See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/231901096

A generalized Drazin inverse

Article in Glasgow Mathematical Journal August 1996
Impact Factor: 0.33 DOI: 10.1017/S0017089500031803

CITATIONS

READS

139

256

1 author:
J. J. Koliha
University of Melbourne
94 PUBLICATIONS 1,533 CITATIONS
SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate,

letting you access and read them immediately.

Available from: J. J. Koliha

Retrieved on: 10 June 2016

A generalized Drazin inverse

J. J. Koliha, University of Melbourne

Abstract
The paper studies the Drazin inverse in associative rings and in Banach
algebras with aplications to bounded linear operators on a Banachspace. The
Drazin inverse in a Banach algebra is dened for all elements for which 0
is not an accumulation point of the spectrum. Properties of the inverse are
discussed, and series and exponential representations are derived.

Introduction

The main theme of this paper can be described as a study of the Drazin inverse for
bounded linear operators in a Banach space X when 0 is an isolated spectral point
of the operator. This inverse is useful for instance in the solution of dierential
equations formulated in a Banach space X. Since the elements of X rarely enter
into our considerations, the exposition seems to gain in clarity when the operators
are regarded as elements of the Banach algebra L(X).
An element a of a complex Banach algebra A is called regular (or relatively
regular ) if there is x A such that axa = a. Relatively regular elements have
been extensively researched in the case that A is the Banach algebra L(X) of all
bounded linear operators on a complex Banach space X; they have been shown to
generalize in certain aspects invertible operators.
If a is relatively regular, then it has a generalized inverse, which is an element
b A satisfying the equations aba = a and bab = b (see [18] for a comprehensive
account of generalized inverses). A relation between a relatively regular element
and its generalized inverse is reexive in the sense that if b is a generalized inverse
of a, then a is a generalized inverse of b.
In 1958 Drazin [7] introduced a dierent kind of a generalized inverse in associative rings and semigroups, one which does not have the reexivity property, but
commutes with the element:
1

J. J. Koliha

Denition 1.1. Let a, b be elements of a semigroup. An element b is a Drazin

inverse of a, written b = ad , if
ab = ba,

b = ab2 ,

ak = ak+1 b

(1.1)

for some nonnegative integer k. The least nonnegative integer k for which these
equations hold is the Drazin index i(a) of a.
The Drazin inverse is an important tool in ring theory (see for instance Hartwig
[12]). In spite of not being reexive, the Drazin inverse is very useful in matrix
theory and computations and in various applications of matrices [4, 5, 1, 21, 19],
primarily because it has a very desirable spectral property: The nonzero eigenvalues
of the Drazin inverse are the reciprocals of the nonzero eigenvalues of the given
matrix, and the corresponding generalized eigenvectors have the same grade [1,
p.167]. A generalized inverse a
of an element a of a complex Banach algebra
with the property that the nonzero spectrum of a
consists of the reciprocals of
the nonzero spectral points of a is often called a spectral inverse [6]. The inverse
matrix is a (commuting) spectral inverse.
Caradus [6], King [14] and Lay [16] investigated the Drazin inverse in the setting
of bounded linear operators on complex Banach spaces. Caradus [6] proved that
a bounded linear operator T on a complex Banach space has a Drazin inverse if
and only if 0 is a pole of the resolvent (I T )1 of T ; the order of the pole
y [17]
is equal to the Drazin index of T . (See also King [14].) Marek and Zitn
discuss the Drazin inverse in detailfor operators as well as for elements of a
Banach algebra. Harte [9] associated with each quasipolar operator T an operator
T , which is an equivalent of the generalized Drazin inverse studied here. He also
investigated quasipolar elements of a normed algebra [10] and extended the concept
to this setting. In [11], Harte oered two candidates for quasinilpotent elements
in a general associative ring, and introduced the concept of a quasipolar element
and its Drazin inverse. Nashed and Zhao [19] investigated the Drazin inverse for
closed linear operators and applied it to singular evolution equations and partial
dierential operators. Drazin [8] investigated extremal denitions of generalized
inverses that give a generalization of the original Drazin inverse.
In the theory of the Drazin inverse for matrices [1, 4] an important role is played
by the index of a matrix A, which is the least nonnegative integer k for which the
nullspaces of Ak and Ak+1 coincide. The index of a matrix coincides with its Drazin
index [4].

A generalized Drazin inverse

In operator theory, the concept corresponding to the index of a nite matrix is

the ascent (and descent) of a chain-nite bounded linear operator T [13, 22]. An
operator T is chain-nite with the ascent (=descent) k if and only if 0 is a pole of
the resolvent (I T )1 of order k [13, Proposition 50.2]. If we want to translate
results involving the index or the chain-niteness condition to a Banach algebra A
we must interpret the index of a A to be 0 if a is invertible, and k if 0 is a pole
of (e a)1 of order k.

