Adjoint functors between categories of Hilbert

arXiv:1409.8656v3 [math.OA] 23 Feb 2016

C ∗-modules
∗ †
Pierre Clare Tyrone Crisp Nigel Higson

Abstract
Let E be a (right) Hilbert module over a C ∗ -algebra A. If E is
equipped with a left action of a second C ∗ -algebra B, then tensor
product with E gives rise to a functor from the category of Hilbert
B-modules to the category of Hilbert A-modules. The purpose of this
paper is to study adjunctions between functors of this sort. We shall
introduce a new kind of adjunction relation, called a local adjunction,
that is weaker than the standard concept from category theory. We
shall give several examples, the most important of which is the func-
tor of parabolic induction in the tempered representation theory of
real reductive groups. Each local adjunction gives rise to an ordinary
adjunction of functors between categories of Hilbert space representa-
tions. In this way we shall show that the parabolic induction functor
has a simultaneous left and right adjoint, namely the parabolic re-
striction functor constructed in [CCH16].

Keywords: Hilbert C ∗ -modules; adjoint functors; parabolic induction

MSC2010: 46L08, 18A40

1 Introduction
Let A be a C ∗ -algebra and denote by A H the category of non-degenerate
Hilbert space representations of A. This is obviously a category of interest
∗
Partially supported by the Danish National Research Foundation through the Centre
for Symmetry and Deformation (DNRF92).
†
Partially supported by the US National Science Foundation DMS-1101382.

1
when A is a group C ∗ -algebra, since it is isomorphic to the category of unitary
representations of the group. Similarly, if A is the reduced C ∗ -algebra of a real
reductive group, then A H is isomorphic to the category of tempered unitary
representations of the group.
In addition, denote by HA the category of right Hilbert A-modules (a
Hilbert module is a Banach module whose norm is obtained from an associ-
ated A-valued inner product; see [BLM04], [Lan95], or Section 2 for a quick
review). The role of this category in representation theory is a bit less clear,
but for example if A is the reduced C ∗ -algebra of a real reductive group G,
then each discrete series representation of G determines an object in HA .
More generally, the category HA captures the topology of the dual space A b
of irreducible representations of A (this is the tempered dual in our reduced
group C ∗ -algebra example). In contrast, the category A H of Hilbert space
representations is more closely related to the structure of the dual as a set
or measurable space.
In this paper we shall study the categories A H and HA with a particular
view to the case of the reduced C ∗ -algebra of a real reductive Lie group G.
We shall further develop the analysis of the parabolic induction functor

IndG
P: Cr∗ (L) H −→ Cr∗ (G) H

from the perspective of Hilbert modules that was begun in [Cla13] and
[CCH16]. Here L is a Levi subgroup of a parabolic subgroup P ⊆ G. We
shall examine the relationship between parabolic induction and the functor
of parabolic restriction

ResG
P: Cr∗ (G) H −→ Cr∗ (L) H

that we introduced in [CCH16]. We shall prove that parabolic induction and

restriction are left and right adjoints of one another. In fact we shall prove
this as a consequence of a stronger statement involving the Hilbert module
categories HCr∗ (G) and HCr∗ (L) . See Theorems 5.1 and 5.2.
If A and B are C ∗ -algebras, then by a correspondence from A to B we shall
mean a Hilbert B-module equipped with a nondegenerate representation of
A as adjointable Hilbert B-module endomorphisms. Associated to such a
correspondence F there are functors

BH −→ A H and HA −→ HB

2
that are constructed using Hilbert module tensor products. The functors
of parabolic induction and restriction from [Cla13] and [CCH16] are of this
type.
In the algebraic setting, where A and B are rings with unit, it is well
known that if F is an A-B-bimodule, then the associated tensor product
functors
B Mod −→ A Mod and ModA −→ ModB
between left and right module categories always admit right adjoints, namely
X 7→ HomA (F, X) and Y 7→ HomB (F, Y ) respectively. But in the C ∗ -setting
the extra symmetry imposed by the ∗-operation means that every right ad-
joint is also a left adjoint, and the existence of an adjoint functor in these
circumstances is a much more delicate matter.
Kajiwara, Pinzari and Watatani [KPW04] have obtained necessary and
sufficient conditions on a correspondence F for the tensor-product functor
between Hilbert module categories to admit an adjoint. Unfortunately their
conditions, which we shall review in Section 2.4, immediately rule out many
naturally-occurring examples of correspondences. Most notably from our
point of view, the conditions are not satisfied by the correspondences asso-
ciated to parabolic induction and restriction. Comparing with the theory
of smooth representations of p-adic reductive groups—where, as Bernstein
showed ([Ber87], cf. [Ren10]), the parabolic-induction functor admits both
left and right adjoints, and where those adjoints are close to being identical—
we were led to look for a weaker notion of adjunction for Hilbert modules
that would apply in particular to parabolic induction. The main novelty of
this paper is to define such a notion, and to study some of its properties.
If F is a correspondence from A to B, and E is a correspondence from B
to A, we shall say that the associated tensor-product functors
A⊗ F B ⊗ E
HA −→ HB and HB −→ HA

are locally adjoint if there are natural isomorphisms

∼
=
KB (X ⊗A F, Y ) −→ KA (X, Y ⊗B E)

between spaces of compact Hilbert module operators. See Definition 3.1 for
details. Not every morphism between Hilbert modules is compact, so our
definition is not the usual definition of an adjunction. Nor is it obtained
from the usual definition by adjusting to take into account the C ∗ -structure

3
[GLR85] on the categories HA and HB . But if two tensor product functors
are adjoint in the usual sense, then the natural isomorphisms coming from
the adjunction restrict to isomorphisms between spaces of compact operators,
as above, and so adjunction implies local adjunction. See Corollary 3.9. The
converse is rarely true: in general local adjunction is a genuinely weaker
condition.
The commutative case illustrates the distinction between adjoint and local
adjoint quite well. If Y is a quotient of a compact Hausdorff space X, then
the pullback functor HC(Y ) → HC(X) has an adjoint if and only if X is a
finite cover of Y , while it has a local adjoint if and only if X is a finite
branched cover. See Section 4.1. In addition, if U is an open subset of X,
then the pushforward HC0 (U ) → HC(X) always has a local adjoint, but it has
an adjoint if and only if U is both open and closed in X.
In Sections 3.3 and 3.7 we characterize locally adjoint pairs of tensor prod-
uct functors, and show that the condition of local adjointability is equivalent
to the condition of “finite numerical index” considered in [KPW04]. Return-
ing to categories of Hilbert space representations, we show that if two tensor
product functors are locally adjoint, then the associated functors between the
Hilbert space representation categories A H and B H are (two-sided) adjoints
in the usual sense. See Theorem 3.15.
On the basis of all this it is easy to analyze several examples, and we shall
do so in Section 4. Our main example of parabolic induction is considered in
Section 5. As a particular consequence we show that the parabolic restric-
tion functor constructed in [CCH16] is a two-sided adjoint to the parabolic
induction functor from tempered unitary representations of a Levi factor of
a reductive group G to tempered unitary representations of G. This adjoint
functor appears to be new.

2 Background and Notation

2.1 Hilbert Modules and Correspondences
Let A be a C ∗ -algebra. Recall that a Hilbert A-module is a right A-module
X that is equipped with an inner product

h , i : X × X −→ A

4
that is C-antilinear in the first variable, C-linear in the second variable, and
satisfies

hx, yai = hx, yia, hx, yi∗ = hy, xi and hx, xi ≥ 0

for all x, y ∈ X and a ∈ A. In addition the formula

kxk2X := khx, xikA

is required to define a complete norm on X. See for example [BLM04, Chap-

ter 8] or [Lan95] for an introduction to this concept. (But note that the
name “Hilbert C ∗ -module” is used in [BLM04] to refer to what we are call-
ing Hilbert module, while the name “Hilbert module” is used in [BLM04] to
refer to a different concept.)
A map T : X → Y between Hilbert A-modules is adjointable if there is a
map T ∗ : Y → X (necessarily unique) satisfying

hT x, yi = hx, T ∗ yi

for all x ∈ X and y ∈ Y . Adjointable operators are automatically A-linear

and bounded, but the converse is not true, in general. We shall denote by
BA (X, Y ) the Banach space of all adjointable operators from X to Y . The
space BA (X, X) of adjointable operators from X to itself is a C ∗ -algebra in
the operator norm.

Definition 2.1. Let B be a second C ∗ -algebra and let F be a Hilbert B-

module. A ∗-homomorphism A → BB (F, F ) is nondegenerate if the elements
af (a ∈ A, f ∈ F ) span a dense subspace of F .

Definition 2.2. By a correspondence from A to B we shall mean a Hilbert

B-module F equipped with a nondegenerate ∗-homomorphism from A into
BB (F, F ).

Remarks 2.3. A correspondence from A to B is in particular an A-B bi-

module, but the definition is asymmetric in that no A-valued inner product
is implied. The condition of nondegeneracy will not play a crucial role in
what follows. Nevertheless we shall assume it since it in any case holds in
the main examples of interest to us.

5
2.2 Compact Operators
If X and Y are Hilbert A-modules, and if x ∈ X and y ∈ Y , then the formula

(2.4) y ⊗ x∗ : z 7→ yhx, zi

defines an adjointable operator from X to Y ; the adjoint is x ⊗ y ∗. The

operator-norm closure of the linear span of the operators

(2.5) y ⊗ x∗ ∈ BA (X, Y )

is by definition the subspace of A-compact operators, denoted KA (X, Y ).

The composition, on either side, of a compact operator with an adjointable
operator is compact.
To add some algebraic substance to the notation used in (2.4) and (2.5),
we introduce the following concept:

Definition 2.6. Let X be a Hilbert A-module. The conjugate of X, denoted

X ∗ , is the complex conjugate of the vector space X, equipped with left A-
module structure defined by

a · x∗ = (xa∗ )∗ .

Here x∗ ∈ X ∗ denotes the image of x ∈ X under the obvious conjugate

C-linear isomorphism X → X ∗ .

Remark 2.7. The term adjoint would usually be a more appropriate name
for X ∗ than conjugate, but the word will be used quite heavily in other senses
in this paper.

Using the left A-module structure on X ∗ we find that

y · a ⊗ x∗ = y ⊗ a · x∗ ∈ BA (X, Y ),

and so the formula (2.4) defines a C-linear map

(2.8) Y ⊗alg ∗
A X −→ BA (X, Y ).

The compact operators therefore constitute a completion of the algebraic ten-

sor product Y ⊗alg ∗
A X within BA (X, Y ). We shall return to this perspective
on compact operators in a little while.

