SUBDIRECT SUMSOFNONSINGULAR
M
-MATRICESAND OF THEIR INVERSES∗
RAFAEL BRU
†
, FRANCISCO PEDROCHE
†
, AND DANIEL B.SZYLD
‡
Abstrat. Thequestionofwhenthe subdiretsumoftwo nonsingular
M
-matriesisa non-singularM
-matrixisstudied. Suientonditionsaregiven. TheaseofinversesofM
-matriesis alsostudied. Inpartiular,itisshownthatthesubdiretsumofoverlappingprinipalsubmatries ofa nonsingularM
-matrixisanonsingularM
-matrix. Someexamplesillustratingthe onditions presentedarealsogiven.AMSsubjet lassiations.15A48.
Keywords.Subdiretsum,
M
-matries,InverseofM
-matrix,Overlappingbloks.1. Introdution. Subdiretsumofmatriesaregeneralizationsoftheusualsum of matries (a
k
-subdiret sum is formally dened below in Setion 2). Theywere introduedbyFallatandJohnsonin[3℄,wheremanyoftheirpropertieswereanalyzed. Forexample,theyshowedthatthesubdiret sumofpositivedenitematries,orof symmetriM
-matries, arepositivedenite orsymmetriM
-matries,respetively. TheyalsoshowedthatthisisnottheaseforM
-matries: thesumoftwoM
-matries maynotbeanM
-matrix. Onegoalofthepresentpaperistogivesuientonditions sothatthesubdiretsumofnonsingularM
-matriesisanonsingularM
-matrix.We alsotreattheaseofthesubdiretsumofinversesofM
-matries.Subdiret sums of two overlapping prinipal submatries of a nonsingular
M
-matrixappearnaturallywhenanalyzingadditiveShwarzmethodsforMarkovhains or other matries [2℄, [4℄. In this paper we show that the subdiret sum of two overlapping prinipal submatries of a nonsingularM
-matrix is a nonsingularM
-matrix.Thepaperisstruturedasfollows. InSetion2wefousonthenonsingularityof thesubdiret sum ofanypairof nonsingularmatries, givinganexpliit expression for the inverse. In Setion 2.1 we study the
k
-subdiret sum of two nonsingularM
-matries and in partiular, the ase of subdiret sums of overlapping bloks ofnonsingular
M
-matries. InSetion2.3weextendsomeresultstothesubdiretsum ofmorethantwononsingularM
-matries. InSetion3weanalyzethesubdiretsum of twoinverses. Finally, in Setion 4we mentionsomeopen questions on subdiret sums ofP
-matries. Throughout the paper we give exampleswhih help illustrate thetheoretialresults.∗
Reeivedbytheeditors 16February 2005. Aepted forpubliation22June 2005. Handling Editor: MihaelNeumann.
†
InstitutdeMatemàtiaMultidisiplinar,UniversitatPolitèniadeValènia.CamídeVeras/n. 46022Valènia. Spain(rbrumat.upv.es, pedrohemat.upv.es). SupportedbySpanishDGIand FEDERgrantMTM2004-02998 andbythe Oina deCiéiaiTenologiadelaPresidéia deLa GeneralitatValenianaunderprojetGRUPOS03/062.
‡
2. Subdiret sums of nonsingular matries. Let
A
andB
be two square matriesofordern
1
andn
2
,respetively,andletk
beanintegersuhthat1
≤
k
≤
min(
n
1
, n
2)
. LetA
andB
bepartitionedinto2
×
2
bloksasfollows:A
=
A
11
A
12
A
21
A
22
,
B
=
B
11
B
12
B
21
B
22
,
(2.1)where
A
22
andB
11
aresquarematriesoforderk
. Following[3℄,weallthefollowing squarematrixof ordern
=
n
1
+
n
2
−
k
,C
=
A
11
A
12
0
A
21
A
22
+
B
11
B
12
0
B
21
B
22
(2.2)the
k
-subdiretsumofA
andB
anddenoteitbyC
=
A
⊕
k
B
.Weareinterestedin theasewhen
A
andB
arenonsingularmatries. We parti-tiontheinversesofA
andB
onformablyto(2.1)anddenoteitsbloksasfollows:A
−
1
=
ˆ
A
11
A
ˆ
12
ˆ
A
21
A
ˆ
22
,
B
−
1
=
ˆ
B
11
B
ˆ
12
ˆ
B
21
B
ˆ
22
,
(2.3)where,asbefore,
A
ˆ
22
andB
ˆ
11
aresquareoforderk
.In the following result we show that nonsingularity of matrix
A
ˆ
22
+ ˆ
B
11
is a neessaryandsuientonditionforthek
-subdiret sumC
tobenonsingular. The proof is basedon theuse of therelationn
=
n
1
+
n
2
−
k
to properly partition the indiatedmatries.Theorem 2.1. Let
A
andB
be nonsingular matries of ordern
1
andn
2
, re-spetively, and letk
be an integer suhthat1
≤
k
≤
min(
n
1
, n
2)
. LetA
andB
be partitionedasin(2.1)andtheirinverses bepartitionedasin(2.3). LetC
=
A
⊕
k
B
. ThenC
isnonsingularif andonlyifH
ˆ
= ˆ
A
22
+ ˆ
B
11
isnonsingular.Proof. Let
I
m
betheidentitymatrixof orderm
. Thetheorem followsfrom the followingrelation:A
−
1
O
O
I
n
−
n
1
C
I
n
−
n
2
O
O
B
−
1
=
I
n
−
n
2
A
ˆ
12
O
O
H
ˆ
B
ˆ
12
O
O
I
n
−
n
1
.
