectured that ev-
ery analytic free semigroup algebra has wandering vectors. Recently,
this conjecture has been settled in [20]. It is shown there that a free
semigroup algebra either has a wandering vector, or is a von Neumann
algebra. Motivated by the classical case, the notion of absolute con-
tinuity was introduced by Davidson-Li-Pitts in [10], and provided a
major device there. By Theorem 2.2 below and the definitions, analyt-
icity implies absolute continuity. It turns out that these two notions
coincide in higher-dimensional cases [21]. However, this is not true in
the classical case (cf. [21, 29]).
6
D. YANG
In what follows, we record some results of analytic free semigroup
algebras for later reference. Recall that a weak*-closed operator al-
gebra A on H has property A1(1), if given a weak*-continuous linear
functional τ on S and ǫ > 0, there are vectors x, y ∈ H such that
kxkkyk ≤ kτ k + ǫ and τ = [xy∗]. In particular, the weak* and weak
topologies on such an algebra coincide.
Theorem 2.2. ([8]) If S is an analytic free semigroup algebra gener-
ated by an isometric n-tuple S = [S1 · · · Sn], then there is a canonical
weak*-homeomorphic and completely isometric isomorphism φ from S
onto Ln, which maps Si to Li for i = 1, ..., n.
Theorem 2.3. ([3] for n = 1 and [21] for n ≥ 2) Any analytic free
semigroup algebra S has property A1(1).
Dual algebras and operator spaces. The main sources of this sub-
section are [16, 17, 28].
In this paper, a dual algebra is a unital weak*-closed subalgebra
of B(H) for some Hilbert space H. Usually, this is called a (unital)
concrete dual algebra. There is a characterization for abstract dual al-
gebras, and it turns out that every abstract dual algebra has a concrete
realization. We will not need those facts in this paper. Observe that
free semigroup algebras are dual algebras.
Let A ⊆ B(H) be a dual algebra. Then A has the standard predual
B(H)∗/A⊥, where, as usual, B(H)∗ is the space of all bounded weak*-
continuous linear functionals on B(H), and A⊥ is the preannihilator of
A in B(H)∗. Of course, A may have more than one predual. We will
use the notation AH∗, or simply A∗ if the context is clear, to denote
its standard predual. This will not cause any confusion in the context
of free semigroup algebras because of the uniqueness of their preduals
(cf. [14]).
It is very useful to notice that dual algebras are intimately connected
with the theory of operator spaces. Let V be an operator space. Then
its dual V ∗ is also an operator space, which is called the operator space
dual of V . An operator space W is said to be a dual operator space if W
is completely isometrically isomorphic to V ∗ for some operator space
V . In this case, we also say that V is an operator predual of W . The
normal spatial tensor product of dual operator spaces (resp. algebras)
is again a dual operator space (resp. algebra). Any dual algebra A
is a dual operator space. Furthermore, A∗ inherits a natural operator
space structure from A∗, and so A∗ itself is an operator space. In fact,
this paper heavily relies on the operator space structure of A∗.
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
7
Given dual operator spaces V ∗ and W ∗ in B(H) and B(K), respec-
tively, the normal Fubini tensor product V ∗⊗F W ∗ is defined by
V ∗⊗F W ∗ = {A ∈ B(H ⊗ K) : (ω1 ⊗ id)(A) ∈ W ∗, (id ⊗ ω2)(A) ∈ V ∗
for all ω1 ∈ B(H)∗, ω2 ∈ B(K)∗},
where ω1 ⊗ id and id ⊗ ω2 are the right and left slice mappings deter-
mined by ω1 and ω2, respectively. Let V ∗⊗W ∗ and V b⊗W stand for the
normal spatial tensor product of V ∗ and W ∗, and the operator projec-
tive tensor product of V and W , respectively. Since ω1 ⊗ id and id ⊗ ω2
are weak*-continuous, we have V ∗⊗W ∗ ⊆ V ∗⊗F W ∗. The following
result characterizes when they are equal.
Theorem 2.4. ([17, 28]) (i) We have the following weak*-homeomorphic
completely isometric isomorphism:
(ii) V ∗⊗F W ∗ = V ∗⊗W ∗ if and only if (V ∗⊗W ∗)∗
V ∗⊗F W ∗ ∼= (V b⊗W )∗.
∼= V b⊗W .
Notation. For a Hilbert space H, we use Hk and H⊗k to denote the
direct sum of k copies of H and the tensor product of k copies of H,
respectively. For A ∈ B(H), by A(k) ∈ B(Hk) and A⊗k ∈ B(H⊗k),
we mean the direct sum of k copies of A acting on Hk and the tensor
product of k copies of A acting on H⊗k, respectively.
Let S = [S1 · · · Sn] be an arbitrary isometric n-tuple acting on H.
We think of S either as n isometries acting on the common Hilbert
space H or as an isometry from Hn to H. Also, for k ≥ 1, we set
S⊗k = [S⊗k
· · · S⊗k
1
n ].
Let A and B be two dual algebras. We use A⊗B to denote the
normal spatial tensor product of A and B.
3. Ln revisited
In this section, we will take closer look at the non-commutative an-
alytic Toeplitz algebras, and prove some results which are vital in con-
structing Hopf algebra structures of an analytic free semigroup algebra
and its predual.
By Lm⊗wLn, we mean the wot-closed algebra generated by spa-
tial tensor product of Lm and Ln. The notation Rm⊗wRn is defined
It is easy to see that Rm⊗wRn ⊆ (Lm⊗wLn)′. Then, by
similarly.
[2, Theorem 4.3], Lm⊗wLn has property A1(1) if m > 1 or n > 1, so
that the weak* and wot topologies on it coincide. This is also the
case if m = n = 1 (cf., e.g., [4]). Thus Lm⊗wLn = Lm⊗Ln for all
m, n ∈ N. By [23], Lm⊗Ln can be identified with the higher rank
8
D. YANG
analytic Toeplitz operator algebra LΛ associated with a simple rank 2
graph Λ. Using a general result on rank 2 graphs there, one has the
following result (refer to [23, Section 1 and Section 3]).
Lemma 3.1. (Lm⊗Ln)′ = Rm⊗Rn
and (Lm⊗Ln)′′ = Lm⊗Ln.
In what follows, we identify the standard predual of Lm⊗Ln.
Proposition 3.2. We have the following completely isometric isomor-
phism
(Lm⊗Ln)∗
∼= Lm∗b⊗Ln∗.
(1)
Proof. Although [16, Theorem 7.2.4] handles the selfadjoint dual alge-
bras (i.e., von Neumann algebras), the proof of our proposition follows
the same line there. For self-containedness, it is also included here.
Clearly, it suffices to show that Lm⊗F Ln ⊆ Lm⊗Ln. By Lemma
3.1, we just need to check that each T ∈ Lm⊗F Ln commutes with all
operators in the commutant (Lm⊗Ln)′ = Rm⊗Rn. Thus it is sufficient
to show that
T (I ⊗ B) = (I ⊗ B)T (B ∈ Rn)
and
T (A ⊗ I) = (A ⊗ I)T (A ∈ Rm).
For the first identity, notice the following identification ([16])
π : B(Hm ⊗ Hn) ∼= CB(B(Hm)∗, B(Hn))
X 7→ π(X) : ω1 7→ (ω1 ⊗ id)(X).
So it suffices to show that
(ω1 ⊗ id)(T (I ⊗ B)) = (ω1 ⊗ id)((I ⊗ B)T )
for all w1 ∈ B(Hm)∗, namely,
(ω1 ⊗ id)(T )B = B(ω1 ⊗ id)(T ).
The above identity holds true since T ∈ Lm⊗F Ln implies that (ω1 ⊗
id)(T ) ∈ Ln.
The second identity can be proved similarly.
Before giving our next result, let us recall that, for 1 ≤ k ≤ ∞, an
isometric n-tuple S is said to be pure of multiplicity k if S is unitarily
equivalent to L(k). In particular, any pure isometric tuple is analytic.
Also, a subspace W is wandering for an isometric n-tuple S if the
subspaces {SuW : u ∈ F+
n } are pairwise orthogonal.
Proposition 3.3. Let L be the left regular representation of F+
for any k > 1, L⊗k is a pure isometric n-tuple with multiplicity ∞.
n . Then
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
9
Proof. 1 Obviously, L⊗k is an isometric n-tuple. In what follows, we
prove that L⊗k is pure of multiplicity ∞. Set
K = span{ξu1 ⊗ · · · ⊗ ξuk : u1, ..., un have no common prefix}.
That is, K is spanned by the basis vectors ξu1 ⊗ · · · ⊗ ξuk , where ui = ∅
for some i ∈ {1, ..., k}, or ui = kiu′
i with ki 6= kj for some i 6= j.
It is easy to check that K is wandering for L⊗k. Suppose now ξu1 ⊗
n with ξu1 ⊗ · · · ⊗ ξuk 6∈ K.
n of maximal length such that ui = wu′
i
n (i = 1, ..., k). Notice that w 6= ∅ as ξu1 ⊗· · ·⊗ξuk 6∈ K.
uk ∈ K.
· · · ⊗ ξuk is an arbitrary basis vector for H⊗k
One can choose a word w ∈ F+
for some u′
On the other hand, since w is maximal, it follows that ξ′
Clearly,
u1⊗· · ·⊗ξ′
i ∈ F+
ξu1 ⊗ · · · ⊗ ξuk = L⊗k
w (ξ′
u1 ⊗ · · · ⊗ ξ′
uk).
Thus, for any basis vector ξu1 ⊗ · · · ⊗ ξuk ∈ H⊗k
n , we have
ξu1 ⊗ · · · ⊗ ξuk ∈ M
L⊗k
w K,
and so
w∈F+
n
n = M
H⊗k
L⊗k
w K.
w∈F+
n
Obviously, dim K = ∞ as k > 1. Therefore, it follows from [25] that
L⊗k is a pure isometric tuple with multiplicity ∞.
As an immediately consequence of the above proposition, we have
that L⊗k is analytic for all k ∈ N.
4. Analytic free semigroup algebras are Hopf algebras
Following Effros-Ruan in [17], we define the term Hopf algebras from
analysts' point of view as follows. A Hopf algebra (A, m, ∆) consists of
a linear space A with norms or matrix norms, an associative bilinear
multiplication m : A × A → A, and a coassociative comultiplication
∆ : A → Ae⊗A, where e⊗ is a suitable tensor product, and ∆ is an
algebra homomorphism. The maps are assumed to be bounded in some
appropriate sense.
In this paper, we are interested in two classes of Hopf algebras -- Hopf
dual algebras and Hopf convolution dual algebras, where the latter is
induced from the former. The former is investigated in this section,
and the latter will be studied in next section. Some definitions are
first.
1I am indebted to Adam Fuller for showing me this proof.
10
D. YANG
Definition 4.1. (i) A dual algebra A is said to be a Hopf dual algebra
if there is an injective unital weak*-continuous completely contractive
homomorphism ∆ : A → A⊗A such that it is also coassociative:
(id ⊗ ∆)∆ = (∆ ⊗ id)∆.
That is, the following diagram commutes:
A
∆
∆
A⊗A
∆ ⊗ id
A⊗A
id ⊗ ∆
A⊗A⊗A.
The homomorphism ∆ is called a comultiplication or coproduct on
A.
(ii) A Hopf dual algebra is counital
if there is a non-zero weak*-
continuous homomorphism ǫ : A → C such that
(ǫ ⊗ id) ◦ ∆ = id = (id ⊗ ǫ) ◦ ∆.
The homomorphism ǫ is called a counit of A.
(iii) A Hopf dual algebra A is said to be cocommutative if σ ◦ ∆ = ∆,
where σ is the flip map a ⊗ b 7→ b ⊗ a.
We shall use the pair (A, ∆), or (A, ∆A) if A needs to be stressed,
or simply A if the context is clear, to denote the Hopf dual algebra A
with the comultiplication ∆.
Definition 4.2. A morphism of two Hopf dual algebras (A, ∆A) and
(B, ∆B) is a unital weak*-continuous completely contractive homomor-
phism π : A → B that makes ∆A and ∆B compatible in the following
sense:
∆B ◦ π = (π ⊗ π) ◦ ∆A.
If π is weak*-weak* homeomorphic and completely isometrical iso-
morphic, then A and B are said to be isomorphic as dual Hopf algebras,
and denoted as (A, ∆A) ∼= (B, ∆B).
Definition 4.3. Let (A, ∆) be a Hopf dual algebra. A unital weak*-
continuous linear functional ϕ : A → C is called a left (resp. right)
integral on (A, ∆) if it is left-invariant (resp. right-invariant), that is,
(id ⊗ ϕ)(∆(a)) = ϕ(a)I
(resp. (ϕ ⊗ id)(∆(a)) = ϕ(a)I).
A left and right integral is briefly called an integral.
Some remarks for the above definitions are in order.
/
/
/
/
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
11
Remark 4.4. Note that if A in Definition 4.1 is selfadjoint (i.e., A is
a von Neumann algebra), then (A, ∆) is nothing but a Hopf von Neu-
mann algebra ([18]). This is due to the fact that a unital completely
contractive homomorphism between two C*-algebras is automatically
a *-homomorphism ([24]). So, as one expects, the notions of Hopf dual
and Hopf von Neumann algebras coincide in the selfadjoint case. Then,
in this case, the notion of morphisms in Definition 4.2 is the same as
the one for Hopf von Neumann algebras.
Remark 4.5. All tensor maps involved in Definitions 4.1, 4.2 and 4.3
have unique weak*-continuous extensions. For example, ∆ ⊗ id has
a unique weak*-continuous extension, still denoted by ∆ ⊗ id, which
maps A⊗A to A⊗A⊗A. This can be seen, e.g., from [17] (also cf. the
proof of Theorem 4.9 below).
Remark 4.6. Notice that integrals in Definition 4.3 are normalized.
It is not hard to check that integrals of a Hopf dual algebra are unique.
If A is a von Neumann algebra, then an integral is nothing but a Haar
state.
We are now ready to give the first main theorem in this section. It
seems to have been overlooked in the classical case.
Theorem 4.7. (Ln, ∆) is a cocommutative Hopf dual algebra with the
integral ϕ0 = [ξ∅ξ∗
7→ Li ⊗ Li for
i = 1, ..., n.
∅], where ∆ is determined by Li
Proof. Since L⊗2 is an isometric n-tuple, there is a representation ∆
of the non-commutative disk algebra An determined by
∆ : An → Ln⊗Ln, Li 7→ Li ⊗ Li
(i = 1, ..., n).
By Proposition 3.3, L⊗2 is analytic and so absolutely continuous. Thus
∆ can be extended to a weak*-continuous representation of Ln, which
is still denoted as ∆ by abusing notation:
∆ : Ln → Ln⊗Ln, Li 7→ Li ⊗ Li
(i = 1, ..., n).
12
D. YANG
Let A ∼ Pw awLw be an arbitrary element in Ln. Then
(∆ ⊗ id) ◦ ∆(A) = ∆ ⊗ id(∆(sot−lim
k
(Σk(A))))
= ∆ ⊗ id(∆(wot−lim
k
(Σk(A))))
= ∆ ⊗ id(∆(weak*−lim
k
(Σk(A))))
= weak*−lim
k
∆ ⊗ id(∆(Σk(A)))
= weak*−lim
k
= weak*−lim
k
X
w<k
X
w<k
(1 −
(1 −
w
k
w
k
)aw∆ ⊗ id(Lw ⊗ Lw)
)aw(Lw ⊗ Lw ⊗ Lw)
= wot−lim
k
X
w<k
(1 −
w
k
)aw(Lw ⊗ Lw ⊗ Lw),
(2)
where the above third "=" uses the fact that wot and weak* topologies
coincide on Ln by Theorem 2.3, the fourth "=" is from the weak*-
continuity of ∆ ⊗ id and ∆, and the last second one comes from the
definition of ∆. Some obviously slight changes of the above proof yield
(∆ ⊗ id) ◦ ∆(A) = wot−lim
k
X
w<k
(1 −
w
k
)aw(Lw ⊗ Lw ⊗ Lw).
Thus ∆ is coassociative.
Similarly, one can show the cocommutativity of Ln. Thus Ln is a
cocommutative Hopf dual algebra.
Finally, we verify that ϕ0 = [ξ∅ξ∗
∅] is the integral of Ln. Clearly ϕ0 is
a weak*-continuous linear functional and ϕ0(I) = 1. Then using some
calculations similar to the above, one obtains
(ϕ0 ⊗ id) ◦ ∆(A) = ϕ0(A)I = (id ⊗ ϕ0) ◦ ∆(A)
for all A ∈ Ln. Therefore, ϕ0 is the integral on Ln.
Remark 4.8. (i) By Proposition 3.3, L⊗3 is an analytic isometric n-
tuple. From the above proof, one actually obtains that the map (∆ ⊗
id)◦∆, or (id⊗∆)◦∆, is nothing but the representation of Ln extended
from that of An induced by L⊗3:
(∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆ : Ln → Ln⊗Ln⊗Ln
Li 7→ L⊗3
i
(i = 1, ..., n).
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
13
Also, the operator in (2) is the one with the Fourier expansion Pw aw(Lw⊗
Lw ⊗ Lw). (Refer to the proof of Theorem 6.6 below for some related
details.)
(ii) It is worth noticing that, in the free group case with n ≥ 2,
the above integral ϕ0 is the canonical faithful trace of the group von
Neumann algebra vN(Fn), which is a II1 factor.
One can now generalize the above result to an arbitrary analytic
isometric tuple.
Theorem 4.9. Suppose that S is an analytic free semigroup algebra
generated by an isometric n-tuple S = [S1 · · · Sn]. Then (S, ∆S) is a
cocommutative Hopf dual algebra with an integral, where ∆S maps Si
to Si ⊗ Si for all 1 ≤ i ≤ n.
Proof. Let φ : S → Ln be the canonical weak*-homeomorphic com-
pletely isometric isomorphism. From [16, Corollary 4.1.9], we have
that φ∗ : Ln∗ → S∗ is completely isometrically isomorphic, and so is
φ∗ ⊗φ∗ : Ln∗b⊗Ln∗ → S∗b⊗S∗ by [16, Proposition 7.1.7]. Applying [16,
Corollary 4.1.9] again gives that (φ∗ ⊗ φ∗)∗ : (S∗b⊗S∗)∗ → (Ln∗b⊗Ln∗)∗
is a weak*-homeomorphic completely isometric isomorphism. Hence
(φ∗ ⊗ φ∗)∗ is a weak*-homeomorphic completely isometric isomorphism
between S⊗F S and Ln⊗F Ln by Theorem 2.4. But it follows from
Theorem 2.4 and Proposition 3.2 that Ln⊗Ln = Ln⊗F Ln. Hence one
has
S⊗S = S⊗F S.
(3)
Let Φ be the inverse of (φ∗ ⊗ φ∗)∗ (which is actually the extension of
φ−1 ⊗ φ−1). Then Φ is a weak*-homeomorphic completely isomorphism
from Ln⊗Ln onto S⊗S.
Set ∆S := Φ ◦ ∆ ◦ φ, where ∆ is the comultiplication on Ln given in
Theorem 4.7. By [15], ∆ is weak*-continuous completely isometrically
homomorphic. Then ∆S : S → S⊗S is a unital weak*-continuous
completely isometric homomorphism. Also, from the above analysis,
∆S maps Si to Si ⊗ Si (i = 1, ..., n). Then, as in the proof of Theorem
4.7, it is not hard to check the coassociativity of ∆S and the cocom-
mutativity of S. Therefore, S is a cocommutative Hopf dual algebra.
Let ǫ = ϕ0 ◦ φ. Then one can easily verify that ǫ is an integral of S.
As a byproduct of the above proof, one can construct analytic iso-
metric tuples from a given one as follows.
Corollary 4.10. Suppose that S is an analytic isometric tuple. Then
so is S⊗k for any k ≥ 1.
14
D. YANG
Proof. Let k ≥ 1 and S be an analytic isometric n-tuple. As ob-
taining Φ in the proof of Theorem 4.9, one has that there is a weak*-
homeomorphic completely isometric isomorphism eΦ between ⊗k
i=1Ln
and ⊗k
S, the normal spatial tensor product of k copies of S. Let
π : Ln → ⊗k
i=1Ln be the weak*-continuous, completely isometric ho-
momorphism determined by Proposition 3.3. Then the composition
eΦ ◦ π is the weak*-continuous extension of the representation of An
induced by S⊗k. Therefore, S⊗k is analytic.
i=1
Suppose that S is an analytic isometric n-tuple. Then "Ln = S"
as operator algebras. As before, one can check that the canonical
homomorphism σ : S → Ln makes the comultiplications ∆ of Ln and
∆S of S compatible: ∆ ◦ φ = (φ ⊗ φ) ◦ ∆S. Thus, "Ln = S" as Hopf
dual algebras. We record this as a corollary.
Corollary 4.11. Suppose that S is an analytic free semigroup algebra
generated by an isometric n-tuple. Then (Ln, ∆) ∼= (S, ∆S).
5. Hopf Convolution Dual Algebras
Let S be an analytic free semigroup algebra generated by an isomet-
ric n-tuple. In this section, we shall show that its predual S∗ is also
a Hopf algebra. Naturally, S∗ and Ln∗ are canonically isomorphic as
Hopf algebras. Analogous to the Fourier algebra A(Fn), the algebra
Ln∗ is non-unital.
In order to equip S∗ with an algebra structure, let mS∗ = (∆S)∗,
where ∆S is the comultiplication of S constructed in the previous
section. Then it follows essentially from Proposition 3.2 and Theorem
4.9 that mS∗ gives a completely contractive multiplication of S∗. Thus
S∗ is a completely contractive Banach algebra.
To construct a coalgebra structure of S∗, we follow the same line
with Effros-Ruan [17, Section 7]. The idea is sketched as follows. In
order to endow S∗ with a comultiplication ∆S∗, one needs to use the
σh
⊗ S, because it is suitable for
normal Haagerup tensor product S
linearizing the multiplication m on S in the following sense: The mul-
tiplication m : S × S → S extends uniquely to a weak*-continuous
σh
⊗ S → S. Then it turns out that
completely contractive map mS : S
its preadjoint (mS)∗ produces a comultiplication of S∗. To prove that
it is indeed an algebra homomorphism, we need two important ingredi-
ents: (i) a Shuffle Theorem, and (ii) (S⊗S)∗
we have both (i) and (ii): (i) holds true "automatically" because the
Shuffle Theorem [17, Theorem 6.1] holds true for arbitrary operator
∼= S∗b⊗S∗. Fortunately,
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
15
spaces, and (ii) can be proved based on some results obtained in pre-
vious sections.
Recall that, given operator spaces V and W , the normal and ex-
eh
tended Haagerup tensor products are connected by V ∗
⊗
W )∗. Also, a complex algebra A is called a completely contractive
Banach algebra if A is an operator space and the multiplication m :
A × A → A is a completely contractive bilinear mapping, namely, it
σh
⊗ W ∗ = (V
determines a completely contractive linear mapping m : Ab⊗A → A.
Theorem 5.1. Suppose that S is an analytic free semigroup algebra.
Then (S∗, mS∗, ∆S∗) is a Hopf algebra, where the multiplication mS∗ :
S∗b⊗S∗ → S∗ and the comultiplication ∆S∗ : S∗ → S∗
eh
⊗ S∗ are
completely contractive, and they are given by mS∗ = (∆S)∗ and ∆S∗ =
(mS)∗, respectively.
Proof. By Proposition 3.2 and the identity (3), one has
(S⊗S)∗
∼= S∗b⊗S∗.
(4)
Since the comultiplication ∆S : S → S⊗S is a weak*-continuous
complete isometry (from the proof of Theorem 4.9), it follows from
[16, Corollary 4.1.9] that the preadjoint (∆S)∗ is a completely quo-
tient mapping from S∗b⊗S∗ to S∗. Hence (S∗, (∆S)∗) is a completely
contractive Banach algebra. More precisely, for any ϕ, ψ ∈ S∗, the
multiplication ϕ ∗ ψ := (∆S)∗(ϕ ⊗ ψ) is defined by
ϕ ∗ ψ(x) = ϕ ⊗ ψ(∆S(x))
(x ∈ S).
It is now straightforward to check that ϕ ∗ ψ = ψ ∗ ϕ, implying the
commutativity of (S∗, ∗). Therefore (S∗, (∆S)∗) is a completely con-
tractive commutative Banach algebra.
Using the isomorphism (4) and the Shuffle Theorem [17, Theorem
6.1], one can obtain that (mS)∗ is a completely contractive (algebra)
homomorphism. The proof is completely similar to that of [17, The-
orem 7.1] (also refer to the discussion preceding the statement of this
theorem), and so it is omitted here.
Therefore (S∗, (∆S)∗, (mS)∗) is a Hopf algebra.
Applying Theorems 4.7 and 5.1 to the case of n = 1, we immedi-
ately obtain the following result in the classical case, which seems new.
Recall that (H ∞)∗ = L1(T)/H1(0), where H1(0) = zH1(T) ([22]).
Corollary 5.2. H ∞ and L1(T)/H1(0) are Hopf algebras.
16
D. YANG
Following Effros-Ruan [17], Hopf algebras obtained from Theorem
5.1 are called Hopf convolution dual algebras. Of course, counits, mor-
phisms and integrals can be defined as those in Definitions 4.1, 4.2 and
4.3, but without the weak*-continuity requirement.
Since Fn is a free group, it is well known that the Fourier algebra
A(Fn) is not unital. Actually, for n ≥ 2, more is true: A(Fn) does not
have bounded approximate identities. In our case, so far we can show
that Ln∗ is not unital. Before proving this result, let us first recall
from [6, 11] that the set M(Ln) of wot-continuous multiplicative
linear functionals consists of those linear functionals having the form
ϕλ = [νλν∗
λ], where λ ∈ Bn (the open unit ball of Cn) and νλ is a
certain unit vector in Hn (whose precise definition is not important for
us and so is omitted here). It is known that ϕλ(p(L)) = p(λ) for any
polynomial p = Pw aww in the semigroup algebra CF+
n .
Proposition 5.3. If S is an analytic free semigroup algebra, then S∗
is non-unital.
Proof. It is not hard to see that it suffices to show that S is not
counital. To the contrary, assume that it has a counit ǫS. Then ǫ :=
φ−1◦ǫS is a counit of Ln. So ǫ = ϕλ for some λ ∈ Bn. From the identity
required in the definition of a counit, it is easy to check that ǫ(Lw) = 1
for all w ∈ F+
n . However, as mentioned above, ϕλ(Lw) = w(λ), where
w(λ) = λk1
n . This particularly forces λi = 1
for all i = 1, ..., n. Obviously, this is impossible as λ ∈ Bn.
n for w = ik1
1 · · · λkn
1 · · · ikn
It is worthy to notice that M(Ln) is also closed under the multipli-
cation ∗ of Ln∗ and has an involution †. In fact, for ϕλ, ϕµ ∈ {ϕν :
ν ∈ Bn}, then ϕλ ∗ ϕµ = ϕλ∗µ, where λ ∗ µ = (λ1µ1, ..., λnµn) for
λ = (λ1, ..., λn) and µ = (µ1, ..., µn) in Bn. Also, † can be defined by
ϕ†
λ = ϕλ.
If S is an analytic free semigroup algebra generated by an isometric
n-tuple, then S and Ln are completely isometrically isomorphic. So
∼= Ln∗ as operator spaces. The following result tells us this is also
S∗
the case as Hopf convolution algebras.
Proposition 5.4. Suppose that S is an analytic free semigroup algebra
∼= Ln∗ as Hopf convolution
generated by an isometric n-tuple. Then S∗
algebras.
Proof. For brevity, we use ∗ for the multiplication of Ln∗ and that
of S∗ constructed above.
It follows from [16, Corollary 4.1.9] that
φ∗ : Ln∗ → S∗ is completely isometric. We now show that it is also
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
17
an algebra homomorphism, namely, φ∗(f ∗ g) = φ∗(f ) ∗ φ∗(g) for all
f, g ∈ Ln∗. Indeed, we have
φ∗(f ∗ g) = (f ∗ g) ◦ φ = (f ⊗ g) ◦ ∆ ◦ φ
= (f ⊗ g) ◦ Φ ◦ ∆S
= (f ⊗ g) ◦ (φ ⊗ φ) ◦ ∆S
= (φ∗ ⊗ φ∗) ◦ (f ⊗ g) ◦ ∆S
= (φ∗(f ) ⊗ φ∗(g)) ◦ ∆S
= φ∗(f ) ∗ φ∗(g).
Also, it is easy to check that (φ∗)−1 = (φ−1)∗. Therefore, φ∗ is an
algebra isomorphism.
In order to show that φ∗ is compatible with the comultiplications on
Ln∗ and S∗, consider the following commuting diagram
S
mS
φ
Ln
mLn
S
σh
⊗ S
φ ⊗ φ
σh
⊗ Ln
Ln
and then take its preadjoint
S∗
∆S∗
eh
⊗ S∗
S∗
φ∗
Ln∗
∆Ln ∗
Ln∗
eh
⊗ Ln∗
.
eh
⊗ φ∗
φ∗
6. Completely Bounded Representations of S∗
Let S be an analytic free semigroup algebra. In this section, we shall
study completely bounded representations of S∗. Through those rep-
resentations, S and S∗ are closely connected with each other. On one
hand, we show that there is a bijection between the set of completely
bounded representations of S∗ and the set of corepresentations of S.
On the other hand, S can be recovered by completely bounded repre-
sentations of S∗: The algebra induced from the coefficient operators
of completely bounded representations of S∗ is precisely S. This gives
us a new duality between S∗ and S.
/
/
O
O
O
O
/
/
o
o
o
o
18
D. YANG
As an amusing application of the former bijection, we prove that
the spectrum of S∗, as a commutative Banach algebra, is precisely
F+
n . In order to achieve the latter duality, one needs to introduce the
notion of tensor products of completely bounded representations of S∗.
Fortunately, for this one can borrow it from [17, Section 7]. Actually,
the entire subsection on tensor products is from there. It should be
mentioned that we have the equality C(S∗) = S as Theorem 6.7, rather
than just an inclusion C(S∗) ⊆ S in [17, Theorem 7.2], because we
are dealing with only "Fourier algebras", rather than general Hopf
convolution dual algebras.
Corepresentations of S. We begin this subsection with the following
definition.
Definition 6.1. Let (A, ∆) be a Hopf dual algebra acting on H. A
corepresentation of (A, ∆) on a Hilbert space K is an (arbitrary) oper-
ator V ∈ A ⊗ B(K) satisfying
V1,3V2,3 = (∆ ⊗ id)(V ).
Here V1,3 and V2,3 are the standard leg notation (cf. [1]): V1,3 is a linear
operator on the Hilbert space H ⊗ H ⊗ K, which acts as V on the first
and third tensor factors and as the identity on the second one. The
notation V2,3 is defined similarly.
Before stating our first main theorem in this section, we need the
following lemma, which is probably a folklore. Here we give a direct
proof based on special properties of Ln.
Lemma 6.2. Suppose that S is an analytic free semigroup algebra.
Then we have a completely isometric isomorphism
CB(S∗, B(K)) ∼= S⊗B(K).
Proof. We first claim that this lemma holds true when S = Ln:
CB(Ln∗, B(K)) ∼= Ln⊗B(K).
(5)
To this end, observe that
(Ln⊗B(K))′ = Rn⊗CI
and (Ln⊗B(K))′′ = Ln⊗B(K).
Then, applying some obvious modifications to the proof of Proposition
isometrically isomorphic. On the other hand, there is a natural com-
3.2, one can show that (Ln∗b⊗B(K)∗)∗ and Ln⊗B(K) are completely
pletely isometric isomorphism (Ln∗b⊗B(K)∗)∗ ∼= CB(Ln∗, B(K)) by [16,
Corollary 7.1.5]. This proves our claim.
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
19
Now suppose S is an arbitrary analytic free semigroup algebra. Us-
ing a proof completely similar to that of Theorem 4.9, we obtain that
(S∗b⊗B(K)∗)∗ ∼= (Ln∗b⊗B(K)∗)∗
S⊗B(K) ∼= Ln⊗B(K).
and
Therefore
CB(S∗, B(K)) ∼= (S∗b⊗B(K)∗)∗ ∼= (Ln∗b⊗B(K)∗)∗
∼= CB(Ln∗, B(K)) ∼= Ln⊗B(K)
∼= S⊗B(K),
where the first ∼= is from [16, Corollary 7.1.5], while the fourth one is
from (5).
Theorem 6.3. Suppose that S is an analytic free semigroup algebra
acting on a Hilbert space H. Then there is a bijection between the
set of all corepresentations of S and the set of all completely bounded
representations of S∗.
More precisely, if V is a corepresentation of S on K, then the cor-
responding representation πV of S∗ is given by
πV (ϕ) = (ϕ ⊗ id)(V )
(ϕ ∈ S∗);
(6)
if π is a completely bounded representation of S∗ on K, then the cor-
responding corepresentation Vπ of S is given by
(Vπ(ξ ⊗ x), η ⊗ y) = (π([ξη∗])x, y)
(7)
for all ξ, η ∈ H, x, y ∈ K.
Proof. The following calculations show that, for a corepresentation V
of S, the map πV defined in (6) indeed yields a representation of S∗
on K: For ϕ ∈ S∗, we have
(ϕ ⊗ id)(V )(ψ ⊗ id)(V ) = (ϕ ⊗ ψ ⊗ id)(V1,3V2,3)
= (ϕ ⊗ ψ ⊗ id)((∆S ⊗ id)(V ))
= ϕ ∗ ψ(V ).
The fact that πV is completely bounded follows from the completely
isometric isomorphism in Lemma 6.2.
For π ∈ CB(S∗, B(K)), the identification given in Lemma 6.2 deter-
mines an operator Vπ ∈ S⊗B(K) via π(ϕ) = (ϕ⊗id)(Vπ) for all ϕ ∈ S∗.
Reversing the above proof, we see that Vπ is a corepresentation of S.
In order to get (7), it suffices to make use of Theorem 2.3 stating
that S has property A1(1).
20
D. YANG
It is time to look at some examples.
Example 6.4. Let W ∈ B(Hn ⊗ Hn) be defined by W (ξu ⊗ ξv) =
ξvu⊗ξv. Then one can show that W is a proper isometry and intertwines
Lw ⊗ Lw and id ⊗ Lw:
(Lw ⊗ Lw)W = W (id ⊗ Lw)
for all w ∈ F+
n .
Also one can check that W is a corepresentation of Ln. In fact, checking
on the basis vectors, it is easy to see that it satisfies the identity required
in Definition 6.1. In order to check W ∈ Ln⊗B(Hn), it suffices to check
that W commutes each element in the commutant (Ln⊗B(Hn))′ =
Rn ⊗ CI. For Ru ∈ Rn (u ∈ F+
n ), then a simple calculation shows that
W (Ru ⊗ id) = (Ru ⊗ id)W . We are done.
It is not hard to check that the completely bounded representation
πW of Ln∗ corresponding to the above W in Theorem 6.3 is nothing
but the left multiplication representation:
πW (ϕ)ξu = ϕ(Lu)ξu
(ϕ ∈ Ln∗).
Two points are worth being mentioned here. Firstly, in the group
case, the above corepresentation W is known as the fundamental op-
erator in the theory of Hopf von Neumann algebras. Secondly,
in
spite of the fact that the above W is a proper isometry, the tuples
[L1 ⊗ L1 · · · Ln ⊗ Ln] and [id ⊗ L1 · · · id ⊗ Ln] are unitarily equivalent
because both of them are pure of multiplicity ∞ (cf. [25] and Section
3).
Example 6.5. Fix w ∈ F+
n . Let ρw : Ln∗ → C be defined by ρw(ϕ) =
ϕ(Lw). Then it is a character of Ln∗. In fact, for all ϕ, ψ ∈ Ln∗, we
have
ρw(ϕ ∗ ψ) = ϕ ∗ ψ(Lw) = ϕ ⊗ ψ(∆(Lw))
= ϕ ⊗ ψ(Lw ⊗ Lw) = ϕ(Lw)ψ(Lw)
= ρw(ϕ)ρw(ψ).
In particular, it is a completely contractive representation. It is easy
to check that its corresponding corepresentation Vρw of Ln in Theorem
6.3 is given by
Vρw = Lw ⊗ idC ∼= Lw.
In what follows, as an amusing application of Theorem 6.3, we iden-
tify the spectrum of S∗. The proof also takes advantage of the useful
idea of Fourier expansions of operators in Ln and Ln⊗Ln.
Theorem 6.6. Let S be an analytic free semigroup algebra generated
by an isometric n-tuple. Then the (Gelfand) spectrum of S∗ is F+
n .
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
21
Proof. Obviously, it suffices to show that the spectrum ΣLn ∗ of Ln∗ is
F+
n .
By Example 6.5 above, one has F+
n ⊆ ΣLn ∗. In order to prove ΣLn ∗ ⊆
F+
n , we first observe that every operator A ∈ Ln⊗Ln is completely
determined by the image A(ξ∅ ⊗ ξ∅) of the "vacuum vector" ξ∅ ⊗ ξ∅
under A. Indeed, assume that
A(ξ∅ ⊗ ξ∅) = X
au,v ξu ⊗ ξv
with au,v ∈ C. Then for all α, β ∈ F+
n , one has
u,v∈F+
n
A(ξα ⊗ ξβ) = ARα ⊗ Rβ(ξ∅ ⊗ ξ∅)
= Rα ⊗ RβA(ξ∅ ⊗ ξ∅)
au,v ξu ⊗ ξv)
= Rα ⊗ Rβ(X
= X
u,v
au,v ξu eα ⊗ ξv eβ,
u,v
where eα is the word in F+
n obtained by reversing the order of α, and
similarly for eβ. So similar to Ln, we heuristically view A as its Fourier
expansion:
A ∼ X
au,v Lu ⊗ Lv,
u,v∈F+
n
where the (u, v)-th Fourier coefficient is given by au,v = (A(ξ∅⊗ξ∅), ξu⊗
ξv).
Now assume that ϕ : Ln∗ → C is a character of Ln∗. Let cϕ
1,1 ∈ Ln
be the coefficient operator defined by
hf, cϕ
1,1i = ϕ(f )
(f ∈ Ln∗).
Since ϕ is a character, it is automatically a completely bounded repre-
sentation. By Theorem 6.3, we have that V := cϕ
1,1 ∈ Ln is a corepre-
sentation of Ln. That is,
V1,3V2,3 = (∆ ⊗ idC)(V ).
(8)
Since V ∈ Ln, one can assume that it has the following Fourier expan-
sion:
V ∼ X
aw Lw.
w∈F+
n
Then it follows from (8) that
X
awLw ⊗ X
awLw ∼ X
w
w
w
aw(Lw ⊗ Lw) ∈ Ln⊗Ln.
(9)
22
D. YANG
Then taking the (w, w)-th Fourier coefficients of both sides of (9) gives
awLw ⊗ X
((X
= (X
w
awLw)(ξ∅ ⊗ ξ∅), ξw ⊗ ξw)
w
aw(Lw ⊗ Lw)(ξ∅ ⊗ ξ∅), ξw ⊗ ξw).
w
This implies aw = 0 or 1 for every w ∈ F+
n . On one hand, there is a
word w ∈ F+
n such that aw 6= 0 as ϕ 6= 0 implies V 6= 0. On the other
hand, suppose that there are u 6= v ∈ F+
n such that both au and av are
non-zero. Then au = av = 1. Taking the (u, v)-th Fourier coefficients
of both sides of (9), we have
1 = auav = ((X
= (X
w
awLw ⊗ X
w
awLw)(ξ∅ ⊗ ξ∅), ξu ⊗ ξv)
aw(Lw ⊗ Lw)(ξ∅ ⊗ ξ∅), ξu ⊗ ξv)
w
= auδu,v = 0.
This is an obvious contradiction. So V = Lw for some w ∈ F+
fore, ΣLn ∗ ⊆ F+
n .
n . There-
Tensor products of completely bounded representations of S∗.
This subsection borrows from [17, Section 7], and is rather sketched.
Refer to [17, Section 7] for details.
Suppose that πi : S∗ → B(Hi) (i = 1, 2) are two completely bounded
representations of S∗. Define their multiplication π1 × π2 as the com-
position of the following mappings
∆S∗−→ S∗
eh
⊗ S∗
S∗
π1⊗π2−→ B(H1)
⊆B(H1)
eh
⊗ B(H2)
σh
⊗ B(H2) θ→ B(H1 ⊗ H2),
where θ is determined by the product of
B(H1) → B(H1 ⊗ H2) : S 7→ S ⊗ IH2
and
Let
B(H2) → B(H1 ⊗ H2) : T 7→ IH1 ⊗ T.
C(S∗) = (cid:8)cπ
ξ,η : cπ
ξ,η(f ) = (π(f )ξ, η), f ∈ S∗, ξ, η ∈ Hπ(cid:9)
be the set of coefficient operators of completely bounded representa-
tions of S∗. Then we have the following duality result.
FREE SEMIGROUP ALGEBRAS AND HOPF ALGEBRAS
23
Theorem 6.7. Suppose that S is an analytic free semigroup algebra.
Then C(S∗) = S.
Proof. This directly follows from [17, Theorem 7.2] and Example 6.5.
Let us end this paper with two remarks. Firstly, most of the results
on Hopf convolution dual algebras of this paper hold true for more
general ones, e.g., those induced from Hopf dual algebras A with the
property (A⊗A)∗
there is a certain analogy between the Fourier algebra A(Fn) and Ln∗.
So, basically, for whatever property A(Fn) has, one could ask if it
holds true for Ln∗ as long as it makes sense. This and more will be
investigated in the future.
∼= A∗b⊗A∗. Secondly, as we have seen from above,
Acknowledgements. I would like to thank my colleague Prof. Zhiguo
Hu for several useful conversations at the early stage of this work, and
Prof. Matthew Kennedy for some useful discussions after I gave a talk
at the CMS 2011 Winter Meeting, in which some results of this paper
were presented. Thanks also go to Prof. Laurent Marcoux for showing
me some references.
References
[1] S. Baaj and G. Skandalis, Unitaires multiplicatifs et dualit´e ppour les produits
crois´es de C*-alg`ebres, Ann. Sci. ´Ecole Norm. Sup. (4) 26 (1993), 425 -- 488.
[2] H. Bercovici, Hyper-reflexivity and the factorization of linear functionals, J.
Funct. Anal. 158 (1998), 242 -- 252.
[3] H. Bercovici, Operator theory and Arithmetic in H ∞, American Mathematical
Society, Providence, 1988.
[4] H. Bercovici, Hari and D. Westwood, The factorization of functions in the
polydisc, Houston J. Math. 18 (1992), 1 -- 6.
[5] J. B. Conway, A Course in Operator Theory, Graduate Studies in Mathematics,
Volume 21, American Mathematical Society, Providence, Rhode Island, 2000.
[6] K. R. Davidson, Free semigroup algebras: a survey, Systems, approximation,
singular integral operators, and related topics (Bordeaux, 2000), 209 -- 240,
Oper. Theory Adv. Appl. 129, Birkhauser, Basel, 2001.
[7] B(H) is a free semigroup algebra, Proc. Amer. Math. Soc. 134 (2006), 1753 --
1757.
[8] K. R. Davidson, E. Katsoulis and D. R. Pitts, The structure of free semigroup
algebras, J. reine angew. Math. (Crelle) 533 (2001), 99 -- 125.
[9] K. R. Davidson, D. W. Kribs and M. Shpigel, Isometric Dilations of non-
commuting finite rank n-tuples, Canad. J. Math. 53 (2001), 506 -- 545.
[10] K. R. Davidson, J. Li and D. R. Pitts, Absolutely continuous representations
and a Kaplansky theorem for free semigroup algebras, J. Func. Anal. 224
(2005), 160 -- 191.
24
D. YANG
[11] K. R. Davidson and D. R. Pitts, The algebraic Structure of non-commutative
analytic Toeplitz algebras, Math. Annalen 311 (1998), 275 -- 303.
[12] K. R. Davidson and D. R. Pitts, Nevanlinna-Pick Interpolation for non-
commutative analytic Toeplitz algebras, Integral Equations and Operator The-
ory 31 (1998), 321 -- 337.
[13] K. R. Davidson and D. R. Pitts, Invariant subspaces and hyper-reflexivity for
free semigroup algebras, Proc. London Math. Soc. 78 (1999), 401 -- 430.
[14] K. R. Davidson and A. Wright, Operator algebras with unique preduals, Canad.
Math. Bull. 54 (2011), 411 -- 421.
[15] K. R. Davidson and D. Yang, A note on absolute continuity in free semigroup
algebras, Houston J. Math. 34 (2008), 283 -- 288.
[16] E. G. Effros and Z.-J. Ruan, Operator Spaces, London Mathematical Society
Monographs. New Series, 23. The Clarendon Press, Oxford University Press,
New York, 2000.
[17] E. G. Effros andZ.-J. Ruan, Operator space tensor products and Hopf convolu-
tion algebras J. Operator Theory 50 (2003),131 -- 156.
[18] M. Enock and J.-M. Schwartz, Kac Algebras and Duality of Locally Compact
Groups, Springer-Verlag, 1992.
[19] P. Eymard, L'alg`ebre de Fourier d'un groupe localement compact, Bull. Soc.
Math. France 92 (1964), 181 -- 236.
[20] M. Kennedy, Wandering vectors and the reflexivity of free semigroup algebras,
J. Reine Angew. Math. 653 (2011), 47 -- 73.
[21] M. Kennedy, The structure of an isometric tuple, arXiv:1001.3182.
[22] P. Koosis, Introduction to H
p spaces, Second Edition, Cambridge Tracts in
Mathematics 115, Cambridge University Press, 1998.
[23] D. W. Kribs and S. C. Power, The analytic algebras of higher rank graphs,
Math. Proc. R. Ir. Acad. 106A (2006), 199 -- 218.
[24] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge
Studies in Advanced Mathematics, 78. Cambridge University Press, Cam-
bridge, 2002.
[25] G. Popescu, Isometric dilations for infinite sequences of noncommuting oper-
ators, Trans. Amer. Math. Soc. 316 (1989), 523 -- 536.
[26] G. Popescu, Von Neumann inequality for (B(H)n)1, Math.Scand. 68 (1991),
292 -- 304.
[27] G. Popescu, Non-commutative disc algebras and their representations, Proc.
Amer. Math. Soc. 124 (1996), 2137 -- 2148.
[28] Z.-J. Ruan, On the predual of dual algebras, J. Operator Theory 27 (1992),
179 -- 192.
[29] J. Wermer, On invariant subspaces of normal operators, Proc. Amer. Math.
Soc. 3 (1952), 270 -- 277.
Dilian Yang, Department of Mathematics & Statistics, University
of Windsor, Windsor, ON N9B 3P4, CANADA
E-mail address: [email protected]
|
1706.05654 | 2 | 1706 | 2017-11-27T10:26:15 | Minimal and maximal matrix convex sets | [
"math.OA"
] | To every convex body $K \subseteq \mathbb{R}^d$, one may associate a minimal matrix convex set $\mathcal{W}^{\textrm{min}}(K)$, and a maximal matrix convex set $\mathcal{W}^{\textrm{max}}(K)$, which have $K$ as their ground level. The main question treated in this paper is: under what conditions on a given pair of convex bodies $K,L \subseteq \mathbb{R}^d$ does $\mathcal{W}^{\textrm{max}}(K) \subseteq \mathcal{W}^{\textrm{min}}(L)$ hold? For a convex body $K$, we aim to find the optimal constant $\theta(K)$ such that $\mathcal{W}^{\textrm{max}}(K) \subseteq \theta(K) \cdot \mathcal{W}^{\textrm{min}}(K)$; we achieve this goal for all the $\ell^p$ unit balls, as well as for other sets. For example, if $\overline{\mathbb{B}}_{p,d}$ is the closed unit ball in $\mathbb{R}^d$ with the $\ell^p$ norm, then \[ \theta(\overline{\mathbb{B}}_{p,d}) = d^{1-|1/p - 1/2|}. \] This constant is sharp, and it is new for all $p \neq 2$. Moreover, for some sets $K$ we find a minimal set $L$ for which $\mathcal{W}^{\textrm{max}}(K) \subseteq \mathcal{W}^{\textrm{min}}(L)$. In particular, we obtain that a convex body $K$ satisfies $\mathcal{W}^{\textrm{max}}(K) = \mathcal{W}^{\textrm{min}}(K)$ if and only if $K$ is a simplex.
These problems relate to dilation theory, convex geometry, operator systems, and completely positive maps. We discuss and exploit these connections as well. For example, our results show that every $d$-tuple of self-adjoint operators of norm less than or equal to $1$, can be dilated to a commuting family of self-adjoints, each of norm at most $\sqrt{d}$. We also introduce new explicit constructions of these (and other) dilations. | math.OA | math | MINIMAL AND MAXIMAL MATRIX CONVEX SETS
BENJAMIN PASSER, ORR MOSHE SHALIT, AND BARUCH SOLEL
Abstract. To every convex body K ⊆ Rd, one may associate a minimal matrix convex
set W min(K), and a maximal matrix convex set W max(K), which have K as their ground
level. The main question treated in this paper is: under what conditions on a given pair of
convex bodies K, L ⊆ Rd does W max(K) ⊆ W min(L) hold? For a convex body K, we aim
to find the optimal constant θ(K) such that W max(K) ⊆ θ(K) · W min(K); we achieve this
goal for all the (cid:96)p unit balls, as well as for other sets. For example, if Bp,d is the closed unit
ball in Rd with the (cid:96)p norm, then
θ(Bp,d) = d1−1/p−1/2.
This constant is sharp, and it is new for all p (cid:54)= 2. Moreover, for some sets K we find a
minimal set L for which W max(K) ⊆ W min(L). In particular, we obtain that a convex body
K satisfies W max(K) = W min(K) if and only if K is a simplex.
These problems relate to dilation theory, convex geometry, operator systems, and com-
pletely positive maps. We discuss and exploit these connections as well. For example, our
results show that every d-tuple of self-adjoint operators of norm less than or equal to 1,
can be dilated to a commuting family of self-adjoints, each of norm at most
d. We also
introduce new explicit constructions of these (and other) dilations.
√
7
1
0
2
v
o
N
7
2
]
.
A
O
h
t
a
m
[
2
v
4
5
6
5
0
.
6
0
7
1
:
v
i
X
r
a
1. Introduction
1.1. Overview. This paper treats containment problems for matrix convex sets. A matrix
convex set in d-variables is a set S = ∪nSn, where every Sn consists of d-tuples of n ×
n matrices, that is closed under direct sums, unitary conjugation, and the application of
completely positive maps. Matrix convex sets are closely connected to operator systems
and have been investigated for several decades. Recently, they appeared in connection with
the interpolation problem for UCP maps (see, e.g., [6, 11, 24]), and also in the setting of
relaxation of spectrahedral containment problems [11, 13].
Given a closed convex set K ⊆ Rd, one can define several matrix convex sets S such that
S1 = K. One may ask, to what extent does the "ground level" S1 determine the structure
and the size of S? For example: given two matrix convex sets S = ∪nSn and T = ∪nTn,
what does containment at the first level S1 ⊆ T1 imply about the relationship between S
and T ? Of course, there is (usually) no reason that S ⊆ T would follow as a consequence
of S1 ⊆ T1, but in many cases - given some conditions on S1 - one can find a constant C
such that
S1 ⊆ T1 =⇒ S ⊆ C · T .
2010 Mathematics Subject Classification. 47A20, 47A13, 46L07, 47L25.
Key words and phrases. matrix convex set; dilation; abstract operator system; matrix range.
The work of B. Passer is partially supported by a Zuckerman Fellowship at the Technion.
The work of O.M. Shalit is partially supported by ISF Grants no. 474/12 and 195/16.
1
One way to treat the problem follows from the observation that given a closed convex set
K ⊆ Rd, there exist a minimal matrix convex set W min(K), and a maximal matrix convex
set W max(K), which have K as their ground level: K = W min
(K). To solve
the above problem, one can let K = S1, and concentrate on the problem of finding the best
constant C for which
(K) = W max
1
1
W max(K) ⊆ C · W min(K).
This problem was treated in [6, 8, 13]. For example, if K enjoys some symmetry properties,
then it was shown that C = d works. If K is the Euclidean ball B2,d ⊆ Rd, then C = d is
the optimal constant.
In this paper we find the optimal constant for a large class of sets. For example, it was
known that for the cube K = [−1, 1]d (which is just the unit ball of the (cid:96)∞ norm), the
√
constant C = d works, but optimality was not known. We find that the best constant for
d, as part of a more general technique which can assign unequal
the cube is actually C =
norms to the dilations. Moreover, we find sharp constants for the (cid:96)p balls, p ∈ [1,∞].
We also treat the closely related problem of finding conditions on two convex bodies K, L
such that W max(K) ⊆ W min(L). In particular, we show that W min(K) = W max(K) if and
only if K is a simplex (somewhat improving a result from [8], where this was essentially
obtained under the assumption that K is a polytope).
In order to describe our main results more clearly, we now turn to setting the notation
and reviewing some preliminaries.
1.2. Notation and preliminaries.
General background and notation. Let d ∈ N, and let M d
sa denote d-tuples of
matrices or self-adjoint matrices, respectively. We consider the disjoint unions Md = ∪nM d
n
and Md
n)sa. The algebra of bounded operators on a Hilbert space H is denoted
by B(H), B(H)d denotes d-tuples of operators, and B(H)d
sa denotes d-tuples of self-adjoint
operators.
sa = ∪n(M d
n and (Mn)d
We will use basic results on C*-algebras and operator systems freely (see [5] and [20],
respectively), but let us recall a few definitions. If A and B are C*-algebras, a linear map
φ : A → B is said to be positive if φ(a) ≥ 0 whenever a ≥ 0. If A and B have units, then φ is
said to be unital if φ(1) = 1. The n× n matrix algebra over A is denoted Mn(A), and is also
a C*-algebra. A linear map φ : A → B can be promoted to a map φn : Mn(A) → Mn(B) by
acting componentwise, and a map φ is said to be completely positive if φn is positive for all
n. A unital and completely positive map will be called, briefly, a UCP map, and the set of
all UCP maps between A and B will be denoted UCP(A, B).
By Stinespring's theorem [23], for every UCP map φ : A → B(H), there exists a Hilbert
space K, an isometry v : H → K, and a unital ∗-representation π : A → B(K), such that
φ(a) = v∗π(a)v , a ∈ A.
Moreover, one can arrange that the closed subspace spanned by π(A)vH is equal to K;
under this condition (K, v, π) is determined uniquely, and is referred to as the (minimal)
Stinespring dilation of φ.
2
UCP maps between matrix algebras Mn = Mn(C) can be characterized using the unital
case of Choi's theorem (see [4]): a map φ : Mn → Mk is UCP if and only if it has the form
N(cid:88)
φ(a) =
tiat∗
i , a ∈ Mn,
j=1
1 + . . . + tN t∗
for some N ≤ kn and t1, . . . , tN ∈ Mk,n with t1t∗
N = IMk. Choi's theorem is a
special case of Stinespring dilation.
An operator system is a vector subspace S of a unital C*-algebra A such that 1A ∈ S and
S = S∗. Given a tuple M = (M1, . . . , Md) of elements in the same unital C*-algebra, we let
S(M ) or SM denote the operator system generated by M1, . . . , Md, and we write C∗(SM )
or C∗(M ) for the unital C*-algebra generated by M1, . . . , Md. Note that since an operator
system S is defined in reference to a larger unital C∗-algebra, we may discuss positivity of
elements in S; in fact, S is spanned by its positive elements. Since the same reasoning applies
to the operator systems Mn(S) ⊆ Mn(A), the definitions of positive, completely positive, and
UCP maps carry over to the discussion of linear maps between operator systems. Arveson's
extension theorem [2, Theorem 1.2.3] states that if S is an operator system contained in a
unital C*-algebra A, then every UCP map of S into B(H) extends to a UCP map from A
into B(H).
Matrix convex sets. We now give some background on matrix convex sets; see [6, Sections
2,3] for some more details.
A free set (in d free dimensions) S is a disjoint union S = ∪nSn ⊆ Md, where Sn ⊆ M d
n.
In this paper, the focus will be on free sets in Md
sa. Containment is defined in the obvious
way: we say that S ⊆ T if Sn ⊆ Tn for all n. A free set S is said to be open/closed/convex
if Sn is open/closed/convex for all n. It is said to be bounded if there is some C such that
for all n and all A ∈ Sn, it holds that (cid:107)Ai(cid:107) ≤ C for all i.
An nc set is a free set that is closed under direct sums and under simultaneous unitary
conjugation.
An nc set is said to be matrix convex if it is closed under the application of UCP
maps, meaning that whenever X is in Sn and φ ∈ UCP(Mn, Mk), the tuple φ(X) :=
(φ(X1), . . . , φ(Xd)) is in Sk. The main examples of matrix convex sets are given by free
spectrahedra and matrix ranges.
A monic linear pencil is a free function of the form
(cid:88)
L(x) = LA(x) = I +
Ajxj,
where A ∈ B(H)d; in this paper we will concentrate on the case where A ∈ B(H)d
pencil L acts on a d-tuple X = (X1, . . . , Xd) in M d
sa. The
n(cid:88)
n by
Aj ⊗ Xj.
L(X) = I ⊗ In +
We write
where
j=1
DL = DLA = ∪nDL(n) = ∪nDLA(n),
DL(n) = {X = (Xj) ∈ (Mn)d
sa : L(X) ≥ 0}.
3
d(cid:88)
The set DL is said to be a free spectrahedron. Some authors use "free spectrahedron" for
pencils with matrix coefficients, but we allow operator coefficients. There is also a nonself-
adjoint version, but we do not discuss it in this paper.
We shall require the homogenization of a monic pencil LA, which gives a so called truly
linear pencil hLA in d + 1 variables, defined by
hLA(X1, . . . , Xd, Xd+1) =
Aj ⊗ Xj + I ⊗ Xd+1,
and its corresponding positivity set DhLA = {X ∈ Md+1
: hLA(X) ≥ 0} (see [24]).
The matrix range [3, Section 2.4] of a tuple A in B(H)d is defined to be the set
sa
j=1
W(A) = ∪nWn(A),
where
Wn(A) = {(φ(A1), . . . , φ(Ad)) : φ ∈ UCP(C∗(SA), Mn)}.
By [6, Proposition 3.5], a set S = ∪nSn ⊆ Md (respectively, in Md
sa) is a closed and bounded
matrix convex set if and only if S = W(A) for some A ∈ B(H)d (respectively, A ∈ B(H)d
sa).
Minimal and maximal matrix convex sets. Given a closed convex set K ⊆ Rd (or K ⊆ Cd),
one may ask whether it is equal to S1 for some matrix convex set S. The minimal matrix
convex set W min(K) and the maximal matrix convex set W max(K) for which
W min
1
(K) = W max
1
(K) = K
were described in several places; we follow the conventions of [6, Section 4] (see also [8], [12]
and [21], noting that analogous constructions in the theory of operator spaces have appeared
as far back as [19]). First, if we let
d(cid:88)
i=1
(cid:88)
(1.1) W max
n
(K) = {X ∈ (Mn)d
sa :
αiXi ≤ aIn whenever
αixi ≤ a for all x ∈ K},
n
W max
n
then W max(K) = ∪W max
(K) is clearly seen to be a closed matrix convex set, which satisfies
all the linear inequalities that determine K. Since satisfying the linear inequalities that
determine K is a necessary requirement for any matrix convex set that has ground level equal
to K, we see that W max(K) is indeed maximal. An alternative description for W max(K) is
(1.2)
(K) = {X ∈ (Mn)d
sa : W1(X) ⊆ K}.
The minimal matrix convex set W min(K) (called the matrix convex hull of K in [12])
clearly exists, as the intersection of all matrix convex sets containing K, but here is a more
useful description. A d-tuple X ∈ B(H1)d is said to be a compression of A ∈ B(H2)d if
there is an isometry V : H1 → H2 such that Xi = V ∗AiV for 1 ≤ i ≤ d. Conversely, A is
said to be a dilation of X if X is a compression of A. We will write X ≺ A when X is a
compression of A. A tuple N = (N1, . . . , Nd) will be called a normal tuple if N1, . . . , Nd are
normal commuting operators, and in this case, the unital C∗-algebra generated by the Ni is
commutative. In other words, if N is normal, C∗(SN ) is of the form
C(X) = {f : X → C : f is continuous}
4
for a unique compact Hausdorff space X, which is the maximal ideal space of C∗(SN ). The
joint spectrum of N = (N1, . . . , Nd) is then the compact set
σ(N ) := {(N1(x), . . . , Nd(x)) : x ∈ X} ⊆ Cd,
and if N acts on a finite dimensional space, σ(N ) is nothing but the finite set of all d-tuples
of joint eigenvalues. For a normal tuple N , the individual operators N1, . . . , Nd are all self-
adjoint if and only if the joint spectrum of N satisfies σ(N ) ⊆ Rd. If a normal tuple N
happens to consist of self-adjoint operators, we still call it a normal tuple, and avoid calling
it a self-adjoint tuple. In [6, Proposition 4.3], it was shown that
(1.3)
W min
n
(K) = {X ∈ (Mn)d
(cid:110)(cid:88)
sa : ∃N normal, s.t. X ≺ N and σ(N ) ⊆ K}.
(cid:111)
sa : x(j) ∈ K, Pj ≥ 0 and
(cid:88)
Pj = I
.
(1.4)
W min
n
(K) =
x(j) ⊗ Pj ∈ (Mn)d
In [8, Section 4], a slightly different version was given, which we write here as
(See also [21, Definition 3.8], with min and max reversed.)
In [8] this was shown to be
the minimal matrix convex set containing K, so this also defines W min(K). One can also
see directly - using Naimark's dilation theorem - that the definitions (1.3) and (1.4) are
equivalent.
1.3. Main results. At this point we can finally state our main results. In Section 3 we
introduce the dilation constants
θ(K, L) := inf{C > 0 : W max(K) ⊆ C · W min(L)},
θ(K) := θ(K, K),
θ(K) := inf{θ(T (x + K)) : x ∈ Rd, T ∈ GLd(R)}.
In particular, θ(K) is a shift invariant version of θ(K). A leitmotif in this paper is that
the numerical values of these constants are determined by the geometry of convex sets. To
this end, we first catalog the natural relationship between these constants and the familiar
Banach-Mazur distance ρ(K, L) between convex bodies K and L. Namely,
(1.5)
θ(K) ≤ ρ(K, L) θ(L),
2d−1(K) = W min
and a corresponding inequality holds for the shift invariant constants (Proposition 3.4).
When K is symmetric (K = −K), one would expect that the constants θ(K) and θ(K) are
equal. We show that this is indeed the case in Proposition 3.5.
Our first main result is that for a compact convex set K ⊂ Rd, θ(K) = 1 if and only if K
is a simplex, which holds if and only if W max
2d−1(K) (Theorem 4.1). This will be
used in Section 7 to improve the corresponding part of [8, Theorem 4.7], which is a result
of Fritz, Netzer and Thom. (After a translation from the setting of matrix convex sets over
cones to convex bodies, the corresponding part of [8, Theorem 4.7] says that for polyhedral
K, θ(K) = 1 if and only if K is a simplex, and the result does not give control over the matrix
level). Just as in [8], bounds on the Banach-Mazur distances between simplices and convex
bodies show that θ(K) ≤ d + 2 for any convex body K ⊂ Rd, and θ(K) ≤ d if K is also
symmetric (Corollary 4.4). This implies Corollary 4.6: if K ⊆ Rd is a symmetric convex set,
H is a Hilbert space, and A ∈ B(H)d
sa has W1(A) ⊆ K, then there exists a normal dilation
A ≺ N with σ(N ) ⊆ d· K. In contrast, the scale d was seen in [6] under conditions which did
not appear to capture all symmetric sets, but were also not limited to the symmetric setting.
5
In Theorem 4.5, we give an alternative proof of the fact that θ(K) ≤ d for symmetric K
using the techniques of [6].
The next two sections are devoted to studying a family of convex sets of special interest.
In Section 5 we completely resolve the dilation problem for the (cid:96)2-ball. We show that if
B1 = x + C1 · B2,d and B2 = y + C2 · B2,d are (cid:96)2-balls in Rd, then W max(B1) ⊆ W min(B2) if
and only if there is a d-simplex Π with B1 ⊆ Π ⊆ B2. This result appears in Theorem 5.6,
together with explicit conditions on the centers and radii of the balls:
W max(B1) ⊆ W min(B2) ⇐⇒ C2 ≥(cid:113)y − x2 + C 2
1 (d − 1)2 + C1.
Two special cases of the above are the computations θ(B2,d) = d and θ(x + B2,d) = ∞ for
x = 1, which were obtained in [6]. A key role in the analysis of the dilation problem for
the ball is played by a universal d-tuple F = (F1, . . . , Fd) of 2d−1×2d−1 self-adjoint, mutually
anti-commuting unitary matrices.
After solving the dilation problem for (cid:96)2-balls, we turn to treat the (cid:96)p-balls
Bp,d := {x ∈ Rd : (cid:107)x(cid:107)p ≤ 1},
as well as their positive sections
p,d := Bp,d ∩ [0, 1]d.
B+
Note that B∞,d = [−1, 1]d and that B+
1,d is a simplex. By the results in [6], it was known that
each of these sets has dilation constant at most d, but except for the (cid:96)2-ball B2,d, optimality
of the constant d was not claimed. In Theorems 6.4 and 6.9, we find that
and
√
θ(B+
p,d) = d1−1/p
θ(Bp,d) = d1−1/p−1/2,
1,d) = d1−1/p. How-
among several other sharp variants of these constants, such as θ(B+
√
ever, note that not all of these dilation constants are obtained from simplex containment.
For example, while θ([−1, 1]2) =
2]2.
The calculation of θ(Bp,d) for all p relies on the special cases p = 1, 2,∞. While the
constants θ(B1,d) and θ(B∞,d) must be equal by duality, we provide explicit dilation con-
structions that give rise to both constants (Theorems 6.6 and 6.7), and these constructions
also include positive parameters a1, . . . , ad. Namely, if D(a1, . . . , ad) denotes the convex hull
of (±a1, 0, . . . , 0), . . . , (0, . . . , 0,±ad), then
2, there is no simplex Π with [−1, 1]2 ⊆ Π ⊆ [−√
W max(B1,d) ⊆ W min(D(a1, . . . , ad)) ⇐⇒ W max(B∞,d) ⊆ W min(cid:16)(cid:89)
(cid:17) ⇐⇒(cid:88) 1
p,d, B+
2,
[−aj, aj]
≤ 1.
a2
j
In particular, every d-tuple of self-adjoint contractions can be dilated to a commuting d-tuple
d. The unit balls of the complex (cid:96)p spaces inside R2d
of self-adjoints with norm at most
are also briefly treated; see Corollary 6.11 and the surrounding discussion.
√
Remark 1.1. The constants we compute in this paper depend on the geometry of the convex
sets, and in particular, they depend on the number of variables. On the other hand, they are
independent of the rank of the operators to which one may apply them. In contrast, the lion's
share in [13] is dedicated to finding constants that are rank dependent, but independent of
the number of variables. By one of the main results of [13], any d self-adjoint n× n matrices
can be dilated to d commuting self-adjoints of norm less than or equal to ϑ(n), which is a
6
constant independent of d. It is known that ϑ(n) behaves like
to n, the constant ϑ(n) from [13] is better than θ([−1, 1]d) =
√
√
n, so when d is large relative
d, and inversely.
We were very much inspired by the paper [8] of Fritz, Netzer and Thom. In that paper,
the main objects of study were matrix convex sets living over cones, whereas we usually
consider convex bodies. In Section 7 we catalog a simple device for passing from one setting
to the other. We use it to import results or send some of our results to their setting, with
improvements and alternative proofs when possible. First, we consider the translation of the
problem of finite dimensional realizability treated in [8]. Importing results shows that for a
compact convex set K ⊂ Rd, W min(K) = W(A) for A ∈ Md
sa if and only if K is a polytope,
and if K is a polytope, then W max(K) = W(A) for A ∈ Md
sa if and only if K is a simplex.
However, for the statement regarding W min(K), we offer an alternative proof. In Corollary
7.6, we translate our Theorem 4.1 to show for a salient convex cone with nonempty interior
C ⊆ Rd, it holds that W min(C) = W max(C) if and only if C is a simplicial cone. This is an
improvement of the corresponding part of [8, Theorem 4.7] in that we drop the assumption
that C is polyhedral.
If K, L ⊆ Rd are compact convex sets and W max(K) ⊆ W min(L), then we say that L is a
dilation hull of K. In Section 8 we consider the collection of all dilation hulls of a fixed K,
and we search for minimal elements of this collection, which we call minimal dilation hulls.
First, we are able to show that minimal dilation hulls always exist in Proposition 8.1, and
that the diamond d · B1,d is a minimal dilation hull of the (cid:96)2-ball B2,d in Proposition 8.3.
Therefore, the dilation of an (cid:96)2-ball to an (cid:96)1-ball is not characterized by simplex containment,
in contrast to the case when both sets are (cid:96)2-balls. On the other hand, certain simplices
over B2,d are also minimal dilation hulls (Proposition 8.5), so the shape of minimal dilation
hulls is not unique. Finally, the simplex is a minimal dilation hull of "simplex-pointed sets"
(Theorem 8.8), and from this it follows that d1−1/p · B+
1,d is a minimal dilation hull of B+
(Corollary 8.9). Many natural questions remain.
p,d
2. Additional background material
2.1. Connection to completely positive maps. A major motivation to study matrix
convex sets is their relation to unital completely positive (UCP) maps between operator
systems. By [6, Theorem 5.1] (which essentially goes back to Arveson [3, Theorem 2.4.2]),
W(B) ⊆ W(A) if and only there exists a UCP map φ : SA → SB such that φ(Ai) = Bi for
all i. As a consequence, one obtains the following result (which has already been elaborated
on in [6, 13]).
Proposition 2.1. Let K1, K2 ⊆ Rd be compact convex sets, A ∈ B(H1)d
Suppose that W1(A) ⊇ K1 and W1(B) ⊆ K2.
sa, and B ∈ B(H2)d
sa.
(1) If W max(K2) ⊆ W min(K1), then there exists a UCP map φ : SA → SB such that
(2) If θ(K1) < ∞, K1 = W1(A), and there exists a unital positive map with Ai (cid:55)→ Bi,
C Bi (i = 1, . . . , d)
(3) If θ(K2) < ∞, K2 = W1(B), and there exists a unital positive map with Ai (cid:55)→ Bi,
C Bi (i = 1, . . . , d)
φ(Ai) = Bi for all i.
then for all C ≥ θ(K1), the unital map that sends I (cid:55)→ I and Ai (cid:55)→ 1
is completely positive.
then for all C ≥ θ(K2), the unital map that sends I (cid:55)→ I and Ai (cid:55)→ 1
is completely positive.
7
Proof. (1) follows from the above remarks together with the inclusions W(B) ⊆ W max(K2) ⊆
W min(K1) ⊆ W(A). For (2), note that if W1(A) = K1 and there is a unital positive map
Ai (cid:55)→ Bi, then W1(B) ⊆ W1(A) = K1. Therefore, for C ≥ θ(K1), there are containments
W(B) ⊆ W max(K1) ⊆ C · W min(K1) ⊆ W(C · A), and (2) follows. Likewise, if W1(B) = K2,
then the existence of a unital positive map gives K2 ⊆ W1(A). Therefore, for C ≥ θ(K2),
there are containments W(B) ⊆ W max(K2) ⊆ C · W min(K2) ⊆ W(C · A).
2.2. Polar duality. If S is an nc subset of Md
define the (matrix) polar dual of S to be the set S• = ∪nS•
sa then, following Effros and Winkler [7], we
n, where
S•
n = {X ∈ (Mn)d
sa :
Aj ⊗ Xj ≤ I for all A ∈ S}.
(cid:88)
(cid:88)
j
Recall that the usual (scalar) polar dual C(cid:48) of a set C ⊆ Rd is defined to be
C(cid:48) = {x ∈ Rd :
xjyj ≤ 1 for y ∈ C}.
By [6, Theorem 4.7],
and if 0 ∈ C, then
j
W min(C)• = W max(C(cid:48)),
W max(C)• = W min(C(cid:48)).
Lemma 2.2. Let 0 ∈ K ⊆ L be closed convex subsets of Rd. If W max(K) ⊆ W min(L), then
W max(L(cid:48)) ⊆ W min(K(cid:48)). In particular,
W max(K) ⊆ C · W min(K) =⇒ W max(K(cid:48)) ⊆ C · W min(K(cid:48)).
Proof. Applying the polar dual to W max(K) ⊆ W min(L) reverses the inclusion, and we
obtain
W max(L(cid:48)) = (W min(L))• ⊆ (W max(K))• = W min(K(cid:48)).
C · K(cid:48).
The final assertion follows from (C · K)(cid:48) = 1
2.3. Dilations for operators on finite and on infinite dimensional spaces. Let
K, L ⊆ Cd be compact convex sets. Consider the following three related - but not tauto-
logically equivalent - dilation problems:
(1) For every d-tuple of matrices A with W1(A) ⊆ K, find a d-tuple of normal commuting
operators N on some Hilbert space, such that A ≺ N and σ(N ) ⊆ L.
(2) For every d-tuple of matrices A with W1(A) ⊆ K, find a d-tuple of normal commuting
operators N , acting on a finite dimensional Hilbert space, such that A ≺ N and
σ(N ) ⊆ L.
(3) For every d-tuple of operators A with W1(A) ⊆ K, find a d-tuple of normal commut-
ing operators N on some Hilbert space, such that A ≺ N and σ(N ) ⊆ L.
It turns out that these problems are equivalent; thus, whenever we prove a dilation theorem
for matrices we obtain one for operators, and whenever we prove a dilation theorem with
operators, we obtain, in fact, a theorem that says that we can dilate matrices to matrices.
For the record, we state this fact here.
Proposition 2.3. Let K, L ⊆ Cd be compact convex sets. Then the following conditions are
equivalent:
8
(1) For every d-tuple A ∈ B(H)d with joint numerical range W1(A) ⊆ K, there is a nor-
mal d-tuple N of bounded operators acting on some Hilbert space, such that σ(N ) ⊆ L
and A ≺ N .
(2) For every d-tuple A ∈ Md with joint numerical range W1(A) ⊆ K, there is a normal
d-tuple N , acting on a finite dimensional Hilbert space, such that σ(N ) ⊆ L and
A ≺ N .
(3) W max(K) ⊆ W min(L).
Moreover, if A ∈ (Mn)d has a normal dilation A ≺ N such that σ(N ) ⊆ L, then there exists
a normal dilation A ≺ N of bounded operators, acting on some Hilbert space of dimension
at most 2n3(d + 1) + 1, such that σ( N ) ⊆ ext(L).
Proof. (1) ⇒ (2) follows from Theorem 7.1 of [6], which says that if there exists any normal
dilation A ≺ N for A ∈ (Mn)d, then there exists a normal dilation A ≺ N , where N is a
tuple of operators on a space of dimension 2n3(d + 1) + 1, and σ( N ) ⊆ σ(N ).
(2) ⇒ (3) is immediate from the characterizations (1.2) and (1.3).
Now for (3) ⇒ (1), fix A ∈ B(H)d such that W1(A) ⊆ K, and assume dim H = ∞,
as this is the only case requiring consideration. Let C(L) denote the space of continuous
complex-valued functions on L, and let Z = (Z1, . . . , Zd) ∈ C(L)d be the normal tuple of
coordinate functions on L. We first prove that there exists a UCP map φ : S(Z) → B(H)
that maps every Zi onto Ai. Let Pn be an increasing net of finite dimensional projections
weakly converging to the identity, and put A(n) = PnAPn, which we may view as an operator
on either H or PnH. By assumption, for all n there is a normal tuple N (n) with σ(N (n)) ⊆ L
such that A(n) ≺ N (n). By [6, Corollary 4.4], the matrix range of a normal tuple N is equal
to W min(σ(N )). Thus, W(N (n)) ⊆ W min(L) = W(Z).
It follows from [6, Theorem 5.1]
that there exists a UCP map S(Z) → S(N (n)) mapping Zi to N (n)
, and after compressing
we obtain a UCP map ψn : S(Z) → S(A(n)) that maps Zi to A(n)
for all i. By Arveson's
extension theorem, ψn extends to a UCP map ψn : C(L) → B(PnH). Enlarging the Hilbert
space gives a contractive and completely positive (CCP) map φn : C(L) → B(H) which
sends the constant function 1 to Pn and Zi to A(n)
for all i. Any bounded point-weak
limit point φ of the net {φn} in CCP(C(L),B(H)) is a UCP map that sends Zi to Ai for
N := π(Z) = (π(Z1), . . . , π(Zd)) is a normal dilation for A with σ(N ) ⊆ σ(Z) = L. This
completes the proof of (3) ⇒ (1).
We now prove the final claim. Given A ∈ (Mn)d and a normal dilation N ∈ B(K)d
with σ(N ) ⊆ L, we note that as above, there is a UCP map φ : C(L) → B(K) that sends
Zi to Ni, where Z1, . . . , Zd are the coordinate functions in C(L). The operator system
S(Z) generated by Z1, . . . , Zd ∈ C(L) is completely isometrically isomorphic to the operator
system S(W ) generated by the coordinate functions W1, . . . , Wd ∈ C(ext(L)). Thus, there
is a UCP map ψ : S(W ) → S(Z) mapping Wi to Zi. This gives rise to a UCP map
θ = φ ◦ ψ : S(W ) → B(K) mapping Wi to Ni. Applying Arveson extension and then
all i. Finally, we let π : C(L) → B((cid:101)H) be the Stinespring dilation of φ, so the tuple
Stinespring dilation to θ, we obtain a unital ∗-representation π : C(ext(L)) → B((cid:101)K) with
i
i
i
Ni = θ(Wi) = PKπ(Wi)PK. Setting Mi = π(Wi) shows that there is a normal dilation M of
N , and hence of A, with σ(M ) ⊆ ext(L). Finally, we apply [6, Theorem 7.1] to the normal
dilation A ≺ M to obtain a normal dilation A ≺ N , where N acts on a space of dimension
at most 2n3(d + 1) + 1 and σ( N ) ⊆ σ(M ) ⊆ ext(L).
9
We will also have need of the following reformulation of [6, Theorem 7.11]. Recall that a
set K ⊆ Rd is a convex body if it is compact, convex, and has nonempty interior.
Theorem 2.4. Let K and L be convex bodies in Rd. Assume that there is a k-tuple of real
d × d rank one matrices λ such that Id ∈ conv{λ(1), . . . , λ(k)} and such that λ(m)K ⊆ L for
all 1 ≤ m ≤ k. Then W max(K) ⊆ W min(L). That is, for every X such that W1(X) ⊆ K,
there is a d-tuple T = (T1, . . . , Td) of self-adjoint matrices such that
(1) {T1, . . . , Td} is a commuting family of operators,
(2) σ(T ) ⊆ L,
(3) X ≺ T .
It is worth noting that, given X ∈ B(H)d
construction of the normal dilation T acting on H ⊗ Ck.
Remark 2.5. Applying the above theorem to B1,d and B∞,d, respectively, one obtains The-
orems 10.6 and and 10.10 from [6], which show that W max(B1,d) ⊆ W min(B∞,d) and that
W max(B∞,d) ⊆ d · W min(B1,d), respectively.
sa, the proof of this theorem provides a concrete
3. Change of variables and convex geometry
3.1. Affine change of variables. For A = (A1, . . . , Ad) and M ∈ Mm,d(C), let M A be the
tuple B = (B1, . . . , Bm) given by
d(cid:88)
Bi =
MijAj.
Likewise, given y ∈ Cm, we consider the affine transformation F (x) = M x + y and its action
on tuples
j=1
F (A) = M A + y,
where y is shorthand for (y1I, . . . ymI).
Now fix A ∈ B(H)d and put B = F (A), where F is an affine map F (x) = M x + y as
above. Then C∗(SB) ⊆ C∗(SA), so every φ ∈ UCP(C∗(SA), Mn) restricts to a UCP map
C∗(SB) → Mn. Also, every φ ∈ UCP(C∗(SB), Mn) extends to a UCP map C∗(SA) → Mn
by Arveson's extension theorem. If X = φ(A) ∈ W(A), then
F (X) = M φ(A) + y = φ(M A + y) = φ(B) ∈ W(B).
On the other hand, if X = φ(B) then
X = φ(M A + y) = M φ(A) + y = F (φ(A)).
We conclude
(3.1)
Lemma 3.1. Let F : Rd → Rd be an affine and invertible map and let n ∈ N. If W max
W min
W max(K) ⊆ C · W min(K), then W max(F (K)) ⊆ C · W min(F (K)).
(K) ⊆
(F (L)). In particular, if F is linear and invertible with
(F (K)) ⊆ W min
(L), then W max
n
n
n
n
W(B) = F (W(A)).
10
Proof. Let X ∈ W max
n
(F (K)). This means that W1(X) ⊆ F (K). Then by (3.1),
W1(F −1(X)) = F −1(W1(X)) ⊆ K,
n
so Y := F −1(X) ∈ W max
But then F (N ) is a normal dilation for F (Y ) = X, and σ(F (N )) ⊆ F (L).
(K). By assumption, Y has a normal dilation N with σ(N ) ⊆ L.
Let H be an affine subspace in Rd and let K be a closed convex subset in Rd, so H ∩ K
is a closed and convex set. If the dimension of H is k, then H can be identified with Rk.
One can also define the projection onto H, in the direction orthogonal to H; we let PH(K)
denote the projection of K under this orthogonal projection.
Lemma 3.2. Let K, L ⊆ Rd be closed convex sets, and let H be an affine subspace of Rd. If
W max(K) ⊆ W min(L), then
W max(H ∩ K) ⊆ W min(PH(L)).
Proof. By Lemma 3.1, we may assume that
H = {(x1, . . . , xk, 0, . . . , 0) : (x1, . . . , xk) ∈ Rk}.
Assume that X ∈ W max(H ∩ K). Then W1(X) ⊆ H ∩ K ⊆ K, and necessarily X has the
form X = (X1, . . . , Xk, 0, . . . , 0). By assumption, there is a dilation X ≺ Y = (Y1, . . . , Yd)
consisting of commuting self-adjoints such that σ(Y ) ⊆ L. But then
Y (cid:48) := (Y1, . . . , Yk, 0, . . . , 0) = PH(Y )
is a dilation of X, and σ(Y ) ⊆ PH(L).
Remark 3.3. Note that nothing changes in the above result if we think of X and Y (cid:48) as
k-tuples, rather than d-tuples padded with 0s.
3.2. Convex Geometry. If K and L are closed convex sets in Rd, then there are two
established notions of distance which compare K and L. First,
(3.2)
ρ(K, L) := inf{C > 0 : ∃T ∈ GLd(R), K ⊆ T (L) ⊆ C · K}.
Note that ρ(K, L) = ρ(L, K). Even if K and L have interior, it is still possible that ρ(K, L) =
∞ (which is understood to mean that there are no C, T such that K ⊆ T (L) ⊆ C · K).
However, if K and L are compact with 0 in the interior of both sets, then ρ(K, L) is finite,
and if K and L are also symmetric with respect to the origin, then ρ(K, L) is just the
Banach-Mazur distance of the corresponding norms on Rd. Second,
(3.3) ρ(K, L) := inf{C > 0 : ∃x, y ∈ Rd, T ∈ GLd(R), y + K ⊆ T (x + L) ⊆ C · (y + K)}.
Note that ρ(K, L) = ρ(L, K).
Recall that a convex body is a compact convex subset of Rd with nonempty interior, and
that every point in a convex body may be approximated by interior points. For any convex
bodies K and L, ρ(K, L) is finite, as we can shift the interiors of K and L to include the
origin. Moreover, if B2,d denotes the closed unit ball of real (cid:96)2
d space, then it is known that
(3.4)
K ⊆ Rd is a convex body =⇒ ρ(K, B2,d) ≤ d,
and in [17] Leichtweiss showed that the simplex is the unique maximizer (see also [18]):
(3.5)
K ⊆ Rd is a convex body with ρ(K, B2,d) = d ⇐⇒ K is a d-simplex.
Here and below, a k-simplex is the convex hull of k + 1 affinely independent points.
11
The distances ρ and ρ may also be tied to (not necessarily symmetric!) scaling constants
that arise naturally in dilation theory:
(3.6)
(3.7)
(3.8)
θ(K, L) := inf{C > 0 : W max(K) ⊆ C · W min(L)},
θ(K) := θ(K, K),
θ(K) := inf{θ(T (x + K)) : x ∈ Rd, T ∈ GLd(R)}.
One can use Proposition 2.3 to show that if K and L are compact and convex, then these
infima are actually obtained. It was also shown in Lemma 3.1 and the preceding discussion
that an invertible affine transformation factors through W max and W min.
T ∈ GLd(R), x ∈ Rd =⇒
W max(x + T (K)) = x + T (W max(K))
W min(x + T (K)) = x + T (W min(K))
It follows that θ is invariant under invertible linear transformations.
T ∈ GLd(R) =⇒ θ(K, L) = θ(T (K), T (L))
θ(K) = θ(T (K))
Note, however, there is no reason to believe that θ is invariant under invertible affine trans-
formations, as translations and linear maps do not generally commute. In fact, there exist
numerous counterexamples to affine invariance of θ in [6] and this manuscript. Applying
reductions to θ yields
θ(K) = inf{θ(a + K) : a ∈ Rd}.
Similarly, since
W max(a + K) ⊆ C · W min(a + K) ⇐⇒ W max(K) ⊆ W min((C − 1)a + C · K),
it also holds that
(3.9)
θ(K) = inf{C > 0 : ∃b ∈ Rd,W max(K) ⊆ W min(b + C · K)}.
The following proposition gives a way of estimating the constants θ(K) and θ(K), when
K can be approximated by a set L for which we know the dilation constants.
Proposition 3.4. Let K and L be closed convex subsets of Rd. Then
θ(K) ≤ ρ(K, L) θ(L)
and
θ(K) ≤ ρ(K, L) θ(L).
Proof. We may certainly assume that θ(L), θ(L), ρ(K, L), and ρ(K, L) are finite.
Let C1 > θ(L) and let C2 > 0, T ∈ GLd(R) be such that K ⊆ T (L) ⊆ C2 · K. Since
θ(T (L)) = θ(L), it follows that
W max(K) ⊆ W max(T (L)) ⊆ C1 · W min(T (L)) ⊆ C1C2 · W min(K),
which shows that θ(K) ≤ C1C2. Taking an infimum over all applicable C1 and C2 yields
θ(K) ≤ ρ(K, L) θ(L).
12
Similarly, following the reduction in (3.9), let C1 > 0, b ∈ Rd be such that
W max(L) ⊆ W min(b + C1 · L) = b + C1 · W min(L)
and choose x, y ∈ Rd, C2 > 0 with
y + L ⊆ T (x + K) ⊆ C2 · (y + L).
Then it follows that
W max(T (x + K)) ⊆ C2 · W max(y + L)
= C2y + C2 · W max(L)
⊆ C2y + C2 · (b + C1 · W min(L))
= C2y + C2b − C2C1y + C2C1 · W min(y + L)
⊆ C2y + C2b − C2C1y + C2C1 · W min(T (x + K))
= C2y + C2b − C2C1y + C2C1T (x) + C2C1 · W min(T (K)).
Setting z := C2y + C2b − C2C1y + C2C1T (x) and rearranging the given containment
W max(T (x + K)) ⊆ z + C2C1 · W min(T (K)) shows that
W max(K) ⊆ x + T −1(z) + C2C1 · W min(K),
so by (3.9), θ(K) ≤ C2C1. Taking an infimum over C2 and C1 gives θ(K) ≤ ρ(K, L) θ(L).
Letting
Bp,d = {x ∈ Rd : xp ≤ 1}
d space, we know from [6, Example 7.24] that θ(B2,d) = d.
denote the closed unit ball of real (cid:96)p
Since the Banach-Mazur distance between (cid:96)p norms is governed by ρ(Bp,d, Bq,d) ≤ d1/p−1/q,
the estimate
θ(Bp,d) ≥
d1/p−1/2 = d1−1/p−1/2
holds from setting q = 2 and applying Proposition 3.4. Note that the inequality ρ(Bp,d, Bq,d) ≤
d1/p−1/q follows from examining the identity map from real (cid:96)p
d space, and it
is generally not an equality, as seen in Chapter 1, Section 8 of [15]. However, such compli-
cations only arise when p < 2 < q or q < 2 < p, and we will not need to use this case to
estimate dilation constants. Theorem 6.9 shows that θ(Bp,d) is exactly d1−1/p−1/2.
d space to real (cid:96)q
d
The distance ρ of (3.2) is calculated using many different invertible linear transformations
T . Alternatively, one may use only scaling maps, and from calculations similar to those in
Proposition 3.4, it follows that
α · K1 ⊆ K2,
L2 ⊆ β · L1
=⇒ θ(K1, L1) ≤ β
α
· θ(K2, L2).
(3.10)
Proposition 3.5. Let K ⊆ Rd be a closed convex set with K = −K. Then θ(K) = θ(K).
Proof. The main idea is reminiscent of the proof of [8, Theorem 5.8]. First, rewrite (3.9) as
θ(K) = inf{C > 0 : ∃b ∈ Rd,W max(K + b) ⊆ W min(C · K)}.
Now let C > θ(K) and b ∈ Rd be such that W max(K + b) ⊆ W min(C · K). Since K = −K,
W max(K − b) ⊆ W min(C · K).
13
Next, let X ∈ W max(K). Then X ± b ∈ W max(K ± b), so by the previous paragraph,
X ± b ∈ W min(C · K). By convexity of W min(C · K), we find that X = 1
2(X − b) ∈
W min(C · K). Thus θ(K) ≤ C, and taking an infimum gives θ(K) ≤ θ(K). The reverse
inequality is trivial, so we are done.
2(X + b) + 1
4. The simplex is the only convex body that determines a unique matrix
convex set
It is natural to ask, for what compact convex sets K ⊆ Rd does it hold that
(4.1)
W max(K) = W min(K)?
We shall now prove that simplices are the unique compact convex sets such that (4.1)
In Section 7, we see that this may be used to improve the analogous uniqueness
holds.
statement [8, Corollary 5.3] by removing a polyhedral assumption.
Theorem 4.1. Let K ⊆ Rd be a compact convex set. Then the following are equivalent.
(1) K is a simplex.
(2) W max(K) = W min(K).
(3) W max
2d−1(K).
2d−1(K) = W min
Proof. If K is contained in a proper affine subspace H, then H ∩ K = PH(K). By Lemma
3.2 and Remark 3.3, we may assume, without loss of generality, that K is not contained in
a proper subspace, and therefore it has non-empty interior.
To show that (1) implies (2), we may first apply Lemma 3.1 and assume that K is the stan-
dard simplex span{0, e1, . . . , ed}. Now, every X ∈ W max(K) is just a tuple X = (X1, . . . , Xd)
sure on {1, . . . , d}. By Naimark's theorem, X can be dilated to a spectral measure P on
{1, . . . , d}, which is just a tuple P = (P1, . . . , Pd) of mutually orthogonal (hence commuting)
projections dilating X (see also Theorem 6.1 for a more explicit dilation).
with Xi ≥ 0 and(cid:80) Xi ≤ I, and thus can be considered as a positive operator valued mea-
Since (2) clearly implies (3), it remains to prove that W max
2d−1(K) implies that
K is a simplex. Suppose that (3) holds, and that K is a convex body (so it has non-empty
interior) in Rd that is not a simplex. From (3.5), the fact that K is not a simplex implies
that ρ(K, B2,d) < d. This means that after applying an invertible affine transformation, we
may assume that B2,d ⊆ K ⊆ C · B2,d for some C ∈ (0, d). By Lemma 3.1, the equation
W max
2d−1(K) = W min
2d−1(K) = W min
W max
2d−1(K) = W min
2d−1(K) is not affected by an affine transformation. Thus,
2d−1(C · B2,d).
2d−1(B2,d) ⊆ W max
2d−1(K) ⊆ W min
(cid:80) viXi ≤ I for all v ∈ B2,d, then X has a normal dilation N with spectrum σ(N ) ⊆ C · Bd.
By definition of W max and W min, for every d-tuple of self-adjoint 2d−1 × 2d−1 matrices X, if
But this contradicts [6, Example 7.24], which shows that there is a d-tuple X ∈ W max
2d−1(Bd)
that has no such normal dilation (see also Theorem 5.6 below). This contradiction shows
that K must be a simplex.
Theorem 4.1 also implies the following.
Corollary 4.2. Suppose that K, L ⊂ Rd are compact convex sets with W max
for at least one n ≥ 2d−1. Then K = L is a simplex.
n
(K) = W min
n
(L)
14
Proof. Given x ∈ L, we know that X :=
(K). Therefore X satisfies
the linear inequalities that characterize K, which proves x ∈ K by restricting to the diagonal,
so we may conclude L ⊆ K. Similarly, the claim that W max
(L)
may be used to show that K ⊆ L, so we have that K = L.
(K) = W min
(L) ⊆ W max
(L) = W max
x ∈ W min
i=1
n
n
n
n
n
(K) for some n ≥ 2d−1. Fix T ∈ W max
2d−1(K), so that for any
Now, W max
x ∈ K, S = T ⊕ n−2d−1(cid:76)
(K) = W min
n
n
x ∈ W max
(K) = W min
n
n
(K), meaning S has a normal dilation with
i=1
joint spectrum in K. Certainly this also applies to T , and T ∈ W min
2d−1(K) = W min
W max
Of course, with any changes in the minimum matrix dimension 2d−1 used in the statement
2d−1(K), we may conclude that K is a simplex using Theorem 4.1.
2d−1(K). Since this proves
of Theorem 4.1, Corollary 4.2 changes as well.
Problem 4.3. Is there some n < 2d−1 such that if K ⊂ Rd is compact and convex with
W max
(K), then K is a simplex?
(K) = W min
n
n
In Section 8 we will see that there is a large family C of convex bodies, which we call
"simplex-pointed sets," for which it suffices to check up to level two:
K ∈ C, W max
2
(K) = W min
2
(K) =⇒ K is a simplex.
We also recover the following reformulations of Theorems 5.7 and 5.8 from [8]. As in [8],
a crucial part of the proof is the application of scaling results from convex geometry.
Corollary 4.4. If K ⊆ Rd is a convex body, then θ(K) ≤ d + 2. If, in addition, K is
symmetric with respect to the origin (i.e., K = −K), then θ(K) ≤ d.
Proof. Let ∆ be a nondegenerate simplex in Rd. By [16], ρ(K, ∆) ≤ d + 2. Moreover, if
K = −K, then by [9, Corollary 5.8], ρ(K, ∆) ≤ d. The proof is completed by invoking
Theorem 4.1, together with Propositions 3.4 and 3.5.
sa satifies W1(X) ⊆ K, then there exist k ≤ d(d+1)
We also give an alternative proof of the fact that θ(K) ≤ d for symmetric sets, using
the techniques of [6].
In [6], the dilation scale of d was seen under conditions which did
not appear to capture all symmetric sets, but which were also not limited to the symmetric
setting.
Theorem 4.5. If K ⊆ Rd is a symmetric convex body, then θ(K) ≤ d. In fact, if X ∈
B(H)d
2 + 1 points x1, . . . , xk ∈ K and a
normal dilation X ≺ T acting on H ⊗ Ck such that σ(T ) ⊆ d · conv{x1, . . . , xk}.
Proof. We may assume that K is in John position, meaning that the ellipsoid of maximal
volume contained in K is the unit ball (see [1, Section 2.1]). Indeed, if K is not in John
position then it may be moved to John position by a linear transformation (K is symmetric,
so no translation is required). By [1, Theorem 2.1.10], there exist k ≤ d(d+1)
∂K ∪ ∂Bd, and positive numbers c1, . . . , cm such that(cid:80)
2 + 1, x1, . . . , xk ∈
m cm = d and
n(cid:76)
k(cid:88)
m=1
15
Id =
cmxm ⊗ xm,
where xm ⊗ xm denotes the rank one operator xm ⊗ xm(x) = (cid:104)x, xm(cid:105)xm. (This result goes
back to F. John, and it appears in passing in [14, Section 4]).
Now set λ(m) = dxm ⊗ xm, so Id ∈ conv{λ(1), . . . , λ(k)}. Moreover, the fact that Bd ⊆ K
implies that xm +{xm}⊥ is a supporting hyperplane for K for all m; in other words λ(m)K ⊆
In fact, λ(m)K ⊆ L for all m, where L = d · conv{x1, . . . , xk}.
dK for all m = 1, . . . , k.
Applying Theorem 2.4, we obtain, for every W1(X) ⊆ K, the normal dilation T as claimed.
From the dilation-theoretic characterization of W min, we get the following dilation result.
Corollary 4.6. Let K ⊆ Rd be a symmetric compact convex set and H a Hilbert space. If
A ∈ B(H)d
sa has W1(A) ⊆ K, then there exists a normal dilation A ≺ N with σ(N ) ⊆ d · K.
The constant d cannot be improved in general, because the ball B2,d is symmetric but
satisfies θ(B2,d) = d, as seen in [6]. On the other hand, we will see in the remaining sections
that for certain symmetric sets, the constant can be significantly improved.
5. The Euclidean ball
In this section, we solve the dilation problem for the ball: if K and L are (cid:96)2-balls in Rd
(perhaps with different centers and radii), then W max(K) ⊆ W min(L) if and only if there is
a d-simplex Π with K ⊆ Π ⊆ L. This information is catalogued, along with a numerical
estimate, in Theorem 5.6. Crucial to this pursuit is a modification of [6, Lemma 7.23], which
considers the self-adjoint unitary 2d−1 × 2d−1 matrices F [d]
d , defined as follows.
1 , . . . , F [d]
⊗ F [d−1]
1 ≤ j ≤ d − 1,
,
j
(5.1)
F [d]
j =
F [d]
d =
F [1]
1 = 1
(cid:19)
(cid:18)0 1
(cid:18)1
1 0
(cid:19)
0
0 −1
⊗ I2d−2.
When the number of matrices d is not needed as a superscript (i.e., whenever induction in
d is not needed), we shall denote the matrices as F1, . . . , Fd.
It was shown in [6, Lemma 7.23] that the tuple (F1, . . . , Fd) is in W max(B2,d), and also that
F1 ⊗ F1 + . . . + Fd ⊗ Fd has d as an eigenvalue. From the construction, it is evident that the
self-adjoint matrices F1, . . . , Fd pairwise anti-commute and are unitary, though these facts
were not emphasized. Consider also the universal unital C∗-algebra
A := C∗(x1, . . . , xd xj = x∗
j = 1, xixj = −xjxi for i (cid:54)= j)
j , x2
generated by pairwise anti-commuting self-adjoint unitaries. The relations on the generators
of A impose that A is spanned by 1 and by monomial terms xi1xi2 . . . xij where i1 < i2 <
. . . < ij. Thus the vector space dimension of A is at most
= 2d. By definition,
there is a unital C∗-homomorphism φ : A → C∗(F1, . . . , Fd) defined by φ : xj (cid:55)→ Fj, but from
direct examination (and inductive proof) we see that for d ≥ 2, the vector space dimension
of C∗(F1, . . . , Fd) is also 2d. This means the kernel of φ is trivial, so
(5.2) C∗(F1, . . . , Fd) ∼= C∗(x1, . . . , xd xj = x∗
j = 1, xixj = −xjxi for i (cid:54)= j) for d ≥ 2.
d(cid:80)
k=0
k
(cid:18) d
(cid:19)
j , x2
16
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1 , . . . , F [d]
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥(cid:112)y2 + (d − 1)2 + 1.
(cid:18)0 1
(cid:19)
Below we modify the inductive step of [6, Lemma 7.23] to include a parameter y ∈ Rd,
which will help us to strengthen the dilation estimates. Note, however, that much like (5.2),
this lemma must start at d = 2.
Lemma 5.1. Fix d ≥ 2. Then the tuple (F1, . . . , Fd) = (F [d]
d ) of pairwise anti-
commuting, self-adjoint, unitary matrices of dimension 2d−1 × 2d−1 defined by (5.1) has the
property that for any y ∈ Rd,
(Fj − yj) ⊗ Fj
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d(cid:80)
j=1
Proof. First, we prove the case (y1, . . . , yd) = (0, . . . , 0, a) inductively in d. Now, [6, Lemma
has (d − 1) as an eigenvalue, with eigenvector
7.23] shows that T :=
⊗ F [d−1]
F [d−1]
denoted xt. (Note that the case d − 1 = 1 is included here.) Since F [d]
for 1 ≤ j ≤ d − 1 and F [d]
⊗ I2d−2, it follows that
j =
⊗ F [d−1]
j
1 0
d−1(cid:80)
j=1
j
(cid:18)1
j
(cid:19)
0
0 −1
d =
(1 − a)I
d−1(cid:88)
has an eigenvalue (cid:112)a2 + (d − 1)2 + 1 with eigenvector
d − aI) ⊗ F [d]
j ⊗ F [d]
F [d]
j + (F [d]
d =
0
0
T
j=1
0
T
0
(−1 − a)I
0
T
(−1 + a)I
0
(cid:32)(cid:112)a2 + (d − 1)2 − a
d − 1
T
0
0
(1 + a)I
(cid:33)t
· x, 0, 0, x
.
Therefore the theorem holds for y of the form (0, . . . , 0, a).
Next, consider any y ∈ Rd \ {0}, d ≥ 2. There is an orthogonal matrix U ∈ Od(R)
1y y, and we may apply the transformation U to (F1, . . . , Fd) to produce
whose final row is
a tuple (H1, . . . , Hd). A simple calculation shows that because U is orthogonal, the matrices
H1, . . . , Hd pairwise anti-commute and are self-adjoint unitaries. Moreover, the vector space
dimension of C∗(H1, . . . , Hd) is still 2d, so just as in (5.2), H1, . . . , Hd is universal for these
relations, and there is a unital C∗-homomorphism φ : Hj (cid:55)→ Fj. It follows that φ ⊗ φ is a
contraction, so
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d−1(cid:88)
j=1
(5.3)
Hj ⊗ Hj + (Hd − y) ⊗ Hd
Fj ⊗ Fj + (Fd − y) ⊗ Fd
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d−1(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≥(cid:112)y2 + (d − 1)2 + 1.
j=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Finally, because the last row of U is
1
y y, we know that by definition,
d(cid:88)
j=1
Hd =
1
y yj · Fj,
17
and because U is orthogonal, we also know that
d(cid:88)
d(cid:88)
Hj ⊗ Hj =
Fj ⊗ Fj.
j=1
j=1
Therefore, it follows that
d−1(cid:88)
Hj ⊗ Hj + (Hd − y) ⊗ Hd =
(Fj − yj) ⊗ Fj,
j=1
so by (5.3), we have reached the estimate
(Fj − yj) ⊗ Fj
d(cid:88)
j=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d(cid:80)
j=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥(cid:112)y2 + (d − 1)2 + 1.
Since F1, . . . , Fd are anti-commuting self-adjoint unitaries such that F1⊗ F1 + . . . + Fd⊗ Fd
has an eigenvalue d, it follows from the diagonalization of the normal tuple N = (F1 ⊗
F1, . . . , Fd⊗Fd) and the triangle inequality that (1, . . . , 1) is in the joint spectrum of N . This
claim may certainly be strengthened to include other tuples with ±1 entries by examining
conjugations, but we will not need it. From
(1, . . . , 1) ∈ σ(F1 ⊗ F1, . . . , Fd ⊗ Fd)
it follows that for t1, . . . , td > 0,
(5.4)
t1F1 ⊗ F1 + . . . + tdFd ⊗ Fd = t1 + . . . + td.
Moreover, the claim in [6] that
(5.5)
(F1, . . . , Fd) ∈ W max(B2,d)
may be viewed as a direct result of pairwise anti-commutation.
Lemma 5.2. If x1, . . . , xn are self-adjoint, pairwise anti-commuting elements of a C∗-
algebra, then it follows that
Proof. Apply the C∗-norm identity A =(cid:112)AA∗ and cancel xixj + xjxi = 0 for i (cid:54)= j:
x1 + . . . + xn =(cid:112)(x1 + . . . + xn)2 =
n ≤(cid:112)x12 + . . . + xn2.
n ≤(cid:112)x12 + . . . + xn2.
(cid:113)x2
x1 + . . . + xn =
(cid:113)x2
1 + . . . + x2
1 + . . . + x2
Indeed, (F1, . . . , Fd) ∈ W max(B2,d) follows easily, as whenever (λ1, . . . , λd) ∈ B2,d, the
inequality −1 ≤ λ1F1 + . . . + λdFd ≤ 1 holds due to anti-commutation and Lemma 5.2. The
properties of the tuple (F1, . . . , Fd) are sufficient to characterize exactly when the maximal
matrix convex set over a ball is contained in the minimal matrix convex set over another
ball. First, we obtain a numerical estimate.
Lemma 5.3. Let C1, C2 > 0, x, y ∈ Rd be such that x+C1·(F1, . . . , Fd) ∈ W min
where F1, . . . , Fd are as in Lemma 5.1. Then C2 ≥(cid:112)y − x2 + C 2
2d−1(y+C2·B2,d),
1 (d − 1)2 + C1.
18
2d−1(y + C2 · B2,d). We will show that C2 ≥(cid:112)y2 + (d − 1)2 + 1.
Proof. After translation and scaling, it suffices to consider x = 0, C1 = 1, and (F1, . . . , Fd) ∈
W min
If there is a normal dilation (N1, . . . , Nd) of (F1, . . . , Fd) with joint spectrum in y +C2·B2,d,
then by Lemma 5.2,
(N1 − y1) ⊗ F1 + . . . + (Nd − yd) ⊗ Fd ≤ C2.
On the other hand, by Lemma 5.1,
(F1 − y1) ⊗ F1 + . . . + (Fd − yd) ⊗ Fd ≥(cid:112)y2 + (d − 1)2 + 1.
Comparing these norms shows that C2 ≥(cid:112)y2 + (d − 1)2 + 1.
We also show that if C2 ≥ (cid:112)y − x2 + C 2
1 (d − 1)2 + C1, then there is a simplex in
between x + C1 · B2,d and y + C2 · B2,d. The following lemma provides an estimate that will
help prove this claim.
Lemma 5.4. Suppose b > 1, K ⊆ Rd−1 is compact and convex, and Π ⊆ Rd is the convex
hull of K × {−1} and (0, . . . , 0, b). Then
B2,d ⊆ Π ⇐⇒
b + 1
b − 1
· B2,d−1 ⊆ K
Proof. Parametrize Π by its final coordinate xd ∈ [−1, b], Π(xd) = {y ∈ Rd−1 : (y, xd) ∈ Π}.
Then B2,d ⊆ Π holds if and only if for each xd ∈ [−1, 1],(cid:112)1 − x2
d · B2,d−1 ⊆ Π(xd). If
C(xd) = max{C ≥ 0 : C · B2,d−1 ⊆ Π(xd)},
then C(xd) is linear, with C(b) = 0. Setting α = C(−1), we have C(xd) =
Therefore, B2,d ⊆ Π if and only if for each xd ∈ [−1, 1],
≥(cid:112)1 − x2
α(b − xd)
d, or rather,
α(b − xd)
.
b + 1
(cid:115)
(cid:26) b + 1
(cid:113)
b − xd
(cid:27)
b + 1
1 − x2
d
.
α ≥ max
xd∈[−1,1]
(cid:113)
(cid:115)
b
b2 =
1 − 1
The maximum occurs at the critical point xd = 1/b, so the inequality is equivalent to
α ≥ b + 1
b + 1
. Since α = C(−1) is exactly the maximum scale of B2,d−1
b − 1
b − 1
sitting inside K, we are done.
Lemma 5.5. Let x, y ∈ Rd, C1 > 0, and C2 =(cid:112)y − x2 + C 2
C := C2 = (cid:112)y − x2 + (d − 1)2 + 1. We will produce a simplex Π with B2,d ⊂ Π ⊂
1 (d − 1)2 + C1. Then there
is a simplex Π with x + C1 · B2,d ⊂ Π ⊂ y + C2 · B2,d.
Proof. The result is trivial for d = 1, so we assume d ≥ 2. After a rotation, translation,
and scaling, we may suppose that x = (0, . . . , 0), y = (0, . . . , 0, a) for a > 0, C1 = 1, and
(0, . . . , 0, a) + C · B2,d.
d−1 · B2,d−1 ⊂ ∆ ⊂
r · B2,d−1, as an application of (3.5) in dimension d− 1. We wish to find r so that the convex
hull
For any r > d − 1, let ∆ = ∆(r) ⊆ Rd−1 be a (d − 1)-simplex with r
Π = Π(r) = Conv(∆ × {−1}, (0, . . . , 0, a + C)),
19
which is itself a d-simplex, satisfies
B2,d ⊆ Π ⊆ (0, . . . , 0, a) + C · B2,d.
For the claim Π ⊆ (0, . . . , 0, a) + C · B2,d, we must have
(a + 1)2 + r2 ≤ C 2.
On the other hand, for the claim B2,d ⊆ Π, Lemma 5.4 demands
(cid:115)
r
d − 1
≥
(a + C) + 1
(a + C) − 1
.
The two inequalities can be satisfied simultaneously if and only if
≤ C 2,
(a + 1)2 + (d − 1)2 · a + C + 1
a + C − 1
which is equivalent to
(d − 1)2 · a + C + 1
a + C − 1
≤ C 2 − (a + 1)2.
Canceling linear terms (which are positive for C > 1) gives the equivalent form
(d − 1)2 ≤ (C − (a + 1))(a + C − 1) = (C − 1)2 − a2,
or
Since this inequality is satisfied by definition, we know B2,d ⊂ Π ⊂ (0, . . . , 0, a) + C · B2,d.
(C − 1)2 ≥ a2 + (d − 1)2.
Finally, we have reached a solution of the dilation problem for (cid:96)2-balls.
Theorem 5.6. Fix C1, C2 > 0, x, y ∈ Rd, and let F1, . . . , Fd be as in Lemma 5.1. Then the
following are equivalent.
(1) W max(x + C1 · B2,d) ⊆ W min(y + C2 · B2,d)
2d−1(x + C1 · B2,d) ⊆ W min
2d−1(y + C2 · B2,d)
(2) W max
(3) x + C1 · (F1, . . . , Fd) ∈ W min
2d−1(y + C2 · B2,d)
(5) There is a simplex Π with x + C1 · B2,d ⊂ Π ⊂ y + C2 · B2,d
(4) C2 ≥(cid:112)y − x2 + C 2
1 (d − 1)2 + C1
Proof. First, (1) =⇒ (2) and (2) =⇒ (3) are trivial, (3) =⇒ (4) is exactly Lemma 5.3,
and (4) =⇒ (5) is exactly Lemma 5.5. Finally, (5) =⇒ (1) holds because any simplex Π
has W max(Π) = W min(Π), as in Theorem 4.1.
This allows us to compute the dilation constant θ(·) for any (cid:96)2-ball, generalizing [6, Ex-
amples 7.22 and 7.24].
Corollary 5.7. Fix d ≥ 2, A > 0, x ∈ Rd, and consider the ball x + A · B2,d. Then
θ(x + A · B2,d) =
∞,
(cid:112)A2 − x2
A(d − 1)
x ≥ A
x < A
+ 1,
and consequently θ(B2,d) = θ(B2,d) = d.
20
that is,
x2
A2 + (d − 1)2.
(5.6)
(cid:19)
Since C ≥ 1, squaring and solving (5.6) yields the equivalent form
(C − 1)2
C − 1 ≥
≥ (d − 1)2.
(C − 1)2
1 − x2
A2
(cid:115)
(cid:18)
W max(cid:0)x + A · B2,d
W min(cid:0)Cx + CA · B2,d
Proof. Consider C ≥ 1. By definition, the claim θ(x + A · B2,d) ≤ C holds if and only if
(cid:1) ⊆ W min(cid:0)Cx + CA · B2,d
(cid:1) if and only if
CA ≥(cid:112)Cx − x2 + A2(d − 1)2 + A,
(cid:1). From Theorem 5.6, W max(cid:0)x + A · B2,d
(cid:1) ⊆
+ 1.
For finite C, this is impossible unless x < A, in which case we reach the desired estimate
(cid:112)A2 − x2
C ≥ A(d − 1)
Remark 5.8. If K is a compact convex set and 0 /∈ K, then θ(K) = ∞ unless K is a simplex.
Indeed, if K is not a simplex, then θ(K) > 1 by Theorem 4.1, but the assumption 0 /∈ K
produces an issue at the ground level: K (cid:42) C · K for C ∈ (1,∞), and thus W max(K) (cid:42)
C · W min(K) for C ∈ (1,∞). This gives θ(x + A · B2,d) = ∞ for (cid:107)x(cid:107) > A, as in the
previous corollary. For (cid:107)x(cid:107) = A, there is also a "positive" interpretation for the value
θ(x + A · B2,d) = ∞. Consider the closed half-space
H = {y ∈ Rd : (cid:104)x, y(cid:105) ≥ 0} = ∪C≥1C · (x + A · B2,d)
as the infinite inflation of x + A · B2,d, i.e. "H = ∞ · W min(x + A · B2,d)." The positive
interpretation of θ(x + A · B2,d) = ∞ is then
Indeed, one can always find a simplex Π with x + A · B2,d ⊂ Π ⊂ H.
W max(x + A · B2,d) ⊆ W min(H).
As in [6], there are also consequences for the existence of UCP maps.
Corollary 5.9. Fix d ≥ 2, let F1, . . . , Fd ∈ M2d−1(C) be as in Lemma 5.1, and fix C > 0,
y ∈ Rd. Equip Sd−1 = ∂B2,d with any positive measure of full support, and let Mx1, . . . , Mxd ∈
B(L2(Sd−1)) be the multiplication operators for the coordinate functions xj.
• There is a UCP map φ : B(L2(Sd−1)) → M2d−1(C) with
if and only if C ≥(cid:112)y2 + (d − 1)2 + 1.
φ : Mxj (cid:55)→ 1
C
Fj − yj
C
I
• There is a unital positive (not necessarily UCP) map ψ : span{1, Mx1, . . . , Mxd} →
M2d−1(C) with
if and only if C ≥ y + 1.
ψ : Mxj (cid:55)→ 1
C
Fj − yj
C
I
21
Proof. Because the chosen measure on Sd−1 has full support, we know that (Mx1, . . . , Mxd)
has joint spectrum equal to Sd−1, and since
holds for a linear map if and only if
(5.7)
Mxj (cid:55)→ 1
C
Fj − yj
C
I
CMxj + yj I (cid:55)→ Fj,
If C ≥(cid:112)y2 + (d − 1)2 + 1, then by Theorem 5.6 we know that W max(B2,d) ⊆ W min(y +
2d−1(y + C · B2,d), so by Theorem 5.6, C ≥(cid:112)y2 + (d − 1)2 + 1, as desired.
we will use the second formulation. The joint spectrum of (CMx1 + y1I, . . . , CMxd + ydI) is
y + C ·Sd−1, so, as for any normal tuple, the matrix range of this normal tuple is the minimal
matrix convex set containing the joint spectrum, i.e. W min(y + C · B2,d).
C·B2,d), and therefore W(F1, . . . , Fd) ⊆ W min(y+C·B2,d) = W(CMx1 +y1I, . . . , CMxd +ydI).
By [6, Theorem 5.1], the UCP map (5.7) exists.
On the other hand, suppose the UCP map (5.7) exists. Using [6, Theorem 5.1] in the other
direction, we obtain that W(F1, . . . , Fd) ⊆ W min(y + C · B2,d). In particular, (F1, . . . , Fd) ∈
W min
Finally, if the map CMxj + yj I (cid:55)→ Fj need only be unital and positive, on the domain
span{1, Mx1, . . . , Mxd}, we may repeat the above argument with the matrix range W replaced
by the numerical range W1. Since W1(CMx1 + y1 I, . . . , CMxd + yd I) = y + C · B2,d and
W1(F1, . . . , Fd) = B2,d, the positive unital map exists if and only if B2,d ⊆ y + C · B2,d. That
is, C ≥ y + 1.
Remark 5.10. Note that the maps φ and ψ in Corollary 5.9 are defined on different domains.
By Arveson's extension theorem, any UCP map φ : span{1, Mx1, . . . , Mxd} → M2d−1(C)
extends to a UCP map φ : B(L2(Sd−1)) → M2d−1(C). However, there is no corresponding
extension theorem for unital positive maps. If y + 1 ≤ C <(cid:112)y2 + (d − 1)2 + 1, then
it is actually guaranteed that the unital positive map ψ in Corollary 5.9 cannot extend to
a unital positive map ψ : C∗(Mx1, . . . , Mxd) → M2d−1(C).
In particular, if the extension
exists, then it is a unital positive map whose domain is a commutative C*-algebra, so the
extension is actually UCP. Applying Arveson extension to the domain B(L2(Sd−1)) would
then contradict the first part of Corollary 5.9.
Given a convex body K, if 0 is an interior point of K, then we have an estimate for θ(K).
Namely, since it is impossible to dilate an ellipse E centered at the origin with smaller scale
than θ(E) = θ(B2,d) = d, we know from Proposition 3.4 that d = θ(B2,d) ≤ ρ(B2,d, K) θ(K).
That is, θ(K) ≥
occurs, we may claim that the
optimal dilation scale among all translations of K is found when the translation is zero.
Corollary 5.11. Let K ⊂ Rd be a convex body with 0 in the interior. Then it follows that
θ(K) ≥
, then θ(K) = θ(K).
. In particular, if θ(K) =
. If equality θ(K) =
ρ(B2,d, K)
d
ρ(B2,d, K)
d
d
ρ(B2,d, K)
d
ρ(B2,d, K)
Proof. By Proposition 3.4 and Corollary 5.7,
ρ(B2,d, K) θ(K) ≥ θ(B2,d) = θ(B2,d) = d,
22
that is,
θ(K) ≥
d
ρ(B2,d, K)
≥
d
.
ρ(B2,d, K)
The computation of Banach-Mazur distance
(5.8)
ρ(B2,d, K) = inf{C > 0 : ∃T ∈ GLd(R), K ⊆ T (B2,d) ⊆ C · K}
includes many possible transformations T . In certain circumstances, the infimum may be
attained for a scaling map T = aI, so the ellipse that most closely resembles K is a ball. Since
the ball is orthogonally invariant, we may then extend Corollary 5.11 to include orthogonal
images of K in addition to translations.
Corollary 5.12. Let K ⊆ Rd be a convex body with 0 in the interior, such that the infimum
of (5.8) is attained for a scaling map T = aI. Then for any U ∈ Od(R) and y ∈ Rd,
θ(K, y + U (K)) ≥
d
.
ρ(B2,d, K)
Proof. Let α · B2,d ⊆ K ⊆ β · B2,d, where α, β > 0 are such that ρ(B2,d, K) =
choice exists due to the assumption that the infimum of (5.8) is attained at a scaling map.
Using orthogonal invariance of the ball, we reach the two separate containments α·B2,d ⊆ K
and y + U (K) ⊆ β ·
. From (3.10) and Theorem 5.6, we find that
(cid:18) 1
. Such a
· y + B2,d
(cid:19)
β
α
≤ β
α
· θ(K, y + U (K)) = ρ(B2,d, K) · θ(K, y + U (K)),
(cid:18)
(cid:19)
β
· y + B2,d
d ≤ θ
as desired.
B2,d,
1
β
As above, if the dilation constant θ(K) is equal to the estimate
d
ρ(B2,d, K)
, then under the
assumptions of Corollary 5.12 (i.e. that the Banach-Mazur distance between K and the ball
is attained using a scaling map), we may conclude that the best dilation of K to a translated,
scaled, orthogonal image of K occurs when the translation is zero. We will see in Section 6
that if K is the unit ball of (cid:96)p-space on Rd or on Cd ∼= R2d, then θ(K) is indeed equal to the
estimate
d
ρ(B2,d, K)
or
2d
ρ(B2,2d, K)
, respectively.
Problem 5.13. Let K ⊆ Rd be a convex body with 0 in the interior. Under what circum-
stances must θ(K) =
? Is it sufficient to assume that K = −K?
d
ρ(B2,d, K)
6. Sharp constants for (cid:96)p balls and their positive sections
Let
p,d := Bp,d ∩ [0, 1]d
B+
23
denote the positive section of the closed unit ball in real (cid:96)p
d space. The methods of [6] (e.g.,
Proposition 8.1 there) show that the balls and positive sections admit dilations governed by
W max(Bp,d) ⊆ d · W min(Bp,d)
and
so θ(Bp,d) ≤ d and θ(B+
since B+
1,d is a simplex,
(6.1)
and hence θ(B+
W max(B+
p,d) ⊆ d · W min(B+
p,d),
p,d) ≤ d. In this section we find sharp constants for all p. Note that
W max(B+
1,d) = W min(B+
1,d),
1,d) = 1 (see Theorem 4.1).
We begin with an explicit dilation for (6.1), which may be found by following the first
part of the proof of [22, Theorem 2.1]. We record this with details for completeness.
Theorem 6.1. Suppose that (X1, . . . , Xd) ∈ B(H)d
(cid:80)d
i=1 Xi ≤ I. Then there exists a normal dilation (P1, . . . , Pd) ∈ B(H ⊗ Cd+1)d
sa is a tuple of positive operators with
sa such that
the Pi are mutually orthogonal projections. Moreover, the Pi may be chosen with block entries
in the real unital C∗-algebra generated by X1, . . . , Xd.
Proof. Consider the row operator
Xd
,
which satisfies T T ∗ ≤ IH and consequently T ∗T ≤ IH⊗Cd. Let
T :=
X2
(cid:105)
(cid:112)
(cid:104)(cid:112)
(cid:112)
∆ :=(cid:112)IH⊗Cd − T ∗T ,
X1
. . .
so that when ∆2 is viewed as a d × d block matrix with entries in B(H),
i (cid:54)= j =⇒ (∆2)ij = −(cid:112)
Xi
(cid:112)Xj.
(∆2)ii = IH − Xi
Define another row operator
and
(cid:104)(cid:112)
(cid:105)
,
so that
ViV ∗
i = Xi +
Vi :=
. . . ∆id
∆ik∆∗
ik = Xi +
Xi ∆i1 ∆i2
(cid:88)
i Vi is a projection in B(H ⊗ Cd+1), and examination of the top left
(cid:88)
(cid:112)Xj + (∆2)ij = 0
(cid:88)
(cid:112)Xj +
∆ik∆ki = Xi + (∆2)ii = IH.
(cid:88)
∆ik∆∗
∆ik∆kj =
(cid:112)
(cid:112)
jk =
Xi
Xi
k
k
It follows that Pi := V ∗
block entry shows that Pi dilates Xi. For i (cid:54)= j, the identity
(cid:112)
(cid:112)Xj +
ViV ∗
j =
Xi
k
implies that the projections are mutually orthogonal.
k
Remark 6.2. Recall that it is possible to dilate any contraction S to a unitary
(cid:18)
U :=
√
S
I − S∗S
24
(cid:19)
,
√
I − SS∗
−S∗
as in [10]. In fact, S is not required to be a square operator. If this dilation is applied to
for Xi as in Theorem 6.1, then the first d block rows of U are the same
S = T ∗ =
√
X1
...√
Xd
as the row operators Vi obtained in the proof.
Theorem 6.1 demonstrates that (6.1) may be realized with dilations that annihilate each
other, which in general is much stronger than commutation. This is not surprising, as distinct
extreme points of B+
1,d never have matching nonzero entries (recall that by Proposition 2.3, the
spectrum of a normal dilation can be pushed to the extreme points). On the other hand, the
following lemma shows that we may reach annihilating dilations of certain tuples (X1, . . . , Xd)
of positive operators in a very specific way, which will aid us in later computations.
sa be a tuple of positive operators such that if i (cid:54)= j,
Lemma 6.3. Let (X1, . . . , Xd) ∈ B(H)d
then the only operator T with 0 ≤ T ≤ Xi and 0 ≤ T ≤ Xj is T = 0. Suppose that
(Y1, . . . , Yd) ∈ B(K)d
sa is a normal dilation of (X1, . . . , Xd) such that Yi ≥ 0 for each i. Then
there is a normal dilation (Z1, . . . , Zd) ∈ B(K)d
sa of (X1, . . . , Xd) which satisfies 0 ≤ Zi ≤ Yi,
ZiZj = 0 for i (cid:54)= j, and σ(Z1, . . . , Zd) ⊆ σ(Y1, . . . , Yd) ∪ {0}.
Proof. Since Y1, . . . , Yd are positive, bounded, and commute, by the spectral theorem there is
a positive, bounded spectral measure F and a collection of nonnegative measurable functions
gi such that
Let V : H → K be an isometric embedding satisfying Xi = V ∗YiV . The operator-valued
measure E(·) = V ∗F (·)V , which is positive and bounded but not necessarily spectral, satisfies
(cid:82) 1
Fix n ≥ 1, i (cid:54)= j, and let Si,j,n = {t : gi(t) > 1/n and gj(t) > 1/n}, so that 1
n E(Si,j,n) =
n ISi,j,ndE is a positive operator that is bounded above by both Xi and Xj. Therefore, by
the assumptions about Xi and Xj, E(Si,j,n) = 0, and from a countable union we see that
S := {t : ∃i (cid:54)= j with gi(t) > 0 and gj(t) > 0} also has E(S) = 0. We are then free to change
the value of gi on S:
(cid:90)
(cid:90)
Yi =
gi dF.
Xi =
gi dE.
(cid:90)
(cid:90)
gi · ISc dE.
gi · ISc dF
It follows that
Xi =
Zi :=
are the desired positive dilations of Xi. The Zi annihilate each other because the functions
gi·ISc are never simultaneously nonzero, and the Zi are bounded above by Yi since gi·ISc ≤ gi.
Finally,
σ(Z1, . . . , Zd) = EssRan(g1·ISc, . . . , gd·ISc) ⊆ EssRan(g1, . . . , gd)∪{0} = σ(Y1, . . . , Yd)∪{0}.
25
d ball B+
1,d) ⊆ d1−1/p · W min(B+
Any d self-adjoint operators Z1, . . . , Zd with Zi ≤ M and ZiZj = 0 for all i and j (cid:54)= i
automatically satisfy Z1 + . . . + Zd ≤ M , so Lemma 6.3 may be used to relate dilation of
positive sets to the positive diamond.
Theorem 6.4. For all p ∈ [1,∞] and all d ∈ Z+, the positive (cid:96)p
p,d satisfies
W max(B+
p,d) ⊆ d1−1/p · W min(B+
p,d).
p,d, B+
p,d) = θ(B+
sa of positive operators with W1(X) ⊆ B+
1,d) = d1−1/p. Moreover, given a tuple
p,d, there is a normal dilation
sa with Ni = d1−1/pPi, where P1, . . . , Pd are mutually orthogonal
(6.2)
The scale d1−1/p is optimal: θ(B+
(X1, . . . , Xd) ∈ B(H)d
(N1, . . . , Nd) ∈ B(H ⊗ Cd+1)d
projections with block entries in the real unital C∗-algebra generated by X1, . . . , Xd.
Proof. Suppose (X1, . . . , Xd) ∈ B(H)d
p,d. Since this set is contained
in d1−1/p · B+
1,d, it follows that the positive operators Xi satisfy X1 + . . . + Xd ≤ d1−1/pI.
From Theorem 6.1, there is a normal dilation (N1, . . . , Nd) of the form Ni = d1−1/pPi, where
P1, . . . , Pd are mutually orthogonal projections in B(H ⊗ Cd+1) with block entries in the
real unital C∗-algebra generated by X1, . . . , Xd. For tuples of matrices, the above dilation
procedure is the explicit witness to the computation
sa has W1(X) ⊆ B+
W max(B+
p,d) ⊆ d1−1/p · W max(B+
1,d) = d1−1/p · W min(B+
1,d) ⊆ d1−1/p · W min(B+
p,d),
p,d) ⊆ C · W min(B+
which establishes both desired containments.
For optimality of the constant d1−1/p, suppose that W max(B+
0 < ε < 1/d and let Pj ∈ M2(C) be the projection onto the span of ((cid:112)1 − j2ε2, jε), so that
p,d). Fix
the top left entry of Pj is 1−j2ε2 and the Pj have ranges with pairwise trivial intersection. It
follows that the only positive semidefinite matrix bounded above by two distinct Pi and Pj is
zero. Define Xj = d−1/pPj, so that 0 ≤ Xj ≤ d−1/pI2. Since (d−1/p, . . . , d−1/p) has (cid:96)p norm 1,
it follows that (X1, . . . , Xd) ∈ W max(B+
p,d), there is then
an isometry V : C2 (cid:44)→ H and a normal dilation (Y1, . . . , Yd) ∈ B(H)d
sa of positive operators
with Yj ≤ C. By Lemma 6.3, there is a different normal dilation (Z1, . . . , Zd) ∈ B(H)d
of positive operators which annihilate each other and have Zj ≤ C. The joint spectrum
sa
of (Z1, . . . , Zd) is then contained in C · B+
p,d). From W max(B+
p,d) ⊆ C · W min(B+
1,d, and
Z1 + . . . + Zd ≤ C.
From applying V ∗(·)V we see that
X1 + . . . + Xd ≤ C,
and from the top-left entry of Xj we reach
d(cid:88)
d−1/p(1 − j2ε2) ≤ C.
j=1
Taking ε → 0+ gives d · d−1/p = d1−1/p ≤ C. This shows that θ(B+
have d1−1/p ≥ θ(B+
p,d) ≥ d1−1/p, the proof is complete.
1,d) ≥ θ(B+
p,d, B+
p,d) ≥ d1−1/p, and since we
26
That is, the dilation constant for B+
p,d is the highest (cid:96)1 norm attained in that set. This
computation will be refined and generalized in Section 8. We may also use the explicit
dilation of the positive diamond B+
1,d found in Theorem 6.1 to form an explicit dilation of
the full diamond, improving the estimate obtained in [6]. Moreover, these estimates may
also be phrased in terms of the self-adjoint matrix ball
B(d) = {X ∈ Md
d ≤ I}
sa : X 2
1 + . . . + X 2
sa has W1(Y ) ⊆ B1,d, then(cid:80) Y 2
after a quick computation.
Lemma 6.5. If Y = (Y1, . . . , Yd) ∈ B(H)d
lar, it holds that W max(B1,d) ⊆ B(d).
Proof. Assume W1(Y ) ⊆ B1,d. Then for all σ ∈ {0, 1}d,
i ≤ I. In particu-
−I ≤ (−1)σ1Y1 + (−1)σ2Y2 + . . . + (−1)σdYd ≤ I.
Squaring this inequality (which is valid due to the C∗-norm identity) for a fixed σ leads to
0 ≤ Y 2
1 + . . . + Y 2
d +
(−1)σj +σkYjYk ≤ I,
(cid:88)
j(cid:54)=k
and averaging over all choices of σ cancels out all cross terms, so
0 ≤ Y 2
1 + . . . + Y 2
d ≤ I.
Consequently, W max(B1,d) ⊆ B(d).
In [6], shortly before Corollary 9.9, it is shown that B(d) ⊆ √
d·W min(B2,d). The following
theorem improves upon this result (in particular, by using the minimal matrix convex set
over a diamond instead of a ball) and contains the optimal dilation scale for B1,d as well.
Theorem 6.6. For a list of positive numbers a1, . . . , ad, let D(a1, . . . , ad) be the convex hull
of (±a1, 0, . . . , 0), . . . , (0, . . . , 0,±ad). Then
(6.3)
W max(B1,d) ⊆ W min(D(a1, . . . , ad))
holds if and only if(cid:80) 1
≤ 1. In fact, if(cid:80) 1
≤ 1, then the stronger claim
a2
i
a2
i
B(d) ⊆ W min(D(a1, . . . , ad))
sa satisfies (cid:80) X 2
(6.4)
holds, and if (X1, . . . , Xd) ∈ B(H)d
i ≤ I, then there is a normal dilation
(N1, . . . , Nd) ∈ B(H ⊗ C2d+1)d
sa with joint spectrum in D(a1, . . . , ad). Moreover, we may
insist that the block entries of Nj are in the real unital C∗-algebra generated by X1, . . . , Xd
and that NiNj = 0 for i (cid:54)= j.
Proof. Suppose W max(B1,d) ⊆ W min(D(a1, . . . , ad)), and let F1, . . . , Fd be as in Lemma 5.1.
If a normal tuple (N1, . . . , Nd) has joint spectrum in D(a1, . . . , ad), then
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
a2
1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1
· N 2
1 + . . . +
27
· N 2
d
1
a2
d
(cid:18) 1
a1
(cid:19)
Nd
1
ad
follows because
N1, . . . ,
is a normal tuple with joint spectrum in B1,d ⊆ B2,d.
Now, since F1, . . . , Fd pairwise anti-commute, we also see from Lemma 5.2 that
· N1 ⊗ F1 + . . . +
· Nd ⊗ Fd
1
ad
· N1 ⊗ F1 + . . . +
(6.5)
· N 2
1 ⊗ I + . . . +
· N 2
d ⊗ I
1 + . . . +
Applying Lemma 5.2 to the tuple
F1, . . . ,
shows that
1
a2
j
(cid:18) 1
(cid:118)(cid:117)(cid:117)(cid:116) d(cid:88)
a1
1
a2
j
j=1
(cid:19)
∈
Recalling (5.4) and (6.5), it follows that
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
a2
1
a1
(cid:115)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
(cid:18) 1
F1, . . . ,
a1
so
(cid:18) 1
d(cid:88)
a1
1
a2
j
=
j=1
F1, . . . ,
1
ad
(cid:18) 1
a1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
a1
Fd
1
ad
Fd
(cid:19)
(cid:19)
(cid:118)(cid:117)(cid:117)(cid:116) d(cid:88)
F1
j=1
(cid:19)2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
1
ad
· Nd ⊗ Fd
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1.
· N 2
d
1
a2
d
(cid:118)(cid:117)(cid:117)(cid:116)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
1
ad
a1
Fd
a2
1
· N 2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18) 1
(cid:115)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
(cid:19)
(cid:118)(cid:117)(cid:117)(cid:116) d(cid:88)
(cid:115) d(cid:80)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
Nd ⊗ Fd
1
a2
j
1
a2
j
1
ad
a1
j=1
j=1
⊗ Fd
· W max(B1,d) ⊆
· W min(D(a1, . . . , ad)),
has a normal dilation M =
· N , σ(N ) ⊆ D(a1, . . . , ad).
(cid:19)
Fd
(cid:18) 1
ad
1
ad
N1 ⊗ F1 + . . . +
⊗ F1 + . . . +
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1
a1
·
1
a2
j
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =
Md ⊗ Fd
1
ad
M1 ⊗ F1 + . . . +
(cid:118)(cid:117)(cid:117)(cid:116) d(cid:88)
j=1
1
a2
j
.
That is,
+ . . . +
1
a2
1
≤ 1.
1
a2
d
For the other direction, consider any tuple (X1, . . . , Xd) ∈ B(H)d
d = X12 + . . . + Xd2 ≤ I.
1 + . . . + X 2
0 ≤ X 2
sa which satisfies
First, applying the positive linear functional (cid:104)(·)v, v(cid:105) for arbitrary v ∈ H shows that
(6.6)
0 ≤ X1 v2 + . . . + Xd v2 ≤ v2.
Given any list a1, . . . , ad of positive numbers, the Cauchy-Schwarz Inequality and (6.6) give
that
(cid:115)
(6.7)
0 ≤ 1
a1
· X1 v + . . . +
· Xd v ≤
1
ad
1
a2
1
+ . . . +
· v.
1
a2
d
From the Cauchy-Schwarz Inequality once more we also know that
0 ≤ (cid:104)Xjv, v(cid:105) ≤ Xj v · v,
28
which combined with (6.7) gives that
(cid:28) 1
a1
(cid:29)
(cid:28) 1
ad
0 ≤
· X1v, v
+ . . . +
· Xdv, v
(cid:115)
≤
1
a2
1
+ . . . +
· v2.
1
a2
d
Since v is arbitrary, this implies
(cid:29)
(cid:115)
(6.8)
· Xd ≤
0 ≤ 1
a1
· X1 + . . . +
If (cid:80) 1
≤ 1, then we have reached (cid:80) 1
1
a2
1
In this case, write Xj = Aj − Bj,
Xj = Aj + Bj, where Aj and Bj are positive and determined from the functional calculus
(using real coefficients). We then know that
Xj ≤ I.
+ . . . +
1
a2
d
· I.
1
ad
a2
j
aj
· A1 +
· Ad +
(cid:18) 1
· Bd ≤ I.
0 ≤ 1
a1
· B1 + . . . +
1
a1
· A1,
1
ad
1
· B1, . . . ,
From Theorem 6.1, the tuple
ad
to a tuple (K1, L1, . . . , Kd, Ld) ∈ B(H ⊗ C2d+1)2d
sa of positive contractions which pairwise
annihilate each other. It follows that Nj = aj · (Kj − Lj) is a self-adjoint dilation of Xj
such that NiNj = 0 for i (cid:54)= j and Nj ≤ aj. The joint spectrum of (N1, . . . , Nd) is then in
D(a1, . . . , ad). Finally, tracing the steps used shows that the block entries of the Nj are in
the real unital C∗-algebra generated by X1, . . . , Xd. With Lemma 6.5, this explicit dilation
1
ad
1
· Bd
ad
sa may be dilated
∈ B(H)2d
· Ad,
(cid:19)
1
a1
a1
≤ 1, then W max(B1,d) ⊆ B(d) ⊆ W min(D(a1, . . . , ad)).
procedure shows that if(cid:80) 1
a2
j
√
d, we obtain
From the constant sequence a1 = . . . = ad =
√
d · W min(B1,d),
√
d. It follows that the dilation scale
W max(B1,d) ⊆
and this constant is optimal. That is, θ(B1,d) =
also optimal in the containment
√
This result improves the claim that B(d) ⊆ √
Consider now the dual problem of dilating the cube B∞,d = [−1, 1]d. First, using duality
d · W min(B2,d) from [6].
d · W min(B1,d).
√
d is
B(d) ⊆
(Lemma 2.2), we immediately obtain
W max([−1, 1]d) ⊆ W min(cid:0)[−aj, aj]d(cid:1) ⇐⇒ 1
+ . . . +
≤ 1,
1
a2
d
a2
1
and in particular
W max([−1, 1]d) ⊆
√
d · W min([−1, 1]d),
√
with θ([−1, 1]d) =
d optimal. So, we know that given a d-tuple X of self-adjoint contrac-
√
tions, we should be able to find a d-tuple N of commuting self-adjoints each of norm at
d, such that X ≺ N . Our next goal is to find such a dilation explicitly; this will be
most
29
achieved in Theorem 6.7 below. A crucial step in this process is the dilation technique of
Halmos in [10]; if a self-adjoint operator X has X ≤ 1, then
(6.9)
Y :=
(cid:18) X
√
I − X 2
(cid:19)
√
I − X 2
−X
is a self-adjoint dilation of X which has Y 2 = I. That is, Y is self-adjoint and unitary. The
block entries of Y also belong to the real unital C∗-algebra generated by X.
Theorem 6.7. Let a1, . . . , ad be a list of positive numbers. Then
W max([−1, 1]d) ⊆ W min
[−aj, aj]
(cid:33)
(cid:32) d(cid:89)
j=1
(6.10)
holds if and only if(cid:80) 1
a2
j
≤ 1. In this case, if (X1, . . . , Xd) ∈ B(H)d
sa is such that Xj ≤ 1
for each j, then there is a normal dilation (N1, . . . , Nd) ∈ B(H ⊗ C4d−1)d
sa with Nj ≤ aj
for each j. The Nj may be chosen with block entries in the real unital C∗-algebra generated
by X1, . . . , Xd.
Proof. First, suppose that W max([−1, 1]d) ⊆ W min(cid:16)(cid:81)d
(cid:112)t2
holds. From Lemma
5.1, (5.4), and (5.5), the tuple (F1, . . . , Fd) ∈ W max([−1, 1]d) is such that for all t1, . . . , td > 0,
t1F1 ⊗ F1 + . . . + tdFd ⊗ Fd = t1 + . . . + td. On the other hand, if (M1, . . . , Md) is a normal
tuple of self-adjoints with Mj ≤ aj, then by Lemma 5.2, t1M1 ⊗ F1 + . . . + tdMd ⊗ Fd ≤
(cid:17)
j=1[−aj, aj]
d. Therefore we must have
1 + . . . + t2
da2
1a2
(cid:113)
t2
1a2
1 + . . . + t2
d ≥ t1 + . . . + td.
da2
Plugging in tj =
1
a2
j
yields (cid:115)
1
a2
1
+ . . . +
1
a2
d
≥ 1
a2
1
+ . . . +
1
a2
d
,
that is,
+ . . . +
1
a2
1
≤ 1.
1
a2
d
that
d(cid:80)
We will show the converse dilation procedure inductively, noting that d = 1 is trivial.
Therefore, we may suppose the theorem holds for all lists b1, . . . , bd of positive numbers such
≤ 1. Since
≤ 1. Suppose a1, . . . , ad+1 is a list of positive numbers with
ad+1 > 1, write ad+1 =(cid:112)1 + 1/s2 for s > 0, and define bj = aj√
(cid:19)
1+s2 . Then
(cid:19)
(cid:18)
(cid:18)
1
a2
j
1
b2
j
d+1(cid:80)
j=1
j=1
1
b2
1
+. . .+
= (1+s2)·
1
b2
d
+ . . . +
1
a2
d
≤ (1+s2)·
1 − 1
a2
d+1
= (1+s2)·
1 −
1
1 + 1/s2
(cid:19)
= 1.
(cid:18) 1
a2
1
30
d(cid:80)
Let X1, . . . , Xd+1 ∈ B(H) be self-adjoint with Xj ≤ 1. Apply the Halmos dilation (6.9):
Y1, . . . , Yd+1 ∈ B(H ⊗ C2) are self-adjoint dilations with Y 2
j = I. The block operators
Yj
s
2
s
2
Zj =
· (YjYd+1 − Yd+1Yj)
· (Yd+1Yj − YjYd+1)
· I
· I −Yd+1
are self-adjoint operators in B(H ⊗ C4) such that Zj commutes with Zd+1 for each j. More-
over, the block diagonal and off-diagonal entries of Zj anti-commute, so the norms of Zj are
then bounded by Lemma 5.2, which gives
Zd+1 =
1
s
1
s
Yj
, 1 ≤ j ≤ d,
Yd+1
Zj ≤
Yj2 +
· (Yd+1Yj − YjYd+1)
1 ≤ j ≤ d,
(cid:114)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) s
(cid:115)
2
Zd+1 ≤
Yd+12 +
· I
√
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ≤
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ≤(cid:112)1 + 1/s2 = ad+1.
1 + s2,
√
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1
s
j=1
1
1 + s2) shows that (Z1, . . . , Zd) ∈
≤ 1, the inductive assumption (scaled by
Since
b2
j
sa admits a normal dilation (N1, . . . , Nd) ∈ B(H ⊗ C4 ⊗ C4d−1)d
sa with Nj ≤
B(H ⊗ C4)d
√
4d−1(cid:76)
1 + s2 · bj = aj such that the block entries of each Nj are in the real unital C∗-algebra
generated by Z1, . . . , Zd. The Nj therefore also commute with Nd+1 :=
Zd+1, which
has Nd+1 = Zd+1 ≤ ad+1. Tracing back the steps used shows that the block en-
tries of N1, . . . , Nd+1 are in the real unital C∗-algebra generated by the original self-adjoints
X1, . . . , Xd+1, so the induction is complete. Applied to tuples of matrices, the above dilation
≤ 1, then W max([−1, 1]d) ⊆ W min ((cid:81)[−aj, aj]).
procedure shows that if(cid:80)d
m=1
1
a2
j
j=1
√
As before, the constant sequence a1 = . . . = ad =
W max([−1, 1]d) ⊆
(6.11)
and this constant is optimal. That is, θ(B∞,d) =
Remark 6.8. If one takes the containment of Theorem 6.7,
√
d.
d gives
√
d · W min([−1, 1]d),
(cid:33)
[−aj, aj]
(cid:32) d(cid:89)
⇐⇒ (cid:88) 1
⊆ W min([−1, 1]d) ⇐⇒ (cid:88) 1
xj, the result is
a2
j
j=1
aj
(cid:21)(cid:33)
≤ 1,
≤ 1.
and applies a scaling transformation xj (cid:55)→ 1
W max([−1, 1]d) ⊆ W min
W max
(cid:32) d(cid:89)
(cid:104)− 1
j=1
(cid:20)
(cid:105)
− 1
aj
,
1
aj
(6.12)
Since each set(cid:81)d
≤ 1, it is very
tempting to replace the left hand side of (6.12) with W max(B2,d) in order to construct and
, 1
aj
j=1
a2
j
aj
is contained in B2,d due to the inequality(cid:80) 1
a2
j
31
prove a more general result. However, if d ≥ 2, it holds that
W max(B2,d) (cid:54)⊆ W min([−1, 1]d),
(6.13)
as [−1, 1]d ⊆ √
We have now computed θ(B1,d) =
d · B2,d and θ(B2,d) = d >
√
√
d.
√
d = θ(B∞,d), where we note that the proof of op-
d essentially returns to the claim θ(B2,d) = d from [6, Example
timality of the constant
7.24]. Indeed, we may use the explicit dilations of the diamond and the cube, as well as the
minimal dilation constant of the ball, to interpolate θ(Bp,d) for all 1 ≤ p ≤ ∞.
Theorem 6.9. For 1 ≤ p ≤ ∞, θ(Bp,d) = d1−1/p−1/2.
Proof. It is known from Theorems 6.6 and 6.7 that θ(B∞,d) =
d = θ(B1,d), and the
Banach-Mazur distance between (cid:96)p norms is bounded by ρ(Bp,d, Bq,d) ≤ d1/p−1/q. We first
use the boundary cases p = 1 and p = ∞ to obtain dilation estimates from Proposition 3.4:
√
1 ≤ p ≤ 2 =⇒ θ(Bp,d) ≤ ρ(Bp,d, B1,d) θ(B1,d) ≤ d1−1/p · d1/2 = d3/2−1/p = d1−1/p−1/2,
2 ≤ p ≤ ∞ =⇒ θ(Bp,d) ≤ ρ(Bp,d, B∞,d) θ(B∞,d) ≤ d1/p · d1/2 = d1/2+1/p = d1−1/p−1/2.
Then we use the fact that θ(B2,d) = d (see [6, Example 7.24] or Corollary 5.7) for minimality:
1 ≤ p ≤ ∞ =⇒ d = θ(B2,d) ≤ ρ(B2,d, Bp,d) θ(Bp,d) ≤ d1/p−1/2 · θ(Bp,d) =⇒
θ(Bp,d) ≥ d1−1/p−1/2.
Because the constants θ(Bp,d) given in Theorem 6.9 are achieved by pushing to the bound-
ary cases p = 1 and p = ∞, note that different properties of the explicit dilations are
inherited. For example, the dilation technique bearing witness to
1 ≤ p ≤ 2 =⇒ W max(Bp,d) ⊆ d3/2−1/p · W min(Bp,d)
produces dilations which annihilate each other. This places tighter bounds on the joint
spectrum, so that
1 ≤ p ≤ 2 =⇒ W max(Bp,d) ⊆ d3/2−1/p · W min(B1,d)
(6.14)
holds. Thus,
θ(Bp,d, B1,d) = θ(Bp,d) = d3/2−1/p,
Moreover, for 1 ≤ p ≤ 2, the technique dilates B(H)d
dimension is not proved minimal.
p ∈ [1, 2].
for
sa to B(H ⊗ C2d+1)d
On the other hand, we know that the dilation scheme
p > 2 =⇒ W max(Bp,d) ⊆ d1/2+1/p · W max(Bp,d)
sa, although this
cannot always be achieved with annihilating operators, as otherwise the joint spectrum of a
dilation would actually be in d1/2+1/p·B1,d ⊂ d1/2+1/p·B2,d. Since d1/2+1/p is strictly less than
d and B2,d ⊆ Bp,d, such a claim would contradict the minimality of θ(B2,d) = d. However,
this dilation is obtained from the cube:
p ≥ 2 =⇒ W max(Bp,d) ⊆ W max(B∞,d) ⊆ d1/2 · W min(B∞,d).
(6.15)
Thus,
θ(Bp,d, B∞,d) = θ(B∞,d) = d1/2,
for
p ∈ [2,∞].
32
Similarly, for p ≥ 2, the technique dilates B(H)d
know whether this dimension is necessarily minimal.
A similar computation holds for the closed unit balls of (cid:96)p space in Cd ∼= R2d.
Corollary 6.10. Let Bp,d(C) denote the closed unit ball of Cd in the (cid:96)p norm, viewed as a
subset of R2d. Then
sa to B(H ⊗ C4d−1)d
sa, but again we do not
θ(Bp,d(C)) = 2 θ(Bp,d) = 2 d1−1/p−1/2.
√
Proof. We proceed as in the real case, continuing to use the boundary cases from real (cid:96)p
balls: θ(B∞,2d) =
2d = θ(B1,2d). Note that the Banach-Mazur distance is governed by
√
√
2 ρ(Bp,d(C), B∞,d(C)) ≤
ρ(Bp,d(C), B∞,2d) ≤
2 d1/p,
√
√
2 ρ(Bp,d(C), B1,d(C)) ≤
ρ(Bp,d(C), B1,2d) ≤
2 d1−1/p,
√
√
2d = 2 d3/2−1/p,
2 d1−1/p
√
√
2d = 2 d1/2+1/p.
2 d1/p
1 ≤ p ≤ 2 =⇒ θ(Bp,d(C)) ≤ ρ(Bp,d(C), B1,2d) θ(B1,2d) ≤
2 ≤ p ≤ ∞ =⇒ θ(Bp,d(C)) ≤ ρ(Bp,d(C), B∞,2d) θ(B∞,2d) ≤
so we have
Finally, since real and complex (cid:96)2-balls (of the appropriate dimension) are identical, we have
ρ(Bp,d(C), B2,2d) = ρ(Bp,d(C), B2,d(C)) ≤ d1/p−1/2,
which along with θ(B2,2d) = 2d gives that
1 ≤ p ≤ ∞ =⇒ 2d = θ(B2,2d) ≤ ρ(B2,2d, Bp,d(C)) θ(Bp,d(C)) = d1/p−1/2 · θ(Bp,d) =⇒
θ(Bp,d(C)) ≥ 2 d1−1/p−1/2 =
(cid:26) 2 d3/2−1/p,
2 d1/2+1/p,
1 ≤ p ≤ 2
2 ≤ p ≤ ∞ .
(cid:16)Dd(cid:17)
√
= 2
d,
A special case of the above is that the optimal dilation constant for the polydisk B∞,d(C) =
is given by
Dd
θ
√
) = 2
(6.16)
which improves the estimate 2d of [6, Corollary 7.10]. Now, θ(Dd
d implies that
√
every d-tuple T of (not necessarily normal) contractions dilates to a normal tuple N such
that (cid:107)Ni(cid:107) ≤ 2
Indeed, since any tuple of contractions T = (T1, . . . , Td) has
d for all i.
, we may conclude from (6.16) that T
(Re(T1), Im(T1), . . . , Re(Td), Im(Td)) ∈ W max(cid:16)Dd(cid:17)
√
admits a normal dilation consisting of operators with norm at most 2
In particular,
it does not matter if H is infinite-dimensional, as the explicit dilations presented capture
this case (alternatively, see subsection 2.3). However, the converse argument does not hold;
knowing that every d-tuple of contractions T dilates to a normal tuple N such that (cid:107)Ni(cid:107) ≤ C
for all i, does not imply that θ
such that X + iY is not a contraction. Similarly, if T = T1 is a contraction, then the Halmos
dilation
(cid:16)Dd(cid:17) ≤ C. Even for d = 1, W max(D) includes tuples (X, Y )
(cid:18)
(cid:19)
d.
(6.17)
√
I − T T ∗
−T ∗
T
√
I − T ∗T
33
is unitary, so it has norm 1, which is strictly less than θ(cid:0)D(cid:1).
We caution the reader that while the mathematical content and constants of [6, Corollary
7.10] are correct, the presentation of the corollary suggests that this faulty converse may
hold. (Indeed, while two facts are presented in the corollary, bridged by an implication, the
proof provides the implication in reverse order.) Regardless, we give a refinement of that
result.
Corollary 6.11. Fix d ≥ 1. Then
W max(cid:16)Dd(cid:17) ⊆ 2
√
d · W min(cid:16)Dd(cid:17)
,
and this constant is optimal for all d. Further, if (T1, . . . , Td) ∈ B(H)d is a tuple of (not
√
necessarily normal) contractions, then there exists a normal dilation N = (N1, . . . , Nd) such
that Nj ≤ min{d, 2
d}. This constant is not necessarily optimal for d ≥ 2.
√
Proof. The first statement is just (6.16), and the existence of a normal dilation with norms
d was shown directly after that equation.
bounded by 2
Alternatively, one may apply the Halmos dilation (6.17) to the members of (T1, . . . , Td) to
obtain a dilation (U1, . . . , Ud) of unitaries. Let P1, . . . , Pd be mutually orthogonal projections
in Md(C) such that each Pj has 1/d in the top left entry; for example we may take
(cid:32)
(cid:33)
,
Pj = projection onto
1√
,
d
ωj√
d
, . . . ,
ωj(d−1)√
d
where ω is a primitive dth root of unity. Then (U1⊗dP1, . . . , Ud⊗dPd) is a normal dilation of T
(in particular, the normal operators Uj⊗dPj annihilate each other) such that Uj⊗dPj ≤ d
for each j.
√
d for d ∈ {1, 2, 3}, it is evident that θ
Since d < 2
d is not the same as the
optimal constant for dilating tuples of contractions. On the other hand, because dilations of
self-adjoint contractions are governed by θ([−1, 1]d) =
d (where we note that if a dilation
√
is not self-adjoint, we can remove the imaginary component), we know we can achieve no
better scale than
d. At the moment we do not see a clear gap between the self-adjoint
and general cases. Indeed, since the Halmos dilation (6.17) gives normality of the individual
operators Ti for free, it is conceivable that no gap exists.
√
(cid:16)Dd(cid:17)
√
= 2
Problem 6.12. Let T = (T1, . . . , Td) be a tuple of contractions. Must there exist a normal
dilation N = (N1, . . . , Nd) of T such that Nj ≤ √
d for all j?
Finally, we note that the real and complex (cid:96)p-balls satisfy the hypotheses of Corollaries 5.11
and 5.12, and the estimate for θ(Bp,d) or θ(Bp,d(C)) given by computing the Banach-Mazur
distance to an (cid:96)2-ball is actually attained. Therefore we can conclude the following.
Corollary 6.13. Let K, L ⊂ Rd or K, L ⊂ R2d be orthogonal images of the same real or
complex (cid:96)p unit ball. If there is some C > 0 and a translation b such that
W max(K) ⊆ C · W min(b + L)
holds, then C ≥ θ(K). That is, θ(K) ≤ θ(K, b + L), and in particular, θ(K) = θ(K).
34
Remark 6.14. Since all the explicit dilations in this section dilate operators X1, . . . , Xd ∈
B(H) to block-operators whose entries are in the real unital C∗-algebra generated by X1, . . . , Xd,
it follows that such dilations can always be achieved continuously. For example, we may con-
sider C∗-algebras such as C(X, Mk(C)) for any compact Hausdorff space X.
7. Connections with cones and operator systems
In [6], the focus was on closed and bounded matrix convex sets - these correspond to
matrix ranges of d-tuples of bounded operators. In [8], the focus was on closed and salient
matrix convex cones with an interior point, which were referred to by the authors abstract
operator systems - as these correspond to finite dimensional operator systems.
In this
section, we describe how to go back and forth between both points of view, and see how the
minimal and maximal matrix convex sets over a set behave with respect to this.
7.1. Operator systems and matrix convex cones. Recall that a (d + 1)-dimensional
abstract operator system is a free set C such that
(1) C = ∪C(n) ⊆ Md+1
is a closed matrix convex set,
(2) each C(n) is a closed cone,
(3) each C(n) is salient: C(n) ∩ (−C(n)) = {0}, and
(4) there exists u ∈ Rd such that u ⊗ In is an interior point for all n.
sa
The assumption C(1) ∩ (−C(1)) = {0} implies that after an affine change of coordinates, we
may assume that C(1) \ {0} ⊆ {x ∈ Rd+1 : xd+1 > 0}. By applying another affine change of
coordinates that fixes e1 = (1, 0, . . . , 0), . . . , ed = (0, . . . , 0, 1, 0), we may assume in addition
that the unit is given by u = ed+1 = (0, . . . , 0, 1). Thus we shall make this a standing
assumption.
Standing assumption. All (d+1)-dimensional abstract operator systems considered below
will be assumed to satisfy
C(1) \ {0} ⊆ {x ∈ Rd+1 : xd+1 > 0},
(7.1)
and
(7.2)
the order unit of C is given by u = ed+1 = (0, . . . , 0, 1).
A realization of C consists of a concrete operator system in B(H) with an ordered basis
T1, . . . , Td, I, such that a (d + 1)-tuple X = (X1, . . . , Xd+1) is in C(n) if and only if
hLT (X) := T1 ⊗ X1 + . . . + Td ⊗ Xd + I ⊗ Xd+1 ≥ 0,
which means that C = DhLT := {X ∈ Md+1
sa
: hLT (X) ≥ 0}.
It is easy to see why a matrix convex set C as above is called an operator system: for every
finite dimensional operator system one can find a basis T1, . . . , Td, I consisting of self-adjoint
elements. Then letting S = ST be the operator system generated by T1, . . . , Td, I, we see that
elements in Mn(S) correspond to sums of the form hLT (X) := T1⊗X1+. . .+Td⊗Xd+I⊗Xd+1,
where X ∈ Md+1. An element hLT (X) ∈ Mn(S) is then in Mn(S)+ if and only if hLT (X) ≥ 0.
Note that since T1, . . . , Td are all self-adjoint and linearly independent, hLT (X) can be self-
adjoint only if X ∈ Md+1
(If we used a basis
T1, . . . , Td, I that is not required to be self-adjoint, we would need to consider tuples X ∈
Md+1).
sa , i.e., if X1, . . . , Xd+1 are all self-adjoints.
35
Lemma 7.1. Let S be an operator system, spanned by a basis of self-adjoints {I, T1, . . . , Td}.
Then there exists a basis of self-adjoints {I, A1, . . . , Ad} for S such that 0 ∈ intW(A) and
such that DLA is bounded.
Proof. By [6, Lemma 3.4], the conditions (i) 0 ∈ intW(A), and (ii) DLA is bounded, are
both equivalent to (iii) 0 ∈ intW1(A). We therefore consider W1(T ). If the convex set W1(T )
had no interior point, then it would have to be contained in a hyperplane. This means that
there are constants a0, a1, . . . , ad ∈ R, not all zero, such that
(cid:32) d(cid:88)
i=1
(cid:33)
d(cid:88)
i=1
φ
aiTi
=
aiφ(Ti) = a0 = φ (a0I) .
Choose an interior point c = (c1, . . . , cd) ∈ int(W1(T )). Now putting Ai = Ti − ciI, we
for every state φ. As {I, T1, . . . , Td} is a linearly independent set, this is impossible.
obtain a basis {I, A1, . . . , Ad} such that 0 ∈ intW1(A).
7.2. The operations C (cid:55)→ C and S (cid:55)→ S. Given an abstract operator system C ⊆ Md+1
sa ,
we define C ⊆ Md
sa by
C(n) = {A = (A1, . . . , Ad) ∈ (Mn)d
sa : (A, In) ∈ C(n)}.
Note that this is the same as defining
C(n) = {A = (A1, . . . , Ad) ∈ (Mn)d
sa : ∃B ∈ (Mn)sa,(cid:107)B(cid:107) ≤ 1 and (A, B) ∈ C(n)}.
It is straightforward to check that C is a closed matrix convex set with 0 ∈ int C (this
makes use of the standing assumption (7.2)). It follows from the definition and from the
comments above that, if {T1, . . . , Td, I} is a realization for C, then C is the free spectrahedron
determined by the pencil LT , defined by
LT (A) := T1 ⊗ A1 + . . . + Td ⊗ Ad + I ⊗ I.
sa : LT (A) ≥ 0}.
Thus C = (DhLT )∨ = DLT := {A ∈ Md
In general, given an abstract operators system C, the matrix convex set C constructed
as above might be unbounded; however the standing assumption (7.1) implies that C(1) is
bounded, and hence C is too (by, e.g., [6, Lemma 3.4]).
Conversely, if S is a matrix convex set with 0 in the interior, then S = DLT for some T ,
i=1 Ti ⊗ Ai + I ⊗ I
i.e., S is the free spectrahedron determined by some pencil LT (A) =(cid:80)d
(see [6, Proposition 3.5]). But then the homogeneous pencil
hLT (X) :=
Ti ⊗ Xi + I ⊗ Xd+1
d(cid:88)
i=1
determines a free spectrahedron which is a closed matrix convex salient cone in Md+1
sa , that
is, a (d+1)-dimensional operator system. We write S for this free spectrahedron DhLT . Note
that if S is bounded, then S(1)\{0} is contained in the open halfspace {x ∈ Rd+1 : xd+1 > 0}.
Moreover, given any (d + 1)-dimensional operator system S, Lemma 7.1 says that we can
find a basis {I, T1, . . . , Td} such that DLT is bounded, so there is no loss of generality.
Therefore, starting from a (d + 1)-dimensional operator system, and choosing a basis
{I, T1, . . . , Td} such that DLT is bounded, we find that the abstract operator system C =
DhLT = (DLT )∧ satisfies the assumptions (7.1) and (7.2), that is, that C(1)\{0} ⊆ {x ∈ Rd+1 :
36
xd+1 > 0} and that u = ed+1 = (0, . . . , 0, 1) is the order unit, i.e., the point corresponding
to I in the realization.
By a Theorem of Zalar ([24, Theorem 2.5]), under the assumption that DLT is bounded,
DLT1
⊆ DLT2
⇐⇒ DhLT1
⊆ DhLT2
.
(7.3)
Since a matrix convex set has the form DLT if and only if it contains 0 in its interior [6,
Proposition 3.5], we conclude that the maps C (cid:55)→ C and S (cid:55)→ S are mutual inverses between
the set of (d+1)-dimensional abstract operators systems satisfying (7.1) and (7.2), and closed
and bounded matrix convex sets in d variables that contain the origin in their interior.
7.3. The operations C (cid:55)→ C and K (cid:55)→ K at the scalar level. W min and W max for
cones. Given a closed salient cone C ⊆ {x ∈ Rd+1 : xd+1 > 0} with (0, . . . , 0, 1) ∈ int C, we
define
C = {x ∈ Rd : (x, 1) ∈ C}.
Conversely, given a closed convex set K ⊆ Rd, we let K denote the closed convex cone
generated by 0 ∈ Rd+1 and {(x, 1) : x ∈ K} ⊆ Rd+1.
Recall from Section 1.2 the definitions of the polar dual of a closed convex or matrix
convex set. For cones, a slightly different description of the polar duals is available: if C is
a closed convex cone, then one sees that
C(cid:48) = {y : (cid:104)x, y(cid:105) ≤ 0 for all x ∈ C};
if C is an operator system, then it is also easy to check that C• is also given by
(7.4)
C• = {Y :
Xi ⊗ Yi ≤ 0 for all X ∈ C}.
(cid:88)
i
It is not hard to check that - up to a sign change in the inequality - this polar duality
corresponds with the duality of operator spaces defined by Paulsen, Todorov and Tomforde
in [21, Definition 4.1].
Proposition 7.2. Let K ⊆ Rd be a compact convex set with 0 ∈ int K. Then
Proof. Recall that K = K(cid:48)(cid:48) = {x ∈ Rd :(cid:80)d
W min(K)∧ = W min( K)
(cid:40)
W max(K) =
X = (X1, . . . , Xd) :
and W max(K)∧ = W max( K).
i=1 xiyi ≤ 1 for all y ∈ K(cid:48)}, and therefore
d(cid:88)
yiXi ≤ I for all y ∈ K(cid:48)
(cid:41)
.
Then
W max(K)∧ =
On the other hand,
K =
i=1
(cid:40)
X = (X1, . . . , Xd+1) : Xd+1 − d(cid:88)
(cid:40)
(x1, . . . , xd+1) : xd+1 − d(cid:88)
i=1
xiyi ≥ 0 for all y ∈ K(cid:48) ≥ 0
,
i=1
37
(cid:41)
.
yiXi ≥ 0 for all y ∈ K(cid:48)
(cid:41)
so
(cid:40)
(X1, . . . , Xd+1) : Xd+1 − d(cid:88)
W max( K) =
(cid:41)
.
yiXi ≥ 0 for all y ∈ K(cid:48)
i=1
That proves W max(K)∧ = W max( K).
For the other equality, first let us note that W min(K)∧ ⊇ W min( K) is immediate, since
K ⊆ W min(K)∧. Now, suppose that W min(K) = DLT . Then W min(K)∧ = DhLT , and
note that DhLT (1) = K. The set W min( K) is also a matrix convex cone, which is salient and
contains an order unit (0, . . . , 0, 1), so it is therefore realized as the operator system generated
by a (d + 1)-tuple (S1, . . . , Sd, I). Equivalently, it is the free spectrahedron determined by a
homogeneous linear pencil: W min( K) = DhLS . We claim that
DLT ⊆ DLS .
Assuming the claim, we invoke [24, Theorem 2.5] to find that
W min(K)∧ = DhLT ⊆ DhLS = W min( K).
It remains to prove the claim. Now, DLS is a matrix convex set, and to prove that it
contains DLT = W min(K), it suffices to show that K = DLT (1) ⊆ DLS (1). But if x ∈ K,
then (x, 1) ∈ K = DhLS (1), so hLS(x, 1) = LS(x) ≥ 0. Whence x ∈ DLS (1), and the proof is
complete.
Below we will require the following closely related lemma, the proof of which is omitted.
Lemma 7.3. Let K be a compact convex subset of Rd, and assume that 0 ∈ int K. Then
(− K)(cid:48) = (−K(cid:48))∧.
7.4. Inclusions of cones and finite-dimensional representations. Following [8], for a
closed salient cone C ⊆ {x ∈ Rd+1 : xd+1 > 0} with (0, . . . , 0, 1) ∈ int C, and a positive
number ν > 0, let us define ν ↑ C := (ν C)∧. In [8, Section 5], Fritz, Netzer and Thom found
constants ν such that W max(ν ↑ C) ⊆ W min(C). For example, in [8, Theorem 5.8] they show
that if C is symmetric with respect to 0, then W max( 1
d ↑ C) ⊆ W min(C). They note that for
(cid:19)
applications one is interested in finding the largest ν for which this inclusion occurs. Our
results apply to this setting: for example, we see that
C = [−1, 1]d =⇒ W max
↑ C
⊆ W min(C)
(cid:18) 1√
d
(7.5)
holds by Theorem 6.7 and Proposition 7.2, and this improves upon the previously known
constant ν = 1
d.
The problem of containment of matrix convex cones is relevant to interpolation problems
of UCP maps as well. By [24, Theorem 2.5], under assumptions (7.1) and (7.2), whenever
C = DhLT and C(cid:48) = DhLT(cid:48) , then C ⊆ C(cid:48) if and only if there exists a UCP map S(T ) → S(T (cid:48))
that maps Ti to T (cid:48)
i . This also follows from (7.3) together with results (such as Corollary
5.10 and Remark 5.11 in [6]) saying that DLT ⊆ DLT(cid:48) if and only if there exists a UCP map
S(T ) → S(T (cid:48)) that maps Ti to T (cid:48)
i .
sa has the form S = W(A) for some
tuple A ∈ B(H)d
sa ([6, Proposition 3.5]). It is natural to ask, under what geometric conditions
on S can this tuple A be chosen to act on a finite dimensional Hilbert space H? A dual
Any closed and bounded matrix convex set S ⊆ Md
38
version of this problem has been treated for cones in [8, Theorems 3.2 and 4.7]. Their results
show that the maximal operator system over a cone C is finite-dimensionable realizable,
i.e. can be realized as DhLA for A acting on a finite-dimensional space, if and only if C
is polyhedral. Further, under the assumption that C is a salient polyhedral cone, it holds
that the minimal operator system over C is finite-dimensional realizable if and only if C
is simplicial, which holds if and only if the minimal and maximal operator systems over
C are equal. (Recall that a cone C ⊆ Rd+1 is said to be a simplicial cone if it is affinely
isomorphic to a positive orthant {x ∈ Rd+1 : x1 ≥ 0, . . . , xd+1 ≥ 0}. A cone C such that
C \ {0} ⊆ {x ∈ Rd+1 : xd+1 > 0} is simplicial if and only C is a simplex.)
In this subsection we treat the corresponding problems for W min(K) and W max(K), giving
alternative proofs and improvements of [8] whenever possible. The following theorem may be
proved by importing the results of [8] (where we note that "min" and "max" are interchanged
when translating the results to our setting), but we give an alternative proof for the statement
regarding W min(K).
Theorem 7.4. Let K be a compact convex subset of Rd. Then W min(K) = W(A) for
sa if and only if K is a polytope. If K is a polytope, then W max(K) = W(A) for
A ∈ Md
A ∈ Md
sa if and only if K is a simplex.
Proof. By [6, Corollary 2.8], if A is a normal tuple, then W(A) is the smallest matrix convex
set containing σ(A), so W(A) = W min(conv(σ(A)). If K is a polytope, then letting A be the
normal tuple that has joint eigenvalues at the vertices of K, we have that W min(A) = W(A).
Moreover, if K is a simplex, this still holds, and by Theorem 4.1, W max(K) = W min(K) =
W(A), and we are done with one direction.
sa, then A ∈ W min(K), so
by Proposition 2.3 A has a normal dilation N ∈ Md
sa with σ(N ) ⊆ K. But then W(A) ⊆
W(N ), so K ⊆ W1(A) ⊆ W1(N ) = conv σ(N ). We conclude that conv σ(N ) = K. Since
the spectrum of N has finitely many points, K is polytope.
Suppose now that K is polytope. Assume without loss of generality that 0 ∈ int K, and
that W(A) = W max(K). Then, using Proposition 7.2, W(A)∧ = W max(K)∧ = W max( K)
is an abstract operator system. Taking polar duals and applying Lemma 7.3, we find that
W max( K)• = W min(( K)(cid:48)) = −W min((−K(cid:48))∧). Since W(A) = W max(K), we get W(A)• =
W min(K(cid:48)). On the other hand, W(−A)• = DLA (this follows from [6, Proposition 3.1]; note
the change in sign convention). Thus, −K(cid:48) = DLA(1), and using Proposition 7.2 again,
Now for the converse direction. If W min(K) = W(A) for A ∈ Md
W min((−K(cid:48))∧) = W min(−K(cid:48))∧ = (DLA)∧ = DhLA.
So A is a finite dimensional realization for W min((−K(cid:48))∧). By [8, Theorem 4.7], (−K(cid:48))∧
must be simplicial, therefore K = K(cid:48)(cid:48), together with K(cid:48), must be simplices.
Remark 7.5. Dual to the comments in [8], note that even if K is not a polytope, it is still
possible that W max(K) could equal the matrix range of a tuple of matrices. For this, let
sa be as in Section 5. By [13, Corollary 14.15], DLF = W min(B2,2). It
F = (F [2]
follows that W(F ) = W max(B2,2).
2 ) ∈ (M2)2
1 , F [2]
While Theorem 7.4 is a direct translation of results from [8], the following theorem is a
slight improvement of its corresponding result, having removed the assumption that C is
polyhedral.
39
Corollary 7.6. Let C ⊆ Rd+1 be a closed salient convex cone with nonempty interior. Then
W min(C) = W max(C) if and only if C is a simplicial cone.
Proof. As we discussed in the beginning of the section, we may assume that C \ {0} ⊆
{x ∈ Rd+1 : xd+1 > 0} and (0, . . . , 0, 1) ∈ int C. Suppose that W min(C) = W max(C). By
Proposition 7.2,
W min( C)∧ = W min(C) = W max(C) = W max( C)∧.
Because the maps C (cid:55)→ C and S (cid:55)→ S are mutual inverses, we find that W min( C) = W max( C).
By Theorem 4.1, C is a simplex, so C is a simplicial cone.
The converse follows from Theorem 4.1 and Proposition 7.2 in a similar way.
Problem 7.7. Characterize the tuples A for which there exists some K such that W(A) =
W min(K). Likewise, characterize the tuples A for which there exists some K such that
W(A) = W max(K).
8. Dilation constants versus minimal dilation hulls
Given compact convex sets K, L ⊆ Rd, we have considered the dilation constants θ(K, L),
θ(K) and θ(K) (see equations (3.6) (3.7) and (3.8)). We know from Theorems 6.4 and 6.9
p,d) = d1−1/p and θ(Bp,d) = d1−1/2−1/p. However, the explicit dilations which were
that θ(B+
used to compute θ(·) often had joint spectrum in smaller sets, which were merely contained
in a multiple of Bp,d or B+
p,d. We now consider a sense of minimality among dilation hulls
which can see this distinction.
Define
E(K) = {L ⊇ K compact and convex : W max(K) ⊆ W min(L)}.
(8.1)
From the definitions (1.2) and (1.3) of W max and W min, we know that a compact convex set
L ⊆ Rd is in E(K) if and only if for every X ∈ (Mn)d
sa with W1(X) ⊆ K, there exists a tuple
sa of commuting self-adjoint operators such that σ(N ) ⊆ L and N is a dilation of
N ∈ B(H)d
X. (Recall that N is a dilation of X if and only if
Xi = V ∗NiV , i = 1, . . . , d,
holds for some isometry V : Cn → H.) If L ∈ E(K), then we say that L is a dilation hull
of K. It is worth pointing out that E(K) is not closed under intersections (e.g., by using
Theorem 6.7).
Proposition 8.1. Let K ⊆ Rd be a compact convex set. If E(K) is ordered by inclusion,
then E(K) has a minimal element.
Proof. Suppose that {Lα} is a chain (decreasing by inclusion) in E(K). We must show that
L := ∩αLα ∈ E(K). Let X ∈ W max
(K). For every α, there is an isometry V (α) : Cn → H (α)
and a normal tuple N (α) ∈ B(H (α))d
sa such that X = V (α)∗N (α)V (α) and σ(N (α)) ⊆ Lα. By
Proposition 2.3, we may assume that H (α) = CN , for N = 2n3(d + 1) + 1. Therefore, one
can choose convergent subnets V (β) → V , N (β) → N , so that X = V ∗N V and σ(N ) ⊆
∩βLβ = ∩αLα = L.
n
Let us call a minimal element of E(K) a minimal dilation hull and define
md(K) := {L ∈ E(K) : L is minimal}.
40
First, there is no reason for elements in md(K) to resemble K in any way; the (cid:96)2-ball provides
a counterexample.
Proposition 8.2. There is no (cid:96)2-ball in md(B2,d).
Proof. Suppose a ball L = x + C · B2,d is in E(B2,d). Then from Theorem 5.6 it follows that
there is a simplex Π with B2,d ⊂ Π ⊂ L, and certainly Π ∈ E(Π) ⊆ E(B2,d). Therefore L is
not minimal: L (cid:54)∈ md(B2,d).
However, note that while Theorem 5.6 shows that if an (cid:96)2-ball belongs to E(B2,d), then
there is a smaller simplex which is also a dilation hull, the theorem does not apply for other
shapes. For example, the d-diamond d · B1,d is a minimal dilation hull of B2,d.
Proposition 8.3. It holds that d · B1,d ∈ md(B2,d).
Proof. The containment (6.14) shows that d· B1,d ∈ E(B2,d). Suppose that there is a proper
compact convex subset L ⊂ d · B1,d such that W max(B2,d) ⊆ W min(L). Then one of the
extreme points of the diamond is missing, so by rotational symmetry of the ball we may
suppose that (−d, 0, . . . , 0) (cid:54)∈ L. Therefore
L ⊆ M = {(cid:126)x ∈ d · B1,d : −d + ε < x1 ≤ d}
(cid:113)
(x1 − ε)2 +(cid:80)d
for some ε > 0, which we may suppose has ε < 1. Now, M is the convex hull of points of the
form (−d + ε, b) with b(cid:96)1 ≤ ε, as well as the extreme points of d·B1,d, except (−d, 0, . . . , 0).
√
d2 + ε2 on these points, so as any (cid:96)2-
The expression
ball is convex, it follows that L ⊆ M is contained in the ball of radius C =
d2 + ε2 centered
at y = (ε, 0, . . . , 0). From Theorem 5.6, we conclude
k is bounded by
k=2 x2
√
d2 + ε2 ≥(cid:112)ε2 + (d − 1)2 + 1,
√
but this is a contradiction.
√
d,
d,
√
√
d]d ∈ E(B2,d), [−√
√
Certainly this also implies that there is no simplex Π with B2,d ⊂ Π ⊂ d · B1,d. On the
other hand, while [−√
d]d is not generally a minimal dilation hull.
Proposition 8.4. The claim [−√
d]d ∈ md(B2,d) is true for d = 2 and false for d = 4.
Proof. If d = 2, then [−√
2]2 is an orthogonal image of the diamond 2 · B1,2, which is a
minimal dilation hull of the ball B2,2. Since the ball is orthogonally invariant, [−√
2]2 is
On other other hand, if d = 4, then while [−√
also minimal.
4]4 = [−2, 2]4 is a dilation hull of B2,4,
so is every orthogonal image of the diamond 4 · B1,4. Such an image may be found properly
inside [−2, 2]4; consider the convex hull of the mutually orthogonal vectors
(2,−2,−2, 2),
(2, 2,−2,−2),
(2,−2, 2,−2),
(2, 2, 2, 2),
√
√
2,
d,
2,
4,
and their opposites, all of which have norm 4.
We note that the shape of minimal dilation hulls of K is not necessarily unique, once again
seen when K is an (cid:96)2-ball. Namely, Proposition 8.3 and the following proposition produce
diamonds and simplices as minimal dilation hulls, respectively.
41
Proposition 8.5. Let ∆ ⊂ Rd be any d-simplex such that B2,d ⊂ ∆ and each vertex of ∆
lies on d · ∂B2,d. Then ∆ ∈ md(B2,d).
Proof. Since W max(∆) = W min(∆), it is clear that ∆ is a dilation hull of B2,d. Suppose a
proper compact convex subset L ⊂ ∆ is a dilation hull, so one of the vertices of ∆ is missing
from L. It follows that L may be contained in a ball of radius d centered at some x (cid:54)= 0, so
W max(B2,d) ⊆ W min(x + d · B2,d). This contradicts Theorem 5.6.
An example of such a simplex ∆ = ∆(d) can be defined by the base case ∆(1) = [−1, 1]
(cid:16)(cid:113) d+1
d−1 · ∆(d − 1)
(cid:17) × {−1} and
and the recursive definition of ∆(d) as the convex hull of
(0, . . . , 0, d). By repeated application of Lemma 5.4, it is seen that B2,d ⊆ ∆. Further, all of
the vertices of ∆(d) have norm d by the computation
Problem 8.6. For what types of collections K,L of compact convex sets does it hold that if
K ∈ K, L ∈ L, and W max(K) ⊆ W min(L), then there is a simplex Π with K ⊆ Π ⊆ L?
(cid:16)(cid:113) d+1
(cid:17)2
d−1 · (d − 1)
+ (−1)2 = d2.
If K = L is the collection of all (cid:96)2-balls in Rd, then interpolating simplices can be found,
but they cannot necessarily be found when K = L is the collection of cubes, or diamonds,
and so on. Can this be generalized?
We now consider another special case in which a circumscribed simplex over K is a minimal
dilation hull of K. This also gives a partial answer to Problem 4.3.
Definition 8.7. We say that a convex body K ⊂ Rd is simplex-pointed at x if x ∈ K and
there is an open set O ⊂ Rd such that x ∈ O and O ∩ K is a d-simplex with x as a vertex.
Equivalently, there is an invertible affine transformation A on Rd such that
A(x) = (0, . . . , 0),
A(K) ⊂ [0,∞)d,
and
∃ε > 0 with {(x1, . . . , xd) ∈ Rd : 0 ≤ xj ≤ ε for all j} ⊆ K.
Figure 1. K is simplex-pointed
at x, but not polyhedral.
Figure 2. K is polyhedral, but
not simplex-pointed at any x.
Theorem 8.8. Let K ⊂ Rd be a convex body which is simplex-pointed at x ∈ K. Suppose ∆
is a d-simplex such that K ⊆ ∆, x is a vertex of ∆, and the edges of ∆ which emanate from
x point in the same direction as the edges of the simplex O ∩ K based at x. Let F ⊂ ∆ be
42
the face of ∆ which does not include x, and suppose that the interior of the face F includes
a point y ∈ K. Then ∆ ∈ md(K).
Moreover, if K meets these conditions and W max
particular, K is a simplex.
Proof. After an affine transformation, we may suppose K ⊂ [0,∞)d, x = (0, . . . , 0) ∈ K,
K ⊆ ∆ = B+
(K), then K = ∆, and in
1,d, and there exists α > 0 such that
(K) = W min
2
2
(8.2)
0 ≤ aj ≤ α ∀j ∈ {1, . . . , d} =⇒ (a1, . . . , ad) ∈ K.
Suppose L ⊆ B+
Moreover, there is a point y = (y1, . . . , yd) ∈ K such that yj > 0 for all j, and y(cid:96)1 = 1.
1,d is a compact convex set such that W max(K) ⊆ W min(L). Since
(0, . . . , 0) ∈ K, it certainly follows that (0, . . . , 0) ∈ L. We will show that each standard
1,d. Fix 1 ≤ j ≤ d and note that since
basis vector ej is also in L, which implies L = B+
y(cid:96)1 = 1 and each yj is strictly greater than zero, we can fix r ∈ (0, 1) such that for each
yk < r < 1. Approximate y with an interior point c of K such that
j, (cid:80)
k(cid:54)=j
(8.3)
2 · 1 − c(cid:96)1
1 − r
< 1
and
(cid:88)
k(cid:54)=j
ck < r.
Also, fix any ε > 0 small enough that a strengthening of (8.3) holds:
(8.4)
δ := 2 · 1 − c(cid:96)1 · (1 − ε)
1 − r
< 1.
Next, choose projections P1, . . . , Pd ∈ M2(C) (dependent on ε) such that Pj − Pk < ε
for each j and k, but P1, . . . , Pd project onto lines which pairwise intersect trivially. The
tuple (c1P1, c2P2 . . . , cdPd) has joint numerical range which is contained in [0,∞)d and is
within ε of the joint numerical range of (c1P1, c2P1, . . . , cdP1), i.e. the line segment between
c and 0. Since c is an honest interior point of K and (8.2) shows that there is a nonnegative
neighborhood of 0 within K, it follows that for sufficiently small ε > 0, the joint numerical
range of (c1P1, . . . , cdPd) is within K. That is, (c1P1, . . . , cdPd) ∈ W max(K). Since we have
assumed W max(K) ⊆ W min(L), there is a normal dilation N of (c1P1, . . . , cdPd) with joint
spectrum in L. By design, (c1P1, . . . , cdPd) satisfies the conditions of Lemma 6.3, and we may
find a different normal dilation Z such that ZiZj = 0 for i (cid:54)= j and σ(Z) ⊆ σ(N ) ∪ {0} ⊆ L.
Because the operators Zk annihilate each other and have norm bounded by 1 (as L ⊆ B+
1,d),
it follows that for each j,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Zj + (1 − δ)
(cid:88)
k(cid:54)=j
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ max{Zj, (1 − δ)Zk} ≤ max{Zj, 1 − δ}.
Zk
43
On the other hand,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)cjPj + (1 − δ)
(cid:88)
k(cid:54)=j
ckPk
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:33)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88)
(cid:32) d(cid:88)
≥ c(cid:96)1 − c(cid:96)1 · ε − δ ·(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) −
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
k(cid:54)=j
Pj
k=1
ck
≥ c(cid:96)1 · (1 − ε) − δ · r.
ck
k(cid:54)=j
ck(Pk − Pj)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) − δ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88)
k(cid:54)=j
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ckPk
We may therefore conclude that
max{Zj, 1 − δ} ≥ c(cid:96)1 · (1 − ε) − δ · r.
The definition of δ = 2 · 1 − c(cid:96)1 · (1 − ε)
1 − δ. It follows that
1 − r
shows that this inequality cannot be satisfied by
Zj ≥ c(cid:96)1 · (1 − ε) − δ · r = c(cid:96)1 · (1 − ε) − 2r · 1 − c(cid:96)1 · (1 − ε)
:= bj(c, ε).
1 − r
As (Z1, . . . , Zd) is a tuple of mutually annihilating operators with joint spectrum in L, and
0 ∈ L, it follows that bj(c, ε)·ej ∈ L. Since this claim holds for all ε > 0 which are sufficiently
small, it follows that
(cid:18)
c(cid:96)1 − 2r · 1 − c(cid:96)1
1 − r
(cid:19)
· ej ∈ L.
Similarly, if W max
(K) = W min
As c approaches y, and consequently c(cid:96)1 approaches 1, this implies ej ∈ L. Finally,
L = B+
1,d.
2
2
(K), then we may apply the same affine transformation to
suppose K ⊆ B+
1,d is positioned as above, repeating the argument with K = L. Because the
argument above only dilates 2 × 2 matrices, we may similarly conclude that K is equal to
B+
1,d, which in particular is a simplex.
Theorem 8.8 describes a particular scenario in which a circumscribing simplex over K is a
minimal dilation hull of K. It may be used to refine the dilation constants found in Theorem
6.4.
Corollary 8.9. For 1 ≤ p ≤ ∞, d1−1/p · B+
Proof. The simplex d1−1/p · B+
p,d, so it is certainly a dilation hull. Further,
p,d is simplex-pointed at (0, . . . , 0), and the simplex d1−1/p · B+
B+
1,d is positioned over B+
p,d
exactly as demanded in Theorem 8.8, with base point x = (0, . . . , 0) and opposite point
y = (d−1/p, . . . , d−1/p).
Remark 8.10. In Theorem 8.8, it is crucial that y ∈ K ∩ ∆ is in the interior of the face F
of ∆ which excludes x, and that the edges of ∆ based at x point in the same directions as
the edges of the simplex O ∩ K. Indeed, without these assumptions it is possible that there
is another simplex Π such that K ⊆ Π (cid:40) ∆. Because we know that W max(Π) = W min(Π),
this fact would certainly rule out ∆ as a minimal dilation hull of K. The figures below show
one example which meets the conditions of Theorem 8.8, and two examples which do not
meet the conditions.
1,d contains B+
1,d ∈ md(B+
p,d).
44
Figure 3. K and ∆ meet the conditions of Theorem 8.8.
Figure 4. ∆ is not a minimal dilation hull of K. Theorem 8.8 does not
apply, as the chosen intersection point y is in ∂F , where F is the face of ∆
that excludes x.
Figure 5. ∆ is not a minimal dilation hull of K. Theorem 8.8 does not apply,
as the edges of ∆ do not point in the same direction as the edges of K that
are defined near x.
45
We conclude with the following problem, which could potentially generalize Theorem 8.8
and Proposition 8.5.
Problem 8.11. Let K be a convex body and let ∆ be a simplex with K ⊆ ∆. If ∆ is a
circumscribing simplex, must ∆ ∈ md(K)?
acknowledgments
We would like to thank Emanuel Milman for giving us the lead to the reference [18]. We
also thank the referee for helpful feedback.
References
ematical Surveys and Monographs 202 (2015). 15
1. S. Artstein-Avidan A. Giannopoulos, and V.D. Milman, "Asymptotic geometric analysis, Part I" Math-
2. W. Arveson, Subalgebras of C∗-algebras, Acta Math. 123 (1969), 141–224. 3
3. W. Arveson, Subalgebras of C*-algebras. II., Acta Math. 128 (1972), 271–308. 4, 7
4. M.D. Choi, Completely Positive Linear Maps on Complex Matrices Linear Algebra Appl. 10 (1975),
285–290. 3
5. K.R. Davidson, C*-algebras by example, American Mathematical Soc., 1996. 2
6. K.R. Davidson, A. Dor-On, O.M. Shalit and B. Solel, Dilations, inclusions of matrix convex sets, and
completely positive maps, to appear in Int. Math. Res. Not. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16,
17, 18, 20, 21, 22, 24, 27, 29, 32, 33, 34, 35, 36, 37, 38, 39
7. E.G. Effros and S. Winkler, Matrix convexity: Operator analogues of the bipolar and Hahn-Banach
theorems, J. Funct. Anal. 144 (1997), 117–152. 8
8. T. Fritz, T. Netzer and A. Thom Spectrahedral containment and operator systems with finite-dimensional
realization, preprint, arXiv:1609.07908. 2, 4, 5, 7, 13, 14, 15, 35, 38, 39
9. Y. Gordon, A.E. Litvak, M. Meyer and A. Pajor, John's decomposition in the general case and applica-
tions, J. Diff. Geom., 68 (2004), 99–119. 15
10. P. R. Halmos, Normal dilations and extensions of operators, Summa Brasil. Math., 2 (1950), 125–134.
25, 30
11. J.W. Helton, I. Klep and S. McCullough, The matricial relaxation of a linear matrix inequality, Math.
Program. 138 (2013), 401–445. 1
12. J.W. Helton, I. Klep, and S. McCullough, Matrix convex hulls of free algebraic sets, Trans. Amer. Math.
Soc. 368 (2016), 3105–3139. 4
13. J.W. Helton, I. Klep, S. McCullough and M. Schweighofer, Dilations, Linear Matrix Inequalities, the
Matrix Cube Problem and Beta Distributions, preprint arXiv:1412.1481 , 2015. 1, 2, 6, 7, 39
14. F. John, Extremum problems with inequalities as subsidiary conditions, Courant Anniversary Volume,
Interscience, New York, 1948. MATHEMATICAL INSTITUTE, HUNGARIAN ACADEMY OF SCI-
ENCES 1053: 187–204. 16
15. W. B. Johnson and J. Lindenstrauss, Handbook of the geometry of Banach spaces. Vol. 2, North-Holland,
Amsterdam, 2003. 13
16. M. Lassak, Approximation of convex bodies by inscribed simplices of maximum volume, Beitrage zur
Algebra und Geometrie/Contributions to Algebra and Geometry 52:2 (2011), 389–394. 15
17. K. Leichtweiss, Uber die affine Exzentrizitat konvexer Korper, Archiv fur Mathematische Logik und
Grundlagenforschung 10 (1959), 187–199. 11
18. O. Palmon, The only convex body with extremal distance from the ball is the simplex, Israel J. Math. 80
(1992), 337–349. 11, 46
19. V.I. Paulsen, Representations of function algebras, abstract operator spaces, and Banach space geometry,
J. Funct. Anal 109 (1992), 113–129. 4
20. V.I. Paulsen, Completely bounded maps and operator algebras, Cambridge University Press, 2002. 2
21. V.I. Paulsen, I.G. Todorov and M. Tomforde, Operator system structures on ordered spaces, Proc. Lond.
Math. Soc. 102 (2011), 25–49. 4, 5, 37
46
22. G. Popescu, Isometric Dilations for Infinite Sequences of Noncommuting Operators Trans. Amer. Math.
Soc. 316 (1989) 523–536. 24
23. W.F. Stinespring, Positive Functions on C*-algebras, Proc. Amer. Math. Soc. 6 (1955), 211–216. 2
24. A. Zalar. Operator positivestellensatze for noncommutative polynomials positive on matrix convex sets,
J. Math. Anal. Appl. 445 (2017): 32–80. 1, 4, 37, 38
Faculty of Mathematics, Technion - Israel Institute of Technology, Haifa 3200003, Is-
rael
E-mail address: [email protected]
Faculty of Mathematics, Technion - Israel Institute of Technology, Haifa 3200003, Is-
rael
E-mail address: [email protected]
Faculty of Mathematics, Technion - Israel Institute of Technology, Haifa 3200003, Is-
rael
E-mail address: [email protected]
47
|
1107.3190 | 1 | 1107 | 2011-07-16T01:50:46 | Jordan higher all-derivable points in triangular algebras | [
"math.OA"
] | Let ${\mathcal{T}}$ be a triangular algebra. We say that $D=\{D_{n}: n\in N\}\subseteq L({\mathcal{T}})$ is a Jordan higher derivable mapping at $G$ if $D_{n}(ST+TS)=\sum_{i+j=n}(D_{i}(S)D_{j}(T)+D_{i}(T)D_{j}(S))$ for any $S,T\in {\mathcal{T}}$ with $ST=G$. An element $G\in {\mathcal{T}}$ is called a Jordan higher all-derivable point of ${\mathcal{T}}$ if every Jordan higher derivable linear mapping $D=\{D_{n}\}_{n\in N}$ at $G$ is a higher derivation. In this paper, under some mild conditions on ${\mathcal{T}}$, we prove that some elements of ${\mathcal{T}}$ are Jordan higher all-derivable points. This extends some results in [6] to the case of Jordan higher derivations. | math.OA | math |
Jordan higher all-derivable points in triangular algebras 1
Jinping Zhao 2, Jun zhu3
Institute of Mathematics, Hangzhou Dianzi University, Hangzhou 310018, People's Republic of China
Abstract
Let T be a triangular algebra. We say that D = {Dn : n ∈ N } ⊆ L(T ) is a Jordan higher
derivable mapping at G if Dn(ST + T S) =Pi+j=n(Di(S)Dj(T ) + Di(T )Dj(S)) for any S, T ∈ T
with ST = G. An element G ∈ T is called a Jordan higher all-derivable point of T if every Jordan
higher derivable linear mapping D = {Dn}n∈N at G is a higher derivation. In this paper, under
some mild conditions on T , we prove that some elements of T are Jordan higher all-derivable
points. This extends some results in [6] to the case of Jordan higher derivations.
AMS Classification:16W25, 47B47
Keywords : Jordan higher all-derivable point; triangular algebra; Jordan higher derivable linear
mapping at G
1. Introduction and preliminaries
Let A be a ring (or algebra) with the unit I. An additive linear mapping δ from A into itself is
called a derivation if δ(ST ) = δ(S)T + Sδ(T ) for any S, T ∈ A and is said to be a Jordan derivation
if δ(ST + T S) = δ(S)T + Sδ(T ) + δ(T )S + T δ(S) for any S, T ∈ A. We say that a mapping δ is
Jordan derivable at a given point G ∈ A if δ(ST + T S) = δ(S)T + Sδ(T ) + δ(T )S + T δ(S) for any
S, T ∈ A with ST = G, and G is called a Jordan all-derivable point of A if every Jordan derivable
mapping at G is a derivation. We say that D = {Dn} ⊆ L(A) is a Jordan higher derivable mapping
at G if Dn(ST + T S) =Pi+j=n(Di(S)Dj(T ) + Di(T )Dj(S)) for any S, T ∈ A with ST = G. An
element G ∈ A is called a Jordan higher all-derivable point of A if every Jordan higher derivable
linear mapping D = {Dn} at G is a higher derivation. There have been a number of papers on the
study of conditions under which derivations of operator algebras can be completely determined by
the action on some sets of operators. In [3], W. Jing showed that I is a Jordan all-derivable point
of B(H) with H is a Hilbert space. In [7], J. Zhu proved that every invertible operator in nest
algebra is an all-derivable point in the strong operator topology. Also it was showed that every
element in the algebra of all upper triangular matrices is a Jordan all-derivable point by Z. Sha
and J. Zhu in [6].
With the development of derivation, higher derivation has attracted much attention of math-
ematicians as an active subject of research in algebras. In [4] Z. Xiao and F. Wei showed that
any Jordan higher derivation on a triangular algebra is a higher derivation. In this paper we will
extend the conclusion of [6] to the case of Jordan higher derivations.
Let A and B be two unital rings (or algebras) with the unit I1, I2, and M be a unital (A,
B)-bimodule, which is faithfull as a left A-module and as a right B-module. The ring(or algebra)
1This work is supported by the National Natural Science Foundation of China (No 10771191)
2E-mail address: [email protected]
3E-mail address: zhu [email protected]
1
T = {(cid:20) a m
b (cid:21) : a ∈ A, m ∈ M, b ∈ B},
0
under the usual matrix operations is said to be a triangular algebra. We mainly proved that 0 and
(cid:20) I1 X0
I2 (cid:21) are Jordan higher all-derivable points for any given point X0 ∈ M.
0
2. Jordan higher all-derivable points in ring algebras
In this section, we always assume that the characteristics of A and B are not 2 and 3, and for
any X ∈ A, Y ∈ B, there are some integers n1, n2 such that n1I1 − X and n2I2 − Y are invertible.
The following two theorems are the main results in this paper.
Theorem 2.1 Let D = (Dn)n∈N be a family of additive linear mappings on T that D0 = iDT
(identical mapping on T ). If D is Jordan higher derivable at 0, then D is a higher derivation.
Proof. For any T =(cid:20) X Y
0 Z (cid:21) ∈ T , we can write
Dn((cid:20) X Y
0 Z (cid:21)) =(cid:20) δ11
n (X) + ϕ11
n (Y ) + τ 11
0
n (Z)
δ12
n (X) + ϕ12
n (X) + ϕ22
δ22
n (Y ) + τ 12
n (Y ) + τ 22
n (Z)
n (Z) (cid:21) ,
n : A → Aij , ϕij
where δij
A11 = A, A12 = M, A22 = B. It follows from the fact D0 = iDT that when i = j = 1, δij
else δij
0 = iτB, else τ ij
n : B → Aij , 1 ≤ i ≤ j ≤ 2 are additive maps with
0 = iδA,
0 = 0.
0 = 0; when i = 1, j = 2, ϕij
0 = 0; when i = j = 2 , τ ij
0 = iϕM, else ϕij
n : M → Aij , τ ij
0 (cid:21) and T = (cid:20) X 0
0 (cid:21) for every X ∈ A, W ∈ M. Then ST = 0 and
0
(Di(S)Dj(T ) + Di(T )Dj(S))
i (W ) ϕ12
ϕ22
0
n (XW )
n (XW ) (cid:21) = Dn(ST + T S) = Pi+j=n
j (X) (cid:21)
j (W ) (cid:21))
i (W ) (cid:21)(cid:20) δ11
δ12
j (X)
δ22
j (W )
i (X) (cid:21)(cid:20) ϕ11
j (W ) ϕ12
ϕ22
i (W )
j (X)
δ12
i (X)
δ22
0
0
((cid:20) ϕ11
i (X)
0
We set S = (cid:20) 0 W
(cid:21). So
0
0
n (XW ) ϕ12
ϕ22
0
0
T S =(cid:20) 0 XW
(cid:20) ϕ11
= Pi+j=n
+(cid:20) δ11
= Pi+j=n
This implies that
ϕ11
i (W )δ11
j (X) + δ11
i (X)ϕ11
j (W )
ϕ11
+ϕ12
i (W )δ12
i (W )δ22
j (X) + δ11
j (X) + δ12
i (X)ϕ12
i (X)ϕ22
j (W )
j (W )
0
ϕ11
n (XW ) = Xi+j=n
ϕ22
i (W )δ22
j (X) + δ22
i (X)ϕ22
j (W )
(ϕ11
i (W )δ11
j (X) + δ11
i (X)ϕ11
j (W )),
.
(1)
(2)
ϕ12
n (XW ) = Xi+j=n
(ϕ11
i (W )δ12
j (X) + δ11
i (X)ϕ12
j (W ) + ϕ12
i (W )δ22
j (X) + δ12
i (X)ϕ22
j (W )),
2
and
ϕ22
n (XW ) = Xi+j=n
(ϕ22
i (W )δ22
j (X) + δ22
i (X)ϕ22
j (W ))
for any X ∈ A, W ∈ M. One obtains that
ϕ11
n (W ) = Xi+j=n
n (W ) = Xi+j=n
ϕ22
(ϕ11
i (W )δ11
j (I1) + δ11
i (I1)ϕ11
j (W )),
(ϕ22
i (W )δ22
j (I1) + δ22
i (I1)ϕ22
j (W ))
(3)
(4)
(5)
by taking X = I1 in Eq. (1) and Eq. (3). Now we prove the fact that ϕ11
by induction on n. When n = 0, it is easily verified that ϕ11
characterizations of ϕ11
0 . When n = 1, ϕ11
the proof in [6, Theorem 2.1]. We assume that ϕ11
0 = iδA, we have ϕ11
In fact, by the Eq. (4) and δ11
ϕ11
1 (W ) = 0 and ϕ22
m (W ) = 0 and ϕ22
n (W ) = ϕ11
n (W ) = 0. Similarly combining Eq. (5) with the fact that δ22
n (W ) = 0 and ϕ22
n (W ) = 0
0 (W ) = 0 from the
1 (W ) = 0 can be obtained by
m (W ) = 0 for all 1 ≤ m < n.
n (W ). Thus
n (W ) = 0
n (W ) = 2ϕ11
0 = 0, we can get ϕ22
0 (W ) = 0 and ϕ22
n (W ) + ϕ11
0 and ϕ22
for any W ∈ M and n ∈ N . For any X ∈ A, W ∈ M and Y ∈ B, setting S = (cid:20) 0 W
T =(cid:20) X 0
0 (cid:21), then ST = 0, T S =(cid:20) 0 XW
0
0
0 Y (cid:21) and
(6)
(7)
(8)
(9)
for any X ∈ A, W ∈ M. One can see that
+δ11
i (X)ϕ12
j (W ) + δ11
i (X)τ 12
j (Y ) + δ12
i (X)τ 22
j (Y )
(τ 11
i (Y )δ11
j (I1) + δ11
i (I1)τ 11
j (Y )) = 0
Xi+j=n
by taking X = I1 in Eq. (6). Using Eq. (9) and induction, one has τ 11
Similarly taking Y = I2 in Eq. (7), by inducting and using the fact that τ 22
δ22
n (X) = 0 for every n ∈ N and X ∈ A.
n (Y ) = 0 for every n ∈ N .
0 (Y ) = 0, we get
3
(Di(S)Dj(T ) + Di(T )Dj(S))
0
(cid:21). One gets
(cid:21) = Dn(ST + T S) = Pi+j=n
(cid:21)(cid:20) δ11
i (W ) + τ 12
τ 22
i (Y )
i (Y )
0
j (X)
δ12
i (X)
δ22
i (X) (cid:21)(cid:20) τ 11
0
j (Y ) ϕ12
j (W ) + τ 12
j (Y )
τ 22
j (Y )
δ12
j (X)
δ22
j (X) (cid:21)
(cid:21)).
(τ 11
i (Y )δ11
j (X) + δ11
i (X)τ 11
j (Y )) = 0,
(τ 22
i (Y )δ22
j (X) + δ22
i (X)τ 22
j (Y )) = 0,
(τ 11
i (Y )δ12
j (X) + ϕ12
i (W )δ22
j (X) + τ 12
i (Y )δ22
j (X)
Hence the following three equations hold
n (XW )
0
0
(cid:20) 0 ϕ12
((cid:20) τ 11
0
= Pi+j=n
+(cid:20) δ11
i (X)
0
i (Y ) ϕ12
Xi+j=n
Xi+j=n
n (XW ) = Pi+j=n
ϕ12
We can obtain that
Xi+j=n
(δ11
i (X)τ 12
j (Y ) + δ12
i (X)τ 22
j (Y )) = 0
(10)
by δ22
i (X) = 0, τ 11
By Eq. (2) and the fact that δ22
i (Y ) = 0 and taking W = 0 in Eq. (8).
n (X) = 0, ϕ11
n (W ) = 0, ϕ22
n (W ) = 0 and ϕ12
0 = iϕM, we have
ϕ12
n (XW ) = Xi+j=n
δ11
i (X)ϕ12
j (W ).
(11)
We claim that δ = {δ11
a derivation by Theorem 2.1 in [6]. It follows that δ11
X1, X2 in A. Now we assume that δ11
n : n ∈ N } is a higher derivation on A. In fact, we know that δ1 is
1 (X1)X2 + X1δ11
1 (X2) for any
j (X2) for any 1 ≤ m < n with
1 (X1X2) = δ11
i (X1)δ11
δ11
m ∈ N . Summing up Eq. (11) and ϕ12
m (X1X2) = Pi+j=m
0 = iϕM , we get
ϕ12
n (X1(X2W )) = Pi+j=n
δ11
i (X1)ϕ12
j (X2W )
δ11
i (X1)δ11
e (X2)W +
δ11
i (X1)δ11
e (X2)ϕ12
k (W )
= Pi+e=n
Pi+e+k=n,k>0
for any X1, X2 ∈ A and W ∈ M. On the other hand
ϕ12
n ((X1X2)W ) = Pi+j=n,j>0
Pe+k+j=n,j>0
e (X1)δ11
δ11
=
δ11
i (X1X2)ϕ12
j (W ) + δ11
n (X1X2)W
k (X2)ϕ12
j (W ) + δ11
n (X1X2)W
(12)
(13)
Pe+i=n
δ = {δ11
for any X1, X2 ∈ A and W ∈ M. Combining Eq. (12) with Eq. (13), we get [δ11
δ11
i (X1)δ11
e (X2)]W = 0. Since M is faithful, we get δ11
δ11
i (X1)δ11
n (X1X2) −
j (X2), i.e.
n : n ∈ N } is a higher derivation.
Letting S =(cid:20) 0 −X −1W Y
Y
0
(cid:21) and T =(cid:20) X W
0 (cid:21) for any Y ∈ B, W ∈ M, and invertible
0
X ∈ A. Then ST = T S = 0. So we get
n (X1X2) = Pi+j=n
(Di(S)Dj(T ) + Di(T )Dj(S))
(cid:20) 0 0
0 0 (cid:21) = Dn(ST + T S) = Pi+j=n
((cid:20) 0 −ϕ12
= Pi+j=n
+(cid:20) δ11
i (X −1W Y ) + τ 12
δ12
i (X) + ϕ12
τ 22
i (Y )
i (W )
i (X)
0
0
0
i (Y )
(cid:21)(cid:20) δ11
(cid:21)(cid:20) 0 −ϕ12
0
j (X)
0
δ12
j (X) + ϕ12
j (W )
0
(cid:21)
j (X −1W Y ) + τ 12
j (Y )
τ 22
j (Y )
(cid:21)).
The above equation implies that
[δ11
i (X)(−ϕ12
j (X −1W Y ) + τ 12
j (Y )) + (δ12
i (X) + ϕ12
i (W ))τ 22
j (Y )].
(14)
0 = Xi+j=n
4
By replacing W by λW in the above equation, dividing the equation by λ and letting λ → +∞,
we obtain that
[−δ11
i (X)ϕ12
j (X −1W Y ) + ϕ12
i (W )τ 22
j (Y )].
So we can get
[−δ11
i (I1)ϕ12
j (W Y ) + ϕ12
i (W )τ 22
j (Y )]
0 = Xi+j=n
0 = Xi+j=n
(15)
(16)
by setting X = I1 in the above equation. Since δ = {δ11
when n ≥ 1. It follows from Eq. (16) that
n : n ∈ N } is a higher derivation, δ11
n (I1) = 0
ϕ12
n (W Y ) = Xi+j=n
ϕ12
i (W )τ 22
j (Y ).
(17)
We claim that τ = {τ 22
Theorem 2.1] we know that τ1 is a higher derivation. This implies that τ 22
Y1τ 22
n : n ∈ N } is a higher derivation on B. In fact, by the proof of [6,
1 (Y1)Y2 +
j (Y2) for all
1 (Y2) for any Y1, Y2 ∈ B. We now assume that τ 22
1 (Y1Y2) = τ 22
τ 22
i (Y1)τ 22
1 ≤ m < n with m ∈ N . It follows from Eq. (17) that
m (Y1Y2) = Pi+j=m
ϕ12
n (W Y1Y2) = ϕ12
n (W (Y1Y2))
= W τ 22
= W τ 22
n (Y1Y2) + Pi+j=n,j<n
Pi+e+k=n,i>0
n (Y1Y2) +
ϕ12
i (W )τ 22
j (Y1Y2)
ϕ12
i (W )τ 22
e (Y1)τ 22
k (Y2)
(18)
for any Y1, Y2 ∈ B and W ∈ M. On the other hand by Eq. (17) and the fact that M is a (A,
B)-bimodule, we have
ϕ12
n (W Y1Y2) = ϕ12
n ((W Y1)Y2)
ϕ12
Pe+k+j=n,e>0
n (Y1Y2) − Pk+j=n
ϕ12
i (W Y1)τ 22
j (Y2) = Pe+k+j=n
ϕ12
e (W )τ 22
k (Y1)τ 22
j (Y2)
(19)
τ 22
e (Y1)τ 22
j (Y2) +
e (W )τ 22
k (Y1)τ 22
j (Y2).
= Pi+j=n
= W Pk+j=n
Combining Eq. (18) with Eq. (19), we get W [τ 22
τ 22
e (Y1)τ 22
j (Y2)]W = 0. Since M
is faithful, we get τ 22
τ 22
i (Y1)τ 22
j (Y2).
n (Y1Y2) = Pi+j=n
0
(cid:20) X2 W2
Now we prove that (Dn)n∈N is a higher derivation. For any S = (cid:20) X1 W1
Y1 (cid:21) , T =
Y2 (cid:21) ∈ T , where X1, X2 ∈ A, W1, W2 ∈ M and Y1, Y2 ∈ B. Summing up the above
Dn(ST ) = Dn((cid:20) X1X2 X1W2 + W1Y2
(cid:21))
Y1Y2
results and using the definition of Dn, we obtain that
0
n (X1X2)
δ12
n (X1X2) + ϕ12
n (X1W2 + W1Y2) + τ 12
n (Y1Y2)
0
τ 22
n (Y1Y2)
(cid:21) ,
= (cid:20) δ11
0
5
and
Pi+j=n
(cid:20) δ11
=
Di(S)Dj(T ) = Pi+j=n
i (X1)
((cid:20) δ11
0
δ12
i (X1) + ϕ12
i (W1) + τ 12
τ 22
i (Y1)
i (Y1)
(cid:21)
j (X2)
0
j (W2) + τ 12
j (X2) + ϕ12
δ12
τ 22
j (Y2)
j (Y2)
(cid:21))
δ11
n (X1X2) Pi+j=n
(δ11
i (X1)δ12
j (X2) + δ11
i (X1)τ 12
j (Y2) + δ12
i (X1)τ 22
j (Y2)
+τ 12
i (Y1)τ 22
j (Y2)) + ϕ12
n (X1W2 + W1Y2)
0
τ 22
n (Y1Y2)
by Eq. (17) and the fact that both δ and τ are higher derivations. So D is a higher derivations if
and only if the equation
n (X1X2) + ϕ12
δ12
n (X1W2 + W1Y2) + τ 12
n (Y1Y2)
(δ11
i (X1)δ12
j (X2) + δ11
i (X1)τ 12
j (Y2)
= Pi+j=n
+δ12
i (X1)τ 22
j (Y2) + τ 12
i (Y1)τ 22
j (Y2)) + ϕ12
n (X1W2 + W1Y2)
holds.
We get that τ 22
n (I2) = 0(n ≥ 1) from [4, lemma 2.2]. So we can write
δ12
n (X) = − Xi+j=n
δ11
i (X)τ 12
j (I2)
by setting Y = I2 in Eq. (10). Letting X = I1 in the above equation, one gets δ12
So
n (I1) = −τ 12
n (I2).
Similarly by taking X = I1 in Eq. (10) and noting the fact δ11
n (I1) = 0(n ≥ 1), we have
δ12
n (X) = Pi+j=n
n (Y ) = − Pi+j=n
τ 12
i (X)δ12
δ11
j (I1).
δ12
i (I1)τ 22
j (Y ).
Thus it follows from Eq. (20) and Eq. (21) that
n (X1X2) + τ 12
δ12
n (Y1Y2) = Pi+j=n
δ11
i (X1X2)δ12
j (I1) − Pi+j=n
δ12
i (I1)τ 22
j (Y1Y2)
δ11
k (X1)δ11
l (X2)δ12
j (I1) − Pi+k+l=n
δ12
i (I1)τ 22
k (Y1)τ 22
l (Y2).
(δ11
i (X1)δ12
j (X2) + δ11
i (X1)τ 12
j (Y2) + δ12
i (X1)τ 22
j (Y2) + τ 12
i (Y1)τ 22
j (Y2))
On the other hand
(20)
(21)
(22)
(23)
=
Pk+l+j=n
Pi+j=n
= Pi+j=n Pk+l=j
+ Pi+j=n Pk+l=i
Pi+k+l=n
=
δ11
i (X1)δ11
k (X2)δ12
δ11
i (X1)δ12
k (I1)τ 22
l (Y2)
δ11
k (X1)δ12
l (I1)τ 22
δ11
k (I1)τ 22
l (Y1)τ 22
j (Y2)
l (I1) − Pi+j=n Pk+l=j
j (Y2) − Pi+j=n Pk+l=i
l (I1) − Pj+k+l=n
6
δ11
i (Xk)δ11
k (X2)δ12
δ12
k (I1)τ 22
l (Y1)τ 22
j (Y2).
Thus combining Eq. (22) with Eq. (23), we arrive at
δ12
n (X1X2) + ϕ12
n (X1W2 + W1Y2) + τ 12
n (Y1Y2)
= Pi+j=n
+τ 12
(δ11
i (X1)δ12
j (X2) + δ11
i (X1)τ 12
j (Y2) + δ12
i (X1)τ 22
j (Y2)
i (Y1)τ 22
j (Y2)) + ϕ12
n (X1W2 + W1Y2).
Finally we obtain the desired result.
Theorem 2.2 Let D = {Dn} be a family of additive mappings on T that D0 = iDT . If D is
Jordan higher derivable at G =" I1 X0
Proof. We set S = (cid:20) X 0
and Y ∈ B. Then ST = G and T S =(cid:20) I1 X −1X0Y
Y (cid:21) and T = (cid:20) X −1 X −1X0
Y −1
I2
0
0
0
0
I2 #, then D is a higher derivation.
(cid:21), so we obtain
(cid:21) for every invertible element X ∈ A
2δ12
+ϕ12
n (I1) + 2τ 12
n (I2)+
n (X0 + X −1X0Y )
2δ22
n (I1) + ϕ22
n (X0 + X −1X0Y ) + 2τ 22
n (I2)
(Di(S)Dj(T ) + Di(T )Dj(S))
δ12
j (X) + τ 12
δ22
j (X) + τ 22
j (Y )
j (Y ) (cid:21)).
So according to the above matrix equation, we get
2δ11
n (I1) + 2τ 11
n (I2) + ϕ11
n (X0 + X −1X0Y )
[(δ11
i (X) + τ 11
i (Y ))(δ11
j (X −1) + ϕ11
j (X −1X0) + τ 11
j (Y −1))
= Pi+j=n
+(δ11
i (X −1) + ϕ11
i (X −1X0) + τ 11
i (Y −1))(δ11
j (X) + τ 11
j (Y ))],
7
2δ11
+ϕ11
n (I1) + 2τ 11
n (X0 + X −1X0Y )
n (I2)
0
0
i (X) + τ 11
((cid:20) δ11
j (X −1) + ϕ11
δ11
= Dn(ST + T S) = Pi+j=n
= Pi+j=n
+
(cid:20) δ11
i (X −1) + ϕ11
δ11
j (X) + τ 11
j (Y −1)
j (Y )
+τ 11
+τ 11
i (Y −1)
0
0
0
i (Y )
δ12
i (X) + τ 12
i (X) + τ 22
δ22
i (Y )
i (Y ) (cid:21)
j (X −1X0)
j (X −1) + ϕ12
δ12
j (X −1X0)
+τ 12
j (Y −1)
j (X −1) + ϕ22
δ22
j (X −1X0) + τ 22
j (Y −1)
i (X −1) + ϕ12
δ12
i (X −1X0)
+τ 12
i (Y −1)
i (X −1) + ϕ22
δ22
i (X −1X0) + τ 22
i (Y −1)
i (X −1X0)
(24)
2δ12
n (I1) + 2τ 12
n (I2) + ϕ12
n (X0 + X −1X0Y )
[(δ11
i (X) + τ 11
i (Y ))(δ12
j (X −1) + ϕ12
j (X −1X0) + τ 12
j (Y −1))
= Pi+j=n
+(δ12
i (X) + τ 12
i (Y ))(δ22
j (X −1) + ϕ22
j (X −1X0) + τ 22
j (Y −1))
+(δ11
i (X −1) + ϕ11
i (X −1X0) + τ 11
i (Y −1))(δ12
j (X) + τ 12
j (Y ))
+(δ12
i (X −1) + ϕ12
i (X −1X0) + τ 12
i (Y −1))(δ22
j (X) + τ 22
j (Y ))],
2δ22
n (I1) + 2τ 22
n (I2) + ϕ22
n (X0 + X −1X0Y )
[(δ22
i (X) + τ 22
i (Y ))(δ22
j (X −1) + ϕ22
j (X −1X0) + τ 22
j (Y −1))
= Pi+j=n
i (X −1) + ϕ22
i (X −1X0) + τ 22
i (Y −1))(δ22
j (X) + τ 22
j (Y ))].
+(δ22
n (I1) = τ 11
We claim that δ11
n (I2) = ϕ11
n (X0) = 0 when n ≥ 1 . In fact, we could obtain
2δ11
n (I1) + 2τ 11
n (I2) + ϕ11
n (X0 + X0)
[(δ11
i (I1) + τ 11
i (I2))(δ11
j (I1) + ϕ11
j (X0) + τ 11
j (I2))
= Pi+j=n
+(δ11
i (I1) + ϕ11
i (X0) + τ 11
i (I2))(δ11
j (I1) + τ 11
j (I2))]
(25)
(26)
(27)
by setting X = I1 and Y = I2 in Eq. (24). When n = 1, the result that δ11
1 (X0) = 0 holds according to the [6, Theorem 2.2]. So we assume that δ11
ϕ11
m (X0) = 0 for all 1 ≤ m < n, m ∈ N . Combining Eq. (27) with the fact δ11
ϕ11
and using the induction hypothesis, we have
1 (I1) = τ 11
m (I1) = τ 11
1 (I2) =
m (I2) =
0 (I2) = 0
0 (I1) = I1, τ 11
2δ11
n (I1) + 2τ 11
n (I2) + 2ϕ11
n (X0) = δ11
n (I1) + τ 11
n (I2) + δ11
n (I1) + τ 11
n (I2)
+2δ11
n (I1) + 2τ 11
n (I2) + 2ϕ11
n (X0).
n (I1) + τ 11
Hence δ11
(24). Using the induction hypothesis, we get δ11
above equations we get 2δ11
n (I2) = ϕ11
n (I2) = 0(n ≥ 1). Similarly we also can set that X = I1 and Y = −I2 in Eq.
n (X0). Summing up the
n (I1) − τ 11
n (X0).
n (I1) = −2τ 11
n (I2) = −ϕ11
Setting X = 1
2 I1 and Y = I2 in Eq. (24) and using δ11
i (I2))(2δ11
i (I1) + τ 11
n (I1) + τ 11
j (I1) + τ 11
n (I2) = 0, we have
j (I2) + 2ϕ11
j (X0))
2 δ11
[( 1
3ϕ11
n (X0) = Pi+j=n
+(2δ11
i (I1) + τ 11
i (I2) + 2ϕ11
i (X0))( 1
2 δ11
j (I1) + τ 11
j (I2))].
Thus combining 2δ11
n (I1) = −2τ 11
n (I2) = ϕ11
n (X0) with the assumption and using δ11
0 (I1) = I1,
one obtains
3ϕ11
n (X0) = 1
2 (2δ11
n (I1) + τ 11
n (I2) + 2ϕ11
n (X0))
+2(δ11
n (I1) + τ 11
n (I2)) + 2(δ11
n (I1) + τ 11
n (I2))
.
+ 1
2 (2δ11
n (I1) + τ 11
n (I2) + 2ϕ11
n (X0))
8
n (X0) = 4δ11
So ϕ11
n (I1) + 5τ 11
Eq. (24) can be rewritten into
n (I2). We can claim that δ11
n (I1) = τ 11
n (I2) = ϕ11
n (X0) = 0. Hence the
ϕ11
n (X −1X0Y ) = Pi+j=n
[(δ11
i (X) + τ 11
i (Y ))(ϕ11
j (X −1X0) + τ 11
j (Y −1) + δ11
j (X −1))
(28)
+(δ11
i (X −1) + ϕ11
i (X −1X0) + τ 11
i (Y −1))(δ11
j (X) + τ 11
j (Y ))].
Similarly by setting X = I1 and Y = I2 in Eq. (26) and using the induction, we can get
n (X0) = 0 if we take X = I1 and
n (I2) = 0. We also can obtain δ22
n (I2) = ϕ22
n (I1) = τ 22
δ22
n (I1) + τ 22
Y = 1
2 I2 in Eq. (27). Thus
ϕ22
n (X −1X0Y ) = Pi+j=n
[(δ22
i (X) + τ 22
i (Y ))(ϕ22
j (X −1X0) + τ 22
j (Y −1) + δ22
j (X −1))
(29)
+(δ22
i (X −1) + ϕ22
i (X −1X0) + τ 22
i (Y −1))(δ22
j (X) + τ 22
j (Y ))].
We take X = I1 and Y = I2 in Eq. (25), then we can get δ12
n (I1) + τ 12
n (I2) = 0. Letting
respectively Y = I2 and Y = 1
2 I2 in Eq. (25) and using the above equation we have
ϕ12
n (X0 + X −1X0) = Pi+j=n
[δ11
i (X)(δ12
j (X −1) + ϕ12
j (X −1X0) + τ 12
j (I2))
+(δ12
i (X) + τ 12
i (I2))(δ22
j (X −1) + ϕ22
j (X −1X0))
+(δ11
i (X −1) + ϕ11
i (X −1X0))(δ12
j (X) + τ 12
j (I2))
+(δ12
i (X −1) + ϕ12
i (X −1X0) + τ 12
i (I2))δ22
j (X)]
+δ12
n (X) + τ 12
n (I2) + δ12
n (X −1) + ϕ12
n (X −1X0) + τ 12
n (I2),
ϕ12
n (X0 + 1
2 X −1X0) = Pi+j=n
[δ11
i (X)(δ12
j (X −1) + ϕ12
j (X −1X0) + 2τ 12
j (I2))
+(δ12
i (X) + 1
2 τ 12
i (I2))(δ22
j (X −1) + ϕ22
j (X −1X0))
+(δ11
i (X −1) + ϕ11
i (X −1X0))(δ12
j (X) + 1
2 τ 12
j (I2))
+(δ12
i (X −1) + ϕ12
i (X −1X0) + 2τ 12
i (I2))δ22
j (X)]
(30)
(31)
+2δ12
n (X) + τ 12
n (I2) + 1
2 δ12
n (X −1) + 1
2 ϕ12
n (X −1X0) + τ 12
n (I2),
which implies that
1
2 ϕ12
n (X −1X0) = Pi+j=n
[−δ11
i (X)τ 12
j (I2)
+ 1
2 τ 12
i (I2)(δ22
j (X −1) + ϕ22
j (X −1X0)) + 1
2 (δ11
i (X −1) + ϕ11
i (X −1X0))τ 12
j (I2)
−τ 12
i (I2)δ22
j (X)] − δ12
n (X) + 1
2 δ12
n (X −1) + 1
2 ϕ12
n (X −1X0).
9
So
1
2 Pi+j=n
[τ 12
i (I2)δ22
j (X −1) + δ11
i (X −1)τ 12
j (I2)
+τ 12
i (I2)ϕ22
j (X −1X0) + ϕ11
i (X −1X0)τ 12
j (I2)] + 1
2 δ12
n (X −1)
= Pi+j=n
Thus we get
[δ11
i (X)τ 12
j (I2) + τ 12
i (I2)δ22
j (X)] + δ12
n (X).
[τ 12
i (I2)δ22
j (X) + δ11
i (X)τ 12
j (I2)
1
2 Pi+j=n
+τ 12
i (I2)ϕ22
j (XX0) + ϕ11
i (XX0)τ 12
j (I2)] + 1
2 δ12
n (X)
[δ11
i (X −1)τ 12
j (I2) + τ 12
i (I2)δ22
j (X −1)] + δ12
n (X −1)
= Pi+j=n
for any invertible X ∈ A by replacing X −1 by X in Eq.(32). It follows that
[τ 12
i (I2)δ22
j (X) + δ11
i (X)τ 12
j (I2)
1
2 [ 1
2 Pi+j=n
+τ 12
i (I2)ϕ22
j (XX0) + ϕ11
i (XX0)τ 12
j (I2)] + 1
2 δ12
n (X)]
[τ 12
i (I2)ϕ22
j (X −1X0) + ϕ11
i (X −1X0)τ 12
j (I2)]
[δ11
i (X)τ 12
j (I2) + τ 12
i (I2)δ22
j (X)] + δ12
n (X).
1
+ 1
2 Pi+j=n
= Pi+j=n
4 [ Pi+j=n
4 Pi+j=n
2 Pi+j=n
+ 1
+ 1
So
[τ 12
i (I2)δ22
j (X) + δ11
i (X)τ 12
j (I2)] + δ12
n (X)]
[τ 12
i (I2)ϕ22
j (XX0) + ϕ11
i (XX0)τ 12
j (I2)]
[τ 12
i (I2)ϕ22
j (X −1X0) + ϕ11
i (X −1X0)τ 12
j (I2)]
for any invertible X ∈ A.
= Pi+j=n
[τ 12
i (I2)δ22
j (X) + δ11
i (X)τ 12
j (I2)] + δ12
n (X)
Similarly by letting X = I1 and X = 2I1 in Eq. (25), it is easily checked that
ϕ12
n (X0 + X0Y ) = Pi+j=n
[τ 11
i (Y )(ϕ12
j (X0) + τ 12
j (Y −1) + δ12
j (I1))
+(δ12
i (I1) + τ 12
i (Y ))τ 22
j (Y −1) + τ 11
i (Y −1)(δ12
j (I1) + τ 12
j (Y ))
+(ϕ12
i (X0) + τ 12
i (Y −1) + δ12
i (I1))τ 22
j (Y )]
+ϕ12
n (X0) + τ 12
n (Y −1) + 2δ12
n (I1) + τ 12
n (Y ),
10
(32)
(33)
(34)
(35)
ϕ12
n (X0 + 1
2 X0Y ) = Pi+j=n
[τ 11
i (Y )( 1
2 ϕ12
j (X0) + τ 12
j (Y −1) + 1
2 δ12
j (I1))
+(2δ12
i (I1) + τ 12
i (Y ))τ 22
j (Y −1) + τ 11
i (Y −1)(2δ12
j (I1) + τ 12
j (Y ))
+( 1
2 ϕ12
i (X0) + τ 12
i (Y −1) + 1
2 δ12
i (I1))τ 22
j (Y )]
+ϕ12
n (X0) + 2τ 12
n (Y −1) + 2δ12
n (I1) + 1
2 τ 12
n (Y ),
which implies that
1
2 ϕ12
n (X0Y ) = Pi+j=n
[ 1
2 τ 11
i (Y )(ϕ12
j (X0) + δ12
j (I1)) − δ12
i (I1)τ 22
j (Y −1)
(36)
(37)
−τ 11
i (Y −1)δ12
j (I1) + 1
2 (ϕ12
i (X0) + δ12
i (I1))τ 22
j (Y )] + 1
2 τ 12
n (Y ) − τ 12
n (Y −1).
By considering Eq. (28) and ϕ11
n (X0) = 0 and letting X = I1 and X = 2I1 respectively, it is
easily verified that
[τ 11
i (Y )τ 11
j (Y −1) + τ 11
i (Y −1)τ 11
j (Y )] + 2τ 11
n (Y −1) + 2τ 11
n (Y ),
ϕ11
n (X0Y ) = Pi+j=n
n (X0Y ) = Pi+j=n
1
2 ϕ11
[τ 11
i (Y )τ 11
j (Y −1) + τ 11
i (Y −1)τ 11
j (Y )] + 4τ 11
n (Y −1) + τ 11
n (Y ).
(38)
(39)
1 (Y ) = 0 according to [6, Theorem 2.2]. We assume
m (Y ) = 0 for any Y ∈ B and 1 ≤ m < n. So combining Eq. (38) with Eq. (39) and using
0 (Y ) = 0. When n = 1, τ 11
When n = 0, τ 11
that τ 11
the induction hypothesis, we have
ϕ11
n (X0Y ) = 2τ 11
n (Y −1) + 2τ 11
n (Y ),
1
2
ϕ11
n (X0Y ) = 4τ 11
n (Y −1) + τ 11
n (Y ).
(40)
(41)
By direct computation, one can verify that τ 11
is invertible for any Y ∈ B and τ 11
n (I2) = 0, so τ 11
n (Y ) = 0 for any Y ∈ B .
n (Y −1) = 0. There exists n ∈ N such that nI2 − Y
When n = 0, δ22
n = 1, δ22
respectively Y = I2 and Y = 2I2 in Eq. (29) and using τ 22
1 (X) = 0. So now we assume that δ22
0 (X) = 0 for any X ∈ A. By [6, Theorem 2.2], we can claim that When
m (X) = 0 for all 1 ≤ m < n and X ∈ A. Taking
n (I2) = 0, n ≥ 1, τ 22
0 = iτB we have
ϕ22
n (X −1X0) = Pi+j=n
[δ22
i (X)(δ22
j (X −1) + ϕ22
j (X −1X0))
+(ϕ22
i (X −1X0) + δ22
i (X −1))δ22
j (X)]
+2δ22
n (X) + 2ϕ22
n (X −1X0) + 2δ22
n (X −1),
and
2ϕ22
n (X −1X0) = Pi+j=n
[δ22
i (X)(δ22
j (X −1) + ϕ22
j (X −1X0))
+(ϕ22
i (X −1X0) + δ22
i (X −1))δ22
j (X)]
+δ22
n (X) + 4ϕ22
n (X −1X0) + 4δ22
n (X −1).
11
(42)
(43)
Combining the assumption and the above equations, we have the following equations:
−ϕ22
n (X −1X0) = 2δ22
n (X) + 2δ22
n (X −1),
−2ϕ22
n (X −1X0) = δ22
n (X) + 4δ22
n (X −1).
By direct computation, one can verify that δ22
n (X) = 0 for any invertible X ∈ A and n ∈ N .
Because there is some integer n such that nI1 − X is invertible for every X ∈ A, the conclusion of
δ22
n (X) = 0 holds for every X ∈ A.
We set S = (cid:20) X XW
Y
0
(cid:21) and T = (cid:20) X −1 X −1X0 − W Y −1
Y −1
0
M, and for any invertible X ∈ A, then ST = G and T S = (cid:20) I1 X −1X0Y
I2
0
δ12
n (I1) + τ 12
n (I2) = 0 with the characterization of D, we obtain the following when n ≥ 1
(cid:21) for any Y ∈ B, W ∈
(cid:21). So combining
n (X −1X0Y ) ϕ12
(cid:20) ϕ11
0
n (X0 + X −1X0Y )
ϕ22
n (X −1X0Y )
(cid:21)
(Di(S)Dj(T ) + Di(T )Dj(S))
i (XW )
δ12
i (X) + ϕ12
i (XW ) + τ 12
i (XW )
τ 22
i (Y ) + ϕ22
i (Y )
(cid:21)
j (X −1X0 − W Y −1)
j (X −1) + ϕ12
δ12
j (Y −1) + ϕ22
τ 22
j (X −1X0 − W Y −1) + τ 12
j (X −1X0 − W Y −1)
j (Y )
(cid:21)
i (X −1X0 − W Y −1)
i (X −1) + ϕ12
δ12
i (Y −1) + ϕ22
τ 22
i (X −1X0 − W Y −1) + τ 12
i (X −1X0 − W Y −1)
i (Y )
0
i (X) + ϕ11
j (X −1) + ϕ11
((cid:20) δ11
= Dn(ST + T S) = Pi+j=n
= Pi+j=n
(cid:20) δ11
+(cid:20) δ11
(cid:20) δ11
i (X −1) + ϕ11
j (X) + ϕ11
j (XW )
0
0
0
δ12
j (X) + ϕ12
j (XW ) + τ 12
j (XW )
j (Y ) + ϕ22
τ 22
j (Y )
(cid:21) ,
which implies the following three equations
ϕ11
n (X −1X0Y ) = Pi+j=n
[(δ11
i (X) + ϕ11
i (XW ))(δ11
j (X −1) + ϕ11
j (X −1X0 − W Y −1))
(δ11
i (X −1) + ϕ11
i (X −1X0 − W Y −1))(δ11
j (X) + ϕ11
j (XW ))],
(cid:21)
(44)
ϕ12
n (X0 + X −1X0Y ) = Pi+j=n
[(δ11
i (X) + ϕ11
i (XW ))(δ12
j (X −1) + ϕ12
j (X −1X0 − W Y −1) + τ 12
j (Y −1))
+(δ12
i (X) + ϕ12
i (XW ) + τ 12
i (Y ))(τ 22
j (Y −1) + ϕ22
j (X −1X0 − W Y −1))
+(δ11
i (X −1) + ϕ11
i (X −1X0 − W Y −1))(δ12
j (X) + ϕ12
j (XW ) + τ 12
j (Y ))
+(δ12
i (X −1) + ϕ12
i (X −1X0 − W Y −1) + τ 12
i (Y −1))(τ 22
j (Y ) + ϕ22
j (XW ))],
(45)
12
ϕ22
n (X −1X0Y ) = Pi+j=n
[(τ 22
i (Y ) + ϕ22
i (XW ))(τ 22
j (Y −1) + ϕ22
j (X −1X0 − W Y −1))
(46)
+(τ 22
i (Y −1) + ϕ22
i (X −1X0 − W Y −1))(τ 22
j (Y ) + ϕ22
j (XW ))].
Now we take X = 2I1 and Y = I2 in Eq. (44) and Eq. (46), it is checked that
1
2 ϕ11
n (X0) = Pi+j=n
[(2δ11
i (I1) + 2ϕ11
i (W ))( 1
2 δ11
j (I1) + ϕ11
j ( 1
2 X0 − W ))
( 1
2 δ11
i (I1) + ϕ11
i ( 1
2 X0 − W ))(2δ11
j (I1) + 2ϕ11
j (W ))],
1
2 ϕ22
n (X0) = Pi+j=n
[(τ 22
i (I2) + 2ϕ22
i (W ))(τ 22
j (I2) + ϕ22
j ( 1
2 X0 − W ))
+(τ 22
i (I2) + ϕ22
i ( 1
2 X0 − W ))(τ 22
j (I2) + 2ϕ22
j (W ))].
By the fact that δ11
any n ≥ 0, it follows that
n (I1) = 0(n ≥ 1), τ 22
n (I2) = 0(n ≥ 1) and ϕ11
n (X0) = 0, ϕ22
n (X0) = 0 for
0 = 2ϕ11
0 = 2ϕ22
n (W ) + 4 Xi+j=n
n (W ) + 4 Xi+j=n
ϕ11
i (W )ϕ11
j (W ),
ϕ22
i (W )ϕ22
j (W ).
When n = 0, ϕ11
m (W ) = ϕ22
that ϕ11
the assumption, we get that ϕ11
0 (W ) = ϕ22
1 (W ) = 0, So we assume
m (W ) = 0 for all 1 ≤ m < n and W ∈ M. Combining the above equation with
0 (W ) = 0, When n = 1, ϕ11
1 (W ) = ϕ22
n (W ) = ϕ22
n (W ) = 0 for all 1 ≤ m < n.
By setting respectively Y = 1
2 I2 and Y = I2 in Eq. (45), the following two equations hold
ϕ12
n (X0 + 1
2 X −1X0) = Pi+j=n
[δ11
i (X)(δ12
j (X −1) + ϕ12
j (X −1X0 − 2W ) + 2τ 12
j (I2))
+δ11
i (X −1)(δ12
j (X) + ϕ12
j (XW ) + 1
2 τ 12
j (I2))] + 2δ12
n (X)
+2ϕ12
n (XW ) + τ 12
n (I2) + 1
2 δ11
n (X −1) + 1
2 ϕ12
n (X −1X0 − 2W ) + τ 12
n (I2),
ϕ12
n (X0 + X −1X0) = Pi+j=n
[δ11
i (X)(δ12
j (X −1) + ϕ12
j (X −1X0 − W ) + τ 12
j (I2))
+δ11
i (X −1)(δ12
j (X) + ϕ12
j (XW ) + τ 12
j (I2))] + δ12
n (X)
+ϕ12
n (XW ) + τ 12
n (I2) + δ11
n (X −1) + ϕ12
n (X −1X0 − W ) + τ 12
n (I2).
Which implies that
− 1
2 ϕ12
n (X −1X0) = Pi+j=n
[−δ11
i (X)ϕ12
j (W ) + δ11
i (X)τ 12
j (I2)
+ 1
2 δ11
i (X −1)τ 12
j (I2)] + δ12
n (X)
+ϕ12
n (XW ) − 1
2 δ11
n (X −1) − 1
2 ϕ12
n (X −1X0).
13
(47)
(48)
(49)
It follows from Eq. (34) and the fact δ22
n (X) = ϕ11
n (W ) = ϕ22
n (W ) = 0, we have
δ12
n (X) = − Xi+j=n
δ11
i (X)τ 12
j (I2).
(50)
Hence combing Eq. (49) with Eq. (50), we can see that
ϕ12
n (XW ) = Xi+j=n
δ11
i (X)ϕ12
j (W )
for any invertible X ∈ A. There exists some n ∈ N such that nI1 − X is invertible for every
X ∈ A, one can check that
ϕ12
n (XW ) = Xi+j=n
δ11
i (X)ϕ12
j (W )
for any X ∈ A.
Now we take respectively X = I1 and X = 2I1 in Eq. (45), one gets
ϕ12
n (X0 + X0Y ) = Pi+j=n
[(δ12
i (I1) + ϕ12
i (W ) + τ 12
i (Y ))τ 22
j (Y −1)
+(δ12
i (I1) + ϕ12
i (X0 − W Y −1) + τ 12
i (Y −1))τ 22
j (Y )] + δ12
n (I1)
+ϕ12
n (X0 − W Y −1) + τ 12
n (Y −1) + δ12
n (I1) + τ 12
n (Y ) + ϕ12
n (W ),
ϕ12
n (X0 + 1
2 X0Y ) = Pi+j=n
[(2δ12
i (I1) + 2ϕ12
i (W ) + τ 12
i (Y ))τ 22
j (Y −1)
+( 1
2 δ12
i (I1) + ϕ12
i ( 1
2 X0 − W Y −1) + τ 12
i (Y −1))τ 22
j (Y )] + δ12
n (I1)
+2ϕ12
n ( 1
2 X0 − W Y −1) + 2τ 12
n (Y −1) + δ12
n (I1) + 1
2 τ 12
n (Y ) + ϕ12
n (W ),
which implies that
1
2 ϕ12
n (X0Y ) = Pi+j=n
[−(δ12
i (I1) + ϕ12
i (W ))τ 22
j (Y −1)
+ 1
2 τ 12
n (Y ).