The Drazin inverse in a ring

In this section A is an associative ring with a unit e. First we give an equivalent

denition of the Drazin inverse. By Inv(A) and N(A) we denote the set of all
invertible and nilpotent elements in A, respectively.
Lemma 2.1. In a ring A with unit, (1.1) is equivalent to
ab = ba,

b = ab2 ,

a a2 b N(A).

(2.1)

The Drazin index i(a) is equal to the nilpotency index of a a2 b.

Proof. Suppose that the rst two equations in (2.1) hold. Then the element p =
e ab is an idempotent as (ab)2 = a(ab2 ) = ab. So
ak ak+1 b = ak p = (ap)k = (a(e ab))k = (a a2 b)k
for any k 1, and (1.1) is equivalent to (2.1).
Harte [11] dened quasinilpotent (and quasi-quasinilpotent) elements in an arbitrary associative ring A. This concept can be used to dene a generalized Drazin
inverse in rings.
Denition 2.2. [11, Denition 2] An element a of a ring A is quasinilpotent if, for
every x commuting with a, we have e xa Inv(A). The set of all quasinilpotent
elements in a ring A will be denoted by QN(A).
Note that N(A) QN(A). In a Banach algebra A, the preceding denition
coincides with the usual denition an 1/n 0, which is equivalent to e a
Inv(A) for all complex = 0.

J. J. Koliha

Denition 2.3. Let a A. An element b A is a Drazin inverse b = aD of a if

ab = ba,

a a2 b QN(A).

b = ab2 ,

(2.2)

The Drazin index i(a) of a is the nilpotency index of a a2 b if a a2 b N(A),

and i(a) = otherwise.
The inverse ad introduced by Denition 1.1 will be called the nite index Drazin
inverse.
Lemma 2.4. In a ring A with unit, an element a has a Drazin inverse aD if and
only if there is an idempotent p commuting with a such that
ap QN(A),

a + p Inv(A).

(2.3)

The Drazin inverse aD is unique and is given by

aD = (a + p)1 (e p).

(2.4)

Proof. Suppose that there is an idempotent p commuting with a and satisfying

(2.3). Set b = (a + p)1 (e p). Then ab = ba,
ab = a(a + p)1 (e p) = (a + p)(a + p)1 (e p) = e p,
and ab2 = (ab)b = (ep)b = b(ep) = b. Finally, aa2 b = a(eab) = ap QN(A),
so that (2.2) hold.
Conversely, suppose that b satises (2.2). Set p = eab. Since (ab)2 = a(ab2 ) =
ab, p is an idempotent commuting with a, and ap = a a2 b QN(A). Further,
(a + p)(b + p) = ab + ap + bp + p = e + ap Inv(A)
as ap QN(A); then also a + p Inv(A). From (a + p)b = e p it follows that
b = (a + p)1 (e p), which proves the uniqueness of b.
We compare our denition of the Drazin inverse with that given by Harte [11]
for quasipolar elements of a ring A: An element a of a ring A is quasipolar [11,
Denition 5] if there exists b A such that
b comm2 (a),

ab = (ab)2 ,

a(e ab) QN(A);

(2.5)

here comm2 (a) denotes the double commutant of a. An element b satisfying (2.5)
and b = ab2 is called a Drazin inverse of a in [11]. Clearly, a Drazin inverse in the

A generalized Drazin inverse

sense of Harte is also a Drazin inverse in the sense of Denition 2.3. Conversely
however, it cannot be guaranteed that the Drazin inverse b = aD dened according
to Denition 2.3 is in the double commutant of a. If A is not only a ring, but
a Banach algebra, this fact follows from topological and spectral considerations
discussed in the next section (see also [11, Theorem 7.5.3]).

Isolated spectral points

From this point on, A denotes a complex unital Banach algebra with unit e; Inv(A),
QN(A) and N(A) denote the set of all invertible elements of A, the set of all
quasinilpotent elements of A and the set of all nilpotent elements of A, respectively.
For each a A, (a) denotes the spectrum of a, (a) the resolvent set of a and
r(a) the spectral radius of a. We write iso (a) for the set of all isolated spectral
points of a and acc (a) for the set of all accumulation points of (a). By H(a) we
denote the set of all complex valued functions f , each dened and holomorphic in
an open neighbourhood (f ) of (a). If is an open subset of the complex plane
and K a compact subset of , we dene a cycle to be a (, K)-cycle if lies
in \K, ind (; ) = 0 for every
/ , and ind (; ) = 1 for every K. The
functional calculus for a A is dened for any function f H(a) by

