Home Search Collections Journals About Contact us My IOPscience

Spectral analysis of linear relations and degenerate operator semigroups

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

2002 Sb. Math. 193 1573

(http://iopscience.iop.org/1064-5616/193/11/A01)

View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 128.59.62.83
The article was downloaded on 15/03/2013 at 19:00

Please note that terms and conditions apply.

Sbornik: Mathematics 193:11 1573–1610
2002
c RAS(DoM) and LMS
Matematicheskiı̆ Sbornik 193:11 3–42 DOI 10.1070/SM2002v193n11ABEH000696

Spectral analysis of linear relations

and degenerate operator semigroups

A. G. Baskakov and K. I. Chernyshov

Abstract. Several problems of the spectral theory of linear relations in Banach

spaces are considered. Linear differential inclusions in a Banach space are studied.
The construction of the phase space and solutions is carried out with the help of
the spectral theory of linear relations, ergodic theorems, and degenerate operator
semigroups.
Bibliography: 29 titles.

§ 1. Introduction
This paper is concerned with certain questions of the spectral theory of linear
relations (multivalued linear operators) and with the construction of solutions of
linear differential inclusions by means of degenerate semigroups of bounded linear
operators. Interest in the theory of linear relations has become apparent in recent
years; one must mention in this connection the two recent monographs [1] and [2],
which contain an extensive bibliography and fortunately complement each other:
the first discusses general issues of the theory of linear relations and, in partic-
ular, their spectral theory, while the second considers applications to differential
equations. Note that in the Russian mathematical literature the theory of linear
relations has never been adequately addressed; we point out the paper [3], which
considers linear relations in Hilbert space and, in particular, questions relating to
indefinite metrics.
We now list the concepts of the theory of linear relations used in what follows.
Aiming at a self-contained presentation we shall not always use the terminology
of [1], [2] (for instance, we deliberately keep clear of the concept of multivalued
linear operator) and, guided by applications to spectral theory, mainly content
ourselves with the discussion of linear relations in one Banach space, although
one can introduce many concepts, similarly to [1], for linear relations between two
normed linear spaces.
Let X and Y be complex Banach spaces. An arbitrary linear subspace A ⊂ X×Y
is called a linear relation between X and Y . If A is closed in X × Y , then it is called
a closed linear relation.
This research was carried out with the financial support of the Russian Foundation for Basic
Research (grant no. 01-01-00408).
AMS 2000 Mathematics Subject Classification. Primary 47A06, 47A10, 47D06; Secondary
34A60.
1574 A. G. Baskakov and K. I. Chernyshov

The subspace D(A) = {x ∈ X | there exists y ∈ Y such that (x, y) ∈ A},

which is the projection of A onto X, is called the domain of the linear relation
A ⊂ X × Y . We shall denote by Ax, where x ∈ D(A), the set {y ∈ Y | (x, y) ∈ A};
in addition, Ker A = {x ∈ D(A) | (x, 0) ∈ A} is the kernel of the

relation A and
Im A = {y ∈ Y | there exists x ∈ D(A) such that (x, y) ∈ A} = x∈D(A) Ax is its
range, the projection of A onto Y . Note that Ax = y + A0 for each y ∈ Ax.
By the sum of linear relations A, B ⊂ X × Y we mean the linear subspace of
X × Y of the form A + B = {(x, y) ∈ X × Y | x ∈ D(A) ∩ D(B), y ∈ Ax + Bx}.
Hence D(A + B) = D(A) ∩ D(B), and by Ax + Bx we shall mean the algebraic
sum of the sets Ax and Bx.
By the product of linear relations A ⊂ X × Y, B ⊂ Y × Z, where Z is a Banach
space, we mean the linear subspace of X × Z of the form
BA = {(x, z) ∈ X × Z | ∃ y ∈ D(B) ⊂ Y such that (x, y) ∈ A, (y, z) ∈ B}.
By the inverse of a linear relation A ⊂ X × Y we mean the linear relation
A−1 = {(y, x) ∈ Y × X | (x, y) ∈ A} ⊂ Y × X.
Each linear relation A⊂X×Y is the graph of a set-valued map A  : D(A)⊂X→2Y ,
where Ax = Ax ∈ 2Y . We identify them in what follows and use the same symbol A
for both.
We denote the set of closed relations in X×Y by LR(X, Y ); if X = Y , then we set
LR(X) = LR(X, X). Here we regard the set LO(X, Y ) of closed linear operators
from X into Y as a subset of LR(X, Y ) (by identifying an operator and its graph).
Thus, LR(X, Y ) contains the Banach space Hom(X, Y ) of bounded linear operators
(homomorphisms) from X into Y . If X = Y , then LO(X) = LO(X, X) and End X
is the Banach algebra of bounded linear operators (endomorphisms) on X. Hence
End X ⊂ LO(X) ⊂ LR(X).
In what follows, unless otherwise stated we write M ⊂ N to indicate that the
set M is a (not necessarily proper) subset of the set N .
Definition 1.1. A relation A in LR(X, Y ) is said to be injective if Ker A = {0}
and it is surjective if Im A = Y .
Defnition 1.2. A relation A in LR(X, Y ) is said to be continuously invertible if
it is both injective and surjective; in that case A−1 ∈ Hom(Y, X).
In the next definition and in what follows we denote by I the identity operator
in the Banach space under consideration. In the expression ‘a linear relation’ we
shall usually drop ‘linear’.
Definition 1.3. The resolvent set of a relation A ∈ LR(X) is the set ρ(A) of
λ ∈ C such that (A − λI)−1 ∈ End X. The spectrum of A ∈ LR(X) is the set
σ(A) = C \ ρ(A).
The set ρ(A) is open, whereas the spectrum σ(A) of a relation A ∈ LR(X) is
closed.
Definition 1.4. The map
R( · , A) : ρ(A) → End X, R(λ, A) = (A − λI)−1 , λ ∈ ρ(A),
is called the resolvent of the relation A ∈ LR(X).
 Spectral analysis of linear relations and degenerate operator semigroups 1575

The resolvent of a relation A ∈ LR(X) is a pseudoresolvent in the standard

meaning of the term (see, for instance, [4], § 4.8 and Definition 2.1), and we have
Ker R(λ0 , A) = A0 and Im R(λ0 , A) = D(A) for each λ0 ∈ ρ(A). In this connec-
tion, in what follows we use the notation Ker R, Im R in place of Ker R(λ0 , A),
Im R(λ0 , A), respectively.
If B ∈ End X is a quasinilpotent operator, then σ(B −1 ) = ∅ (see § 2). To keep
clear of complications related to the possibly empty spectrum of a relation we shall
use the following definition.
Definition 1.5. The extended spectrum of a relation A ∈ LR(X) is the sub-
(A) of the extended complex plane C
set σ  = C ∪ {∞} that is equal to σ(A) if the
following three conditions are satisfied:
(1) A0 = {0}, that is, A ∈ LO(X);
(2) the resolvent R( · , A) of the operator A can be analytically continued to ∞;
(3) lim|λ|→∞ R(λ, A) = 0.
Otherwise one sets σ (A) = σ(A) ∪ {∞}. The set ρ(A) = C  \σ(A) is called the
extended resolvent set of the relation A.
By Definition 1.5 and Liouville’s theorem, if X = {0}, then σ  (A) = ∅ for each
A ∈ LR(X); furthermore, ∞ ∈ σ (A) if dim A0
1, that is, for A ∈ LR(X)\LO(X).
The definition of the extended spectrum of a linear operator can be found in
the monographs [4] and [5]. Note that the results presented in §§ 2–5 point to a
profoundly more important role of the concept of extended spectrum in the case of
linear relations.
The importance of the investigation of linear relations is shown by the examples
of problems in §§ 1.1–1.8 in which they can occur. We shall combine the presentation
with the introduction of new concepts and definitions.
1.1. If A ∈ LO(X) and Ker A = {0}, then A−1 is a linear relation in LR(X). In
particular, an ordinary differential operator A : C (n) [a, b] ⊂ C[a, b] → C[a, b] of the
kind
(Ax)(t) = x(n) (t) + a1 (t)x(n−1)(t) + · · · + an (t)x(t),
where ak ∈ C[a, b], k = 1, . . . , n, that acts in the Banach space C[a, b] of bounded
continuous complex functions on [a, b] ⊂ R has a finite-dimensional kernel Ker A of
dimension n
1, and therefore A−1 ∈ LR(C[a, b]) is a linear relation. If Ax = x ,
then 
A−1 x = x(t) dt, x ∈ C[a, b].

1.2. Let A ∈ LO(X) be a linear operator with non-dense domain: D(A) = X.

Then its adjoint operator is not defined. Nevertheless, one can define in the natural
fashion (see also [2], § 1.5) the linear relation A∗ ⊂ X ∗ ×X ∗ adjoint to A, where X ∗
is the Banach space dual to X:

A∗ = {(η, ξ) ∈ X ∗ × X ∗ | ξ(y) = η(x) for all (x, y) ∈ A}. (1.1)

Clearly, A∗ 0 = {η ∈ X ∗ | η(x) = 0 for each x ∈ D(A)} = D(A)⊥ . It is

important to note that this definition of the relation A∗ is consistent also for
A ∈ LR(X) \ LO(X).
1576 A. G. Baskakov and K. I. Chernyshov

The first to define the adjoint linear relation was von Neumann [6]; his work
apparently triggered the development of the theory of linear relations.
1.3. Each pseudoresolvent R : U ⊂ C → End X defined on an open subset U of C
is the resolvent of the relation A = (R(λ0 ))−1 + λ0 I, where λ0 ∈ U ; in addition,
ρ(A) ⊃ U and the definition of A is independent of one’s choice of the point λ0
in U (see § 2 for greater detail).
1.4. We say that a sequence of operators An in LO(X) is  convergent if the resol-
vent sets ρ(An ), n
1, have a non-empty intersection n
1 ρ(An ) containing a
connected open set U and for some λ0 ∈ U the sequence R(λ0 , An ), n
1, is
Cauchy with respect to the operator norm in End X. Then for each λ ∈ U there
exists a limit limn→∞ R(λ, An ) = R(λ) ∈ End X and the function R : U → End X is
a pseudoresolvent, although not necessarily the resolvent of an operator in LO(X).
Taking account of § 1.3 we conclude that in the general case the limit A0 of a
sequence of closed operators is a linear relation.
1.5. A strongly continuous semigroup of operators {T (t); t
0} in the algebra
End X is called a degenerate semigroup if T (0) = I.
Let {T (t); t
0} be a degenerate semigroup in End X. Then T (0) = P is a
non-trivial projection and X = X0 ⊕ X1 , where X0 = Im P , X1 = Ker P . Hence
T (t) = T0 (t) ⊕ T1 (t), t
0, where T0 (t) = 0 ∈ End X0 for t
0 and {T1 (t); t
0}
is a C0 -semigroup in End X. Hence there exist constants M1
1 and ω1 ∈ R such
that T1 (t)  M1 exp(ω1 t), t
0. This estimate allows one to consider the Laplace
transform R : {λ ∈ C | Re λ > ω1 } → End X of the semigroup {T (t); t
0}, which
is the resolvent (see § 1.3) of a linear relation A ∈ LR(X) \ LO(X), and we have
Ker R(λ0 ) = X0 if Re λ0 > ω1 . We call the linear relation A the generator of the
degenerate operator semigroup {T (t); t
0}.
Degenerate operator semigroups occur, for instance, in the analysis of sequences
of C0 -semigroups {Tn (t); t
0}, the convergence of which means that the following
conditions are simultaneously satisfied:
(1) Tn (t)  M exp(ωt), t
0, for all n ∈ N and some M
1, ω ∈ R;
(2) limn→∞ R(λ0 , An )x = R(λ0 )x exists for some Re λ0 > ω and all x ∈ X,
where An is the generator of the semigroup {Tn (t); t
0}.
Under these assumptions R is the resolvent of a relation A ∈ LR(X), which is
the generator of a degenerate operator semigroup {T (t); t
0} in End X. A study
of the convergence of degenerate semigroups was carried out in [7]. It must be
pointed out that the theory of linear relations enables one to simplify some proofs.
For instance, Theorem 3.6 in [7] is an easy consequence of § 1.3 (see § 2 for a more
detailed discussion).
1.6. A linear operator A : D(A) ⊂ X → X is said to admit a closed extension
(see [8], Chapter III, § 1.3) if the conditions (1) limn→∞ xn = 0, xn ∈ D(A), and
(2) there exists limn→∞ Axn = y, imply that y = 0. An equivalent definition: a
linear operator A : D(A) ⊂ X → X admits a closed extension if the closure of its
graph Γ(A) = {(x, Ax) ∈ X × X | x ∈ D(A)} is the graph of a linear operator. In
the general case the closure A = Γ(A) of the graph of an operator A is a relation
in LR(X).
 Spectral analysis of linear relations and degenerate operator semigroups 1577

1.7. Let A, B ∈ Hom(X, Y ). Then the function P(λ) = A + λB, λ ∈ C, is called a

linear pencil. Many problems in mathematical physics are known to reduce to the
study of invertibility conditions for the operators P(λ), λ ∈ C. Linear pencils are
also results of certain transformations of polynomial pencils and S. Kreı̌n pencils.
In turn, the investigation of linear pencils often reduces to the study of the spectral
properties of the linear relations B−1 A, AB−1 (see § 6).

1.8. Consider the Cauchy problem

x(0) = x0 ∈ X (1.2)

for the linear differential equation

F ẋ(t) = Gx(t), t ∈ R+ = [0; +∞), (1.3)

containing a pair of closed linear operators from the Banach space X into the
Banach space Y , with Ker F = {0}.
There exist two approaches to the solubility and the construction of solutions
of (1.3). The first is based on the spectral theory of ordered pairs of linear operators
(see, for instance, [9]–[12]). The second is based on the use of a linear differential
inclusion of the following form:

ẏ(t) ∈ Ay(t), t ∈ R+ , y(0) = y0 ∈ D(A), (1.4)

where A ∈ LR(X) is the relation of the form A = F −1 G. This approach is used in [2]
(see also [13] and Remark 7.1). For instance, the construction of solutions in [2] uses
a semigroup of linear operators generated by the relation A. Incidentally, con-
straints on A in [2] do not allow one to consider important classes of differential
inclusions even in a finite-dimensional space X (see Remark 4.3).
In the present paper we consider several questions of the spectral theory of linear
relations scantily discussed in the monographs [1] and [2]. Some of them prove to be
very useful in applications to the theory of differential inclusions of the form (1.4).
In § 2 we present some results on pseudoresolvents that can be obtained with the
use of linear relations and also establish results on the spectral decomposition
for a linear relation and the spectrum of the inverse relation. In § 3 we obtain
algebraic conditions (in terms of stabilization of domains and the images of zero
under powers of a linear relation) ensuring that the point ∞ is isolated in the
extended spectrum of a linear relation. In § 4 we use ergodic theorems to describe
the phase space for the differential inclusion (1.4). Using certain analogues of the
assumptions of the Hille–Phillips–Yosida–Feller– Miyadera theorem [4] we construct
strongly continuous degenerate operator semigroups corresponding to fixed linear
relations. In § 5 we construct analytic degenerate operator semigroups correspond-
ing to sectorial linear relations. We present applications of the spectral theory of
linear relations to the spectral theory of ordered pairs of linear operators in § 6, and
to the problem (1.2)–(1.3) in § 7.
1578 A. G. Baskakov and K. I. Chernyshov

§ 2. On pseudoresolvents, the spectral decomposition of a

linear relation, and the spectrum of the inverse relation
The broadest class of semigroups of bounded operators strongly continuous for
t > 0 in [4] is the class E (see [4], Definition 18.4.1). Fairly complete informa-
tion about semigroups in the class E can be extracted from pseudoresolvents, con-
structed in [4] with the use of the Laplace transformation for semigroups. The
study of pseudoresolvents is given much attention in [4], Chapter 18 (Theorems
5.8.3–5.8.6, 5.9.1–5.9.3) and in several recent papers (see [1], [2], [7], [14], [15]).
Linear relations, particularly their spectral theory, can be of substantial help in the
analysis of pseudoresolvents because pseudoresolvents are the resolvents of linear
relations, which are not considered in [4].
We now describe this approach in greater detail. In the process we address
several well-known results and establish them in a simpler fashion, sometimes with
refinements. We believe that in this way many results on pseudoresolvents look
more natural and more clearly defined.
Definition 2.1. A function R : Ω ⊂ C → End X is called a pseudoresolvent if the
following equality (Hilbert’s identity) holds for all λ1 , λ2 ∈ Ω:

R(λ1 ) − R(λ2 ) = (λ1 − λ2 )R(λ1 )R(λ2 ).