6
 We shall need one more idea about compact operators. Let X be a Hilbert
A-module. The space KA (A, X) of compact adjointable operators from the
Hilbert module A to X is itself a Hilbert A-module under the inner product
hS, T i = S ∗ T.
Here the operator S ∗ T : A → A is left multiplication by some unique element
of A, and we identify S ∗ T with that element so as to obtain an A-valued
inner product. Each operator in KA (A, X) is of the form
T : a 7→ xa
for some unique element x ∈ X. Moreover all these operators are adjointable,
with adjoint
T ∗ : y 7→ hx, yi.
Finally, T and T ∗ are compact, as can be seen using an approximate iden-
tity in A. We arrive at the following results (see also [BLM04, Proposition
8.1.11]).
Lemma 2.9. The Hilbert A-module X is isomorphic to the Hilbert A-module
KA (A, X) via the map that associates to x ∈ X the compact operator a 7→
xa.
Lemma 2.10. The conjugate X ∗ of a Hilbert A-module X is isomorphic to
the A-module KA (X, A) via the map that associates to x∗ ∈ X ∗ the compact
operator y 7→ hx, yi.

2.3 Functors on Hilbert Modules

Definition 2.11. If A is a C ∗ -algebra, then denote by HA the category whose
objects are right Hilbert A-modules and whose morphisms are adjointable
maps between Hilbert modules.
Within the algebraic context, bimodules give rise to functors between
module categories via tensor product. This is so in the Hilbert module con-
text, too, thanks to the following construction.
Definition 2.12. Let X be a Hilbert A-module and let F be a correspon-
dence from A to B. The (internal ) tensor product X ⊗A F , which is a Hilbert
B-module, is the completion of the algebraic tensor product X ⊗alg
A F in the
norm induced by the B-valued inner product
hx1 ⊗ f1 , x2 ⊗ f2 iX⊗A F := hf1 , hx1 , x2 iX · f2 iF .

7
 See [Lan95, Chapter 4] for basic information on the internal tensor prod-
uct construction. One has natural isomorphisms
∼
= ∼
=
A ⊗A F −→ F and F ⊗B B −→ F

via multiplication.
If F is a correspondence from A to B, then internal tensor product with
F gives rise to a tensor product functor

F : HA −→ HB

(as indicated, we shall use the same letter for the bimodule and the functor),
since the tensor product of an adjointable operator with the identity operator
on F is an adjointable operator between tensor product modules.
It is interesting to note that subject to a natural continuity condition
and compatibility with the adjoint operation, every functor between Hilbert
module categories is a tensor product functor. We shall not use this fact, but
here is a short summary.

Definition 2.13. A functor F between categories of Hilbert modules is called

a ∗-functor if it is C-linear on morphisms, and satisfies F (T ∗ ) = F (T )∗ for
every adjointable operator T .

Definition 2.14. A ∗-functor F : HA → HB is strongly continuous if for

every object X ∈ HA , the ∗-homomorphism

F : KA (X, X) → BB (F (X), F (X))

is nondegenerate.

Theorem 2.15. [Ble97, Theorem 5.3] The category of strongly continuous

∗-functors HA → HB (and natural transformations) is equivalent to the cat-
egory of correspondences from A to B (and adjointable operators compatible
with the A-B bimodule structure).

The equivalence is given in one direction by sending a functor F to the

correspondence F (A), and in the other direction by sending a correspondence
F to the associated tensor product functor.

8
Remark 2.16. In addition a strongly continuous ∗-functor HA → HB is
an equivalence (with inverse a strongly continuous ∗-functor) if and only if
the associated correspondence is a Morita equivalence between A and B: see
[BLM04, 8.1.2, 8.2.20]. If A and B are Morita equivalent, then the dual
spaces Ab and Bb are homeomorphic (moreover the converse also holds if A
and B are commutative). So we see that HA carries information about the
b as a topological space, as remarked in the introduction.
structure of A

2.4 Adjunctions
We shall start with the standard definition from category theory.

Definition 2.17. Let A and B be C ∗ -algebras, and let E and F be corre-

spondences from B to A and from A to B, respectively, determining tensor
product functors
⊗A F ⊗B E
HA −→ HB and HB −→ HA .
An adjunction between F and E is a natural isomorphism
∼
=
(2.18) ΦX,Y : BB (X ⊗A F, Y ) −→ BA (X, Y ⊗B E),

or in other words a natural equivalence between the left and right sides in
op
(2.18), considered as functors from the product category HA × HB to the
category of sets.

But unlike the ordinary situation in category theory, there is no real

distinction between left adjoints and right adjoints in the context of Hilbert
modules: given an adjunction ΦX,Y as above, the formula

Φ∗X,Y : T 7→ (ΦX,Y )−1 (T ∗ )∗

defines an adjunction
∼
=
(2.19) Φ∗X,Y : BA (Y ⊗B E, X) −→ BB (Y, X ⊗A F )

that reverses the role of E and F in Definition 2.17.

Definition 2.20. Let Φ be an adjunction, as in Definition 2.17.

9
(a) A unit for Φ is a bounded, adjointable A-bimodule map

η : A −→ F ⊗B E

that defines ΦX,Y by means of the commuting diagram

idX ⊗η
X ⊗A A / X ⊗A F ⊗B E
∼
= T ⊗idE
 
X ΦX,Y (T )
/ Y ⊗B E.

(b) A counit for Φ is a bounded, adjointable B-bimodule map

ε : E ⊗A F −→ B

that defines the inverse of the isomorphism ΦX,Y in (2.18) by means of

the commuting diagram
idY ⊗ε
Y ⊗B B o Y ⊗B EO ⊗A F
∼
= T ⊗idF

Y o X ⊗B F.
Φ−1
X,Y (T )

Proposition 2.21. Every adjunction admits a unique unit and counit.

Proof. This is standard. For instance the unit is the image of the identity
operator on F under the map
∼
=
ΦA,F : BB (F, F ) −→ BA (A, F ⊗B E),

See [ML98, IV.1], for example.

No continuity conditions are imposed on the isomorphisms ΦX,Y in an
adjunction. But in fact continuity is automatic, as the following calculation
shows.
Lemma 2.22. Let Φ be an adjunction, as in Definition 2.17. Each map
∼
=
ΦX,Y : BB (X ⊗A F, Y ) −→ BA (X, Y ⊗B E),

is a linear, topological isomorphism.

10
Proof of the Lemma. It is clear from the definition of unit that ΦX,Y is linear,
with norm bounded by the operator norm of the unit map from A into
F ⊗B E. The inverse is likewise linear, with norm bounded by the norm of
the counit.
The following definition very slightly elaborates on Definition 2.6.
Definition 2.23. Let A and B be C ∗ -algebras and let F be a correspondence
from A to B. The conjugate of F is the Hilbert module conjugate F ∗ , as in
Definition 2.6, equipped with right A-module structure defined by

f ∗ · a = (a∗ f )∗ .

The conjugate F ∗ of a correspondence from A to B is a B-A-bimodule,

but, as it stands, is not a correspondence from B to A: there is no obvious
A-valued inner product. Kajiwara, Pinzari and Watatani [KPW04] relate the
existence of such an inner product on F ∗ to the existence of adjoint functors,
as follows.
Theorem 2.24. [KPW04, Theorem 4.4(2), Theorem 4.13] A tensor product
functor from HA to HB , induced from a correspondence F , has an adjoint
tensor product functor if and only if all the following conditions are met:
(a) The conjugate B-A-bimodule F ∗ carries an A-valued inner product mak-
ing it into a correspondence from B to A.

(b) The conjugate operator space structure on F ∗ is completely boundedly

equivalent to the Hilbert A-module operator space structure on F ∗ .
(c) The left action of A on F is through a ∗-homomorphism from A into
KB (F, F ).
(d) The left action of B on F ∗ is through a ∗-homomorphism from B into
KA (F ∗ , F ∗ ).
When these conditions are met, the adjoint functor from HB to HA is given
by tensor product with the correspondence F ∗ .
Condition (b) will be explained in the next section, where we shall also
give a proof of Theorem 2.24 in the course of our study of the weaker notion
of local adjunction. The full relationship of our results to those of [KPW04]
will be discussed in detail in Section 3.7.

11
Example 2.25. Consider the case of unital C ∗ -algebras A and B. If F is a
correspondence satisfying condition (a) of Theorem 2.24, then the conditions
(c) and (d) are equivalent to requiring that F be finitely generated (and
hence projective, cf. [BLM04, Theorem 8.1.27]) as a module over B and
A, respectively, while the analytic condition (b) follows automatically from
(c) and (d): see [KPW04, Example 2.31]. (We note that in the non-unital
setting, the conditions (b), (c) and (d) are independent of one another: see
the examples in Sections 4.5.3 and 4.5.1.) Combined with a theorem of
Morita on adjunctions in the algebraic setting [Mor65, Theorem 4.1], we find
that the following are equivalent for unital C ∗ -algebras A and B:

(a) The tensor product functor ⊗A F : HA → HB has an adjoint.

(b) The algebraic tensor product functor ⊗alg

A F : ModA → ModB has a
two-sided adjoint.

This situation is reminiscent of Beer’s result, that two unital C ∗ -algebras are
Morita equivalent as C ∗ -algebras if and only they are Morita equivalent as
rings [Bee82, Theorem 1.8].

2.5 Hilbert Modules as Operator Spaces

Recall that an operator space structure on a vector space X is a sequence
of Banach space norms on the spaces Mn (X) of n × n matrices with entries
from X such that

(a) the norm of a block diagonal matrix is the largest of the norms of its
diagonal blocks; and

(b) the norm of a three-fold product ABC, where A and C are scalar n × n
matrices, and B is an n × n matrix with entries from X, is no more
than the product of the norms of the matrices A, B and C (we use the
operator norm for the scalar matrices).

See [ER00] or [BLM04]. An operator space is of course a vector space with

an operator space structure.

Example 2.26. The above conditions hold when X is a closed subspace of

a C ∗ -algebra B and the norms on Mn (X) are inherited from the C ∗ -algebra
norm on Mn (B).

12
 A linear map T : X → Y between operator spaces is completely bounded
if
kT kcb = sup kMn (T )k < ∞,
n

where Mn (T ) : Mn (X) → Mn (Y ) is defined by applying T entrywise. An

isomorphism of operator spaces will mean, for us, a linear isomorphism T :
X → Y such that both T and T −1 are completely bounded. We shall use
the term complete isometric isomorphism when

kT kcb = kT −1kcb = 1.

Example 2.27. Every operator space is completely isometrically isomorphic

to a closed subspace of a C ∗ -algebra. See for example [ER00, Chapter 2] for
an exposition.

Suppose now that X is a right Hilbert A-module. Then Mn (X) is a

Hilbert Mn (A)-module in a natural way, namely the right action of Mn (A)
on Mn (X) is given by matrix multiplication, and the Mn (A)-valued inner
product is
n
X
hS, T ii,j = hSkj , Tkj i.
k=1

This observation gives X an operator space structure. See [BLM04, 8.2.1].

Example 2.28. If A is a C ∗ -algebra, then the operator space norms on

Mn (A) induced by the inner product ha, bi = a∗ b on A are the canonical
C ∗ -algebra norms.