1
f ()(e a)1 d,
f (a) =
2i
where is a ((f ), (a))-cycle. We recall that, in commutative Banach algebras,
the spectral radius is submultiplicative, subadditive and continuous. This fact will
be used throughout the paper usually without a further comment. For an element
a A we dene the spectral index s-ind (a) to be 0 if a is invertible and k if 0 is a
pole of order k of the resolvent (ea)1 . (We do not use the word index because
in operator theory it is reserved for the dierence dim T 1 (0) codim T (X), where
T is a bounded linear operator on a Banach space X [13].) For spectral theory and
functional calculus in Banach algebras see Heuser [13, Chapter VII].
We start with a characterization of isolated spectral points of an element of
A that will provide the main tool for the development of the generalized Drazin
inverse. The result is a special case of a theorem obtained by the author [15].
Theorem 3.1. Let a A. Then 0
/ acc (a) if and only if there is an idempotent
p A commuting with a such that
ap QN(A),

p + a Inv(A).

(3.1)

J. J. Koliha

Moreover, 0 iso (a) if and only if p = 0, in which case p is the spectral idempotent
of a corresponding to = 0.
Proof. Clearly, a A is invertible if and only if (3.1) holds with p = 0.
Let 0 iso (a). The spectral idempotent p of a is dened by p = f (a),
where f H(a) is such that f = 1 in a neighbourhood of 0 and f = 0 in a
neighbourhood of (a)\{0}. Then p2 = p = 0, p commutes with a, and ap = h(a),
where h() = f (); as (ap) = (h(a)) = h((a)) = {0} by the spectral mapping
theorem, ap QN(A). The function g() = f () + is in H(a) and is nonzero on
the spectrum of a; so g(a) = p + a Inv(A).
Conversely, assume that there is a nonzero idempotent p commuting with a
such that (3.1) hold. For any ,
e a = (e ap)p + (e (p + a))(e p).
There is r > 0 (e.g. r = (p + a)1 1 ) such that e (p + a) Inv(A) if || < r.
Since ap QN(A), e ap Inv(A) for all = 0. So
(e a)1 = (e ap)1 p + (e (p + a))1 (e p)

(3.2)

whenever 0 < || < r. Since p = 0, 0 iso (a). To show that p is the spectral
idempotent of a corresponding to a, we pick a function f H(a) with f = 1 in a
neighbourhood of 0 and f = 0 in a neighbourhood of (a)\{0}. By (3.2),

1
f (a) =
(e a)1 d
2i


1
1
1
(e ap) p d +
(e (p + a))1 (e p) d
=
2i
2i
 

1
=
n1 an p + 0 = p
2i
n=0

where is a ({ : || < r}, {0})-cycle.

Harte [10] obtained a similar result using the concept of a quasipolar element
in a Banach algebra. An element a of a Banach algebra A is quasipolar [10, p. 257]
if there is an idempotent q A commuting with a such that
a(e q) QN(A),

q (Aa) (aA).

(3.3)

We give a proof using the preceding characterization of isolated spectral points.

A generalized Drazin inverse

Theorem 3.2. (Harte [10, Theorem 9.7.6]) An element a A is quasipolar if

and only if 0
/ acc (a).
Proof. Suppose that a is quasipolar, satisfying (3.3). Write p = e q. Then there
are u, v A such that e p = ua = av. A simple calculation shows that
(a + p)(uav + p) = e + ap = (uav + p)(a + p).
Since e + ap Inv(A), we have a + p Inv(A). Then 0
/ acc (a) by the preceding
theorem.
Conversely, if 0
/ acc (a), the conclusions of the preceding theorem hold. Since
(a + p)(e p) = a(e p), we have
a(a + p)1 (e p) = e p = (a + p)1 (e p)a.
Then (3.3) holds for q = e p.
Note 3.3. It is often useful to know that (3.1) can be replaced by
ap QN(A),

a + p Inv(A), = 0.

Changes to the preceding proof are obvious.

Note 3.4. In view of the Laurent expansion for the resolvent (e a)1 , Theorem 3.1 gives the following criterion: 0 is a pole of (e a)1 if and only if there
is a nonzero idempotent p commuting with a such that
ap N(A),

a + p Inv(A).

The Drazin inverse in a Banach algebra

In this section A is a complex unital Banach algebra. By Inv(A) and QN(A) we

denote the sets of all invertible and quasinilpotent elements of A, respectively. By
denition, an element x of a Banach algebra A is quasinilpotent if xn 1/n 0.
As mentioned earlier, this is equivalent to the Harte denition of a quasinilpotent
(Denition 2.2).
Denition 4.1. Let a A. An element b A is called a Drazin inverse of a, and
is denoted by aD , if it satises conditions
ab = ba,

b = ab2 ,

a a2 b QN(A).

