Abstract
The Drazin inverse, introduced in [1], named after Michael P. Drazin in 1958 in the setting of an abstract ring, is a kind of generalized inverse of a matrix.
Access provided by CONRICYT-eBooks. Download chapter PDF
Similar content being viewed by others
Keywords
- Generalized Drazin Inverse
- Partially Commutative
- Bauer-Fike Theorem
- Complex Banach Algebra
- Spectral Idempotent
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.
The Drazin inverse, introduced in [1], named after Michael P. Drazin in 1958 in the setting of an abstract ring, is a kind of generalized inverse of a matrix. Many interesting spectral properties of the Drazin inverse make it as a concept that is extremely useful in various considerations in topics such as Markov chains, multibody system dynamics, singular difference and differential equations, differential-algebraic equations and numerical analysis ([1,2,3,4,5,6]).
In this chapter we will focus our attention on the behavior of the Drazin inverse of a sum of two Drazin invertible elements in the setting of matrices as well as in Banach algebras, where we will also consider the concept of the generalized Drazin inverse. In 1958, while considering the question of Drazin invertibility of a sum of two Drazin invertible elements of a ring Drazin proved that
provided that \(AE=EA=0\). After that this topic received considerable interest with many authors working on this problem [4, 7,8,9,10], which in turn lead to a number of different formulae for the Drazin inverse \((A+E)^\mathrm{D}\) as a function of \(A, E, A^\mathrm{D}\) and \(E^\mathrm{D}\).
6.1 Additive Results for the Drazin Inverse
Although it was already even in 1958 that Drazin [1] pointed out that computing the Drazin inverse of a sum of two elements in a ring was not likely to be easy, this problem remains open to this day even for matrices. It is precisely this problem when considered in rings of matrices that will be the subject of our interest in this section, i.e., under various conditions we will compute \((P+Q)^\mathrm{D}\) as a function of P, Q, \(P^\mathrm{D}\) and \(Q^\mathrm{D}\). We will extend Drazin’s result in the sense that only one of the conditions \(PQ=0\) or \(PQ=QP\) is assumed. The results obtained will be then used to analyze a special class of perturbations of the type \(A-X.\)
Throughout the section, we shall assume familiarity with the theory of Drazin inverses (see [11]). Also, for \(A\in {\mathbb C}^{n\times n}\), we denote \(Z_A=I-AA^\mathrm{D}\).
First, we will give a representation of \((P+Q)^\mathrm{D}\) under the condition \(PQ=0\) which was considered in [10, Theorem 2.1]:
Theorem 6.1
Let \(P,Q\in {\mathbb C}^{n\times n}\). If \(PQ=0\), then
and
where \(\max \{\mathrm{Ind}(P),\mathrm{Ind}(Q)\} \le k \le \mathrm{Ind}(P)+\mathrm{Ind}(Q).\)
Proof
Under the assumption \(PQ=0\), we have
Using Cline’s Formula [12], \((AB)^\mathrm{D}=A[(BA)^\mathrm{D}]^{2}B,\) we have
Now, by Theorem 1 of [4], we have that
for
and
where \(\max \{\mathrm{Ind}(P), \mathrm{Ind}(Q)\} \le k \le \mathrm{Ind}(P)+\mathrm{Ind}(Q).\)
Hence
Substituting R in the above equality, we get (6.1). It is straightforward to prove (6.2) from (6.1) and (6.3).\(\square \)
Now we list some special cases of the previous result:
Corollary 6.1
Let \(P,Q\in {\mathbb C}^{n\times n}\) be such that \(PQ=0\) and let k be such that \(\max \{\) \(\mathrm{Ind}(P), \mathrm{Ind}(Q)\} \le k \le \mathrm{Ind}(P)+\mathrm{Ind}(Q).\)
-
(i)
If Q is nilpotent, then \( (P+Q)^\mathrm{D}=P^\mathrm{D}+Q(P^\mathrm{D})^{2}+ \cdots +Q^{k-1}(P^\mathrm{D})^{k}.\)
-
(ii)
If \(Q^{2}=0,\) then \((P+Q)^\mathrm{D}=P^\mathrm{D}+Q(P^\mathrm{D})^{2}.\)
-
(iii)
If P is nilpotent, then \( (P+Q)^\mathrm{D}=Q^\mathrm{D}+(Q^\mathrm{D})^{2}P+ \cdots +(Q^\mathrm{D})^{k}P^{k-1}.\)
-
(iv)
If \(P^{2}=0,\) then \( (P+Q)^\mathrm{D}=Q^\mathrm{D}+(Q^\mathrm{D})^{2}P.\)
-
(v)
If \(P^{2}=P,\) then \( (P+Q)^\mathrm{D}=(I-QQ^\mathrm{D})(I+Q+ \cdots +Q^{k-1})P+Q^\mathrm{D}(I-P),\) and \( (P+Q)^\mathrm{D}(I-P)=Q^\mathrm{D}(I-P).\)
-
(vi)
If \(Q^{2}=Q,\) then \( (P+Q)^\mathrm{D}=(I-Q)P^\mathrm{D}+Q(I+P+ \cdots +P^{k-1})(I-PP^\mathrm{D}),\) and \( (I-Q)(P+Q)^\mathrm{D}=(I-Q)P^\mathrm{D}.\)
-
(vii)
If \(PR=0,\) then \( (P+Q)^\mathrm{D}R=(I-QQ^\mathrm{D})P^\mathrm{D}R+Q^\mathrm{D}R=Q^\mathrm{D}R.\)
Theorem 6.1 may be used to obtain several additional perturbation results concerning the matrix \(\varGamma =A-X.\) Needless to say these are rather special, since addition and inversion rarely mix. First a useful result.
Lemma 6.1
Let \(A,F,X\in {\mathbb C}^{n\times n}\). If \(AF=FA\) and \(FX=X,\) then
Proof
Since \(AF=FA\) and \((I-F)X=0,\) we have that
Now the assertion is proved by induction. The case \(k=1\) is trivial. Suppose \((AF-X)^{k}X=(A-X)^{k}X.\) Then by (6.5),
Now we present a perturbation result.
Corollary 6.2
Let \(A,F,X\in {\mathbb C}^{n\times n}\) and let F be an idempotent matrix which commutes with A. Let \(\varGamma =A-X\) and let \(\max \{ \mathrm{Ind}(A), \mathrm{Ind}(X) \} \le k \le \mathrm{Ind}(A)+ \mathrm{Ind}(X).\) If \(FX=X\) and \(R=\varGamma F=AF-XF,\) then
Proof
Let \(\varGamma =A-X=P+Q,\) where \(P=A(I-F)\) and \(Q=AF-FX.\) Since \(F^2=F\) we have that \((I-F)^2=I-F\) and \((I-F)^\mathrm{D}=(I-F).\) Since
after applying Theorem 6.1 we get
where \(V=[P^\mathrm{D}+Q(P^\mathrm{D})^{2}+ \cdots +Q^{k-1}(P^\mathrm{D})^{k}]\) and \(W=[Q^\mathrm{D}+(Q^\mathrm{D})^{2}P + \cdots +(Q^\mathrm{D})^{k}P^{k-1}]\). Put \(T_{1}=(I-QQ^\mathrm{D})V\) and \(T_{2}=W(I-PP^\mathrm{D})\). So we see that we need to compute \(Q^\mathrm{D}\) and \(P^\mathrm{D}.\) The latter is easily found because A and F commute:
On the other hand, in order to compute \(Q^\mathrm{D}\), we split Q further as
where \(R=(A-X)F=AF-FXF\) and \(S=FX(I-F).\) Since
and \(S^{2}=FX[(I-F)F]X(I-F)=0\), by (iv) of Corollary 6.1 we get
Since \(SR^\mathrm{D}=SR=0,\) it follows that \(QQ^\mathrm{D}=RR^\mathrm{D}-R^\mathrm{D}S.\) Also, \(R^\mathrm{D}P=0,\) because
So \(Q^\mathrm{D}P=-(R^\mathrm{D})^{2}SP= -(R^\mathrm{D})^{2}XP.\) Similarly, since \(SR^\mathrm{D}=0\), we get
Repeating the process, we obtain
which when substituted yields the second term:
Let us next examine the first term
We compute the powers \(Q^{i}(P^\mathrm{D})^{i+1}=(AF-X)^{i}(I-F)(A^\mathrm{D})^{i+1}.\) For \(i=1,\) this becomes \((AF-X)(I-F)(A^\mathrm{D})^{2}=-X(I-F)(A^\mathrm{D})^{2},\) while for higher powers of i we may use Lemma 6.1 to obtain
Now
for all i, and \(R^\mathrm{D}(I-F)=(R^\mathrm{D})^2R(I-F)=(R^\mathrm{D})^2(A-X)[F(I-F)]=0,\) so
completing the proof. \(\square \)
Using the previous result we will analyze some special types of perturbations of the matrix \(A-X.\) We shall thereby extend earlier work by several authors [13,14,15,16] and partially solve a problem posed in 1975 by Campbell and Meyer [17], who considered it difficult to establish norm estimates for the perturbation of the Drazin inverse.