(2.4)2.1. Nonsingular
M
-matries. GivenA
=
{
a
ij
} ∈
R
m
×
n
, wewrite
A > O
(A
≥
O
), to indiatea
ij
>
0
(a
ij
≥
0
), fori
= 1
, . . . , m
,j
= 1
, . . . , n
, and suh matriesarealledpositive(nonnegative). Similarly,A
≥
B
whenA
−
B
≥
O
. Square matrieswhihhavenonpositiveo-diagonalentriesarealledZ
-matries. WeallaZ
-matrixM
anonsingularM
-matrixifM
−
1
≥
O
. Wereallsomepropertiesofthese matries;see[1℄,[8℄:
(i)Thediagonalofanonsingular
M
-matrixispositive.(iii) Amatrix
M
is anonsingularM
-matrixifand onlyif eah prinipal submatrix ofM
isanonsingularM
-matrix.(iv) A
Z
-matrixM
is anonsingularM
-matrixif and only ifthere exists apositive vetorx >
0
suh thatM x >
0
.Werstonsiderthe
k
-subdiretsumofnonsingularZ
-matries. >From(2.4)we anexpliitlywriteC
−
1
=
I
n
−
n
2
O
O
B
−
1
I
n
−
n
2
−
A
ˆ
12
H
ˆ
−
1
A
ˆ
12
H
ˆ
−
1
B
ˆ
12
O
H
ˆ
−
1
−
H
ˆ
−
1
B
ˆ
12
O
O
I
n
−
n
1
A
−
1
O
O
I
n
−
n
1
fromwhihweobtain
C
−
1
=
ˆ
A
11
−
A
ˆ
12
H
ˆ
−
1
A
ˆ
21
A
ˆ
12
−
A
ˆ
12
H
ˆ
−
1
A
ˆ
22
A
ˆ
12
H
ˆ
−
1
B
ˆ
12
ˆ
B
11
H
ˆ
−
1
A
ˆ
21
B
ˆ
11
H
ˆ
−
1
A
ˆ
22
−
B
ˆ
11
H
ˆ
−
1
B
ˆ
12
+ ˆ
B
12
ˆ
B
21
H
ˆ
−
1
A
ˆ
21
B
ˆ
21
H
ˆ
−
1
A
ˆ
22
−
B
ˆ
21
H
ˆ
−
1
B
ˆ
12
+ ˆ
B
22
(2.5)andthereforeweanstatethefollowingimmediateresult.
Theorem 2.2. Let
A
andB
be nonsingularZ
-matries of ordern
1
andn
2
, respetively,and letk
be an integersuh that1
≤
k
≤
min(
n
1
, n
2)
. LetA
andB
be partitionedasin(2.1)andtheir inversesbepartitionedasin(2.3). LetC
=
A
⊕
k
B
. LetH
ˆ
= ˆ
A
22
+ ˆ
B
11
benonsingular. ThenC
isanonsingularM
-matrixifandonly if eah ofthe nine bloksofC
−
1
in(2.5)isnonnegative.
We onsider nowthe ase where
A
andB
are nonsingularM
-matries. It was shown in [3℄ that even ifH
=
A
22
+
B
11
is a nonsingularM
-matrix, this doesnot guaranteethatC
=
A
⊕
k
B
isanonsingularM
-matrix. WepointoutthatthismatrixH
is notthematrixH
ˆ
obtainedfromA
−
1
and
B
−
1
andused in Theorem 2.1. The fatthat
H
isanonsingularM
-matrixisaneessarybutnotasuientonditionforC
tobeanonsingularM
-matrix. Suientonditionsarepresentedinthefollowingresult.
Theorem 2.3. Let
A
andB
benonsingularM
-matriespartitionedasin (2.1). Letx
1
>
0
∈
R
(
n
1
−
k
)
×
1
,
y
1
>
0
∈
R
k
×
1
,
x
2
>
0
∈
R
k
×
1
and
y
2
>
0
∈
R
(
n
2
−
k
)
×
1
be suhthatA
x
1
y
1
>
0
,
B
x
2
y
2
>
0
.
(2.6)Let
H
=
A
22
+
B
11
beanonsingularM
-matrixandlety
=
H
−
1
(
A
22
y
1
+
B
11
x
2)
.
(2.7)Then if
y
≤
y
1
andy
≤
x
2
thek
-subdiret sumC
=
A
⊕
k
B
is a nonsingularM
-matrix.Proof. Wewillshowthatthereexists
u >
0
suhthatCu >
0
. Werstnotethat from(2.6)wegetA
11
x
1
+
A
12
y
1
>
0
A
21
x
1
+
A
22
y
1
>
0
,
B
11
x
2
+
B
12
y
2
>
0
B
21
x
2
+
B
22
y
2
>
0
Taking
u
=
x
1
y
y
2
andpartitioningC
asin(2.2)weobtainCu
=
A
11
x
1
+
A
12
y
A
21
x
1
+ (
A
22
+
B
11)
y
+
B
12
y
2
B
21
y
+
B
22
y
2
.
