Electronic Journal of Linear Algebra ISSN 1081-3810 A Publication of The International Linear Algebra Society Volume 23, Pp. 88-114, January 2012

ELA
ON MULTIPOINT BOUNDARY VALUE PROBLEMS FOR INDEX-2

LINEAR SINGULAR DIFFERENCE EQUATIONS
LE CONG LOI
Abstract. On the background of a careful analysis of index-2 linear singular dierence equations
with both constant and varying coecients cases, multipoint boundary value problems for these equa-
tions are considered. Necessary and sucient conditions for the solvability of multipoint boundary
value problems are established. Further, general solution formulae are explicitly constructed.
Key words. Index, Matrix pencil, Linear singular dierence equations, Multipoint boundary
value problems.
AMS subject classications. 15A06, 39A05, 39A10.
1. Introduction. In recent years, there has been considerable interest in study-
ing linear singular dierence equations (LSDEs) of the form
A
n
x
n+1
= B
n
x
n
+q
n
, n 0, (1.1)
where A
n
, B
n
R
mm
, q
n
R
m
are given and rankA
n
= r (1 r m 1) for all
n 0 (see [2][8] and references therein). The index notion of a matrix pencil was
introduced to investigate Eq. (1.1) with constant coecients. Further, the solvability
of initial value problems (IVPs) has been studied thoroughly [4][6]. However, as
far as we know the qualitative questions such as the existence, uniqueness, etc. of
multipoint boundary value problems (MPBVPs) for (1.1) with constant coecients
have not been discussed. In the varying coecients case, the index-1 concept of Eq.
(1.1) was also introduced in [2, 8] and the solvability of IVPs as well as MPBVPs
for index-1 LSDEs has been considered in [2, 3, 8]. Later on, the index-2 concept
of Eq. (1.1) has been proposed, and basing on this index-2 notion, the condition of
solvability as well as the solution formula of IVPs for index-2 LSDE (1.1) have been
established in [7]. As discussed in [7], many valid results for index-1 case can be
extended to index-2 case, however, the extension meets with some diculties.
The main goal of this paper is studying MPBVPs for index-2 LSDE (1.1) in both
constant and varying coecients cases. The index-2 of a matrix pencil and index-2
Received by the editors on September 18, 2010. Accepted for publication on November 25, 2011.
Handling Editor: Daniel Szyld.
Department of Mathematics, Vietnam National University, 334 Nguyen Trai, Hanoi, Vietnam
(loilc@vnu.edu.vn). Supported by NAFOSTED grant no. 101.02.65.09.
88
Electronic Journal of Linear Algebra ISSN 1081-3810
A publication of the International Linear Algebra Society
Volume 23, pp. 88-114, January 2012
http://math.technion.ac.il/iic/ela
ELA
Multipoint Boundary Value Problems 89
of Eq. (1.1) turn to be the keystone in the analysis of MPBVPs. For index-2 LSDEs
with constant coecients, similarly as in [4][6], one can solve Eq. (1.1) by means
of index of a matrix pencil and Drazin inverse. It is well known that many results
for constant coecients LSDEs cannot be directly generalized to varying coecients
LSDEs (ref. [2, 3, 7, 8]). Thus, in the varying coecients case, our approach to
LSDEs is based on index-2 notion of Eq. (1.1) and projections. We shall develop
some techniques of index-1 LSDEs in [3, 8] for index-2 LSDEs.
The paper is organized as follows. In Section 2 we recall some denitions and
preliminary results, as well as give some simple results concerning index-2 LSDE (1.1).
Necessary and sucient conditions for the solvability and a general formula solution
of MPBVPs for index-2 LSDE (1.1) will be established in Section 3.
2. Preliminaries. We start this section by recalling the Drazin inverse of a
matrix and the index notion of a matrix pencil, which have been studied in [4, 6].
Firstly, if M R
mm
, the index of M, denoted by ind(M), is the least non-negative
integer such that kerM
= kerM
+1
. It is worth noting that the following theorem
plays an important role to study autonomous LSDEs.
Theorem 2.1. [4] Suppose that M R
mm
, ind(M)= and rankM
= t. Then
there exists a nonsingular matrix S R
mm
such that
M = S
_
W 0
0 N
_
S
1
, (2.1)
where W is a nonsingular t t matrix and N is a nilpotent (mt) (mt) matrix
with =ind(N).
If M R
mm
is given in the form (2.1), then the Drazin inverse of M, denoted
by M
D
, is dened by
M
D
= S
_
W
1
0
0 0
_
S
1
.
It is easy to verify that
MM
D
= M
D
M, M
D
MM
D
= M
D
, M
k+1
M
D
= M
k
for k Ind(M)
and the Drazin inverse is unique.
In what follows, we consider A, B R
mm
and always assume that the matrix
pencil (A, B) is regular (i.e., there exists a scalar C such that A + B is non-
singular) and let

A
:= (A + B)
1
A,

B
:= (A + B)
1
B,

f
:= (A + B)
1
f for
f R
m
. Observe that

B
= I
, hence,

A
and

B
commute.
Theorem 2.2. [4] Suppose that the matrix pencil (A, B) is regular and f R
m
.
ELA
90 L.C. Loi
Then for all , C for which (A + B)
1
and (A + B)
1
exist, the following
statements hold:
(i) ind(
) = ind(
),
(ii)

A

A
D
=

A

A
D
,
(iii)

A
D
=

A
D
and

B
D
=

B
D
,
(iv)

A
D
=

A
D
and

B
D
=

B
D
.
If (A, B) is regular and det(A + B) = 0, then ind(
) is called the index of

the pencil (A, B), denoted by ind(A, B), i.e., ind(A, B) := ind(
). Theorem 2.2
guarantees that the denition of the index of the matrix pencil does not depend on
the chosen value .
Next, to study the index-2 LSDE (1.1) with variable coecients, we start with
some basic denitions for non-autonomous LSDEs (see [2, 3, 8, 7]). Let Q
n
be any
projection onto kerA
n
and T
n
GL(R
m
) for all n 0 such that T
n
|
kerAn
is an
isomorphism from kerA
n
onto kerA
n1
, here we put A
1
:= A
0
. Denote again by T
n
the matrix induced by the operator T
n
.
Lemma 2.3. [7] The matrix G
n
:= A
n
+B
n
T
n
Q
n
is nonsingular if and only if
kerA
n1
S
n
= {0},
where S
n
:= {z R
m
: B
n
z imA
n
}.
Definition 2.4. [7] The LSDE (1.1) is said to be of index-1 if
(i) rankA
n
r,
(ii) kerA
n1
S
n
= {0}.
Now we suppose that the matrices G
n
are singular for all n 0, i.e., Eq. (1.1) is
of higher index. Put P
n
:= I Q
n
for all n 0 and let A
+
n
denote the Moore-Penrose
generalized inverse of A
n
.
Lemma 2.5. [7] The following relation
(T
n
+T
n
P
n
A
+
n
B
n
T
n
Q
n
)kerG
n
= kerA
n1
S
n
is valid.
It is worth noting that the matrices (T
n
+T
n
P
n
A
+
n
B
n
T
n
Q
n
) are nonsingular for
all n 0, consequently, we come to the following corollary.
Corollary 2.6. [7] dim(kerG
n
)= dim(ker A
n1
S
n
), n 0.
Lemma 2.7. [7] Let Q
n
,

Q
n
be two projections onto kerA
n
and T
n
,

T
n
GL(R
m
)
such that T
n
|
kerAn
,

T
n
|
kerAn
are two isomorphisms between kerA
n
and kerA
n1
. Put
ELA
G
n
:= A
n
+B
n
T
n
Q
n
,

G
n
:= A
n
+B
n
T
n

Q
n
and
S
1,n
:= {z R
m
: B
n
P
n1
z imG
n
},

S
1,n
:= {z R
m
: B
n

P
n1
z im
G
n
}.
Then, the following relations hold:
G
n
= G
n
(P
n
+T
1
n
T
n

Q
n
), n 0, (2.2)
S
1,n
= (
P
n1
+

T
1
n1
T
n1
Q
n1
)S
1,n
, n 0, (2.3)
ker
G
n

S
1,n+1
= (
P
n
+

T
1
n
T
n
Q
n
)(kerG
n
S
1,n+1
), n 0. (2.4)
Remark that the identity (2.4) ensures that the following denition does not
depend on the choice of the projections onto kerA
n
and the isomorphisms between
kerA
n
and kerA
n1
. For well-denedness, we put G
1
:= G
0
.
Definition 2.8. [7] The LSDE (1.1) is said to be of index-2 if the following
conditions
(i) rankA
n
r, 1 r m1,
(ii) dim(kerA
n1
S
n
) ms, 1 s m1,
(iii) kerG
n1
S
1,n
= {0}
hold for all n 0.
From Corollary 2.6, we get that rankG
n
does not depend on the choice of the
projections onto kerA
n
and the isomorphisms between kerA
n
and kerA
n1
, hence we
can suppose that rankG
n
s, 1 s m 1. Here, Q
1,n
denotes a projection onto
kerG
n
and let T
1,n
be a nonsingular operator with the restriction T
1,n
|
kerGn
is an
isomorphism between kerG
n
and kerG
n1
. We also denote again by T
1,n
the matrix
induced by the operator T
1,n
.
Lemma 2.9. [7] The matrix G
1,n
:= G
n
+B
n
P
n1
T
1,n
Q
1,n
is nonsingular if and
only if
kerG
n1
S
1,n
= {0}.
Moreover, if G
1,n
is nonsingular then
Q
1,n1
:= T
1,n
Q
1,n
G
1
1,n
B
n
P
n1
is a projection from R
m
onto kerG
n1
along S
1,n
.
ELA
92 L.C. Loi
Recall that notations ker
G
n
and

