Projective geometry of the Poincaré disk of a C ∗-algebra

E. Andruchow, G. Corach, L. Recht

August 13, 2019

Abstract
We study the Poincaré disk D = {a ∈ A : kak < 1} of a C∗ -algebra A from a projective
point of view: D is regarded as an open subset of the projective line AP, the space of
complemented rank one submodules of A2 . We introduce the concept of cross ratio of
four points in AP. Our main result establishes the relation between the exponential map
Expz0 (z1 ) of D (z0 , z1 ∈ D) and the cross ratio of the four-tuple
δ(−∞), δ(0) = z0 , δ(1) = z1 , δ(+∞),
where δ is the unique geodesic of D joining z0 and z1 at times t = 0 and t = 1, respectively.

2010 MSC: 46L05, 58B20, 22E65, 46L08

Keywords: Projective line, Poincaré disk, C∗ -algebra.

1 Introduction
Let A be a unital C∗ -algebra and G the group of invertible elements of A and UA the subgroup
of unitary elements. The set G+ of all positive elements of G can be thought as a family of
inner products on a Hilbert space H where A is faithfully represented. G+ is a homogeneous
reductive space of the group G, by means of the action (g, a) 7→ (g ∗ )−1 ag −1 . The differential
geometry of G+ has been studied in [5]. If we consider that G+ is the configuration space of
a mechanical system, then its tangent bundle T G+ is the space of states of the system. These
ideas motivated the paper [1]. The immersion T G+ → A given by (X, a) 7→ X + ia, where
X = X ∗ ∈ (T G+ )a , identifies T G+ with the Poincaré half-space of A
1
H = {h ∈ A : Im(z) = (h − h∗ ) ∈ G+ }.
2i
H is a homogeneous reductive  space of unitary group U(θH ) of the quadratic form θH :
 the 
x1 y1
A2 × A2 → A given by θH ( , ) = −ix∗1 y2 + ix∗2 y1 . The geometry of H as a
x2 y2
homogeneous space of U(θH ) is the theme of [1]. The space H is equivalent, as a homogeneous
reductive space, to the Poincaré disk
D = {z ∈ A : kzk < 1},
with its corresponding group of movements U(θ), where
   
x1 y1
θ( , ) = x∗1 y1 − x∗2 y2 .
x2 y2

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 In this paper we study the disk D from a projective viewpoint. One can say that in [1] we
studied the intrinsic geometry of H or D, as a homogeneous reductive space. We study the
relative geometry of D as a subspace of the projective line AP, and its relation to the intrinsic
geometry mentioned above. The main tool is the notion of an operator cross ratio, defined and
studied by M. I. Zelikin in [17]. This notion of a cross ratio has an interesting geometrical
interpretation when one studies the geodesic metric in D (see Theorems 9.9 and 10.1). In order
to study D from a projective viewpoint we introduce the projective line AP of A, and regard
D as an open subset of AP. The group U(θ) appears as a group of projective transformations
which preserve D. The concept of a cross ratio of four points in AP is based on Zelikin’s study
[17].

Figure 1

Here the plane represents A2 and the lines are elements of AP, i.e. complemented submodules
of A2 generated by a non-zero element. The cross ratio of `∞ , `0 , `1 and `2 (in this precise order)
has the following geometric meaning: in the projective line AP we choose `∞ as the point of
infinity, then AP \ {`∞ } turns into an affine line (L in the figure above). In this affine line, the
line `0 defines a point 0, and the affine line turns into a vector line. In this vector line we choose
a point u, and the vector line L turns into a scalar line (identified with A by means of the basis
u). Finally, if in this scalar line we choose a fourth element `2 6= `0 , which has a coordinate in the
basis u. Classically, this scalar is the cross ratio of `∞ , `0 , `1 , `2 . These features can be observed
in Figure 1, in the line L. In Zelikin’s construction, the cross ratio is the endomorphism of `1
that sends x to y = ϕ(x) = xλ, λ ∈ A and this scalar λ is the coordinate of y in the basis u of
L. In this paper, the cross ratio of these four lines will be an endomorphism ϕ of `1 . The map
ϕ associates to each x ∈ `1 the point ϕ(x) = y, obtained by two succesive projections: first to
`0 , parallel to `∞ ; next to `1 , parallel to `2 . This is clear in the figure above.
We introduce a fibre bundle Γ whose fibers are C∗ -algebras (isomorphic to A), over the base
space D. In the tangent bundle T D, we define an inner product with values in Γ (similar to a
Riemannian structure). If z ∈ D and X, Y ∈ (T D)z , we denote this inner product by hX, Y iz .
It is an element in Γz (the fiber of Γ over z). Then, by definition, the cross ratio of four points
in D is an element of Γ.
On the other hand, we compute explicitly the exponential map of the natural connection
over T D, which is bijective at every tangent space, and its inverse Logz : D → (T D)z . We
1/2
consider the element hLogz (w), Logz (w)iz ∈ Γz as a distance operator from z to w in D.
As in the classic case (A = C), in order to obtain the Poincaré distance d(0, z) in the disk,
as a cross ratio of lines, we have to compute the limit points at t = ±∞ of the geodesic joining

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 0 and z, and perform the cross ratio of the four points in the adequate order. In the C∗ -algebra
case, these points at infinity fall off from A, and belong to the enveloping von Neumann algebra
A∗∗ . Nevertheless, cross ratio morphisms exist in A. The problem remains, that they may not
be unique. We show that there is a natural choice. If z is invertible, it is the unique possible
choice. If A is a von Neumann algebra, by a result by Dixmier and Maréchal [8], invertible
elements are dense in D in the strong operator topology. We show that the cross ratio is strong
operator continuous in the appropriate sense, which justifies our choice in this setting (A a von
Neumann algebra).
In Theorems 10.1 and 10.2 we prove our main result, which states the relationship between
1/2
hLogz (w), Logz (w)iz and the cross ratio of `−∞ , `z , `w , `∞ . Here ±∞ denote the limits

SOT − lim γ(t),

t→±∞

where γ is the geodesic from z to w in time 1, and the algebra A is considered in its universal
representation. The formulas in Theorems 10.1 and 10.2 show a fundamental connection between
the differential geometry of the Poincaré disk D as a homogeneous reductive space of the group
U(θ), with the projective geometry of D as a subspace of the projective line AP.
The contents of the paper are the following. In Section 2 we define the form θ, the group
U(θ) of θ-unitary elements, and the unit sphere K of θ, and we study their basic properties. In
Section 3 we study the space Qρ of θ-symmetric idempotents of M2 (A) which decompose the
form θ. We examine the structure of Qρ , in particular, a natural fibration K → Qρ . In Section
4 we study the Poincaré disk as a homogeneous space of U(θ) and establish an equivariant
diffeomorphism between Qρ and D. In Section 5 we define the hyperbolic part APθ of the
projective line AP define the projective line AP of A, and introduce the operator cross ratio of
four submodules, following ideas introduced by Zelikin [17] for closed linear subspaces. Here it
is defined as a set of module homomorphisms, which may be empty or contain more than one
element, without further assumptions on the submodules. We use a natural bijection from Qρ
onto APθ to understand the geometry of the latter space. In Section 6 we recall from [1] the
geodesics of D. In Section 7 we compute the limit points of geodesics at t = ±∞. In particular,
we show that the partial isometry of z ∈ D coincides with the SOT − limt→∞ δ(t), where δ is
the geodesic joining 0 and z. To this effect, the Borel subgroup Bθ of U(θ) will be useful,

g − ĝ
 
g + ĝ
− ĝx − ĝx
 2 2 ∗
Bθ = { g −  : g ∈ G, x = −x}.

ĝ g + ĝ
+ ĝx + ĝx
2 2
In Section 8 we introduce the U(θ)-invariant Finsler metric in APθ .
In Section 9 we consider the cross ratio of the points `−∞ , `0 , `, `+∞ , where `0 is the submod-
ule corresponding to the origin in D, ` 6= `0 is an arbitrary point in APθ , and `∓∞ are the ∓∞-
limit points of the geodesic starting at `0 at t = 0 and reaching ` at t = 1. We show that this set
is always non empty, and that there is a natural choice of a specific homomorphism cr(`0 , `), or
cr(0, z) if ` ' z ∈ D (which is the only possible choice for a strongly dense subset of ` ∈ APθ , if
A is a von Neumann algebra). We finish Section 9 by giving a first glimpse of our main result,
which relates cr(0, z) with the metric geometry of D (or equivalently, APθ ). Namely,
1
kcr(0, z)kB(`z ) = d(0, z),
2

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 where kcr(0, z)kB(`z ) stands for the norm of the endomorphism cr(0, z) : `z → `z .
In Section 10 we briefly introduce the coefficient bundle over APθ (using the model Qρ ),
and an Hermitian structure on APθ , with values in the coefficient bundle. This is done in order
to present the main results of this paper, which are the formulas given in Theorem 10.1 and
Theorem 10.2, relating the cross ratio cr(`1 , `2 ) with the logarithm map Log`1 `2 .
In Section 11 we consider a relevant example, namely, when there exists a C∗ -subalgebra B
of the center of A and a conditional expectation tr : A → B satisfying tr(xy) = tr(yx) for all
x, y ∈ A; an important particular case is when A is commutative. In this case the computations
are much simpler, and the formula relating the cross ratio and the logarithm is

e|Log`1 `2 | = cr(`1 , `2 ).

2 Preliminaries
Let A be a unital C∗ -algebra and consider A2 = A × A as a right C∗ -module over A, with
A-valued inner product
hx, yi = x∗1 y1 + x∗2 y2 ,
   
x1 y1
for x = ,y= ∈ A2 . The space A2 is endowed with the usual C ∗ -module norm
x2 y2
kxk = khx, xik1/2 = kx∗1 x1 + x∗2 x2 k1/2 . We shall use the same symbol for norms of scalars a ∈ A
or vectors x ∈ A2 .
The C∗ -algebra of adjointable operators of A2 identifies with the algebra M2 (A) of 2 × 2
matrices with entries in A, acting by left multiplication in A2 . We refer the reader to the books
by Lance [11] and Manuilov and Troitsky [12] for all matters concerning C ∗ -modules. The book
[4] by Beltita will be our reference for basic facts on infinite dimensional homogeneous spaces.
The group of invertible elements of M2 (A) is denoted by Gl2 (A) and U2 (A) denotes the
subgroup of unitary elements of M2 (A).
A main ingredient in this paper is the form θ : A2 × A2 → A,

θ(x, y) = x∗1 y1 − x∗2 y2 .

