Professional Documents
Culture Documents
ELEMENTS OF
THE THEORY
OF GENERALIZED
INVERSES FOR
MATRICES
Randall E. Cline
Mathematics Department
University of Tennessee
Knoxville, Tennessee 37916
ISBN-13: 978-0-8176-3013-3
DOl: 10.1007/978-1-4684-6717-8
e-ISBN-13: 978-1-4684-6717-8
TABLE OF CONTENTS
1.
2.
Preface .
INTRODUCTION
1.1
Preliminary Remarks
1.2
1.3
2
4
3.
4.
5.
9
9
12
21
27
29
3.1
3.2
3.3
29
40
56
Basic Properties of A+ . .
Applications with Matrices of Special Structure
Miscellaneous Exercises
DRAZIN INVERSES
4.1 The Drazin Inverse of a Square Matrix
4.2 An Extension to Rectangular Matrices
4.3 Expressions Relating Ad and A+
4.4 Miscellaneous Exercises
64
71
5.1
71
79
85
87
-iii-
57
57
68
70
Preface
have appeared in a number of excellent graduate level textbooks since 1971, material for use at the undergraduate
level remains fragmented.
per week was for a lecture on the subject matter, and another
meeting was to permit students to present solutions to
exercises.
-v-
At the suggestion of
The
-vi-
1
Introduction
1.1
Preliminary Remarks
The material in this monograph requires a knowledge of
Fundamental definitions
Therefore, it may
It is recom-
Although perhaps
-1-
1.2
are
The special
and the m by n null matrix, respectively, with sizes determined by the context when no subscripts are used.
If it is
L x.y ..
i=l
I Ixl I
I IxI I
= + v'{X;XT = +
IIi=l Ix. I
1
2.
Although
b.
1.2
AB
(z ,w )
p m
and also as
AB
L x.y.
i=l
b.
c.
o.
BAH = CAH.
*1.3
Let A be any normal matrix with eigenvalues Al, ... ,A n and orthonormal eigenvectors xl' ... ,x n .
- 3-
a.
b.
If E.
L A.x.x.
i=l I I I
c.
1.3
Loosely
For this
(1. l)
AkXA
(1. 2)
XAX
(1. 3)
(AX}H
AX,
(1. 4)
(XA}H
XA,
(1. 5)
AX
XA.
X,
(It should be noted that (l.l) with k > 1 and (1.5) implicitly
assume A and X are square matrices, whereas (l.l) with k
(1.2),
1,
(1.3), and (1.4) require only that X has the size of AH.
-4-
For any A, let N(A) denote the null space of A, that is,
N(A)
a.
= {zlAz = OJ.
A is nonsingular,
(ii)
(i i i)
Rank (A)
(iv)
(v)
Ax
=b
n,
side b.
b.
1.5
Let
2
2
2
-/
-01
o
o
1
-0 _41]
0-1
-5-
-1
a.
b.
Show that X = A4
-1
2' i f i = j - 1 ,
a .. = {
I J
1, otherwi se,
form AS- l
More generally, given An of this form for any
n ~ 2, what is A -1 7
n
c.
2, that is,
Idet An I = 1.
d.
1.6
Let x be any vector with I Ixl I = 1 and let k be any real number.
Show that A
b.
r'
+ kxx
a.
if i
= j,
= I, otherwise,
construct A-I.
c.
d.
*e.
-2.
= -1.
Show that A has one eigenvalue equal to l+k and all other
eigenvalues equal to unity.
(Hint:
y orthogonal to x.)
*f.
1.7
a.
A
= [1
-1
_22J,
X. "10[:
-6-
-'o] ;
-2
b.
A:
c.
1.8
A:
[~
-1 \ '
[~
:J
21
2
X : 1/51
~615
lit
-6
J
X : 1/50G
-7
3
!] .
A:
r-,'
l~
-5
-6
-It
2
[-'
, X 1/2 ;
-5
-It
It
-7-
!] ;
2
Systems of Equations
and the
Moore -Penrose Inverse
of a Matrix
2 .1
Now by definition,
= {z IAz = O}
= n.
to construct ALb.
However, when r
<
xl + az is also a solu-
(Z .1)
2 a. z.
i=l
x,
Exe rc i ses
2. I
1-\
L4
13
(i i i)
A
3
(v)
2.2
l~
A, - [;
-3
3
-2
0
7] ,
bl
r-']
(; ; )
A2
l-64 ;
:1 b,' r: 1'
3J
-:1'
I
3
['
(;,)
A4
3
I
b,{J
(v i)
A6
[2
I
3
2 0
_:J b[:j,
4'
2.4
2.5
= b6
are written as
where
xl
IT [0
-1
2) ,
for every solution such that IIxl1 2 = 25124.
then
2.6
Show that A, AHA and AAH have the same rank for every matrix A.