Example 2.29. If X is a closed subspace of BA (Y, Z), where Y and Z are

Hilbert A-modules, then the operator space structure on X associated to the
BA (Y )-valued inner product hS, T i = S ∗ T coincides with the one given by
the natural embedding of BA (Y, Z) as a corner of the C ∗ -algebra BA (Y ⊕Z).
For instance, if X is a Hilbert A-module then the isomorphism X ∼ =
KA (A, X) of Lemma 2.9 is completely isometric for the operator space struc-
ture induced on X by the A-valued inner product, and the operator space
structure on KA (A, X) induced by its embedding into KA (A ⊕ X).

Example 2.30. If H is a Hilbert space, then the operator space structure

associated to the inner product on H is the column Hilbert space structure
[BLM04, 1.2.23] coming from the identification H ∼
= B(C, H).

13
 The assignment of an operator space structure to each Hilbert module,
as above, gives a functorial embedding of the category of Hilbert modules
and adjointable maps into the category of operator spaces and completely
bounded maps:
Theorem 2.31. [BLM04, Proposition 8.2.2] If X and Y are right Hilbert
A-modules, then every bounded A-linear map X → Y is completely bounded.
We are going to use operator spaces to treat Hilbert A-modules X and
their conjugate modules X ∗ introduced in Definition 2.6 on a somewhat equal
footing. To this end, we recall the following concept:
Definition 2.32. If X is any operator space, then its conjugate X ∗ is the
complex conjugate vector space, equipped with the norms
k[xij ]kMn (X ∗ ) = k[xji ]kMn (X) ,
which endow it with the structure of an operator space.
Remark 2.33. The operator space X ∗ is usually called the adjoint of X, but
once again we shall try to avoid over-using this word in this paper. However
we warn the reader that the term conjugate as it is used in [KPW04] refers
to adjoint functors.
Example 2.34. In the case of a Hilbert A-module X, the operator space
structure on X ∗ provided by Definition 2.32 is the operator space structure
we would obtain by viewing X ∗ as a left Hilbert A-module (a concept that
we are avoiding in this paper). See [BLM04, 1.2.25 and 8.2.3(2)].
Example 2.35. Let H be a complex Hilbert space, equipped with its column
operator space structure. The conjugate operator space H ∗ is, as a vector
space, the same as the complex conjugate Hilbert space H, and the conjugate
operator space structure on H ∗ is the same as the row Hilbert space structure
[BLM04, 1.2.23] on H. Being a Hilbert space in its own right, H also carries
a column Hilbert space structure. The identity map I : H → H, considered
as a map from the row operator space to the column operator space, has
kIk2cb = dim H. In particular, this map is not completely bounded if H is
infinite-dimensional. See [ER00, p.56].
Example 2.36. Let X be a Hilbert A-module, and view X ∗ as a conjugate
operator space as above. The isomorphism X ∗ ∼= KA (X, A) of Lemma 2.10 is
completely isometric, where KA (X, A) is viewed as a subspace of KA (X ⊕ A)
as in Example 2.29.

14
2.6 Operator Modules
If a C ∗ -algebra A (or more generally a Banach algebra of operators on a
Hilbert space) acts on a Banach space X, then there are natural induced
actions
Mn (A) × Mn (X) −→ Mn (X)
that combine the given action with matrix multiplication. When X is an
operator space we shall always assume that these actions are completely
contractive in the sense that

ka · xkMn (X) ≤ kakMn (A) kxkMn (X) .

The same will go for right actions instead of left actions, and we shall use the
term operator module to describe this situation (the term h-module is used
in [BLM04, Section 3.1.3]; this is a reference to the Haagerup tensor product
that we shall review below).

2.7 The Haagerup Tensor Product

There are several notions of tensor product for operator spaces. Here we
shall need only the Haagerup tensor product, which is defined as follows. Let
A be a C ∗ -algebra, let X be a right operator A-module, and let Y be a left
operator A-module. The Haagerup tensor product X ⊗hA Y is the completion
of the algebraic tensor product X ⊗alg
A Y that is characterized by the following
universal property: each bilinear map

Φ : X × Y −→ Z

into an operator space for which

(a) Φ(xa, y) = Φ(x, ay), and

(b) the matrix extensions

Φn : Mn (X) × Mn (Y ) −→ Mn (Z)

satisfy
kΦn (x, y)kMn(Z) ≤ kxkMn (X) kykMn(Y ) .
for all n

15
extends to a complete contraction from X ⊗hA Y to Z. See [BLM04, 3.4.2].
The Haagerup tensor product is associative [BLM04, Theorem 3.4.10] and
functorial in both variables with respect to completely bounded module maps
[BLM04, Lemma 3.4.5], and the natural isomorphism on algebraic tensor
products extends to a completely isometric isomorphism
∼
=
(X ⊗hA Y )∗ −→ Y ∗ ⊗hA X ∗ .
See [BLM04, 1.5.9].
The following theorems of Blecher relate the Haagerup tensor product to
the tensor product and compact operators on Hilbert modules:
Theorem 2.37. [BLM04, Theorem 8.2.11] Let X be a Hilbert A-module,
and let F be a correspondence from A to B. The identity map on X ⊗alg
A F
extends to a completely isometric natural isomorphism
∼
=
X ⊗hA F −→ X ⊗A F
from the Haagerup tensor product to the internal Hilbert module tensor prod-
uct.
Theorem 2.38. [BLM04, Corollary 8.2.15] Let X and Y be Hilbert A-
modules. There is a completely isometric isomorphism of operator spaces
∼
=
Y ⊗hA X ∗ −→ KA (X, Y )
mapping each elementary tensor y ⊗x∗ to the corresponding compact operator
y ⊗ x∗ defined in (2.5).
As in Example 2.29, KA (X, Y ) is a closed subspace of the C ∗ -algebra
KA (X ⊕ Y ) and it is to be viewed as an operator space in that way. Note
that the operator space Y ⊗hA X ∗ does not depend on the inner products on
X and Y , but only on the induced operator space structures. In contrast,
the action of Y ⊗hA X ∗ as operators X → Y appearing in Theorem 2.38 does
depend on the A-valued inner product on X.

3 Local Adjunctions for Hilbert Modules

3.1 Definitions and Basic Properties
We are ready now to introduce the main concept of the paper.

16
Definition 3.1. Let A and B be C ∗ -algebras, and let E and F be corre-
spondences from B to A and from A to B, respectively, determining tensor
product functors
⊗A F ⊗B E
HA −→ HB and HB −→ HA .
A local adjunction between these functors is a natural isomorphism
∼
=
ΦX,Y : KB (X ⊗A F, Y ) −→ KA (X, Y ⊗B E)
that is, for each X and Y , a continuous linear map.
Theorem 3.2. Let Φ be a local adjunction, as in Definition 3.1. Each of the
linear maps
∼
=
ΦX,Y : KB (X ⊗A F, Y ) −→ KA (X, Y ⊗B E)
is an isomorphism of operator spaces.
We will see later (Corollary 3.14) that the matrix norms of ΦX,Y are in
fact bounded independently of X and Y .
Proof. For a Hilbert A-module X, we let X ∞ denote the orthogonal direct
sum of countably many copies of X (see [BLM04, 8.1.9]). For each n ≥ 1
there is a natural isometric embedding
Mn (KA (X, Z)) −→ KA (X ∞ , Z ∞ )
that is defined by letting each n × n matrix act by matrix multiplication on
the first n copies of X inside X ∞ , and by zero on the remaining copies. One
also has isometric isomorphisms (X ⊗A F )∞ ∼ = X ∞ ⊗A F . The diagram
Mn (ΦX,Y )
Mn (KB (X ⊗A F, Y )) / Mn (KA (X, Y ⊗A E))

 ΦX ∞ ,Y ∞ 
KB (X ∞ ⊗A F, Y ∞ ) / KA (X ∞ , Y ∞ ⊗B E)

commutes by the naturality of Φ, showing that kMn (ΦX,Y )k ≤ kΦX ∞ ,Y ∞ k

for every n.
As with adjoints, there is no distinction between left and right local ad-
junctions: given a local adjunction Φ as above, we may define a second local
adjunction,
∼
=
(3.3) Φ∗X,Y : KA (Y ⊗B E, X) −→ KA (Y, X ⊗A F )

17
by means of the formula

Φ∗X,Y : T 7→ (ΦX,Y )−1 (T ∗ )∗ .

This interchanges the roles played by the correspondences E and F in Defini-

tion 3.1. And as with adjoints, it is very relevant to study units and counits
associated to a local adjunction. The following definition merely repeats
Definition 2.20 in the present context.

Definition 3.4. Let Φ be a local adjunction, as in Definition 3.1.

(a) A unit for Φ is a bounded, adjointable A-bimodule map

η : A −→ F ⊗B E

that defines Φ by means of the commuting diagram

idX ⊗η
X ⊗A A / X ⊗A F ⊗B E
∼
= T ⊗idE
 
X / Y ⊗B E.
ΦX,Y (T )

(b) A counit for Φ is a bounded, adjointable B-bimodule map

ε : E ⊗A F −→ B

that defines the inverse of the isomorphism ΦX,Y by means of the com-
muting diagram
idY ⊗ε
Y ⊗B B o Y ⊗B EO ⊗A F
∼
= T ⊗idF

Y o X ⊗A F.
Φ−1
X,Y (T )

Once again, these definitions are symmetric with respect to the transpo-
sition Φ ↔ Φ∗ given in (3.3) above. If η is a unit for Φ, then the adjoint
operator η ∗ is a counit for Φ∗ , while if ε is a counit for Φ, then ε∗ is a unit
for Φ∗ .

18
Lemma 3.5. A local adjunction admits at most one unit and at most one
counit.

Proof. Let η : A → F ⊗B E be a unit, and let uλ be an approximate unit in

the C ∗ -algebra KB (F, F ). The map η is the strong-operator limit of the net
(uλ ⊗ idE ) ◦ η, and the unit property of η identifies this net with ΦA,F (uλ ).
Thus η is uniquely determined by Φ. The uniqueness of counits follows by
symmetry from the uniqueness of units.
A local adjunction need admit neither a unit nor a counit, or it might
admit one without the other. But in the examples of interest to us at least
one will exist. It is also the case that every local adjunction admits a bounded
(but not necessarily adjointable) counit ε : F ⊗A E → B. See Section 3.5.
If a functor has an adjoint, then the adjoint is unique up to a canoni-
cal natural isomorphism. Local adjoints are, in general, only unique in the
following weaker sense:

Lemma 3.6. Suppose that a correspondence F from A to B is a local adjoint

to a correspondence E from B to A, and also to a second correspondence G
from B to A. There is a canonical completely bounded isomorphism E ∼=G
of B-A operator bimodules.

Proof. The two local adjunctions give completely bounded isomorphisms

= KA (A, E) ∼
E∼ = KA (A, G) ∼
= KB (F, B) ∼ =G

of B-A-bimodules.