(2.9)Sine
A
21
≤
O
andB
12
≤
O
, from (2.8) it followsthatA
22
y
1
>
0
andB
11
x
2
>
0
. SineH
−
1
≥
O
, from(2.7)wehavethat
y
ispositive,and onsequently,soisu
,i.e.,u >
0
. We will show thatCu >
0
oneblokof rowsin (2.9) at atime. Ify
≤
y
1
,as
A
12
≤
0
, we have thatA
12
y
≥
A
12
y
1
and again using (2.8) weobtain that the rstblokofrowsofCu
ispositive. Inasimilarway, theonditiony
≤
x
2
together with the last equation of (2.8) allows to onlude that the third blok of rows ofCu
is positive. Finally, substitutingy
given by(2.7) in the seond row ofCu
andonsidering(2.8)weonludethat theseondblokofrowsof
Cu
isalsopositive. Note thatA
andB
arenonsingularM
-matries and thereforethe positive ve-tors(
x
1
, y
1)
and(
x
2
, y
2)
of(2.6)alwaysexist. Thistheorem givessuientbut not neessaryonditionsforC
=
A
⊕
k
B
to beanonsingularM
-matrix,asillustratedin Example2.5furtherbelow.Example2.4. Thematries
A
=
3
−
2
−
1
−
1
/
2
2
−
3
−
1
−
1
4
andB
=
1
−
2
−
1
/
3
−
3
9
0
−
2
−
1
/
2
6
,
andthevetors
x
1
y
1
=
1
.
8
2
1
andx
2
y
2
=
2
.
5
1
1
satisfy the inequalities (2.6), and omputing the vetor
y
from (2.7) we gety
≈
(1
.
95
,
0
.
87)
T
,whih satisfy
y
≤
y
1
andy
≤
x
2
. Thereforethe2
-subdiretsumC
=
3
−
2
−
1
0
−
1
/
2
3
−
5
−
1
/
3
−
1
−
4
13
0
0
−
2
−
1
/
2
6
isanonsingular
M
-matrixinaordanewith Theorem2.3. Example2.5. ThematriesA
=
5
−
1
/
2
−
1
/
3
−
1
4
−
2
−
1
−
6
10
andB
=
1
−
2
−
1
/
3
−
3
9
0
−
2
−
1
/
2
6
andthevetors
x
1
y
1
=
1
1
1
andx
2
y
2
=
2
.
5
1
1
satisfytheinequalities(2.6),butomputingvetor
y
from(2.7)weobtainy
≈
(1
.
18
,
0
.
85)
T
, whihdoesnotsatisfytheonditionsof Theorem2.3. Never-thelessthe
2
-subdiret sumC
=
A
⊕
2
B
=
5
−
1
/
2
−
1
/
3
0
−
1
5
−
4
−
1
/
3
−
1
−
9
19
0
0
−
2
−
1
/
2
6
isanonsingular
M
-matrix.In the speial ase of
A
andB
blok lower and upper triangular nonsingularM
-matries, respetively, theresultsof Theorems 2.2 and 2.3 areeasy to establish.Let
A
=
A
11
0
A
21
A
22
,
B
=
B
11
B
12
0
B
22
,
(2.10)with
A
22
andB
11
squarematriesoforderk
.Theorem 2.6. Let
A
andB
be nonsingular lower and upper blok triangular nonsingularM
-matries, respetively,partitioned asin (2.10). ThenC
=
A
⊕
k
B
is anonsingularM
-matrix.Proof. We anrepeat the same argumentas in the proof of Theorem 2.3 with the advantage of having
A
12
=
O
andB
21
=
O
. Note that onditionsy
≤
y
1
andy
≤
x
2
arenotneessaryherebeausetherstand lastblokofrowsofCu
in (2.9) areautomatiallypositivein thisase.Remark 2.7. Theexpressionof
C
−
1
isgivenby(2.5). Inthis partiularaseof bloktriangularmatrieswehave
A
ˆ
12
=
O
,B
ˆ
21
=
O
,A
ˆ
22
=
A
−
1
22
,B
ˆ
11
=
B
−
1
11
,fromwhih
H
ˆ
=
A
−
1
22
+
B
−
1
11
. If,inaddition,A
22
=
B
11
,thenweobtainC
−
1
=
A
−
1
11
O
O
−
1
2
A
−
1
22
A
21
A
−
11
1
1
2
A
−
1
22
−
1
2
A
−
1
22
B
12
B
22
−
1
O
O
B
−
1
22
≥
O.