2.7
a.
2.2
(AHA)-IAHb.
Suppose m = n?
=
b if
1, (1.2),
AX, (XA)H
For if
XA,
Now if A has full row rank, then with X any right inverse of A,
A, XA
C2.2) hold.
column rank.
Now X
AHCAAH)-l is a right
inverse of A, and
CXA)H
(AHCAAH)-lA)H
AHCAAH)-lA
XA.
Such
C2.3)
[~
[~
and so
-1]
2
-13 -
1:.[
2 -1]
3 -1 2
Example 2.2
If Y is the column vector
then
+
IS
[1
2-i
3iJ.
that AXA = A, it follows at once that the system is consistent if and only if
(2.4)
AXb = b.
Conversely,
Suppose now
Xb + Zy
with Z any matrix with n-m columns which form a basis of N(A)
and y an arbitrary vector.
-14-
Proof:
With X = A+ in (2.5),
(x,x) = (A+b+Zy,A+b+Zy)
(A+b,A+b) + (Zy,Zy)
IIA+bI1 2 + IIZyl12
since
o.
(A+b,Zy)
Thus
o.
J,
[_54
then
122
= -3-'
It was noted in Section 2.1 that the least squares solution of an inconsistent system of equations Ax = b when A has
full column rank is x = (AHA)-lAHb. From (2.3) we have,
therefore, that x = A+b is the least squares solution in this
case. Although this result can be established by use of
calculus (Exercise 2.7), the following derivation in terms of
norms is more direct.
THEOREM 3: For any system of equations Ax = b where A has
full column rank, x = A+b is the unique vector with I Ib-Axl 12
minimal.
If A is square or if m > n and Ax = b is consistent,
then with A+
(AHA) -1 AH
a left.Inverse of A and AA + b = b, the
vector x = A+b is the unique solution with Ilb-Axl1 2 = O. On
the other hand, if m > n and Ax = b is inconsistent,
Proof:
lib-Axil
II (I-AA )b-A(x-A b) II
then
1[3
IT
-4
3
7
-2]
-1 '
AA+b
and
b with I Ib - Ax I I 2
144
11
linearly independent.
indices for which ajl' ... ,ajr are linearly independent, and
let E be the m by r matrix
E =
[a.
)1
, ... ,a. 1.
)r
Suppose r < n.
))
we then have
I in each case.
Example 2.5
Let
-5
-1
6
-15
-1
-1
1 -1
Then
-17-
[4 6]
1
-1
-5 -15
4/5
-1/5
3/5
7/5
-3
-2/5
- 3/5
-7]
3
Using full rank factorization, the existence of the MoorePenrose inverse of any matrix follows at once.
The following
A, XAX
AXA
(AX)H
= AX, (XA)H = XA
A+.
If A = 0
is the m by n null
mn
If A is not the null matrix, then for any
nm
full rank factorization EFG of A, A+ = G+F-IE+
matrix, A+
Proof:
after (2.2).
If A = 0 ,then XAX = X implies X = A+ = 0
If A is
nm
not the null matrix, then for any full rank factorization
Thus E+ = (EHE)-lE H
Moreover, AXA
= A and
XAX = X, so that X = A+ .
It should be noted that although the existence of a full
rank factorization A
Such a procedure
tion, and if y
solution to Ey
b, by Theorem 3.
-18-
b with I Ixl
12
A+b is the
minimal.
Show that xl
2.10
[I ,u]+ --
2.11
a.
Given any real numbers bl, ... ,b n , show that all solutions to
the equations
1, .. ,n,
can be written as
X.
b. - _1_
L b.
n+l i=l
Cl,
1, ... ,n,
and
1
= -1
L b.I
n+ i=l
where
b.
Cl
Cl,
is arbitrary.
L x.
i=l
Cl
01
-19-
c.
= -
L b ..
ni=l
d.
2.12
Given any real numbers b l , ... ,b n , show that the mean is the least
squares solution to the equations
= 1, ... ,no
2.13
uv H a matrix of rank
Ax = b,
A+b = x -
n-r
L a.z.
i=l
with a. = (x,z.),
I
2.15
1, ... , n-r.
(Continuation):
2.16
Use the results of Exercises 2.4 and 2.15 to construct A6+ starting
with the right inverse
X
2 -3]2
[-10
0
o
0
-20-
2.3
. . J ' - - - - - - f - - - - - - xl
Figure I.
x.
