A generalization of the Moore–Penrose inverse related to matrix subspaces of
-
Upload
antonio-suarez -
Category
Documents
-
view
217 -
download
0
Transcript of A generalization of the Moore–Penrose inverse related to matrix subspaces of
Applied Mathematics and Computation 216 (2010) 514–522
Contents lists available at ScienceDirect
Applied Mathematics and Computation
journal homepage: www.elsevier .com/ locate /amc
A generalization of the Moore–Penrose inverse relatedto matrix subspaces of Cn�m
Antonio Suárez, Luis González *
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:Moore–Penrose inverseFrobenius normS-Moore–Penrose inverseApproximate inverse preconditioningConstrained least squares problem
0096-3003/$ - see front matter � 2010 Elsevier Incdoi:10.1016/j.amc.2010.01.062
* Corresponding author.E-mail address: [email protected] (L. Gonzá
a b s t r a c t
A natural generalization of the classical Moore–Penrose inverse is presented. The so-calledS-Moore–Penrose inverse of an m� n complex matrix A, denoted by AyS, is defined for anylinear subspace S of the matrix vector space Cn�m. The S-Moore–Penrose inverse AyS is char-acterized using either the singular value decomposition or (for the full rank case) theorthogonal complements with respect to the Frobenius inner product. These results areapplied to the preconditioning of linear systems based on Frobenius norm minimizationand to the linearly constrained linear least squares problem.
� 2010 Elsevier Inc. All rights reserved.
1. Introduction
Let Cm�n ðRm�nÞ denote the set of all m� n complex (real) matrices. Throughout this paper, the notations AT; A�; A�1, r(A)
and tr(A) stand for the transpose, conjugate transpose, inverse, rank and trace of matrix A, respectively. The general reciprocalwas described by Moore in 1920 [1] and independently rediscovered by Penrose in 1955 [2], and it is nowadays called theMoore–Penrose inverse. The Moore–Penrose inverse of a matrix A 2 Cm�n (sometimes referred to as the pseudoinverse of A)is the unique matrix Ay 2 Cn�m satisfying the four Penrose equations
AAyA ¼ A; AyAAy ¼ Ay; ðAAyÞ� ¼ AAy; ðAyAÞ� ¼ AyA: ð1:1Þ
For more details on the Moore–Penrose inverse, see [3]. The pseudoinverse has been widely studied during the last decadesfrom both theoretical and computational points of view. Some of the most recent works can be found, e.g., in [4–10] and inthe references contained therein.
As is well-known, the Moore–Penrose 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
minM2Cn�m
kMA� InkF ; ð1:2Þ
minM2Cn�m
kAM � ImkF ; ð1:3Þ
where In denotes the identity matrix of order n and k � kF stands for the matrix Frobenius norm.Another way to characterize the pseudoinverse is by using the singular value decomposition (SVD). If A ¼ URV� is an SVD
of A then
Ay ¼ VRyU�: ð1:4Þ
. All rights reserved.
lez).
A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522 515
Also, the pseudoinverse of A can be defined via a limiting process as [11]
Ay ¼ limd!0
A�ðAA� þ dImÞ�1 ¼ limd!0ðA�Aþ dInÞ�1A�
and by the explicit algebraic expression [3]
Ay ¼ C�ðCC�Þ�1ðB�BÞ�1B�
based on the full rank factorization
A ¼ BC; B 2 Cm�r ; C 2 Cr�n; rðAÞ ¼ rðBÞ ¼ rðCÞ ¼ r:
In particular, if A 2 Cm�n 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 ¼ ðiÞ A� AA�ð Þ�1 iff rðAÞ ¼ m;
ðiiÞ A�Að Þ�1A� iff rðAÞ ¼ n;
(ð1:5Þ
so that Ay is a right inverse (left inverse, respectively) of A if r(A) = m (r(A) = n, respectively).A natural generalization of the Moore–Penrose inverse, the so-called left or right S-Moore–Penrose inverse, arises simply
by taking in Eq. (1.2) or (1.3), respectively, the minimum over an arbitrary matrix subspace S of Cn�m, instead of over thewhole matrix space Cn�m. This is the starting point of this paper, that has been organized as follows.
