Sunteți pe pagina 1din 19

Mathematical and Computer Modelling 44 (2006) 10251043

www.elsevier.com/locate/mcm

Laguerre matrix polynomial series expansion: Theory and computer


applications
J. Sastre a, , E. Defez b , L. Jodar b
a Instituto de Telecomunicaciones y Aplicaciones Multimedia, Universidad Politecnica de Valencia, Spain
b Instituto de Matematica Multidisciplinar, Universidad Politecnica de Valencia, Spain

Received 13 March 2006; accepted 14 March 2006

Abstract

This paper deals with the expansion of matrix functions in series of Laguerre matrix polynomials of a complex matrix parameter.
A new asymptotic expression for positive and large x and n is obtained for this family of matrix polynomials that improves previous
existing expressions. This expression is applied to improving the conditions imposed on the matrix function that are to be expanded
in a series of non-hermitian Laguerre matrix polynomials, providing a new series expansion theorem that is more general than
previous works. An application to solving linear differential systems without matrix exponentials is included.
c 2006 Elsevier Ltd. All rights reserved.

Keywords: Laguerre matrix polynomial; Asymptotic expression; Series expansion; Differential systems

1. Introduction

Laguerre matrix polynomials were introduced and studied in [16]. As in the corresponding problem for scalar
functions (see [7]), the problem of the development of matrix functions in series of Laguerre matrix polynomials
requires some new results about the matrix operational calculus that are not availablein the literature. Recently, the
asymptotics of normalized Laguerre matrix polynomials for n/x = o(1) and x/ n = O(1) as n were
studied in [6]. In this paper, a new asymptotic expression for large x and n is obtained and applied to provide a new
series expansion theorem that improves the recent results of [5].
Throughout this paper, for a complex number z, R(z) and =(z) denote its real and imaginary parts, respectively.
For a matrix A in Cr r , its spectrum (A) denotes the set of all the eigenvalues of A, (A) = max{R(z); z (A)},
(A) = min{R(z); z (A)}. The spectral radius of A denoted by (A) is the maximum of the set {|z|; z (A)}.
The 2-norm of A will be denoted by
kAxk2
kAk = sup , (1.1)
x6=0 kxk2

Corresponding author.
E-mail addresses: jorsasma@iteam.upv.es (J. Sastre), edefez@imm.upv.es (E. Defez), ljodar@imm.upv.es (L. Jodar).

c 2006 Elsevier Ltd. All rights reserved.


0895-7177/$ - see front matter
doi:10.1016/j.mcm.2006.03.006
1026 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

where, for a vector y in Cr , kyk2 = (y H y)1/2 is its euclidean norm [8]. If f (z) and g(z) are holomorphic
functions of the complex variable z, which are defined in an open set of the complex plane, and A is a matrix
in Cr r with (A) , then, from the properties of the matrix functional calculus [9, p. 558], it follows that
f (A)g(A) = g(A) f (A). The reciprocal gamma function denoted by 1 (z) = 1/ (z) is an entire function of the
complex variable z. Then the image of 1 (z) acting on A, denoted by 1 (A), is a well defined matrix. Furthermore,
if A + n I is invertible for every integer n 0, where I is the identity matrix in Cr r , then (A) is invertible, its
inverse coincides with 1 (A), and, by [10, p. 253], it follows that A(A + I ) (A + (n 1)I ) = (A +
n I ) 1 (A), n 1. If we denote by (A) the logarithmic norm of A, and by (A) the number (A), then [11]
n h io n h io
(A) = max z; z (A + A H )/2 , (A) = min z; z (A + A H )/2 . (1.2)

By (vi) [11, p. 647], it follows that

(A) = min{/ke At k e t , t 0}, (1.3)


and, taking into account the Schur decomposition of A, by [8, p. 192193] one gets

r 1
k
(A)t
X kAkr 1/2 t
At
ke k e , t 0. (1.4)
k=0
k!