Example 2.8. Thematries
A
=
3
0
0
−
1
5
−
1
−
1
−
9
5
andB
=
6
−
2
−
1
−
4
3
−
3
0
0
2
satisfythehypothesesofTheorem 2.6. Thematries
C
=
A
⊕
2
B
andC
−
1
are
C
=
3
0
0
0
−
1
11
−
3
−
1
−
1
−
13
8
−
3
0
0
0
2
,
C
−
1
=
1
/
3
0
0
0
11
/
147
8
/
49
3
/
49
17
/
98
8
/
49
13
/
49 11
/
49 23
/
49
0
0
0
1
/
2
Insome appliations, suh asin domain deomposition [6℄, [7℄, matries
A
andB
partitioned as in (2.1)arise with aommon blok, i.e.,A
22
=
B
11
. In the nextexample we show that even if
A
andB
are nonsingularM
-matries, and so is the ommonblok,weannotensurethatC
=
A
⊕
k
B
isanonsingularM
-matrix.Example2.9. Thematries
A
=
370
−
342
−
318
−
448
737
−
107
−
46
−
190
444
,
B
=
737
−
107
−
134
−
190
444
−
440
−
885
−
182
603
arenonsingular
M
-matrieswithA
22
=
B
11
anM
-matrix,butC
=
A
⊕
2
B
isnotanM
-matrix,sinewehaveC
=
370
−
342
−
318
0
−
448
1474
−
214
−
134
−
46
−
380
888
−
440
0
−
885
−
182
603
and
C
−
1
≈
−
0
.
0291
−
0
.
0242
−
0
.
0204
−
0
.
0203
−
0
.
0145
−
0
.
0109
−
0
.
0098
−
0
.
0096
−
0
.
0214
−
0
.
0163
−
0
.
0132
−
0
.
0133
−
0
.
0277
−
0
.
0210
−
0
.
0183
−
0
.
0164
.
Inthenext setionwe shallsee that when
A
andB
share ablok andthey are submatriesofagivennonsingularM
-matrix,theresultingk
-subdiretsumisinfat anonsingularM
-matrix.2.2. Overlapping
M
-matries. InthissetionwerestritA
andB
tobe prin-ipalsubmatriesofagivennonsingularM
-matrixandsuhthattheyhaveaommon blok. LetM
=
M
11
M
12
M
13
M
21
M
22
M
23
M
31
M
32
M
33
(2.11)beanonsingular
M
-matrixwithM
22
squarematrixoforderk
≥
1
andletA
=
M
11
M
12
M
21
M
22
and
B
=
M
22
M
23
M
32
M
33
(2.12)
beoforder
n
1
andn
2
,respetively. Thek
-subdiretsumofA
andB
isthusgivenbyC
=
A
⊕
k
B
=
M
11
M
12
O
M
21
2
M
22
M
23
O
M
32
M
33
.
(2.13)Theorem 2.10. Let
M
be a nonsingularM
-matrix partitioned as in (2.11), andletA
andB
be twooverlapping prinipal submatries given by (2.12). Then thek
-subdiretsumC
=
A
⊕
k
B
is anonsingularM
-matrix. Proof. Letusonstrutann
×
n Z
-matrixT
asfollows:T
=
M
11
2
M
12
M
13
M
21
2
M
22
M
23
M
31
2
M
32
M
33
.
(2.14)Then
T
=
M
diag(
I,
2
I, I
)
andwegetT
−
1
= diag(
I,
(1
/
2)
I, I
)
M
−
1
≥
O
. Then
T
is anonsingularM
-matrix. FinallysineC
isaZ
-matrixandC
≥
T
weonludethatC
isanonsingularM
-matrix.Example2.11. Thefollowingnonsingular
M
-matrixispartitionedasin (2.11):M
=
13
/
14
−
4
/
23
−
3
/
20
−
1
/
42
−
19
/
186
−
3
/
46
−
3
/
7
21
/
23
−
1
/
5
−
1
/
21
−
1
/
93
−
6
/
23
−
1
/
7
−
7
/
46
17
/
20
−
1
/
14
−
1
/
186
−
2
/
23
−
4
/
21
−
27
/
92
−
1
/
15
4
/
7
−
58
/
93
−
27
/
92
−
1
/
14
−
9
/
46
−
3
/
10
−
1
/
7
53
/
62
−
9
/
46
−
2
/
21
−
9
/
92
−
2
/
15
−
2
/
7
−
7
/
62
83
/
92
.
(2.15)Takingoverlappingsubmatries
A
andB
asin(2.12)the3
-subdiretsumC
=
A
⊕
3
B
isgivenbyC
=
13
/
14
−
4
/
23
−
3
/
20
−
1
/
42
−
19
/
186
0
−
3
/
7
21
/
23
−
1
/
5
−
1
/
21
−
1
/
93
0
−
1
/
7
−
7
/
46
17
/
10
−
1
/
7
−
1
/
93
−
2
/
23
−
4
/
21
−
27
/
92
−
2
/
15
8
/
7
−
116
/
93
−
27
/
92
−
1
/
14
−
9
/
46
−
3
/
5
−
2
/
7
53
/
31
−
9
/
46
0
0
−
2
/
15
−
2
/
7
−
7
/
62
83
/
92
anditisanonsingular
M
-matrixaordingtoTheorem2.10. Infat,wehavethatC
−
1
≈
1
.
3500 0
.
3977 0
.
2624 0
.
1609 0
.
2103 0
.
1232
0
.
7628 1
.
4108 0
.
3383 0
.
2085 0
.
2185 0
.