In Section 2, we define the S-Moore–Penrose inverse and we provide an explicit expression based on the SVD of matrix A(which generalizes Eq. (1.4) for Ay), 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 Section 3 we apply theseresults 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 Section 4 we present our conclusions.
2. The S-Moore–Penrose inverse
Eqs. (1.2) and (1.3), regarding the least squares characterization of the Moore–Penrose inverse, can be generalized asfollows.
Definition 2.1. Let A 2 Cm�n and let S be a linear subspace of Cn�m. Then
(i) The left Moore–Penrose inverse of A with respect to S or, for short, the left S-Moore–Penrose inverse of A, denoted byAyS;l, is the minimum Frobenius norm solution to the matrix minimization problem
minM2SkMA� InkF : ð2:1Þ
(ii) The right Moore–Penrose inverse of A with respect to S or, for short, the right S-Moore–Penrose inverse of A, denotedby AyS;r , is the minimum Frobenius norm solution to the matrix minimization problem
minM2SkAM � ImkF : ð2:2Þ
Remark 2.1. Note that for S ¼ Cn�m, the left and right S-Moore–Penrose inverses become the Moore–Penrose inverse. That is,the left and right S-Moore–Penrose inverses generalize the definition, via the matrix minimization problems (1.2) and (1.3),respectively, of the classical pseudoinverse. Moreover, when A 2 Cn�n is nonsingular and A�1 2 S then the left and right S-Moore–Penrose inverses are just the inverse. Briefly:
AyCn�m ;l ¼ Ay ¼ Ay
Cn�m ;r and AyS;l ¼ A�1 ¼ AyS;r for all S s:t: A�1 2 S:
Remark 2.2. In particular, if matrix A has full row rank (full column rank, respectively) then it has right inverses (leftinverses, respectively). Thus, due to the uniqueness of the orthogonal projection of the identity matrix In (Im, respectively)onto the matrix subspace SA # Cn�n (AS # Cm�m, respectively), we conclude that, for these special cases, Definition 2.1 simplyaffirms that AyS;l (AyS;r , respectively) is just the unique solution to problem (2.1) (to problem (2.2), respectively). The followingsimple example illustrates this fact.
Example 2.1. For n ¼ m ¼ 2 and
A ¼1 12 0
� �; S ¼ span
1 00 0
� �� �¼
a 00 0
� �� �a2C
;
on one hand, we have
516 A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522
AyS;lA� I2
��� ���2
F¼min
M2SkMA� I2k2
F ¼mina2Cja� 1j2 þ jaj2 þ 1;
so that
AyS;l ¼12 00 0
!; AyS;lA� I2
��� ���F¼ 3
2
and on the other hand, we have
AAyS;r � I2
��� ���2
F¼min
M2SkAM � I2k2
F ¼mina2Cja� 1j2 þ j2aj2 þ 1;
so that
AyS;r ¼15 00 0
!; AAyS;r � I2
��� ���F¼ 9
5:
Remark 2.3. Example 2.1 illustrates three basic differences between the Moore–Penrose and the S-Moore–Penrose inverses.First, we have that, in general, AyS;l – AyS;r . Second, neither AyS;l, nor AyS;r satisfies, in general, none of the Penrose equations givenin (1.1). Finally, although matrix A has full rank, neither AyS;l is a left inverse, nor AyS;r is a right inverse of matrix A.
Throughout this paper we address only the case of the right S-Moore–Penrose inverse AyS;r , but analogous results can beobtained for the left S-Moore–Penrose inverse AyS;l.
Next theorem provides us with an explicit expression for AyS;r from an SVD of matrix A.