Note that it is possible in Definition 2.1 that Ω = {λ0 } is a singleton and in that
case R(λ0 ) can be an arbitrary endomorphism of the algebra End X.
Definition 2.2. A pseudoresolvent Rmax : Ωmax ⊂ C → End X is called a max-
imal extension of a pseudoresolvent R : Ω ⊂ C → End X if it is an extension of
every other extension R. We say that such a pseudoresolvent is maximal. The set
Sing R = C/Ωmax is called the singular set of the pseudoresolvent R.
Definition 2.3. Let Q ∈ End X, λ0 ∈ C, and let R : Ω ⊂ C → End X be a
pseudoresolvent. We say that the operator Q is embedded in R at a point λ0 if
λ0 ∈ Ω and Q = R(λ0 ).
It is an immediate consequence of Definitions 2.1–2.3 that the question of the
embedding of an operator in the range of a pseudoresolvent can be reduced to
the question of the construction of maximal extensions of pseudoresolvents.
Theorem 2.1. Each pseudoresolvent R : Ω ⊂ C → End X has a unique maximal
extension. It is the resolvent of some linear relation A and Sing R = σ(A). In
particular, if Q ∈ End X and λ0 ∈ C, then there exists a unique maximal pseudo-
resolvent R0 : Ω ⊂ C → End X such that λ0 ∈ Ω and Q is embedded in R0 at the
point λ0 .
Proof. Consider two points λ1 , λ2 ∈ Ω and two linear relations Ak = (R(λk ))−1 +λk I,
k = 1, 2, with the same domain D(Ak ) = Im R = Im R(λk ), k = 1, 2. Since
Ak = {(R(λk )y, y + λk R(λk )y) | y ∈ X}, it follows by Hilbert’s identity that
A1 = A2 = A. Hence R(λ) = R(λ, A) for each λ ∈ Ω and it follows from the
above that R( · , A) : ρ(A) → End X is the unique maximal extension of R. Hence
Sing R = C \ ρ(A) = σ(A).
 Spectral analysis of linear relations and degenerate operator semigroups 1579

The result of Theorem 2.1 on the existence of a maximal extension was obtained
in [4] (Theorem 5.8.6); but it required greater effort. The result of the embedding
of a bounded operator in a pseudoresolvent was established in [7], Theorem 3.6. In
view of the estimate

R(λ, A)
r(R(λ, A))
(dist(λ, σ(A)))−1 for each λ ∈ ρ(A)

(see Corollary 2.1 to Theorem 2.4), Theorem 2.1 covers also Proposition 3.5 in [7]
on the growth of the norm of a pseudoresolvent as one approaches Sing R.
The concept of singular set of a pseudoresolvent has been introduced in [15]. In
view of Theorem 2.1, it is the spectrum of the linear relation the resolvent of which
extends the pseudoresolvent in question. This is very important for the analysis of
pseudoresolvents on the basis of the spectral theory of linear relations.
The next result refines Assertion 5.8.4 in [4] and is an easy consequence of Theo-
rem 2.1. Under its assumptions we denote by Sp A the spectrum of a commutative
Banach algebra A with unity, that is, Sp A is a compact topological algebra of non-
trivial complex homomorphisms of the algebra A,  a : Sp A → C, and  a(χ) = χ(a),
χ ∈ Sp A, is the Gel’fand transform of the element a of Sp A (see [16]).
Theorem 2.2. Let R : Ω ⊂ C → End X be a maximal pseudoresolvent and let A be
the minimal closed subalgebra of the Banach algebra End X containing all the oper-
ators R(λ), λ ∈ Ω, and the identity operator I. Then the spectrum Sp A is homeo-
morphic to the extended spectrum σ  (A) of the linear relation A ∈ LR(X) whose
resolvent is R : Ω = ρ(A) → End X. Moreover, there exists a homeomorphism
α : Sp A → σ (A) such that

 1
R(λ, A)(χ) = , χ ∈ Sp A, λ ∈ ρ(A).
λ − α(χ)

In addition, α(χ∞ ) = ∞ for the character χ∞ ∈ Sp A defined by the following

conditions: R(λ, A) ∈ Ker χ∞ for each λ ∈ ρ(A) and χ∞ (I) = 1.
Thanks to Theorem 2.1 the definitions below are consistent and the results proved
on their basis are well established.
Definition 2.4. A closed linear subspace X0 of X is said to be invariant with
respect to a relation A ∈ LR(X) with non-empty ρ(A) if X0 is invariant with respect
to all operators R(λ, A), λ ∈ ρ(A). By the restriction of the relation A ∈ LR(X)
to the subspace X0 we shall mean the relation A0 ∈ LR(X) whose resolvent is the
restriction R0 : ρ(A) → End X0 , R0 (λ) = R(λ, A)  X0 , λ ∈ ρ(A), of the resolvent
R( · , A) : ρ(A) → End X to X0 ; we denote it by A0 = A  X0 .
Definition 2.5. Let
X = X0 ⊕ X1 (2.1)
be a direct sum of subspaces invariant with respect to A∈LR(X) and let A0 =A  X0 ,
A1 = A  X1 . Then we say that the relation A is the direct sum of the relations
A0 and A1 , and we write
A = A0 ⊕ A1 . (2.2)
1580 A. G. Baskakov and K. I. Chernyshov

In addition, A0 = A0 0 ⊕ A1 0, and equalities (2.1) and (2.2) mean that for each
x ∈ D(A) the set Ax is defined by the formula

Ax = A0 x0 + A1 x1 , x = x0 + x1 ,

where xi ∈ D(Ai ) ⊂ Xi , i = 0, 1, and Ax is the algebraic sum of the sets A0 x0

and A1 x1 .
Lemma 2.1. If a relation A ∈ LR(X) satisfies equalities (2.1) and (2.2), then
(A) = σ
σ (A0 ) ∪ σ
(A1 ), where Ai is the restriction of A to Xi , i = 0, 1.
Proof. Definitions 2.4 and 2.5 immediately yield the equality

R(λ, A) = R(λ, A0 ) ⊕ R(λ, A1 ), λ ∈ ρ(A), (2.3)

which ensures the required result.

Lemma 2.2. Let A ∈ LR(X). Then ∞ ∈
/σ(A) ⇔ A ∈ End X.
Proof. If A ∈ End X, then obviously, ∞ ∈  (A).
/σ
Conversely, let ∞ ∈ (A). Then by Definition 1.5 the relation A is a closed
/ σ
linear operator with resolvent satisfying the condition lim|λ|→∞ R(λ, A) = 0, and
(A) = σ(A) = σ0 is a compact subset of C. Hence λ → (A−λI)−1 is a holomorphic
σ
function outside a  sufficiently large disc and can be represented as a convergent
∞
series R(λ, A) = − n=0 Kn λ−n−1 , where Kn ∈ End X, n
0. Hence there exists
a limit lim|λ|→∞(−λR(λ, A)) = K0 . The ergodic Theorem 18.8.3 in [4] shows that
K0 = P0 is a projection onto X0 = D(A) = D(A).
We claim that P0 = I. Note that, as follows from the above, the restriction
A0 = A  X0 satisfies the equality lim|λ|→∞(−λR(λ, A0 )) = I0 , where I0 is the iden-
tity operator in X0 . Hence the R(λ, A0 ) are invertible operators in End X0 for suffi-
ciently large |λ|. If X = X0 , then for x ∈ X\X0 we set y0 = R(λ, A)x ∈ D(A) = X0 .
There exists x0 ∈ X such that R(λ, A0 )x0 = y0 . Since R(λ, A)(x0 − x) = 0, it fol-
lows that x0 = x ∈ X0 . We have arrived at a contradiction, which yields the
equality D(A) = X0 = X.
Note that in [1] the condition ∞ ∈ (A) for a relation A ∈ LR(X) means by
/ σ
/ σ(A−1 ).
definition that 0 ∈
Theorem 2.3. Let A ∈ LR(X) be a relation with extended spectrum σ
(A) repre-
sentable in the following form:

(A) = σ0 ∪ σ1 ,
σ (2.4)

 and σ0 ∩ σ1 = ∅. Then
where σ0 is a compact subset of C, σ1 a closed subset of C,
there exist decompositions (2.1) and (2.2) in which the A-invariant closed subspaces
X0 , X1 and the restrictions A0 = A  X0 , A1 = A  X1 of the relation A have the
following properties:
(1) A0 ∈ End X0 , σ(A0 ) = σ(A0 ) = σ0 ;
(2) A1 0 = A0 = Ker R( · , A) = Ker R( · , A1) ⊂ X1 , D(A) = X0 ⊕ D(A1 ),
(A1 ) = σ1 .
σ
 Spectral analysis of linear relations and degenerate operator semigroups 1581

Proof. Let γ be a closed Jordan curve (or several such curves) lying in ρ(A) so
that σ0 lies in its interior and σ1 lies outside it. We consider the Riesz projection

1
P0 = − R(λ, A) dλ ∈ End X, (2.5)
2πi γ

which defines the decomposition of X into a direct sum of subspaces:

X = X0 ⊕ X1 , X0 = Im P0 , X1 = Ker P0 .

Definition 2.4 compels us to verify that R(µ, A)Xi ⊂ Xi , µ ∈ ρ(A), i = 0, 1. To

this end it is sufficient to demonstrate that R(µ, A)P0 = P0 R(µ, A) for all µ ∈ ρ(A),
which follows in effect from (2.5) and the commutativity of the range of the resolvent
of the relation A.
Consider now the restrictions R0 and R1 of the resolvent R( · , A) of A to X0
and X1 , respectively. If x0 ∈ Ker R, then it is an immediate consequence of (2.5)
that P0 x0 = 0 and therefore Ker R ⊂ X1 . Hence R0 is a resolvent of an operator
A0 ∈ LO(X).
The functions R0 and R1 are defined on ρ(A). We claim that R0 can be ana-
lytically continued to C \ σ0 and lim|λ|→∞ R(λ, A0 ) = 0, so that A0 ∈ End X by
Lemma 2.2. Using Hilbert’s identity we represent the function R0 : ρ(A) → End X0
in the following form:

1
R0 (µ)x0 = R(µ, A0 )x0 = R(µ, A)P0 x0 = − R(µ, A)R(λ, A)x0 dλ
2πi γ

1 R(µ, A) − R(λ, A)
=− x0 dλ.
2πi γ µ−λ

dλ
If µ lies outside γ, then = 0. Since γ is an arbitrary curve with the
γ µ−λ
above-prescribed properties, it follows that the function R0 ( · , A) admits an analytic
 \ σ0 ; moreover, it follows from the above equality
continuation to the set C

1
R0 (µ)x0 = R(µ, A0 )x0 = (µ − λ)−1 R(λ, A)x0 dλ
2πi γ

that lim|µ|→∞ R(µ, A0 ) = 0. It remains to apply Lemma 2.2, which shows that
A0 ∈ End X0 .
Let A1 ⊂ LR(X1 ) be a relation with resolvent R1 . We can prove in a similar
way that the function R( · , A1 ) = R1 : ρ(A) → End X1 has an analytic continuation
 \ σ1 . Hence σ
into the set C 1 ⊂ σ1 , and by Lemma 2.1 we arrive at the equalities
(A0 ) = σ(A0 ) = σ0 and σ
σ (A1 ) = σ1 . The other properties of A1 are immediate
consequences of the inclusion A0 ∈ End X0 .
Theorem 2.4. If A ∈ LR(X), then the extended spectrum σ (A−1 ) of the relation
−1
A ∈ LR(X) inverse to A can be represented in the following form:

(A−1 ) = {λ−1 | λ ∈ σ
σ (A)}.
1582 A. G. Baskakov and K. I. Chernyshov

Proof. For 0 = λ ∈ C the relations A − λI and A−1 − λ−1 I can be represented in

the following form:
A − λI = {(x, y − λx) | (x, y) ∈ A}, A−1 − λ−1 I = {(y, x − λ−1 y) | (x, y) ∈ A}.
We prove first that for λ = 0 one of these relations is injective if and only if so
is the other. If 0 = x0 ∈ Ker(A − λI), then there exists y0 such that y0 − λx0 = 0,
that is, y0 = λx0 = 0, and therefore (y0 , x0 − λ−1 y0 ) = (y0 , 0) ∈ A−1 − λ−1 I. The
converse also holds because the relation A − λI can be obtained from A−1 − λ−1 I
by means of the same procedure as A−1 − λ−1 I is obtained from A − λI.
Next, it is an immediate consequence of the definitions of the relations A − λI
and A−1 − λ−1 I, λ = 0, that their ranges have the form {y − λx | (x, y) ∈ A} and
therefore are the same. Hence if one of them is surjective, then so is the other.
It follows from the above that the linear relations A − λI and A−1 − λ−1 I are
simultaneously invertible for λ = 0. Thus, the sets σ(A) \ {0} and σ(A−1 ) \ {0}
are taken into each other by the map λ → 1/λ.
Consider now the case λ = 0. Assume that 0 ∈ / σ(A). Then A−1 ∈ End X
and in accordance with the definition of the extended spectrum one must con-
sider its resolvent R(λ, A−1 ) = (A−1 − λI)−1 in the neighbourhood of ∞. Since
lim|λ|→∞ R(λ, A−1 ) = 0, it follows that ∞ ∈ /σ (A−1 ) by Definition 1.5. Conversely,
assume that ∞ ∈ /σ (A−1 ). Then it follows by Lemma 2.2 that A−1 ∈ End X. Hence
0∈ / σ(A) ⊂ σ  (A).
The same arguments can be applied to A−1 , which as a result yields the required
equality.
Corollary 2.1. If A ∈ LR(X) and µ ∈ ρ(A), then
σ(R(µ, A)) = {(µ − λ)−1 | λ ∈ σ
(A)}.
The result of this corollary was proved in [1], Theorem V.4.2. Note that the
definition of the spectrum of a linear relation is due to Favini and Yagi [17]. How-
ever, neither there nor in the monograph [2] did they use the concept of extended
spectrum of a linear relation. The definition of the extended spectrum for linear
relations on normed spaces was introduced by Cross [1], but little used in fact.
Corollary 2.2. The following conditions are equivalent for A ∈ LR(X):
(1) A ∈ End X; (2) ∞∈
/σ(A); (3) / σ(A−1 ).
0∈
Corollary 2.3. For A ∈ LR(X) the equality σ
(A) = {∞} holds if and only if
A−1 ∈ End X is a quasinilpotent operator.
Corollary 2.4. If λ ∈ ρ(A), λ = 0, then the following equalities hold:
(A−1 − λ−1 I)−1 = −λI − λ2 (A − λI)−1 , (2.6)
−1 −1 −2 −1 −1 −1
(A − λI) = −λ I −λ (A −λ I) . (2.7)
Proof. It follows from the description of the range of the relation A − λI given in
the proof of Theorem 2.4 that the operator (A − λI)−1 ∈ End X is constructed
as follows: each vector z ∈ X can be represented in the form z = y − λx, where
(x, y) ∈ A; then (A − λI)−1 z = x. Since x − (1/λ)y = −(1/λ)z, it follows that
(A−1 −λ−1 I)−1 z = −λy = −λ(z+λx) = −λz−λ2 (A−λI)−1 z. Hence equality (2.6)
holds. Equality (2.7) follows from (2.6).
 Spectral analysis of linear relations and degenerate operator semigroups 1583