(4.1)

J. J. Koliha

If 0 (a), that is, if a is invertible, then aD = a1 . Even though the dening

relations (4.1) in a Banach algebra are the same as the dening relations (2.1) in a
ring, topology enters here via the denition of quasinilpotents which is dependent
on the norm of A. In particular, we shall see that in a Banach algebra the Drazin
inverse of a is in the double commutant of a, which is not always the case in a ring.
Theorem 4.2. The following conditions on an element a A are equivalent:
(i) 0
/ acc (a),
(ii) there is an idempotent p A commuting with a satisfying (3.1),
(iii) a has a Drazin inverse.
In this case the Drazin inverse is unique, and is given by
aD = (a + p)1 (e p),

(4.2)

where p is the spectral idempotent of a corresponding to 0.

Proof. Conditions (i) and (ii) are equivalent by Theorem 3.1, and (ii) and (iii)
are equivalent by Lemma 2.4. The explicit formula (4.2) was also obtained in
Lemma 2.4.
Harte obtained the following result, dening in eect a Drazin inverse a for a
quasipolar element a A (see the denition following Theorem 3.1). We give an
alternative proof based on our denition of the Drazin inverse.
Theorem 4.3. Harte [10, Theorem 7.5.3] If a A is quasipolar, then there are
unique elements a,
a A such that
(e a)a
= a(e a)
QN(A),

a a = aa = a = a 2 ,

a a = aa
= a . (4.3)

Proof. A quasipolar element a satises 0

/ acc (a) (Theorem 3.2). Let p be the
spectral idempotent of a at = 0, and dene a = e p, a = aD (exists by the
preceding theorem); (4.3) is then veried using Theorems 3.1 and 4.2, which proves
the existence. Further, if a and a satisfy (4.3), then
= 0,
a a(a )2 = a (e aa ) = a (e a)
and
QN(A).
a a2 a = a(e aa ) = a(e a)
Then a = aD and a = aaD = e p, which proves the uniqueness.

A generalized Drazin inverse

The explicit formula (4.2) was given for matrices by Rothblum [20] in the form
= (a p)1 (e p). It can be deduced from the next theorem that aD satises
a more general equation
ad

aD = (a + p)1 (e p),

= 0.

(4.4)

Theorem 4.2 shows that, in the case that 0 is a pole rather than just an isolated
spectral point of (ea)1 , the Drazin inverse aD reduces to the nite index Drazin
inverse ad . In this case, the Drazin index of a is given by
i(a) = s-ind (a).
The next result generalizes [17, Theorem 13.4]. It shows that aD is a spectral
inverse of a which lies in the double commutant of a.
Theorem 4.4. If 0 iso (a), then
aD = f (a),
where f H(a) is such that f = 0 in a neighbourhood of 0 and f () = 1 in
a neighbourhood of (a)\{0}. Hence aD commutes with every element of A that
commutes with a, and
(aD )\{0} = {1 : (a)\{0}}.

(4.5)

Proof. Write the spectral idempotent p of a at 0 as p = g(a), where g H(a),

g = 1 in a neighbourhood of 0, g = 0 in a neighbourhood of (a)\{0}, and set
f () = ( + g())1 (1 g()).
By (4.2), aD = (a + p)1 (e p) = f (a); it is easy to check that f has the required
property. Equality (4.5) follows from the spectral mapping theorem.

Properties of the Drazin inverse

We start with a version of the Laurent series for the resolvent of an element a
of a Banach algebra A in a neighbourhood of an isolated spectral point 0. The
theorem generalizes a result of Caradus [6] which was obtained for operators under
the additional assumption that 0 is a pole of the operator resolvent.

10

J. J. Koliha

Theorem 5.1. Let 0 iso (a), and let b be the Drazin inverse of a. Then, on
some punctured disc 0 < || < r,
1

(e a)

n n1

(e ab)

n=1

n bn+1 .