Hence t A = exp(A log t) verifies


 (A)
t , t 1,
kt A k (A) (1.5)
t , 0 < t 1,
r 1
k
(A)
X kAkr 1/2 log t
A
kt k t , t 1, (1.6)
k=0
k!
r 1
k

(A)
X kAkr 1/2 log t
, 0 < t 1.
A A( log t)
kt k = e t
k=0
k!

Note that, by (i) [11, p. 647] and Bendixons theorem [12, p. 395], it follows that
kAk (A) (A) R(z) (A) (A) kAk , z (A), (1.7)
and, taking x = log t in 3.381 (4) of [13, p. 364], one gets
1 (1)n (n + 1)
Z
t m logn t dt = < , m > 1, n 0. (1.8)
0 (m + 1)n+1
This paper is organized as follows. Section 2 deals with some preliminary results. Section 3 is concerned with
functional relations satisfied by Laguerre matrix polynomials. Section 4 deals with the asymptotics of Laguerre matrix
polynomials for large n and x. The application of Section 4 to Laguerre matrix polynomial series expansion of matrix
functions is given in Section 5. Using functional relations of Section 3 and the results about series expansion of
Section 5, Section 6 provides an application for solving matrix differential systems without matrix exponentials.

2. Preliminaries

For the sake of clarity in presentation, we introduce the following definition:

Definition 2.1. Given a fixed matrix A in Cr r , a positive scalar function g(n), and a sequence of functions f n (z)
defined in an open set with (A) , we say that the sequence { f n (A)} behaves O(g(n), A), if k f n (A)k
M(A)g(n) for some positive constant M(A) and f n (A) commutes with A for n 1.
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1027

If is a complex number with R() > 0 and A is a matrix in Cr r with A + n I invertible for every integer
(A,) (A,)
n 1, the nth Laguerre matrix polynomial L n (x) and the nth normalized Laguerre matrix polynomial n (x)
are defined by [1, p. 58] and [2, p. 25]:
n  1
(1)k k (A + (n + 1)I )

2
L (A,) n(A,) (x) = L (A,)
X A+I
(x) = (A + I )n [(A + I )k ]1 x k , 2 (x),
n
k=0
k!(n k)! (n + 1) n

(2.1)
where (A + I )n = (A + I )(A + 2I ) (A + n I ), n 1, (A + I )0 = I . From (2.1), it is easy to show that

L (A,
n
2 ) (x) = L (A,1 ) ( x/ ).
n 2 1 (2.2)
The generating function of Laguerre matrix polynomials is given in [1, p. 57]:
  X
(A+I ) xt
G(x, t, , A) = (1 t) exp = L (A,)
n (x)t n , t C, |t| < 1, x C, (2.3)
1t n0

and Lemma 2.1 of [6] provides an integral expression for normalized Laguerre matrix polynomials (see [7, p. 107])
for (A) > 1/2 in terms of Hermite scalar polynomials:
s Z /2
(A,) A+I (A + (n + 1)I ) (1)n n! 1
n (x) = 2 (A + I /2) 2 (sin )2A H2n ( x cos )d. (2.4)
(n + 1) (2n)! 0

From (2.9) of [6], the following asymptotic expression of Hermite scalar polynomials holds:
x2 (2n)!
|H2n (x)| O(1)e 2 n s , (2.5)
n!

where s = 0 if |x| (4 )n, 0 < < 4, or s = 1/6 if x R.
The next lemma and results provide an asymptotic expression for an integral and other expressions that will be
used below.