Corollary 2.5. If A = B −1 , where B is a quasinilpotent operator in End X, then

(A) = {∞} and (A − λI)−1 = B + λB 2 + λ2 B 3 + · · · . In particular, R( · , A) is a
σ
polynomial if B is a nilpotent operator in End X.
Theorem 2.4 and relations (2.6), (2.7) allow us to say that, as concerns the
extended spectrum of a linear relation, all points in the extended complex plane C, 
including ∞, ‘have equal rights’ in a certain sense. On the other hand if the point ∞
belongs to the extended spectrum of a closed linear operator and is isolated there,
then it can only be an essential singularity of its resolvent (see [4], Theorem 5.9.4).
Let ∆ be an open subset of the extended complex plane C  containing the
extended spectrum σ (A) of a relation A. We denote by F(∆) the algebra of ana-
lytic complex functions on ∆. Let γ be a closed Jordan  curve enclosing σ (A) and
let f ∈ F(∆) be a function such that the integral f(λ)R(λ, A) dλ converges
γ
absolutely. Then the formula

1
f(A) = δf(∞)I − f(λ)R(λ, A) dλ, (2.8)
2πi γ

where δ = 1 or δ = 0 depending on whether λ = ∞ lies inside or outside γ, defines a

bounded operator in the algebra End X. Moreover, it belongs to the commutative
subalgebra A introduced before Theorem 2.2. This allows us to prove the following
result.
Theorem 2.5 (on maps of the spectrum). The following equality holds for opera-
tors A ∈ LR(X): σ(f(A)) = f(
σ (A)) = {f(λ) | λ ∈ σ
(A)}.
Proof. Since the spectrum of the operator f(A) is equal to the range of its Gel’fand
transform, the required representation is a consequence of the following equalities,
in which we use Cauchy’s theorem, Theorem 2.2, and formula (2.8):
σ(f(A)) = {χ(f(A)) | χ ∈ Sp A}
  
1
= δf(∞) − f(λ)χ(R(λ, A)) dλ | χ ∈ Sp A
2πi γ
  
1 f(λ)
= δf(∞) − dξ | ξ ∈ σ
(A) = f(σ (A)).
2πi γ λ − ξ

Theorem 2.5 is an analogue of Theorem 5.12.1 in [4] established for closed linear
operators. The above proof is simpler thanks to Theorem 2.2, which is in its turn
a refined analogue of Theorem 5.8.4 in [4].

§ 3. Conditions for compactness of the spectrum of linear relations

Throughout what follows we assume that the linear relations under consideration
are non-singular in the sense of the following condition.
Assumption 3.1. The resolvent set ρ(A) of the relation A ∈ LR(X) is non-empty.
The following result is an immediate consequence of the definition of the inverse
relation and the properties of linear relations stated in § 2 (see also the properties
of relations listed in § 6).
1584 A. G. Baskakov and K. I. Chernyshov

Lemma 3.1. Let A ∈ LR(X). Then the following equalities hold for arbitrary
λ0 ∈ ρ(A):

Ker(R(λ0 , A))k = Ak 0, Im(R(λ0 , A))k = D(Ak ), k ∈ N.

This lemma ensures the legitimacy of the notation Ker Rk and Im Rk , k ∈ N, for
powers of the resolvent of a relation A ∈ LR(X).
Definition 3.1. We say that a relation A ∈ LR(X) has the property of stabiliza-
tion of powers at infinity if there exists m ∈ N such that

Am−1 0 ⊂ Am 0 = Am+1 0, D(Am−1 ) ⊃ D(Am ) = D(Am+1 ), (3.1)

where all inclusions are strict. The integer m is called the order of stabilization.
Note that for m = 1 we set {0} ⊂ A0 = A2 0 and X ⊃ D(A) = D(A2 ).
Assumption 3.2. The relations A ∈ LR(X) have the property of stabilization of
powers at infinity of order m.
Theorem 3.1. Let m
2 be an integer. Then the following conditions are equi-
valent for a linear relation A ∈ LR(X) \ LO(X):
(1) the point ∞ is a pole of order m − 2 of the function R( · , A) for m
3 or
a removable singularity of R( · , A) for m = 2;
(2) the Banach space X can be represented as the direct sum X = X0 ⊕ X∞
of the A-invariant closed subspaces X0 = D(Am ), X∞ = Am 0, the restric-
tion A0 of A to X0 belongs to End X0 , σ(A0 ) = σ(A), and in addition,
(A∞ ) = {∞}, A−1
σ ∞ ∈ End X∞ , (A∞ )
−1 m
= 0 for the restriction A∞ of the
−1 m−1
relation A to X∞ , and (A∞ ) = 0;
(3) the assumptions of Proposition 3.2 hold.

Proof. (1)⇒(2). By assumption, ∞ is an isolated point of σ (A), therefore we have

(A) = σ0 ∪ {∞}, where σ0 = σ(A) is a compact subset of C. Setting σ1 = {∞}
σ
we obtain the representation (2.4), after which, applying Theorem 2.3 we arrive at
the decompositions

X = X0 ⊕ X∞ , A = A0 ⊕ A∞ , (3.2)

where X∞ = X1 , A∞ = A1 . It also follows from Theorem 2.3 that σ(A0 ) = σ0 ,

A0 ∈ End X0 , and σ(A∞ ) = {∞}, and by Corollary 2.3 the operator A−1∞ ∈ End X∞
is quasinilpotent. Since σ(A∞ ) = {∞}, it follows by Corollaries 2.4 and 2.5 that
the resolvent of the relation A∞ can be represented as a series:
∞

R(λ, A∞ ) = λi (A−1
∞ )
i+1

i=0

in a neighbourhood of ∞. By assumption it has a pole of order m − 2 at this point,

therefore (A−1
∞ )
m
= 0 and (A−1
∞ )
m−1
= 0. Note that R(λ, A∞ ) = A−1 ∞ for m = 2,
that is, ∞ is a removable singularity of the function R( · , A). All the above ensures
the properties required in part (2).
 Spectral analysis of linear relations and degenerate operator semigroups 1585

(2)⇒(3). It follows from the decompositions (3.2) with the above-indicated

properties that
Ak 0 = 0 ⊕ Ak∞ 0, D(Ak ) = X0 ⊕ D(Ak∞ ), k ∈ N.
Since (R(0, A∞ )) =m
(A−1
∞ )
m
= 0 and (R(0, A∞)) = (A−1
m−1
∞ )
m−1
= 0, prop-
erty (3) is a consequence of Lemma 1. Note that the relations A and A∞ possess
the property of stabilization of powers at infinity of order m simultaneously.
(3)⇒(1). It follows from Assumption 3.2 and Lemma 3.1 that
Ker B m−1 ⊂ Ker B m = Ker B m+1 , Im B m−1 ⊃ Im B m = Im B m+1 ,
where B = R(λ0 , A) ∈ End X, λ0 ∈ ρ(A). We set V = B m ∈ End X; then
Ker V = Ker V 2 , Im V = Im V 2 . We must prove that
X = X0 ⊕ X∞ , (3.3)
where X∞ = Ker V and X0 = Im V . Note that X∞ and X0 are V -invariant
subspaces and the restriction V0 : X0 → X0 of V0 to X0 is an invertible operator.
The proof of the decomposition (3.3) can be found, for instance, in [19], Chap-
ter 6.
We claim that 0 is an isolated point in σ(V ) and X0 is a closed subspace of X.
To this end we consider the quotient space X = X/X∞ and the factor operator

V ∈ End X,
 V x
 = V
x,
where y = y+X∞ is the coset containing the vector y ∈ X. Since X/X∞ = X0 /X∞ ,
it follows that V is an invertible operator in End V . Hence 0 ∈/ σ0 = σ(V ). Assume
that 0 = µ0 ∈ / σ0 . We shall prove that µ0 ∈ / σ(V ). If x0 ∈ Ker(V − µ0 I), then
(V − µ0 I)x0 = 0, and therefore (V − µ0 I)
x0 =  0 ∈ X, where x
0 = x0 + X∞ .
The operator V − µ0 I is invertible, therefore x
0 = 0, that is, x0 ∈ X∞ . The
restriction V∞ of the operator V to X∞ is zero and µ0 = 0, therefore x0 = 0.
Hence Ker(V − µ0 I) = {0}.
We now prove that the operator V −µ0 I is surjective. Let y be an arbitrary vector
in X. Since the operator V − µ0 I is invertible, there exists a coset x  = x + X∞

such that (V − µ0 I) x = y, that is, y0 = (V − µ0 I)x − y ∈ X∞ . Since V y0 = 0, it
follows that (V − µ0 I)y0 = y0 for y0 = −µ−1 
0 y0 . Hence we have (V − µ0 I)(x − y0 ) =
y + y0 − y0 = y and V − µ0 I is surjective.
By the above σ(V ) ⊂ {0} ∪ σ1 and moreover, 0 ∈ / σ1 . Since V = (R(λ0 , A))m ,
it follows that 0 is an isolated point in σ(R(λ0 , A)). It is now clear that the sub-
space X0 in the decomposition (3.3) is the range of the Riesz projection constructed
for the operator R(λ0 , A) and the spectral set σ(R(λ0 , A)) \ {0}, and therefore it is
closed.
As already pointed out, the subspaces X∞ and X0 are independent of one’s choice
of λ0 ∈ ρ(A) and are equal to the closed subspaces Am 0 and D(Am ), respectively.
From Corollary 2.1 to Theorem 2.1 we see that {∞} is an isolated point in σ  (A),
∞ ∈ / σ(A  X0 ), and σ (A  X∞ ) = {∞}, that is, A  X0 ∈ End X0 , where
(R(λ, A))m−1 = 0, and (R(λ, A))m = 0 for each λ ∈ ρ(A). Hence condition (1)
holds.
We now define eigenvectors and associated vectors of linear relations correspond-
ing to the point ∞. Here we shall take Theorems 2.4 and 3.1 into account.
1586 A. G. Baskakov and K. I. Chernyshov

Definition 3.2. Each non-zero vector x0 in A0 ⊂ X is called an eigenvector of the

relation A ∈ LR(X) corresponding to the point ∞. A vector x1 ∈ X is called a root
vector of the relation A corresponding to ∞ if there exists k ∈ N such that x1 ∈ Ak 0.
A positive integer k is called the height of the root vector x1 if x1 ∈ Ak 0 \ Ak−1 0.
Definition 3.2 immediately shows that the closed subspace Ak 0 = Ker(A−1 )k
consists of the root vectors of the relation A corresponding to the point ∞ and of
height at most k.
Definition 3.3. We say that a relation A ∈ LR(X) possesses a finite Jordan chain
x0 , x1 , . . . , xk−1 of length k corresponding to the point ∞ if x0 is an eigenvector for A
corresponding to ∞ and the xi, 2  i  k − 1, are root vectors corresponding to
the same point such that

x0 ∈ A0, xi ∈ Axi−1 , 1  i  k − 1, xk−1 ∈

/ Ak 0

(so that the vector xi, 0  i  k − 1, has height i). The vectors x1 , . . . , xk−1 are
said to be associated with the eigenvector x0 .
Lemma 3.2. A relation A ∈ LR(X) possesses a finite Jordan chain x0 , . . . , xk−1
of length k corresponding to the point ∞ if and only if the relations

R(λ0 , A)x0 = 0, x0 ∈ A0, R(λ0 , A)xi = xi−1 , 0  i  k − 1,

hold for some (and therefore for all ) λ0 ∈ ρ(A) and there exists no vector x such
that R(λ0 , A)x = xk−1 .
Definition 3.4. We say that a relation A ∈ LR(X) is Fredholm at infinity if D(A)
is a closed subspace of X and, in addition, A0 and X/D(A) are finite-dimensional
linear spaces. The integer ind A = dim A0 − dim(X/D(A) is called the index of the
Fredholm relation A corresponding to the point ∞.
It is an immediate consequence of Definition 3.4 that the relation A is Fred-
holm at infinity if and only if the operator R(λ0 , A), λ0 ∈ ρ(A), is Fredholm; the
corresponding indices are equal.
The next result follows by Theorem 3.1, Definitions 3.2–3.4, and Lemma 3.2 and
explains the concepts involved in their statements. In addition, in the proof of the
equivalence of properties (2) and (3) one must bear in mind that if the index of a
relation A is zero, then so are the indices of all its powers.
Theorem 3.2. Let m ∈ N and let A ∈ LR(X) \ LO(X) be a relation that is
Fredholm at infinity with index zero. Then the following conditions are equivalent:
(1) all Jordan chains of the relations A corresponding to ∞ have lengths at
most m ∈ N and there exists a Jordan chain of length m;
(2) Am−1 0 ⊂ Am 0 = Am+1 0;
(3) D(Am−1 ) ⊃ D(Am ) = D(Am+1 );
(4) ∞ is a pole of order m − 2 of the resolvent R(λ, A) (λ ∈ ρ(A)) of the
relation A if m
3, and it is a removable singular point of it for m = 2.
 Spectral analysis of linear relations and degenerate operator semigroups 1587

Corollary 3.1. If A ∈ LR(X) \ LO(X), where X is a finite-dimensional space,

 (A) contains finitely many points and their number does not exceed the dimen-
then σ
sion n = dim X of the space X; moreover, the following spectral decomposition holds
for the relation A:

m
A= λi Pi + Q + A∞ , m + 1  n,
i=1

where the Pi ∈ End X are the Riesz projections corresponding to the one-point
 (A∞ ) = {∞}, and Q and A−1
sets {λi }, 1  i  m, σ(A) = {λ1 , . . . λm }, σ ∞ are
nilpotent operators in the algebra End X commuting with each other and with the
projections Pi, 1  i  m.
Remark 3.1. Corollary 3.1, along with other results of §§ 3–5, holds under Assump-
tion 3.1. If this assumption fails, then it is possible that σ (A) = C even for
1  dim X < ∞. This occurs, for instance, for the relation A = X × X.

