JMM Volume 5 Issue 1 Pages 1-14

Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

 

Journal of Mathematical Modeling


Vol. 5, No. 1, 2017, pp. 1-14 
JMM 

GGMRES: A GMRES–type algorithm


for solving singular linear equations with
index one
Alireza Ataeia∗ and Faezeh Toutounianb
a
Mathematics Department, Faculty of Science, Persian Gulf University, Iran
b
Department of Applied Mathematics, School of Mathematical Sciences,
Ferdowsi University of Mashhad, Iran
Emails: [email protected], [email protected]

Abstract. In this paper, an algorithm based on the Drazin generalized


conjugate residual (DGMRES) algorithm is proposed for computing the
group-inverse solution of singular linear equations with index one. Numeri-
cal experiments show that the resulting group-inverse solution is reasonably
accurate and its computation time is significantly less than that of group-
inverse solution obtained by the DGMRES algorithm.
Keywords: singular linear systems, DGMRES method, group-inverse solution,
Drazin-inverse solution, Krylov subspace methods.
AMS Subject Classification: 15A06, 15A09, 65F10, 65F50.

1 Introduction
Consider the linear system

Ax = b, (1)

where A ∈ CN ×N is a singular matrix and ind(A) is arbitrary. Here ind(A),


the index of A is the size of the largest Jordan block corresponding to the
zero eigenvalue of A. We recall that the Drazin-inverse solution of (1) is the
vector AD b, where AD is the Drazin-inverse of the singular matrix A. For

Corresponding author.
Received: 27 October 2016 / Revised: 6 November 2016 / Accepted: 6 November 2016.


c 2017 University of Guilan http://jmm.guilan.ac.ir
2 A. Ataei and F. Toutounian

the Drazin-inverse and its properties, we refer to [2] or [5]. In the important
special case k = 1, this matrix is called the group inverse of A and denoted
by A] .

The Drazin-inverse has various applications in the theory of finite Markov


chains [5], the study of singular differential and difference equations [5], the
investigation of Cesaro-Neumann iterations [9], cryptography [8], iterative
methods in numerical analysis [6, 7], multibody system dynamics [13] and
others.

It is well known that the representations of the Drazin (group) inverse


of matrices are very important not only in matrix theory, but also in singu-
lar differential and difference equations, probability statistical, numerical
analysis, game theory, econometrics, control theory and so on [2, 5], and
also singular systems with index one arise naturally in Markov chain mod-
elling [3, 11].

The problem of finding the solution of the form AD b for (1) is very com-
mon in the literature and many different techniques have been developed in
order to solve it. In [14], Sidi proposed a general approach to Krylov sub-
space methods for computing Drazin-inverse solution. And then, he gave
several Krylov subspace methods of Arnoldi, the Drazin generalized con-
jugate residual (DGCR) and Lancoze types. Moreover in [15, 16], Sidi has
continued to drive two Krylov subspace methods for computing AD b. One is
DGMRES method, which is implementation of the DGCR method for sin-
gular systems that is analogues to GMRES for non-singular systems. The
other one is the Drazin biconjugate gradient algorithm (DBI-CG) which is
Lanczos type algorithm. DGMRES, just like, GMRES method, is a stable
numerically and economical computationally and storage wise. DBI-CG
method, also just like the biconjugate gradient (BI-CG) for non–singular
systems, is a fast algorithm, but when we need a high accuracy, the algo-
rithm is invalid. In the present paper, we develop the group generalized
minimal residual (GGMRES) algorithm which is another implementation
of DGMRES, for solving the singular linear system (1) with ind(A) = 1.
By numerical examples, we show that the computation time of GGMRES
algorithm is substantially less than that of DGMRES algorithm.

