moorepenrose inverse.pdf
TRANSCRIPT
-
7/22/2019 MoorePenrose inverse.pdf
1/9
A generalization of the MoorePenrose inverse relatedto matrix subspaces ofCnm
Antonio Surez, Luis Gonzlez *
Department of Mathematics, University of Las Palmas de Gran Canaria, 35017 Las Palmas de Gran Canaria, Spain
a r t i c l e i n f o
Keywords:
MoorePenrose inverse
Frobenius norm
S-MoorePenrose inverse
Approximate inverse preconditioning
Constrained least squares problem
a b s t r a c t
A natural generalization of the classical MoorePenrose inverse is presented. The so-called
S-MoorePenrose inverse of anm n complex matrixA, denoted by AyS, is defined for any
linear subspaceSof the matrix vector space Cnm. TheS-MoorePenrose inverseAySis char-
acterized using either the singular value decomposition or (for the full rank case) the
orthogonal complements with respect to the Frobenius inner product. These results are
applied to the preconditioning of linear systems based on Frobenius norm minimization
and to the linearly constrained linear least squares problem.
2010 Elsevier Inc. All rights reserved.
1. Introduction
Let Cmn Rmndenote the set of allmncomplex (real) matrices. Throughout this paper, the notations AT; A; A1,r(A)andtr(A) stand for the transpose, conjugate transpose, inverse, rank and trace of matrix A, respectively. Thegeneral reciprocal
was described by Moore in 1920[1]and independently rediscovered by Penrose in 1955[2], and it is nowadays called the
MoorePenrose inverse. The MoorePenrose inverse of a matrix A 2 Cmn (sometimes referred to as the pseudoinverse ofA)
is the unique matrix Ay 2 Cnm satisfying the four Penrose equations
AAyA A; A
yAA
y Ay; AAy AAy; A
yA AyA: 1:1
For more details on the MoorePenrose inverse, see[3]. The pseudoinverse has been widely studied during the last decades
from both theoretical and computational points of view. Some of the most recent works can be found, e.g., in[410]and in
the references contained therein.
As is well-known, the MoorePenrose inverse Ay can be alternatively defined as the unique matrix that gives the mini-
mum Frobenius norm among all solutions to any of the matrix minimization problems
minM2Cnm kMAInkF; 1:2min
M2CnmkAMImkF; 1:3
whereIn denotes the identity matrix of order n and k kFstands for the matrix Frobenius norm.
Another way to characterize the pseudoinverse is by using the singular value decomposition (SVD). IfA URV is an SVD
ofA then
Ay VRyU: 1:4
0096-3003/$ - see front matter 2010 Elsevier Inc. All rights reserved.doi:10.1016/j.amc.2010.01.062
* Corresponding author.
E-mail address: [email protected](L. Gonzlez).
Applied Mathematics and Computation 216 (2010) 514522
Contents lists available at ScienceDirect
Applied Mathematics and Computation
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c at e / a m c
http://dx.doi.org/10.1016/j.amc.2010.01.062mailto:[email protected]://www.sciencedirect.com/science/journal/00963003http://www.elsevier.com/locate/amchttp://www.elsevier.com/locate/amchttp://www.sciencedirect.com/science/journal/00963003mailto:[email protected]://dx.doi.org/10.1016/j.amc.2010.01.062 -
7/22/2019 MoorePenrose inverse.pdf
2/9
Also, the pseudoinverse ofA can be defined via a limiting process as [11]
Ay lim
d!0A
AA dIm1 lim
d!0AA dIn
1A
and by the explicit algebraic expression[3]
Ay CCC1BB1B
based on the full rank factorization
A BC; B2 Cmr; C2 Crn; rA rB rC r:
In particular, ifA 2 Cmn has full row rank (full column rank, respectively) then Ay is simply the unique solution to prob-
lem(1.2)(to problem(1.3), respectively), given by the respective explicit expressions[3]
Ay
i A AA
1iff rA m;
ii AA 1
A
iff rA n;
( 1:5
so thatAy is a right inverse (left inverse, respectively) ofA ifr(A) =m (r(A) = n, respectively).