§ 4. Several analogues of the Hille–Phillips–Yosida–Feller–Miyadera

conditions for linear relations
Let A be a relation in LR(X). We consider the Cauchy problem for the differ-
ential inclusion

ẋ(t) ∈ Ax(t), t ∈ R+ = [0; +∞), (4.1)

x(0) = x0 ∈ D(A). (4.2)

We call a differentiable function x : R+ → X such that x(0) = x0 , x(t) ∈ D(A)

for all t
0, a solution of the Cauchy problem (4.1), (4.2) if it satisfies the inclu-
sion (4.1).
Definition 4.1. We call the closure in X of the set of initial conditions of the
form (4.2) such that the problem (4.1), (4.2) is soluble the phase space of
the differential inclusion (4.1). We denote it by Φ(A).
Remark 4.1. In homogeneous linear differential equations, similarly to semigroups
of differential operators, setting x(t) = exp(λ0 t)y(t), where λ0 ∈ ρ(A), trans-
forms (4.1) into a differential inclusion defined by the continuously invertible rela-
tion A − λ0 I. Hence in what follows we shall assume without loss of generality that
0 ∈ ρ(A), that is, A−1 ∈ End X.
Remark 4.2. If x : R+ → X is a solution of the differential inclusion (4.1), B = A−1 ,
and 0 ∈ ρ(A), then x : R+ → X is a solution of the differential equation (see
Lemma 7.1)
Bẋ = x. (4.3)
In this section we consider linear relations for which ∞ is not necessarily an iso-
lated point in the spectrum. Using ergodic theorems we construct certain subspaces
containing the phase space of a differential inclusion and after that, with the help of
analogues for linear relations of the conditions of the Hille–Phillips–Yosida–Feller–
Miyadera theorem, construct degenerate semigroups of bounded linear operators.
As before, Assumption 3.1 holds throughout.
1588 A. G. Baskakov and K. I. Chernyshov

Definition 4.2. Let m ∈ N. We say that the mth power of the resolvent of a
relation A ∈ LR(X) has the property of minimal growth at infinity if there exists
a sequence {λn } ⊂ ρ(A) such that

(1) lim |λn | = ∞; (2) sup{|λm

n | (R(λn , A)) } < ∞.
m
(4.4)
n→∞ n
1

Assumption 4.1. The resolvent of the relation A ∈ LR(X) has properties (4.4) from
Definition 4.2.
Theorem 4.1. If Assumption 3.2 holds for a relation A ∈ LR(X), then Assump-
tion 4.1 also holds.
Proof. We use the notation of Theorem 3.1 and its result on the equivalence of three
conditions. We see that condition (2) holds. Hence R(λ; A) = R(λ; A∞) ⊕ R(λ; A0 )
if |λ|
r0 > 0 for sufficiently large r0 . Since A0 ∈ End X0 , lim|λ|→∞ R(λ, A0 ) = 0,
and (A−1∞ )
m
= 0, it follows by the representation of the resolvent R( · , A∞) of the
relation A∞ obtained in the proof of Theorem 3.1 that (R(λ; A∞ ))m = 0 for all
λ ∈ C. Thus, (R(λ; A))m = (R(λ; A0 ))m ⊕ 0 for all λ ∈ C, and Assumption 4.1
holds therefore for each sequence {λn } ⊂ ρ(A) such that limn→∞ |λn | = ∞.
Lemma 4.1. If Assumption 4.1 holds, then the lengths of all Jordan chains of the
relation A ∈ LR(X) corresponding to the point ∞ are at most m and they all lie
in X∞ = Am 0.
Proof. Assume the contrary and consider an eigenvector x0 of A ∈ LR(X) corre-
sponding to λ = ∞ for which there exists a Jordan chain of length k
m + 1. We
consider the k-dimensional subspace X ∞ that is the linear span of x0 and all its

associated vectors. Let A∞ be the restriction of the relation A ∈ LR(X) to X ∞ ;
 ∞ ) = {∞}. By Lemma 3.2 we obtain (R(λn , A))m xm = x0 = {0},
 (A
note that σ
n
1. Hence limn→∞ |λm n | (R(λn , A)) xm  = ∞ as |λn | → ∞.
m

Under Assumption 4.1 we consider a bounded sequence of operators in End X

of the following form:

An = I − (−λn R(λn , A))m , n ∈ N, (4.5)

and the closed subspace

 = {x ∈ X : ∃ lim An x}.
X
n→∞

For the construction of the phase space Φ(A) of the differential inclusion (4.1) we
shall apply ergodic theorems of [20] to the sequence An . First, we present several
concepts and results from [20] used here (not in their most general form).
Let A be the minimal closed subalgebra of the Banach algebra End X containing
all operators R(λ; A), λ ∈ ρ(A), and the identity operator I. Then A is a commu-
tative Banach algebra with unity and the sequence (An ) lies in A. Let m ∈ N. We
consider the minimal closed ideal J = Jm of A containing the operators (R(λ; A))m ,
λ ∈ ρ(A).
 Spectral analysis of linear relations and degenerate operator semigroups 1589

Definition 4.3. We call a bounded sequence of linear operators (An ) in A a

J-sequence if it satisfies the following two conditions:
(1) limn→∞ An F  = 0 for each F ∈ J;
(2) An x − x ∈ Jx = {F x; F ∈ J} for each x ∈ X.
Let (An ) be a J-sequence. We denote by Erg(X, (An )) the (closed) subspace
Erg(X, (An )) = {x ∈ X : there exists lim An x};
n→∞

we call it the ergodic subspace corresponding to the J-sequence (An ).

Lemma 4.2. Under Assumption 4.1 the sequence (An ) is a J-sequence and there-
 = Erg(X, (An )).
fore X
Proof. Since the sequence (An ) is bounded, for the verification of condition (1) in
Definition 4.2 it is sufficient to prove that limn→∞ An (R(µ, A))m  = 0 for each
fixed integer µ in ρ(A). This follows from the inequalities
An (R(µ, A))m   const(µ)|λ−1
n |, n
1,
which, in their turn, follow (by the repeated use of Hilbert’s identity) from the
estimates
(R(λn , A))k (R(µ, A))m−k   const |λn |−k , 1  k  m, n
1.
Condition (2) in Definition 4.2 obviously holds.
Since the subspace X  is invariant with respect to all the operators R(λ, A),
λ ∈ ρ(A), it is invariant with respect to the relation A, and therefore one can
consider the restriction A of A to X (see Definition 2.4).
The next result is a consequence of Lemma 1 in [20] and is, in fact, a specialization
of it.
Theorem 4.2. Under Assumption 4.1 the subspace X  is the direct sum
 = X0 ⊕ X∞
X (4.6)
of two closed A-invariant subspaces X0 and X∞ , where X0 = D(Am ), X∞ = Am 0,
 ∈ LR(X),
and the corresponding decomposition of the relation A 
 = A0 ⊕ A∞ ,
A (4.7)
has the following properties: σ(A∞ ) = {∞}, (A−1
∞ )
m
= 0, A0 : D(A0 ) ⊂ X0 → X0
 = σ(A), and D(Am ), the
is a closed linear operator with spectrum σ(A0 ) = σ(A) 0
domain of the operator Am 0 , is dense in X0 .
 is performed by the projection P0
Remark 4.3. The decomposition (4.4) of X
defined by the following relations:

P0 x = lim (I − An )x = lim (−λn R(λn , A))m x, 

x ∈ X,
n→∞ n→∞
Im P0 = X0 , Ker P0 = X∞ ,
and independent of one’s choice of a sequence (λn ) in ρ(A) satisfying Assumption 4.1
(see [20], Lemma 2). Note also that D(Ak0 ) is dense in X0 for each k ∈ N (see [20]).
1590 A. G. Baskakov and K. I. Chernyshov

Corollary 4.2. A linear relation A ∈ LR(X) satisfying Assumption 4.1 and such
that the domain D(Am ) of the operator Am is dense in X is a linear operator.
Corollary 4.3. Under Assumption 4.1 the subspace Am 0 is finite-dimensional
if D(Am ) has a finite codimension in X.
The next result is a consequence of Theorem 1 in [20].
Theorem 4.3. If Assumption 4.1 holds, then X  = X if and only if vectors in the
subspace Am 0 separate functionals in the subspace (A∗ )m 0 of the Banach space X ∗
dual to X (A∗ ⊂ X ∗ × X ∗ is the linear relation adjoint to A (see § 1.2)).
 = X if one of the following conditions holds:
In particular, X
(1) X is a reflexive Banach space;
(2) R(λ0 , A) ∈ End X is a weakly compact operator for some λ0 ∈ ρ(A);
(3) dim Am 0 = dim(A∗ )m 0 < ∞.

Note that the result of Theorem 4.2 for a reflexive Banach space and m = 1 can
be found in [2]; § 1.3 and [13].
Corollary 4.4. If a linear relation A ∈ LR(X) satisfying Assumption 4.1 has a
compact resolvent, then X = X0 ⊕ X∞ = Am 0 ⊕ D(Am ).
Assumption 4.2. There exist quantities M > 0, ω ∈ R, and m ∈ N such that the
following estimates hold for all λ ∈ C with Re λ > ω and all n ∈ N:

M
(R(λ, A))mn   , n ∈ N. (4.8)
(Re λ − ω)mn

We now construct the phase space Φ(A) and the degenerate semigroups of linear
operators used for the definitions of solutions of the problem (4.1), (4.2). We carry
out this construction under Assumption 4.2 and the assumption that dim A0
1,
that is, A ∈ LR(X) \ LO(X). Assumption 4.2 shows that Assumption 4.1 holds,
therefore by Lemma 4.1 we can consider the ergodic subspace X  = Erg(X, (An ))
constructed for the bounded sequence (An ) ∈ End X. It is defined by formula (4.5),
where (λn ) is an arbitrary sequence in R+ ∩ ρ(A) such that limn→∞ λn = ∞. Thus,
we have the result of Theorem 4.2 on the decomposition of the space X,  in which
the subspaces X∞ = A 0 and X0 = D(A ) are A-invariant. The restriction
m m

A0 = A  X0 ∈ LO(X0 ) of the relation A to X0 continues to satisfy Assumption 4.2.

We shall see that with its help one can construct a C0 -semigroup {T0 (t); t
0}
in X0 with generator A0 , which has the domain D(A0 ) dense in X0 by Theorem 4.2.
For the construction of the semigroup we use an analogue of the Yosida approx-
imation (see [4], Theorem 12.3.1) of the following form:
 
λn
An = −
0
(I − (−λn R(λn , A0 ))m ) ∈ End X0 , n
1.
m

Lemma 4.3. The following estimate holds for each µ ∈ ρ(A) under Assump-
tion 4.1:

(A0n − (−λn R(λn , A0 ))m A0 )(R(µ, A0 ))m   const |λn |−1 , n
1. (4.9)
 Spectral analysis of linear relations and degenerate operator semigroups 1591

Proof. We transform the left-hand side of (4.9) as follows:

(A0n − (−λn R(λn , A0 ))m A0 )(R(µ, A0 ))m

  
λn
= − (I − (−λn R(λn , A0 )) ) − (−λn R(λn , A0 )) A0 (R(µ, A0 ))m
m m
m
= (−λn R(λn , A0 ))m
    
λn −1 λn −1
× − (I − λn A0 ) +
m
(I − mλn A0 ) (R(µ, A0 ))m
m m
 
m
λn k −k k
= − (−λn R(λn , A0 ))m Cm λn A0 (R(µ, A0 ))m .
m
k=2

The estimate (4.9) is now a consequence of the above representation of its left-
hand side and Assumption 4.1.
Theorem 4.4. Let A ∈ LR(X) be a linear relation satisfying Assumption 4.2 such
that dim A0
1. Then
Φ(A) ∩ X = D(Am ) = X0

and there exists a unique degenerate operator semigroup {T(t); t
0} ⊂ End X 

  
generated by the relation A ∈ LR(X) defined by the equalities A = A0 on X0 ,
 = X0 ∩ D(A), and A0
D(A)  = X∞ . The semigroup {T(t); t
0} has the following
properties:
(1) its restriction {T0 (t); t
0} ⊂ End X0 to X0 is a C0 -semigroup, and each
solution x : R+ → X of the problem (4.1), (4.2) with x0 ∈ D(A0 ) ⊂ X0 has
the form x(t) = T0 (t)x0 , t
0;
(2) T(0) ∈ End X  is the projection onto the subspace X0 parallel to X∞ .
∗
If vectors in X∞ = Am 0 ⊂ X separate functionals in X∞ = (A∗ )m 0 ⊂ X ∗ (for
instance, if one of the three conditions in Theorem 4.3 holds), then X  = X and
Φ(A) = X0 .
Proof. We start from the case when Assumption 4.2 holds for M = 1 and ω = 0.
Then (−λn R(λn , A0 ))m   1 and therefore

exp(A0n t)  1, n
1, t
0. (4.10)

We claim that the following results hold under Assumption 4.2 with M = 1 and
ω = 0:
(i) T0 (t)x = limn→∞ exp(A0n t)x is well defined for each x ∈ X0 ;
(ii) {T0 (t); t
0} is a C0 -semigroup on X0 ;
(iii) the generator of this semigroup is the operator A0 .
We now prove (i). Let Tn0 (t) = exp(A0n t), n
1, t
0. Then for x ∈ X0 we have
 t
d 0
Tn0 (t)x − Tk0 (t)x = T (t − s)Tn0 (s)x ds
0 ds k
 t
= Tk0 (t − s)Tn0 (s)(A0k x − A0n x) ds.
0
1592 A. G. Baskakov and K. I. Chernyshov

Hence

Tn0 (t)x − Tk0 (t)x  tA0n x − A0k x, t
0, x ∈ X0 .