(5.1)

n=0

Proof. Let p be the spectral idempotent of a corresponding to = 0; then b =

(a + p)1 (e p) by (4.2). In some disc || < r, e (a + p) is invertible; ap is
quasinilpotent. According to (3.2), if 0 < || < r, then
(e a)1 = (e ap)1 p + (e (a + p))1 (e p)



n n1
a
p
n (a + p)n1 (e p)
=
=

n=1

n=0

n n1

(e ab)

n=1

n bn+1 .

n=0

/ Inv(A). Then
Theorem 5.2. Let b = aD be the Drazin inverse of an element a
s-ind (b) = 1.
Proof. The element p = e ab is the (nonzero) spectral idempotent of a corresponding to = 0 with ap QN(A) and a + p Inv(A). Then
bp = b(e ab) = b ab2 = 0,
and
(a + p)(b + p) = ab + ap + p = e + ap Inv(A),
so that b + p Inv(A). By Theorem 3.1, 0 iso (b) with the spectral idempotent
p; since bp = 0, = 0 is a simple pole of (e b)1 .
Theorem 5.3. Let a
/ Inv(A). Then (aD )D = a if and only if s-ind (a) = 1.
Proof. Suppose that b = aD and that 0 is a simple pole of (e a)1 with the
spectral idempotent p. Then a a2 b = a(e ab) = ap = 0, and
ab = ba,

b ab2 = 0,

a a2 b = 0.

This shows that bD = a. Conversely, if aD = b and bD = a, then the preceding

equations hold and signify that 0 is a simple pole of (e a)1 .

A generalized Drazin inverse

11

In the case that 0 is a pole of (e a)1 , the Drazin inverse aD and the nite
index Drazin inverse ad of a coincide; if 0 is a simple pole, ad is then called the
group inverse of a, and in matrix theory it is usually denoted by a# . The name
was chosen because the elements a and a# generate a multiplicative Abelian group
in A; the idempotent aa# plays the role of the group identity.
The properties of the Drazin inverse stated in the following theorem are easily
veried by using Theorems 3.1, 4.2 and 4.4.
Theorem 5.4. Suppose that a A has the Drazin inverse aD and that p is the
spectral idempotent of a corresponding to 0. Then
(i) (an )D = (aD )n for all n = 1, 2, . . . ,
(ii) (aD )D = a2 aD = a(e p),
(iii) ((aD )D )D = aD ,
(iv) aD (aD )D = aaD = e p.
The next result deals with the inverse of a product, and gives an indication of
the usefulness of the Drazin inverse as it provides a short proof of the fact that 0 is
an isolated spectral point of ab if a, b commute and 0 is an isolated spectral point
for both a and b. The result is well known for matrices [4, Theorem 7.8.4].
Theorem 5.5. Let a, b be commuting elements of A such that aD and bD exist.
Then (ab)D exists and
(ab)D = aD bD .
Proof. The elements a, b, aD , bD all commute. Then
ab(aD bD )2 = a(aD )2 b(bD )2 = aD bD .
Further,
ab (ab)2 aD bD = (a a2 aD )(b b2 bD ) + a2 aD (b b2 bD ) + b2 bD (a a2 aD )
and r(ab (ab)2 aD bD ) = 0 as the spectral radius is subadditive and submultiplicative in commutative subalgebras of A. So ab (ab)2 aD bD QN(A).

12

J. J. Koliha

This result is implicit in Harte [10, Theorem 7.5.4], whereit is proved that the
product of commuting quasipolar elements is quasipolar.
The proof of the preceding theorem can be adapted to show that the product
of two commuting elements of a Banach algebra with a nite Drazin index has a
nite index Drazin inverse. Indeed, it is enough to observe that the product of
commuting elements, one of which is nilpotent, is also nilpotent, and the sum of
commuting nilpotent elements is nilpotent.
Bouldin [2] gives an example of two noncommuting bounded linear operators
A, B on a complex Banach space which have the nite index Drazin inverses Ad ,
B d , but whose product AB is not Drazin invertible [2, Example 4.1]. Bouldin
also gives some conditions under which the equation (AB)d = B d Ad holds in the
noncommutative case.
Theorem 5.6. If aD exists, b QN(A) and ab = ba, then also (a + b)D exists and
(a + b)D = (a + b + p)1 (e p),
where p is the spectral idempotent of a at = 0.
Proof. Let p be the spectral idempotent of a at = 0; the set {a, b, p} is commutative, a + p Inv(A) and ap QN(A). Then
a + b + p = (a + p) + b Inv(A),

(a + b)p = ap + bp QN(A).

So 0
/ acc (a + b) by Theorem 3.1, and the explicit formula for the Drazin inverse
of a + b follows from (4.2).
The preceding result shows that if 0 iso (a), b QN(A) and ab = ba, then
0 iso (a+b). This is implicitly contained in Harte [10, Theorem 7.5.4] formulated
in terms of quasipolarity.
Theorem 5.7. If aD , bD exist and ab = ba = 0, then also (a + b)D exists and
(a + b)D = aD + bD .
Proof. As in [7] we observe that the elements a, b, aD , bD all commute, that abD =
ab(bD )2 = 0 and that aD b = ab(aD )2 = 0. Hence
(a + b)(aD + bD )2 = a(aD )2 + b(bD )2 = aD + bD ,

A generalized Drazin inverse

13

and
(a + b) (a + b)2 (aD + bD ) = (a a2 aD ) + (b b2 bD ) QN(A),
which shows that (a + b)D = aD + bD .