S
1,n+1
have been introduced in Lemma 2.7. We
now come to the following lemma which states the relationship between projections
Q
1,n
and

Q
1,n
.
Lemma 2.10. Suppose that LSDE (1.1) is of index-2 and let

Q
1,n
be a projection
from R
m
onto ker
G
n
along

S
1,n+1
. Then the following relation holds:
Q
1,n
= (
P
n
+

T
1
n
T
n
Q
n
)
Q
1,n
. (2.5)
Proof. Putting
Q
1,n
:= (
P
n
+

T
1
n
T
n
Q
n
)
Q
1,n
(P
n
+T
1
n
T
n

Q
n
)
and noting that (P
n
+ T
1
n
T
n

Q
n
)(
P
n
+

T
1
n
T
n
Q
n
) = I,

Q
2
1,n
=

Q
1,n
, we obtain
Q
2
1,n
=

Q
1,n
, i.e.,

Q
1,n
is a projection.
Applying the relation (2.2) and observing that G
n

Q
1,n
= 0, we have
G
n

Q
1,n
= G
n

Q
1,n
(P
n
+T
1
n
T
n

Q
n
) = 0.
On the other hand, let x R
m
such that

Q
1,n
x = 0, or equivalently,

Q
1,n
(P
n
+
T
1
n
T
n

Q
n
)x = 0. Since

Q
1,n
is the projection onto kerG
n
along S
1,n+1
, it follows
(P
n
+T
1
n
T
n

Q
n
)x S
1,n+1
. This leads to x (
P
n
+

T
1
n
T
n
Q
n
)S
1,n+1
. Hence, using
the relation (2.3), we get x

S
1,n+1
. Thus,

Q
1,n
is a projection onto ker
G
n
along
S
1,n+1
meaning that
Q
1,n
= (
P
n
+

T
1
n
T
n
Q
n
)
Q
1,n
(P
n
+T
1
n
T
n

Q
n
).
Furthermore, observing that

Q
1,n
= T
1,n+1
Q
1,n+1
G
1
1,n+1
B
n+1
P
n
and
T
1
n
T
n

Q
n
= Q
n
T
1
n
T
n

Q
n
yields
Q
1,n
(P
n
+T
1
n
T
n

Q
n
) =

Q
1,n
.
Thus, we obtain Eq. (2.5).
From now on, we put P
1,n
:= I Q
1,n
and

P
1,n
:= I

Q
1,n
.
Lemma 2.11. [7] Suppose that the LSDE (1.1) is of index-2 and

G
1,n
:= G
n
+
B
n
P
n1
T
1,n

Q
1,n
. Then the following relations hold:
G
1
1,n
G
n
=

P
1,n
,

G
1
1,n
A
n
=

P
1,n
P
n
, (2.6)
G
1
1,n
B
n
=

G
1
1,n
B
n
P
n1

P
1,n1
+T
1
1,n
Q
1,n1
+

P
1,n
T
1
n
Q
n1
. (2.7)
ELA
Suppose that the LSDE (1.1) is of index-2. We also introduce an operator

T
1,n

GL(R
m
) whose restriction

T
1,n
|
ker
Gn
is an isomorphism between ker
G
n
and ker
G
n1
.
Put

G
1,n
:=

G
n
+ B
n

P
n1
T
1,n
Q
1,n
. A similar result of the relation (2.2) can be
established for index-2 LSDEs, namely, we obtain the following lemma.
Lemma 2.12. Let the LSDE (1.1) be of index-2. Then the identity
G
1,n
=

G
1,n
_
P
n
+T
1
n
T
n

Q
n
+T
1
n
Q
n1

P
n1
T
1,n
Q
1,n
(2.8)
+T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n

Q
1,n
_
is valid for each n 0.
Proof. Since T
1
n
Q
n1
= Q
n
T
1
n
Q
n1
and P
n1

P
n1
= P
n1
, we have that
G
n
T
1
n
Q
n1

P
n1
T
1,n
Q
1,n
= B
n

P
n1
T
1,n
Q
1,n
B
n
P
n1
T
1,n
Q
1,n
. (2.9)
Observing that

Q
1,n
Q
n
= 0, we come to the following identity
Q
1,n
(P
n
+T
1
n
T
n

Q
n
) =

Q
1,n
.
This gives
P
n
+T
1
n
T
n

Q
n

Q
1,n
=

P
1,n
(P
n
+T
1
n
T
n

Q
n
).
Thus, we obtain
B
n
P
n1
T
1,n

Q
1,n
(P
n
+T
1
n
T
n

Q
n

Q
1,n
) = 0. (2.10)
Using the relation (2.2), we can easily see that
T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n
=

Q
1,n
T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n
.
Therefore, we have
G
n
T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n
= 0. (2.11)
Further, since T
1
n1
T
n1

Q
n1
= Q
n1
T
1
n1
T
n1

Q
n1
, we get
B
n
P
n1
T
1,n

Q
1,n
T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n
= B
n
P
n1
T
1,n
Q
1,n
. (2.12)
Finally, combining the relation (2.2) with Eqs. (2.9)(2.12), and observing that
G
n

Q
1,n
= 0 and

Q
1,n
Q
n
= 0 implies that
G
1,n
_
P
n
+T
1
n
T
n

Q
n
+T
1
n
Q
n1

P
n1
T
1,n
Q
1,n
ELA
94 L.C. Loi
+T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n

Q
1,n
_
= (G
n
+B
n
P
n1
T
1,n

Q
1,n
)
_
P
n
+T
1
n
T
n

Q
n
+T
1
n
Q
n1

P
n1
T
1,n
Q
1,n
+T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n

Q
1,n
_
=

G
n
+B
n

P
n1
T
1,n
Q
1,n
B
n
P
n1
T
1,n
Q
1,n
+B
n
P
n1
T
1,n
Q
1,n
=

G
1,n
,
which is Eq. (2.8) as to be proved.
The following fact easily follows from Lemma 2.12.
Corollary 2.13. Suppose that the LSDE (1.1) is of index-2. Then the matrix
P
n
+T
1
n
T
n

Q
n
+T
1
n
Q
n1

P
n1
T
1,n
Q
1,n
+T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n
Q
1,n
is nonsingular. Moreover,
_
P
n
+T
1
n
T
n

Q
n
+ T
1
n
Q
n1

P
n1
T
1,n
Q
1,n
(2.13)
+ T
1
1,n
_
P
n1
+T
1
n1
T
n1

Q
n1
_
T
1,n
Q
1,n

Q
1,n
_
1
=

P
n
+

T
1
n
T
n
Q
n
+

T
1
n
Q
n1
P
n1
T
1,n

Q
1,n
+

T
1
1,n
_
P
n1
+

T
1
n1
T
n1
Q
n1
_
T
1,n

Q
1,n
Q
1,n
.
3. Multipoint boundary value problems.
3.1. Constant coecients case. We shall consider the LSDEs with constant
coecients
Ax
i+1
= Bx
i
+q
i
, i = 0, N 1, (3.1)
together with the boundary conditions
N
i=0
C
i
x
i
= , (3.2)
where A, B, C
i
R
mm
, q
i
, R
m
are given and suppose that :=ind(A, B) is
greater than one.
We suppose that C such that det(A + B) = 0. Multiply Eq. (3.1) by
(A +B)
1
from the left to obtain
x
i+1
=

B
x
i
+ q
i
, i = 0, N 1, (3.3)
ELA
where q
i
:= (A+B)
1
q
i
for all i = 0, N 1. According to Theorem 2.1, there exists
a nonsingular matrix T R
mm
such that
T
1

