Subspaces with or without a common complement

Esteban Andruchow and Eduardo Chiumiento

arXiv:2412.18113v1 [math.FA] 24 Dec 2024

December 25, 2024

Abstract
Let H be a separable complex Hilbert space. Denote by Gr(H) the Grassmann manifold
of H. We study the following sets of pairs of elements in Gr(H):

∆ = {(S, T ) ∈ Gr(H) × Gr(H) : ∃Z ∈ Gr(H) such that S +̇Z = T +̇Z = H},

which are pairs of subspaces that have a common complement, and

Γ = Gr(H) × Gr(H) \ ∆,

which are pairs of subspaces that do not admit a common complement. We identify S ∼ PS ,
the subspace S with the orthogonal projection PS onto S. Thus we may regard ∆ and
Γ as subsets of B(H) × B(H) (here B(H) denotes the algebra of bounded linear operators
in H). We show that ∆ is open, and its connected components are parametrized by the
dimension and codimension of the subspaces. The connected component of ∆ having both
infinite dimensional and co-dimensional subspaces is dense in the corresponding component
of Gr(H) × Gr(H). On the other hand, Γ is a (closed) C ∞ submanifold of B(H) × B(H), and
we characterize the connected components of Γ in terms of dimensions and semi-Fredholm
indices. We study the role played by the geodesic structure of the Grassmann geometry of
H in the geometry of both ∆ and Γ. Several examples of pairs in ∆ and the connected
components of Γ are given in Hilbert spaces of functions.

2020 MSC: 47B02, 58B10, 58B20, 46E20

Keywords: common complement; Grassmann manifold; semi-Fredholm operator; geodesic;
Hilbert space of functions.

1 Introduction
These notes reflect on the paper [22] by M. Lauzon and S. Treil. These authors characterize
pairs of (closed) subspaces S and T of a Hilbert space H, which have a common complement
Z. That is, there exists a closed subspace Z ⊂ H such that

S +̇Z = T +̇Z = H,

where the symbol +̇ stands for direct sum. In this paper we investigate geometric aspects of the
set of all the pairs of subspaces having a common complement, and the set of all pairs that do
not admit a common complement,

1
 Denote by B(H) the algebra of bounded linear operators in a complex separable Hilbert
space H, P(H) the subset of orthogonal projections, and Gr(H) the Grassmann manifold of H.
Throughout, we identify subspaces in Gr(H) with their corresponding orthogonal projections in
P(H). Our objects of study will be the complementary sets

∆ := {(PS , PT ) ∈ P(H) × P(H) : S and T have a common complement},

and

Γ := {(PS , PT ) ∈ P(H) × P(H) : S and T do not have a common complement}.

The space P(H) is a complemented C∞ submanifold of B(H) (see [13, 28]). Therefore it is
natural to ask about the geometric structure of the sets ∆ and Γ. By an elementary argument,
or using a result by J. Giol [20], it can be shown that ∆ is an open subset of P(H) × P(H). Its
complement Γ is shown to be a (non complemented) closed C ∞ submanifold of P(H) × P(H).
Thus, both spaces have differentiable structure. We study how the geometry of these spaces
relates to geometry of P(H). Specifically, concerning the geodesics of P(H), which have been
thoroughly studied. We are also interested in the study of concrete examples of pairs of subspaces
in ∆ and Γ in the context of Hilbert spaces of functions.
Let us describe the contents of this paper. In Section 2 we recall preliminaries on common
complemented subspaces obtained by M. Lauzon and S. Treil [22], and also the results that were
later proved by J. Giol [20]. We also state basic facts on the following aspects of projections: the
geometry of P(H) [13, 28], Halmos’ model for a pair of subspaces [18], and the index of a pair
of projections [9].
In Section 3 we study the structure of the set ∆, and show that ∆ is an open subset of
P(H) × P(H). This is a consequence of a result by Giol [20]; we give though an elementary proof
of this fact. We characterize the connected components of ∆ in terms of ranks and co-ranks,
and the there is only one connected component ∆∞ having subspaces of both infinite dimension
and codimension. We show by using Halmos’ model that ∆∞ is dense in the corresponding
connected component of the (product) Grassmann manifold.
In Section 4 we consider the structure of the set Γ. The main result states that Γ is a (closed)
non complemented C ∞ submanifold of B(H) × B(H). For 1 ≤ i, j ≤ +∞, denote by Pi,j the
connected component of P(H) consisting of projections with rank i and co-rank j. We show that
Γ consists of the following parts:

• Pi,j × Pk,l , for i 6= k or j 6= l.

• For pairs of projections in P∞,∞ ,

Γ1 = {(PS , PT ) ∈ Γ : dim S ∩ T ⊥ < ∞, dim S ⊥ ∩ T < ∞}.

and

Γ∞ = {(PS , PT ) ∈ Γ : only one of S ∩ T ⊥ , S ⊥ ∩ T has infinite dimension}.

2
 Moreover, Γ1 splits in the following submanifolds, which are its connected components para-
matrized by the integers as follows:

Γn1 = {(PS , PT ) ∈ Γ1 : PS T
: T → S is a Fredholm operator of index n}, n ∈ Z.

The connected components of the submanifold Γ∞ are

Γl∞ := {(PS , PT ) ∈ P(H) × P(H) : dim S ∩ T ⊥ < ∞, dim S ⊥ ∩ T = +∞}

and
Γr∞ := {(PS , PT ) ∈ P(H) × P(H) : dim S ∩ T ⊥ = ∞, dim S ⊥ ∩ T < +∞}.

These results are based on previous results obtained in [5].

In Section 5 we discuss when geodesics of P(H) × P(H) remain inside ∆ or Γ along their
paths. We give an example which shows that arbitrary close points in ∆ may not be joined by
a minimizing geodesic of P(H) × P(H). On the other hand, cases where geodesics remain in
∆ or Γ can be are obtained by imposing conditions related to the ranges or nullspaces, or by
taking special kinds of projections such as the ones defined by the restricted Grassmannian (see
[12, 29, 30]).
In Section 6 we present several examples of subspaces with or without a common comple-
ment related to Hilbert spaces of functions. The existence of geodesics between shift-invariant
subspaces of the usual Lebesgue space L2 (T) (T the unit circle) has been related to the injec-
tivity problem for Toeplitz operators (see [6, 25, 27]). We now examine when several of those
examples discussed in relation to the existence of geodesics satisfy the more general condition
of having a common complement. Also we consider other examples related to Blaschke products
in the Hardy space of the unit disk [7] or the uncertainty principle [4]. In the case of pairs of
subspaces without a common complement, we further determine to which of the above described
component of Γ the pair belongs.

2 Basic facts
Let H be a separable complex Hilbert space. We denote by B(H) the algebra of bounded op-
erators acting on H. For T ∈ B(H), let R(T ) and N (T ) be the range and nullspace of T ,
respectively. The symbol +̇ stands for direct sum of subspaces of H; while we reserve the symbol
⊕ for orthogonal sums of subspaces. We write PS ∈ B(H) for the orthogonal projection with
range S. Given a decomposition H = S +̇Z, we will denote by PSkZ ∈ B(H) the idempotent
with range T and nullspace Z. Let us state below several preliminary facts.
Subspaces with a common complement. Two closed subspaces S, T ⊂ H have a common com-
plement if there exists a closed subspace Z ⊂ H such that S +̇Z = T +̇Z = H.

Remark 2.1. Let us state the following results from [22] by M. Lauzon and S. Treil.
1. The subspaces S, T have a common complement if and only if there exist a bounded (not
necessarily orthogonal) projection E onto one of the subspaces (say T ), such that

E S
:S→T

3
is an isomorphism ([22, Prop. 1.3]). This result is valid even in the context of Banach spaces.

2. Denote by G = PT S , regarded as an operator G : S → T (so that G∗ : T → S is PS T

).
Denote by E the projection valued spectral measure of G∗ G. Note that

N (G) = S ∩ T ⊥ and N (G∗ ) = S ⊥ ∩ T .

Then S and T have a common complement if and only if

dim N (G) + dim E(0, 1 − ǫ)S = dim N (G∗ ) + dim E(0, 1 − ǫ)S (1)

for some ǫ > 0 (equivalently, for all sufficiently small ǫ > 0). This characterization also holds for
non separable Hilbert spaces ([22, Thm. 0.1]).

3. As a straightforward consequence of the previous statement, Lauzon and Treil oberved that
the subspaces S and T do not have a common complement in a separable Hilbert space H if
and only if dim S ∩ T ⊥ 6= dim S ⊥ ∩ T and the operator (1 − G∗ G) N (G)⊥ is compact ([22, Rem.
0.5]).

Remark 2.2. Later on J. Giol proved the following equivalences (see [20, Prop. 6.2]):

i) S and T are subspaces with a common complement.

ii) There exists P ∈ P(H) such that kPS − P k < 1 and kP − PT k < 1.

iii) There exist four idempotents P1 , P1′ , P2 , P2′ such that S = R(P1 ), R(P1′ ) = R(P2′ ), R(P2 ) =
T and S ⊥ = N (P1′ ), N (P1 ) = N (P2 ), N (P2′ ) = T ⊥ .

The Grassman manifold. The Grassmann manifold Gr(H) of H is defined as the set of all the
closed subspaces of H. We identify the Grassmann manifold with the manifold of all orthogonal
projections in H given by

P(H) = {P ∈ B(H) : P = P 2 = P ∗ }.

We observe that this identification can be made more precise: there is a diffeomorpshim between
the presentation as subspaces and orthogonal projections (see [1]). P(H) is a complemented
submanifold of B(H) with tangent space (T P(H))P at P ∈ P(H) given by

(T P(H))P = {XP − P X : X = −X ∗ ∈ B(H)}.

Notice that tangent vectors are co-digonal selfadjoint operators. For each projection P , one can
define the idempotent EP : B(H) → (T P)P , EP (X) = P X(1 − P ) + (1 − P )XP . The family
{EP }P ∈P(H) induces a linear connection in P(H): for X(t) a tangent vector field along a curve
γ(t) ∈ P(H), set
DX
= Eγ (Ẋ).
dt
Let Gl(H) be the group of all the invertible operators, and U (H) the subgroup consisting of
unitary operators. Recall that U (H) acts on P(H) by U · PS = PU (S) = U PS U ∗ , U ∈ U (H).

4
 The geodesics of this connection can be computed (see [28], [13]): they are induced by one
parameter unitary groups, namely, the geodesics starting a S are of the form:

δ(t) = eitZ PS e−itZ ,

where Z ∗ = Z is co-diagonal with respect to S, meaning that Z(S) ⊂ S ⊥ and Z(S ⊥ ) ⊂ S (i.e.,
Z has co-diagonal matrix in terms of the decomposition H = S ⊕ S ⊥ ). It is useful to pick also
kZk ≤ π/2, we shall call these normalized geodesics. It is known (see [13, 28] or also [2]) that
there exists a geodesic between S and T if and only if

dim S ∩ T ⊥ = dim S ⊥ ∩ T . (2)

Moreover, there exists a unique normalized geodesic joining S and T if and only if S ∩ T ⊥ =
{0} = S ⊥ ∩ T .
We shall measure the distance between closed subspaces S, T ⊂ H by means of the corre-
sponding orthogonal projections and the usual operator norm:

d(S, T ) := kPS − PT k.