Representations of the Drazin inverse

From the Laurent expansion we get a limit representation of the Drazin inverse in
a Banach algebra A:
Theorem 6.1. Let 0 iso (a) with the corresponding spectral idempotent p. Then
aD = lim (a e)1 (e p).
0

(6.1)

Proof. If we multiply both sides of (5.1) by e p we get

(e a)1 (e p) =

n bn+1 ,

n=0

where b = aD . The result then follows.

We give a representation of the Drazin inverse in two special cases. For the rst
case we recall that much of the elementary spectral theory in a Banach algebra A
depends on the expansion
(e a)1 =

an

n=0

valid in the case that an 0. There is a generalization of this expansion in the

case that an p, where p need not be 0.
Theorem 6.2. Let an p in A. Then s-ind (e a) 1 and
(e a) =
D

an (e p).

(6.2)

n=0

Proof. We observe that p2 = p; if c is dened by c = a p, then cn = (a p)n =

an p 0, and cp = pc = 0. Further, p is an idempotent commuting with e a
such that (e a)p = 0 and e a + p = e c Inv(A); by Theorem 3.1, = 0

14

J. J. Koliha

is a resolvent point of e a or a simple pole of its resolvent. As an (e p) =

(cn p)(e p) = cn (e p),

an (e p) =

n=0

cn (e p) = (e c)1 (e p)

n=0

= (e a + p)1 (e p) = (e a)D .

For the second case note that exp(ta) 0 (as t ) implies


1
a =
exp(ta) dt.
0

This generalizes to elements a for which exp(ta) converges, but not necessarily to
0.
Theorem 6.3. Let exp(ta) p as t . Then s-ind (a) 1, and

D
a =
exp(ta)(e p) dt.

(6.3)

Proof. We have p2 = p. If c = a p, then

exp(tc) = exp(ta) exp(tp) = exp(ta)(e p + et p) p(e p) = 0;
hence (c) lies in the left half-plane, and c is invertible. Further,

1 s
ap = a lim
exp(ta) dt = lim (exp(sa) e) = 0,
s s 0
s
which implies pc = cp = p. To conclude that 0
/ acc (a), it is enough to observe
that a p = c Inv(A) (Note 3.3).
Finally, since exp(ta)(e p) = exp(tc)(e p) and since exp(tc) 0 as t ,
we have



exp(ta)(e p) dt =
exp(tc)(e p) dt = (e p)
exp(tc) dt
0

= c1 (e p) = (a p)1 (e p)
= aD
by (4.4).

A generalized Drazin inverse

15

The following theorem generalizes the so-called corenilpotent decomposition

A = A(I P ) + AP of a square matrix [1, p.175] to elements of a Banach algebra.
An operator version was given by King [14] in the special case that 0 is a pole of
the operator resolvent.
Theorem 6.4. Let a A. Then 0 iso (a) if and only if there are x, y A such
that
a = x + y,

xy = yx = 0,

s-ind (x) = 1,

y QN(A).

(6.4)

Such a decomposition is unique.

Proof. If (6.4) hold, then aD = (x + y)D = xD + y D = xD by Theorem 5.7. So
0 iso (a) by Theorem 4.2. The decomposition is unique as x = (xD )D = (aD )D
by Theorem 5.3.
Conversely, let 0 be an isolated spectral point of a with the corresponding
spectral idempotent p, and let
x = a(e p),

y = ap.

Then y = ap QN(A), xp = 0 and x + p = (a + p) ap Inv(A) (invertible minus

quasinilpotent). The last two conditions imply that s-ind (x) = 1.
The preceding result shows that the calculation of the Drazin inverse can be
reduced to the calculation of the nite index Drazin inverse, in fact of the group
inverse, as
aD = (a(e p))# .
Thus we have a complete description of elements which possess aD :
Corollary 6.5. An element a of a Banach algebra A has the Drazin inverse aD
if and only if it is the sum of an element x with a nite Drazin index and a
quasinilpotent element y such that xy = yx = 0. Then
aD = xd .
In addition, aD = ad if and only if y is nilpotent.