 
1 0
Note that θ(x, y) = hx, ρyi, where ρ is the reflection ∈ U2 (A). Observe that for
0 −1
every ã ∈ M2 (A) there exists a unique ã] ∈ M2 (A) such that θ(x, ãy) = θ(ã] x, y) for all
x, y ∈ A2 . We call ã] the θ-adjoint of ã. Clearly ã] = ρã∗ ρ.

Definition 2.1. We say that g̃ ∈ Gl2 (A) is θ-unitary if ã] = ã−1 , i.e., if ρã∗ ρã = 1 = ãρã∗ ρ.
The subgroup of Gl2 (A) of all θ-unitary matrices is denoted by U(θ). These matrices are called
Moebius transformations or general symplectic matrices in classical contexts [15], [14], [9], [7],
[16], [13].

The factors of the polar decomposition of elements of U(θ) remain in U(θ). Before we prove
this fact, we state two identities which will be useful.
 
u1 0
Proposition 2.2. 1. The unitary elements of U(θ) are the matrices of the form
0 u2
for u1 , u2 ∈ UA .

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 2. The set U(θ)+ of positive elements in U(θ) is the set of all matrices

(1 + b∗ b)1/2 b∗
 
, for b ∈ A.
b (1 + bb∗ )1/2

3. If ã ∈ U(θ)+ and n is a positive integer, then ã1/n ∈ U(θ).

Proof. For the first assertion, note that if ã ∈ Gl2 (A) satisfies ρã∗ ρ = ã−1 = ã∗ , then it follows
that ρã = ãρ, which proves the statement.
The second assertion is proved in a straightforward manner, using the algebraic relations
involved.
If ã ∈ U(θ)+ then ρãρ = ã−1 . Then ρã1/n ρ is a positive invertible matrix such that
(ρã ρ)n = ρ(ã1/n )n ρ = ã−1 . By uniqueness of the positive nth root, it follows that ρã1/n ρ =
1/n

ã−1/n .

Corollary 2.3. If ã ∈ U(θ), then both factors in the polar decomposition ã = |ã|ũ belong to
U(θ).

Proof. First note that if ã ∈ U(θ), then ã∗ ∈ U(θ). Then |ã|2 = ã∗ ã ∈ U(θ). Also

(ρ|ã|2 ρ)2 = ρkã|2 ρ = ρã∗ ãρ = (ρã∗ ρ)(ρãρ) = ã−1 (ã∗ )−1 = (ã∗ ã)−1 = |ã|−2 .

Then ρ|ã|ρ = |ã|−1 , i.e., |ã| ∈ U(θ). Thus, ũ = |ã|−1 ã ∈ U(θ).

We define the unit sphere K of the form θ,

K = {x ∈ A2 : θ(x, x) = 1, x1 ∈ G} = {x ∈ A2 : x∗1 x1 − x∗2 x2 = 1, x1 ∈ G}.

We collect several elementary facts which will prove useful in the sequel.

Proposition 2.4. 1. If x ∈ K, then kx2 x−1

1 k < 1.

2. If ã ∈ U(θ) and x ∈ K, then ãx ∈ K; in particular, ãe1 ∈ K.

3. If x ∈ K then there exists y ∈ A2 with θ(y, y) = −1, such that θ(x, y) = 0; moreover,
 
x1 y1
ãx,y = ∈ U(θ).
x2 y2

4. U(θ)e1 = K. In particular, U(θ) acts transitively on K.

Proof. 1. If x ∈ K, x∗1 x1 = 1 + x∗2 x2 , and then x∗2 x2 < x∗1 x1 . Then (x2 x−1 ∗ −1
1 ) (x2 x1 ) =
(x∗1 )−1 x∗2 x2 x−1 −1
1 < 1, which implies that kx2 x1 k < 1.
 
a11
2. First note that the first column of ã ∈ U(θ) belongs to K (indeed, note a11 is
a21
invertible). Thus ãe1 ∈ K. If x ∈ K,

1 = θ(x, x) = θ(ãx, ãx).

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 In order to prove that ãx ∈ K it remains to show that a11 x1 + a12 x2 is invertible. Note
that
a11 x1 + a12 x2 = a11 (1 + a−1 −1
11 a12 x2 x1 )x1 .

The result follows because

ka−1 −1 −1 −1
11 a12 x2 x1 k ≤ ka11 a12 kkx2 x1 k < 1.
 ∗ 
−1 ∗ ∗ −1 ∗ a11
Indeed, ka11 a12 k = ka12 (a11 ) k < 1 because ã e1 = and ã∗ ∈ U(θ).
a∗12
 ∗ −1 ∗ 
(x1 ) x2
3. If x ∈ K, let y0 = , and observe that θ(x, y0 a) = θ(x, y0 )a = 0. Note that
1
θ(y0 , y0 ) is negative and invertible: θ(y0 , y0 ) = x2 x−1 −1 ∗ −1 −1 ∗
1 (x2 x1 ) − 1 with kx2 x1 (x2 x1 ) k =
−1 2 0
kx2 x1 k < 1. Thus we can find a ∈ A such that y = y a satisfies θ(y, y) = −1 (note that
we do not claim, nor need,  that y ∈K: the first coordinate of y may not be invertible).
x1 y1
It is easy to see that ã = ∈ U(θ).
x2 y2

4. The inclusion U(θ)e1 ⊂ K was proved in 2. The other inclusion follows from 3.: if x ∈ K
and y satisfies the properties of 3., then ã ∈ U(θ) and ãe1 = x.

3 The space Qρ
Recall that every ã ∈ Gl2 (A) admits a factorization ã = ũ|ã|, where |ã| = (ã∗ ã)1/2 and
ũ ∈ U2 (A). It also admits the factorization ã = |ã∗ |ũ (the same unitary matrix in both factoriza-
tions). These factorizations are called the polar decompostions of ã, and they have the following
uniqueness property: if ã = ṽr̃ = s̃w̃ for positives r̃, s̃ and unitaries ṽ, w̃, then ṽ = w̃ = ũ, r̃ = |ã|
and s̃ = |ã∗ |.

Notation 3.1. In what follows we write µ for the unitary part of ã ∈ Gl2 (A) and λ2 = |ã∗ |,
that is λ = |ã∗ |1/2 .

The following facts were proven in [6], Section 3.

Proposition 3.2. If ã =  ∈ M2 (A) is a reflection, i.e., 2 = 1, then the factors of  = λ2 µ

satisfy

1. µ = µ∗ = µ−1 , i.e. µ is a Hermitian reflection (also called a symmetry).

2. λµ = µλ−1 .

3. || = λ−1 = |∗ |−1 .

In this paper we mainly deal with idempotent matrices q ∈ M2 (A). Note that q 2 = q if and
only if 2q − 1 is a reflection.

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Corollary 3.3. If q ∈ M2 (A) with q 2 = q, then
2q − 1 = µ|2q − 1| = |2q − 1|−1 µ
for a unique Hermitian reflection µ, and
q = |2q − 1|−1/2 p|2q − 1|,
where p = 12 (1 + µ) is an Hermitian idempotent.
Notation 3.4. For q 2 = q ∈ M2 (A), we write λq = |2q − 1|−1/2 ; observe that λq∗ = |2q − 1|1/2 .
Definition 3.5. We denote by Qρ the set of all idempotents q ∈ M2 (A) such that
1. q ] = q, or explicitly, ρq ∗ ρ = q;
2. q decomposes the form θ: θ is positive in R(q) and negative in N (q).
It can be proven that these properties are equivalent to the condition (2q − 1)ρ is positive
(see [6], Section 3). Qρ is a closed and complemented submanifold of Q = {q ∈ M2 (A) : q 2 = q},
which has a well-studied differentiable structure as a complemented submanifold of M2 (A).
Remark 3.6. Note that if q ∈ Qρ , then 2q − 1 ∈ U(θ), and therefore |2q − 1|−1/2 ∈ U(θ) by
Proposition 2.2 and Corollary 2.3. Therefore one can regards the elements of Qρ as idempotents
or as reflections. Here we choose the former (in [1] the other viewpoint was taken).
Proposition 3.7. The map
ϕρ : K → Qρ , ϕρ (x) = px := xx∗ ρ
is a differentiable surjection with a global cross section. It holds that px = py , for x, y ∈ K if
and only if there exists u ∈ UA such that y = xu.
Proof. Clearly ϕρ is differentiable. Every q ∈ Qρ can be written [6]
1 + b∗ b −(1 + b∗ b)1/2 b∗
 
q= .
−b(1 + b∗ b)1/2 −bb∗
Then the map
(1 + b∗ b)1/2
 
q 7→ = qe1 ab ,
−b
where ab = θ(qe1 , qe1 )−1/2 = θ(qe1 , e1 )−1/2 , is a cross section for ϕρ .
Observe that if q = px for some x ∈ K, then b = x2 x∗1 |x∗1 |−1 .
If U ∈ UA and x ∈ K, it is clear that xu ∈ K and pxu = px . Conversely, suppose that
x, y ∈ K satisfy that px = py . Then clearly x = ya and y = xb for a, b ∈ A. Then 1 = θ(x, x) =
a∗ θ(y, y)a = a∗ a. Also y = xb = yab; comparing the first coordinates one gets y1 = y1 ab, and
since y1 is invertible, 1 = ab. Then a is a unitary element in A.
Proposition 3.8. U(θ) acts transitively on Qρ by means of
ã · q = ãqã−1 .
Proof. This is clearly a left action. In order to see that it is transitive,
 it suffices
 to show that
−1 1 0
any q ∈ Qρ is of the form q = ãp1 ã for some ã ∈ U(θ), where p1 = . This was shown
0 0
above in Corollary 3.3 for ã = λq = |2q − 1|−1/2 , which belongs to U(θ) (Remark 3.6).

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 4 The Poincaré disk as a homogeneous space of U(θ)
As an open subset of A, the disk D is naturally an open submanifold of A, with its tangent
spaces (T D)z identified with A. In this section we show that D is a homogeneous space of U(θ).
This is done by establishing a natural U(θ)-equivariant diffeomorphism between D and Qρ .
We have seen before that if x ∈ K then x2 x−11 ∈ D = {z ∈ A : kzk < 1}. Note that kzk < 1
if and only if z ∗ z < 1.