In the following five special cases, we assume \(FX=X\) and \(R=AF-XF.\)
Case (1) \(XF=0.\)
Clearly \((R^\mathrm{D})^{i}=(A^\mathrm{D})^{i}F\) and \(S=X.\) Moreover \((A-X)^{i}FX=A^{i}X\) for \(i \ge 0.\) Thus (6.6) reduces to
Case (1a) \(XF=0\) and \(F=AA^\mathrm{D}.\)
If we in addition assume that \(F=AA^\mathrm{D},\) then \(XA^\mathrm{D}=0\) and (6.7) is reduced to
Case (1b) \(XF=0\) and \(F=I-AA^\mathrm{D}.\)
In this case, \(A^\mathrm{D}X=0\) and (6.7) becomes
Case (2) \(F=AA^\mathrm{D}.\)
Now \(AA^\mathrm{D}X=X\), \(R=A^{2}A^\mathrm{D}(I-A^\mathrm{D}XAA^\mathrm{D})\) and (6.6) simplifies to
If we set \(U=I-A^\mathrm{D}XAA^\mathrm{D}\) and \(V=I-AA^\mathrm{D}XA^\mathrm{D}\), then \(UA^\mathrm{D}=A^\mathrm{D}V\) and \(R=A^{2}A^\mathrm{D}U=VA^{2}A^\mathrm{D}.\) Now if we assume that U is invertible, then so will be V and \(U^{-1}A^\mathrm{D}=A^\mathrm{D}V^{-1}.\) It is now easily verified that \(R^{\#}\) exists and
In fact \(RR^{\#}=A^{2}A^\mathrm{D}UU^{-1}A^\mathrm{D} =AA^\mathrm{D}=A^\mathrm{D}V^{-1}VA^{2}A^\mathrm{D}=R^{\#}R\) and \(R^{2}R^{\#}=RAA^\mathrm{D}=R\) and \(R^{\#}RR^{\#}=U^{-1}A^\mathrm{D}AA^\mathrm{D}=U^{-1}A^\mathrm{D}=R^{\#}.\) We then have two sub-cases.
Case (2a) \(F=AA^\mathrm{D},\) and \(U=I-A^\mathrm{D}XAA^\mathrm{D}\) is invertible.
In this case, (6.10) is just
where \(R=A^{2}A^\mathrm{D}U,\) \(R^{\#}=U^{-1}A^\mathrm{D}.\) In general, \((R^{\#})^{i} \ne U^{-i}(A^\mathrm{D})^{i}.\)
Remark 6.1.1
The matrix \(U=I-A^\mathrm{D}XAA^\mathrm{D}\) is invertible if and only if \(I-A^\mathrm{D}X\) is invertible. This result generalizes the main results from [13,14,15,16].
Case (2b) \(F=I-AA^\mathrm{D}.\)
We have \(A^\mathrm{D}X=A^\mathrm{D}F=0,\) and (6.6) becomes
where \(R=A(I-AA^\mathrm{D})-(I-AA^\mathrm{D})X(I-AA^\mathrm{D}).\)
Case (3) \(AA^\mathrm{D}XF=XFAA^\mathrm{D}=XF,\) \(U=I-A^\mathrm{D}XF\) is invertible and \((AF)^{\#}\) exists.
Now \(R=AF-XF=AF-AA^\mathrm{D}FXF\!=\!AF(I-A^\mathrm{D}XF)\!=\!AFU=VFA,\) where \(V=I-XFA^\mathrm{D}.\) Furthermore \(A^\mathrm{D}FV=UA^\mathrm{D}F.\) We may now conclude that U is invertible exactly when V is, in which case \(Y=U^{-1}A^\mathrm{D}F=A^\mathrm{D}FV^{-1}.\)
We then have \(RY=AFU(U^{-1}A^\mathrm{D}F)=AA^\mathrm{D}F= A^\mathrm{D}FV^{-1}(VFA)=YR.\) Lastly,
and \(R^{2}Y=RAA^\mathrm{D}F=A^{2}A^\mathrm{D}F-AA^\mathrm{D}FXFAA^\mathrm{D}=A^{2}A^\mathrm{D}F-XF.\)
If \((AF)^{\#}\) exists then \(AF=AF(AF)^{\#}AF=AFF^{\#}A^\mathrm{D}AF=AFF^{\#}FAA^\mathrm{D}= A^{2}A^\mathrm{D}F\), so \(R^2Y=AF-XF=R,\) i.e., \(Y=R^{\#}\) and (6.6) becomes
Case (4) \(FX=XF=X.\)
In this case, (6.6) reduces to
If in addition to \(F=AA^\mathrm{D},\) the matrix \(U=I-A^\mathrm{D}X\) is invertible, this reduces further to [15]
Case (5) If \(X=A^{2}A^\mathrm{D}\) then \(\varGamma \) is nilpotent and \(\varGamma ^\mathrm{D}=0.\)
Although Theorem 6.1 solves our problem under the assumption that \(PQ=0\), the condition can be relaxed and the result therefore generalized as follows: Since
we may extend the considerations above to the case when \(P(P+Q)^{k-1}Q=0.\)
In fact
This requires computation of \([P(P+Q)^{k-1}]^\mathrm{D}\) and \([(P+Q)^{k-1}Q]^\mathrm{D},\) which may actually be easier than that of \((P+Q)^\mathrm{D}.\)
A second attempt to generalize Theorem 6.1 would be to assume only that \(P^{2}Q=0.\) Needless to say, this is the best attempted via the block form, which in turn should give a suitable formula.
Now, we will investigate explicit representations for the Drazin inverse \((A+E)^\mathrm{D}\) in the case when \(AE=EA\), which was considered in [18, Theorem 2]. For \(A\in {\mathbb C}^{n\times n}\) with Ind\((A) = k\) and rank\((A^k)=r\), there exists an nonsingular matrix \(P\in {\mathbb C}^{n\times n}\) such that
where \(C\in {\mathbb C}^{r\times r}\) is a nonsingular matrix, N is nilpotent of index k and \({\hbox {Ind}}(N)={\hbox {Ind}}(A)=k\). In that case
If \(P=I,\) then the block-diagonal matrices A and \(A^\mathrm{D}\) are written as \(A=C\oplus N\) and \(A^\mathrm{D}=C^{-1}\oplus 0\).
Now we state the following result which was obtained by Hartwig and Shoaf [19] and Meyer and Rose [20], since it will be used in the theorem to follow.