A natural generalization of the MoorePenrose inverse, the so-called left or rightS-MoorePenrose inverse, arises simply
by taking in Eq.(1.2)or (1.3), respectively, the minimum over an arbitrary matrix subspaceSofCnm, instead of over the
whole matrix space Cnm. This is the starting point of this paper, that has been organized as follows.
In Section2, we define theS-MoorePenrose inverse and we provide an explicit expression based on the SVD of matrixA
(which generalizes Eq.(1.4)forA
y
), as well as an alternative expression for the full rank case, in terms of the orthogonal com-plements with respect to the Frobenius inner product (which generalizes Eq.(1.5)for Ay). Next, in Section3we apply these
results to the preconditioning of linear systems based on Frobenius norm minimization and to the linear least squares prob-
lem subject to some linear restrictions. Finally, in Section4 we present our conclusions.
2. The S-MoorePenrose inverse
Eqs. (1.2) and (1.3), regarding the least squares characterization of the MoorePenrose inverse, can be generalized as
follows.
Definition 2.1. Let A 2 Cmn and letSbe a linear subspace ofCnm. Then
(i) The left MoorePenrose inverse ofA with respect to Sor, for short, the left S-MoorePenrose inverse ofA, denoted byAyS;l, is the minimum Frobenius norm solution to the matrix minimization problem
minM2S
kMAInkF: 2:1
(ii) The right MoorePenrose inverse ofA with respect to Sor, for short, the rightS-MoorePenrose inverse ofA, denoted
byAyS;r, is the minimum Frobenius norm solution to the matrix minimization problem
minM2S
kAMImkF: 2:2
Remark 2.1. Note that forS Cnm, the left and rightS-MoorePenrose inverses become the MoorePenrose inverse. That is,
the left and right S-MoorePenrose inverses generalize the definition, via the matrix minimization problems(1.2) and (1.3),
respectively, of the classical pseudoinverse. Moreover, when A 2 Cnn is nonsingular and A1 2Sthen the left and right S-
MoorePenrose inverses are just the inverse. Briefly:
Ay
Cnm ;l
Ay
Ay
Cnm ;r
and AyS;l A
1 AyS;r for allSs:t: A
1 2S:
Remark 2.2. In particular, if matrix A has full row rank (full column rank, respectively) then it has right inverses (left
inverses, respectively). Thus, due to the uniqueness of the orthogonal projection of the identity matrix In (Im, respectively)
onto the matrix subspaceSA#Cnn (AS#Cmm, respectively), we conclude that, for these special cases,Definition 2.1simply
affirms thatAyS;l(AyS;r, respectively) is just the unique solution to problem (2.1)(to problem(2.2), respectively). The following
simple example illustrates this fact.
Example 2.1. Forn m 2 and
A 1 1
2 0
; S span
1 0
0 0
a 0
0 0
a2C
;
on one hand, we have
A. Surez, L. Gonzlez / Applied Mathematics and Computation 216 (2010) 514522 515
http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?- -
7/22/2019 MoorePenrose inverse.pdf
3/9
AyS;l
AI2
2F
minM2S
kMAI2k2Fmin
a2Cja 1j2 jaj2 1;
so that
AyS;l
12
0
0 0
!; A
yS;lAI2
F
3
2
and on the other hand, we have
AAyS;r I2
2F
minM2S
kAMI2k2Fmin
a2Cja 1j2 j2aj2 1;
so that
AyS;r
15
0
0 0
!; AA
yS;r I2
F
9
5:
Remark 2.3. Example 2.1illustrates three basic differences between the MoorePenrose and the S-MoorePenrose inverses.
First, we have that, in general, AyS;l AyS;r. Second, neitherA
yS;l, norA
yS;rsatisfies, in general, none of the Penrose equations given
in(1.1). Finally, although matrix A has full rank, neither AyS;l is a left inverse, nor AyS;ris a right inverse of matrix A.