Lemma 4.1. The map

ϕD : K → D, ϕD (x) = x2 x−1
1

is a C ∞ surjection with a global cross section.

Proof. This map is clearly C ∞ , and

 
1
σD : D → K, σD (z) = (1 − z ∗ z)−1
z

is a cross section for ϕD .

The next Lemma allows us to find the transitive action of U(θ) on D.

Lemma 4.2. If ã ∈ U(θ) and z ∈ D, then a11 + a12 z is invertible.

Proof. a11 + a12 z = a11 (1 + a−1 −1

11 a12 z) is invertible because kzk < 1 and ka11 a12 k < 1 (this last
fact was shown in the proof of part 2. of Proposition 2.4).

Proposition 4.3. If ã ∈ U(θ) and z ∈ D then

ã · z = ϕD (ãσD (z) = (a21 + a22 z)(a11 + a12 z)−1 (1)

defines a C ∞ transitive left action of U(θ) on D.

Proof. By the above Lemma ã · z is well defined and clearly belongs to D; it is a straightforward
computation that ã · (b̃ · z) = (ãb̃) · z. To prove the transitivity it suffices to show that for every
z ∈ D there exists ã ∈ U(θ) such that ã · 0 = z. The matrix

(1 − z ∗ z)−1/2 z ∗ (1 − zz ∗ )−1/2
 
ã =
z(1 − z ∗ z)−1/2 (1 − zz ∗ )−1/2

does the feat, and clearly belongs to U(θ).

Theorem 4.4. There exists a diffeomorphism Φ : Qρ → D such that the following diagram
commutes
K (2)
ϕρ ϕD

~ 
Qρ
Φ / D.

All maps in this diagram are U(θ)-equivariant.

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Proof. Recall that ϕρ (x)px = xx∗ ρ. Put Φ(px ) = x2 x−1
1 . Φ is well defined because px = py if
and only if y = xu for some unitary u ∈ A. Bijectivity of Φ follows by the direct computation
of Φ−1 , namely
Φ−1 (z) = pσD (z) ,
 
1
where (recall that) σD (z) = (1 − z ∗ z)−1/2 .
z
For the U(θ)-equivariance, observe that for ã ∈ U(θ) and px ∈ Qρ , ãpx ã−1 is the unique θ
orthogonal projection onto the module [ãx]. So

Φ(ãpx ã−1 ) = (a21 x1 + a22 )(a11 x1 + a12 x2 )−1 = ã · x2 x−1

1 .

5 The hyperbolic part of AP

We say that an element x ∈ A2 is regular if x ∈ Gl2 (A)e1 , i.e., if x is the first column of a
matrix in Gl2 (A). A regular rank one submodule of A2 is a right A-module ` ⊂ A2 of the form
` = [x] := {xa; a ∈ A} for some regular x ∈ A2 .

Notation 5.1. We denote by AP the set of all regular rank one submodules of A2 , and call it
the projective line of A.

We define the cross ratio CR(`1 , `2 , `3 , `4 ) of four submodules `1 , `2 , `3 , `4 in AP, following

ideas by M. I. Zelikin [17].

Definition 5.2. We denote by CR(`1 , `2 , `3 , `4 ) the (possibly empty) set of module homomor-
phisms ϕ : `3 → `3 of the form ϕ = ψη, where the homomorphisms η : `3 → `2 , ψ : `2 → `3
satisfy x − ψ(x) ∈ `4 and y − η(y) ∈ `1 , for all x ∈ `2 , y ∈ `3 .

This set may contain more than one element. In Section 9 below, we discuss existence and
uniqueness of elements in CR(`1 , `2 , `3 , `4 ).
In this paper we study the subset APθ of AP, which we call the hyperbolic part of AP:

APθ = {[x] : x ∈ K}.

Note that the elements of K are regular, they are the first columns of matrices in U(θ). It is
easy to see that APθ is properly contained in AP.

Proposition 5.3. The group U(θ) acts transitively on APθ .

Proof. For ã ∈ U(θ) and [x] ∈ APθ , define

ã · [x] = [ãx].

This operation is well defined: if x, x0 span the same submodule then x0 = xu for some unitary
u ∈ A. Then ãx0 = (ãx)u and [ãx0 ] = [ãx]. To finish the proof, it suffices to recall that U(θ)
acts transitively on K (Proposition 2.4).

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 There exist many natural ways of endowing AP with a differentiable structure, see for in-
stance [2], [3]. Here we choose to define the differentiable structure only on APθ . For this task,
we consider the set Qρ of all θ-orthogonal idempotents such that θ is a positive form in R(q)
and a negative form in N (q). We have shown that Qρ = {px : x ∈ K}, where px is the unique
θ-orthogonal projection with range [x]. The geometry of Qρ has been extensively studied in [6]
and, more recently, [1].
Proposition 5.4. The mapping R : Qρ → APθ defined by R(q) = R(q) is a bijection, with
R−1 ([x]) = px . Morever, R(ãqã−1 ) = ã · R(q).
Proof. Straightforward computation.

The natural bijection R allows one to transfer the differentiable structure of Qρ to APθ . For
instance, the tangent space at [x] ∈ APθ , for x ∈ K, is identified as
(T APθ )[x] = {X ∈ M2 (A) : X ] = X, px X = X(1 − px )}

' {[y] : y ∈ A2 , θ(x, y) = 0, θ(y, y) = 1} = N (px ).

The isomorphism implementing ' above is the map X 7→ Xx, from (T APθ )[x] onto N (px ).
Definition 5.5. If ` ∈ APθ , denote by `⊥θ ,
`⊥θ = {y ∈ A2 : θ(y, x) = 0 for all x ∈ `}.
Note that (T APθ )[x] ' [x]⊥θ .

6 Geodesics of D
The geometry of the Poincaré disk D was studied in [1]. It is induced on D from the geometry
of Qρ , by means of the equivariant diffeomorphism Φ (Theorem 4.4) . In particular, we proved
that given two points z0 , z1 ∈ D there exists a unique geodesic joining them. In [1] we computed
the velocity of this unique geodesic in the case z0 = 0 and z1 = z is an arbitrary element of D.
The geodesic is given by  
0 α∗ 
t
α 0
δ(t) = e · 0,
where
∞
X 1
α=z (z ∗ z)k . (3)
2k + 1
k=0
t2k+1
The curve δ satisfies that δ(0) = 0 and δ(1) = z. Note here that the series ∞
P
k=0 2k+1 corre-
sponds, in the interval (−1, 1), to the function f (t) = 21 log( 1+t
1−t ). We shall compute below the
explicit form of δ in D.
Let us compute explicitly the form of the geodesic δ joining 0 and z ∈ D at time t = 1:
Lemma 6.1. Given z ∈ D, the unique geodesic δ of D with δ(0) = 0 and δ(1) = z is given by
δ(t) = ω tanh(t|α|), (4)
where α is given in (3) above, and ω is the partial isometry in the polar decomposition of α:
α = ω|α|, performed in A∗∗ .

10

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 0 α∗
 
Proof. Straightforward computations show that the even and odd powers of are,
α 0
respectively
2k  ∗ k 2k+1 
0 α∗ 0 α∗ α∗ (αα∗ )k
   
(α α) 0 0
= and = .
α 0 0 (αα∗ )k α 0 α(α∗ α)k 0
 
0 α∗
t 
α 0
We are interested in the first column of e , which is
P∞ t2k ∗ k ! 
(α α)

k=0 cosh(t|α|)
P∞ (2k)!t2k+1 = .
α k=0 (2k+1)! (α∗ α)k ω sinh(t|α|)

Then
δ(t) = ω sinh(t|α|)(cosh(t|α|))−1 = ω tanh(t|α|).

Notice that ω ∈ A∗∗ need not belong to A. However δ(t) ∈ A for all t.
Let us relate the polar decompositions of z and α.
Proposition 6.2. If z ∈ D and α as in (3), then
1
α = ω(log(1 + |z|) − log(1 − |z|)), and z = ω|z|,
2
i.e., the partial isometry ω ∈ A∗∗ is the same for α and z.
Proof. First note that
∞ ∞ ∞
!2
X 1 X 1 X 1
|α|2 = α∗ α = (z ∗ z)k z ∗ z (z ∗ z)k = |z|2k+1 ,
2k + 1 2k + 1 2k + 1
k=0 k=0 k=0

i.e., |α| = 21 (log(1P

+ |z|) − log(1 − |z|)). Next, put z = µ|z| the polar decomposition of z. Note
that, since α = z ∞ 1 2k
k=0 2k+1 |z| , we have that
∞
X 1
α = µ|z| |z|2k = µ|α|.
2k + 1
k=0

Thus, in order to prove our claim, it suffices to show that both partial isometries µ, ω have the
same initial and final spaces (the result follows, then, by the uniqueness property of the polar
decomposition). The partial isometry µ maps N (z)⊥ = N (|z|)⊥ onto R(z), whereas ω maps
N (α)⊥ onto R(α). If zξ = 0, then
∞ ∞
X 1 X 1
αξ = z (z ∗ z)k ξ = (zz ∗ )k zξ = 0,
2k + 1 2k + 1
k=0 k=0
P∞ 1 ∗ 2k
i.e., N (z) ⊂ N (α). Conversely, in the last expression of α, α = k=0 2k+1 |z | z; observe that
∞
X 1
|z ∗ |2k = g(|z ∗ |),
2k + 1
k=0

11

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 1
where g(t) = 2t (log(1+t)−log(1−t)) (which can be extended continuously as g(0) = 1), is defined
in σ(|z |) ⊂ [0, 1), and is nonvanishing there. Therefore, g(|z ∗ |) is invertible. Thus, αξ = 0
∗

implies zξ = 0. Again, using the function g, we get α = zg(|z|), and thus R(z) = R(α).

Corollary 6.3. The exponential map Exp0 of D at 0, and its inverse Log0 can be written
explicitly as follows: if z = ω|z| ∈ D
1
Log0 : D → (T D)0 , Log0 (z) = ω log (1 + |z|)(1 − |z|)−1 .


2
If α = ω|α| ∈ A ' (T D)0 , then

Exp0 : (T D)0 → D , Exp0 (α) = ω tanh(|α|).