2 (δ12
i (I1) + ϕ12
i (X0))τ 22
j (Y )] + ϕ12
n (W Y −1) − τ 12
n (Y −1) + 1
Combining the above equation with Eq. (37) and the fact τ 11
n (Y ) = 0, we get
[−δ12
i (I1)τ 22
j (Y −1) + 1
2 δ12
i (I1)τ 22
j (Y ) + 1
2 ϕ12
i (X0)τ 22
j (Y )] + 1
2 τ 12
n (Y ) − τ 12
n (Y −1)
Pi+j=n
= Pi+j=n
− Pi+j=n
[−δ12
i (I1)τ 22
j (Y −1) + 1
2 δ12
i (I1)τ 22
j (Y ) + 1
2 ϕ12
i (X0)τ 22
j (Y )]
ϕ12
i (W )τ 22
j (Y −1) + 1
2 τ 12
n (Y ) − τ 12
n (Y −1) + ϕ12
n (W Y −1).
So
ϕ12
n (W Y −1) = Xi+j=n
14
ϕ12
i (W )τ 22
j (Y −1).
(51)
(52)
(53)
(54)
(55)
(56)
Replacing Y by Y −1 in the above equation, we obtain for any invertible Y ∈ B
ϕ12
n (W Y ) = Xi+j=n
ϕ12
i (W )τ 22
j (Y ).
(57)
Since there is some integer n such that nI2 − Y is invertible for every Y ∈ B, it is easy to see that
Eq. (57) is true for every Y ∈ B and W ∈ M, Summing up Eq. (54) and Eq. (56), we obtain that
Thus
Xi+j=n
Xi+j=n
δ12
i (I1)τ 22
j (Y −1) + τ 12
n (Y −1) =
1
2
[ Xi+j=n
δ12
i (I1)τ 22
j (Y ) + τ 12
n (Y )].
δ12
i (I1)τ 22
j (Y ) + τ 12
n (Y ) =
1
2
[ Xi+j=n
δ12
i (I1)τ 22
j (Y −1) + τ 12
n (Y −1)]
by replacing Y −1 by Y in the Eq. (58). Combining Eq. (58) with Eq. (59), we can obtain
1
2
[ Xi+j=n
δ12
i (I1)τ 22
j (Y −1) + τ 12
n (Y −1)] = 2[ Xi+j=n
δ12
i (I1)τ 22
j (Y −1) + τ 12
n (Y −1)].
So using the direct computation, we can claim that
τ 12
n (Y ) = − Xi+j=n
δ12
i (I1)τ 22
j (Y ).
(58)
(59)
(60)
Now summing up all the above equations and using similar arguments as that in the proof of
n }n∈N are higher derivations. There-
Theorem 2.1, it is easily checked that both {δ11
fore it is also an easy computation to see that {Dn}n∈N is a higher derivation. ✷
n }n∈N and {τ 22
Reference
[1 ] R.L. An, J.C. Hou, Characterization of derivations on triangular rings: Additive maps derivable at
idempotents, 431 (2009) 1070-1080.
[2 ] R.L. An, J.C. Hou, Additivity of Jordan multiplicative maps on Jordan operator algebras. Taiwanese
J. Math. 10 (2006) 45 -- 64.
[3 ] W. Jing, On Jordan all-derivable points of B(H), Linear Algebra Appl. 430 (2009) 941-946.
[4 ] Z.K. Xiao, F. Wei, Higher derivations of triangular algebras and its generations, Linear Algebra Appl.
432(2010) 2615-2622.
[5 ] J. Zhang, W. Yu, Jordan derivations of triangular algebras, Linear Algebra Appl. 419 (2006) 251-255.
[6 ] S. Zhao, J. Zhu, Jordan all-derivable points in the algebra of all upper triangular matrices, Linear
Algebra Appl. 433 (2010) 1922-1938.
[7 ] J. Zhu, All-derivable points of operator algebras, Linear Algebra Appl. 427 (2007) 1-5.
15
|
1605.06057 | 2 | 1605 | 2017-02-24T16:58:45 | Classification of a family of non almost periodic free Araki-Woods factors | [
"math.OA"
] | We obtain a complete classification of a large class of non almost periodic free Araki-Woods factors $\Gamma(\mu,m)"$ up to isomorphism. We do this by showing that free Araki-Woods factors $\Gamma(\mu, m)"$ arising from finite symmetric Borel measures $\mu$ on $\mathbf{R}$ whose atomic part $\mu_a$ is nonzero and not concentrated on $\{0\}$ have the joint measure class $\mathcal C(\bigvee_{k \geq 1} \mu^{\ast k})$ as an invariant. Our key technical result is a deformation/rigidity criterion for the unitary conjugacy of two faithful normal states. We use this to also deduce rigidity and classification theorems for free product von Neumann algebras. | math.OA | math |
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC
FREE ARAKI -- WOODS FACTORS
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
Abstract. We obtain a complete classification of a large class of non almost periodic free
Araki -- Woods factors Γ(µ, m)′′ up to isomorphism. We do this by showing that free Araki --
Woods factors Γ(µ, m)′′ arising from finite symmetric Borel measures µ on R whose atomic
part µa is nonzero and not concentrated on {0} have the joint measure class C(Wk≥1 µ∗k) as an
invariant. Our key technical result is a deformation/rigidity criterion for the unitary conjugacy
of two faithful normal states. We use this to also deduce rigidity and classification theorems
for free product von Neumann algebras.
1. Introduction
Free Araki-Woods factors are a free probability analog of the type III hyperfinite factors, just
like free group factors are free probability analogs of the hyperfinite II1 factor. The classification
of hyperfinite type III factors has a beautiful history originating in the work of Powers [Po67].
Through the works of Connes [Co72], Haagerup [Ha85] and Krieger [Kr75], the classification
question was ultimately reduced to the classification of ergodic actions of the additive group of
real numbers R, i.e., to classification of virtual subgroups of R in the sense of Mackey.
Following [Sh96], to every orthogonal representation (Ut)t∈R of R on a real Hilbert space HR
is associated the free Araki -- Woods factor Γ(U, HR)′′. For almost periodic representations, i.e.
when (Ut) is a direct sum of finite dimensional representations, the free Araki -- Woods factors
were completely classified in [Sh96] by Connes' Sd invariant [Co74], which is in this case equal
to the subgroup of (R∗
+, ·) generated by the eigenvalues of (Ut). Beyond the almost periodic
case, the classification of free Araki -- Woods factors is a very intriguing open problem and there
is not even a conjectural classification statement. So far, one could only distinguish between
families of non almost periodic free Araki -- Woods factors by computing their invariants, like
Connes' τ -invariant (see [Sh97b, Sh02]), or by structural properties of their continuous core
(see [Sh02, Ho08b, Ha15]). In this paper, we prove the first complete classification theorem for
a large family of non almost periodic free Araki -- Woods factors.
Orthogonal representations of R are classified by their spectral measure and multiplicity func-
tion. So, to any finite symmetric Borel measure µ on R and to any symmetric Borel multiplicity
function m : R → N ∪ {+∞} (that we always assume to satisfy m ≥ 1 µ-almost everywhere),
we associate the free Araki -- Woods factor Γ(µ, m)′′, which comes equipped with the free quasi-
free state ϕµ,m. The almost periodic case corresponds to µ being an atomic measure and then,
by [Sh96], Γ(µ1, m1)′′ is isomorphic with Γ(µ2, m2)′′ if and only if the sets of atoms of µ1 and
µ2 generate the same subgroup of (R, +).
In this paper, we fully classify the free Araki -- Woods factors in the case where the atomic
part µa is nonzero and not concentrated on {0} and where the continuous part µc satisfies
2010 Mathematics Subject Classification. 46L10, 46L54, 46L36.
Key words and phrases. Free Araki -- Woods factors; Free product von Neumann algebras; Popa's deforma-
tion/rigidity theory; Type III factors.
CH is supported by ERC Starting Grant GAN 637601.
DS is supported by NSF Grant DMS-1500035.
SV is supported by ERC Consolidator Grant 614195, and by long term structural funding -- Methusalem
grant of the Flemish Government.
1
2
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
µc ∗ µc ≺ µc. We find in particular that in that case, the free Araki -- Woods factor does not
depend on the multiplicity function m. But we also show that in other cases, Γ(µ, m)′′ does
depend on m.
In order to state our main results, we first introduce some terminology. For every σ-finite Borel
measure µ on R, we denote by C(µ) the measure class of µ, defined as the set of all Borel sets
U ⊂ R with µ(U ) = 0. Note that C(µ) = C(ν) if and only if µ ∼ ν, while C(µ) ⊂ C(ν) if
and only if ν ≺ µ. For any sequence of measures (µk)k∈N, we denote by Wk∈N µk any measure
with the property that C(Wk∈N µk) = Tk∈N C(µk). We denote by µ = µc + µa the unique
decomposition of a measure µ as the sum of a continuous and an atomic measure.
We show that free Araki -- Woods factors Γ(µ, m)′′ arising from finite symmetric Borel measures
µ on R whose atomic part µa is nonzero and not concentrated on {0} have the joint measure
class C(Wk≥1 µ∗k) as an invariant. More precisely, we obtain the following result.
Theorem A. Let µ, ν be finite symmetric Borel measures on R and m, n : R → N ∪ {+∞}
symmetric Borel multiplicity functions. Assume that ν has at least one atom not equal to 0.
If the free Araki -- Woods factors Γ(µ, m)′′ and Γ(ν, n)′′ are isomorphic, then there exists an
isomorphism that preserves the free quasi-free states. In particular, the joint measure classes
C(Wk≥1 µ∗k) and C(Wk≥1 ν∗k) are equal.
Denote by S(R) the set of all finite symmetric Borel measures µ = µc + µa on R satisfying the
following two properties:
(i) µc ∗ µc ≺ µc and
(ii) µa 6= 0 and supp(µa) 6= {0}.
Denote by Λ(µa) the countable subgroup of R generated by the atoms of µa and by δΛ(µa) a
finite atomic measure on R whose set of atoms equals Λ(µa).
Combining our Theorem A with the isomorphism Theorem 4.1 below, we obtain a complete
classification of the free Araki -- Woods factors arising from measures in S(R). Here and else-
where in this paper, we call isomorphism between von Neumann algebras M and N any bijective
∗-isomorphism. Even when M and N are equipped with distinguished faithful normal states,
isomorphisms are not assumed to preserve these states.
Corollary B. The set of free Araki -- Woods factors
(cid:8)Γ(µ, m)′′ : µ ∈ S(R) and m : R → N ∪ {+∞} is a symmetric Borel multiplicity function(cid:9)
is exactly classified, up to isomorphism, by the countable subgroup Λ(µa) and the measure class
C(µc ∗ δΛ(µa)).
Note that the measure class C(µc ∗ δΛ(µa)) equals the set of Borel sets U ⊂ R satisfying
µc(x + U ) = 0 for all x ∈ Λ(µa) and, in particular, does not depend on the choice of δΛ(µa).
The family S(R) is large and provides many nonisomorphic free Araki -- Woods factors having
the same Connes' invariants and in particular the same τ -invariant, see Example 5.4. Note that
previously, only two non almost periodic free Araki -- Woods factors having the same τ -invariant
could be distinguished, see [Sh02, Theorem 5.6].
Combining our Corollary B with [HI15, Theorem A], we also obtain a complete classification
for tensor products of free Araki -- Woods factors arising from measures in S(R).
We then show that free Araki -- Woods factors Γ(µ, m)′′ arising from continuous finite symmetric
Borel measures µ on R have all their centralizers amenable, i.e. the centralizer of any faithful
normal state is amenable. More generally, we obtain the following result.
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
3
Corollary C. Let µ be any finite symmetric Borel measure on R and m : R → N ∪ {+∞}
any symmetric Borel multiplicity function. The free Araki -- Woods factor Γ(µ, m)′′ has all its
centralizers amenable if and only if the atomic part µa of µ is either zero or is concentrated on
{0} with m(0) = 1.
By [Ho08a], all free Araki -- Woods factors M satisfy Connes' bicentralizer conjecture (see [Co80])
and thus, by [Ha85, Theorem 3.1], admit faithful normal states ϕ such that M ϕ ⊂ M is an
irreducible subfactor. So, having all centralizers amenable is the smallest centralizers can be
in general.
In the setting of Corollary C and under the additional assumption that the Fourier transform of
the continuous finite symmetric Borel measure µc vanishes at infinity, it was shown in [Ho08b,
Theorem 1.2] that the continuous core of the corresponding free Araki -- Woods factor Γ(µ, m)′′
is solid (see [Oz03]), meaning that the relative commutant of any diffuse subalgebra that is
the range of a faithful normal conditional expectation is amenable. Any type III1 factor whose
continuous core is solid has all its centralizers amenable. Observe that there are many free
Araki -- Woods factors arising in Corollary C whose Connes' τ -invariant (see [Co74]) is not the
usual topology on R. In particular, these free Araki -- Woods factors have a continuous core that
is not full (see [Co74, Sh02]) and hence not solid (see [Oz03, Proposition 7] with N0 = M).
Therefore, Corollary C provides many new examples of type III1 factors whose centralizers are
all amenable.
The following is an immediate consequence of Corollary C.
Corollary D. Let λ be the Lebesgue measure on R. Then Γ(λ + δ0, 1)′′ 6∼= Γ(λ + δ0, 2)′′. So in
certain cases, the isomorphism class of Γ(µ, m)′′ depends on the multiplicity function m.
Our main technical tool to prove the results mentioned so far is a deformation/rigidity criterion
for the unitary conjugacy of two faithful normal states on a von Neumann algebra M .
In
Corollary 3.2 below, we prove that a corner of the state ψ is unitarily conjugate with a corner
of the state ϕ if and only if in the continuous core c(M ), there is a Popa intertwining bimodule
(in the sense of [Po02, Po03]) between the canonical subalgebras Lψ(R) and Lϕ(R) of c(M )
given by realizing c(M ) as respectively M ⋊σψ R and M ⋊σϕ R (see Section 2 for details).
More generally, when P ⊂ M is a von Neumann subalgebra that is the range of a faithful normal
conditional expectation EP : M → P , we provide in Theorem 3.1 below a deformation/rigidity
criterion describing when a state ψ on M has a corner that is unitarily conjugate with a corner
of a state of the form θ ◦EP . Applying this criterion to a free product von Neumann algebra M ,
we obtain the following complete characterization of when M has all its centralizers amenable.
Theorem E. For i = 1, 2, let (Mi, ϕi) be a von Neumann algebra with a faithful normal
state. Denote by (M, ϕ) = (M1, ϕ1) ∗ (M2, ϕ2) their free product. Then M has all its central-
izers amenable if and only if both M1 and M2 have all their centralizers amenable and M ϕ is
amenable.
Finally, we use the same criterion, in combination with methods of [Sh97a], to prove the
following classification result for a free product with a free Araki -- Woods factor.
Theorem F. Let µ be a continuous finite symmetric Borel measure on R. Fix the free Araki --
Woods factor (M, ϕ) = (Γ(µ, +∞)′′, ϕµ,+∞), where +∞ denotes the multiplicity function equal
to +∞ everywhere.
(i) If (A, τ ) and (B, τ ) are nonamenable II1 factors with their tracial states, then the free
products (M, ϕ) ∗ (A, τ ) and (M, ϕ) ∗ (B, τ ) are isomorphic (not necessarily in a state
preserving way) if and only if there exists a t > 0 such that A ∼= Bt.
4
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
(ii) If (Ai, ψi), i = 1, 2, are full type III factors with almost periodic states having a facto-
rial centralizer Aψi
i , then the free products (M, ϕ) ∗ (A1, ψ1) and (M, ϕ) ∗ (A2, ψ2) are
isomorphic (again, not necessarily in a state preserving way) if and only if A1 ∼= A2.
By Theorem F, the free Araki -- Woods factors Γ((λ, +∞) + (δ0, m))′′ are isomorphic for all
2 ≤ m < +∞, but the question whether these are isomorphic with Γ((λ, +∞) + (δ0, +∞))′′ is
equivalent with the free group factor problem L(Fm) ∼=? L(F∞).
Acknowledgment. We are grateful to the Mittag-Leffler Institute for their hospitality during
the program Classification of operator algebras: complexity, rigidity, and dynamics, where part
of the work on this paper was done.
Disclaimer. Some of the results in this paper, in particular Corollary D, Theorem 4.1 and
Proposition 7.5, were obtained in the preprint [Sh03] by the second named author. That
preprint will remain unpublished and has been incorporated in this article.
Contents
Isomorphisms of free Araki -- Woods factors
Introduction
1.
2. Preliminaries
3. A criterion for the unitary conjugacy of states
4.
5. Proofs of Theorem A and Corollaries B, C, D
6. Proof of Theorem E
7. Further structural results and proof of Theorem F
References
1
4
6
9
13
15
18
23
2. Preliminaries
For any von Neumann algebra M , we denote by U (M ) its group of unitaries. For any (possibly
unbounded) positive selfadjoint closed operator A on a separable Hilbert space H, we denote
by C(A) the measure class on R of the spectral measure of log(A), i.e. the set of all Borel sets
U ⊂ R such that the spectral projection 1U (log(A)) equals 0. Here 1U denotes the function
that is equal to 1 on U and equal to 0 elsewhere. Note that we can always choose a measure µ
on R such that C(A) = C(µ).
Free Araki -- Woods factors. Following [Sh96], we associate to every orthogonal represen-
tation (Ut)t∈R of R on the real Hilbert space HR the free Araki -- Woods factor Γ(HR, Ut)′′,
equipped with the free quasi-free state ϕU .
Denoting by H = HR + iHR the complexification of HR, define the positive nonsingular
operator ∆ on H such that Ut = ∆it for all t ∈ R. Also define the anti-unitary operator
J : H → H : J(ξ + iη) = ξ − iη for all ξ, η ∈ HR. Then, J∆J = ∆−1. Therefore, the measure
class C(∆) of the spectral measure of log(∆) and the multiplicity function m : R → N ∪ {+∞}
of log(∆) are symmetric. This measure class and multiplicity function completely classify
orthogonal representations of R on separable real Hilbert spaces. We therefore use the notation
(Γ(µ, m)′′, ϕµ,m) to denote the free Araki -- Woods factor and its free quasi-free state associated
with the unique orthogonal representation with spectral invariant (µ, m).
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
5
Background on σ-finite von Neumann algebras. Let M be any σ-finite von Neumann
algebra with predual M∗ and ϕ ∈ M∗ any faithful normal state. We denote by σϕ the modular
automorphism group of the state ϕ defined by the formula σϕ
ϕ) for all t ∈ R.
The centralizer M ϕ of the state ϕ is by definition the fixed point algebra of (M, σϕ). The
continuous core of M with respect to ϕ, denoted by cϕ(M ), is the crossed product von Neumann
algebra M ⋊σϕ R. The natural inclusion πϕ : M → cϕ(M ) and the unitary representation
λϕ : R → cϕ(M ) satisfy the covariance relation
t = Ad(∆it
λϕ(t)πϕ(x)λϕ(t)∗ = πϕ(σϕ
t (x))
for all x ∈ M and all t ∈ R.
Put Lϕ(R) := λϕ(R)′′. There is a unique faithful normal conditional expectation ELϕ(R) :
cϕ(M ) → Lϕ(R) satisfying ELϕ(R)(πϕ(x)λϕ(t)) = ϕ(x)λϕ(t) for all x ∈ M and all t ∈ R.
The faithful normal semifinite weight defined by f 7→ RR exp(−s)f (s) ds on L∞(R) gives rise
to a faithful normal semifinite weight Trϕ on Lϕ(R) via the Fourier transform. The formula
Trϕ = Trϕ ◦ ELϕ(R) extends it to a faithful normal semifinite trace on cϕ(M ).
Because of Connes' Radon -- Nikodym cocycle theorem [Co72, Th´eor`eme 1.2.1] (see also [Ta03,
Theorem VIII.3.3]), the semifinite von Neumann algebra cϕ(M ) together with its trace Trϕ
If ψ ∈ M∗ is another
does not depend on the choice of ϕ in the following precise sense.
faithful state, there is a canonical isomorphism Πϕ,ψ : cψ(M ) → cϕ(M ) of cψ(M onto cϕ(M )
such that Πϕ,ψ ◦ πψ = πϕ and Trϕ ◦ Πϕ,ψ = Trψ. Note however that Πϕ,ψ does not map the
subalgebra Lψ(R) ⊂ cψ(M ) onto the subalgebra Lϕ(R) ⊂ cϕ(M ) (and hence we use the symbol
Lϕ(R) instead of the usual L(R)). We have Πϕ,ψ(λψ(t)) = πϕ(wt)λϕ(t) for every t ∈ R, where
wt = [Dψ : Dϕ]t is Connes' Radon -- Nikodym cocycle between ψ and ϕ.
Lemma 2.1. Let M be any σ-finite von Neumann algebra and ϕ ∈ M∗ any faithful state such
that M ϕ is a II1 factor. Let M ϕ ⊂ P ⊂ M be any intermediate von Neumann subalgebra that
is globally invariant under the modular automorphism group σϕ. Denote by EP : M → P the
unique ϕ-preserving conditional expectation and write M ⊖ P := ker(EP ). Let p ∈ M ϕ be any
nonzero projection and put ϕp := ϕ(p · p)
ϕ(p) ∈ (pM p)∗.
Then we have
C(∆ϕ) = C(∆ϕp)
and C(∆ϕL2(M )⊖L2(P )) = C(∆ϕpL2(pM p)⊖L2(pP p)).
Proof. Fix a standard representation M ⊂ B(H) and denote by ξϕ ∈ H the canonical unit
vector that implements ϕ. For every x ∈ M , denote by µϕ
x the unique finite Borel measure on
R that satisfies
ϕ(x∗σϕ
t (x)) = h∆it
ϕ(xξϕ), xξϕi =ZR
exp(ist) dµϕ
x (s)
for all t ∈ R.
For any Borel subset U ⊂ R, we have U ∈ C(∆ϕ) (resp. U ∈ C(∆ϕL2(M )⊖L2(P ))) if and only if
µϕ
x (U ) = 0 for all x ∈ M (resp. for all x ∈ M ⊖P ). Since pM p ⊂ M and p(M ⊖P )p ⊂ M ⊖P , it
is clear that C(∆ϕ) ⊂ C(∆ϕp) (resp. C(∆ϕL2(M )⊖L2(P )) ⊂ C(∆ϕpL2(pM p)⊖L2(pP p))). It remains
to prove that C(∆ϕp) ⊂ C(∆ϕ) (resp. C(∆ϕpL2(pM p)⊖L2(pP p)) ⊂ C(∆ϕL2(M )⊖L2(P ))).
Up to shrinking p ∈ M ϕ if necessary and since we have p2M p2 ⊂ p1M p1 and p2(M ⊖ P )p2 ⊂
p1(M ⊖ P )p1 whenever p2 ≤ p1 with p1, p2 nonzero projections in M ϕ, we may assume without
loss of generality that ϕ(p) = m−1 with m ∈ N. Since M ϕ is a II1 factor, we may find partial
isometries u1, . . . , um ∈ M ϕ such that u1 = p, u∗
j=1 uju∗
j = 1.
Let x ∈ M (resp. x ∈ M ⊖ P ). Since ϕ(x∗σϕ
i,j=1 ϕp((u∗
j xui)) for all
t ∈ R, we have µϕ
j xui ∈ p(M ⊖P )p for all 1 ≤ i, j ≤
m. This implies that C(∆ϕp) ⊂ C(∆ϕ) (resp. C(∆ϕpL2(pM p)⊖L2(pP p)) = C(∆ϕL2(M )⊖L2(P ))) and
finishes the proof.
(cid:3)
j uj = p for all 1 ≤ j ≤ m and Pm
j xui)∗σϕp
t (x)) = ϕ(p)Pm
x = ϕ(p)Pm
t (u∗
i,j=1 µϕp
u∗
j xui
. If x ∈ M ⊖P , then u∗
6
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
Popa's intertwining-by-bimodules. Popa introduced his intertwining-by-bimodules theory
in [Po02, Po03]. In the present work, we make use of this theory in the context of semifinite von
Neumann algebras. We introduce the following terminology. Let M be any σ-finite semifinite
von Neumann algebra endowed with a fixed faithful normal semifinite trace Tr. Let 1A and 1B
be any nonzero projections in M and let A ⊂ 1AM 1A and B ⊂ 1BM 1B be any von Neumann
subalgebras. Assume that Tr(1A) < +∞ and that TrB is semifinite.
We say that A embeds into B inside M and write A ≺M B if there exist a projection e ∈ A,
a finite trace projection f ∈ B, a nonzero partial isometry v ∈ eM f and a unital normal
homomorphism θ : eAe → f Bf such that av = vθ(a) for all a ∈ eAe. We use the following
useful characterization [Po02, Po03] (see also [HR10, Lemma 2.2]).
Theorem 2.2. Keep the same notation as above. Denote by EB : 1BM 1B → B the unique
trace preserving conditional expectation. Then the following conditions are equivalent.
(i) A ≺M B.
(ii) There exists no net (wi)i∈I of unitaries in U (A) such that limi kEB(y∗wix)k2 = 0 for
all x, y ∈ 1AM 1B.
3. A criterion for the unitary conjugacy of states
Recall that for any von Neumann algebra N , any θ ∈ N∗ and any a, b ∈ N , we define (aθb)(y) :=
θ(bya) for every y ∈ N .
Theorem 3.1. Let M be a von Neumann algebra with a faithful normal state ϕ ∈ M∗ and
P ⊂ M a von Neumann subalgebra that is the range of a ϕ-preserving conditional expectation
EP : M → P . Let ψ ∈ M∗ be another faithful normal state and q ∈ M ψ a nonzero projection.
Then the following statements are equivalent.
(i) There exists a nonzero finite trace projection r ∈ Lψ(R) such that
Πϕ,ψ(Lψ(R)qr) ≺cϕ(M ) cϕ(P ) .
(ii) There exist a faithful normal positive functional θ ∈ P∗ and a nonzero partial isometry
v ∈ M such that p = vv∗ ∈ M θ◦EP , q0 = v∗v ∈ qM ψq and ψq0 = v∗(θ ◦ EP )v.
Proof. Assume that (i) holds. Write wt = [Dψ : Dϕ]t. We claim that there exists a δ > 0 and
x1, . . . , xk ∈ qM such that
(3.1)
kXi,j=1
ϕ(cid:0) EP (x∗
i wt σϕ
t (xj)) EP (x∗
i wt σϕ
t (xj))∗ (cid:1) ≥ δ
for all t ∈ R. Assuming that the claim is false, we prove that (i) does not hold. Take a net
ti ∈ R such that
lim
i
ϕ(cid:0) EP (x∗ wti σϕ
ti(y)) EP (x∗ wti σϕ
ti(y))∗ (cid:1) = 0
for all x, y ∈ qM . Using the 2-norm k · k2 w.r.t. the canonical trace on cϕ(M ), we get that for
all finite trace projections p, p′ ∈ Lϕ(R) and all x, y ∈ qM ,
(cid:13)(cid:13)p EP (x∗ wti σϕ
ti(y)) p′(cid:13)(cid:13)2
2 ≤(cid:13)(cid:13)p EP (x∗ wti σϕ
ti(y))(cid:13)(cid:13)2
= Tr(p) ϕ(cid:0) EP (x∗ wti σϕ
2
ti(y)) EP (x∗ wti σϕ
ti(y))∗ (cid:1) → 0 .
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
7
We then also get for all finite trace projections p, p′ ∈ Lϕ(R), all s, s′ ∈ R and all x, y ∈ M
that
(cid:13)(cid:13)Ecϕ(P )(cid:0)p λϕ(s)∗ x∗ Πϕ,ψ(λψ(ti)q) y λϕ(s′) p′(cid:1)(cid:13)(cid:13)2
=(cid:13)(cid:13)λϕ(s)∗ p EP(cid:0)(qx)∗wtiσϕ
=(cid:13)(cid:13)p EP(cid:0)(qx)∗wtiσϕ
ti(qy)(cid:1) p′k2 → 0 .
ti(qy)(cid:1) p′ λϕ(ti + s′)k2
The linear span of such elements xλϕ(s)p is dense in L2(cϕ(M ), Tr). So whenever r ∈ Lψ(R) is
a finite trace projection and a, b ∈ cϕ(M ), we first approximate Πϕ,ψ(r) b in k · k2 by a linear
combination of yλϕ(s′)p′ and conclude that
for all finite trace projections p ∈ Lϕ(R) and all s ∈ R, x ∈ M . We then approximate Πϕ,ψ(r) a
in k · k2 by a linear combination of xλϕ(s)p and conclude that
(cid:13)(cid:13)Ecϕ(P )(cid:0)p λϕ(s)∗ x∗ Πϕ,ψ(λψ(ti)qr) b(cid:1)(cid:13)(cid:13)2 → 0
(cid:13)(cid:13)Ecϕ(P )(cid:0)a∗ Πϕ,ψ(λψ(ti)qr) b(cid:1)(cid:13)(cid:13)2 → 0 .
Applying Theorem 2.2, we conclude that (i) does not hold. This concludes the proof of the
claim.
Fix δ > 0 and x1, . . . , xk such that (3.1) holds for all t ∈ R. Denote by hM, eP i the basic
construction for P ⊂ M , i.e. the von Neumann algebra acting on L2(M, ϕ) generated by M
acting by left multiplication and the orthogonal projection eP : L2(M, ϕ) → L2(P, ϕ). As
in [ILP96, Section 2.1], denote by ϕ the canonical faithful normal semifinite weight on hM, eP i
characterized by
σ ϕ
t (xeP y) = σϕ
t (x)eP σϕ
ϕT ∆−it
t (y) and
ϕ for all T ∈ hM, eP i and all t ∈ R. In particular,
t (x) for all x ∈ M . Denote by TM the unique faithful normal semifinite operator
for all x, y ∈ R. We have σ ϕ
σ ϕ
t (x) = σϕ
valued weight from hM, eP i to M such that ϕ = ϕ ◦ TM . Note that TM (eP ) = 1.
Define ψ = ψ ◦ TM . So, ψ is a faithful normal semifinite weight on hM, eP i and by [Ha77,
Theorem 4.7], we have
ϕ(xeP y) = ϕ(xy)
t (T ) = ∆it
for all t ∈ R. In particular, we get that
[D ψ : D ϕ]t = [Dψ : Dϕ]t = wt
(3.2)
σ
for all x, y ∈ M and t ∈ R.
ψ
t (xeP y) = wtσϕ
t (x)eP σϕ
t (y)w∗
t
Define the positive element X ∈ qhM, eP iq by
X =
kXj=1
xjeP x∗
j .
Also define the normal positive functional Ω ∈ hM, eP i∗ given by
Ω(T ) =
kXi=1
ϕ(eP x∗
i T xieP )
for all T ∈ hM, eP i. Using (3.2) and (3.1), we get for every t ∈ R,
Ω(σ
ψ
t (X)) =
=
kXi,j=1
kXi,j=1
ϕ(cid:0)eP x∗
ϕ(cid:0)EP (x∗
i wtσϕ
t (xj)eP σϕ
t (xj)∗w∗
i wtσϕ
t (xj)) EP (x∗
i wtσϕ
t xieP(cid:1)
t (xj))∗(cid:1) ≥ δ .
8
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
ψ
t (X) : t ∈ R} inside qhM, eP iq. Note that
Define K as the σ-weakly closed convex hull of {σ
kY k ≤ kXk for all Y ∈ K. Also, every Y ∈ K is positive and satisfies ψ(Y ) ≤ ψ(X) < +∞,
by the σ
ψ
t -invariance and σ-weak lower semicontinuity of ψ. We then also have
ψ(Y ∗Y ) = ψ(Y 2) ≤ kY k ψ(Y ) ≤ kXk ψ(X)
ψ
t -invariant, it follows that σ
ψ
t (X0) = X0 for all T . Since Ω(σ
for all Y ∈ K. By [HI15, Lemma 4.4], the image of K in L2(hM, eP i, ψ) is norm closed. So,
there is a unique element X0 ∈ K where the function Y 7→ ψ(Y ∗Y ) attains its minimal value.
ψ
t (X)) ≥ δ
Since this function is σ
t ◦ TM and since TM is
for all t ∈ R, also Ω(X0) ≥ δ, so that X0 6= 0. Since TM ◦ σ
σ-weakly lower semicontinuous, we get that kTM (Y )k ≤ kTM (X)k < +∞ for all Y ∈ K. In
particular, kTM (X0)k < +∞.
Take ε > 0 small enough such that the spectral projection e = 1[ε,+∞)(X0) is nonzero. It follows
ψ
that e is a projection in qhM, eP iq satisfying σ
t (e) = e for all t ∈ R and kTM (e)k < +∞.
By Lemma 3.3 below, we may assume that e ≺ eP inside hM, eP i. Take a partial isometry
V ∈ hM, eP i such that V ∗V = e and V V ∗ ≤ eP . Let p0 ∈ P be the unique projection such
that V V ∗ = p0eP . We get that V = p0V . Since e ≤ q, we also have V = V q.
Since kTM (V ∗V )k = kTM (e)k < +∞, it follows from the push down lemma [ILP96, Proposition
2.2] (where the factoriality assumption is unnecessary) that
ψ
t = σψ
V = eP V = eP TM (eP V ) = eP TM (V ) .
Write v = TM (V ). Then, v ∈ M and V = eP v. By construction, v ∈ p0M q.
Since eP hM, eP ieP = P eP , we can uniquely define ut ∈ P such that
for all t ∈ R. Since V ∗V = e and wtσ ϕ
t ut = σϕ
u∗
t (V ∗V )w∗
t (p0) for all t ∈ R. Also, t 7→ ut is strongly continuous and
utσϕ
t (V wsσ ϕ
uteP = V wtσ ϕ
t (V ∗)
ψ
t (e) = e, we get that utu∗
t = σ
t = p0 and
t (us) eP = uteP σ ϕ
= V wt σ ϕ
= V wt+s σ ϕ
t (useP ) = V wtσ ϕ
t (ws) σ ϕ
t (V ∗V ) σϕ
t+s(V ∗) = ut+seP
t (V ∗) σ ϕ
t+s(V ∗) = V σ
s (V ∗))
ψ
t (e) wtσϕ
t (ws) σ ϕ
t+s(V ∗)
for all s, t ∈ R. So, (ut)t∈R is a 1-cocycle for ϕP . By [Co72, Th´eor`eme 1.2.4] (see also [Ta03,
Theorem VIII.3.21] for a formulation adapted to non faithful states), there is a unique faithful
normal semifinite weight θ on p0P p0 such that [Dθ : DϕP ]t = ut for all t ∈ R. Define the
faithful normal semifinite weight θ1 on p0M p0 by θ1 = θ ◦ EP . By [Ha77, Theorem 4.7], we
have that [Dθ1 : Dϕ]t = ut for all t ∈ R.
Since ut ∈ P , we get that
eP utσϕ
t (v) = uteP σϕ
= V wtσ ϕ
t (v) = uteP σ ϕ
t (V )
ψ
t (V ∗V ) = V σ
t (V ∗V ) wt = V wt = eP vwt .
Applying TM , we conclude that utσϕ
t (v) = vwt for all t ∈ R. Replacing v by its polar
part, we may assume that v ∈ M is a partial isometry such that p1 = vv∗ ∈ (p0M p0)θ1 and
q1 = v∗v ∈ qM ψq. We then get that
[D(v∗θ1v) : Dϕ]t = v∗[Dθ1 : Dϕ]tσϕ
t (v) = v∗utσϕ
t (v) = q1[Dψ : Dϕ]t = [D(ψq1) : Dϕ]t
for all t ∈ R. We conclude that ψq1 = v∗θ1v.
In particular, θ(EP (p1)) = θ1(p1) = ψ(q1) < +∞. Also, σθ
t (p1)) = EP (p1),
for all t ∈ R. We can thus find a nonzero spectral projection f of EP (p1) such that f v 6= 0,
t (EP (p1)) = EP (σθ1
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
9
θ(f ) < +∞ and f ∈ (p0P p0)θ. Since utσϕ
get that
t (v) = vwt, [Dθ : DϕP ]t = ut and f ∈ (p0P p0)θ, we
utσϕ
t (f v) = f utσϕ
t (v) = f vwt .
We then replace θ by θf and v by the polar part of f v. Then, θ is a faithful normal positive
functional on f P f , the projection p = vv∗ belongs to (f M f )θ◦EP , the projection q0 = v∗v
belongs to qM ψq and ψq0 = v∗(θ ◦ EP )v. Adding to θ an arbitrary faithful normal state on
(1 − f )P (1 − f ), it follows that (ii) holds.
Conversely, assume that (ii) holds. Take θ, q0, p and v as in the statement of (ii). Define
wt = [Dψ : Dϕ]t and ut = [Dθ : DϕP ]t. Since ψq0 = v∗(θ ◦ EP )v, we get that vwt = utσϕ
t (v)
for all t ∈ R. This means that
for all t ∈ R. Also, v Πϕ,ψ(qr) = v Πϕ,ψ(r) 6= 0 for every nonzero finite trace projection
r ∈ Lψ(R). We conclude that
v Πϕ,ψ(λψ(t)) = ΠϕP ,θ(λθ(t)) v
Πϕ,ψ(Lψ(R)qr) ≺cϕ(M ) cϕ(P )
for every nonzero finite trace projection r ∈ Lψ(R), so that (i) holds.
(cid:3)
Applying Theorem 3.1 to the case P = C1, we get the following result.
Corollary 3.2. Let M be a von Neumann algebra with faithful normal states ψ, ϕ ∈ M∗ and
let q ∈ M ψ be a nonzero projection. Then the following statements are equivalent.
(i) There exists a nonzero finite trace projection r ∈ Lψ(R) such that
Πϕ,ψ(Lψ(R)qr) ≺cϕ(M ) Lϕ(R) .
(ii) There exists a nonzero partial isometry v ∈ M such that p = vv∗ belongs to M ϕ,
q0 = v∗v belongs to qM ψq, and
1
ψ(q0)
ψq0 =
1
ϕ(p)
v∗ϕv .
Lemma 3.3. Let ψ be a faithful normal semifinite weight on a von Neumann algebra N and
e ∈ N ψ a projection satisfying 0 < ψ(e) < +∞. Let e1 ∈ N be any projection with central
support equal to 1. Then there exists a nonzero projection e0 ∈ eN ψe satisfying e0 ≺ e1 inside
N .
Proof. Since the central support of e1 equals 1, we can find a nonzero projection f ∈ N such
that f ≤ e and f ≺ e1. Define the faithful normal state θ on eN e given by θ(x) = ψ(e)−1ψ(x)
for all x ∈ eN e. By [HU15, Lemma 2.1], there exists a projection e0 ∈ (eN e)θ such that e0 ∼ f
inside eN e. Then, e0 is a nonzero projection in eN ψe and e0 ≺ e1 inside N .
(cid:3)
4. Isomorphisms of free Araki -- Woods factors
The isomorphism part of Corollary B follows from the following result that we deduce from
[Sh96].
Theorem 4.1. Let µ be any finite symmetric Borel measure on R and m : R → N ∪ {+∞}
any symmetric Borel multiplicity function. Denote by Λ the subgroup of R generated by the
atoms of µ and assume that Λ 6= {0}. There is an isomorphism
Γ(µ, m)′′ ∼= Γ(µ ∗ δΛ, +∞)′′
preserving the free quasi-free states, where δΛ denotes any atomic finite symmetric Borel mea-
sure on R with set of atoms equal to Λ.
10
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
Proof. For every 0 < a < 1, we denote by Ba the von Neumann algebra B(ℓ2(N)) equipped
with the faithful normal state θa given by
θa(T ) = (1 − a)
∞Xk=0
akhT (δk), δki ,
where (δk)k∈N is the standard orthonormal basis of ℓ2(N). Throughout the proof of the theo-
rem, we always assume that a free Araki -- Woods factor comes with its free quasi-free state and
that all free products are taken w.r.t. the canonical states that we fixed. We always equip a
free product with the free product state and a tensor product with the tensor product state.
We first prove that for every 0 < a < 1, for all finite symmetric Borel measures µ on R and
all symmetric Borel multiplicity functions m : R → N ∪ {+∞}, there exists a state preserving
isomorphism
(4.1)
Γ(µ, m)′′ ∗ Ba ∼= Γ(µ ∗ δZ log(a), +∞)′′ ⊗ Ba .
To prove (4.1), fix an orthogonal representation (Ut)t∈R of R on a real Hilbert space HR having
(µ, m) as its spectral invariant. Denote by H = HR + iHR the complexification of HR. Define
the positive operator ∆ on H such that Ut = ∆it and denote by J : H → H the anti-unitary
operator given by J(ξ + iη) = ξ − iη for all ξ, η ∈ HR. Define H1 = H ⊗ ℓ2(N2). On H1, we
consider the positive operator ∆1 and anti-unitary operator J1 given by
∆1(ξ ⊗ δij) = aj−i ∆(ξ) ⊗ δij
and J1(ξ ⊗ δij) = J(ξ) ⊗ δji
for all i, j ∈ N and ξ ∈ D(∆). Here, (δij ) denotes the standard orthonormal basis of ℓ2(N2).
Note that J1∆1J1 = ∆−1
1 .
Denote by F(H1) the full Fock space of H1 and by θ1 the vector state on B(F(H1)) implemented
by the vacuum vector. For every ξ ∈ H, define the element L(ξ) ∈ B(F(H1)) ⊗ Ba given by
L(ξ) =
∞Xi,j=0
ℓ(ξ ⊗ δij) ⊗ eijq(1 − a)ai .
By [Sh96, Theorem 5.2], we can realize Γ(µ, m)′′ ∗Ba as the von Neumann algebra M generated
by
{L(ξ) + L(J∆1/2ξ)∗ : ξ ∈ D(∆1/2)}
and 1 ⊗ Ba .
Moreover, the free product state on Γ(µ, m)′′ ∗ Ba is given by the restriction of θ1 ⊗ θa to M.
To conclude the proof of (4.1), it thus suffices to show that there is a state preserving isomor-
phism
(4.2)
The left hand side of (4.2) is generated by the operators
(cid:0)(1 ⊗ e00)M(1 ⊗ e00), θ1) ∼=(cid:0)Γ(µ ∗ δZ log(a), +∞)′′, ϕµ∗δZ log(a),+∞(cid:1) .
(1 ⊗ e0i) (L(ξ) + L(J∆1/2ξ)∗) (1 ⊗ ej0) =q(1 − a)ai (cid:16)ℓ(cid:0)ξ ⊗ δij(cid:1) + ℓ(cid:0)J1∆1/2
(ξ ⊗ δij)(cid:1)∗(cid:17) .
1
So, the left hand side of (4.2) equals (Γ(µ1, m1)′′, ϕµ1,m1) where µ1 and m1 are chosen so that
the measure class and multiplicity function of log(∆1) equal C(µ1) and m1. One checks that
C(µ1) = C(µ ∗ δZ log(a)) and m1 = +∞ a.e. So we have proved the existence of the state
preserving isomorphism (4.1).
We next prove that for every t > 0, there is a state preserving isomorphism
(4.3)
Γ(µ, m)′′ ∗ Γ(δt + δ−t, 1)′′ ∼= Γ(µ ∗ δZt ∨ δZt, +∞)′′ .
By [Sh96, Theorem 4.8], there is a state preserving isomorphism
Γ(δt + δ−t, 1)′′ ∼= L∞([0, 1]) ∗ Bexp(−t) .
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
11
By [Sh96, Theorem 2.11], the free Araki -- Woods functor Γ turns direct sums into free products.
Writing µ1 = µ + δ0, m1 = m + δ0 and using (4.1), we obtain the state preserving isomorphisms
Γ(µ, m)′′ ∗ Γ(δt + δ−t, 1)′′ ∼= Γ(µ1, m1)′′ ∗ Bexp(−t)
∼= Γ(µ1 ∗ δZt, +∞)′′ ⊗ Bexp(−t) .
Applying this to (µ, m) = (µ1 ∗ δZt, +∞), we also have the state preserving isomorphisms
Γ(µ1 ∗ δZt, +∞)′′ ∼= Γ(µ1 ∗ δZt, +∞)′′ ∗ Γ(δt + δ−t, 1)′′ ∼= Γ(µ1 ∗ δZt, +∞)′′ ⊗ Bexp(−t) .
Combining both, it follows that (4.3) holds.
We are now ready to prove the theorem. Fix an atom t > 0 of µ. Writing (µ, m) = (µ0, m0) +
(δt + δ−t, 1), we get from (4.3) the state preserving isomorphism
(4.4) Γ(µ, m)′′ ∼= Γ(µ0, m0)′′ ∗ Γ(δt + δ−t, 1)′′ ∼= Γ(µ0 ∗ δZt ∨ δZt, +∞)′′ ∼= Γ(µ ∗ δZt, +∞)′′ .
Let {tn : n ≥ 0} be the positive atoms of µ, with repetitions if there are only finitely many of
them. For every n ≥ 0, define
µn = µ ∗ δZt0 ∗ · · · ∗ δZtn .
For every n, tn+1 is an atom of µn. Repeatedly applying (4.4), we find the state preserving
isomorphisms
Γ(µ, m)′′ ∼= Γ(µ0, +∞)′′ ∼= Γ(µn, +∞)′′ .
So, we also get state preserving isomorphisms
Γ(µ, m)′′ ∼= Γ(µ0, +∞)′′ ∼= ∗
Γ(µ0, +∞)′′ ∼= ∗
n∈N
n∈N
Γ(µn, +∞)′′ ∼= Γ(∨n∈Nµn, +∞) .
Since ∨n∈Nµn is equivalent with µ ∗ δΛ, the theorem follows.
(cid:3)
We deduce the isomorphism part of Theorem F from the following result, generalizing [Sh97a,
Theorem 5.1] and proved using the same methods. For every faithful normal state ψ on a
von Neumann algebra A and for every nonzero projection p ∈ A, we denote by ψp the faithful
normal state on pAp given by ψp(a) = ψ(p)−1ψ(a) for all a ∈ pAp.
Proposition 4.2. Let µ be a finite symmetric Borel measure on R and fix the free Araki --
Woods factor (M, ϕ) = (Γ(µ, +∞)′′, ϕµ,+∞). Let A be a von Neumann algebra with a faithful
normal state ψ having a factorial centralizer Aψ. For every nonzero projection p ∈ Aψ, there
is a state preserving isomorphism
(cid:16)p(cid:0)(M, ϕ) ∗ (A, ψ)(cid:1)p, (ϕ ∗ ψ)p(cid:17) ∼=(cid:16)(M, ϕ) ∗ (pAp, ψp), ϕ ∗ ψp(cid:17) .
To prove Proposition 4.2, we need the following lemma. It is a direct consequence of [Sh97a,
Corollary 2.5]. To formulate the lemma, we use yet another convention for the construction of
free Araki -- Woods factors. We call involution on a Hilbert space H any closed densely defined
antilinear operator S satisfying S(ξ) ∈ D(S) and S(S(ξ)) = ξ for all ξ ∈ D(S). Taking the
polar decomposition S = J∆1/2 of such an involution, we obtain an anti-unitary operator J
and a nonsingular positive selfadjoint operator ∆ satisfying J∆J = ∆−1. Denoting by Ut
the restriction of ∆it to the real Hilbert space HR = {ξ ∈ H : J(ξ) = ξ}, we obtain an
orthogonal representation (Ut)t∈R. Every orthogonal representation of R arises in this way.
The associated free Araki -- Woods factor can be realized on the full Fock space F(H) as the
von Neumann algebra generated by the operators ℓ(ξ) + ℓ(S(ξ))∗, ξ ∈ D(S). We denote this
realization of the free Araki -- Woods factor as Γ(H, S)′′.
Lemma 4.3. Let K be a Hilbert space and Ω ∈ K a unit vector. Let H be a Hilbert space and
H0 ⊂ H a total subset. Assume that
• A ⊂ B(K) is a von Neumann subalgebra and h · Ω, Ωi defines a faithful state ψ on A,
• for every ξ ∈ H0, we are given an operator L(ξ) ∈ B(K),
such that the following conditions hold:
12
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
(i) L(ξ1)∗aL(ξ2) = ψ(a) hξ2, ξ1i 1 for all ξ1, ξ2 ∈ H0 and a ∈ A,
(ii) L(ξ)∗aΩ = 0 for all ξ ∈ H0 and a ∈ A,
(iii) denoting by A the ∗-algebra generated by A and {L(ξ) : ξ ∈ H0}, we have that AΩ is
dense in K.
Then, L can be uniquely extended to a linear map L : H → B(K) such that the above properties
remain valid. For every involution S on H with associated free Araki -- Woods factor Γ(H, S)′′,
there is a unique normal homomorphism
π :(cid:0)Γ(H, S)′′, ϕ(H,S)(cid:1) ∗ (A, ψ) → B(K)
satisfying π(cid:0)ℓ(ξ) + ℓ(S(ξ))∗(cid:1) = L(ξ) + L(S(ξ))∗ for all ξ ∈ D(S) and π(a) = a for all a ∈ A.
Also, hπ( · )Ω, Ωi equals the free product state ϕ(H,S) ∗ ψ.
Using Lemma 4.3, we can prove Proposition 4.2.
Proof of Proposition 4.2. Since Aψ is a factor, we can choose partial isometries vi ∈ Aψ, i ≥ 1,
such that v∗
i vi) = ψ(p)/ni for some integers ni ≥ 1. We can
then also choose partial isometries wis ∈ pAψp, s = 1, . . . , ni, such that wisw∗
i vi for all s
i vi ≤ p, P∞
i = 1 and ψ(v∗
i=1 viv∗
is = v∗
and Pni
s=1 w∗
iswis = p.
Since (M, ϕ) is a free Araki -- Woods factor with infinite multiplicity, we can choose an involution
S0 on a Hilbert space H0 and realize (M, ϕ) as Γ(H, S)′′, where H = H0 ⊗ ℓ2(N2) and S is
given by S(ξ ⊗ δkl) = S0(ξ) ⊗ δlk for all ξ ∈ D(S0) and all k, l ≥ 1. We then consider the
standard free product representation for Γ(H, S)′′ ∗ A on the Hilbert space K with vacuum
vector Ω. Note that p(cid:0)Γ(H, S)′′ ∗ A(cid:1)p is generated by
(4.5)
pAp ∪nv∗
i(cid:0)ℓ(ξ ⊗ δkl) + ℓ(S0(ξ) ⊗ δlk)∗(cid:1)vj (cid:12)(cid:12)(cid:12) i, j, k, l ≥ 1, ξ ∈ D(S0)o .
For all k, l ≥ 0, i, j ≥ 1 and ξ ∈ H, define
Lijkl(ξ) = ψ(p)−1/2
niXs=1
njXt=1
A direct computation shows that
w∗
isv∗
i ℓ(ξ ⊗ δnik+s,njl+t) vj wjt .
Li′j′k′l′(ξ′)∗ a Lijkl(ξ) = δijkl,i′j′k′l′ hξ, ξ′i ψp(a) p
for all i, j, i′, j′ ≥ 1, k, l, k′, l′ ≥ 0, ξ, ξ′ ∈ H and a ∈ pAp.
Applying Lemma 4.3 to the Hilbert space H1 = H ⊗ ℓ2(N2 × N2
0) with involution S1(ξ ⊗ δijkl) =
S0(ξ) ⊗ δjilk, it follows that Γ(H1, S1)′′ ∗ pAp can be realized as the von Neumann algebra N
generated by
pAp ∪nLijkl(ξ) + Ljilk(S0(ξ))∗ (cid:12)(cid:12)(cid:12) i, j ≥ 1, k, l ≥ 0, ξ ∈ D(S0)o ,
with the free product state being implemented by ψ(p)−1/2pΩ.
Note that
jt = ψ(p)−1/2 v∗
wis (cid:0)Lijkl(ξ)+Ljilk(S0(ξ))∗(cid:1) w∗
i (cid:0)ℓ(ξ⊗δnik+s,njl+t)+ℓ(S0(ξ)⊗δnj l+t,nik+s)∗(cid:1) vj .
that p(cid:0)Γ(H, S)′′ ∗A(cid:1)p equals Γ(H1, S1)′′ ∗pAp in a state preserving way. Since also Γ(H1, S1)′′ ∼=
For fixed i, j ≥ 1, the parameters nik + s and njl + t with k, l ≥ 0, s = 1, . . . , ni and t =
1, . . . , nj exactly run through N2. So, we find back the generating set of (4.5) and conclude
(M, ϕ) in a state preserving way, this concludes the proof of the proposition.
(cid:3)
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
13
5. Proofs of Theorem A and Corollaries B, C, D
Combining Corollary 3.2 with the deformation/rigidity theorems for free Araki -- Woods factors
and for free product factors obtained in [HR10, HU15], we get the following theorem.
Theorem 5.1. Let (M, ϕ) be either a free Araki -- Woods factor with its free quasi-free state or
a free product ∗n(Mn, ϕn) of amenable von Neumann algebras equipped with the free product
state. Let ψ ∈ M∗ be any faithful normal state on M and denote by [Dψ : Dϕ]t Connes'
Radon -- Nikodym 1-cocycle between ψ and ϕ. Let z ∈ Z(M ψ) be the central projection such that
M ψ(1 − z) is amenable and M ψz has no amenable direct summand.
There exists a sequence of partial isometries vn ∈ M such that the projection qn = vnv∗
to M ψ, the projection pn = v∗
n belongs
nvn belongs to M ϕ, Pn qn = z,
n vn σϕ
t (v∗
qn [Dψ : Dϕ]t = λit
n) and ψqn = λn vnϕv∗
n ,
with λn = ψ(qn)/ϕ(pn).
Proof. Let q ∈ M ψ be a nonzero projection such that qM ψq has no amenable direct summand.
Let r0 ∈ Lψ(R) be a nonzero finite trace projection. Put r = Πϕ,ψ(qr0). Then r is a nonzero
finite trace projection in the core cϕ(M ) and Πϕ,ψ(Lψ(R))r commutes with qM ψqr. Since
qM ψqr has no amenable direct summand, it follows from [HR10, Theorem 5.2] (in the case
where M is a free Araki -- Woods factor) and [HU15, Theorem 4.3] (in the case where M is a
free product of amenable von Neumann algebras) that Πϕ,ψ(Lψ(R))r ≺cϕ(M ) Lϕ(R).
By Theorem 3.1, we find a nonzero partial isometry v ∈ qM such that the projection q0 = vv∗
belongs to M ψ, the projection p = v∗v belongs to M ϕ and ψq = λ vϕv∗.
In particular,
λ = ψ(q0)/ϕ(p) and q0 [Dψ : Dϕ]t = λit vσϕ
Since q ∈ M ψ was an arbitrary nonzero projection such that qM ψq has no amenable direct
summand, the theorem follows by a maximality argument.
(cid:3)
t (v∗).
In order to apply Theorem 5.1 to the classification of free Araki -- Woods factors, we need the
following description of the centralizer of the free quasi-free state.
Remark 5.2. When M = Γ(µ, m)′′ is an arbitrary free Araki -- Woods factor with free quasi-free
state ϕ = ϕµ,m, the centralizer M ϕ can be described as follows. Denote by Ma = Γ(µa, m)′′
the almost periodic part of M . First note that M ϕ = M ϕ
a . So if µa = 0, we have M ϕ =
If µa is concentrated on {0}, we conclude that M ϕ = Ma = L(Fm(0)), where the last
C1.
isomorphism follows because the free Araki -- Woods factor associated with the m-dimensional
trivial representation, i.e. Γ(δ0, m)′′, is isomorphic with L(Fm). When µa(log λ) > 0 for some
0 < λ < 1, there is a state preserving inclusion Tλ ⊂ Ma, where Tλ is the unique free Araki --
Woods factor of type IIIλ (see [Sh96, Section 4]). It then follows from [Sh96, Corollary 6.8] that
M ϕ is a factor that contains a copy of the free group factor L(F∞), so that M ϕ is nonamenable.
Actually, using [Dy96], we get that M ϕ ∼= L(F∞) in this case.
Theorem A is a particular case of the following more general result.
Theorem 5.3. Let µ, ν be finite symmetric Borel measures on R and m, n : R → N ∪ {+∞}
symmetric Borel multiplicity functions. Assume that νa 6= 0 and either supp(νa) 6= {0} or
supp(νa) = {0} with n(0) ≥ 2.
If the free Araki -- Woods factors Γ(µ, m)′′ and Γ(ν, n)′′ are isomorphic then there exist nonzero
projections p ∈ (Γ(µ, m)′′)ϕµ,m and q ∈ (Γ(ν, n)′′)ϕν,n and a state preserving isomorphism
(cid:0)p Γ(µ, m)′′ p, (ϕµ,m)p(cid:1) ∼=(cid:0)q Γ(ν, n)′′ q, (ϕν,n)q(cid:1)
where (ϕµ,m)p = ϕµ,m(p · p)
ϕµ,m(p) and (ϕν,n)q = ϕν,n(q · q)
ϕν,n(q) .
14
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
In particular, the joint measure classes C(Wk≥1 µ∗k) and C(Wk≥1 ν∗k) are equal.
Moreover, in case supp(νa) 6= {0} or supp(νa) = {0} with n(0) = +∞, there exists a state
preserving isomorphism (Γ(µ, m)′′, ϕµ,m) ∼= (Γ(ν, n)′′, ϕν,n).
Proof. Put (M, ϕ) := (Γ(µ, m)′′, ϕµ,m) and (N, θ) := (Γ(ν, n)′′, ϕν,n). Let π : M → N be any
isomorphism between M and N . Put ψ := θ ◦ π. By our assumptions on ν and Remark
5.2, the centralizer M ψ is nonamenable. By Theorem 5.1, we find a nonzero partial isometry
v ∈ M such that p = v∗v ∈ M ϕ, q = vv∗ ∈ M ψ and Ad(v) : (pM p, ϕp) → (qM q, ψq) is
It follows in particular that pM ϕp = (pM p)ϕp ∼= (qM q)ψq = qM ψq is a
state preserving.
nonamenable II1 factor. So M ϕ cannot be abelian and Remark 5.2 implies that M ϕ is a II1
factor. Applying Lemma 2.1 twice, we have
C(_k∈N
µ∗k) = C(∆ϕ) = C(∆ϕp) = C(∆ψq ) = C(∆ψ) = C(_k∈N
ν∗k).
This implies that C(Wk≥1 µ∗k) = C(Wk≥1 ν∗k).
Assume now that either supp(νa) 6= {0} or supp(νa) = {0} with n(0) = +∞. In the latter
case where ν({0}) > 0 and n(0) = +∞, we use that the free Araki -- Woods functor Γ turns
direct sums into free products (see [Sh96, Theorem 2.11]) and conclude that there exists a state
preserving isomorphism
(5.1)
(N, θ) ∼= (N, θ) ∗ (L(F∞), τ ) .
In the case where ν has at least one atom different from 0, it follows similarly from the classi-
fication of almost periodic free Araki -- Woods factors (see [Sh96]) that (5.1) holds.
Put q0 = π(q). Above, we have proved that there exists a state preserving isomorphism
(pM p, ϕp) ∼= (q0N q0, θq0). Taking a smaller p if needed, we may assume that ϕ(p) = 1/k
for some integer k ≥ 1. Combining (5.1) with Proposition 4.2 and the fact that the funda-
mental group of L(F∞) equals R∗
+ (see [Ra91]), it follows that there exists a state preserving
isomorphism (q0N q0, θq0) ∼= (q1N q1, θq1) whenever q1 ∈ N θ is a nonzero projection.
Choose a projection q1 ∈ N θ with θ(q1) = 1/k. So, there exists a state preserving isomorphism
(pM p, ϕp) ∼= (q1N q1, θq1). Since ϕ(p) = 1/k = θ(q1) and since both M ϕ and N θ are factors,
taking k × k matrices, we find a state preserving isomorphism (M, ϕ) ∼= (N, θ).
(cid:3)
Proof of Corollary B. Let µ, ν ∈ S(R) such that Λ(µa) = Λ(νa) =: Λ and C(µc∗δΛ) = C(νc∗δΛ).
Let m, n : R → N ∪ {+∞} be any symmetric Borel multiplicity functions. Then we have
C(µ ∗ δΛ) = C(ν ∗ δΛ). By Theorem 4.1, there is a state preserving isomorphism Γ(µ, m)′′ ∼=
Γ(ν, n)′′.
Conversely, let µ, ν ∈ S(R) and m, n : R → N ∪ {+∞} be any symmetric Borel multiplic-
ity functions such that Γ(µ, m)′′ ∼= Γ(ν, n)′′. By Theorem A, we have that C(Wk≥1 µ∗k) =
C(Wk≥1 ν∗k). Since for every k ≥ 1, we have µ∗k
c ≺ µc and ν∗k
c ≺ νc, it follows that
C(µc ∗ δΛ(µa) ∨ δΛ(µa)) = C(_k≥1
= C(_k≥1
µ∗k)
ν∗k)
= C(νc ∗ δΛ(νa) ∨ δΛ(νa)).
This implies that Λ(µa) = Λ(νa) and C(µc ∗ δΛ(µa)) = C(νc ∗ δΛ(νa)).
(cid:3)
Proof of Corollary C. Put (M, ϕ) := (Γ(µ, m)′′, ϕµ,m). Let ψ ∈ M∗ be a faithful normal state
such that M ψ is nonamenable. By Theorem 5.1, we find a nonzero partial isometry v ∈ M such
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
15
that q = vv∗ ∈ M ψ, p = v∗v ∈ M ϕ, qM ψq has no amenable direct summand and ψq = λ vϕv∗
with λ = ψ(q)/ϕ(p). It follows that pM ϕp ∼= qM ψq has no amenable direct summand. By
Remark 5.2, this means that either µa has an atom different from 0 or µa is concentrated on
{0} with m(0) ≥ 2. Conversely, if µa satisfies these properties, it follows from Remark 5.2 that
the centralizer of the free quasi-free state is nonamenable.
(cid:3)
Proof of Corollary D. By Corollary C, the von Neumann algebra Γ(λ + δ0, 1)′′ has amenable
centralizers while Γ(λ + δ0, 2)′′ does not.
(cid:3)
Example 5.4. Many different measures in the family S(R) of Corollary B can be constructed
as follows. Let K ⊂ R be an independent Borel set, meaning that every n-tuple of distinct
elements in K generates a free abelian group of rank n. By [Ru62, Theorems 5.1.4 and 5.2.2],
there exist compact independent K ⊂ R such that K is homeomorphic to a Cantor set. Fix
such a K ⊂ R and put L = K ∪ (−K). Also fix a countable subgroup Λ < R.
For every continuous symmetric probability measure µ on R that is concentrated on L, define
the measure class eµ on R given by
(x + µ∗n) .
eµ = _x∈Λ,n≥1
By construction, each eµ is a continuous symmetric measure class on R that is invariant under
translation by Λ and that satisfies eµ ∗eµ ≺ eµ.
claim that C(fµ1) = C(fµ2) if and only if C(µ1) = C(µ2). One implication is obvious. The other
implication is a consequence of the following result contained in [LP97, Corollary 1] : if η1 and
η2 are concentrated on L and η1 ⊥ η2, then also η1 ⊥ (x + η∗k
Given continuous symmetric probability measures µ1 and µ2 that are concentrated on L, we
2 ) for all x ∈ R and all k ≥ 1.
Choosing Λ to be a nontrivial subgroup of R and applying Corollary B, for all continuous
symmetric probability measures µ1 and µ2 concentrated on the Cantor set L, we find that
Γ(fµ1 ∨ δΛ, m1)′′ ∼= Γ(fµ2 ∨ δΛ, m2)′′
iff C(µ1) = C(µ2) .
iff C(µ1) = C(µ2) .
Adding the Lebesgue measure to eµ, we claim that we also have
Γ(λ ∨fµ1 ∨ δΛ, m1)′′ ∼= Γ(λ ∨fµ2 ∨ δΛ, m2)′′
By [Sh97b, Corollary 8.6], for all these free Araki -- Woods factors, the τ -invariant equals the
usual topology on R, so that they cannot be distinguished by Connes' invariants.
To prove the claim, define Ln as the n-fold sum Ln = L + · · · + L and put S =Sn≥1 Ln. Below
we prove that λ(S) = 0. The claim then follows from Corollary B : if C(λ ∨fµ1) = C(λ ∨fµ2),
restricting to S, we get that C(fµ1) = C(fµ2). As proven above, this implies that C(µ1) = C(µ2).
It remains to prove that λ(Ln) = 0 for all n. If for some n ≥ 1, we have λ(Ln) > 0, then
L2n = Ln − Ln contains a neighborhood of 0. Every nonzero x ∈ L2n can be uniquely written
as x = α1y1 + · · · + αkyk with k ≥ 1, y1, . . . , yk distinct elements in K and αi ∈ Z \ {0} with
αi ≤ 2n for all i. So if x ∈ L2n is nonzero, we have that (2n + 1)x 6∈ L2n. Therefore, L2n does
not contain a neighborhood of 0 and it follows that λ(Ln) = 0 for all n ≥ 1.
6. Proof of Theorem E
To prove Theorem E, we combine [HU15, Theorem 4.3] and Theorem 3.1 with the following
lemma. Whenever θ is a faithful normal state on a von Neumann algebra M , we denote by
Map,θ the von Neumann subalgebra of M generated by the almost periodic part of (σθ
t ).
16
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
Lemma 6.1. For i = 1, 2, let (Mi, ϕi) be von Neumann algebras with a faithful normal state.
Denote by (M, ϕ) = (M1, ϕ1) ∗ (M2, ϕ2) their free product. Denote by EM1 : M → M1 the
unique ϕ-preserving conditional expectation. Let θ1 be a faithful normal state on M1 and define
θ = θ1 ◦ EM1. Let q ∈ M θ be a projection.
There exist projections q0, q1, . . . with q0 ∈ M θ1
1 and qi ∈ M θ for all i ≥ 1 such that
(i) P∞
i=0 qi = q,
(ii) q0Map,θq0 = q0M1,ap,θ1q0 and q0M θq0 = q0M θ1
(iii) for every i ≥ 1, there exists a partial isometry vi ∈ M with viv∗
1 q0,
i = qi, v∗
i vi ∈ M ϕ and
1
θ(qi)
θqi =
1
ϕ(v∗
i vi)
viϕv∗
i .
Proof. Fix standard representations Mi ⊂ B(Hi). For every faithful normal state µ on Mi,
denote by ξµ ∈ Hi the canonical unit vector that implements µ.
Define ut = [Dθ1 : Dϕ1]t ∈ U (M1). Note that also [Dθ : Dϕ]t = [Dθ1 ◦ E : Dϕ1 ◦ E]t = ut for all
t ∈ R. Let e1, e2, . . . be a maximal sequence of nonzero projections in M θ1
such that eiej = 0
1
whenever i 6= j and such that for every i ≥ 1, there exists a partial isometry wi ∈ M1 and a
λi > 0 with wiw∗
i=1 ei. Then
e0 ∈ M θ1
1 . By construction, the unitary representation (Ut)t∈R on H1 given by Ut(xξϕ1) =
utσϕ1
i wi for all t ∈ R. Define e0 = 1 −P∞
t (x)ξϕ1 for all x ∈ M1 is weakly mixing on e0H1.
i = ei and utσϕ1
t (wi) = λit
For i = 1, 2, define
◦
H i = Hi ⊖ Cξϕi. For every k ≥ 1, define the Hilbert space
Kk = H1 ⊗
◦
H 2 ⊗
◦
H 1 ⊗
◦
H 1 ⊗
⊗
◦
H2
⊗H1 .
k times
◦
H 2 and k − 1 times
◦
H 1, alternatingly
· · ·
{z
}
We can then identify the standard Hilbert space H for M with
H = H1 ⊕
∞Mk=1
Kk .
Under this identification, ξϕ = ξϕ1 ∈ H1 and ξθ = ξθ1 ∈ H1. Denote by (Vt)t∈R the unitary
representation on H1 given by Vt(xξθ1) = σϕ1
t ξθ1 for all x ∈ M1. Under the above
identification of H, we get that
t (x)u∗
∆it
θ = ∆it
θ1 ⊕
ϕ2 ⊗ ∆it
ϕ1 ⊗ · · · ⊗ ∆it
ϕ1 ⊗ ∆it
∞Mk=1(cid:16)Ut ⊗ ∆it
ϕ2 ⊗ Vt(cid:17) .
Since (Ut)t∈R is weakly mixing on e0H1, we conclude that (∆it
e0H1. It follows that
θ )t∈R is weakly mixing on e0H ⊖
(6.1)
e0Map,θe0 = e0M1,ap,θ1e0
and e0M θe0 = e0M θ1
1 e0 .
and such that statement (iii) in the lemma holds for every i ≥ 1. Define q0 = q −P∞
Let q1, q2, . . . be a maximal sequence of nonzero projections in qM θq such that qiqj = 0 if i 6= j
i=1 qi.
Then q0 ∈ M θ. We prove that q0 ≤ e0. Once this is proven, it follows from (6.1) that q0 ∈ M θ1
1
and that q0Map,θq0 = q0M1,ap,θ1q0, so that the lemma follows.
If q0 6≤ e0, we find j ≥ 1 such that q0ej 6= 0. Then the polar part v of q0wj is a nonzero partial
isometry in M satisfying vv∗ ≤ q0 and utσϕ1
j v for all t ∈ R. So, the projection vv∗
could be added to the sequence q1, q2, . . ., contradicting its maximality. Therefore, q0 ≤ e0 and
the lemma is proved.
(cid:3)
t (v) = λit
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
17
Theorem E will be an immediate consequence of the following more technical proposition that
will also be used in Section 7 below.
Proposition 6.2. For i = 1, 2, let (Mi, ϕi) be von Neumann algebras with a faithful normal
state. Denote by (M, ϕ) = (M1, ϕ1) ∗ (M2, ϕ2) their free product and by EMi : M → Mi the
unique ϕ-preserving conditional expectation. Let ψ be a faithful normal state on M . Define the
set of projections P ⊂ M ψ given by P = P1 ∪ P2 ∪ P3 where
• for i = 1, 2, Pi consists of the projections q ∈ M ψ for which there exists a partial
isometry v ∈ M and a faithful normal state θi on Mi with v∗v = q, e = vv∗ ∈ M θi
i ,
1
vψv∗ =
ψ(q)
vMap,ψv∗ = eMi,ap,θie
θi(e)
1
(θi ◦ EMi)e
,
and
vM ψv∗ = eM θi
i e ,
• P3 consists of the projections q ∈ M ψ for which there exists a partial isometry v ∈ M
with v∗v = q, e = vv∗ ∈ M ϕ,
1
ψ(q)
vψv∗ =
1
ϕ(e)
ϕe
and
vM ψv = eM ϕe .
If q ∈ M ψ is a projection such that qM ψq has no amenable direct summand, then q can be
written as a sum of projections in P.
Proof. Let ψ be a faithful normal state on M and q ∈ M ψ a projection such that qM ψq has
no amenable direct summand. It suffices to prove that q dominates a nonzero projection in P,
since then a maximality argument can be applied.
Fix any nonzero finite trace projection r0 ∈ Lψ(R) and put r = Πϕ,ψ(qr0). Define the von
Neumann subalgebra Q ⊂ rcϕ(M )r given by
Note that Q has no amenable direct summand. By [HU15, Theorem 4.3],
Q = Πϕ,ψ(cid:0)Lψ(R)qr0 ∨ qM ψqr0(cid:1) .
either Q′ ∩ rcϕ(M )r ≺cϕ(M ) Lϕ(R)
or Q ≺cϕ(M ) cϕ(Mi) for i = 1 or i = 2.
Since Πϕ,ψ(Lψ(R)qr0) belongs to both Q and Q′ ∩ rcϕ(M )r, it follows that
for i = 1 or i = 2.
Πϕ,ψ(Lψ(R)qr0) ≺cϕ(M ) cϕ(Mi)
By Theorem 3.1, we find a faithful normal state θi on Mi and a partial isometry v ∈ M such
that q0 = v∗v is a nonzero projection in qM ψq, p = vv∗ belongs to M θi◦EMi and
1
ψ(q0)
vψv∗ =
1
θi(EMi(p))
(θi ◦ EMi)p .
Write θ = θi ◦ EMi. By Lemma 6.1, we either find a nonzero projection e ≤ p such that e ∈ M θi
i
and eMap,θe = eMi,ap,θie and eM θe = eM θi
i e, or we find a nonzero projection p0 ∈ pM θp and
a partial isometry w ∈ M such that ww∗ = p0, e = w∗w belongs to M ϕ and
1
θ(p0)
w∗θw =
1
ϕ(e)
ϕe .
In the first case, we get that the projection v∗ev ≤ q belongs to Pi, while in the second case,
the projection v∗p0v ≤ q belongs to P3.
(cid:3)
18
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
Proof of Theorem E. Denote by EMi : M → Mi the unique ϕ-preserving conditional expecta-
tion. If M ϕ is nonamenable, then obviously, M does not have all its centralizers amenable. If
Mi admits a faithful normal state θi such that M θi
is nonamenable, then θi ◦ EMi is a faithful
i
normal state on M with M θi
i ⊂ M θi◦EMi , so that again, M does not have all its centralizers
amenable.
Conversely, assume that ψ is a faithful normal state on M such that M ψ is nonamenable. Take
a nonzero projection q ∈ M ψ such that qM ψq has no amenable direct summand. By 6.2, we
either find i ∈ {1, 2} and a faithful normal state θi on Mi such that M θi
is nonamenable, or
i
we find that M ϕ is nonamenable.
(cid:3)
7. Further structural results and proof of Theorem F
We start by showing that the invariant of Theorem A is not a complete invariant for the family
of free Araki -- Woods factors Γ(µ, m)′′ arising from finite symmetric Borel measures µ on R
whose atomic part µa is nonzero and not supported on {0}.
Theorem 7.1. Let Λ < R be any countable subgroup such that Λ 6= {0} and denote by δΛ
a finite atomic measure on R whose set of atoms equals Λ. Let η be any continuous finite
symmetric Borel measure on R such that C(η) = C(η ∗ δΛ) and such that the measures (η∗k)k≥1
are pairwise singular.
Put µ = δΛ + η and ν = δΛ + η + η ∗ η. Then,
Γ(µ, 1)′′ 6∼= Γ(ν, 1)′′
and C(_k≥1
µ∗k) = C(_k≥1
ν∗k).
Proof. By construction, we have C(Wk≥1 µ∗k) = C(Wk≥1 ν∗k). We denote M := Γ(µ, 1)′′ and
N := Γ(ν, 1)′′. We denote by Q ⊂ N the canonical von Neumann subalgebra given by Q :=
Γ(δΛ + η, 1)′′. Put ϕ := ϕµ,1 and ψ := ϕν,1. Observe that the inclusion Q ⊂ N is globally
invariant under the modular automorphism group σψ.
Assume by contradiction that M ∼= N . By Theorem A, there exists a state preserving iso-
morphism π : (M, ϕ) → (N, ψ) of M onto N . Then, π extends to a unitary operator
U : L2(M, ϕ) → L2(N, ψ) satisfying U ∆ϕU ∗ = ∆ψ. Define the real Hilbert space
H µ
R :=nf ∈ L2
C(R, µ) : f (−s) = f (s) for µ-almost every s ∈ Ro
and the orthogonal representation
U µ : R y H µ
Denote by s(ξ) := ℓ(ξ) + ℓ(ξ)∗, ξ ∈ H µ
M and satisfy σϕ
µ is singular w.r.t. η∗k for all k ≥ 2, it follows that π(s(ξ)) ∈ Q for all ξ ∈ H µ
π(M ) ⊂ Q, which is impossible because π is surjective.
t ξ). By construction, C(∆ψL2(N )⊖L2(Q)) = Tk≥2 C(η∗k). Since
R : U µ
R, the canonical semicircular elements that generate
s (f )(t) = exp(ist)f (t) .
R. But then,
(cid:3)
t (s(ξ)) = s(U µ
Note that Example 5.4 provides many measures η satisfying the assumptions of Theorem 7.1.
We could not prove or disprove that the measure class C(µc ∗ δΛ(µa)) is an invariant for the
family of free Araki -- Woods factors Γ(µ, m)′′ arising from finite symmetric Borel measures µ
on R whose atomic part µa is nonzero and not supported on {0}.
Using Theorem 5.1 in combination with the results of [BH16], we can also clarify the rela-
tion between free Araki -- Woods factors and free products of amenable von Neumann algebras.
Combining [Sh96, Theorems 2.11 and 4.8], it follows that every almost periodic free Araki --
Woods factor is isomorphic with a free product of von Neumann algebras of type I. Conversely,
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
19
by [Ho06], many free products of type I von Neumann algebras are isomorphic with free Araki --
Woods factors. In [Dy92, Theorem 4.6], it is proved that a free product of two amenable von
Neumann algebras w.r.t. faithful normal traces is always isomorphic to the direct sum of an in-
terpolated free group factor and a finite dimensional algebra. It is therefore tempting to believe
that every type III factor arising as a free product of amenable von Neumann algebras w.r.t.
faithful normal states is a free Araki -- Woods factor. The following example shows however that
this is almost never the case if one of the states fails to be almost periodic.
Theorem 7.2. Let (P, θ) = ∗n(Pn, θn) be a free product of amenable von Neumann algebras.
Assume that the centralizer P θ has no amenable direct summand and that at least one of the
θn is not almost periodic.
Then P is not isomorphic to a free Araki -- Woods factor. Even more:
there is no faithful
normal homomorphism π of P into a free Araki -- Woods factor M such that π(P ) ⊂ M is with
expectation.
The same conclusions hold if (P, θ) is any von Neumann algebra with a faithful normal state
satisfying the following three properties: the centralizer P θ has no amenable direct summand,
θ is not almost periodic and P is generated by a family of amenable von Neumann subalgebras
Pn ⊂ P that are globally invariant under the modular automorphism group (σθ
t ).
Proof. Let (M, ϕ) be a free Araki -- Woods factor with its free quasi-free state. Let (P, θ) be
a von Neumann algebra with a faithful normal state θ such that the centralizer P θ has no
amenable direct summand and such that P is generated by a family of amenable von Neumann
subalgebras Pn ⊂ P that are globally invariant under the modular automorphism group (σθ
t ).
Let π : P → M be a normal homomorphism and E : M → π(P ) a faithful normal conditional
expectation. We prove that θ is almost periodic.
Define the faithful normal state ψ ∈ M∗ given by ψ = θ ◦ π−1 ◦ E. Since π(P θ) ⊂ M ψ, we
get that M ψ has no amenable direct summand. By Theorem 5.1, we find partial isometries
vn ∈ M such that qn = vnv∗
n ∈ M ψ, pn = v∗
nvn ∈ M ϕ, Pn qn = 1 and
qn [Dψ : Dϕ]t = λit
n vn σϕ
t (v∗
n)
where λn = ψ(qn)/ϕ(pn).
Replacing (M, ϕ) by the free product of (M, ϕ) and the appropriate almost periodic free Araki --
Woods factor, we still get a free Araki -- Woods factor and we may assume that M ϕ is a factor
and that each λn is an eigenvalue for the free quasi-free state. We can then choose partial
isometries wn ∈ M such that σϕ
nwn belongs
n = pn and such that en = w∗
n wn, wnw∗
t (wn) = λ−it
to M ϕ with Pn en = 1. So, we find that
for all n and all t ∈ R. We conclude that v = Pn vnwn is a unitary in M satisfying [Dψ :
t (v∗). This means that ϕ = ψ ◦ Ad v and that the homomorphism η = Ad v∗ ◦ π
t ◦ η = η ◦ σθ
t .
qn [Dψ : Dϕ]t = vnwn σϕ
Dϕ]t = v σϕ
satisfies σϕ
t (w∗
nv∗
n)
So for every n, the subalgebra η(Pn) ⊂ M is amenable and globally invariant under the modular
automorphism group (σϕ
t ). By [BH16, Theorem 4.1], it follows that η(P ) lies in the almost
periodic part of (M, ϕ). This implies that the restriction of (σθ
t ) to Pn is almost periodic. Since
this holds for every n, we conclude that θ is almost periodic.
(cid:3)
From Proposition 6.2, we get the following rigidity results for free product von Neumann al-
gebras. Roughly, the result says that an arbitrary free product of a "very much non almost
periodic" M1 with an almost periodic M2 remembers the almost periodic part M2 up to am-
plification.
20
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
Recall that a faithful normal positive functional ϕ on a von Neumann algebra M is said to
be weakly mixing if the unitary representation σϕ
t ( · ) on L2(M, ϕ) ⊖ C1 is weakly mixing, and
that ϕ is said to be almost periodic if the unitary representation σϕ
t ( · ) on L2(M, ϕ) is almost
periodic.
Proposition 7.3. For i = 1, 2, let (Mi, ϕi) and (Ni, ψi) be von Neumann algebras with faithful
normal states. Denote by (M, ϕ) = (M1, ϕ1) ∗ (M2, ϕ2) and (N, ψ) = (N1, ψ1) ∗ (N2, ψ2) their
free products. Assume that
• M1 and N1 have all their centralizers amenable and ϕ1, ψ1 are weakly mixing states,
• M ϕ2
2 have no amenable direct summand and ϕ2, ψ2 are almost periodic.
2 and N ψ2
If M ∼= N , there exist nonzero projections e ∈ M2 and q ∈ N2 such that eM2e ∼= qN2q.
Proof. Whenever µ is a faithful normal state on M and q ∈ M µ is a projection such that
qM µq has no amenable direct summand, we can apply Proposition 6.2. Since M1 has all its
centralizers amenable, the set P1 in Proposition 6.2 equals {0}. Since ϕ1 is weakly mixing, the
almost periodic part of a state of the form θ2 ◦ EM2 (and, in particular, of ϕ) is contained in
M2. It thus follows from Proposition 6.2 that there exist sequences of projections qi ∈ M µ and
ei ∈ M2, as well as partial isometries vi ∈ M and faithful normal positive functionals θi on
eiM2ei such that v∗
i equals θi ◦ EM2 on eiM ei for all
i ≥ 0.
i = ei, P∞
i=0 qi = q and viµv∗
i vi = qi, viv∗
Let π : M → N be an isomorphism of M onto N . We first apply the result in the first paragraph
to µ = ψ ◦π. Since ψ1 is weakly mixing, M µ = π−1(N ψ2
2 ). We find nonzero projections q ∈ N ψ2
and e ∈ M2, a partial isometry v ∈ M and a faithful normal positive functional θ on eM2e
such that v∗v = π−1(q), vv∗ = e and vµv∗ = θ ◦ EM2 on eM e. Since ϕ1 is weakly mixing, the
almost periodic part of θ ◦ EM2 equals eM2,ap,θe. Since ψ1 is weakly mixing and ψ2 is almost
periodic, the almost periodic part of µ equals π−1(N2). It follows that
2
(7.1)
vπ−1(N2)v∗ = eM2,ap,θe .
2 . So conjugating e and θ, we may assume that e ∈ M ϕ2
By [HU15, Lemma 2.1], every projection in M2 is equivalent, inside M2, with a projection in
M ϕ2
2 . We then apply the result of the first
paragraph of the proof to the free product N = N1 ∗ N2, the faithful normal state µ′ = ϕ ◦ π−1
and the projection f = π(e) in N µ′
2 ).
We thus find projections ei ∈ eM ϕ2
2 e summing up to e, projections pi ∈ N2, partial isometries
wi ∈ N and faithful normal positive functionals Ωi on piN2pi such that w∗
i = pi
and wiµ′w∗
i = Ωi ◦ EN2 on piN pi for all i. In particular, Ω(pi) = ϕ2(ei) and Pi Ω(pi) = ϕ2(e).
Define the projection p ∈ B(ℓ2(N)) ⊗ N2 given by p = Pi eii ⊗ pi. Define the faithful normal
. Since ϕ1 is weakly mixing, we have N µ′
positive functional Ω on p(B(ℓ2(N)) ⊗ N2)p given by
= π(M ϕ2
i wi = π(ei), wiw∗
Ω(T ) =Xi
Ωi(Tii) .
Finally define W ∈ ℓ2(N) ⊗ N given by W = Pi ei ⊗ wi.
W W ∗ = p and
It follows that W ∗W = π(e),
W µ′W ∗ = Ω ◦ EB(ℓ2(N))⊗N2
on p(B(ℓ2(N)) ⊗ N )p .
As above, Nap,µ′ = π(Map,ϕ) = π(M2). Since ψ1 is weakly mixing, the almost periodic part
of the functional Ω ◦ EB(ℓ2(N))⊗N2 on p(B(ℓ2(N)) ⊗ N )p is contained in p(B(ℓ2(N)) ⊗ N2)p. It
follows that
(7.2)
W π(eM2e)W ∗ ⊂ p(B(ℓ2(N)) ⊗ N2)p .
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
21
Write V = W π(v). Then V ∈ ℓ2(N) ⊗ N is a partial isometry with V V ∗ = p ∈ B(ℓ2(N)) ⊗ N2
and V ∗V = q ∈ N2. Using (7.1) and (7.2), we find that
(7.3)
V qN2qV ∗ = W π(eM2,ap,θe)W ∗ ⊂ W π(eM2e)W ∗ ⊂ p(B(ℓ2(N)) ⊗ N2)p .
Since N ψ2
2 has no amenable direct summand, it follows in particular that N2 is diffuse. Then
(7.3) implies that V ∈ ℓ2(N) ⊗ N2 and therefore, all the inclusions in (7.3) are equalities. In
particular, eM2,ap,θe = eM2e so that (7.1) implies that qN2q ∼= eM2e. This concludes the proof
of the proposition.
(cid:3)
Combining Propositions 4.2 and 7.3, we can easily prove Theorem F.
.
∼= At
Proof of Theorem F. Put (M, ϕ) = (Γ(µ, +∞)′′, ϕµ,+∞) as in the formulation of the Proposi-
tion. Then, ϕ is weakly mixing and by Corollary C, the free Araki -- Woods factor M has all its
centralizers amenable. For i = 1, 2, let (Ai, ψi) be von Neumann algebras with almost periodic
faithful normal states having a nonamenable factorial centralizer Aψi
i
If the free products (Mi, ϕi) = (M, ϕ) ∗ (Ai, ψi) satisfy M1 ∼= M2, it follows from Proposition
7.3 that there exist nonzero projections pi ∈ Ai such that p1A1p1 ∼= p2A2p2. In the first case,
2 for some t > 0. In the second case,
where the Ai are II1 factors, this implies that A1
where the Ai are type III factors, this implies that A1 ∼= A2.
For the converse, first assume that the (Ai, ψi) are II1 factors with their tracial states and
2 for some t > 0. Take nonzero projections pi ∈ Ai such that p1A1p1 ∼= p2A2p2. By
A1 ∼= At
the uniqueness of the trace, we have (p1A1p1, (ψ1)p1) ∼= (p2A2p2, (ψ2)p2). It then follows from
Proposition 4.2 that p1M1p1 ∼= p2M2p2. Since the Mi are type III factors, this further implies
that M1 ∼= M2.
Finally assume that the (Ai, ψi) are full type III factors with almost periodic states having a
factorial centralizer Aψi
i and that π : A1 → A2 is an isomorphism of A1 onto A2. Denote by
Γ = Sd(A1) = Sd(A2) the Sd-invariant of A1 ∼= A2. Define (Bi, θi) = (B(ℓ2(N)) ⊗ Ai, Tr ⊗ ψi).
By [Co74, Lemma 4.8], the weight θi on Bi is a Γ-almost periodic weight. By [Co74, Theorem
4.7], there exists a unitary U ∈ B2 and a constant α > 0 such that θ2 ◦ Ad(U ) ◦ (id ⊗ π) = α θ1.
Since Aψ2 is a factor, after a unitary conjugacy of π, we find nonzero projections pi ∈ Aψi
such
i
that π(p1) = p2 and (ψ2)p2 ◦ π = (ψ1)p1 on p1A1p1. As in the previous paragraph, we can use
Proposition 4.2 to conclude that M1 ∼= M2.
(cid:3)
We finally consider two further structural properties of free Araki -- Woods factors: the free
absorption property and the structure of its continuous core. We say that a von Neumann
algebra M with a faithful normal state ϕ has the free absorption property if the free product
(N, ψ) = (M, ϕ) ∗ (L(F∞), τ ) satisfies N ∼= M . One of the key results in [Sh96] is the free
absorption property for the almost periodic free Araki -- Woods factors. In general, we get the
following result.
Proposition 7.4. Let (M, ϕ) = (Γ(µ, +∞)′′, ϕµ,+∞) be a free Araki -- Woods factor with infinite
multiplicity. Then (M, ϕ) has the free absorption property if and only if the atomic part µa is
nonzero.
Proof. If µ({0}) > 0, then (M, ϕ) freely splits off (L(F∞), τ ) and the free absorption property
If µ({a}) > 0 for some a 6= 0, then (M, ϕ) freely splits off an almost
immediately holds.
periodic free Araki -- Woods factor of type III and the free absorption property follows from
[Sh96, Theorem 5.4]. Conversely, if µa = 0, it follows from Corollary C that M has all its
centralizers amenable. But then M cannot have the free absorption property.
(cid:3)
22
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
One of the most intriguing isomorphism questions for free Araki -- Woods factors, well outside
the scope of our methods, is whether Γ(λ, 1)′′ ∼= Γ(λ + δ0, 1)′′ ? In [Sh97a, Theorem 4.8], it was
shown that the continuous core of Γ(λ, 1)′′ is isomorphic with B(ℓ2(N)) ⊗ L(F∞). We prove
that the same holds for Γ(λ + δ0, 1)′′. Note here that in [Ha15, Corollary 1.10], it is proved
that if µ is singular w.r.t. the Lebesgue measure λ, then the continuous core of Γ(µ, m)′′ is
never isomorphic with B(ℓ2(N)) ⊗ L(F∞). Under the stronger assumption that all convolution
powers µ∗n are singular w.r.t. the Lebesgue measure, this was already shown in [Sh02].
Proposition 7.5. The continuous core of Γ(λ + δ0, 1)′′ is isomorphic with B(ℓ2(N)) ⊗ L(F∞).
Proof. In [Sh97a, Sh97b], von Neumann algebras generated by A-valued semicircular elements
are introduced. In the special case where A is semifinite and equipped with a fixed faithful
normal semifinite trace Tr, this construction can be summarized as follows.
Let H be a Hilbert A-bimodule, meaning that H is a Hilbert space equipped with a normal
homomorphism A → B(H) and a normal anti-homomorphism A → B(H) having commuting
images. We denote the left and right action of A on H as a · ξ · b for all a, b ∈ A, ξ ∈ H. Further
assume that S is an A-anti-bimodular involution on H. More precisely, S is a closed, densely
defined operator on H such that S(ξ) ∈ D(S) with S(S(ξ)) = ξ for all ξ ∈ D(S) and such that
for all ξ ∈ D(S) and all a, b ∈ A, we have a · ξ · b ∈ D(S) and S(a · ξ · b) = b∗ · S(ξ) · a∗. Define
FA(H) = L2(A, Tr) ⊕
∞Mk=1(cid:0)H ⊗A H ⊗A · · · ⊗A H
(cid:1) .
}
{z
k factors
A vector ξ ∈ H is called right bounded if there exists a κ > 0 such that kξ · ak ≤ κkak2,Tr for
all a ∈ nTr. For every ξ ∈ H, there exists an increasing sequence of projections pn ∈ A such
that pn → 1 strongly and ξ · pn is right bounded for all n. So, the subspace H0 ⊂ H defined as
H0 = {ξ ∈ D(S) : ξ and S(ξ) are right bounded }
is dense. For every right bounded vector ξ ∈ H, we have a natural left creation operator
ℓ(ξ) ∈ B(FA(H)). Then we define
Φ(A, Tr, H, S) =(cid:0)A ∪ {ℓ(ξ) + ℓ(S(ξ))∗ : ξ ∈ H0}(cid:1)′′ .
There is a normal conditional expectation E : Φ(A, Tr, H, S) → A given by E(x)P = P xP ,
where P : FA(H) → L2(A, Tr) is the orthogonal projection. By [Sh97b, Proposition 5.2], we
have that E is faithful. By construction, if A = C1 and Tr(1) = 1, and using the notation
introduced before Lemma 4.3, we find the free Araki -- Woods factor Φ(C1, Tr, H, S) ∼= Γ(H, S)′′.
The above construction can be applied to a normal completely positive map ϕ : A → A
satisfying Tr(ϕ(a)b) = Tr(aϕ(b)) for all a, b ∈ mTr. To such a map ϕ, we associate the Hilbert
A-bimodule Hϕ by separation and completion of A ⊗ nTr with inner product
ha ⊗ϕ b, c ⊗ϕ di = Tr(b∗ϕ(a∗c)d) .
We also define the anti-unitary involution S(a ⊗ϕ b) = b∗ ⊗ϕ a∗. We denote the resulting von
Neumann algebra Φ(A, Tr, H, S) as Φ(A, Tr, ϕ).
Given a trace preserving inclusion (A, Tr) ⊂ (D, Tr), we denote by E : D → A the unique
Tr-preserving conditional expectation and define ψ : D → D : ψ(d) = ϕ(E(d)) for all d ∈ D.
Then, the functor Φ satisfies
(7.4)
Φ(A, Tr, ϕ) ∗A D ∼= Φ(D, Tr, ψ)
where the amalgamated free product is taken w.r.t. the canonical conditional expectations.
Let (M, ϕ) = (Γ(λ, 1)′′, ϕλ,1). We can then reformulate [Sh97a, Theorem 4.1] as
(7.5)
cϕ(M ) ∼= Φ(A, Tr, ϕ)
CLASSIFICATION OF A FAMILY OF NON ALMOST PERIODIC FREE ARAKI -- WOODS FACTORS
23
where A = L∞(R), Tr(f ) = RR f (x) exp(−x) dx and ϕ : A → A is such that the associated
A-bimodule H is isomorphic with the coarse A-bimodule L2(R2) with anti-unitary involution
(Sξ)(x, y) = ξ(y, x). Under the identification Lϕ(R) ∼= L∞(R), the isomorphism in (7.5)
respects the canonical conditional expectations cϕ(M ) → Lϕ(R) and Φ(A, Tr, ϕ) → A.
By [Sh97a, Theorem 4.8], we have in this particular case that Φ(A, Tr, ϕ) ∼= B(ℓ2(N)) ⊗ L(F∞).
Let now (D, TrD) be an arbitrary diffuse abelian von Neumann algebra with a faithful normal
semifinite trace satisfying Tr(1) = +∞ and let ψ : D → D be any normal completely positive
map satisfying TrD(ψ(c)d) = TrD(cψ(d)) for all c, d ∈ mTrD such that the associated D-
bimodule H and anti-unitary involution S are isomorphic with L2(D, TrD) ⊗ L2(D, TrD) with
S(c ⊗ d) = d∗ ⊗ c∗. Since there exists an isomorphism α : D → A of D onto A satisfying
Tr ◦ α = TrD, it follows that
(7.6)
Φ(D, TrD, ψ) ∼= Φ(D, TrD, L2(D, TrD) ⊗ L2(D, TrD), S)
∼= Φ(A, Tr, L2(A, Tr) ⊗ L2(A, Tr), S)
∼= Φ(A, Tr, ϕ) ∼= B(ℓ2(N)) ⊗ L(F∞) .
Write (M1, ϕ1) = (Γ(λ + δ0, 1)′′, ϕλ+δ0,1). Since (M1, ϕ1) ∼= (M, ϕ) ∗ (B, τ ) for some diffuse
abelian von Neumann algebra B with faithful normal state τ , it follows that
cϕ1(M1) ∼= cϕ(M ) ∗Lϕ(R) (Lϕ(R) ⊗ B) .
Since the isomorphism in (7.5) respects the conditional expectations, we conclude that
cϕ1(M1) ∼= Φ(A, Tr, ϕ) ∗A (A ⊗ B) ,
where the conditional expectation A ⊗ B → A is given by id ⊗ τ . Write D = A ⊗ B and define
ψ : D → D : ψ(d) = ϕ((id ⊗ τ )(d)) ⊗ 1. By (7.4), we get that
Φ(A, Tr, ϕ) ∗A (A ⊗ B) ∼= Φ(D, Tr ⊗ τ, ψ) .
Since D is diffuse abelian and the D-bimodule associated with ψ is isomorphic with the coarse
D-bimodule, it follows from (7.6) that Φ(D, Tr⊗τ, ψ) ∼= B(ℓ2(N))⊗L(F∞). So, the proposition
is proved.
(cid:3)
References
[BH16] R. Boutonnet, C. Houdayer, Structure of modular invariant subalgebras in free Araki -- Woods factors.
Anal. PDE 9 (2016), 1989 -- 1998.
[Co72] A. Connes, Une classification des facteurs de type III. Ann. Sci. ´Ecole Norm. Sup. 6 (1973), 133 -- 252.
[Co74] A. Connes, Almost periodic states and factors of type III1. J. Funct. Anal. 16 (1974), 415 -- 445.
[Co80] A. Connes, Classification des facteurs. In "Operator algebras and applications, Part 2 (Kingston,
1980)", Proc. Sympos. Pure Math. 38, Amer. Math. Soc., Providence, 1982, pp. 43-109.
[Dy92] K. Dykema, Free products of hyperfinite von Neumann algebras and free dimension. Duke Math. J. 69
(1993), 97 -- 119.
[Dy96] K. Dykema, Free products of finite-dimensional and other von Neumann algebras with respect to non-
tracial states. In "Free probability theory (Waterloo, ON, 1995)", Fields Inst. Commun. 12, Amer.
Math. Soc., Providence, 1997, pp. 41 -- 88.
[Ha77] U. Haagerup, Operator-valued weights in von Neumann algebras. I. J. Funct. Anal. 32 (1979), 175 -- 206.
[Ha85] U. Haagerup, Connes' bicentralizer problem and uniqueness of the injective factor of type III1. Acta
Math. 69 (1986), 95 -- 148.
[Ha15] B. Hayes, 1-bounded entropy and regularity problems in von Neumann algebras. Int. Math. Res. Not.
IMRN, to appear. arXiv:1505.06682.
[Ho06] C. Houdayer, On some free products of von Neumann algebras which are free Araki-Woods factors.
Int. Math. Res. Not. IMRN 23 (2007), art. id. rnm098.
[Ho08a] C. Houdayer, Free Araki-Woods factors and Connes' bicentralizer problem. Proc. Amer. Math. Soc.
137 (2009), 3749-3755.
[Ho08b] C. Houdayer, Structural results for free Araki -- Woods factors and their continuous cores. J. Inst. Math.
Jussieu 9 (2010), 741 -- 767.
24
CYRIL HOUDAYER, DIMITRI SHLYAKHTENKO, AND STEFAAN VAES
[HI15] C. Houdayer, Y. Isono, Unique prime factorization and bicentralizer problem for a class of type III
factors. Adv. Math. 305 (2017), 402 -- 455.
[HR10] C. Houdayer, ´E. Ricard, Approximation properties and absence of Cartan subalgebra for free Araki --
Woods factors. Adv. Math. 228 (2011), 764 -- 802.
[HU15] C. Houdayer, Y. Ueda, Rigidity of free product von Neumann algebras. Compos. Math. 152 (2016),
2461 -- 2492.
[ILP96] M. Izumi, R. Longo, S. Popa, A Galois correspondence for compact groups of automorphisms of von
Neumann algebras with a generalization to Kac algebras. J. Funct. Anal. 155 (1998), 25 -- 63.
[Kr75] W. Krieger, On ergodic flows and the isomorphism of factors. Math. Ann. 223 (1976), 19 -- 70.
[LP97] M. Lemanczyk, F. Parreau, On the disjointness problem for Gaussian automorphisms. Proc. Amer.
Math. Soc. 127 (1999), 2073 -- 2081.
[Oz03] N. Ozawa, Solid von Neumann algebras. Acta Math. 192 (2004), 111 -- 117.
[Po02] S. Popa, On a class of type II1 factors with Betti numbers invariants. Ann. of Math. 163 (2006),
809 -- 899.
[Po03] S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, I. Invent. Math.
165 (2006), 369 -- 408.
[Po67] R.T. Powers, Representations of uniformly hyperfinite algebras and their associated von Neumann
rings. Ann. of Math. 86 (1967), 138 -- 171.
[Ra91] F. Radulescu, The fundamental group of the von Neumann algebra of a free group with infinitely many
generators is R+ \ {0}. J. Amer. Math. Soc. 5 (1992), 517 -- 532.
[Ru62] W. Rudin, Fourier analysis on groups. Interscience Tracts in Pure and Applied Mathematics 12, John
Wiley and Sons, New York, London, 1962.
[Sh96] D. Shlyakhtenko, Free quasi-free states. Pacific J. Math. 177 (1997), 329 -- 368.
[Sh97a] D. Shlyakhtenko, Some applications of freeness with amalgamation. J. Reine Angew. Math. 500
(1998), 191 -- 212.
[Sh97b] D. Shlyakhtenko, A-valued semicircular systems. J. Funct. Anal. 166 (1999), 1 -- 47.
[Sh02] D. Shlyakhtenko, On the classification of full factors of type III. Trans. Amer. Math. Soc. 356 (2004),
4143 -- 4159.
[Sh03] D. Shlyakhtenko, On multiplicity and free absorption for free Araki -- Woods factors. Preprint.
arXiv:math/0302217
[Ta03] M. Takesaki, Theory of operator algebras. II. Encyclopaedia of Mathematical Sciences, 125. Operator
Algebras and Non-commutative Geometry, 6. Springer-Verlag, Berlin, 2003. xxii+518 pp.
Laboratoire de Math´ematiques d'Orsay, Universit´e Paris-Sud, CNRS, Universit´e Paris-Saclay,
91405 Orsay, France
E-mail address: [email protected]
Mathematics Department, UCLA, Los Angeles, CA 90095-1555, United States
E-mail address: [email protected]
KU Leuven, Department of Mathematics, Celestijnenlaan 200B, B-3001 Leuven, Belgium
E-mail address: [email protected]
|
1412.8240 | 3 | 1412 | 2016-02-03T16:55:13 | Crossed products by endomorphisms of $C_0(X)$-algebras | [
"math.OA"
] | In the first part of the paper, we develop a theory of crossed products of a $C^*$-algebra $A$ by an arbitrary (not necessarily extendible) endomorphism $\alpha:A\to A$. We consider relative crossed products $C^*(A,\alpha;J)$ where $J$ is an ideal in $A$, and describe up to Morita-Rieffel equivalence all gauge invariant ideals in $C^*(A,\alpha;J)$ and give six term exact sequences determining their $K$-theory. We also obtain certain criteria implying that all ideals in $C^*(A,\alpha;J)$ are gauge invariant, and that $C^*(A,\alpha;J)$ is purely infinite.
In the second part, we consider a situation where $A$ is a $C_0(X)$-algebra and $\alpha$ is such that $\alpha(f a)=\Phi(f)\alpha(a)$, $a\in A$, $f\in C_0(X)$ where $\Phi$ is an endomorphism of $C_0(X)$. Pictorially speaking, $\alpha$ is a mixture of a topological dynamical system $(X,\varphi)$ dual to $(C_0(X),\Phi)$ and a continuous field of homomorphisms $\alpha_x$ between the fibers $A(x)$, $x\in X$, of the corresponding $C^*$-bundle.
For systems described above, we establish efficient conditions for the uniqueness property, gauge-invariance of all ideals, and pure infiniteness of $C^*(A,\alpha;J)$. We apply these results to the case when $X=$Prim$(A)$ is a Hausdorff space. In particular, if the associated $C^*$-bundle is trivial, we obtain formulas for $K$-groups of all ideals in $C^*(A,\alpha;J)$. In this way, we constitute a large class of crossed products whose ideal structure and $K$-theory is completely described in terms of $(X,\varphi,\{\alpha_{x}\}_{x\in X};Y)$ where $Y$ is a closed subset of $X$. | math.OA | math |
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
B. K. KWAŚNIEWSKI
Abstract. In the first part of the paper, we develop a theory of crossed products of a
C -algebra A by an arbitrary (not necessarily extendible) endomorphism α : A Ñ A. We
consider relative crossed products C pA, α; Jq where J is an ideal in A, and describe up
to Morita-Rieffel equivalence all gauge-invariant ideals in C pA, α; Jq and give six term
exact sequences determining their K-theory. We also obtain certain criteria implying that
all ideals in C pA, α; Jq are gauge-invariant, and that C pA, α; Jq is purely infinite.
In the second part, we consider a situation where A is a C0pXq-algebra and α is such
that αpf aq " Φpf qαpaq, a P A, f P C0pXq where Φ is an endomorphism of C0pXq.
Pictorially speaking, α is a mixture of a topological dynamical system pX, ϕq dual to
pC0pXq, Φq and a continuous field of homomorphisms αx between the fibers Apxq, x P X,
of the corresponding C -bundle.
For systems described above, we establish efficient conditions for the uniqueness prop-
erty, gauge-invariance of all ideals, and pure infiniteness of C pA, α; Jq. We apply these
results to the case when X " PrimpAq is a Hausdorff space. In particular, if the associated
C -bundle is trivial, we obtain formulas for K-groups of all ideals in C pA, α; Jq. In this
way, we constitute a large class of crossed products whose ideal structure and K-theory is
completely described in terms of pX, ϕ, tαxuxPX ; Y q where Y is a closed subset of X.
Introduction.
Crossed products by endomorphisms proved to be one of the major model examples in
classification of simple C -algebras. The first instances of such crossed products, infor-
mally introduced in [9], were Cuntz algebras On. Rørdam [49] and Rørdam and Elliot [12]
established the range of K-theoretical invariant for all Kirchberg algebras by showing that
crossed products by endomorphisms of AT-algebras of real rank zero contain classifiable
Kirchberg algebras with arbitrary K-theory.
In particular, by Kirchberg-Phillips classi-
fication, every Kirchberg algebra is isomorphic to such a crossed product. Significantly,
Elliott's classification of (not necessarily simple) AT-algebras of real rank zero [11] implies
that all unital simple AT-algebras of real rank zero with K1 equal to integers are modeled
by crossed products associated to Cantor systems, studied by Putnam [48], see also [17].
Another milestone in the classification of non-simple C -algebras is Kirchberg's classifica-
tion of strongly purely infinite, nuclear, separable C -algebras via ideal related KK-theory
[26]. Nevertheless, this invariant is fairly complicated and there is still a lot of effort put
into classifying certain non-simple purely infinite C -algebras by means of apparently less
elaborated invariants, cf. [41], [7], [50]. Accordingly, it is of interest to establish non-trivial
but still accessible examples of C -algebras whose ideal structure and K-theory of all ideals
2000 Mathematics Subject Classification. 46L05; Secondary 46L35.
Key words and phrases. crossed product, endomorphism, C0pXq-algebra, pure infiniteness, K-theory,
ideal structure.
1
2
B. K. KWAŚNIEWSKI
and quotients can be controlled. An overall aim of the present paper is to develop tools to
construct and analyze a large class of crossed products by endomorphisms that fulfill these
requirements. Another source of motivation comes from potential applications to spectral
analysis of certain non-local operators [4], [22], [5].
In section 3, we introduce and study C0pXq-dynamical systems. A C0pXq-dynamical
system is a pair pA, αq where A is a C0pXq-algebra and α : A Ñ A is an endomorphism
compatible with the C0pXq-structure (we give several characterization of this notion). Such
a system can be viewed as a convenient combination of topological and noncommutative
dynamics; encoded in a pair pϕ, tαxuxP∆q where ϕ : ∆ Ñ X is a continuous proper mapping
defined on an open set ∆ Ď X, and αx : Apϕpxqq Ñ Apxq, x P ∆, is a homomorphism
between the corresponding fibers of the C0pXq-algebra A, so that
αpaqpxq " αxpapϕpxqqq,
a P A, x P ∆.
We refer to the pair pϕ, tαxuxP∆q as to a morphism of the corresponding C -bundle A :"
Apxq (Definition 3.1). In particular, if every fiber Apxq is trivial (equal to C) we get a
ŮxPX
topological dynamical system. In the case when X is trivial (a singleton), α : A Ñ A is just
an endomorphism of A, and we call pA, αq simply a C -dynamical system. An important
non-trivial example arises when A is a unital C -algebra, C Ď ZpAq is a non-degenerate
C -subalgebra of the center of A and α 'almost preserves' C, that is αpCq Ď Cαp1q.
Then pA, αq is naturally a C0pXq-dynamical system with C0pXq -- C. Analysis of crossed
products associated to such C0pXq-dynamical systems, in the case α is an automorphism,
played an important role in the study of non-local operators, cf.
[22], such as (abstract)
weighted shift operators [4], or singular integral operators with shifts [5]. If C " ZpAq and
the primitive ideal space PrimpAq of A is a Hausdorff space, then X -- PrimpAq, and this
will be our model example.
We associate to any C -dynamical system pA, αq and an ideal J in A the relative crossed
product C pA, α; Jq introduced (for an arbitrary completely positive map) in [35]. As
explained in detail in [35, Section 3.4], these crossed products include as special cases those
studied in [44], [54], [40], [37], [33]. In the present paper, we consider only the case when A
embeds into C pA, α; Jq, equivalently when J is contained in the annihilator pker αqK of the
kernel of α (the general case may be covered by passing to a quotient C -dynamical system,
cf. Remark 2.8 below). The unrelative crossed product is C pA, αq :" C pA, α; pker αqKq. If
pA, αq is a C0pXq-dynamical system with the related morphism pϕ, tαxuxP∆q, then among
our main results we list the following:
‚ Isomorphism theorem. We show that for certain continuous C0pXq-algebras, if the
map ϕ is topologically free outside a set Y related to the ideal J, then every injective
representation of pA, αq whose ideal of covariance is maximal possible, give rise to a faithful
representation of C pA, α; Jq (see Theorem 4.11).
‚ Description of the ideal structure. We prove that, if ϕ is free, then we have a bijective
correspondence between ideals I in C pA, α; Jq and certain pairs pI, I 1q of ideals in A,
called J-pairs for pA, αq (see Theorem 4.12 and Definition 2.17). Moreover, the quotient
of C pA, α; Jq by I is naturally isomorphic a crossed product associated to the quotient of
pA, αq, and the ideal I is Morita-Rieffel (strongly Morita) equivalent either to the crossed
product associated to the restricted endomorphism αI or to an endomorphism constructed
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
3
from pI, I 1q and α (see Theorem 2.19 and Proposition 2.25). In the case A has a Hausdorff
primitive ideal space and X " PrimpAq, we describe ideals in C pA, α; Jq in terms of pairs
pV, V 1q of closed subsets of X, called Y -pairs for pX, ϕq (see Proposition 5.6 and Definition
5.1). Hence the ideal structure of C pA, α; Jq is completely described in terms of the
topological dynamical system pX, ϕq. In particular, in this case we characterize simplicity
of crossed products (Proposition 5.9).
‚ Pure infinitneness. It seems that amongst the existing technics of showing pure in-
finiteness of crossed products there are two types of approaches. In the first one the corre-
sponding crossed product is simple [39], [19], [20]. In the second one the initial algebra A
is assumed to have the ideal property [52], [16], [43], [46], [38]. We cover these two lines of
research in our context by showing that if ϕ is free, then
A is purely infinite and
has the ideal property
ùñ the same is true for C pA, α, Jq.
and
A is purely infinite and there are
finitely many J-pairs for pA, αq
ùñ
C pA, α, Jq is purely infinite and
has finitely many ideals
(see Theorem 4.12). If X " PrimpAq and A is purely infinite, this leads us to necessary
and sufficient conditions for C pA, αq to be a Kirchberg algebra (Corollary 5.10). We
recall that in the presence of the ideal property, pure infiniteness is equivalent to strong
pure infiniteness. We also point out that using conditions introduced recently in [38], the
aforementioned results could be potentially generalized to the case when A is not necessarily
purely infinite. Moreover, in recent papers [28], [29] Sierakowski and Kirchberg introduced
a new machinery that gives strong pure infiniteness criteria for crossed products by discrete
group actions, without passing (explicitly) through pure infiniteness. It seems plausible that
combining their technics with tools of the present paper one could also obtain permanence
results for strong pure infiniteness of C pA, α, Jq. Nevertheless, we do not pursue these
issues here.
‚ K-theory.
In the case when the corresponding C -bundle is trivial, that is when
A " C0pX, Dq for a C -algebra D, and under the assumptions that X is totally discon-
nected, K0pDq is torsion free and K1pDq " 0, we give formulas for K-groups of C pA, α, Jq
formulated in terms of pX, ϕ, tαxuxP∆; Y q (Proposition 5.13). These formulas can be viewed
as a far reaching generalization of those given by Putnam in [48]. If additionally D is sim-
ple and ϕ is free we get formulas for K-groups of all ideals in C pA, α, Jq (Theorem 5.15).
We show by concrete examples that not only the dynamical system pX, ϕq (which deter-
mines the ideal structure of C pA, α, Jq) but also endomorphisms αx, x P ∆, contribute to
K-theory, thus giving us a lot of flexibility in constructing interesting algebras.
The aforementioned results are based on general facts for crossed products C pA, α; Jq,
which we develop in section 2. One of the main tools is a description of a reversible J-
extension pB, βq of pA, αq, introduced in [33]: we show that if pA, αq is a C0pXq-dynamical
system, then pB, βq is a C0prXq-dynamical system induced by a morphism prϕ, tβxuxPr∆q
where pX,rϕq is a reversible Y -extension of pX, ϕq introduced in [30] (see Theorem 4.9).
[44], [54], [40], [37], [33], crossed products by
endomorphisms where studied either in the case A is unital or under the assumption that
It has to be emphasized that so far, cf.
4
B. K. KWAŚNIEWSKI
the endomorphism α : A Ñ A is extendible [1], i.e. that it extends to an endomorphism
of the multiplier algebra MpAq of A. However, these assumptions exclude a number of
important applications. For instance, a restriction of an extendible endomorphism to an
invariant ideal in general is not extendible. Thus, we are forced to develop a large part
of theory of crossed products by not necessarily extendible endomorphisms. We do it in
section 2. The established results are interesting in their own right.
More specifically, we generalize one of the main results of [33] and describe the gauge-
invariant ideals I in C pA, α; Jq by J-pairs pI, I 1q of ideals in A. Additionally, we show
that I is Morita-Rieffel equivalent either to the crossed product associated to the restricted
endomorphism αI or to an endomorphism constructed from pI, I 1q and α (Theorem 2.19
and Proposition 2.25). We generalize the classic Pimsner-Voiculescu sequence so that it
applies to the crossed product C pA, α; Jq (Proposition 2.26). As a consequence we get six-
term exact sequences for K-groups of all gauge-invariant ideals in C pA, α; Jq (Theorem
2.27). We extend the terminology of [33] and say that pA, αq is a reversible C -dynamical
systems if α has a complemented kernel and a hereditary range. For an arbitrary C -
dynamical system pA, αq we generalize the construction of a reversible J-extension pB, βq
introduced in [33], see also [36]. We show that
C pA, α; Jq -- C pB, βq
(Theorem 2.29). This is a powerful tool because for reversible systems pA, αq the crossed
product C pA, αq has an accessible structure, very similar to that of classical crossed prod-
uct by an automorphism.
In particular, for such systems we have natural criteria for
uniqueness property, gauge-invariance of all ideals, and pure infiniteness of C pA, αq (see
Propositions 2.35 and 2.46).
We note that, in contrast to [33] where more direct methods where used, in the present
paper we base our more general analysis on certain results for relative Cuntz-Pimsner
algebras and an identification of C pA, α; Jq as such an algebra. We present the relevant
facts in Appendix A.
The content is organized as follows: We recall the relevant notions and facts concern-
ing C0pXq-algebras in section 1. General crossed products are studied in section 2.
In
section 3 we introduce and analyze C0pXq-dynamical systems. Section 4 contains general
main results for C0pXq-dynamical systems. We apply them to C0pXq-dynamical systems
with X " PrimpAq in section 5, where our results attain a particularly nice form. We
finish the paper with Appendix A, which contains relevant facts from the theory of C -
correspondences and relative Cuntz-Pimsner algebras, as well as a discussion of a particular
case of the C -correspondence Eα associated to pA, αq.
0.1. Notation and conventions. The set of natural numbers N starts from zero. All
ideals in C -algebras are assumed to be closed and two-sided. All homomorphisms be-
tween C -algebras are by definition -preserving. For actions γ : A B Ñ C such as
multiplications, inner products, etc., we use the notation:
γpA, Bq " spantγpa, bq : a P A, b P Bu.
If A is a C -algebra 1 denotes the unit in the multiplier C -algebra MpAq. The enveloping
von Neumann algebra of A is denoted by A. We recall, see [27, Theorem 4.16], that
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
5
a C -algebra A is purely infinite if and only if every a P A`zt0u is properly infinite, e.g.
a ' a À a ' 0 in M2pAq, where À Cuntz comparison of positive elements. We recall that for
a, b P A`, a À b in A if and only if for every ε ą 0 there is x P A` such that }a ´ xbx} ă ε.
A C -algebra A has the ideal property [47], [45], if every ideal in A is generated (as an
ideal) by its projections.
1. Preliminaries on C0pXq-algebras and C -bundles
In this section, we gather certain facts concerning C0pXq-algebras. We find it beneficial to
use two equivalent pictures of such objects: as C -algebras with a C0pXq-module structure
and as C -algebras of sections of C -bundles. Thus we implement both of these viewpoints.
As a general reference we use [55, Section C], but cf. also, for instance, [18], [6].
1.1. C -bundles and section C -algebras. Let X be a locally compact Hausdorff space.
Apxq such that
An upper semicontinuous C -bundle over X is a topological space A " ŮxPX
the natural surjection p : A Ñ X is open continuous, each fiber Apxq is a C -algebra, the
mapping A Q a Ñ }a} P R is upper semicontinuous, and the -algebraic operations in each
of the fibers are continuous in A, for details see [55, Definition C.16]. If additionally, the
mapping A Q a Ñ }a} P R is continuous, A is called a continuous C -bundle over X. For
each x P X, we denote by 0x the zero element in the fiber C -algebra Apxq, and by 1x
the unit in the multiplier algebra MpApxqq of Apxq. A C -bundle A is trivial if there is
Apxq onto X D which intertwines p
a C -algebra D and homeomorphism from A " ŮxPX
with the projection onto the first coordinate.
We denote by ΓpAq :" ta P CpX, Aq : ppapxqq " xu the set of continuous sections of
the upper semicontinuous C -bundle A. It is a -algebra with respect to natural pointwise
operations. Moreover, the set of continuous sections that vanish at infinity
Γ0pAq :" ta P ΓpAq : @εą0
tx P X : }apxq} ě εu is compactu
is a C -algebra with the norm }a} :" supxPX }apxq}. We call Γ0pAq the section C -algebra
of A. The section algebra Γ0pAq determines the topology of the C -bundle A. In particular,
we have the following lemma (see, for instance, the proof of [55, Theorem C.25]).
Lemma 1.1. A net tbiu converges to b in the C -bundle A if and only if ppbiq Ñ ppbq
and for each ε ą 0 there is a P Γ0pAq such that }apppbqq ´ b} ă ε and we eventually have
}apppbiqq ´ bi} ă ε.
The algebra Γ0pAq is naturally equipped with the structure of C0pXq-algebra given by
pf aqpxq :" f pxqapxq for f P C0pXq and a P Γ0pAq.
1.2. C0pXq-algebras. A C0pXq-algebra is a C -algebra A endowed with a nondegenerate
homomorphism µA from C0pXq into the center ZpMpAqq of the multiplier algebra MpAq
of A. When X is compact A is also called a CpXq-algebra. The C0pXq-algebra A is viewed
as a C0pXq-module where
f a :" µApf qa,
f P C0pXq, a P A.
Accordingly, the structure map µA : C0pXq Ñ ZpMpAqq is often suppressed. Using the
Dauns-Hofmann isomorphism we may identify ZpMpAqq with CbpPrim Aq, and then µA
6
B. K. KWAŚNIEWSKI
becomes the operator of composition with a continuous map σA : Prim A Ñ X. This map,
called the base map, is determined by the equivalence:
Apxq such that A becomes
(1)
C0pXztxuq A Ď P ðñ σApP q " x,
P P PrimpAq.
Let us fix a C0pXq-algebra A and consider a bundle A :" ŮxPX
It can be shown that there is a unique topology on A :" ŮxPX
Apxq :" A{´C0pXztxuq A¯,
x P X.
Apxq where
an upper semicontinuous C -bundle and the C0pXq-algebra A can be identified with Γ0pAq
by writing apxq for the image of a P A in the quotient algebra Apxq. Moreover, A is a
continuous C -bundle if and only if σA : Prim A Ñ X is an open map. In the latter case,
A is called a continuous C0pXq-algebra. In other words, we have the following statement,
see [55, Theorem C.26].
Theorem 1.2. A C -algebra A is a C0pXq-algebra if and only if A -- Γ0pAq where A is an
upper semicontinuous C -bundle. Moreover, A is a continuous C0pXq-algebra if and only
if A is a continuous C -bundle.
Convention 1.3. In the sequel we will freely pass (often without a warning) between the
above equivalent descriptions. Thus for any C0pXq-algebra A we will write A " Γ0pAq
where A is the associated C -bundle.
A pxq P " C0pXztxuq
A whenever x P σApPrim Aq, and Apxq " t0u if and only if x R σApPrim Aq. Thus if
σApPrimpAqq is locally compact (which is always the case when A is unital, or when A
Remark 1.4. Let A be a C0pXq-algebra. In view of (1), we haveŞP Pσ´1
is a continuous C0pXq-algebra), then we may treat A as a C0`σApPrimpAqq-algebra; in
other words, we may assume that σA is surjective, or equivalently that all fibers Apxq are
non-trivial.
1.3. Multiplier algebra of a C0pXq-algebra. We say that a C0pXq-algebra A has local
units if all fibers Apxq, x P X, are unital, and for any x P X there is a P A such that
apyq " 1y is the unit in Apyq for all y in a neighborhood of x.
Lemma 1.5. A C0pXq-algebra A is unital if and only if A has local units and the range of
σA is compact.
Proof. If 1 is the unit in A then σApPrimpAqq " tx P X : }1pxq} ě 1{2u is compact,
because 1 P Γ0pAq, and clearly the (global) unit 1 is a local unit for any point in X.
Conversely, suppose that σApPrimpAqq is compact and A has local units. Consider the
Apxq where for each x P A we let 1pxq :" 1x to be the unit in
function 1 : X Ñ A " ŮxPX
Apxq. Using Lemma 1.1 and local units one readily sees that 1 is a continuous section of
A. For any ε the set tx P X : }1pxq} ě εu is compact, as a closed subset of σApPrimpAqq.
Thus 1 P Γ0pAq " A.
(cid:3)
We have the following natural description of the multiplier algebra MpAq of a C0pXq-
MpApxqq, see [55, Lemma C.11]. We emphasize
algebra A as sections of the set MpAq :" ŮxPX
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
7
however, that in general (even when X is compact) MpAq can not be equipped with a
topology making it an upper semicontinuous C -bundle such that MpAq Ď ΓpMpAqq, see
[55, Example C.13].
Proposition 1.6. Suppose that A is a C0pXq-algebra. The multiplier algebra MpAq can
be naturally identified with the set of all functions m on X such that mpxq P MpApxqq,
for all x P X, and the functions x ÞÑ mpxqapxq, x ÞÑ apxqmpxq are in A " Γ0pAq for any
a P A. Then the C -algebraic structure of MpAq is given by the pointwise operations and
the supremum norm }m} " supxPX }mpxq}.
1.4. Ideals and quotients of a C0pXq-algebra. Fix a C0pXq-algebra A and let I be an
ideal in A. Assuming the standard identifications Prim I " tP P Prim A : I Ę P u and
PrimpA{Iq " tP P Prim A : I Ď P u, we see that both I and A{I are C0pXq-algebras with
base maps σA : PrimpIq Ñ X and σA : PrimpA{Iq Ñ X respectively. Moreover, we have
natural isomorphisms pA{Iqpxq -- Apxq{Ipxq where Ipxq " tapxq : a P I Ď Au, x P X.
Suppose that A is a continuous C0pXq-algebra. Then the ideal I is naturally a continuous
C0pY q-algebra for any locally compact set Y containing the open set σApPrimpIqq, because
a restriction of an open map to an open set is open (independently of the codomain). The
situation is quite different when dealing with a restriction to a closed set, and thus the
case of the quotient A{I is more delicate. Nevertheless, the set Y " σApPrimpA{Iqq is
locally compact, and the mapping σA : PrimpA{Iq Ñ Y is open for instance when I is
complemented or σA is injective. Translating this to the language of C -bundles we get the
following lemma.
Lemma 1.7. Suppose that I is an ideal in a C -algebra A " Γ0pAq of continuous sections
Apxq. The ideal I and the quotient algebra
of an upper semicontinuous C -bundle A " ŮxPX
A{I can be naturally treated as algebras of continuous sections of I " ŮxPX
ŮxPX
Apxq{Ipxq (equipped with unique topologies), respectively. Moreover, we have
Ipxq and A{I "
tx P X : Ipxq ‰ t0uu " σApPrimpIqq,
(2)
(3)
tx P X : Ipxq ‰ Apxqu " σApPrimpA{Iqq.
If A is a continuous bundle, then I is continuous over the set (2) and A{I is continuous
over the set (3) whenever I is complemented or σA is injective.
Proof. In view of the above discussion we only need to show (2) and (3). The equivalences
P ðñ DP Pσ´1
A pxqI Ę P ðñ x P σApPrimpIqq
P¯ ‰ A ðñ DP0PPrimpAq´I ` čP Pσ´1
A pxq
P¯ Ď P0
prove (2). To see (3) notice that using (1) we get
Ipxq ‰ t0u ðñ I Ę čP Pσ´1
Ipxq ‰ Apxq ðñ ´I ` čP Pσ´1
A pxq
A pxq
ðñ DP0PPrimpAq I Ď P0 and čP Pσ´1
A pxq
ðñ x P σApPrimpA{Iqq.
P Ď P0
8
B. K. KWAŚNIEWSKI
(cid:3)
Let I be an ideal in the C0pXq-algebra A. The annihilator I K " ta P A : aI " 0u of
I is also a C0pXq-algebra with the base map σA : PrimpI Kq Ñ X. Moreover, since I K is
the biggest ideal in A with the property that I X I K " t0u it follows that PrimpI Kq "
Int pPrimpA{Iqq. If A is a continuous C0pXq-algebra then I K is a continuous C0pUq-algebra
where U " σApInt pPrimpA{Iqqq. In terms of C -bundles, I K can be viewed as the algebra
IpxqK where IpxqK is contained in the
of continuous sections of the C -bundle I K :" ŮxPX
annihilator of Ipxq in Apxq. In particular, IpxqK " t0u if and only if x R U.
2. General crossed products by endomorphisms
In this section, we define crossed products C pA, α; Jq for an arbitrary (not necessar-
ily extendible) endomorphism α : A Ñ A. We establish basic results concerning the
structure of these C -algebras, including description of all gauge-invariant ideals and 'Pim-
sner Voiculescu sequences' determining their K-theory. We also construct a reversible
J-extension pB, βq of pA, αq, discuss the notion of topological freeness for systems which
are reversible or commutative, and give a general pure infiniteness criteria for reversible
systems.
2.1. C -dynamical systems and their crossed products. A C -dynamical system is
a pair pA, αq where A is a C -algebra and α : A Ñ A is an endomorphism. We say
that α, or that the system pA, αq, is extendible [1] if α extends to a strictly continuous
endomorphism α : MpAq Ñ MpAq. It is known to hold exactly when for some (and hence
any) approximate unit tµλu in A the net tαpµλqu converges strictly in MpAq. In contrast
to [33], in the present paper in general we do not assume that pA, αq is extendible.
Definition 2.1 (Definition 2.4 in [33]). A C -dynamical system pA, αq is called reversible
if ker α is a complemented ideal in A and αpAq is a hereditary subalgebra of A (briefly, α
has a complemented kernel and a hereditary range).
Remark 2.2. An extendible endomorphism α : A Ñ A has a hereditary range if and only
if it is a corner endomorphism, that is if αpAq is a corner in A (we then necessarily have
αp1qAαp1q " αpAq). In particular, an extendible C -dynamical system pA, αq is reversible
if and only if α is a corner endomorphism with complemented kernel.
Suppose that pA, αq is a reversible C -dynamical system. Then α : pker αqK ÞÑ αpAqAαpAq
is an isomorphism and we denote its inverse by α´1. If pA, αq is extendible, then αpAqAαpAq "
αp1qAαp1q and α´1 extends to a completely positive map α : A Ñ A given by the formula
(4)
αpaq " α´1pαp1qaαp1qq,
a P A.
The map α is a transfer operator for pA, αq in the sense of Exel [13], that is we have
αpαpaqbq " aαpbq, for all a, b P A. Moreover, α is regular, which means that α α is
a conditional expectation onto αpAq. In fact, α is a unique regular transfer operator for
pA, αq, see [35, Proposition 4.15]. Transfer operators satisfying (4) appear in a natural way
in a number of papers, see for instance [13], [2], [34], [33].
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
9
Example 2.3. If A " C0pXq where X is a locally compact Hausdorff space, then every
endomorphism α : A Ñ A is of the form
αpaqpxq "#apϕpxqq, x P ∆,
x R ∆,
0,
where ϕ : ∆ Ñ X is a continuous proper mapping defined on an open subset ∆ Ď X. Note
that, properness of ϕ implies that ϕp∆q is closed in X. We call the pair pX, ϕq a partial
dynamical system dual to pA, αq. The endomorphism α is extendible if and only if ∆ is
closed. The kernel of α is a complemented ideal in A if and only if ϕp∆q is open. The pair
pA, αq is a reversible C -dynamical system if and only if both ∆ and ϕp∆q are open in X
and ϕ : ∆ Ñ ϕp∆q is a homeomorphism. If the latter conditions are satisfied we say that
pX, ϕq is a reversible partial dynamical system.
Now, we turn to the definition of crossed products. For more details, in the case A is
unital or α is extendible, see [37] and [33].
Definition 2.4. A representation pπ, Uq of a C -dynamical system pA, αq on a Hilbert space
H consists of a non-degenerate representation π : A Ñ BpHq and an operator U P BpHq
such that
U πpaqU " πpαpaqq,
(5)
We will occasionally deal with representations of pA, αq in a C -algebra B by which mean a
pair pπ, Uq where π : A Ñ B is a non-degenerate homomorphism and U P B (an element
of the enveloping von Neumann algebra of B) satisfies (5). If π is injective then we say
pπ, Uq is injective.
for all a P A.
Let pπ, Uq be a representation of pA, αq in a C -algebra B. Then U is necessarily a
partial isometry. Indeed, if tµλu is an approximate unit in A then by non-degeneracy of
π, tπpµλqu converges σ-weakly to the unit in B and therefore tπpαpµλqqu " tU πpµλqU u
converges σ-weakly to U U . Hence, using multiplicativity of α, we get that
pU U q2 " σ- lim
λ
σ- lim
λ1
πpαpµλqqπpαpµλ1qq " σ- lim
λ
πpαpµλqq " U U
is a projection, cf.
[35, Proposition 3.21]. Moreover, see [35, Proposition 3.21] or the
proof of [37, Lemma 1.2], multiplicativity of α implies that the initial projection U U of U
commutes with the elements of πpAq. In particular,
Ipπ,U q :" ta P A : U U πpaq " πpaqu
is an ideal in A. If an ideal J in A is contained in Ipπ,U q we say that the representation
pπ, Uq is J-covariant. If pπ, Uq is pker αqK-covariant, that is if
a P pker αqK ùñ πpaq " U U πpaq
we say that pπ, Uq is a covariant representation. Note that if α is injective, then the
representation pπ, Uq is covariant if and only if U is an isometry. Thus if α is injective and
non-degenerate then the representation pπ, Uq is covariant if and only if U is a unitary. The
special role of pker αqK is also indicated in the following fact.
Lemma 2.5. Suppose that pπ, Uq is an injective representation of pA, αq. Then Ipπ,U q Ď
pker αqK. In particular, Ipπ,U q " pker αqK if and only if pπ, Uq is a covariant representation.
10
B. K. KWAŚNIEWSKI
Proof. Let a P Ipπ,U q and b P ker α. Then ab P Ipπ,U q and
πpabq " U U πpabqU U " U πpαpabqqU " U πpαpaqqπpαpbqqU " 0.
Hence ab " 0 because π is injective. Accordingly, Ipπ,U q Ď pker αqK.
(cid:3)
Combining the above lemma and the following proposition one can see that pA, αq admits
an injective J-covariant representation if and only if J Ď pker αqK.
Proposition 2.6. For any C -dynamical system pA, αq and any ideal J in pker αqK there
exists a C -algebra C pA, α; Jq containing A as a non-degenerate C -algebra and an oper-
ator u P C pA, α; Jq such that
a) C pA, α; Jq is generated (as a C -algebra) by A Y uA,
αpaq " uau for each a P A and
J " ta P A : uua " au,
b) for every J-covariant representation pπ, Uq of pA, αq there is a representation π ¸ U
of C pA, α; Jq determined by relations pπ¸Uqpaq " πpaq, a P A, and pπ¸Uqpuq " U.
Moreover, if α is extendible, then u P MpC pA, α; Jqq and C pA, α; Jq " C pA Y Auq.
Proof. Existence of C pA, α; Jq with the prescribed properties can be deduced from Propo-
sitions A.8 and A.1. It also follows from [35, Proposition 3.26] which in essence states that
C pA, α; Jq is a special case of the crossed product defined in [35, Definition 3.5] (note that
the identification of the aforementioned algebras goes thorough the equality s " u). In
particular, [35, Remark 3.11] implies that u P C pA, α; Jq and when α is extendible then
u P MpC pA, α; Jqq. If α is extendible then C pA, α; Jq " C pA Y Auq by [35, Lemma
3.23].
(cid:3)
Universal properties of the C -algebra C pA, α; Jq imply that, up to a natural isomor-
phism, it is uniquely determined by the triple pA, α, Jq.
Definition 2.7. We define the relative crossed product associated to a C -dynamical system
pA, αq and an ideal J in pker αqK to be the C -algebra described in Proposition 2.6. We
also write C pA, αq :" C pA, α, pker αqKq and call it the (unrelative) crossed product of A
by α.
Remark 2.8. In the case A is unital or α is extendible, the C -algebra C pA, α; Jq was
studied respectively in [37] and [33]. In general, the crossed product C pA, α; Jq is a special
case of the one defined in [35, Definition 3.5] where α is treated as a completely positive
map, or the one introduced in [31, Definition 4.9] where α is treated as a partial morphism
of A. In particular, one could consider the crossed product C pA, α; Jq for an arbitrary
ideal J in A, not necessarily contained in pker αqK, cf. [35], or [31]. However, if J Ę pker αqK
the algebra A does not embed into C pA, α; Jq. Moreover, as described in [37, Section 5.3],
see also [31, Example 6.24], or [33, Remark 4.4], by passing to a quotient C -dynamical
system, one can always reduce this seemingly more general situation to that of Definition
2.7.
By universal property of the crossed product C pA, α; Jq, there is a circle action T " tz P
C : z " 1u Q z ÞÝÑ γz P AutpC pA, α; Jqq determined by relations γzpaq " a, γzpuq " zu,
a P A, z P T. We call γ " tγzuzPT the gauge action on C pA, α; Jq. We say that a
representation pπ, Uq of pA, αq admits a gauge action if the relations γzpπpaqq " πpaq,
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
11
γzpUq " zU, a P A, z P T, determine a circle action on the C -algebra generated by
πpAq Y U πpAq. We have the following version of the gauge-uniqueness theorem.
Proposition 2.9. For any injective J-covariant representation pπ, Uq the homomorphism
π ¸ U of C pA, α; Jq is injective if and only if Ipπ,U q " J and pπ, Uq admits a gauge action.
In particular, if pπ, Uq is a covariant representation of pA, αq then the homomorphism
π ¸ U of C pA, αq is injective if and only if π is injective and pπ, Uq and admits a gauge
action.
Proof. The first part of the assertion follows from Propositions A.2 and A.8. For the second
part apply Lemma 2.5.
(cid:3)
We list certain general permanence properties for the crossed products C pA, α; Jq.
Proposition 2.10. Let pA, αq be a C -dynamical system and let ideal J be an ideal in
pker αqK.
(i) A is exact ðñ C pA, α; Jq is exact.
(ii) A is nuclear ùñ C pA, α; Jq is nuclear.
(iii) If A is separable, nuclear, and both A and J satisfy the UCT, then C pA, α; Jq
satisfies the UCT.
Proof. By Proposition A.8, we have C pA, α; Jq -- OpJ, Eαq. Since OpJ, Eαq is the quotient
of the Toeplitz algebra TEα " Cpt0u, Eαq, item (i) follows from [24, Theorem 7.1]. Similarly,
[24, Theorem 7.2] implies (ii). The argument leading to [24, Proposition 8.8] gives (iii). (cid:3)
2.2. Algebraic structure of crossed products. The -algebraic structure underlying
the crossed product C pA, α; Jq, that could actually be used to construct C pA, α; Jq, cf.
[31, Example 2.20 and Definition 4.9], is described in the following proposition.
Proposition 2.11. For any C -dynamical system pA, αq and any ideal J in pker αqK the
universal operator u P C pA, α; Jq is a power partial isometry, that is tunununPN and
tunununPN are decreasing sequences of mutually commuting projections. Moreover,
(6)
so the projections unun commute with elements of A. The elements
for all a P A, n P N,
una " αnpaqun,
(7)
a "
unan,mum,
an,m P αnpAqAαmpAq, n, m " 1, ..., N, N P N,
form a dense -subalgebra of C pA, α; Jq and their products are determined by the formula
Nÿn,m"1
(8)
punan,mumq`um`kam`k,lul " un`kαkpan,mqam`k,lul,
Proof. The first part of the assertion follows from [35, Proposition 3.21] and the fact that
if pπ, Uq is a representation of pA, αq, then pπ, U nq is a representation of pA, αnq, n P N.
Let us show (8). Since an,m P AαmpAq we have an,mumum " an,m and by (6) we get
an,muk " ukαkpan,mq. Thus
punan,mum q`um`kam`k,l ul " unan,mukam`k,lul " un`kαkpan,mqam`k,lul.
Now, using (8), one readily sees that elements (7) form a -algebra generated by A and
uA " αpAqu.
(cid:3)
n, m, k, l P N.
12
B. K. KWAŚNIEWSKI
Corollary 2.12. The initial projection uu of the universal partial isometry u P C pA, α; Jq
belongs to the multiplier algebra MpC pA, α; Jqq of C pA, α; Jq.
Proof. Let an,m P αnpAqAαmpAq, n, m P N. If n ą 0, then puuqunan,mum " unan,mum P
C pA, α; Jq. If n " 0 then
puuqunan,mum " puuqa0,mum " uαpa0,mqum`1 P C pA, α; Jq.
By Proposition 2.11 we get puuqC pA, α; Jq Ď C pA, α; Jq and consequently (since uu is
self-adjoint) uu P MpC pA, α; Jqq.
(cid:3)
If the kernel of α is complemented then the initial projection uu of the universal partial
isometry u P C pA, αq can also be treated as a multiplier of A:
Lemma 2.13. Suppose that pA, αq is a C -dynamical system such that the kernel of α is
a complemented ideal in A. Let u P C pA, αq be the universal partial isometry. Then
uuA " pker αqK and uαpaqu " uua,
a P A.
If pA, αq is a reversible C -dynamical system then for every n P N the system pA, αnq is
also reversible and
(9)
ununA " pker αnqK and unαnpaqun " ununa,
a P A.
Proof. We have pker αqK " ta P A : uua " au Ď uuA and upqu maps both of the C -
algebras pker αqK and uuA isomorphically onto αpAq " uAu. This implies that pker αqK "
uuA. In particular, uua " uuauu " uαpaqu for every a P A.
Now assume that pA, αq is reversible. Note that a composition of two homomorphisms
f : A Ñ B and g : B Ñ C with hereditary ranges have a hereditary range. The latter
holds because
gpf pAqqCgpf pAqq " gpf pAqqgpBqCgpBqgpf pAqq " gpf pAqqgpBqgpf pAqq
" gpf pAqBf pAqq " gpf pAqq.
Thus for every n P N the range of αn is a hereditary subalgebra of A. We prove that ker αn
is complemented and that (9) holds by induction on n. For n " 1 we have already seen it.
Assume that the assertion holds for some n P N. Let θ be the inverse to the isomorphism
αn : pker αnqK Ñ αnpAq. Then clearly θpαnpAq X pker αqKq Ď pker αn`1qK. However, since
αnpAq is hereditary in A we have αnpAq X pker αqK " αnpAqpker αqKαnpAq. Hence αn`1
maps θpαnpAq X pker αqKq onto αn`1pAq. Since αn`1pker αn`1qK Ñ αn`1pAq is isometric it
follows that it is actually an isomorphism and we have θpαnpAq X pker αqKq " pker αn`1qK.
For any element αnpaq in pker αqK " uuA, by the induction hypothesis, we have
θpαnpaqq " unαnpaqun " unpuuqαnpaqun " un`1un`1aunun " un`1un`1a.
Hence pker αn`1qK Ď un`1un`1A, and the argument we used to show that pker αqK " uuA
implies that we actually have pker αn`1qK " un`1un`1A. Thus un`1un`1 P MpAq is the
projection onto pker αn`1qK.
(cid:3)
Integration over the Haar measure on T gives the (faithful) conditional expectation
(10)
γzpaqdµ,
a P C pA, α; Jq,
Epaq "żT
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
13
from C pA, α; Jq onto the fixed point C -algebra B for the gauge action. We refer to the
C -algebra B as the core C -subalgebra of C pA, α; Jq. In view of Proposition 2.11, we
have
B " spantunaun : a P αnpAqAαnpAq, n P Nu.
If the C -dynamical system pA, αq is reversible, then the core of C pA, αq coincides with
A and C pA, αq has a similar structure to that of classical crossed product by an automor-
phism.
Proposition 2.14. Suppose that pA, αq is a reversible C -dynamical system. T The crossed
product C pA, αq is the closure of a dense -algebra consisting of the elements of the form
(11)
a "
uka
´k ` a0 `
akuk,
ak P AαkpAq, k " 0, 1, ..., n.
nÿk"1
nÿk"1
The coefficients ak P AαkpAq in (11) are uniquely determined by a.
Proof. To see that any element (7) can be presented in the form (11) let us consider an
element unan,mum where an,m P αnpAqAαmpAq and put k :" m ´ n. Suppose that k ě 0.
Then αnpAqAαn`kpAq " αnpAqAαnpAqαn`kpAq " αnpAqαn`kpAq " αnpαkpAqq. Thus there
is ak P αkpAq X pker αkqK such that an,m " αnpakq. Hence, by Lemma 2.13, we get
unan,mum " unαnpakqunuk " akuk.
If k ă 0 by passing to adjoints we get unan,mum " u´kak. In view of Proposition 2.11,
this proves the first part of the assertion. For the last part notice that if a is of the form
(11), then for k ě 0 we have Epukaq " a
´k and Epaukq " ak. Hence the coefficients ak
are uniquely determined by a.
(cid:3)
2.3. Gauge-invariant ideals. Let pA, αq be a fixed C -dynamical system and let J be an
ideal in pker αqK. Ideals in C pA, α; Jq that are invariant under the gauge action are called
gauge-invariant. In this subsection we describe these ideals in terms of pairs of ideals in A.
Definition 2.15 (Definitions 3.2 and 3.3 in [33]). We say that an ideal I in A is a positively
invariant ideal in pA, αq if αpIq Ď I. We say that I is J-negatively invariant ideal in pA, αq
if J X α´1pIq Ď I. If I is both positively invariant and J-negatively invariant we say that
I is J-invariant, and if J " pker αqK we drop the prefix 'J-'.
Let I be a positively invariant ideal in pA, αq. It induces two C -dynamical systems: the
restricted C -dynamical system pI, αIq and the quotient C -dynamical system pA{I, αIq
where αIpa ` Iq :" αpaq ` I for all a P A. Note that if pA, αq is extendible then so is the
quotient pA{I, αIq, but pI, αIq in general fails to be extendible. For instance, if A " C0pRq,
αpaqpxq " apx ´ 1q and I " C0p0, 8q then α is extendible but αI is not, see also [1].
Lemma 2.16. Let I be an invariant ideal in pA, αq.
(i) If the kernel of α is a complemented ideal in A then αI and αI have complemented
kernels in A{I and I respectively, and
pker αIqK " qIppker αqKq,
pker αIqK " pker αqK X I.
(ii) If pA, αq is reversible then pA{I, αIq and pI, αIq are reversible.
14
B. K. KWAŚNIEWSKI
Proof. (i). Since pker αqK X α´1pIq Ď I we get ker αI " qI pα´1pIqq " qIppker αqK X α´1pIq `
ker αq " qI pker αq. Thus qIppker αqKq " pker αIqK and qIpker αq is a complemented ideal in
A{I. Since ker αI " ker α X I we see that pker αIqK " pker αqK X I and therefore these
ideals are complementary in I.
(ii). In view of part (i) it suffices to note that both αI and αI have hereditary ranges.
The former is straightforward and the latter follows from the following relations
αpIqIαpIq " αpIAqIαpAIq " αpIqαpAqIαpAqαpIq Ď αpIqαpAqαpIq " αpIq.
(cid:3)
Definition 2.17. Let I, I 1, J be ideals in A where J Ď pker αqK. We say that pI, I 1q is a
J-pair for a C -dynamical system pA, αq if
I is positively invariant, J Ď I 1
and I 1 X α´1pIq " I.
The set of J-pairs for pA, αq is equipped with a natural partial order induced by inclusion:
pI1, I 1
def
ðñ I1 Ď I2 and I 1
1q Ď pI2, I 1
2q
1 Ď I 1
2.
Lemma 2.18. If pπ, Uq is a J-covariant representation then pker π, Ipπ,U qq is a J-pair.
Proof. It is clear that ker π is positively invariant, J Ď Ipπ,U q and ker π Ď Ipπ,U q.
In
particular, ker π Ď Ipπ,U q X α´1pker πq. For the reverse inclusion, note that for any a P
Ipπ,U q X α´1pker πq we have πpaq " U U πpaq " U U πpaqU U " U πpαpaqqU " 0.
(cid:3)
Clearly, if pI, I 1q is a J-pair, then I is J-invariant and I `J Ď I 1. Note that then pI, I `Jq
is also a J-pair, but in general I ` J ‰ I 1, cf. [33, Remark 3.2 and Example 3.1]. We have
the following relationship between gauge-invariant ideals and J-pairs for pA, αq.
Theorem 2.19. Let pA, αq be a C -dynamical system and let J an ideal in pker αqK. The
relations
I " A X I,
(12)
establish an order preserving bijective correspondence between J-pairs pI, I 1q for pA, αq and
gauge-invariant ideals I in C pA, α; Jq. Moreover, for objects satisfying (12) we have a
natural isomorphism
I 1 " ta P A : p1 ´ uuqa P Iu
C pA, α; Jq{I -- C pA{I, αI ; qIpI 1qq
and if I 1 " I ` J (equivalently I is generated by its intersection with A), then I is Morita-
Rieffel equivalent to C pI, αI; I X Jq.
Proof. We use Proposition A.8 to identify C pA, α; Jq with OpJ, Eαq. By Theorem A.4
and Proposition A.10 the relations I " A X I and I 1 " A X pI ` EαE
αq establish a
bijective correspondence between J-pairs pI, I 1q for pA, αq and gauge-invariant ideals I in
C pA, α; Jq. Note that EαE
α " uαpAqAαpAqu and recall that p1 ´ uuq is a multiplier of
C pA, α; Jq by Corollary 2.12. Thus a P I 1 implies that p1 ´ uuqa P I. Conversely, if a P A
is such that p1´uuqa P I then, since p1´uuqa " a´uαpaqu, we have a P I`EαE
α. Hence
I 1 " ta P A : p1 ´ uuqa P Iu. Since EαI -- EI we get C pA, α; Jq{I -- C pA{I, αI ; qIpI 1qq,
see Theorem A.4 and Proposition A.10.
Clearly, the bijective correspondence pI, I 1q ÐÑ I preserves order. Thus I is generated
by I if and only if I 1 " I ` J. In this case we see that I is Morita-Rieffel equivalent to
C pI, αI; I X Jq by Theorem A.4, because EαI -- IEα.
(cid:3)
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
15
Remark 2.20. The pairs pt0u, Jq and pt0u, pker αqKq are always J-pairs. Thus Theo-
rem 2.19 implies that C pA, α; Jq is never simple unless J " pker αqK, that is unless
C pA, α; Jq " C pA, αq. More detailed necessary conditions and certain sufficient con-
ditions for C pA, αq to be simple can be found in [33, Theorem 4.2]. The only simplicity
result we explicitly state in this paper is Proposition 5.9 below.
Corollary 2.21. If the kernel of α is a complemented ideal in A then the relations
(13)
I " A X I,
I is generated by I
establish a bijective correspondence between invariant ideals I for pA, αq and gauge-invariant
ideals I in C pA, αq, under which we have C pA, αq{I -- C pA{I, αIq and I is Morita-
Rieffel equivalent to C pI, αIq.
Proof. Let pI, I 1q be a pker αqK-pair and let I be the corresponding gauge-invariant ideal in
C pA, αq. Using (12) and Lemma 2.13 for any such pair we get
I 1 " ta P A : p1 ´ uuqa P Iu " ta ' b P ker α ' pker αqK : a P Iu
" pker α X Iq ' pker αqK.
Hence I 1 " I ` pker αqK, that is I is generated by I. By Lemma 2.16, ker αI " qIpker αq
and pker αqK X I " pker αIqK. In particular, we get qIpI 1q " qIppker αqKq " pker αIqK. Now
the assertion follows from Theorem 2.19.
(cid:3)
For crossed products of reversible C -dynamical systems we can actually identify gauge-
invariant ideals up to isomorphism.
Proposition 2.22. If pA, αq is a reversible C -dynamical system and I is a gauge-invariant
ideal in C pA, αq, then I -- C pI, αIq where I " A X I.
Proof. In view of Propositions A.8 and A.11 it suffices to apply the general result for Hilbert
bimodules [25, Theorem 10.6.]. Alternatively, the assertion can be proved directly using
Lemma 2.16 and Propositions 2.14 and 2.9.
(cid:3)
2.4. Extensions with complemented kernel. We can use Corollary 2.21 to describe
up to Morita-Rieffel equivalence all gauge-invariant ideals in an arbitrary crossed product
C pA, α; Jq. More specifically, there is a canonical construction of a C -dynamical system
pAJ , αJ q such that C pA, α; Jq -- C pAJ , αJq and the kernel of αJ is complemented. The
system pAJ , αJ q was considered in [37, Subsection 6.1], in the case A is unital, but the
construction works also in our general context.
Definition 2.23. For every C -dynamical system pA, αq and an ideal J in pker αqK we put
and define an endomorphism αJ : AJ Ñ AJ by the formula
AJ :"`A{ ker α '`A{J
AJ Q pa ` ker αq ' pb ` Jq αJ
ÝÑ pαpaq ` ker αq ' pαpaq ` Jq P AJ .
The system pAJ , αJq extends pA, αq, in the sense that the map
A Q a ιJ
ÞÝÑ`a ` ker α '`a ` J P AJ
16
B. K. KWAŚNIEWSKI
is an injective homomorphism that intertwines α and αJ . Moreover, the kernel of αJ
coincides with the direct summand A{J in AJ . Hence pker αJqK corresponds to the A{ ker α
summand. We also note that ιJ : A Ñ AJ is an isomorphism if and only if ker α is a
complemented ideal in A and J " pker αqK.
Lemma 2.24. If pI, I 1q is a J-pair for pA, αq then
(14)
pI, I 1qJ :" qker αpIq ' qJ pI 1q Ÿ AJ
is an invariant ideal in pAJ , αJq such that
(15)
I " ι´1
J ppI, I 1qJ q.
Proof. Let us prove (15) first. Since I Ď I 1 we have I Ď ι´1
J ppI, I 1qJ q,
then a " i ` k for some i P I, k P ker α, and a P I 1. This implies that k P I 1 X ker α Ď
I 1 X α´1pIq " I. Hence a P I and (15) holds.
Now using (15) and the equality αJ pAJ q " ιJ pαpAqq we get
J ppI, I 1qJ q. If a P ι´1
αJ pAJ q X pI, I 1qJ Ď ιJ pIq Ď αJppI, I 1qJ q.
On the other hand, since αpIq Ď I Ď I 1 we have αJppI, I 1qJ q Ď pI, I 1qJ . Therefore
αJ ppI, I 1qJ q " αJpAJ q X pI, I 1q. It follows that αJ : pI, I 1qJ X pker αJ qK Ñ αJ pAJ q X pI, I 1q
is an isomorphism and this implies that pI, I 1qJ is invariant in C pAJ , αJq.
(cid:3)
Proposition 2.25. Let pA, αq be a C -dynamical system and let J be an ideal in pker αqK.
The embedding ιJ extends to a gauge-invariant isomorphism
C pA, α; Jq -- C pAJ , αJq.
If I is a gauge-invariant ideal in C pA, α; Jq corresponding to a J-pair pI, I 1q for pA, αq,
then I is mapped by the above isomorphism onto a gauge-invariant ideal in C pAJ , αJq
which is generated by the ideal pI, I 1qJ given by (14). In particular, I is Morita-Rieffel
equivalent to C ppI, I 1qJ , αJpI,I 1qJ q.
Proof. Let us denote by u and v the universal partial isometries in C pA, α; Jq and C pAJ , αJ q
respectively. It is clear that pιJ , vq is an injective representation of pA, αq in C pAJ , αJ q that
admits a gauge action. Using Lemma 2.13 we get that ta P A : pvvqιJ paq " ιJ paqu " J. By
virtue of Proposition 2.9 we see that ιJ ¸ v : C pA, α; Jq Ñ C pAJ , αJq is a gauge-invariant
isomorphism. Note, again using Lemma 2.13, that for any a, b P A we have
and
p1 ´ vvq`pa ` ker αq ' pb ` Jq " 0 ' pb ` Jq " p1 ´ vvqιJ pbq
pvvq`pa ` ker αq ' pb ` Jq " pa ` ker αq ' 0 " pvvqιJ paq.
Let us now fix a J-pair pI, I 1q in pA, αq and let pI, I 1qJ be the corresponding invariant ideal
in pAJ , αJq given by (14). In view of the above equalities, we have
(16)
Let I J be the ideal in in C pAJ , αJq generated by pI, I 1qJ and let I :" pιJ ¸ vq´1pI J q. Then
pI, I 1qJ " pvvqιJ pIq ` p1 ´ vvqιJ pI 1q.
ta P A : p1 ´ uuqa P Iu " ta P A : p1 ´ vvqιJ paq P I J u
" ta P A : p1 ´ vvqιJ paq P pI, I 1qJ u " I 1.
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
17
This, together with (15), shows that I is the gauge-invariant ideal corresponding to the
J-pair pI, I 1q. Hence I is Morita-Rieffel equivalent to C ppI, I 1qJ , αJ pI,I 1qJ q, by Corollary
2.21.
(cid:3)
2.5. K-theory of gauge-invariant ideals. We have the following generalization of the
classical Pimsner-Voiculescu sequence.
Proposition 2.26. For an ideal J in pker αqK we have the following exact sequence
K0pJq
K0pιq´K0pαJ q
/ K0pAq
K0pιq
/ K0pC pA, α; Jqq
,
K1pC pA, α; Jqq
K1pιq
K1pAq
K1pιq´K1pαJ q
K1pJq
where ι stands for inclusion.
Proof. Using Lemma A.13 we see that in the sequence (48) we may replace the maps
Kipι22q´1 Kipι11 φJ q with KipαJ q, i " 0, 1. This results with the desired sequence. (cid:3)
One can combine results from previous subsections with Proposition 2.26 to get exact
six-term sequences for K-theory of all gauge-invariant ideals and relevant quotients in the
crossed product C pA, α; Jq. We state explicitly only results for gauge-invariant ideals.
Theorem 2.27. Let I be a gauge-invariant ideal in C pA, α; Jq where pA, αq is a C -
dynamical system and J is an ideal in pker αqK. Let pI, I 1q be the J-pair for pA, αq given
by (12). We have
where pI, I 1qJ is given by (14), and in particular if K1ppI, I 1qJ q " 0 then
KpIq -- K`C ppI, I 1qJ , αJ pI,I 1qJ q ,
(17) K0pIq -- coker`K0pιq ´ K0pαJ qker αpIqq ,
K1pIq -- ker`K0pιq ´ K0pαJ qker αpIqq ,
where αJqker αpIq : qker αpIq ' t0u Ñ pI, I 1qJ is the restriction of αJ and ι : qker αpIq Ñ pI, I 1qJ
is the inclusion. If I is generated by I, that is if I 1 " I ` J, then
and if additionally K1pIq " K1pI X Jq " 0, then
KpIq -- K pC pI, αI; I X Jqq ,
K0pIq " cokerpK0pιq ´ K0pαIXJ qq,
K1pIq " kerpK0pιq ´ K0pαIXJqq
where αIXJ : I X J Ñ I is the restriction of α and ι : I X J Ñ I is the inclusion.
Proof. By the last part of Proposition 2.25, I is Morita-Rieffel equivalent to the crossed
product C ppI, I 1qJ , αJ pI,I 1qJ q. Hence the corresponding K-groups are isomorphic by [24,
Proposition B.5], see also [24, Remark B.6]. If K1ppI, I 1qJ q " 0 then also pαJ pI,I 1qJ qK "
qker αpIq ' t0u has K1-group equal to zero. Thus applying Proposition 2.26 to the system
ppI, I 1qJ , αJpI,I 1qJ q and the ideal pker αJ pI,I 1qJ qK we get the second part of the assertion and
(17). In view of the second part of Theorem 2.19, the above argument proves also the first
part of the assertion.
(cid:3)
/
/
O
O
o
o
o
o
18
B. K. KWAŚNIEWSKI
Corollary 2.28. If ker α is a complemented ideal in A then for every gauge-invariant ideal
I in C pA, αq we have
KpIq -- K pC pI, αIqq ,
where I :" I X A.
If additionally K1pIq " 0, then
K0pIq " cokerpK0pιq ´ K0pαIXpker αqKqq,
K1pIq " kerpK1pιq ´ K1pαIXpker αqKqq
where αIXpker αqK : I X pker αqK Ñ I is restriction of α and ι : I X pker αqK Ñ I is the
inclusion.
Proof. It suffices to combine the second parts of Theorem 2.27 and Corollary 2.21.
(cid:3)
2.6. Reversible extensions. We fix a C -dynamical system pA, αq and an ideal J in
pker αqK. We generalize a construction of a reversible C -dynamical system pB, βq asso-
ciated to the triple pA, α; Jq in [33, Subsection 3.1], see also [36, Section 4], to the case
when α is not necessarily extendible. The system pB, βq can be viewed as a direct limit
of approximating C -dynamical systems pBn, βnq, n P N. We denote by q : A Ñ A{J the
quotient map and for each n P N we put
An :" αnpAqAαnpAq.
The C -algebra Bn, n P N, is a direct sum of the form
and the endomorphism βn : Bn Ñ Bn is given by the formula
Bn " qpA0q ' qpA1q ' ... ' qpAn´1q ' An,
βnpa0 ' a1 ' ... ' anq " a1 ' a2 ' ... ' qpanq ' αpanq,
where ak P qpAkq, k " 0, ..., n´ 1, and an P An, n ą 0. Thus we get a sequence pBn, βnq, n P
N, of C -dynamical systems where pB0, β0q " pA, αq. We consider bonding homomorphisms
αn : Bn Ñ Bn`1, n P N, whose action is presented by the diagram
Bn
" qpA0q
' ... ' qpAn´1q
' An
αn
id
id
q
.
α
#●●●●●●●●●
Bn`1 " qpA0q ' ... ' qpAn´1q ' qpAnq ' An`1
In other words, αn is given by the formula
αnpa0 ' ... ' an´1 ' anq " a0 ' ... ' an´1 ' qpanq ' αpanq,
where ak P qpAkq, k " 0, ..., n ´ 1, and an P An. Plainly, the homomorphisms αn are
injective and we have
αn βn " βn`1 αn,
n P N.
Accordingly, we get the direct sequence of C -dynamical systems:
pB0, β0q α0ÝÑ pB1, β1q α1ÝÑ pB2, β2q α2ÝÑ ... .
We denote by pB, βq a direct limit of the above direct sequence. More precisely, B "
limÝÝÑtBn, αnu is the C -algebraic direct limit, and β is determined by the formula βpφnpaqq "
#
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
19
φnpβnpaqq where φn : Bn Ñ B is the natural (injective) homomorphism, a P Bn and n P N.
That is we have
(18)
βpφnpa0 ' a1 ' ... ' anqq " φn´1pa1 ' a2 ' ... ' anq.
We now extend the main parts of [33, Theorem 3.1 and Proposition 4.7], see also [33,
Remark 3.3].
Theorem 2.29. The C -dynamical system pB, βq described above is reversible and we may
assume a natural identification
under which we have
C pA, α; Jq " C pB, βq
B " spantukauk : a P αkpAqAαkpAq, k P Nu
and
βpbq " ubu, b P B.
k"0 U kπpakqU k, ak P αkpAqAαkpAq, es-
tablishes a one-to-one correspondence between J-covariant representations pπ, Uq of pA, αq
k"0 ukakukq " řn
In particular, the relation rπpřn
and covariant representations prπ, Uq of pB, βq.
Proof. Let us prove first that pB, βq is reversible. To this end, take a " a0 'a1'...'an P Bn
and b " b0 ' b1 ' ... ' bn P Bn´1 for n ą 1. Then, in view of (18), we get
βpφnpaqqφn´1pbqβpφnpaqq " βpφnp0 ' a1b0a1 ' ... ' anbn´1anqq.
This implies that βpBqBβpBq " βpBq. Hence βpBq is a hereditary C -subalgebra in B.
The ideal ker β is complemented in B as Bn X ker β " tφnpa0 ' 0 ' ... ' 0q : a0 P qpA0qu is
complemented in Bn for every n ą 0. Thus pB, βq is reversible.
k"0 ukakuk : ak P αkpAqAαkpAq, k " 0, ..., nu Ď
Now, for each n P N, we define Cn :" třn
C pA, α; Jq. We also put C :" ŤnPN Cn. It follows from (8) that Cn, n P N, and C are
of C. Exactly as in the proof of [36, Statement 1], one checks that if a "řn
bk P αkpAqAαkpAq, k " 0, ..., n, then putting ak "řk
C -algebras. Recall, see Proposition 2.11, that tukukukPN is a decreasing sequence of or-
thogonal projections that commute with elements of A. Hence they commute with elements
k"0 ukbkuk P Cn,
i"0 αk´ipbiq we get that
(19)
a "
ukp1 ´ uuqakuk ` unanun,
n´1ÿk"0
n´1àk"0
where 1 is the unit in MpC pA, α; Jqq. Hence (19) is a general form of an element in Cn.
In particular, since ukp1 ´ uuquk " pukuk ´ uk`1uk`1q, k " 0, ..., n ´ 1, and unun are
mutually orthogonal projections commuting with elements of Cn, we see that Cn admits
the following direct sum decomposition
Cn "
pukuk ´ uk`1uk`1qCn ' ununCn
Since uk is a partial isometry it follows that
αkpAqCαkpAq " ukAukCukAuk Q a Ñ ukauk P C,
k " 1, ..., n,
is a -homomorphic isometry. Since J " uuA X A we also see, cf. for instance [21, Lemma
10.1.6], that
p1 ´ uuqA Q a Ñ qpaq P qpAq
20
B. K. KWAŚNIEWSKI
is an isomorphism of C -algebras. Combining these facts we get that the formula
Φn pqpa0q ' qpa1q ' ... ' qpan´1q ' anq "
ukp1 ´ uuqakuk ` unanun,
n´1ÿk"0
defines an isomorphism Φn : Bn Ñ Cn. If a P Cn is given by (19), then using equality
un`1αpanqun`1 " unpuuqanun we get
a "
ukp1 ´ uuqakuk ` un`1αpanqun`1.
nÿk"0
Therefore Φn`1 αn " Φn, n P N. Hence the isomorphisms Φn induce the isomorphism
Φ : B Ñ C between the inductive limit C -algebras B and C.
We claim that Φpβpbqq " ubu, for b P C. Indeed, let a P Cn is given by (19). Notice that
for k ą 0 we have uuk " puuquk´1 " uk´1uk´1puuquk´1 " uk´1ukuk. Therefore,
since ukukak " ak, we get
u`ukp1 ´ uuqakuk u " uk´1p1 ´ uuqakuk´1.
Clearly, up1 ´ uuqa0u " ua0u ´ ua0u " 0. Accordingly, Φpβpφnpaqqq " uΦpφnpaqqu,
which proves our claim.
It readily follows from the definition of Φn that
uuΦnpBnq " tφnp0 ' a1 ' ... ' anq : 0 ' a1 ' ... ' an P Bnu " Bn X pker βqK.
This implies that pker βqK " tb P B : uuΦpbq " Φpbqu.
Concluding the pair pΦ, uq is an injective covariant representation of pB, βq in C pA, α; Jq
that admits gauge-action. Thus, by Proposition 2.9, Φ ¸ u : C pB, βq Ñ C pA, α; Jq is an
isomorphism which we may use to assumed the described identification. The last part of
the assertion follows from the universal properties of the crossed products.
(cid:3)
Definition 2.30 (Definition 3.1 in [33]). Suppose that pA, αq is a C -dynamical system
and J is an ideal in pker αqK. We call the C -dynamical system pB, βq constructed above
the natural reversible J-extension of pA, αq.
Let pB, βq be a natural reversible J-extension of pA, αq and suppose that A " C0pXq
is commutative. Then, in view of our construction, B is also commutative and thus we
may identify it with C0prXq where rX is a locally compact Hausdorff space. With this
identification, β is given by the formula
βpbqprxq "#bprϕprxqq, rx P r∆,
rx R r∆,
where rϕ : r∆ Ñ rϕpr∆q is a homeomorphism, r∆ Ď rX is open and rϕpr∆q Ď rX is clopen. The
pair prX,rϕq is uniquely determined by pX, ϕq and the closed set
of prX,rϕq.
which necessarily contains Xzϕp∆q. Similarly as in [33, Proposition 4.7], cf. also [30, The-
orem 3.5], using the above construction of pB, βq one can deduce the following description
Y " tx P X : apxq " 0 for all a P Ju,
(20)
0
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
21
Proposition 2.31. Up to conjugacy with a homeomorphism, the above partial dynamical
system prX,rϕq can be described as follows:
8ďN "0
rX "
where
XN Y X8
XN " tpx0, x1, ..., xN , 0, ...q : xn P ∆, ϕpxnq " xn´1, n " 1, ..., N, xN P Y u,
X8 " tpx0, x1, ...q : xn P ∆, ϕpxnq " xn´1, n ě 1u.
The topology on rX is the product one inherited fromśnPNpX Y t0uq where t0u is a clopen
singleton and Y is given by (20). The homeomorphism rϕ : r∆ Ñ rϕpr∆q is given by the
formula
rϕpx0, x1, ...q " pϕpx0q, x0, x1, ...q,
r∆ " tpx0, x1, ...q P rX : x0 P ∆u.
Proof. We omit the proof as the assertion will follow from a much more general result we
prove below, see Theorem 4.9.
(cid:3)
Definition 2.32 (cf. Definition 3.5 [30]). Let Y be a closed subset of X that contains
2.31 the natural reversible Y -extension of pX, ϕq.
Xzϕp∆q. We call the dynamical system prX,rϕq described in the assertion of Proposition
Note that for the natural reversible Y -extension prX,rϕq of pX, ϕq, the map Φ : rX Ñ X
given by Φprxq " x0 is surjective and intertwines rϕ and ϕ. This justifies the name.
2.7. Topological freeness and freeness. We turn to a discussion of certain conditions
implying uniqueness property and gauge-invariance of all ideals in the crossed products.
For reversible and extendible systems the relevant statements in [33, Subsection 4.5] were
deduced from [34, Theorem 2.20]. We will extend them by applying general results from
[32] and facts presented in Appendix A.
Definition 2.33. Let ϕ be a partial homeomorphism of a topological (not necessarily
Hausdorff) space X with domain being an open set ∆ Ď X. We say that ϕ is topologically
free if the set of its periodic points of any given period n ą 0 has empty interior. A set
V Ď X is invariant if ϕpV X ∆q " V X ϕp∆q. We say that ϕ is (essentially) free, if it is
topologically free when restricted to any closed invariant set.
Definition 2.34. Let pA, αq be a reversible C -dynamical system. Since pker αqK is an
ideal in A and αpAq " αpAqAαpAq is a hereditary subalgebra of A we have the natural
identifications:
{pker αqK " tπ P pA : πppker αqKq ‰ 0u, zαpAq " tπ P pA : πpαpAqq ‰ 0u.
Thus we treat zαpAq and {pker αqK as open subsets of pA. With these identifications the
homeomorphism pα : zαpAq Ñ {pker αqK dual to the isomorphism α : pker αqK Ñ αpAq
becomes a partial homeomorphism of the spectrum of pA, cf. [33]. We refer to pα as to the
Proposition 2.35. Let pA, αq be a reversible C -dynamical system.
partial homeomorphism dual to pA, αq.
22
B. K. KWAŚNIEWSKI
pA, αq give rise to a faithful representation of C pA, αq.
(i) If pα is topologically free, then every injective covariant representation pπ, Uq of
(ii) If pα is free, then all ideals in C pA, αq are gauge-invariant; hence they are in one-
to-one correspondence with invariant ideals in pA, αq, cf. Corollary 2.21.
Proof. By Proposition A.8 and Lemma A.12, Theorem A.6 translates to the desired asser-
tion.
(cid:3)
One can apply the above proposition to an arbitrary crossed product C pA, α; Jq using
the identification C pA, α; Jq " C pB, βq from Theorem 2.29, and the following lemma.
Lemma 2.36. Let pA, αq be a C -dynamical system, J an ideal in pker αqK, and pB, βq the
natural reversible J-extension of pA, αq.
(i) For any injective J-covariant representation pπ, Uq of pA, αq the corresponding co-
variant representation prπ, Uq of pB, βq is injective if and only if J " ta P A :
(ii) Relations (12) establish a bijective correspondence between with J-pairs pI, I 1q in
U U πpaq " πpaqu.
pA, αq and invaraint ideals in pB, βq.
Proof. (ii). Since C pA, α; Jq " C pB, βq we get the assertion by applying Theorem 2.19
to C pA, α; Jq and Corollary 2.21 to C pB, βq.
(i). It follows from item (ii) and Lemma 2.18.
(cid:3)
In practice, in order to use Proposition 2.35 and Lemma 2.36, one has to determine
if A is commutative.
topological freeness and freeness ofpβ in terms of pA, αq and J. This can be readily achieved
Definition 2.37 (Definition 4.8 in [33]). Let ϕ be a partial mapping of a locally compact
Hausdorff space X defined on an open set ∆ Ď X. We say that a periodic orbit O "
tx, ϕpxq, ..., ϕn´1pxqu of a periodic point x " ϕnpxq has an entrance y P ∆ if y R O and
ϕpyq P O. We say ϕ is topologically free outside a set Y Ď X if the set of periodic points
whose orbits do not intersect Y and have no entrances have empty interior.
Y is a closed set containing Xzϕp∆q, see Definition 2.32. Then
Lemma 2.38. Let prX,rϕq be the Y -extension of a partial dynamical system pX, ϕq where
(i) rϕ is topologically free if and only if ϕ is topologically free outside Y ,
(ii) rϕ is free if and only if ϕ is free (has no periodic points).
Proof. Item (i) can be proved exactly as [33, Lemma 4.2]. Item (ii) is straightforward. (cid:3)
One of the aims of the present paper is to obtain effective conditions implying the prop-
erties of crossed products described in Proposition 2.35 for a class of C -dynamical systems
on C0pXq-algebras. This is achieved in Theorems 4.11 and 4.12 below.
2.8. Pure infinite crossed products for reversible C -dynamical systems. In this
subsection, we fix a reversible C -dynamical system pA, αq. The property that we are
about to introduce appears (without a name) in a number of proofs of pure infiniteness for
crossed products. As we explain in more detail below, in the context of crossed products,
this property is formally weaker than spectral freeness [46], topological freenees, proper
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
23
outerness [10] and aperiodicity [38], but the general relationship between these notions is
not completely clear.
Definition 2.39. Let A be a C -subalgebra of a C -algebra B. We say that A` supports
elements of B` if for every if for every b P B`zt0u there exists a P A` such that a À b.
We say that that A` residually supports elements of B` if for every ideal I of B, qIpAq`
supports elements of qIpBq.
Remark 2.40. If A` is a filling family for B in the sense of [28, Definition 4.2] then A`
residually supports elements of B` (it is not clear whether the converse implication holds).
Thus, if A is commutative or seperable and α : A Ñ A is a residually properly outer
authomorphism, then [29, Theorem 3.8] implies that A` residually supports elements of
C pA, αq`. In [46, Proposition 3.9] it is shown that A` supports elements of B` if and
only if for every b P B`zt0u there is z P B such that zaz is a non-zero element of
In particular, [46, Lemma 3.2] implies that if α : A Ñ A is an automorphism and
A.
the corresponding Z-action is spectrally free in the sense of [46, Definition 1.3], then A`
residually supports elements of C pA, αq`.
In connection with Remark 2.40 we show that the notion of residual aperiodicity intro-
duced in [38, Definition 8.19], for (a semigroup version of) extendible reversible systems,
implies that A` residually supports elements of C pA, αq`.
Definition 2.41. We say that an extendible reversible C -dynamical system pA, αq is
aperiodic if for each n ą 0, each a P A and every hereditary subalgebra D of A
inft}daαnpdq} : d P D`, }d} " 1u " 0.
We say that pA, αq is residually aperiodic if the quotient system pA{I, αIq is aperiodic for
every invariant ideal I in pA, αq.
Proposition 2.42. If pA, αq is residually aperiodic then A` residually supports elements
of C pA, αq`.
Proof. By [38, Lemmas 8.18], [38, Corollary 4.7] every ideal in C pA, αq is generated by
it intersection with A. Let I be an ideal in C pA, αq. By Corollary 2.21, we have the
isomorphism C pA, αq{I -- C pA{I, αIq where I :" AXI is an invariant ideal in pA, αq. The
system pA{I, αIq is reversible by Lemma 2.16 ii). Fix a positive element b in C pA, αq{I.
We may assume that }b} " 1. Applying to pA{I, αIq [38, Lemmas 4.2 and 8.18], we may
find a positive contraction h P A{I such that
}hEpbqh ´ hbh} ď 1{4,
(21)
where E is the conditional expectation from C pA{I, αIq onto A{I. Putting a :" phEpbqh ´
1{2q` P A{I we have that a ‰ 0 because }hEpbqh} ą 1{2. Moreover, by [27, Proposition
2.2], relations }hEpbqh} ą 1{2 and }hEpbqh ´ hbh} ď 1{4 imply that a À hbh relative to
C pA{I, αIq -- C pA, αq{I.
(cid:3)
}hEpbqh} ě }Epbq} ´ 1{4 " 3{4
Before we prove the main result of this subsection we need two lemmas.
Lemma 2.43. If A` residually supports elements of C pA, αq` then every ideal in C pA, αq
is gauge-invariant.
24
B. K. KWAŚNIEWSKI
Proof. Let I be an ideal in C pA, αq and let xIy be the smallest ideal in C pA, αq containing
I :" I X A. By Corollary 2.21, we may identify C pA, αq{xIy with C pA{I, αIq. We have
a natural epimorphism Φ : C pA, αq{xIy Ñ C pA, αq{I which is injective on A{I. For any
non-zero positive element b in C pA, αq{xIy there is a non-zero positive element a in A{I
such that a À b. Since 0 ‰ Φpaq À Φpbq, we conlude that Φpbq ‰ 0. Thus ker Φ " t0u and
therefore I " xIy is gauge-invariant.
(cid:3)
Lemma 2.44. Let A Ď B be C -algebras and let A be of real rank zero. The following
conditions are equivalent
(i) Every non-zero positive element in A is properly infinite in B.
(ii) Every non-zero projection in A is properly infinite in B.
Proof. Implication (i)ñ(ii) is trivial. Assume that (ii) holds and let a P A be a non-zero
positive element. By [8, Theorem 2.6] there is an approximate unit tpλ : λ P Λu in aAa
consisting of projections. Thus, by [27, Proposition 2.7(i)], pλ À a for all λ, in A and all
the more in B. Applying [27, Lemma 3.17(ii)] we see that tpλ : λ P Λu Ď Jpaq :" tx P
B : a ' x À au. Thus Btpλ : λ P ΛuB Ď Jpaq because Jpaq is an ideal, see [27, Lemma
3.12(i)]. On the other hand Jpaq Ď BaB by [27, Lemma 3.12(iii)] and since we clearly have
BaB Ď Btpλ : λ P ΛuB it follows that Jpaq " BaB. Hence [27, Lemma 3.12(iv)] tells us
that a is properly infinite in B.
(cid:3)
Remark 2.45. The equivalence of (i) and (ii) in Lemma 2.44 answers the question posed
in the proof of [16, Theorem 4.4]: it shows that [16, Theorem 4.4] can be deduced from [16,
Theorem 4.2].
Proposition 2.46 (pure infiniteness criterion). Let pA, αq be a reversible C -dynamical
system such that A` residually supports elements of C pA, αq`. Suppose also that either
A has the ideal property or that A is separable and there finitely many invariant ideals in
pA, αq. The following statements are equivalent:
(i) Every non-zero positive element in A is properly infinite in C pA, αq.
(ii) C pA, αq is purely infinite.
(iii) C pA, αq is purely infinite and has the ideal property.
(iv) Every non-zero hereditary C -subalgebra in any quotient C pA, αq contains an infi-
nite projection.
If A is of real rank zero, then each of the above conditions is equivalent to
(i') Every non-zero projection in A is properly infinite in C pA, αq.
In particular, if A is purely infinite then C pA, αq is purely infinite and has the ideal
property.
Proof. Implications (iv)ô(iii)ñ(ii)ñ(i) are general facts, see respectively [47, Propositions
2.11], [27, Proposition 4.7] and [27, Theorem 4.16]. If A is if real rank zero the equivalence
(i)ô(i') is ensured by Lemma 2.44. Thus it suffices to show that (i) implies (iii) or (iv).
Let us then assume that every element in A`zt0u is properly infinite in C pA, αq.
Suppose first that A has the ideal property. We will show (iv). Let I be an ideal in
C pA, αq and let B be a non-zero hereditary C -subalgebra in the quotient C pA, αq{I.
Fix a non-zero positive element b in B. Since A` residually supports elements of C pA, αq`
is non-zero positive element a in qIpAq such that a À b. Note that a is properly infinite
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
25
in C pA, αq{I by [27, Proposition 3.14]. Since A has the ideal property we can find a
projection q P A that belongs to the ideal in A generated by the preimage of a in A but
not to I :" A X I. Then q ` I belongs to the ideal in C pA, αq{I generated by a, whence
q ` I À a À b, by [27, Proposition 3.5(ii)]. From the comment after [27, Proposition 2.6] we
can find z P C pA, αq{I such that q ` I " zbz. With v :" b
2 z it follows that vv " q ` I,
whence p :" vv " b
2 is a projection in B, which is equivalent to q ` I. By our
assumption q ` I and hence also p is properly infinite.
2 zzb
1
1
1
Suppose now that A is separable and there are finitely many, say n, invariant ideals in
pA, αq. By Lemma 2.43 and Corollary 2.21 they are in one-to-one correspondence with ideals
in C pA, αq. Hence by [27, Proposition 2.11], the conditions (ii) and (iii) are equivalent.
We will prove (ii). The proof goes by induction on n.
Assume first that n " 2 so that C pA, αq is simple. For any b P C pA, αq`zt0u take
a P A`zt0u such that a À b. Then b P C pA, αqaC pA, αq " C pA, αq and as a is properly
infinite we get b À a by [27, Proposition 3.5]. Hence b is properly infinite as it is Cuntz
equivalent to a. Thus C pA, αq purely infinite.
Now suppose that our claim holds for any k ă n. Let I be any non-trivial ideal in
C pA, αq and put I " I X A. By Lemma 2.16 ii), the systems pA{I, αIq and pI, αIq are
reversible, and by Corollary 2.21 and Proposition 2.22 we have C pA{I, αIq -- C pA, αq{I
and C pI, αIq -- I. Clearly, the system pA{I, αIq satisfies the assumptions of the assertion
(a non-zero image of properly infinite element is properly infinite by [27, Proposition 3.14])
and there are less than n invariant ideals in pA{I, αIq. Hence C pA{I, αIq is purely infinite.
Similar, argument works for C pI, αIq; in particular note that if a À b for b P I `zt0u
and a P A`zt0u, then a P I. Also if a P I `zt0u is properly infinite in C pA, αq, then it is
[27, Proposition 3.3]. Concluding, both I and C pA, αq{I are
properly infinite in I, cf.
purely infinite, and since pure infiniteness is closed under extensions [27, Theorem 4.19] we
get that C pA, αq is purely infinite.
(cid:3)
Remark 2.47. We recall, see [47, Propositions 2.11, 2.14], that in the presence of the ideal
property pure infiniteness of a C -algebra is equivalent to strong pure infiniteness, weak
pure infiniteness, and many other notions of infiniteness appearing in the literature. Thus
the list of equivalent conditions in Proposition 2.46 can be considerably extended.
Remark 2.48. In the case when there are finitely many invariant ideals in pA, αq, we
used separability of A in the proof Proposition 2.46 only to get the equivalence (ii)ô(iii).
Accordingly, in this case, the conditions (i) and (ii) are equivalent even for non-separable
C -algebras.
Remark 2.49. Certain properties that imply condition (i) in Proposition 2.46 were intro-
duced in [38]. In particular, Proposition 2.46 can be readily used to obtain a generalization
of [38, Theorem 8.22] so that it covers not necessarily extendible systems on C -algebras
not necessarily possessing the ideal property.
3. Category of C0pXq-algebras and C0pXq-dynamical systems
In this section, we introduce morphisms of upper semicontinuous C -bundles which in-
duce certain homomorphisms of C0pXq-algebras. We give several characterizations of such
homomorphisms, and study basic properties of C -dynamical systems pA, αq where A is
26
B. K. KWAŚNIEWSKI
a C0pXq-algebra and α is induced by a morphism. We show that the arising category of
C0pXq-algebras has direct limits, and in some cases such limits exist in the subcategory of
continuous C0pXq-algebras.
Bpyq be upper semicontinuous C -bundles. We wish to view morphism between
3.1. Morphism of C -bundles and C0pXq-dynamical systems. Let A " ŮxPX
B " ŮyPY
C -bundles as a common generalization of proper mappings and C -homomorphisms. Mim-
icking the definition of morphisms of vector bundles, one can imagine such a morphism as
a pair of continuous mappings α : B Ñ A and ϕ : X Ñ Y such that the following diagram
Apxq and
B
p
α
/ A
p
Y
Xϕo
commutes and for each x P X, α : Bpϕpxqq Ñ Apxq is a homomorphism. Since some of
these homomorphisms might be zero we will allow ϕ to be defined on an open subset ∆ of
X.
Definition 3.1. A morphism (of upper semicontinuous C -bundles) from B to A is a pair
pϕ, tαxuxP∆q consisting of
1) a continuous proper map ϕ : ∆ Ñ Y defined on an open set ∆ Ď X, and
2) a continuous bundle of homomorphisms tαxuxP∆ between the corresponding fibers,
i.e.:
a) for each x P ∆, αx : Bpϕpxqq Ñ Apxq is a homomorphism;
b) if txiuiPΛ Ď ∆ and tbiuiPΛ Ď B are nets such that xi Ñ x P ∆, bi Ñ b and
ppbiq " ϕpxiq, for i P Λ, then αxipbiq Ñ αxpbq.
The above definition is born to work well with section algebras.
Proposition 3.2. Let ϕ : ∆ Ñ Y be a proper continuous mapping where ∆ Ď X is an open
set. For each x P ∆ let αx : Bpϕpxqq Ñ Apxq be a homomorphism. The pair pϕ, tαxuxP∆q
is a morphism from B to A if and only if the formula
(22)
b P Γ0pBq, x P X,
αpbqpxq "#αxpbpϕpxqq, x P ∆,
x R ∆,
0x
yields a well defined homomorphism α : Γ0pBq Ñ Γ0pAq between the section C -algebras.
Proof. Suppose that pϕ, tαxuxP∆q is a morphism. Clearly, it suffices to show that the map
(22) is well defined, equivalently, that for any b P Γ0pBq the mapping
(23)
X Q x ÞÝÑ αpbqpxq P Apxq Ď A
is in Γ0pAq. Condition 2b) from Definition 3.1 readily implies that the map ∆ Q x ÞÝÑ
αpbqpxq P Apxq Ď A is continuous (consider elements bi :" bpϕpxiqq). In particular, ∆ Q
x ÞÝÑ }αpbqpxq} P R is upper semicontinuous. Thus for any ε ą 0 the set tx P X :
}αpbqpxq} ě εu " tx P ∆ : }αpbqpxq} ě εu is closed. Actually it is compact because
tx P X : }αpbqpxq} ě εu Ď tx P ∆ : }bpϕpxqq} ě εu
/
o
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
27
and the latter set is compact as ϕ is proper and b vanishes at infinity. Thus the map
(23) is vanishing at infinity. To conclude that αpbq P Γ0pAq we need to show that αpbq is
continuous on the boundary B∆ of ∆. But if txiuiPΛ Ď X is a net convergent to x0 P B∆,
then for every ε ą 0 the point x0 belongs to the open set tx P X : }αpbqpxq} ă εu and hence
αpbqpxiq converges to 0 by Lemma 1.1 (consider bi " αpbqpxiq and a " 0).
Conversely, assume that α : Γ0pBq Ñ Γ0pAq is a homomorphism satisfying (22). We
need to show condition 2b) in Definition 3.1. Let txiuiPΛ Ď ∆ and tbiuiPΛ Ď B be nets such
that xi Ñ x P ∆, bi Ñ b and ppbiq " ϕpxiq. Take arbitrary ε ą 0. By Lemma 1.1 there
is a P Γ0pAq such that }apppbqq ´ b} ă ε and we eventually have }apϕpxiqq ´ bi} ă ε. This
implies that }αpaqpxq ´ αxpbq} ă ε and we eventually have }αpaqpxiq ´ αxipbiq} ă ε. Since
αpaq P Γ0pAq we have αxipbiq Ñ αxpbq by Lemma 1.1.
(cid:3)
Definition 3.3. To indicate that a homomorphism α : Γ0pBq Ñ Γ0pAq is given by (22) for
a certain morphism pϕ, tαxuxP∆q of upper semicontinuous C -bundles we will say that α is
induced by a morphism.
Let A " Γ0pAq and B " Γ0pBq. Note that for an induced homomorphism α : B Ñ A
the underlying mapping ϕ : ∆ Ñ Y is uniquely determined by α on the set
(24)
∆0 :" tx P X : αx ‰ 0u Ď ∆,
which coincides with ∆ when all endomorphisms αx, x P ∆, are non-zero. Sometimes we
can assume that ∆ " ∆0 using the following lemma.
Lemma 3.4. Let α : B Ñ A be a homomorphism induced by a morphism pϕ, tαxuxP∆q
from a continuous C0pY q-algebra B to a continuous C0pXq-algebra A and let ∆0 be given
by (24). Suppose also that Bpϕpxqq ‰ t0u, for x P ∆, and that every αx, x P ∆0, is
injective. Then ∆0 is a clopen in ∆ and in particular pϕ∆0, tαxuxP∆0q is a morphism that
induces α.
Proof. Since ∆0 " ŤbPBtx P X : }αpbqpxq} ą 0u and A is a continuous C0pXq-algebra,
we see that set ∆0 is open. Suppose that x0 is a point in the boundary of ∆0 in ∆.
Take a net txiui Ď ∆0 converging to x0 and an element b P B such that }bpϕpx0qq} " 1.
Since the homomorphism αxi are isometric, and the mappings ∆ Q x ÞÑ }αxpbpϕpxqqq} and
∆ Q x ÞÑ }bpϕpxqq} are continuous, we get
}αx0pbpϕpx0qqq} " lim
i
}αxipbpϕpxiqqq} " lim
i
}bpϕpxiqq} " 1.
Hence αx0 ‰ 0, that is x0 P ∆0. Thus ∆0 is closed in ∆ and therefore ϕ∆0 : ∆0 Ñ X is a
proper map. Clearly, α satisfies (22) with ∆0 in place of ∆. Accordingly, α is induced by
the morphism pϕ∆0, tαxuxP∆0q by Proposition 3.2.
(cid:3)
We have the following characterizations of homomorphism induced by morphisms phrased
in terms of C0pXq-algebras.
Proposition 3.5. Let A be C0pXq-algebra and B a C0pY q-algebra. For any homomorphism
α : B Ñ A the following conditions are equivalent:
(i) α is induced by a morphism from B " ŮyPY
Bpyq to A " ŮxPX
Apxq,
28
B. K. KWAŚNIEWSKI
(ii) there is a homomorphism Φ : C0pY q Ñ C0pXq such that
αpf bq " Φpf q αpbq,
f P C0pY q, b P B.
If additionally B is unital and A is a continuous C0pXq-algebra then the above conditions
are equivalent to the following one:
(iii) α maps C0pY q 'almost into' C0pXq, that is
αpC0pY q 1q Ď C0pXq αp1q.
If the additional assumptions and condition (iii) are satisfied, then the corresponding mor-
phism pϕ, tαxuxP∆q can be chosen so that ∆ is compact and each αx, x P ∆, is non-zero.
Proof. (i) ñ (ii). It suffices to put Φpaq :" a ϕ for a P C0pY q.
(ii) ñ (i). Note that Φ : C0pY q Ñ C0pXq is given by the formula
(25)
Φpbqpxq "#bpϕpxqq, x P ∆,
x R ∆,
0
b P C0pY q
where ϕ : ∆ Ñ Y is a continuous proper mapping defined on an open set ∆ Ď X. Let
x P ∆. We define a homomorphism αx : Bpϕpxqq Ñ Apxq as follows. For any b0 P Bpϕpxqq
there is b P B such that bpϕpxqq " b0, and we claim that the element
(26)
αxpb0q :" αpbqpxq
is well defined (does not depend on the choice of b).
Indeed, letrb, b P B be such that
rbpϕpxqq " bpϕpxqq " b0. Then bpϕpxqq ´rbpϕpxqq " 0. Upper semicontinuity of the C -
bundle B " ŮyPY
Bpyq imply that for every ε ą 0 there is an open neighbourhood U of ϕpxq
such that
}bpyq ´rbpyq} ă ε,
for all y P U.
Let us choose a function h P C0pY q such that hpϕpxqq " 1, 0 ď h ď 1 and hpyq " 0 outside
U. We get
}αpbqpxq ´ αprbqpxq} " }`Φphqαpbq ´ Φphqαprbqpxq} " }αphb ´ hrbqpxq}
ď }αphb ´ hrbq} ď }hb ´ hrb} ď ε.
This proves our claim. Now it is straightforward to see that (26) gives the desired homo-
morphism αx : Bpϕpxqq Ñ Apxq. Moreover, for the above defined pair pϕ, tαxuxP∆q the
formula (22) holds. Hence in view of Proposition 3.2, α is induced by a morphism.
Let us now assume that B is a unital and A is a continuous C0pXq-algebra.
(ii) ñ (iii). It is obvious.
(iii)ñ (ii). Since αpC0pY q 1q Ď C0pXq αp1q, for every f P C0pY q there exists g P C0pXq
such that
αpf 1qpxq " gpxqαp1qpxq,
x P X.
Clearly, the function g is uniquely determined by f on the set ∆ :" tx P X : αpBqpxq ‰
0u " tx P X : αp1qpxq ‰ 0u. Since the mapping X Q x Ñ }αp1qpxq} P t0, 1u is continuous
and vanishing at infinity, ∆ is open and compact. Now it is straightforward to see that
the formula Φpf q " g∆ defines a homomorphism Φ : C0pY q Ñ Cp∆q Ď C0pXq satisfying
condition (ii).
(cid:3)
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
29
Example 3.6. Suppose that qI : A Ñ A{I is a quotient map and A is a C0pXq-algebra.
We may treat A{I as a C0pV q-algebra for any closed set V containing σApPrimpA{Iqq, cf.
Lemma 1.7. Then we have
qI pf aq " f V qIpaq,
f P C0pXq, a P A.
Hence condition (ii) in Proposition 3.5 is satisfied.
morphism pid, tqI,xuxPV q where qI,x : Apxq Ñ Apxq{Ipxq, x P V , are the quotient maps.
In particular, qI is induced by the
Let us consider a category of C0pXq-algebras with morphisms being homomorphisms
satisfying the equivalent conditions in Proposition 3.5. In this paper, we are interested in
properties of systems pA, αq where A is an object and α is a morphism in this category.
Definition 3.7. We say that a C -dynamical system pA, αq is a C0pXq-dynamical system,
if A is a C0pXq-algebra and α is induced by a morphism. If additionally A is a continuous
C0pXq-algebra, we say that pA, αq is a continuous C0pXq-dynamical system
In section 5, we will study crossed products associated to continuous C0pXq-dynamical
systems introduced in the following example.
Example 3.8 (Endomorphisms of C -algebras with Hausdorff primitive ideal space). If A
is a C -algebra and its primitive ideal space X :" PrimpAq is Hausdorff, then using Dauns-
Hofmann isomorphism we may naturally treat A as a continuous C0pXq-algebra where the
[6, 2.2.2]. In particular, for x P X " PrimpAq the fiber
structure map σA is identity, cf.
Apxq " A{x is a simple (non-zero) C -algebra. Thus if α : A Ñ A is an endomorphism
induced by a morphism pϕ, tαxuxP∆q, then Lemma 3.4 applies and we may assume that
each αx, x P ∆, is injective. Moreover, if A is unital then we may identify CpXq with ZpAq
and by Proposition 3.5 an endomorphism α : A Ñ A is induced by a morphism if and only
if αpZpAqq Ď ZpAqαp1q.
For trivial C -bundles we have the following description of endomorphisms induced by
morphisms. We equip the set EndpDq of all endomorphisms of a C -algebra D with the
topology of point-wise convergence.
Proposition 3.9. Let ϕ : ∆ Ñ X be a proper continuous mapping defined on an open set
∆ Ď X and let a continuous mapping ∆ Q x ÝÑ αx P EndpDq where D is a C -algebra.
We treat A :" C0pX, Dq as a C0pXq-algebra in an obvious way. The formula
(27)
αpaqpxq :"#αxpapϕpxqq, x P ∆,
x R ∆,
0
a P C0pX, Dq, x P X.
defines an endomorphism of A induced by a morphism. Every endomorphism of A induced
by a morphism arises in this way. If D is simple, or if A is unital, then we can choose the
corresponding morphism in such a way that each αx, x P ∆, is non-zero.
Proof. The corresponding C -bundle A " ŮxPX
together with its product topology. In this case condition 2b) from Definition 3.1 translates
to the following: If txiuiPΛ Ď ∆ and tbiuiPΛ Ď D are nets such that xi Ñ x P ∆ and
D can be identified with the product X D,
30
B. K. KWAŚNIEWSKI
bi Ñ b P D, then αxipbiq Ñ αxpbq. The latter condition is equivalent to the continuity of
the map ∆ Q x ÝÑ αx P EndpDq, which can be readily deduced from the inequality:
}αxipbiq ´ αxpbq} ď }bi ´ b} ` }αxipbq ´ αxpbq}.
Thus the assertion follows by Proposition 3.2. The last remark follows by Lemma 3.4 and
the last part of Proposition 3.2.
(cid:3)
3.2. Quotients and restrictions of C0pXq-dynamical systems. Restrictions and quo-
tients of C0pXq-dynamical systems can be treated as C0pXq-dynamical systems in the
following sense.
Proposition 3.10. Suppose that α : A Ñ A is an endomorphism induced by a morphism
pϕ, tαxuxP∆q. Let I be a positively invariant ideal in pA, αq. Then pI, αIq and pA{I, αIq
are naturally C0pXq-dynamical systems where αI is induced by pϕ, tαxIpϕpxqquxP∆q and αI
is induced by pϕ, tαI,xuxP∆q where
(28)
a P Apϕpxqq, x P ∆.
αI,x`a ` Ipϕpxqq :" αxpaq ` Ipxq,
Proof. Note that positive invariance of I implies that αxpIpϕpxqqq Ď Ipxq, x P ∆.
In
particular, (28) gives a well defined homomorphism αI,x. Now, Proposition 3.2 readily
implies that pϕ, tαxIpϕpxqquxP∆q is a morphism that induces αI and that pϕ, tαI,xuxP∆q is
the description of the quotient C -bundle in Lemma
a morphism that induces αI (cf.
1.7).
(cid:3)
We note that even when the structure map µA : C0pXq Ñ ZpMpAqq is injective, the
structure maps for I and A{I treated as C0pXq-algebras as in the above proposition, will
hardly ever be injective. Moreover, A{I might not be a continuous C0pXq-algebra, even if
A is, cf. Lemma 1.7. In certain situations, these problems can be circumvented by using
the following proposition.
Proposition 3.11. Suppose that α : A Ñ A is an endomorphism induced by a morphism
pϕ, tαxuxP∆q and I is positively invariant ideal in pA, αq. Put
and treat A{I as a C0pV q-algebra and I as a C0pUq-algebra.
V :" σApPrimpA{Iqq,
U :" σApPrimpIqq
(i) If ϕpV X ∆q Ď V , which is automatic when for each x P ∆ the range of αx is a full
subalgebra of Apxq, then the quotient endomorphism αI : A{I Ñ A{I is induced by
the morphism pϕV X∆, tαI,xuxPV X∆q, cf. (28).
(ii) If U is open and ϕ´1pUq Ď U, which is automatic when A is a continuous C0pXq-
algebra and each αx, x P ∆, is injective, then the restricted C -dynamical system
pI, αIq is a naturally induced by a morphism pϕϕ´1pU q, tαxIpϕpxqquxPϕ´1pU qq.
Proof. (i). Suppose that ϕpV X ∆q Ď V . Then the restriction ϕV : V X ∆ Ñ V is a
proper map and A can be naturally treated as a C0pV q-algebra by Lemma 1.7. Hence the
morphism pϕ, tαI,xuxP∆q from Proposition 3.10 restricts to a morphism pϕV X∆, tαI,xuxPV X∆q
that induces αI.
Now, we show that ϕpV X ∆q Ď V , if for each x P ∆ the range of αx is a full subalgebra of
Apxq. To this end, let V0 " σApPrimpA{Iqq and recall that x P V0 if and only if Ipxq ‰ Apxq,
see (3). Let x P ∆ X V0. We claim that ϕpxq P V0. Indeed, assume on the contrary that
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
31
Ipϕpxqq " Apϕpxqq. Then by positive invariance of I we have αxpApϕpxqqq " αxpIpϕpxqq Ď
Ipxq. Since αxpApϕpxqq is full in Apxq we get
Ipxq " ApxqIpxqApxq Ě ApxqαxpApϕpxqqqApxq " Apxq.
This contradicts the fact that x P Y0, cf. (3). Accordingly, ϕpV0 X ∆q Ď V0. By continuity
of ϕ we get ϕpV X ∆q Ď V .
(ii). If U is open and ϕ´1pUq Ď U, then ϕϕ´1pU q : ϕ´1pUq Ñ U is a proper map and I
is naturally a C0pUq-algebra. Since U " tx P X : Ipxq ‰ t0uu by (2), for any a P I and
x R ϕ´1pUq we have αpaqpxq " 0. Thus pϕϕ´1pU q, tαxIpϕpxqquxPϕ´1pU qq is a morphism that
induces pI, αIq, by Proposition 3.5.
Now, suppose that A is a continuous C0pXq-algebra and each αx, x P ∆, is injective.
Then U is open. By (2), for each x P ϕ´1pUq we may find an element a P I such that
apϕpxqq ‰ 0. Since αx is injective, we have αpaqpxq " αxpapϕpxqqq ‰ 0, and thus x P U,
again by (2). Hence ϕ´1pUq Ď U.
(cid:3)
The above proposition implies that the quotient and the restriction of continuous C0pXq-
dynamical systems described in Example 3.8 are naturally C -dynamical systems of the
same type, see Lemma 5.3 below.
3.3. Extendible morphisms and reversible C0pXq-dynamical systems. In the fore-
going lemma we use the description of multiplier algebras given in Proposition 1.6.
Lemma 3.12. Suppose that A is a C0pXq-algebra, B is a C0pY q-algebra, and α : B Ñ A is
an extendible homomorphism induced by a morphism pϕ, tαxuxP∆q. Each αx : Bpϕpxqq Ñ
Apxq, x P ∆, is extendible and α : MpBq Ñ MpAq is given by
(29)
m P MpBq, x P X.
αpmqpxq "#αx`mpϕpxqq, x P ∆,
x R ∆,
0x,
If additionally either A has local units or A is locally trivial then the set ∆0 defined in (24)
is closed in X.
Proof. Let tµλu be an approximate unit in B and let x P ∆. Then tµλpϕpxqqu is an
approximate unit in Bpϕpxqq and αxpµλpϕpxqqq " αpµλqpxq converges strictly in Apxq.
Hence the homomorphisms αx, x P ∆, are extendible. Recall that α is determined by the
formula αpmqa " limλ αpmµλqa where a P A, m P MpBq. It follows that for any x P ∆ we
have
pαpmqaqpxq " lim
λ
" lim
λ
αpmµλqpxqapxq " lim
λ
αx`pmµλqpϕpxqqapxq
αx`mpϕpxqqµλpϕpxqqapxq " αx`mpϕpxqqapxq.
Thus we get (29).
Now suppose that x0 is a point in the boundary of the set ∆0 " tx P X : αpBqpxq ‰ 0u,
but x0 R ∆0. If A has local units, or if A is locally trivial, then we may choose a P A,
such that }apxq} " 1 for every x in an open neighbourhood U of x0. Then the compact set
tx P X : }pαp1qaqpxq} ě 1{2u contains ∆0 X U but does not contain x0. This leads to a
contradiction, since x0 is in the closure of ∆0 X U.
(cid:3)
32
B. K. KWAŚNIEWSKI
Suppose now that pA, αq is a reversible C -dynamical system where A is a C0pXq-algebra
and α is induced by a morphism pϕ, tαxuxP∆q. In general all we can say about the kernel
of α and its annihilator is that
ker α " ta P A : apyq P čxPϕ´1pyq
ker αx for all y P ϕp∆qu
and pker αqK is contained in
(30)
ta P A : aXzϕp∆q " 0 and apyq P´ čxPϕ´1pyq
ker αx¯K
for y P ϕp∆qu.
Nevertheless, we have the following statement.
Proposition 3.13. Suppose that pA, αq is a reversible C -dynamical system where A is a
C0pXq-algebra and α is induced by a morphism pϕ, tαxuxP∆q.
(i) If all of the endomorphisms αx, x P ∆, are injective then ϕ is injective.
(ii) If ϕ is injective and pA, αq is extendible then the unique regular transfer operator
for pA, αq is determined by the formula
αpaqpxq "#α,x`apϕ´1pxqq, x P ϕp∆q,
x R ϕp∆q,
0x,
(31)
a P A, x P X,
where for each x P ϕp∆q, α,x : Apϕ´1pxqq Ñ Apxq is a completely positive gen-
eralized inverse to αϕ´1pxq : Apxq Ñ Apϕ´1pxqq. The mappings α,x have strictly
continuous extensions α,x and the strictly continuous extension α of α is given
by the formula
αpaqpxq "#α,xpapϕ´1pxqqq, x P ϕp∆q,
x R ϕp∆q,
0x,
a P MpAq, x P X,
where we use the description of multipliers given in Proposition 1.6.
Proof. (i). Injectivity of αx's imply that ker α " ta P A : apxq " 0 for all x P ϕp∆qu and
therefore pker αqK Ď ta P A : apxq " 0 for all x R ϕp∆qu. Let x, y P ∆ be two different
points. Take any b P αpAq such that bpxq ‰ 0 and any h P C0pXq such that hpxq " 1 and
hpyq " 0. Since αpAq " αpAqAαpAq we see that c :" hb is in αpAq. Obviously, cpxq ‰ 0 and
cpyq " 0. Thus for any a P A with αpaq " c we have αxpapϕpxqq ‰ 0 and αypapϕpyqq " 0.
Injectivity of αx and αy implies that ϕpxq ‰ ϕpyq. Hence ϕ is injective.
(ii). Fix x P ∆ and b0 P αxp1ϕpxqqApxqαxp1ϕpxqq (we use the notation of Lemma 3.12).
Take b P αpAq " αp1qAαp1q such that bpxq " b0. Let a P pker αqK be the unique element
such that αpaq " b. Accordingly, b0 " bpxq " αxpapϕpxqqq where apϕpxqq P pker αxqK (here
we use injectivity of ϕ and that pker αqK is contained in the set (30)). It follows that the
range of αx : Apϕpxqq Ñ Apxq is the corner αxp1ϕpxqqApxqαxp1ϕpxqq and αx : pker αxqK Ñ
αxpApϕpxqqq is an isomorphism. The latter fact implies that ker αx is a complemented ideal
in Apϕpxqq. We define the map
(32)
α,ϕpxqpaq :" α´1
x ´αxp1ϕpxqqaαxp1ϕpxqq¯,
a P Apxq,
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
33
where α´1
is the inverse to the isomorphism αx : pker αxqK Ñ αxp1ϕpxqqApxqαxp1ϕpxqq.
x
The maps α,x have strictly continuous extensions which are given by (32) with α´1
x
replaced by the inverse to the strictly continuous isomorphism αx : Mppker αxqKq Ñ
αxp1ϕpxqqMpApxqqαxp1ϕpxqq, cf. Lemma 3.12. Now it is immediate to see that the ho-
momorphisms α,x and α,x, x P ϕp∆q, fulfill the requirements of the assertion.
(cid:3)
Injectivity of the map ϕ in the second part of the above proposition is essential.
Example 3.14. Consider a reversible C -dynamical system pA, αq where A " C3 and
αpaq " pa2, 0, a3q for a " pa1, a2, a3q P A. Then the regular transfer operator αpaq "
p0, a1, a3q for pA, αq is actually an endomorphism. Treating A as a Cpt1, 2uq-algebra where
ap1q " a1 P C and ap2q " pa2, a3q P C2, for a P A, the endomorphism α is induced by the
morphism pϕ, tα1, α2uq where
ϕp1q " ϕp2q " 2,
α1pa2, a3q " a2, α2pa2, a3q " p0, a3q.
But α is not induced by a morphism because the fiber αpaqp2q " pa1, a3q of αpaq depends
on two fibers of a.
3.4. Direct limits. In this subsection, we show that a direct limit of C0pXnq-algebras is
this result, in subsection 4.2, to show that a reversible extension of a C0pXq-dynamical
naturally a C0prXq-algebra, and in certain cases a continuous C0prXq-algebra. We will use
system is a C0prXq-dynamical system.
α1ÝÑ ..., where for each n P N, Bn is a C0pXnq-
algebra and αn : Bn Ñ Bn`1 is a homomorphism induced by a morphism pϕn, tαxuxPXn`1q
(we may assume that the sets Xn are disjoint, so a homomorphism αx is uniquely determined
by the point x P Xn`1). We show that the C -algebraic direct limit B :" limÝÝÑtBn, αnu is
Let us consider a direct sequence B0
α0ÝÑ B1
naturally a C0prXq-algebra over the the topological inverse limit space
To this end, we attach to eachrx " px0, x1, ...q P rX the direct limit C -algebra
rX :" limÐÝÝtXn`1, ϕnu.
Bprxq :" limÝÝÑtBnpxnq, αxn`1u
of the direct sequence B0px0q
and φn : Bn Ñ B be the natural homomorphisms:
αx1ÝÑ B1px1q
αx2ÝÑ B2px2q
αx3ÝÑ ... . We let φrx,n : Bnpxnq Ñ Bprxq
, bn, αn`1pbnq, ...qs,
φnpbnq " r0, ..., 0
loomoonn
φrx,npbnq " r0, ..., 0
loomoonn
, bnpxnq, αxn`1pbnpxnqq, ...qs
where bn P Bn andrx P rX. The following statement can be viewed as a generalization of [18,
Proposition 1.7] to the non-unital case. In contrast to [18] we prove it using the C -bundle
approach.
(33)
continuous.
rx P rX, bn P Bn,
34
B. K. KWAŚNIEWSKI
Proposition 3.15. Retain the above notation and suppose additionally that the mappings
αrx,rn,ms :" αxm ... αxn`2 αxn`1
These are the bonding homomorphisms from Bpxnq to Bpxmq and from Bn to Bm, re-
and
αrn,ms :" αm´1 ... αn`1 αn.
If additionally, all the algebras Bn are continuous C0pXnq-algebras and all the endo-
establishes the natural isomorphism from B " limÝÝÑtBn, αnu onto Γ0pBq.
B Q φnpbnq ÝÑ φnpbnqprxq :" φrx,npbnpxnqq,
ϕn are surjective, n P N. There is a unique topology on B "ŮrxPrX Bprxq making it an upper
semicontinuous C -bundle over rX such that
morphisms αx, x P Xn`1, n P N, are injective, then the C -bundle B " ŮrxPrX Bprxq is
Proof. Forrx P rX and m ą n we put
spectively. Let rx P rX, bn P Bn. To check that the map (33) is well defined assume that
all the more }αrx,rn,mspbnpxnqq} " }αrm,nspbnqpxmq} ă ε. This implies that φrx,npbnpxnqq " 0.
Bprxq.
is upper semicontinuous. Suppose that rx P rX is such that }φnpbnqprxq} ă K. Then there
is m ą n such that }αrx,rn,mspbnpxnqq} ă K. Since αrm,nspbnqpxmq " αrx,rn,mspbnpxnqq and
Xm Q x Ñ }αrm,nspbnqpxq} is upper semicontinuous, there is an open neighborhood U of xm
such that }αrm,nspbnqpxq} ă K for all x P U. It follows that the set
rX Qrx ÞÑ }φnpbnqprxq} P C
φnpbnq " 0. Then for any ε ą 0 and sufficiently large m we have }αrn,mspbnq} ă ε, and
Hence (33) is well defined and clearly it yields a surjective homomorphism from B onto
We show that for a fixed φnpbnq P B, the mapping
(34)
is an open neighborhood ofrx such that forry P rU we have
rU :" try " py0, y1, ...q P rX : ym P Uu
}φnpbnqpyq} ď }αry,rn,mspbnpynqq} ă K.
This proves the upper semicontinuity of (34).
We wish to show that (34) vanishes at infinity. Let ε ą 0. By upper semicontinuity of
is compact as well.
(34) the set trx P rX : }φnpbnqprxq} ě εu is closed, and clearly, it is a subset of trx P rX :
}bnpxnq} ě εu. However, the latter set is compact because the map rX Q rx Ñ xn P Xn is
proper and Xn Q x Ñ }bnpxq} is vanishing at the infinity. Hence trx P rX : }φnpbnqprxq} ě εu
Now, by Fell's theorem, see [55, Theorem C.25], there is a unique topology on B such
that (33) defines a surjective homomorphism from B onto Γ0pBq. We still need to show
that this homomorphism is injective.
To this end, assume that φnpbnq is non zero. Then there exists ε ą 0 such that }αrm,nspbnq} ą
ε for all m ą n. Thus, for each m ą n, the set
Dm :" tx P Xm : }αrm,nspbnqpxq} ě εu
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
35
is nonempty, and it is compact because Xm Q x Ñ }αrm,nspbnqpxq} vanishes at infinity. Note
that ϕmpDm`1q Ď Dm. Thus the sets
rDm :" trx P rX : xm P Dmu
}φnpbnqprx0q} ě ε ą 0.
form a decreasing sequence of compact nonempty sets (non-emptiness follows from surjec-
This finishes the proof of the first part of the assertion.
Assume now that for each n P N, Bn is a continuous C0pXnq-algebra and all of the
tivity of the mappings ϕm). Hence there isrx0 PŞmąn rDm and plainly
endomorphisms αx, x P Xn, are injective. Then }φnpbnqprxq} " }bnpxnq} for all rx P rX,
rX Qrx Ñ xn P Xn and Xn Q xn Ñ }bnpxnq}, is continuous.
bn P Bn, n P N. Hence mapping (34), as a composition of two continuous mappings
(cid:3)
Injectivity of the endomorphisms αx, x P Xn`1, in the second part of Proposition 3.15 is
essential.
Example 3.16. Consider the stationary inductive limit given by the continuous C0pNq-
algebra A :" C0pN, C2q and the endomorphism α : A Ñ A induced by the morphism
pϕ, tαnunPNq where
φp0q " 0, α0 " id,
and
φpnq " n ´ 1, αnpa, bq " pa, 0q, for n ą 0.
The resulting direct limit B " limÝÝÑtA, αu can be viewed as a C0pt´8u Y Zq-algebra with
the obvious topology on t´8u Y Z and fibers B´8 " C2 and Bn " C, n P Z. The image
of the constant function N Q n Ñ pa, bq P C2 (treated as an element of A) in the algebra B
corresponds to the section f with f p´8q " pa, bq and f pnq " a for n P Z. If a ă b, the
function t´8u Y Z Q x Ñ }f pxq} is not lower semicontinuous at ´8.
4. Crossed products of C0pXq-dynamical systems
In this section, we fix a C0pXq-dynamical system pA, αq and denote by pϕ, tαxuxP∆q
a morphism that induces α. We also fix an ideal J in pker αqK and make the following
standing assumption:
‚ σA : PrimpAq Ñ X is surjective, equivalently Apxq ‰ t0u for all x P X.
The above assumption allows us to treat C0pXq as a subalgebra of MpZpAqq. We will
study crossed products C pA, α; Jq using the following tactic. Firstly, we consider reversible
systems. Then we show that the natural reversible J-extension pB, βq of pA, αq is induced
by a morphism, which will immediately lead us to general results.
4.1. The case of a reversible system. In this subsection, we assume that pA, αq is a
cf. Definition 2.34, factors through to the partial homeomorphism of the primitive ideal
reversible C -dynamical system. The dual partial homeomorphism pα : zαpAq Ñ {pker αqK,
space PrimpAq. We denote the latter mapping byqα : PrimpαpAqq Ñ Primppker αqKq. Thus
we have qαpker πq " kerpπ αq for π P zαpAq. With the identifications PrimpαpAqq " tP P
PrimpAq : αpAq Ę P u and Primppker αqKq " tP P PrimpAq : pker αqK Ę P u we have
P P PrimpαpAqq.
(35)
qαpP q " α´1pP q,
36
B. K. KWAŚNIEWSKI
Lemma 4.1. With the above notation, the following diagram
PrimpαpAqq
σA
∆
qα
ϕ
/ Primppker αqKq
σA
/ ϕp∆q
commutes. In particular, if ϕ is free then qα is free, and if A is a continuous C0pXq-algebra
and ϕ is topologically free, then qα is also topologically free.
Proof. Using, among other things, (2) and (30) we see that
σApPrimpαpAqqq " σApPrimpAαpAqAqq " tx P X :`AαpAqApxq ‰ 0u
" tx P X : αpAqpxq ‰ 0u Ď ∆,
σApPrimppker αqKqq " tx P pker αqKpxq ‰ 0u Ď ϕp∆q.
Now let P P PrimpαpAqq. Then x :" σApP q is in ∆. By (1), C0p∆ztxuq A Ď P . Applying
to this inclusion α´1 we get C0´ϕp∆qz tϕpxqu¯α´1pAq Ď α´1pP q. This in view of (35) and
(1) means that σApqαpP qq " ϕpxq " σApP q. The last part of the assertion is straightforward.
Clearly, (topological) freeness ofqα implies (topological) freeness ofpα. Hence Lemma 4.1
Corollary 4.2. Suppose that A is a continuous C0pXq-algebra and ϕ is topologically free.
A representation of the crossed product C pA, αq is faithful if and only if it is faithful on
A.
and Proposition 2.35 give us the following results.
(cid:3)
Corollary 4.3. If ϕ is free then every ideal in C pA, αq is gauge-invariant.
In order to use our pure infiniteness criterion - Proposition 2.46, we need to show that
freeness of ϕ implies that A` residually supports elements of C pA, αq. To this end we
use a technical device introduced in the following lemma, which will allow us to adapt the
arguments of [14] to our setting. Recall that any C -algebra B is a MpBq-bimodule where
pm bq :" mb and pb mq :" pmbq, for m P MpBq, b P B.
Lemma 4.4. The action of h P C0pXq on A as a multiplier of A extends to the action on
C pA, αq as a multiplier of C pA, αq which is uniquely determined by the formulas
(36)
h paunq :" ph aq un,
paαnpbqunq h :" aunpb hq " aαnpb hqun,
where a, b P A, n P N.
Proof. Recall that A is a non-degenerate subalgebra of C pA, αq. In other words, multiplica-
tion from the left defines a non-degenerate homomorphism from A into MpC pA, αqq. This
homomorphisms extends uniquely to the homomorphism from MpAq into MpC pA, αqq.
Composing the latter with µA : C0pXq Ñ ZpMpAqq we get a multiplier action of C0pXq
on C pA, αq that clearly satisfies (36). In view of Proposition 2.14 formulas (36) determine
this action uniquely.
(cid:3)
/
/
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
37
In the following statements we use the C0pXq-bimodule structure on C pA, αq described
in the previous lemma (to increase readability we will suppress the symbol '').
Lemma 4.5 (cf. Lemma 2.3 in [14]). Let k ą 0 and a P AαkpAq. Suppose that x0 P X is
not fixed by ϕk. For every ε ą 0 there is h P C0pXq such that 0 ď h ď 1, hpx0q " 1 and
}hpaukqh} ď ε.
Proof. The proof of [14, Lemma 2.3] is readily adapted to our case; it suffices to replace
the partial crossed product convolution formula with (36).
(cid:3)
Proposition 4.6 (cf. Proposition 2.4 in [14]). Suppose that either ϕ is topologically free
and A is a continuous C0pXq-algebra, or that ϕ is free. Then for every a P C pA, αq and
every ε ą 0 there is h P C0pXq such that
(i) }hEpaqh} ě }Epaq} ´ ε,
(ii) }hEpaqh ´ hah} ď ε,
(iii) h ě 0 and }h} " 1,
where E is the conditional expectation (10).
Proof. We adapt the proof of [14, Proposition 2.4]. A simple approximation argument
implies that we may assume that a is of the form (11). Then Epaq " a0. Let us consider
the non-empty set V " tx P X : }a0pxq} ą }a0} ´ εu and notice that there exists x0 P V
such that x0 is not a fixed point for ϕk for all k " 1, ..., n. Indeed, if ϕ is free, existence of
such x0 is obvious. If A is a continuous C0pXq-algebra then V is open and the existence
of x0 is guaranteed by topological freeness of ϕ. Applying Lemma 4.5 we see that for each
k " 1, ..., n there exists hk P C0pXq such that
hkpx0q " 1,
}hkpakukqhk} ď
,
and 0 ď hk ď 1.
ε
2n
Let h :"śk"1,...,n hk. Then (iii) is immediate, and (i) holds because }ha0h} ě }a0px0q} ą
}a0} ´ ε. For (ii), we have
}ha0h ´ hah} ď ÿk"1,...,n
}hpakukqh} ď ÿk"1,...,n
}hkpakukqhk} ă ε.
(cid:3)
Proposition 4.7. If ϕ is free then A` residually supports elements of C pA, αq`.
Proof. Let I be an ideal in C pA, αq. By Corollaries 4.3 and 2.21, we have the isomorphism
C pA, αq{I -- C pA{I, αIq where I :" A X I is an invariant ideal in pA, αq. The system
pA{I, αIq is reversible by Lemma 2.16 ii). By Proposition 3.10, pA{I, αIq is induced by the
morphism pϕ, tαI,xuxP∆q. Fix a positive element b in C pA, αq{I. Without loss of generality
we may assume that }b} " 1. Applying Proposition 4.6 to pA{I, αIq , we may find a positive
contraction h P MpA{Iq such that (21) holds. Now the last part of the proof of Proposition
2.42 shows that a :" phEpbqh ´ 1{2q` P A{I is non-zero element such that a À b relative to
C pA, αq{I -- C pA{I, αIq.
(cid:3)
Corollary 4.8. Suppose that ϕ is free.
(i) If A is has the ideal property and is purely infinite, then the same holds C pA, αq.
38
B. K. KWAŚNIEWSKI
(ii) If there are finitely many invariant ideals in pA, αq and A is purely infinite, then
C pA, αq is purely infinite.
Proof. The assertion follows from Proposition 4.7 and Proposition 2.46 modulo Remark
2.48.
(cid:3)
4.2. Reversible extension of a C -dynamical system induced by a morphism. Let
us now get back to the case of a not necessarily reversible C0pXq-dynamical system pA, αq.
Let pB, βq denote the reversible J-extension of pA, αq, cf. Definition 2.30. We put
In view of our standing assumption, equality (3) and the fact that J is contained in the
Y " σApPrimpA{Jqq.
of the partial dynamical system pX, ϕq, see Definition 2.32. Our aim is to use Proposition
set (30) we see that Y contains Xzϕp∆q. We denote by prX,rϕq the reversible Y -extension
3.15 to describe B as a C0prXq-algebra and β as an endomorphism induced by a morphism
prϕ, tβrxurxP∆q for a certain field of homomorphisms βrx,rx P ∆.
We start by fixing indispensable notation. Let ∆n :" ϕ´np∆q be the domain of ϕn,
n P N. For x P ∆n we put
αpx,nq :" αx αϕpxq ... αϕn´1pxq,
n ą 0,
and αpx,0q :" id. To each n P N and x P X belonging to the domain of ϕn, we associate the
hereditary subalgebra in Apxq generated by the range of αpx,nq:
(37)
Anpxq :" αpx,nqpApϕnpxqqqApxqαpx,nqpApϕnpxqqq.
We construct fibre C -algebras Bprxq as follows. Ifrx " px0, x1, ...q P X8, we let
αx1ÝÑ A1px1q
αx2ÝÑ A2px3q
Bprxq :" limÝÝÑtAnpxnq, αxn`1u
Bprxq " AN pxN q{JpxN q.
to be the inductive limit of the sequence A0px0q
px0, x1, ..., xN , 0, ...q P XN , we simply put
αx3ÝÑ ... . If rx "
In other words, Bprxq " qxN pAN pxN qq where pid, tqxuxPY q is the morphism that induces the
We will represent the dense -subalgebra ŤnPN Bn of B as an algebra of sections of
B "ŮrxPrX Bprxq. For every a " pa0 ` Jq ' ... ' pan´1 ` Jq ' an P Bn we define the section
πpφnpaqqprxq "$'''&'''%
rx P XN , N ď n,
rx P XN , N ą n,
, anpxnq, αxn`1panpxnqq, αpxn`2,2qpapxnqq, ...s, rx P X8.
We define endomorphisms βrx : Bprϕprxqq Ñ Bprxq,rx P r∆, as follows. We put
and ifrx P XN Xr∆ we let βrx be the inclusion map corresponding to the inclusion Bprϕprxqq "
qxN pAN `1pxN qq Ď Bprxq " qxN pAN pxN qq.
loomoonn
βrxra0, a1, a2, ...s :" ra1, a2, ...s,
aN pxN q ` JpxN q,
αpxN ,N ´nqpanpxnqq ` JpxN q,
r0, ..., 0
ifrx P X8 Xr∆,
quotient map q : A Ñ A{J, see Example 3.6.
πpφnpaqq of B by the formula
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
39
If the system pA, αq happens to be extendible, then by Lemma 3.12 the homomorphisms
αx, and hence also αpx,nq, are extendible. In this case we put
ppx,0q :" 1x, x P X,
and
ppx,nq :" αpx,nqp1ϕnpxqq
for n ą 0, x P ϕ´np∆q.
With this notation the algebras Anpxq are corners ppx,nqApxqppx,nq. We define positive linear
β,rϕprxqra0, a1, a2, ...s :" r0, ppx0,1q a0 ppx0,1q, ppx1,2q a1 ppx1,2q, ...s.
isomorphism B -- Γ0pBq. Identifying B with the algebra of continuous sections of B we
have
maps β,rϕprxq : Bprxq Ñ Bprϕprxqq,rx P r∆, as follows. For X8 Xr∆ we set
Ifrx P XN Xr∆, we put β,rϕprxqpqxN paqq :" qxN pppxN ,N `1q a ppxN ,N `1qq.
Theorem 4.9. Retain the above notation. There is a unique topology on B "ŮrxPrX Bprxq
making it into an upper semicontinuous C -bundle over rX such that π establishes the
βpaqprxq "#βrx`a`rϕprxq, rx P r∆,
rx R r∆,
βpaqprxq "#β,rx`a`rϕ´1prxq, rx P rϕpr∆q,
rx R rϕpr∆q,
where β is the unique regular transfer operator for pB, βq. Moreover, if A is a continuous
C0pXq-algebra and either
and if pA, αq is extendible, then pB, βq is extendible and
0,
0,
(i) J is a complemented ideal, every ideal Jpxq is trivial (i.e. either t0u or Apxq), and
every homomorphism αx is injective, or
(ii) σA is injective, i.e. X -- PrimpAq, cf. Example 3.8,
then B "ŮrxPrX Bprxq is a continuous C -bundle.
Proof. Notice that Bn, n P N, is naturally a C0pYnq-algebra with
Yn :" Y \ Y X ∆1 \ ... \ Y X ∆n´1 \ ∆n
where \ denotes the disjoint sum of topological spaces. Moreover, the bonding homomor-
phism αn : Bn Ñ Bn`1 is induced by a morphism. Indeed, consider the map ϕn : Yn`1 Ñ Yn
given by the diagram
Yn " Y \ ... \ Y X ∆n´1 \
∆n
ϕn
id
id
id
Yn`1
" Y
\ ... \ Y X ∆n´1
\ Y X ∆n
d❏❏❏❏❏❏❏❏❏❏
ϕ
\ ∆n`1.
Let y P Yn`1. Define αx,n : Bnpϕnpyqq Ñ Bn`1pyq to be identity if y P Y X ∆k belongs
to the k-th summand of Xn`1, k ď n ´ 1, to be qy if y P Y X ∆n belongs to the n-th
summand of Xn`1, and to be αy if y P ∆n`1 belongs to the last summand of Yn`1. Then
αn : Bn Ñ Bn`1 is induced by pϕn, tαx,nuxPYn`1q.
O
O
O
O
O
O
O
O
d
40
B. K. KWAŚNIEWSKI
Since ϕp∆q Y Y " X, the mappings ϕn : Yn`1 Ñ Yn are surjective and we may ap-
ply Proposition 3.15 to the inductive system tBn, αnunPN. The arising direct limit C -
bundle can be identified with the one described above. Indeed, we have a natural home-
are naturally isomorphic because they arise as direct limits of direct sequences that can be
the point in the last direct summand of Yn, for n ď N, and xN P Y X ∆N as the point
omorphism Φ : rX Ñ limÐÝÝtYn`1, ϕnu. Namely, for rx " px0, x1, x2, ...q P X8 we define
Φprxq " px0, x1, x2, ...q P limÐÝÝtYn`1, ϕnu where in the latter we treat xn P ∆n as the point
in last direct summand of Yn for all n P N. Forrx " px0, x1, ..., xN , 0, 0, ...q P XN we define
Φprxq " px0, x1, ..., xN , xN , xN , ...q P limÐÝÝtYn`1, ϕnu where in the latter we treat xn P ∆n as
in N-the direct summand in Yn for n ě N. For rx P X8, the algebras Bprxq and BpΦprxqq
naturally identified. Forrx " px0, x1, ..., xN , 0, 0, ...q P XN , BpΦprxqq is naturally isomorphic
to BN pxN q where xN P YN lies in the last direct summand of Y . Thus BpΦprxqq is naturally
Then we get πpφnpaqqprxq " φnpaqpΦprxqq for any a P Bn and rx P rX. This proves the first
Let a " a0 ` J ' ... ' an´1 ` J ' an P Bn andrx P rX. Note that πpφnpβnpaqqqprxq is equal
isomorphic to AN pxN q{JpxN q. Hence we may identify the corresponding fibers of bundles.
part of the assertion.
to
$'''&'''%
aN `1pxN q ` JpxN q,
αpxN ,N `1´nqpanpxn´1qq ` JpxN q,
r0, ..., 0
loomoonn
, αxnpanpxn´1qq, αpxn`1,1qpapxn´1qq, ...s, rx P X8.
rx P XN , N ` 1 ď n,
rx P XN , N ` 1 ą n,
one sees that
Suppose first thatrx " px0, ..., xn, ...q R r∆. Since x0 P Xz∆ we either have xN P ∆N z∆N `1,
when rx P XN for N ă n, or xn P ∆nz∆n`1, otherwise. Thus πpβnpaqqprxq " 0 because for
any x R ∆k, using that ak P αkpAqAαkpAq, we get akpxq " 0. On the other hand, forrx P r∆
Hence, in view of Proposition 3.2, β : B Ñ B is induced by the morphism prϕ, tβrxu
rxPr∆q.
If pA, αq is extendible then pB, βq is extendible by [33, Proposition 2.4]. Moreover,
invoking Proposition 3.13 and formula (32) one concludes that the corresponding transfer
operator β satisfies the formula described in the assertion with the mappings
πpφnpβnpaqqqprxq " βrxpπpφnpaqqprϕprxqqq.
β,rϕprxqpbq :" β´1
is the inverse to the isomorphism βrx : pker βrxqK Ñ βrxp1rϕprxqqBprxqβrxp1rϕprxqq. We
where β´1
leave it to the reader to check that these maps coincide with the maps we have previously
described. This proves the second part of the assertion.
b P Bprxq, rx P r∆,
rx ´βrxp1rϕprxqqaβrxp1rϕprxqq¯,
For the last part of the assertion it suffices to apply the second part of Proposition
3.15. Indeed, if all ideals Jpxq are trivial and all homomorphisms αx are injective, then all
homomorphisms αx,n are injective. Moreover, if J is complemented or σA is injective, then
Bn, n P N, is a continuous C0pYnq-algebra by the last part of Lemma 1.7.
(cid:3)
rx
Remark 4.10. The morphism prϕ, tβrxurxP∆q constructed above can be considered canonical.
In particular, rϕ is always a partial homeomorphism and all the homomorphisms βrx are
injective. Thus even when the initial system pA, αq is already reversible and J " pker αqK,
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
41
so that we have pA, αq " pB, βq, the morphism pϕ, tαxuxP∆q may differ from prϕ, tβrxurxP∆q.
For instance, for the reversible dynamical system pA, αq described in Example 3.14 we
obtain (omitting zero fibers) that B " A " C3 is naturally a C0pt1, 2, 3uq-algebra and
β " α is induced by the morphism pϕ, tα1, α3uq where ϕp1q " 2, ϕp3q " 3, α1 " α3 " id.
4.3. General results. Now, we put together the results of the previous subsections. For
the first statement recall Definition 2.37.
Theorem 4.11. Let pA, αq be a continuous C0pXq-dynamical system. Suppose that ϕ is
topologically free outside the set Y " σApPrimpA{Jqq and either (i) or (ii) in Theorem 4.9
holds. Every injective representation pπ, Uq of pA, αq such that J " ta P A : U U πpaq "
πpaqu give rise to a faithful representation π ¸ U of C pA, α; Jq.
Proof. By Theorem 4.9 the reversible J-extension pB, βq of pA, αq is induced by a morphism
Corollary 4.2 applied to pB, βq.
(cid:3)
based on the reversible Y -extension prX,rϕq of pX, ϕq. Moreover, B is a continuous C0prXq-
algebra and rϕ is topologically free by Lemma 2.38. Hence the assertion follows from
The first part of the following result also implies the uniqueness property described in
It does not require continuity of the C0pXq-algebra A, still it requires
Theorem 4.11.
freeness of ϕ which is a much stronger condition than topological freeness.
Theorem 4.12. Suppose that ϕ is free. Then all ideals in C pA, α; Jq are gauge-invariant
and, in particular, they are in one-to-one correspondence with J-pairs for pA, αq. Moreover,
(i) If A has the ideal property and is purely infinite, then the same holds C pA, α; Jq.
(ii) If there are finitely many J-pairs for pA, αq and A is purely infinite, then C pA, α; Jq
has finitely many ideals and is purely infinite.
Proof. Let pB, βq be the natural J-extension of pA, αq. We show that pB, βq satisfies the
assertion follows by Corollary 4.3, applied to pB, βq.
assumptions of Corollary 4.8, when treated us induced by the morphism prϕ, tβrxu
rxPr∆q de-
scribed in Theorem 4.9. Plainly, rϕ is free, cf. Lemma 2.38. Hence the first part of the
Suppose that A is purely infinite. Since pure infiniteness is preserved under taking direct
sums, quotients, hereditary subalgebras and direct limits, see [27, Propositions 4.3, 4.17
and 4.18], we conclude that B is purely infinite.
(i). It is easy to see that the ideal property is preserved under taking direct sums and
quotients. It is also preserved when passing to direct limits [45, Proposition 2.2] and in the
presence of pure infiniteness it also passes to hereditary subalgebras, see [47, Proposition
2.10]. Thus B is purely infinite and has the ideal property. Accordingly, Corollary 4.8 (i)
applies.
(ii). By Lemma 2.36 (ii) there are finitely many invariant ideals in pB, βq. Hence Corol-
(cid:3)
lary 4.8 (ii) applies to pB, βq.
5. Crossed products of C -algebras with Hausdorff primitive ideal space
In this section, we fix a C -algebra with a Hausdorff primitive ideal space X " PrimpAq
and consider a C0pXq-dynamical system pA, αq described in Example 3.8. Let pϕ, tαxuxP∆q
be the morphism determining α. By Lemma 3.4, without loss of generality we may assume
42
B. K. KWAŚNIEWSKI
that every αx, x P ∆, is nonzero (and thus injective), so that pϕ, tαxuxP∆q is uniquely
determined by α. Thus, we make the following standing assumptions:
‚ X " PrimpAq is a Hausdorff space, σA " id, and every αx, x P ∆, is non-zero.
In particular, we have a bijective correspondence
(38)
X Ě V ÞÝÑ IV :" ta P A : apxq " 0 for all x P V u Ÿ A
between closed subsets of X and ideals in A. We use it to describe ideal structure of the
crossed product C pA, α; Jq in terms of the dynamical system pX, ϕq. We also give some
criteria for C pA, α; Jq to be purely infinite or a Kirchberg algebra. We finish this section
by describing the K-theory of all ideals and quotients in C pA, α; Jq when A " C0pX, Dq
with D a simple C -algebra.
5.1. Ideal structure of C pA, α; Jq. In this subsection, we generalize results proved in
the commutative case (i.e., when D " Cq in [33, Subsection 4.6]. Let us fix an ideal J in
pker αqK. Since pker αqK " IXzϕp∆q we have
J " IY where Y is a closed subset of X such that Y Y ϕp∆q " X.
We have the following dual version of Definitions 2.15 and 2.17.
Definition 5.1 (Definition 4.9 in [33]). A closed set V Ď X is positively invariant under
ϕ if ϕpV X ∆q Ď V , and V is Y -negatively invariant if V Ď Y Y ϕpV X ∆q. If V is both
positively and Y -negatively invariant, we call it Y -invariant. We say that V is invariant if
it is Xzϕp∆q-invariant. A pair pV, V 1q of closed subsets of X satisfying
V is positively ϕ-invariant, V 1 Ď Y
and V 1 Y ϕpV X ∆q " V
is called a Y -pair for pX, ϕq.
Lemma 5.2. An ideal IV in A is positively (resp. J-)invariant if and only if V is positively
(resp. Y -)invariant). A pair pIV , IV 1q of ideals in A is a J-pair for pA, αq if and only if
pV, V 1q is a Y -pair for pX, ϕq.
Proof. We recall that ϕ : ∆ Ñ X is necessarily a closed map. In particular, if V is closed
then ϕpV X ∆q is also closed. Since the endomorphisms αx, x P ∆, are injective, one readily
sees that α´1pIV q " IϕpV X∆q. Using this observation we get the following equivalences
αpIV q Ď IV ðñ IV Ď α´1pIV q ðñ ϕpV X ∆q Ď V,
J X α´1pIV q Ď IV ðñ IY YϕpV X∆q Ď IV ðñ V Ď Y Y ϕpV X ∆q,
This proves the initial part of the assertion. Similarly as above, we get
IV 1 X α´1pIV q " IV ðñ V 1 Y ϕpV X ∆q " V.
Since IV 1 Ď J if and only if V 1 Ď Y , this completes the proof.
(cid:3)
We note that the class C -dynamical systems satisfying our standing assumptions is
closed under taking quotients and restrictions.
Lemma 5.3. If V is a positively invariant closed set, then the quotient endomorphism αIV
is induced by the morphism pϕ∆XV , tαxuxP∆XV q where we treat A{IV as a C0pV q-algebra,
and the restricted endomorphism αIV is induced by the morphism pϕ∆zϕ´1pV q, tαxuxP∆zϕ´1pV qq
where we treat IV as a C0pXzV q-algebra.
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
Proof. It readily follows from Proposition 3.11.
43
(cid:3)
Now we describe the dual topological version of the system introduced in Definition 2.23,
it's action is schematically presented in [30, Figure 1].
Definition 5.4. We define a partial dynamical system pX Y , ϕY q by putting
X Y :" ϕp∆q \ Y,
∆Y :" pϕp∆q X ∆q \ pY X ∆q Ď X Y
and letting ϕY : ∆Y Ñ X Y to map a point x from ∆Y to the point ϕpxq lying in the first
disjoint summand ϕp∆q of X Y .