If kPS − PT k < 1, then one can find a unitary operator U = US,T which is a C ∞ expression in
terms of PS , PT such that U S = T . There is more than one way to do this, but if one chooses
the geometric argument developed in [28] and [13], based on the existence of a unique minimal
geodesic of the Grassmann manifold, one has
 
arcsin(kPS − PT k)
kUS,T − 1k = 2 sin .
2

We remark that the connected components of P(H) are parametrized by the rank and the
co-rank. We denote by Pi,j the connected component of P(H) consisting of projections with
rank i and corank j, where the indices satisfy 0 ≤ i, j ≤ ∞ and i + j = ∞ (usual convention if
both are infinite).
Halmos’ model for a pair of projections / subspaces. Several papers consider the following de-
composition of a Hilbert space in the presence of two projections (or subspaces): mainly Dixmier
[15], Halmos [18]. We transcribe Halmos’ model. Let P , Q be ortohogonal projections in H, then
the following five space decomposition reduces P and Q:

H = R(P ) ∩ R(Q) ⊕ N (P ) ∩ N (Q) ⊕ R(P ) ∩ N (Q) ⊕ N (P ) ∩ R(Q) ⊕ H0 , (3)

where H0 is the orthogonal complement of the sum of the former four, and is called the generic
part of P and Q. The subspace H0 reduces both P and Q, we shall denote the reductions by
P0 := P H0 , Q0 := Q H0 .
There exists a unitary isomorphism between H0 and a product space L × L, and a posi-
tive operator X in L with N (X) = {0} and kXk ≤ π/2, such that the projections P and Q
correspond, under this isomorphism, with the matrices (in L × L)
! !
1 0 C 2 CS
P0 ∼ and Q0 ∼ .
0 0 CS S 2
Here C = cos(X) and S = sin(X).

5
Remark 2.3. The subspace L is determined by the following decomposition

R(P ) = R(P ) ∩ R(Q) ⊕ R(P ) ∩ N (Q) ⊕ L.

Take S = R(P ), T = R(Q) in Remark 2.1, then N (G)⊥ = S ∩ T ⊕ L and the operator
(1 − G∗ G)|N (G)⊥ is compact if and only if S 2 = sin(X) is compact in L. Equivalently, X is
compact in L. Thus, (S, T ) ∈ ∆ if and only if dim S ⊥ ∩ T = dim S ∩ T ⊥ or X is not compact.
This characterization was observed in [20, Rem. 4.2].

Index of a pair of projections. The following facts were taken from [9]. A pair (P, Q) of orthogonal
projections is called a Fredholm pair if

Q : R(P ) → R(Q)
R(P )

is a Fredholm operator. Its index is

 
index(P, Q) = index Q : R(P ) → R(Q) = dim R(P ) ∩ N (Q) − dim N (P ) ∩ R(Q).
R(P )

Among the various characterizations of Fredholm pairs we transcribe the following.

Remark 2.4. (P, Q) is a Fredholm pair if and only if

1. +1 and −1 are isolated in the spectrum σ(P − Q) of P − Q.

2. R(P ) ∩ N (Q) and N (P ) ∩ R(Q) are finite dimensional.

Notice that a condition for a pair of projections (P, Q) that is stronger than being a Fredholm
pair, which shall appear later on, is that P − Q is compact. It is an easy exercise that this is in
fact so.

3 Structure of the set of pairs of subspaces with a common

complement
In this section we discuss elementary properties of ∆. We begin by determining its connected
components.

Remark 3.1. Recall that Pi,j denotes de connected components of P(H). If i = k < ∞
or j = l < ∞, then Pi,j × Pk,l ⊆ ∆. Indeed, take (P, Q) ∈ Pi,j × Pk,l with i = k. Since
T := Q|R(P ) : R(P ) → R(Q) is an operator defined in finite-dimensional spaces, we have k =
dim N (T ) + dim R(T ) = dim N (T ∗ ) + dim R(T ∗ ). From dim R(T ) = dim N (T )⊥ = dim R(T ∗ ), it
follows that dim R(P )∩N (Q) = dim N (T ) = dim N (T ∗ ) = dim R(Q)∩N (P ). Hence (P, Q) ∈ ∆.
Next we observe that the case where j = l < ∞ follows by using the previous case and the
map ⊥: P(H) × P(H) → P(H) × P(H), ⊥ (P, Q) = (P ⊥ , Q⊥ ). This is an isomorphism that
preserves connected components, and from the characterization in Remark 2.2, one easily sees
that ⊥ (∆) = ∆.

6
 On the other hand, assume now that i 6= k or j 6= l, and take (P, Q) ∈ Pi,j × Pk,l with
i 6= k. Thus, R(P ) nad R(Q) cannot be isomorphic, and according to Remark 2.1 i), we get
(P, Q) ∈/ ∆. One can consider also the case where (P, Q) ∈ Pi,j × Pk,l with j 6= l, which follows
again that (P, Q) ∈
/ ∆ by using the map ⊥ as above.
From these facts, we obtain that

∆ij := ∆ ∩ (Pi,j × Pi,j ) = Pi,j × Pi,j ,

whenever i < ∞ or j < ∞, are the only connected components of ∆ with finite dimensional
rank or corank.
To analize the case of P∞,∞ we need the following (known) property:
Lemma 3.2. Let P, Q be orthogonal projections such that kP − Qk < 1, and let P (t) be the
unique minimal geodesic of P(H) such that P (0) = P and P (1) = Q. Then for all t ∈ [0, 1] we
have that kP − P (t)k < 1.
Proof. P (t) is of the form P (t) = eitX P e−itX , for X ∗ = X, P -co-diagonal with kXk < π/2 (see
Section 2). That X is P -co-diagonal means that X anti-commutes with 2P − 1. Then

2P (t) − 1 = eitX (2P − 1)e−itX = ei2tX (2P − 1),

and thus
1
kP − P (t)k = k(2P − 1) − (2P (t) − 1)k = k(ei2tX − 1)(2P − 1)k = kei2tX − 1k < 1,
2
because k2tXk ≤ 2kXk < π, by a standard functional calculus argument.

Theorem 3.3. The subset ∆∞ of ∆, consisting of pairs of pojections in P∞,∞ with a common
complement, is arcwise connected. Therefore, in view of the above remark, ∆∞ is the connected
component of such pairs.
Proof. Let (PS , PT ) ∈ ∆∞ . We shall prove that there is a continuous path inside ∆ connecting
(PS , PT ) with (PS , PS ). We proceed in two steps. First we show that there is a continuous path
inside ∆ connecting (PS , PT ) with a pair (PS , E) such that kPS − Ek < 1. Indeed, by Remark
2.2 there exists E ∈ P(H) such that kPS − Ek < 1 and kPT − Ek < 1. Let E(t) be the minimal
geodesic of P(H) with E(0) = E and E(1) = PT . Then the curve (PS , E(t)) remains inside ∆
for t ∈ [0, 1]. This follows again using the result by Giol, for we have the intermediate projection
E satisfying kPS − Ek < 1 and kE(t) − Ek < 1 (by Lemma 3.2). Next, we find a continuous path
inside ∆ connecting (PS , E) with (PS , PS ). Let P (t) be the minimal geodesic joining P (0) = PS
and P (1) = E. Then the curve (PS , P (t)) remains inside ∆ for t ∈ [0, 1], since by Lemma 3.2
we know that kPS − P (t)k ≤ kP − Ek < 1.
The proof of the theorem follows showing that any two pairs (PS , PS ) and (PS ′ , PS ′ ) with
S, S ′ infinite and co-infinite, can be joined by a continuous path inside ∆. To this effect, note
that if S and T have a common complement Z and U is a unitary operator, then U S and U T also
have a common complement (namely U Z), i.e., (PS , PT ) ∈ ∆ implies that (U PS U ∗ , U PT U ∗ ) =
(PU S , PU T ) ∈ ∆. Then, since PS , PS ′ ∈ P∞,∞ , there exists a continuous path of unitaries U (t)
such that U (0) = 1 and U (1)S = S ′ . Then (U (t)PS U ∗ (t), U (t)PS U ∗ (t)) is a continuous curve in
∆ wich joins (PS , PS ) and (PS ′ , PS ′ ).

7
Theorem 3.4. The set ∆ is open in P(H) × P(H), and consequently, ∆ is a submanifold of
P(H) × P(H).

Proof. The statement follows immediately from Giol’s characterization in Remark 2.2 ii), which
is clearly an open condition. We also give here an elementary proof based on the first item
of Remark 2.1. Pick (PS , PT ) ∈ ∆. Let S ′ be close to S and T ′ close to T . Suppose that
kPS ′ −PS k < 1 and kPT ′ −PT k < 1. Then there exists a unitary operator U = US,S ′ (a continuous
explicit expression in terms of PS and PS ′ ) such that U S = S ′ , and kU − 1k is controlled by
kPS ′ − PS k. Similarly, there exists a unitary operator V with analogous properties for T and
T ′ (e.g., V T = T ′ , etc.). Then there exists a projection E onto T such that E S : S → T is an
isomorphism. We fix E, and we claim that for T ′ sufficiently close to T , EV ∗ S : S → T is also
an isomorphism. Indeed, note that

kEV ∗ − Ek = kE(V ∗ − 1)k ≤ kEkkV − 1k.

In particular, we can restrict both EV ∗ and E to S, and regard these operators as elements
in B(S, T ). Since E is an isomorphism by the open mapping theorem there exists a constant
rE = rE,S,T (depending on E fixed, S and T ) such that for any T ∈ B(S, T ), kT − Ek < rE
rE
implies that T is an isomorphism. It follows that if T ′ is close enough to T so that kV −1k < kEk ,
∗ ′ ∗
then EV S : S → T is an isomorphism. Since V maps T onto T , V EV S : S → T is also an ′

isomorphism. Note that V EV ∗ is a (not necessarily orthogonal) projection onto T ′ . We claim

that V EV ∗ S ′ : S ′ → T ′ is also an isomorphism for S ′ close enough to S. Since U S = S ′ and
V T = T ′ , this is equivalent to saying that

EV ∗ S′
U∗ S
= EV ∗ U ∗ S
:S→T

is an isomorphism. Note that

kEV ∗ U − Ek ≤ kEkk1 − V ∗ U k = kEkkV − U k ≤ kEk(kV − 1k + k1 − U k).

rE rE
It follows that if kV − 1k < 2kEk and kU − 1k < 2kEk , then EV ∗ U S
is an isomorphism between
S and T . Thus, S ′ and T ′ have a common complement.

As we have observed above the connected components of ∆ij of ∆, where i < ∞ or j < ∞,
coincide with the whole corresponding component of P(H)×P(H). For the connected component
∆∞ we have the following.

Theorem 3.5. ∆∞ is dense in P∞,∞ × P∞,∞ .

