EVOLUTION SEMIGROUPS, TRANSLATION ALGEBRAS, AND EXPONENTIAL DICHOTOMY OF COCYCLES YURI LATUSHKIN AND ROLAND SCHNAUBELT Abstract. We study the exponential dichotomy of an exponentially bounded, strongly continuous cocycle over a continuous flow on a locally compact metric space Θ acting on a Banach space X. Our main tool is the associated evolution semigroup on C 0 (Θ; X). We prove that the cocycle has exponential dichotomy if and only if the evolution semi- group is hyperbolic if and only if the imaginary axis is contained in the resolvent set of the generator of the evolution semigroup. To show the latter equivalence, we establish the spectral mapping/annular hull theorem for the evolution semigroup. In addition, dichotomy is characterized in terms of the hyperbolicity of a family of weighted shift operators defined on c 0 (Z; X). Here we develop Banach algebra techniques and study weighted translation algebras that contain the evolution operators. These results imply that dichotomy persists under small perturbations of the cocycle and of the underly- ing compact metric space. Also, exponential dichotomy follows from pointwise discrete dichotomies with uniform constants. Finally, we extend to our situation the classical Perron theorem which says that dichotomy is equivalent to the existence and uniqueness of bounded, continuous, mild solutions to the inhomogeneous equation. 1. Introduction Exponential dichotomy of cocycles and nonautonomous Cauchy problems is a classi- cal and well-studied subject, see [13, 14, 17, 18, 30, 35, 42, 43] and the literature cited therein. Recently, important contributions were made in the infinite dimensional case, see [10, 11, 12, 28, 44], and in the applications, [46, 47]. In this paper we continue the investigation, begun in [21, 25], of the dichotomy of cocycles by means of the so-called evolution semigroup. Let {ϕ t } t∈R be a continuous flow on a locally compact metric space Θ and let {Φ t } t∈R + be an exponentially bounded, strongly continuous cocycle over {ϕ t } with values in the set L(X ) of bounded linear operators on a Banach space X . Cocycles arise, for instance, as solution operators for variational equations ˙ u(t)= A(ϕ t θ)u(t) with possibly unbounded operators A(θ), see e.g. [21, 42, 44]. On the space F = C 0 (Θ; X ) of continuous functions f :Θ → X vanishing at infinity endowed with the sup-norm, we 1991 Mathematics Subject Classification. 47D06, 34G10, 34C35, 34D10, 47A10. Key words and phrases. Dichotomy, hyperbolicity, cocycle, evolution semigroup, spectral mapping theorem, translation algebras. YL: Supported by the NSF grant DMS-9622105 and by the Research Board of the University of Missouri. RS: Support by the Deutsche Forschungsgemeinschaft is gratefully acknowledged. 1
36
Embed
EVOLUTION SEMIGROUPS, TRANSLATION ALGEBRAS, AND ...schnaubelt/media/aht.pdf · EVOLUTION SEMIGROUPS, TRANSLATION ALGEBRAS, AND EXPONENTIAL DICHOTOMY OF COCYCLES YURI LATUSHKIN AND
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
EVOLUTION SEMIGROUPS, TRANSLATION ALGEBRAS, ANDEXPONENTIAL DICHOTOMY OF COCYCLES
YURI LATUSHKIN AND ROLAND SCHNAUBELT
Abstract. We study the exponential dichotomy of an exponentially bounded, strongly
continuous cocycle over a continuous flow on a locally compact metric space Θ acting on
a Banach space X. Our main tool is the associated evolution semigroup on C0(Θ;X).
We prove that the cocycle has exponential dichotomy if and only if the evolution semi-
group is hyperbolic if and only if the imaginary axis is contained in the resolvent set of
the generator of the evolution semigroup. To show the latter equivalence, we establish
the spectral mapping/annular hull theorem for the evolution semigroup. In addition,
dichotomy is characterized in terms of the hyperbolicity of a family of weighted shift
operators defined on c0(Z;X). Here we develop Banach algebra techniques and study
weighted translation algebras that contain the evolution operators. These results imply
that dichotomy persists under small perturbations of the cocycle and of the underly-
ing compact metric space. Also, exponential dichotomy follows from pointwise discrete
dichotomies with uniform constants. Finally, we extend to our situation the classical
Perron theorem which says that dichotomy is equivalent to the existence and uniqueness
of bounded, continuous, mild solutions to the inhomogeneous equation.
1. Introduction
Exponential dichotomy of cocycles and nonautonomous Cauchy problems is a classi-
cal and well-studied subject, see [13, 14, 17, 18, 30, 35, 42, 43] and the literature cited
therein. Recently, important contributions were made in the infinite dimensional case,
see [10, 11, 12, 28, 44], and in the applications, [46, 47]. In this paper we continue the
investigation, begun in [21, 25], of the dichotomy of cocycles by means of the so-called
evolution semigroup.
Let ϕtt∈R be a continuous flow on a locally compact metric space Θ and let Φtt∈R+
be an exponentially bounded, strongly continuous cocycle over ϕt with values in the
set L(X) of bounded linear operators on a Banach space X. Cocycles arise, for instance,
as solution operators for variational equations
u(t) = A(ϕtθ)u(t)
with possibly unbounded operators A(θ), see e.g. [21, 42, 44]. On the space F = C0(Θ;X)
of continuous functions f : Θ→ X vanishing at infinity endowed with the sup-norm, we
1991 Mathematics Subject Classification. 47D06, 34G10, 34C35, 34D10, 47A10.Key words and phrases. Dichotomy, hyperbolicity, cocycle, evolution semigroup, spectral mapping
theorem, translation algebras.YL: Supported by the NSF grant DMS-9622105 and by the Research Board of the University of
Missouri.
RS: Support by the Deutsche Forschungsgemeinschaft is gratefully acknowledged.1
define the evolution semigroup Ett∈R+ by
(Etf)(θ) = Φt(ϕ−tθ)f(ϕ−tθ) for θ ∈ Θ, t ≥ 0, f ∈ F . (1.1)
The semigroup Et is strongly continuous on C0(Θ;X). Its generator is denoted by Γ.
We briefly review the history of the subject before we outline the results and methods
presented in the current paper. Starting with the pioneering paper [31] by J. Mather,
spectral properties of evolution semigroups have been used to characterize exponential
dichotomy of the underlying cocycle. In fact, for the finite dimensional case dimX <∞and compact Θ, the following Dichotomy Theorem is contained in [31]:
Φt has exponential dichotomy on Θ if and only if Et is hyperbolic;
that is, σ(E1) ∩ T = ∅ for the spectrum σ(E1) and the unit circle T. Moreover, in the
finite dimensional, compact setting it was proved in [7, 8, 9, 19] that the semigroup Etsatisfies the spectral mapping theorem
σ(Et)\0 = exp(tσ(Γ)) for t ≥ 0 (1.2)
provided that the flow ϕt is aperiodic (i.e., Intθ ∈ Θ : ϕT θ = θ for some T > 0 = ∅),and the annular hull theorem
(σ(Et)\0) · T = exp(tσ(Γ)) · T for t ≥ 0 (1.3)
with no restrictions on the flow ϕt. We point out that the spectral mapping theorem
does not hold without the assumption of the aperiodicity of the flow, see [9, Ex. 2.3].
As a consequence, the dichotomy of the cocycle is equivalent to the spectral condition
σ(Γ) ∩ iR = ∅. An essential step in the proof of these results was a proposition from [29]
known as the Mane Lemma, see also [7, 8, 9]. It says that 1 belongs to the approximate
point spectrum σap(E1) if and only if there exists a point θ0 ∈ Θ, called the Mane point,
such that sup|Φt(θ0)x0| : t ∈ R is finite for a nonzero vector x0. We stress that,
in general, the Mane Lemma does not hold in the infinite dimensional or/and locally
compact setting due to the noncompactness of the unit sphere in X or/and Θ and to the
fact that Φt is defined only for t ≥ 0.
The operator E1 =: E can be expressed as a weighted translation operator E = aV on
F , where a(θ) = Φ1(ϕ−1θ) and (V f)(θ) = f(ϕ−1θ). An important C∗-algebra technique
for the study of weighted translation operators E on the space L2(Θ;X) was developed
in [1, 2]. In particular, the hyperbolicity of the operator E was related to that of a family
of weighted shift operators πθ(E) acting on a space of X-valued sequences x = (xn)n∈Zby the rule
πθ(E) : (xn)n∈Z 7→ (Φ1(ϕn−1θ)xn−1)n∈Z = diag((a(ϕnθ))n∈Z S (xn)n∈Z (1.4)
for θ ∈ Θ, where S(xn)n∈Z = (xn−1)n∈Z.
For compact Θ and infinite dimensionalX, the evolution semigroup (1.1) was considered
in [25] for eventually norm-continuous in θ cocycles, cf. Section 3.3, and in [40] for strongly
continuous cocycles with invertible values. The proof of the Dichotomy Theorem from
[40] requires just a straightforward modification to work in our setting.
The special case of the translation flow ϕtθ = θ + t on Θ = R is well understood
by now, see [4, 5, 20, 22, 34, 36, 37, 39]. Here, the cocycle is given by Φt(θ) = U(θ +
t, θ) for a strongly continuous evolution family U(t, s)t≥s on a Banach space X which2
can be thought as solution operator for the evolution equation u(t) = A(t)u(t) with
unbounded operators A(t). We refer to [45] for applications of the spectral theory of
evolution semigroups to parabolic evolution equations.
In [21] and [24] the general locally compact and infinite dimensional setting was inves-
tigated under the additional and restrictive assumption that the underlying flow ϕt is
aperiodic. In particular, the spectral mapping theorem (1.2) was established in [21] for
F = C0(Θ;X) and in [24] for F = Lp(Θ;X). Also, the following Discrete Dichotomy
Theorem was proved in [21]: The semigroup Et is hyperbolic on C0(Θ;X) if and only
if σ(πθ(E))∩T = ∅ for each θ ∈ Θ on the space c0(Z;X) of X-valued sequences vanishing
at infinity and
supz∈T,θ∈Θ
‖[z − πθ(E)]−1‖L(c0(Z;X)) <∞.
Via the Dichotomy Theorem, the Discrete Dichotomy Theorem relates “global” dichotomy
of Φt on Θ and “pointwise” dichotomies at θ ∈ Θ, cf. [10, 18].
The proof of the Discrete Dichotomy Theorem in [21] required a further development of
the algebraic techniques from [2, 20, 22, 25], see also [16]. Namely, the weighted translation
operator E was “immersed” in the algebra B of operators b ∈ L(C0(Θ;X)) of the form
b =∞∑
k=−∞
akVk with ‖b‖1 :=
∞∑k=−∞
‖ak‖L(C0(Θ;X)) <∞,
where ak are the multiplication operators induced by strongly continuous, bounded func-
tions ak : Θ → L(X). The map πθ : E 7→ πθ(E) was extended to a representation of B
in L(c0(Z;X)). We stress that without the aperiodicity assumption on ϕt the algebra
B cannot even be defined. Thus, a generalization to arbitrary flows requires a radical
modification of the method.