The paper is organized as follows. In Section 2, we will give a review of


DGMRES. In Section 3, we will derive the GGMRES algorithm. In Section
4 the results of some numerical examples are given. Section 5 is devoted to
concluding remarks.
GGMRES: A GMRES–type algorithm for singular systems 3

2 DGMRES algorithm
DGMRES method is a Krylov subspace method for computing the Drazin–
inverse solution of consistent or inconsistent linear systems (1) [14, 16]. In
this method, there is no restriction on the matrix A. Thus, in general,
A is non-Hermitian, a := ind(A) is arbitrary, and the spectrum of A can
have any shape. DGMRES starts with an initial vectors x0 and generates
a sequence of vectors x0 , x1 , . . ., as
m−a
X
xm = x0 + ci Aa+i−1 r0 , r0 = b − Ax0 .
i=1

Then
m−a
X
rm = b − Axm = b − ci Aa+i r0 .
i=1
The Krylov subspace we will use is

Km−a {A; Aa r0 } = span{Aa r0 , Aa+1 r0 , . . . , Am−1 r0 }.

The vector xm produced by DGMRES satisfies

kAa rm k = min kAa (b − Ax)k2 . (2)


x∈x0 +Km−a {A;Aa r0 }

As xm = x0 + m−a a+i−1 r , we start by orthogonalizing the krylov


P
i=1 ci A 0
a a+1
vectors A r0 , A r0 , . . . , using the Arnoldi–Gram–Schmidt process, see
[1, 12], carried out numerically like the modified Gram–Schmidt process:
• For i = 1, 2, . . . , do
• Compute hji = (vj , Aυi ), j = 1, 2, . . . , i.
Compute v̂i = Avi − ij=1 vj hji .
P

• Let hi+1,i = kv̂i k2 and set vi+1 = v̂i /hi+1,i.
• EndDo
Consequently, we have a set of orthonormal vectors v1 , v2 , . . . , that satisfies
i+1
X
Avi = vj hji , i = 1, 2, . . . , (3)
j=1

as long as i ≤ q − 1, where q is the degree of the minimal polynomial of A


with respect to Aa r0 , hence with respect to v1 . Furthermore, for each k,

span{v1 , v2 , . . . , vk } = span{Aa r0 , Aa+1 r0 , . . . , Ak +a−1 r0 } = Kk (A; Aa r0 ).


(4)
4 A. Ataei and F. Toutounian

If we now define the N × k matrix V̂k by

V̂k = [v1 |v2 | . . . |vk ], k = 1, 2, . . . , (5)

then, for m ≤ m0 (for definition of m0 see [14] and [16]), we can write

xm = x0 + V̂m−a ξm , for some ξm ∈ Cm−a (6)

where we need to determine ξm . Since rm = r0 + AV̂m−a ξm , we have

Aa rm = Aa r0 + Aa+1 V̂m−a ξm = βv1 − Aa+1 V̂m−a ξm . (7)

Next, provided k ≤ q − 1, from (3) we can write


 
h11 h12 · · · ··· h1k
 h21 h22 · · · ··· h2k 
 
 0 h32 . . .
 .. .. 
. . 
AV̂k = V̂k+1 H̄k , H̄k = . . (8)
 
.. .. .. ..
 .. . . . . 
 ..
 
. .. . .. .. 
 . . hkk 
0 ··· ··· 0 hk+1,k

Note that H̄k ∈ C(k+1)×k and H̄k has full rank when k ≤ q −1 [16]. Now, by

using (7), (8), and V̂m+1 V̂m+1 = I(m+1)×(m+1) we can reduce the n×(m−a)
least squares problem of (2) to the (m + 1) × (m − a) least squares problem

kAa rm k = kV̂m+1 (βe1 − Ĥm ξm )k (9)


= min kV̂m+1 (βe1 − Ĥm ξ)k
ξ∈Cm−a

= min kβe1 − Ĥm ξk,


ξ∈Cm−a

where
Ĥm = H̄m H̄m−1 . . . H̄m−a , (10)

and Ĥm ∈ C(m+1)×(m−a) . In general, the value of n is very large and m  n,


which implies that the problem in (9) is very small. The minimization
problem (9) is accomplished by using the QR decomposition of Ĥm . For
more details we refer the reader to [14] and [16].
We now summarize the steps of DGMRES for the solution (1) where
a = ind(A) is known.
GGMRES: A GMRES–type algorithm for singular systems 5