Lemma 5.5. Let pAJ , αJq be the C -dynamical system described in Definition 2.23. We
may assume the identification PrimpAJ q " X Y , and treating AJ as a C0pX Y q-algebra the
endomorphism αJ is induced by the morphism pϕY , tαxuxP∆Y q. Moreover, if pV, V 1q is Y -
pair for pX, ϕq, then
(39)
is a closed invariant set in pX Y , ϕY q that corresponds to a positively invariant ideal in
pAJ , αJq given by
pV, V 1qY :" pϕp∆q X V q \ V 1 Ď X Y
(40)
IpV,V 1qY " ta P AJ : apxq " 0 for x P pV, V 1qY u.
Proof. In particular, assuming the identification PrimpAJ q " X Y , cf. Lemma 5.5, for the
corresponding Y -pair pV, V 1q we have pIV , IV 1qJ " qker αpIV q ' qJ pIV 1q. Thus the assertion
follows by Proposition 2.25.
(cid:3)
The following proposition generalizes [33, Proposition 4.9] (proved in the commutative
case) and in addition it describes up to Morita-Rieffel equivalence all ideals in C pA, α; Jq.
Proposition 5.6. If ϕ is free then all ideals in C pA, α; Jq are gauge-invariant. In general,
we have a bijective correspondence between gauge-invariant ideals I in C pA, α; Jq and Y -
pairs pV, V 1q for pX, ϕq established by relations
IV " A X I,
IV 1 " ta P A : p1 ´ uuqa P Iu.
Moreover, for the corresponding objects we have an isomorphism
V qq,
C pA, α; Jq{I -- C pA{IV , αIV ; qI 1
pI 1
(41)
and if V 1 " V X Y (equivalently if I is generated by IV ), then I is Morita-Rieffel equivalent
to C pIV , αIV ; IV YY q. In general, I is Morita-Rieffel equivalent to
V
C pIpV,V 1qY , αJ IpV,V 1qY ,Dqq
where IpV,V 1qY is given by (40) and αJ : AJ Ñ AJ is induced by the morphism pϕY , tαxuxP∆Y q.
Proof. By Theorem 4.12 every ideal in C pA, α; Jq is gauge-invariant. Hence the assertion
follows from Theorem 2.19 and Lemmas 5.2 and 5.5.
(cid:3)
Corollary 5.7. Suppose that ϕ is free and that ϕp∆q is open in X (equivalently ker α is a
complemented ideal in A). We have a bijective correspondence between ideals I in C pA, αq
and invariant sets V for pX, ϕq established by the relation IV " A X I. Moreover, for every
ideal I and the corresponding invariant set V we have an isomorphism
C pA, αq{I -- C pA{IV , αIV q
44
B. K. KWAŚNIEWSKI
and I is Morita-Rieffel equivalent to C pIV , αIV q.
Proof. In the proof of Proposition 5.6, instead of Theorem 2.19 we may apply Corollary
2.21.
(cid:3)
5.2. Pure infiniteness and simplicity. For separable C -algebras we have the following
result concerning permanence of pure infiniteness and the ideal property.
Proposition 5.8. Suppose that A is separable and purely infinite and assume that ϕ is
free. If either X is totally disconnected or there are finitely many Y -pairs for pX, ϕq, then
C pA, α; Jq is purely infinite and has the ideal property.
Proof. If X is totally disconnected, then A has the ideal property by [27, Proposition 2.11].
Thus C pA, α; Jq is purely infinite and has the ideal property by Theorem 4.12 (i). If there
are finitely many Y -pairs for pX, ϕq, then C pA, α; Jq is purely infinite and has finitely
many ideals by Theorem 4.12 (ii). Hence [27, Proposition 2.11] implies that C pA, α; Jq
has the ideal property.
(cid:3)
The following characterization of simplicity of C pA, αq, cf. Remark 2.20, is a far reaching
generalization of [33, Theorem 4.4], proved in the case A is commutative.
Proposition 5.9. The crossed product C pA, αq is simple if and only one of the two possible
cases hold:
(i) X is discrete and pX, ϕq is (up to conjugacy) either a truncated shift on t1, ..., nu,
one sided shift on N, or a two-sided shift on Z,
(ii) X is not discrete and ϕ : X Ñ X is a surjection such that there are no non-trivial
closed subsets V of X satisfying ϕpV q " V .
Proof. If (i) or (ii) holds then there are no non-trivial closed invariant set V in pX, ϕq and
ϕ is free. Hence C pA, αq is simple by Corollary 5.7.
Conversely, suppose that C pA, αq is simple. By Proposition 5.6 there are no non-trivial
closed invariant sets in pX, ϕq. The argument in the proof of [33, Theorem 4.4] shows that
either (i) or (ii) holds. In particular, if ϕ : X Ñ X is a surjection then a closed set V is
invariant in pX, ϕq if and only if ϕpV q " V .
(cid:3)
Corollary 5.10. Suppose that A is purely infinite, nuclear and separable. The crossed
product C pA, αq is a Kirchberg C -algebra if and only if one of the conditions (i) or (ii)
in Proposition 5.9 is satisfied.
If C pA, αq is a Kirchberg C -algebra, and A satisfies the UCT then C pA, αq satisfies
the UCT.
Proof. Plainly, C pA, α; Jq is separable, and it is nuclear by Proposition 2.10 (ii). Propo-
sition 5.8 implies that C pA, α; Jq is purely infinite if one of the conditions in Proposition
5.9 holds. Hence the assertion follows from Proposition 5.9.
Now suppose that C pA, αq is a Kirchberg C -algebra and A satisfies the UCT. If ϕ
is surjective then pker αqK " A and if ϕ is not surjective then Xztϕp∆qu " tx0u, cf.
Proposition 5.9 (i), and A " pker αqK ' Apx0q. In both cases pker αqK satisfies the UCT
and thus C pA, αq satisfies the UCT by Proposition 2.10 (iii).
(cid:3)
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
45
5.3. K-theory in the case of a trivial bundle. In this subsection, we assume that the
associated C -bundle A is trivial. In other words, we assume that A " C0pX, Dq where D is
a simple C -algebra. By Proposition 3.9, α is given by the formula (27) where ϕ : ∆ Ñ X is
proper continuous map defined on an open subset ∆ Ď X, and ∆ Q x ÝÑ αx P EndpDqzt0u
is a continuous map. Actually, we make the following standing assumptions:
‚ A " C0pX, Dq where D is a simple C -algebra,
‚ X is totally disconnected, G :" K0pDq is torsion free and K1pDq " t0u.
We treat G as a discrete group and denote by C0pX, Gq the set of continuous functions
f : X Ñ G such that f ´1pGzt0uq is compact.
In other words, any f P C0pX, Gq is of
i"1 χXiτi where Xi's are compact and open subsets of X and τi P G. We
consider C0pX, Gq an abelian group with the group operation defined pointwise. We also
put C0pH, Gq :" t0u.
the form f "řn
Lemma 5.11. For each τ P G, the function ∆ Q x ÞÑ K0pαxqpτ q P G is continuous.
Proof. For any projection p in MnpDq, the function x ÞÑ αxppq P MnpDq is continuous.
Hence the function x ÞÑ rαxppqs0 P G is locally constant. This implies the assertion.
(cid:3)
Definition 5.12. Let δα be a group homomorphism δα : C0pX, Gq Ñ C0pX, Gq given by
δαpf qpxq "#f pxq ´ K0pαxqpf pϕpxqqq, x P ∆
x R ∆.
0
Note that δα is well defined by Lemma 5.11. We define δY
the restriction of δα. We put
α : C0pXzY, Gq Ñ C0pX, Gq to be
K0pX, ϕ, tαxuxP∆; Y q :" cokerpδY
α q,
K1pX, ϕ, tαxuxP∆; Y q :" kerpδY
α q.
If Y " Xzϕp∆q then we write KipX, ϕ, tαxuxP∆q :" KipX, ϕ, tαxuxP∆; Y q, i " 0, 1.
Proposition 5.13. We have the following isomorphism
(42)
KpC pC0pX, Dq, α; Jqq -- KpX, ϕ, tαxuxP∆; Y q.
Proof. Since G " K0pDq is torsion free, K1pC0pXqq " 0 and K1pDq " 0, by Künneth
formulas, see for instance [53, Proposition 2.11], we get
K0pC0pX, Dqq -- K0pC0pXqq b G,
K0pC0pX, Dqq " t0u,
where the isomorphism Ψ : K0pC0pX, Dqq Ñ K0pC0pXqq b K0pDq is determined by the
natural identifications
MrpC0pXq b Dq " C0pXq b MrpDq,
r P N.
It is well known, that the maps ProjpMrpC0pXqqq Q p ÞÑ Tr p P C0pX, Zq determine the
isomorphism K0pC0pXqq -- C0pX, Zq, cf. [51, Exercise 3.4]. The formula
C0pX, Zq b G Q f b τ ÞÝÑ fτ P C0pX, Gq, where fτ pxq :" f pxqτ, x P X,
determines an isomorphism Φ : C0pX, Zq b G Ñ C0pX, Gq, and to see it is enough to note
i"1 χXi b τi
that any element in C0pX, Zq b G can be presented as a sum of the form řn
where Xi's are compact-open and pairwise disjoint subsets of X and τi P G.
46
B. K. KWAŚNIEWSKI
Composing the aforementioned isomorphisms we conclude that we have the isomorphism
K0pC0pX, Dqq -- C0pX, Gq
whose inverse sends a function f "řn
disjoint, and pi P ProjpbMrpDqq, to the element rřn
i"1 χXirpis0 P C0pX, Gq, with Xi compact-open and
i"1 χXipis0 P K0pC0pX, Dqq. We recall
that A " C0pXq b D and J " C0pXzY q b D. The above analysis shows that the following
diagram
K0pιq´K0pαJ q
K0pJq
K0pAq
,
K0pC0pXzY, Dqq
δα
/ K0pC0pX, Dqq
where the vertical arrows are isomorphisms, commutes. Since K1pAq " K1pJq " 0, by
Proposition 2.26, we get
K0pC pA, α; Jqq -- coker pK0pιq ´ K0pαJ qq -- K0pX, ϕ, tαxuxP∆; Y q,
K1pC pA, α; Jqq -- kerpK0pιq ´ K0pαJ qq -- K1pX, ϕ, tαxuxP∆; Y q.
(cid:3)
Remark 5.14. We note that neither Definition 5.12 nor the proof of Theorem 5.13 makes
use of the assumption that D is simple.
We are ready to give formulas for K-theory of all gauge-invariant ideals and corresponding
quotients in C pC0pX, Dq, α; Jq.
Theorem 5.15. If ϕ is free then all ideals in C pC0pX, Dq, α; Jq are gauge-invariant. In
general, if I is a gauge-invariant ideal in C pC0pX, Dq, α; Jq and pV, V 1q is the correspond-
ing Y -pair for pX, ϕq, as described in Proposition 5.6, then
(43)
and
KpC pC0pX, Dq, α; Jq{Iq -- KpV, ϕ∆XV , tαxuxP∆XV ; V 1q
KpIq -- KpU, ϕY ∆zpϕY q´1pU qu, tαxuxP∆Y zpϕY q´1pU qq,
(44)
where pX Y , ϕY q is the system described in Definition 5.4 and U :" X Y zpV, V 1qY where
pV, V 1qY is given by (39). If V 1 " V X Y , then
(45)
KpIq -- KpXzV, ϕ∆zϕ´1pV q, tαxuxP∆zϕ´1pV q; XzpV Y Y qq.
Proof. We get (43) by combining isomorphisms (41) and (42). Since I is Morita-Rieffel
equivalent to C pC0pU, Dq, αJC0pU,Dqq see Proposition 5.6, we get (44) by Theorem 5.13
and [24, Proposition B.5]. Similarly, in view of the last part of Proposition 5.6, we obtain
(45).
(cid:3)
Corollary 5.16. Suppose that ϕ is free and ϕp∆q is open in X. For any ideal I in
C pC0pX, Dq, αq we have
KpC pC0pX, Dq, αq{Iq -- KpV, ϕ∆XV , tαxuxP∆XV q
and
where IV " C0pX, Dq X I.
KpIq -- KpXzV, ϕ∆zϕ´1pV q, tαxuxP∆zϕ´1pV qq
/
/
/
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
47
Proof. In the proof of Theorem 5.15 apply Corollary 5.7 instead of Proposition 5.6.
(cid:3)
We illustrate the above results by indicating how our construction can be used to pro-
duce non-simple classifiable C -algebras from simple ones, only by adding an appropriate
'dynamical ingredient'. Starting from an arbitrary Kirchberg algebra we construct a classi-
fiable C -algebra with a non-Hausdorff primitive ideal space with two points. Such algebras
were the first to be considered in classification of non-simple infinite C -algebras [50], [7].
Example 5.17. Let D be a Kirchberg algebra that satisfies the UCT and let X :" C Ytx0u
be a disjoint sum of the Cantor set C and a clopen singleton tx0u. Then A " CpX, Dq is
a nuclear, separable C -algebra satisfying the UCT. Suppose that ϕ : X Ñ X is such that
ϕpXq " C and ϕ : C Ñ C is a minimal homeomorphism. Then C is the only non-trivial
closed set invariant in pX, ϕq. Hence C pA, αq has the only one non-trivial ideal I (in
particular, PrimpC pA, αqq has two elements and is non-Hausdorff). Note that IC -- D
and αIC " 0. Hence C pIC, αIC q -- D. By Theorem 2.19, we conclude that I is Morita-
Rieffel equivalent to D, and thus KpIq " KpDq. The latter fact can be deduced using
our formulas for K-theory: since XzC " tx0u and ϕXzϕ´1pCq " ϕH is the empty map,
the domain of the corresponding group homomorphism δtx0u
is the zero group C0pH, Gq "
αIC
t0u and the codomain is C0ptx0u, Gq -- G. Moreover, both C pA, αq and C pA, αq{I --
C pIC, αIC q satisfy the UCT, see Proposition 2.10. Note that I is not a complemented ideal
in A and hence has no unit. Concluding, cf. Proposition 5.8, we get
C pA, αq is a strongly purely infinite, nuclear and separable C -algebra with
only one non-trivial ideal I, which is a unique non-unital Kirchberg algebra,
satisfying the UCT, with the same K-theory as D. The non-trivial quo-
tient C pA, αq{I is stably isomorphic to the unique stable Kirchberg algebra
satisfying the UCT with the K-theory equal to KpC, ϕC, tαxuxPCq.
Let us comment on the groups KpC pA, αq{Iq -- KpC, ϕC, tαxuxPCq. In the case G " Z
(that is, for instance, when D " O8) and K0pαxq " id for every x P C, these groups
coincide with K-groups of crossed products studied by Putnam in [48]. In particular, by
[17, Theorem 6.2], K0pC pA, αq{Iq might be any group which can be equipped with a
structure of a simple dimension group. This indicates that allowing the endomorphisms
K0pαxq to be non-trivial or G not to be equal to Z, we have a lot of flexibility for constructing
systems with different K0pC pA, αq{Iq.
Turning to K1pC pA, αq{Iq -- K1pC, ϕC, tαxuxPCq, we note that f P K1pC, ϕC, tαxuxPCq if
and only if
(46)
f pxq " K0pαxqpf pϕpxqqq,
for every x P C.
Thus by minimality of ϕ : C Ñ C, we see that f P K1pC, ϕC, tαxuxPCq is uniquely determined
by its value in a fixed point (0 P C, for instance). Therefore we have
K1pC pA, αq{Iq -- tg P G : there is f P C0pC, Gq such that f p0q " g and (46) holdsu.
If K0pαxq " id for every x P C, then K1pC pA, αq{Iq -- G. If K0pαxq " 0 for at least one
point x P C, then K1pC pA, αq{Iq " t0u (by Lemma 5.11 and minimality of the system).
For the particular case when G " Z, we have K0pαxqpτ q " mx τ for all τ P Z, where
mx P Z is fixed for every x P C . If at least one of the numbers mx is different than 1 then
K1pC pA, αq{Iq " t0u. Indeed, then there is a non-empty open set U Ď C and m P Zzt1u
48
B. K. KWAŚNIEWSKI
such that K0pαxqpτ q " m τ for all x in U. We may assume that m ‰ 0. By minimality, the
orbit of 0 visits U infinitely many times. Thus for any f P K1pC, ϕC, tαxuxPCq the integer
g :" f p0q is divisible by any power of m. This implies that g " 0.
Finally, we include a simple example showing explicitly the dependence of K0pC pA, αqq
on the choice of endomorphisms αx, x P X.
Example 5.18 (K-theory for finite minimal systems). Let pX, ϕq be given by the relations:
X " t1, 2, ..., nu, n ą 1, ∆ " Xzt1u, ϕpiq " i ´ 1, for i " 2, ..., n. Let Y " Xzϕp∆q " tnu.
Assume also that G " Z. Then there are integers m2, ..., mn such that K0pαiqpkq " mik,
for all i " 2, ..., n and k P Z. In particular, identifying C0pX, Zq and C0pXzY, Zq with Zn
and Zn´1 respectively, we see that δY
α : Zn´1 Ñ Zn is given by the formula
δY
α pk1, ..., kn´1q " p0, k2 ´ m2k1, ..., kn´1 ´ mn´1kn´2, ´mnkn´1q.
A moment of thought yields that
K1pC pA, αqq -- kerpδY
α q -- #Z
are non-zero then the map pg1, ..., gnq ÞÑ pg1,řn
The reader may check that cokerpδY
isomorphism cokerpδY
α q -- Z ' Z{pm2m3...mnqZ. Thus we get
if mi " 0 for some i,
otherwise.
t0u
α q -- Z2 if mi " 0 for some i, and if the numbers mi
j qgiq factors through to the
i"2pśj"i`1 mn
K0pC pA, αqq -- Z ' Z{pm2m3...mnqZ.
Appendix A. Relative Cuntz-Pimsner algebras
A C -correspondence over a C -algebra A is a right Hilbert A-module E with a left
action φE : A Ñ LpEq of A on E via adjointable operators. We let JpEq :" φ´1
E pKpEqq
to be the ideal in A consisting of elements that act from the left on E as generalized
compact operators. For any ideal J in JpEq the relative Cuntz-Pimnser algebra OpJ, Eq is
constructed as a quotient of the C -algebra generated by Fock representation of E, see [42,
Definition 2.18] or [37, Definition 4.9]. The C -algebra OpJ, Eq is universal with respect to
appropriately defined representations of E, see [15, Remark 1.4] or [37, Proposition 4.10].
Namely, a representation pπ, πEq of a C -correspondence E consists of a representation
π : A Ñ BpHq in a Hilbert space H and a linear map πE : E Ñ BpHq such that
πEpax bq " πpaqπEpxqπpbq,
πEpxqπEpyq " πpxx, yyAq,
a, b P A, x P E.
Then πE is automatically bounded. If π is faithful, then πE is isometric and we say that
pπ, πEq is injective. The C -subalgebra KpEq Ď LpEq of generalized compact operators is
the closed linear span of the operators Θx,y where Θx,ypzq " xxy, zyA for x, y, z P E. Any
representation pπ, πEq of E induces a homomorphism pπ, πEqp1q : KpEq Ñ BpHq which
satisfies
pπ, πEqp1qpΘx,yq " πEpxqπEpyq,
pπ, πEqp1qpT qπEpxq " πEpT xq
for x, y P E and T P KpEq. The set
Ipπ,πEq :" ta P JpEq : pπ, πEqp1qpφEpaqq " πpaqu
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
49
is an ideal in JpEq. We call Ipπ,πEq the ideal of covariance for pπ, πEq. For any ideal J
contained in JpEq a representation pπ, πEq of E is said to be J-covariant if J Ď Ipπ,πEq.
Note that if pπ, πEq is injective, then we have
Ipπ,πEq " ta P A : πpaq P pπ, πEqp1qpKpEqqu Ď pker φEqK
cf. [24, Page 143].
Proposition A.1. Let E be a C -correspondence over A and let J be an ideal in JpEq.
Then there is a J-covariant representation pι, ιEq of E such that
i) OpJ, Eq is generated as a C -algebra by ιpAq Y ιEpEq,
ii) for any J-covariant representation pπ, πEq of E there is a homomorphism π ¸J πE
of OpJ, Eq such that pπ ¸J πEq ι " π and pπ ¸J πEq ιE " πE.
Moreover, the representation pι, ιEq is injective if and only if J Ď pker φEqK.
Proof. The first part of the assertion is [15, Proposition 1.3]. The second part follows from
[42, Proposition 2.21] and [24, Proposition 3.3].
(cid:3)
It follows from the above proposition that OpJ, Eq is equipped with a gauge circle action
which acts as identity on the image of A in OpJ, Eq. We say that a representation pπ, πEq of
E admits a gauge action if there exists a group homomorphism β : T Ñ AutpC pπpAqYπEqq
such that βzpπpaqq " πpaq and βzpπEpxqq " zπEpxq for all a P A, x P E and z P T.
Proposition A.2 (Corollary 11.8 in [25]). Let us assume that J is an ideal JpEq X
pker φEqK. For any injective J-covariant representation pπ, πEq the homomorphism π ¸J πE
of OpJ, Eq is injective if and only if Ipπ,πEq " J and pπ, πEq admits a gauge action.
Katsura, in [25], described ideals in OpJ, Eq that are invariant under the gauge action in
the following way. For any ideal I in A we define two another ideals
EpIq :" spantxx, a yyA P A : a P I, x, y P Eu,
E´1pIq :" ta P A : xx, a yyA P I for all x, y P Eu.
If EpIq Ď I, then the ideal I is said to be positively invariant, [25, Definition 4.8]. For
any positively invariant ideal I we have a naturally defined quotient C -correspondence
EI " E{EI over A{I. Denoting by qI : A Ñ A{I the quotient map one puts
JEpIq :" ta P A : φEI pqIpaqq P KpEI q, aE´1pIq Ď Iu.
Definition A.3 (Definition 5.6 in [25]). Let E be a C -correspondence over a C -algebra
A. A T -pair of E is a pair pI, I 1q of ideals I, I 1 of A such that I is positively invariant and
I Ď I 1 Ď JEpIq.
Exploiting the results of [25] we get the following theorem.
Theorem A.4. Let E be a C -correspondence over a C -algebra A and J be an ideal of A
contained in pker φEqK X JpEq. Then relations
(47)
establish a bijective correspondence between T -pairs pI, I 1q for E with J Ď I 1 and gauge-
invariant ideals I in OpJ, Eq. Moreover, for objects satisfying (47) we have
I 1 " A X pI ` EEq
I " A X I,
OpJ, Eq{I -- OpqI pI 1q, EIq,
50
B. K. KWAŚNIEWSKI
and if I is generated (as an ideal) by I then I is Morita-Rieffel equivalent to OpJ X I, IEq.
Proof. The first part of the assertion follows from [25, Proposition 11.9]. Now, let pI, I 1q
and I be the corresponding objects satisfying (47) and let q : OpJ, Eq Ñ OpJ, Eq{I be the
quotient map. Put
πpa ` Iq :" qpaq,
πEI px ` IEq :" qpxq,
a P A, x P E.
Since I " A X I, this yields a well defined representation pπ, πEI q of pI, EIq. Since I 1 Ď
pI ` EEq we have qIpI 1q Ď Ipπ,πEI q. Thus pπ, πEI q gives rise to a surjection OpqIpI 1q, EIq Ñ
OpJ, Eq{I. To see it is an isomorphism note that pπ, πEI q admits a gauge action, because
I is gauge-invariant. Moreover, I 1 " A X pI ` EEq implies that a P I 1 if and only if
a ` I P I ` EE, for any a P A. Thus we get
tqIpaq P A{I : πpqIpaqq P pπ, πEqp1qpKpEI qu " ta ` I P A{I : qpaq P qpEEqu
" ta ` I P A{I : a P I 1u " qIpI 1q.
Hence by [25, Corollary 11.8] we get OpqI pI 1q, EIq -- OpJ, Eq{I.
Suppose now that I is generated (as an ideal) by I. The embeddings of I and IE into
OpJ, Eq give rise to a faithful representation pπ, πIEq of pI, IEq in OpJ, Eq. Clearly, pπ, πIEq
admits a gauge action and we have
Ipπ,πIEq " ta P I : a P pIEqpIEqu " ta P I : a P EEu
" ta P A : a P EEu X I " J X I.
Hence by [25, Corollary 11.8] we see that the C -subalgebra B of OpJ, Eq generated by
I and IE is isomorphic to OpJ X I, IEq. It is not difficult to see, cf. the proof of [25,
Proposition 9.3], that B " I OpJ, EqI is the hereditary subalgebra of OpJ, Eq generated by
I. Hence B -- OpJ X I, IEq is Morita-Rieffel equivalent to the ideal I generated by I. (cid:3)
We recall Katsura's version of the Pimsner-Voiculescu exact sequence for a C -correspondence
E. We consider the linking algebra DE " KpE ' Aq in the following matrix representation
A ,
DE "KpEq E
rE
where rE " KpE, Aq is the dual Hilbert bimodule of E -- KpA, Eq. Let ι : J Ñ A,
0 a.
0 0, ι22paq "0 0
ι11 : KpEq Ñ DE and ι22 : A Ñ DE be inclusion maps; ι11paq "a 0
By [24, Proposition B.3], Kipι22q : KipAq Ñ KipDEq, i " 0, 1, are isomorphisms.
Theorem A.5 (Theorem 8.6 in [24]). Within the above notation, the following sequence is
exact:
(48)
K0pJq
K0pιq´K0pι22q´1K0pι11φE J q
K1pOpJ, Eqq
K1piAq
/ K0pAq
K1pAq
K0piAq
/ K0pOpJ, Eqq
K1pιq´K1pι22q´1K1pι11φE J q
K1pJq
/
/
O
O
o
o
o
o
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
51
By a Hilbert A-B bimodule we mean E which is both a left Hilbert A-module and a right
Hilbert B-module with respective inner products x, yB and Ax, y satisfying the condition:
xxy, zyB " Axx, yy z, for all x, y, z P E. If additionally, B " xB, ByB and A " AxA, Ay, we
say that A and B are Morita-Rieffel equivalent (or strongly Morita equivalent). A Hilbert
A-A bimodule is also called a Hilbert bimodule over A. If E is a Hilbert bimodule over A
then it is also a C -correspondence and Katsura's algebra OE associated to E coincides
with the C -algebra associated to E in [3], see [23, Proposition 3.7].
Suppose that E is a Hilbert bimodule over A. Then E induces a partial homeomorphism
pE of pA dual to E, see [32, Definition 1.1]. More specifically, xE, EyA and AxE, Ey are
ideals in A and pE : {xE, EyA Ñ {AxE, Ey is a homeomorphism, which factors through the
induced representation functor E -Ind. The latter is defined as follows: if π : A Ñ BpHq is a
representation, then E -Indpπq : A Ñ BpE bπ Hq is a representation where the Hilbert space
E bπ H is generated by simple tensors xbπ h, x P E, h P H, satisfying xx1 bπ h1, x2 bπ h2y "
xh1, πpxx1, x2yAqh2y, and
E -Indpπqpaqpx bπ hq " paxq bπ h,
a P A.
By [32, Theorems 2.2 and 2.5] we have the following result.
Theorem A.6. Let pE be the partial homeomorphism of pA associated to a Hilbert bimodule
E. If pE is topologically free, then every non-zero ideal in OE has a non-zero intersection
of A. If pE is free, then every ideal in OE is gauge-invariant.
C -correspondences associated to C -dynamical systems. Let us now fix C -dynamical
system pA, αq. We associate to pA, αq the C -correspondence given by
Eα :" αpAqA,
a x :" αpaqx,
x a :" xa,
xx, yyA :" xy,
where a P A, x, y P E. Clearly, we have ker φEα " ker α.
Lemma A.7. We have JpEαq " A and the map KpEαq Q Θx,y ÞÑ xy P αpAqAαpAq yields
an isomorphism of C -algebras.
Proof. The proof is straightforward and thus left to the reader.
(cid:3)
Proposition A.8. For any ideal J in pker αqK there is a natural isomorphism C pA, α; Jq --
OpJ, Eαq. More precisely, the relation
πEαpxq " U πpxq,
x P αpAqA,
yields a one-to-one correspondence between representations pπ, Uq of pA, αq and represen-
tations pπ, πEαq of Eα where π : A Ñ BpHq is a non-degenerate representation. For the
corresponding representations we have Ipπ,πEα q " Ipπ,U q.
Proof. By [35, Proposition 3.26] crossed product by α treated as a completely positive map
coincides with the crossed product considered in the present paper (note that an operator S
in [35] plays the role of U ). By [35, Lemma 3.25] the GNS C -correspondence associated to
α (treated as a completely positive map) is naturally isomorphic to Eα. Thus the assertion
follows from [35, Propositions 3.10].
(cid:3)
Some of the following facts were stated without proof in [33, Appendix A].
52
B. K. KWAŚNIEWSKI
Proposition A.9. An ideal I in A is positively invariant for Eα if and only if I is pos-
itively invariant in pA, αq. Moreover, if I is positively invariant, then we have natural
identifications:
IEα " EαI ,
pEαqI " EαI .
Proof. Clearly, αpIq Ď I if and only if EαpIq " AαpIqA Ď I, which proves the first part
of the assertion. If I is positively invariant then αpIqA " αpIqIA " αpIqI, which allows
us the identification IEα " EαI . The natural algebraic isomorphism EαI " αIpA{IqA{I "
qIpαpAqAq -- αpAqA{αpAqI " pEαqI intertwines the operations of C -correspondences.
Hence it is an isomorphism that allows us the identification pEαqI " EαI .
(cid:3)
Proposition A.10. Suppose that I and I 1 are ideals in A.
pI, I 1q is a T -pair for Eα with J Ď I 1 ðñ pI, I 1q is a J-pair for pA, αq.
Proof. We have E´1
α pIq " ta P A : xαpaqy P I for all x, y P αpAqAu " α´1pIq. By
Proposition A.9 we may identify pEαqI with EαI . Hence JppEαqIq " A{I and we get
JEαpIq " ta P A : aα´1pIq Ď Iu . Thus if we assume that I is positively invariant, cf.
Proposition A.9, then we get
I Ď I 1 Ď JEαpIq ðñ I 1 X α´1pIq " I.
This implies the assertion.
(cid:3)
We recall, see [23, Subsection 3.3] or [31, Proposition 1.11], that a C -correspondence E
over A is a Hilbert bimodule over A if and only if φE : pker φEqK X JpEq Ñ KpEq is an
isomorphism, and then Axx, yy " φ´1
Proposition A.11. The C -correspondence Eα is a Hilbert bimodule over A if and only
if pA, αq is reversible and then
E pΘx,yq.
Axx, yy " α´1pxyq
where α´1 is the inverse to the isomorphism α : pker αqK Ñ αpAq.
Proof. Clearly, we have pker φEαqK " pker αqK. Thus Lemma A.7 implies that φEα :
pker φEαqK X JpEαq " pker αqK Ñ KpEq -- αpAqAαpAq is an isomorphism if and only if
the system pA, αq is reversible.
(cid:3)
Let us now consider a reversible C -dynamical system pA, αq and the corresponding
Hilbert bimodule Eα. Clearly, AxEα, Eαy " pker αqK and xEα, EαyA " AαpAqA. Under the
homeomorphism dual to Eα can be identified with the one described in Definition 2.34:
Lemma A.12. Let pA, αq be a reversible C -dynamical system. The homeomorphisms
standard identifications we have zαpAq " trπs P pA : πpαpAqq ‰ 0u " {xE, EyA. The partial
pα : zαpAq Ñ {pker αqK and xEα : {xEα, EαyA Ñ {AxEα, Eαy coincide.
pαprπsq is the equivalence class of the representation π α : A Ñ BpπpαpAqqHq. Since
πpαpAqqH " πpαpAqAqH and }ři ai bπ hi}2 " }ři,jxhi, πpa
i ajqhjy} " }ři πpaiqhi}2, for
Proof. Let π : A Ñ BpHq be an irreducible representation such that πpαpAqq ‰ 0. Then
ai P Eα " αpAqA, hi P H, i " 1, ..., n, we see that a bπ h ÞÑ πpaqh yields a unitary operator
U : Eα bπ H Ñ πpαpAqqH. Furthermore, for a P A, b P αpAq and h P H we have
rEα -IndpπqpaqU sπpbqh " Eα -Indpπqpaq b bπ h " pαpaqbq bπ h " rU pπ αqpaqsπpbqh.
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
53
(cid:3)
Hence U intertwines Eα -Ind and π α. This proves that xEα "pα.
by using (48) and the following lemma.
We get the exact sequence for crossed products by endomorphisms, see Proposition 2.26,
Lemma A.13. Let J be an ideal in pker αqKq. With the notation preceding Theorem A.5
applied to Eα we have Kipι22 αJ q " Kipι11 φEαJ q, i " 1, 2.
Proof. For brevity, we put E :" Eα. Let 5 : E Ñ rE be the canonical antilinear isomorphism,
and let α´1 be the inverse to the isomorphism α : pker αqK Ñ αppker αqKq. Plainly, the map
M2pαpJqq Qa11 a12
a21 a22 Φ
ÞÝÑφEpα´1pa11qq a12
a22 P DE,
5pa
21q
is a homomorphism of C -algebras. The following diagram commutes:
J
ι11α
z✉✉✉✉✉✉✉✉✉✉
Φ
M2pαpJqq
ι11φE
❆❆❆❆❆❆❆❆
/ DE
.
Therefore
Kipι11 φEJ q " KipΦ ι11 αJ q,
i " 0, 1.
Recall that for any C -algebra B the homomorphisms ιii : B Ñ M2pBq, i " 1, 2, induce
the same mappings on K-groups. Thus Kipι11 αJq " Kipι22 αJ q, i " 1, 2. By the form
of Φ we see that Φ ι22 α " ι22 α on J. Concluding, for i " 0, 1 we get
Kipι11 φEJ q " KipΦ ι11 αJq " KipΦ ι22 αJ q " Kipι22 αJ q.
(cid:3)
Acknowledgments
The author would like to thank the referee for her/his valuable comments and sugges-
tions that resulted in rewriting of the previous version of the manuscript and including
K-theory considerations. The author also thanks Jamie Gabe for his comments and an
additional source of motivation for the present work. This research was supported by the
7th European Community Framework Programme FP7-PEOPLE-2013-IEF: Marie-Curie
Action: Project 'OperaDynaDual' number 621724; and in part by NCN; grant number
DEC-2011/01/D/ST1/04112.
References
[1] S. Adji, Crossed products of C -algebras by semigroups of endomorphisms, Ph.D. thesis, University
of Newcastle, 1995.
[2] A. B. Antonevich, V. I. Bakhtin, A. V. Lebedev, Crossed product of C -algebra by an endomorphism,
coefficient algebras and transfer operators, Sb. Math. 202(9) (2011), 1253 -- 1283.
[3] B. Abadie, S. Eilers, R. Exel, Morita equivalence for crossed products by Hilbert C -bimodules. Trans.
Amer. Math. Soc. 350(8) (1998), 3043 -- 3054.
[4] A. B. Antonevich, A. V. Lebedev, Functional differential equations: I. C -theory, Longman Scientific
& Technical, Harlow, Essex, England, 1994.
[5] M. A. Bastos, C. A. Fernandes, Y. I. Karlovich, A C -algebra of singular integral operators with shifts
admitting distinct fixed points, J. Math. Anal. Appl. 413 (2014), 502 -- 524.
z
/
54
B. K. KWAŚNIEWSKI
[6] E. Blanchard, E. Kirchberg, Non-simple purely infinite C -algebras:
the Hausdorff case, J. Funct.
Anal. 207 (2004), 461 -- 513.
[7] A. Bonkat, Bivariante K-Theorie fur Kategorien projektiver Systeme von C*-Algebren, Ph.D. Thesis,
Westf. Wilhelms-Universitat Munster, 2002.
[8] L. G. Brown, G. K. Pedersen, C -algebras of real rank zero, J. Funct. Anal. 99 (1991), 131 -- 149.
[9] J. Cuntz, Simple C -algebras generated by isometries, Comm. Math. Phys., 57 (1977), 173 -- 185.
[10] G. A. Elliot, Some simple C -algebras constructed as crossed products with discrete outer automor-
phism groups, Publ. Res. Inst. Math. Sci. 16 (1980), no. 1, 299 -- 311.
[11] G. A. Elliot, On the classification of C -algebras of real rank zero, J. Reine Angew. Math. 443 (1993),
179 -- 219.
[12] G. A. Elliot, M. Rørdam, Classification of certain infinite simple C -algebras, II, Comment. Math.
Helv. 70 (1995), 615 -- 638.
[13] R. Exel, A new look at the crossed-product of a C -algebra by an endomorphism, Ergodic Theory
Dynam. Systems, 23 (2003), 1733 -- 1750.
[14] R. Exel, M. Laca, J. Quigg, Partial dynamical systems and C -algebras generated by partial isometries,
J. Operator Theory, 47 (2002), 169 -- 186.
[15] N. J. Fowler, P. S, Muhly, I. Raeburn, Representations of Cuntz-Pimsner algebras, Indiana Univ.
Math. J. 52 (2003) 569 -- 605.
[16] T. Giordano, A. Sierakowski, Purely infinite partial crossed products, J. Funct. Anal. 266(9) (2014),
5733 -- 5764.
[17] R. Herman, I. F. Putnam, C. F. Skau Ordered Bratteli diagrams, dimension groups and topological
dynamics, Internat. J. Math. 3 (1992), 827 -- 864.
[18] I. Hirshberg, M. Rørdam, W. Winter, C0pXq-algebras, stability and strongly self-absorbing C -algebras,
Math. Ann. 339 (2007), 695 -- 732.
[19] Ja A Jeong, Purely infinite simple C -crossed products, Proc. Amer. Math. Soc. 123 (1995), 3075 --
3078.
[20] P. Jolissaint and G. Robertson, Simple purely infinite C -algebras and n-filling actions, J. Funct. Anal.
175 (2000), 197 -- 213.
[21] R.V. Kadison, J.R. Ringrose, Fundamentals of the Theory of Operator Algebras. Vol. 2. Advanced
Theory, Academic Press, 1986.
[22] Yu. I. Karlovich, A local-trajcetory method and iosmorphism theorems for non-local C -algebras, In
Operator Theory: Advances and Applications 170, 137-166, Birkhauser, Basel, 2006.
[23] T. Katsura, A construction of C -algebras from C -correspondences. Contemp. Math. vol. 335, pp.
173 -- 182, Amer. Math. Soc., Providence (2003).
[24] T. Katsura, On C -algebras associated with C -correspondences, J. Funct. Anal., 217(2) (2004), 366 --
401.
[25] T. Katsura, Ideal structure of C -algebras associated with C -correspondences, Pacific J. Math. 230
(2007), 107 -- 145.
[26] E. Kirchberg, Das nicht-kommutative Michael-Auswahlprinzip und die Klassifikation nicht-einfacher
Algebren, C -Algebras (Munster, 1999), Springer, Berlin, 2000, 92 -- 141.
[27] E. Kirchberg, M. Rørdam, Non-simple purely infinite C -algebras, Amer. J. Math. 122 (2000), 637 --
666.
[28] E. Kirchberg and A. Sierakowski, Filling families and strong pure infiniteness, preprint, arX-
ive:1503.08519v1.
[29] E. Kirchberg and A. Sierakowski, Strong pure infiniteness of crossed products, preprint, arX-
ive:1312.5195v2.
[30] B. K. Kwaśniewski, C -algebras associated with reversible extensions of logistic maps, Sb. Math.
203(10) (2012), 1448 -- 1489.
[31] B. K. Kwaśniewski, C -algebras generalizing both relative Cuntz-Pimsner and Doplicher-Roberts alge-
bras, Trans. Amer. Math. Soc. 365 (2013), 1809 -- 1873.
[32] B. K. Kwaśniewski, Topological freeness for Hilbert bimodules. Israel J. Math. 199(2) (2014), 641 -- 650.
CROSSED PRODUCTS BY ENDOMORPHISMS OF C0pXq-ALGEBRAS
55
[33] B. K. Kwaśniewski, Ideal structure of crossed products by endomorphisms via reversible extensions of
C -dynamical systems, Internat. J. Math. 26 (2015), no. 3, 1550022 [45 pages].
[34] B. K. Kwaśniewski, Crossed products by interactions and graph algebras, Integral Equations Operator
Theory 80(3) (2014), 415-451.
[35] B. K. Kwaśniewski, Exel's crossed products and crossed products by completely positive maps, to appear
in Houston J. Math., arXiv:1404.4929.
[36] B. K. Kwaśniewski, Extensions of C -dynamical systems to systems with complete transfer operators,
Math. Notes, 98(3) (2015), 419 -- 428.
[37] B. K. Kwaśniewski, A. V. Lebedev, Crossed products by endomorphisms and reduction of relations in
relative Cuntz-Pimsner algebras, J. Funct. Anal. 264(8) (2013), 1806-1847.
[38] B. K. Kwaśniewski, W. Szymański, Pure infiniteness and ideal structure of C -algebras associated to
Fell bundles, preprint, arxiv: 1505.05202 v1.
[39] M. Laca and J. S. Spielberg, Purely infinite C -algebras from boundary actions of discrete groups, J.
reine angew. Math. 480 (1996), 125 -- 139.
[40] J. Lindiarni, I. Raeburn, Partial-isometric crossed products by semigroups of endomorphisms, J. Op-
erator Theory, 52 (2004), 61 -- 87.
[41] R. Meyer, R. Nest, C -algebras over topological spaces: filtrated K-theory, Canad. J. Math. 64(2)
(2012), 368 -- 408.
[42] P. S. Muhly and B. Solel, Tensor algebras over C -correspondences (representations, dilations, and
C -envelopes), J. Funct. Anal. 158 (1998), 389 -- 457.
[43] E. Ortega, E. Pardo, Purely infinite crossed products by endomorphisms, J. Math. Anal. Appl. 412
(2014) 466 -- 477.
[44] W. L. Paschke, The crossed product of a C -algebra by an endomorphism, Proc. Amer. math. Soc. 80
(1980), 113 -- 118.
[45] C. Pasnicu, On the AH algebras with the ideal property, J. Operator Theory 43(2) (2000), 389 -- 407.
[46] C. Pasnicu and N. C. Phillips, Crossed products by spectrally free actions, J. Funct. Anal. 269 (2015),
915 -- 967.
[47] C. Pasnicu, M. Rørdam, Purely infinite C -algebras of real rank zero, J. Reine Angew. Math. 613
(2007), 51-73.
[48] I. Putnam, The C -algebras associated with minimal homeomorphisms of the Cantor set, Pacific J.
Math. 136 (1989), 329 -- 353.
[49] M. Rørdam, Classification of certain infinite simple C -algebras. J. Funct. Anal. 131 (1995), 415 -- 458.
[50] M. Rørdam, Classification of extensions of certain C -algebras by their six term exact sequences in
K-theory, Math. Ann. 308(1) (1997), 93 -- 117.
[51] M. Rørdam, F. Larsen, N. J. Laustsen, An introduction to K-theory for C -algebras, London Math.
Society Student Texts 49, Cambridge University Press, 2000.
[52] M. Rørdam, A. Sierakowski, Purely infinite C -algebras arising from crossed products. Ergodic Theory
Dynam. Systems 32 (2012), 273 -- 293.
[53] C. Schochet, Topological methods for -algebras. II. Geometry resolutions and the Künneth formula.
Pacific J. Math. 98 (1982), no. 2, 443 -- 458.
[54] P. J. Stacey, Crossed products of C -algebras by -endomorphisms, J. Aust. Math. Soc. 54 (1993),
204 -- 212.
[55] D. P. Williams, Crossed Products of C -Algebras, Math. Surveys and Monographs v. 134, Amer.
Math. Soc., Providence, Rh.I., 2007.
Department of Mathematics and Computer Science, The University of Southern Den-
mark, Campusvej 55, DK -- 5230 Odense M, Denmark // Institute of Mathematics, University
of Bialystok, ul. K. Ciolkowskiego 1M, 15-245 Bialystok, Poland
E-mail address: [email protected]
|
1603.09165 | 1 | 1603 | 2016-03-30T13:02:04 | Partial transformation groupoids attached to graphs and semigroups | [
"math.OA",
"math.DS"
] | We introduce the notion of continuous orbit equivalence for partial dynamical systems, and give an equivalent characterization in terms of Cartan-isomorphisms for partial C*-crossed products. Both graph C*-algebras and semigroup C*-algebras can be described as C*-algebras attached to partial dynamical systems. As applications, for graphs, we generalize and explain a result of Matsumoto and Matui relating orbit equivalence and Cartan-isomorphism, and for semigroups, we strengthen several structural results for semigroup C*-algebras concerning amenability, nuclearity as well as simplicity of boundary quotients. We also discuss pure infiniteness for partial transformation groupoids arising from graphs and semigroups. | math.OA | math |
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS
AND SEMIGROUPS
XIN LI
Abstract. We introduce the notion of continuous orbit equivalence for partial dynamical sys-
tems, and give an equivalent characterization in terms of Cartan-isomorphisms for partial C*-
crossed products. Both graph C*-algebras and semigroup C*-algebras can be described as C*-
algebras attached to partial dynamical systems. As applications, for graphs, we generalize and
explain a result of Matsumoto and Matui relating orbit equivalence and Cartan-isomorphism,
and for semigroups, we strengthen several structural results for semigroup C*-algebras concern-
ing amenability, nuclearity as well as simplicity of boundary quotients. We also discuss pure
infiniteness for partial transformation groupoids arising from graphs and semigroups.
1. Introduction
Recently, in the setting of ordinary topological dynamical systems, the notion of continuous or-
bit equivalence was introduced, and a C*-algebraic characterization was given involving Cartan-
isomorphisms (see [22]). We present a generalization to partial dynamical systems. This step
is important as many C*-algebras appear naturally as crossed products attached to partial dy-
namical systems, while the setting of ordinary dynamical systems seems rather restricted. Our
main motivation stems from graph C*-algebras and semigroup C*-algebras, both of which can
be described in a very natural way as C*-algebras of partial dynamical systems. This description
turns out to be very helpful for the study of structural properties of these C*-algebras.
In the case of graphs, we obtain a very easy explanation why all graph C*-algebras are nuclear.
This goes back to the observation that non-abelian free semigroups embed into amenable groups.
Furthermore, we generalize and explain results in [25] about orbit equivalence and Cartan
isomorphism from shifts of finite type to general graphs. We note that a generalization of the
results in [25] has been established independently in [4].
For semigroup C*-algebras and their boundary quotients, we are able to generalize several struc-
tural results from [20]. The key idea is that using partial transformation groupoids, we obtain
structural results for our C*-algebras without assuming independence or the Toeplitz condition,
which were crucial in our previous approach (see [19, 20]). We obtain general characterizations
for nuclearity of semigroup C*-algebras and simplicity of boundary quotients.
2010 Mathematics Subject Classification. Primary 46L05; Secondary 37B05, 37A20.
Research supported by EPSRC grant EP/M009718/1.
1
2
XIN LI
Furthermore, we study which partial transformation groupoids of graphs and semigroups are
purely infinite, in the sense of [27].
In the case of graphs, we are able to prove that the
partial system of a graph is residually topologically free and purely infinite if and only if the
corresponding graph C*-algebra is purely infinite. For semigroups, we show that the partial
transformation groupoid corresponding to a boundary quotient is purely infinite as long as
the semigroup is not left reversible. We are also able to identify a class of integral domains
whose ax + b-semigroups have purely infinite partial transformation groupoids. These results
strengthen and explain previous results in [20, 21].
2. Partial actions, transformation groupoids, C*-algebras and Cartan
subalgebras
In the following, groups are discrete and countable, and topological spaces are locally compact,
Hausdorff and second countable.
Definition 2.1. Let G be a group with identity e, and let X be a topological space. A partial
action α of G on X consists of
• a collection {Ug}g∈G of open subsets Ug ⊆ X,
• a collection {αg}g∈G of homeomorphisms αg : Ug−1 → Ug, x 7→ g.x such that
-- Ue = X, αe = idX;
-- for all g1, g2 ∈ G, we have g2.(U(g1g2)−1 ∩Ug−1
for all x ∈ U(g1g2)−1 ∩ Ug−1
2
.
) = Ug2 ∩Ug−1
1
, and (g1g2).x = g1.(g2.x)
2
We call such a triple (X, G, α) a partial system, and denote it by α : G y X or simply G y X.
Let α : G y X be a partial system. The dual action α∗ of α is the partial action (in the sense
of [28]) of G on C0(X) given by α∗
The transformation groupoid attached to the partial system α : G y X is given by
g : C0(Ug−1) → C0(Ug), f 7→ f (g−1.⊔).
G α⋉ X := {(g, x) ∈ G × X: g ∈ G, x ∈ Ug−1} ,
with source map s(g, x) = x, range map r(g, x) = g.x, composition (g1, g2.x)(g2, x) = (g1g2, x)
and inverse (g, x)−1 = (g−1, g.x). We equip G α⋉ X with the subspace topology from G × X.
Usually, we write G ⋉ X for G α⋉ X if the action α is understood. The unit space of G ⋉ X
coincides with X. Since G is discrete, G ⋉ X is an ´etale groupoid. Actually, if we set Gx :=
{g ∈ G: x ∈ Ug−1} and Gx := {g ∈ G: x ∈ Ug} for x ∈ X, then we have canonical identifications
s−1(x) ∼= Gx, (g, x) 7→ g and r−1(x) ∼= Gx, (g, g−1.x) 7→ g.
Let us now recall the construction (from [28]) of the reduced crossed product C0(X) ⋊α∗,r G
attached to our partial system α : G y X. As with groupoids, we omit α∗ in our notation for the
crossed product. First of all C0(X) ⋊ℓ1 G :=nPg fgδg ∈ ℓ1(G, C0(X)): fg ∈ C0(Ug)o becomes
a *-algebra under component-wise addition, multiplication given by(cid:16)Pg fgδg(cid:17) ·(cid:16)Ph
fhδh(cid:17) :=
Pg,h α∗
g−1(fg) fh)δgh and involution(cid:16)Pg fgδg(cid:17)∗
:=Pg α∗
g(f ∗
g−1)δg.
g(α∗
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
3
As in [28], we construct a representation of C0(X) ⋊ℓ1 G. Viewing X as a discrete set, we
define ℓ2X and the representation M : C0(X) → L(ℓ2X), f 7→ M(f ), where M(f ) is the
multiplication operator M(f )(ξ) := f · ξ for ξ ∈ ℓ2X. M is obviously a faithful representation
of C0(X). Every g ∈ G leads to a twist of M, namely Mg : C0(X) → L(ℓ2X) given by
Mg(f )ξ := f Ug(g.⊔) · ξUg−1 . Here we view f Ug(g.⊔) as an element in Cb(Ug−1), and Cb(Ug−1)
acts on ℓ2Ug−1 just by multiplication operators. Given ξ ∈ ℓ2X, we set ξUg−1 (x) := ξ(x) if
x ∈ Ug−1 and ξUg−1 (x) := 0 if x /∈ Ug−1. In other words, ξUg−1 is the component of ξ in ℓ2Ug−1
with respect to the decomposition ℓ2X = ℓ2Ug−1 ⊕ ℓ2U c
g−1. So we have Mg(f )ξ(x) = f (g.x)ξ(x)
if x ∈ Ug−1 and Mg(f )ξ(x) = 0 if x /∈ Ug−1.
Consider now the Hilbert space H := ℓ2(G, ℓ2X) ∼= ℓ2G ⊗ ℓ2X, and define the representation
µ : C0(X) → L(H) given by µ(f )(δg ⊗ ξ) := δg ⊗ Mg(f )ξ. For g ∈ G, let Eg be the orthogonal
projection onto µ(C0(Ug−1))H. Moreover, let λ denote the left regular representation of G on
ℓ2G, and set Vg := (λg ⊗ I) · Eg. Here I is the identity operator on H.
We can now define the representation µ × λ : C0(X) ⋊ℓ1 G → L(H),Pg fgδg 7→Pg µ(fg)Vg.
Following the original definition in [28], we set C0(X) ⋊r G := C0(X) ⋊ℓ1 G
k·kµ×λ.
The following result follows from [1, Theorem 3.3], but we give a direct and short proof.
Proposition 2.2. The canonical homomorphism
(1)
Cc(G ⋉ X) → C0(X) ⋊ℓ1
G, θ 7→Xg
θ(g, g−1.⊔)δg,
where θ(g, g−1.⊔) is the function Ug−1 → C, x 7→ θ(g, g−1.x), extends to an isomorphism C ∗
X)
∼=−→ C0(X) ⋊r G.
r (G⋉
Proof. As above, let µ × λ be the representation C0(X) ⋊ℓ1 G → L(H) which we used to define
C0(X) ⋊r G. Our first observation is
since for x /∈ h−1.(Uh ∩ Ug−1) = U(gh)−1 ∩ Uh−1, f Uh(h.x) = 0 for f ∈ C0(Ug−1). Therefore,
To see this, observe that for all g ∈ G, Im (Eg) ⊆ Lh δh ⊗ ℓ2(Uh−1 ∩ U(gh)−1). This holds
π(C0(Ug−1))(δh ⊗ ℓ2X) ⊆ δh ⊗ ℓ2(Uh−1 ∩ U(gh)−1). Hence Im (Eg) ⊆Lh δh ⊗ ℓ2(Uh−1 ∩ U(gh)−1),
and thus, Im (Vg) ⊆Lh δgh ⊗ ℓ2(Uh−1 ∩ U(gh)−1) ⊆Lh δh ⊗ ℓ2Uh−1. This shows "⊆" in (2). For
"⊇", note that for f ∈ C0(X), (µ × λ)(f δe) = µ(f )Ee, and for ξ ∈ ℓ2Uh−1, µ(f )Ee(δh ⊗ ξ) =
δh ⊗ f Uh(h.⊔)ξ. So (µ × λ)(f δe)(H) contains δh ⊗ f · ξ for all f ∈ C0(Uh−1) and ξ ∈ ℓ2Uh−1,
hence also δh ⊗ ℓ2Uh−1. This proves "⊇".
For x ∈ X, let Gx = {g ∈ G: x ∈ Ug−1} as before. Our second observation is that for every
x ∈ X, the subspace Hx := ℓ2Gx ⊗ δx is (µ × λ)-invariant.
It is clear that µ(f ) leaves Hx
invariant for all f ∈ C0(X). For g, h ∈ G, Eg(δh ⊗ δx) = δh ⊗ δx if x ∈ Uh−1 ∩ U(gh)−1, and if
that is the case, then Vg(δh ⊗ δx) = δgh ⊗ δx ∈ Hx.
Therefore, H =(cid:0)Lx∈X Hx(cid:1) ⊕ (µ × λ)(C0(X) ⋊ℓ1 G)(H)⊥ is a decomposition of H into µ × λ-
invariant subspaces. For x ∈ X, set ρx := (µ × λ)Hx. Then C0(X) ⋊r G = C0(X) ⋊ℓ1 G
k·kLx ρx .
(2)
Im (µ × λ)(H) =Mh∈G
δh ⊗ ℓ2Uh−1.
4
XIN LI
Moreover, we have for x ∈ Uh−1,
ρx Xg
fgδg! (δh ⊗ δx) = Xg
µ(fg)Vg(δh ⊗ δx) = Xg: x∈U(gh)−1
δgh ⊗ fg(gh.x)δx = Xk∈Gx
= Xg: x∈U(gh)−1
µ(fg)(δgh ⊗ δx)
δk ⊗ fkh−1(k.x)δx
(3)
(4)
Let us now compare this construction with the construction of the reduced groupoid C*-
algebra of G ⋉ X. Obviously, (1) is an embedding of Cc(G ⋉ X) as a subalgebra which is
k·kℓ1-dense in C0(X) ⋊ℓ1 G. Therefore, C0(X) ⋊r G = Cc(G ⋉ X)
Now, to construct the reduced groupoid C*-algebra C ∗
r (G ⋉ X), we follow [35, § 2.3.4] and
construct for every x ∈ X the representation πx : Cc(G ⋉ X) → L(ℓ2(s−1(x))) by setting
πx(θ)(ξ)(ζ) := Pη ∈ s−1(x) θ(ζη−1)ξ(η). In our case, using s−1(x) = Gx × {x}, we obtain for
ξ = δh ⊗ δx with h ∈ Gx: πx(θ)(δh ⊗ δx)(k, x) = θ((k.x)(h, x)−1) = θ(kh−1, h.x). Thus,
k·kLx ρx .
πx(θ)(δh ⊗ δx)(k, x) = Xk∈Gx
θ(kh−1, h.x)δk ⊗ δx.
r (G ⋉ X) = Cc(G ⋉ X)
By definition, C ∗
and
coincide on Cc(G ⋉ X), it suffices to show that for every x ∈ X, πx and the restriction
k·kLx ρx
of ρx to Cc(G ⋉ X) are unitarily equivalent. Given x ∈ X, using s−1(x) = Gx × {x}, we obtain
the canonical unitary ℓ2(s−1(x)) ∼= Hx = ℓ2(Gx) ⊗ δx, so that we may think of both ρx and πx
as representations on ℓ2(Gx) ⊗ δx. We then have for x ∈ X, θ ∈ Cc(G ⋉ X) and h ∈ Gx:
k·kLx πx . Therefore, in order to show that k·kLx ρx
ρx(θ)(δh ⊗ δx)
θ(g, g−1.⊔)δg)(δh ⊗ δx)
δk ⊗ θ(kh−1, h.x)δx
(4)
= πx(θ)(δh ⊗ δx).
(1)
= ρx(Xg
(3)
= Xk∈Gx
This yields the canonical identification C0(X) ⋊r G ∼= C ∗
r (G ⋉ X), as desired.
(cid:3)
Following [15], we define topological freeness as follows:
Definition 2.3. A partial system G y X is called topologically free if for every e 6= g ∈ G,
{x ∈ Ug−1: g.x 6= x} is dense in Ug−1.
Lemma 2.4. A partial system G y X is topologically free if and only if
{x ∈ X: g.x 6= x f or all e 6= g ∈ Gx} is dense in X.
Proof. The direction "⇐" is simple: If {x ∈ X: g.x 6= x f or all e 6= g ∈ Gx} is dense in X,
then in particular, for every fixed e 6= g ∈ G, the set {x ∈ X: g.x 6= x if g ∈ Gx} is dense in
X. Hence {x ∈ Ug−1: g.x 6= x} = {x ∈ X: g.x 6= x if g ∈ Gx} ∩ Ug−1 is dense in Ug−1.
For "⇒", note that because of topological freeness, we know that for every e 6= g ∈ G, the
open set {x ∈ Ug−1: g.x 6= x} ∪ Ug−1
is dense in X. Therefore,
c
{x ∈ X: g.x 6= x f or all e 6= g ∈ Gx} = \e6=g∈G
{x ∈ X: g.x 6= x if g ∈ Gx}
must be dense in X since it contains Te6=g∈G(cid:0){x ∈ Ug−1: g.x 6= x} ∪ Ug−1
X by the Baire category theorem.
c(cid:1) which is dense in
(cid:3)
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
5
Corollary 2.5. A partial system G y X is topologically free if and only if the transformation
groupoid G ⋉ X is topologically principal.
Proof. Recall (see [34]) that a topological groupoid G is called topologically principal if the
set of points in G(0) with trivial isotropy is dense in G(0). Here, x ∈ G(0) is said to have
trivial isotropy if for all γ ∈ G, s(γ) = t(γ) = x already implies γ = x. In the case of the
transformation groupoid G = G ⋉ X, x ∈ X has trivial isotropy if whenever g ∈ G satisfies
g ∈ Gx and g.x = x, then we must have g = e. Hence the set of points with trivial isotropy
is nothing else but {x ∈ X: g.x 6= x f or all e 6= g ∈ Gx}. With this observation, our corollary
follows immediately from Lemma 2.4.
(cid:3)
Let us now introduce the notion of continuous orbit equivalence for partial systems, generalizing
[22, Definition 2.5].
Definition 2.6. Partial systems G y X and H y Y are called continuously orbit equivalent
if there exists a homeomorphism ϕ : X
∼=−→ Y and continuous maps a : Sg∈G {g} × Ug−1 → H,
b : Sh∈H {h} × Vh−1 → G (where Vh−1 is the domain of the partial homeomorphism attached to
h ∈ H) such that
(5)
(6)
Implicitly, we require here that a(g, x) ∈ Hϕ(x) and b(h, y) ∈ Gϕ−1(y).
ϕ−1(h.y) = b(h, y).ϕ−1(y).
ϕ(g.x) = a(g, x).ϕ(x),
Note that in particular, ϕ(Gx.x) = Hϕ(x).ϕ(x).
In analogy to [22, Theorem 1.2], we obtain
Theorem 2.7. Let G y X and H y Y be topologically free partial systems. Then the following
are equivalent:
(i) G y X and H y Y are continuously orbit equivalent,
(ii) the transformation groupoids G ⋉ X and H ⋉ Y are isomorphic as topological groupoids,
∼=−→ C0(Y ) ⋊r H with Φ(C0(X)) = C0(Y ).
(iii) there exists an isomorphism Φ : C0(X) ⋊r G
Moreover, "(ii) ⇒ (i)" holds in general (i.e., without the assumption of topological freeness).
Proof. The proof is completely analogous to the one of [22, Theorem 1.2]. Therefore, we refer
the reader to [22] for details, and only mention the key ideas.
For "(i) ⇒ (ii)", observe that G ⋉ X → H ⋉ Y, (g, x) 7→ (a(g, x), ϕ(x)) and H ⋉ Y →
G⋉X, (h, y) 7→ (b(h, y), ϕ−1(y)) are continuous homomorphisms of groupoids which are inverse
to each other. Here ϕ, a and b are as in Definition 2.6. This uses topological freeness as in [22].
For "(ii) ⇒ (i)", let χ : G ⋉ X → H ⋉ Y be an isomorphism of topological groupoids. Set
χ−→ H ⋉ Y → H, and b as the
χ−1
−→ G ⋉ X → G. Then it is easy to check that ϕ, a
ϕ := χX, a as the composition Sg∈G {g} × Ug−1 → G ⋉ X
compositionSh∈H {h} × Vh−1 → H ⋉ Y
and b satisfy the conditions in Definition 2.6. This does not use topological freeness.
For "(ii) ⇔ (iii)", observe that by Corollary 2.5 and [34, Theorem 5.2], (C0(X)⋊rG, C0(X)) ∼=
r (G ⋉ X), C0(X)) and (C0(Y ) ⋊r H, C0(Y )) ∼= (C ∗
r (H ⋉ Y ), C0(Y )) are Cartan pairs, in the
(cid:3)
(C ∗
sense of [34]. Then apply [34, Proposition 4.13].
6
XIN LI
3. Examples: Inverse semigroups, graphs, and subsemigroups of groups
3.1. Inverse semigroups. Let S be an inverse semigroup with zero 0 and E the semilattice
of idempotents of S. Assume that σ is a partial homomorphism from S to a group G which is
idempotent pure, i.e., σ is a map S× → G, where S× = S \ {0}, such that σ(st) = σ(s)σ(t) for
all s, t ∈ S× with st 6= 0, and that σ−1(e) = E×(= E \ {0}). Existence of such an idempotent
pure partial homomorphism is equivalent to saying that S is strongly 0-E-unitary.
In this situation, we describe a partial action G y C ∗(E) such that the (left) reduced
C*-algebra C ∗
λ(S) of S is canonically isomorphic to C ∗(E) ⋊r G. This is the analogue of [29,
Theorem 5.2], but without the assumption that G is the universal group. Our observation gives
us more freedom in the choice of idempotent pure partial homomorphisms. This will lead to a
very simple criterion for nuclearity of inverse semigroup C*-algebras.
λ(S): For s ∈ S, let λs : ℓ2S× → ℓ2S×, δx 7→ δsx if s∗s ≥ xx∗ and
First, recall the definition of C ∗
δx 7→ 0 else. The reduced C*-algebra C ∗
λ(S) of S is the sub-C*-algebra of L(ℓ2S×) generated
by {λs: s ∈ S}. Note that by construction, λ0 = 0. This is why we work with ℓ2S× instead of
λ(S). In the following, we write e for λe ∈ C ∗(E). The
ℓ2S. Let C ∗(E) := C ∗({λe: e ∈ E}) ⊆ C ∗
full inverse semigroup C*-algebra C ∗(S) of an inverse semigroup S is the universal C*-algebra
for *-representations of S by partial isometries. We mod out 0 if S has a zero, as in [31].
Now let us describe the partial action G y C ∗(E). For g ∈ G,
let Dg be the sub-C*-
algebra (actually ideal) of C ∗(E) given by Dg−1 = span({s∗s: s ∈ S×, σ(s) = g}). As σ is
idempotent pure, we have De = C ∗(E). For every g ∈ G, we have a C*-isomorphism
α∗
g : Dg−1 → Dg, s∗s 7→ ss∗. The corresponding dual action is given as follows: Let bE =
Spec (C ∗(E)), and for every g ∈ G, set Ug = Spec (Dg) ⊆ bE. It is easy to see that Ug−1 =
nχ ∈ bE: χ(s∗s) = 1 for some s ∈ S× with σ(s) = go. αg : Ug−1 → Ug is given by αg(χ) =
the partial transformation groupoid G ⋉ bE, in the sense of [32, § 4.3]. Hence we obtain
g−1. Given χ ∈ Ug−1 and s ∈ S× with σ(s) = g and χ(s∗s) = 1, we have αg(χ)(e) = χ(s∗es).
The same proof as in [29] gives an explicit identification of the universal groupoid of S with
χ◦α∗
Proposition 3.1. We have C ∗
isomorphisms C ∗
λ(S) ∼= C ∗
λ(S) → C ∗(E) ⋊r G, λs 7→ (ss∗)uσ(s) and C ∗(S) → C ∗(E) ⋊ G, s 7→ (ss∗)δσ(s).
r (G ⋉ bE) and C ∗(S) ∼= C ∗(G ⋉ bE), and there are
In many situations, one is interested not only in the reduced C*-algebra of the inverse semigroup,
but also in its boundary quotient. Let us describe the partial system corresponding to this
quotient. Given a semilattice E, let bEmax be the subset of bE consisting of those χ ∈ bE
such that {e ∈ E: χ(e) = 1} is maximal. We then set ∂bE := bEmax ⊆ bE. Now let E be the
semilattice of idempotents in an inverse semigroup S. As ∂bE ⊆ bE is closed, we obtain a short
exact sequence 0 → I → C0(bE) → C0(∂bE) → 0. Viewing I as a subset of C ∗
λ(S), we form the
λ(S) generated by I. The boundary quotient in Exel's sense (see [9, 10, 11, 13])
λ(S) := C ∗
ideal hIi of C ∗
is given by ∂C ∗
λ(S)/ hIi.
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
7
Lemma 3.2. Let S be an inverse semigroup as above, with an idempotent pure partial homo-
there exists f ∈ E with ψ(f ) = 1 but χ(s∗f s) = 0. Then ψ ∈ Ug since χ(ss∗) = 1, which
implies ψ(ss∗) = 1. Consider g−1.ψ given by g−1.ψ(e) = ψ(ses∗). Then for every e ∈ E,
χ(e) = 1 implies χ(s∗ses∗s) = 1, hence χ(s∗(ses∗)s) = 1, so that g−1.ψ(e) = ψ(ses∗) = 1. But
morphism σ from S to a group G, with semilattice of idempotents bE and partial system G y bE.
Then ∂bE is G-invariant.
Proof. Let us first show that for every g ∈ G, g.(Ug−1 ∩bEmax) ⊆ Ug ∩bEmax. Take χ ∈ bEmax with
χ(s∗s) = 1 for some s ∈ S with σ(s) = g. Then g.χ(e) = χ(s∗es). Assume that g.χ /∈ bEmax.
This means that there is ψ ∈ bEmax such that ψ(e) = 1 for all e ∈ E with χ(e) = 1, and
χ(s∗f s) = 0 and g−1.ψ(s∗f s) = ψ(ss∗f ss∗) = ψ(f ) = 1. This contradicts χ ∈ bEmax. Hence
g.(Ug−1 ∩ bEmax) ⊆ Ug ∩ bEmax. To see that g.(Ug−1 ∩ ∂bE) ⊆ Ug ∩ ∂bE, let χ ∈ Ug−1 ∩ ∂bE and
choose a net (χi)i in bEmax with limi χi = χ. As Ug−1 is open, we may assume that all the χi lie
in Ug−1. Then g.χi ∈ bEmax, and limi g.χi = g.χ. This implies g.χ ∈ ∂bE.
λ(S) ∼= C0(bE) ⋊r G from Proposition 3.1
λ(S) ∼= C0(bE) ⋊r G from Proposition 3.1,
Lemma 3.3. Let S be an inverse semigroup with an idempotent pure partial homomorphism σ
from S to an exact group G. Then the identification C ∗
identifies ∂C ∗
Proof. It is easy to see that under the identification C ∗
hIi corresponds to I ⋊r G. Since G is exact, we obtain a short exact sequence 0 → I ⋊r G →
C ∗
(cid:3)
λ(S) → 0 by [12, Theorem 22.9].
λ(S) with C0(∂bE) ⋊r G.
λ(S) → ∂C ∗
(cid:3)
Corollary 3.4. Let S be an inverse semigroup with an idempotent pure partial homomorphism
σ from S to a group G. If G is amenable, then both C ∗
λ(S) are nuclear. If G is
exact, then both C ∗
λ(S) and ∂C ∗
λ(S) are exact.
λ(S) and ∂C ∗
Proof. By Proposition 3.1 and Lemma 3.3, both C ∗
λ(S) can be described as reduced
partial crossed products by G, hence as reduced C*-algebras of Fell bundles over G. Therefore,
our claims follow from [12, Theorem 20.7, Theorem 25.10 and Theorem 25.12].
(cid:3)
λ(S) and ∂C ∗
3.2. Graphs. Let E = (E 0, E 1, r, s) be a (countable) graph with vertices E 0, edges E 1 and
range and source maps r, s : E 1 → E 0. Let E ∗ be the set of finite paths in E, and let l(µ)
denote the length of a path µ ∈ E ∗. The graph inverse semigroup SE is given by SE =
{(µ, ν) ∈ E ∗ × E ∗: s(µ) = s(ν)} ∪ {0}, where (µ, ν)∗ = (ν, µ) and
The semilattice EE of idempotents of SE is given by {(µ, µ) ∈ E ∗ × E ∗} ∪ {0}, hence can be
identified canonically with E ∗ ∪ {0}. Multiplication in EE is given by
(µ, ν)(ζ, η) =
µ · ν =
(µ, ν′η)
(µζ ′, η)
0
if ν = ζν′,
if ζ = νζ ′,
else.
µ if µ = νµ′,
if ν = µν′,
ν
0
else.
8
XIN LI
Note that we write µν for concatenation of paths and µ · ν for the product in EE.
It is easy to see that C ∗
∂C ∗
λ(SE) is canonically isomorphic to the graph C*-algebra of E.
λ(SE) is canonically isomorphic to the Toeplitz C*-algebra of E, and that
Remark 3.5. Note that we are using the convention that the partial isometry sµ for µ ∈ E 1
has source projection es(µ) corresponding to the source of µ, and range projection dominated
by er(µ), the projection corresponding to the range of µ. This is the same convention as in [37],
but different from the one in [4, 17].
Let us construct an idempotent pure partial homomorphism on SE. Let FE 1 be the free group
generated by E 1. We view µ ∈ E ∗ with l(µ) ≥ 1 as elements in FE 1 in a canonical way. Define
σ : S×
E → FE 1 by setting σ(µ, ν) = µν−1 if l(µ), l(ν) ≥ 1, σ(µ, ν) = µ if l(µ) ≥ 1 and l(ν) = 0,
σ(µ, ν) = ν−1 if l(µ) = 0 and l(ν) ≥ 1, and σ(µ, ν) = e if l(µ) = l(ν) = 0. Here e is the identity
of FE 1. It is easy to check that σ is an idempotent pure partial homomorphism from SE to FE 1.
Clearly, σ is the universal one, in the sense of [29].
E 1 for the free semigroup generated by E 1. Let F+
Remark 3.6. Let us show how a modification of σ produces an easy argument for nuclearity of
graph C*-algebras. We write F+
2 and F2 be the
free semigroup and the free group on two generators. Clearly, there is a semigroup embedding
F+
E 1 ֒→ F+
2 ֒→
F2/F′′
2 is the second commutator subgroup of F2. Hence we obtain an embedding
F+
E 1 ֒→ F+
2. By universal property of FE 1, we obtain a homomorphism FE 1 → F2/F′′
2,
which has to factorize as FE 1 → FE 1/F′′
2 (E 1 is countable by assumption). Moreover, by [18], we have an embedding F+
2, such that the diagram
2, where F′′
2 ֒→ F2/F′′
E 1 → F2/F′′
F+
E 1
FE 1
/ F+
2
/ F2/F′′
2
/ FE 1/F′′
E 1
commutes. Thus, the canonical homomorphism F+
E 1 → FE 1/F′′
E 1 is injective.
Now let σ′′ be the composition S×
E
E 1. σ′′ is again a partial homomorphism,
and σ′′ is idempotent pure because F+
E 1 is solvable,
in particular amenable. Hence Corollary 3.4 implies that both the Toeplitz C*-algebra as well
as the graph C*-algebra of E are nuclear.
E 1 is injective. Moreover, FE 1/F′′
σ−→ FE 1 → FE 1/F′′
E 1 → FE 1/F′′
Let us now come back to σ : S×
E → FE 1, and describe the corresponding partial actions
FE 1 y cEE and FE 1 y ∂E, where ∂E := ∂cEE .
We start with FE 1 y cEE. For g ∈ FE 1, Ug−1 is empty unless g ∈ Im (σ), i.e., g = αβ−1
such g, Ug−1 consists of those χ ∈ cEE such that there exists (µ, ν) ∈ S×
for some paths α, β ∈ (E ∗ \ E 0) ∪ {e}, and if α, β both lie in E ∗ \ E 0, then s(α) = s(β). For
E with χ(ν) = 1 and
σ(µ, ν) = αβ−1. For such χ with (µ, ν) as above, (g.χ)(ζ) = χ(νζ ′) if ζ = µζ ′ and (g.χ)(ζ) =
0 otherwise. By Proposition 3.1, the Toeplitz C*-algebra of E is canonically isomorphic to
λ(SE) ∼= C ∗
C ∗
r (FE 1 ⋉cEE) ∼= C0(cEE ) ⋊r FE 1.
_
/
/
/
O
O
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
9
Following [37] (see also [14]), the partial system FE 1 y ∂E can be described explicitly as
0 }, where E ∞ is the set of infinite
follows: As a set, we identify ∂E with E ∞ ∪ {α ∈ E ∗: s(α) /∈ E 0
paths in E. To describe the topology, let µ ∈ E ∗ and set
Z(µ) = {ν ∈ E ∗ ∪ E ∞: ν = µν′ for some ν′ ∈ E ∗ ∪ E ∞} .
Given a finite subset F of r−1(s(µ)), let Z(µ \ F ) = Z(µ) \Sν∈F Z(µν). Then
∂E ∩ Z(µ \ F ), µ ∈ E ∗, F ⊆ r−1(s(µ)) finite
is a basis for the topology of ∂E. The FE 1-action is just given by the restriction of the partial
system FE 1 y cEE , which has been described above. By Lemma 3.3, the graph C*-algebra
of E is canonically isomorphic to ∂C ∗
Theorem 20.9]).
λ(SE) ∼= C ∗
r (FE 1 ⋉ ∂E) ∼= C0(∂E) ⋊r FE 1 (compare [9,
It is known (see [4, Proposition 2.3], and also Lemma 3.7) that FE 1 ⋉ ∂E is topologically
principal, or equivalently (see Corollary 2.5), that FE 1 y ∂E is topologically free, if and only if
E satisfies condition (L). Recall that E satisfies condition (L) if every loop has an entry, i.e., for
every µ = µ1 · · · µn ∈ E ∗ with µi ∈ E 1, n ≥ 1 and s(µn) = r(µ1), there is ν ∈ E 1 and 1 ≤ i ≤ n
with r(ν) = r(µi) and ν 6= µi.
Let us now relate our transformation groupoid to the groupoid attached to topological Markov
shifts or graphs (see [25] and [4]), and our notion of continuous orbit equivalence for partial
systems to continuous orbit equivalence for topological Markov shifts or graphs (see [25] and
[4]). Once these relations are established, we will see that Theorem 2.7 applied to graphs
generalizes [25, Theorem 2.3] and gives an alternative interpretation for [4, Theorem 5.1].
Let E be a graph as above. We compare the transformation groupoid FE 1 ⋉∂E with the groupoid
GE from [4, § 2.3] (see [25, § 2.2] for the case of topological Markov shifts). The groupoid GE
is given by GE =(cid:8)(α, n, β) ∈ ∂E × Z × ∂E: n = k − l f or k, l ∈ Z+ and σk(α) = σl(β)(cid:9). Here,
we identify ∂E with E ∞∪{α ∈ E ∗: s(α) /∈ E 0
0 } and define for µ = µ1µ2µ3 . . . in E ∞ with l(µ) ≥ 2
(µi ∈ E 1) σ(µ) = µ2µ3 . . . , and σ(µ) = s(µ) if µ ∈ E 1. For µ ∈ E 0, σ is not defined. An equation
like σk(α) = σl(β) always implicitly means that σk(α) and σl(β) are defined, i.e., l(α) ≥ k and
l(β) ≥ l. Moreover, we write Z+ = {z ∈ Z: z ≥ 0}.
Lemma 3.7. FE 1 ⋉ ∂E and GE are isomorphic as topological groupoids.
Proof. It is easy to see that GE → FE 1 ⋉ ∂E, (λν, n, µν) 7→ (λµ−1, µν) and FE 1 ⋉ ∂E →
GE, (λµ−1, µν) 7→ (λν, l(λ) − l(µ), µν) are mutually inverse (continuous) groupoid homomor-
phisms. (Note that these expressions for the maps only make sense for l(µ), l(ν) ≥ 1. But there
is an obvious way to define these maps for l(µ) = 0 or l(ν) = 0.)
(cid:3)
Let us also compare the notion of (continuous) orbit equivalence for graphs introduced in [4,
Definition 3.1] (see also [25, § 2.1] for the case of topological Markov shifts) with continuous
orbit equivalence of the corresponding partial systems. Let E and F be two graphs, with σE
and σF as above, and partial systems FE 1 y ∂E and FF 1 y ∂F . E and F are (continuously)
orbit equivalent in the sense of [4, Definition 3.1] if there exists a homeomorphism ϕ : ∂E → ∂F
10
XIN LI
together with continuous maps k, l : ∂E → Z+ and k′, l′ : ∂F → Z+ such that
(7)
(8)
σk(ζ)
F (ϕ(σE(ζ))) = σl(ζ)
F (ϕ(ζ)) f or all ζ ∈ ∂E,
σk′(η)
E
(ϕ−1(σF (η))) = σl′(η)
E
(ϕ−1(η)) f or all η ∈ ∂F .
Lemma 3.8. If FE 1 y ∂E and FF 1 y ∂F are continuously orbit equivalent, then E and F are
orbit equivalent.
If E and F satisfy condition (L), then the converse holds, i.e., if E and F are orbit equivalent,
then FE 1 y ∂E and FF 1 y ∂F are continuously orbit equivalent.
Proof. Assume that FE 1 y ∂E and FF 1 y ∂F are continuously orbit equivalent via ϕ : ∂E ∼= ∂F
0 , it suffices to
and continuous maps a, b as in Definition 2.6. As ∂E =(cid:16)Fζ∈E 1 ζ∂E(cid:17) ⊔ E 0 \ E 0
define k and l on Fζ∈E 1 ζ∂E (on the remaining part, just set k and l to be 0). For ζ ∈ E 1
and ζξ ∈ ζ∂E, we know that a(ζ −1, ζξ) has the reduced form λκ−1, so that we can define k
and l on ζ∂E by setting k(ζξ) := l(λ) and l(ζξ) := l(κ). These maps k and l are obviously
locally constant, hence continuous. Moreover, we know that ϕ(ζ −1.(ζξ)) = (λκ−1).ϕ(ζξ), which
means that there exists ω ∈ ∂F with ϕ(ζξ) = κω and ϕ(ξ) = ϕ(ζ −1.(ζξ)) = λω. Therefore,
σl(λ)
F (ϕ(σE(ζξ))) = σl(λ)
F (ϕ(ζξ)). Thus, (7) holds. k′
and l′ are defined in a similar way, using b.
F (λω) = ω = σl(κ)
F (ϕ(ξ)) = σl(λ)
F (κω) = σl(κ)
Now assume that E and F satisfy condition (L), and suppose conversely that E and F are
orbit equivalent. Because of condition (L), our partial systems FE 1 y ∂E and FF 1 y ∂F are
topologically free. Therefore, to prove that they are continuously orbit equivalent, all we have
to show is that for every g ∈ FE 1 and x ∈ Ug−1, there exists an open neighbourhood U of x and
h ∈ FF 1 such that ϕ(g.¯x) = h.ϕ(¯x) for all ¯x ∈ U, and analogously for ϕ−1. In our case, since
FE 1 is generated by E 1, it suffices to consider g ∈ (E 1)−1. Take ζ ∈ E 1 and x ∈ ζ∂E. Choose
λ, κ ∈ F ∗ with l(λ) ≥ k(x), l(κ) ≥ l(x), such that ϕ(x) ∈ κ∂F and ϕ(ζ −1.x) ∈ λ∂F , and for
all ζξ ∈ ζ∂E with ϕ(ζξ) ∈ κ∂F and ϕ(ξ) = ϕ(ζ −1.(ζξ)) ∈ λ∂F , we have k(ζξ) = k(x) and
l(ζξ) = l(x). Such λ and κ exist because k and l are continuous, hence locally constant. Set
U := {ζξ ∈ ζ∂E: ϕ(ζξ) ∈ κ∂F and ϕ(ξ) ∈ λ∂F }. U is obviously an open neighbourhood of x.
Set k := k(x) and l := l(x). For all ζξ ∈ U, we have σk
F (ϕ(ζξ)).
Thus, when we write λ = λ′λ′′ with l(λ′) = k and κ = κ′κ′′ with l(κ′) = l, we even know that
(λ′)−1.ϕ(ξ) = σk
F (ϕ(ζξ)) = (κ′)−1.ϕ(ζξ), and hence ϕ(ζ −1.(ζξ)) = (λ′(κ′)−1).ϕ(ζξ)
for all ζξ ∈ U. Thus ϕ has the desired property. The proof for ϕ−1 is analogous.
(cid:3)
F (ϕ(σE (ζξ))) = σl
F (ϕ(ξ)) = σk
F (ϕ(ξ)) = σl
Lemma 3.7, Lemma 3.8 and Theorem 2.7 imply the following
Theorem 3.9. Let E and F be graphs. Consider the statements
a) FE 1 ⋉ ∂E and FF 1 ⋉ ∂F are isomorphic as topological groupoids,
b) GE and GF are isomorphic as topological groupoids,
c) there exists an isomorphism Φ : C ∗(E)
d) FE 1 y ∂E and FF 1 y ∂F are continuously orbit equivalent,
e) E and F are orbit equivalent.
∼=−→ C ∗(F ) with Φ(C ∗(EE )) = C ∗(EF ),
We always have a) ⇔ b), a) ⇒ c) as well as a) ⇒ d) ⇒ e). If E and F satisfy condition (L),
then all these statements are equivalent.
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
11
This generalizes [25, Theorem 2.3] and gives an alternative proof for the corresponding parts
of [4, Theorem 5.1]. Note that in [4], it is proven that a), b) and c) are always equivalent.
3.3. Subsemigroups of groups. Let us discuss C*-algebras attached to subsemigroups of
groups. Given a subsemigroup P of a group G (or any left cancellative semigroup), the left
reduced semigroup C*-algebra C ∗
λ(P ) of P is defined as the sub-C*-algebra of L(ℓ2P ) generated
by the left multiplication operators Vp : ℓ2P → ℓ2P, δx 7→ δpx. Here δx is the delta function in
x ∈ P , and these δx form the canonical orthonormal basis of ℓ2P . The canonical commutative
subalgebra Dλ(P ) is given by Dλ(P ) = C ∗
λ(P ) ∩ ℓ∞(P ). Here is an alternative description
of Dλ(P ): Let IV be the inverse semigroup of partial isometries on ℓ2P generated by Vp,
p ∈ P , i.e., IV =(cid:8)V ∗
pnVqn: n ∈ N, pi, qi ∈ P(cid:9). We can define a partial homomorphism
σ : IV \ {0} → G, V ∗
n qn. To see that σ is well-defined, note that
every V ∈ IV has the property that there exists g ∈ G such that for every x ∈ P , either
V δx = 0 or V δx = δgx. And σ is defined in such a way that σ(V ) = g. Now the closed linear
span span(σ−1(e)) of σ−1(e), where e is the identity of G, coincides with Dλ(P ). All this is
explained in [19, 20] (see [19, Remark 3.12] for the alternative description of Dλ(P )).
p1Vq1 · · · V ∗
p1Vq1 · · · V ∗
pnVqn 7→ p−1
1 q1 · · · p−1
Let us now describe the canonical partial action G y Dλ(P ). We first describe the dual action
α∗. For g ∈ G, let Dg−1 := span({V ∗V : V ∈ IV \ {0} , σ(V ) = g}). Dg−1 is an ideal of Dλ(P ),
and it follows from the alternative description of Dλ(P ) that De = Dλ(P ). We then define α∗
g as
g : Dg−1 → Dg, V ∗V → V V ∗ for V ∈ IV \ {0} with σ(V ) = g. This is well-defined: If we view
α∗
ℓ2P as a subspace ℓ2G and let λ be the left regular representation of G, then every V ∈ IV \ {0}
with σ(V ) = g satisfies V = λgV ∗V . Therefore, V V ∗ = λgV ∗V λ∗
g is just
conjugation with the unitary λg. This also explains why α∗
g is an isomorphism. Of course, we can
also describe the dual action α. Set ΩP := Spec (Dλ(P )) and for every g ∈ G, let Ug−1 := [Dg−1.
It is easy to see that Ug−1 = {χ ∈ ΩP : χ(V ∗V ) = 1 f or some V ∈ IV \ {0} with σ(V ) = g}.
We then define αg by setting αg(χ) := χ ◦ α∗
g. This shows that α∗
g−1.
Proposition 3.10. There is a canonical isomorphism C ∗
Vp 7→ VpV ∗
defined above, and ug denote the canonical partial isometries in Dλ(P ) ⋊r G.
λ(P ) ∼= Dλ(P ) ⋊r G determined by
p up. Here we form the partial crossed product for the partial action G y Dλ(P )
Proof. Our strategy is to describe both C ∗
algebras of Fell bundles, and then to identify the underlying Fell bundles.
λ(P ) and Dλ(P ) ⋊r G as reduced (cross sectional)
Let us start with C ∗
λ(P ). We have already defined IV and σ. Now we set Bg := span(σ−1(g))
for every g ∈ G. We want to see that (Bg)g∈G is a grading for C ∗
λ(P ), in the sense of [8,
Definition 3.1]. Conditions (i) and (ii) are obviously satisfied. For (iii), we use the faithful
conditional expectation E : C ∗
λ(P ) ։ Dλ(P ) = Be from [19, § 3.1]. Given a finite sum x =
gxh,
gxg (here we used that EBg = 0 if g 6= e). This implies that
xg = 0 for all g. Therefore, the subspaces Bg are independent. It is clear that the linear span of
all the Bg is dense in C ∗
λ(P ). This proves (iii). If we let B be the Fell bundle given by (Bg)g∈G,
then [8, Proposition 3.7] implies C ∗
λ(P ) ։ Dλ(P ) = Be is a faithful
conditional expectation satisfying EBe = idBe and EBg = 0 if g 6= e.
λ(P ) of elements xg ∈ Bg such that x = 0, we conclude that 0 = x∗x =Pg,h x∗
Pg xg ∈ C ∗
and hence 0 = E(x∗x) = Pg x∗
r (B) because E : C ∗
λ(P ) ∼= C ∗
12
XIN LI
r (B′).
g)g∈G satisfy (i), (ii) and (iii) in [8, Definition 3.1]. Moreover, B′
Let us also describe Dλ(P ) ⋊r G as a reduced algebra of a Fell bundle. We denote by Wg the
partial isometry in Dλ(P ) ⋊r G corresponding to g ∈ G, and we set B′
g := DgWg. Recall that
we defined Dg−1 = span({V ∗V : V ∈ IV \ {0} , σ(V ) = g}) earlier on. It is easy to check that
(B′
e = De = Dλ(P ), and it
follows immediately from the construction of the reduced partial crossed product that there is
a faithful conditional expectation Dλ(P ) ⋊r G ։ Dλ(P ) = B′
e and 0 on
B′
g)g∈G, then [8, Proposition 3.7]
implies Dλ(P ) ⋊r G ∼= C ∗
To identify C ∗
g for g 6= e. Hence if we let B′ be the Fell bundle given by (B′
λ(P ) and Dλ(P )⋊r G, it now remains to identify B with B′. We claim that the map
i Wg is well-defined
span({V : σ(V ) = g}) → span({V V ∗Wg: σ(V ) = g}),Pi αiVi 7→Pi αiViV ∗
All we have to show is that our map is isometric. We have kPi αiVik2 =(cid:13)(cid:13)(cid:13)Pi,j αiαjViV ∗
j(cid:13)(cid:13)(cid:13)Dλ(P )
and kPi αiViV ∗
j . Hence, indeed, kPi αiVik2 = kPi αiViV ∗
i Wgk2 = (cid:13)(cid:13)(cid:13)Pi,j αiαjViV ∗
and extends to an isometric isomorphism Bg → B′
j = λg−1VjV ∗
j ,
i Wgk2,
we have ViV ∗
and we are done.
j(cid:13)(cid:13)(cid:13)Dλ(P )
e which is identity on B′
. Since Vi = ViV ∗
i VjV ∗
i VjV ∗
g, for all g ∈ G.
i λgλg−1VjV ∗
i λg and V ∗
j = ViV ∗
j = ViV ∗
(cid:3)
All in all, we have proven that C ∗
r (B′) ∼= Dλ(P ) ⋊r G.
Remark 3.11. A straightforward computation shows that actually, VpV ∗
Thus the isomorphism in the Proposition 3.10 is given by Vp 7→ up for all p ∈ P .
λ(P ) ∼= C ∗
r (B) ∼= C ∗
p up = up for all p ∈ P .
Our next goal is to write C ∗
λ(P ) as a quotient of a C*-algebra of an inverse semigroup in a
canonical way. Let S := Il(P ) be the inverse semigroup of partial bijections of P generated
∼=−→ P, x 7→ px (p ∈ P ). The semilattice of idempotents E of S is given by the
by p : P
set of constructible ideals J = (cid:8)p−1
is canonically isomorphic to the inverse semigroup IV constructed above. The homomorphism
IV \ {0} → G yields an idempotent pure partial homomorphism from S to G such that for
every s ∈ S×, s(x) = σ(s) · x if x ∈ dom(s).
n qnP : pi, qi ∈ P(cid:9) ∪ {∅}. It is easy to see that S
1 q1 · · · p−1
As explained in [31, Corollary 3.2.13], the isometry ℓ2P → ℓ2S×, δp 7→ δp induces surjective
homomorphisms C ∗(E) ։ Dλ(P ) and C ∗
λ(P ). The first surjection allows us to view
λ(S) ։ C ∗
some 1 ≤ i ≤ n.
The following lemma is easy to check:
i=1 Xi, χ(X) = 1 implies χ(Xi) = 1 for
ΩP = Spec (Dλ(P )) as a closed subspace of bE. More precisely, χ ∈ bE lies in ΩP if for all
constructible ideals X, X1, . . . , Xn of P with X = Sn
Lemma 3.12. ΩP is a G-invariant subspace with respect to the canonical G-action on bE. The
obvious map C(bE) ⋊r G ։ C(ΩP ) ⋊r G under the identifications given by Proposition 3.1 and
corresponding partial system G y ΩP coincides with α, the dual action of α∗ : G y Dλ(P )
constructed before Proposition 3.10.
λ(P ) from [31, Corollary 3.2.13] corresponds to the
Moreover, the surjection C ∗
Proposition 3.10.
λ(S) ։ C ∗
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
13
Here is a sufficient condition for topological freeness of G y ΩP :
Proposition 3.13. If P contains the identity e ∈ G, and if the group of units P ∗ in P is
trivial, i.e., P ∗ = {e}, then G y ΩP is topologically free.
i=1 Xi, and then χp ∈ U(X; X1, . . . , Xn).
It turns out that {χp: p ∈ P } is dense in ΩP . Basic open sets in ΩP
Obviously, χp ∈ ΩP .
are of the form U(X; X1, . . . , Xn) = {χ ∈ ΩP : χ(X) = 1, χ(Xi) = 0 for all 1 ≤ i ≤ n}. Here
i=1 Xi.
Thus, for a non-empty basic open set U(X; X1, . . . , Xn), we may choose p ∈ X such that
Proof. For p ∈ P , let χp ∈ bE be defined by χp(X) = 1 if and only if p ∈ X, for X ∈ J .
X, X1, . . . , Xn are constructible ideals of P . Clearly, U(X; X1, . . . , Xn) is empty if X =Sn
p /∈Sn
Let p ∈ P and g ∈ G satisfy g.χp = χp. This equality only makes sense if χp ∈ Ug−1, i.e.,
there exists Y ∈ J with Y ⊆ g−1 · P and p ∈ Y . Then g.χp(X) = χp(g−1 · (X ∩ g · Y )). So
g.χp(X) = 1 if and only if p ∈ g−1 · (X ∩ g · Y ) if and only if gp ∈ X ∩ g · Y if and only if gp ∈ X,
while χp(X) = 1 if and only if p ∈ P . Hence g.χp = χp implies that for every X ∈ J , gp ∈ X if
and only if p ∈ X. For X = pP , we obtain gp ∈ pP , and for X = gpP , we get p ∈ gpP . Hence
there exist x, y ∈ P with gp = px and p = gpy. So p = gpy = pxy and gp = px = gpyx. Thus
xy = yx = e. Hence x, y ∈ P ∗. Since P ∗ = {e} by assumption, we must have x = y = e, and
hence gp = p. This implies g = e. In other words, for every e 6= g ∈ G, g.χp 6= χp for all p ∈ P
such that χp ∈ Ug−1. Hence it follows that {χ ∈ Ug−1: g.χ 6= χ} contains {χp ∈ Ug−1: p ∈ P },
and the latter set is dense in Ug−1 as {χp: p ∈ P } is dense in ΩP .
(cid:3)
Coming back to the comparison of C ∗
the following are equivalent:
• the canonical map C ∗
λ(P ) is injective,
• the canonical map C ∗(E) ։ Dλ(P ) is injective,
• J is independent.
λ(S) and C ∗
λ(S) ։ C ∗
λ(P ), it was shown in [31, Theorem 3.2.14] that
Recall that J is called independent if for all X, X1, . . . , Xn ∈ J , X =Sn
for some 1 ≤ i ≤ n.
i=1 Xi implies X = Xi
In view of [19, 20], it makes sense to view the full C*-algebra C ∗(S) of the inverse semigroup
S = Il(P ) as the full semigroup C*-algebra of P . Thus we set C ∗(P ) := C ∗(S), and let
λ : C ∗(P ) ։ C ∗
λ(P ). Because of our
descriptions as transformation groupoids (see Proposition 3.1), the following generalizations of
[20, Theorem 6.1] are immediate.
λ(P ) be the composite C ∗(P ) = C ∗(S) ։ C ∗
λ(S) ։ C ∗
Theorem 3.14. If G ⋉bE or G ⋉ ΩP is amenable, then λ : C ∗(P ) ։ C ∗
if and only if P is independent.
Theorem 3.15. Consider the following statements:
λ(P ) is an isomorphism
(i) C ∗(P ) is nuclear,
(ii) C ∗
λ(P ) is nuclear,
(iii) G ⋉ ΩP is amenable,
(iv) λ : C ∗(P ) ։ C ∗
λ(P ) is an isomorphism.
14
XIN LI
We always have (i) ⇒ (ii) ⇔ (iii). If P is independent, then (iii) ⇒ (iv) and (iii) ⇒ (i), so
that (i), (ii) and (iii) are equivalent.
Corollary 3.16. If P is a subsemigroup of an amenable group G, then (i), (ii) and (iii) from
Theorem 3.15 hold, and (iv) is true if and only if P is independent.
Proof. By Proposition 3.1, both C ∗(P ) and C ∗
λ(P ) can be written as (full or reduced) partial
crossed products by G, and hence as (full or reduced) C*-algebras of Fell bundles over G.
Therefore, by [12, Theorem 20.7 and Theorem 25.10], C ∗(P ) and C ∗
λ(P ) are nuclear. (iii) holds
because we know that (ii) ⇔ (iii). Finally, our claim about (iv) follows from Theorem 3.14. (cid:3)
Note that the statement in the last corollary was also obtained in [31].
Lemma 3.17. We have ∂ΩP ⊆ ΩP .
Now set ∂ΩP := ∂bE, where E is the semilattice of idempotents of S = Il(P ).
Proof. Let X, X1, . . . , Xn ∈ J satisfy X =Sn
χ(X) = 0. This shows bEmax ⊆ ΩP . As ΩP is closed, we conclude that ∂bE ⊆ ΩP .
i, 1 ≤ i ≤ n be as above. Then for X ′ =Tn
i=1 Xi. Then for χ ∈ bEmax, χ(Xi) = 0 implies that
i = ∅. Thus if χ(Xi) = 0 for all 1 ≤ i ≤ n,
i, χ(X ′) = 1 and X ∩ X ′ = ∅. Thus
(cid:3)
there exists X ′
then let X ′
i=1 X ′
i ∈ J with χ(X ′
i) = 1 and Xi ∩ X ′
As ∂ΩP is G-invariant, the following makes sense.
Definition 3.18. The boundary quotient of C ∗
λ(P ) is given by ∂C ∗
λ(P ) := C(∂ΩP ) ⋊r G.
By construction and because of Proposition 3.10, there is a canonical projection C ∗
∂C ∗
∂C ∗
λ(P ). Morever, if P is a subsemigroup of an exact group G, then ∂C ∗
λ(S) ∼= C(∂ΩP ) ⋊r G. Let us now generalize the results in [20, § 7.3].
λ(P ) ։
λ(P ) is isomorphic to
we know that p.χ ∈ C. We have p.χ(pP ) = χ(P ) = 1, so that p.χ(X) = 1 as p ∈ X implies
pP ⊆ X (X is a right ideal). Set χX := p.χ. Consider the net (χX)X indexed by X ∈ J with
χ(X) = 1, ordered by inclusion. Passing to a convergent subnet if necessary, we may assume
Lemma 3.19. ∂bE = ∂ΩP is the minimal non-empty closed G-invariant subspace of bE.
Proof. Let C ⊆ bE be non-empty, closed and G-invariant. Let χ ∈ bEmax be arbitrary, and
choose X ∈ J with χ(X) = 1. Choose p ∈ X and χ ∈ C. As Up−1 = bE, we can form p.χ, and
that limX χX exists. But it is clear because of χ ∈ bEmax that limX χX = χ. As χX ∈ C for all
X, we deduce that χ ∈ C. Thus bEmax ⊆ C, and hence ∂bE ⊆ C.
In particular, ∂ΩP is the minimal non-empty closed G-invariant subspace of ΩP . Another
immediate consequence is
(cid:3)
Corollary 3.20. The transformation groupoid G ⋉ ∂ΩP is minimal.
To discuss topological freeness of G y ∂ΩP , let
G0 =(cid:8)g ∈ G: X ∩ g · P 6= ∅ 6= X ∩ g−1 · P for all ∅ 6= X ∈ J(cid:9) ,
as in [20, § 7.3]. Clearly, G0 = {g ∈ G: pP ∩ g · P 6= ∅ 6= pP ∩ g−1 · P for all p ∈ P }. Further-
more, [20, Lemma 7.19] shows that G0 is a subgroup of G.
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
15
Proposition 3.21. G y ∂ΩP is topologically free if and only if G0 y ∂ΩP is topologically free.
Proof. "⇒" is clear. For "⇐", assume that G0 y ∂ΩP is topologically free, and suppose that
G y ∂ΩP is not topologically free, i.e., there exists g ∈ G and U ⊆ Ug−1 ∩ ∂ΩP such that
g.χ = χ for all χ ∈ U. As bEmax = ∂ΩP , we can find χ ∈ Ug−1 ∩ bEmax with g.χ = χ.
For every X ∈ J with χ(X) = 1, choose x ∈ X and ψX ∈ bEmax with ψX (xP ) = 1, so that
ψX (X) = 1. Consider the net (ψX )X indexed by X ∈ J with χ(X) = 1, ordered by inclusion.
Passing to a convergent subnet if necessary, we may assume that limX ψX = χ. As U is open,
we may assume that ψX ∈ U for all X. Then ψX(xP ) = 1 implies that ψX ∈ Ux ∩ U.
Hence for sufficiently small X ∈ J with χ(X) = 1, there exists x ∈ X such that x−1.(Ux ∩ U)
is a non-empty open subset of ∂ΩP . We conclude that (x−1gx).ψ = ψ for all ψ ∈ x−1.(Ux ∩ U).
This implies that x−1gx /∈ G0 as G0 y ∂ΩP is topologically free. So there exists p ∈ P with
If
pP ∩ x−1gx · P = ∅, then xpP ∩ gx · P = ∅, so that xpP ∩ g−1xp · P = ∅. Hence g.χX 6= χX if
χX ∈ Ug−1. If pP ∩ x−1g−1x · P = ∅, then xpP ∩ g−1x · P = ∅, so that xpP ∩ g−1xp · P = ∅.
Again, g.χX 6= χX if χX ∈ Ug−1.
pP ∩ x−1gx · P = ∅ or pP ∩ x−1g−1x · P = ∅. Let χX ∈ bEmax satisfy χX(xpP ) = 1.
For every sufficiently small X ∈ J with χ(X) = 1, we can find x ∈ X and χX as above. Hence
we can consider the net (χX )X as above, and assume after passing to a convergent subnet that
limX χX = χ. As χ ∈ U ⊆ Ug−1 ∩ ∂ΩP , it follows that χX ∈ U ⊆ Ug−1 ∩ ∂ΩP for sufficiently
small X. So we obtain g.χX 6= χX, although g acts trivially on U. This is a contradiction. (cid:3)
Corollary 3.22. If G0 y ∂ΩP is topologically free, then ∂C ∗
λ(P ) is simple.
Proof. This follows from Lemma 3.19, Proposition 3.21 and [33, Chapter II, Proposition 4.6].
(cid:3)
Theorem 3.23. Let P = hS, Ri+ be a monoid given by a positive r-complete presentation
(S, R) in the sense of [7]. Assume that for all u ∈ S, there is v ∈ S such that R does not
contain any relation of the form u · · · = v · · · . Also, suppose that P embeds into a group G such
that (G, P ) is quasi-lattice ordered in the sense of [30]. Then G0 = {e} and ∂C ∗
λ(P ) is simple.
Proof. In view of Corollary 3.22, it suffices to prove G0 = {e}. Let g ∈ G0. Assume that
gP ∩ P 6= P . Then g ∈ G0 implies that this intersection is not empty. Hence, we must
have gP ∩ P = pP for some p ∈ P because (G, P ) is quasi-lattice ordered. As p 6= e, there
exists u ∈ Σ with pP ⊆ uP . By assumption, there exists v ∈ Σ such that no relation in R
is of the form u · · · = v · · · . By r-completeness, uP ∩ vP = ∅ (see [7, Proposition 3.3]), so
that gP ∩ vP = ∅. This contradicts g ∈ G0. Hence, we must have gP ∩ P = P , and similarly,
g−1P ∩P = P . These two equalities imply g ∈ P ∗. But P ∗ = {e} because (G, P ) is quasi-lattice
ordered. Thus g = e.
(cid:3)
Remark 3.24. By going over to the opposite semigroup, we obtain analogues results for the
right versions C ∗
ρ (P ) and ∂C ∗
ρ (P ).
Examples 3.25. Theorem 3.23 implies that for every graph-irreducible right-angled Artin
monoid A+
Γ ) is simple. Also, for the Thompson monoid F + =
hx0, x1, . . . xnxk = xkxn+1 for k < ni+, we get that ∂C ∗
Γ in the sense of [5], ∂C ∗
ρ (F +) is simple.
λ(A+
16
XIN LI
Corollary 3.22 and [21, Corollary 5.10] imply that for every countable Krull ring R with
P(R)inf 6= ∅ or P(R)fin infinite (see [21] for details), ∂C ∗
λ(R ⋊ R×) is simple.
4. Purely infinite groupoids
Our goal is to exhibit examples of purely infinite groupoids. More precisely, we study groupoids
attached to graphs, groupoids corresponding to boundary quotients of semigroup C*-algebras,
and groupoids underlying semigroup C*-algebras of ax + b-semigroups.
The following lemma is easy to see, and will be used several times:
i=1 Ui of basic open sets Ui of the form
U(e; e1, . . . , en) =nχ ∈ bE: χ(e) = 1, χ(e1) = . . . = χ(en) = 0o , e, e1, . . . , en ∈ E.
Lemma 4.1. Let E be a semilattice. Every compact open subset A ⊆ bE can be written as a
disjoint union A =Fm
This lemma will be helpful because given a partial system G y bE (or G y X for any G-
invariant subspace X ⊆ bE), and we want to show that every compact open subset is (G, CO)-
of the form U(e; e1, . . . , en). Here CO is the set of compact open subsets of bE or X.
paradoxical in the sense of [16, Definition 4.3], then it suffices to show this for basic open sets
4.1. Graphs. Let us first recall a necessary and sufficient condition from [17], in terms of
graphs, for pure infiniteness of graph C*-algebras. Let E = (E 0, E 1, r, s) be a graph, and we
use the same notation as in § 3.2. Note that since our notation differs from the one in [17] (see
Remark 3.5), we have to reverse all the arrows in the condition for pure infiniteness.
For v, w ∈ E 0, we write w ← v if there exists a path µ ∈ E ∗ from v to w,
i.e., with
r(µ) = w and s(µ) = v, and we write w 6← v if there is no such path. For v ∈ E 0, let
Ω(v) = {w ∈ E 0: w 6= v, w 6← v}. v ∈ E 0 is called a breaking vertex if r−1(v) = ∞ and
0 < r−1(v) \ s−1(Ω(v)) < ∞.
Moreover, we call µ ∈ E ∗ a loop if l(µ) ≥ 1 and r(µ) = s(µ). We say that E satisfies condition
(K) if for all v ∈ E 0, whenever there exists a loop µ with r(µ) = v = s(µ), there exists another
loop µ′ with r(µ′) = v = s(µ′) and µ · µ′ = 0 in EE .
Furthermore, a subset M ⊆ E 0 is called a maximal tail if
• for every v ∈ E 0, whenever there is w ∈ M with v ← w, then v ∈ M;
• for every v ∈ M with 0 < r−1(v) < ∞, there exists η ∈ E 1 with r(η) = v and s(η) ∈ M;
• for all v, w ∈ M, there exist y ∈ M with v ← y and w ← y.
We say that v ∈ E 0 connects to a loop if there is a loop µ with r(µ) = w = s(µ) with v ← w.
[17, Theorem 2.3] says that C ∗(E) is purely infinite if and only if E satisfies the following
condition:
(PI) There exist no breaking vertices in E, E satisfies condition (K), and every vertex in each
maximal tail M connects to a loop in M.
Lemma 4.2. Suppose that E satisfies condition (PI). Let v ∈ E 0 satisfy r−1(v) = ∞. Then
there exist infinitely many loops µ1, µ2, . . . ∈ E ∗ with r(µi) = v = s(µi) of the form µi = ζiηi for
pairwise distinct ζi ∈ E 1.
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
17
Proof. Since v is not a breaking vertex, we must have r−1(v) \ s−1(Ω(v)) ∈ {0, ∞}.
If r−1(v) \ s−1(Ω(v)) = ∞, then there exist infinitely many pairwise distinct ζi ∈ E 1 with
r(ζi) = v, and there is ηi ∈ E ∗ with r(ηi) = s(ζi), s(ηi) = v. Then ζiηi are the required loops.
If r−1(v) \ s−1(Ω(v)) = 0, then consider the subset M := {w ∈ E 0: w ← v}. It is easy to
see that M is a maximal tail, with v ∈ M. By condition (PI), v has to connect to a loop µ in
M. This means that there is w ∈ E 0 such that w lies on µ and v ← w. But then, w has to
lie in M, so that w ← v by definition of M; hence v lies on a loop v ← w ← v. Let ζ1 · · · ζn
be such a loop. Then ζ1 ∈ r−1(v), but ζ1 /∈ s−1(Ω(v)) as s(ζ1) ← v, for instance via ζ2 · · · ζn.
Hence 0 < r−1(v) \ s−1(Ω(v)), which is a contradiction.
(cid:3)
Let F ⋉ ∂E be the transformation groupoid attached to F y ∂E as in § 3.2.
Theorem 4.3. If E satisfies condition (PI), then F ⋉ ∂E is purely infinite.
Proof. By Lemma 4.1, it suffices to show that basic open sets of the form U(µ; µ1, . . . , µn) are
(F, CO)-paradoxical. By passing even further to finite unions if necessary, it suffices to treat
the basic open sets U(µ; µ1, . . . , µn) with r−1(s(µ)) = ∞ (and µ1, . . . , µn ≤ µ), or basic open
sets of the form U(µ).
Let us consider the first case. Set U := U(µ; µ1, . . . , µn). Let v = s(µ). Since r−1(v) = ∞,
Lemma 4.2 tells us that there are infinitely many loops µ1, µ2, . . . ∈ E ∗ with r(µi) = v = s(µi)
of the form µi = ζiηi for pairwise distinct ζi ∈ E 1. Hence we can find, among the ζi, edges
ζ, ζ ′ ∈ E 1 with r(ζ) = r(ζ ′) = v and µζ · µ1 = . . . = µζ · µn = 0 = µζ ′ · µ1 = . . . = µζ ′ · µn = 0.
Therefore, for χ ∈ ∂E, χ(µζ) = 1 implies χ(µi) = 0 for all 1 ≤ i ≤ n, so that χ ∈ U, and
similarly for ζ ′. Moreover, µζ · µζ ′ = 0 yields that χ(µζ) = 1 implies χ(µζ ′) = 0 and vice versa.
Setting g = µζµ−1 ∈ F, h = µζ ′µ−1 ∈ F, we obtain g.U ⊆ U, h.U ⊆ U and g.U ∩ h.U = ∅.
In the second case, let v = s(µ). By Lemma 4.2 and condition (PI), we can find α =
α1 · · · αn ∈ E ∗ with v = r(α) and r−1(αi) < ∞ (1 ≤ i ≤ n) such that there is a loop β ∈ E ∗
with r(β) = s(α) = s(β). Therefore, by passing to finite unions if necessary, we may assume
by condition (K) that there exist two loops ζ and ζ ′ with r(ζ) = v = s(ζ), r(ζ ′) = v = s(ζ ′)
and ζ · ζ ′ = 0. Again, setting g = µζµ−1 ∈ F, h = µζ ′µ−1 ∈ F, we obtain g.U ⊆ U, h.U ⊆ U
and g.U ∩ h.U = ∅.
(cid:3)
It is clear that condition (PI) implies that F y ∂E is residually topologically free, in the
[16, Theorem 4.4] tells us that if F y ∂E is
sense of [16, Definition 3.4 (ii)]. Moreover,
residually topologically free and if every compact open subset of ∂E is (F, CO)-paradoxical,
then C ∗(E) ∼= C(∂E) ⋊r F must be purely infinite (using that F is free and [12, Theorem 22.9]).
All in all, we obtain from Theorem 4.3:
Corollary 4.4. For a graph E, F y ∂E is residually topologically free and purely infinite if
and only if C ∗(E) is purely infinite.
4.2. Boundary quotients of semigroup C*-algebras. Let P be a subsemigroup of a group
G. Let E be the semilattice of idempotents of S = Il(P ). We write Ω for ΩP , Ωmax for bEmax
and ∂Ω for ∂ΩP = ∂bE.
Theorem 4.5. If there exist p, q ∈ P with pP ∩ qP = ∅, then G ⋉ ∂Ω is purely infinite.
Proof. Let U = {ψ ∈ ∂Ω: ψ(X) = 1, ψ(X1) = . . . = ψ(Xn) = 0} be a basic open subset for some
X, X1, . . . , Xn ∈ J . By Lemma 4.1, it suffices to show that U is (G, CO)-paradoxical. Since
18
XIN LI
that there exists Yi ∈ J with Xi ∩ Yi = ∅ and χ(Yi) = 1. Let Y := X ∩Tn
Ωmax is dense in ∂Ω, there exists χ ∈ Ωmax with χ ∈ U. As χ lies in Ωmax, χ(Xi) = 0 implies
i=1 Yi. Certainly,
Y 6= ∅ as χ(Y ) = 1. Moreover, for every ψ ∈ ∂Ω, ψ(Y ) = 1 implies ψ ∈ U. Now choose x ∈ Y .
By assumption, we can find p, q ∈ P with pP ∩ qP = ∅. For ψ ∈ ∂Ω, xp.ψ(xpP ) = ψ(P ) = 1.
Similarly, for all ψ ∈ ∂Ω, xq.ψ(xqP ) = 1. Thus xp.U ⊆ xp.∂Ω ⊆ U, xq.U ⊆ xq.∂Ω ⊆ U and
(xp.U) ∩ (xq.U) ⊆ (xp.∂Ω) ∩ (xq.∂Ω) = ∅ since xpP ∩ xqP = ∅.
(cid:3)
This strengthens and explains [20, Corollary 7.23].
4.3. Groupoids from ax + b-semigroups. Let R be an integral domain with quotient field
K. Consider the ax + b-semigroup R ⋊ R×, viewed as a subsemigroup of K ⋊ K ×. Let E be
the semilattice of idempotents in S = Il(R ⋊ R×).
Theorem 4.6. Let R be an integral domain with Jac(R) = (0) and R× 6= R∗. Then (K ⋊
K ×) ⋉ bE is purely infinite.
Proof. Let U be the basic open subset
U(x + I; x1 + I1, . . . , xn + In)
= nχ ∈ bE: χ((x + I) × I ×) = 1; χ((x1 + I1) × I ×
Set J := Tn
where Ii ⊆ I ( R. By Lemma 4.1, it suffices to show that U is (K ⋊ K ×, CO)-paradoxical.
i=1 Ii. Obviously, J 6= (0). Therefore 1 + J * R∗. Otherwise, by [3, Proposi-
tion 1.9], J ⊆ Jac(R) = (0), which is a contradiction. So we can choose a ∈ (1 + J) \ R∗, a 6= 0,
as R× 6= R∗. Then J * aR. Otherwise, there would exist r ∈ R with a − 1 = ar, so that
a(1 − r) = 1 contradicting a /∈ R∗. Choose b1 ∈ J, δ ∈ J \ aR and set b2 := b1 + δ.
1 ) = . . . = χ((x + I) × I ×) = 0o ,
For k = 1, 2, (bk, a).U = U(bk +ax+aI; bk +ax1+aI1, . . . , bk +axn +aIn). Since bk +ax+aI ⊆
ax + aI + J ⊆ a(x + I) + J ⊆ (1 + J)(x + I) + J ⊆ x + I, every χ ∈ (bk, a).U satisfies
χ((x + I) × I ×) = 1. Moreover, we have (xi + Ii) ∩ (bk + ax + aI) = bk + (xi + Ii) ∩ (ax + aI) =
bk + (axi + Ii) ∩ (axi + aI) since xi − axi = (1 − a)xi ∈ J.
We claim that Ii ∩ aI = aIi. "⊇" is clear. If r ∈ Ii ∩ aI, then r = as for some s ∈ I, so that
s = (1 − a + a)s = (1 − a)s + as ∈ J + Ii ⊆ Ii. This proves "⊆".
Hence (xi + Ii) ∩ (bk + ax + aI) = bk + axi + aIi. Therefore, every χ ∈ (bk, a).U satisfies
χ((xi + Ii) × I ×
i ) = 0. This shows that for k = 1, 2, (bk, a).U ⊆ U.
Also, (b1+ax+aI)∩(b2+ax+aI) = ∅ as b2−b1 = δ /∈ aR. Hence (b1, a).U ∩(b2, a).U = ∅. (cid:3)
This strengthens and explains [21, Theorem 1.3].
4.4. Almost finite groupoids. Apart from purely infinite groupoids, Matui also introduced
almost finite ones in [26]. We would like to end with the following obervation concerning the
relation between almost finiteness and amenability:
Proposition 4.7. Let G be a group acting on a compact space X. Assume that there exists
x ∈ X with trivial stabilizer group, i.e., Gx = {g ∈ G: g.x = x} = {e}. If the transformation
groupoid G ⋉ X is almost finite, then G is amenable.
PARTIAL TRANSFORMATION GROUPOIDS ATTACHED TO GRAPHS AND SEMIGROUPS
19
Proof. Let E ⊆ G be a finite subset with E = E−1. For every ε > 0, we have to find a finite
subset F ⊆ G such that for every s ∈ E, sF △F / F < ε.
Set C := E × X ⊆ G ⋉ X. Let x ∈ X be as above,
i.e., {g ∈ G: g.x = x} = {e}.
Since G ⋉ X is almost finite, we can find an elementary subgroupoid K of G ⋉ X such that
CKx \ Kx / Kx < ε
2. Let F = {g ∈ G: (g, x) ∈ K}. F is contained in the image of K under
the canonical projection G ⋉ X → G, (g, x) 7→ g. Therefore, F is a finite subset of G. We have
Kx = {(g, x): g ∈ F } and CKx \ Kx = {(g, x): g ∈ EF \ F } = Ss∈E {(g, x): g ∈ sF \ F }.
Therefore, ε
F \ (F ∩ sF ) = s(s−1F \ (s−1F ∩ F ) = s(s−1F \ F ). Hence F \ sF = s−1F \ F < ε
our computation above. (Note that by assumption, E = E−1.)
2 > CKx \ Kx / Kx ≥ sF \ F / F for all s ∈ E. For s ∈ E, F \ sF =
2 F by
Therefore, for every s ∈ E, sF △F / F = sF \ F / F + F \ sF / F < ε.
(cid:3)
References
[1] F. Abadie, On partial actions and groupoids, Proc. Am. Math. Soc. 132 (2003), 1037 -- 1047.
[2] P. Ara and R. Exel, Dynamical systems associated to separated graphs, graph algebras, and paradoxical
decompositions, Adv. in Math. 252 (2014), 748 -- 804.
[3] M.F. Atiyah and I.G. MacDonald, Introduction to Commutative Algebra, Addison-Wesley Publishing
Company, Reading, Massachusetts, 1969.
[4] N. Brownlowe, T. Carlsen and M. Whittaker, Graph algebras and orbit equivalence, to appear in
Ergod. Th. Dyn. Syst.
[5] J. Crisp and M. Laca, Boundary quotients and ideals of Toeplitz algebras of Artin groups, J. Funct. Anal.
242 (2007), 127 -- 156.
[6] J. Cuntz, S. Echterhoff and X. Li, On the K-theory of crossed products by automorphic semigroup
actions, Quart. J. Math. 64 (2013), 747 -- 784.
[7] P. Dehornoy, Complete positive group presentations, Journal of Algebra 268 (2003), 156 -- 197.
[8] R. Exel, Amenability for Fell bundles, J. Reine Angew. Math. 492 (1997), 41 -- 73.
[9] R. Exel, Inverse semigroups and combinatorial C*-algebras, Bull. Braz. Math. Soc. (N.S.) 39 (2008),
191 -- 313.
[10] R. Exel, Tight representations of semilattices and inverse semigroups, Semigroup Forum 79 (2009), 159 --
182.
[11] R. Exel, Reconstructing a totally disconnected groupoid from its ample semigroup, Proc. Am. Math. Soc.
138 (2010), 2991 -- 3001.
[12] R. Exel, Partial Dynamical Systems, Fell Bundles and Applications, available on R. Exel's homepage.
[13] R. Exel, D. Goncalves and C. Starling, The tiling C*-algebra viewed as a tight inverse semigroup
algebra, Semigroup Forum 84 (2012), 229 -- 240.
[14] R. Exel and M. Laca, Cuntz-Krieger algebras for infinite matrices, J. reine angew. Math. 512 (1999),
119 -- 172.
[15] R. Exel, M. Laca and J. Quigg, Partial dynamical systems and C*-algebras generated by partial isome-
tries, J. Op. Th. 47 (2002), 169 -- 186.
[16] T. Giordano, A. Sierakowski , Purely infinite partial crossed products, J. Funct. Anal. 266 (2014),
5733 -- 5764.
[17] J.H. Hoang and W. Syzmanski, Purely infinite Cuntz-Krieger algebras of directed graphs, Bull. London
Math. Soc. 35 (2003), 689 -- 696.
[18] M. Hochster, Subsemigroups of amenable groups, Proc. Amer. Math. Soc. 21 (1969), 363 -- 364.
[19] X. Li, Semigroup C*-algebras and amenability of semigroups, J. Funct. Anal. 262 (2012), 4302 -- 4340.
[20] X. Li, Nuclearity of semigroup C*-algebras and the connection to amenability, Adv. in Math. 244 (2013),
626 -- 662.
[21] X. Li, Semigroup C∗-algebras of ax + b-semigroups, Trans. Amer. Math. Soc. 368 (2016), 4417 -- 4437.
20
XIN LI
[22] X. Li, Continuous orbit equivalence rigidity, preprint, arXiv:1503.01704.
[23] X. Li and M. D. Norling, Independent resolutions for totally disconnected dynamical systems I: Algebraic
case, J. Alg. 424 (2015), 98 -- 125.
[24] X. Li and M.D. Norling, Independent resolutions for totally disconnected dynamical systems II: C*-
algebraic case, J. Op. Th. 75 (2016), 163 -- 193.
[25] K. Matsumoto and H. Matui, Continuous orbit equivalence of topological Markov shifts and Cuntz-
Krieger algebras, Kyoto J. Math. 54 (2014), 863 -- 877.
[26] H. Matui, Homology and topological full groups of ´etale groupoids on totally disconnected spaces, Proc.
Lond. Math. Soc. 104 (2012), 27 -- 56.
[27] H. Matui, Topological full groups of one-sided shifts of finite type, J. Reine Angew. Math. 705 (2015),
35 -- 84.
[28] K. McClanahan, K-theory for partial crossed products by discrete groups, J. Funct. Anal. 130 (1995),
77 -- 117.
[29] D. Milan and B. Steinberg, On inverse semigroup C*-algebras and crossed products, Groups Geom. Dyn.
8 (2014), 485 -- 512.
[30] A. Nica, C*-algebras generated by isometries and Wiener-Hopf operators, J. Operator Theory 27 (1992),
17 -- 52.
[31] M.D. Norling, Inverse semigroup C*-algebras associated with left cancellative semigroups, Proc. Edinb.
Math. Soc. (Series 2) 57 (2014), 533 -- 564.
[32] A.L.T. Paterson, Groupoids, Inverse Semigroups, and their Operator Algebras, Progress in mathematics
170, Birkhauser, Boston, 1999.
[33] J. Renault, A groupoid approach to C*-algebras, LNM 793, Springer, Berlin Heidelberg New York, 1980.
[34] J. Renault, Cartan subalgebras in C*-algebras, Irish Math. Soc. Bull. 61 (2008), 29 -- 63.
[35] J. Renault, C*-algebras and Dynamical Systems, Publicacoes matematicas, IMPA, 2009.
[36] M. Rørdam, A. Sierakowski , Purely infinite C*-algebras arising from crossed products, Ergod. Th. Dyn.
Syst. 32 (2012), 273 -- 293.
[37] S.B.G. Webster, The path space of a directed graph, Proc. Amer. Math. Soc. 142 (2014), 213 -- 225.
Xin Li, School of Mathematical Sciences, Queen Mary University of London, Mile End Road,
London E1 4NS, United Kingdom
E-mail address: [email protected]
|
1510.01862 | 1 | 1510 | 2015-10-07T08:51:33 | $q$-invariance of quantum quaternion spheres | [
"math.OA"
] | The $C^*$-algebra of continuous functions on the quantum quaternion sphere $H_q^{2n}$ can be identified with the quotient algebra $C(SP_q(2n)/SP_q(2n-2))$. In commutative case i.e. for $q=1$, the topological space $SP(2n)/SP(2n-2)$ is homeomorphic to the odd dimensional sphere $S^{4n-1}$. In this paper, we prove the noncommutative analogue of this result. Using homogeneous $C^*$-extension theory, we prove that the $C^*$-algebra $C(H_q^{2n})$ is isomorphic to the $C^*$-algebra $C(S_q^{4n-1})$. This further implies that for different values of $q \in [0,1)$, the $C^*$-algebras underlying the noncommutative space $H_q^{2n}$ are isomorphic. | math.OA | math |
q-invariance of quantum quaternion spheres
Bipul Saurabh
July 16, 2018
Abstract
The C ∗-algebra of continuous functions on the quantum quaternion sphere H 2n
can be identified with the quotient algebra C(SPq(2n)/SPq(2n − 2)). In commuta-
tive case i.e. for q = 1, the topological space SP (2n)/SP (2n − 2) is homeomorphic
to the odd dimensional sphere S4n−1. In this paper, we prove the noncommuta-
tive analogue of this result. Using homogeneous C ∗-extension theory, we prove
that the C ∗-algebra C(H 2n
). This fur-
ther implies that for different values of q ∈ [0, 1), the C ∗-algebras underlying the
noncommutative space H 2n
q ) is isomorphic to the C ∗-algebra C(S4n−1
q are isomorphic.
q
q
AMS Subject Classification No.: 58B34, 46L80, 19K33
Keywords. Homogeneous extension, Quantum double suspension.
1
Introduction
Quantization of Lie groups and their homogeneous spaces have played an important role
in linking the theory of compact quantum group with noncommutative geometry. Many
authors (see [16], [12], [3], [10]) have studied different aspects of the theory of quantum
homogeneous spaces. However, in these papers, main examples have been the quotient
spaces of the compact quantum group SUq(n). Neshveyev & Tuset ([9]) studied quantum
homogeneous spaces in a more general set up and gave a complete classification of the
irreducible representations of the C ∗-algebra C(Gq/Hq) where Gq is the q-deformation
of a simply connected semisimple compact Lie group and Hq is the q-deformation of a
closed Poisson-Lie subgroup H of G. Moreover, Neshveyev & Tuset ([9]) proved that
C(Gq/Hq) is KK-equivalent to the classical counterpart C(G/H). Quantum symplectic
group SPq(2n) and its homogeneous space C(SPq(2n)/SPq(2n − 2)) have been studied
1
by the author in [14] and K-groups of quotient space C(SPq(2n)/SPq(2n − 2)) with
explicit generators were obtained.
The C ∗-algebra C(H 2n
q ) of continuous functions on the quantum quaternion sphere is
defined as the universal C ∗-algebra given by a finite set of generators and relations (see
[14]). In [14], the isomorphism between the quotient algebra C(SPq(2n)/SPq(2n−2)) and
C(H 2n
q ) has been established. Now several questions arise about this noncommutative
space H 2n
q .
1. Is H 2n
q
topologically same as S4n−1
q
isomorphic?
, i.e. are the C ∗-algebras C(H 2n
q ) and C(S4n−1
q
)
2. Are the C ∗-algebras C(H 2n
3. Does the quantum quaternion sphere admit a good spectral triple equivariant
q ) isomorphic for different values of q?
under the SPq(2n)-group action?
We attempt the first two questions in this paper. In commutative case i.e. for q = 1, the
quotient space SP (2n)/SP (2n − 2) can be realized as the quaternion sphere H 2n. It can
be easily verified that the quaternion sphere H 2n is homeomorphic to the odd dimensional
sphere S4n−1. One can now expect the quotient algebra C(SPq(2n)/SPq(2n − 2)) or
q ) to be isomorphic to the C ∗-algebra underlying the
equivalently the C ∗-algebra C(H 2n
. In this paper, using homogeneous C ∗-extension
odd dimensional quantum sphere S4n−1
theory, we show that this is indeed the case.
q
The remarkable work done by L. G. Brown, R. G. Douglas and P. A. Fillmore ([2])
on extensions of commutative C ∗-algebras by compact operators has led many authors
to extend this theory further in order to provide a tool for analysing the structure of
C ∗-algebras. For a nuclear, separable C ∗-algebra A and a separable C ∗-algebra B, G. G.
Kasparov ([8]) constructed the group Ext(A, B) consisting "stable equivalence classes"
of C ∗-algebra extensions of the form
0 → B ⊗ K → E → A → 0
Here E will be called the middle C ∗-algebra. One of the important features of this con-
struction is that the group Ext(A, B) coincides with the group KK 1(A, B). Another
important aspect is that it does not demand much. It does not require the extensions
to be unital or essential. But at the same time, it does not provide much informa-
tions about the middle C ∗-algebras. Since elements of the group Ext(A, B) are stable
equivalence classes and not unitary equivalence classes of extensions, two elements in
the same class may have nonisomorphic middle C ∗-algebras. For a nuclear C ∗-algebra
A and a finite dimensional compact metric space Y (i.e. a closed subset of S n for
2
some n ∈ N), M. Pimsner, S. Popa and D. Voiculescu ([11]) constructed another group
ExtP P V (Y, A) consisting of unitary equivalence classes of unital homogeneous extensions
of A by C(Y ) ⊗ K. For y0 ∈ Y , the subgroup ExtP P V (Y, y0, A) consists of those ele-
ments of ExtP P V (Y, A) that split at y0. For a commutative C ∗-algebra A, the group
ExtP P V (Y, A) was computed by Schochet in [15]. Further Rosenberg & Schochet ([13])
showed that ExtP P V (Y, A+) = Ext(A, C(Y )) and ExtP P V (Y +, +, A+) = Ext(A, C(Y ))
where Y is a finite dimensional locally compact Hausdorff space, + is the point at infinity
and A+ is the C ∗-algebra obtained by adjoining unity to A.
q
q ) is isomorphic to C(S4n−1
To show that the C ∗-algebra C(H 2n
), we first exhibit
an isomorphism between the group ExtP P V (Y, y0, A) and the group ExtP P V (Y, y0, Σ2A)
under certain assumptions on the topological space Y where Σ2A is the quantum dou-
ble suspension of A and y0 ∈ Y . Using this, we describe all elements of the group
ExtP P V (T, C(S2ℓ+1
)) explicitly. We then prove that all nonisomprphic middle C ∗- al-
gebras that occur in all the extensions of the group ExtP P V (T, C(S2ℓ+1
)) have different
K-groups. Then using representation theory of C(H 2n
q ), we show that the following
extension
0
0
0 → C(T) ⊗ K → C(H 2n
q ) → C(S4n−3
0
) → 0
is unital and homogeneous. Now by comparing the K-groups, we prove that the above
extension is unitarily equivalent to either the following extension
0 → C(T) ⊗ K → C(S4n−1
0
) → C(S4n−3
0
) → 0
0
0
q ) and C(S4n−1
or its inverse in the group ExtP P V (T, C(S2ℓ+1
)). This proves that the C ∗-algebras
C(H 2n
) are isomorphic. For q = 0, it follows directly from the defin-
ing relations. In [10], it was proved that for different values of q ∈ [0, 1) the C ∗-algebras
C(S4n−1
) are
isomorphic for all q ∈ [0, 1). Also, this shows that the C ∗-algebras C(H 2n
q ) are isomor-
phic for different values of q which establishes q-invariance of the quantum quaternion
spheres.
) are isomorphic. As a consequence, the C ∗-algebras C(H 2n
q ) and C(S4n−1
q
q
We now set up some notations. The standard bases of the Hilbert spaces L2(N) and
L2(Z) will be denoted by {en : n ∈ N} and {en : n ∈ Z} respectively. We denote the left
shift operator on L2(N) and L2(Z) by the same notation S. For m < 0, (S∗)m denotes
the operator S−m. Let pi denote the rank one projection sending ei to ei and p denote
the operator p0. We write L(H) and K(H) for the sets of all bounded linear operators on
H and compact operators on H respectively. We denote by K the C ∗-algebra of compact
operators. For a C ∗-algebra A, Σ2A and M(A) are used to denote the quantum double
3
suspension of A and multiplier algebra of A respectively. The map π will denote the
canonical homomorphism from M(A) to Q(A) := M(A)/A and for a ∈ M(A), [a] stands
for the image of a under the map π. For a locally compact Hausdorff space Y , we write
Y + to denote one point compactification of Y . For a C ∗-algebra A, A+ denotes the
C ∗-algebra obtained by adjoining unity to A. Both the symbols S n and Tn will denote
the n-dimensional sphere. Unless otherwise stated, q will denote a real number in the
interval (0, 1).
2
C ∗-algebra extensions
We first recall some notions related to the C ∗-extension theory. Let A be a unital
separable nuclear C ∗-algebra. Let B be a stable C ∗- algebra. An extension of A by B
j
is a short exact sequence 0 → B i→ E
→ A → 0. In such case, there exists a unique
homomorphism σ : E → M(B) such that σ(i(b)) = b for all b ∈ B. We can now define
j
the Busby invariant for the extension 0 → B i→ E
→ A → 0 as the homomorphism
τ : A → M(B)/B given by τ (a) = π ◦ σ(e) where e is a preimage of a and π is the
canonical map M(B) → M(B)/B. It is easy to see that τ is well defined. An extension
τ is called an essential extension if τ is injective or equivalently image of B is an essential
ideal of E. We call an extension unital if it is a unital homomorphism or equivalently
E is a unital C ∗-algebra. An extension τ is called a trivial (or split) extension if there
exists a homomorphism λ : A → M(B) such that τ = π ◦ λ. Two extensions τ1 and
τ2 are said to be weakly unitarily equivalent if there exists a unitary u in Q(B) such
that uτ1(a)u∗ = τ2(a) for all a ∈ A. They are said to be unitarily equivalent if there
exists a unitary U in M(B) such that π(U)τ1(a)π(U ∗) = τ2(a) for all a ∈ A. We
denote unitarily equivalence relation by ∼u. Let Ext∼u(A, B) denote the set of unitary
equivalence classes of extensions of A by B. One can put a binary operation "+" on
Ext∼u(A, B) as follows. Since M(B) is a stable C ∗-algebra, we can get two isometries ν1
and ν2 in M(B) such that ν1ν∗
2 = 1. Let τ1 and τ2 be two elements in Ext∼u(A, B).
Define τ1 + τ2 : A → Q(B) by
1 +ν2ν∗
(τ1 + τ2)(a) := π(ν1)τ1(a)π(ν∗
1 ) + π(ν2)τ2(a)π(ν∗
2 ).
(2.1)
This makes Ext∼u(A, B) a commutative semigroup. Moreover, the set of trivial exten-
sions forms a subsemigroup of Ext∼u(A, B). We denote the quotient of Ext∼u(A, B)
with the set of trivial extensions by Ext(A, B). For a separable nuclear C ∗-algebra
A, the set Ext(A, B) under the operation + is a group (see [1]). Two extensions
4
τ1 and τ2 represent the same element in Ext(A, B) if there exists two trivial exten-
sions φ1 and φ2 such that τ1 + φ1 ∼u τ2 + φ2. One can show that for a stable C ∗-
algebra B, Ext(A, B) = Ext(A, B ⊗ K). Now for an arbitrary C ∗-algebra B, define
Ext(A, B) := Ext(A, B ⊗ K). We denote an equivalent class in the group Ext(A, B) of
an extension τ by [τ ]s. For B = C, we denote the group Ext(A, C) by Ext(A). Note
that in this case, two unital essential extensions τ1 and τ2 are in the same equivalence
class (i.e. [τ1]s = [τ2]s) if and only if they are unitarily equivalent.
Suppose that Y is a finite dimensional compact metric space i.e. a closed subset
of S n for some n ∈ N. Let M(Y ), Q(Y ) and Q be the C ∗-algebras M(C(Y ) ⊗ K),
M(C(Y ) ⊗ K)/C(Y ) ⊗ K and L(H)/K(H) (Calkin algebra) respectively. It is easy to
show that M(Y ) is the set of all continuous functions from Y to L(H) where continuity
is with respect to ∗-strong operator topology on L(H). We call an extension τ of A by
C(Y ) ⊗ K homogeneous if for all y ∈ Y , the map evy ◦ τ : A → Q is injective where
evy : Q(Y ) → Q is the evaluation map at y. Let ExtP P V (Y, A) be the set of unitary
equivalence classes of unital homogeneous extensions of A by C(Y ) ⊗ K. For a nuclear
C ∗-algebra A, Pimsner, Popa and Voiculescu ([11]) showed that ExtP P V (Y, A) is a group
with the additive operation defined in 2.1. We denote the equivalence class in the group
ExtP P V (Y, A) of an extension τ by [τ ]u. For y0 ∈ Y , define the set
ExtP P V (Y, y0, A) = {[τ ]u ∈ ExtP P V (Y, A) : evy0 ◦ τ is trivial} .
The set ExtP P V (Y, y0, A) is a subgroup of ExtP P V (Y, A).
2.1 The groups ExtP P V (Y, A) and ExtP P V (Y, Σ2A)
Here we will show that for a separable nuclear C ∗-algebra A and a finite dimensional
compact metric space Y such that K-groups of C(Y ) are free groups with finite genera-
tors, the groups ExtP P V (Y, A) and ExtP P V (Y, Σ2A) are isomorphic. Let us recall some
definitions. We say that two elements a and b in Q(A) are unitarily equivalent if there
exists a unitary U ∈ M(A) such that [U]a[U ∗] = b. They are weakly unitarily equivalent
if there exists unitary u ∈ Q(A) such that uau∗ = b. We call an element a in a C ∗-algebra
B norm-full if it is not contained in any proper closed ideal in B. Suppose that A and
B are separable C ∗-algebras. An extension τ : A → Q(B ⊗ K) is said to be norm-full if
for every nonzero element a ∈ A, τ (a) is norm full element of Q(B ⊗ K).
Definition 2.1. Let B be a separable stable C ∗-algebra. Then B is said to have the
corona factorization property if every norm-full projection in M(B) is Murray-von Neu-
mann equivalent to unit element of M(B).
5
It is easy to see that a C ∗-algebra A with corona factorization property, any norm-
full projection in Q(B) is Murray-von Neumann equivalent to 1 of Q(B). Further, one
can show that for a finite dimensional compact metric space Y , C(Y ) ⊗ K has corona
factorization property (see [11]).
Proposition 2.2. Let A be a unital separable nuclear C ∗-algebra which satisfies the
Universal Coefficient Theorem. Suppose that Y is a finite dimensional compact metric
space. Then the map
i : ExtP P V (Y, A) −→ KK 1(A, C(Y ))
[τ ]u 7→ [τ ]s
is an injective homomorphism.
Proof : Since unitarily equivalence implies stable equivalence, the map i is well defined.
Any unital homogeneous extension is a purely large extension and hence a norm-full
extension (see page 19, [4]). Therefore from Theorem 2.4 in [7], it follows that i is
injective.
(cid:3)
From now on, without loss of generality, we will assume that the Hilbert space H is
L2(N). Let τ be a unital homogeneous extension of A by C(Y ) ⊗ K(H). Define τ : A →
Q(C(Y ) ⊗ K(H) ⊗ K(H)) by : τ (a) = [τa ⊗ p] where [τa] = τ (a). By universal property
of quantum double suspension (see proposition 2.2, [5]), we have a homomorphism
Σ2τ : Σ2A → Q(C(Y ) ⊗ K(H) ⊗ K(H))
(2.2)
such that Σ2τ (a ⊗ p) = τ (a) = [τa ⊗ p] and Σ2τ (1 ⊗ S) = [1 ⊗ 1 ⊗ S]. Clearly Σ2τ is
a unital extension. Since τ is homogeneous, the map evy ◦ Σ2τ is injective on the C ∗-
algebra A ⊗ p for all y ∈ Y . Making use of the fact that (1 ⊗ p)A ⊗ K(1 ⊗ p) = A ⊗ p,
one can prove that the map evy ◦ Σ2τ is injective on A ⊗ K. Since A ⊗ K is an essential
ideal of Σ2A, we conclude that the map evy ◦ Σ2τ is injective on Σ2A and hence Σ2τ is
a homogeneous extension. Moreover, if τ1 and τ2 are unitarily equivalent by a unitary
U ∈ M(C(Y ) ⊗ K(H)) then so are Σ2τ1 and Σ2τ2 by the unitary U ⊗ 1 ∈ M(C(Y ) ⊗
K(H) ⊗ K(H)). This gives a well defined map
β : ExtP P V (Y, A) −→ ExtP P V (Y, Σ2A)
[τ ]u 7→ [Σ2τ ]u
Proposition 2.3. The map β : ExtP P V (Y, A) −→ ExtP P V (Y, Σ2A) given above is in-
jective group homomorphism.
6
Proof : It follows from straightforward calculations.
(cid:3)
To get surjectivity of the map β, we need to put certain assumptions on the topological
space Y .
Proposition 2.4. Let Y be a finite dimensional compact metric space. Assume that
K0(C(Y )) and K1(C(Y )) are free groups with finite number of generators. Let V ∈
Q(C(Y ) ⊗ K(H) ⊗ K(H)) be a isometry such that V V ∗ and 1 − V V ∗ both are norm full
projections. Then V is weakly unitarily equivalent to [1 ⊗ 1 ⊗ S].
Proof : We assume that V is not weakly unitarily equivalent to [1⊗1⊗S]. Since C(Y )⊗K
has corona factorization property, it follows that V V ∗ and 1 − V V ∗ both are Murray-von
Neumann equivalent to [1] of Q(C(Y ) ⊗ K(H) ⊗ K(H)). Also, one can easily verify that
[1 ⊗ 1 ⊗ p] and [1 − 1 ⊗ 1 ⊗ p] = [1 ⊗ 1 ⊗ (1 − p)] are Murray-von Neumann equivalent to
[1] of Q(C(Y ) ⊗ K(H) ⊗ K(H)). This implies that V V ∗ is weakly unitarily equivalent to
1 − [1 ⊗ 1 ⊗ p]. So, without loss of generality, we can assume that V has final projection
1 − [1 ⊗ 1 ⊗ p]. Take a split unital homogeneous extension τ of C(T) by C(Y ) ⊗ K(H).
V τ be a unital homogeneous extension of Σ2C(T) by C(Y ) ⊗ K(H ⊗ K(H) given
Let Σ2
by Σ2
V τ (a ⊗ p) = [τa ⊗ p] and Σ2
V τ (1 ⊗ S) = V where [τa] = τ (a). From Corollary 3.8
([7]) and the fact that V is not weakly unitarily equivalent to [1 ⊗ 1 ⊗ S], it follows that
[Σ2
V τ ]u = β([φ]u) for some n ∈ Z − {0}
and for some unital homogeneous extension φ of C(T) by C(Y ) ⊗ K(H). It is easy to
see that φ must be split and in that case n[Σ2
V τ ] is the class of split extensions. By
proposition 2.2 and the fact that KK 1(Σ2C(T), C(Y )) is free group, we get that Σ2
V τ is
a split extension. This contradicts the fact that [Σ2
V τ ]u is not in the image of the map β.
So, for any n ∈ Z−{0}, n[Σ2
V τ ]u is not in the image of the map β. This shows that image
of ExtP P V (Y, Σ2C(T)) in the group KK 1(Σ2C(T), C(Y )) has one more free generator
than the group ExtP P V (Y, C(T)) in KK 1(C(T), C(Y )) ≡ KK 1(Σ2C(T), C(Y )). Since
for all n ∈ N, KK 1(Σ2nC(T), C(Y )) ≡ K0(C(Y )) ⊕ K1(C(Y )) are free groups with finite
generators, iterating above process will lead to a contradiction. This proves that V is
weakly unitarily equivalent to [1 ⊗ 1 ⊗ S].
V τ ]u is not in the image of the map β. Let n[Σ2
(cid:3)
Remark 2.5. Here we should point out that above proposition may hold for any finite
dimensional compact metric space Y . But since we could not find it in literature, we
prove the proposition under certain assumptions on Y .
Corollary 2.6. Let Y and V be as in the above proposition. Then V is unitarily equiv-
alent to [1 ⊗ 1 ⊗ S].
7
Proof : Consider the unital extension Σ2
3.8 ([7]), it follows that Σ2
A = C(T). Hence V is unitarily equivalent to [1 ⊗ 1 ⊗ S].
V τ constructed in proposition 2.4. From Corollary
V τ is unitarily equivalent to Σ2τ defined in equation 2.2 with
(cid:3)
The following lemma establishes the isomorphism between the groups ExtP P V (Y, A)
and ExtP P V (Y, Σ2A) under certain assumptions on the space Y .
Lemma 2.7. Let Y be a finite dimensional compact metric space. Assume that the
groups K0(C(Y )) and K1(C(Y )) are free groups with finite number of generators. Then
the map β : ExtP P V (Y, A) −→ ExtP P V (Y, Σ2A) given above is an isomorphism.
Proof : We only need to show that β is surjective thanks to proposition 2.3. Let φ be
a unital homogeneous extension of Σ2A by C(Y ) ⊗ K(H) ⊗ K(H). Let φ(1 ⊗ S) = V .
Since φ is a unital homogeneous extension and hence a norm full extension, it follows
that V V ∗ and 1 −V V ∗ are norm full projections. Therefore by Corollary 2.6, there exists
a unitary U ∈ M(C(Y ) ⊗ K(H) ⊗ K(H)) such that [U]V [U ∗] = [1 ⊗ 1 ⊗ S]. So without
loss of generality, we can assume that φ maps 1 ⊗ S to [1 ⊗ 1 ⊗ S]. This implies that
φ(1 ⊗ p) = [1 ⊗ 1 ⊗ p]. But then φ(A ⊗ p) ⊂ (1 ⊗ p)φ(A ⊗ p)(1 ⊗ p) ⊂ Q(C(Y ) ⊗ K(H)) ⊗ p
which induces a map τ : A → Q(C(Y ) ⊗ K(H)) by omitting the projection p. Therefore
we get a unital homogeneous extension of A such that β([τ ]u) = [φ]u. Hence β is
surjective.
(cid:3)
Corollary 2.8. For y0 ∈ Y , the map
βExtP P V (Y,y0,A) : ExtP P V (Y, y0, A) −→ ExtP P V (Y, y0, Σ2A)
is an isomorphism.
Proof : It is easy to check that if evy0 ◦ τ is split then so is evy0 ◦ Σ2τ and vice versa.
Now the claim will follow by Lemma 2.7.
(cid:3)
3
Elements of ExtP P V (T, C(S2ℓ+1
0
))
In the present section, we will write down all elements of the groups Ext(C(S2ℓ+1
ExtP P V (T, C(S2ℓ+1
)) and
)) in terms of their Busby invariants. Define the ∗-homomorphisms
0
0
8
ϕm as follows:
ϕm : C(S2ℓ+1
) → Q(cid:16)K(cid:0) L2(N) ⊗ · · · ⊗ L2(N)
}
{z
S∗ ⊗ 1 ⊗ · · · ⊗ 1 7→ S∗ ⊗ 1 ⊗ · · · ⊗ 1
ℓ+1 copies
0
(cid:1)(cid:17)
p ⊗ S∗ ⊗ 1 ⊗ · · · ⊗ 1 7→ p ⊗ S∗ ⊗ 1 ⊗ · · · ⊗ 1
· · ·
p ⊗ p ⊗ · · · ⊗ p ⊗ S∗ ⊗ 1 7→ p ⊗ p ⊗ · · · ⊗ S∗ ⊗ 1
p ⊗ p ⊗ · · · ⊗ t 7→ p ⊗ p ⊗ · · · ⊗ p ⊗ (S∗)m
It is easy to verify that ϕm's are essential unital extensions of C(S2ℓ+1
operators. Hence [ϕm]s ∈ Ext(C(S2ℓ+1
Ext(C(S2ℓ+1
L2(N) ⊗ · · · ⊗ L2(N)
) by compact
)). We shall show that each element in the group
)) is of the form [ϕm]s for some m ∈ Z. Let H0 be the Hilbert space
) given
⊗L2(Z). For m ∈ Z, let ϑm be the representation of C(S2ℓ+1
0
0
0
0
by
ℓ copies
{z
}
ϑm : C(S2ℓ+1
0
) → L(H0)
S∗ ⊗ 1 ⊗ · · · ⊗ 1 7→ S∗ ⊗ 1 ⊗ · · · ⊗ 1
p ⊗ S∗ ⊗ 1 ⊗ · · · ⊗ 1 7→ p ⊗ S∗ ⊗ 1 ⊗ · · · ⊗ 1
· · ·
p ⊗ p ⊗ · · · ⊗ p ⊗ S∗ ⊗ 1 7→ p ⊗ p ⊗ · · · ⊗ S∗ ⊗ 1
p ⊗ p ⊗ · · · ⊗ t 7→ p ⊗ p ⊗ · · · ⊗ p ⊗ (S∗)m
Let P be the self adjoint projection in L(H0) on the subspace spanned by the ba-
sis elements (cid:8)en1 ⊗ · · · ⊗ enℓ+1 : ni ∈ N for all i ∈ {1, 2, · · · , ℓ + 1}(cid:9). One can check that
Fm := (cid:16)C(S2ℓ+1
), H0, 2P − 1(cid:17) with the underlying representation ϑm is a Fredholm
module. By Proposition 17.6.5 in ([1], page 157), the group Ext(C(S2ℓ+1
)) is isomor-
phic to the group K 1(C(S2ℓ+1
)). Under this identification, one can easily show that the
equivalence class of the Fredholm module Fm corresponds to the equivalence class [ϕm]s.
0
0
0
Proposition 3.1. For ℓ ∈ N, one has
Ext(C(S2ℓ+1
0
)) = {[ϕm]s : m ∈ Z} .
Proof : To prove the claim, we will use the index pairing between the groups K1(C(S2ℓ+1
and K 1(C(S2ℓ+1
0
)) given by Kasparov product (see [1]). The group K1(C(S2ℓ+1
))
)) is
0
0
9
generated by the unitary u := p ⊗ · · · ⊗ p
⊗t + 1 − p ⊗ · · · ⊗ p
⊗1. For m ∈ Z, let
ℓ copies
{z
}
}
Rm : P H0 → P H0 be the operator P ϑm(u)P . Hence we get
ℓ copies
{z
hu, Fmi = Index(Rm) = m.
This completes the proof.
To describe all elements of ExtP P V (T, C(S2ℓ+1
follows:
0
)), we define the ∗-homomorphisms φm as
(cid:3)
0
φm : C(S2ℓ+1
) → Q(cid:16)K(cid:0) L2(N) ⊗ · · · ⊗ L2(N)
}
{z
S∗ ⊗ 1 ⊗ · · · ⊗ 1 7→ S∗ ⊗ 1 ⊗ · · · ⊗ 1
ℓ+1 copies
(cid:1) ⊗ C(T)(cid:17)
p ⊗ S∗ ⊗ 1 ⊗ · · · ⊗ 1 7→ p ⊗ S∗ ⊗ 1 ⊗ · · · ⊗ 1
· · ·
p ⊗ p ⊗ · · · ⊗ p ⊗ S∗ ⊗ 1 7→ p ⊗ p ⊗ · · · ⊗ S∗ ⊗ 1 ⊗ 1
p ⊗ p ⊗ · · · ⊗ t 7→ p ⊗ p ⊗ · · · ⊗ p ⊗ (S∗)m ⊗ 1
It is easy to verify that φm's are essential unital extensions. Since last component is
1, these extensions are also homogeneous. Let Am be the C ∗-subalgebra of C(S2ℓ+3
)
generated by the operators
0
S∗ ⊗ 1 ⊗ · · · ⊗ 1 ⊗ 1,
p ⊗ S∗ ⊗ 1 ⊗ · · · ⊗ 1 ⊗ 1
· · ·
p ⊗ p ⊗ · · · ⊗ S∗ ⊗ 1 ⊗ 1,
p ⊗ p ⊗ · · · ⊗ p ⊗ (S∗)m ⊗ 1
(cid:1) ⊗ C(T). Then for each m ∈ Z, we have the following exact
and K(cid:0) L2(N) ⊗ · · · ⊗ L2(N)
}
ℓ+1 copies
{z
sequence
0 −→ K(cid:0) L2(N) ⊗ · · · ⊗ L2(N)
}
ℓ+1 copies
{z
(cid:1) ⊗ C(T) −→ Am −→ C(S2ℓ+1
0
) −→ 0
with the Busby invariant φm. By using the six term sequence, one can show that
K0(Am) = Z ⊕ Z/mZ,
K1(Am) = Z.
(3.1)
10
Lemma 3.2. For ℓ ∈ N and t0 ∈ T, one has
ExtP P V(cid:0)T, t0, C(S2ℓ+1
0
)(cid:1) = {0} ,
ExtP P V(cid:0)T, C(S2ℓ+1
0
)(cid:1) = Z.
Proof : It follows from Theorem 1.5 in [13] that
ExtP P V(cid:0)T, t0, C(T)(cid:1) = ExtP P V(cid:0)R+, t0, C0(R)+(cid:1) = Ext(cid:0)C0(R), C0(R)(cid:1) = {0} .
The C ∗-algebra C(S2ℓ+1
C(T) repeatedly (see [6]). Therefore from Corollary 2.8, we have
) can be obtained by applying quantum double suspension on
0
ExtP P V(cid:0)T, t0, C(S2ℓ+1
0
)(cid:1) = ExtP P V(cid:0)T, t0, C(T)(cid:1) = {0} .
Further from Theorem 1.4 in [13], we get
ExtP P V(cid:0)T, C(T)(cid:1) = ExtP P V(cid:0)T, C0(R)+(cid:1) = Ext(cid:0)C0(R), C(T)(cid:1) = Z.
Hence by applying Lemma 2.7, we get the claim.
(cid:3)
The following lemma says that each element of the group ExtP P V (T, C(S2ℓ+1
form [φm]u for some m ∈ Z.
0
)) is of the
Lemma 3.3. For ℓ ∈ N, one has
ExtP P V (T, C(S2ℓ+1
0
)) = {[φm]u : m ∈ Z} .
Proof : Fix t0 ∈ T. Define a homomorphism Ψ as follows:
Ψ : ExtP P V(cid:0)T, C(S2ℓ+1
0
[τ ]u 7−→ [evt0 ◦ τ ]s
)(cid:1) −→ Ext(cid:0)C(S2ℓ+1
0
)(cid:1)
Clearly ker Ψ = ExtP P V(cid:0)T, t0, C(S2ℓ+1
0
)(cid:1) = {0}. Therefore Ψ is an injective group
homomorphism. Since for all m ∈ Z, evt0 ◦ φm = ϕm, it follows that the homomorphism
Ψ is surjective. This proves the claim.
(cid:3)
4 Quantum quaternion sphere
In this section, we first recall the definition and representation theory of the C ∗-algebra
C(H 2n
q ) of continuous functions on the quantum quaternion sphere. Then we prove our
main result that the C ∗-algebra C(H 2n
q ) is isomorphic to the C ∗-algebra C(S4n−1
).
q
11
Definition 4.1. The C ∗-algebra C(H 2n
q ) of continuous functions on the quantum quater-
nion sphere is defined as the universal C ∗-algebra generated by elements z1, z2, ....z2n
satisfying the following relations:
for i > j, i + j 6= 2n + 1
(4.1)
qi−kzkzk
′
for i > n
(4.2)
(4.3)
zizj = qzjzi
′ = q2zi
k>i
zizi
′ zi − (1 − q2)X
′ z∗
′ = q2zi
z∗
i zi
i
z∗
i zj = qzjz∗
z∗
i zj = qzjz∗
z∗
i zi = ziz∗
i
i zi = ziz∗
z∗
2n
X
i=1
ziz∗
i = 1
i + (1 − q2)ǫiǫjqρi+ρj zi′ z∗
j ′
i + (1 − q2)X
i + (1 − q2)q2ρizi′ z∗
zkz∗
k
k>i
i′ + (1 − q2)X
k>i
for i + j > 2n + 1, i 6= j
(4.4)
for i + j < 2n + 1, i 6= j
(4.5)
for i > n
zkz∗
k
for i ≤ n
(4.6)
(4.7)
(4.8)
m, u2n
In [14], we showed that the C ∗-algebra C(H 2n
q ) is isomorphic to the quotient algebra
C(SPq(2n)/SPq(2n − 2)) that can also be described as the C ∗-subalgebra of C(SPq(2n))
generated by {u1
m : m ∈ {1, 2, · · · 2n}} i.e. elements of first and last row of funda-
mental matrix of C(SPq(2n)). Here we briefly describe all irreducible representations of
C(H 2n
q ). For a detailed treatment on this, we refer the reader to [14]. Let N be the num-
ber operator given by N : en 7→ nen and S be the shift operator given by S : en 7→ en−1 on
L2(N). For i = 1, 2, · · · , n − 1, let πsi denote the following representation of C(SPq(2n)),
πsi(uk
l ) =
p1 − q2N +2S
S∗p1 − q2N +2
−qN +1
qN
qN +1
−qN
δkl
if (k, l) = (i, i) or (2n − i, 2n − i),
if (k, l) = (i + 1, i + 1) or (2n − i + 1, 2n − i + 1),
if (k, l) = (i, i + 1),
if (k, l) = (i + 1, i),
if (k, l) = (2n − i, 2n − i + 1),
if (k, l) = (2n − i + 1, 2n − i),
otherwise .
12
For i = n,
πsn(uk
l ) =
p1 − q4N +4S
S∗p1 − q4N +4
−q2N +2
q2N
δkl
if (k, l) = (n, n),
if (k, l) = (n + 1, n + 1),
if (k, l) = (n, n + 1),
if (k, l) = (n + 1, n),
otherwise .
Each πsi is an irreducible representation and is called an elementary representation
of C(SPq(2n)). For any two representations ϕ and ψ of C(SPq(2n)) define, ϕ ∗ ψ :=
(ϕ ⊗ ψ) ◦ ∆ where ∆ is the co-multiplication map of C(SPq(2n)). Let W be the Weyl
group of sp2n and ϑ ∈ W such that si1si2...sik is a reduced expression for ϑ. Then πϑ =
πsi1
is an irreducible representation which is independent of the reduced
expression. Now for t = (t1, t2, · · · , tn) ∈ Tn, define the map τt : C(SPq(2n) −→ C by
∗ · · · ∗ πsik
∗ πsi2
τt(ui
j) =
tiδij
t2n+1−iδij
if i ≤ n,
if i > n,
Then τt is a ∗-algebra homomorphism. For t ∈ Tn, ϑ ∈ W , let πt,ϑ = τt ∗ πϑ. Define the
representation ηt,ϑ of C(H 2n
q ). Denote by ωk the following
word of Weyl group of sp2n,
q ) as πt,ϑ restricted to C(H 2n
I
ωk =
s1s2 · · · sk−1
s1s2 · · · sn−1snsn−1 · · · s2n−k+1
if k = 1,
if 2 ≤ k ≤ n,
if n < k ≤ 2n.
For k = 1, define ηt,I : C(H 2n
q ) → C such that ηt,I(zj) = tδ1j. The set {ηt,I : t ∈ T }
gives all one dimensional irreducible representations of C(H 2n
q ).
Theorem 4.2. ([14]) The set {ηt,ωk : 1 ≤ k ≤ 2n, t ∈ T} gives a complete list of irre-
ducible representations of C(H 2n
q ).
Define ηωk : C(H 2n
q ). Let C 2n
C(H 2n
q ) → C(T) ⊗ T ⊗k−1 such that ηωk(a)(t) = ηt,ωk (a) for all a ∈
1 = C(T) and for 2 ≤ k ≤ 2n, C 2n
k = ηωk (C(H 2n
q )).
Corollary 4.3. The set {ηt,ωl : 1 ≤ l ≤ k, t ∈ T} gives a complete list of irreducible rep-
resentations of C 2n
k .
Corollary 4.4. ηωk is a faithful representation of C 2n
k .
13
By Corollary 4.3, one can find all primitive ideals i.e. kernels of irreducible represen-
k generated
k(cid:9). For t ∈ T, let Ct(T) be the set of all continuous functions on T
l := ηωk(zl) for 1 ≤ l ≤ k. Let I k
tations of C 2n
by (cid:8)yk
l be the ideal of C 2n
k . Define yk
l+1, · · · , yk
l , yk
vanishing at the point t. Then
(cid:8)Ct(T) ⊗ K(L2(N))⊗(k−1)(cid:9)t∈T ⊂ I k
k ⊂ I k
k−1 ⊂ · · · ⊂ I k
1 = C 2n
k
(4.9)
is a complete list of primitive ideals of C 2n
following exact sequence
k . In Lemma 5.1 in [14], we established the
0 −→ C(T) ⊗ K(L2(N))⊗(k) −→ C 2n
k+1
σk+1−→ C 2n
k −→ 0
where σk+1 is the restriction of (1⊗(k) ⊗ σ) to C 2n
k+1, the map σ : T → C is the homo-
morphism such that σ(S) = 1 and T is the toeplitz algebra. The following lemma says
that this exact sequence is a unital homogeneous extension of C 2n
k by C(T) ⊗ K.
Lemma 4.5. For 1 ≤ k ≤ 2n, the following exact sequence
0 −→ C(T) ⊗ K(L2(N))⊗(k) −→ C 2n
k+1
σk+1−→ C 2n
k −→ 0
is a unital homogeneous extension of C 2n
k by C(T) ⊗ K.
k+1 is unital, the given extension is unital. Let τ : C 2n
Proof : Since C 2n
k → Q(T) be
the Busby invariant corresponding to this extension. For t0 ∈ T, let τt0 : C 2n
k → Q
be the map evt0 ◦ τ where evt0 : Q(T) → Q is the evaluation map at t0. Assume that
Jt0 = ker(τt0). To show that the given short exact sequence is a homogeneous extension,
we need to prove that Jt0 = {0} for all t0 ∈ T.
Case 1: n < k < 2n
We have
τt0(yk
)
k) = τt0(t ⊗ qN ⊗ · · · ⊗ qN
}
(n−1) copies
{z
= t0[qN ⊗ · · · ⊗ qN
}
⊗q2N ⊗ qN ⊗ · · · ⊗ qN
}
{z
⊗q2N ⊗ qN ⊗ · · · ⊗ qN
}
(k−n−1) copies
(k−n−1) copies
(n−1) copies
{z
{z
6= 0.
⊗p1 − q2N S∗]
(4.10)
This shows yk
k /∈ Jt0. Since Jt0 is intersection of all primitive ideals that contains Jt0, we
conclude that Jt0 is equal to CF (T) ⊗ K for some closed subset F of T where CF (T) is
14
set of all continuous functions on T vanishing on F . From equation 4.10, we get
τt0((yk
k)(yk
(n−1) copies
k)∗) = [q2N ⊗ · · · ⊗ q2N
}
= [q2N ⊗ · · · ⊗ q2N
}
(n−1) copies
{z
{z
⊗q4N ⊗ q2N ⊗ · · · ⊗ q2N
}
(k−n−1) copies
⊗q4N ⊗ q2N ⊗ · · · ⊗ q2N
}
(k−n−1) copies
{z
{z
⊗(1 − q2N )]
⊗1].
Therefore
Hence
τt0(1 ⊗ p ⊗ · · · p
(k−1) copies
{z }
) = [ p ⊗ · · · ⊗ p
⊗1].
{z
(k−1) copies
}
τt0(t ⊗ p ⊗ · · · ⊗ p
) = t0[ p ⊗ · · · ⊗ p
(k−1) copies
{z
}
(k−1) copies
= t0[ p ⊗ · · · ⊗ p
(k−1) copies
{z
{z
⊗p1 − q2N S∗]
⊗S∗].
}
}
Consider the function χ : C(T) → Q such that χ(t) = [S∗]. Since [S∗] is unitary in Q
with spectrum equal to T, it follows that the map χ is injective. This shows that for any
) 6= 0 which further implies that F = T
nonzero function f on T, τt0(f (t) ⊗ p ⊗ · · · ⊗ p
(k−1) copies
{z
}
and Jt0 = {0}.
Case 2: 1 ≤ k ≤ n
For k = n,
For 1 ≤ k < n,
τt0(yn
n) = t0[qN ⊗ · · · ⊗ qN
}
(n−1) copies
{z
τt0(yk
k) = t0[qN ⊗ · · · ⊗ qN
}
(k−1) copies
{z
⊗p1 − q4N S∗].
⊗p1 − q2N S∗].
Similar calculations as done in the case 1 shows that Jt0 = {0}. This establishes the
claim.
(cid:3)
We now state the main result of this paper.
Theorem 4.6. For all n ∈ N, n ≥ 2 and 1 ≤ k ≤ 2n, the C ∗-algebra C 2n
k
to the C ∗-algebra C(S2k−1
In particular, C(H 2n
is isomorphic
) of continuous functions on odd dimensional quantum sphere.
) or equivalently to C(S4n−1
q ) is isomorphic to C(S4n−1
).
0
q
0
15
Proof : Fix n. To prove the theorem, we use induction on k. For k = 1, C 2n
the claim is true for k = 1. Assume that the claim is true for k i.e. C 2n
k
C(S2k−1
). From Lemma 4.5, it follows that following short exact sequence
1 = C(T). So
is isomorphic to
0
0 −→ C(T) ⊗ K −→ C 2n
k+1 −→ C 2n
k −→ 0
(4.11)
is a unital homogeneous extension. Therefore it can be viewed as an element of the group
ExtP P V (T, C(S2k−1
)). This implies that it is unitarily equivalent to φm or equivalently
to the following exact sequence
0
0 −→ C(T) ⊗ K −→ Am −→ C(S2k−1
0
) −→ 0
for some m ∈ Z. From Theorem 5.3 in [14] and equation (3.1), we have
K0(C 2n
k+1) = Z,
K0(Am) = Z ⊕ Z/mZ.
Since unitary equivalence gives an isomorphism of the middle C ∗ algebras and hence an
isomorphism of the K-groups of middle C ∗-algebras, it follows that the exact sequence
4.11 is unitarily equivalent to φ1 or φ−1. This implies that C 2n
k+1 is isomorphic to A1 or
A−1. Since A1 = A−1 = C(S2k+1
). Hence
by induction, it follows that C(H 2n
). From Lemma 3.2 in [10],
it follows that the C ∗-algebras C(S4n−1
) for q ∈ (0, 1). This
proves that C(H 2n
), it follows that C 2n
q ) is isomorphic to C(S4n−1
k+1 is isomorphic to C(S2k+1
) are isomorphic to C(S4n−1
q ) is isomorphic to C(S4n−1
(cid:3)
).
q
0
0
q
0
0
Remark 4.7. In case of q = 0, we need to be slightly careful to get the defining relations
of C(H 2n
0 ). In the relation (4.2), we first start with i = 2n which gives z2nz1 = 0. Then
we take i = 2n − 1 and so on and get the relation zizi′ = 0 for i < n. Further in the
relation (4.5), it is easy to check that for i + j < 2n + 1, ρi + ρj > 0. Now by putting
q = 0 in the relations (4.3), (4.4) and (4.4), we get z∗
i zj = 0 for i 6= j. Other relations
are obtained by putting q = 0 in the remaining relations. By looking at the relations,
one can see that the defining relations of C(H 2n
).
These facts together with Theorem 4.6 prove that for different values of q ∈ [0, 1), the
C ∗-algebras C(H 2n
0 ) are exactly same as those of C(S4n−1
q ) are isomorphic.
0
Acknowledgement: I would like to thank Arup Kumar Pal, my supervisor, for his constant support
and for his valuable suggestions.
16
References
[1] Bruce Blackadar. K-theory for operator algebras, volume 5 of Mathematical Sciences
Research Institute Publications. Cambridge University Press, Cambridge, second
edition, 1998.
[2] L.G. Brown, R. G. Douglas and P. A. Fillmore. Extensions of C ∗-algebras and
K-homology. Ann. of Math. (2) 105 (1977), no. 2, 265-324.
[3] Partha Sarathi Chakraborty and Arupkumar Pal. Characterization of SUq(ℓ + 1)-
equivariant spectral triples for the odd dimensional quantum spheres. J. Reine
Angew. Math., 623:25 -- 42, 2008.
[4] George A. Elliott and Dan Kucerovsky. An abstract Voiculescu-Brown-Douglas-
Fillmore absorption theorem. Pacific J. Math. 198 (2001), no. 2, 385-409.
[5] Jeong Hee Hong and Wojciech Szyma´nski. Noncommutative balls and mirror quan-
tum spheres. J. Lond. Math. Soc. (2) 77 (2008), no. 3, 607-626.
[6] Jeong Hee Hong and Wojciech Szyma´nski. Quantum spheres and projective spaces
as graph algebras. Comm. Math. Phys., 232(1):157 -- 188, 2002.
[7] Huaxin Lin.
Unitary equivalences for essential extensions of
C ∗-algebras.
arXiv:math/0403236, 2004.
[8] G. G. Kasparov. The K-functors in the theory of extensions of C ∗-algebras. (Rus-
sian) Funktsional. Anal. i Prilozhen. 13 (1979), no. 4, 73-74.
[9] Sergey Neshveyev and Lars Tuset. Quantized algebras of functions on homogeneous
spaces with Poisson stabilizers. Comm. Math. Phys., 312(1):223 -- 250, 2012.
[10] Arupkumar Pal and S. Sundar. Regularity and dimension spectrum of the equiv-
ariant spectral triple for the odd-dimensional quantum spheres. J. Noncommut.
Geom., 4(3):389 -- 439, 2010.
[11] M. Pimsner, S. Popa and D. Voiculescu. Homogeneous C ∗-extensions of C(X) ⊗
K(H). 1. J. Operator Theory 1 (1979), no. 1, 55-108.
[12] G. B. Podkolzin and L. I. Vainerman. Quantum Stiefel manifold and double cosets
of quantum unitary group. Pacific J. Math., 188(1):179 -- 199, 1999.
17
[13] Jonathan Rosenberg and Claude Schochet. Comparing functors classifying exten-
sions of C ∗-algebras. J. Operator Theory 5 (1981), no. 2, 267-282.
[14] Bipul Saurabh. On quantum quaternion spheres, 2015. To appear in Proc. Indian
Acad. Sci. Math. Sci.
[15] Claude Schochet. Homogeneous extensions of C ∗-algebras and K-theory. I. Bull.
Amer. Math. Soc. (N.S.) 3 (1980), no. 1, part 1, 715-718.
[16] L. L. Vaksman and Ya. S. Soibelman. Algebra of functions on the quantum group
SU(n + 1), and odd-dimensional quantum spheres. Algebra i Analiz, 2(5):101 -- 120,
1990.
Bipul Saurabh ([email protected])
Indian Statistical Institute, 7, SJSS Marg, New Delhi -- 110 016, INDIA
18
|
1712.09551 | 2 | 1712 | 2018-09-23T12:04:20 | On the K-theory of C*-algebras for substitution tilings (a pedestrian version) | [
"math.OA",
"math.DS",
"math.KT"
] | Under suitable conditions, a substitution tiling gives rise to a Smale space, from which three equivalence relations can be constructed, namely the stable, unstable, and asymptotic equivalence relations. We denote with $S$, $U$, and $A$ their corresponding $C^*$-algebras in the sense of Renault. In this article we show that the $K$-theories of $S$ and $U$ can be computed from the cohomology and homology of a single cochain complex with connecting maps for tilings of the line and of the plane. Moreover, we provide formulas to compute the $K$-theory for these three $C^*$-algebras. Furthermore, we show that the $K$-theory groups for tilings of dimension 1 are always torsion free. For tilings of dimension 2, only $K_0(U)$ and $K_1(S)$ can contain torsion. | math.OA | math |
On the K-theory of C∗-algebras for substitution tilings
(a pedestrian version)
DANIEL GONC¸ ALVES∗ MARIA RAMIREZ-SOLANO∗∗
In memory of Uffe Haagerup†
Abstract. Under suitable conditions, a substitution tiling gives rise to a
Smale space, from which three equivalence relations can be constructed, namely
the stable, unstable, and asymptotic equivalence relations. We denote with S,
U, and A their corresponding C∗-algebras in the sense of Renault.
In this
article we show that the K-theories of S and U can be computed from the
cohomology and homology of a single cochain complex with connecting maps
for tilings of the line and of the plane. Moreover, we provide formulas to com-
pute the K-theory for these three C∗-algebras. Furthermore, we show that the
K-theory groups for tilings of dimension 1 are always torsion free. For tilings
of dimension 2, only K0(U) and K1(S) can contain torsion.
1. Introduction
The study of aperiodic order has gained impetus after the discovery of quasicrys-
tals by Dan Shechtman in 1982, for which he was awarded the 2011 Nobel Prize
in Chemistry. The quasicrystal structure can be explained by an aperiodic tiling
[41], [17], [39], and physical properties of quasicrystals are related to mathematical
aspects of the corresponding tiling. An important mathematical aspect to be un-
derstood is the K-theory of C∗-algebras associated with an aperiodic tiling, since
K-theory enters physics through Bellissard's formulation of gap labelling [5], [6], [7].
Moreover, K-theory has applications to deformations, spaces of measures, and exact
regularity [9], [37]. One can associate the so-called stable and unstable C∗-algebras
to a tiling. The first author computed the K-theory for the stable C∗-algebra
for a few examples in an ad-hoc manner in [16], and from that the first author
raised the conjecture that there is a relationship between the stable and unstable
C∗-algebras. In this paper we show that there is indeed a relationship. To this
end, we introduce the so-called stable cohomology and stable-transpose homology,
from which one can compute the K-theories of the stable and unstable C∗-algebras,
respectively. In particular, for tilings of dimension 1, and for tilings of dimension
space.
Date: 2 September 2018.
2010 Mathematics Subject Classification. 52C23, 46L80, 37D15.
Key words and phrases. Aperiodic tiling; Computing K-theory; C∗-algebra; Groupoid; Smale
∗ Partially supported by CNPq.
∗∗ Supported by the Villum Foundation under the project "Local and global structures of
groups and their algebras" at University of Southern Denmark, and by CNPq at Universidade
Federal de Santa Catarina.
† Dedicated to a special friend whose presence will never be forgotten.
1
2
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
2, in the absence of torsion, we show that the cohomology and homology can be
computed as a direct limit of a matrix and its transpose. Our new method for com-
puting the K-theories of the stable and unstable C*-algebras is highly computable,
and we have implemented this new method in Mathematica. The stable-transpose
homology arguably provides a simpler method than other methods found in the
literature for computing the K-theory of the unstable C∗-algebra. We start with
some background theory.
A tiling of Rd is a subdivision of Rd into so-called tiles. These tiles can only
intersect on their boundaries, and they are homeomorphic to the closed unit disk.
There are several ways to construct these tilings, see for example [40].
In this
paper, we consider only those built via a substitution rule on a finite number of
tiles. For such tilings, one can construct a tiling space (Ω, d) and a substitution
map ω : Ω → Ω with inflation factor λ > 1. We will assume that the tiling space
has FLC (finite local complexity), and that the substitution map ω is primitive and
injective. In such case, the tiling space, which is also called the continuous hull, is
a nonempty compact metric space, and it has no periodic tilings. Moreover, ω is a
homeomorphism, and the triple (Ω, d, ω) is a topologically mixing Smale space. For
more details see [1]. Using the Smale space structure of (Ω, d, ω), one can construct
three equivalence relations on Ω, namely, Rs, Ru, Ra -- the stable, unstable, and
asymptotic equivalence relations. Each of these equivalence relations are defined as
an increasing union of subspaces of Ω× Ω with the subspace product topology, and
Rs, Ru, Ra are given the inductive limit topology. They are locally compact and
Hausdorff. Moreover, the equivalence classes [T]Rs, [T]Ru , [T]Ra of T ∈ Ω are each
dense in Ω. A characterization of the stable and unstable equivalence relations is
given in the following proposition
Proposition 1.1 (Characterization of the equivalence relations Rs, Ru). For tilings
T, T 0 ∈ Ω, the following holds
• T ∼Rs T 0 ⇐⇒ ∃ n ∈ N0: ωn(T)(0) = ωn(T 0)(0)
• T ∼Ru T 0 ⇐⇒ ∃ x ∈ Rd: T 0 = T + x,
where N0 := N ∪ {0}, and T(0) is the set of tiles in T that contain the origin.
(cf. Section 3). Recall that an equivalence relation is in particular a groupoid, and
that the definition of a groupoid C∗-algebra [33] requires a topology on the groupoid
and a Haar system for the groupoid. Rs and Ru have the already-mentioned in-
ductive limit topologies and the Lebesgue measures. The inductive limit topology
of the asymptotic equivalence relation Ra is an ´etale topology and thus Ra has the
Haar system of counting measures. Let
S := C∗
r (Rs),
A := C∗
r (Ra)
U := C∗
r (Ru),
denote their C∗-algebras, in the sense of Renault. These are all UCT, simple, and
amenable -- UCT holds for the C∗-algebra of any amenable groupoid, by J. L. Tu's
[44, Proposition 10.7 and Lemma 3.5]. Moreover, A is strongly Morita equivalent
to U ⊗ S. For more details, see [30], [31]. We should note that "the C∗-algebra of
a tiling" in the literature usually refers to the C∗-algebra U.
The K-theory of U is related to the Cech cohomology of Ω (cf. [1, Theorem 6.3]).
This cohomology can be computed in many different ways [1], [23], [25], [38], [42],
[3]. An alternative way of computing the K-groups of U is as follows: The transver-
sal equivalence relation Rpunc is an ´etale equivalence relation. Moreover, the C∗-
algebras U and C∗
r (Rpunc) are strongly Morita equivalent by Theorem 2.8 in [26],
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)3
and hence their K-theories coincide. For more details see [20], [21], [22], [24]. One
can also compute the K-theory of U for some tilings via AF-algebras [18]. In gen-
eral, there exists a homology theory for Smale spaces [28] and a duality theory for
the Ruelle algebras associated to a Smale space [19], but there is no general theory
linking the K-theory of the stable and unstable C∗-algebras associated to a Smale
space. We show nonetheless in this article that the K-theories of S and U can be
computed from the cohomology and homology of a single cochain complex with
connecting maps.
The K-theory of S is computed via a transversal to Rs as follows: Let T ∈ Ω be
a fixed tiling. The equivalence relation
(1.1)
R0
s := [
n∈N0
Rn
is equipped with the inductive limit topology, where
Rn := 1
λn R(ωn(T)) = 1
λn{(x, y) ∈ Rd × Rd ωn(T)(x) − x = ωn(T)(y) − y}
is given the subspace topology of Rd × Rd. Then R0
s is topologically groupoid
isomorphic to a transversal of Rs. Since R0
s is an ´etale equivalence relation, i.e. the
range and source maps are local homeomorphisms, we can define the C∗-algebra
S0 := C∗
r (R0
s)
in the sense of Renault. Since S is strongly Morita equivalent to S0, their K-theory
coincide. For more details see Section 3.
In this paper we only focus on tilings of the line and of the plane. So assume
that the fixed tiling T is of dimension d ∈ {1, 2}. Although many of our definitions
can be generalized immediately to higher dimensions, we will keep them in their
simplest form, for the purpose of clarity.
The equivalence relation R0 induces an equivalence relation on the cells. The
R0-equivalence classes of the vertices, edges, and faces are called the stable vertices,
stable edges, and stable faces, respectively, and there is a finite number of them.
These numbers are denoted by sV , sE, sF, respectively. It is convenient to identify
the stable cells with its representatives. This proves quite useful when doing explicit
computations. We thus make the following informal definition
Definition 1.2 (Stable cells). If v is a vertex, e an edge, f a face of the fixed tiling
T of dimension d ≤ 2, then
• stable vertex: sv := T(v).
• stable edge: se := T(◦
• stable face: sf := f.
e), where ◦
e is the edge minus its vertices.
Here T(X), for X ⊂ Rd, is the set of tiles that contain one or more points of X.
We remark that the stable vertices are collared vertices (cf. Definition 2.2).
The cochain map of abelian groups with connecting maps is given as the following
commutative diagram
(1.2)
0
0
/ ZsV
WV
/ ZsV
δ0
δ0
ZsE
WE
/ ZsE
δ1
δ1
ZsF
WF
/ ZsF
0
/ 0,
/
/
/
/
/
/
/
/
/
/
/
4
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
where the coboundary maps δ0, δ1, and the connecting maps WV , WE, WF are
given below, right after the statement of Theorem 1.5. The stable cohomology
groups for T are given by
(1.3)
/ coker δ1 WF /
/ ···
S(T) = lim→ coker δ1 WF /
H2
H1
S(T) = lim→
S(T) = lim→ ker δ0 WV /
H0
ker δ1
Im δ0
WE /
WE /
/ ker δ1
Im δ0
/ ker δ0 WV /
/ coker δ1 WF
/ ···
/ ··· .
/ ker δ1
Im δ0
/ ker δ0 WV
WE /
and the stable-transpose homology groups for T are given by
(1.4)
H ST
2
H ST
1
H ST
0
F /
(T) = lim→ ker(δ1t) W t
(T) = lim→
(T) = lim→ coker (δ0t) W t
ker(δ0t)
Im (δ1 t)
E /
W t
/ ker(δ1t) W t
F /
/ ker(δ1t) W t
F
/ ···
/ ker(δ0 t)
Im (δ1t)
W t
E /
/ ker(δ0 t)
Im (δ1t)
W t
E
/ ···
V /
/ coker (δ0t) W t
V /
/ coker (δ0t) W t
V
/ ··· .
Then by Section 2, Section 4 and Section 5 we have
Theorem 1.3 (K-theory of S). For tilings of dimension 1 or 2 the following holds
line:
K0(S) = H0
/ K0(S)
S(T)
/ H0
K1(S) = Z
K1(S) = H1
/ Z
/ 0
S(T)
plane: 0
S(T) is torsion free for
where the sequence is a short exact sequence. Moreover, H0
both dimensions. In particular, tilings of the line always have torsion free K-theory
groups.
Theorem 1.4 (K-theory of U). For tilings of the plane with convex tiles, or tilings
of the line, the following holds.
S(T).
K1(U) = Z
(T) K1(U) = H ST1
(T).
line: K0(U) = H ST0
(T)
plane: K0(U) = Z ⊕ H ST0
Moreover, H ST1
H ST0
groups.
(T) is torsion free for both dimensions. For tilings of the line
(T) is torsion free. Thus tilings of the line always have torsion free K-theory
H k
transpose! (cf. Corollary 5.27).
It is worth noting that in the absence of torsion, e.g. for tilings of dimension 1,
S(T) and H ST
k (T) are given alone by the direct limit of a single matrix and its
Since S is U CT, and A is strongly Morita equivalent to U ⊗ S, we can use the
Kunneth formula to express the K-theory of A in terms of the K-theory of U and
S (cf. Section 6)
Theorem 1.5 (K-theory of A). For tilings of dimension 1 or 2, the following holds
line: K0(A) =(cid:16)
K0(S) ⊗ K0(U)(cid:17) ⊕ Z
K1(A) = K0(S) ⊕ K0(U)
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)5
plane: If Ki(S) and Ki(U), i = 1, 2 are torsion free then
K0(A) =(cid:16)
K1(A) =(cid:16)
K0(S) ⊗ K0(U)(cid:17) ⊕(cid:16)
K0(S) ⊗ K1(U)(cid:17) ⊕(cid:16)
K1(S) ⊗ K1(U)(cid:17)
K1(S) ⊗ K0(U)(cid:17)
.
To define the coboundary maps δ0, δ1 we need to assume that the stable faces
and stable edges have orientation, e.g. one can put counterclockwise orientation on
the faces. By translation, all cells of T have orientation. For vertex v, edge e, face
f of T let
(1.5)
δe,v :=
(1.6)
δf,e :=
−1,
0,
Then δ0 : ZsV → ZsE is given by
(1.7)
and δ1 : ZsE → ZsF is given by
(1.8)
v is the initial vertex of e
v is the final vertex of e
else
e ∈ ∂f same orient.
e ∈ ∂f opps. orient.
else
1,
0,
−1,
1,
δ0([v]) := X
δ1([e]) := X
e∈T (v)
f∈T (◦
e)
δe,v [e]
δf,e [f] = [f1] − [f2],
where f1 and f2 are the faces in T that contain the edge e, and such that e and the
corresponding edge in f1 have same orientation.
To define the connecting maps, we need a homotopy h on the prototiles which
extends to Rd by translation of the prototiles (cf. Definition 4.6). For edges e ∈ T,
e0 ∈ ω(T), let
1,
h1( 1
−1, h1( 1
0,
else
δe,e0 :=
λ e0) = e (matching orientation)
λ e0) = e (opposite orientation)
Then the connecting maps, which are also named substitution-homotopy matrices,
are given by
WF ([f])
WE([e])
WV ([v])
f0∈ω(f)
λ f0)=f
h1( 1
:= X
:= X
:= X
e0∈ω(T (◦
e))
λ e0)=e
h1( 1
v0∈ω(T (v))
λ v0)=v
h1( 1
[f0]
δe,e0 [e0]
[v0].
Compare these formulas with their unstable counterparts given in Section 2. (The
reason that we use the letter W is that it looks similar to ω).
6
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
λ], Z[
In Section 7 we computed the K-theory of a number of tilings of dimension 1
and dimension 2 using the above formulas. The results are listed in Tables 1, 2, 3,
4. It is interesting to note that in all of our examples of dimension 1, we had the
equality Ki(U) = Ki(S), i = 0, 1. Some of these groups are direct sums of groups of
the form Z[ 1
1
det A], where λ denotes the eigenvalues of a matrix A coming from
the substitution matrices. But we also include one example, namely, the tiling that
we named Pathologic, whose stable and unstable K0-groups are rank-3 subgroups
of Z ⊕ Z[1/17]2. Here it is impossible to write the rank-2 subgroup of Z[1/17]2
nontrivially as a direct sum! Yet those two K0-groups are isomorphic. For tilings
of the plane, it is worth noting that the torsion of the stable and unstable K-groups
occurred precisely in the groups predicted by Corollary 5.26.
We are in the process of extending this article to higher dimensions using spectral
sequences. It is also our hope that we can remove the homotopy.
s on Rd.
The paper is organized in the following way: In the next section we summarize a
description of the tiling space as an inverse limit using an alternative definition of
the so-called collared tiles. Moreover, we present some remarks on cohomology. In
Section 3, we prove that the stable equivalence relation Rs has a transversal which
is isomorphic to the inductive equivalence relation R0
In Section 4, we
introduce the C∗-algebras C∗(Rn), and calculate their K-theory groups in terms of
the compacts. Then we introduce a homotopy of the tiling and use it to construct
connecting maps which we show are ∗-homomorphisms. We calculate the generators
of the K-theory groups and formulas for the connecting maps in K-theory, and
the K-theory of S.
In Section 5, we show a relationship between the K-theory
of the stable C∗-algebra S (as constructed in this paper) and the K-theory of the
unstable C∗-algebra U via Cech cohomology, PE-cohomology, PE-homology, stable-
transpose homology and stable homology, plus the homotopy introduced in Section
4. Moreover, we use a universal coefficient theorem to derive properties of the
stable cohomology and the stable-transpose homology. In Section 6 we investigate
the asymptotic C∗-algebra A and show that the stable C∗-algebra S is amenable.
In Section 7 we give explicit calculations of the K-theory for a number of tilings
of the line and of the plane. In Appendix A, we provide methods for calculating
the direct limits of automorphisms of finitely generated free abelian groups.
In
particular, we provide formulas to compute the kernel, cokernel, and homology of
integer matrices. Such formulas are based on the Smith normal form of integer
matrices, and they can easily be implemented for example in Mathematica. For
convenience to the reader we provide in the rest of the appendix some background
results on topological spaces and C∗-algebras, which we assume are well-known.
This paper is a continuation of the work initiated by the first author in the
papers [15], [16]. On the one hand, we provide detailed proofs not found in those
papers, for example that R0
s is isomorphic to a transversal of Rs. We furthermore
give a solid foundation to the concept of a homotopy defined on a tiling, which
can be used to define ∗-homomorphisms called connecting maps. These connecting
maps are used to give explicit formulas for the K-theory of the stable C∗-algebra.
We should note that the idea of the homotopy was borrowed from an example in
[16]. Of course, we avoid repeating results from [15], [16], but in a few instances we
provide alternative short proofs using machinery found in the literature for the full
C∗-algebra (instead of the reduced C∗-algebra as was done in [15], [16]). Moreover,
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)7
we work mostly with connected components of Rn, instead of working directly with
Rn, in order to make the proofs simpler.
On the other hand, we show the results mentioned in the introduction such as
a relationship between the K-theories of the stable and unstable C∗-algebras, and
we provide interesting examples and formulas to compute the K-theory groups of
the C∗-algebras S, U and A.
2. Preliminaries
We start with some remarks on the definition of a tiling T of Rd. On the one
hand, T is a CW-complex with underlying space Rd, such that the closed d-cells are
homeomorphic to the closed unit disk. The closed d-cells are called tiles, and we
allow the possibility that our tiles carry labels. On the other hand, we will consider
T to be the collection of its tiles, which is a cover of Rd satisfying that two tiles
can only intersect on their boundaries. (cf. [1, p. 6]). More generally, we consider
the closed k-cells of T as a cover of the k-skeleton of T, where k = 0, . . . , d. We
will always assume that the cells of T are closed.
The tiling space Ω is the set of all the tilings whose tiles are translated copies of
a finite (fixed) set of prototiles t1, . . . , tN, such that if T ∈ Ω, and P ⊂ T is a patch
(i.e. a finite subset) then P is contained in a translated copy of ωk(tj) for some
k ∈ N, j ∈ {1, . . . , N}. Since the equivalence class [T]Ru is dense in Ω, it holds as
well that
Ω = T + Rd
d
i.e. Ω is the completion under the metric d of the set of all the translations of T.
Moreover, one can also write the tiling space as an inverse limit
(2.1)
where Γ := Ω×Rd
Definition 2.1 (Collared equivalence relation Rc). Two pointed tilings (T, x),
(T 0, x0) ∈ Ω × Rd are said to be collared equivalent (or Rc-equivalent) if
, with Rc being given as follows:
Ω = lim← Γ o
··· ,
Γ o
Γ o
Rc
T(σ) − x = T 0(σ0) − x0,
where σ ∈ T (resp. σ0 ∈ T 0) is the (necessarily unique) closed cell whose interior
contains x (resp. x0).
We remark that by a pointed tiling we just mean a pair consisting of a tiling and
a point. It is easy to see that Rc is indeed an equivalence relation, and so there
is no need to take the transitive closure. All the proofs in [1, Section 4] carry out
almost immediately using Rc instead of the relation ∼1 defined in that section, and
thus Eq. (2.1) holds (using Rc instead of ∼1). We should remark that our definition
of Rc is equivalent to the one in [13, Definition 2.2] and they also show Eq. (2.1)
using their definition. Since Cech cohomology behaves well under inverse limits we
get
H(Ω) = lim→
(2.2)
and since Γ is a finite CW-complex, H k(Γ) = H k
much easier to compute.
/ H k(Γ)
H k(Γ)
/ H k(Γ)
/ ··· ,
cell(Γ), and cellular cohomology is
o
o
o
o
o
o
o
o
o
/
/
/
8
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
The equivalence relation Rc induces an equivalence relation on the cells. The Rc-
equivalence classes of the vertices, edges, and faces are called the collared vertices,
collared edges, and collared faces, respectively, and there is a finite number of
them. These numbers are denoted by cV , cE, cF, respectively. We make though
the following informal definition, just as we did for their stable counterparts.
Definition 2.2 (Collared cells). If v is a vertex, e an edge, f a face of a tiling T,
then
• collared vertex: cv := T(v).
• collared edge: ce := T(e).
• collared face: cf := T(f).
(i.e. cv is the set of tiles that contain the vertex v).
(i.e. ce is the set of tiles that contain the edge e or its vertices).
(i.e. cf is the set of tiles that share an edge or vertex with the tile f).
Using that Γ is a finite CW-complex, we can define the following commutative
diagram
(2.3)
0
0
/ ZcV
ωt
V
/ ZcV
∂t
1
∂t
1
ZcE
ωt
E
/ ZcE
∂t
2
∂t
2
ZcF
ωt
F
/ ZcF
0
/ 0
where the exponent t stands for the transpose, the boundary maps ∂1, ∂2 are the
standard cellular boundary maps, and the substitution matrices ωV , ωE, ωF are
given below. The Cech cohomology groups are given by
(2.4)
H2(Ω) = lim→ coker ∂t1
H1(Ω) = lim→
H0(Ω) = Z .
ker ∂t
2
Im ∂t
1
ωt
F /
/ coker ∂t1
ωt
F /
ωt
F
/ ···
/ coker ∂t1
/ ···
ωt
E
ωt
E /
/ ker ∂t
2
Im ∂t
1
ωt
E /
/ ker ∂t
2
Im ∂t
1
By Theorems 6.1, 6.3, 7.1 in [1] we get
Theorem 2.3 (K-theory of U). For tilings of dimension 1 or 2, the following
holds.
line: K0(U) = H1(Ω)
plane: K0(U) = Z ⊕ H2(Ω) K1(U) = H1(Ω).
To define the boundary maps ∂1, ∂2, we need to assume that the prototiles and
the collared edges have orientation. By translation, all the cells of T ∈ Ω have
orientation.
The boundary map ∂1 : ZcE → ZcV is given by
K1(U) = Z
and the boundary map ∂2 : ZcF → ZcE is given by
∂1([e]) := X
∂2([f]) := X
v∈e
T (v)⊂T (e)
e∈f
T (e)⊂T (f)
δe,v [v],
δf,e [e],
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)9
where δe,v and δf,e are defined in Eq. (1.5) and Eq. (1.6).
The collared substitution matrices are given by
f0∈ω(f)
T (f0)⊂ω(T (f))
ωF ([f]) := X
ωE([e]) := X
ωV ([v]) := X
e0∈ω(e)
T (e0)⊂ω(T (e))
v0∈ω(v)
T (v0)⊂ω(T (v))
[f0],
[e0],
[v0].
2.1. Cohomology notation. Given a cochain complex
C• := ··· δ−1
/ C0
δ0
/ C1
δ1
/ C2
δ2
/ ··· ,
we will denote the cohomology groups by
H n(C•) := ker δn
Im δn−1
n ∈ Z.
3. Dynamics
The construction of the Smale space (Ω, d, ω) for a substitution tiling, and the
dynamics for the unstable equivalence relation Ru, were developed in [1, Section 3].
In this section, we consider the stable equivalence relation Rs, describe a transversal
to it, and set up the ground for computation of the K-theory of S.
Recall that two tilings T, T 0 are said to be stable equivalent if
n→∞ d(ωn(T), ωn(T 0)) = 0.
lim
The stable equivalence class for tiling T can be written as (see [31, p.181])
(3.1)
ω−n(V s(ωn(T), ε0)),
[T]Rs = V s(T) = [
n∈N0
where V s(ωn(T), ε) is the stable local canonical coordinate of T and it is defined
in [1, p. 9] as
V s(T, ε) = {T 0 ∈ Ω [T, T 0] = T 0 and d(T, T 0) < ε}
for 0 < ε ≤ ε0.
Moreover, the constant ε0 < 1√
2 is fixed, and it depends on the inflation factor λ
and on some separation constant of the diagonal of R0 to the other components
ε0 − ε0 > 0. We
of R0, and it is defined right after Lemma 3.2 in [1]. Note that 1
remark that the union in (3.1) is an increasing union (see [31, p.181])
V s(T, ε0) ⊂ ω−1(V s(ω(T), ε0)) ⊂ ω−2(V s(ω2(T), ε0)) ⊂ ··· .
The equivalence relation Rs is considered as a topological groupoid as follows.
Define
0 ⊂ Gs
Gs
1 ⊂ ···
by
Gs
0
Gs
n
:= {(T, T 0) ∈ Ω × Ω T 0 ∈ V s(T, ε0)}
:= (ω × ω)−n(Gs
0),
/
/
/
/
10
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
and equip Gs
n ⊂ Ω × Ω with the subspace topology, and give the increasing union
Rs = [
Gs
n
n∈N0
the inductive limit topology. A simple characterization of the stable equivalence
relation is
Proposition 3.1. Two tilings T, T 0are stable equivalent if and only if their tiles
containing the origin agree after a finite number of substitutions. That is,
T ∼Rs T 0 ⇐⇒ ∃ n ∈ N0 : ωn(T)(0) = ωn(T 0)(0).
Proof. (⇒). By [1, p. 9], we have
V s(T, ε) ⊂ {T 0 ∈ Ω T and T 0 agree on B 1
ε −ε(0)},
for any 0 < ε ≤ ε0.
(We would like to point out that we replaced the strict
inequality ε < ε0 given in [1, p. 9] with ε ≤ ε0. This can be done since a Smale
space with constant ε0 is also a Smale space with smaller constant ε0. We work
with 0 < ε ≤ ε0, and we rename ε0 as ε0.) Recall that two tilings T, T 0 ∈ Ω are
said to agree on a set U ⊂ Rd if T(U) = T 0(U), where T(U) denotes the smallest
patch in T containing the set U. (cf. [1, p. 4]). Suppose that T 0 ∈ [T]Rs. Then by
Eq. (3.1) there exists a n ∈ N0 such that
ωn(T 0) ∈ V s(ωn(T), ε0).
Thus ωn(T 0) and ωn(T) agree on a ball of radius 1
In particular they agree at the origin, i.e.
ε0 − ε0 > 0 centered at the origin.
ωn(T 0)(0) = ωn(T)(0).
(⇐). Suppose that ωn(T)(0) = ωn(T 0)(0) for some n ∈ N0, and define the patch
P := ωn(T)(0). Let r := d(0, ∂P) be the largest radius such that Br(0) ⊂ P. Note
that 0 is an interior point of P, and hence r > 0. Then for k ∈ N, ωn+k(T) and
ωn+k(T 0) agree on the patch ωk(P). Thus they agree on the ball of radius λkr.
Hence
d(ωn+k(T), ωn+k(T 0)) ≤ 1
λkr
→ 0
for k → ∞.
Thus T and T 0 are stable equivalent.
(cid:3)
By [30] the unstable equivalence class [T]Ru is dense in Ω and [T]Ru = T + Rd
is a transversal to Ω with respect to Rs. We elaborate on this in the rest of the
paragraph. The unstable equivalence class [T]Ru is considered as a topological
space as follows. Define
n (T) := ωn(V u(ω−n(T), ε0)) = T + B2ε0λn(0) ⊂ Ω,
V u
where the equality is given in [1, p.9], and the set V u(T, ε0) is known as the unstable
local canonical coordinate of T. By [1, p.10], we can write
[T]Ru = V u(T) = [
n∈N0
n (T) = T + Rd.
V u
(3.2)
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)11
n (T) ⊂ Ω with the subspace topology, and give the increasing union in
Equip V u
Eq. (3.2) the inductive limit topology. Define the equivalence relation
Hs(T)
:= Rs ∩(cid:0)(T + Rd) × (T + Rd)(cid:1)
= {(T − x, T − y) ∈ Rs x, y ∈ Rd}
= {(T − x, T − y) ∈ Ω × Ω ∃n ∈ N0 : ωn(T − x)(0) = ωn(T − y)(0)}.
By [30, Theorem 4.2, Theorem 3.7] Hs(T) is groupoid equivalent to Rs in the sense
of [26]. Moreover, by [30, Theorem 3.7] Hs(T) is ´etale. Note that the set T + Rd is
a (generalized) transversal to Ω with respect to Rs since the set [T 0]Rs ∩ (T + Rd)
is countable for any T 0 ∈ Ω.
By [26, Theorem 2.8] the C∗-algebras C∗(Hs(T)) and C∗(Rs) are strongly Morita
equivalent, and hence their K-theories coincide. It is worth noting that T +Rd ⊂ Ω
has been given the inductive limit topology in Eq. (3.2) and not the subspace topol-
ogy of Ω. Moreover, Hs(T) is equipped with the topology stated in [30, Lemma 3.3].
Convergence of sequences in Hs(T) can be stated as follows by [30, p.10]:
(T − xn, T − yn) →Hs(T ) (T − x, T − y)
if and only if
(T − xn, T − yn) →Rs (T − x, T − y)
and T − xn →T +Rd T − x
and T − yn →T +Rd T − y.
Proposition 3.2. Equip T + Rd with the inductive limit topology as in Eq. (3.2).
Then the map Rd → T + Rd given by
x 7→ T − x
is a homeomorphism.
Proof. The map α : Rd → T + Rd defined by
α(x) := T − x
is bijective because T − x = T − y implies x = y as the tiling T is aperiodic.
Moreover, α is continuous since d(T − x, T − y) ≤ 1
2x − y, where the inequality
follows since T −x+ x−y
2 agree everywhere. Let Bn := B2ε0λn(0) ⊂
Rd, n ∈ N0 be the open ball of radius 2ε0λn. Since Bn ⊂ Rd is bounded and α is
continuous, the restriction map
2 and T −y− x−y
αn := α : Bn → T + Bn
is a homeomorphism, where Bn and T + Bn have the subspace topology of Rd and
Ω, respectively. (It is the restriction of the homeomorphism ¯αn : Bn → T + Bn
defined in the same way). Then, since the following diagram commutes,
we get that
B0
α0 ∼=
T + B0
Rd = [
B1
α1 ∼=
/ T + B1
∼= [
B2
α2 ∼=
/ T + B2
T + Bn = T + Rd,
Bn
n∈N0
n∈N0
/
/
/
/
/
/
/
/
/
/
12
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
where the first equality is by Lemma B.2.
(cid:3)
We should remark that the above proposition would be false if we had equipped
T + Rd with the subspace topology of Ω, as shown in Proposition E.1.
Using Proposition 3.2, we can "see" the transversal Hs as an equivalence relation
on Rd. We provide the details below.
3.1. Equivalence relation R0
the associated equivalence relation on Rd given by
s on Rd. For a tiling T of Rd, d ∈ N, let R(T) be
R(T) := {(x, y) ∈ Rd × Rd T(x) − x = T(y) − y}.
Let T ∈ Ω be a fixed tiling, and define the sequence of tilings with shrunk prototiles
and equivalence relations
Tn := 1
λn ωn(T)λn,
Rn := R(Tn) = 1
λn R(ωn(T)),
n ∈ N0
n ∈ N0,
where ω is the substitution map with inflation factor λ > 1. Note that Tn, as a
function on Rd, is given by
It is easy to check that
Tn(x) := 1
for x ∈ Rd.
λn (ωn(T)(λnx)),
R1 ⊂ R2 ⊂ R3 ⊂ . . . ,
so we can define the equivalence relation R0
s on Rd by
∞[
n=0
R0
s :=
Rn.
We equip Rn ⊂ R2d with the subspace topology, and R0
s with the inductive limit
topology, which, as we show next, is ´etale. Since Rn is an equivalence relation on
Rd, it is in particular a groupoid with unit space Rd. Every connected component
C in Rn is homeomorphic onto its image in Rd via both the range r and source s
maps given by r(x, y) := x and s(x, y) := y. In particular, we have that r and s
are local homeomorphisms. Hence Rn is ´etale. Moreover, r(C) and s(C) are both
open in Rd. It follows that Rn is open in Rn+1: if x ∈ Rn, and Cn (resp. Cn+1) is
the connected component of x in Rn (resp. Rn+1), then r(Cn) is open in r(Cn+1)
since r(Cn) is open in Rd and Cn ⊂ Cn+1. Thus, Cn is open in Cn+1 because
r : Cn+1 → r(Cn+1) is a homeomorphism. As a consequence we have that R0
s is
s then x ∈ Rn for some n ∈ N0. As above, let Cn be the connected
´etale: If x ∈ R0
component of x in Rn. Since Rn is open in Rk for k ≥ n and since R0
s has the
inductive limit topology, then Cn is open in R0
s. Furthermore, since Rn is open in
Rn+1, Cn as a subspace of Rn has the same topology as Cn as a subspace of R0
s.
Since r and s restricted to Cn (as a subspace of Rn and hence also as a subspace
of R0
s) are homeomorphisms onto their images in Rd, we see that r and s are local
homeomorphisms on R0
s.
It is convenient to write Rn in different ways:
Rn = R(Tn)
= {(x, y) ∈ Rd × Rd Tn(x) − x = Tn(y) − y}
= {(x, y) ∈ Rd × Rd ωn(T − x)(0) = ωn(T − y)(0)}.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)13
Moreover, Rn induces an equivalence relation on the cells of the shrunk tiling Tn,
namely
Definition 3.3 (Stable cells). Let n ∈ N0 be fixed. If v is a vertex, e an edge, and
f a face of the tiling Tn, then we define the equivalence relations:
• vertices: v ∼Rn v0 ⇐⇒ ∃x ∈ Rd : Tn(◦
v0) = Tn(◦
e0) = Tn(◦
• edges: e ∼Rn e0 ⇐⇒ ∃x ∈ Rd : Tn(◦
e) + x
◦
◦
• faces: f ∼Rn f0 ⇐⇒ ∃x ∈ Rd : Tn(
f0) = Tn(
f) + x,
σ denotes the interior of a cell σ. We remark that the interior of a vertex is
v) + x
where ◦
by convention the vertex itself.
The number of equivalence classes of k-cells, k = 0, 1, 2, is always finite and
independent of n, and is denoted by sV, sE, sF, respectively. In this paper we are
◦
only considering tilings of dimension d ≤ 2, and thus we have that Tn(
f) = f. The
equivalence classes of vertices, edges, and faces are called stable vertices, stable
edges, and stable faces, respectively, but informally it is more convenient to define
them as follows
Definition 3.4 (Stable cells). Let n ∈ N0 be fixed. If v is a vertex, e an edge, f a
face of tiling Tn, then
• stable vertex: sv := Tn(◦
v) = Tn(v).
• stable edge: se := Tn(◦
e), where ◦
◦
• stable face: sf := Tn(
f) = f.
e is the edge minus its vertices.
Note that 1
λn ωn(T) and Tn are the same as tilings, i.e. one can draw Tn by first
1
λn . Also
drawing the tiling ωn(T) and then shrinking all its cells by the factor
λn ωn(T)(0) = Tn(0), but in general 1
1
s - Hs relationship. Given a fixed tiling T ∈ Ω, the following proposition
3.2. R0
shows that the transversal equivalence relation Hs := Hs(T) is groupoid isomorphic
to the equivalence R0
s.
Proposition 3.5. The map R0
λn ωn(T)(x) 6= Tn(x) for x ∈ Rd, n ∈ N.
s → Hs given by
(x, y) 7→ (T − x, T − y)
is a topological groupoid isomorphism.
Proof. By the proof of Proposition 3.2, the map α : Rd → T + Rd given by α(x) :=
T − x is bijective. Since
x ∼R0
s
y ⇐⇒ α(x) ∼Hs α(y),
s → Hs given by
α induces the (non-topological) groupoid isomorphism φ : R0
φ(x, y) := (T−x, T−y). It remains to show that φ is a homeomorphism. We start by
showing that ψ := φ−1 is continuous. Suppose that (T −xk, T −yk) → (T −x, T −y)
in Hs. That is, suppose that
(T − xk, T − yk) → (T − x, T − y) in Rs
and
T − xk → T − x in T + Rd (ind)
14
and
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
T − yk → T − y in T + Rd (ind),
where 'ind' stands for convergence in the inductive limit topology of T + Rd. By
Proposition B.1, there exists a n0, k0 ∈ N0 such that
(T − x, T − y), (T − xk, T − yk) ∈ Gs
n0 ,
k ≥ k0.
n ∩(cid:0)(T + Rd) × (T + Rd)(cid:1), n ∈ N0.
G0
n := Gs
Define the set
Since also (T − xk, T − yk) ∈ (T + Rd) × (T + Rd), then
n0 ⊂ G0
(T − x, T − y), (T − xk, T − yk) ∈ G0
ε0 − ε0 and
where ε1 := 1
n0,max := {(T − x, T − y) ∈ (T +Rd)2 Tn0 − x and Tn0 − y agree on B 1
G0
n0,max) ⊂ Rn0 as sets. Thus
Since ε1 > 0 (as ε0 < 1√
k ≥ k0,
n0,max,
2), we get (α × α)−1(G0
(x, y), (xk, yk) ∈ Rn0 ,
k ≥ k0.
ε1λn (0)}.
Now, since by Proposition 3.2
xk = α−1(T − xk) → α−1(T − x) = x
and
in Rd, and Rn0 ⊂ R2d has the subspace topology, then, by Lemma B.3,
yk = α−1(T − yk) → α−1(T − y) = y
(xk, yk) → (x, y) in Rn0
s, because the inclusion Rn0 ,→ R0
We now show that φ is continuous. Suppose that (xk, yk) → (x, y) in R0
and hence in R0
s is continuous. This completes
the proof that ψ is continuous.
s. We
need to show that (T − xk, T − yk) → (T − x, T − y) in Hs. That is, we need to
show the following three statements
(3.3)
(3.4)
(3.5)
Note that R0
continuous. By this and since (xk, yk) → (x, y) in R0
T − xk → T − x in T + Rd (ind),
T − yk → T − y
in T + Rd (ind).
s ,→ Rd × Rd is continuous since all the inclusions Rn ,→ Rd × Rd are
(T − xk, T − yk) → (T − x, T − y)
in Rs,
xk → x
and
yk → y
s, we get that
in Rd.
Hence, Eq. (3.4) and Eq. (3.5) hold by Proposition 3.2. It remains to show Eq. (3.3).
By Proposition B.1, there exists m0, k0 ∈ N0 such that
and by Lemma B.3 (xk, yk) → (x, y) in Rm0. Let
(x, y), (xk, yk) ∈ Rm0
k ≥ k0,
min := [
G0
n∈N0
G0
n,min,
where
n,min := {(T − x, T − y) ∈ (T + Rd)2 Tn − x and Tn − y agree on B 1
G0
ε0λn (0)}.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)15
For n ∈ N0, give G0
n,min the subspace topology of (T +Rd)×(T +Rd), where T +Rd
has the inductive limit topology. By a proof similar to the one of Proposition 3.1,
0 ∈ N0 such that (T − x, T − y) ∈
we have Hs = G0
min as sets. Hence there is an n0
G0
0,min. Let
n0
n0 := max(n0
0 + 1, m0 + 1),
G be the connected component of (T − x, T − y) in G0
and let C n0
R be
the connected component of (x, y) in Rn0. Before we proceed we need the following
auxiliary result
Lemma 3.6. Let α : Rd → T + Rd be the homeomorphism from Proposition 3.2.
G be the connected component of (T − x, T − y) in G0
Let C n
R be
n,min, and let C n
the connected component of (x, y) in Rn. Then (α × α)−1(C n
G) is a subset of C n
R.
R -interior of (α × α)−1(C n+1
Moreover, (α × α)−1(C n
G ),
n ∈ N0.
G) is a subset of the C n+1
n0,min. Let C n0
n,min) ⊂ Rn as sets. Since (α × α)−1 : G0
Proof. The proof of the lemma is illustrated in Figure 1. Since ε0 > 0, we have that
(α × α)−1(G0
n,min → Rn is continuous by
G is connected, we get that (α× α)−1(C n
G)
restriction and corestriction, and since C n
G) ⊂ C n
is connected. Thus (α × α)−1(C n
R. This proves the first statement of the
lemma. Define
G := {(T − x, T − y) ∈ (T + Rd)2 (x, y) ∈ C n
C n
= {(T − x, T − y) ∈ (T + Rd)2 (x, y) ∈ C n
R, Brn(x) ⊂ r(C n
R, d(x, ∂Rd(r(C n
R)}
R))) ≥ rn},
G is obtained from C n
ε0λn and Brn(x) is the open ball of Rd of radius rn and with center
R by removing a neighborhood
where rn := 1
x ∈ Rd. Roughly speaking, C n
of the boundary. We claim that
(3.6)
Then, since C n
(3.7)
We now show our claim. We start by showing Eq. (3.6)(⊆) : Let (x0, y0) ∈ (α ×
R and Brn(x0) ⊂ r(C n
α)−1( C n
R).
Let x00 ∈ Brn(x0) ⊂ r(C n
R ⊂ Rn. Thus
G) = (α × α)−1(G0
n,min and (α × α)−1(C n
(α × α)−1(C n
R, we get by the claim that
G) ⊂ C n
R.
G) ⊂ C n
G) ⊂ (α × α)−1( C n
G). Then (T − x0, T − y0) ∈ C n
(α × α)−1( C n
G ⊂ G0
R). Let (x00, y00) := (rCn
G. Then (x0, y0) ∈ C n
)−1(x00) ∈ C n
n,min) ∩ C n
R (as sets).
R
r0(C n+1
r0(C n+1
G )
G )
r0(C n
r0(C n
G) ⊆ r0( C n
G) ⊆ r0( C n
G)
G)
rn
rn+1
r(C n
R)
r(C n
R)
r0( C n
r0( C n
G)
G)
Figure 1. Subsets r(C n
r0 := r ◦ (α × α)−1.
R), r0(C n
G), r0(C n+1
G ), r0( C n
G) of Rd, where
16
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
y00
y0
Rn
C n
R
r(C n
R)
x0
x00
Figure 2. (x00, y00) 6∈ Rn.
(Tn − x00)(0) = (Tn − y00)(0). Since (x00, y00) ∈ C n
equalities together with the identity (cf. [1, p.9])
R, y00 = x00 + y0 − x0. The last two
(Tn − x00)(0) + x00 = Tn(x00) = (Tn − x0)(x00 − x0) + x0
give
(3.8)
(Tn − x0)(x00 − x0) = (Tn − x00)(0) + (x00 − x0)
= (Tn − y00)(0) + (y00 − y0)
= (Tn − y0)(y00 − y0).
R, the inclusion ⊆ holds.
i.e. (Tn − x0) and Tn − y0 agree on Brn(0). Thus (x0, y0) ∈ (α× α)−1(G0
(x0, y0) ∈ C n
(x0, y0) 6∈ (α × α)−1( CG), we want to show that (x0, y0) 6∈ (α × α)−1(G0
this assumption we have Brn(x0) 6⊂ r(C n
We now show Eq. (3.6)(⊇) by contraposition: Suppose that (x0, y0) ∈ C n
R). We claim that this implies that
R but
n,min). By
n,min). Since
Brn(x0) ∩ ∂r(C n
Suppose for contradiction that Brn(x0) ∩ ∂r(C n
R) 6= ∅.
R) = ∅. Then
and Brn(x0) ∩(cid:0)Rd\r(C n
Brn(x0) = (cid:0)Brn(x0) ∩ r(C n
= (cid:0)Brn(x0) ∩ r(C n
R)(cid:1) ∪(cid:0)Brn(x0) ∩(cid:0)Rd\r(C n
R)(cid:1) ∪(cid:0)Brn(x0) ∩ ∂r(C n
R)(cid:1)◦(cid:1)
R)(cid:1) ∪(cid:0)Brn(x0) ∩(cid:0)Rd\r(C n
R)(cid:1)◦(cid:1).
R) and Brn(x0) ∩(cid:0)Rd\r(C n
R)(cid:1)◦ are both open in Brn(x0)
R)(cid:1)◦ is not empty because Brn(x0) 6⊂ r(C n
R ⇐⇒ (cid:0)y − y0 = x − x0 and x ∈ r(C n
R)(cid:1)
Moreover, Brn(x0) ∩ r(C n
with the subspace topology of Rd. Since they are complement of each other, they
are clopen in Brn(x0). Moreover, Brn(x0)∩ r(C n
R) is not empty since it contains x0,
R). Hence Brn(x0)
is not connected, a contradiction, which proves our claim.
R), and let
y00 = x00 + y0 − x0. Then since r(C n
(x, y) ∈ C n
we get that (x00, y00) 6∈ C n
R) we have (x00, y00) 6∈ Rn. Thus
(Tn − x00)(0) 6= (Tn − y00)(0). Hence (Tn − x0) and (Tn − y0) do not agree on Brn(0)
We illustrate this paragraph in Figure 2. Pick x00 ∈ Brn(x0) ∩ ∂r(C n
R. Since also x00 ∈ ∂r(C n
R) is open and since
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)17
as y00 − y0 = x00 − x0 ∈ Brn(0) and by Eq. (3.8). Hence (x0, y0) 6∈ (α× α)−1(G0
and thus Eq. (3.6)(⊇) has been shown.
α)−1(C n
r(C n+1
(x0, y0) is in the interior of (α× α)−1(C n+1
G ) as subsets of C n+1
Then for any x00 ∈ Br(x0) we get by the triangle inequality that
n,min),
We will now show the second statement of the lemma. Let (x0, y0) ∈ (α ×
R) ⊂
and Brn(x0) ⊂ r(C n
R ), where the last inclusion is because Rn ⊂ Rn+1. We need to show that
R . Let r := rn − rn+1.
G). Then (x0, y0) ∈ C n
G) ⊂ (α × α)−1( C n
R ⊂ C n+1
R
Brn+1(x00) ⊂ Brn(x0) ⊂ r(C n+1
R ),
which shows that x00 ∈ r0( C n+1
nected, where r0 := r ◦ (α × α)−1. Thus
G ) and hence that x00 ∈ r0(C n+1
G ) as Br(x0) is con-
U := (rCn+1
R
)−1(Br(x0)) ⊂ (α × α)−1(C n+1
G )
is a neighborhood of (x0, y0) and is open in C n+1
R .
G
G ) ⊂ C n0
R is open, we get that (xk, yk) → (x, y) in C n0
(cid:3)
Continuation of the proof of Proposition 3.5. Since (xk, yk) → (x, y) in Rn0, and
R for k ≥ k1 for some k1 ≥ k0.
since C n0
By the lemma above, the point (x, y), which is in (α × α)−1(C n0−1
), is also in the
G ) for k ≥ k2 for
R -interior of (α × α)−1(C n0
G ), and thus (xk, yk) ∈ (α × α)−1(C n0
C n0
R for k ≥ k2.
some k2 ≥ k1. By Lemma B.3 (xk, yk) → (x, y) in (α×α)−1(C n0
Applying the homeomorphism α × α we get (T − xk, T − yk) → (T − x, T − y) in
G for k ≥ k2. Since the inclusion f : T + Rd → Ω is injective and continuous
C n0
by [30, Lemma 4.1, p.16], where T + Rd has the inductive limit topology, then
the restriction f × f : C n0
G → Gn0 is continuous, and hence (T − xk, T − yk) →
(T − x, T − y) in Gn0 for k ≥ k2. Since the inclusion Gn0 ,→ Rs is continuous, we
get (T − xk, T − yk) → (T − x, T − y) in Rs.
(cid:3)
Corollary 3.7. Let α : Rd → T + Rd be the homeomorphism from Proposition 3.2.
For n ∈ N0 let
Hn
:= (α × α)(Rn)
= {(T − x, T − y) ∈ (T + Rd) × (T + Rd) (Tn − x)(0) = (Tn − y)(0)}.
Give Hn the subspace topology of (T +Rd)×(T +Rd), where T +Rd has the inductive
limit topology. Equip the increasing union
with the inductive limit topology. Then
H0 ⊂ H1 ⊂ ···
as topological spaces.
Proof. Let
Hs = [
s := [
H00
n∈N0
n∈N0
Hn
Hn.
Clearly Hs = H00
s as sets. Since φn : Rn → Hn given by the restriction
φn := α × α : Rn → Hn
18
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
is a homeomorphism, and since the following is a commutative diagram
R0
φ0 ∼=
R1
φ1 ∼=
/ H1
R2
φ2 ∼=
/ H2
we get
H0
∼= R0
Since the map from Hs to H00
s is the identity map, which by the above expression
(cid:3)
is a homeomorphism, the corollary follows.
∼= lim→ Hn = H00
s = lim→ Rn
Hs
s .
4. Stable C∗-algebra S
In this section we build a cochain complex with connecting maps from which we
define the so-called stable cohomology. We describe the K-theory of S in terms of
this stable cohomology.
For fixed tiling T ∈ Ω, recall the definition of the tiling Tn with shrunk prototiles
given by
That is Tn(x) := 1
Tn := 1
n ∈ N0.
λn ωn(T)λn
λn (ωn(T)(λnx)) for x ∈ Rd.
Rn := {(x, y) ∈ R2d Tn(x) − x = Tn(y) − y}.
Recall as well the associated equivalence relation Rn (for Tn) on Rd given by
Then the increasing union
R0
s :=
∞[
n=0
Rn
is an ´etale equivalence relation, where Rn ⊂ R2d is given the subspace topology and
R0
s the inductive limit topology. For more details see subsection 3.1. That subsec-
tion also comments on the connected components of Rn, which play an important
role in some of the proofs of this section.
i.e. if x ∈ X Tn
k and x ∼Rn y then y ∈ X Tn
s(r−1(X Tn
k be the k-skeleton of Tn, n ∈ N0, k = 0, . . . , d. Then X Tn
k , or more succinctly,
k )) ⊂ X Tn
k ,
is Rn-invariant,
Let X Tn
k
where r and s denote the range and source map of the groupoid Rn. The restriction
Rn
X
Tn
k
k ) ∩ s−1(X Tn
:= r−1(X Tn
k ),
= {(x, y) ∈ Rn x, y ∈ X Tn
k }
n ∈ N0, k = 0, . . . , d
is a sub-equivalence relation of Rn and hence a subgroupoid of Rn with unit space
k . Moreover, the complement
X Tn
k − X Tn
X Tn
k−1, n ∈ N0,
k−1 := X Tn
k \X Tn
is also Rn-invariant, and the restriction
Rn X Tn
k −X Tn
k−1
,
n ∈ N0, k = 1, . . . d
is a sub-equivalence relation of Rn and thus a subgroupoid of Rn with unit space
k−1. Since the equivalence relations Rn are ´etale, i.e. r : Rn → Rd and
k − X Tn
X Tn
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)19
s : Rn → Rd are local homeomorphisms, by restriction these maps are still local
homeomorphisms, and hence the above sub-equivalence relations are also ´etale. We
can now define the following reduced groupoid C∗-algebras in the sense of Renault
[33].
Definition 4.1 (Groupoid C∗-algebras). For n ∈ N0, let
An := C∗
Bn := C∗
Cn := C∗
In := C∗
Jn := C∗
r (Rn)
r (RnX Tn
r (RnX Tn
r (RnX Tn
r (RnX Tn
1
)
)
0
2 −X Tn
1
1 −X Tn
0
)
).
By the following proposition the algebras Cn, In, Jn, n ∈ N0, can be written in
terms of the compacts. Using the theory for the compact operators, we can then
compute their corresponding K-theory groups. For convenience to the reader, we
provide in the appendix a short proof to the following proposition.
Proposition 4.2. [15, Prop. 3.7-9] For n ∈ N0,
(1) Cn
(2) Jn
(3) In
∼=
∼=
∼=
k=1
sVL
sEL
sFL
k=1
k=1
K('2([vk]))
C0(◦
ek, K('2([ek])))
◦
f k, K('2([fk])))
C0(
where sV, sE, sF are the number of stable vertices, stable edges, stable faces, respec-
tively, and vk, ek, fk are representatives of the Rn-equivalence classes.
Proof. Proposition F.1.
(cid:3)
Taking the K-theory of these algebras, we get (cf. [36])
Corollary 4.3. For n ∈ N0
K1(Cn) = 0.
K1(Jn) ∼= ZsE
K1(In) = 0
(1) K0(Cn) ∼= ZsV
(2) K0(Jn) = 0
(3) K0(In) ∼= ZsF
Since In, Jn and Cn are given in terms of the compacts, they are type I and
hence amenable. Using that Cn is amenable we get by [34, Remark 4.10 p.32] the
short exact sequence
(4.1)
Since the class of type I C∗-algebras is closed under extensions we see that Bn is
type I, hence amenable. Using that Bn is amenable we get by arguing as above the
short exact sequence
(4.2)
and that An is type I, hence amenable. We would like to remark that it is crucial
that the quotient algebras be amenable for the above statements to hold, otherwise,
one can find counterexamples in [34, Remark 4.10 p.32] and [43].
/ Bn
/ Bn
/ An
/ Cn
/ Jn
/ In
/ 0.
/ 0,
0
0
/
/
/
/
/
/
/
/
20
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
We are interested in computing the K-theory of the C∗-algebra (cf. Proposition
B.4)
S0 = lim→ A0
ι0
/ A1
ι1
/ A2
ι2
,
where the inclusion map ιn : An ,→ An+1, n ∈ N0, is a ∗-homomorphism by
Remark 4.24. We will first compute the abelian groups K0(An), K1(An), and then
we will construct a homotopy of ∗-homomorphisms in order to compute the group
homomorphisms K0(ιn), K1(ιn).
4.1. The K-theory groups of An. To compute the K-theory groups K0(An) and
K1(An), we need to put an orientation on the cells of the tiling. For tilings of the
plane, we can put on the faces the counterclockwise orientation, and for the edges
we can put the orientation of the vectors
(cos θ, sin θ),
π
θ ∈ (− π
2 ,
2 ].
Each short exact sequence of C∗-algebras induces in K-theory a six-term exact
sequence of abelian groups. From Eq. (4.1) we get the six term exact sequence
(4.3)
∼= ZsV
/ K0(Bn)
/ K0(Cn)
0
0
K1(Bn)
δ0
K1(Jn)
∼= ZsE.
because K0(Jn) = 0 and K1(Cn) = 0. For n = 0, the exponential map δ0 : ZsV →
ZsE is computed to be (cf. [16, Prop. 2.1])
(4.4)
δ0([v]) := X
δe,v [e],
e∈T (v)
where δe,v was defined in Eq. (1.5), and T(v) is the set of all tiles in T that contain
the vertex v, and e ∈ T(v) means all the edges in the patch T(v). We should
remark that we are using a different convention for δ0 than the one in [16] (one is
the negative of the other). Our δ0 corresponds to the standard differential defining
cellular cohomology. There is no ambiguity in calling δ0 : K0(Cn) → K1(Jn) and
δ0 : ZsV → ZsE with the same name, as the latter is a computation of the former
via the isomorphisms K0(Cn) ∼= ZsV, K1(Jn) ∼= ZsE. For convenience to the reader
we provide a proof of the description of the exponential map for tilings of dimension
1 in Proposition F.2.
From Eq. (4.2) we get the six term exact sequence
(4.5)
/ K0(An)
/ K0(Bn)
ZsF ∼= K0(In)
δ1
K1(Bn)
From Eq. (4.3), we get
K1(An)
because K1(In) = 0, where δ1 is the index map.
K0(Bn) ∼= ker δ0
ZsE
K1(Bn) ∼=
Im δ0
(4.6)
0.
/
/
/
/
/
/
O
O
o
o
o
o
o
o
/
/
/
/
O
O
?
_
o
o
o
o
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)21
Thus, Eq. (4.5) can be rewritten as
(4.7)
ZsE ∼= K1(Jn)
δ1
q
ZsF
∼=
K0(In)
δ1
K1(Bn)
∼=
ZsE
Im δ0
/ K0(An)
/ K0(Bn)
∼= ker δ0 ⊂ ZsV
K1(An)
0,
where q is the quotient map, and δ1 := δ1 ◦ q. Since q is surjective, Im δ1 = Im δ1,
and therefore by Lemma A.20 we get the short exact sequence
(4.8)
0
/ ZsF
Im δ1
/ K0(An)
/ ker δ0
/ 0 .
Since ker δ0 is free abelian, the short exact sequence splits
(4.9)
Note that δ1(Im δ0) = 0, i.e.
(4.10)
From Eq. (4.7) we get
K0(An) ∼=
ZsF
Im δ1 ⊕ ker δ0.
δ1 ◦ δ0 = 0.
K1(An) ∼= ker δ1 ∼=
ker δ1
ker q
= ker δ1
Im δ0 .
(4.11)
For n = 0, the (extended) index map δ1 : ZsE → ZsF is computed in [16, Prop. 2.4]
to be
(4.12)
δ1([e]) := X
δf,e [f],
f∈T (◦
e)
where δf,e was defined in Eq. (1.6), and where T(◦
the edge e without its two vertices, and f ∈ T(◦
T(◦
e).
Lemma 4.4. For tilings of the plane,
e) means all the tiles that contain
e) means all the faces in the patch
Im δ1 = {(x1, . . . , xsF) ∈ ZsF x1 + . . . + xsF = 0}.
Proof. For each edge e,
δ1([e]) = [f1] − [f2],
where f1, f2 are the two faces adjacent to e. Thus, whenever two faces f1, f2 share
an edge, [f1] − [f2] ∈ Im δ1, i.e. [f1] ∼Im δ1 [f2]. By connectedness of the tiling, all
faces are mapped to the same element in ZsF
Im δ1 , i.e. all faces are Im δ1-equivalent,
i.e. if f1 and f2 are arbitrary faces then [f1]− [f2] ∈ Im δ1. For instance if f1f2f3 is
a chain of faces (i.e. f1 shares an edge with f2, and f2 shares an edge with f3) then
Im δ1 3 ([f1] − [f2]) + ([f2] − [f3]) = [f1] − [f3].
/
/
/
/
/
/
/
/
:
:
O
O
?
_
o
o
o
o
/
/
/
/
/
/
22
Thus
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Im δ1 = spanZ{(1,−1, 0, . . . , 0), (0, 1,−1, 0, . . . , 0), (0, . . . , 0, 1,−1)}
= {(x1, . . . , xsF) ∈ ZsF x1 + . . . + xsF = 0}.
The second equality is because
y1(1,−1, 0, . . . , 0) + y2(0, 1,−1, 0, . . . , 0) + ysF−1(0, . . . , 0, 1,−1)
= (y1, y2 − y1, y3 − y2, y4 − y3, . . . , ysF−1 − ysF−2,−ysF−1)
= (x1, x2, x3, . . . , xsF),
and thus
x1 + x2 + ··· + xsF = y1 + y2 − y1 + y3 − y2 + ··· + ysF−1 − ysF−1 = 0.
Finally, notice that from the xi's one gets the yi's and vice-versa.
Lemma 4.5. For tilings of the plane, it holds
∼= Z
(cid:3)
K0(An) ∼= Z ⊕ ker δ0
ZsF
Im δ1
ZsF
ker φ
Proof. Let φ : ZsF → Z be the linear map φ(x1, . . . , xsF) = x1+. . .+xsF. Note that
φ has matrix (1, . . . , 1). By Lemma 4.4, ker φ = Im δ1. Hence since φ is surjective,
ZsF
Im δ1 =
∼= Im φ = Z.
(cid:3)
The second equation in the lemma follows by Eq. (4.9).
4.2. Homotopy. Recall that T is a fixed tiling of dimension d, and that T0 := T,
λ ω(T)λ. Let t1, . . . , tN ∈ T be the prototiles (then for dimension d = 1,
and T1 := 1
N=sE is the number of stable edges, and for d = 2, N = sF is the number of stable
faces).
Roughly speaking, one can say that the tiles of T0 can be obtained from the
shrunk tiles of T1 simply by joining them, i.e. for a tile t ∈ T, one joins the shrunk
λ(ω(t)) ⊂ T1 to get t. We need a homotopy hs of Rd such that
tiles of the patch 1
the restriction hs : t → t homotopes a unique shrunk tile t0 ∈ 1
λ(ω(t)) to t, and
the rest of the cells will collapse cellularly, i.e. the vertices of 1
λ ω(t) collapse to
vertices of t, the edges of 1
λ ω(t) collapse to edges or vertices of t, and the faces of
λ ω(t)\{t0} collapse to edges or vertices of t. Since T is a substitution tiling, we
1
define h : ti → ti on the prototiles and extend it to the whole Rd by translation.
Of course, one has to ensure that the extension by translation to the whole Rd is
possible, for one can easily construct a homotopy on the prototiles which will not
extend by translation, e.g. if "pulling" occurs in opposite directions when gluing
together two copies of the patches 1
Formally, we assume that there is a homotopy satisfying the following assump-
tions
Definition 4.6 (Homotopy hs). Suppose that h : Rd × [0, 1] → Rd, d ≤ 2, is a
homotopy that satisfies the following items with the assumption that the homotopy
ignores the orientation of the cells, i.e. as if the cells had no orientation,
λ ω(ti).
(1) h0 = id
(2) hs, 0 < s < 1 is a homeomorphism where for each prototile ti ∈ T0
(3) h1 : T1 → T0, where for each prototile ti ∈ T0 the following holds
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)23
(a) hs : ti → ti is a homeomorphism and is cellular. Hence, in particular,
hs(v) = v for every vertex v ∈ ti, and hs(e) = e for every edge e ∈ ti.
λ ω(ti) → ti is cellular and surjective. Moreover, for every edge
λ ω(ti), if e ∈ h1(e0) then h1(e0) = e,
λ ω(ti) which homotopes to ti and such that
e ∈ ti, and every edge e0 ∈ 1
i.e. h1(e0) cannot contain more than one edge e.
(ii) there is a unique tile t0
(i) h1 : 1
i ∈ 1
◦
i → ◦
t0
ti is a homeomorphism.
(iii) for each vertex v ∈ ti, we require that h1(v) = v.
h1 :
For each edge e ∈ ∂ti, there is a unique shrunk edge e0 ∈ 1
λ ω(∂ti) that
homotopes to e. For such shrunk edge e0 we require that e0 ⊂ e, and
that h1 : ◦
λ ω(ti), and each edge e ∈ ti, if e ∈ h1(t00)
1) = e = h1(e0
then there is exactly two edges e0
2)
when t00 6= t0
i. Moreover,
h−1
1 (e) is connected.
e is a homeomorphism.
1, e0
i, and there is exactly one edge when t00 = t0
(iv) for each shrunk tile t00 ∈ 1
2 ∈ t00 with h1(e0
e0 → ◦
(4) Items (2) and (3) extend to the rest of the tiles of T0 by translation of the
scaled substituted prototiles 1
λ ω(ti + x0) = 1
λ ω(ti) + x0.
Remark 4.7. We should note that some parts of item 3 follow from item 2(a)
by continuity, but for clarity we prefer to spell them out. Note that for s ∈ [0, 1],
λ ω(ti)) is ti. Hence for any patch P ⊂ T and any
the underlying space of hs( 1
λ ω(P)) is P. For s ∈ [0, 1], the continuous
s ∈ [0, 1], the underlying space of hs( 1
map hs : T → T is by construction cellular, i.e. hs(X T1 ) ⊂ X T1 and hs(X T0 ) ⊂ X T0 ,
where X T1 and X T0 are the 1- and 0-skeleton of T, respectively. Item (3)(iii) plays
an important role in Section 5. We include item (3)(iv) in the definition of our
homotopy to simplify our proofs of Section 5 even though it is not strictly needed.
For tilings of the line, items (3)(iii) and (3)(iv), though still true, can be ignored.
Definition 4.8. Define tiling
Ts := h1−s(T1),
0 ≤ s ≤ 1.
Note that the family of tilings Ts continuously deform the tiling T1 = h0(T1) to
the tiling T0 = h1(T1). We illustrate this in Figure 3.
t0
T1
T0.5
T0.25
Figure 3. Tilings Ts =: h1−s(T1), s ∈ [0, 1].
t
T0
Since the C∗-algebra An is amenable, it is equal to the full C∗-algebra C∗(Rn).
Given a locally compact space X, we denote by Cc(X) the space of complex-valued,
continuous functions on X with compact support.
In what follows we consider
groupoids and their C∗-algebras. The results concern the full C∗-algebra C∗(G),
24
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
r (G). The goal is to construct a homotopy of ∗-
not the reduced C∗-algebra C∗
homomorphisms φs.
Remark 4.9. The fact that the φs defined below is a ∗-homomorphism relies on
the properties of our special homotopy. The construction of φs is inspired by the
fact that any groupoid homomorphism π : G → H, which is continuous and proper,
induces the map π∗ : Cc(H) → Cc(G) given by π∗(f) = f ◦ π. However, this map
is usually not a ∗-homomorphism. For instance, consider the case when G := ∆R0
is the diagonal of H := R0 for the Fibonacci tiling with proto-edges a, b. Then G
is a connected component and a subgroupoid of the ´etale equivalence relation H.
Note that the product in Cc(G) is just pointwise multiplication, and G is trivially
´etale. Let i : G → H be the inclusion map, which is obviously continuous and
proper. Then the restriction map i∗ is not a ∗-homomorphism since the product
is not preserved. For example, choose a1 = [0, 1], a2 = [1, 2] with label a, and let
f, g ∈ Cc(H) be such that f is zero outside (a1 × a2)◦ ∩ H, and f(0.5, 1.5) 6= 0, and
g is zero outside (a2 × a1)◦ ∩ H and g(1.5, 0.5) 6= 0. Then i∗(f) = f ◦ i = 0 and thus
i∗(f) ∗ i∗(g) = 0. On the other hand, i∗(f ∗ g)(0.5, 0.5) = f(0.5, 1.5)g(1.5, 0.5) 6= 0,
and hence i∗(f ∗ g) 6= i∗(f) ∗ i∗(g).
Definition 4.10 (Homotopy φs). Let φs : Cc(R0) → Cc(R0), 0 ≤ s ≤ 1, be defined
as
(cid:26) g ◦ (hs × hs)(x, y)
0
hs(x) ∼R0 hs(y)
else,
φs(g)(x, y) :=
where g ∈ Cc(R0) and x ∼R0 y.
In [14, p.92-93], it was shown that φs is well-defined as stated in the next lemma.
For the convenience of the reader, we provide an alternative proof of this lemma in
terms of connected components. Moreover, many of the ideas in the proof of the
next lemma will also be used in later proofs of this section.
Lemma 4.11. The map φs : Cc(R0) → Cc(R0), s ∈ [0, 1] is well-defined.
Proof. Since g ∈ Cc(R0) is evaluated only when hs(x) ∼R0 hs(y), the map φs(g) :
R0 → C is well-defined. For s ∈ [0, 1] we have
(4.13)
because
supp φs(g) = (hs × hs)−1(supp g)
(x0, y0) ∈ supp φs(g) ⇐⇒ g(hs(x0), hs(y0)) 6= 0 and hs(x0) ∼R0 hs(y0)
⇐⇒ (hs(x0), hs(y0)) ∈ supp g
⇐⇒ (x0, y0) ∈ (hs × hs)−1(supp g).
By assumption
K := supp(g)R0 ⊂ R0
is compact in R0. For showing that φs(g) is continuous with compact support, it
is enough to consider the restrictions gC, φs(g)C to a connected component C of
R0. This is because the support of g is in K, and by compactness, K intersects
only a finite number of connected components. Moreover, K ∩ C is compact in C
because C is clopen. Also, a compact set in C is compact in R0 by continuity of the
inclusion map C ,→ R0. Furthermore, a finite union of compact sets is compact.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)25
Since the range map r : R0 → Rd restricted to the component C is a homeomor-
phism onto its image, it is enough to consider the maps
f : r(C) → C,
s(f) : r(C) → C
φ0
given by
Note that
f := gC ◦ r−1,
s(f) := φs(g)C ◦ r−1.