In the present paper, we consider evolution semigroups Et and cocycles Φt without
any additional assumptions on ϕt. We recall necessary background material and prove
the Dichotomy Theorem for strongly continuous cocycles in Section 2. The persistence of
dichotomy under small perturbations of the cocycle follows immediately from this result,
cf. [10].
A modification of the Banach algebra techniques mentioned above is developed in Sec-
tion 3. Here, we replace the algebra B by a certain convolution algebra B. This allows
to prove the Discrete Dichotomy Theorem. In addition, for θ ∈ Θ, we introduce a family
of strongly continuous semigroups Πtθt∈R+ on C0(R;X) by setting
(Πtθh)(s) = Φt(ϕs−tθ)h(s− t) for s ∈ R, t ≥ 0, h ∈ C0(R;X).
Each of the semigroups Πtθ is an evolution semigroup in the sense of [20, 37]. If Φt(θ) is
the solution operator for the variational equation u(t) = A(ϕtθ)u(t), then the generator
Γθ of Πtθ is the closure of the operator h 7→
[− dds
+ A(ϕ·θ)]h, see [21]. We give an
analogue of the Discrete Dichotomy Theorem for Πtθ. As an application, we derive
for eventually norm-continuous in θ cocycles an infinite dimensional generalization of the
Sacker-Sell Perturbation Theorem [43, Thm. 6] on the semicontinuity of the Sacker-Sell
dynamical spectrum as a function of the underlying compact set Θ.
In Section 4, we prove the spectral mapping theorem for aperiodic flows and the an-
nular hull theorem provided that the function of prime periods is strictly separated from3
zero. (In fact, we assume somewhat weaker properties of the flow.) Note that the last
assumption is needed for the Annular Hull Theorem 1.3 if dimX =∞. Counterexamples
are furnished by the identical flow on a singleton and the cocycle given by a semigroup
violating the annular hull theorem, see [33, 34] for such semigroups. We stress that the
annular hull theorem does not hold even for the semigroup related to a first-order pertur-
bation of the two-dimensional wave equation, see [41]. In the proof of the annular hull and
spectral mapping theorems for the evolution semigroups we give an explicit construction
of approximate eigenfunctions for Γ supported in a tube along certain trajectories of the
flow. This localization method can be traced as far as to [31]. The idea of our construc-
tion goes back to the finite-dimensional case [7, 8, 9], but uses essentially different infinite
dimensional methods of [20, 37]. The choice of the trajectories is related to the existence
of Mane sequences θm ⊂ Θ which we define in this paper as an infinite dimensional
generalization of the Mane points mentioned above. It is easy to see that σap(E)∩T 6= ∅implies the existence of a Mane sequence. As a generalization of the Mane Lemma, we
show that the existence of a Mane sequence implies σap(E) ∩ T 6= ∅ under an additional
technical condition related to the residual spectrum.
Finally, we characterize in Section 5 the dichotomy of Φt by the existence and unique-
ness of bounded, continuous solutions to the mild inhomogeneous equation
u(ϕtθ) = Φt(θ)u(θ) +
∫ t
0
Φt−τ (ϕτθ)g(ϕtθ)dθ, t ≥ 0, θ ∈ Θ,
in C0(Θ;X) or Cb(Θ;X). In C0(Θ;X), this equation is just a reformulation of Γu = −g.
Our result gives an affirmative answer to a question remained unsolved in [23] and can
be regarded as a far-reaching generalization of a classical theorem by O. Perron, see
[13, 14, 23, 27, 32] for further comments and related results.
Acknowledgment. We thank A. Stepin for the suggestion to use ε-pseudo orbits which
was crucial in the proof of the perturbation theorem in Subsection 3.3, see [26].
Notation. We denote: T – unit circle in C; D – open unit disc in C; X – Banach space
with norm | · | ; L(X) – bounded linear operators on X; Ls(X) – bounded linear operators
on X with the strong topology; I – the identity operator; ρ(A), σ(A), σap(A), σr(A) –
resolvent set, spectrum, approximate point spectrum, residual spectrum of A.
2. The Dichotomy Theorem
We briefly introduce exponential dichotomy of cocycles (see the classical sources [17, 18,
42, 43] and [10, 44]) and evolution semigroups (see [21, 25, 40]), and prove the equivalence
of the exponential dichotomy of a cocycle and the hyperbolicity of the corresponding
evolution semigroup.
2.1. Exponential dichotomy. Let ϕtt∈R be a continuous flow on a locally compact
metric space Θ; that is, the map R × Θ 3 (t, θ) 7→ ϕtθ ∈ Θ is continuous and ϕt+τθ =
ϕt(ϕτθ), ϕ0θ = θ for all t, τ ∈ R and θ ∈ Θ. Let Φtt≥0 be an exponentially bounded,
strongly continuous cocycle over ϕtt∈R. This means that the map R+ × Θ 3 (t, θ) 7→Φt(θ) ∈ L(X) is strongly continuous, ‖Φt(θ)‖ ≤ Neωt for t ≥ 0, θ ∈ Θ, and some constants
N and ω, and the following cocycle property holds:
Φt+τ (θ) = Φt(ϕτθ)Φτ (θ) and Φ0(θ) = I for t, τ ≥ 0 and θ ∈ Θ. (2.1)4
We will also consider cocycles for discrete time t ∈ N which are defined analogously.
Cocycles Φtt≥0 are in one-to-one correspondence with continuous linear skew-product
flows (LSPF) ϕtt≥0 defined by
ϕt : Θ×X → Θ×X : (θ, x) 7→ (ϕtθ,Φt(θ)x) for t ≥ 0.
Therefore, all assertions below concerning cocycles can be reformulated in terms of the
corresponding linear skew-product flows.
Definition 2.1. The cocycle Φtt≥0 has exponential dichotomy on Θ if there exists a
bounded, strongly continuous, projection valued function P : Θ→ Ls(X) satisfying
(a) P (ϕtθ)Φt(θ) = Φt(θ)P (θ),
(b) ΦtQ(θ) is invertible as an operator from ImQ(θ) to ImQ(ϕtθ) (with inverse
Φ−tQ (ϕtθ)),
(c) there exist constants β > 0 and M = M(β) ≥ 0 such that
‖ΦtP (θ)‖ ≤Me−βt and ‖[Φt
Q(θ)]−1‖ ≤Me−βt
for all θ ∈ Θ and t ≥ 0. The function P (·) is called dichotomy projection and the
constants β,M are the dichotomy constants.
Here and below, ΦtP (θ) and Φt
Q(θ) denote the restrictions Φt(θ)P (θ) : ImP (θ) →ImP (ϕtθ) and Φt(θ)Q(θ) : ImQ(θ)→ ImQ(ϕtθ), respectively, and we set Q = I − P for
a projection P . Notice that Φt(θ)Q(θ) is strongly continuous on R×Θ and satisfies (2.1)
with I replaced by Q(θ). Further, we define dichotomy along single orbits.
Definition 2.2. The cocycle Φtt∈K+ has exponential dichotomy (over K = R or Z) at
a point θ0 ∈ Θ if there are uniformly bounded projections Pτ ∈ L(X) for τ ∈ K which
depend strongly continuous on τ if K = R and satisfy
(a) Pτ+tΦt(ϕτθ0) = Φt(ϕτθ0)Pτ ,
(b) ΦtQ(ϕτθ0) is invertible from ImQτ to ImQτ+t,
(c) there exist constants β = β(θ0) > 0 and M = M(β) ≥ 0 such that
‖ΦtP (ϕτθ0)‖ ≤Me−βt and ‖[Φt
Q(ϕτθ0)]−1‖ ≤Me−βt
for t, τ ∈ K and t ≥ 0.
Clearly, dichotomy on Θ implies dichotomy at θ0. If ImQ(θ) and ImQ(ϕtθ) are finite
dimensional, then (b) and the second estimate in (c) follow from
|Φt(θ)x| ≥ 1Meβt|x| for x ∈ ImQ(θ), t ≥ 0.
If each operator Φt(θ) is invertible, exponential dichotomy at θ0 is equivalent to the
existence of a projection P such that
‖Φt(θ0)P [Φτ (θ0)]−1‖ ≤Me−β(t−τ), t ≥ τ,
‖Φt(θ0)Q[Φτ (θ0)]−1‖ ≤Me−β(τ−t), t ≤ τ,
cf. [40]. Then one has Pτ = Φτ (θ)P0[Φτ (θ)]−1. This is the classical definition of expo-
nential dichotomy, see [13, 14, 42]. We also recall the analogous concept for a single
operator.
Definition 2.3. A operator E ∈ L(X) is called hyperbolic if σ(E) ∩ T = ∅.5
Observe that the powers Enn∈N or a semigroup Ett≥0 have exponential dichotomy
if and only if E or Et0 for some t0 > 0 are hyperbolic, respectively. In this case, the
dichotomy projection is given by the Riesz projection
P =1
2πi
∫T(λI − E)−1 dλ.
We denote by EtP = EtP and Et
Q = EtQ the restriction to the stable and the unstable sub-
space, ImP and ImQ, respectively; and analogously for the generator of the semigroup,
cf. [33, A-III].
For an exponentially bounded, strongly continuous cocycle Φt and λ ∈ C, we define
the rescaled cocycle Φtλ(θ) = e−λtΦt(θ) and the rescaled linear skew-product flow
ϕtλ : (θ, x) 7→ (ϕtθ, e−λtΦt(θ)x) for θ ∈ Θ, x ∈ X, t ≥ 0.
Note that the cocycle Φtλ corresponds to the equation u(t) = [A(ϕtθ)− λ]u(t).
Definition 2.4. The dynamical (or Sacker-Sell) spectrum Σ is defined by
Σ = λ ∈ R| Φtλt≥0 does not have exponential dichotomy.
2.2. Evolution semigroups. By F = C0(Θ;X) we denote the space of continuous func-
tions f : Θ→ X vanishing at infinity endowed with the sup-norm. For a cocycle Φtt≥0,
we introduce on C0(Θ;X) the evolution semigroup Ett≥0 by setting
(Etf)(θ) = Φt(ϕ−tθ)f(ϕ−tθ) for θ ∈ Θ, t ≥ 0, f ∈ C0(Θ;X).
One checks the following simple fact, cf. [21, Thm. 2.1], [40, Prop. 3].
Proposition 2.5. Ett≥0 is a strongly continuous semigroup on C0(Θ;X) if and only if
Φtt≥0 is an exponentially bounded, strongly continuous cocycle.
The following result is due to R. Rau, [40, Lemma 7], see [2, 20, 21, 25, 36] for earlier
and different versions.
Theorem 2.6 (Spectral Projection Theorem). Let Φtt≥0 be an exponentially bounded,
strongly continuous cocycle over a continuous flow ϕtt∈R on Θ. If the induced evolution
semigroup Et is hyperbolic on C0(Θ;X), then there exists a bounded, strongly continu-
ous, projection valued function P : Θ → Ls(X) such that the Riesz projection P for Et
corresponding to σ(Et) ∩ D is given by the formula (Pf)(θ) = P (θ)f(θ) for θ ∈ Θ and
f ∈ C0(Θ;X).
The idea of the proof is to check, see Lemma 3.2 in [20], that P commutes with the
multiplication operators induced by continuous, bounded functions χ : Θ → R. This
allows to define P (θ0)x = (Pf)(θ0) for θ0 ∈ Θ and x ∈ X, where f ∈ C0(Θ;X) is an
arbitrary function such that f(θ0) = x.