Algorithm 1 DGMRES algorithm


1. Pick x0 and compute r0 = b − Ax0 and Aa r0 .
2. Compute β = kAa r0 k and set v1 = β −1 (Aa r0 ).
3. Orthogonalize the Krylov vectors Aa r0 , Aa+1 r0 , . . . , via the Arnoldi-
Gram-Schmidt process carried out like the modified Gram-Schmidt pro-
cess:
For i = 1, 2, . . . , do
Compute hij = (vj , Av Pi ), j = 1, 2, . . . , i.
i
Compute v̂i = Avi − j=1 vj hij .
Let hi+1,i = kv̂i k and set vi+1 = v̂i /hi+1,i .
4. EndDo
5. For k = 1, 2, . . . , form the matrices V̂k ∈ CN ×k and H̄k ∈ C(k+1)×k
as defined in (5) and (8), respectively.
6. For m = a + 1, ..., form the matrix Ĥm = H̄m H̄m−1 . . . H̄m−a .
7. Compute the QR factorization of Ĥm : Ĥm = Qm Rm ; Qm ∈
C(m+1)×(m−a) and Rm ∈ C(m−a)×(m−a) .
8. Solve the (upper triangular) system Rm ξm = β(Q∗m e1 ), where e1 =
[1, 0, . . . , 0]. p
9. Compute xm = x0 + V̂m−a ξ (then kA a r k = β 1 − kQ∗ e k2 ).
p m m m 1
10. Compute kAa rm k = β 1 − kQ∗m e1 k2 .

3 GGMRES algorithm
In this section, we develop a new implementation of the DGMRES algo-
rithm for the case ind(A) = 1.
(m) (m)
Let H̄i and H̄j represent the row i and the column j of H̄m , re-
spectively. By partitioning H̄m and Ĥm = H̄m H̄m−1 as
" #
(m)  T 
H̄1 dm
H̄m = and Hm =
b , (11)
R̄m Fm

respectively, where R̄m is an m × m upper triangular matrix, Fm is an


m × (m − 1) upper Hessenberg matrix, and dTm ∈ R1×(m−1) , we see that
(2) (1) (m+1) (m)
dT2 = H̄1 H̄1 , dTm+1 = (dTm |H̄1 H̄m ), m = 2, 3, . . . (12)
and
(1) (m)
F2 = R̄2 H̄1 , Fm+1 = (F̃m |R̄m+1 H̄m ), m = 2, 3, . . . (13)
where  
Fm
F̃m = .
0
6 A. Ataei and F. Toutounian

If m steps of the Arnoldi process have been taken and V̂m+1 , the Arnoldi
basis associated with DGMRES is of full rank, then Fm is of full rank. This
result follows from the fact that for the elements ĥi+2,i , i = 1, 3, . . . , m − 1,
of Ĥ, we have
ĥi+2,i = h̄i+2,i+1 h̄i+1,i 6= 0.
In what follows we suppose that Fm is of full rank.
In order to get the solution ξm of the least squares problem (9), we can
consider the normal equation
T T
Ĥm Ĥm ξm = β Ĥm e1 . (14)

The use of (11) implies that

dm dTm + Fm
T T

Fm ξm = β H
bm e1 = βdm .