1478
0
.
3007 0
.
2845 0
.
7422 0
.
2006 0
.
1824 0
.
1763
1
.
1024 1
.
1571 0
.
8927 1
.
6092 1
.
3118 0
.
8940
0
.
4854 0
.
5256 0
.
5116 0
.
4379 0
.
9664 0
.
4013
0
.
4543 0
.
4743 0
.
4564 0
.
5941 0
.
5634 1
.
4679
.
2.3.
k
-subdiret sum ofp M
-matries. In this setion we extend Theo-rems 2.3 and 2.10 to the subdiret sum of several nonsingularM
-matries. Exam-ple2.14laterinthesetionillustratesthenotationusedin theproofs.Theorem 2.12. Let
A
i
∈
R
n
i
×
n
i
,
i
= 1
, . . . p
,be nonsingularM
-matries parti-tionedasA
i
=
A
i,
11
A
i,
12
A
i,
21
A
i,
22
with
A
i,
11
asquarematrixoforderk
i
−
1
≥
1
andA
i,
22
asquarematrixoforderk
i
≥
1
, i.e.,n
i
=
k
i
−
1
+
k
i
. SineA
i
are nonsingularM
-matries we have that there existx
i
>
0
∈
R
(
n
i
−
k
i
)
×
1
andy
i
>
0
∈
R
k
i
×
1
suhthat
A
i
x
i
y
i
>
0
,
i
= 1
, . . . , p.
(2.17)Let
C
0
=
A
1
anddene the followingp
−
1
k
i
-subdiretsums:C
i
=
C
i
−
1
⊕
k
i
A
i
+1
,
i
= 1
, . . . , p
−
1
,
(2.18)i.e.,
C
1
=
A
1
⊕
k
1
A
2
,
C
2
=
(
A
1
⊕
k
1
A
2)
⊕
k
2
A
3
=
C
1
⊕
k
2
A
3
,
. . .
C
p
−
1
=
A
1
⊕
k
1
A
2
⊕
k
2
· · · ⊕
k
p
−2
A
p
−
1
⊕
k
p
−1
A
p
=
C
p
−
2
⊕
k
p
−
1
A
p
.
Eahsubdiretsum
C
i
isof orderm
i
,suhthatm
0
=
n
1
andm
i
=
m
i
−
1
+
n
i
+1
−
k
i
=
m
i
−
1
+
k
i
+1
,
i
= 1
, . . . , p
−
1
.
Letuspartition
C
i
in the formC
i
=
C
i,
11
C
i,
12
C
i,
21
C
i,
22
,
i
= 1
, . . . , p
−
1
,
(2.19)with
C
i,
22
asquarematrixof orderk
i
+1
. LetH
i
=
C
i
−
1
,
22
+
A
i
+1
,
11
,
i
= 1
, . . . p
−
1
,
benonsingular
M
-matriesandletz
i
=
H
i
−
1
(
C
i
−
1
,
22
y
i
+
A
i
+1
,
11
x
i
+1)
,
i
= 1
, . . . p
−
1
.
Then,if
z
i
≤
y
i
andz
i
≤
x
i
+1
,thesubdiretsumsC
i
given by(2.18)are nonsingularM
-matriesfori
= 1
, . . . , p
−
1
.Proof. It is easy to see that applying Theorem 2.3 to eah onseutive pair of matries
C
i
wehavethatC
1
,C
2
,...,C
p
−
1
arenonsingularM
-matries. Thisanbe shownbyindution.WenowextendTheorem 2.10tothesub-diretsum of
p
submatriesof agiven nonsingularM
-matrixM
. Tothatend,werstdeneM
(
S
)
aprinipalsubmatrixofM
withrowsandolumnswithindiesinthesetofindiesS
=
{
i, i
+ 1
, i
+ 2
, . . . , j
}
. In[2℄wealltheseonseutiveprinipalsubmatries. Forexample,matriesA
andB
givenby(2.12)anbeexpressedasasubmatriesofM
givenby(2.11)asA
=
M
(
S
1)
,Theorem 2.13. Let
M
be a nonsingularM
-matrix. LetA
i
=
M
(
S
i
)
,i
=
1
, . . . , p
,beprinipal onseutivesubmatriesofM
andonsiderthep
−
1
k
i
-subdiretsumsgiven by
C
i
=
C
i
−
1
⊕
k
i
A
i
+1
,
i
= 1
, . . . , p
−
1
,
(2.20)inwhih
C
0
=
A
1
. Theneahofthek
i
-subdiretsumsC
i
isanonsingularM
-matrix. Proof. It is easy to relate the struture of eahC
i
to that of the submatriesA
i
involved. WeonsiderthatA
i
areoverlappingprinipal submatriesoftheform(2.12)butallowingthateah
A
i
hasdierentnumberofbloks. LetM
bepartitioned asM
=
M
11
M
12
M
13
· · ·
M
1
n
M
21
M
22
M
23
· · ·
M
2
n
M
31
M
32
M
33
· · ·
M
3
n
. . .
. . .
. . .
. .
. . . .