Theorem 6.2
If \(M=\left[ \begin{array}{cc}A\ \ \ &{}C\\ 0\ \ \ &{}B\end{array}\right] \), where \(A\in \mathbb {C}^{n\times n}\) and \(B\in \mathbb {C}^{m\times m}\) with Ind\((A)=k\) and Ind\((B)=l\), then \(M^\mathrm{D}=\left[ \begin{array}{cc}A^\mathrm{D}\ \ \ &{}X\\ 0&{}\ \ \ B^\mathrm{D}\end{array}\right] , \)
where
Theorem 6.3
If \(A, E\in \mathbb {C}^{n\times n}\), \(AE=EA\) and Ind\((A)=k\), then
and
Proof
Let \(A\in \mathbb {C}^{n\times n}\) be given by (6.16). Without loss of generality, we assume that \(P=I\) and \(A=C\oplus N\), where C is invertible and N is nilpotent with \(N^k=0\). From \(AE=EA,\) we have \(A^kE=EA^k\). Now \(E=E_1\oplus E_2\), \(CE_1=E_1C\) and \(NE_2=E_2N.\) Hence
Since C and \( I+C^{-1}E_1\) commute, we get
Notice that \((I+T)^{-1}=\sum \limits _{i=0}^{k-1} (-T)^i\) if \(T^k=0.\) Applying Lemma 4 [19] we get \((N+E_2)^\mathrm{D}= E_2^\mathrm{D}(I+E_2^\mathrm{D}N)^{-1}\) and
Hence
and
\(\square \)
From Theorem 6.3, we can see that the generalized Schur complement \(I+A^\mathrm{D}E\) [21] plays an important role in the representation of the Drazin inverse \((A+E)^\mathrm{D}\). In some special cases, it is possible to give an expression for \((I+A^\mathrm{D}E)^\mathrm{D}.\)
Theorem 6.4
Let \(A, E\in \mathbb {C}^{n\times n}\) be such that \(AE=EA\) and let Ind\((A)=k\) and Ind\((E)=l\).
(1) If \(A^\mathrm{D}E^\mathrm{D}=0\), then
(2) If \(A^\mathrm{D}E=0\), then \( (A+E)^\mathrm{D}=A^\mathrm{D}+(I-AA^\mathrm{D})\sum \limits _{i=0}^{k-1} (E^\mathrm{D})^{i+1}(-A)^i.\)
(3) If Ind\((A)=1,\) then \((A+E)^\mathrm{D} = (I+A^\#E)^\mathrm{D}A^\#+(I-AA^\#)E^\mathrm{D}.\)
Proof
We use the notations from the proof of Theorem 6.3.
(1) If \(A^\mathrm{D}E^\mathrm{D}=0,\) then \(E_1\) is nilpotent with \(E_1^{l}=0.\) So we have
The result now follows from Theorem 6.3.
(2)–(3) Note that if \(A^\mathrm{D}E=0\), then \( E_1=0;\) if Ind\((A)=1,\) then \(N=0.\) The results follow directly from the proof of Theorem 6.3.\(\square \)
Let \(A,E\in {\mathbb C}^{n\times n}\). If there exists a nonzero idempotent matrix \(P=P^2\) such that \(AEP=EAP\) (or \(PAE=PEA\)), then A and E are partially commutative. For \(A, E\in \mathbb {C}^{n\times n}\), let \(A^\pi =I-AA^\mathrm{D} \) and Ind\((A)=k\) and suppose \(E^2=0\). In [22], Castro-González proved that if \(A^\pi E=E\) and \(AEA^\pi =0,\) then
But no representations of \((A+E)^\mathrm{D}\) assuming only partial commutativity are known. Under the conditions \(A^\pi E=E\) and \(AEA^\pi =EAA^\pi ,\) we are able to give an expression for \((A+E)^\mathrm{D}\).
Theorem 6.5
Let \(A\in \mathbb {C}^{n\times n}\) with Ind\((A)=k\) and \(E\in \mathbb {C}^{n\times n}\) be nilpotent of index l. If \(EA^\mathrm{D}=0\) and \(A^\pi AE=A^\pi EA,\) then
where \( T(i)=(I-AA^\mathrm{D}) \sum \limits _{j=0}^{i} {j\atopwithdelims ()i}A^jE^{i-j}.\)
Proof
Similarly as in the proof of Theorem 6.3, let \(A=C\oplus N\), where C is invertible and N is nilpotent with \(N^k=0\). It follows from \(EA^\mathrm{D}=0\) that E can be written as \(E=\left[ \begin{array}{cc}0\ \ \ &{}E_1\\ 0\ \ \ &{}E_2\end{array}\right] \) with \(E_2^l=0.\) Also by \(A^\pi AE=A^\pi EA\), we get \(E_2N=NE_2.\) Thus
We observe that \(N+E_2\) is nilpotent of index \(k+l-1.\) From Theorem 6.2, we further obtain
where
Hence
\(\square \)
The following result generalizes Theorems 6.3 and 6.5 to the case of partial commutativity.
Theorem 6.6
Let \(A, E\in \mathbb {C}^{n\times n}\) and Ind\((A)=k\). Also let \(Q\in {\mathbb C}^{n\times n}\) be an idempotent matrix such that \(QA=AQ\) and \(EQ=0\). If \((I-Q)AE=(I-Q)EA\), then
where \(\varPsi =(I+A^\mathrm{D}E)^\mathrm{D}A^\mathrm{D}+(I-AA^\mathrm{D})\sum \limits _{i=0}^{k-1} (E^\mathrm{D})^{i+1}(-A)^i\) and \(h=\) Ind\([(I-Q)(A+E)].\)
Proof
Suppose that \(Q=I_{r\times r}\oplus 0_{(n-r)\times (n-r)}\), where \(r\le n\). If \(QA=AQ\), \(EQ=0\) and \((I-Q)AE=(I-Q)EA,\) then \(A= A_1\oplus A_2\) and \(E=\left[ \begin{array}{cc}0\ \ \ &{}E_1\\ 0\ \ \ &{}E_2\end{array}\right] \quad {\text{ with } } A_2E_2=E_2A_2.\) Using Theorems 6.2 and 6.3, we have
where
and Ind\((A_2+E_2)=h.\)
If we write \(\varPsi =(I+A^\mathrm{D}E)^\mathrm{D}A^\mathrm{D}+(I-AA^\mathrm{D})\sum \limits _{i=0}^{k-1} (E^\mathrm{D})^{i+1}(-A)^i,\) then \((I-Q)\varPsi =0\oplus (A_2+E_2)^\mathrm{D}.\) We can simplify the expression for \((A+E)^\mathrm{D}\) using the block decomposition above. We deduce
and \( \varSigma _3=QA^\mathrm{D}E\varPsi =\left[ \begin{array}{cc} 0&{}A_1^\mathrm{D}E_1(A_2+E_2)^\mathrm{D}\\ 0&{}0 \end{array}\right] . \)
Thus
\(\square \)
Now a few special cases follow immediately.
Corollary 6.3
Let \(A, E \in \mathbb {C}^{n\times n}\) with \( \mathrm{Ind}(A)=k\) and \(\mathrm{Ind}(E)=l\).
(1) If \(EA^\pi =0\) and \((I-A^\pi ) AE=(I-A^\pi ) EA,\) then
where \(\varPsi =(I+A^\mathrm{D}E)^\mathrm{D}A^\mathrm{D}+(I-AA^\mathrm{D})\sum \limits _{i=0}^{k-1} (E^\mathrm{D})^{i+1}(-A)^i\).
(2) If E is nilpotent, \(EA^\pi =E\) and \(A^\pi AE=A^\pi EA,\) then
Proof
We adopt the notations from Theorem 6.6.
(1) Let \(Q=I-AA^\mathrm{D}\) in Theorem 6.6 and apply \(QA^\mathrm{D}=0\) to (6.19).
(2) Let \(Q=AA^\mathrm{D}\) in Theorem 6.6. Since \(EA^\pi =E\), we obtain \(EA^\mathrm{D}=EA^\pi A^\mathrm{D}=0\). Thus \((A^\mathrm{D}E)^2=A^\mathrm{D}EA^\mathrm{D}E=0\) and \((I+A^\mathrm{D}E)^\mathrm{D}A^\mathrm{D}=(I+A^\mathrm{D}E)^{-1}A^\mathrm{D}=A^\mathrm{D}.\) Note that E is nilpotent so that \(\varPsi =A^\mathrm{D}\). Hence
The result follows directly from (6.19). \(\square \)
Let A be an \(n \times n\) complex matrix and \(B=A+E\) be a perturbation of A. The classical Bauer-Fike theorem on eigenvalue perturbation gives a bound on the distance between an eigenvalue \(\mu \) of B and the closest eigenvalue \(\lambda \) of A, which is required to be diagonalizable.