16

J. J. Koliha

The Drazin inverse for bounded linear operators

Let X be a complex Banach space and L(X) the Banach algebra of all bounded
linear operators on X with the operator norm
A = sup Ax.
x=1

We write N (A) for the null space of A L(X), and R(A) for its range. The
following result was obtained by Lay [16] for the nite index Drazin inverse.
Theorem 7.1. Let A L(X) be such that 0
/ acc (A). If AD is the Drazin
inverse of A as an element of the algebra L(X), then X = R(AD A) N (AD A),
A = A1 A2 with respect to this direct sum and
AD = A1
1 0.
Proof. Let P be the corresponding spectral idempotent, so that P = I AD A,
and X = N (P ) R(P ) = R(AD A) N (AD A). From Theorem 3.1 we can deduce
that A = A1 A2 with A1 invertible in L(R(AD A)) and A2 quasinilpotent in
L(N (AD A)). Observe that P = 0 I. Then by (4.2)
AD = (A + P )1 (I P ) = (A1 (A2 + I))1 (I 0) = A1
1 0.

Campbell [3] gives an alternative generalization of the Drazin inverse of bounded

linear operators. Suppose that A L(X) is such that the hyperrange
R(A ) =

R(An )

n=0

is closed and complemented by a closed A-invariant subspace M . Then

A=U V
with respect to the decomposition X = R(A ) M . If U is invertible, Campbell
denes
A = U 1 0.
Then
AA = A A,

A = A(A )2 ,

A A2 A = 0 V.

(7.1)

A generalized Drazin inverse

17

/ iso (A); in this case A need not be unique [3, Example

A may exist when 0
3].
If 0 is a pole of (I A)1 , then R(AD A) = R(Ak ) = R(A ) for k = s-ind (A).
In this case the inverses A , AD and Ad all exist and are equal.
If 0 is an essential singularity of (I A)1 , then AD exists, but A need not. If,
in addition, R(A ) = R(AD A), then A exists and A = (A1 A2 ) = A1
1 0=
D
A .
From (7.1) we see that if A exists and V is quasinilpotent, then 0 iso (A) and
AD = A . If AD and A both exist and are equal, then V in (7.1) is quasinilpotent.
Note that R(A ) = R(AD A) whenever AD = A .
We close the section with an application to stability of dierential equations in
Banach spaces which generalizes a result stated in [17, p.111].
Theorem 7.2. Suppose A L(X) is such that (An ) converges, and let
B = (I + A)(A I)D .
Then the dierential equation
x(t)

= Bx(t) + u,

x(0) = u

(t 0, u X)

has a steady state solution as t , that is, limt x(t) exists.

Proof. In the proof of Theorem 6.2 we established that the limit P of (An ) is
the spectral idempotent of A corresponding to = 1. Write P = g(A), where
g H(A) is such that g = 1 in a neighbourhood of 1 and g = 0 in a neighbourhood
of (A)\{1}. Dene also f H(A) such that f is equalto 0 in a neighbourhood of
1 and to (1 + )/( 1) in a neighbourhood of (A)\{1}. Then B = f (A). We
show that
BP = 0,

B + P Inv(L(X)).

Indeed, BP = f (A)g(A) = (f g)(A) = 0 as f g = 0. Also, f + g = 0 on (A), so

that B + P = f (A) + g(A) is invertible. This ensures that 0 is a simple pole of the
resolvent of B. The function f maps the open unit disc D onto the open left half
plane H of C. Since (A)\{1} D, the spectral mapping theorem implies that
(B)\{0} = f ((A)\{1}) f (D) H.
The condition (B)\{0} H together with s-ind (B) 1 guarantees that exp(sB)
converges as s . This proves the result as the solution to the dierential
equation above is given by x(t) = exp(tB)u.

18

J. J. Koliha

Examples

Example 8.1. Dene an operator B on the Banach space /1 by the innite matrix

0
1
0
0
0
...

0
0

0
0
0

1
2

0
0

1
3

0
...

0
0
0
0
1
4

...
...
...
...
...
...

Then B is a quasinilpotent (but not nilpotent) element of the Banach algebra L(/1 )
of all bounded linear operators on /1 [22, p.280]. The nite index Drazin inverse
B d does not exist, the Drazin inverse B D is the zero operator.
Example 8.2. Dene a diagonal operator A on the Banach space /1 by the innite
matrix

0
0
0
0 ...
0 2 0
0 ...

0
0 3 0 . . .

,
0
0
0 4 . . .
...
...
...
where 0 < |n | M for all n = 2, 3, . . . Then A is an element of the Banach
algebra L(/1 ) with the spectrum (A) = {0} cl {n : n = 2, 3, . . . }, and, for any
f H(A),
f (A)(n ) =

n n en ,

n=2

where e1 , e2 , . . . is the standard Schauder basis in /1 [22]. It follows that 0 is a

simple pole of A with the spectral idempotent P (n ) = 1 e1 . By Theorem 4.4 the
inverse AD = A# is given by AD = f (A), where f () is 0 in a neighbourhood of 0
and 1 in a neighbourhood of (A)\{0}; hence AD is dened by the matrix

0
0
0
0
...