Theorem 2.1. Let A 2 Cm�n and let S be a linear subspace of Cn�m. Let A ¼ URV� be an SVD of matrix A. Then
AyS;r ¼ VRyV�SU;rU� and AAyS;r � Im
��� ���F¼ RRyV�SU;r � Im
��� ���F: ð2:3Þ
Moreover,
Ay � AyS;r��� ���
F¼ Ry � RyV�SU;r
��� ���F: ð2:4Þ
Proof. For all M 2 S, we have N ¼ V�MU 2 V�SU and M ¼ VNU�. Using the invariance property of the Frobenius norm undermultiplication by unitary matrices, we get
kA eM � ImkF ¼minM2SkAM � ImkF ¼ min
N2V�SUkURV�VNU� � ImkF ¼ min
N2V�SUkUðRN � ImÞU�kF
¼ minN2V�SU
kRN � ImkF ¼ kReN � ImkF ; ð2:5Þ
that is, eN 2 V�SU is a solution to the matrix minimization problem
minN2V�SU
kRN � ImkF ð2:6Þ
if and only if eM ¼ V eNU� 2 S is a solution to problem (2.2).Now, let eN be any solution to problem (2.6). Since RyV�SU;r is the minimum Frobenius norm solution to problem (2.6), we
have
RyV�SU;r
��� ���F6 keNkF
and then, using again the fact that the Frobenius norm is unitarily invariant, we have
VRyV�SU;rU�
��� ���F6 kV eNU�kF for any solution eN to problem ð2:6Þ; i:e:;
VRyV�SU;rU�
��� ���F6 k eMkF for any solution eM to problem ð2:2Þ:
This proves that the minimum Frobenius norm solution AyS;r to problem (2.2) is VRyV�SU;rU�, 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 expression Ay ¼ VRyU� and the left-hand equality of Eq. (2.3),
that is,
Ay � AyS;r��� ���
F¼ VRyU� � VRyV�SU;rU
���� ���
F¼ Ry � RyV�SU;r
��� ���F: �
Remark 2.4. Formula (2.3) for the right S-Moore–Penrose inverse generalizes the well-known expression (1.4) for theMoore–Penrose inverse in terms of an SVD of A, since (see Remark 2.1)
A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522 517
S ¼ Cn�m ) V�SU ¼ Cn�m ) Ay ¼ AyS;r ¼ VRyV�SU;rU� ¼ VRyU�:
For full column rank matrices, the next theorem provides an expression for AyS;r involving the orthogonal complement ofsubspace S (denoted, as usual, by S?) with respect to the Frobenius inner product h�; �iF .
Theorem 2.2. Let A 2 Cm�n with rðAÞ ¼ n and let S be a linear subspace of Cn�m. Then
AyS;r ¼ ðA�AÞ�1ðA� þ QÞ; for some Q 2 S?: ð2:7Þ
Proof. Let
N ¼ ðA�AÞAyS;r 2 Cn�m and Q ¼ N � A� 2 Cn�m:
Since rðA�AÞ ¼ n, we get
AyS;r ¼ ðA�AÞ�1N ¼ ðA�AÞ�1ðA� þ QÞ; for some Q 2 Cn�m ð2:8Þ
and we only need to prove that Q 2 S?.Using the fact that AAyS;r is the orthogonal projection of the identity matrix onto the subspace AS, we have
AAyS;r � Im;AMD E
F¼ 0 for all M 2 S; i:e:;
tr AAyS;r � Im
� �M�A�
� �¼ 0 for all M 2 S; i:e:;
tr A�AAyS;r � A�� �
M�� �
¼ 0 for all M 2 S;
but using Eq. (2.8), we have
A�AAyS;r � A� ¼ ðA�AÞðA�AÞ�1ðA� þ QÞ � A� ¼ Q
and then
trðQM�Þ ¼ 0; i:e:; hQ ;MiF ¼ 0 for all M 2 S; i:e:; Q 2 S?: �
Remark 2.5. Eq. (2.7) generalizes the well-known formula (1.5)(ii) for the Moore–Penrose inverse of full column rank matri-ces. Indeed (see Remark 2.1)
S ¼ Cn�m ) S? ¼ f0g ) Q ¼ 0 ) AyCn�m ;r ¼ ðA
�AÞ�1A� ¼ Ay:
An application of Theorem 2.2 is given in the next corollary.
Corollary 2.1. Let A 2 Cm�n with rðAÞ ¼ n. Let s 2 Cm � f0g and let ½0=s� be the annihilator subspace of s in Cn�m, i.e.,
½0=s� ¼ fM 2 Cn�mjMs ¼ 0g:
Then
Ay½0=s�;r ¼ ðA�AÞ�1A� Im � ksk�2
2 ss�� �
; ð2:9Þ
where k � k2 stands for the usual vector Euclidean norm.