~+------ Xl
H+
that the solution of the form x = a l b l is the point on
Pl(bl) with minimal distance from the origin. Also, since
the vector ~ is perpendicular to the planes Pl(O),Pl(b l ),
II II is the distance between PI (0) and PI (b l ). The representation of any solution x as x = alH+b l + alz, corresponding
to (2.1) in this case, is illustrated in Figure 3.
A
x3
~----r------- Xl
Figure 3.
The representation x
-22-
al
H+
b l + a1z.
-23-
Y = AA b have components YI'Y2'Y3' and let a l ,a 2 ,a 3 desIgnate rows of A. Now if Pi(Yi) is the plane of all solutions
of aiHx = Yi' i = 1,2,3, then the set of all solutions of
Ax = y is the line 12 = PI (YI) n P 2 (Y2) as shown in Figure Sa,
A+y = A+b is the point on 12 of minimal norm and Ily-bl1 2 is
minimal, as shown in Figure Sb.
-24-
~~------------_ bl
(a)
Figure 5.
(b)
(a) Solutions of Ax
=y
where y
= AA+b.
> 0
Exerc ises
2.17
Pi(b i ) = {xi a i x = b i , a i
= (ail,aiZ,ai3)}'
then
z.18
Given any points (xi'Yi)' i = O,l, ... ,n, in the (x,y) plane with
xO, ... ,x n distinct, it is well known that there is a unique
interpolating polynomial Pn(x) of degree 2 n (that is, Pn(x i ) = Yi
fo raIl i = 0, ... , n), and i f
then aO, ... ,a n can be determined by solving the system of equations Aa = y where
, a
, y
-26-
For k
polynomial
which minimizes
n
L [yo
i=O
a.
- Pk(x.)]
b.
c.
L y.
i=O
then SSR
-5
-4
-3
10
What
2.4
Miscellaneous Exercises
2.19
X, AX
ZI
a.
X.
c.
AH.
AH, then
X and XAA H
The matrices AHA, (AHA)2, (AHA)3, ... , are not all linearly
independent, so that there exists scalars d l , ... ,d k not all
zero, for which
H
H 2
H
dlA A + d2 (A A) + ... + dk(A A) = O.
(Note that if A has n columns, k
d.
n +1.
Why?)
2.21
Show that if Ax
(Continuation):
Why?)
=B
is a consistent
where Y is arbitrary.
-28-
3
More on
Moore -Penrose Inverses
3.1
Basic Properties of A+
(a)
A m by n implies A+ n by m;
(b)
(c)
A++
A',
(d)
AH+
A+H.,
(e)
A+
(f)
(AHA)+
A+
mn implies
(AHA)+AH
nm'
AH(AAH)+;
= A+AH+;
-29-
Then
(g)
+ = f/a, i f a 'f 0,
0, if a = o,
(h)
(i)
If A
L A.
i=l
o whenever
where A. A.
(UAV) +
vHA+U H;
i 'f j, A+ =
L A.
i=l
(j)
If A is normal, A+A
(k)
A, A+, A+A and AA+ all have rank equal to trace (A+A).
AA+;
Proof:
AH+AHA = A.
(AHA)+AHA(AHA)+
(AHA)+.
A.HA.
J
0 since
A.+A.+HA.HA.
O.
(AHA)+AHA
(AAH)+AAH
-30-
= AH+AH =
(AA+)H
AA+.
To show that A, A+, A+A and AA+ all have the same rank,
we can apply the fact that the rank of a product of matrices
never exceeds the rank of any factor to the equations
AA+A = A and A+AA+ = A+. Then rank (A) = trace (A+A) holds
since rank (E) = trace (E) for any idempotent matrix E [7) .
Observe in Lemma See) that these expressions for A+
reduce to the expressions in (2.3) whenever A has full row
rank or full column rank. Moreover, observe that the relationship (EG)+ = G+E+ which holds for full rank factorizations
EG, by Theorem 4, also holds for AHA where A is any matrix,
by Lemma S(f). The following example shows, however, that
the relation (BA)+ = A+B+ need not hold for arbitrary matrices
A and B.
Example 3.1
Let
[~
1-1].
1 -1
Then
[~
BA
~J
(BA) + ,
(AHA)-lAH
and
B+
1/2 [ 2
-2
-!] [~
ll'[: :]
BH(BBH)-l
-1 -1
so that
A+B +
r-11
-1]
1
Also, we have
1
1
~]
[ 2 - 2]
-2
1/2 [ 2
-2
1/2
~]
[2o -']1 ,
o -1
(BA)+.
Then
AQ
[a"
Jl
, ... , a"
In
1.