In particular, if α = Log0 (z) (or, equivalently, z = Exp0 (α)), then z and α have the same partial
isometry in the polar decomposition. Also,
1
log (1 + |z|)(1 − |z|)−1 .


|Exp0 (α)| = tanh(|α|) and |Log0 (z)| =
2

7 Limit points of geodesics

One of our concerns in computing the geodesic δ, and the above results on the polar decompo-
sition, is to establish the following result:

Theorem 7.1. For z ∈ D, let δ be the unique geodesic of D such that δ(0) = 0 and δ(1) = z.
Put z = ω|z| the polar decomposition (i.e., ω ∈ A∗∗ ); then

SOT − lim δ(t) = ω and SOT − lim δ(t) = −ω.

t→∞ t→−∞

Proof. By formula (4), we only need to compute the limit of tanh(t|z|) when t → ±∞. The
spectrum σ(|z|) is contained in [0, 1). Clearly, for any s ∈ [0, 1),

1 if s ∈ (0, 1)
lim tanh(ts) =
t→∞ 0 if s = 0.

By Lebesgue’s bounded convergence theorem, and the Borel functional calculus for bounded
selfadjoint operators, we have that

lim tanh(t|α|) = χ(0,1) (|α|) = PN (α)⊥ .

t→∞

Then,
lim δ(t) = ωPN (α)⊥ = ω.
t→∞

−1 if s ∈ (0, 1)
Similarly, using that limt→−∞ tanh(st) = , we get that
0 if s = 0

lim δ(t) = −ωPN (α)⊥ = −ω.

t→−∞

12

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 This geometric role of the partial isometry ω in the polar decomposition of z ∈ D (or, more
generally, of every z ∈ A \ {0}) has not been noticed before, to the authors’ knowledge.
In order to compute the limit points of arbitrary geodesics, it will be useful to extend the
action of U(θ) to the strong operator border

∂D := {a ∈ A∗∗ : kak = 1}

of D, i.e., to define g̃ · a for a ∈ A∗∗ with kak = 1.

Lemma 7.2. If g̃ ∈ U(θ) and a ∈ ∂D, then g11 + g12 a is invertible in A∗∗ .

Proof. Note that  

1
g11 + g12 a = he1 , g̃ i.
a
 
u1 0
Using the polar decomposition g̃ = ũ|g̃|, ũ = ,
0 u2
   
1 1
he1 , g̃ i = u∗1 he1 , |g̃| i,
a a

(1 + b∗ b)1/2 b∗
 
i.e., we may suppose g̃ ≥ 0, g̃ = , for b ∈ A. Then
b (1 + bb∗ )1/2

g11 + g12 a = (1 + b∗ b)1/2 + b∗ a = (1 + b∗ b)1/2 (1 + (1 + b∗ b)−1/2 b∗ a).

It suffices to show that k(1 + b∗ b)−1/2 b∗ ak < 1. Note that, since kak = 1,

k(1+b∗ b)−1/2 b∗ ak2 ≤ k(1+b∗ b)−1/2 b∗ k2 = k(1+b∗ b)−1/2 b∗ b(1+b∗ b)−1/2 k = max{f (t) : t ∈ σ(b∗ b)},
t
for f (t) = (1+t) . Clearly, this number is strictly less than 1.

Proposition 7.3. If a ∈ ∂D, and g̃ ∈ U(θ), then

g̃ · a := (g21 + g22 a)(g11 + g12 a)−1 ∈ ∂D,

defines a left action of U(θ) on ∂D.

Proof. A density argument (or a proof similar as in the previous lemma), shows that if x ∈ A∗∗
with kxk < 1, then g̃ · x, defined as above, also satifies kg̃ · xk < 1, and defines an action on the
unit ball of A∗∗ . Let a ∈ ∂D. Then, by Kaplansky’s density theorem, there exists a sequence
an ∈ D such that an → a in the strong operator topology. We claim that
Lemma 7.4. g̃ · an → g̃ · a strongly.

Proof. Considerthe polar decomposition g̃ =  ũ|g̃|. We check first that |g̃| · an → |g̃| · a strongly.
(1 + b∗ b)1/2 b∗
As before, g̃ = . Then
b (1 + bb∗ )1/2

|g̃| · an = (b + (1 + bb∗ )1/2 an )((1 + b∗ b)1/2 + b∗ an )−1 .

13

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Clearly, b + (1 + bb∗ )1/2 an → b + (1 + bb∗ )1/2 a and (1 + b∗ b)1/2 + b∗ an → (1 + b∗ b)1/2 + b∗ a strongly.
Moreover,
((1 + b∗ b)1/2 + b∗ an )−1 = (1 + (1 + b∗ b)−1/2 b∗ an )−1 (1 + b∗ b)−1/2 .
Let us show that the norms of these inverses are uniformly bounded. It suffices to see that the
norms k(1+(1+b∗ b)−1/2 b∗ an )−1 k are uniformly bounded. Denote dn = (1+b∗ b)−1/2 b∗ an . Then,
as seen above,
kdn k2 ≤ max{f (t) : t ∈ σ(b∗ b)} = r2 < 1.
Thus,
1
k(1 + (1 + b∗ b)−1/2 b∗ an )−1 k ≤ .
1−r
Therefore the inverses (1 + (1 + b∗ b)−1/2 b∗ an )−1 converges strongly to (1+ (1 + b∗ b) 
−1/2 b∗ a)−1 .

u1 0
Then, clearly, cn := |g̃| · an converge strongly to c := |g̃| · a. Since ũ = , it is clear
0 u2
that
g̃ · an = ũ · (|g̃| · an ) = u2 cn u∗1 → u2 cu∗1 = g̃ · a.

Let us proceed with the proof of Proposition 7.3. Since kan k < 1, we know that kg̃ · an k < 1.
From Lemma 7.4, it follows that kg̃ · ak ≤ 1. Suppose that kg̃ · ak < 1. Using again the fact that
U(θ) acts on D, this would imply that

g̃ −1 · (g̃ · a) = a ∈ D,

a contradiction. Thus, kg̃ · ak = 1.

The fact that this rule defines, indeed, a left action, follows from similar density arguments.

Using this result, we can compute the limit points of arbitrary geodesics. Since the action
of U(θ) is transitive, given z1 , z2 ∈ D, there exists g̃ ∈ U(θ) such that g̃ · z1 = 0.

Corollary 7.5. Let z0 , z1 ∈ D and g̃ ∈ U(θ) such that g̃ · z0 = 0. Let δ be the unique geodesic
of D such that δ(0) = z0 and δ(1) = z1 . Denote by δ̇0 the initial velocity of δ. Then

SOT − lim δ(t) = g̃ · ω0 and SOT − lim δ(t) = g̃ · (−ω0 ),

t→+∞ t→−∞

where ω0 ∈ A∗∗ is the partial isometry in the polar decomposition of δ̇0 : δ̇0 = ω0 |δ̇0 |.

Remark 7.6. In order to identify these  limit ∗points in D, following the notation
 of the above
(1 + b b) 1/2 b∗ v1 0
Corollary, note that if g̃ = |g̃ ∗ |ṽ = (the reversed
b (1 + bb∗ )1/2 0 v2
polar decomposition), then ṽ · ω0 = v2 ω0 v1∗ is a partial isometry. Therefore, the limit points of
geodesics are elements in ∂D of the form

(b + (1 + bb∗ )1/2 ω)((1 + b∗ b)1/2 + b∗ ω)−1 and (b − (1 + bb∗ )1/2 ω)((1 + b∗ b)1/2 − b∗ ω)−1 ,

where b ∈ A is arbitrary and ω ∈ A∗∗ is a partial isometry.

14

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Note that not every partial isometry in A∗∗ occurs in the polar decomposition of an element
in D. For instance, if A = C([0, 1]) (continuous functions in the unit interval),
 the polar decom-
f (t)/|f (t)| if f (t) 6= 0
position of f ∈ A is f = w|f |, where w ∈ L∞ (0, 1) is given by w(t) = .
0 if f (t) = 0
An arbitrary partial isometry in L∞ (0, 1) is a measurable function whose values are zero or
complex numbers of modulus 1. The set of zeros of such a function is an arbitrary measurable
set, whereas the set of zeros of partial isometries which occur in the polar decomposition of a
continuous function, are closed subsets of [0, 1].
Another way to study the limit points of geodesics, is by using the Borel subgroup Bθ ⊂ U(θ)
instead. Indeed, since the action of this group is transitive in D (a fact which will be proven in
the next section), any limit point of a geodesic is either of the form g̃ · v or g̃ · (−v), for g̃ ∈ Bθ .
Consider the following example:
Example  7.7. Suppose that A is a von Neumann algebra, and let p 6= 0 be a projection in A.
g + ĝ g − ĝ
 2 − ĝx 2
− ĝx 
For g̃ =  g − ĝ g + ĝ , let us compute g̃ ·p. After straightforward computations,
+ ĝx + ĝx
2 2
g̃ · p = (g(1 + p) + ĝ (−1 + p + 2x(1 + p))) (g(1 + p) + ĝ (1 − p − 2x(1 + p)))−1 .
Note that 1 + p is invertible and that (1 − p)(1 + p)−1 = 1 − p. Then

−1
g̃ · p = 1 + ĝ (p − 1 + 2x(1 + p)) (1 + p)−1 g −1 1 + ĝ (1 − p − 2x(1 + p)) (1 + p)−1 g −1


=

−1
= 1 + ĝ(p − 1 + 2x)g −1 1 + ĝ(1 − p − 2x)g −1


.
Denote α = ĝ(p − 1)g −1 and β = 2ĝxg −1 . Observe that α is a non-invertible selfadjoint element,
α ≤ 0 and its range is proper and closed; β is an arbitrary anti-selfadjoint element. Then
g̃ · p = (1 + α + β) (1 − (α + β))−1 .
Note that 1 − (α + β) is invertible because Re(1 − (α + β)) = 1 − α ≥ 1. If one picks p = 1, then
α = 0 and
g̃ · 1 = (1 + β)(1 − β)−1 ,
which is a unitary operator such that −1 does not belong to its spectrum. In particular, this
shows that the action of Bθ ceases to be transitive in ∂D.
Our next result shows a necessary condition for an element a ∈ A∗∗ with kak = 1 to be a
limit point of a geodesic of D
Proposition 7.8. If a ∈ A∗∗ is the limit point at +∞ of a geodesic in D, then 1 − a∗ a = hqh,
where h ∈ G+ and q ∈ A∗∗ is a projection. In particular, not every element of norm 1 in A∗∗
is the limit point of a geodesic: such elements satisfy that the defect element 1 − a∗ a has closed
range.
Proof. If a is the limit point of a geodesic if and only if there exists g̃ ∈ U(θ) and a partial
isometry ω ∈ A∗∗ such that a = g̃ · a. The definition of the action implies that this equality can
be read as an usual matrix equality
   