φ0
• r(C) can be unbounded (e.g. r(C) = Rd when C is the diagonal of R0).
Abusing language, r(C) is a (not necessarily finite) connected patch minus
its boundary; the patch is made of tiles of T.
• For s < 1, hs maps r(C) to r(C) homeomorphically.
• h1 maps r(C) to r(C)
Rd, thus r(C)
Rd\h1(r(C)) can be nonempty. Actually
for any patch P of T, h1 :
◦
P)\ ◦
1 (∂Rd r(C)) → r(C) is surjective.
◦
P → P, and h1(
• h1 : r(C)\ h−1
P can be nonempty.
Let K be the closure of the support of f in r(C). Then, by Eq. (4.13), K0
contains the closure of the support of φ0
(4.14)
s (supp f)r(C) ⊂ h−1
h−1
s(f) in r(C) i.e.
K = supp f
r(C)
s (K) = K0
s.
s := h−1
s (K)
If s < 1, then h−1
and
s (supp f)r(C) = h−1
s (K) = K0
s is compact in r(C) because hs is
a homeomorphism of r(C) when restricted to r(C). Moreover, since
s < 1
s(f) is continuous. This completes the
s(f) = f ◦ hs,
φ0
is a composition of continuous functions, φ0
proof for s < 1.
Now suppose that s = 1. Then
(cid:26) f ◦ h1(x)
h1(x) ∈ r(C)
else
0
φ0
1(f)(x) =
We start by showing that φ0
1(f)(x) ∈ C.
1(f) is continuous. Suppose that (xn)∞
n=1 ⊂ r(C) is
a sequence that converges to x ∈ r(C). We need to show that φ0
1(f)(xn) converges
to φ0
Assume that h1(x) ∈ r(C). Then by continuity of h1 we have h1(xn) → h1(x).
Since r(C) is open, and h1(x) ∈ r(C), h1(xn) ∈ r(C) for n ≥ n0 for some n0 ∈ N.
Thus by continuity of f on r(C),
1(f)(x),
1(f)(xn) = f(h1(xn)) → f(h1(x)) = φ0
φ0
(n ≥ n0)
1(f) is continuous at h1(x) whenever h1(x) ∈ r(C).
(i.e. h1(x) is in the boundary of r(C)
which proves that φ0
Rd.) Then
Assume that h1(x) 6∈ r(C).
1(f)(x) = 0. We claim that h1(xn) 6∈ K,
h1(x) 6∈ K because K ⊂ r(C). Thus φ0
1(f)(xn) = 0 → 0 = φ0
for all n ≥ n0 for some n0 ∈ N. Hence φ0
1(f)(x), and thus
φ0
1(f) is continuous. We proceed to the claim. For contradiction, assume that the
claim is false. Then by compactness of K, there exists a convergent subsequence
h1(xnk) ∈ K. Since h1(xn) → h1(x), we have that h1(xnk) → h1(x). Since K is
closed, h1(x) ∈ K, a contradiction, which proves the claim. This completes the
proof of continuity of φ0
We will now show that K0
1 (K) is r(C)-compact.
1 = h−1
1(f).
26
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
maps to boundary
h−1
1 (K)
t0
i
K
ti
T(K)
Figure 4. h−1
1 (K) ⊂ T(K).
Since K is compact in r(C), it is compact in Rd by continuity of the inclusion
map r(C) ,→ Rd. Hence K is closed and bounded in Rd. Let P := T(K) be the
smallest patch of T that contains K, i.e. P is the necessarily finite set of tiles in
1 (K) ⊆ P, as the homotopy
T that intersect K. By definition of our homotopy, h−1
moves points from a tile of T to the tile itself. See Figure 4. More precisely, if
x 6∈ P then h1(x) 6∈ ◦
P (although it could be in the boundary of P), and since
K ⊂ ◦
1 (K)) ⊆ T(P). Hence
1 ⊂ P is bounded. Since h1 is Rd-continuous, h−1
K0
1 (K) is Rd-closed and hence
1 = h−1
K0
Since h1 is cellular, (e.g. h1 maps the Rd-boundary of r(C) to itself), and K ⊂
◦
P then h1(x) 6∈ K, i.e. x 6∈ h−1
1 (K) is Rd-compact.
1 (K). Hence T(h−1
T(K) ⊂ r(C), we have that
1 = h−1
K0
1 (K) ⊂ r(C).
This is because
h1(∂r(C)) ⊂ ∂r(C)
⇔ ∀x ∈ r(C) : x ∈ ∂r(C) =⇒ h1(x) ∈ ∂(r(C))
⇔ ∀x ∈ r(C) : h1(x) 6∈ ∂(r(C)) =⇒ x 6∈ ∂(r(C))
⇔ ∀x ∈ r(C) : h1(x) ∈ r(C) =⇒ x ∈ r(C)
⇔ ∀x ∈ r(C) : x ∈ h−1
1 (r(C)) =⇒ x ∈ r(C)
⇔ h−1
1 (r(C)) ⊂ r(C),
1 is contained in r(C). By Lemma C.2, K0
where the boundary and closure are done in Rd. Thus we have shown that K0
Rd-compact and that K0
1 is
1 is r(C)-compact.
(cid:3)
Remark 4.12. Using the notation from the above proof (of Lemma 4.11), the
i ∈ T1 and
map φ0
expands continuously (cylindrically) the graph of f on the boundary ∂ti to the set
ti\t0. We illustrate this in Figure 5.
1(f) shrinks the graph of f on prototile ti ∈ T to the shrunk tile t0
We next introduce the auxiliary map ψ whose extension (see Lemma 4.18) is
used in the proof of Proposition 4.20 given later in this section. The proof of the
next lemma is similar to the one of Lemma 4.11.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)27
Figure 5. (i) φ0
1(f), (ii) f.
Lemma 4.13. The map ψ : Cc(R0) → Cc([0, 1] × R0) given by
(4.15)
(cid:26) f(hs(x), hs(y)) hs(x) ∼R0 hs(y)
ψ(f)(s, (x, y)) :=
else
0
0
is well-defined.
Proof. Recall that the product of two groupoids is a groupoid, and [0, 1] is the
trivial groupoid where each x ∈ [0, 1] is a unit. As explained in the proof of Lemma
4.11, it is enough to consider the restriction to a connected component C of R0
via the range map r : R0 → Rd, because a R0-compact set can only intersect a
finite number of connected components, and because the range map restricted to
a connected component is a homeomorphism onto its image. Thus, it suffices to
show that ψ0 : Cc(r(C)) → Cc([0, 1] × r(C)) given by
(cid:26) f(hs(x)) hs(x) ∈ r(C)
ψ0(f)(s, x) :=
else
is well-defined. The case s < 1 is easy as hs is a homeomorphism of r(C), so we just
show the case s = 1. We start by showing continuity of ψ0(f) for f ∈ Cc(r(C)) with
r(C). Suppose that the sequence (sn, xn) converges to
compact support K = supp f
(1, x) and that h(1, x) ∈ r(C). Then by continuity of h we have h(sn, xn) → h(1, x)
for n → ∞. Since r(C) is open and h(1, x) ∈ r(C), h(sn, xn) ∈ r(C) for n ≥ n0,
for some n0 ∈ N. Then by continuity of f on r(C), we get
ψ0(f)(sn, xn) = f(h(sn, xn)) → f(h(1, x)) = ψ0(f)(1, x),
(n ≥ n0).
Thus ψ0(f) is continuous at (1, x) whenever h(1, x) ∈ r(C). Now assume that
the sequence (sn, xn) converges to (1, x), and that h1(x) 6∈ r(C).
(i.e. h1(x) ∈
∂Rd(r(C))). Thus h(1, x) 6∈ K because K ⊂ r(C). We claim that h(sn, xn) 6∈ K for
n ≥ n0 for some n0 ∈ N. Then
ψ0(f)(sn, xn) = 0 → 0 = ψ0(1, x),
(n ≥ n0),
and hence ψ0(f) is continuous at (1, x) whenever h1(x) 6∈ r(C). We now prove the
claim. Suppose that it is false, i.e. suppose that for all n0, there exists an n ≥ n0
such that h(sn, xn) ∈ K. Since K is compact, there exists a convergent subsequence
h(snj , xnj) → h(1, x) ∈ K, which contradicts h(1, x) 6∈ K.
28
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
(cid:3)
Lemma 4.14. Equip Cc(R0) and Cc([0, 1] × R0) with the inductive limit topology
(cf. [10, Example 5.10, p .118]). Then ψ : Cc(R0) → Cc([0, 1] × R0) is continuous.
Proof. As explained in the proof of Lemma 4.11, and used in the previous lemma, it
is enough to consider the restriction to a connected component C of R0 via the range
map r : R0 → Rd. Thus, it suffices to show that ψ0 : Cc(r(C)) → Cc([0, 1] × r(C))
given by
(cid:26) f(hs(x)) hs(x) ∈ r(C)
ψ0(f)(s, x) :=
is continuous in the inductive limit topology. Suppose that fn ∈ Cc(r(C)) converges
to f ∈ Cc(r(C)) in the inductive limit topology. That is, there is a compact subset
K ⊂ r(C) such that for all n ∈ N
0
else
supp fn, supp f ⊂ K,
and such that
as n → ∞.
Note that the supremum is taken over K. We claim that
fn − fsup → 0
K0 := h−1(K) ⊂ r(C)
is compact and that
Then
supp ψ0(fn), supp ψ0(f) ⊂ K0.
ψ0(fn) − ψ0(f)sup =
=
ψ0(fn)(s, x0) − ψ0(f)(s, x0)
fn(hs(x0)) − f(hs(x0))
sup
(s,x0)∈K0
sup
(s,x0)∈K0
hs(x0)∈r(C)
fn(x) − f(x)
≤ sup
x∈K
= fn − fsup → 0 as n → ∞,
where the inequality holds because we can only get nonzero values if hs(x0) ∈ K.
To check that the claim holds, notice that, by the proof of Lemma 4.11,
K0
s := h−1
s (K) ⊂ r(C),
s ∈ [0, 1].
Since
we have
K0 = h−1(K) = [
(s, x) ∈ h−1(K) ⇐⇒ hs(x) ∈ K,
s (K) = [
s∈[0,1]
{s} × h−1
s∈[0,1]
{s} × K0
s ⊂ [0, 1] × r(C).
s (supp f). Hence (s, x) ∈ {s} × K0
Moreover, supp ψ0(f) ⊂ K0 because if ψ0(f)(s, x) 6= 0 then f(hs(x)) 6= 0 and thus
hs(x) ∈ supp f, i.e. x ∈ h−1
s ⊂ K0. Similarly, it
holds that supp ψ0(fn) ⊂ K0. By definition of our homotopy K0 is Rd+1-bounded,
and by continuity of h, it is Rd+1-closed. Hence K0 is Rd+1-compact, and since it
is a subset of [0, 1] × r(C), K0 is ([0, 1] × r(C))-compact by Lemma C.2.
(cid:3)
We show next that φs is continuous, the proof of which is a "pointwise" version
of the previous proof (continuity of ψ).
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)29
Lemma 4.15. Equip Cc(R0) with the inductive limit topology (cf. [10, Ex. 5.10,p.118]).
Then φs : Cc(R0) → Cc(R0) is continuous for any s ∈ [0, 1].
Proof. As before, it is enough to consider the restriction to a connected component
C of R0 via the range map r : R0 → Rd.
Suppose that fn ∈ Cc(r(C)) converges to f ∈ Cc(r(C)) in the inductive limit
topology. That is, there is a compact subset K ⊂ r(C) such that for all n ∈ N
supp fn, supp f ⊂ K,
and such that
as n → ∞.
Note that the supremum is taken over K. We claim that
fn − fsup → 0
K0 := h−1
s (K) ⊂ r(C)
is compact and that
s(f) ⊂ K0,
s was given in the proof of Lemma 4.11. Then
s(fn), supp φ0
supp φ0
where φ0
φ0
s(fn) − φ0
s(f)sup = sup
x0∈K0
=
fn(hs(x0)) − f(hs(x0))
s(f)(x0)
φ0
s(fn)(x0) − φ0
sup
x0∈K0
fn(x) − f(x)
hs(x0)∈r(C)
≤ sup
x∈K
= fn − fsup → 0 as n → ∞,
where the inequality holds because we can only get nonzero values if hs(x0) ∈ K.
To check the claim notice that, since by Eq. (4.13) supp φ0
s (supp fn) and
supp φ0
s(fn) = h−1
s(f) = h−1
s (supp f), it is clear that
s(fn), supp φ0
supp φ0
s(f) ⊂ h−1
s (K) = K0.
Now, K0 = h−1
Lemma 4.11.
s (K) ⊂ r(C) is compact, by the same argument as in the proof of
(cid:3)
The next lemma shows that φs preserves the involution and the product. Note
that the product is convolution, not pointwise multiplication. As a corollary it will
follow that ψ is also a ∗-homomorphism.
Lemma 4.16. Let s ∈ [0, 1]. Then the map φs : Cc(R0) → Cc(R0) given in
Definition 4.10 is a ∗-homomorphism.
Proof. The ∗-operation holds because
φs(g)∗(x, y) = φs(g)(y, x) = g(hs(y), hs(x)) = φs(g∗)(x, y)
when hs(x) ∼R0 hs(y). Now
(4.16)
φs(g ? g0)(x, y) = X
φs(g) ? φs(g0)(x, y) = X
hs(x)∼R0 z0
x∼R0 z
when hs(x) ∼R0 hs(y), else the sum is zero. On the other hand,
g(hs(x), hs(z))g0(hs(z), hs(y)),
g(hs(x), z0)g0(z0, hs(y)),
D. GONC¸ ALVES
30
when hs(x) ∼R0 hs(z) and hs(z) ∼R0 hs(y), else the sum is zero. i.e.
(4.17)
φs(g) ? φs(g0)(x, y) = X
M. RAMIREZ-SOLANO
g(hs(x), hs(z))g0(hs(z), hs(y)),
hs(x)∼R0 hs(z)
x∼R0 z
when hs(x) ∼R0 hs(y) else the sum is zero.
Assume s < 1. Since hs is a homeomorphism, there is a unique z ∈ Rd such
that z0 = hs(z). Moreover, by item 2(a) in the definition of our homotopy,
hs(x) ∼R0 hs(z) if and only if x ∼R0 z. Hence φs(g ? g0) = φs(g) ? φs(g0) for
s < 1. (See picture 6).
a0
b0
a0
a0
x
hs(x)
a0
b0
b0
z
hs(z)
a
b
a
a
Figure 6. hs(x) ∼R0 hs(z) if and only if x ∼R0 z for s < 1.
T1
T1−s
T0
Now assume that s = 1. If h1(x) 6∼R0 h1(y) then both sums are zero, so assume
h1(x) ∼R0 h1(y). Define
S1 := {z0 h1(x) ∼R0 z0}
S2 := {z x ∼R0 z, h1(x) ∼R0 h1(z)}.
Let Φ : S2 → S1 be given by
Φ(z) := h1(z).
Then clearly, S1 = [h1(x)]R0, S2 = [x]R0 ∩ h−1
1
(4.18)
φ1(g ? g0)(x, y) = X
φ1(g) ? φ1(g0)(x, y) = X
(4.19)
z0∈S1
(cid:0)[h1(x)]R0
(cid:1) and
g(h1(x), z0)g0(z0, h1(y)),
g(h1(x), h1(z))g0(h1(z), h1(y)).
z∈S2
We claim that Φ is bijective. Then φ1(g ? g0) = φ1(g) ? φ1(g0) and hence φ1 is a
∗-homomorphism. We start by showing surjectivity. Suppose that z0 ∈ Rd satisfies
that h1(x) ∼R0 z0. We thus have T(h1(x)) − h1(x) = T(z0) − z0. By definition of
our homotopy we also have x ∈ T(h1(x)), and note that h1(x) ∈ T(x) ⊆ T(h1(x)),
where the inclusion holds because the map h1 is cellular. Thus z ∈ T(z0) given by
h1(x) − x = z0 − z
satisfies h1(z) = z0, since the definition of the homotopy on T(z0) is the same as
that on T(h1(x)) (up to translation). We illustrate this in the following figure.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)31
•h1(x)
•x
T(h1(x))
•z0 = h1(z)
•z
T(z0)
By construction of z, we have T(x) − x = T(z) − z, i.e. x ∼R0 z, and hence Φ is
surjective.
We now show that Φ is injective. Suppose that Φ(z) = h1(z) = z0 = h1(z) = Φ(z)
and that x ∼ z, x ∼ z, h1(x) ∼ h1(z), h1(x) ∼ h1(z). Thus we have,
T(z) − z = T(x) − x = T(z) − z.
In particular, the definition of h1 on T(z) is the same as that on T(z). By definition
of our homotopy, we must have z, z0 ∈ T(z), z, z0 ∈ T(z), from where we note that
z0 ∈ T(z) ∩ T(z). From the two last statements we get z0 − z = z0 − z, i.e. z = z.
(Compare with Figure 7.) Thus Φ is injective.
•z
•
z0
•
z
•z
•
z0
•z
h1(z)=z0=h1(z)
but
z6∼R0 z
z∼R0 z
but
h1(z)6=z0=h1(z)
Figure 7. h1(z) = z0 = h1(z) and z ∼R0 z implies z = z.
(cid:3)
Corollary 4.17. The map ψ : Cc(R0) → Cc([0, 1] × R0) given in Eq. (4.15) is a
∗-homomorphism.
Proof. Note that ψ(f)(s, (x, y)) = φs(f)(x, y) for f ∈ Cc(R0), s ∈ [0, 1], (x, y) ∈ R0.
We have ψ(f∗) = ψ(f)∗ because
ψ(f∗)(s, (x, y)) = φs(f∗)(x, y) = φs(f)(y, x),
and
ψ(f∗)(s, (x, y)) = ψ(f)(s−1, (x, y)−1) = ψ(f)(s, (y, x)) = φs(f)(y, x).
We also have ψ(f ? g) = ψ(f) ? ψ(g) because
ψ(f ? g)(s, (x, y)) = φs(f ? g)(x, y) = φs(f) ? φs(g)(x, y)
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
32
and
ψ(f) ? ψ(g)((s, s), (x, y)) = X
= X
= X
z∼x
(s0,z)∼(s,x)
ψ(f)((s, s0), (x, z))ψ(g)((s0, s), (z, y))
ψ(f)(s, (x, z))ψ(g)(s, (z, y))
φs(f)(x, z)φs(g)(z, y)
z∼x
= φs(f) ? φs(g)(x, y),
where we identify the interval [0, 1] with the "trivial" equivalence relation, where
each s ∈ [0, 1] is equivalent only to itself, i.e. [s] = {s} for s ∈ [0, 1]. We also see it
as the "trivial" groupoid where each s ∈ [0, 1] is a unit.
(cid:3)
In the next two lemmas we apply Theorem D.5 to extend the ∗-homomorphisms
ψ and φs to their completions under the full C∗-norm.
Lemma 4.18. The map ψ defined in Eq. (4.15) extends to a (continuous) ∗-
homomorphism
ψ : C∗(R0) → C∗([0, 1] × R0).
Proof. By Lemma 4.14 and Corollary 4.17, the map ψ : Cc(R0) → Cc([0, 1] × R0)
is continuous and it is a ∗-homomorphism. The statement of the lemma follows by
(cid:3)
Theorem D.5.
Lemma 4.19. The map φs : Cc(R0) → Cc(R0) in Definition 4.10 extends to
∗-homomorphism
φs : C∗(R0) → C∗(R0),
s ∈ [0, 1].
Proof. By Lemma 4.15 and Lemma 4.16, the map φ : Cc(R0) → Cc(R0) is contin-
uous and it is a ∗-homomorphism. By Theorem D.5 the lemma follows.
(cid:3)
We have now arrived to the main result of this subsection, namely that φs
gives a homotopy of ∗-homomorphisms of C∗-algebras. We should note that in the
statement of the proposition, we use the same notation for φs as for its extension.
Proposition 4.20. The map φ : [0, 1] × Cc(R0) → Cc(R0) given by
extends to a homotopy of ∗-homomorphisms
φ(s, g) := φs(g)
φs : C∗(R0) → C∗(R0),
where φs is given in Definition 4.10, Cc(R0) has the inductive limit topology, and
In particular, id ∼h φ1 (a path of ∗-
multiplication in Cc(R0) is convolution.
homomorphisms).
Proof. Let φ : [0, 1] × C∗(R0) → C∗(R0) be the map given by
φ(s, a) := φs(a),
where φs was constructed in Lemma 4.19. By Lemma 4.18,
ψ : C∗(R0) → C∗([0, 1] × R0)
is a (continuous) ∗-homomorphism. It is well known that
C∗([0, 1] × R0) ∼= C([0, 1], C∗(R0))
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)33
Via the above isomorphism, we have
ψ = (a 7→ φ(·, a))
(up to the isomorphism), and hence φ(·, a) ∈ C([0, 1], C∗(R0)) for a ∈ C∗(R0).
It follows that the map ρ : a 7→ φ(·, a) is a ∗-homomorphism from C∗(R0) to
C([0, 1], C∗(R0)). We are now ready to show that φ is continuous. Suppose that
(sn, an) converges to (s, a). Then
φ(sn, an) − φ(s, a) = φ(sn, an) − φ(sn, a) + φ(sn, a) − φ(s, a),
and
φ(sn, an) − φ(sn, a)C∗(R0) ≤ ρ(an − a)sup ≤ an − a → 0,
because ρ is bounded, and
φ(sn, a) − φ(s, a)C∗(R0) → 0,
(cid:3)
because φ(·, a) ∈ C([0, 1], C∗(R0)) and sn → s.
4.3. Connecting maps. In this subsection we do the preparations for computing
the group homomorphisms K0(ιn), K1(ιn), n ∈ N0. To accomplish this task, we
need to use the filtration of Rn by the skeletons of the tiling T, i.e. we need to use
the short exact sequences in Eq. (4.1) and Eq. (4.2). However, we cannot simply
"restrict" the inclusion map ιn to the ideal and quotient C∗-algebras of these short
exact sequences. The reason is that we would get a non-commutative diagram, as
it is going to be explained in Remark 4.21. But, one gets a commutative diagram
when one instead "restricts" the ∗-homomorphism ι0 ◦ φ1. Since φs is a homotopy
of ∗-homomorphisms, and φ0 is the identity, we get by homotopy invariance of
K-theory (cf. [36, Prop. 4.1.4, p. 61]), that Ki(ι0) = Ki(ι0 ◦ φ1), i = 1, 2. Recall
that homotopy invariance of K-theory uses the property that the homotopy is a
homotopy of ∗-homomorphisms φs, since we need the fact that φs, extended to the
unitization, preserves projections and unitaries. We only need to consider ι := ι0,
since, in the natural bases in terms of stable cells, each matrix for the K-theory of
ιn ◦ φ1 "restricted" to the ideal and quotient C∗-algebras is equal to that of ι0 ◦ φ1.
We end this subsection with Lemma 4.25, where we show the construction of a
"computable" cochain complex with connecting maps. It is "computable" since it
is given in terms of the compacts.
Similarly, from Eq. (4.2) we get the diagram (which will be shown to be com-
mutative in Lemma 4.22)
(4.20)
0
0
/ I0
α0
/ I1
A0
α
/ A1
/ B0
β
/ B1
/ 0
/ 0,
where the connecting maps α, β, α0 are defined in terms of φ1 as follows:
The map α := ι0 ◦ φ1 : A0 → A1 is given by
(cid:26) g ◦ (h1 × h1)(x0, y0)
0
α(g)(x0, y0) :=
h1(x0) ∼R0 h1(y0)
else,
/
/
/
/
/
/
/
/
/
/
/
/
/
M. RAMIREZ-SOLANO
D. GONC¸ ALVES
34
for g ∈ Cc(R0) and x0 ∼R1 y0.
The map α0 := αI0 : I0 → I1 is given by
α0(g)(x0, y0) :=
h1(x0) ∼R0 h1(y0)
else,
(cid:26) g ◦ (h1 × h1)(x0, y0)
(cid:26) g ◦ (h1 × h1)(x0, y0)
0
0
for g ∈ Cc(R0X T0
The map β : B0 → B1 is given by
2 −X T0
1
) and x0 ∼R1 y0. Note that g vanishes on the edges of T0.
β(g)(x0, y0) :=
for g ∈ Cc(R0X T0
is defined only on the edges of T0.
) and x0 ∼R1 y0 and x0, y0 ∈ X T1
1
1
h1(x0) ∼R0 h1(y0)
else,
lie in edges of T1. Note that g
From Eq. (4.1) we get the diagram (which will be shown to be commutative in
Lemma 4.22)
(4.21)
0
0
/ J0
β0
/ J1
B0
β
/ B1
/ C0
γ
/ C1
/ 0
/ 0,
where the connecting maps β0, γ are defined as follows:
The map β0 := βJ0 is given by
(cid:26) g ◦ (h1 × h1)(x0, y0)
β0(g)(x0, y0) :=
0
h1(x0) ∼R0 h1(y0)
else,
) and x0 ∼R1 y0 and x0, y0 are in the edges of T1. Note that
for g ∈ Cc(R0X T0
1 −X T0
0
g is defined only on the edges of T0 and vanishes on the vertices of T0.
The map γ is given by
γ(g)(x0, y0) :=
(cid:26) g ◦ (h1 × h1)(x0, y0)
h1(x0) ∼R0 h1(y0)
else,
0
0 are vertices of T1. Note that g
0 ) and x0 ∼R1 y0 and x0, y0 ∈ X T1
for g ∈ Cc(R0X T0
is defined only on the vertices of T0.
Remark 4.21. Roughly speaking, the connecting maps of the above two diagrams
are just the inclusion maps composed with the homotopy φ1. By the next lemma,
these diagrams commute. Had we defined the above diagrams solely in terms of
the inclusion map (without homotopy) then the restriction maps α0, β0 would not
map into the ideals I1, J1 respectively, and the connecting maps β, γ would make
the diagrams to not commute (to make sense of this see for example Figure 8).
Lemma 4.22 (Connecting maps). The diagrams in Eq. (4.20) and Eq. (4.21) are
commutatative diagrams of ∗-homomorphisms.
Proof. The map αI0 is a ∗-homomorphism because it is the composition I0 ,→
α−→ A1 of ∗-homomorphisms. The left square of the diagram in Eq. (4.20)
A0
commutes because α0 is just a corestriction of the ∗-homomorphism αI0. That is,
α0 maps I0 into I1. This follows since the map h1 : T1 → T0 is cellular. Indeed,
if g ∈ Cc(R0X T0
)
2 −X T1
1
) vanishes on the edges of T0, then α0(g) ∈ Cc(R1X T1
2 −X T0
1
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)35
Figure 8. A commutative diagram for the fibonacci tiling.
vanishes on the edges of T1. That is, α0(g)(x0, y0) = g(h1(x0), h1(y0)) = 0 for
x0, y0 ∈ X T1
We will now show that β is a ∗-homomorphism making the right square in
1 because h1(x0), h1(y0) are in the edges of T0.
Eq. (4.20) commute.
We start by showing that
n
n
qn /
/ A0
/ B0
in /
n := Cc(RnX Tn
/ 0
), A0
0
/ I0
n
1
2 −X Tn
1
n := Cc(Rn), B0
is a short exact sequence, where I0
n :=
Cc(RnX Tn
). To show that qn is surjective one uses Tietze's extension theorem
applied to functions with compact support, cf. [29, Theorem 3.4.8 p.124]. That qn
is a ∗-homomorphism follows from the fact that the 1-skeleton X Tn1
is Rn-invariant.
The map in extends functions with compact support to be zero outside RnX Tn
2 −X Tn
such that the extended function is continuous, which can be done as RnX Tn
1
2 −X Tn
is open in Rn. It follows that in is injective. That in is a ∗-homomorphism follows
1
by the fact that X Tn2 − X Tn1
is Rn-invariant. It is easy to see that Im in = ker qn.
If g, g0 ∈ Cc(R0) are equal on the 1-skeleton of T0, i.e. q0(g) = q0(g0) then
(q1 ◦ α)(g) = (q1 ◦ α)(g0) because h1 is cellular. More precisely, for x0, y0 ∈ X T1
1 we
get by cellularity of h1 that h1(x0), h1(y0) are in the 1-skeleton of T0. Then
(q1 ◦ α)(g)(x0, y0) = g(h1(x0), h1(y0)) = g0(h1(x0), h1(y0)) = (q1 ◦ α)(g0)(x0, y0)
1 descends to
1, making the diagram
0 is the map shown above. (In this proof, β is
0, while its extension to the C∗-algebra is denoted by β. We do this
if h1(x0) ∼R0 h1(y0) else they are both zero. Hence q1 ◦ α : A0
the quotient giving a unique ∗-homomorphism β : B0
commute. The evaluation of β on B0
defined on B0
for notational reasons). It is a general fact that β preserves multiplication:
β(q0(g)q0((g0)) = β(q0(gg0)) = q1(α(gg0)) = q1(α(g))q1(α(g0)) = β(q0(g))β(q0(g0)),
0 → B0
0 → B0
/
/
/
/
36
for g, g0 ∈ Cc(R0). It is clear that β is ∗-preserving.
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
To continue with our proof we need the following result.
1
1
) is continuous.
) → Cc(R1X T1
Lemma 4.23. Equip Cc(R0) with the inductive limit topology (cf. [10, Ex. 5.10,p.118]).
Then β : Cc(R0X T0
Proof. As explained in the proof of Lemma 4.11, it is enough to consider the re-
striction to a connected component C of R0 via the range map r : R0 → Rd because
a R0-compact set can only intersect a finite number of connected components, be-
cause the range map restricted to a connected component is a homeomorphism onto
its image, and because β(g) is 0 outside R0 for g ∈ Cc(R0X T0
). Thus it suffices
to show that the maps β and ι defined below are continuous in the inductive limit
topology. The map ι : Cc(R0X T1
) → Cc(R1X T1
) is given by
1
1
(cid:26) g(x0, y0) x0 ∼R0 y0
1
0
x0 6∼R0 y0
,
ι(g)(x0, y0) :=
where x0, y0 ∈ X T1
1 and x0 ∼R1 y0. Since R0 is open in R1, we get by the gluing
lemma for open subsets that one can always extend a function with compact support
to be zero outside R0 such that the extended function is continuous. Hence ι is
well-defined. Moreover, ι is continuous in the inductive limit topology because the
support of a function is unchanged when extending it by zeroes.
The map β : Cc(r(C) ∩ X T0
β(f)(x0) :=
1 ) → Cc(r(C) ∩ X T1
(cid:26) f(h1(x0)) h1(x0) ∈ r(C)
1 ) is given by
h1(x0) 6∈ r(C)
0
(and thus h1(x0) ∈ X T0
1 ).
Suppose fn ∈ Cc(r(C) ∩ X T0
where x0 ∈ X T1
1
limit topology. That is, there is a compact subset K ⊂ r(C) ∩ X T0
1
all n ∈ N
1 ) converges to f ∈ Cc(r(C) ∩ X T0
1 ) in the inductive
such that for
supp fn, supp f ⊂ K
and such that
as n → ∞.
Note that the supremum is taken over K. We claim that
1 ⊂ r(C) ∩ X T1
1
fn − fsup → 0
1 (K) ∩ X T1
K0 := h−1
is compact and that
Then
supp β(fn), supp β(f) ⊂ K0.
β(fn) − β(f)sup = sup
x0∈K0
β(fn)(x0) − β(f)(x0)
fn(h1(x0)) − f(h1(x0))
=
h1(x0)∈r(C)
sup
x0∈K0
fn(x) − f(x)
≤ sup
x∈K
= fn − fsup → 0 as n → ∞,
where the inequality holds because we can only get nonzero values if h1(x0) ∈ K.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)37
Next we check the claim. By an argument similar to Eq. (4.13), supp β(fn) =
1 (supp fn) ∩ X T1
h−1
1 and supp β(f) = h−1
1 . It follows that
1 (supp f) ∩ X T1
supp β(fn), supp β(f) ⊂ K0.
,→ r(C) is continuous.
Now, K is compact in r(C) since the inclusion r(C) ∩ X T0
1
By the same argument as in the proof of Lemma 4.11, h−1
1 (K) is compact in r(C).
is Rd-closed, K0 ⊂ r(C) ∩ X T1
is compact in Rd hence
Since the 1-skeleton X T1
1
1
also compact in r(C) ∩ X T1
(cid:3)
1 by Lemma C.2.
1
), from which it follows that the right square commutes.
We continue with the proof of Lemma 4.22. By the above lemma, Lemma 4.23,
β is continuous in the inductive limit topology. By Theorem D.5, β extends to a
∗-homomorphism β : B0 → B1. By continuity of β, we get β(lim bi) = lim β(bi),
bi ∈ Cc(R0X T0
the maps are ∗-homomorphisms.
A similar argument shows that the diagram in Eq. (4.21) commutes, and that
(cid:3)
The inclusion map ιn was shown to be a ∗-homomorphism in [14, Prop. IV.9,
p. 57] using the reduced norm. In the next remark, we include an alternative proof
using the full norm.
Remark 4.24. Recall that ιn : Cc(Rn) → Cc(Rn+1) is given by
(cid:26) g(x0, y0) x0 ∼Rn y0
0
ιn(g)(x0, y0) :=
x0 6∼Rn y0
,
where x0 ∼Rn+1 y0. The map ιn is well-defined and continuous in the inductive
limit topology by the same proof as for ι in the proof of Lemma 4.23. It is easy to
check that ιn is a ∗-homomorphism. By Theorem D.5, ιn extends to a continuous
∗-homomorphism ιn : C∗(Rn) → C∗(Rn+1).
Lemma 4.25. Each row in the following diagram is a cochain complex of abelian
groups (not a short exact sequence).
(4.22)
0
0
/ K0(C0)
K0(γ)
/ K0(C1)
δ0
K1(J0)
K1(β0)
δ0
/ K1(J1)
δ1
δ1
/ K0(I0)
K0(α0)
/ K0(I1)
/ 0
/ 0.
Moreover the diagram commutes, (i.e. the diagram is a cochain map).
Proof. From the six-term exact sequence in Eq. (4.3) we get the diagram with exact
rows
K0(C0)
δ0
K1(J0)
/ K1(B0)
K0(γ)
K1(β0)
K1(β)
K0(C1)
δ0
/ K1(J1)
/ K1(B1)
/ 0
/ 0.
The diagram commutes because of the naturality of the exponential map, because
the diagram in Eq. (4.21) commutes, and because Kj(φ ◦ ψ) = Kj(φ) ◦ Kj(ψ),
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
38
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
j = 1, 2. Similarly, from the six-term exact sequence in Eq. (4.5) we get the
commutative diagram
K1(B0)
K1(β)
K1(B1)
δ1
δ1
K0(I0)
K0(α0)
/ K0(I1).
Putting these two diagrams together and using Eq. (4.7) we get
δ1
(4.23)
K0(C0)
δ0
K1(J0)
K0(γ)
K1(β0)
K0(C1)
δ0
/ K1(J1)
K1(B0)
K1(β)
K1(B1)
δ1
δ1
/ K0(I0)
K0(α0)
/ K0(I1).
δ1
By Eq. (4.10), δ1 ◦ δ0 = 0. Thus each row in Eq. (4.22) is a cochain complex of
(cid:3)
abelian groups.
4.4. Computation of K0(γ), K0(α0), K1(β0). In this subsection, we compute
the connecting maps of the diagram in Lemma 4.25. That is, we compute the
connecting maps K0(γ), K0(α0), K1(β0) in terms of the generators (i.e. the natural
basis elements) of the K-groups. This can be done since they are expressed in terms
of the compact operators. The following lemma will be given up to the isomorphism
given in Proposition 4.2, which we denote by λ.
Lemma 4.26. The generators of K0(C0) are
[euu]0,
where u is a vertex in T. Take a vertex u ∈ T from each R0-equivalence class
(i.e. each stable class) to get all generators.
Proof. Let
(cid:26) δu
0
v = w
else
euv(δw) :=
be a matrix unit (e.g. it sends δv to δu), where u ∼R0 v ∼R0 w are vertices in T,
and δw ∈ '2([u]R0) is the standard basis element of the Hilbert space '2([u]R0), i.e.
for w0 ∈ [u]R0,
(cid:26) 1 w = w0
0
else.
δw(w0) :=
Recall that a 1-dimensional projection generates the K0-group of the compact
operators K and that K0(K) is isomorphic to Z (cf. [36, Cor. 6.4.2, p.103]). One
can choose any 1-dimensional projection because they are all equivalent (cf. Lemma
G.2). Since euu is a 1-dimensional projection, and since C0 can be written in terms
of the compacts by Proposition 4.2(1), the lemma follows by taking a vertex u ∈ T
(cid:3)
from each R0-equivalence class.
/
/
/
/
/
/
/
/
/
%
%
/
/
/
/
/
/
9
9
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)39
Lemma 4.27. With notation as in the previous lemma, K0(γ) : K0(C0) → K0(C1)
evaluated on the generators of K0(C0) is given by
K0(γ)([euu]0) = [γ(euu)]0 = X
[eu0u0]0,
where the sum is over all vertices u0 in T1 = 1
Proof. Take vertex u ∈ T. Then by definition of K0 on ∗-homomorphisms we get
(cf. [36, p. 61])
h1(u0)=u
λ ω(T)λ.
where λ is the isomorphism in Proposition 4.2(1). We have λ−1(euu) = g, where
K0(λγλ−1)([(euu)]0) = [λγλ−1(euu)]0,
g(x, y) :=
(cid:26) 1 x = y = u
(cid:26) 1 x = y and h1(x) = u
else.
0
Then since
γ(g)(x, y) =
0
where x, y are vertices in T1 such that x ∼R1 y, we get
eu0u0,
λγλ−1(euu) = X
else
h1(u0)=u
where the sum is over all vertices u0 ∈ T1.
(cid:3)
The following lemma will be given up to the isomorphism given in Proposition
4.2, which we denote by λ.
Lemma 4.28. The generators of K0(I0) are
where u ∈ T is a face, b ∈ M2
is its scalar part, and
[b ⊗ euu + s(b) ⊗ (I − euu)]0 − [s(b) ⊗ I]0,
(cid:0) (cid:94)
C0(◦
u)(cid:1) is the (reparametrized) Bott projection, s(b)
I =X
evv,
v
where the sum is over all faces v ∈ T.
Choose face u ∈ T0 from each R0-equivalence class (i.e. from each stable face) to
get all generators.
Proof. Choose protoface fn ∈ T, (n fixed), and let Kn := K('2([fn])) be the C∗-
algebra of compact operators on the Hilbert space Hn := '2([fn]). Since [fn]R0 is
an infinite set, Kn is non-unital, i.e. In := IHn 6∈ Kn. Let
be the open unit disk. Since D is homeomorphic to
C0(D, Kn). Recall that C0(D, Kn) is isomorphic to C0(D) ⊗ Kn via the map
◦
fn, we have C0(
◦
fn, Kn) ∼=
(cf [46, p.173]). Let euv ∈ Kn be the standard matrix unit
D := {z ∈ C z < 1}
f ⊗ a 7→ (t 7→ f(t)a).
(cid:26) δu w = v
0
else
euv(δw) =
D. GONC¸ ALVES
40
where u, v ∈ [fn], i.e. u, v ∈ T are faces R0-equivalent to face fn. Consider the
projection euu ∈ Kn, and let φu : C0(D) → C0(D) ⊗ Kn be the ∗-homomorphism
M. RAMIREZ-SOLANO
φu(a) := a ⊗ euu.
zp1 − z2
sFM
C0(D),
Then K0(C0(D) ⊗ Kn) is canonically isomorphic to K0(C0(D)) via the homomor-
phism K0(φu), (but the inverse is not explicit). Moreover, K0(C0(D)) is isomorphic
to Z via the map 1 7→ [b]0 − [s(b)]0 , where b ∈ M2((cid:94)C0(D)) is the Bott projection,
and s(b) is its scalar part (cf. [46, p.156,173], [36, Example 11.3.2 p.200], [36, Exer-
cise 9.5, p.171]). We remark that the Bott projection and its scalar part correspond
(cid:21)
to the elements in M2(C(¯D)) given by
b = (z 7→
1 − z2
Define the non-unital C∗-algebras
¯zp1 − z2
(cid:20) 1C(¯D)
(cid:20) 1
= (z 7→
s(b) =
z2
(cid:21)
(cid:21)
(cid:20)
0
0
0
0
),
0
0
).
A :=
A0 :=
C0(D) ⊗ Kn.
i=1
As usual, A, fA0 will denote their unitization, respectively. Let p ∈ M2(fA0) be the
p := (cid:0)(b − s(b)) ⊗ euu, 0, . . . , 0(cid:1) +(cid:0)s(b) ⊗ I1, . . . , s(b) ⊗ IsF
projection (assuming n=1)
(cid:1)
i=1
= b ⊗ euu + s(b) ⊗ (I − euu),
where
I :=
sFX
i=1
Ii =
sFX
X
i=1
v∼fi
evv =X
evv,
v
sFM
and the last sum is over all faces v of T. (We use the rule of tensors a⊗ x + a⊗ y =
a⊗(x+ y)). Thus the scalar part of p is s(p) = s(b)⊗ I. Note that the unit element
offA0 is
where 1(cid:94)C0(D) is the constant function z 7→ 1, z ∈ ¯D. Let bA ∈ M2( A) be the
projection
1eA0 = 1(cid:94)C0(D) ⊗ I,
bA := (cid:0)(b − s(b)), 0, . . .(cid:1) +(cid:0)s(b), s(b), . . . , s(b))
= (b, s(b), . . . , s(b)).
We claim that
(4.24)
where φ2 is the usual extension of φu to M2( A) → M2(fA0). Then by [36, Prop. 4.2.2, p. 63]
φ2(bA) = p,
we have the following identifications
1 7→ [b]0 − [s(b)]0 7→ [bA]0 − [s(bA)]0 7→ [p]0 − [s(p)]0,
The proof of the claim, which is inspired by the splitting lemma, follows from
from which the lemma follows.
the facts that
(4.25)
φ2
(cid:0)bA − s(bA)(cid:1) = p − s(p)
and
ψ2(π2(bA)) = π0
2(p),
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)41
where φ2 is the extension of φu to M2(A) → M2(A0), and ψ2, π2, π0
2 are defined
below. Here we use the fact that we have a commutative diagram of short exact
sequences that split as C-vector spaces:
0
0
r
/ A i
φ
/ A0
r0
i0
A
φ
/fA0
σ
π
σ0
π0
C
ψ
/ C
0
/ 0,
where φ := φu, and the retractions are r(a) = a − s(a) and r0(a0) = a0 − s(a0), and
the sections σ(π(a)) = s(a) and σ0(π0(a0)) = s(a0), and ψ = idC, and we have the
identities
(4.26)
r ◦ i = idA,
r0 ◦ i0 = idA0,
π ◦ σ = idC,
π0 ◦ σ0 = idC,
i ◦ r + σ ◦ π = id A
i0 ◦ r0 + σ0 ◦ π0 = ideA0.
By functoriality and additivity of M2 we get the commutative diagram
0
/ M2(A)
φ2
r2
r0
2
0
/ M2(A0)
M2( A)
φ2
/ M2(fA0)
σ2
π2
σ0
2
π0
2
M2(C)
ψ2
0
/ M2(C)
/ 0,
where the operations are done coordinatewise. In particular, the equations corre-
assuming φ2 ◦ r2(bA) = r0
sponding to Eq. (4.26) hold here as well. Since bA ∈ M2( A) and p ∈ M2(fA0) and
2(p) then φ2(bA) = p because
2(p) = ideA0(p) = p.
2(p) and ψ2(π2(bA)) = π0
φ2(bA) = φ2(i2 ◦ r2(bA) + σ2 ◦ π2(bA))
2 ◦ π0
2 ◦ ψ2 ◦ π2(bA) = i0
The converse, although not needed in this proof, follows by commutativity of the
diagram. Thus Eq. (4.24) is equivalent to Eq. (4.25).
It remains to show that
Eq. (4.25) holds. The first equality follows immediately:
2 ◦ φ2 ◦ r2(bA) + σ0
2(p) + σ0
2 ◦ r0
= i0
φ2(bA − s(bA)) = φ2
(cid:0)b − s(b), 0, . . . , 0(cid:1)
= (cid:0)(b − s(b)) ⊗ euu, 0, . . . , 0(cid:1)
= p − s(p).
s(bA) = (s(b), . . . , s(b))
Let 1 := 1(cid:94)C0(D). Then s(bA) ∈ M2( A) is
(cid:20) 1
(cid:21)
(cid:20) (1, . . . , 1)
(cid:20) 1 A 0
(cid:21)
= (
0
0
=
0
0
0
0
, . . . ,
=
0
0
,
(cid:21)
)
0
0
(cid:20) 1
(cid:21)
0
/
/
/
/
/
/
/
/
/
z
z
/
{
{
/
/
/
/
/
/
v
v
/
/
u
u
/
/
/
/
v
v
/
t
t
/
42
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
where we use the identifications
i=1
(cid:0) sFM
(cid:94)C0(D)(cid:1) ⊃ M2( A).
(cid:21)
(cid:20) 1
(cid:20) 1 ⊗ IsF
(cid:21)
(cid:21)
∈ M2(C).
(cid:21)
0
0
0
0
0
0
)
(cid:0)(cid:94)C0(D)(cid:1) ∼= M2
M2
sFM
i=1
π2(bA) = π2(s(bA)) =
s(p) = (
0
0
0
(cid:20) 1 ⊗ I1
(cid:20) (1 ⊗ I1, . . . , 1 ⊗ IsF)
(cid:20) 1eA0
, . . . ,
(cid:21)
0
,
0
0
0
0
0
=
=
π2(p) = π2(s(p)) =
∈ M2(C).
(cid:20) 1
0
(cid:21)
0
0
Hence
Similarly,
yields
(cid:3)
Hence the second equality in Eq. (4.25) holds as well.
Lemma 4.29. With notation as in the previous lemma, K0(α0) : K0(I0) → K0(I1)
evaluated on the generators of K0(I0) is
K0(α0)(cid:16)[b ⊗ euu + s(b) ⊗ (I − euu)]0 − [s(b) ⊗ I]0
b ⊗ euu + s(b) ⊗ (I − euu)(cid:17)]0 − [α0(cid:16)
(cid:17)
s(b) ⊗ (I − euu)(cid:17)]0
= [α0(cid:16)
= [b ⊗ eu0u0 + s(b) ⊗ (I − eu0u0)]0 − [s(b) ⊗ I]0.
In the above expression u0 ∈ T1 is the unique face of T1 = 1
to u.
Proof. Let g : R0 → M2(C) be given by g = λ
λ is the ∗-isomorphism induced by λ. Note that here b denotes the Bott projection
reparametrized to face u and note that g is (up to isomorphism) a projection in M2
of the unitization of I0 = C∗(R0X2−X1). Evaluating g we get
−1(cid:0)b ⊗ euu + s(b) ⊗ (I − euu)(cid:1) where
λ ω(T)λ, which homotopes
g(x, y) =
.
b(x)
0
x = y ∈ u
s(b)(x) x = y 6∈ u
b(h1(x)) x = y ∈ u0
x = y 6∈ u0
x 6= y
x 6= y
c
0
,
Note that g is zero outside the diagonal of R0, and it is constant on the diagonal
intersected with X1 × X1. Next, we evaluate α0(g) and get
α0(g)(x, y) =
(cid:20) 1 0
(cid:21)
∈ M2(C) is the constant value of the constant function
where c :=
s(b). Then, denoting the constant function with value c by ζ(x) := c (i.e. ζ is
a reparametrization of s(b)), we have that
0 0
λ(α0(g)) = b ◦ h1 ◦ γu0
i,u0 ⊗ eu0u0 + ζ ⊗ (I − eu0u0)
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)43
is a projection in the unitization ofLsF
i are the rep-
resentatives of the R1-equivalence classes, i.e., the stable faces of T1) and its scalar
part is
i=1 C0( ◦
0) ⊗ K('2([u0
ui
i]R1)), (u0
s(λ(α0(g))) = ζ ⊗ I.
ζu0
i
i=1
i
v0∼u0
v06=u0
i, and ζu0
The map γu0
x0 being the translation of the two faces u0 = u0
function ζ has different faces as domain depending on the context. That is
i → u0 is the translation map given by γu0
sFX
i,u0(x) := x + x0, with
i + x0. Note that the constant
i,u0 : u0
sFX
ev0v0),
ζ ⊗ (I − eu0u0) =
⊗ (X
ev0v0),
ζ ⊗ I =
ζu0
i
i=1
⊗ (X
v0∼u0
i
is the
i. By definition of K0 on ∗-homomorphisms and
where the inner sum is over all faces v0 ∈ T1 equivalent to u0
constant function c with domain u0
Proposition 4.2.2 p.63 in [36], we get
[λ(α0(g))]0 − [s(λ(α0(g)))]0 = [b ◦ h1 ◦ γu0
i,u0 ⊗ eu0u0 + ζ ⊗ (I − eu0u0)]0 − [ζ ⊗ I)]0.
Moreover, the scalar part map s and λ commute. Since all faces of T1 and T0 are
oriented counter clockwise, the map h0u = idu is orientation preserving and since h
is a homotopy, also h1 : u0 → u is orientation preserving. Hence (up to orientation
preserving reparametrization) we can replace b ◦ h1 ◦ γu0
i,u0 with b and ζ with s(b)
and we get
i
[λ(α0(g))]0 − [s(λ(α0(g)))]0 = [b ⊗ eu0u0 + s(b) ⊗ (I − eu0u0)]0 − [s(b) ⊗ I]0.
(cid:3)
The following lemma will be given up to the isomorphism given in Proposition
4.2, which we denote by λ. Since the proofs of the next two lemmas are similar to
the previous two, we omit some technical details.
Lemma 4.30. The generators of K1(J0) are
[b ⊗ euu +X
v6=u
1 ⊗ evv]1
where u is an edge in T, the sum is over all vertices of v in T, and b is the map
s 7→ ei2πs, s ∈ [0, 1] (up to reparametrization from domain u to domain [0, 1]).
Choose edge u ∈ T from each R0-equivalence class to get all generators.
Proof. Let T := {z ∈ C z = 1} be the unit circle, and u ∈ T be an edge.
Recall that Z is isomorphic to K1(C(T)) mapping 1 7→ [b0]1, where b0 is the identity
map z 7→ z, z ∈ T. Moreover, K1(C(T)) is isomorphic to K1(C0(◦
u)) mapping
[b0]1 7→ [b]1. Note that b is in the unitization of C0(◦
u). Recall that C0((0, 1), K) is
isomorphic to C0((0, 1)) ⊗ K via the map
f ⊗ a 7→ (t 7→ f(t)a).
Recall that K1(C0(◦
u)⊗K), where K is the C∗-algebra of
compact operators on the Hilbert space '2([u]R0). Moreover, C0(◦
u) ⊗ K embeds in
J0 via the isomorphism given in Proposition 4.2(2), where for simplicity, we assume
u)) is isomorphic to K1(C0(◦
D. GONC¸ ALVES
44
that u is one of the representatives ei in the proposition. Let euv ∈ K('2([u])) be
the standard matrix unit
M. RAMIREZ-SOLANO
(cid:26) δu w = v
0
else
,
euv(δw) =
where u, v, w ∈ T are edges R0-equivalent to edge u. Thus we have the following
identifications
1 7→ [b0]1 7→ [b]1 7→ [b ⊗ euu + 1 ⊗ (I − euu)]1,
i=1
evv,
I :=
!
sEX
X
IB(Hi) =X
where the third map follows by the same argument as for the faces, the 1 in the
expression 1 ⊗ (I − euu) denotes the constant function 1, and
sEX
where the sum is over all edges v ∈ T, and 1ei ∈ (cid:94)
C0( ◦
ofLsE
ei) is the unit element -- the
constant function 1. Note that b⊗ euu +1⊗(I − euu) is a unitary in the unitization
(cid:3)
Lemma 4.31. With notation as in the previous lemma, K1(β0) : K1(J0) → K1(J1)
evaluated on the generators of K1(J0) is
ei) ⊗ K('2([ei])).
i=1 C0( ◦
1 ⊗ I :=
1ei ⊗
v∼ei
evv
i=1
v
,
K1(β0)(cid:16)[b ⊗ euu +X
1 ⊗ ev0v0]1 − X
v6=u
[b ⊗ eu0u0 + X
(cid:17) =
[b ⊗ eu0u0 + X
1 ⊗ evv]1
1 ⊗ ev0v0]1.
X
h1(u0)=u
v06=u0
h1(u0)=u
v06=u0
where the sums are over all edges u0, v0 in T1. Here u is the edge u with reversed
orientation, and for this lemma equality h1(u0) = u is understood to include that
h1 : u0 → u is orientation preserving.
We remark that by Whitehead's lemma
1 ⊗ ew0w0]1 = [b ⊗ ev0v0 + X
[b ⊗ eu0u0 + X
K1(λβ0λ−1)(cid:16)[b ⊗ euu + 1 ⊗ (I − euu)]1
w06=u0
w06=v0
1 ⊗ ew0w0]1
for u0 ∼R1 v0.
(cid:17) = [λ β0λ−1(b ⊗ euu + 1 ⊗ (I − euu))]1,
Proof. Take edge u ∈ T. Then by definition of K1 on ∗-homomorphisms we get
where 1 is the constant function 1.
Let g be the unitary in the unitization of J0 given by
where λ is the isomorphism in Proposition 4.2 extended to the unitization. We
evaluate g and get
g := λ−1(cid:0)b ⊗ euu + 1 ⊗ (I − euu)(cid:1),
b(x) x = y ∈ u
x = y 6∈ u
g(x, y) :=
else.
1
0
β0(g)(x, y) =
b(h1(x))
b(h1(x))
1
0
x = y ∈ u0, h1(u0) = u
x = y ∈ u0, h1(u0) = u
x = y ∈ u0, u 6= h1(u0) 6= u
x 6= y
(4.27)
Since
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)45
we get
λ β0(g) = X
h1(u0)=u
b ⊗ eu0u0 + X
h1(u0)=u
b∗ ⊗ eu0u0 + X
h1(v0)6=u
h1(v0)6=u
1 ⊗ ev0v0,
where the sums is over all edges u0, v0 ∈ T1 = 1
s ∈ [0, 1], up to reparametrization). Thus
λ ω(T)λ (and b∗ is the function e−2πis,
[λ β0λ−1(b ⊗ euu +X
b ⊗ eu0u0 + X
1 ⊗ evv)]1 =
v6=u
b∗ ⊗ eu0u0 + X
h1(u0)=u
[ X
b ⊗ eu0u0 + X
X
[b ⊗ eu0u0 + X
h1(v0)6=u
[ X
h1(u0)=u
h1(u0)=u
v06=u0
1 ⊗ ev0v0]1 =
h1(u0)=u
1 ⊗ ev0v0]1 − [ X
1 ⊗ ev0v0]1 − X
h1(u0)=u
h1(v0)6=u
h1(v0)6=u
b ⊗ eu0u0 + X
[b ⊗ eu0u0 + X
h1(v0)6=u
h1(u0)=u
v06=u0
1 ⊗ ev0v0]1 =
1 ⊗ ev0v0]1,
where the equalities hold by [36, Prop. 8.1.4(iv), p.135].
Note that by Whitehead's lemma
[b ⊗ eu0u0 + c2]1 = [b ⊗ ev0v0 + c3]1,
whenever u0 ∼R1 v0. This equality is equivalent to [b ⊗ eu0u0 + b∗ ⊗ ev0v0 + c1]1 = 0
by [36, Prop. 8.1.4(iv), p.135], which can easily be verified:
[b ⊗ eu0u0 + b∗ ⊗ ev0v0 + c1]1 = [bb∗ ⊗ eu0u0 + c2]1 = [(1 ⊗ eu0u0 + c2)]1 = [1]1 = 0,
where the first equality is by Whitehead's lemma, and
1 ⊗ ew0w0
1 ⊗ ew0w0
c2 := X
w06=u0
c3 := X
w06=v0
c1 := X
w06=u0
w06=v0
1 ⊗ ew0w0.
(cid:3)
4.5. Stable cohomology and K-theory. In this subsection we introduce the
stable cohomology groups, from which we compute the K-theory of the C∗-algebra
S0. We start with the following theorem, which is one of our main results. It shows
that
δ0
δ1
/ ZsF
/ ZsV
/ ZsE
C•
S := 0
(4.28)
is a cochain complex, and the collection of connecting maps
(4.29)
form a cochain map, which we denote by W • : C•
S, and whose explicit
computations are given in Lemma 4.27, Lemma 4.31, Lemma 4.29. Moreover, it
relates the group homomorphisms K0(ι), K1(ι) with the connecting maps.
WF := K0(α0)
S → C•
WE := K1(β0),
WV := K0(γ),
/ 0
/
/
/
/
46
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Theorem 4.32. With the above notation, the rows of the following commutative
diagram are cochain complexes of abelian groups
(4.30)
0
0
/ ZsV
WV
/ ZsV
δ0
δ0
ZsE
WE
/ ZsE
δ1
δ1
ZsF
WF
/ ZsF
0
/ 0.
Moreover, the group homomorphism K1(ι) : K1(A0) → K1(A1) is given by
K1(ι) = H1(WE),
The group homomorphism K0(ι) : K0(A0) → K0(A1) fits into the commutative
Im δ0 is given by H1(WE)([z]) := [WE(z)], z ∈ ker δ1.
Im δ0 → ker δ1
(4.31)
where H1(WE) : ker δ1
diagram
(4.32)
0
0
/ H2(C•
S)
H2(WF )
/ H2(C•
S)
K0(A0)
K0(ι)
/ K0(A1)
/ H0(C•
S)
H0(WV )
/ H0(C•
S)
/ 0
/ 0,
S) = ker δ0, H0(WV ) = WV ker δ0. For tilings of dimension 1, H2(C•
where H0(C•
0. For tilings of dimension 2, H2(C•
Proof. The first diagram is simply restating the one in Lemma 4.25 but with explicit
formulas for the connecting maps, which were computed in the lemmas 4.27, 4.31,
4.29. We now show the second statement of the theorem, namely that K1(ι) =
H1(WE). From the six-term exact sequence in Eq. (4.7) we get the following
diagram with exact rows
S) = Z and H2(WF ) = id.
S) =
0
0
/ K1(A0)
K1(α)
/ K1(A1)
/ K1(B0)
K1(β)
/ K1(B1)
δ1
δ1
/ K0(I0)
K0(α0)
/ K0(I1)
as K1(In) = 0. The diagram commutes because of naturality of δ1, because the
diagram in Eq. (4.20) commutes, and because Kj(φ◦ ψ) = Kj(φ)◦ Kj(ψ), j = 1, 2.
(cf. [36, Prop. 9.1.5, p. 157]). Hence by this diagram
K1(ι) = K1(α) = K1(β)ker δ1 ,
where the first equality follows from the invariance under homotopy of K-theory,
and the second equality is up to the isomorphism K1(An) ∼= ker δ1. By the diagram
in Eq. (4.23) we get
K1(β)ker δ1 = [K1(β0)ker δ1]Im δ0
= [WEker δ1]Im δ0
= H1(WE),
where the first equality is up to the isomorphism ker δ1 ∼= ker δ1
denotes the map [x]Im δ0 7→ [WE(x)]Im δ0 for x ∈ ker δ1.
Im δ0 , and [WEker δ1]Im δ0
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)47
We now show the last statement of the theorem. By Eq. (4.7) and Eq. (4.8)
and naturality of the index map (and hence naturality of δ1) we get the following
commutative diagram with exact rows
0
0
and note that
/ K0(I0)
Im δ1
K0(A0)
/ K0(B0)
[K0(α0)]Im δ1
K0(ι)
/ K0(A1)
K0(β)
/ K0(B1)
/ K0(I1)
Im δ1
/ 0
/ 0,
[K0(α0)]Im δ1 = [WF ]Im δ1 = H2(WF ).
By the six-term exact sequence in Eq. (4.3) and naturality of the exponential map
we get the following commutative diagram with exact rows
0
0
/ K0(B0)
K0(β)
/ K0(B1)
/ K0(C0)
K0(γ)
/ K0(C1)
δ0
δ0
/ K1(J0)
K1(β0)
/ K1(I1)
Thus
where the first equality is up to the isomorphism K0(Bn) ∼= ker δ0.
For tilings of dimension 2, it holds, by Lemma 4.5, that
K0(β) = K0(γ)ker δ0 = WV ker δ0 = H0(WV ),
∼=
K0(In)
Im δ1
[WF ]Im δ1 = 1.
ZsF
Im δ1 =
ZsF
Im δ1
∼= Z
Note that in the last isomorphism, any face f yields a generator of Z. Now
WF ([f]) = [f0] in the notation of the definition of homotopy (cf. Definition 4.6),
corresponds to the identity map Z → Z, which we denote by 1.
(cid:3)
Definition 4.33 (Stable(S)). For k ∈ {0, 1, 2}, we define the stable cohomology
groups for the fixed tiling T as
H k
S
where the chain complex C•
4.32. Here the notation is
:= lim→ (H k(C•
S), H k(W •)),
S and chain map W • were defined right above Theorem
lim→ (X, A) := lim→ X A /
/ X A /
/ X A /
/ .
The K-theory groups K0(S0), K1(S0) for the C∗-algebra S0 can be expressed in
S as is shown in the following
terms of the stable cohomology groups H0
theorem.
Theorem 4.34. For tilings of dimension 1 or 2, the group K1(S0) is given by
S, H2
S, H1
K1(S0) = H1
S,
and the group K0(S0) fits into the short exact sequence
/ H0
/ K0(S0)
/ H2
0
S
S
/ 0.
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
48
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Moreover, in dimension 1 it holds, H2
S = Z.
H2
Proof. By Proposition B.4 we have that
S = 0, H1
S = Z, and in dimension 2 it holds
S0 = lim→ A0
ι0
/ A1
ι1
/ . . .
.
By [36, Theorem 6.3.2]
Kj(S0) = lim→ Kj(A0) Kj(ι0)/
/ Kj(A1)Kj(ι1) /
/ . . .
j = 0, 1.
Then, by Theorem 4.32, we get the statements of the theorem, where for the second
statement, we use the fact that the direct limit of a directed system of short exact
sequences of abelian groups is a short exact sequence because the direct limit is an
(cid:3)
exact functor in the category of abelian groups.
Unlike in the unstable case (cf. Theorem 1.4), we do not know whether the
S is not necessarily projective,
short exact sequence in Theorem 4.34 splits, since H0
i.e. isomorphic to Z' for some ' ∈ N0.
4.6. Properties.
Proposition 4.35. The short exact sequence in Eq. (4.1) does not split.
Proof. Suppose for contradiction that the short exact sequence splits, that is, there
exists a section σ : Cn → Bn. Since σ is in particular a ∗-homomorphism, σ(p) is a
projection for any projection p ∈ M'( Cn), ' ∈ N. Then by the continuous functional
calculus, exp(2πiσ(p)) = I. By [36, Prop. 12.2.2(i)], δ0([p]0 − [s(p)]0) = [I]1 = 0,
i.e. the exponential map δ0 is the zero map. This, together with Eq. (4.4), implies
(cid:3)
that there is only 1 prototile, a contradiction since the tiling T is aperiodic.
Proposition 4.36. For tilings of dimension 2, the short exact sequence in Eq. (4.2)
does not split.
Proof. Suppose for contradiction that the short exact sequence splits, that is, there
exists a section σ : Bn → An. Since σ is in particular a ∗-homomorphism, v := σ(u)
is a unitary for any unitary u ∈ M'( Bn), ' ∈ N. Thus p := 1 − v∗v = 0 and
q := 1− vv∗ = 0. By [36, Prop. 9.2.2], δ1([u]1) = [p]0 −[q]0 = 0, i.e. the exponential
map δ1 is the zero map, and hence also δ1 is the zero map. This together with
Eq. (4.12) implies that there is only 1 prototile, a contradiction since the tiling T
(cid:3)
is aperiodic.
Proposition 4.37. The C∗-algebra S0 is not unital. (But this does not rule out
that S might be unital).
Proof. It is well-known that the C∗-algebra of compact operators K is never unital
on an infinite dimensional Hilbert space H, since the identity I ∈ B(H) is not a
compact operator. The quotient of a unital C∗-algebra is unital because the class
[I] is the identity in the quotient. Since, by Eq. (4.1), the compacts K is a quotient
of Bn = C∗
r (Rn) is non-
unital. Hence S0 must be non-unital, since all the algebras forming the limit are
non unital. Cf. [36, Exc.6.7(iii)] which uses that the inclusion map An → An+1 is
injective. This is because an element of a unital C∗-algebra with distance less than
1 to the identity is invertible, so if the limit algebra is unital then the individual
r (RnX1), Bn is not unital. Similarly, by Eq. (4.2), An = C∗
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)49
algebras must be unital from a certain step. Tensoring with the compact operators
will turn a unital C∗-algebra into an non-unital C∗-algebra Morita equivalent to the
first. So S0 non unital does not rule out the possibility that S might be unital. (cid:3)
Proposition 4.38. Let n ∈ N. Then, the ideal Jn of Bn (d=1) in Definition 4.1
does not contain any projections other than the zero projection. Hence Bn is not
AF. (The same holds for d=2). This does not prove though that S0 is not AF.
However, we can decide this from K1(S0).
Proof. If p ∈ C0((0, 1), K('2([edge]))) is a projection, then p : (0, 1) → K('2([edge]))
is continuous vanishing at the endpoints. Since the map p(t) 7→ p(t) is continu-
ous and since p(t) is a projection, its norm is 0 or 1. Since (0, 1) is connected, and
since the image of connected sets under continuous maps is connected, the map
p(t) 7→ p(t) is either 0 or 1. Since it vanishes at the endpoints, it has to be zero.
In particular Jn is not AF. Any ideal in an AF algebra is AF. Hence Bn is not AF.
(Recall that an AF-algebra is the norm closed linear span of its projections. This is
because it is an inductive limit of finite dimensional C∗-algebras (which are direct
sums of matrix algebras), and every finite dimensional C∗-algebra is the linear span
(cid:3)
of its projections.)
5. Stable-Unstable relationship
In this section we relate the K-theory of U to the K-theory of S. On the one hand,
the K-theory of S, as shown in Section 4, is given in terms of the stable cohomology
via a skeletal decomposition. On the other hand, the K-theory of U is well-known
to be given in terms of the Cech cohomology also via a skeletal decomposition. We
will relate the so-called stable-transpose homology to Cech cohomology via PE-
cohomology and PE-homology. In particular, the stable-transpose homology is a
simpler method for computing the K-theory of U.
Recall that Ω denotes the continuous hull, ω : Ω → Ω denotes the substitution
map, which is a homeomorphism, and λ > 1 denotes the inflation factor. Define
the tiling space
with inflation factor λ and substitution map ωn : Ωn → Ωn given by
Ωn := λ−nΩλn,
n ∈ Z,
ωn := Ad(λ−n)ωAd(λn)
where Ad(λ) : Ω → λΩλ−1 is defined by
Ad(λ)(T) := λT λ−1.
n ∈ Z,
Fix a substitutional tiling T ∈ Ω. We can then construct the sequence of tilings
Tn := λ−nωn(T)λn ∈ Ωn,
n ∈ Z,
whose tiles shrink as n increases. Furthermore we can construct the following
commutative diagram
ω−2
ω−1
ω0
8
T1
ω1
T2;
···
···
/ T−2
T−1
θ1
θ2
T0
_
id
T0
θ−1
θ−2
/
/
/
/
/
/
/
#
#
/
/
/
/
{
{
o
o
/
/
50
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
where T0 := T, θ0 := id, and
(5.1)
ωn := Ad( 1
)ωn,
θn := ω−nAd(λn),
λ
n ∈ Z.
n ∈ Z.
If p is a prototile of T0 then λ−np is a prototile of tiling Tn, n ∈ Z. Moreover, the
prototile λnp coincides with ωn(p) as sets.
Example 5.1. Consider the Fibonacci tiling with protoedges a, b of lengths a = 1,
b = 1/φ, (λ = φ=golden ratio) and substitution rule ω(a) = ab, ω(b) = a. The
tilings T1, T0, T−1, T−2 are shown in the following picture.
T1 := 1
λ ω(T)λ
a b a a
b
a
λ
b
λ
a
λ
a
λ
b
λ
a
a0
b
a
b0
a
b
a
λa
λb
T0 := T
T−1 := λω−1(T) 1
λ
T−2 := λ2ω−2(T) 1
λ2
a00
λ2a
Note that the edge a0 = λa ∈ T−1 coincides with the patch ω(a) = ab ⊂ T0 as sets,
i.e. a0 is "tiled" with ω(a).
We then define the equivalence relation Rn (for Tn) on Rd as
Rn := R(Tn) = 1
λn R(ωn(T)),
n ∈ Z,
where
R(Tn) := {(x, y) ∈ R2d Tn(x) − x = Tn(y) − y},
and recall that Tn(x) is the patch made of all the tiles in tiling Tn that contain x.
We equip these equivalence relations with the subspace topology of R2d. Then
··· ⊂ R−2 ⊂ R−1 ⊂ R0 ⊂ R1 ⊂ R2 ⊂ ···
and Rn is open in Rn+1, n ∈ Z (cf. Subsection 3.1). We have the directed system
of chain complexes indexed by Z
···
S(R−1) W •
/ C•
−1 /
S(R0) W •
/ C•
0
S(R1) W •
/ C•
1
S(R2) W •
/ C•
2 /
··· ,
S(R0) and W •
where C•
0 are defined in the diagram of Lemma 4.25 in terms of the
compacts and of the stable cells of T0 and simplified in Eq. (4.28). The remaining
S(Rn) and W •
C•
n are defined similarly in terms of the compacts and of the stable
cells of Tn. They are independent of n because Tn has the same stable cells as T0
up to shrinking. Applying the contravariant Hom( -- , Z) functor we get
0
t /
/ C ST•
(R1) W •
/ C ST•
(Rn) := C•
··· ,
···
S(Rn)t. Note that we used the transpose instead of Hom( -- , Z)
where C ST•
since the groups we take the transpose of are finitely generated free abelian groups
with a fixed basis (the stable cells of Tn). Recall that the category of abelian groups
(R−1) W •
(R0) W •
(R−2) W •
−3
/ C ST•
/ C ST•
−2
t /
−1
t/
t /
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)51
is cocomplete, i.e. all small colimits (in particular all direct limits) exist. Hence,
we can define the abelian groups
H k
S := lim→ (··· H k(W •
−1) /
:= lim→ (··· Hk(W •
t)/
0
/ H k(C•
S(R0)) H k(W •
0 ) /
/ H k(C•
S(R1)) H k(W •
1 )/
··· ),
−1
H ST
k
(R0)) Hk(W •
/ Hk(C ST•
··· ).
Since all chain complexes are independent of n, it makes sense to define them more
succinctly as follows
Definition 5.2 (Stable-Transpose(ST), Stable(S)). For k ∈ {0, 1, 2}, define the
homology and cohomology groups
t)/
/ Hk(C ST•
(R−1))H k(W •
−2
t)/
H ST
k
H k
S
:= lim→ (Hk(C ST•
:= lim→ (H k(C•
), Hk(W •t))
S), H k(W •)),
where we repeat the definition of H k
emphasize the tiling T we will write H ST
is
S from Definition 4.33. When we want to
S(T) instead. Here the notation
k (T) and H k
lim→ (X, A) := lim→ X A /
/ X A /
/ X A /
/ .
In Section 4 we related the stable cohomology H k
S to the K-theory of the stable
C∗-algebra S. Next we present a series of definitions, which are necessary for us
to the K-theory of the unstable C∗-
to relate the stable-transpose homology H ST
algebra U via the Cech cohomology of Ω (cf. Section 2).
k
Define the relation
Definition 5.3 (Rn-equivalent sets). Let n ∈ Z. Two bounded subsets σ1, σ2 ⊂ Rd
are said to be Rn-equivalent, σ1 ∼Rn σ2, if there exists x ∈ Rd such that σ1 = σ2+x
and Tn(σ◦
2) + x. (Recall that for tiling T 0, T 0(σ) denotes the patch made
of all the tiles containing at least a point of σ).
Note that x in the definition is unique: If σ1 = σ2 + x then we can write
1) = Tn(σ◦
xi = sup pi(σ1) − sup pi(σ2),
: Rd → R is the projection onto the i-th coordinate. Since the sets
where pi
σ1, σ2 are bounded, their supremums on each coordinate are unique and hence
x = (x1, . . . , xn) is uniquely determined by σ1 and σ2.
It is straightforward to check that the above Rn-equivalence is an equivalence
relation on the set of bounded subsets of Rd. For instance, transitivity is just
addition of vectors. Moreover, this definition reduces to Definition 3.3(stable cells)
when applying it to the k-cells of Tn. We use the standard convention that a vertex
has no boundary.
Definition 5.4 (Combinatorial ball T 0m(σ)). Let m ∈ N0, let T 0 be a tiling, and let
σ ⊂ Rd be a bounded subset. Define the combinatorial ball T 0m(σ) of combinatorial
radius m and combinatorial center σ to be the patch T 0(··· (T 0(σ))) done m times.
The combinatorial ball T 0m(σ) induces the following relation:
/
/
D. GONC¸ ALVES
52
Definition 5.5 (T 0m-equivalent sets). Let m ∈ N0, and T 0 a tiling. Two bounded
subsets σ1, σ2 ⊂ Rd are said to be T 0m-equivalent, which we denote by σ1 ∼T 0m σ2,
if there exists x ∈ Rd such that σ1 = σ2 + x and T 0m(◦
σ1) = T 0m(◦
M. RAMIREZ-SOLANO
σ2) + x.
It is straighforward to check that the above relation ∼T 0m is actually an equiva-
lence relation on the set of bounded subsets of Rd and that the vector x is unique.
The proof is the same as for the Rn-equivalence relation.
5.1. Borel-Moore chains. We start by defining some subchain complexes of the
Borel-Moore chain complex on a tiling T 0. These definitions are equivalent to
definitions in [42].
Definition 5.6 (BM k-chain for tiling T 0). Let k ∈ {0, 1, 2} be fixed, and let
ξ :=
σ closed k-cell of T 0
Kσ σ,
X
X
where Kσ ∈ Z for all k-cells σ ∈ T 0. We say that the k-chain ξ is Borel-Moore (BM
for short).
We remark that all integers Kσ in the above sum can be non-zero. This is
in contrast with the definition of a standard cellular k-chain where only a finite
number of the Kσ's are allowed to be nonzero.
Definition 5.7 (PE k-chain for tiling T 0). Let k ∈ {0, 1, 2} be fixed, and let
ξ :=
σ closed k-cell of T 0
Kσ σ,
where Kσ ∈ Z for all k-cells σ ∈ T 0. We say that the k-chain ξ is pattern-equivariant
(PE for short) if the following condition is satisfied:
(5.2)
∃m ∈ N0 : ∀σ, σ0 k-cells of T' : σ ∼T 0m σ0 =⇒ Kσ = Kσ0.
If ξ is a PE k-chain, then by definition there exists an m0 such that the condition
in Eq. (5.2) is satisfied. We would like to remark that the condition is also satisfied
for any integer m > m0.
We define the chain complex
(T 0)
BM,PE
2
/ C
0
(5.3)
∂2
/ C
BM,PE
1
(T 0)
∂1
/ C
BM,PE
0
(T 0)
/ 0
BM,PE
k
(T 0), k ∈ {0, 1, 2}, is an abelian group whose elements are exactly all
where C
the PE k-chains, and where the differentials ∂k are the standard cellular boundary
maps.
Example 5.8. For the Fibonacci tiling, an example of a sequence (Ke)e∈T of edges
that yield a PE 1-chain with m ≥ 0 is
2
a
Ke
T
2
a
2
a
2
a
2
a
2
a
3
3
3
b
3
2
a
and with m ≥ 1 is (m=0 does not work here)
5
a
3
a
5
a
2
a
?
a
4
4
4
b
b
b
b
b
b
3
a
4
b
?
a
Ke
T
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)53
the center of the combinatorial ball.
where Kaab = 3, Kaba = 4, Kbaa = 5, Kbab = 2, and the underlined letter denotes
An example of a sequence (Kv)v∈T of vertices that yield a PE 0-chain with m ≥ 0
is
?
5
5
5
5
5
5
5
b
a
a
and with m ≥ 1 is (m=0 does not work here)
a
a
a
b
b
?
1
2
0
1
2
1
2
5
0
5
5
a
b
a
1
2
?
?
a
b
a
a
b
a
b
a
a
b
a
Kv
T
Kv
T
where Ka.a = 0, Ka.b = 1, Kb.a = 2.
Definition 5.9 ((T−', R−n) k-chain). Let n ∈ Z and let ' ≤ n. Let k ∈ {0, 1, 2}
be fixed, and let
X
ξ :=
Kσσ,
σ closed k-cell of T−'
where Kσ ∈ Z for all k-cells σ ∈ T. We say that the k-chain ξ is a (T−', R−n)
k-chain if the following condition is satisfied
(5.4)
∀σ, σ0 k-cells of T−': σ ∼R−n σ0 =⇒ Kσ = Kσ0.
A (T−', R−n) k-chain complex is a PE-chain complex for tiling T−'. This
amounts to showing that there exists an m ≥ 0 such that the map
σ ∈ T−' is a k-cell,
) := [σ]R−n ,
fm,n,k([σ]T m−'
is well-defined. For instance consider the Fibonacci tiling and ' = 0, n = 1: We get
a well-defined map on the edges (k = 1) if we choose a combinatorial radius m = 1,
and on vertices (k = 0) if we choose a combinatorial radius m = 2. Namely, using
the notation of Example 5.28, fm,n,k on the 1-balls of edges is given by
baa 7→ b0
e0
bab 7→ a0
e0
aab 7→ a0
e0
aba 7→ a0
e1,
where the underlined letter denotes the center of the combinatorial ball. On the
2-balls of vertices, fm,n,k is
aa.ba 7→ a0
v1
ab.aa 7→ a0.b0 ab.ab 7→ a0.a0
ba.ab 7→ b0.a0
ba.ba 7→ a0
v1.
Definition 5.10 (C•(T−', R−n)). Define the abelian group Ck(T−', R−n) to be
the subgroup of C BM,P E
(T−') whose elements are the (T−', R−n) k-chains. Here
k ∈ {0, 1, 2}, n ∈ Z, ' ≤ n. The chain complex C•(T−', R−n) is then a subchain
complex of the chain complex in Eq. (5.3) (with T 0 = T−') by restricting the
differentials.
k
54
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
For the fixed tiling T, we now construct the approximate sequence of chain maps
(not a short exact sequence)
(5.5)
0
0
0
0
/ C2(T, R0)
∂2
C1(T, R0)
∂1
C0(T, R0)
q2 i2
q1 i1
q0 i0
/ C2(T, R−1)
∂2
C1(T, R−1)
∂1
C0(T, R−1)
(1)
q
2
(1)
i
2
/ C2(T, R−2)
(1)
q
1
(1)
i
1
C1(T, R−2)
(1)
q
0
(1)
i
0
C0(T, R−2)
∂1
∂2
0
0
0
(2)
q
2
(2)
i
2
(2)
q
1
(2)
i
1
(2)
q
0
(2)
i
0
/ C∞
2 (T)
∂2
/ C∞
1 (T)
∂1
/ C∞
0 (T)
/ 0.
The direct limit chain complex
(1)
•
(2)
•
(0)
•
i
i
/ C•(T, R−1)
C∞
• (T) := lim→ C•(T, R0)
k (T) := ∪n∈N0 Ck(T, R−n), is quasi-isomorphic to the chain complex in
where C∞
Eq. (5.3) by [42, Lemma 4.13]. Recall that a quasi-isomorphism is by definition a
morphism of chain complexes that becomes an isomorphism after taking homology.
We will only describe in detail the chain maps between C•(T, R0) and C•(T, R−1)
because the maps for higher n behave similarly.
A basis element for Ck(T−', R−n), n ∈ Z, ' ≤ n, is a (T−', R−n) k-chain
/ C•(T, R−2) i
,
I([σ]R−n) := X
σ0,
σ0∈[σ]R−n
where σ ∈ T−' is a closed k-cell. Choose k-cell σ ∈ T' from each R−n equiva-
lence class to get a whole basis. Unless ambiguity arises, we will denote I([σ]R−n)
simply as [σ]R−n. Since R−(n+1) ⊂ R−n then for any k-cell σ ∈ T−', the set
[σ]R−n is partitioned with the equivalence relation R−(n+1). Thus a basis element
of Ck(T−', R−n) equals a (finite) sum of basis elements of Ck(T−', R−(n+1)). Hence
Ck(T−', R−n) is a subgroup of Ck(T−', R−(n+1)). Define C•(T−', R−n) to be the
transpose of C•(T−', R−n) with respect to the above basis, that is the abelian group
C k(T−', R−n) can and will be given the basis of Ck(T−', R−n), but the differentials
of C•(T−', R−n) are
(5.6)
5.2. Stable Transpose homology (ST). In this subsection we relate the stable
transpose homology with Cech cohomology. Moreover, we show that the stable
cohomology and stable transpose homology are always torsion free for tilings of
can contain
the line. For tilings of the plane, we show that only H1
torsion. In the absence of torsion, we show that H k
k =
lim→ (Znk , (Ak)t) for some matrix Ak ∈ Mnk(Z). We start with some technical
definitions and lemmas.
S = lim→ (Znk , Ak) and H ST
S and H ST0
δk := ∂t
k+1.
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)55
Definition 5.11 (replacing equivalence relation). We define the cochain map r• :
C•(T, R−1) → C•(T, R0) by
rk([σ]R−1) := [σ]R0
σ ∈ T is a k-cell,
for k ∈ {0, 1, 2}.
Definition 5.12 (inclusion). Let k ∈ {0, 1, 2}. Let σ ∈ T be a k-cell, and suppose
that [σ]R0 = [σ1]R−1 t ··· t [σj]R−1 for some j ∈ N. The inclusion chain map
i• : C•(T, R0) → C•(T, R−1) in the basis given above and evaluated at σ is
ik(I([σ]R0)) := I([σ1]R−1) + ··· + I([σj]R−1).
Since rk([σi]R−1) = [σi]R0 = [σ]R0 for i = 1, . . . , j, the inclusion ik is the trans-
pose of rk
(5.7)
Definition 5.13 (relabeled parent map). Let k, j ∈ {0, 1, 2}. Since T−1 and
λω−1(T) are the same as tilings, for every j-cell σ0 ∈ T−1 there is a cell σ ∈ ω−1(T),
such that σ0 = λσ. In particular ω(σ) ⊂ T. Moreover, σ0 and ω(σ) are equal as
sets. We say that σ0 is the parent of a k-cell τ (k ≤ j) if ◦
σ). Note that
the parent of a cell is unique and the parent always has at least the dimension of
the cell. We define the cochain map g0• : C•(T, R−1) → C•(T−1, R−1) by
τ is in ω(◦
ik = (rk)t.
(cid:26) δσ,τ · [λσ]R−1
0
g0k([τ]R−1) :=
if λσ is a k-cell of T−1
else,
where τ ∈ T is a k-cell such that ◦
σ). Informally, and ignoring signs, the map
g0k maps a cell τ to its parent λσ if λσ has the same dimension as τ, else it maps
τ to 0 (i.e. when λσ has a higher dimension).
τ ∈ ω(◦
We define the cochain map g• : C•(T, R−1) → C•(T, R0) by
gk := Λ0k ◦ g0k,
k ∈ {0, 1, 2},
where (Λ0k)−1 : C k(T0, R0) → C k(T−1, R−1) is the isomorphism that identifies the
k-stable cells σ ∈ T with the expanded k-stable cells λσ + x ∈ T−1
(Λ0k)−1([σ]R0) := [λσ + x]R1,
(5.8)
where x is a translational vector such that λσ + x ∈ T−1.
Definition 5.14 (relabeled children map). The chain map q• : C•(T, R0) →
C•(T, R−1) is defined as
k ∈ {0, 1, 2},
i.e. qk is the transpose of the relabeled parent map gk.
qk := (gk)t
This definition of q• coincides with that of q• in [42, Lemma 4.14], and by the
same lemma, q• is a quasi-isomorphism, i.e. it is an isomorphism when one takes
homology.
1 : T → T−1 be the homotopy defined in Sub-
Definition 5.15 (a section). Let h0
section 4.2 but for tiling T−1 instead of tiling T0. That is, on the (expanded)
prototiles of T−1, the definition of h0
1 is the same as the definition of h1 on the
56
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
prototiles of T0 up to expanding by the factor λ. We define the cochain map
s0• : C•(T−1, R−1) → C•(T, R−1) by
s0k([σ0]R−1) := X
δσ0,τ · [τ]R−1,
τ∈ω−1 (T−1(◦
h0
1(τ)=σ0
σ0))
where σ0 is k-cell of T−1, and τ is k-cell of T, and ω−1 : T−1 → T is the map defined
in Eq. (5.1). Note that ω−1(T−1(◦
σ0); this
set is tiled with prototiles of T0 via ω−1.
σ0)) ⊂ T0 is, as a set, the stable cell T−1(◦
We define the cochain map s• : C•(T, R0) → C•(T, R−1) as
sk := s0k ◦ (Λ0k)−1,
k ∈ {0, 1, 2},
where Λ0k was defined in Eq. (5.8).
By the following two lemmas, the maps sk commute with the differentials δk.
Lemma 5.16. For dimension d ≤ 2, the following diagram commutes
C1(T−1, R−1)
δ1
C2(T−1, R−1)
s01
s02
C1(T0, R−1)
δ1
/ C2(T0, R−1).
Proof. We will assume that d = 2, otherwise the lemma trivially holds. To help the
reader with the notation, we will use Fig. 9 as intuition. Moreover, in this proof [·]
denotes [·]R−1. Let T−1(◦
2} be a stable edge, for some edge e0 ∈ T−1 of
2 ∈ T−1, and assume e0 has same orientation as one of the edges of t0
two tiles t0
1,
and e0 has opposite orientation as one of the edges of t0
e0) = {t0
1, t0
1, t0
2. Then
δ1([e0]) = [t0
1] − [t0
2].
We remark that if t0 ∈ T−1 is a prototile (i.e. a stable face), then all the shrunk
tiles in ω−1(t0) ⊂ T0 are representatives of the (T0, R−1)-equivalence classes. By
definition of h0
1(t2) =
t0
2. Hence
1, there are unique tiles t1, t2 ∈ T0 such that h0
1(t1) = t0
1 and h0
s02(δ1([e0])) = [t1] − [t2].
1) ⊂ T0 such that e ∈ e0. Again, by definition of h0
1, there is a chain of tiles τ0τ1τ2···τn from
On the other hand, by definition of h0
τ0 := t1 to τn := t2 such that the edges i := τiτi+1 ∈ T0 homotope to e0, i.e.
1(i) = e0. Indeed, by definition of h0
h0
1 (cf. Definition 4.6(3(iii))) there is a unique
edge e ∈ ω−1(t0
1 (cf. Definition
1) ⊂ T0 that contains the edge e contains exactly two
4.6(3(iv))), a tile τ ∈ ω−1(t0
edges that homotope to e0 whenever τ 6= t1, else τ = t1. There is a finite number of
cells to go through, so the chain ends eventually in an edge of t1. By connectedness
of h−1
1) which homotope to e0 are contained in the chain.
A similar argument holds for tile t2. We can assume that the edges homotoping
to e0 preserve the orientation, since we can replace [] with [] = −[] without
changing the resulting value of δ1(s1([e0])), where is the edge but with opposite
orientation. Then
1 (e0), all the edges in ω−1(t0
s01([e0]) = [1] + ··· + [n−1].
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)57
t0
1
e0
t0
2
T−1(◦
e0)
τ0 := t1
τ1
τ2
e
τ3
τ4
τ5 := t2
ω−1(T−1(◦
e0))
Figure 9. The stable edge T−1(◦
e0)
ω−1(T−1(◦
e0)) ⊂ T0.
tiled with the patch
Thus
δ1(s1([e0])) = ([t1] − [τ1]) + ([τ1] − [τ2]) + ··· + ([τn−1] − [t2]])
= [t1] − [t2].
(cid:3)
Lemma 5.17. For dimension d ≤ 2, the following diagram commutes
C0(T−1, R−1)
δ0
C1(T−1, R−1)
s00
s01
C0(T0, R−1)
δ0
/ C1(T0, R−1).
Proof. In this proof, [·] denotes [·]R−1. Let T−1(v0) be a stable vertex, for some
vertex v0 ∈ T−1. Suppose that e0
1, . . . , e0
n are all the edges in T−1(v0) for which one
of its vertices is v0. We will assume that e0
j, j = 1, . . . , n, is oriented so that v0
j] with [e0
is a final vertex since we can replace [e0
j] without changing the
j is the edge e0
resulting value of s01(δ0([v0])), where e0
j with reversed orientation.
Then
1]) + ··· s01([e0
j] = −[e0
n]).
Suppose that T−1(◦
e1
the edges of t0
Suppose that ej1, . . . ejmj are all the edges whose interior is in ω−1(
that homotope to e0
homotoping to e0
s01(δ0([v0])) = s01([e0
0) = {t0
1 matches the orientation of one of
1 with reverse orientation matches one of the edges of t0
2.
◦
j) ⊂ T0 and
t0
1, where j = 1, 2. We assume here as well that the edges
1 preserve the orientation, since the resulting value would not
2} such that e0
1, and e0
1, t0
/
/
/
58
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
change by similar reasons as above. Then
s01([e0
1]) = [e1] + [e11] + ··· + [e1m1] + [e21] + ··· + [e2m2],
1) that homotopes to e0
1.
where e1 is the unique edge in ω−1(e0
On the other hand, consider all the vertices in ω−1(T−1(v0)) ⊂ T0 which homo-
tope to v0: If two of these vertices are connected by an edge e of T0, then, in the
resulting value of δ0(s01([v0])), we get the term +[e] (coming from the final vertex)
and the term −[e] (coming from the initial vertex), and hence, after cancellation,
[e] does not occur. After removing such edges, all the edges in δ0(s01([v0])) must
homotope to one of the e0
i's. We conclude that the terms in δ0(s01([v0])) coming
from edges homotoping to e0
1 are
[e1] + [e11] + ··· + [e1m1] + [e21] + ··· + [e2m2].
It follows that s01(δ0([v0])) = δ0(s01([v0])).
(cid:3)
Next we will prove four auxiliary results, that will be necessary in our main
theorems.
Lemma 5.18. For k ∈ {0, 1, 2}, the map sk is a section of the relabeled parent
map gk, that is
gk ◦ sk = id.
Proof. By the standard isomorphism Λ0 that identifies stable cells T with stable
cells of T−1, it suffices to show that g0k ◦ s0k = id. We get
g0k(s0k([σ0]R−1)) =
δσ0,τ · g0k([τ]R−1) = [σ0]R−1 ,
X
τ∈ω−1 (T−1(◦
σ0)), h0
1(τ)=σ0
σ0)) with h0
where the last equality is because of the following: First, σ0 is a k-cell in T−1.
Second, all the k-cells τ ∈ ω−1(T−1(◦
1(τ) = σ0 that don't lie inside σ0 (as
sets) vanish under g0k because the parent of τ is a j-cell of T−1 of dimension j > k.
Third, exactly one k-cell τ0 ⊂ σ0 homotopes to σ0 (cf. Subsection 4.2 adapted to
1). Hence τ0 is the only k-cell for which g0k is nonzero. Since g0k([τ0]R−1) = [σ0]R−1
h0
(cid:3)
the equality follows.
Lemma 5.19. We have
(5.9)
Proof. Recalling that C•(T, R0) := C•(T, R0)t where the right hand side is given
the standard basis, it is easy to see that
C•
S = C•(T, R0)
C ST• = C•(T, R0).
and
C•
S = C•(T, R0).
S are the stable classes [σ]R0, where σ ∈ T0 is a
Namely, the basis elements of C k
k-cell. As explained at the end of Subsection 5.1, the basis elements of C k(T, R0)
are also the equivalence classes [σ]R0 where σ ∈ T0 is a k-cell. Moreover, the
differentials for C•
S in Eq. (1.7) and Eq. (1.8) agree with those for C•(T, R0) given
in Eq. (5.6). Applying the transpose to both sides, we get
C ST• = C•
S
t = C•(T, R0)t = (C•(T, R0)t)t = C•(T, R0).
(cid:3)
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)59
Lemma 5.20. For k ∈ {0, 1, 2}, the connecting maps W k : C k
S can be
obtained from the replacing-equivalence-relation chain map rk and the section chain
map sk by the following equality:
S → C k
W k = rk ◦ sk,
k ∈ {0, 1, 2}.
Proof. By the standard isomorphism Λ0k in Eq. (5.8) that identifies stable k-cells of
T−1 with stable k-cells of T, and by Λk defined similarly, that identifies the stable
k-cells of T with stable k-cells of T1, it suffices to show that W k0 : C k(T0, R0) →
C k(T1, R1) given by
W k
0 ([σ]R0) :=
δσ,τ [τ]R1
τ∈ω0(T (σ◦)), h1(τ)=σ
has the same matrix as rk◦s0k : C k(T−1, R−1) → C k(T, R0) in their standard basis.
This holds because the composition is
X
X
X
rk(s0k([σ0]R−1)) =
=
τ∈ω−1 (T−1(◦
σ0)), h0
1(τ)=σ0
τ∈ω−1 (T−1(◦
σ0)), h0
1(τ)=σ0
δσ0,τ · rk([τ]R−1)
δσ0,τ · [τ]R0.
(cid:3)
Lemma 5.21. For k ∈ {0, 1, 2}
Hk(W •t) = Hk(q•)−1Hk(i•).
Proof. Applying the functor Hk(( )t) to the cochain map W • : C•
the following map
S → C•
S we get
Hk(W •t) : Hk(C•
S
t) → Hk(C•
S
t).
Since W • = r• ◦ s• and g• ◦ s• = id• and r• = it• and g• = qt•, and q• is a
quasi-isomorphism we have that
Hk(W •t) = Hk(s•t)Hk(r•t) = Hk(q•)−1Hk(i•).
(cid:3)
Recall that the tiling space Ω is assumed to be FLC, and that the substitution
map ω is primitive and injective.
Theorem 5.22. Let T ∈ Ω be a tiling of dimension less or equal to two, with convex
prototiles. For k ∈ {0, 1, 2}, the stable-transpose-homology groups are related to the
Cech cohomology groups by
k (T) = H d−k(Ω).
H ST
Proof. Let
H∞
k := lim→ (Hk(C•(T, R0)), Hk(q•)−1Hk(i•)).
Then, by Lemma 5.19 and Lemma 5.21, we have that
k ∈ {0, 1, 2}.
k = H∞
H ST
k
The theorem follows by the previous equality and by the isomorphisms
H∞
k
∼= H P E
k
∼= H d−k
P E
∼= H d−k(Ω),
60
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
where the first isomorphism is by [42, Thm. 4.5], the second is by [42, Thm. 2.2], and
the third isomorphism is in [23], [25], [38]. We should remark that the first isomor-
phism factors through Hk(C∞
• (T)) are
obtained by taking direct limit of the rows in the following commutative diagram
whose vertical maps are isomorphisms
(5.10)
• (T)). Namely, the groups H∞
k and Hk(C∞
Hk(C•(T, R0)) Hk(W • t) /
id
Hk(C•(T, R0)) Hk(W • t) /
Hk(q•)
Hk(q
/ Hk(C•(T, R−2)) Hk(i
Hk(C•(T, R0)) Hk(W • t) /
• ◦q•)
(1)
(2)
• ) /
(1)
• ) /
/ Hk(C•(T, R−1)) Hk(i
Hk(C•(T, R0)) Hk(i•) /
(cid:3)
and one also uses that C∞
Remark 5.23. For tilings with non-convex prototiles, one can either refine the
substitution using convex tiles, or adapt a recipe similar to the Kites-Darts Penrose
tiling shown in [42, Example 4.3].
(T) is a quasi-isomorphism.
• (T) ,→ C BM,P E
/ ,
•
We should remark that we denote cohomology groups in the following theorem
using the notational convention given in Subsection 2.1.
Theorem 5.24. With notation from the beginning of this section, the following
diagram commutes, and the rows in the diagram are short exact sequences which
split (non-canonically)
0
0
/ Ext1Z(H k+1(C•
S), Z)
Hk(C ST•
)
Hom(H k(C•
S), Z)
Ext1Z(H k+1(W •),Z)
Hk(W • t)
Hom(H k(W •),Z)
/ Ext1Z(H k+1(C•
S), Z)
/ Hk(C ST•
)
/ Hom(H k(C•
S), Z)
0
/ 0.
S → C•
Proof. Applying the universal coefficient theorem (UCT) for cohomology to the
chain map W • : C•
s (using the notation of Subsection 2.1) we get the
following commutative diagram, where the rows are short exact sequences which
split (non-canonically)
/ Ext1Z(H k+1(C•
0
Hom(H k(C•
Hk(Hom(C•
S), Z)
S, Z))
S), Z)
Ext1Z(H k+1(W •),Z)
Hk(Hom(W •,Z))
Hom(H k(W •),Z)
0
/ 0.
(cid:3)
S), Z)
/ Ext1Z(H k+1(C•
0
The result follows by identifying Hom(C•
Definition 5.25 (S-tor, ST-torFree). Define the following homology groups
S, Z) with C ST•
/ Hk(Hom(C•
/ Hom(H k(C•
S), Z)
S, Z))
.
:= lim→
:= lim→
(cid:0)Ext1Z(H k(C•
(cid:0)Hom(H k(C•
) ∼= Ext1Z(H k+1(C•
HS-tor
HST-torFree
k
k
Hk(C ST•
S), Z), Ext1Z(H k(W •), Z)(cid:1)
S), Z), Hom(H k(W •), Z)(cid:1).
S), Z) ⊕ Hom(H k(C•
S), Z),
Corollary 5.26. We have
where the isomorphism is non-canonical. Note that the stable-torsion part moves
from level k + 1 to level k of the unstable-torsion part.
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)61
Moreover, HS-tor
k
is a torsion group, HST-torFree
k
is a torsion free group, and
0
/ HS-tor
k+1
/ H ST
k
/ HST-torFree
k
/ 0
is a short exact sequence (might not split). In particular,
(1) if T is a one-dimensional tiling, then there is no stable nor unstable torsion
part.
S = Z are torsion free, but H1
H2
the groups H ST2 = Z, H ST1
(2) if T is a two-dimensional tiling, then in the stable case, the groups H0
S and
S can contain torsion; in the unstable case,
can contain torsion.
Proof. The direct sum follows from the short exact sequence of the top row of the
diagram in the theorem, which splits (non-canonically).
are torsion free, but H ST0
The group HST-torFree
S), Z), we get that HS-tor
k+1
Recall that the direct limit of a directed system of short exact sequences of
abelian groups is a short exact sequence. This is because the direct limit is an
exact functor in the category of abelian groups. From this fact and the theorem,
the short exact sequence of the corollary follows.
direct sum of G's is a torsion group. Since Ext1Z(H k+1(C•
and the direct limit HS-tor
k+1
Ext1Z(H k+1(C•
group).
for some n ∈ N (here we used that Hom(H k(C•
abelian group). (cf. Appendix A).
Recall that if G is an abelian torsion group then any quotient of an infinite
S), Z) is a torsion group
is in particular a quotient of an infinite direct sum of
is a torsion group (it might be the zero
is torsion free since it is isomorphic to a subgroup of Qn
S), Z) is a finitely generated free
S) = Z by a similar proof to
Lemma 4.5, and thus by the direct sum in this corollary, there is no torsion to
move to H0(C ST•
) = 0, there can be no torsion to
) = H0(Ω) = Z. Since before taking the direct
move from H0(C•
limit there is no torsion in any of the groups, after taking the direct limit there is
nor in H•
S.
still no torsion in any of the groups, that is, there is no torsion in H ST•
The conclusion for tilings of dimension 2 is deduced similarly from the following
table
). Furthermore, since H−1(C ST•
S). Finally, H1(C ST•
If the tiling T is of dimension d = 1 then H1(C•
k
Unstable
Stable
H2(C ST•
) = Z
H2(C•
S) = Z
H1(C ST•
)
torsion free
H0(C ST•
)
H1(C•
S)
H0(C•
S).
(cid:3)
Corollary 5.27. If T ∈ Ω is a tiling of dimension 1, or if T is of dimension 2
with H1(C•
S) torsion free then
H ST
k
H k
S
∼= lim→
∼= lim→
Zn
At
k
/ Zn
At
k /
/ Zn
At
k
Zn Ak
/ Zn Ak /
/ Zn Ak
k ∈ {0, 1, 2}
k ∈ {0, 1, 2}.
/
/
/
/
s
s
s
s
/
/
/
/
/
/
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
k is the transpose of Ak.
62
where Ak ∈ Mn(Z) is the matrix of H k(W •) with respect to a fixed basis of H k(C•
S),
and At
Proof. By assumption and by Corollary 5.26, H k(C•
direct limits cannot introduce torsion, H k
sequence in Corollary 5.26 we get that
S) is torsion free. Since taking
S is also torsion free. By the short exact
(cid:0)Hom(H k(C•
S), Z), Hom(H k(W •), Z)(cid:1),
k ∈ {0, 1, 2}.
k = lim→
H ST
Since H k(C•
S) is a finitely generated torsion free abelian group, it is isomorphic
to Zn for some n ∈ N0. Choosing such an isomorphism, that is, giving a basis to
H k(C•
S), then in the dual basis of H k(C•
S), the matrix for the dual map
Hom(H k(W •), Z) : Hom(H k(C•
S), Z) → Hom(H k(C•
S), Z)
is well-known (and easily shown) to be the transpose of the matrix. That is
Hom(H k(W •), Z) = (H k(W •))t. The corollary follows immediately.
Note that,
(5.11)
H ST
2
H ST
1
H ST
0
= lim→ (coker δ1)t
= lim→ ( ker δ1
Im δ0 )t
= lim→ (ker δ0)t
W t
F /
/ (coker δ1)t
W t
F
/ ···
W t
E /
/ ( ker δ1
Im δ0 )t
W t
E /
/ ( ker δ1
Im δ0 )t
W t
E /
/ ···
W t
V /
/ (ker δ0)t
W t
V /
/ (ker δ0)t
W t
V /
/ ···
(cid:3)
5.3. Examples. We denote with (T0, R0) the tiling T0 with cells partitioned by
the equivalence relation R0. The equivalence classes [σ]R0, where σ is a k-cell of
T, are the stable vertices if k = 0, the stable edges if k = 1, and the stable faces
if k = 2. For tilings of the line, we denote the stable edges with the labels of the
prototiles, e.g. a, b. The stable vertices we denote them by pairs of edges with a
dot in the middle to denote the vertex, e.g. a.b.
Similarly, we denote with (T−1, R−1) the tiling T−1 with cells partitioned by the
equivalence relation R−1. The R−1-equivalence classes of cells of T−1 are the stable
cells of T−1. For tilings of the line we denote the stable edges and stable vertices
of T−1 with primes on the labels of the prototiles, e.g. a0, b0, a0.b0.
Similarly, we denote with (T0, R−1) the tiling T0 with cells partitioned by the
equivalence relation R−1. However, the similarity with the above ends here. Each
cell in T0 has a parent cell in T−1. The R−1 equivalence classes of cells in T0 are
denoted by the label of the stable parent cell (which lives in T−1) together with a
unique sub-label to distinguish the siblings from each other.
Example 5.28 (Fibonacci tiling). Consider the Fibonnacci tiling with protoedges
a, b of lengths a = 1, b = 1/φ, (λ = φ=golden ratio) and substitution rule
ω(a) = ab, ω(b) = a. The pairs (T0, R0), (T−1, R−1) (T0, R−1) are drawn below:
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)63
a
b
a
a
b
a
b
a
a
b
a
(T0, R0)
a0
v1 a0.b0
a0
e1
a0
e0
a0
b0.a0
b0
e0
b0
a0
v1 a0.a0
a0
e1
a0
e0
a0
e0
a0
v1 a0.b0
a0
e1
a0
a0
b0.a0
b0
e0
b0
a0
v1 a0.b0
a0
e0
a0
e1
a0
b0
e0
b0
(T0, R−1)
h0
t
(T−1, R−1)
The matrices for the replacing-the-equivalence-relation maps rk are
a0.a0 a0.b0
0
0
0
0
1
1
b0.a0 a0
v1
1
0
1
0
0
0
!
r0 =
a.a
a.b
b.a
For instance r0(a0.a0) = ω(a).ω(a) = ab.ab = b.a.
(cid:18)
a0
e0
1
0
a0
e1
0
1
(cid:19)
b0
e0
1
0
r1 =
a
b
For instance r1(a0
e1) = r1(ω(a)e1) = b because the second edge of ω(a) = ab is b.
The matrices for the parent maps g0k are
a0.a0 a0.b0
0
1
1
0
0
0
!
b0.a0 a0
v1
0
0
0
0
1
0
g00 =
a0.a0
a0.b0
b0.a0
(cid:18)
a0
e0
1
0
a0
b0
(cid:19)
b0
e0
0
1 .
a0
e1
1
0
g01 =
The matrices for the relabeled parent maps gk are
a0.a0 a0.b0
0
1
1
0
0
0
b0.a0 a0
v1
0
0
0
0
1
0
!
g0 =
a.a
a.b
b.a
(cid:18)
a0
e0
1
0
a
b
a0
e1
1
0
(cid:19)
b0
e0
0
1 .
g1 =
64
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
The matrices for the section maps s0k are
s00 =
a0.a0
a0.b0
b0.a0
a0
v1
s01 =
a0.a0 a0.b0
0
1
1
0
0
0
1
1
b0.a0
0
0
1
0
a0
1
0
0
!
.
b0
0
0
1
a0
a0
b0
e0
e1
e0
The matrices for the section maps sk are
s0 =
a0.a0
a0.b0
b0.a0
a0
v1
s1 =
b.a
0
0
1
0
!
.
a0
a0
b0
e0
e1
e0
a.a a.b
0
1
1
0
0
0
1
1
a b
1
0
0
0
0
1
It is easy to check that gk ◦ sk = id, as expected by Lemma 5.18. It is also easy
to check that r0 ◦ s0 = WV in Eq. (7.2) and r1 ◦ s1 = WE in Eq. (7.1), as expected
by Lemma 5.20.
6. Asymptotic C∗-algebra A
r (Rn), ιn).
n→∞(C∗
In this section we compute the K-theory of the asymptotic C∗-algebra A. To do
so we will use the Kunneh formula, and so we will first prove that both S0 and S
are UCT.
Recall that, by Section 3 and Section 4, the C∗-algebra S is Morita equivalent
to the transversal S0 = lim
Theorem 6.1. For tilings of dimension d = 1, 2, the C∗-algebras C∗
I. Hence S0 and S are both UCT and amenable.
r (RnX0) is isomorphic to a finite direct sum of the
Proof. By Proposition 4.2(1), C∗
r (RnX1−X0) is isomorphic
compacts. Hence it is type I. By Proposition 4.2(2) C∗
to a finite direct sum of C0((0, 1), K). Hence it is also type I. Since the class of type
r (RnX1) is type I by Eq. (4.1). This
I C∗-algebras is closed under extensions, C∗
proves that C∗
r (RnX2−X1) is type I. Hence, by Eq. (4.2)
By Proposition 4.2(3) the ideal C∗
r (Rn) is type I. Hence C∗
C∗
Since type I is UCT (cf. [8, p.229]), C∗
r (Rn) is UCT. Since the class of UCT
C∗-algebras is closed under direct limits (cf. [8, p.229]) the stable transversal C∗-
algebra S0 is UCT as well. Since type I C∗-algebras are amenable, and amenability
r (Rn) is also type I for tilings of dimension 2.
r (Rn) is type I for tilings of dimension 1.
r (Rn) are type
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)65
is preserved under direct limits, S0 is amenable. By the third line in [8, p.229], UCT
is preserved by Morita equivalence. Hence S is UCT. Amenability is also preserved
by Morita equivalence. Hence S is also amenable. (Note: In [8] the UCT class of
(cid:3)
C∗-algebras is called class N, cf. [8, Definition 22.3.4]).
Corollary 6.2. For tilings of dimension 1,
K0(A) = (cid:0)K0(S) ⊗Z K0(U)(cid:1) ⊕ Z,
K1(A) = K0(S) ⊕ K0(U),
where A, S and U are the asymptotic, stable, and unstable C∗-algebras, respectively.
Proof. By [31, Theorem 3.1] the C∗-algebras A and S ⊗max U are strongly Morita
equivalent. The Kunneth formula [8, Theorem 23.1.3, p.234] holds here because:
1) Ki(S), Ki(U) i = 0, 1 are torsion free (Corollary 5.26, Theorem 4.34).
2) S is UCT by the theorem.
The explicit formulas for the Kunneth formula are
K0(S ⊗ U) =(cid:0)K0(S) ⊗Z K0(U)(cid:1) ⊕(cid:0)K1(S) ⊗Z K1(U)(cid:1),
K1(S ⊗ U) =(cid:0)K0(S) ⊗Z K1(U)(cid:1) ⊕(cid:0)K1(S) ⊗Z K0(U)(cid:1).
We do not need to worry about the kind of tensor product S ⊗ U since S and U
(cid:3)
are amenable. Since K1(S) = Z and K1(U) = Z the corollary follows.
For tilings of the plane we have, by Theorem 4.34, that K1(S0) has torsion only
S(T) has torsion. Note that K0(S0) cannot have torsion, because if an element
if H1
has torsion in K0(S0) then by injectivity of the inclusion in the short exact sequence
in Theorem 4.34, the element does not come from the ideal, as the ideal has no
torsion, and thus the element must map to the quotient, with torsion. This is a
contradiction since H0
S has no torsion by Corollary 5.26.
By the Kunneth formula, Corollary 5.26, and by a similar proof as the one above,
where A, S and U are the asymptotic, stable, and unstable C∗-algebras, respectively.
7. Examples
In this section, we calculate the stable and unstable K-theories for a number of
tilings of the line and of the plane. We then use Corollary 6.2 and Corollary 6.3
to obtain the asymptotic K-theories. The simplification of the tensor products are
done using Proposition A.15, Corollary A.16 and Corollary A.17.
Many of the following computations of direct limits could be done by hand. We
choose however to use the general formulas stated in Appendix A, which rely heavily
on the Smith normal form of integer matrices, in order to illustrate the use of them.
Moreover, we have programmed these formulas in Mathematica. For instance, we
have written functions in Mathematica that, in the absence of torsion, yield the
∼= Zr for integer matrices A, B. The
isomorphisms ker A ∼= Zr, coker A ∼= Zr, ker A
Mathematica files for the examples found in this section can be downloaded at
Im B
we have
Corollary 6.3. For tilings of dimension 2, if H1
then
S(T) and H ST0
(T) are torsion free,
K0(A) = (K0(S) ⊗Z K0(U)) ⊕ (K1(S) ⊗Z K1(U)),
K1(A) = (K0(S) ⊗Z K1(U)) ⊕ (K1(S) ⊗Z K0(U)),
66
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
https://github.com/mariars/Tilings-Ktheory
We should remark that in this section we use the notation for direct limit
lim→ (A, X) := lim→ X A /
/ X A /
/ X A /
/ ,
where we write the matrix first, in order to emphasize it, in contrast to our previous
notation where we wrote the group first.
7.1. One dimensional tilings. Let T be a tiling of the line and let e1, . . . , eN ∈ T
be the prototiles(=proto-edges). For these tilings we can always put the homotopy
that homotopes the leftmost edge in 1
λ ω(ei) to ei. Morever, we put the orientation
of the vector (1, 0) ∈ R2 to all the edges of the tiling.
Example 7.1 (Fibonacci tiling). (cf. [1, Ex. 1, p.30]). Let T be the Fibonacci
tiling with proto-edges a, b ∈ T and substitution rule ω(a) := ab, ω(b) := a. The
length of the interval a is 1, and the length of b is 1/φ, where φ is the golden ratio.
The inflation factor λ = φ ≈ 1.618. We illustrate this and the homotopy in the
following figure:
a
λ
b
λ
a
a
λ
b
a
λ
b
λ
a
λ
b
λ
a
a
a
λ
b
a
λ
b
λ
a
a
λ
b
T1 = 1
λ ω(T)λ
hs
T0 := T
• stable edges (2): a, b.
• stable vertices (3): a.a, a.b, b.a
Note that b.b never occurs. Recall that the stable edges are always the proto-edges
for 1-dimensional tilings. Since ω3(a) and ω4(a) both contain precisely the stable
vertices a.a, a.b, b.a and w(b) = a, these are all the stable vertices in the tiling.
The prototiles of the shrunk tiling T1 := 1
The substitution-homotopy map WE : ZsE → ZsE is given by WE(a) = a0,
WE(b) = a0. Thus its matrix is
λ, b0 := b
λ.
λ ω(T)λ are a0 := a
(cid:18)
(cid:19)
a b
1
1
0 .
0
a0
b0
(7.1)
WE =
The substitution-homotopy map WV : ZsV → ZsV is given by
WV (a.a) = a0.b0 + b0.a0,
WV (a.b) = a0.b0 + b0.a0,
WV (b.a) = a0.a0
Thus its matrix is
(7.2)
WV =
a0.a0
a0.b0
b0.a0
a.a a.b
0
0
1
1
1
1
!
.
b.a
1
0
0
The exponential map δ0 : ZsV → ZsE is given by
δ0(a.b) = a − b
δ0(a.a) = a − a = 0,
δ0(b.a) = b − a.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)67
Thus its matrix is
By Proposition A.10,
(cid:18)
δ0 =
a
b
(cid:19)
a.a a.b
0
0 −1
b.a
1 −1
1 .
where p : ker δ0 → Zn−r and q : Zn−r → ker δ0 are Z-isomorphisms defined in the
proposition. Since
lim→ (WV , ker δ0) = lim→ (p WV q, Zn−r),
(cid:18) 0
1
(cid:19)
1
1
W 0
V := p WV q =
is symmetric (i.e. W 0
and zero stable-transpose homology group are the same:
V is equal to its transpose) the zero stable cohomology group
S(T) = lim→ (WV , ker δ0) = lim→ (W 0
V , Zn−r) =
H0
lim→ ((W 0
V )t, Zn−r) = lim→ ((WV )t, (ker δ0)t) = H ST
Thus the zero K-groups for the stable and unstable C∗-algebras are the same:
(T).
0
K0(S) = H0
S(T) = H ST
0
(T) = K0(U).
Since W 0
V has determinant 1, it is Z-invertible and thus
K0(U) = K0(S) = H0
S(T) = lim→ (W 0
V , Z2) = Z2.
Example 7.2 (Morse Tiling). (cf. [1, p.33]). Let T be the Morse tiling with proto-
edges a, b ∈ T and substitution rule ω(a) = ab, ω(b) = ba. The length of the edges
a, b are 1, and inflation factor is λ = 2. We follow the same procedure as for the
Fibonacci tiling so we omit most of the details.
• stable edges (2): a, b
• stable vertices (4): a.a, a.b, b.a, b.b
0 0
1 1
1 0
0 1
,
1
0
1
0
0
1
1
0
WV =
By Proposition A.10,
(cid:18) 1
0
(cid:19)
,
0
1
δ0 =
(cid:18) 0
1 −1
1
0 −1
(cid:19)
.
0
0
WE =
lim→ (WV , ker δ0) = lim→ (p WV q, Zn−r),
where p : ker δ0 → Zn−r and q : Zn−r → ker δ0 are the Z-isomorphisms defined in
the proposition. Since
0
W 0
V := p WV q =
1
1
1
(T) = K0(U).
Using Proposition A.9 we remove the zero eigenvalues of W 0
is equal to its transpose, K0(S) = H0
1
0
S(T) = H ST0
V , ker δ0) = lim→ (p0 W 0
V q0, Zn−r),
lim→ (W 0
0
1
0
V , i.e.
68
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
where p0 : W 0
tion, and computing we get
Zn → Zr and q0 : Zr → W 0
V
V
Zn are the maps defined in the proposi-
W 00
V := p0 W 0
V q0 =
(cid:18) 1
2
(cid:19)
.
1
0
/ lim→ (W 00
V , ker q(W 00
V ))
Using Proposition A.14 we extract the eigenvalue -1 of W 00
0
where p(x) = (x + 1)(x − 2) = (x + 1)q(x) is the minimal polynomial of W 00
Proposition A.10, lim→ (W 00
A.9 lim→ (W 00
exact sequence
V . By
V )) = lim→ (2, Z) = Z[1/2], and by Proposition
V )) = lim→ (−1, Z) = Z. Since the short
V , i.e.
/ lim→ (λIn, q(W 00
V )) = lim→ (−I, Im q(W 00
V , ker q(W 00
V , Im q(W 00
/ lim→ (W 00
V , Zn)
V )Zn)
/ 0 ,
0
/ Z[ 1
2]
/ lim→ (W 00
V , Zn)
/ Z
/ 0
splits, we get
K0(U) = H ST
0
(T) = H0
S(T) = K0(S) = Z[1
2] ⊕ Z.
Example 7.3 (Pathologic). Let T be the Pathologic tiling with proto-edges a, b ∈ T
and substitution rule ω(a) = babbaaa, ω(b) = abbbbb. The length of edge a is
√
≈ 6.30. (To
compute the lengths see [1, Section 8, p.26]). We follow same procedure as for the
Fibonacci tiling and Morse tiling so we omit most of the details.
≈ 1.30 and of b is 1, and the inflation factor is λ = 9+
13−1
2
√
2
13
• stable edges (2): a, b
• stable vertices (4): a.a, a.b, b.a, b.b
2 3
WV =
0
1
1
4
By Proposition A.10,
2 1
2 2
1 1
0
1
0
5
,
(cid:18) 0
1
(cid:19)
,
1
0
δ0 =
(cid:18) 0
1 −1
1
0 −1
(cid:19)
.
0
0
WE =
where p : ker δ0 → Zn−r and q : Zn−r → ker δ0 are the Z-isomorphisms defined in
the proposition, and we get
lim→ (WV , ker δ0) = lim→ (p WV q, Zn−r),
2
2
1
.
3
2
6
0
1
4
W 0
V := p WV q =
We now extract the eigenvalue -1 of W 0
V as follows. By the proof of Proposition
A.14, (and Propositions A.10, A.9) we get the following commutative diagram with
exact rows
0
0
/ Z2
/ Z3
/ Z
/ 0
W 00
V
W 0
V
/ Z2
/ Z3
−1
/ Z
/ 0,
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)69
where
W 00
V :=
(cid:18) 7 −1
3
2
(cid:19)
.
We now get by the proof of Proposition A.14 that,
V , Z3) = Z ⊕ lim→ (W 00
(cid:18) 1 −1
(cid:18) 7 −1
lim→ (W 0
(cid:19)
(cid:18) 1 −1
Since
(cid:19)
1
0
.
3
2
.
1
0
W 000
V :=
V , Z2).
(cid:19)−1
(cid:18) 3
1
=
(cid:19)
1
6
we get
.
1
1
3
4
.
1
3
4
−3
−3
(cid:19)
(cid:19)
, Z2).
V , Z2).
V and get
(cid:18) 3
lim→ (W 00
(cid:19)
(cid:19)
Since
we get
V )t, Z3) = Z ⊕ lim→ (
lim→ ((W 0
(cid:18) 4 −1
(cid:19)−1
(cid:18) 5
Similarly, using Proposition A.14, we extract the eigenvalue -1 of the transpose of
W 0
V , Z3) = Z ⊕ lim→ (W 000
(cid:18) 5
(cid:18) 4 −1
(cid:18) 3
(T) = K0(U) = Z⊕lim→ (
V , Z2).
V , Z2) cannot be written as a direct sum even though it
V , Z2) is of rank 2, as the matrix
V are 9±√
, hence two distinct
2
V is the same as the characteristic
V , and is irreducible over Q. Let λ := 9+
. By [11, Prop. 4]
K0(S) = H0
We now show that lim→ (W 000
has rank two! First note that the group lim→ (W 000
has determinant 17. Since the eigenvalues of W 000
irrational numbers, the minimal polynomial for W 000
polynomial for W 000
the direct limit lim→ (W 000
, Z2) = Z⊕lim→ (W 000
S(T) = H ST
0
V , Z2) is quasi-isomorphic (as abelian groups) to
Lλ := R[ 1
λn q ∈ R, n ∈ Z},
] = { q
λ
=
(cid:19)
13
√
2
1
6
1
1
1
6
13
1
,
where R is the ring of algebraic integers in the quadratic extension Q[λ] = Q[
{q1 + q2
13 q1, q2 ∈ Q}. That is,
√
√
13] =
√
R = { m + n
2
13
m, n ∈ Z},
13
= 9 +
13
= λ
√
mp + np
2
√
√
2
√
mp + np
2
13
= 47 + 9
2
13
= λ2
since 13 is congruent to 1 mod 4. Note that λ ∈ R. Since 13 is a prime number,
it is a square free integer. Moreover, it is not a unit in R since λ−1 = 9−√
6∈ R.
Furthermore there is no prime number p such that λ ∈ pR because the equation
13
34
implies that p must divide 9 and 1. Similarly, there is no prime number p such that
λ2 ∈ pR because the equation
70
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
V , Z2) is strongly indecomposable, since lim→ (W 000
implies that p must divide 47 and 9. Hence, by [11, Prop. 9] λ is strong. By [11,
Thm. 6] and the remark below the theorem, Lλ is an E-ring whose additive group is
strongly indecomposable (cf. [2, Thm. 14.3, p. 163]). Hence by [32, Thm. 6, p. 49]
lim→ (W 000
V , Z2) is quasi-isomorphic to
the additive group of Lλ. Recall that a ring R is said to be an E-ring if R+, the
additive group of R, is group isomorphic to End(R) via the map r 7→ (x 7→ rx).
A group G is said to be strongly indecomposable if for all integers n 6= 0 and all
abelian subgroups A, B:
(nG ⊂ A ⊕ B ⊂ G) =⇒ (A = 0 or B = 0).
A group G is said to be indecomposable if for all abelian subgroups A, B:
(G = A ⊕ B) =⇒ (A = 0 or B = 0).
Thus lim→ (W 000
sum of two nonzero abelian subgroups!
V , Z2) is an indecomposable group, i.e. it cannot be written as a direct
In the following example we discuss briefly the computation of the stable coho-
mology and stable-transpose homology for several more tilings of the line.
Example 7.4 (more 1-dimensional tilings).
Tiling OneFifth:(cf. [13, Example 2.6]).
√
5
2 = 3.618
√
• Inflation factor: λ = 5+
• Proto-edges(=stable edges) (2): a,b, a =
• substitution: ω(a) = aba, ω(b) = bbab
• Stable vertices (3): a.b, b.a, b.b
• K0(U) = H ST0
(cid:19)
S(T) = K0(S) = lim→ (
(T) = H0
2 = 0.618, b = 1
5−1
(cid:18) 3
1
(cid:19)
1
2
, Z2) = Z[ 1
5]2.
(cid:18) 3 1
The substitution-homotopy matrix WV is calculated in a similar way as for the
Fibonacci tiling. We then use Proposition A.10 to compute the matrix W 0
V :=
. Since it is equal to its transpose, the stable 0-cohomology group equals
1 2
the stable-transpose 0-homology group. By Proposition A.2,
lim→ (W 0
V , Z2) ⊂ Z[1
5]2.
We then check with a computer that the powers 5kW 0−k
are integer matrices for
small values of k ∈ N. Diagonalizing the matrix W 0
V in C, we compute explicitly
W 0−k
, and using an induction argument we show that the above inclusion is actu-
ally an equality. We learn from this example that this direct limit is of the form
Z[
V
V
1
det A]2.
√
Tiling OneSixth:(cf. [13, Example 1.21]).
3 ≈ 4.73
• Inflation factor: λ = 3 +
• Proto-edges(=stable edges) (2): a,b, a =
• substitution: ω(a) = bbaaab, ω(b) = bbab
• Stable vertices (4): a.a, a.b, b.a, b.b
• K0(U) = H ST0
(T) = H0
S(T) = K0(S) = lim→ (
√
3 ≈ 1.73, b = 1
(cid:18) 6 −2
(cid:19)
3
0
, Z2) = Z[ 1
6]2
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)71
(cid:18) 6 −2
(cid:18) 6
0
−2
3
(cid:19)
3
0
(cid:19)
, Z2) = Z[ 1
6]2
, Z2) = Z[ 1
6]2.
• H0
• K0(U) = H ST0
S(T) = K0(S) = lim→ (
(T) = lim→ (
(cid:19)
(cid:18) 6 −2
V :=
The substitution-homotopy matrix WV is calculated in a similar way as for the
Fibonacci tiling. We then use Proposition A.10 and Proposition A.9 to compute
the matrix W 0
. The direct limit is computed the same way as we
did for the tiling OneFifth and is also of the form Z[
1
det A]2. Moreover, this example
shows that the collared equivalence relation Rc given in Definition 2.2 is different
from the equivalence relation ∼1 defined in [1, Section 4], even though both yield
the same Cech cohomology (cf. Section 2).
3
0
Tiling nonReducible-4-Letter:
• Inflation factor: λ ≈ 2.508
• Proto-edges(=stable edges) (4): a, b, c, d a ≈ 1.966, b ≈ 0.542, c ≈
• substitution: ω(a) = aad, ω(b) = cd, ω(c) = cb, ω(d) = ab
• Stable vertices (9): a.a, a.b, a.d, b.a, b.c, c.b, c.d, d.a, d.c
• K0(U) = H ST0
S(T) = K0(S) = Z5
0.359, d = 1
(T) = H0
We compute the substitution-homotopy matrix WV in a similar way as we did for
the Fibonacci tiling. We then use Proposition A.10 and Proposition A.9 to remove
the eigenvalue 0. The result matrix is a 5 × 5 matrix with determinant 1.
Tiling PeriodDoubling:
• Inflation factor: λ = 2
• Proto-edges(=stable edges) (2): a, b, a = 1, b = 1
• substitution: ω(a) = bb, ω(b) = ba
• Stable vertices (3): a.b, b.a, b.b
• K0(U) = H ST0
S(T) = K0(S) = Z ⊕ Z[ 1
2]
(T) = H0
We compute the substitution-homotopy matrix WV in a similar way as we did for
the Fibonacci tiling. We then use Proposition A.10 to get the matrix lim→ (
which also occurs in the Morse tiling example, hence the same direct limit is ob-
tained.
1
0
2
, Z2),
(cid:18) 1
(cid:19)
Tiling Rauzy:
• Inflation factor: λ ≈ 1.839
• Proto-edges(=stable edges) (3): a, b, c a ≈ 1.839, b ≈ 1.543, c = 1
• substitution: ω(a) = ab, ω(b) = ac, ω(c) = a
• Stable vertices (5): a.a, a.b, a.c, b.a, c.a
• K0(U) = H ST0
S(T) = K0(S) = Z3
(T) = H0
We compute the substitution-homotopy matrix WV in a similar way as we did for
the Fibonacci tiling. We then use Proposition A.10 to get a 3 × 3 matrix with
determinant 1.
Tiling Rudin-Shapiro:
where
W 0
V :=
−2 1 −1
−1 2
1
0
0
2
,
Since
By Proposition A.6 and Proposition A.8,
V , Z3) = Z[1
2] ⊕ lim→ (
lim→ (W 0
(cid:18) 1 1
(cid:18) 3
(cid:19)
(cid:18) 1
(cid:19)
(cid:18) 3 1
(cid:18) 4
(cid:19)
(cid:19)
, Z2) = lim→ (
1
we get, by Proposition A.4, that
lim→ (
1 3
0 1
1
3
1
0
.
.
(cid:18) 4
W 0
V
2 =
1
(cid:18) 3
(cid:19)−1
(cid:19)
1
1
0
2
.
0
3
1
0
1
3
0
0
2
(cid:19)
(cid:18) 4
=
1
1
3
, Z2).
(cid:19)
0
2
, Z2) = lim→ (A, Z2)
72
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
• Inflation factor: λ = 2
• Proto-edges(=stable edges) (4): a, b, c, d a = 1, b = 1, c = 1, d = 1
• substitution: ω(a) = ab, ω(b) = ac, ω(c) = db, ω(d) = dc
• Stable vertices (8): a.b, a.c, b.a, b.d, c.a, c.d, d.b, d.c
• K0(U) = H ST0
S(T) = K0(S) = Z ⊕ Z[ 1
2]3
(T) = H0
We compute the substitution-homotopy matrix WV in a similar way as we did for
the Fibonacci tiling. We then use Proposition A.10 and Proposition A.9 to remove
the eigenvalue 0. We then extract the eigenvalue -1 with Proposition A.14 and get
lim→ (WV , ker δ0) = Z ⊕ lim→ (W 0
V , Z3),
0
2
1
. It is clear that A−nZ2 is a subset of Z[ 1
2]2 because every
where A =
entry is of that form. The other inclusion is by the following argument: The second
coordinate A−n.(0, k) is obviously all the dyadic numbers for all integers k and all
n. The first coordinate A−n.(k, 0) gives all the dyadics in the first entry and some
numbers in the second entry, which we know are dyadics, so we just subtract that
and in this way we get all dyadic numbers in the first entry. Hence by Proposition
A.2, lim→ (A, Z2) = Z[ 1
S(T) = lim→ (WV , ker δ0) = Z ⊕ Z[ 1
2]3.
2]2, and thus H0
The K-theory groups of the above examples are summarized in Table 1 and
Table 2. For computing the K-theory of the asymptotic C∗-algebra A we used the
formula given in Proposition A.15. We are the first ones to compute the stable
and unstable K-theories of the "Pathologic" tiling. The unstable K-theories of
the rest of the above examples are already well-known, and they agree with our
computations.
7.2. Stable cells - Collared cells relationship. Recall that for tilings of dimen-
sion 1, we can always use the homotopy that maps the leftmost edge of ω(e) to
e (the orientation of the edges is from left to right). In such case, we can write
the substitution-homotopy matrices WV , WE in terms of the collared-substitution
matrices and the forgetful and inclusion maps defined as follows
FE(ee0e00) := e0.e00
FV (e.e0) := e0
iE(e0.e00) := ee0e00
iV (e0) := e.e0,
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)73
Table 1. K0-groups for tilings of the line.
(T)
Z2
Tiling
Fibonnaci
Morse
Nonreducible 4 letter
Period doubling
Rauzi (tribonacci)
Rudin Shapiro
OneFifth
OneSixth
Pathologic Z ⊕ lim→
K0(U) = H ST0
Z ⊕ Z[ 1
2]
Z ⊕ Z[ 1
2]
Z3
Z ⊕ Z[ 1
2]3
Z[ 1
5]2
Z[ 1
6]2
1
6
(cid:18) 3
(cid:19)
Z5
1
(cid:36) Z ⊕ Z[ 1
Z2
Z5
K0(S) = H0
S(T)
Z ⊕ Z[ 1
2]
Z ⊕ Z[ 1
2]
Z3
Z ⊕ Z[ 1
2]3
Z[ 1
5]2
Z[ 1
6]2
17]2 Z ⊕ lim→
1
(cid:18) 3
1
6
(cid:19)
Z5
K0(A)
Z2 ⊕ Z[ 1
2]3
Z26
Z2 ⊕ Z[ 1
2]3
Z10
Z2 ⊕ Z[ 1
2]15
Z ⊕ Z[ 1
5]4
Z ⊕ Z[ 1
6]4
Note: some authors prefer to write for example Z[1/4] instead of its reduced form Z[1/2].
Table 2. K1-groups for tilings of the line.
Tiling K1(U) = H ST1
Fibonnaci
Morse
nonreducible 4 letter
Period doubling
Rauzi (tribonacci)
Rudin Shapiro
OneFifth
OneSixth
Pathologic
(T) K1(S) = H1
S(T)
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z4
K1(A)
Z2 ⊕ Z[ 1
2]2
Z10
Z2 ⊕ Z[ 1
2]2
Z2 ⊕ Z[ 1
2]6
Z[ 1
5]6
Z[ 1
6]4
Z6
Z
Z
Z
Z
Z
Z
Z
Z
Z
where ee0e00 is a collared edge, e0 is a stable edge, and e.e0 is a stable or collared
vertex. Then it is easy to see that
Theorem 7.5 (Stable cells - Collared cells relationship). For any tiling T ∈ Ω of
dimension 1 with the homotopy defined above, the following relations hold
• WE = FV ◦ ωV ◦ iV
• WV = FE ◦ ωE ◦ iE
• δ0 = FV ◦ ∂1 ◦ iE
• FV ◦ ∂1 = δ0 ◦ FE
• FV ◦ ωV = WE ◦ FV
• FE ◦ ωE = WV ◦ FE.
The last relation in the above theorem gives rise to the following commutative
diagram
(7.3)
coker ∂t1
ωt
E /
/ coker ∂t1
ωt
E
/ ···
F t
E
coker δt0
W t
V /
F t
E
coker δt0
W t
V
··· .
/
/
O
O
/
/
O
O
74
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
All the above one dimensional tilings satisfied
H ST0
(T)
lim(coker δt0, W t
K0(U)
V ) = lim(coker ∂t1, ωt
K0(S)
E) = lim(ker δ0, WV ).
The first equality holds in general for any 1-dimensional tiling by Theorem 1.4. See
also Corollary 5.27. The second equality holds at least for the above examples, but
we don't know if it is valid in general. In all these 1-dimensional examples, except
the tiling called OneSixth, the first equality was verified using diagram (7.3) after
removing some of the 0-eigenvalues; the second equality was verified empirically
with the following 4 steps:
(1) Let Au : Zr → Zr be the matrix for ωt
E : coker ∂t1 → coker ∂t1 with some of
(2) Let As : Zr → Zr be the matrix for WV : ker δ0 → ker δ0 with some of the
(3) Let the following be C-diagonalization of the matrices Au, As:
the zero eigenvalues removed.
zero eigenvalues removed.
PuDP −1
u = Au, PsDP −1
s = As.
(4) If PuP −1
s
is a Z-invertible integer matrix then Au and As are Z-similar.
s
in Step 4 always yielded a Z-invertible matrix for the above
The matrix PuP −1
examples except for the tiling called OneSixth.
7.3. Two dimensional tilings. [Block/Rectangular tilings.] Let T be a tiling
of the plane, and let f1, . . . , fN ∈ T be the prototiles(=protofaces).
If all the
prototiles are rectangles or blocks (cf. [1, Section 8]) then one can always generalize
the procedure that was used for the one-dimensional tilings. Namely, for these
"block" tilings one can always put the homotopy that homotopes the left-most
bottom-most rectangle in 1
λ ω(fi) to fi. Moreover, we put the counterclockwise
orientation on all the rectangles of the tiling, on its horizontal edges we put the
orientation from left to right, and on its vertical edges we put the orientation from
bottom to top.
Figure 10. Tri-square tiling.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)75
Example 7.6 (Tri-square tiling). (cf. [45, Section 5.4,p.99]). Let T be the tri-
square tiling with protofaces a, b, c ∈ T and substitution rule given as follows:
a
ω
b
c
c
b
b
ω
b
a
a
b
c
ω
a
c
c
a
The length of the horizontal edge and of the vertical edge is 1. The inflation factor
is λ = 2. We illustrate the homotopy hs, 0 ≤ s ≤ 1, on the vertices in the following
figure:
hs
•
•
•
hs
•
•
•
hs
hs
hs
•
•
•
• stable faces (3): a, b, c
• vertical stable edges (7): aa, ab, ac, ba, bc, ca, cb
• horizontal stable edges (7):
a , c
a , b
c , b
b , a
b , c
a , a
c
• stable vertices (21): a
a
a , b
a
a , c
b
a , a
a
a , a
c
a , b
c
b , b
c
a
c
a
a
a , b
a , a
a , a
b , b
a
c , a
b
c , a
b
c , c
a
c , c
b
a
b
c
b
a
a
b
b
c
a
a
a
a
a
a
b , a
a
c
b , a
a
b
c , c
a
b
c , c
b
a
a , c
b
b
a ,
a
c
c
c
We list the stable edges in the above order starting with the vertical stable edges.
We denote them as se1, . . . , se14. The stable vertices are also listed in the above
order and we denote them as sv1, . . . , sv21. The exponential map δ0 : ZsV → ZsE
is given by the matrix
76
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Figure 11. WV (sv2) = sv15 + sv16 + 2 sv21
For instance δ0(sv3) = se1 − se7 − se9 + se10. The index map δ1 : ZsE → ZsF is
given by the matrix
0
δ1 =
0 −1
1 −1
1
0 −1
1
1
0
0 −1
0 −1
0
0
0 −1
1
1
0 −1 −1
0
0
1
0
1
0
0
0 −1 −1
1
1 −1 −1
1
1
0
0
For controlling computational errors, it is always good to check that δ1 ◦ δ0 = 0.
The substitution-homotopy map WV : ZsV → ZsV is given by the matrix
For instance, WV (sv2) = sv15 + sv16 +2 sv21, and it is illustrated in Figure 11. The
substitution-homotopy map WE : ZsE → ZsE is given by the matrix
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)77
b
c
c
b
•
•
b
a
•
•
a
b
Figure 12. WE(se2) = se4 + se7
For instance, WE(se2) = se4 + se7, and it is illustrated in Figure 12.
The substitution-homotopy map WF : ZsF → ZsF is given by the matrix
0
0
1
1
0
0
0
0
1
WF =
For controlling computational errors it is always good to check that the diagram
(1.2) commutes, i.e. that WE◦δ0 = δ0◦WV and WF ◦δ1 = δ1◦WE. The computation
of the direct limits is done similarly as the ones done for the 1-dimensional tilings,
so here we skip most of the steps. By Proposition A.10 we replace ker δ0 with Z11.
By Proposition A.9 we remove the zero-eigenvalues of the matrix. By Proposition
A.14 we extract the ±1-eigenvalues and get
Note that
−1
0
0
2 −1
1
0
0 −1
S(T) = lim→ (WV , ker δ0) = Z4 ⊕ lim→ (
H0
8 −18
2 −6
.
18 −38
18 −38 16
8 −18
2 −6
8
4
0
0
16
8
4
.
−1
, Z3) = lim→ (
4
lim→ (
is a lower triangular matrix with eigenvalues 4,−2, 2 and, by Proposition A.4, we
get that
8 −18
2 −6
18 −38
−1
2 −1
1
0
0 −1
0
−8 −2
2
2
0
0
2
16
8
4
=
, Z3).
4
, Z3).
0
−8 −2
2
2
0
0
2
78
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
By the same argument as the one used for the lower triangular matrix in Example
7.4(Rudin-Shapiro) we conclude that
4
, Z3) = Z4 ⊕ Z[1
0
−8 −2
2
2
E, coker B) for some matrices W 0
Im δ0 ) = lim→ (W 0
0
0
2
2]3.
E, B.
S(T) = lim→ (WV , ker δ0) = Z4 ⊕ lim→ (
H0
By Proposition A.12, lim→ (WE, ker δ1
Then by Proposition A.11 we get rid of the coker and get
K1(S) = H1
S(T) = lim→ (WE,
(cid:18) 2
ker δ1
Im δ0 ) = (
0
(cid:19)
0
2
, Z2) = Z[1
2]2.
We are the first ones to compute the stable and unstable K-theories of the Tri-
square tiling. The computation of K(U) for this example and for the rest of the
examples was done in Mathematica. Since the computations are similar to the ones
for K(S), we do not explain them in this paper.
The Table tiling and its unstable K-theory was communicated by Franz Gahler.
However, we have refined its substitution rule in order to be able to apply our
formulas.
Example 7.7 (Table Tiling). Let T be the Table tiling with proto-faces a, b, c, d ∈
T and substitution rule given below:
Figure 13. Table tiling.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)79
a
b
ω
ω
d
c
a
a
b
b
d
c
d
c
ω
ω
a
d
c
a
b
d
c
b
The length of each of the edges of the proto-faces is 1, and the inflation factor is
λ = 2. The homotopy is the same as the one of the previous example.
• stable faces (4): a, b, c, d
• vertical stable edges (10): ab, ba, bc, bd, ca, cc, cd, da, dc, dd
• horizontal stable edges (10): a
d , c
b , d
d , b
c , a
b , c
b , b
a , a
a , c
a , b
• stable vertices (24): b
a , b
c
a , c
b
b , a
c
b , c
c
b , b
b
b , a
a
b , c
a
d
c , c
a
a
a
a
a
b
b
b
a
d , d
d
d
b
c
d , c
a
a , a
a , b
c
a , b
c , c
c , b
a
d , a
b
a
a , d
c
a , d
d
c , d
c
a
d , a
b
d
c , c
b
d
d
d
c
d
d
d
d
d
c
b
b
d , b
b
c
d ,
c
c
d
d
We list the stable edges in the above order starting with the vertical stable edges.
We denote them as se1, . . . , se20. The stable vertices are also listed in the above
order and we denote them as sv1, . . . sv24. The exponential map δ0 : ZsV → ZsE is
given by the matrix
The index map δ1 : ZsE → ZsF is given by the matrix
80
The substitution-homotopy map WV : ZsV → ZsV is given by the matrix
M. RAMIREZ-SOLANO
D. GONC¸ ALVES
The substitution-homotopy map WE : ZsE → ZsE is given by the matrix
The substitution-homotopy map WF : ZsF → ZsF is given by the matrix
.
0
1
0
0
1
0
0
0
0
0
0
1
0
0
1
0
WF =
The computation of the direct limits is done similar to the one from the previous
example. By Proposition A.10 we replace ker δ0 with Z9. By Proposition A.9
we remove the zero-eigenvalue of the matrix. By Proposition A.14 we extract the
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)81
±1-eigenvalues and get
0
2
0
0
0
, Z5).
8
2
18
0 −6 −10
0 −10 −16
0
0
−8
8
16
14 −10
4
0
8
2
S(T) = lim→ (WV , ker δ0) = Z3 ⊕ lim→ (
H0
Note that
is a lower triangular matrix with eigenvalues 4, -2, -2, 2, 2, and by Proposition A.4
we get that
2
0
18
2 0 −6 −10
0 −10 −16
0
0
0
0 0
0
2
0
0
2 −2
0
0
2
0
0
0
0 −6 −2
0
0 −2
0
0
−8
8
16
14 −10
4
0
lim→ (
0
0
0
0
4
8
2
8
By the same argument as the one used for the lower triangular matrix in Example
7.4(Rudin-Shapiro) we conclude that
, Z5) = lim→ (
, Z5).
, Z5) = Z3⊕Z[1
, Z2 ⊕ Z2) = Z2 ⊕ Z[1
0
0
0
0
4
0
2
0
0
2 −2
0
0
0
2
0
0
0 −6 −2
0
0 −2
0
0
E, coker B) for some matrices W 0
2]5.
E, B.
S(T) = lim→ (WV , ker δ0) = Z3⊕lim→ (
H0
By Proposition A.12, lim→ (WE, ker δ1
Then by Proposition A.11 we get rid of the coker and get
Im δ0 ) = lim→ (W 0
K1(S) = H1
S(T) = lim→ (WE,
ker δ1
Im δ0 ) = lim→ (
0
0
2
7.4. Two dimensional tilings. [more general tilings.]
Example 7.8 (Half-hex tiling). Let T be the Half-hex tiling with proto-faces
a, b, c, d, e, f ∈ T and substitution rule
2]2.
0
0
2
0
1 −1
82
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Figure 14. Half-hex tiling.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)83
The inflation factor is λ = 2. There are 24 stable edges and 20 stable vertices.
84
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
The exponential map δ0 : ZsV → ZsE is given by the matrix
The index map δ1 : ZsE → ZsF is given by the matrix
We illustrate the homotopy hs, 0 ≤ s ≤ 1, on the vertices in the following figure:
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)85
The substitution-homotopy map WV : ZsV → ZsV is given by the matrix
The substitution-homotopy map WE : ZsE → ZsE is given by the matrix
86
The substitution-homotopy map WF : ZsF → ZsF is given by the matrix
M. RAMIREZ-SOLANO
D. GONC¸ ALVES
1
2
1
2
, Z3).
, Z3) = Z2 ⊕ lim→ (4, Z) = Z2 ⊕ Z[1
2].
2
By Proposition A.10,
1
1
2
By Proposition A.14 we extract the 1-eigenvalues and get
H0
S(T) = lim→ (WV , ker δ0) = lim→ (
1
1
H0
S(T) = lim→ (
1
1
2
Im δ0 ) = lim→ (W 0
By Proposition A.12, lim→ (WE, ker δ1
Then by Proposition A.11 we get rid of the coker and get
1
2
1
1
1
K1(S) = H1
S(T) = lim→ (WE,
ker δ1
Im δ0 ) = lim→ (
E, coker B) for some matrices W 0
E, B.
0 −2 −2
, Z2 ⊕ Z2) = Z[1
2]2.
0
0
2
0
0
2
Example 7.9 (Chair tiling). Let T be the Chair tiling with proto-faces a, b, c, d ∈ T
and substitution rule given by:
Figure 15. Chair tiling.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)87
The inflation factor is λ = 2. There are 44 stable edges and 47 stable vertices
88
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)89
We remark that we will write big matrices in the format:
{{row1},{row2}, . . .}.
The exponential map δ0 : ZsV → ZsE is given by the matrix:
{{1, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, -1, 0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, -1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {-1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, -1, 0,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, -1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, -1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0}, {0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0,
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0,
90
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0,
0, 0, -1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, 0, 0, 0, 0, 0, 0,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 1, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, -1, -1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0,
0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, -1, 0, 0, 0, 1, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, -1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, -1, -1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1, 0, 0, 1, 0, 0, 0, 0, 0, 0, -1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, -1, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 1, 0, -1, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 1, 0, 0, 0,
0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 1, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, -1, 0, 0, 0}}
The index map δ1 : ZsE → ZsF is given by the matrix:
{{1, 0, 0, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, 0, 0, -1, 0, 0, 1, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {-1, 0, 0, 0,
-1, -1, 0, 0, 1, 1, -1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, -1, 1, -1, -1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, -1, -1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-1, 0, 0, 0, 0, 1, 0, 1, -1, 1, 0, 0, -1, -1, 1, 1, -1, 0, 0, -1, 1, 0, -1, 1, 0, -1, -1, 1, 1}, {0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 1, -1, -1, 0, 0, 0, 0, 0, -1, -1, 0, 0, 0,
0, 0, 0, 0, 0, -1, -1, 1, -1, 1, 1, -1, 0, 1, 0, 0, 1, 1, 0, 0}}
We illustrate the homotopy hs, 0 ≤ s ≤ 1, on the vertices in the following figure:
The substitution-homotopy map WV : ZsV → ZsV is given by the matrix:
{{1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 2, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0}, {1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {1, 0, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0,
0}, {0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 1, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0,
0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0,
0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1}, {0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0,
0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0}}
The substitution-homotopy map WE : ZsE → ZsE is given by the matrix:
{{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 2, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 2, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1,
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0}, {0, 0,
0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)91
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0,
0, 0, 2, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0}, {0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1}}
The substitution-homotopy map WF : ZsF → ZsF is given by the matrix
By Proposition A.10,
0
0
0
0
0
0
0
0
0 −1
1
4
0
0
1 −1
1
2
0
0
0
1
0
0 −1
1
0
0
0
0
0
0
1
0
1 −1
0
0
0
0
1 −1 −1
1
0
0
0
2
0
0
1 −1
1
0
0
1
1
0
0
, Z8).
0
0
2 −1
.
1
2
2
, Z3).
=
2
, Z3) = Z[1
, Z2
0
0
2
0 −1
1
1
0
0
4
0 −1
0 −1
0
0
0
4
0
0 −1
E, coker B) for some matrices W 0
2]3.
0
0
2
2 ⊕ Z2) = Z[1
2]2
,
E, B.
H0
S(T) = lim→ (WV , ker δ0) = lim→ (
0
0
1
0
0
By Proposition A.9 we remove the zero-eigenvalues and get
1
5 −3
1
1
H0
S(T) = lim→ (WV , ker δ0) = lim→ (
Since 1
0 −1
1
1
0 −1
0
0
−1
2 −1
.
0
0
1
5 −3
1
1
S(T) = lim→ (WV , ker δ0) = lim→ (
H0
Im δ0 ) = lim→ (W 0
By Proposition A.12, lim→ (WE, ker δ1
Then by Proposition A.11 we get rid of the coker and get
0
0
0
2
ker δ1
Im δ0 ) = lim→ (
S(T) = lim→ (WE,
K1(S) = H1
0
2
0
0
0
0
2
0
0
0
0
because 2 mod 2 is zero.
Example 7.10 (Octagonal tiling). (cf. [24]). Let T be the Octagonal tiling with
proto-faces a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t,∈ T and substitution rule
given by:
92
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Figure 16. Octagonal tiling.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)93
The tiles a, i form a unit square, and the angles in the tile q are 45◦, 135◦. The
2. There are 56 stable edges and 49 stable vertices.
inflation factor is λ = 1 +
√
94
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)95
96
The exponential map δ0 : ZsV → ZsE is given by the matrix:
M. RAMIREZ-SOLANO
D. GONC¸ ALVES
{{-1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, -1, -1},
{0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, -1, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 1, 1}, {0, 0, -1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0}, {0, -1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1,
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0,
0, 0, 0}, {0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, -1, 0, 0, -1, 0, 0, 0, 0, 0, 0}, {0, 0, 0,
0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {-1, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1, 0, -1, 0, 0, -1, -1, 0, -1}, {1, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, -1, 0, 0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, 0,
-1, 0, 0, -1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, -1, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1,
0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, -1, 0, 0, 0, -1, 0, 1, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, -1, 0, -1, 0, 0, 0}, {0, 0, 0, 0, -1, 0, 0, 0, -1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, -1, 0, 0, -1, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, -1, 0, -1, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0}, {-1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, -1, -1, 0, -1, 0, 0, 0, -1, 0, -1}, {0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {-1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, -1, 0, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, -1, 0, -1, 0, -1}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0,
0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, -1, 0, 0, 0, 0, 0, 0, -1, 0, 0,
0, 0, -1, 0, -1, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0,
-1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0,
-1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, -1, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, -1, 0, -1, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, -1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0,
0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0}, {0, -1, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, -1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, -1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, -1, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, -1,
0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1, 0, 0, 0, -1, -1, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 1, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
The index map δ1 : ZsE → ZsF is given by the matrix:
{{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, -1, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0,
0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, -1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0},
{0, 1, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0},
{0, 0, 1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, -1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, -1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, -1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0}, {0,
0, 0, 0, 0, 0, 0, 1, 0, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0}, {0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0}, {0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0,
0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0,
1, 0, 1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, -1, -1, 0, 0, 0, 0, 0, 1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {-1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 0, 1, 1, 0, 0, 0, 0, -1, 0, -1, 0, 0, 0, 0}, {1, -1,
-1, 1, 0, 0, 0, 0, 0, 0, 1, -1, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0,
0, -1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 1, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 1, 1, 0, -1, 0, 0, 0, 0, 0, 0, 0}}
We illustrate the homotopy hs, 0 ≤ s ≤ 1, on the vertices in the following figure:
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)97
The substitution-homotopy map WV : ZsV → ZsV is given by the matrix:
{{1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {1, 0, 0, 0, 1, 1, 1,
0, 1, 0, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {1, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 0,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1}, {1, 0, 0, 0, 1, 1, 1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1,
0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 1, 1, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1,
1, 1}, {1, 0, 1, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {0, 0, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 1, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {1, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1,
0, 0, 1, 1, 0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 1, 0}, {1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1,
1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1}, {1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1}, {1, 0, 0,
0, 1, 1, 0, 0, 1, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0,
0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 0, 1, 1, 0, 0, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}, {1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0,
0, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 1}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0,
0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 0, 0, 0, 0, 1, 0, 0,
0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {1, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 1,
0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1}, {0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
The substitution-homotopy map WE : ZsE → ZsE is given by the matrix:
{{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 1, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0,
0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, -1, 0,
0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0}, {1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0}, {1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0,
0}, {0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1, 0, -1, 0, 0, 0, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, -1, 0, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 0,
1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 1, 0, 0,
0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0}, {0, 0, 1, 0, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0}, {0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0}, {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
1, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0}, {0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0}, {0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0,
0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {-1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0,
98
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0}, {0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-1, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}}
The substitution-homotopy map WF : ZsF → ZsF is given by the matrix
By Proposition A.10 we replace ker δ0 with Z17. By Proposition A.9 we remove the
zero-eigenvalues of the corresponding matrix and get
H0
S(T) = lim→ (WV , ker δ0)
= Z3 ⊕ lim→ (
= Z9
−19 −63
8
10 −24 −74
−2
5
2
−15 −48
3
21
6
0
−2
7
1
7
1
0
1
1
7
−4
−1
0
1
12 −14 −7
15 −15 −6
−1
0
10 −10 −5
−4
3
−2
−1
−1
2
−1
2
−3
0
5
−1
2
2
−3
12 −1
−1
8
0
0
−1
7
−5
1
−1
5
−4
1
−4
1
−2
8
7
4
0
3
−2
1
−2
−2
3
, Z9)
because the determinant of the 9×9 matrix is -1. By Proposition A.12, lim→ (WE, ker δ1
lim→ (W 0
E, B. Then by Proposition A.11 we get rid
of the coker and get
E, coker B) for some matrices W 0
Im δ0 ) =
0 −1
1
0
0
0
1
3
2 −1 −2
2
0 −1
2
1 −1
3
1
1
0
1
0
1
1
, Z5) = Z5
K1(S) = H1
S(T) = lim→ (WE,
ker δ1
Im δ0 ) = lim→ (
because the determinant of the 5 × 5 matrix is -1.
The K-theory groups of the above examples are summarized in Table 3 and
Table 4. We are the first ones to compute the stable and unstable K-theories of
the tri-square tiling. The unstable K-theories of the rest of the above examples are
already well-known, and they agree with our computations.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)99
Table 3. K0-groups for tilings of the plane.
Tiling K0(U) = Z ⊕ H ST0
Octagonal
Chair
Tri-square
Half-hex
Table
Z10
Z ⊕ Z[ 1
2]3
Z5 ⊕ Z[ 1
2]3
Z3 ⊕ Z[ 1
2]
Z4 ⊕ Z[ 1
2]5 ⊕ Z2
(T) K0(S)/Z = H0
S(T) K0(A)
Z200
Z9
Z[ 1
2]3
Z4 ⊕ Z[ 1
2]3
Z2 ⊕ Z[ 1
2]
Z3 ⊕ Z[ 1
2]5
--
Table 4. K1-groups for tilings of the plane.
Tiling K1(U) = H ST1
(T) K1(S) = H1
S(T) K1(A)
Z200
Octagonal
Chair
Tri-square
Half-hex
Table
Z5
Z[ 1
2]2
Z[ 1
2]2
Z[ 1
2]2
Z[ 1
2]2
Z5
Z[ 1
2]2
Z[ 1
2]2
Z[ 1
2]2
2]2 ⊕ Z2
Z[ 1
--
Acknowledgments. The authors would like to thank Jean Renault, Erik Chris-
tensen, Rohit Dilip Holkar, Uffe Haagerup, and Toke Nørgard-Solano for the fruitful
discussions and valuable suggestions. The first author was partially supported by
CNPq. The second author was supported by the Villum Foundation under the
project "Local and global structures of groups and their algebras" at University of
Southern Denmark, and by CNPq grant at Universidade Federal de Santa Catarina.
Appendix A. Direct limits of abelian groups
In this section we give some tools to compute direct limits of integer matrices.
Definition A.1 (Direct limit notation). Suppose that A : X → X is an endomor-
phism of some abelian group X (i.e. a Z-module). Then we introduce the notation
for the direct limit
lim→ (A, X) := lim→ X A /
/ X A /
/ X A /
/ .
Note that we write the homomorphism A first, which we do in order to emphasize
it.
Proposition A.2. Let A ∈ Mn(Z) be an n× n integer matrix such that det A 6= 0.
Then
S∞
k=0 A−kZn = lim→
Zn
A /
/ Zn
A /
/ Zn A /
/ .
Note that
∞[
k=0
A−kZn ⊂ Z[
1
det A
]n.
D. GONC¸ ALVES
100
Proof. Since AZn ⊂ Zn we get that Zn = A−1(AZn) ⊂ A−1Zn. Thus the following
diagram is well defined
M. RAMIREZ-SOLANO
Zn
I∼=
Zn
A
Zn
∼=
A−1
/ A−1Zn
A
Zn
∼=
A−2
/ A−2Zn
A
Zn
∼=
A−3
/ A−3Zn
A
A
/ .
Moreover, since the diagram commutes, and the vertical maps are isomorphisms,
and the maps on the bottom row are the inclusion maps the statement of the
(cid:3)
proposition holds.
Corollary A.3. Let A ∈ Mn(Z) be an integer matrix such that det A = ±1. Then
lim→ (A, Zn) = Zn.
Proof. If det A = ±1 then A−1Zn = Zn. The statement of the corollary then
(cid:3)
follows by the proposition.
Proposition A.4. Let A, G ∈ Mn(Z) be integer matrices, and suppose that det G =
±1. Then
lim→ (A, Zn) = lim→ (GAG−1, Zn).
Proof. Since the diagram
A /
A /
A /
Zn
G∼=
Zn
Zn
G∼=
/ Zn
Zn
G∼=
/ Zn
Zn A /
G∼=
/ Zn
GAG−1
GAG−1
GAG−1
GAG−1
commutes, and its vertical maps are isomorphisms, the statement of the proposition
(cid:3)
holds.
Proposition A.5. Let A ∈ Mn(Z) be an integer matrix. Then
lim→ (A, Zn) = lim→ (A, AZn).
Proof. Since the diagram
Zn
A
AZn
A
ι
A
Zn
A
AZn
A
ι
A
Zn
A
AZn
A
ι
A
Zn A
A
AZn
ι
A
commutes and we can go up and down in the diagram, the statement of the propo-
(cid:3)
sition holds.
Proposition A.6. Let A ∈ Mn(Z) be an integer matrix. Then
lim→ (A, Zn) = lim→ (A2, Zn).
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
-
;
;
/
/
-
;
;
/
/
-
;
;
/
/
.
>
>
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)101
Proof. Since the diagram
Zn
A /
Zn
A /
Zn
A /
Zn A /
I
Zn
A
A2
A
Zn
I
I
I
Zn
A
A2
A
Zn
I
I
commutes, and we can go up and down in the diagram, the statement of the propo-
(cid:3)
sition holds.
Proposition A.7. Let X, Y be Z-modules. Let A : X → X be an endomorphism,
and let X and Y be Z-isomorphic with isomorphisms p : X → Y , q : Y → X. Then
lim→ (A, X) = lim→ (pAq, Y ).
Proof. This is because the following diagram commutes and we can go up and down
the diagram.
X A /
X A /
X A /
X A /
p
q
p
q
p
q
p
q
Y
Y
Y
Y
pAq
pAq
pAq
pAq
(cid:3)
Proposition A.8 (direct sum). Let A ∈ Mm(Z), B ∈ Mn(Z) be integer matrices.
Then
(cid:18) A 0
0 B
(cid:19)
, Zm+n) = lim→ (A, Zm) ⊕ lim→ (B, Zn).
L∞
i=1 Zm
i
where ei = (0, . . . , 0, 1, 0, . . .) has 1 at the i-th position, and Zi := Z. For instance
if
span{xiei − (Axi)ei+1 i ∈ N} = {[zei] i ∈ N, z ∈ Zm},
x = x1ei1 + x2ei2 + x3ei3 ∈M
Zm
i
lim→ (
Proof. Recall that
lim→ (A, Zm
i ) =
i∈N
with i1 < i2 < i3 then
We claim thatL
x ∼ (Ai3−iix1 + Ai3−i2x2 + x3)ei3 .
i∈N(Zm
i,a ⊕ Zn
i,b)
h(x, y)ei − (Ax, By)ei+1i =
L
i∈N Zm
i − (Ax)e
(a)
=
xe
i,a
(a)
i+1
hxe
!
i,a) ⊕ (L
(L
i∈N Zm
L
i − (Ax)e
(a)
(a)
i+1, ye
i∈N Zn
⊕
i − (By)e
(b)
ye
j,b
.
(b)
j+1
i∈N Zn
i,b)
!
j − (By)e
(b)
j+1i
(b)
The second equality is clear since the relations do not intertwine. The proof of
the first equality is as follows. Note that the left numerator of the equality can be
identified with the right numerator. Hence it suffices to show that the denominators
are equal. Proof that the left denominator is contained in the right denominator:
/
/
/
/
/
/
=
=
/
/
=
=
/
/
=
=
/
/
?
?
/
/
/
/
/
/
O
O
/
/
O
O
/
/
O
O
/
/
O
O
D. GONC¸ ALVES
102
Let (x, y)ei − (Ax, By)ei+1 be given. This element is in the right denominator
because it is the sum of two generators of the right denominator with i = j, namely
M. RAMIREZ-SOLANO
(xe
i − (Ax)e
(a)
(a)
i+1) + (ye
i − (By)e
(b)
i+1) = (x, y)ei − (Ax, By)ei+1.
(b)
Proof that the left denominator contains the right denominator: Let xe
be given. This element is in the left denominator since
i −(Ax)e
(a)
(a)
i+1
(x, 0)ei − (Ax, 0)ei+1 = xe
i − (Ax)e
(a)
(a)
i+1.
j − (By)e(b) is in the
(b)
By the same argument as before, we see that the element ye
(cid:3)
left denominator.
The next proposition tells us how to remove the zero-eigenvalues of a matrix.
Proposition A.9 (Im A ∼= Zr). Let A ∈ Mm×n(Z) be an m × n integer matrix
and let D = P AQ be its Smith normal form. In particular, all the nonzero entries
in the diagonal of
D = diag(d1, d2, ..., dr, 0, ...0)
are given first. Moreover, di divides di+1, i = 1, . . . , r − 1.
first r rows of the identity matrix In.
Define π : Zn → Zr to be the projection onto the first r coordinates i.e. π is the
Let q : Zr → AZn be defined by
q := P −1Dπt,
where πt is the transpose of π, and let p : AZn → Zr be defined by
where the pseudo-inverse D† is the diagonal matrix
p := πD†P
D† = diag( 1
d1
,
1
d2
, ...,
1
dr
, 0, ..., 0).
Then AZn and Zr are Z-isomorphic with isomorphisms p, q. In particular, (for
computer testing)
pq = Ir,
where Ir is the r × r identity matrix.
Proof. Observe that p is a rational matrix, not necessarily an integer matrix. We
have
qpA = A,
qZr = P −1DπtZr = P −1(Dπtπ)Zn = P −1DZn = A(QZn) = AZn.
Thus q is surjective. The map pq is the identity on Zr because
pq = πD†P P −1Dπt = π(D†Dπt) = ππt = Ir.
We now show that the map qp is the identity on AZn. Let x ∈ AZn. Since q is
surjective x = q(y) for some y ∈ Zr. Then
Thus AZn and Zr are Z-isomorphic with isomorphisms p, q.
(cid:3)
qp(x) = qpq(y) = q(y) = x.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)103
Proposition A.10 (ker A ∼= Zn−r). Let A ∈ Mm×n(Z) be an m× n integer matrix
and let D = P AQ be its Smith normal form. In particular, all the nonzero entries
in the diagonal of
D = diag(d1, d2, ..., dr, 0, ...0)
are given first. Moreover, di divides di+1, i = 1, . . . , r−1. If r = n then ker A = {0},
so assume r < n. Define π : Zn → Zn−r to be the projection onto the last n − r
coordinates i.e. π is the last n − r rows of the identity square matrix In. Let
q : Zn−r → ker A be defined by
where πt is the transpose of π, and let p : ker A → Zn−r be defined by
q := Q ◦ πt,
p := π ◦ Q−1.
Then ker A and Zn−r are Z-isomorphic, with isomorphisms p, q. In particular, (for
computer testing)
pq = In−r,
Aqp = Aq = 0.
Proof.
ker A = {x ∈ Zn Ax = 0}
= {x ∈ Zn P −1DQ−1x = 0}
= {x ∈ Zn Dy = 0, y = Q−1x}
= {Qy λ1y1e1 + ··· + λryrer = 0}
= {Qy y1 = 0, . . . , yr = 0}
= span(Qer+1, . . . , Qen)
= QπtZn−r
= qZn−r.
Thus q is surjective. The map pq is the identity on Zn−r because
pq = πQ−1Qπt = ππt = In−r.
We now show that the map qp is the identity on ker A. Let x ∈ ker A. Since q is
surjective, x = q(y) for some y ∈ Zn−r. Then
qp(x) = qpq(y) = q(y) = x.
(cid:3)
Thus ker A and Zn−r are Z-isomorphic, with isomorphisms p, q.
Proposition A.11 (coker A ∼= Z
(for m ≤ n)). Let A ∈ Mm×n(Z) be
an m × n integer matrix such that m ≤ n, and let D = P AQ be its Smith normal
form. In particular, all the nonzero entries in the diagonal of
ds+1 ⊕···⊕ Z
dm
D = diag(±1, . . . ,±1, ds+1, ds+2, ..., dr, 0, ...0)
are given first. Moreover, di divides di+1, i = 1, . . . , r − 1, and there are s ≥ 0
ones. If s = m then coker A = {0}. So assume s < m. Define π : Zm → Zm−s to
be the projection onto the last m − s coordinates i.e. π is the last m − s rows of the
identity square matrix Im. Let q : Zm−s → Zm be defined by
where πt is the transpose of π, and let p : Zm → Zm−s be defined by
q := P −1 ◦ πt,
p := π ◦ P.
104
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
ds+1 ⊕ ··· ⊕ Z
Z
Then coker A and
··· ⊕ Z
In particular, (for computer testing)
→ coker A and ¯p : coker A → Z
dm
dm
ds+1 ⊕ ··· ⊕ Z
dm
are Z-isomorphic, with isomorphisms ¯q :
ds+1 ⊕
Z
induced by q, p, respectively.
pq = Im−s,
qpA = 0.
Proof. Consider the following diagram whose rows are short exact sequences of
abelian groups.
0
0
0
/ Im A
/ Im D
/ Zm
P∼=
/ Zm
π
Zm
Im A
¯P ∼=
Zm
Im D
¯π
g
/ Zds+1 ⊕ ··· ⊕ Zdm
/ Zm−s
h /
Zds+1 ⊕ ··· ⊕ Z
Z
Zdm
0
0
/ 0
Since h is a quotient map and π is a projection, they are both surjective and hence
hπ is surjective. Since
ker(hπ) = π−1(ker h) = π−1(Zds+1 ⊕···⊕ Zdm) = Zs ⊕ Zds+1 ⊕···⊕ Zdm = Im D
we get by the first isomorphism theorem of groups that
Zm
Im D
=
Zm
ker hπ
∼=
Z
Zds+1
⊕ ··· ⊕ Z
Zdm
,
where the isomorphism is induced by π and it is denoted by ¯π. Now since ππt =
Im−s and since
ker(gπt) = (πt)−1(ker g) = (πt)−1(Im D) = (πt)−1(ker hπ) = ker(hππt) = ker h
we get that πt induces an isomorphism πt :
πt = ¯π−1.
Similarly, since gP is surjective and
Zds+1 ⊕ ··· ⊕ Z
Z
Zdm
∼−→ Zm
Im D and that
ker gP = P −1(ker g) = P −1(Im D) = P −1DZn = AQZn = AZn = Im A,
we get that ¯P exists and that it is an isomorphism with inverse P −1. Thus q :=
P −1πt and p := πP induce the isomorphisms ¯q, ¯p :
¯q = P −1 ◦ πt = P −1πt,
¯p = ¯π ◦ ¯P = πP .
(cid:3)
Proposition A.12 ( ker δ1
Im δ0 ∼= coker πQ−1δ0). Let
0
/ ZE
/ ZV
δ1
δ0
/ ZF
/ 0
be a cochain complex (i.e. δ1 ◦ δ0 = 0) for some V, E, F ∈ N. Let D = P δ1Q be the
Smith normal form of δ1. In particular, all the nonzero entries in the diagonal of
D = diag(d1, ..., dr, 0, ...0)
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)105
are given first. Moreover, di divides di+1, i = 1, . . . , r − 1. Define π : ZE → ZE−r
to be the projection onto the last E − r coordinates, i.e. π is the last E − r rows of
the identity square matrix IE. Let q : ZE−r → ZE be defined by
q := Q ◦ πt,
where πt is the transpose of π, and let p : ZE → ZE−r be defined by
p := π ◦ Q−1.
Im δ0 and coker πQ−1δ0 are isomorphic, with isomorphisms q0 : coker πQ−1δ0 →
Im δ0 → coker πQ−1δ0 induced by q, p, respectively. In particular, (for
Then ker δ1
Im δ0 , p0 : ker δ1
ker δ1
computer testing)
pq = IE−r,
δ1qp = 0.
Proof. From the commutative diagram of cochain complexes (not a short exact
sequence)
0
0
0
/ ZE
∼=
ZE
Q−1
δ0
/ ZV
id∼=
/ ZV Q−1δ0 /
id∼=
/ ZV πQ−1δ0/
π
πt
/ ZE−r
δ1
D /
/ ZF
P∼=
ZF
0
/ 0
/ 0
0
/ 0
we get the isomorphisms
H1(Q−1) : ker δ1
Im δ0
∼ /
ker D
Im Q−1δ0 ,
H1(π) :
ker D
Im Q−1δ0
∼ /
ZE−r
πQ−1δ0ZV = coker πQ−1δ0 ,
where the inverse of the last isomorphism is described next. Note that Im Q−1δ0 ⊂
ker D = 0r ⊕ ZE−r because DQ−1δ0ZV = P δ1δ0ZV = 0. Hence
πtπQ−1δ0 = Q−1δ0,
because πtπ = 0r ⊕ IE−r. So πt induces the map H1(πt) and, since ππt = IE−r,
we get that H1(π)−1 = H1(πt). Thus q := Qπt and p := πQ−1 induce the isomor-
phisms q0, p0 :
q0 = H1(Q) ◦ H1(πt) = H1(q)
p0 = H1(π) ◦ H1(Q−1) = H1(p).
(cid:3)
Proposition A.13. Let A ∈ Mn(Z) be an n × n integer matrix. Suppose that 0
is an eigenvalue of A and let p(x) = xrq(x) be the characteristic polynomial of A.
Then
lim→ (Ar, Zn) = lim→ (Ar, ker q(A)).
/
/
/
/
/
/
/
/
/
/
O
O
/
/
/
/
D. GONC¸ ALVES
106
Proof. Since q(A)Ar = Arq(A) = p(A) = 0, we get ArZn ⊂ ker q(A). Moreover,
A ker q(A) ⊂ ker q(A) because if x ∈ ker q(A) then q(A)x = 0 and thus q(A)Ax =
A(q(A)x) = 0. Hence the following diagram is well-defined
M. RAMIREZ-SOLANO
Zn
Ar
Zn
Ar
Zn
Ar
ι
Ar
ker q(A) Ar
ι
Ar
ker q(A) Ar
ι
Ar
ker q(A) Ar
(cid:3)
and since it commutes lim→ (Ar, Zn) = lim→ (Ar, ker q(A)).
Proposition A.14 (extract the 1-eigenvalues). Let A ∈ Mn(Z) be an n× n integer
matrix and let p(x) be its minimal polynomial. Suppose that λ ∈ Z is an eigenvalue
of A and let q(x) be defined from the equation p(x) = (x − λ)q(x). Let q(x) have
integer coefficients. Then
0
/ lim→ (A, ker q(A))
/ lim→ (A, Zn)
/ lim→ (λIn, q(A)Zn)
/ 0
is a short exact sequence, where In ∈ Mn(Z) is the identity matrix.
Proof. Recall that an eigenvalue of A is a root of the minimal polynomial p(x) of
A, and thus by the assumption it makes sense to write p(x) = (x − λ)q(x). By
definition of minimal polynomial 0 = p(A) = (A − λI)q(A), i.e. Aq(A) = λq(A).
Moreover, A ker q(A) ⊂ ker q(A) because if x ∈ ker q(A) then q(A)x = 0 and thus
q(A)Ax = A(q(A)x) = 0. Hence the following diagram is well-defined
0
0
/ ker q(A)
A
A
/ ker q(A)
/ Zn
q(A) /
/ q(A)Zn
A
/ Zn
λI
q(A) /
/ q(A)Zn
/ 0
/ 0.
Moreover, the rows are clearly short exact sequences and the diagram is commu-
tative. Recall that the direct limit of a directed system of short exact sequences
of abelian groups is a short exact sequence because the direct limit is an exact
functor in the category of abelian groups. Hence the statement of the proposition
(cid:3)
holds.
Proposition A.15. Let m, n ∈ N. Then
] ⊗Z Z[ 1
] = Z[ 1
Z[ 1
m
n
].
mn
Proof. We use the formula "localization of a module" (see [12, Lemma 2.4, p.66]):
S−1M = M ⊗Z S−1Z
( m, s) ∼ (n, t) ⇐⇒ ∃u ∈ S : u(sn − t m) = 0
m
s
:= [( m, s)]∼
with S = {ni i ∈ N0} and M = Z[ 1
that
m]. Notice that 1 ∈ S and ninj ∈ S. Recall
S−1M = {[( m, s)]∼ m ∈ M, s ∈ S}
where
/
/
/
/
/
/
/
/
+
8
8
/
/
+
8
8
/
/
-
<
<
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)107
and S−1M is a module with operations
:= t m + sn
+ n
t
st
m
s
,
a · m
s
:= am
s
.
Thus,
mi
S−1M = { m
s
= { a
= { a
= {
a
= Z[ 1
m
m ∈ Z[ 1
], s ∈ S}
nj a ∈ Z, i, j ∈ N0}
minj a ∈ Z, i, j ∈ N0}
(mn)i a ∈ Z, i ∈ N0}
mn
]
(cid:3)
m
n
(cid:18)
Z[ 1
Corollary A.16. Let m, n, a, b,∈ N. Then
]a ⊗Z Z[ 1
(cid:19)
]b = Z[ 1
mn
Corollary A.17. Let k, ', m, n, a, b, c, d ∈ N. Then
Z[ 1
A.1. Short exact sequences.
Lemma A.18. Let
]a ⊕ Z[1
]c ⊕ Z[ 1
= Z[ 1
Z[ 1
(cid:19)
(cid:18)
]d
n
]b
'
km
⊗
m
k
]ab.
]ac ⊕ Z[ 1
]ad ⊕ Z[ 1
'm
]bc ⊕ Z[ 1
'n
]bd
kn
be a short exact sequence of abelian groups. If Ext1(B, Z) = 0 then the transpose
0
/ I
ι
/ A r /
/ B
/ 0
0
/ Bt
rt
/ At
ιt
/ I t
/ 0
is also a short exact sequence.
Proof. From the given short exact sequence we get the long exact sequence
0
Hom(r,Z) /
/ Hom(B, Z)
Ext1(B, Z) Ext1(r,Z) /
Ext2(B, Z) Ext2(r,Z) /
/ Hom(A, Z) Hom(ι,Z) /
/ Ext1(A, Z) Ext1(ι,Z) /
/ Ext2(A, Z) Ext2(ι,Z) /
Hence if Ext1(B, Z) = 0 we get a short exact sequence.
Corollary A.19. Let
/ Hom(I, Z)
/ Ext1(I, Z)
/ Ext2(I, Z)
be a short exact sequence of abelian groups. Then the transpose
0
/ I
ι
/ A r /
/ Zn
/ 0
0
/ Zn
rt
/ At
ιt
/ I t
/ 0
is also a short exact sequence.
/ ··· .
(cid:3)
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
108
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Proof. We start by constructing a projective resolution
/ 0
/ 0
/ Zn
0
id
/ Zn
−1
/ 0.
Applying Hom(−, Z) to this projective resolution of abelian groups (without the -1
term) we get
0
0
01
Hom0
(Zn, Z)
0
.
Taking homology of this chain complex we get Ext0(Zn, Z) = Hom(Zn, Z) and
Exti(Zn, Z) = 0 for i ∈ N. In particular Ext1(Zn, Z) = 0. Thus by the lemma, the
(cid:3)
corollary follows.
Lemma A.20. Let
/ A
f
/ B
g
/ C h /
/ D
/ 0
be an exact sequence of abelian groups. Then
0
/ B
Im f
ι
/ C h /
/ D
/ 0
is a short exact sequence. In particular,
C
ι( B
Im f )
∼= D.
Proof. Since the sequence is exact, Im f = ker g, and Im g = ker h. Thus
B
Im f
The map induced by g from B
Im f to C is denoted by ι, and it is injective. Since
Im g = ker h, the short sequence is exact. In particular, since h is surjective
= B
ker g
∼= Im g ⊂ C.
C
ι( B
Im f )
= C
ker h
∼= D.
(cid:3)
Appendix B. Direct limits of topological spaces and C∗-algebras
Proposition B.1. Let
be an increasing sequence of T1-spaces, and equip
X0 ⊂ X1 ⊂ X2 ⊂ . . .
X := [
Xn
n∈N0
with the inductive limit topology. Suppose that (xk)∞
k=0 is a sequence in X that
converges to x ∈ X. Then there exists an n0 ∈ N0 such that x ∈ Xn0 and such that
the sequence is eventually in Xn0. That is, there exists a k0 ∈ N0 such that
x, xk ∈ Xn0
for k ≥ k0.
/
/
/
/
/
o
o
o
o
o
o
o
o
o
o
/
/
/
/
/
/
/
/
/
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)109
Proof. Define X−1 := ∅, and assume that xk → x ∈ X. For contradiction, assume
that the conclusion of the proposition is false. Then, without loss of generality, we
can assume that there exists a subsequence (xk')∞
and xk' ∈ X'\X'−1,
(We simply dropped all the spaces not containing an
'=0 such that
for ' ∈ N0,
x ∈ X0
and such that xk0 6= x.
element of the subsequence). Then
xk0 ∈ X0
xk1 ∈ X1,
xk2 ∈ X2,
xk1 6∈ X0
xk2 6∈ X0 ∪ X1,
etc.
Choose a neighborhood x ∈ U0 ⊂ X0 open in X0 such that xk0 6∈ U0. Such
neighborhood exists because X0 is a T1-space. Notice also that xk' 6∈ U0 for ' ≥ 1
because xk' 6∈ X0. Now, choose V1 ⊂ X1 open in X1 such that
U0 = V1 ∩ X0.
This is possible since X0 has the subspace topology of X1. Let
U1 := V1\{xk1} = V1 ∩ {xk1}c,
which is open in X1 since all one point sets in T1-spaces are closed. Then
U1 ∩ X0 = V1 ∩ X0 = U0
since xk1 6∈ X0. Moreover, xk' 6∈ U1 for ' ∈ N0 because U1 ⊂ X1, because we have
removed xk1, and because xk0 6∈ U0 = U1 ∩ X0, hence also xk0 6∈ U1 (as xk0 ∈ X0).
Inductively, we can construct the sets Um ⊂ Xm, m ∈ N0, where
and,
(B.1)
and such that
Let U :=S∞
Um is open in Xm
Um0 = Um ∩ Xm0 ∀ m0 ≤ m,
xk' 6∈ Um ∀ ' ∈ N0.
m=0 Um. Then U is open in X because by Eq. (B.1), U ∩ Xn = Un for
all n ∈ N0, and Un is open in Xn for all n ∈ N0. Since
x ∈ U0 ⊂ U1 ⊂ U2 ⊂ ··· ,
U is a neighborhood of x in X. But xk' 6∈ U for all ' ∈ N0 because xk' 6∈ Um for all
', m ∈ N0. Hence xk' 6→ x, a contradiction. Thus the conclusion of the proposition
(cid:3)
is true.
Lemma B.2. Let Bn(0) be the open ball in Rd of radius n ∈ N, and equip it with
the subspace topology of Rd. Let
X := [
n∈N
Bn(0)
be given the inductive limit topology. Then X = Rd as topological spaces.
Proof. The inclusion map i : X ,→ Rd, is continuous because the inclusion in :
Bn(0) ,→ Rd is continuous for all n ∈ N. We now show that i−1 is continuous. Let
U ⊂ X be open in X. That is, U ∩ Bn(0) is open in Rd for all n ∈ N0. Hence
(cid:3)
n∈N U ∩ Bn(0) is open in Rd.
U =S
110
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
The following lemma is straightforward and well-known.
Lemma B.3. Let Y be a subspace of a topological space X, and suppose that
xn → x in X as n → ∞. If xn, x ∈ Y for all n ∈ N, then xn → x in Y .
Proposition B.4. Let R0 ⊂ R1 ⊂ R2 ⊂ ··· , be the increasing sequence of ´etale
equivalence relations defined in Section 4, and recall that Rn ⊂ Rn+1 is open for
n ∈ N0. Recall as well that An := C∗(Rn), n ∈ N0 is amenable, (thus C∗(Rn) =
C∗
r (Rn)). Then
∞[
C∗(
Rn) = lim→ A0
Proof. Consider the following diagram
n=0
ι0
/ A1
ι1
/ . . .
C∗(R0)
ι0
C∗(R1)
ι1
C∗(R2)
ι3
λ0
λ1
C∗(cid:16) ∞S
Rn
n=0
(cid:17)
.
λ2
n=0Rn), since {Rn}∞
where λn is the inclusion map. For f ∈ Cc(∪∞
n=0 is an open
cover, and the support of f is compact, a finite subset covers the support of f.
Moreover, since the union is increasing, the support of f is in RN for some N ∈ N0.
Hence f = λN( f) for some f ∈ Cc(RN). By slight abuse of notation, we will write
f ∈ Cc(RN) instead of f ∈ Cc(RN).
By a similar proof to Remark 4.24, λn : C∗(Rn) → C∗(R0
s) is continuous. By an
adaptation of the proof of [14, Prop. IV.9], λn is injective, which uses C∗(Rn) =
r (Rn) (amenability). Hence the C∗-algebra C∗(Rn), n ∈ N can be viewed as a
C∗
subalgebra of C∗(R0
s). Then by [36, Exc. 6.2, p.104], the direct limit
⊂ C∗(R0
s)
is a subalgebra. We now show the inclusion ⊃. Let a ∈ C∗(R0
sequence (ak)∞
C∗(Rn)
n=1
s). Then there is a
∞[
·C∗(R0
s)
lim→ (C∗(Rn), ιn) =
k=1 ⊂ Cc(S∞
k=1 ⊂S∞
ak
n=0 Rn) such that
·C∗(R0
s) /
/ a .
By compactness of the support of ak, there exists a nk ∈ N such that ak ∈ Cc(Rnk).
(cid:3)
Hence (ak)∞
n=1 C∗(Rn) and thus a ∈S∞
·C∗(R0
s).
n=1 C∗(Rn)
Appendix C. Compactness in Rd
Lemma C.1. Let U be an open subset of Rd, and let its Rd-boundary be non-empty.
Let K ⊂ U be compact in U. Then
d(K, ∂U) > 0.
Proof. Note K∩∂U is empty because K is a subset of U and U∩∂U = U∩(U\ ◦
U) = ∅
as U is open. Suppose for contradiction that d(K, ∂U) = 0. Then there is a sequence
(xn)∞
n=1 ⊂ K such that
d(xn, ∂U) → 0
as n → ∞.
/
/
/
/
%
%
/
/
/
/
y
y
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)111
Since K is compact, there exists a convergent subsequence xnj → x ∈ K. Since
d(xnj , ∂U) = inf
u∈∂U
xnj − u,
we have
d(x, ∂U) ≤ x − xnj + d(xnj , ∂U) → 0.
n=1 ⊂ ∂U converging to x. Since
Hence d(x, ∂U) = 0. Thus there is a sequence (un)∞
∂U = U ∩ Rd\U is closed, x ∈ ∂U. Thus x ∈ K ∩ ∂U = ∅, a contradiction.
(cid:3)
Lemma C.2. Let X be a subset of Rd, and let K ⊂ X. Then K is X-compact if
and only if K is Rd-compact.
Proof. If K is X-compact, then it is Rd-compact because the inclusion map X ,→ Rd
is continuous. If K is Rd-compact then it is Rd-closed and bounded. Since it is
Rd-closed, it is complete. Since it is bounded, then by the Archimedean property
(cid:3)
it is totally bounded. Thus K is X-compact.
Appendix D. Cc(G)
Let G be a locally compact groupoid with left Haar system {λu}u∈G0. Then the
set Cc(G) of continuous C-valued functions with compact support is equipped with
the inductive limit topology (cf. [10, Example 5.10, p.118]), and with operations
Z
:=
f ? g(x)
f∗(x)
y∈r−1(s(x))
:= f(x−1).
f(xy)g(y−1)dλs(x)(y)
We remark that the names Renault-representation and Renault-bounded (given
Under these operations Cc(G) is a ∗-algebra (cf. [33, p. 48]).
below) are not standard terminology, but we use them for the sake of brevity.
Definition D.1 (Renault-representation). ([33, p.50]) A Renault-representation of
Cc(G) on a Hilbert space H is a ∗-homomorphism
L : (Cc(G), ind) → (B(H), WO)
which is continuous, where Cc(G) has the inductive limit topology, and B(H) has
the WO-topology (weak-operator topology). Moreover, the linear span {L(f)ξ
f ∈ Cc(G), ξ ∈ H} is dense in H.
By [33, Prop. I.4, p.50], the topology given by the one-norm · I on Cc(G) is
coarser than the inductive limit topology, i.e.
where
fI = max(cid:16) sup
Z
ind /
/ f
fn
=⇒ fn − fI → 0,
f(x) dλu(x) ,
Z
sup
u∈G0
u∈G0
x∈r−1(u)
x−1∈r−1(u)
for f ∈ Cc(G). Moreover, · I is a ∗-algebra norm on Cc(G).
Definition D.2 (Renault-bounded). (cf. [33, p. 50-51]) A Renault-representation
f(x) dλu(x−1)(cid:17)
,
L : (Cc(G), ind) → (B(H), WO)
is said to be Renault-bounded if
L(f)B(H) ≤ fI ∀f ∈ Cc(G),
D. GONC¸ ALVES
112
where · B(H) is the operator norm of B(H).
(Thus L : (Cc(G), · I) → (B(H), · B(H)) is continuous).
M. RAMIREZ-SOLANO
For f ∈ Cc(G) define
f := sup
L(f)B(H),
L
where the supremum is taken over all Renault-bounded Renault-representations
(L, H) of Cc(G). This norm on Cc(G) is a C∗-norm, and the C∗-algebra C∗(G) of
the groupoid G is defined as the completion
C∗(G) := Cc(G)·
.
(cf. [33, p.51-58]). Observe that for f ∈ Cc(G),
f ≤ fI .
(D.1)
Lemma D.3. If
L : (Cc(G), ind) → (B(H), WO)
is a Renault-bounded Renault-representation, then
L : (Cc(G), · ) → (B(H), · B(H))
is continuous.
Proof. By definition of f, f ∈ Cc(G) we have L(f)B(H) ≤ f. Hence the
(cid:3)
linear operator L(f) is bounded (with norm at most 1) and thus continuous.
Lemma D.4. If L : C∗(G) → B(H) is a faithful representation (in the standard
way) then L restricted to Cc(G) is a Renault-bounded Renault-representation.
n=1 ⊂ Cc(G) converges to f ∈ Cc(G) in the in-
Proof. Suppose that a sequence (fn)∞
ductive limit topology. Since the I-norm is coarser than the inductive limit topology
we have
fn − fI → 0.
Hence by Eq. (D.1)
fn − f → 0.
Since L is faithful, we get by [47, Theorem 10.7] that
L(fn) − L(f)B(H) = fn − f → 0.
Since convergence in the norm-topology implies convergence in the weak-operator
topology we get
L(fn) WO /
/ L(f).
Thus L is a Renault-representation. It is Renault-bounded since
L(f)B(H) = f ≤ fI ,
by [47, Theorem 10.7] and Eq. (D.1).
(cid:3)
For convenience to the reader, we provide a proof of the following well-known
result:
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)113
Theorem D.5 (extension by continuity for groupoids). Let G0 and G be locally
compact groupoids with Haar systems, and assume that G0 is second countable and
´etale. Let ψ : Cc(G0) → Cc(G) be a continuous ∗-homomorphism. Then ψ extends
to a (continuous) ∗-homomorphism
ψ : C∗(G0) → C∗(G).
Proof. Let (L, H) be a faithful representation of C∗(G). By Lemma D.4, L is
a Renault-representation Renault-bounded when restricted to Cc(G). Since ψ is
continuous in the inductive limit topology and it is a ∗-homomorphism, L ◦ ψ is
a Renault-representation. Then by [33, Corollary 1.22, p. 72], L ◦ ψ is Renault-
bounded. By Lemma D.3,
L ◦ ψ : (Cc(G0), · C∗(G0)) → (B(H), · B(H))
is bounded. Hence we can extend L◦ ψ by continuity to C∗(G0) and we denote the
extended map by (cid:93)L ◦ ψ. Thus
is a (continuous) ∗-homomorphism (as L is a ∗-isomorphism onto its image in
(cid:3)
B(H)).
ψ = L−1 ◦ (cid:93)L ◦ ψ
Appendix E. On T + Rd with the subspace topology of Ω
Proposition E.1. Let T + Rd ⊂ Ω be given the subspace topology. Then the map
T + Rd → Rd given by
T + x 7→ x
is not continuous.
Proof. Let Pn := T(Bn(0)), n ∈ N, be the patch containing the ball Bn(0) of radius
n, and note that Pn ⊂ Pn+1. Choose patch Pn + xn in T such that xn ≥ n. This
can be done by [24, Thm 2.3, p.3] and [1, Cor. 3.5, p.11]. Since T and T − xn agree
on the patch Pn and thus on Bn(0), we have
d(T, T − xn) ≤ 1
n
Thus T − xn converges to T, but xn 6→ 0 because
xn → ∞.
.
(cid:3)
Appendix F. Projections in C∗
r (RX0)
for fixed n ∈ N0.
We use the notation of Section 4. For simplicity of notation, let T := Tn, R := Rn
Let f ∈ Cc(RX0) be defined by
f(x, y) =
if x = y = v0
otherwise
i
where x ∼ y. Since f∗(x, y) = f(y, x) = f(y, x) = f(x, y), f is self-adjoint. For
x ∼ z we have
(cid:26) 1
(f ? f)(x, z) =X
0
x∼y
f(x, y)f(y, z).
114
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
f(x, y) =
0
Thus (f ? f)(x, z) 6= 0 only if x = z = v0
f(y, z) =
if x = y = v0
otherwise
i
if y = z = v0
otherwise
i . In this case,
i
0
(cid:26) 1
(cid:26) 1
i ) = X
i
y∼v0
= f(v0
= 1.
i , v0
i )f(v0
i , v0
i )
(f ? f)(v0
i , v0
f(v0
i , y)f(y, v0
i )
Hence, f ? f = f. Hence f is a projection in C∗
r (RX0).
Below we give a different proof of [15, Prop. 3.7-9].
Proposition F.1. [15, Prop. 3.7-9]
sVL
r (RX0) ∼=
r (RX1−X0) ∼=
r (RX2−X1) ∼=
k=1
(1) C∗
(2) C∗
(3) C∗
K('2([vk]))
sEL
sFL
k=1
k=1
C0(◦
C0(◦
ek, K('2([ek])))
σk, K('2([σk])))
where sV, sE, sF are the number of stable vertices, stable edges, stable faces, respec-
tively, and vk, ek, σk are representatives of the R-equivalence classes.
Proof. Let N := sV be the number of stable vertices,
i.e. the number of R-
equivalence classes. Then the underlying space of the 0-skeleton of T is the set
of vertices
X0 = {v ∈ Rd v ∈ [v1] ∪ ··· ∪ [vN]},
where the vertices v0, . . . , vN ∈ T are (fixed) representatives of the R-equivalence
classes. We remark that each of these equivalence classes has countable many
vertices. The restriction of R to the vertices is
R X0= {(x, y) ∈ Rd × Rd (x, y) ∈ [v1] × [v1] ∪ ··· ∪ [vN] × [vN]}.
The C∗-algebra C∗
r (RX0) is the completion inside B('2(X0)) of
{λ(f) f ∈ Cc(RX0)},
(λ(f)ξ)(x) =X
y∼x
f(x, y)ξ(y).
where
For f ∈ Cc(R X0) let
Then
fi := f[vi]×[vi],
i = 1, . . . , N.
f = f1 + ··· + fN ,
supp(fi) ⊂ [vi] × [vi]
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)115
and
[v1]
f1(x, y)
0
0
0
0
λ(f) =
We have
[v2]
0
f2(x, y)
0
0
0
···
0
0
...
0
0
r (RX0) ∼=
C∗
[vN]
0
0
0
0
fN(x, y)
[v1]
[v2]
...
[vN−1]
[vN]
[vN−1]
0
0
0
0
fN−1(x, y)
sVM
K('2([vi]))
i=1
sVM
i=1
because λ : Cc(RX0) ,→ B('2(X0)) is an injective ∗-homomorphism and
r (RX0) = λ(Cc(RX0))·B(H) =
C∗
K('2([vi])),
where H is the Hilbert space '2(X0). This is fairly easy to show.
Now let N := sE be the number of stable edges, and J := Jn. Then the
underlying space of the 1-skeleton of T is
X1 = {x ∈ e e ∈ [e1] ∪ ··· ∪ [eN]},
where the edges e1, . . . , eN ∈ T are (fixed) representatives of the R-equivalence
classes. The restriction of R to the 1-skeleton minus its vertices is
R X1−X0= {(x, γe,e0(x)) x ∈ ◦
e, (e, e0) ∈ [e1] × [e1] ∪ . . . ∪ [eN] × [eN]},
where γe,e0(x) := x+ x0, with x0 being the translation of the two edges: e0 = e+ x0.
For f ∈ Cc(R X1−X0) let
fi := fR[ei]×[ei]−X0
,
i = 1, . . . , N,
R [ei]×[ei]−X0= {(x, γe,e0(x)) x ∈ ◦
e, (e, e0) ∈ [ei] × [ei]}.
where
Then
f = f1 + ··· + fN ,
supp(fi) ⊂ R[ei]×[ei]−X0 .
Since f has compact support and we have removed the vertices, f is zero in a neigh-
e0 ⊂ RX1−X0
borhood of the vertices. (Recall that the topology of the graph γ◦
is the topology of ◦
e, which by definition is homeomorphic to R.) Thus fi has
LN
r (RX1−X0) is the completion inside
compact support as well. The C∗-algebra C∗
(cid:0)◦
ei, B('2([ei]))(cid:1) of
e,
◦
i=1 C0
{λ(f) f ∈ Cc(RX1−X0)},
where
and λ(fi) : ◦
λ(f) = (λ(f1), . . . , λ(fN)),
ei → B('2([ei])) is defined on s ∈ ◦
λ(fi)(s)(ξ)(e) = X
ei as
(cid:0)γei,e(s), γei,e0(s)(cid:1)ξ(e0),
fi
e∼e0
116
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
, and it is zero on a
Since fi is non-zero only on a finite number of the graphs γ◦
◦
e
ei,
neighborhood of its vertices, λ(fi) is zero in a smaller neighborhood of its vertices
as well. Thus λ(fi) has also compact support. We have
sEM
r (RX1−X0) ∼=
C∗
(cid:0)◦
ei, B('2([ei]))(cid:1) is an injective ∗-homomorphism
(cid:0)◦
ei, K('2([ei]))(cid:1)
i=1
C0
i=1 Cc
because λ : Cc(RX1−X0) ,→LN
NM
r (RX1−X0) =
C∗
and
i=1
·r
sEM
Cc(R[ei]×[ei]−X0)
=
C0( ◦
ei, K('2([vi]))).
The proof of (3) is similar to the proof of (2).
i=1
(cid:3)
It follows that
K0(C∗
(cid:16) sVM
r (RX0)) ∼= K0
(cid:16)
sVM
i=1
=
K('2([vi]))(cid:17)
K('2([vi]))(cid:17) ∼= ZsV ,
K0
i=1
where the equality in the second line follows because the number of stable vertices
sV is finite.
It is well-known that the K0-group of the compact operators K('2([v])) is iso-
morphic to Z and that it has generator [p]0, for any 1-dimensional projection p in
K('2([v])). Take
(F.1)
fi(x, y) =
Now we show that λ(fi) is a one dimensional projection in K('2([vi])): Let
ξ ∈ '2([v0
i ]). For x ∈ [v0
i ],
(cid:26) 1
0
x = y = v0
otherwise.
i
fi(x, y)ξ(y)
(λ(fi)ξ)(x) = X
(cid:26) ξ(v0
y∈[v0
i ]
=
i
i
x = v0
x 6= v0
i )
0
. If ξ ∈ '2([vj]), j 6= i, then for x ∈ [vj],
j ]. Hence λ(fi) is a one dimensional
i ])). Therefore,
So λ(fi)'2([v0
i ]) is a projection onto Cδv0
i
i for all x ∈ [v0
(λ(fi)ξ)(x) = 0 because x 6= v0
projection in K('2([v0
i ]∈V K('2([v0
[v0
K0(C∗
r (RX0))
i ])) ⊂L
has generators
where f1, . . . , fn are defined in (F.1).
[f1]0, . . . , [fn]0
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)117
F.1. Exponential map. Recall that we are using the notation of Section 4 and
T := Tn and R := Rn, n ∈ N0.
Proposition F.2 ([14, Proposition III.28]). Assume that the dimension of tiling T
is 1. The exponential map δ0 : ZsV → ZsE is given by
δ0(e1.e2) = e1 − e2.
Proof. Recall that a 1-dimensional projection generates the K0-group of the com-
pact operators K. One can choose any 1-dimensional projection because all are
equivalent (cf. Lemma G.2). Let v ∈ T be a vertex in the stable vertex e1.e2; let p
be the 1-dimensional projection in C = C∗
p(x, y) =
r(RX0) given by
(x, y) = (v, v)
else.
Let a ∈ B be the self-adjoint operator given by
a(x, y) =
x = y = s ∈ e1
1 − s x = y = s ∈ e2
0
else.
(where we ignore reparametrization of the edge, e.g. we identify e2 with the unit
interval [0, 1]). Note that p is a projection that corresponds to the stable edge e1.e2.
Moreover, the support of p and a is on the diagonal of R.
0
(cid:26) 1
s
•1
1 − s
x = y
R
s
y
e2
•
v
e1
•
v
e1
a(x, y)
e2
x
Then p := (p, 0) is a projection in the unitization C of C, and a := (a, 0) is a selfad-
r (RX1). By [36, Proposition 12.2.2],
joint operator in the unitization B of B = C∗
there exists a unique unitary u ∈ J, where J is the unitization of J = C∗
r (RX1−X0),
such that
i(u)(x, y) = e2πia(x, y)
and such that
δ0([p]0) = δ0([p]0 − [s(p)]0) = [u]1.
and note that the scalar part s(p) = 0 is zero. We would like to remark that we are
using a different convention than Rørdam's (he defines δ0([p]0) = −[u]1). Our δ0
118
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
corresponds to the standard differential defining cellular cohomology. We evaluate
e2πia(x, y) and get
e2πia(x, y) =
e2πis
e2πi(1−s) = e−2πis
1
0
s = x = y ∈ e1
s = x = y ∈ e2
x = y 6∈ e1 ∪ e2
else.
Note that the evaluation of the exponential function e2πia(x, y) is done pointwise
since a has support on the diagonal ∆R of R and C∗
We want to describe u in terms of the compacts, that is, up to the isomorphism
given in Proposition 4.2. But first note that u is in the unitization of J, and recall
that the unitization of the compacts is K(H) + CIH. Suppose that e1 6∼R e2 as
stable edges (we are only working on the diagonal of R). Then
r (∆R) = C0(∆R).
δ0([p]0) = [e2πis ⊗ ee1e1 + e−2πis ⊗ ee2e2 + X
= [e2πis ⊗ ee1e1 − e2πis ⊗ ee2e2 + X
e36∼e1,e36∼e2
1 ⊗ ee3e3]1.
1 ⊗ ee3e3]1,
e36∼e1,e36∼e2
(where the second equality is because for unitary u it holds 0 = [uu−1]1 = [u]1 +
[u−1]1 by Proposition 8.1.4 in [36]). Remark: e2πis is just a unitary in the unitiza-
tion of C0((0, 1)) i.e. is in A := {f ∈ C[0, 1] f(0) = f(1)}. [e2πis]1 is a generator
for K1(A) (see Bott element in [36] 11.3). Thus
If e1 ∼R e2 as stable edges then we get, by Whitehead's lemma (Lemma 2.1.5 in
[36]), that
δ0(e1.e2) = e1 − e2.
(cid:19)
(cid:18) e2πis
(cid:19)
δ0([p]0) = [e2πis ⊗ ee1e1 + e−2πis ⊗ ee2e2 + X
(cid:18) 1
e−2πis
∼h
0
0
1
0
0
.
1 ⊗ ee3e3]1 = 0.
e36∼e1
Hence
Thus
where the last equality is because e1 and e2 are R-equivalent.
δ0(e1.e2) = 0 = e1 − e2,
(cid:3)
Appendix G. Background review
The following lemma is closely related to the splitting lemma.
Lemma G.1. Let A, B, C be R-modules, where R is a commutative ring. Let
r
0
/ A i
/ B
s
π
/ C
/ 0
be a split short exact sequence. Thus the injective map i, the surjetive map π, the
retraction map r, and the section map s are all homomorhism, and they satisfy
r ◦ i = 1A,
i ◦ r + s ◦ π = 1B,
π ◦ s = 1C,
π ◦ i = 0,
r ◦ s = 0.
/
/
}
}
/
/
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)119
This is equivalent to B being isomorphic to the direct sum A ⊕ C. More precisely,
the following diagram commutes
/ A
1A∼=
/ A
/ B
φ∼=
/ A ⊕ B π0
/ C
1C∼=
/ C
/ 0,
0
0
i
i0
π
/ 0
where i0(a) = (a, 0) and π0(a, c) = c.
Proof. We will show that B ∼= A ⊕ C. Define
φ(b) := (r(b), π(b))
φ−1(a, c) := i(a) + s(c).
Then
φ−1 ◦ φ(b) = φ−1(r(b), π(b)) = i ◦ r(b) + s ◦ π(b) = b
φ◦φ−1(a, c) = φ(i(a)+s(c)) = (r(i(a)+s(c)), π(i(a)+s(c))) = (a+0, 0+c) = (a, c).
The first square commutes since
φ ◦ i(a) = (r(i(a)), π(i(a))) = (a, 0) = i0(a).
The second square commutes since
π0 ◦ φ(b) = π0(r(b), π(b)) = π(b) = 1C ◦ π(b).
Conversely, assume B ∼= A ⊕ C. Define
0
/ A
r
i /
/ A ⊕ C
s
π
/ C
/ 0
as
i(a) = (a, 0),
π(a, c) = c,
r(a, c) = a,
s(c) = (0, c).
Moreover, we get
ker π = Im i,
ker r = Im s,
i ◦ r + s ◦ π = 1A⊕C.
r ◦ i = 1A,
π ◦ s = 1C
Define r0 := 1A ◦ r ◦ φ−1, s0 := φ ◦ s ◦ 1C. Then we have the commutative diagram
(clockwise and counterclockwise)
0
0
/ A
1A
/ A
/ B
φ∼=
/ A ⊕ B
r
i
i0
r0
s
π
π0
s0
/ C
1C
/ C
/ 0
/ 0.
(cid:3)
If A is nuclear and B is any arbitrary C∗-algebra, then only the minimal tensor
product is possible to make A ⊗ B into a C∗-algebra. The unitization of the
compacts K := K(H) is K(H) + CIH. Moreover (cid:94)A ⊗ K ⊂ A ⊗ K. Equality
holds if A or K is unital. (K is unital if K is finite dimensional). Otherwise the
inclusion is proper since a ⊗ 1 and 1 ⊗ k are not in the unitization of A ⊗ K.
/
/
/
/
/
/
/
/
/
z
z
/
y
y
/
/
/
z
z
/
z
z
/
/
/
c
c
/
d
d
/
120
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
Lemma G.2. Let E, F be 1-dimensional projections on a Hilbert space H, and let
K(H) be the compact operators. Then E ∼ F in K(H) i.e. there is a V ∈ K(H)
such that V ∗V = E and V V ∗ = F.
Proof. K := E(H) and L := F(H) are 1-dimensional subspaces of H. Thus, there
is a V0 : K → L isometry of K onto L. Extend to H by putting
Then V ∈ B(H), V ∗V = E, V V ∗ = F. Since V (H) = F(H) are 1-dimensional, V
is finite dimensional, hence compact i.e. V ∈ K(H).
(cid:3)
V x := V0Ex.
ON THE K-THEORY OF C∗-ALGEBRAS FOR SUBSTITUTION TILINGS (A PEDESTRIAN VERSION)121
References
[1] J. E. Anderson, I. F. Putnam. Topological Invariants for Substitution Tilings and their Associ-
ated C∗-algebras. Department of Mathematics and Statistics, University of Victoria, Victoria
B.C. Canada. 1-45. (1995)
[2] D. M. Arnold. Finite Rank Torsion Free Abelian Groups and Rings. Lecture Notes in Math-
ematics 931. Springer-Verlag Berlin Heidelberg New York. (1982)
[3] M. Barge, B. Diamond. Cohomology in one-dimensional
Proc. Amer. Math. Soc. 136, no. 6, 2183-2191. (2008)
substitution tiling spaces.
[4] M. Barge, B. Diamond, J. Hunton, L. Sadun. Cohomology of substitution tiling spaces. Ergodic
Theory Dynam. Systems, 30(6):1607-1627. (2010)
[5] J. Bellissard. K-theory of C*-algebras in solid state physics. In Statistical Mechanics and Field
Theory: Mathematical Aspects, eds. T. C. Dorlas, N. M. Hugenholtz and M. Winnik. Lecture
Notes in Physics, vol. 257, Springer, Berlin. 99-156. (1986)
[6] J. Bellissard. Gap labelling theorems for Schrodinger's Operators. In From Number Theory
to Physics, eds. M. Waldschmidt, P. Moussa, J. M. Luck and C. Itzykson, Springer, Berlin.
539-630. (1993)
[7] J. Bellissard, R. Benedetti, J. M. Gambaudo. Spaces of tilings, finite telescopic approximations
and gap-labelling. Comm. Math. Phys., 261(1), 1-41. (2006)
[8] B. Blackadar. K-Theory for operator algebras. MSRI Publ. 5, Cambridge University Press.
[9] A. Clark, L. Sadun. When shape matters: deformations of tiling spaces. Ergodic Theory
Dynam. Systems, 26 (1), 69-86. (2006)
[10] J. B. Conway. A course in functional analysis. Springer-Verlag. (1990)
[11] M. Dugas. Torsion-Free abelian groups defined by an integral matrix. Int. J. Algebra. Vol 6.
[12] D. Eisenbud. Commutative algebra, with a view toward algebraic geometry. Springer-Verlag.
no. 1-4, 85-99. (2012)
(1995)
(1998)
Topology Appl. 160, no. 5, 703-719. (2013)
[13] F. Gahler, G. Maloney. Cohomology of one-dimensional mixed substitution tiling spaces
[14] D. Gon¸calves. C∗-algebras from Substitution Tilings: A New Approach. Ph.D. Thesis,
[15] D. Gon¸calves. New C∗-algebras from substitution tilings. J. Operator Th., 2, 57, 391-407.
the stable C∗-algebras from substitution tilings.
Univ. Of Victoria, Victoria, Canada. (2005)
(2007)
[16] D. Gon¸calves. On the K-theory of
J. Func. Anal. 4, 260, 998-1019. (2011)
[17] C. Janot. Quasicrystals: A Primer. Clarendon press, Oxford. (1994)
[18] A. Julien, J. Savinien. K-theory of the chair tiling via AF-algebras. Journal of Geometry and
Physics, Vol. 106, 314-326. (2016)
[19] J. Kaminker, I. F. Putnam, M. Whittaker. K-Theoretic duality for hyperbolic dynamical
system. J. Reine Angew. Math., 730, 263-299. (2017)
[20] J. Kellendonk. Non-commutative geometry of tilings and gap labelling. Rev. Math. Phys. 7,
[21] J. Kellendonk. The local structure of tilings and their integer group of coinvariants. Com-
[22] J. Kellendonk. Integer groups of coinvariants associated to octagonal tilings. Fields Insitute
1133-1180. (1995)
mun. Math. Phys. 187, 115-157. (1997)
Communications 13, 155-169. (1997)
[23] [K] J. Kellendonk, Pattern-equivariant functions and cohomology, J. Phys. A 36 (2003) 5765-
5772. J. Kellendonk, Pattern-equivariant functions and cohomology, J. Phys. A 36 (2003),
no. 21, 5765-5772.
[24] J. Kellendonk, I. F. Putnam. Tilings, C∗-algebras and K-theory. Directions in mathematical
quasicrystals, CRM Monogr. Ser., 13, Amer. Math. Soc., Providence, RI. (2000)
no. 3, 693-711. (2006)
[25] J. Kellendonk, I. F. Putnam. The Ruelle-Sullivan map for actions of Rn, Math. Ann. 334,
[26] P. S. Muhly, J. N. Renault, D. P. Williams. Equivalence and isomorphism for groupoid C∗-
[27] P. S. Muhly, J. N. Renault, D. P. Williams. Continuous-trace groupoid C∗-algebras III.
algebras. J. Operator Th. 17, 3-22. (1987)
Trans. Amer. Math. Soc. 348,9, 3621-3641. (1996)
122
D. GONC¸ ALVES
M. RAMIREZ-SOLANO
pages. (2014)
algebras.pdf. (2014)
systems. J. Func. Anal., 163:279-299. (1999)
[28] I. F. Putnam, A homology theory for Smale spaces. Memoirs A.M.S., Vol 232(1094), 122
[29] I. F. Putnam. Lecture notes on C∗-algebras. http://www.math.uvic.ca/faculty/putnam/ln/C*-
[30] I. F. Putnam, J. Spielberg. The structure of C∗-algebras associated with hyperbolic dynamical
[31] I. F. Putnam, C∗-algebras from Smale spaces. Canad. J. Math, 48, 175-195. (1996)
[32] J. D. Reid, Abelian groups finitely generated over their endomorphism rings, Springer-Verlag
Lecture Notes 874, 41-52 in Abelian Group Theory (Proceedings of the Oberwolfach Confer-
ence). (1981).
[33] J. Renault, A groupoid approach to C∗-algebras. Lecture Notes in Mathematics, No.793,
[34] J. Renault, The ideal structure of groupoid crossed product C∗-algebras. J. Operator Theory
[35] M. A. Rieffel. Induced representations of C∗-algebras, Adv. Math. 13, 176-257. (1974).
[36] M. Rørdam, F. Larsen, N.J. Laustsen. An introduction to K-Theory for C∗-algebras. J. Op-
Springer-Verlag, Berlin-New York, (1980).
25,3-36. (1991).
erator Theory 25,3-36. (1991).
ical Systems 31, 1819-1834. (2011)
[37] L. Sadun. Exact regularity and the cohomology of tiling spaces. Ergodic Theory and Dynam-
[38] L. Sadun. Pattern-equivariant cohomology with integer coefficients. Ergodic Theory and Dy-
namical Systems 27(6), 1991âĂŞ1998. (2007)
[39] L. Sadun. Tilings, tiling spaces and topology. Philos. Mag. 86, 875-881 (2006)
[40] L. Sadun. Topology of Tiling Spaces. University Lecture Series Vol. 46, Providence, Rhode
Island. (2008)
[41] P. J. Steinhardt, S. Ostlund. The Physics of quasicrystals. World Scientific. (1987)
[42] J. J. Walton. Pattern-equivariant homology arxiv.org:1401.8153. (2014)
[43] R. Willett. A non-amenable groupoid whose maximal and reduced C∗-algebras are the same.
arxiv.org:1504.05615. (2015)
215-264. (1999)
[44] J. L. Tu. La conjecture de Baum-Connes pour les feuilletages moyennables. K-Theory 17,
br/pos/Gustavo_Felisberto_Valente.pdf. Master Thesis. Advisor D. Gon¸calves (2013)
[45] G. F. Valente. Cohomologia associada a ladrilhamentos de substitui¸cao. http://mtm.ufsc.
[46] N.E. Wegge-Olsen. K-theory and C∗-algebras. Oxford university press. (1993)
[47] K. Zhu. An introduction to operator algebras. CRC press. (1993)
Daniel Gon¸calves, Departamento de Matem´atica, Universidade Federal de Santa Cata-
rina, Florian´opolis, 88.040-900, Brazil.
E-mail address: [email protected]
Maria Ramirez-Solano, Department of Mathematics, University of Southern Denmark,
Odense, Denmark.
E-mail address: [email protected]
|
1807.01686 | 1 | 1807 | 2018-07-04T17:07:12 | $C^*$-algebras of self-similar graphs over arbitrary graphs | [
"math.OA"
] | In this note we extend the construction of a $C^*$-algebra associated to a self-similar graph to the case of arbitrary countable graphs. We reduce the problem to the row-finite case with no sources, by using a desingularization process. Finally, we characterize simplicity in this case. | math.OA | math | C ∗-ALGEBRAS OF SELF-SIMILAR GRAPHS OVER ARBITRARY
GRAPHS
RUY EXEL, ENRIQUE PARDO, AND CHARLES STARLING
Abstract. In this note we extend the construction of a C ∗-algebra associated to a self-
similar graph to the case of arbitrary countable graphs. We reduce the problem to the
row-finite case with no sources, by using a desingularization process. Finally, we characterize
simplicity in this case.
8
1
0
2
l
u
J
4
]
.
A
O
h
t
a
m
[
1
v
6
8
6
1
0
.
7
0
8
1
:
v
i
X
r
a
Introduction
C ∗-algebras of self-similar graphs were introduced by the first and second authors in [8]
for the case of discrete countable groups and finite graphs with no sources. Lately, B´edos,
Kaliszewski and Quigg [2] extended the definition to arbitrary groups and (topological)
graphs, using a Cuntz-Pimsner picture of these algebras; unfortunately, they have not pro-
vided characterizations of simple or purely infinite for these algebras under this point of
view. Also, no one has developed directly the results in [8] to the context of countable
discrete groups and arbitrary graphs. This note we fill the gap, by reducing the problem
row-finite graphs with no sources, and then showing how characterizations stated in [8] works
correctly under this restrictions. The results in this note are essential to extend the scope of
the results obtained in [8] for Katsura algebras over finite matrices to the general case; this
guarantees, by [9], that every Kirchberg algebra in the UCT is the full groupoid C ∗-algebra
of a second countable amenable ample groupoid.
1. The case of finite graphs
In this section we will recall the essential items needed to understand the algebras OG,E
associated to triples (G, E, ϕ), introduced in [8]. Let us recall the construction.
1.1. The basic data for our construction is a triple (G, E, ϕ) composed of:
(1) A finite directed graph E = (E0, E1, r, s) without sources.
(2) A discrete group G acting on E by graph automorphisms.
(3) A 1-cocycle ϕ : G × E1 → G satisfying the property
ϕ(g, a) · x = g · x for every g ∈ G, a ∈ E1, x ∈ E0.
2010 Mathematics Subject Classification. 46L05, 46L55.
Key words and phrases. Self-similar graph,
inverse semigroup, tight representation, tight groupoid,
groupoid C ∗-algebra.
The first-named author was partially supported by CNPq. The third named author was partially supported
by PAI III grant FQM-298 of the Junta de Andaluc´ıa, and by the DGI-MINECO and European Regional
Development Fund, jointly, through grants MTM2014-53644-P and MTM2017-83487-P. The third named
author was partially supported by a Carleton University internal research grant.
1
2
RUY EXEL, ENRIQUE PARDO, AND CHARLES STARLING
The property (3) required on ϕ is tagged (2.3) in [8]. As remarked in [2], this condition
can be weakened to
without pain.
ϕ(g, a) · s(a) = g · s(a) for every g ∈ G, a ∈ E1,
Definition 1.2. Given a triple (G, E, ϕ) as in (1.1), we define OG,E to be the universal
C ∗-algebra as follows:
(1) Generators:
(2) Relations:
{px : x ∈ E0} ∪ {sa : a ∈ E1} ∪ {ug : g ∈ G}.
(a) {px : x ∈ E0} ∪ {sa : a ∈ E1} is a Cuntz-Krieger E-family in the sense of [11].
(b) The map u : G → OG,E defined by the rule g 7→ ug is a unitary ∗-representation
of G.
(c) ugsa = sg·auϕ(g,a) for every g ∈ G, a ∈ E1.
(d) ugpx = pg·xug for every g ∈ G, x ∈ E0.
Notice that the relation (2a) in Definition 1.2 implies that there is a natural representation
map
φ : C ∗(E) → OG,E
7→ px
7→ sa
px
sa
which is injective [8, Proposition 11.1].
Recall from [8, Definition 4.1] that given a triple (G, E, ϕ) as in (1.1), we define an inverse
semigroup SG,E as follows:
(1) The set is
SG,E = {(α, g, β) : α, β ∈ E∗, g ∈ G, s(α) = gs(β)} ∪ {0}},
where E∗ denotes the set of finite paths in E.
(2) The operation is defined by:
(α, g, β) · (γ, h, δ) :=
and (α, g, β)∗ := (β, g−1, α).
(α, gϕ(h, ε), δhε),
if β = γε
(αgε, ϕ(g, ε)h, δ),
if γ = βε
0,
otherwise,
Then, we can construct the groupoid of germs of the action of SG,E on the space of tight
filters bEtight(SG,E) of the semilattice E(SG,E) of idempotents of SG,E. In our concrete case,
bEtight(SG,E) turns out to be homeomorphic to the compact space E∞ of one-sided infinite
paths on E; in particular, bEtight(SG,E) = bE∞(SG,E). Hence, the action of (α, g, β) ∈ SG,E on
η = βbη is given by the rule (α, g, β) · η = α(gbη). Thus, the groupoid of germs is
G(G,E)
tight = {[α, g, β; η] : η = βbη},
ARBITRARY SELF-SIMILAR GRAPHS
3
where [s; η] = [t; µ] if and only if η = µ and there exists 0 6= e2 = e ∈ SG,E such that e · η = η
and se = te. The unit space
is identified with the one-sided infinite path space E∞, via the homeomorphism [α, 1, α; η] 7→
η. Under this identification, the range and source maps on G(G,E)
tight are:
(0)
G(G,E)
tight
= {[α, 1, α; η] : η = αbη}
A basis for the topology on G(G,E)
tight
s([α, g, β; βbη]) = βbη
and
r([α, g, β; βbη]) = α(gbη).
is given by compact open bisections of the form
Θ(α, g, β; Z(γ)) := {[α, g, β; ξ] ∈ G(G,E)
tight
: ξ ∈ Z(γ)}
where γ ∈ E∗ and Z(γ) := {γbη : bη ∈ E∞}. Thus G(G,E)
[8] characterizations are given for when G(G,E)
is Hausdorff [8, Theorem 12.2], amenable [8,
tight
Corollary 10.18], minimal [8, Theorem 13.6] or effective [8, Theorem 14.10] in terms of the
properties of the triple (G, E, ϕ) and the action of SG,E on E∞.
is locally compact and ample. In
tight
2. Extending to the countable case
In this section we will look at the problem of extending our class to the case of countably
infinite graphs with no restrictions (i.e. sources, sinks and infinite receivers are admitted).
Notice that, if E0 is countably infinite, then the algebra C ∗(E) is not longer unital. Thus,
if we pretend to get a unitary representation of the group G associated to the algebra, we
need to consider unitary representations of G in the multiplier algebra M(OG,E). This is
not a good idea for giving an intrinsic definition of the object, but it is very helpful to deal
with the details in the definition. Nevertheless, the model we will follow is that of Katsura
algebras [9], where the unitary associated to an element of Z is written in terms of partial
unitaries associated to the projections px for x ∈ E0.
2.1. The basic data for our construction is a triple (G, E, ϕ) composed of:
(1) A countable directed graph E = (E0, E1, r, s).
(2) A discrete group G acting on E by graph automorphisms.
(3) A 1-cocycle ϕ : G × E1 → G satisfying the property
ϕ(g, a) · x = g · x for every g ∈ G, a ∈ E1, x ∈ E0.
As remarked before, this condition can be weakened to
ϕ(g, a) · s(a) = g · s(a) for every g ∈ G, a ∈ E1,
without pain. Notice that all the results in [8, Section 2] are true for triples (G, E, ϕ) as in
(2.1). So, assume that we have a such triple (G, E, ϕ). To any x ∈ E0 and any g ∈ G we
ug,x converges to a unitary
will associate an element ug,x ∈ OG,E. Since we want that Px∈E 0
1M(OG,E ) = Px∈E 0
ug ∈ M(OG,E) in the strong topology, our idea is to think that ug,x = ugpx (recall that
px). Using the relations enjoyed by ugpx in the original definition, we conclude
that ug,x is a partial isometry with:
(1) ug,xu∗
g,x = pg·x.
4
RUY EXEL, ENRIQUE PARDO, AND CHARLES STARLING
g,xug,x = px.
(2) u∗
(3) ug,s(a)sa = sg·s(a)uϕ(g,a),s(a).
(4) ug,xpx = pg·xug,x.
With this definition, Px∈E 0
ug,x converges to an element ug ∈ M(OG,E), and it is easy to
see that it is a unitary. Since we are interested in getting a unitary representation of G in
M(OG,E), we need to be sure that uguh = ugh for every g, h ∈ G. A simple computation
shows that this occurs whenever
ugh,h−1·x = ug,xuh,h−1·x
for every g, h ∈ G and every x ∈ E0.
definition
In view of all that facts, we obtain the following
Definition 2.2. Given a triple (G, E, ϕ) as in (2.1), we define OG,E to be the universal
C ∗-algebra as follows:
(1) Generators:
{px x ∈ E0} ∪ {sa a ∈ E1} ∪ {ug,x g ∈ G, x ∈ E0}.
(2) Relations:
(a) {px x ∈ E0} ∪ {sa a ∈ E1} is a Cuntz-Krieger E-family in the sense of [11].
(b) ug,x is a partial isometry with:
(i) ug,xu∗
(ii) u∗
g,x = pg·x.
g,xug,x = px.
(c) ugh,h−1·x = ug,xuh,h−1·x for every g, h ∈ G and x ∈ E0.
(d) ug,s(a)sa = sg·auϕ(g,a),s(a) for every g ∈ G, a ∈ E1.
(e) ug,xpx = pg·xug,x for every g ∈ G, x ∈ E0.
Remark 2.3. When E0 is finite, Definition 2.2 coincides with Definition 1.2. Moreover, we
have:
(1) ug := Px∈E 0
ug,x is a unitary in M(OG,E).
(2) The map u : G → M(OG,E) defined by the rule g 7→ ug is a unitary ∗-representation
of G.
(3) ug,x = ugpx for every g ∈ G and x ∈ E0.
(4) ugpx = pgxuϕ(g,a) for every g ∈ G and x ∈ E0.
(5) ugsa = sgauϕ(g,a) for every g ∈ G and a ∈ E1.
Following the same structure, we will need to associate an abstract inverse semigroup to
this algebra. The natural one will be
Definition 2.4. Given a pair (G, E) as in (1.1), we define a ∗-semigroup dSG,E as follows. As
a set,
dSG,E = {(α, (g, x), β) α, β ∈ E∗, g ∈ G, g · x = s(α) = g · s(β)} ∪ {0}.
The operation is defined as in [8, Definition 4.1], and the semilattice of idempotents is
E(dSG,E) = {α, (1, x), α) α ∈ E∗, s(α) = x}.
ARBITRARY SELF-SIMILAR GRAPHS
5
Since x ∈ E0 in the definition must to coincide with s(α) = g · s(β), we conclude that dSG,E
and SG,E are essentially identical as ∗-semigroups, and thus we will switch the notation of
the semigroup to SG,E. So, with small adaptations, [8, Sections 3 & 4] hold. Moreover, there
is an semigroup homomorphism
π :
SG,E
The essential point is the following result
(α, (g, x), β)
−→ OG,E
7→ sαug,xs∗
β
.
Theorem 2.5. Let (G, E, ϕ) be a triple as in (2.1) such that E is a row-finite graph with no
sources. Then:
(1) The semigroup homomorphism π : SG,E → OG,E is a universal tight representation of
SG,E.
(2) OG,E
∼= C ∗
tight(SG,E) ∼= C ∗(G(G,E)
tight ).
Proof. (1) If E is a row-finite graph with no sources, then the proofs of [8, Proposition 6.2]
and [8, Theorem 6.3] works correctly.
(2) Because of (1), we can use the proof of [6, Theorem 2.4] to conclude the desired
(cid:3)
result.
In order to extend the results in [8] to this context, we will need to reduce ourselves to a
situation in which the graph E may be assumed to be row-finite without sources. We will
show that this is possible via a "desingularization" process, inspired in the one developed in
[4] for graph C ∗-algebras.
3. Desingularizing triples
Suppose we have a triple (G, E, ϕ) as in (2.1), and let F denote the desingularized graph
of E obtained in [4]. In this section we will show that we can define an action G y F and a
1-cocycle bϕ : G × F → G extending the ones in the original triple such that OG,E and OG,F
are strong Morita equivalent C ∗-algebras.
Remark 3.1. Let (G, E, ϕ) be a triple as in (2.1), and let x ∈ E0 be a vertex. Then:
(1) If x is a source, then so is gx for every g ∈ G.
(2) If x is an infinite receiver, then so is gx for every g ∈ G.
In view of that, when defining the desingularization we need to keep track of the fact that,
for any vertex in the orbit o a singular vertex, we must define the tails added to it as orbit-
connected parts of the graph.
3.1. Desingularizing a source. Now, we will explain how to construct a triple (G, F, ϕ)
that desingularize a source in a triple (G, E, ϕ).
Let x ∈ E0 be a source. Then:
(1) We define a tail {ei}i≥1 ⊂ F 1 so that s(ei) = r(ei+1) for every i ≥ 1, and r(e1) = x.
(2) For any h ∈ StG(x), we define hei = ei for every i ≥ 1.
(3) If bG is a set of representatives of the orbits of x under the action G y E, then for
each g ∈ bG we define {ei,g}i≥1 ⊂ F 1 so that s(ei,g) = r(ei+1,g) for every i ≥ 1, and
r(e1,g) = gx, while gei = ei,g for every i ≥ 1.
6
RUY EXEL, ENRIQUE PARDO, AND CHARLES STARLING
What we are doing is applying the Drinen-Tomforde desingularization construction in such a
way that is coherent with the action G y E.
Next step is to extend the 1-cocycle. To do this, for any g, h ∈ G and any i ≥ 1 we define
bϕ(h, ei,g) := h. Once this is done, what we have obtained is:
(4) A graph F extending E, constructed using the Drinen-Tomforde desingularization
process.
(5) An action G y F extending the original action G y E, taking care of Remark 3.1.
(6) A 1-cocycle bϕ : G × F → G extending ϕ : G × E → G.
Then, the triple (G, F, bϕ) is the desingularization of (G, E, ϕ) on to the source x.
3.2. Desingularizing an infinite receiver. Now, we will explain how to construct a triple
(G, F, ϕ) that desingularize an infinite receiver in a triple (G, E, ϕ).
Let x ∈ E0 be an infinite receiver, and list r−1(x) = {ai}i≥1. Then:
(1) We define a tail {e1}i≥1 ⊂ F 1 so that s(ei) = r(ei+1) for every i ≥ 1, and r(e1) = x.
(2) We define v0 = x and for any i ≥ 1, vi = s(ei).
(3) For each aj ∈ r−1(x), we define an edge fj ∈ F 1 from s(aj) to vj−1 = s(ej).
(4) We remove the edges {ai}i≥1 from F .
(5) For each j ≥ 1 we define paths αj := e1e2 . . . ej−1fj.
(6) For any h ∈ StG(x), we define hei = ei and hfi = fi for every i ≥ 1.
(7) If bG is a set of representatives of the orbits of x under the action G y E, then for each
g ∈ bG we define {ei,g}i≥1 ∪ {fi,g}i≥1 ⊂ F 1 so that s(ei,g) = r(ei+1,g), r(fi,g) = gr(ei)
and s(fi,g) = gs(ai) for every i ≥ 1, r(e1,g) = gx, while gei = ei,g and gfi = fi,g for
every i ≥ 1.
What we are doing is applying the Drinen-Tomforde desingularization construction in such a
way that is coherent with the action G y E.
Next step is to extend the 1-cocycle. What we do is, to any g, h ∈ G and any i ≥ 1, we
define:
(7) bϕ(h, ei,g) := h.
(8) bϕ(h, fi,g) = ϕ(h, aj).
In particular, bϕ(h, αi,g) = ϕ(h, aj). Once this is done, what we have obtained is:
(9) A graph F extending E, constructed using the Drinen-Tomforde desingularization
process.
(10) An action G y F extending the original action G y E, taking care of Remark 3.1.
(11) A 1-cocycle bϕ : G × F → G extending ϕ : G × E → G.
Then, the triple (G, F, bϕ) is the desingularization of (G, E, ϕ) on to the infinite receiver x.
3.3. The desingularization result. Now, we will check that the results of Drinen and
Tomforde about their desingularization process for C ∗(E) extend to this context.
A simple inspection shows that [4, Lemmas 2.9 & 2.10] extend to our context. Now, we
will arrange the proof of [4, Theorem 2.11] in order to obtain the desired Morita equivalence
between OG,E and OG,F . We will follow the notation of the proof of [4, Theorem 2.11]. Let
E be a graph with a singular vertex v0, let
{te, qv e ∈ F 1, v ∈ F 0}
ARBITRARY SELF-SIMILAR GRAPHS
7
be the canonical set of generators for C ∗(F ), and let
{se, pv e ∈ E1, v ∈ E0}
be the Cuntz-Krieger E-family constructed into C ∗(F ) in [4, Lemma 2.9]. Recall that {se, pv}
is defined as follows:
(1) For every v ∈ E0, pv := qv.
(2) For every e ∈ E1 such that r(e) ∈ E0
(3) For every e ∈ E1 such that r(e) 6∈ E0
rg, se := te.
rg, we have that e = aj for some j ≥ 1, and thus
se := tαj .
Define B := C ∗({se, pv}) and p := Pv∈E 0
qv ∈ M(C ∗(F )). Then, [4, Theorem 2.11] shows
that
C ∗(E) ∼= B ∼= pC ∗(F )p
and that p ∈ M(C ∗(F )) is a full projection.
Now, observe that:
(1) For every g ∈ G and x ∈ E0 we have that ugpx = ugqx = qgxug = pgxug.
(2) If r(e) ∈ E0
rg, then ugse = ugte = tgeu bϕ(g,e) = sgeuϕ(g,e), since ϕ and bϕ matches on
G × E.
(3) If r(e) 6∈ E0
rg, then
ugse = ugtαj = ugte1te2 · · · tej−1tfj = tge1tge2 · · · tgej−1ugtfj =
tge1tge2 · · · tgej−1tgfj u bϕ(g,fj ) = tgαj u bϕ(g,fj) = sgeuϕ(g,e)
by definition of bϕ in this case.
Thus, the C ∗-algebra isomorphism
Φ : C ∗(E) → C ∗({se, qx})
Px
Te
7→
7→
qx
se
satisfies that:
(1) For every g ∈ G and x ∈ E0, ugΦ(Px) = Φ(Pgx)ug.
(2) For every g ∈ G and e ∈ E1, ugΦ(Te) = Φ(Tge)u bϕ(g,e).
Hence, by the universal property of OG,E, P hi extends to a C ∗-algebra isomorphism
Moreover, following [4, Theorem 2.11], we have a C ∗-isomorphism
Φ : OG,E → C ∗({ug, se, qx}) ⊆ OG,F .
Ψ : Φ(C ∗(E)) → pC ∗(F )p
βp
7→ psαs∗
sαs∗
β
.
Notice that in pC ∗(F )p, for every g ∈ G and x ∈ E0 we have that pugpxp = ppgxugp,
and since ugp = pug by definition of p, this is equal to ppgxpug. Since a homomorphism
from OG,E to pOG,F p extending Ψ should send ugpx 7→ pugpxp and pgxug 7→ ppgxpug, a
such extension will be compatible with the defining relations of OG,E. Similarly for every
g ∈ G and a ∈ E1 we have that ugsa and sgau bϕ(g,a) will map to the same element in pOG,F p.
So, again by the universal property of OG,E, the isomorphism Ψ extends to a C ∗-algebra
8
RUY EXEL, ENRIQUE PARDO, AND CHARLES STARLING
isomorphism OG,E
full projection.
Summarizing
∼= pOG,F p. Also, as in [4, Theorem 2.11], p ∈ M(C ∗(F )) ⊂ M(OG,F ) is a
Theorem 3.2. Let (G, E, ϕ) be a triple, and let (G, F, bϕ) be its desingularization. Then,
∼= pOG,F p. In particular, OG,E
there exists a full projection p ∈ M(OG,F ) such that OG,E
and OG,F are strongly Morita equivalent.
Hence, up to Morita equivalence, we can assume that E is a countable, row-finite graph
with no sources, and thus is a full groupoid C ∗-algebra by Theorem 2.5.
4. Characterizing properties of OG,E
Now, we are ready to extend the characterizations of various properties of G(G,E)
tight
(and
thus, the simplicity of OG,E) obtained in [8] to the case of triples (G, E, ϕ) with E countable
arbitrary graph; in this sense, recall that properties like Conditions (L) and (K), or cofinality,
are preserved through the desingularization process, as shown in [4].
Since several arguments in [8] uses the fact that bEtight(SG,E) ∼= E∞ when E is a finite
graph without sources, we need to prove this fact in the case of E being infinite. Because
of Theorem 3.2 and the previous remark, we can assume without loss of generality that E is
row-finite without sources. So, we will keep that fact in force for the remain of the section.
4.1. A technical issue. Suppose that (G, E, ϕ) with E row-finite graph without sources.
Then, it is easy to see that bE∞(SG,E) ∼= E∞. But since E is infinite, the space of filters bE0 is
locally compact but not compact, whence we cannot guarantee that bE∞(SG,E) = bEtight(SG,E);
the only we know is that bE∞(SG,E) is a dense subspace of bEtight(SG,E).
Let us recall the characterization of ultrafilters given in [5, Lemma 12.3]:
"A filter ξ is an ultrafilter if and only if for f ∈ E(SG,E), if f ⋓ e for every e ∈ ξ, then f ∈ ξ".
Also, we need to recall the characterization of tight filters given in [5, Theorem 12.9]:
"A filter ξ is a tight filter if and only if for every X, Y ⊂ E(SG,E) finite subsets and for every
Z ⊂ E(SG,E)X,Y finite cover one has that X ⊂ ξ and Y ∩ ξ = ∅ implies that Z ∩ ξ 6= ∅".
Given any ω ∈ E∞ and any n ∈ N, we denote fωn = (ωn, (1, s(ωn)), ωn) ∈ E(SG,E), and
we define Fω := {fωn n ∈ N}. Then, the map
τ : E∞ → bE∞(SG,E)
Fω
ω
7→
is a homeomorphism. In particular, if ξ ∈ bE∞(SG,E), then ξ is an infinite set of idempotents.
Now, suppose that ξ ∈ bEtight(SG,E) \ bE∞(SG,E). We have two options:
(1) If ξ is an infinite set, for each n ∈ N define ξn := {e ∈ ξ e = (α, (1, s(α)), α) with α =
n}. Given e, f ∈ ξn ⊂ ξ, 0 6= ef because ξ is a filter, and then e = f . Thus, for any
n ∈ N we have that ξn ≤ 1.
Now, for n ≤ m natural numbers, suppose that ξn = {(α, (1, s(α)), α)} and ξm =
{(β, (1, s(β)), β)}. Again for the fact that ξ is a filter, we conclude that α is the prefix
βn of β. In particular, if for some m ∈ N we have ξm = {(β, (1, s(β)), β)}, then for
ARBITRARY SELF-SIMILAR GRAPHS
9
any n ≤ m we have ξm = {(βn, (1, s(βn)), βn)}. Thus, for any n ∈ N we have that
ξn = 1.
So, we can construct ω ∈ E∞ such that, for each n ∈ N, ξn = {fωn}. Hence, Fω ⊆ ξ,
and since Fω is maximal, we conclude that Fω = ξ, contradicting the assumption.
(2) If ξ is a finite set, and ξ = n, the same argument as in case (1) states that the
idempotents in ξ are associated to prefixes of a fixed finite path α. Now, take any
X ⊆ ξ and any finite subset Y ⊆ E(SG,E) \ ξ. Since E has no sources, be can chose
a finite path β of positive length such that f := (αβ, (1, s(β)), αβ) 6∈ Y . Thus,
Z := {f } ⊂ E(SG,E)X,Y = \
x∈X
Jx ∩ \
y∈Y
J ⊥
y .
But f ⋓ e for every e ∈ ξ, and since ξ is a tight filter, we conclude that f ∈ ξ,
contradicting the choice of f .
Summarizing
Proposition 4.1. If (G, E, ϕ) is a triple with E row-finite graph without sources, then
bEtight(SG,E) = bE∞(SG,E).
4.2. The properties. Finally, we are ready to obtain the desired characterizations. Notice
that, since the final aim is to characterize simplicity of OG,E, and this property is Morita
invariant, using Theorem 3.2 we can reduce the problem to the case E is a row-finite graph
without sources. Moreover, under this restriction Proposition 4.1 holds. So, we can use the
arguments in [8] involving actions SG,E y E∞ in this context. Thus, we can look at the
results in [8, Sections 12-15] and fix the hypotheses to make work them in this context.
First, with respect to Hausdorffness of G(G,E)
tight , we have the following result.
Theorem 4.2 (c.f. [8, Theorem 12.2]). Let (G, E, ϕ) a triple with E being a row-finite graph
without sources. Then, the following are equivalent:
(1) For every g ∈ G and for every x ∈ E0 there exists a finite number of minimal strongly
fixed paths for g with range x.
(2) G(G,E)
tight
is Hausdorff.
Proof. It is the same proof as this of [8, Theorem 12.2].
(cid:3)
Next, we characterize minimality, as follows.
Theorem 4.3 (c.f. [8, Theorem 13.6]). Let (G, E, ϕ) a triple with E being a row-finite graph
without sources. Then, the following are equivalent:
(1) The standard action G y E∞ is irreducible.
(2) G(G,E)
(3) E is weakly G-transitive.
is minimal.
tight
Proof. It is the same proof as this of [8, Theorem 13.6].
(cid:3)
Finally, we characterize the groupoid being essentially principal, as follows.