Proof. Pick (P, Q) ∈ P∞,∞ \ ∆∞ . Notice that we can take a unitary conjugation of both pro-
jections to prove the statement, so by Halmos’ model we may assume

P = 1 ⊕ 0 ⊕ 1 ⊕ 0 ⊕ P0 ,
Q = 1 ⊕ 0 ⊕ 0 ⊕ 1 ⊕ Q0 ,

where we follow the order of the subspaces in (3). The projections P0 , Q0 acting on L × L, are
given by ! !
1 0 C 2 CS
P0 = and Q0 = ,
0 0 CS S 2

8
where C = cos(X), S = sin(X), and kXk ≤ π2 . Then, (P, Q) ∈ / ∆ implies that X is compact
(see Remark 2.3). We consider two cases.
In the first case, we assume that L is infinite dimensional. Recall that the operators S =
sin(X) and C = cos(X) are injective in Halmos’model, hence we get that neither 0 nor π2 are
eigenvalues of X. Therefore, its spectral decomposition must be
∞
X
X= λk Pk ,
k=1

where {λk }k≥1 is a sequence of positive eigenvalues in (0, π2 ) that converges to 0 and {Pk }k≥1 is
the corresponding sequence of finite-rank spectral projections. For ǫ > 0 sufficiently small, the
operators of the form
X
Xǫ := ǫ1L + λk Pk
k : k≥2ǫ
π
have spectrum contained in (0, 2 ).
This implies that Sǫ = sin(Xǫ ) and Cǫ = cos(Xǫ ) are positive
injective operators, and the projections acting on L defined by
!
Cǫ2 Cǫ Sǫ
Qǫ,0 =
Cǫ Sǫ Sǫ2

clearly satisfy Qǫ,0 → Q0 in the operator norm as ǫ → 0. Next take the projections acting on H
defined by
Qǫ := 1 ⊕ 0 ⊕ 0 ⊕ 1 ⊕ Qǫ,0 .

By using that Sǫ and Cǫ are injective operators, one can easily verify that R(P0 ) ∩ R(Qǫ,0 ) =
R(P0 ) ∩ N (Qǫ,0 ) = {0}. This implies R(P ) ∩ R(Qǫ ) = R(P ) ∩ R(Q) and R(P ) ∩ N (Qǫ ) =
R(P ) ∩ N (Q), and consequently, we obtain the following decomposition

R(P ) = R(P ) ∩ R(Qǫ ) ⊕ R(P ) ∩ N (Qǫ ) ⊕ L,

for ǫ ≥ 0 small enough. Consider the operators Gǫ = Qǫ |R(P ) introduced in Remark 2.1. Since
N (Gǫ )⊥ = R(P ) ∩ R(Q) ⊕ L by our previous computations, it follows that (1 − G∗ǫ Gǫ )|N (Gǫ )⊥
is not compact because Sǫ , or equivalently Xǫ , are not compact for ǫ > 0 sufficiently small. We
thus get (P, Qǫ ) ∈ ∆ and (P, Qǫ ) → (P, Q).
In the second case, we assume that L is finite dimensional. Recall that L is defined by
R(P ) = R(P ) ∩ R(Q) ⊕ R(P ) ∩ N (Q) ⊕ L. Since P, Q ∈ P∞,∞ , it must be dim R(P ) ∩ R(Q) = ∞
or dim R(P ) ∩ N (Q) = ∞. But (P, Q) ∈ Γ, so that dim R(P ) ∩ N (Q) 6= dim N (P ) ∩ R(Q). Then,
using that H is separable, we may assume dim R(P ) ∩ N (Q) < ∞ because we can interchange
the roles of P and Q if necessary. Hence we obtain dim R(P ) ∩ R(Q) = ∞.
As we have observed in Remark 3.1 the sets Γ and ∆ are invariant by taking orthogo-
nal complements . Therefore, (P ⊥ , Q⊥ ) ∈ Γ. There is a corresponding subspace L′ defined by
N (P ) = N (P ) ∩ N (Q) ⊕ N (P ) ∩ R(Q) ⊕ L′ for the Halmos’ model of (P ⊥ , Q⊥ ). If dim L′ = ∞,
then (P ⊥ , Q⊥ ) can be approximated by pairs in ∆ by the first case we have considered,
and consequently, the same holds for (P, Q) by taking orthogonal complements. Thus, we
assume dim L′ < ∞. By the same argument given in the previous paragraph, we also get
dim N (P ) ∩ N (Q) = ∞.

9
 We take the decomposition H = K1 ⊕ K2 ⊕ K3 , where K1 = R(P )∩ R(Q), K2 = N (P )∩ N (Q)
and K3 is the orthogonal complement of the first two subspaces. Since K1 and K2 have both
infinite dimension and co-dimension, we can conjugate by a unitary operator the projections P
and Q, and write them in terms of H = K1 × K1 × K3 as follows
   
1 0 0 1 0 0
P = 0 0 0  , Q = 0 0 0  ,
   

0 0 P3 0 0 Q3

for some projections P3 and Q3 . Then, we define the following projections

cos2 (ǫ)
 
cos(ǫ) sin(ǫ) 0
Qǫ := cos(ǫ) sin(ǫ) sin2 (ǫ) 0  , ǫ > 0.
 

0 0 Q3

Clearly, we have Qǫ → Q. Take the corresponding Gramian operators Gǫ for P and Qǫ . Since
N (Gǫ )⊥ = K1 ⊕ M, for some subspace M of K3 , and 1 − cos2 (ǫ) = sin2 (ǫ) = sin2 (ǫ)1K1 is not
compact because dim K1 = ∞, we find that (1 − P Qǫ P )|N (Gǫ )⊥ is neither compact. We conclude
that (P, Qǫ ) ∈ ∆ for all sufficiently small ǫ > 0.

4 Pairs of subspaces without a common complement

.4.1 Geometric structure
Let us consider now

Γ := {(PS , PT ) ∈ P(H) × P(H) : S, T do not have a common complement}.

Remark 4.1. According to Remark 3.1, we have that

Γijkl := Γ ∩ (Pi,j × Pk,l ) = Pi,j × Pk,l ,

whenever i 6= k or j 6= l, are the only connected components of Γ with finite dimensional rank
or corank. Hence we are left to understand the structure of pairs in P∞,∞ × P∞,∞ without a
common complement.

The facts recalled in Remark 2.1 3. can be used to rewrite

Γ = {(PS , PT ) : PS PT⊥ is compact in (S ∩ T ⊥ )⊥ and dim S ∩ T ⊥ 6= dim S ⊥ ∩ T }.

Denote by P0 and Q0 the reductions of PS and PT to H0 . We shall see below (in Subsection 4.2)
that the condition
PS PT⊥ is compact in (S ∩ T ⊥ )⊥ ,

can be replaced by the condition

A0 := P0 − Q0 is compact;

10
or by the condition
P0 Q⊥
0 is compact;

or also by
either PS PT⊥ or PT PS⊥ is compact.

Given a L is a Hilbert space, we denote by K(L) ⊂ B(L) the ideal of compact operators in
L, and by
πL : B(L) → B(L)/K(L) := C(L)

the ∗-epimorphism onto the Calkin algebra C(L). For a projection P , denote by r(P ) =
dim R(P ) ≤ +∞ the rank of P . In [5], the set

C = {(P, Q) ∈ P(H) × P(H) : P Q ∈ K(H)}

was studied.

Remark 4.2. We recall the basic facts from [5] needed here:

1. Let (P, Q) ∈ C such that both P and Q have infinite rank. The projections π(P ) and π(Q)
are mutually orthogonal, and therefore can be written as 2 × 2 matrices in terms of π(P )
as ! !
1 0 0 0
π(P ) = and π(Q) = .
0 0 0 q
Since both P and Q have infinite rank, we have (in terms of the above 2 × 2 matrix
description):
C1 := {(P, Q) : q = 1 in C(R(P )⊥ )},

and
C∞ := {(P, Q) : q is a proper projection (6= 0, 1) in C(R(P )⊥ )}.

The elements (P, Q) in the class C1 have finite Fredholm index.

We recall the following results from [5].

• (Theorems 6.5 and 7.6) The connected components of C are

C∞ and C1n = {(P, Q) ∈ C1 : index(P, Q) = n}.

• (Theorem 8.5) the set C is a C ∞ (non complemented) submanifold of B(H) × B(H). There-
fore C∞ C1 are sumanifolds of B(H) × B(H).

• (Proposition 6.6) Let (P, Q) ∈ C, then (P, Q) ∈ C1 if and only if dim N (P ) ∩ N (Q) < ∞.

Then we have the following:

Theorem 4.3. Let S, T be closed subspaces of both infinite dimension and codimension. The
following are equivalent

11
 1. (PS , PT ) ∈ Γ with
dim S ∩ T ⊥ < ∞ and dim S ⊥ ∩ T < ∞.

2. (PS , PT⊥ ) ∈ C1n with n 6= 0.

Proof. 2) =⇒ 1): By Proposition 6.6 in [5] (cited above), we have that (PS , PT⊥ ) ∈ C1 if and
only if
dim N (PS ) ∩ N (PT⊥ ) = dim S ⊥ ∩ T < ∞.

On the other hand, the fact that PS PT⊥ is compact forces that the intersections of the ranges
(where this product acts as the identity) be of finite dimension:

dim R(PS ) ∩ R(PT⊥ ) = dim S ∩ T ⊥ < ∞.

Since the index n 6= 0, these numbers are different, and therefore S and T do not have a common
complement. 1) =⇒ 2): straightforward, using the above computations.

Denote by
Γ1 = {(PS , PT ) ∈ Γ : (PS , PT⊥ ) ∈ C1 },

which due to the above result above coincides with

Γ1 = {PS , PT ) ∈ Γ : dim S ∩ T ⊥ < ∞ and dim S ⊥ ∩ T < ∞}.

Moreover, put
Γn1 = {(PS , PT ) ∈ Γ : (PS , PT⊥ ) ∈ C1n }.

Corollary 4.4. Γ1 is a C ∞ (non complemented) submanifold of P(H) × P(H). Its connected

components are Γn1 , n ∈ Z.

Proof. Straightforward, using the previous theorem, and the results from [5] cited above: the
global diffeomorphism t : P(H) × P(H) → P(H) × P(H),

t(P, Q) = (P, Q⊥ )

maps Γn1 onto C1n .

We have the following elementary corollaries to Theorem 4.3:

Corollary 4.5. (P, Q) ∈ Γn1 if and only if (Q, P ) ∈ Γ−n

1

Proof. We have seen that

Γ1 = {(PS , PT ) ∈ Γ : dim S ∩ T ⊥ < ∞, dim S ⊥ ∩ T < ∞}.

Clearly these conditions are symmetric, and also

index(PS , PT ) = −index(PT , PS ).

12
 Recall from Section 2 the definition of Fredholm pairs.

Corollary 4.6. (PS , PT ) ∈ Γn1 if and only if (P, Q) is a Fredholm pair with index n.

Proof. Clearly we have that (PS , PT ) ∈ Γ1 if and only if (PS , PT⊥ ) ∈ C and (PT , PS⊥ ) ∈ C are
compact, i.e.,
PS − PS PT PS and PT − PT PS PT are compact,

i.e.,
PT S
:S→T

is a Fredhom operator. Its index is dim S ∩ T ⊥ − dim S ⊥ ∩ T .

Let us now consider the complementary part

Γ∞ := {(PS , PT ) ∈ Γ : dim S ∩ T ⊥ = +∞ or dim S ⊥ ∩ T = +∞}.

Clearly only one of the two dimensions can be infinite in each case, and then this set parts into
two disjoint subsets
Γ∞ = Γl∞ ∪ Γr∞ ,

where
Γl∞ := {(PS , PT ) ∈ P(H) × P(H) : dim S ∩ T ⊥ < ∞, dim S ⊥ ∩ T = +∞}

or equivalently

Γl∞ = {(PS , PT ) ∈ Γ : PS PT⊥ compact and PT PS⊥ non compact},

and

Γr∞ := {(PS , PT ) ∈ P(H) × P(H) : dim S ⊥ ∩ T < ∞, dim S ∩ T ⊥ = +∞}

= {(PS , PT ) ∈ Γ : PT PS⊥ compact and PS PT⊥ non compact}.