We now come to the first main result; our proof is a modification of [40, Thm. 10,12].
Theorem 2.7 (Dichotomy Theorem). Let Φtt≥0 be an exponentially bounded, strongly
continuous cocycle over a continuous flow ϕtt≥0 on a locally compact metric space Θ,
and let Ett≥0 be the induced evolution semigroup on C0(Θ;X). Then the cocycle has
exponential dichotomy on Θ if and only if the evolution semigroup is hyperbolic. In this
case, the Riesz projection P corresponding to the spectral set σ(E1)∩D and the dichotomy6
projection P : Θ → Ls(X) of Φt are related by (Pf)(θ) = P (θ)f(θ) for θ ∈ Θ and
f ∈ C0(Θ;X).
Proof. (1) Assume that Φt has exponential dichotomy with projections P (θ). Define
P = P (·) on C0(Θ;X). Then PEt = EtP by (a) in Definition 2.1. The inequality
‖EtP‖L(C0(Θ;X)) ≤ Me−βt follows from the first inequality in (c). Also, Et
Q on ImQ has
the inverse (Rtf)(θ) = [ΦtQ(θ)]−1f(ϕtθ). Now, (c) yields ‖[Et
Q]−1‖L(C0(Θ;X)) ≤ Me−βt for
t ≥ 0. This means that Et is hyperbolic.
(2) Assume that Et is hyperbolic with Riesz projection P . The Spectral Projection
Theorem 2.6 shows that (Pf)(θ) = P (θ)f(θ) for some bounded, strongly continuous,
projection valued function P : Θ→ Ls(X). Since P is the Riesz projection for E1, EtQ is
invertible on ImQ and there exist constants β > 0 and M > 0 such that
‖EtP‖L(C0(Θ;X)) ≤Me−βt, t ≥ 0, (2.2)
‖EtQQf‖∞ ≥ 1
Meβt‖Qf‖∞, t ≥ 0. (2.3)
Clearly, (a) in Definition 2.1 holds. Let θ ∈ Θ and x ∈ X. Take a function f ∈ C0(Θ;X)
such that f(θ) = x and ‖f‖∞ = |x|. Using (2.2), we obtain
|ΦtP (θ)x| ≤ ‖Φt
P (·)f(·)‖∞ = ‖EtPf‖∞ ≤Me−βt‖f‖∞ = Me−βt|x|
which gives the first inequality in (c). To prove (b), fix t ≥ 0, θ ∈ Θ, and x ∈ Im Q(θ).
Recall that for any f ∈ Im Q we have ‖EtQf‖∞ = supθ∈Θ |Φt
Q(θ)f(θ)|. For any ε > 0
choose f ∈ Im Q such that f(θ) = x and ‖EtQf‖∞ ≤ |Φt(θ)x| + ε. This function can be
chosen as f = χx where χ is a continuous bump-function with χ(θ) = 1 and a sufficiently
small support. Then (2.3) implies
1Meβt|x| ≤ 1
Meβt‖f‖∞ ≤ ‖Et
Qf‖∞ ≤ |Φt(θ)x|+ ε,
and hence
|ΦtQ(θ)x| ≥ 1
Meβt|x|. (2.4)
In particular, the operator ΦtQ(θ) : ImQ(θ) → ImQ(ϕtθ) is injective. To see that it is
surjective, take y ∈ ImQ(ϕtθ). For a continuous, compactly supported function α : Θ 7→[0, 1] with α(ϕtθ) = 1, set f = α(·)Q(·)y. Then f ∈ ImQ. Since Et
Q is invertible on ImQ,
there exists g ∈ ImQ such that EtQg = f For x := g(θ) ∈ ImQ(θ0), this gives
ΦtQ(θ)x = Φt
Q(θ)g(θ) = f(ϕtθ) = α(ϕtθ)Q(ϕtθ)y = y.
Thus, (b) in Definition 2.1 holds. Together with (2.4) this establishes the second inequality
in (c) of this definition.
Corollary 2.8. The dichotomy projection is uniquely determined by Definition 2.1(a)–
(c).
Theorem 2.7 combined with the argument used in part (1) of its proof shows that the
existence of a dichotomy over R and Z for the cocycle Φt are equivalent:
Corollary 2.9. Given a cocycle Φtt≥0 the following assertions are equivalent:
(1) Φtt≥0 has an exponential dichotomy over R;
(2) the discrete time cocycle Φtt∈N has dichotomy.7
The Dichotomy Theorem 2.7 relates our approach with the Sacker-Sell spectral theory,
[43], for infinite dimensional cocycles or linear skew-product flows. In fact, by rescaling
one obtains
Σ = log |σ(E1) \ 0|,cf. [25, 40]. One can also derive robustness of dichotomy from Theorem 2.7, see also
[10, 12, 13, 14, 18, 21, 22, 30, 45]. We present the following result in this direction.
Corollary 2.10. Assume that the exponentially bounded, strongly continuous cocycle
Φtt≥0 has exponential dichotomy. There exists ε > 0 such that every exponentially
bounded, strongly continuous cocycle Ψtt≥0 with
supθ∈Θ‖Φt(θ)−Ψt(θ)‖L(X) < ε
for some t > 0 also has exponential dichotomy.
Proof. Let Et and F t denote the evolution semigroups induced by Φt and Ψt,respectively. Assume that
supθ∈Θ‖Φt(θ)−Ψt(θ)‖L(X) < inf
λ∈T‖(λ− Et)−1‖−1 =: ε.
Then, using λ− F t = [1− (F t − Et)(λ− Et)−1] (λ− Et) and
‖Et − F t‖L(C0(Θ;X)) = supθ∈Θ‖Φt(θ)−Ψt(θ)‖L(X),
we derive that F t, and hence Ψt, has exponential dichotomy.
3. Global and Pointwise Dichotomies
We now want to characterize the (global) dichotomy of Φt on Θ by the (local) di-
chotomy at each θ0 ∈ Θ with uniform dichotomy constants. To that purpose, we develop
Banach algebra techniques which allow one to describe the dichotomy of Φt (or, equiv-
alently, the hyperbolicity of Et) in terms of the hyperbolicity of the discrete operators
πθ(E) on c0(Z;X) defined in (1.4).
3.1. Weighted translation algebras. For a continuous flow ϕtt∈R on a locally com-
pact metric space Θ, let us define the translation group (V tf)(θ) = f(ϕ−tθ) for t ∈ Ron the space C0(Θ;X). Set at(θ) = Φt(ϕ−tθ) for t ≥ 0, θ ∈ Θ, and an exponentially
bounded, strongly continuous cocycle Φtt≥0 over ϕtt∈R. We abbreviate V = V 1,
E = E1, ϕ = ϕ1, Φ = Φ1.
We denote by A = Cb(Θ;Ls(X)) the algebra of bounded, strongly continuous, operator
valued functions a = a(·) on Θ with the sup-norm ‖a‖A = ‖a‖∞ = supθ∈Θ ‖a(θ)‖L(X) and
point-wise multiplication. Clearly, A can be isometrically identified with the subalgebra A
of L(C0(Θ;X)) which contains the multiplication operators (denoted simply by a) given
by (af)(θ) = a(θ)f(θ) for θ ∈ Θ and f ∈ C0(Θ;X).
Let B denote the set of all sequences (ak)k∈Z with entries ak ∈ A satisfying (‖ak‖A)k∈Z ∈`1(Z). Equipped with the norm
‖(ak)k∈Z‖1 :=∞∑
k=−∞
‖ak‖A =∞∑
k=−∞
supθ∈Θ‖ak(θ)‖L(X) ,
8
B is a Banach space. For b′ = (a′n)n∈Z and b′′ = (a′′n)n∈Z in B, we define b := b′ ∗ b′′ by
b = (an)n∈Z and an(θ) =∑
k∈Z a′k(θ)a
′′n−k(ϕ
−kθ) for θ ∈ Θ and n ∈ Z. Notice that
‖an(θ)‖L(X) ≤ ‖b′‖1 ‖b′′‖1.
The function θ 7→ a′k(θ)a′′n−k(ϕ
−kθ)x is continuous for x ∈ X and k, n ∈ Z. We have
|[an(θ1)− an(θ2)]x| ≤∑|k|≤N
| [a′k(θ1)a′′n−k(ϕ−kθ1)− a′k(θ2)a′′n−k(ϕ
−kθ2)]x |
+ 2 |x| ‖b′′‖1
∑|k|>N
‖a′k‖∞ .
for θ1, θ2 ∈ Θ and N ∈ N. Thus an ∈ A. We further estimate
‖b‖1 ≤∑
n
∑k‖a′k‖∞ ‖a′′n−k ϕ−k‖∞ =
∑k
∑n‖a′k‖∞ ‖a′′n−k‖∞ ≤ ‖b′‖1 ‖b′′‖1.
Therefore, (B, ∗, ‖ · ‖1) is a Banach algebra with the unit element e = e⊗ δn0. Here and
below, e(θ) = IX for θ ∈ Θ and δnk is the Kronecker delta. In other words, e = (en) where
e0 = I and en = 0 for n 6= 0.
To relate the algebra B with L(C0(Θ;X))), we introduce the (algebra) homomorphism
ρ : B→ L(C0(Θ;X))) : b = (ak)k∈Z 7→∞∑
k=−∞
akVk. (3.1)
Clearly, ρ(e) = I and ‖ρ(b)‖L(C0(Θ;X)) ≤ ‖b‖1 . We stress that, in general, ρ is not injective.
As an example, let ϕ1θ = θ for all θ ∈ Θ. Then, V = I and each a ∈ A ⊂ L(C0(θ;X))
can be represented as
a = 12a+ 1
2aV = 1
3a+ 2
3aV etc.
However, ρ is injective provided that ϕ is aperiodic, i.e., the set of aperiodic points for ϕ
is dense in Θ. Assuming, for a moment, the aperiodicity of ϕ, let B be the subspace of
L(C0(Θ;X)) containing the operators of the form
b =∞∑
k=−∞
akVk where ak ∈ A and ‖b‖1 :=
∞∑k=−∞
‖ak‖A <∞. (3.2)
For aperiodic ϕ, it was proved in Prop. 2.14 and 2.15 of [21] that the representation
b ∈ B as in (3.2) is unique and that (B, ‖ · ‖1) is a Banach algebra (with respect to the
composition of operators). Thus, ρ : B→ B is an isomorphism if ϕ is aperiodic.