" "
H .. wmH deS1.gnate
"
In a s1.m1.1ar
manner, 1."f wI"
t h e rows 0 fA ,
and if P is the permutation matrix obtained by permuting
rows of 1m in any order {i l , .. ,i m}, then
H
PA
that if
or colmanner,
Moreover,
Go
1
B =
PAQ
[~ ~]
Go
1
~]
i[ ~ -~].
-1
[~
o
1
:1[: : :]
We
(I
[Ak_l,a k 1 , let
+
Ak-lA k - l )a k
and let
Yk
H
H+
+
a k Ak - l Ak - l a k
Then
(3.1)
A +
k
where
bk
r;'
(l+Yk)
if
-1
ck f
0,
H
H+
+
a k Ak _ l Ak - l '
-33-
if c k
O.
Proof:
0 are exhaustive.
and c k
Then to
bka k
Continuing, using (3.2) gives
(3.4)
and
(30 5)
since
Suppose now that c k
0 and b k
c k . Then
thus c k a k
1, then
XA
in (3.3) is Hermitian.
:]
Moreover,
-34-
in (3.4), and
x
in (3.5).
f O.
in (3.2) is Hermitian.
-1
H
H+
+
a k Ak - l
Ak - l
Then
we have also
in (3.3) Hermitian.
since c k
= 0 implies Ak-1A k - l a k = a k ,
1, ... ,n,
[~~ -~l
-1
-3
1/5[2
-1],
and
Hence
1/29[3
10
6]
and so
[1/5 [2
-1]
1/29[3
1/145 [
55
-10
15
50
1/145 [3
10
10
6]
11
- 35]
= 1/29 [ 3
30
Continuing,
1/29 [ 29]
-58
and
o.
Thus, with Y3
5 and
-36-
-2
10
H H+ +
a 3 A2 A 2
+
H +
(A 2 a 3 ) A2
1/29[5
-22
-19] ,
we have
b3
1/174[5
-22
-19]
so that
A3
1/29
[113
-2
10
:]
1/174[5
[61
1/174 2:
-22
5
-10
-22
44
-19]
38
-19]
-23]
-2 .
10
16
-22
1/174[
-19
A,
with A+ known.
A and
3. I
Let A be any matrix with columns a l , ... ,a n , and let A+ have rows
H
H
wI , . ,w n Prove that if K denotes any subset of the indices
I , ... ,n such that a i = 0, then wi H = 0 for all iEK.
3.2
c.
Construct A+.
-37-
3.3
(Continuation):
A=
trapezoidal.
trapezoidal decomposition of
b.
A=
EG is called a
A. )
A,
obtained from
2
-1
Hint:
A
[;
Start with
0
,.
...
:] [:
-1
* *
0
*
:]
"
Compute A+.
[B,Rl be any mby n upper trapezoidal matrix wi th n
m+ 2
3.5
3.6
Show that the two forms for [Ak_I,a k ) + in Theorem 6 can be written
in a single expression as
[Ak_I,a k)
[Ak-I-Adkk-H/akdkHj
where
d H
k
3.7
Let [Ak_I,a k)
be partitioned as
[Ak_I,a k)+ =
['H]
d H
k
wi th d kH a row vector
'1
f d k Ha .,.
..J. I ,
k
1.
b.
Construct A+ if
0
A [:
Hint:
3.8
*3.9
:].
-I
A = L.:
I I
r- xXi H
1
-39-
c.
If \
Xl
~l
=...!..[
/2 -1
x2
1/3[~1
2
x3
=_1
312
[-~l
1
3.2
i = 1.2.3.
special structure is such that we can obtain information concerning the set of all solutions.
i = l, ... ,n,
In this section
Q is the ms by nt matrix, P
-40-
XQ,
of the form
2'
[: oJ,
[~
2i
-1
-3
3
-:]
then
pXQ
[~
[~
4
i
12
3i
4
0
4
12
i
3i
-lJ
and
Q xP
-1
3
-3
2i
-;1
[P X Q]H
H
(qn P )
pH X QH.
-41-
[PXQ1[RXS1
L q oS olPR
j=l mJ J
L q oSotPR
j=l mJ J
Therefore,
[P X Q1 [R X Sl = PR X QS.
(3. 7)
ExamEle 3.5
To construct A+ i f
[1
-1
-3
12
;]
4 '
-4
"3
-~J,
Q =
[~ ~] .