A
T =
_
C 0
0 U
_
, T
1

B
T =
_
I C 0
0 I U
_
, (3.4)
where C R
rr
is nonsingular with r := rank
and U R
(mr)(mr)
is nilpotent
of the order . Letting x
i
= Ty
i
and f
i
= T
1
q
i
, then we can rewrite Eq. (3.3) as
_
C 0
0 U
_
_
y
(1)
i+1
y
(2)
i+1
_
=
_
I C 0
0 I U
_
_
y
(1)
i
y
(2)
i
_
+
_
f
(1)
i
f
(2)
i
_
, i = 0, N 1,
where y
(1)
i
, f
(1)
i
R
r
, y
(2)
i
, f
(2)
i
R
mr
. Note that when = 1 then U = 0, and in
this case, we easily obtain solutions of the above dierence equation. The problem of
solving (3.1), (3.2) is not dicult, hence, it is omitted here due to lack of space. In
this paper, we consider the case 2, i.e., U = 0. However, it is easy to see that
these results are still valid for the case = 1. Since U has only the eigenvalue 0, it
yields that I U is nonsingular. Besides, noting that C is a nonsingular matrix, we
nd that all solutions of Eq. (3.1) are given by
x
i
=
_
A
D
_
i
A
D
x
0
+
_
B
D
_
Ni
(I

A
D
) x
N
(3.5)
+
i1
l=0
_

A
D
_
l
A
D
q
il1
(I

A
D
)
Ni1
l=0
_
B
D
_
l
B
D
q
i+l
, i = 0, N,
where x
0
, x
N
R
m
are arbitrary vectors. Here it is assumed that
1
l=0
= 0.
Notice that the formula (3.5) has also been established in [4]. Further, applying
Theorem 2.2 we see that the solution formula (3.5) is independent of the chosen value
.
Remark 3.1. An important special case is when A is nonsingular. To study
MPBVP (3.1), (3.2), instead of (3.5), we usually use the following solution formula
x
i
= (A
1
B)
i
x
0
+
i1
l=0
(A
1
B)
l
A
1
q
il1
, i = 0, N,
where x
0
R
m
is an arbitrary vector. In another important special case, when B is
invertible, the solution to (3.1) is given by
x
i
= (B
1
A)
Ni
x
N

Ni1
l=0
(B
1
A)
l
B
1
q
i+l
, i = 0, N,
ELA
96 L.C. Loi
where x
N
R
m
is an arbitrary vector. These results were discussed in the theory of
boundary value problems for ordinary dierence equations, we refer the reader to [1]
for more details. The purpose of this paper is to study the MPBVP (3.1), (3.2) in
the case, where A and B are both singular.
Let X
i
(i = 0, N) be the fundamental solution of Eq. (3.1), i.e.,
AX
i+1
= BX
i
, i = 0, N 1.
It is clear that
X
i
=
_

A
D
_
i
A
D
+
_
B
D
_
Ni
(I

A
D
), i = 0, N.
Put X
(1)
i
:=
_

A
D
_
i
A
D
, X
(2)
i
:=
_
B
D
_
Ni
(I

A
D
) (i = 0, N), D
1
:=
N
i=0
C
i
X
(1)
i
, D
2
:=
N
i=0
C
i
X
(2)
i
and
:=
N
i=0
C
i
z
i
, where
z
i
:=
i1
l=0
_
A
D
_
l
A
D
q
il1
(I

A
D
)
Ni1
l=0
_
B
D
_
l
B
D
q
i+l
, i = 0, N.
In what follows, we shall deal with the (m 2m) matrix (D
1
, D
2
) with columns of
D
1
and D
2
and the (2m2m) matrix
R :=
_

A
D
0
0 I

A
D
_
.
From Theorem 2.2 it follows that the matrices (D
1
, D
2
) and R do not depend on the
chosen value .
Theorem 3.2. Suppose that the matrix pencil (A, B) is regular and ind(A, B)
2. Then the MPBVP (3.1) and (3.2) has a unique solution for every q
i
R
m
(i =
0, N 1) and every R
m
if and only if
ker(D
1
, D
2
) = kerR (3.6)
and it can be represented as
x
i
= X
(1)
i
+X
(2)
i
+z
i
, i = 0, N, (3.7)
where (
T
,
T
)
T
= (D
1
, D
2
)
+
with (D
1
, D
2
)
+
the generalized inverse in Moore-
Penroses sense of (D
1
, D
2
).
Proof. Due to our construction, the relation
kerR ker(D
1
, D
2
)
ELA
is valid.
Assume that the MPBVP (3.1), (3.2) is uniquely solvable, and let ( x
T
0
, x
T
N
)
T
ker(D
1
, D
2
). Then
D
1
x
0
+D
2
x
N
= 0.
Putting x
i
:= X
(1)
i
x
0
+ X
(2)
i
x
N
(i = 0, N), we nd that {x
i
}
N
i=0
is a solution of the
homogeneous MPBVP (3.1), (3.2) with q
i
= 0, (i = 0, N 1) and = 0. Since the
homogeneous MPBVP (3.1) and (3.2) has only a trivial solution, it follows x
i
= 0 for
all i = 0, N. In particular, we have x
0
= 0 and x
N
= 0, hence,
A
D
x
0
+ (
B
D
)
N
(I

A
D
) x
N
= 0 (3.8)
and
(
A
D
)
N

A
D
x
0
+ (I

A
D
) x
N
= 0. (3.9)
From Eq. (3.4) and the facts that
A
D
= T
_
C
1
0
0 0
_
T
1
,

B
D
= T
_
(I C)
D
0
0 (I U)
1
_
T
1
,
it follows that
A
D
= T
_
I 0
0 0
_
T
1
, I

A
D
= T
_
0 0
0 I
_
T
1
, (3.10)
and
A
D
= T
_
C
1
(I C) 0
0 0
_
T
1
, (3.11)
B
D
= T
_
(I C)
D
C 0
0 (I U)
1
U
_
T
1
. (3.12)
Next, applying formulae (3.10)(3.12) and putting
( y
(1)
T
0
, y
(2)
T
0
)
T
:= T
1
x
0
, ( y
(1)
T
N
, y
(2)
T
N
)
T
:= T
1
x
N
with y
(1)
0
, y
(1)
N
R
r
, we can reduce the equalities (3.8), (3.9) to
_
y
(1)
0
= 0,
_
(I U)
1
U
_
N
y
(2)
N
= 0
and
_
_
C
1
(I C)
_
N
y
(1)
0
= 0,
y
(2)
N
= 0,
ELA
98 L.C. Loi
respectively. Thus, we obtain
x
0
= T
_
0
_
, x
N
= T
_

0
_
,
where R
mr
and R
r
are arbitrary vectors, or x
0
ker(
A
D
) and x
N

ker(I

A
D
), hence ( x
T
0
, x
T
N
)
T
kerR. This means that the inclusion
ker(D
1
, D
2
) kerR
must be true, and consequently, (3.6) holds.
Conversely, let (3.6) be valid. Then for each q
i
R
m
(i = 0, N 1) and R
m
a solution of the MPBVP (3.1), (3.2) is determined by (3.5) and
D
1
x
0
+D
2
x
N
=
.
Let q
i
= 0 for all i = 0, N 1 and = 0. Then x
0
and x
N
satisfy the following
equality
D
1
x
0
+D
2
x
N
= 0.
Therefore, we have ( x
T
0
, x
T
N
)
T
ker(D
1
, D
2
) = kerR. Now (3.5) ensures that the
homogeneous MPBVP (3.1), (3.2) has only a trivial solution.
According to the formula (3.5), any solution of (3.1) can be expressed as (3.7)
where , R
m
are constant vectors. This solution satises the boundary condition
(3.2) if and only if
D
1
+D
2
=
which means that (