Let

λm = β − dTm ξm , (15)

then, we have
T
Fm Fm ξm = λm dm . (16)
By assumption, Fm is of full rank and λm = 0 implies that ξm = 0, which
is not the desired solution. Therefore, λm must be nonzero. By defining
ξm
um = , (17)
λm
the equation (16) can be written as
T
Fm Fm um = dm . (18)

For solving this positive definite system, we form the QR factorization


 
Rm
Qm Fm = ,
o

where Rm is an (m − 1) × (m − 1) nonsingular upper triangular matrix and


Q ∈ Rm×m is an unitary matrix. This gives the following equation:
T
Rm Rm um = dm . (19)

With setting zm = Rm um , the vector um can be computed by solving


Tz
the lower and upper triangular systems Rm m = dm and Rm um = zm ,
respectively.
GGMRES: A GMRES–type algorithm for singular systems 7

From (19), we have uTm dm = kRm um k2 ≥ 0. So, the relations (15) and
(17) imply that

β
λm = >0 and ξm = λm um , (20)
1 + dTm u

which can be used for computing λm and ξm .


We note that Fm+1 can be obtained as a simple update of Fm by first
appending a row of zeros at the bottom of Fm and following that by ap-
(m)
pending the (m + 1)-vector R̄m+1 H̄m as the mth column. After forming
Fm , for obtaining Rm , we can factorize Fm by appling a series of Givens
rotations to the columns of Fm .
Now, we show that the norm kArm k can be obtained without actually
having to form xm and rm . From (9) and (11) there holds

kArm k22 = kβe1 − 


Ĥm ξm k22
dTm
= kβe1 − ξm k22
Fm
β − dTm ξm 2
 
= −Fm ξm


2
2
 
λ m
= −λm Fm um

2
= λ2m (1
" + kF
2
m um k2 )
  2 #
R m

= λ2m 1 + Q m um

o
2
= λ2m (1 + kRm um k22 )
= λ2m (1 + kzm k22 ).

So, we have
q
kArm k2 = λm 1 + kzm k22 .

This relation enables us to obtain the norms kArm k2 , m ≥ 1, without


actually having to form xm and rm . Now, we summarize the steps of the
new method, called GGMRES method, for the solution group of a linear
system (1) when A is singular and ind(A) = 1, as follows.
8 A. Ataei and F. Toutounian

Algorithm 2 GGMRES algorithm


1. Pick x0 and compute r0 = b − Ax0 and Ar0 .
2. Compute β = kAa r0 k and set v1 = β −1 (Aa r0 ).
3. Orthogonalize the Krylov vectors Aa r0 , Aa+1 r0 , . . . , via the Arnoldi-
Gram-Schmidtprocess carried out like the modified Gram-Schmidt pro-
cess:
For i = 1, 2, . . . , do
Compute hij = (vj , Av Pi ), j = 1, 2, . . . , i.
i
Compute v̂i = Avi − j=1 vj hij .
Let hi+1,i = kv̂i k and set vi+1 = v̂i /hi+1,i .
4. EndDo
5. For k = 1, 2, . . . , form the matrices V̂k ∈ CN ×k and H̄k ∈ C(k+1)×k as
defined in (5)and (8), respectively.
6. Form the vector dm and the matrix Fm by using the recursive formula
(2) (1) (k+1) (k)
dT2 = H̄1 H̄1 , dTk+1 = (dTk |H̄1 H̄k ), k = 2, 3, . . . , m − 1

and
(1) (k)
F2 = R̄2 H̄1 , Fk+1 = (F̃k |R̄k+1 H̄k ), k = 2, 3, . . . , m − 1,

Fk
where F̃k = and R̄k is defined in (11).
0
7. Compute the QR factorization of Fm : Fm = Qm Rm ; Qm ∈ Rm×m−1
and Rm ∈ R(m−1)×(m−1) .
T z = d and R u = z .
8. Solve Rm m m m m m
9. Compute λm = 1+dβT um and ξm = λm um .
m
10. Compute xm = x0 + V̂m−1
p ξm .
11. Compute kArm k2 = λm 1 + kzm k22 .

It is possible to implement the GGMRES algorithm in a progressive


manner. The columns of H̄m , Fm , and Rm can be computed step by step
and kArm k2 can be computed with no additional operations. In addition,
as DGMRES algorithm, the computation of the vector xm requires that
the m vectors v1 , v2 , . . . , vm−1 and x0 (all of dimension n) to be stored.