1 b1
g̃ = ,
ω b2

15

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 with b1 ∈ G and b2 b−1
1 = a. Using the form θ, and the fact that g̃ ∈ U(θ),
       
b1 b1 1 1
θ( , ) = θ( , ),
b2 b2 ω ω

i.e.
b∗1 b1 − b∗2 b2 = 1 − ω ∗ ω = q 0 ,
which is a projection in A∗∗ . Let b1 = u|b1 | be the polar decomposition, with u unitary. Then
au = b2 b−1 −1
1 u = b2 |b1 | . Thus,

q 0 = b∗1 b1 − b∗2 b2 = |b1 |(1 − |b1 |−1 b∗2 b2 |b1 |−1 )|b1 | =, |b1 |(1 − u∗ a∗ au)|b1 |,

i.e.,
1 − a∗ a = u|b1 |−1 q 0 |b1 |−1 u∗ = hqh,
for q = uq 0 u∗ a projection in A∗∗ and h = u|b1 |−1 u∗ ∈ G+ .

Remark 7.9. The characterization of the partial isometries which appear in the polar decom-
positions of all limit points a ∈ A∗∗ is an interesting open problem.

8 The invariant metric in APθ

Given ` ∈ APθ and V ∈ (T APθ )` , fix a generator x0 ∈ K for `, i.e., [x0 ] = ` and θ(x0 , x0 ) = 1.
Recall that x0 is determined up to a unitary element of A: if x00 is another such generator,
then there exists u ∈ UA such that x00 = x0 u. Having fixed a generator for `, as we saw above,
(T APθ )` identifies with `⊥θ , and to the tangent vector V corresponds an element v ∈ `⊥θ . We
define
|V |` := kθ(v, v)k1/2 . (5)
Note that | |` does not depend on the choice of the generator. If we choose x00 = x0 u instead,
the tangent vector V is represented by v0 = vu ∈ `⊥θ , and therefore

kθ(v0 , v0 )k = kθ(vu, vu)k = ku∗ θ(v, v)uk = kθ(v, v)k.

Next, recall from Lemma that θ is negative definite (non degenerate), and therefore the expres-
sion (5) above defines a proper norm in (T APθ )` .
 ∗ −1 
θ ⊥ (x1 ) x2
Remark 8.1. If V ∈ (T AP )[x0 ] , V is represented by some v ∈ [x0 ] ; since y0 =
θ ,
1
 
x1
for x0 = , is a generator of [x0 ]⊥θ , there exists a ∈ A such that v = y0 a. Then
x2

|V |[x0 ] = kθ(y0 a, y0 a)k1/2 = ka∗ (1 − x2 x−1 ∗ 1/2

1 ) ak = k(1 − |(x2 x−1 ∗ 2 1/2
1 ) | ) ak.

Note also that the norm of v = y0 a in A2 is

ky0 ak = khy0 a, y0 aik1/2 = k(1 + |(x2 x−1 ∗ 2 1/2

1 ) | ) ak.

Proposition 8.2. For any ` ∈ APθ , the norm | |` of (T APθ )` is complete.

16

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Proof. With the current notations, if V = y0 a ∈ (T APθ )[x0 ] ,

|V |[x0 ] = k(1−|(x2 x−1 ∗ 2 1/2

1 ) | ) ak = k(1−|(x2 x−1 ∗ 2 1/2
1 ) | ) (1+|(x2 x−1 ∗ 2 −1/2
1 ) | ) (1+|(x2 x−1 ∗ 2 1/2
1 ) | ) ak

≤ k{(1 − |(x2 x−1 ∗ 2 −1 ∗ 2 −1 1/2

1 ) | )(1 + |(x2 x1 ) | ) } kky0 ak.
Similarly
ky0 ak ≤ k{(1 + |(x2 x−1 ∗ 2 −1 ∗ 2 −1 1/2
1 ) | )(1 − |(x2 x1 ) | ) } k|V |[x0 ] .
It follows that on (T APθ )`] ' [y0 ], the metric | |` and the norm of A2 are equivalent. Since [y0 ]
is closed in A2 , it is complete, and the proof follows.

The distribution APθ 3 ` 7→ | |` is clearly continuous. Thus, APθ is endowed with a Finsler
metric.
The following result is tautological, but of the utmost importance for our discussion:

Theorem 8.3. The Finsler metric defined in (5) is invariant under the action of U(θ).

Proof. Pick ` = [x0 ] ∈ APθ , V ∈ (T APθ )` (as before, V is identified to ∼ y0 a) and g̃ ∈ U(θ).
The action of U(θ) on APθ induces an action on the tangent spaces. Since g̃ preserves θ,

g̃([y0 ]⊥θ ) = [g̃y0 ]⊥θ .

Then
|g̃V |q̃[x0 ] = kθ(g̃(y0 a), g̃(y0 a))k1/2 = kθ(y0 a, y0 a)k1/2 = |V |[x0 ] .

8.1 Invariant metric in D

We need to compute the differential of the map APθ → D at [e1 ]. To do so, we use the diagram
(2) in Theorem 4.4.

Lemma 8.4. The differential of the map APθ → D, [x] 7→ x2 x−1

1 , at [e1 ] is the map
 
0
7−→ x , x ∈ A.
x

Proof. Fix x ∈ A. We use the commutative diagram (2). Let x(t)  ∈ K bea smooth curve
0
such that x(0) = e1 , and the derivative of [x(t)] (in APθ ) at t = 0 is . Then x1 (0) =
x
1, x2 (0) = 0, ẋ2 (0) = x. If we map x(t) onto D, and differentiate at t = 0 we get:

d
x2 (t)x−1 −1 −1 −1
1 (t)|t=0 = ẋ2 (0)x1 (0) − x2 (0)x1 (0)ẋ1 (0)x1 (0) = x,
dt
which proves our claim.

Theorem 8.5. The identification [x] ←→ x2 x−1 θ

1 between AP and D, is isometric.

17

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Proof. The group U(θ) acts isometrically on both APθ and D. The action is isometric in both
spaces. Therefore, it suffices to show that the differential at [e1 ] is an isometry. By Lemma 8.4,
this map is  
0
7−→ x.
x
 
0
The norm of is computed via the identification (T APθ )[e1 ] ' [e1 ]⊥θ = [e2 ], which
x
 
0
sends to e2 x. The norm (in A2 ) of this element is
x

kx∗ θ(e2 , e2 )xk1/2 = kx∗ xk1/2 = kxk.

On the other hand, the norm of x as a tangent element of D at 0 is the usual norm kxk (in
A).

8.2 The action of the Borel subgroup Bθ

As we mentioned before, the group Bθ , which we call the Borel group of the form θ, is a subgroup
of U(θ) which acts transitively in D. Since it is a subgroup of U(θ), the action is also isometric
(recall also that the action is free in the hyperboloid K). Therefore it acts transitively and
isometrically in APθ .
Given z ∈ D, let us find an element g̃ ∈ Bθ such that g̃ · 0 = z. First, we describe the action
of U(θ) on D, which factors through the hyperboloid K:

Remark 8.6. Given g̃ ∈ U(θ) and z ∈ D, we lift z to K by means of the global section

(1 − z ∗ z)−1/2
 
D 3 z 7→ .
z(1 − z ∗ z)−1/2

(1 − z ∗ z)−1/2
 
Next, we multiply g̃ , and finally we compose with the fibration
z(1 − z ∗ z)−1/2
 
z1
K3 7→ z2 z1−1 ∈ D.
z2

Definition 8.7. If z ∈ D, we define

(1 − z ∗ z)−1/2 (1 + z ∗ )−1 z ∗ (z + 1)(1 − z ∗ z)−1/2

 
g̃z =
z(1 − z ∗ z)−1/2 (1 + z ∗ )−1 (z + 1)(1 − z ∗ z)−1/2

(1 + z ∗ )−1 1 + z∗ z ∗ (z + 1) (1 − z ∗ z)−1/2
   
0 0
= .
0 (1 + z ∗ )−1 ∗
(1 + z )z z+1 0 (1 − z ∗ z)−1/2

Observe that the diagonal matrices on the right and left-hand sides do not belong to U(θ),
so that this is not a proper factorization of g̃z .

Lemma 8.8. If z ∈ D, then g̃z ∈ Bθ and satisfies g̃z · 0 = z.

18

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Proof. The origin 0 ∈ D lifts to e1 ∈ K. Recall the form of the elements in Bθ :

g − ĝ
 
g + ĝ
 2 − ĝx 2
− ĝx 
∗
Bθ = { g − ĝ g + ĝ  : g ∈ G, x = −x}.
+ ĝx + ĝx
2 2
Thus, we are looking for g ∈ G and x ∈ A with x∗ = −x such that
1 1
(g + ĝ) − ĝx = (1 − z ∗ z)−1/2 and (g − ĝ) + ĝx = z(1 − z ∗ z)−1/2 .
2 2

That is, g = (1 + z)(1 − z ∗ z)−1/2 , thus ĝ = (1 + z ∗ )−1 (1 − z ∗ z)1/2 and

1
ĝx = (1 + z ∗ )−1 (z − z ∗ )(1 − z ∗ z)−1/2 .
2
Then,
1 1
x = g ∗ (1 + z ∗ )−1 (z − z ∗ )(1 − z ∗ z)−1/2 = (1 − z ∗ z)−1/2 (z − z ∗ )(1 − z ∗ z)−1/2 ,
2 2
which is clearly anti-Hermitian, and thus g̃z ∈ Bθ .