T
,
T
)
T
= (D
1
, D
2
)
+
. Thus, the unique solution of (3.1), (3.2)

has the representation (3.7).
It is easy to see that dim(kerR) = m. Denote p := dim
_
ker(D
1
, D
2
)
_
. We now
consider a case, when (3.6) does not hold, i.e., p > m and the problem (3.1), (3.2)
has either no solution or an innite number of solutions. We denote by {w
0
i
}
m
i=1
certain base of kerR. Using the fact that kerR ker(D
1
, D
2
), we can extend {w
0
i
}
m
i=1
to a basis {w
0
i
}
p
i=1
of ker(D
1
, D
2
). Let u
0
i
, v
0
i
R
m
be the rst and the second
groups of components of w
0
i
, i.e., w
0
i
= (u
0
T
i
, v
0
T
i
)
T
, (i = 1, p). We construct the
column matrices
i
:= X
(1)
i
U + X
(2)
i
V (i = 0, N), where U := (u
0
m+1
, . . . , u
0
p
), V :=
(v
0
m+1
, . . . , v
0
p
) R
m(pm)
. To represent solutions of the MPVBP (3.1), (3.2) we
introduce a linear operator L acting in R
m(N+1)
, dened by
L(x
T
0
, . . . , x
T
N
)
T
:=
_
(Ax
1
Bx
0
)
T
, . . . , (Ax
N
Bx
N1
)
T
,
_
N
i=0
C
i
x
i
_
T
_
T
.
ELA
Lemma 3.3. kerL =
__
(
0
a)
T
, . . . , (
N
a)
T
_
T
: a R
pm
_
.
Proof. Suppose that x = (x
T
0
, . . . , x
T
N
)
T
__
(
0
a)
T
, . . . , (
N
a)
T
_
T
: a R
pm
_
,
i.e., there exists a vector a R
pm
such that x
i
=
i
a, i = 0, N. This leads to
x
i
= X
(1)
i
Ua +X
(2)
i
Va. From Eqs. (3.10)(3.12), by simple computations we nd
AX
(1)
i+1
= BX
(1)
i
and AX
(2)
i+1
= BX
(2)
i
, i = 0, N 1.
Using the above equations, we have
Lx :=
_
(Ax
1
Bx
0
)
T
, . . . , (Ax
N
Bx
N1
)
T
_
T
=
_
_
(AX
(1)
1
BX
(1)
0
)Ua
_
T
, . . . ,
_
(AX
(1)
N
BX
(1)
N1
)Ua
_
T
_
T
+
_
_
(AX
(2)
1
BX
(2)
0
)Va
_
T
, . . . ,
_
(AX
(2)
N
BX
(2)
N1
)Va
_
T
_
T
= 0.
Denote x :=
N
i=0
C
i
x
i
= D
1
Ua + D
2
Va. Since U and V are column matrices whose
columns are u
0
i
, v
0
i
and (u
0
T
i
, v
0
T
i
)
T
ker(D
1
, D
2
) (i = m + 1, p), it gives that D
1
U +
D
2
V = 0, which immediately implies x = 0. Thus, we obtain Lx = 0, which means
that
__
(
0
a)
T
, . . . , (
N
a)
T
_
T
: a R
pm
_
kerL.
Conversely, assume that x = (x
T
0
, . . . , x
T
N
)
T
kerL, i.e,
_
_
_
Ax
i+1
= Bx
i
, i = 0, N 1,
N
i=0
C
i
x
i
= 0.
Due to the formula (3.5), x
i
= X
(1)
i
+ X
(2)
i
(i = 0, N), where vectors , R
m
satisfy the relation D
1
+ D
2
= 0, hence we have (
T
,
T
)
T
ker(D
1
, D
2
). Since
(u
0
T
k
, v
0
T
k
)
T
(k = 1, p) is the basis of ker(D
1
, D
2
), there exists a sequence {
k
}
p
k=1
such
that (
T
,
T
)
T
=
p
k=1
k
(u
0
T
k
, v
0
T
k
)
T
, hence =
p
k=1
k
u
0
k
and =
p
k=1
k
v
0
k
. Thus,
x
i
=
m
k=1
k
_
X
(1)
i
u
0
k
+X
(2)
i
v
0
k
_
+
p
k=m+1
k
_
X
(1)
i
u
0
k
+X
(2)
i
v
0
k
_
, i = 0, N.
Observing that (u
0
T
k
, v
0
T
k
)
T
kerR, i.e.,

A
D
u
0
k
= 0 and (I

A
D
)v
0
k
= 0 for all
k = 1, m, we nd X
(1)
i
u
0
k
= 0 and X
(2)
i
v
0
k
= 0 for all k = 1, m, i = 0, N. Thus,
x
i
= X
(1)
i
p
k=m+1
k
u
0
k
+X
(2)
i
p
k=m+1
k
v
0
k
, i = 0, N.
ELA
100 L.C. Loi
Taking a := (
m+1
, . . . ,
p
)
T
R
pm
, we get x
i
= X
(1)
i
Ua + X
(2)
i
Va (i = 0, N), i.e.,
x
i
=
i
a for all i = 0, N where a R
pm
. Thus, we obtain
x
__
(
0
a)
T
, . . . , (
N
a)
T
_
T
: a R
pm
_
,
or
kerL
__
(
0
a)
T
, . . . , (
N
a)
T
_
T
: a R
pm
_
.
Next, we let q := dim
_
ker(D
1
, D
2
)
T
_
and denote by {w
i
}
q
i=1
certain base of
ker(D
1
, D
2
)
T
. Letting W R
qm
be a row matrix whose rows are vectors w
i
(i =
1, q), we come to the following theorem.
Theorem 3.4. Let the matrix pencil (A, B) be regular and ind(A, B) 2. Then,
the problem (3.1), (3.2) is solvable if and only if
W
= 0. (3.13)
Moreover, a general solution of (3.1), (3.2) has the following form
x
i
= X
(1)
i
+ X
(2)
i
+z
i
+
i
a, i = 0, N, (3.14)
where a R
pm
is an arbitrary vector and (
T
,
T
)
T
= (D
1
, D
2
)
+
with (D
1
, D
2
)
+
the generalized inverse in Moore-Penroses sense of (D
1
, D
2
).
Proof. The problem (3.1), (3.2) is solvable if and only if
(q
T
0
, . . . , q
T
N1
,
T
)
T
imL,
i.e., there exists x = (x
T
0
, . . . , x
T
N
)
T
R
m(N+1)
satisfying Lx = (q
T
o
, . . . , q
T
N1
,
T
)
T
.
Equivalently, there exist vectors , R
m
such that x
i
= X
(1)
i
+X
(2)
i
+z
i
(i = 0, N)
and
N
i=0
C
i
x
i
= . Thus, the system (3.1), (3.2) possesses a solution if and only
if there exist vectors , R
m
such that D
1
+ D
2
=
. Using the fact that

im(D
1
, D
2
) =
_
ker(D
1
, D
2
)
T
)
we come to the conclusion that the MPBVP (3.1),

(3.2) is solvable if and only if
_
ker(D
1
, D
2
)
T
)
. Thus, the problem (3.1), (3.2)

possesses a solution if and only if (3.13) is valid.
Finally, thanks to Lemma 3.3 and the formula (3.5), to show that (3.14) is a
general solution formula of the problem (3.1), (3.2) we only need to prove that x
i
is given by x
i
= X
(1)
i
+ X
(2)
i
+ z
i
(i = 0, N) with (
T
,
T
)
T
= (D
1
, D
2
)
+
, is a
particular solution of the above mentioned problem.
Theorem 3.2 and Theorem 3.4 imply the following corollary.
Corollary 3.5. (Fredholm alternative) Suppose that the matrix pencil (A, B)
is regular, ind(A, B) 2 and let p := dim
_
ker(D
1
, D
2
)
_
. Then
ELA
(i) either p = m and the MPBVP (3.1), (3.2) is uniquely solvable for any data
q
i
(i = 0, N 1) and ;
(ii) or p > m and the MPBVP (3.1, (3.2) is solvable if and only if the condition
(3.13) is valid.
Moreover, the solution formula (3.14) holds.
3.2. Varying coecients case. In this subsection, we shall deal with the MP-
BVPs for the non-autonomous LSDEs as follows
A
i
x
i+1
= B
i
x
i
+q
i
, i = 0, N 1, (3.15)
N
i=0
C
i
x
i
= , (3.16)
where A
i
, B
i
, C
i
R
mm
, q
i
, R
m
are given and suppose that the LSDE (3.15) is
of index-2 in the sense that the following relations hold:
(i) rankA
i
r, 1 r m1,
(ii) dim(kerA
i1
S
i
) ms, 1 s m1,
(iii) kerG
i1
S
1,i
= {0}
for all i = 0, N 1. Further, here it is assumed that A
1
:= A
0
, G
1
:= G
0
and
Q
1,N1
is projection onto kerG
N1
such that

Q
1,N1
Q
N1
= 0.
Now, we describe shortly the decomposition technique for index-2 LSDEs (see [7]
for details). We decompose the index-2 LSDE solution x
i
into
x
i
= Q
i1
x
i
+P
i1

P
1,i1
x
i
+P
i1

Q
1,i1
x
i
=: w
i
+u
i
+P
i1
v
i
.
Multiplying Eq. (3.15) by P
i

P
1,i

G
1
1,i
, Q
i
P
1,i

G
1
1,i
, and

Q
1,i

G
1
1,i
, respectively, using
the relations (2.6), (2.7) and the fact that

Q
1,i
Q
i
= 0, and carrying out some technical
computations, we decouple the index-2 LSDE (3.15) into the system
_
_
u
i+1
= P
i