Theorem 4.4 (c.f. [8, Theorem 14.10]). Let (G, E, ϕ) a triple with E being a row-finite graph
without sources. Then, the standard action G y E∞ is topologically free (equivalently, G(G,E)
is essentially principal) if and only if the following two conditions hold:
tight
10
RUY EXEL, ENRIQUE PARDO, AND CHARLES STARLING
(1) Every G-circuit has an entry.
(2) Given a vertex x ∈ E0 and a group element g ∈ G, if g fixes Z(x) pointwise then g is
slack at x.
Proof. It is the same proof as this of [8, Theorem 14.10].
(cid:3)
Hence, we conclude the following characterization of simplicity.
Theorem 4.5 (c.f.
let (G, F, bϕ) be its desingularization. If G is amenable and G(G,F )
simple if and only if the following conditions are satisfied:
tight
[8, Theorem 16.1]). Let (G, E, ϕ) a triple with E countable graph, and
is Hausdorff, then OG,E is
(1) F is weakly G-transitive.
(2) Every G-circuit in F has an entry.
(3) Given a vertex x ∈ F 0 and a group element g ∈ G, if g fixes ZF (x) poinwise then g is
slack at x.
Proof. This is because of Theorem 3.2, Theorem 4.2, Theorem 4.3, Theorem 4.4 and [3,
Theorem 5.1].
(cid:3)
Unfortunately, when E is an infinite graph, [8, Theorem 15.1] is false, because the impli-
cation (iv) ⇒ (i) do not work. Fortunately, there is a condition, generalizing [8, Theorem
15.1(iv)], which allows to show an analog result.
[8, Theorem 16.1]). Let (G, E, ϕ) a triple with E countable, row finite
Theorem 4.6 (c.f.
graph with no sinks. Then, the following are equivalent:
(1) SG,E is a locally contracting inverse semigroup.
(2) The standard action θ : G y E∞ is locally contracting.
(3) G(G,E)
(4) For every x ∈ E0 there exists αx ∈ E∗ with r(αx) = x and there exists a G-circuit
is a locally contracting groupoid.
tight
(g, γ) with entries such that s(αx) = r(γ).
Proof. (i) ⇔ (ii) By Proposition 4.1 and [7, Theorem 6.5].
(ii) ⇒ (iii) It follows immediately by [7, Proposition 6.3].
(iii) ⇒ (iv) First, suppose that (g, γ) is a G-circuit with no entries. Then, ω := γ1γ2 · · · ∈
E∞ is an isolated point in E∞. Thus,
∅ 6= U = {ω} ⊂ E∞
is an open set, and there are no nonempty open set V ⊆ U and an open bisection S ⊆ G(G,E)
tight
such that V ⊆ S−1S and SV S−1 ( V . So, every G-circuit in E must to have an entry.
Next, suppose that x ∈ E0 do not connect with any G-circuit. Consider the subtree H of
E with root x. By assumption, it do not contain any G-circuit (in particular, any circuit).
Now, consider bH the subgraph of E generated by H under the action of G (on E), and notice
that bH cannot contain any G-circuit. Moreover, for every α ∈ bH ∗ and for every g ∈ G, we
have gα ∈ bH ∗. Thus, (G, bH, ϕ bH) is a subtriple of (G, E, ϕ), and the inclusion ι : bH ֒→ E
induces a natural inclusion bι : G(G, bH)
tight of topological groupoids. Now, given any
nonempty open subset U ⊆ G(G,E)
, there exists α ∈ E∗ such that ∅ 6= Z(α) ⊆ U. Hence,
tight
the subtree rooted on s(α) is acyclic, and since the action of G do not generate G-circuits
֒→ G(G,E)
tight
(0)
ARBITRARY SELF-SIMILAR GRAPHS
11
is
there is no open bisection S ⊆ G(G,E)
tight
not locally contracting, and then so does G(G,E)
such that V ⊆ S−1S and SV S−1 ( V . Thus, G(G, bH)
tight
tight , contradicting the assumption.
(iv) ⇒ (i) Let 0 6= e ∈ E(SG,E), written e = (µ, (1, s(µ)), µ) for some µ ∈ E∗. By
hypothesis, there exist α ∈ E∗ with r(α) = s(µ) and a G-circuit (g, γ) with entry τ such that
s(α) = r(γ). If ω := γ1γ2 · · · ∈ E∞, there exists k ∈ N such that s(γk−1) = r(τ ) = r(γk) and
τ 6= γk. Define bγ := γ1γ2 · · · γk, β := µα and bβ := µαbγ. These are well-defined finite paths
in E, and moreover s(α) = r(bγ) = gk+1s(bγ). Hence, s := (bβ, (g−1
f1 := (β, (1, s(α)), β), and notice that
k+1, s(α)), β) ∈ SG,E. Take
sf1s∗ = (bβ, (g−1
k+1, s(α)), β)(β, (1, s(α)), β)(β, (g−1
k+1, s(α)), bβ) = (bβ, (1, s(bγ), bβ) ≤ f1.
Also, s∗s = (β, (1, s(α)), β) = f1 ≤ e, so that f1 ≤ s∗s ≤ e = e · s∗s. Now, define
γ′ := γ1γ2 · · · γk−1τ 6= bγ, and f0 := (µαγ′, (1, s(τ )), µαγ′). Hence, f0 ≤ f1 and f0 · s =
(µαγ′, (1, s(τ )), µαγ′)(bβ, (g−1
k+1, s(α)), β) = 0 (because γ′ 6= bγ), whence f0sf1 = 0. Thus, SG,E
is locally contracting by [7, Proposition 6.7].
(cid:3)
In particular, if CG is minimal, effective, and E contains at least one G-circuit, then G(G,E)
tight
is locally contracting. Hence,
Corollary 4.7 (c.f. [8, Corollary 16.3]). Let (G, E, ϕ) a triple with E countable graph, and
is Hausdorff, then whenever
let (G, F, bϕ) be its desingularization. If G is amenable and G(G,F )
OG,E is simple, it is necessarily also purely infinite (simple).
tight
Acknowledgments
Part of this work was done during a visit of the second author to the Departamento de
Matem´atica da Universidade Federal de Santa Catarina (Florian´opolis, Brasil). This author
thanks the host center for its warm hospitality.
References
[1] C. Anantharaman-Delaroche, Purely infinite C ∗-algebras arising form dynamical systems, Bull.
Soc. Math. France 125 (1997), no. 2, 199–225.
[2] E. B´edos, S. Kaliszewski, J. Quigg, On Exel-Pardo algebras, J. Operator Th. 78(2) (2017), 309–345.
[3] J. Brown, L.O. Clark, C. Farthing, A. Sims, Simplicity of algebras associated to ´etale groupoids,
Semigroup Forum 88 (2014), 433–452.
[4] D. Drinen, M. Tomforde, The C ∗-algebras of arbitrary graphs, Rocky Mount. J. Math. 35 (2005),
105–135.
[5] R. Exel, Inverse semigroups amb combinatorial C ∗-algebras, Bull. Braz. Math. Soc. (N.S.) 39 (2008),
191-313.
[6] R. Exel, Reconstructing a totally disconnected groupoid from its ample semigroup, Proc. Amer. Math.
Soc. 138 (2010), 2991–3001.
[7] R. Exel, E. Pardo, The tight groupoid of an inverse semigroup, Semigroup Forum 92 (2016), 274 –
303.
[8] R. Exel, E. Pardo, Self-similar graphs, a unified treatment of Katsura and Nekrashevych C*-algebras,
Adv. Math. 306 (2017), 1046 – 1129.
[9] T. Katsura, A construction of actions on Kirchberg algebras which induce given actions on their K-
groups, J. reine angew. Math. 617 (2008), 27–65.
[10] M. Laca, I. Raeburn, J. Ramagge, M. Whittaker, Equilibrium states on operator algebras asso-
ciated to self-similar actions of groupoids on graphs, Adv. Math. 331 (2018), 268–325.
12
RUY EXEL, ENRIQUE PARDO, AND CHARLES STARLING
[11] I. Raeburn, "Graph Algebras", CBMS Reg. Conf. Ser. Math., vol. 103, Amer. Math. Soc., Providence,
RI, 2005.
Ruy Exel, Departamento de Matem´atica, Universidade Federal de Santa Catarina, 88040-
970 Florian´opolis SC, Brazil
E-mail address: [email protected]
URL: http://www.mtm.ufsc.br/~exel/
Enrique Pardo, Departamento de Matem´aticas, Facultad de Ciencias, Universidad de C´adiz,
Campus de Puerto Real, 11510 Puerto Real (C´adiz), Spain.
E-mail address: [email protected]
URL: https://sites.google.com/a/gm.uca.es/enrique-pardo-s-home-page/
Charles Starling, Carleton University, School of Mathematics and Statistics, 4302 Herzberg
Laboratories, 1125 Colonel By Drive, Ottawa, ON, Canada, K1S 5B6.
E-mail address: [email protected]
URL: https://carleton.ca/math/people/charles-starling/
|
1906.07397 | 2 | 1906 | 2019-09-11T02:09:24 | Infinite families of potential modular data related to quadratic categories | [
"math.OA",
"math.QA"
] | We present several infinite families of potential modular data motivated by examples of Drinfeld centers of quadratic categories. In each case, the input is a pair of involutive metric groups with Gauss sums differing by a sign, along with some conditions on the fixed points of the involutions and the relative sizes of the groups. From this input we construct $S$ and $T$ matrices which satisfy the modular relations and whose Verlinde coefficients are non-negative integers. We also check certain restrictions coming from Frobenius-Schur indicators.
These families generalize Evans and Gannon's conjectures for the modular data associated to generalized Haagerup and near-group categories for odd groups, and include the modular data of the Drinfeld centers of almost all known quadratic categories. In addition to the subfamilies which are conjecturally realized by centers of quadratic categories, these families include many examples of potential modular data which do not correspond to known types of modular tensor categories. | math.OA | math |
INFINITE FAMILIES OF POTENTIAL MODULAR DATA
RELATED TO QUADRATIC CATEGORIES
PINHAS GROSSMAN AND MASAKI IZUMI
Abstract. We present several infinite families of potential modular data moti-
vated by examples of Drinfeld centers of quadratic categories. In each case, the
input is a pair of involutive metric groups with Gauss sums differing by a sign,
along with some conditions on the fixed points of the involutions and the relative
sizes of the groups. From this input we construct S and T matrices which satisfy
the modular relations and whose Verlinde coefficients are non-negative integers. We
also check certain restrictions coming from Frobenius-Schur indicators.
These families generalize Evans and Gannon's conjectures for the modular data
associated to generalized Haagerup and near-group categories for odd groups, and
include the modular data of the Drinfeld centers of almost all known quadratic cat-
egories. In addition to the subfamilies conjecturally realized by centers of quadratic
categories, these families include many examples of potential modular data which
do not correspond to known types of modular tensor categories.
1. Introduction
In this paper we construct several infinite families of potential modular data, mo-
tivated by examples of quadratic categories which appeared in the classification of
small-index subfactors. Modular data are numerical invariants of modular tensor cat-
egories, expressed as a pair of matrices S and T which are the images of the canonical
generators
(cid:18) 0 −1
0 (cid:19) and (cid:18) 1 1
0 1 (cid:19)
1
in a projective unitary representation of the modular group SL2(Z) which is associ-
ated to the category. The matrix S is symmetric, the matrix T is diagonal with finite
order, and S2 is a permutation matrix commuting with T . The S and T matrices are
indexed by the simple objects of the category, with the Verlinde coefficients
(1.1)
giving the fusion rules
N k
ij =Xr
Si,rSj,rSk,r
S0,r
(see [Ver88, MS90, BK01]).
dim(Hom(Xi ⊗ Xj, Xk)) = N k
ij
2010 Mathematics Subject Classification. Primary 46L37; Secondary 18D10.
Key words and phrases. modular data, subfactors, fusion categories.
Supported in part by JSPS KAKENHI Grant Number JP15H03623 and ARC grant DP170103265.
1
2
PINHAS GROSSMAN AND MASAKI IZUMI
Modular data has proven to be a useful invariant of modular tensor categories,
with the first examples of distinct modular categories sharing the same modular
data discovered only recently [MS17]. It is therefore natural to consider construc-
tion/classification of modular data as a first step towards construction/classification
of modular tensor categories. The properties enjoyed by modular data - in particu-
lar non-negative integrality of the Verlinde coefficients - place severe restrictions on
the matrices involved, and the modular data in turn encode a wealth of information
about any categories which may realize them.
Modular tensor categories first appeared in conformal field theory [MS90], and play
an important role in quantum topology [Tur92]. The category of representations of
a rational vertex operator algebra is a modular tensor category [Hua05], and it is
a major open problem whether every modular tensor category can be realized this
way. Modular tensor categories also arise as categories of representations of quantum
groups at roots of unity, and it is also a major open problem whether such categories
generate the Witt group of unitary modular tensor categories [DMNO13].
These two problems reflect the paucity of known examples and constructions of
modular tensor categories. Besides quantum groups, one source of examples is the
Drinfeld center construction. If C is a spherical fusion category, its Drinfeld center
Z(C) is a modular tensor category (by definition belonging to the trivial Witt class).
A number of examples of fusion categories not arising from representations of groups
or quantum groups have been discovered through the classification of small-index sub-
factors (see [Haa94, JMS14, AMP15]). With one exception (the Extended Haagerup
subfactor [BPMS12]), these examples are all related to quadratic categories.
A quadratic category is a fusion category C whose set of simple objects has a unique
non-trivial orbit under the tensor product action of the group of invertible objects
Inv(C). (Such categories are also sometimes called generalized near-group categories
[Tho12].) Given a quadratic category, one can consider the stabilizer subgroup of
Inv(C) for a given non-invertible simple object X. If this stabilizer subgroup is all of
Inv(C), then X is the only non-invertible simple object, and the category is called a
near-group category [Sie03]. At the other extreme, if the stabilizer subgroup is trivial,
then the number of non-invertible simple objects is equal to Inv(C). Examples of
such categories are the generalized Haagerup categories introduced in [Izu01].
A general method for constructing various types of quadratic categories from
endomorphisms of Cuntz algebras was developed by the second named author in
[Izu93, Izu01, Izu17, Izu18]. This construction was used to describe the tube alge-
bra and thereby compute the modular data of the Drinfeld center for a number of
examples, including the Haagerup subfactor [Izu01].
Evans and Gannon observed that the modular data for the Drinfeld center of the
Haagerup categories, as well for the centers of other generalized Haagerup categories
associated to odd groups, can be viewed as composed of two distinct blocks "grafted"
together [EG11]. They gave a definition of grafting, noting that the concept could
be "massively generalized." In a subsequent paper, they considered modular data for
near-group categories, and conjectured a remarkably simple formula for the modular
INFINITE FAMILIES OF MODULAR DATA
3
data for a certain type of near-group category when the associated group is odd
[EG14].
Their formula for odd near-group modular data is expressed entirely in terms a
pair of metric groups. A metric group is a finite abelian group equipped with a
non-degenerate quadratic form. The category of metric groups is equivalent to the
category of non-degenerate braided pointed fusion categories [EGNO15, Section 8.4].
By [Izu17, Theorem 12.9], generalized Haagerup categories for odd groups are de-
equivariantizations of near-group categories, and their modular data could also be
determined from this fornula (if the conjecture is correct).
In this paper, we study a number of generalizations of the Evans-Gannon conjec-
tures motivated by examples of generalized Haagerup categories for even groups and
their (de)-equivariantizations, which are discussed in the accompanying paper [GI19].
A new ingredient is the introduction of involutions on the metric groups. For even
groups, one natural involution is the inverse operation, whose fixed points form the
subgroup of order two elements; but other involutions are of interest as well.
The families of potential modular data which we construct from the pairs of met-
ric groups are infinite, but we do not address the difficult question of whether such
modular data actually arise from modular tensor categories, beyond known exam-
ples. For certain choices of input data (i.e. particular types of groups equipped with
particular quadratic forms), we conjecture that the resulting modular data are real-
ized by Drinfeld centers of infinite families of quadratic fusion categories. But it is
not known whether the corresponding infinite families of quadratic categories exist.
There are also many other choices of input data which do not correspond to any
known modular tensor categories.
The paper is organized as follows. After this introduction and a section of pre-
liminaries on modular data and metric groups, there are five chapters, each dealing
with a different family of potential modular data. In each case, the starting point is
a pair of (involutive) metric groups with Gauss sums differing by a sign, and some
conditions on the involutive fixed points. From this input, potential modular data
are constructed and the modular relations are checked. Then the Verlinde coefficients
are computed, and non-negative integrality gives constraints on the difference in size
of the groups. We also check consistency with certain conditions for Frobenius-Schur
indicators of modular data, which in some cases give further restrictions on the ad-
missible groups and quadratic forms.
Then we consider examples of this potential modular data for small groups and
various quadratic forms and compare to known examples.
In some cases, we can
formulate conjectures for the realization of the modular data based on the known
examples.
Briefly, the five families are as follows.
(1) The first family (Section 3) generalizes the Evans-Gannon odd near-group
conjecture to an arbitrary finite abelian group. Here the input is a pair of
involutive metric groups G and Γ which are assumed to have isomorphic
involutive fixed point subgroups. For the Drinfeld center of a near group
category with A = Inv(C) and multiplicity A, G is conjectured to be A × A,
4
PINHAS GROSSMAN AND MASAKI IZUMI
where the involution is the flip, with Γ having order A(A + 4). But there
are potentially other examples where G is not of the form A × A but rather
some other group; such examples would not necessarily correspond to centers
of anything (this phenomenon is already present in the odd case).
(2) For the second family (Section 4), the involution on each metric group is the
inverse map. It is assumed that the fixed point subgroups (consisting of the
order two elements) are both Z2, and there are a couple of other assumptions.
Here the integrality of coefficients requires that Γ = G + 2. We conjecture
that examples of modular categories realizing this family are given by Drinfeld
centers of Z2-de-equivariantizations of near-group categories for Z2n+1 × A
(with one known instance).
(3) For the third family (Section 5), it is assumed that the involutive fixed point
subgroups both have order 4, but that the quadratic forms only agree on a
common order 2 subgroup (and some other assumptions). Here the integrality
of coefficients requires that Γ = G + 4. There are several cases to consider
here, according to whether the nontrivial invertible object is a boson or a
fermion, and some other conditions. We conjecture that examples of mod-
ular tensor categories realizing this family are given by Drinfeld centers of
generalized Haagerup categories for A with A2 = 2, and G = A × A (with
two known instances, and numerical evidence for other instances); or more
generally by Drinfeld centers of quadratic categories with two non-invertible
simple objects which are stabilized by an odd group of invertible objects.
(4) For the fourth family (Section 6), the group of order two elements of G has
size 4, and Γ = G + 1. This family unifies the modular data of the Drin-
feld centers of the Asaeda-Haagerup categories (which come from a Z2-de-
equivariantization of a generalized Haagerup category for Z4×Z2 [GIS18]), the
generalized Haagerup category for Z2 × Z2, and the Z2-de-equivariantizations
of generalized Haagerup categories for Z4 and Z8. These four types of mod-
ular data correspond to different forms of the even part of G and different
quadratic forms, and can all be generalized to infinite subfamilies.
(5) The fifth family (Section 7) comes from the fourth family by switching the
roles of G and Γ. We conjecture that examples of this family can be realized
by Drinfeld centers of near-group categories with multiplicity twice the order
of the group.
In [BGH+17], a construction called zesting is introduced whereby a spin-modular
category (a modular category containing a fermion) can be twisted to give 8 different
modular closures of its supermodular subcategory (out of a total of 16). A formula
for the zested modular data is given there.
For a group of the form Z2×A, with A odd, we have conjectural modular data for
the centers of both generalized Haagerup categories and near-group categories which
contain fermions. In both cases the metric group G used to construct the modular
data has even part Z2 × Z2. It turns out that in each case, the eight zested modular
data belong to the same family where Z2 × Z2 is possibly replaced by Z4 and the
quadratic forms vary (see Remarks 3.13 and 5.10.)
INFINITE FAMILIES OF MODULAR DATA
5
The verification of the modular relations and the computations of the Verlinde
coefficients and Frobenius-Schur indicators for the various families are lengthy and
tedious, so we defer them to an online appendix, which is included in arxiv version
of the paper. There is a separate section of the appendix for the calculations in
each of Sections 3-6 of the paper, which correspond to the first four families (the
fifth family is related to the fourth family so new calculations are not needed.) The
appendix also contains a computation of modular data for Drinfeld centers of Z2-de-
equivariantizations of Z4-near-group categories, verifying an instance of Conjecture
4.5.
We also include with the arxiv source the Mathematica notebook ifpmd.nb which
constructs the modular data for one example from each family, verifies the modular
relations, and computes the Verlinde coefficients and Frobenius-Schur indicators. The
notebook is annotated so that the reader can modify the input (which consists of
specifying the pair of involutive metric groups and some associated data, depending
on the family) to look at modular data for other examples.
Acknowledgements. We would like to thank the American Institute of Mathe-
matics for its generous hospitality during the SQuaRE Classifying Fusion Categories
and the Isaac Newton Institute for its generous hospitality during the semester pro-
gram Operator algebras: Subfactors and their Applications. We would like to thank
Eric Rowell for pointing out that all sixteen rank 10 modular data related to the
center of the near-group category for Z2 with multiplicity 2 can be realized though
zesting and complex conjugation (see Section 3.2.1), and we would like to thank Mar-
cel Bischoff for first pointing out the realization of the c = −1 case. We would like to
thank Terry Gannon for showing us his further generalizations of grafting in [Gan18].
We would like to thank Andrew Schopieray for helpful comments on the preprint.
2.1. Modular data. Throughout the paper, we use the notation
2. Preliminaries
T = {z ∈ C; z = 1}.
2πi
n .
For n ∈ N, we use the notation ζn = e
Let C be a unitary modular tensor category with J the set of isomorphism classes
of simple objects in C (for the definition and basic properties of modular tensor
categories, see [BK01] or [M03]). The modular data of (S, T ) of C are a pair of
J by J matrices. Since there are several different normalizations for (S, T ) in the
literature (see [BGH+17] and [Gan05] for example), we first fix ours. We assume that
S is a symmetric unitary matrix satisfying S0,j > 0 for every j ∈ J and S2 = C
with Cij = δi,j, and T is a unitary diagonal matrix of finite order with T0,0 = 1 and
CT = T C, where j is the dual object of j ∈ J, and 0 is the unit object. We are
mainly concerned with the following known constraints for the modular data (S, T )
(see [BGH+17], [Gan05], and [NS07] for example).
6
PINHAS GROSSMAN AND MASAKI IZUMI
1
(1) (ST )3 = cC, where c ∈ T is given by
S00 Xi∈J
Nijk =Xa∈J
(2) For i, j, k ∈ J, let
c =
S2
0iTii.
SaiSajSa,k
.
S0,a
Then N k
is a non-negative integer.
ij = Nijk is the fusion coefficient dim Hom(k, i⊗ j). In particular Nijk
(3) For k ∈ J and n ∈ N, let
νn(k) = Xi,j∈J
N k
ijS0iS0j(
Tjj
Tii
)n.
Then νn(k) is the n-th Frobenius-Schur indicator of k.
In particular,
(FS2) We have ν2(k) = 0 if k 6= k, and ν2(k) ∈ {1,−1} if k = k.
(FS3) The third indicator ν3(k) is a sum of Nkkk numbers belonging to {1, ζ3, ζ 2
3}.
In this paper we propose a number of potential modular data: pairs of unitary
matrices (S, T ) satisfying the above conditions, without knowing whether the under-
lying modular tensor categories exist. More precisely, we take the above formulae (2)
and (3) as the definitions of Nijk and νn(k) for a given (S, T ) satisfying (1), and check
whether these numbers satisfy the conditions in (2) and (3). Although there are other
number theoretic constraints known for (S, T ), we do not discuss them as they suit
for case-by-case analysis and it is relatively easy to apply them to our examples.
2.2. Metric groups. Every finite abelian group G is canonically decomposed as
G = Ge × Go, where Ge is a 2-group and Go is an odd group. We set
G2 = {g ∈ G; 2g = 0},
which is a subgroup of Ge.
For a quadratic form q on a finite abelian group G, we adopt multiplicative no-
tation. Namely, a quadratic form q on G is a map q : G → T = {z ∈ C; z = 1}
satisfying q(−g) = q(g) for all g ∈ G such that the map h·,·iq : G × G → T defined
by
hg, hiq = q(g + h)q(g)−1q(h)−1
is a bicharacter. We say the q is non-degenerate if the associated bicharacter is
non-degenerate. We often suppress q in h·,·iq if there is no possibility of confusion.
We define the Gauss sum for non-degenerate q and k ∈ Z by
G(q, k) =
q(g)k ∈ C,
1
pGXg∈G
and we denote G(q) = G(q, 1) ∈ T. For a given non-degenerate symmetric bicharacter
h·,·i, we can always find a quadratic form q giving h·,·i as above, and if moreover G
INFINITE FAMILIES OF MODULAR DATA
7
is an odd group, it is unique and of the form q(g) = hg, gi1/2. For the classification
of non-degenerate quadratic forms, the reader is referred to [BJ15]. We often use
the following facts. We have G(q) ∈ {1,−1} if G ≡ 1 mod 4 and G(q) ∈ {i,−i} if
G ≡ 3 mod 4. For odd r and integer n ≥ 1, we have
2n+1 = (−1)n r2−1
ζ rx2
8 ζ r
8.
1
√2n
2n−1Xx=0
We say that a pair (G, q) is a metric group if G is a finite abelian group and q is a
non-degenerate quadratic form on G. For such a pair, we define unitary matrices
S(G,q)
g,g′ =
1
g,g′ = δg,g′q(g), C (G,q)
pGhg, g′iq, T (G,q)
(S(G,q)T (G,q))3 = G(q)C (G,q), C (G,q)T (G,q) = T (G,q)C (G,q).
g,g′ = δg+g′,0.
Then they satisfy the relations
(S(G,q))2 = C (G,q),
It is known that modular data of a pointed modular tensor category is always of
the form (S(G,q), T (G,q)) (see [EGNO15, Section 8.4]). Every metric group (G, q) is
canonically decomposed as (G, q) = (Ge, qe) × (Go, qo).
Definition 2.1. We say that (G, q, θ) is an involutive metric group if (G, q) is a metric
group and θ ∈ Aut(G) is an involution (an order 2 automorphism of G) preserving
q. For an involution θ ∈ Aut(G), we set
Gθ = {g ∈ G; θ(g) = g}.
We recall the following easy but useful lemma.
Lemma 2.2. Assume that (G, q) is a metric group and H is a subgroup of G. If the
restriction of q to H is non-degenerate, we have G = H × H⊥ and q = qH × qH⊥,
where
H⊥ = {g ∈ G; ∀h ∈ H, hg, hiq = 1}.
If moreover θ ∈ Aut(G) is an involution preserving q and θ(H) = H, then we also
have θ = θH × θH⊥.
2.3. When Gθ = Z2 × Z2. For later use we classify involutive metric groups whose
fixed point subgroups are Z2 × Z2.
Lemma 2.3. Let (A, q, θ) be an involutive metric group with a 2-group A. Assume
Aθ = {0, a} ∼= Z2.
(1) If q(a) = ±i, we have A = Z2, q(x) = i±x2, and θ = −1.
(2) If q(a) = −1, one of the following holds:
• A = Z4, q(x) = ζ rx2
• A = Z2×Z2, q is a flip invariant non-degenerate quadratic form, and θ(x, y) =
(y, x).
8 with odd r, and θ = −1.
(3) If q(a) = 1, we have A = Z2n with n ≥ 3, q(x) = ζ rx2
2n+1 with odd r, and θ = −1.
8
PINHAS GROSSMAN AND MASAKI IZUMI
Proof. If q(a) = ±i, the restriction of q to B = {0, a} is non-degenerate. Thus we
If B⊥ is not
have the factorization A = B × B⊥ as an involutive metric group.
trivial, the involution θ has a non-trivial fixed point as B⊥ is an even group, which
is a contradiction. Thus (1) holds.
Let ϕ : A → Aθ be a group homomorphism given by ϕ(x) = x + θ(x), and let ϕ0
be the restriction of ϕ to A2. Then ker ϕ0 = Aθ, and A2/Aθ = ϕ0(A) ≤ 2, which
shows A2 ≤ 4. Thus the 2-rank of A is at most 2.
If the 2-rank of A is 1, the group A is cyclic, say Z2n, and we have q(x) = ζ rx2
2n+1
with odd r and a = 2n−1. Since θ preserves q, there exists an integer s with s2−1 ≡ 0
mod 2n+1 with θ(x) = sx. Since Aθ = {0, 2n−1}, the number (s − 1)/2 is odd and
s + 1 ≡ 0 mod 2n, which implies θ = −1.
Assume that the 2-rank of A is 2. Then A2/Aθ = 2, and ϕ0 is a surjection. Thus
there exists b ∈ A2 satisfying b + ϕ(b) = a. If q(a) = −1, we have A2 = {0, a, b, θ(b)},
and {q(a), q(b), q(θ(a))} = {−1, q(b), q(b)}. This shows that q restricted to A2 is non-
degenerate, and we get A = A2 as before. Assume q(a) = 1 now. Since A = 2 ker ϕ
and ker ϕ∩{0, b} = {0}, we have the decomposition A = ker ϕ×{0, b}. Let x ∈ ker ϕ.
Then we have q(x + b) = q(x)q(b)hx, bi, and on the other hand, we have
q(θ(x + b)) = q(−x + a + b)
= q(−x + a)q(b)h−x + a, bi = q(x)q(a)q(b)h−x, aih−x + a, bi
= q(x)q(b)hx, aihx, biha, bi.
We have
ha, bi =
q(a + b)
q(a)q(b)
=
q(θ(b))
q(b)
= 1.
Since q(θ(x + b)) = q(x + b), we get hx, ai = 1, and we also have ha, bi = 1. Since A is
generated by ker ϕ and b this means that q is degenerate, which is a contradiction. (cid:3)
Theorem 2.4. Let (A, q, θ) be an involutive metric group with a 2-group A. Assume
Aθ = {0, a0, a1, a2} ∼= Z2 × Z2. Then the following hold up to group automorphism.
(1) If q(a0) = −1, q(a1) = q(a2) = 1, we have A = Z2
2, q(x, y) = (−1)xy, and θ = −1.
2, q(x, y) = (−1)x2+xy+y2, and
(2) If q(a0) = q(a1) = q(a2) = −1, we have A = Z2
θ = −1.
2, q(x, y) = i±(x2+y2), and
(3) If q(a0) = −1, q(a1) = q(a2) = ±i, we have A = Z2
θ = −1.
(4) If q(a0) = 1, q(a1) = i, q(a2) = −i, we have A = Z2
θ = −1.
(5) If q(a0) = −1, q(a1) = i, and q(a2) = −i, one of the following holds:
2, q(x, y) = ix2−y2, and
ζ r2y2
• A = Z2 × Z4, q(x, y) = ζ r1x2
8
• A = Z2×Z2×Z2, q(x, y, z) = ζ r1x2
where q′ is a flip invariant non-degenerate quadratic form on Z2 × Z2.
with odd r1, r2, and θ = −1.
q′(y, z) with odd r, and θ(x, y, z) = (x, z, y),
4
4
(6) If q(a0) = 1, q(a1) = −1, q(a2) = −1, one of the following holds:
ζ r2y2
• A = Z4 × Z2m with m ≥ 2, q(x, y) = ζ r1x2
2m+1 with odd r1, r2, and θ = −1
for n = 2, and θ = −1 or θ(x, y) = (−x + 2y, 2n−1x + (−1 + 2n−1)y) for n ≥ 3.
8
INFINITE FAMILIES OF MODULAR DATA
9
• A = Z2 × Z2 × Z2m with m ≥ 2, q(x, y, z) = q′(x, y)ζ rz2
2m+1 with odd r, and
θ(x, y, z) = (y, x,−z), where q′ is a flip invariant non-degenerate quadratic
form on Z2 × Z2.
2, q(x1, x2, x3, x4) = q′(x1, x3)q′′(x2, x4), where q′ and q′′ are flip in-
• A = Z4
variant non-degenerate quadratic forms on Z2 × Z2, and θ(x1, x2, x3, x4) =
(x3, x4, x1, x2).
(7) If q(a0) = 1, q(a1) = q(a2) = ±i, we have A = Z2 × Z2n with n ≥ 3, q(x, y) =
i±x2ζ ry2
(8) If q(a0) = q(a1) = q(a2) = 1, one of the following holds:
2n+1, and θ = −1.
2n or q(x, y) = ζ x2+xy+y2
2n
2n with n ≥ 2, q(x, y) = ζ xy
• A = Z2
n = 2, and θ = −1 or θ = −1 + 2n−1 for n ≥ 3.
• A = Z2m × Z2n with 3 ≤ m ≤ n, q(x, y) = ζ r1x2
θ = −1 or θ(x, y) = (−x + 2m−1y, 2n−1x − y).
• A = Z4
and θ(x1, x2, x3, x4) = (x3, x4, x1, x2).
, and θ = −1 for
2n+1 with odd r1, r2, and
2m+1ζ r2y2
2, q(x1, x2, x3, x4) = (−1)x1x4+x2x3 or q(x1, x2, x3, x4) = (−1)x1x4+x2x3i(x1+x3)2,
Proof. In cases (1)-(4), the restriction of q to {0, a0, a1, a2} is non-degenerate, and we
get the statement. If a1 = ±i, the restriction of q to B = {0, a1} is non-degenerate,
and we get A = B × B⊥ as an involutive metric group. Thus the restriction of θ
to B⊥ has exactly one non-trivial fixed point, and the statement follows from the
previous lemma. The remaining cases are (6) and (8).
Let ϕ : A → Aθ be a group homomorphism given by ϕ(x) = x + θ(x), and let ϕ0
be the restriction of ϕ to A2. Then ker ϕ0 = Aθ and A2/Aθ = ϕ0(A2) ≤ 4. Thus
A2 ≤ 16, and the 2-rank of A is at most 4. When it is 4, the map ϕ0 is surjection,
and when it is 3, we get ϕ0(A2) = 2. If it is 2, we have A2 = Aθ, and the possible
structure of the metric group (A, q) is determined from the classification of metric
groups.
8
(6) Assume that the 2-rank of A is 2. Then we may assume that A = Z4× Z2n with
ζ r2y2
n ≥ 2 and q(x, y) = ζ r1x2
2n+1 with odd r1, r2. Since ϕ(A) ⊂ A2, the involution θ is of
the form θ(x, y) = (εx+2sy, 2n−1tx+(−1+2n−1u)y) with ε ∈ {1,−1}, s, t, u ∈ {0, 1}.
Since θ preserves q and Aθ = A2, we get the statement.
Assume that the 2-rank of A is 4. Then there exists b ∈ A2 satisfying b + θ(b) = a1.
Let B = {0, a1, b, θ(b)} ∼= Z2 × Z2. Since {q(a1), q(b), q(b)} = {−1, q(b), q(b)}, the
restriction of q to B is non-degenerate, and we get the factorization A = B × B⊥ as
an involutive metric group. Now the statement follows from the previous lemma.
Now assume that the 2-rank of A is 3. If ϕ0(A2) = {0, a1} or {0, a2}, the statement
follows from the same argument as above, and we assume ϕ0(A) = {0, a0}. If there
exists d ∈ ker ϕ satisfying 2d = a1, the restriction of q to B = hdi ∼= Z4 is non-
degenerate as q(2d) = −1, and B is globally preserved by θ. Thus A = B × B⊥ as an
involutive metric group, and we get the statement. Therefore we assume that neither
a1 nor a2 has a square root in ker ϕ. We choose b ∈ A2 satisfying b + θ(b) = a0. Then
A2 is generated by a0, a1, and b. For ϕ(A), we have two possibilities: {0, a0} and
{0, a0, a1, a2}.
10
PINHAS GROSSMAN AND MASAKI IZUMI
First we assume ϕ(A) = {0, a0}. Then A = 2 ker ϕ. Since ker ϕ ∩ {0, b} = {0},
we get A = ker ϕ × {0, b}. Let x ∈ ker ϕ. Then q(x + b) = hx, biq(x)q(b). On the
other hand,
q(θ(x + b)) = q(−x + a0 + b)
= h−x + a0, biq(−x + a0)q(b)
= hx, biha0, bih−x, a0iq(a0)q(b)
= hx, biha0, bihx, a0iq(b)
We can show that ha0, bi = 1 as before, and since θ preserves q, we get hx, a0i =
1. Since A is generated by ker ϕ and b, we see that q is degenerate, which is a
contradiction. Thus ϕ(A) = {0, a0} does not occur.
Assume ϕ(A) = Aθ. Then there exists c ∈ A\A2 satisfying c+θ(c) = a1. Note that
we have 2c ∈ ker ϕ \ {0}, and A is generated by ker ϕ, b, and c. Let ψ : A → ker ϕ
be a group homomorphism given by ψ(x) = x − θ(x). Then ker ψ = Aθ. Thus
ψ(A) = A/4, and we also have ker ϕ = A/ϕ(A) = A/4. Thus ψ is a surjection
from A onto ker ϕ. We have ψ(ker ϕ) = 2 ker ϕ, ψ(b) = a0, and ψ(c) = 2c − a1. Since
ker ϕ∩A2 = Aθ = 4, the 2-rank of ker ϕ is 2, and we get ker ϕ = ha0i×h2c+a1i. We
express a1 as ma0 + n(2c + a1) with integers m, n. Then a2 = (m + 1)a0 + n(2c + a1).
Since neither a1 nor a2 has a square root in ker ϕ, the number n is odd, and we
also have ker ϕ = ha0i × ha1i ∼= Z2 × Z2. Thus we have 2c + a1 = a2 as 2c 6= 0, and
2c = a0. This means that we have A = ha1i×hbi×hci ∼= Z2× Z2× Z4 with θ(a1) = a1,
θ(b) = b + a0 = b + 2c, and θ(c) = a1 − c. We express xa1 + yb + zc as (x, y, z). Then
since q(a1) = −1, we have q(x, y, z) = (−1)xir1y2ζ r2z2
(−1)sxy+txz+uyz with r1 ∈ {0, 1},
r2 ∈ {0,±1,±2,±3}, and s, t, u ∈ {0, 1}. Since q(a0) = 1 and q(a2) = −1, we
have 1 = q(0, 0, 2) = (−1)r2, and −1 = q(1, 0, 2) = −(−1)r2, which shows that r2 is
even. However this implies q(x, y, z + 2) = q(x, y, z), and q is degenerate, which is a
contradiction.
8
(8) If the 2-rank of A is 2, we can show the statement as in (6).
Assume the 2-rank of A is 4. Then we have ϕ0(A2) = Aθ. Thus we can choose
b0, b1 ∈ A2 satisfying b0 + θ(b0) = a0 and b1 + θ(b1) = a1. Then b2 = b0 + b1 satisfies
b2 + θ(b2) = a2. Since ker ϕ = A/4 and ker ϕ ∩ {0, b0, b1, b2} = {0}, we have a
factorization A = ker ϕ × {0, b0, b1, b2}. Since ker ϕ ∩ A2 = Aθ, the 2-rank of ker ϕ is
2. Thus we may assume that ker ϕ = Z2m × Z2n with 1 ≤ m ≤ n and a0 = (2m−1, 0)
and a1 = (0, 2n−1). Identifying {0, b0, b1, b2} with hb0i × hb1i, we are in the following
situation: A = Z2m × Z2n × Z2 × Z2 and
θ(x1, x2, x3, x4) = (−x1 + 2m−1x3,−x2 + 2n−1x4, x3, x4).
If n ≥ 2, we have (0, 2n−1, 0, 1) ∈ Aθ, which is a contradiction. Thus m = n = 2.
Since q(1, 0, 0, 0) = q(0, 1, 0, 0) = q(1, 1, 0, 0) = 1, we have
q(x1, x2, x3, x4) = ir1x2
3+r2x2
4(−1)s13x1x3+s14x1x4+s23x2x3+s24x2x4+s34x3x4,
with r1, r2 ∈ {0,±1, 2} and s13, s14, s23, s24, s34 ∈ {0, 1}. Since q is preserved by θ, we
get s13 = s24 = 0 and s14 = s23, and then
q(x1, x2, x3, x4) = ir1x2
3+r2x2
4(−1)sx3x4(−1)t(x1x4+x2x3),
INFINITE FAMILIES OF MODULAR DATA
11
with s, t ∈ {0, 1}. The Gauss sum is
3+r2x2
1
4 Xx3,x4
ir1x2
G(q) =
=(cid:26) 1 + ir1 + ir2 + ir1+r2(−1)s,
1,
4(−1)sx3x4(1 + (−1)tx3)(1 + (−1)tx4)
t = 0
t = 1
,
which shows that q is non-degenerate if and only if t = 1. The centralizer of θ in
Aut(A) = GL(4, 2) is
CAut(A)(θ) = {(cid:18) X Y
0 X (cid:19) ∈ GL(4, 2); X ∈ GL(2, 2)},
and there are exactly two CAut(A)(θ)-orbits in the non-degenerate quadratic forms
preserved by θ. The two orbits are represented by q(x1, x2, x3, x4) = (−1)x1x4+x2x3
and q(x1, x2, x3, x4) = ix2
3(−1)x1x4+x2x3. Expressing q and θ with respect to the basis
{(0, 0, 1, 0), (0, 0, 0, 1), (1, 0, 1, 0), (0, 1, 0, 1)},
we get the statement.
Assume that the 2-rank of A is 3. Since ϕ0(A2) = 2, we may assume that there
exists b ∈ A2 satisfying b + θ(b) = a0. We can show that ϕ(A) = 2 does not occur as
in (6), and we assume ϕ(A) = Aθ. Then there exists c ∈ A\A2 satisfying c+θ(c) = a1,
2c ∈ ker ϕ, and A is generated by ker ϕ, b, and c. As before we have ψ(A) = ker ϕ,
and ker ϕ is generated by ψ(ker ϕ) = 2 ker ϕ, ψ(b) = a0, and ψ(c) = 2c + a1. Since
ker ϕ∩A2 = Aθ, the 2-rank of ker ϕ is 2, and we get ker ϕ = hbi×h2c+a2i ∼= Z2×Z2m.
We can show that m = 1 does not occur as in (6), and we assume m ≥ 2. Then we
have either a1 = 2m−1(2c + a1) = 2mc or a1 = a0 + 2m−1(2c + a1) = a0 + 2mc. In any
case, we have a1 ∈ ha0, ci, and
Since q(1, 0, 0) = q(0, 0, 2m) = (1, 0, 2m) = 1, we get
A = ha0i × hbi × hci ∼= Z2 × Z2 × Z2m+1.
q(x, y, z) = ir2y2
ζ r3z2
2m+2(−1)sx1x2+tx1x3+ux2x3,
with r2 ∈ {0,±1, 2}, r3 ∈ {0,±1 ± 2,±3}, and s, t, u ∈ {0, 1}.
Assume a1 = 2mc first. Then since θ(a0) = a0, θ(b) = b + a0, and θ(c) = −c + a1 =
(2m − 1)c, we have θ(x, y, z) = (x + y, y, (2m − 1)z). Since θ preserves q, we see
that r3 is even and q(x, y, z) = q(x, y, z + 2m). This shows that q is degenerate, and
this case does not occur. Assume a1 = a0 + 2mc now. Then we have θ(x, y, z) =
(x + y + z, y, (2m − 1)z). Again we can see that r3 is even and q is degenerate. Thus
this case does not occur either.
(cid:3)
2.4. Quadratic categories. Let P and Q be finite abelian groups and let m be a
natural number. We say that a fusion category C is a quadratic category of type
(P, Q, m) if the set of equivalence classes of simple objects in C are represented by
{αp ⊗ βq}p∈P, q∈Q ∪ {αp ⊗ ρ}p∈P ,
12
PINHAS GROSSMAN AND MASAKI IZUMI
and they satisfy the following fusion rules,
[αp][αp′] = [αp+p′],
[βq][βq′] = [βq+q′],
[αp][βq] = [βq][αp],
[αp][ρ] = [ρ][α−p],
[ρ2] =Xq∈Q
[βq] + mQXp∈P
[βq][ρ] = [ρ],
[αp][ρ].
Thanks to [βq][ρ] = [ρ], we can see that {βq}q∈Q has a trivial associator. Irrational
near-group categories are of type ({0}, Q, m) and generalized Haagerup categories
are of type (P,{0}, 1).
We can also consider a non-self-dual version. Namely, we say that a fusion category
C is a quadratic category of type (P, Q, m) with non-self-dual ρ if the set of equivalence
classes of simple objects in C is as above, but satisfying the following fusion rules:
[αp][αp′] = [αp+p′],
[βq][βq′] = [βq+q′],
[αp][βq] = [βq][αp],
[αp][ρ] = [ρ][α−p],
[ρρ] =Xq∈Q
[βq] + mQXp∈P
[βq][ρ] = [ρ],
[αp][ρ];
and there exists p0 ∈ P such that
[ρ] = [αp0][ρ].
The Z2-de-equivariantizaiton of a near-group category for Z4m with multiplicity 4m is
a quadratic category of type (Z2, Zm, 1) with non-self-dual ρ ([Izu17, Theorem 12.5]).
3. Near-group family
3.1. General formulae. The modular data (S, T ) of the Drinfeld center Z(C) of
a near-group category C with a finite abelian group A and multiplicity A (i.e. a
quadratic category of type ({0}, A, 1) was computed in [Izu01] and [EG14]. Evans-
Gannon [EG14, Conjecture 2] conjectured an explicit formula of (S, T ) for odd A,
and they verified the conjecture for A ≤ 13. Their formula is surprising simple and
it involves only two metric groups (A, q) and (A′, q′) satisfying A′ = A + 4 and
G(q′) = −G(q). It is immediate to describe the first metric group (A, q) in terms of
C, while the second metric group (A′, q′) is rather mysterious, and it is captured only
through heavy computation. In this section, we generalize their formula to general
finite abelian groups. Our family of potential modular data also include examples
which are not associated to Drinfeld centers of fusion categories.
Throughout this section, we assume that (G, q1, θ1) and (Γ, q2, θ2) are involutive
metric groups satisfying c := G(q1) = −G(q2) and
(U, q0) := (Gθ1, q1Gθ1 ) ∼= (Γθ2, q2Γθ2 ).
We fix an isomorphism of Gθ1 and Γθ2 as above, and regard U as a common subgroup
of G and Γ.
We set a = 1/pG, b = 1/pΓ. We choose and fix subsets G∗ ⊂ G and Γ∗ ⊂ Γ
satisfying
G = U ⊔ G∗ ⊔ θ1(G∗), Γ = U ⊔ Γ∗ ⊔ θ2(Γ∗),
INFINITE FAMILIES OF MODULAR DATA
13
and set
J = (U × {0, π}) ⊔ G∗ ⊔ Γ∗.
We use letters u, u′, u′′, v for elements of U, g, g′, g′′, h for elements of G and γ, γ′, γ′′, ξ
for elements of Γ. We introduce an involution of J by setting (u, 0) = (−u, 0),
(u, π) = (−u, π), and
g =(cid:26) −g,
θ1(−g),
γ =(cid:26) −γ,
θ2(−γ),
if −g ∈ G∗
otherwise
if −γ ∈ Γ∗
otherwise
,
.
We often suppress the subscript i in θi and qi in h·,·iqi if there is no possibility of
confusion.
Definition 3.1. Let S, T , and C be J by J matrices defined by
S =
a+b
a−b
2 hu, u′i
2 hu, u′i
a−b
a+b
2 hu, u′i
2 hu, u′i
ahg, u′i
ahg, u′i
bhγ, u′i −bhγ, u′i
ahu, g′i
ahu, g′i
a(hg, g′i + hθ(g), g′i)
0
bhu, γ′i
−bhu, γ′i
0
−b(hγ, γ′i + hθ(γ), γ′i)
,
T = Diag(q0(u), q0(u), q1(g), q2(γ)),
and Cx,y = δx,y (where the four blocks of S and T are indexed by U × {0}, U × {π},
G∗, and Γ∗, respectively.)
Lemma 2.2 immediately implies the following.
Lemma 3.2. Let the notation be as in Definition 3.1. Assume that L is a subgroup
of U such that the restriction of q0 to L is non-degenerate, and let
G′ = {g ∈ G; ∀l ∈ L hg, li = 1},
Γ′ = {γ ∈ Γ; ∀l ∈ L hγ, li = 1},
q′1 = q1G′, θ′1 = θ1G′, q′2 = q2Γ′, θ′2 = θ2Γ′, and q′0 = q0L. Then (G′, q′1, θ′1) and
(Γ′, q′2, θ′2) also satisfy the assumptions of this section. If we let (S′, T ′) denote the
unitary matrices defined in Definition 3.1 for G′ and Γ′, we have
(S, T ) = (S(L,q′0) ⊗ S′, T (L,q′0) ⊗ T ′).
Remark 3.3. Recall that G (and hence q1 and θ1 too) is canonically decomposed as
Ge×Go, where Ge is a 2-group and Go is an odd group. Moreover, we can diagonalize
θ1 on Go. Thus we have the canonical decomposition
(G, q1, θ1) = (Ge, q1,e, θ1,e) × (Go,+, q1,o,+, 1) × (Go,−, q1,o,−,−1),
(Γ, q2, θ2) = (Γe, q2,e, θ2,e) × (Γo,+, q2,o,+, 1) × (Γo,−, q2,o,−,−1).
Under the identification (U, q0) = (Gθ, q1Gθ1 ) = (Γθ2, q2Γθ2 ), we get
(U, q0) = (Gθ1,e
e
, q1G
θ1,e
e
) × (Go,+, q1,o+) = (Γθ2,e
e
, q2Γ
θ2,e
2
) × (Γo,+, q2,o+).
Thus we can always apply Lemma 3.2 to L = Go,+ = Γo,+. Therefore for the purposes
of constructing new modular data, we may assume that Go,+ = Γo,+ = {0}.
14
PINHAS GROSSMAN AND MASAKI IZUMI
Direct computation shows the following.
Lemma 3.4. Let the notation be as in Definition 3.1. Then S, T , and C are unitary
matrices satisfying S2 = C, (ST )3 = cC, and T C = CT .
Lemma 3.5. Let the notation be as Definition 3.1. Then we have
N(u,0),(u′,0),(u′′,0) = N(u,0),(u′,π),(u′′,π) = δu+u′+u′′,0,
N(u,0),(u′,0),(u′′,π) = N(u,0),(u′,0),g = N(u,0),(u′,0),γ = N(u,0),(u′,π),g = N(u,0),(u′,π),γ = 0,
N(u,0),g,g′ = δu+g+g′,0+δu+θ(g)+g′,0, N(u,0),g,γ = 0, N(u,0),γ,γ′ = δu+γ+γ′,0+δu+θ(γ)+γ′,0,
N(u,π),(u′,π),(u′′,π) =
N(u,π),(u′,π),g =
N(u,π),(u′,π),γ =
4U
Γ − G
4U
1
Γ − G
4U
Γ − G
1
UXv∈U
UXv∈U
UXv∈U
1
hu + u′ + u′′, vi,
hu + u′ + g, vi,
hu + u′ + γ, vi,
N(u,π),g,g′ =
N(u,π),γ,γ′ =
4U
Γ − G
N(u,π),g,γ =
4U
Γ − G
4U
Γ − G
1
UXv∈U
4U
Γ − G
1
UXv∈U
UXv∈U
1
1
hu + g + g′, vi + δu+g+g′,0 + δu+θ(g)+g′,0,
UXv∈U
hu + γ + γ′, vi − δu+γ+γ′,0 − δu+θ(γ)+γ′,0,
hu, vihg, vihγ, vi,
Ng,g′,g′′ =
hg + g′ + g′′, vi
+ δg+g′+g′′,0 + δθ(g)+g′+g′′,0 + δg+θ(g′)+g′′,0 + δg+g′+θ(g′′),0,
Ng,g′,γ =
Ng,γ,γ′ =
4U
Γ − G
1
1
UXv∈U
UXv∈U
hγ + γ′ + γ′′, vi
4U
Γ − G
4U
Γ − G
1
UXv∈U
hg + g′, vihγ, vi,
hg, vihγ + γ′, vi,
Nγ,γ′,γ′′ =
− δγ+γ′+γ′′,0 − δθ(γ)+γ′+γ′′,0 − δγ+θ(γ′)+γ′′,0 − δγ+γ′+θ(γ′′),0.
The above computation shows that in order for (S, T ) to be modular data, the
number 4U/(Γ − G) has to be a positive integer. On the other hand, since U
is a common subgroup of G and Γ, the number (Γ − G)/U is a positive integer
too. Thus we have only three possibilities:
Γ = G + U, Γ = G + 2U, and
Γ = G + 4U.
INFINITE FAMILIES OF MODULAR DATA
15
Theorem 3.6. Let the notation be as in Definition 3.1. All the fusion coefficients in
Lemma 3.5 are non-negative integers if and only of one of the following occurs:
• Γ = G + 4U,
• Γ = G + 2U, and G/U is an even number.
Proof. We first assume Γ = G + U. Then either Γ/U or G/U is an odd
number. We assume that G/U is odd as the other case can be treated in the same
way. Let
(G, q1, θ1) = (Ge, q1,e, θ1,e) × (Go,+, q1,o+, 1) × (Go,−, q1,o−,−1)
be the decomposition as in Remark 3.3. Then G/U being odd implies that θ1,e is
trivial, and
In particular q0 is non-degenerate, and Lemma 3.2 implies that we have a factorization
(U, q0) = (G1,e, q1,e) × (Go,+, q1,o+).
(Γ, θ2) = (U, q0) × (Γ′, θ′2),
with θ′2 a fixed-point-free involution. However Γ′ = G/U + 1 implies that Γ′ is an
even group, which never allows a fixed-point-free involution, and we get contradiction.
Assume now that Γ = G + 2U, and G/U is odd. Then Γ/U is odd too,
and we have the decompositions
(G, q1, θ1) = (Ge, q1,e, θ1,e) × (Go,+, q1,o+, 1) × (Go,−, q1,o−,−1),
(Γ, q2, θ2) = (Γe, q2,e, θ2,e) × (Γo,+, q1,o+, 1) × (Γo,−, q1,o−,−1),
with trivial θ1,e and θ2,e. Thus
(G, q1) = (U, q0) × (Go,−, q1,o−),
(Γ, q2) = (U, q0) × (Γo,−, q2,o−),
and G(q1,o−) = −G(q2,o−). However, this is impossible as Go,− and Γo,− are odd
groups with Γo,− = Go,− + 2.
(cid:3)
Recall that a near-group category for a finite abelian group A with multiplicity
A comes with a non-degenerate symmetric bicharacter, say hh·,·ii (we avoid the
symbol h·,·i used in [Izu01], [Izu17] to prevent possible confusion). We generalize
Evans-Gannon's conjecture. It is easy to see from Lemma 3.2 and Remark 3.3 that
the following statement actually generalizes [EG14, Conjecture 2] from the odd group
case to the general case.
Conjecture 3.7. The modular data of the Drinfeld center of a near-group category
for a finite abelian group A with multiplicity A and a non-degenerate symmetric
bicharacter hh·,·ii is given by (S, T ) of Definition 3.1 with G = A × A, q1(a1, a2) =
hha1, a2ii, θ1(a1, a2) = (a2, a1), and
Γ = G + 4U = A(A + 4).
16
PINHAS GROSSMAN AND MASAKI IZUMI
A = Z4.
The conjecture is true for odd groups A with A ≤ 13, as well as A = Z2 and
We compute the Frobenius-Schur indicators now. For u ∈ U and m ∈ Z, let
hu, γiq2(γ)m,
hu, giq1(g)m, G(q2, u, m) =
G(q1, u, m) =
1
pGXg∈G
1
pΓXγ∈Γ
and
G(q1, q2, v, m) =
hv, giq1(g)m +
Lemma 3.8. Let the notation be as above. Then
1
pGXg∈G
1
pΓXγ∈Γ
hv, γiq2(γ)m.
νm((u, 0)) = δmu,0q0(u)m,
1
hu, viG(q1, q2, v, m)2,
Γ − GXv∈U
hg, viG(q1, q2, v, m)2 + δmg,0q1(g)m,
hγ, viG(q1, q2, v, m)2 − δmγ,0q1(γ)m.
νm((u, π)) =
νm(g) =
νm(γ) =
1
Γ − GXv∈U
Γ − GXv∈U
1
Note that if G is an odd group and (S, T ) comes from a modular tensor category,
we have Γ = G + 4, and then
νm(π) =
1
4G(q1, m) + G(q2, m)2,
νm(g) =
νm(γ) =
1
4G(q1, m) + G(q2, m)2 + δmg,0q1(g)m,
1
4G(q1, m) + G(q2, m)2 − δmγ,0q1(γ)m.
Lemma 3.9. Assume G is an odd group with θ1 = −1 and Γ = G + 4.
(1) We have ν2(x) = 1 for any x ∈ J, and this does not give any restriction for
(2) If the 3-rank of G and that of Γ are both less than or equal to 1, the values of
the pair (G, q1) and (Γ, q2).
ν3(x) are consistent with Nx,x,x for any x ∈ J .
(3) If either the 3-rank of G or that of Γ is greater than or equal to 3, the value of
ν3(π) is not consistent with Nπ,π,π = 1, and the pair (S, T ) never comes from
a modular tensor category.
Proof. (1) We have
2
ν2(π) = G(q1, 2) + G(q2, 2)
= G(q1)(−1) G2−1
= G(q1) − G(q2)
2
8 + G(q2)(−1)
2 = G(q1)2 = 1.
2
2
(G+4)2−1
8
2
INFINITE FAMILIES OF MODULAR DATA
17
We can similarly show ν2(x) = 1 for the other x ∈ J from this computation.
Jacobi symbol and the quadratic reciprocity law, we get
(2) If either G or Γ has no 3-component (i.e.
G ≡ 1 mod 3), then using the
2
G(q1)( 3
G
ν3(π) = G(q1, 3) + G(q2, 3)
) + G(q2)( 3
Γ
=
2
= G(q1) − G(q2)
2
2 = G(q1)2 = 1.
2
2 = G(q1)(−1) G−1
2
)
3 ) + G(q2)(−1) G+3
(G
2
2
(G+4
3
)
2
If G has 3-rank 1, we have either G(q1, 3) ∈ {√3,−√3} and G(q2, 3) ∈ {i,−i}, or
G(q1, 3) ∈ {√3i,−√3i} and G(q2, 3) ∈ {1,−1}.In either case we have ν3(π) = 1. We
(3) If G has 3-rank larger than or equal to 3, we get G(q1, 3) ≥ 3√3 and G(q2, 3) =
get the same conclusion if Γ has 3-rank 1. Consistency of ν3(x) and Nx,x,x for other
x follows from the same computation.
1. Then ν3(π) > 1, which is not consistent with Nπ,π,π = 1. We get the same
conclusion if Γ has 3-rank larger than equal to 3.
(cid:3)
Remark 3.10. The most interesting case is of course when either the 3-rank of G or
that of Γ is 2. In this case, our formulae for ν3(π) and Nπ,π,π = 1 give a restriction
of quadratic forms. For example, assume G = Z5 and Γ = Z3 × Z3. Then
ν3(π) = G(q1, 3) + 3
2
2 = −G(q1) + 3
2
2 = G(q2) + 3
2
2.
Thus G(q2) = 1 is forbidden, and only G(q2) = −1 (and consequently G(q1) = 1)
is consistent with Nπ,π,π = 1. This explains why there are only 3 solutions for the
polynomial equations for the near-group category for Z5 with multiplicity 5 (see
In fact there are two solutions corresponding to Γ = Z9 with
[EG14, Table 2]).
9 , ζ−x2
q2(x) = ζ x2
, and one solution corresponding to Γ = Z3 × Z3 with q2(x, y) =
ζ x2+y2
. In a similar way, for G = Z3 × Z3 and Γ = Z13, only G(q1) = 1 is possible,
3
which gives the modular data of the Drinfeld center of the Haagerup category.
9
Corollary 3.11. If Conjecture 3.7 is true, there exists no generalized Haagerup cat-
egory with an odd abelian group A whose 3-rank is larger than or equal to 2.
Proof. If C is such a category, C ⊠ VecA is Morita equivalent to a near-group category
for A×A with multiplicity A2 (see [Izu17, Theorem 12.9]), which contradicts Lemma
3.9,(3).
(cid:3)
We get back to the general case. Recall that if m is an odd number and (A, q) is
a metric group with A = 2n, we have
G(q, m)
G(q)m = (−1)n m2−1
8
.
Lemma 3.12. Let the notation be as above. Let m ∈ Z and let u ∈ U.
18
PINHAS GROSSMAN AND MASAKI IZUMI
(1) For odd m, we have
G(q1, q2, u, m) = G(q1, q2, 0, m) = (−1)n1
8 G(q2,o, m)
m2−1
G(q2,o)m ,
where Ge = 2n1 and Γe = 2n2. If moreover (m,G) = (m,Γ) = 1, we
have
G(q1,o)m − (−1)n2
8 G(q1,o, m)
m2−1
G(q1, q2, u, m) = (−1)n1
m ) is the Jacobi symbol.
where (Go
(2) For m = 2, we have
m2−1
8
(Go
m
) − (−1)n2
m2−1
8
(Γo
m
),
G(q1, q2, u, 2) = G(q1,e, u, 2)
G(q1,e)
(−1) Go2−1
8 − G(q2,e, u, 2)
G(q2,e)
(−1) Γo2−1
8
.
Proof. (1) Since U is a 2-group, for any u ∈ U there exists a unique element u′ ∈ U
satisfying mu′ = u. Thus
G(q1, q2, u, m)
=
1
1
pGe Xg∈Ge
pGe Xg∈Ge
pGe Xg∈Ge
1
=
hu, giq1,eq1,e(g)mG(q1,o, m) +
hu, γiq2,eq2,e(γ)mG(q2,o, m)
1,eqm
1,e(g)G(q1,o, m) +
hu′, giqm
1,e(g + u′)
qm
1,e(u′) G(q1,o, m) +
qm
2,e(γ)G(q2,o, m)
2,eqm
hu′, γiqm
2,e(γ + u′)
qm
2,e(u′) G(q2,o, m)
qm
1
1
pΓe Xγ∈Γe
pΓe Xγ∈Γe
pΓe Xγ∈Γe
1
=
= G(q1, m) + G(q2, m).
G(q1, m)
G(q1)m − G(q2, m)
Since G(q1) = −G(q2), this is equal to
which shows the first part.
G(q1,e)m G(q1,o, m)
G(q2)m = G(q1,e, m)
G(q2,e)m G(q2,o, m)
G(q2,o)m ,
If moreover (m,G) = (m,Γ) = 1, the quadratic reciprocity law implies
) = (Go
),
m
G(q1,o)m − G(q2,e, m)
2 (Go
m
(Go
m
) = (
Go−1
m−1
m−1
G(q1,o, m)
G(q1,o)m =
2
2
(−1)
G(q1,o)m−1
2
(−1) Go−1
G(q1,o)2 )
and we get the second part.
(2) Since G(q1) = −G(q2), we get
G(q1, u, 2) + G(q2, u, 2)
= G(q1,e, u, 2)G(q1,o, 2)
= G(q1,e, u, 2)
(−1) Go2−1
G(q1,e)
G(q1,e)G(q1,o)
G(q2,e)G(q2,o)
− G(q2,e, u, 2)G(q2,o, 2)
8 − G(q2,e, u, 2)
G(q2,e)
(−1) Γo2−1
8
.
(cid:3)
INFINITE FAMILIES OF MODULAR DATA
19
3.2. Examples with Γ = G + 4U. We give examples of involutive metric groups
(G, q1, θ1) and (Γ, q2, θ2) satisfying the assumptions of Definition 3.1 withΓ = G +
4U such that (FS2) and (FS3) from Section 2.1 are satisfied. Taking Lemma 3.2
and Remark 3.3 into account, we assume
(G, q1, θ1) = (Ge, q1,e, θ1,e) × (Go, q1,o,−1),
(Γ, q2, θ2) = (Γe, q2,e, θ2,e) × (Γo, θ2,o,−1).
We also assume that q0 is degenerate on any non-trivial subgroup of U.
Lemma 3.5 implies that
N(u,π),(u,π),(u,π) =
1
UXv∈U
hu, vi,
Ng,g,g =
Nγ,γ,γ =
1
UXv∈U
UXv∈U
1
hg, vi + δ3g,0 + 3δ2g+θ(g),0,
hγ, vi − δ3γ,0 − 3δ2γ+θ(γ),0.
On the other hand, Lemma 3.8 and 3.12 show that
ν3(u, π) = G(q1, q2, 0, 3)2
4
ν3(g) = G(q1, q2, 0, 3)2
4
ν3(γ) = G(q1, q2, 0, 3)2
4
1
UXv∈U
UXv∈U
1
1
hu, vi,
UXv∈U
hg, vi + δ3g,0q1(g)3,
hγ, vi − δ3γ,0q2(γ)3.
Thus the condition (FS3) is equivalent to G(q1, q2, 0, 3) = 2. Direct computation
using Lemma 3.12 shows that restriction coming from (FS3) in the following examples
is very similar to that discussed in Lemma 3.9 and Remark 3.10, and we will not
mention it in what follows.
In the following subsections we consider different forms of Ge and θ1,e.
3.2.1. Assume Ge = Z2 × Z2 with θ1,e(x, y) = (y, x). In this case, we have U ∼= Z2
and Γ = 4Go + 8 = 4(Go + 2), and hence Γe = 4, Γo = Go + 2. Thus
G(q2,o)/G(q1,o) ∈ {i,−i}, and so G(q2,e)/G(q1,e) ∈ {i,−i} too. This implies Γe =
Z2 × Z2. Since θ2,e is non-trivial, we may assume θ2,e(x, y) = (y, x). Note that the
only flip-invariant quadratic forms of Z2 × Z2 are q(x, y) = (−1)xy, (−1)x2+xy+y2, and
i±(x2+y2). We consider these quadratic forms separately.
20
PINHAS GROSSMAN AND MASAKI IZUMI
3.2.1.1. Assume q1,e(x, y) = (−1)xy. In this case we have G(q1,e) = 1 and q2,e(x, y) =
i±(x2+y2). To fulfill G(q2) = −G(q1), we need G(q2,o) = ±iG(q1,o). Direct computation
shows that (FS2) gives no restriction. In the simplest example G = Z2 × Z2, Γ =
Z2 × Z2 × Z3, the pair (S, T ) is the modular data of the Drinfeld center of the near-
group category for Z2 with multiplicity 2 (see [Izu01]). More generally, we conjecture
that the modular data of the Drinfeld center of a near-group category for Z2 × Go
with multiplicity 2Go is given by
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T ).
3.2.1.2. Assume q1,e(x, y) = (−1)x2+xy+y2. In this case we have G(q1,e) = −1 and
q2,e(x, y) = i±(x2+y2). To fulfill G(q2) = −G(q1), we need G(q2,o) = ∓iG(q1,o). Direct
computation shows that (FS2) gives no restriction.
3.2.1.3. Assume q1,e(x, y) = i±(x2+y2). In this case G(q1,e) = ±i and q2,e(x, y) is either
(−1)xy or (−1)x2+xy+y2. We have G(q2,o) = ∓iG(q1,o) for q2,e(x, y) = (−1)xy, and
G(q2,o) = ±iG(q1,o) for q2,e(x, y) = (−1)x2+xy+y2. Direct computation shows that
(FS2) gives no restriction.
8
, q2,e(x) = ζ r2x2
8 with odd r, and G(q) = ζ r
, and G(q2,o) = −ζ r1−r2
3.2.2. Assume Ge = Z4. Since θ1,e is non-trivial, we have θ1,e(x) = −x. In this case
we have U ∼= Z2, and Γ = 4Go+8 = 4(Go+2), and hence Γe = 4, Γo = Go+2.
Thus G(q2,o)/G(q1,o) ∈ {i,−i}, and so G(q2,e)/G(1, e) ∈ {i,−i} too. This implies that
Γe = Z4. Since θ2,e is non-trivial, we have θ2,e(x) = −x. Non-degenerate quadratic
forms on Z4 are of the form q(x) = ζ rx2
8. Possible
combinations are q1,e(x) = ζ r1x2
G(q1,o), with
r1 − r2 ≡ 2 mod 4. Direct computation shows that (FS2) gives no restriction.
Remark 3.13. In the above examples with Ge = 4, we fix two odd metric groups
(Go, q1,o) and (Γo, q2,o). Then we can show that all the above potential modular
data (S, T ) can be obtained from that for q1,e(x, y) = (−1)xy by applying the zesting
construction introduced in [BGH+17, Theorem 3.15].
3.2.3. Assume Ge = Z4 × Z4 and θ1,e(x, y) = (y, x). In this case we have U ∼= Z4
and Γ = 24(Go + 1). Thus there exists n ≥ 3 satisfying Γe = 2n+2, and we have
2n+2ζ r(1−2n−2)y2
Go = 2n−2Γo − 1. We assume that Γe = Z2n+1 × Z2, q2,e(x, y) = ζ rx2
,
and
4
8
8
θ2,e(x, y) = ((2n−1 − 1)x + 2ny, x + y),
with r ∈ {1,−1, 3,−3}. Then θ2,e preserves q2,e and Γθ2,e
e =< (2n−1, 1) >∼= Z4. We
have
= ζ r
4.
q2,e(2n−1, 1) = ζ r22n−2
2n+2 ζ r(1−2n−2)
4
We have
To fulfill the condition (Gθ1, q1Gθ1 ) ∼= (Γθ2, q2Γθ2 ), we have 3 possibilities: q1,e(x, y) =
ζ rxy
4
. We consider these quadratic forms separately.
, ζ−r(x2+xy+y2)
4
8
G(q2,e) = (−1)n r2−1
, or ζ r(x2+y2)
8 (−1)
2n−3 (2n−3−1)
2
ir(1−2n−3).
INFINITE FAMILIES OF MODULAR DATA
21
3.2.3.1. Assume q1,e(x, y) = ζ rxy
4
. In this case G(q1,e) = 1 and
G(q1,o) = −(−1)n r2−1
8 (−1)
2n−3(2n−3−1)
2
ir(1−2n−3)G(q2,o).
For n = 3, we have G(q1,o)/G(q2,o) ∈ {1,−1} and Go ≡ Γo mod 4. This to-
gether with Go = 2Γo − 1 implies that Go ≡ 1 mod 8. For n ≥ 4, we have
G(q1,o)/G(q2,o) ∈ {i,−i} and Go ≡ Γo − 2 mod 4. This together with Go =
2n−2Γo − 1 implies that Go ≡ 3 mod 4 and Γo ≡ 1 mod 4.
Direct computation shows that (FS2) gives no restriction.
For the smallest example G = Z4 × Z4, Γ = Z16 × Z2, there are two possibilities:
r = ±3. These give the modular data for the Drinfeld centers of the two complex-
conjugate near-group categories for Z4 with multiplicity 4. This can be verified using
Theorem 6.8 of [Izu01], after solving Equations (6.18)-(6.21) for Z4 using the structure
constants found in Example A.2 there.
More generally, we conjecture that the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T ),
is the modular data of the Drinfeld center of a near-group category for Z4 × Go with
multiplicity 4Go.
3.2.3.2. Assume q1,e(x, y) = ζ−r(x2+xy+y2)
tion is very similar to the previous case.
. In this case G(q1,e) = 1, and the computa-
4
3.2.3.3. Assume q1,e(x, y) = ζ r(x2+y2)
G(q2,o) = −(−1)n r2−1
8
. In this case G(q1,e) = ir and
8 (−1)
2n−3 (2n−3−1)
2
ir2n−3
G(q1,o).
For n = 3, we have G(q1,e)/G(q2,e) ∈ {i,−i} and Go ≡ Γe − 2 mod 4. This
together with Go = 2Γe − 1 shows that Go ≡ 5 mod 8 and Γo ≡ 3 mod 8. For
n ≥ 4, we have G(q1,e)/G(q2,e) ∈ {1,−1} and Go ≡ Γe mod 4. This together with
Go = 2Γe − 1 shows that Go ≡ Γe ≡ 3 mod 4.
Direct computation shows that (FS2) gives no restriction.
For the smallest example G = Z4× Z4× Z5, Γ = Z16× Z2× Z3, there are 8 possibil-
irG(q1,o).
ities: two choices of q1,o and four choices of r. We have G(q2,o) = −(−1)
r2−1
8
2 and θ1,e(x1, x2, y1, y2) = (y1, y2, x1, x2). In this case we have
3.2.4. Assume Ge = Z4
U ∼= Z2 × Z2 and Γ = 24(Go + 1). Again we consider different quadratic forms.
3.2.4.1. Assume q1,e(x1, x2, y1, y2) = (−1)x1y1+x2y2. We choose n ≥ 3 with Γe = 2n+2
and Go = 2n−2Γo − 1. Since G(q2,e)4 = G(q1,e)4 = 1, it follows from Theorem
2.4 that Γe = Z2n × Z4 is the only possibility. We have q2,e(x, y) = ζ rx2
8 with
r, s ∈ {1,−1, 3,−3}, and θ2,e(x, y) = (−x,−y) or ((−1 + 2n−1)x + 2n−1y, 2n−1x − y).
We denote the latter by τ (x, y).
2n+1ζ sy2
22
PINHAS GROSSMAN AND MASAKI IZUMI
We have q2(2n−1, 0) = 1, q2(0, 2) = −1, q2(2n−1, 2) = −1, and (Gθ1, q1) ∼= (Γθ2, q2).
We have G(q1,e) = 1 and G(q2,e) = (−1)n r2−1
G(q1,o) + (−1)n r2−1
8 G(q2,o) = 0.
The group automorphism (x, y) → (x + 2n−2y,±x + y) of Z2n × Z4 transforms (r, s)
to (r + 2n−2s, 2n−2r + s).
, and so
8 ζ r+s
8 ζ r+s
8
Direct computation shows that G(q1, q2, k, 2)2 = 4 for k 6= 0, and
G(q1, q2, 0, 2)2 = 4 − 2(−1)
r2−1
8 (−1)
s2−1
8 (−1) Go2−1
8
(−1) Γo2−1
8
2,
which is compatible with ν2(0, π) = ±1 if and only if
r2−1
8 (−1)
Assuming this condition, we have
(−1)
(−1) Γo2−1
8 = 1.
8
s2−1
8 (−1) Go2−1
UXu∈U
1
ν2(γ) =
hγ, ui − δ2γ,0q2(γ)2.
This shows that (FS2) is satisfied only for θ2,e = τ .
For the smallest example G = Z4
(−3, 3), which are transformed to each other by a group automorphism.
We conjecture that the pair
2, Γ = Z8 × Z4, we have either (r, s) = (3,−3) or
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
with θ2,e = τ is the modular data of the Drinfeld center of a near-group category for
Z2 × Z2 × Go with multiplicity 4Go.
3.2.4.2. Assume q1,e(x1, x2, y1, y2) = (−1)x1y2+x2y1. Note that we have q0(k) = 1 in
this case. Theorem 2.4 shows that only Γe = Z2m × Z2n with 3 ≤ m ≤ n is possible
for Γe, and we have
Γθ2,e
e = {0, (2m−1, 0), (0, 2n−1), (2m−1, 2n−1)}.
Since Go = 2m+n−4Γo − 1, we have Go ≡ 3 mod 4. Thus c = G(q1,o) =
−G(q2,e)G(q2,o) implies G(q2,e)2G(q2,o)2 = −1. We have the following three possi-
bilities for (q2,e, θ2,e):
2m+1ζ sy2
• q2,e(x, y) = ζ rx2
θ2,e(x, y) = (−x + 2m−1y, 2n−1x − y). We denote the latter by τ′(x, y).
• m = n, q2,e(x, y) = ζ xy
• m = n, q2,e(x, y) = ζ x2+xy+y2
2n+1 with r, s ∈ {1,−1, 3,−3}, and θ2,e(x, y) = (−x,−y) or
2m, and θ = −1 or θ = −1 + 2m−1.
, and θ = −1 or θ = −1 + 2m−1.
2m
If Γo ≡ 1 mod 4, only the first of these is possible, and we have
8 G(q2,o) = 0.
G(q1,o) + (−1)m r2−1
8 (−1)n s2−1
8 ζ r+s
Direct computation shows that G(q1, q2, l, 2)2 = 4 for l 6= 0, and
G(q1, q2, 0, 2)2 = 4 − 2(−1)
r2−1
8 (−1)
s2−1
8 (−1) Go2−1
8
(−1) Γo2−1
8
2,
INFINITE FAMILIES OF MODULAR DATA
23
which is possible only if
Assuming this condition, we see that (FS2) is satisfied only for θ2,e = τ′.
(−1)
r2−1
8 (−1)
s2−1
8
(−1) Γo2−1
8 (−1) Go2−1
2 × Z3, we have m = n = 3 and Γo = {0}. Up to
8 = 1.
For the smallest example G = Z4
group automorphism, there are two possibilities: (G(q1,o), r, s) = ±(i, 3,−1).
We conjecture that the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
with θ2,e = τ′ is the modular data of the Drinfeld center of a near-group category for
Z2 × Z2 × Go with multiplicity 4Go.
3.2.4.3. Other quadratic forms. There are various other θ1,e-invariant non-degenerate
quadratic forms on Z4
2 (see Theorem 2.4).
3.2.5. Assume Ge = Z2n × Z2n with n ≥ 3 and θ1,e(x, y) = (y, x). In this case we
have U ∼= Z2n and Γ = 22n + 2n+2Go, and hence Γe = 2n+2 and Γo = Go + 2n−2.
Let Γe = Z2n+1 × Z2, let
q2,e(x, y) = ζ sx2
2n+2ζ−(1+2n−2)sy2
4
with odd s, and let θ2,e(x, y) = ((1 + 2n−1)x + 2ny, x + y). Then θ2,e is an involution
preserving q2,e, and
e =< (2, 1) >∼= Z2n.
Γθ2,e
We have q2,e(2, 1) = ζ s(1−2n−2−22n−4)
2n
and
G(q2,e) = i−2n−3s(−1)
2n−3 (2n−3−1)
2
(−1)n s2−1
8
.
Thus G(q2,e) ∈ {i,−i} for n = 3, and G(q2,e) ∈ {1,−1} for n ≥ 4. Again we consider
different quadratic forms.
3.2.5.1. Assume q1,e(x, y) = ζ rxy
2n with r ∈ {1,−1, 3,−3}. We set
s = (1 + 2n−2)r + 2nt,
with t ∈ {0, 1, 2, 3}. Then
s(1 − 2n−2 − 22n−4) ≡ r mod 2n,
and q1,e(l, l) = q2,e(2l, l). Since the automorphism (x, y) 7→ ((1+2n)x, y) of Z2n+1 × Z2
commutes with θ2,e and transforms t to t + 2, we need to consider only t = 0 and
t = 1. We have
G(q1,o) = −i−2n−3s(−1)
2n−3(2n−3−1)
2
(−1)n s2−1
8 G(q2,o).
Since Γe = 2n−2Go + 1, and Γo ≡ Go + 2 mod 4 for n = 3 and Γo ≡ Go for
n ≥ 4, we have Go ≡ 1 mod 4, and Γo ≡ 3 mod 4 for n = 3 and Γo ≡ 1 mod 4
for n ≥ 4.
Direct computation shows that (FS2) gives no restriction.
We give a few examples.
24
PINHAS GROSSMAN AND MASAKI IZUMI
(i) Go = {0} with n = 3.
In this case we have Γo = Z3. There are four
(ii) Go = {0} with n = 4.
possibilities up to group automorphism. We have G(q2,o) = −(−1)
possibilities, and we have G(q2,o) = −1.
In this case we have Γo = Z5. There are four
8
r2−1
ir.
We conjecture that the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T ).
is the modular data of the Drinfeld center of a near-group category for Z2n × Go with
multiplicity 2nGo.
3.2.5.2. Assume q1,e(x, y) = ζ r(x2+xy+y2)
the previous case.
. This case can be treated in a similar way as
2n
3.2.5.3. Assume q1,e(x, y) = ζ r(x2+y2)
2n+1
with r ∈ {1,−1, 3,−3}. We set
s = (1 + 2n−2)r + 2nt,
with t ∈ {0, 1, 2, 3}. Then
s(1 − 2n−2 − 22n−4) ≡ r mod 2n,
and q1,e(l, l) = q2,e(2l, l). Since the automorphism (x, y) 7→ ((1+2n)x, y) of Z2n+1 × Z2
commutes with θ2,e and transforms t to t + 2, again we only need to consider t = 0
and t = 1. Since G(q1,e) = ir, we have
G(q1,o) = −i−(1+2n−3+22n−5)r(−1)
2n−3(2n−3−1)
2
(−1)n (1+2n−2)2 r2−1
8
G(q2,o).
Thus Γo ≡ Go mod 4 for n = 3 and Γo ≡ Go + 2 mod 4 for n ≥ 4. Now
Γo = 2n−2Go + 1 implies that Go ≡ 3 mod 4, and Γo ≡ 3 mod 4 for n = 3 and
Γo ≡ 1 mod 4 for n ≥ 4. The rest of computation is similar to 3.2.5.1 above.
3.2.6. Assume Ge = Z2n with n ≥ 3, and θ1,e(x) = −x. In this case U ∼= Z2, and
Γ = 2nGo + 23 = 23(2n−3Go + 1). Let q1,e(x) = ζ rx2
2n+1 with r ∈ {1,−1, 3,−3}. We
have G(q1,e) = ζ r
2n+1 = 1. We consider the cases n = 3
and n > 3 separately.
, and q1(2n−1) = ζ r22n−2
8 (−1)n r2−1
8
3.2.6.1. Assume n=3. In this case Ge = Z8, Γe = Z2m, and Go = 2m−3Γo − 1 with
m ≥ 4. Assume q2,e(x) = ζ sx2
2m+1 with s ∈ {1,−1, 3,−3} and θ2(x) = −x. Then
G(q1,e) = (−1)
8. Thus
8 ζ r
r2−1
8 and G(q2,e) = (−1)m s2−1
8 ζ s
G(q2,o) = −(−1)
8 (−1)m s2−1
r2−1
8 ζ r−s
8 G(q1,o).
Direct computation shows that (FS2) gives no restriction.
For the smallest example G = Z8, Γ = Z16, there are four possibilities for (r, s):
(1,−3), (−1, 3), (3, 3), and (−3,−3).
INFINITE FAMILIES OF MODULAR DATA
25
3.2.6.2. Assume n > 3. In this case Γe = 8 and Γo = 2n−3Go + 1. Assume
Γe = Z8, q2,e(x) = ζ sx2
8G(q1,o) and
G(q2) = (−1)
16 , and θ2,e(x) = −x. Then G(q1) = (−1)n r2−1
8 ζ r
8 ζ s
s2−1
8G(q2,o), and hence
(−1)n r2−1
8 ζ r
8G(q1,o) + (−1)
8G(q2,o) = 0.
s2−1
8 ζ s
Direct computation shows that (FS2) gives no restriction.
For the smallest example G = Z16, Γ = Z8 × Z3, there are four possibilities, we
have r − s ≡ 2 mod 4, and G(q2,o) = (−1)
s2−1
8 ζ r−s
.
8
4. Near-group over Z2 family
4.1. General formulae. Throughout this section we assume that (G, q1) and (Γ, q2)
are metric groups satisfying G2 ∼= Γ2 ∼= Z2 and c := G(q1) = −G(q2). We denote G2 =
{0, g0}, Γ2 = {0, γ0}, and assume that hg0, g0i = −hγ0, γ0i and c2 = q1(g0)−1q2(γ0).
We set
q1(g0)−1 + q2(γ0)−1
.
s = c
√2
Since q1(g0)2 = hg0, g0i and q2(γ0)2 = hγ0, γ0i, we get
s2 = c2hg0, g0i + 2q1(g0)−1q2(γ0)−1 + hγ0, γ0i
2
= q1(g0)−2 = hg0, g0i.
We set a = 1/pG, b = 1/pΓ. We choose and fix subsets G∗ ⊂ G and Γ∗ ⊂ Γ
satisfying
We set
G = G2 ⊔ G∗ ⊔ −G∗, Γ = Γ2 ⊔ Γ∗ ⊔ −Γ∗.
J = {0, π} ⊔ ({g0} × {1,−1}) ⊔ G∗ ⊔ ({γ0} × {1,−1}) ⊔ Γ∗.
We denote J1 = {g0}×{1,−1} and J2 = {γ0}×{1,−1}. We use letters, g, g′, g′′, h· · ·
for elements of G, and γ, γ′, γ′′, ξ · · · for elements of Γ.
We introduce an involution of J by setting (g0, ε) = (g0,hg0, g0iε), (γ0, ε) =
(γ0,hg0, g0iε), and leaving the other indices fixed.
Definition 4.1. Under the above assumptions, we define J by J matrices S, T , and
C by
S =
a−b
2
a+b
2
a
2
a
b
2
b
a+b
2
a−b
2
a
2
a
− b
−b
2
a
2
a
2
2 + εε′ s
( a
2√2
ahg, g0i
)hg0, g0i
εε′ s
2√2
0
ahg0, g′i
a(hg, g′i + hg, g′i)
a
a
0
0
b
2
− b
2
εε′ s
2√2
0
2 − εε′ s
2√2
−( b
)hγ0, γ0i
−bhγ, γ0i
b
−b
0
0
−bhγ0, γ′i
−b(hγ, γ′i + hγ, γ′i)
,
T = Diag(1, 1, q1(g0), q1(g), q2(γ0), q2(γ)),
and Cj,j′ = δj,j′ = δj,j′ (where the six blocks in S and T are indexed by {0},{π},J1,
G∗, J2, and Γ∗, respectively).
Direct computation shows the following.
26
PINHAS GROSSMAN AND MASAKI IZUMI
Lemma 4.2. Let the notation be as in Definition 4.1. Then S, T , and C are unitary
matrices satisfying S2 = C, (ST )3 = cC and T C = CT .
Lemma 4.3. Let the notation be as in Definition 4.1. Then we have
4
2
4
Nπ,π,π =
, Nπ,π,(g0,ε) =
, Nπ,π,g =
2
Γ − G
1
+
Γ − G
1
Γ − G
1
2
εε′
2
Γ − G
Nπ,π,(γ0,ε) =
Nπ,(g0,ε),(g0,ε′) =
Nπ,(g0,ε),(γ0,ε′) =
4
1
Γ − G
Γ − G −
N(g0,ε),(g0,ε′),(g′0,ε′′) =
1
2
N(g0,ε),(g0,ε′),g =
,
Γ − G
Γ − G
, Nπ,π,γ =
4
Γ − G
,
2
, Nπ,(g0,ε),g =
,
Γ − G
2
Γ − G
,
4
+
, Nπ,(g0,ε),γ =
2
Γ − G
2
Γ − G
+
4
Γ − G
Γ − G − δγ,γ′,
Nπ,g,g′ =
+ δg,g′, Nπ,g,(γ0,ε) =
, Nπ,g,γ =
,
Nπ,(γ0,ε),(γ0,ε′) =
, Nπ,(γ0,ε),γ =
, Nπ,γ,γ′ =
N(g0,ε),(g0,ε′),(γ0,ε′′) =
N(g0,ε),(g0,ε′),γ =
N(g0,ε),g,g′ =
N(g0,ε),g,(γ0,ε′) =
N(g0,ε′′),(γ0,ε),(γ0,ε′) =
1
εε′ + ε′ε′′ + ε′′ε
,
4
2(Γ − G)
+
1
Γ − G
1
+
2(Γ − G)
Γ − G −
1
2
,
2
εε′hg0 + g, g0i
ε′′(ε + ε′)hg0, g0i + εε′
εε′hg0, g0ihγ0, γi
4
,
,
2
Γ − G
1
+
+ δg0+g+g′,0 + δg0+g−g′,0,
εε′hg, g0i
2
Γ − G
1
2(Γ − G)
N(g0,ε),γ,γ′ =
2
,
, N(g0,ε),g,γ =
Γ − G
εε′ − ε′′(ε + ε′)hg0, g0i
,
2
4
+
,
Γ − G
1
+
εε′hγ, γ0i
,
2
Γ − G
N(g0,ε),(γ0,ε′),γ =
+ δg+g′+g′′,0 + δ−g+g′+g′′,0 + δg−g′+g′′,0 + δg+g′−g′′,0,
Ng,g′,(γ0,ε) =
, Ng,g′,γ =
2
Γ − G
1
Γ − G
4
Γ − G
εε′hg0 + g, g0i
+
2
,
,
Ng,(γ0,ε),(γ0,ε′) =
Ng,g′,g′′ =
4
Γ − G
INFINITE FAMILIES OF MODULAR DATA
27
Ng,(γ0,ε),γ =
, Ng,γ,γ′ =
2
Γ − G
1
2(Γ − G)
+
1
Γ − G
2
4
,
Γ − G
εε′ + ε′ε′′ + ε′′ε
,
+
εε′hγ0 + γ, γ0i
4
,
2
N(γ0,ε),(γ0,ε′),(γ0,ε′′) =
N(γ0,ε),(γ0,ε′),γ =
N(γ0,ε),γ,γ′ =
Γ − G − δγ0+γ+γ′,0 − δγ0+γ−γ′,0,
4
Nγ,γ′,γ′′ =
Γ − G − δγ+γ′+γ′′,0 − δ−γ+γ′+γ′′,0 − δγ−γ′+γ′′,0 − δγ+γ′−γ′′,0.
Theorem 4.4. Let the notation be as in Definition 4.1. Then all the fusion coeffi-
cients Nijk are non-negative integers if and only if Γ − G = 2.
We assume that Γ = G + 2 in the rest of this section.
Conjecture 4.5. Let A be an odd abelian group and let G = Z2n × A with n ≥ 1.
Then the pair
(S(G,q1) ⊗ S, T (G,q1) ⊗ T )
is the modular data of the Drinfeld center of the Z2-de-equivariantization of a near-
group category for Z2n+1 × A with multiplicity 2n+1A.
The conjecture is true for n = 1 and trivial A (see the online appendix). More
generally, we have the following.
Conjecture 4.6. Let C be a quadratic category of type (Z2, Q, 1) with non-self-dual ρ
and with the VecZ2 subcategory having non-trivial associator. Then the modular data
of the Drinfeld center Z(C) is given by
(SG,q1 ⊗ S, T (G,q1) ⊗ T )
with G = Z2 × Q.
Note that we have
Nπ,π,π = 2, N(g0,ε),(g0,ε),(g0,ε) = 1, Ng,g,g = 2 + δ3g,0,
Direct computation shows the following.
N(γ0,ε),(γ0,ε),(γ0,ε) = 1, Nγ,γ,γ = 2 − δ3γ,0.
Lemma 4.7. We have
νm(π) = G(q1, m) + G(q2, m)2
2
,
νm((g0, ε)) = G(q1, m) + G(q2, m)2
νm(g) = G(q1, m) + G(q2, m)2
νm((γ0, ε)) = G(q1, m) + G(q2, m)2
2
4
4
+
δmg0,0q1(g0)m
2
+ δmg,0q1(g)m,
δmγ0,0q1(γ0)m
2
−
,
,
28
PINHAS GROSSMAN AND MASAKI IZUMI
νm(γ) = G(q1, m) + G(q2, m)2
2
− δmγ,0q2(γ)m.
In particular, (FS2) and (FS3) are equivalent to the following conditions, respectively:
(4.1)
(4.2)
G(q1, 2) + G(q2, 2) = √2,
G(q1, 3) + G(q2, 3) = 2.
4.2. Examples. We give a list of metric groups (G, q1), (Γ, q2) satisfying G2 =
{0, g0} ∼= Γ2 = {0, γ0} ∼= Z2, Γ = G + 2, c := G(q1) = −G(q2), c2 = q1(g0)−1q2(γ0),
hg0, g0i = −hγ0, γ0i, and Eq.(4.1)-(4.2).
A similar computation as in the proof of Lemma 3.12 shows that Eq.(4.1) is equiv-
alent to
G(q1,e, 2)
G(q1,e)
(−1) Go2−1
8 − G(q2,e, 2)
G(q2,e)
and Eq.(4.2) is equivalent to
(−1) Γo2−1
8
= √2,
(−1)n1 G(q1,o, 3)
G(q1,o)3 − (−1)n2 G(q2,o, 3)
G(q2,o)3 = 2,
where Ge = 2n1 and Γe = 2n2. Moreover, if neither G nor Γ has a 3-component,
Eq.(4.2) is further equivalent to
(−1)n1(Go
3
) − (−1)n2(Γo
3
) = 2.
Direct computation shows that the restriction coming from (FS3) in the following
examples is very similar to that discussed in Lemma 3.9 and Remark 3.10, and we
will not mention it in what follows.
We consider separately the cases Ge = Z2 and Ge = Z2n with n > 1.
4.2.1. Assume Ge = Z2. In this case there exists n ≥ 2 satisfying Γe = Z2n and
Go = 2n−1Γo − 1. We have n = 2 for Go ≡ 1 mod 4, and n ≥ 3 for Go ≡ 3
mod 4. We may assume q1,e(x) = ir1x2 with r1 ∈ {1,−1} and q2,e(x) = ζ r2x2
2n+1 with
r2 ∈ {1,−1, 3,−3}. Then G(q1,e) = ζ r1
, and
r2
2−1
8
c = ζ r1
8 G(q1,o) = −ζ r2
Since G(q1,e, 2) = 0 and G(q2,e, 2) = √2ζ r2
(±1,∓3). These two modular data
8 , G(q2,e) = ζ r2
8 (−1)n
8 (−1)(n−1)
8 (−1)n
8 G(q2,o).
r2
2−1
r2
2−1
8
For the smallest example G = Z2, Γ = Z4, there are two possibilities: (r1, r2) =
, Eq.(4.1) always holds.
(S(Z2,q1) ⊗ S, T (Z2,q1) ⊗ T ),
are the modular data of the Drinfeld center of the two categories constructed in
[Izu17], [LMP15] (see the online appendix for a proof of this).
INFINITE FAMILIES OF MODULAR DATA
29
4.2.2. Assume Ge = Z2n with n ≥ 2. In this case we have Γe = Z2 and Γo =
2n−1Go + 1. Thus Γo ≡ 3 mod 4 for n = 2, and Γo ≡ 1 mod 4 for n ≥ 3.
We may assume q1,e(x) = ζ r1x2
2n+1 with r1 ∈ {1,−1, 3,−3} and q2,e(x) = ir2x2. Then
G(q1,e) = ζ r1
8 (−1)n
8 , and
r2
1−1
8
8 G(q1,o) = −ζ r2
As above, Eq.(4.1) always holds. For the smallest example G = Z4, Γ = Z6, there
are four possibilities determined by G(q2,o) = −ζ r1−r2
and r1 − r2 ≡ 2 mod 4.
8 G(q2,o).
8
, G(q2,e) = ζ r2
c = ζ r1
8 (−1)n
r2
1−1
5. Even generalized Haagerup family
5.1. General formulae. Throughout this section we assume that (G, q1, θ1) and
(Γ, q2, θ1) are involutive metric groups satisfying the following conditions:
c := G(q1) = −G(q2),
Gθ1 = Γθ2 = 4,
and there exist order 2 elements k0 ∈ Gθ1 \ {0}, σ0 ∈ Γθ2 \ {0} with q1(k0) = q2(σ0).
We identify {0, k0} and {0, σ0}, and denote them by U = {0, u0}, which is regarded
as a common subgroup of Gθ1 and Γθ2. We denote by q0 the restriction of q1 (and
q2 too) to U. We denote by h·,·ii the bicharacter associated with qi, and we often
suppress subscript i in h·,·ii and θi if there is no possibility of confusion.
We denote K∗ := Γ2 \ {0, k0} = {k1, k2} and Σ∗ := Γ2 \ {0, σ0} = {σ1, σ2}. We
assume that one of the following holds:
(A1) q1(k1) = q1(k2) and q2(σ1) = −q2(σ2). In this case, we have
hk0, k1i = hk0, k2i = q0(u0)−1 and hσ0, σ1i = hσ0, σ2i = −q0(u0)−1.
We set s = c/q1(k1), and assume s4 = 1.
(A2) q1(k1) = −q1(k2) and q2(σ1) = q2(σ2). In this case, we have
hk0, k1i = hk0, k2i = −q0(u0)−1 and hσ0, σ1i = hσ0, σ2i = q0(u0)−1.
We set s = c/q2(σ1), and assume s4 = 1.
In either case, we have
(5.1)
(5.2)
(5.3)
(5.4)
hu0, u0i = 1,
hk1, k1i = hk2, k2i,
hσ1, σ1i = hσ2, σ2i,
q0(u0) ∈ {1,−1}.
s(q1(k1) + q1(k2) + q2(σ1) + q2(σ2)) = 2c.
Indeed, since k0 + k1 = k2, Eq.(5.1) is an easy consequence of the assumptions (A1)
and (A2) above. Since hg,−gi = q1(0)/q1(g)2, we have q1(g)2 = hg, gi, and so Eq.(5.2)
and Eq.(5.3) hold. Eq.(5.4) is obvious.
30
PINHAS GROSSMAN AND MASAKI IZUMI
Later, we will assume
(5.5)
and
(1 + hu0, ki)(1 − s2hk, ki) = 0,
∀k ∈ K∗,
(5.6)
(1 + hu0, σi)(1 + s2hσ, σi) = 0,
too, but for the moment we do not assume these.
∀σ ∈ Σ∗,
We set a = 1/pG, b = 1/pΓ. We choose and fix subsets G∗ ⊂ G and Γ∗ ⊂ Γ
satisfying
and set
G = G2 ⊔ G∗ ⊔ θ1(G∗), Γ = Γ2 ⊔ Γ∗ ⊔ θ2(Γ∗),
J = (U × {0, π}) ⊔ (K∗ × {1,−1}) ⊔ G∗ ⊔ (Σ∗ × {1,−1}) ⊔ Γ∗.
We denote J1 = K∗ × {1,−1} and J2 = Σ∗ × {1,−1}. We use letters u, u′, u′′, v,· · ·
for elements in U, k, k′, k′′, l · · · for elements of K∗, g, g′, g′′, h· · · for elements of G,
σ, σ′σ′′, τ,· · · for elements in Σ∗, and γ, γ′, γ′′, ξ · · · for elements of Γ.
We choose a function f : K∗×Σ∗ → {s,−s} satisfying f (k +k0, σ) = f (k, σ)hu0, σi,
f (k, σ + σ0) = f (k, σ)hk, u0i. Such a choice is possible thanks to q1(k0) = q2(σ0).
(We may assume f (k1, σ1) = s if it is convenient.) We have f (k, σ)2 = s2 for any
choice of f . We introduce an involution of J by setting (u, 0) = (u, 0), (u, π) = (u, π),
(k, ε) = (−k, s2ε), (σ, ε) = (−σ, s2ε), and
g =(cid:26) −g,
θ1(−g),
γ =(cid:26) −γ,
θ2(−γ),
if −g ∈ G∗
otherwise
if −γ ∈ Γ∗
otherwise
,
.
Definition 5.1. Under the above assumptions, we define J by J matrices S, T , and
C by
S =
a
a−b
2 hu, u′i
a+b
2 hu, u′i
2hk, u′i
ahg, u′i
2hσ, u′i
bhγ, u′i
b
a+b
2 hu, u′i
a−b
2 hu, u′i
a
2 hk, u′i
ahg, u′i
− b
2hσ, u′i
−bhγ, u′i
a
2hu, k′i
2hu, k′i
a
2 + εε′s
( a
4 )hk, k′i
ahg, k′i
εε′ f (k′,σ)
4
0
ahu, g′i
ahu, g′i
ahk, g′i
0
0
a(hg, g′i + hg, θ(g′)i)
b
2hu, σ′i
− b
2hu, σ′i
εε′ f (k,σ ′)
4
0
−( b
2 + −εε′ s
−bhγ, σ′i
4
)hσ, σ′i
bhu, γ′i
−bhu, γ′i
0
0
−bhσ, γ′i
−b(hγ, γ′i + hγ, θ(γ′)i)
,
T = Diag(q0(u), q1(k), q1(g), q2(σ), q2(γ)),
and Cj,j′ = δj,j′ = δj,j′ (where the six blocks in S and T are indexed by U × {0},
U × {π}, J1, G∗, J2, and Γ∗, respectively. )
Remark 5.2. As an alternative to assumptions (A1) and (A2) above, we could con-
sider a third possibility corresponding to hu0, u0i = −1 together with an appropriate
condition. In this case the quadratic form q0 is non-degenerate and S and T factorize
as S = S(U,q0) ⊗ S′ and T = T (U,q0) ⊗ T ′. The pair (S′, T ′) arising in this way is none
other than the one that we discussed in the previous section. It is more convenient
to treat this case separately.
Lemma 5.3. Let the notation be as above. Then the following equations hold.
INFINITE FAMILIES OF MODULAR DATA
31
Xl∈K∗
hk′ − k, li + Xσ∈Σ∗
hσ′ − σ, τi + Xk∈K∗
Xτ∈Σ∗
sXl∈K∗
hk, lif (l, σ) + s Xτ∈Σ∗
f (k, σ)f (k′, σ) = 4δk,k′.
f (k, σ)f (k, σ′) = 4δσ,σ′,
hσ, τif (k, τ ) = 0.
s2 Xl∈K∗
hk + k′, liq1(l) + Xσ∈Σ∗
f (k, σ)f (k′, σ)q2(σ) = 2cshk, k′iq1(k)q1(k′).
s(Xl∈K∗
f (l, σ)hk, liq1(l) + Xτ∈Σ∗
f (k, τ )hσ, τiq2(τ )) = 2cf (k, σ)q1(k)q2(σ).
(1)
(2)
(3)
(4)
(5)
(6)
Xl∈K∗
f (l, σ)f (l, σ′)q1(l) + s2 Xτ∈Σ∗
Proof. (1) Let u = k′ − k ∈ U. Then
hσ + σ′, τiq2(τ ) = 2cshσ, σ′iq2(σ)q2(σ′).
Xl∈K∗
hk′ − k, li + Xσ∈Σ∗
f (k, σ)f (k′, σ) = Xl∈K∗
hu, li + Xσ∈Σ∗
hu, σi = 4δu,0.
Eq. (2) can be shown in the same way.
(3) Note that we have sf (k, σ) = sf (k, σ).
hσ, τif (k, τ )
hk, lif (l, σ) + s Xτ∈Σ∗
hk, k + uif (k + u, σ) + sXv∈U
hk, k + uif (k, σ)hu, σi + sXv∈U
sXl∈K∗
= sXu∈U
= sXu∈U
= sf (k, σ)(hk, kiXu∈U
hk, uihu, σi + hσ, σiXv∈U
hσ, σ + vif (k, σ + v)
hσ, σ + vif (k, σ)hk, vi
hσ, vihk, vi)
= sf (k, σ)(hk, ki + hσ, σi)(1 + hk, u0ihσ, u0i) = 0.
32
PINHAS GROSSMAN AND MASAKI IZUMI
(4) Let u = k′ − k.
f (k, σ)f (k′, σ)q2(σ)
hk + k′, liq1(l) + Xσ∈Σ∗
q1(k + k′ − l)q1(k + k′) + Xσ∈Σ∗
q1(l) + Xσ∈Σ∗
q1(l) + hk, 2k′i Xσ∈Σ∗
s2 Xl∈K∗
= s2 Xl∈K∗
= s2(hk, k′iq1(k)q1(k′)Xl∈K∗
= hk, k′iq1(k)q1(k′)s2(Xl∈K∗
= hk, k′iq1(k)q1(k′)s2(q1(k1) + q2(k2) + hk, 2k′i(q2(σ1) + q2(σ2))).
f (k, σ)2hu, σiq2(σ)
q2(u − σ)q2(u))
q2(σ))
If Gθ1 ⊂ G2, the statement follows from Eq.(5.4). If not, we have Gθ1 ∼= Z4, and
k1 = −k2. Hence q1(k1) = q1(k2) and q2(σ1) + q2(σ2) = 0, and the statement holds
too. Eq. (6) can be shown in the same way.
(5) We have:
f (k, τ )hσ, τiq2(τ ))
f (k, τ )q2(σ − τ )q2(σ))
f (k, σ + v)q2(v))
f (l, σ)hk, liq1(l) + Xτ∈Σ∗
f (l, σ)q1(l − k)q1(k) + Xτ∈Σ∗
s(Xl∈K∗
= s(Xl∈K∗
= s(q1(k)Xu∈U
= sf (k, σ)(q1(k)Xu∈U
= sf (k, σ)(q1(k)Xu∈U
f (k + u, σ)q1(u) + q2(σ)Xv∈U
hu,−σiq0(u) + q2(σ)Xv∈U
q2(u − σ)q2(σ) + q2(σ)Xv∈U
h−k, viq0(v))
q1(v − k)q1(k))
= sf (k, σ)q1(k)q2(σ)(q1(k1) + q1(k2) + q2(σ1) + q2(σ2))
= 2cf (k, σ)q1(k)q2(σ).
(cid:3)
Using Lemma 5.3, we can show the following by direct computation.
Lemma 5.4. Let the notation be as in Definition 5.1. Then S, T , and C are unitary
matrices satisfying S2 = C, (ST )3 = cC and T C = CT .
Lemma 5.5. Let the notation be as in Definition 5.1. Then we have
N(u,0),(u′,0),(u′′,0) = N(u,0),(u′,π),(u′′,π) = δu+u′+u′′,0,
N(u,0),(u′,0),(u′′,π) = N(u,0),(u′,0),(k,ε) = N(u,0),(u′,0),g = N(u,0),(u′,0),(σ,0) = N(u,0),(u′,0),γ = 0,
N(u,0),(u′,π,),(k,ε) = N(u,0),(u′,π,),g = N(u,0),(u′,π,),(σ,ε) = N(u,0),(u′,π),γ = 0,
N(u,0),(k,ε),(k′,ε′) = δu+k+k′,0δε,ε′, N(u,0),g,g′ = δu+g+g′,0 + δu+g−g′,0,
N(u,0),(σ,ε),(σ′,ε′) = δu+σ+σ′,0δε,ε′, N(u,0),γ,γ′ = δu+γ+γ′,0 + δu+γ−γ′,0,
INFINITE FAMILIES OF MODULAR DATA
33
N(u,0),(k,ε),g = N(u,0),(k,ε),(σ,ε′) = N(u,0),(k,ǫ),γ = N(u,0),g,(σ,ε) = N(u,0),g,γ = N(u,0),(σ,ε),γ = 0,
N(u,π),(u′,π),(u′′,π) =
N(u,π),(u′,π),(k,ε) =
N(u,π),(u′,π),g =
N(u,π),(u′,π),(σ,ε) =
N(u,π),(u′,π),γ =
8
Γ − G
,
2
Γ − G
4
Γ − G
2
Γ − G
4
Γ − G
(1 + hk, u0i),
(1 + hg, u0i),
(1 + hσ, u0i),
(1 + hγ, u0i),
N(u,π),(k,ε),(k,ε′)
=
2
Γ − G
+
δu+k+k′,0
2
+
εε′s2
4
N(u,π),(k,ε),g =
(hu + k + k′, k1i − hu + k + k′, σ1ih2k, σ1i).
N(u,π),g,g′ =
(1 + hg + g′, u0i) + δu+g+g′,0 + δu+g+θ(g′),0,
2
(1 + hk + g, u0i),
Γ − G
N(u,π),(k,ε),(σ,ε′) = 0,
2
Γ − G
(1 + hk, u0ihγ, u0i),
2
Γ − G
4
Γ − G
(1 + hg, u0ihσ, u0i),
(1 + hg, u0ihγ, u0i),
N(u,π),(k,ε),γ =
4
Γ − G
N(u,π),g,(σ,ε) =
N(u,π),g,γ =
N(u,π),(σ,ε),(σ′,ε′)
=
2
Γ − G −
δu+σ+σ′,0
2
+
εε′s2
4
(hu + σ + σ′, k1ih2σ, k1i − hu + σ + σ′, σi),
N(u,π),(σ,ε),γ =
2
Γ − G
(1 + hσ + γ, u0i),
N(u,π),γ,γ′ =
4
Γ − G
(1 + hγ + γ′, u0i) − δu+γ+γ′,0 − δu+γ+θ(γ′),0,
N(k,ε),(k′,ε′),(k′′,ε′′)
= (1 + hu0, k + k′ + k′′i)(
1
2(Γ − G)
+
s2(εε′ + ε′ε′′ + ε′′ε)
8
hk + k′ + k′′, k + k′ + k′′i),
N(k,ε),(k′,ε′),g = (
1
Γ − G
+
s2εε′hk, k + k′ + gi
4
)(1 + hg, u0i),
34
PINHAS GROSSMAN AND MASAKI IZUMI
N(k,ε),(k′,ε′),(σ,ε′′)
= (1 + hσ, u0i)(
1
2(Γ − G)
ε′′(ε + ε′)s
+
8
N(k,ε),(k′,ε),γ = (1 + hγ, u0i)(
N(k,ε),g,g′ =
2
Γ − G
N(k,ε),g,(σ,ε′) = (
Γ − G
2
N(k,ε),g,γ =
s2εε′
8 hk − k′, σihσ, σi),
hk + k′, kif (k, σ) −
1
Γ − G −
εε′s2
4 hk − k′, σ1ihγ, σ1i),
(1 + hk + g + g′, u0i) + δk+g+g′,0 + δk+g+θ(g′),0,
1
εε′sf (k, σ)hk + g, ki
+
)(1 − hg, u0i),
4
(1 + hk + g, u0ihγ, u0i),
Γ − G
)(1 + hk, u0i),
N(k,ε′′),(σ,ε),(σ′,ε′)
= (
1
2(Γ − G)
+
N(k,ε),(σ,ε′),γ = (
εε′s2hσ − σ′, kihk, ki
sε′′(ε + ε′)f (k, σ)hσ + σ′, σi
8
1
−
8
εε′sf (k, σ)hσ + γ, σi
2
4
Γ − G −
Γ − G
(1 + hg + g′ + g′′, u0i)
4
N(k,ε),γ,γ′ =
(1 + hk, u0ihγ + γ′, u0i),
)(1 − hγ, u0i),
+ δg+g′+g′′,0 + δθ(g)+g′+g′′,0 + δg+θ(g′)+g′′,0 + δg+g′+θ(g′′),0,
Ng,g′,g′′ =
Γ − G
Ng,(σ,ε),(σ′,ε′) = (
)(1 + hg, u0i).
Ng,g′,(σ,ε) =
Ng,g′,γ =
2
Γ − G
4
Γ − G
1
+
Γ − G
2
(1 + hg + g′, u0ihσ, u0i),
(1 + hg + g′, u0ihγ, u0i),
εε′s2hg, k1ihk1, σ − σ′i
4
Ng,(σ,ε),γ =
Ng,γ,γ′ =
(1 + hg, u0ihσ + γ, u0i),
(1 + hg, u0ihγ + γ′, u0i),
Γ − G
4
Γ − G
N(σ,ε),(σ′ ,ε′),(σ′′,ε′′)
= (1 + hσ + σ′ + σ′′, u0i)(
1
2(Γ − G) −
s2(εε′ + ε′ε′′ + ε′′ε)
8
hσ + σ′ + σ′′, σ + σ′ + σ′′i),
N(σ,ε),(σ′,ε′),γ = (
1
Γ − G −
εε′s2hσ + σ′ + γ, σi
4
)(1 + hσ + σ′ + γ, u0i),
INFINITE FAMILIES OF MODULAR DATA
35
(1 + hσ + γ + γ′, u0i) − δσ+γ+γ′,0 − δσ+γ+θ(γ′),0,
N(σ,ε),γ,γ′ =
Nγ,γ′,γ′′ =
2
Γ − G
4
Γ − G
(1 + hγ + γ′ + γ′′, u0i)
− δγ+γ′+γ′′,0 − δθ(γ)+γ′+γ′′,0 − δγ+θ(γ′)+γ′′,0 − δγ+γ′+θ(γ′′),0.
From the above computations (in particular from the formulas for N(k,ε),(k′,ε′),(k",ε′′)
and N(σ,ε),(σ′,ε′),(σ",ε′′)), we find the following.
Theorem 5.6. Under the assumptions of Definition 5.1, all the fusion coefficients
are non-negative integers if and only if the following five conditions are satisfied.
(1) Γ = G + 4.
(2) (1 + hu0, ki)(1 − s2hk, ki) = 0 for all k ∈ K∗.
(3) (1 + hu0, σi)(1 + s2hσ, σi) = 0 for all σ ∈ Σ∗.
(4) If Gθ1 ∼= Z4, we have q0(u0) = −1.
(5) If Γθ2 ∼= Z4, we have q0(u0) = −1.
Under these conditions, the fusion coefficients are
N(u,0),(u′,0),(u′′,0) = N(u,0),(u′,π),(u′′,π) = δu+u′+u′′,0,
N(u,0),(u′,0),(u′′,π) = N(u,0),(u′,0),(k,ε) = N(u,0),(u′,0),g = N(u,0),(u′,0),(σ,0) = N(u,0),(u′,0),γ = 0,
N(u,0),(u′,π,),(k,ε) = N(u,0),(u′,π,),g = N(u,0),(u′,π,),(σ,ε) = N(u,0),(u′,π),γ = 0,
N(u,0),(k,ε),(k′,ε′) = δu+k+k′,0δε,ε′, N(u,0),g,g′ = δu+g+g′,0 + δu+g−g′,0,
N(u,0),(σ,ε),(σ′,ε′) = δu+σ+σ′,0δε,ε′, N(u,0),γ,γ′ = δu+γ+γ′,0 + δu+γ−γ′,0,
N(u,0),(k,ε),g = N(u,0),(k,ε),(σ,ε′) = N(u,0),(k,ǫ),γ = N(u,0),g,(σ,ε) = N(u,0),g,γ = N(u,0),(σ,ε),γ = 0,
N(u,π),(u′,π),(u′′,π) = 2, N(u,π),(u′,π),(k,ε) =
1 + hσ, u0i
εε′s2
N(u,π),(u′,π),(σ,ε) =
1 + δu+k+k′,0
2
N(u,π),(k,ε),(k,ε′) =
N(u,π),(k,ε),g =
2
1 + hk + g, u0i
1 + hk, u0i
2
, N(u,π),(u′,π),g = 1 + hg, u0i,
, N(u,π),(u′,π),γ = 1 + hγ, u0i,
(hu + k + k′, k1i − hu + k + k′, σ1i),
4
+
1 + hk, u0ihγ, u0i
2
,
2
, N(u,π),(k,ε),(σ,ε′) = 0, N(u,π),(k,ε),γ =
N(u,π),g,g′ = 1 + hg + g′, u0i + δu+g+g′,0 + δu+g+θ(g′),0,
N(u,π),g,(σ,ε) =
, N(u,π),g,γ = 1 + hg, u0ihγ, u0i,
1 + hg, u0ihσ, u0i
1 − δu+σ+σ′,0
+
2
εε′s2
N(u,π),(σ,ε),(σ′,ε′) =
N(u,π),(σ,ε),γ =
1 + hσ + γ, u0i
2
N(k,ε),(k′,ε′),(k′′,ε′′) =
N(k,ε),(k′,ε′),g =
4
(hu + σ + σ′, k1i − hu + σ + σ′, σi,
2
, N(u,π),γ,γ′ = 1 + hγ + γ′, u0i − δu+γ+γ′,0 − δu+γ+θ(γ′),0,
(1 + hu0, k + k′ + k′′i)(1 + εε′ + ε′ε′′ + ε′′ε)
(1 + s2εε′hk, k + k′ + gi)(1 + hg, u0i)
8
,
,
4
36
PINHAS GROSSMAN AND MASAKI IZUMI
N(k,ε),(k′,ε′),(σ,ε′′) =
(1 + hσ, u0i)(1 + εε′ + ε′′(ε + ε′)hk + k′, kisf (k, σ))
8
,
N(k,ε),(k′,ε),γ =
(1 + hγ, u0i)(1 − εε′s2hk − k′, σ1ihγ, σ1i)
4
,
N(k,ε),g,g′ =
1 + hk + g + g′, u0i
2
+ δk+g+g′,0 + δk+g+θ(g′),0,
(1 + εε′sf (k, σ)hk + g, ki)(1 − hg, u0i)
,
N(k,ε),g,(σ,ε′) =
1 + hk + g, u0ihγ, u0i
N(k,ε),g,γ =
(1 + εε′ − ε′′(ε + ε′)sf (k, σ)hσ + σ′, σi)(1 + hk, u0i)
2
,
8
,
N(k,ε′′),(σ,ε),(σ′ ,ε′) =
N(k,ε),(σ,ε′),γ =
(1 − εε′sf (k, σ)hσ + γ, σi)(1 − hγ, u0i)
4
,
N(k,ε),γ,γ′ =
1 + hk, u0ihγ + γ′, u0i
,
2
4
4
Ng,g′,g′′ = 1 + hg + g′ + g′′, u0i + δg+g′+g′′,0 + δθ(g)+g′+g′′,0 + δg+θ(g′)+g′′,0 + δg+g′+θ(g′′),0,
Ng,g′,(σ,ε) =
1 + hg + g′, u0ihσ, u0i
Ng,(σ,ε),(σ′,ε′) =
Ng,(σ,ε),γ =
1 + hg, u0ihσ + γ, u0i
, Ng,g′,γ = 1 + hg + g′, u0ihγ, u0i,
2
(1 + εε′s2hg, k1ihk1, σ − σ′i)(1 + hg, u0i)
,
N(σ,ε),(σ′,ε′),(σ′′,ε′′) =
N(σ,ε),(σ′,ε′),γ =
N(σ,ε),γ,γ′ =
, Ng,γ,γ′ = 1 + hg, u0ihγ + γ′, u0i,
2
(1 + hσ + σ′ + σ′′, u0i)(1 + εε′ + ε′ε′′ + ε′′ε))
,
(1 − εε′s2hσ + σ′ + γ, σi)(1 + hσ + σ′ + γ, u0i)
1 + hσ + γ + γ′, u0i
− δσ+γ+γ′,0 − δσ+γ+θ(γ′),0,
8
4
2
,
Nγ,γ′,γ′′ = 1 +hγ + γ′ + γ′′, u0i− δγ+γ′+γ′′,0 − δθ(γ)+γ′+γ′′,0 − δγ+θ(γ′)+γ′′,0 − δγ+γ′+θ(γ′′),0.
Proof. Note that G and Γ are multiples of 4. Since hk, u0ihσ, u0i = −1, either
N(π,u),(π,u′),(k,ε) or N(π,u),(π,u′),(σ,ε) is
, and so (1) is necessary for the fusion coef-
ficients to be non-negative integers. We assume (1) for the rest of the proof.
Γ−G
4
Necessity of (2) and (3) follows from the formulas for N(k,ε),(k,ε),(−k,−ε) and N(σ,ε),(σ,ε),(−σ,−ε)
respectively. For (4), we have
N(u,π),(k,ε),(k′,ε′) =
1 + δu+k+k′
2
+
εε′s2
4
(hu + k + k′, k1i − hu + k + k′, σ1ih2k, σi),
which is 1 + εε′s2
4 (1 − h2k, σi) if u + k + k′ = 0, and is
εε′s2
(hu0, k1i − hu0, σ1ih2k, σi) =
4
1
2
+
1
2
εε′s2hu0, k1i
+
4
(1 + h2k, σi),
INFINITE FAMILIES OF MODULAR DATA
37
if u+k+k′ = u0. In the both cases N(u,π),(k,ε),(k′,ε′) is a non-negative integer if and only
if h2k, σi = 1. If Gθ1 ∼= Z2 × Z2, this does not give any restriction. Assume Gθ1 ∼= Z4.
Then we have 2k = u0. Since k1 = −k2, we have q1(k1) = q2(k2), and we are in the
case (A1). Thus h2k, σi = hu0, σi = −1/q0(u0). This shows that h2k, σi = 1 if and
only if q0(u0) = −1, and we get (4). Necessity of (5) follows from the formula for
N(u,π),(σ,ε),(σ,ε′).
Assume (1)-(5) on the other hand. Then it is easy to show that the fusion coeffi-
cients are non-negative integers except for
N(k,ε),(k′,ε′),(σ,ε′′), N(k,ε),g,(σ,ε′), N(k,ε),(σ,ε′),(σ′,ε′′), N(k,ε),(σ,ε′),γ.
We show that the first two are non-negative integers, and the last two can be treated
in the same way. We have
1 + ε′′(ε + ε′)sf (k, σ)hk + k′, ki − εε′hk′ − k, σis2hσ, σi
N(k,ε),(k′,ε′),(σ,ε′′) = (1+hu0, σi)
If hu0, σi = −1, there is nothing to show, so we assume hu0, σi = 1. Then hk′ − k, σi =
1 and (3) implies that s2hσ, σi = −1. Thus
8
.
1 + εε′ + ε′′(ε + ε′)sf (k, σ)hk + k′, ki
N(k,ε),(k′,ε′),(σ,ε′′) =
4
(1 + εε′)(1 + εε′′sf (k, σ)hk + k′, ki)
,
=
4
which is either 0 or 1.
Since
N(k,ε),g,(σ,ε′) =
(1 − hg, u0i)(1 + εε′sf (k, σ)hk + g, ki)
4
,
to show that this is a non-negative integer, it suffices to show that hg, u0i = −1
implies that hk + g, ki ∈ {1,−1}, or equivalently hk + g, ki2 = 1. If Gθ1 ∼= Z2 × Z2,
there is nothing to show. Assume Gθ1 ∼= Z4. Since 2k = u0 and q1(u0) = −1, the
restriction of q1 to Gθ1 is non-degenerate, and we get hk, ki2 = −1. Thus
hk + g, ki2 = hk, ki2hg, u0i = 1.
(cid:3)
In the rest of this section, we assume that the conditions in Theorem 5.6 are
satisfied.
Lemma 5.7. We have the following formulas for the Frobenius-Schur indicators:
νm((u, 0)) = δmu,0q0(u)m,
νm((u, π)) =
1
1
hu, viG(q1, q2, v, m)2,
4Xv∈U
hk, viG(q1, q2, v, m)2 +
8Xv∈U
4Xv∈U
hg, viG(q1, q2, v, m)2 + δmg,0q1(g)m,
1
2
δmk,0q1(k)m
,
νm((k, ε)) =
νm(g) =
38
PINHAS GROSSMAN AND MASAKI IZUMI
νm((σ, ε)) =
1
1
δmσ,0q2(σ)m
,
2
hσ, viG(q1, q2, v, m)2 −
8Xv∈U
4Xv∈U
hγ, viG(q1, q2, v, m)2 − δmγ,0q1(γ)m,
pGXg∈G
pΓXγ∈Γ
hv, giq1(g)m +
1
hv, γiq2(γ)m.
νm(γ) =
where
G(q1, q2, v, m) =
1
In particular, (FS2) implies
G(q1, q2, 0, 2)2 + G(q1, q2, u0, 2)2 = 4.
Proof. Direct computation gives the formulae for νm(x). Since (0, π) is self-dual, we
get the last formula.
(cid:3)
Recall that we have
N(u,π),(u,π),(u,π) = 2, N(k,ε),(k,ε),(k,ε) =
1 + hk, u0i
2
,
Ng,g,g = 1 + hg, u0i + δ3g,0, N(σ,ε),(σ,ε),(σ,ε) =
1 + hσ, u0i
2
,
We can show the following lemma as in the proof of Lemma 3.12.
Nγ,γ,γ = 1 + hγ, u0i − δ3γ,0.
Lemma 5.8. Let the notation be as above.
(1) If m is an odd number, then
G(q1, q2, u, m) = G(q1, m) + G(q2, m) = G(q1, m)
G(q1)m − G(q2, m)
G(q2)m .
(2) The condition (FS3) is equivalent to
(−1)n1 G(q1,o, 3)
G(q1,o)3 − (−1)n2 G(q2,o, 3)
G(q2,o)3 = 2,
where Ge = 2n1 and Γe = 2n2.
further equivalent to
If neither G nor Γ has a 3-component, this is
(−1)n1+n2(Go
3
)(Γo
3
) = −1.
(3) We have
G(q1, q2, u, 2) = G(q1,e, u, 2)
G(q1,e)
(−1) Go2−1
8 − G(q2,e, u, 2)
G(q2,e)
(−1) Γo2−1
8
.
INFINITE FAMILIES OF MODULAR DATA
39
5.2. Examples. We give a list of examples satisfying the conditions in Definition
5.1, Theorem 5.5, (FS2), and (FS3). Note that the restriction of θ1 (resp. θ2) to Go
(resp. Γo) is always multiplying by −1. Direct computation using Lemma 5.8 shows
that the constraint coming from (FS3) in the following examples is very similar to
that discussed in Lemma 3.9 and Remark 3.10, and we will not mention it in what
follows.
Recall that we have either hu0, k1i = hu0, k2i = −1 or hu0, σ1i = hu0, σ2i = −1.
Lemma 5.9. Let the notation be as above.
(1) Assume hu0, k1i = hu0, k2i = −1. Then we have either Ge = Z2 × Z2 or Ge = Z4,
and Γe is a multiple of 8. The 2-rank of Γ is either 2 or 4 for Ge = Z2 × Z2 and it
is 3 for Ge = Z4.
(2) Assume hu0, σ1i = hu0, σ2i = −1. Then we have either Γe = Z2 × Z2 or Γe = Z4,
and Ge is a multiple of 8. The 2-rank of G is either 2 or 4 for Γe = Z2 × Z2 and it
is 3 for Γe = Z4.
Proof. (1) If hu0, k1i = hu0, k2i = −1, since we have hk1, k1i = hk2, k2i, the restriction
of q1 to Gθ1 is non-degenerate, and we have the factorization G = Gθ1 × Gθ1⊥ as an
involutive metric group. If Gθ1⊥ were an even group, the restriction of θ1 to it would
have a non-trivial fixed point, which is a contradiction. Thus we get Ge = Gθ1, and
Ge is either Z2 × Z2 or Z4. Since Γ = 4(Go + 1) in this case, the order of Γe is a
multiple of 8. Since G(q1) = −G(q2), we have G(q1,e)4 = G(q2,e)4, and the difference
of the 2-rank of G and that of Γ is even. Part (2) can be shown in the same way. (cid:3)
Division into cases: Since there are a number of different cases to consider,
we briefly outline the division. We consider four different general cases according
to whether (A1) or (A2) holds, and whether u0 is a fermion, i.e. q0(u0) = −1, or
a boson, i.e. q0(u0) = 1. These four cases are treated in subsections 5.2.1-5.2.4.
Each of these four cases is further subdivided into two different subcases according
to values of the quadratic forms on the fixed point subgroups of G and Γ. These
subdivisions are treated in paragraphs, e.g. 5.2.1.1 and 5.2.1.2. Finally, each of these
subcases is further subdivided according to the forms of the even groups Ge and Γe
and the associated quadratic forms q1,e and q2,e, as in previous sections. These further
subcases are treated in subparagraphs, e.g. 5.2.1.1.1.
5.2.1. Assume (A1) and q0(u0) = −1. We necessarily have hu0, k1i = hu0, k2i = −1
and hu0, σ1i = hu0, σ2i = 1. Thus Ge = 4 and the restriction of θ1 to Ge is trivial.
Since s = c/q1(k), Theorem 5.6 implies s2hσ, σi = −1 for σ ∈ Σ∗, and so c2 =
−hk, kihσ, σi for k ∈ K∗ and σ ∈ Σ∗. Since
we have the following possibilities:
{q1(k0), q1(k1), q1(k2)} = {−1, q1(k1), q1(k1)},
• If Ge = Z2 × Z2, we may assume k0 = (1, 1), k1 = (1, 0), and k2 = (0, 1). We
have four possibilities for q1,e(x, y): (1) (−1)xy, (2) (−1)x2+xy+y2, (3) ix2+y2,
and (4) i−(x2+y2).
40
PINHAS GROSSMAN AND MASAKI IZUMI
• If Ge = Z4, we may assume k1 = 1, k0 = 2, k2 = 3, and q1,e(x) = ζ rx2
r ∈ {1,−1, 3,−3}.
8 with
The set {q1(σ0), q1(σ1), q1(σ2)} is either {−1, i,−i}, or {−1, 1,−1}, and since q2(σ1) 6=
q2(σ2), the group Γθ2 is isomorphic to Z2 × Z2. Thus Theorem 2.4 shows that the
possibilities for Γe are as follows:
• Go ≡ 1 mod 4 and Γe = Z2 × Z4 or Γe = Z2 × Z2 × Z2.
• Go ≡ 3 mod 4 and Γe = Z4 × Z2m or Γe = Z2 × Z2 × Z2m with m ≥ 2 or Z4
2.
We now consider the two possibilities for q2(Γθ) separately.
5.2.1.1. q2(σ1) = i and q2(σ2) = −i. In this case, we have s2 = 1, c2 = hk, ki for
k ∈ K∗. We have Go ≡ 1 mod 4 and Go = 2Γo − 1. More precisely, Γo ≡ 1
mod 4 if and only if G ≡ 1 mod 8, and Γo ≡ 3 mod 4 if and only if Go ≡ 5
mod 8. We have either Ge = Z2 × Z2 and Γe = Z2 × Z4, or Ge = Z4 and Γe = Z3
2.
If Γe = Z2 × Z4, we may assume σ0 = (0, 2), σ1 = (1, 0), σ2 = (1, 2), q2,e(x, y) =
ixζ ry2
8 with r ∈ {1,−1, 3,−3}, and θ = −1 up to group automorphism. If Γe = Z3
2,
we may assume σ0 = (1, 1, 0), σ1 = (0, 0, 1), σ2 = (1, 1, 1), q(x, y, z) = q′(x, y)iz2,
and θ(x, y, z) = (y, x, z), where q′(x, y) is one of the following: (−1)xy, (−1)x2+xy+y2,
i±(x2+y2).
Since G(q2,e, 0, 2) = G(q2,e, σ0, 2) = 0, direct computation shows that (FS2) gives
We now consider the different possibilities for Ge separately.
no restriction.
5.2.1.1.1. Assume Ge = Z2×Z2 and q1,e(x, y) = (−1)xy. In this case we have G(q1,e) =
1, c2 = 1, and c = G(q1,o) = −ζ 1+r
8 G(q2,o). We conjecture that the modular data of
the Drinfeld center of a generalized Haagerup category for Z2×A with odd A is (S, T )
with Ge = Z2 × Z2, q1,e(x, y) = (−1)xy, Go = A × A, and q1,o(p, χ) = χ(p). There
are at least two possibilities of (Γ, q2) as we will see in (i) below: q2,e(x, y) = ix2ζ−y2
with G(q2,o) = −1 and q2,e(x, y) = ix2ζ 3y2
8 with G(q2,o) = 1 (with the exception of
A = {0}, where the first case does not occur). The conjecture is true for A = {0},
and numerical evidence is given for Z3, Z5 (see [GI19, Section 4.2]). More generally,
we conjecture that the pair
8
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
is the modular data of the Drinfeld center Z(C) of a quadratic category C of type
(Z2, Go, 1) with self-dual ρ such that α1 lifts to a fermion in Z(C) . We can consider
the following two possibilities:
(i) Assume Go ≡ 1 mod 8. In this case we have Γo ≡ 1 mod 4, and r ∈ {−1, 3}.
The smallest example is G = Z2 × Z2 and Γ = Z2 × Z4, and the only possibility is
r = 3. This indeed comes from the Drinfeld center of the unique generalized Haagerup
category for Z2, or equivalently the even part of the A7 subfactor (see [GI19, Example
4.2]).
INFINITE FAMILIES OF MODULAR DATA
41
(ii) Assume Go ≡ 5 mod 8.
In this case we have Γo ≡ 3 mod 4, and r ∈
{1,−3}. Thus G(q1,o)/G(q2,o) ∈ {i,−i} and so G(q2,e) = {i,−i}. The smallest exam-
ple is Go = Z5 and Γo = Z3, and there are four possibilities.
5.2.1.1.2. Assume Ge = Z2×Z2 and q1,e(x, y) = (−1)x2+xy+y2. Then we have G(q1,e) =
−1, and c2 = 1, and c = −G(q1,o) = −ζ 1+r
8 G(q2,o). We have r ∈ {−1, 3} for Go ≡ 1
mod 8, and r ∈ {1,−3} for Go ≡ 5 mod 8.
5.2.1.1.3. Assume Ge = Z2 × Z2 and q1,e(x, y) = ix2+y2. Then we have G(q1,e) = i,
c2 = −1, and c = iG(q1,o) = −ζ 1+r
8 G(q2,o). We have r = 1,−3 for Go ≡ 1 mod 8,
and r = −1, 3 for Go ≡ 5 mod 8.
5.2.1.1.4. Assume Ge = Z2 × Z2 and q1,e(x, y) = i−(x2+y2). This is the complex con-
jugate of the previous case.
8, c2 = ir, and c = ζ r
8 with r ∈ {1,−1, 3,−3}. Then we have
8G(q1,o) = −G(q′)ζ8G(q2,o). We have ir−1 = G(q′)2 for
5.2.1.1.5. Assume Ge = Z4 and q1,e(x) = ζ rx2
G(q1,e) = ζ r
Go ≡ 1 mod 8, and ir−1 = −G(q′)2 for Go ≡ 5 mod 8.
Remark 5.10. In the above examples with Ge = 4, fix two odd metric groups
(Go, q1,o) and (Γo, q2,o). Then we can show that all the above potential modular
data (S, T ) can be obtained from that for q1,e(x, y) = (−1)xy by applying the zesting
construction introduced in [BGH+17, Theorem 3.15].
5.2.1.2. Assume q2(σ1) = 1 and q2(σ2) = −1. In this case we have s2 = −1, c2 =
−hk, ki for k ∈ K∗, and there is an n ≥ 2 such that Γe = 2n+2 and Go = 2nΓo− 1.
In particular, we have Go ≡ 3 mod 8 for n = 2, i.e. Γe = 16, and Go ≡ 7 mod 8
for n ≥ 3.
Lemma 5.11. Under the assumptions of this section, the following hold.
(1) If Ge = Z2
additional assumption, we have the following.
2, then (FS2) implies that the 2-rank of Γ is 2 and θ2 = −1. Under this
• If either q1,e(x, y) = (−1)xy or q1,e(x, y) = (−1)x2+xy+y2, then (FS2) is equiv-
alent to
G(q2,e, σ0, 2) = 2 and G(q2,e, 0, 2)
G(q2,e)
=
2
G(q1,e)
(−1) Go2−1
8
+ Γo2−1
8
.
• If q1,e(x, y) = i±(x2+y2), then (FS2) is equivalent to
G(q2,e, 0, 2) = 2 and G(q2,e, σ0, 2)
G(q2,e)
= 2(∓i)(−1) Go2−1
8
+ Γo2−1
8
.
(2) If Ge = Z4, then (FS2) implies G(q1, q2, 0, 2) = G(q1, q2, u0, 2) = √2. Under
this additional assumption, (FS2) is equivalent to the restriction coming from
ν2(γ) =
1 + hγ, u0i
2
− δ2γ,0q2(γ)2.
42
PINHAS GROSSMAN AND MASAKI IZUMI
Proof. Since s2 = −1, we have ν2(k, ε) = 0. Since hk, u0i = −1 for k ∈ K∗, Lemma
5.7 implies that
G(q1, q2, 0, 2)2 = 2(1 − q1,e(k)2δ2k,0) and G(q1, q2, u0, 2)2 = 2(1 + q1,e(k)2δ2k,0).
(1) Assume Ge = Z2
2. If either q1,e(x, y) = (−1)xy or (−1)x2+xy+y2, we have q1(k)2 =
1, which implies
0 = G(q1, q2, 0, 2) =
2
(−1) Go2−1
8 − G(q2,e, 0, 2)
G(q2,e)
2 = G(q1, q2, u0, 2) = G(q2,e, σ0, 2).
G(q1,e)
(−1) Γo2−1
8
,
Assume that this holds. Since G(q2,e, 0, 2) = G(q2,e, σ0, 2) = 2, the case Γe = Z4
2
in the list of Theorem 2.4 never occurs, and the 2-rank of Γe is 2. Thus Γθ2 = Γ2.
Lemma 5.7 shows that ν2(k, ε) = 0, ν2(g) = hk0, gi, and ν2(γ) = hγ, σ0i − δ2γ,0q2(γ)2.
If θ2,e 6= −1, there would exist a non-self-dual γ ∈ Γ∗ which satisfies 2γ 6= 0. However,
this implies ν2(γ) = 0, which is contradiction. Therefore we get (1) in the case of
q1,e(x, y) = (−1)xy, (−1)x2+xy+y2.
If Ge = Z2
2 and q1,e(x, y) = i±(x2+y2), we get
2 = G(q1, q2, 0, 2) = G(q2,e, 0, 2),
0 = G(q1, q2, u0, 2) = 2(∓i)(−1) Go2−1
8 − G(q2,e, u0, 2)
G(q2,e)
(−1) Γo2−1
8
.
In the same way as above, we can show the statement.
(2) Since 2k 6= 0 and q1(k)2 = ±i, we get 2 = G(q1, q2, 0, 2) = G(q1, q2, u0, 2).
Under this condition, Lemma 5.7 implies
ν2(g) =
1 + hg, u0i
2
, ν2(k, ε) = 0, and ν2(γ) =
1 + hγ, u0i
2
− δ2γ,0q2(γ)2,
which shows the statement.
(cid:3)
Theorem 2.4, Lemma 5.9, and Lemma 5.11 show that we have only the following
possibilities for Ge, which we treat separately.
5.2.1.2.1. Assume Ge = Z2 × Z2 and q1,e(x, y) = (−1)xy. We may assume Γe =
ζ r2y2
Z4× Z2n with n ≥ 2, q2,e(x, y) = ζ r1x2
2n+1 with r1, r2 ∈ {1,−1, 3,−3}, and θ2,e = −1.
We may also assume σ0 = (2, 0) up to involutive metric group isomorphism. We have
c2 = −1 and
8
c = G(q1,o) = −(−1)n
r2
2−1
8 ζ r1+r2
1−1+r2
8
G(q2,o).
= 1.
Then (FS2) is equivalent to (−1) Go2−1+Γo2−1+r2
The smallest example is G = Z2 × Z2 × Z3 and Γ = Z4 × Z4, and there are four
possibilities for (r1, r2): (3,−1), (−1, 3),(1,−3),and (−3, 1), which give two metric
group isomorphism classes (consider the transformation (x, y) 7→ (x + 2y, 2x + y)).
2−1
8
We conjecture that the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
INFINITE FAMILIES OF MODULAR DATA
43
is the modular data of the Drinfeld center Z(C) of a quadratic category C of type
(Z2, Go, 1) with non-self-dual ρ such that α1 lifts to a fermion in Z(C). We also
conjecture that there are exactly two such fusion categories C for Go = Z3.
5.2.1.3. Other quadratic forms for Ge = Z2 × Z2. We can treat the other quadratic
forms for Ge = Z2 × Z2, namely q1,e(x, y) = (−1)x2+xy+y2, q1,e(x, y) = ix2+y2, and
q1,e(x, y) = i−x2−y2, in a similar way.
5.2.1.4. Assume Ge = Z4, Γe = Z2×Z2×Z2n with n ≥ 2, q1,e(x) = ζ r1x2
, q2,e(x, y, z) =
q′(x, y)ζ r2z2
2n+1 with r1, r2 ∈ {1,−1, 3,−3} and a flip invariant non-degenerate quadratic
form q′ on Z2 × Z2, and θ2,e(x, y, z) = (y, x,−z). We have σ0 = (1, 1, 0) or (0, 0, 2) for
n = 2, and σ0 = (1, 1, 0) or (1, 1, 2) for n ≥ 3. Then c2 = −ir1, and
8
c = ζ r1
8 G(q1,o) = −(−1)n
Lemma 5.11 implies that
r2
2−1
8 ζ r2
8 G(q′)G(q2,o).
and
G(q1, q2, 0, 2)2 = G(q1, q2, u0, 2)2 = 2,
ν2(γ) =
1 + hγ, u0i
2
− δ2γ,0q2(γ)2.
We choose γ0 ∈ Γ0 \ {0}, and set γ1 = (1, 0, 0, γ0). Then γ1 ∈ Γ∗ and it is not
self-dual. Thus 0 = ν2(γ), and h(1, 0, 0), σ0i = −1. Let γ2 = (0, 0, 1, 0) ∈ Γ∗. Then
θ2(γ2) = −γ2, and γ2 is self-dual. Since 2γ2 6= 0, we get
ν2(γ2) =
Thus we get σ0 = (1, 1, 0).
The relation
1 + h(0, 0, 1), σ0i
2
and h(0, 0, 1), σ0i = 1.
√2 = G(q1, q2, 0, 2) =
√2(−1)
r2
8 + Go2−1
1−1
8 −
√2G(q′, 2)
G(q′)
(−1)
r2
8 + Γo2−1
2−1
8
implies that either G(q′, 2) = 0 or
and the relation
G(q′, 2) = 2G(q′)(−1)
√2i−r1(−1)
√2 = G(q1, q2, u0, 2) =
implies that either G(q′, (1, 1), 2) = 0 or
G(q′, (1, 1), 2) = 2i−r1G(q′)(−1)
In summary, we get the following conditions:
r2
1−1
8 +
r2
8 + Go2−1
2−1
8
+ Γo2−1
8
,
r2
8 + Go2−1
1−1
8 −
√2G(q′, (1, 1), 2)
G(q′)
(−1)
r2
8 + Γo2−1
2−1
8
r2
8 + Go2−1
2−1
8
+ Γo2−1
8
.
r2
1−1
8 +
• c2 = −ir1, c = ζ r1
• G(q′) = (−1)
• G(q′) = ir1(−1)
r2
1−1
8 +
r2
1−1
8 +
8 G(q1,o) = −(−1)n
+ Γe2−1
r2
8 + Ge2−1
2−1
r2
8 + Ge2−1
2−1
8
8
+ Γe2−1
8
8
8 G(q′)G(q2,o).
r2
2−1
8 ζ r2
if q′(x, y) = (−1)xy or q′(x, y) = (−1)x2+xy+y2.
if q′(x, y) = i±(x2+y2).
44
PINHAS GROSSMAN AND MASAKI IZUMI
The smallest example is Go = Z3, and we have n = 2 and Γo = {0} in this case.
There are 16 possibilities, and 8 isomorphism classes of involutive metric groups
(consider the transformation (x, y, z) 7→ (x + z, y + z, 2(x + y) + z).)
5.2.2. Assume (A2) and q0(u0) = −1. We necessarily have hu0, ki = 1 for k ∈ K∗
and hu0, σi = −1 for σ ∈ Σ∗. Thus one of the following holds:
• Γe = Z2×Z2, q2,e(x, y) is (−1)xy, (−1)x2+xy+y2, ix2+y2, or i−x2−y2, and θ2 = −1.
We may assume σ0 = (1, 1), σ1 = (1, 0), and σ2 = (0, 1).
• Γe = Z4, q2,e(x) = ζ rx2
8 with r ∈ {1,−1, 3,−3}, and θ2,e = 1. We may assume
σ0 = 2, σ1 = 1, and σ2 = 3.
Since s = c/q2(σ), Theorem 5.6 implies s2 = hk, ki and c2 = hk, kihσ, σi. The or-
der Ge is a multiple of 8 and the set {q1(k0), q1(k1), q1(k2)} is either {−1, i,−i} or
{−1,−1, 1}. As before we treat these two cases separately.
5.2.2.1. Assume q1(k1) = i and q1(k2) = −i. In this case we have s2 = −1 and
c2 = −hσ, σi. Since
−hσ, σi = c2 = G(q2.e)2G(q2.o)2,
we have G(q2.o)2 = −1 and Γo ≡ 3 mod 4. Since G(q1,e)4 = G(q2,4)4, we have only
the following combinations:
• Ge = Z4 × Z2 and Γe = Z2 × Z2.
• Ge = Z3
2 and Γe = Z4.
Since
we can show that (FS2) does not give any restriction. The smallest examples are
G(q1,e, 0, 2) = G(q1,e, k0, 2) = 0,
and
G = Z4 × Z2 and Γ = Z2 × Z2 × Z3
G = Z3
2 and Γ = Z4 × Z3.
5.2.2.2. Assume q1(k1) = −1 and q1(k2) = 1. In this case we have s2 = 1 and c2 =
hσ, σi.
Lemma 5.12. Under the assumptions of this section, the following hold.
(1) If Γe = Z2
additional assumption, we have the following.
2, then (FS2) implies that the 2-rank of G is 2 and θ1 = −1. Under this
• If either q2,e(x, y) = (−1)xy or (−1)x2+xy+y2, then (FS2) is equivalent to
G(q1,e, k0, 2) = 2 and G(q1,e, 0, 2)
G(q1,e)
=
(−1) Go2−1
8
+ Γo2−1
8
.
2
G(q2,e)
• If q2,e(x, y) = i±(x2+y2), then (FS2) is equivalent to
G(q1,e, 0, 2) = 2 and G(q1,e, k0, 2)
G(q1,e)
= 2(∓i)(−1) Go2−1
8
+ Γo2−1
8
.
INFINITE FAMILIES OF MODULAR DATA
45
(2) If Γe = Z4, (FS2) implies G(q1, q2, 0, 2) = G(q1, q2, u0, 2) = √2. Under this
additional assumption, (FS2) is equivalent to the restriction coming from
ν2(g) =
1 + hg, u0i
2
+ δ2g,0q2(g)2.
Proof. Since s2 = 1, we have ν2(σ, ε) = ±1, and since hσ, u0i = −1 for σ ∈ Σ∗,
Lemma 5.7 implies
G(q1, q2, 0, 2)2 = 2(1 + 2ν2(σ, ε) + q2,e(σ)2δ2σ,2),
G(q1, q2, u0, 2)2 = 2(1 − 2ν2(σ, ε) − q2,e(σ)2σ, 2).
If q2,e(x, y) = (−1)xy or (−1)x2+xy+y2, we get
G(q1, q2, 0, 2)2 = 4(1 + ν2(σ, ε)),
G(q1, q2, u0, 2)2 = −4ν2(σ, ε),
which implies ν2(σ, ε) = −1, and
0 = G(q1, q2, 0, 2) = G(q1,e, 0, 2)
G(q1,e)
(−1) Go2−1
8 −
2
G(q2,e)
(−1) Γo2−1
8
,
If q2,e(x, y) = i±(x2+y2), we get
2 = G(q1, q2, u0, 2) = G(q1,e, k0, 2).
G(q1, q2, 0, 2)2 = 4ν2(σ, ε),
G(q1, q2, u0, 2)2 = 4(2 − ν2(σ, ε)),
which implies ν2(σ, ε) = 1, and
2 = G(q1, q2, 0, 2) = G(q1,e, 0, 2),
0 = G(q1, q2, u0, 2) = G(q1,e, k0, 2)
G(q1,e)
(−1) Go2−1
8 − 2(∓i)(−1) Γo2−1
8
The rest of the proof is the same as the proof of Lemma 5.11.
.
(cid:3)
Theorem 2.4 and Lemma 5.9 show that we have the following possibilities:
• Ge = Z4 × Z2n with n ≥ 2 and Γe = Z2 × Z2.
• Ge = Z2 × Z2 × Z2n with n ≥ 2 and Γe = Z4.
5.2.3. (A1) and q0(u0) = 1. We necessarily have hu0, ki = 1 for k ∈ K∗ and hu0, σi =
−1 for σ ∈ Σ∗, and hence Γe = Z2 × Z2. Since s = c/q1(k), Theorem 5.6 implies that
c2 = 1. Since
we have the following two possibilities:
{q2(σ0), q2(σ1), q2(σ2)} = {1, q2(σ1),−q2(σ1)},
• q2,e(x, y) = (−1)xy with σ0 = (1, 0), σ1 = (0, 1), σ2 = (1, 1).
• q2,e(x, y) = ix2−y2 with σ0 = (1, 1), σ1 = (1, 0), σ2 = (0, 1).
Either way, we have G(q2,e) = 1. Also, c2 = 1 implies that Γo ≡ 1 mod 4. Thus
Ge is a multiple of 16. For {q1(k0), q1(k1), q1(k2)} we have 4 possibilities: {1, 1, 1},
{1,−1,−1}, {1, i, i}, and {1,−i,−i}. As before, we treat these cases separately.
46
PINHAS GROSSMAN AND MASAKI IZUMI
5.2.3.1. Assume q1(k1) = q1(k2) = ±i. Then we have s2 = −1. We may assume Ge =
Z2×Z2m with m ≥ 3, q1,e(x, y) = ir1x2ζ r2y2
2m+1 with r1 ∈ {1,−1} and r2 ∈ {1,−1, 3,−3},
θ1 = −1, and k1 = (0, 2n−1). Hence
2m−1G1,o + 1 = Γ2,o and c = ζ r±1
8
(−1)m r2−1
8 G(q1,o) = −G(q2,o).
Since G(q1,e, 2) = G(q1,e, k0, 2) = 0, we can show that (FS2) does not give any restric-
tion. The smallest example is G = Z2 × Z8 and Γ = Z2 × Z2 × Z5.
5.2.3.2. Assume q1(k1) = q1(k2) = ±1. Then we have s2 = 1, which implies that
(σ, ε) is self-dual, and hence ν2(σ, ε) = ±1. We can show the following Lemma as
before.
Lemma 5.13. Under the above assumption, (FS2) implies the 2-rank of G is 2 and
θ1,e = −1. Under this additional assumption, we have the following:
• If q2,e(x, y) = (−1)xy, (FS2) is equivalent to
G(q1,e, k0, 2) = 2 and G(q1,e, 0, 2)
G(q1,e)
• If q2,e(x, y) = ix2−y2 (FS2) is equivalent to
G(q1,e, 0, 2) = 2 and G(q1,e, k0, 2)
G(q1,e)
= 2(−1) Go2−1
8
= 2(−1) Go2−1
8
+ Γo2−1
8
+ Γo2−1
8
.
.
to isomorphism of involutive metric groups are the following:
Among involutive metric groups in Theorem 2.4, the only cases with θ1,e = −1 up
2m, and k0 = (2m−1, 2m−1) or
, and k0 = (2m−1, 2m−1) or
2m+1ζ r2y2
2n+1 with r1, r2 ∈
• Ge = Z2m × Z2m with m ≥ 2, q1,e(x, y) = ζ xy
(2m−1, 0).
• Ge = Z2m × Z2m with m ≥ 2, q1,e(x, y) = ζ x2+xy+y2
(2m−1, 0).
• Ge = Z2m × Z2n with n ≥ m ≥ 3, q1,e(x, y) = ζ r1x2
{1,−1, 3,−3}, and k0 = (2m−1, 2n−1) or k0 = (0, 2n−1).
• Ge = Z4 × Z2n with n ≥ 3, q1,e(x, y) = ζ r1x2
and k0 = (0, 2n−1).
• Ge = Z4 × Z4, q1,e(x, y) = ζ r1x2+r2y2
2m
8
8
(2, 2).
ζ r2y2
2n+1 with r1, r2 ∈ {1,−1, 3,−3},
with r1, r2 ∈ {1,−1, 3,−3}, and k0 =
We only look at the first case in detail. Assume Ge = Z2m × Z2m with m ≥ 2 and
2m. Then we have 22m−2Go + 1 = Γo, G(q1,e) = 1, and c = G(q1,o) =
q1,e(x, y) = ζ xy
−G(q2,o), with c2 = 1. We have
G(q1,e, 0, 2) = G(q1,e, (2m−1, 0), 2) = 2 and G(q1,e, (2m−1, 2m−1), 2) = 2(−1)2m−2
.
So we get G(q1,e, 2) = G(q2,e, k0, 2) = 2. We consider different quadratic forms
separately.
INFINITE FAMILIES OF MODULAR DATA
47
5.2.3.2.1. Assume q2,e(x, y) = ix2−y2. Lemma 5.14 implies
G(q1,e, k0, 2) = 2(−1) Go2−1
8
+ Γo2−1
8 = 2(−1) Go2−1
8
(−1)22(m−2)
.
On the other hand, for k0 = (2m−1, 2m−1), we have G(q1,e, k0, 2) = 2(−1)2m−2, and we
get (−1) Go2−1
For σ0 = (2m−1, 0), we have G(q1,e, k0, 2) = 2, and
8 = 1. Thus (FS2) is satisfied if and only if Go ≡ 1 mod 8.
Go ≡(cid:26) 5 mod 8,
1 mod 8,
m = 2
m ≥ 3.
We conjecture that the modular data of the Drinfeld center of a generalized Haagerup
category for Z2m × A with odd A is given by (S, T ) in this case with Go = A × A,
q1,o(p, χ) = χ(p) and k0 = (2m−1, 2m−1) for non-trivial ǫh(g) and k0 = (2m−1, 0) for
trivial ǫh(g) (see [GI19, Example 4.3]). The conjecture is true for m = 2 and A = {0}.
More generally, the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
may possibly arise from the Drinfeld center of a quadratic category of type (Z2m, Go, 1).
5.2.3.2.2. Assume q2,e(x, y) = (−1)xy. Then Lemma 5.14 implies
and
8
G(q1,e, 2) = 2(−1) Go2−1+Γo2−1
Go ≡(cid:26) 5 mod 8,
1 mod 8,
= 2(−1) Go2−1
8
+22(m−2)
,
m = 2
m ≥ 3.
5.2.4. (A2) and q0(u0) = 1. We necessarily have hu0, ki = −1 for k ∈ K∗ and
hu0, σi = 1 for σ ∈ Σ∗. Thus Ge = Z2 × Z2. Theorem 5.6 implies s2hσ, σi = −1, and
since s = c/q2(σ), we get c2 = −1. Since
{q1(k0), q1(k1), q1(k2)} = {1, q1(k1),−q1(k1)}.
we have two possibilities for q1,e:
• q1,e(x, y) = (−1)xy with k0 = (1, 0), k1 = (0, 1), k2 = (1, 1).
• q1,e(x, y) = ix2−y2 with k0 = (1, 1), k1 = (1, 0), k2 = (0, 1).
Either way we have G(q1,e) = 1. Since c2 = −1, we get G(q1,o)2 = −1 and Go ≡ 3
mod 4. Therefore Γe is a multiple of 16.
{1, 1, 1}. As before, we treat these separately.
For {q2(σ0), q2(σ1), q2(σ2)}, we have the following possibilities: {1,±i,±i}, {1,−1,−1},
48
PINHAS GROSSMAN AND MASAKI IZUMI
5.2.4.1. Assume q2(σ1) = q2(σ2) = ±i. Then s2 = 1. We may assume that Γe =
ζ r2y2
Z2 × Z2n with n ≥ 2, q2,e(x, y) = ζ r1x2
2n+1 with r1 ∈ {1,−1} and r2 ∈ {1,−1, 3,−3},
θ2 = −1, and σ0 = (0, 2n−1). We have
4
c = G(q1,o) = −(−1)n
r2
2−1
8 ζ r1+r2
8
G(q2,o).
Since G(q2,e, 2) = G(q2,e, σ0, 2) = 0, we can show that (FS2) does not give any restric-
tion.
The smallest example is G = Z2 × Z2 × Z3 and Γ = Z2 × Z8. There are four
possibilities: (r1, r2) = (1, 1), (−1,−1), (1,−3), or (−1, 3); and 2 isomorphism classes
(consider the transformation (x, y) 7→ (x + y, 2x + y)).
We conjecture that for q1,e(x, y) = (−1)xy, the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T ),
is given by the Drinfeld center Z(C) of a quadratic category C of type (Z2, Go, 1) with
self-dual ρ such that α1 lifts to a boson in Z(C). We also conjecture that there are
exactly two such categories C for Go = Z3.
5.2.4.2. Assume q2(σ1) = q2(σ2) = ±i. Then s2 = −1, and ν2(k, ε) = 0. As before,
we can show the following lemma.
Lemma 5.14. Under the assumptions of this section, (FS2) implies that the 2-rank
of Γ is 2 and θ2,e = −1. Under this additional assumption, we have the following:
• If q1,e(x, y) = (−1)xy, then (FS2) is equivalent to
G(q2,e, σ0, 2) = 2 and G(q2,e, 0, 2)
G(q2,e)
= 2(−1) Go2−1
8
+ Γo2−1
8
• If q1,e(x, y) = ix2−y2 then (FS2) is equivalent to
G(q2,e, 0, 2) = 2 and G(q2,e, σ0, 2)
G(q2,e)
= 2(−1) Go2−1
8
+ Γo2−1
8
.
.
Among the involutive metric groups in Theorem 2.4, the only possible cases with
θ2,e = −1 are the following, up to isomorphism:
(2, 2).
2m, and σ0 = (2m−1, 2m−1) or
• Γe = Z2m × Z2m with m ≥ 2, q2,e(x, y) = ζ xy
(2m−1, 0).
• Γe = Z2m × Z2m with m ≥ 2, q2,e(x, y) = ζ x2+xy+y2
(2m−1, 0).
• Γe = Z2m × Z2n with n ≥ m ≥ 3, q2,e(x, y) = ζ r1x2
{1,−1, 3,−3}, and σ0 = (2m−1, 2n−1) or (0, 2n−1).
ζ r2y2
• Γe = Z4 × Z2n with n ≥ 3, q2,e(x, y) = ζ r1x2
2n+1 with r1, r2 ∈ {1,−1, 3,−3},
and σ0 = (0, 2n−1).
• Γe = Z4 × Z4, q2,e(x, y) = ζ r1x2+r2y2
with r1, r2 ∈ {1,−1, 3,−3}, and σ0 =
2m+1ζ r2y2
2n+1 with r1, r2 ∈
2m
, and σ0 = (2m−1, 2m−1) or
8
8
INFINITE FAMILIES OF MODULAR DATA
49
We conjecture that for q1,e(x, y) = (−1)xy, the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
8
is the modular data of the Drinfeld center Z(C) of a quadratic category C of type
(Z2, Go, 1) with non-self-dual ρ such that α1 lifts to a boson in Z(C). We also con-
jecture that there are exactly two such categories C for Go = Z3.
The first test case is G = Z2 × Z2 × Z3, Γ = Z4 × Z4, q1,e(x, y) = (−1)xy, and
q2,e(x, y) = ζ r1x2+r2y2
with (r1, r2) = (1,−3) or (r1, r2) = (−1, 3).
5.3. Conjectures. We summarize the conjectures stated so far in this section.
Conjecture 5.15. Let C be a generalized Haagerup category with an abelian group A
satisfying A2 ∼= Z2. We denote A2 = {0, a0} and ( A)2 = {0, χ0}. Then the modular
data of the Drinfeld center Z(C) is given by (S, T ) in Definition 5.1 with G = A× A,
q1(a, χ) = χ(a), k0 = (a0, χ0) for non-trivial ǫh(g) and k0 = (a0, 0) for trivial ǫh(g).
If moreover Ae = Z2m with m ≥ 2, we have Γ2 = Z2 × Z2 and q2,e(x, y) = ix2−y2.
Remark 5.16. If the above conjecture is true, it is easy to compute the Frobenius-
Schur indicator as
νk(ρ) = G(q1, k) + G(q2, k)
2
= Ak + G(q2,e, k)G(q2,o, k)
2
,
using [NS07]. For k = 2, we know ν2(ρ) = 1, and the above formula would imply
G(q2,e, 2) = 0. This shows that if A2 = Z2m with m ≥ 2, then q2,e(x, y) = (−1)xy is
not consistent. For k = 3, we can directly show ν3(ρ) = 1.
Recall that a generalized Haagerup category C for an odd group A can be equiv-
ariantized by the conjugation action of Inv(C) to produce a near group category
CA (of quadratic type ({0}, A × A, 1)), and the Drinfeld center splits as Z(CA) =
Z(C) ⊠ Z(VecA). In a similar way, a generalized Haagerup category C for the group
Z2 × A with A odd can be equivariantized by the action of A. The resulting cat-
egory Z(CA) is of quadratic type (Z2, A × A, 1). Again the Drinfeld center splits as
Z(CA) = Z(C) ⊠ Z(VecA). Therefore the above conjecture is a special case of the
following one.
Conjecture 5.17. Let A be a finite abelian group of odd order, and let C be a qua-
dratic category of type (Z2, A, 1) with the VecZ2 subcategory having trivial associator.
Then the modular data of the Drinfeld center Z(C) is given by
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T ),
where (S, T ) is as in Definition 5.1 with Ge = Z2 × Z2, Go = A, q1,e(x, y) = (−1)xy.
Moreover,
(1) If α1 lifts to a fermion in Z(C) and ρ is self-dual, we have A ≡ 1 mod 4 and
k0 = (1, 1).
(2) If α1 lifts to a fermion in Z(C) and ρ is non-self-dual, we have A ≡ 3 mod 4
and k0 = (1, 1). For A = Z3, there are exactly two such categories.
(3) If α1 lifts to a boson in Z(C), we have A ≡ 3 mod 4 and k0 = (1, 0). For
A = Z3, there are exactly 2 such categories for self-dual ρ, and 2 for non-self-dual ρ.
50
PINHAS GROSSMAN AND MASAKI IZUMI
6. Asaeda-Haagerup family
Throughout this section, we assume that (G, q1) and (Γ, q2) are metric groups
satisfying c := G(q1) = −G(q2). We assume that G is an even group and Γ is an odd
group. We denote K = G2 for simplicity. We will eventually show that K ∼= Z2 × Z2
is the only relevant case for our purposes, but for the moment we do not make this
assumption.
We set a = 1/pG and b = 1/pΓ. We choose subsets G∗ ⊂ G and Γ∗ ⊂ Γ
satisfying
and set
G = K ⊔ G∗ ⊔ (−G∗), Γ = {0} ⊔ Γ∗ ⊔ (−Γ∗),
J = {0, π} ⊔ J1 ⊔ G∗ ⊔ Γ∗,
where J1 = (K\{0})×{1,−1}. We use letters k, k′, k′′, l for elements of K, g, g′, g′′, h
for elements of G and γ, γ′, γ′′, ξ for elements of Γ. We introduce an involution of J
by setting (k, ε) = (k, c2hk, kiε) for (k, ε) ∈ J1, and leaving the other elements fixed.
Note that since Γ is odd, we have G(q2)4 = 1 and c2 ∈ {1,−1}.
Definition 6.1. Let S, T , and C be J by J matrices defined by
S =
a−b
2
a+b
a+b
2
a−b
2
2
a
a
2
2
a
a
b −b
a
2
a
2
2
0
ahk,k′i+cεε′q1(k)δk,k′
ahg, k′i
a
a
0
ahk, g′i
a(hg, g′i + hg, g′i)
b
−b
0
0
−b(hγ, γ′i + hγ, γ′i)
,
T = Diag(1, 1, q1(k), q1(g), q2(γ)),
and Cx,y = δx,y (where the block in S and T are indexed by {0}, {π}, J1, G∗, and
Γ∗.)
Direct computation gives the following results.
Lemma 6.2. Let the notation be as in Definition 6.1. Then S, T , and C are unitary
matrices satisfying S2 = C, (ST )3 = cC and T C = CT .
Lemma 6.3. Let the notation be as in Definition 6.1. Then we have
Nπ,π,π = Nπ,π,γ = Nπ,π,g = Nπ,g,γ = Nγ,γ′,g =
Nπ,π,(k.ε) = Nπ,γ,(k.ε) = Nγ,γ′,(k.ε) =
4
Γ − G
,
2
,
Γ − G
Nπ,(k,ε),(k′,ε′) = δ(k,ε),(k′,ε′) +
, Nγ,(k,ε),(k′,ε′) =
1
Γ − G
,
4
Nπ,γ,γ′ =
Γ − G − δγ,γ′,
Γ − G
Nπ,(k,ε),g = Nγ,(k,ε),g =
1
2
,
Γ − G
INFINITE FAMILIES OF MODULAR DATA
51
Nπ,g,g′ = δg,g′ +
4
4
Γ − G
, Nγ,g,g′ =
4
Γ − G
,
Nγ,γ′,γ′′ =
Γ − G − (δγ+γ′+γ′′,0 + δγ+γ′,γ′′ + δγ′+γ′′,γ + δγ′′+γ,γ′),
N(k,ε),(k′,ε′),(k′′,ε′′) =
+ c2 εε′hk, k + k′′iδk,k′ + ε′ε′′hk′, k′ + kiδk′,k′′ + ε′′εhk′′, k′′ + k′iδk′′,k
Γ − G
+ δk+k′+k′′,0)
,
1
1
2
(
1
Γ − G
2
2
+ c2εε′hk, k + giδk,k′,
+ δk+g+g′,0 + δk+g−g′,0,
Γ − G
+ (δg+g′+g′′,0 + δg+g′,g′′ + δg′+g′′,g + δg′′+g,g′).
N(k,ε),(k′,ε′),g =
N(k,ε),g,g′ =
4
Γ − G
Ng,g′,g′′ =
Theorem 6.4. Let the notation be as above. Then all the fusion coefficients Nijk are
non-negative integers if and only if Γ = G + 1 and K ∼= Z2 × Z2.
Proof. The above computation shows that Γ − G = 1 is necessary to have non-
negative integer fusion coefficients. Assume this condition. Since the condition
G(q1) = −G(q2) is not fulfilled by K = Z2, we get K ∼= Zs
2 with s > 1. Let
k, k′, k′′ ∈ K \ {0} be mutually distinct elements. Then we have
N(k,ε),(k′,ε′),(k′′,ε′′) =
1 + δk+k′+k′′,0
.
2
For N(k,ε),(k′,ε′),(k′′,ε′′) to be a non-negative integer, the only possibility is k+k′+k′′ = 0.
This means that the only possible case is K ∼= Z2 × Z2.
are non-negative integers if Γ = G + 1 and K ∼= Z2 × Z2.
On the other hand it is routine work to show that all the fusion coefficients Nijk
(cid:3)
The explicit formula for (S, T ) in Definition 6.1 stemmed out of an attempt to
unify formulae appearing in the following three conjectures.
Conjecture 6.5. Let C be a generalized Haagerup category with an abelian group A
satisfying Ae ∼= Z2 × Z2. Then the modular data of the Drinfeld center Z(C) is given
by
with G = Z2
2 × Ao × Ao, q1,e(x, y) = (−1)x2+xy+y2, and q1,o(a, χ) = χ(a).
(S(Ge,q1,e) ⊗ S, T (Ge,q1,e) ⊗ T )
This conjecture is true for trivial Ao = {0}.
Conjecture 6.6. Let C be a generalized Haagerup category with an abelian group
A satisfying Ae ∼= Z2n+1. Then the modular data of the Drinfeld center Z(C/Z2) of
2n × Ao × Ao,
the Z2-de-equivariantization C/Z2 of C is given by (S, T ) with G = Z2
q1,e(x, y) = ζ x2−y2
2n+1
, and q1,o(a, χ) = χ(a).
This conjecture is true for A = Z4 and A = Z8.
52
PINHAS GROSSMAN AND MASAKI IZUMI
Conjecture 6.7. Let C be a generalized Haagerup category with an abelian group
A satisfying Ae ∼= Z2n × Z2 with n ≥ 2. Then the modular data of the Drinfeld
center Z(C/Z2) of the Z2-de-equivariantization C/Z2 of C is given by (S, T ) with
G = Z2
2n , and q1,o(a, χ) = χ(a). The modular data of the
Drinfeld center Z(C) is given by
2n × Ao × Ao, q1,e(x, y) = ζ xy
(S(Z2×Z2,Q) ⊗ S, T (Z2×Z2,Q) ⊗ T ),
where Q(x, y) = (−1)xy.
This conjecture is true for n = 2 and Ao = {0}.
We assume Γ = G + 1 and K ∼= Z2 × Z2 in the rest of this section. Note that
we have
Nπ,π,π = 4, N(k,ε),(k,ε),(k,ε) = 2, Ng,g,g = 4 + δ3g,0, Nγ,γ,γ = 4 − δ3γ,0.
Direct computation shows the following.
Lemma 6.8. We have
νm(π) = G(q1, m) + G(q2, m)2,
νm((k, ε)) = G(q1, m) + G(q2, m)2 + δmk,0q1(k)m
2
,
νm(g) = G(q1, m) + G(q2, m)2 + δmg,0q1(g)m,
νm(γ) = G(q1, m) + G(q2, m)2 − δmγ,0q2(γ)m.
In particular, (FS2) and (FS3) are equivalent to the following conditions respectively:
(6.1)
(6.2)
G(q1, 2) + G(q2, 2) = 1,
G(q1, 3) + G(q2, 3) = 2.
6.1. Examples. We give a list of metric groups (G, q1), (Γ, q2) satisfying conditions
in Theorem 6.4 and Lemma 6.8. More precisely, the conditions are: G2 ∼= Z2 × Z2,
Γ = G + 1, G(q1) = −G(q2), and Eq.(6.1)-(6.2). The condition G2 ∼= Z2× Z2 means
that G is of the form Z2m × Z2n × Go with m, n > 0.
As in Lemma 3.12, Eq.(6.1) is equivalent to
(6.3)
G(q1,e, 2)
G(q1,e) − (−1) Go2−1
8
(−1)2m+n−2
= 1,
and Eq.(6.2) is equivalent to
(6.4)
(−1)m+nG(q1,o, 3)
G(q1,o)3 − G(q2, 3)
G(q2)3 = 2.
INFINITE FAMILIES OF MODULAR DATA
53
If neither G nor Γ has a 3-component (that is Go ≡ 1 mod 3 and Γ ≡ 2 mod 3
for even m + n or Go ≡ Γ ≡ 2 mod 3 for odd m + n), then we get
G(q1,o)3 − G(q2, 3)
G(q2)3
(Go
3 )
(−1)m+nG(q1,o, 3)
= (−1)m+n (−1) Go−1
G(q1,o)2
= (−1)m+n(Go
) − (Γ
3
2
) = 2.
3
2 (Γ
3 )
(−1) Γ−1
G(q2)2
−
Direct computation with the above formulae shows that restriction coming from (FS3)
in the following examples is very similar to that discussed in Lemma 3.9 and Remark
3.10, and we will not mention it in what follows except for one case.
Since
Γ = 2m+nGo + 1 ≡ 1 mod 4,
we have G(q2) ∈ {1,−1}, and G(q1,e)G(q1,o) ∈ {1,−1}. Thus we have either G(q1,e) ∈
{1,−1} and Go ≡ 1 mod 4, or G(q1,e) ∈ {i,−i} and Go ≡ 3 mod 4.
We treat the different forms of Ge separately.
6.1.1. Ge = Z2 × Z2, q1,e(x, y) = (−1)x2+xy+y2. In this case we have G(q1,e) = −1
and G(q1,e, 2) = 2. Thus Go ≡ 1 mod 4 and Eq.(6.3) implies (−1) Go2−1
8 = 1. This
is possible if and only if Go ≡ 1 mod 8.
The smallest example is G = Z2 × Z2 and Γ = Z5 with G(q2) = −1. The pair
(S(G,q1) ⊗ S, T (G,q1) ⊗ T )
is the modular data of the Drinfeld center of the Z2 × Z2 generalized Haagerup
category (see [GI19, Example 5.2]).
We conjecture that the modular data of the Drinfeld center of the generalized
Haagerup category for Z2 × Z2 × A with odd A is
(S(Z2×Z2,q1,e) ⊗ S, T (Z2×Z2,q1,e) ⊗ T )
with Go = A × A and q1,o(a, χ) = χ(a). More generally, the pair
(S(G,q1) ⊗ S, T (G,q1) ⊗ T )
possibly arises from the Drinfeld center of a quadratic category of type (Z2×Z2, Go, 1)
with the VecZ2×Z2 subcategory having trivial associator.
6.1.2. Ge = Z2 × Z2, q1,e(x, y) = (−1)xy. In this case we have G(q1,e) = 1 and
G(q1,e, 2) = 2. Thus Go ≡ 1 mod 4 and Eq.(6.3) implies (−1) Go2−1
8 = −1. This is
possible if and only if Go ≡ 5 mod 8.
The smallest example is Go = Z5 and Γ = Z21. There are four possible combina-
tions of q1,o and q2, up to group automorphism.
54
PINHAS GROSSMAN AND MASAKI IZUMI
6.1.3. Ge = Z2 × Z2, q1,e(x, y) = ix2−y2. In this case we have G(q1,e) = 1 and
G(q1,e, 2) = 0. Thus Go ≡ 1 mod 4 and (FS2) gives no restriction.
The smallest example is G = Z2 × Z2 and Γ = Z5 with G(q2) = −1. The pair
(S, T ) is the modular data of the Drinfeld center of the Z2-de-equivariantization of
the generalized Haagerup category for Z4 (see [GI19, Example 6.9]).
We conjecture that the modular data of the Drinfeld center of the Z2-de-equivariantization
of a generalized Haagerup category for Z4 × A with odd A is (S, T ) with Go = A× A
and q1,o(a, χ) = χ(a). More generally, the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
2n
8
8 =
. In this case we have
may possibly arise from the Drinfeld center of the Z2-de-equivariantization a qua-
dratic category of type (Z4, Go, 1).
6.1.4. Ge = Z2n × Z2n with n ≥ 2 and q1,e(x, y) = ζ x2+xy+y2
G(q1,e) = (−1)n and G(q1,e, 2) = 2(−1)n−1. Thus Go ≡ 1 mod 4, and (−1) Go2−1
−1. This is possible if and only if Go ≡ 5 mod 8.
The smallest example is Go = Z4 × Z4 × Z5, and Eq.(6.2) implies that Γ is either
Z81, Z3 × Z27, or Z9 × Z9 with restricted possibilities of quadratic forms coming from
(FS3).
6.1.5. Ge = Z2n×Z2n with n ≥ 2 and q1,e(x, y) = ζ xy
and G(q1,e, 2) = 2. Thus Go ≡ 1 mod 4, and (−1) Go2−1
and only if Go ≡ 1 mod 8.
The smallest example is G = Z4 × Z4 and Γ = Z17 with G(q2) = −1. The pair
(S, T ) is the modular data of the Drinfeld center of the Asaeda-Haagerup category,
namely the Z2 de-equivariantization of the Z4× Z2 generalized Haagerup category (or
a quadratic category of type (Z4,{0}, 2) with the VecZ4 subcategory having trivial
associator) (see [GI19, Example 5.5]).
We conjecture that the modular data of the Drinfeld center Z(C/Z2) of the Z2-de-
equivariantization of a generalized Haagerup category for Z2n × Z2 × A with odd A
(or a quadratic category C of type (Z2n × A,{0}, 2) with the VecZ2n×A subcategory
having trivial associator) is (S, T ) with Go = A × A and q2(a, χ) = χ(a). More
generally, the pair
2n . In this case we have G(q1,e) = 1
= 1. This is possible if
may possibly arises from a quadratic category of type (Z2n, Go, 2) with the VecZ2n
subcategory having trivial associator.
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
assume r = 1, up to group automorphism.
6.1.6. Ge = Z2n × Z2n with n ≥ 2 and q1,e(x, y) = ζ r(x2−y2)
G(q1,e, 2) = 2. Thus Go ≡ 1 mod 4 and (−1) Go2−1
only if Go ≡ 1 mod 8.
The smallest example is G = Z4 × Z4 and Γ = Z17 with G(q2) = −1. The pair
(S, T ) is the modular data of the Drinfeld center of the Z2-de-equivariantization of
with odd r. We may
In this case we have G(q1,e) = 1 and
= 1. This is possible if and
2n+1
8
INFINITE FAMILIES OF MODULAR DATA
55
the Z8 generalized Haagerup category (or a quadratic category of type (Z4,{0}, 2)
with the VecZ4 subcategory having non-trivial associator) - see [GI19, Example 6.10].
We conjecture that the modular data of the Drinfeld center Z(C/Z2) of the Z2-de-
equivariantization C/Z2 of a generalized Haagerup category C for Z2n+1 × A with odd
A is (S, T ) with Go = A × A and q2(a, χ) = χ(a). More generally, the pair
(S(Go,q1,o) ⊗ S, T (Go,q1,o) ⊗ T )
possibly arises from a quadratic category of type (Z2n, Go, 2) with the VecZ2n subcat-
egory having non-trivial associator.
8
6.1.7. Ge = Z2 × Z2n with n ≥ 1 and q1,e(x, y) = irx2ζ sy2
2n+1 with r ∈ {1,−1} and
(−1)n s2−1
s ∈ {1,−1, 3,−3}. In this case we have G(q1,e) = ζ r+s
and G(q1,e, 2) = 0.
Thus Go ≡ 1 mod 4 if r + s ≡ 0 mod 4, and Go ≡ 3 mod 4 if r + s ≡ 2
mod 4. Eq,(6.3) gives no restriction. The case with n = 1 and s = −r was already
treated, and the first new example is Go = Z2 × Z2 × Z3, and Γ = Z13.
In this
example r = s = ±1, and there are two possibilities determined by the relation
irG(q1,o) = −G(q2).
6.1.8. Ge = Z2m × Z2n with 2 ≤ m ≤ n and q1,e(x, y) = ζ rx2
{1,−1, 3,−3}. In this case we have
8 +n s2−1
2n+1 with r, s ∈
2m+1ζ sy2
8
8
8 +(n−1) s2−1
8
,
G(q1,e) = ζ r+s
8
(−1)m r2−1
(−1)(m−1) r2−1
and G(q1,e, 2) = 2ζ r+s
8 = (−1) Go2−1
r2−1
8 + s2−1
8
8
and Eq.(6.3) implies that (−1)
cases:
(1) Go ≡ 1 mod 8; r + s = 0.
(2) Go ≡ 5 mod 8; (r, s) = (1, 3), (−1,−3), (3, 1), (−3,−1).
(3) Go ≡ 3 mod 8; (r, s) = (1,−3), (−1, 3), (3,−1), (−3, 1).
(4) Go ≡ 7 mod 8; (r, s) = (1, 1), (−1,−1), (3, 3), (−3,−3).
7. Switching the roles of G and Γ
. We get the following sub-
As in the last section, we assume that (G, q1) and (Γ, q2) are metric group satisfying
c := G(q1) = −G(q2). We assume that G is an even group and Γ is an odd group.
We denote K = G2 for simplicity.
2
−a+b
a+b
2
,−1Γ∗
Definition 7.1. Let S and T be the unitary matrices defined in Definition 6.1, and
). We set S′ = −RSR, T ′ = T , and C′ = C;
let R = Diag(1,−1,−1J1,−1G∗
that is
−a(hg, g′i + hg, g′i)
b(hγ, γ′i + hγ, γ′i)
a+b
−a+b
2
−a
2
−a
b
ahk,k′i+cεε′q1(k)δk,k′
−ahg, k′i
−ahk, g′i
a
−a
b
b
0
0
a
2
−a
2
2
a
2
a
b
2
0
S =
−
Note that we have S′0,x > 0 if Γ < G. The following are consequences of the
calculations in the previous section.
0
.
56
PINHAS GROSSMAN AND MASAKI IZUMI
Corollary 7.2. We have S′2 = C′, (S′T ′)3 = −cC′ and T ′C′ = C′T ′.
Proof. Since R commutes with C and T , we get S′2 = RCR = C and
S′T ′S′ = RSRT RSR = RST SR = cRT ST R = cT RSRT = −cT ′S′T ′.
(cid:3)
Let
Corollary 7.3. We have
N′u,v,w =Xx
S′u,xS′v,xS′w,x
S′0,x
.
N′π,π,π = N′π,π,γ = N′π,π,g = N′π,γ,g = N′γ,γ′,g =
N′π,π,(k.ε) = N′π,γ,(k.ε) = N′γ,γ′,(k.ε) =
N′π,(k,ε),(k′,ε′) = −δ(k,ε),(k′,ε′) +
+ δγ,γ′,
4
N′π,γ,γ′ =
G − Γ
G − Γ
N′π,(k,ε),g = N′γ,(k,ε),g =
1
4
G − Γ
,
2
,
G − Γ
1
G − Γ
,
, N′γ,(k,ε),(k′,ε′) =
2
G − Γ
, N′γ,g,g′ =
,
4
,
G − Γ
N′π,g,g′ = −δg,g′ +
4
G − Γ
4
G − Γ
1
(
2
1
N′γ,γ′,γ′′ =
+ (δγ+γ′+γ′′,0 + δγ+γ′,γ′′ + δγ′+γ′′,γ + δγ′′+γ,γ′),
N′(k,ε),(k′,ε′),(k′′,ε′′) =
− c2 εε′hk, k + k′′iδk,k′ + ε′ε′′hk′, k′ + kiδk′,k′′ + ε′′εhk′′, k′′ + k′iδk′′,k
G − Γ − δk+k′+k′′,0)
2
,
N′(k,ε),(k′,ε′),g =
N′(k,ε),g,g′ =
1
G − Γ − c2εε′hk, k + giδk,k′,
G − Γ − δk+g+g′,0 − δk+g−g′,0,
2
N′g,g′,g′′ =
4
G − Γ − (δg+g′+g′′,0 + δg+g′,g′′ + δg′+g′′,g + δg′′+g,g′).
Proof. Note that we have S′x,y = −f (x)f (y)Sx,y, where f (0) = 1 and f (j) = −1 for
the other j ∈ J. Thus we get
N′u,v,w = f (u)f (v)f (w)Xx
Su,xSv,xSw,x
S0,x
= f (u)f (v)f (w)Nu,v,w.
(cid:3)
As in the proof of Theorem 6.4, we can show the following.
INFINITE FAMILIES OF MODULAR DATA
57
Theorem 7.4. Let the notation be as in Definition 7.1. Then all the fusion coeffi-
cients N′ijk are non-negative integers if and only if Γ = G − 1 and G2 ∼= Z2 × Z2.
In the rest of this section, we assume Γ = G − 1 and G2 ∼= Z2 × Z2.
Let
ν′n(k) = Xi,j∈J
N k
ijS′0iS′0j(
)n.
T ′jj
T ′ii
As before, we have the following.
Theorem 7.5. Let the notation be as above. We have
ν′m(π) = G(q1, m) + G(q2, m)2,
ν′m((k, ε)) = G(q1, m) + G(q2, m)2 − δmk,0q1(k)m
ν′m(g) = G(q1, m) + G(q2, m)2 − δmg,0q1(g)m,
ν′m(γ) = G(q1, m) + G(q2, m)2 + δmγ,0q2(γ)m.
2
,
In particular, (FS2) and (FS3) are equivalent to the following respectively:
G(q1, 2) + G(q2, 2)2 = 1,
G(q1, 3) + G(q2, 3)2 = 4.
Conjecture 7.6. The modular data of the Drinfeld center of a near-group category
with a finite abelian group A with multiplicity 2A is given by
with Γ = A.
(S(Γ,q2) ⊗ S′, T (Γ,q2) ⊗ T ′)
7.1. Examples. We give a list of pairs of metric groups (G, q1) and (Γ, q2) satisfying
the conditions in Theorem 7.4 and Theorem 7.5. More precisely, the conditions are:
G2 ∼= Z2 × Z2, Γ = G − 1, G(q1) = −G(q2), Eq.(6.1) and Eq.(6.2). The condition
G2 ∼= Z2 × Z2 means that G is of the form Z2m × Z2n × Go with m, n > 0. As in the
last section, Eq.(6.1) is equivalent to Eq.(6.3).
Direct computation with the above formulae shows that the restriction coming
from (FS3) through in the following examples is very similar to that discussed in
Lemma 3.9 and Remark 3.10, and we will not mention it in what follows except for
one case.
Since
Γ = 2m+nGo − 1 ≡ 3 mod 4,
we have G(q2) ∈ {i,−i}, and G(q1,e)G(q1,o) ∈ {i,−i}. Thus we have either G(q1,e) ∈
{1,−1} and Go ≡ 3 mod 4 or G(q1,e) ∈ {i,−i} and Go ≡ 1 mod 4.
7.1.1. Ge = Z2 × Z2, q1,e(x, y) = (−1)x2+xy+y2. In this case we have G(q1,e) = −1
and G(q1,e, 2) = 2. Thus Go ≡ 3 mod 4 and Eq.(6.3) implies that (−1) Go2−1
8 = 1.
This is possible if and only if Go ≡ 7 mod 8. The smallest example is Go = Z7,
and Eq.(6.2) allows only Γ = Z27 and Γ = Z3 × Z9 with restricted quadratic forms.
58
PINHAS GROSSMAN AND MASAKI IZUMI
8
2n
. In this case we have
2n . In this case we have G(q1,e) = 1,
= 1. This is possible if and
7.1.2. Ge = Z2 × Z2, q1,e(x, y) = (−1)xy. In this case we have G(q1,e) = 1 and
G(q1,e, 2) = 2. Thus Go ≡ 3 mod 4 and Eq.(6.3) implies that (−1) Go2−1
= −1.
This is possible if and only if Go ≡ 3 mod 8. The smallest example is Go =
Z3 and Γ = Z11. There are two possible combinations of q1,o and q2 up to group
automorphism.
7.1.3. Ge = Z2 × Z2, q1,e(x, y) = ix2−y2. In this case we have G(q1,e) = 1 and
G(q1,e, 2) = 0. Thus Go ≡ 3 mod 4 and Eq.(6.3) gives no restriction. The smallest
example is Go = Z3 and Γ = Z11, and there are two possible combinations of q1,o and
q2.
7.1.4. Ge = Z2n × Z2n with n ≥ 2 and q1,e(x, y) = ζ x2+xy+y2
G(q1,e) = (−1)n and G(q1,e, 2) = 2(−1)n−1. Thus Go ≡ 3 mod 4, and (−1) Go2−1
8 =
−1. This is possible if and only if Go ≡ 3 mod 8. The smallest example is n = 2,
Go = Z3, and Γ = Z47.
7.1.5. Ge = Z2n×Z2n with n ≥ 2 and q1,e(x, y) = ζ xy
G(q1,e, 2) = 2. Thus Go ≡ 3 mod 4 and (−1) Go2−1
only if Go ≡ 7 mod 8. The smallest example is n = 2, Go = Z7, and Γ = Z111.
7.1.6. Ge = Z2n × Z2n with n ≥ 2 and q1,e(x, y) = ζ r(x2−y2)
with odd r. We may
assume r = 1, up to group automorphism. In this case, we have G(q1,e) = 1 and
G(q1,e, 2) = 2. Thus Go ≡ 3 mod 4, and (−1) Go2−1
= 1. This is possible if and
only if Go ≡ 7 mod 8. The smallest example is n = 2, Go = Z7, and Γ = Z111.
7.1.7. Ge = Z2 × Z2n with n ≥ 1 and q1,e(x, y) = irx2ζ sy2
s ∈ {1,−1, 3,−3}. If n = 1, we further assume r + s 6= 0.
G(q1,e) = ζ r+s
mod 4, and Go ≡ 1 mod 4 if r + s ≡ 2 mod 4. Eq,(6.3) gives no restriction.
(1) n = 1, Go = {0}, Γ = Z3, r = s = ±1, and there are two possibilities determined
by the relation G(q2) = −ir.
(2) n = 1, Go = Z5, Γ = Z19, r = s = ±1, and there are two possibilities determined
by the relation G(q2) = −irG(q1,o).
(3) n = 2, Go = {0}, and Γ = Z7. We have G(q2) = −ζ s+t
(4) n = 3, Go = {0}, and Γ = Z15. We have G(q2) = −(−1)
mod 4.
7.1.8. Ge = Z2m × Z2n with 2 ≤ m ≤ n and q1,e(x, y) = ζ rx2
2n+1 with r, s ∈
{1,−1, 3,−3}. If m = n, we further assume r + s 6= 0. In this case we have G(q1,e) =
ζ r+s
8
2n+1 with r ∈ {1,−1} and
In this case we have
and G(q1,e, 2) = 0. Thus Go ≡ 3 mod 4 if r + s ≡ 0
8 with r + s ≡ 2 mod 4.
8 with r + s ≡ 2
We give a few examples.
(−1)n s2−1
8
2m+1ζ sy2
s2−1
8 ζ r+s
2n+1
8
8
8
(−1)m r2−1
8 +n s2−1
8
,
and Eq.(6.3) implies that (−1)
G(q1,e, 2) = 2ζ r+s
r2−1
8 + s2−1
8
(−1)(m−1) r2−1
8 = (−1) Go2−1
8
.
8 +(n−1) s2−1
8
,
INFINITE FAMILIES OF MODULAR DATA
59
We get the following sub-cases.
(1) Go ≡ 1 mod 8 and r = s. The smallest example is n = 2, Go = {0}, and
Γ = Z15.
(2) Go ≡ 3 mod 8 and (r, s) = (1, 3), (−1,−3), (3, 1), or (−3,−1).
(3) Go ≡ 5 mod 8. and (r, s) = (1,−3), (−1, 3), (3,−1), or (−3, 1).
(4) Go ≡ 7 mod 8 and r + s = 0.
The first four sections of this (online only) appendix contain calculations for check-
ing the modular relations and computing Verlinde coefficients and Frobenius-Schur
indicators for the families of modular data in Sections 3-6, respectively. The last
section of the appendix computes the modular data for the Z2-de-equivariantizations
of the Z4-near-group categories referred to in Section 4.
Appendix A. Calculations for Section 3
Proof of Lemma 3.4.
hu′ − u, vi + a2 Xg∈G∗
hu′ − u, gi + b2 Xγ∈Γ∗
hu′ − u, γi
S(u,π),xS(u′,π),x
b2
2 Xγ∈Γ
hu′ − u, γi
Xx
=
=
a2 + b2
S(u,0),xS(u′,0),x =Xx
2 Xv∈U
2 Xg∈G
hu′ − u, gi +
a2
= δu,u′.
=
S(u,0),xS(u′,π),x
Xx
2 Xv∈U
a2 − b2
2 Xg∈G
hu′ − u, gi −
a2
=
hu′ − u, vi + a2 Xg∈G∗
hu′ − u, gi − b2 Xγ∈Γ∗
hu′ − u, γi
b2
2 Xγ∈Γ
hu′ − u, γi =
δu,u′
2 −
δu,u′
2
= 0.
S(u,π),xSg,x
S(u,0),xSg,x =Xx
Xx
= a2Xv∈U
= a2Xh∈G
hg − u, vi + a2 Xh∈G∗
hg − u, hi = δg,u = 0.
(hg − u, hi + hθ(g) − u, hi)
60
PINHAS GROSSMAN AND MASAKI IZUMI
S(u,π),xSγ,x
Xx
S(u,0),xSγ,x = −Xx
= −b2Xv∈U
= −b2Xξ∈Γ
hγ − u, vi − b2Xξ∈Γ∗
hγ − u, ξi = −δγ,u = 0.
(hγ − u, ξi + hθ(γ) − u, ξi)
Sg,xSg′,x
Xx
= 2a2Xv∈U
= a2(Xh∈G
hg′ − g, vi + a2 Xh∈G∗
hg′ − g, vi +Xh∈G
hg′ − θ(g), hi
(h−g, hi + h−θ(g), hi1)(hg′, hi + hθ(g′), hi)
= δg,g′ + δθ(g),g′ = δg,g′.
Px Sg,xSγ,x = 0 is obvious.
Sγ,xSγ′,x
Xx
= 2b2Xv∈U
hγ′ − γ, vi + b2Xξ∈Γ∗
= δγ,γ′ + δθ(γ),γ′ = δγ,γ′.
(h−γ, ξi + h−θ(γ), ξi)(hγ′, ξi + hθ(γ′), ξi)
Thus S is unitary. It is obvious that T and C are unitary. Since Sx,y = Sx,y = Sx,y
and S is symmetric, we have S2 = C. Since Tx,x = Tx,x, we have CT = T C.
The only remaining condition is (ST )3 = cC, and it suffices to verify
Xx
Sj,xSj′,xTx,x = cSj,j′Tj,jTj′,j′
for all j, j′ ∈ J.
Xx
S(u,π),xS(u′,π),xTx,x
a2 + b2
S(u,0),xS(u′,0),xTx,x =Xx
2 Xv∈U
2 Xg∈G
q1(u + u′ − g)q0(u + u) +
hu + u′,−viq0(v) + a2 Xg∈G∗
2 Xγ∈Γ
a2
b2
=
=
= (
a
2G(q1) +
b
2G(q2))hu, u′iq0(u)q0(u′)
= cS(u,0),(u′,0)T(u,0),(u′,0) = cS(u,π),(u′,π)T(u,π),(u′,π).
hu + u′,−giq1(g) + b2 Xγ∈Γ∗
q2(u + u′ − γ)q0(u + u′)
hu + u′,−γiq2(γ)
INFINITE FAMILIES OF MODULAR DATA
61
S(u,0),xS(u′,π),xTx,x
Xx
2 Xv∈U
a2 − b2
2 Xg∈G
q1(u + u′ − g)q0(u + u) −
hu + u′,−viq1(v) + a2 Xg∈G∗
2 Xγ∈Γ
a2
b2
=
=
hu + u′,−giq1(g) − b2 Xγ∈Γ∗
q2(u + u′ − γ)q0(u + u′)
hu + u′,−γiq2(γ)
b
2G(q2))hu, u′iq0(u)q0(u′)
bG(q2)
= cS(u,0),(u′,π)T(u,0),(u,0)T(u′,π),(u′,π).
= (
=
2
a
2G(q1) −
aG(q1)
2 −
Xx
= a2Xv∈U
= a2Xh∈G
S(u,0),xSg,xTx,x = S(u,π),xSg,xTx,x
hu + g,−viq1(v) + a2 Xh∈G∗
(hu + g,−hi + hu + θ(g),−hi)q1(h)
q1(u + g − h)q1(u + g) = aG(q1)hu, giq0(u)q1(g)
= cS(u,0),gT(u,0),(u,0)Tg,g = cS(u,π),gT(u,π),(u,π)Tg,g.
S(u,π),xSγ,xTx,x
Xx
S(u,0),xSγ,xTx,x = −Xx
= −b2Xv∈U
= −b2Xξ∈Γ
hu + γ,−viq0(v) − b2Xξ∈Γ∗
q2(u + γ − ξ)q2(u + γ) = −bG(q2)hu, γiq0(u)q2(γ)
(hu + γ,−ξi + hu + θ(γ),−ξi)q2(ξ)
= cS(u,0),γT(u,0),(u,0)Tγ,γ = −cS(u,π),γT(u,π),(u,π)Tγ,γ.
Sg,xSg′,xTx,x
Xx
= 2a2Xv∈U
= a2Xh∈G
= ca(hg, g′iq1(g)q1(g′) + hθ(g), g′iq1(g)q1(g′)
= cSg,g′Tg,gTg′,g′.
hg + g′,−vi1q1(v) + a2 Xh∈G∗
(q1(g + g′ − h)q1(g + g′) + q1(θ(g) + g′ − h)q1(θ(g) + g′))
(hg,−hi + hθ(g),−hi)(hg′,−hi + hθ(g′),−hi)q1(h)
Xx
Sg,xSγ,xTx,x = 0 = cSg,γTg,gTγ,γ.
62
PINHAS GROSSMAN AND MASAKI IZUMI
Sγ,xSγ′,xTx,x
Xx
= 2b2Xv∈U
= b2Xξ∈Γ
= bG(q2)(q2(γ + γ′) + q2(γ − γ′))
= cSγ,γ′Tγ,γTγ′,γ′.
hγ + γ′,−vi + b2Xξ∈Γ∗
(q2(γ + γ′ − ξ)q2(γ + γ′) + q2(θ(γ) + γ′ − ξ)q2(θ(γ) + γ′))
(hγ,−ξi + hθ(γ),−ξi)(hγ′,−ξihθ(γ′),−ξi)q2(ξ)
Proof of Lemma 3.5.
N(u,0),(u′,0),(u′′,0) = N(u,0),(u′,π),(u′′,π)
(cid:3)
a + b
2
)2Xv∈U
hu + u′ + u′′,−vi
= (
a − b
2
+ a2 Xg∈G∗
2 Xg∈G
a2
=
)2Xv∈U
hu + u′ + u′′,−vi + (
hu + u′ + u′′,−gi + b2 Xγ∈Γ∗
2 Xγ∈Γ
hu + u′ + u′′,−gi +
b2
= δu+u′+u′′,0.
hu + u′ + u′′,−γi
hu + u′ + u′′,−γi
N(u,0),(u′,0),(u′′,π)
=
4 Xv∈U
a2 − b2
hu + u′ + u′′,−vi +
hu + u′ + u′′,−gi − b2 Xγ∈Γ∗
+ a2 Xg∈G∗
2 Xg∈G
2 Xγ∈Γ
hu + u′ + u′′, gi −
a2
b2
=
hu + u′ + u′′,−vi
4 Xv∈U
a2 − b2
hu + u′ + u′′,−γi
hu + u′ + u′′, γi = 0.
N(u,0),(u′,0),g = N(u,0),(u′,π),g
a(a + b)
hu + u′ + g,−vi +
2 Xv∈U
2 Xv∈U
hu + u′,−hi(hg,−hi + hθ(g),−hi)
hu + u′ + g, hi = δu+u′+g,0 = 0.
a(a − b)
=
+ a2 Xh∈G∗
= a2Xh∈G
hu + u′ + g,−vi
INFINITE FAMILIES OF MODULAR DATA
63
hu + u′ + γ,−vi
N(u,0),(u′,0),γ = −N(u,0),(u′,π),γ
=
b(a − b)
hu + u′ + γ,−vi −
2 Xv∈U
2 Xv∈U
hu + u′,−ξi(hγ,−ξi + hθ(γ),−ξi)
b(a + b)
− b2Xξ∈Γ∗
= −b2Xξ∈Γ
hu + g + g′,−vi + a2 Xh∈G∗
(hu + g + g′, hi + hu + θ(g) + g′, hi)
hu + u′ + γ, ξi = −δu+u′+γ,0 = 0.
hu,−hi(hg,−hi + hθ(g),−hi)(hg′,−hi + hθ(g′),−hi)
N(u,0),g,g′
= 2a2Xv∈U
= a2Xh∈G
N(u,0),γ,γ′
= 2b2Xv∈U
= b2Xξ∈Γ
= δu+g+g′,0 + δu+θ(g)+g′,0.
N(u,0),g,γ = abXv∈U
hu,−vihg,−vihγ,−vi − abXv∈U
hu,−vihg,−vihγ,−vi = 0.
hu + γ + γ′,−vi + b2Xξ∈Γ∗
(hu + γ + γ′, ξi + hu + θ(γ) + γ′, ξi)
hu,−ξi(hγ,−ξi + hθ(γ),−ξi)(hγ′,−ξi + hθ(γ′),−ξi)
= δu+γ+γ′,0 + δu+θ(γ)+γ′,0.
N(u,π),(u′,π),(u′′,π)
hu + u′ + u′′,−vi
=
−
= (
(a + b)3
4(a − b)Xv∈U
+ a2 Xg∈G∗
(a + b)4 + (a − b)4
2 Xg∈G
a2 − b2 Xv∈U
hu + u′ + u′′,−vi +
hu + u′ + u′′,−gi − b2 Xγ∈Γ∗
2 Xv∈U
a2 − b2
4(a2 − b2)
2 Xγ∈Γ
hu + u′ + u′′, gi −
hu + u′ + u′′, vi +
UXv∈U
1
hu + u′ + u′′, vi.
Γ − G
4U
4a2b2
=
=
a2
+
b2
4(a + b)Xv∈U
(a − b)3
hu + u′ + u′′,−γi
hu + u′ + u′′, vi
hu + u′ + u′′, γi
δu+u′+u′′,0
δu+u′+u′′,0
−
2
2
64
PINHAS GROSSMAN AND MASAKI IZUMI
hu + u′ + g,−vi
− a2)Xv∈U
hu + u′ + g, vi + a2Xh∈G
hu + u′ + g, hi
N(u,π),(u′,π),g
a(a + b)2
2(a − b) Xv∈U
+ a2 Xh∈G∗
hu + u′ + g,−vi +
2(a + b) Xv∈U
a(a − b)2
hu + u′,−hi(hg,−hi + hθ(g),−hi)
= (a
2(a2 − b2)
(a + b)3 + (a − b)3
a2 − b2 Xv∈U
4a2b2
=
=
hu + u′ + g, vi + δu+u′+g,0
UXv∈U
1
hu + u′ + g, vi.
4U
Γ − G
=
N(u,π),(u′,π),γ
b(a + b)2
2(a − b) Xv∈U
− b2Xξ∈Γ∗
hu + u′ + γ,−vi −
2(a + b) Xv∈U
b(a − b)2
hu + u′ + γ,−vi
hu + u′,−ξi(hγ,−ξi + hθ(γ),−ξi)
+ b2)Xv∈U
hu + u′ + γ, vi − b2Xξ∈Γ
hu + u′ + γ, ξi
= (b
2(a2 − b2)
(a + b)3 − (a − b)3
a2 − b2 Xv∈U
4a2b2
=
=
hu + u′ + γ, vi − δu+u′+γ,0
UXv∈U
1
hu + u′ + γ, vi.
4U
Γ − G
INFINITE FAMILIES OF MODULAR DATA
65
(hu + g + g′, hi + hu + θ(g) + g′, hi)
N(u,π),g,g′
= a2 a + b
+ a2 Xh∈G∗
= a2(
a − bXv∈U
a + bXv∈U
hu + g + g′,−vi + a2 a − b
hu + g + g′,−vi
hu,−hi(hg,−hi + hθ(g),−hi)(hg′,−hi + hθ(g′),−hi)
4a2b2
− 2)Xv∈U
(a + b)2 + (a − b)2
a2 − b2 Xv∈U
hu + g + g′, vi + a2Xh∈G
a2 − b2
hu + g + g′, vi + δu+g+g′,0 + δu+θ(g)+g′,0
UXv∈U
1
hu + g + g′, vi + δu+g+g′,0 + δu+θ(g)+g′,0.
Γ − G
4U
=
=
hu,−vihg,−vihγ,−vi −
ab(a − b)
a + b Xv∈U
hu,−vihg,−vihγ,−vi
N(u,π),g,γ
= ab
=
=
=
ab(a + b)
a − b Xv∈U
(a + b)2 − (a − b)2
a2 − b2 Xv∈U
a2 − b2 Xv∈U
hu, vihg, vihγ, vi
1
4a2b2
4U
Γ − G
UXv∈U
hu, vihg, vihγ, vi.
hu, vihg, vihγ, vi
N(u,π),γ,γ′
= b2 a + b
− b2Xξ∈Γ∗
= b2(
a + bXv∈U
hu + γ + γ′,−vi + b2 a − b
a − bXv∈U
hu + γ + γ′,−vi
hu,−ξi(hγ,−ξi + hθ(γ),−ξi)(hγ′,−ξi + hθ(γ′),−ξi)
4a2b2
a2 − b2
+ 2)Xv∈U
(a + b)2 + (a − b)2
a2 − b2 Xv∈U
hu + γ + γ′, vi − b2Xξ∈Γ
hu + γ + γ′, vi − δu+γ+γ′,0 − δu+θ(γ)+γ′,0
UXv∈U
1
hu + γ + γ′, vi − δu+γ+γ′,0 − δu+θ(γ)+γ′,0.
Γ − G
4U
=
=
(hu + γ + γ′, ξi + hu + θ(γ) + γ′, ξi)
66
PINHAS GROSSMAN AND MASAKI IZUMI
hg + g′ + g′′,−vi +
2a3
a + bXv∈U
hg + g′ + g′′,−vi
(hg,−hi + hθ(g),−hi)(hg′,−hi + hθ(g′),−hi)(hg′′,−hi + hθ(g′′),−hi)
hg + g′ + g′′,−vi
Ng,g′,g′′
2a3
=
= (
4a4
a − bXv∈U
+ a2 Xh∈G∗
a2 − b2 − 4a2)Xv∈U
+ a2Xh∈G
a2 − b2 Xv∈U
4a2b2
=
=
(hg + g′ + g′′, hi + hθ(g) + g′ + g′′, hi + hg + θ(g′) + g′′, hi + hg + g′ + θ(g′′), hi)
hg + g′ + g′′, vi + δg+g′+g′′,0 + δθ(g)+g′+g′′,0 + δg+θ(g′)+g′′,0 + δg+g′+θ(g′′),0
UXv∈U
1
hg + g′ + g′′, vi
4U
Γ − G
+ δg+g′+g′′,0 + δθ(g)+g′+g′′,0 + δg+θ(g′)+g′′,0 + δg+g′+θ(g′′),0.
2a2b
a + bXv∈U
hg + g′,−vihγ,−vi
Ng,g′,γ
=
=
=
2a2b
a − bXv∈U
a2 − b2 Xv∈U
4a2b2
4U
Γ − G
hg + g′,−vihγ,−vi −
hg + g′, vihγ, vi
UXv∈U
1
hg + g′, vihγ, vi.
2ab2
a + bXv∈U
hg,−vihγ + γ′,−vi
Ng,γ,γ′
=
=
=
2ab2
a − bXv∈U
a2 − b2 Xv∈U
4a2b2
4U
Γ − G
hg,−vihγ + γ′,−vi +
hg, vihγ + γ′, vi
UXv∈U
1
hg, vihγ + γ′, vi.
INFINITE FAMILIES OF MODULAR DATA
67
Nγ,γ′,γ′′
2b3
=
4b4
= (
a − bXv∈U
− b2Xξ∈Γ∗
a2 − b2 + 4b2)Xv∈U
− b2Xξ∈Γ
a2 − b2 Xv∈U
4a2b2
=
=
4U
Γ − G
hγ + γ′ + γ′′,−vi −
2b3
a + bXv∈U
hγ + γ′ + γ′′,−vi
(hγ,−ξi + hθ(γ),−ξi)(hγ′,−ξi + hγ′θ,−ξi)(hγ,−ξi + hθ(γ′′),−ξi)
hγ + γ′ + γ′′, vi
(hγ + γ′ + γ′′, vi + hθ(γ) + γ′ + γ′′, vi + hγ + θ(γ′) + γ′′, vi + hγ + γ′ + θ(γ′′), vi)
hγ + γ′ + γ′′, vi − δγ+γ′+γ′′,0 − δθ(γ)+γ′+γ′′,0 − δγ+θ(γ′)+γ′′,0 − δγ+γ′+θ(γ′′),0
UXv∈U
1
hγ + γ′ + γ′′, vi
− δγ+γ′+γ′′,0 − δθ(γ)+γ′+γ′′,0 − δγ+θ(γ′)+γ′′,0 − δγ+γ′+θ(γ′′),0.
(cid:3)
(a + b)2
q0(u′)m
4
q0(u − u′)m
Proof of Lemma 3.8.
νm((u, 0))
N (u,0)
(u′,0),(u−u′,0)
(a − b)2
4
N (u,0)
(u′,π),(u−u′,π)
q0(u′)m
q0(u − u′)m + Xu′∈U
N (u,0)
γ,u−γb2
q2(γ)m
q2(u − γ)m
= Xu′∈U
= Xg∈G∗
2 Xg∈G
a2
=
N (u,0)
g,u−ga2
q1(g)m
q1(u − g)m + Xγ∈Γ∗
2 Xγ∈Γ
b2
q0(g)m
q1(γ)m
q0(u − g)m +
q2(u − γ)m .
Since q1(g)/q1(u − g) = hu, giq0(u), we have
a2
2 Xg∈G
q0(g)m
q0(u − g)m =
1
2GXg∈G
hmu, giq0(u)−m =
δmu,0q0(u)m
2
,
and νm((u, 0)) = δmu,0q0(u)m.
Note that we have G(q1, q2, v, m) = G(q1, q2,−v, m) and G(q1, q2, v,−m) = G(q1, q2, v, m).
We set
A(g, m) = 2a Xh∈G∗
hg, hiq1(h)m, B(γ, m) = 2bXξ∈Γ∗
C(v, m) = (a + b)Xu∈U
hv, uiq0(u)m.
hγ, ξiq2(ξ)m.
68
Then
PINHAS GROSSMAN AND MASAKI IZUMI
G(q1, q2, v, m) = A(v, m) + B(v, m) + C(v, m).
For νm((u, π)), we have
νm((u, π))
N (u,π)
(u′,0),(u−u′,π)
N (u,π)
(u′,π),g
q0(u′)m
a2 − b2
(
(
4
2
a(a + b)
q0(u − u′)m +
q0(u′)m
q1(g)m +
q0(u′)m
q2(γ)m +
q1(h)m + Xg∈G∗, γ∈Γ∗
b(a + b)
(
q1(g)m
q0(u′)m )
q2(γ)m
q0(u′)m )
N (u,π)
(u′,π),γ
2
g,h a2 q1(g)m
N (u,π)
= Xu′∈U
+ Xu′∈U, g∈G∗
+ Xu′∈U, γ∈Γ∗
+ Xg,h∈G∗
and
q0(u′)m ) + Xu′,u′′∈U
q0(u − u′)m
N (u,π)
(u′,π),(u′′,π)
(a + b)2
4
q0(u′)m
q0(u′′)m
N (u,π)
g,γ ab(
q1(g)m
q2(γ)m +
q2(γ)m
q1(g)m ) + Xγ,ξ∈Γ∗
γ,ξ b2 q2(γ)m
N (u,π)
q2(ξ)m ,
=
N (u,π)
Xu′∈U
4 Xu′∈U
a2 − b2
Xu′,u′′∈U
(u′,0),(u−u′,π)
a2 − b2
4
(
q0(u′)m
q0(u − u′)m +
q0(u − u′)m
q0(u′)m ) =
q0(u − u′)m
q0(u′)m )
2 Xu′∈U
a2 − b2
q0(u′)m
(
q0(u − u′)m +
q0(u′)m
q0(u − u′)m ,
N (u,π)
(u′,π),(u′′,π)
(a + b)2
4
q0(u′)m
q0(u′′)m
=
=
(a + b)2
4 Xu′,u′′∈U
Γ − GXv∈U
1
4
hu′ + u′′ − u, vi
Γ − GXv∈U
h−u, viC(u′, m)C(u′′,−m),
q0(u′)m
q0(u′′)m
q1(g)m
q0(u′)m )
2
(
q0(u′)m
q1(g)m +
Γ − GXv∈U
4
Xu′∈U, g∈G∗
a(a + b)
N (u,π)
(u′,π),g
=
=
a(a + b)
2 Xu′∈U, g∈G∗
Γ − GXv∈U
1
h−u, vi(C(v, m)A(v,−m) + A(v, m)C(v,−m)),
hu′ − u + g, vi(
q0(u′)m
q1(g)m +
q1(g)m
q0(u′)m )
INFINITE FAMILIES OF MODULAR DATA
69
q2(γ)m
q0(u′)m )
2
(
q0(u′)m
q2(γ)m +
Γ − GXv∈U
4
Xu′∈U, γ∈Γ∗
b(a + b)
N (u,π)
(u′,π),γ
=
=
b(a + b)
2 Xu′∈U, γ∈Γ∗
Γ − GXv∈U
1
h−u, vi(C(v, m)B(v,−m) + B(v, m)C(v,−m)),
hu′ − u + γ, vi(
q0(u′)m
q2(γ)m +
q2(γ)m
q0(u′)m )
g,h a2 q1(g)m
N (u,π)
q1(h)m
(
1
=
Xg,h∈G∗
= a2 Xg,h∈G∗
Γ − GXv∈U
Γ − GXv∈U
Γ − GXv∈U
Γ − GXv∈U
=
=
=
1
1
1
hg + h − u, vi + δg+h,u + δg+θ(h),u)
q1(g)m
q1(h)m
4
Γ − GXv∈U
h−u, viA(v, m)A(v,−m) + a2 Xg,h∈G∗
h−u, viA(v, m)A(v,−m) + a2 Xg∈G∗
2 Xg∈G
h−u, viA(v, m)A(v,−m) +
a2
h−u, viA(v, m)A(v,−m) +
2
δmu,0q0(u)m
(δg+h,u + δg+θ(h),u)
q1(g)m
q1(h)m
q1(g)m
q1(u − g)m
q1(g)m
q1(u − g)m −
2 Xg∈U
−
a2
q1(g)m
q1(u − g)m
a2
2 Xg∈U
q1(u − g)m .
q1(g)m
q1(g)m
q2(γ)m +
Γ − GXv∈U
4
g,γ ab(
N (u,π)
Xg∈G∗, γ∈Γ∗
= ab Xg∈G∗, γ∈Γ∗
Γ − GXv∈U
=
1
q2(γ)m
q1(g)m )
h−u, vihg, vihγ, vi(
q1(g)m
q2(γ)m +
q2(γ)m
q1(g)m )
h−u, vi(A(v, m)B(v,−m) + B(v, m)A(v,−m)),
70
PINHAS GROSSMAN AND MASAKI IZUMI
γ,ξ b2 q2(γ)m
N (u,π)
q2(ξ)m
Xγ,ξ∈Γ∗
= b2 Xγ,ξ∈Γ∗
(
1
Γ − G
=
=
4
Γ − GXv∈U
B(v, m)B(v,−m) − b2 Xγ∈Γ∗
1
Γ − G
B(v, m)B(v,−m) −
h−u + γ + ξ, vi − δγ+ξ,u − δγ+θ(ξ),u)
q2(γ)m
q2(ξ)m
q2(γ)m
δmu,0q0(u)m
q2(u − γ)m
2 Xγ∈U
b2
+
2
q2(γ)m
q2(u − γ)m .
Thus we get
νm((u, π))
=
1
Γ − GXv∈U
h−u, viG(q1, q2, v, m)G(q1, q2, v,−m) =
1
Γ − GXv∈U
hu, viG(q1, q2, v, m)2.
For νm(g), we have
νm(g)
N g
(u,π),(u′,π)
(a + b)2
4
q0(u)m
q0(u′)m
(u,0),g−u
N g
(u,π),h
N g
(u,π),γ
q0(u)m
a(a − b)
(
(
2
2
a(a + b)
q1(g − u)m +
q0(u)m
q1(h)m +
q0(u)m
q2(γ)m +
q1(h′)m + Xh∈G∗, γ∈Γ∗
b(a + b)
2
(
h,h′a2 q1(h)m
N g
q0(u)m ) + Xu,u′∈U
q1(g − u)m
q1(h)m
q0(u)m )
q2(γ)m
q0(u)m )
N g
h,γab(
q1(h)m
q2(γ)m +
N g
=Xu∈U
+ Xu∈U h∈G∗
+ Xu∈U γ∈Γ∗
+ Xh,h′∈G∗
Xu∈U
and
N g
(u,0),g−u
a(a − b)
2
(
q0(u)m
q1(g − u)m +
q1(g − u)m
q0(u)m ) =
a(a − b)
q0(u)m
q1(g − u)m +
q1(g − u)m
q0(u)m ),
γ,ξb2 q2(γ)m
N g
q2(ξ)m ,
q2(γ)m
q1(h)m ) + Xγ,ξ∈Γ∗
2 Xu∈U
(
Xu,u′∈U
N g
(u,π),(u′,π)
(a + b)2
4
q0(u)m
q1(u′)m
=
=
(a + b)2
4 Xu,u′∈U
Γ − GXv∈U
1
4
Γ − GXv∈U
hu + u′ − g, vi
q0(u)m
q0(u′)m
h−g, viC(v, m)C(v,−m).
INFINITE FAMILIES OF MODULAR DATA
71
hu + h − g, vi + δg−h,u + δg−θ(h),u)(
q0(u)m
q1(h)m +
q1(h)m
q0(u)m )
q1(h)m
q0(u)m )
2
(
(
q0(u)m
q1(h)m +
Γ − GXv∈U
4
Xu∈U h∈G∗
a(a + b)
N g
(u,π),h
=
=
+
1
a(a + b)
2 Xu∈U h∈G∗
Γ − GXv∈U
2 Xu∈U
a(a + b)
(
(C(v, m)A(v,−m) + A(v, m)C(v,−m))
q0(u)m
q1(g − u)m +
q1(g − u)m
q0(u)m )
q2(γ)m
q0(u)m )
2
(
q0(u)m
q2(γ)m +
Γ − GXv∈U
4
Xu∈U γ∈Γ∗
b(a + b)
N g
(u,π),γ
=
=
b(a + b)
2 Xu∈U γ∈Γ∗
Γ − GXv∈U
1
(C(v, m)B(v,−m) + B(v, m)C(v,−m).
hu, vihγ, vih−g, vi(
q0(u)m
q2(γ)m +
q2(γ)m
q0(u)m )
72
PINHAS GROSSMAN AND MASAKI IZUMI
hh + h′ − g, vi)
q1(h)m
q1(h′)m
(δh+h′−g,0 + δθ(h)+h′−g,0 + δh+θ(h′)−g,0 + δh+h′−θ(g),0)
(δh+h′−g,0 + δθ(h)+h′−g,0 + δh+θ(h′)−g,0 + δh+h′−θ(g),0)
A(v, m)A(v,−m)
A(v, m)A(v,−m)
q1(h)m
q1(h′)m
q1(h)m
q1(h′)m
q1(h)m
q1(h′)m
q1(h)m
q1(h′)m
4
1
1
=
+
=
+
a2
a2
a2
−
h,h′a2 q1(h)m
N g
q1(h′)m
Γ − GXv∈U
Xh,h′∈G∗
= a2 Xh,h′∈G∗
+ a2 Xh,h′∈G∗
Γ − GXv∈U
4 Xh,h′∈G\U
Γ − GXv∈U
4 Xh,h′∈G
4 Xh∈U ;h′∈G
4 Xh,h′∈U
Γ − GXv∈U
4 Xh,h′∈G\U
Γ − GXv∈U
Γ − GXv∈U
Xh∈G∗, γ∈Γ∗
= ab Xh∈G∗, γ∈Γ∗
Γ − GXv∈U
N g
a2
a2
1
1
+
=
=
+
=
=
1
1
(δh+h′−g,0 + δθ(h)+h′−g,0 + δh+θ(h′)−g,0 + δh+h′−θ(g),0)
(δh+h′−g,0 + δθ(h)+h′−g,0 + δh+θ(h′)−g,0 + δh+h′−θ(g),0)(
q1(h)m
q1(h′)m +
q1(h′)m
q1(h)m )
(δh+h′−g,0 + δθ(h)+h′−g,0 + δh+θ(h′)−g,0 + δh+h′−θ(g),0)
A(v, m)A(v,−m)
(δh+h′−g,0 + δθ(h)+h′−g,0 + δh+θ(h′)−g,0 + δh+h′−θ(g),0)
q1(h)m
q1(h′)m
q1(h)m
A(v, m)A(v,−m) + a2Xh∈G
A(v, m)A(v,−m) + δmg,0q1(g)m − a2Xh∈U
q1(g − h)m − a2Xh∈U
(
q1(h)m
(
q1(g − h)m +
q1(g − h)m
q1(h)m )
q1(h)m
q1(g − h)m +
q1(g − h)m
q1(h)m ),
q2(γ)m
q1(h)m )
h,γab(
q1(h)m
q2(γ)m +
Γ − GXv∈U
4
hh − g, vihγ, vi(
q1(h)m
q2(γ)m +
q2(γ)m
q1(h)m )
h−g, vi(A(v, m)B(v,−m) + B(v, m)A(v,−m)),
INFINITE FAMILIES OF MODULAR DATA
73
Xγ,ξ∈Γ∗
= b2 Xγ,ξ∈Γ∗
γ,ξb2 q2(γ)m
N g
q2(ξ)m
Γ − GXv∈U
4
Thus we get
νm(g)
h−g, vihγ + ξ, vi
q2(γ)m
q2(ξ)m =
1
Γ − GXv∈U
h−g, viB(v, m)B(v,−m).
=
=
1
Γ − GXv∈U
Γ − GXv∈U
1
h−g, viG(q1, q2, v, m)G(q1, q2, v,−m) + δmg,0q1(g)m
hg, viG(q1, q2, v, m)2 + δmg,0q1(g)m.
In a similar way we can compute νm(γ).
(cid:3)
Appendix B. Calculations for Section 4
Proof of Lemma 4.2.
a2 + b2
+
a2
2
+ a2G∗ +
b2
2
+ b2Γ∗ = 1.
a2 − b2
+ a2G∗ −
b2
2 − b2Γ∗ = 0.
2
a2
2
+
Sπ,xS(g0,ε),x
Sπ,xSπ,x =
Xx
S0,xSπ,x =
S0,xS0,x =Xx
Xx
S0,xS(g0,ε),x =Xx
εδs
2√2
Xx
a
2
+
a
2
(
=
+
a2
2
2Xδ
S0,xSg,x =Xx
Xx
= a2 + a2h−g, g0i + a2 Xh∈G∗
Sπ,xSg,x
)h−g0, li + a2 Xg∈G∗
h−g0, gi =
δg0,0
2
= 0.
(h−g, hi + hg, hi) = a2Xh∈G
h−g, hi = δ0,g = 0.
Xx
S0,xS(γ0,ε),x = −Xx
b2
b
εδs
2√2
= −
2 −
2 −
2Xδ
b
(
Sπ,xS(γ0,ε),x
)h−γ0, γ0i − b2 Xγ∈Γ∗
h−γ0, γi = −
b2
2 Xγ∈Γ∗
hu − γ0, γi = −
δu,γ0
2
= 0.
74
PINHAS GROSSMAN AND MASAKI IZUMI
Xx
S0,xSγ,x = −Xx
= −b2 − b2Xδ
Sπ,xSγ,x
h−γ, γ0i − b2Xξ∈Γ∗
(h−γ, ξi + hγ, ξi) = −b2Xξ∈Γ
h−γ, ξi = −δγ,0 = 0.
Xx
S(g0,ε),xS(g0,ε′),x
=
a2
2
+Xδ
(
a
2
+
εδs
2√2
)(
a
2
+
ε′δs
2√2
) + a2G∗ +
εε′
4
=
1
2
+
εε′
2
= δε,ε.
Px S(γ0,ε),xS(γ0,ε′),x = δγ0,γ0δε,ε′ can be shown in the same way.
(
a
2
+
εδ
2√2
)hg0 − g, g0i + a2 Xh∈G∗
(hg0 − g, hi + hg0 + g, hi)
hg0 − g, hi = δg0,g = 0.
Px S(γ0,ε),xSγ0,x = 0 can be shown in the same way.
S(g0,ε),xS(γ0,ǫ′),x
hg0, g0i(
a
2
+
εδs
2√2
)
ε′δs
4 −Xδ
εδs
4 h−γ0, γ0i(
b
2 −
ε′δs
2√2
)
(hg0, g0i + hγ0, γ0i) = 0.
Px S(g0,ε),xSγ,x = 0, Px Sg,xS(γ0,ε),x = 0, and Px Sg,xSγ,x = 0 are obvious.
S(g0,ε),xSg,x
Xx
= a2 + aXδ
= a2Xh∈G
Xx
=Xδ
εε′
4√2
=
Sg,xSg′,x
Xx
= 2a2 + 2a2hg − g′, g0i + a2 Xh∈G∗
= a2(Xh∈G
hg − g′, hi)
hg + g′, hi +Xh∈G
= δg+g′,0 + δg,g′ = δg,g′.
(hg, hi + h−g, hi)(hg′, hi + h−g′, hi)
Px Sγ,xSγ′,x = δγ,γ′ can be shown in the same way. Thus S is unitary.
It is obvious that T and C are unitary. Since Sx,y = Sx,y = Sx,y and S is symmetric,
we have S2 = C. Since Tx,x = Tx,x, we have CT = T C.
INFINITE FAMILIES OF MODULAR DATA
75
The only remaining condition is (ST )3 = cC, and it suffices to verify
Xx
Sj,xSj′,xTx,x = cSj,j′Tj,jTj′,j′
for all j, j′ ∈ J.
Xx
S0,xS0,xTx,x =Xx
Sπ,xSπ,xTx,x
=
=
a2 + b2
2
+
a2
2
q1(g0) + a2 Xg∈G∗
q1(g) +
b2
2
q2(γ0) + b2 Xγ∈Γ∗
a
2G(q1) +
b
2G(q2) = cS0,0T0,0T0,0 = cSπ,πTπ,πTπ,π.
q2(γ)
S0,xSπ,xTx,x
Xx
a2 − b2
=
2
+
a2
2
q1(g0) + a2 Xg∈G∗
q1(g) −
b2
2
q2(γ0) − b2 Xγ∈Γ∗
q2(γ)
= (
a
2G(q1) −
S0,xS(g0,ε),xTx,x =Xx
Xx
b
2G(q2)) = cS0,πT0,0Tπ,π.
Sπ,xS(g0,ε),xTx,x
=
=
a2
2
+
a2
2 hg0, g0iq1(g0) + a2 Xg∈G∗
hg0, giq1(g) =
hg0, giq1(g) =
a2
2 Xg∈G
a
2G(q1)q1(g0) = cS0,(g0,ε)T0,0T(g0,ε),(g0,ε) = cSπ,(g0,ε)Tπ,πT(g0,ε),(g0,ε).
a2
2 Xg∈G
q1(g0 + g)q1(g0)
In a similar way, we can show
Xx
S0,xS(γ0,ε),xTx,x = −Xx
= cS0,(γ0,ε)T0,0T(γ0,ε),(γ0,ε) = −cSπ,(γ0,ε)Tπ,πT(γ0,ε),(γ0,ε).
Sπ,xS(γ0,ε),xTx,x
S0,xSg,xTx,x = Sπ,xSg,xTx,x
Xx
= a2 + a2hg, g0iq1(g0) + a2 Xh∈G∗
(hg, hi + h−g, hi)q1(h) = a2Xh∈G
q1(g + h)q1(g)
In a similar way, we can show
= aG(q1)q1(g) = cS0,gT0,0Tg,g = cSπ,gTπ,πTg,g.
Xx
S0,xSγ,xTx,x = −Xx
Sπ,xSγ,xTx,x = cS0,γT0,0Tγ,γ = −cSπ,γTπ,πTγ,γ.
76
PINHAS GROSSMAN AND MASAKI IZUMI
Xx
S(g0,ε),xS(g0,ε′),xTx,x
=
=
a2
2
+Xδ
2 Xg∈G
a2
(
a
2
+
q1(g) +
εδs
2√2
εε′s2
4
)(
a
2
+
ε′δs
2√2
)q1(g0) + a2 Xg∈G∗
q1(g) +
εε′s2
4
q2(γ0)
(q1(g0) + q2(γ0)) = c
a
2
+
εε′s2
√2sc
4
= cS(g0,ε),(g0,ε′)T(g0,ε),(g0,ε)T(g0,ε′),(g0,ε′).
S(g0,ε),xSg,xTx,x
Xx
= a2 + 2a2hg0 + g, g0iq1(g0) + a2 Xh∈G∗
= a2Xh∈G
= cS(g0,ε),gT(g0,ε),(g0,ε)Tg,g.
(hg0 + g, hi + hg0 − g, hi)q1(h)
hg0 + g, hiq1(h) = aG(q1)q1(g0 + g) = cahg0, giq1(g0)q1(g)
In a similar way, we can show Px S(γ0,ε),xSγ,xTx,x = cS(γ0,ε),γT(γ0,ε),(γ0,ε)Tγ,γ.
Xx
=Xδ
=
=
εε′s2
4
εε′s2
2√2
S(g0,ε),xS(γ0,ε′),xTx,x
(
a
2
+
εδs
2√2
ε′δs
2√2
)hg0, g0i
q1(g0) −Xδ
(hg0, g0iq1(g0) + hγ0, γ0iq2(γ0)) =
q1(g0) + q2(γ0)
εδs
2√2
εε′s2
2√2
(
b
2 −
ε′δs
2√2
)hγ0, γ0iq2(γ0)
q1(g0) + q2(γ0)
√2
√2
q1(g0)q2(γ0) = cS(g0,ε),(γ0,ε′)T(g0,ε),(g0,ε)T(γ0,ε′),(γ0,ε′).
It is easy to show
Xx
Xx
S(g0,ε),xSγ,xTx,x = 0 = cS(g0,ε),γT(g0,ε),(g0,ε)Tγ,γ,
Sg,xS(γ0,ε),xTx,x = 0 = cS(g,(γ0,ε)Tg,gTγ0,ε),(γ0,ε).
Xx
Sg,xSγ,xTx,x = 0 = cSg,γTg,gTγ,γ.
INFINITE FAMILIES OF MODULAR DATA
77
(hg, hi + h−g, hi)(hg′, hi + h−g′, hi)q1(h)
Sg,xSg′,xTx,x
Xx
= 2a2 + 2a2hg + g′, g0i1q1(g0) + a2 Xh∈G∗
= a2Xh∈G
= ca(hg, g′iq1(g)q1(g′) + hg,−g′iq1(g)q1(g′)
= cSg,g′Tg,gTg′,g′.
In a similar way, we can show Px Sγ,xSγ′,xTx,x = cSγ,γ′Tγ,γTγ′,γ′.
(q1(g + g′ + h)q1(g + g′) + q1(g − g′ + h)q1(g − g′))
S(γ0,ε),xS(γ0,ε′),xTx,x
Xx
=
=
b2
2
+
εε′s2
4
q1(g0) +Xδ
(
b
2 −
εδs
2√2
)(
b
2 −
ε′δs
2√2
)q2(γ0) + b2 Xγ∈Γ∗
q2(γ)
b2
2 Xγ∈Γ
q2(γ) +
εε′s2
4
(q1(g0) + q2(γ0))
= −c
b
2
+ c
εε′s
2√2
= cS(γ0,ε),(γ0,ε′)T(γ0,ε),(γ0,ε)T(γ0,ε′),(γ0,ε′).
(cid:3)
Proof of Lemma 4.3.
Nπ,π,π
+
(a + b)3
(a − b)3
4(a − b)
4(a + b)
(a + b)4 + (a − b)4
= (
= (
4(a2 − b2)
) +
a2
2
+ a2G∗ −
b2
2 − b2Γ∗
−
a2 − b2
2
) +
a2
2 G −
b2
2 Γ =
4a2b2
a2 − b2 =
4
Γ − G
.
Nπ,π,(g0,ε)
= (
= (a
+
a(a + b)2
a(a − b)2
4(a − b)
4(a + b)
(a + b)3 + (a − b)3
4(a2 − b2)
) +
a2
2
−
a2
2 hg0, g0i + a2 Xh∈G∗
2 Xh∈G
hg0, hi =
a2
) +
hg0, hi
2a2b2
a2 − b2 +
1
2
δg0,0 =
2
Γ − G
.
Nπ,π,g
= (
= (a
+
a(a − b)2
a(a + b)2
2(a − b)
2(a + b)
(a + b)3 + (a − b)3
2(a2 − b2)
) + a2hg, g0i + a2 Xh∈G∗
− a2) + a2Xh∈G
hg, hi =
(hg, hi + hg,−hi)
4a2b2
a2 − b2 + δg,0 =
4
.
Γ − G
78
PINHAS GROSSMAN AND MASAKI IZUMI
Nπ,π,(γ0,ε)
= (
= (
b(a + b)2
4(a − b) −
b2(3a2 + b2)
2(a2 − b2)
b(a − b)2
4(a + b)
+
b2
2
) −
Nπ,π,γ
hγ0, γi
b2
) −
b2
2 hγ0, γ0i − b2 Xγ∈Γ∗
2 Xγ∈Γ
a2 − b2 −
hγ0, γi =
2a2b2
δγ,0
2
=
2
Γ − G
.
= (
= (b
b(a + b)2
b(a − b)2
2(a − b) −
2(a + b)
(a + b)3 − (a − b)3
2(a2 − b2)
) − b2 Xγ0∈Σ∗
+ b2) − b2Xξ∈Γ
hγ, γ0i − b2Xξ∈Γ∗
(hγ, ξi + hγ,−ξi)
hγ, ξi =
4a2b2
a2 − b2 − δγ,0 =
4
Γ − G
.
= (
Nπ,(g0,ε),(g0,ε′)
a2(a + b)
4(a − b)
a2(a − b)
4(a + b)
= (a2 (a + b)2 + (a − b)2
+
4(a2 − b2)
Nπ,(g0,ε),g
) +Xδ
−
a2
2
) +
(
a
2
+
εδs
2√2
)(
a
2
+
1
2
=
1
Γ − G
ε′δs
2√2
1
2
+
) + a2G∗ −Xδ
εδs
2√2
ε′δs
2√2
= (
+
a2(a + b)
a2(a − b)
2(a − b)
2(a + b)
(a + b)2 + (a − b)2
= a2(
2(a2 − b2)
) + aXδ
− 1) + a2Xh∈G
hg0 + g, g0i(
a
2
+
εδs
2√2
) + a2 Xh∈G∗
hg0 + g, hi =
2
.
Γ − G
hg0, hi(hg, hi + h−g, hi)
Nπ,(g0,ε),(γ0,ε′)
ab(a + b)
= (
4(a − b) −
a2 − b2 +
a2b2
4
εε′s2
=
)hg0, g0i
1
ε′δs
2√2
εε′s2
(
ε′δs
2√2
b
2 −
εδs
2√2
+Xδ
(hg0, g0i − hγ0, γ0i)
4
+
)hγ0, γ0i
ab(a − b)
4(a + b)
) +Xδ
(
a
2
+
εδs
2√2
(hg0, g0i − hγ0, γ0i) =
Γ − G
Nπ,(g0,ε),γ =
ab
2
(
a + b
a − b −
a − b
a + b
) + bXδ
εδs
4 hγ, γ0i =
2
Γ − G
.
INFINITE FAMILIES OF MODULAR DATA
79
Nπ,g,g′
= (a2 a + b
a − b
(a + b)2 + (a − b)2
+ a2 a − b
a + b
= a2(
a2 − b2
4a2b2
) + 2a2 + a2 Xh∈G∗
− 2) + a2Xh∈G
4
=
a2 − b2 + δg+g′,0 + δg−g′,0 =
Γ − G
+ δg,g′.
(hg, hi + hg,−hi)(hg′, hi + hg′,−hi)
(hg + g′, hi + hg − g′, hi)
Nπ,g,(γ0,ε) =
ab
2
(
a + b
a − b −
a − b
a + b
) + aXδ
hg, g0i
εδs
2√2
=
2
Γ − G
.
Nπ,g,γ = (
ab(a + b)
a − b −
ab(a − b)
a + b
) = ab
(a + b)2 − (a − b)2
a2 − b2
=
4a2b2
a2 − b2 =
4
Γ − G
.
=
+
b2
4
a − b
a + b
Nπ,(γ0,ε),(γ0,ε′)
a + b
a − b
a2 + b2
a2 − b2 + 1) −
b2
2
=
(
(
) +Xδ
1
2
=
εδs
2√2
1
Γ − G −
1
2
.
ε′δs
2√2 −Xδ
(
b
2 −
εδs
2√2
)(
b
2 −
εδs
2√2
) − b2Γ∗
Nπ,(γ0,ε),γ
=
b2
2
= (
(
(
+
a − b
a + b
a + b
a − b
) − bXδ
a2 − b2 + b2) − b2Xξ∈Γ
b2(a2 + b2)
hγ0 + γ, ξi =
2
Γ − G
.
b
2 −
εδs
4
)hγ0 + γ, γ0i − b2Xξ∈Γ∗
hγ0, ξi(hγ, ξi + hγ,−ξi)
Nπ,γ,γ′
= (b2 a + b
a − b
(a + b)2 + (a − b)2
+ b2 a − b
a + b
= b2(
a2 − b2
) − 2b2hγ + γ′, γ0i − b2Xξ∈Γ∗
+ 2) − b2Xξ∈Γ
4a2b2
=
a2 − b2 − δγ+γ′,0 − δγ−γ′,0 =
4
Γ − G − δγ,γ′.
(hγ + γ′, ξi + hγ − γ′, ξi)
(hγ, ξi + hγ,−ξi)(hγ′, ξi + hγ′,−ξi)
80
PINHAS GROSSMAN AND MASAKI IZUMI
N(g0,ε),(g0,ε′),(g0,ε′′)
a
2
+
εδs
2√2
)(
a
2
+
ε′δs
2√2
)(
a
2
+
ε′′δs
2√2
)hg0, g0i
ε′′δs
2√2
(εε′ + ε′ε′′ + ε′′ε)s2
hg0, gi +
4
=
1
2(Γ − G)
hg0, g0i
εε′ + ε′ε′′ + ε′′ε
.
4
+
4
= (
a3
4(a − b)
+
a3
4(a + b)
+ a2 Xg∈G∗
2(a2 − b2) −
hg0, gi +
a2
a4
2
= (
2
bXδ
) +
(
2
) +
aXδ
ε′δs
εδs
2√2
2√2
2 Xg∈G
a2
εε′ + ε′ε′′ + ε′′ε
=
a2b2
2(a2 − b2)
+
N(g0,ε),(g0,ε′),g
= (
= (
a3
2(a − b)
+
a3
2(a + b)
a4
a2 − b2 − a2) + a2Xh∈G
(
) + 2Xδ
hg, hi +
a
2
+
εδs
2√2
)(
a
2
+
εε′s2
2 hg, g0i =
N(g0,ε),(g0,ε′),(γ0,ε′′)
)hg, g0i + a2 Xh∈G∗
ε′δs
2√2
1
εε′hg0 + g, g0i
.
+
2
Γ − G
(hg, hi + hg,−hi)
a2b
4(a + b)
) +
ε′δs
2√2
(
b
2 −
ε′′δs
2√2
ε′′(ε + ε′)s2
= (
−
=
a2b
4(a − b) −
bXδ
εδs
2√2
2
a2b2
2(a2 − b2)
+
4
a
2
(
2
+
)(
a
2
εδs
2√2
aXδ
)hγ0, γ0i
εε′s2
4 hγ0, γ0i =
−
+
ε′δs
2√2
)
ε′′δs
2√2
1
2(Γ − G)
+
ε′′(ε + ε′)hg0, g0i + εε′
4
.
N(g0,ε),(g0,ε),γ
=
=
a2b
2(a − b) −
a2 − b2 −
a2b2
ε′δs
2√2hγ, γ0i
a2b
2(a + b) − 2Xδ
εε′s2
2 hγ0, γi =
εδs
2√2
1
Γ − G −
εε′hg0, g0ihγ0, γi
2
.
INFINITE FAMILIES OF MODULAR DATA
81
+
a3
a + b
) + 2aXδ
(
a
2
+
εδs
2√2
)hg0 + g + g′, g0i
hg0, hi(hg, hi + hg,−hi)(hg′, hi + hg′,−hi)
a2 − b2 − 2) + a2Xh∈G
2a2
(hg0 + g + g′, hi + hg0 + g − g′, hi)
+ δg0+g+g′,0 + δg0+g−g′,0.
N(g0,ε),g,g′
a3
a − b
= (
+ a2 Xh∈G∗
= a2(
2
=
Γ − G
N(g0,ε),g,(γ0,ε′)
= (
a2b
2(a − b) −
=
1
Γ − G
+
)
ε′δs
2√2hg0 + g, g0i
εε′hg, g0i
2
.
+
εδs
2√2
1
Γ − G
(
a
2
a2b
2(a + b)
) + 2Xδ
εε′s2
2 hg0 + g, g0i =
a2b
a + b
a − b −
a2b
+
2
) =
.