-42-
Then we have
Q-l
Q+
-~] ,
-l/Z [ 4
-3
1:]
ir [22-9
p+
17
-8
so that
p+
XQ+
88
16
-44
-8
-36
60
18
-30
68
1
lZ2 -66
-3Z
-34
16
-12
ZZ
27
-45
-9
15
51
24
17
-8
ExamEle 3.6
To construct A+ if
A =
["3
u3
13
u3
13
0
u
13
0
II
["3
u3
13
u3
13
13
u3
XU Z '
I'J
Next observe
XI 2
Therefore
A can be wyitten as
and thus
(3.9)
Permuting rows of [I,u)+ in Exercise 2.10 now gives
[u i ' Ii)
so that
1
i+l
(3.9) becomes
(3.10)
The additional
integer p let G ;
p
(pI +WIlW)-l.
n
-44-
Then
Proof:
II
[G p Xu,
p w X I p - Gp W Xu p u p J.
(3.11)
then
(3.12)
Also, observe that with (pI +WHW)W+ = pW+ + WH, the relation
n
(3.13)
is Hermitian, implies
II'G IV
P
is Hermitian.
Let
l
rI
XU H1
:XIP ,
p
and let
+
X= [G p XU p ,W XI p -GW
Xuu
p
p p J.
Then it follows from (3.12) and (3.7) that
XA=G Xuu
P P
+WWXI
-GWWXuu
G (I - W+W) Xu u H + W+W X I
-(1
is Hermitian.
P p
-W W) Xu u
+ W WX I
p,
P P
Xu H
A.
AXA
-W (I - W W) Xu u
P
n
P p
+ WW W X Ip
Continuing, we have
XA(G
1
+
+
-(I -W W)G Xpu +W WG XU = GpXu p '
P
p n
p
p
p
Xu)
p
leI
XA(W+XI)
p
-w+W)w+Xu u H+W+WW+XI
P p
W+XI
p'
and
1
--0
- W+W) Gp W+
p n
XA (G W+ Xu u H)
P P
lienee, XAX = X.
AX
WG
Finally,
pG
H +
+
H
X pu u
+W WG W Xu u
p p
p
p p
forming AX gives
W+ X u p H - Gp W+ X p u PH
: u
WW+ X I
- WG W+ Xu u H
P P -
Now
W+ Xu H - G W+ X pu H
G WH Xu H
p'
implies
A+ which establishes
(3.11) .
ExamEle 3.7
lI
xu 3
W X I3
li
-46-
and
['n XX 3
U
W 13
G3
W+-G W+
3
+
-1,3 W
G3
-G W+
3
W+-G W+
3
-G 3W+
G3
-G 11'+
3
-G 11'+
3
W+-G 3W+J
~G3W'1
I'
T = T(p,W) =
-I Xu Hj
[n p .
WXI
x =
with x(l~ ... ,x(p) and b(O)n-tuples and b(l), ... ,b(P) m-tuples,
then x is a solution if and only if
(3.14)
Ix (j)
j=l
b(O)
and
(3.15)
Wx(j) = b(j) ,
j = 1, ... ,p.
3.11
Neglect-
ing other factors which may effect yield, a model for this
experiment has the form
(3.16)
m + t. + e ..
I
where y .. is the yield of the jth plot to which ferti lizer i has
IJ
i =1 j=l
a.
e ..
I
and
where x
y =
Ax + e,
and
-48-
b.
c.
x=
It.o.
i=1
x from
Starting wi th
(Continuation):
Ij
Ij
I j
j=1 , ... ,q, designate the effect of the interaction between treatment i and block j.
a.
Y = Alx + e,
where now
and
AI =
b.
Show that the data for the model in (3.19) can be represented
as
where
x = (m, t I ' ... , t P , b I ' ... ,bq , ( t b) II ' ... , ( t b) I q , ... , ( t b) pi' ... , ( t b) pq)
- 4 9-
c.
x = A/Y.
and thus
by assuming
p
L t.
i=l
q
=
L b.
j=l J
and also
L(tb) ..
i=l
IJ
L (tb) ..
j=l
L t.
i=l
*3.13
L b.
j=l J
O. )
3.14
a.
'~b.
3.15
w, !-T+r
= (I -W+W)
X(I p -u p
+).
I-n.
Prove directly that for any mbyn matrix W of rank r and any
p,
a.
-50-
Xu
(3.20)
and
(3.21 )
b.
WX = B.
(Continuation):
YJ'
-T+T, where
I
J
np
has components Y'I'"'Y' , let Z.. denote the matrix
Jp
JP
1 if and only if
(3.22)
Z .. U
I
J P
and
(3.23)
3.18
WZ = O.
IJ
Consider
L a.
i=l
L boO
j=l J
j, i=l , ... ,n and j=l , ... ,p, then how should the shipments be allocated in order to minimize total transportation cost?