Remark 8.9. Denote by the d the distance defined by the the Finsler metric introduced in D.
In [1] (Section 8) it is shown that

1 1 + kzk
d(0, z) = log( ).
2 1 − kzk

Using the fact that the action of Bθ on D is isometric, and the above construction of g̃z , for
arbitrary z1 , z2 ∈ D, we can compute d(z1 , z2 ) as follows:

1 1 + kg̃z−1 · z2 k
d(z1 , z2 ) = d(0, g̃z−1
1
· z2 ) = log( 1
−1 ). (6)
2 1 − kg̃z1 · z2 k

In order to compute g̃z−1 , recall that elements g̃ in U(θ) are characterized by the relation ρg̃ ∗ ρ =
ρ−1 . Then

(1 − z ∗ z)−1/2 −z ∗ (1 + z) (1 + z)−1
   
−1 0 1+z 0
g̃z = .
0 (1 − z ∗ z)−1/2 −(1 + z ∗ )z 1 + z∗ 0 (1 + z)−1

Then, after straightforward computations,

g̃z−1
1
· z2 = (1 − z1∗ z1 )−1/2 (1 + z1∗ )(1 + z1 )−1 (z2 − z1 )(1 − z1∗ z2 )−1 (1 − z1∗ z1 )1/2 .

9 Operator cross ratio in the hyperbolic part of the projective

line
Here we state our main result, relating the metric of APθ introduced in Section 8, with the so
called operator cross ratio, as defined in the Grassmann manifold of a Hilbert space by M.I.

19

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Zelikin [17]. We shall apply these ideas to the rank one submodules in APθ . To this effect, the
isometry betweenAPθ and
 the disk D will be important.
1
Consider ` = ∈ APθ , for z ∈ D. Let z = ω|z| be the polar decomposition. Let δ be
z
   
θ 1 1
the geodesic of AP such that δ(0) = and δ(1) = . Equivalently, regarded in
0 z
D: δ(0) = 0 and δ(1) = z. As seen in Section 6,

SOT − lim δ(t) = ω and SOT − lim δ(t) = −ω.

t→+∞ t→−∞

Four points are determined: −ω, 0, z, ω, or better, four rank one submodules
       
1 1 1 1
`−∞ := , `0 := ,` = , `+∞ := ,
−ω 0 z ω

where the limit lines lie in ∂APθ .

In Section 3 we defined the operator cross ratio of four elements in AP, as a (possibly empty)
set of module endomorphisms, following ideas of Zelikin [17]. Here we compute the operator
cross ratio CR(`−∞ , `0 , `, `+∞ ), proving that it is non empty, and that there exists a natural
`-endomorphism to choose from this set.
Recall that elements of CR(`−∞ , `0 , `, `+∞ ) are (module) endomorphisms of `, defined as the
composition of the projection from ` to `0 parallel to `−∞ , followed by the projection from `0 to
` parallel to `+∞ .
In coordinates, by choosing generators in the respective submodules
       
1 1 1−λ 1
7→ λ, = µ.
z 0 z −ω

Then 1 − λ = µ and z = −ωµ. Then ω|z| = −ωµ. If z is invertible (and then ω is unitary)
this implies µ = −|z|, otherwise this is just one possible solution. Non uniqueness of solutions
of these equations reflect the geometric fact that the modules `0 and `∞ may not be in direct
sum. Explicitly, all solutions of these equations are of the form

λ = 1 + |z| − Ω , µ = −|z| + Ω,

where Ω ∈ A∗∗ is such that ωΩ = 0. In particular, |z|Ω = |z|ω ∗ ωΩ = 0. We choose the solution
with Ω = 0. Note that λ = 1 + |z|, and therefore the first projection in the above composition
is given by    
1 1 + |z|
7→ .
z 0
Next        
1 + |z| 1 1 + |z| − γ 1
7→ γ, = .
0 z −zγ ω
So that 1 + |z| − γ =  and −zγ = ω, and then (the unique solution if ω is unitary, or a possible
solution that we choose, otherwise) 1 + |z| − γ = −|z|ω, i.e.,

γ = (1 − |z|)−1 .

20

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Other solutions of the above equation are of the form
γ = (1 − |z|)−1 + (1 − |z|)−1 Ω0 ,
where Ω0 ∈ A∗∗ is such that ωΩ0 = 0. In general, the possible endomorphisms ` → ` are given
(in these coordinates) by
   
1 1
7→ (1 + |z| + Ω)(1 − |z|)−1 (1 + Ω0 ) = (1 + |z|)(1 − |z|)−1 + Ω0 + Ω(1 − |z|)−1 + ΩΩ0 ,
z z
where we use that ωΩ = ωΩ0 = 0, and thus (1 ± |z|)±1 Ω = Ω (and the same for Ω0 ).
As noted, if z is invertible, there is a unique solution with Ω = Ω0 = 0. Our choice of cross
ratio, picking Ω = Ω0 = 0 in any case, is justified below.
Remark 9.1. The following fact is known (see for instance [10]). Let an ∈ A with kan k ≤ 1. If
an → a strongly, then |an | → |a| strongly.
Remark 9.2. Dixmier and Marechal [8] proved that the set on invertible elements of a von
Neumann algebra is strong operator dense in the algebra. The argument in [8] proceeds as
follows. Let a = u|a| be the polar decomposition of a (u ∈ A∗∗ ). First, the algebra A∗∗ is
factored in its finite and properly infinite parts. In the finite part u can be chosen unitary. In
the properly infinite part, one readily sees that it suffices to consider the cases in which u is an
isometry or a co-isometry. In the case that u is an isometry, Dixmier and Maréchal prove that
u is the strong limit of unitaries un . If u is a co-isometry, they show that there exist invertible
elements gn which converge strongly to u, with norms kgn k = 1 (this is clear in the proof, though
it is not stated in their result). Summarizing, if A is a von Neumann algebra, and a ∈ A, there
exist gn ∈ G with kgn k ≤ kak such that
SOT − lim gn = a.
n→∞

Using this fact, it is clear that, if A is a von Neumann algebra, and z ∈ D, then there exists
and zn ∈ G with kzn k ≤ kzk, such that zn → z strongly.
Proposition 9.3. Let zn , z ∈ D such that zn → z strongly and kzn k ≤ kzk. Then
(1 + |zn |)(1 − |zn |)−1 → (1 + |z|)(1 − |z|)−1 strongly.
Proof. First note that |zn | → |z| strongly. Then 1 ± |zn | converges strongly to 1 ± |z|. Clearly
k1 + |zn |k ≤ 2. Let us check that also k(1 − |zn |)−1 k are uniformly bounded:
∞ ∞
−1
X
k
X 1
k(1 − |zn |) k=k |zn | k ≤ kznk k ≤ < ∞.
1 − kzk
k=0 k=0

Therefore, (1 − |zn |)−1 → (1 − |z|)−1

strongly, and since the product is strongly continuous on
norm bounded sets, the proof follows.
 
1
Definition 9.4. Let z ∈ D. We define cr(0, z) ∈ CR(`−∞ , `0 , `z , `∞ ), for `z = , as the
z
endomorphism
   
1 1
(1 + |z|)(1 − |z|)−1 a,


cr(0, z) : `z → `z , cr(0, z) a =
z z
for a ∈ A.

21

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 We use the action of U(θ) to extend this definition to any pair z0 6= z1 ∈ D.

Definition 9.5. Let z0 , z1 ∈ D, z0 6= z1 . Pick g̃ ∈ U(θ) such that g̃ · 0 = z0 and denote

z = g̃ −1 · z1 . We define
cr(z0 , z1 ) = g̃ cr(0, z)g̃ −1 .

Before checking that the definition does not depend on the choice of g̃, we remark the
following. Let δ be the unique geodesic of D such that δ(0) = z0 and δ(1) = z1 , and let

z−∞ = SOT − lim δ(t) and z+∞ = SOT − lim δ(t).

t→−∞ t→+∞

Then cr(z0 , z1 ) ∈ CR(`z−∞

 , `z0 , `z1 , `z+∞ ), because g̃ is a module homomorphism which maps `z
1
onto `z1 . Indeed, if x = a ∈ `z , then clearly
z
   
1 1
g̃x = (g11 + g12 z)a = (g11 + g12 z)a ∈ `z1 .
g̃ · z z1

Let us check that cr(z1 , z2 ) is well defined, i.e., that it does not depend on the
 choice ofg̃.
u1 0
To prove this, recall that if k̃ ∈ U(θ) satisfies k̃ · 0 = 0, then k̃ = , with
0 u2
u1 , u2 ∈ UA .

Proposition 9.6. With the above notations, the endomorphism

cr(z0 , z1 ) ∈ CR(`z−∞ , `z0 , `z1 , `z+∞ )

does not depend on the choice of g̃. Namely, if h̃ ∈ U(θ) satisfies h̃ · 0 = z0 , and z 0 = h̃−1 · z1 ,
then
h̃ cr(0, z 0 )h̃−1 = g̃ cr(0, z)g̃ −1 .
 
−1 u1 0
Proof. Since h̃ · 0 = g̃ · 0, it follows that (g̃ h̃) · 0 = 0, and therefore h̃ = g̃ , for
0 u2
u1 , u2 ∈ UA . Then
 ∗   ∗ 
0 −1 u1 0 −1 u1 0
z = h̃ · z1 = · (g̃ · z1 ) = · z = u∗2 zu1 ,
0 u∗2 0 u∗2

and
u∗1 0
   
0 −1 u1 0
h̃ cr(0, z )h̃ = g̃ cr(0, u∗2 zu1 ) g̃ −1 .
0 u2 0 u∗2
u∗1 0
   
u1 0
Thus, we must show that cr(0, u∗2 zu1 )
= cr(0, z). Let us see how the
0 u2 0 u∗
 2
1
left-hand side endomorphism transforms the element a ∈ `z . First, it is sent to
z

u∗1 0
    
1 1
a= u∗1 a.
0 u∗2 z u∗2 zu1

22

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 The map cr(0, z 0 ) maps this element to
 
1
(1 + |u∗2 zu1 |)(1 − |u∗2 zu1 |)−1 u∗1 a.
u∗2 zu1

Note that |u∗2 zu1 | = ((u∗2 zu1 )∗ u∗2 zu1 )1/2 = (u∗1 z ∗ zu1 )1/2 = u∗1 |z|u1 . Therefore, the above element
equals  ∗ 
u1
(1 + |z|)(1 − |z|)−1 a.
u∗2 z
 
u1 0
Finally, multiplying on the left by the matrix yields
0 u2
   
1 −1 1
(1 + |z|)(1 − |z|) a = cr(0, z) a.
z z

Suppose that A is a von Neumann algebra. If z ∈ D is non invertible, there exist zn ∈ D

which are invertible such that zn → z strongly and kzn k ≤ kzk. Then cr(0, zn ), cr(0, z) are
endomorphisms of different submodules. In order to compare them, we can regard them as
A-module morphisms of A2 , embedding each module in A2 using the θ-orthogonal projections
p`zn , p`z onto the submodules `zn , `z , respectively. For z 0 ∈ D,
   
0 2 −1/2 1 1
p`z0 (x) = (1 − |z | ) θ( , x) (1 − |z 0 |2 )−1/2 .
z0 z0

Proposition 9.7. Let A be a von Neumann algebra.