Γ − G
N(g0,ε),g,γ = (
N(g0,ε′′),(γ0,ε),(γ0,ε′)
ab2
= (
+
ab2
4(a + b)
) +
2
4(a − b)
bXδ
ε′′δs
2√2
a2b2
2(a2 − b2)
+
=
(
2
+
a
2
ε′′δs
2√2
aXδ
ε′δs
b
2√2
2 −
ε′′(ε + ε′)s2
)
4
(
)(
εδs
2√2
b
2 −
εε′s2
4 hg0, g0i −
+
)hg0, g0i
εδs
2√2
ε′δs
2√2
=
1
2(Γ − G)
+
εε′ − ε′′(ε + ε′)hg0, g0i
4
.
N(g0,ε),(γ0,ε′),γ
= (
=
ab2
+
2(a − b)
Γ − G −
1
εδs
2√2
ab2
2(a + b)
) + 2Xδ
εε′s2
2 hγ0 + γ, γ0i =
(
b
2 −
1
ε′δs
2√2
+
Γ − G
)hγ0 + γ, γ0i
εε′hγ, γ0i
.
2
+
ab2
a + b
+ 2bhγ0, γ + γ′iXδ
sεδ
2√2
N(g0,ε),γ,γ′
ab2
a − b
2
=
=
.
Γ − G
82
PINHAS GROSSMAN AND MASAKI IZUMI
Ng,g′,g′′
) + 4a2hg + g′ + g′′, g0i
(hg, hi + h−g, hi)(hg′, hi + h−g′, hi)(hg′′, hi + h−g′′, hi)
= (
2a3
a − b
+
2a3
a + b
4a4
+ a2 Xh∈G∗
a2 − b2 − 4a2)
+ a2Xh∈G
= (
4
=
Γ − G
(hg + g′ + g′′, hi + h−g + g′ + g′′, hi + hg − g′ + g′′, hi + hg + g′ − g′′, hi)
+ δg+g′+g′′,0 + δ−g+g′+g′′,0 + δg−g′+g′′,0 + δg+g′−g′′,0.
Ng,g′,(γ0,ε) = (
a2b
a − b −
a2b
a + b
) + 2aXδ
hg + g′, g0i
δεs
2√2
=
2
Γ − G
.
Ng,g′,γ = (
2a2b
a − b −
2a2b
a + b
) =
4a2b2
a2 − b2 =
4
Γ − G
.
Ng,(γ0,ε),(γ0,ε′)
= (
=
ab2
2(a − b)
1
Γ − G
+
+
ab2
2(a + b)
) + 2Xδ
εε′s2
2 hg, g0i =
hg, g0i
1
εδs
2√2
ε′δs
2√2
εε′hg0 + g, g0i
.
+
Γ − G
ab2
a + b
) =
2
.
2
Γ − G
4
) =
4a2b2
a2 − b2 =
.
Γ − G
+
ab2
a − b
2ab2
a + b
+
Ng,(γ0,ε),γ = (
Ng,γ,γ′ = (
2ab2
a − b
N(γ0,ε),(γ0,ε′),(γ0,ε′′)
4(a + b)
= (
−
= (
b3
2
4(a − b) −
bXδ
b
2 −
b4
(
2(a2 − b2)
εδs
2√2
b2
2
+
=
1
2(Γ − G)
+
2
)(
)(
b3
) +
a X(l,δ)∈J1
ε′δs
b
2√2
2 −
2 Xγ∈Γ
b2
) −
εε′ + ε′ε′′ + ε′′ε
b
2 −
.
4
εδs
2√2
ε′′δs
2√2
ε′′δs
2√2
ε′δs
2√2
)hγ0, γ0i − b2 Xγ∈Γ∗
(εε′ + ε′ε′′ + ε′′ε)s2
hγ0, γi −
4
hγ0, γi
hγ0, γ0i
INFINITE FAMILIES OF MODULAR DATA
83
N(γ0,ε),(γ0,ε′),γ
=
= (
(
b3
b3
2(a + b) − 2Xδ
hγ, ξi −
2(a − b) −
a2 − b2 + b2) − b2Xξ∈Γ
b4
b
2 −
)(
b
2 −
εδs
2√2
εε′s2
2 hγ, γ0i =
ε′δs
2√2
)hγ, γ0i − b2Xξ∈Γ∗
1
Γ − G
+
εε′hγ0 + γ, γ0i
.
2
(hγ, ξi + hγ,−ξi)
N(γ0,ε),γ,γ′
)hγ0 + γ + γ′, γ0i
(
=
b3
b3
b
2 −
εδs
2√2
a + b − 2bXδ
hγ0, ξi(hγ, ξi + hγ,−ξi)(hγ′, ξi + hγ′,−ξi)
a − b −
− b2Xξ∈Γ∗
a2 − b2 + 2b2) − b2Xξ∈Γ
Γ − G − δγ0+γ+γ′,0 − δγ0+γ−γ′,0.
= (
2b4
=
2
(hγ0 + γ + γ′, ξi + hγ0 + γ − γ′, ξi)
Nγ,γ′,γ′′
2b3
2b3
a + b
4b4
= (
= (
) − 4b2hγ + γ′ + γ′′, γ0i
(hγ, ξi + h−γ, ξi)(hγ′, ξi + h−γ′, ξi)(hγ, ξi + h−γ′′, ξi)
a − b −
− b2Xξ∈Γ∗
a2 − b2 + 4b2)
− b2Xξ∈Γ
Γ − G − δγ+γ′+γ′′,0 − δ−γ+γ′+γ′′,0 − δγ−γ′+γ′′,0 − δγ+γ′−γ′′,0.
=
4
(hγ + γ′ + γ′′, ξi + h−γ + γ′ + γ′′, ξi + hγ − γ′ + γ′′, ξi + hγ + γ′ − γ′′, ξi)
(cid:3)
Proof of Lemma 4.7. We set
Am = 2a Xg∈G∗
q1(g)m, Bm = 2bXγ∈Γ∗
q2(γ)m.
Then
G(q1, m) + G(q2, m) = a(1 + q1(g0)m) + b(1 + q2(γ0)m) + Am + Bm.
84
PINHAS GROSSMAN AND MASAKI IZUMI
We have
νm(π)
= N π
0,π
a2 − b2
2
+ N π
π,π
(a + b)2
4
+Xε
N π
π,(g0,ε)
a(a + b)
4
(q1(g0)m + q1(g0)−m)
a(a + b)
b(a + b)
(
a2
2
) + N π
q1(g)m
q1(g0)
(q1(g)m + q1(g)−m) +Xε
(q2(γ)m + q2(γ)−m) +Xε,ε′
q1(g0)m
q1(g)m +
q1(g0)m
q2(γ)m +
q1(g0)m
q2(γ0)m +
q1(h)m + Xg∈G∗, γ∈Γ∗
q2(γ)m
q1(g0)m ) + N π
q2(γ0)m
q1(g0)m )
g,γab(
ab
2
N π
(
(
(g0,ε),(γ0,ε′)
ab
4
g,ha2 q1(g)m
N π
N π
π,(γ0,ε)
b(a + b)
4
(q2(γ0)m + q2(γ0)−m)
(γ0,ε),(γ0,ε′)
b2
4
N π
(g0,ε),(g0,ε′)
N π
a2
4
+Xε,ε′
q1(g)m
q2(γ0)m
q2(γ0)m ))
q1(g)m +
q2(γ)m
q2(γ0)m
q2(γ)m +
q2(γ0)m ))
(
ab
2
(
ab
2
(γ0,ε),g
(γ0,ε),γ
q1(g)m
q2(γ)m +
q2(γ)m
q1(g)m ) + Xγ,ξ∈Γ∗
γ,ξb2 q2(γ)m
N π
q2(ξ)m
2
2
N π
(g0,ε),g
(g0,ε),γ
(N π
(N π
N π
π,g
N π
π,γ
+ Xg∈G∗
+ Xγ∈Γ∗
+ Xg∈G∗,ε
+ Xγ∈Γ∗,ε
+Xε,ε′
+ Xg,h∈G∗
a2 − b2
Γ − G Xg∈G∗
Γ − G Xγ∈Γ∗
+ Xg∈G∗
+ Xγ∈Γ∗,ε
Γ − G
2ab
(
2a(a + b)
2b(a + b)
2a2
+
+
+
=
2
(
+
+
=
+
(a + b)2
Γ − G
+
a(a + b)
Γ − G
(q1(g0)m + q1(g0)−m)
(q1(g)m + q1(g)−m) +
(q2(γ0)m + q2(γ0)−m)
b(a + b)
Γ − G
1
(q2(γ)m + q2(γ)−m) + a2(
(
q1(g0)m
q1(g)m +
q1(g0)m
q2(γ)m +
(
Γ − G
2ab
(
Γ − G
2ab
Γ − G
q1(g)m
q1(g0)
) +
q2(γ)m
q1(g0)m ) +
4a2
Γ − G
(
ab
Γ − G
q1(g0)m
q2(γ0)m +
Γ − G Xg∈G∗, γ∈Γ∗
4ab
(
q2(γ0)m
q1(g0)m ) +
q1(g)m
q2(γ)m +
Γ − G Xg,h∈G∗
q2(γ)m
q1(g)m ) +
4b2
Γ − G Xγ,ξ∈Γ∗
1
2
)
1
+
1
2
(
) + b2(
Γ − G −
q1(g)m
q2(γ0)m ))
q2(γ)m
q2(γ0)m ))
q2(γ0)m
q1(g)m +
q2(γ0)m
q2(γ)m +
q1(g)m
q1(h)m + a2G∗
q2(γ)m
q2(ξ)m − b2Γ∗
(a + b)2 + (a + b)(q1(g0)m + q1(g0)−m + q2(γ0)m + q2(γ0)−m + Am + A−m + Bm + B−m)
Γ − G
a2 + b2 + AmA−m + AmB−m + BmA−m + BmB−m
= G(q1, m) + G(q2, m)2
Γ − G
Γ − G
.
INFINITE FAMILIES OF MODULAR DATA
85
2νm((g0, ε)) − νm(π)
a(a − b)
= 2N (g0,ε)
0,(g0,ε)
4
(2N (g0,ε)
π,(g0,δ) − N π
(q1(g0)m + q1(g0)−m) − N π
π,(g0,δ))
a(a + b)
0,π
(q1(g0)m + q1(g0)−m)
a2 − b2
2
4
(2N (g0,ε)
π,(γ0,δ) − N π
π,(γ0,δ))
b(a + b)
4
(q1(γ0)m + q1(γ0)−m)
(2N (g0,ε)
(g0,δ),(g0,δ′) − N π
(g0,δ),(g0,δ′))
(2N (g0,ε)
(g0,δ),(γ0,δ′) − N π
(g0,δ),(γ0,δ′))
(
(γ0,δ),(γ0,δ′))
b2
4
(γ0,δ),(γ0,δ′) − N π
q2(γ0)m
q1(g0)m )
+Xδ
+Xδ
+Xδ,δ′
+Xδ,δ′
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh,h′∈G∗
(2N (g0,ε)
(g0,δ),h − N π
(g0,δ),h)
(2N (g0,ε)
(g0,δ),ξ − N π
(g0,δ),ξ)
(2N (g0,ε)
h,(γ0,δ) − N π
h,(γ0,δ))
(2N (g0,ε)
(γ0,δ),γ − N π
(γ0,δ),γ)
a2
4
ab
4
(2N (g0,ε)
+Xδ,δ′
q1(g0)m
q2(γ0)m +
q1(h)m
q1(g0)m )
q2(ξ)m
q1(g0)m )
q2(γ0)m
q1(h)m )
q2(γ)m
q2(γ0)m )
(
(
(
a2
2
ab
2
ab
2
q1(g0)m
q1(h)m +
q1(g0)m
q2(ξ)m +
q1(h)m
q2(γ0)m +
q2(γ0)m
q2(γ)m +
q1(h′)m + Xξ,ξ′∈Γ∗
a2 − b2
b2
2
(
(2N (g0,ε)
h,h′ − N π
h,h′)a2 q1(h)m
(2N (g0,ε)
ξ,ξ′ − N π
ξ,ξ′)b2 q2(ξ)m
q2(ξ′)m
a(a + b)
a(a − b)
2
2
=
+
(q1(g0)m + q1(g0)−m) −
+ Xh,h′∈G∗
(2δg0+h+h′,0 + 2δg0+h−h′,0 − δh,h′)a2 q1(h)m
= a2(q1(g0)m + q1(g0)−m) − a2 + b2 + 2a2 Xh∈G∗
2
δξ,ξ′b2 q2(ξ)m
q2(ξ′)m
q1(h′)m + Xξ,ξ′∈Γ∗
q1(g0 + h)m − a2G∗ + b2Γ∗
q1(h)m
(q1(g0)m + q1(g0)−m) −
a2 − b2
2
= δmg0,0q1(g0)m.
86
PINHAS GROSSMAN AND MASAKI IZUMI
νm(g) − νm(π)
a(a − b)
= N g
0,g
(q1(g)m + q1(g)−m) − N π
0,π
a2 − b2
2
π,(g0,δ))
a(a + b)
4
(q1(g0)m + q1(g0)−m)
2
π,(g0,δ) − N π
(N g
(N g
π,h − N π
π,h)
a(a + b)
2
(q1(h)m + q1(h)−m)
(N g
π,(γ0,δ) − N π
π,(γ0,δ))
b(a + b)
4
(q1(γ0)m + q1(γ0)−m)
+Xδ
+ Xh∈G∗
+Xδ
+Xδ,δ′
+Xδ,δ′
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh,h′∈G∗
(N g
(g0,δ),(g0,δ′) − N π
(g0,δ),(g0,δ′))
(γ0,δ),(γ0,δ′) − N π
(γ0,δ),(γ0,δ′))
b2
4
(N g
(g0,δ),(γ0,δ′) − N π
(g0,δ),(γ0,δ′))
(
q2(γ0)m
q1(g0)m )
(N g
(g0,δ),h − N π
(g0,δ),h)
(N g
(g0,δ),ξ − N π
(g0,δ),ξ)
(N g
h,(γ0,δ) − N π
h,(γ0,δ))
(N g
(γ0,δ),ξ − N π
(γ0,δ),ξ)
(N g
h,h′ − N π
h,h′)a2 q1(h)m
a2
4
ab
4
(N g
+Xδ,δ′
q1(g0)m
q2(γ0)m +
q1(h)m
q1(g0)m )
q2(ξ)m
q1(g0)m )
q2(γ0)m
q1(h)m )
q2(ξ)m
q2(γ0)m )
(
(
(
a2
2
ab
2
ab
2
q1(g0)m
q1(h)m +
q1(g0)m
q2(ξ)m +
q1(h)m
q2(γ0)m +
q2(γ0)m
q2(ξ)m +
q1(h′)m + Xξ,ξ′∈Γ∗
a2 − b2
q1(g0)m
q1(h)m +
b2
2
2
(
ξ,ξ′)b2 q2(ξ)m
q2(ξ′)m
(N g
ξ,ξ′ − N π
a(a + b)
(q1(g)m + q1(g)−m) −
a2
2
+
b2
2
(δg+h+h′,0 + δ−g+h+h′,0 + δg−h+h′,0 + δg+h−h′,0 − δh,h′)
a(a − b)
2
=
+
2
q1(h)m
q1(g0)m )
(q1(g)m + q1(g)−m) −
(δg0+g+h,0 + δg0+g−h,0)(
+ a2 Xh∈G∗
+ a2 Xh,h′∈G∗
− b2 Xξ,ξ′∈Γ∗
= a2(q1(g)m + q1(g)−m) − a2 + b2 − a2G∗ + b2Γ∗ + a2(
+ a2 Xh,h′∈G∗
(δg+h+h′,0 + δ−g+h+h′,0 + δg−h+h′,0 + δg+h−h′,0)
q2(ξ)m
q2(ξ′)m
δξ,ξ′
= δmg,0q1(g)m.
q1(h)m
q1(h′)m
q1(g0 + g)m
q1(g0)m )
q1(g0)m
q1(g0 + g)m +
q1(h)m
q1(h′)m
INFINITE FAMILIES OF MODULAR DATA
87
2νm((γ0, ε)) − νm(π)
b(a − b)
= 2N (γ0,ε)
0,(γ0,ε)
4
(2N (γ0,ε)
π,(g0,δ) − N π
(q1(γ0)m + q1(γ0)−m) − N π
π,(g0,δ))
a(a + b)
0,π
(q1(g0)m + q1(g0)−m)
a2 − b2
2
4
(2N (γ0,ε)
π,(γ0,δ) − N π
π,(γ0,δ))
b(a + b)
4
(q1(γ0)m + q1(γ0)−m)
(2N (γ0,ε)
(g0,δ),(g0,δ′) − N π
(g0,δ),(g0,δ′))
(2N (γ0,ε)
(g0,δ),(γ0,δ′) − N π
(g0,δ),(γ0,δ′))
(
(γ0,δ),(γ0,δ′))
b2
4
(γ0,δ),(γ0,δ′) − N π
q2(γ0)m
q1(g0)m )
+Xδ
+Xδ
+Xδ,δ′
+Xδ,δ′
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh,h′∈G∗
(2N (γ0,ε)
(g0,δ),h − N π
(g0,δ),h)
(2N (γ0,ε)
(g0,δ),ξ − N π
(g0,δ),ξ)
(2N (γ0,ε)
h,(γ0,δ) − N π
h,(γ0,δ))
(2N (γ0,ε)
(γ0,δ),ξ − N π
(γ0,δ),ξ)
a2
4
ab
4
(2N (γ0,ε)
+Xδ,δ′
q1(g0)m
q2(γ0)m +
q1(h)m
q1(g0)m )
q2(ξ)m
q1(g0)m )
q2(γ0)m
q1(h)m )
q2(ξ)m
q2(γ0)m )
(
(
(
a2
2
ab
2
ab
2
q1(g0)m
q1(h)m +
q1(g0)m
q2(ξ)m +
q1(h)m
q2(γ0)m +
q2(γ0)m
q2(ξ)m +
q1(h′)m + Xξ,ξ′∈Γ∗
a2 − b2
b2
a2
2
2
b2
2
+
2
(
(2N (γ0,ε)
h,h′ − N π
h,h′)a2 q1(h)m
(2N (γ0,ε)
ξ,ξ′ − N π
ξ,ξ′)b2 q2(ξ)m
q2(ξ′)m
b(a − b)
b(a + b)
2
2
=
−
(q1(γ0)m + q1(γ0)−m) −
(q1(γ0)m + q1(γ0)−m) −
(δγ0+ξ+ξ′,0 + δγ0+ξ−ξ′,0)
− a2G∗ − 2b2 Xξ,ξ′∈Γ∗
= −b2(q1(γ0)m + q1(γ0)−m) − a2 + b2 − a2G∗ + b2Γ∗ − 2b2Xξ∈Γ∗
= −δmγ0,0q2(γ0)m.
q2(ξ)m
q2(ξ′)m + b2Γ∗
q2(ξ)m
q2(γ0 + ξ)m
88
PINHAS GROSSMAN AND MASAKI IZUMI
νm(γ) − νm(π)
b(a − b)
= N γ
0,γ
(q2(γ)m + q2(γ)−m) − N π
0,π
a2 − b2
2
π,(g0,δ))
a(a + b)
4
(q1(g0)m + q1(g0)−m)
2
π,(g0,δ) − N π
(N γ
(N γ
π,h − N π
π,h)
a(a + b)
2
(q1(h)m + q1(h)−m)
(N γ
π,(γ0,δ) − N π
π,(γ0,δ))
b(a + b)
4
(q1(γ0)m + q1(γ0)−m)
(N γ
π,ξ − N π
π,ξ)
b(a + b)
2
(q1(ξ)m + q1(ξ)−m)
(N γ
(g0,δ),(g0,δ′) − N π
(g0,δ),(g0,δ′))
(γ0,δ),(γ0,δ′) − N π
(γ0,δ),(γ0,δ′))
b2
4
(N γ
(g0,δ),(γ0,δ′) − N π
(g0,δ),(γ0,δ′))
(
q2(γ0)m
q1(g0)m )
+Xδ
+ Xh∈G∗
+Xδ
+Xξ∈Γ∗
+Xδ,δ′
+Xδ,δ′
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh∈G∗,δ
+ Xξ∈Γ∗,δ
+ Xh,h′∈G∗
a2
4
ab
4
(N γ
+Xδ,δ′
q1(g0)m
q2(γ0)m +
q1(h)m
q1(g0)m )
q2(ξ)m
q1(g0)m )
q2(γ0)m
q1(h)m )
q2(γ)m
q2(γ0)m )
(
(
(
a2
2
ab
2
ab
2
q1(g0)m
q1(h)m +
q1(g0)m
q2(ξ)m +
q1(h)m
q2(γ0)m +
q2(γ0)m
q2(γ)m +
q1(h′)m + Xξ,ξ′∈Γ∗
a2 − b2
2 −
q2(γ0)m
q2(ξ)m +
b2
2
(
(N γ
ξ,ξ′ − N π
b2
a2
4 −
2
+
q2(ξ)m
q2(γ0)m )
(N γ
(g0,δ),h − N π
(g0,δ),h)
(N γ
(g0,δ),ξ − N π
(g0,δ),ξ)
(N γ
h,(γ0,δ) − N π
h,(γ0,δ))
(N γ
(γ0,δ),γ − N π
(γ0,δ),γ)
(N γ
h,h′ − N π
h,h′)a2 q1(h)m
b(a − b)
2
=
b2
2
(q2(γ)m + q2(γ)−m) −
(
(δγ+γ0+ξ,0 + δγ+γ0−ξ)
− Xξ∈Γ∗,δ
− a2G∗ + b2Γ∗ − b2 Xξ,ξ′∈Γ∗
= −b2(q2(γ)m + q2(γ)−m) − b2(
− b2 Xξ,ξ′∈Γ∗
= −δmγ,0q2(γ0)m.
ξ,ξ′)b2 q2(ξ)m
q2(ξ′)m
b(a + b)
2
(q2(γ)m + q2(γ)−m)
(δγ,ξ,ξ′,0 + δ−γ+ξ+ξ′,0 + δ−γ−ξ+ξ′,0 + δγ+ξ−ξ′,0)
q2(ξ)m
q2(ξ′)m
q2(γ0)m
q2(γ + γ0)m +
q2(γ + γ0)m
q2(γ0)m )
q2(ξ)m
q2(ξ′)m
(δγ,ξ,ξ′,0 + δ−γ+ξ+ξ′,0 + δ−γ−ξ+ξ′,0 + δγ+ξ−ξ′,0)
INFINITE FAMILIES OF MODULAR DATA
89
(cid:3)
Appendix C. Calculations for Section 5
Proof of Lemma 5.4.
S(u,π),xS(u′,π),x
Xx
=
+
=
a2 + b2
S(u,0),xS(u′,0),x =Xx
2 Xv∈U
hu′ − u, vi +
hu′ − u, σi + b2 Xγ∈Γ∗
2 Xσ∈Σ∗
2 Xγ∈Γ
2 Xg∈G
hu′ − u, gi +
a2
b2
b2
= δu,u′.
hu′ − u, ki + a2 Xg∈G∗
a2
2 Xk∈K∗
hu′ − u, γi
hu′ − u, γi
=
S(u,0),xS(u′,π),x
Xx
2 Xv∈U
a2 − b2
hu′ − u, vi +
hu′ − u, σi − b2 Xγ∈Γ∗
2 Xσ∈Σ∗
2 Xγ∈Γ
2 Xg∈G
hu′ − u, gi −
−
a2
b2
b2
=
= 0.
hu′ − u, ki + a2 Xg∈G∗
a2
2 Xk∈K∗
hu′ − u, γi
hu′ − u, γi
hu′ − u, gi
hu′ − u, gi
Xx
=
=
a2
S(u,0),xS(k,ε),x =Xx
2 Xv∈U
hk − u, vi +
2 Xg∈G∗
hk − u, gi =
a2
δu,k
2
= 0.
S(u,π),xS(k,ε),x
a2
2 Xl∈K∗
hk − u, li + a2 Xg∈G∗
hk − u, gi
90
PINHAS GROSSMAN AND MASAKI IZUMI
S(u,π),xSg,x
Xx
S(u,0),xSg,x =Xx
= a2Xv∈U
= a2Xh∈G
hg − u, hi = δu,g = 0.
hg − u, vi + a2 Xk∈K∗
hg − u, ki + a2 Xh∈G∗
(hg − u, hi + hg − u, θ(h)i)
b2
S(u,0),xS(σ,ε),x = −Xx
Xx
2 Xv∈U
hσ − u, vi −
= −
2 Xγ∈Γ∗
hσ − u, γi = −
= −
b2
b2
δu,σ
2
= 0.
S(u,π),xS(σ,ε),x
2 Xτ∈Σ∗
hσ − u, τi − b2 Xγ∈Γ∗
hσ − u, γi
S(u,π),xSγ,x
S(u,0),xSγ,x = −Xx
Xx
= −b2Xv∈U
= −b2Xξ∈Γ
hγ − u, vi − b2 Xσ∈Σ∗
hγ − u, ξi = −δu,γ = 0.
hγ − u, σi − b2Xξ∈Γ∗
(hγ − u, ξi + hγ − u, ξi)
Xx
S(k,ε),xS(k′,ε′),x
=
a2
2 Xu∈U
+ a2 Xg∈G∗
2 Xg∈G
a2
=
=
δk,k′
2
+
hk′ − k, ui + X(l,δ)∈J1
hk′ − k, gi + Xσ∈Σ∗
(Xl∈K∗
hk′ − k, gi +
(Xl∈K∗
εε′
8
εε′
8
hk′ − k, li + Xσ∈Σ∗
8
= δk,k′δε,ε′.
hk′ − k, li(
a
2
+
εδs
4
)(
a
2
+
ε′δs
4
)
εε′f (k, σ)f (k′, σ)
hk′ − k, li + Xσ∈Σ∗
f (k, σ)f (k′, σ))
f (k, σ)f (k′, σ))
INFINITE FAMILIES OF MODULAR DATA
91
Px S(σ,ε),xS(σ′,ε′),x = δσ,σ′δε,ε′ can be shown in the same way.
S(k,ε),xSg,x
Xx
= a2Xu∈U
= a2Xh∈G
hk − g, ui + a2 Xl∈K∗
hk − g, li + a2 Xh∈G∗
hk − g, hi = δk,g = 0.
(hk − g, hi + hk + g, hi)
Px S(σ,ε),xSγ,x = 0 can be shown in the same way.
S(k,ε),xS(σ,ǫ′),x
Xx
= X(l,δ)∈J1
sXl∈K∗
εε′
8
=
hk, li(
a
2
+
εsδ
4
)
ε′δf (l, σ)
4
εδf (k, τ )
4
hσ, τi(
b
2
+ −ε′ε′′s
4
)
− X(τ,δ)∈J2
hk, lif (l, σ) +
εε′
8
s Xτ∈Σ∗
hσ, τif (k, τ ) = 0.
Sg,xSg′,x
Px S(k,ε),xSγ,x = 0, Px Sg,xS(σ,ε),x = 0, and Px Sg,xSγ,x = 0 are obvious.
Xx
= 2a2Xu∈U
= a2(Xh∈G
hg′ − g, ui + 2a2 Xl∈K∗
hg′ − g, hi +Xh∈G
hg′ − g, li + a2 Xh∈G∗
hg′ − θ(g), hi)
= δg′,g + δg′,θ(g) = δg,g′.
(h−g, hi + h−g, θ(h)i)(hg′, hi + hg′, θ(h)i)
Px Sγ,xSγ′,x = δγ,γ′ can be shown in the same way. Thus S is unitary.
It is obvious that T and C are unitary. Since Sx,y = Sx,y = Sx,y and S is symmetric,
we have S2 = C. Since Tx,x = Tx,x, we have CT = T C.
The only remaining condition is (ST )3 = cC, and it suffices to verify
Xx
Sj,xSj′,xTx,x = cSj,j′Tj,jTj′,j′
92
PINHAS GROSSMAN AND MASAKI IZUMI
S(u,π),xS(u′,π),xTx,x
for all j, j′ ∈ J.
Xx
b2
a2
a2 + b2
S(u,0),xS(u′,0,),xTx,x =Xx
2 Xv∈U
hu + u′, viq0(v) +
2 Xσ∈Σ∗
hu + u′, σiq2(σ) + b2 Xγ∈Γ∗
2 Xγ∈G
2 Xg∈G
2 Xg∈G
q1(g − u − u′)q0(u + u′) +
h−(u + v), giq1(g) +
a2
a2
b2
=
+
=
=
= (
a
2G(q1) +
b
2G(q2))hu, u′iq0(u)q0(u′)
hu + u′, kiq1(k) + a2 Xg∈G∗
2 Xk∈K∗
hu + u′, γiq2(γ)
hu + u′, giq1(g)
h−(u + v), γiq2(γ)
b2
2 Xγ∈Γ
q2(γ − u − u′)q0(u + u′)
= cS(u,0),(u′,0)T(u,0),(u,0)T(u′,0),(u′,0) = cS(u,π),(u′,π)T(u,π),(u,π)T(u′,π),(u′π).
=
a2
S(u,0),xS(u′,π,),xTx,x
Xx
2 Xv∈U
a2 − b2
hu + u′, viq0(v) +
2 Xσ∈Σ∗
hu + u′, σiq2(σ) − b2 Xγ∈Γ∗
2 Xg∈G
q1(g − u − u′)q0(u + u′) −
−
a2
b2
=
2 Xk∈K∗
hu + u′, γiq2(γ)
2 Xγ∈Γ
b2
= (
a
2G(q1) −
b
2G(q2))hu, u′iq0(u)q0(u′)
= cS(u,0),(u′,π)T(u,0),(u,0)T(u′,π),(u′,π).
hu + u′, kiq1(k) + a2 Xg∈G∗
hu + u′, giq1(g)
q2(γ − u − u′)q0(u + u′)
S(u,π),xS(k,ε),xTx,x
a2
2 Xl∈K∗
hu + k, liq1(l) + a2 Xg∈G∗
hu + k, giq1(g)
Xx
=
=
=
a2
S(u,0),xS(k,ε),xTx,x =Xx
2 Xv∈U
hu + k, viq0(v) +
2 Xg∈G
a
2G(q1)hu, kiq0(u)q1(k)
a2
q1(g − u − k)q1(u + k)
= cS(u,0),(k,ε)T(u,0),(u,0)T(k,ε),(k,ε) = cS(u,π),(k,ε)T(u,π),(u,π)T(k,ε),(k,ε).
INFINITE FAMILIES OF MODULAR DATA
93
In a similar way, we can show
S(u,0),xS(σ,ε),xTx,x = −Xx
Xx
= cS(u,0),(σ,ε)T(u,0),(u,0)T(σ,ε),(σ,ε) = −cS(u,π),(σ,ε)T(u,π),(u,π)T(σ,ε),(σ,ε).
S(u,π),xS(σ,ε),xTx,x
Xx
= a2Xv∈U
= a2Xh∈G
S(u,0),xSg,xTx,x = S(u,π),xSg,xTx,x
hu + g, viq0(v) + a2 Xl∈K∗
hu + g, liq1(l) + a2 Xh∈G∗
(hu + g, hi + hu + g, θ(h)i)q1(h)
q1(h − u − g)q1(u + g) = aG(q1)hu, giq0(u)q1(g)
= cS(u,0),gT(u,0),(u,0)Tg,g = cS(u,π),gT(u,π),(u,π)Tg,g.
In a similar way, we can show
S(u,0),xSγ,xTx,x = −S(u,π),xSγ,xTx,x
Xx
= cS(u,0),γT(u,0),(u,0)Tγ,γ = −cS(u,π),γT(u,π),(u,π)Tγ,γ.
Xx
S(k,ε),xS(k′,ε′),xTx,x
=
a2
2 Xu∈U
+ a2 Xg∈G∗
2 Xg∈G
a2
=
εε′
hk + k′, giq1(g) +
hk + k′, uiq0(u) + X(l,δ)∈J1
8 XσΣ∗
8 Xl∈K∗
2 Xl∈K∗
hk + k′, giq1(g) +
εε′
4
εε′s2
s2
(
= c
a
2hk, k′iq1(k)q1(k′) +
= cS(k,ε),(k′,ε′)T(k,ε),(k,ε)T(k′,ε′),(k′,ε′)
hk + k′, li(
a
2
+
εδs
4
)(
a
2
+
ε′δs
4
)q1(l)
f (k, σ)f (k′, σ)q2(σ)
hk + k′, liq1(l) +
1
hk + k′, liq1(l) +
εε′
8 Xσ∈Σ∗
2 Xσ∈Σ∗
f (k, σ)f (k′, σ)q2(σ)
f (k, σ)f (k′, σ)q2(σ))
S(k,ε),xSg,xTx,x
Xx
= a2Xu∈U
= a2Xh∈G
hk + g, uiq0(u) + a2 Xl∈K∗
hk + g, liq1(l) + a2 Xh∈G∗
(hk + g, hi + hk + g, θ(h)i)q1(h)
hk + g, hiq1(h) = aG(q1)q1(k + g) = cahk, giq1(k)q1(g)
= cS(k,ε),gT(k,ε),(k,ε)Tg,g.
94
PINHAS GROSSMAN AND MASAKI IZUMI
S(k,ε),xS(σ,ε′),xTx,x
In a similar way, we can show Px S(σ,ε),xSγ,xTx,x = cS(σ,ε),γT(σ,ε),(σ,ε)Tγ,γ.
Xx
= X(l,δ)∈J1
(sXl∈K∗
f (l, σ)hk, liq1(l) + s Xτ∈Σ∗
q1(l) − X(τ,δ)∈J2
f (k, τ )hσ, τiq2(τ ))
+ −ε′δs
ε′δf (l, σ)
)hk, li
(
a
2
+
εδs
4
=
εε′
8
(
b
2
4
εδf (k, τ )
4
)hσ, τiq2(τ )
4
= cS(k,ε),(σ,ε′)T(k,ε),(k,ε)T(σ,ε′),(σ,ε′).
It is easy to show
Xx
Xx
S(k,ε),xSγ,xTx,x = 0 = cS(k,ε),γT(k,ε),(k,ε)Tγ,γ,
Sg,xS(σ,ε),xTx,x = 0 = cS(g,(σ,ε)Tg,gTσ,ε),(σ,ε).
Xx
Sg,xSγ,xTx,x = 0 = cSg,γTg,gTγ,γ.
Sg,xSg′,xTx,x
hg + g′, liq1(l)
hg + g′, uiq0(u) + 2a2 Xl∈K∗
(hg, hi + hg, θ(h)i)(hg′, hi + hg′, θ(h)i)q1(h)
Xx
= 2a2Xu∈U
+ a2 Xh∈G∗
= a2Xh∈G
= a2Xh∈G
= ca(hg, g′iq1(g)q1(g′) + hg, θ(g′)iq1(g)q1(g′)
= cSg,g′Tg,gTg′,g′.
(hg + g′, hi + hg + θ(g′), hi)q1(h)
(q1(g + g′ − h)q1(g + g′) + q1(g + θ(g′) − h)q1(g + θ(g′)))
INFINITE FAMILIES OF MODULAR DATA
95
In a similar way, we can show Px Sγ,xSγ′,xTx,x = cSγ,γ′Tγ,γTγ′,γ′.
Xx
S(σ,ε),xS(σ′,ε′),xTx,x
εε′
8 Xl∈K∗
)(
b
2 −
f (l, σ)f (l, σ′)q1(l)
ε′δs
4
)q2(τ )
=
b2
b
2 −
hσ + σ′, τi(
hσ + σ′, uiq0(u) +
εδs
4
2 Xu∈U
+ X(τ,δ)∈J2
+ b2 Xγ∈Γ∗
2 Xγ∈Γ
b
2hσ, σ′iq2(σ)q2(σ′) +
hσ + σ′, γiq2(γ) +
hσ + σ′, γiq2(γ)
= −c
b2
=
= cS(σ,ε),(σ′,ε′)T(σ,ε),(σ,ε)T(σ′,ε′),(σ′,ε′).
εε′
8 Xl∈K∗
2 Xl∈K∗
1
(
εε′
4
f (l, σ)f (l, σ′)q1(l) +
f (l, σ)f (l, σ′)q1(l) +
εε′
8
s2 Xτ∈Σ∗
s2 Xτ∈Σ∗
1
2
hσ + σ′, τiq2(τ )
hσ + σ′, τiq2(τ ))
(cid:3)
Proof of Lemma 5.5. Since S(u,0),x = hu, xiS(0,0),x in an obvious sense, those equations
involving (u, 0) can be shown as in the proof of the unitarity of S. Recall hu, vi = 1
for all u, v ∈ U.
N(u,π),(u′,π),(u′′,π)
hu + u′ + u′′, ki
hu + u′ + u′′, γi
+
= (
(a + b)3
4(a − b)
(a − b)3
4(a + b)
)Xv∈U
hu + u′ + u′′, gi −
= (
+ a2 Xg∈G∗
(a + b)4 + (a − b)4
2 Xg∈G
a2 − b2 Xv∈U
4(a2 − b2)
hu + u′ + u′′, gi −
4a2b2
−
a2
=
+
a2
2 Xk∈K∗
hu + u′ + u′′, vi +
b2
2 Xσ∈Σ∗
hu + u′ + u′′, σi − b2 Xγ∈Γ∗
)Xv∈U
a2 − b2
hu + u′ + u′′, vi
2 Xγ∈Γ
hu + u′ + u′′, γi
δu+u′+u′′,0
δu+u′+u′′,0
b2
2
hu + u′ + u′′, vi +
2
−
2
8
=
.
Γ − G
96
PINHAS GROSSMAN AND MASAKI IZUMI
N(u,π),(u′,π),(k,ε)
a(a + b)2
4(a − b)
= (
+
+ a2 Xh∈G∗
= (a
hu + u′ + k, hi
(a + b)3 + (a − b)3
a2 − b2 Xv∈U
)Xv∈U
hu + u′ + k, vi +
4(a2 − b2)
2a2b2
a2
2
−
1
2
=
a(a − b)2
4(a + b)
)Xv∈U
hu + u′ + k, vi +
a2
2 Xl∈K∗
hu + u′ + k, li
hu + u′ + k, vi +
2
δu+u′+k,0 =
a2
2 Xh∈G
hu + u′ + k, hi
(1 + hk, u0i).
Γ − G
+
N(u,π),(u′,π),g
a(a + b)2
2(a − b)
(
+ a2 Xh∈G∗
a(a − b)2
2(a + b)
)Xv∈U
hu + u′ + g, vi + a2 Xl∈K∗
hu + u′ + g, li
(hu + u′ + g, hi + hu + u′ + g, θ(h)i)
− a2)Xv∈U
hu + u′ + g, vi + a2Xh∈G
hu + u′ + g, hi
= (a
2(a2 − b2)
(a + b)3 + (a − b)3
a2 − b2 Xv∈U
4a2b2
4
Γ − G
(1 + hg, u0i).
=
=
hu + u′ + g, vi + δu+u′+g,0
b(a − b)2
4(a + b)
)Xv∈U
hu + u′ + σ, vi −
b2
2 Xτ∈Σ∗
hu + u′ + σ, τi
hu + u′ + σ, vi −
b2
2 Xγ∈Γ
hu + u′ + σ, γi
hu + u′ + σ, vi −
δu+u′+γ,0
2
N(u,π),(u′,π),(σ,ε)
b(a + b)2
= (
4(a − b) −
− b2 Xγ∈Γ∗
= (
=
=
b2(3a2 + b2)
2(a2 − b2)
a2 − b2 Xv∈U
2a2b2
2
Γ − G
hu + u′ + σ, γi
)Xv∈U
b2
2
+
(1 + hσ, u0i).
INFINITE FAMILIES OF MODULAR DATA
97
b(a − b)2
2(a + b)
)Xv∈U
hu + u′ + γ, vi − b2 Xτ∈Σ∗
hu + u′ + γ, τi
(hu + u′ + γ, ξi + hu + u′ + γ, θ(ξ)i)
+ b2)Xv∈U
hu + u′ + γ, vi − b2Xξ∈Γ
hu + u′ + γ, ξi
hu + u′ + γ, vi − δu+u′+γ,0
N(u,π),(u′,π),γ
b(a + b)2
= (
2(a − b) −
− b2Xξ∈Γ∗
= (b
2(a2 − b2)
(a + b)3 − (a − b)3
a2 − b2 Xv∈U
4a2b2
4
Γ − G
(1 + hγ, u0i).
=
=
hu + k + k′, li(
a
2
+
εδs
4
)(
a
2
+
ε′δs
4
)
hu + k + k′, hi
N(u,π),(k,ε),(k,ε′)
a2(a + b)
4(a − b)
= (
+ a2 Xh∈G∗
= (a2 (a + b)2 + (a − b)2
+
−
a2
2
a2(a − b)
4(a + b)
)Xv∈U
hu + k + k′, hi − X(σ,δ)∈J2
)Xv∈U
4(a2 − b2)
hu + k + k′, li − Xσ∈Σ∗
(1 + hk + k′, u0i) +
hu + k + k′, li − Xσ∈Σ∗
δu+k+k′,0
+
εε′s2
+
εε′
8
(s2 Xl∈K∗
Γ − G
εε′s2
(Xl∈K∗
8
1
2
+
=
+
=
hu + k + k′, vi + X(l,δ)∈J1
εδf (k, σ)
ε′δf (k′, σ)
hu, σi
4
4
a2
hu + k + k′, vi +
2 Xh∈G
f (k, σ)f (k′, σ)hu, σi)
δu+k+k′,0
2
hk′ − k, σihu, σi)
Γ − G
2
(hu + k + k′, k1i − hu + k + k′, σ1ih2k, σ1i).
4
98
PINHAS GROSSMAN AND MASAKI IZUMI
N(u,π),(k,ε),g
= (
a2(a + b)
2(a − b)
+
a2(a − b)
2(a + b)
)Xv∈U
hu + k + g, vi + a X(l,δ)∈J1
hu + k + g, li(
a
2
+
εδs
4
)
hu + k, hi(hg, hi + hg, θ(h)i)
= a2(
(a + b)2 + (a − b)2
− 1)Xv∈U
hu + k + g, vi + a2Xh∈G
hu + k + g, hi
2(a2 − b2)
(1 + hk + g, u0i) + δu+k+g,0
(1 + hk + g, u0i).
ab(a − b)
4(a + b)
)Xv∈U
hu, vihk, vihσ, vi + X(l,δ)∈J1
(
a
2
+
εδs
4
)hu + k, li
ε′δf (l, σ)
4
+ a2 Xh∈G∗
=
=
2
2
Γ − G
Γ − G
N(u,π),(k,ε),(σ,ε′)
ab(a + b)
= (
4(a − b) −
4
=
a2b2
+ X(τ,δ)∈J2
a2 − b2 Xv∈U
Γ − GXv∈U
Γ − GXv∈U
=
=
1
1
εδf (k, τ )
(
b
2 −
ε′δs
4
hk, vihσ, vi +
)hu + σ, τi
sεε′
8
(Xl∈K∗
8 Xv∈U
sεε′
f (l, σ)hu + k, li − Xτ∈Σ∗
f (k, τ )hu + σ, τi)
hk, vihσ, vi +
(f (k + v, σ)hu + k, k + vi − f (k, σ + v)hu + σ, σ + vi)
hk, vihσ, vi +
1
sf (k, σ)εε′
8
(hk, kihu, ki − hσ, σihu, σi)Xv∈U
hv, kihv, σi
sf (k, σ)εε′
+
8
(hk, kihu, ki − hσ, σihu, σi))
= (1 + hu0, kihu0, σi)(
= 0.
Γ − G
N(u,π),(k,ε),γ
a + b
=
ab
2
(
a − b −
=
2
Γ − G
(1 + hk, u0ihγ, u0i).
a − b
a + b
)Xv∈U
hu + k, vihγ, vi + b X(σ,δ)∈J2
εδf (k, σ)
4
hγ, σi
INFINITE FAMILIES OF MODULAR DATA
99
+ a2 a − b
a + b
)Xv∈U
hu + g + g′, vi + 2a2 Xl∈K∗
hu + g + g′, li
hu, hi(hg, hi + hg, θ(h)i)(hg′, hi + hg′, θ(h)i)
− 2)Xv∈U
hu + g + g′, vi + a2Xh∈G
(hu + g + g′, hi + hu + g + θ(g′), hi)
N(u,π),g,g′
= (a2 a + b
a − b
+ a2 Xh∈G∗
= a2(
4a2b2
a2 − b2
(a + b)2 + (a − b)2
a2 − b2 Xv∈U
Γ − GXv∈U
4
=
=
hu + g + g′, vi + δu+g+g′,0 + δu+g+θ(g′),0
hu + g + g′, vi + δu+g+g′,0 + δu+g+θ(g′),0.
N(u,π),g,(σ,ε)
a + b
=
ab
2
(
a − b −
=
2
Γ − G
(1 + hg, u0ihσ, u0i).
a − b
a + b
)Xv∈U
hg, vihu + σ, vi + a X(l,δ)∈J1
hg, lihu, li
εδf (l, σ)
4
N(u,π),g,γ
hu, vihg, vihγ, vi
hu, vihg, vihγ, vi
= (
= ab
ab(a + b)
)Xv∈U
a2 − b2 Xv∈U
ab(a − b)
a − b −
a + b
(a + b)2 − (a − b)2
a2 − b2 Xv∈U
Γ − GXv∈U
hu, vihg, vihγ, vi
4a2b2
4
hu, vihg, vihγ, vi.
=
=
100
PINHAS GROSSMAN AND MASAKI IZUMI
N(u,π),(σ,ε),(σ′,ε′)
=
b2
4
(
− X(τ,δ)∈J2
+
a + b
a − b
b
(
2 −
a2 + b2
a − b
a + b
εδs
4
)(
)Xv∈U
b
2 −
hu + σ + σ′, vi + X(l,δ)∈J1
)hu + σ + σ′, τi − b2Xξ∈Γ∗
εδs
4
εδf (l, σ)
ε′δf (l, σ′)
4
4
hu + σ + σ′, ξi
hu, li
=
+
=
+
=
(
b2
2
a2 − b2 + 1)Xv∈U
8 Xl∈K∗
f (l, σ)f (l, σ′)hu, li −
hu + σ + σ′, vi −
8 Xτ∈Σ∗
εεs2
εε′
hu + σ + σ′, ξi
b2
2 Xξ∈Γ
hu + σ + σ′, τi
(1 + hu + σ + σ′, u0i) −
hu + σ + σ′, lih2σ, li − Xτ∈Σ∗
δu+σ+σ′,0
εε′s2
δu+σ+σ′,0
2
hu + σ + σ′, τi)
1
8
Γ − G
εε′s2
(Xl∈K∗
Γ − G −
2
2
+
(hu + σ + σ′, k1ih2σ, k1i − hu + σ + σ′, σi)
4
N(u,π),(σ,ε),γ
a + b
a − b
b2
2
=
(
− b2Xξ∈Γ∗
2
=
Γ − G
+
a − b
a + b
)Xv∈U
hu + σ + γ, vi − b X(τ,δ)∈J2
(
b
2 −
εδs
4
)hu + σ + γ, τi
hu + σ, ξi(hγ, ξi + hγ, θ(ξ)i)
(1 + hu + σ + γ, u0i) − δu+σ+γ,0.
INFINITE FAMILIES OF MODULAR DATA
101
N(u,π),γ,γ′
= (b2 a + b
a − b
− b2Xξ∈Γ∗
= b2(
+ +b2 a − b
a + b
hu + γ + γ′, vi − 2b2 Xτ∈Σ∗
hu, ξi(hγ, ξi + hγ, θ(ξ)i)(hγ′, ξi + hγ′, θ(ξ)i)
)Xv∈U
hu + γ + γ′, τi
4a2b2
a2 − b2
(a + b)2 + (a − b)2
a2 − b2 Xv∈U
Γ − GXv∈U
4
=
=
+ 2)Xv∈U
hu + γ + γ′, vi − b2Xξ∈Γ
hu + γ + γ′, vi − δu+γ+γ′,0 − δu+γ+θ(γ′),0
hu + γ + γ′, vi − δu+γ+γ′,0 − δu+γ+θ(γ′),0.
(hu + γ + γ′, ξi + hu + γ + θ(γ′), ξi)
+
a3
4(a + b)
)Xv∈U
hk + k′ + k′′, vi
εδs
4
)(
a
2
+
ε′δs
4
)(
a
2
+
ε′′δs
4
N(k,ε),(k′,ε′),(k′′,ε′′)
a3
= (
+
+
4(a − b)
2
a
(
2
a X(l,δ)∈J1
+ a2 Xg∈G∗
2(a2 − b2) −
= (
2
b X(τ,δ)∈J2
4
a2
2
hk + k′ + k′′, gi +
)Xv∈U
a4
hk + k′ + k′′, vi +
Xl∈K∗
hk + k′ + k′′, li
8
s2(εε′ + ε′ε′′ + ε′′ε)
+
=
a2b2
2(a2 − b2)Xv∈U
= (1 + hu0, k + k′ + k′′i)(
2
+
1
2(Γ − G)
)hk + k′ + k′′, li
ε′δf (k′, τ )
εδf (k, τ )
ε′′δf (k′′, τ )
4
4
hk + k′ + k′′, gi
a2
2 Xg∈G
hk + k′ + k′′, vi +
δk+k′+k′′,0
+
s2(εε′ + ε′ε′′ + ε′′ε)
8
Xl∈K∗
hk + k′ + k′′, li
s2(εε′ + ε′ε′′ + ε′′ε)
8
hk + k′ + k′′, k + k′ + k′′i)
102
PINHAS GROSSMAN AND MASAKI IZUMI
N(k,ε),(k′,ε′),g
+
a3
2(a + b)
)Xv∈U
+
)(
+
hv, k + k′ + gi
a
2
ε′δs
4
)hl, k + k′ + gi + a2 Xh∈G∗
hv, k + k′ + gi + a2Xh∈G
hh, k + k′ + gi
hh, k + k′i(hh, gi + hh, θ(g)i)
a3
= (
a4
= (
2(a − b)
a
(
2
s2εε′
εδs
4
+ 2 X(l,δ)∈J1
a2 − b2 − a2)Xv∈U
4 Xl∈K∗
Γ − GXv∈U
Γ − GXv∈U
1
1
=
=
+
= (
1
Γ − G
+
hl, k + k′ + gi
hl, k + k′ + gi
s2εε′
4 Xl∈K∗
hk + v, k + k′ + gi
hv, k + k′ + gi + δk+k′+g,0 +
4 Xv∈U
hv, k + k′ + gi +
s2εε′hk, k + k′ + gi
s2εε′
4
)(1 + hg, u0i).
INFINITE FAMILIES OF MODULAR DATA
103
N(k,ε),(k′,ε′),(σ,ε′′)
a2b
4(a + b)
)Xv∈U
hk + k′, vihσ, vi
hk + k′, li(
a
2
+
εδs
4
)(
a
2
+
ε′δs
4
)
ε′′δf (l, σ)
4
εδf (k, τ )
ε′δf (k′, τ )
4
(
b
2 −
ε′′δs
4
)hσ, τi
ε′′(ε + ε′)s
8 Xl∈K∗
hk + k′, vihσ, vi +
hk + k′, lif (l, σ)
a2b
2
2
4(a − b) −
a X(l,δ)∈J1
b X(τ,δ)∈J2
2(a2 − b2)Xv∈U
8 Xτ∈Σ∗
a2b2
εε′
4
= (
+
−
=
−
=
−
= (1 + hσ, u0i)(
= (1 + hσ, u0i)(
N(k,ε),(k′,ε′),γ
f (k, τ )f (k′, τ )hσ, τi
1
2(Γ − G)
εε′
8 Xv∈U
ε′′(ε + ε′)s
(1 + hσ, u0i) +
8 Xv∈U
f (k, σ + v)f (k′, σ + v)hσ, σ + vi.
ε′′(ε + ε′)s
1
hk + k′, k + vif (k + v, σ)
+
+
8
ε′′(ε + ε′)s
8
hk + k′, kif (k, σ) −
hk + k′, kif (k, σ) −
f (k, σ)f (k′, σ)hσ, σi)
εε′
8
s2εε′
8 hk′ − k, σihσ, σi).
2(Γ − G)
1
2(Γ − G)
εδf (k, τ )
ε′δf (k′, τ )
4
4
hτ, γi
= (
=
=
a2b
a2b2
2(a − b) −
a2 − b2 Xv∈U
Γ − GXv∈U
1
= (1 + hγ, u0i)(
a2b
2(a + b)
)Xv∈U
hv, k + k′ihv, γi − 2 X(τ,δ)∈J2
εε′
4 Xτ∈Σ∗
hk′ − k, τihτ, γi
hv, k + k′ihv, γi −
4 Xτ∈Σ∗
εε′s2
4 hk′ − k, σ1ihγ, σ1i).
hv, γi −
εε′s2
Γ − G −
1
f (k, τ )f (k′, τ )hτ, γi
104
PINHAS GROSSMAN AND MASAKI IZUMI
N(k,ε),g,g′
a3
a − b
= (
+ a2 Xh∈G∗
= a2(
2
=
Γ − G
+
a3
a + b
)Xv∈U
hk + g + g′, vi + 2a X(l,δ)∈J1
(
a
2
+
εδs
4
)hk + g + g′, li
hk, hi(hg, hi + hg, θ(h)i)(hg′, hi + hg′, θ(h)i)
a2 − b2 − 2)Xv∈U
2a2
hk + g + g′, vi + a2Xh∈G
(1 + hk + g + g′, u0i) + δk+g+g′,0 + δk+g−g′,0.
(hk + g + g′, hi + hk + g + θ(g′), hi)
N(k,ε),g,(σ,ε′)
a2b
= (
2(a − b) −
Γ − GXv∈U
1
=
= (
1
Γ − G
+
a2b
2(a + b)
)Xv∈U
hk + g, vihσ, vi + 2 X(δ,l)∈J1
(
a
2
+
εδs
4
)
ε′δf (l, σ)
4
hk + g, li
εε′s
4 Xl∈K∗
hk + g, vihσ, vi +
εε′sf (k, σ)hk + g, ki
4
f (l, σ)hk + g, li
)(1 + hk + g, u0ihσ, u0i).
N(k,ε),g,γ = (
a2b
a − b −
a2b
a + b
)Xv∈U
hk + g, vihγ, vi =
2
Γ − G
(1 + hk + g, u0ihγ, u0i).
INFINITE FAMILIES OF MODULAR DATA
105
N(k,ε′′),(σ,ε),(σ′,ε′)
+
ab2
4(a + b)
)Xv∈U
ε′′δs
)hk, li
4
hk, uihσ + σ′, ui
ε′δf (l, σ′)
εδf (l, σ)
4
4
ε′′δf (k, τ )
(
b
2 −
εδs
4
)(
b
2 −
ε′δs
4
)hσ + σ′, τi
f (l, σ)f (l, σ′)hk, li −
sε′′(ε + ε′)
8 Xτ∈Σ∗
f (k, τ )hσ + σ′, τi
hk, uihσ + σ′, ui
hk, uihσ + σ′, ui
f (k, σ)f (k, σ′)hσ + σ′, vihk, k + vi −
1
εε′s2hσ′ − σ, kihk, ki
+
8
sε′′(ε + ε′)
8 Xv∈U
f (k, σ)hk, vihσ + σ′, σ + vi
sε′′(ε + ε′)f (k, σ)hσ + σ′, σi
8
−
)(1 + hk, u0i).
ab2
= (
2
4
+
a2b2
4(a − b)
2
a
(
2
a X(l,δ)∈J1
b X(τ,δ)J2
2(a2 − b2)Xv∈U
8 Xl∈K∗
2(Γ − G)Xv∈U
8 Xv∈U
εε′
εε′
1
+
+
=
+
=
+
= (
2(Γ − G)
N(k,ε),(σ,ε′),γ
ab2
= (
2(a − b)
=
= (
1
Γ − GXv∈U
Γ − G −
1
+
ab2
2(a + b)
)Xv∈U
hk, vihσ + γ, vi + 2 X(τ,δ)∈J2
εδf (k, τ )
4
(
b
2 −
ε′δs
4
)hσ + γ, τi
εε′s
4 Xτ∈Σ∗
hk, vihσ + γ, vi −
εε′sf (k, σ)hσ + γ, σi
4
f (k, τ )hσ + γ, τi
)(1 + hk, u0ihσ + γ, u0i).
= (
N(k,ε),γ,γ′
ab2
a − b
2
=
Γ − G
+
ab2
a + b
)Xv∈U
hk, vihγ + γ′, vi +
2b2
a X(σ,δ)∈J2
εδf (k, σ)
4
hγ + γ′, σi
(1 + hk, u0ihγ + γ′, u0i).
PINHAS GROSSMAN AND MASAKI IZUMI
hg + g′ + g′′, vi + 4a2 Xl∈K∗
hg + g′ + g′′, ki
(hg, hi + hg, θ(h)i)(hg′, hi + hg′, θ(h)i)(hg′′, hi + hg′′, θ(h)i)
hg + g′ + g′′, vi
106
Ng,g′,g′′
= (
2a3
a − b
+
2a3
a + b
)Xv∈U
= (
4a4
+ a2 Xh∈G∗
a2 − b2 − 4a2)Xv∈U
+ a2Xh∈G
a2 − b2 Xv∈U
4a2b2
=
4
=
Γ − G
(hg + g′ + g′′, hi + hθ(g) + g′ + g′′, hi + hg + θ(g′) + g′′, hi + hg + g′ + θ(g′′), hi)
hg + g′ + g′′, vi + δg+g′+g′′,0 + δθ(g)+g′+g′′,0 + δg+θ(g′)+g′′,0 + δg+g′+θ(g′′),0
(1 + hg + g′ + g′′, u0i) + δg+g′+g′′,0 + δθ(g)+g′+g′′,0 + δg+θ(g′)+g′′,0 + δg+g′+θ(g′′),0.
Ng,g′,(σ,ε)
a2b
= (
a − b −
=
2
Γ − G
a2b
a + b
)Xv∈U
hg + g′, vihσ, vi + 2a X(l,δ)∈J1
hg + g′, li
δεf (l, σ)
4
(1 + hg + g′, u0ihσ, u0i).
Ng,g′,γ
hg + g′, vihγ, vi
2a2b
a − b −
a2 − b2 Xv∈U
4a2b2
2a2b
a + b
)Xv∈U
hg + g′, vihγ, vi
= (
=
=
4
Γ − G
(1 + hg + g′, u0ihγ, u0i).
INFINITE FAMILIES OF MODULAR DATA
107
Ng,(σ,ε),(σ′,ε′)
εδf (l, σ)
ε′δf (l, σ′)
4
4
hg, li
ab2
= (
+
ab2
2(a + b)
2(a − b)
1
Γ − GXv∈U
Γ − GXv∈U
1
=
=
= (
1
Γ − G
+
hg, vi +
εε′
hg, vihσ + σ′, vi + 2 X(l,δ)∈J1
)Xv∈U
4 Xl∈K∗
hg, lif (l, σ)f (l, σ′)
4 Xv∈U
εε′s2
hg, k1 + vihk1 + v, σ′ − σi
hg, vi +
εε′s2hg, k1ihk1, σ′ − σi
4
)(1 + hg, u0i).
Ng,(σ,ε),γ = (
ab2
a − b
+
ab2
a + b
)Xv∈U
hg, vihσ + γ, vi =
2
Γ − G
(1 + hg, u0ihσ + γ, u0i).
Ng,γ,γ′
= (
+
2ab2
a − b
4a2b2
a2 − b2 Xv∈U
4
Γ − G
hg, vihγ + γ′, vi
2ab2
a + b
)Xv∈U
hg, vihγ + γ′, vi
(1 + hg, u0ihγ + γ′, u0i).
=
=
N(σ,ε),(σ′ ,ε′),(σ′′,ε′′)
= (
+
−
b3
2
4(a − b) −
a X(l,δ)∈J1
b X(τ,δ)∈J2
2
(
b
2 −
b3
4(a + b)
)Xv∈U
εδf (l, σ)
ε′δf (l, σ′)
hσ + σ′ + σ′′, vi
ε′′δf (l, σ′′)
4
4
4
εδs
4
)(
b
2 −
ε′δs
4
)(
b
2 −
ε′′δs
4
+
b4
= (
b2
2
2(a2 − b2)
s2(εε′ + ε′ε′′ + ε′′ε)
)Xv∈U
hσ + σ′ + σ′′, vi −
Xτ∈Σ∗
= (1 + hσ + σ′ + σ′′, u0i)(
2(Γ − G) −
hσ + σ′ + σ′′, τi
−
8
1
)hσ + σ′ + σ′′, τi − b2 Xγ∈Γ∗
2 Xγ∈Γ
hσ + σ′ + σ′′, γi
b2
hσ + σ′ + σ′′, γi
s2(εε′ + ε′ε′′ + ε′′ε)
8
hσ + σ′ + σ′′, σ + σ′ + σ′′i).
108
PINHAS GROSSMAN AND MASAKI IZUMI
N(σ,ε),(σ′ ,ε′),γ
b3
2(a + b)
)Xv∈U
hσ + σ′ + γ, vi − 2 X(τ,δ)∈J2
(
b
2 −
εδs
4
)(
b
2 −
ε′δs
4
)hσ + σ′ + γ, τi
hσ + σ′, ξi(hγ, ξi + hγ, θ(ξ)i)
hσ + σ′ + γ, vi − b2Xξ∈Γ
hσ + σ′ + γ, ξi −
εε′s2
4 Xτ∈Σ∗
hσ + σ′ + γ, τi
εε′s2hσ + σ′ + γ, σi
4
)(1 + hσ + σ′ + γ, u0i).
b3
)Xv∈U
hσ + γ + γ′, vi − 2b X(τ,δ)∈J2
(
b
2 −
εδs
4
)hσ + γ + γ′, τi
hσ, ξi(hγ, ξi + hγ, θ(ξ)i)(hγ′, ξi + hγ′, θ(ξ)i)
hσ + γ + γ′, vi − b2Xξ∈Γ
(hσ + γ + γ′, ξi + hσ + γ + θ(γ′), ξi)
(1 + hσ + γ + γ′, u0i) − δσ+γ+γ′,0 − δσ+γ+θ(γ′),0.
b3
= (
2(a − b) −
− b2Xξ∈Γ∗
a2 − b2 + b2)Xv∈U
Γ − G −
= (
= (
b4
1
N(σ,ε),γ,γ′
b3
= (
a + b
a − b −
− b2Xξ∈Γ∗
a2 − b2 + 2b2)Xv∈U
= (
2b4
2
=
Γ − G
Nγ,γ′,γ′′
2b3
2b3
a + b
4b4
= (
= (
a − b −
− b2Xξ∈Γ∗
a2 − b2 + 4b2)Xv∈U
− b2Xξ∈Γ
a2 − b2 Xv∈U
4a2b2
=
4
=
Γ − G
)Xv∈U
hγ + γ′ + γ′′, vi − 4b2 Xτ∈Σ∗
hγ + γ′ + γ′′, τi
(hγ, ξi + hγ, θ(ξ)i)(hγ′, ξi + hγ′, θ(ξ)i)(hγ, ξi + hγ′′, θ(ξ)i)
hγ + γ′ + γ′′, vi
(hγ + γ′ + γ′′, ξi + hθ(γ) + γ′ + γ′′, ξi + hγ + θ(γ′) + γ′′, ξi + hγ + γ′ + θ(γ′′), ξi)
hγ + γ′ + γ′′, vi − δγ+γ′+γ′′,0 − δθ(γ)+γ′+γ′′,0 − δγ+θ(γ′)+γ′′,0 − δγ+γ′+θ(γ′′),0
(1 + hγ + γ′ + γ′′, u0i) − δγ+γ′+γ′′,0 − δθ(γ)+γ′+γ′′,0 − δγ+θ(γ′)+γ′′,0 − δγ+γ′+θ(γ′′),0.
(cid:3)
INFINITE FAMILIES OF MODULAR DATA
109
Proof of Lemma 5.7.
νm((u, 0))
N (u,0)
(u′,0),(u−u′,0)
(a − b)2
4
q0(u′)m
q0(u − u′)m + Xu′∈U
N (u,0)
(u′,π),(u−u′,π)
(a + b)2
q0(u′)m
4
q0(u − u′)m
q2(γ)m
q2(u − γ)m
N (u,0)
(σ,ε),(σ−u,ε)
b2
4
q2(σ)m
q1(σ + u)m
= Xu′∈U
+ Xg∈G∗
+ X(k,ε)∈J1
2 Xg∈G
a2
=
N (u,0)
g,u−ga2
q1(g)m
q1(u − g)m + Xγ∈Γ∗
q1(k)m
N (u,0)
γ,u−γb2
q1(k + u)m + X(σ,ε)∈J2
2 Xγ∈Γ
q2(u − γ)m .
q1(γ)m
b2
N (u,0)
(k,ε),(k−u,ε)
a2
4
q0(g)m
q0(u − g)m +
Since q1(g)/q1(u − g) = hu, giq0(u), we have
a2
2 Xg∈G
q0(g)m
q0(u − g)m =
1
2GXg∈G
hmu, giq0(u)−m =
δmu,0q0(u)m
2
,
and νm((u, 0)) = δmu,0q0(u)m.
Note that we have G(q1, q2, v, m) = G(q1, q2,−v, m) and G(q1, q2, v,−m) = G(q1, q2, v, m).
We set
A(g, m) = 2a Xh∈G∗
hg, hiq1(h)m, B(γ, m) = 2bXξ∈Γ∗
hγ, ξiq2(ξ)m.
C(g, m) = a Xk∈K∗
hg, kiq0(k)m, D(γ, m) = b Xσ∈Σ∗
hγ, σiq2(σ)m.
Then
G(q1, q2, v, m) = (a + b)(1 + q0(u0)m) + A(v, m) + B(v, m) + C(v, m) + D(v, m).
In the computation below, we often use the following facts:
(1) hu, vi = 1 for any u, v ∈ U,
(2) hk, u0ihσ, u0i = −1 for any k ∈ K∗ and σ ∈ Σ∗,
(3) hu, k + li = 1 for any u ∈ U and k, l ∈ K∗,
(4) hu, σ + τi = 1 for any σ, τ ∈ Σ∗.
110
PINHAS GROSSMAN AND MASAKI IZUMI
For νm((u, π)), we have
νm((u, π))
a2 − b2
(
a(a + b)
4
2
q0(v)m
(
q0(u − v)m +
q0(v)m
q1(g)m +
q0(v)m
q1(l)m +
(
4
a(a + b)
(v,π),(l,ε)
b(a + b)
4
(
q0(v)m
q2(τ )m +
q1(g)m
q0(v)m ) + Xv,v′∈U
q0(u − v)m
q0(v)m ) + Xv∈U, ξ∈Γ∗
q1(l)m
q0(v)m )+
q2(τ )m
q0(v)m ) + X(l,ε),(l′,ε′)∈J1
(τ,ε),(τ′,ε′)
b2
4
q2(τ )m
q2(τ′)m + X(l,ε)∈J1, (τ,ε′)∈J2
N (u,π)
(l,ε),(τ,ε′)
q1(l)m ) + Xh∈G∗, (τ,ε)∈J2
q1(l)m ) + X(τ,ε)∈J2, ξ∈Γ∗
(
(
a2
2
ab
2
q2(ξ)m
q1(h)m
q1(l)m
q2(ξ)m +
q1(l)m
q1(h)m +
q1(h′)m + Xh∈G∗, ξ∈Γ∗
q0(u − v)m
q0(v)m ) +
q0(u − v)m +
q0(v)m
h,h′ a2 q1(h)m
N (u,π)
N (u,π)
(v,π),(v′,π)
(a + b)2
4
q0(v)m
q0(v′)m
N (u,π)
(v,π),ξ
b(a + b)
2
(
q0(v)m
q2(ξ)m +
q2(ξ)m
q0(v)m )
N (u,π)
(l,ε),(l′,ε′)
a2
4
q1(l)m
q1(l′)m +
q2(τ )m
q1(l)m )
(
(
N (u.π)
h,(τ,ε)
ab
4
q1(l)m
q2(τ )m +
q1(h)m
ab
q2(τ )m +
2
q2(τ )m
q2(ξ)m +
q1(h)m ) + Xξ,ξ′∈Γ∗
q2(ξ)m
N (u,π)
(τ,ε),ξ
b2
2
(
q2(τ )
q1(h)m )
q2(ξ)m
q2(τ )m )
ξ,ξ′ b2 q2(ξ)m
N (u,π)
q2(ξ′)m
q0(v)m
q0(v′)m
(
=
(v,π),(τ,ε)
N (u,π)
N (u,π)
N (u,π)
N (u,π)
N (u,π)
(l,ε),ξ
N (u,π)
(l,ε),h
N (u,π)
(v,π),g
(v,0),(u−v,π)
=Xv∈U
+ Xv∈U, g∈G∗
+ Xv∈U, (l,ε)∈J1
Xv∈U, (τ,ε)∈J2
X(τ,ε),(τ′,ε′)∈J2
+ X(l,ε)∈J1, ξ∈Γ∗
+ X(l,ε)∈J1, h∈G∗
+ Xh,h′∈G∗
4 Xv∈U
a2 − b2
Γ − G Xv∈U, h∈G∗
Γ − G Xv∈U, ξ∈Γ∗
Γ − G Xv∈U, l∈K∗
Γ − G Xv∈U, τ∈Σ∗
+ a2 Xl,l′∈K∗
Γ − G Xl∈K∗, ξ∈Γ∗
Γ − G Xh∈G∗, τ∈Σ∗
Γ − G Xl∈K∗, h∈G∗
Γ − G Xτ∈Σ∗, ξ∈Γ∗
Γ − G
2a(a + b)
2b(a + b)
a(a + b)
b(a + b)
2a2
2ab
2ab
2b2
+
+
+
+
+
+
+
2
(
N (u,π)
h,ξ ab(
2(a + b)2
q1(h)m
q2(ξ)m +
Γ − G Xv,v′∈U
q0(v)m
q1(h)m
q1(h)m +
q0(v)m )
q0(v)m
q2(ξ)m
q2(ξ)m +
q0(v)m )
q0(v)m
q1(l)m
q1(l)m +
q0(v)m )
q0(v)m
q2(τ )m
q2(τ )m +
q0(v)m )
q1(l′)m + b2 Xτ,τ′∈Σ∗
q1(l)m
q1(l)m
q2(ξ)m +
q1(h)m
q2(τ )m +
2
(
Γ − G −
q2(ξ)m
q1(l)m )
q2(τ )
q1(h)m )
q1(l)m
q1(h)m +
q2(τ )m
q1(h)m
q1(l)m )
q2(ξ)m
(1 + hu0, hi)(
(1 + hu0, ξi)(
(1 + hu0, li)(
(1 + hu0, τi)(
+
δu+l+l′,0
2
)
(1 + hu0, lihu0, ξi)(
(1 + hu0, hihu0, τi)(
(1 + hu0, l + hi)(
δu+τ +τ′,0
2
)
q2(τ )m
q2(τ′)m
INFINITE FAMILIES OF MODULAR DATA
111
and
a2
2 Xl∈K∗
q1(l)m
q1(u + l)m −
b2
2 Xτ∈Σ∗
q2(τ )m
q2(u + τ )m
q0(u − v)m
q0(v)m ) +
q2(ξ)m
q2(u + ξ)m
(q0(v)m(A(w,−m) + B(w,−m)) + (A(w, m) + B(w, m))q0(v)−m)
(q0(v)m(C(w,−m) + D(w,−m)) + (C(w, m) + D(w, m))q0(v)−m)
(C(w, m)C(w,−m) + D(w, m)D(w,−m))
(
+
+
+
+
=
a + b
a + b
q0(v)m
q1(h)m
νm((u, π))
4 Xv∈U
a2 − b2
q0(u − v)m +
+ a2 Xh∈G∗
q1(u + h)m − b2Xξ∈Γ∗
Γ − G Xw∈UXv∈U
Γ − G Xw∈UXv∈U
Γ − G Xw∈U
Γ − G Xw∈U
Γ − G Xw∈U
Γ − G Xw∈U
2 Xh∈G
Γ − G Xw∈U
q1(u + h)m −
q1(h)m
1
1
1
1
+
+
a2
=
=
1
(A(w, m)C(w,−m) + C(w, m)A(w,−m) + A(w, m)D(w,−m) + D(w, m)A(w,−m))
(B(w, m)C(w,−m) + C(w, m)B(w,−m) + B(w, m)D(w,−m) + D(w, m)B(w,−m))
(A(w, m)A(w,−m) + A(w, m)B(w,−m) + B(w, m)A(w,−m) + B(w, m)B(w,−m))
q2(ξ)m
q2(u + ξ)m +
1
Γ − G Xw∈U
G(q1, q2, w, m)2
b2
2 Xξ∈Γ
G(q1, q2, w, m)2.
112
PINHAS GROSSMAN AND MASAKI IZUMI
νm((k, ε))
N (k,ε)
=Xv∈U
+ Xv∈U, g∈G∗
+ Xv∈U, (l,δ)∈J1
+ Xv∈U, (τ,δ)∈J2
+ X(τ,δ),(τ′,δ′)∈J2
+ X(l,δ)∈J1, ξ∈Γ∗
+ X(l,δ)∈J1, h∈G∗
+ Xh,h′∈G∗
(v,0),(k−v,ε)
N (k,ε)
(v,π),g
a(a − b)
4
a(a + b)
2
(
(
a(a + b)
N (k,ε)
(v,π),(l,δ)
N (k,ε)
(v,π),(τ,δ)
b(a + b)
q0(v)m
q1(g)m
q0(v)m ) + Xv,v′∈U
q0(k − v)m
q0(v)m ) + Xv∈U, ξ∈Γ∗
q1(l)m
q0(v)m )
q2(τ )m
q0(k − v)m +
q0(v)m
q1(g)m +
q0(v)m
q1(l)m +
q0(v)m
q2(τ )m +
q0(v)m ) + X(l,δ),(l′,δ′)∈J1
N (k,ε)
(v,π),ξ
(
(
4
4
N (k,ε)
(v,π),(v′,π)
(a + b)2
4
q0(v)m
q0(v′)m
b(a + b)
2
(
q0(v)m
q2(ξ)m +
q2(ξ)m
q0(v)m )
N (k,ε)
(l,δ),(l′,δ′)
a2
4
q1(l)m
q1(l′)m
N (k,ε)
(τ,δ),(τ′,δ′)
b2
4
q2(τ )m
q2(τ′)m + X(l,δ)∈J1, (τ,δ′)∈J2
N (k,ε)
(l,δ),(τ,δ′)
q2(ξ)m
q1(l)m ) + Xh∈G∗, (τ,δ)∈J2
q1(l)m ) + X(τ,δ)∈J2, ξ∈Γ∗
q1(h)m
N (k,ε)
h,ξ ab(
N (k,ε)
(l,δ),ξ
N (k,ε)
(l,δ),h
(
(
a2
2
ab
2
q1(l)m
q2(ξ)m +
q1(l)m
q1(h)m +
q1(h′)m + Xh∈G∗, ξ∈Γ∗
q1(k − v)m +
q0(v)m
h,h′ a2 q1(h)m
N (k,ε)
(
(
ab
4
N (u,π)
h,(τ,δ)
q1(l)m
q2(τ )m +
q1(h)m
ab
q2(τ )m +
2
q2(τ )m
q2(ξ)m +
q1(h)m ) + Xξ,ξ′∈Γ∗
q2(τ )m
q1(l)m )
q2(τ )m
q1(h)m )
q2(ξ)m
q2(τ )m )
ξ,ξ′ b2 q2(ξ)m
N (k,ε)
q2(ξ′)m
N (k,ε)
(τ,δ),ξ
b2
2
(
(1 + hk, u0i)
q0(v)m
q0(v′)m
q2(ξ)m
q1(h)m
q2(ξ)m +
2(Γ − G) Xv,v′∈U
(a + b)2
q1(g)m
q0(v)m )
q2(ξ)m
q0(v)m )
q0(v)m
q1(l)m +
q1(k − v)m
q0(v)m ) +
q0(v)m
q1(g)m +
q0(v)m
q2(ξ)m +
(1 + hk + g, u0i)(
(1 + hk + ξ, u0i)(
2(1 + hu0, k + li)
Γ − G
(1 + hk + l + l′, u0i)
q1(l)m
q1(l′)m
(1 + hk, u0ihτ + τ′, u0i)
q2(τ )m
q2(τ′)m
+ δv−k+l,0)(
q1(l)m
q0(v)m )
(1 + hk + l, u0ihτ, u0i)(
q1(l)m
q2(τ )m +
q2(τ )m
q1(l)m )
(1 + hk + l, u0ihξ, u0i)(
(1 + hk + h, u0ihτ, u0i)(
q1(l)m
q2(ξ)m +
q1(h)m
q2(τ )m +
q2(ξ)m
q1(l)m )
q2(τ )m
q1(h)m )
q1(l)m
q1(h)m
=
+
+
+
+
+
+
+
+
(
(
a2
b(a + b)
a(a + b)
a(a + b)
a(a − b)
4 Xv∈U
Γ − G Xv∈U, g∈G∗
Γ − G Xv∈U, ξ∈Γ∗
4 Xv∈U, l∈K∗
2(Γ − G) Xl,l′∈K∗
2(Γ − G) Xτ,τ′∈Σ∗
2(Γ − G) Xl∈K∗,τ∈Σ∗
Γ − G Xl∈K∗, ξ∈Γ∗
Γ − G Xh∈G∗, τ∈Σ∗
Γ − G Xl∈K∗, h∈G∗
a2
ab
ab
ab
b2
INFINITE FAMILIES OF MODULAR DATA
113
and
νm((k, ε))
q0(v)m ) + a2 Xh∈G∗
q1(k − v)m
q1(h)m
q1(k + h)m
(1 + hk, u0i)
q0(v)m
q0(v′)m
hk, wi((A(w, m) + B(w, m) + C(w, m) + D(w, m))q0(w)−m)
hk, wi((A(w,−m) + B(w,−m) + C(w,−m) + D(w,−m))q0(w)m)
=
+
+
+
+
+
=
+
=
hk, wi
hk, wi
(
1
1
1
a2
q0(v)m
(a + b)2
2 Xv∈U
q1(k − v)m +
2(Γ − G) Xv,v′∈U
2(Γ − G) Xw∈U
2(Γ − G) Xw∈U
2(Γ − G) Xw∈U
2(Γ − G) Xw∈U
2(Γ − G) Xw∈U
2 Xv∈U
q0(0)m
q1(k)m +
2(Γ − G) Xw∈U
2 Xh∈G
2(Γ − G) Xw∈U
2(Γ − G) Xw∈U
q1(k + h)m
q1(h)m
a2
1
1
1
hk, wi
(A(w, m)C(w,−m) + C(w, m)A(w,−m) + A(w, b)D(w,−m) + D(w, m)A(w,−m))
+
1
(B(w, m)C(w,−m) + C(w, m)B(w,−m) + B(w, b)D(w,−m) + D(w, m)B(w,−m))
+
1
(A(w, m)A(w,−m) + A(w, m)B(w,−m) + B(w, b)B(w,−m) + B(w, m)B(w,−m))
=
q0(u0)m
q1(h)m
a2
a2
(
q1(k − u0)m +
q1(k − u0)m
q0(u0)m ) +
2 Xh∈G\G2
q1(k)m
q0(0)m +
q1(k + h)m
hk, wiG(q1, q2, w, m)2
hk, wiG(q1, q2, w, m)2
hk, wiG(q1, q2, w, m)2 +
δmk,0q1(k)m
2
.
114
PINHAS GROSSMAN AND MASAKI IZUMI
νm(g)
a(a − b)
2
(
(v,0),g−v
N g
(v,π),h
a(a + b)
2
a(a + b)
N g
(v,π),(l,δ)
b(a + b)
N g
(v,π),(τ,δ)
q0(v)m
q1(h)m
q0(v)m ) + Xv,v′∈U
q1(g − v)m
q0(v)m ) + Xv∈U, ξ∈Γ∗
q1(l)m
q0(v)m )
q2(τ )m
(
q1(g − v)m +
q0(v)m
q1(h)m +
q0(v)m
q1(l)m +
q0(v)m
q2(τ )m +
4
4
(
(
q0(v)m ) + X(l,δ),(l′,δ′)∈J1
N g
(v,π),ξ
N g
(v,π),(v′ ,π)
(a + b)2
4
q0(v)m
q0(v′)m
b(a + b)
2
(
q0(v)m
q2(ξ)m +
q2(ξ)m
q0(v)m )
N g
(l,δ),(l′,δ′)
a2
4
q1(l)m
q1(l′)m
N g
=Xv∈U
+ Xv∈U, h∈G∗
+ Xv∈U, (l,δ)∈J1
+ Xv∈U, (τ,δ)∈J2
+ X(τ,δ),(τ′,δ′)∈J2
+ X(l,δ)∈J1, ξ∈Γ∗
+ X(l,δ)∈J1, h∈G∗
+ Xh,h′∈G∗
N g
(τ,δ),(τ′,δ′)
b2
4
q2(τ )m
q2(τ′)m + X(l,δ)∈J1, (τ,δ)∈J2
N g
(l,δ),(τ,δ)
ab
4
(
N g
N g
(l,δ),ξ
(l,δ),h
(
a2
2
ab
2
q1(l)m
q2(ξ)m +
q1(l)m
q1(h)m +
q1(h′)m + Xh∈G∗, ξ∈Γ∗
(
q2(ξ)m
q1(l)m ) + Xh∈G∗, (τ,δ)∈J2
q1(l)m ) + X(τ,δ)∈J2, ξ∈Γ∗
q1(h)m
N g
h,ξab(
q1(h)m
q2(ξ)m +
h,h′a2 q1(h)m
N g
N g
h,(τ,δ)
N g
(τ,δ),ξ
(
q1(l)m
q2(τ )m +
q1(h)m
ab
q2(τ )m +
2
q2(τ )m
q2(ξ)m +
q2(τ )m
q1(l)m )
q2(τ )m
q1(h)m )
q2(ξ)m
q2(τ )m )
ξ,ξ′b2 q2(ξ)m
N g
q2(ξ′)m
b2
2
(
q2(ξ)m
q1(h)m ) + Xξ,ξ′∈Γ∗
INFINITE FAMILIES OF MODULAR DATA
115
q0(v)m
(
q1(g − v)m +
q1(g − v)m
q0(v)m ) +
(a + b)2
Γ − G Xv,v′∈U
4(1 + hu0, g + hi)
(
Γ − G
+ δv−g+h,0 + δv−g+θ(h),0)(
(1 + hu0, gi)
q0(v)m
q1(h)m +
q0(v)m
q0(v′)m
q1(h)m
q0(v)m )
(1 + hu0, gihu0, ξi)(
q0(v)m
q2(ξ)m +
q2(ξ)m
q0(v)m )
(1 + hu0, l + gi)(
(1 + hu0, gihu0, τi)(
q0(v)m
q1(l)m +
q0(v)m
q2(τ )m +
q1(l)m
q0(v)m )
q2(τ )m
q0(v)m )
(1 + hu0, gi)
(1 + hu0, gi)
q1(l)m
q1(l′)m
q2(τ )m
q2(τ′)m
(1 + hu0, l + gihu0, τi)(
(1 + hu0, l + gihu0, ξi)(
(1 + hu0, g + hihu0, τi)(
q1(l)m
q2(τ )m +
q1(l)m
q2(ξ)m +
q1(h)m
q2(τ )m +
q2(τ )m
q1(l)m )
q2(ξ)m
q1(l)m )
q2(τ )m
q1(h)m )
a2
b2
ab
=
+
+
+
+
+
+
+
b(a + b)
a(a + b)
a(a + b)
2b(a + b)
a(a − b)
2 Xv∈U
2 Xv∈U, h∈G∗
Γ − G Xv∈U, ξ∈Γ∗
Γ − G Xv∈U, l∈K∗
Γ − G Xv∈U, τ∈Σ∗
Γ − G Xl,l′∈K∗
Γ − G Xτ,τ′∈Σ∗
Γ − G Xl∈K∗, τ∈Σ∗
Γ − G Xl∈K∗, ξ∈Γ∗
Γ − G Xh∈G∗, τ∈Σ
+ a2 Xl∈K∗, h∈G∗
Γ − G Xτ∈Σ∗, ξ∈Γ∗
+ a2 Xh,h′∈G∗
Γ − G Xh∈G∗, ξ∈Γ∗
Γ − G Xξ,ξ′∈Γ∗
2ab
2ab
4ab
2b2
+
+
4b2
+
+
+
2(1 + hu0, k + g + hi)
(
Γ − G
+ δl−g+h,0 + δl−g+θ(h),0)(
q1(l)m
q1(h)m +
q1(h)m
q1(l)m )
(1 + hu0, gihu0, τ + ξi)(
q2(τ )m
q2(ξ)m +
q2(ξ)m
q2(τ )m )
4(1 + hu0, g + h + h′i)
(
Γ − G
) + δ−g+h+h′,0 + δ−θ(g)+h+h′,0 + δ−g+θ(h)+h′,0 + δ−g+h+θ(h′))
q1(h)m
q1(h′)m
(1 + hu0, g + hihu0, ξi)(
q1(h)m
q2(ξ)m +
q2(ξ)m
q1(h)m )
(1 + hu0, gihu0, ξ + ξ′i)
q2(ξ)m
q2(ξ′)m ,
PINHAS GROSSMAN AND MASAKI IZUMI
116
and
νm(g)
(
q0(v)m
q0(v)m ) + a2 Xl∈K∗
q1(g − v)m
q1(g − v)m +
(δ−g+h+h′,0 + δ−θ(g)+h+h′,0 + δ−g+θ(h)+h′,0 + δ−g+h+θ(h′),0)
q1(g − l)m +
q1(l)m
q1(g − l)m
q1(l)m )
(
q1(h)m
q1(h′)m
+
+
a2
b2
ab
+
+
+
+
b(a + b)
a(a + b)
(a + b)2
2b(a + b)
2a(a + b)
= a2Xv∈U
+ a2 Xh,h′∈G∗
Γ − G Xv,v′∈U
Γ − G Xv∈U, h∈G∗
Γ − G Xv∈U, ξ∈Γ∗
Γ − G Xv∈U, l∈K∗
Γ − G Xv∈U, τ∈Σ∗
Γ − G Xl,l′∈K∗
Γ − G Xτ,τ′∈Σ∗
Γ − G Xl∈K∗, τ∈Σ∗
Γ − G Xl∈K∗, ξ∈Γ∗
Γ − G Xh∈G∗, τ∈Σ
Γ − G Xl∈K∗, h∈G∗
Γ − G Xτ∈Σ∗, ξ∈Γ∗
Γ − G Xh,h′∈G∗
Γ − G Xh∈G∗, ξ∈Γ∗
Γ − G Xξ,ξ′∈Γ∗
4a2
2a2
2ab
4ab
+
+
+
+
+
+
+
+
2ab
4b2
2b2
+
(1 + hu0, gi)
q0(v)m
q0(v′)m
(1 + hu0, g + hi)(
(1 + hu0, gihu0, ξi)(
(1 + hu0, l + gi)(
q0(v)m
q1(h)m +
q0(v)m
q2(ξ)m +
q1(h)m
q0(v)m )
q2(ξ)m
q0(v)m )
q0(v)m
q1(l)m +
q0(v)m
q2(τ )m +
q1(l)m
q0(v)m )
q2(τ )m
q0(v)m )
(1 + hu0, gihu0, τi)(
(1 + hu0, gi)
(1 + hu0, gi)
q1(l)m
q1(l′)m
q2(τ )m
q2(τ′)m
(1 + hu0, l + gihu0, τi)(
(1 + hu0, l + gihu0, ξi)(
(1 + hu0, g + hihu0, τi)(
(1 + hu0, k + g + hi)(
q1(l)m
q2(τ )m +
q1(l)m
q2(ξ)m +
q1(h)m
q2(τ )m +
q1(l)m
q1(h)m +
q2(τ )m
q2(ξ)m +
q2(τ )m
q1(l)m )
q2(ξ)m
q1(l)m )
q2(τ )m
q1(h)m )
q1(h)m
q1(l)m )
q2(ξ)m
q2(τ )m )
(1 + hu0, gihu0, τ + ξi)(
(1 + hu0, g + h + h′i)
q1(h)m
q1(h′)m
(1 + hu0, g + hihu0, ξi)(
q1(h)m
q2(ξ)m +
q2(ξ)m
q1(h)m )
(1 + hu0, gihu0, ξ + ξ′i)
q2(ξ)m
q2(ξ′)m
INFINITE FAMILIES OF MODULAR DATA
117
(
= a2Xl∈K
+ a2 Xh,h′∈G∗
q1(l)m
q1(g − l)m
q1(l)m ) +
q1(g − l)m +
(δ−g+h+h′,0 + δ−θ(g)+h+h′,0 + δ−g+θ(h)+h′,0 + δ−g+h+θ(h′),0)
hg, wiG(q1, q2, w, m)2
q1(h)m
q1(h′)m
1
Γ − G Xw∈U
1
Γ − G Xw∈U
hg, wiG(q1, q2, w, m)2.
= δmg,0q1(g)m +
νm((σ, ε))
(v,0),(σ−v,ε)
N (σ,ε)
(v,π),h
b(a − b)
4
a(a + b)
2
(
(
N (σ,ε)
(v,π),(l,δ)
a(a + b)
N (σ,ε)
(v,π),(τ,δ)
b(a + b)
q0(v)m
q2(σ − v)m +
q0(v)m
q1(h)m +
q0(v)m
q1(l)m +
q0(v)m
q2(τ )m +
q1(h)m
q0(v)m ) + Xv,v′∈U
q2(σ − v)m
q0(v)m ) + Xv∈U, ξ∈Γ∗
q1(l)m
q0(v)m )
q2(τ )m
q0(v)m ) + X(l,δ),(l′,δ′)∈J1
N (σ,ε)
(v,π),ξ
(
(
4
4
N (σ,ε)
(v,π),(v′,π)
(a + b)2
4
q0(v)m
q0(v′)m
b(a + b)
2
(
q0(v)m
q2(ξ)m +
q2(ξ)m
q0(v)m )
N (σ,ε)
(l,δ),(l′,δ′)
a2
4
q1(l)m
q1(l′)m
N (σ,ε)
=Xv∈U
+ Xv∈U, h∈G∗
+ Xv∈U, (l,δ)∈J1
+ Xv∈U, (τ,δ)∈J2
+ X(τ,δ),(τ′,δ′)∈J2
+ X(l,δ)∈J1, ξ∈Γ∗
+ X(l,δ)∈J1, h∈G∗
+ Xh,h′∈G∗
N (σ,ε)
(l,δ),(τ,δ)
N (σ,ε)
(τ,δ),(τ′,δ′)
b2
4
q2(τ )m
q2(τ′)m + X(l,δ)∈J1, (τ,δ)∈J2
N (σ,ε)
(l,δ),ξ
N (σ,ε)
(l,δ),h
(
a2
2
ab
2
q1(l)m
q2(ξ)m +
q1(l)m
q1(h)m +
q1(h′)m + Xh∈G∗, ξ∈Γ∗
(
q2(ξ)m
q1(l)m ) + Xh∈G∗, (τ,δ)∈J2
q1(l)m ) + X(τ,δ)∈J2, ξ∈Γ∗
q1(h)m
N (σ,ε)
h,ξ ab(
q1(h)m
q2(ξ)m +
(
(
ab
4
N (σ,ε)
h,(τ,δ)
q1(l)m
q2(τ )m +
q1(h)m
ab
q2(τ )m +
2
q2(τ )m
q2(ξ)m +
q1(h)m ) + Xξ,ξ′∈Γ∗
q2(τ )m
q1(l)m )
q2(τ )m
q1(h)m )
q2(ξ)m
q2(τ )m )
ξ,ξ′ b2 q2(ξ)m
N (σ,ε)
q2(ξ′)m
N (σ,ε)
(τ,δ),ξ
q2(ξ)m
b2
2
(
h,h′ a2 q1(h)m
N (σ,ε)
118
PINHAS GROSSMAN AND MASAKI IZUMI
q0(v)m
(
q2(σ − v)m +
(1 + hσ, u0i)
q2(σ − v)m
q0(v)m )
q0(v)m
q0(v′)m
=
+
b2
ab
+
+
+
+
+
a2
b(a + b)
b(a + b)
a(a + b)
(a + b)2
b(a − b)
4 Xv∈U
2(Γ − G) Xv,v′∈U
Γ − G Xv∈U, h∈G∗
Γ − G Xv∈U, ξ∈Γ∗
4 Xv∈U, τ∈Σ∗
2(Γ − G) Xl,l′∈K∗
2(Γ − G) Xτ,τ′∈Σ∗
2(Γ − G) Xl∈K∗, τ∈Σ∗
Γ − G Xl,∈K∗, ξ∈Γ∗
Γ − G Xh∈G∗, τ∈Σ∗
Γ − G Xl∈K∗, h∈G∗
Γ − G Xτ∈Σ∗, ξ∈Γ∗
Γ − G Xh,h′∈G∗
Γ − G Xh∈G∗, ξ∈Γ∗
+ b2 Xξ,ξ′∈Γ∗
+
+
2a2
+
+
+
2ab
a2
+
+
ab
ab
b2
2
(1 + hh, u0ihσ, u0i)(
q0(v)m
q1(h)m +
q1(h)m
q0(v)m )
q2(ξ)m
q0(v)m )
q0(v)m
q2(ξ)m +
q0(v)m
q2(τ )m +
q2(τ )m
q0(v)m )
(1 + hσ + ξ, u0i)(
4
(
Γ − G − δv−σ+τ )(
(1 + hu0, σi)
q1(l)m
q1(l′)m
q2(τ )m
q2(τ′)m
(1 + hu0, σ + τ + τ′i)
q1(l)m
q2(τ )m +
(1 + hu0, li)(
q2(τ )m
q1(l)m )
q2(ξ)m
q1(l)m )
q1(h)m
q1(l)m )
(1 + hu0, lihu0, σ + ξi)(
(1 + hu0, hi)(
q1(h)m
q2(τ )m +
(1 + hu0, l + hihu0, σi)(
(1 + hu0, γi)(
q2(τ )m
q2(ξ)m +
q1(l)m
q2(ξ)m +
q2(τ )m
q1(h)m )
q1(l)m
q1(h)m +
q2(ξ)m
q2(τ )m )
q1(h)m
q1(h′)m
(1 + hu0, h + h′ihu0, σi)
(1 + hu0, hihu0, σ + ξi)(
(
Γ − G − δ−σ+ξ+ξ′,0 − δ−σ+ξ+θ(ξ),0)
q2(ξ)m
q1(h)m )
q1(h)m
q2(ξ)m +
q2(ξ)m
q2(ξ′)m .
INFINITE FAMILIES OF MODULAR DATA
119
N γ
(v,0),γ−v
b(a − b)
2
(
N γ
(v,π),h
a(a + b)
2
(
a(a + b)
N γ
(v,π),(l,δ)
q0(v)m
q2(γ − v)m +
(v,π),ξ
N γ
q1(h)m
q0(v)m ) + Xv,v′∈U
q2(γ − v)m
q0(v)m ) + Xv∈U, ξ∈Γ∗
q1(l)m
q0(v)m )
q2(τ )m
q0(v)m ) + X(l,δ),(l′,δ′)∈J1
q0(v)m
q1(h)m +
q0(v)m
q1(l)m +
q0(v)m
q2(τ )m +
(
(
4
4
b(a + b)
N γ
(v,π),(τ,δ)
N γ
(v,π),(v′ ,π)
(a + b)2
4
q0(v)m
q0(v′)m
b(a + b)
2
(
q0(v)m
q2(ξ)m +
q2(ξ)m
q0(v)m )
νm(γ) =Xv∈U
+ Xv∈U, h∈G∗
+ Xv∈U, (l,δ)∈J1
+ Xv∈U, (τ,δ)∈J2
+ X(τ,δ),(τ′,δ′)∈J2
+ X(l,δ)∈J1, ξ∈Γ∗
+ X(l,δ)∈J1, h∈G∗
+ Xh,h′∈G∗
N γ
(τ,δ),(τ′,δ′)
b2
4
q2(τ )m
q2(τ′)m + X(l,δ)∈J1, (τ,δ)∈J2
N γ
(l,δ),(τ,δ)
ab
4
(
N γ
N γ
(l,δ),ξ
(l,δ),h
(
a2
2
ab
2
q1(l)m
q2(ξ)m +
q1(l)m
q1(h)m +
q1(h′)m + Xh∈G∗, ξ∈Γ∗
(
q2(ξ)m
q1(l)m ) + Xh∈G∗, (τ,δ)∈J2
q1(l)m ) + X(τ,δ)∈J2, ξ∈Γ∗
q1(h)m
N γ
h,ξab(
q1(h)m
q2(ξ)m +
h,h′a2 q1(h)m
N γ
N γ
h,(τ,δ)
N γ
(τ,δ),γ
q2(ξ)m
q1(h)m ) + Xξ,ξ′∈Γ∗
N γ
(l,δ),(l′,δ′)
a2
4
q1(l)m
q1(l′)m
(
q1(l)m
q2(τ )m +
q1(h)m
ab
q2(τ )m +
2
q2(τ )m
q2(ξ)m +
q2(τ )m
q1(l)m )
q2(τ )m
q1(h)m )
q2(ξ)m
q2(τ )m )
ξ,ξ′b2 q2(ξ)m
N γ
q2(ξ′)m
b2
2
(
120
PINHAS GROSSMAN AND MASAKI IZUMI
q0(v)m
(
q2(γ − v)m +
q2(γ − v)m
q0(v)m ) +
(a + b)2
Γ − G Xv,v′∈U
q1(h)m
q0(v)m )
q0(v)m
q1(h)m +
(1 + hu0, γi)
q0(v)m
q0(v′)m
=
a2
ab
+
+
+
+
+
+
b(a + b)
b(a + b)
a(a + b)
2a(a + b)
b(a − b)
2 Xv∈U
Γ − G Xv∈U, h∈G∗
2 Xv∈U, ξ∈Γ∗
Γ − G Xv∈U,l∈K∗
Γ − G Xv∈U,τ∈Σ∗
Γ − G Xl,l′∈K∗
Γ − G Xl∈K∗,τ∈Σ∗
Γ − G Xl∈K∗,ξ∈Γ∗
Γ − G Xh∈G∗,τ∈Σ∗
Γ − G Xl∈K∗,h∈G∗
+ b2 Xτ∈Σ∗,ξ∈Γ∗
Γ − G Xh,h′∈G∗
Γ − G Xh∈G∗, ξ∈Γ∗
+ b2 Xξ,ξ′∈Γ∗
2a2
4a2
2ab
2ab
4ab
+
+
+
+
+
q0(v)m
q2(ξ)m +
q2(ξ)m
q0(v)m )
q2(τ )m
q2(τ′)m
(1 + hu0, hihu0, γi)(
4(1 + hu0, γ + ξi)
(
Γ − G
(1 + hu0, lihu0, γi)(
(1 + hu0, τ + γi)(
q1(l)m
q1(l′)m +
(1 + hu0, γi)
(1 + hu0, lihu0, τ + γi)(
(1 + hu0, lihu0, γ + ξi)(
(1 + hu0, hihu0, γ + τi)(
(1 + hu0, l + hihu0, γi)(
− δv−γ+ξ,0 − δv−γ+θ(ξ))(
q1(l)m
q0(v)m
q1(l)m +
q0(v)m )
q0(v)m
q2(τ )m
q2(τ )m +
q0(v)m )
Γ − G Xτ,τ′∈Σ∗
b2
q1(l)m
q2(τ )m +
q1(l)m
q2(ξ)m +
q1(h)m
q2(τ )m +
q1(l)m
q1(h)m +
(1 + hu0, γi)
q2(τ )m
q1(l)m )
q2(ξ)m
q1(l)m )
q2(τ )m
q1(h)m )
q1(h)m
q1(l)m )
2(1 + hu0, γ + τ + ξi)
(
Γ − G
− δ−γ+τ +ξ,0 − δ−γ+τ +θ(ξ),0)(
q2(τ )m
q2(ξ)m +
q2(ξ)m
q2(τ )m )
(1 + hu0, h + h′ihu0, γi)
q1(h)m
q1(h′)m
(1 + hu0, hihu0, γ + ξi)(
q1(h)m
q2(ξ)m +
q2(ξ)m
q1(h)m )
4(1 + hu0, γ + ξ + ξ′i)
(
Γ − G
− δ−γ+ξ+ξ′ − δ−θ(γ)+ξ+ξ′ − δ−γ+θ(ξ)+ξ′ − δγ+ξ+θ(ξ′))
q2(ξ)m
q2(ξ′)m .
The rest of computation is the same as in the previous cases.
(cid:3)
INFINITE FAMILIES OF MODULAR DATA
121
Appendix D. Calculations for Section 6
Proof of Lemma 6.2. We first check the unitarity of S.
Sπ,x2 =
a2 + b2
2
+
a2
2
(K − 1) + a2G − K
2
+ b2Γ − 1
2
= 1.
S0,x2 =Xx
Xx
S0,xSπ,x =
Xx
Xx
a2 − b2
2
+
a2
2
(K − 1) + a2G − K
2
− b2Γ − 1
2
= 0.
Xx
=
S0,xS(k,ε),x =Xx
2 Xl∈K\{0}
a2
a2
2
+
Sπ,xS(k,ε),x
hk, li1 + a2 Xg∈G∗
hk, gi1 =
a2Gδk,0
2
= 0.
Sπ,xSg,x
Xx
S0,xSg,x =Xx
= a2 + a2Xl6=0
hk, li1 + a2 Xh∈G∗
(hg, hi1 + hg, hi1) = a2Gδg,0 = 0.
Sπ,xSγ,x = −b2 − b2 Xγ∈Γ∗
(hγ, ξi2 + hγ, ξi2) = −b2Γδγ,0 = 0.
S0,xSγ,x = −Xx
Xx
S(k,ε),xS(k′,ε′),x
(a2hk − k′, li + εε′q1(k)q1(k′)δk,lδk′,l) + a2 Xg∈G∗
hk − k′, gi
1
+
4 X(l,t)∈J2
a2
2
a2Gδk,k′
+
=
=
2
εε′δk,k′
2
= δk,k′δε,ε′.
S(k,ε),xSg,x
Xx
= a2 + a2 Xl∈K\{0}
= a2Gδk,g = 0.
hk − g, li1 + a2 Xh∈G∗
(hk + g, hi1 + h−k + g, hi1)
122
PINHAS GROSSMAN AND MASAKI IZUMI
Px S(k,ε),xSγ,x = 0 is obvious.
hg − g′, li1
Sg,xSg′,x = 2a2 + 2a2 Xl∈K\{0}
Xx
+ a2 Xh∈G∗
= a2(Xl∈K
+ Xh∈G∗
(hg, hi1 + hg, hi1)(hg′, hi1 + hg′, hi1)
hg − g′, li1 +Xl∈K
(hg + g′, hi1 + hg + g′, hi1 + hg − g′, hi1 + hg − g′, hi1)
hg + g′, li1
= δg+g′,0 + δg,g′ = δg,g′.
Px Sg,xSγ,x = 0 is obvious.
Sγ,xSγ′,x
Xx
= 2b2 + b2Xξ∈Γ∗
= δγ+γ′,0 + δγ,γ′ = δγ,γ′.
(hγ, ξi2 + hγ, ξi2)(hγ′, ξi2 + hγ′, ξi2)
Thus S is unitary. It is obvious that T and C are unitary.
Since Tx,x = Tx,x, we have CT = T C. We claim Sx,y = Sx,y = Sx,y, which implies
S2 = C. Indeed, the only block of S that may not be real is the J1-J1 block, and for
other blocks, we have
Sx,y = Sx,y = Sx,y = Sx,y.
Using c2, q1(k)2 = hk, ki1 ∈ {1,−1}, we get
ahk, k′i1 + cεε′q1(k)δk,k′
S(k,ε),(k′,ε′) =
2
=
ahk, k′i1 + c2hk, ki1cεε′q1(k)δk,k′
2
,
and we get
S(k,ε),(k′,ε′) = S(k,c2hk,ki1ε),(k′,ε′) = S(k,ε),(k′,c2hk′,k′i1ε′).
The claim is shown.
The only remaining condition is (ST )3 = cC, or equivalently ST S = cC(T ST )∗.
Since S is symmetric and C commutes with S and T , we always have
Thus it suffices to verify
C(T ST )∗ = CT ST = T ST .
Xx
Sj,xSj′,xTx,x = cSj,j′Tj,jTj′,j′
INFINITE FAMILIES OF MODULAR DATA
123
for all j, j′ ∈ J.
Xx
S2
0,xTx,x
=
=
a2 + b2
2
+
aG(q1)
2
+
a2
2 Xk∈K\{0}
q1(k) + a2 Xg∈G∗
q1(g) + b2 Xγ∈Γ∗
q2(γ)
bG(q2)
2
= c
a − b
2
= cS0,0T 2
0,0.
S0,xSπ,xTx,x
Xx
a2 − b2
=
2
+
a2
2 Xk∈K\{0}
q1(k) + a2 Xg∈G∗
q1(g) − b2 Xγ∈Γ∗
q2(γ)
=
aG(q1)
2 −
bG(q2)
2
= c
a + b
2
= cS0,πT0,0Tπ,π.
Xx
S0,xS(k,ε),xTx,x = Sπ,xS(k,ε),xTx,x
=
=
a2
2
+
a2
2 Xl∈K\{0}
hk, liq1(l) + a2 Xh∈G∗
hk, hi1q1(h) =
a2
2 Xh∈G
q1(k + h)q1(k)
aG(q1)q1(k)
2
= c
a
2
q1(k)
= cS0,(k,ε)T0,0T(k,ε),(k,ε) = cSπ,(k,ε)Tπ,πT(k,ε),(k,ε).
S0,xSg,xTx,x = Sπ,xSg,xTx,x
Xx
= a2 + a2 Xl∈K\{0}
= a2Xh∈G
hg, liq1(l) + a2 Xh∈G∗
(hg, hi1 + hg, hi1)q1(h)
q1(g + h)q1(g) = aG(q1)q1(g)
= cS0,gT0,0Tg,g = cSπ,gTπ,πTg,g.
S0,xSγ,xTx,x = −Xx
Xx
= −b2 − b2Xξ∈Γ∗
(hγ, ξi2 + hγ, ξi2)q2(ξ) = −b2Xξ∈Γ
Sπ,xSγ,xTx,x
= cS0,γT0,0Tγ,γ = −cSπ,γT0,0Tγ,γ.
hξ, γi2q2(ξ) = −bG(q2)q2(γ)
PINHAS GROSSMAN AND MASAKI IZUMI
124
Xx
S(k,ε),xS(k′,ε′),xTx,x
a2
2
+
1
2 Xl∈K\{0}
(a2hk + k′, li1 + c2εε′q1(k)q1(k′)δk,lδk′,l)q1(l) + a2 Xh∈G∗
hk + k′, hi1q1(h)
=
= c
c2εε′q1(k)3δk,k′
aG(q1)q1(k + k′)
ahk, k′i1q1(k)q1(k′) + cεε′q1(k)3δk,k′
+
2
2
2
= cS(k,ε),(k′,ε′)T(k,ε),(k,ε)T(k′,ε′),(k′,ε′),
= c
ahk, k′i1 + cεε′q1(k)δk,k′
2
q1(k)q1(k′)
where we used hk, k′i1 ∈ R and q1(k)4 = hk, ki2
1 = 1.
S(k,ε),xSg,xTx,x
Xx
= a2 + a2 Xl∈K\{0}
= a2Xh∈G
hk + g, li1q1(l) + a2 Xh∈G∗
hk, hi1(hg, hi + hg, hi)q1(h)
q1(k + g + h)q1(k + g) = aG(q1)q1(k + g)
= cahk, gi1q1(k)q1(g)
= cS(k,ε),gT(k,ε),(k,ε)Tg,g.
hg + g′, li1q1(l)
Xx
Sg,xSg′,xTx,x = 2a2 + 2a2 Xl∈K\{0}
+ a2 Xh∈G∗
= a2Xh∈G
= aG(q1)(q1(g + g′) + q1(g − g′) = ca(hg, g′i1 + hg, g′i1)q1(g)q1(g′)
= cSg,g′Tg,gTg′,g′.
(hg, hi + hg, hi)(hg′, hi + hg′, hi)q1(h)
(q1(g + g′ + h)q1(g + g′) + q1(g − g′ + h)q1(g − g′))
Sγ,xSγ′,xTx,x
Xx
= 2b2 + b2Xξ∈Γ∗
= bG(q2)(q2(γ + γ′) + q2(γ − γ′)) = −cb(hγ, ξi2 + hγ, ξi2)q2(γ)q2(γ′)
= cSγ,γ′Tγ,γTγ′,γ′.
(hγ, ξi2 + hγ, ξi2)(hγ′, ξi2 + hγ′, ξi2)q2(ξ)
INFINITE FAMILIES OF MODULAR DATA
125
Xx
S(k,ε),xSγ,xTx,x =Xx
Sg,xSγ,xTx,x
= S(k,ε),γT(k,ε),(k,ε)Tγ,γ = Sg,γTg,gTγ,γ = 0.
(cid:3)
Proof of Lemma 6.3. Since S0,x = Sπ,x for x 6= 0, π, γ, we have
Nπ,y,z − N0,y,z =Xx
(Sπ,0 − S0,0)Sy,0Sz,0
=
S0,0
(Sπ,x − S0,x)Sy,xSz,x
S0,x
= b(
Sy,0Sz,0
S0,0 −
Sy,πSz,π
S0,π
) − 2Xγ∈Γ∗
Sy,γSz,γ.
+
(Sπ,π − S0,π)Sy,πSz,π
S0,π
+ Xγ∈Γ∗
(Sπ,γ − S0,γ)Sy,γSz,γ
S0,γ
Thus
Nπ,π,π = 1 + b(
S2
0,π
S0,0 −
S2
0,0
S0,π
= b2 S2
0,0 + S0,0S0,π + S2
0,π
Nπ,π,γ = b2(
S0,0S0,π
Sπ,0
S0,0
+
Sπ,π
S0,π
4a2b2
a2 − b2 =
a2 − b2 =
4a2b2
4
Γ − G
4
Γ − G
,
.
=
S0,0S0,π
) − b2(Γ − 1)
+ b2 = b2 (S0,0 + S0,π)2
) − 2b2Xξ∈Γ∗
1
S0,0 −
1
S0,π
(hγ, ξi + hγ, ξi) =
) − 2Xξ∈Γ∗
b4
Nπ,γ,γ′ = δγ,γ′ + b3(
Sγ,ξSγ′,ξ
= δγ,γ′ +
S0,0S0,π − 2(δγ,γ′ − Sγ,0Sγ′,0 − Sγ,πSγ′,π)
4a2b2
=
a2 − b2 − δγ,γ′ =
4
Γ − G − δγ,γ′.
For z 6= 0, π, γ, we get
Sπ,0
Nπ,π,z = b(
S0,0 −
S0,0
S0,π
)Sz,0 = b2 S0,0 + S0,π
S0,0S0,π
Sz,0 =
4ab2
a2 − b2 Sz,0,
Nπ,γ,z = Nγ,γ′,z = b2(
1
S0,0
+
1
S0,π
)Sz,0 =
and in particular, we get
4ab2
a2 − b2 Sz,0,
Nπ,π,(k,ε) = Nπ,γ,(k,ε) = Nγ,γ′,(k,ε) =
Nπ,π,g = Nπ,γ,g = Nγ,γ′,g =
2a2b2
a2 − b2 =
2
,
Γ − G
4
.
4a2b2
a2 − b2 =
Γ − G
126
PINHAS GROSSMAN AND MASAKI IZUMI
For y, z 6= 0, π, γ,
Nπ,y,z = δy,z + b(
1
S0,0 −
1
S0,π
)Sy,0Sz,0 = δy,z +
4b2
b2 − a2 Sy,0Sz,0,
Nγ,y,z = b(
1
S0,0 −
1
S0,π
)Sy,0Sz,0 =
4b2
b2 − a2 Sy,0Sz,0,
and in particular,
Nπ,(k,ε),(k′,ε′) = δ(k,ε),(k′,ε′) +
1
Γ − G
, Nγ,(k,ε),(k′,ε′) =
1
Γ − G
,
Nπ,(k,ε),g = Nγ,(k,ε),g =
2
Γ − G
,
Nπ,g,g′ = δg,g′ +
4
Γ − G
, Nγ,g,g′ =
4
Γ − G
.
Nγ,γ′,γ′′ = b3(
1
S0,0 −
1
S0,π
)
=
4b4
(hγ, ξi2 + hγ, ξi2)(hγ′, ξi2 + hγ′, ξi2)(hγ′′, ξi2 + hγ′′, ξi2)
− b2Xξ∈Γ∗
a2 − b2 − b2(Γ(δγ+γ′+γ′′,0 + δγ+γ′,γ′′ + δγ′+γ′′,γ + δγ′′+γ,γ′) − 4)
a2 − b2 − (δγ+γ′+γ′′,0 + δγ+γ′,γ′′ + δγ′+γ′′,γ + δγ′′+γ,γ′)
Γ − G − (δγ+γ′+γ′′,0 + δγ+γ′,γ′′ + δγ′+γ′′,γ + δγ′′+γ,γ′).
4a2b2
4
=
=
For y, z, w 6= 0, π, γ, we have
4a
a2 − b2 Sy,0Sz,0Sw,0
Sy,(l,t)Sz,(l,t)Sw,(l,t) +
Ny,z,w =
+
2
a X(l,t)∈J1
1
a Xh∈G∗
Sy,hSz,hSw,h,
INFINITE FAMILIES OF MODULAR DATA
127
and
N(k,ε),(k′,ε′),(k′′,ε′′) =
a4
2(a2 − b2)
1
+
4a X(l,ε)∈J1
× (ahk′′, li1 + cε′′tq1(k′′)δk′′,l) + a2 Xh∈G∗
hk + k′ + k′′, li1
2(a2 − b2)
2 Xl∈K
a2b2
a2
+
=
(ahk, li1 + cεtq1(k)δk,l)(ahk′, li1 + cε′tq1(k′)δk′,l)
hk + k′ + k′′, hi1
+ c2 εε′hk, k + k′′i1δk,k′ + ε′ε′′hk′, k′ + ki1δk′,k′′ + ε′′εhk′′, k′′ + k′i1δk′′,k
+ a2 Xh∈G∗
2
hk + k′ + k′′, hi
1
=
1
2
(
Γ − G
+ δk+k′+k′′,0)
+ c2 εε′hk, k + k′′i1δk,k′ + ε′ε′′hk′, k′ + ki1δk′,k′′ + ε′′εhk′′, k′′ + k′i1δk′′,k
,
2
(ahk, li1 + cεtq1(k)δk,l)(ahk′, li1 + cε′tq1(k′)δk′,l)hg, li1
hk + k′, hi1(hg, hi1 + hg, hi1)
hk + k′ + g, li1 + a2 Xh∈G∗
(hk + k′ + g, hi1 + hk + k′ + g,−hi1)
a4
a2 − b2
N(k,ε),(k′,ε′),g =
1
+
2 X(l,t)∈J1
+ a2 Xh∈G∗
a2 − b2 + a2Xl∈K
+ c2εε′hk, k + gi1δk,k′
=
a2b2
=
1
Γ − G
+ c2εε′hk, k + gi1δk,k′,
128
PINHAS GROSSMAN AND MASAKI IZUMI
2a4
a2 − b2
(ahk, li1 + cεtq1(k)δk,l)hg + g′, li1
hk, hi1(hg, hi1 + hg, hi1)(hg′, hi1 + hg′, hi1)
N(k,ε),g,g′ =
+ a X(l,t)∈J1
+ a2 Xh∈G∗
a2 − b2 + 2a2Xl∈K
2a2b2
=
2
=
Γ − G
hk + g + g′, li1 + a2 Xh∈G∗
hk, hi1(hg, hi1 + hg, hi1)(hg′, hi1 + hg′, hi1))
+ δk+g+g′,0 + δk+g−g′.
(hg, hi1 + hg, hi1)(hg′, hi1 + hg′, hi1))(hg′, hi1 + hg′, hi1))
hg + g′ + g′′, li1
Ng,g′,g′′ =
4a4
a2 − b2
=
4a2b2
+ 4a2 Xl∈K\{0}
+ a2 Xh∈G∗
a2 − b2 + 4a2 Xl∈K\{0}
+ a2 Xh∈G∗
+ a2 Xh∈G∗
4
=
Γ − G
hg + g′ + g′′, li1 + a2 Xh∈G∗
(hg + g′ − g′′, hi + hg + g′ − g′′, hi) + a2 Xh∈G∗
(h−g + g′ + g′′, hi + h−g + g′ + g′′, hi)
+ δg+g′+g′′,0 + δg+g′,g′′ + δg′+g′′,g + δg′′+g,g′.
(hg + g′ + g′′, hi + hg + g′ + g′′, hi)
(hg − g′ + g′′, hi + hg − g′ + g′′, hi)
(cid:3)
q2(γ)m.
Proof of Lemma 6.8. We set
Am = 2a Xg∈G∗
q1(g)m, Bm = 2bXγ∈Γ∗
Cm = a Xk∈K\{0}
q0(k)m.
Then
G(q1, m) + G(q2, m) = a + b + Am + Bm + Cm.
INFINITE FAMILIES OF MODULAR DATA
129
For νm((k, π)), we have We have
νm(π)
= N π
0,π
a2 − b2
2
+ N π
π,π
(a + b)2
4
2
2
=
+
N π
N π
N π
π,g
(k,ε),(k′,ε′)
a(a + b)
2a(a + b)
g,ha2 q1(g)m
N π
(a + b)2
Γ − G
+ Xg∈G∗
+ X(k,ε),(k′,ε′)∈J1
+ X(k,ε)∈J1, γ∈Γ∗
+ Xg,h∈G∗
a2 − b2
Γ − G Xg∈G∗
4 X(k,ε),(k′,ε′)∈J1
Γ − G Xk∈K\{0}, g∈G∗
+ a2 Xg,h∈G∗
Γ − G Xg∈G∗, γ∈Γ∗
a2 − b2
Γ − G
2a2
4ab
a2
+
+
+
+
+
1
4
(
(
+ X(k,ε)∈J1
N π
π,(k,ε)
a(a + b)
4
(q1(k)m + q1(k)−m)
N π
π,γ
(q1(g)m + q1(g)−m) + Xγ∈Γ∗
q1(k′)m + X(k,ε)∈J1, g∈G∗
q1(k)m
a2
4
b(a + b)
2
(q2(γ)m + q2(γ)−m)
N π
(k,ε),g
a2
2
(
q1(k)m
q1(g)m +
q1(g)m
q1(k)
)
q2(γ)m
q1(k)m )
(
(k,ε),γ
ab
2
q1(k)m
q2(γ)m +
q1(h)m + Xg∈G∗, γ∈Γ∗
+
a(a + b)
Γ − G Xk∈K\{0}
N π
g,γab(
q1(g)m
q2(γ)m +
q2(γ)m
q1(g)m ) + Xγ,ξ∈Γ∗
γ,ξb2 q2(γ)m
N π
q2(ξ)m
(q1(k)m + q1(k)−m)
(q1(g)m + q1(g)−m) +
(q2(γ)m + q2(γ)−m)
2b(a + b)
Γ − G Xγ∈Γ∗
q1(k)m
q1(k′)m
+ δ(k,ε),(k,ε))
Γ − G
(
) +
q1(m)
q1(k)
q1(k)m
q1(g)m +
q1(h)m + b2 Xγ,ξ∈Γ∗
q1(g)m
(
q1(k)m
q2(γ)m +
q2(γ)m
q1(k)m )
2ab
Γ − G Xk∈K\{0}, γ∈Γ∗
Γ − G − δγ,ξ)
4
(
q2(γ)m
q2(ξ)m
+ δg,h)
(
q1(g)m
q2(γ)m +
q2(γ)m
q1(g)m )
=
+
+
2
CmC−m
Γ − G
AmA−m
Γ − G
(a + b)2
Γ − G
3a2
+
+
2
+ a2G∗ +
(a + b)(Cm + C−m)
(a + b)(Am + A−m)
+
+
Γ − G
CmA−m + AmC−m
Γ − G
CmB−m + BmC−m
+
Γ − G
BmB−m
Γ − G − b2Γ∗ +
Γ − G
AmB−m + BmA−m
Γ − G
= a + b + Am + Bm + Cm2
.
Γ − G
+
(a + b)(Bm + B−m)
Γ − G
130
PINHAS GROSSMAN AND MASAKI IZUMI
νm((k, ε))
= N (k,ε)
0,(k,ε)
a(a − b)
(q1(k)m + q1(k)−m) + N (k,ǫ)
π,π
(a + b)2
4
4
N (k,ε)
π,(l,δ)
a(a + b)
4
b(a + b)
2
N (k,ε)
π,γ
+ X(l,δ)∈J1
+ Xγ∈Γ∗
+ X(l,δ)∈J1, g∈G∗
+ Xg,h∈G∗
π,g
N (k,ε)
(q1(l)m + q1(l)−m) + Xg∈G∗
(q2(γ)m + q2(γ)−m) + X(l,δ),(l′,δ′)∈J1
q1(l)m ) + X(l,δ)∈J1, γ∈Γ∗
q1(g)m
(
N (k,ε)
(l,δ),g
a2
2
g,h a2 q1(g)m
N (k,ε)
q1(l)m
q1(g)m +
q1(h)m + Xg∈G∗, γ∈Γ
N (k,ε)
(l,δ),γ
ab
2
(
q1(l)m
q2(γ)m +
q2(γ)m
q1(l)m )
γ,ξ b2 q2(γ)m
N (k,ε)
q2(ξ)m
N (k,ε)
g,γ ab(
q1(g)m
q2(γ)m +
q2(γ)m
q1(g)m ) + Xγ,ξ∈Γ∗
a(a + b)
2
(q1(g)m + q1(g)−m)
N (k,ε)
(l,δ),(l′,δ′)
a2
4
q1(l)m
q1(l′)m
a(a − b)
4
(q1(k)m + q1(k)−m) +
(a + b)2
2(Γ − G)
+ δ(k,ε),(l,δ))(q1(l)m + q1(l)−m) +
(q1(g)m + q1(g)−m)
a(a + b)
Γ − G Xg∈G∗
(q2(γ)m + q2(γ)−m) +
a2
8 X(l,δ),(l′,δ′)∈J1
(
1
Γ − G
+ δl+l′,k)
q1(l)m
q1(l′)m
(δδ′hl, l + kiδl,l′ + δ′εhl′, l′ + liδl′,k + εδhk, k + l′iδk,l)
q1(l)m
q1(l′)m
(
q2(γ)m
q1(l)m
q1(g)m +
q1(l)m ) + a2 Xg,h∈G∗
Γ − G Xγ,ξ∈Γ∗
+ c2εδhk, l + giδl,k)(
q1(l)m
q2(γ)m +
q2(γ)m
q1(g)m ) +
(a + b)2
2(Γ − G)
CmC−m
2(Γ − G)
2b2
+
+
q1(g)m
q1(l)m )
2
Γ − G
q2(γ)m
q2(ξ)m
+ δg+h,k + δg−h,k)
q1(g)m
q1(h)m
(a + b)(Cm + C−m + Am + A−m + Bm + B−m)
a2
2
q1(l)m
q1(k − l)m −
2(Γ − G)
a2
(q1(k)m + q1(k)−m)
2
Γ − G
(
1
a2c2
b(a + b)
a(a + b)
4 X(l,δ)∈J1
Γ − G Xγ∈Γ∗
8 X(l,δ),(l′,δ′)∈J1
2 X(l,δ)∈J1, g∈G∗
2(Γ − G) X(l,δ)∈J1, γ∈Γ∗
Γ − G Xg∈G∗, γ∈Γ
2ab
Γ − G
a2
ab
1
(
(
(
q1(g)m
q2(γ)m +
a(a − b)
4
a(a + b)
4
(q1(k)m + q1(k)−m) +
(q1(k)m + q1(k)−m) +
=
+
+
+
+
+
+
=
+
+
+
AmC−m + CmA−m + BmC−m + CmB−m
AmA−m + AmB−m + BmA−m + BmB−m
2(Γ − G)
2(Γ − G)
+ a2 Xg∈G∗
q1(g)m
q1(g − k)m
4
N g
π,h
a(a + b)
2
(q1(h)m + q1(h)−m)
N g
(k,ε),(k′,ε′)
a2
4
q1(k)m
q1(k′)m
νm(g) = N g
0,g
a(a − b)
(q1(g)m + q1(g)−m) + N g
π,π
2
a(a + b)
4
(q1(k)m + q1(k)−m) + Xh∈G∗
(q2(γ)m + q2(γ)−m) + X(k,ε),(k′,ε′)
π,(k,ε)
b(a + b)
2
N g
N g
N g
π,γ
+ X(k,ε)∈J1
+ Xγ∈Γ∗
+ X(k,ε)∈J1;h∈G∗
+ Xh,h′∈G∗
(k,ε),h
(
a2
2
q1(k)m
q1(h)m +
q1(h′)m + Xh∈G∗, γ∈Γ∗
h,h′a2 q1(h)m
N g
=
a(a − b)
2
(q1(g)m + q1(g)−m) +
q1(h)m
q1(k)m ) + X(k,ε)∈J1;γ∈Γ∗
N g
(k,ε),γ
ab
2
(
N g
h,γab(
q1(h)m
q2(γ)m +
a(a + b)
q2(γ)m
q1(h)m ) + Xγ,ξ∈Γ∗
(a + b)2
Γ − G
+
Γ − G Xk∈K\{0}
q1(k)m
q2(γ)m +
q2(γ)m
q1(k)m )
γ,ξb2 q2(γ)m
N g
q2(ξ)m
(q1(k)m + q1(k)−m)
+ δg,h)(q1(h)m + q1(h)−m) +
(q2(γ)m + q2(γ)−m)
2b(a + b)
Γ − G Xγ∈Γ∗
INFINITE FAMILIES OF MODULAR DATA
131
= G(q1, m) + G(q2, m)2
2(Γ − G)
= G(q1, m) + G(q2, m)2
2(Γ − G)
+
+
a2
2 Xg∈G
q1(g)m
q1(k − g)m
δmk,0q1(k)m
.
2
(a + b)2
Γ − G
(
(
(
1
4
+
+
+
+
a2
a2
a(a + b)
Γ − G
2
2 Xh∈G∗
4 X(k,ε),(k′,ε′)
2 X(k,ε)∈J1;h∈G∗
Γ − G Xk∈K\{0};γ∈Γ∗
+ a2 Xh,h′∈G∗
Γ − G Xh∈G∗, γ∈Γ∗
Γ − G
4ab
2ab
+
4
(
Γ − G
+ c2εε′hk, k + giδk,k′)
q1(k)m
q1(k′)m
+ δk+g+h,0 + δk+g−h,0)(
q1(k)m
q1(h)m +
q1(h)m
q1(k)m )
(
q1(k)m
q2(γ)m +
q2(γ)m
q1(k)m )
+ δg+h+h′,0 + δ−g+h+h′,0 + δg−h+h′,0 + δg+h−h′,0)
q1(h)m
q1(h′)m
(
q1(h)m
q2(γ)m +
q2(γ)m
q1(h)m ) +
4b2
Γ − G Xγ,ξ∈Γ∗
q2(γ)m
q2(ξ)m
PINHAS GROSSMAN AND MASAKI IZUMI
132
=
+
a(a − b)
2
(q1(g)m + q1(g)−m) +
a(a + b)
2
(q1(g)m + q1(h)−g) +
+
(a + b)(Cm + C−m + Am + A−m + Bm + B−m)
(a + b)2
Γ − G
CmC−m + CmA−m + AmC−m + CmA−m + AmC−m
Γ − G
Γ − G
+
a2
2 X(k,ε)∈J1;h∈G∗
+ a2 Xh,h′∈G∗
(δk+g+h,0 + δk+g,h)(
q1(k)m
q1(h)m +
q1(h)m
q1(k)m )
(δg+h+h′,0 + δh+h′,g + δg+h′,h + δg+h,h′)
q1(h)m
q1(h′)m
AmA−m + AmB−m + A−mBm + BmB−m
+
(
q1(k)m
q1(g + k)m +
q1(g + k)m
q1(k)m )
q1(h)m
q1(h′)m
(δg+h+h′,0 + δh+h′,g + δg+h′,h + δg+h,h′)
Γ − G
+ a2Xk∈K
Γ − G
= G(q1) + G(q2)2
+ a2 Xh,h′∈G∗
= G(q1, m) + G(q2, m)2
Γ − G
+ δmg,0q1(g)m.
INFINITE FAMILIES OF MODULAR DATA
133
νm(γ) = N γ
0,γ
b(a − b)
(q2(γ)m + q2(γ)−m) + N γ
π,π
(a + b)2
4
2
a(a + b)
π,(k,ε)
b(a + b)
2
4
N γ
π,g
(q1(k)m + q1(k)−m) + Xg∈G∗
(q2(ξ)m + q2(ξ)−m) + X(k,ε),(k′,ε′)∈J1
q1(k)m ) + X(k,ε)∈J1, ξ∈Γ∗
q1(g)m
N γ
N γ
(
(k,ε),g
a2
2
q1(k)m
q1(g)m +
q1(h)m + Xg∈G∗, ξ∈Γ∗
g,ha2 q1(g)m
N γ
(q2(γ)m + q2(γ)−m) +
N γ
N γ
π,ξ
+ X(k,ε)∈J2
+Xξ∈Γ∗
+ X(k,ε)∈J1, g∈G∗
+ Xg,h∈G∗
=
b(a − b)
2
a(a + b)
2
(q1(g)m + q1(g)−m)
(k,ε),(k′,ε′)
a2
4
q1(k)m
q1(k′)m
N γ
(k,ε),ξ
ab
2
(
q2(ξ)m
q1(k)m )
q1(k)m
q2(ξ)m +
ξ,ξ′b2 q2(ξ)m
N γ
q2(ξ′)m
(q1(k)m + q1(k)−m)
N γ
g,ξab(
q1(g)m
q2(ξ)m +
q2(ξ)m
q1(g)m ) + Xξ,ξ′∈Γ∗
+
b(a + b)
a(a + b)
(a + b)2
Γ − G
Γ − G Xk∈K\{0}
2 Xξ∈Γ∗
Γ − G Xk∈K\{0}, g∈G∗
4
(
Γ − G − δξ,γ)(q2(ξ)m + q2(ξ)−m)
q1(k)m
q1(g)m +
q1(g)m
q1(k)m )
(
(q1(g)m + q1(g)−m) +
(
2a2
q1(k)m
q1(k′)m +
q1(k)m
q2(ξ)m
q2(ξ)m +
q1(k)m )
Γ − G Xg∈G∗, ξ∈Γ∗
4ab
q1(g)m
q1(h)m +
+
+
+
a2
2ab
2a(a + b)
Γ − G Xg∈G∗
Γ − G Xk,k′∈K\{0}
Γ − G Xk∈K\{0}, ξ∈Γ∗
Γ − G Xg,h∈G∗
+ b2 Xξ,ξ′∈Γ∗
4a2
+
4
(
b(a − b)
2
(
q1(g)m
q2(ξ)m +
q2(ξ)m
q1(g)m )
Γ − G − δγ+ξ+ξ′ − δ−γ+ξ+ξ′ − δγ−ξ+ξ′ − δγ+ξ−ξ′)
(q2(γ)m + q2(γ)−m) +
(a + b)2 + (a + b)(Cm + C−m + Am + A−m + Bm + B−m)
q2(ξ)m
q2(ξ′)m
CmC−m + CmA−m + AmC−m + CmB−m + BmC−m
Γ − G
Γ − G
=
−
+
b(a + b)
2
(q2(γ)m + q2(γ)−m) +
AmA−m + AmB−m + BmA−m + BmB−m
Γ − G
(δγ+ξ+ξ′ + δ−γ+ξ+ξ′ + δγ−ξ+ξ′ + δγ+ξ−ξ′)
− b2(q2(γ)m + q2(γ)−m)
Γ − G
(δγ+ξ+ξ′ + δ−γ+ξ+ξ′ + δγ−ξ+ξ′ + δγ+ξ−ξ′)
− b2 Xξ,ξ′∈Γ∗
= G(q1, m) + G(q2, m)2
− b2 Xξ,ξ′∈Γ∗
= G(q1, m) + G(q2, m)2
Γ − G
− δmγ,0q2(γ)m.
q2(ξ)m
q2(ξ′)m
q2(ξ)m
q2(ξ′)m
134
PINHAS GROSSMAN AND MASAKI IZUMI
Appendix E. Verification of Conjecture 4.5 for G = Z4.
E.1. Realization in O3 ⋊ Z2. Assume that a fusion category C with the following
fusion rules is realized in End(M) for a type III factor M.
(cid:3)
[α2] = [id],
[α][ρ] = [ρ][α],
[ρ2] = [α] + [α][ρ].
We also assume that α has a non-trivial associator. Such a fusion category certainly
exists (see [Izu17], [LMP15]).
We may assume that there exists a unitary U ∈ M satisfying α2 = AdU, α(U) =
−U. Moreover, we may assume U 2 = 1, and α as a Z4-action is stable (see [Con77]).
From [α][ρ] = [ρ][α], there exists a unitary V ∈ M satisfying AdV ◦α◦ρ = ρ◦α. Then
multiplying V by a complex number of modulus 1 if necessary, we may assume that V
is a α-cocycle as a Z4-action, and stability of α implies that it is a cobounday. Thus
there exists a unitary W ∈ M satisfying V = W ∗α(W ), which means that AdW ◦ ρ
commutes with α. Therefore from the beginning we may assume α ◦ ρ = ρ ◦ α.
We choose isometries s0 ∈ (α, ρ2), s1 ∈ (ρ, ρ2), s2 ∈ (α ◦ ρ, ρ2). Since α commutes
with ρ, we have α(s0) ∈ (α, ρ2), α(s1) ∈ (ρ, ρ2), α(s2) ∈ (α ◦ ρ, ρ2), and there exist
4-th roots of unity ξi, i = 0, 1, 2, satisfying α(s0) = ξ0s0, α(s1) = ξ1s1, α(s2) = ξ2s2.
Choosing a unitary Z ∈ M satisfying α(Z) = ξ−1
2 Z, and replacing ρ with AdZ ◦ ρ if
necessary, we may and do always assume ξ2 = 1, and hence α(s2) = s2.
Since Uρ(U) ∈ (ρ, ρ), there exists ǫ ∈ {1,−1} satisfying ρ(U) = ǫU. On one hand,
we have ρ2(U) = ǫ2U = U, and on the other hand,
which implies ξ2
ρ2(U) = s0α(U)s∗0 + s1ρ(U)s∗1 + s2α(ρ(U))s∗2 = −s0Us∗0 + ǫs1Us∗1 − ǫs2Us∗2,
1 = −1.
2 = −ǫ, and so we get ǫ = −1 and ξ2
Let ρ = α−1 ◦ ρ. Then ρ is the dual of ρ, and we may choose Rρ = s0. Since
ρ ◦ ρ = ρ ◦ ρ, we have Rρ = ζs0 with ζ = 1. Let d = 1 + √2. Then since
R∗ρρ(Rρ) = 1/d, we get s∗0ρ(s0) = ζ/d, which implies
0 = −1, ξ2
1 = ǫ, ξ2
0 = ξ2
ζ
d
s0 = ρ(s0)∗ρ2(s0)s0 = ρ(s0)∗s0α(s0) =
ζξ
d
s0.
Thus we get ξ1 = ζ 2, and ζ 4 = −1.
Note that Frobenius reciprocity implies √ds∗1ρ(s0) ∈ (ρ, ρρ) = (αρ, ρ2) and √ds∗2ρ(s0) ∈
(αρ, ρρ) = (ρ, ρ2) are isometries. We may assume √ds∗1ρ(s0) = s2 by redefining s2,
and there exists ν ∈ T satisfying √ds∗2ρ(s0) = νs1. Applying α to these, we get
ξ0 = ξ1, and ξ1 = ζ 2.
So far we have shown that we can choose α, ρ, U, si, i = 0, 1, 2, satisfying α(s0) =
ζ 2s0, α(s1) = ζ 2s1, α(s2) = s2, α(U) = ρ(U) = −U, α2 = AdU, Us0 = −s0U,
Us1 = −s1U, Us2 = s2U, and
ρ(s0) =
s0 +
(s1s2 + νs2s1),
ζ
d
1
√d
INFINITE FAMILIES OF MODULAR DATA
135
η1√d
where ζ is a number satisfying ζ 4 = −1.
Frobenius reciprocity implies that √dρ(s1)∗s0 ∈ (ρ, ρρ) = (αρ, ρ2) is an isom-
etry, and there exists η1 ∈ T satisfying √ds∗0ρ(s1) = η1s∗2. Since ρ(s1)s0 is an
isometry in (id, ρ3α−1) = Cs1s0, there exists η2 ∈ T satisfying ρ(s1)s0 = η2s1s0.
Since s∗1ρ(s1) ∈ (ρ2, ρ2) and s∗2ρ(s1) ∈ (ρ2, αρ2) = Cs1s∗2 + Cs2Us∗1, there exist
A, B, C, D ∈ C satisfying
ρ(s1) =
s0s∗2 + η2s1s0s∗0 + As1s1s∗1 + Bs1s2s∗2 + Cs2s1s∗2 + Ds2s2Us∗1.
=
s0s∗1 + ζ 2η1s2s0s∗0 + ζ 2Cs2s2s∗2 + νBs2s1s∗1 + As1s2s∗1 − ζ 2νDs1s1Us∗2.
Since s2 = √ds∗1ρ(s0), we get
ρ(s2) = √dρ(s∗1)ρ2(s0)
= √dρ(s1)∗(ζ 2s0s0s∗0 + s1ρ(s0)s∗1 + ζ 2s2ρ(s0)s∗2)
= ζ 2η1s2s0s∗0 + √d(η2s0s∗0 + As1s∗1 + Bs2s∗2)ρ(s0)s∗1 + ζ 2√d(Cs2s∗1 + Ds1Us∗2)ρ(s0)s∗2
ζη2√d
On the other hand, since s1 = ν√ds∗2ρ(s0),
ρ(s1) = ν√dρ(s2)∗ρ2(s0)
= ν√dρ(s2)∗(ζ 2s0s0s∗0 + s1ρ(s0)s∗1 + ζ 2s2ρ(s0)s∗2)
= νζη2s1s0s∗0 + ν√d(As1s∗2 − ζ
νDs2Us∗1)ρ(s0)s∗1
+ ζ 2ν√d(ζ
2
η1s0s∗0 + νBs1s∗1 + ζ
Cs2s∗2)ρ(s0)s∗2
ζνη1√d
s0s∗2 + ζνη2s1s0s∗0 + ζνAs1s1s∗1 + ζ 2ν 2Bs1s2s∗2 + Cs2s1s∗2 − ζ
ν2s2Us2s∗1,
=
2
2
2
which shows ν = ζ.
We summarize what we have gotten so far:
ρ(s0) =
ζ
d
s0 +
1
√d
(s1s2 + ζs2s1),
ρ(s1) =
ρ(s2) =
ζη2√d
s0s∗2 + η2s1s0s∗0 + As1s1s∗1 + Bs1s2s∗2 + Cs2s1s∗2 + Ds2s2Us∗1,
η1√d
s0s∗1 + ζ 2η1s2s0s∗0 + ζ 2Cs2s2s∗2 + ζBs2s1s∗1 + As1s2s∗1 + ζDs1s1Us∗2.
= ζ±11
16√2+√2
,
4 , Us0 = −s0U, Us1 = −s1U, Us2 = s2U, U 2 = 1, and α(s0) = ζ 2s0,
Theorem E.1. Let the notation be as above. Then ζ = ζ±1
D = 2− 1
α(s1) = ζ 2s1, α(s2) = s2, α(U) = −U, ρ(U) = −U,
8 , A = 1−√2∓i
2
ρ(s0) =
s0 +
(s1s2 + ζs2s1),
ζ
d
1
√d
ρ(s1) =
ζ
√d
s0s∗2 + ζ 3s1s0s∗0 + As1s1s∗1 + ζ 3As1s2s∗2 + ζAs2s1s∗2 + Ds2s2Us∗1,
136
PINHAS GROSSMAN AND MASAKI IZUMI
ρ(s2) =
2
ζ
√d
s0s∗1 + ζ 3s2s0s∗0 + ζ 3As2s2s∗2 + ζ
2
As2s1s∗1 + As1s2s∗1 + ζDs1s1Us∗2,
α2 = AdU, U 2 = 1.
Proof. Since s1 = ζη2√dρ(s2)∗s0,
ζ
d
s0 +
ρ(s1) = ζη2√dρ2(s∗2)ρ(s0)
= ζη2√dρ2(s∗2)(
1
√d
ζ 2η2√d
ζ 2η2√d
s0s∗2 + ζη2s1(ζ
=
=
2
s1s2 +
ζ
√d
s2s1)
s0α(s∗2) + ζη2s1ρ(s2)∗s2 + ζ 2η2s2ρ(α(s2))∗s1
η1s0s∗0 + ζBs1s∗1 + ζ
2
Cs2s∗2) + ζ 2η2s2(As1s∗2 + ζDs2Us∗1).
Thus we get η1 = ζ 2η2, η2 = ζη2η1, A = η2B, B = ζη2C, C = ζ 2η2A, D = ζ 3η2D.
This implies A = B = C, and ρ(s1)∗ρ(s1) = 1 implies
1 = A2 + D2.
+ 2A2,
1 =
1
d
Since neither A nor D is 0, and we get η1 = ζ, η2 = ζ 3, B = ζ 3A, and C = ζA.
Now the Cuntz algebra relation of ρ(s0), ρ(s1), ρ(s2) is equivalent to
1
d
1
d
+ (ζ 5 + 1)A = 0,
+ (ζ 2 + 1)A2 = 0,
1
,
A2 =
2 + √2
A2 + D2 = 1.
Solving these, we get the statement.
(cid:3)
Remark E.2. The phase of D cannot be determined from C because replacing s1 with
cs1 and s2 with cs2 changes D into c4D though other coefficients do not change. In
other words, we may always assume D = 2−1/4.
E.2. TubeC. Let
Λ = dimC = 2 + 2d2 = 2(2 + 2d) = 4(2 + √2).
It is easy to show that α has a unique half-braiding Eα(α) = ζ 2, Eα(ρ) = ζ 2,
Eα(αρ) = −1, and there is a unique extension α ∈ Z(C) of α. For simplicity, we
denote α = α if there there is no possibility of confusion. Since
(cid:18) S0,0 S0,α
Sα,0 Sα,α (cid:19) =
1
Λ(cid:18) 1
1
1 −1 (cid:19) ,
INFINITE FAMILIES OF MODULAR DATA
137
the category generated by α is non-degenerate. We set D = Z(C) ∩ {α}′, which is
the modular tensor category with
Z(C) = {0, α} ⊠ D.
Thus the modular data (S, T ) of Z(C) has the tensor product decomposition
(S, T ) = ((cid:18) 1√2
1√2
1√2 − 1√2 (cid:19) ⊗ S′,(cid:18) 1
0
0 ζ 2 (cid:19) ⊗ T ′),
where (S′, T ′) are the modular data of D. For X, Y ∈ Irr(D), we have S′X,Y = √2SX,Y .
We set s(α, X) = Sα,X
Sα,X
. Then by definition,
Irr(D) = {X ∈ Irr(Z(C)); s(α, X) = 1}.
2(10 + (0 α1α 0)), p(0, 1) = 1
p(0, 0) = 1
There exists a unique half-braiding for id, and
We determine the structure of A0, which is an abelian algebra of dimension 4. Let
2(10 − (0 α1α 0)), which are projections.
(10 + (0 α1α 0) + (1 + √2)(0 ρ1ρ 0) + (1 + √2)(0 αρ1αρ) 0)
4(2 + √2)
z(0) =
1
is a minimal central projection of TubeC. Let
E(0, 0) =
(10 + (0 α1α 0) + (1 −
1
4(2 − √2)
√2)(0 ρ1ρ 0) + (1 −
√2)(0 αρ1αρ) 0),
E(0, 1)ǫ =
1
2
p(0, 1)(1 + εζ 2(0 ρ1ρ 0)),
ε ∈ {1,−1},
which are projections. Then we have
A0 = Cz(0) ⊕ CE(0, 0) ⊕ CE(0, 1)+ ⊕ CE(0, 1)−
as an algebra, and E(0, 0)(0 ρ1ρ 0) = (1 − √2)E(0, 0), E(0, 1)ǫ(0; ρ1ρ 0) =
−ǫζ 2E(0, 1)ǫ.
The left A0 modules A0,ρ and A0,αρ have bases {p(0, 0)(0 ρs1ρ ρ), p(0, 1)(0 ρs1ρ ρ)}
and {p(0, 0)(0 ρs2Uρ αρ), p(0, 1)(0 ρs2Uρ αρ)} respectively. Since z(0) acts on
A0,ρ and A0,αρ trivially, E(0, 0) acts on p(0, 0)A0,ρ and p(0, 0)A0,αρ with multiplicity
1.
On the other hand, we have
(0 ρ1ρ 0)p(0, 1)(0 ρs1ρ ρ) = p(0, 1)(0 ρ1ρ 0)(0 ρs1ρ ρ)
= p(0, 1)((0 ρs∗1ρ(s1)s1ρ ρ) + (0 αρs∗2ρ(s1)s2αρ ρ))
= Ap(0, 1)((0 ρs1ρ ρ) + ζ(0 αρs1αρ ρ))
= Ap(0, 1)((0 ρs1ρ ρ) + ζ 5(0 α1α 0)(0 ρs1ρ ρ))
= A(1 − ζ 5)p(0, 1)(0 ρs1ρ ρ)
= −ζ 2p(0, 1)(0 ρs1ρ ρ)
138
and
PINHAS GROSSMAN AND MASAKI IZUMI
(0 ρ1ρ 0)p(0, 1)(0 ρs2Uρ αρ) = p(0, 1)(0 ρ1ρ 0)(0 ρs2Uρ αρ)
= p(0, 1)((0 ρs∗1ρ(s2U)s1ρ αρ) + (0 αρs∗2ρ(s2U)s2αρ αρ))
= p(0, 1)((0 ρs∗1ρ(s2)s1Uρ αρ) + (0 αρ − s∗2ρ(s2)s2Uαρ αρ))
= Ap(0, 1)((0 ρs2Uρ αρ) − ζ 3(0 αρs2Uαρ αρ))
= Ap(0, 1)((0 ρs2Uρ αρ) + ζ 3(0 α1α 0)(0 ρs2Uρ αρ))
= A(1 − ζ 3)p(0, 1)((0 ρs2Uρ αρ))
= ζ 2p(0, 1)((0 ρs2Uρ αρ)).
Summing up the above argument, we have the following.
Lemma E.3. Let π = id ⊕ ρ ⊕ αρ, ϕ = id ⊕ ρ, and ψ = id ⊕ αρ.
(1) π has a unique half-braiding with e(π)0,0 = E(0, 0), and Eπ(0)0,0 = 1, Eπ(α)0,0 =
(2) ϕ has a unique half-braiding with e( ϕ)0,0 = E(0, 1)+, and Eϕ(α)0,0 = −1,
(3) ψ has a unique half-braiding with e( ψ)0,0 = E(0, 1)−, and Eψ(α)0,0 = −1,
1, Eπ(ρ)0,0 = Eπ(αρ)0,0 = − 1
d2 .
Eϕ(ρ)0,0 = ζ 2/d.
Eψ(ρ)0,0 = −ζ 2/d.
The entries of S and T for these half-braidings are
2 + √2
S0,0 = Sπ,π =
, S0,π =
2 − √2
8
, S0, ϕ = S0, eψ =
1
4
,
8
1
4
,
S ϕ, ϕ =
Seπ, ϕ = Seπ, eψ =
1 ± i
4
, S ϕ, eψ =
1 ∓ i
4
T0,0 = Teπ,eπ = T ϕ, ϕ = T eψ, eψ = 1.
, S eψ, eψ =
1 ∓ i
4
.
The half-braiding equations for ρ are
Eρ(α)2 = −1,
ρ(s0)Eρ(α) = Eρ(ρ)ρ(Eρ(ρ))s0,
ρ(s1)Eρ(ρ) = Eρ(ρ)ρ(Eρ(ρ))s1,
ρ(s2)Eρ(αρ) = Eρ(ρ)ρ(Eρ(ρ))s2,
where Eρ(α) ∈ T, Eρ(ρ) ∈ (ρ2, ρ2) = Cs0s∗0+Cs1s∗1+Cs2s∗2, and Eρ(αρ) = Eρ(α)α(Eρ(ρ)) =
Eρ(α)Eρ(ρ). Direct computation shows the following.
Lemma E.4. There are exactly two half-braidings
ρ (α) = −ζ 2,
E 0
E 1
ρ (α) = ζ 2,
E 0
ρ (ρ) = ζ 3s0s∗0 + ζ 3s1s∗1 + ζ 6s2s∗2,
ρ (ρ) = ζ 5s0s∗0 + s1s∗1 + ζ 5s2s∗2,
E 1
INFINITE FAMILIES OF MODULAR DATA
139
for ρ. The entries of S and T for them are
1
4√2
,
s(α, ρ0) = 1,
S0,ρ0 = S0,ρ1 =
s(α, ρ1) = −1, Seπ,ρ0 = Seπ,ρ1 = −
, S eψ,ρi = ±i
S ϕ,ρi = ∓i
4
4
, Sρ0,ρ1 = −1 ∓ √2i
4√2
Tρ1,ρ1 = ∓i.
Tρ0,ρ0 = −1,
So far we have obtained simple objects
Sρ0,ρ0 = −1 ± √2i
4√2
1 ∓ √2i
4√2
, Sρ1,ρ1 =
.
1
4√2
,
0,eπ, α ϕ, αeψ, ρ0, α ρ1, α, αeπ, ϕ, eψ, α ρ0, ρ2
of Z(C), and the first half belong to D. Since dimAρ = dimαρ = 8 and dimAρ,αρ = 4,
there are only two missing, and they are eµ0 and eµ1 with µ = ρ⊕ αρ. We may assume
eµ0 ∈ D.
Using S(αX, Y ) = s(α, Y )SX,Y and TαX,αX = Tα,αTX,Xs(α, X), we can show from
our computation so far that S′ and T ′ with respect to
Theorem E.5. Let the notation be as above. Then
4
1
4
0,eπ, αeψ, α ϕ, ρ0, α ρ1,eµ0
− 1
4
±i
2√2
∓i
2√2
4 + ±i
4 + ∓i
√2+1
√2−1
4
1
2√2
1
2√2
− 1
4
− 1
− 1
T ′ = Diag(1, 1,∓i,∓i,−1,−1,∗).
1
2√2
1
2√2
−1±i
2√2
−1∓i
2√2
∓i
2√2 − 1
2√2 − 1
±i
0
1
2√2
1
2√2
−1∓i
2√2
−1±i
2√2
±i
2√2
∓i
2√2
0
2√2 − 1
2√2 − 1
∗
2
4
− 1
4
∓i
2√2
±i
2√2
4 + ∓i
2√2
4 + ±i
2√2
1
4
∗
1
4
4
1
4
1
2√2
1
2√2
−1∓i
2√2
−1±i
2√2
±i
2√2
∓i
2√2
0
1
2√2
1
2√2
−1±i
2√2
−1∓i
2√2
∓i
2√2 − 1
±i
2√2 − 1
0
√2+1
√2−1
4
1
2√2
1
2√2
− 1
4
− 1
− 1
T ′ = Diag(1, 1,∓i,∓i,−1,−1, ζ±3
8 ).
− 1
4
±i
2√2
∓i
2√2
4 + ±i
4 + ∓i
2√2 − 1
2√2 − 1
1
2
2
4
− 1
4
∓i
2√2
±i
2√2
4 + ∓i
2√2
4 + ±i
2√2
1
2
1
2
− 1
2
0
0
∗
∗
∗
1
2
− 1
2
0
0
1
2
1
2
0
,
,
are as follows:
S′ =
√2−1
√2+1
4
4
1
2√2
1
2√2
1
4
1
4
1
2
√2−1
√2+1
4
4
1
2√2
1
2√2
1
4
1
4
1
2
S′ =
Proof. Since S is a symmetric unitary matrix, the remaining entries are determined
as above. The Gauss sum formula for T determines the last corner of T ′.
(cid:3)
140
PINHAS GROSSMAN AND MASAKI IZUMI
References
[AMP15] Narjess Afzaly, Scott Morrison, and David Penneys. The classification of subfactors with
index at most 5 1
4 , 2015. arXiv:1509.00038.
[BJ15]
[BGH+17] Paul Bruillard, C´esar Galindo, Tobias Hagge, Siu-Hung Ng, Julia Yael Plavnik, Eric C.
Rowell, and Zhenghan Wang. Fermionic modular categories and the 16-fold way. J.
Math. Phys., 58(4):041704, 31, 2017.
Tathagata Basak and Ryan Johnson. Indicators of Tambara-Yamagami categories and
Gauss sums. Algebra Number Theory, 9(8):1793 -- 1823, 2015.
Bojko Bakalov and Alexander Kirillov, Jr. Lectures on tensor categories and modu-
lar functors, volume 21 of University Lecture Series. American Mathematical Society,
Providence, RI, 2001.
[BK01]
[BPMS12] Stephen Bigelow, Emily Peters, Scott Morrison, and Noah Snyder. Constructing the
[Con77]
extended Haagerup planar algebra. Acta Math., 209(1):29 -- 82, 2012.
A. Connes. Periodic automorphisms of the hyperfinite factor of type II1. Acta Sci. Math.
(Szeged), 39(1-2):39 -- 66, 1977.
[EG11]
[DMNO13] Alexei Davydov, Michael Muger, Dmitri Nikshych, and Victor Ostrik. The Witt group
of non-degenerate braided fusion categories. J. Reine Angew. Math., 677:135 -- 177, 2013.
David E. Evans and Terry Gannon. The exoticness and realisability of twisted Haagerup-
Izumi modular data. Comm. Math. Phys., 307(2):463 -- 512, 2011.
David E. Evans and Terry Gannon. Near-group fusion categories and their doubles. Adv.
Math., 255:586 -- 640, 2014.
[EG14]
[GIS18]
[Haa94]
[Hua05]
[Gan05]
[Gan18]
[GI19]
[EGNO15] Pavel Etingof, Shlomo Gelaki, Dmitri Nikshych, and Victor Ostrik. Tensor categories,
volume 205 of Mathematical Surveys and Monographs. American Mathematical Society,
Providence, RI, 2015.
Terry Gannon. Modular data: the algebraic combinatorics of conformal field theory. J.
Algebraic Combin., 22(2):211 -- 250, 2005.
Terry Gannon. The hypothetical smashed sum construction, 2018. Unpublished note.
Pinhas Grossman and Masaki Izumi. Drinfeld centers of fusion categories arising from
generalized haagerup subfactors, 2019. arxiv:1501.07679.
Pinhas Grossman, Masaki Izumi, and Noah Snyder. The Asaeda-Haagerup fusion cate-
gories. J. Reine Angew. Math., 743:261 -- 305, 2018.
Uffe Haagerup. Principal graphs of subfactors in the index range 4 < [M : N ] < 3 +√2.
In Subfactors (Kyuzeso, 1993), pages 1 -- 38. World Sci. Publ., River Edge, NJ, 1994.
Yi-Zhi Huang. Vertex operator algebras, the Verlinde conjecture, and modular tensor
categories. Proc. Natl. Acad. Sci. USA, 102(15):5352 -- 5356, 2005.
Masaki Izumi. Subalgebras of infinite C ∗-algebras with finite Watatani indices. I. Cuntz
algebras. Comm. Math. Phys., 155(1):157 -- 182, 1993.
Masaki Izumi. The structure of sectors associated with Longo-Rehren inclusions. II.
Examples. Rev. Math. Phys., 13(5):603 -- 674, 2001.
Masaki Izumi. A Cuntz algebra approach to the classification of near-group categories. In
Proceedings of the 2014 Maui and 2015 Qinhuangdao conferences in honour of Vaughan
F. R. Jones' 60th birthday, volume 46 of Proc. Centre Math. Appl. Austral. Nat. Univ.,
pages 222 -- 343. Austral. Nat. Univ., Canberra, 2017.
Masaki Izumi. The classification of 3n subfactors and related fusion categories. Quantum
Topol., 9(3):473 -- 562, 2018.
Vaughan F. R. Jones, Scott Morrison, and Noah Snyder. The classification of subfactors
of index at most 5. Bull. Amer. Math. Soc. (N.S.), 51(2):277 -- 327, 2014.
Zhengwei Liu, Scott Morrison, and David Penneys. 1-supertransitive subfactors with
index at most 6 1
Michael Muger. On the structure of modular categories. Proc. London Math. Soc. (3),
87(2):291 -- 308, 2003.
5 . Comm. Math. Phys., 334(2):889 -- 922, 2015.
[LMP15]
[JMS14]
[Izu93]
[Izu01]
[Izu17]
[Izu18]
[M03]
INFINITE FAMILIES OF MODULAR DATA
141
[MS90]
[MS17]
[NS07]
[Sie03]
[Tho12]
[Tur92]
[Ver88]
Gregory Moore and Nathan Seiberg. Lectures on RCFT. In Physics, geometry, and
topology (Banff, AB, 1989), volume 238 of NATO Adv. Sci. Inst. Ser. B Phys., pages
263 -- 361. Plenum, New York, 1990.
Michael Mignard and Peter Schauenburg. Modular categories are not determined by
their modular data, 2017.
Siu-Hung Ng and Peter Schauenburg. Frobenius-Schur indicators and exponents of
spherical categories. Adv. Math., 211(1):34 -- 71, 2007.
Jacob Siehler. Near-group categories. Algebr. Geom. Topol., 3:719 -- 775, 2003.
Josiah E. Thornton. Generalized near-group categories. ProQuest LLC, Ann Arbor, MI,
2012. Thesis (Ph.D.) -- University of Oregon.
Vladimir G. Turaev. Modular categories and 3-manifold invariants. Internat. J. Modern
Phys. B, 6(11-12):1807 -- 1824, 1992. Topological and quantum group methods in field
theory and condensed matter physics.
Erik Verlinde. Fusion rules and modular transformations in 2D conformal field theory.
Nuclear Phys. B, 300(3):360 -- 376, 1988.
School of Mathematics and Statistics, University of New South Wales, Sydney
NSW 2052, Australia
E-mail address: [email protected]
Department of Mathematics, Graduate School of Science, Kyoto University,
Sakyo-ku, Kyoto 606-8502, Japan
E-mail address: [email protected]
|
1703.04486 | 8 | 1703 | 2018-03-20T18:01:01 | Euler totient of subfactor planar algebras | [
"math.OA",
"math.CO",
"math.GR",
"math.RT"
] | We extend the Euler's totient function (from arithmetic) to any irreducible subfactor planar algebra, using the Mobius function of its biprojection lattice, as Hall did for the finite groups. We prove that if it is nonzero then there is a minimal 2-box projection generating the identity biprojection. We explain a relation with a problem of K.S. Brown. As an application, we define the dual Euler totient of a finite group and we show that if it is nonzero then the group admits a faithful irreducible complex representation. We also get an analogous result at depth 2, involving the central biprojection lattice. | math.OA | math |
EULER TOTIENT OF SUBFACTOR PLANAR ALGEBRAS
S´EBASTIEN PALCOUX
Abstract. We extend the Euler's totient function (from arithmetic) to any
irreducible subfactor planar algebra, using the Mobius function of its bipro-
jection lattice, as Hall did for the finite groups. We prove that if it is nonzero
then there is a minimal 2-box projection generating the identity biprojection.
We explain a relation with a problem of K.S. Brown. As an application, we
define the dual Euler totient of a finite group and we show that if it is nonzero
then the group admits a faithful irreducible complex representation. We also
get an analogous result at depth 2, involving the central biprojection lattice.
1. Introduction
Any finite group G acts outerly on the hyperfinite II1 factor R, and the group
subfactor (R ⊆ R ⋊ G), of index G, remembers the group [5]. Jones proved in [6]
that the set of possible values for the index M : N of a subfactor (N ⊆ M ) is
) n ≥ 3} ⊔ [4,∞].
{4cos2(
π
n
By Galois correspondence [13], the lattice of intermediate subfactors of (R ⊆ R⋊G)
is isomorphic to the subgroup lattice of G. Moreover, Watatani [21] extended the
finiteness of the subgroup lattice to any irreducible finite index subfactor. Then, the
subfactor theory can be seen as an augmentation of the finite group theory, where
the indices are not necessarily integers. The notion of subfactor planar algebra [8]
is a diagrammatic axiomatization of the standard invariant of a finite index II1
subfactor [7]. Bisch [2] proved that the intermediate subfactors are given by the
biprojections (see Definition 3.1) in the 2-box space of the corresponding planar
algebra. The recent results of Liu [12] on the biprojections are also crucial for this
paper (see Section 3).
The usual Euler's totient function ϕ(n) counts the number of positive integers
up to n that are relatively prime to n. Let G be a finite group and µ the Mobius
function (see Section 4, Definition 4.1) of its subgroup lattice L(G). Hall proved in
[4] that the Euler totient of G (defined below) is the cardinal of {g ∈ G hgi = G}.
ϕ(G) := XH∈L(G)
µ(H, G)H
So if ϕ(G) is nonzero then G is cyclic, and ϕ(Cn) = ϕ(n). This implication will be
generalized in Section 5, as the author did with Ore's theorem in [15, 16]. Let P be
2010 Mathematics Subject Classification. 46L37 (Primary), 05E10, 05E15, 06B15, 20C15,
20D30 (Secondary).
Key words and phrases. von Neumann algebra; subfactor; planar algebra; biprojection; lattice;
Mobius function; Euler totient; Boolean algebra.
1
ϕ(G) := XH∈L(G)
µ(1, H)G : H
2
S ´EBASTIEN PALCOUX
an irreducible subfactor planar algebra and µ the Mobius function of its biprojection
lattice [e1, id]. We use notations of Definition 3.3. Let the Euler totient of P be
ϕ(P) := Xb∈[e1,id]
µ(b, id)b : e1.
Theorem 1.1. If ϕ(P) is nonzero then P is w-cyclic (i.e. there is a minimal 2-box
projection generating the identity biprojection).
Then, for any finite group G, by considering P(RG ⊂ R), we get that if the dual
Euler totient
is nonzero then G has a faithful irreducible complex representation. It is a dual
version of the initial implication. As a general application, we get a non-trivial
upper bound for the minimal number of minimal central projections generating the
identity biprojection. By applying this result to any finite group G, we deduce a
non-trivial upper bound for the minimal number of irreducible components for a
faithful complex representation of G. It involves the subgroup lattice and indices
only. It is a link between combinatorics and representations in finite group theory.
These results were not known to group theorists and the author can provide a group
theoretical translation of the proofs as he did in [17] for the dual Ore's theorem.
This path of research conducts to (Section 6) a generalization of a problem of K.S.
Brown (considered hard in [19]) and to a possible counter-example of Watatani's
problem [21] on whether any finite lattice has a biprojection lattice representation.
We finally prove an additional result for the irreducible subfactor planar algebras
of depth 2, involving the central biprojection lattice. The dual group case recovers
a known result involving the normal subgroup lattice.
2. Basics on lattice theory
A lattice (L,∧,∨) is a poset L in which every two elements a, b have a unique
supremum (or join) a ∨ b and a unique infimum (or meet ) a ∧ b. Let G be a finite
group. The set of subgroups K ⊆ G forms a lattice, denoted by L(G), ordered by
⊆, with K1 ∨ K2 = hK1, K2i and K1 ∧ K2 = K1 ∩ K2. A sublattice of (L,∧,∨) is a
subset L′ ⊆ L such that (L′,∧,∨) is also a lattice. If a, b ∈ L with a ≤ b, then the
interval [a, b] is the sublattice {c ∈ L a ≤ c ≤ b}. Any finite lattice is bounded,
i.e. admits a minimum and a maximum, denoted by 0 and 1. The atoms are the
minima of L\{0}. The coatoms are the maxima of L\{1}. Consider a finite lattice,
b the join of its atoms and t the meet of its coatoms, then let call [0, b] and [t, 1] its
bottom and top intervals. A lattice is distributive if the join and meet operations
distribute over each other. A distributive bounded lattice is called Boolean if any
element b admits a unique complement b∁ (i.e. b ∧ b∁ = 0 and b ∨ b∁ = 1). The
subset lattice of {1, 2, . . . , n}, with union and intersection, is called the Boolean
lattice Bn of rank n. Any finite Boolean lattice is isomorphic to some Bn.
Remark 2.1. A finite lattice is Boolean if and only if it is uniquely atomistic, i.e.
every element can be written uniquely as a join of atoms. It follows that if [a, b]
and [c, d] are intervals in a Boolean lattice, then
[a, b] ∨ [c, d] := {k ∨ k′ k ∈ [a, b], k′ ∈ [c, d]},
is the interval [a ∨ c, b ∨ d].
EULER TOTIENT OF SUBFACTOR PLANAR ALGEBRAS
3
We refer to [20] for more details.
3. Subfactor planar algebras and biprojections
For the notions of subfactor, subfactor planar algebra and basic properties, we
refer to [7 -- 9]. See also [14, Section 3] for a short introduction. Let (N ⊆ M )
be a finite index irreducible subfactor. The n-box spaces Pn,+ and Pn,− of the
planar algebra P = P(N ⊆ M ), are N ′ ∩ Mn−1 and M ′ ∩ Mn. Let R(a) be the
range projection of a ∈ P2,+. We define the relations a (cid:22) b by R(a) ≤ R(b),
and a ∼ b by R(a) = R(b). Let e1 := eM
M be the Jones and the
identity projections in P2,+. Let tr be the normalized trace (i.e. tr(id) = 1). Then
tr(e1) = M : N−1 = δ−2. Let F : P2,± → P2,∓ be the Fourier transform (1-click
or 90◦ rotation) and let a∗ b = F (F −1(a)·F −1(b)) be the coproduct of a, b ∈ P2,±.
If a, b are positive then so is a ∗ b by [12, Theorem 4.1].
Definition 3.1 ([12] Def. 2.14). A biprojection is a nonzero projection b ∈ P2,±
with F (b) a multiple of a projection.
N and id = eM
M are biprojections.
Note that e1 = eM
N and id := eM
Theorem 3.2 ([2] p212). A projection b ∈ P2,+ is a biprojection if and only if it
is the Jones projection eM
K of an intermediate subfactor N ⊆ K ⊆ M .
Therefore, the set of biprojections is a lattice of the form [e1, id].
Definition 3.3. Consider the intermediate subfactors N ⊆ K1 ⊆ K2 ⊆ M , and let
bi ∈ [e1, id] be the biprojection eM
Ki. We define P(b1, b2) := P(K1 ⊆ K2) and
b2 : b1 := tr(b2)/ tr(b1) = K2 : K1.
Definition 3.4. Consider a ∈ P2,+ positive, and let pn be the range projection of
Pn
k=1 a∗k. By finiteness, there exists N such that for all m ≥ N , pm = pN , which
is a biprojection [12, Lemma 4.14], denoted hai, called the biprojection generated
by a. It is the smallest biprojection b (cid:23) a. For S a finite set of positive elements
of P2,+, let hSi be hPs∈S si.
Proposition 3.5 ([15]). Let p ∈ P2,+ be a minimal central projection. Then, there
exists u ≤ p minimal projection such that hui = hpi.
Definition 3.6 ([15]). A planar algebra P is weakly cyclic (or w-cyclic) if it
satisfies one of the following equivalent assertions:
• ∃u ∈ P2,+ minimal projection such that hui = id,
• ∃p ∈ P2,+ minimal central projection such that hpi = id.
4. Euler totient
We define a notion of Euler totient on the irreducible subfactor planar algebras
as an extension of the usual Euler's totient function on the positive integers.
Definition 4.1. The Mobius function µ of a finite poset P is defined inductively
as follows. For a, b ∈ P with a ≤ b,
µ(a, b) :=( 1
otherwise.
if a = b,
−Pc∈(a,b] µ(c, b)
4
S ´EBASTIEN PALCOUX
The following result can be seen as a Boolean representation for the Mobius
function of a finite lattice.
Theorem 4.2 (Crosscut Theorem). Let L be a finite lattice and a1, . . . , an its
coatoms. Consider the (order-reversing) map m : Bn → L such that
Then for any a ∈ L,
m(I) =( 1
Vi∈I ai
if I = ∅,
otherwise.
µ(a, 1) = XI∈m−1({a})
(−1)I
Proof. It is a reformulation of [20, Corollary 3.9.4] on [a, 1] with its coatoms. (cid:3)
Definition 4.3. Let P be an irreducible subfactor planar algebra, with biprojection
lattice [e1, id] and Mobius function µ. For any b1, b2 ∈ [e1, id] with b1 ≤ b2, consider
ϕ(b1, b2) := Xb∈[b1,b2]
The Euler totient of P is ϕ(P) := ϕ(e1, id).
Remark 4.4. Let P be the dual subfactor planar algebra of P. Then
µ(b, b2)b : b1
ϕ(P) = Xb∈[e1,id]
µ(b, id)b : e1 and ϕ( P) = Xb∈[e1,id]
µ(e1, b) id : b,
because the biprojection lattice of P is the reversed from that of P.
Lemma 4.5 ([4]). The Euler totient of a finite group G,
ϕ(G) := XH∈L(G)
µ(H, G)H,
is the cardinal of {g ∈ G hgi = G}.
Proof. By Theorem 4.2 with its map m : Bn → L(G),
(−1)αH = Xα∈Bn
ϕ(G) = XH∈L(G) Xα∈m−1({H})
(−1)αm(α).
Then, by the inclusion-exclusion principle, ϕ(G) = G \Si m({i}).
Corollary 4.6. A finite group G is cyclic if and only if ϕ(G) is nonzero.
Proof. If G is cyclic then G = Cn for some positive integer n, and ϕ(G) = ϕ(n) 6= 0.
Conversely, if ϕ(G) 6= 0, then G is cyclic by Lemma 4.5.
(cid:3)
(cid:3)
Note that ϕ(P(R ⊆ R ⋊ G)) = ϕ(G) and ϕ(Cn) = ϕ(n), the usual Euler's totient
function of n. Thus, we can see P 7→ ϕ(P) as an extension from the positive integers
to the irreducible subfactor planar algebras. The following proposition extends the
usual formula that if n =Qi pni
is the prime factorization of n, then
i
pni−1
i
ϕ(n) =Yi
(pi − 1).
·Yi
Proposition 4.7. Let [t, id] be the top interval of [e1, id]. Then
ϕ(P) = t : e1 · ϕ(t, id).
EULER TOTIENT OF SUBFACTOR PLANAR ALGEBRAS
5
Proof. If b 6∈ [t, id] then µ(b, id) = 0 by Theorem 4.2 because m−1({b}) = ∅. Next
if b ∈ [t, id] then b : e1 = b : t · t : e1, since the index is multiplicative [6].
(cid:3)
5. Main result
In this section, we generalize one way of Corollary 4.6 by the following theorem.
We will see later with Remark 7.12 that the converse is false in general.
Theorem 5.1. An irreducible subfactor planar algebra P is w-cyclic if its Euler
totient ϕ(P) is nonzero.
Proof. Let p1, . . . , pr be the minimal central projections of P2,+. Consider the sum
S(i) := Xb∈[e1,id]
µ(b, id) tr(bpi).
By Theorem 4.2 with its map m : Bn → [e1, id],
S(i) = Xb∈[e1,id] Xβ∈m−1({b})
(−1)β tr(bpi) = Xβ∈Bn
(−1)β tr(m(β)pi).
Recall that the map m is order-reversing and the image of the atoms of Bn are the
coatoms of [e1, id]; let call these latter ones b1, . . . , bn. Let Ai be the set of atoms
α of Bn satisfying pi ≤ m(α), and Bi the set of atoms not in Ai. Let αi (resp. βi)
be the join of all the elements of Ai (resp. Bi).
Claim: For α ∈ Bn, pi ≤ m(α) ⇔ α ∈ [αi, 1].
Proof: Just observe that pi ≤Vj∈α bj if and only if ∀j ∈ α, pi ≤ bj.
Now by Remark 2.1, we have
(cid:4)
Bn = [∅, αi] ∨ [∅, βi] = Gα∈[∅,αi]
α ∨ [∅, βi],
Consider the following sum
T (i) := Xβ∈[∅,βi]
(−1)β tr(m(β)pi)
For any α ∈ [∅, αi] and β ∈ [∅, βi], we have (−1)α∨β = (−1)α(−1)β and
m(α ∨ β)pi = m(α)pi ∧ m(β)pi = m(β)pi.
So we get that
S(i) = Xα∈[∅,αi]
(−1)αT (i) = T (i) · (1 − 1)Ai.
Claim: The planar algebra P is w-cyclic if and only if ∃i with Ai = 0.
Proof: It is w-cyclic if and only if ∃i with hpii = id, if and only if for any coatom
b ∈ [e1, id], pi 6≤ b, if and only if Ai = 0.
If P is not w-cyclic, then ∀i Ai 6= 0, so S(i) = 0; but b : e1 = tr(b)/ tr(e1),
tr(b) =Pi tr(bpi) and tr(e1) = δ−2, so ϕ(e1, id) = δ2Pr
It is a double generalization of [1, Theorem 3.10] from groups to subfactors and
from Boolean lattice to any lattice. It is also a purely combinatorial criterion for a
subfactor planar algebra to be w-cyclic.
i=1 S(i) = 0.
(cid:4)
(cid:3)
6
S ´EBASTIEN PALCOUX
6. Generalization of a problem of K.S. Brown
We will explain how this path of research conducts to a generalization of a
problem of K.S. Brown in finite group theory and to a possible counter-example of
the problem of Watatani on representing any finite lattice as a biprojection lattice.
Proposition 6.1. Assume that every biprojection is central. Then
ϕ(P) = δ2 Xhpii=id
tr(pi)
with p1, . . . , pr the minimal central projections of P2,+.
Proof. First of all, b : e1 = tr(b)/ tr(e1) and tr(e1) = δ−2, so
ϕ(P) = δ2 Xb∈[e1,id]
µ(b, id) tr(b).
By Theorem 4.2 with its map m : Bn → [e1, id],
ϕ(P) = δ2 Xβ∈Bn
(−1)β tr(m(β)),
but every biprojection is central, so by the inclusion-exclusion principle
ϕ(P) = δ2[tr(id) − tr(R[
n
Xi=1
m({i})])] = δ2 Xhpii=id
tr(pi).
(cid:3)
It follows that the converse of Theorem 5.1 is true if every biprojection is central;
if moreover they form a Boolean lattice, then by [15, Theorem 4.26], P is w-cyclic,
so ϕ(P) is nonzero; and without assuming the biprojections to be central, P is still
w-cyclic by [16, Theorem 5.9], and ϕ(P) is suspected to be still nonzero:
Conjecture 6.2. If [e1, id] is Boolean then ϕ(P) is nonzero.
Remark 6.3. If [e1, id] is Boolean of rank n + 1, we wonder whether ϕ(P) ≥ φn,
with φ the golden ratio. If this lower bound is correct, then it is optimal as realized
by the tensor product T LJ (√2) ⊗ T LJ (φ)⊗n, thanks to [15, Theorem 4.8].
By Theorem 5.1, a proof of Conjecture 6.2 would be stronger than [16]. This
conjecture can be seen as the Boolean restriction of the following planar algebraic
generalization of a problem of K.S. Brown (considered hard in [19]):
Conjecture 6.4. Let P be an irreducible subfactor planar algebra. Then its Euler
characteristic (defined below) is nonzero.
χ(P) := − Xb∈[e1,id]
µ(b, id) id : b
Gaschutz proved the usual Brown's problem, i.e. for P(R ⊆ R ⋊ G), if G is solvable.
It is an extension of Conjecture 6.2 because by Remark 4.4, χ(P) = ±ϕ( P) if [e1, id]
is Boolean (of rank n), because then µ(e1, b) = (−1)nµ(b, id). This last equality
holds more generally for any Eulerian lattice, which is a graded lattice such that
µ(a, b) = (−1)b−a for a ≤ b, with a 7→ a the rank function. This leads us to:
Conjecture 6.5. If [e1, id] is Eulerian then ϕ(P) is nonzero, and so P is w-cyclic.
EULER TOTIENT OF SUBFACTOR PLANAR ALGEBRAS
7
The last part of Conjecture 6.5, deduced from the first together with Theorem 5.1,
would provide an extension of Ore's theorem [16, Theorem 5.9] from top Boolean
to top Eulerian lattices (notation: a lattice is called top X if its top interval is X).
Note that the face lattice of any convex polytope is Eulerian [20, Proposition 3.8.9].
Proposition 6.6. An Eulerian subgroup lattice (of a finite group) is Boolean.
Proof. If L(G) is Eulerian then µ(1, G) = ±1, but [11, Th´eor`eme 3.1] states that
µ(1, G) ∈
Z
G
G : G′0
with G′ the commutator subgroup of G and G : G′0 the square-free part of G : G′.
Then G = G : G′0, and so G′ = 1. It follows that G is abelian with G square-
free, so G is cyclic of square-free order and L(G) is Boolean.
It is unknown whether Proposition 6.6 extends to any interval. This leads us to:
(cid:3)
Question 6.7. Is there an irreducible subfactor planar algebra with a non-Boolean
Eulerian biprojection lattice?
It is unknown whether any finite lattice admits a biprojection lattice representation
[21, 22], even in the group-subgroup case, so the smallest non-Boolean Eulerian
lattice (the face lattice of the square polytope, below) could be a counter-example.
•
•
•
•
•
•
•
•
•
•
7. Applications
We will deduce a non-trivial upper bound from Theorem 5.1 providing a link
between combinatorics and representations in finite group theory by translating.
Definition 7.1. The Euler totient of an interval of finite groups [H, G] is
µ(K, G)K : H.
ϕ(H, G) := XK∈[H,G]
Note that ϕ(P(R ⋊ H ⊆ R ⋊ G)) = ϕ(H, G).
Corollary 7.2. There is g ∈ G with hHgi = G if and only if ϕ(H, G) is nonzero.
Proof. By Proposition 6.1; or directly, let M1, . . . , Mn be the coatoms of [H, G].
Then, as for the proof of Lemma 4.5,
so that ϕ(H, G) is the cardinal of {Hg g ∈ G and hHgi = G}.
Hϕ(H, G) = G \[ Mi,
(cid:3)
8
S ´EBASTIEN PALCOUX
Corollary 7.3. The minimal cardinal for a generating set of a finite group G, is
the minimal length ℓ for an ordered chain of subgroups
{e} = H0 < H1 < ··· < Hℓ = G
such that ϕ(Hi, Hi+1) is nonzero.
Proof. A subset {g1, . . . , gn} is generating iff there is an ordered chain of subgroups
{e} = K0 < K1 < ··· < Kn = G
with Ki+1 = hKi, gi+1i = hKigi+1i, iff ϕ(Ki, Ki+1) is nonzero by Corollary 7.2. (cid:3)
So, the minimal cardinal for a generating set of a finite group G depends only on
the subgroup lattice and indices (known to [4]). We will generalize Corollary 7.3
by a non-trivial upper bound. Let P be an irreducible subfactor planar algebra.
Lemma 7.4 ([15]). Let b1 < b2 be biprojections. If P(b1, b2) is w-cyclic, then there
is a minimal projection u ∈ P2,+ such that b2 = hb1, ui.
Theorem 7.5. The minimal number r of minimal projections generating the iden-
hu1, . . . , uri = id) is at most the minimal length ℓ for an
tity biprojection (i.e.
ordered chain of biprojections
e1 = b0 < b1 < ··· < bℓ = id
such that ϕ(bi, bi+1) is nonzero.
Proof. Consider a chain as above. By Theorem 5.1, P(bi, bi+1) is w-cyclic, so by
Lemma 7.4, there is a minimal projection ui+1 ∈ P2,+ such that bi+1 = hbi, ui+1i.
It follows that hu1, . . . , uℓi = id.
(cid:3)
We deduce weak dual versions of Corollaries 7.2, 4.6 and 7.3, giving the link
between combinatorics and representation theory:
Definition 7.6. The dual Euler totient of the interval [H, G] is
ϕ(H, G) := XK∈[H,G]
µ(H, K)G : K.
Note that ϕ(P(RG ⊆ RH)) = ϕ(H, G).
Definition 7.7. Let W be a representation of a group G, K a subgroup of G, and
X a subspace of W . Let the fixed-point subspace be
and the pointwise stabilizer subgroup be
W K := {w ∈ W kw = w ,∀k ∈ K}
G(X) := {g ∈ G gx = x ,∀x ∈ X}
Definition 7.8. An interval of finite group [H, G] is said to be linearly primitive
if there is an irreducible complex representation V of G with G(V H ) = H.
Theorem 7.9 ([15]). The planar algebra P(RG ⊆ RH ) is w-cyclic if and only if
the interval of finite groups [H, G] is linearly primitive.
Corollary 7.10. If the dual Euler totient ϕ(H, G) is nonzero, then the interval of
finite groups [H, G] is linearly primitive.
Proof. Apply Theorem 5.1 to P(RG ⊆ RH ), and then Theorem 7.9.
(cid:3)
EULER TOTIENT OF SUBFACTOR PLANAR ALGEBRAS
9
In particular, let the dual Euler totient of G be ϕ(G) := ϕ(1, G). Then:
Corollary 7.11. A finite group G admits a faithful irreducible complex represen-
tation if its dual Euler totient ϕ(G) is nonzero.
Proof. It follows from Corollary 7.10 because G(V ) = ker(πV ).
(cid:3)
It is a purely combinatorial criterion for a finite group to have an irreducible faithful
complex representation.
Remark 7.12. The converse is false. The modular maximal-cyclic group M4(2) (of
order 16) has a faithful irreducible complex representation, whereas ϕ(M4(2)) = 0.
This is not surprising because according to Proposition 4.7, ϕ(e1, id) 6= 0 if and
only if ϕ(t, id) 6= 0 with [t, id] the top interval of [e1, id]; and the bottom interval of
[1, M4(2)] is [1, C2
2 ]. Even if we assume that [1, G] is its own bottom interval, the
converse is still false: there are exactly two counter-examples of index ≤ 100, given
by G = D8 ⋊ C2
Theorem 7.13 ([3] §226). A complex representation V of a finite group G is faithful
if and only if for any irreducible complex representation W there is an integer n
such that W (cid:22) V ⊗n.
Corollary 7.14. The minimal number of irreducible components for a faithful
complex representation of G, is at most the minimal length ℓ for an ordered chain
of subgroups
2 or D8 ⋊ S3 (of order 64 and 96 respectively).
{e} = H0 < H1 < ··· < Hℓ = G
such that ϕ(Hi, Hi+1) is nonzero.
Proof. Consider a chain as above. By Corollary 7.5 applied to P(RG ⊆ R), we have
hu1, . . . , uℓi = id, with ui minimal projection. Let pi be the central support of ui.
Then hp1 + ··· + pℓi = id. But the coproduct of two minimal central projections
is given by the tensor product of the associated irreducible representations of G
[16, Corollary 7.5]. So by Definition 3.4 and Theorem 7.13, the representation
V1 ⊕···⊕ Vℓ (with Vi the irreducible complex representation im(pi)) is faithful. (cid:3)
Note that Corollary 7.14 extends to any finite dimensional Kac algebra as for [16,
Remark 6.14]. It can also be improved by taking for H0 any core-free subgroup of
H1 (instead of just {e}), thanks to [16, Lemma 6.13]. In particular:
Corollary 7.15. A finite group G admits a faithful irreducible complex represen-
tation if there is a core-free subgroup H < G with ϕ(H, G) nonzero.
This criterion is more efficient than Corollary 7.11 (consider for example G simple),
but it is no more purely combinatorial.
Question 7.16. Is the converse of Corollary 7.15 true?
8. Additional result for the depth 2
We can prove an additional result in the depth 2 case (involving the central
biprojection lattice) coming from the fact that the irreducible depth 2 subfactor
planar algebras correspond to the Kac algebras [10]. Note that the group subfactors
are depth 2, since the group algebras are examples of Kac algebras.
10
S ´EBASTIEN PALCOUX
Theorem 8.1 (Splitting, [10] p39). Let P be an irreducible depth 2 subfactor planar
algebra. Any element x ∈ P2,+ splits as follows:
x = x(1)
x(2)
and
x = x(1)
x(2)
Note that ∆(x) = x(1)⊗x(2) is the sumless Sweedler notation for the comultiplication
of the corresponding Kac algebra.
Corollary 8.2. If a, b ∈ P2,+ are central, then so is the coproduct a ∗ b.
Proof. This diagrammatic proof by splitting is due to Vijay Kodiyalam.
(a ∗ b) · x =
x
=
a
b
a
b
x(1) x(2)
a
b
a
b
=
=
x(1) x(2)
x
= x · (a ∗ b)
(cid:3)
Corollary 8.3. The set of central biprojections is a sublattice of [e1, id].
Proof. Let b1, b2 be central biprojections. Then b1 ∧ b2 is central, and b1 ∨ b2 is the
range projection of (b1 ∗ b2)∗k for k large enough, so is central by Corollary 8.2. (cid:3)
Let C be the central biprojection lattice and µC its Mobius function. Let the
central Euler totient of P be
Theorem 8.4. Let p1, . . . , pr be the minimal central projections of P2,+. Then
µC(b, id)b : e1.
ϕC(P) :=Xb∈C
ϕC(P) = δ2 Xhpii=id
tr(pi).
Proof. By Corollary 8.2, a central projection generates a central biprojection. The
rest of the proof is identical to that of Proposition 6.1.
(cid:3)
Corollary 8.5. Let P be an irreducible subfactor planar algebra of depth 2. Then
P is w-cyclic if and only if ϕC(P) is nonzero.
Remark 8.6. The last three results extend to any irreducible subfactor planar
algebra in which any central projection generates a central biprojection.
The rest is only translation. Let G be a finite group, N (G) its normal subgroup
lattice and µN the Mobius function of N (G). Let the dual normal Euler totient be
ϕN (G) := XH∈N (G)
µN (1, H)G : H
Corollary 8.7. A finite group G has a faithful irreducible complex representation
if and only if ϕN (G) is nonzero.
Proof. Apply Corollary 8.5 on P(RG ⊆ R); or Theorem 8.4 with Vi = im(pi) then
(cid:3)
ϕN (G) = XVi faithful
dim(Vi)2.
The above equality can be proved directly from the content of the page 97 in [18].
EULER TOTIENT OF SUBFACTOR PLANAR ALGEBRAS
11
9. Acknowledgments
The author would like to thank the Isaac Newton Institute for Mathematical
Sciences, Cambridge, for support and hospitality during the programme Operator
Algebras: Subfactors and their applications, where work on this paper was under-
taken. This work was supported by EPSRC grant no EP/K032208/1; thanks to
David Evans for the invitation; an anonymous referee for her/his interest in this
work and useful suggestions; Richard Stanley for showing the crosscut theorem;
John Shareshian for Kratzer-Th´evenaz theorem and Eulerian lattices; Benjamin
Steinberg and Frieder Ladisch for P´alfy's formula.
References
[1] Mamta Balodi and Sebastien Palcoux, On Boolean intervals of finite groups, J. Comb. Theory,
Ser. A 157 (2018), 49-69, DOI 10.1016/j.jcta.2018.02.004.
[2] Dietmar Bisch, A note on intermediate subfactors, Pacific J. Math. 163 (1994), no. 2, 201 --
216, DOI 10.2140/pjm.1994.163.201. MR1262294 (95c:46105)
[3] William Burnside, Theory of groups of finite order, 2d ed, Cambridge University Press, 1911.
[4] Hall Philip, The Eulerian functions of a group., Q. J. Math., Oxf. Ser. 7 (1936), 134 -- 151.
[5] Vaughan F. R. Jones, Actions of finite groups on the hyperfinite type II1 factor, Mem. Amer.
Math. Soc. 28 (1980), no. 237, v+70, DOI 10.1090/memo/0237. MR587749 (81m:46094)
, Index for subfactors, Invent. Math. 72 (1983), no. 1, 1 -- 25, DOI 10.1007/BF01389127.
[6]
[7] Vaughan F. R. Jones and V. S. Sunder, Introduction to subfactors, London Mathematical
Society Lecture Note Series, vol. 234, Cambridge University Press, Cambridge, 1997.
[8] Vaughan F. R. Jones, Planar algebras, I, arXiv:math/9909027 (1999), 122pp. to appear in
New Zealand Journal of Mathematics.
[9] Vijay Kodiyalam and V. S. Sunder, On Jones' planar algebras, J. Knot Theory Ramifications
13 (2004), no. 2, 219 -- 247, DOI 10.1142/S021821650400310X. MR2047470 (2005e:46119)
[10] Vijay Kodiyalam, Zeph Landau, and V. S. Sunder, The planar algebra associated to a Kac al-
gebra, Proc. Indian Acad. Sci. Math. Sci. 113 (2003), no. 1, 15 -- 51, DOI 10.1007/BF02829677.
[11] Charles Kratzer and Jacques Th´evenaz, Fonction de Mobius d'un groupe fini et anneau de
Burnside, Commentarii Mathematici Helvetici 59 (1984), no. 1, 425 -- 438.
[12] Zhengwei Liu, Exchange relation planar algebras of small rank, Trans. Amer. Math. Soc. 368
(2016), no. 12, 8303 -- 8348, DOI 10.1090/tran/6582. MR3551573
[13] Masahiro Nakamura and Ziro Takeda, On the fundamental theorem of the Galois theory for
finite factors., Proc. Japan Acad. 36 (1960), 313 -- 318. MR0123926 (23 #A1247)
[14] Sebastien Palcoux, Ore's theorem for cyclic subfactor planar algebras and applications,
arXiv:1505.06649v10, 50pp.
, Ore's theorem for cyclic subfactor planar algebras and beyond, Pacific Journal of
Mathematics 292 (2018), no. 1, 203 -- 221, DOI 10.2140/pjm.2018.292.203.
, Ore's theorem on subfactor planar algebras, 14 pp., available at arXiv:1704.00745v3.
Under review.
, Dual Ore's theorem on distributive intervals of finite groups, 8 pp., available at
[15]
[16]
[17]
arXiv:1708.02565. Accepted in the Journal of Algebra.
[18] P´eter P´al P´alfy, On faithful irreducible representations of finite groups, Studia Sci. Math.
Hungar. 14 (1979), no. 1-3, 95 -- 98.
[19] John Shareshian and Russ Woodroofe, Order complexes of coset posets of finite groups are not
contractible, Advances in Mathematics 291 (2016), 758 -- 773, DOI 10.1016/j.aim.2015.10.018.
[20] Richard P. Stanley, Enumerative combinatorics. Volume 1, 2nd ed., Cambridge Studies in
Advanced Mathematics, vol. 49, Cambridge University Press, Cambridge, 2012. MR2868112
[21] Yasuo Watatani, Lattices of intermediate subfactors, J. Funct. Anal. 140 (1996), no. 2, 312 --
334, DOI 10.1006/jfan.1996.0110. MR1409040 (98c:46134)
[22] Feng Xu, On representing some lattices as lattices of intermediate subfactors of finite index,
Advances in Mathematics 220 (2009), no. 5, 1317 - 1356, DOI 10.1016/j.aim.2008.11.006.
Institute of Mathematical Sciences, Chennai, India
E-mail address: [email protected]
|
0810.0658 | 5 | 0810 | 2010-02-11T14:07:53 | Quantum Isometry groups of the Podles Spheres | [
"math.OA",
"math.QA"
] | For $\mu \in (0,1), c> 0,$ we identify the quantum group $SO_\mu(3)$ as the universal object in the category of compact quantum groups acting by `orientation and volume preserving isometries' in the sense of \cite{goswami2} on the natural spectral triple on the Podles sphere $S^2_{\mu, c}$ constructed by Dabrowski, D'Andrea, Landi and Wagner in \cite{{Dabrowski_et_al}}. | math.OA | math |
Quantum isometry groups of the Podles Spheres
by
Jyotishman Bhowmick 1
and
Debashish Goswami 2
Abstract
For µ ∈ (0, 1), c ≥ 0, we identify the quantum group SOµ(3) as the universal object in the
category of compact quantum groups acting by 'orientation and volume preserving isometries' in
the sense of [4] on the natural spectral triple on the Podles sphere S 2
µ,c constructed by Dabrowski,
D'Andrea, Landi and Wagner in [8].
Mathematics 1991 Subject Classification: Primary 58H05 , Secondary 16W30, 46L87, 46L89.
1
Introduction
In a series of articles initiated by [12] and followed by [3], [4], we have formulated and studied a
quantum group analogue of the group of Riemannian isometries of a classical or noncommutative
manifold. This was motivated by previous work of a number of mathematicians including Wang,
Banica, Bichon and others (see, e.g. [21], [22], [1], [2], [5], [23] and references therein), who have de-
fined quantum automorphism and quantum isometry groups of finite spaces and finite dimensional
algebras. Our theory of quantum isometry groups can be viewed as a natural generalization of such
quantum automorphism or isometry groups of 'finite' or 'discrete' structures to the continuous or
smooth set-up. Clearly, such a generalization is crucial to study the quantum symmetries in non-
commutative geometry, and in particular, for a good understanding of quantum group equivariant
spectral triples.
The group of Riemannian isometries of a compact Riemannian manifold M can be viewed as
the universal object in the category of all compact metrizable groups acting on M , with smooth
and isometric action. Moreover, assume that the manifold has a spin structure (hence in particular
orientable, so we can fix a choice of orientation) and D denotes the conventional Dirac operator
acting as an unbounded self-adjoint operator on the Hilbert space H of square integrable spinors.
Then, it can be proved that the action of a compact group G on the manifold lifts as a unitary
representation (possibly of some group G which is topologically a 2-cover of G, see [7] and [9] for
more details) on the Hilbert space H which commutes with D if and only if the action on the
manifold is an orientation preserving isometric action. Therefore, to define the quantum analogue
of the group of orientation-preserving Riemannian isometry group of a possibly noncommutative
manifold given by a spectral triple (A∞,H, D), it is reasonable to consider a category Q′(D)
of compact quantum groups having unitary (co-) representation, say U , on H, which commutes
with D, and the action on B(H) obtained via conjugation by U maps A∞ into its weak closure.
A universal object in this category, if it exists, should define the 'quantum group of orientation
preserving Riemannian isometries' of the underlying spectral triple. Indeed (see [4]), if we consider a
1The support from National Board of Higher Mathematics, India, is gratefully acknowledged.
2 Partially supported by a project on 'Noncommutative Geometry and Quantum Groups' funded by Indian Na-
tional Science Academy.
1
classical spectral triple, the subcategory of the category Q′(D) consisting of groups has the classical
group of orientation preserving isometries as the universal object, which justifies our definition of
the quantum analogue. Unfortunately, if we consider quantum group actions, even in the finite-
dimensional (but with noncommutative A) situation the category Q′(D) may often fail to have
It turns out, however, that if we fix any suitable faithful functional τR on
a universal object.
B(H) (to be interpreted as the choice of a 'volume form') then there exists a universal object in
the subcategory Q′
R(D) of Q′(D) obtained by restricting the object-class to the quantum group
actions which also preserve the given functional. The subtle point to note here is that unlike the
classical group actions on B(H) which always preserve the usual trace, a quantum group action
may not do so. In fact, it was proved by one of the authors in [11] that given an object (Q, U ) of
Q′(D) (where Q is the compact quantum group and U denotes its unitary co-representation on H),
we can find a suitable functional τR (which typically differs from the usual trace of B(H) and can
have a nontrivial modularity) which is preserved by the action of Q. This makes it quite natural
to work in the setting of twisted spectral data (as defined in [11]). It may also be mentioned that
in [4] we have actually worked in slightly bigger category QR(D) of so called ' quantum family of
orientation and volume preserving isometries ' and deduced that the universal object in QR(D)
exists and coincides with that of Q′
R(D).
It is very important to explicitly compute the (orientation and volume preserving) quantum
group of isometries for as many examples as possible. This programme has been successfully
carried out for a number of spectral triples, including classical spheres and tori as well as their
Rieffel deformations. The aim of the present article is to identify SOµ(3) as the quantum group
of orientation and volume preserving isometries for the spectral triples on the Podles spheres S2
µ,c,
constructed by Dabrowski et al in [8]. Let us mention here that although the quantum groups
SOµ(3) are 'deformations' of the classical SO(3), these are not Rieffel deformations and so the
results and techniques of [4] do not apply.
Our characterization of SOµ(3) as the quantum isometry group of a noncommutative Rieman-
nian manifold generalizes the classical description of the group SO(3) as the group of orientation
preserving isometries of the usual Riemannian structure on the 2-sphere.
It may be mentioned
here that in a very recent article ([20]), P. M. Soltan has characterized SOµ(3) as the universal
compact quantum group acting on the finite dimensional C ∗-algebra M2(C) such that the action
preserves a functional ωµ defined in [20]. In the classical case, we have three equivalent descriptions
of SO(3): (a) as a quotient of SU (2), (b) as the group of (orientation preserving) isometries of S2,
and (c) as the automorphism group of M2. In the quantum case the definition of SOµ(3) is an
analogue of (a), so the characterization of SOµ(3) obtained in this paper as the quantum isometry
group, together with Soltan's characterization, completes the generalization of all three classical
descriptions of SO(3).
2 Notations and preliminaries
2.1 Basics of the theory of compact quantum groups
We begin by recalling the definition of compact quantum groups and their actions from [26], [25].
A compact quantum group (to be abbreviated as CQG from now on) is given by a pair (S, ∆),
where S is a unital separable C ∗ algebra equipped with a unital C ∗-homomorphism ∆ : S → S ⊗S
(where ⊗ denotes the injective tensor product) satisfying
(ai) (∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆ (co-associativity), and
(aii) the linear spans of ∆(S)(S ⊗ 1) and ∆(S)(1 ⊗ S) are norm-dense in S ⊗ S.
2
It is well-known (see [26], [25]) that there is a canonical dense ∗-subalgebra S0 of S, consisting of
the matrix coefficients of the finite dimensional unitary (co)-representations ( to be defined below
) of S, and maps ǫ : S0 → C (co-unit) and κ : S0 → S0 (antipode) defined on S0 which make S0 a
Hopf ∗-algebra.
A CQG (S, ∆) is said to (co)-act on a unital C ∗ algebra B, if there is a unital C ∗-homomorphism
(called an action) α : B → B ⊗ S satisfying the following :
(bi) (α ⊗ id) ◦ α = (id ⊗ ∆) ◦ α, and
(bii) the linear span of α(B)(1 ⊗ S) is norm-dense in B ⊗ S.
For a Hilbert B-module E, (where B is a C ∗ algebra) we shall denote the set of adjointable
B-linear maps on E by L(E). The norm-closure of the linear span of the finite-rank B-linear maps
on E, to be called the set of compact operators on E, will be denoted by K(E). We note that
L(E) = M(K(E)), where M(C) denotes the multiplier algebra of a C ∗-algebra C. We shall also
need the 'leg-numbering' notation: for an operator X in B(H1 ⊗ H2), X(12) and X(13) will denote
the operators X ⊗ IH2 in B(H1 ⊗ H2 ⊗ H2), and Σ23X12Σ23 respectively, where Σ23 is the unitary
on H1 ⊗ H2 ⊗ H2 which flips the two copies of H2.
A unitary ( co ) representation of a CQG (S, ∆) on a Hilbert space H is given by a unitary
element U of M(K(H) ⊗ S) ≡ L(H ⊗ S) satisfying
(id ⊗ ∆)(U ) = U(12)U(13).
Given a unitary representation U we shall denote by αU the ∗-homomorphism αU (X) = U (X ⊗
1S )U ∗ for X belonging to B(H). We shall sometimes identify U with the isometric map from the
Hilbert space H to the Hilbert module H ⊗ S which sends a vector ξ of H to U (ξ ⊗ 1), and may
even denote U (ξ ⊗ 1) by U ξ by a slight abuse of notation. We say that a (possibly unbounded)
operator T on H commutes with U if T ⊗ I (with the natural domain) commutes with U . Such an
operator will also be called U -equivariant or S-equivariant if U is understood.
2.2 The quantum group of orientation preserving Riemannian isometries
We briefly recall the definition of the quantum group of orientation preserving Riemannian isome-
tries for a spectral triple (of compact type) (A∞,H, D) as in [4]. We consider the category
Q′(A∞,H, D) ≡ Q′(D) whose objects (to be called orientation preserving isometries) are the
triplets (S, ∆, U ), where (S, ∆) is a CQG with a unitary representation U in H, satisfying the
following:
(i) U commutes with D,
(ii) for every state ω on S, (id ⊗ ω) ◦ αU (a) belongs to (A∞)′′ for all a in A∞.
The category Q′(D) may not have a universal object in general, as pointed out in [4]. In case
^
QISO+(D), with the corresponding representation
there is a universal object, we shall denote it by
^
QISO+(D) generated by the
U , say, and we denote by QISO+(D) the Woronowicz subalgebra of
elements of the form < ξ ⊗ 1, αU (a)(η ⊗ 1) >, where ξ, η belong to H, a belongs to A∞ and < ·,· >
^
QISO+(D). The quantum group QISO+(D) will
is the
be called the quantum group of orientation-preserving Riemannian isometries of the spectral triple
(A∞,H, D).
QISO+(D)-valued inner product of H ⊗
Although the category Q′(D) may fail to have a universal object, we can always get a universal
object in suitable subcategories which will be described now. Suppose that we are given an invertible
^
3
positive (possibly unbounded) operator R on H which commutes with D. Then we consider the
full subcategory Q′
R(D) of Q′(D) by restricting the object class to those (S, ∆, U ) for which αU
satisfies (τR ⊗ id)(αU (X)) = τR(X)1 for all X in the ∗-subalgebra generated by operators of the
form ξ >< η, where ξ, η are eigenvectors of the operator D which by assumption has discrete
spectrum, and τR(X) = Tr(RX) =< η, Rξ > for X = ξ >< η. We shall call the objects of Q′
R(D)
orientation and (R-twisted) volume preserving isometries. It is clear (see Remark 2.9 in [4]) that
when Re−tD2
is trace-class for some t > 0, the above condition is equivalent to the condition that
αU preserves the bounded normal functional Tr(· Re−tD2
) on the whole of B(H). It is shown in
[4] that the category Q′
R(D), and
the Woronowicz subalgebra generated by {< ξ ⊗ 1, αW (a)(η ⊗ 1) >: ξ, η ∈ H, a ∈ A∞} (where
R(D) and called
W is the unitary representation of
the quantum group of orientation and (R-twisted) volume preserving Riemannian isometries of the
spectral triple.
R(D) in H) will be denoted by QISO+
R(D) always admits a universal object, to be denoted by
^
QISO+
^
QISO+
2.3 SUµ(2) and the Podles spheres
Fix µ in (0, 1). The C ∗ algebra underlying the CQG SUµ(2) is defined as the universal unital
C ∗ algebra generated by α, γ such that α∗α + γ∗γ = 1, αα∗ + µ2γγ∗ = 1, γγ∗ = γ∗γ, µγα =
αγ, µγ∗α = αγ∗.
The CQG structure is given by the following fundamental representation: (cid:18) α −µγ∗
α∗ (cid:19) .
γ
The coproduct is defined by :
∆(α) = α ⊗ α − µγ∗ ⊗ γ,
∆(γ) = γ ⊗ α + α∗ ⊗ γ.
The Haar state of SUµ(2) will denoted by h and the corresponding G.N.S. Hilbert space will be
denoted by L2(SUµ(2)). We will call the unital ∗-subalgebra of SUµ(2) (without any norm-closure)
generated by α, γ the 'co-ordinate Hopf ∗-algebra' of SUµ(2) and denote it by O(SUµ(2)) as in [18].
We now recall the definition of the Podles sphere from [8] (see also the original article [16] by
Podles).
For c ≥ 0, let t in (0, 1] be given by c = t−1 − t. Let [n] ≡ [n]µ = µn−µ−n
The Podles sphere S2
µ,c is defined to be the universal unital C ∗ algebra generated by elements
µ−µ−1 , n ∈ IN .
x−1, x0, x1 satisfying the relations:
x−1(x0 − t) = µ2(x0 − t)x−1,
x1(x0 − t) = µ−2(x0 − t)x1,
−[2]x−1x1 + (µ2x0 + t)(x0 − t) = [2]2(1 − t),
−[2]x1x−1 + (µ−2x0 + t)(x0 − t) = [2]2(1 − t).
µ,c is given by
The involution on S2
x∗
−1 = −µ−1x1, x∗
0 = x0.
4
We note that S2
µ,c as defined above is the same as χq,α′,β in page 124 of [14] with q = µ, α′ =
t, β = t2 + µ−2(µ2 + 1)2(1 − t).
realized as a ∗-subalgebra of SUµ(2) by setting:
Thus, from the expressions of x−1, x0, x1 given in page 125 of [14], it follows that S2
µ,c can be
x−1 =
µα2 + ρ(1 + µ2)αγ − µ2γ2
1
µ(1 + µ2)
2
,
x0 = −µγ∗α + ρ(1 − (1 + µ2)γ∗γ) − γα∗,
x1 =
µ2γ∗2 − ρµ(1 + µ2)α∗γ∗ − µα∗2
,
1
(1 + µ2)
2
(1)
(2)
(3)
where ρ2 =
Taking
µ2t2
(µ2+1)2(1−t) .
A =
1 − t−1x0
1 + µ2
, B = µ(1 + µ2)− 1
2 t−1x−1,
one obtains ( see [8] ) that the C ∗ algebra S2
µ,c coincides with the original description given in
[16], i.e, the universal C ∗ algebra generated by elements A and B satisfying the relations:
A∗ = A, AB = µ−2BA,
B∗B = A − A2 + cI, BB∗ = µ2A − µ4A2 + cI.
We will denote by O(S2
by A, B.
µ,c) the co-ordinate ∗-algebra of S2
µ,c, i.e. the unital ∗-subalgebra generated
We recall from [18] the Hopf ∗-algebra Uµ(su(2)) which is generated by elements F, E, K, K −1
with defining relations :
KK −1 = K −1K = 1, KE = µEK, F K = µKF, EF − F E = (µ − µ−1)−1(K 2 − K −2)
with involution E∗ = F, K ∗ = K and comultiplication :
∆(E) = E ⊗ K + K −1 ⊗ E, ∆(F ) = F ⊗ K + K −1 ⊗ F, ∆(K) = K ⊗ K.
The counit is given by ǫ(E) = ǫ(F ) = ǫ(K − 1) = 0 and antipode S(K) = K −1, S(E) =
−µE, S(F ) = −µ−1F.
pairing among the generators are given below:
There is a dual pairing h., .i of Uµ(su(2)) and O(SUµ(2)), for which the nonzero values of the
(cid:10)K ±1, α∗(cid:11) = (cid:10)K ∓1, α(cid:11) = µ± 1
The left action ⊲ and right action ⊳ of Uµ(su(2)) on SUµ(2) are given by:
f ⊲ x = (cid:10)f, x(2)(cid:11) x(1), x ⊳ f = (cid:10)f, x(1)(cid:11) x(2), x ∈ O(SUµ(2)), f ∈ Uµ(su(2)) where we have used
2 , hE, γi = hF,−µγ∗i = 1.
the Sweedler notation ∆(x) = x(1) ⊗ x(2).
The actions satisfy the following :
(f ⊲ x)∗ = S(f )∗ ⊲x∗, (x ⊳ f )∗ = x∗ ⊳S(f )∗, f ⊲xy = (f(1) ⊲x)(f(2) ⊲y), xy ⊳f = (x⊳f(1))(y ⊳f(2)).
The action on generators is given by :
5
0, K ⊲ α = µ− 1
E ⊲ α = −µγ∗, E ⊲ γ = α∗, E ⊲ γ∗ = E ⊲ α∗ = 0, F ⊲ (−µγ∗) = α, F ⊲ α∗ = γ, F ⊲ α = F ⊲ γ =
γ ⊳ E = α, α∗ ⊳ E = −µγ∗, α ⊳ E = γ∗ ⊳ E = 0, α ⊳ F = γ, − µγ∗ ⊳ F = α∗, γ ⊳ F = α∗ ⊳ F =
2 γ∗, K ⊲ γ = µ− 1
2 α, K ⊲ (γ∗) = µ
2 γ, K ⊲ α∗ = µ
2 α∗.
1
1
1
1
0, α ⊳ K = µ− 1
2 α, γ∗ ⊳ K = µ− 1
2 γ∗, γ ⊳ K = µ
2 γ, α∗ ⊳ K = µ
2 α∗.
We recall an alternative description of S2
Let
Xc = µ
1
2 (µ−1 − µ)
−1
µ,c from [19] which we are going to need.
c− 1
2 (1 − K 2) + EK + µF K, c > 0,
X0 = 1 − K 2.
One has ∆(Xc) = 1 ⊗ Xc + Xc ⊗ K 2. Moreover, we have the following ([19], page 9):
Theorem 2.1 We have,
O(S2
A basis of the vector space O(S2
Thus, any element of O(S2
Ak, AkBl, AkB∗l.
µ,c) = {x ∈ O(SUµ(2)) : x ⊳ Xc = 0}.
µ,c) is given by {Ak, AkBl, AkB∗m, k ≥ 0, l, m > 0}.
µ,c) can be written as a finite linear combination of elements of the form
Let ψ be the densely defined linear map on L2(SUµ(2)) defined by ψ(x) = x ⊳ Xc.
Lemma 2.2 The map ψ is closable and we have S2
of ψ and S2
µ,c denotes the Hilbert subspace generated by S2
µ,c ⊆ Ker(ψ) where ψ is the closed extension
µ,c) =
µ,c in L2(SUµ(2)). Moreover, O(S2
O(SUµ(2))T Ker(ψ) = O(SUµ(2))T Ker(ψ).
closable, hence Ker(ψ) is closed. The lemma now follows from the observation that O(S2
Ker(ψ) ⊆ Ker(ψ). ✷
Proof : From the expression of Xc, it is clear that O(SUµ(2)) ⊆ Dom(ψ∗) implying that ψ is
µ,c) =
We end this subsection with a discussion on the CQG SOµ(3) as described in [17].
It is the universal unital C ∗ algebra generated by elements M, N, G, C, L satisfying :
L∗L = (I − N )(I − µ−2N ), LL∗ = (I − µ2N )(I − µ4N ), G∗G = GG∗ = N 2, M ∗M = N −
N 2, M M ∗ = µ2N − µ4N 2, C ∗C = N − N 2, CC ∗ = µ2N − µ4N 2, LN = µ4N L, GN =
N G, M N = µ2N M, CN = µ2N C, LG = µ4GL, LM = µ2M L, M G = µ2GM, CM =
M C, LG∗ = µ4G∗L, M 2 = µ−1LG, M ∗L = µ−1(I − N )C, N ∗ = N.
This CQG can be identified with a Woronowicz subalgebra of SUµ(2) by taking:
N = γ∗γ, M = αγ, C = αγ∗, G = γ2, L = α2.
The canonical action of SUµ(2) on S2
µ,c, i.e. the action obtained by restricting the coproduct of
µ,c, is actually a faithful action of SOµ(3). On the subspace spanned
SUµ(2) to the subalgebra S2
by {x−1, x0, x1} this action is given by the following SOµ(3)-valued 3 × 3-matrix:
Z ′ :=
µ2γ∗2
2 αγ I − µ(µ + µ−1)γ∗γ −µ(1 + µ−2)
α∗2
−µ(1 + µ−2)
(1 + µ−2)
(1 + µ−2)
2 γ∗α∗
2 αγ∗
2 γα∗
α2
γ2
.
1
1
1
1
6
3
Spectral triples on the Podles spheres and their quantum isom-
etry groups
3.1 Description of the spectral triples
We now recall the spectral triples on S2
µ,c discussed in [8] (see also [18] for the case c = 0).
1
2 .
2 ± (c + 1
4 )
2 λ−, λ± = 1
2 IN ,
Let s = −c− 1
For all j in 1
uj = (α∗ − sγ∗)(α∗ − µ−1sγ∗)......(α∗ − µ−2j+1sγ∗),
wj = (α − µsγ)(α − µ2sγ)........(α − µ2j sγ),
u−j = E2j ⊲ wj ,
u0 = w0 = 1,
y1 = (1 + µ−2)
N l
2 µ2γ∗2 − µγ∗α∗ − µc
l−juj)(cid:13)(cid:13)
.
k,jF l−k ⊲ (yl−j
2 α∗2),
2 (c
−1
1
1
1
N ,
Set
N + 1, ........, m = −l, .......l}.
kj = (cid:13)(cid:13)F l−k ⊲ (y1
Define vl
Let MN be the Hilbert subspace of L2(SUµ(2)) with the orthonormal basis {vl
2 IN 0, j, k = −l,−l + 1, ......l.
uj), l ∈ 1
k,j = N l
1
m,N : l =
and define a representation π of S2
µ,c on H by
H = M− 1
2 ⊕ M 1
2
,
π(xi)vl
m,N = α−
i (l, m; N )vl−1
m+i,N + α0
i (l, m; N )vl
m+i,N + α+
i (l, m; N )vl+1
m+i,N ,
where α−
i , α0
i , α+
i are as defined in [8].
µ,c) with S2
We will often identify π(S2
Finally by Proposition 7.2 of [8], the following Dirac operator D gives a spectral triple (O(S2
µ,c.
µ,c),H, D)
which we are going to work with :
D(vl
m,± 1
2
) = (c1l + c2)vl
m,∓ 1
2
where c1, c2 belong to R, c1 6= 0.
a unitary representation, say U0 on H.
It is easy to see that the action of SUµ(2) on itself keeps the subspace H invariant and so induces
We define a positive, unbounded operator R on H by R(vn
) = µ−2ivn
i,± 1
i,± 1
.
2
2
Proposition 3.1 αU0 preserves the R-twisted volume. In particular, for x belonging to π(S2
and t > 0, we have h(x) = τR(x)
Haar state of SUµ(2) to the subalgebra S2
µ,c)
), and h denotes the restriction of the
µ,c, which is the unique SUµ(2)-invariant state on S2
τR(1) , where τR(x) := Tr(xRe−tD2
µ,c.
Proof : It is enough to prove that τR is αU0-invariant. Define R0(vn
that it has been observed in [13] that Tr(R0e−tD2
) = µ−2i∓1vn
) < ∞ ( for all t > 0 ) and one has
i,± 1
2
i,± 1
2
, and note
(τR0 ⊗ id)(U0(x ⊗ 1)U0
∗) = τR0(x).1,
7
for all x in B(H), where τR0(x) = Tr(xR0e−tD2
).
Let us denote by P 1
2} respectively. Moreover, let τ± be the functionals defined by τ±(x) = Tr(xR0P± 1
the projections onto the closed subspaces generated by {vl
i, 1
e−tD2
, P− 1
2
2
2
2} and
). We
i,− 1
{vl
observe that R0, e−tD2
and U0 commute with P± 1
2
(τ± ⊗ id)(αU0(x)) = (τR0 ⊗ id)(αU0(xP± 1
i.e. τ± are αU0-invariant. Moreover, since we have RP± 1
with µ−1τ+ + µτ−, hence is αU0-invariant. ✷
2
2
so that for x belonging to B(H),
)1 = τ±(x)1,
)) = τR0(xP± 1
2
= µ±R0P± 1
2
, the functional τR coincides
Theorem 3.2 (SUµ(2), ∆, U0) is an object in Q′
R(D).
Proof : The above spectral triple is equivariant with respect to this representation (see [8]) and
it preserves τR by Proposition 3.1, which completes the proof. ✷
We now note down some useful facts for later use.
Remark 3.3 Using the definition of vl
i,j and ⊲, we observe :
1. The eigenspace of D corresponding to (c1l + c2) and −(c1l + c2) are span{vl
m, 1
2
+ vl
m,− 1
2
: −l ≤
2 + c2) is span{α, γ, α∗, γ∗}.
m ≤ l} and span{vl
m, 1
2 − vl
m,− 1
2
: −l ≤ m ≤ l} respectively.
2. The eigenspace of D corresponding to the eigenvalue (c1. 1
m,N},
m,N belongs to Span{vl−1
m,N , vl+1
Remark 3.4 1. π(A)vl
π(B)vl
π(B∗)vl
m,N belongs to Span{vl−1
m,N belongs to Span{vl−1
m−1,N , vl
m+1,N , vl
3.π(Am′
2.π(Ak)(vl
m,N ) belongs to Span{vl−k
Bn′
)(vl
m,N , vl−k+1
m,N ) belongs to Span{vl−m′−n′
m,N ) belongs to Span{vl−s−r
We shall now proceed to show that QISO+
4. π(ArB∗s)(vl
m,N , vl
m−1,N , vl+1
m+1,N , vl+1
m−1,N},
m+1,N}.
m,N , ......., vl+k
m,N}.
m−n′,N , vl−(n′+m′−1)
m+s,N , vl−s−r+1
m−n′,N
, ......., vl+n′+m′
m−n′,N }.
m+s,N , .......vl+s+r
m+s,N}.
R(D) is isomorphic with SOµ(3). Let ( Q, U ) be an
object in the category Q′
R(D) of CQG s acting by orientation and R-twisted volume preserving
isometries on this spectral triple and Q be the Woronowicz C ∗ subalgebra of Q generated by
µ,c (where < ·,· > Q is the Q valued inner product
< (ξ ⊗ 1), αU (a)(η ⊗ 1) > Q, for ξ, η in H, a in S2
of H ⊗ Q). We shall denote αU by φ from now on.
The proof has two main steps: first, we prove that φ is 'linear', in the sense that it keeps the
span of {1, A, B, B∗} invariant, and then we shall exploit the facts that φ is a ∗-homomorphism
and preserves the canonical volume form on S2
µ,c, i.e. the restriction of the Haar state of SUµ(2).
Remark 3.5 The first step does not make use of the fact that φ preserves the R-twisted volume,
so linearity of the action follows for any object in the bigger category Q′(D).
8
3.2 Linearity of the action
For a vector v in H, we shall denote by Tv the map from B(H) to L2(SUµ(2)) given by Tv(x) =
xv ∈ H ⊂ L2(SUµ(2)). It is clearly a continuous map with respect to the strong operator topology
on B(H) and the Hilbert space topology of L2(SUµ(2)).
For an element a in SUµ(2), we consider the right multiplication Ra as a bounded linear map on
L2(SUµ(2)). Clearly the composition RaTv is a continuous linear map from B(H) (with the strong
operator topology) to the Hilbert space L2(SUµ(2)). We now define
T = Rα∗ Tα + µ2RγTγ ∗ .
µ,c, we have T (φω(x)) = φω(x) ≡ R1(φω(x)) belonging
Lemma 3.6 For any state ω on Q and x in S2
to S2
µ,c ⊆ L2(SUµ(2)), where φω(x) = (id ⊗ ω)(φ(x)).
Proof : It is clear from the definition of T (using αα∗ + µ2γγ∗ = 1) that T (x) = x ≡ R1(x) for
µ,c ⊂ B(H), where x in the right hand side of the above denotes the identification of x ∈ S2
µ,c
µ,c, φω(x) belongs to
µ,c in the strong operator topology, and the continuity of T in this
x in S2
as a vector in L2(SUµ(2)). Now, the lemma follows by noting that for x in S2
(S2
topology discussed before. ✷
µ,c)′′, which is the closure of S2
Let
V l = Span{vl′
i,± 1
2
,−l′ ≤ i ≤ l′, l′ ≤ l}.
2
i,± 1
,−l ≤ i ≤ l} is the eigenspace of D corresponding to the eigenvalue c1l + c2, U
Since Span{vl
and U ∗ must keep V l invariant for all l.
Lemma 3.7 There is some finite dimensional subspace V of O(SUµ(2)) such that Rα∗(φω(A)v
j,± 1
1
2
2
),
1
2
) belong to V for all states ω on Q.
Rγ(φω(A)v
The same holds when A is replaced by B or B∗.
j,± 1
2
Proof : We prove the result for A only, since a similar argument will work for B and B∗.
We have φ(A)(v
1
2
j,± 1
2 ⊗ 1) = U (π(A) ⊗ 1)U ∗(v
j,± 1
2 ⊗ 1).
1
2
1
3
2
j,± 1
Now, U ∗(v
2 ⊗ 1) belongs to V
3.4, we observe that (π(A) ⊗ 1)U ∗(v
2 ⊗ Q. Again, U keeps V
V
Similarly, Rγ(φω(A)(v
V = Span{vα∗, vγ : v ∈ V
± 1
2
2
1
3
1
2
1
j,± 1
2 ⊗ 1) belongs to Span{vl′
2 ⊗ Q, and then using the definition of π as well as the Remark
2} ⊗ Q =
2}.
2}. So, the lemma follows for A by taking
) belongs to Span{vα∗ : v ∈ V
: −l′ ≤ j ≤ l′, l′ ≤ 3
j,± 1
± 1
3
3
2
1
2
2
2 ⊗ Q invariant, so Rα∗(φω(A)v
)) belongs to Span{vγ : v ∈ V
2} ⊂ O(SUµ(2)). ✷
3
Since α, γ∗ belong to Span{v
1
2
j,± 1
2}, we have the following immediate corollary:
Corollary 3.8 There is a finite dimensional subspace V of O(SUµ(2)) such that for every state
(hence for every bounded linear functional) ω on Q, we have T (φω(A)) belongs to V. A similar
conclusion holds for B and B∗ as well.
9
Proposition 3.9 φ(A), φ(B), φ(B∗) belong to O(S2
µ,c) ⊗alg Q.
µ,c = V TO(S2
µ,c ⊂ O(SUµ(2))T Ker(ψ) and hence V T S2
Proof : We give the proof for φ(A) only, the proof for B, B∗ being similar. From the Corollary
3.8 and Lemma 3.6 it follows that for every bounded linear functional ω on Q, T (φω(A)) belongs
to V T S2
µ,c), where V is the finite dimen-
sional subspace mentioned in Corollary 3.8. Clearly, V TO(S2
µ,c) is a finite dimensional subspace
of O(S2
µ,c) implying that there must be finite m, say, such that for every ω, T (φω(A)) belongs
to Span{Ak, AkBl, AkB∗l : 0 ≤ k, l ≤ m}. Denote by W the (finite dimensional) subspace of
B(H) spanned by {Ak, AkBl, AkB∗l : 0 ≤ k, l ≤ m}. Since for every state (and hence for every
bounded linear functional) ω on Q, we have T (φω(A)) = R1(φω(A)) ≡ φω(A)1, it is clear that
φω(A) belongs to W for every ω in Q∗. Now, let us fix any faithful state ω on the separable uni-
tal C ∗-algebra Q and embed Q in B(L2(Q, ω)) ≡ B(K). Thus, we get a canonical embedding of
L(H ⊗ Q) in B(H ⊗ K). Let us thus identify φ(A) as an element of B(H ⊗ K), and then by choos-
ing a countable family of elements {q1, q2, ...} of Q which is an orthonormal basis in K = L2(ω),
we can write φ(A) as a weakly convergent series of the form P∞
i,j=1 φij(A) ⊗ qi >< qj. But
φij(A) = (id ⊗ ωij)(φ(A)), where ωij(·) = ω(q∗
i · qj). Thus, φij(A) belongs to W for all i, j, and
hence the sequence Pn
i,j=1 φij(A) ⊗ qi >< qj ∈ W ⊗ B(K) converges weakly, and W being finite
dimensional (hence weakly closed), the limit, i.e. φ(A), must belong to W ⊗B(K). In other words,
if A1, ..., Ak denotes a basis of W, we can write φ(A) = Pk
We claim that each Bi must belong to Q. For any trace-class positive operator ρ in H, say of the
form ρ = Pj λjej >< ej, where {e1, e2, , ...} is an orthonormal basis of H and λj ≥ 0,Pj λj < ∞,
let us denote by ψρ the normal functional on B(H) given by x 7→ Tr(ρx), and it is easy to see
that it has a canonical extension ψρ := (ψρ ⊗ id) on L(H ⊗ Q) given by ψρ(X) = Pj λj <
ej ⊗ 1, X(ej ⊗ 1) > Q, where X belongs to L(H ⊗ Q) and < ·,·, > Q denotes the Q-valued inner
product of H ⊗ Q. Clearly, ψρ is a bounded linear map from L(H ⊗ Q) to Q. Now, since A1, ..., Ak
in the expression of φ(A) are linearly independent, we can choose trace class operators ρ1, ..., ρk
such that ψρi(Ai) = 1 and ψρi(Aj) = 0 for j 6= i. Then, by applying ψρi on φ(A) we conclude
that Bi belongs to Q. But by definition, Q is the Woronowicz subalgebra of Q generated by
< ξ ⊗ 1, φ(x)(η ⊗ 1) > Q, with η, ξ belonging to H and x in O(S2
µ,c), and hence it follows that Bi
belongs to Q. ✷
Proposition 3.10 φ keeps the span of 1, A, B, B∗ invariant.
i=1 Ai ⊗ Bi for some Bi ∈ B(K).
Proof : We prove the result for φ(A) only, the proof for the other cases being quite similar.
Using Proposition 3.9, we can write φ(A) as a finite sum of the form :
r,s.
Pk≥0 Ak ⊗ Qk +Pm′,n′,n′6=0 Am′
Let ξ = vl
We have that U (ξ) belongs to Span{vl
⊗ Rm′,n′ +Pr,s,s6=0 ArB∗s ⊗ R′
m,N , m = −l, ......l, N = ± 1
2}. Let us write
Bn′
m0,N0
.
U (ξ ⊗ 1) = Xm=−l,....l,N =± 1
2
vl
m,N ⊗ ql
(m,N ),(m0,N0),
where ql
(m,N ),(m0,N0) belong to Q. Since αU preserves the R-twisted volume, we have :
(m,N ), (m′,N ′)ql∗
ql
(m,N ), (m′,N ′) = 1.
Xm′,N ′
(4)
10
It also follows that U (Aξ) belongs to Span{vl′
m,N , m = −l′, ........l′, l′ = l − 1, l, l + 1, N = ± 1
2}.
Akvl
m,N ⊗ Qkql
2
(m,N ), (m0,N0) +
Recalling Remark 3.4, we have φ(A)U (ξ⊗ 1) = Pk, m=−l,....l,N =± 1
Pm′, n′,n′6=0, m=−l,....l,N =± 1
+Pr,s, s6=0, m=−l,...l, N =± 1
m,N ⊗ Rm′, n′ql
vl
m,N ⊗ R′
r,sql
(m,N ), (m0,N0)
(m,N ), (m0,N0).
ArB∗svl
Am′
Bn′
2
2
Let m′
the expression of φ(A). We claim that the coefficient of v
0 denote the largest integer m′ such that there is a nonzero coefficient of Am′
m−n′,N in φ(A)U (ξ⊗1) is Rm′
0−n′
l−m′
Bn′
0,n′ql
, n′ ≥ 1 in
(m,N ), (m0,N0).
0−n′
l−m′
m−n′,N can arise in three ways:
it can come from a term of the form
Am′′
Indeed, the term v
Bn′′
In the first case, we must have l − m′
m,N or ArB∗svl
vl
m,N or Akvl
implying m′′ = m′
0 + t, and since m′
only have the possibility t = 0, i.e. v
m.N for some m′′, n′′, k, r, s.
0 − n′ = l − m′′ − n′′ + t, 0 ≤ t ≤ 2m′′ and m− n′ = m− n′′
appears in φ(A), we
0 Bn′
0 is the largest integer such that Am′
l−m′
m−n′,N appears only in Am′
0 Bn′
0−n′
.
In the second case, we have m − n′ = m implying n′ = 0 - a contradiction. In the last case,
we have m − n′ = m + s so that −n′ = s which is only possible when n′ = s = 0 which is again a
contradiction.
It now follows from the above claim, using Remark 3.4 and comparing coefficients in the equality
U (Aξ ⊗ 1) = φ(A)U (ξ ⊗ 1), that Rm′
0 ≥ 1.
Now varying (m0, N0), we conclude that the above holds for all ( m0, N0 ). Using (4), we conclude
that
(m,N ), (m0,N0) = 0 for all n′ ≥ 1, for all m, N when m′
0,n′ql
0, we deduce Rm′,n′ = 0
Rm′
that is, Rm′
0,n′ Pm′,N ′ ql
for all m′ ≥ 1, n′ ≥ 1.
(m,N ), (m′,N ′)ql∗
(m,N ), (m′,N ′) = 0 for all n′ ≥ 1,
0,n′ = 0 for all n′ ≥ 1 if m′
0 ≥ 1. Proceeding by induction on m′
Similarly, we have Qk = 0 for all k ≥ 2 and R′
Thus, φ(A) belongs to Span{1, A, B, B∗, B2, ..., Bn, B∗2, ..., B∗m}. But the coefficient of vl−n′
r,s = 0 for all r ≥ 1, s ≥ 1.
in φ(A)U (ξ ⊗ 1) is R0,n′ . Arguing as before, we conclude that R0,n′ = 0 for all n′ ≥ 2. In a similar
way, we can prove R′
m−n′,N
0,n′ = 0 for all n′ ≥ 2. ✷
In view of the above, let us write:
φ(A) = 1 ⊗ T1 + A ⊗ T2 + B ⊗ T3 + B∗ ⊗ T4,
φ(B) = 1 ⊗ S1 + A ⊗ S2 + B ⊗ S3 + B∗ ⊗ S4,
(5)
(6)
for some Ti, Si in Q.
3.3
Identification of SOµ(3) as the quantum isometry group
In this subsection, we shall use the facts that φ is a ∗-homomorphism and it preserves the R-twisted
volume to derive relations among Ti, Si in ( 5 ), ( 6 ).
Lemma 3.11
T1 =
1 − T2
1 + µ2 ,
S1 = −S2
1 + µ2 .
11
Proof : We have the expressions of A and B in terms of the SUµ(2) elements from the equations
( 1 ), ( 2 ) and ( 3 ). From these, we note that h(A) = (1 + µ2)−1 and h(B) = 0. By recalling
Proposition 3.1, we use (h ⊗ id)φ(A) = h(A).1 and (h ⊗ id)φ(B) = h(B).1 to have the above two
equations. ✷
Lemma 3.12 T ∗
1 = T1, T ∗
2 = T2, T ∗
4 = T3.
Proof :
It follows by comparing the coefficients of 1, A and B respectively in the equation
φ(A∗) = φ(A). ✷
Lemma 3.13
S∗
2 S2 + c(1 + µ2)2S∗
3 S3 + c(1 + µ2)
2
S∗
4 S4
= (1 − T2)(µ2 + T2) − c(1 + µ2)2T3T ∗
3 − c(1 + µ2)
2
T ∗
3 T3 + c(1 + µ2)2.1,
(7)
− 2S∗
2 S2 + (1 + µ2)S∗
3 S3 + µ2(1 + µ2)S∗
4 S4 = (µ2 + 2T2 − 1)T2 − µ2(1 + µ2)T3T ∗
3 − (1 + µ2)T ∗
3 T3, (8)
S∗
2 S2 − S∗
S∗
2 S4 + S∗
3 S3 − µ4S∗
4 S4 = −T 2
3 S2 = −(µ2 + T2)T ∗
3 + T ∗
2 + µ4T3T ∗
3 T3,
3 + T ∗
3 (1 − T2),
2 S3 + µ2S∗
4 S2 = −T2T3 − µ2T3T2,
S∗
4 S3 = −T 2
S∗
3 .
(9)
(10)
(11)
(12)
Proof : It follows by comparing the coefficients of 1, A, A2, B∗, AB and B2 in the equation
φ(B∗B) = φ(A) − φ(A2) + c φ(I) and then using Lemma 3.11 and Lemma 3.12. ✷
Lemma 3.14
−S2(1 − T2) + c(1 + µ2)
2
S3T ∗
2
3 + c(1 + µ2)
S4T3
= −µ2(1 − T2)S2 + cµ2(1 + µ2)
2
T3S4 + cµ2(1 + µ2)2T ∗
3 S3,
(13)
S2 − 2S2T2 + (1 + µ2)(µ2S3T ∗
3 + S4T3) = µ2S2 − 2µ2T2S2 + µ4(1 + µ2)T3S4 + µ2(1 + µ2)T ∗
3 S3, (14)
− S2T3 + S3(1 − T2) = −µ2T3S2 + µ2(1 − T2)S3,
− S2T ∗
3 S2,
3 + S4(1 − T2) = µ2(1 − T2)S4 − µ2T ∗
S2T3 + µ2S3T2 = µ2(T2S3 + µ2T3S2),
S3T3 = µ2T3S3,
S4T ∗
3 = µ2T ∗
3 S4.
(15)
(16)
(17)
(18)
(19)
Proof : It follows by equating the coefficients of 1, A, B, B∗, AB, B2 and B∗2 in the equation
φ(BA) = µ2φ(AB) and then using Lemma 3.11 and Lemma 3.12. ✷
12
Lemma 3.15
− S2S∗
4 − S3S∗
2 = µ2(1 + µ2)T3 − µ4(1 − T2)T3 − µ4T3(1 − T2),
S2S∗
4 + µ2S3S∗
S3S∗
2 = −µ4T2T3 − µ6T3T2,
4 = −µ4T 2
3 .
(20)
(21)
(22)
Proof : The Lemma is proved by equating the coefficient of B, AB, B2 in the equation φ(BB∗) =
µ2φ(A) − µ4φ(A2) + c φ(I) and then using Lemma 3.11 and Lemma 3.12. ✷
Now, we compute the antipode, say κ of Q.
To begin with, we note that {x−1, x0, x1} is a set of orthogonal vectors. Moreover, they have
the same norm. The first assertion being easier, we prove below the second one.
Lemma 3.16 h(x∗
t(−µ4 − 2µ2 − 1)].
−1x−1) = h(x∗
0x0) = h(x∗
1x1) = t2(1 − µ2)(1 − µ6)−1[µ2 + t−1(µ4 + 2µ2 + 1) +
Proof : We have x∗
−1x−1 = t2µ−2(1+µ2)(A−A2+cI), x∗
0x0 = t2(1−2(1+µ2)A+(1 + µ2)2
A2), x∗
1x1 =
t2(1 + µ2)(µ2A − µ4A2 + cI).
We recall from [19] that for all bounded Borel function f on σ(A),
h(f (A)) = γ+
∞
Xn=0
f (λ+µ2n)µ2n + γ−
∞
Xn=0
f (λ−µ2n)µ2n,
where λ+ = 1
µ2)λ−(λ− − λ+)−1.
2 + (c + 1
4 )
2 , λ− = 1
1
2 − (c + 1
4 )
1
2 , γ+ = (1 − µ2)λ+(λ+ − λ−)−1, γ− = (1 −
The Lemma follows by applying this relation to the above expressions of x∗
−1x−1, x∗
0x0, x∗
1x1.
✷
If x′
−1, x′
0, x′
) along with the fact that each of the vectors x−1, x0, x1 has the same norm, it follows that
1 is the normalized basis corresponding to {x−1, x0, x1}, then from ( 5 ) and ( 6
2 S2 + x′
1
0 ⊗ −µ−1(1 + µ2)− 1
0 ⊗ T2 + x′
2 + x′
2 T3 + x′
0 ⊗ (1 + µ2)− 1
2 S∗
1 ⊗ (1 + µ2)
1 ⊗ S∗
3 .
1 ⊗ −µ−1S4,
2 T4,
1
φ(x′
φ(x′
−1) = x′
0) = x′
1) = x′
−1 ⊗ S3 + x′
−1 ⊗ −µ(1 + µ2)
4 + x′
−1 ⊗ −µS∗
0, x′
0, x′
φ(x′
Since φ is kept invariant by the Haar state h of SUµ(2) and φ keeps the span of the orthonormal
1} invariant too, we get a unitary representation of the CQG Q on the span
1}. If we denote by Z the M3(Q)-valued unitary corresponding to this unitary
3 from
set {x′
of {x′
representation with respect to the ordered basis {x′
Lemma 3.12 the following:
1}, we get by using T4 = T ∗
−1, x′
−1, x′
−1, x′
0, x′
S3
−S2
−µp1 + µ2T3 −µS∗
2√1+µ2
S∗
3
T2
S ∗
4
Z =
µ√1+µ2
−µ−1S4 p1 + µ2T ∗
unitary representation U α ≡ (uα
3
.
Recall that ( see, for example, [15] ), the antipode κ on the matrix elements of a finite-dimensional
qp)∗. Thus, the antipode κ is given by :
pq) is given by κ(uα
pq) = (uα
13
κ(T2) = T2, κ(T3) =
S ∗
2
µ2(1+µ2) , κ(S2) = µ2(1 + µ2)T ∗
3 , κ(S3) = S∗
3 , κ(S4) = µ2S4, κ(T ∗
3 ) =
S2
1+µ2 , κ(S∗
2 ) = (1 + µ2)T3, κ(S∗
3 ) = S3, κ(S∗
4 ) = µ−2S∗
4 .
Now we derive some more relations by applying the anti-homomorphism κ on the relations
obtained earlier.
Lemma 3.17
−2µ4(1 + µ2)
3
2
T ∗
3 T3 + µ2(1 + µ2)
= µ2(1 + µ2)T2(µ2 + 2T2 − 1) − µ2S2S∗
2
3 S3 + µ4(1 + µ2)
S∗
2 − S∗
2 S2,
S4S∗
4
4
µ4(1 + µ2)
2
2
2
S4S∗
4
S∗
3 S3 − µ6(1 + µ2)
2 S2,
T ∗
3 T3 − µ2(1 + µ2)
2 + µ4S2S∗
T 2
= −µ2(1 + µ2)
T ∗
S4T3 + µ2(1 + µ2)
3 S3 = −S2(µ2 + T2) + (1 − T2)S2,
µ2(1 + µ2)2 .
S4S3 = −
2 + S∗
−S2
2
2
µ2(1 + µ2)
2
Proof : The relations follow by applying κ on ( 8 ), ( 9 ), ( 10 ) and ( 12 ) respectively. ✷
Lemma 3.18
− µ2(1 − T2)T ∗
3 + cS2S∗
3 + cS∗
3 (1 − T2) + cµ2S4S∗
2 + cµ2S∗
3 S2,
2 S4 = −µ4T ∗
S3S2 = µ2S2S3,
S2S4 = µ2S4S2,
3 S∗
3 = −µ2T ∗
− S∗
2 T ∗
− S2T ∗
3 + (1 − T2)S∗
3 + (1 − T2)S4 = µ2S4(1 − T2) − µ2T ∗
3 (1 − T2),
3 S2.
2 + µ2S∗
(23)
(24)
(25)
(26)
(27)
(28)
(29)
(30)
(31)
(32)
(33)
(34)
Proof : The relations follow by applying κ on ( 13 ), ( 18 ), ( 19 ), ( 15 ) and ( 16 ) respectively.