Going back to a not necessarily aperiodic flow ϕ, let D ⊂ L(c0(Z;X)) be the algebra
of all bounded diagonal operators d = diag(d(n))n∈Z. Define
C := D =∞∑
k=−∞
dkSk ∈ L(c0(Z;X)) : dk ∈ D, ‖D‖C :=
∞∑k=−∞
‖dk‖L(c0(Z;X)) <∞,
where S : (xn)n∈Z 7→ (xn−1)n∈Z is the shift operator on c0(Z;X). The representation of
D ∈ C as D =∑dkS
k is unique since, for m0, n0 ∈ Z and x ∈ X with |x| = 1, one has
‖D‖L(c0(Z;X)) ≥ ‖D(x⊗ δn,n0)‖c0(Z;X) =∥∥∥∑
kdiag(d
(n)k )n∈Z (x⊗ δn−k,n0)n∈Z
∥∥∥c0(Z;X)
≥ |∑
kd
(m0)k x δm0−k,n0| = |d
(m0)m0−n0
x|.9
We further need, for each θ ∈ Θ, the map
πθ :B→ L(c0(Z;X)) : b = (ak)n∈Z 7→∞∑
k=−∞
diag[ak(ϕnθ)]n∈Z S
k. (3.3)
Observe that πθ : B→ C and
‖πθ(b)‖C =∞∑
k=−∞
‖ diag(ak(ϕnθ))n∈Z‖L(c0(Z;X))
=∞∑
k=−∞
supn∈Z
supθ∈Θ‖ak(ϕnθ)‖L(X) =
∞∑k=−∞
‖ak‖∞ = ‖b‖1 .
Also, we compute for b = b′ ∗ b′′
πθ(b) =∞∑
k=−∞
diag
(∞∑
l=−∞
a′l(ϕnθ)a′′k−l(ϕ
n−lθ)
)n∈Z
Sk
=∞∑
l=−∞
[diag(a′l(ϕ
nθ))n∈Z SlS−l
∞∑k=−∞
diag(a′′k−l(ϕn−lθ))n∈Z S
k
]
=∞∑
l=−∞
[diag(a′l(ϕ
nθ))n∈Z Sl
∞∑k=−∞
diag(a′′k−l(ϕnθ))n∈Z S
k−l
]= πθ(b
′) · πθ(b′′).
As a result, πθ is a continuous homomorphism. Its importance relies on the following fact.
Proposition 3.1.⋂θ∈Θ ker πθ = 0.
Proof. Assume that πθ(b) = 0 for all θ ∈ Θ and some b = (an)n∈Z ∈ B. This means that
0 = πθ(b) =∑
kdiag(ak(ϕ
nθ))n∈Z Sk ∈ C
for θ ∈ Θ. Since the representation D =∑
k diag(d(n)k )Sk for D ∈ C is unique, we conclude
that ak(ϕnθ) = 0 for each k, n ∈ Z and θ ∈ Θ. Thus, b = 0 in B.
For a ∈ A, we define the operator b = I − aV ∈ L(C0(Θ;X)) and the sequence
b = (ak)k∈Z ∈ B by a0 = e, a1 = −a, and ak = 0 otherwise. Then ρ(b) = I − aV = b by
(3.1). Also, given b = I−aV , we introduce for each θ ∈ Θ the operator πθ(b) ∈ L(c0(Z;X))
by the rule
πθ(b) = πθ(I − aV ) := I − diag(a(ϕnθ))n∈ZS. (3.4)
To relate these concepts to the evolution semigroup Et on C0(Θ;X) induced by the
cocycle Φt, we remark that the operator I −E1 is of the form b = I −E = I − aV with
a(θ) = Φ(ϕ−1θ) for θ ∈ Θ. We can now give a sufficient condition for the invertibility of
I − E in terms of the operators πθ(b).
Theorem 3.2. Let b = I − aV for a ∈ A. Assume that πθ(b) is invertible in L(c0(Z;X))
for each θ ∈ Θ and that there exists a constant B > 0 such that
‖[πθ(b)]−1‖L(c0(Z;X)) ≤ B for all θ ∈ Θ. (3.5)
Then b−1 exists in L(C0(Θ;X)).10
Proof. Let b = (ak)k∈Z ∈ B be given by a0 = I, a1 = −a, and ak = 0 for k 6= 0, 1. The
definitions (3.3) and (3.4) yield
πθ(b) = πθ(b) for θ ∈ Θ. (3.6)
Also, ρ(b) = b by (3.1). Setting d0(θ) = I, d1(θ) = diag(a(ϕnθ))n∈Z , and dk(θ) = 0 for
k 6= 1, 2 and θ ∈ Θ, we see that
πθ(b) =∑
kdk(θ)S
k ∈ C.
Next, we show that the inverse of πθ(b) also belongs to C. The proof of this fact is close
to that of Lemma 1.6 in [22].
Proposition 3.3. Under the assumptions of Theorem 3.2, the operator [πθ(b)]−1 =
[πθ(b)]−1 belongs to C for each θ ∈ Θ; that is,
[πθ(b)]−1 =
∞∑k=−∞
Ck(θ)Sk where Ck(θ) = diag(C
(n)k (θ))n∈Z satisfies (3.7)
∞∑k=−∞
supθ∈Θ‖Ck(θ)‖L(c0(Z;X)) <∞. (3.8)
Proof of Proposition 3.3. Set Dθ = πθ(b) = I − diag(a(ϕnθ))n∈ZS. Assumption (3.5) and
identity (3.6) yield
supθ∈Θ‖D−1
θ ‖ ≤ B. (3.9)
For γ > 0, we define the operator
Dθ(γ) = I − diag(eγ(|n|−|n−1|)a(ϕnθ))n∈Z S
on c0(Z;X). Note that Dθ(γ) = J−1γ DθJγ for Jγ = diag(e−γ|n|)n∈Z. Clearly,
Then ‖Dθ − Dθ(γ)‖ · ‖D−1θ ‖ ≤ q for all θ ∈ Θ and 0 ≤ γ ≤ γ0. Since Dθ is invertible
in L(c0(Z;X)), we conclude that Dθ(γ) is invertible and ‖[Dθ(γ)]−1‖ ≤ R := B1−q for all
0 ≤ γ ≤ γ0 and θ ∈ Θ.
Consider D−1θ as an operator matrix D−1
θ = [Ckj(θ)]k,j∈Z, where Ckj(θ) ∈ L(X) are
defined for x ∈ X by
Ckj(θ)x := (D−1θ (xj))k for xj = (δjnx)n∈Z.
From Dθ(γ)−1 = J−1γ D−1
θ Jγ we derive Dθ(γ)−1 = [eγ(|k|−|j|)Ckj(θ)]k,j∈Z. Now,
R ≥ ‖Dθ(γ)−1‖L(c0(Z;X)) ≥ eγ(|k|−|j|)‖Ckj(θ)‖L(X)
for k, j ∈ Z yields
‖Ck,0(θ)‖L(X) ≤ Re−γ|k| for θ ∈ Θ and k ∈ Z.
Observe that, for every j, k,m ∈ Z, one has
Dϕmθ = S−mDθSm and Ckj(ϕ
mθ) = Ck+m,j+m(θ). (3.10)11
In particular, from Ck+m,m(θ) = Ck,0(ϕmθ) it follows that
‖Ck+m,m(θ)‖L(X) ≤ Re−γ|k| for k,m ∈ Z and θ ∈ Θ. (3.11)
So we can write D−1θ as
D−1θ =
∑kCk(θ)S
k =∑
kdiag(C
(n)k (θ))n∈ZS
k
for C(n)k (θ) = Cn,n−k(θ), and we estimate∑
k∈Z
supθ∈Θ‖Ck(θ)‖L(c0(Z;X)) =
∑k∈Z
supθ∈Θ
supn∈Z‖Cn,n−k(θ)‖L(X) ≤
∑k∈Z
Re−γ|k| <∞.
To conclude the proof of Theorem 3.2 we need the following fact.
Claim 1. θ 7→ C(0)k (θ), see (3.7), is a continuous, bounded function from Θ to Ls(X)
with ∑k∈Z
supθ∈Θ‖C(0)
k (θ)‖L(X) <∞.
We finish the proof of the theorem and show Claim 1 later. First, notice that (3.10)
implies
C(n)k (θ) = Cn,n−k(θ) = C0,−k(ϕ
nθ) = C(0)k (ϕnθ) (3.12)
for k, n ∈ Z and θ ∈ Θ. The sequence d = (C(0)k )k∈Z belongs to B by Claim 1. Thus, we
can apply to d the homomorphism ρ : B→ L(C0(Θ;X)). Let d = ρ(d). We want to show
that db = bd = I.
Consider the sequence r = d ∗ b − e ∈ B. Fix θ ∈ Θ and apply πθ : B → C. Using
(3.3), (3.12), and Proposition 3.3, we compute
πθ(r) = πθ(d) · πθ(b)− I =[∑
kdiag(C
(0)k (ϕnθ))n∈Z S
k]· πθ(b)− I
=[∑
kdiag(C
(n)k (θ))n∈Z S
k]· πθ(b)− I
=[∑
kCk(θ)S
k]· πθ(b)− I = 0.
Since θ ∈ Θ was arbitrary, Proposition 3.1 shows that r = 0 in B. As a consequence,
0 = ρ(r) = ρ(d ∗ b− e) = ρ(d) · ρ(b)− I = d · b− I.
Similarly, one obtains 0 = bd− I, and the theorem is proved.
Proof of Claim 1. Let k ∈ Z, x ∈ X, and θ0 ∈ Θ. Define x = (xn) ∈ c0(Z;X) by xn = x
if n = −k and xn = 0 if n 6= −k. Then, for every θ ∈ Θ,∣∣∣[C(0)k (θ)− C(0)
k (θ0)]x∣∣∣ =
∣∣∣∣∣∞∑
l=−∞
[C
(0)l (θ)− C(0)
l (θ0)]x−l
∣∣∣∣∣≤ sup
n∈Z
∣∣∣∑l
[C
(n)l (θ)− C(n)
l (θ0)]xn−l
∣∣∣=∥∥∥(∑
ldiag
(C
(n)l (θ)− C(n)
l (θ0))n∈Z
Sl)x∥∥∥c0(Z;X)
=∥∥∥([πθ(b)]
−1 − [πθ0(b)]−1)x∥∥∥c0(Z;X)
≤∥∥∥[πθ(b)]
−1∥∥∥L(c0(Z;X))
·∥∥∥[πθ0(b)− πθ(b)]y
∥∥∥c0(Z;X)
, (3.13)
12
where y := [πθ0(b)]−1x ∈ c0(Z;X). It is easy to see that
Θ 3 θ 7→ πθ(b) ∈ Ls(c0(Z;X))
is (strongly!) continuous. So (3.13) and (3.9) imply the strong continuity of θ 7→ C(0)k (θ) ∈
Ls(X). ¿From (3.12) and (3.8) we derive∑k∈Z
supθ∈Θ‖C(0)
k (θ)‖L(X) =∑k∈Z
supθ∈Θ
supn∈Z‖C(0)
k (ϕnθ)‖L(X) =∑k∈Z
supθ∈Θ‖Ck(θ)‖L(c0(Z;X)) <∞,
and Claim 1 is proved.
Remark 3.4. Assume that ϕ is aperiodic. Then, ρ : B → B is an isomorphism and
πθ = πθ ρ−1. The proof of Theorem 3.2 shows that in this case b−1 belongs to B, cf. [21,
Lemma 4.4]. ♦
In the course of the proof of Theorem 3.2 we showed that for b = I−aV the element b is
invertible in B under condition (3.5). Its inverse is given by d = (C(0)k )k∈Z. Later, dealing
with eventually norm-continuous in θ cocycles, we will need the following refinement of
this fact. Let Anorm be the closed subalgebra Cb(Θ;L(X)) of A of all norm continuous,
bounded, operator valued functions a : Θ → L(X). Let Bnorm denote the corresponding
closed subalgebra of B, that is, Bnorm = (ak)k∈Z ∈ B : ak ∈ Anorm.