Corollary 4.7. Both sets ΓL r

∞ , Γ∞ are C
∞ (non complemented) submanifolds of P(H) × P(H),

which are diffeomorphic to C∞ . Namely:

(PS , PT ) 7→ (PS , PT⊥ ) maps Γl∞ onto C∞

and
(PS , PT ) 7→ (PS⊥ , PT ) maps Γl∞ onto C∞ .

Proof. Use Theorem 8.5 of [5].

Remark 4.8. The pairs (PS , PT ) ∈ Γl∞ are also characterized by the condition that PS T → S
is a semi-Fredholm operator of index +∞. Indeed, PS − PS PT PS is compact, and thus PS PT PS
is a compact perturbation of the identity, as an operator in S; i.e., PT |S : S → T has a left
inverse, but not a right inverse.
Similarly, (PS , PT ) ∈ Γr∞ if and only if PT S : S → T has a right inverse, but not a left
inverse.

13
 Consequently, we have the following result.

Corollary 4.9. The connected components of Γ∞ are Γl∞ and Γr∞ .

Proof. Let (PS(t) , PT (t) ) be a continuous path in Γ∞ , t ∈ [0, 1]. In particular, PS(t) and PS(t) are
continuous paths in P(H). Since for each P ∈ P(H) the map

U (H) → P(H), U 7→ U P U ∗

is a fiber bundle, there exist continuous curves U (t), V (t) in U (H), t ∈ [0, 1], such that

U (t)PS(0) U ∗ (t) = PS(t) and V (t)PT (0) V ∗ (t) = PT (t) for t ∈ [0, 1].

In particular, U (t) maps S(0) onto S(t), and V (t) maps T (0) onto T (t). Then
 
U ∗ (t) PS(t) T (t) V (t) = U ∗ (t)P (t)V (t) t(0) : T (0) → S(0)

is a continuous path of semi-Fredholm operators between the (fixed) spaces T (0) and S(0). It
folllows that the index (be it +∞ or −∞) remains constant, i.e. the curve (PS(t) , PT (t) ) remains
entirely inside Γl∞ or Γr∞ .
Clearly, since both Γl∞ and Γr∞ are homeomorphic to C∞ , they are connected.

Proposition 4.10. The subsets Γn1 (n 6= 0), Γl∞ and Γr∞ are relatively open in Γ.

Proof. Fix (PS0 , PT0 ) ∈ Γn1 , for n 6= 0. Let (PS , PT ) ∈ Γ1 such that kPS − PS0 k < r < 1 and
kPT − PT0 k < r < 1. Then there exist unitaries U0 , V0 with kU0 − 1k, kV0 − 1k < δr , and δr → 0 if
r → 0, such that U0 S0 = S and V0 T0 = T (i.e., U0 PS0 U0∗ = PS and V0 PT0 V0∗ = PT ). Then, since
PS determines a Fredholm operator between T and S, U0∗ PS V0 determines a Fredholm operator
between T0 and S0 , with the same index. Note that (since U0∗ PS U0 = PS0 )

kU0∗ PS V0 − PS0 k ≤ kU0∗ PS V0 − U0∗ PS U0 k = k(U0∗ PS )(U0 − V0 )k ≤ kU0 − V0 k

≤ kU0 − 1k + k1 − V0 k < 2δr .

In particular, this implies that kU0∗ PS V0 T0 − PS0 T0 k < 2δr in B(S0 , T0 ). This implies that for r
small enough (depending on PS0 T0 ), U0∗ PS V0 T0 has the same index as PS0 T0 . It follows that (for
such values of r) (PS , PT ) ∈ Γn1 . The argument for the components Γl∞ and Γr∞ is similar.

Corollary 4.11. Γ is a smooth (non complemented) submanifold of P(H) × P(H).

4.2 Other characterizations of Γ

We present other characterizations of Γ in terms of sums or products of projections. In [14]
Chandler Davis studied operators A = P − Q, where P and Q are orthogonal projections. They
are selfadjoint contractions with certain spectral symmetries which we describe below.

Remark 4.12. Let P, Q ∈ P(H) and denote A = P − Q. The first two facts are elementary.
The third asertion can be found in [14].

14
 • N (A) = N (P ) ∩ N (Q) ⊕ R(P ) ∩ R(Q).

• N (A − 1) = R(P ) ∩ N (Q) and N (A + 1) = N (P ) ∩ R(Q).

• Denote by H0 = H⊖(N (A) ⊕ N (A − 1) ⊕ N (A + 1)) (usually named the generic part of P

and Q). Note that H0 reduces (P , Q and) A, denote by A0 = A H0 . Then Davis proved that
there exists a symmetry V = VP,Q (by symmetry we mean an operator V = V ∗ = V −1 ,
V P V = Q and V QV = P ). In particular, V A0 V = −A0 . The spectrum σ(A0 ) of A0 is
contained in [−1, 1]. The existence of this symmetry V implies that σ(A0 ) is symmetric
with respect to the origin, λ ∈ σ(A0 ) if and only if −λ ∈ σ(A0 ), and the multiplicity
function is symmetric.

The spectral symmetry of the reduction A0 may break for A at the extremes of the spectrum:
N (A − 1) and N (A + 1) may have different dimensions.
Chandler Davis in fact showed much more: that the decompositions of A0 (as diference of
projections) are in one to one correspondence with the operators V implementing this symmetry,
with explicit formulas to recover P and Q from V .

We claim that the condition that S and T admit a common complement is in fact a property
of the operator A = PS − PT . To this purpose, we shall need the following facts from [3], which
relate the difference A with the product B := PS PT .
We shall say that T is S-decomposable if it has a singular value (or Schmidt) decomposition,
X X
T = sn h , ξn iψn = sn ψn ⊗ ξn , (4)
n≥1 n≥1

where {ξn : n ≥ 1} and {ψn : n ≥ 1} are orthonormal systems, and sn > 0. In this case,
{ψn }, {ξn } are orthonormal bases of R(T ), N (T )⊥ , respectively and T ξn = sn ψn , T ∗ ψn = sn ξn ,
T ∗ T ξn = s2n ξn , T T ∗ ψn = s2n ψn for all n ≥ 1.

Remark 4.13. (see [3, Thms. 2.2, 2.4 and Rem. 2.3]) Let P, Q be orthogonal projections.

1. Suppose that B = P Q is S-decomposable with singular values sn . Then A = P − Q is

diagonalizable, with eigenvalues ±(1 − s2n )1/2 , n ≥ 1, and maybe 0, −1 and 1.

2. If A = P − Q is diagonalizable with (non zero) eigenvalues ±λn (0 < |λn | < 1) and ±1 ,
then B = P Q is S-decomposable with singular values (1 − λ2n )1/2 and 1.

3. Except for 1 and −1, the eigenvalues (1 − s2k )1/2 and −(1 − s2k )1/2 of A have the same
multiplicity. For the eigenvalues in (−1, 1) (or the singular values in (0, 1) we have the
following relation: the multiplicity of (1 − λ2n )1/2 as a singular value of B is 2 times the
multiplicity of λn as an eigenvalue of A.

Remark 4.14. An elementary observation concerning this property, is that P Q is S-decomposable

if and only if P (1 − Q) also is. And since clearly in this case also QP has the same property, it
follows that also (1 − P )Q and (1 − P )(1 − Q) are S-decomposable (see [3, Prop. 2.7]). Therefore
the following condition can be added to the list above:

15
 4. (see [3, Corol. 2.8]) P − Q is diagonalizable if and only if P + Q is diagonalizable. In that
case, λn is an eigenvalue of P − Q with 0 < |λn | < 1 if and only if 1 ± (1 − λn )2 is an eigenvalue
of P + Q, with the same multiplicity.

Putting these fact together, we have the following:

Theorem 4.15. Let S, T be closed subspaces of H. Then S and T do not have a common
complement if and only if the following conditions hold:

1. dim N (A − 1) 6= dim N (A + 1)

2. A0 = A H0
is compact, where H0 = (N (A − 1) ⊕ N (A + 1))⊥ .

Proof. The proof is just the translation of the two conditions proven by Lauzon and Treil [22],
transcribed here in Remark 2.1.3. The first condition is just the translation of dim S ∩ T ⊥ 6=
dim S ⊥ ∩ T . With respect to the second condition (compacteness of 1 − G∗ G as an operator on
N (G)⊥ ) we use Halmos’ decomposition:

S ∩ T ⊕ S ⊥ ∩ T ⊥ ⊕ S ∩ T ⊥ ⊕ S ⊥ ∩ T ⊕ H0 .

Note that 1 − G∗ G in S coincides with PS − PS PT PS = PS PT⊥ PS in S. Then clearly 1 − G∗ G is

compact in S ⊖ (S ∩ T ⊥ ) if and only PS PT⊥ PS is compact in S ⊥ ⊕ (S ⊖ (S ∩ T ⊥ )) = (S ∩ T ⊥ )⊥ ,
(because PS PT⊥ PS is trivial in S ⊥ ). Note that PS PT⊥ PS in the decomposition

(S ∩ T ⊥ )⊥ = S ∩ T ⊕ S ⊥ ∩ T ⊥ ⊕ S ⊥ ∩ T ⊕ H0 ,

is
PS PT⊥ PS = 0 ⊕ 0 ⊕ 0 ⊕ P0 Q⊥
0 P0 ,

where P0 , Q0 denote the reductions of PS and PT to H0 , respectively. Therefore the second

condition is equivalent to P0 Q⊥ ⊥
0 P0 being compact. This is equivalent to P0 Q0 being compact.
This in turn is equivalent to P0 Q⊥
0 being S-decomposable with singular values sk such that sk
has finite multiplicity and the sequence is finite or else sk → 0. By the above facts, this is
equivalent to
P0 + Q⊥0 = A0 + 1

being diagonalizable, with eigenvalues λk of finite multiplicity, and the relations above (between
eigenvalues of the sum of two projections, namely P0 and Q⊥ 0 , the eigenvalues of the difference
and the singular values of the product) mean that λk have finite multiplicity, are finite or else
λk → 1 (an important observation here is that neither +1 nor −1 are eigenvalues of A0 , because
P0 and Q0 are in generic position). This is clearly equivalent to A0 being diagonalizable with
eigenvalues αk of finite multiplicity, being finitely many, or else αk → 0, i.e., A0 compact.

Remark 4.16. One can further elaborate on the argument of the above theorem. Suppose that
S and T do not have a common complement. Then since S ∩ T ⊥ and S ⊥ ∩ T have different
dimensions, one of the two dimensions must be finite. If dim S ∩ T ⊥ < ∞, then it follows that
(with the argument above) PS PT⊥ PS (or equivalently, PS PT⊥ ) is compact in in the whole space
H. Similarly, if dim S ⊥ ∩ T < ∞, PT PS⊥ is compact.
Summarizing, we have

16
Corollary 4.17. S and T do not have a common complement if and only if

• dim S ∩ T ⊥ 6= dim S ⊥ ∩ T , and

• either PS PT⊥ or PT PS⊥ is compact.

Let us cite the following result from [5] which is relevant to our situation:

Remark 4.18. ([5, Thm. 4.1]) Let P, Q be orthogonal projections. Then P Q is compact if and
only if there exist orthonormal basis {ξn : n ≥ 1} of R(P ) and {ψn : n ≥ 1} of R(Q) such that

hξn , ψk i = 0 if n 6= k

and
hξn , ψn i → 0 (n → ∞).
In the literature (see for instance [17]), in the context of example 6.8 below, these are called
bi-orthogonal bases.