Remarks 3.7. The converse of Lemma 3.6 is also true: Theorems 2.37
and 2.38 together imply that up to natural isomorphism, KA (X, Y ⊗B E)
depends only on the operator bimodule structure of E. In the course of
proving Theorem 3.10, below, we will in fact establish a bijection between
the set of local adjunctions between F and E, and the set of B-A operator
∼
=
bimodule isomorphisms F ∗ − → E.
On the question of uniqueness, we will later see (Proposition 3.34) that
if there exists a counit
ε : E ⊗A F −→ B,
then the canonical isomorphism E ∼ = G of Lemma 3.6 is adjointable, and so
furnishes a natural isomorphism between the tensor product functors E and

19
G. In particular, if the tensor product functor F has an adjoint, then it has
a unique local adjoint. In the absence of a counit, a tensor product functor
may admit several local adjoints that are not isomorphic to one another as
correspondences. See Section 4.5.

3.2 Local Adjunctions from Adjunctions

Lemma 3.8. If F : HA → HB is a tensor product functor with an adjoint
E : HB → HA , then for all X ∈ HA and all Y ∈ HB , the natural isomorphism

ΦX,Y : BB (X ⊗A F, Y ) → BA (X, Y ⊗B E)

maps KB (X ⊗A F, Y ) isomorphically onto KA (X, Y ⊗B E).

Proof. The space KB (X ⊗A F, Y ) is densely spanned by operators of the form

L = y ⊗ (Kx ⊗ f )∗ ,

where y ∈ Y , x ∈ X, f ∈ F , K ∈ KA (X, X), and we are using the notation

of (2.4). The naturality of Φ gives



ΦX,Y (L) = ΦX,Y (y ⊗ (x ⊗ f )∗ ) ◦ (K ∗ ⊗ 1F ) = ΦX,Y y ⊗ (x ⊗ f )∗ ◦ K ∗ ,

which is compact because K ∗ is. This shows that Φ maps compact operators
into compact operators, and a similar argument applied to Φ−1 shows that
the map is a bijection.

Corollary 3.9. Every adjunction of tensor product functors

∼
=
ΦX,Y : BB (X ⊗A F, Y ) −→ BA (X, Y ⊗B E)

restricts to a local adjunction

∼
=
ΦX,Y : KB (X ⊗A F, Y ) −→ KA (X, Y ⊗B E).

20
3.3 Characterization of Local Adjunctions
Theorem 3.10. A tensor product functor from HA to HB , induced from a
correspondence F , has a locally adjoint tensor product functor if and only if
both of the following conditions are met:
(a) The conjugate B-A-bimodule F ∗ carries an A-valued inner product mak-
ing it into a correspondence from B to A.
(b) The conjugate operator space structure on F ∗ is completely boundedly
equivalent to the Hilbert A-module operator space structure on F ∗ .
Proof. First suppose that E is a local adjoint to F . The isomorphism
∼
=
ΦA,B : KB (F, B) −→ KA (A, E)

gives, using Lemma 2.9, an isomorphism of B-A-bimodules

∼
=
F ∗ −→ E

that is, in addition, an operator space isomorphism. The A-valued inner

product on F ∗ inherited from E via this isomorphism satisfies conditions (a)
and (b).
Conversely, suppose we are given a compatible A-valued inner product
making F ∗ into a correspondence from B to A. Let us introduce a second
symbol, E, for this correspondence, and give E the operator space structure
it inherits from its Hilbert A-module structure (in contrast, we assign to F ∗
the operator space structure it receives as the conjugate of F ). We have
sequences of natural isomorphisms of operator spaces

KB (X ⊗A F, Y ) ∼
= Y ⊗hB (X ⊗A F )
∗
(1)
∼
= Y ⊗hB (X ⊗hA F )
∗
(2)
∼
= Y ⊗hB F ∗ ⊗hA X ∗
(3)

and

Y ⊗hB E ⊗hA X ∗ ∼
= Y ⊗B E ⊗hA X ∗
(4)
∼
= KA (X, Y ⊗B E),
(5)

21
as follows. The isomorphisms (1) and (5) come from Theorem 2.38, while
(2) and (4) come from Theorem 2.37. The isomorphism (3) is a result of the
compatibility of the Haagerup tensor product with the conjugation operation
on operator spaces. If we assume that the identity map from F ∗ to E is a
completely bounded isomorphism of operator spaces, then of course
Y ⊗hB F ∗ ⊗hA X ∗ ∼
= Y ⊗hB E ⊗hA X ∗ ,
and we can link all of the displayed isomorphisms together to obtain a local
adjunction between F and E.
Since a local adjunction between F and E determines canonically an
operator space isomorphism E ∼ = F ∗ —and conversely—we will usually just
write F ∗ instead of E from now on, keeping in mind that this implies the
choice of a suitable A-valued inner product on F ∗ .
Remarks 3.11. The proof gives a very simple formula for the adjunction
isomorphism
ΦX,Y : KB (X ⊗A F, Y ) → KA (X, Y ⊗B F ∗ ),
namely


(3.12) ΦX,Y y ⊗ (x ⊗ f )∗ = (y ⊗ f ∗ ) ⊗ x∗ .
We also note that, given a local adjunction between F and F ∗ , the equivalence
of the two operator space structures on F ∗ allows us to combine Theorems
2.37 and 2.38 to obtain canonical isomorphisms
(3.13) KB (F, F ) ∼
= F ⊗B F ∗ and KA (F ∗ , F ∗) ∼
= F ∗ ⊗A F
of A-A-bimodules and B-B-bimodules, respectively.
Corollary 3.14. Let Φ be a local adjunction, as in Definition 3.1. The
matrix norms of the isomorphisms ΦX,Y are bounded independently of X and
Y:
kΦX,Y kcb ≤ kΦA,B kcb ≤ kΦA∞ ,B∞ k.
Proof. Using Theorem 2.38 to identify spaces of compact operators with
Haagerup tensor products, the map ΦX,Y is given by
1Y ⊗ ΦA,B ⊗ 1X ∗ : Y ⊗hB F ∗ ⊗hA X ∗ → Y ⊗hB E ⊗hA X ∗ .
The functoriality of the Haagerup tensor product (see [BLM04, Lemma
3.4.5]) then gives kΦX,Y kcb ≤ kΦA,B kcb . Theorem 3.2 gives the second in-
equality, kΦA,B kcb ≤ kΦA∞ ,B∞ k.

22
3.4 Local Adjunctions and Representations
In addition to the tensor product functors F : HA → HB studied in the
previous sections, every correspondence F from A to B induces a functor
F = F ⊗B : B H −→ A H
between the categories of (nondegenerate) Hilbert space representations of A
and B. In this section we shall prove the following result.
Theorem 3.15. Let F be a correspondence from A to B, and let E be a
correspondence from B to A. Every local adjunction between F and E gives
rise to a (two-sided) adjunction between the tensor product functors
F : B H −→ A H and E : A H −→ B H
on categories of Hilbert space representations.
Remark 3.16. Under the equivalences explained in Section 3.7, this theorem
corresponds to [KPW04, Theorem 4.4(1)]. Compare also [CH16], where an
adjunction theorem is formulated in the context of operator modules.
Before beginning the proof let us collect some preliminary facts.
Lemma 3.17. Let F be a correspondence from A to B. The formula
(f1∗ , f2 ) 7→ hf1 , f2 i
defines a completely contractive map of operator B-B-bimodules
F ∗ ⊗hA F −→ B.
Proof. If S ∈ Mn (F ∗ ) (we shall avoid writing S ∗ to avoid confusion with the
matrix adjoint operation) and if T ∈ Mn (F ), then the map
h , in : Mn (F ∗ ) × Mn (F ) −→ Mn (B)
induced from the inner product map in the statement of the lemma, as in
Section 2.7, sends the pair (S, T ) to the Hilbert module inner product
hS ⊤ , T i ∈ Mn (B),
where S ∈ Mn (F ) is image of S ∈ Mn (F ∗ ) under the conjugate linear iso-
morphism Mn (F ∗ ) → Mn (F ) and S ⊤ is the transpose matrix. So the lemma
follows from the universal property of the Haagerup tensor product (cf. Sec-
tion 2.7) and the Cauchy-Schwarz inequality [BLM04, p.297]:
khS ⊤ , T ikMn(B) ≤ kS ⊤ kMn (F ) kT kMn (F ) = kSkMn (F ∗ ) kT kMn (F ) .

23
 Suppose next we are given a local adjunction
∼
=
ΦX,Y : KB (X ⊗A F, Y ) −−→ KA (X, Y ⊗B E).

Use it to identify E with F ∗ as in the proof of Theorem 3.10, and in this way
equip F ∗ with the structure of a Hilbert B-A-bimodule.

Lemma 3.18. Let F be a correspondence from A to B, and assume that F ∗

has been equipped with a Hilbert A-module structure making F ∗ and F local
adjoints. The formulas

ε : f1∗ ⊗ f2 7→ hf1 , f2 i and δ : f1 ⊗ f2∗ 7→ hf1∗ , f2∗i

define completely bounded bimodule maps

ε : F ∗ ⊗A F → B and δ : F ⊗B F ∗ → A.

Proof. The assertion about ε follows from the previous lemma and Theo-
rem 2.37. The assertion about δ follows from these results together with
Theorem 3.10.
The completely bounded map δ : F ⊗B F ∗ → A given by Lemma 3.18
induces, by Theorem 2.37, a natural bounded Hilbert space operator

(3.19) δX : F ⊗B F ∗ ⊗A X −→ X

for each X ∈ A H. Being a bounded operator between Hilbert spaces, δX has

an adjoint operator

(3.20) ηX : X −→ F ⊗B F ∗ ⊗A X,

and we obtain a natural transformation from the identity functor on A H to

the tensor product functor

F ◦ F ∗ : X 7→ F ⊗B FA∗ ⊗ X.