Throughout this paper we address only the case of the right S-MoorePenrose inverseAyS;r, but analogous results can be
obtained for the left S-MoorePenrose inverseAy
S;l.Next theorem provides us with an explicit expression for AyS;rfrom an SVD of matrix A.
Theorem 2.1. Let A2 Cmn and let S be a linear subspace ofCnm. Let A URV be an SVD of matrix A. Then
AyS;rVR
yVSU;r
U and AAyS;r Im
F
RRyVSU;r
Im
F: 2:3
Moreover,
Ay AyS;r
F
Ry RyV SU;r
F: 2:4
Proof. For allM2 S, we have N VMU2 VSUand M VNU. Using the invariance property of the Frobenius norm under
multiplication by unitary matrices, we get
kA eMImkFminM2S
kAMImkF minN2V SU
kURVVNU ImkF minN2VSU
kURNImUkF
minN2VSU
kRNImkF kReN ImkF; 2:5
that is, eN2VSUis a solution to the matrix minimization problem
minN2VSU
kRNImkF 2:6
if and only if eMVeNU 2Sis a solution to problem(2.2).Now, let eNbe any solution to problem(2.6). Since Ry
VSU;r is the minimum Frobenius norm solution to problem(2.6), we
have
RyVSU;r
F6 k
eNkF
and then, using again the fact that the Frobenius norm is unitarily invariant, we have
VRyV SU;r
U
F6 kVeNUkF for any solutioneNto problem 2:6; i:e:;
VRyV SU;r
U
F6 k eMkF for any solutioneMto problem 2:2:
This proves that the minimum Frobenius norm solution AyS;rto problem(2.2)is VRyV SU;r
U, and the right-hand equality of
Eq.(2.3) immediately follows from Eq.(2.5).
Finally, Eq.(2.4)immediately follows using the well-known expressionAy VRyU and the left-hand equality of Eq.(2.3),
that is,
Ay AyS;r
F
VRyU VRyVSU;rU
F
Ry RyV SU;r
F:
Remark 2.4. Formula (2.3) for the right S-MoorePenrose inverse generalizes the well-known expression (1.4) for the
MoorePenrose inverse in terms of an SVD ofA, since (seeRemark 2.1)
516 A. Surez, L. Gonzlez/ Applied Mathematics and Computation 216 (2010) 514522
-
7/22/2019 MoorePenrose inverse.pdf
4/9
S Cnm ) VSU Cnm ) Ay AyS;rVRyVSU;r
U VRyU:
For full column rank matrices, the next theorem provides an expression for AyS;r involving the orthogonal complement of
subspaceS(denoted, as usual, by S?) with respect to the Frobenius inner product h; iF.
Theorem 2.2. Let A2 Cmn with rA n and let S be a linear subspace ofCnm. Then
AyS;r A
A
1A
Q; for some Q2S?: 2:7
Proof. Let
N AAAyS;r2 Cnm and QNA 2 Cnm:
SincerAA n, we get
AyS;r A
A1N AA1A Q; for someQ2 Cnm 2:8
and we only need to prove that Q2S?.
Using the fact thatAAyS;ris the orthogonal projection of the identity matrix onto the subspace AS, we have
AAyS;r Im;AM
D EF0 for all M2 S; i:e:;
tr AAyS;r Im MA 0 for all M2 S; i:e:;tr A
AA
yS;r A
M
0 for all M2 S;
but using Eq.(2.8), we have
AAA
yS;r A
AAAA1A Q A Q
and then
trQM 0; i:e:; hQ;MiF0 for all M2 S; i:e:; Q2S?:
Remark 2.5. Eq.(2.7)generalizes the well-known formula(1.5)(ii) for the MoorePenrose inverse of full column rank matri-
ces. Indeed (seeRemark 2.1)
S Cnm ) S? f0g ) Q0 ) AyC
nm ;r AA1A Ay:
An application ofTheorem 2.2is given in the next corollary.