Corollary 3.5. Assume that a ∈ Cb(Θ;L(X)) is norm continuous. If the conditions of
Theorem 3.2 hold, then d = b−1 belongs to Bnorm.
Proof. We have to check that θ 7→ C(0)k (θ) ∈ L(X) is norm continuous (cf. Claim 1 in
the proof of Theorem 3.2). Fix k ∈ Z. Recall the estimate (3.13)
|[C(0)k (θ)− C(0)
k (θ0)]x| ≤ B ‖[πθ0(b)− πθ(b)]y‖c0(Z;X),
for x ∈ X and the constant B given in (3.5), and
y = [πθ0(b)]−1x =
∞∑l=−∞
diag(C(n)l (θ0))n∈Z S
lx = (C(n)n−k(θ0)x)n∈Z = (Cn,k(θ0)x)n∈Z,
where x = δn,−k ⊗ x and we have used (3.12). The inequality (3.11) yields ‖Cn,k(θ0)‖ ≤Re−γ|n−k| for n ∈ Z. Fix ε > 0 and take N = Nε such that Re−γ|n−k| < ε for |n| ≥ N .
Using the continuity of a : Θ→ L(X), choose δ > 0 such that
‖a(ϕnθ)− a(ϕnθ0)‖L(X) < ε provided |n| < N and d(θ, θ0) < δ.
For d(θ, θ0) < δ, we now conclude that
‖[πθ0(b)− πθ(b)]y‖c0(Z;X)
≤ max sup|n|<N
|[a(ϕnθ)− a(ϕnθ0)]Cn,k(θ0)x|, sup|n|≥N
|[a(ϕnθ)− a(ϕnθ0)]Cn,k(θ0)x|
≤ εmaxR, 2 supθ∈Θ‖a(θ)‖L(X) |x|.
13
3.2. Discrete Dichotomy Theorem. The algebraic method developed above now en-
ables us to prove the main result of this section. This result connects global dichotomy
on Θ as defined in Definition 2.1, dichotomy at points θ ∈ Θ as defined in Definition 2.2,
the hyperbolicity of the operator given by (Ef)(θ) = Φ(ϕ−1θ)f(ϕ−1θ) on C0(Θ;X), and
the hyperbolicity of the operators πθ(E) = diag(Φ(ϕn−1θ))n∈Z S on c0(Z;X).
We start with some elementary facts. First, observe that
πθ(E)k = πθ(Ek) = diag(Φk(ϕn−kθ))n∈Z S
k (3.14)
for k ∈ N and θ ∈ Θ. We set ‖T‖•,Y = inf‖Ty‖ : ‖y‖ = 1 for a bounded operator T
on a Banach space Y . Observe that the operator πθ(E) is also defined on `∞(Z, X), the
space of bounded sequences endowed with the sup-norm.
Proposition 3.6. Let z ∈ T and θ ∈ Θ. We have ‖I − πθ(E)‖• = ‖zI − πθ(E)‖•on c0(Z;X) or `∞(Z, X). Further, the spectrum σ(πθ(E)) in c0(Z;X) or `∞(Z, X) is
invariant with respect to rotations centered at the origin, and if σ(πθ(E)) ∩ T = ∅, then
‖(I − πθ(E))−1‖ = ‖(zI − πθ(E))−1‖.
Proof. Define Lξ = diag(e−inξ)n∈Z on c0(Z;X) or `∞(Z, X) for ξ ∈ R. Then Lξ is an
invertible isometry on both spaces. So the identity
Recall that due to Corollary 2.9 the dichotomies of a cocycle for continuous and discrete
times are equivalent. Thus, we can work with discrete time.
Theorem 3.7 (Discrete Dichotomy Theorem). Let ϕ : Θ→ Θ be a homeomorphism on a
locally compact metric space Θ and let Φtt∈N with Φ1 = Φ ∈ Cb(Θ;Ls(X)) be a discrete
time cocycle over ϕ. The following assertions are equivalent.
(a) Φt has exponential dichotomy on Θ over Z.
(b) Φt has exponential dichotomy over Z at each point θ ∈ Θ with dichotomy con-
stants β(θ) ≥ β > 0 and M(β(θ)) ≤M .
(c) πθ(E) is hyperbolic on c0(Z;X) for each θ ∈ Θ and
supz∈T
supθ∈Θ‖[zI − πθ(E)]−1‖L(c0(Z;X)) <∞.
(d) E is hyperbolic on C0(Θ;X).
If (a)-(d) hold, let P (·) be the dichotomy projection for Φt on Θ, let Pn, n ∈ Z, be the
family of the dichotomy projections for Φt at θ ∈ Θ, let pθ ∈ L(c0(Z;X)), θ ∈ Θ, be the
Riesz projection for πθ(E) on c0(Z;X) that corresponds to σ(πθ(E))∩D, and let P be the
Riesz projection for E on C0(Θ;X) that corresponds to σ(E) ∩ D. Then,
P = P (·), Pn = P (ϕnθ), and pθ = diag(Pn)n∈Z.
Proof. (a) ⇒ (b). Let Φtt≥0 have exponential dichotomy for discrete time with projec-
tions P (·) ∈ Cb(Θ;Ls(X)) and constants β,M > 0, see Definition 2.1. For θ ∈ Θ, the
projections Pn := P (ϕnθ), n ∈ Z, clearly satisfy Definition 2.2 with constants β(θ) = β
and M(β(θ)) = M .14
(b) ⇒ (c) Fix θ ∈ Θ. Let Pn, n ∈ Z, be the family of dichotomy projections for Φtat θ as in Definition 2.2. Define pθ = diag(Pn)n∈Z on c0(Z;X). By (a) in Definition 2.2,
we have pθπθ(E) = πθ(E)pθ. The identity (3.14) and Definition 2.2(b) imply that
(πθ(E)pθ)k = diag(Φk(ϕn−kθ)Pn−k)n∈Z S
k and
(πθ(E)qθ)−k = diag([Φk
Q(ϕnθ)Qn]−1)n∈Z S−k = diag(Φ−kQ (ϕn+kθ)Qn+k)n∈Z S
−k,
for k ∈ N. Thus, the estimates in Definition 2.2(c) give
‖(πθ(E)pθ)k‖L(c0(Z;X)) ≤M e−βk and ‖(πθ(E)qθ)
−k‖L(c0(Z;X)) ≤M e−βk.
As a result, πθ(E) is hyperbolic and (z, θ) 7→ ‖[zI − πθ(E)]−1‖L(c0(Z;X)) is bounded.
(c)⇒ (d). For every z ∈ T, apply Theorem 3.2 to b = I−aV , where a(θ) = z−1Φ(ϕ−1θ)
and (V f)(θ) = f(ϕ−1θ) for θ ∈ Θ.
(d) ⇒ (a) and P = P (·) can be shown exactly as Theorem 2.7.
One can improve the equivalence (b)⇔(c) as follows.
Corollary 3.8. The cocycle Φtt∈N has exponential dichotomy at θ ∈ Θ over Z if and
only if πθ(E) is hyperbolic on c0(Z;X).
Proof. Necessity was proved, in fact, in Theorem 3.7 (b)⇒(c). So let πθ(E) be hyperbolic.
As in [20, Lemma 3.2], we see that pθχ = χpθ for every χ = (χn) ∈ `∞(Z). By arguments
similar to those used in the proof of [40, Lemma 7], one obtains that pθ0 = diag(Pn)n∈Z for
a family of uniformly bounded projections Pn ∈ L(X), n ∈ Z. From pθπθ(E) = πθ(E)pθwe derive (a) in Definition 2.2. Further, Φk(ϕn−kθ) : ImQn−k → ImQn is surjective due
to (3.14) and the surjectivity of πθ(E)k on Im qθ. Finally, the estimates in (c) follow as in
the proof of Theorem 2.7.
In addition, for the case K = R, we can characterize pointwise dichotomy of the cocycle
by means of evolution semigroups along trajectories, Πtθt≥0 on C0(R;X), which are
defined by
(Πtθh)(s) = Φt(ϕs−tθ)h(s− t), s ∈ R, t ≥ 0, h ∈ C0(R;X),
for each θ ∈ Θ, cf. Theorem 5.9 of [21]. Since Φt is an exponentially bounded, strongly
continuous cocycle, the operators Uθ(s, τ) := Φs−τ (ϕτθ), s ≥ τ , yield an exponentially
bounded, strongly continuous evolution family, see [21, p.110]. In particular, (Πtθh)(s) =
Uθ(s, s − t)h(s − t) and Πtθ is an evolution semigroup on C0(R;X), see [20, 37] for
the definitions. By the results in [20, 36, 39], the evolution family Uθ(s, τ)s≥τ has
exponential dichotomy on R if and only if Πtθt≥0 is hyperbolic if and only if Γθ is
invertible. But dichotomy of Uθ(s, τ)s≥τ means exactly the dichotomy of Φtt≥0 at θ
over R. The next result now follows from the Dichotomy Theorem 2.7 and the techniques
used in the proof of Theorem 5.9 of [21].
Theorem 3.9. The following assertions are equivalent for an exponentially bounded,
strongly continuous cocycle Φtt≥0 over a continuous flow ϕtt∈R on Θ.
(a) Φtt≥0 has exponential dichotomy on Θ.
(b) σ(Π1θ) ∩ T = ∅ for each θ ∈ Θ and
supz∈T
supθ∈Θ‖[z − Π1
θ]−1‖L(C0(R;X)) <∞.
15
(c) σ(Γθ) ∩ iR = ∅ for each θ ∈ Θ and
supξ∈R
supθ∈Θ‖[iξ − Γθ]
−1‖L(C0(R;X)) <∞.
(d) Φtt≥0 has exponential dichotomy over R at each point θ ∈ Θ with dichotomy
constants β(θ) ≥ β > 0 and M(β(θ)) ≤M .
Moreover, the Riesz projection P for the operator E, the Riesz projection pθ for the
operator Π1θ, the dichotomy projection P (·) for the dichotomy of Φt on Θ, and the
dichotomy projections P θτ τ∈R for the dichotomy of Φt at a point θ ∈ Θ are related as
follows:
P = P (·), P θτ = P (ϕτθ), (pθh)(τ) = P θ
τ h(τ), h ∈ C0(R;X).
3.3. Perturbation theorem. In this subsection we give an infinite-dimensional general-
ization of the Sacker-Sell Perturbation Theorem (see Theorem 6 in [43]) which establishes
the semicontinuity of the dynamical spectrum as a function of Θ. To this end, we assume
that Θ is a compact metric space and that the cocycle Φt is eventually norm continuous
in θ. This means (cf. [33, p. 38]) that for some t0 > 0 the map Θ 3 θ 7→ Φt0(θ) ∈ L(X)
is continuous in operator norm. By rescaling the time, we can assume that t0 = 1. Also,
we work with discrete time t ∈ N. We start with a preliminary fact.
Proposition 3.10. Assume that Θ 3 θ 7→ Φ1(θ) ∈ L(X) is norm continuous. If
σ(E)∩T = ∅, then Φnn∈N has exponential dichotomy with a norm continuous dichotomy
projection P (·).