We need just such a natural transformation in order to prove that the

tensor product functor F ∗ : A H → B H is left adjoint to F (namely the unit
of the adjunction). We also need a natural transformation

εY : F ∗ ⊗A F ⊗B Y −→ Y

24
for every Y ∈ B H (this is the counit of the adjunction), and then we need to
show that the compositions
ηF (Y ) F (εY )
(3.21) F ⊗B Y −−−→ F ⊗B F ∗ ⊗A F ⊗B Y −−−→ F ⊗B Y
and
F ∗ (ηX ) εF ∗ (X)
(3.22) F ∗ ⊗A X −−−−→ F ∗ ⊗A F ⊗B F ∗ ⊗A X −−−−→ F ∗ ⊗A X
are the identity. See [ML98, IV.1 Theorem 2].
We shall define εY to be the map
ε⊗id ∼
=
F ∗ ⊗A F ⊗B Y −−−→
Y
A ⊗A Y −→ Y
obtained from the completely bounded inner product map ε : F ∗ ⊗A F → B in
Lemma 3.18. Having done so, (3.21) and (3.22) become effectively equivalent:
just reverse the roles of F and F ∗ to get from one to the other. So it remains
to prove (3.21).
Lemma 3.23. Let uλ be an approximate unit for the C ∗ -algebra KB (F, F ),
viewed as a net in F ⊗B F ∗ using the isomorphism of Theorem 2.38. The
∗
adjoint operator ηX = δX in (3.20) is given by the formula
ηX (x) = lim uλ ⊗ x,
λ→∞

where the limit exists in the weak topology on the Hilbert space F ⊗B F ∗ ⊗A X.
Proof. We need to prove that
lim huλ ⊗ x, vi = hx, δX vi
λ→∞

for all v ∈ F ⊗B F ∗ ⊗A X. To this end, let us show that

(3.24) hT ⊗ x, vi = hx, δX T ∗ vi,
where T ∈ KB (F, F ) and where on the right hand side of the identity the
adjoint operator T ∗ acts on the triple tensor product F ⊗B F ∗ ⊗A X by
acting on the first factor alone. To prove (3.24) it suffices to calculate with
elementary tensors
T = f1 ⊗ f2∗ and v = f3 ⊗ f4∗ ⊗ x1 ,
and this straightforward using the formulas for the Hilbert space inner prod-
ucts given in Definition 2.12.

25
Proof of Theorem 3.15. Let us show that for every Y ∈ B H the composition
(3.21) is equal to the identity map. Let f ∈ F , y ∈ Y and z ∈ F ⊗B Y . It
suffices to show that

(3.25) (F (εY )ηF (Y ) (f ⊗ y), z = f ⊗ y, z .

If we write w = F (εY )∗ z then (3.25) becomes the identity

(3.26) ηF (Y ) (f ⊗ y), w = f ⊗ y, z .

Using the formula for ηF (Y ) proved in the lemma, the left hand side of (3.26)
is
lim uλ ⊗ f ⊗ y, w ,
λ→∞

or in other words

(3.27) lim F (εY )(uλ ⊗ f ⊗ y), z .

λ→∞

From the definition of εY we have

F (εY ) : f1 ⊗ f2∗ ⊗ f3 ⊗ y 7→ f1 hf2 , f3 i ⊗ y,

and this implies that for every T ∈ KB (F, F ) ∼

= F ⊗B F ∗ one has

F (εY )(T ⊗ f ⊗ y) = T f ⊗ y.

Applying this to (3.27) we find that

(F (εY )ηF (Y ) (f ⊗ y), z = lim uλ f ⊗ y, z ,

λ→∞

and (3.25) follows from this.

The uniqueness of adjoint functors in the usual context of category theory
implies:

Corollary 3.28. If F admits local adjoints E and G, then the tensor product
functors
E⊗A , G⊗A : A H −→ B H
are canonically isomorphic.

26
3.5 Existence of Units and Counits
We continue to work with locally adjoint Hilbert modules F and F ∗ as in
Theorem 3.10, but let us return now from Hilbert space representations back
to Hilbert modules. For each Hilbert B-module Y we have a bounded, B-
linear map
idY ⊗ε : Y ⊗B F ∗ ⊗A F → Y ⊗B B
defined by identifying the Hilbert module tensor products with Haagerup
tensor products, and using the functoriality of the latter with respect to
completely bounded maps. A short computation using (3.12) shows that for
each Hilbert A-module X, and each T ∈ KA (X, Y ⊗B F ∗ ), the diagram

idY ⊗ε
Y ⊗B B o Y ⊗B FO ∗ ⊗A F
∼
= T ⊗idF

Y o X ⊗B F.
Φ−1
X,Y (T )

is commutative: thus ε is almost a counit for the adjunction, its only defect
being that it might not be adjointable. The action homomorphism η : A →
BB (F, F ) may similary be considered a kind of generalised unit, a point of
view justified by the following proposition.

Proposition 3.29. The following are equivalent:

(a) There exists a unit A → F ⊗B F ∗ for the local adjunction.

(b) The natural isomorphism ΦX,Y : KB (X ⊗A F, Y ) → KA (X, Y ⊗B F ∗ )

extends to a natural transformation BB (X ⊗A F, Y ) → BA (X, Y ⊗B F ∗ ).

(c) The action of A on F is through a ∗-homomorphism from A into KB (F, F ).

(d) The map δ : F ⊗B F ∗ → A of Lemma 3.18 is adjointable.

If A is unital, one may add a fifth equivalent condition:

(e) F is finitely generated as a right B-module.

When these conditions hold, the unit in (a) is equal to δ ∗ ; the natural trans-
formation in (b) is 1–1; and the ∗-homomorphism in (c) corresponds to η
under the canonical identification of KB (F, F ) with F ⊗B F ∗ .

27
Proof. That (a) implies (b) is clear. To prove (b) implies (c), let us first show
that ΦX,Y is one-to-one on BB (X ⊗A F, Y ). Fix a nonzero T ∈ BB (X ⊗A
F, Y ), and choose y ∈ Y such that T ∗ y 6= 0. Then the compact operator
(T ∗ y)∗ = y ∗ ◦ T ∈ KB (X ⊗A F, B) is nonzero, so its image under ΦX,B is
nonzero. By naturality we have

ΦX,B (y ∗ ◦ T ) = (y ∗ ⊗ 1F ∗ ) ◦ ΦX,Y (T ),

and so ΦX,Y (T ) 6= 0.
Now let a be an element of A = KA (A, A). By naturality, ΦA,F (a ⊗ 1F ) =
a ◦ ΦA,F (1F ) ∈ KA (A, F ⊗B F ∗ ). Since ΦA,F is 1–1, and maps KB (F, F )
surjectively onto KA (A, F ⊗B F ∗ ), we conclude that a ⊗ 1F is compact: i.e.,
A acts on F by compact operators.
Next, we show that (c) implies (d). Identifying F ⊗B F ∗ with KB (F, F )
as in (3.13), one finds the following formula for the A-valued inner product:

(3.30) hK1 , K2 iKB (F,F ) = δ(K1∗ K2 ).

Letting η : A → KB (F, F ) be the action homomorphism, we find that

hη(a), KiKB (F,F ) = δ(η(a∗ )K) = a∗ δ(K) = ha, δ(K)iA ,

where we have used that δ is a map of A-bimodules. Thus δ is adjointable,

with adjoint η.
Finally (d) implies (a), as follows. If δ is adjointable, then Lemma 3.18
implies that δ is a counit for the local adjunction Φ∗ , and so δ ∗ is a unit for
Φ.
If A is unital, then the equivalence of (c) and (e) follows from the well-
known fact ([BLM04, 8.1.27]) that the identity operator on a Hilbert B-
module F is compact if and only if F is finitely generated.

Corollary 3.31. Let F be a correspondence from A to B, admitting a local

adjoint F ∗ and a unit η : A → F ⊗B F ∗ . The C ∗ -algebra A decomposes as
the direct sum of two-sided ideals

A = AF ∗ ⊕ A⊥
F∗

where AF ∗ = span{hf1∗, f2∗ i | f1∗ , f2∗ ∈ F ∗ } and A⊥

F ∗ = ker(A → BB (F, F )).

28
Proof. By Proposition 3.29, the existence of a unit η implies that δ is ad-
jointable, with δ ∗ = η. Being a homomorphism of C ∗ -algebras, η has closed
range and so [Lan95, Theorem 3.2] gives

A = image δ ⊕ ker η = AF ∗ ⊕ A⊥
F∗

as (right) Hilbert A-modules. Both δ and ε are left A-linear, so the above is
a decomposition of C ∗ -algebras.
Recalling the relationship between units and counits, Proposition 3.29
and Corollary 3.31 immediately give:
Proposition 3.32. The following are equivalent:
(a) There exists an (adjointable) counit F ∗ ⊗A F → B.

(b) The natural isomorphism Φ−1 ∗

X,Y : KA (X, Y ⊗B F ) → KB (X ⊗A F, Y )
extends to a natural transformation BA (X, Y ⊗B F ∗ ) → BB (X ⊗A F, Y ).

(c) The action of B on F ∗ is through a ∗-homomorphism from B into the

C ∗ -algebra KA (F ∗ , F ∗ ).

(d) The map ε : F ∗ ⊗A F → B of Lemma 3.18 is adjointable.

If B is unital, one may add a fifth equivalent condition:
(e) F is finitely generated as a left A-module.
When these conditions hold, the counit in (a) is equal to ε; natural trans-
formation in (b) is 1–1; the ∗-homomorphism in (c) corresponds to ε∗ under
the canonical identification of KA (F ∗ , F ∗ ) with F ∗ ⊗A F ; and the C ∗ -algebra
B decomposes as the direct sum of two-sided ideals B = BF ⊕ BF⊥ .
Corollary 3.33. A local adjunction extends to an adjunction if and only if
it admits both a unit and a counit.
Proof. As we observed above (Proposition 2.21), the “only if” direction is a
standard fact about adjoint functors. For the converse, assume that a local
adjunction
ΦX,Y : KB (X ⊗A F, Y ) → KA (X, Y ⊗B F ∗ )
admits a unit and a counit. Proposition 3.29 gives an extension of Φ to a 1–1
natural transformation ΦX,Y : BB (X ⊗A F, Y ) → BA (X, Y ⊗B F ∗ ), while

29
Proposition 3.32 gives a 1–1 natural transformation ΨX,Y : BA (X, Y ⊗F ∗ ) →
BB (X ⊗A F, Y ) such that Φ and Ψ are mutually inverse on the subspaces of
compact operators. For each T ∈ BB (X ⊗A F, Y ) and y ∈ Y we have

y ∗ ◦ ΨX,Y ΦX,Y (T ) = ΨX,B ΦX,B (y ∗ ◦ T ) = y ∗ ◦ T

in KB (X ⊗A F, B). Since the maps y ∗ : Y → B separate the points of Y ,

the above calculation implies that ΨX,Y ◦ ΦX,Y = id. A similar computation
shows that ΦX,Y ◦ ΨX,Y = id, and so Φ is a natural isomorphism.
Combining the above results, we recover Theorem 2.24.

3.6 Uniqueness of Local Adjoints

Local adjoints are unique up to completely bounded bimodule isomorphism,
by Lemma 3.6. They are not, in general, unique as correspondences: see
Section 4.5. However, the stronger form of uniqueness does hold in the
presence of a counit:

Proposition 3.34. Let F be a correspondence from A to B admitting a local

adjoint E and an (adjointable) counit

ε : E ⊗A F −→ B.

If G is a second local adjoint to F , then E and G are isomorphic as corre-

spondences from B to A.

The proposition is an easy consequence of the following two lemmas.

Lemma 3.35. Let E and G be Hilbert A-modules and let T : E → G be a

bounded, A-linear operator (not necessarily adjointable). If K is any compact
(and hence adjointable) operator on the Hilbert module E, and if g ∈ G, then
the operator

SK,g : E −→ A
SK,g : e 7−→ hg, T KeiG

is adjointable (and indeed compact).