1. If z ∈ D ∩ G, the set CR(`−∞ , `0 , `z , `∞ ) consists of a single element cr(0, z).

2. If z ∈ D is non invertible, there exist zn ∈ D which are invertible such that zn → z strongly
and kzn k ≤ kzk. Then, with the above notations,

cr(0, zn )p`zn (x) → cr(0, z)p`z (x)

strongly in A2 .

Proof. By Remark 9.3, we know that (1 + |zn |)(1 − |zn |)−1 → (1 + |z|)(1 − |z|)−1 strongly. By a
similar argument, it also holds that (1−|zn |2 )−1/2 → (1−|z|2 )−1/2 strongly. Also these operators
are uniformly bounded. Therefore, using that the product is strongly continuous in bounded
sets, our claim follows.

Observe that in the von Neumann algebra case cr(0, z) is then uniquely defined when z is
invertible, and is defined by strong continuity in the remaining cases.

Remark 9.8. As a corollary we get that, even if the set CR(`1 , `2 , `3 , `4 ) may be empty for
general `1 , `2 , `3 , `4 , the particular set CR(`−∞ , `0 , `z , `∞ ) is not, and cr(0, z) is a distinguished
element of this set.

23

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 As a first approximation of the deep relationship between the cross ratio and the metric
in APθ , we can state the following result. It can be regarded of the scalar version of our
main result, here we state the equality between the norm of the endomorphism cr(0, z) and the
distance between 0 and z in the Poincaré disk D.

Theorem 9.9. Let z ∈ D, then

1
kcr(0, z)kB(`z ) = d(0, z),
2
where k kB(`z ) denotes the norm of operators acting in `z ⊂ A2 .

(1 − z ∗ z)−1/2
 
Proof. Choose for `z the unital basis ez = . Then for any x = ez a ∈ `z ,
z(1 − z ∗ z)−1/2

cr(0, z)x = ez log((1 + |z|)(1 − |z|)−1 a,

and thus

hcr(0, z)x, cr(0, z)xi = a∗ log((1 + |z|)(1 − |z|)−1 hez , ez i log((1 + |z|)(1 − |z|)−1 a

= a∗ (log((1 + |z|)(1 − |z|)−1 )2 a.

Since (log((1 + |z|)(1 − |z|)−1 )2 ≤ k log((1 + |z|)(1 − |z|)−1 k2 , it follows that

a∗ (log((1 + |z|)(1 − |z|)−1 )2 a ≤ a∗ ak log((1 + |z|)(1 − |z|)−1 k2 ,

and therefore

khcr(0, z)x, cr(0, z)xik1/2 ≤ k log((1 + |z|)(1 − |z|)−1 kka∗ ak1/2

= k log((1 + |z|)(1 − |z|)−1 kkxk.

This implies that kcr(0, z)kB(`z ) ≤ k log((1 + |z|)(1 − |z|)−1 k. Note that

cr(0, z)ez = ez log((1 + |z|)(1 − |z|)−1 ,

so that
kcr(0, z)ez k2 = khez log((1 + |z|)(1 − |z|)−1 , ez log((1 + |z|)(1 − |z|)−1 k
= k log((1 + |z|)(1 − |z|)−1 k2 ,
i.e.,
kcr(0, z)kB(`z ) = k log((1 + |z|)(1 − |z|)−1 k = 2d(0, z).

24

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 10 The coefficient bundle
Consider the slight variant of the commutative diagram in (2)

K
πˆθ π̃θ

~ 
Qρ
' / D,

Recall (from the end of Section 4), that Qρ denotes the space of θ-orthogonal rank one projections
(considered here the coordinate free version of APθ ). Let us introduce the canonical bundle

ξ → Qρ ,

whose fiber over q ∈ Qρ is the module R(q). This is a fiber bundle of right A-modules, which
has a canonical connection. Elements of ξ are pairs (q, x), with q ∈ Qρ and q(x) = x.
Let q ∈ Qρ and ϕ : R(q) → R(q) a right module endomorphism. Pick a normalized generator
x ∈ R(q): θ(x, x) = 1 (i.e., an element of R(q) in K). Then the endomorphism ϕ is determined
by the value ϕ(x) = xa. That is, for any element y = xλ ∈ R(q), ϕ(y) = xaλ. In other words,
if we regard x as a basis for R(q), ϕ can be expressed as λ 7→ aλ. We call a ∈ A the matrix of ϕ
in the basis x. If x0 is another basis of R(q) in K, then there exists a unitary u ∈ UA such that
x0 = xu. If b is the matrix of ϕ in the basis x0 (i.e., ϕ(x0 ) = x0 b), then

xub = x0 b = ϕ(x0 ) = ϕ(xu) = ϕ(x)u = x0 au.

Then ub = au, which means that the matrix of ϕ in the basis x0 is b = u∗ au .

This shows that we can regard the set End(R(q)) of endomorphisms of R(q), as the set of
pairs (x, a), where x ∈ K with q(x) = x, and a ∈ A, with the identification

(x, a) ∼ (xu, u∗ au), u ∈ UA .

The map Γ → Qρ defined by (q, ϕ) 7→ q for q ∈ Qρ and ϕ ∈ End(R(q)), is a fiber bundle

which we call the coefficient bundle; alternatively (x, a) ∼ (xu, u∗ au) 7→ [x]. Each fiber of Γ is
a C∗ -algebra, which is isomorphic to A. The canonical connection of the bundle ξ induces a
connection in Γ, by the rule

(DX ϕ)(y) = (DX ϕ)y + ϕ(DX y).

Here, ϕ is a local cross section of Γ and y is a local cross section of ξ. We remark that the
connections of ξ and Γ are compatible with the action of U(θ).
We define the basic 1-form. Given x ∈ K and X ∈ (T Qρ )q , with q = xθ(x, ), put

κx (X) = Xx.

Note that X is a matrix in M2 (A), θ-symmetric and q-codiagonal: Xx ∈ x⊥θ = N (q).

Given X, Y ∈ (T Qρ )q , we define the product

hX, Y ix = −θ(κx (X), κx (Y )) = −θ(Xx, Y x).

25

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 If the generator x is changed for x0 = xu, we have

hX, Y ix0 = −θ(X(xu), Y (xu)) = −u∗ θ(Xx, Y x)u = u∗ hX, Y ix u.

This means that, given q = [x] = [x0 ], the product hX, Y iq is well defined as an element of the
fiber Γq .
This product is therefore a Hilbertian product in T Qρ , with values in Γ. To this effect, note
that T Qρ is a right module over the bundle Γ of coefficients. Indeed, if we fix x ∈ K with
q = xθ(x, ), the map X 7→ κx (X) = Xx from (T Qρ )q to N (q) is one to one. If we change
x into xu, κx (X) changes to κxu (X) = κx (X)u. If X ∈ (T Qρ )q and ϕ ∈ Γq , we define Xϕ as
κx (Xϕ) = Xxa, where ϕ is represented by the class of (x, a). With this definition we have

hX, Y ϕiq = hX, Y iq ϕ.

10.1 The cross ratio, the logarithm and the exponential

We have just defined a Hilbertian Γ-valued structure in Qρ ' APθ , or, equivalently, in D. In
particular, the product
hLog0 (z), Log0 (z)i0
 
1
takes values in the set of endomorphisms of `0 = , where Log0 is defined in Corollary
0
6.3. It is a positive module endomorphism (given by multiplying the generator e1 by a posi-
1/2
tive element of a). Thus, it has a unique positive square root hLog0 (z), Log0 (z)i0 , which we
denote by mod0 (Log0 (z)) of Log0 (z). Explicitly, in the

generator e1 , mod0 (Log0 (z)) consists in
multiplying the generator by log (1 + |z|)(1 − |z|)−1 .
On the other hand, we saw that, for z ∈ D, the endomorphism of `z denoted by cr(0,  z),
 is
1
given by the same coefficient log (1 + |z|)(1 − |z|)−1 , which multiplies the generator


of
z
`z .
We shall translate the endomorphism cr(0, z) from `z to `0 by means of the parallel transport
of APθ , along the geodesic δ, with δ(0) = `0 and δ(1) = `z (i.e., the same former δ, which under
the identification D ' APθ joins δ(0) = 0 and δ(1) = z in D: δ(t) = ω tanh(t|α|)).  
0 α∗ 
t
α 0
The parallel transport of elements of D (or APθ ) along the geodesic δ(t) = e · 0,
where α is, as in Remark 6.1.2
∞
X 1
α=z (z ∗ z)k ,
2k + 1
k=0
is given by 
the left action of the invertible matrix
0 α ∗
t 
α 0
e : `0 → `δ(t) . The endomorphism cr(0, z) of `z is transported to `0 as
   
0 α∗ 0 α∗
−   
α 0 α 0
cr(0, z)0 := e cr(0, z)e : `0 → `0 .