Let \(A=X \varSigma X^{-1}\) be an eigendecomposition, where \(\varSigma \) is a diagonal matrix, and X is an eigenvector matrix. The Bauer-Fike theorem [23, Theorem IIIa] states that for any eigenvalue \(\mu \) of B, there exists an eigenvalue \(\lambda \) of A such that \( |\mu - \lambda | \le \kappa (X) \Vert E\Vert , \) where \(\kappa (X)= \Vert X\Vert ~\Vert X^{-1}\Vert \) is the condition number of X.
The relative perturbation version of the Bauer-Fike theorem [24, Corollary 2.2] below requires, in addition, that A be invertible. That is, if A is diagonalizable and invertible, then for any eigenvalue \(\mu \) of B, there exists an eigenvalue \(\lambda \) of A such that
Without the assumption of diagonalizability and invertibility of A, we refine the bound (6.20) under the condition that \(AE=EA\).
Theorem 6.7
Let \(B=A+E \in \mathbb {C}^{n\times n}\) be such that A is not nilpotent and \(AE=EA\). For any eigenvalue \(\mu \) of B, there exists a nonzero eigenvalue \(\lambda \) of A such that
where \(\rho (A^\mathrm{D}E)\) is the spectral radius of \(A^\mathrm{D}E\).
Proof
Assume that \(AE=EA\) and that A is not nilpotent. Then for any nonzero eigenvalue \(\lambda \) of A, there exits a common eigenvector x [25, p.250] such that
Therefore
whence
\(\square \)
Recently, the perturbation of the Drazin inverse has been studied by several authors ([6, 9, 22, 26,27,28,29,30,31,32,33]). As one application of our results in Theorem 6.3, we can establish upper bounds for the relative error \(\Vert B^\mathrm{D}\Vert \) and \(\Vert B^\mathrm{D}-A^\mathrm{D}\Vert /\Vert A^\mathrm{D}\Vert \) under the assumption that \(AE=EA\).
Theorem 6.8
If \(B=A+E\in \mathbb {C}^{n\times n}\), \(AE=EA\) and \(\max \{\Vert A^\mathrm{D}E\Vert , \Vert A^\pi AE^\mathrm{D}\Vert \}<1\), then
and
Proof
Note that the assumption \(\max \{\Vert A^\mathrm{D}E\Vert , \Vert A^\pi AE^\mathrm{D}\Vert \}<1\) implies invertibility of \(I+A^\mathrm{D}E\) and \(I+A^\pi AE^\mathrm{D}.\) It follows directly from Theorem 6.3 that
and
\(\square \)
Remark 6.1.2
For any non-zero eigenvalue \(\mu \) of the spectral set \(\sigma (A+E)\), we can estimate its lower bound: let \(\mu \in \sigma (A+E)\). We have \(1/\mu \in \sigma [(A+E)^\mathrm{D}]\) and \( |1/\mu | \le \rho [(A+E)^\mathrm{D}] \le \Vert (A+E)^\mathrm{D}\Vert ,\) i.e.,
Next we will apply Theorem 6.5 to obtain a perturbation bound in terms of \(A^\mathrm{D}\) and \({\mathscr {E}}_l=B^l-A^l\) for some positive integer l.
Theorem 6.9
Let \(B=A+E \in \mathbb {C}^{n\times n}\) with Ind\((A)=k\) and Ind\((B)=s\). Denote \({\mathscr {E}}_l=B^l-A^l\), where \(l=\max \{k,s\}\). Assume that the conditions in Theorem 6.5 hold. Then
Proof
Since \(l=\max \{k,s\}\), using the notations in the proof of Theorem 6.5, we have
Then
and
We then have
and
The proof is complete.\(\square \)
Generalizations of the results of this section to linear operators on Banach spaces can be found in [9, 34,35,36] while their generalizations to Banach algebra elements can be found in [37] and some will also be given in the next section where the generalized Drazin inverse will be considered.
6.2 Additive Results for the Generalized Drazin Inverse in Banach Algebra
Let \({\mathscr {A}}\) be a complex Banach algebra with the unit 1. By \({\mathscr {A}}^{-1}\), \({{\mathscr {A}}}^{\mathsf{nil}}\), \({{\mathscr {A}}}^{\mathsf{qnil}}\) we denote the sets of all invertible, nilpotent and quasi-nilpotent elements in \({\mathscr {A}}\), respectively. Let us recall that the Drazin inverse of \(a\in {\mathscr {A}}\) [1] is the (unique) element \(x\in {\mathscr {A}}\) (denoted by \(a^\mathsf{D}\)) which satisfies
for some nonnegative integer k. The least such k is the index of a, denoted by \(\mathrm{ind}(a)\). When \(\mathrm{ind}(a)= 1\) then the Drazin inverse \(a^\mathsf{D}\) is called the group inverse and it is denoted by \(a^{\#}.\) The conditions (6.23) are equivalent to
The concept of the generalized Drazin inverse in a Banach algebra was introduced by Koliha [38]. The condition \(a-a^2x\in {\mathscr {A}}^\mathsf{nil}\) was replaced by \(a-a^2x\in {{\mathscr {A}}}^{\mathsf{qnil}}\). Hence, the generalized Drazin inverse of a is the (unique) element \(x\in {\mathscr {A}}\) (written \(a^\mathsf{d}\)) which satisfies
We mention that an alternative definition of the generalized Drazin inverse in a ring is also given in [39,40,41]. These two concepts of the generalized Drazin inverse are equivalent in the case when the ring is actually a complex Banach algebra with a unit. It is well known that \(a^\mathsf{d}\) is unique whenever it exists [38]. The set \({\mathscr {A}}^\mathsf{d}\) consists of all \(a\in {\mathscr {A}}\) such that \(a^\mathsf{d}\) exists. For many interesting properties of the Drazin inverse see [1, 38, 42].
This section is a continuation of the previous one with the difference that here we investigate additive properties of the generalized Drazin inverse in a Banach algebra and find explicit expressions for the generalized Drazin inverse of the sum \(a+b\) under various conditions.
Hartwig et al. [10] for matrices and Djordjević and Wei [9] for operators used the condition \(AB= 0\) to derive a formula for \((A+B)^\mathsf{d}\). After that Castro and Koliha [43] relaxed this hypothesis by assuming the following complimentary condition symmetric in \(a, b\in {\mathscr {A}}^\mathsf{d}\),
thus generalizing the results from [9]. It is easy to see that \(ab= 0\) implies (6.26), but the converse is not true (see [43, Example 3.1]).
In the first part of the section we will find some new conditions, which are not equivalent with the conditions from [43], allowing for the generalized Drazin inverse of \(a+b\) to be expressed in terms of \(a, a^\mathsf{d}, b,b^\mathsf{d}\). It is interesting to note that in some cases the same expression for \((a+b)^\mathsf{d}\) are obtained as in [43]. In the rest of the section we will generalize some recent results from [43].
Let \(a\in {\mathscr {A}}\) and let \(p\in {\mathscr {A}}\) be an idempotent (\(p= p^2\)). Then we can write
and use the notations
Every idempotent \(p\in {\mathscr {A}}\) induces a representation of an arbitrary element \(a\in {\mathscr {A}}\) given by the following matrix
Let \(a^\pi \) be the spectral idempotent of a corresponding to \(\{0\}\). It is well known that \(a\in {\mathscr {A}}^\mathsf{d}\) can be represented in the matrix form:
relative to \(p= aa^\mathsf{d}= 1-a^\pi \), where \(a_{11}\) is invertible in the algebra \(p{\mathscr {A}}p\) and \(a_{22}\) is quasi-nilpotent in the algebra \((1-p){\mathscr {A}}(1-p)\). Then the generalized Drazin inverse is given by
The following result is proved in [4, 20] for matrices, extended in [44] for a bounded linear operator and in [43] for arbitrary elements in a Banach algebra.
Theorem 6.10
Let \(x,y\in {\mathscr {A}}\) and
relative to the idempotent \(p\in {\mathscr {A}}\).