0
0
0
0
0
21
0
31
0
0
0
41
...

...
...
...
...
...

A generalized Drazin inverse

19

Example 8.3. Let X = /1 /1 and let T = A B, where A and B are dened in

the preceding two examples. Set U = A 0 and V = 0 B; then T, U, V L(X)
and
T = U + V,

U V = V U = 0.

U has a simple pole at 0 and V is quasinilpotent in L(X).

The spectrum of T is (T ) = (A) (B) = (A) [22, p.289]. By Theorem 6.4,
T has an isolated spectral point at 0 which is an essential singularity. T has the
Drazin inverse T D but not the nite index Drazin inverse T d . We have
T D = U d = A# 0.
Example 8.4. Let A be the Banach algebra C[0, 1] of all continuous complex
valued functions on [0, 1] equipped with the supremum norm. The zero element
0 has the Drazin inverse 0D = 0. Every ball centred at 0 contains an element
x(t) = t for a suciently small > 0. Since (x) = x([0, 1]) = [0, ], 0
/ iso (a)
D
and x does not exist. This shows that the set of all elements of a Banach algebra
which possess a Drazin inverse is not necessarily open in the algebra.

References
[1] A. Ben-Israel and T. N. E. Greville, Generalized Inverses: Theory and Applications, Wiley-Interscience, New York, 1974.
[2] R. H. Bouldin, Generalized inverses and factorizations, in Recent Applications
of Generalized Inverses, S. L. Campbell, ed., Research Notes in Mathematics
66, Pitman, London, 1982, pp. 233249.
[3] S. L. Campbell, The Drazin inverse of an operator, in Recent Applications of
Generalized Inverses, S. L. Campbell, ed., Research Notes in Mathematics 66,
Pitman, London, 1982, pp.250259.
[4] S. L. Campbell and C. D. Meyer, Generalized Inverses of Linear Transformations, Pitman, London, 1979.
[5] S. L. Campbell, C. D. Meyer and N. J. Rose, Applications of the Drazin inverse
to linear systems of dierential equations with singular constant coecients,
SIAM J. Appl. Math. 31 (1976), 411425.

20

J. J. Koliha

[6] S. R. Caradus, Generalized inverses and operator theory, Queens Papers in

Pure and Appl. Math. 50, Queens University, Kingston, Ontario, 1978.
[7] M. P. Drazin, Pseudo-inverse in associative rings and semigroups, Amer. Math.
Monthly 65 (1958), 506514.
[8] M. P. Drazin, Extremal denitions of generalized inverses, Lin. Alg. Appl. 165
(1992), 185196.
[9] R. E. Harte, Spectral projections, Irish Math. Soc. Newsletter 11 (1984), 10
15.
[10] R. E. Harte, Invertibility and Singularity for Bounded Linear Operators,
Monographs and Textbooks in Pure and Applied Mathematics vol. 109, Marcel
Dekker, New York, 1988.
[11] R. E. Harte, On quasinilpotents in rings, PanAm. Math. J. 1 (1991), 1016.
[12] R. E. Hartwig, Schurs theorem and the Drazin inverse, Pacic J. Math. 78
(1978), 133138.
[13] H. G. Heuser, Functional Analysis, Wiley, New York, 1982.
[14] C. F. King, A note on Drazin inverses, Pacic. J. Math. 70 (1977), 383390.
[15] J. J. Koliha, Isolated spectral points, Proc. Amer. Math. Soc. 124 (1996),
34173424.
[16] D. C. Lay, Spectral properties of generalized inverses of linear operators, SIAM
J. Appl. Math. 29 (1975), 103109.
y, Matrix Analysis for Applied Sciences, Vol. 2, Teubner[17] I. Marek and K. Zitn
Texte zur Mathematik Band 84, Teubner, Leipzig, 1986.
[18] M. Z. Nashed, ed., Generalized Inverses and Applications, Academic Press,
New York, 1976.
[19] M. Z. Nashed and Y. Zhao, The Drazin inverse for singular evolution equations
and partial dierential equations, World Sci. Ser. Appl. Anal. 1 (1992), 441
456

A generalized Drazin inverse

21

[20] U. G. Rothblum, A representation of the Drazin inverse and characterizations

of the index, SIAM J. Appl. Math. 31 (1976), 646648.
[21] B. Simeon, C. F
uhrer and P. Rentrop, The Drazin inverse in multibody system
dynamics, Numer. Math. 64 (1993), 521539.
[22] A. E. Taylor and D. C. Lay, Introduction to Functional Analysis, 2nd edition,
Wiley, New York, 1980.