Proof. Using Eq. (2.7) for S ¼ ½0=s�, we have
Ay½0=s�;r ¼ ðA�AÞ�1ðA� þ QÞ; for some Q 2 ½0=s�?
and since the orthogonal complement of the annihilator subspace ½0=s� is given by [12]
½0=s�? ¼ fvs�jv 2 Cng;
we get
Ay½0=s�;r ¼ ðA�AÞ�1ðA� þ vs�Þ; for some v 2 Cn: ð2:10Þ
Multiplying both sides of Eq. (2.10) by vector s, we have
Ay½0=s�;rs ¼ ðA�AÞ�1ðA� þ vs�Þs
and, since Ay½0=s�;r 2 ½0=s�, we get
518 A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522
0 ¼ ðA� þ vs�Þs ) v ¼ � 1
ksk22
A�s
and substituting v by its above expression in Eq. (2.10), we get
Ay½0=s�;r ¼ ðA�AÞ�1 A� � 1
ksk22
A�ss� !
¼ ðA�AÞ�1A�ðIm � ksk�22 ss�Þ: �
Remark 2.6. According to formula (1.5)(ii) for the full column rank case, Eq. (2.9) can be simply rewritten as
Ay½0=s�;r ¼ Ay Im �ss�
s�s
� �;
a curious representation of the right S-Moore–Penrose inverse of A (S being the annihilator subspace ½0=s� of s) as the productof the Moore–Penrose inverse of A and an element of S, namely the elementary annihilator Im � ss�
s�s of vector s.
Regarding the full column rank case, let us mention that, in addition to formulas (2.3) and (2.7), other explicit expressionsfor AyS;r – more appropriate for computational purposes – can be given using an arbitrary basis of subspace S. The basic ideasimply consists of expressing the orthogonal projection AAyS;r of Im onto the subspace AS by its expansion with respect to anorthonormal basis of AS (after using the Gram–Schmidt orthonormalization procedure, if necessary). These expressions havebeen 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;r that will be used in the next section.
Lemma 2.1. Let A 2 Cm�n and let S be a linear subspace of Cn�m. Then
AAyS;r � Im
��� ���2
F¼ m� tr AAyS;r
� �; ð2:11Þ
AAyS;r��� ���2
F¼ tr AAyS;r
� �: ð2:12Þ
Lemma 2.2. Let A 2 Cm�n and let S be a linear subspace of Cn�m. Let fkigmi¼1 and frigm
i¼1 be the sets of eigenvalues and singularvalues, respectively, of matrix AAyS;r arranged, as usual, in nonincreasing order of their modules. Then
Xm
i¼1
r2i ¼
Xm
i¼1
ki 6 m: ð2:13Þ
We omit the proofs of Lemmas 2.1 and 2.2, since these proofs are respectively similar to those of Lemmas 2.1 and 2.3given in [18] in the context of the spectral analysis of the orthogonal projections of the identity onto arbitrary subspacesof Rn�n.
3. Applications
3.1. Applications to preconditioning
In numerical linear algebra, the convergence of the iterative Krylov methods [14–16] used for solving the linear system
Ax ¼ b; A 2 Rn�n; A nonsingular; x; b 2 Rn�1 ð3:1Þ
can be improved by transforming it into the right preconditioned system
AMy ¼ b; x ¼ My: ð3:2Þ
The approximate inverse preconditioning (3.2) of system (3.1) is performed in order to get a preconditioned matrix AM asclose as possible to the identity in some sense, taking into account the computational cost required for this purpose. Thecloseness of AM to In may be measured by using a suitable matrix norm like, for instance, the Frobenius norm. In thisway, the optimal (in the sense of the Frobenius norm) right preconditioner of system (3.1), among all matrices M belongingto a given subspace S of Rn�n, is precisely the right S-Moore–Penrose inverse AyS;r , i.e., the unique solution to the matrix min-imization problem (Remark 2.2)
minM2SkAM � InkF ; A 2 Rn�n; rðAÞ ¼ n; S # Rn�n: ð3:3Þ
Alternatively, we can transform system (3.1) into the left preconditioned system MAx = Mb, which is now related to the left S-Moore–Penrose inverse AyS;l.