M
n
1
M
n
2
M
n
3
· · ·
M
nn
(2.21)
aordingwith the size of theprinipal submatries
A
i
. Eah blokM
ij
may beasubmatrixofmorethanone
A
m
,m
= 1
, . . . , p
. Letb
(
l
)
ij
≥
0
bethenumberofmatriesA
m
suhthatM
ij
isasubmatrixofA
m
,form
= 1
, . . . , l
+ 1
. Ofourseweanhaveb
(
ij
l
)
= 0
. Letusonsiderthel
thsubdiretsumC
l
,1
≤
l
≤
p
−
1
,whihisoftheformC
l
=
b
11
(
l
)
M
11
b
(
12
l
)
M
12
b
13
(
l
)
M
13
· · ·
b
(
1
l
l
)
M
1
l
b
(
21
l
)
M
21
b
(
l
)
22
M
22
b
(
l
)
23
M
23
· · ·
b
(
l
)
2
l
M
2
l
b
(
31
l
)
M
31
b
(
l
)
32
M
32
b
(
l
)
33
M
33
· · ·
b
(
l
)
3
l
M
3
l
. . .
. . .
. . .
. .
. . . .
b
(
l
1
l
)
M
l
1
b
(
l
2
l
)
M
l
2
b
(
l
3
l
)
M
l
3
· · ·
b
(
ll
l
)
M
ll
.
(2.22)Observe that
C
l
is aZ
-matrix and thatb
(
l
)
ii
>
0
. Furthermore, for eah olumn itholdsthat
b
(
l
)
ii
≥
b
(
l
)
ji
, j
= 1
, . . . , l
.The proofproeeds in amanner similar to that of Theorem 2.10. Considerthe
Z
-matrix(partitionedinthesamemannerasM
)T
l
=
M
l
diag(
b
(
l
)
11
I, b
(
l
)
22
I, b
(
l
)
33
I, . . . , b
(
l
)
ll
I
)
,
where
M
l
istheprinipalsubmatrixof(2.21)withrowandolumnbloksfrom1
tol
. It followsthatT
−
1
l
≥
O
and thereforeT
l
isanonsingularM
-matrix. Finally, sineC
l
≥
T
l
,weonludethatC
l
isanonsingularM
-matrix,l
= 1
, . . . , p
.Example 2.14. Given the nonsingular
M
-matrixM
of Example 2.11, let us onsiderthefollowingoverlappingbloksA
1
=
M
(
{
1
,
2
,
3
}
) =
13
/
14
−
4
/
23
−
3
/
20
−
3
/
7
21
/
23
−
1
/
5
−
1
/
7
−
7
/
46
17
/
20
A
2
=
M
(
{
2
,
3
,
4
,
5
}
) =
21
/
23
−
1
/
5
−
1
/
21
−
1
/
93
−
7
/
46
17
/
20
−
1
/
14
−
1
/
186
−
27
/
92
−
1
/
15
4
/
7
−
58
/
93
−
9
/
46
−
3
/
10
−
1
/
7
53
/
62
,
A
3
=
M
(
{
3
,
4
,
5
,
6
}
) =
17
/
20
−
1
/
14
−
1
/
186
−
2
/
23
−
1
/
15
4
/
7
−
58
/
93
−
27
/
92
−
3
/
10
−
1
/
7
53
/
62
−
9
/
46
−
2
/
15
−
2
/
7
−
7
/
62
83
/
92
.
Thenwehavethe
2
-subdiret sumC
1
=
A
1
⊕
2
A
2
=
13
/
14
−
4
/
23
−
3
/
20
0
0
−
3
/
7
42
/
23
−
2
/
5
−
1
/
21
−
1
/
93
−
1
/
7
−
7
/
23
17
/
10
−
1
/
14
−
1
/
186
0
−
27
/
92
−
1
/
15
4
/
7
−
58
/
93
0
−
9
/
46
−
3
/
10
−
1
/
7
53
/
62
,
whih isanonsingular
M
-matrix,andthe3
-subdiretsumC
2
=
C
1
⊕
3
A
3
=
13
/
14
−
4
/
23
−
3
/
20
0
0
0
−
3
/
7
42
/
23
−
2
/
5
−
1
/
21
−
1
/
93
0
−
1
/
7
−
7
/
23
51
/
20
−
1
/
7
−
1
/
93
−
2
/
23
0
−
27
/
92
−
2
/
15
8
/
7
−
116
/
93
−
27
/
92
0
−
9
/
46
−
3
/
5
−
2
/
7
53
/
31
−
9
/
46
0
0
−
2
/
15
−
2
/
7
−
7
/
62
83
/
92
,
whihisalsoanonsingular
M
-matrixinaordanewithTheorem2.13. Observethat in this example we havek
1
= 2
andk
2
= 3
. Note also that, for example, wehaveb
(1)
22
= 2
,b
(1)
33
= 2
,b
(1)
14
= 0
,b
(2)
22
= 2
,b
(3)
22
= 2
,b
(2)
33
= 3
,b
(2)
14
= 0
.3. Subdiret sumsof inverses. Let
A
andB
be nonsingularmatries parti-tionedasin (2.1). Inthis setionweonsider thek
-subdiret sumof theirinverses. We will establish ounterparts to some of results in the previous setions. Let us denotebyG
=
A
−
1
⊕
k
B
−
1
,withA
−
1
and
B
−
1
partitionedasin(2.3),i.e.,
G
=
ˆ
A
11
A
ˆ
12
0
ˆ
A
21
A
ˆ
22
+ ˆ
B
11
B
ˆ
12
0
B
ˆ
21
B
ˆ
22
.