Corollary 2.1. Let A2 Cmn with rA n. Let s2 Cm f0gand let0=s be the annihilator subspace of s in Cnm, i.e.,
0=s fM2 CnmjMs 0g:
Then
Ay0=s;r A
A1A Im ksk
22 ss
; 2:9
wherek k2 stands for the usual vector Euclidean norm.
Proof. Using Eq.(2.7)for S 0=s, we have
Ay0=s;r A
A1A Q; for someQ2 0=s
?
and since the orthogonal complement of the annihilator subspace0=sis given by[12]
0=s?
fvsjv2 Cng;
we get
Ay0=s;r A
A
1A
vs; for some v2 Cn: 2:10
Multiplying both sides of Eq.(2.10)by vector s, we have
Ay0=s;rs A
A1A vss
and, sinceAy0=s;r2 0=s, we get
A. Surez, L. Gonzlez / Applied Mathematics and Computation 216 (2010) 514522 517
-
7/22/2019 MoorePenrose inverse.pdf
5/9
0 A
vss ) v 1
ksk22A
s
and substituting vby its above expression in Eq. (2.10), we get
Ay0=s;r A
A1 A
1
ksk22A
ss
! AA1AIm ksk
22 ss
:
Remark 2.6. According to formula(1.5)(ii) for the full column rank case, Eq. (2.9)can be simply rewritten as
Ay0=s;rA
yIm
ss
ss
;
a curious representation of the right S-MoorePenrose inverse ofA(Sbeing the annihilator subspace0=sofs) as the product
of the MoorePenrose inverse ofA and an element ofS, namely the elementary annihilatorImss
ssof vector s.
Regarding the full column rank case, let us mention that, in addition to formulas (2.3) and (2.7), other explicit expressions
forAyS;r more appropriate for computational purposes can be given using an arbitrary basis of subspaceS. The basic idea
simply consists of expressing the orthogonal projection AAyS;rofIm onto the subspaceASby its expansion with respect to an
orthonormal basis ofAS(after using the GramSchmidt orthonormalization procedure, if necessary). These expressions have
been developed in[13]for the special case of real n n nonsingular matrices, and they have been applied to the precondi-
tioning of large linear systems and illustrated with some numerical experiments corresponding to real-world cases.
The next two lemmas state some spectral properties of the matrix product AAyS;rthat will be used in the next section.
Lemma 2.1. Let A2 Cmn and let S be a linear subspace ofCnm. Then
AAyS;r Im
2F
mtr AAyS;r
; 2:11
AAyS;r
2F
tr AAyS;r
: 2:12
Lemma 2.2. Let A2 Cmn and let S be a linear subspace ofCnm. Letfkigmi1 and frig
mi1 be the sets of eigenvalues and singular
values, respectively, of matrix AAyS;rarranged, as usual, in nonincreasing order of their modules. Then
Xmi1
r2i Xmi1
ki 6 m: 2:13
We omit the proofs ofLemmas 2.1 and 2.2, since these proofs are respectively similar to those of Lemmas 2.1 and 2.3
given in[18]in the context of the spectral analysis of the orthogonal projections of the identity onto arbitrary subspaces
ofRnn.
3. Applications
3.1. Applications to preconditioning
In numerical linear algebra, the convergence of the iterative Krylov methods[1416]used for solving the linear system
Ax b; A2R
nn
; A nonsingular; x; b2R
n1
3:1can be improved by transforming it into the right preconditioned system
AMy b; x My: 3:2
The approximate inverse preconditioning(3.2)of system(3.1)is performed in order to get a preconditioned matrix AMas
close as possible to the identity in some sense, taking into account the computational cost required for this purpose. The
closeness of AM to In may be measured by using a suitable matrix norm like, for instance, the Frobenius norm. In this
way, the optimal (in the sense of the Frobenius norm) right preconditioner of system (3.1), among all matrices Mbelonging
to a given subspaceSofRnn, is precisely the rightS-MoorePenrose inverseAyS;r, i.e., the unique solution to the matrix min-
imization problem (Remark 2.2)
minM2S
kAMInkF; A2 Rnn; rA n; S#Rnn: 3:3
Alternatively, we can transform system (3.1)into the left preconditioned system MAx=Mb, which is now related to the left S-
MoorePenrose inverseAyS;l .