P
1,i

G
1
1,i
B
i
u
i
+P
i

P
1,i

G
1
1,i
q
i
,
Q
i
v
i+1
= Q
i
P
1,i

G
1
1,i
B
i
u
i
+T
1
i
w
i
+Q
i

P
1,i

G
1
1,i
q
i
,
0 = T
1
1,i
v
i
+

Q
1,i
G
1
1,i
q
i
,
i = 0, N 1.
Thus, we obtain
v
i
= T
1,i

Q
1,i

G
1
1,i
q
i
, i = 0, N 1,
w
i
= T
i
Q
i

P
1,i

G
1
1,i
B
i
u
i
T
i
Q
i
v
i+1
T
i
Q
i
P
1,i

G
1
1,i
q
i
, i = 0, N 1,
u
i+1
= P
i

P
1,i

G
1
1,i
B
i
u
i
+P
i

P
1,i

G
1
1,i
q
i
, i = 0, N 1.
We denote
i
:=
_
I T
i
Q
i

P
1,i

G
1
1,i
B
i
_
P
i1

P
1,i1
, i = 0, N 1
ELA
102 L.C. Loi
and
M
(i)
k
:=
k
l=0
P
il1

P
1,il1

G
1
1,il1
B
il1
, i = 1, N, k = 1, i 1,
where it is assumed that
1
l=0
= I. Observing that (P
i1

P
1,i1
)
2
= P
i1

P
1,i1
, we get
the solution of the LSDE (3.15) as follows
x
i
=
i
_
M
(i)
i1
x
0
+
i1
k=0
M
(i)
ik2
P
k

P
1,k

G
1
1,k
q
k
_
(3.17)
+T
i
Q
i
T
1,i+1

Q
1,i+1

G
1
1,i+1
q
i+1
_
T
i
Q
i

P
1,i

G
1
1,i
+P
i1
T
1,i

Q
1,i

G
1
1,i
_
q
i
, i = 0, N 2,
x
N1
=
N1
_
M
(N1)
N2
x
0
+
N2
k=0
M
(N1)
Nk3
P
k

P
1,k

G
1
1,k
q
k
_
(3.18)
T
N1
Q
N1

Q
1,N1
x
N
_
T
N1
Q
N1

P
1,N1

G
1
1,N1
+P
N2
T
1,N1

Q
1,N1

G
1
1,N1
_
q
N1
and
x
N
= M
(N)
N1
x
0
+
N1
k=0
M
(N)
Nk2
P
k

P
1,k

G
1
1,k
q
k
+Q
N1
x
N
+P
N1

Q
1,N1
x
N
, (3.19)
where x
0
, x
N
R
m
are arbitrary vectors and it is assumed that
1
k=0
= 0.
Remark 3.6. Similar to Remark 3.1, if A
i
(resp., B
i
) is nonsingular for each
i = 0, N 1 then we will use the corresponding solution formulae for (3.15), (3.16)
x
i
=
_
i1
l=0
A
1
il1
B
il1
_
x
0
+
i1
k=0
_
ik2
l=0
A
1
il1
B
il1
_
A
1
k
q
k
, i = 0, N,
or
x
i
=
_
N1
l=i
B
1
l
A
l
_
x
N

N1
k=i
_
k1
l=i
B
1
l
A
l
_
B
1
k
q
k
, i = 0, N,
where x
0
, x
N
R
m
are arbitrary vectors instead of the formulae (3.17)(3.19). See
[1] for details. The formulae (3.17)(3.19) are useful in the case, where A
i
and B
i
are
ELA
both singular, and Eq. (3.15) is of index-2. Further, it is clear that these formulae
are an extension of the above mentioned formulae.
Lemma 3.7. Let the LSDE (3.15) be of index-2. Then the matrices
i
,
i
M
(i)
i1
,
i
M
(i)
ik2
P
k

P
1,k

G
1
1,k
, T
i
Q
i
T
1,i+1

Q
1,i+1

G
1
1,i+1
,
(T
i
Q
i
P
1,i
+P
i1
T
1,i

Q
1,i
)
G
1
1,i
are independent of the choice of the T
i
, Q
i
and T
1,i
.
Proof. Let

T
i
be another transformation, whose restriction

T
i
kerAi
is an isomor-
phism from kerA
i
onto kerA
i1
and

Q
i
be another projection onto kerA
i
,

P
i
:= I
Q
i
.
We denote

G
i
:= A
i
+ B
i
T
i

Q
i
. Let

Q
1,i
be a projection onto ker
G
i
along S
1,i+1
,
P
1,i
:= I

Q
1,i
,

T
1,i
ker
Gi
denote an isomorphism from ker
G
i
onto ker
G
i1
and put
G
1,i
:=

G
i
+B
i

P
i1
T
1,i
Q
1,i
,

i
:=
_
I

T
i

Q
i
P
1,i
G
1
1,i
B
i
_
P
i1
P
1,i1
.
First, we put
Z
i
:=

P
i
+
T
1
i
T
i
Q
i
+
T
1
i
Q
i1
P
i1
T
1,i
Q
1,i
+
T
1
1,i
_
P
i1
+
T
1
i1
T
i1
Q
i1
_
T
1,i

Q
1,i
Q
1,i
.
From the identities (2.8) and (2.13), we have
T
i

Q
i
P
1,i
G
1
1,i
=

T
i

Q
i
P
1,i
Z
i

G
1
1,i
. (3.20)
Using the facts that

Q
1,i
Q
i
= 0,

T
1
i
T
i
Q
i
=

Q
i
T
1
i
T
i
Q
i
and
T
1
1,i
_
P
i1
+

T
1
i1
T
i1
Q
i1
_
T
1,i

Q
1,i
=

Q
1,i
T
1
1,i
_
P
i1
+

T
1
i1
T
i1
Q
i1
_
T
1,i
Q
1,i
,
we see that
T
i

Q
i
P
1,i

P
i
=
T
i

Q
i
Q
1,i
P
i
=
T
i
Q
i
Q
1,i
,
T
i
Q
i
P
1,i
T
1
i
T
i
Q
i
=

T
i
Q
i
(I

Q
1,i
)
Q
i
T
1
i
T
i
Q
i
= T
i
Q
i
,
T
i

Q
i
P
1,i
T
1
i
Q
i1
P
i1
T
1,i

Q
1,i
=

T
i

Q
i
(I

Q
1,i
)
Q
i
T
1
i
Q
i1
P
i1
T
1,i

Q
1,i
=

Q
i1
P
i1
T
1,i

Q
1,i
and
T
i

Q
i
P
1,i
T
1
1,i
_
P
i1
+

T
1
i1
T
i1
Q
i1
_
T
1,i

Q
1,i
= 0.
ELA
104 L.C. Loi
Combining the above relations with Eq. (2.5), it follows that
T
i