30
Proof. Since the compact operators are a norm-closed subspace of all oper-
ators, it suffices to prove the lemma in the special case where K = e1 ⊗ e∗2 .
In this case the operator in the lemma is the compact operator

SK,g = hg, T e1iG ⊗ e∗2 : E −→ A,

and so the proof is complete.

Lemma 3.36. Let E and G be correspondences from B to A. If the action

of B on E is through compact operators, then every bounded, A-B-linear
operator T : E → G is adjointable.

Proof. We need to show that for every g ∈ G there is some f ∈ E such that

hg, T eiG = hf, eiE

for every e ∈ E. By the factorization theorem (applied twice), the element

g may be written as a product b1 g1 a1 with b1 ∈ B and g1 ∈ G, and a1 ∈ A.
Next, if e ∈ E, then

hb1 g1 , T eiG = hg1, b∗1 T eiG = hg1 , T b∗1 eiG ,

where in the last step we are using the B-linearity of T . Since b∗1 is acting
on E as a compact operator, it follows from the previous lemma that we can
write
a∗ hg1 , T b∗1 eiG = hSa, eiE
for some operator S : A → E and every a ∈ A. Hence if we set f = Sa1 ,
then
hg, T eiG = hb1 g1 a1 , T eiG = a∗1 hg1 , T b∗1 eiG = hf, eiE ,
as required.
Proof of Proposition 3.34. This is an immediate consequence of the previous
lemma, together with Proposition 3.32, which tells us that B acts on E
through compact operators, and Lemma 3.6, which tells us that E and G are
isomorphic as B-A-operator bimodules.

Corollary 3.37. If F admits an adjoint E, then E is the unique local adjoint

to F .

31
3.7 Left and Right Indexes
The purpose of this section is to explicate the relationship between our results
and those of [KPW04]. Throughout this section, F denotes a correspondence
from A to B, and we assume that F ∗ has been equipped with an A-valued
inner product making it into a correspondence from B to A.
Following [KPW04, Definitions 2.8 and 2.9], we say that F is of finite
numerical index if there are positive constants l and r such that for all
f1 , . . . , fn ∈ F ,
X X
hfi , fi i ≤ l fi∗ ⊗ fi
B KA (F ∗ ,F ∗ )

and X X
hfi∗ , fi∗ i ≤r fi ⊗ fi∗ .
A KB (F,F )

The smallest l and r for which these inequalities hold are called, respectively,
the left numerical index and the right numerical index of F .
Lemma 3.38. The correspondence F is of finite numerical index if and only
if F ∗ is a local adjoint to F .
Proof. If F ∗ is a local adjoint to F then Lemma 3.18, combined with the
= F ⊗B F ∗ and KA (F ∗ , F ∗ ) ∼
identifications KB (F, F ) ∼ = F ∗ ⊗A F , implies that
the map
ε : KA (F ∗ , F ∗ ) → B, f1∗ ⊗ f2 7→ hf1 , f2 i
is bounded, and this ensures that F has finite left numerical index (equal
to kεk). Reversing the roles of F and F ∗ shows that F also has finite right
numerical index, equal to the norm of the map

δ : KB (F, F ) → A, f1 ⊗ f2∗ 7→ hf1∗ , f2∗ i.

To prove the converse, denote by E the bimodule F ∗ equipped with the

operator space structure coming from the A-valued inner product; F ∗ will
denote the same bimodule with its conjugate operator space structure. If F
has finite right numerical index r, then for f ∗ ∈ F ∗ one has
1/2 1/2 1/2
kf ∗kE = khf ∗ , f ∗ ikA ≤ rkf ⊗ f ∗ kKB (F,F ) = rkhf, f ikB = rkf ∗kF ∗ ,

showing that the identity map E → F ∗ has norm at most r. Similarly, if F

has finite left numerical index l, then the identity map F ∗ → E has norm at

32
most l. Two successive applications of [KPW04, Corollary 2.10] show that, if
F has left numerical index l and right numerical index r, then so does Mn (F )
(viewed as a correspondence from Mn (A) to Mn (B)). The above argument
then shows that the identity map Mn (E) → Mn (F ∗ ) has norm at most r, and
its inverse has norm at most l. Therefore the identity E → F ∗ is completely
bounded, and so E is a local adjoint to F by Theorem 3.10.
Suppose F has finite numerical index. The maps ε and δ may be extended
to normal maps
ε′′ : KA (F ∗ , F ∗ )′′ → B ′′ and δ ′′ : KB (F, F )′′ → A′′
between the enveloping von Neumann algebras. Following [KPW04, Defini-
tion 2.17], we say that F has finite left index if the image ε′′ (1) of the identity
lies in the multiplier algebra of B. The condition of having finite right index
is defined analogously, in terms of the map δ ′′ . Say F has finite index if it
has finite left index and finite right index.
Lemma 3.39. Suppose F has finite numerical index, so that F ∗ is a local
adjoint to F .
(a) F has finite left index if and only if there is a counit F ∗ ⊗A F → B for
the local adjunction.
(b) F has finite right index if and only if there is a unit A → F ⊗B F ∗ for
the local adjunction.
(c) F has finite index if and only if the local adjunction is an adjunction.
Proof. It is shown in [KPW04, Theorem 2.22] that F has finite left (respec-
tively, right) index if and only if B (resp. A) acts on F ∗ (resp. F ) by compact
operators. The asserted equivalences now follow from Propositions 3.29 and
3.32 and from Corollary 3.33
These results are summarised in the following table, which relates our
terminology to that of [KPW04]:
The functor ⊗A F has if and only if the correspondence F has
a local adjoint finite numerical index
a local adjoint with unit finite numerical index and finite right index
a local adjoint with counit finite numerical index and finite left index
an adjoint finite index

33
4 Examples
4.1 Commutative Examples
Let X be a compact Hausdorff space. The category of Hilbert modules over
the C ∗ -algebra C(X) is equivalent to the category of continuous fields of
Hilbert spaces over X (and adjointable operators between continuous fields)
via the functor that associates to a continuous field its Hilbert module of
continuous sections.
A continuous map ϕ : X → Y between compact Hausdorff spaces deter-
mines a homomorphism of C ∗ -algebras from C(Y ) to C(X). Let

F := C(X)

be the associated correspondence from C(Y ) to C(X). From the point of

view of continuous fields the tensor product functor

ϕ∗ := ⊗F : HC(Y ) → HC(X)

associated to F is given by pullback ϕ∗ of continuous fields of Hilbert spaces

along the map ϕ.

Coverings. Let us examine the correspondence F above in the case of a

surjective map between compact Hausdorff spaces.

Definition 4.1. Let π : X → Y be a continuous surjection of compact

Hausdorff spaces.

(a) The map π is a covering if it is an open map, and if #π −1 [y] (the number
of points in the pre image of y) is a finite and locally constant function
of y ∈ Y .

(b) The map π is a branched covering if it is open and if the number #π −1 [y]
is finite and uniformly bounded over y ∈ Y .

The following is a consequence of results of Pavlov and Troitsky [PT11],

combined with Theorem 2.24 and Proposition 4.7 (below):

Proposition 4.2. Let π : X → Y be a continuous surjection of compact

Hausdorff spaces.

34
(a) The pullback functor π ∗ admits an adjoint if and only if π is a covering.

(b) The functor π ∗ admits a local adjoint if and only if π is a branched

covering.

Proof. Theorem 2.24 implies that π ∗ admits an adjoint if and only if F =

C(X) is finitely generated and projective over C(Y ), which by [PT11, Theo-
rem 1.3] is equivalent to π being a covering. Proposition 4.7 (below) implies
that π ∗ admits a local adjoint if and only if there is a finite-index conditional
expectation C(X) → C(Y ), which by [PT11, Theorem 1.1] is equivalent to
π being a branched covering.
The C(Y )-valued inner product on the conjugate module F ∗ = C(X)∗ is
given by a formula of the kind
X
hf1∗ , f2∗i(y) = µy (x)f1 (x)f2 (x)
x∈π −1 [y]

where the weight functions µy : π −1 [y] → [0, 1] are determined (usually not
uniquely) by the branching of the cover over the point y. For example, if π is
the quotient map for a finite group action, one can take µy to be the constant
function 1/#π −1 [y] (see Proposition 4.8). For a construction of suitable µy
for a general branched covering, see [PT11, Proof of Theorem 4.3].

Open Subsets. Now let X be a compact Hausdorff space, and let U ⊂

X be an open subset. Let F = C0 (U). This is a correspondence from
C(X) to C0 (U), whose associated tensor-product functor HC(X) → HC0 (U ) is
restriction of continuous fields of Hilbert spaces from X to U (even though
U is not compact it is still true that HC0 (U ) is equivalent to the category of
continuous fields of Hilbert spaces over U).

Proposition 4.3. The functor of restriction to U has a local adjoint, namely

extension of continuous fields by zero, which is represented by E = C0 (U)
viewed as a correspondence from C0 (U) to C(X). The local adjoint is an
adjoint if and only if U is both open and closed in X.

Proof. This is a special case of Proposition 4.5, below.

35
Infinite-Dimensional Fibers. In the above examples, the bimodules that
occur, viewed as continuous fields of Hilbert spaces, have finite fiber dimen-
sions, namely fiber dimensions zero or one. Interesting examples with infi-
nite fiber dimension do not occur in the purely commutative context. For
instance, if H is a Hilbert space, viewed as a correspondence from C to C,
then H has a local adjoint if and only if H is finite-dimensional (in which
case the local adjoint is an adjoint).

4.2 Relation with Morita Equivalences

Let F be a correspondence from A to B, and suppose that F ∗ has been given
an A-valued inner product making it into a correspondence from B to A,
such that
(4.4) hx∗ , y ∗iz = xhy, zi
for all x, y, z ∈ F . This condition says that F and F ∗ restrict to mutually
inverse (strong) Morita equivalences between the ideals AF ∗ ⊆ A and BF ⊆
B: see [BLM04, 8.1.2].
Proposition 4.5. The correspondences F and F ∗ are local adjoints. They
are adjoints if and only if the ideals AF ∗ and BF are complemented.
Proof. The equality (4.4) implies that the two norms on F ∗ are equal: see
[BLM04, Lemma 8.1.15]. The equality (4.4) is also satisfied by the induced
inner products on Mn (F ) and Mn (F ∗ ), and so the identity map on F ∗ is a
complete isometry between the two operator structures; thus F and F ∗ are
local adjoints, by Theorem 3.10.
If F and F ∗ are adjoints, then the ideals AF ∗ and BF are complemented
by Corollary 3.31 and Proposition 3.32. Conversely, suppose that AF ∗ is
complemented in A. The formula (4.4) implies that the elements of AF ∗
act on F by B-compact operators. On the other hand, (4.4) implies that
F = KB (F, F )F = AF ∗ F , showing that the complementary ideal A⊥ F ∗ acts
by zero on F . Therefore A acts by compact operators on F , and Proposition
3.29 thus implies that the local adjunction admits a unit; a similar argument
proves the existence of a counit.
The local adjunctions for which the inner products satisfy (4.4) are pre-
cisely those for which the adjunction isomorphisms
∼
=
Φ : KB (X ⊗A F, Y ) −→ KA (X, Y ⊗B F ∗ )

36
are isomorphisms of ternary rings of operators, meaning that

Φ(RS ∗ T ) = Φ(R)Φ(S)∗ Φ(T )

for every R, S, T ∈ KB (X ⊗A F, Y ). (This is an immediate consequence of the

formula (3.12) for Φ; we owe this observation to an anonymous referee.) This
is in turn equivalent to the condition that the isomorphisms Φ be isometries,
or equivalently complete isometries, as follows from a theorem of Hamana,
Kirchberg and Ruan [BLM04, Corollary 4.4.6] and a naturality argument as
in Theorem 3.2.