518 A. Surez, L. Gonzlez/ Applied Mathematics and Computation 216 (2010) 514522
-
7/22/2019 MoorePenrose inverse.pdf
6/9
In the next theorem, we apply the spectral property (2.13)to the special case of the optimal preconditioner AyS;r, i.e., the
solution to problem(3.3). We must assume that matrix AyS;ris nonsingular, in order to get a nonsingular coefficient matrix
AAyS;rof the preconditioned system(3.2). Hence, all singular values of matrix AAyS;rwill be greater than zero, and they will be
denoted as
r1 P r2 P P rn> 0:
Theorem 3.1. Let A2 Rnnbe nonsingular and let S be a vector subspace ofRnn such that the solution Ay
S;r
to problem(3.3)is
nonsingular. Then the smallest singular value rnof the preconditioned matrix AAyS;rlies in the interval (0,1]. Moreover, we have the
following lower and upper bounds on the distance dIn;AS
1 rn26 AA
yS;rIn
2F6 n 1 r2n
: 3:4
Proof. UsingLemma 2.2for m =n, we get
nr2n 6Xni1
r2i 6 n ) r2n 6 1 ) rn 6 1:
To prove the left-hand inequality of Eq.(3.4), we use the well-known fact that singular values continuously depend on their
arguments[17], i.e., for all A; B2 Rnn and for all i 1; 2;. . .; n
jriA riBj 6 kABkF ) jrn 1j 6 AAyS;r In
F:
Finally, to prove the right-hand inequality of Eq.(3.4), we use Eqs.(2.11) and (2.12)for m =n
AAyS;rIn
2F
ntr AAyS;r
n AAyS;r
2F
nXni1
r2i 6 n 1 r2n
:
Remark 3.1. Eq.(3.4)states that AAyS;r I
F
decreases to 0 at the same time as the smallest singular valuernof the precon-
ditioned matrixAAyS;rincreases to 1. That is, the closeness ofAAyS;rto the identity is determined by the closeness ofrn to the
unity. In the extremal case, rn 1 iffAyS;r A
1 iffA1 2S. A more complete version ofTheorem 3.1, involving not only the
smallest singular value, but also the smallest eigenvalues modulus, of the preconditioned matrix, can be found in [18]. The
proof presented here is completely different, more direct and simpler than the one given in [18].
Remark 3.2. In numerical linear algebra, the theoretical effectiveness analysis of the preconditioners Mto be used, is basedon the spectral properties of the preconditioned matrix AM (e.g., condition number, clustering of eigenvalues and singular
values, departure from normality, etc.). For this reason, the spectral properties of matrix AAyS;r, stated inLemmas 2.1 and
2.2 and in Theorem 3.1, are especially useful when the left S-MoorePenrose inverse is the optimal preconditioner AyS;rdefined by problem(3.3). We refer the reader to [18]for more details about the theoretical effectiveness analysis of these
preconditioners.
3.2. Applications to the constrained least squares problem
TheS-MoorePenrose inverse can be applied to the linear least squares problem subject to a set of homogeneous linear
constraints. This is in accordance with the well-known fact that the MoorePenrose inverse can be applied to the linear least
squares problem without restrictions[19], the major application of the MoorePenrose inverse[20]. More precisely, let us
recall that, regarding the linear system Ax b; A2 Rmn
, the unique solution (ifA has full column rank) or the minimumnorm solution (otherwise) of the least squares problem
minx2Rn
kAxbk2
is given by
x Ayb:
Here we consider as well the linear systemAx b; A2 Rmn, but now subject to the conditionx2 T, whereTis a given vector
subspace ofRn. Denoting byBx = 0 the system of Cartesian equations of subspace T, this problem is usually formulated as
minx2Rn
kAxbk2 subject to Bx 0 3:5
and we refer the reader to[11]and to the references contained therein, for more details about the linearly constrained linear
least squares problem.