(3.1)As aorollaryto, andin analogyto Theorem 2.1, the nextstatementindiates thatthenonsingularityof
A
22
+
B
11
isaneessaryonditiontoobtainG
nonsingular. Theorem 3.1. LetA
andB
benonsingular matriespartitionedasin(2.1)and let their inverses be partitioned as in (2.3). LetG
=
A
−
1
We remark that in analogyto the expression (2.5) of
C
−
1
, the expliit form of
G
−
1
isG
−
1
=
A
11
−
A
12
H
−
1
A
21
A
12
−
A
12
H
−
1
A
22
A
12
H
−
1
B
12
B
11
H
−
1
A
21
B
11
H
−
1
A
22
−
B
11
H
−
1
B
12
+
B
12
B
21
H
−
1
A
21
B
21
H
−
1
A
22
−
B
21
H
−
1
B
12
+
B
22
.
(3.2)Corollary3.2. When
A
andB
arenonsingularM
-matrieswith the ommon blokA
22
=
B
11
asquarematrixof orderk
,i.e., ofthe formA
=
A
11
A
12
A
21
A
22
,
B
=
A
22
B
12
B
21
B
22
,
(3.3)then
H
= 2
A
22
isnonsingularandthereforeG
=
A
−
1
⊕
k
B
−
1
isnonsingular.We note that this isthe asewhen
A
andB
are overlappingsubmatriesof anM
-matrix,i.e.,oftheform(2.12)and(2.11)onsideredinSetion2.2,wherewewereinterestedinthesubdiretsumof
A
andB
. Hereweonludethatthesubdiretsum oftheirinversesisalwaysnonsingular.Example3.3. Let
A
andB
bethematriesofExample2.11,thenaordingto Corollary3.2,the3
-subdiretsumoftheinversesG
=
A
−
1
⊕
3
B
−
1
≈
1
.
5033 0
.
5513 0
.
5547 0
.
2757 0
.
3912
0
0
.
9540 1
.
5996 0
.
7158 0
.
3635 0
.
4038
0
0
.
6004 0
.
5636 2
.
9750 0
.
8144 0
.
7407 0
.
3708
2
.
0383 2
.
1242 3
.
5729 6
.
5498 5
.
3372 2
.
0139
0
.
8953 0
.
9650 2
.
0470 1
.
8025 3
.
9062 0
.
9048
0
0
0
.
8551 1
.
3803 1
.
2652 1
.
9143
isanonsingularmatrix.
Intheaboveexampleadiretomputationshowsthat
G
−
1
isnotan
M
-matrix:G
−
1
≈
0
.
8900
−
0
.
2337
−
0
.
0750
−
0
.
0119
−
0
.
0511
0
.
0512
−
0
.
4682
0
.
8566
−
0
.
1000
−
0
.
0238
−
0
.
0054
0
.
0470
−
0
.
0714
−
0
.
0761
0
.
4250
−
0
.
0357
−
0
.
0027
−
0
.
0435
−
0
.
0952
−
0
.
1467
−
0
.
0333
0
.
2857
−
0
.
3118
−
0
.
1467
−
0
.
0357
−
0
.
0978
−
0
.
1500
−
0
.
0714
0
.
4274
−
0
.
0978
0
.
1242
0
.
2045
−
0
.
0667
−
0
.
1429
−
0
.
0565
0
.
7123
whih is nota
Z
-matrix. Note that whenA
andB
areM
-matries we havefrom (3.1) thatG
=
A
−
1
⊕
B
−
1
is nonnegative. Therefore assuming that
G
−
1
exists we have
(
G
−
1
)
−
1
≥
O
. Thenitisanaturalquestiontoseekonditionssothat
G
−
1
isa nonsingular
M
-matrix. Westudythisquestionnext.Theexpressions (3.1)of
G
and (3.2)ofG
−
1
, Theorem3.1, and theobservation thatfornonsingular
M
-matrieswehave(
G
−
1
)
−
1
≥
O
, implythefollowingresult. Theorem 3.4. Let
A
andB
benonsingularM
-matriespartitioned asin (2.1) andtheir inverses partitioned asin (2.3). LetG
=
A
−
1
⊕
k
B
−
1
withk
≥
1
, andletH
=
A
22
+
B
11
benonsingular. ThenG
−
1
isa nonsingular
M
-matrixif andonly ifG
−
1
Corollary 3.5. Let
A
andB
be lower and upper bloktriangular nonsingularM
-matries, respetively, partitioned asin (2.10) withA
22
andB
11
square matriesof order
k
andH
=
A
22
+
B
11
nonsingular. ThenG
−
1
= (
A
−
1
⊕
k
B
−
1
)
−
1
is anonsingular
M
-matrixifandonly ifthe followingonditionshold: i)B
11
H
−
1
A
21
≤
O
. ii)B
11
H
−
1
A
22
isaZ
-matrix.iii)
−
B
11
H
−
1
B
12
+
B
12
≤
O
.Proof. >From(3.2)and(2.10)wehavethat
G
−
1
=
A
11
0
0
B
11
H
−
1
A
21
B
11
H
−
1
A
22
−
B
11
H
−
1
B
12
+
B
12
0
0
B
22
(3.4)andtherefore
G
−
1
isa
Z
-matrixifandonlyiftheonditionsi),ii)andiii) hold. Conditionsi),ii)andiii)intheorollaryarenotasstringentastheymayappear. Forexample,letA
andB
bebloktriangularnonsingularM
-matriespartitionedas in(2.10)with aommonblokA
22
=
B
11
,asquarematrixoforderk
, i.e.,A
=
A
11
0
A
21
A
22
and
B
=
A
22
B
12
0
B
22
.