4.3 Conditional Expectations

Let A be a ∗-subalgebra of a C ∗ -algebra C, and assume that A contains an
approximate unit for C. Recall that a conditional expectation from C to A
is a positive, idempotent, A-A-bimodule map ϕ : C → A.
Following Frank and Kirchberg [FK98, Theorem 1], we make the following
definition.

Definition 4.6. A conditional expectation ϕ : C → A is of finite index if

one of the following equivalent conditions is satisfied:

(a) ϕ(c∗ c) 6= 0 whenever c 6= 0, and C is complete in the norm

kck2ϕ = kϕ(c∗ c)k.

(b) There is a λ ≥ 0 such that λϕ − idC : C → C is a positive map.

(c) There is a κ ≥ 0 such that κϕ − idC : C → C is a completely positive

map.

Proposition 4.7. Let F be a correspondence from A to B and suppose that

in fact A ⊆ KB (F, F ). The correspondence F admits a local adjoint if and
only if there is a finite-index conditional expectation ϕ : KB (F, F ) → A.

Proof. Suppose given a finite-index conditional expectation as in the state-

ment of the proposition. Define an A-valued inner product on F ∗ by

hf1∗ , f2∗ i = ϕ(f1 ⊗ f2∗ ).

37
Condition (a) in Definition 4.6 ensures that F ∗ is a correspondence from B
to A. Denote by E this correspondence, with the operator space structure
it inherits from the A-valued inner product. The canonical map F ∗ → E
is completely bounded because ϕ, like all conditional expectations, is com-
pletely bounded. The finite index condition (c) implies that the inverse map
E → F ∗ is also completely bounded, since for each f ∗ = [fi,j
∗
] ∈ Mn (F ∗ ) we
have the estimate
κkf ∗ k2Mn (E) = kMn (κϕ)(f ⊗Mn (B) f ∗ )kMn (A) ≥ kf ⊗Mn (B) f ∗ kMn (KB (F,F ))
= kf ∗ k2Mn (F ∗ ) .
Thus Theorem 3.10 implies that F and E are local adjoints.
Conversely, suppose that F has a local adjoint, and again denote by E
the conjugate space F ∗ with the operator space structure it inherits from
the given A-valued inner product. The A-valued inner product determines a
map of operator spaces
F ⊗B F ∗ −→ A
thanks to Lemma 3.18, and so thanks to Theorem 2.38, a (completely) pos-
itive map
ψ : KB (F, F ) −→ A
of A-A-bimodules that sends the elementary compact operator f1 ⊗ f2∗ on F
to hf1∗ , f2∗ iA . P
Now let T = ni=1 xi ⊗ yi∗ ∈ K(F ). We compute that
n
X
∗
ψ(T T ) = hxi , xj iyj∗, yi∗ A
i,j=1

where the inside inner product is the B-valued inner product of F , and hence,
after a further computation, that kψ(T ∗ T )k is the norm of the element
 1  
hx1 , x1 i . . . hx1 , xn i 2 y1∗
 .. ..   .. 
 . .   .  ∈ Mn (E)
hxn , x1 i . . . hxn , xn i yn∗
On the other hand a third computation shows that the Mn (F ∗ )-norm of this
element is kT ∗ T k. Using the fact that the norms on E and F ∗ are completely
equivalent we find that
kψ(T ∗ T )k ≥ constant · kT ∗ T k

38
for some constant independent of T .
The restriction of the map ψ to A ⊆ KB (F, F ) is multiplication by some
positive and central element of the multiplier algebra of A. The computation
above shows that this element is invertible. Adjusting ψ by multiplying with
the inverse of this element we obtain a finite-index conditional expectation,
as required.
Here is an example that we shall use when analyzing parabolic induction
in Section 5. Let W be a finite group acting by ∗-automorphisms on a
C ∗ -algebra B, and also acting projectively by twisted automorphisms on a
Hilbert B-module F . This means that associated to each w ∈ W there is a
C-linear operator
Uw : F −→ F
such that

(i) Uw (f b) = Uw (f )w(b) for all f ∈ F , all b ∈ B, and

(ii) hUw (f1 ), Uw (f2 )i = w(hf1 , f2 i) for all f1 , f2 ∈ F .

and such that moreover

(iii) Uw1 (Uw2 (f )) = Uw1 w2 (f )u(w1 , w2 )

for some unitary u(w1 , w2 ) in the multiplier algebra of B. The formula

w(T ) = Uw T Uw−1 : F −→ F

defines a genuine action of W by automorphisms on the C ∗ -algebra KB (F, F ).

Proposition 4.8. Let W be a finite group acting projectively by twisted au-

tomorphisms on a Hilbert B-module F , as above, and let A be the fixed-point
algebra KB (F, F )W . Then F , considered as a correspondence from A to B,
admits a unique local adjoint E for which the canonical action of W is iso-
metric. The isomorphisms

ΦX,Y : KB (X ⊗A F, Y ) → KA (X, Y ⊗B E)

for this local adjunction satisfy

1/2
kΦX,Y kcb = 1 and kΦ−1
X,Y kcb ≤ |W | .

39
Remark 4.9. The canonical action on E referred to in the statement of the
proposition is the one coming from the identification of E with F ∗ that the
local junction implies. And the term isometric refers to the given Banach
space structure on E.
Proof of the Proposition. The conditional expectation
1 X
ϕ : KB (F, F ) → A, ϕ(T ) := w(T )
|W | w∈W

has finite index, since if T ≥ 0 then

X
|W |ϕ(T ) − T = w(T ) ≥ 0.
w6=1

Therefore the correspondence F admits a local adjoint, namely the B-A-

bimodule F ∗ , with A-valued inner product

hf1∗ , f2∗ i = ϕ(f1 f2∗ ).

The action of W on E is isometric, because ϕ is W -invariant.

Suppose G is another local adjoint to F , and let ψ : KB (F, F ) → A denote
the finite-index conditional expectation induced by the inner product on G
as in Proposition 4.7. The action of W on G is isometric if and only if ψ is
W -invariant, which is to say, if and only if ψ = ϕ. Thus E is unique.
For the bounds on the norms, recall from Corollary 3.14 that

kΦX,Y kcb ≤ kΦA,B kcb .

The same argument shows that kΦ−1 −1

X,Y kcb ≤ kΦA,B kcb . The conditional ex-
pectation ϕ is completely bounded, with cb-norm equal to 1, and this implies
that ΦA,B has cb-norm 1. The map |W |ϕ−id : B → B is completely positive,
and this implies that
1/2
kΦ−1
A,B kcb ≤ |W | ,
as required.

4.4 Direct Sums

Direct sums of local adjoints are not, in general, local adjoints: for example,
C is an adjoint correspondence from C to C, but the countable direct sum

40
C∞ ∼= ℓ2 (N) does not possess a local adjoint as correspondence from C to C.
We may
L∞however consider C∞ as a correspondence on the C ∗ -algebraic direct
sum C∼ = C0 (N), and viewed in this way C∞ does have a local adjoint,
as we shall see.
Suppose we are given two C ∗ -algebra direct sums
M M
A= Aα and B = Bβ .
α β

Suppose we are given a map β 7→ ι(β) from the index set for the decompo-
sition of B to the index set for the decomposition of A. For each index β let
Fβ be a correspondence from Aι(β) to Bβ . The external direct sum is
M Y
F = Fβ = {(fβ ) ∈ Fβ : kfβ k → 0 as β → ∞},
β

with the obvious A-B-bimodule structure and B-valued inner product.

We shall assume throughout the rest of this section that the map ι on
indices is finite-to-one; indeed we shall assume that it is uniformly finite-to-
one, in the sense that

(4.10) sup #ι−1 [β] < ∞.

β

Proposition 4.11. Let Fβ be a family of correspondences from Aι(β) to Bβ ,

as above, and assume that condition (4.10) holds. Suppose that each Fβ has
a local adjoint Eβ , such that the isomorphisms ϕβ : Fβ∗ → Eβ satisfy

sup{kϕβ kcb , kϕ−1

β kcb } < ∞.
β

Then F has a local adjoint.

Proof. Let
M Y
E= Eβ = {(eβ ) ∈ Eβ : keβ k → 0 as β → ∞},
β

This is a correspondence from A to B, and the map

M M M
ϕ= ϕβ : Fβ∗ → Eβ
β β

41
is an isomorphism of B-A-bimodules. Moreover
M
F∗ = Fβ∗ ,
β

and so ϕ is also an isomorphism of operator spaces.

4.5 Miscellany
4.5.1 Units and Counits
Let B be an ideal in A, and consider F = B as a correspondence from A to
A, with the obvious bimodule structure and the inner product hb1 , b2 i = b∗1 b2 .
Then F is locally adjoint to itself, but the local adjunction admits neither a
unit nor a counit unless B is complemented in A. Replacing one of the two
acting copies of A by B gives an example of a local adjunction with a unit
but no counit, or vice versa.

4.5.2 Distinct Local Adjoints

Let B = C[−1, 1], and let A = C[0, 1], considered as a subalgebra of B via
the surjective map x 7→ |x|, [−1, 1] → [0, 1]. Let F = B, considered as a
correspondence from A to B. Define finite-index conditional expectations
ϕ, ψ : B → A by
1 1 2 1
ϕ(b)(y) := b(y) + b(−y) and ψ(b)(y) := b(y) + b(−y).
2 2 3 3
The associated correspondences E = Bϕ and G = Bψ from B to A are both
locally adjoint to F , but E and G are not isomorphic as Hilbert modules.
Indeed, if they were isomorphic, then by polar decomposition we could find
a unitary isomorphism of bimodules U : E → G. Unitarity means that
ϕ(b) = ψ(bU(1)∗ U(1)) for every b ∈ B, but this equality implies that the
function U(1)∗ U(1) is discontinuous at the origin.