It follows by inequality (4.9) and Lemma 4.3 (see also Remark 4.3) that the
sequence (A0n x) is Cauchy for each vector x in D(Am+1 0 ) and converges to A0 x.
Since D(Am+10 ) is dense in X0 and Tn
0
(t)  1 for all t
0, the function sequence
ϕn (t) = Tn0 (t)x, where x ∈ X0 , converges uniformly on each interval [0, t0] in R+ .
We now prove (ii). It follows from the pointwise convergence of the semi-
groups {Tn0 (t); t
0}, n
1, that their limit is a contraction C0 -semigroup
{T0 (t); t
0} ⊂ End X0 .
We prove (iii). Let B0 : D(B0 ) ⊂ X0 → X0 be the operator generating the semi-
group {T0 (t); t
0}. We claim that B0 = A0 . If x ∈ D(Am+1 0 ), then by the above
the function sequence ϕn (t) = Tn0 (t)x, t
0, n
1, converges uniformly on each
compact interval [0, t0 ] to a function ϕ : R+ → X0 . Since ϕn (t) = Tn0 (t)A0n x, t
0,
the ϕn (t) converge uniformly to the function T0 (t)A0 x. Hence x ∈ D(B0 ), and
A0 x = B0 x for x ∈ D(Am+1 0 ), so that A0 (R(λ0 , A0 ))m+1 = B0 (R(λ0 , B0 ))m+1
for each λ0 > 0 (note that R+ ⊂ ρ(A0 ) ∩ ρ(B0 )). This equality shows that
(R(λ0 , A0 ))m = (B0 − λ0 I)(R(λ0 , A0 ))m+1 , and therefore R(λ0 , B0 ) = R(λ0 , A0 ),
that is, B0 = A0 .
We consider now the general case in Assumption 4.2. If ω = 0, then proceeding
if necessary from A to the relation A − ωI we assume that ω = 0 (see Remark 4.1).
Hence
(−λR(λ, A))mn   M for all λ > 0, n ∈ N.
For each µ > 0 we define a new norm in X by the formula

xµ = sup (−µR(µ, A))mn x, x ∈ X.

n
0

With respect to this norm we have

(iv) x  xµ  M x, x ∈ X;
(v) (−λR(λ, A))m µ  1 for each λ > 0.
Hence the norms  · µ and  ·  are equivalent and the relation A satisfies
Assumption 4.2 with M = 1, ω = 0, and the norm  · µ . It remains to use
properties (i)–(iii).
Next, we extend the operator semigroup {T0 (t); t
0} ⊂ End X0 constructed
above to a degenerate operator semigroup {T(t); t
0} ⊂ End X  by setting
 
T (t)x = T0 (t)x, x ∈ X0 , and T (t)x = 0 for each x ∈ X∞ . It is clear that all
solutions of the problem (4.1), (4.2) with x0 ∈ X0 can be defined by means of the
semigroup {T0 (t);
0} on the subspace X0 , and the generator of {T(t); t
0} is
the linear relation A  on X
 with the properties mentioned in the statement of The-
orem 4.4. Obviously, the degenerate semigroup {T(t); t
0} with the prescribed
properties is unique.
We claim that Φ(A) ∩ X∞ = {0}. Assume that there exists a vector x1 in
Φ(A) ∩ X∞ distinct from zero and let x 1 : R+ → X be the solution of (4.1) such
that x 
1 (0) = x1 ∈ D(A). If P∞ ∈ End X is the projection onto X∞ parallel to X0 ,
 Spectral analysis of linear relations and degenerate operator semigroups 1593

then P∞ x 1 : R+ → X∞ is a solution of the inclusion ẋ(t) ∈ A∞ x(t), t
0, and

x∞ (0) = x1 ∈ X∞ . Taking account of Remark 4.2 and the fact that 0 ∈ ρ(A∞ ) we
obtain A−1 −1
∞ ẋ∞ (t) = x∞ (t) for all t
0. Hence x∞ (t) ∈ Im A∞ for all t
0 and
−1 m
therefore x∞ (t) ∈ Im(A∞ ) = {0} for t
0, that is, x1 (0) = x1 = 0, which is a
contradiction.
 by Theorem 4.3.
If vectors in Am 0 separate functionals in (A∗ )m 0, then X = X
Hence Φ(A) = X0 .
Remark 4.4. The results of Theorem 4.4 for m = 1 can be found in the mono-
graph [2], Chapter II. Note that even if A is a linear relation on a finite-dimensional
space X and A2 0 = A0, these results from [2] cannot be of use. The decomposition
X = D(A) ⊕ A0 was established in [2] only for a reflexive Banach space. For m > 1
differential inclusions of the form (4.1) were studied in [13] by the method of n times
integrated semigroups. However, the main results were announced in [13] under the
a priori assumption that the decomposition X = D(Am ) ⊕ Am 0 exists; this decom-
position was pointed out to exist in the case of a reflexive Banach space X and
m = 1.
Corollary 4.5. Let A ∈ LR(X) be a linear relation satisfying Assumption 4.1 and
assume that there exist quantities M > 0, ω ∈ R such that the following estimates
hold for all x ∈ X0 , λ ∈ C with Re λ > ω, and all n ∈ N:
M x
(R(λ, A))n x  .
(Re λ − ω)n
Then all the results of Theorem 4.4 also hold.
Remark 4.5. Under the assumptions of Theorem 4.4 one can show that the phase
space Φ(A) contains no vectors in X \ X  with the following properties:
(1) x0 is a vector such that the Cauchy problem (4.1), (4.2) has a solution with m
derivatives;
(2) x0 is a vector such that the Cauchy problem (4.1), (4.2) has a solution x(t)
with the estimate x(t)  M exp(ωt), t
0.
One immediate consequence of Theorem 4.4 is as follows.
Theorem 4.5. Let A ∈ LO(X) be a linear operator with D(A) = X such that
Assumption 4.2 holds. Then A is a generator of a C0 -semigroup.
From Theorems 3.1 (condition (2)), 4.1, and 4.4 we deduce the following result.
Theorem 4.6. If the relation A ∈ LR(X) satisfies one of the conditions in The-
orem 3.1, then Φ(A) = X0 ; moreover, each solution x : R+ → X of the problem
(4.1), (4.2) with x0 ∈ X0 can be defined by means of the analytic group of operators
{exp A0 t, t ∈ R} and has the form x(t) = (exp A0 t)x0 , t ∈ R.
Note that the semigroup {T(t); t
0} constructed in Theorem 4.4 for the
decomposition (3.2) of the space X has the form T(t) = exp A0 t ⊕ 0. It can also
be expressed as T0 (t) = (exp A0 t)P0 , t
0. For all t
0 each solution x of the
inclusion (4.1) can be represented in the form x(t) = T0 (t)x0 , x0 ∈ X0 = Φ(A).
Remark 4.6. Under Assumption 4.2 the subspace X0 is the phase space Φ(A) (the
subspace of initial data) for mild solutions [14].
1594 A. G. Baskakov and K. I. Chernyshov

Remark 4.7. The subspace X∞ arising under Assumption 4.1 makes no contribution
to the phase space Φ(A) by Theorem 4.1, because A−1 ∞ is a nilpotent operator.
However, if σ  (A) = {∞}, that is, if A−1 ∈ End X is a quasinilpotent operator, then
it is possible that Φ(A) = X. For a corresponding example one can take an arbitrary
operator A ∈ LO(X) generating a C0 -semigroup with σ (A) = {∞}; for instance,
the generator of a nilpotent C0 -semigroup (see [14]).

§ 5. Sectorial linear relations and

degenerate analytic operator semigroups
In this section we define sectorial linear relations and construct for them degen-
erate analytic semigroups of linear operators. In the proofs of the main results we
also make essential use of ergodic theorems.
Definition 5.1. A relation A ∈ LR(X) is said to be sectorial with angle θ ∈ (π/2, π)
if for some a ∈ R the sector
Ω = Ωa,θ = {λ ∈ C | | arg(λ − a)| < θ, λ = a}
lies in the resolvent set ρ(A) of A and for each δ ∈ (0, θ − π/2) there exist m ∈ N
and Mδ
1 such that
sup ((a − λ)R(λ, A))m  = Mδ < ∞. (5.1)
λ∈Ωa,θ−δ

The following assumption holds throughout this section.

Assumption 5.1. A ∈ LR(X) is a sectorial relation.
In view of Remark 4.1, we can assume without loss of generality that a = 0 in
the definition of a sectorial relation. Conditions (5.1) take the following form for
a = 0:
sup (−λR(λ, A))m  = Mδ < ∞. (5.10 )
λ∈Ω0,θ−δ

In what follows we carry out our constructions under assumption (5.10 ); however,
we do not include the constraint a = 0 in the statements of results.
For the construction of an analytic operator semigroup generated by a sectorial
relation A we require the following result.
Lemma 5.1. There exists a constant C > 0 such that the resolvent of the sectorial
relation A ∈ LR(X) satisfies the following estimate for each δ ∈ (0, θ − π/2):
R(λ, A)  C(1 + |λ|)m−2 , λ ∈ Ωa,θ−δ . (5.2)
d
Proof. Since (R(λ, A))n = n(R(λ, A))n+1 , n ∈ N, it follows that for each fixed
dλ
point λ0 ∈ Ωa,θ−δ ⊂ ρ(A),
 λ
(R(λ, A))n = n (R(µ, A))n+1 dµ + (R(λ0 , A))n , (5.3)
λ0
where the integration proceeds over an arbitrary piecewise-smooth curve γ con-
necting points λ0 and λ in Ωa,θ−δ and lying entirely in Ωa,θ−δ . The estimate in
question follows from the representation (5.3), which must be repeatedly used for
n = m − 1, m − 2, . . . , 1, so that the function µ → (R(µ, A))m will be integrated m
times.
 Spectral analysis of linear relations and degenerate operator semigroups 1595

Definition 5.2. Let A be a sectorial relation in LR(X) with angle θ. Then for
z ∈ Ω0,θ−δ , where δ ∈ (0, θ − π/2), we set

1
T (z) = − eλz R(λ, A) dλ. (5.4)
2πi γ

For the curve γ = γ(r, ε) in Definition 5.2 we can take the union of three
curves γk (r, ε), k = 1, 2, 3, of the following form:

γ1 (r, ε) = {−ρe−i(θ−ε) | −∞  −ρ  −r},

γ2 (r, ε) = {reiα | −(θ − ε)  α  θ − ε},
γ3 (r, ε) = {ρei(θ−ε) | r  ρ  ∞},

where ε = (δ0 − δ)/2, δ0 = θ − π/2 and r = 1/|z|.

The convergence of the integral in the uniform operator topology for all z ∈ Ω0,δ ,
δ ∈ (0, δ0 ), is a consequence of Lemma 5.1.
Now let δ0 = θ − π/2.
Theorem 5.1. Let A ∈ LR(X) be a sectorial linear relation (that is, condi-
tion (5.1) in Assumption 5.1 holds). Then equality (5.4) defines an operator sub-
semigroup of the algebra End X that is analytic in Ω0,δ0 ⊂ ρ(A) and bounded for
t > 0.
Proof. In the standard way (see [21]; [22], Chapter 1, § 3.4), one immediately estab-
lishes from the definition the analyticity of the function T in the open sector
Ω0,δ ∈ ρ(A) for t > 0, the property

T (z1 + z2 ) = T (z1 )T (z2 ), z1 , z2 ∈ Ω0,δ0 , (5.5)

and the strong continuity of the semigroup {T (t); t > 0}.

We now integrate T (z), z ∈ Ω0,δ , in (5.4) m times by parts. As a result, for
t > 0 we obtain the representation

(−1)m (m − 1)!
T (t) = eλt (R(λ, A))m dλ. (5.6)
2πitm−1 γ(1/t,ε)

Setting µ = λt in (5.6) we arrive at the formula

   m
(−1)m (m − 1)! µ
T (t) = R ,A eµ dµ.
2πitm γ(1,ε) t

Hence it follows by (5.1) that


(m − 1)! Mδ |eµ |
T (t)  |dµ|, (5.7)
2π γ(1,ε) |µ|m

which ensures the boundedness of the semigroup for t > 0.

1596 A. G. Baskakov and K. I. Chernyshov

We investigate the semigroup {T (t); t > 0} with the help of two operator-valued
functions
A(λ) = I − ((−λR(λ, A))m ), λ > 0, and B(t) = I − T (t), t > 0. (5.8)
In what follows we consider two arbitrary sequences (λn ) and (tn ), n
1, in R+
with the following properties:
λn ∈ Ω ⊂ ρ(A), lim λn = ∞, lim tn = 0,
n→∞ n→∞

and we consider the corresponding bounded operator sequences An = A(λn ) and

Bn = B(tn ), n ∈ N, in the Banach algebra End X.
Lemma 5.2. Under Assumption 5.1 the following relations hold for each fixed
λ0 ∈ Ω0,θ ⊂ ρ(A):
(1) Im(I − An ) ⊂ Im(R(λ0 , A))m , Im(I − Bn ) ⊂ Im(R(λ0 , A))m ;
(2) limn→∞ An (R(λ0 , A))m = 0, limn→∞ Bn (R(λ0 , A))m = 0.
Proof. The assertions concerning the sequence (An ) have been proved in Lemma 4.2.
The second inclusion in part (1) for the sequence (Bn ) is an immediate consequence
of (5.6) for t > 0. It remains to prove that for each fixed λ0 ∈ Ω0,θ ,
lim Bn (R(λ0 , A))m = 0. (5.9)
n→∞

In fact, by (5.4) and the equality


(m − 1)! eλt
1=− dλ (5.10)
2πitm−1 γ(1/t,ε) λm
we obtain 
(m − 1)! eλt
B(t) = I − T (t) = − A(λ) dλ. (5.11)
2πitm−1 γ(1/t,ε) λm
Fixing t > 0 we set µ = λt in (5.11), which yields
   µ
(m − 1)! µ e
B(t) = − A dµ. (5.12)
2πi γ(1,ε) t µm

Since T (t) commutes with R(λ, A), for t = tn equality (5.12) takes the following
form:
  
(m − 1)! µ eµ
B(tn )(R(λ0 , A))m = − A (R(λ0 , A))m m dµ. (5.13)
2πi γ(1,ε) tn µ

Using (4.9) and (5.13) we arrive at the estimates


|eµ |
B(tn )(R(λ0 , A))   ctn
m
|dµ|, n
1,
γ(1,ε) |µ|
m+1

which lead to the limit relation (5.9).

Thus, in accordance with the terminology of § 4 (Definition 4.3) both (An )
and (Bn ) are J-sequences for the ideal J ⊂ A considered in § 4.
 Spectral analysis of linear relations and degenerate operator semigroups 1597

Theorem 5.2. The following equalities hold under Assumption 5.1:

 = X0 ⊕ X∞ ,
Erg(X, (An )) = Erg(X, (Bn )) = X (5.14)

and the spaces X0 and X∞ are closed. The equalities

P∞ x = lim An x = lim Bn x, 
x ∈ X,
n→∞ n→∞

 with the following properties:

define a bounded projection P∞ ∈ End X
(1) P∞  min{limn→∞ An , limn→∞ Bn };
(2) Im P∞ = X∞ , Ker P∞ = X0 .
All the assertions of Theorem 5.2 are consequences of Lemma 1 in [20], for the
application of which one must use Lemma 5.2.
We set P0 = I − P∞ and observe that

P0 x = lim (−λn R(λn , A))m x = lim T (tn )x, 

x ∈ X.
n→∞ n→∞

These limits are well defined and equal whatever the choice of the sequences (λn )
and (tn ), n ∈ N, with the above-mentioned properties. This ensures the following

representations of the subspaces X0 and X∞ of X:

X0 = {x ∈ X | lim T (t)x = x}, X∞ = Ker T (t). (5.15)
t→+0
t>0

These representations mean that the restriction {T0 (t); t
0} of the semigroup T (t)
to the subspace X0 is a semigroup strongly continuous at the origin and analytic
in the sector Ω0,δ0 . The restriction {T(z); z ∈ Ω0,δ0 } of the function T : Ω0,δ0 →
End X to the subspace X  is also analytic in this sector.
We denote the inverse relation to the operator zero on X∞ by O∞ = 0−1 ∞ .