Q
i
P
1,i
Z
i
=

T
i
Q
i
Q
1,i
+T
i
Q
i
+

Q
i1
P
i1
T
1,i

Q
1,i
(3.21)
=

T
i
Q
i
(
P
i
+

T
1
i
T
i
Q
i
)
Q
1,i
+T
i
Q
i
+

Q
i1
P
i1
T
1,i
Q
1,i
= T
i
Q
i
P
1,i
+

Q
i1
P
i1
T
1,i
Q
1,i
.
Thus,
i
=
_
I T
i
Q
i
P
1,i

G
1
1,i
B
i

Q
i1
P
i1
T
1,i

Q
1,i

G
1
1,i
B
i
_
P
i1
P
1,i1
.
Observe that
P
i1
P
1,i1
=

P
i1

P
i1
(
P
i1
+

T
1
i1
T
i1
Q
i1
)
Q
1,i1
=

P
i1

P
1,i1
and
Q
1,i1

P
i1

P
1,i1
=

Q
1,i1
(I

Q
i1
)
P
1,i1
= 0.
Further, we note that
G
1
1,i
B
i
Q
i1
=

G
1
1,i
(A
i
+B
i
T
i
Q
i
)Q
i
T
1
i
Q
i1
=

P
1,i
T
1
i
Q
i1
,
implying that
Q
1,i1
:= T
1,i

Q
1,i
G
1
1,i
B
i
P
i1
= T
1,i

Q
1,i
G
1
1,i
B
i
T
1,i

Q
1,i

G
1
1,i
B
i
Q
i1
= T
1,i

Q
1,i
G
1
1,i
B
i
.
This leads to
i
=
_
IT
i
Q
i

P
1,i

G
1
1,i
B
i
_
P
i1
P
i1

P
1,i1
=
i
_
IT
i
Q
i
P
1,i

G
1
1,i
B
i
_
Q
i1
P
i1

P
1,i1
.
Since
_
I T
i
Q
i
P
1,i

G
1
1,i
B
i
_
Q
i1
=

Q
i1
T
i
Q
i
P
1,i

G
1
1,i
B
i
Q
i1

Q
i1
=

Q
i1
T
i
Q
i
P
1,i

P
1,i
T
1
i
Q
i1

Q
i1
=

Q
i1
T
i
Q
i
(I

Q
1,i
)Q
i
T
1
i
Q
i1
= 0,
we have

i
=
i
.
Applying the identities (2.8), (2.13) we get
P
i1
P
1,i1
G
1
1,i1
=

P
i1
P
1,i1
Z
i1

G
1
1,i1
.
ELA
On the other hand, since
P
i1
P
1,i1

P
i1
=

P
i1

P
i1
Q
1,i1
(I

Q
i1
) =

P
i1
P
1,i1
=

P
i1

P
1,i1
,
P
i1
P
1,i1
T
1
i1
T
i1
Q
i1
=

P
i1
(I

Q
1,i1
)
Q
i1
T
1
i1
T
i1
Q
i1
= 0,
P
i1
P
1,i1
T
1
i1
Q
i2
P
i2
T
1,i1

Q
1,i1
=

P
i1
(I

Q
1,i1
)
Q
i1
T
1
i1
Q
i2
P
i2
T
1,i1

Q
1,i1
= 0
and
P
i1
P
1,i1
T
1
1,i1
(
P
i2
+

T
1
i2
T
i2
Q
i2
)T
1,i1

Q
1,i1
=

P
i1
P
1,i1
Q
1,i1
T
1
1,i1
(
P
i2
+

T
1
i2
T
i2
Q
i2
)T
1,i1

Q
1,i1
= 0,
we have that
P
i1
P
1,i1
G
1
1,i1
=

P
i1

P
1,i1

G
1
1,i1
= P
i1

P
1,i1

G
1
1,i1

Q
i1
P
i1

P
1,i1

G
1
1,i1
.
Observing that
i

Q
i1
=
_
I T
i
Q
i

P
1,i

G
1
1,i
B
i
_
P
i1

P
1,i1
Q
i1

Q
i1
= 0, we obtain
i

P
i1
P
1,i1
G
1
1,i1
=
i
P
i1

P
1,i1

G
1
1,i1

i

Q
i1
P
i1

P
1,i1

G
1
1,i1
=
i
P
i1

P
1,i1

G
1
1,i1
,
hence
i

P
i1
P
1,i1
G
1
1,i1
=
i
P
i1

P
1,i1

G
1
1,i1
.
Since

G
1
1,i1
B
i1
Q
i2
=

P
1,i1
T
1
i1
Q
i2
and P
i1

P
1,i1
Q
i1
= 0, we have
i
P
i1

P
1,i1

G
1
1,i1
B
i1

Q
i2
P
i2

P
1,i2

G
1
1,i2
=
i
P
i1

P
1,i1

G
1
1,i1
B
i1
Q
i2

Q
i2
P
i2

P
1,i2

G
1
1,i2
=
i
P
i1

P
1,i1

P
1,i1
T
1
i1
Q
i2

Q
i2
P
i2

P
1,i2

G
1
1,i2
=
i
P
i1

P
1,i1
Q
i1
T
1
i1
Q
i2
P
i2

P
1,i2

G
1
1,i2
= 0.
Thus,
i

P
i1
P
1,i1
G
1
1,i1
B
i1

P
i2
P
1,i2
G
1
1,i2
=
i
P
i1

P
1,i1

G
1
1,i1
B
i1
P
i2

P
1,i2

G
1
1,i2
.
ELA
106 L.C. Loi
Thus, the matrices
i
M
(i)
i1
,
i
M
(i)
ik2
P
k

P
1,k

G
1
1,k
do not depend on the choice of the
T
i
, Q
i
and T
1,i
, as it was to be proved.
From Eqs. (2.8), (2.13) it follows that
T
i

Q
i
T
1,i+1
Q
1,i+1
G
1
1,i+1
=

T
i

Q
i
T
1,i+1
Q
1,i+1
Z
i+1

G
1
1,i+1
.
Besides,

Q
1,i+1
=

Q
1,i+1
(
P
i+1
+

T
1
i+1
T
i+1
Q
i+1
), therefore, we get
T
i
Q
i
T
1,i+1
Q
1,i+1
Z
i+1

G
1
1,i+1
=

T
i

Q
i
T
1,i+1
Q
1,i+1
_
P
1,i+1
(
P
i+1
+

T
1
i+1
T
i+1
Q
i+1
)
+
T
1
i+1
Q
i
P
i
T
1,i+1

Q
1,i+1
+

T
1
1,i+1
(
P
i
+

T
1
i
T
i
Q
i
)T
1,i+1

Q
1,i+1
_
G
1
1,i+1
= T
i
Q
i
T
1,i+1

Q
1,i+1

G
1
1,i+1
,
or equivalently,
T
i

Q
i
T
1,i+1
Q
1,i+1
G
1
1,i+1
= T
i
Q
i
T
1,i+1

Q
1,i+1

G
1
1,i+1
.
From Eqs. (3.20)(3.21), it follows that
T
i
Q
i
P
1,i
G
1
1,i
= T
i
Q
i
P
1,i

G
1
1,i
+

Q
i1
P
i1
T
1,i
Q
1,i
G
1
1,i
,
further,
P
i1
T
1,i
Q
1,i
G
1
1,i
=

P
i1
T
1,i
Q
1,i
_
P
1,i
(
P
i
+

T
1
i
T
i
Q
i
) +

Q
i
T
1
i
Q
i1
P
i1
T
1,i
Q
1,i
+
T
1
1,i
(
P
i1
+

T
1
i1
T
i1
Q
i1
)T
1,i

Q
1,i
_
G
1
1,i
=

P
i1
T
1,i

Q
1,i

G
1
1,i
,
giving
T
i

Q
i
P
1,i
G
1
1,i
+

P
i1
T
1,i
Q
1,i
G
1
1,i
= T
i
Q
i
P
1,i

G
1
1,i
+

Q
i1
P
i1
T
1,i
Q
1,i
G
1
1,i
+
P
i1
T
1,i

Q
1,i

G
1
1,i
= T
i
Q
i
P
1,i

G
1
1,i
+
_
(I

P
i1
)P
i1
+

P
i1
_
T
1,i
Q
1,i

G
1
1,i
= T
i
Q
i
P
1,i

G
1
1,i
+P
i1
T
1,i
Q
1,i
G
1
1,i
.
Thus, we obtain
_
T
i
Q
i
P
1,i
+

P
i1
T
1,i
Q
1,i
_
G
1
1,i
=
_
T
i
Q
i

P
1,i
+P
i1
T
1,i
Q
1,i
_
G
1
1,i
.
From the formulae (3.17)(3.19), it follows that the fundamental solution of
equations
A
i
X
i+1
= B
i
X
i
, i = 0, N 1
ELA
can be determined as X
i
:=
i
M
(i)
i1
, i = 0, N 1 and X
N
:= M
(N)
N1
. We dene
R := diag(P
1

P
1,1
, I P
N1

P
1,N1
) and the matrix (D
1
, D
2
), whose columns are
the columns of the matrices D
1
:=
N
i=0
C
i
X
i
and
D
2
:= C
N1
T
N1
Q
N1

Q
1,N1
+C
N
Q
N1
+C
N
P
N1

Q
1,N1
.
Lemma 3.8. Suppose that the LSDE (3.15) is of index-2. Then the following
condition
ker(D
1
, D
2
) = kerR (3.22)
does not depend on the chosen T
i
, Q
i
and T
1,i
.
Proof. Assume that

Q
i
is another projection onto kerA
i
and

T
i
(resp.,

T
1,i
) is
another transformation with

T
i
|
kerAi
(resp.,

T
1,i
|
ker
Gi
) being an isomorphism from
kerA
i
onto kerA
i1
(resp., ker
G
i
onto ker
G
i1
). Here, the matrices

P
i
,

G
i
,

Q
1,i
,

P
1,i
,
G
1,i
and

i
are dened in the proof of Lemma 3.7. We put
D
1
:=
N
i=0
C
i

X
i
,

X
i
:=

M
(i)
i1
, i = 0, N 1,

X
N
:=

M
(N)
N1
,
D
2
:= C
N1
T
N1

Q
N1
Q
1,N1
+C
N

Q
N1
+C
N

P
N1
Q
1,N1
and
R := diag(
P
1
P
1,1
, I

P
N1
P
1,N1
),
where

M
(i)
i1
:=
i1
l=0
P
il1
P
1,il1
G
1
1,il1
B
il1
(i = 0, N).
Lemma 3.7 ensures that X
i
=

X
i
for all i = 0, N 1. Besides, using Eq. (2.7)
and the facts that

P
N1
P
1,N1
T
1
N1
Q
N2
= 0,

P
N2
P
1,N2
N1
=

P
N2
P
1,N2
we have
M
(N)
N1
=

P
N1
P
1,N1
G
1
1,N1
B
N1
N1
M
(N1)
N2
.
Applying Lemma 3.7 again and noting that
P
N1
P
1,N1
G
1
1,N1
= P
N1