-
(1)
If \(a\in (p{\mathscr {A}}p)^\mathsf{d}\) and \(b\in ((1-p){\mathscr {A}}(1-p))^\mathsf{d}\), then x and y are Drazin invertible and
$$\begin{aligned} x^\mathsf{d}= \left[ \begin{array}{cc} a^\mathsf{d}&{}u\\ 0&{}b^\mathsf{d}\end{array}\right] _p,\ \ \ y^\mathsf{d}= \left[ \begin{array}{cc} b^\mathsf{d}&{}0\\ u&{}a^\mathsf{d}\end{array}\right] _{(1-p)} \end{aligned}$$(6.28)where \(u= \sum \limits _{n= 0}^{\infty }(a^\mathsf{d})^{n+2}c b^n b^\pi +\sum \limits _{n= 0}^{\infty }a^\pi a^n c (b^\mathsf{d})^{n+2}-a^\mathsf{d} c b^\mathsf{d}.\)
-
(2)
If \(x\in {\mathscr {A}}^\mathsf{d}\) and \(a\in (p{\mathscr {A}}p)^\mathsf{d}\), then \(b\in ((1-p){\mathscr {A}}(1-p))^\mathsf{d}\) and \(x^\mathsf{d}, y^\mathsf{d}\) are given by (6.28).
We will need the following auxiliary result.
Lemma 6.2
Let \(a,b\in {{\mathscr {A}}}^{\mathsf{qnil}}\). If \(ab= ba\) or \(ab= 0\), then \(a+b\in {{\mathscr {A}}}^{\mathsf{qnil}}\).
Proof
If \(ab= ba\), we have that
which gives \(a+b\in {{\mathscr {A}}}^{\mathsf{qnil}}\). The case when \(ab= 0\) follows from the equation
\(\Box \)
In view of the previous lemma, the first approach to the problem addressed in this section was to replace the condition \(ab= 0\) used in [9, 10] by \(ab= ba\). As expected, this alone was not enough to derive a formula for \((a+b)^\mathsf{d}\). We will thus impose the following three conditions on \(a,b\in {\mathscr {A}}^\mathsf{d}\):
Instead of the condition \(ab= ba\) we are thus assuming the weaker condition \(b^\pi a^\pi ba= b^\pi a^\pi ab\). Notice that
where for \(u\in {\mathscr {A}}\), \(u^\circ = \{x\in {\mathscr {A}}\ :\ ux= 0\}\).
For matrices and bounded linear operators on a Banach space the conditions (6.30)–(6.32) are equivalent to
Remark that, unlike the conditions (3.1) from [43], the conditions (6.29) are not symmetric in a, b so our expression for \((a+b)^\mathsf{d}\) will not be symmetric in a, b.
In the next theorem, under the assumption that (6.29) holds, we can give an expression for \((a+b)^\mathsf{d}\) as follows.
Theorem 6.11
Let \(a,b\in {\mathscr {A}}^\mathsf{d}\) be such that (6.29) is satisfied. Then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
Before proving Theorem 6.11, we first have to prove the special case of it given below.
Theorem 6.12
Let \(a\in {{\mathscr {A}}}^{\mathsf{qnil}}\), \(b\in {\mathscr {A}}^\mathsf{d}\) satisfy \(b^\pi ab= b^\pi ba\) and \(a= ab^\pi \). Then (6.29) is satisfied, \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
Proof
First, suppose that \(b\in {{\mathscr {A}}}^{\mathsf{qnil}}\). Then \(b^\pi = 1\) and from \(b^\pi ab= b^\pi ba\) we obtain \(ab= ba\). Using Lemma 6.2, \(a+b\in {{\mathscr {A}}}^{\mathsf{qnil}}\) and (6.28) holds. Now, we assume that b is not quasi-nilpotent and consider the matrix representations of a and b relative to \(p= 1-b^\pi \). We have
where \(b_1\in (p{\mathscr {A}}p)^{-1}\) and \(b_2\in ((1-p){\mathscr {A}}(1-p))^\mathsf{qnil}\subset {\mathscr {A}}^\mathsf{qnil} \). From \(a= ab^\pi \), it follows that \(a_{11}= 0\) and \(a_{21}= 0\). We denote \(a_1= a_{12}\) and \(a_2= a_{22}\). Hence
The condition \(b^\pi ab= b^\pi ba\) implies that \(a_2b_2= b_2a_2\). Hence, using Lemma 6.2, we get \(a_2+b_2\in ((1-p){\mathscr {A}}(1-p))^\mathsf{qnil}\). Now, by Theorem 6.10, we obtain \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
\(\Box \)
Let us observe that the expression for \((a+b)^\mathsf{d}\) in (6.28) and that in (3.6) of Theorem 3.3 in [43] are exactly the same. If we assume that \(ab= ba\) instead of \(b^\pi ab= b^\pi ba\), we get a much simpler expression for \((a+b)^\mathsf{d}\).
Corollary 6.4
Suppose \(a\in {{\mathscr {A}}}^{\mathsf{qnil}}\), \(b\in {\mathscr {A}}^\mathsf{d}\) satisfy \(ab= ba\) and \(a= ab^\pi \). Then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
Proof
From \(a= ab^\pi \), as we mentioned before, it follows that \(ab^\mathsf{d}= 0\). Because the Drazin inverse \(b^\mathsf{d}\) is a double commutant of a, we have
\(\Box \)
Proof of Theorem 6.11: If b is quasi-nilpotent we can apply Theorem 6.12. Hence, we assume that b is neither invertible nor quasi-nilpotent and consider the matrix representations of a and b relative to \(p= 1-b^\pi \):
where \(b_1\in (p{\mathscr {A}}p)^{-1}\) and \(b_2\in ((1-p){\mathscr {A}}(1-p))^\mathsf{qnil}\). As in the proof of Theorem 6.12, from \(a= ab^\pi \) it follows that
From the conditions \(b^\pi a^\pi ba= b^\pi a^\pi ab\) and \(b^\pi ba^\pi = b^\pi b\), we obtain \(a_2^\pi b_2a_2= a_2^\pi a_2 b_2\) and \(b_2= b_2a_2^\pi \). Now, from Theorem 6.12 it follows that \((a_2+b_2)\in ((1-p){\mathscr {A}}(1-p))^\mathsf{d}\) and
By Theorem 6.10, we get
where \(u= \sum \limits _{n= 0}^{\infty }b_1^{-(n+2)}a_1(a_2+b_2)^n(a_2+b_2)^\pi -b_1^{-1}a_1(a_2+b_2)^\mathsf{d}\) and \(b_1^{-1}\) is the inverse of \(b_1\) in the algebra \(p{\mathscr {A}}p\). Using (6.35), we have
By a straightforward manipulation, (6.33) follows. \(\Box \)
Corollary 6.5
Suppose \(a,b\in {\mathscr {A}}^\mathsf{d}\) are such that \(ab= ba\), \(a= ab^\pi \) and \(b^\pi = ba^\pi = b^\pi b\). Then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
If a is invertible and b is group invertible, then conditions (6.31) and (6.32) are satisfied, so we only have to assume \(a= ab^\pi \). In the remaining case when b is invertible we get \(a= 0\).
It is interesting to remark that conditions (6.26) and (6.29) are independent, i.e., neither of them implies the other, but in some cases the same expressions for \((a+b)^\mathsf{d}\) are obtained.
If we consider the algebra \({\mathscr {A}}\) of all complex \(3\times 3\) matrices and \(a,b\in {\mathscr {A}}\) which are given in the Example 3.1 [43], we can see that condition (6.26) is satisfied, whereas condition (6.29) fails. In the following example we have the opposite case. We construct a, b in the algebra \({\mathscr {A}}\) of all complex \(3\times 3\) matrices such that (6.29) is satisfied but (6.26) is not. If we assume that \(ab= ba\) in Theorem 6.11 the expression for \((a+b)^\mathsf{d}\) will be exactly the same as that in [43, Theorem 3.5] (which is Corollary 6.7 there).