✷
Lemma 3.19
− µ2(1 + µ2)
2
µ2S2
2
S3S4 = −
(1 + µ2)2 ,
3 = µ2(1 + µ2)S∗
T3S∗
2
2
3 + µ2(1 + µ2)
T3S∗
4 T ∗
S∗
3 − µ2(1 + µ2)
2
S∗
4 T ∗
(1 + µ2)
2 − µ4S∗
3 = −µ2S∗
2(1 − T2) − µ4(1 − T2)S∗
2 ,
2 T2 − µ4T2S∗
2 .
Proof : The relations follow by applying κ on ( 22 ), ( 20 ) and ( 21 ) respectively. ✷
Remark 3.20 It follows from ( 26 ) and ( 32 ) that µ4S4S3 = S3S4.
Lemma 3.21 S∗
2 S2 = (1 − T2)(µ2 + T2).
14
Proof : Subtracting the equation obtained by multiplying c(1 + µ2) with ( 8 ) from ( 7 ), we
have
(1 + 2c(1 + µ2))S∗
2
2 S2 + c(1 + µ2)
(1 − µ2)S∗
4 S4
= (1 − T2)(µ2 + T2) − c(1 + µ2)(µ2 + 2T2 − 1)T2 + c(1 + µ2)
Again, by adding ( 7 ) with c(1 + µ2)2 times ( 9 ) gives
2
(µ2 − 1)T3T ∗
2
3 + c(1 + µ2)
.1.
(35)
2
(1 + c(1 + µ2)
)S∗
2 S2 + c(1 − µ4)(1 + µ2)
2
S∗
4 S4
= (1 − T2)(µ2 + T2) − c(1 + µ2)
2
2
T 2
2 + c(1 + µ2)
(µ4 − 1)T3T ∗
3 + c(1 + µ2)
2
.1.
(36)
Subtracting the equation obtained by multiplying (µ2 + 1) with ( 35 ) from ( 36 ) we obtain
−(µ2 + c(1 + µ2)
2
)S∗
2 S2 = (1 − T2)(µ2 + T2) − c(1 + µ2)
2
T 2
2
−(1 + µ2)(1 − T2)(µ2 + T2) − cµ2(1 + µ2)
2
2
.1 + c(1 + µ2)
(µ2 + 2T2 − 1)T2.
The right hand side can be seen to equal −(µ2 + c(1 + µ2)2)(1 − T2)(µ2 + T2).
Thus, S∗
2 S2 = (1 − T2)(µ2 + T2). ✷
Lemma 3.22
2
µ2(1 + µ2)
2
(1 + µ2)
S2S∗
3 T3 = (1 − T2)(µ2 + T2),
T ∗
3 = (1 − T2)(1 + µ2T2),
T3T ∗
2 = µ2(1 − T2)(1 + µ2T2).
(37)
(38)
(39)
Proof : Applying κ on Lemma 3.21, we obtain ( 37 ).
Unitarity of the matrix Z ( ( 2, 2 ) position of the matrix Z ∗Z ) gives µ2(1 + µ2)T ∗
3 T3 + T 2
2 +
(1 + µ2)T3T ∗
3 = 1.
Using ( 37 ) we deduce −(1 + µ2)2T3T ∗
Applying κ on ( 38 ), we deduce ( 39 ). ✷
3 = (T2 − 1)(1 + µ2T2). Thus we obtain ( 38 ).
Lemma 3.23 S∗
4 S4 = S4S∗
4 = (1 + µ2)−2
µ2(1 − T2)2.
Proof : Adding ( 23 ) and ( 24 ), we have :
2
3 T3+µ4(1 + µ2)
(1−µ2)S4S∗
4 = −µ2(1+µ2)(1−µ2)T2(1−T2)−µ2(1−µ2)S2S∗
2 .
3
(1−µ2)T ∗
−µ4(1 + µ2)
Using µ2 6= 1, we obtain,
−µ4(1 + µ2)
3
T ∗
3 T3 + µ4(1 + µ2)
2
S4S∗
4 = −µ2(1 + µ2)T2(1 − T2) − µ2S2S∗
2 .
Now using ( 37 ) and ( 39 ), we reduce the above equation to
S4S∗
4
2
µ4(1 + µ2)
= −µ2(1 − T2)(T2 + µ2T2 + µ2 + µ4T2) + µ2(1 + µ2)(1 − T2)(µ2 + T2)
= µ6(1 − T2)2.
15
Thus,
S4S∗
4
=
=
µ6
µ4(1 + µ2)2 (1 − T2)2
(1 + µ2)2 (1 − T2)2.
µ2
Applying κ, we have S∗
Thus, S∗
4 S4 = S4S∗
4 S4 = µ2
4 = µ2
(1+µ2)2 (1 − T2)2.
(1+µ2)2 (1 − T2)2. ✷
Lemma 3.24 µ2(1 + µ2)2
S∗
3 S3 = (µ2 + T2)[µ2(1 + µ2) − (1 − T2)].
Proof : Using Lemma 3.21 in ( 7 ), we have
S∗
3 S3 + T ∗
3 T3 + T3T ∗
3 + S∗
4 S4 = 1.
(40)
The lemma is derived by substituting the expressions of T ∗
and Lemma 3.23 in the equation ( 40 ). ✷
3 T3, T3T ∗
3 and S∗
4 S4 from ( 37 ), ( 38 )
Lemma 3.25 (1 + µ2)2
S3S∗
3 = (1 + µ2T2)(1 + µ2 − µ4(1 − T2)).
Proof : By unitarity of the matrix Z, in particular equating the ( 1, 1 ) th entry of ZZ ∗ to 1
4 S4 = 1. Then the Lemma follows by using ( 38 ) and Lemma
3 + µ2S∗
we get S3S∗
3 + µ2(1 + µ2)T3T ∗
3.23 in the above equation. ✷
Lemma 3.26 −S∗
2 S3 = (µ2 + T2)T3.
Proof : By applying the adjoint and then multiplying by µ2 on ( 10 ) we have µ2S∗
−µ2T3(µ2 + T2) + µ2(1 − T2)T3. Subtracting this from ( 11 ) we have (1 − µ2)S∗
µ2T3T2 + µ2T3(µ2 + T2) − µ2(1 − T2)T3 which implies −S∗
Lemma 3.27 S2(1 − T2) = µ2(1 − T2)S2.
2 S3 = (µ2 + T2)T3 as µ2 6= 1. ✷
2 S3 +µ2S∗
4 S2 =
2 S3 = −T2T3 −
Proof : Applying κ to Lemma 3.26 and then taking adjoint, we have
2
µ2(1 + µ2)
T ∗
3 S3 = −(µ2 + T2)S2.
Adding ( 33 ) and ( 34 ) and then taking adjoint, we get ( by using µ2 6= 1 )
2
µ2(1 + µ2)
T3S4 = µ4(1 − T2)S2.
Moreover, ( 25 ) gives
(41)
(42)
2
µ2(1 + µ2)
S4T3 = −S2(µ2 + T2) + (1 − T2)S2 − µ2(1 + µ2)
Using ( 41 ), the right hand side of this equation turns out to be S2(1 − T2).
2
T ∗
3 S3.
16
Thus,
2
(1 + µ2)
S4T3 = µ−2S2(1 − T2).
Again, application of adjoint to the equation ( 33 ) gives :
µ2(1 + µ2)
2
S3T ∗
3 = −µ2(1 + µ2)
2
T3S4 − µ2(1 + µ2)S2 + µ4(1 − T2)S2 + µ4S2(1 − T2).
Using ( 42 ), we get
Using ( 41 ) - ( 44 ) to the equation ( 14 ), we obtain :
(1 + µ2)
2
S3T ∗
3 = −S2(1 + µ2T2).
(43)
(44)
S2 − 2S2T2 − (1 + µ2)−1µ2S2(1 + µ2T2) + µ−2(1 + µ2)−1S2(1 − T2) = µ2S2 − 2µ2T2S2 + (1 +
µ2)−1µ6(1 − T2)S2 − (1 + µ2)−1(µ2 + T2)S2.
This gives
µ2(1 + µ2)[(S2 − S2T2)− (µ2S2 − µ2T2S2)]− µ2(1 + µ2)(S2T2 − µ2T2S2)− µ4S2 − µ6S2T2 + S2(1−
Thus, µ2(1 + µ2)[S2(1 − T2) − µ2(1 − T2)S2] + S2(1 − T2) − µ2(S2 − T2S2) + µ6[S2(1 − T2) −
On simplifying, (µ6 + 2µ4 + 2µ2 + 1)(S2(1 − T2) − µ2(1 − T2)S2) = 0, which proves the lemma
T2) − µ8(S2 − T2S2) + µ4S2 + µ2T2S2 = 0.
µ2(1 − T2)S2] − µ6(1 − T2)S2 + µ4S2(1 − T2) + µ2(S2(1 − T2) − µ2(1 − T2)S2) = 0.
as 0 < µ < 1. ✷
Lemma 3.28
T3(1 − T2) = µ2(1 − T2)T3,
S3S∗
4 = µ4S∗
4 S3.
(45)
(46)
Proof : The equation ( 45 ) follows by applying κ on Lemma 3.27 and then taking adjoint.
We have S∗
3 from ( 12 ). On the other hand we have S3S∗
3 from ( 22 ).
4 S3 = −T 2
4 = −µ4T 2
Combining these two, we get ( 46 ). ✷
Lemma 3.29 S4T2 = T2S4.
Proof : Subtracting ( 31 ) from ( 16 ) we get the required result. ✷
Lemma 3.30 T3S2 = S2T3.
Proof : By applying adjoint on ( 30 ) and then subtracting it from ( 15 ) we obtain S2T3−T3S2 =
0. ✷
Lemma 3.31 S3(1 − T2) = µ4(1 − T2)S3.
Proof : By adding ( 15 ) with ( 17 ) we obtain
Thus, using µ2 6= 1,
S3(1 − T2) + µ2S3(T2 − 1) = µ2(µ2 − 1)T3S2.
Moreover, by taking adjoint of ( 30 ), we obtain µ2(1 − T2)S3 = µ2S2T3 − T3S2 + S3(1 − T2).
S3(1 − T2) = −µ2T3S2.
(47)
17
Thus,
µ4(1 − T2)S3 = µ4S2T3 − µ2T3S2 + µ2S3(1 − T2).
Hence, to prove the Lemma it suffices to prove:
S3(1 − T2) = µ4S2T3 − µ2T3S2 + µ2S3(1 − T2).
After using T3S2 = S2T3 obtained from Lemma 3.30 we get this to be the same as (1 − µ2)S3(1 −
T2) = µ2(µ2 − 1)T3S2. This is equivalent to S3(1 − T2) = −µ2T3S2 ( as µ2 6= 1 ) which follows from
( 47 ). ✷
Proposition 3.32 The map SOµ(3) → Q sending M, L, G, N, C to −(1+µ2)−1S2, S3, −µ−1S4, (1+
µ2)−1(1 − T2), µT3 respectively is a CQG homomorphism.
Proof : It is enough to check that the map is ∗-homomorphic, since the coproducts on SOµ(3)
and Q are determined in terms of the fundamental unitaries Z ′ and Z respectively, and the map
described in the statement of the proposition sends (ij)-th entry of Z ′ to the (ij)-th entry of Z for
all (ij).
Now, it can easily be checked that the proof of the homomorphic property of the given map
reduces to verification of the relations on Q as derived in Lemmas 3.21 - 3.31 along with the
following equations :
(48)
(49)
(50)
(51)
S3S4 = µ4S4S3,
S3S2 = µ2S2S3,
S2S4 = µ2S4S2,
S3S4 = −
µ2
(1 + µ2)2 S2
2 ,
which follow from Remark 3.20, ( 28 ), ( 29 ), ( 32 ) respectively. ✷
Theorem 3.33 We have the isomorphism:
QISO+
R(O(S2
µ,c), H, D) ∼= SOµ(3).
^
QISO+
^
QISO+
R(D) to SUµ(2) which clearly maps QISO+
Proof : SUµ(2) is an object in
morphism from
latter as a quantum subgroup of QISO+
SOµ(3) by Π. On the other hand, Proposition 3.32 implies that QISO+
of SOµ(3), and the corresponding surjective CQG morphism from SOµ(3) onto QISO+
seen to be the inverse of Π, thereby completing the proof. ✷
R(D) as remarked before, and thus one gets a surjective
R(D) onto SOµ(3), identifying the
R(D) to
R(D) is a quantum subgroup
R(D) is clearly
R(D). Let us denote the surjective map from QISO+
Remark 3.34 Theorem 3.33 shows that for a fixed µ, the quantum isometry group QISO+
R(D)
of S2
µ,c does not depend on c. This may appear somewhat surprising, but let us remark that in
the classical situation ( that is for µ = 1 ), c corresponds to the radius of the sphere and S2
1,c
are isomorphic as C ∗ algebras for all c ≥ 0. We refer the reader to [14], page 126, for the details
regarding this. Although in the noncommutative case, that is, when µ 6= 1, we do get non-isomorphic
C ∗-algebras S2
µ,c for different choices of c, one may still think that the parameter c in some sense
18
determines the 'radius' of the noncommutative sphere, and thus one should get the same (quantum)
isometry group for different choices of c.
In view of the above, it seems impossible to 'reconstruct' the quantum homogeneous spaces S2
µ,c
from the quantum isometry groups SOµ(3). In this context, it may be mentioned that for µ 6= 1,
although all S2
µ,0 arises as a
quotient of SOµ(3) by a quantum subgroup (see [17] for more details). Thus, it is perhaps possible
to somehow 'reconstruct' S2
µ,c are quantum homogeneous spaces corresponding to SOµ(3), only S2
µ,0 from the quantum group SOµ(3).
3.4
Existence of ^QISO+(D)
For the above spectral triple, we have been unable to settle the issue of the existence of ^QISO+(D)
which is the universal object ( if it exists ) in the category Q′(D) mentioned in Section 1. Nev-
ertheless, we shall show that if a universal object in Q′(D) exists, then QISO+(D) must coincide
with SOµ(3).
Lemma 3.35 If ^QISO+(D) exists, its induced action on S2
on the subspace spanned by {1, A, B, B∗, AB, AB∗, A2, B2, B∗2
µ,c, say α0, must preserve the state h
}.
2
2
2
2
as well. Let W 3
= Span{1, A, B, B∗},
= Span{1, A, B, B∗, AB, AB∗, A2, B2, B∗2}.
Proof : Let W0 = C.1, W 1
W 3
We note that the proof of Proposition 3.10 and the lemmas preceding it do not use the assump-
tion that the action is R-twisted volume preserving, so the proof of Proposition 3.10 goes through
verbatim implying that α0 keeps invariant the subspace spanned by {1, A, B, B∗} and hence it
2 ⊕ W ′ be the orthogonal decomposition with respect to the
preserves W 3
Haar state (say h0) of QISO+(D). Since SOµ(3) is a sub-object of QISO+(D), there is a CQG
morphism π from QISO+(D) onto SOµ(3) satisfying (id ⊗ π)α0 = ∆, where ∆ is the SOµ(3) ac-
tion on S2
µ,c. It follows from this that any QISO+(D)-invariant subspace (in particular W ′) is also
, the SOµ(3)-action decomposes
SOµ(3)-invariant. On the other hand, it is easily seen that on W 3
2 ⊕ W ′′, (orthogonality with respect to h, the Haar state of SOµ(3))), where W ′′ is a five
as W 1
dimensional irreducible subspace.
We claim that W ′ = W ′′, which will prove that the QISO+(D)-action α0 has the same h-
, so preserves C.1 and its h-orthogonal
= W 1
orthogonal decomposition as the SOµ(3)-action on W 3
complements. This will prove that α0 preserves the Haar state h on W 3
2
We now prove the claim. Observe that V := W ′TW ′′ is invariant under the SOµ(3)-action but
due to the irreducibility of ∆ on the vector space W ′ or W ′′, it has to be zero or W ′ = W ′′. Now,
dim(V) = 0 implies dim(W ′) + dim(W ′′) = 5 + 5 > 9 = dim(W 3
) which is a contradiction unless
W ′ = W ′′.
2
2
.
2
✷
Theorem 3.36 If ^QISO+(D) exists, then we must have that QISO+(D) ∼= SOµ(3).
Proof :
In the proof of Lemma 3.35, it was noted that Proposition 3.10 follows under the
assumption of the present theorem. To complete the proof of the theorem, we just need to observe
that the other lemmas used for proving Theorem 3.33 require the conclusion of Lemma 3.35 as the
only extra ingredient. ✷
19
Let us conclude the article with brief explanation of the technical difficulties regarding the issue
of existence of ^QISO+(D) . Let R′ be a positive, invertible operator commuting with D such
µ,c. The problem of existence of
that τR′
^QISO+(D) is closely related to the question whether it is possible to identify QISO+
R′ (D) as a
quantum subgroup of SOµ(3) for a general R′. By Theorem 3.36, a negative answer of this question
will prove that Q′(D) does not have a universal object.
6= τR and let φ′ denote the action of QISO+
R′ (D) on S2
i , S′
i which generate QISO+
Now, as has been noted in Remark 3.5, φ′ is linear, that is, it keeps the span of {1, A, B, B∗}
invariant and hence it us given by an expression similar to equations ( 5 ) and ( 6 ) with Ti, Si
replaced by some T ′
R′(D) as a C ∗-algebra. We can in principle write
down all the relations satisfied by these generators, proceeding as in the Subsection 3.3. These
relations will be analogous to equations ( 7 )-( 34 ), and in fact, the relations which make use of the
homomorphism property only remain unchanged. However, the ones making use of the fact that
φ′ preserves τR′ will change, since τR′ is in general different from τR. In particular, the expression
of the antipode will change, which will affect all the relations starting from ( 23 ). We need to
i for a general R′, and
have a deeper and systematic understanding of the relations satisfied by T ′
possibly study their representations in concrete Hilbert spaces, to decide whether QISO+
R′(D) is a
quantum subgroup of SOµ(3) or not. We are not yet able to do this.
i , S′
Moreover, even if ^QISO+(D) exists, although we can identify QISO+(D) with the well-known
quantum group SOµ(3), it is not so easy to explicitly compute ^QISO+(D).
If U denotes the
unitary representation corresponding to ^QISO+(D), the fact that U commutes with D implies
that U must preserve each of the two-dimensional eigenspaces span{vl
2} and
span{vl
2} of D. Suppose that (qij)i,j=1,2 and (rij)i,j=1,2 are the matrices
(with entries in ^QISO+(D) ) of U corresponding to these two spaces respectively. Then it is clear
that as a C ∗ algebra ^QISO+(D) will be generated by qij, rij's as well as the generators Ti, Si of
SOµ(3). However, the mutual relations among these generating elements have to be determined
from the fact that U preserves each of the eigenspaces of D. In principle one gets infinitely many
such relations which are quite complicated and it is not clear how to simplify them.
: m = ± 1
: m = ± 1
2 − vl
+ vl
m,− 1
2
m, 1
2
m, 1
m,− 1
2
References
[1] Banica, T.: Quantum automorphism groups of small metric spaces, Pacific J. Math.
219(2005), no. 1, 27 -- 51.
[2] Banica, T.: Quantum automorphism groups of homogeneous graphs, J. Funct. Anal.
224(2005), no. 2, 243 -- 280.
[3] Bhowmick, J. and Goswami, D.: Quantum isometry groups : examples and computations,
Comm. Math. Phys. 285 ( 2009 ), no. 2, 421-444, arXiv 0707.2648.
[4] J. Bhowmick and D. Goswami, Quantum group of orientation preserving Riemannian
Isometries, J. Funct. Anal. 257 ( 2009 ), no. 8, 2530-2572.
20
[5] Bichon, J.: Quantum automorphism groups of finite graphs, Proc. Amer. Math. Soc.
131(2003), no. 3, 665 -- 673.
[6] Connes, A.: "Noncommutative Geometry", Academic Press, London-New York (1994).
[7] Connes, Alain; Dubois-Violette, Michel. : Noncommutative finite-dimensional manifolds.
I. Spherical manifolds and related examples. Comm. Math. Phys. 230 (2002), no. 3, 539 -- 579.
[8] Dabrowski, L., D'Andrea, F., Landi, G., Wagner, E.: Dirac operators on all Podles
quantum spheres, J. Noncomm. Geom. 1 (2007) 213-239.
[9] Dabrowski, L.: Spinors and Theta Deformations, math.QA/0808.0440.
[10] Frohlich, J.; Grandjean, O.; Recknagel, A.: Supersymmetric quantum theory and non-
commutative geometry, Comm. Math. Phys. 203 (1999), no. 1, 119 -- 184.
[11] Goswami, D.: Twisted entire cyclic cohomology, JLO cocycles and equivariant spectral
triples, Rev. Math. Phys. 16( 2004 ), no.5, 583-602.
[12] Goswami, D.: Quantum Group of isometries in Classical and Non Commutative Geometry,
Comm. Math. Phys. 285 ( 2009 ), no. 1, 141-160, , arXiv 0704.0041.
[13] Goswami, D.: Some Noncommutative Geometric aspects of SUq(2), arXiv:0108003v4.
[14] Klimyk, A. and Schmudgen, K. : Quantum Groups and their Representations, Springer,
New York, 1998.
[15] Maes, A. and Van Daele, A.: Notes on compact quantum groups, Nieuw Arch. Wisk.
(4)16(1998), no. 1-2, 73 -- 112.
[16] Podles, P.: Quantum Spheres, Letters in Mathematical Physics 14 ( 1987 ) 193 - 202.
[17] Podles, P.: Symmetries of quantum spaces. Subgroups and quotient spaces of quantum
SU (2) and SO(3) groups, Commun. Math. Phys. 170 (1995), 1-20.
[18] Schmudgen, K. and Wagner, E. : Dirac operators and a twisted cyclic cocycle on the
standard Podles' quantum sphere , J. Reine Angew. Math. 574 (2004), 219 -- 235.
21
[19] Schmudgen, K. and Wagner, E. : Representation of cross product algebras of Podles
quantum spheres, math.QA/0305309.
[20] Soltan, P. M.: Quantum SO(3) Groups and Quantum Group Actions on M2, arXiv:
0810.0398v1.
[21] Wang, S.: Free products of compact quantum groups, Comm. Math. Phys. 167 (1995), no.
3, 671 -- 692.
[22] Wang, S.: Quantum symmetry groups of finite spaces, Comm. Math. Phys. 195(1998),
195 -- 211.
[23] Wang, S.: Structure and isomorphism classification of compact quantum groups Au(Q) and
Bu(Q), J. Operator Theory 48 (2002), 573 -- 583.
[24] Wang, S. : Ergodic actions of universal quantum groups on operator algebras. Comm.
Math. Phys. 203 (1999), no. 2, 481 -- 498.
[25] Woronowicz, S. L.: Compact matrix pseudogroups, Comm. Math. Phys. 111(1987), no. 4,
613 -- 665.
[26] Woronowicz, S. L.: "Compact quantum groups", pp. 845 -- 884 in Sym´etries quantiques
(Quantum symmetries) (Les Houches, 1995), edited by A. Connes et al., Elsevier, Amster-
dam, 1998.
[27] Woronowicz, S. L.: Pseudogroups, pseudospaces and Pontryagin duality, Proceedings of
the International Conference on Mathematical Physics, Lausane (1979), Lecture Notes in
Physics 116, pp. 407-412.
Jyotishman Bhowmick : Stat-Math Unit, Indian Statistical Institute, 203, B. T. Road, Kolkata
700 108 E-mail address: [email protected]
Debashish Goswami: Stat-Math Unit, Indian Statistical Institute, 203, B. T. Road, Kolkata 700
108 E-mail address: [email protected]
22
|
1709.08839 | 3 | 1709 | 2018-10-07T12:11:01 | The inner structure of boundary quotients of right LCM semigroups | [
"math.OA"
] | We study distinguished subalgebras and automorphisms of boundary quotients arising from algebraic dynamical systems $(G,P,\theta)$. Our work includes a complete solution to the problem of extending Bogolubov automorphisms from the Cuntz algebra in $2 \leq p<\infty$ generators to the $p$-adic ring $C^*$-algebra. For the case where $P$ is abelian and $C^*(G)$ is a maximal abelian subalgebra, we establish a picture for the automorphisms of the boundary quotient that fix $C^*(G)$ pointwise. This allows us to show that they form a maximal abelian subgroup of the entire automorphism group. The picture also leads to the surprising outcome that, for integral dynamics, every automorphism that fixes one of the natural Cuntz subalgebras pointwise is necessarily a gauge automorphism. Many of the automorphisms we consider are shown to be outer. | math.OA | math |
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT
LCM SEMIGROUPS
VALERIANO AIELLO, ROBERTO CONTI, STEFANO ROSSI, AND NICOLAI STAMMEIER
Abstract. We study distinguished subalgebras and automorphisms of boundary quo-
tients arising from algebraic dynamical systems (G, P, θ). Our work includes a complete
solution to the problem of extending Bogolubov automorphisms from the Cuntz alge-
bra in 2 ≤ p < ∞ generators to the p-adic ring C ∗-algebra. For the case where P
is abelian and C ∗(G) is a maximal abelian subalgebra, we establish a picture for the
automorphisms of the boundary quotient that fix C ∗(G) pointwise. This allows us to
show that they form a maximal abelian subgroup of the entire automorphism group.
The picture also leads to the surprising outcome that, for integral dynamics, every au-
tomorphism that fixes one of the natural Cuntz subalgebras pointwise is necessarily a
gauge automorphism. Many of the automorphisms we consider are shown to be outer.
Contents
Introduction
1.
2. Preliminaries and Notations
3. Maximal abelian subalgebras
3.1. The group C*-algebra
3.2. The diagonal
4. The relative commutant of a family of generating isometries
5. Extendability of Bogolubov automorphisms
6. Automorphisms preserving the group C*-algebra
6.1. Automorphisms fixing the group C*-algebra
6.2. A closer look at integral dynamics
6.3. Outerness
References
1
5
8
8
9
11
15
19
19
22
24
27
1. Introduction
Introduced as long ago as the late 1970s, [Cun77], the Cuntz algebras have since been
an undeniably interesting field of research. A particularly fascinating area within this
field is the structure of the endomorphisms and the automorphisms of Cuntz algebras,
which provide fertile grounds for a deep interplay between C∗-algebra theory, ergodic
theory, and combinatorics. Over the last decade, remarkable progress has been made
towards a better understanding of certain key features and challenging problems, see
[Con10, CRS10, CS11, CHS12a, CHS12b, CHS12c, CHS12d, CHS15] for a brief selection.
Date: October 9, 2018.
1
2
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
Inspired by [Cun77], a wealth of constructions of C∗-algebras associated with various
sorts of input data has been constructed. It is thus quite natural to ask to what extent
the endomorphism structure of these C∗-algebras resembles the case of the original
Cuntz algebras. For instance, through the works [Cun08, CL10, CL11, Li10, Li12, CDL13,
BRRW14] there are classes of C∗-algebras that are associated with purely algebraic
objects, such as rings, integral domains, fields, and arbitrary left cancellative semigroups.
Especially the study of the C∗-algebras constructed in [Li12] has led to an extensive
list of impressive results, from which we would like at least to mention [Li13, CEL13,
CEL15, ELR16, ABLS17], and refer the reader to [Cun16] for a recent survey as well as
to [CELY17] for a more detailed exposition with an emphasis on K-theory.
In [LL12], Larsen and Li performed a detailed case study on the representation theory
of one particular C∗-algebra of the former type: the 2-adic ring C∗-algebra Q2. This
algebra is the universal C∗-algebra generated by a unitary u and an isometry s subject
to su = u2s and ss∗ + uss∗u∗ = 1, which can also be described as the ring C∗-algebra
associated to Z ⋊ h2i ⊂ Z ⋊ Z×. One reason for choosing Q2 is that the algebra is a
balanced version of O2 in the sense that it is a unital UCT Kirchberg algebra whose
K-groups are both equal to Z. In addition, O2 appears quite naturally as the subalgebra
C∗({s, us}) of Q2. This served as the motivation for the first three authors to investigate
the inner structure of Q2, see [ACR18a, ACR18b]. Shortly after the first version of
[ACR18a] was being circulated, we realized that a good part of the questions, answers
and techniques entering the proofs in [ACR18a] have analogues in the much broader
setting of boundary quotients of right LCM semigroups. In hindsight, it is fair to say
that most of those results and proofs have in fact become clearer and more conceptual
as a benefit of the higher level of abstraction.
Let us now describe the C∗-algebras to which we extend the line of research started
in [ACR18a]: We shall focus on universal C∗-algebras associated to particularly well-
behaved examples of right LCM semigroups, that is, left cancellative monoids in which
the intersection of any two principal right ideals is either empty, or another principal
right ideal again. There are two main types of examples that we have in mind here:
(a) integral dynamics Z ⋊ P ⊂ Z ⋊ N×, where P is generated by a family of mutually
(b) semidirect products G⋊θ N for an injective endomorphism θ of a discrete, abelian
coprime positive integers, as in [BOS18], see Example 2.1;
group G with finite cokernel as appearing in [CV13], see Example 2.2.
These are elementary examples of right LCM semigroups of the form G ⋊θ P built from
algebraic dynamical systems (G, P, θ) in the sense of [BLS], that is, G is a countable
discrete group, P is a right LCM semigroup, and θ : P y G is an action by injective
group endomorphisms such that pP ∩ qP = rP implies θp(G) ∩ θq(G) = θr(G) for all
p, q ∈ P . Let us also recall that (G, P, θ) is an algebraic dynamical system of finite type
if the index [G : θp(G)] is finite for all p ∈ P . Some results will be proven for more
general algebraic dynamical systems (G, P, θ) than the ones specified in (a) and (b),
see for instance Section 4 and Theorem 6.5. However, we assume that the right LCM
semigroup P is directed with respect to p ≥ q
⇔ p ∈ qP . In this case, the boundary
quotient Q(G⋊θ P ), that is, the C∗-algebras we intend to study, is the quotient of the full
semigroup C∗-algebra C∗(G ⋊θ P ) from [Li12] by the relationsP[g]∈G/θp(G) v(g,p)v∗(g,p) = 1
def
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
3
for all p ∈ P for which θp(G) has finite index in G, where the v(g,p) denote the standard
generating isometries in C∗(G ⋊θ P ), see [BS16, Proposition 4.1].
By virtue of [BS16, Proposition 4.3], directedness of P is also crucial to have access
to a natural representation π : Q(G ⋊θ P ) → L(ℓ2(G)), to which we shall refer as the
canonical representation. This canonical representation is faithful under moderate as-
sumptions and we start by addressing its irreducibility in Proposition 2.3. In addition
Proposition 2.4 gives the decomposition of the induced representation of the torsion
subalgebra AS for integral dynamics into irreducible subrepresentations.
In light of the renewed interest in the study of maximal abelian subalgebras and Car-
tan subalgebras, see for instance [BL17, BL, LR17], we take pain to spot mild conditions
ensuring that C∗(G) is a maximal abelian subalgebra of Q(G ⋊θ P ), see Theorem 3.5,
and that the diagonal D generated by the range projections of the generating isometries
in Q(G ⋊θ P ) is a Cartan subalgebra, see Theorem 3.10. The proofs we give rely on
the canonical representation and conditional expectations in both cases. In the case of
C∗(G), we observe a rigidity phenomenon for conditional expectations from L(ℓ2(G))
onto the group von Neumann algebra W ∗(G), see Proposition 3.3, which leads to a
uniqueness result for conditional expectations Q(G ⋊θ P ) → C∗(G), see Remark 3.4.
In Section 4, we consider general algebraic dynamical systems (G, P, θ) with P abelian
and a strong form of minimality, namely Tp∈P θp(G) = {1G}. In this setting, we es-
tablish a generalization of the classical Fourier-coefficient technique from [Cun77], see
Lemma 4.4. This is a key ingredient to prove Theorem 4.6, which asserts that the rel-
ative commutant of the generating isometries for P inside Q(G ⋊θ P ) is as small as
possible, namely C∗(P ∗), where P ∗ is the subgroup of invertible elements in P . We
remark that we need to assume P ∗ to be finite for Theorem 4.6 for technical reasons,
but this seems likely to be unnecessary.
In contrast to the generality of Section 4, we focus on the intersection of the two
classes described in Example 2.1 and Example 2.2 in Section 5. That is to say, we study
the p-adic ring C∗-algebras Qp = Q(Z ⋊ hpi) for 2 ≤ p < ∞. The torsion subalgebra of
Qp is then the Cuntz algebra Op generated by (umsp)0≤m≤p−1. Understanding the way
Op sits inside Qp is a very natural and rewarding task. For example, it was shown in
[BOS18] that the inclusion induces a split-injection onto the torsion part in K-theory.
Here, we show that the representation σ : Op → L(L2([0, 1])) known as the interval
unitarily equivalent to the canonical representation, see Proposition 5.2.
picture extends in a unique way to representation eσ : Qp → L(L2([0, 1])), which is not
More importantly, we use this result and a fact about the canonical representation
to completely solve the problem of extending Bogolubov automorphisms of Op to an
endomorphism of Qp, see Theorem 5.8: A Bogolubov automorphism is extendible if and
only if it is a gauge automorphism, the exchange automorphism, or a composition of the
former two. Moreover, a Bogolubov automorphism admits at most one extension, which
either fixes the unitary u (gauge automorphism) or sends it to u∗ (exchange automor-
phism is present). In particular, every extension of a Bogolubov automorphism is an
automorphism of Qp. This generalization of [ACR18a, Theorem 4.14] hints at a remark-
able rigidity for extensions of certain automorphism groups of the torsion subalgebra to
endomorphisms of the boundary quotient Q(G ⋊θ P ). Towards this goal, the contribu-
tion of Section 5 is to provide a refined version of the argument given in [ACR18a] that
4
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
may pave the way to similar results, for instance in the context of Zappa-Sz´ep products
associated to self-similar group actions, see [BRRW14].
In Section 6, we focus on algebraic dynamical systems (G, P, θ) with abelian group
G. In this case, the boundary quotient Q(G ⋊θ P ) contains a copy of the group C∗-
algebra C∗(G), and we investigate the structure of automorphisms of Q(G ⋊θ P ) that
preserve C∗(G) globally. To describe the subgroup of all automorphisms that fix C∗(G)
pointwise, which we shall denote by AutC ∗(G)Q(G ⋊θ P ), we introduce the notion of
of G, see Definition 6.1. We denote the group formed under pointwise multiplication
a θ-twisted homomorphism ψ : P → C(bG, T), where bG denotes the Pontryagin dual
by Homθ(P, C(bG, T)). In Theorem 6.2, we then show that Homθ(P, C(bG, T)) embeds
into AutC ∗(G)Q(G ⋊θ P ). Under the additional assumption that C∗(G) ⊂ Q(G ⋊θ
P ) is maximal abelian, this embedding is also surjective and EndC ∗(G)Q(G ⋊θ P ) =
AutC ∗(G)Q(G ⋊θ P ). In particular, we deduce that every element in EndC ∗(G)Q(G ⋊θ
P ) is characterized by a family of unitaries in C∗(G), corresponding to a θ-twisted
homomorphism ψ.
As a first application of this result, we prove that AutC ∗(G)Q(G ⋊θ P ) is a maximal
abelian subgroup of AutQ(G ⋊θ P ) in Theorem 6.5, assuming that (G, P, θ) is an alge-
braic dynamical system of finite type such that G and P are abelian, C∗(G) ⊂ Q(G⋊θ P )
is maximal abelian, and Q(G ⋊θ P ) is simple. This result also applies to our motivating
examples (a) and (b), see Corollary 6.6. A second application of the picture established
in Theorem 6.2 is given by Theorem 6.11: For every integral dynamics (Z, P, θ), every au-
tomorphism of Q(Z⋊P ) that fixes a Cuntz subalgebra On = C∗({uksn 0 ≤ k ≤ n−1})
pointwise for some n ≥ 2, n ∈ P , necessarily belongs to AutC ∗(Z)Q(Z ⋊ P ). In addition,
we determine the subgroup of Homθ(P, C(T, T)) corresponding to AutOnQ(Z ⋊ P ). One
important consequence of these findings is that two automorphism of Q(Z⋊P ) are equal
if and only if they agree on the torsion subalgebra AS, see Corollary 6.12. A heuris-
tic explanation for this phenomenon comes from Theorem 6.11 in combination with the
crossed product descriptions Q(G ⋊θ P ) ∼= lim−→ Mp(C∗(Z)) ⋊P and AS ∼= lim−→ Mp(C) ⋊P ,
see [BOS18] for details.
Within the final Subsection 6.3, we show that many of the automorphisms we consid-
ered previously yield outer actions: In the case of (b), every automorphism of Q(G ⋊θ N)
f (1 bG) 6= 1, see Proposition 6.14. A criterion for the outerness of the gauge action based
on functional equations is given in Theorem 6.15 for the case where C∗(G) ⊂ Q(G ⋊θ P )
is maximal abelian and P is abelian. This applies readily to integral dynamics, see Corol-
lary 6.16. However, the latter result also follows easily from the outerness result for the
that fixes C∗(G), and hence is given by some f ∈ C(bG, T) due to Theorem 6.2, is outer if
gauge action in Corollary 6.17, which assumes P to be abelian, butTp∈P θp(G) = {1G}
In Theorem 6.18, we then prove that, for integral
in place of an assumption on G.
dynamics, every automorphism that inverts the generating unitary u is outer. Combin-
ing this with Corollary 6.16 and Theorem 5.8 shows that the extensions of Bogolubov
automorphisms of Op yield an outer action T × Z/2Z y Q(Z ⋊ P ).
Acknowledgements: This work was initiated during a visit of the fourth author to
the first three authors in Rome in June 2016. He thanks his collaborators for quite an
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
5
enjoyable and productive week, and Sapienza University for its hospitality and financial
support. The fourth author was supported by RCN through FRIPRO 240362.
2. Preliminaries and Notations
In what follows we shall be dealing with C∗-algebras associated to semigroups built
from algebraic dynamical systems. All our semigroups will have an identity, and hence
are monoids. By an algebraic dynamical system we mean a triple (G, P, θ), where
(a) G is a countable discrete group;
(b) P is a right LCM semigroup, that is, a countable left cancellative monoid in
which the intersection of two principal right ideals is either empty or another
principal right ideal; and
(c) θ is an action of P upon G through injective homomorphisms that respects the
order, that is, pP ∩ qP = rP implies θp(G) ∩ θq(G) = θr(G).
Basics on algebraic dynamical systems (G, P, θ) are to be found in [BLS17, BLS]. It is
known that the last two conditions are equivalent to the right LCM condition for G⋊θ P ,
given that G is a group. This semidirect product is the actual right LCM semigroup we
are interested in.
For the dynamical system (G, P, θ), the (full) semigroup C∗-algebra C∗(G ⋊θ P ) in
the sense of Li provides a natural object to study, see [BLS]. This C∗-algebra admits a
canonical representation on ℓ2(G ⋊θ P ), which is faithful in a number of cases, e.g. when
G is amenable and P is left Ore with amenable enveloping group P −1P . However,
C∗(G ⋊θ P ) rather resembles C∗-algebras of Toeplitz type. In fact, it can be viewed as
a Nica-Toeplitz algebra for a discrete product system of Hilbert bimodules over P , see
[BLS, Theorem 7.9].
Inspired by the approach of Crisp and Laca for right-angled Artin groups, see [CL07],
a boundary quotient Q(S) was defined in [BRRW14] for general right LCM semigroups
S. This quotient of C∗(S) was then studied in connection with structure results for
C∗-algebras associated to tight groupoids of inverse semigroups, see [Star15] and the
references therein for details. In [BS16], the boundary relation for Q(S) was analyzed in
order to identify this quotient with a previously known C∗-algebra in important cases.
If P is directed with respect to reverse inclusion of the associated principal right ideals,
that is, p ≥ q ⇔ p ∈ qP , then Q(G ⋊θ P ) is the universal C∗-algebra generated by
a unitary representation u of the group G and a representation s of the monoid P by
isometries satisfying the relations
(I) spug = uθp(g)sp,
(II) s∗pugsq =(ug1sp′s∗q′ug2
(III) Pg∈G/θp(G) eg,p = 1 if Np < ∞,
.
= ugsps∗pu∗g and Np
0
, if g = θp(g1)θq(g2) and pP ∩ qP = pp′P, pp′ = qq′
, otherwise.
.
= [G : θp(G)], see [BS16, Proposition 4.1]. We remark
where eg,p
that eg,p is independent of the choice of g ∈ g = gθp(G) by (I).
This result allows for an identification of Q(G ⋊θ P ) with the C∗-algebra O[G, P, θ]
from [Sta15] for irreversible algebraic dynamical systems (G, P, θ) in the sense of [Sta15],
see [BS16, Corollary 4.2]. The latter C∗-algebra was constructed as a universal model
for the natural realization of the dynamics on ℓ2(G). More generally, it is shown in
6
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
[BS16, Proposition 4.3] that for a general algebraic dynamical system (G, P, θ), the C∗-
algebra Q(G ⋊θ P ) admits a canonical representation π : Q(G ⋊θ P ) → L(ℓ2(G)), ugsp 7→
UgSp given by UgSpξh = ξgθp(h) if and only if P is directed. In addition, we remark that
π is known to be faithful under moderate assumptions, which in fact guarantee that
Q(G ⋊θ P ) is even simple, see [BS16, Theorem 4.17].
Instead of continuing with this increasingly involved discussion of the structure for
general algebraic dynamical systems, we have chosen to focus on two natural general-
izations of the dynamics considered in [ACR18a]. There, the authors studied the case
of (G, P, θ) = (Z, N, 2), i.e. multiplication by 2 on the integers. We will consider the
following:
Example 2.1. Let S ⊂ N× \ {1} be a family of relatively prime natural numbers, and
P ⊂ N× the free abelian monoid generated by S, i.e. P = hSi. Then P is right LCM,
and acts on Z by multiplication θp(n) = pn for p ∈ P, n ∈ Z. This defines an action θ
that respects the order if and only if S consists of relatively prime numbers. Thus we
obtain an (irreversible) algebraic dynamical system (Z, P, θ), and the resulting boundary
quotient Q(Z ⋊θ P ) has the following features:
[Sta15, Lemma 3.4].
(i) s∗psq = sqs∗p for all p, q ∈ S, p 6= q.
(ii) Q(Z ⋊θ P ) is the closed linear span of {umsps∗qu−n m, n ∈ Z, p, q ∈ P}, see
(iii) Q(Z ⋊θ P ) is a unital UCT Kirchberg algebra, see [Sta15, Example 3.29(a)].
(iv) The canonical representation π : Q(Z ⋊θ P ) → L(ℓ2(Z)) is faithful.
(v) The unitary u in Q(Z ⋊θ P ) generates a copy of C∗(Z).
(vi) The diagonal subalgebra DS of Q(Z⋊θ P ) is generated by the family of commuting
projections {unsps∗pu−n (n, p) ∈ Z ⋊θ P}. Its spectrum is given by the S-adic
completion of Z.
(vii) There is a natural gauge action γ of the S-dimensional torus on Q(Z ⋊θ P )
given by γχ(u) = u, γ(sp) = χ(p)sp for p ∈ S. The fixed-point algebra F for γ is
the Bunce-Deddens algebra of type (Qp∈S p)∞, see [Sta15, Example 3.29(a)].
(viii) There is another distinguished subalgebra:
.
= C∗({unsp p ∈ S, 0 ≤ n ≤ p − 1}) ⊂ Q(Z ⋊θ P ).
AS
It is known through [BOS18, Corollary 5.2 and Corollary 5.4] that AS is also a
unital UCT Kirchberg algebra like Q(Z ⋊θ P ), and that the canonical inclusion
AS ֒→ Q(Z ⋊θ P ) yields a split-injection onto the torsion part of K∗(Q(Z ⋊θ P )).
For this reason, AS was named the torsion subalgebra in [BOS18]. It was also
divisor of S − 1 ⊂ N× is 1, see [BOS18, Theorem 6.4].
shown that AS is isomorphic to Np∈S Op if S ≤ 2 or the greatest common
While Example 2.1 promotes the direction of considering actions of higher dimensional
semigroups P on the same group, we can equally well stay with the case of a single
endomorphism, and allow the group to be more complicated than Z:
Example 2.2. Suppose G is a discrete, abelian group, and θ is an injective group en-
domorphism of G with finite cokernel. Then (G, N, θ) form an (irreversible) algebraic
dynamical system. In addition, we shall assume that (G, N, θ) is minimal in the sense of
[Sta15], that is,Tn∈N θn(G) = {1G}. This is equivalent to minimality of the dynamical
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
7
The C∗-algebra Q(G ⋊θ N) then has the following features:
(i) Q(G ⋊θ N) is the closed linear span of {ugsmsn∗uh g, h ∈ G, m, n ∈ N}, see
system formed by the Pontryagin dual bG of G and the dual endomorphism θ, see [Sta].
(ii) Q(G ⋊θ N) is a unital UCT Kirchberg algebra if and only ifTn∈N θn(G) = {1G},
see [Sta15, Corollary 3.28] and [BS16, Corollary 4.14].
[Sta15, Lemma 3.4].
abelian, hence amenable.
(iii) The canonical representation π : Q(G ⋊θ N) → L(ℓ2(G)) is faithful.
(iv) The unitary representation u in Q(G ⋊θ N) generates a copy of C∗(G) as G is
(v) The diagonal subalgebra Dθ of Q(G ⋊θ N) is generated by the commuting pro-
jections {ugsnsn∗u∗g (g, n) ∈ G ⋊θ N}. Its spectrum Gθ is a completion of G
with respect to θ. This Cantor space is actually a compact abelian group, see
[Sta15, Remark 4.1 c)].
(vi) There is a natural gauge action γ of the torus on Q(G ⋊θ N) given by γz(ug) = ug
for g ∈ G, and γ(s) = zs. The fixed-point algebra F for γ is a generalized Bunce-
Deddens algebra, see [Sta15, Proposition 4.2].
(vii) For each transversal T for G/θ(G), that is, a minimal complete set of representa-
.
= C∗({ugs g ∈ T }) of Q(G ⋊θ N). We
tives, we can consider the subalgebra AT
remark that, unlike the case of G = Z from Example 2.1, there is no canonical
choice for T in general. On the other hand, AT is always isomorphic to ONθ
.
with Nθ
= [G : θ(G)] because the generators form a Cuntz family of isometries
due to (III).
We observe that the p-adic ring C∗-algebras form a family of algebras being in the
N)
intersection of the two above mentioned cases. In fact, we have that Qp = Q(Z ⋊θp
(cf. [BOS18, Definition 2.1 and Proposition 2.12]).
The following two propositions about the canonical representation parallel the pre-
liminary results we needed in [ACR18a].
Proposition 2.3. Suppose (G, P, θ) is an algebraic dynamical system for which P is
directed. If (G, P, θ) is minimal, that is, Tp∈P θp(G) = {1G} holds, then the canonical
representation π : Q(G ⋊θ P ) → L(ℓ2(G)) is irreducible.
Proof. Let (ξg)g∈G denote the standard orthonormal basis for ℓ2(G), and M ⊂ ℓ2(G) be
a π(Q(G ⋊θ P ))-invariant closed subspace. Then the orthogonal projection Q onto M
belongs to π(Q(G ⋊θ P ))′. Hence we have SpQξ1G = Qξ1G, i.e. Qξ1G is an eigenvector
Qξ1G ∈ Cξ1G, so that either ξ1G ∈ M or ξ1G ∈ M⊥. In the first case, the invariance of
M under π(ug), g ∈ G, yields ξg ∈ M for all g ∈ G, so that M = ℓ2(G). The second
case is analogous as M⊥ is necessarily π(Q(G ⋊θ P ))-invariant as well.
for each Sp with eigenvalue 1. The condition Tp∈P θp(G) = {1G} now implies that
(cid:3)
In the case of Example 2.1, we can also describe the decomposition of the induced
representation πAS of AS into irreducible subrepresentations. Indeed, the arguments
from [ACR18a, Propositon 2.6 -- Corollary 2.9] carry over verbatim, where one has to
replace O2 by AS. Letting E+, E− ∈ L(ℓ2(Z)) denote the orthogonal projection onto
ℓ2(Z
≥0) and ℓ2(Z<0), respectively, we arrive at:
8
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
Proposition 2.4. For (G, P, θ) as in Example 2.1, the representation of AS obtained as
the restriction of the canonical representation π : Q(Z ⋊ P ) → L(ℓ2(Z)) decomposes into
.
two disjoint, irreducible representations π+
= E−πE−. In particular,
we have π(AS)′ = CE+ ⊕ CE− and π(AS)′′ = L(ℓ2(Z
.
= E+πE+ and π−
≥0)) ⊕ L(ℓ2(Z<0)).
3. Maximal abelian subalgebras
3.1. The group C*-algebra. In this subsection, we first restrict to the following setup
for showing that C∗(G) is maximal abelian in Q(G ⋊θ P ), see Theorem 3.5. Recall that
.
an algebraic dynamical system (G, P, θ) is said to be of finite type if Np
= [G : θp(G)] is
finite for all p ∈ P .
Example 3.1. Suppose (G, P, θ) is a commutative algebraic dynamical system of finite
type with directed P , such that the canonical representation π : Q(G ⋊θ P ) → L(ℓ2(G))
is faithful. In addition, assume that for all p, q ∈ P, p 6= q there exists g ∈ G of infinite
.
= θp(h)θq(h−1) is
order such that the group endomorphism Φp,q of G given by Φp,q(h)
injective on hgi ∼= Z.
The last part of the assumptions in Example 3.1 holds for instance if G is not a
torsion group and Φp,q is injective for all distinct p, q ∈ P . All the assumptions in
Example 3.1 are satisfied by all examples described in Example 2.1 or Example 2.2 (due
to minimality). From now on, we assume that (G, P, θ) is as specified in Example 3.1.
Lemma 3.2. For all distinct p, q ∈ P , there exists g ∈ G such that the point spectrum
of π(uΦp,q(g)) ∈ L(ℓ2(G)) is empty.
Proof. Let g ∈ G have the property described in Example 3.1 for given p 6= q. Then the
element Φp,q(g) = θp(g)θq(g−1) still has infinite order, as 1G = (θq(g−1)θp(g))n = Φp,q(gn)
forces gn = 1G. The conclusion then follows from the general observation that the point
spectrum of π(uk) is empty whenever k ∈ G is of infinite order. Indeed, suppose there
6= 0. Then we would
Thus we would have ckmh′ = ch′ 6= 0 for all m ∈ Z. Since k is of infinite order, this
contradicts ξ ∈ ℓ2(G), and we apply this to Φp,q(g).
was an eigenvector ξ =Ph∈G chξh ∈ ℓ2(G) of π(uk), and say ch′
get Ph∈G ck−1hξh = π(uk)ξ = λξ = Ph∈G λchξh for some λ ∈ T (as uk is a unitary).
(cid:3)
Denote by W ∗(G) ⊂ L(ℓ2(G)) the group von Neumann algebra of G. Then there
exists a conditional expectation E1 : L(ℓ2(G)) → W ∗(G), see [KS59].
Proposition 3.3. If (G, P, θ) is as described in Example 3.1, then every conditional
expectation E : L(ℓ2(G)) → W ∗(G) satisfies
(3.1)
In particular, every conditional expectation E : L(ℓ2(G)) → W ∗(G) restricts to the con-
ditional expectation E0 : π(Q(G ⋊θ P )) → π(C∗(G)) given by
E ◦ π(ugsps∗qu∗h) = δp,q N−1
for all g, h ∈ G, p, q ∈ P.
p ugh−1
Proof. Since G is abelian, W ∗(G) = π(C∗(G))′ is abelian so that E ◦ π(ugsps∗qu∗h) =
π(ugh−1)E ◦ π(sps∗q) for all g, h ∈ G, p, q ∈ P . In particular, (III) implies
1 = E ◦ π(cid:0) Pg∈G/θp(G)
ugsps∗pu∗g(cid:1) = Np E ◦ π(sps∗p),
E0 ◦ π(ugsps∗qu∗h) = δp,q N−1
p π(ugh−1)
for all g, h ∈ G, p, q ∈ P.
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
9
so that E ◦ π(ugsps∗pu∗g) = N−1
p .
On the other hand, for p 6= q, (I) yields
π(uθq(g))E(π(sps∗q)) = E(π(sps∗quθq(g))) = E(π(uθp(g)sps∗q)) = π(uθp(g))E(π(sps∗q)),
which is equivalent to (1− π(uΦp,q(g)))E(π(sps∗q)) = 0. By Lemma 3.2, the value 1 is not
an eigenvalue of π(uΦp,q(g)) ∈ L(ℓ2(G)) for g ∈ G chosen according to the hypothesis in
Example 3.1 for p 6= q. Thus 1 − π(uΦp,q(g)) is injective, which allows us to conclude
that E ◦ π(sps∗q) = 0.
Remark 3.4. Due to the faithfulness of the canonical representation, the proof of Propo-
sition 3.3 in fact shows that π−1 ◦ E0 ◦ π : Q(G ⋊θ P ) → C∗(G) is the unique conditional
expectation from Q(G ⋊θ P ) onto C∗(G).
Theorem 3.5. Suppose (G, P, θ) satisfies the conditions in Example 3.1. If (G, P, θ) is
minimal, then the C∗-algebra C∗(G) is maximal abelian in Q(G ⋊θ P ).
Proof. It will be convenient to work in the representation π : Q(G ⋊θ P ) → L(ℓ2(G)).
Let x ∈ C∗(G)′ ∩ Q(G ⋊θ P ) and choose a sequence (xk)k∈N ⊂ span{ugsps∗quh g, h ∈
G, p, q ∈ P} with kx − xkk → 0. Then Proposition 3.3 entails
kπ(x) − E0(π(xk))k = kE1(π(x)) − E0(π(xk))k = kE0(π(x − xk))k ≤ kx − xkk → 0
is closely linked to the condition Tp∈P θp(G) = {1G} in the situation of Example 3.1,
because π(x) ∈ π(C∗(G))′ = W ∗(G) and kE0 ◦ πk ≤ 1. Since E0 ◦ π expects onto
π(C∗(G)) and π is faithful, we conclude that x ∈ C∗(G).
Remark 3.6. The canonical representation is clearly faithful if Q(G⋊θP ) is simple, which
see for instance [BS16, Remark 4.15].
In the case where P is free abelian, minimal-
ity of (G, P, θ) was already known to imply that C∗(G) ⊂ Q(G ⋊θ P ) is a maximal
abelian subalgebra, see [Sta, Corollary 2.7, Theorem 5.6 and Proposition 6.1]. More-
over, minimality of (G, P, θ) is equivalent to simplicity of Q(G ⋊θ P ) in this situation,
see [Sta, Corollary 6.2].
(cid:3)
(cid:3)
It is certainly an interesting task to investigate under what conditions C∗(G) is not
only maximal abelian, see Theorem 3.5, but in fact a Cartan subalgebra of Q(G ⋊θ P ).
In the case of the diagonal D, this question has a clear and affirmative answer, as we
shall see in the next subsection.
3.2. The diagonal. Let us now turn our attention to the diagonal subalgebra D of
Q(G ⋊θ P ), that is, the commutative subalgebra of Q(G ⋊θ P ) generated by the projec-
tions eg,p
.
= ugsps∗pu∗g with (g, p) ∈ G ⋊θ P , see [Sta15, Lemma 3.5]. We note that
D = span{ugsps∗pug−1 (g, p) ∈ G ⋊θ P}.
In many cases, there exists a faithful conditional expectation Θ : Q(G ⋊θ P ) → D given
by ugsps∗qu∗h 7→ δg,hδp,qeg,p, for instance if G and P are abelian so that we obtain Θ as
the composition of two faithful conditional expectations obtained from averaging over
the gauge actions of the duals of the Grothendieck group of P and of the dual group of
G using Q(G ⋊θ P ) ∼= (D ⋊ G) ⋊ P , see [Sta15, Corollary 3.22]. In order to recover the
maximality of D in this general setting we need to make an extra assumption:
(3.2)
For all p, q ∈ P, p 6= q, every g ∈ G has finitely many preimages under Φp,q,
10
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
where Φp,q(h) = θp(h)θq(h)−1, see Example 3.1. Equation 3.2 holds for example for the
integral dynamics (Z, P, θ) described in Example 2.1: given m ∈ Z and p, q ∈ N× with
p > q, the only solution to m = pn− qn is of course n = (p− q)−1m (which belongs to Z
in case m ∈ (p − q)Z). Thus the sets are not only finite, but singletons. However, (3.2)
does not hold as soon as there is p ∈ P \{1P} for which the endomorphism θp fixes some
group element h 6= 1G of infinite order. Indeed, we then get pp 6= p by left cancellation
for P , but 1G = θpp(h)θp(h)−1. For instance, this happens for (Z2, θ, N) given by the
diagonal matrix θ1 = diag(p, 1) with p ∈ Z \ {−1, 0, 1}.
Lemma 3.7. Let (G, P, θ) be an algebraic dynamical system with directed P , and let E
.
be the conditional expectation from L(ℓ2(G)) onto ℓ∞(G) given by E(V )ξg
= hV ξg, ξgiξg.
If (3.2) holds, then the restriction of Eπ(Q(G⋊θP )) coincides with π ◦ Θ up to compact
operators.
Proof. We observe that for g, h, k ∈ G and p, q ∈ P ,
E(π(ugsps∗qu∗h))ξk =(ξk
0
, if h−1k ∈ θq(G) and k = gθp(θ−1
, otherwise.
q (h−1k)),
In the case of p = q, the second part in the first condition amounts to g = h so that the
operator needs to be diagonal. This is exactly what happens for Θ. For p 6= q, assume
k = hθq(ℓ) for some ℓ ∈ G. Then the second part turns into g−1h = θp(ℓ)θq(ℓ)−1.
By (3.2), there are only finitely many ℓ ∈ G that satisfy this, so E(π(ugsps∗qu∗h)) and
π(Θ(ugsps∗qu∗h)) differ by a finite rank operator. Hence the claim follows from Q(G ⋊θ
P ) = span{ugsps∗qu∗h (g, p), (h, q) ∈ G ⋊θ P}.
Remark 3.8. In particular, if in the situation of Lemma 3.7 the algebra π(Q(G ⋊θ P ))
does not contain nontrivial compacts, then E(π(x)) = π(Θ(x)) for every x ∈ Q(G ⋊θ P )
with E(π(x)) ∈ π(Q(G ⋊θ P )). This is for instance true if Q(G ⋊θ P ) is purely infinite
and simple.
(cid:3)
Lemma 3.9. If (G, P, θ) is a minimal algebraic dynamical system with directed P , then
π(D)′ = π(D)′′ = ℓ∞(G).
Proof. By the minimality assumption, the net of projections (π(ugsps∗pu∗g))p∈P ⊂ D con-
verges strongly to the projection onto Cξg for every g ∈ G. Since π(D)′′ is a von Neu-
mann algebra and hence strongly closed, we deduce that all these minimal projections
belong to π(D)′′, and hence π(D)′′ = ℓ∞(G). Now π(D)′ = π(D)′′′ = ℓ∞(G)′ = ℓ∞(G)
proves the remaining assertion.
(cid:3)
Theorem 3.10. Suppose (G, P, θ) is a minimal algebraic dynamical system with directed
P such that Θ : Q(G ⋊θ P ) → D is a faithful conditional expectation and (3.2) holds.
Then the diagonal D is a Cartan subalgebra of Q(G ⋊θ P ).
Proof. By assumption, Θ is a faithful conditional expectation. Moreover, this allows
us to deduce simplicity and pure infiniteness for Q(G ⋊θ P ) out of minimality, that is,
In particular, the canonical representation π is faithful. To show that D is maximal
abelian in Q(G ⋊θ P ), we follow the strategy of [ACR18a, Theorem 3.9]: Suppose
x ∈ D′∩Q(G ⋊θ P ). By Lemma 3.9, the operator π(x) belongs to ℓ∞(G)∩π(Q(G ⋊θ P )),
T(g,p)∈G⋊θP gθp(G)g−1 = {1G}, very much in the way [Sta15, Theorem 3.26] is proven.
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
11
so that Remark 3.8 implies that π(x) = E(π(x)) = π(Θ(x)). Finally, the normalizer of
D in Q(G ⋊θ P ) generates Q(G ⋊θ P ) since one has, for all h, g ∈ G, p, q ∈ P ,
uh(ugsps∗pu∗g)u∗h = ughsps∗pu∗gh,
sq(ugsps∗pu∗g)s∗q = uθq(g)spqs∗pqu∗θq(g)
s∗q(ugsps∗pu∗g)sq =(ug1sq′s∗p′ug2(ug1sq′s∗p′ug2)∗
0
=(ug1sq′s∗q′u∗g1 ∈ D if g = θq(g1)θp(g2), pP ∩ qP = pp′P, pp′ = qq′
otherwise
0
showing that uh, sq ∈ ND(Q(G ⋊θ P )).
(cid:3)
For the sake of completeness we would like to observe that an elegant, but less elemen-
tary, proof of a strengthening of the above result can be achieved as follows, compare
[ACR18a, end of Section 3.1]: Under the assumptions that G is amenable and P embeds
into an amenable group H (for simplicity take H = hPi), the algebra Q(G ⋊θ P ) is a
reduced partial transformation groupoid C∗-algebra. If G ⋊θ P happens to be minimal,
then the corresponding partial action of G ⋊ H is topologically free. The diagonal D is
then seen to be a Cartan subalgebra in Q(G ⋊θ P ) as an application of [Ren08, Propo-
sition 3.1]. In addition, let us mention that [LR17] applies to our situation, showing
that there are infinitely many pairwise inequivalent Cartan subalgebras of Q(G ⋊θ P ),
because the latter will be a UCT Kirchberg algebra in many of the above cases.
In hindsight, we thus see that (3.2) is an artifact of our strategy of proof here, but
we still find it interesting that a variant of it enters the stage for both commutative
subalgebras D and C∗(G), see Example 3.1.
4. The relative commutant of a family of generating isometries
Throughout this section we shall assume that (G, P, θ) is an algebraic dynamical
system with P abelian and Tp∈P θp(G) = {1G}. Requiring P to be abelian grants us
access to the Grothendieck group P −1P of P , which is a discrete abelian group. We
denote by T its Pontryagin dual, which is a compact abelian group. The group T acts
on Q(G ⋊θ P ) by a gauge action γ with γχ(ugsp) = χ(p) ugsp, and we denote by F the
fixed-point algebra for γ, see Example 2.1 (vii) and Example 2.2 (vi). We remark that
F = span{ugsps∗pu∗h g, h ∈ G, p ∈ P}.
possible, see Theorem 4.6.
The aim of this section is to show that C∗({sp p ∈ P})′ ∩ Q(G ⋊θ P ) is as small as
If P is abelian, it is in particular directed with respect to p ≥ q ⇔ p ∈ qP , so there
is a sequence (pn)n∈N ⊂ P with pn+1 ∈ pnP for all n such that for every p ∈ P we have
pn ∈ pP for n large enough. We shall need the following result in Lemma 4.5.
Lemma 4.1. For x ∈ F , the sequence (s∗pmxspm)m∈N converges to a limit in C.
Proof. First, let x ∈ span{ugsps∗puh g, h ∈ G, p ∈ P}. As x is a finite linear combination
of elements of the form ugsps∗puh, there is m0 ∈ N such that pm ∈ pP for all p that appear
and all m ≥ m0. Therefore, we get s∗pmxspm ∈ C[G] ⊂ Q(G ⋊θ P ) for all m ≥ m0 by
(II), more precisely s∗pmxspm = Pg∈F dgug with dg ∈ C for a finite set F ⊂ G. As
Tp∈P θp(G) = {1G}, we can choose p ∈ P such that g /∈ θp(G) for all g ∈ F \ {1G}. Let
12
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
.
= cx ∈ C, using (II).
.
= {(k, m) ∈ N2 m ≥ mk} and xk,m
m′ ∈ N be large enough so that pm ∈ pm0pP for all m ≥ m′. For such m, we then have
s∗pmxspm = d1
Now let x ∈ F and pick a sequence (xk)k∈N ⊂ span{ugsps∗puh g, h ∈ G, p ∈ P} with
xk → x. Let mk ∈ N be an m′ from the first part for xk, and define a net (xk,m)(k,m)∈Λ
.
= s∗pmxkspm. Then (xk,m)(k,m)∈Λ is a Cauchy
by Λ
net (in C) as
kxk,m − xℓ,nk = cxk − cxℓ = kxk,m+n − xℓ,m+nk = ks∗pm+n(xk − xℓ)spm+nk ≤ kxk − xℓk
and xk → x. Therefore, xk,m → c for some c ∈ C because C is complete. From this we
easily deduce s∗pmxspm → c ∈ C as ks∗pmxspm − xk,mk ≤ kx − xkk → 0.
(cid:3)
Recall that, given p, q ∈ P , an element r ∈ P satisfying pP ∩ qP = rP is called a
right least common multiple (right LCM) of p and q. Commutativity in P implies that
we do not need to distinguish between right and left multiples, but more importantly,
the notion of a greatest common divisor (GCD) also makes sense for elements in P : For
p, q ∈ P , an element r ∈ P is a common divisor if p, q ∈ rP . An element d ∈ P is
said to be the greatest common divisor (GCD) for p and q, if every common divisor d
of p and q satisfies d ∈ dP . For every pair p and q, the GCD exists by an application
of Zorn's Lemma, and it is unique up to multiplication by the subgroup of invertible
elements P ∗. The GCD relates to the (right) LCM as follows:
Proposition 4.2. Let P be an abelian right LCM semigroup. If d is a greatest common
divisor and m a least common multiple for p, q ∈ P , then mP ∗ = d−1pqP ∗.
Proof. As P is abelian and d−1p, d−1q ∈ P , we have d−1pq ∈ pP ∩qP = mP . Conversely,
let a, b ∈ P with m = pa = qb. Since pq ∈ mP , there is c ∈ P such that pq = mc =
pac = qbc. By left cancellation, this says that q = ac and p = bc, that is, c is a common
divisor of p and q. Thus we get d = ce for some e ∈ P , and hence m = pd−1qe ∈ pd−1qP .
As P is right LCM, this shows mP ∗ = d−1pqP ∗.
(cid:3)
If P = N×, the claim in the previous proposition is nothing but the well-known
equality nm = GCD(n, m) · lcm(n, m) for all n, m ∈ N.
The set P (⊥) ⊂ P × P and the maps F(p,q) : Q(G ⋊θ P ) → F for (p, q) ∈ P (⊥)
introduced in the following remark will be crucial to the proof of Lemma 4.4, which is
the heart of the whole section and goes all the way back to [Cun77, Proposition 1.10].
.
Remark 4.3. Let P (⊥)
= {(p, q) ∈ P × P pP ∩ qP = pqP} denote the collection of all
relatively prime pairs in P . The terminology alludes to the fact that pP ∩ qP = pqP is
equivalent to the GCD of p and q being in P ∗, see Proposition 4.2. For all (p, q) ∈ P (⊥),
the isometries sp and sq in Q(G ⋊θ P ) doubly commute, i.e.
(4.1)
s∗psq = sqs∗p,
as easily follows by (II), applied to g = g1 = g2 = 1G, p′ = q and q′ = p.
Two pairs (p, q), (p, q) ∈ P (⊥) satisfy p−1q = p−1 q if and only if there is x ∈ P ∗ such
that p = xp and q = xq. This relation defines an equivalence relation on P (⊥) × P (⊥)
that we denote by ∼. For each (p, q) ∈ P (⊥), we define a contractive, linear map
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
13
F(p,q) : Q(G ⋊θ P ) → F by a 7→RT γχ(spas∗q) dχ, where the integration is with respect
to the normalized Haar measure on the compact abelian group T . We observe that
(4.2)
F(p,q)(a) = sps∗pF(p,q)(a)sqs∗q
for all a ∈ Q(G ⋊θ P ), (p, q) ∈ P (⊥).
Moreover, we note that
(4.3)
Finally, for every ugsp′s∗q′uh with g, h ∈ G, p′, q′ ∈ P , there is a unique [(p, q)] ∈ P (⊥)/∼
such that
s∗pF(p,q)(a)sq = s∗pF(p,q)(a)sq whenever (p, q) ∼ (p, q).
ugsp′s∗q′uh = s∗pspugsp′s∗q′uhs∗qsq = s∗pF(p,q)(ugsp′s∗q′uh)sq,
obtained as p′P ∩ q′P = p′pP, p′p = q′q (a least common multiple of p′ and q′). Thus for
every a ∈ span{ugsp′s∗q′uh g, h ∈ G, p′, q′ ∈ P}, there is a uniquely determined finite
set A(a) ⊂ P (⊥)/∼ with the property that
(4.4)
s∗pF(p,q)(a)sq,
a = P[(p,q)]∈A(a)
which is well defined by (4.3).
Lemma 4.4. If a ∈ Q(G ⋊θ P ) satisfies F(p,q)(a) = 0 for all (p, q) ∈ P (⊥), then a = 0.
Proof. Suppose a ∈ Q(G ⋊θ P ) satisfies F(p,q)(a) = 0 for all (p, q) ∈ P (⊥). We fix a
faithful representation ϕ : Q(G ⋊θ P ) → L(H) on some Hilbert space H. Precomposing
by a gauge automorphism ϕ◦γχ : Q(G⋊θP ) → L(H) is then still a faithful representation
for every χ ∈ T . Given ξ, η ∈ H with kξk = kηk = 1, define f : T → C by
f (χ)
.
= hϕ(γχ(a))ξ, ηi.
fk(χ)
.
= hϕ(γχ(ak))ξ, ηi.
Pick a sequence (ak)k∈N ⊂ span{ugsp′s∗q′uh g, h ∈ G, p′, q′ ∈ P} with ak → a, and
define fk : T → C by
cp−1q,kχ(p−1q)
As kf − fkk∞ ≤ ka − akk → 0, (fk)k∈N converges uniformly to f on T . According to
(4.4), there is a uniquely determined sequence of finite subsets (A(ak))k∈N ⊂ P (⊥)/∼
such that ak =P[(p,q)]∈A(ak) s∗pF(p,q)(ak)sq for all k ≥ 1. This leads us to
fk(χ) = P[(p,q)]∈A(ak)hϕ(γχ(s∗pF(p,q)(ak)sq))ξ, ηi = P[(p,q)]∈A(ak)
.
.
= hϕ(s∗pF(p,q)(ak)sq)ξ, ηi as F(p,q)(ak) ∈ F . We set cp−1q,k
= 0 for [(p, q)] /∈
for cp−1q,k
A(ak). Note that if [(p1, q1)], [(p2, q2)] ∈ A(ak) are distinct, then p−1
2 q2 in the
group P −1P as (pi, qi) ∈ P (⊥) for i = 1, 2. Therefore, we can interpret cp−1q,k as the
Fourier coefficient of ak for p−1q. Every element g ∈ P −1P can be described as p−1q for
some (p, q) ∈ P (⊥) by removing the GCD from any given expression as a quotient of
two elements from P . Since (fk)k∈N converges uniformly to f , the Fourier coefficients
.
= hϕ(s∗pF(p,q)(a)sq)ξ, ηi of f for all
(cp−1q,k)k∈N converge to the Fourier coefficients cp−1q
p−1q ∈ P −1P . But then F(p,q)(a) = 0 for all (p, q) ∈ P (⊥) forces cp−1q,k → cp−1q = 0
for all p−1q ∈ P −1P , so that f (χ) = limk→∞ fk(χ) = 0 for all χ ∈ T . As ξ and η were
arbitrary and ϕ ◦ γχ was faithful, we get a = 0.
1 q1 6= p−1
We denote by C∗(P ∗) the C∗-algebra of the abelian group P ∗, and think of C∗(P ∗)
(cid:3)
as the subalgebra of Q(G ⋊θ P ) generated by the unitaries sp, p ∈ P ∗.
14
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
Lemma 4.5. Suppose that P ∗ is finite.
wspw∗ = zpsp with zp ∈ T for all p ∈ P , then w ∈ C∗(P ∗) and zp = 1 for all p ∈ P .
Proof. For every r ∈ P , (4.1) allows us to compute
If w ∈ Q(G ⋊θ P ) is a unitary satisfying
s∗rγχ(spws∗q)sr = zpzqγχ(s∗rs∗qwspsr) = zrzpzqγχ(s∗qs∗rsrwsp) = zrγχ(spws∗q)
for all (p, q) ∈ P (⊥). Therefore we get s∗rF(p,q)(w)sr = zrF(p,q)(w) for all r ∈ P and all
(p, q) ∈ P (⊥). As F(p,q)(w) ∈ F by definition, Lemma 4.1 implies that
(s∗pmF(p,q)(w)spm)m∈N = (zpm)m∈NF(p,q)(w)
converges to a limit in C. However, for every (p, q) ∈ P (⊥) with p /∈ P ∗ or q /∈ P ∗,
(4.2) forces F(p,q)(w) = 0 as sps∗p or sqs∗q is a proper subprojection of 1 and hence not in
s∗pF(p,q)(w)sq,
C. Since P ∗ is finite, we can consider the element a
which then satisfies F(p,q)(a) = 0 for all (p, q) ∈ P (⊥). Lemma 4.4 now implies
= w −P[(p,q)]∈P (⊥)/∼,
p,q∈P ∗
.
w =
P[(p,q)]∈P (⊥)/∼,
p,q∈P ∗
s∗pF(p,q)(w)sq ∈ C∗(P ∗)
as F(p,q)(w) ∈ C by the previous part. Since P is abelian, w commutes with every sp,
so zp = 1 for all p ∈ P .
(cid:3)
We are now ready to prove the announced theorem.
Theorem 4.6. Suppose that P ∗ is finite. The relative commutant C∗({sp p ∈ P})′ ∩
Q(G ⋊θ P ) equals C∗(P ∗). In particular, if there are no non-trivial elements in P ∗, then
C∗({sp p ∈ P})′ ∩ Q(G ⋊θ P ) = C.
Proof. Being a unital C∗-algebra, the relative commutant C∗({sp p ∈ P})′∩Q(G ⋊θ P )
is the linear span of its unitaries. According to Lemma 4.5, every such unitary belongs
to C∗(P ∗). The reverse inclusion is clear as P is abelian.
(cid:3)
Remark 4.7. The assumption in Theorem 4.6 that the subgroup P ∗ of the invertible
elements in P needs to be finite is likely to be dispensable. For instance, this hypothesis
could be got rid of by establishing a Fej´er-type theorem for the series
P[(p,q)]∈P (⊥)/∼,
p,q∈P ∗
s∗pF(p,q)(w)sq
which will in general fail to be norm convergent, even if w is as above. Nevertheless,
we would expect the series to be Ces`aro summable once a suitable way to count the
elements of P ∗ is introduced. This aspect lies outside the scope of the present work, but
we plan to address it in future studies.
Remark 4.8. In the context of integral dynamics, see Example 2.1, the relative commu-
tant C∗({sp p ∈ T})′∩Q(Z ⋊ P ) for T ⊂ S might also be worth computing. We would
expect it to equal C∗({sp p ∈ S \ T}).
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
15
5. Extendability of Bogolubov automorphisms
For every 2 ≤ p < ∞ the C∗-algebra Qp = Q(Z ⋊θp
N) is by now known as the p-adic
ring C∗-algebra. Inside each Qp, there is a copy of the Cuntz algebra Op, generated by
(uisp)0≤i≤p−1. For convenience, let us from now on denote these generating isometries
uisp of Op ⊂ Qp by Ti. The case p = 2 has already been studied in detail in [LL12,
ACR18a]. A special feature of Q2 is that the two isometries T0 and T1 are intertwined
by u, i.e. T0u = uT1. For a general p, however, T0 and Tp−1 are still intertwined by u,
but we cannot expect all the isometries Ti to be unitarily equivalent to one another. For
instance, consider p = 3 and the faithful canonical representation π : Q3 → L(ℓ2(Z)).
Here, both π(T0) and π(T2) have eigenvalue 1 corresponding to the eigenvectors ξ0
and ξ−1, respectively. However, the point spectrum of π(T1) is empty. Indeed, if ξ =
Pm∈Z cmξm ∈ ℓ2(Z) satisfies π(T1)ξ = λξ for some λ ∈ T, then Pm∈Z cmξ3m+1 =
Pm∈Z λcmξm. Therefore, cm = c3m+1 for all if m ∈ Z, which forces ξ = 0.
We start our discussion by introducing a distinguished representation of Op, occasion-
ally referred to as the interval picture of Op, which will come in useful in Lemma 5.4.
For 2 ≤ p < ∞, we define a map σ : Op → L(L2([0, 1])) by σ(Ti)(f )(t) = √pf ◦ h−1
i (t)
for t ∈ [i/p, (i + 1)/p] and 0 otherwise, for f ∈ L2([0, 1]) and i = 0, . . . , p − 1, where hi
is the compression
hi : [0, 1] → [i/p, (i + 1)/p]
t
7→ (i + t)/p
where α ranges over multi-indices of length k in the letters {0, 1, . . . , p − 1} and Tα =
Tα1Tα2 · · · Tαk , because ϕ(Tp−1) is pure.
Proposition 5.2. For 2 ≤ p < ∞, the isometries (σ(Ti))0≤i≤p−1 are pure and the repre-
sentation σ of Op extends uniquely to a representationeσ : Qp → L(L2([0, 1])). Moreover,
eσ is not unitarily equivalent to the canonical representation π : Qp → L(ℓ2(Z)).
Proof. We start by observing that the isometry σ(Ti) is pure since σ(Ti)kσ(Ti)∗k is
the projection onto L2(I), where I ⊂ [0, 1] is an interval of length p−k. According to
Remark 5.1 σ thus admits at most one extension to a representation of Qp. In order to
obtain an extension eσ, we note that the each of the following two families of mutually
orthogonal projections
{σ(T j
p−1) 0 ≤ i ≤ p− 2, j ≥ 0} and {σ(T j
0 ) 1 ≤ i ≤ p− 1, j ≥ 0}
p−1Ti T ∗i T ∗j
0 Ti T ∗i T ∗j
and the Hilbert space L2([0, 1]) is defined w.r.t.
the Lebesgue measure. Observing
that the adjoint of σ(Ti) is given by σ(Ti)∗(f ) = 1√pf ◦ hi, we see that σ(Ti)∗σ(Ti) =
δi,j idL2([0,1]) and σ(Ti)σ(Ti)∗ = idL2([i/p,(i+1)/p]). Hence σ defines a representation of Op.
Remark 5.1. If ϕ : Op → L(H) is a unital representation on some Hilbert space H such
that ϕ(Tp−1) is a pure isometry (namely the range projections of its powers converge
strongly to 0), then ϕ has at most one extension to a representation of Qp. Obviously,
any extension is completely determined by the image of the unitary u. Now if U, W ∈
U(H) are such that either of them yields an extension of ϕ to a representation of Qp,
then W ϕ(Ti) = ϕ(Ti+1) = Uϕ(Ti) for all 0 ≤ i ≤ p − 2, W ϕ(Tp−1) = ϕ(T0)W , and
Uϕ(Tp−1) = ϕ(T0)U lead to
(U − W )ϕ(Tα)ϕ(Tα)∗ = lim
k→∞
(U − W )ϕ(Tp−1)kϕ(Tp−1)∗k = 0,
U − W = lim
k→∞ Pα=k
16
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
sums up to the identity due to the pureness of Tp−1 and T0, respectively. We then define
U ∈ L(L2([0, 1])) by
(5.1)
.
= σ(T j
Uσ(T j
0 Ti+1T ∗i T ∗j
p−1)
for 0 ≤ i ≤ p − 2, j ≥ 0.
p−1Ti T ∗i T ∗j
p−1)
As
UU∗ = P1≤i≤p−1,
j≥0
0 Ti T ∗i T ∗j
σ(T j
U∗U = U∗ P1≤i≤p−1,
j≥0
σ(T j
0 Ti T ∗i T ∗j
0 ) = 1,
and
0 )U = P0≤i≤p−2,
j≥0
σ(T j
p−1Ti T ∗i T ∗j
p−1) = 1,
j≥0
σ(T j+1
σ(T j+1
0 Ti T ∗i T ∗j
p−1)U∗ = P1≤i≤p−1,
the operator U is a unitary. For i = 0, . . . , p − 2, we clearly have Uσ(Ti) = σ(Ti+1) by
(5.1), and Uσ(Tp−1) = σ(T0)U follows from
p−1 Ti T ∗i T ∗j
Uσ(Tp−1)U∗ = U P0≤i≤p−2,
Finally, the representationeσ is not unitarily equivalent to π because the point spectrum
of the pure isometries is empty, whereas π(Tp−1) has eigenvalue 1 corresponding to the
eigenspace spanned by ξ−1 ∈ ℓ2(Z).
Remark 5.3. We provide an explicit description of the operator U for p = 2. As it
.
= [0, 1
turns out to be defined piecewise, we need to set some notation. We define I0
2 ],
2j ], k ≥ 0. Given any f ∈ L2([0, 1]) we
Ik
= [Pk
can write f =P∞k=0 fk, where fk = f χIk. Then Uf is given by the function g =P∞k=0 gk,
where gk = gχJk with
0 ) = σ(T0).
.
= [ 1
2j+1 , 1
j≥0
l=1
(cid:3)
.
1
1
l=1
2l ], k ≥ 1, and Jk
2l ,Pk+1
gk(s) = fk(s +Pk−1
g0(s) = f0(s − 1
2)
g1(s) = f1(s + 1
4)
l=1
2 , 1]
4 , 1
2 ]
s ∈ [ 1
s ∈ [ 1
s ∈ Jk, k ≥ 2
1
2l + 1
2k+1 )
In the sequel, we shall refer to the representationeσ : Qp → L(L2([0, 1])) from Propo-
sition 5.2 as the interval picture of Qp.
Pp−1
A distinguished family of automorphisms of Op is given by the Bogolubov auto-
.
morphisms λA associated with a unitary matrix A = (ai,j) ∈ Up(C), where λA(Ti)
=
j=0 aj,iTj. As λA ◦ λB = λAB, the map Up(C) ∋ A 7→ λA ∈ Aut(Op) is a repre-
sentation of the group Up(C). Gauge automorphisms are obviously a special case of
Bogolubov automorphisms, i.e. those coming from diagonal matrices with entries in T.
The aim of this section is to determine which Bogolubov automorphisms of Op extend
to automorphisms of Qp. In addition, we show that all these Bogolubov automorphisms
admit a unique extension to an automorphism of Qp, see Theorem 5.8: The group of
extendible Bogolubov automorphisms is generated by the gauge automorphisms and
the exchange automorphism described in Definition 5.6. Moreover, we show that the
extensions of all nontrivial gauge automorphisms of Op are outer automorphisms of Qp,
see Theorem 6.15.
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
17
Lemma 5.4. For all 2 ≤ p < ∞, every Bogolubov automorphism λA of Op admits at
extension of ϕ ◦ λA, which is unique due to Remark 5.1. But if ϕ is non-zero, it is
Proof. Let A = (ai,j) ∈ Up(C). We intend to invoke Remark 5.1, so we need to work
most one extension fλA to an automorphism of Qp.
in a representation eϕ of Qp that extends a non-zero representation ϕ of Op for which
ϕ(λA(Tp−1)) is pure. For if fλA is an extension of λA to Qp, then eϕ ◦fλA will be an
actually an isomorphism, and so is eϕ as both Op and Qp are simple. Therefore, the
extension fλA is also unique.
note the range projection of the k-th power. We observe that λA(Tp−1)k =Pα=k cαTα,
.
= max{ai,p−1 i = 0, 1, . . . p−1} ∈ [0, 1] equals 1, then ai,p−1 = δi,i0 for a unique
0 ≤ i0 ≤ p − 1. According to Proposition 5.2, we can then pick ϕ = σ. The case M < 1
is handled in the canonical representation ϕ = π instead: Let Qk
p−1)) de-
where α is a multi-index with values in {0, 1, . . . , p − 1} and cα = aj1,p−1 . . . ajk,p−1 for
α = (j1, . . . , jk). Then the inequality cα ≤ M k forces kQkξmk ≤ M k → 0 as k → ∞
for every m ∈ Z because M < 1 and T ∗αTβ = 0 for all multi-indices α 6= β of length k.
Therefore, π(λA(Tp−1)) is pure and the proof is complete.
.
= π(λA(T k
p−1T ∗k
If M
(cid:3)
We will now show that very few Bogolubov automorphisms are extendible.
Lemma 5.5. If λA is an extendible Bogolubov automorphism, then fλA(u) ∈ C∗(Z).
Proof. As an intermediate step, we show that fλA(u) ∈ F , for which we proceed by
contradiction. Suppose there exists some z ∈ T \ {1} such that γz(fλA(u)) 6=fλA(u). Set
= γz(fλA(u)). We want to show that Λ is an automorphism of
Λ(sp)
Qp that extends λA. The calculation
.
= λA(sp) and Λ(u)
.
for Ti = uisp ensures that ΛOp = λA. The defining relation of Qp corresponding to (I)
and (III) are satisfied as
Λ(Ti) = Λ(ui)Λ(sp) = γz(fλA(ui))λA(sp) = zγz(λA(Ti)) = λA(Ti)
Λ(sp)Λ(u) = zγz(fλA(spu)) = zγz(fλA(upsp)) = Λ(u)pΛ(sp), and
p−1Pm=0
Λ(u)mΛ(sp)Λ(sp)∗Λ(u)−m =
p−1Pm=0
zzγz(λA(em,p)) = γz(λA(1)) = 1.
0 )xλA(T k
Next, we note that for all x ∈ F′k
Now Lemma 5.4 yields a contradiction. Hence we get fλA(u) ∈ F .
.
= span{umspks∗pku−n m, n ∈ Z}, the element
λA(T ∗k
sequence (xk)k≥1 with xk ∈ F′f (k) for some f : N → N. As F′k ⊂ F′k+1 for all k, we can
assume that f is monotone increasing. Now if λA(T ∗k
p−1) → y in Qp, then
y ∈ C∗(Z) as
p−1) belongs to C∗(Z). Let x ∈ F =Sk≥1 F′k and choose an approximating
0 )xλA(T k
0
)xkλA(T f (k)
and λA(T ∗f (k)
0
)xkλA(T f (k)
ky − λA(T ∗f (k)
)xλA(T f (k)
p−1 ) ∈ C∗(Z). Now, we observe that T ∗k
p−1 )k ≤ ky − λA(T ∗f (k)
that fλA(u) ∈ F by the first part and that (λA(T ∗k
sequence equal to fλA(u), we deduce that fλA(u) ∈ C∗(Z).
0 )fλA(u)λA(T k
0
p−1 )k + kx − xkk → 0
0 uT k
p−1 = u. Thus, using
p−1))k≥1 is a constant
(cid:3)
18
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
Definition 5.6. For the exchange matrix E ∈ Mp({0, 1}), that is, E = (δp−i+1,j)1≤i,j≤p,
the associated Bogolubov automorphism λE of Op is called the exchange automorphism.
For the proof of our main result Theorem 5.8, we need the following small variation of
[ACR18a, Proposition A.1], for which we remark that the proof carries over verbatim.
Here we limit ourselves to pointing out that the integer n is the winding number of f ,
which is well defined thanks to compactness of T along with continuity of f .
Proposition 5.7. For all 2 ≤ p < ∞, every f ∈ C(T, T) satisfying f (zp) = f (z)p for
all z ∈ T is of the form f (z) = zn for some n ∈ Z.
Theorem 5.8. A Bogolubov automorphism λA of Op admits an extension to an auto-
morphism of Qp if and only if A belongs to the subgroup {zB z ∈ T, B ∈ {1, E}} ∼=
T × Z/2Z of Up(C). If A = z1, then fλA(u) = u, whereas A = zE implies fλA(u) = u∗.
Proof. By Lemma 5.5, there is f ∈ C(T, T) ∼= U(C∗(Z)) such that f (u) = fλA(u).
Applying fλA to both sides of (I) and using the fact that f (up)sp = spf (u), we easily get
f (u)pλA(Ti) = fλA(u)p+iλA(T0) = fλA(u)iλA(T0)fλA(u) = fλA(u)iλA(T0)f (u)
for all 0 ≤ i ≤ p − 1. Thus relation (III) implies that f satisfies the functional equation
from Proposition 5.7 for p, and hence there is n ∈ Z such that f (z) = zn for all z ∈ T.
= fλA(u)if (up)λA(T0) = f (up)λA(Ti)
Next, we observe that each 1 ≤ i ≤ p − 1 yields an equation
aj,iujsp =
p−1Pj=0
which, applied to ξ0 gives
(5.2)
p−1Pj=0
p−1Pj=0
aj,iTj = λA(Ti) =fλA(ui)λA(T0) =
p−1Pk=0
ak,0ξk+ni
for all 1 ≤ i ≤ p − 1.
aj,iξj =
p−1Pj=0
aj,0uj+nisp,
n = 0.
n = 0: The algebra Qp is simple and fλA 6= 0, so fλA has to be faithful, which excludes
n ≥ 2: The fact that j 6= k + n(p − 1) for all 0 ≤ j, k ≤ p − 1 forces λA(Tp−1) = 0 due
to (5.2), but this is impossible as λA(Tp−1) needs to be an isometry.
n = 1: As j 6= k + n(p − 1) for all 0 ≤ j, k ≤ p − 1 except j = p − 1 and k = 0, (5.2)
implies that ak,0 = zδk,0 for some z ∈ T, as λA(T0) needs to be an isometry. But
then we get
n = −1: Similar to the case of n = 1, we get ak,0 = zδk,p−1 for some z ∈ T by looking
λA(Ti) =fλA(ui)λA(T0) n=1= zuiT0 = γz(Ti)
for all 1 ≤ i ≤ p − 1, so that λA = γz.
at i = p − 1. From this we get
λA(Ti) =fλA(ui)λA(T0) n=−1= zu−iTp−1 = γz(λE(Ti))
for all 1 ≤ i ≤ p − 1, so that λA = γz ◦ λE.
(cid:3)
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
19
Remark 5.9. According to Theorem 5.8, the only way an extendible non-trivial Bogol-
ubov automorphism λA can fix one of the generating isometries (Ti)0≤i≤p−1 is that p is
odd and A is the exchange matrix E, in which case the fixed isometry is T(p−1)/2.
Remark 5.10. In the case of algebraic dynamical systems (G, P, θ) where P has more
than one generator, it is not clear what the right definition of a Bogolubov automorphism
ought to be. A good place to start appears to be the case of integral dynamics, see
Example 2.1, the natural higher-dimensional version of Qp where P ⊂ N× is generated
by a family S of mutually relatively prime natural numbers. In such a situation, there
is a notion of a torsion subalgebra AS = C∗({uisp p ∈ S, 0 ≤ i ≤ p − 1}) ⊂ Q(Z ⋊ P ),
which plays the role of Op ⊂ Qp in many ways, see [BOS18] for details.
6. Automorphisms preserving the group C*-algebra
6.1. Automorphisms fixing the group C*-algebra. In this subsection we consider
automorphisms of Q(G ⋊θ P ) that fix the group C∗-algebra C∗(G) ⊂ Q(G ⋊θ P ) point-
wise. The corresponding subgroup of AutQ(G⋊θ P ) shall be denoted by AutC ∗(G)Q(G⋊θ
P ). We restrict our attention to the situation where (G, P, θ) is an algebraic dynamical
system of finite type such that G and P are abelian. We note that this covers both
Example 2.2 and Example 2.1. In Theorem 6.2 we show that, under this hypothesis,
every element of AutC ∗(G)Q(G⋊θ P ) arises from a suitable family of unitaries in the com-
mutative group C∗-algebra C∗(G), given that C∗(G) is maximal abelian in Q(G ⋊θ P ).
This result generalizes [ACR18a, Theorems 6.13 and 6.14], and is used in Theorem 6.5
to show that AutC ∗(G)Q(G ⋊θ P ) is a maximal abelian subgroup of the automorphism
group of Q(G ⋊θ P ). These results point towards a generalization of the notion of quasi-
freeness that appears in [DS01, Zac00], but we shall not pursue this line of research
here.
In the sequel, we will make use of commutativity of G to identify C∗(G) with C(bG)
via ug 7→ [χ 7→ χ(g)], where bG is the Pontryagin dual of G.
Definition 6.1. A map ψ : P → C(bG, T), p 7→ ψp is said to be a θ-twisted homomor-
phism if it satisfies
(6.1)
ψpq = ψp θp(ψq)
for all p, q ∈ P,
where θp(ug)
.
= uθp(g) for g ∈ G. The collection of all θ-twisted homomorphisms is
denoted by Homθ(P, C(bG, T)).
We note that Homθ(P, C(bG, T)) is an abelian group under pointwise multiplication
.
= (ψp)−1 because θ consists of group homomorphisms of
with inverses given by (ψ−1)p
G.
Theorem 6.2. The map ψ 7→ βψ with βψ(ugsp)
is maximal abelian in Q(G ⋊θ P ), then this homomorphism is an isomorphism and
an injective group homomorphism Homθ(P, C(bG, T)) → AutC ∗(G)Q(G ⋊θ P ). If C∗(G)
.
= ugψpsp for (g, p) ∈ G ⋊θ P defines
EndC ∗(G)Q(G ⋊θ P ) = AutC ∗(G)Q(G ⋊θ P ) ∼= Homθ(P, C(bG, T)).
20
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
Proof. We first verify the defining relations (I) -- (III) from Section 2 for βψ(ug) = ug and
βψ(sp): Relation (I) holds as
βψ(sp)ug = ψpspug = ψpuθp(g)sp = uθp(g)βψ(sp).
For (II), let p, q ∈ P, g ∈ G. Since P is abelian, we let r ∈ P denote a greatest common
divisor of p and q, see Proposition 4.2, and pick q′, p′ ∈ P such that p = rq′, q = rp′, so
that pP ∩ qP = rq′p′P . As q′ and p′ are relatively prime, that is, (q′, p′) ∈ P (⊥) in the
notation of Remark 4.3, and (G, P, θ) is of finite type, [Sta15, Proposition 1.1] implies
θq′(G)θp′(G) = G. By (6.1), we have
ψ∗pψq = θr(ψ∗q′)ψ∗r ψr
θr(ψp′) = θr(ψ∗q′ψp′)
so that, after some computations,
βψ(s∗p)ugβψ(sq) = s∗p
θr(ψ∗q′ψp′)ugsq
= (ug1s∗q′ψ∗q′ψp′sp′ug2
0
, if g = θp(g1)θq(g2) (⇔ g ∈ θr(G))
, otherwise.
Now we observe that s∗q′ψ∗q′ψp′sp′ equals ψp′sp′s∗q′ψ∗q′ as
ψp′sp′s∗q′ψ∗q′ = ψp′s∗q′sp′ψ∗q′ = s∗q′ θq′(ψp′)θp′(ψ∗q′)sp′
and
ψ∗q′ψp′ = θq′(ψp′)θp′(ψ∗q′) ⇐⇒ ψp′ θp′(ψq′) = ψq′p′ = ψq′ θq′(ψp′).
This shows that relation (II) is satisfied. Checking the summation relation (III) amounts
to
Pg∈G/θp(G)
βψ(eg,p) = ψp Pg∈G/θp(G)
eg,p! ψ∗p = 1
because ψp is a unitary. Thus βψ defines an endomorphism of Q(G ⋊θ P ). It is evident
In particular, βψ−1 ◦ βψ = id = βψ ◦ βψ−1 implies that, βψ is
that βψ ◦ βϕ = βψϕ.
an automorphism, and hence an element of AutC ∗(G)Q(G ⋊θ P ) because βψ(ug) = ug
βϕ(sp) = ϕpsp. Multiplying the equation by ug on the left, by s∗pu∗g on the right, where
g ∈ G ranges over a transversal for G/θp(G) leads to
as C∗(G) is abelian. If βψ = βϕ for ψ, ϕ ∈ Homθ(P, C(bG, T)), then ψpsp = βψ(sp) =
ψp = ψp Pg∈G/θp(G)
eg,p! = ϕp Pg∈G/θp(G)
eg,p! = ϕp.
injective.
Thus the group homomorphism Homθ(P, C(bG, T)) → AutC ∗(G)Q(G ⋊θ P ), ψ 7→ βψ is
for some ψ ∈ Homθ(P, C(bG, T)). Note that β is necessarily unital as 1 ∈ C∗(G). For all
As for the second claim, suppose C∗(G) is maximal abelian in Q(G ⋊θ P ), and let
β ∈ EndC ∗(G)Q(G ⋊θ P ). Then both parts of the claim will follow if we show β = βψ
g, h ∈ G and p ∈ P we have
s∗pu∗gβ(sp)uh = s∗pu∗guθp(h)β(sp) = uhs∗pu∗gβ(sp)
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
21
.
= s∗pu∗gβ(sp) is an element
because G is abelian and β fixes C∗(G) pointwise. Thus wg,p
of C∗(G), as the latter is assumed maximal abelian. Then the finite linear combination
wp
.
=Pg∈G/θp(G) ug θp(wg,p) ∈ C∗(G) satisfies
ugsps∗pu∗gβ(sp) = β(sp).
The element wp is a unitary in C∗(G) since wpugsp = ugwpsp = ugβ(sp) = β(ugsp) and
the summation relation for p allow us to compute
ug θp(wg,p)sp = Pg∈G/θp(G)
wpsp = Pg∈G/θp(G)
w∗pwp = wpw∗p = wp Pg∈G/θp(G)
eg,p! w∗p = β Pg∈G/θp(G)
eg,p! = β(1) = 1.
Finally, we claim that ψ(p)
Indeed, given p, q ∈ P we deduce ψ(pq) = ψ(p)θp(ψ(q)) from the equation
.
= wp defines a θ-twisted homomorphism so that β = βψ.
ψ(pq)spq = β(spq) = β(sp)β(sq) = ψ(p)θp(ψ(q))spq
in the same way as we proved injectivity of the group homomorphism. Thus we arrive
at β = βψ.
(cid:3)
The proof of Theorem 6.2 shows that a family of unitaries in C∗(G) defines a ∗-
homomorphism of Q(G ⋊θ P ) if and only if it comes from a θ-twisted homomorphism.
As a strengthening of Theorem 6.2, we will now prove that AutC ∗(G)Q(G ⋊θ P ) is not
only abelian but maximal abelian in AutQ(G ⋊θ P ) in a number of important cases, see
Theorem 6.5.
Lemma 6.3. Let K be a metrizable compact abelian group. If χ ∈ K has the property
that for every f ∈ C(K, T) there exists λf ∈ T with f (χξ) = λf f (ξ) for all ξ ∈ K, then
χ = 1K.
Proof. In fact, this is a particular case of a more general result proved in Theorem 3.6
of [ACR18b] that the same is true for every continuous map Φ on a metrizable compact
space X.
(cid:3)
Remark 6.4. The result recalled in the proof of Lemma 6.3 extends to general C∗-
algebras: every endomorphism ϕ of a unital C∗-algebra A such that ϕ(u) = χuu for
every u ∈ U(A), where χu ∈ T, must be the identity map of A. This is seen as follows.
2(f+(a) + f−(a)) with f+(t) = t + i√1 − t2 and
If a ∈ Asa and kak ≤ 1, then a = 1
f−(t) = t − i√1 − t2. Now ϕ(a) = 1
2(χ+f+(a) + χ−f−(a)) says that C∗(a), which is of
course both separable and commutative, is globally invariant under ϕ. Therefore, the
endomorphism restricts to C∗(a) trivially. This concludes the proof as a ∈ Asa was an
arbitrary self-adjoint contraction.
The following result is a generalization of [ACR18a, Theorem 6.16].
Theorem 6.5. Suppose that (G, P, θ) is an algebraic dynamical system such that G is
abelian. If C∗(G) ⊂ Q(G ⋊θ P ) is maximal abelian and Q(G ⋊θ P ) is simple, then the
group AutC ∗(G)Q(G ⋊θ P ) is maximal abelian in AutQ(G ⋊θ P ).
22
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
deduce that AutC ∗(G)Q(G ⋊θ P ) ∼= Homθ(P, C(bG, T)), which is abelian. Let α be such
Proof. Since C∗(G) ⊂ Q(G ⋊θ P ) is maximal abelian, we can apply Theorem 6.2 to
that α ◦ β = β ◦ α for all β ∈ AutC ∗(G)Q(G ⋊θ P ).
In particular, for β = Ad(x)
with x ∈ U(C∗(G)) we get Ad(α(x)) = α ◦ Ad(x) ◦ α−1 = Ad(x). Thus we have
Ad(x−1α(x)) = id, which implies that x−1α(x) belongs to the center of Q(G ⋊θ P ). As
Q(G ⋊θ P ) is simple, its center is trivial, so there is λx ∈ T with α(x) = λxx for every
x ∈ U(C∗(G)). Since α is a ∗-homomorphism, the values λx arise from a character
χ ∈ bG via α(ug) = χ(g)ug for all g ∈ G. Therefore, reinterpreting α(x) = λxx for
x ∈ U(C∗(G)) as an equation for functions f ∈ C(bG, T) ∼= U(C∗(G)), we arrive at
Since G is countable, its compact abelian Pontryagin dual bG is a metrizable compact
abelian group. Thus Lemma 6.3 implies χ = 1 bG so that α ∈ AutC ∗(G)Q(G ⋊θ P ).
for all ξ ∈ bG, f ∈ C(bG, T).
f (χξ) = λf f (ξ)
(cid:3)
We have chosen this abstract formulation of Theorem 6.5 as we do not have a com-
plete answer to the question when C∗(G) is maximal abelian in Q(G ⋊θ P ) for arbitrary
algebraic dynamical systems, and the precise condition for simplicity is somewhat tech-
nical in the greatest generality, see for instance [Star15, BS16]. Yet the proof we provide
here does not need any of the relatively strong extra hypotheses that enter the proof of
Theorem 3.5.
Corollary 6.6. If (G, P, θ) belongs to the class described in Example 3.1 and Q(G ⋊θ P )
is simple, then AutC ∗(G)Q(G ⋊θ P ) is maximal abelian in AutQ(G ⋊θ P ).
Proof. In the case of Example 2.1, Q(G ⋊θ P ) is known to be simple and (G, P, θ)
satisfies the conditions of Example 3.1. Thus C∗(G) is maximal abelian in Q(G ⋊θ P )
by Theorem 3.5, and Theorem 6.5 yields the claim.
(cid:3)
As a special case of Corollary 6.6, we would like to highlight the case of integral
dynamics:
Corollary 6.7. Suppose that (G, P, θ) satisfies the assumptions of Example 2.1. Then
AutC ∗(Z)Q(Z ⋊ P ) ∼= Homθ(P, C(T, T)) is a maximal abelian subgroup of AutQ(Z ⋊ P ).
Remark 6.8. We would like to point out that a general statement can be made here about
arbitrary inclusions of C∗-algebras B ⊂ A with B is abelian and A having trivial center.
Indeed, one has the equality CAutB A(Aut A) = Z(AutB A), where CAutB A(Aut A) is the
centralizer of AutB A in Aut A and Z(AutB A) the center of AutB A. As a result, the
group AutB A is maximal abelian whenever it is abelian.
6.2. A closer look at integral dynamics. In this subsection, we assume that (G, P, θ)
is of the form specified in Example 2.1 so that we are dealing with subdynamics of Z⋊N×.
In this situation, (6.1) takes the more explicit form
ψp(z)ψq(zp) = ψq(z)ψp(zq)
(6.2)
We observe that if one chooses ψp(z) = f (z) for all generators p ∈ S of P for some
f ∈ C(T, T), then f (z) = c for some constant c ∈ T unless S = 1.
It is not difficult to exhibit non-trivial solutions of this system of functional equations,
but to the best of the authors' knowledge a complete description of all solutions of (6.2)
has yet to be found.
for all z ∈ T and p, q ∈ P.
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
23
.
= u and γt(sp)
The largest specimen for Example 2.1 yields the C∗-algebra QN introduced in [Cun08],
that is, we consider the case P = N×. For every other integral dynamics, the semidirect
product Z ⋊ P embeds into Z ⋊ N×, along with an embedding of the corresponding
C∗-algebra Q(Z ⋊θ P ) into Q(Z ⋊ N×) ∼= QN.
Cuntz remarked in [Cun08, Section 4] that QN is acted upon by a one-parameter
.
= pitsp for p ∈ N×, t ∈ R,
group of automorphisms, given by γt(u)
to which he refers as the canonical action of R. Here we consider the slightly larger
class of automorphisms AutC ∗(Z)Q(Z ⋊ P ). Due to Theorem 3.5, which applies to
Z ⋊ P , Theorem 6.2 implies that we are dealing with automorphisms of the form βψ for
ψ ∈ Homθ(P, C(T, T)).
Proposition 6.9. Let ψ ∈ Homθ(P, C(T, T)). If ψp is a monomial in the generator z
for all p ∈ P , then it is of the form ψp(z) = zk(p−1) for some k ∈ Z.
.
= zkp, then (6.2) turns into kpq = kp +pkq for all p, q ∈ N×.
Proof. Indeed, if we set ψp(z)
It is then straightforward to check that all solutions of this system of integral equations
are of the form kp = k(p − 1) for all p, where k ∈ Z is arbitrary.
(cid:3)
The automorphisms βψk arising from the monomial solutions obtained in Proposi-
tion 6.9 are inner as βψk = Ad(u−k).
Proposition 6.10. Let ψ ∈ Homθ(P, C(T, T)).
constant, then ψp is constant for all p ∈ P .
Proof. By (6.2), we get ψp(z)ψn(zp) = ψn(z)ψp(zn), which reduces to ψp(z) = ψp(zn) for
all p ∈ P as ψn is constant. But the only solutions to this type of functional equation
are constant since n ≥ 2.
If there is n ≥ 2 such that ψn is
(cid:3)
In the second part of this final section, we show that every automorphisms α of
Q(Z ⋊ P ) that fixes a natural subalgebra On = C∗({uksn 0 ≤ k ≤ n − 1}) pointwise
for some n ∈ P, n ≥ 2 is necessarily of the form βψ appearing in Theorem 6.2 for some
ψ ∈ Hom(P, T), so that, in particular, α fixes C∗(Z) pointwise.
Theorem 6.11. For every n ∈ P , 2 ≤ n < ∞, the group AutOnQ(Z⋊P ) is a subgroup of
AutC ∗(Z)Q(Z ⋊ P ) ∼= Homθ(P, C(T, T)). For ψ ∈ Homθ(P, C(T, T)), the automorphism
βψ belongs to AutOnQ(Z⋊P ) if and only if ψ ∈ Hom(P, T) with ψn = 1, that is, βψ = γχ
for some χ ∈ T with χ(n) = 1.
Proof. Let n ∈ P , 2 ≤ n < ∞ and α ∈ AutOnQ(Z ⋊ P ). We work in the canonical
representation π : Q(Z ⋊ P ) → L(ℓ2(Z)) to show that α(u) = u. Noting that em,nk =
umsnks∗nku−m ∈ On for k ≥ 1 if 0 ≤ m ≤ nk − 1, we get
α(u)em,nk = α(uem,nk) = um+1snks∗nku−m = uem,nk
for all 0 ≤ m ≤ nk − 2. Therefore, we obtain π(α(u)) and π(u) coincide outside
ℓ2(Tk≥1 nk−1+nkZ) = Cξ−1 ⊂ ℓ2(Z). Now Q(Z⋊P ) is purely infinite and simple, so its
image under π does not contain any compact operator other than 0. Thus faithfulness
of π allows us to conclude that α(u) = u, so that AutOnQ(Z ⋊ P ) is a subgroup of
AutC ∗(Z)Q(Z ⋊ P ) ∼= Hom(P, C(T, T)), where we refer to Theorem 6.2 for the latter
identification.
24
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
Now suppose ψ ∈ Homθ(P, C(T, T)) satisfies βψ ∈ AutOnQ(Z ⋊ P ). Then sn =
βψ(sn) = ψnsn forces ψn = 1 as in the proof for injectivity in Theorem 6.2. But then
Proposition 6.10 implies ψp ∈ T for all p ∈ P , so that ψ ∈ Hom(P, T) and hence βψ is a
gauge automorphism γχ.
(cid:3)
This result has an interesting immediate consequence for which we recall from Exam-
ple 2.1 that S ⊂ N× is a family of mutually coprime numbers and P is the submonoid
of N× generated by S.
.
Corollary 6.12. For ∅ 6= S′ ⊂ S, let Q
=
C∗({uksp p ∈ S′, 0 ≤ k ≤ p − 1}) ⊂ Q(Z ⋊ P ). Then AutAS ′Q(Z ⋊ P ) is isomor-
phic to the group of gauge automorphisms of R on Q(Z ⋊ Q). In particular, the only
automorphism of Q(Z ⋊ P ) that fixes the torsion subalgebra AS is the identity.
.
= hS \ S′i ⊂ N×, R
= \Q−1Q, and AS ′
Corollary 6.12 is remarkable as the torsion subalgebra plays a central role for the
.
K-theory of Q(Z ⋊ P ). We remark that in view of
Q(Z ⋊ P ) ∼= lim−→ Mp(C∗(Z)) ⋊ P whereas AS ∼= lim−→ Mp(C) ⋊ P,
this rigidity is already apparent from Theorem 6.11 because it tells us that the group
C∗-algebra has to be fixed pointwise as well.
Remark 6.13. Thanks to Theorem 6.11, it is now easy to see that the group AutOn Q(Z⋊
P ) is topologically isomorphic with TS−1, where S is the rank of P , see Example 2.1,
irrespective of what n is. Here, TS−1 is equipped with the product topology and
AutOn Q(Z ⋊ P ) with the topology of pointwise convergence in norm.
6.3. Outerness. We start with a simple observation for AutC ∗(G)Q(G ⋊θ P ) as dis-
following generalization of [ACR18a, Proposition 6.5]:
Proof. In the following computations, we work in the representation obtained out of the
Proposition 6.14. Let (G, P, θ) be an algebraic dynamical system as in Example 2.2.
cussed in Subsection 6.1 in the case of P = N, where every ψ ∈ Homθ(P, C(bG, T))
is determined by a single unitary f ∈ C(bG, T). By writing βf for βψ, we obtain the
If f ∈ C(bG, T) satisfies f (1 bG) 6= 1, then βf is an outer automorphism.
canonical representation via Fourier transform. This is the representation on L2(bG, µ)
= ξ(θ(χ)) for χ ∈ bG, where θ is the surjective group endomorphism
of bG corresponding θ via Pontryagin duality, π(ug)ξ(χ)
abelian subalgebra of Q(G ⋊ N), the unitary V is in C∗(G) ∼= C(bG, T), which means
that V = Mg with g ∈ C(bG, T). In our representation the equality f (u)s = V sV ∗ reads
Suppose that βF = Ad(V ) for some V ∈ U(Q(G ⋊ N)). Since C∗(G) is a maximal
= χ(ug)ξ(χ), ξ ∈ L2(bG, µ).
f (χ)ξ(θ(χ)) = f (χ)(sξ)(χ) = (Mf s)(ξ)(χ) = (MgsMg)(ξ)(χ) = g(χ)(sMg)(ξ)(χ)
given by π(s)ξ(χ)
.
as
.
= g(χ)(Mg)(ξ)(θ(χ)) = g(χ)g(θ(χ))ξ(θ(χ))
In particular, if we choose χ = ξ = 1 bG we get
f (1 bG) = g(1 bG)g(θ(1 bG)) = g(1 bG)g(1 bG) = 1.
(cid:3)
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
25
The next result is a generalization of [ACR18a, Theorem 5.1] for Q2, but the proof
we give here is quite different. Recall that Np = [G : θp(G)], and if P is an abelian
right LCM semigroup, we denote by T the Pontryagin dual of its Grothendieck group
P −1P and by γ : T y Q(G ⋊θ P ) the gauge action given by γχ(ugsp) = χ(p)ugsp for
Theorem 6.15. Suppose that (G, P, θ) is an algebraic dynamical system with G and P
abelian, and 2 ≤ [G : θp(G)] < ∞ for some p ∈ P . If C∗(G) ⊂ Q(G ⋊θ P ) is maximal
χ ∈ T, g ∈ G, p ∈ P . For (G, P, θ) with abelian G, we denote by bθp the surjective group
endomorphism of bG corresponding to θp for p ∈ P .
abelian and the only χ ∈ bG for which the system of functional equations
admits a solution f in C(bG, T) is χ = 1, then γ : T y Q(G ⋊θ P ) is an outer action.
Proof. Let χ ∈ bG such that γχ = Ad(w∗) for some w ∈ U(Q(G ⋊θ P )). As ug = γχ(ug) =
w∗ugw for all g ∈ G, we get w ∈ U(C∗(G)) ∼= C(bG, T) because C∗(G) is maximal abelian
in Q(G ⋊θ P ). If f ∈ C(bG, T) corresponds to w, then relation (I) yields
χ(p)f = f ◦ bθp
for all p ∈ P with Np < ∞
(6.3)
χ(p)sp = γχ(sp) = f spf = f (f ◦ bθp)sp
for all p ∈ P.
Multiplying the above equations by f on the left and using (III) for p ∈ P with Np < ∞,
we arrive at (6.3). By assumption, this forces χ = 1.
(cid:3)
Corollary 6.16. If (Z, P, θ) is an integral dynamics as described in Example 2.1, then
γ : TS y Q(Z ⋊ P ) is outer.
Proof. Note that G and P are abelian, the system is of finite type, and C∗(Z) ⊂ Q(Z⋊P )
is maximal abelian due to Theorem 3.5, see the discussion following Example 3.1. For
w ∈ T = TS, (6.3) gives wpf (z) = f (zp) for all z ∈ T and all p ∈ S. But it is well
known that each of these individual equations only has a solution for wp = 1, so that
w = 1. Now the result follows from Theorem 6.15.
(cid:3)
In fact, Corollary 6.16 can be easily deduced from the following result that competes
with Theorem 6.15:
Corollary 6.17. If (G, P, θ) is an algebraic dynamical system for which P is abelian,
P ∗ is finite, andTp∈P θp(G) = {1G}, then γ : T y Q(G ⋊θ P ) is an outer action.
Proof. Suppose γχ is inner, that is, γχ = Ad w, for some unitary w ∈ Q(G ⋊θ P ). Then
Lemma 4.5 applies as wspw∗ = γχ(sp) = χ(p)sp with χ(p) ∈ T, and implies χ = 1. (cid:3)
Our choice to include both Theorem 6.15 and Corollary 6.17 is based on the difference
in their approaches: While Theorem 6.15 makes seemingly restrictive assumptions like
C∗(G) ⊂ Q(G⋊θP ) being maximal abelian and the non-existence of non-trivial solutions
this assumption. On the other hand, Corollary 6.17 only asks for P to be abelian,
to (6.3), it does not require Tp∈P θp(G) = {1G}, even though our Theorem 3.5 needs
26
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
finiteness of P ∗, andTp∈P θp(G) = {1G}. So unless the assumptions in Theorem 6.15
forceTp∈P θp(G) = {1G} and P ∗ to be finite, the two results remain independent.
We stay with integral dynamics as described in Example 2.1 and prove outerness of
all automorphisms of α ∈ AutQ(Z ⋊ P ) that invert the unitary u, that is, α(u) = u∗,
see Theorem 6.18. The two particular examples we have in mind as motivation are
(a) λ−1 given by u 7→ u∗, sp 7→ sp; and
(b) φ given by u 7→ u∗, sp 7→ up−1sp for p ∈ P .
In order to check that φ is a morphism of QN we only need to check that φ(sp)φ(sq) =
φ(spq) (II) holds, which follows from
φ(sp)φ(sq) = up−1spuq−1sq = up−1+p(q−1)spq = φ(spq).
The reason is that the relations (I) and (II) for Q(Z ⋊ P ) are satisfied by φ(u) and φ(sp)
for all p ∈ P because the restriction φQp coincides with the unique extension of the
exchange automorphism on Qp ⊂ Q(Z ⋊ P ), see Definition 5.6 and Theorem 5.8. In
fact, φ is determined completely by the collection of all these exchange automorphisms.
We note that φ and λ−1 are unitarily equivalent as φ = Ad(u∗) ◦ λ−1.
Theorem 6.18. Every automorphism α of Q(Z ⋊ P ) with the property α(u) = u∗ is
outer. In particular, λ−1 and φ are outer.
Proof. The proof is an adaptation of the proof of [ACR18a, Theorem 5.9], so we rather
explain the strategy and point out modifications than go through the proof in full detail:
.
The symmetry P ∈ L(ℓ2(Z)) given by Pξn
= ξ−n for n ∈ Z satisfies Pπ(u)P = π(u∗), see
[ACR18a, Remark 5.5]. Then [ACR18a, Theorem 5.8] shows that P /∈ Q2, which we shall
explain here for Q(Z ⋊ P ): The C∗-algebra Q(Z ⋊ P ) is the closure of the linear span
with ci ∈ C, mi, ni ∈ Z, pi, qi ∈ P for i = 1, . . . , j for some j ∈ N. For every n ≥
max1≤i≤jmi ∨ ni we then get
of operators of the form umsps∗qu∗n with p, q ∈ P, m, n ∈ Z. So letP1≤i≤j ciumispis∗qiuni
ciumispis∗qiuni(ξn)(cid:13)(cid:13) ≥ 1
as
(cid:13)(cid:13)(cid:0)P − P1≤i≤j
ciumispis∗qiuni(cid:1)ξn(cid:13)(cid:13) =(cid:13)(cid:13)ξ−n − P1≤i≤j
umispis∗qiuni(ξn) =(ξmi+ pi
(ni+n)
qi
0
, if ni + n ∈ qiZ,
, otherwise.
qi
The reason is that the choice of n and P ⊂ N× force pi
(ni + n) to have the same sign as
n. Thus, again by the choice of n, we have mi + pi
(ni + n) 6= −n for all i, and therefore
qi
P /∈ Q(Z ⋊ P ).
Now suppose α ∈ AutQ(Z ⋊ P ) with α(u) = u∗ is inner, that is, there is w ∈
U(Q(Z⋊P )) such that α = Ad(w). Then we get Pπ(wuw∗)P = π(u), so that the unitary
Pπ(w) commutes with π(u). Thus we deduce Pπ(w) ∈ U(π(C∗(Z))′) ∼= U(L∞(T, µ)) =
L∞((T, µ), T), where µ denotes the Haar measure (or Lebesque measure) on T. Since P
is a symmetry, we arrive at π(w) ∈ P · L∞((T, µ), T). If we could take f ∈ C(T, T) ⊂
L∞((T, µ), T) with π(w) = Pπ(f (u)) instead of an essentially bounded function, the
contradiction would follow readily as P = Pπ(f (u)f (u)) = π(wf (u)) ∈ Q(Z ⋊ P )
contradicts P /∈ Q(Z ⋊ P ). However, the general case is more technical, and we refer to
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
27
the proof of [ACR18a, Theorem 5.9] for details, remarking only some constants within
the estimations will change.
(cid:3)
Remark 6.19. It is crucial to Theorem 6.18 that there is no p ∈ N× such that the monoid
P ⊂ Z× contains both p and −p because then P =P0≤m≤p−1 π(ums
−ps∗pu−m) belongs
to π(Q(Z ⋊ P )) ∼= Q(Z ⋊ P ), and the proof of Theorem 6.18 breaks down. In fact, we
have π ◦ λ−1 = Ad(P) ◦ π, and thus both λ−1 and also φ = Ad(u∗) ◦ λ−1 are inner. One
may therefore ask for new examples of outer automorphisms of Q(Z ⋊ P ), especially
with an eye on the case where P ⊂ Z× contains −1.
Corollary 6.20. For every 2 ≤ p < ∞, the group of extendible Bogolubov automor-
phisms of Op defines an outer action of T × Z/2Z on Qp.
Proof. This follows immediately from combining Theorem 5.8 with Corollary 6.16 and
Theorem 6.18.
(cid:3)
References
[ABLS17] Zahra Afsar, Nathan Brownlowe, Nadia S. Larsen, and Nicolai Stammeier, Equilibrium
states on right LCM semigroup C ∗-algebras, Int. Math. Res. Not., posted on 2017 (advance
article), DOI 10.1093/imrn/rnx162.
[ACR18a] Valeriano Aiello, Roberto Conti, and Stefano Rossi, A look at the inner structure of the 2-
adic ring C ∗-algebra and its automorphism groups, Publ. Res. Inst. Math. Sci. 54 (2018),
no. 1, 45 -- 87, DOI 10.4171/PRIMS/54-1-2 .
[ACR18b]
, Diagonal automorphisms of the 2-adic ring C ∗-algebra, Q. J. Math. 69 (2018),
815 -- 833, DOI 10.1093/qmath/hax064.
[BL17] Sel¸cuk Barlak and Xin Li, Cartan subalgebras and the UCT problem, Adv. Math. 316
(2017), 748 -- 769, DOI 10.1016/j.aim.2017.06.024.
[BL]
, Cartan subalgebras and the UCT problem, II. preprint, arxiv:1704.04939.
[BOS18] Sel¸cuk Barlak, Tron Omland, and Nicolai Stammeier, On the K-theory of C ∗-algebras aris-
ing from integral dynamics, Ergodic Theory Dynam. Systems 38 (2018), no. 3, 832 -- 862,
DOI 10.1017/etds.2016.63.
[BLS17] Nathan Brownlowe, Nadia S. Larsen, and Nicolai Stammeier, On C ∗-algebras associ-
ated to right LCM semigroups, Trans. Amer. Math. Soc. 369 (2017), no. 1, 31 -- 68, DOI
10.1090/tran/6638.
[BLS]
, C ∗-algebras of algebraic dynamical systems and right LCM semigroups, to appear
in Indiana Univ. Math. J. preprint available under IUMJ/Preprints/7527.
[BRRW14] Nathan Brownlowe, Jacqui Ramagge, David Robertson, and Michael F. Whittaker, Zappa-
Sz´ep products of semigroups and their C ∗-algebras, J. Funct. Anal. 266 (2014), no. 6, 3937 --
3967, DOI 10.1016/j.jfa.2013.12.025.
[BS16] Nathan Brownlowe and Nicolai Stammeier, The boundary quotient for algebraic dynamical
systems, J. Math. Anal. Appl. 438 (2016), no. 2, 772 -- 789, DOI 10.1016/j.jmaa.2016.02.015.
[Con10] Roberto Conti, Automorphisms of the UHF algebra that do not extend to the Cuntz algebra,
J. Aust. Math. Soc. 89 (2010), 309 -- 315.
[CHS12a] Roberto Conti, Jeong Hee Hong, and Wojciech Szyma´nski, The restricted Weyl group of
the Cuntz algebra and shift endomorphisms, J. Reine Angew. Math. 667 (2012), 177 -- 191.
[CHS12b] Roberto Conti, J. H. Hong, and Wojciech Szyma´nski, Endomorphisms of graph algebras,
J. Funct. Anal. 263 (2012), 2529 -- 2554.
[CHS12c]
[CHS12d]
, The Weyl group of the Cuntz algebra, Adv. Math. 231 (2012), no. 6, 3147 -- 3161.
, Endomorphisms of the Cuntz Algebras, Banach Center Publ. 96 (2012), 81 -- 97.
[CHS15] Roberto Conti, Jeong Hee Hong, and Wojciech Szyma´nski, On conjugacy of maximal abelian
subalgebras and the outer automorphism group of the Cuntz algebra, Proc. Roy. Soc. Edin-
burgh 145 A (2015), 269 -- 279.
28
V. AIELLO, R. CONTI, S. ROSSI, AND N. STAMMEIER
[CRS10] Roberto Conti, Mikael Rørdam, and Wojciech Szyma´nski, Endomorphisms of On which
[CS11] Roberto Conti and Wojciech Szyma´nski, Labeled trees and localized automorphisms of the
preserve the canonical UHF-subalgebra, J. Funct. Anal. 259 (2010), 602 -- 617.
Cuntz algebras, Trans. Amer. Math. Soc. 363 (2011), no. 11, 5847 -- 5870.
[CL07] John Crisp and Marcelo Laca, Boundary quotients and ideals of Toeplitz C ∗-algebras of
Artin groups, J. Funct. Anal. 242 (2007), no. 1, 127 -- 156, DOI 10.1016/j.jfa.2006.08.001.
[Cun77] Joachim Cuntz, Simple C*-algebras generated by isometries, Comm. Math. Phys. 57 (1977),
no. 2, 173 -- 185.
[Cun08]
, C ∗-algebras associated with the ax + b-semigroup over N, K-theory and noncom-
mutative geometry, 2008, pp. 201 -- 215.
[Cun16]
, C∗-Algebras associated with algebraic actions, Operator Algebras and Applications,
2016, pp. 145 -- 159.
[CDL13] Joachim Cuntz, Christopher Deninger, and Marcelo Laca, C*-algebras of Toeplitz type
associated with algebraic number fields, Math. Ann. 355 (2013), no. 4, 1383 -- 1423.
[CEL13] Joachim Cuntz, Siegfried Echterhoff, and Xin Li, On the K-theory of crossed prod-
ucts by automorphic semigroup actions, Q. J. Math. 64 (2013), no. 3, 747 -- 784, DOI
10.1093/qmath/hat021.
[CEL15]
, On the K-theory of the C*-algebra generated by the left regular representa-
tion of an Ore semigroup, J. Eur. Math. Soc. (JEMS) 17 (2015), no. 3, 645 -- 687, DOI
10.4171/JEMS/513.
[CELY17] Joachim Cuntz, Siegfried Echterhoff, Xin Li, and Guoliang Yu, K-Theory for Group C ∗-
Algebras and Semigroup C ∗-Algebras, Oberwolfach Seminars, Birkhauser Basel, 2017.
[CL10] Joachim Cuntz and Xin Li, The regular C*-algebra of an integral domain, Quanta of maths,
2010, pp. 149 -- 170.
[CL11]
, C ∗-algebras associated with integral domains and crossed products by actions on
adele spaces, J. Noncommut. Geom. 5 (2011), no. 1, 1 -- 37.
[CV13] Joachim Cuntz and Anatoly Vershik, C ∗-algebras associated with endomorphisms and poly-
morphisms of compact abelian groups, Comm. Math. Phys. 321 (2013), no. 1, 157-179, DOI
10.1007/s00220-012-1647-0.
[DS01] Kenneth J. Dykema and Dimitri Shlyakhtenko, Exactness of Cuntz -- Pimsner C*-algebras,
Proceedings of the Edinburgh Mathematical Society (Series 2) 44 (2001), no. 02, 425 -- 444.
[ELR16] Søren Eilers, Xin Li, and Efren Ruiz, The isomorphism problem for semigroup C ∗-algebras
of right-angled Artin monoids, Doc. Math. 21 (2016), 309 -- 343.
[KS59] Richard V. Kadison and Isadore M. Singer, Extensions of pure states, Amer. J. Math. 81
(1959), 383 -- 400.
[LL12] Nadia S. Larsen and Xin Li, The 2-adic ring C ∗-algebra of the integers and its representa-
tions, J. Funct. Anal. 262 (2012), no. 4, 1392 -- 1426.
[LR17] Xin Li and Jean Renault, Cartan subalgebras in C∗-algebras. Existence and uniqueness,
preprint, arXiv:1703.10505 (2017).
[Li10] Xin Li, Ring C∗-algebras, Math. Ann. 348 (2010), no.
4,
859 -- 898, DOI
10.1007/s00208-010-0502-x.
[Li12]
[Li13]
, Semigroup C ∗-algebras and amenability of semigroups, J. Funct. Anal. 262 (2012),
no. 10, 4302 -- 4340, DOI 10.1016/j.jfa.2012.02.020.
, Nuclearity of semigroup C*-algebras and the connection to amenability, Adv. Math.
244 (2013), 626 -- 662, DOI 10.1016/j.aim.2013.05.016.
[Ren08] Jean Renault, Cartan subalgebras in C ∗-algebras, Irish Math. Soc. Bull. 61 (2008), 29 -- 63.
[Sta] Nicolai Stammeier, Topological freeness for ∗-commuting covering maps, to appear in Hous-
, On C ∗-algebras of irreversible algebraic dynamical systems, J. Funct. Anal. 269
ton J. Math. arxiv:1311.0793.
[Sta15]
(2015), no. 4, 1136 -- 1179, DOI 10.1016/j.jfa.2015.02.005.
[Star15] Charles Starling, Boundary quotients of C ∗-algebras of right LCM semigroups, J. Funct.
Anal. 268 (2015), no. 11, 3326 -- 3356, DOI 10.1016/j.jfa.2015.01.001.
THE INNER STRUCTURE OF BOUNDARY QUOTIENTS OF RIGHT LCM SEMIGROUPS
29
[Zac00] Joachim Zacharias, Quasi-free automorphisms of Cuntz-Krieger-Pimsner algebras, C*-
Algebras, 2000, pp. 262 -- 272.
Section de Math´ematiques, Universit´e de Gen`eve, 2-4 rue du Li`evre, Case Postale
64, 1211 Gen`eve 4, Switzerland
E-mail address: [email protected]
Dipartimento di Scienze di Base e Applicate per l'Ingegneria, Sapienza Universit`a di
Roma, Italy
E-mail address: [email protected]
Dipartimento di Matematica, Universit`a di Roma Tor Vergata, Italy
E-mail address: [email protected]
Department of Mathematics, University of Oslo, P.O. Box 1053, Blindern, NO-0316
Oslo, Norway
E-mail address: [email protected]
|
1102.1229 | 3 | 1102 | 2012-02-28T23:50:35 | The path space of a higher-rank graph | [
"math.OA"
] | We construct a locally compact Hausdorff topology on the path space of a finitely aligned $k$-graph $\Lambda$. We identify the boundary-path space $\partial\Lambda$ as the spectrum of a commutative $C^*$-subalgebra $D_\Lambda$ of $C^*(\Lambda)$. Then, using a construction similar to that of Farthing, we construct a finitely aligned $k$-graph $\wt\Lambda$ with no sources in which $\Lambda$ is embedded, and show that $\partial\Lambda$ is homeomorphic to a subset of $\partial\wt\Lambda$ . We show that when $\Lambda$ is row-finite, we can identify $C^*(\Lambda)$ with a full corner of $C^*(\wt\Lambda)$, and deduce that $D_\Lambda$ is isomorphic to a corner of $D_{\wt\Lambda}$. Lastly, we show that this isomorphism implements the homeomorphism between the boundary-path spaces. | math.OA | math |
THE PATH SPACE OF A HIGHER-RANK GRAPH
SAMUEL B.G. WEBSTER
Abstract. We construct a locally compact Hausdorff topology on the path
space of a finitely aligned k-graph Λ. We identify the boundary-path space ∂Λ
as the spectrum of a commutative C ∗-subalgebra DΛ of C ∗(Λ). Then, using a
construction similar to that of Farthing, we construct a finitely aligned k-graph
eΛ with no sources in which Λ is embedded, and show that ∂Λ is homeomorphic
to a subset of ∂eΛ . We show that when Λ is row-finite, we can identify C ∗(Λ)
with a full corner of C ∗(eΛ), and deduce that DΛ is isomorphic to a corner of D eΛ.
Lastly, we show that this isomorphism implements the homeomorphism between
the boundary-path spaces.
1. Introduction
Cuntz and Krieger's work [2] on C ∗-algebras associated to (0, 1)-matrices and
the subsequent interpretation of Cuntz and Krieger's results by Enomoto and
Watatani [4] were the foundation of the field we now call graph algebras. Directed
graphs and their higher-rank analogues provide an intuitive framework for the anal-
ysis of this broad class of C ∗-algebras; there is an explicit relationship between the
dynamics of a graph and various properties of its associated C ∗-algebra. Kumjian
and Pask in [7] introduced higher-rank graphs (or k-graphs) as analogues of di-
rected graphs in order to study Robertson and Steger's higher-rank Cuntz-Krieger
algebras [18] using the techniques previously developed for directed graphs. Higher-
rank graph C ∗-algebras have received a great deal of attention in recent years, not
least because they extend the already rich and tractable class of graph C ∗-algebras
to include all tensor products of graph C ∗-algebras (and thus many Kirchberg alge-
bras whose K1 contains torsion elements [7]), as well as (up to Morita equivalence)
the irrational rotation algebras and many other examples of simple AT-algebras
with real rank zero [8].
Although the definition of a k-graph (Definition 2.1) isn't quite as straightfor-
ward as that of a directed graph, k-graphs are a natural generalisation of directed
graphs: Kumjian and Pask show in [7, Example 1.3] that 1-graphs are precisely
the path-categories of directed graphs. Like directed graph C ∗-algebras, higher-
rank graph C ∗-algebras were first studied using groupoid techniques. Kumjian and
Date: 29th March, 2011.
2010 Mathematics Subject Classification. Primary 46L05.
Key words and phrases. Graph algebra, k-graph, higher-rank graph.
This research was supported by the ARC Discovery Project DP0984360.
1
2
S.B.G. WEBSTER
Pask defined the k-graph C ∗-algebra C ∗(Λ) to be the universal C ∗-algebra for a
set of Cuntz-Krieger relations among partial isometries associated to paths of the
k-graph Λ. Using direct analysis, they proved a version of the gauge-invariant
uniqueness theorem for k-graph algebras. They then constructed a groupoid GΛ
from each k-graph Λ, and used the gauge invariant uniqueness theorem to prove
that the groupoid C ∗-algebra C ∗(GΛ) is isomorphic to C ∗(Λ). This allowed them
to make use of Renault's theory of groupoid C ∗-algebras to analyse C ∗(Λ). The
unit space G(0)
Λ of GΛ, which must be locally compact and Hausdorff, is a collection
of paths in the graph: for a row-finite graph with no sources, G(0)
Λ is the collec-
tion of infinite paths in Λ (the definition of an infinite path in a k-graph is not
straightforward, see Remark 2.4). For more complicated graphs, the infinite paths
are replaced with boundary paths (Definition 2.9).
In [12], Raeburn, Sims and Yeend developed a "bare-hands" analysis of k-graph
C ∗-algebras. They found a slightly weaker alternative to the no-sources hypothesis
from Kumjian and Pask's theorems called local convexity (Definition 2.7). The
same authors later introduced finitely aligned k-graphs in [13], and gave a direct
analysis of their C ∗-algebras. This remains the most general class of k-graphs to
which a C ∗-algebra has been associated and studied in detail.
Many results for row-finite directed graphs with no sources can be extended to
arbitrary graphs via a process called desingularisation. Given an arbitrary directed
graph E, Drinen and Tomforde show in [3] how to construct a row-finite directed
graph F with no sources by adding vertices and edges to E in such a way that
the C ∗-algebra associated to F contains the C ∗-algebra associated to E as a full
corner. The modified graph F is now called a Drinen-Tomforde desingularisation
of E. Although no analogue of a Drinen-Tomforde desingularisation is currently
available for higher-rank graphs, Farthing provided a construction in [5] analogous
to that in [1] for removing the sources in a locally convex, row-finite higher-rank
graph. The statements of the results of [5] do not contain the local convexity
hypothesis, but Farthing alerted us to an issue in the proof of [5, Theorem 2.28]
(see Remark 6.2), which arises when the graph is not locally convex.
The goal of this paper is to explore the path spaces of higher-rank graphs and
investigate how these path spaces interact with desingularisation procedures such
as Farthing's.
In Section 2, we recall the definitions and standard notation for higher-rank
graphs. In Section 3, following the approach of [9], we build a topology on the path
space of a higher-rank graph, and show that the path space is locally compact and
Hausdorff under this topology.
In Section 4, given a finitely aligned k-graph Λ, we construct a k-graph eΛ with
no sources which contains a subgraph isomorphic to Λ. Our construction is mod-
elled on Farthing's construction in [5], and the reader is directed to [5] for several
proofs. The crucial difference is that our construction involves extending elements
of the boundary-path space ∂Λ, whereas Farthing extends paths from a different
THE PATH SPACE OF A HIGHER-RANK GRAPH
3
set Λ≤∞ (see Remark 2.10). Interestingly, although ∂Λ and Λ≤∞ are potentially
different when Λ is row-finite and not locally convex (Proposition 2.12), our con-
struction and Farthing's yield isomorphic k-graphs except in the non-row-finite
case (Examples 4.10 and Proposition 4.12). We follow Robertson and Sims' nota-
tional refinement [17] of Farthing's desourcification: we construct a new k-graph in
which the original k-graph is embedded, whereas Farthing's construction adds bits
onto the existing k-graph. This simplifies many arguments involving eΛ; however,
the main reason for modifying Farthing's construction is that Λ≤∞ is not as well
behaved topologically as ∂Λ (see Remark 3.5) and in particular, no analogue of
Theorem 5.1 holds for Farthing's construction.
In Section 5, we prove that given a row-finite k-graph Λ, there is a natural
homeomorphism from the boundary-path space of Λ onto the space of infinite
discussion showing that the topological basis constructed in Section 3 is the one
we want.
In Section 6 we recall the definition of the Cuntz-Krieger algebra C ∗(Λ) of a
with no sources obtained by applying the construction of Section 4 to Λ, then the
paths in eΛ with range in the embedded copy of Λ. We provide examples and
higher-rank graph Λ. We show that if Λ is a row-finite k-graph andeΛ is the graph
embedding of Λ in eΛ induces an isomorphism π of C ∗(Λ) onto a full corner of
C ∗(eΛ).
Section 7 contains results about the diagonal C ∗-subalgebra of a k-graph C ∗-
algebra: the C ∗-algebra generated by range projections associated to paths in
the k-graph. We identify the boundary-path space of a finitely aligned higher-
rank graph with the spectrum of its diagonal C ∗-algebra. We then show that
the isomorphism π of Section 6 restricts to an isomorphism of diagonals which
implements the homeomorphism of Section 5.
Acknowledgements. The work contained in this paper is from the author's PhD
thesis, and as such I extend thanks to my PhD supervisors Iain Raeburn and Aidan
Sims for their support and willingness to proofread and guide my work.
2. Preliminaries
Definition 2.1. Given k ∈ N, a k-graph is a pair (Λ, d) consisting of a countable
category Λ = (Obj(Λ), Mor(Λ), r, s) together with a functor d : Λ → Nk, called the
degree map, which satisfies the factorisation property: for every λ ∈ Mor(Λ) and
m, n ∈ Nk with d(λ) = m + n, there are unique elements µ, ν ∈ Mor(Λ) such that
λ = µν, d(µ) = m and d(ν) = n. Elements λ ∈ Mor(Λ) are called paths. We follow
the usual abuse of notation and write λ ∈ Λ to mean λ ∈ Mor(Λ). For m ∈ Nk
we define Λm := {λ ∈ Λ : d(λ) = m}. For subsets F ⊂ Λ and V ⊂ Obj(Λ), we
write V F := {λ ∈ F : r(λ) ∈ V } and F V := {λ ∈ F : s(λ) ∈ V }. If V = {v},
we drop the braces and write vF and F v. A morphism between two k-graphs
(Λ1, d1) and (Λ2, d2) is a functor f : Λ1 → Λ2 which respects the degree maps. The
4
S.B.G. WEBSTER
(0, 2)
(1, 2)
(2, 2)
(0, 1)
(1, 1)
(2, 1)
(0, 0)
(1, 0)
(2, 0)
Figure 1. The 2-graph Ω2,2.
factorisation property allows us to identify Obj(Λ) with Λ0. We refer to elements
of Λ0 as vertices.
Remark 2.2. To visualise a k-graph we draw its 1-skeleton: a directed graph with
i=1 Λei. To each edge we assign a colour determined by the
edge's degree. We tend to use 2-graphs for examples, and we draw edges of degree
(1, 0) as solid lines, and edges of degree (0, 1) as dashed lines.
vertices Λ0 and edgesSk
Example 2.3. For k ∈ N and m ∈ (N∪{∞})k, we define k-graphs Ωk,m as follows.
Set Obj(Ωk,m) = {p ∈ Nk : pi ≤ mi for all i ≤ k},
Mor(Ωk,m) = {(p, q) : p, q ∈ Obj(Ωk,m) and pi ≤ qi for all i ≤ k},
r(p, q) = p, s(p, q) = q and d(p, q) = q − p, with composition given by (p, q)(q, t) =
(p, t). If m = (∞)k, we drop m from the subscript and write Ωk. The 1-skeleton
of Ω2,2 is depicted in Figure 1.
Remark 2.4. The graphs Ωk,m provide an intuitive model for paths: every path
λ of degree m in a k-graph Λ determines a k-graph morphism xλ : Ωk,m → Λ.
To see this, let p, q ∈ Nk be such that p ≤ q ≤ m. Define xλ(p, q) = λ′′, where
λ = λ′λ′′λ′′′; and d(λ′) = p, d(λ′′) = q − p and d(λ′′′) = m − q. In this way, paths
in Λ are often identified with the graph morphisms xλ : Ωk,m → Λ. We refer to
the segment λ′′ of λ (as factorized above) as λ(p, q), and for n ≤ m, we refer to
the vertex r(λ(n, m)) = s(λ(0, n)) as λ(n). By analogy, for m ∈ (N ∪ {∞})k we
define Λm := {x : Ωk,m → Λ : x is a graph morphism.}. For clarity of notation, if
m = (∞)k we write Λ∞.
Define
WΛ := [n∈(N∪{∞})k
Λn.
We call WΛ the path space of Λ. We drop the subscript when confusion is unlikely.
For m, n ∈ Nk, we denote by m ∧ n the coordinate-wise minimum, and by m ∨ n
the coordinate-wise maximum. With no parentheses, ∨ and ∧ take priority over
the group operation: a − b ∧ c means a − (b ∧ c).
Since finite and infinite paths are fundamentally different, that one can compose
them isn't immediately obvious.
THE PATH SPACE OF A HIGHER-RANK GRAPH
5
Lemma 2.5 ([19, Proposition 3.0.1.1]). Let Λ be a k-graph. Suppose λ ∈ Λ and
suppose that x ∈ W satisfies r(x) = s(λ). Then that there exists a unique k-graph
morphism λx : Ωk,d(λ)+d(x) → Λ such that (λx)(0, d(λ)) = λ and (λx)(d(λ), n +
d(λ)) = x(0, n) for all n ≤ d(x).
Definition 2.6. For λ, µ ∈ Λ, write
Λmin(λ, µ) := {(α, β) ∈ Λ × Λ : λα = µβ, d(λα) = d(λ) ∨ d(µ)}
for the collection of pairs which give minimal common extensions of λ and µ, and
denote the set of minimal common extensions by
MCE(λ, µ) := {λα : (α, β) ∈ Λmin(λ, µ)} = {µβ : (α, β) ∈ Λmin(λ, µ)}.
Definition 2.7. A k-graph Λ is row-finite if for each v ∈ Λ0 and m ∈ Nk, the set
vΛm is finite; Λ has no sources if vΛm 6= ∅ for all v ∈ Λ0 and m ∈ Nk.
We say that Λ is finitely aligned if Λmin(λ, µ) is finite (possibly empty) for all
λ, µ ∈ Λ.
As in [12, Definition 3.1], a k-graph Λ is locally convex if for all v ∈ Λ0, all
i, j ∈ {1, . . . k} with i 6= j, all λ ∈ vΛei and all µ ∈ vΛej , the sets s(λ)Λej
and s(µ)Λei are non-empty. Roughly speaking, local convexity stipulates that Λ
contains no subgraph resembling:
µ
u
v
w
λ
Definition 2.8. For v ∈ Λ0, a subset E ⊂ vΛ is exhaustive if for every µ ∈ vΛ
there exists a λ ∈ E such that Λmin(λ, µ) 6= ∅. We denote the set of all finite
exhaustive subsets of Λ by F E(Λ).
Definition 2.9. An element x ∈ W is a boundary path if for all n ∈ Nk with
n ≤ d(x) and for all E ∈ x(n)F E(Λ) there exists m ∈ Nk such that x(n, m) ∈ E.
We write ∂Λ for the set of all boundary paths.
Define the set Λ≤∞ as follows. A k-graph morphism x : Ωk,m → Λ is an element
of Λ≤∞ if there exists nx ≤ d(x) such that for n ∈ Nk satisfying nx ≤ n ≤ d(x)
and ni = d(x)i, we have x(n)Λei = ∅.
Remark 2.10. Raeburn, Sims and Yeend introduced Λ≤∞ to construct a nonzero
Cuntz-Krieger Λ-family [13, Proposition 2.12]. Farthing, Muhly and Yeend intro-
duced ∂Λ in [6]; in order to construct a groupoid to which Renault's theory of
groupoid C ∗-algebras [15] applied, they required a path space which was locally
compact and Hausdorff in an appropriate topology, and Λ≤∞ did not suffice. The
differences between ∂Λ and Λ≤∞ can be easily seen if Λ contains any infinite re-
ceivers (e.g. any path in a 1-graph Λ with source an infinite receiver is an element
of ∂Λ \ Λ≤∞), but can even show itself in the row-finite case if Λ is not locally
convex.
6
S.B.G. WEBSTER
Example 2.11. Suppose Λ is the 2-graph with the skeleton pictured below.
•
•
•
•
f0
v0
ω0
f1
v1
x0
f2
v2
x1
f3
. . .
v3
x2
ω1
•
ω2
•
ω3
•
. . .
•
Consider the paths x = x0x1 . . . , and ωn = x0x1 . . . xn−1ωn for n = 0, 1, 2, . . . .
for each n ∈ N, we have d(x)2 = 0 = (n, 0)2, and
Observe that x /∈ Λ≤∞:
x((n, 0))Λe2 = vnΛe2 6= ∅.
We claim that x ∈ ∂Λ. Fix m ∈ N and E ∈ vmF E(Λ). Since E is exhaustive,
for each n ≥ m, there exists λn ∈ E such that MCE(λn, xm . . . xn−1ωn) 6= ∅. Since
E is finite, it can not contain xm . . . xn−1ωn for every n ≥ m, so it must contain
xm . . . xp for some p ∈ N. So x((m, 0), (m + p)) = xm . . . xp belongs to E.
The 2-graph of Example 2.11 first appeared in Robertson's honours thesis [16]
to illustrate a subtlety arising in Farthing's procedure [5] for removing sources in
k-graphs when the k-graphs in question are not locally convex.
It was for this
reason that only locally convex k-graphs in the main results of [16, 17].
Proposition 2.12. Suppose Λ is a finitely aligned k-graph. Then Λ≤∞ ⊂ ∂Λ. If
Λ is row-finite and locally convex, then Λ≤∞ = ∂Λ.
To prove this we use the following lemma.
Lemma 2.13. Let Λ be a row-finite, locally convex k-graph, and suppose that
v ∈ Λ0 satisfies vΛei 6= ∅ for some i ≤ k. Then vΛei ∈ vF E(Λ).
Proof. Since Λ is row-finite, vΛei is finite. To see that it is exhaustive, let µ ∈ vΛ.
If d(µ)i > 0, then g = µ(0, ei) ∈ vΛei implies that Λmin(µ, g) 6= ∅. Suppose that
d(µ)i = 0. Let µ = µ1 . . . µn be a factorisation of µ such that d(µj) = 1 for each
j ≤ n. Since Λ is locally convex, s(µ)Λei = s(µn)Λei 6= ∅. Fix g ∈ s(µ)Λei. Let
f := (µg)(0, ei). Then f ∈ vΛei. Since d(µi) = 0, we have d(µg) = d(µ) ∨ d(f ).
Hence (g, (µg)(ei, d(µg))) ∈ Λmin(µ, f ) as required.
(cid:3)
Proof of Proposition 2.12. Fix x ∈ Λ≤∞. Let m ≤ d(x) and E ∈ x(m)F E(Λ).
Define t ∈ Nk by
if d(x)i < ∞,
if d(x)i = ∞.
ti :=(d(x)i
max
λ∈E (cid:0)nx ∨ (m + d(λ))(cid:1)i
Then x(m, t) ∈ x(m)Λ, so there exists λ ∈ E such that Λmin(x(m, t), λ) is non-
empty. Let (α, β) ∈ Λmin(x(m, t), λ). We first show that d(α) = 0. Since x ∈ Λ≤∞
and nx ≤ t ≤ d(x), if d(x)i < ∞ then x(t)Λei = ∅. So for each i such that
d(x)i < ∞, we have d(α)i = 0. Now suppose that d(x)i = ∞. Then d(x(m, t))i =
ti − mi ≥ d(λ)i. So d(x(m, t)α)i = max{d(x(m, t))i, d(λ)i} = d(x(m, t))i, giving
d(α)i = 0. Then we have x(m, t) = λβ, so x(m, m + d(λ)) = λ.
THE PATH SPACE OF A HIGHER-RANK GRAPH
7
Now suppose that Λ is row-finite and locally convex. We want to show ∂Λ ⊂
Λ≤∞. Fix x ∈ ∂Λ, and n ∈ Nk such that n ≤ d(x) and ni = d(x)i. It suffices
to show that x(n)Λei = ∅. Since ni = d(x)i, we have x(n)Λei /∈ x(n)F E(Λ).
Lemma 2.13 then implies that x(n)Λei = ∅.
(cid:3)
3. Path Space Topology
Following the approach of Paterson and Welch in [9], we construct a locally
compact Hausdorff topology on the path space W of a finitely aligned k-graph
Λ. The cylinder set of µ ∈ Λ is Z(µ) := {ν ∈ W : ν(0, d(µ)) = µ}. Define
α : W → {0, 1}Λ by α(w)(y) = 1 if w ∈ Z(y) and 0 otherwise. For a finite subset
G ⊂ s(µ)Λ we define
(3.1)
Z(µ \ G) := Z(µ) \ [ν∈G
Z(µν).
Our goals for this section are the following two theorems. The basis we end up
with is slightly different to that in [9, Corollary 2.4], revealing a minor oversight
of the authors.
Theorem 3.1. Let Λ be a finitely aligned k-graph. Then the collection
nZ(µ \ G) : µ ∈ Λ and G ⊂
(s(µ)Λei) is finiteo
k[i=1
is a base for the initial topology on W induced by {α}.
Theorem 3.2. Let Λ be a finitely aligned higher-rank graph. With the topology
described in Theorem 3.1, W is a locally compact Hausdorff space.
Let F be a set of paths in a k-graph Λ. A path β ∈ W is a common extension
of the paths in F if for each µ ∈ F , we can write β = µβµ for some βµ ∈ W .
µ
paths in F . We denote the set of all minimal common extensions of the paths in
F by MCE(F ). Since MCE({µ, ν}) = MCE(µ, ν), this definition is consistent with
Definition 2.6.
If in addition d(β) = Wµ∈F d(µ), then β is a minimal common extension of the
Remark 3.3. If F ⊂ Λ is finite, thenTµ∈F Z(µ) =Sβ∈MCE(F ) Z(β).
µ ∈ G and {0, 1} otherwise. Then the sets N(F, G) := Qµ∈Λ U F,G
Proof of Theorem 3.1. We first describe the topology on {0, 1}Λ. Given disjoint
finite subsets F, G ⊂ Λ and µ ∈ Λ, define sets U F,G
to be {1} if µ ∈ F , {0} if
µ where F, G
range over all finite disjoint pairs of subsets of Λ form a base for the topology on
{0, 1}Λ.
8
S.B.G. WEBSTER
Clearly, α is a homeomorphism onto its range, and hence the sets α−1(N(F, G))
are a base for a topology on W . Routine calculation shows that
Z(ν)! ,
α−1(N(F, G)) = [µ∈MCE(F )
Z(µ) \ [ν∈G
so the sets Z(µ) \Sν∈G Z(µν) = Z(µ \ G) are a base for our topology.
i=1(cid:0)s(α)Λei(cid:1) such that
and λ ∈ Z(µ \ G), there exist α ∈ Λ and a finite F ⊂ Sk
λ ∈ Z(α \ F ) ⊂ Z(µ \ G). Let N := (cid:0)Wν∈G d(µν)(cid:1) ∧ d(λ) and α = λ(0, N). To
define F , we first define a set Fν associated to each ν ∈ G, then take F =Sν∈G Fν.
To finish the proof, it suffices to show that for µ ∈ Λ, a finite subset G ⊂ s(µ)Λ
Fix ν ∈ G. We consider the following cases:
(1) If N ≥ d(µν) or MCE(α, µν) = ∅, let Fν = ∅.
(2) If N (cid:3) d(µν) and MCE(α, µν) 6= ∅, define Fν as follows:
Since N (cid:3) d(µν), there exists jν ≤ k such that Njν < d(µν)jν . Hence each
γ ∈ MCE(α, µν) satisfies d(γ)jν = (N ∨ d(µν))jν > Njν . Define Fν = {γ(N, N +
ejν ) : γ ∈ MCE(α, µν)}. Since Λ is finitely aligned, Fν is finite.
We now show that λ ∈ Z(α \ F ). We have λ ∈ Z(α) by choice of α. If F = ∅
If not, then fix ν ∈ G such that Fν 6= ∅, and fix e ∈ Fν. Then
we are done.
e = γ(N, N + ejν ) for some γ ∈ MCE(α, µν). Then d(λ)jν = Njν < (N + ejν )jν =
d(αe)jν . So λ /∈ Z(αe), hence λ ∈ Z(α \ F ).
We now show that Z(α \ F ) ⊂ Z(µ \ G). Fix β ∈ Z(α \ F ). Since α ∈ Z(µ),
we have β ∈ Z(µ). Fix ν ∈ G. We show that β /∈ Z(µν) in cases:
(1) Suppose that N ≥ d(µν). Since β ∈ Z(α) = Z(λ(0, N)) and λ /∈ Z(µν), it
follows that β /∈ Z(µν).
(2) If N (cid:3) d(µν), then either
(a) MCE(α, µν) = ∅, in which case β ∈ Z(α) implies that β /∈ Z(µν); or
(b) MCE(α, µν) 6= ∅. Then for each γ ∈ MCE(α, µν), we know β(N, N +
(cid:3)
ejν ) 6= γ(N, N + ejν ). It then follows that β /∈ Z(µν).
Lemma 3.4. Let {ν(n)} be a sequence of paths in Λ such that
(i) d(ν(n+1)) ≥ d(ν(n)) for all n ∈ N, and
ν(n) for all n ∈ N.
(ii) ν(n+1)(cid:0)0, d(ν(n))(cid:1) = ν(n) for all n ∈ N.
Then there exists a unique ω ∈ W with d(ω) =Wn∈N d(ν(n)) and ω(cid:0)0, d(ν(n))(cid:1) =
Proof. Let m =Wn∈N d(ν(n)) ∈ (N ∪ {∞})k. Then
For a ∈ Nk with a ≤ m, there exists Na ∈ N such that d(ν(Na)) ≥
a.
(3.2)
For each (p, q) ∈ Ωk,m apply (3.2) with a = q and define ω(p, q) = ν(Nq)(p, q).
Routine calculations using (3.2) show that ω : Ωk,m → Λ is a well-defined graph
morphism with the required properties.
(cid:3)
THE PATH SPACE OF A HIGHER-RANK GRAPH
9
Proof of Theorem 3.2. Fix v ∈ Λ0. We follow the strategy of [9, Theorem 2.2]
to show Z(v) is compact: since α is a homeomorphism onto its range, and since
{0, 1}Λ is compact, it suffices to prove that α(Z(v)) is closed in {0, 1}Λ. Suppose
that (ω(n))n∈N is a sequence in Z(v) such that converging to f ∈ {0, 1}Λ. We seek
ω ∈ Z(v) such that f = α(ω). Define A = {ν ∈ Λ : α(ω(n))(ν) → 1 as n → ∞}.
Then A 6= ∅ since v ∈ A. Let d(A) :=Wν∈A d(ν).
Claim 3.2.1. There exists ω ∈ vΛd(A) such that:
• d(ω) ≥ d(µ) for all µ ∈ A, and
• ω(0, n) ∈ A for all n ∈ Nk with n ≤ d(A).
Proof. To define ω we construct a sequence of paths and apply Lemma 3.4. We
first show that for each pair µ, ν ∈ A, MCE(µ, ν) ∩ A contains exactly one element.
Fix µ, ν ∈ A. Then for large enough n, there exist βn ∈ MCE(µ, ν) such that
ωn = βn(ωn)′. Since MCE(µ, ν) is finite, there exists M such that ωn = βM (ωn)′
for infinitely many n. Define βµ,ν := βM . Then βµ,ν ∈ A. For uniqueness, suppose
that φ ∈ MCE(µ, ν) ∩ A. Then for large n we have βµ,ν = ωn(0, d(µ) ∨ d(ν)) = φ.
Since A is countable, we can list A = {ν1, ν2, . . . , νm, . . . }. Let y1 := ν1, and
iteratively define yn = βyn−1,νn. Then d(yn) = d(yn−1) ∨ d(νn) ≥ d(yn−1), and
yn(0, yn−1) = yn−1. By Lemma 3.4, there exists a unique ω ∈ W satisfying d(ω) =
d(A) and ω(0, d(yn)) = yn for all n. It then follows from (3.2) that ω(0, n) ∈ A for
all n ≤ d(A).
(cid:3)Claim
To see α(Z(v)) is closed, fix λ ∈ Λ. We show that α(ω(n))(λ) → α(ω)(λ). If
α(ω)(λ) = 1, then λ = ω(0, d(λ)) ∈ A by Claim 3.2.1, and thus α(ω(n))(λ) → 1
If d(λ) (cid:2) d(ω), then λ /∈ A by
as n → ∞. Now suppose that α(ω)(λ) = 0.
Claim 3.2.1, forcing α(ω(n))(λ) → 0. Suppose that d(λ) ≤ d(ω). Since ω(0, d(λ)) ∈
A, we have ω(n)(0, d(λ)) = ω(0, d(λ)) for large n. Then α(ω)(λ) = 0 implies
that ω(0, d(λ)) 6= λ. So for large enough n we have ω(n)(0, d(λ)) 6= λ, forcing
α(ω(n))(λ) → 0.
(cid:3)
Remark 3.5. It has been shown that ∂Λ is a closed subset of W [6, Lemma 5.12].
Hence ∂Λ, with the relative topology, is a locally compact Hausdorff space. Con-
sider the 2-graph of Example 2.11. For each n ∈ N, we have ωn ∈ Λ≤∞. Notice
that ωn → x /∈ Λ≤∞. So Λ≤∞ is not closed in general, and hence is not locally
compact.
4. Removing Sources
Theorem 4.1. Let Λ be a finitely aligned k-graph. Then there is a finitely aligned
k-graph eΛ with no sources, and an embedding ι of Λ in eΛ. If Λ is row-finite, then
so is eΛ.
Definition 4.2. Define a relation ≈ on VΛ := {(x; m) : x ∈ ∂Λ, m ∈ Nk} by:
(x; m) ≈ (y; p) if and only if
10
S.B.G. WEBSTER
(V1) x(m ∧ d(x)) = y(p ∧ d(y)); and
(V2) m − m ∧ d(x) = p − p ∧ d(y).
Definition 4.3. Define a relation ∼ on PΛ := {(x; (m, n)) : x ∈ ∂Λ, m ≤ n ∈ Nk}
by: (x; (m, n)) ∼ (y; (p, q)) if and only if
(P1) x(m ∧ d(x), n ∧ d(x)) = y(p ∧ d(y), q ∧ d(y));
(P2) m − m ∧ d(x) = p − p ∧ d(y); and
(P3) n − m = q − p.
It is clear from their definitions that both ≈ and ∼ are equivalence relations.
Lemma 4.4. Suppose that (x; (m, n)) ∼ (y; (p, q)). Then n−n∧d(x) = q−q∧d(y).
Proof. It follows from (P1) and (P3) that
n − n ∧ d(x) − (m − m ∧ d(x)) = q − q ∧ d(y) − (p − p ∧ d(y)).
The result then follows from (P2).
(cid:3)
To define the range and source maps, observe that if (x; (m, n)) ∼ (y; (p, q)),
then (x; m) ≈ (y; p) by definition, and (x; n) ≈ (y; q) by Lemma 4.4. We define
range and source maps as follows.
Let fPΛ := PΛ/ ∼ and fVΛ := VΛ/ ≈. The class in fPΛ of (x; (m, n)) ∈ PΛ is
denoted [x; (m, n)], and similarly the class in fVΛ of (x; m) ∈ VΛ is denoted [x; m].
Definition 4.5. Defineer,es :fPΛ →fVΛ by:
Sn≥m Λn → Λ by σm(λ)(p, q) = λ(p + m, q + m).
elements of fPΛ satisfying [x; n] = [y; p]. Let z := x(0, n ∧ d(x))σp∧d(y)y. Then
(1) z ∈ ∂Λ;
(2) m ∧ d(x) = m ∧ d(z) and n ∧ d(x) = n ∧ d(z);
(3) x(m ∧ d(x), n ∧ d(x)) = z(m ∧ d(z), n ∧ d(z)) and y(p ∧ d(y), q ∧ d(y)) =
We now define composition. For each m ∈ Nk, we define the shift map σm :
Proposition 4.6. Suppose that Λ is a k-graph and let [x; (m, n)] and [y; (p, q)] be
er([x; (m, n)]) = [x; m]
and
es([x; (m, n)]) = [x, n].
z(n ∧ d(z), (n + q − p) ∧ d(z)).
Proof. Part (1) follows from [6, Lemma 5.13], and (2) and (3) can be proved as in
[5, Proposition 2.11].
(cid:3)
d(x))σp∧d(y)y. That the formula
Fix [x; (m, n)], [y; (p, q)] ∈ fPΛ such that [x; n] = [y; p], and let z = x(0, n ∧
[x; (m, n)] ◦ [y; (p, q)] = [z; (m, n + q − p)]
(4.1)
determines a well-defined composition follows from Proposition 4.6.
Define id :fVΛ →fPΛ by id[x;m] = [x; (m, m)].
Proposition 4.7 ([5, Lemma 2.19]). eΛ := (fVΛ,fPΛ,er,es, ◦, id) is a category.
THE PATH SPACE OF A HIGHER-RANK GRAPH
11
Definition 4.8. Define ed :eΛ → Nk by ed(v) = ⋆ for all v ∈fVΛ, and ed([x; (m, n)]) =
n − m for all [x; (m, n)] ∈fPΛ.
Proposition 4.9 ([5, Theorem 2.22]). The map ed defined above satisfies the fac-
torisation property. Hence with eΛ as in Proposition 4.7, (eΛ,ed) is a k-graph with
Example 4.10. If we allow infinite receivers, our construction yields a different k-
graph to Farthing's construction in [5, §2]: consider the 1-graph E with an infinite
number of loops fi on a single vertex v:
no sources.
...
fi
v
Here we have E≤∞ = ∅, so Farthing's construction yields a 1-graph E ∼= E.
Since v belongs to every finite exhaustive set in E, we have ∂E = E. Furthermore
[fj; p] = [fi; p] = [v; p] for all i, j, p ∈ N, and
[fj; (p, q)] = [fi; (p, q)] = [v; (p − 1, q − 1)]
for all i, j, p, q such that 1 < p ≤ q. Thus there is exactly one path between any
two of the added vertices, resulting in a head at v, yielding the graph illustrated
below
...
fi
v
It is intriguing that following Drinen and Tomforde's desingularisation, a head
is also added at infinite receivers like this, and then the ranges of the edges fi
are distributed along this head -- we cannot help but wonder whether this might
suggest an approach to a Drinen-Tomforde desingularisation for k-graphs.
4.1. Row-finite 1-graphs. While one expects this style of desourcification to
agree with adding heads to a row-finite 1-graph as in [1], this appears not to have
been checked anywhere.
Proposition 4.11. Let E be a row-finite directed graph and F be the graph obtained
by adding heads to sources, as in [1, p4]. Let Λ be the 1-graph associated to E.
Then eΛ ∼= F ∗, where F ∗ is a the path-category of F .
Proof. Define η′ : PΛ → F ∗ as follows. Fix x ∈ ∂E and m, n ∈ N. Then either
x ∈ E∞, or x ∈ E∗ and s(x) is a source in E. If x ∈ E∞, define η′((x; (m, n))) =
12
S.B.G. WEBSTER
x(m, n). For x ∈ E∗, let µx be the head added to s(x), and define η′((x; (m, n))) =
(xµx)(m, n). It is straightforward to check that η′ respects the equivalence relation
calculations show that η is a graph morphism.
∼ on PΛ. Define η : eΛ → F ∗ by η([x; (m, n)]) = η′((x; (m, n))). Easy but tedious
We now construct a graph morphism ξ : F ∗ → eΛ. Let ν ∈ F ∗. To define ξ
we first need some preliminary notation. ξ will be defined casewise, broken up as
follows:
(i) ν ∈ E∗,
(ii) r(ν) ∈ E∗ and s(ν) ∈ F ∗ \ E∗, or
(iii) r(ν), s(ν) ∈ F ∗ \ E∗.
If ν ∈ E∗, fix αν ∈ s(ν)∂E.
If ν has r(ν) ∈ E∗ and s(ν) ∈ F ∗ \ E∗, let
pν = max{p ∈ N : ν(0, p) ∈ E∗}. Then ν(pν) is a source in E∗, and ν(0, pν) ∈ ∂E.
If ν ∈ F ∗ \ E∗, then ν is a segment of a head µν added to a source in E∗, and we
let qν be such that ν = µν(qν, qν + d(µ)).
We then define ξ by
ξ(ν) =
[ναν; (0, d(ν))]
[ν(0, pν); (0, d(ν))]
[r(µν); (qν, qν + d(ν))]
if ν ∈ E∗
if r(ν) ∈ E∗ and s(ν) /∈ E∗
if r(ν), s(ν) ∈ F ∗ \ E∗.
Again, tedious but straightforward calculations show that ξ is a well-defined
(cid:3)
graph morphism, and that ξ ◦ η = 1 eΛ and η ◦ ξ = 1F ∗.
When Λ is row-finite and locally convex, Proposition 2.12 implies that Λ≤∞ =
∂Λ. In this case our construction is essentially the same as that of Farthing [5, §2],
with notation adopted as in [17]. If Λ is row-finite but not locally convex, then
Λ≤∞ ⊂ ∂Λ (Example 2.11 shows that this may be a strict containment). Thus it
is reasonable to suspect that our construction could result in a larger path space
than Farthing's. Interestingly, this is not the case.
Proposition 4.12. Let Λ be a row-finite k-graph. Suppose that x ∈ ∂Λ \ Λ≤∞ and
m ≤ n ∈ Nk. Then there exists y ∈ Λ≤∞ such that (x; (m, n)) ∼ (y; (m, n)).
Proof. Since x /∈ Λ≤∞, there exists q ≥ n ∧ d(x) and i ≤ k such that q ≤ d(x),
qi = d(x)i, and x(q)Λei 6= ∅. Let
J := {i ≤ k : qi = d(x)i and x(q)Λei 6= ∅}.
Since x ∈ ∂Λ, for each E ∈ x(q)F E(Λ) there exists t ∈ Nk such that x(q, q +t) ∈ E.
Since qi = d(x)i for all i ∈ J, the set Si∈J x(q)Λei contains no such segments of
x, and thus cannot be finite exhaustive. Since Λ is row-finite, Si∈J x(q)Λei
finite, so Si∈J x(q)Λei is not exhaustive. Thus there exists µ ∈ x(q)Λ such that
MCE(µ, ν) = ∅ for all ν ∈Si∈J x(q)Λei. By [13, Lemma 2.11], s(µ)Λ≤∞ 6= ∅. Let
z ∈ s(µ)Λ≤∞, and define y := x(0, q)µz. Then y ∈ Λ≤∞ by [13, Lemma 2.10].
Now we show that (x; (m, n)) ∼ (y; (m, n)). Condition (P3) is trivially satisfied.
To see that (P1) and (P2) hold, it suffices to show that n ∧ d(x) = n ∧ d(y). Firstly,
is
THE PATH SPACE OF A HIGHER-RANK GRAPH
13
let i ∈ J. If d(µz)i 6= 0, then (µz)(0, d(µ) + ei) ∈ MCE(µ, ν) for ν = (µz)(0, ei) ∈
r(µ)Λei = x(q)Λei, a contradiction. So for each i ∈ J, d(µz)i = 0, and hence
d(y)i = d(x)i. Now suppose that i /∈ J. Then either x(q)Λei = ∅ or qi < d(x)i. If
x(q)Λei = ∅ then d(y)i = d(x)i. So suppose that qi < d(x)i. Since n ∧ d(x) ≤ q, it
follows that ni < d(x)i and ni ≤ qi ≤ d(y)i, hence (n ∧ d(x))i = ni = (n ∧ d(y))i.
So n ∧ d(x) = n ∧ d(y).
(cid:3)
Proposition 4.13. Suppose that Λ is a k-graph, and that λ ∈ Λ. Then s(λ)∂Λ 6=
∅. If x, y ∈ s(λ)∂Λ, then λx, λy ∈ ∂Λ and (λx; (0, d(λ))) ∼ (λy; (0, d(λ))). More-
The following result allows us to identify Λ with a subgraph of eΛ.
over, there is an injective k-graph morphism ι : Λ →eΛ such that for λ ∈ Λ
ι(λ) = [λx; (0, d(λ))] for any x ∈ s(λ)∂Λ.
Proof. By [6, Lemma 5.15], we have v∂Λ 6= ∅ for all v ∈ Λ0. In particular, we have
s(λ)∂Λ 6= ∅. Let x, y ∈ s(λ)∂Λ. Then [6, Lemma 5.13(ii)] says that λx, λy ∈ ∂Λ.
It follows from the definition of ∼ that (λx; (0, d(λ))) ∼ (λy; (0, d(λ))). Then
straightforward calculations show that that ι is an injective k-graph morphism. (cid:3)
We want to extend ι to an injection of WΛ into W eΛ. The next proposition shows
that any injective k-graph morphism defined on Λ can be extended to WΛ.
Proposition 4.14. Let Λ, Γ be k-graphs and φ : Λ → Γ be a k-graph morphism.
Let x ∈ WΛ \ Λ, then φ(x) : Ωk,d(x) → WΓ defined by φ(x)(p, q) = φ(x(p, q)) belongs
to WΓ.
Proof. Follows from φ being a k-graph morphism.
(cid:3)
In particular, we can extend ι to paths with non-finite degree. We need to know
that composition works as expected for non-finite paths.
Proposition 4.15. Let Λ, Γ be k-graphs and φ : Λ → Γ be a k-graph morphism.
Let λ ∈ Λ, x ∈ s(λ)WΛ, and suppose that n ∈ Nk satisfies n ≤ d(x). Then
(1) φ(λ)φ(x) = φ(λx); and
(2) σn(φ(x)) = φ(σn(x)).
Proof. Follows from φ being a k-graph morphism.
(cid:3)
Remark 4.16. It follows that the extension of an injective k-graph morphism to
ι : WΛ → W eΛ.
WΛ is also injective. In particular, the map ι : Λ → eΛ has an injective extension
We need to be able to 'project' paths fromeΛ onto the embedding ι(Λ) of Λ. For
π([y; (m, n)]) = [y; (m ∧ d(y), n ∧ d(y))].
y ∈ ∂Λ define
(4.2)
Straightforward calculations show that π is a surjective functor, and is a projec-
tion in the sense that π(π([y; (m, n)])) = π([y; (m, n)]) for all [y; (m, n)] ∈ eΛ. In
particular, πι(Λ) = idι(Λ).
14
S.B.G. WEBSTER
Lemma 4.17. Let Λ be a k-graph. Suppose that λ, µ ∈eΛ, and that λ ∈ Z(µ). Then
π(λ) ∈ Z(π(µ)). If d(π(λ))i > d(π(µ))i for some i ≤ k, then d(µ)i = d(π(µ))i.
Proof. Write λ = [x; (m, m + d(λ))]. Then µ = [x; (m, m + d(µ))], so
π(λ) = [x; (m ∧ d(x), (m + d(λ)) ∧ d(x))], and
π(µ) = [x; (m ∧ d(x), (m + d(µ)) ∧ d(x))].
Since d(λ) ≥ d(µ), it follows that π(λ) ∈ Z(π(µ)).
If d(π(λ))i > d(π(µ))i, then d(x)i > mi + d(µ)i, so
d(π(µ))i = mi + d(µ)i − mi = d(µ)i.
(cid:3)
Lemma 4.18. Let Λ be a k-graph and µ, ν ∈eΛ. Then
π(MCE(µ, ν)) ⊂ MCE(π(µ), π(ν)).
Proof. Suppose that λ ∈ MCE(µ, ν). By Lemma 4.17 we have π(λ) ∈ Z(π(µ)) ∩
Z(π(ν)), hence d(π(λ)) ≥ d(π(µ)) ∨ d(π(ν)).
It remains to prove that d(π(λ)) = d(π(µ)) ∨ d(π(ν)). Suppose for a contra-
diction that there is some i ≤ k such that d(π(λ))i > max{d(π(µ))i, d(π(ν))i}.
By Lemma 4.17 we then have d(π(µ))i = d(µ)i and d(π(ν))i = d(ν)i. Then
d(λ)i ≥ d(π(λ))i > max{d(µ)i, d(ν)i}, contradicting that λ ∈ MCE(µ, ν).
(cid:3)
and π(λ) = π(µ). Then λ = µ.
Lemma 4.19. Let Λ be a k-graph, and let µ, λ ∈ ι(Λ0)eΛ be such that d(λ) = d(µ)
Proof. Since µ, λ ∈ ι(Λ0)eΛ and d(λ) = d(µ), we can write λ = [x; (0, n)] and µ =
[y; (0, n)] for some x, y ∈ ∂Λ and n ∈ Nk. We will show that (x; (0, n)) ∼ (y; (0, n)).
Conditions (P2) and (P3) are trivially satisfied. Since
[x; (0, n ∧ d(x))] = π(λ) = π(µ) = [y; (0, n ∧ d(y))],
we have (x; (0, n ∧ d(x))) ∼ (y; (0, n ∧ d(y))). Hence x(0, n ∧ d(x)) = y(0, n ∧ d(y)),
and (P1) is satisfied.
(cid:3)
embedding from Proposition 4.13.
Proof of Theorem 4.1. The existence of eΛ follows from Proposition 4.9, and the
To check that eΛ is finitely aligned, fix µ, ν ∈ eΛ, and α ∈ ι(Λ0)eΛr(µ). Then
MCE(µ, ν) = MCE(αµ, αν). Since Λ is finitely aligned, MCE(π(αµ), π(αν))
is finite. We will show that MCE(αµ, αν) = MCE(π(αµ), π(αν)).
It follows from Lemma 4.18 that MCE(αµ, αν) ≥ MCE(π(αµ), π(αν)). For
the opposite inequality, suppose λ, β are distinct elements of MCE(αµ, αν). Then
d(λ) = d(β). Since r(αµ), r(αν) ∈ ι(Λ0), Lemma 4.19 implies that π(λ) 6= π(β).
So MCE(αµ, αν) = MCE(π(αµ), π(αν)).
For the last part of the statement, we prove the contrapositive. Suppose that eΛ
is not row-finite. Let [x; m] ∈ eΛ0 and i ≤ k be such that [x; m]eΛei = ∞. Then
THE PATH SPACE OF A HIGHER-RANK GRAPH
15
(4.3)
[y; (n, n + ei)] only if (P1) fails. That is,
for each [y; (n, n + ei)] ∈ [x; m]eΛei we have [y; n] = [x; m], so [x; (m, m + ei)] 6=
Since [x; m]eΛei = ∞, there are infinitely many [y; (n, n+ei)] ∈ [x; m]eΛei satisfying
x(m ∧ d(x), (m + ei) ∧ d(x)) 6= y(n ∧ d(y), (n + ei) ∧ d(y)).
Remark 4.20. Suppose that Λ is a finitely aligned k-graph, that x ∈ ∂Λ and
that E ⊂ x(0)Λ. Since ι : Λ → ι(Λ) is a bijective k-graph morphism, we have
E ∈ x(0)F E(Λ) if and only if ι(E) ∈ [x; 0]F E(ι(Λ)).
(4.3). Hence x(m ∧ d(x))Λei = ∞.
(cid:3)
The following results show how sets of minimal common extensions and finite
Proposition 4.21 ([5, Lemma 2.25]). Suppose that Λ is a finitely aligned k-graph,
exhaustive sets in a k-graph Λ relate to those in eΛ.
and that v ∈ ι(Λ0). Then E ∈ vF E(ι(Λ)) implies that E ∈ vF E(eΛ).
Proof. Since ι(Λ) ⊂ eΛ, we have MCEι(Λ)(µ, ν) ⊂ MCE eΛ(µ, ν). Suppose that λ ∈
Lemma 4.22. Let Λ be a finitely aligned k-graph and let µ, ν ∈ ι(Λ). Then
MCEι(Λ)(µ, ν) = MCE eΛ(µ, ν).
MCE eΛ(µ, ν). It suffices to show that λ ∈ ι(Λ). Write µ = [x; (0, n)], ν = [y; (0, q)]
and λ = [z; (0, n ∨ q)]. Then λ ∈ Z(µ) ∩ Z(ν) implies that d(z) ≥ n ∨ q, hence
λ ∈ ι(Λ).
(cid:3)
Remark 4.23. Since there is a bijection from Λmin(µ, ν) onto MCE(µ, ν), it follows
5. Topology of Path Spaces under Desourcification
composition of x with the inclusion of Ωk,px in Ωk,d(x). Then π(x) is a k-graph
morphism. Our goal for this section is the following theorem.
from Lemma 4.22 that eΛmin(µ, ν) = ι(Λ)min(µ, ν) for all µ, ν ∈ ι(Λ).
We extend the projection π defined in (4.2) to the set of infinite paths in eΛ,
and prove that its restriction to ι(Λ0)eΛ∞ is a homeomorphism onto ι(∂Λ). For
x ∈ ι(Λ0)eΛ∞, let px = W{p ∈ Nk : x(0, p) ∈ ι(Λ)}, and define π(x) to be the
Theorem 5.1. Let Λ be a row-finite k-graph. Then π : ι(Λ0)eΛ∞ → ι(∂Λ) is a
Proposition 5.2. Let Λ be a finitely aligned k-graph. Let x ∈ ι(Λ0)eΛ∞. Suppose
that {yn : n ∈ Nk} ⊂ ∂Λ satisfy [yn; (0, n)] = x(0, n). Then
We first show that the range of π is a subset of ι(∂Λ).
homeomorphism.
ι(yn) = π(x) in W eΛ; and
(i) lim
n∈Nk
(ii) there exists y ∈ ∂Λ such that π(x) = ι(y), and for m, n ∈ Nk with m ≤ n ≤
px we have π(x)(m, n) = ι(y(m, n)).
16
S.B.G. WEBSTER
Proof. For part (i), fix a basic open set Z(µ \ G) ⊂ W eΛ containing π(x). Fix
have µ ∈ ι(Λ). Since n ≥ d(µ), we have [yn; (0, d(µ))] = µ.
Let α = ι−1(µ) and z ∈ s(α)∂Λ. Then [yn; (0, d(µ))] = µ = [αz; (0, d(µ))], and
n ≥ N :=Wν∈G d(µν). We first show that ι(yn) ∈ Z(µ). Since π(x) ∈ Z(µ), we
We now show that ι(yn) /∈ Sν∈G Z(µν). Fix ν ∈ G. If d(yn) (cid:3) d(µν), then
trivially we have ι(yn) /∈ Z(µν). Suppose that d(yn) ≥ d(µν). Since n ≥ d(µν),
we have
(P1) gives ι(yn(0, d(µ) ∧ d(yn))) = ι((αz)(0, d(µ))) = ι(α) = µ. So ι(yn) ∈ Z(µ).
x(0, d(µν)) = [yn; (0, n)](0, d(µν)) = ι(yn)(0, d(µν)) ∈ ι(Λ).
So ι(yn)(0, d(µν)) = x(0, d(µν)) = π(x)(0, d(µν)) 6= µν.
For part (ii), recall that ι is injective, then we can define y : Ωk,px → Λ by
ι(y(m, n)) = π(x)(m, n). So ι(y) = π(x). To see that y ∈ ∂Λ, fix m ∈ Nk such that
m ≤ d(y) and fix E ∈ y(m)F E(Λ). We seek t ∈ Nk such that y(m, m + t) ∈ E.
Let p := m +Wµ∈E d(µ). Then since m ≤ d(y) = px
[yp; (0, m)] = x(0, m) = π(x)(0, m) = ι(y(0, m)) = [y(0, m)y′; (0, m)]
for some y′ ∈ y(m)∂Λ. So (yp; (0, m)) ∼ (y(0, m)y′; (0, m)), hence
yp(0, m ∧ d(yp)) = (y(0, m)y′)(0, m ∧ d(y(0, m)y′)) = y(0, m)
by (P1). In particular, this implies that yp(m) = y(m). Since yp ∈ ∂Λ, there exists
t ∈ Nk such that yp(m, m + t) ∈ E. So m + t ≤ p, and we have
ι(yp(m, m + t)) = [yp; (0, p)](m, m + t) = x(0, p)(m, m + t) = x(m, m + t).
So x(m, m + t) ∈ ι(Λ), giving
ι(yp(m, m + t)) = x(m, m + t) = π(x)(m, m + t) = ι(y(m, m + t)).
Finally, injectivity of ι gives y(m, m + t) = yp(m, m + t) ∈ E.
(cid:3)
(4.2) when we regard finite paths as k-graph morphisms. The following lemma is
The next few lemmas ensure that our definition of π on eΛ∞ is compatible with
also crucial in showing that π is injective on ι(Λ0)eΛ∞.
Lemma 5.3. Let Λ be a finitely aligned k-graph. Let x ∈ ι(Λ0)eΛ∞. Suppose that
w ∈ ∂Λ satisfies π(x) = ι(w). Then x(0, n) = [w; (0, n)] for all n ∈ Nk.
Proof. Fix n ∈ Nk. Let z ∈ ∂Λ be such that x(0, n) = [z; (0, n)]. We aim to show
that (z; (0, n)) ∼ (w; (0, n)). That (P2) and (P3) hold follows immediately from
their definitions. It remains to verify condition (P1):
(5.1)
z(0, n ∧ d(z)) = w(0, n ∧ d(ω)).
Since π(x) = ι(w) we have d(w) = px. Thus
[w; (0, n ∧ px)] = ι(w(0, n ∧ px)) = x(0, n ∧ px) = [z; (0, n ∧ px)].
So (w; (0, n ∧ px)) ∼ (z; (0, n ∧ px)). It then follows from (P1) that
(5.2)
w(0, n ∧ px) = z(0, n ∧ px).
THE PATH SPACE OF A HIGHER-RANK GRAPH
17
Hence n ∧ d(z) ≥ n ∧ px. Furthermore,
x(0, n ∧ d(z)) = [z; (0, n ∧ d(z))] = ι(z(0, n ∧ d(z))) ∈ ι(Λ)
implies that n ∧ px ≥ n ∧ d(z). So n ∧ d(z) = n ∧ px, and (5.2) becomes (5.1), as
required.
(cid:3)
Remark 5.4. Suppose that Λ be a finitely aligned k-graph, and that y ∈ ∂Λ and
m, n ∈ Nk satisfy m ≤ n ≤ d(y). Then
[y; (m, n)] = [σm(y); (0, n − m)] = ι(σm(y)(0, n − m)) = ι(y(m, n)),
So [y; (m, n)] = ι(y(m, n)).
The next proposition shows that our definitions of π for finite and infinite paths
are compatible:
Proposition 5.5. Let Λ be a finitely aligned k-graph. Suppose that x ∈eΛ∞, and
m ≤ n ∈ Nk. Then π(x(m, n)) = π(x)(m ∧ px, n ∧ px).
Proof. Fix y ∈ ∂Λ such that π(x) = ι(y). Then
π(x(m, n)) = π([y; (m, n)]
by Lemma 5.3
= [y; (m ∧ px, n ∧ px)]
= ι(y(m ∧ px, n ∧ px))
= π(x)(m ∧ px, n ∧ px)
since d(y) = px
by Remark 5.4
by Proposition 5.2(ii) .
(cid:3)
ι(∂Λ) is a bijection.
We first show that it is a bijection, then show it is continuous. Openness is the
trickiest part, and the proof of it completes this section.
We can now show that π restricts to a homeomorphism of ι(Λ0)eΛ∞ onto ι(∂Λ).
Proposition 5.6. Let Λ be a finitely aligned k-graph. Then the map π : ι(Λ0)eΛ∞ →
let w ∈ ∂Λ and define x : Ωk → eΛ by x(p, q) = [w; (p, q)]. Then px = d(w), and
r(x) ∈ ι(Λ). To see that π(x) = ι(w), fix m, n ∈ Nk with m ≤ n ≤ d(w). Then
Proof. That π is injective follows from Lemma 5.3. To see that π is onto ι(∂Λ),
π(x)(m, n) = x(m, n)
by Proposition 5.5
= [w; (m, n)]
by Lemma 5.3
= ι(w(m, n))
by Remark 5.4
= ι(w)(m, n)
by Proposition 4.14.
(cid:3)
Proposition 5.7. Let Λ be a finitely aligned k-graph. Then π : ι(Λ0)eΛ∞ → ι(∂Λ)
is continuous.
18
S.B.G. WEBSTER
If Z(µ \ G) ∩ ι(∂Λ) = ∅, then
Proof. Fix a basic open set Z(µ \ G) ⊂ W eΛ.
π−1(Z(µ \ G) ∩ ι(∂Λ)) = ∅ is open. Suppose that Z(µ \ G) ∩ ι(∂Λ) 6= ∅, and fix
y ∈ Z(µ \ G) ∩ ι(∂Λ). Let F = G ∩ ι(Λ). We will show that
(5.3)
π−1(y) ∈ Z(µ \ F ) ∩(cid:0)eΛ∞ ∩ r−1(ι(Λ))(cid:1) ⊂ π−1(Z(µ \ G) ∩ ι(∂Λ)).
Since y ∈ Z(µ), it follows that π−1(y) ∈ Z(µ). To see that π−1(y) /∈Sβ∈F Z(µβ),
fix β ∈ F . First suppose that d(µβ) (cid:2) d(y). Then π−1(y)(0, d(µβ)) /∈ ι(Λ). Since
µβ ∈ ι(Λ), we have π−1(y)(0, d(µβ)) 6= µβ. Now suppose that d(µβ) ≤ d(y), then
π−1(y)(0, d(µβ)) = y(0, d(µβ)) 6= µβ.
We now show that Z(µ \ F ) ∩ ι(Λ0)eΛ∞ ⊂ π−1(Z(µ \ G) ∩ ι(∂Λ)). Let z ∈
Z(µ\F )∩ι(Λ0)eΛ∞. It suffices to show that π(z) ∈ Z(µ\G). Firstly, π(z)(0, d(µ)) =
z(0, d(µ)) = µ ∈ ι(Λ). To see that π(z) /∈ Sν∈G Z(µν), fix ν ∈ G. If d(µν) (cid:2)
d(π(z)), then trivially π(z) /∈ Z(µν). Suppose that d(µν) ≤ d(π(z)).
If ν /∈
ι(Λ), then π(z)(0, d(µν)) 6= µν. Otherwise, ν ∈ ι(Λ), then ν ∈ F and we have
π(z)(0, d(µν)) = z(0, d(µν)) 6= µν.
(cid:3)
open.
Proposition 5.8. Let Λ be a row-finite k-graph. Then π : ι(Λ0)eΛ∞ → ι(∂Λ) is
Proof. Fix π(y) ∈ π(Z(µ \ G) ∩ ι(Λ0)eΛ∞). Let ω ∈ ∂Λ be such that π(y) = ι(ω).
Define λ := y(0,Wν∈G d(µν)), and
F :=[{s(π(λ))ι(Λei) : d(λ)i > d(π(y))i}.
We claim that
First we show that π(y) ∈ Z(π(λ)). It follows from Lemma 5.3 that π(λ) =
π(y) ∈ Z(π(λ) \ F ) ∩ ι(∂Λ) ⊂ π(Z(µ \ G) ∩ ι(Λ0)eΛ∞).
[ω; (0, d(λ) ∧ d(ω))]. Since d(ω) = d(π(y)), Proposition 5.5 implies that
π(y)(0, d(π(λ))) = π(y)(0, d(λ) ∧ d(ω)) = π(y(0, d(λ))) = π(λ).
Now we show that π(y) /∈ Sf ∈F Z(π(λ)f ). Fix f ∈ F ; say d(f ) = ei. Then by
definition of F , d(λ)i > d(π(y))i = d(ω)i, and thus
d(π(λ))i = min{d(λ)i, d(ω)i} = d(ω)i = d(π(y))i.
So d(π(y)) (cid:3) d(π(λ)f ), and hence π(y) /∈ Z(π(λ)f ) as required.
Now we show that Z(π(λ) \ F ) ∩ ι(∂Λ) ⊂ π(Z(µ \ G) ∩ ι(Λ0)eΛ∞). Let π(β) ∈
Z(π(λ) \ F ) ∩ ι(∂Λ). We aim to show that β ∈ Z(µ \ G). Since Z(λ) ⊂ Z(µ \ G),
it suffices to show that β ∈ Z(λ). Clearly β ∈ Z(π(λ) \ F ). If d(λ) = d(π(λ))
then π(λ) = λ and we are done. Suppose that d(λ) > d(π(λ)), and let τ =
β(d(π(λ)), d(λ)). We know that β ∈ Z(π(λ)). We aim to use Lemma 4.19 to show
that τ = λ(d(π(λ)), d(λ)). Fix i ≤ k such that d(λ)i > d(π(λ))i. Then since
d(π(λ)) = d(λ) ∧ d(ω), we have d(λ)i > d(ω)i = d(π(y))i. Now β ∈ Z(π(λ) \ F )
implies that τ (0, ei) /∈ F . In particular, τ (0, ei) /∈ ι(Λ). We claim that d(π(τ )) = 0.
THE PATH SPACE OF A HIGHER-RANK GRAPH
19
Suppose, for a contradiction, that d(π(τ ))j > 0 for some j ≤ k. Then π(τ )(0, ej) =
τ (0, ej) /∈ ι(Λ). But π(τ ) ∈ ι(Λ) by definition of π. So we must have d(π(τ )) = 0,
which implies that
π(τ ) = r(τ ) = s(π(λ)) = π(λ(d(π(λ)), d(λ))).
Now Lemma 4.19 implies that τ = λ(d(π(λ)), d(λ)). Then
β(0, λ) = β(0, d(π(λ)))τ = π(λ)λ(d(π(λ)), d(λ)) = λ.
(cid:3)
Example 5.9. We can see that π is not open for non-row-finite graphs by con-
sidering the 1-graph E from Example 4.10 with 'desourcification' eE. Observe that
Z(µ1) ∩ ι(E0)eΛ∞ = {µ1µ2 · · · } is open in eE, and π(Z(µ1) ∩ ι(E0)eE∞) = {v}.
Since ∂E = E, any basic open set in ∂E containing v is of the form Z(v \ G)
for some finite G ⊂ E1. Since E1 is infinite, there is no finite G ⊂ E1 such that
Z(v \ G) ⊂ {v}. Hence {v} is not open in E, and π is not an open map.
Proof of Theorem 5.1. Propositions 5.6, 5.7 and 5.8 say precisely that π is a bijec-
tion, is continuous, and is open.
(cid:3)
Remark 5.10. Although πι(Λ0)eΛ∞ is open for all row-finite k-graphs, it behaves
particularly well with respect to cylinder sets for locally convex k-graphs. The
following discussion and example arose in preliminary work on a proof that π is
open when Λ is row-finite and locally convex. We have retained this example since
it helps illustrate some of the issues surrounding the map π.
Denote our standard topology for a finitely k-graph by τ1. The collection {Z(µ) :
µ ∈ Λ} of cylinder sets also form a base for a topology: they cover WΛ, and if
x ∈ Z(λ) ∩ Z(ν), then x ∈ Z(x(0, d(λ) ∨ d(ν))) ⊂ Z(λ) ∩ Z(ν). This topology,
denoted τ2, is not necessarily Hausdorff: we cannot separate any edge from its
range: if r(f ) ∈ Z(µ) then µ = r(f ), and thus f ∈ Z(µ).
It may seem reasonable to expect that {Z(µ) ∩ ∂Λ : µ ∈ Λ} is a base for the
restriction of τ1 to ∂Λ. However, this is not so. To see why, consider the 2-graph
of Example 2.11. Let y be the boundary path beginning with f0. So x, y ∈ ∂Λ.
Let µ be such that x ∈ Z(µ). Then µ = x0 . . . xn for some n ∈ N, so y ∈ Z(µ)
also. So the topology τ1 is not Hausdorff even when restricted to ∂Λ. Endowed
with τ2, it is easy to see how to separate these two points: y ∈ Z(f0) ∩ ∂Λ and
x ∈ Z(r(x) \ {f0}) ∩ ∂Λ, and these two sets are disjoint.
If we restrict ourselves to locally convex k-graphs, τ1 and τ2 do restrict to the
same topology on ∂Λ: certainly, for each µ ∈ Λ, we can realise a cylinder set Z(µ)
as a set of the form Z(µ\G) by taking G = ∅. Now suppose that x ∈ Z(µ\G)∩∂Λ.
We claim that with
νx := x(0,(cid:0)_α∈G
d(µα)(cid:1) ∧ d(x)),
we have x ∈ Z(νx) ∩ ∂Λ ⊂ Z(µ \ G) ∩ ∂Λ. Clearly we have x ∈ Z(νx) ∩ ∂Λ.
The containment requires a little more work. Clearly y ∈ Z(µ). Fix α ∈ G. We
will show that y /∈ Z(µα). If d(y) (cid:3) d(µα), then trivially y /∈ Z(µα). Suppose
20
S.B.G. WEBSTER
that d(y) ≥ d(µα). We claim that d(x) ≥ d(µα): suppose, for a contradiction,
that d(x) (cid:3) d(µα). Then there exists i ≤ k such that d(x)i < d(µα)i. Then
d(x)i = d(νx)i. Since x ∈ ∂Λ, we must have x(d(νx))Λei /∈ x(d(νx))F E(Λ). Since
Λ is locally convex, Lemma 2.13 implies that y(d(νx))Λei = x(d(νx))Λei = ∅. So
d(y)i = d(νx)i = d(x)i < d(µα)i, a contradiction. Hence d(x) ≥ d(µα). This
implies that d(νx) ≥ d(µα). So
y(0, d(µα)) = vx(0, d(µα)) = x(0, d(µα)) 6= µα.
Proposition 5.11. Suppose that Λ is a row-finite, locally convex k-graph, and let
µ ∈ ι(Λ0)eΛ. Then π(Z(µ) ∩ ι(Λ0)eΛ∞) = Z(π(µ)) ∩ ι(∂Λ). In particular, π is open.
Proof. We first show that π(Z(µ) ∩ ι(Λ0)eΛ∞) ⊂ Z(π(µ)) ∩ ι(∂Λ). Suppose that
π(y) ∈ π(Z(µ \ G) ∩ ι(Λ0)eΛ∞). Trivially π(y) ∈ ι(∂Λ). We will show that π(y) ∈
Z(π(µ) \ π(G)). Since y(0, d(µ)) = µ, we have
π(µ) = π(y(0, d(µ))) = π(y)(0, d(µ) ∧ d(π(y))).
So π(y) ∈ Z(π(µ)). Furthermore, d(π(µ)) = d(µ) ∧ d(π(y)).
Fix ν ∈ G. We will show that π(y) /∈ Z(π(µν)). Since y ∈ Z(µ \ G), we
have y(0, d(µν)) 6= µν. Since d(y(0, d(µν))) = d(µν) and r(y) = r(µν) ∈ ι(Λ0),
Lemma 4.19 implies that
π(µν) 6= π(y(0, d(µν))) = π(y)(0, d(µν) ∧ d(π(y))).
So π(Z(µ \ G) ∩ ι(Λ0)eΛ∞) ⊂ Z(π(µ) \ π(G)) ∩ ι(∂Λ).
Now suppose that ι(ω) ∈ Z(π(µ)) ∩ ι(∂Λ), and let y = π−1(ι(ω)). We show
that y ∈ Z(µ). Write µ = [z; (0, d(µ))]. Then π(µ) = [z; (0, d(µ) ∧ d(z))] and
y(0, d(µ)) = [ω; (0, d(µ))]. We claim that (z; (0, d(µ))) ∼ (ω; (0, d(µ))). That (P2)
and (P3) hold follows immediately from their definition. To show that (P1) is
satisfied, we must show that z(0, d(µ) ∧ d(z)) = w(0, d(µ) ∧ d(w)). Since π(y) ∈
Z(π(µ)), we have y ∈ Z(π(µ)). Then
[ω; (0, d(π(µ)))] = y(0, d(π(µ))) = π(µ) = [z; (0, d(µ) ∧ d(z))].
So (ω; (0, d(π(µ)))) ∼ (z; (0, d(µ) ∧ d(z))). Then (P1) implies that
ω(0, d(π(µ))) = ω(0, d(π(µ)) ∧ d(ω)) = z(0, d(µ) ∧ d(z)),
and d(π(µ)) = d(µ) ∧ d(z). We will show d(µ) ∧ d(w) = d(π(µ)). Fix i ≤ k. We
argue the following cases separately:
(1) If d(π(µ))i = d(µ)i, we have d(w) ≥ d(π(µ)) = d(µ)i. Hence (d(µ) ∧
d(w))i = d(µ)i = d(π(µ))i.
(2) If d(π(µ))i < d(µ)i, it requires a little more work:
Since d(µ)i > d(π(µ))i = min{d(µ)i, d(z)i}, we have d(π(µ))i = d(z)i. Then
z ∈ ∂Λ implies that z(d(π(µ)))Λei /∈ z(d(π(µ)))F E(Λ). By Lemma 2.13, we have
z(d(π(µ)))Λei = ∅, and hence ω(d(π(µ)))Λei = ∅. So d(ω)i = d(π(µ))i < d(µ)i,
giving (d(µ) ∧ d(ω))i = d(ω)i = d(π(µ))i.
(cid:3)
THE PATH SPACE OF A HIGHER-RANK GRAPH
21
6. High-Rank Graph C ∗-algebras
Definition 6.1. Let Λ be a finitely aligned k-graph. A Cuntz-Krieger Λ-family in
a C ∗-algebra B is a collection {tλ : λ ∈ Λ} of partial isometries satisfying
(CK1) {sv : v ∈ Λ0} is a set of mutually orthogonal projections;
(CK2) sµsν = sµν whenever s(µ) = r(ν);
(CK3) s∗
β for all µ, ν ∈ Λ; and
µ) = 0 for every v ∈ Λ0 and E ∈ vF E(Λ).
µsν =P(α,β)∈Λmin(µ,ν) sαs∗
(CK4) Qµ∈E(sv − sµs∗
The C ∗-algebra C ∗(Λ) of a k-graph Λ is the universal C ∗-algebra generated by
a Cuntz-Krieger Λ-family {sλ : λ ∈ Λ}.
Remark 6.2. The following Theorem appears as [5, Theorem 2.28]. Farthing alerted
us to an issue in the proof of the theorem. It contains a claim which is proved in
cases, and in the proof of Case 1 of the claim (on page 189), there is an error when
i0 is such that mi0 = d(x)i0 + 1. Then ai0 = d(x)i0, and [5, Equation (2.13)] gives
ti0 ≤ d(z)i0; not ti0 ≥ d(z)i0 as stated.
Theorem 6.3. Let Λ be a row-finite k-graph. Let C ∗(Λ) and C ∗(eΛ) be generated by
the Cuntz-Krieger families {sλ : λ ∈ Λ} and {tλ : λ ∈eΛ}. Then the sumPv∈ι(Λ0) tv
converges strictly to a full projection p ∈ M(C ∗(eΛ)) such that pC ∗(eΛ)p = C ∗({tι(λ) :
λ ∈ Λ}), and sλ 7→ tι(λ) determines an isomorphism ς : C ∗(Λ) ∼= pC ∗(eΛ)p.
{tλ : λ ∈eΛ} is a Cuntz-Krieger eΛ-family, then {tλ : λ ∈ ι(Λ)} is a Cuntz-Krieger
Proposition 6.4 ([5, Theorem 2.26]). Let Λ be a finitely aligned k-graph.
Before proving Theorem 6.3, we need the following results.
ι(Λ)-family.
If
Remark 6.5. Let Λ be a finitely aligned k-graph.
properties of C ∗(Λ) and C ∗(ι(Λ)) that C ∗(Λ) ∼= C ∗(ι(Λ)).
It follows from the universal
Proposition 6.6 ([5, Theorem 2.27]). Let Λ be a finitely aligned k-graph, and let
{tλ : λ ∈eΛ} be the universal Cuntz-KriegereΛ-family which generates C ∗(eΛ). Then
C ∗(Λ) is isomorphic to the subalgebra of C ∗(eΛ) generated by {tλ : λ ∈ ι(Λ)}.
Lemma 6.7. Suppose that Λ is a finitely aligned k-graph. Let λ ∈eΛ, and let λ′ =
λ(d(π(λ)), d(λ)). Suppose that x ∈ ∂Λ satisfies ι(r(x)) = r(λ′) and d(x)∧d(λ′) = 0.
Then λ′ = [x; (0, d(λ′))].
Proof. Write λ = [y; (0, d(λ))], then λ′ = [y; (d(λ)∧d(y), d(λ))]. We must show that
(y; (d(λ) ∧ d(y), d(λ)) ∼ (x; (0, d(λ′)). That conditions (P2) and (P3) hold follows
immediately from their definitions. It remains to show that (P1) is satisfied. Since
d(x) ∧ d(λ′) = 0, it suffices to show that y(d(λ) ∧ d(y)) = x(0). We have
ι(x(0)) = ι(r(x)) = r(λ′) = [y; d(λ) ∧ d(y)] = ι(y(d(λ) ∧ d(y))).
Injectivity of ι then gives y(d(λ) ∧ d(y)) = x(0).
(cid:3)
22
S.B.G. WEBSTER
Lemma 6.8. Let λ ∈eΛ. Let λ′ = λ(d(π(λ)), d(λ)) and define
{α ∈ s(π(λ))ι(Λ)ei : MCE(α, λ′) = ∅}.
Gλ :=
k[i=1
It suffices to show that MCE(µν, λ′) 6= ∅. Write µν = [z; (0, d(µν))].
Then Gλ ∪ {λ′} ∈ s(π(λ))F E(eΛ).
Proof. Fix µ ∈ s(π(λ))eΛ, and suppose that MCE(µ, α) = ∅ for all α ∈ Gλ. We
will show that MCE(µ, λ′) 6= ∅. Fix ν ∈ s(µ)eΛd(µ)∨d(λ′ )−d(µ). Then d(µν) ≥ d(λ′).
We first show that d(λ′) ∧ d(π(µν)) = 0. Suppose for a contradiction that
d(λ′) ∧ d(π(µν)) > 0. So we have d(λ′) ∧ d(µν) ∧ d(z) > 0, hence there exists i ≤ k
such that d(λ′)i, d(µν)i, and d(z)i are all greater than zero. Then (µν)(0, ei) =
[z; (0, ei)] = ι(z)(0, ei) ∈ ι(Λ). Since πι(Λ) = idι(Λ) and π(λ′) = s(π(λ)) 6= λ′, we
have λ′ /∈ ι(Λ). This implies that (µν)(0, ei) 6= λ′(0, ei). So MCE((µν)(0, ei), λ′) =
∅, and thus (µν)(0, ei) ∈ Gλ. But MCE(µν(0, ei), µν) 6= ∅, which implies that
MCE(µ, µν(0, ei)) 6= ∅. This contradicts our supposition that MCE(µ, α) = ∅ for
all α ∈ Gλ. So d(λ′) ∧ d(π(µν)) = 0.
Since d(µν) ≥ d(λ′), we have
d(z) ∧ d(λ′) = d(z) ∧ d(µν) ∧ d(λ′) = d(π(µν)) ∧ d(λ′) = 0
Since r(λ′) = r(µν) = ι(r(z)), it follows from Lemma 6.7 that λ′ = [z; (0, λ′)].
(cid:3)
Thus µν = [z; (0, µν)] ∈ MCE(µν, λ′).
Proof of Theorem 6.3. Let A := C ∗({tλ : λ ∈ ι(Λ)}). Then A ∼= C ∗(Λ) by Propo-
ptλt∗
sition 6.6. We will show that A is a full corner of C ∗(eΛ).
Following the argument of [10, Lemma 2.10], the sum Pv∈ι(Λ0) tv converges
strictly in M(C ∗(eΛ)) to a projection p satisfying
ifer(λ),er(µ) ∈ ι(Λ0);
µp =(tλt∗
otherwise.
The standard argument shows that p is a full projection in M(C ∗(eΛ)). It follows
from (6.1) that A ⊂ pC ∗(eΛ)p. Now suppose that λ, µ ∈ ι(Λ0)eΛ. We will show that
µp ∈ A. Ifes(λ) 6=es(µ), then (CK1) implies that ptλt∗
es(λ) =es(µ). We first show that
λ(d(π(λ)), d(λ)) = µ(d(π(µ)), d(µ)).
µp = 0 ∈ A. Suppose that
ptλt∗
µ
0
(6.1)
(6.2)
Let x, y ∈ ∂Λ such that λ = [x; (0, d(λ))] and µ = [y; (0, d(µ))]. Let
λ′ = λ(d(π(λ)), d(λ)) = [x; (d(λ) ∧ d(x), d(λ)]
µ′ = µ(d(π(µ)), d(µ)) = [y; (d(µ) ∧ d(y), d(µ))].
and
THE PATH SPACE OF A HIGHER-RANK GRAPH
23
We claim that λ′ = µ′. Condition (P2) is trivially satisfied, and (P1) and (P3)
λ′ = µ′.
follow from the vertex equivalence [x; d(λ)] = es(λ) = es(µ) = [y; d(µ)]. Hence
Claim 6.3.1. Let Gλ :=Sk
i=1{α ∈ s(π(λ))ι(Λ)ei : MCE(α, λ′) = ∅}. Then
tλ′t∗
Proof. Lemma 6.8 implies that Gλ ∪{λ′} is finite exhaustive, so (CK4) implies that
λ′ = Yα∈Gλ(cid:0)ts(π(λ)) − tαt∗
α(cid:1)
Yβ∈Gλ∪{λ′}(cid:0)ts(π(λ)) − tβt∗
β(cid:1) = 0.
β(cid:1) =(cid:16) Yα∈Gλ(cid:0)ts(π(λ)) − tαt∗
α)(cid:17) −(cid:16)tλ′t∗
λ′ Yα∈Gλ
λ′ − Xγ∈MCE(λ′,α)
β(cid:1) = Yα∈Gλ(cid:0)ts(π(λ)) − tαt∗
α) = tλ′t∗
(ts(π(λ)) − tαt∗
Furthermore,
Yβ∈Gλ∪{λ′}(cid:0)ts(π(λ)) − tβt∗
=(cid:16) Yα∈Gλ
Fix α ∈ Gλ. By [13, Lemma 2.7(i)],
So
0 = Yβ∈Gλ∪{λ′}(cid:0)ts(π(λ)) − tβt∗
Now we put the pieces together:
ptλt∗
µp = tλt∗
µ
tλ′t∗
λ′(ts(π(λ)) − tαt∗
tγt∗
γ = tλ′t∗
λ′.
λ′)
α(cid:1)(cid:17)(ts(π(λ)) − tλ′t∗
α)(cid:17).
(ts(π(λ)) − tαt∗
α(cid:1) − tλ′t∗
λ′.
(cid:3)Claim
pC ∗(eΛ)p.
= tπ(λ)tλ′t∗
λ′t∗
π(µ)
by (6.2)
= tπ(λ) Yα∈Gλ(cid:0)ts(π(λ)) − tαt∗
α(cid:1)t∗
π(µ)
by Claim 6.3.1.
which is an element of A since π(λ), π(µ), α ∈ ι(Λ) for all α ∈ Gλ. So A =
(cid:3)
7. The Diagonal and the Spectrum
For k-graph Λ, we call C ∗{sµs∗
µ : µ ∈ Λ} ⊂ C ∗(Λ) the diagonal C ∗-algebra of Λ
and denote it DΛ, dropping the subscript when confusion is unlikely. For a commu-
tative C ∗-algebra A, denote by ∆(A) the spectrum of A. Given a homomorphism
π : A → B of commutative C ∗-algebras, define by π∗ the induced map from ∆(B)
to ∆(A) such that π∗(f )(y) = f (π(y)) for all f ∈ ∆(B) and y ∈ A.
24
S.B.G. WEBSTER
Theorem 7.1. Let Λ be a row-finite higher-rank graph. Let p ∈ M(C ∗(eΛ)) and
ς : C ∗(Λ) ∼= pC ∗(eΛ)p be from Theorem 6.3. Then the restriction ςDΛ =: ρ is an
isomorphism of DΛ onto pD eΛp. Let π : ι(Λ0)eΛ∞ → ι(∂Λ) be the homeomorphism
ι(Λ0)eΛ∞ → ∆(pD eΛp) such that the following diagram commutes.
from Theorem 5.1, then there exist homeomorphisms hΛ : ∂Λ → ∆(DΛ) and η :
ι(∂Λ)
π
ι(Λ0)eΛ∞
η
∆(pD eΛp)
hΛ ◦ ι−1
∆(DΛ)
ρ∗
As in [11], for a finite subset F ⊂ Λ, define
∨F := [G⊂F
MCE(G) = [G⊂F(cid:8)λ ∈ \µ∈G
µΛ : d(λ) = _µ∈G
d(µ)(cid:9).
Lemma 7.2. Let Λ be a finitely aligned k-graph and let F be a finite subset of Λ.
Suppose that r(λ) ∈ F for each λ ∈ F . For µ ∈ F , define
q∨F
µ
:= sµs∗
(sµs∗
µ − sµµ′s∗
µµ′).
µ Yµµ′∈∨F \{µ}
µ are mutually orthogonal projections in span{sµs∗
µ : µ ∈ ∨F }, and for
Then the q∨F
each ν ∈ ∨F
(7.1)
Proof. Since
sµs∗
µ Yµµ′∈∨F \{µ}
sνs∗
ν = Xνν ′∈∨F
q∨F
νν ′
sνs∗
ν = Xνν ′∈∨F
q∨F
νν ′
(sµs∗
µ − sµµ′s∗
µµ′) = sµs∗
(sr(µ) − sµµ′s∗
µµ′),
µ Yµµ′∈∨F,d(µ′)6=0
[11, Proposition 8.6] says precisely that the q∨F
µ
tions. That
are mutually orthogonal projec-
is established in the proof of [11, Proposition 8.6] on page 421.
(cid:3)
Remark 7.3. We have
q∨F
µ = sµ(cid:16) Yµ′∈s(µ)Λ\{s(µ)}
µµ′∈∨F
(ss(µ) − sµ′s∗
µ′)(cid:17)s∗
µ.
This follows from a straightforward induction on ∨ F .
The following lemma can be verified through routine calculation. The reader is
referred to the author's PhD thesis for details.
THE PATH SPACE OF A HIGHER-RANK GRAPH
25
Lemma 7.4 ([19, Lemma A.0.7]). Let A be a C ∗-algebra, let p be a projection in
A, let Q be a finite set of commuting subprojections of p and let q0 be a nonzero
subprojection of p. Then Qq∈Q(p − q) is a projection. If q0 is orthogonal to each
q ∈ Q, then q0Qq∈Q(p − q) = q0, so in particular,Qq∈Q(p − q) 6= 0.
Proposition 7.5. Let Λ be a finitely aligned k-graph. Then D = span{sµs∗
Λ}, and for each x ∈ ∂Λ there exists a unique h(x) ∈ ∆(D) such that
µ : µ ∈
h(x)(sµs∗
µ) =(1
0
if x = µµ′
otherwise.
Moreover, x 7→ h(x) is a homeomorphism h : ∂Λ → ∆(D).
Proof. Let µ, ν ∈ Λ. It follows from (CK3) that
(sµs∗
µ)(sνs∗
ν) = Xλ∈MCE(µ,ν)
sλs∗
λ,
to zero we can assume that each path in F has its range in F , and write
µ ∈ span{sµs∗
µ : µ ∈ Λ}. By setting extra coefficients
hence D = span{sµs∗
µ : µ ∈ Λ}.
Fix x ∈ ∂Λ andPµ∈F bµsµs∗
Xµ∈F
bµsµs∗
µ = Xµ∈∨F
bµsµs∗
µ.
Let n = W{p ∈ Nk : x(0, p) ∈ ∨F }. Since ∨F is a finite set of finite paths, n
is finite. Since ∨F is closed under minimal common extensions, x(0, n) ∈ ∨F .
Furthermore, since x ∈ ∂Λ, we have
Fx := {µ′ ∈ x(n)Λ \ {x(n)} : x(0, n)µ′ ∈ ∨F } /∈ x(n)F E(Λ).
So there exists ν ∈ x(n)Λ such that for each µ′ ∈ Fx, MCE(ν, µ′) = ∅. Then
sνs∗
ν and
Q = {sµ′s∗
µ′ = 0 for all µ′ ∈ Fx. Applying Lemma 7.4 with p = sx(n), q0 = sνs∗
νsµ′s∗
µ′) 6= 0. So
µ′ : µ′ ∈ Fx}, we haveQµ′∈Fx(sx(n) − sµ′s∗
qF
x(0,n) = sx(0,n) Yµ′∈Fx
(sx(n) − sµ′s∗
µ′)s∗
x(0,n) 6= 0.
26
We have
(cid:13)(cid:13)(cid:13) Xν∈∨F
Hence the formula
(7.2)
bµsµs∗
by (7.1)
S.B.G. WEBSTER
ν
=
ν∈Z(µ)
ν∈Z(µ)
max
{ν∈∨F :q∨F
ν (cid:13)(cid:13)(cid:13)
µ(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13) Xν∈∨F(cid:16) Xµ∈∨F
bµ(cid:17)q∨F
bµ(cid:12)(cid:12)(cid:12)
6=0}(cid:12)(cid:12)(cid:12) Xµ∈∨F
bµ(cid:12)(cid:12)(cid:12)
≥(cid:12)(cid:12)(cid:12) Xµ∈∨F
bµ(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12) Xµ∈F
h(x)(cid:16)Xµ∈F
bµsµs∗
x(0,n)∈Z(µ)
x(0,n)∈Z(µ)
µ(cid:17) = Xµ∈F
x∈Z(µ)
bµ,
since q∨F
x(0,n) 6= 0
since bµ = 0 for µ ∈ ∨F \ F .
determines a norm-decreasing linear map on span{sµs∗
µ : µ ∈ Λ}.
To see that h(x) is a homomorphism, it suffices to show that
(7.3)
h(x)(sµs∗
µsαs∗
α) = h(x)(sµs∗
µ) h(x)(sαs∗
α).
Calculating the right hand side of (7.3) yields
h(x)(sµs∗
µ) h(x)(sαs∗
Calculating the left hand side of (7.3) gives
h(x)(sµs∗
µsαs∗
0 otherwise.
α) =(1 if x ∈ Z(µ) ∩ Z(α)
α) = h(x)(cid:16) Xλ∈MCE(µ,α)
λ(cid:17).
sλs∗
There exists at most one λ ∈ MCE(µ, α) such that x ∈ Z(λ). Such a λ exists if
and only if x ∈ Z(µ) ∩ Z(α), so
h(x)(sµs∗
µsαs∗
α) =(1 if x ∈ Z(α) ∩ Z(µ)
0 otherwise.
Thus we have established (7.3), hence h(x) is a homomorphism, and thus extends
uniquely to a nonzero homomorphism h(x) : D → C.
We claim the map h : ∂Λ → ∆(D) is a homeomorphism. The trickiest part is
to show h is onto:
Claim 7.5.1. The map h is surjective.
THE PATH SPACE OF A HIGHER-RANK GRAPH
27
Proof. Fix φ ∈ ∆(D). We seek x ∈ ∂Λ such that h(x) = φ. For each n ∈ Nk,
{sµs∗
µ : d(µ) = n} are mutually orthogonal projections. It follows that for each
n ∈ Nk there exists at most one νn ∈ Λn such that φ(sνns∗
νn) = 1. Let S denote
the set of n for which such νn exist. If ν = µν′ and φ(sνs∗
ν) = 1, then
1 = φ(sνs∗
ν) = φ(sνs∗
νsµs∗
µ) = φ(sνs∗
ν)φ(sµs∗
µ) = φ(sµs∗
µ).
This implies that if n ∈ S and m ≤ n, then m ∈ S and νm = νn(0, m). Set
N := ∨S, and define x : Ωk,N → Λ by x(p, q) = νq(p, q). Then since each νq is a
k-graph morphism, so is x.
We now show that x ∈ ∂Λ. Fix n ∈ Nk such that n ≤ d(x), and E ∈ x(n)F E(Λ).
We seek m ∈ Nk such that x(n, n + m) ∈ E. Since E is finite exhaustive, (CK4)
λ) = 0. Multiplying on the left by sx(0,n) and on the
implies thatQλ∈E(sx(n) − sλs∗
x(0,n) yields
right by s∗
(sx(0,n)s∗
x(0,n) − sx(0,n)λs∗
x(0,n)λ) = 0.
Yλ∈E
Thus, since φ is a homomorphism, there exists λ ∈ E such that
0 = φ(sx(0,n)s∗
x(0,n)) − φ(sx(0,n)λs∗
x(0,n)λ)
νn) − φ(sx(0,n)λs∗
x(0,n)λ)
= φ(sνns∗
= 1 − φ(sx(0,n)λs∗
x(0,n)λ)
So φ(sx(0,n)λs∗
x ∈ ∂Λ.
x(0,n)λ) = 1. Thus x(0, n)λ = νn+d(λ) = x(0, n + d(λ)), and hence
Now we must show that h(x) = φ. For each µ ∈ Λ we have
φ(sµs∗
µ) = 1 ⇐⇒ d(µ) ∈ S and νd(µ) = µ
⇐⇒ x(0, d(µ)) = µ
⇐⇒ h(x)(sµs∗
µ) = 1.
Since φ(sµs∗
µ) and h(x)(sµs∗
µ) take values in {0, 1}, we have h(x) = φ.
(cid:3)Claim
To see that h is injective, suppose that h(x) = h(y). Then for each n ∈ Nk, we
have
h(y)(sx(0,n∧d(x))s∗
x(0,n∧d(x))) = h(x)(sx(0,n∧d(x))s∗
x(0,n∧d(x))) = 1.
Hence y(0, n ∧ d(x)) = x(0, n ∧ d(x)). By symmetry, we also have y(0, n ∧ d(y)) =
x(0, n ∧ d(y)) for all n.
In particular, d(x) = d(y) and y(0, n) = x(0, n) for all
n ≤ d(x). Thus x = y.
Recall that ∆(D) carries the topology of pointwise convergence. For openness,
it suffices to check that h−1 is continuous. Suppose that h(xn) → h(x). Fix a basic
open set Z(µ) containing x, so h(x)(sµs∗
µ) ∈ {0, 1} for all
n, for large enough n, we have h(xn)(sµs∗
µ) = 1. So xn ∈ Z(µ). For continuity,
a similarly straightforward argument shows that if xn → x, then h(xn)(sµs∗
µ) →
µ) = 1. Since h(xn)(sµs∗
28
S.B.G. WEBSTER
h(x)(sµs∗
D by an ε/3 argument.
µ). This convergence extends to span{sµs∗
µ : µ ∈ Λ} by linearity, and to
We can now prove our main result.
(cid:3)
Proof of Theorem 7.1. Let Λ be a row-finite k-graph, andeΛ be the desourcification
described in Proposition 4.9. Let {sλ : λ ∈ Λ} and {tλ : λ ∈ eΛ} be universal
Cuntz-Krieger families in C ∗(Λ) and C ∗(eΛ). Let A be the C ∗-subalgebra of C ∗(eΛ)
generated by {tλ : λ ∈ ι(Λ)}, and define the diagonal subalgebra of A by DA :=
span{tλt∗
λ in the proof Theorem 6.3 yields
∼= pD eΛp as
DA
required.
∼= pD eΛp. Since A ∼= C ∗(Λ), it follows that DA
λ : λ ∈ ι(Λ)}. Replacing tλt∗
∼= DΛ. Thus DΛ
µ with tλt∗
We now construct η and show that it is a homeomorphism. That p commutes
with D eΛ implies that pD eΛp is an ideal in D eΛ. Then [14, Propositions A26(a) and
A27(b)] imply that map k : φ 7→ φpD eΛp is a homeomorphism of {φ ∈ ∆(D eΛ) :
∆(pD eΛp).
φpD eΛp 6= 0} onto ∆(pD eΛp). Since eΛ is row finite with no sources, ∂eΛ = eΛ∞. Let
h eΛ : eΛ∞ → ∆(D eΛ) be the homeomorphism obtained from Proposition 7.5. Then
h eΛ(x) ∈ dom(k) for all x ∈ ι(Λ0)eΛ∞. Define η := k ◦ h eΛι(Λ0)eΛ∞ : ι(Λ0)eΛ∞ →
fix x ∈ ι(Λ0)eΛ∞ and µ ∈ Λ and show that
We now show that hΛ ◦ ι−1 ◦ π = ρ∗ ◦ η. Since ρ is an isomorphism, it suffices to
(7.4)
(hΛ ◦ ι−1 ◦ π)(x)(sµs∗
µ) = (ρ∗ ◦ η)(x)(sµs∗
µ).
Let ω ∈ ∂Λ be such that π(x) = ι(ω). Then the left-hand side of (7.4) becomes
(hΛ ◦ ι−1 ◦ π)(x)(sµs∗
µ) = hΛ(w)(sµs∗
µ) =(1 if ω ∈ Z(µ)
0 otherwise.
Since r(x) ∈ ι(Λ0), the right-hand side of (7.4) simplifies to
(ρ∗ ◦ η)(x)(sµs∗
µ) = η(x)(ρ(sµs∗
µ)) = h eΛ(x)(tι(µ)t∗
ι(µ)) =(1 if x ∈ Z(ι(µ))
0 otherwise.
We claim that x ∈ Z(ι(µ)) if and only if ω ∈ Z(µ). Suppose that x ∈ Z(ι(µ)).
Since µ ∈ Λ and π(x) = ι(ω), we have π(x(0, d(µ))) = π(ι(µ)) = ι(µ). So
d(π(x(0, d(µ)))) = d(µ), and thus d(x) ∧ d(w) ≥ d(µ). So d(ω) ≥ d(µ). Then
THE PATH SPACE OF A HIGHER-RANK GRAPH
29
we have
x ∈ Z(ι(µ)) ⇐⇒ x(0, d(µ)) = ι(µ)
since ι preserves degree
⇐⇒ [ω; (0, d(µ))] = ι(µ)
⇐⇒ ι(ω(0, d(µ))) = ι(µ)
⇐⇒ ω(0, d(µ)) = µ
⇐⇒ ω ∈ Z(µ).
by Lemma 5.3
by Remark 5.4
since ι is a injective
So equation (7.4) holds, and we are done.
(cid:3)
References
[1] T. Bates, D. Pask, I. Raeburn, and W. Szyma´nski, The C ∗-algebras of row-finite graphs,
New York J. Math. 6 (2000), 307 -- 324.
[2] J. Cuntz and W. Krieger, A class of C ∗-algebras and topological Markov chains, Invent.
Math. 56 (1980), 251 -- 268.
[3] D. Drinen and M. Tomforde, The C ∗-algebras of arbitrary graphs, Rocky Mountain J. Math.
35 (2005), 105 -- 135.
[4] M. Enomoto and Y. Watatani, A graph theory for C ∗-algebras, Math. Japon. 25 (1980),
435 -- 442.
[5] C. Farthing, Removing sources from higher-rank graphs, J. Operator Theory 60 (2008),
165 -- 198.
[6] C. Farthing, P. S. Muhly, and T. Yeend, Higher-rank graph C ∗-algebras: an inverse semi-
group and groupoid approach, Semigroup Forum 71 (2005), 159 -- 187.
[7] A. Kumjian and D. Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20.
[8] D. Pask, I. Raeburn, M. Rørdam, and A. Sims, Rank-two graphs whose C ∗-algebras are direct
limits of circle algebras, J. Funct. Anal. 239 (2006), 137 -- 178.
[9] A. L. T. Paterson and A. E. Welch, Tychonoff 's theorem for locally compact spaces and an
elementary approach to the topology of path spaces, Proc. Amer. Math. Soc. 133 (2005),
2761 -- 2770.
[10] I. Raeburn, Graph algebras, Published for the Conference Board of the Mathematical Sci-
ences, Washington, DC, 2005, vi+113.
[11] I. Raeburn and A. Sims, Product systems of graphs and the Toeplitz algebras of higher-rank
graphs, J. Operator Theory 53 (2005), 399 -- 429.
[12] I. Raeburn, A. Sims, and T. Yeend, Higher-rank graphs and their C ∗-algebras, Proc. Edinb.
Math. Soc. (2) 46 (2003), 99 -- 115.
[13] I. Raeburn, A. Sims, and T. Yeend, The C ∗-algebras of finitely aligned higher-rank graphs,
J. Funct. Anal. 213 (2004), 206 -- 240.
[14] I. Raeburn and D. P. Williams, Morita equivalence and continuous-trace C ∗-algebras, Amer-
ican Mathematical Society, Providence, RI, 1998, xiv+327.
[15] J. Renault, A groupoid approach to C ∗-algebras, Springer, Berlin, 1980, ii+160.
[16] D. Roberston, Simplicity of C ∗-algebras associated to row-finite locally convex higher-rank
graphs, Honours thesis, University of Newcastle, 2006.
[17] D. Robertson and A. Sims, Simplicity of C ∗-algebras associated to row-finite locally convex
higher-rank graphs, Israel J. Math. 172 (2009), 171 -- 192.
[18] G. Robertson and T. Steger, Affine buildings, tiling systems and higher rank Cuntz-Krieger
algebras, J. Reine Angew. Math. 513 (1999), 115 -- 144.
[19] S. B. G. Webster, Directed Graphs and k-graphs: Topology of the Path Space and How It
Manifests in the Associated C ∗-algebra, PhD Thesis, University of Wollongong, 2010.
30
S.B.G. WEBSTER
Samuel Webster, School of Mathematics and Applied Statistics, University of
Wollongong, NSW 2522, AUSTRALIA
E-mail address: [email protected]
|
1604.02911 | 2 | 1604 | 2016-11-15T20:01:35 | Inner amenability, property Gamma, McDuff II_1 factors and stable equivalence relations | [
"math.OA",
"math.DS",
"math.GR"
] | We say that a countable group $G$ is McDuff if it admits a free ergodic probability measure preserving action such that the crossed product is a McDuff II_1 factor. Similarly, $G$ is said to be stable if it admits such an action with the orbit equivalence relation being stable. The McDuff property, stability, inner amenability and property Gamma are subtly related and several implications and non implications were obtained in [Ef73,JS85,Va09,Ki12a,Ki12b]. We complete the picture with the remaining implications and counterexamples. | math.OA | math |
Inner amenability, property Gamma,
McDuff II1 factors and stable equivalence relations
by Tobe Deprez1 and Stefaan Vaes2
Abstract
We say that a countable group G is McDuff if it admits a free ergodic probability measure
preserving action such that the crossed product is a McDuff II1 factor. Similarly, G is said
to be stable if it admits such an action with the orbit equivalence relation being stable. The
McDuff property, stability, inner amenability and property Gamma are subtly related and
several implications and non implications were obtained in [E73, JS85, V09, K12a, K12b].
We complete the picture with the remaining implications and counterexamples.
1
Introduction
Murray and von Neumann proved in [MvN43] that the hyperfinite II1 factor R is not isomorphic
to the free group factor L(F2), by showing that R has property Gamma while L(F2) has not.
Recall that a tracial von Neumann algebra (M, τ ) has property Gamma if there exists a net of
unitaries un ∈ M such that un → 0 weakly and limn kxun − unxk2 = 0 for all x ∈ M , where
kxk2 =pτ (x∗x).
We say that a countable group G has property Gamma when L(G) does. In [E73], it was shown
that if G has property Gamma, then G is inner amenable : there exists a conjugation invariant
mean m on G such that m(U ) = 0 for every finite subset U ⊂ G. The converse does not
hold : an inner amenable group G with infinite conjugacy classes (icc) but not having property
Gamma was constructed in [V09].
If a II1 factor M admits non commuting central sequences, more precisely if the central sequence
algebra M ′ ∩ M ω is non abelian, then M ∼= M ⊗ R and M is said to be McDuff, see [McD69].
The counterpart of the McDuff property for equivalence relations was introduced in [JS85]
and is called stability. A countable ergodic probability measure preserving (pmp) equivalence
relation R is called stable if R ∼= R × R0, where R0 is the unique hyperfinite ergodic pmp
In [JS85, Theorem 3.4], stability is characterized in terms of central
equivalence relation.
sequences. Clearly, if R is stable, the II1 factor L(R) is McDuff.
Again, both the McDuff property and stability are closely related to inner amenability. If G is
a countable group and G y (X, µ) is a free ergodic pmp action (X, µ), we consider the group
measure space II1 factor M = L∞(X) ⋊ G and the orbit equivalence relation R = R(G y X).
By [JS85], if R is stable, the group G is inner amenable. Actually, already if M is McDuff, G
must be inner amenable (see Proposition 4.1 below). We say that a group G is stable, resp.
McDuff, if G admits a free ergodic pmp action G y (X, µ) such that the orbit equivalence
relation R(G y X) is stable, resp. the crossed product L∞(X) ⋊ G is McDuff. Here and in
what follows all probability spaces are assumed to be standard.
We thus have the following subtly related notions for a countable group G : inner amenability,
property Gamma, the McDuff property and stability. Figure 1 summarizes all implications and
1KU Leuven, Department of Mathematics, Leuven (Belgium), [email protected]
Supported by a PhD fellowship of the Research Foundation Flanders (FWO).
2KU Leuven, Department of Mathematics, Leuven (Belgium), [email protected]
Supported in part by European Research Council Consolidator Grant 614195, and by long term structural
funding – Methusalem grant of the Flemish Government.
1
non implications between these properties. The main contribution of this article is to fill in
all the arrows that were unknown so far. So, we prove that Kida's counterexamples of [K12a]
provide icc groups with property Gamma that are not McDuff. We also adapt the example
of [PV08] of a (necessarily non icc) property (T), but yet McDuff group G to obtain an icc
McDuff group that is not stable.
G has property Gamma
[E73]
[V09]
G is inner amenable
[K12b]
Thm 2.1
[K12a]
[K12b]
Thm 2.1
[K12a]
[JS85]
Prop 4.1
G is McDuff
G is stable
Thm 3.2
Figure 1: Implications and non implications for a countable discrete icc group G. The dashed
arrows are proven in this article.
For later use, recall from [J99, Theorem 1.2] that a subgroup H of a countable group G has
relative property (T) if and only if for every unitary representation π : G → U (K) and every
bounded sequence of vectors ξn ∈ K satisfying limn kπ(g)ξn − ξnk = 0 for all g ∈ G, we have
that limn kPH (ξn) − ξnk = 0, where PH : K → KH denotes the orthogonal projection of K onto
the closed subspace of π(H)-invariant vectors.
In the context of tracial von Neumann algebras, this characterization of relative property (T)
gives the following well known lemma. For completeness, we include a proof.
Lemma 1.1. Let (M, τ ) be a tracial von Neumann algebra with von Neumann subalgebra
N ⊂ M . Let G be a countable group and H < G a subgroup with the relative property (T).
Let π : G → U (M ) be a homomorphism satisfying π(g)N π(g)∗ = N for all g ∈ G. Define
B = π(H)′ ∩ N = {x ∈ N π(h)x = xπ(h) ∀h ∈ H}. Denote by EB : N → B the unique trace
preserving conditional expectation.
If (an) is a bounded sequence in N satisfying limn kπ(g)an − anπ(g)k2 = 0 for all g ∈ G, then
limn kan − EB(an)k2 = 0.
Proof. It suffices to observe that (Ad π(g))g∈G defines a unitary representation of G on L2(N )
and that EB is the orthogonal projection of L2(N ) onto the subspace of H-invariant vectors.
2 An icc group with property Gamma that is not McDuff
In [K12a], Kida constructed the first icc groups G that have property Gamma (and in particular,
are inner amenable) but that are not stable, meaning that they do not admit a stable free
2
ergodic pmp action. We prove here that these groups are not even McDuff, meaning that they
do not admit a free ergodic pmp action with the crossed product being McDuff.
Fix a property (T) group H that admits a central element a ∈ Z(H) of infinite order. As
in [C05], one could take n ≥ 3, define H ⊂ SL(n + 2, Z) given by
H =
b
1
c
0 A d
0
1
0
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
A ∈ SL(n, Z) , b ∈ Z1×n ,
c ∈ Z , d ∈ Zn×1
and take a =
0
1
1
0 In 0
0
1
0
.
Given two non zero integers p, q with p 6= q, we define G as the HNN extension
G = hH, t t−1apt = aqi .
(2.1)
Theorem 2.1. Kida's group G defined in (2.1) is icc, has property Gamma, but is not McDuff.
Proof. Analyzing reduced words in HNN extensions, it is easy to check that G is icc and that
H < G is a relatively icc subgroup, meaning that {hgh−1 h ∈ H} is infinite for every g ∈ G\H
(see e.g. [S05]).
Ozawa already proved that L(G) has property Gamma and his proof can be found in [K12a,
Theorem 1.4]. For completeness, we include the following slightly easier proof. We claim that
there exists a sequence of Borel functions fn : R → T satisfying fn(x + 1) = fn(x) for all x ∈ R,
n Z 1
lim
0
fn(x) dx = 0 and
n Z 1
lim
0
fn(px) − fn(qx)2 dx = 0 .
(2.2)
Denote by (ug)g∈G the canonical unitary elements of L(G). Given such a sequence of functions
0 fn(x)e−2πikx dx the Fourier coefficients of fn, we can define the
fn and denoting by cfn(k) =R 1
unitary elements un ∈ L(H) given by
un =Xk∈Zcfn(k) upk
a .
Then (2.2) says that τ (un) → 0 and ku∗
t unut − unk2 → 0. Since moreover un ∈ Z(L(H)), it
follows that (un) is a non trivial central sequence in L(G), so that L(G) has property Gamma.
To prove the existence of fn satisfying (2.2), define Λ = Z[p−1, q−1]⋊Z as the semidirect product
of the abelian group Z[p−1, q−1] with Z acting by the automorphisms given by multiplication
with a power of p/q. Define the action Λ yα R given by (x, k) · y = x + (p/q)ky and equip R
with the Lebesgue measure class. Since Λ is an amenable group, it follows from [CFW81] that
the orbit equivalence relation R(Λ y R) is hyperfinite and thus, not strongly ergodic. This
provides us with a sequence of unitary elements fn ∈ L∞(R) tending to 0 weakly and being
approximately Λ-invariant in the sense that α(x,k)(fn) − fn → 0 strongly for all (x, k) ∈ Λ.
Restricting fn to the interval [0, 1) and then extending fn periodically, we find a sequence
satisfying (2.2).
Choose a free ergodic pmp action G y (X, µ). We prove that the II1 factor M = L∞(X) ⋊ G
is not McDuff. Let un ∈ U (M ) be a central sequence. Put B = L(H)′ ∩ M . Since H
has property (T), by Lemma 1.1, we have that kun − EB(un)k2 → 0, where EB : M →
B is the unique trace preserving conditional expectation. Since H < G is relatively icc, it
follows that B ⊂ B1, where B1 = L∞(X) ⋊ H. We also define B2 = L∞(X) ⋊ tHt−1. Since
kun − EB1 (un)k2 → 0 and kun − u∗
t unutk2 → 0, we get that kun − EB2 (un)k2 → 0. Since
H ∩ tHt−1 ⊂ Z(H), we conclude that kun − ED(un)k2 → 0, where D = L∞(X) ⋊ Z(H).
3
Define C = L∞(X)H ⋊Z(H), where L∞(X)H denotes the von Neumann algebra of H-invariant
functions. We claim that the subalgebras B, D ⊂ M form a commuting square with B ∩D = C.
with ag ∈ L∞(X) for all g ∈ G be the Fourier decomposition of a. Since a commutes with
To prove this claim, it suffices to take a ∈ B and prove that ED(a) ∈ C. Let a = Pg∈G agug
L(H), we get in particular that ah ∈ L∞(X)H for all h ∈ Z(H). Since ED(a) =Ph∈Z(H) ahuh,
the claim follows. Since kun − EB(un)k2 → 0 and kun − ED(un)k2 → 0, we conclude that
kun − EC(un)k2 → 0. Because C is abelian, M is not McDuff.
3 An icc McDuff group that is not stable
There are two obvious obstructions for stability of a group G : non inner amenability and
property (T). In [PV08, Theorem 6.4.2] using [E10], an example of a McDuff property (T)
group G is constructed. In particular, G is an example of a McDuff group that is not stable.
However by [PV08, Theorem 6.4.1], a McDuff property (T) group can never be icc.
In this section, we construct examples of icc McDuff groups that are not stable. We do this by
giving the following obstruction for stability and by combining the methods of [PV08] and [V09].
Proposition 3.1. If a countable group G admits a subgroup H < G with the relative prop-
erty (T) and the relative icc property, then G is not stable.
Proof. Let G y (X, µ) be a free ergodic pmp action and denote by R = R(G y X) the
associated orbit equivalence relation. Write A = L∞(X) and M = A ⋊ G. We denote by
NM (A) = {u ∈ U (M ) uAu∗ = A} the normalizer of A inside M . To prove that R is not
stable, by [JS85, Theorem 3.4], we have to show that all central sequences (un) and (pn) in M
with un ∈ NM (A) and pn a projection in A for all n satisfy kunpn − pnunk2 → 0.
Put B = L(H)′ ∩ M and also define D = AH as the algebra of H-invariant functions in A.
Denote by EB : M → B and ED : A → D the unique trace preserving conditional expectations.
Since H < G has the relative property (T), by Lemma 1.1, we get that kun − EB(un)k2 → 0
and kpn − ED(pn)k2 → 0. Since H < G is relatively icc, we have B ⊂ A ⋊ H. Therefore, B
and D commute. It follows that kunpn − pnunk2 → 0.
For every countable group H, we denote by Hf its FC-radical, defined as the normal subgroup
of H consisting of all elements that have a finite conjugacy class. Also recall that a group is
said to be virtually abelian if it admits an abelian subgroup of finite index. We denote by
CH(K) the centralizer in H of a subset K ⊂ H.