P
1,N1

G
1
1,N1

Q
N1
P
N1

P
1,N1

G
1
1,N1
,
we obtain
M
(N)
N1
= M
(N)
N1

Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
N1
M
(N1)
N2
,
ELA
108 L.C. Loi
or
X
N
= X
N

Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
X
N1
.
This implies that
D
1
= D
1
C
N

Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
X
N1
. (3.23)
According to Eq. (2.5), we get
T
N1

Q
N1
Q
1,N1
= T
N1
Q
N1

Q
1,N1
and

P
N1
Q
1,N1
=

P
N1

Q
1,N1
.
Therefore,
D
2
= D
2
+C
N
(
Q
N1
Q
N1
) +C
N
(
P
N1
P
N1
)
Q
1,N1
,
or equivalently,
D
2
= D
2
+C
N
(
Q
N1
Q
N1
)
P
1,N1
. (3.24)
Now suppose that Eq. (3.22) is valid. Let ( x
T
0
, x
T
N
)
T
ker(
D
1
,

D
2
) then
D
1
x
0
+

D
2
x
N
= 0.
From Eqs. (3.23)(3.24), the above equation can be rewritten as follows
D
1
x
0
+D
2
x
N
C
N

Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
X
N1
x
0
+C
N
(
Q
N1
Q
N1
)
P
1,N1
x
N
= 0.
Put
:=
Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
X
N1
x
0
and
:= (
Q
N1
Q
N1
)
P
1,N1
x
N
.
Using the facts that

Q
N1
= Q
N1

Q
N1
and

Q
1,N1
Q
N1
= 0, we have
D
2
( +) = C
N

Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
X
N1
x
0
+C
N
(
Q
N1
Q
N1
)
P
1,N1
x
N
.
This gives
D
1
x
0
+D
2
( x
N
+ +) = 0.
ELA
This means that
_
x
T
0
, ( x
N
+ + )
T
_
T
ker(D
1
, D
2
) = kerR. It ensures that
P
1

P
1,1
x
0
= 0 and (I P
N1

P
1,N1
)( x
N
+ +) = 0. Now applying Eq. (2.5) we
come to the conclusion that

P
1
P
1,1
=

P
1

P
1,1
. Further,

P
1
=

P
1
P
1
, hence
P
1
P
1,1
=

P
1
P
1

P
1,1
. This implies that
P
1
P
1,1
x
0
= 0. (3.25)
On the other hand, since P
1

P
1,1
x
0
= 0 and X
N1
= X
N1
P
1

P
1,1
, it follows that
X
N1
x
0
= 0. Thus, = 0 and we obtain (I P
N1

P
1,N1
)( x
N
+) = 0. Observing
that = Q
N1
and P
N1

P
1,N1
Q
N1
= 0, we get
x
N
+Q
N1
= P
N1

P
1,N1
x
N
.
This relation leads to that

Q
1,N1
x
N
= 0, hence = (
Q
N1
Q
N1
) x
N
. This implies
that
Q
N1
x
N
+Q
N1
(
Q
N1
Q
N1
) x
N
= 0
or we have

Q
N1
x
N
= 0. Thus,
Q
N1
x
N
+

P
N1

Q
1,N1
x
N
= 0.
This means that
Q
N1
x
N
+

P
N1
Q
1,N1
x
N
= 0.
The last equation is equivalent to
(I

P
N1
P
1,N1
) x
N
= 0. (3.26)
Combining Eqs. (3.25)(3.26), we come to the conclusion that ( x
T
0
, x
T
N
)
T
ker
R.
Thus, the inclusion ker(
D
1
,

D
2
) ker
R is proved. To show the converse inclu-

sion, we observe that for arbitrary ( x
T
0
, x
T
N
)
T
ker
R, i.e,

P
1
P
1,1
x
0
= 0 and
(I

P
N1
P
1,N1
) x
N
= 0. Due to

X
i
=

X
i
P
1
P
1,1
(i = 0, N), it implies that
D
1
x
0
= 0. Notice that the equality (I

P
N1
P
1,N1
) x
N
= 0 is equivalent to the
following relation
Q
N1
x
N
+

P
N1
Q
1,N1
x
N
= 0.
Moreover, since

Q
1,N1

P
N1
P
1,N1
= 0 it follows

Q
1,N1
x
N
= 0. Therefore, we
obtain

D
2
x
N
= 0. It implies that

D
1
x
0
+

D
2
x
N
= 0 or ( x
T
0
, x
T
N
)
T
ker(
D
1
,

D
2
).
ELA
110 L.C. Loi
We denote by
z
i
:=
i
i1
k=0
M
(i)
ik2
P
k

P
1,k

G
1
1,k
q
k
+T
i
Q
i
T
1,i+1

Q
1,i+1

G
1
1,i+1
q
i+1
T
i
Q
i
P
1,i

G
1
1,i
q
i
P
i1
T
1,i

Q
1,i

G
1
1,i
q
i
, i = 0, N 2,
z
N1
:=
N1
N2
k=0
M
(N1)
Nk3
P
k

P
1,k

G
1
1,k
q
k
T
N1
Q
N1
P
1,N1

G
1
1,N1
q
N1
P
N2
T
1,N1

Q
1,N1

G
1
1,N1
q
N1
,
z
N
:=
N1
k=0
M
(N)
Nk2
P
k

P
1,k

G
1
1,k
q
k
and
:=
N
i=0
C
i
z
i
.
Note that Lemma 3.8 guarantees that the following theorem does not depend on the
chosen T
i
, Q
i
and T
1,i
.
Theorem 3.9. Let the LSDE (3.15) be of index-2. Then the MPBVP (3.15),
(3.16) is uniquely solvable for every q
i
R
m
(i = 0, N 1) and every R
m
if and
only if the condition (3.22) holds. Moreover, the unique solution can be represented
as
_
_
x
i
= X
i
+z
i
, i = 0, N 2,
x
N1
= X
N1
+z
N1
T
N1
Q
N1

Q
1,N1
,
x
N
= X
N
+z
N
+Q
N1
+P
N1

Q
1,N1
,
(3.27)
where (
T
,
T
)
T
= (D
1
, D
2
)
+
with (D
1
, D
2
)
+
the generalized inverse in Moore-
Penroses sense of (D
1
, D
2
).
Proof. The proof of Theorem 3.9 is quite similar to that of Theorem 3.2, hence
it will be outlined only.
First, observe that the equations corresponding to (3.8) and (3.9) are
(I T
0
Q
0

P
1,0

G
1,0
B
0
)P
1

P
1,1
x
0
= 0 (3.28)
and
X
N
x
0
+Q
N1
x
N
+P
N1

Q
1,N1
x
N
= 0. (3.29)
Using the facts that P
1

P
1,1
= P
0

P
1,0
, T
0
= I and P
0
Q
0
= 0, we obtain that Eq.
(3.28) yields
P
0

P
1,0
x
0
= 0.
ELA
Moreover, X
N
x
0
= X
N
P
0

P
1,0
x
0
= 0, hence, Eq. (3.29) implies that
Q
N1
x
N
+ P
N1

Q
1,N1
x
N
= 0,
or equivalently,
(I P
N1

P
1,N1
) x
N
= 0.
In the proof of the converse part, we note that the solution formula (3.5) is
replaced with the formulae (3.17)(3.19). Since X
i
= X
i
P
1

P
1,1
, i = 0, N and
P
1

P
1,1
x
0
= 0, it gives X
i
x
0
= 0 (i = 0, N). Using the fact that P
N1
Q
N1
=
0, we can conclude that the equality Q
N1
x
N
+ P
N1

Q
1,N1
x
N
= 0 implies that
Q
N1
x
N
= 0 and P
N1

Q
1,N1
x
N
= 0. Besides, since
x
N
= Q
N1
x
N
+P
N1

Q
1,N1
x
N
+P
N1

P
1,N1
x
N
,
we get x
N
= P
N1

P
1,N1
x
N
. This leads to
Q
1,N1
x
N
=

Q
1,N1
P
N1

P
1,N1
x
N
=

Q
1,N1
(I Q
N1
)
P
1,N1
x
N
= 0.
Next, we have the following useful lemma.
Lemma 3.10. The dimensions of kerR and ker(D
1
, D
2
) are independent of the
choice of T
i
, Q
i
and T
1,i
, moreover, dim(kerR) = m and dim
_
ker(D
1
, D
2
)
_
=: p
m.
Proof. Firstly, we observe that
ker(P
i

P
1,i
) = kerA
i
kerG
i
, i = 1, N 1. (3.30)
Indeed, let ker(P
i
P
1,i
) means that P
i

P
1,i
= 0. We write as =

P
1,i
+

Q
1,i
.
Clearly,

Q
1,i
kerG
i
. Furthermore, since P
i

P
1,i
= 0, it implies that

P
1,i
=
(P
i
+Q
i
)
P
1,i
= Q
i
P
1,i
kerA
i
. Thus, we get
ker(P
i

P
1,i
) kerA
i
+ kerG
i
, i = 1, N 1.
Conversely, for arbitrary = y+z kerA
i
+ kerG
i
, we see that P
i