4 Numerical examples
To compare the behavior of the proposed GGMRES method discussed in
the previous section with the DGMRES method, in this section, we present
GGMRES: A GMRES–type algorithm for singular systems 9

numerical results for two examples. Our examples, which have a singular
coefficient matrix, are derived by the finite difference method for elliptic
partial differential equations. The numerical computations are performed in
MATLAB (R213a) with double precision. All computations were performed
running the code on an Intel (R) Core (TM) i7-2600, 3.40 GHz machine
with 8 GB of RAM memory using Windows 7 professional 64-bit operating
system. The initial vector x0 is the zero vector. All the tests were stopped
as soon as kArm k2 ≤ 10−12 .

Example 1. We form the linear system Ax = b by discretizing Poisson


equation with Neumann boundary conditions:

( 2
∂ ∂2
( ∂x 2 + ∂y 2
)u(x, y) = f (x, y), (x, y) ∈ Ω = [0, 1] × [0, 1],

∂n u(x, y) = ϕ(x, y), x, y ∈ ∂Ω.

This linear system also has been formed by Sidi [16] for testing DGMRES
algorithm. The problem has also been considered by Hank and Hochbruck
[10] for testing the Chebyshev-type semi-iterative method.
Let M be an odd integer, we discretize the Poisson equation on a uni-
form grid of mesh size h = 1/M via central differences, and then by taking
the unknowns in the red-black order we obtain the system Ax = b, where
the (M + 1)2 × (M + 1)2 nonsymmetric matrix A is as follows

 
A1 A2
A= , (21)
A3 A1

where A1 = 4I,

T2 −2I o ··· ··· ··· ··· o


 
..
−I T1 −I o .
 
 
  ..
o −I T2 −I o .
 
 
 .. ..  ..

. o −I T1 −I . 
.
A2 =  ,
 
.. .. .. .. .. ..

 . . . . . 
 .
 .. .. .. .. .. 

 . . . . . o 

 .. 
 . o −I T2 −I 
o ··· · · · · · · · · · o −2I T1
10 A. Ataei and F. Toutounian

and
T1 −2I o ··· ··· ··· ··· o
 
..
−I T2 −I o .
 
 
  ..
o −I T1 −I o .
 
 
 .. ..  ..

. o −I T2 −I . 
.
A3 =  .
 
.. .. .. .. .. ..

 . . . . . 
 .
 .. .. .. .. .. 

 . . . . . o 

 .. 
 . o −I T1 −I 
o ··· · · · · · · · · · o −2I T2
Here, I and o denote, respectively, the (M + 1)/2 × (M + 1)/2 identity and
zero matrices and the (M + 1)/2 × (M + 1)/2 matrices T1 and T2 are given
by
   
−2 o · · · · · · o −1 −1 o · · · o
 −1 −1 . . .
 ..   o −1 . . . . . .
 .. 
 . 
  . 

T1 =  o . . . . . . . . . .
..  , T2 =  ... .. .. .. .
   
. . . o
 ..  ..
   
..  
 . . −1 o   . −1 −1 
o · · · o −1 −1 o · · · · · · o −2
The numerical experiment is performed for M = 31, 63, 127.
Example 2. As shown in [4], applying 5−point central differences to the
partial differential equation
∂2U ∂2U ∂U
2
+ 2
+d = f (x, y), 0 < x, y < 1,
∂x ∂y ∂x
over the unit square Ω = (0, 1)×(0, 1) with the periodic boundary condition:
u(x, 0) = u(x, 1), u(0, y) = u(1, y),
yields a singular system with a nonsymmetric coefficient matrix. The mesh
size is chosen as h = 1/m for Ω, so that the resulting system has the
following n × n coefficient matrix (where n = m2 ):
 
Dm Im · · · Im
 Im Dm Im 
 
 . . . . . . 
1  . . . 
A := 2  
. . .
. (22)
h  .. .. .. 