Proof. Assume that E is hyperbolic. By the Discrete Dichotomy Theorem 3.7 we know
that the Riesz projection pθ for πθ(E) = πθ(aV ) is given by pθ = diag(P (ϕnθ))n∈Z. Notice
that a = Φ1 ϕ−1 ∈ Anorm, see the notations before Corollary 3.5. Define a = (an)n∈Z ∈Bnorm by a1 = a and an = 0 otherwise. Let bz = e − z−1a for z ∈ T. Recall that
πθ(bz) = πθ(I − z−1aV ) = I − z−1πθ(E) by (3.6). The Discrete Dichotomy Theorem 3.7
yields
‖(I − z−1πθ(E))−1‖L(c0(Z;X)) ≤ C for all θ ∈ Θ and z ∈ T.Applying Corollary 3.5 to bz = I − z−1aV , we see that for each z ∈ T the element
ze− a ∈ Bnorm is invertible in Bnorm. Define the idempotent p in the algebra Bnorm by
p =1
2πi
∫T(ze− a)−1dz.
Since πθ : B→ C is a continuous homomorphism for all θ ∈ Θ, we obtain
πθ(p) =1
2πi
∫Tπθ([ze− a]−1) dz =
1
2πi
∫T[zI − πθ(aV )]−1 dz = pθ
for every θ ∈ Θ. On the order hand, let p = (Pk)k∈Z for Pk ∈ Anorm. Then we have
0 = πθ(p)− pθ = πθ(P0)− pθ +∑k 6=0
πθ(Pk)Sk
by the definition (3.3) of πθ and πθ(Pk) = diag[Pk(ϕnθ)]n. Since the representation of
πθ(p) − pθ ∈ C as a series in powers of S is unique, we conclude that πθ(P0) = pθ and
πθ(Pk) = 0 for k 6= 0 and each θ ∈ Θ. So Proposition 3.1 yields Pk = 0 for k 6= 0, and
hence P = P0 ∈ Anorm (use the sequence (δn0Pk)n ∈ B). 16
Let K be the set of all closed, ϕ-invariant subsets of Θ. For fixed Θ0 ∈ K and the
metric d(·, ·) on Θ, define
distΘ0(Θ1) = supθ1∈Θ1
infθ0∈Θ0
d(θ0, θ1) for Θ1 ∈ K.
Further, denote by Σ(Θ1) = Σ(ϕn|Θ1 × X) the Sacker-Sell dynamical spectrum, see
Definition 2.4, for the cocycle Φnn∈N restricted on Θ1, that is, for Φnn∈N = Φn(θ) :
n ∈ N, θ ∈ Θ1, or for the linear skew-product flow ϕn restricted to Θ1 ×X. The next
result implies the upper semicontinuity of the function Θ1 7→ Σ(Θ1) on K.
Theorem 3.11 (Perturbation Theorem). Let Φtt∈Z+ with Φ1 = Φ ∈ Cb(Θ;L(X)) be
a discrete time cocycle over a homeomorphism ϕ on a compact metric space Θ. Assume
that Φtt∈N has exponential dichotomy on the set Θ0 ∈ K in the sense of Definition
2.1. There exists δ∗ > 0 such that the cocycle Φtt∈N has exponential dichotomy on each
Θ1 ∈ K with the property
distΘ0(Θ1) ≤ δ for some δ ∈ (0, δ∗). (3.15)
The proof of the theorem uses nothing but elementary calculations with Neumann series
and the Discrete Dichotomy Theorem. Let us sketch our argument. For ε > 0, a sequence
θ = θnn∈Z ⊂ Θ is called ε-pseudo-orbit for ϕ if d(ϕθn, θn+1) ≤ ε for all n ∈ Z. As an
analogue to the operators πθ(E), we define the operator Tθ : (xn)n∈Z 7→ (Φ(θn−1)xn−1)n∈Zon c0(Z;X) for every ε-pseudo-orbit θ = θnn∈Z ⊂ Θ0.
Assuming that Φt has exponential dichotomy on Θ0, we will show that for sufficiently
small ε the operators I − Tθ are invertible on c0(Z, X) with uniformly bounded norms
‖(I − Tθ)−1‖ for all ε-pseudo-orbits θ ⊂ Θ0. Fix θ ∈ Θ1 and choose an ε-pseudo-orbit
θ = θn ⊂ Θ0 such that d(θn, ϕnθ) is small for each n ∈ Z. We use the continuity of Φ(·)
to show that for this choice of θn the norm ‖πθ(E)−Tθ‖L(c0(Z;X)) is small for sufficiently
small ε > 0. Since I−Tθ is invertible for θ ⊂ Θ0, we conclude that I−πθ(E) is invertible
and ‖(I −πθ(E))−1‖ ≤ C for each θ ∈ Θ1. Hence, Φt has exponential dichotomy on Θ1
by the Discrete Dichotomy Theorem.
Proof. Assume that Φt has exponential dichotomy on Θ0 ∈ K with projection P (·) and
constants M , β. Set c = maxθ∈Θ0‖P (θ)‖, 1. Choose γ > 0 such that
γ ≤ 116c. (3.16)
Fix N ∈ N such that Me−βN ≤ γ. Set ψ = ϕN and Ψ(θ) = ΦN(θ). We derive for all
θ ∈ Θ0 that
Ψ(θ)P (θ) = P (ψθ)Ψ(θ), (3.17)
ΨQ(θ) : ImQ(θ)→ ImQ(ψθ) is invertible, (3.18)
‖ΨP (θ)‖ ≤ γ, (3.19)
‖Ψ−1Q (θ)‖ ≤ γ. (3.20)
17
Recall that P (·) and Ψ(·) are norm-continuous and Θ0 and Θ are compact. Fix ε∗ > 0
such that for every ε ∈ (0, ε∗) the following inequalities hold:
supd(θ,η)≤ε, θ,η∈Θ ‖Ψ(θ)−Ψ(η)‖ ≤ 3/16, (3.21)
supd(θ,η)≤ε, θ,η∈Θ0‖P (θ)− P (η)‖ ≤ 1
2(1+2c), (3.22)
c supd(θ,η)≤ε, θ,η∈Θ0‖P (θ)− P (η)‖ ·maxθ∈Θ ‖Ψ(θ)‖ ≤ 3/8. (3.23)
For a sequence θ = θnn∈Z ⊂ Θ0, we define by
Tθ : (xn) 7→ (Ψ(θn−1)xn−1), Pθ = diag P (θn), Qθ = I − Pθ, pθ = diag P (ψθn−1)
operators on c0(Z;X). Then (3.17) implies
TθPθ = pθTθ and TθQθ = (I − pθ)Tθ. (3.24)
The following lemma shows that exponential dichotomy, (3.17)–(3.20), implies the in-
vertibility of I−Tθ for every ε-pseudo-orbit θ for ψ in Θ0, where ε is given by (3.21)–(3.23).
Lemma 3.12. Let θ = θn ⊂ Θ0 be an ε-pseudo-orbit for ψ, where ε satisfies (3.21)–
(3.23). Then I − Tθ is invertible on c0(Z, X) and
‖(I − Tθ)−1‖ ≤ 83· (3.25)
We prove the lemma later and finish the proof of Theorem 3.11 first. Fix ε satisfying
(3.21)–(3.23). Fix δ∗ ≤ ε/4 such that d(ψθ, ψη) ≤ ε/2 as soon as d(θ, η) < 2δ∗. Let δ < δ∗
and let Θ1 ∈ K satisfy (3.15). For θ ∈ Θ1, define the operator Tθ := πθ(EN) : (xn) 7→
(Ψ(ψn−1θ)xn−1).
For a given θ ∈ Θ1 and each n ∈ Z, we use (3.15) to find θn ∈ Θ0 such that d(θn, ψnθ) <
2δ. The sequence θ = θn ⊂ Θ0 is an ε-pseudo-orbit for ψ because of
Hence, |xsm| → ∞ and |xum| → 0 as m→∞. This contradicts (4.2) with k = 0.
We now give sufficient conditions for the existence of a Mane sequence, cf. [8, 29], using
the following refinement of Theorem 3.7 (c)⇒(d).
Lemma 4.3. Assume there exists a constant c > 0 such that
‖I − πθ(E)‖•,`∞(Z;X) ≥ c for all θ ∈ Θ.
Then we have ‖zI − E‖•,F ≥ c for all z ∈ T.
Proof. Let f ∈ F and z ∈ T. Notice that for each θ ∈ Θ the sequence xθ := (f(ϕkθ))k∈Zis bounded. The assumption and Proposition 3.6 yield
‖zf − Ef‖F = supθ∈Θ|zf(θ)− Φ1(ϕ−1θ)f(ϕ−1θ)| = sup
θ∈Θsupk∈Z|zf(ϕkθ)− Φ1(ϕk−1θ)f(ϕk−1θ)|
= supθ∈Θ‖[zI − πθ(E)]xθ‖`∞(Z;X) ≥ c sup
θ∈Θ‖xθ‖`∞(Z;X) = c ‖f‖F .
21
Proposition 4.4. A Mane sequence θm∞1 exists provided one of the following conditions
holds.
(a) infθ∈Θ ‖I − πθ(E)‖•,`∞ = 0.
(b) σap(E) ∩ T 6= ∅.
Proof. In view of Lemma 4.3, it suffices to consider (a). Since Φt is exponentially
bounded, N := supθ ‖πθ(E)‖ is finite. Fix m ∈ N and let ε := (4∑2m
k=0 Nk)−1. By the
assumption, there exists σm ∈ Θ and y(m) = (y(m)l )l∈Z ∈ `∞(Z;X) such that ‖y(m)‖∞ = 1
and ‖[I − πσm(E)]y(m)‖∞ < ε. Hence,
12< ‖[πσm(E)]ky(m)‖`∞(Z;X) < 2 for k = 0, 1, . . . 2m.
For a fixed k ∈ 0, 1, . . . , 2m, we have
‖[πσm(E)]ky(m)‖l∞(Z;X) = supl∈Z|Φk(ϕl−kσm)y
(m)l−k | = sup
l∈Z|Φk(ϕl−mσm)y
(m)l−m|.
Select l such that 12≤ |Φm(ϕl−mσm)y
(m)l−m|. Note that |Φk(ϕl−mσm)y
(m)l−m| ≤ 2 holds for
k = 0, 1, . . . 2m. Therefore, θm := ϕl−mσm is a Mane sequence with the vectors xm :=
y(m)l−m.
In the construction of the approximate eigenfunctions we will use some basic facts
concerning the period function and flow boxes, or tubes.
A point θ ∈ Θ is called periodic with respect to the flow ϕtt∈R if ϕT θ = θ for some
T > 0, and aperiodic if there is no such T . The prime period function p : Θ→ R∪∞ is
defined as p(θ) = infT > 0 : ϕT θ = θ for periodic points θ and p(θ) =∞ for aperiodic
θ. Further, set
B(Θ) = θ ∈ Θ : p is bounded in a neighborhood of θ and
BC(Θ) = θ ∈ B(Θ) : p is continuous at θ.