A. Surez, L. Gonzlez / Applied Mathematics and Computation 216 (2010) 514522 519
-
7/22/2019 MoorePenrose inverse.pdf
7/9
Next theorem provides us with a new, different approach in terms of the S-MoorePenrose inverse to problem(3.5).
Our theorem covers all possible situations that one can imagine for both the full rank and the rank deficient cases, e.g., Ax=b
consistent but Ax b; x2 T inconsistent; Ax=b consistent and underdetermined while Ax b; x2 T consistent and un-
iquely determined; bothAx =b and Ax b; x2 T inconsistent, etc.
Theorem 3.2. Let A2 Rmn and b2 Rm. Let T be a linear subspace ofRn of dimension p. Then the minimum norm solution to the
constrained least squares problem
minx2T
kAxbk2
3:6
is given by
xTAySP;r
b; 3:7
where P is the n p matrix whose columns are the vectors of an arbitrary basis of T and SPPRpm
#Rnm. In particular, if A has
full column rank then xTis the unique least squares solution to problem(3.6).
Proof. LetB fv1;. . . ;vpgbe a given orthonormal basis ofTand letPbe thenpmatrix whose columns arev1;. . . ;vp. For a
clearer exposition, we subdivide the proof into the following three parts:
(i) Writing any vector x 2 T asx Pa a2 Rp, we get
kA~xbk2 minx2T
kAxbk2 mina2Rp
kAPa bk2 kAP~a bk2;
that is, ~a2 Rp is a solution to the unconstrained least squares problem
mina2Rp
kAPa bk2 3:8
if and only if ~x P~a2 Tis a solution to the constrained least squares problem (3.6).
Now, let ~abe any solution to problem (3.8). Since a APyb is the minimum norm solution to problem(3.8), we have
kAPybk2 6 k~ak2
and then, using the fact that Phas orthonormal columnsPTP Ip, we have
kPak22 PaTPa aTa kak22 for all a2 R
p
and then
kPAPy
bk2 6 kP~ak2 for any solution ~a to problem 3:8; i:e:;kPAPybk2 6 k~xk2 for any solution ~x to problem 3:6:
This proves that the minimum norm solution xT to problem(3.6)can be expressed as
xTPAPyb for any orthonormal basisB of T: 3:9
(ii) Writing any matrixM2 SPPRpm asM PN N2 Rpm, we get
kA eMImkFminM2SP
kAMImkF minN2Rpm
kAPNImkF kAPeN ImkF;
that is, eN2 Rpm is a solution to the unconstrained matrix minimization problem
minN2Rpm
kAPNImkF 3:10
if and only if eMPeN2SPis a solution to the constrained matrix minimization problemmin
M2SPPRpm
kAMImkF: 3:11
Now, let eNbe any solution to problem(3.10). Since APy is the minimum Frobenius norm solution to problem (3.10), wehave
kAPykF6 keNkF
and then, using again the fact that Phas orthonormal columns PTP Ip, we have
kPNk2FtrPNTPN trNTN kNk2F for allN2 R
pm
and then
kPAPykF6 kPeNkF for any solutione
Nto problem 3:10; i:e:;
kPAPykF6 k eMkF for any solutioneMto problem 3:11:
520 A. Surez, L. Gonzlez/ Applied Mathematics and Computation 216 (2010) 514522
-
7/22/2019 MoorePenrose inverse.pdf
8/9
This proves that the minimum Frobenius norm solution AySP;rto problem(3.11)can be expressed as
AySP;r
PAPy
for any orthonormal basis B of T: 3:12
(iii) Using Eqs.(3.9) and (3.12), we get
xTAySP;r
b for any orthonormal basisB of T: 3:13
To conclude the derivation of formula(3.7), we must proof that, although the assumption that B is an orthonormal basisofTis (in general) essential for Eqs.(3.9) and (3.12)to be hold, formula(3.13)for xTis valid not only for orthonormal bases,