A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522 519
In the next theorem, we apply the spectral property (2.13) to the special case of the optimal preconditioner AyS;r , i.e., thesolution to problem (3.3). We must assume that matrix AyS;r is nonsingular, in order to get a nonsingular coefficient matrixAAyS;r of the preconditioned system (3.2). Hence, all singular values of matrix AAyS;r will be greater than zero, and they will bedenoted as
r1 P r2 P � � �P rn > 0:
Theorem 3.1. Let A 2 Rn�nbe nonsingular and let S be a vector subspace of Rn�n such that the solution AyS;r to problem (3.3) isnonsingular. Then the smallest singular value rn of the preconditioned matrix AAyS;r lies in the interval (0,1]. Moreover, we have thefollowing lower and upper bounds on the distance dðIn;ASÞ
ð1� rnÞ2 6 AAyS;r � In
��� ���2
F6 n 1� r2
n
: ð3:4Þ
Proof. Using Lemma 2.2 for m = n, we get
nr2n 6
Xn
i¼1
r2i 6 n ) r2
n 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 theirarguments [17], i.e., for all A; B 2 Rn�n and for all i ¼ 1;2; . . . ;n
jriðAÞ � riðBÞj 6 kA� BkF ) 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;r � In
��� ���2
F¼ n� tr AAyS;r
� �¼ n� AAyS;r
��� ���2
F¼ n�
Xn
i¼1
r2i 6 n 1� r2
n
: �
Remark 3.1. Eq. (3.4) states that AAyS;r � I��� ���
Fdecreases to 0 at the same time as the smallest singular value rn of the precon-
ditioned matrix AAyS;r increases to 1. That is, the closeness of AAyS;r to the identity is determined by the closeness of rn to theunity. In the extremal case, rn ¼ 1 iff AyS;r ¼ A�1 iff A�1 2 S. A more complete version of Theorem 3.1, involving not only thesmallest singular value, but also the smallest eigenvalue’s modulus, of the preconditioned matrix, can be found in [18]. Theproof 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 M to be used, is basedon the spectral properties of the preconditioned matrix AM (e.g., condition number, clustering of eigenvalues and singularvalues, departure from normality, etc.). For this reason, the spectral properties of matrix AAyS;r , stated in Lemmas 2.1 and2.2 and in Theorem 3.1, are especially useful when the left S-Moore–Penrose 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 thesepreconditioners.
3.2. Applications to the constrained least squares problem
The S-Moore–Penrose inverse can be applied to the linear least squares problem subject to a set of homogeneous linearconstraints. This is in accordance with the well-known fact that the Moore–Penrose inverse can be applied to the linear leastsquares problem without restrictions [19], the major application of the Moore–Penrose inverse [20]. More precisely, let usrecall that, regarding the linear system Ax ¼ b; A 2 Rm�n, the unique solution (if A has full column rank) or the minimumnorm solution (otherwise) of the least squares problem
minx2RnkAx� bk2
is given by
�x ¼ Ayb:
Here we consider as well the linear system Ax ¼ b; A 2 Rm�n, but now subject to the condition x 2 T , where T is a given vectorsubspace of Rn. Denoting by Bx = 0 the system of Cartesian equations of subspace T, this problem is usually formulated as
minx2RnkAx� bk2 subject to Bx ¼ 0 ð3:5Þ
and we refer the reader to [11] and to the references contained therein, for more details about the linearly constrained linearleast squares problem.
520 A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522
Next theorem provides us with a new, different approach – in terms of the S-Moore–Penrose 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 = bconsistent but Ax ¼ b; x 2 T inconsistent; Ax = b consistent and underdetermined while Ax ¼ b; x 2 T consistent and un-iquely determined; both Ax = b and Ax ¼ b; x 2 T inconsistent, etc.
Theorem 3.2. Let A 2 Rm�n and b 2 Rm. Let T be a linear subspace of Rn of dimension p. Then the minimum norm solution to theconstrained least squares problem
minx2TkAx� bk2 ð3:6Þ
is given by
�xT ¼ AySP ;rb; ð3:7Þ
where P is the n� p matrix whose columns are the vectors of an arbitrary basis of T and SP ¼ PRp�m # Rn�m. In particular, if A hasfull column rank then �xT is the unique least squares solution to problem (3.6).