This problem
The problem now is to determine the x .. , i=l, ... ,n and j=l, ... ,p
to minimize the total shipping cost.
-51-
0.
bJ
0.
~x ..
x ll
x lj
xi 1
nl
bl
Figure 6.
al
x lp
a.
X.
IP
nj
b.
J
a
np
La.
= Lb.
j J
(3.24)
L L c . .x . . ,
i=l j=l
IJ
IJ
L x ..
j=l
aj
1, .. . ,n,
and
n
(3.26)
L x ..
i =1
b.,
J
1, . ,p.
Also, we must have x .. > 0, for all i and j, and, assuming fracIJ -
integers.
a.
t6)--
become
Prove that if T
[[r n -
) and
T(p,u n ), then
_1
n+p
u u
n n
HJ Xu p+ ,
+X
[1 p - _1
u u HJ]
n+p p p
.
x = T+b,
and column
then
1
x .. = -a.
IJ
P I
L a.
np i=l
*c.
n '
only if all row sums and column sums in the tableau form are
zero.
d.
The vector g = (1
inner product
(c,x)
np
L L c . . x ..
i=l j=l IJ I J
in (3.24) .
L
i=l
Llc ij
j=
(Continuation):
Suppose that we
A "three-dimensional" transportation
L a k
i=l I
L b k ,
j=l j
Jla ik
q
L b k
k=l j
L d .. ,
j=l Ij
n
L d .. ,
i=l Ij
Show that the conditions which the x ijk must satisfy if there
is a solution can be written as Tx = b where T = T{q,W) with
-54-
W; T
c.
Show that
- _I
u u HJ XI]
[1-[1
qp
p+qpp
n
and that G T
q np
u;
+; [u,Vl where
I (I p -
n {n+q}
-55-
and
v
3.3
3.20
= u
Xn(p+q)
1 [1
n
n + 2p + q u u HJ
(n+p) (n+p+q) n n .
Miscellaneous Exercises
Prove that a necessary and sufficient condition that the equations
AX
= C,
XS
= CS,
a particular solution.
3.21
Prove Lemma 7.
3.22
Prove that
-56-
4
Drazin Inverses
4.1
(4. 2)
(4.3)
Observe first that if A = 0 is the null matrix, then
A and X = 0 satisfy (4.1), (4.2) and (4.3).
Proof:
i=l
d.A 1
0,
-57-
d.A
dk1 [I
i=k+l ~
i - k - l ).
Also, multi-
... ,
and thus
(4.5)
for all m > 1.
Let X = AkU k +l .
and
X~+lAk
X3AZ
xk+lAk+ly
XAk+lyk+l = Akyk+l
XAY
AYZ
yZA = Y
to establish uniqueness .
-58-
D~azin
Also, we
will call the smallest k such that (4.1) holds the index of
A.
That Ad is a generalized inverse of A is apparent by
noting that (4.1) holds with k
also (4.2) and (4.3) hold.
BC, Ad
2
B(CB)d C.
Proof:
Then
-59-
The
same argument now applies to CZB Z ' that is, either CZB Z is
nonsingular and
and thus
(4.7)
Ad
in Lemma 10.
A2
and
B1ClBlC l
BlBZCZC l , ... , Am
-60-
= BlB Z
... BmCmC m_ l
... Cl
(4.8)
m-l
m-l
BmCm
and
+
m
CmBm.
m
m+l
Therefore, in both cases we have rank (A ) = rank (A
).
Furthermore, it follows in both cases that (4.1) holds for
k = m and does not hold for any k < m. That is to say, k in
.
k
k+l
(4.1) is the smallest positive Integer such that A and A
have the same rank.
Example 4.1
If A is the singular matrix
[:
~l
-1
BlC l
[:
1" [~
2 -2]
1 '
then
ClB l
[:
:]
-61-
BleClB1)
Ad
-Z
0'.[:
Cl
-Z6
35
3
"]
-33
-1
ExamEle 4. Z
I f A is the matrix
li
with A
11
l:
BIC l
where
1
1
0
o
1
-1
][:
B3C3
[~] [1
1
-1
1]
~] ,
[~
-1
~]
CZB Z
Continuing,
~]
so that
CZB2
OJ
7.
Hence A has
Ad
BIB2B3(C3B3)
-4
C3 C2 C1
1
1
24 01
343
. [1
0]
oJ
7"
II
10
~r
AAdA
A,
AdAAd
Ad'
AAd
AdA,
so that
(4.10)
by the duality in the roles of A and Ad.
Conversely, if
Consequently,
Thus X
= BC, A# = B(CB)-2C.