Theorem 5.3. Let A ∈ LR(X)\LO(X) be a sectorial relation. Then {T(t); t
0}

is an operator subsemigroup of End X  analytic in the sector Ω0,δ ; this is a degen-
0

erate semigroup with generator A  = A0 ⊕ O∞ ∈ LR(X),  where O∞ ∈ LR(X∞ ),

D(O∞ ) = {0}, O∞ 0 = X∞ , the operator A0 = A   X0 ∈ LO(X0 ) is the gen-
erator of a strongly continuous operator semigroup {T0 (t); t
0} analytic in the
sector Ω0,δ0 , and Φ(A) ∩ X  = X0 . Each solution of the problem (4.1), (4.2) with
x0 ∈ D(A0 ) has the form x(t) = T0 (t)x0 , t
0.
If vectors in the subspace Am 0 ⊂ X separate functionals in the subspace (A∗ )m 0
of X ∗ (in particular, if one of the three conditions in Theorem 4.3 holds), then
X = X and Φ(A) = X0 .

Proof. We shall prove that the relation A = A0 ⊕ O∞ is the generator of the

degenerate semigroup {T(t); t
0}. To this end it is sufficient to show (see
Theorem 4.1) that the closed linear operator A0 ∈ LO(X0 ) is the generator of the
operator semigroup {T0 (t); t
0}. Clearly, Assumption 5.1 holds for A0 (with
the same constants from Definition 5.1). Assume that B0 : D(B0 ) ⊂ X0 → X0 is
1598 A. G. Baskakov and K. I. Chernyshov

the generator of the C0 -semigroup {T0 (t); t
0}. To prove that A0 = B0 it is

sufficient to show that R(λ, A0 ) = R(λ, B0 ) for λ = |ω0 | + 2, where ω0 is a quantity
such that T0 (t)  const exp(ω0 t), t
0.
Consider now a sequence of curves γn = γ(rn , ε) each of which is the union of
three curves γk (rn , ε), k = 1, 2, 3, of the same form as in Definition 5.2, where the
radii rn of the circles with centre at 0 approach infinity. Then by the formulae
 
R(µ, A0 )
dµ = 2πiR(λ, A0 ), eµt R(µ, A0 )x dµ = −2πiT0 (t)x
γn µ−λ γn

and Fubini’s theorem, for all x ∈ X0 we obtain

 τ 
1 1 − eτ(µ−λ)
e−λt T0 (t)x dt = R(µ, A0 )x dµ
0 2πi γn µ−λ

1 eτ(µ−λ)
= R(λ, A0 )x − R(µ, A0 )x dµ.
2πi γn µ − λ

Since Re(µ − λ)  −1 for µ ∈ γn , ε = (δ0 − δ)/2 and


lim e−|τ| µm−2 d|µ| = 0,
τ→∞ γn

it follows by Lemma 5.1 that


eτ(µ−λ)
lim R(µ, A0 ) dµ = 0
τ→∞ γn µ−λ
and therefore
 τ
R(λ, B0 )x = lim e−λt T0 (t)x dt = R(λ, A0 )x, x ∈ X0 .
τ→∞ 0

Note that the remaining assertions of Theorem 5.3 were already proved in § 4.
Remark 5.1. Since the semigroup {T0 (t); t
0} constructed under Assumption 5.1
is analytic, it follows by well-known results (see [21], [22]) that Assumption 5.1 holds
for the operator A0 ∈ LO(X) with m = 1 (perhaps with another constant Mδ ).

§ 6. Applications to the spectral theory

of ordered pairs of linear operators
In this section we apply the spectral theory of linear relations to the spectral
theory of ordered pairs (G, F ) of closed linear operators F : D(F ) ⊂ X → Y and
G : D(G) ⊂ X → Y from a complex Banach space X into another complex Banach
space Y . We assume that the domains D(F ) and D(G) satisfy one of the following
conditions:
(i) D(F ) = X, D(G) = X;
(ii) D(F ) = X, D(G) = X;
(iii) D(F ) = X, D(G) = X.
Let D = D(G, F ) be the subspace D(F ) ∩ D(G); we call it the domain of the
ordered operator pair (G, F ).
 Spectral analysis of linear relations and degenerate operator semigroups 1599

Definition 6.1. We include in the resolvent set ρ(G, F ) of an ordered operator

pair (G, F ) all complex numbers λ = 0 such that the operator G−λF : D ⊂ X → Y
has a continuous inverse and also include the point λ = 0 if G : D → Y has a
continuous inverse operator and D(F ) = X. We call the set σ(G, F ) = C \ ρ(G, F )
the spectrum of the pair.
We call the operator-valued function

R( · ; G, F ) : ρ(G, F ) ⊂ C → Hom(Y, X),

R(λ; G, F ) = (G − λF )−1 , λ ∈ ρ(G, F ),

the resolvent of the ordered pair (G, F ). It is defined and analytic on the open
set ρ(G, F ).
Remark 6.1. If one treats condition 0 ∈ ρ(G, F ) formally as the continuous invert-
ibility of the operator G, then in case (ii) the point 0 is isolated in ρ(G, F ) and
therefore ρ(G, F ) is not an open set. Other complications will also occur in that
case.
Similarly to linear relations it is very useful in this context to introduce the
concept of extended spectrum of an ordered operator pair.
Definition 6.2. The subset σ  equal to σ(G, F ) if the function R( · ; G, F )
 (G, F ) of C
has an analytic continuation at ∞ such that lim|λ|→∞ R(λ; G, F ) = 0 and to
σ(G, F ) ∪ {∞} otherwise is called the extended spectrum of the ordered pair (G, F ).
 \
We call the set ρ(G, F ) = C σ(G, F ) the extended resolvent set of the pair (G, F ).
In the reduction of the problem (1.2), (1.3) to (4.1), (4.2), in the natural way one
comes across two linear relations, Al = F −1 G ⊂ X × X and Ar = GF −1 ⊂ Y × Y ,
which we call the left and the right linear relations for the ordered pair (G, F ).
We now present several simple properties of linear relations (see [1], Chapter I),
which we shall often use in what follows without special mention.
Let X, Y , Z, and W be Banach spaces and let A ∈ LR(X, Y ), B ∈ LR(Y, Z).
Then the following relations hold:
(1) (BA)−1 = A−1 B−1 ;
(2) D(BA) = A−1 D(B) = A−1 (Im B−1 ), Im BA = B(Im A);
(3) C(BA) = (CB)A for each C ∈ LR(Z, W ).
We also point out two further properties: for A1 , A2 ∈ LR(X, Y ), C ∈ LR(Y, Z),
and T ∈ LR(W, X) we have the equalities
(4) C(A1 + A2 ) = CA1 + CA2 if D(C) ⊃ Im A1 ∩ Im A2 ;
(5) (A1 + A2 )T = A1 T + A2 T if Im T ⊂ D(A1 ) ∩ D(A2 ).
It follows from properties (1) and (2) that for the left and the right linear rela-
tions Al and Ar constructed for an ordered pair (G, F ) we have the following
representations:

D(Al ) = G−1 (Im F ), Im Al = F −1 (Im G), (6.1)

D(Ar ) = F (D(G)), Im Ar = G(D(F )), (6.2)
−1
R(λ, Al ) = (G − λF ) F, λ ∈ ρ(G, F ), (6.3)
R(λ, Ar ) = F (G − λF )−1 , λ ∈ ρ(G, F ). (6.4)
1600 A. G. Baskakov and K. I. Chernyshov

Here formula (6.3) holds only if D(G) and D(F ) satisfy condition (i) or con-
dition (iii). On the other hand, if condition (ii) holds, then (6.3) fails because
D(F ) = X. Note that in that case one can apply to G−1 F formula (2.7) with
0 = λ ∈ ρ(G, F ), which yields the relation

R(λ, Al ) = −λ−1 I − λ−2 (A−1 −1 −1

l − λ I)
= −λ−1 I − λ−2 (G−1 F − λ−1 I)−1 = −λ−1 (I − (G − λF )−1 G).
(6.5)

It follows from (6.3)–(6.5) that if ∞ ∈ ρ(G, F ), then ∞ ∈ (Al ) ∪ σ

/ σ (Ar ), so that
Al ∈ End X and Ar ∈ End Y . We have thus established the following result.
Lemma 6.1. The following inclusions hold:

ρ(G, F ) ⊂ ρ(Al ) ∩ ρ(Ar ), (G, F ) ⊃ σ

σ (Al ) ∪ σ
(Ar ). (6.6)

We call the resolvents of the relations Al and Ar the left and the right resolvents
of the ordered operator pair (G, F ); we denote them by Rl ( · ; G, F ) and Rr ( · ; G, F ),
respectively. By definition, these functions range in the algebras End X and
End Y , respectively.
In what follows we assume that the following condition (of the non-singularity
of the pair (G, F )) holds.
Assumption 6.1. For the ordered pair (G, F ) the set ρ(G, F ) is non-empty.
Remark 6.2. The following conditions were assumed to hold in the monograph [2]:
D(F ) ⊂ D(G) ⊂ X = Y . In [11], [12] the authors considered operators F and G
with the following properties: F ∈ Hom(X, Y ), D(G) = X. The most general case
is considered in [23], where D(F ) and D(G) can both be distinct from X at the
same time. However, the analysis there was carried out under the assumption that
D = D(F ) ∩ D(G) = {0} and, moreover, under our Assumption 6.1. This enables
one to consider in D the norm x∗ = (G − λ0 F )x, x ∈ D, where λ0 is a scalar
from ρ(G, F ). This norm makes D a Banach space isomorphic to Y . Considering D
in place of X one can assume that condition (iii) holds. In that case for the Banach
space containing D one can take one of the subspaces D(F ) and D(G) with the
corresponding graph norm (one’s choice of a subspace is usually determined by
the particular context; the subspace of X not taken into account here is discarded).
Alongside conditions (i)–(iii) we distinguish the following ones:
(iv) D(F ) ⊂ D(G) = X;
(v) D(G) ⊂ D(F ) = X.
If condition (iv) holds, then one can introduce in D(G) the norm of the graph
of the operator G: x∗ = x + Gx, x ∈ D(G), and considering D(G) in
place of X one can assume that condition (ii) holds. If condition (v) is satisfied,
then one considers in D(F ) the norm of the graph of the operator F , and we find
ourselves in the situation when (i) holds.
Theorem 6.1. The following relations hold for an ordered operator pair (G, F ):
(G, F ) = σ
(1) σ (Al );
(G, F ) \ {0, ∞} = σ
(2) σ (Ar ) \ {0, ∞};
 Spectral analysis of linear relations and degenerate operator semigroups 1601

(3) (G, F ) = σ
σ (Al ) = σ (Ar ) if D = X;
(4) 0 ∈ ρ(Al ) ⇔ G−1 F ∈ End X;
(5) 0 ∈ ρ(Ar ) ⇔ F G−1 ∈ End Y ;
(6) 0 ∈ ρ(Al ) ∩ ρ(Ar ) ⇔ 0 ∈ ρ(G, F );
(7) ∞ ∈ ρ(Al ) ⇔ Al = F −1 G ∈ End X;
(8) ∞ ∈ ρ(Ar ) ⇔ Ar = GF −1 ∈ End Y ;
(9) ∞ ∈ ρ(Al ) ∩ ρ(Ar ) ⇔ ∞ ∈ ρ(G, F ).

Proof. We prove the inclusion

(ρ(Al ) ∪ ρ(Ar )) \ {0} ⊂ ρ(G, F ). (6.7)

Assume that 0 = λ ∈ ρ(Al ) and let λ0 = λ be a fixed number in ρ(G, F ), which

exists by Assumption 6.1. We represent the operator G − λF in the following form:

G − λF = (G − λ0 F )(I − (λ − λ0 )R(λ0 , Al ))
= (I − (λ − λ0 )R(λ0 , Ar ))(G − λ0 F ). (6.8)

It follows by Corollary 2.1 that 1/(λ − λ0 ) ∈ / σ(R(λ0 , Al )) and therefore the

operator G − λF : D ⊂ X → Y is continuously invertible as the product of two
operators with this property. Hence λ ∈ ρ(G, F ) and therefore ρ(Al ) ⊂ ρ(G, F ). In
a similar way we prove that ρ(Ar ) ⊂ ρ(G, F ). This completes the proof of (6.7).
Corollaries (4) and (5) are consequences of the definition of a continuously invert-
ible relation. If 0 ∈ ρ(Al ) ∩ ρ(Ar ), then G−1 F ∈ End X and F G−1 ∈ End Y . These
are bounded operators, therefore D(F ) = X, Ker G = {0}, and Im G = Y . Hence
0 ∈ ρ(G, F ). We have thus proved relation (6).
Relations (7)–(9) are consequences of relations (5)–(7) and Theorem 2.4. Rela-
tion (2) is now a consequence of (6.7), (6.8), and (4)–(9).
We now prove relation (1) for the points 0 and ∞. The relations 0 ∈ ρ(Al )
and A−1l ∈ End X are equivalent by definition. Condition A−1 l = G−1 F ∈ End X
means that D(F ) = X, Ker G = {0}, and Im F ⊂ Im G. Assumption 6.1 yields
the equality Im F = Y , therefore F : D(F ) ⊂ X → Y is a continuously invertible
operator and 0 ∈ ρ(G, F ).
In a similar way we prove that the relations ∞ ∈ ρ(Al ) and Al ∈ End X are
equivalent, while the inclusion F −1 G ∈ End X means that D(G) = X, Ker F = {0},
and Im G ⊂ Im F . Proposition 6.1 yields Im G = Y , therefore G : D(G) ⊂ X → Y
is an operator with a continuous inverse and ∞ ∈ ρ(G, F ).
If D = X, then for λ0 ∈ ρ(G, F ) the operator G − λ0 F is an isomorphism of the
Banach spaces X and Y and we have the equality

Rr (λ; G, F )(G − λ0 F ) = (G − λ0 F )Rl (λ; G, F ), λ ∈ C, (6.9)

which shows that the left and the right resolvents of the pair (G, F ) are similar,
so that σ(Al ) = σ(Ar ). From properties (1) and (2) we now see that relation (3)
holds.
Corollary 6.1. If both points 0 and ∞ belong to ρ(G, F ), then D = X and the
operators G, F ∈ Hom(X, Y ) are continuously invertible.
1602 A. G. Baskakov and K. I. Chernyshov

Corollary 6.2. An ordered pair (G, F ) has the following properties:

(G, F ) = {0} ⇔ the operator F : D → Y is continuously invertible,
(a) σ
D(G) = X, and the operators Al = F −1 G ∈ End X, Ar = GF −1 ∈ End Y
are quasinilpotent.
(G, F ) = {∞} ⇔ the operator G : D → Y is continuously invertible,
(b) σ
D(F ) = X, and the operators A−1
l =G
−1
F ∈ End X, A−1
r =FG
−1
∈ End Y
are quasinilpotent.
We set S0 (δ) = {0 = λ ∈ C : |λ| < δ, δ > 0}.
Remark 6.3. The following two conditions are equivalent:
(a) D(F ) = X and G : D(G) ⊂ X → Y is a continuously invertible operator;
(b) there exists δ>0 such that S0 (δ)⊂ρ(G, F ) and supλ∈S0 (δ) (G −λF )−1<∞.
The proof of the equivalence of these two conditions can be reduced to Lemma 2.2
by means of Theorem 6.2.
We point out that the sets σ(G, F ) = σ(Al ) and σ
(Ar ) can be distinct.
Example 6.1. Let H be a complex Hilbert space and let T : H → H be a non-
invertible isometry. Then the operator T ∗ has a non-trivial kernel, T T ∗ = I, and
T ∗ T = I.
First we consider the operator pair (G, F ), where G = T ∗ ∈ End H and
F = T −1 : Im T ⊂ H → H. We have 0 ∈ σ(G, F ). Incidentally, Al = T T ∗ = I
and Ar = T ∗ T = I, so that 0 ∈ σ(Al ), but 0 ∈ / σ(Ar ) = {1}.
Consider next the ordered pair (F, G). Since the left and right relations for it are
A−1
l and A−1r , respectively, it follows by Theorems 2.4 and 6.1 that ∞ ∈ σ(F, G) =
−1
(Al ), but ∞ ∈
σ /σ(A−1r ).
Finally, we can combine both cases by considering a suitable operator pair in
the product H × H.
Theorem 6.2. The following relation holds for the extended spectra of the ordered
pairs (G, F ) and (F, G):
 
1
(F, G) =
σ :λ∈σ(G, F ) . (6.10)
λ

Proof. Since the left and right linear relations for the pair (F, G) are the inverses
of Al and Ar , respectively, (6.10) is a consequence of Theorem 2.4 and property (1)
in Theorem 6.1.
Definition 6.3. An ordered pair of subspaces (X1 , Y1 ), where X1 ⊂ X and Y1 ⊂ Y ,
is said to be invariant with respect to (G, F ) if GX1 ⊂ Y1 and F X1 ⊂ Y1 .
Definition 6.4. Let
X = X0 ⊕ X1 , Y = Y0 ⊕ Y1 (6.11)
be direct sums of closed subspaces such that (X0 , Y0 ) and (X1 , Y1 ) are (G, F )-
invariant pairs. Let Gi , Fi : D(G, F ) ∩ Xi = Di → Yi , i = 0, 1, be the restrictions
of the operators G and F to the Xi , i = 0, 1. Then we write
(G, F ) = (G0 , F0) ⊕ (G1 , F1) (6.12)
 Spectral analysis of linear relations and degenerate operator semigroups 1603

and say that the ordered operator pair (G, F ) has the representation (6.12) with
respect to the decompositions (6.11) of the spaces and is the direct sum of the pairs
(G0 , F0 ) and (G1 , F1 ).
(G, F ) of an ordered pair
Theorem 6.3. Assume that the extended spectrum σ
(G, F ) has a representation

(G, F ) = σ0 ∪ σ1 ,
σ (6.13)

where σ0 is a compact set, σ1 is closed, and σ0 ∩ σ1 = ∅.

Then there exist (G, F )-invariant pairs of subspaces (X0 , Y0 ) and (X1 , Y1 ) such
that the decompositions (6.11) and (6.12) hold and, in addition,
(1) the projections Pi ∈ End X and Qi ∈ End Y , i = 0, 1, performing the
decompositions (6.11) (that is, such that Im Pi = Xi and Im Qi = Yi for
i = 0, 1) are defined by the formulae
 
1 1
P0 = − R(λ, Al ) dλ, Q0 = − R(λ, Ar ) dλ, (6.14)
2πi γ0 2πi γ0
P1 = I − P0 , Q1 = I − Q0 , (6.15)

where γ0 is a closed Jordan curve (or finitely many such curves) lying
in ρ(A), so that σ0 lies inside it and σ1 lies outside;
(G0 , F0 ) = σ(G0 , F0 ) = σ0 , σ
(2) σ (G1 , F1) = σ1 ;
(3) D(G0 ) = X0 , F0 : D(F0 ) = D(F ) ∩ X0 ⊂ X0 → Y0 is a continuously
invertible operator and Al = F0−1 G0 ∈ End X0 , Ar = G0 F0−1 ∈ End Y0 ;
(0) (0)

(0) (0)
(4) the left Rl ( · ; G0, F0 ) = R( · , Al ) and the right Rr ( · ; G0 , F0) = R( · , Ar )
resolvents of the pair (G0 , F0 ) are similar and
(0) (0)
 (Al ) = σ(Al ) = σ
σ (A(0) (0)
(G0 , F0);
r ) = σ(Ar ) = σ

(5) D(F1 ) = X1 , G1 : D(G1 ) = D(G) ∩ X1 ⊂ X1 → Y1 is a continuously

invertible operator if 0 ∈/ σ1 , in which case Rl (0; G1, F1 ) = G−1
1 F1 ∈ End X1
−1
and Rr (0; G1, F1 ) = F1 G1 ∈ End Y1 ; in addition, Rl ( · ; G1 , F1) = 0 and
Rr ( · ; G1 , F1) = 0 if X1 = Al 0.

Proof. If condition (6.13) holds, then Theorem 6.1 yields the equality

(Al ) ∪ σ
σ (Ar ) = σ0 ∪ σ1 , (6.16)

which in accordance with Theorem 2.3 enables one to define the projections Pi
and Qi , i = 0, 1, by formulae (6.14) and (6.15) and to consider the corresponding
space decompositions. It follows from the easily verified equalities

GR(λ, Al ) = R(λ, Ar )G, F R(λ, Al ) = R(λ, Ar )F, λ ∈ ρ(G, F ),

and (6.14), (6.15) that

GPi = Qi G, F Pi = QiF, i = 0, 1. (6.17)

1604 A. G. Baskakov and K. I. Chernyshov

These equalities ensure the (G, F )-invariance of the (Xi , Yi), i = 0, 1; therefore,
(G, F ) = (G0 , F0) ⊕ (G1 , F1 ) and, clearly, R(λ; G, F ) = R(λ; G0 , F0 ) ⊕ R(λ; G1 , F1 ),
Al = Al ⊕Al , and Ar = Ar ⊕Ar with Al = F0−1 G0 and Al = G0 F0−1 being
(0) (1) (0) (1) (0) (0)

operators from End X0 and End Y0 , respectively, R(0, Al ) = G−1

(1)
1 F1 ∈ End X1 , and
R(0, Ar ) = F1 G−1
(1)
1 ∈ End Y 1 . It remains to use Theorems 2.3 and 6.1, which yield
the equalities σ(Gi , Fi) = σi , i = 0, 1, and all the other assertions of the lemma.
Remark 6.4. If in accordance with Remark 6.2 we take for X the subspace D(F )
or D(G) with the corresponding graph norm, then in place of (6.11) in Theorem 6.3
we obtain a decomposition of D(F ) or D(G).
We did not know Theorem 6.3 in such a general form before. Some of its asser-
tions were earlier obtained under certain conditions on the domain D(G, F ) of
the pair (G, F ). For instance, Rutkas [24] states without proofs some assertions
from parts (1)–(3) of Theorem 6.3 in the case when D(G, F ) = X. The case
σ0 = {0}, D(G, F ) = D(F ) has been considered in detail in [25] and in [26] after
that. Ditkin [27] obtained a result covered by assertions (1)–(3) of Theorem 6.3
under the assumption D(F ) ⊂ D(G) ⊂ X and for X = Y (see Remarks 6.2 and 6.4).
The most general result for a pair (G, F ) of closed operators was discussed by Rad-
bel’ [23] (see Remark 6.4). However, she used formulae of the form (6.15), where
the resolvent of the relation Al was replaced by the right-hand side of (6.3), which
does not allow one to consider the case D(F ) = X.
In fact, the above-mentioned authors have not seized on the opportunities and
the advantages that the use of the extended spectrum of a pair offers for a more
thorough analysis of the decomposition of a pair.
We pointed out in Example 6.1 that the sets σ (G, F ) and σ
(Ar ) can differ only
by the presence or the absence of the points 0 and ∞. The distinction between
them is more precisely described by the following result.
Theorem 6.4. The following two equivalences hold:
(1) 0 ∈ σ(G, F ) \ σ(Ar ) ⇔ Ker G = {0}, Y = Im G, D(F ) = D(F ), and
X = Ker G ⊕ D(F );
(2) ∞ ∈ σ(G, F ) \ σ
(Ar ) ⇔ Ker F = {0}, Y = Im F , D(G) = D(G), and
X = Ker F ⊕ D(G).
Proof. By Theorem 6.2 it is sufficient to prove only one of the two assertions, because
they can be obtained from each other by changing the ordering of the operators.
We shall prove (1). Let 0 ∈ σ(G, F ) \ σ(Ar ). Then it follows by Theorem 6.1
that 0 is an isolated point in the set σ(G, F ) = σ(Al ). Using Theorem 6.3 for
σ0 = {0} and σ1 = σ (G, F )\{0} we obtain decompositions of the form (6.11), (6.12),
(1) (1)
in which σ(G0 , F0) = σ(G0 , F0 ) = {0} and 0 ∈ (G1 , F1 ) = σ
/ σ (Al ) ⊃ σ (Ar ),
where Al = F1−1 G1 , Ar = G1 F1−1 . Since 0 ∈
(1) (1)
/ σ(Ar ), it follows by formula (6.14)
for the projection Q0 that Q0 = 0, that is, Y0 = {0}. It follows from the above
and Theorems 6.1 and 6.3 that D(G0 ) = X0 , G0 = 0, and D(F0 ) ∩ X0 = {0} (if
D(F0 ) ∩ X0 = {0}, Proposition 6.1 would fail), D(F1 ) = D(F ) = X1 , and G1 is
an invertible operator. Hence X = Ker G ⊕ D(F ). Since G0 = 0, it follows that
Im G1 = Im G = Y1 = Y .
Conversely, if the conditions in part (1) hold, then, clearly, 0 ∈ σ(G, F ) \ σ(Ar ).
 Spectral analysis of linear relations and degenerate operator semigroups 1605

Theorem 6.5. Let B1 ∈ Hom(X, Y ), B2 ∈ Hom(Y, X), and let Ker B2 = {0}.
Then σ(B1 B2 ) \ {0} = σ(B2 B1 ) \ {0}. Moreover, 0 ∈ σ(B2 B1 ) \ σ(B1 B2 ) ⇔
Ker B1 = {0}, Im B1 = Y , Im B2 = Im B2 , and X = Ker B1 ⊕ Im B2 .
Proof. We consider the ordered pair (G, F ) with D = D(G, F ) = Im B2 ⊂ X,
where G = B1 ∈ Hom(X, Y ) and F = B2−1 : Im B2 ⊂ X → Y . Since we have
G − λF = B2−1 (B2 B1 − λI), it follows that λ ∈ ρ(G, F ) once |λ| > B2 B1 . Hence
ρ(G, F ) = ∅ and ∞ ∈ ρ(G, F ). Since Al = B2 B1 and Ar = B1 B2 , the assertion of
the theorem follows from relation (3) in Theorem 6.1 and part (1) of Theorem 6.4.
Note that for elements of Banach algebras the result of Theorem 6.5 is cited in
many monographs (see, for instance, [16], Chapter 1, § 1). On the other hand, they
do not give a detailed analysis of spectral properties similar to the second part of
Theorem 6.5.
In the rest of this section we present results closely connected to results of § 3
and immediately following from them.
Given the left Al and the right Ar relations of the pair (G, F ) we consider the
sequences of linear subspaces

Xk = Akl 0, X(k) = D(Akl ), Yk = Akr 0, Y(k) = D(Akr ), k ∈ N. (6.18)

By properties (1)–(3) of linear relations pointed out in the beginning of the

section we obtain their representations in terms of the operators F and G:

X0 = {0}, X1 = Ker F, . . ., Xn = F −1 (GXn−1), n ∈ N, ...,

−1 −1
X (0)
= X, X (1)
=G (Im F ), . . ., X (n)
=G (F X (n−1)
), n ∈ N, ...,
−1
Y0 = {0}, Y1 = G(Ker F ), . . ., Yn = G(F Yn−1 ), n ∈ N, ...,
−1 (n−1)
Y (0)
= Y, Y (1)
= F (D(G)), . . ., Y (n)
= F (G Y ), n ∈ N, ... .

Clearly, the pairs (Xn , Yn ) and (X(n), Y(n)) of subspaces are invariant with respect
to the operator pair (G, F ).
In the next statement we have m ∈ N, m
2, and all the inclusions are strict.
Theorem 6.6. If Ker F = {0}, then the following conditions are equivalent for an
ordered pair (G, F ):
(1) the point ∞ is a pole of order m − 1 of the resolvent of the left relation Al
of the pair (G, F ) for m
2; ∞ is a removable singularity of it for m = 1;
(2) there exist invariant pairs (X0 , Y0 ) and (X1 , Y1 ) of subspaces such that the
representations (6.11) and (6.12) hold and, in addition,
(G0 , F0 ) = σ(G, F ), σ
(a) σ (G1 , F1 ) = {∞};
(b) D(G0 ) = X0 , F0−1 ∈ Hom(Y0 , X0 ), D(F1 ) = X1 , G−1 1 ∈ Hom(Y1 , X1 );
(c) (G−1
1 F1 )
m−1
= 0, (G−1
1 F1 )
m
= 0;
(3) stabilization of subspaces occurs: Xm−1 ⊂Xm =Xm+1 , X(m−1)⊃Xm =X(m+1) .
Proof. This follows from Theorem 3.1 applied to the left relation Al of (G, F ). One
must also take into account Theorem 6.3 (which is an equivalent of Theorem 2.3
used in the proof of Theorem 3.1) and formula (6.18) for the subspaces Xk , k
1.
1606 A. G. Baskakov and K. I. Chernyshov

§ 7. Linear differential equation with

non-invertible operator coefficient of the derivative
In the last section we consider the Cauchy problem

F ẋ(t) = Gx(t), t ∈ R+ = [0, ∞), (7.1)

x(0) = x0 ∈ X, (7.2)

where the operators F and G satisfy one of conditions (i)–(iii) from the previous
section.
Definition 7.1. A solution of the problem (7.1), (7.2) is a differentiable function
x : R+ → X such that
x(t) ∈ D = D(G) for all t ∈ R+ if condition (i) or (iii) holds,
ẋ(t) ∈ D = D(F ) for all t ∈ R+ if condition (ii) holds,
and, in addition, equality (7.2) holds and F ẋ(t) = Gx(t) for all t ∈ R+ .
We call the closure of the set of vectors x0 ∈ X for which the problem (7.1), (7.2)
is soluble the phase space of equation (7.1) and denote it by Φ(G, F ).
There exist many papers concerned with the solubility of the problem (7.1), (7.2);
many of them are mentioned in [2]. Here we use the theory of linear relations and
linear differential inclusions to obtain most of the results announced in [10].
Alongside (7.1), (7.2) we consider the problem