Proof. Let B ¼ fv1; . . . ;vpg be a given orthonormal basis of T and let P be the n� p matrix whose columns are v1; . . . ;vp. For aclearer exposition, we subdivide the proof into the following three parts:
(i) Writing any vector x 2 T as x ¼ Pa ða 2 RpÞ, we get
kA~x� bk2 ¼minx2TkAx� bk2 ¼min
a2RpkðAPÞa� bk2 ¼ kðAPÞ~a� bk2;
that is, ~a 2 Rp is a solution to the unconstrained least squares problem
mina2RpkðAPÞa� bk2 ð3:8Þ
if and only if ~x ¼ P~a 2 T is a solution to the constrained least squares problem (3.6).Now, let ~a be any solution to problem (3.8). Since �a ¼ ðAPÞyb is the minimum norm solution to problem (3.8), we have
kðAPÞybk2 6 k~ak2
and then, using the fact that P has orthonormal columns ðPT P ¼ IpÞ, we have
kPak22 ¼ ðPaÞTðPaÞ ¼ aTa ¼ kak2
2 for all a 2 Rp
and then
kPðAPÞybk2 6 kP~ak2 for any solution ~a to problem ð3:8Þ; i:e:;
kPðAPÞybk2 6 k~xk2 for any solution ~x to problem ð3:6Þ:
This proves that the minimum norm solution �xT to problem (3.6) can be expressed as
�xT ¼ PðAPÞyb for any orthonormal basis B of T: ð3:9Þ
(ii) Writing any matrix M 2 SP ¼ PRp�m as M ¼ PN ðN 2 Rp�mÞ, we get
kA eM � ImkF ¼minM2SP
kAM � ImkF ¼ minN2Rp�m
kðAPÞN � ImkF ¼ kðAPÞeN � ImkF ;
that is, eN 2 Rp�m is a solution to the unconstrained matrix minimization problem
minN2Rp�m
kðAPÞN � ImkF ð3:10Þ
if and only if eM ¼ P eN 2 SP is a solution to the constrained matrix minimization problem
minM2SP¼PRp�m
kAM � ImkF : ð3:11Þ
Now, let eN be any solution to problem (3.10). Since ðAPÞy is the minimum Frobenius norm solution to problem (3.10), wehave
y
kðAPÞ kF 6 keNkFand then, using again the fact that P has orthonormal columns ðPT P ¼ IpÞ, we have
kPNk2F ¼ trððPNÞTðPNÞÞ ¼ trðNT NÞ ¼ kNk2
F for all N 2 Rp�m
and then
kPðAPÞykF 6 kP eNkF for any solution eN to problem ð3:10Þ; i:e:;
kPðAPÞykF 6 k eMkF for any solution eM to problem ð3:11Þ:
A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522 521
This proves that the minimum Frobenius norm solution AySP ;rto problem (3.11) can be expressed as
AySP ;r¼ PðAPÞy for any orthonormal basis B of T: ð3:12Þ
(iii) Using Eqs. (3.9) and (3.12), we get
�xT ¼ AySP ;rb for any orthonormal basis B of T: ð3:13Þ
To conclude the derivation of formula (3.7), we must proof that, although the assumption that B is an orthonormal basisof T is (in general) essential for Eqs. (3.9) and (3.12) to be hold, formula (3.13) for �xT is valid not only for orthonormal bases,but in fact for any basis (not necessarily orthonormal) of subspace T. In other words, the vector AySP ;r
b is independent of thechosen basis. Indeed, let B, B0 be two arbitrary bases of T and let P ,P0 be the n� p matrices whose columns are the vectors of Band B0, respectively. Writing matrix P0 as P0 ¼ PC, C being a p� p nonsingular matrix, we have
SP0 ¼ P0Rp�m ¼ PðCRp�mÞ ¼ PRp�m ¼ SP ) AySP ;rb ¼ AySP0 ;r
b:
Finally, if in particular A has full column rank then A is left invertible and, due to the uniqueness of the orthogonal pro-jection A�xT of vector b onto the subspace AT # Rm, we derive that the minimum norm solution �xT ¼ AySP ;r
b to problem (3.6) isindeed 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 �xT
of problem (3.6), is valid for any basis (not necessarily orthonormal) of T. Indeed, let B, B0 be two arbitrary bases of T and let P,P0 be the n� p matrices whose columns are the vectors of B and B0, respectively. Writing matrix P0 as P0 ¼ PC, C being a p� pnonsingular matrix, we have
P0ðAP0Þy ¼ PCððAPÞCÞy ¼ PCCyðAPÞy ¼ PCC�1ðAPÞy ¼ PðAPÞy;
where the reverse order law for the product (AP)C holds because AP has full column rank, i.e.,
p ¼ rðAÞ þ rðPÞ � n 6 rðAPÞ 6 minðrðAÞ; rðPÞÞ ¼ p
and C has full row rank; see, e.g., [3]. In conclusion, the minimum norm solution to problem (3.6) is given by
�xT ¼ PðAPÞybfor any orthonormal basis of T if rðAÞ < n;for any basis of T if rðAÞ ¼ n
�
and also, for both cases ðrðAÞ 6 nÞ, by �xT ¼ AySP ;rb for any basis of T.