Corollary 4.19. Let S, J be closed subspaces of H. Then there exist no common complements
for S and T if and only if dim S ∩ T ⊥ 6= dim S ⊥ ∩ T and, if the first of these two numbers is
finite there exist orthonormal bases {ξn : n ≥ 1} of S and {ψn : n ≥ 1} of T ⊥ such that

hξn , ψk i = 0 if n 6= k

and
hξn , ψn i → 0 (n → ∞).
If instead dim S ⊥ ∩ T < ∞, then there exist ohonormal bases of S ⊥ and T with the analogous
property.

Proof. The condition PS PT⊥ PS compact is equivalent to PS PT⊥ compact, and thus it the result
in the above remark applies verbatim.

5 Geodesics in ∆ and Γ
We first observe the following facts about geodesics and common complements.

Remark 5.1. 1. The condition that dim N (G) = dim S ∩ T ⊥ equals dim N (G∗ ) = dim S ⊥ ∩ T
(which is stronger than condition (1)) is equivalent to the existence of a geodesic of P(H) joining
PS and PT . Moreover, following the explicit construction of these geodesics (see for instance [2]),
it can be shown that if P (t) is a minimal geodesic joining P (0) = PS with P (1) = PT , then for
any pair t, t′ ∈ [0, 1], R(P (t)) and R(P (t′ )) have a common complement.
2. As the condition for the existence of a geodesic is stronger than the condition for a common
complement, it follows that two projections PS , PT such that (PS , PT ) ∈ ∆ may or may not
be joined by a geodesic. However, in this case one can always join PS and PT by a polygonal
consisting in two minimizing geodesics in P(H). This is a consequence of Giol’s characterization
in Remark 2.2 i). Indeed, the existence of a projection P such that kPS −P k < 1 and kP −PT k < 1
implies that there is a minimizing geodesic joining PS and P , and another minimizing geodesic
joining P and PT .

17
 Since the sets ∆ and Γ are included in P(H) × P(H), we may also ask for the existence
of minimizing geodesics joining points of ∆ or Γ. This leads us to study when geodesics of
P(H) × P(H) remain in ∆ or Γ along their paths. We endow P(H) × P(H) with the product
Finsler metric. The length of a piecewise C 1 curve δ : [0, 1] → P(H) × P(H), δ = (δ1 , δ2 ), is
given by Z 1
L(δ) = (kδ˙1 k2 + kδ˙2 k2 )1/2 dt.
0
Thus, the curve δ is a (minimizing) geodesic in P(H) × P(H) if and only if δ1 and δ2 are
(minimizing) geodesics in P(H).

Remark 5.2. We first discuss the case of projections with range or nullspace of finite dimension.

1. By Remark 3.1 the connected components with finite rank or corank of ∆ are given by
∆ij = Pi,j × Pi,j (i < ∞ or j < ∞). In this case every pair of points in Pi,j × Pk,l can be
joined by a minimizing geodesic contained in ∆.

2. By Remark 4.1 for the case of Γ we known that these connected components are given
by Γijkl = Pi,j × Pk,l , when i 6= k or j 6= l. If i 6= k, then there is a minimizing geodesic
contained in Γ joining two points in Pi,j × Pk,l if and only there is a minimizing geodesic
joining them in P(H) × P(H). Indeed, given (P0 , Q0 ), (P1 , Q1 ) ∈ Pi,j × Pk,l , the condition
for the existence of a geodesic joining them does not necessarily hold. But when this
condition is satisfied, any geodesic must be contained in Γ, since corank(Pt ) = corank(P0 )
and corank(Qt ) = corank(Q0 ) for all t. The case with j 6= l follows again by using the map
⊥.

Now we turn to the case of projections in P∞,∞ × P∞,∞ . In contrast with the previous
situation, the following example illustrates that, in general, there is no minimizing geodesic
contained in ∆∞ joining two given points. Furthermore, the example exhibits arbitrary close
points in ∆∞ without a minimizing geodesic (contained in ∆∞ ) that joins them.

Example 5.3. We begin by giving a pair of projections (P0 , Q0 ) ∈ Γ. Let L be an infinite

dimensional closed subspace such that L × L = H. Take c an orthogonal projection on L such
that dim R(c) = dim N (c) = ∞. We set
! !
1 0 1 0
P0 = , Q0 = .
0 0 0 c
Since P0 − P0 Q0 P0 = 0, dim R(P0 ) ∩ N (Q0 ) = 0 and dim R(Q0 ) ∩ N (P0 ) = dim R(c) = ∞, we
know that (P0 , Q0 ) ∈ Γ.
Now we construct a geodesic δ(t) = (Pt , Qt ), t ∈ R, such that δ(0) = (P0 , Q0 ). Take an
operator Z ≥ 0 on L and put c⊥ = 1 − c, then we define the following antihermitian operators
! !
0 −Z 0 − π2 c⊥
X= , Y = π ⊥ .
Z 0 2c 0
For t ∈ R, we have the unitary operators
! !
cos(tZ) − sin(tZ) c + cos(t π2 )c⊥ − sin(t π2 )c⊥
etX = , etY = .
sin(tZ) cos(tZ) sin(t π2 )c⊥ c + cos(t π2 )c⊥

18
According to the results in [2], since X is a P0 -codiagonal operator and Y is a Q0 -codiagonal
operator, two geodesics of P(H) passing through the point (P0 , Q0 ) are given by
!
tX −tX cos2 (tZ) cos(tZ) sin(tZ)
Pt := e P0 e = ,
cos(tZ) sin(tZ) sin2 (tZ)
!
c + cos 2 ( π t)c⊥ cos( π
t) sin( π
t)c ⊥
Qt := etY Q0 e−tY = 2 2 2 .
cos( π2 t) sin( π2 t)c⊥ c + sin2 ( π2 t)c⊥

In what follows, we further assume t ∈ (− 21 , 12 ), t 6= 0, cZ = Zc and σ(Z) ⊆ (0, π2 ). We observe

the following facts:
1. R(Pt ) ∩ N (Qt ) = {0}. To see this, set λt = sin( π2 t) cos( π2 t)−1 , and note
( ! ! ! !)
π ⊥ π ⊥
f c + cos(t )c − sin(t )c f 0
N (Qt ) = N (Q0 e−tY ) = : 2 2 =
g 0 c g 0
= {(λt g, g) : g ∈ N (c)} .

Since cos(tZ) and sin(tZ) are invertible operators by the previous assumptions on t and Z, we
set Tt = sin(tZ)−1 cos(tZ) to get
( ! ! )
cos(tZ) 0 f
R(Pt ) = R(etX P0 ) = : f, g ∈ L
sin(tZ) 0 g
= {(Tt g, g) : g ∈ L}.

Using that λt > 0 and Tt > 0, we get

R(Pt ) ∩ N (Qt ) = {(Tt g, g) : g ∈ N (c), g ∈ N (Tt − λt )} = {0}.

To prove the last equality, note that Tt g = λt g if and only if

π π
cos(tZ) sin−1 (tZ)g = sin( t) cos−1 ( t)g.
2 2
By the spectral theorem, this means that there is µ ∈ σ(Z) such that 1 = tan(t π2 ) tan(tµ). But
this condition does not hold since t π2 , tµ ∈ (− π4 , π4 ).
2. dim(N (Pt ) ∩ R(Qt )) = dim R(c) = ∞. For we note that

R(Qt ) = N (Qt )⊥ = (R(c) × R(c)) ⊕ {(f, −λt f ) : f ∈ N (c)},

N (Pt ) = R(Pt )⊥ = {(f, −Tt f ) : f ∈ L}.

Recalling that cZ = Zc, we find that

N (Pt ) ∩ R(Qt ) = {(f, −Tt f ) : f ∈ N (c⊥ (Tt + λt ))}

⊇ {(f, −Tt f ) : f ∈ R(c)},

which gives dim(N (Pt ) ∩ R(Qt )) = ∞.

3. Set Gt = Qt |R(Pt ) : R(Pt ) → R(Qt ). Notice that N (Gt )⊥ = (R(Pt ) ∩ N (Qt ))⊥ = R(Pt )
by the first item. We claim that 1 − G∗t Gt |R(Pt ) : R(Pt ) → R(Pt ) is compact if and only if

19
c⊥ sin2 (t( π2 c⊥ − Z)) is compact. For we observe that XY = Y X, so that e−tX etY = et(Y −X) .
Therefore, we have that Pt − G∗t Gt = etX [P0 − P0 et(Y −X) Q0 e−t(Y −X) P0 ]e−tX is compact if and
only if P0 − P0 et(Y −X) Q0 e−t(Y −X) P0 is compact. Set Pt = t( π2 c⊥ − Z), we compute
! ! !
t(Y −X) −t(Y −X) cos(Pt ) − sin(Pt ) 1 0 cos(Pt ) sin(Pt )
P0 e Q0 e P0 = P0 P0
sin(Pt ) cos(Pt ) 0 c − sin(Pt ) cos(Pt )
!
cos2 (Pt ) + c sin2 (Pt ) 0
= .
0 0

Therefore, P0 − P0 et(Y −X) Q0 e−t(Y −X) P0 is compact if and only if c⊥ sin2 (Pt ) is compact, which
proves our claim. Also observe that one can construct examples, where Z satisfies the previous
conditions, such that Xt := c⊥ sin2 (Pt ) is compact or not. For instance, if Z satisfies N (Z)⊥ =
R(Z) ⊆ N (c), then this depends on when π2 c⊥ − Z is compact or not.
From these observations, we obtain that (Pt , Qt ) ∈ ∆ for all t ∈ (− 21 , 21 ), t 6= 0, if and only if
we take Z such that Xt is not compact. Thus, we can choose Z satisfying this condition to get
a family of curves δt (s) = (δ1t (s), δ2t (s)), s ∈ [0, 1], defined by

δ1t (s) := e2stX e−tX P0 etX e−2stX ,

δ2t (s) := e2stY e−tY Q0 etY e−2stY ,

which are geodesics in P(H) × P(H) joining (P−t , Q−t ) ∈ ∆ and (Pt , Qt ) ∈ ∆ such that
δt (s) ∈ ∆, s 6= 21 and δt ( 21 ) = (P0 , Q0 ) ∈ Γ. Furthermore, we know that δt has minimal length
since t ∈ (− 21 , 21 ), kZk ≤ π2 , X is co-diagonal with respect to P0 and Y is co-diagonal with
respect to Q0 . Minimizing geodesics of P(H) are uniquely determined for endpoints lying at
distance less than 1. Hence, the points (P−t, Q−t ), (Pt , Qt ) ∈ ∆ can be made arbitrary close to
each other if one takes t small enough, and they cannot be joined by a minimizing geodesic
contained in ∆.

Properties of the Fredholm index can be used to derive sufficient conditions to obtain min-
imizing geodesics contained in ∆ or Γ that join pairs in P∞,∞ × P∞,∞ . We now recall the
definition and some properties of the restricted Grassmannian; see for instance [12, 29, 30]. The
restricted Grassmannian is given by

Grres = Grres (H+ , H− ) := {P ∈ P(H) : P − P+ is compact}, (5)

for some fixed projections P± associated to a decomposition H+ ⊕ H− = H, where H± are both

infinite dimensional subspaces. Recall from Remark 2.4 that given two projections P, Q such
that P − Q is compact, then Q|R(P ) : R(P ) → R(Q) is a Fredholm operator, with index

index(P, Q) := index(QP |R(P ) : R(P ) → R(Q))

= dim(R(P ) ∩ N (Q)) − dim(R(Q) ∩ N (P )).