(3.5)Then
G
−
1
= (
A
−
1
⊕
k
B
−
1
)
−
1
is anonsingularM
-matrix,sine we havefrom (3.4)that
G
−
1
=
A
11
O
O
1
2
A
21
1
2
A
22
1
2
B
12
O
O
B
22
,
andtherefore
G
−
1
isa
Z
-matrix. Infat,inthisase,wehaveG
=
A
−
1
11
O
O
−
A
−
1
22
A
21
A
−
1
11
2
A
−
1
22
−
A
−
1
22
B
12
B
−
1
22
O
O
B
−
1
22
≥
O.
Thenextexampleillustratesthissituation.
Example3.6. Let
A
andB
bethematriesofExample2.8,thenG
=
A
−
1
⊕
2
B
−
1
=
1
/
3
0
0
0
1
/
8
49
/
80 21
/
80
9
/
20
7
/
24 77
/
80 73
/
80 11
/
10
0
0
0
1
/
2
,
and
G
−
1
=
3
0
0
0
−
18
/
49 146
/
49
−
6
/
7
−
39
/
49
−
4
/
7
−
22
/
7
2
−
11
/
7
0
0
0
2
isanonsingular
M
-matrixinaordanewithCorollary3.5.Note that if the hypotheses of Corollary 3.5 are satised, and realling Theo-rem2.6,wehavethat eahofthematries
C
=
A
⊕
k
B
andG
−
1
= (
A
−
1
⊕
k
B
−
1
)
−
1
arebothnonsingular
M
-matries.4.
P
-matries. A square matrix is aP
-matrix if all its prinipal minors are positive. As aonsequene we havethat all the diagonal entries of aP
-matrix are positive. ItisalsofollowsthatanonsingularM
-matrixisaP
-matrix. Itanalsobe shownthat ifA
isanonsingularM
-matrix,thenA
−
1
isa
P
-matrix;see,e.g.,[5℄. In [3℄ it is shown that thek
-subdiret sum (withk >
1
) of twoP
-matries is not neessarilyaP
-matrix. Our results in Setions 2.1 and 3 hold for nonsingularM
-matriesandinversesofM
-matries,respetively. Asthesetwolassesofmatriesare subsetsof
P
-matries, it is naturalto ask ifsimilar suient onditionsan be foundsothatthek
-subdiretsumofP
-matriesisaP
-matrix. Thefollowingexample indiates thattheanswermaynotbeeasytoobtain,sineevenin thesimplestase ofdiagonalsubmatriesthek
-subdiretsummaynotbeaP
-matrix.Example 4.1. Giventhe
P
-matriesA
=
543 388 322
69
160
0
368
0
375
,
B
=
136
0
219
0
225 159
61
177 230
wehavethatthe
2
-subdiretsumC
=
A
⊕
2
B
=
543 388 322
0
69
296
0
219
368
0
600 159
0
61
177 230
isnota
P
-matrix,sinedet(
C
)
<
0
.Aknowledgment. Wethanktherefereeforaveryarefulreadingofthemanusript andforhisomments.
REFERENCES
[1℄ A.BermanandR.J.Plemmons.NonnegativeMatriesintheMathematialSienes,Aademi Press,NewYork,1979.Reprintedandupdated,SIAM,Philadelphia,1994.
[2℄ R.Bru, F.Pedrohe,andD.B.Szyld.AdditiveShwarzIterationsforMarkovChains,SIAM J.MatrixAnal.Appl.,toappear.
[3℄ S.M.FallatandC.R.Johnson.Sub-diretsumsandpositivitylassesofmatries.LinearAlgebra Appl.,288:149173,1999.
[4℄ A. Frommer and D.B. Szyld. Weighted Max Norms, Splittings,and Overlapping Additive ShwarzIterations,NumerisheMathematik,83:259278,1999.
[5℄ R.A.HornandC.R.Johnson.MatrixAnalysis.CambridgeUniversityPress,NewYork,1985. [6℄ B.F.Smith,P.E.Bjørstad,andW.D.Gropp.Domaindeomposition: Parallelmultilevel meth-odsforelliptipartialdierentialequations.CambridgeUniversityPress,Cambridge,1996. [7℄ A.ToselliandO.B.Widlund.DomainDeomposition:AlgorithmsandTheory.SpringerSeries
inComputationalMathematis,vol.34. Springer,NewYork,2005.