Exercises
4.1
Al
4.2
5
7
1] ,
A2
[:
2
0
0
-:J.
A{
3
-2
-ll
0
0
-2
-3
Given any matrices Band C of the same size where B has full
column rank, we will say that C is aZias to B if B+ = (CHB)+C H.
a.
b.
;j'
c.
Prove that Ad
= A+
= A+
when A is Hermitian.
4.4
Prove that Ad
4.5
=0
4.2
1 such
arrangements of the n columns of B (taken without any permutations) and the m-n columns of zeroes, and that Ad can be
different in each case.
Example 4.3
If
B
[: J
1
[:
and
:]
-1
AZ
"
[:
-64-
0
0
-1
z
A3
[:
-1
:]
then
:]
-2
1
-13
l]
10
3
}[l
[:
t[:
(A 2 ) d
0
0
0
-1
BC.
o
o
o
1
(BW)k
(4.12)
XWBWX
X,
(4.13)
BWX
XWB.
Proof: Let X
B(WB)d 2 . Then with XW = B(WB)d2W = (BW)d'
by Lemma 10, (4.11) holds with k the index of BW. Also,
and
XlWBWX l
BWXlWX l
(BW)Z(XlW)ZX l
XZWBW(BW)k(XlW)k Xl
then
~
Xz(WXZ)k+l(WB)k+lWBWXl
Xz(WXZ)k+lW(BW)k+lXlWB
(BW)d'
(BW)d
:J
I~
l3
WI
II
l3
-1
-:]
w =
2
[1
0
~] ,
then
(BW )d
1
_1_
(91) 2
[16960 169
338] ,
87
-169
(BW ) d
2
"6 [:
:]
Exercises
4.6
Verify that Band (BW)d satisfy the defining equations in Theorem
11 for W = Cl , C2 , C3 in Example 4.3 and for W = Wl ' W2 in
Example 4.4.
4.7
= W+ (WB) -1 .
4.8
4.9
(Note
that this result follows at once from Lemma 5(f) and Exercise
4.8 if B has full column rank, whereas Lemma 5(f) and Exercise
4.2d can be used for the general case.)
-67-
4.3
W
then
B+ = W(BW)dW.
Thus, using W-weighted Drazin inverses with W = BH,(BW)d and
B+ are related directly in terms of products of matrices
which implies that (BW)d and B+ have the same rank.
In con-
rank (A
k+l
),
There-
Proof:
> k.
In this case,
AXA = A ,
XAX = X
and
(4.17)
(XA)H
XA ,
AX = AA d
(AX)H
AX,
XA = AA
d
4.12
4.13
4.14
A2
B1B2C2Cl , ... ,
(Continuation):
Show that
-69-
and
4.15
2
2
Construct (A )L and (A )R for the matrix A in Example 4.1
4.16
4.4
Miscellaneous Exercises
4.17
XWB,
= WB(WB)/,
= B(WB)q[WB)rW)(B(WB)q)]d 2 .
AAd
= BB d
-70-
Bd ' then
5
Other Generalized Inverses
5.1
-71-
(5.1)
CYZB = I r ,
and XAX
X implies
(5.2)
ZBCY = Is.
YZBC
Y(Cy)-lC
(YC)#YC,
by Lemma 10.
#
A(X+UV)A = A.
Now
(5.4)
where the first matrix on the right-hand side has full column
rank (Exercise 5.6). Thus, for any choice of V such that
the second matrix on the right-hand side of (5.4) has full
row rank, (5.3) holds and rank (X+UV) > rank (A).
The following example illustrates the construction of
matrices, X, of prescribed rank such that A and X satisfy at
least one of the conditions AXA = X and XAX = X.
Example 5.1
Let A be the matrix
BC
2
1
-2]
1 .
=} [-1Z
BL
-1
is a left inverse of B.
Cy
c:],
:]
Now if yH
194[-1
(Cy)+ =
[3
9],
zH
5], then
k[l9
-11
0]
so that
-33
-44
-55
z of
[1 -1]
1
1
1.
-1
Then
CY [05 -3'
4] (Cy) -1 201[35
=
z
10
so that
-4
6
-4
-74-
where
Then
u~N(C).
-4
-4
~]
-3
A and XAX
X is
(S.Z)
In the following
= AX i
, i
= 1,2,3.
Example 5.2
Given the matrix A and full rank factorization A
in Example 5.1, again let yH = [3
5].
BC
Then
1
804[17
8]
and
yz
8~4
51
21
68
85
28
24]
32
35
40
1.