Also, we denote by O(θ) = ϕtθ : t ∈ R the orbit through θ ∈ Θ. The following result
is taken from [6, Thm. IV.2.11].
Lemma 4.5. Let ϕtt∈R be a continuous flow on a locally compact, metric space Θ with
prime period function p. Suppose that θ0 ∈ Θ and τ are such that 0 < τ < p(θ0)/4. Then
there exists an open set U and a set Σ such that θ0 ∈ Σ ⊂ U , the closure U is compact,
and for each θ ∈ U there exists a unique number t = t(θ) in (−τ, τ) such that the point
σ := ϕ−tθ belongs to Σ. Moreover, the map θ 7→ t(θ) is continuous on U .
The set U as in Lemma 4.5 is called flow-box of length τ with cross-section Σ at θ0 ∈ Θ.
Lemma 4.6. For a continuous flow ϕtt∈R on a locally compact metric space Θ with
prime period function p, we have
(a) p is lower semicontinuous, in particular, the set θ : p(θ) ≤ d is closed for each
d ≥ 0;
(b) the points of continuity of p are dense in Θ;
(c) for θ0 ∈ BC(Θ) with p(θ0) > 0, there exists a relatively compact, open set Uand a set Σ 3 θ0 such that O(θ0) ⊂ U ⊂ BC(Θ), for θ ∈ U there is unique
t = t(θ) ∈ [0, p(θ)) with ϕ−tθ ∈ Σ, and U 3 θ 7→ t(θ) is continuous.22
Proof. Assertion (a) is shown in [3, p.314]. The second assertion follows from (a) and [15,
pp.87]. To show (c), we use Lemma 4.5 to find a relatively compact, open flow box U0 at
θ0 of length τ ∈ (0, p(θ0)/4) with cross section Σ0 such that θ 7→ t(θ) is continuous, the
function p takes values in (3p(θ0)/4, 5p(θ0)/4) on U0, and
O(θ0) ∩ Σ0 = θ0. (4.3)
Let Σ1 be a relatively open subset of Σ0 containing θ0 such that Σ1 ⊂ U0. We claim that
O(θ0) has an open neighborhood U ⊂ U0 satisfying
for all θ ∈ U the set O(θ) ∩ Σ1 contains exactly one element. (4.4)
Suppose that no such neighborhood exists. Then there were points θn converging to a
point ϕtθ0 for some 0 ≤ t < p(θ0) such that O(θn)∩Σ1 is empty or contains more than one
element. But, since ϕ−tθn → θ0, we obtain ϕ−tθn ∈ U0 and, hence, points σn ∈ O(θn)∩Σ0
for large n. Moreover, t(ϕ−tθn) → t(θ0) = 0 so that σn → θ0. In particular, σn ∈ Σ1 for
large n. Then there must exist σ′n = ϕsnσn ∈ Σ1 for some sn with
0 < τ ≤ sn ≤ p(σn)− τ < p(θ0)− τ
2
and n sufficiently large, where we have used the continuity of p at θ0. Taking a subse-
quence, we may assume that sn → s so that
σ′n → ϕsθ0 ∈ Σ1 ⊂ U0 .
Again by the continuity of t(·), we derive ϕsθ0 ∈ Σ0. Due to (4.3), this yields ϕsθ0 = θ0.
But this contradicts 0 < s < p(θ0), and so there exists a neighborhood U of O(θ0)
satisfying (4.4).
It remains to show the continuity of p(·) and t(·) on U . So let θn → θ in U . Suppose
that p(θn) does not converge to p(θ). But, p(θn) → p0 for a subsequence, and so θn =
ϕp(θn)θn → ϕp0θ. Hence, p0 = kp(θ) for k 6= 1. This is impossible since p(θ), p0 ∈[3p(θ0)/4, 5p(θ0)/4]. Further, ϕrθn → ϕrθ in U0 for some r and sufficiently large n. This
yields the continuity of θ 7→ t(θ) on U .
Recall that Γ denotes the infinitesimal generator of the evolution semigroup Ett≥0 on
F = C0(Θ;X) defined by formula (1.1). Our next goal is to construct the approximate
eigenfunctions for Γ on F for a given Mane sequence. We first treat the case that the
Mane sequence consists of aperiodic points.
Lemma 4.7. Assume θm∞1 is a Mane sequence such that θm /∈ B(Θ) or p(θm) →∞. Then, for each ξ ∈ R, there exist functions fn ∈ D(Γ) with ‖fn‖∞ = 1 such that
‖(iξ − Γ)fn‖∞ → 0 as n→∞.
Proof. Without loss of generality, we assume that p(θm) → ∞ as m → ∞. Denote
n = minm,[
15p(θm)
], where [ · ] is the integer part. Note that n > 0 for sufficiently
large m. Set σ0 = ϕm(θm) and x = Φm−n(θm)xm. By the assumption there are constants
C, c > 0 such that c ≤ |Φm(θm)xm| and |Φk(θm)xm| ≤ C for k = 0, 1, . . . , 2m. This implies
that
0 < c ≤ |Φn(ϕ−nσ0)x| and |Φk(ϕ−nσ0)x| ≤ C for k = 0, 1, . . . , 2n. (4.5)23
Using Lemma 4.5, we find an open flow box U of length n at σ0 with cross-section Σ 3 σ0
such that
|Φt(ϕ−nσ)x− Φt(ϕ−nσ0)x| ≤ 1 for 0 ≤ t ≤ 2n and σ ∈ Σ. (4.6)
For fixed n, choose “bump”-functions β ∈ C(U) with 0 ≤ β ≤ 1, β(σ0) = 1, and compact
support and γ ∈ C1(R) with 0 ≤ γ ≤ 1, supp γ ⊂ (0, 2n), γ(n) = 1, and ‖γ′‖∞ ≤ 2/n.
For ξ ∈ R, define
fn(θ) = e−iξtβ(σ)γ(t)Φt(ϕ−nσ)x if θ = ϕt−nσ ∈ U with t ∈ (0, 2n), σ ∈ Σ, (4.7)
and fn(θ) = 0 if θ /∈ U . Clearly, fn ∈ F and ‖fn‖∞ ≥ |fn(σ0)| ≥ c > 0 by (4.5). For
h > 0 so small that γ = 0 on [2n− h, 2n], one has
(Ehfn)(θ) = e−iξ(t−h)β(σ)γ(t− h)Φt(ϕ−nσ)x
for θ = ϕt−nσ ∈ U and (Ehfn)(θ) = 0 for θ /∈ U . Thus, fn ∈ D(Γ) and
for θ = ϕt−nσ with t ∈ (0, 2n) and σ ∈ Σ and (Γfn)(θ) = 0 for θ /∈ U . Let N =
sup‖Φτ (θ)‖ : τ ∈ [0, 1], θ ∈ Θ. Then (4.6) and (4.5) imply
‖Γfn − iξfn‖∞ ≤2N
nmax
k=0,...,2n|Φk(ϕ−nσ)x| ≤ 2N(C + 1)
n.
If the Mane sequence consists of periodic points, we will make use of the following fact.
Lemma 4.8. Assume there is a sequence θm ⊂ BC(Θ) with 0 < d0 ≤ p(θm) ≤ d1 and
xm ∈ X such that |xm| = 1 and |[z − Φp(θm)(θm)]xm| → 0 as m → ∞ for some z ∈ T.
Then there exist ξ ∈ R and fn ∈ D(Γ) such that ‖fn‖∞ = 1 and ‖(iξ − Γ)fn‖∞ → 0 as
n→∞.
Proof. We may assume that z = eiη = eiη
p(θn)p(θn) and p(θn)→ p0 ∈ [d0, d1]. Set ξ = η/p0.
By Proposition 4.6, there exist open tubular neighborhoods Un of θn with sections Σn 3 θnsuch that on Un the function p is continuous and takes values in [d0/2, 2d1] and
|xn − e−iξp(σ)Φp(σ)(σ)xn| → 0 for σ ∈ Σn as n→∞. (4.9)
Take a continuous “bump”-function β : Un → [0, 1] with β(θn) = 1 and compact support.
Choose a function α ∈ C1[0, 1] satisfying 0 ≤ α ≤ 1, α = 0 on [0, 1/3], α = 1 on [2/3, 1],
for θ = ϕtσ ∈ Un, where σ ∈ Σn and 0 ≤ t < p(σ), and fn(θ) = 0 for θ /∈ Un. Clearly,
fn ∈ F with
‖fn‖∞ ≥ |fn(θn)| = |Φp(θn)(θn)xn| ≥ 1/2
for large n. For 0 < h < d0/6, we compute
(Ehfn)(θ) =
e−iξ(t−h)β(σ)Φt(σ)[α( t−h
p(σ))xn + (1− α( t−h
p(σ)))e−iξp(σ)Φp(σ)(σ)xn], t ≥ h,
e−iξ(t−h)β(σ)Φt(σ)e−iξp(σ)Φp(σ)(σ)xn, t < h,24
for θ = ϕtσ ∈ Un and (Ehfn)(θ) = 0 otherwise. (Write ϕ−hθ = ϕp(θ)+t−hσ if θ ∈ Un and
t− h < 0.) Thus, fn ∈ D(Γ) and
Γfn(θ)− iξfn(θ) = −e−iξtβ(σ)1
p(σ)α′( t
p(σ))Φt(σ)[xn − e−iξp(σ)Φp(σ)(σ)xn] (4.11)
for θ = ϕtσ ∈ Un and zero otherwise. Let N = sup‖Φτ (θ)‖ : θ ∈ Θ, 0 ≤ τ ≤ 2d1.Finally we obtain
‖(Γ− iξ)fn‖∞ ≤8N
d0
|xn − e−iξp(σ)Φp(σ)(σ)xn| → 0
as n→∞ thanks to (4.9).
We are now in the position to prove the annular hull theorem and the spectral mapping
theorem for the evolution semigroup
(Etf)(θ) = Φt(ϕ−tθ)f(ϕ−tθ) on F = C0(Θ;X)
with generator Γ. The spectral mapping theorem says that
σ(Et)\0 = exp tσ(Γ) for t ≥ 0. (4.12)
In general, this result does not hold if ϕt is not aperiodic, see e.g. [9, Ex. 2.3]. The annular
hull theorem states that
exp tσ(Γ) ⊆ σ(Et)\0 ⊆ H(exp tσ(Γ)) for t ≥ 0,
where H(·) is the annular hull of the set (·), that is, the union of all circles centered at the
origin that intersect the set (·). In other words, the annular hull theorem can be written
as
T · (σ(Et)\0) = T · exp tσ(Γ) for t ≥ 0. (4.13)
Observe that the annular hull theorem can be violated if there are fixed points. Indeed,
for the one-point set Θ = θ and identical flow ϕt(θ) ≡ θ define a cocycle Φt(θ) = etA
for a strongly continuous semigroup etA on X which does not satisfy the annular hull
theorem (see [33, A-III] for examples). Clearly, in this case (4.13) does not hold for the
corresponding evolution semigroup.