Example 6.1
Let
Then
and \(b^\pi = 1\). We can see that \(a= ab^\pi \), \(a^\pi ab= a^\pi = ba\) and \(ba^\pi = b\), i.e., (6.29) holds. Also, \(a^\pi b= 0\ne b\), so (6.26) is not satisfied.
In the rest of the section, we present a generalization of the results from [43]. We use some weaker conditions than those in [43]. For example in the next theorem, which generalizes [43, Theorem 3.3], we assume that \(e= (1-b^\pi )(a+b)(1-b^\pi )\in {\mathscr {A}}^\mathsf{d}\) instead of \(ab^\pi = a\). If \(ab^\pi = a\), then \(e= (1-b^\pi )b= \left[ \begin{array}{cc} b_1&{}0\\ 0&{}0\end{array}\right] _p\) for \(p= 1-b^\pi \) and \(e^\mathsf{d}= b^\mathsf{d}\).
Theorem 6.13
Let \(b\in {\mathscr {A}}^\mathsf{d}\), \(a\in {{\mathscr {A}}}^{\mathsf{qnil}}\) be such that
Then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
Proof
The case when \(b\in {{\mathscr {A}}}^{\mathsf{qnil}}\) follows from Lemma 6.2. Hence, we assume that b is not quasi-nilpotent. Then
where \(p= 1-b^\pi \). From \(b^\pi ab= 0\) we have \(b^\pi a(1-b^\pi )= 0\), i.e., \(a_{21}= 0\). Put \(a_1= a_{11}\), \(a_{22}= a_2\) and \(a_{12}= a_{3}\). Then,
Also, \(b^\pi ab= 0\) implies that \(a_2b_2= 0\), so \(a_2+b_2\in ((1-p){\mathscr {A}}(1-p))^\mathsf{qnil}\), according to Lemma 6.2. Applying Theorem 6.10, we obtain
where \(u= \sum \limits _{n= 0}^{\infty }((a_1+b_1)^\mathsf{d})^{n+2}a_3(a_2+b_2)^n\).
By direct computation, we verify that
\( \Box \)
Now, as a corollary we obtain Theorem 3.3 from [43].
Corollary 6.6
Let \(b\in {\mathscr {A}}^\mathsf{d}\), \(a\in {{\mathscr {A}}}^{\mathsf{qnil}}\) and \(ab^\pi = a\), \(b^\pi ab= 0\). Then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
The next result is a generalization of [43, Theorem 3.5]. For simplicity, we use the following notation:
for given \(a,b\in A^\mathsf{d}\).
Theorem 6.14
Let \(a,b\in {\mathscr {A}}^\mathsf{d}\) be such that \((1-a^\pi ) b(1-a^\pi )\in {\mathscr {A}}^\mathsf{d}\), \(f\in {\mathscr {A}}_1^{-1}\) and \(e\in {\mathscr {A}}_2^\mathsf{d}\). If
then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
where by \((f)_{{\mathscr {A}}_1}^{-1}\) we denote the inverse of f in \({\mathscr {A}}_1\).
Proof
Obviously, if a is invertible, then the statement of the theorem holds. If a is quasi-nilpotent, then the result follows from Theorem 6.13. Hence, we assume that a is neither invertible nor quasi-nilpotent. As in the proof of Theorem 6.11, we have
where \(p= 1-a^\pi \), \(a_1\in (p{\mathscr {A}}p)^{-1}\) and \(a_2\in ((1-p){\mathscr {A}}(1-p))^\mathsf{qnil}\). From \((1-a^\pi )ba^\pi = 0\), we have that \(b_{12}= 0\). Let \(b_1= b_{11}\), \(b_{22}= b_2\) and \(b_{21}= b_{3}\). Then,
The condition \(a^\pi b^\pi ab a^\pi = 0\) expressed in the matrix form yields
Similarly, \(a^\pi a(1-b^\pi )= 0\) implies that \(a_2b_2^\pi = a_2\). From Corollary 6.6, we get \(a_2+b_2\in {\mathscr {A}}^\mathsf{d}\) and
Using Theorem 6.10, we obtain \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
where
By straightforward computation, the desired result follows.\(\Box \)
Corollary 6.7
Suppose \(a,b\in {\mathscr {A}}^\mathsf{d}\) satisfy condition (6.26). Then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
Proof
We have that \(f= (1-a^\pi )a\), so \((f)_{{\mathscr {A}}_1}^{-1}= a^\mathsf{d}\). \(\square \)
Next we generalize the results from [45] to the Banach algebra case.
Theorem 6.15
Let \(a,b\in {\mathscr {A}}^\mathsf{d}\) and \(ab=ba\). Then \(a+b\in {\mathscr {A}}^\mathsf{d}\) if and only if \(1+a^\mathsf{d} b\in {\mathscr {A}}^\mathsf{d}\). In this case, we have
and
Moreover, if \(\Vert b\Vert \Vert a^\mathsf{d}\Vert <1\) and \(\Vert a\Vert \Vert b^\mathsf{d}\Vert <1\), then we have
and
Proof
Since a is generalized Drazin invertible, and
relative to \(p=1-a^\pi \), where \(a_{11}\) is invertible in the algebra \(p{\mathscr {A}}p\) and \(a_{22}\) is a quasi-nilpotent element of the algebra \((1-p) {\mathscr {A}}(1-p)\). Let \(b= \left[ \begin{array}{cc} b_{11}&{}b_{12}\\ b_{21}&{}b_{22}\end{array}\right] _p\).
From \(ab=ba\), we have \(b_{12}=(a_{11})^{-1}_{p{\mathscr {A}}p}b_{12}a_{22}\) which implies that \(b_{12}=(a_{11})^{-n}_{p{\mathscr {A}}p} b_{12}a_{22}^n\), for arbitrary \(n\in \mathbb {N}\). Since \(a_{22}\) is a quasi-nilpotent, we obtain \(b_{12}=0\). Similarly, from \(ab=ba\) it follows that \(b_{21}=a_{22}b_{21}(a_{11})^{-1}_{p{\mathscr {A}}p}\), i.e., \(b_{21}=0\). Also, \(a_{11}b_{11}=b_{11}a_{11}\) and \(a_{22}b_{22}=b_{22}a_{22}\).
Since, \(b\in {\mathscr {A}}^\mathsf{d}\) and \(\sigma (b)=\sigma (b_1)_{p{\mathscr {A}}p}\cup \sigma (b_2)_{(1-p){\mathscr {A}}(1-p)}\), using Theorem 4.2 from [38], we deduce \(b_1\in {p{\mathscr {A}}p}\) and \(b_2\in {(1-p){\mathscr {A}}(1-p)}\), so \(b_{11},b_{22}\in {\mathscr {A}}^\mathrm{d}\) and we can represent \(b_{11}\) and \(b_{22}\) as
where \(p_1=1-b_{11}^\pi \) and \(p_2=1-b_{22}^\pi \), \(b'_{11}, b''_{11}\) are invertible in the algebras \(p_1{\mathscr {A}}p_1\) and \(p_2{\mathscr {A}}p_2\) respectively, and \(b'_{22}, b''_{22}\) are quasi-nilpotent. Since \(b_{11}\) commutes with an invertible \(a_{11}\) and \(b_{22}\) with a quasi-nilpotent \(a_{22}\), we prove as before that
Since \(p_1p=pp_1=p_1\), from the fact that \(a_{11}\) is invertible in the sub-algebra \(p{\mathscr {A}}p\), we prove that \(a'_{11}\) and \(a'_{22}\) are invertible in the algebras \(p_1{\mathscr {A}}p_1\) and \((p-p_1){\mathscr {A}}(p-p_1)\), respectively. Also, \(a''_{11}\) and \(a''_{22}\) are quasi-nilpotent, thus \(a'_{ii}\) commutes with \(b'_{ii}\) and \(a''_{ii}\) with \(b''_{ii}\), for \(i=1,2\).
Since \(a'_{22}\) is invertible and \(b'_{22}\) is quasi-nilpotent and they commute, we have that \((a'_{22})^{-1}_{(1-p_1){\mathscr {A}}(1-p_1)}b'_{22}\) is quasi-nilpotent, so \((1-p_1)+(a'_{22})^{-1}_{(1-p_1){\mathscr {A}}(1-p_1)}b'_{22}\) is invertible in \((1-p_1){\mathscr {A}}(1-p_1)\) and \(a'_{22}+b'_{22}\in {\mathscr {A}}^\mathsf{d}\).