Our main result (for the origin) is the following:

26

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Theorem 10.1. With the current notation, if z ∈ D (or `z ∈ APθ ),

emod0 (Log0 (z)) = cr(0, z)0 or, equivalently, mod0 (Log0 (z)) = log(cr(0, z)0 ), (7)

where the exponential in the first equality is the usual exponential of A, log in the second equality
is the usuallogarithm
 of G, and each endomorphism of `0 is identified with its coefficient in the
1
basis e1 = .
0
Proof. Let us prove the first equality. Since we are comparing endomorphisms of `0 , it suffices
to show that they carry the generator e1 to the same element in A2 . Note that
   
0 α∗ 0 α∗
−    
α 0 α 0 1
cr(0, z)0 (e1 ) = e cr(0, z)e
0
 
0 α∗
−   
α 0 cosh(|α|)
=e cr(0, z) ;
ω sinh(|α|)
using that ω tanh(|α|) = δ(1) = z (Lemma 6.1), this yields
   
0 α∗ 0 α∗
−    −  
α 0 1 α 0 1
e cr(0, z) cosh(|α|) = e (1 + |z|)(1 − |z|)−1 cosh(|α|).
z z
 

0 α∗    
α 0 1 1
By the same computation that showed that e = cosh(|α|) (see the proof
0 z
of Lemma 6.1), we have that
 
0 α∗
−    
α 0 1 1
e = cosh(|α|)−1 ,
z 0
i.e.,  
1
cr(0, z)0 (e1 ) = (1 + |z|)(1 − |z|)−1 .
0
On the other hand, the endomorphism mod0 (Log0 (z) sends e1 to e1 log (1 + |z|)(1 − |z|)−1 ,

and thus  
1
emod0 (Log0 (z))
(e1 ) = (1 + |z|)(1 − |z|)−1 .
0

As in Definition 9.5, let z0 6= z1 ∈ D (`z0 6= `z1 ∈ APθ ). Pick g̃ ∈ U(θ) such that g̃ · 0 = z0 ,
and denote by z = g̃ −1 · z1 as before. Let δ be the geodesic such that δ(0) = 0 and δ(1) = z1 .
Then δz0 ,z1 = g̃ · δ is the geodesic which joins z0 and z1 at t = 0 and t = 1, respectively. Recall
that cr(z0 , z1 ) = g̃cr(z0 , z1 )g̃ −1 . Likewise, we put

Logz0 (z1 ) := g̃Log0 (z)g̃ −1 , and modz0 (z1 ) = hLogz0 (z1 ), Logz0 (z1 )i1/2
z0 ,

27

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 where hϕ, ψiz0 = g̃hg̃ϕg̃ −1 , g̃ψg̃ −1 i0 g̃ −1 , and Logz0 is the inverse of the exponential expz0 :
(T D)z0 → D. It is not difficult to verify that these definitions do not depend on the choice of g̃.
Finally, let us denote by cr(z0 , z1 )z0 the parallel transport of cr(z0 , z1 ) from `z1 to `z0 along
the geodesic δz0 ,z1 (obtained by conjugation as in the case of the origin, by the value at t = 1
of the one parameter group in U(θ) which determines δz0 ,z1 ). The U(θ)-covariance of the data
involved enables us to prove the following:

Theorem 10.2. With the current notations,

modz0 (Logz0 (z1 )) = log (cr(z0 , z1 )z0 ) .

In particular, kLogz0 (z1 )kz0 = k log(cr(z0 , z1 )z0 )k.

Figure 2

Remark 10.3. If we think of modz0 (Logz0 (z1 )) as an distance operator from z0 to z1 , then the
identity in the above theorem shows a projective way of computing this operator distance.

11 An example
Suppose that the algebra A has a trace tr onto a central subalgebra, that is, there exists a C∗ -
subalgebra B ⊂ Z(A) of the center of A and a conditional expectation tr : A → B satisfying
tr(xy) = tr(yx) for all x, y ∈ A. This happens, for instance, if A is a finite von Neumann
algebra.
A relevant case of this situation is the following. Consider a complex vector bundle E → B
with compact base space B, endowed with a Riemannian metric he, e0 ib , b ∈ B, e, e0 ∈ Eb (the
fiber of E over b). Consider the fiber bundle End(E) → B of endomorphisms of the vector bundle
E, and let A be the algebra Γ(End(E)) of the continuous global cross sections of End(E). Since

28

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 each Eb is a (finite dimensional) Hilbert space, End(Eb ) is a C∗ -algebra. The space Γ(End(E))
of cross sections has therefore the norm kϕk = supb∈B kϕb k, where ϕb : Eb → Eb and kϕb k is
the usual norm of linear operators. With this norm, A is a C∗ -algebra. The center Z(A) of
this algebra is the space of scalar sections λ in End(E) (homotetic in each fiber). The central
trace is given by tr : A → Z(A), tr(σ)b = T r(σb ), b ∈ B, with T r the usual trace of Eb . More
specifically, B could be a compact manifold, and E the complexification of its tangent bundle,
with an Hermitian metric. This case is interesting due to the following observation: in our
previous work [1], we noticed the equivalence, as homogeneous spaces, of the disk D and the
Poincaré halfspace H of the algebra A. This homogeneous space can be thought as the tangent
bundle T G+ of the space G+ of positive and invertible elements of A, as explained in [1]. In
this context, an element of H is a pair (X, a) with a ∈ G+ and X ∈ (T G+ )a . The element
a ∈ G+ represents a Riemannian metric in B, and a possible vector X (a selfadjoint element of
A) could be the Ricci curvature of the metric a. In this manner, the geometry of H is linked to
the deformation of the pairs (Riemannian metric, Ricci curvature), viewed as elements of T G+ .
Back to the general case (of this example):
trA → B ⊂ Z(A),
we can define a Hilbertian B-valued inner product, by means of
hX, Y itr,q = −tr(θ(Xx, Y y)).
Indeed, since tr is tracial, the value of −tr(θ(Xx, Y y)) is independent of the choice of x ∈ K
satisfying q = xθ(x, ). On the other hand, cr(z0 , z1 ) is an element of Γz0 , which has matrix a in
a unital base x ∈ R(q), as explained before. We put cr(z0 , z1 )tr , for tr(a). Clearly, cr(z0 , z1 )tr
does not depend on the basis x. With these notations, the formula in Corollary 10.2, can be
written
1/2
hLogz0 z1 , Logz0 z1 itr = log cr(z0 , z1 )tr , (8)
which is an identity involving elements in B.
More specifically, if A is commutative, we can choose tr the identity A = B, and we have

|Logz0 z1 | = log cr(z0 , z1 ), (9)

as elements in A.

References
[1] Andruchow, E.; Corach, G.; Recht, L., Poincaré half-space of a C ∗ -algebra, Rev. Mat.
Iberoam. (to appear), preprint arXiv:1711.08802.
[2] Andruchow, E.; Corach, G.; Stojanoff, D. Projective spaces of a C ∗ -algebra, Integral Equa-
tions Operator Theory 37 (2000), 143–168.
[3] Andruchow, E.; Corach, G.; Stojanoff, D., Projective space of a C ∗ -module, Infin. Dimens.
Anal. Quantum Probab. Relat. Top. 4 (2001), 289–307.
[4] Beltita, D., Smooth homogeneous structures in operator theory. Chapman & Hall/CRC
Monographs and Surveys in Pure and Applied Mathematics, 137. Chapman & Hall/CRC,
Boca Raton, FL, 2006.

29

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 [5] Corach, G.; Porta, H.; Recht, L., The geometry of the space of selfadjoint invertible elements
in a C ∗ -algebra, Integral Equations Operator Theory 16 (1993), 333–359.

[6] Corach, G.; Porta, H.; Recht, L., The geometry of spaces of projections in C ∗ -algebras,
Adv. Math. 101 (1993), 59–77.

[7] de la Harpe, P., The real symplectic group of a Hilbert space is essentially simple, J. London
Math. Soc. (2) 24 (1981), 293–307.

[8] Dixmier, J.; Maréchal, O., Vecteurs totalisateurs d’une algèbre de von Neumann, Comm.
Math. Phys. 22 (1971), 44–50.

[9] Harris, L. A., Bounded symmetric homogeneous domains in infinite dimensional spaces.
Proceedings on Infinite Dimensional Holomorphy (Internat. Conf., Univ. Kentucky, Lex-
ington, Ky., 1973), 13–40. Lecture Notes in Math., Vol. 364, Springer, Berlin, 1974.

[10] Kadison, R. V.; Ringrose, J. R., Fundamentals of the theory of operator algebras. Vol.
I. Elementary theory. Reprint of the 1983 original. Graduate Studies in Mathematics, 15.
American Mathematical Society, Providence, RI, 1997.

[11] Lance, E. C., Hilbert C ∗ -modules. A toolkit for operator algebraists. London Mathematical
Society Lecture Note Series, 210. Cambridge University Press, Cambridge, 1995.

[12] Manuilov, V. M.; Troitsky, E. V., Hilbert C ∗ -modules. Translated from the 2001 Russian
original by the authors. Translations of Mathematical Monographs, 226. American Mathe-
matical Society, Providence, RI, 2005.

[13] Ostrovskii, M. I.; Shulman, V. S.; Turowska, L., Unitarizable representations and fixed
points of groups of biholomorphic transformations of operator balls, J. Funct. Anal. 257
(2009), 2476–2496.

[14] Phillips, R. S., On symplectic mappings of contraction operators, Studia Math. 31 (1968),
15–27.

[15] Siegel, C.L., Symplectic geometry, Amer. J. Math. 65 (1943), 1–86.

[16] Young, N. J., Orbits of the unit sphere of L(H, K) under symplectic transformations, J.
Operator Theory 11 (1984), 171–191.

[17] Zelikin, M. I., Geometry of the cross ratio of operators. (Russian) ; translated from Mat.
Sb. 197 (2006), 39–54 Sb. Math. 197 (2006), 37–51.

Esteban Andruchow
Instituto de Ciencias, Universidad Nacional de Gral. Sarmiento,
J.M. Gutierrez 1150, (1613) Los Polvorines, Argentina
and Instituto Argentino de Matemática, ‘Alberto P. Calderón’, CONICET,
Saavedra 15 3er. piso, (1083) Buenos Aires, Argentina.
e-mail: eandruch@ungs.edu.ar

Gustavo Corach
Instituto Argentino de Matemática, ‘Alberto P. Calderón’, CONICET,

30

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.
 Saavedra 15 3er. piso, (1083) Buenos Aires, Argentina,
and Depto. de Matemática, Facultad de Ingenierı́a, Universidad de Buenos Aires, Argentina.
e-mail: gcorach@fi.uba.ar

Lázaro Recht
Departamento de Matemática P y A, Universidad Simón Bolı́var
Apartado 89000, Caracas 1080A, Venezuela
e-mail: recht@usb.ve

31

14 Aug 2019 12:33:09 PDT

Version 1 - Submitted to Adv. Math.