ẋ(t) ∈ Al x(t) = F −1 Gx(t), t ∈ R+ , (7.3)

x(0) = x0 ∈ X. (7.4)

Lemma 7.1. The problem (7.1), (7.2) is equivalent to the problem (7.3), (7.4), and
Φ(G, F ) = Φ(Al ) ⊂ D(Al ) = G−1 (Im F ).
Proof. Assume first that the pair (G, F ) satisfies one of conditions (i), (iii). If
x : R+ → X is a solution of the problem (7.1), (7.2), then it follows by the equalities
F ẋ(t) = Gx(t) for t
0 that x(t) ∈ G−1 (Im F ) = D(Al ) for all t
0. In particular,
x(0) ∈ D(Al ). Thus, x is a solution of the problem (7.3), (7.4). If x : R+ → X
is a solution of (7.3), (7.4), then x(t) ∈ D(Al ) for all t
0. Hence x(t) ∈ D(G)
for all t
0 and F ẋ(t) = Gx(t), that is, x : R+ → X is a solution of the problem
(7.1), (7.2).
Assume now that the pair (G, F ) satisfies condition (ii) and x : R+ → X is a
differentiable function. Then the following conditions are clearly equivalent:
(1) ẋ(t) ∈ D(F ) for all t
0, F ẋ(t) = Gx(t) for each t
0;
(2) x(t) ∈ D(Al ) = G−1 (Im F ) for all t
0 and ẋ(t) ∈ Al x(t) for all t
0.
Remark 7.1. In the monograph [2], which is remarkable in that there are a large
number of examples of concrete classes of differential equations, the authors consider
the Cauchy problem

dF x(t)
= Gx(t), t ∈ R+ , (7.5)
dt
F x(0) = u0 ∈ X. (7.6)
 Spectral analysis of linear relations and degenerate operator semigroups 1607

Setting y(t) = F x(t) one reduces it to a Cauchy problem for the differential inclusion
with linear relation Ar = GF −1 . The results that we present below for the problem
(7.1), (7.2) can be formulated also for the problem (7.5), (7.6) in terms of the right
relation Ar . At the same time, the results of the previous section give one serious
reason to believe that the properties of the differential inclusion (7.3) with the left
linear relation Al = F −1 G reflect much more fully the properties of the differential
equations in question.
We assume in what follows that the mth power of the left resolvent Al ∈ LR(X)
of the ordered pair (G, F ) has for some m ∈ N the property of minimal growth at
infinity, that is, Assumption 4.1 holds.
We consider now the subspace X  = Erg(X, (An )) corresponding to the bounded
sequence An = I − (−λn R(λn , Al ))m ∈ End X, n ∈ N. The relation adjoint to Al
has the form A∗l = G∗ (F ∗)−1 , that is, it is the right relation of the pair (G∗ , F ∗). In
the statement of the next theorem we use the subspaces X∗m =(X∗ )m =(A∗l )m 0⊂X ∗ .
Acting similarly to the description of this kind of subspace presented at the end
of § 6 we arrive at the sequence

X∗0 = {0}, X∗1 = G∗ (Ker F ∗), . . . , X∗n = G∗((F ∗ )−1 X∗n−1 ), n ∈ N, . . . . (7.7)

As an immediate consequence of Theorem 4.3, taking into account the description

in § 6 of the subspaces Am m
l 0 and D(Al ), we obtain the following result.

Theorem 7.1. Under Assumption 4.1, X  = X if and only if vectors in the sub-
space Xm = Al 0 separate functionals in the subspace (X∗ )m = (A∗l )m 0 ⊂ X ∗ . In
m

 = X if one of the following conditions holds:

particular, X
(1) X is a reflexive Banach space;
(2) R(λ0 , Al ) ∈ End X is a weakly compact operator for some λ0 ∈ ρ(Al );
(3) dim Xm = dim(X∗ )m < ∞.

A similar result involving the right relation Ar and the corresponding conditions
on R( · , Ar ) holds for the Banach space Y . Such a result is necessary for the analysis
of the problem (7.5), (7.6).
In the next statement we assume that the linear relation A = Al ∈ LR(X) of
the pair (G, F ) satisfies Assumption 4.2 and the condition Ker F = {0}, that is,
Al ∈/ LO(X). Hence for Al we have the result of Theorem 4.2 on the decomposi-
tion (4.6) of the space X  and the decomposition (4.7) of the restriction A  l of the

relation Al to X. In what follows we keep the notation of Theorem 4.2, but, for
instance, formula (4.7) must be interpreted as follows: A  l = A0 ⊕ A∞ .

Theorem 7.2. Let Al ∈ LR(X) be a linear relation satisfying Assumption 4.2 and
assume that Ker F = {0}. Then

 = D(Am ) = X (m) = X0
Φ(G, F ) ∩ X l

and there exists a unique degenerate operator semigroup {T(t); t
0} ⊂ End X

  
with generator A ∈ LR(X) defined by the equalities A = Al (the restriction of Al
0
1608 A. G. Baskakov and K. I. Chernyshov

to X0 ), D(A) = X0 , A0 = X∞ . The semigroup {T(t); t
0} has the following

properties:
(1) its restriction {T0 (t); t
0} ⊂ End X0 to X0 is a C0 -semigroup and each
solution x : R+ → X of (7.1), (7.2) with x0 ∈ D(A0l ) ⊂ G−1 (Im F ) has the
form x(t) = T0 (t)x0 , t
0;
(2) T(0) is the projection onto the subspace X0 parallel to X∞ .
If vectors in the subspace X∞ = Am l 0 ⊂ X separate vectors in the subspace
∗ m ∗
(Al ) 0 ⊂ X (or one of the three conditions in Theorem 7.1 holds), then X  =X
and Φ(G, F ) = X0 .
All the assertions of Theorem 7.2 follow from Lemma 7.1 and Theorem 4.2 in
view of the description of the subspaces D(Am ) and Am 0 presented at the end of § 6
(see also formulae (7.7)).
Corollary 7.1. If in place of Assumption 4.2 Al satisfies condition (3) in Theo-
rem 6.6, then all the results of Theorem 7.2 hold with A0 ∈ End X0 .
In a similar fashion one can formulate an analogue of Theorem 7.2, involving
the right relation Ar of the ordered pair (G, F ) in place of the left relation Al ; in
terms of it one constructs a degenerate operator semigroup used for the definition
of solutions of the problem (7.5), (7.6).
Remark 7.2. An analogue of Theorem 7.2 was obtained by Sviridyuk [28] and
Fedorov [12] under the following assumptions about the ordered pair (G, F ) (since
the terminology in [28] and [12] is considerably different from the standard one, we
shall formulate these assumptions using the notation of the present paper):
(1) D(F ) = X, D(G) = X;
(2) (a, ∞) ⊂ ρ(G, F ) for some a ∈ R;
(3) there exists M > 0 such that for all λ = (λ1 , . . . , λm ) ∈ (a, ∞)m and all
n ∈ N,
  n    n 
 m   m 
max  R(λ ; A ) ,  R(λ ; A )   m M .
 k l   k r 
k=1 |λk − a|
n
k=1 k=1

These assumptions are a substitute for Assumption 4.2. If the above conditions
hold, then obviously Assumption 4.2 also holds for Al . The equality X  = X was
established by Fedorov [12] for a reflexive Banach space X under the above-stated
conditions (1)–(3) (more stringent than Assumption 4.1) with n = 1. In [12] the
author also obtained (cumbersome and only sufficient) conditions for the equality
between X  and X for n > 1.
Note that one of the first authors who presented solutions of the problem (7.1),
(7.2) by means of operator semigroups was Rutkas [24]. He considered a pair of
operators (G, F ) with assumptions about the resolvent corresponding to the case
m = 1.
Definition 7.2. † We say that an ordered operator pair (G, F ) is left-sectorial
(right-sectorial) with angle θ ∈ (π/2, π) if the left relation Al (respectively, the
right relation Ar ) is sectorial with angle θ ∈ (π/2, π).
† Editor’s note. This was (erroneously) called “Definition 7.1” in the original Russian text.
 Spectral analysis of linear relations and degenerate operator semigroups 1609

It follows immediately from Definition 7.2 that if an operator pair (G, F ) is

left- (right-)sectorial, then so is the left (right) relation Al (respectively, Ar ) of this
pair (with the same angle). Thus, Lemma 7.1 and Theorem 5.1 yield the following
result.
Theorem 7.3. If a pair of operators (G, F ) is left-sectorial with angle θ ∈ (π/2, π)
(Al satisfies condition (5.1) in Assumption 5.1), then the equality

1
T (z) = − eλz R(λ, Al ) dλ
2πi γ

defines an operator semigroup in the algebra End X that is analytic in the sector
Ω0,δ0 ⊂ ρ(Al ), δ0 = θ − π/2, and bounded for t > 0. In addition, all the results of
Theorem 5.3 hold for the relation A = Al along with relation (1) in Theorem 7.2.
Remark 7.3. A similar result holds for right-sectorial pairs (G, F ).
Remark 7.4. The result closest to Theorem 7.3 is established in [11]. There in the
definition of a sectorial pair (G, F ) it is required that the product of operators
that are the values of the left and right resolvents of the pair (G, F ) satisfy certain
estimates of the same type as conditions (1)–(3) in Remark 7.2 in the corresponding
sector.

Bibliography
[1] R. Cross, Multivalued linear operators, Dekker, New York 1998.
[2] A. Favini and A. Yagi, Degenerate differential equations in Banach spaces, Dekker, New
York 1998.
[3] V. S. Ritsner, “Theory of linear relations”, Manuscript no. 846–82 deposited at VINITI
(1982).
[4] E. Hille and R. Phillips, Functional analysis and semigroups, Amer. Math. Soc., Providence,
RI 1974.
[5] N. Dunford and J. Schwartz, Linear operators. I. General theory, Interscience, New York
1958.
[6] J. von Neumann, “Über adjungierte Funktionaloperatoren”, Ann. of Math. (2) 33 (1932),
294–310.
[7] W. Arendt, “Approximation of degenerate semigroups”, Tübinger Berichte zur
Funktionalanalysis 9 (1999/2000), 33–46.
[8] S. G. Kreın (ed.), Functional analysis, Nauka, Moscow 1972; English transl. of part of
papers by M. Sh Birman, N. Ya. Vilenkin, E. A. Gorin, et al., Functional analysis, Wolters-
Noordhof, Groningen 1972.
[9] A. G. Baskakov and K. I. Chernyshov, “Ordered operator pairs and semigroups”, Izv. Ross.
Akad. Estestv. Nauk Mat. Mat. Model. Inform. Uprav. 2:3 (1998), 39–69. (Russian)
[10] A. G. Baskakov and K. I. Chernyshov, “Construction of the phase space and solutions for
differential equations not resolved with respect to the derivative”, Dokl. Ross. Akad. Nauk
371 (2000), 295–298; English transl. in Dokl. Math. 61 (2000).
[11] G. A. Sviridyuk and V. E. Fedorov, “On the identities of analytic semigroups of operators
with kernels”, Sibirsk. Mat. Zh. 39:3 (1998), 604–616; English transl. in Siberian Math. J.
39 (1998).
[12] V. E. Fedorov, “Degenerate strongly continuous semigroups of operators”, Algebra i Analiz
12:3 (2000), 173–200; English transl. in St. Petersburg Math. J. 12 (2001).
[13] I. V. Mel’nikova and A. V. Gladchenko, “Well-posedness of the Cauchy problem for inclusions
in Banach spaces”, Dokl. Ross. Akad. Nauk 361 (1998), 736–739; English transl. in Dokl.
Math. 58 (1998).
1610 A. G. Baskakov and K. I. Chernyshov

[14] K.-J. Engel and R. Nagel, One-parameter semigroups for linear evolution equations,
Springer-Verlag, New York 2000.
[15] F. Räbiger and M. P. H. Wolff, “Spectral and asymptotic properties of resolvent-dominated
operators”, Tübinger Berichte zur Funktionalanalysis 7 (1997/1998), 217–235.
[16] N. Bourbaki, Éléments de mathématique. Fasc. XXXII. Théories spectrales, Hermann, Paris
1967.
[17] A. Favini and A. Yagi, “Multivalued linear operators and degenerate evolution equations”,
Ann. Mat. Pura Appl. (4) 163 (1993), 353–384.
[18] A. G. Baskakov and K. I. Chernyshov, “Conditions for compactness of the spectrum of
ordered pairs of linear operators”, Izv. Ross. Akad. Estestv. Nauk Mat. Mat. Model. Inform.
Uprav. 3:3 (1999), 5–24. (Russian)
[19] A. Pietsch, Operator ideals, North-Holland, Amsterdam 1980.
[20] A. G. Baskakov, “Operator ergodic theorems and complementability of subspaces of Banach
spaces”, Izv. Vyssh. Uchebn. Zaved. Mat. 11(318) (1988), 3–11; English transl. in Soviet
Math. (Iz. VUZ ) 32 (1988).
[21] S. G. Kreı̌n, Linear differential equations in Banach space, Nauka, Moscow 1967; English
transl., Amer. Math. Soc., Providence, RI 1971.
[22] D. Henry, Geometric theory of semilinear parabolic equations, Springer-Verlag, Berlin 1981.
[23] N. I. Radbel’, “Ordered pairs of linear operators and the Cauchy problem for the equation
Aẋ(t) + Bx(t) = 0 in Banach space”, Kandidat thesis, Institute of Applied Mathematics and
Mechanics, Donetsk 1984. (Russian)
[24] A. G. Rutkas, “Cauchy’s problem for the equation Ax
(t) + Bx(t) = f (t)”, Differentsial’nye
Uravneniya 11 (1975), 1996–2010; English transl. in Differential Equations 11 (1975).
[25] K. I. Chernyshov, “Operator differential equations not resolved with respect to the
derivative”, Kandidat Thesis, Institute of Mathematics, Academy of Sciences of UkrSSR,
Kiev 1979.
[26] S. G. Kreı̌n and K. I. Chernyshov, “Singularly perturbed differential equations in Banach
space”, IX International conference on non-linear oscillations, vol. 1, Naukova Dumka, Kiev
1984, pp. 193–197. (Russian)
[27] V. V. Litkin, “Certain spectral properties of a pencil of linear bounded operators”, Mat.
Zametki 31:1 (1982), 75–79; English transl. in Math. Notes 31 (1982).
[28] G. A. Sviridyuk, “On the general theory of operator semigroups”, Uspekhi Mat. Nauk 49:4
(1994), 47–74; English transl. in Russian Math. Surveys 49 (1994).
[29] A. G. Baskakov and K. I. Chernyshov, “Ergodic subspaces and analytic semigroups”, Spectral
and evolutionary problems, vol. 11 (11th Crimean Autumn Math. School), National Taurida
University, Simferopol 2001, pp. 136–143.

Voronezh State University;

Voronezh State Academy of Forestry Engineering
E-mail address: [email protected], [email protected]
Received 12/MAR/02
Translated by IPS(DoM)

Typeset by AMS-TEX