but in fact for any basis (not necessarily orthonormal) of subspaceT. In other words, the vectorAySP;rbis independent of the
chosen basis. Indeed, letB,B0 be two arbitrary bases ofTand letP,P0 be thenpmatrices whose columns are the vectors ofB
andB 0, respectively. Writing matrix P0 asP0 PC, Cbeing ap p nonsingular matrix, we have
SP0 P0R
pm PCRpm PRpm SP ) AySP;r
b AySP0 ;r
b:
Finally, if in particularA has full column rank then A is left invertible and, due to the uniqueness of the orthogonal pro-
jectionAxTof vectorb onto the subspaceAT#Rm, we derive that the minimum norm solution xTA
ySP;r
bto problem(3.6)is
indeed the unique solution to this problem. h
Remark 3.3. In particular, for the special full rank case r(A) =n, the expression(3.9)for the unique least squares solution xTof problem(3.6), is valid for any basis (not necessarily orthonormal) ofT. Indeed, letB,B0 be two arbitrary bases ofTand letP,
P0 be thenpmatrices whose columns are the vectors ofB and B0, respectively. Writing matrixP0 asP0 PC,Cbeing appnonsingular matrix, we have
P0AP0
yPCAPC
yPCCyAP
yPCC1AP
yPAP
y;
where the reverse order law for the product (AP)Cholds because APhas full column rank, i.e.,
p rA rP n 6 rAP 6 minrA; rP p
andChas full row rank; see, e.g.,[3]. In conclusion, the minimum norm solution to problem (3.6)is given by
xTPAPyb
for any orthonormal basis of T if rA< n;
for any basis of T if rA n
and also, for both cases rA 6 n, by xTA
ySP;r
b for any basis ofT.
The following simple example illustratesTheorem 3.2.
Example 3.1. Let A 1 1 1 2 R13 and b 2 2 R11. Consider the subspace T x3 0 of R3. Then m 1; n 3;
p 2 and taking the (nonorthonormal) basis ofT :B f1; 0; 0; 1; 1; 0g, we get
P
1 1
0 1
0 0
0B@1CA ) SP PR21 ab
0
0B@1CA
8>:9>=>;
a;b2R
R31 ) AySP;r
1=2
1=2
0
0B@1CA
and then the minimum norm solution to problem(3.6)is given by
xTAySP;r
b
1
1
0
0B@
1CA:
4. Concluding remarks
The left and rightS-MoorePenrose inverses,AyS;land AyS;r, extend the idea of the standard inverse from nonsingular square
matrices not only to rectangular and singular square matrices (as the MoorePenrose inverse does), but also to invertible
matrices, themselves, by considering a subspace SofCnm not containingA1. Restricted to the right case, the two expres-
sions given forAyS;r(namely, the first one based on an SVD ofA; the second one involving the orthogonal complement of sub-
spaceS), respectively generalize the well-known corresponding formulae (1.4) and (1.5)(ii) for the MoorePenrose inverse.
TheS-MoorePenrose inverseAyS;rcan be seen as an useful alternative to Ay: it provides us with a computationally feasible
approximate inverse preconditioner for large linear systems (in contrast to the unfeasible computation ofAy A1), and it
also provides the minimum norm solution AySP;rb to the least squares problem subject to homogeneous linear constraints
(in contrast to the minimum norm solution Ayb to the unconstrained least squares problem). The behavior of the S-
MoorePenrose inverse with respect to both theoretical results and other applications, already known for the MoorePen-
rose inverse, can be considered for future researches.