 
 Im Dm Im 
Im Im Dm
GGMRES: A GMRES–type algorithm for singular systems 11

Here Im is the m × m unit matrix and Dm the m × m matrix is given by


 
−4 α+ α−
 α− −4 α+ 
 
Dm := 
 .. .. .. ,

 . . . 
 α− −4 α+ 
α+ α− −4

where α± = 1 ± dh 2 . The numerical experiment is done for d = 0.1, d = 0.3,


d = 0.5, and m = 60.
For the matrix A of both (21) and (22) the identity Ae = AT e = 0 holds,
so that Null(A) = Null(AT ) = Span{e}, where e = (1, 1, . . . , 1)T . Further-
more, ind(A) = 1, as mentioned in [10, 16]. Even if the continuous problem
has a solution, the discretized problem need not to be consistent. Here, we
consider only the Group-inverse solution of the system for arbitrary right
side b, not necessarily related to f and ϕ.
As [17], we first construct a consistent system with known solution
ŝ ∈ R(A) via ŝ = Ay, where y = [0, . . . , 0, 1]T . Then we perturb Aŝ, the
right-hand side of Ax = Aŝ = b̂, with a constant multiple of the null space
vector e and we obtain the right-hand side
e
b = b̂ + δ .
kek2

e
Consequently the system Ax = b̂ + δ kek 2
is solved for x. The perturbation
−2
parameter δ is selected as 10 in our experiments.
For these examples, the solution we are looking for is the vector ŝ, whose
components are zeros except

ŝ2M̂ 2 −M̂ = −1, ŝ2M̂ 2 −1 = −1, ŝ2M̂ 2 = −2, ŝ4M̂ 2 = 4,

where M̂ = (M + 1)/2 for Example 1 and except

ŝm = 1, ŝm2 −m = 1, ŝm2 −m+1 = α− , ŝm2 −1 = α+ , ŝm2 = 4,

for Example 2.
In Tables 1-4, we give the number of iterations (Its), the CPU time
(Time) required for convergence, and the error (Error) for the DGMRES
and GGMRES methods. As shown in Tables 1-4 the GGMRES algorithm
is effective and less expensive than the DGMRES algorithm.
12 A. Ataei and F. Toutounian

Table 1: Application of GGMRES implementation to the consistent singu-


lar system for Example 1.

Size of A 1024 × 1024 4096 × 4096 16384 × 16384


Method Its Time Error Its Time Error Its Time Error
DGMRES 164 0.33 8.87e − 013 310 4.20 9.92e − 13 471 29.46 9.00e − 13
GGMRES 164 0.24 8.67e − 013 310 3.23 9.88e − 13 471 24.34 9.83e − 13

Table 2: Application of GGMRES implementation to the inconsistent sin-


gular system for Example 1.

Size of A 1024 × 1024 4096 × 4096 16384 × 16384


Method Its Time Error Its Time Error Its Time Error
DGMRES 164 0.38 8.74e − 013 310 4.04 9.90e − 13 471 29.48 9.96e − 13
GGMRES 164 0.31 8.67e − 013 310 3.35 9.88e − 13 471 22.99 9.83e − 13

Table 3: Application of GGMRES implementation to the consistent singu-


lar system for Example 2 with m = 60(n=3600).

A 0.1 0.3 0.5


Method Its Time Error Its Time Error Its Time Error
DGMRES 217 1.74 9.97e − 13 240 2.11 9.70e − 13 246 2.27 9.35e − 13
GGMRES 217 1.38 9.93e − 13 240 1.60 9.68e − 13 246 1.64 9.30e − 13

Table 4: Application of GGMRES implementation to the inconsistent sin-


gular systemfor Example 2 with m = 60(n=3600).