We observe that the sets

(k)
Grres := {P ∈ Grres : index(P+ , P ) = k}, k ∈ Z,

describe all the connected components of Grres .

20
Proposition 5.4. The following assertions hold:
(k) (k)
i) For every k ∈ Z, Grres × Grres ⊆ ∆, and there is a minimizing geodesic contained in ∆
(k) (k)
joining any pair of points in Grres × Grres .
(k) (j)
ii) For k 6= j, Grres × Grres ⊆ Γ, and there is a minimizing geodesic contained in Γ joining
(k) (j)
any pair of points in Grres × Grres .

Proof. i) Given three projections P1 , P2 , P3 such that (P1 , P2 ) and (P2 , P3 ) are Fredholm pairs,
and either P1 − P2 is compact or P2 − P3 is compact, then (P1 , P3 ) is a Fredholm pair and
index(P1 , P3 ) = index(P1 , P2 )+index(P2 , P3 ) (see [9, Thm. 3.4]). Thus, if one has two projections
P, Q ∈ Grres satisfying index(P+ , P ) = index(P+ , Q) = k, then index(P, Q) = index(P, P+ ) +
index(P+ , Q) = −index(P+ , P ) + index(P+ , Q) = 0. That is, we obtain that (P, Q) ∈ ∆. Take
(k) (k)
now two pairs (P0 , Q0 ), (P1 , Q1 ) ∈ Grres × Grres . By the same properties of the index, we find
that index(P0 , P1 ) = index(Q0 , Q1 ) = 0. It was shown in [8] that there exists a minimizing
geodesic joining every pair of points in the same connected of the restricted Grassmannian.
Actually this results was proved for the restricted Grassmannian relative to the ideal of Hilbert-
Schmidt operators; but it is not difficult to check that the same holds for the above defined
restricted Grassmannian relative to the ideal of compact operators. Thus, there is a minimizing
geodesic δ(t) = (Pt , Qt ) ∈ ∆, t ∈ [0, 1], joining the points (P0 , Q0 ) and (P1 , Q1 ).

ii) Similarly as in the previous item, we see that index(P, Q) 6= 0, whenever index(P+ , P ) = k and
index(P+ , Q) = j (k 6= j). Also note that P − P QP = P (P − Q)P = P (P − P+ )P + P (P+ − Q)P
is compact since we assume that P, Q ∈ Grres . Hence (P, Q) ∈ Γ. Again the same argument as
in the previous item shows that there is a minimizing geodesic δ(t) = (Pt , Qt ) ∈ Γ, t ∈ [0, 1],
(k) (j)
joining any pairs of points (P0 , Q0 ) and (P1 , Q1 ) with (P0 , Q0 ), (P1 , Q1 ) ∈ Grres × Grres .

6 Examples in function spaces

In previous works [4, 6, 7] we studied the condition for the existence of minimal geodesics joining
two subspaces in certain functional spaces. In this section we discuss if those subspaces and other
related subspaces satisfy the weaker condition of having a common complement.
Let L2 and L∞ be the spaces of square-integrable and essentially bounded complex valued
functions with respect to the Lebesgue measure on the unit circle T = {z ∈ C : |z| = 1}. For
p = 2, ∞, we consider the corresponding Hardy spaces

H p = {f ∈ Lp : fˆ(n) = 0 for n < 0}.

Here {fˆ(n)}n∈Z are the Fourier coefficients of f ∈ L2 . These are closed subspaces of Lp . As usual
we identify H 2 with the space H 2 (D) consisting of all the functions f analytic in the unit disk
D satisfying
 Z 2π 1/2
1 it 2
kf kH 2 (D) := sup |f (re )| dt < ∞.
0≤r<1 2π 0

Given f ∈ H 2 (D), one has that limr→1− f (reit ) = f ∗ (eit ) exists for the L2 norm, and kf ∗ kL2 =
kf kH 2 (D) . Analogously the subspace H ∞ is identified with the space H ∞ (D) consisting of

21
bounded analytic functions on D endowed with the norm kf kH ∞ (D) = supz∈D |f (z)|; though
in this case the limit defining f ∗ has to be taken in the weak* - topology. These classical results
and others mentioned without references in this section can be found for instance in the books
[10, 16, 26].
Our first set of examples concerns subspaces of the form f H 2 , f ∈ L∞ . These are closed
subspaces of L2 if and only if f ∈ (L∞ )× , where (L∞ )× are the invertible functions in the algebra
L∞ (see [6, Lemma 3.1]). In such case, f H 2 is a (single) shift-invariant subspace. Conversely,
the Beurling-Helson theorem gives that a closed (single) shift-invariant subspace must be of the
form f H 2 , for some f ∈ L∞ , |f | = 1 a.e. We consider the Riesz projection
!
fˆ(n)e fˆ(n)eint .
X X
2 2 int
P+ : L → H , P+ =
n∈Z n≥0

We recall that the Toeplitz operator Tf : H 2 → H 2 with symbol f ∈ L∞ is defined by Tf h =

P+ (f h), for h ∈ H 2 .

Example 6.1 (Sarason algebra and restricted Grassmannian of the Hardy space).
Let C denote the space of continuous complex valued functions on T. The Sarason algebra is
defined as
H ∞ + C = {f + g : f ∈ H ∞ , g ∈ C}.

It is the smallest closed subalgebra of L∞ that properly contains H ∞ . We write (H ∞ + C)×

for the invertible functions of the algebra H ∞ + C, which has a characterization in terms of the
harmonic extension. Recall that the harmonic extension to the unit disk of a function f ∈ L∞
is given by
Z 2π
1
fˆ(n)r |n| eint , 0 ≤ r < 1, 0 ≤ t < 2π,
X
it
(hf )(re ) = kr (t − s)f (eis )ds =
2π 0
n∈Z

where kr (t) = (1−r 2 )/(1−2r cos(t)+r 2 ) is the Poisson kernel. Then, one has that f ∈ (H ∞ +C)×
if and only if there exist δ, ǫ > 0 such that |(hf )(reit )| ≥ ǫ for 1 − δ < r < 1. For f ∈ (H ∞ + C)× ,
one can define an index ind(f ) as minus the winding number with respect to the origin of the
curve (hf )(reit ) for 1 − δ < r < 1. This index is stable under small perturbations and it is an
homomorphism of (H ∞ + C)× onto the group of integers.
In this example, we take L2 as the underlying Hilbert space, and subspaces of the form

S = f H 2 , T = gH 2 , f, g ∈ (H ∞ + C)× .

Let Grres the corresponding restricted Grassmannian associated with the Riesz projection P+
(see (5)). According to [6], for f an invertible function in L∞ , we have f ∈ (H ∞ +C)× if and only
(k)
if Pf H 2 ∈ Grres , where k = −ind(f ). Applying Proposition 5.4 we obtain that the subspaces
f H 2 , gH 2 admit a common complement if and only if ind(f ) = ind(g). Notice that 1−G∗ G|N (G)⊥
is always compact since PS − P+ and PT − P+ are compact operators for PS , PT ∈ Grres . In the
case in which ind(f ) 6= ind(g), we observe that (Pf H 2 , PgH 2 ) ∈ Γn1 , where n = ind(g) − ind(g).

22
Remark 6.2. We observe some relevant cases contained in the previous example, and we recall
several ways to compute the index, which can be useful to construct concrete examples.

i) Let (H ∞ )× and C × be the invertible functions in H ∞ and C, respectively. Of course, C × are

the non-vanishing continuous functions on T; meanwhile, for f ∈ H ∞ , one has f ∈ (H ∞ )× if
and only if |(hf )(z)| ≥ ǫ > 0, for all z ∈ D. Since (H ∞ )× ⊂ (H ∞ + C)× and C × ⊂ (H ∞ + C)× ,
we can give as particular examples subspaces of the form f H 2 , where f ∈ (H ∞ )× or f ∈ C × .
At this point, it can be of interest to recall another characterization of the invertible functions
in the Sarason algebra. For f ∈ H ∞ + C, one has that f ∈ (H ∞ + C)× if and only if f = f1 f2 ,
for some functions f1 ∈ (H ∞ )× and f2 ∈ C × (see [26, Section 2.5.3]).

ii) The index of a non-vanishing differentiable function f on T is given by

Z 2π ′ it
1 f (e ) it
ind(f ) = e dt.
2π 0 f (eit )

When f is C 1 and |f (eit )| = 1, 0 ≤ t < 2π, it can be rewritten as

n|fˆ(n)|2 .
X
ind(f ) =
n∈Z

In the case of a rational function f which does not vanish on T, ind(f ) = z − p, where z and p
are the number of zeros and poles of f in D counted with multiplicities.

We now recall the inner-outer factorization. A function f ∈ H 2 is called inner if |f (eit )| = 1

a.e. on T. A function f ∈ H 2 is outer if span{f χn : n ≥ 0} = H 2 , where χn are the functions
on T defined by χn (eit ) = eint (n ∈ Z). Every f ∈ H 2 can be expressed as f = finn fout , where
finn is an inner function and fout is an outer function. This factorization is unique up to a
multiplicative constant.
An inner function f can also be factorized uniquely as f = λbs, where λ ∈ T, b is a Blaschke
product and s is a singular inner function. This is known as the canonical factorization. We
recall that a Blaschke product is an inner function b with zeros {aj }n1 in D (1 ≤ n ≤ ∞) defined
by
n
Y āj aj − z
b(z) = , z ∈ D.
|aj | 1 − āj z
j=1

For aj = 0 we interpret āj /|aj | = −1. In the case where n = ∞, the sequence of zeros {aj }j≥1
P
satisfies the Blaschke condition j≥1 1 − |aj | < ∞, and the product converges uniformly on
compact subsets of D. On the other hand, a singular inner function s = sµ is defined by
 Z 
ψ+z
s(z) = exp − dµ(ψ) , z ∈ D,
T ψ−z

for µ a positive finite measure on T that is singular with respect to the Lebesgue measure. Thus,
for f an inner function, f = λbs, b is the Blaschke product is associated with the zero set of f
and s > 0 on D.

Example 6.3 (Inner functions with distinct supports). The support of an inner function
f denoted by supp(f ) is the set of all points in T that are limits of zeros of f or belong to the

23
support of µ (here s = sµ ). In this example, we consider H = L2 , S = f H 2 and T = gH 2 ,
for f , g inner functions with distinct support. Thus, there is point z0 ∈ supp(f ) \ supp(g) or
z0 ∈ supp(g) \ supp(f ). We may assume that the first of these conditions holds true for our
purpose of determining if f H 2 and gH 2 have a common complement. Then from the proof of
[23, Thm. 1] we know that Tf ḡ is not left invertible in H 2 , which is equivalent to say that Tf ḡ
is not injective or R(Tf ḡ ) is not closed. Below we consider these two cases.
In the first case, Tf ḡ is not injective, so by Coburn’s lemma we know that Tf∗ḡ = Tf¯g is
injective. Hence dim f H 2 ∩(gH 2 )⊥ = dim N (Tf ḡ ) > 0 and dim gH 2 ∩(f H 2 )⊥ = dim N (Tgf¯) = 0.
Then, we have to discuss whether 1 − G∗ G|N (G)⊥ is compact or not, where we take the operator
G as G := PS |T : T → S. Note that PS = Mf P+ Mf¯ and PT = Mg P+ Mḡ . Here Mh is the
multiplication operator by a function h ∈ L∞ acting on L2 . Then

G∗ G = PT PS PT |T = Mg P+ Mf ḡ P+ Mgf¯P+ Mḡ |gH 2 ,

and
N (G) = N (G∗ G) = Mg N (P+ Mf ḡ P+ Mgf¯P+ ).