To construct an icc McDuff group that is not stable, we start with the property (T) group
of [PV08, Theorem 6.4] : a residually finite property (T) group H such that the FC-radical Hf
is not virtually abelian. In [E10], it was proved that such a group indeed exists. From [PV08,
Theorem 6.4], we know that H is McDuff. We then perform the same iterative amalgamated
free product construction as in [V09] in order to embed H as a relatively icc subgroup of a
larger group G in such a way that G remains McDuff.
Fix a decreasing sequence of finite index normal subgroups Ln ⊳ H withTn Ln = {e}. Enumer-
ate Hf = {hn n ∈ N}. Note that for every n ∈ N, CH(h1, . . . , hn) is a finite index subgroup
of H. Then,
Hn := Ln ∩
\g∈H
g CH(h1, . . . , hn) g−1
4
is a decreasing sequence of finite index normal subgroups of H with Tn Hn = {e}. Put Kn =
Hf ∩ Hn.
Define inductively H = G0 ⊂ G1 ⊂ G2 ⊂ · · · as the amalgamated free product Gn+1 =
Gn ∗Kn (Kn × Z), where we view Kn ⊂ H = G0 ⊂ Gn. Define G as the direct limit of
G0 ⊂ G1 ⊂ · · · .
Theorem 3.2. The group G constructed above is icc, has the McDuff property, but is not
stable.
Proof. We first prove that H < G is relatively icc. Take g ∈ G \ H. Take n ≥ 0 such that
g ∈ Gn+1 \ Gn. We have Gn+1 = Gn ∗Kn (Kn × Z) and Kn < H < Gn. Since Hf is infinite
amenable and H has property (T), Hf < H has infinite index. So Kn has infinite index in H
and it follows that {hgh−1 h ∈ H} is an infinite set. To prove that G is icc, it suffices to prove
that every h ∈ H with h 6= e has an infinite conjugacy class. Take n ≥ 0 such that h ∈ H \ Kn.
Let t be the generator of Z in the description Gn+1 = Gn ∗Kn (Kn × Z). Since h ∈ Gn \ Kn,
the elements tkht−k, k ∈ Z, are all distinct.
Since H has property (T), it follows from Proposition 3.1 that G is not stable. It remains to
construct a free ergodic pmp action G y (X, µ) such that L∞(X) ⋊ G is McDuff.
For every n ≥ 0, we have Gn+1 = Gn ∗Kn (Kn × Z) and we define the homomorphism πn,n+1 :
Gn+1 → Gn given by the identity map on Gn and by mapping Z to {e}. Writing πn,m = πn,n+1◦
· · · ◦ πm−1,m for all n ≤ m, we obtain a compatible system of homomorphisms πn,m : Gm → Gn
that combine into a homomorphism πn : G → Gn. In particular, π0 : G → H. By induction
on n, one checks that Km is a normal subgroup of Gn+1 for all n ≥ 0 and all m ≥ n, and that
gπ0(g)−1 commutes with Kn for every g ∈ Gn+1.
Denote by H yβ (Y, ν) the profinite action given as the inverse limit of H y H/Hn. Define
G yα Y by α(g) = β(π0(g)). Note that α is an ergodic pmp action. For every n ≥ 1, we
consider the Bernoulli action βn of Gn/Kn−1 on Xn = [0, 1]Gn/Kn−1 equipped with the product
measure µn = λGn/Kn−1. We define G yαn Xn by αn(g) = βn(πn(g)). Note that αn is a
weakly mixing pmp action. We define (X, µ) as the product of (Y, ν) and all (Xn, µn), n ≥ 1.
We define G y X as the diagonal action. Since each αn is weakly mixing and α is ergodic, the
diagonal action G y (X, µ) is ergodic. When g ∈ G \ {e}, take n ≥ 1 such that g ∈ Gn \ Kn−1.
Since the Bernoulli action βn is essentially free, the set {x ∈ Xn αn(g)(x) = x} has measure
zero and thus also {x ∈ X g ·x = x} has measure zero. We conclude that G y X is essentially
free.
Denote M = L∞(X) ⋊ G and let τ : M → C be the tracial state on M . To prove that M
is McDuff, we use the method of [PV08, Theorem 6.4]. Denote by θ : L∞(Y ) → L∞(X) and
θn : L∞(Xn) → L∞(X) the natural ∗-homomorphisms induced by the coordinate maps X → Y
and X → Xn. The definition of Hk implies that for every h ∈ Hf , we have Hk ⊂ CH(h) for all
k large enough. So for every h ∈ Hf , we can define the element
vh = Xs∈H/Hk
θ(1sHk ) ush−1s−1
whenever k is such that Hk ⊂ CH(h). A direct computation shows that τ (vh) = 0 for every
h ∈ Hf \ {e} and that h 7→ vh is a homomorphism from Hf to U (L(H)′ ∩ M ).
When g ∈ Gn+1, we have that gπ0(g)−1 commutes with Kn. It follows that vh commutes with
L(Gn+1) whenever h ∈ Kn. When h ∈ Kn, we also get that vh commutes with θ(1rHm) for
all r ∈ H and all m ≤ n and that vh commutes with θm(L∞(Xm)) for all m ≤ n + 1. So,
5
choosing any sequence hn ∈ Kn \ {e}, we get that (vhn)n is a central sequence of unitaries in
M with τ (vhn) = 0 for all n. Since Kn is a finite index subgroup of Hf , we know that Kn is
n 6= h′
n ∈ Kn such that hnh′
non abelian. We can thus choose sequences hn, h′
nhn for all n. The
hnv∗
) = 0 for all n ≥ 0. So,
corresponding central sequences (vhn) and (vh′
h′
n
M is a McDuff II1 factor.
n) satisfy τ (vhnvh′
nv∗
4 All McDuff groups are inner amenable
For icc groups G, the following result is a consequence of the main theorem in [C81]. In the
non icc case, we need a small extra argument, because we follow the convention that a group
G is inner amenable if there exists a conjugation invariant mean on G that gives weight zero
to all finite subsets of G and not only to the subset {e}.
Proposition 4.1. Every McDuff group G is inner amenable.
Proof. Let G be a group that is not inner amenable. Let G yα (X, µ) be any free ergodic pmp
action. Denote A = L∞(X) and M = A ⋊ G. Let xn ∈ U (M ) be an arbitrary central sequence.
We prove that limn kxn − EA(xn)k2 = 0.
xn pk2 = 0 for all p ∈ A, we have limn kp an(g) − an(g) αg(p)k2 = 0 for all p ∈ A. Let p1 ∈ A be
a maximal projection satisfying p1αg(p1) = 0. Put p2 = αg(p1) and p3 = 1 − (p1 + p2). Since
Write xn =Pg∈G an(g) ug with an(g) ∈ A for every g ∈ G. Fix g ∈ G \ {e}. Since limn kp xn −
the action α is essentially free, we get that Pi pi = 1 and pi αg(pi) = 0 for all i. It follows that
limn kpi an(g)k2 = 0 for every i, so that limn kan(g)k2 = 0 for all g ∈ G \ {e}.
Define the unit vectors ξn ∈ ℓ2(G) given by ξn(g) = kan(g)k2. In the previous paragraph, we
proved that limn ξn(g) = 0 for every g ∈ G \ {e}. As in [C81], the vectors ξn are approximately
conjugation invariant. Since G is not inner amenable, the functions ξn must satisfy limn ξn(e) =
1, meaning that limn kxn − EA(xn)k2 = 0.
So, we have proved that every central sequence xn ∈ U (M ) satisfies limn kxn − EA(xn)k2 = 0.
It follows that M ′ ∩ M ω is abelian, so that M is not McDuff.
References
[C81]
[CFW81] A. Connes, J. Feldman and B. Weiss, An amenable equivalence relation is generated by a
M. Choda, Inner amenability and fullness. Proc. Amer. Math. Soc. 86 (1982), 663-666.
[C05]
[E73]
[E10]
[J99]
[JS85]
[K12a]
single transformation. Ergodic Theory Dynam. Systems 1 (1981), 431-450.
Y. de Cornulier, Finitely presentable, non-Hopfian groups with Kazhdan's property (T) and
infinite outer automorphism group. Proc. Amer. Math. Soc. 135 (2007), 951-959.
E.G. Effros, Property Γ and inner amenability. Proc. Amer. Math. Soc. 47 (1975), 483-486.
M. Ershov, Kazhdan groups whose FC-radical is not virtually abelian. Preprint. (2010), avail-
able at http://people.virginia.edu/~mve2x/Research/Kazhdan_infFC0706.pdf.
P. Jolissaint, On property (T) for pairs of topological groups. Enseign. Math. 51 (2005),
31-45.
V.F.R. Jones and K. Schmidt, Asymptotically invariant sequences and approximate finiteness.
Amer. J. Math. 109 (1987), 91-114.
Y. Kida, Inner amenable groups having no stable action. Geom. Dedicata 173 (2014), 185-192.
6
[K12b]
Y. Kida, Stability in orbit equivalence for Baumslag-Solitar groups and Vaes groups. Groups
Geom. Dyn. 9 (2015), 203-235.
[McD69] D. McDuff, Central sequences and the hyperfinite factor. Proc. London Math. Soc. (3) 21
(1970), 443-461.
[MvN43] F.J. Murray and J. von Neumann, Rings of operators IV, Ann. Math. 44 (1943), 716-808.
[PV08]
S. Popa and S. Vaes, On the fundamental group of II1 factors and equivalence relations
arising from group actions. In Quanta of Maths, Proceedings of the Conference in honor of
A. Connes' 60th birthday, Clay Mathematics Institute Proceedings 11 (2011), pp. 519-541.
Y. Stalder, Moyennabilit´e int´erieure et extensions HNN. Ann. Inst. Fourier (Grenoble) 56
(2006), 309-323.
S. Vaes, An inner amenable group whose von Neumann algebra does not have property
Gamma. Acta Math. 208 (2012), 389-394.
[S05]
[V09]
7
|
1101.4803 | 2 | 1101 | 2011-10-11T12:34:32 | Lifting defects for nonstable K_0-theory of exchange rings and C*-algebras | [
"math.OA",
"math.CT",
"math.RA"
] | The assignment (nonstable K_0-theory), that to a ring R associates the monoid V(R) of Murray-von Neumann equivalence classes of idempotent infinite matrices with only finitely nonzero entries over R, extends naturally to a functor. We prove the following lifting properties of that functor: (1) There is no functor F, from simplicial monoids with order-unit with normalized positive homomorphisms to exchange rings, such that VF is equivalent to the identity. (2) There is no functor F, from simplicial monoids with order-unit with normalized positive embeddings to C*-algebras of real rank 0 (resp., von Neumann regular rings), such that VF is equivalent to the identity. (3) There is a {0,1}^3-indexed commutative diagram D of simplicial monoids that can be lifted, with respect to the functor V, by exchange rings and by C*-algebras of real rank 1, but not by semiprimitive exchange rings, thus neither by regular rings nor by C*-algebras of real rank 0. By using categorical tools from an earlier paper (larders, lifters, CLL), we deduce that there exists a unital exchange ring of cardinality aleph three (resp., an aleph three-separable unital C*-algebra of real rank 1) R, with stable rank 1 and index of nilpotence 2, such that V(R) is the positive cone of a dimension group and V(R) is not isomorphic to V(B) for any ring B which is either a C*-algebra of real rank 0 or a regular ring. | math.OA | math |
LIFTING DEFECTS FOR NONSTABLE K0-THEORY OF
EXCHANGE RINGS AND C*-ALGEBRAS
FRIEDRICH WEHRUNG
Abstract. The assignment (nonstable K0-theory), that to a ring R associates
the monoid V(R) of Murray-von Neumann equivalence classes of idempotent
infinite matrices with only finitely nonzero entries over R, extends naturally
to a functor. We prove the following lifting properties of that functor:
(i) There is no functor Γ, from simplicial monoids with order-unit with nor-
malized positive homomorphisms to exchange rings, such that V ◦ Γ ∼= id.
(ii) There is no functor Γ, from simplicial monoids with order-unit with
normalized positive embeddings to C*-algebras of real rank 0 (resp., von
Neumann regular rings), such that V ◦ Γ ∼= id.
(iii) There is a {0, 1}3-indexed commutative diagram ~D of simplicial monoids
that can be lifted, with respect to the functor V, by exchange rings and
by C*-algebras of real rank 1, but not by semiprimitive exchange rings,
thus neither by regular rings nor by C*-algebras of real rank 0.
By using categorical tools (larders, lifters, CLL) from a recent book from
the author with P. Gillibert, we deduce that there exists a unital exchange ring
of cardinality ℵ3 (resp., an ℵ3-separable unital C*-algebra of real rank 1) R,
with stable rank 1 and index of nilpotence 2, such that V(R) is the positive
cone of a dimension group but it is not isomorphic to V(B) for any ring B
which is either a C*-algebra of real rank 0 or a regular ring.
1. Introduction
While attending the August 2010 "New Trends in Noncommutative Algebra"
conference in Seattle, the author of the present paper was told about the question
whether the K0 functor from AF C*-algebras to dimension groups splits, that is,
has a (categorical) right inverse. We describe here three reasons why this is not
the case. The first and the third reason are more conveniently expressed via the
nonstable K0-theory (often also called "nonstable K-theory") of a ring R, mainly
described by a commutative monoid, usually denoted by V(R).
If R is unital,
then V(R) is the monoid of all isomorphism types of finitely generated projective
right R-modules and K0(R) is the Grothendieck group of V(R).
Our first reason, Corollary 6.6, implies that the more general question, whether
the functor V has a right inverse from simplicial monoids (cf. Section 3) to exchange
rings (cf. Warfield [30], and Ara [1] for the non-unital case), has a negative answer.
Date: August 24, 2018.
2000 Mathematics Subject Classification. 19A49, 46L80, 16B50, 16B70, 16E20, 16E50, 16N60,
18A30, 18C35, 06B20, 08B20, 03E05.
Key words and phrases. Ring; exchange property; regular; C*-algebra; real rank; stable rank;
index of nilpotence; semiprimitive; V-semiprimitive; weakly V-semiprimitive; simplicial monoid;
dimension group; commutative monoid; order-unit; o-ideal; refinement property; nonstable; K-
theory; idempotent; orthogonal; projection; functor; diagram; lifting; premeasure; measure; larder;
lifter; condensate; CLL.
1
2
F. WEHRUNG
Actually, Corollary 6.6 is stated for an even more general class of rings that we call
weakly V-semiprimitive rings. It also suggests that the original question might get
more interesting if one required the morphisms between dimension groups be one-
to-one. Our second and third reason will address the modified K0-splitting
question.
Our second reason, Proposition 7.1, involves the author's construction, intro-
duced in Wehrung [31], of a dimension group with order-unit, of cardinality ℵ2,
which is not isomorphic to V(R) for any (von Neumann) regular ring R. If K0 had
a right inverse from simplicial groups (cf. Section 3) to AF C*-algebras, then it
would also have a right inverse from simplicial groups to regular rings. However,
due to the abovementioned counterexample, this is impossible. There are some
complications, due to the requirement that the maps between simplicial groups be
one-to-one, that we solve by using tools from lattice theory and universal algebra.
Our third reason, Theorem 10.1, is the combinatorial core of our second reason.
It involves a commutative diagram ~D of simplicial monoids, indexed by a cube
(i.e., the powerset lattice {0, 1}3 of a three-element set), that has no lifting, with
respect to the functor V, by semiprimitive exchange rings. In particular, it can be
lifted neither by C*-algebras of real rank 0 nor by regular rings. In particular, the
negative result of Proposition 7.1 extends both to regular rings (this could have
been deduced directly from [31]) and to C*-algebras of real rank 0 (this is new).
Lifting the diagrams ~D and ~L. Besides those three reasons, we also provide
two positive representation results of the abovementioned diagram ~D: while The-
orem 10.1 implies that ~D cannot be lifted, with respect to the functor V, by any
commutative diagram of semiprimitive exchange rings, Theorem 11.2 implies that
a certain "collapsed" image, denoted by ~L, of the diagram ~D, can be lifted with
exchange rings, while Theorem 12.5 implies that ~L can be lifted with C*-algebras
(of real rank and stable rank both equal to 1). In order to be able to state that ~L is
really a "diagram", we need the extended notion of diagram given in Definition 2.1.
By "unfolding" ~L back to ~D, we obtain (Proposition 11.1) that ~D can also be lifted
with exchange rings on the one hand, and with C*-algebras on the other hand.
There is no "best of two worlds", because a C*-algebra is an exchange ring iff it
has real rank 0, and ~D cannot be lifted by C*-algebras of real rank 0.
The journey back to the transfinite: separating the nonstable K0-theory
of exchange rings and the one of C*-algebras from the ones of regular
rings and C*-algebras of real rank 0. The reason why we need a diagram
indexed by a finite lattice lies in a quite technical result of categorical algebra by
Gillibert and the author [13] called the Condensate Lifting Lemma, CLL in short.
CLL makes it possible to turn diagram counterexamples to object counterexamples,
with possible jumps of cardinality. It works only on diagrams indexed by so-called
almost join-semilattices, in particular for finite lattices.
By applying various instances of CLL to the negative lifting versus positive lifting
results of the cube ~D, we obtain (cf. Theorems 13.6 and 14.1) that there exists a
unital exchange ring (resp., a unital C*-algebra of real rank 1) R, with stable rank 1
and index of nilpotence 2, such that V(R) is the positive cone of a dimension group
and V(R) is not isomorphic to V(B) for any ring B which is either a C*-algebra
of real rank 0 or a regular ring. Due to the cardinality-jumping properties of the
LIFTING DEFECTS IN NONSTABLE K0-THEORY
3
"condensate" construction underlying CLL, the exchange ring that we construct
has ℵ3 elements, while the C*-algebra that we construct is ℵ3-separable (it has also
real rank 1). Both the ring and the C*-algebra are stably finite, have stable rank 1
and index of nilpotence 2, while the nonstable K0-theory is the positive cone of a
dimension group with order-unit of index 2.
Although the present paper is written in the language of ring theory, a substantial
part of our results and proofs draw their basic inspiration from related results in
universal algebra and lattice theory. For example,
• Corollary 6.6 (the "first reason") originates in Tuma and Wehrung [29,
Section 8], which is a non-lifting result by universal algebras with respect
to the congruence lattice functor.
• Proposition 7.1 (the "second reason") originates in Wehrung [31], which
solves, in the negative, representation problems with respect to various
functors, including K0, by using the lattice structure of the set of all prin-
cipal right ideals in a regular ring.
• The cube ~D (cf. Figure 10.1), which is the key ingredient of Theorem 10.1
(the "third reason"), originates in a cube of Boolean semilattices intro-
duced in Tuma and Wehrung [29, Section 3].
Acknowledgment. The author is indebted to Pere Ara, Francesc Perera, Enrique
Pardo, George Elliott, Chris Phillips, and Ken Goodearl for stimulating interac-
tion and helpful comments at various stages of this paper's preparation. Special
thanks are due to Ken Goodearl for his improvement of the author's original argu-
ment, which led to Theorem 6.5, and to Chris Phillips for most inspiring conver-
sations about C*-algebras and noncommutative rings at the Seattle conference in
August 2010.
2. Organization of the paper
All our rings will be associative, but not necessarily unital. We denote by 1R,
or 1 if R is understood, the unit of a ring R.
Although we shall use standard commutative diagrams such as the one, denoted
by ~D, introduced in Figure 10.1, we shall also use sometimes diagrams of a more
general kind, a formal definition of which follows.
Definition 2.1. A quiver is a quadruple I = (V, E, s, t) where V and E (the
"vertices" and the "edges", respectively) are sets and s, t : E → V (the "source"
and "target" map, respectively). An element e ∈ E is thus viewed as an "arrow"
from s(e) to t(e).
A diagram in a category C, indexed by I, is a functor from I to C. Formally,
it is given by a pair (Φo, Φm), where Φo (resp., Φm) is a map from V (resp., E)
to the object class (resp., morphism class) of C, such that Φm(e) is a morphism
from Φo(s(e)) to Φo(t(e)) for any e ∈ E. There is no composition of arrows in I,
thus nothing to worry about at that level.
The usual definitions of a natural transformation and a natural equivalence carry
over to diagrams without modification. For diagrams Φ, Ψ : I → C, we denote by
Φ ∼= Ψ the natural equivalence of Φ and Ψ. For categories A and B, we say
that a diagram Γ : I → A lifts a diagram Ψ : I → B with respect to a functor
Φ : A → B if Ψ ∼= Φ Γ (cf. Figure 2.1). In all our lifting problems, we will need
to specify additional composition relations among the arrows in the range of Γ.
4
F. WEHRUNG
(The first such instance in this paper is Theorem 6.5: the additional requirement
is h ◦ s = h.)
I
Γ
$IIIIIIIIIII
Ψ
A
Φ
B
Figure 2.1. The diagram Γ lifts the diagram Ψ with respect to Φ
In many of our applications (but not all of them), A will be a category of rings, B
will be a category of commutative monoids, and Φ will be the nonstable K0-theory
functor V.
We shall now give a brief section by section overview of the paper.
In Section 3, we recall the basic monoid-theoretical concepts used in the paper.
In Section 4, we recall some basic facts about the functor V, from rings to com-
mutative monoids, describing nonstable K0-theory, for either rings or C*-algebras.
We also recall basic facts about semiprimitive rings.
In Section 5, we recall some basic facts about exchange rings, including the
non-unital case.
In Section 6, we introduce a notion of orthogonality of elements in a commu-
tative monoid, then the notion of a weakly V-semiprimitive ring (V-semiprimitive
rings will be introduced in Section 9). We prove that every exchange ring is weakly
V-semiprimitive. We also introduce a diagram (in the sense of Definition 2.1) ~K
of simplicial monoids with order-unit that has no lifting, with respect to the func-
tor V, by weakly V-semiprimitive rings (Theorem 6.5). It follows (Corollary 6.6)
that the functor V has no right inverse from simplicial monoids with order-unit,
with normalized monoid homomorphisms, to weakly V-semiprimitive rings.
In Section 7, we prove (cf. Proposition 7.1) that the functor V has no right in-
verse from simplicial monoids with unit, and their embeddings, to AF C*-algebras.
The proof uses universal algebra, lattice theory, and a counterexample of cardinal-
ity ℵ2 constructed in Wehrung [31]. As Proposition 7.1 will be later superseded by
stronger results, we give only an outline of the proof. Nevertheless we chose to keep
it there, because a large part of the material in further sections originates from the
proof of that result.
In Section 8 we introduce a tool that will prove essential for our work, the notion
of a (pre)measured ring. By using some results of Section 4 and 5, we prove (cf.
Lemma 8.8) that the category of measured exchange rings is a reflective subcategory
of the one of premeasured exchange rings. We also prove (cf. Lemma 8.9) an
analogous result for C*-algebras of real rank 0.
In Section 9, we introduce a common generalization of semiprimitive exchange
rings, principal ideal domains, and polynomial rings over fields, that we call V-
semiprimitive rings. We also observe there that not every exchange ring, and not
every C*-algebra, is V-semiprimitive. The C*-algebra D constructed in Example 9.4
will play a crucial role in further sections.
In Section 10, we introduce the main combinatorial object of the paper, namely
the diagram ~D. It is a {0, 1}3-indexed commutative diagram of simplicial monoids,
/
/
$
LIFTING DEFECTS IN NONSTABLE K0-THEORY
5
and it originates in the proof of Proposition 7.1. Theorem 10.1 implies that ~D
has no lifting, with respect to the functor V, by any commutative diagram of V-
semiprimitive rings. This is the central non-representability result of the paper: all
the other negative lifting results of the paper, with the exception of Theorem 6.5,
follow from that one. The proof of Theorem 10.1 is direct:
it does not involve
any universal algebra, lattice theory, or transfinite cardinals. We also introduce
an auxiliary diagram, denoted there by ~E, of simplicial monoids and monoid em-
beddings, such that any lifting of ~E, with respect to the functor Π introduced in
Definition 8.6, gives rise to a lifting of ~D. This implies, in particular, that the
functor V has no right inverse from simplicial monoids with order-unit, and their
embeddings, to C*-algebras of real rank 0 (resp., regular rings), see Corollary 10.8.
In Section 11, we take advantage of the many symmetries underlying the di-
agram ~D to collapse it to another diagram (now in the sense of Definition 2.1),
denoted by ~L, much simpler looking than ~D, such as every lifting of ~L gives rise
to a lifting of ~D (this is shown in the argument of the proof of Proposition 11.1).
We also prove, by a direct construction, that ~L (thus also ~D) can be lifted, with
respect to the functor V, by exchange rings. In a sense, everything in that section
boils down to finding a matrix c (given in (11.4)) satisfying the constraints given
in (11.5) and (11.6).
In Section 12, we construct a lifting of ~L by C*-algebras. These C*-algebras
have both real rank and stable rank equal to 1, while they have index of nilpotence 2.
By using Theorem 10.1, it can be seen that these values are optimal.
In Section 13, we take advantage of the conflict between Theorem 10.1 (a non-
lifting result of ~D) and Proposition 11.3 (a lifting result of ~D by exchange rings)
to construct, for every field K, a unital exchange K-algebra RK such that V(RK),
although being the positive cone of a dimension group, is never isomorphic to V(B)
for either a C*-algebra of real rank 0 or a regular ring B. Furthermore, RK has at
most ℵ3 + card K elements, and it has index of nilpotence 2. In order to achieve
this, we are using a fair amount of heavy machinery established earlier in the book
Gillibert and Wehrung [13]. That work introduces tools of categorical algebra and
infinite combinatorics, called larders and lifters, that are designed to turn diagram
counterexamples to object counterexamples. These tools are effective mainly on
lattice-indexed diagrams, which is the reason why we need to keep the cube ~D
instead of the better looking ~L. The results of [13] being now used as a toolbox, it
should be possible to read the proofs of Section 13 without needing to ingest large
amounts of larder technology. In particular, the ring-theoretical and C*-algebraic
arguments used there are elementary.
In Section 14, we show how to extend to C*-algebras what we did for exchange
rings in Section 13. In particular, we construct a unital C*-algebra E, of real rank
and stable rank both equal to 1, such that V(E), although being the positive cone
of a dimension group, is never isomorphic to V(B) for either a C*-algebra of real
rank 0 or a regular ring B. Furthermore, E is ℵ3-separable, and it has index of
nilpotence 2.
In Section 15, we list some open problems.
We represent on Figure 2.2 the implications between the main attributes of rings
used in the paper.
6
F. WEHRUNG
exchange
V-semiprimitive
semiprimitive
Weakly V-semiprimitive
9lllllllllllll
lllllllllllll
m RRRRRRRRRRRRR
RRRRRRRRRRRRR
semiprimitive exchange
C*-algebra
8iiiiiiiiiiiiiiiii
iiiiiiiiiiiiiiiii
n UUUUUUUUUUUUUUUUU
UUUUUUUUUUUUUUUUU
regular
C*-algebra of real rank 0
Figure 2.2. Attributes of rings
3. Basic notions about commutative monoids, refinement monoids,
dimension groups
We refer to Goodearl [14] for partially ordered abelian groups, interpolation
groups, dimension groups. We set Z+ := {0, 1, 2, . . . } and N := Z+ \ {0}.
We shall write all our commutative monoids additively. A submonoid I of a
commutative monoid M is an o-ideal of M if x + y ∈ I implies that x ∈ I and
y ∈ I, for all x, y ∈ I. We say that M is conical
if {0} is an o-ideal. Every
commutative monoid M can be endowed with its algebraic preordering ≤, defined
by x ≤ y if there exists z ∈ M such that y = x + z. We say that an element e ∈ M
is an order-unit of M if for each x ∈ M there exists n ∈ N such that x ≤ ne. For
n ∈ Z+, an element a ∈ M has index at most n if (n + 1)x ≤ a implies that x = 0,
for each x ∈ M .
If I is an o-ideal of a commutative monoid M , the binary relation ≡I on M
defined by
a ≡I b ⇐⇒ (∃x, y ∈ I)(a + x = b + y) ,
for all a, b ∈ M ,
(3.1)
is a monoid congruence of M , and M/I := M/≡I is a conical commutative monoid.
The Grothendieck group of M is the initial object in the category of all monoid
homomorphisms from M to a group.
It consists of an abelian group G with a
monoid homomorphism ε : M → G, and we shall always endow it with the unique
compatible preordering with positive cone ε(M ) (i.e., x ≤ y iff y − x ∈ ε(M )).
A commutative monoid M has the refinement property, or is a refinement monoid
(cf. Dobbertin [9]), if for all a0, a1, b0, b1 ∈ M such that a0 + a1 = b0 + b1, there are
cj,k ∈ M (for j, k ∈ {0, 1}) such that aj = cj,0 + cj,1 and bj = c0,j + c1,j for each
j ∈ {0, 1}. A very special class of refinement monoids consists of the positive cones
of dimension groups. A partially ordered abelian group G is a dimension group if it
is directed (i.e., G = G++(−G+)), unperforated (i.e., mx ≥ 0 implies x ≥ 0 for each
m ∈ N and each x ∈ G) and the positive cone G+ satisfies refinement. A simplicial
group is a group of the form Zn, for a natural number n, ordered componentwise.
Every simplicial group is a dimension group. Conversely, every dimension group
is a direct limit of simplicial groups (this follows from Effros, Handelman, and
Shen [10], but the form of this result established earlier in Grillet [18] is easily seen
1
K
S
e
0
K
S
K
S
K
S
f
K
S
LIFTING DEFECTS IN NONSTABLE K0-THEORY
7
to be equivalent). We shall also call a simplicial monoid the positive cone of a
simplicial group (i.e., (Z+)n for some n ∈ Z+).
Definition 3.1. For a subcategory C of the category CMon of conical commuta-
tive monoids with monoid homomorphisms, we shall denote by C(1) the category
of monoids with order-unit in C: the objects of C(1) are the pairs (M, a), where M
is an object of C and a is an order-unit of M , and a morphism from (M, a) to (N, b)
is a monoid homomorphism f : M → N such that f (a) = b (i.e., f is normalized ).
4. Basic notions about rings: nonstable K0-theory, semiprimitivity
Definition 4.1. Not all of our rings will be unital, and consequently not all our ring
homomorphisms will preserve the unit. For a subcategory R of the category Ring
of rings and ring homomorphisms, we shall denote by R(1) the subcategory of R
consisting of all unital rings in R, with unital ring homomorphisms, possibly with
any additional structure present in R.
For example, we shall denote by RR0 the category of all C*-algebras of real
rank 0 with C*-homomorphisms, and by RR0(1) the category of all unital C*-
algebras of real rank 0 with unital C*-homomorphisms. A useful subcategory
of RR0 is the full subcategory AF of all the approximately finite, or AF, C*-
algebras. By definition, a C*-algebra is AF if it is a (possibly uncountable) C*-direct
limit of finite-dimensional C*-algebras (not all our AF algebras will be separable).
For basic concepts about (stable or nonstable) K-theory of rings and C*-algebras,
we refer to Goodearl [17, Section 4] for the unital case, Ara [1, Section 3] for the
general case, Blackadar [6, Chapter 3] for the case of C*-algebras. For basic notions
about C*-algebras we refer to Murphy [23].
For a ring R (associative but not necessarily unital), we shall often identify
the ring Mn(R) of all n × n matrices over R with its image, via the embedding
0
x 7→(cid:18)x 0
0(cid:19), in Mn+1(R). Hence the elements of the (non-unital) ring M∞(R) :=
Sn∈N Mn(R) can be identified with the countably infinite matrices with entries
from R and only finitely many nonzero entries.
We define In(R) as the set of all idempotent elements of Mn(R), for each n ∈
N ∪ {∞}. For a ring homomorphism f : R → S, we denote by In(f ) the map from
In(R) to In(S) defined by the rule
In(f )(cid:0)(aj,k)1≤j,k≤n(cid:1) :=(cid:0)f (aj,k)(cid:1)1≤j,k≤n ,
We shall sometimes write f (a) instead of In(f )(a), for a ∈ In(R).
for each (aj,k)1≤j,k≤n ∈ In(R) .
Idempotents a, b ∈ I∞(R) are orthogonal if ab = ba = 0.
The (Murray-von Neumann, algebraic) equivalence of idempotents a, b ∈ M∞(R)
is defined by a ∼ b iff there are x, y ∈ M∞(R) such that a = xy and b = yx.
Replacing x by axb and y by bya, we see that we can assume that x = axb and
y = bya. We denote by [a]R, or [a] if R is understood, the ∼-equivalence class of a.
For all a, b ∈ I∞(R) there exists b′ ∼ b such that ab′ = b′a = 0; it follows that
equivalence classes can be added, via the formula
[a] + [b] := [a + b] if ab = ba = 0 , for all a, b ∈ I∞(R) .
(4.1)
The set V(R) := {[a] a ∈ I∞(R)}, endowed with the addition defined in (4.1), is
a conical commutative monoid. If R is unital, then [1R] is an order-unit of V(R)
and V(R) is isomorphic to the monoid of all isomorphism types of finitely generated
8
F. WEHRUNG
projective right (resp., left) R-modules, with addition defined by [X]+[Y ] = [X ⊕Y ]
(where [X] now denotes the isomorphism class of X). We shall also set V1(R) :=
(V(R), [1R]).
The assignment R 7→ V(R) can be extended to a functor from Ring to CMon:
for a homomorphism f : R → S of rings (resp., of unital rings), there is a unique
monoid homomorphism V(f ) : V(R) → V(S) such that V(f )([a]R) = [I∞(f )(a)]S
for each a ∈ I∞(R). The functor V preserves all direct limits (i.e., in categori-
Rj) ∼=
cal language, directed colimits) and finite products: for instance, V(lim
V(Rj) (for direct limits) and V(R × S) ∼= V(R) × V(S). All these facts ex-
lim
−→j∈I
tend to the unital case, giving a functor V1 : Ring(1) → CMon(1), with V1(R) :=
(V(R), [1R]) for every unital ring R.
−→j∈I
If R is unital, K0(R) is defined as the (preordered) Grothendieck group of V(R).
For a two-sided ideal I of a ring R and the inclusion map f : I ֒→ R, the
map V(f ) embeds V(I) into V(R). We shall thus often identify V(I) with its
image, under V(f ), in V(R). It is an o-ideal of V(R) (cf. Section 3).
Lemma 4.2 (folklore). Let R be a ring, let c ∈ I∞(R), and let α, β ∈ V(R). If
[c] = α + β, then there are orthogonal idempotents a, b ∈ I∞(R) such that c = a + b,
[a] = α, and [b] = β.
Note. Observe, in particular, that if c ∈ R, then a, b ∈ R.
Proof. There are orthogonal idempotents u, v ∈ I∞(R) such that [u] = α and
[v] = β. As [u + v] = α + β = [c], there are x, y ∈ M∞(R) such that c = xy while
u + v = yx. The matrices a := xuy and b := xvy are as required.
(cid:3)
The following result is observed, for unital exchange rings R and E ⊆ R, in the
course of the proof of Pardo [25, Teorema 4.1.7]. It is contained, in full generality,
in Ara and Goodearl [2, Proposition 10.10].
Lemma 4.3. Let R be a ring, let E ⊆ I∞(R), and denote by I the two-sided ideal
of R generated by the entries of all the elements of E. Then V(I) is the o-ideal
of V(E) generated by {[x]R x ∈ E}.
A ring R has index of nilpotence at most n if xn+1 = 0 implies that xn = 0 for
all x ∈ R. It is proved in Yu [33, Corollary 4] that every unital exchange ring with
finite index of nilpotence has stable rank 1. The following easy result gives some
more information.
Lemma 4.4. Let n be a positive integer and let R be a unital ring with index of
nilpotence at most n. Then [1R] has index at most n in V(R). Furthermore, if V(R)
satisfies the refinement property, then it is the positive cone of a dimension group.
idempotents e0, . . . , en, a of R such that 1 = a + Pn
Proof. Let ξ ∈ V(R) such that (n + 1)ξ ≤ [1]. By Lemma 4.2, there are orthogonal
i=0 ej while [ej] = ξ for
each j ≤ n. For each j < n, let ϕj : ejR → ej+1R be an isomorphism of right
R-modules, and denote by ϕ the unique endomorphism of RR extending all the ϕj
such that ϕ(a) = ϕ(en) = 0. Setting x := ϕ(1), we obtain that ϕn+1 = 0, thus
xn+1 = 0, and thus, by assumption, xn = 0, so ϕn = 0, and so, as ϕn(e0R) = enR,
we get en = 0, therefore ξ = 0. This shows that [1] has index at most n in V(R).
Suppose now that V(R) is a refinement monoid. As [1] is an order-unit in that
monoid, it follows that every element of V(R) has finite index (cf. Wehrung [32,
LIFTING DEFECTS IN NONSTABLE K0-THEORY
9
Corollary 3.12]), and so V(R) is the positive cone of a dimension group (cf. Wehrung
[32, Proposition 3.13]).
(cid:3)
In case R is a C*-algebra, each Mn(R), for n ∈ N, is also a C*-algebra, and it
is often convenient to replace algebraic equivalence of idempotents by *-equivalence
of projections (a projection is a self-adjoint idempotent), namely,
a ∗∼ b
if
(∃x)(a = xx∗ and b = x∗x) ,
for all projections a, b .
This can be done, because every idempotent in a C*-algebra is equivalent to a
projection (and even generates the same right ideal, cf. Kaplansky [21, Theorem 26,
p. 34]) and two projections are equivalent iff they are *-equivalent (cf. Kaplansky
[21, Theorem 27, p. 35]).
Lemma 4.5. Let L be a left ideal in a C*-algebra A. Then every projection of L
is equivalent to some projection of L.
Proof. Let e be a projection of L. There exists a self-adjoint a ∈ L such that
ke − ak ≤ 1/3. By Rørdam, Larsen, and Laustsen [27, Lemma 2.2.3], the spectrum
of a is contained in [−1/3, 1/3] ∪ [2/3, 4/3]. Let f : R → R the unique continuous
function such that f (x) = 0 for each x ≤ 1/3, f (x) = 1/x for each x ≥ 2/3, and f is
affine on [1/3, 2/3]. Observe that f (x)x = 0 for x ∈ [−1/3, 1/3] and f (x)x = 1 for
x ∈ [2/3, 4/3], thus, using continuous function calculus, b := f (a)a is a projection.
It belongs to L as a ∈ L, and ka − bk ≤ 1/3. Now ke − bk ≤ 2/3, so e ∼ b.
(cid:3)
Semiprimitive rings. It is known (cf. Jacobson [20, Theorem 1], Herstein [19,
Theorem 1.2.3]) that the Jacobson radical J(R) of a ring (not necessarily unital) R
is the largest two-sided ideal J of R such that (∀x ∈ J)(∃s ∈ R)(x + s − sx = 0).
This concept is left-right symmetric. A ring R is semiprimitive (semi-simple in
Jacobson [20, Section 2] or Herstein [19, page 16], but we are using the seemingly
more common terminology) if J(R) = {0}. Not every exchange ring is semiprimitive
(consider the upper triangular matrices over a field, cf. Example 9.3).
The following is contained in Jacobson [20, Theorem 8].
Proposition 4.6. Every regular ring is semiprimitive.
The following is contained in Dixmier [8, Th´eor`eme 2.9.7].
Proposition 4.7. Every C*-algebra is semiprimitive.
5. Exchange rings
Exchange rings have been first defined in the unital case in Warfield [30], then
in the general case by Ara [1]. A ring R is an exchange ring if for all x ∈ R, there
are an idempotent e ∈ R and r, s ∈ R such that e = rx = x + s − sx. This condition
is left-right symmetric. Not every exchange ring is a two-sided ideal in a unital
exchange ring, see [1, Example 4, page 412]. A ring R is (von Neumann) regular
if for all x ∈ R, there exists y ∈ R such that xyx = x. Every regular ring is an
exchange ring, and the converse fails (cf. Example 9.3).
The following result is established in the unital case in Ara, Goodearl, O'Meara,
and Pardo [3, Theorem 7.2], and then extended to the non-unital case in Ara [1,
Theorem 3.8].
Proposition 5.1. A C*-algebra has real rank 0 iff it is an exchange ring.
10
F. WEHRUNG
The following is contained in Proposition 1.3 and Theorem 1.4 in Ara [1].
Lemma 5.2. The following statements hold, for any exchange ring R.
(i) eRe is an exchange ring, for any idempotent e ∈ R.
(ii) Mn(R) is an exchange ring, for any positive integer n.
It is well-known that Lemma 5.2 extends to C*-algebras of real rank 0 (cf. Corol-
lary 2.8 and Theorem 2.10 in Brown and Pedersen [7]).
The following result is proved in [1, Proposition 1.5].
Proposition 5.3. The monoid V(R) has refinement, for every exchange ring R.
The first part of the following result is proved in Ara [1, Theorem 2.2]. The
second part is proved, in the unital case, in Ara, Goodearl, O'Meara, and Pardo [3,
Proposition 1.4]. The proof can be trivially extended to the non-unital case, by
using Lemma 5.2.
Proposition 5.4. Let I be a two-sided ideal in a ring R. Then R is an exchange
ring iff I and R/I are both exchange rings and idempotents can be lifted modulo I.
Furthermore, if R is an exchange ring, then the canonical homomorphism V(R) →
V(R/I) is surjective, and it induces an isomorphism V(R)/ V(I) → V(R/I).
In the statement of Proposition 5.4, V(I) is identified with its image in V(R),
which is an o-ideal of V(R). The quotient V(R)/ V(I) is defined in (3.1) and the
comment following.
From now on we shall denote by Reg (resp., Exch) the category of all regular
rings (resp., exchange rings) with ring homomorphisms.
6. Weakly V-semiprimitive rings; an unliftable diagram
Definition 6.1. Elements α and β in a conical commutative monoid M are or-
thogonal, in notation α ⊥ β, if the following statement holds:
Equivalently, the o-ideals generated by α and β, respectively, meet in {0}.
(∀n ∈ N)(∀γ ∈ M )(cid:0)(γ ≤ nα and γ ≤ nβ) ⇒ γ = 0(cid:1) .
Lemma 6.2. Let R be an exchange ring, let a, b ∈ I∞(R) with [a]R ⊥ [b]R, and
denote by A and B the two-sided ideals of R generated by the entries of a and b,
respectively. Then A ∩ B is contained in J(R).
Proof. Set I := A ∩ B. It follows from Lemma 4.3 that for each e ∈ I∞(I), there
exists n ∈ N such that [e] ≤ n[a] and [e] ≤ n[b]. As [a] ⊥ [b], it follows that e = 0.
Hence V(I) = {0}. As I is an exchange ring without nonzero idempotents, it must
be contained in J(R).
(cid:3)
Definition 6.3. A ring R is weakly V-semiprimitive if for all a, b ∈ I∞(R), the
relation [a]R ⊥ [b]R implies that ab ∈ J(M∞(R)).
As an immediate consequence of Lemma 6.2, we record the following.
Proposition 6.4. Every exchange ring is weakly V-semiprimitive.
The converse of Proposition 6.4 fails: not every weakly V-semiprimitive ring is
an exchange ring (e.g., the ring Z of all integers).
Consider the diagram ~K represented in the left hand side of Figure 6.1: by
definition, s(x, y) = (y, x) and h(x, y) = x + y, for all (x, y) ∈ (Z+)2.
LIFTING DEFECTS IN NONSTABLE K0-THEORY
11
s
(Z+)2
h
/ Z+
s
R
h
/ S
Figure 6.1. The diagram ~K and a lifting ~R of ~K
The diagram ~K and Theorem 6.5 were both suggested to the author by Ken
Goodearl, thus improving the author's original formulation of Corollary 6.6 (which
was stated for functors Γ : S(1) → Exch).
Theorem 6.5. The diagram ~K has no lifting, with respect to the functor V, by
any diagram labeled as on the right hand side of Figure 6.1 in such a way that R
is weakly V-semiprimitive and h ◦ s = h.
Proof. Let η : V ~R → ~K be a natural equivalence, with component isomorphisms
ηR : V(R) → (Z+)2 and ηS : V(S) → Z+. We obtain a commutative diagram as on
Figure 6.2.
V(R)
ηR
(Z+)2
V(s)
s
V(R)
ηR
/ (Z+)2
V(h)
h
V(S)
ηS
/ Z+
Figure 6.2. A commutative diagram of commutative monoids
There are a, b ∈ I∞(R) such that ηR([a]R) = (1, 0) and ηR([b]R) = (0, 1). As ηR
is an isomorphism, [a]R ⊥ [b]R. Chasing around the diagram of Figure 6.2, we get
ηR([b]R) = (0, 1) = s(1, 0) = (s ◦ ηR)([a]R) = (ηR ◦ V(s))([a]R) = ηR([s(a)]R) ,
thus [b]R = [s(a)]R, and thus [a]R ⊥ [s(a)]R, and so, as R is weakly V-semiprimitive,
as(a) ∈ J(M∞(R)). Hence there exists x ∈ M∞(R) such that as(a)+x−as(a)x = 0.
By applying h to each side of this equation and observing that a2 = a while h◦s = h,
we obtain that h(a) + h(x) − h(a)h(x) = 0, thus, multiplying this equation on the
left by h(a), we get h(a) = 0. Therefore,
1 = h(1, 0) = (h ◦ ηR)([a]R) = (ηS ◦ V(h))([a]R) = ηS([h(a)]S) = 0
in Z+ ,
a contradiction.
(cid:3)
Corollary 6.6. Denote by S(1) the category of all simplicial monoids with order-
unit and normalized monoid homomorphisms, and by wVSem the category of all
weakly V-semiprimitive rings and ring homomorphisms. Then there is no functor
Γ : S(1) → wVSem such that V ◦ Γ ∼= id.
We shall see in Proposition 12.7 that while the diagram ~K has no lifting, with
respect to the functor V, by weakly V-semiprimitive rings with h ◦ s = h, it has
one by unital C*-algebras.
:
:
/
=
=
/
/
/
/
/
/
/
12
F. WEHRUNG
7. From a transfinite counterexample to non-splitting of K0
As the following result will be superseded by Corollary 10.8, we shall only outline
its proof.
Proposition 7.1. Denote by Semb
grp (1) the category of all simplicial groups with
order-unit with normalized positive group embeddings. Then there is no functor
Γ : Semb
grp (1) → AF such that K0 ◦ Γ ∼= id.
Outline of proof. We shall first invoke some ideas from universal algebra and lattice
theory (cf. McKenzie, McNulty, and Taylor [22] for background). A bounded lattice
is a lattice with smallest element and largest element (usually denoted by 0 and 1,
respectively). A variety of bounded lattices is the class of all bounded lattices that
satisfy a given set of identities, written in the language (∨, ∧, 0, 1). For a variety V
of bounded lattices and a set X, we denote by FV(X) the free object on X within
the variety V. We shall denote by M3 the lattice of length two with three atoms (cf.
Figure 7.1) and by M3 the variety that it generates. As M3 is finite, the variety M3
Figure 7.1. The lattice M3
is locally finite, that is, FM3 (X) is finite whenever X is finite (this is a classical
argument of universal algebra, see, for example, the proof of McKenzie, McNulty,
and Taylor [22, Lemma 4.98]). Furthermore,
FM3(X) = lim
−→
(FM3(Y ) Y ⊆ X finite) ,
(7.1)
with canonical transition maps and limiting maps, for any set X.
We proceed by invoking a lattice-theoretical tool. The dimension monoid Dim L
of a lattice L, defined in Wehrung [32, Definition 1.1], is always a conical commu-
tative monoid; furthermore, the assignment L 7→ Dim L extends to a functor. As
in [32], we define K0(L) as the (preordered) Grothendieck group of Dim L.
Now fix any set X with at least ℵ2 elements. By [32, Corollary 10.30], there is no
regular ring R such that Dim FM3(X) ∼= V(R). Furthermore, as the Dim functor
preserves direct limits (cf. [32, Proposition 1.4]), it follows from (7.1) that
Dim FM3(X) = lim
−→
(Dim FM3 (Y ) Y ⊂ X finite) ,
(7.2)
and thus
(7.3)
For each finite Y ⊂ X, the lattice FM3 (Y ) is finite modular, thus its dimension
monoid Dim FM3(Y ) is a finitely generated simplicial monoid (cf.
[32, Proposi-
−→(cid:0)K0(cid:0)FM3(Y )(cid:1) Y ⊂ X finite(cid:1) .
K0(cid:0)FM3(X)(cid:1) = lim
thermore, FM3(Y ) is a retract of FM3(X) (send every element of X \ Y to 0), thus
tion 5.5]); in particular, it embeds into its Grothendieck group K0(cid:0)FM3(X)(cid:1). Fur-
K0(cid:0)FM3(Y )(cid:1) is a retract of K0(cid:0)FM3(X)(cid:1). Finally, all the transition morphisms
in (7.2) are trivially seen to preserve the canonical order-units (namely, using the
notation of [32], ∆(0, 1)). Therefore, the direct limit in (7.3) is, in fact, a direct
LIFTING DEFECTS IN NONSTABLE K0-THEORY
13
union of dimension groups with order-unit. Hence, by assumption on Γ, we can
define an AF C*-algebra as a C*-direct limit,
A := lim
(7.4)
Let R be a dense locally matricial algebra (over C) in A. Then K0(R) ∼= K0(A).
Furthermore, as the functor K0 preserves C*-direct limits and K0 ◦ Γ ∼= id, it follows
from (7.4) that
−→(cid:0)Γ K0(cid:0)FM3 (Y )(cid:1) Y ⊂ X finite(cid:1)
for C*-algebras .
K0(A) = lim
−→(cid:0)K0(cid:0)FM3(Y )(cid:1) Y ⊂ X finite(cid:1)
in the category of dimension groups ,
so K0(A) ∼= K0(cid:0)FM3(X)(cid:1), and so V(R) ∼= Dim FM3(X). However, as R is von Neu-
mann regular, this is a contradiction.
(cid:3)
The proof of Proposition 7.1 may feel quite outlandish to a number of readers:
while its statement is, essentially, combinatorial, its proof involves the counterex-
ample of cardinality ℵ2 introduced in [31]. Extracting the gist of that proof will
lead us to the proof of Theorem 10.1, which is, to a large extent, a finitary result,
involving neither transfinite cardinals, nor universal algebra, nor lattice theory. It
will also extend Proposition 7.1 to C*-algebras of real rank 0.
8. The category of premeasured rings
Definition 8.1. Let M be a conical commutative monoid and let R be a ring. An
M -valued premeasure on R is a map µ : I∞(R) → M such that
(M0) µ(0) = 0;
(M1) µ(a + b) = µ(a) + µ(b), for all orthogonal idempotents a, b ∈ I∞(R);
(M2) a ∼ b implies that µ(a) = µ(b), for all a, b ∈ I∞(R). (Recall that ∼ stands
for Murray-von Neumann equivalence, see Section 4.)
We say that the premeasure µ is a measure if it satisfies the condition
(M0+) µ(e) = 0 iff e = 0, for each e ∈ I∞(R).
We say that the premeasure µ is a V-premeasure if it satisfies the following V-
condition:
(M3) For all c ∈ I∞(R) and all α, β ∈ M such that µ(c) = α + β, there are
orthogonal idempotent a, b ∈ I∞(R) such that c = a + b, µ(a) = α, and
µ(b) = β.
We say that µ is a V-measure if it is both a measure and a V-premeasure.
We shall use the notation M = Mµ (the codomain of µ).
A fundamental example of a V-measure is the following.
Example 8.2. Let R be a ring. Then the assignment τR : I∞(R) → V(R), e 7→ [e]R
is a measure on R. Due to Lemma 4.2, τR is actually a V-measure. We call it the
canonical V-measure on R.
Definition 8.3. Let M and N be conical commutative monoids. A homomorphism
f : M → N is a pre-V-homomorphism if whenever a, b ∈ N and c ∈ M with
f (c) = a + b, there are x, y ∈ M such that c = x + y, f (x) = a, and f (y) = b.
If, in addition, f −1 {0N } = 0M , then we say that f is a V-homomorphism (cf.
Dobbertin [9, Definition 1.2]).
The following result relates the definition of a measure given above, the canonical
V-measure, and the notion of V-homomorphism. It expresses the universality of τR
as a premeasure on R, and its proof is a straightforward exercise.
14
F. WEHRUNG
Proposition 8.4. Let R be a ring, let M be a conical commutative monoid, and
let µ be an M -valued premeasure on R. Then there exists a unique monoid homo-
morphism µ : V(R) → M such that µ = µ ◦ τR. Furthermore,
(i) µ is a measure iff µ−1 {0} = {0};
(ii) µ is a V-premeasure iff µ is a pre-V-homomorphism.
Definition 8.5. For a category R of rings and ring homomorphisms, we shall
denote by Rpmeas the category where the objects (premeasured rings) are the pairs
(R, µ), where µ is a premeasure on the ring R ∈ R (with values in some conical
commutative monoid), and, for objects (R, µ) and (S, ν), a morphism from (R, µ)
to (S, ν) is a pair (f, f ), where f : R → S is a ring homomorphism, f : Mµ → Mν is
a monoid homomorphism, and ν ◦ I∞(f ) = f ◦ µ (cf. Figure 8.1). The composition
R
Mµ
f
f
/ S
/ Mν
I∞(R)
µ
Mµ
I∞(f )
f
I∞(S)
ν
/ Mν
Figure 8.1. A morphism from (R, µ) to (S, ν)
of morphisms is defined componentwise (i.e., (f, f ) ◦ (g, g) = (f ◦ g, f ◦ g)).
We define similarly Rmeas, RVpmeas, RVmeas, using measures, V-premeasures,
and V-measures, respectively, instead of premeasures.
Next, we introduce two functors: one to Ringpmeas, and one from Ringpmeas.
Definition 8.6. The canonical functor τ : Ring → Ringpmeas is defined by τ (R) :=
(R, τR) (cf. Example 8.2) and τ (f ) := (f, V(f )), for any rings R and S and any
ring homomorphism f : R → S.
The projection functor Π : Ringpmeas → CMon is defined by Π(R, µ) := Mµ
and Π(f, f ) := f , for any premeasured rings (R, µ) and (S, ν) and any morphism
(f, f ) : (R, µ) → (S, ν).
Remark 8.7. It is trivial that V = Π ◦ τ . In particular, if a diagram can be lifted
by rings with respect to the functor V, then it can be lifted by premeasured rings
with respect to the functor Π.
From now on, we shall often omit the second coordinate f in a morphism (f, f )
from Ringpmeas, thus simply writing f : (R, µ) → (S, ν) and specifying what is f
in case needed.
Lemma 8.8. The category Exchmeas is a reflective subcategory of Exchpmeas, with
reflector Θr : Exchpmeas → Exchmeas satisfying Π ◦ Θr = Π. Furthermore,
(i) Θr sends surjective premeasures to surjective measures;
(ii) Θr sends V-premeasures to V-measures;
(iii) Θr sends regular rings to regular rings.
Proof. Let µ be a premeasure, with values in a conical commutative monoid M ,
on an exchange ring R, and denote by µ : V(R) → M the canonical monoid ho-
momorphism. Denote by I the two-sided ideal of R generated by the entries of all
/
/
/
/
/
LIFTING DEFECTS IN NONSTABLE K0-THEORY
15
x ∈ I∞(R) such that µ(x) = 0. By Lemma 4.3 together with the conicality of M ,
µ(e) = 0 for each e ∈ I∞(I); so µ↾V(I) = 0. Therefore, denoting by ρ : V(R) ։
V(R)/ V(I) the canonical projection, there exists a unique monoid homomorphism
µ : V(R)/ V(I) → M such that µ = µ ◦ ρ. Then R∗ := R/I is an exchange ring,
while, composing µ with the canonical isomorphism V(R/I) ∼= V(R)/ V(I) (cf.
Proposition 5.4) and using the fact that idempotents can be lifted modulo I, we
obtain a (necessarily unique) monoid homomorphism µ∗ : V(R/I) → M such that
µ∗([e + M∞(I)]R/I ) = µ(e) ,
for each e ∈ I∞(R) .
The map µ∗ defined by the rule µ∗(e) := µ∗([e]R∗ ) for each e ∈ I∞(R∗) is an
M -valued premeasure on R∗, and
µ∗(e + M∞(I)) = µ(e) ,
for each e ∈ I∞(R) .
(8.1)
For each e ∈ I∞(R) such that µ∗(e + M∞(I)) = 0, that is, µ(e) = 0, it follows from
the definition of I that e ∈ M∞(I). Therefore, µ∗ is an M -valued measure on R∗.
Claim. (R∗, µ∗) is the Exchmeas-reflection of (R, µ), with Exchmeas-reflection
morphism the canonical projection π : R ։ R∗ with π := idM .
Proof of Claim. We must prove that for every measured exchange ring (S, ν) and
every morphism ϕ : (R, µ) → (S, ν),
exists a unique morphism
ϕ∗ : (R∗, µ∗) → (S, ν) such that ϕ = ϕ∗ ◦ π. For each x ∈ I, ν(ϕ(x)) = ϕ(µ(x)) =
ϕ(0) = 0, thus, as ν is a measure, ϕ(x) = 0. It follows that there exists a unique
ring homomorphism ϕ∗ : R∗ → S such that ϕ = ϕ∗ ◦ π. Setting ϕ∗ := ϕ, we obtain
that ϕ is as desired.
(cid:3) Claim.
there
Denote by Θr : (R, µ) 7→ (R∗, µ∗) the Exchmeas-reflection functor. As µ and µ∗
have the same codomain and ϕ∗ := ϕ in the proof of the Claim above, Π ◦ Θr = Θr
and the surjectivity of µ implies the one of µ∗. If µ is a V-premeasure, then, as
idempotents can be lifted modulo I, so is µ∗ (cf. (8.1)). Finally, if R is a regular
ring, then R∗ = R/I is also regular.
(cid:3)
The following analogue of Lemma 8.8 for C*-algebras of real rank 0 is valid.
Lemma 8.9. The category RR0
pmeas → RR0
reflector Θa : RR0
meas is a reflective subcategory of RR0
meas satisfying Π ◦ Θa = Π. Furthermore,
pmeas, with
(i) Θa sends surjective premeasures to surjective measures;
(ii) Θa sends V-premeasures to V-measures.
Proof. Using Proposition 5.1, the proof is almost identical to the one of Lemma 8.8.
The main difference is the need to replace the ideal I of R by its topological closure I
(so R/I is still a C*-algebra, necessarily of real rank 0, see Brown and Pedersen [7,
Theorem 3.14]). We need to prove that µ(x) = 0 for each projection x of I. By
Lemma 4.5, x is equivalent to a projection y of I. By the proof of Lemma 8.8,
meas
µ(y) = 0; thus µ(x) = 0. Now the reflection morphism Θa : RR0
is defined by Θa(R, µ) := (R∗, µ∗), where this time
pmeas → RR0
µ∗(e + M∞(I)) = µ(e) ,
for each e ∈ I∞(R) ,
(8.2)
while π : R ։ R/I is the canonical projection and π = idM . The rest of the proof
runs as the one of Lemma 8.8.
(cid:3)
16
F. WEHRUNG
9. V-semiprimitive rings
The following definition introduces a strengthening of the weakly V-semiprimitive
rings introduced in Definition 6.3.
Definition 9.1. A ring R is V-semiprimitive if for all a, b ∈ I∞(R), [a]R ⊥ [b]R
in V(R) implies that ab = 0.
If V(R) is totally ordered with respect to its algebraic preordering, then R is
V-semiprimitive. In particular, this holds if V(R) ∼= Z+. This isomorphism oc-
curs in case R is a principal ideal domain. Due to the Quillen-Suslin Theorem,
it also holds in case R is a polynomial ring over a field. Another important class
of V-semiprimitive rings, which also explains the terminology, is provided by the
following result.
Proposition 9.2. Every semiprimitive exchange ring is V-semiprimitive.
We emphasize here that the rings in question are not necessarily unital.
Proof. Let R be a semiprimitive exchange ring and let a, b ∈ I∞(R) such that
[a] ⊥ [b] in V(R). Denoting by A and B the two-sided ideals of R generated by the
entries of a and b, respectively, it follows from Lemma 6.2 that A ∩ B is contained
in J(R), thus, by assumption, A ∩ B = {0}, so ab = 0.
(cid:3)
The following two examples show that none of the assumptions of semiprimitivity
and exchange property can be dispensed with in the statement of Proposition 9.2.
Example 9.3. Let K be a field (any unital exchange ring would do). Then the
ring R of all upper triangular 2×2 matrices over K is an exchange ring, but it is not
Observe that V(R) ∼= Z+ × Z+, and once the two monoids are identified, [a] = (1, 0)
V-semiprimitive. To show the latter statement, set a :=(cid:18)1 1
while [b] = (0, 1). In particular, [a] ⊥ [b], although ab =(cid:18)0 1
0 0(cid:19) and b :=(cid:18)0
0 0(cid:19) is nonzero.
1(cid:19).
0
0
Example 9.4. We denote by C(X, A) the C*-algebra of all A-valued continuous
functions on a topological space X, for any C*-algebra A. We denote by [0, 1] the
unit interval of R and we set
D := {x ∈ C([0, 1], M2(C)) x(0) is a diagonal matrix} .
This type of construction appears in Su [28, Section 1.2]. The ring D is a key
ingredient in the main construction in Section 12. Then D is a semiprimitive unital
ring (because it is a unital C*-algebra, cf. Proposition 4.7). However, D is not
V-semiprimitive. Indeed, denoting by z : [0, 1] ֒→ C the inclusion map and setting
z
0
1(cid:19) ,
0(cid:19) and b :=(cid:18)0
then a and b are both idempotent and ab =(cid:18)0 z
a :=(cid:18)1
0
0
0
0(cid:19) is nonzero. However, we prove
in Section 12 that V(D) ∼= Z+ × Z+, via an isomorphism that is easily seen to
send [a] to (1, 0) and b to (0, 1). In particular, [a] ⊥ [b] in V(D). Therefore, D is
not V-semiprimitive.
LIFTING DEFECTS IN NONSTABLE K0-THEORY
17
Lemma 9.5. Let R be a V-semiprimitive ring, let M be a conical commutative
monoid, and let µ : I∞(R) → M be a measure. Let a, b, e ∈ I∞(R) with a ≤ e and
b ≤ e while µ(a) ⊥ µ(e − b) and µ(b) ⊥ µ(e − a) in M . Then a = b.
Proof. Denote by µ : V(R) → M the unique V-homomorphism such that µ(x) =
µ([x]) for each x ∈ I∞(R) (cf. Proposition 8.4). We claim that [a] ⊥ [e − b] in V(R).
Indeed, otherwise there would exist a nonzero x ∈ I∞(R) such that [x] ≤ [a] and
[x] ≤ [e − b]. By applying µ to those inequalities, we obtain µ(x) ≤ µ(a) and
µ(x) ≤ µ(e − b), a contradiction as µ(x) 6= 0 and µ(a) ⊥ µ(e − b), and thus proving
our claim. As R is V-semiprimitive, it follows that a(e − b) = (e − b)a = 0, thus, as
a ≤ e, we get that a = ab = ba. By symmetry, a = ab = ba = b.
(cid:3)
10. More unliftable diagrams of simplicial monoids
The main combinatorial object underlying the present paper is a commutative
diagram, which we shall denote by ~D, of simplicial monoids and monoid homo-
morphisms. This diagram is indexed by the cube, that is, the partially ordered
set of all subsets of a three-element set. Its vertices are the commutative monoids
A = B = Z+ and Aj = Bj = Z+×Z+ (j ∈ {0, 1, 2}). The monoid homomorphisms
labeling its edges are the identity map, together with the maps e : Z+ → Z+ × Z+,
s : Z+ × Z+ → Z+ × Z+, and h : Z+ × Z+ → Z+ defined by e(x) := (x, x),
s(x, y) := (y, x), and h(x, y) := x + y for all x, y ∈ Z+.
(The maps s and h
were already introduced in Section 6.)
We shall later consider the diagram, denoted by ~Du, obtained by associating
order-units to the vertices of the diagram ~D: 1 is associated to A, (1, 1) is associated
to both Aj and Bj for each j ∈ {0, 1, 2}, and 2 is associated to B. Observe
that ~Du is a commutative diagram of pointed monoids, which means here that
e(1) = (1, 1) = s(1, 1) and h(1, 1) = 2.
The diagrams ~D and ~Du are represented on the left- and right-hand side of
Figure 10.1, respectively.
B2
B1
B0
(B2, (1, 1))
(B1, (1, 1))
(B0, (1, 1))
A0
A2
(A0, (1, 1))
(A1, (1, 1))
(A2, (1, 1))
h
s^
@
^========
^========
e
B
h
A1
e
A
h
^========
========
========
@
e
s
h
8qqqqqqqqqq
fMMMMMMMMMM
qqqqqqqqqq
qqqqqqqqqq
fMMMMMMMMMM
e
(B, 2)
h
e
(A, 1)
h
fMMMMMMMMMM
MMMMMMMMMM
MMMMMMMMMM
qqqqqqqqqq
qqqqqqqqqq
8qqqqqqqqqq
e
Figure 10.1. The diagrams ~D and ~Du
The following theorem will involve the functor Π : Ringpmeas → CMon intro-
duced in Definition 8.6, together with V-semiprimitivity (cf. Definition 9.1).
@
O
O
^
8
O
O
f
f
^
O
O
@
f
O
O
8
18
F. WEHRUNG
Theorem 10.1. There exists no lifting of ~D, with respect to the functor Π, by any
commutative diagram of premeasured rings satisfying the following conditions:
(i) the lifts of A0, A1, A2 are surjective V-premeasures;
(ii) the lifts of B0, B1, B2 are surjective V-measures;
(iii) the underlying rings of the lifts of B0, B1, B2 are V-semiprimitive.
Proof. Suppose that we are given a lifting ~R, with respect to the functor Π, of the
diagram ~D, labeled as on Figure 10.2 (we remind the reader that f is now dropped
from the notation (f, f ) for morphisms in Ringpmeas). We further assume that the
conditions (i) -- (iii) above hold.
(B, β)
h1
(B1, β1)
f1
h2
:uuuuuuuuu
B
g2:::::::::
dIIIIIIIII
e0
g0
g1
f0
h0
dIIIIIIIII
(B0, β0)
]:::::::::::::::
B
:uuuuuuuuu
(A2, α2)
e1
e2
(B2, β2)
]:::
f2
(A0, α0)
(A1, α1)
(A, α)
Figure 10.2. A commutative diagram ~R of premeasured rings lifting ~D
By composing the premeasures in the diagram ~R with the arrows of the natural
equivalence given by the relation Π ~R ∼= ~D, we may assume that Π ~R = ~D. This
means that the following equalities hold for each j ∈ {0, 1, 2}: Mα = Mβ = Z+,
Mαj = Mβj = Z+ × Z+, ej = e, hj = h, g2 = s, and f0 = f1 = f2 = g0 = g1 =
idZ+×Z+ .
There exists u ∈ I∞(A) such that α(u) = 1. Set uj := ej(u) ∈ I∞(Aj ), for each
j ∈ {0, 1, 2}. Then αj(uj) = ejα(u) = e(1) = (1, 1). Likewise, "propagating" those
new idempotents up the diagram ~R, we obtain idempotent matrices vj ∈ I∞(Bj),
for j ∈ {0, 1, 2}, and v ∈ I∞(B) such that v2 = f2(u0) = g2(u1) and cyclically,
v = h2(v2) = h1(v1) = h0(v0), β2(v2) = β1(v1) = β0(v0) = (1, 1), and β(v) = 2.
Define new matrices xj ∈ I∞(Aj ), for j ∈ {0, 1, 2}, as follows. As αj (uj) =
(1, 0) + (0, 1) in Aj and αj is a V-premeasure, there exists xj ∈ I∞(Aj) such that
xj ≤ uj, αj(xj ) = (1, 0), and αj (uj − xj) = (0, 1). Set
(y0
z0
:= f0(x1)
:= g0(x2)
, (y1
z1
:= f1(x0)
:= g1(x2)
, (y2
z2
:= f2(x0)
:= g2(u1 − x1) .
:
O
O
d
O
O
B
]
B
]
O
O
d
O
O
:
LIFTING DEFECTS IN NONSTABLE K0-THEORY
19
Then yj and zj both belong to I∞(Bj) and yj, zj ≤ vj, for each j ∈ {0, 1, 2}.
Further calculations yield easily
βj(yj) = βj(zj) = (1, 0) and βj(vj − yj) = βj(vj − zj) = (0, 1) ,
for each j ∈ {0, 1, 2} .
(10.1)
For example,
while
β2(y2) = β2(f2(x0)) = f2α0(x0) = f2(1, 0) = (1, 0)
β2(v2 − z2) = β2(g2(x1)) = g2α1(x1) = s(1, 0) = (0, 1) .
By evaluating the equation h1f1 = h2f2 at x0, we obtain h1(y1) = h2(y2). By
evaluating h0f0 = h2g2 at x1, we obtain h0(y0) = h2(v2 − z2). By evaluating
h0g0 = h1g1 at x2, we obtain h0(z0) = h1(z1).
Now, as each Bj is V-semiprimitive and by (10.1), it follows from Lemma 9.5
that yj = zj for each j ∈ {0, 1, 2}. Therefore,
h0(y0) = v − h2(y2) = v − h1(y1) = v − h0(y0) ,
a contradiction as h0(y0) and v are both idempotent with h0(y0) ≤ v and v 6= 0. (cid:3)
We observed in Remark 8.7 that any lifting of a diagram with respect to the
functor V yields a lifting of the same diagram with respect to the functor Π. As
the canonical measure on a ring is a V-measure (cf. Example 8.2), we obtain the
following immediate consequence of Theorem 10.1.
Corollary 10.2. There exists no lifting of ~D, with respect to the functor V, by any
commutative diagram of rings and ring homomorphisms in which the lifts of B0,
B1 and B2 are all V-semiprimitive.
Corollary 10.3. There exists no lifting of ~D, with respect to the functor Π, by any
commutative diagram of V-premeasured regular rings.
Proof. Suppose otherwise, and let ~R be a commutative diagram of V-premeasured
regular rings such that Π ~R ∼= ~D. By Lemma 8.8, the diagram ~R′ := Θr ~R is a
commutative diagram of V-measured regular rings and Π ~R′ = Π ~R ∼= ~D. But
every regular ring is a semiprimitive exchange ring (cf. Proposition 4.6), thus it is
V-semiprimitive (cf. Proposition 9.2). This contradicts Theorem 10.1.
(cid:3)
A similar proof, using Lemma 8.9 instead of Lemma 8.8, yields the following.
Corollary 10.4. There exists no lifting of ~D, with respect to the functor Π, by any
commutative diagram of V-premeasured C*-algebras of real rank 0.
By using Remark 8.7, we obtain the following negative lifting results with respect
to the functor V.
Corollary 10.5. There is no lifting, with respect to the functor V, of the diagram ~D
by any diagram of regular rings (resp., C*-algebras of real rank 0).
Nonetheless, we shall see that ~D can be lifted, with respect to the functor V, by
a diagram of exchange rings (cf. Proposition 11.3) and by a diagram of C*-algebras
(cf. Proposition 12.6).
In order to be able to extend Proposition 7.1 to C*-algebras of real rank 0, we
need to find an analogue of the diagram ~D in which the arrows are all embeddings.
20
F. WEHRUNG
Fortunately, this can be done. We denote by f , g : (Z+)2 ֒→ (Z+)4 and a, b,
c : (Z+)4 ֒→ (Z+)5 the maps defined by
f (x, y) := (x, x, y, y) ,
g(x, y) := (x, y, x, y) ,
a(x1, x2, x4, x4) := (x1, x2, x3, x4, x2 + x3) ,
b(x1, x2, x4, x4) := (x2, x1, x3, x4, x1 + x4) ,
c(x1, x2, x4, x4) := (x2, x3, x1, x4, x1 + x4) ,
for all x, y, x1, x2, x3, x4 ∈ Z+. Further, we denote by ~E the commutative dia-
gram of simplicial monoids and normalized monoid order-embeddings represented
in Figure 10.3.
B′
2 := (Z+)4
f
1 := (Z+)4
g
f
0 := (Z+)4
g
A0 := (Z+)2
A1 := (Z+)2
A2 := (Z+)2
B′ := (Z+)5
g
a
B′
8ppppppppppp
fNNNNNNNNNNN
8ppppppppppp
gOOOOOOOOOOO
e
b
e
c
B′
fNNNNNNNNNNN
8ppppppppppp
fNNNNNNNNNNN
7ooooooooooo
f
e
A := Z+
Figure 10.3. The diagram ~E
Lemma 10.6. There exists a natural transformation ~χ : ~E → ~D in which all the
arrows are surjective pre-V-homomorphisms.
Proof. We describe the components of the natural transformation.
• The morphism χB : B′ ։ B is defined by the rule
χB(x1, x2, x3, x4, x5) := x5.
• The morphism χB2 : B′
2 ։ B2 is defined by the rule
χB2 (x1, x2, x3, x4) := (x2, x3).
• The morphism χB1 : B′
1 ։ B1 is defined by the rule
χB1 (x1, x2, x3, x4) := (x1, x4).
• The morphism χB0 : B′
0 ։ B0 is defined by the rule
χB0 (x1, x2, x3, x4) := (x1, x4).
• The morphism χAj : Aj ։ Aj is the identity on (Z+)2 for each j < 3, and
χA : A ։ A is the identity on Z+.
Each of these arrows is the canonical projection from M × N onto M , for suit-
able (simplicial) monoids M and N ; hence it is a surjective pre-V-homomorphism.
Verifying that ~χ is a natural transformation (which, in reality, amounts to verifying
the commutativity of only five squares) is trivial.
(cid:3)
*
8
?
O
O
T
4
f
?
O
O
*
8
4
T
f
*
8
4
T
f
?
O
O
4
T
g
?
O
O
*
7
LIFTING DEFECTS IN NONSTABLE K0-THEORY
21
Denote by Semb(1) the category of all simplicial monoids with order-unit, with
normalized monoid order-embeddings (the latter means that the maps f satisfy
f (x) ≤ f (y) implies x ≤ y, where ≤ denotes the algebraic preordering, which is
here an ordering, on the monoid). Actually, all arrows of the diagram ~E are not
only order-embeddings but coretractions, that is, they all have a left inverse, so
Corollaries 10.7 and 10.8 extend to the category of all simplicial monoids endowed
with those maps as well.
Corollary 10.7. There is no functor Γ : Semb(1) → RR0
Γ : Semb(1) → RegVpmeas) such that Π ◦ Γ ∼= id.
(resp.,
Vpmeas
Vpmeas. It suffices to prove that the diagram ~E has
Proof. We give the proof for RR0
no lifting, with respect to the functor Π, by a diagram of V-premeasured C*-algebras
of real rank 0. If ~A were such a lifting, then, composing all the V-premeasures in ~A
with the corresponding pre-V-homomorphisms given by Lemma 10.6, we would
obtain a lifting of the diagram ~D, with respect to the functor Π, by a diagram
in RR0
Vpmeas, a contradiction by Corollary 10.4.
The proof for regular rings is similar, using this time Corollary 10.3.
(cid:3)
By using Remark 8.7, we obtain the following extension of Proposition 7.1.
Corollary 10.8. There is no functor Γ : Semb(1) → RR0 (resp., Γ : Semb(1) →
Reg) such that V ◦ Γ ∼= id.
We shall see in the subsequent sections that the V-semiprimitivity assumption
cannot be dispensed with in the statement of Theorem 10.1. We do not know
whether Corollary 10.7 extends to all C*-algebras, or even to all rings, see Prob-
lems 4 and 5.
11. Collapse ~D, collide with exchange rings
Five arrows out of six in the middle part of the diagram ~D (cf. Figure 10.1) are
identities. This suggests that the structure of ~D is highly collapsible. In the present
section we shall take advantage of this observation and define another diagram,
denoted by ~L, obtained by identifying as many "symmetric" parts of ~D as possible
while keeping its main unliftability property (Theorem 10.1). Denote by ~Lu the
diagram obtained by adding the natural order-units to the vertices of ~L. The
diagrams ~L and ~Lu are represented in Figure 11.1. The maps e, h, and s are
already arrows of ~D.
Proposition 11.1. There is no lifting ~R, with respect to the functor V, of the
diagram ~L, such that, labeling ~R as on Figure 11.2,
h ◦ f ◦ s = v ◦ h ◦ f ,
v ◦ h ◦ g = h ◦ g ,
f ◦ e = g ◦ e ,
s ◦ e = e ,
(11.1)
and C is V-semiprimitive. In particular, C cannot be a semiprimitive exchange
ring, and thus it can be neither a regular ring nor a C*-algebra of real rank 0.
Note. Equations such as V(s) = s, V(f ) = idZ+×Z+ , and so on, are, strictly speak-
ing, stated only up to the natural equivalence between V( ~R) and ~Lu.
Proof. Suppose that ~R satisfies the given conditions. Then we can "unfold" ~R to
the commutative diagram represented in Figure 11.3, which lifts ~D with respect to
the functor V; a contradiction by Corollary 10.2.
(cid:3)
22
F. WEHRUNG
Z+
id
(Z+, 2)
id
h
Z+ × Z+
id
id
h
(Z+ × Z+, (1, 1))
h(x, y) := x + y
id
id
Z+ × Z+
s
(Z+ × Z+, (1, 1))
s
s(x, y) := (y, x)
e
Z+
e
(Z+, 1)
e(x) := (x, x)
Figure 11.1. The diagrams ~L and ~Lu
D
h
C
f
g
B
e
A
v
V(v) = idZ+
V(h) = h
s
V(f ) = V(g) = idZ+×Z+ , V(s) = s
V(e) = e
Figure 11.2. A lifting ~R of ~L
h
f s
C
[77777777
C
[77777777
g
e
C
g
B
vh
g
f C
[77777777
[77777777
C
C
e
e
C
g
B
D
vh
C
B
A
Figure 11.3. A lifting of ~D
}
}
x
x
O
O
O
O
I
I
U
U
I
I
U
U
O
O
O
O
}
}
O
O
}
}
I
I
V
V
O
O
C
O
O
[
O
O
C
[
[
O
O
[
O
O
C
LIFTING DEFECTS IN NONSTABLE K0-THEORY
23
Theorem 11.2. Let K be a field. Then the diagram ~Lu has a lifting, with respect
to the functor V1, by unital exchange algebras over K, in such a way that the
equations (11.1) are satisfied and s and v are both involutive automorphisms.
Proof. Denote by K(t) (resp., K[t]) the field of rational functions (resp. the ring
of polynomials) over K in the indeterminate t. We denote by σ the unique auto-
morphism of K(t) such that σ(t) = 1 − t. Observe that σ2 = idK(t).
The subalgebra
K(t)(0) := {p(t)/q(t) p, q ∈ K[t] and q(0) 6= 0}
(11.2)
is a local subring of K(t), with maximal ideal tK(t)(0). In particular, K(t)(0) is
idempotent elements in the K-algebra
an exchange ring. The matrices c0 :=(cid:18)1
C :=(cid:18) K(t)(0)
0
tK(t)(0)
tK(t)(0) K(t)(0)(cid:19) .
0
0(cid:19) and c1 := 1 − c0 =(cid:18)0
0
0
1(cid:19) are both
(11.3)
As c0Cc0 ∼= c1Cc1 ∼= K(t)(0) are exchange rings, it follows from Nicholson [24,
Corollary 2.6] that C is an exchange subring of D := M2(K(t)). We denote by v
the involutive automorphism of D defined by the rule
σx1 σx0(cid:19) ,
x2 x3(cid:19) :=(cid:18)σx3 σx2
v(cid:18)x0 x1
for all x0, x1, x2, x3 ∈ K(t) .
The set J := tM2(cid:0)K(t)(0)(cid:1) is a two-sided ideal of C. Furthermore, the relation
det(1 − x) ≡ 1 (mod tK(t)(0)) holds for each x ∈ J; in particular, 1 − x is invertible
in C. Hence J is contained in the Jacobson radical of C.
Set B := K × K, b0 := (1, 0), and b1 := (0, 1). Then the map s : B → B,
(x, y) 7→ (y, x) is an involutive automorphism of B with s(b0) = b1. Furthermore,
V(B) ∼= Z+ × Z+ with simplicial basis {[b0]B, [b1]B}. We can define a surjective
homomorphism ϕ : C ։ B of K-algebras by the rule
ϕ(cid:18)x0 x1
x2 x3(cid:19) := (x0(0), x3(0)) .
As the Jacobson radical of K × K is zero, it follows that J contains, and thus
is equal to, the Jacobson radical of C. From V(J) = {0} and Proposition 5.4 it
follows that V(C) ∼= V(C/J) via [e]C 7→ [e + M∞(J)]C/J , thus V(C) ∼= Z+ × Z+
with simplicial basis {[c0]C , [c1]C }.
Now the matrix
c :=(cid:18)(1 − t)2(1 + 2t)
t(1 − t)(3 − 2t)
t(1 − t)(1 + 2t)
t2(3 − 2t) (cid:19)
0 0(cid:19) = c0(0), that is, c ≡ c0
(11.4)
is idempotent, and it belongs to C. As c(0) = (cid:18)1 0
(mod J), we get the first key property of c:
[c]C = [c0]C and [1 − c]C = [1 − c0]C .
Furthermore, an elementary calculation yields the second key property of c:
v(c) = c .
(11.5)
(11.6)
24
F. WEHRUNG
Therefore, defining h as the inclusion map from C into D and defining embeddings
of unital K-algebras f : K 2 ֒→ C and g : K 2 ֒→ C by the rules
f (x, y) := xc0 + y(1 − c0) and g(x, y) := xc + y(1 − c) ,
for all x, y ∈ K ,
we obtain the equations hf s = vhf and (using (11.6)) vhg = hg.
Finally, set A := K and e : A ֒→ B, x 7→ (x, x). Obviously, se = e and f e = ge.
We obtain a diagram ~R as on Figure 11.2. Verifying that ~R lifts ~Lu, with respect to
the abovementioned choices of simplicial bases, amounts to verifying the following:
V(e)([1]A) = [b0]B + [b1]B
V(f )([bj]B) = [cj]C for each j < 2
V(g)([bj]B) = [cj]C for each j < 2
V(h)([cj ]C ) = [c0]D = [c1]D for each j < 2
V(v)([1D]D) = [1D]D
(obvious);
(obvious);
(this follows from (11.5))
(obvious);
(obvious).
This concludes the proof.
(cid:3)
The "unfolding" technique used in the proof of Proposition 11.1 yields then
immediately the following consequence of Theorem 11.2.
Proposition 11.3. The diagram ~Du can be lifted, with respect to the functor V1,
by a commutative diagram of unital exchange rings.
12. A lifting of ~Lu by C*-algebras
For any topological space X and any positive integer n, we define the C*-algebra
X := Mn(C(X, C)). In case X is given a distinguished point 0, we set
A(n)
I(n)
X :=nx ∈ A(n)
X := A(n)
X x(0) = 0o ,
X ! ,
I(n)
X
(n)
A
X
(n)
I
X
(n)
D
an ideal of A(n)
X ;
a C*-subalgebra of A
(2n)
X ;
J(n) := M2(I(n)
X ) ,
an ideal of D(n)
X .
We shall often identify A(n)
consisting of all constant functions. This way, the C*-algebra
X with C(X, Mn(C)) and C with the subalgebra of C(X, C)
D
(n)
X :=(cid:18)Mn(C)
0
0
Mn(C)(cid:19)
is a unital subalgebra of D(n)
X .
(n)
Convention. In all the notations A(n)
X , defined above, we
shall drop the subscript X and the superscript (n) in case X = [0, 1] (the unit
interval of R) and n = 1. We introduced the algebra D = D(1)
X , D(n)
X , J(n)
X , I(n)
X , D
[0,1] in Example 9.4.
Proposition 12.1. The algebras A, D, and M2(A) have both real rank and stable
rank equal to 1.
LIFTING DEFECTS IN NONSTABLE K0-THEORY
25
Proof. It follows from Brown and Pedersen [7, Proposition 1.1] that A = C([0, 1], C)
has real rank 1, while, by Rieffel [26, Proposition 1.7], A has stable rank 1. By
Rieffel [26, Theorem 3.3], it follows that M2(A) has also stable rank 1. By Brown
and Pedersen [7, Proposition 1.2], it follows that M2(A) has real rank 1 as well.
As A has stable rank 1, the argument of the second part of the proof of [26,
Theorem 3.3] shows easily that D has stable rank 1. By [7, Proposition 1.2], it
follows that the real rank of D is either 0 or 1. As A is isomorphic to a hereditary
subalgebra of D, it follows from [7, Corollary 2.8] that D has real rank 1.
(cid:3)
For each idempotent e ∈ D(n)
X , the trace tr(e(x)) is a nonnegative integer for
each x ∈ X. As the map tr ◦ e is continuous on X, it follows that if X is connected
(which will always be the case throughout this section), then the value of tr(e(x)),
for x ∈ X, is constant. We shall denote this value by tr(e) and call it the trace of e.
Observe that the trace of e is equal to the rank of e(x), for each x ∈ X.
The following result is contained in Examples 1.4.2 and 5.1.3(c) of Blackadar [6].
Lemma 12.2. Let a, b ∈ R with a ≤ b. Then V(C([a, b], C)) ∼= Z+. Further-
more, for any positive integer n, two idempotent elements of Mn(C([a, b], C)) are
equivalent iff they have the same trace.
Lemma 12.3. Let e ∈ D(n)
alent in D(n)
[0,1].
[0,1] be a projection. Then e and e(0) are unitarily equiv-
Proof. Set e := e(0). The common value r := tr(e) = tr(e) is an integer with
0 ≤ r ≤ 2n. As e is continuous at 0, there is ε ∈ (0, 1] such that ke(t) − e(0)k ≤ 1/2
for each t ∈ [0, ε]. Thus we obtain the inequality (between elements of D(n)
[0,ε])
ke ↾[0,ε] −e↾[0,ε]k ≤ 1/2 .
As e↾[0,ε] and e↾[0,ε] are both projections of D(n)
(n)
[0,ε], that is, there exists a unitary u ∈ D
D
(n)
[0,ε] such that
[0,ε], they are unitarily equivalent in
e↾[0,ε] = ue ↾[0,ε] u∗ .
Extend u to a unitary element of D
By replacing e by u∗eu, we may thus assume the following:
(n)
[0,1] by setting u(t) := u(ε) for each t ∈ [ε, 1].
e(t) = e(0)
for each t ∈ [0, ε] .
(12.1)
As e↾[ε,1] and e↾[ε,1] are projections of M2n(C([ε, 1], C)) with the same rank r, it
follows from Lemma 12.2 that they are unitarily equivalent in M2n(C([ε, 1], C)), so
there exists a unitary v ∈ M2n(C([ε, 1], C)) such that
e↾[ε,1] = ve ↾[ε,1] v∗ .
(12.2)
Extend v to a unitary element of D
By putting (12.1) and (12.2) together, we obtain easily that e = vev∗.
(n)
[0,1] by setting v(t) := v(ε) for each t ∈ [0, ε].
(cid:3)
As in Section 11, we set c0 :=(cid:18)1
0
0
0(cid:19) and c1 := 1 − c0 =(cid:18)0
0
0
1(cid:19). By using the
canonical isomorphism Mn(D) ∼= D(n)
following description of the nonstable K0-theory of D.
[0,1] together with Lemma 12.3, we obtain the
26
F. WEHRUNG
Corollary 12.4. Denote by ρ : C ։ D, x 7→ x(0) the canonical retraction. Then
V(ρ) : V(D) → V(D) ∼= Z+ × Z+ is an isomorphism, and {[c0]C, [c1]C} is a simpli-
cial basis of V(D).
In particular, idempotent matrices a, b ∈ I∞(D) are equivalent in M∞(D) iff
a + M∞(J) and b + M∞(J) are equivalent in M∞(D/J). Observe that D/J ∼= C2.
Theorem 12.5. The diagram ~Lu has a lifting, with respect to the functor V1, by
unital C*-subalgebras of M2(C([0, 1], C)), in such a way that the equations (11.1)
are satisfied and s and v are both involutive automorphisms.
Proof. The argument is similar to the one of the proof of Theorem 11.2. The role
played by K(t) in that proof is taken up here by A := C([0, 1]). We denote by σ
the automorphism of A defined by
σ(x)(t) := x(1 − t) ,
for each x ∈ A and each t ∈ [0, 1] .
We set C := D, D := M2(A), and we denote by v the involutive automorphism
of M2(A) defined by the rule
σx1 σx0(cid:19) ,
x2 x3(cid:19) :=(cid:18)σx3 σx2
v(cid:18)x0 x1
for all x0, x1, x2, x3 ∈ A .
Set B := C × C, b0 := (1, 0), and b1 := (0, 1). Then the map s : B → B,
(x, y) 7→ (y, x) is an involutive C*-automorphism of B, which switches b0 and b1.
Furthermore, V(B) ∼= Z+ × Z+ with simplicial basis {[b0]B, [b1]B}. Denote by
z : [0, 1] ֒→ C the inclusion map. The matrix
c :=(cid:18) 1 − z
pz(1 − z)
is a projection of D. As c(0) =(cid:18)1 0
0 0(cid:19) = c0(0), that is, c ≡ c0 (mod J), it follows
pz(1 − z)
(cid:19)
z
from Corollary 12.4 that
[c]C = [c0]C and [1 − c]C = [1 − c0]C .
Furthermore, an elementary calculation yields the second key property of c, namely
v(c) = c .
(12.3)
Therefore, defining h as the inclusion map from D into M2(A) and defining C*-em-
beddings f : C2 ֒→ D and g : C2 ֒→ D by the rules
f (x, y) := xc0 + y(1 − c0) and g(x, y) := xc + y(1 − c) ,
for all x, y ∈ K ,
we obtain the equations hf s = vhf and (using (12.3)) vhg = hg.
Finally, let e : C ֒→ B, x 7→ (x, x). Obviously, se = e and f e = ge. We obtain
a diagram ~R as on Figure 11.2, with C in place of A. Verifying that ~R lifts ~Lu,
with respect to the abovementioned choices of simplicial bases, is done in a similar
fashion as at the end of the proof of Theorem 11.2.
(cid:3)
The "unfolding" technique used in the proof of Proposition 11.1 yields then
immediately the following consequence of Theorem 12.5.
Proposition 12.6. The diagram ~Du can be lifted, with respect to the functor V1,
by a commutative diagram of unital C*-algebras and unital C*-homomorphisms.
LIFTING DEFECTS IN NONSTABLE K0-THEORY
27
The algebra D can also be applied to find a lifting of the diagram ~K represented
in Figure 6.1.
Proposition 12.7. The diagram ~K can be lifted, with respect to the functor V, by
a diagram of unital C*-algebras and unital C*-homomorphisms such that, labeling
the lifting as on Figure 6.1, h ◦ s = h.
Proof. We shall define unital C*-algebras R and S with unital C*-homomorphisms
s : R → R and h : R → S such that h ◦ s = h and, denoting by ~R the diagram
represented in the right hand side of Figure 6.1, V ~R ∼= ~K.
Set R := D and S := M2(C), with the already introduced c0 and c1. We set
γ(t) :=(cid:18) sin((π/2)t)
− cos((π/2)t)
cos((π/2)t)
sin((π/2)t)(cid:19) ,
for each t ∈ [0, 1] .
Observe that γ is a unitary element of M2(A) and γ /∈ D. Furthermore, set s(x) :=
γ · x · γ∗, for each x ∈ D. Then
s(c0)(0) =(cid:18) 0
−1 0(cid:19)(cid:18)1
1
0
0
0(cid:19)(cid:18)0 −1
0 (cid:19) = c1 ,
1
and, similarly, s(c1)(0) = c0.
It follows that s is an automorphism of D, and,
by Lemma 12.3, s(c0) ∼ c1 and s(c1) ∼ c0 in D. Hence, V(s) switches [c0]D
and [c1]D. Observe that by Corollary 12.4, V(D) is simplicial with simplicial basis
{[c0]C , [c1]C }.
Now let h : C → M2(C), x 7→ x(1). Then h(cj) = cj for each j ∈ {0, 1}.
Furthermore, from γ(1) = 1 it follows that for each x ∈ D,
(h ◦ s)(x) = h(γ · x · γ∗) = γ(1) · x(1) · γ(1)∗ = x(1) = h(x) ,
so h ◦ s = h, and so s and h are as desired.
(cid:3)
We do not know whether it is possible to ensure simultaneously h ◦ s = h and
s2 = idR in a lifting of ~K, with respect to the functor V, by unital C*-algebras.
13. Back to the transfinite: an exchange ring of cardinality ℵ3
Our first non-lifting result, Proposition 7.1, started with a counterexample of
cardinality ℵ2. We applied the combinatorial core of that result in Theorem 10.1.
Now the cycle swings back on itself: we shall get further negative representation re-
sults, algebraically stronger than those that we used in order to get Proposition 7.1,
now in cardinality ℵ3 (the present form of Theorem 10.1 would not yield ℵ2).
Throughout this section we fix a field K, and we denote by K(t)(0) (cf. (11.2))
and CK := (cid:18) K(t)(0)
tK(t)(0) K(t)(0)(cid:19) (cf. (11.3)) the K-algebras introduced in the
proof of Theorem 11.2. We also denote by AK the closure, under finite direct
products and direct limits (of unital K-algebras), of the set
tK(t)(0)
RK := {K} ∪ {CK} ∪ {M2(K(t))} .
(13.1)
As each member of RK is a unital exchange K-algebra with index of nilpotence at
most 2, and as these properties are preserved under finite products and direct limits,
every member R of AK is a unital exchange K-algebra with index of nilpotence at
most 2. As, by Proposition 5.3, V(R) is a (conical) refinement monoid, it follows
from Lemma 4.4 that V(R) is the positive cone of a dimension group in which [1R]
has index at most two. In particular, by Yu [33, Theorem 9], R has stable rank 1.
28
F. WEHRUNG
Now recall that CMon denotes the category of all conical commutative monoids
with monoid homomorphisms (cf. Definition 3.1). We shall also define the sub-
category CMon⇒ with the same objects but where the morphisms are the pre-
V-homomorphisms (cf. Definition 8.3), and the full subcategory CMon6ℵ0 of all
countable (i.e., at most countable) conical commutative monoids.
Denote by Φ : AK → CMon the functor V (i.e., R 7→ V(R), f 7→ V(f )).
We introduce in Definition 3.8.1 and Definition 3.8.2 of Gillibert and Wehrung [13]
attributes of certain collections of categories and functors, called left larderhood and
right larderhood. We shall use the notation used in these two definitions, yet trying
to spell out in detail the verifications to be performed, and we shall start with the
(traditionally easier) left part.
Lemma 13.1. The quadruple (AK , CMon, CMon⇒, Φ) is a left larder.
Proof. We check one after another the items defining left larderhood.
• (CLOS(AK )) and (PROD(AK)): by definition, AK is closed under direct
limits and finite products.
• (CONT(Φ)): it is well-known that the functor V preserves direct limits.
• (PROJ(Φ, CMon⇒)): we must check that V(f ) is a double arrow (i.e.,
a pre-V-homomorphism), for each morphism f : R → S in AK which is
a direct limit (in the category of all arrows of AK) f = lim
fj of
projections fj : Rj × Sj ։ Rj, (x, y) 7→ x, for each j ∈ J (J is a directed
partially ordered set). As V(Rj × Sj) ∼= V(Rj) × V(Sj ), with V(fj) the
canonical projection V(Rj ) × V(Sj) ։ V(Rj), V(fj) is obviously a pre-
V-homomorphism. As the class of pre-V-homomorphisms is easily seen to
be closed under direct limits (within the category of all arrows of CMon)
and the functor V preserves direct limits, it follows that V(f ) is a pre-V-
homomorphism.
(cid:3)
−→j∈J
In order to fill the "right larder" part, we need, for C*-algebras of real rank 0,
• RR0, the category of all C*-algebras of real rank 0 with C*-homomor-
phisms (cf. Definition 4.1).
sep, the category of all separable members of RR0.
• RR0
• Ψa denotes the functor V (i.e., B 7→ V(B), f 7→ V(f )) from RR0 to CMon.
For the side of regular rings, we need
• Reg, the category of all regular rings with ring homomorphisms.
• Reg6ℵ0, the category of all countable regular rings.
• Ψr denotes the functor V (i.e., B 7→ V(B), f 7→ V(f )) from Reg to CMon.
Back to C*-algebras of real rank 0, we prove
sep, CMon, CMon6ℵ0, CMon⇒, Ψa) is a
Lemma 13.2. The 6-uple (RR0, RR0
right ℵ1-larder.
Proof. The part (PRESℵ1(CMon6ℵ0 , Ψa)) involves the notion of a weakly ℵ1-pre-
sented structure introduced in Gillibert and Wehrung [13, Definition 1.3.2]. It is,
sep is separable, thus (cf. the Note on page 28 in
basically, easy: every B ∈ RR0
Blackadar [6]) V(B) is countable, thus (cf. Gillibert and Wehrung [13, Proposi-
tion 4.2.3]) weakly ℵ1-presented.
The part (LSr
ℵ1
(B)), for a given B ∈ RR0, is less obvious. We are given a count-
able conical refinement monoid M and a pre-V-homomorphism ψ : V(B) ⇒ M ,
LIFTING DEFECTS IN NONSTABLE K0-THEORY
29
together with a sequence fn : Bn → B of unital C*-homomorphisms (the fn can
be assumed to be monomorphisms, but this will play no role here). We are try-
ing to find a separable, unital C*-subalgebra C of B, of real rank 0, containing
is a surjective pre-V-homomorphism.
Sn∈Z+ fn(Bn), such that, if eC : C ֒→ B denotes the inclusion map, then ψ ◦ V(eC )
For each n ∈ Z+, denote by rn : R → R the continuous function defined by
n+3 , and rn is affine on the inter-
n+3 , rn(x) = 1 if x ≥ n+2
rn(x) = 0 if x ≤ 1
n+3 , n+2
val (cid:2) 1
n+3(cid:3). We say that a countable subset X of B is FFC-closed ("FFC"
stands for "Flat Function Calculus") if rn(x) ∈ M∞(X) for each self-adjoint el-
ement x ∈ M∞(X) and each n ∈ Z+. Furthermore, denote by F the set of all
countable, FFC-closed, *-closed unital Q[i]-subalgebras of B. A classical argument
about approximating projections on a dense set (see, for example, Blackadar [6,
Proposition 4.5.1]) shows that every member X of F satisfies the following:
For each ε > 0 and each projection e ∈ M∞(X) ,
there exists a projection x ∈ M∞(X) such that ke − xk < ε .
(13.2)
Claim. For all c ∈ I∞(B) and all α, β ∈ M with ψ([c]B) = α + β, there are
orthogonal idempotents a, b ∈ I∞(B) such that c = a + b while ψ([a]B) = α and
ψ([b]B) = β.
Proof of Claim. As ψ is a pre-V-homomorphism, there are α′, β′ ∈ V(B) such that
[c]B = α′ + β′ while ψ(α′) = α and ψ(β′) = β. By Lemma 4.2, there are orthogonal
idempotents a, b ∈ I∞(B) such that c = a + b while [a]B = α′ and [b]B = β′.
Hence a and b are as required.
(cid:3) Claim.
For each n ∈ Z+, pick a countable dense subset Xn ⊆ Bn. As ψ is surjective
{ψ([e]B) e ∈ I∞(Y0)}.
and M is countable, there exists Y0 ∈ F containing Sn∈Z+ fn(Xn) such that M =
Suppose that Yj ∈ F has been constructed for each j ≤ 2k. As B has real
rank 0 and Y2k is countable, there exists Y2k+1 ∈ F containing Y2k such that every
self-adjoint element of Q[i] × Y2k, viewed as a subset of the unitization of B, lies
within 1/(k + 1) of some invertible self-adjoint element of the unitization of Y2k+1.
As Y2k+1 and M are both countable and by the Claim above, there is Y2k+2 ∈ F
containing Y2k+1 such that for all c ∈ I∞(Y2k+1) and α, β ∈ M with ψ([c]B) =
α + β, there are orthogonal idempotents a, b ∈ I∞(Y2k+2) such that c = a + b while
ψ([a]B) = α and ψ([b]B) = β.
By construction, the union Y := Sk∈Z+ Yk belongs to F and the closure Y has
real rank 0. Furthermore, it follows from (13.2) that every projection of M∞(Y )
is equivalent to some projection in M∞(Y ); hence, by the construction of Y2k+2
from Y2k+1, it follows that ψ ◦ V(eY ) is a pre-V-homomorphism. Furthermore,
from Sn∈Z+ fn(Xn) ⊆ Y0 ⊆ Y it follows that Sn∈Z+ fn(Bn) ⊆ Y . Therefore,
C := Y is as required.
(cid:3)
Consider again the set RK of exchange algebras defined in (13.1). As Φ(RK)
is contained in CMon6ℵ0 , we obtain the following consequence of Lemmas 13.1
and 13.2 together with the (obvious) [13, Proposition 3.8.3].
Lemma 13.3. The 8-uple (AK, RR0, CMon, RK, RR0
ℵ1-larder.
sep, CMon⇒, Φ, Ψa) is an
30
F. WEHRUNG
The analogue of Lemma 13.2 for regular rings is the following.
Lemma 13.4. The 6-uple (Reg, Reg6ℵ0, CMon, CMon6ℵ0, CMon⇒, Ψr) is a
right ℵ1-larder.
The proof of Lemma 13.4 is similar, and actually slightly easier (because there
is no topology involved) than the one of Lemma 13.2, thus we omit it.
Lemma 13.3 is modified in a similar fashion.
Lemma 13.5. The 8-uple (AK , Reg, CMon, RK, Reg6ℵ0 , CMon⇒, Φ, Ψr) is an
ℵ1-larder.
Now everything is ready for the proof of the following result.
Theorem 13.6. For any field K, there exists a unital exchange K-algebra RK that
satisfies the following properties:
(i) RK has index of nilpotence 2 (thus stable rank 1).
(ii) For every C*-algebra of real rank 0 (resp., for every regular ring) B, there
is no surjective pre-V-homomorphism from V(B) onto V(RK).
In particular, there is no C*-algebra of real rank 0 (resp., no regular ring) B such
that V(RK) ∼= V(B). Furthermore, RK can be constructed as the (unital ) direct
limit of a system of ℵ3 finite products of members of RK . In particular, if K is
countable (or, more generally, has at most ℵ3 elements), then card RK ≤ ℵ3.
Proof. We apply [13, Lemma 3.4.2], called there CLL, to the following data:
• λ = µ = ℵ1;
• P is the powerset lattice of the three-element set {0, 1, 2};
• ~A is the commutative diagram of unital exchange K-algebras, given in
Proposition 11.3, such that V1 ~A ∼= ~Du. Observe that all objects of ~A
belong to RK (cf. (13.1));
• Λ is one of the ℵ1-larders given by either Lemma 13.3 or Lemma 13.5.
The structure F(X) ⊗ ~A involved in the statement of CLL, which will turn
out to be the desired counterexample RK, is the same for both larders (it
depends only of AK, ~A, P , and X).
In order for the assumptions underlying CLL to be fulfilled, we need P to admit
[13, Definition 3.2.1]) (X, X) such that card X ≤ ℵ3. As P is
an ℵ1-lifter (cf.
a finite lattice, it is a so-called almost join-semilattice (cf.
[13, Definition 2.1.2]),
thus, according to [13, Corollary 3.5.8], it is sufficient to prove that the relation
(ℵ3, <ℵ1) ❀ P (cf. [12, Definition 3.1], also [13, Definition 3.5.1]) holds. According
to the definition of the Kuratowski index kur(P ) given in [12, Definition 4.1], it
suffices to prove that kur(P ) ≤ 3. As P has exactly three join-irreducible elements,
this is a trivial consequence of [12, Proposition 4.2].
The statement of CLL involves a "P -scaled Boolean algebra", denoted there
by F(X) (where (X, X) is the abovementioned ℵ1-lifter), and the "condensate"
F(X) ⊗ ~A. A P -scaled Boolean algebra is a Boolean algebra endowed with a col-
lection, indexed by P , of ideals, subjected to certain constraints. For our present
purposes, neither the exact definition of a P -scaled Boolean algebra (cf. [13, Def-
inition 2.2.3]), nor the exact descriptions of the constructions of F(X) (cf.
[13,
Lemma 2.6.5]) and F(X) ⊗ ~A (cf. [13, Section 3.1]), will matter.
LIFTING DEFECTS IN NONSTABLE K0-THEORY
31
We set RK := F(X) ⊗ ~A. The statement that RK is a direct limit of finite prod-
ucts of members of RK, together with the cardinality bound on RK, are immediate
consequences of the following facts:
• F(X) is a P -scaled Boolean algebra with at most ℵ3 elements.
• Thus F(X) = lim
Bj, for a directed poset I of cardinality ℵ3 and
"compact" (i.e., here, finitely presented) P -scaled Boolean algebras Bj
(cf. [13, Proposition 2.4.6]).
−→j∈I
• F(X) ⊗ ~A = lim
• Each Bj ⊗ ~A is a finite product of members of ~A. This follows from [13,
(Bj ⊗ ~A). This follows from [13, Proposition 3.1.4].
−→j∈I
Definition 3.1.1].
Suppose that there exists a surjective pre-V-homomorphism χ : V(B) ⇒ V(RK),
for a C*-algebra B with real rank 0. As all the assumptions underlying CLL are
satisfied, there are a P -indexed diagram ~B of C*-algebras of real rank 0 and a
natural transformation ~χ : V ~B ⇒ V ~A. As V ~A ∼= ~D (cf. Proposition 11.3), we
get a natural transformation, that we shall denote again by ~χ, now ~χ : V ~B ⇒ ~D.
The double arrow notation for the natural transformation ~χ means here that each
component χp : V(Bp) → Dp (for p ∈ P ) of ~χ is a surjective pre-V-homomorphism.
It follows that the rule µp(e) := χp([e]Bp ), for e ∈ I∞(Bp), defines a Dp-
valued surjective V-measure on Bp. We thus obtain a diagram ~C of surjective
V-premeasures on (unital) C*-algebras of real rank 0 such that Π ~C ∼= ~D. But this
contradicts Corollary 10.4.
This completes the proof that RK is the desired counterexample for C*-algebras
of real rank 0.
We also need to prove that there exists no surjective pre-V-homomorphism
χ : V(B) ⇒ V(RK), for any regular ring B. The proof is similar to the one for
C*-algebras of real rank 0. We need to use Lemmas 13.4 and 13.5 instead of Lem-
mas 13.2 and 13.3, and Corollary 10.5 instead of Corollary 10.4.
(cid:3)
14. An ℵ3-separable C*-algebra
For an infinite cardinal κ, we say that a C*-algebra is κ-separable if it has a
dense subset of cardinality at most κ. In particular, ℵ0-separable means separable.
We claim that what we did in Section 13, using the diagram of exchange rings
constructed in Section 11, can be done for C*-algebras, using the diagram of C*-
algebras constructed in Section 12.
Consider again the C*-algebra D used in Example 9.4 and Section 12, and set
A := {C} ∪ {D} ∪ {M2(C([0, 1), C))} .
(14.1)
The relevant analogue of Theorem 13.6 is the following.
Theorem 14.1. There exists an ℵ3-separable unital C*-algebra E that satisfies the
following properties.
(i) E has index of nilpotence 2, and real rank and stable rank both equal to 1.
(ii) (V(E), [1E]) is the positive cone of a dimension group with order-unit of
index two.
(iii) For every C*-algebra of real rank 0 (resp., every regular ring) B, there
exists no surjective pre-V-homomorphism from V(B) onto V(E).
32
F. WEHRUNG
In particular, there is no C*-algebra of real rank 0 (resp., no regular ring) B such
that V(E) ∼= V(B). Furthermore, E can be constructed as the (unital ) C*-direct
limit of a system of ℵ3 finite products of members of A.
While stating that the nonstable K0-theory of C*-algebras is contained neither
in the one of regular rings nor in the one of C*-algebras of real rank 0 is no big deal
(for V(A) may not have refinement, even for a unital C*-algebra A, cf. Blackadar
[6, Example 5.1.3(e)]), Theorem 14.1 states this fact even for those C*-algebras
whose nonstable K0-theory is a refinement monoid, which is far more difficult.
The proof of Theorem 14.1 is very similar to the one of Theorem 13.6, now getting
the P -indexed diagram ~A from Proposition 12.6 instead of Proposition 11.3. The
right larder part requires no change. The left larder part needs to be modified in
the obvious way: the category AK needs to be replaced by the closure A of A
under finite products and C*-direct limits. As the functor V preserves finite direct
products and C*-direct limits, (V(X), [1X]) is a refinement monoid with order-
unit of index at most two, for each X ∈ A. As in the proof of Lemma 13.1, the
quadruple (A, CMon, CMon⇒, Φ) is a left larder. As every member of A has real
rank either 0 or 1 while it has stable rank 1 (cf. Proposition 12.1), this is also the
case for every direct limit of finite products of members of A, in particular for the
condensate E := F(X) ⊗ ~A. The remaining changes that need to be applied to the
proof of Theorem 13.6 are trivial.
15. Open problems
Problem 1. Does any of the classes {M (∃R regular ring)(M ∼= V(R))} and
{M (∃A C*-algebra of real rank 0)(M ∼= V(A))} contain the other? And on pos-
itive cones of dimension groups?
The "combinatorial" side of the second part of Problem 1 reads as follows.
Problem 2. Are there finite, lattice-indexed commutative diagrams of simplicial
monoids that can be lifted, with respect to the functor V, by diagrams in one of
the classes Reg and RR0 but not the other?
The results and methods of Elliott [11] and Goodearl and Handelman [15] imply
that the nonstable K-theories of regular rings and of C*-algebras of real rank 0
(and also of AF C*-algebras) agree on dimension groups of cardinality at most ℵ1.
Our next question is about the remaining cardinality gap ℵ2.
Problem 3. Can the bound ℵ3 be improved to ℵ2 in Theorems 13.6 and 14.1?
As to the non-representability by regular rings, it looks plausible that a positive
solution to Problem 3 would follow from the methods of Wehrung [31].
In that
paper, the fact that the principal right ideals in a regular ring form a lattice is
crucial. As the analogous result for C*-algebras of real rank 0 does not hold, it
sounds unlikely that the methods of [31] are ready to help finding a solution to
Problem 3 for those algebras. On the other hand, it is plausible that one may be
able to use special functors, similar to those involved in Gillibert and Wehrung
[13, Chapter 5] (which is pure lattice theory!), for which a suitable analogue of
Theorem 10.1 would remain valid.
Every conical commutative monoid with order-unit is isomorphic to V1(R) for
some unital hereditary ring R: this is proved in Theorems 6.2 and 6.4 of Bergman [4]
LIFTING DEFECTS IN NONSTABLE K0-THEORY
33
for the finitely generated case, and in Bergman and Dicks [5, page 315] for the gen-
eral case. The general, non-unital case is proved in Ara and Goodearl [2, Proposi-
tion 4.4]. In light of this result, the following problem is natural.
Problem 4. Does there exist a functor Γ, from the category of conical commu-
tative monoids with monoid homomorphisms to the category of rings and ring
homomorphisms, such that V ◦ Γ ∼= id?
We do not even know the answer to the following restricted version of Problem 4.
Problem 5. Does there exist a functor Γ, from the category of simplicial monoids
with monoid homomorphisms to the category of C*-algebras and C*-homomor-
phisms, such that V ◦ Γ ∼= id?
The lifting results of ~Lu obtained in Theorems 11.2 and 12.5 also suggest the
following problems.
Problem 6. Does there exist a functor Γ, from the category of simplicial monoids
with normalized monoid embeddings to the category of exchange rings and ring
homomorphisms, such that V ◦ Γ ∼= id?
Problem 7. Does there exist a functor Γ, from the category of simplicial monoids
with normalized monoid embeddings to the category of C*-algebras and C*-homo-
morphisms, such that V ◦ Γ ∼= id?
Of course, all the problems above have unital versions, which are open as well.
References
[1] P. Ara, Extensions of exchange rings, J. Algebra 197 (1997), 409 -- 423.
[2] P. Ara and K. R. Goodearl, Leavitt path algebras of separated graphs, J. Reine Angew. Math.,
to appear. Available online at http://arxiv.org/abs/1004.4979 .
[3] P. Ara, K. R. Goodearl, K. C. O'Meara, and E. Pardo, Separative cancellation for projective
modules over exchange rings, Israel J. Math. 105 (1998), 105 -- 137.
[4] G. M. Bergman, Coproducts and some universal ring constructions, Trans. Amer. Math.
Soc. 200 (1974), 33 -- 88.
[5] G. M. Bergman and W. Dicks, Universal derivations and universal ring constructions, Pacific
J. Math. 79 (1978), 293 -- 337.
[6] B. Blackadar, "K-Theory for Operator Algebras, Second edition". Mathematical Sciences
Research Institute Publications 5. Cambridge University Press, Cambridge, 1998. xx+300 p.
ISBN: 0-521-63532-2
[7] L. G. Brown and G. K. Pedersen, C ∗-algebras of real rank zero, J. Funct. Anal. 99 (1991),
131 -- 149.
[8] J. Dixmier, "Les C ∗-Alg`ebres et Leurs Repr´esentations". (French) Deuxi`eme ´edition. Cahiers
Scientifiques, Fasc. XXIX. Gauthier-Villars ´Editeur, Paris 1969. xv+390 p.
[9] H. Dobbertin, On Vaught's criterion for isomorphisms of countable Boolean algebras, Algebra
Universalis 15 (1982), 95 -- 114.
[10] E. G. Effros, D. E. Handelman, and C-L. Shen, Dimension groups and their affine represen-
tations, Amer. J. Math. 102, no. 2 (1980), 385 -- 407.
[11] G. A. Elliott, On the classification of inductive limits of sequences of semisimple finite-
dimensional algebras, J. Algebra 38 (1976), 29 -- 44.
[12] P. Gillibert and F. Wehrung, An infinite combinatorial statement with a poset parameter,
Combinatorica 31, no. 2 (2011), 183 -- 200.
[13] P. Gillibert and F. Wehrung, From objects to diagrams for ranges of functors, Lecture Notes
in Mathematics, Vol. 2029. Springer-Verlag, Heidelberg - Dordrecht - London - New York.
x+158 p. ISBN: 978-3-642-21773-9
34
F. WEHRUNG
[14] K. R. Goodearl, "Partially Ordered Abelian Groups with Interpolation", Mathematical Sur-
veys and Monographs, Vol. 20. American Mathematical Society, Providence, R.I., 1986.
xxii+336 p.
[15] K. R. Goodearl and D. E. Handelman, Tensor products of dimension groups and K0 of unit-
regular rings, Canad. J. Math. 38, no. 3 (1986), 633 -- 658.
[16] K. R. Goodearl, "Von Neumann Regular Rings". Second edition. Robert E. Krieger Publishing
Co., Inc., Malabar, FL, 1991. xviii+412 p. ISBN: 0-89464-632-X
[17] K. R. Goodearl, Von Neumann regular rings and direct sum decomposition problems, Abelian
groups and modules (Padova, 1994), Math. Appl. 343, Kluwer Acad. Publ., Dordrecht 1995,
249 -- 255.
[18] P. A. Grillet, Directed colimits of free commutative semigroups, J. Pure Appl. Algebra 9
(1976), no. 1, 73 -- 87.
[19] I. N. Herstein, "Noncommutative Rings". Second printing of the 1968 original. Carus Math-
ematical Monographs 15. John Wiley and Sons, 1971. xi+199 p.
[20] N. Jacobson, The radical and semi-simplicity for arbitrary rings, Amer. J. Math. 67, no. 2
(April 1945), 300 -- 320.
[21] I. Kaplansky, "Rings of Operators", W. A. Benjamin, Inc., New York - Amsterdam 1968,
viii+151 p.
[22] R. N. McKenzie, G. F. McNulty, and W. F. Taylor, "Algebras, Lattices, Varieties. Volume I."
The Wadsworth & Brooks/Cole Mathematics Series. Monterey, California: Wadsworth &
Brooks/Cole Advanced Books & Software, 1987. xii+361 p. ISBN: 0-534-07651-3
[23] G. J. Murphy, "C*-Algebras and Operator Theory". Academic Press, Inc., Boston, MA, 1990.
x+286 p. ISBN: 0-12-511360-9 .
[24] W. K. Nicholson, Lifting idempotents and exchange rings, Trans. Amer. Math. Soc. 229
(1977), 269 -- 278.
[25] E. Pardo,
(in
Catalan), Universitat Aut`onoma de Barcelona, April 1995. Available online at
https://sites.google.com/a/gm.uca.es/enrique-pardo-s-home-page/phd-thesis .
i anells d'intercanvi", Ph. D. Thesis
"Monoides de
refinament
[26] M. A. Rieffel, Dimension and stable rank in the K-theory of C ∗-algebras, Proc. London Math.
Soc. (3) 46, no. 2 (1983), 301 -- 333.
[27] M. Rørdam, F. Larsen, and N. Laustsen, "An Introduction to K-Theory for C ∗-Algebras".
London Mathematical Society Student Texts 49. Cambridge University Press, Cambridge,
2000. xii+242 p. ISBN: 0-521-78334-8; 0-521-78944-3
[28] H. Su, On the classification of C*-algebras of real rank zero:
inductive limits of matrix
algebras over non-Hausdorff graphs, Mem. Amer. Math. Soc. 114 (1995), no. 547, viii+83 p.
[29] J. Tuma and F. Wehrung, Simultaneous representations of semilattices by lattices with per-
mutable congruences, Internat. J. Algebra Comput. 11, no. 2 (2001), 217 -- 246.
[30] R. B. Warfield, Jr., Exchange rings and decompositions of modules, Math. Ann. 199 (1972),
31 -- 36.
[31] F. Wehrung, Non-measurability properties of interpolation vector spaces, Israel J. Math. 103
(1998), 177 -- 206.
[32] F. Wehrung, The dimension monoid of a lattice, Algebra Universalis 40, no. 3 (1998), 247 --
411.
[33] H.-P. Yu, Stable range one for exchange rings, J. Pure Appl. Algebra 98 (1995), 105 -- 109.
LMNO, CNRS UMR 6139, D´epartement de Math´ematiques, Universit´e de Caen, 14032
Caen Cedex, France
E-mail address: [email protected]
URL: http://www.math.unicaen.fr/~wehrung
|
1511.08420 | 1 | 1511 | 2015-11-26T15:28:55 | Orthogonal forms and orthogonality preservers on real function algebras revisited | [
"math.OA"
] | In 2014, we determine the precise form of a continuous orthogonal form on a commutative real C$^*$-algebra. We also describe the general form of a (not-necessarily continuous) orthogonality preserving linear map between commutative unital real C$^*$-algebras. Among the consequences, we show that every orthogonality preserving linear bijection between commutative unital real C$^*$-algebras is continuous. In this note we revisit these results and their proofs with the idea of filling a gap in the arguments, and to extend the original conclusions. | math.OA | math |
ORTHOGONAL FORMS AND ORTHOGONALITY PRESERVERS
ON REAL FUNCTION ALGEBRAS REVISITED
ANTONIO M. PERALTA
Abstract. In [1] we determine the precise form of a continuous orthogonal
form on a commutative real C∗-algebra. We also describe the general form
of a (not-necessarily continuous) orthogonality preserving linear map between
commutative unital real C∗-algebras. Among the consequences, we show that
every orthogonality preserving linear bijection between commutative unital
real C∗-algebras is continuous. In this note we revisit these results and their
proofs with the idea of filling a gap in the arguments, and to extend the original
conclusions.
1. Introduction
Let A be a real or complex C∗-algebra. Elements a, b in A are said to be orthog-
onal (written a ⊥ b) if ab∗ = b∗a = 0. A bilinear form V : A × A → C is said to
be orthogonal if V (a, b) = 0 whenever a ⊥ b. In [1] we establish a generalization
of a celebrated result due to S. Goldstein (see [2, Theorem 1.10]) by proving the
following result:
Theorem 1.1. [1, Theorem 2.4] Let V : A × A → R be a continuous orthogonal
form on a commutative real C∗-algebra, then there exist ϕ1 and ϕ2 in A∗ satisfying
V (x, y) = ϕ1(xy) + ϕ2(xy∗),
for every x, y ∈ A.
We recently realized the presence of a "gap" affecting some of the technical
results given in [1]. The concrete difficulties appear in the following arguments:
By the Gelfand theory for commutative real C∗-algebras, every commutative unital
real C∗-algebra A is C∗-isomorphic (and hence isometric) to a real function algebra
of the form C(K)τ = {f ∈ C(K) : τ (f ) = f }, where K is a compact Hausdorff
space, τ is a conjugation (i.e. a conjugate linear isometry of period 2) on C(K)
given by τ (f )(t) = f (σ(t)) (t ∈ K), and σ : K → K is a topological involution
(i.e. a period 2 homeomorphism) (compare [4, Proposition 5.1.4]). Let σ : K → K
be a topological involution on a compact Hausdorff space K. Clearly, the sets
N = {t ∈ K : σ(t) 6= t} and F = {t ∈ K : σ(t) = t} are open and closed subsets of
K, respectively. It is established in [1, Lemma 2.1 and its proof] that the family,
F , of all open subsets O ⊆ K such that O ∩ σ(O) = ∅, ordered by inclusion is
an inductive set, and hence, by Zorn's lemma, there exists an open subset O ⊂ F
2010 Mathematics Subject Classification. Primary 46H40; 4J10, Secondary 47B33; 46L40;
46E15; 47B48.
Key words and phrases. Orthogonal form, real C∗-algebra, orthogonality preservers, disjoint-
ness preserver, separating map.
Author partially supported by the Spanish Ministry of Economy and Competitiveness project
no. MTM2014-58984-P and Junta de Andaluc´ıa grant FQM375.
2
A.M. PERALTA
maximal with respect to the property O ∩ σ(O) = ∅. Immediately after [1, Lemma
2.1] it is claimed that,
◦
∪ O
◦
∪ σ(O)."
"by the maximality of O, K = F
Unfortunately, the above equality is not always true. Consider, for example,
K = T the unit sphere of C and σ : K → K, σ(t) = −t (t ∈ T). In this case F = ∅
◦
and O = {t ∈ T : ℑm(t) > 0} is a maximal set in F , but F
∪ σ(O) 6= K. This
gap affects several statements and proofs of technical results in [1, Sections 2 and
3].
◦
∪ O
We recall that a mapping T : A → B between real or complex C∗-algebras is
said to be orthogonality or disjointness preserving if a ⊥ b in A implies T (a) ⊥ T (b)
in B.
In the second main goal studied in [1], we consider orthogonality preserving
linear maps between real function algebras belonging to a special subclass of the
category of commutative unital real C∗-algebras. The algebras in this particular
subclass can be presented as follows: Let F be a closed subspace of a compact
Hausdorff space K. We denote by Cr(K) = Cr(K; F ) the real C∗-algebra of all
continuous functions f : K → C taking real values on F . The main result in [1, §3] is
presented in Theorem 3.2, where we establish a complete description of those linear
(not necessarily continuous) orthogonality preserving operators between Cr(K; F )-
spaces. It should be remarked here that the proof of this result is not affected by
the gap commented above. Among the consequences not affected by the difficulties,
we obtain the following result.
Theorem 1.2. [1, Remark 3.4, proof of Theorem 3.5 and Proposition 3.6] In the
notation above, let T : Cr(K1; F1) → Cr(K2; F2) be an orthogonality preserving
linear bijection. Then T is automatically continuous and T −1 preserves invertible
elements, that is, T −1(g) is invertible whenever g is an invertible element in Cr(L2).
(cid:3)
However, it must be remarked that the main goal expected from these results
(that is, [1, Theorem 3.5]) is directly jeopardized by the mistake introduced after [1,
Lemma 2.1]. The concrete obstacle appears because by applying the wrong identity
commented above, we identify every commutative unital real C∗-algebra with a
real function algebra of the form Cr(K; F ). More concretely, let T : C(K1)τ1 →
C(K2)τ2 be an orthogonality preserving linear mapping, and let σi : Ki → Ki be a
topological involution satisfying τi(f ) = f ◦ σi. According to what is claimed in [1,
page 287]:
". . . keeping in mind the notation in the previous section, we write Li := Oi ∪Fi,
where Oi and Fi = {t ∈ Ki : σi(t) = t} are the subsets of Ki given by Lemma [1,
Lemma 2.1]. The map sending each f in C(Ki)τi to its restriction to Li is a C∗-
isomorphism (and hence a surjective linear isometry) from C(Ki)τi onto the real
C∗-algebra Cr(Li) of all continuous functions f : Li → C taking real values on Fi.
Thus, studying orthogonality preserving linear maps between C(K)τ spaces is equiv-
alent to study orthogonality preserving linear mappings between the corresponding
Cr(L)-spaces."
Unfortunately, since, in general, Ki 6= Oi ∪ Fi ∪ σ(Oi), we cannot guarantee
that Li := Oi ∪ Fi, is a closed subset of Ki (compare the example given in page 2).
Furthermore, there are examples of C(K)τ -spaces which are not real C∗-isomorphic
to a real C∗-algebra of the form Cr(X, F ) (compare 3.1). In summary, the proof of
ORTHOGONAL FORMS AND ORTHOGONALITY PRESERVERS REVISITED
3
[1, Theorem 3.5] is only valid for orthogonality preserving linear bijections between
commutative unital real C∗-algebras which are of the form Cr(K, F ).
Though the main results in [1] remain valid in the form they are stated, the gap
commented above makes invalid some of the arguments given in that paper. It is
necessary to provide a complete and correct argument, which fix all the problems.
We present here a complete revision of these results with new and complete argu-
ments, which allow us to solve and fix the problems caused by the gap in the original
proof. The problems caused in [1, §2] are easily fixable and we shall just comment
briefly the necessary changes in Section 2 here. However, the difficulties caused in
the proof of the result asserting the automatic continuity of every orthogonality
preserving linear bijection between commutative unital real C∗-algebras force us to
present a more detailed revision in Section 3.
2. Orthogonal forms on real C∗-algebras
We revise in this section the proof of [1, Theorem 2.4]. For conciseness reasons,
we keep the notation in [1] without inserting explicit definitions. We shall confine
ourselves to state the minimum changes necessary to fix the difficulties in [1, §2].
Lemma 2.2 in [1] should be rewritten as follows:
Lemma 2.2. In the notation of Lemma 2.1, let B(A) = B(K)τ , let a ∈ B(K)τ
and let b be an element in B(A)skew. Then the following statements hold:
a) bF = 0;
b) For each ε > 0, there exist mutually disjoint Borel sets B1, . . . , Bm and real num-
sa,
c) For each ε > 0, there exist mutually disjoint Borel sets C1, . . . , Cm ⊂ K and
bers λ1, . . . , λm satisfying σ(Bj) ∩ Bj = ∅ and(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
mXj=1
real numbers µ1, . . . , µm satisfying σ(Cj ) = Cj and(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
b −
a −
mXj=1
− χσ(Bj ) )(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
µjχCj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
< ε.
i λj (χBj
< ε;
The arguments given in [1, Comments before Lemma 2.1 and Proof of Lemma
2.2] remain valid here.
The notation in [1, Lemma 2.3] should be replaced with the following: For each
C ⊆ K with C ∩ σ(C) = ∅ we shall write uC = i (χC − χσ(C) ). The symbol u0 will
stand for the projection χK\F . Clearly, 1 = χF + u0 where 1 is the unit element in
B(K)τ . By Lemma 2.2 a), for each b ∈ B(K)τ
skew we have b ⊥ χF , and so b = bu0.
The statement of [1, Proposition 2.3] should be modified in the following sense:
Proposition 2.3. Let K be a compact Hausdorff space, τ a period-2 conjugate-
linear isometric ∗-homomorphism on C(K), A = C(K)τ , and V : A × A → R
be an orthogonal bounded bilinear form whose Arens extension is denoted by V ∗∗ :
A∗∗ × A∗∗ → R. Let σ : K → K be a period-2 homeomorphism satisfying τ (a)(t) =
a(σ(t)), for all t ∈ K, a ∈ C(K). Then the following assertions hold for all Borel
subsets D, B, C of K with σ(B) ∩ B = σ(C) ∩ C = ∅ and σ(D) = D:
a) V (χD , uB ) = V (uB , χD ) = 0, whenever D ∩ B = ∅;
b) V (uB , uC ) = 0, whenever B ∩ C = ∅;
c) V ((u0 − uC u∗
C )uB , uC ) = V (uC , (u0 − uC u∗
C )uB ) = 0.
4
A.M. PERALTA
The proof given in [1] remains valid with the obvious changes in the notation.
We include here an sketch of the changes for completeness reasons.
Proof. By an abuse of notation, we identify V and V ∗∗. Arguing as in the first part
of the proof of [1, Proposition 2.3] we get
V (cid:18)uK2
,
1
2
(χK1
+ χσ(K1 ) )(cid:19) = V (cid:18) 1
2
(χK1
+ χσ(K1 ) ), uK2(cid:19) = 0,
V (cid:0)uK1
, uK2(cid:1) = 0,
whenever K1 and K2 are two compact subsets of K such that K1, K2, σ(K1) and
σ(K2) are pairwise disjoint.
B ∩ σ(B) = ∅. By inner regularity there exist nets of the form (χ
a) Let now D, B be two disjoint Borel subsets of K such that σ(D) = D and
)γ
)γ converge in the weak∗ topology of C(K)∗∗ to χD and
such that (χ
)λ and (χ
)λ and (χ
D
λ
B
γ
K
K
K
D
λ
K
B
γ
D
B
χB , respectively, where each K
By the assumptions made on D and B we have that K
and K
V we have
γ ⊆ B is a compact subset of K.
γ ) = ∅
γ ) = ∅ for all λ and γ. By (1) and the separate weak∗ continuity of
λ ⊆ D and each K
λ ∩ σ(K
γ ∩ σ(K
γ = K
λ ∩ K
D
D
B
B
B
B
V (χD , uB ) = w∗ − lim
λ w∗ − lim
γ
V χ
+ χ
K
D
λ
2
σ(K
D
λ
)
, u
K
γ !! = 0,
B
(1)
and
(2)
(3)
and
(4)
V (uB , χ
D
) = 0.
A similar argument, but replacing (1) with (2), applies to obtain b).
To prove the last statement, we observe that
(u0 − ucu∗
c )uB = (χK\F − χC − χσ(C) )uB = χK\(F ∪C∪σ(C)) uB = u(K\(C∪σ(C)))∩B ,
and hence the statement c) follows from b).
(cid:3)
The statement of Theorem 2.4 in [1] remains unaltered, however, the proof of
this theorem needs a slight modification from line 11.
Proof of Theorem 1.1. By Proposition [1, Proposition 1.5] we have
(5)
V (a1, a2) = V (a1a2, 1),
for every a1, a2 in B(K)τ
sa.
To deal with the skew-symmetric part, let D, B, C be Borel subsets of K with,
D = σ(D), B ∩ σ(B) = ∅ and C ∩ σ(C) = ∅. From Proposition 2.3 a), we have
(6)
V (χD , uB ) = V (χD , uB (1 − χD + χD )) = V (χD , uB∩(K\D) ) + V (χD , uB χD )
= V (χD − 1 + 1, uB χD ) = V (−χ(K\D) + 1, u(B∩D) ) = V (1, uB χD ),
and similarly,
(7)
V (uB , χD ) = V (uB χD , 1).
ORTHOGONAL FORMS AND ORTHOGONALITY PRESERVERS REVISITED
5
If we apply Proposition 2.3 b) and c), repeatedly, we deduce that
(8)
V (uB , uC ) = V (uB u0, uC ) = V (uB (u0 + uC u∗
C − uC u∗
C ), uC )
= V (uB uC u∗
C , uC ) = V (uB uC u∗
C , uC − u0 + u0)
= V (u(B∩C) , −u((K\F )\C) + u0) = V (u(B∩C) , u0) = V (uB uC , u0).
and similarly
(9)
V (uB , uC ) = V (u0, uB uC ).
mlXj=1
plXk=1
Let al =
µl,jχ
Dl
j
, bl =
λl,ku
Bl
k
(l ∈ {1, 2}) be two simple elements in
sa and B(K)τ
B(K)τ
{Dl
of K with σ(Dl
1, . . . , Dl
ml} and {Bl
skew, respectively, where λl,k, µl,j ∈ R, for each l ∈ {1, 2},
pl} are families of mutually disjoint Borel subsets
1, . . . , Bl
j ) = Dl
j and Bl
i ∩ σ(Bl
i) = ∅. By (5), (6), (7), and (8), we have
V (a1 + b1, a2 + b2) = V (a1a2, 1) +
D2
, χ
p2Xk=1
p2Xk=1
m1Xj=1
j(cid:19) +
p1Xk=1
µ1,jλ2,kV (cid:18)1, χ
p2Xk=1
m1Xj=1
, 1(cid:19) +
p1Xk=1
µ1,jλ2,kV (cid:18)χ
λ2,kλ1,kV (cid:16)u
k(cid:19)
λ2,kλ1,kV (cid:16)u
p2Xk=1
D2
j
D1
j
B2
u
+
p1Xk=1
m2Xj=1
µ2,jλ1,kV (cid:18)u
B1
k
= V (a1a2, 1) +
+
p1Xk=1
m2Xj=1
µ2,jλ1,kV (cid:18)u
B1
k
χ
, u
D1
j
B2
, u
B2
B1
k
k(cid:19)
k(cid:17)
u
B2
k
, u0(cid:17)
B1
k
= V (a1a2, 1) + V (1, a1b2) + V (b1a2, 1) + V (b1b2, u0)
= ψ1(a1a2) + ψ2 (a1b2) + ψ1 (b1a2) + ψ4 (b1b2) ,
where ψ1, ψ2, and ψ4 are the functionals in A∗ defined by ψ1(x) = V (x, 1), ψ2(x) =
V (1, x), and ψ4(x) = V (x, u0 ), respectively. Since, by Lemma 2.2, simple elements
of the above form are norm-dense in B(K)τ
skew, respectively, and V is
continuous, we deduce that
sa and B(K)τ
V (a1 + b1, a2 + b2) = ψ1(a1a2) + ψ2 (a1b2) + ψ1 (b1a2) + ψ4 (b1b2) ,
for every a1, a2 ∈ B(K)τ
sa, b1, b2 ∈ B(K)τ
skew.
Now, taking φ1 = 1
4 (2ψ1 + ψ2 + ψ4), φ2 = 1
4 (2ψ1 − ψ2 − ψ4), φ3 = 1
4 (ψ2 − ψ4),
and φ4 = 1
4 (ψ4 − ψ2), we get
V (a1 + b1, a2 + b2) = φ1((a1 + b1)(a2 + b2)) + φ2 ((a1 + b1)(a2 + b2)∗)
+φ3 ((a1 + b1)∗(a2 + b2)) + φ4 ((a1 + b1)∗(a2 + b2)∗) ,
for every a1, a2 ∈ B(K)τ
sa, b1, b2 ∈ B(K)τ
skew.
Finally, defining ϕ1(x) = φ1(x) + φ4(x∗) and ϕ2(x) = φ2(x) + φ3(x∗) (x ∈ A),
(cid:3)
we get the desired statement.
6
A.M. PERALTA
3. Orthogonality preservers between C(K)τ -spaces
In this section we shall study orthogonality preserving linear bijections between
commutative unital real C∗-algebras. The aim is to provide to the reader an ar-
gument to avoid the difficulties in the proof of [1, Theorem 3.5]. We have already
commented in the introduction that the arguments in the proof of [1, Theorem 3.5]
are only valid to show that every orthogonality preserving linear bijection between
Cr(K; F )-spaces is continuous (compare also page 287 and section 3 in the same
paper).
We begin this section with a remark that present a commutative unital real C∗-
algebra which is not C∗-isomorphic to a real function algebra of the form Cr(K, F ).
Remark 3.1. Let K = {t1, t2} equipped with the discrete topology, σ : K → K the
topological involution given by σ(t1) = t2. It is easy to check that C(K)τ ≡ CR , the
complex field regarded as a real space. Suppose there exists a compact Hausdorff
space X and a closed subset F ⊆ X such that C(K)τ ≡ CR is C∗-isomorphic to
Cr(X; F ). Since C(X, R) is a real subspace of Cr(X; F ), we easily deduce from
Urysohn's lemma that ♯X ≤ 2 and hence X = {s1, s2}.
In this case, there are
only three possibilities to consider, namely, F = ∅, F = {s2} and F = X. The
real C∗-algebra Cr(X, F ) coincides with C(K) = C ⊕∞ C, C ⊕∞ R and R ⊕∞ R,
respectively. None of the above real C∗-algebras is C∗-isomorphic to C(K)τ ≡ CR.
The above Remark 3.1 implies that we cannot derived that every orthogonality
preserving linear bijection between commutative unital real C∗-algebras is (auto-
matically) continuous as a consequence of [1, Theorem 3.2, Remark 3.4 and the
proof of Theorem 3.5].
Henceforth, let T : C(K1)τ1 → C(K2)τ2 be an orthogonality preserving real
linear bijection. Following standard notation, for each s ∈ K2, we denote by δs :
C(K2)τ2 → C the linear mapping given by δs(g) = g(s) (g ∈ C(K2)τ2 ). We observe
that T being surjective implies that δsT : C(K1)τ1 → C is a non-zero linear map.
The symbol supp(δsT ) will denote the set of all t ∈ K1 such that for each open
set U = σ1(U ) ⊆ K1 with t ∈ U there exists f ∈ C(K1)τ1 with coz(f ) ⊆ U and
δs(T (f )) 6= 0. Let us recall that the cozero set, coz(f ), of a function f ∈ C(K1)τ1
is the set {t ∈ K1 : f (t) 6= 0}. The equality σ1(coz(f )) = coz(f ) holds for every
f ∈ C(K1)τ1.
Proposition 3.2. (a) For each s ∈ K2 there exists a unique element ts ∈ K1 such
that the set supp(δsT ) = {ts, σ1(ts)};
(b) For every s ∈ K2, we have δsT is continuous if and only if δσ2(s)T is continuous.
Moreover, the equality supp(δsT ) = supp(δσ2(s)T ) holds for every s ∈ K2.
Proof. a) Let us take s ∈ K2. We first show that supp(δsT ) contains at most two
points of the form t and σ1(t). Arguing by contradiction, we assume that there exist
t1, t2 in supp(δsT ) with t1 6= t2, σ1(t2). In this case, we can find two open disjoint
subsets U1 and U2 in K1 with ti ∈ Ui = σi(Ui), and two elements f1, f2 ∈ C(K1)τ1
satisfying coz(fi) ⊆ Ui and δsT (fi) 6= 0, for every i = 1, 2. This is impossible
because f1 ⊥ f2 and T is orthogonality preserving.
We shall show next that supp(δsT ) 6= ∅. Otherwise, supp(δsT ) = ∅. Then, for
each t ∈ K1, there exists an open subset Ut = σ1(Ut) with t ∈ Ut and δs(T (f )) =
0 for every f ∈ C(K1)τ1 with coz(f ) ⊆ Ut. By a compactness argument, we
can find a finite open cover {U1, . . . , Um} of K1 satisfying Uk = σ1(Uk) (∀k) and
ORTHOGONAL FORMS AND ORTHOGONALITY PRESERVERS REVISITED
7
δs(T (f )) = 0 for every f ∈ C(K1)τ1 with coz(f ) ⊆ Uk for some k. Let g1, . . . , gm
be a continuous decomposition of the identity in C(K1) subordinate to U1, . . . , Um.
Since Uk = σ1(Uk), the elements f1 = g1+τ1(g1)
define a
continuous decomposition of the unit in C(K1)τ1 subordinate to U1, . . . , Um. For
each f ∈ C(K1)τ1 we have
, . . . , fm = gm+τ1(gm)
2
2
δsT (f ) =
mXk=1
δsT (f fi) = 0,
which contradicts δsT 6= 0.
b) Let s ∈ K2. Clearly, δs(g) = δσ2(s)(g), for every g ∈ C(K2)τ2 . The first
statement follows from the identity δsT = δσ2(s)T . Let us assume that supp(δsT ) 6=
supp(δσ2(s)T ). In this case there exist t1 ∈ supp(δsT ) and t2 ∈ supp(δσ2(s)T ) with
t1 6= t2, σ1(t2). We can find two open disjoint subsets U1 and U2 satisfying ti ∈
Ui = σ1(Ui) and two elements f1, f2 ∈ C(K1)τ1 with coz(fi) ⊆ Ui, δsT (f1) 6= 0 and
δσ2(s)T (f2) = δsT (f2) 6= 0, which contradicts T (f1) ⊥ T (f2).
(cid:3)
Let us define an equivalence relation on Ki given by t ∼ s if σi(t) = σi(s).
It is known that the quotient space [Ki] = Ki/ ∼ is compact. It is not hard to
check that [Ki] = Ki/ ∼ is a compact Hausdorff space. The equivalence class of
an element t ∈ Ki is denoted by [t] = {t′ ∈ Ki : t′ ∼ t}. Applying Proposition
3.2(a), we can define a map ϕ : K2 → [K1], s 7→ [t] = supp(δsT ). By Proposition
3.2(b), supp(δs1 T ) = supp(δs2 T ), for every s1, s2 in K2 with s1 ∼ s2. Therefore,
the mapping [ϕ] : [K2] → [K1], [s] 7→ [t] = supp(δsT ) is well defined.
Lemma 3.3. Let s be an element in K2, then δsT (f ) = 0 for every f ∈ C(K1)τ1
supp(δsT ) is
with supp(δsT ) ∩ coz(f ) = ∅. In particular, the set Supp(K2) = [s∈K2
dense in K1.
Proof. Suppose we have f ∈ C(K1)τ1 with supp(δsT ) ∩ coz(f ) = ∅. We can find an
open set U ⊆ K1 with supp(δsT ) ⊆ U = σ1(U ) and U ∩coz(f ) = ∅. By assumptions,
there exists g ∈ C(K1)τ1 with coz(g) ⊂ U and δsT (g) 6= 0. T being orthogonality
preserving and f ⊥ g imply that T (f ) ⊥ T (g), and hence δsT (f ) = 0.
For the second statement, suppose we can find t0 ∈ K1\ [s∈K2
observe that [s∈K2
ists f0 ∈ C(K1)τ1 such that 0 ≤ f0 ≤ 1, f0(t0) = 1 and coz(f0)∩Ss∈K2
supp(δsT ) =
∅. We deduce from the first part of the lemma that T (f0)(s) = 0, for every s ∈ K2,
which contradicts the injectivity of T .
(cid:3)
supp(δsT ) is σ1-symmetric. Then, by Urysohn's lemma, there ex-
supp(δsT )!. We
We shall derive next some consequences of the previous results.
Proposition 3.4. Let s be an element in K2 such that δsT : C(K1)τ1 → C is a
continuous linear map. Then, for each t ∈ [ϕ][s] = supp(δsT ), there exist λs, µs ∈ C
such that
δsT (f ) = λsℜeδt(f ) + µsℑmδt(f ),
8
A.M. PERALTA
for every f ∈ C(K1)τ1 . Moreover, λs is unique for every s, while µs is unique
whenever supp(δsT ) contains two points (i.e. σ1(ts) 6= ts). It is also clear that
λs = T (1)(s), for every s as above.
Proof. We already know that supp(δsT ) = {ts, σ1(ts)}, for a unique ts ∈ K1. We
fix this ts.
Let us consider the sets Jsupp(δs T ) := {f ∈ C(K1)τ1 : supp(δsT )∩coz(f ) = ∅} and
Ksupp(δs T ) := {f ∈ C(K1)τ1 : f supp(δs T ) = 0}. Clearly, Jsupp(δs T ) ⊆ Ksupp(δs T ). The
arguments given by K. Jarosz in [3, 141] remain valid to show that Jsupp(δs T ) is norm-
dense in Ksupp(δs T ) . Lemma 3.3 implies that Jsupp(δs T ) ⊆ ker(δsT ) = ker(δσ2 (s) T ).
We deduce from the continuity of δsT that
(10)
ker(δts ) = ker(δσ1 (ts )) = Ksupp(δs T ) ⊆ ker(δsT ) = ker(δσ2(s) T ).
The real linear functionals ℜeδsT , ℑmδsT , ℜeδts, and ℑmδts are all continuous.
Since ker(ℜeδts) ∩ ker(ℑmδts) = ker(δts) and ker(δsT ) = ker(ℜeδsT ) ∩ ker(ℑmδsT ),
we deduce from (10) the existence of α1, α2, β1, β2 ∈ R satisfying
ℜeδsT = α1ℜeδts + α2ℑmδts
and
Taking λs = α1 + iβ1 and µs = α2 + iβ2 we obtain
ℑmδsT = β1ℜeδts + β2ℑmδts.
δsT = λsℜeδts + µsℑmδts .
To prove the second statement suppose there are λ1, λ2, µ1 and µ2 in C such
that
λ2ℜeδts + µ2ℑmδts = δsT = λ1ℜeδts + µ1ℑmδts.
Pick, via Urysohn's lemma, a function f0 ∈ C(K1)τ1 satisfying f0(ts) = 1. Then
λ1 = λ2. When σ(ts) 6= ts, we can find another function f1 ∈ C(K1)τ1 satisfying
f1(ts) = i, and hence µ1 = µ2.
(cid:3)
Corollary 3.5. Under the above conditions, the following statements hold:
(a) Let s be an element in K2 such that δsT : C(K1)τ1 → C is a continuous linear
map and σ2(s) 6= s. Then f (ts) = 0, for every ts ∈ supp(δsT );
(b) Suppose s1, s2 are elements in K2 such that σ2(sj) 6= sj for every j = 1, 2,
δs1 T and δs2 T are continuous. If supp(δs1 T ) = supp(δs2 T ) then s1 = s2 or
s1 = σ2(s2).
Proof. (a) Let us fix ts ∈ supp(δsT ). By Proposition 3.4 there exist λs, µs ∈ C such
that
(11)
δsT (f ) = λsℜeδts(f ) + µsℑmδts(f ),
for every f ∈ C(K1)τ1. The surjectivity of T implies that for each ω ∈ C there are
real numbers α, β such that ω = λsα + µsβ. This proves that {λs, µs} is a basis of
the real space CR. Therefore, by (11), δsT (f ) = 0 implies ℜef (ts) = ℑmf (ts) =
δts(f ) = f (ts) = 0, and also f (σ1(ts)) = 0.
(b) Let us assume that supp(δs1 T ) = supp(δs2 T ) for s1 and s2 as in the hypoth-
esis. By (a), δs1 T (f ) = 0 implies f (ts) = 0, for every ts ∈ supp(δs1 T ), which, from
(11), entails that δs2 T (f ) = 0. Since C(K2)τ2 separates points s1, s2 in K2 with
s1 ≁ s2, the surjectivity of T gives s1 ∼ s2.
(cid:3)
ORTHOGONAL FORMS AND ORTHOGONALITY PRESERVERS REVISITED
9
Given s ∈ K2 we denote by kδsT k the norm of the linear mapping δsT :
C(K1)τ1 → C if the latter map is continuous, we set kδsT k = ∞ otherwise. By
Proposition 3.2(b), kδsT k = kδσ(s)T k, for every s ∈ K2.
Lemma 3.6. Let s ∈ K2 and t ∈ K1 such that [t] = supp(δsT ). Let U = σ1(U ) ⊂
K1 be an open set satisfying [t] ⊂ U . The following statements hold:
(a) If kδsT k < ∞, then for each ε > 0 there exists f ∈ C(K1)τ1 such that kf k ≤ 1,
coz(f ) ⊂ U and δsT (f ) > kδsT k − ε;
(b) If kδsT k = ∞, then for each R > 0 there exists f ∈ C(K1)τ1 such that kf k ≤ 1,
coz(f ) ⊂ U and δsT (f ) > R.
Proof. (a) Let g be an element in C(K1)τ1 such that kgk ≤ 1 and δsT (g) >
kδsT k − ε. Let V ⊆ K1 be an open set satisfying [t] ⊂ V ⊆ V ⊆ U . By Urysohn's
lemma there exists 0 ≤ u ≤ 1 in C(K1)τ1 with uV ≡ 1 and uK1\U ≡ 0. We
observe that t /∈ coz(1 − u), and thus, Lemma 3.3 implies that δsT (g(1 − u)) = 0.
We therefore have
δsT (g) = δsT (g(1 − u) + gu) = δsT (gu),
which gives the desired statement for f = gu ∈ C(K1)τ1 with kguk ≤ 1 and
coz(gu) ⊂ U .
The proof of (b) is very similar.
(cid:3)
Lemma 3.7. Under the above assumptions, let (tn) and (sn) be sequences in K1
and K2, respectively, such that [tn] 6= [tm], for every n 6= m and [tn] = supp(δsn T )
for every n ∈ N. Then sup{kδsnT k : n ∈ N} < ∞.
Proof. Arguing by contradiction, we assume that, sup{kδsnT k : n ∈ N} = ∞. Up
to an appropriate subsequence, we can find a sequence (Un) of mutually disjoint
open subsets in K1 such that [tn] ⊂ Un = σ1(Un) and sup{kδsnT k : n ∈ N} = ∞.
Since sup{kδsn T k : n ∈ N} = ∞, we apply Lemma 3.6 to find a subsequence
(nk) and sequence (fk)k ⊂ C(K1)τ1 such that kfkk ≤ 1, coz(fk) ⊂ Unk and
T (fk) > 2k, for every natural k. The functions in the sequence (fk)k are
δsnk
mutually orthogonal, therefore f0 =
fk defines an element in C(K1)τ1. By
1
k
∞Xk=1
∞Xk=1,k6=k0
orthogonality, for each k0 ∈ N, we have
δsnk0
T (f0) =(cid:12)(cid:12)(cid:12)δsnk0
T(cid:16) 1
k0
fk0 +
which is impossible.
1
k
fk(cid:17)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)δsnk0
T(cid:16) 1
k0
fk0(cid:17)(cid:12)(cid:12)(cid:12) >
2k0
k0
,
(cid:3)
Following a similar notation to that employed in [3], we set Z1 := {s ∈ K2 :
δsT is bounded } and Z2 := {s ∈ K2 : δsT is unbounded }. Clearly σ2(Zi) = Zi,
for every i = 1, 2. The following conclusions can be straightforwardly derived from
the previous results.
Corollary 3.8. (a) The set Supp(Z2) := [s∈Z2
isfies σ1(Supp(Z2)) = Supp(Z2). Furthermore, every point in Supp(Z2) is
non-isolated in K1;
(b) The set {kδsT k : s ∈ Z1, σ2(s) 6= s} is bounded.
supp(δsT ) ⊆ K1 is finite and sat-
10
A.M. PERALTA
Proof. (a) The first statement follows from Lemma 3.7. Suppose t0 ∈ Supp(Z2) is
isolated in K1. We know that t0 ∈ supp(δsT ) for certain s ∈ Z2. In this case, we do
not need the continuity assumptions in Proposition 3.4 to get a similar conclusion.
Indeed, since {t0} is a closed subset, a function f ∈ C(K1)τ1 vanishes at t0 if and
only if t0 /∈ coz(f ). Moreover, g0 = δt0 + δσ1(t0) ∈ C(K1)τ1 and f − f (t0)g0 vanished
at t0, for every f ∈ C(K1)τ1. Lemma 3.3 implies that δsT (f − f (t0)g0) = 0, and
hence δsT (f ) = δsT (f (t0)g0) for every f ∈ C(K1)τ1. We note that the restriction
of the real linear mapping δsT to the two dimensional subspace CRg0 is clearly
continuous, and the same property holds for the linear mapping δt0 g0 : C(K1)τ1 →
Cg0, f 7→ δt0 (f )g0. Therefore, the composed mapping f 7→ δsT (f (t0)g0) = δsT (f )
is linear and continuous, which contradicts that s ∈ Z2.
(b) Arguing by contradiction, we assume that {kδsT k : s ∈ Z1, σ2(s) 6= s} is
unbounded. We can find a sequence (sn) ⊂ Z1 such that 2n ≤ kδsn T k (cid:8) kδsn+1T k.
Clearly, sn 6= sm and sn 6= σ2(sm), for every n 6= m (we recall that kδsT k =
kδσ2(s)T k). Applying Corollary 3.5(b) we deduce that supp(δsn T )∩supp(δsm T ) = ∅.
Finally, Lemma 3.7 gives the desired contradiction.
(cid:3)
Let ∆1 := {(s, t) : s ∈ Z1, t ∈ supp(δsT )}. We define a mapping ϑ : ∆1 → C
given by ϑ(s, t) = 0 if σ1(t) = t and ϑ(s, t) = µs if σ1(t) 6= t, where µs is the unique
element given by Proposition 3.4.
Proposition 3.9. (a) Let (s, t) ∈ ∆1 with σ1(t) 6= t, and let g be any element in
C(K1)τ1 satisfying gU ≡ i for some open neighborhood U of t. Then T (g)(s) =
µs = ϑ(s, t).
(b) The function ϑ is bounded;
(c) The set {kδsT k : s ∈ Z1} is bounded;
(d) Z1 is closed.
Proof. (a) Follows straightforwardly from Proposition 3.4.
(b) If ϑ is unbounded, we can find a sequence (sn, tn) ∈ ∆1 satisfying 2n <
ϑ(sn, tn) < ϑ(sn+1, tn+1). Obviously, sn 6= sm and sn 6= σ2(sm), for every n 6= m
and supp(δsn T ) = {tn, σ1(tn)} with tn 6= σ1(tn). We claim that we can find a
T ) ∩ supp(δsnm T ) = ∅ for every
subsequence (snk , tnk ) ⊂ ∆1 satisfying supp(δsnk
k 6= m. Let us assume, on the contrary, that there exists a natural n0 such that
supp(δsn T ) = {t0, σ1(t0)}, for every n ≥ n0. Let g be any element in C(K1)τ1
satisfying gU ≡ i for some open neighborhood U of t0. It follows from (a) that
2n < ϑ(sn, tn) = T (g)(sn), for every n ≥ n0, which contradicts that T (g) ∈
C(K2)τ2. Having in mind that kδsnk
T k ≥ ϑ(snk , tnk ), the desired contradiction
follows from Lemma 3.7.
(c) We set M1 = sup{ϑ(s, ts) : (s, ts) ∈ ∆1}. Let s be an element in Z1, and
let ts be an element in supp(δsT ). By Proposition 3.4 the identity
δsT (f ) = T (1)(s)ℜeδts(f ) + ϑ(s, ts)ℑmδts(f ),
holds for every f ∈ C(K1)τ1. Therefore, δsT (f ) ≤ kT (1)k + M1 for every f ∈
C(K1)τ1 with kf k ≤ 1.
(d) Let M = sup{kδsT k : s ∈ Z1}. Fix an arbitrary z0 ∈ Z1 and a function
f ∈ C(K1)τ1 with kf k ≤ 1. We can find a net (zµ) ⊂ Z1 converging to z0 in
the topology of K2. Since T (f ) is a continuous function, we have T (f )(zµ) →
T (f )(z0). On the other hand, by (d) we have T (f )(zµ) =≤ M , for every µ.
ORTHOGONAL FORMS AND ORTHOGONALITY PRESERVERS REVISITED
11
Therefore, δz0T (f ) = T (f )(z0) ≤ M , for every f ∈ C(K1)τ1 with kf k ≤ 1, which
proves that δz0T is bounded, and hence z0 ∈ Z1.
(cid:3)
We state next the main result of this note, which provides an antidote to fill the
gap we commented at the introduction.
Theorem 3.10. Every orthogonality preserving linear bijection between commuta-
tive unital real C∗-algebras is continuous.
Proof. Let T : C(K1)τ1 → C(K2)τ2 be an orthogonality preserving linear bijection.
We shall prove that Z2 := {s ∈ K2 : δsT is unbounded } = ∅. Suppose, on the
contrary, that Z2 6= ∅. Since the set Z1 is closed (see Proposition 3.9(d)) the set Z2
is an non-empty open subset of K2 with σ2(Z2) = Z2. Find, via Urysohn's lemma,
a non-zero function g ∈ C(K2)τ2 with coz(g) ⊆ Z2. By the surjectivity of T there
exists 0 6= h ∈ C(K1)τ1 satisfying T (h) = g. According to the notation above,
supp(δsT ), and Supp(Z2) =
supp(δsT ), Supp(Z1) = [s∈Z1
we set Supp(K2) = [s∈K2
[s∈Z2
supp(δsT ). We claim that
(12)
h(t) = 0, for every t ∈ Supp(Z1).
To prove the claim, let t0 be an element in Supp(Z1). Since Supp(Z2) is a σ1-
symmetric finite set in K1 disjoint from Supp(Z1), there exist disjoint open sets
U1, U2 ⊆ K1 and a function k ∈ C(K1)τ1 such that σ1(Uj) = Uj for every j = 1, 2,
t0 ∈ U1, Supp(Z2) ⊆ U2, k∗ = k (i.e. k(K1) ⊆ R), k(t0) = 1, and coz(k) ⊆ U1. We
claim that T (kh) = 0. Indeed, for s ∈ Z2, since Supp(Z2) ∩ coz(k) = ∅, Lemma 3.3
implies that T (kh)(s) = δsT (kh) = 0. For each s ∈ Z1, and ts ∈ supp(δsT ), k(ts) ∈
R and hence kh, k(ts)h ∈ C(K1)τ1 with (kh)(ts) = (k(ts)h)(ts). By Proposition 3.4
we also have:
δsT (kh) = T (1)(s)ℜeδts(kh) + ϑ(s, ts)ℑmδts(kh),
= T (1)(s)ℜeδts(k(ts)h) + ϑ(s, ts)ℑmδts(k(ts)h)
= δsT (k(ts)h) = k(ts)δsT (h) = k(ts)δs(g) = 0,
because s ∈ Z1 and coz(g) ⊆ Z2. This proves the second claim, and hence T (kh) =
0. The injectivity of T implies kh = 0 and hence 0 = (kh)(t0) = h(t0), which
completes the proof of the first claim.
By Lemma 3.3 the set Supp(K2) is dense in K1. The subset Supp(Z2) is finite
and every point in Supp(Z2) is non-isolated in K1 (compare Corollary 3.8(a)).
◦
Since Supp(K2) = Supp(Z2)
∪ Supp(Z1). It is not hard to see, from the normality
of K1, that Supp(Z1) must be also dense in K1. Since, by (12), h(t) = 0 for
every t ∈ Supp(Z1), we deduce from the continuity of h that h = 0, and hence
0 = T (h) = g, which gives the final contradiction.
(cid:3)
A generalized version of [1, Corollary 3.10] for real C∗-algebras read as follows:
Corollary 3.11. The following statements are equivalent:
(a) There exists a bi-orthogonality preserving linear bijection T : C(K1)τ1 → C(K2)τ2;
(b) There exists a real C∗-isomorphism S : C(K1)τ1 → C(K2)τ2;
(c) There exists a C∗-isomorphism eS : C(K1) → C(K2);
(d) K1 and K2 are homeomorphic.
(cid:3)
12
A.M. PERALTA
References
[1] J. Garc´es and A. M. Peralta, Orthogonal
forms and orthogonality preservers on
function algebras, Linear and Multilinear Algebra 62, no. 3, 275-296 (2014).
real
DOI:10.1080/03081087.2013.772998
[2] S. Goldstein, Stationarity of operator algebras, J. Funct. Anal. 118, no. 2, 275-308 (1993).
[3] K. Jarosz, Automatic continuity of separating linear isomorphisms, Canad. Math. Bull. 33,
no. 2, 139-144 (1990).
[4] B.R. Li, Real operator algebras, World Scientific Publishing (Singapore), 2003.
Departamento de An´alisis Matem´atico, Facultad de Ciencias, Universidad de Granada,
18071 Granada, Spain.
E-mail address: [email protected]
|
1111.6750 | 2 | 1111 | 2012-04-21T04:24:57 | Classification of Traces and Associated Determinants on Odd-Class Operators in Odd Dimensions | [
"math.OA"
] | To supplement the already known classification of traces on classical pseudodifferential operators, we present a classification of traces on the algebras of odd-class pseudodifferential operators of non-positive order acting on smooth functions on a closed odd-dimensional manifold. By means of the one to one correspondence between continuous traces on Lie algebras and determinants on the associated regular Lie groups, we give a classification of determinants on the group associated to the algebra of odd-class pseudodifferential operators with fixed non-positive order. At the end we discuss two possible ways to extend the definition of a determinant outside a neighborhood of the identity on the Lie group associated to the algebra of odd-class pseudodifferential operators of order zero. | math.OA | math |
Symmetry, Integrability and Geometry: Methods and Applications
SIGMA 8 (2012), 023, 25 pages
Classif ication of Traces and Associated Determinants
on Odd-Class Operators in Odd Dimensions
Carolina NEIRA JIM ´ENEZ † and Marie Fran¸coise OUEDRAOGO ‡
† Fakultat fur Mathematik, Universitat Regensburg, 93040 Regensburg, Germany
E-mail: [email protected]
URL: http://homepages.uni-regensburg.de/~nec07566/
‡ D´epartment de Math´ematiques, Universit´e de Ouagadougou, 03 BP 7021, Burkina Faso
E-mail: [email protected]
Received November 30, 2011, in final form April 11, 2012; Published online April 21, 2012
http://dx.doi.org/10.3842/SIGMA.2012.023
Abstract. To supplement the already known classification of traces on classical pseu-
dodifferential operators, we present a classification of traces on the algebras of odd-class
pseudodifferential operators of non-positive order acting on smooth functions on a closed
odd-dimensional manifold. By means of the one to one correspondence between conti-
nuous traces on Lie algebras and determinants on the associated regular Lie groups, we give
a classification of determinants on the group associated to the algebra of odd-class pseudo-
differential operators with fixed non-positive order. At the end we discuss two possible ways
to extend the definition of a determinant outside a neighborhood of the identity on the Lie
group associated to the algebra of odd-class pseudodifferential operators of order zero.
Key words: pseudodifferential operators; odd-class; trace; determinant; logarithm; regular
Lie group
2010 Mathematics Subject Classification: 58J40; 47C05
1
Introduction
From the connection between the trace of a matrix with scalar coefficients and its eigenvalues,
one can derive a relation between the trace and the determinant of a matrix, namely
det(A) = exp(tr(log A)).
(1)
At the level of Lie groups, a trace on a Lie algebra is the derivative of the determinant at the
identity on the associated Lie group. Using the exponential mapping between a Lie algebra and
its Lie group, one recovers in this setting the relation (1). This exponential mapping always exists
in the case of finite-dimensional Lie groups, and in the infinite-dimensional case its existence is
ensured by requiring regularity of the Lie group.
On trace-class operators over a separable Hilbert space one can promote the trace on matrices
to an operator trace. Further generalizing to classical pseudodifferential operators one can
consider traces on such operators. In the case of a closed manifold of dimension greater than
one, M. Wodzicki proved that there is a unique trace (up to a constant factor) on the whole
algebra of classical pseudodifferential operators acting on smooth functions on the manifold,
namely the noncommutative residue [31]. As S. Paycha and S. Rosenberg pointed out [23],
this fact does not rule out the existence of other traces when restricting to subalgebras of such
operators. In fact, other traces such as the leading symbol trace, the operator trace and the
canonical trace appear naturally on appropriate subalgebras. The classification of the traces
on algebras of classical pseudodifferential operators of non-positive order has been carried out
in [17] (see also [12]).
2
C. Neira Jim´enez and M.F. Ouedraogo
After the construction of the canonical trace on non-integer order pseudodifferential oper-
ators, M. Kontsevich and S. Vishik [8] introduced the set of odd-class operators, which is an
algebra that contains the differential operators. They also defined a trace on this algebra when
the dimension of the manifold is odd, and it has been proven that this is the unique trace in
this context [14, 22, 26]. In [21], M.F. Ouedraogo gave another proof of this fact based on the
expression of a symbol of such an operator as a sum of derivatives of symbols corresponding to
appropriate operators on the same algebra.
Odd-class operators are one of the rare types of operators in odd dimensions on which the
canonical trace is defined, this fact serving as a motivation to investigate the classification of
traces on the algebra of odd-class operators of order zero. The noncommutative residue is not of
interest here since it vanishes on odd-class operators (Lemma 2). This contrasts with the algebra
of ordinary zero-order operators where the only traces are linear combinations of the leading
symbol trace and the noncommutative residue (see [13]). We supplement that classification of
traces, showing that when restricting to odd-class zero-order operators, the only traces turn
out to be linear combinations of the leading symbol trace and the canonical trace (this is the
particular case a = 0 in Theorem 3).
In this article we present the classification of traces on algebras of odd-class pseudodifferential
operators acting on smooth functions on a closed odd-dimensional manifold. The methods we
implement combine various approaches used in the literature on the classification of traces.
However, a detailed analysis is required here because of the specificity of odd-class operators
(see Proposition 3). We recall the one to one correspondence between continuous traces and
determinants of class C1 on regular Lie groups1, and as in [13] we use this correspondence to
give the classification of determinants on the Lie group associated to the algebra of odd-class
operators of a fixed non-positive order. At the end we discuss two ways to extend the definition
of determinants outside a neighborhood of the identity.
In Section 2 we recall some of the basic notions of classical pseudodifferential operators,
including that of symbols on an open subset of the Euclidean space and odd-class operators on
a closed manifold. Inspired by [14], we use the representation of an odd-class symbol as a sum
of derivatives up to a smoothing symbol (Proposition 1), to express an odd-class operator in
terms of commutators of odd-class operators (Proposition 3), a fact that helps considerably in
the classification of traces.
In Section 3 we give a classification of traces on odd-class operators of non-positive order. For
that we recall the known traces on classical pseudodifferential operators. The noncommutative
residue vanishes on the algebra of odd-class operators in odd dimensions, whereas the canonical
trace is the unique linear form on this set which vanishes on commutators of elements in the
algebra (see [14]). Then using the fact that any odd-class operator can be expressed in terms
of commutators of odd-class operators (Proposition 3), we prove that any trace on an algebra
of odd-class operators of fixed non-positive order can be expressed as a linear combination of
a generalized leading symbol trace and the canonical trace (Theorem 3).
In Section 4 we classify determinants on the Lie groups associated to the algebras of odd-
class operators of non-positive order. We follow some of the work done in [13], concerning the
one to one correspondence between continuous traces on Lie algebras and C1-determinants on
the associated regular Lie groups (this is also discussed in [2] in specific situations). Then, we
combine this correspondence with the classification of traces given in Section 3.3, to provide
the classification of determinants on Lie groups associated to algebras of odd-class operators of
fixed non-positive order (Theorem 4). This classification is carried out for operators in a small
neighborhood of the identity, where the exponential mapping is a diffeomorphism.
At the end of this section, we give two possible ways to extend the definition of a determinant
outside a neighborhood of the identity on the Lie group associated to the algebra of odd-
1Special instances of this one to one correspondence were discussed by P. de la Harpe and G. Skandalis in [2].
Traces and Determinants on Odd-Class Operators
3
class pseudodifferential operators of order zero; the first one (see (22)) using a spectral cut
to define the logarithm of an admissible operator; in this case, for some traces this definition
of determinant depends on the spectral cut; the other one (see (24)) via the definition of the
determinant of an element on the pathwise connected component of the identity, using a path
that connects the element with the identity. Both the first and the second extension of the
definition of a determinant, provide maps which do not depend on the spectral cut and which
satisfy the multiplicativity property, under the condition that the image of the fundamental
group of invertible odd-class pseudodifferential operators of order zero is trivial.
2 Preliminaries on pseudodif ferential operators
Here we recall the basic notions of classical pseudodifferential operators following [29].
2.1 Symbols
Let U be an open subset of Rn. Given a ∈ C, a symbol of order a on U is a complex valued
function σ(x, ξ) in C∞(U×Rn) such that for any compact subset K of U and any two multiindices
α = (α1, . . . , αn), β = (β1, . . . , βn) ∈ Nn there exists a constant CK,α,β satisfying for all (x, ξ) ∈
ξ σ(x, ξ)(cid:12)(cid:12) ≤ CK,α,β(1 + ξ)(cid:60)(a)−β,
K × Rn,(cid:12)(cid:12)∂α
Notice that if (cid:60)(a1) < (cid:60)(a2), then Sa1(U ) ⊂ Sa2(U ). We denote by S−∞(U ) := (cid:84)
xn , β = β1 + ··· + βn, and (cid:60)(a) stands for the real part of a. Let Sa(U )
Sa(U )
the space of smoothing symbols on U . Given σ ∈ Sm0(U ), σj ∈ Smj (U ), where mj → −∞ as
j → ∞, we write
a∈C
x ∂β
where ∂α
denote the set of such symbols.
x = ∂α1
x1 ··· ∂αn
∞(cid:88)
j=0
σ ∼
σj,
if for every N ∈ N
σj ∈ SmN (U ).
σ − N−1(cid:88)
σ1 (cid:63) σ2(x, ξ) ∼ (cid:88)
j=0
α≥0
The product (cid:63) on symbols is defined as follows: if σ1 ∈ Sa1(U ) and σ2 ∈ Sa2(U ),
(−i)α
α!
∂α
ξ σ1(x, ξ)∂α
x σ2(x, ξ).
(2)
In particular, σ1 (cid:63) σ2 ∈ Sa1+a2(U ).
2.1.1 Classical symbols
A symbol σ ∈ Sa(U ) is called classical, and we write σ ∈ CSa(U ), if there is an asymptotic
expansion
ψ(ξ) σa−j(x, ξ).
(3)
σ(x, ξ) ∼
∞(cid:88)
j=0
4
Here σa−j(x, ξ) is a positively homogeneous function in ξ of degree a − j:
C. Neira Jim´enez and M.F. Ouedraogo
σa−j(x, tξ) = ta−jσa−j(x, ξ)
for all t ∈ R+,
ξ (cid:54)= 0,
and ψ ∈ C∞(Rn) is any cut-off function which vanishes for ξ ≤ 1
ξ ≥ 1.
2 and such that ψ(ξ) = 1 for
We denote by
(cid:43)
(cid:42)(cid:91)
a∈C
CS(U ) =
CSa(U )
the algebra generated by all classical symbols on U for the product (cid:63).
2.1.2 Odd-class symbols
The homogeneous components in the asymptotic expansion of a classical symbol may satisfy
some other symmetry relations additional to the positive homogeneity on the second variable.
Now we recall the definition of odd-class symbols introduced first in [8] (see also [5]).
Definition 1 (see [8]). A classical symbol σ ∈ CSa(U ) with integer order a is odd-class if for
each j ≥ 0, the term σa−j in the asymptotic expansion (3) satisfies
σa−j(x,−ξ) = (−1)a−jσa−j(x, ξ)
ξ ≥ 1.
for
(4)
In other words, odd-class symbols have the parity one would expect from differential operators.
Let us denote by CSa
CSodd(U ) =
odd(U ) the set of odd-class symbols of order a ∈ Z on U . We set
CSa
odd(U ).
(cid:91)
a∈Z
Lemma 1 (see [3, 8]). The odd-class symbols satisfy the following:
1. The product (cid:63) of two odd-class symbols is an odd-class symbol, therefore CSodd(U ) is an
algebra.
2. If an odd-class symbol is invertible with respect to the product (cid:63), then its inverse is an
odd-class symbol.
2.1.3 The noncommutative residue on symbols
As before, we consider U an open subset of Rn.
Definition 2 (see [6, 31]). The noncommutative residue of a classical symbol σ ∼ ∞(cid:80)
σa−j ∈
CSa(U ) is defined by
(cid:90)
(cid:90)
U
S∗
xU
res(σ) :=
σ−n(x, ξ)µ(ξ) dx =:
j=0
resx(σ)dx,
(cid:90)
U
where µ is the surface measure on the unit sphere S∗
xU over x in the cotangent bundle T ∗U .
The noncommutative residue clearly vanishes on symbols of order strictly less than −n and
also on symbols of non-integer order.
Lemma 2. In odd dimensions, the noncommutative residue of any odd-class symbol vanishes.
Traces and Determinants on Odd-Class Operators
Proof . Let σ ∈ CSa
odd, we have σ−n(x,−ξ) = (−1)nσ−n(x, ξ) = −σ−n(x, ξ). Then we obtain for any x ∈ U
odd(U ) be with asymptotic expansion σ ∼ ∞(cid:80)
(cid:90)
(cid:90)
(cid:90)
j=0
σ−n(x, ξ)µ(ξ) = −
σ−n(x,−ξ)µ(ξ) = −
σ−n(x, ξ)µ(ξ) = − resx(σ).
ψσa−j as in (3). Since n is
5
S∗
xU
S∗
xU
resx(σ) =
S∗
xU
Therefore res(σ) = 0.
(cid:4)
2.1.4 An odd-class symbol as a sum of derivatives
Proposition 1 (see Lemma 1.3 in [4], and [14]). Let n ∈ Z be odd. For any σ ∈ CSa
there exist τi in CSa+1
odd (U ) such that
odd(U ),
σ ∼ n(cid:88)
∞(cid:88)
σ ∼
i=1
j=0
∂ξiτi.
ψσa−j,
Proof . For a cut-off function ψ as in Section 2.1.1 consider
with σa−j a positively homogeneous function of degree a − j in ξ which satisfies (4).
If a − j (cid:54)= −n, consider the homogeneous function τi,a−j+1 := ξiσa−j (x,ξ)
a−j+n . By Euler's identity
we have
n(cid:88)
∂ξi(τi,a−j+1)(x, ξ) = σa−j(x, ξ).
i=1
The homogeneous functions τi,a−j+1 clearly satisfy (4) for ξ ≥ 1:
τi,a−j+1(x, tξ) = ta−j+1τi,a−j+1(x, ξ),
τi,a−j+1(x,−ξ) = (−1)a−j+1τi,a−j+1(x, ξ).
∀ t > 0,
Let a − j = −n. In polar coordinates (r, ω) ∈ R+ × Sn−1, the Laplacian in ξ reads
∂2
ξi
= −r1−n∂r
∆ = − n(cid:88)
(cid:0)rn−1∂r
∆(cid:0)f (ω)r2−n(cid:1) = r−n∆Sn−1f (ω).
i=1
Therefore, for any function f ∈ C∞(Sn−1),
(cid:1) − r−2∆Sn−1.
Since n is odd and σ ∈ CSa
odd(U ), by Lemma 2 we have res(σ) = 0. Therefore σ−n(x,·) (cid:22)Sn−1 is
orthogonal to the constants which form the kernel ker(∆Sn−1). Hence there exists a unique func-
tion h(x,·) ∈ C∞(Sn−1), orthogonal to the constants, such that ∆Sn−1(h(x,·)) = σ−n(x,·) (cid:22)Sn−1.
The function h(x,−ξ) + h(x, ξ) is constant and orthogonal to the constants, therefore, h(x,·) is
an odd function on Sn−1.
We choose a smooth function χ on R which vanishes for small r and is equal to 1 for r ≥ 1/2.
For r = ξ, we set
b−n(x, ξ) := χ(ξ)ξ2−nh
(cid:18)
x,
ξ
ξ
(cid:19)
.
6
C. Neira Jim´enez and M.F. Ouedraogo
The function b−n is smooth on U × Rn and is homogeneous of degree −n + 2 in ξ for ξ ≥ 1. As
σ−n(x, ξ) vanishes for x outside a compact set, so does b−n(x, ξ). In particular, b−n is a symbol
of order −n + 2 on U . Let us define τi,−n+1 := −∂ξib−n. Since h is odd so is b−n and therefore,
τi,−n+1(x,−ξ) = −(∂ξib−n)(x,−ξ) = −∂ξib−n(x, ξ) = (−1)−n+1τi,−n+1(x, ξ).
Moreover, we have for r = ξ ≥ 1
∆b−n(x,·) = ∆(cid:0)r2−nh(x,·)(cid:1) = r−n (σ−n(x,·) (cid:22)Sn−1) = σ−n(x,·).
Let τi ∼ ∞(cid:80)
ψ τi,a−j+1, then since ∂ξiψ has compact support, the difference σ − n(cid:80)
∞(cid:88)
σ ∼ n(cid:88)
ψ∂ξiτi,a−j+1 ∼ n(cid:88)
j=0
smoothing and
∂ξiτi.
i=1
∂ξiτi is
(cid:4)
i=1
j=0
i=1
2.2 Pseudodif ferential operators
Let U ⊂ Rn be an open subset, and denote by C∞
c (U ) the space of smooth compactly supported
functions on U . To the symbol σ ∈ S(U ), we associate the linear integral operator Op(σ) :
c (U ) → C∞(U ) defined for u ∈ C∞
C∞
c (U ) by
Op(σ)(u)(x) =
ei(x−y)·ξσ(x, ξ)u(y)dyd¯ξ
eix·ξσ(x, ξ)(cid:98)u(ξ)d¯ξ =
(cid:90)
(cid:90)
T ∗
x U
U
(cid:90)
(cid:90)
T ∗
x U
U
=
k(x, y)u(y)dy,
U e−iy·ξu(y) dy is the Fourier transform of u and d¯ξ := (2π)−ndξ. In this expres-
where(cid:98)u(ξ) =(cid:82)
sion k(x, y) =(cid:82) ei(x−y)·ξσ(x, ξ)d¯ξ is seen as a distribution on U × U that is smooth outside the
σ ∼ ∞(cid:80)
diagonal. We say that Op(σ) is a pseudodifferential operator (ψDO) with Schwartz kernel given
by k(x, y). An operator is smoothing if its Schwartz kernel is a smooth function on U × U . If
ψσa−j is a classical symbol of order a, then A = Op(σ) is called a classical pseudodif-
j=0
ferential operator of order a. The homogeneous component σa of σ is called the leading symbol
of A, and will be denoted by σL
A.
A ψDO A on U is called properly supported if for any compact K ⊂ U , the set {(x, y) ∈
supp(kA) : x ∈ K or y ∈ K} is compact, where supp(kA) denotes the support of the Schwartz
kernel of A. Any ψDO A can be written in the form (see [29])
A = P + R,
(5)
where P is properly supported and R is a smoothing operator.
A properly supported ψDO maps C∞
c (U ) into itself. The product (cid:63) on symbols defined in (2)
induces a composition on properly supported ψDOs on U . The composition AB of two properly
supported ψDOs A and B is a properly supported ψDO with symbol σ(AB) = σ(A) (cid:63) σ(B).
The notion of a ψDO can be extended to operators acting on manifolds (see Section 4.3
in [29]). Let M be a smooth closed manifold of dimension n. A linear operator A : C∞(M ) →
C∞(M ) is a pseudodifferential operator of order a on M if in any atlas, A is locally a pseu-
dodifferential operator. This means that given a local coordinate chart U of M , with diffeomor-
phism ϕ : U → V , from U to an open set V ⊆ Rn, the operator ϕ#A defined by the following
Traces and Determinants on Odd-Class Operators
7
diagram is a pseudodifferential operator of order a on V :
C∞
c (V )
ϕ#A
............................................................................................................................................................................... ............
C∞(V )
ϕ∗
..........................................................................................................................................
............
ϕ∗
..........................................................................................................................................
............
(cid:43)
rU ◦ A ◦ iU
............................................................................................................................................................................... ............
C∞(U )
c (U ) → C∞
C∞
c (U )
where iU : C∞
natural restriction.
Let C(cid:96)a(M ) denote the set of classical ψDOs of order a on M , i.e. operators whose symbol
is classical of order a in any local chart of M . If A1 ∈ C(cid:96)a1(M ), A2 ∈ C(cid:96)a2(M ), then A1A2 ∈
C(cid:96)a1+a2(M ), thus, the space C(cid:96)a(M ) is an algebra if and only if a is an integer and a ≤ 0. We
denote by
c (M ) is the natural embedding, and rU : C∞(M ) → C∞(U ) is the
(cid:42)(cid:91)
(cid:92)
a∈C
a∈C
C(cid:96)(M ) :=
C(cid:96)a(M )
the algebra generated by all classical ψDOs on M , and by
C(cid:96)−∞(M ) :=
C(cid:96)a(M )
the ideal of smoothing operators in C(cid:96)(M ).
We will also denote by
C(cid:96) /∈Z
(M ) :=
C(cid:96)a(M )
a /∈Z∩[−n,+∞)
(cid:42) (cid:91)
(cid:43)
the space generated by classical ψDOs on M whose order is non-integer or less than −n.
A classical operator A ∈ C(cid:96)a(M ) of integer order a is odd-class if in any local chart its local
odd(M ) the set of odd-class operators of order a ∈ Z
symbol σ(A) is odd-class. We denote by C(cid:96)a
and we define
(cid:91)
a∈Z
C(cid:96)odd(M ) =
C(cid:96)a
odd(M ).
As in Lemma 1, the following lemma implies that C(cid:96)odd(M ) is an algebra:
Lemma 3 (see Section 4 in [8]). Let A ∈ C(cid:96)a
AB ∈ C(cid:96)a+b
AB−1 ∈ C(cid:96)a−b
odd (M ). If moreover B is an invertible elliptic operator, then B−1 ∈ C(cid:96)−b
odd (M ).
odd(M ) and B ∈ C(cid:96)b
odd(M ), a, b ∈ Z. Then
odd(M ) and
The algebra C(cid:96)odd(M ) contains the differential operators and their parametrices.
Remark 1. Even though the definition of odd-class pseudodifferential operators makes sense
on any closed manifold, in this paper we restrict ourselves to odd-dimensional closed manifolds.
The reason is that the canonical trace (which will be explained below in Section 3.1.2) is well
defined only in that case. So, from now on, the notation C(cid:96)odd(M ) will be used only when the
dimension n of the manifold M is odd.
8
C. Neira Jim´enez and M.F. Ouedraogo
2.2.1 Fr´echet topology on pseudodif ferential operators
For any complex number a, we equip the vector space C(cid:96)a(M ) with a Fr´echet topology as
follows. Let us consider a covering of M by open neighborhoods {Ui}i∈I , a finite subordinated
partition of unity (χi)i∈I and smooth functions ((cid:101)χi)i∈I on M such that supp((cid:101)χi) ⊂ Ui and
(cid:101)χi = 1 near the support of χi. By (5) any ψDO A can be written in the form A =(cid:80)
where the operators Ai := χi · Op(σi) · (cid:101)χi ∈ Cla(M ) are properly supported in Ui with symbols
σ(i)(A) ∼ ∞(cid:80)
a−j(A), and Ri is a smoothing operator with smooth kernel ki which has
(Ai + Ri)
ψσ(i)
i∈I
We equip C(cid:96)a(M ) with the following countable set of semi-norms: for any compact subset
j=0
compact support in Ui × Ui.
K ⊂ Ui for any j ≥ 0, N ≥ 1 and for any multiindices α, β
x ∂β
ξ σ(i)(A)(x, ξ)(cid:12)(cid:12),
(1 + ξ)β−a(cid:12)(cid:12)∂α
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)∂α
σ(i)(A) − N−1(cid:88)
(cid:12)(cid:12)∂α
a−j(A)(x, ξ)(cid:12)(cid:12),
(1 + ξ)β−a+N
(cid:12)(cid:12)∂α
j=0
x ∂β
ξ σ(i)
x ∂β
x ∂β
ξ
sup
x,y∈K
sup
x∈K
sup
ξ∈Rn
sup
x∈K
sup
ξ∈Rn
sup
x∈K
sup
ξ=1
a−j(A)
χ(ξ)σ(i)
y ki(x, y)(cid:12)(cid:12).
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ,
(x, ξ)
2.2.2 The logarithm of a classical pseudodif ferential operator
An operator A ∈ C(cid:96)(M ) with positive order has principal angle θ if for every (x, ξ) ∈ T ∗M \{0},
A(x, ξ) has no eigenvalues on the ray Lθ = {reiθ, r ≥ 0}; in that case A is
the leading symbol σL
elliptic.
Definition 3 (see e.g. [18]). An operator A ∈ C(cid:96)(M ) is admissible with spectral cut θ if A has
principal angle θ and the spectrum of A does not meet Lθ = {reiθ, r ≥ 0}. In particular such
an operator is invertible and elliptic. The angle θ is called an Agmon angle of A.
Let A ∈ C(cid:96)(M ) be admissible with spectral cut θ and positive order a. For (cid:60)(z) < 0, the
complex power Az
θ of A is defined by the Cauchy integral (see [28])
(cid:90)
Az
θ =
i
2π
Γr,θ
θ(A − λ)−1 dλ,
λz
θ = λzeiz(argλ) with θ ≤ arg λ < θ + 2π. Here Γr,θ is a contour along the ray Lθ around
where λz
the (non-zero) spectrum of A, and r is any small positive real number such that Γr,θ does not
meet the spectrum of A. The operator Az
θ is an elliptic classical ψDO of order az; in particular,
for z = 0, we have A0
The definition of complex powers can be extended to the whole complex plane by setting
θ := AkAz−k
for k ∈ N and (cid:60)(z) < k; this definition is independent of the choice of k in N
Az
θ = Ak, for k ∈ Z. Complex powers
and preserves the usual properties, i.e. Az1
of operators depend on the choice of spectral cut. Indeed, if Lθ and Lφ are two spectral cuts
for A outside an angle which contains the spectrum of σL(A)(x, ξ) then
θ = Az1+z2
θ Az2
θ = I.
, Ak
θ
θ
φ =(cid:0)1 − e2iπz(cid:1)Pθ,φ(A)Az
θ,
θ − Az
Az
where the operator
Pθ,φ(A) =
1
2iπ
(cid:90)
Γθ,φ
λ−1A(A − λ)−1 dλ
(6)
(7)
Traces and Determinants on Odd-Class Operators
9
is a projection (see [25, 32]). Here Γθ,φ is a contour separating the part of the spectrum of A
contained in the open sector θ < arg λ < φ from the rest of the spectrum.
The logarithm of an admissible operator A with spectral cut θ is defined in terms of the
derivative at z = 0 of its complex power:
logθ(A) = ∂zAz
θz=0
.
Logarithms of classical ψDOs of positive order are not classical anymore since their symbols
involve a logarithmic term log ξ as the following elementary result shows.
Proposition 2 ([18]). Let A ∈ C(cid:96)a(M, E) be an admissible operator with positive order and
spectral cut θ. In a local trivialization, the symbol of logθ(A) reads:
σlogθ(A)(x, ξ) = a log ξI + σA
0 is a classical symbol of order zero.
0 (x, ξ),
where σA
Remark 2. If A is a classical ψDO of order zero then A is bounded, and if it admits a spectral
cut, then complex powers and the logarithm of A are directly defined using a Cauchy integral
formula, and they are classical ψDOs (see [18] and Remark 2.1.7 in [21]).
Just as complex powers, the logarithm depends on the choice of spectral cut. Indeed, given
two spectral cuts θ, φ of the operator A such that 0 ≤ θ < φ < 2π, differentiation of (6) with
respect to z and evaluation at z = 0 yields
logθ A − logφ A = −2iπPθ,φ(A).
(8)
2.2.3 Pseudodif ferential operators in terms of commutators
In this subsection we use the ψDO analysis techniques similar to the ones implemented in [14]; we
assume that M is an n-dimensional closed manifold and n is odd. Given a function f ∈ C∞(M ),
we also denote by f the zero-order classical ψDO given by multiplication by f .
Proposition 3. If A ∈ C(cid:96)a
C(cid:96)a+1
odd(M ), then there exist functions αk ∈ C∞(M ), operators Bk in
odd (M ) and a smoothing operator RA such that
A =
[αk, Bk] + RA.
(9)
Proof . Let us consider a covering of M by open neighborhoods {Uj}N
nated partition of unity {ϕj}N
in one coordinate neighborhood. We write (see Subsection 2.2.1)
j=1 and a finite subordi-
j=1, such that for every pair (j, k), both ϕj and ϕk have support
N(cid:88)
k=1
(cid:88)
j,k
A =
ϕjAϕk + R.
Each operator ϕjAϕk may be considered as an odd-class ψDO on Rn with symbol σ in CSa
By Proposition 1, there exist odd-class symbols τl of order a + 1 such that
odd(Rn).
σ ∼ n(cid:88)
∂ξlτl.
l=1
For any symbol τ we have,
σ([xl, Op(τ )]) ∼ xl · τ − τ · xl − i−1∂ξlτ = i∂ξlτ,
10
so that
C. Neira Jim´enez and M.F. Ouedraogo
Since σ ∼ n(cid:80)
l=1
Op(∂ξlτ ) = −i[xl, Op(τ )] up to a smoothing operator.
∂ξlτl, there exist smoothing operators R(cid:48)(cid:48), R(cid:48) such that
(cid:32) n(cid:88)
(cid:33)
n(cid:88)
Op(σ) = Op
∂ξlτl
+ R(cid:48)(cid:48) = −i
[xl, Op(τl)] + R(cid:48).
l=1
l=1
c (Uj) be such that ψj ≡ 1 near supp(ϕj). Then for χ in C∞
c (Rn)
ϕj[xl, Op(τl)]ψj + ϕjR(cid:48)ψj = −i
[χxl, ϕj Op(τl)ψj] + ϕjR(cid:48)ψj.
For each index j let ψj ∈ C∞
such that χϕj = ϕj, χψj = ψj, we have
n(cid:88)
j Op(σj) +(cid:80)
tor Rj. We have A =(cid:80)
ϕj Op(σ)ψj = −i
l=1
N(cid:88)
k=1
A =
[αk, Bk] + RA,
n(cid:88)
l=1
As in (5) we write ϕjAψj = Op(σj) + Rj for some σj ∈ CSa
odd(Rn) and some smoothing opera-
j Rj + R, hence A can be written in the form
where αk is a smooth function on M (and represents the operator in C(cid:96)0
by αk), Bk lies in C(cid:96)a+1
Corollary 1. If A ∈ C(cid:96)a
and a smoothing operator R such that
odd (M ), and RA is a smoothing operator.
odd(M ) of multiplication
(cid:4)
odd(M ), smooth functions ai ∈ C∞(M ),
A − Op(cid:0)σL
(cid:1) =
A
odd(M ) then there exist Bi ∈ C(cid:96)a
n(cid:88)
[ai, Bi] + R.
i=1
Proof . It follows by applying Proposition 3 to A − Op(σL
A) ∈ C(cid:96)a−1
odd (M ).
(cid:4)
3 Classif ication of traces on odd-class operators
The classification made in this section is essentially based on Proposition 3, namely the decom-
position of an odd-class operator in terms of commutators of odd-class operators. Let us first
recall the definition of a trace.
Let M be a closed connected manifold of dimension n > 1, and let A ⊆ C(cid:96)(M ). A trace
on A is a map
τ : A → C,
linear in the sense that for all a, b ∈ C, whenever A, B and aA + bB belong to A we have
τ (aA + bB) = aτ (A) + bτ (B),
and such that for any A, B ∈ A, whenever AB, BA ∈ A it satisfies
τ ([A, B]) = 0,
or equivalently,
τ (AB) = τ (BA).
Traces and Determinants on Odd-Class Operators
11
3.1 Examples of traces on pseudodif ferential operators
Interestingly by Lemma 2, the noncommutative residue, which is the only trace on the whole
space C(cid:96)(M ) (see [4, 10, 31]), vanishes on odd-class operators when the dimension of the manifold
is odd. In this section we review the traces which are non-trivial on this class of operators.
3.1.1 The L2-trace
A ψDO A whose order has real part less than −n is a trace-class operator. The L2-trace (also
called operator trace) is the functional
Tr :
C(cid:96)a(M )
→ C,
A (cid:55)→ Tr(A) :=
kA(x, x) dx,
(10)
(cid:42) (cid:91)
(cid:60)(a)<−n
(cid:43)
(cid:90)
M
where kA is the Schwartz kernel of the operator A. This trace is continuous for the Fr´echet
topology on the space of ψDOs of constant order less than −n.
This is the unique trace on the algebra of smoothing operators C(cid:96)−∞(M ), since we have the
exact sequence (see [7])
0 →(cid:2)C(cid:96)−∞(M ), C(cid:96)−∞(M )(cid:3) → C(cid:96)−∞(M ) Tr−→ C → 0.
More precisely we have
Theorem 1 (Theorem A.1 in [7]). If R is a smoothing operator then, for any pseudodifferential
idempotent J, of rank 1, there exist smoothing operators S1, . . . , SN , T1, . . . , TN , such that
N(cid:88)
R = Tr(R)J +
[Sj, Tj].
j=1
Therefore, any smoothing operator with vanishing L2-trace is a sum of commutators in the space
[C(cid:96)−∞(M ), C(cid:96)−∞(M )].
Proposition 4 (see e.g. Introduction in [10] and Proposition 4.4 in [11]). The trace Tr does
not extend to a trace functional neither on the whole algebra C(cid:96)(M ), nor does it on the alge-
bra C(cid:96)0(M ).
3.1.2 The canonical trace
We start with the definition of the cut-off integral of a symbol, as in [22]. Let U be an open
subset of Rn.
Proposition 5 (see [22] and Section 1.2 in [24]). Let σ ∈ CSa(U ), such that for any N ∈ N,
ψ(ξ)σa−j(x, ξ) + σN (x, ξ), with σa−j, ψ and σN ∈
N−1(cid:80)
σ can be written in the form σ(x, ξ) =
Sa−N (U ) as in (3). The integral of σ over B∗
the cotangent space T ∗
j=0
x U ) has the asymptotic expansion
x(0, R) (which stands for the ball of radius R in
αj(σ)(x)Ra−j+n + resx(σ) log R + αx(σ),
(cid:90)
N−1(cid:88)
j=0
σ(x, ξ)dξ ∼
R→∞
B∗
x(0,R)
a−j+n(cid:54)=0
where αx(σ) converges when R → ∞.
(cid:90)
(cid:90)
(cid:90)
12
C. Neira Jim´enez and M.F. Ouedraogo
Proof . We write
ψ(ξ)σa−j(x, ξ)dξ =
B∗
x(0,R)
(cid:90)
B∗
x(0,1)
ψ(ξ)σa−j(x, ξ)dξ +
(cid:90)
x(0,R)\B∗
B∗
x(0,1)
ψ(ξ)σa−j(x, ξ)dξ.
Using the fact that σa−j is positively homogeneous of degree a − j we have
(cid:90) R
(cid:90)
1
ω=1
ψ(ξ)σa−j(x, ξ)dξ =
ra−j+n−1σa−j(x, ω) dωdr
x(0,R)\B∗
B∗
x(0,1)
1
Ra−j+n
=
a − j + n
It follows that the integral(cid:82)
σa−j(x, ω)dω −
1
a − j + n
ω=1
ω=1
σa−j(x, ω) dω.
x(0,R) σ(x, ξ) dξ admits the asymptotic expansion
B∗
(cid:90)
(cid:90)
(cid:90)
N−1(cid:88)
σ(x, ξ) dξ ∼
R→∞
B∗
x(0,R)
1
a − j + n
Ra−j+n
ω=1
σa−j(x, ω) dω
j=0
a−j+n(cid:54)=0
+ resx(σ) log R + αx(σ),
(11)
(cid:4)
where αx(σ) is given by
αx(σ) :=
σN (x, ξ) dξ +
(cid:90)
− N−1(cid:88)
B∗
x(0,R)
j=0
a−j+n(cid:54)=0
(cid:90)
N−1(cid:88)
B∗
x(0,1)
j=0
ψ(ξ)σa−j(x, ξ) dξ
(cid:90)
1
a − j + n
ω=1
σa−j(x, ω) dω,
which may depend on the variable x.
For N sufficiently large, the integral (cid:82)
(cid:90)
N−1(cid:88)
(cid:90)
of N > a + n − 1. Moreover, if a < −n, then −(cid:82)
(cid:90)
− N−1(cid:88)
a − j + n
where S∗
σN (x, ξ) dξ +
σ(x, ξ) dξ :=
j=0
a−j+n(cid:54)=0
T ∗
x U
T ∗
x U
−
1
(cid:90)
ψ(ξ)σa−j(x, ξ) dξ
B∗
x(0,1)
j=0
σa−j(x, ω) dω,
S∗
xU
Rnσ(x, ξ)dξ =(cid:82)
x(0,R) σN (x, ξ)dξ is well defined, so the term αx(σ)
B∗
given by (11) converges when R → ∞, and taking this limit we define the cut-off integral of σ
by
xU stands for the unit sphere in the cotangent space T ∗
x U . This definition is independent
According to Proposition 4, there is no non-trivial trace on C(cid:96)(M ) which extends the L2-trace.
However, the L2-trace does extend to non-integer order operators and to odd-class operators.
Indeed, M. Kontsevich and S. Vishik [8] constructed such an extension, the canonical trace
TR : C(cid:96) /∈Z
(M )
C(cid:96)odd(M ) → C,
A (cid:55)→ TR(A) :=
1
(2π)n
M
σ(A)(x, ξ)dξ,
T ∗
x M
where the right hand side is defined using a finite covering of M , a partition of unity subordinated
to it and the local representation of the symbol, but this definition is independent of such choices.
(cid:91)
Rn σ(x, ξ)dξ.
(cid:90)
(cid:90)
dx −
Traces and Determinants on Odd-Class Operators
13
As we already stated in Remark 1, the canonical trace is well defined on C(cid:96)odd(M ) only when
the dimension n of the manifold is odd (see [5, 10]), which is always our case.
If A ∈ C(cid:96)a(M ), B ∈ C(cid:96)b(M ) and if a, b /∈ Z, then ord(AB) = a + b may be an integer, so the
linear space C(cid:96) /∈Z
(M ) is not an algebra; in spite of this, the canonical trace has the following
properties (see [8], Section 5 in [10], and [20, 22, 24]):
1. For any c ∈ C, if A, B ∈ C(cid:96) /∈Z
(M ) are such that ord(cA + B) /∈ Z ∩ [−n, +∞), or if
A, B ∈ C(cid:96)odd(M ), then TR(cA + B) = c TR(A) + TR(B).
2. For any A ∈ C(cid:96)(M ) such that ord(A) < −n, TR(A) = Tr(A), i.e. the canonical trace
extends the L2-trace defined in (10).
3. If A, B ∈ C(cid:96) /∈Z
(M ) are such that AB ∈ C(cid:96) /∈Z
(M ), or if A, B ∈ C(cid:96)odd(M ), then TR(AB) =
TR(BA).
(M )(cid:83) C(cid:96)odd(M ).
order in C(cid:96) /∈Z
The canonical trace is continuous for the Fr´echet topology on the sets of ψDOs of constant
3.1.3 Generalized leading symbol traces
In [23], S. Paycha and S. Rosenberg introduced the leading symbol traces defined on an algebra
of operators C(cid:96)a(M ) for a ≤ 0; in this section we follow [17] and consider a trace which actually
coincides with a leading symbol trace when a = 0. Let a be a non-positive integer and consider
the projection map πa from C(cid:96)a(M ) to the quotient space C(cid:96)a(M )/C(cid:96)2a−1(M ):
0 → C(cid:96)2a−1(M ) → C(cid:96)a(M ) πa→ C(cid:96)a(M )/C(cid:96)2a−1(M ) → 0.
(12)
One can see in Remark 4.3.2 of [17] that it is possible to construct a splitting
θa : C(cid:96)a(M )/C(cid:96)2a−1(M ) → C(cid:96)a(M ).
Lemma 4. Any continuous linear map ρ on C(cid:96)a(M )/C(cid:96)2a−1(M ) defines a trace on C(cid:96)a(M ) by
ρ ◦ πa called generalized leading symbol trace.
Proof . If A, B ∈ C(cid:96)a(M ), their commutator [A, B] belongs to C(cid:96)2a−1(M ), and since ρ ◦ πa
vanishes on C(cid:96)2a−1(M ), it defines a trace on C(cid:96)a(M ).
(cid:4)
For A ∈ C(cid:96)a(M ), ρ(πa(A)) depends on σa(A), . . . , σ2a(A), where σa−i(A) represents the
homogeneous component of degree a − i in the asymptotic expansion of the symbol of A. Since
ρ ◦ πa is linear in A, it is a linear combination of linear maps ρa−i on C∞(S∗M ), in the terms
σa−i(A), hence it reads,
a(cid:88)
ρ(πa(A)) =
ρa−i(σa−i(A)).
i=0
Remark 3. For a < 0, a leading symbol trace is the particular case of a generalized leading
symbol trace when ρa−i ≡ 0 for all i = 1, . . . ,a.
Generalized leading symbol traces are continuous for the Fr´echet topology on the space of
constant order ψDOs, since they are defined in terms of a finite number of homogeneous com-
ponents of the symbols of the operators.
14
C. Neira Jim´enez and M.F. Ouedraogo
3.2 Trace on C(cid:96)odd(M )
The canonical trace is the unique trace (up to a constant) on its domain. This result was proved
in [14] (which goes back to 2007 but it was published in 2008) using the fact that the operator
corresponding to the derivative of a symbol is, up to a smoothing operator, proportional to
the commutator of appropriate operators. The latter idea was considered in [22] to show the
equivalence between Stokes' property for linear forms on symbols and the vanishing of linear
forms on commutators of operators. In [26] the author uses the Schwartz kernel representation
of an operator to express any non-integer order operator and any operator of regular parity class
as a sum of commutators up to a smoothing operator, and then to give another proof of the
uniqueness of the canonical trace. The following proof, that we find in Proposition 3.2.4 of [21],
is done in the spirit of [14].
Theorem 2. Any trace on C(cid:96)odd(M ) is proportional to the canonical trace.
Proof . By (9) any operator A in C(cid:96)odd(M ) can be written in the form
where αk are smooth functions on M that can be seen as elements of C(cid:96)0
C(cid:96)a+1
odd (M ), and RA is a smoothing operator. By Theorem 1, we can express RA as
odd(M ), Bk belong to
A =
[αk, Bk] + RA,
RA = Tr(RA)J +
[Sj, Tj],
N(cid:48)(cid:88)
j=1
N(cid:48)(cid:88)
N(cid:88)
k=1
N(cid:88)
N(cid:88)
k=1
where J is a pseudodifferential projection of rank 1 and Sj, Tj are smoothing operators. Sum-
ming up, the expression for A becomes
A =
[αk, Bk] + Tr(RA)J +
[Sj, Tj].
(13)
k=1
j=1
Applying the canonical trace TR to both sides of this expression, since TR vanishes on commu-
tators of operators in C(cid:96)odd(M ), we infer that
TR(A) = Tr(RA).
Thus (13) reads
A =
[αk, Bk] + TR(A)J +
N(cid:48)(cid:88)
j=1
[Sj, Tj].
(14)
If τ is a trace on C(cid:96)odd(M ), applying τ to both sides of (14) we reach the conclusion of the
(cid:4)
theorem.
3.3 Traces on C(cid:96)a
odd(M ) for a ≤ 0
In this section we assume as before, that the dimension n is odd, and we prove that any trace on
the algebra of odd-class operators of non-positive order is a linear combination of a generalized
leading symbol trace and the canonical trace.
We can adapt Lemma 4.5 in [12] (see also Lemma 5.1.1 in [17]) in the case of odd-class
operators:
Traces and Determinants on Odd-Class Operators
Lemma 5. If a ∈ Z is non-positive, then there exists an inclusion map
15
[C(cid:96)0
odd(M ), C(cid:96)2a
odd(M )] (cid:44)→ [C(cid:96)a
meaning that any commutator in [C(cid:96)0
in [C(cid:96)a
odd(M ), C(cid:96)a
odd(M )].
odd(M ), C(cid:96)a
odd(M )],
odd(M ), C(cid:96)2a
odd(M )] can be written as a sum of commutators
Proof . By Lemma 3, integer powers of an invertible differential operator are odd-class opera-
tors. Hence we proceed as in the proof of Lemma 4.5 in [12] as follows: Let A ∈ C(cid:96)0
odd(M ),
B ∈ C(cid:96)2a
odd(M ). Consider a first-order positive definite elliptic differential operator Λ. For any
a ∈ R, Λa and Λ−a are operators of order a and −a, respectively, and therefore AΛa, ΛaA, Λa,
BΛ−a, Λ−aB, ABΛ−a, Λ−aBA are operators in C(cid:96)a
odd(M ). Moreover,
[AΛa, Λ−aB] = AB − Λ−aBAΛa,
[ΛaA, BΛ−a] = ΛaABΛ−a − BA,
[ABΛ−a, Λa] = AB − ΛaABΛ−a,
[Λ−aBA, Λa] = Λ−aBAΛa − BA.
(15)
(16)
(17)
(18)
Adding up the expressions in (15) -- (18) yields twice the commutator [A, B], so that the resulting
(cid:4)
expression belongs to the space of commutators [Cla
odd(M ), Cla
odd(M )].
As in (12), for a non-positive integer a we also denote by πa the projection
0 → C(cid:96)2a−1
odd (M ) → C(cid:96)a
odd(M ) πa→ C(cid:96)a
odd(M )/C(cid:96)2a−1
odd (M ) → 0,
odd (M ) → C(cid:96)a
with corresponding splitting θa : C(cid:96)a
C(cid:96)a
odd(M ), A − θa(πa(A)) ∈ C(cid:96)2a−1
The following result adds to the known classification of traces on pseudodifferential ope-
rators, the classification of all traces on odd-class operators with fixed non-positive order in
odd-dimensions.
odd (M ).
odd(M ), so that for any A ∈
odd(M )/C(cid:96)2a−1
We fix a non-positive integer a, and describe any trace on C(cid:96)a
odd(M ) (see Section 5.1.4 in [17]).
Theorem 3. If a ∈ Z is non-positive, any trace on C(cid:96)a
bination of a generalized leading symbol trace and the canonical trace.
Proof . Let A ∈ C(cid:96)a
C(cid:96)2a−1
such that
odd (M ), there exist operators Bi ∈ C(cid:96)0
odd(M ). As in Corollary 1, by Proposition 3 applied to A − θa(πa(A)) ∈
odd(M ), and a smoothing operator R
n(cid:88)
odd(M ), Ci ∈ C(cid:96)2a
odd(M ) can be written as a linear com-
A − θa(πa(A)) =
[Bi, Ci] + R.
By Lemma 5, there exist operators D1, . . . , DN , E1, . . . , EN ∈ C(cid:96)a
odd(M ), such that
i=1
N(cid:88)
k=1
N(cid:88)
k=1
A − θa(πa(A)) =
[Dk, Ek] + R.
(19)
Applying TR to both sides of (19) yields
TR(A − θa(πa(A))) =
TR([Dk, Ek]) + TR(R) = TrL2(R).
16
C. Neira Jim´enez and M.F. Ouedraogo
Hence, as in Theorem 1, for any pseudodifferential idempotent J, of rank 1, there exist smoothing
operators S1, . . . , SN(cid:48), T1, . . . , TN(cid:48), such that (19) becomes
N(cid:88)
N(cid:48)(cid:88)
A − θa(πa(A)) =
[Dk, Ek] + TR(A − θa(πa(A)))J +
[Sj, Tj].
(20)
k=1
j=1
Let τ : C(cid:96)a
odd(M ) → C be a trace on C(cid:96)a
odd(M ). If we apply τ to both sides of (20) we get
τ (A) = τ (θa(πa(A))) + TR(A − θa(πa(A)))τ (J)
= τ (θa(πa(A))) − TR(θa(πa(A)))τ (J) + TR(A)τ (J).
So we conclude that τ is a linear combination of a generalized leading symbol trace and the
(cid:4)
canonical trace.
4 Determinants and traces
We use the classification of traces on algebras of odd-class operators given in Theorem 3 to
classify the associated determinants on the corresponding Fr´echet -- Lie group. Well-known ge-
neral results in the finite-dimensional context concerning determinants associated with traces
generalize to the context of Banach spaces (see [2]) and further to Fr´echet spaces (see [13]).
Definition 4 (Definition 36.8 in [9]). A (possibly infinite-dimensional) Lie group G with Lie
algebra Lie(G) admits an exponential mapping if there exists a smooth mapping Exp : Lie(G) →
G such that t (cid:55)→ Exp(tX) is a one-parameter subgroup, i.e. a Lie group homomorphism (R, +) →
G with tangent vector X at 0.
The existence of a smooth exponential mapping for a Lie group is ensured by a notion of
regularity [9, 16] on the group. Following [9], for J. Milnor [16], a Lie group G modelled on
a locally convex space is a regular Lie group if for each smooth curve u : [0, 1] → Lie(G), there
exists a smooth curve γu : [0, 1] → G (which is unique: Lemma 38.3 in [9]) which solves the
initial value problem γu = γu u with γu(0) = 1G, where 1G is the identity of G, with smooth
evolution map
C∞([0, 1], Lie(G)) → G,
u (cid:55)→ γu(1).
For example, Banach Lie groups (in particular finite-dimensional Lie groups) are regular. If E
is a Banach space, then the Banach Lie group of all bounded automorphisms of E is equipped
with an exponential mapping given by the series
∞(cid:88)
i=0
X i
i!
.
Exp(X) =
In [19] a wider concept of infinite-dimensional Lie groups called regular Fr´echet -- Lie groups
is introduced. In this paper we will consider the regular Fr´echet -- Lie groups of classical ψDOs
of non-positive order (see [33]).
If G admits an exponential mapping Exp and if a suitable inverse function theorem is appli-
cable, then Exp yields a diffeomorphism from a neighborhood of 0 in Lie(G) onto a neighborhood
of 1G in G, whose inverse is denoted by Log.
For our purpose in this section, we assume that the Lie group G is regular.
Traces and Determinants on Odd-Class Operators
17
Definition 5 (see Definition 2 in [13]). Let G be a Lie group and let G be its subgroup of
elements pathwise connected to the identity of G. A determinant on G is a group morphism
Λ : G → C∗,
i.e. for any g, h ∈ G,
Λ(gh) = Λ(g)Λ(h).
We also say that Λ is multiplicative.
A trace on the Lie algebra Lie(G) is a linear map λ : Lie(G) → C, such that for all u, v ∈ G,
λ([u, v]) = 0.
In our examples below [u, v] = uv − vu.
The following lemma gives the construction of a locally defined determinant on G from a trace
on Lie(G).
Lemma 6 (see Proposition 2 and Theorem 3 in [13] which is based on [2]). A continuous trace
λ : Lie(G) → C gives rise to a determinant map Λ : Exp(Lie(G)) ⊂ G → C∗ def ined on the
range of the exponential mapping by
Λ(g) := exp(λ(Log(g))),
where locally Log = Exp−1, making the following diagram commutative, for any small enough
neighborhood U0 of zero in Lie(G):
λ
U0 ⊂ Lie(G)
........................................................................................................................ ............
Exp
..........................................................................................................................................
............
Exp(U0) ⊂ G
........................................................................................................................................................................ ............
Λ
C
..........................................................................................................................................
............
C∗
exp
Moreover, Λ is differentiable (hence of class C1) at 1G, with differential D1G Λ = λ.
Proof . We first observe that for all g ∈ Exp(U0), log(Λ(g)) = λ(Log(g)). Let u ∈ U0 ⊂ Lie(G)
be such that g = Exp(u). Since G is a regular Lie group, we can consider the C1-path γ(t) =
Exp(tu) going from 1G to Exp(u) = g. We have γ(t)−1 γ(t) = u and hence
(cid:90) 1
λ(cid:0)γ(t)−1 γ(t)(cid:1) dt = λ
(cid:18)(cid:90) 1
0
0
(cid:19)
γ(t)−1 γ(t)dt
= λ(u) = λ(Log(g))
(21)
using the continuity of λ and that Log(1G) = 0. It follows that if γ1, γ2 are two C1-paths going
from 1G to g1 and g2 respectively, then γ1γ2 is a C1-path going from 1G to g1g2 and we have
λ(cid:0)(γ1(t)γ2(t))−1
(cid:94)γ1(t)γ2(t)(cid:1) = λ(cid:0)γ2(t)−1γ1(t)−1 γ1(t)γ2(t) + γ2(t)−1 γ2(t)(cid:1)
= λ(cid:0)γ1(t)−1 γ1(t)(cid:1) + λ(cid:0)γ2(t)−1 γ2(t)(cid:1) ,
where we have used the tracial property of λ.
Now, for g1, g2 ∈ Exp(U0) ⊂ G,
log(Λ(g1g2)) = λ(Log(g1g2)) = λ (Log(g1)) + λ (Log(g2)) = log(Λ(g1)) + log(Λ(g2)),
and we can apply the map exp to both sides of this expression to reach the statement.
(cid:4)
18
C. Neira Jim´enez and M.F. Ouedraogo
Conversely, following [13] we give a construction of a trace from a determinant. Our proof
here is different from the one in loc. cit.
Lemma 7 (See Proposition 2 and Theorem 3 in [13]). A determinant Λ : Exp(Lie(G)) → C∗,
which is of class C1 on G, yields a continuous trace λ : Lie(G) → C in the following way: for all
u ∈ Lie(G)
λ(u) := D1G Λ(u) =
d
dt
Λ(Exp(tu)),
which makes the following diagram commutative:
Exp(Lie(G))
.......................................................................................................................................................................... ............
Λ
(cid:12)(cid:12)(cid:12)t=0
C∗
............
..........................................................................................................................................
exp
Exp
............
..........................................................................................................................................
Lie(G) ............................................................................................................................................................................................................. ............
C
Proof . Let u1, u2 ∈ Lie(G). Then
λ
(cid:12)(cid:12)(cid:12)s=0
(cid:12)(cid:12)(cid:12)s=0
(cid:12)(cid:12)(cid:12)s=0
(cid:12)(cid:12)(cid:12)t=0
(cid:12)(cid:12)(cid:12)t=0
(cid:12)(cid:12)(cid:12)t=0
d
dt
d
dt
d
dt
λ([u1, u2]) =
=
=
d
ds
d
ds
d
ds
Λ (Exp(tu1)Exp(su2)Exp(−tu1))
Λ(Exp(tu1))Λ(Exp(su2))Λ(Exp(−tu1))
Λ(Exp(su2)) = 0.
Here we use the fact that Λ is multiplicative, which implies that Λ(g−1) = Λ(g)−1. The linearity
(cid:4)
of λ can be proved using the commutativity of the diagram.
Remark 4. The two lemmata imply that continuous traces on Lie(G) are in one to one corre-
spondence with C1-determinants on the open subset of G given by the range of the exponential
mapping.
In the following we assume that M is an n-dimensional closed manifold and n is odd. We
are going to classify determinants on a neighborhood of the identity in the space of invertible
odd(M ))∗, for a non-positive integer a. For that, let us first
pseudodifferential operators (I + C(cid:96)a
single out the Fr´echet -- Lie groups and Fr´echet -- Lie algebras we use.
The following proposition can be found in Proposition 3 in [13] for the case (C(cid:96)0(M ))∗. See
Proposition 6.1.4 in [21] for the case of odd-class operators. We give here an exhaustive proof
of this result since we did not find it in the literature.
its Fr´echet -- Lie algebra is C(cid:96)0
odd(M )(cid:1)∗
Proposition 6. (cid:0)C(cid:96)0
Proof . By Lemma 3, the composition of two operators in(cid:0)C(cid:96)0
and the same holds for the inverse so that the set(cid:0)C(cid:96)0
odd(M )(cid:1)∗
in (cid:0)C(cid:96)0
odd(M )(cid:1)∗
; we want to build an open neighborhood of A in (cid:0)C(cid:96)0
C(cid:96)0
odd(M ). Let us show that it is also an open subset of C(cid:96)0
is a group in the Fr´echet algebra
odd(M ). Let A be an operator
. The algebra
C(cid:96)0
odd(M ) is contained in C(cid:96)0(M ) which corresponds to all bounded classical ψDOs on L2(M )
and hence it is contained in the Banach algebra L(L2(M )) of bounded linear operators. By
the local inverse function theorem, the set of invertible operators on L(L2(M )) is an open set.
Hence A admits an open neighborhood V in the set of invertible operators in L(L2(M )).
belongs to(cid:0)C(cid:96)0
odd(M )(cid:1)∗
is a Fr´echet -- Lie group which admits an exponential mapping and
odd(M )(cid:1)∗
odd(M )(cid:1)∗
odd(M ).
Traces and Determinants on Odd-Class Operators
19
On the other hand (see [29]), the inclusion C(cid:96)0(M ) → L(L2(M )) is continuous so that the
odd(M ) → C(cid:96)0(M ) → L(L2(M )) is also continuous and the inverse image i−1(V )
is canonically
yields an open neighborhood of A in(cid:0)C(cid:96)0
. It follows that(cid:0)C(cid:96)0
odd(M )(cid:1)∗
odd(M )(cid:1)∗
inclusion i : C(cid:96)0
equipped with a manifold structure which makes it a Lie group.
Let us now construct an exponential mapping on C(cid:96)0
odd(M ), the differential equation
C(cid:96)0
odd(M ). Given any operator B in
t
(cid:90)
A−1
A0 = I
At = B,
has a unique solution in(cid:0)C(cid:96)0(M )(cid:1)∗
order. Let us check that At belongs to (cid:0)C(cid:96)0
exp(tλ)(B − λ)−1 dλ,
given by
i
2π
At =
Γ
odd(M )(cid:1)∗
symbol of At are
σ(At)−j =
i
2π
(cid:90)
Γ
exp(tλ)b−j(λ) dλ,
where Γ is a contour around the spectrum of B. Note that At is bounded since B has zero
. The homogeneous components of the
where b−j(λ) denote the components of the resolvent (B − λ)−1 of B at the point λ. Explicitly
we have
b0(λ) := (σ0(B) − λ)−1,
b−j(λ) := −b0(λ)
(cid:88)
(−i)α
α
l<j
k+l+α=j
ξ σ−k(B)∂α
∂α
x b−l(λ).
It follows that At
lies in
. This defines an exponential mapping
odd(M ) since B lies in C(cid:96)0
odd(M ).
So that (B − λ)−1 lies in C(cid:96)0
(cid:0)C(cid:96)0
odd(M )(cid:1)∗
odd(M ) →(cid:0)C(cid:96)0
smooth curve γu : [0, 1] →(cid:0)C(cid:96)0
Exp : C(cid:96)0
odd(M )(cid:1)∗
odd(M )(cid:1)∗
.
Moreover, it follows that for any smooth curve u : [0, 1] → C(cid:96)0
odd(M ), there exists a unique
[0, 1]
.................................................................................................................................................................................. ............
u
C(cid:96)0
odd(M )
Exp
..................................................................................................................................................................................................................... ............
defined by the following diagram
(cid:0)C(cid:96)0
odd(M )(cid:1)∗
which solves the initial value problem γ−1
u γu = u, γu(0) = I.
(cid:4)
Inspired by Corollary 5.12 in [19], and [33] we have the following
Proposition 7. If a < 0, the space of invertible odd-class ψDOs
odd(M )})∗
odd(M ), which admits an exponential mapping
G := (I + C(cid:96)a
odd(M ))∗ = ({I + A : A ∈ C(cid:96)a
is a Fr´echet -- Lie group with Fr´echet -- Lie algebra C(cid:96)a
from C(cid:96)a
odd(M ) to (I + C(cid:96)a
odd(M ))∗.
Explicitly this exponential mapping is given by
A (cid:55)→ Exp(A) =
odd(M ) → (cid:101)G,
Exp : C(cid:96)a
∞(cid:88)
1
k!
Ak.
a neighborhood of the identity in (cid:101)G. The inverse is given by
∞(cid:88)
Log : (cid:101)G → C(cid:96)a
I + A (cid:55)→ Log(I + A) =
odd(M ),
k=0
(−1)k+1
Ak.
k
k=1
This map restricts to a diffeomorphism from some neighborhood of the identity in C(cid:96)a
odd(M ) to
20
C. Neira Jim´enez and M.F. Ouedraogo
4.1 Classif ication of determinants on (I + C(cid:96)a
odd(M ))∗ for a ≤ 0
As before we consider a non-positive integer number a. From the classification of traces on
C(cid:96)a
odd(M ) derived in Theorem 3, we infer a description of the determinants defined on the
range of the exponential mapping in (I + C(cid:96)a
.
∗
odd(M ))
The following theorem supplements the known classification of determinants on classical
pseudodifferential operators.
Theorem 4. Let a ∈ Z be such that a ≤ 0. Determinant maps on (I + C(cid:96)a
a two-parameter family: for any c1, c2∈C, and for any linear map ρ : Cla
Detc1,c2(·) = exp(cid:0)c1ρ ◦ πa(Log(·)) + c2 TR(Log(·))(cid:1).
odd(M ))∗ are given by
odd (M )→C,
odd(M )/Cl2a−1
Proof . By Theorem 3, any trace τ on C(cid:96)a
and a generalized leading symbol trace:
odd(M ) is a linear combination of the canonical trace
τ (·) = c1ρ ◦ πa(·) + c2 TR(·),
for some c1, c2 ∈ C, and for some linear map ρ : Cla
tinuous for the Fr´echet topology of C(cid:96)a
and Lie(G) = C(cid:96)a
exp(cid:0)c1ρ ◦ πa(Log(·)) + c2 TR(Log(·))(cid:1).
odd(M ). Then, applying Lemma 6 to G = (I + C(cid:96)a
odd (M ) → C. Moreover, τ is con-
∗
odd(M ))
odd(M )/Cl2a−1
odd(M ), it follows that a determinant map on G is of the form
(cid:4)
The determinants given in Theorem 4 differ from the ones sometimes used by physicists
for operators of the type I+Schatten-class operator [15, 30] which in contrast to these are not
multiplicative but do extend the ordinary determinant (Fredholm determinant) for operators of
the type I+trace-class operator.
Here are some relevant specific cases
• Det1,0(·) = exp(ρ◦ πa(Log(·))) are extensions of the leading symbol determinants (see [23]
for the case C(cid:96)0(M )).
• Det0,1(·) = exp(TR(Log(·))) is an extension of the Fredholm determinant (see Lemma 2.1
in [27], and [30]).
4.2 Extension of determinants
Now we consider determinants constructed from a trace as above, by defining both sides of
Equation (21), not only on a neighborhood of the identity in the range of the exponential
mapping, but also on a set of admissible operators and on the pathwise connected component
of the identity.
Remark 5. In this section, the word "extension" is in the sense that the determinants can be
defined on operators which not necessarily lie in the range of the exponential mapping. See
Remark 6 below for the other sense of this word.
4.2.1 First extension
of (21). In the case of G = (cid:0)C(cid:96)0
odd(M )(cid:1)∗
The first way to define these determinants is carried out by considering the right hand side
the logarithm can be defined provided one chooses
a spectral cut θ thereby to fix a determination logθ of the logarithm. We set
Detλ
θ (A) := exp (λ(logθ A)) .
(22)
Traces and Determinants on Odd-Class Operators
21
Recall that if the operator A lies in the odd-class and has even order, then the logarithm logθ A
If φ is another spectral cut of A such that 0 ≤ θ < φ < 2π, by
lies also in the odd-class.
formula (8) we have
logθ A − logφ A = −2iπPθ,φ(A),
where Pθ,φ(A) is an odd-class projection as in (7).
Proposition 8 (Proposition 6.2.3 in [21]). Let λ be any continuous trace on C(cid:96)0
odd(M ). Suppose
that λ takes integer values on the image of Pθ,φ(A) for all θ, φ and A, where A is an admissible
operator with spectral cuts θ and φ as in (8). Then λ gives rise to the map Detλ
θ in (22), on
admissible operators, independent of the choice of the spectral cut θ.
Let us consider this construction for the traces given in Section 3: Let A be an admissible
odd(M ) with spectral cut θ.
operator in C(cid:96)0
λ = ρ ◦ π0 is defined by
With the notation of Subsection 3.1.3, the determinant associated to the leading symbol trace
Detλ
θ (A) := exp (ρ ◦ π0(logθ A)) .
In Example 2 of [13], it is shown that if P is a zero-order pseudodifferential idempotent, then
its leading symbol p is also an idempotent so that the fibrewise trace trx(p(x,·)) is the rank
rk(p(x,·)). Hence
(ρ ◦ π0)(Pθ,φ(A)) = (ρ ◦ π0)(cid:0)trx
(cid:0)σL(Pθ,φ(A)(x, ξ))(cid:1)(cid:1) = rk(cid:0)σL(Pθ,φ(A))(cid:1)(ρ ◦ π0)(I).
It follows that Detλ
odd(M )/Cl−1
ρ : Cl0
odd(M ) → Z.
θ (A) is independent of the choice of the spectral cut θ for any linear map
Observe that if A is a zero-order odd-class operator so is logθ A. Hence, the canonical trace
extends to logarithms of admissible odd-class operators of order zero with its property of cyclicity
in odd dimensions. The determinant associated to the canonical trace is defined by
DetTR
θ
(A) := exp(TR(logθ A)).
(23)
In contrast to the leading symbol trace, the canonical trace does not satisfy the requirement of
Proposition 8 so that the associated determinant depends on the choice of spectral cut.
4.2.2 Second extension
An alternative way to define these determinants is by considering the left hand side of (21) and
the expression for a determinant given in the proof of Lemma 6:
(cid:18)(cid:90) 1
(cid:19)
Λ(g) = exp
λ(γ(t)−1 γ(t)) dt
,
0
where γ : [0, 1] → G is a C1-path with γ(0) = 1G and γ(1) = g.
Such an approach was adopted in [2] by P. de la Harpe and G. Skandalis in the case of
a Banach Lie group. In her thesis [3], C. Ducourtioux adopted this point of view to construct
a determinant associated to a weighted trace with associated Lie algebras C(cid:96)0(M ) and C(cid:96)(M ).
In [13], J.-M. Lescure and S. Paycha showed that such a construction extends to Fr´echet -- Lie
groups with exponential mapping.
Proposition 9 (Proposition 6.2.2 in [21]).
1. Let Pcl(G) denote the space of closed C1-paths (loops) in G. The map
Φ : Pcl(G) → Lie(G),
c (cid:55)→
c ω =(cid:82) 1
(cid:90) 1
ω =
c
0
(cid:90)
c∗ω
(cid:19)
22
As in Definition 5, let (cid:101)G denote the pathwise connected component of the identity 1G of G
and P(G) the set of C1-paths γ : [0, 1] → G starting at 1G (γ(0) = 1G) in (cid:101)G. On P(G) we
C. Neira Jim´enez and M.F. Ouedraogo
introduce the map: Detλ : P(G) → C∗ defined by
(cid:18)(cid:90)
(cid:19)
(cid:18)(cid:90) 1
(cid:19)
λ(γ∗ω)
Detλ(γ) := exp
λ(ω)
= exp
,
(24)
where ω = g−1dg is the Maurer -- Cartan form on G.
γ
0
Since λ satisfies the tracial property, Detλ has the multiplicative property:
Lemma 8. Let γ1, γ2 be two C1-paths in P(G). Then
Detλ(γ1γ2) = Detλ(γ1) Detλ(γ2).
Proof . The same proof applies as in Lemma 6.
(cid:4)
In general the Maurer -- Cartan form ω = g−1dg is not exact on G so that for a C1-path
c : [0, 1] → G with c(0) = c(1), the integral(cid:82)
0 c∗ω does not vanish.
induces a map Φ : Π1(G) → Lie(G) on the fundamental group Π1(G) of G. Consequently,
the map Detλ defined in (24) only depends on the homotopy class of the path γ.
2. If Detλ(Π1(G)) = 1, then it induces a determinant map:
Detλ : (cid:101)G → C∗,
(cid:18)(cid:90) 1
0
g (cid:55)→ exp
λ(γ∗ω)
independently of the choice of path γ.
3. If g lies in the range of the exponential mapping Exp then
Detλ(g) = exp (λ(Log(g))) ,
where Log = Exp−1 is the inverse of the exponential mapping.
Remark 6. Item 3 of this proposition shows that Detλ is an extension of the determinant map
defined in Lemma 6.
Proof . 1. We want to show that two homotopic loops c1 and c2 have common primitive. Let us
first recall the following general construction of a primitive: for ω a differential form on G, let
γ : [0, 1] → G be a C1-path and F : [0, 1] → G be such that for any t ∈ [0, 1], F (cid:48)(t) = ω(γ(t))γ(cid:48)(t).
If ω is an exact form i.e. w = df for some f ∈ G, then F (t) = f (γ(t)) is a primitive of F (cid:48).
If the form ω is closed, then ω is locally exact. In this case, let 0 = t0 < t1 < ··· < tk = 1
be a subdivision of the interval [0, 1] such that γ([ti−1, ti]) is a subset of G, and there exists fi
defined on [ti−1, ti] such that dfi = ω. We can construct a function F (t) on [0, 1] in the following
way: F (t) = f0(γ(t)) on [t0, t1], F (t) = f1(γ(t)) − h1 on [t1, t2] where h1 = f1(γ(t1)) − f0(γ(t1))
and for i = 3, . . . , k, F (t) = fi(γ(t)) − hi on [ti−1, ti] where hi = fi(γ(ti)) − fi−1(γ(ti)) + hi−1.
Now let F (t) and G(t) be primitives of c1 and c2 respectively. Since c1 and c2 are homotopic,
there exists a family of C1-paths (αi)0≤i≤k defined in a neighborhood of 1G such that c1 =
Traces and Determinants on Odd-Class Operators
k(cid:81)
αi. Each path αi is closed so that (cid:82)
i=0
c ω vanishes on αi. It follows that F (t) = G(t), and
c2
therefore the map Φ is well-defined on Π1(G).
2. The multiplicativity of Detλ on G follows from Lemma 8: If g1, g2 are two elements of G
and γ1, γ2 are two C1-paths in P(G) such that γ1(1) = g1 and γ2(1) = g2, then Detλ(g1g2) =
Detλ(g1) Detλ(g2).
3. For g in the range of the exponential mapping and γ a C1-path in P(G) such that γ(1) = g,
Log = Exp−1 is well-defined so that λ(Log(γ(t))) is a primitive of λ(γ(t)−1 γ(t)). Thus we recover
equation (21):
23
(cid:90) 1
λ(γ(t)−1 γ(t)) dt = λ(Log(g)).
(cid:4)
0
odd(M )(cid:1)∗
From Proposition 8 we can see that in the first extension of the definition of a determinant,
the non-dependency on the spectral cut is controlled by the image of the projection Pθ,φ by
the trace λ. From Proposition 9, the map Detλ is well defined if the image of the fundamental
is trivial. The following theorem shows that in both definitions (Subsec-
tions 4.2.1 and 4.2.2) the non-dependency on the spectral cut and the multiplicativity of Detλ
. It provides a way to identify which
of the maps Detλ defined above send this fundamental group to 1 and therefore are determinants
in the sense of Definition 5.
group of(cid:0)C(cid:96)0
rely on the image of the fundamental group of(cid:0)C(cid:96)0
odd(M )(cid:1)∗
(cid:0)C(cid:96)0
Theorem 5 (see Section 4.5 in [8] and Lemma A.5 in [3]). The fundamental group of
is generated by the homotopy classes of loops {Exp(2iπtP )}0≤t≤1, where P is
odd(M )(cid:1)∗
a projector in C(cid:96)0
odd(M ).
Remark 7. Let us give some comments for the case C(cid:96)odd(M ). As seen in Theorem 2, any
trace in C(cid:96)odd(M ) is proportional to the canonical trace. For the subalgebra C(cid:96)0
odd(M ) in (23)
we defined the determinant associated to the canonical trace of an operator A with spectral
cut θ. Unfortunately, this definition cannot be extended a priori to positive order elements
of C(cid:96)odd(M ). Indeed, as we said before (see Subsection 4.2.1), if the order of A is even, for
any spectral cut θ of A, the logarithm logθ A is also odd-class, so the canonical trace extends
to logarithms of odd-class operators with even order and one can extend the determinant by:
DetTR
(A) := exp(TR(logθ A)). This is no longer true if the order of A is odd. M. Braverman
in [1] introduced the notion of symmetrized trace to define a symmetrized determinant on odd-
class operators. It was shown in [20] that this symmetrized trace is in fact the canonical trace.
The idea is to compute the average of the usual terms given by two spectral cuts of A, namely θ
and θ − aπ, where a is the order of A; one obtains the following definition of symmetrized
logarithm:
θ
logsym
θ A :=
1
2
(logθ A + logθ−aπ A).
This symmetrized logarithm is also odd-class and, once again, one can use the canonical trace
to define a determinant by setting:
(A) := exp(cid:0)TR(cid:0) logsym
θ A(cid:1)(cid:1).
DETsym
θ
Notice that if a = 0, DETsym
(A). However this symmetrized determinant also
depends on the spectral cut θ and under suitable assumptions it is multiplicative up to a sign
(see [1]).
(A) = DetTR
θ
θ
24
C. Neira Jim´enez and M.F. Ouedraogo
Acknowledgements
We are very grateful to Professor Sylvie Paycha for her advice and encouragement along an
important part of our careers. We express our gratitude to the Max Planck Institute for
Mathematics in Bonn and the University of Regensburg where part of this work was done.
We would also like to thank the anonymous referees for providing constructive comments on the
manuscript.
References
[1] Braverman M., Symmetrized trace and symmetrized determinant of odd class pseudo-differential operators,
J. Geom. Phys. 59 (2009), 459 -- 474, math-ph/0702060.
[2] De la Harpe P., Skandalis G., D´eterminant associ´e `a une trace sur une alg´ebre de Banach, Ann. Inst. Fourier
(Grenoble) 34 (1984), 241 -- 260.
[3] Ducourtioux C., Weighted traces on pseudodifferential operators and associated determinants, Ph.D. thesis,
Universit´e Blaise Pascal, Clermont-Ferrand, 2001.
[4] Fedosov B.V., Golse F., Leichtnam E., Schrohe E., The noncommutative residue for manifolds with boundary,
J. Funct. Anal. 142 (1996), 1 -- 31.
[5] Grubb G., A resolvent approach to traces and zeta Laurent expansions, in Spectral Geometry of Manifolds
with Boundary and Decomposition of Manifolds, Contemp. Math., Vol. 366, Amer. Math. Soc., Providence,
RI, 2005, 67 -- 93, math.AP/0311081.
[6] Guillemin V., A new proof of Weyl's formula on the asymptotic distribution of eigenvalues, Adv. Math. 55
(1985), 131 -- 160.
[7] Guillemin V., Residue traces for certain algebras of Fourier integral operators, J. Funct. Anal. 115 (1993),
391 -- 417.
[8] Kontsevich M., Vishik S., Determinants of elliptic pseudodifferential operators, hep-th/9404046.
[9] Kriegl A., Michor P.W., The convenient setting of global analysis, Mathematical Surveys and Monographs,
Vol. 53, American Mathematical Society, Providence, RI, 1997.
[10] Lesch M., On the noncommutative residue for pseudodifferential operators with log-polyhomogeneous sym-
bols, Ann. Global Anal. Geom. 17 (1999), 151 -- 187, dg-ga/9708010.
[11] Lesch M., Pseudodifferential operators and regularized traces, in Motives, Quantum Field Theory, and
Pseudodifferential Operators, Clay Math. Proc., Vol. 12, Amer. Math. Soc., Providence, RI, 2010, 37 -- 72,
arXiv:0901.1689.
[12] Lesch M., Neira Jim´enez C., Classification of traces and hypertraces on spaces of classical pseudodifferential
operators, arXiv:1011.3238.
[13] Lescure J.M., Paycha S., Uniqueness of multiplicative determinants on elliptic pseudodifferential operators,
Proc. Lond. Math. Soc. (3) 94 (2007), 772 -- 812.
[14] Maniccia L., Schrohe E., Seiler J., Uniqueness of the Kontsevich -- Vishik trace, Proc. Amer. Math. Soc. 136
(2008), 747 -- 752, math.FA/0702250.
[15] Mickelsson J., Current algebras and groups, Plenum Monographs in Nonlinear Physics, Plenum Press, New
York, 1989.
[16] Milnor J., Remarks on infinite-dimensional Lie groups, in Relativity, Groups and Topology, II (Les Houches,
1983), North-Holland, Amsterdam, 1984, 1007 -- 1057.
[17] Neira Jim´enez C., Cohomology of classes of symbols and classification of traces on corresponding classes
of operators with non positive order, Ph.D. thesis, Universitat Bonn, 2009, available at http://hss.ulb.
uni-bonn.de/2010/2214/2214.htm.
[18] Okikiolu K., The Campbell -- Hausdorff theorem for elliptic operators and a related trace formula, Duke
Math. J. 79 (1995), 687 -- 722.
[19] Omori H., Maeda Y., Yoshioka A., Kobayashi O., On regular Fr´echet -- Lie groups. IV. Definition and fun-
damental theorems, Tokyo J. Math. 5 (1982), 365 -- 398.
[20] Ouedraogo M.F., A symmetrized canonical determinant on odd-class pseudodifferential operators, in Geo-
metric and Topological Methods for Quantum Field Theory, Cambridge Univ. Press, Cambridge, 2010,
381 -- 393.
Traces and Determinants on Odd-Class Operators
25
[21] Ouedraogo M.F., Extension of the canonical trace and associated determinants, Ph.D. thesis, Universit´e
Blaise Pascal, Clermont-Ferrand, 2009.
[22] Paycha S., The noncommutative residue and canonical trace in the light of Stokes' and continuity properties,
arXiv:0706.2552.
[23] Paycha S., Rosenberg S., Traces and characteristic classes on loop spaces, in Infinite Dimensional Groups
and Manifolds, IRMA Lect. Math. Theor. Phys., Vol. 5, de Gruyter, Berlin, 2004, 185 -- 212.
[24] Paycha S., Scott S., A Laurent expansion for regularized integrals of holomorphic symbols, Geom. Funct.
Anal. 17 (2007), 491 -- 536, math.AP/0506211.
[25] Ponge R., Spectral asymmetry, zeta functions, and the noncommutative residue, Internat. J. Math. 17
(2006), 1065 -- 1090, math.DG/0310102.
[26] Ponge R., Traces on pseudodifferential operators and sums of commutators, J. Anal. Math. 110 (2010),
1 -- 30, arXiv:0707.4265.
[27] Scott S., The residue determinant, Comm. Partial Differential Equations 30 (2005), 483 -- 507,
math.AP/0406268.
[28] Seeley R.T., Complex powers of an elliptic operator, in Singular Integrals (Proc. Sympos. Pure Math.,
Chicago, Ill., 1966), Amer. Math. Soc., Providence, R.I., 1967, 288 -- 307.
[29] Shubin M.A., Pseudodifferential operators and spectral theory, 2nd ed., Springer-Verlag, Berlin, 2001.
[30] Simon B., Trace ideals and their applications, London Mathematical Society Lecture Note Series, Vol. 35,
Cambridge University Press, Cambridge, 1979.
[31] Wodzicki M., Noncommutative residue. I. Fundamentals, in K-Theory, Arithmetic and Geometry (Moscow,
1984 -- 1986), Lecture Notes in Math., Vol. 1289, Springer, Berlin, 1987, 320 -- 399.
[32] Wodzicki M., Spectral asymmetry and noncommutative residue, Thesis, Steklov Institute, Soviet Academy
of Sciences, Moscow, 1984.
[33] Yoshioka A., Maeda Y., Omori H., Kobayashi O., On regular Fr´echet -- Lie groups. VII. The group generated
by pseudodifferential operators of negative order, Tokyo J. Math. 7 (1984), 315 -- 336.
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.