Similarly, we conclude that \(a''_{11}+b''_{11}\in {\mathscr {A}}^\mathsf{d}\). Also, \(a_{22}''+b_{22}''\) is generalized Drazin invertible.
Now, we obtain
Since, \(a'_{11}+b'_{11}\in p_1{\mathscr {A}}p_1\) and \(b'_{22}+a''_{11}+b''_{11}+a''_{22}+b''_{22}\in (1-p_1){\mathscr {A}}(1-p_1)\) we have
Next, we inspect generalized Drazin invertibility of \(y=a'_{22}+b'_{22}+a''_{11}+b''_{11}+a''_{22}+b''_{22}\). From \(p_2yp_2=a''_{11}+b''_{11}\) and \((1-p_2)y(1-p_2)y=a'_{22}+b'_{22}+a''_{22}+b''_{22}\), we conclude
Previously, we showed that \(a''_{11}+b''_{11}\in {\mathscr {A}}^\mathsf{d}\), so \(y\in {\mathscr {A}}^\mathsf{d}\) if and only if \(z=a'_{22}+b'_{22}+a''_{22}+b''_{22}\in {\mathscr {A}}^\mathsf{d}\). Notice that \(z=pzp+(1-p)z(1-p)\), where \(pzp=a'_{22}+b'_{22}\in {\mathscr {A}}^\mathsf{d}\) and \((1-p)z(1-p)=a''_{22}+b''_{22}\in {\mathscr {A}}^\mathsf{d}\), so \(z\in {\mathscr {A}}^\mathsf{d}\). Hence, \(y\in {\mathscr {A}}^\mathsf{d}\) and we obtain \(a+b\in {\mathscr {A}}^\mathsf{d}\) if and only if \(a'_{11}+b'_{11}\in {\mathscr {A}}^\mathsf{d}\).
Now,
From the first equation, we obtain
Consequently, we have the estimates
and
\( \Box \)
Corollary 6.8
Let \(a, b\in {\mathscr {A}}^\mathsf{d}\) be such that \(ab=ba\) and \(1+a^\mathsf{d}b\in {\mathscr {A}}^\mathsf{d}\).
-
(1)
If b is quasi-nilpotent, then
$$(a+b)^\mathsf{d} = \sum \limits _{n=0}^\infty (a^\mathsf{d})^{n+1}(-b)^n =(1+a^\mathsf{d}b)^{-1}a^\mathsf{d}.$$ -
(2)
If \(b^k=0,\) then \( (a+b)^\mathsf{d} = \sum \limits _{n=0}^{k-1}(a^\mathsf{d})^{n+1}(-b)^n=(1+a^\mathsf{d}b)^{-1}a^\mathsf{d}.\)
-
(3)
If \(b^k=b~ (k\ge 3)\), then \( b^\mathsf{d}=b^{k-2}\) and
$$\begin{array}{rcl} (a+b)^\mathsf{d} &{}=&{}a^\mathsf{d}(1+a^\mathsf{d}b)^\mathsf{d}b^{k-1}+(1-b^{k-1})a^\mathsf{d} +b^{k-2}\left[ \sum \limits _{n=0}^\infty (b^\mathsf{d})^n(-a)^n\right] (1-aa^\mathsf{d}) \\ {} &{}=&{}a^\mathsf{d}(1+a^\mathsf{d}b)^\mathsf{d}b^{k-1}+(1-b^{k-1})a^\mathsf{d} +b^{k-2}(1+ab^{k-2})^\mathsf{d}(1-aa^\mathsf{d}). \end{array}$$ -
(4)
If \(b^2=b\), then \( b^\mathsf{d}=b\) and
$$\begin{array}{rcl} (a+b)^\mathsf{d}&{}=&{}a^\mathsf{d}(1+a^\mathsf{d}b)^\mathsf{d}b+(1-b)a^\mathsf{d} +b\left[ \sum \limits _{n=0}^\infty (-a)^n\right] (1-aa^\mathsf{d}) \\ {} &{}=&{}a^\mathsf{d}(1+a^\mathsf{d}b)^\mathsf{d}b+(1-b)a^\mathsf{d} +b(1+a)^\mathsf{d}(1-aa^\mathsf{d}).\end{array}$$ -
(5)
If \(a^2=a\) and \(b^2=b\), then \(1+ab\) is invertible and \(a(1+ab)^{-1}b=\frac{1}{2}ab.\) 1n this case,
$$\begin{array}{rcl} (a+b)^\mathsf{d} &{}=&{}a(1+ab)^{-1}b+b(1-a)+(1-b)a \\ &{}=&{}a+b-\frac{3}{2}ab. \end{array}$$
Theorem 6.16
Let \(a,b\in {\mathscr {A}}^\mathsf{d}\) be such that \(\Vert a^\mathsf{d}b\Vert <1\), \(a^\pi ba^\pi =a^\pi b\) and \(a^\pi ab=a^\pi ba\). If \(a^\pi b\in {\mathscr {A}}^\mathsf{d}\), then \(a+b\in {\mathscr {A}}^\mathsf{d}\). In this case,
Moreover, if \(\Vert a\Vert \Vert b^\mathsf{d}\Vert <1\), \(\Vert b\Vert \Vert a^\mathsf{d}\Vert <1\) and \(\frac{\Vert a^\mathsf{d}\Vert \Vert a^\mathsf{d}b\Vert }{1-\Vert a^\mathsf{d}b\Vert }\Vert a+b\Vert <1\), then
Proof
Since \(a\in {\mathscr {A}}^\mathsf{d}\) and \(a^\pi b(I-a^\pi )=0\), we have that for \(p=1-a^\pi \)
where \(a_{1}\) is invertible in the algebra \(p{\mathscr {A}}p\) and \(a_{2}\) is a quasi-nilpotent element of the algebra \((1-p) {\mathscr {A}}(1-p)\). Also from \(a^\pi ab=a^\pi ba\) and the fact that \(a^\pi b\in {\mathscr {A}}^\mathsf{d}\), we conclude that \(a_2b _2=b _2a_2\) and \(b_2\in {\mathscr {A}}^\mathsf{d}\). It follows from \(\Vert a^\mathsf{d} b\Vert <1\) that \(1+a^\mathsf{d}b\) is invertible. Now, from Theorem 6.15, we have
Using Theorem 6.10, we get
where
We know that
and
By computation we obtain
Hence, we have
If \(\Vert a\Vert \Vert b^\mathsf{d}\Vert <1\), \(\Vert b\Vert \Vert a^\mathsf{d}\Vert <1\) and \(\frac{\Vert a^\mathsf{d}\Vert \Vert a^\mathsf{d}b\Vert }{1-\Vert a^\mathsf{d}b\Vert }\Vert a+b\Vert <1\), we obtain
\(\Box \)
Corollary 6.9
Let \(a\in {\mathscr {A}}^\mathsf{d}\) and \(b\in {\mathscr {A}}\) be such that \(\Vert ba^\mathsf{d}\Vert <1, a^\pi b(1-a^\pi )=0 \) and \(a^\pi ab=a^\pi ba\),
-
(1)
If \(baa^\mathsf{d}=0\) and b is quasi-nilpotent, then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
$$(a+b)^\mathsf{d}= \sum ^\infty _{n=0}(a^\mathsf{d})^{n+2}b(a+b)^n+a^\mathsf{d}.$$ -
(2)
If \(a^\pi b=ba^\pi \), \(\sigma (a^\pi b)=0\), then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
$$(a+b)^\mathsf{d}= (1+a^\mathsf{d}b)^{-1}a^\mathsf{d}=a^\mathsf{d}(1+ba^\mathsf{d})^{-1}.$$
The following theorem is a generalization of Theorem 6.16 and Theorem 6 from [45].