4.5.3 Action by Compact Operators

Let Ai = Bi = C for i = 1, 2, 3, . . .. Let Fi = Ci , with its canonical Hilbert
space structure, and let Ei = (Ci )∗ , with its canonical Hilbert space structure.
Considered as a correspondence from C to C, each Fi has Ei as its adjoint

42
 L
(cf. Example 2.25). The direct sum F = Fi does not, however, admit
a local adjoint. Indeed, the universal property of direct sums impliesLthat
the only
L ∗possibility
L for a local adjoint would be the direct sum E = L Ei i,
but Fi and Ei are not isomorphic
L as operator bimodules
L over iC
(cf. Example 2.35). Note that A = Ai and B = Bi act as compact
operators on F and E respectively, showing that condition (b) in Theorem
2.24 is independent of the other conditions.

4.5.4 Simple C*-Algebras

Finally, here is a short remark concerning simple C ∗ -algebras.

Proposition 4.12. Let Φ be a local adjunction as in Definition 3.1.

(a) If A is simple, Φ admits a unit.

(b) If B is simple, Φ admits a counit.

(c) If A and B are simple, Φ extends to an adjunction.

Proof. Item (a) follows from Proposition 3.29 and [KPW04, Corollary 2.26].
Item (b) follows by symmetry from (a). Item (c) follows from (a) and (b) by
Corollary 3.33.

5 Parabolic induction
In this section we shall analyze the correspondence associated in [Cla13] and
[CCH16] to parabolic induction of tempered unitary group representations.
We shall work with a fixed real reductive group G and standard parabolic
subgroup P with Levi factor L and unipotent radical N (thus P = LN is a
semidirect product, with L acting on N). We shall be following the precise
conventions of [CCH16, Section 3], but for instance we might take G to
be GL(n, R), and P to be a block upper triangular subgroup of GL(n, R),
decomposed as a semidirect product of the block upper diagonal matrices
acting on the unipotent block upper triangular matrices (with identity matrix
blocks down the diagonal).
The groups G and L act on the homogenous space G/N on the left and
right, respectively. The vector space Cc∞ (G/N) carries left and right actions

43
of G and L, respectively1 and it may be completed to a Hilbert Cr∗ (L)-module
Cr∗ (G/N) on which Cr∗ (G) acts on the left through bounded adjointable oper-
ators. See [Cla13, Section 2] or [CCH16, Section 4]. It is this correspondence,
from Cr∗ (G) to Cr∗ (L), that we shall study in this section.
The importance of the correspondence Cr∗ (G/N) is that the associated
tensor product functor from Hilbert space representations of Cr∗ (L) to Cr∗ (G)
may be identified with the functor of parabolic induction from tempered
unitary representations of L to tempered unitary representations of G. See
[Cla13, Section 3]. In [CCH16, Definition 8.2] we defined a correspondence
Cr∗ (N\G) from Cr∗ (L) to Cr∗ (G) (the construction is outlined below). Here
we shall prove the following result.

Theorem 5.1. The correspondences Cr∗ (G/N) and Cr∗ (N\G) are local ad-
joints.

Proof. The C ∗ -algebras A = Cr∗ (G) and B = Cr∗ (L) both decompose into
direct sums M M
A= Aα and B = Bβ .
α β

The index sets are “associate classes” of pairs consisting of a parabolic sub-
group in G or L, respectively, and a discrete series representation of the
compactly generated part of the parabolic. See [CCH16, Proposition 5.16
and Theorem 6.8].
There is a finite-to-one map ι from the index set for B to the index set for
A which comes from the enlargement of standard parabolic subgroups of L
to standard parabolic subgroups of G (for instance in the case of GL(n, R),
a parabolic subgroup of L consists of a block diagonal group with each block
itself a block upper triangular group, and this may be naturally extended to
a block upper triangular subgroup of GL(n, R)). The enlargement process
preserves Levi factors. The map ι of associate classes is finite-to-one because
non-associate pairs for B can become associate when enlarged to become
pairs for A. There is a uniform bound on the size of the point inverse images
of ι.
The correspondence F = Cr∗ (G/N) from A to B decomposes into an
1
In the case of the L-action, there is an adjustment to the standard action that reflects
the fact that we should identify Cc (G/N ) with the space of compactly supported half-
densities on G/N .

44
orthogonal direct sum M
F = Fβ
β

with the same index set as B. The inner products of elements in Fβ lie within
the summand Bβ ⊆ B. The action of A on Fβ factors through the projection
onto the summand Aι(β) ⊆ A.
The structure of Fβ is a little complicated to recall here (see [CCH16,
Section 7]). But what is important for us is the structure of the C ∗ -algebra
KB (Fβ , Fβ ), which is given by a formula of the type

KB (Fβ , Fβ ) ∼
= KC0 (Zβ ) (Mβ , Mβ )Wβ (L) ,

where, to use notation that has been streamlined from [CCH16],

(a) Mβ is a continuous field of parabolically induced tempered unitary rep-

resentations of G over a locally compact space Zβ (or in other words a
correspondence from Cr∗ (G) to C0 (Zβ )).

(b) Wβ (L) is finite group acting on Zβ and also acting projectively by twisted
automorphisms on Mβ .

The action of A = Cr∗ (G) on Fβ is through the inclusion of the summand

Aι(β) as a fixed point algebra

Aι(β) = KC0 (Zβ ) (Mβ , Mβ )Wβ (G) ,

where Wβ (G) is a finite group containing Wβ (L) as a subgroup that also acts
on Zβ , and on Mβ projectively, by twisted automorphisms.
It follows from Proposition 4.8 that Fβ has a local adjoint Fβ∗ , whose Aι(β) -
valued inner product is given by averaging over Wβ (G). The finite groups
Wβ (G) are uniformly bounded in size, and so it follows from Proposition 4.11
that F = ⊕β Fβ has a local adjoint too, namely the direct sum ⊕β Fβ∗ . The
correspondence Cr∗ (N\G) was defined in [CCH16] to be precisely this direct
sum.
We obtain from Theorem 3.15 the following consequence:

Theorem 5.2. The functor of parabolic induction, from tempered unitary

Hilbert space representations of L to tempered unitary representations of G,
possesses a two-sided adjoint.

45
 Let us note for comparison that, while there is a C ∗ -correspondence
C (G/N) implementing parabolic induction for full group C ∗ -algebras (see
∗

[Cla13]), there is no analogue of Theorem 5.1 in this setting:

Proposition 5.3. Let G = SL(2, R), and let P = LN be the parabolic
subgroup of upper-triangular matrices. The C ∗ -correspondence C ∗ (G/N),
from C ∗ (G) to C ∗ (L), does not possess a local adjoint.
Proof. Suppose that C ∗ (G/N) does have a local adjoint. Since L is amenable
we in fact have C ∗ (G/N) = Cr∗ (G/N), and the left action of C ∗ (G) fac-
tors through the quotient mapping C ∗ (G) → Cr∗ (G). Now Cr∗ (G) acts
on Cr∗ (G/N) by compact operators (see [CCH16, Proposition 4.4]), and so
Proposition 3.29 and Corollary 3.31 imply that the kernel J of the action map
C ∗ (G) → BC ∗ (L) (C ∗ (G/N), C ∗ (G/N)) is a direct summand in C ∗ (G). The
closed subset of the unitary dual G b corresponding to the ideal J is precisely
the principal series, i.e. the set of irreducible constituents of parabolically in-
duced representations of G. The principal series is not open in G, b as there is
a complementary series of representations whose closure intersects the prin-
cial series (see e.g. [Mil71]). Therefore J is not a direct summand of C ∗ (G),
and C ∗ (G/N) has no local adjoint.
We conclude with a remark. The determination of the structure of the
correspondence Cr∗ (G/N) relies very heavily on the classification results on
the tempered representations due to Harish-Chandra, Langlands and others.
It is an interesting open problem to approach the construction of a locally
adjoint correspondence from a more geometric starting point, without relying
so heavily on representation theory.

References
[Bee82] W. Beer. On Morita equivalence of nuclear C ∗ -algebras. J. Pure
Appl. Algebra, 26(3):249–267, 1982.

[Ber87] J. Bernstein. Second adjointness for

repre-
sentations of reductive p-adic groups. Draft:
http://www.math.uchicago.edu/~mitya/langlands.html, 1987.

[Ble97] D. P. Blecher. A new approach to Hilbert C ∗ -modules. Math.

Ann., 307(2):253–290, 1997.

46
[BLM04] D. P. Blecher and C. Le Merdy. Operator algebras and their
modules—an operator space approach, volume 30 of London Math-
ematical Society Monographs. New Series. The Clarendon Press,
Oxford University Press, Oxford, 2004. Oxford Science Publica-
tions.
[CCH16] P. Clare, T. Crisp, and N. Higson. Parabolic induction and restric-
tion via C ∗ -algebras and Hilbert C ∗ -modules. Compositio Mathe-
matica, FirstView:1–33, 2 2016.
[CH16] T. Crisp and N. Higson. Parabolic induction, categories of repre-
sentations and operator spaces. In to appear in Operator algebras
and their applications: a tribute to Richard V. Kadison, volume
671 of Contemp. Math. Amer. Math. Soc., Providence, RI, 2016.
[Cla13] P. Clare. Hilbert modules associated to parabolically induced rep-
resentations. J. Operator Theory, 69(2):483–509, 2013.
[ER00] E. G. Effros and Z-J. Ruan. Operator spaces, volume 23 of London
Mathematical Society Monographs. New Series. The Clarendon
Press, Oxford University Press, New York, 2000.
[FK98] M. Frank and E. Kirchberg. On conditional expectations of finite
index. J. Operator Theory, 40(1):87–111, 1998.
[GLR85] P. Ghez, R. Lima, and J. E. Roberts. W ∗ -categories. Pacific J.
Math., 120(1):79–109, 1985.
[KPW04] T. Kajiwara, C. Pinzari, and Y. Watatani. Jones index theory for
Hilbert C ∗ -bimodules and its equivalence with conjugation theory.
J. Funct. Anal., 215(1):1–49, 2004.
[Lan95] E. C. Lance. Hilbert C ∗ -modules, volume 210 of London Mathe-
matical Society Lecture Note Series. Cambridge University Press,
Cambridge, 1995. A toolkit for operator algebraists.
[Mil71] D. Miličić. Topological representation of the group C ∗ -algebra of
SL(2, R). Glasnik Mat. Ser. III, 6(26):231–246, 1971.
[ML98] S. Mac Lane. Categories for the working mathematician, volume 5
of Graduate Texts in Mathematics. Springer-Verlag, New York,
second edition, 1998.

47
[Mor65] K. Morita. Adjoint pairs of functors and Frobenius extensions.
Sci. Rep. Tokyo Kyoiku Daigaku Sect. A, 9:40–71 (1965), 1965.

[PT11] A. A. Pavlov and E. V. Troitskii. Quantization of branched cov-

erings. Russ. J. Math. Phys., 18(3):338–352, 2011.

[Ren10] D. Renard. Représentations des groupes réductifs p-adiques,

volume 17 of Cours Spécialisés [Specialized Courses]. Société
Mathématique de France, Paris, 2010.

48