P
1,i
= P
i

P
1,i
Q
i
y+
P
i

P
1,i

Q
1,i
z = 0. Therefore,
kerA
i
+ kerG
i
ker(P
i

P
1,i
), i = 1, N 1,
which yields
ker(P
i

P
1,i
) = kerA
i
+ kerG
i
, i = 1, N 1.
ELA
112 L.C. Loi
On the other hand, since

Q
1,i
Q
i
= 0, it is easy to verify that kerA
i
kerG
i
= {0}.
This leads to the identity (3.30), as it was to be proved.
Noting that rankA
i
= r and rankG
i
= s for all i = 1, N 1 and applying Eq.
(3.30), we obtain that
rank(P
i

P
1,i
) = r +s m, i = 1, N 1.
In particular, we have rank(P
1

P
1,1
) = r+sm and rank(P
N1

P
1,N1
) = r+sm.
It follows that dim(kerR) = m.
Now using notations and the arguments in the proof of Lemma 3.8, we consider
a linear operator F from ker(
D
1
,

D
2
) to ker(D
1
, D
2
), dened by
F(
T
,
T
)
T
:=
_
T
,
_
+ (
Q
N1
Q
N1
)
P
1,N1
Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
X
N1
_
T
_
T
.
Let (y
T
, z
T
)
T
ker(D
1
, D
2
) be arbitrary, i.e., D
1
y + D
2
z = 0. Then we determine
two vectors and by = y and
= z + (Q
N1

Q
N1
)
P
1,N1
z +

Q
N1
P
N1

P
1,N1

G
1
1,N1
B
N1
X
N1
y.
From Eq. (2.5) and the facts that

Q
N1
Q
N1
= Q
N1
and

Q
1,N1
Q
N1
= 0, we
obtain
F(
T
,
T
)
T
= (y
T
, z
T
)
T
.
Moreover, it is easy to see that
D
1
+

D
2
= 0, i.e., (
T
,
T
)
T
ker(
D
1
,

D
2
).
This implies that F is sujective, hence F
_
ker(
D
1
,

D
2
)
_
= ker(D
1
, D
2
). According to
the property of the linear operator, we get dimF
_
ker(
D
1
,

D
2
)
_
dim
_
ker(
D
1
,

D
2
)
_
,
hence
dim
_
ker(D
1
, D
2
)
_
dim
_
ker(
D
1
,

D
2
)
_
.
Similarly, we also have the following inequality
dim
_
ker(
D
1
,

D
2
)
_
dim
_
ker(D
1
, D
2
)
_
,
which implies that
dim
_
ker(
D
1
,

D
2
)
_
= dim
_
ker(D
1
, D
2
)
_
.
ELA
Thus, dim
_
ker(D
1
, D
2
)
_
does not depend on the choice of T
i
, Q
i
and T
1,i
. On the
other hand, the following inclusion
kerR ker(D
1
, D
2
)
is always valid. Therefore, dim
_
ker(D
1
, D
2
)
_
:= p m.
As a direct consequence of Theorem 3.9 and Lemma 3.10 we come to the following
corollary.
Corollary 3.11. Suppose that the LSDE (3.15) is of index-2. Then the MPBVP
(3.15), (3.16) is uniquely solvable if and only if dim
_
ker(D
1
, D
2
)
_
= m.
Now we turn to the case when (3.22) is not valid, i.e., p > m and the MPVBP
(3.15), (3.16) has either no solution or an innite number of solutions. Using the same
notations as in the constant coecients case, we dene column matrices
i
:= X
i
U
(i = 0, N 2),
N1
:= X
N1
U T
N1
Q
N1

Q
1,N1
V and
N
:= X
N
U +Q
N1
V +
P
N1

Q
1,N1
V and a linear operator L acting in R
m(N+1)
,
L(x
T
0
, . . . , x
T
N
)
T
:=
_
(A
0
x
1
B
0
x
0
)
T
, . . . , (A
N1
x
N
B
N1
x
N1
)
T
,
_
N
i=0
C
i
x
i
_
T
_
T
.
The counterpart of Lemma 3.3 is now as follows.
Lemma 3.12. kerL =
__
(
0
a)
T
, . . . , (
N
a)
T
_
T
: a R
pm
_
.
Proof. The proof is similar to that for Lemma 3.3 except that now we note that
T
N1
Q
N1
= Q
N2
T
N1
Q
N1
, A
i
Q
i
= 0 (i = N 2, N 1), G
N1

Q
1,N1
= 0,
X
i
= X
i
P
1

P
1,1
(i = 0, N), and the equality (I P
N1

P
1,N1
)v
0
k
= 0 is equivalent
to the relation Q
N1
v
0
k
+P
N1

Q
1,N1
v
0
k
= 0 and implies that

Q
1,N1
v
0
k
= 0 for each
k = 1, m.
Theorem 3.13. Suppose that the LSDE (3.15) is of index-2 and p > m. Then,
the problem (3.15), (3.16) possesses a solution if and only if
W
= 0. (3.31)
Moreover, a general solution of (3.15), (3.16) can be given by
_
_
x
i
= X
i
+z
i
+
i
a, i = 0, N 2,
x
N1
= X
N1
+z
N1
T
N1
Q
N1

Q
N1
+
N1
a,
x
N
= X
N
+z
N
+Q
N1
+P
N1

Q
N1
+
N
a,
(3.32)
where a R
pm
is an arbitrary vector and (
T
,
T
)
T
= (D
1
, D
2
)
+
with (D
1
, D
2
)
+
the generalized inverse in Moore-Penroses sense of (D
1
, D
2
).
ELA
114 L.C. Loi
Proof. The proof of Theorem 3.13 is similar to that of Theorem 3.4 and will be
omitted.
Combining Theorem 3.9 and Theorem 3.13 we come to the following statement.
Corollary 3.14. (Fredholm alternative) Assume that the LSDE (3.15) is of
index-2 and let p := dim
_
ker(D
1
, D
2
)
_
. Then
(i) either p = m and the problem (3.15), (3.16) is uniquely solvable for any data
q
i
(i = 0, N 1) and ;
(ii) or p > m and the problem (3.15), (3.16) is solvable if and only if the condition
(3.31) is valid.
Moreover, the solution formula (3.32) holds.
Acknowledgment. The author wishes to thank Pham Ky Anh and Nguyen Huu
Du for many insightful discussions. The author also would like to extend appreciations
to the anonymous referee(s) for his/her very helpful suggestions which greatly improve
this paper.
REFERENCES
[1] R.P. Agarwal. On multipoint boundary value problems for discrete equations. Journal of Ma-
thetical Analysis and Applications, 96:520534, 1983.
[2] P.K. Anh, N.H. Du, and L.C. Loi. Singular dierence equations: An overview. Vietnam Journal
of Mathematics, 35(4):339372, 2007.
[3] P.K. Anh and L.C. Loi. On multipoint boundary-value problems for linear implicit non-
autonomous systems of dierence equations. Vietnam Journal of Mathematics, 29(3):281
286, 2001.
[4] S.L. Campbell. Singular Systems of Dierential Equations. Pitman Advanced Publishing Pro-
gram, London, 1980.
[5] S.L. Campbell. Singular Systems of Dierential Equations II. Pitman Advanced Publishing
Program, London, 1982.
[6] S.L. Campbell and D. Meyer. Generalized Inverses of Linear Transformations. Dover Publica-
tions, 1991.
[7] N.H. Du, L.C. Loi, T.K. Duy, and V.T. Viet. On index-2 linear implicit dierence equations.
Linear Algebra and its Applications, 434:394411, 2011.
[8] L.C. Loi, N.H. Du, and P.K. Anh. On linear implicit non-autonomous systems of dierence
equations. Journal of Dierence Equations and Applications, 8(12):10851105, 2002.

Electronic Journal of Linear Algebra ISSN 1081-3810 A Publication of The International Linear Algebra Society Volume 23, Pp. 88-114, January 2012

Загружено:

Сведения о документе

Исходное описание:

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Electronic Journal of Linear Algebra ISSN 1081-3810 A Publication of The International Linear Algebra Society Volume 23, Pp. 88-114, January 2012

Загружено:

Авторское право:

Доступные форматы

ELA

ON MULTIPOINT BOUNDARY VALUE PROBLEMS FOR INDEX-2

) is called the index of

which means that (

. Thus, the unique solution of (3.1), (3.2)

. Using the fact that

we come to the conclusion that the MPBVP (3.1),

. Thus, the problem (3.1), (3.2)

R is proved. To show the converse inclu-

Вам также может понравиться