The unitary operator Mḡ = Mg∗ maps T onto H 2 . Note also that Mḡ maps N (G) onto

N (P+ Mf ḡ P+ Mgf¯P+ ) = N (P+ Mgf¯P+ ).

Then instead of showing that 1 − G∗ G N (G)⊥

is compact, we can show, equivalently, that

1 − P+ Mf ḡ P+ Mgf¯P+ N (P+ Mgf¯P+ )⊥

is compact. Note that P+ Mgf¯P+ on H 2 coincides with the Toeplitz operator Tgf¯. Since we have
assumed that N (Tf¯g ) = {0}, our task now reduces to analyze if the operator

1 − Tf∗¯g Tf¯g (6)

is compact in H 2 . To this end, we need to recall here some facts. The operator Tv Tu −Tvu is called
the semicommutator of the Toeplitz operators Tu and Tv for u, v ∈ L∞ . The Hankel operator
with symbol u ∈ L∞ is the bounded linear operator Hu : H 2 → H− 2 defined by H (f ) = P (uf ),
u −
2 2 2 ⊥
where f ∈ H and P− is the orthogonal projection onto H− := (H ) . The following well-known
formula relates semicommutators and Hankel operators:

Tuv − Tu Tv = Hū∗ Hv

Recall also Hartman’s theorem, which states that Hu is compact if and only u ∈ H ∞ + C.
Applying these results with u = f ḡ, v = f¯g, we find that the operator in (6) is compact if
∗ H
and only if Hfg
¯ f¯g is compact. This is equivalent to say that Hf¯g is compact, or by Hartman’s
theorem, that f¯g ∈ H ∞ + C.
We now show that this is a contradiction with our assumption on the supports of f and g.
Indeed, for z0 ∈ supp(f ) there is a sequence {zn }n≥1 in D such that zn → z0 and (hf )(zn ) → 0
([23, Lemma 2]). On the other hand, if z0 ∈ / supp(g), then |(hg)(zn )| → 1. Notice that the
harmonic extensions hf and hg are the analytic extensions because f , g are inner. Take a ∈

24
H ∞ , b ∈ C and ǫ > 0. Recall that the harmonic extension is asymptotically multiplicative on
H ∞ + C, so there exists a compact set K ⊂ D such that k(hf (a + b))(z) − (hf )(z) (ha)(z) −
(hf )(z) (hb)(z)k < ǫ, for all z ∈ D \ K. Then, we have

kf¯g − (a + b)kL∞ (T) = kg − f (a + b)kL∞ (T)

≥ khg − hf (a + b)kL∞ (D) ≥ khg − hf (a + b)kL∞ (D\K)
≥ khg − (hf )(ha) − (hf )(hb)kL∞ (D\K) −
− khf (a + b) − (hf )(ha) − (hf )(hb)kL∞ (D\K)
≥ 1 − ǫ.

We have used that |(hg)(zn )| → 1, (hf )(zn )(ha)(zn ) → 0 since ha is bounded for a ∈ H ∞
and (hf )(zn )(hb)(zn ) → 0 since hb is continuous on D for b ∈ C. The above arguments imply
d(f¯g, H ∞ +C) = 1, and in particular, we obtain f¯g ∈ / H ∞ +C. Hence the operator 1−G∗ G N (G)⊥
is not compact.
In the second case, we suppose that Tf ḡ is injective and R(Tf ḡ ) is not closed. We change
the definition of G here by setting G := PT |S : S → T . Similar arguments as in the previous
case show that, under the assumption N (Tf ḡ ) = {0}, we have the operator 1 − G∗ G|N (G)⊥ is
compact if and only if 1 − Tḡf ∗ T 2
ḡf is compact in H . But R(Tf ḡ ) is not closed, and consequently,
neither R(|Tf ḡ |2 ) is closed, so that 0 is not an isolated point of its spectrum. Hence 1 is not an
isolated point of the spectrum of 1 − Tḡf ∗ T , which clearly implies that this operator cannot be
ḡf
compact.
From the above two cases, we conclude that f H 2 and gH 2 have a common complement,
whenever f and g are inner functions with distinct support.

Remark 6.4. i) The pair of subspaces considered in Example 6.1 differ from that of Example
6.3. Indeed, if f H 2 and gH 2 are subspaces such that f , g inner functions with supp(f ) 6= supp(g),
then we must have that supp(f ) 6= ∅ or supp(g) 6= ∅. But the unique inner functions invertible
in H ∞ + C are the finite Blaschke products (see [26, Sec. 2.5.3]).

ii) We observe that under the stronger assumption supp(f )\supp(g) 6= ∅ and supp(g)\supp(f ) 6=
∅, it can be shown that N (Tf¯g ) ≃ gH 2 ∩ (f H 2 )⊥ = {0} and N (Tf ḡ ) ≃ f H 2 ∩ (gH 2 )⊥ = {0} (see
[23, Thm. 1-2] and [6, Ex. 5.6]). However, one can give examples of inner functions f , g with
supp(g) ⊂ supp(f ) and N (Tf¯g ) 6= {0}. As a simple example, take h a Blaschke product with
infinite zeros converging to z0 ∈ T, g another Blaschke product with infinite zeros converging to
z1 ∈ T, z0 6= z1 and take f = gh. Clearly, we have N (Tf¯g ) = N (Th̄ ) 6= {0}.

Example 6.5 (Inner functions with same supports). We take H = L2 , S = f H 2 and

gH 2 for f , g inner functions with the same support. Furthermore, we consider only the case
supp(f ) = supp(g) = {1}. This situation is illustrated by the following examples.
For the first type of examples we consider f a singular inner function and g a Blaschke
product. More precisely, let fa (z) = exp(a(z + 1)/(z − 1)), a > 0. A Koosis function is a Blaschke
product g such that fa ḡ ∈ H ∞ + C for all a > 0. A Koosis function is characterized in terms
of the sequence of zeros {ak }k≥1 of the Blaschke product g. Indeed, let λk = i(1 + ak )/(1 − ak ),

25
k ≥ 1, which defines a sequence in the upper-half plane. Then, g is a Koosis function if and only
if Im(λk ) → ∞ and
X Im(λk )
lim = 0, x ∈ R.
|x|→∞ |x − λk |2
k

In particular, g is a Koosis function if it has real zeros converging to 1. It was proved that
N (Tf¯a g ) ≃ gH 2 ∩ (fa H 2 )⊥ is infinite dimensional and R(Tf¯a g ) = L2 for every Koosis function g
and a > 0 (see [23, Thm. 4]). Thus, we have dim gH 2 ∩(fa H 2 )⊥ = ∞ and dim fa H 2 ∩(gH 2 )⊥ = 0.
We can repeat the argument of Example 6.3, where now N (Tf¯a g ) = {0} and G = PgH 2 |fa H 2 :
fa H 2 → gH 2 , to find that the operator 1 − G∗ G|N (G)⊥ must be compact because fa ḡ ∈ H ∞ + C
by definition of a Koosis function. Hence fa H 2 and gH 2 do not admit a common complement,
and (Pfa H 2 , PgH 2 ) ∈ Γr∞ .
The second type of examples concerns two infinite Blaschke products converging to 1. It was
mentioned in [21] that there exist infinite Blaschke products f , g such that f /g ∈ (H ∞ +C)× (i.e.
f , g are codivisible in H ∞ +C). This can be constructed by taking two infinite Blaschke products
with zeros converging to one point in T, where one of the zeros is a suitable perturbation of the
other. Using this idea for studying examples of geodesics in the Grassmann manifold, we proved
in [7, Thm. 5.5] that for every integer n ≥ 0 there exist two disjoint sequences {ak }k≥1 and
{bk }k≥1 in D satisfying Blaschke condition, ak → 1, bk → 1, and their corresponding Blaschke
products f and g are such that dim(f H 2 )⊥ ∩gH 2 = 0 and dim f H 2 ∩(gH 2 )⊥ = n. Furthermore, f
and g are codivisible in H ∞ +C, so again by the previous arguments we have that 1−G∗ G|N (G)⊥
is compact. Hence there is a common complement for f H 2 and gH 2 if and only n = 0. In the
case n 6= 0, we observe that (Pf H 2 , PgH 2 ) ∈ Γn1 .

Example 6.6 (Functions supported on arcs and odd functions). Let again H = L2 =
L2 (T), let J ⊂ T be the arc J = {eit : t ∈ [0, T ]}, and consider the subspace

S = SJ = {f ∈ L2 : sup(f ) ⊂ J }.

Here sup(f ) denotes the (essential) support of f ∈ L2 . We also consider the subspace

T = Todd = {g ∈ L2 (T) : ĝ(m) = 0 for m even}.

Notice that S and T are both infinite dimensional subspaces. The following facts follow from
straightforward computations.
⊥ = {0} =
1. If T = π, then SJ and Todd are in generic position, in particular, SJ ∩ Todd
SJ⊥ ∩ Todd , and therefore SJ and Todd have a common complement.
⊥ = {0}. Set J = {eit : −π ≤ t ≤ −T or T ≤ t ≤ π}, then
2. If T < π, then SJ ∩ Todd T

dim SJ⊥ ∩ Todd = dim{f ∈ L2 : fˆ(m) = 0 for m even, sup(f ) ⊂ JT } = ∞.

Nevertheless in the second case above SJ and Todd have a common complement. Indeed, note
that
1 − PSJ PTodd PSJ S = PSJ − PSJ PTodd PSJ L2 (J ) = PSJ (1 − PTodd ) L2 (J ) .
J

26
Next observe that (1 − PTodd )f (eit ) = 12 (f (eit ) + f (−eit )). If h ∈ L2 is supported in J , then
h(−eit ) = 0 a.e. for t ∈ [0, T ], and one has that
1
(1 − PSJ PTodd PSJ )h = h,
2
which clearly implies that 1 − PSJ PTodd PSJ SJ
is non compact.

The Hardy space H 2 is a reproducing kernel Hilbert space, where the reproducing kernels
are given by the Szego kernels
1
kb (z) = , z, b ∈ D.
1 − b̄z
For our next example we recall that that a sequence {bj }j≥1 in D is said to be an interpolating
sequence if there exists δ > 0 such that
Y bi − bj
δ< , i = 1, 2, . . . . (7)
j≥1
1 − b̄j bi
j6=i

Let ΠB denote the Blaschke product associated to a sequence B = {bj }j≥1 satisfying Blaschke
condition. Carleson proved the following result: the normalized Szego kernels form a Riesz basis
k
{ kkbbi k }i≥1 of the model space KΠB = H 2 ⊖ ΠB H 2 if and only B is an interpolating sequence.
i

Example 6.7 (Blaschke products and odd functions). Let now H = H 2 , and B = {bj }j≥1
satisfying the Blaschke condition. Consider the subspaces T = Todd of odd functions in H 2 , and

S = SB = {f ∈ H 2 : f (bj ) = 0 for j ≥ 1}.

We give two type of examples in this setting.