A. Surez, L. Gonzlez / Applied Mathematics and Computation 216 (2010) 514522 521
-
7/22/2019 MoorePenrose inverse.pdf
9/9
Acknowledgments
This work was partially supported by Ministerio de Ciencia e Innovacin (Spain) and FEDER through Grant contract:
CGL2008-06003-C03-01/CLI.
References
[1] E.H. Moore, On the reciprocal of the general algebraic matrix, Bull. Am. Math. Soc. 26 (1920) 394395.
[2] R. Penrose, A generalized inverse for matrices, Proc. Cambridge Philos. Soc. 51 (1955) 406413.[3] A. Ben-Israel, T.N.E. Greville, Generalized Inverses: Theory and Applications, second ed., Springer-Verlag, NewYork, 2003.[4] J.K. Baksalary, O.M. Baksalary, Particular formulae for the MoorePenrose inverse of a columnwise partitioned matrix, Linear Algebra Appl. 421 (2007)
1623.[5] T. Britz, D.D. Olesky, P. van den Driessche, The MoorePenrose inverse of matrices with an acyclic bipartite graph, Linear Algebra Appl. 390 (2004) 47
60.[6] M.A. Rakha, On the MoorePenrose generalized inverse matrix, Appl. Math. Comput. 158 (2004) 185200.[7] Y. Tian, Using rank formulas to characterize equalities for MoorePenrose inverses of matrix products, Appl. Math. Comput. 147 (2004) 581600.[8] Y. Tian, Approximation and continuity of MoorePenrose inverses of orthogonal row block matrices, Appl. Math. Comput. (2008), doi:10.1016/
j.amc.2008.10.027.[9] F. Toutounian, A. Ataei, A new method for computing MoorePenrose inverse matrices, J. Comput. Appl. Math. (2008), doi:10.1016/j.cam.2008.10.008.
[10] X. Zhang, J. Cai, Y. Wei, Interval iterative methods for computing MoorePenrose inverse, Appl. Math. Comput. 183 (2006) 522532.[11] G.H. Golub, C.F. Van Loan, Matrix Computations, third ed., Johns Hopkins Press, Baltimore, USA, 1996.[12] A. Griewank, A short proof of the DennisSchnabel theorem, BIT 22 (1982) 252256.[13] G. Montero, L. Gonzlez, E. Flrez, M.D. Garca, A. Surez, Approximate inverse computation using Frobenius inner product, Numer. Linear Algebra
Appl. 9 (2002) 239247.[14] O. Axelsson, Iterative Solution Methods, Cambridge University Press, Cambridge, MA, 1994.
[15] A. Greenbaum, Iterative Methods for Solving Linear Systems, Frontiers Appl. Math., vol. 17, SIAM, Philadelphia, PA, 1997.[16] Y. Saad, Iterative Methods for Sparse Linear Systems, PWS Publishing Co., Boston, MA, 1996.[17] R.A. Horn, C.R. Johnson, Matrix Analysis, Cambridge University Press, Cambridge, MA, 1985.[18] L. Gonzlez, Orthogonal projections of the identity: spectral analysis and applications to approximate inverse preconditioning, SIAM Rev. 48 (2006)
6675.[19] R. Penrose, On best approximate solutions of linear matrix equations, Proc. Cambridge Philos. Soc. 52 (1956) 1719.[20] A. Ben-Israel, The Moore of the MoorePenrose inverse, Electron. J. Linear Algebra 9 (2002) 150157.
522 A. Surez, L. Gonzlez/ Applied Mathematics and Computation 216 (2010) 514522
http://dx.doi.org/10.1016/j.amc.2008.10.027http://dx.doi.org/10.1016/j.amc.2008.10.027http://dx.doi.org/10.1016/j.cam.2008.10.008http://dx.doi.org/10.1016/j.cam.2008.10.008http://dx.doi.org/10.1016/j.amc.2008.10.027http://dx.doi.org/10.1016/j.amc.2008.10.027