Theorem 6.17
Let \(a,b\in {\mathscr {A}}^\mathsf{d}\) and q be an idempotent such that \(aq=qa\), \((1-q)bq=0\), \((ab-ba)q=0\) and \((1-q)(ab-ba)=0.\) If \((a+b)q\) and \((1-q)(a+b)\) are generalized Drazin invertible, then \(a+b\in {\mathscr {A}}^\mathsf{d}\) and
where
Proof
The proof is a similar to that of Theorem 6.16. \(\Box \)
References
Drazin, M.P.: Pseudoinverse in associative rings and semigroups. Amer. Math. Monthly 65, 506–514 (1958)
Deng, C.Y.: The Drazin inverses of products and differences of orthogonal projections. J. Math. Anal. Appl. 335, 64–71 (2007)
Eisenstat, S.C.: A perturbation bound for the eigenvalues of a singular diagonalizable matrix. Linear Algebra Appl. 416, 742–744 (2006)
Hartwig, R.E., Li, X., Wei, Y.: Representations for the Drazin inverse of \(2\times 2\) block matrix. SIAM J. Matrix Anal. Appl. 27, 757–771 (2006)
Simeon, B., Fuhrer, C., Rentrop, P.: The Drazin inverse in multibody system dynamics. Numer. Math. 64, 521–539 (1993)
Xu, Q., Song, C., Wei, Y.: The stable perturbation of the Drazin inverse of the square matrices. SIAM J. Matrix Anal. Appl. 31, 1507–1520 (2010)
Castro-González, N., Koliha, J., Rakočević, V.: Continuity and general perturbation of the Drazin inverse for closed linear operators. Abstr. Appl. Anal. 7, 335–347 (2002)
Cvetković-Ilić, D.S.: A note on the representation for the Drazin inverse of \(2 \times 2\) block matrices. Linear Algebra Appl. 429, 242–248 (2008)
Djordjević, D.S., Wei, Y.: Additive results for the generalized Drazin inverse. J. Austral Math. Soc. 73, 115–125 (2002)
Hartwig, R.E., Wang, G., Wei, Y.: Some additive results on Drazin inverse. Linear Algebra Appl. 322, 207–217 (2001)
Ben-Isreal, A., Greville, T.N.E.: Generalized Inverse: Theory and Applications, 2nd edn. Springer, New York (2003)
Cline, R.E.: An application of representation of a matrix, MRC Technical Report, # 592 (1965)
Meyer, C.D.: The condition number of a finite Markov chains and perturbation bounds for the limiting probabilities. SIAM J. Algebraic Discrete Methods 1, 273–283 (1980)
Meyer, C.D., Shoaf, J.M.: Updating finite Markov chains by using techniques of group inversion. J. Statist. Comput. Simulation 11, 163–181 (1980)
Wei, Y., Wang, G.: The perturbation theory for the Drazin inverse and its applications. Linear Algebra Appl. 258, 179–186 (1997)
Wei, Y.: On the perturbation of the group inverse and the oblique projection. Appl. Math. Comput. 98, 29–42 (1999)
Campbell, S.L., Meyer, C.D.: Continuality properties of the Drazin inverse. Linear Algebra Appl. 10, 77–83 (1975)
Wei, Y., Deng, C.: A note on additive results for the Drazin inverse. Linear Multilinear Algebra 59, 1319–1329 (2011)
Hartwig, R.E., Shoaf, J.M.: Group inverse and Drazin inverse of bidiagonal and triangular Toeplitz matrices. Austral J. Math. 24A, 10–34 (1977)
Meyer, C.D., Rose, N.J.: The index and the Drazin inverse of block triangular matrices. SIAM J. Appl. Math. 33, 1–7 (1977)
Wei, Y.: Expressions for the Drazin inverse of a \(2\times 2\) block matrix. Linear Multilinear Algebra 45, 131–146 (1998)
Castro-González, N.: Additive perturbation results for the Drazin inverse. Linear Algebra Appl. 397, 279–297 (2005)
Bauer, F., Fike, C.: Norms and exclusion theorem. Numer. Math. 2, 137–141 (1960)
Eisensat, S.C., Ipsen, I.C.F.: Three absolute perturbation bounds for matrix eigenvalues imply relative bounds. SIAM J. Matrix Anal. Appl. 20, 149–158 (1998)
Prasolov, V.V.: Problems and Theorems in Linear Algebra. American Mathematical Society, Providence, RI (1994)
Wei, Y., Wu, H.: The perturbation of the Drazin inverse and oblique projection. Appl. Math. Lett. 13, 77–83 (2000)
Wei, Y., Wu, H.: Convergence properties of Krylov subspace methods for singular linear systems with arbitrary index. J. Comput. Appl. Math. 114, 305–318 (2000)
Wei, Y.: The Drazin inverse of a modified matrix. Appl. Math. Comput. 125, 295–301 (2002)
Wei, Y., Li, X.: An improvement on perturbation bounds for the Drazin inverse. Numer Linear Algebra Appl. 10, 563–575 (2003)
Wei, Y., Li, X., Bu, F.: A perturbation bound of the Drazin inverse of a matrix by separation of simple invariant subspaces. SIAM J. Matrix Anal. Appl. 27, 72–81 (2005)
Wei, Y., Diao, H.: On group inverse of singular Toeplitz matrices. Linear Algebra Appl. 399, 109–123 (2005)
Wei, Y., Diao, H., Ng, M.K.: On Drazin inverse of singular Toeplitz matrix. Appl. Math. Comput. 172, 809–817 (2006)
Wei, Y., Li, X., Bu, F., Zhang, F.: Relative perturbation bounds for the eigenvalues of diagonalizable and singular matrices-application of perturbation theory for simple invariant subspaces. Linear Algebra Appl. 419, 765–771 (2006)
Castro-González, N., Dopazo, E., Robles, J.: Formulas for the Drazin inverse of special block matrices. Appl. Math. Comput. 174, 252–270 (2006)
Du, H., Deng, C.: The representation and characterization of Drazin inverses of operators on a Hilbert space. Linear Algebra Appl. 407, 117–124 (2005)
Koliha, J.J.: The Drazin and Moore-Penrose inverse in \(C^*\)-algebras. Mathematical Proceedings of the Royal Irish Academy 99A, 17–27 (1999)
Cvetković-Ilić, D.S., Djordjević, D.S., Wei, Y.: Additive results for the generalized Drazin inverse in a Banach algebra. Linear Algebra Appl. 418, 53–61 (2006)
Koliha, J.J.: A generalized Drazin inverse. Glasgow Math. J. 38, 367–381 (1996)
Harte, R.E.: Invertibility and Singularity for Bounded Linear Operators. Marcel Dekker, New York (1988)
Harte, R.E.: On quasinilpotents in rings. PanAm. Math. J. 1, 10–16 (1991)
Harte, R.E.: Spectral projections. Irish Math. Soc. Newsletter 11, 10–15 (1984)
Han, J.K., Lee, H.Y., Lee, W.Y.: Invertible completions of \(2 \times 2\) upper triangular operator matrices. Proc. Amer. Math. Soc. 128, 119–123 (1999)
Castro-González, N., Koliha, J.J.: New additive results for the g-Drazin inverse. Proc. Royal Soc. Edinburgh 134A, 1085–1097 (2004)
Djordjević, D.S., Stanimirović, P.S.: On the generalized Drazin inverse and generalized resolvent. Czechoslovak Math. J. 51(126), 617–634 (2001)
Deng, C.Y., Wei, Y.: New additive results for the generalized Drazin inverse. J. Math. Anal. Appl. 370, 313–321 (2009)
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
Copyright information
© 2017 Springer Nature Singapore Pte Ltd.
About this chapter
Cite this chapter
Cvetković Ilić, D.S., Wei, Y. (2017). Additive Results for the Drazin Inverse. In: Algebraic Properties of Generalized Inverses. Developments in Mathematics, vol 52. Springer, Singapore. https://doi.org/10.1007/978-981-10-6349-7_6
Download citation
DOI: https://doi.org/10.1007/978-981-10-6349-7_6
Published:
Publisher Name: Springer, Singapore
Print ISBN: 978-981-10-6348-0
Online ISBN: 978-981-10-6349-7
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)