A 0.1 0.3 0.5


Method Its Time Error Its Time Error Its Time Error
DGMRES 217 1.80 9.99e − 13 240 2.09 9.76e − 13 246 2.29 9.96e − 13
GGMRES 217 1.45 9.93e − 13 240 1.68 9.68e − 13 246 1.64 9.57e − 13

5 Conclusion

In this paper, we have presented a new method, called GGMRES, for com-
puting the group-inverse solution of singular linear equations with index
one. This method is based on DGMRES algorithm. Numerical experiments
show that the group-inverse solution obtained by this method is reasonably
accurate, and its computation time is less than that of solution obtained by
the DGMRES method. So, we can conclude that the GGMRES algorithm
is a robust and efficient tool for computing the group-inverse solution of
singular linear equations with index one.
GGMRES: A GMRES–type algorithm for singular systems 13

Acknowledgements
We would like to thank the editor and anonymous referees for their carefully
reading and useful comments. We are also grateful to editor in chief of the
journal for his comments that improved our results.

References
[1] W.E. Arnoldi, The principle of minimized iterations in the solution of
the matrix eigenvalue problem, Quart. Appl. Math. 9 (1951) 17–29.

[2] A. Ben-Israel and T.N.E. Grevile, Generalized Inverses: Theory and


Applications, second ed., Springer-Verlag, New York, 2003.

[3] A. Berman and R.J. Plemmons, Nonnegative Matrices in the Mathe-


matical Science, Academic Press, New York, 1979.

[4] P.N. Brown and H.F. Walker, GMRES on (nearly) singular systems,
SIAM J. Matrix Anal. Appl. 18 (1997) 37–51.

[5] S.L. Campell and C.D. Meyer, Generalized Inverses of Linear Trans-
formations, Pitman, London, 1979, Dover, New York, 1991.

[6] M. Eiermann, I. Marek and W. Niethammer, On the solution of sin-


gular linear systems of algebric equations by semiiterative methods,
Numer. Math. 53 (1988) 265–283.

[7] R.W. Freund and M. Hochbruck, On the use of two QMR algorithms
for solving singular systems and applications in Markov chain model-
ing, Numer. Linear algebra Appl. 1 (1994) 403–420.

[8] R.E. Hartwig and J. Levine, Applications of the Drazin inverse to the
Hill cryptographic system, Part III, Cryptologia 5 (1981) 67–77.

[9] R.E. Hartwig and F. Hall, Applications of the Drazin inverse to


Cesaro-Neumann iterations, in: S.L. Campbell (Ed.), Recent Appli-
cations of Generalized Inverses, Pitman, London, Program 66 (1982)
145–195.

[10] M. Hank and M. Hochbruck, A Chebyshev–like semiiteration for in-


consistent linear systems, Electron. Trans. Numer. Anal. 1 (1993) 315–
339.
14 A. Ataei and F. Toutounian

[11] I. Marrek and D.B. Szyld, Iterative and semi–iterative methods for
computing stationary probablity vectors of Markov operators, Math.
Comput. 61 (1993) 719–731.

[12] Y. Saad and M. Schultz, A generalized minimal reseidual algorithm


for solving non–symetric linear systems, SIAM J. Sci. Statist. Comput.
7 (1986) 56–869.

[13] B. Simeon, C. Fuhrer and P. Rentrop, The Drazin inverse in multibody


system dynamics, Numer. Math. 64 (1993) 521–539.

[14] A. Sidi, A unified approach to krylov subspace methods for the Drazin–
inverse solution of singular non-symmetric linear systems, Linear Al-
gebra Appl. 298 (1999) 99–113.

[15] A. Sidi and V. Kluzner, A Bi–CG type iterative method for Drazin
inverse solution of singular inconsistent non-symmetric linear systems
of arbitrary index, Electron. J. Linear Algebra Appl. 6 (1999) 72–94.

[16] A. Sidi, DGMRES: A GMRES–type algorithm for Drazin–inverse so-


lution of singular non–symmetric linear systems, Linear Algebra Appl.
335 (2001) 189–204.

[17] J. Zhou and Y. Wei, DFOM algorithm and error analysis for projection
methods for solving singular linear system, Appl. Math. Comput. 157
(2004) 313–329.

You might also like