Y =
[1 -1]
1
-1
then
AY
BCY
10
[ 5
5
2]
5 , Z = (AY)+
1
-75-
1 [ 16
220 -20
35
and
Xz
[lB
lio - z
YZ
18
-15
ZO
-15
~]
Xz
[lB
uv H = lio -z
18
-13
17
-18
:]
=
-10.
= A+
if AXA
= A,
(AX)H
-76-
A, XAX
AX and (XA)H
X,
= XA.
5.3
= X and (AX)H
= Z+.
Let A
that X2AX 2
X2 and (X 2A) H
= BC
and X
= X2A.
column rank.
5.5
a.
b.
c.
YZBCB and
the special
5.6
5.7
5.8
-77 -
((AY)+AJ+(AY)+.
Appendix 1:
Hints for Certain Exercises
Chapter 1
Ex. 1.lb:
xy
yX H imp1 ies y
= ax
where
Ex. 1.3a:
Ex. 1.5b:
unity,
H
A -1
n
-U n_ 1
1.5c:
-u
n-l
-79-
preceding rows and expand the determinant using cofactors of the first
column.
1.5d:
H
Let X = I + dXX and determine d
Ex. 1 .6a:
A
1
20 H] where x = --u40.
= 6 [ I + TXX
1.6f:
1.6e:
/40
50
Ax
that AX
I.
1.6b:
= (l+k)x and Ay = y.
-1
-2
, ... ,
1
-(n-l)
are orthogonal.
Ex. 1.7b:
Form XA fi rst.
Chapter 2
= 0, and Za = 0 implies a = O.
Ex. 2.2:
AZ
Ex. 2.5:
Ex. 2.6:
rank (AHA)
50
Hence
Now assume
(zi,A Az i ) = (Azi,Az i )
impl ies AZ i = 0,
l, ... ,n-k.
IIAzil1
diction.
Ex. 2.10:
Ex. 2.11a:
zEN(A).
2.11d:
Inthiscase
= [ I
u
Then
un]
0
-80-
Ex. 2.13:
x = A+b +
n-r
L a.z.
i=l
x with
Ex. 2.15:
any vector z ..
I
tion of Ax. = e ..
I
Ex. 2.20e:
if BA
CA.
Chapter 3
Ex. 3.2b:
A=
Ex. 3. 7a:
d k a k = 1 if and only if c k # O.
Ex. 3. lIb:
1 [
- CiTP+iT
u/
A+-
(p+1) I -u u
p
]XUH
q
p P
m=~l L L.y.
q\P+l/i=1 j=l IJ
and
1, ... ,po
hl.!..:
-81-
y can
If
L t.
0, then
i=1 1
= 2P L~
a.
i=1
t.1 = 2pq L~ L~
i=1 y=1
y .
IJ
+::-r:h
g
qWPT+III).
1=1 y=1
iJ
so that
"
I P
m=-L
pq i =1
LY ..
j=1 1J
and
:::
I ~
I P ~
L y .. - L L y ..
q j=1 IJ
pq i=1 j=1 IJ
t. = -
Ex. 3. 14a:
np
= I v I and.!. u H = u +
n 1\ p
P P
P
o X
Ex. 3.17:
Z.
T to
block form
T(z. Xy.) = O.
1
Ex. 3.20:
-82-
C and DB+B
D.
Chapter 4
(BHB)-IB H = (CHB)+C H.
Ex. 4.2a:
4.2b:
Ex. 4.11: Use an argument similar to that one employed to establ ish
uniqueness in (2.2).
Ex. 4.16:
by Exercise 4.14.
Then
An
R
Ex. 4.17:
Then show that the first set of three equations implies the second set,
and that the second set implies X has the given form.
If A 2
Ex. 4.18:
Chapter 5
Ex. 5.2:
Use both the direct and dual form of Exercise 4.10 with
5.4b:
= X and
(AX)H
Then X = A#.
-83-
= AX
gives
1.
Appendix 2:
Selected References
1.
2.
1952.
Statistical
1974.
Generalized
New York:
Wiley.
3.
Drazin, M.P.
1958.
Pseudo-inverses in associative
65:506-513.
4.
Graybill, F.A.
1961.
An Introduction to Linear
Greville, T.N.E.
1961.
-85-
SIAM J.
~.
6.
Massachusetts:
7.
8.
9.
New Jersey:
10.
11.
12.
13.
-86-
Index to
Principal Definitions
Page
aUas . . . . .
63
Drazin inverse
59
16
fundamental equations
group inverse.
63
index . . . . . .
59
Kroneaker produat
40
Moore-Penrose inverse
13
power inverse.
. . .
69
trapezoidal deaomposition
38
66
-87-