Theorem 4.9 (Spectral Mapping/Annular Hull Theorem). Let Ett≥0 be the evolu-
tion semigroup on F = C0(Θ;X) with generator Γ induced by an exponentially bounded,
strongly continuous cocycle Φtt∈R+ over a continuous flow ϕtt∈R on a locally compact
metric space Θ, and let p be the function of prime periods for ϕt.(a) If B(Θ) = ∅, then σ(Γ) is invariant with respect to vertical translations, σ(Et)
is invariant with respect to rotations centered at zero, and the spectral mapping
theorem (4.12) holds.
(b) If p(θ) ≥ d0 > 0 for all θ ∈ B(Θ), then the annular hull theorem (4.13) holds.
Proof. By virtue of the spectral inclusion theorem and the spectral mapping theorem
for the residual spectrum, see e.g. [33, A-III.6], and a standard rescaling, (4.13) holds
provided that
σap(E) ∩ T 6= ∅ implies σ(Γ) ∩ iR 6= ∅25
for E = E1. So let z ∈ σap(E) ∩ T. By Proposition 4.4, there exists a Mane sequence
θm in Θ. We have to consider two cases:
Aperiodic Case: there is a subsequence with θm /∈ B(Θ) or p(θm)→∞ as m→∞;
Periodic Case: θm ∈ B(Θ) and p(θm) < d1 for a constant d1 .
In the aperiodic case, Lemma 4.7 shows that iR ⊆ σap(Γ) which implies (4.12). The
other assertions in (a) then follow from the spectral inclusion theorem [33, A-III.6.2] and
∂σ(Γ) ⊂ σap(Γ).
In the periodic case, using Lemma 4.6, we may and will assume that θm ∈ Ω := θ ∈BC(Θ) : d0 ≤ p(θ) ≤ d1. The set Ω is a ϕt-invariant and locally compact metric space by
Lemma 4.6. On Ω, we define a new flow ψtt∈R by ψt(θ) = ϕtp(θ)(θ) and a new cocycle
Ψtt∈R+ over ψt by Ψt(θ) = Φtp(θ)(θ). Recall that θ 7→ p(θ) is continuous on BC(Θ).
Clearly, ψt is continuous and Ψt is strongly continuous and exponentially bounded
on Ω. So there exists the induced evolution semigroup EtΨ on C0(Ω;X). Let EΨ = E1
Ψ.
We stress that (EΨf)(θ) = Φp(θ)(θ)f(θ) for f ∈ C0(Ω;X) and θ ∈ Ω. The following fact
is stated as a separate lemma for later use.
Lemma 4.10. Let Ω be as defined above. Assume that σ(Φp(θ)(θ))∩T = ∅ for each θ ∈ Ω
and
supθ∈Ω‖[zI − Φp(θ)(θ)]−1‖L(X) <∞ for all z ∈ T.
Then Φt has exponential dichotomy on Ω.
Proof of Lemma 4.10. The assumptions imply σ(EΨ) ∩ T = ∅, where
(zI − EΨ)−1f(θ) = [zI − Φp(θ)(θ)]−1f(θ)
for θ ∈ Ω, z ∈ T, and f ∈ C0(Ω;X). An application of the Dichotomy Theorem 2.7 to
EΨ shows that the cocycle Ψt has exponential dichotomy on Ω with the dichotomy
projection P (·) and the dichotomy constants β,M > 0. Therefore, the cocycle Φt has
exponential dichotomy on Ω with P (·), β/d1, and M .
Since Ω contains a Mane sequence, the cocycle Φt does not have exponential di-
chotomy on Ω due to Proposition 4.2. Therefore the assumptions in Lemma 4.10 cannot
be satisfied. This means that either
(1) ‖zI − Φp(θn)(θn)‖•,X → 0 for some z ∈ T and θn ∈ Ω or
(2) there exist z ∈ T, θ0 ∈ Ω, y ∈ X, and δ > 0 such that |zx − Φp(θ0)(θ0)x − y| ≥ δ
for all x ∈ X.
In case (1), Lemma 4.8 implies σap(Γ)∩ iR 6= ∅. In case (2), take a function g ∈ C0(Θ;X)
that is, z ∈ σr(Ep(θ0))∩T. Hence, σr(Γ)∩ iR 6= ∅ by the spectral mapping theorem for the
residual spectrum [33, A-III.6.3]. Altogether, we have established (b) in Theorem 4.9.
Combining the above result with the Dichotomy Theorem 2.7, we obtain the following.26
Corollary 4.11. Let Φtt∈R+ be an exponentially bounded, strongly continuous cocycle
Φtt∈R+ over a continuous flow ϕtt∈R on a locally compact metric space Θ, let p(θ) ≥d0 > 0 for all θ ∈ B(Θ), and let Γ be the generator of the induced evolution semigroup on
C0(Θ;X). Then the cocycle has exponential dichotomy on Θ if and only if iR ⊂ ρ(Γ). If
B(Θ) = ∅, then these conditions are equivalent to ρ(Γ) ∩ iR 6= ∅.
We have seen in Proposition 4.4 that the condition σap(E)∩T 6= ∅ implies the existence
of a Mane sequence. We can now give two results in the opposite direction.
Corollary 4.12. Assume that σr(E) ∩ T = ∅. Then, σap(E) ∩ T 6= ∅ if and only if there
exists a Mane sequence θm∞1 .
Proof. If σap(E) ∩ T 6= ∅, then a Mane sequence exists by Proposition 4.4. If a Mane
sequence exists, then σ(E) ∩ T 6= ∅ by Proposition 4.2 and Dichotomy Theorem 2.7. So
the assumption, σr(E) ∩ T = ∅, implies the result.
Corollary 4.13. Assume that θm∞1 is a Mane sequence.
(a) If the aperiodic case holds for θm∞1 , that is, there is a subsequence such that
θm /∈ B(Θ) or p(θm)→∞ as m→∞, then T ⊂ σap(E).
(b) If the periodic case holds for θm∞1 , that is, θm ∈ Ω = θ ∈ BC(θ) : d0 ≤p(θ) ≤ d1 for some constant d1, and if σr(Φ
p(θ)(θ)) ∩ T = ∅ for all θ ∈ Ω, then
σap(E) ∩ T 6= ∅.
Proof. In the aperiodic case (a), one has iR ⊂ σap(Γ) for the generator Γ of the evolution
semigroup Et on C0(Θ;X) by Lemma 4.7. So the spectral inclusion theorem yields
T ⊂ σap(E).
In the periodic case (b), the cocycle Φt does not have exponential dichotomy on Ω by
Proposition 4.2. Therefore, the assumptions of Lemma 4.10 do not hold. Since we have
assumed σr(Φp(σ)(σ)) ∩ T = ∅ for all θ ∈ Ω, we conclude that ‖zI − Φp(θn)(θn)‖•,X → 0
for some z ∈ T and θn ∈ Ω. Lemma 4.8 implies iξ ∈ σap(Γ) for some ξ ∈ R, and hence
σap(E) ∩ T 6= ∅ by the spectral inclusion theorem.
5. Dichotomy and Mild Solutions
In the sequel, we relate dichotomy of the cocycle Φt or the linear skew-product flow
ϕt on Θ × X to the existence and uniqueness of the solution u for the mild form of
for m > n. As a consequence, D(γ) has dense range on c0(Z;X) and is invertible by (5.8).
Now, (5.7) implies the invertibility of I − πθ0(E) on c0(Z;X) in case (1).
(2) Assume that p(θ0) ∈ (0,∞). We start with a special case.
(2i) Let p(θ0) = 1. Lemma 5.5 and (5.3) yield σap(Φ1(θ0))∩T = ∅. Further, eλ−Φ1(θ0)
is surjective by Lemma 5.6 for all λ = iξ ∈ iR in case (a) and for λ = 0 in case (b) of the
theorem. Therefore, eλ ∈ ρ(Φ1(θ0)) and the boundary of σ(Φ1(θ0)) does not intersect T.
As a result, Φ1(θ0) is hyperbolic and, by Lemma 5.5, I − πθ0(E) is invertible on c0(Z;X).
(2ii) Let p(θ0) = T ∈ (0,∞). We define the continuous flow ψt(θ) = ϕtT (θ) and the
exponentially bounded, strongly continuous cocycle Ψt(θ) = ΦtT (θ) over ψt on Θ. This
cocycle satisfies condition (MλT , Cb(Θ;X)). Also, for the flow ψt one has pψ(θ0) = 1. An
application of parts (I) and (II.2.i) to the closed ψt-invariant set O(θ0) and the cocycle33
Ψ shows that the corresponding operators I − πθ(EΨ) are invertible on c0(Z;X) and
‖(I − πθ(EΨ))−1‖ ≤ C for θ ∈ O(θ0). Therefore Ψt has exponential dichotomy on O(θ0)
due to the Discrete Dichotomy Theorem 3.7. Hence, Φt has exponential dichotomy on
O(θ0) which implies the invertibility of I − πθ0(E).
(3) Assume that p(θ0) = 0. Then Φt(θ0) = etA is a strongly continuous semigroup
on X. For y ∈ X, choose g ∈ Cb(Θ;X) with g(θ0) = y. By condition (Mλ,F) (for all
λ = iξ ∈ iR in case (a) and for λ = 0 in case (b) of the theorem), there is a function
f ∈ Cb(Θ;X) such that f(θ0) =: x satisfies
eλtx− etAx =
∫ t
0
e(t−τ)Aeλτy dτ.
Consequently, x ∈ D(A) and (λ − A)x = y, i.e., λ /∈ σr(A). Lemma 5.5 and (5.3) yield
σap(eA)∩T = ∅. Hence, σap(A)∩ iR = ∅ by the spectral inclusion theorem. In particular,
λ /∈ σ(A) and the boundary of σ(A) does not intersect iR. So we infer that σ(A)∩ iR = ∅and, by the spectral mapping theorem for the residual spectrum, σr(e
A) ∩ T = ∅. As a
result, Φ1(θ0) is hyperbolic. So Lemma 5.5 implies 1 ∈ ρ(πθ0(E)).
We combine the results of this section in the following characterization.
Corollary 5.9. Let Φtt∈R+ be an exponentially bounded, strongly continuous cocycle
Φtt∈R+ over a continuous flow ϕtt∈R on a locally compact metric space Θ. Let F ∈Cb(Θ, X), C0(Θ, X).
(a) Assume that p(θ) ≥ d0 > 0 for θ ∈ B(θ). Then condition (Mλ,F) holds for all
λ = iξ ∈ iR if and only if Φt has exponential dichotomy on Θ.
(b) Assume that B(Θ) = ∅. Then condition (M0,F) holds if and only if Φt has
exponential dichotomy on Θ.
Note added in proof.
We remark that F. Rabiger [private communication] recently gave a direct proof of the fol-
lowing fact: If B(Θ) = ∅, then condition (M0, Cb(Θ;X)) implies condition (M0, C0(Θ;X)).
In other words, he proved that g ∈ C0(Θ;X) implies R0g ∈ C0(Θ;X) assuming condition
(M0, Cb(Θ;X)). The proof does not use operators πθ(E). It is based on a characterization
theorem for evolution semigroups and their generators similar to, e. g., Theorem 3.4 in
[38], and on the fact that Γu = −g for u, g ∈ C0(Θ;X) if and only if u and g satisfy the
mild integral equation (5.1) with λ = 0 (compare Lemma 1.1 in [32]). An indirect proof
of the fact above is contained in our Theorem 5.7.