The following simple example illustrates Theorem 3.2.
Example 3.1. Let A ¼ ð1 1 1 Þ 2 R1�3 and b ¼ ð2Þ 2 R1�1. Consider the subspace T � x3 ¼ 0 of R3. Then m ¼ 1; n ¼ 3;p ¼ 2 and taking the (nonorthonormal) basis of T : B ¼ fð1; 0;0Þ; ð1;1;0Þg, we get
P ¼1 10 10 0
0B@1CA ) SP ¼ PR2�1 ¼
a
b
0
0B@1CA
8><>:9>=>;
a;b2R
� R3�1 ) AySP ;r¼
1=21=2
0
0B@1CA
and then the minimum norm solution to problem (3.6) is given by
�xT ¼ AySP ;rb ¼
110
0B@1CA:
4. Concluding remarks
The left and right S-Moore–Penrose inverses, AyS;l and AyS;r , extend the idea of the standard inverse from nonsingular squarematrices not only to rectangular and singular square matrices (as the Moore–Penrose inverse does), but also to invertiblematrices, themselves, by considering a subspace S of Cn�m not containing A�1. Restricted to the right case, the two expres-sions given for AyS;r (namely, the first one based on an SVD of A; the second one involving the orthogonal complement of sub-space S), respectively generalize the well-known corresponding formulae (1.4) and (1.5)(ii) for the Moore–Penrose inverse.The S-Moore–Penrose inverse AyS;r can be seen as an useful alternative to Ay: it provides us with a computationally feasibleapproximate inverse preconditioner for large linear systems (in contrast to the unfeasible computation of Ay ¼ A�1), and italso provides the minimum norm solution AySP ;r
b 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-Moore–Penrose inverse with respect to both theoretical results and other applications, already known for the Moore–Pen-rose inverse, can be considered for future researches.
522 A. Suárez, L. González / Applied Mathematics and Computation 216 (2010) 514–522
Acknowledgments
This work was partially supported by Ministerio de Ciencia e Innovación (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) 394–395.[2] R. Penrose, A generalized inverse for matrices, Proc. Cambridge Philos. Soc. 51 (1955) 406–413.[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 Moore–Penrose inverse of a columnwise partitioned matrix, Linear Algebra Appl. 421 (2007)
16–23.[5] T. Britz, D.D. Olesky, P. van den Driessche, The Moore–Penrose inverse of matrices with an acyclic bipartite graph, Linear Algebra Appl. 390 (2004) 47–
60.[6] M.A. Rakha, On the Moore–Penrose generalized inverse matrix, Appl. Math. Comput. 158 (2004) 185–200.[7] Y. Tian, Using rank formulas to characterize equalities for Moore–Penrose inverses of matrix products, Appl. Math. Comput. 147 (2004) 581–600.[8] Y. Tian, Approximation and continuity of Moore–Penrose 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 Moore–Penrose 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 Moore–Penrose inverse, Appl. Math. Comput. 183 (2006) 522–532.[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 Dennis–Schnabel theorem, BIT 22 (1982) 252–256.[13] G. Montero, L. González, E. Flórez, M.D. García, A. Suárez, Approximate inverse computation using Frobenius inner product, Numer. Linear Algebra
Appl. 9 (2002) 239–247.[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. González, Orthogonal projections of the identity: spectral analysis and applications to approximate inverse preconditioning, SIAM Rev. 48 (2006)
66–75.[19] R. Penrose, On best approximate solutions of linear matrix equations, Proc. Cambridge Philos. Soc. 52 (1956) 17–19.[20] A. Ben-Israel, The Moore of the Moore–Penrose inverse, Electron. J. Linear Algebra 9 (2002) 150–157.