In the first case, suppose additionally that B ∩ (−B) = ∅ and that the (disjoint) union
B ∪ (−B) is an interpolating sequence. In what follows we use that kkb k = (1 − |b|2 )−1/2 .
⊥ ∩ T 2 1/2 k
P
Suppose that h ∈ SB odd = KΠB ∩ Todd , then h = j≥1 γj (1 − |bj | ) bj is odd. Since
kb (−z) = k−b (z), we get
X X
γj (1 − |bj |2 )1/2 kbj (z) = h(z) = −h(−z) = − γj (1 − |bj |2 )1/2 k−bj (z),
j≥1 j≥1

Using that B ∩ (−B) = ∅, note

(
X δk = γj for ck = bj ,
0= δk (1 − |ck |2 )1/2 kck , where
k≥1
δk = γi for ck = −bi .

⊥ ∩T
Since {(1 − |ck |2 )1/2 kck }k≥1 form a Riesz basis, it follows that ck = 0 for all k. Hence SB odd =
{0}.
Next observe that Todd ⊥ =T 2
even consists of even function in H . Apparently, if f ∈ Teven and
f (b) = 0, then also f (−b) = 0. Therefore if f ∈ SB ∩ Todd ⊥ , the products

bj − z bj + z b2j − z 2
=
1 − b̄j z 1 + b̄j z 1 − b̄2j z 2

27
divide f . Note that these products are even functions. It follows that

⊥
SB ∩ Todd = {g ΠB∪−B : g ∈ H 2 is even},

which is an infinite-dimensional subspace. Now take G := PSB PTodd : Todd → SB , which has
trivial nullspace. Note also that 1 − G∗ G = 1 − PTodd PSB PTodd T = PTodd PS⊥B PTodd T . Recall
odd odd
that if u is an inner function, then the projection Pu : L2 → Ku ⊂ L2 is given by

Pu f = f − uP+ (ūf ),

where P+ is the Riesz projection (see [19, Prop. 5.14]).

We claim that PTodd PS⊥B PTodd T is non compact, which would imply that SB and Todd have
odd
a common complement. Clearly PTodd PS⊥B PTodd T is compact if and only if PTodd PS⊥B is compact
odd
in H 2 . This operator is compact if and only if its natural extension to L2 , namely Podd PΠB is
compact. Here Podd f (z) = 21 (f (z) − f (−z)) is the orthogonal projection onto odd elements in
L2 . Indeed, (
PTodd PS⊥B in H 2 ,
Podd PΠB =
0 in (H 2 )⊥ .
Clearly Podd PΠB is compact if and only if Podd PΠB MΠB is compact, where MΠB denotes the
(unitary) multiplication operator by the unimodular element ΠB . Note that

Podd PΠB MΠB f = Podd (ΠB f − ΠB P + f ) = Podd MΠB P− f,

where P− is the orthogonal projection onto (H 2 )⊥ . This operator is clearly non compact: consider
for instance gn = Π̄B z 2n−1 for integers n ≤ 0. These elements are orthonormal and belong to
(H 2 )⊥ , and clearly
Podd MΠB P− gn = z 2n−1 .

In the second case, we consider a variation of the above example. Put B0 an interpolating
sequence which is symmetric with respect to the origin, i.e. B0 = B∪−B for some B = {bj }j≥1 ⊆
D. Then, we have again

⊥
SB0 ∩ Todd = {g ΠB0 : g ∈ H 2 is even}.

⊥ ∩ T 2 1/2 k
P
But now SB 0 odd consists of odd functions of the form h = j≥1 γk (1 − |bk | ) bk +
P 2 1/2 k bj k −bj
j≥1 δj (1 − |bj | ) k−bj . Since { kbj k }j≥1 ∪ { kk−b k }j≥1 forms a Riesz basis of KB0 , this means
j
that
γk = −δk , k ≥ 1.

This clearly defines an infinite dimensional subspace of H 2 . Therefore, here SB0 ∩ Todd
⊥ and

⊥ ∩T
SB 0 odd are infinite dimensional, and hence SB0 and Todd have a common complement.

The following example concerns well-known facts on the Uncertainty Principle in Harmonic
Analysis (see for instance [24] or the survey article [17]). We also refer to [4] for the relation
with the geometry of the Grassmann manifold.

28
Example 6.8 (Measurable sets and Fourier-Plancharel transform). Let I, J ⊂ Rn be
measurable sets with finite and positive Lebesgue measure. Consider H = L2 (Rn ) with Lebesgue
measure and the projections PI onto the elements of L2 (Rn ) supported in I and QJ onto the
elements whose Fourier-Plancherel transform is supported in J. The following facts are known:

• R(PI ) ∩ R(QJ ) = R(PI ) ∩ N (QJ ) = N (PI ) ∩ R(QJ ) = {0} and N (PI ) ∩ N (QJ ) is infinite
dimensional ([24]).

• PI QJ PI is compact, in fact, nuclear ([17]).

Therefore we have the following:

1. SI = {f ∈ L2 (Rn ) : sup(f ) ⊂ I} and TJ = {g ∈ L2 (Rn ) : sup(ĝ) ⊂ J} have a common

complement since it is straightforward to check that SI ∩ TJ⊥ = {0} = SI⊥ ∩ TJ .

2. SI and TJ⊥ do not have a common complement. The role of T is reversed: now SI ∩(TJ⊥ )⊥ =
R(PI ) ∩ R(QJ ) = {0}, but SI⊥ ∩ TJ⊥ is infinite dimensional. Moreover

1 − PSI TJ⊥
= PI − PI Q⊥
J PI = PI QJ PI

is compact. Hence (PSI , PT ⊥ ) ∈ Γl∞ .

J

Acknowledgment. This work was supported by the grant PICT 2019 04060 (FONCyT - AN-
PCyT, Argentina), (PIP 2021/2023 11220200103209CO), ANPCyT (2015 1505/ 2017 0883) and
FCE-UNLP (11X974).

References
[1] Abbati, M.C.; Manià , A., A geometrical setting for geometric phases on complex Grass-
mann manifolds, J. Geom. Phys. 57 (2007), 777–797.

[2] E. Andruchow. Operators which are the difference of two projections. J. Math. Anal. Appl.
420 (2014), no. 2, 1634–1653.

[3] E. Andruchow; G. Corach. Schmidt decomposable products of projections. Integral Equa-

tions Operator Theory 89 (2017), no. 4, 557–580.

[4] E. Andruchow; G. Corach. Uncertainty principle and geometry of the infinite Grassmann
manifold. Studia Math. 248 (2019), 31–44.

[5] E. Andruchow; G. Corach. Essentially orthogonal subspaces. J. Operator Theory 79 (2018),

no. 1, 79–100.

[6] E. Andruchow; E. Chiumiento, G. Larotonda. Geometric significance of Toeplitz kernels.

J. Funct. Anal. 275 (2018), no. 2, 329–355.

[7] E. Andruchow; E. Chiumiento; A. Varela. Grassmann geometry of zero sets in reproducing

kernel Hilbert spaces, J. Math. Anal. Appl. 500 (2021), no. 1, 125107.

29
 [8] E. Andruchow; G. Larotonda. Hopf-Rinow theorem in the Sato Grassmannian. J. Funct.
Anal. 255 (2008), no. 7, 1692–1712.

[9] J. Avron; R. Seiler; B. Simon. The index of a pair of projections. J. Funct. Anal. 120
(1994), no. 1, 220–237.

[10] A. Böttcher, B. Silbermann, Analysis of Toeplitz Operators, Springer-Verlag, Berlin, 1990.

[11] E. Buckholtz. Hilbert space idempotents and involutions. Proc. Amer. Math. Soc. 128
(2000), no. 5, 1415–1418.

[12] A.L. Carey. Smooth homogeneous spaces and representations of the Hilbert Lie group
U (H)2 . Rev. Roumaine Math. Pures Appl. 30 (1985), no. 7, 505–520.

[13] G. Corach; H. Porta; L. Recht. The geometry of spaces of projections in C ∗ -algebras. Adv.
Math. 101 (1993), no. 1, 59–77.

[14] C. Davis. Separation of two linear subspaces. Acta Sci. Math. (Szeged) 19 (1958), 172–187.

[15] J. Dixmier, Position relative de deux variétés linéaires fermées dans un espace de Hilbert.
(French) Revue Sci. 86 (1948), 387–399.

[16] R.G. Douglas. Banach algebra techniques in operator theory. Second edition. Graduate
Texts in Mathematics, 179. Springer-Verlag, New York, 1998. xvi+194 pp. ISBN: 0-387-
98377-5

[17] G.B. Folland; A. Sitaram. The uncertainty principle: a mathematical survey. J. Fourier
Anal. Appl. 3 (1997), no. 3, 207–238.

[18] P.R. Halmos. Two subspaces. Trans. Amer. Math. Soc. 144 (1969), 381–389.

[19] S.R. Garcia; J. Mashreghi; W.T. Ross. Introduction to model spaces and their operators.
Cambridge Studies in Advanced Mathematics, 148. Cambridge University Press, Cam-
bridge, 2016.

[20] J. Giol. Segments of bounded linear idempotents on a Hilbert space, J. Funct. Anal. 229
(2005), 405–423.

[21] C. Guillory; D. Sarason, Division in H ∞ + C, Mich. Math. J. 28(2) (1981), 173–181.

[22] M. Lauzon; S. Treil, Common complements of two subspaces of a Hilbert space. J. Funct.
Anal. 212 (2004), no. 2, 500–512.

[23] M. Lee; D. Sarason, The spectra of some Toeplitz operators, J. Math. Anal. Appl. 33
(1971) 529–543.

[24] A. Lenard. The numerical range of a pair of projections. J. Functional Analysis 10 (1972),
410–423.

30
[25] N. Makarov, A. Poltoratski, Beurling–Malliavin theory for Toeplitz kernels, Invent. Math.
180 (2010) 443–480.

[26] N.K. Nikolski, Operators, Functions And Systems: An Easy Reading. Vol.1. Hardy, Hankel,
and Toeplitz, Mathematical Surveys and Monographs, vol.92, American Mathematical
Society, Providence, RI, 2002. Translated from the French by Andreas Hartmann.

[27] A. Poltoratski. Toeplitz order. J. Funct. anal. 257 (2018), no. 3, 660–697.

[28] H. Porta; L. Recht. Minimality of geodesics in Grassmann manifolds. Proc. Amer. Math.
Soc. 100 (1987), no. 3, 464–466.

[29] A. Pressley, G. Segal. Loop Groups. Oxford Math. Monogr., Clarendon/Oxford Univ.
Press, Oxford, 1990.

[30] G. Segal, G. Wilson. Loop groups and equations of KdV type. Inst. Hautes Études Sci.
Publ. Math. No. 61 (1985), 5–65.

E. Andruchow, Instituto Argentino de Matemática, ‘Alberto P. Calderón’, CON-

ICET, Saavedra 15 3er. piso, (1083) Buenos Aires, Argentina and Universidad
Nacional de General Sarmiento, J.M. Gutierrez 1150, (1613) Los Polvorines, Ar-
gentina
e-mail: [email protected]

E. Chiumiento, Instituto Argentino de Matemática, ‘Alberto P. Calderón’, CON-

ICET, Saavedra 15 3er. piso, (1083) Buenos Aires, Argentina and Departamento
de Matemáica y Centro de Matemática La Plata, Universidad Nacional de La
Plata, Calle 50 y 115 (1900) La Plata, Argentina
e-mail: [email protected]

31