Lemma 2.1. Let , and be positive numbers. Then the following asymptotic expression holds as x :
Z x1
v 1 (x v)1 ev dv = O(1)x 1 ex . (2.6)
0

Proof. Let , and be positive numbers, x , and

ex
Z x1
I0 = 1 v 1 (x v)1 ev dv. (2.7)
x 0

Taking u = x v in (2.7), one gets


u 1 1 u
Z x
I0 = 1 u e du. (2.8)
1 x
If 1, then
Z x Z 1
I0 u 1 eu du u 1 eu du = () = O(1). (2.9)
1 0
If 0 < < 1, taking such that 0 < < 1, then I0 = I1 + I2 , where
Z x  u 1 1 u
Z x  u 1 1 u
I1 = 1 u e du, I2 = 1 u e du. (2.10)
1 x x x
1028 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

As x , taking x > G > 0, one gets


 1 Z  1
1 1 u 1 1
I1 1 1 u e du < 1 1
() = O(1). (2.11)
x 0 G
Taking v = x u in integral I2 of (2.10), it follows that
Z xx 
x v 1

1 
I2 = 1 (x v)1 e(xv) dv x 1 e(x ) x x g(x), (2.12)
0 x
where
x 1 , 1,

g(x) = (2.13)
x (1) , 0 < < 1.
Hence I2 0 when x and, using (2.11), it follows that I0 = I1 + I2 = O(1). 
Let f (z) = M z log(z), > 0, M > 0. It is easy to show that f (z) > 0 if z = 1, limz f (z) = , that f (z)
1
has only one local minimum at z 0 = (M) , and that, given > 0, if M e, f (z 0 ) 0. Hence, given > 0, there

exists M > 0 such that M z log(z) and M 0 such that

M 0 z logk z,
0
0 = k, z 1. (2.14)
Note that, for x confined to an interval 0 < a x b, using (1.5) it follows that

kx B k = O(1), B Cr r , 0 < a x b. (2.15)

3. On the functional relations of Laguerre matrix polynomials

Laguerre matrix polynomials satisfy functional relations similar to those of Laguerre scalar polynomials; see 8.971
and 8.974-3 [13, p. 10371038], which will be useful for finding solutions of matrix differential systems using series
(A,)
of Laguerre matrix polynomials. Note that, from (2.1), one gets L 0 (x) = I and, in all this section, we take
(A,)
L 1 (x) = 0. Hence it is easy to prove that Laguerre matrix polynomials satisfy the three-term recurrence relation
(3.10) of [1, p. 59] for n 0; see Theorem 3.1 [1, p. 60],
(A,) (A,)
(n + 1)L n+1 (x) + [x I (A + (2n + 1)I )]L (A,)
n (x) + (A + n I )L n1 (x) = 0, n 0, (3.1)
which will be useful for calculating the nth polynomial in terms of the polynomials of order n 1 and n 2. Using
(2.3), it follows that
G(x, t, , A) X 0 (A,)
L n(A+I,) (x)t n ,
X
= L n (x)t n = t G(x, t, , A + I ) = t |t| < 1, (3.2)
x n0 n0

where 0 denotes differentiation with respect to x. Hence, identifying coefficients in the series of (3.2), one gets
(A,) (A+I,)
L 0n (x) = L n1 (x), n 0. (3.3)
From (2.3), it follows that
(1 t)G(x, t, , A + I ) = G(x, t, , A). (3.4)
Hence
L (A+I,) L (A+I,) L (A,)
X X X
n (x)t n n (x)t n+1 = n (x)t n . (3.5)
n0 n0 n0

Identifying coefficients in the last series, one gets


(A+I,)
L (A,)
n (x) = L (A+I,)
n (x) L n1 (x), n 0. (3.6)
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1029

Using (3.3) and (3.6), it follows that


1  0 (A,) (A,)

L (A,) (x) = L n (x) L 0 n+1 (x) , n 1. (3.7)
n

From (3.16) of [1, p. 59], one gets
(A,) (A,)
x L 0n (x) = n L n(A,) (x) (A + n I )L n1 (x), n 0. (3.8)
Taking n = n + 1 in the last expression and using (3.7), it follows that
(A,) (A,)
x L 0 n+1 (x) = (n + 1)L n+1 (x) (A + (n + 1)I )L (A,)
n (x)
(A,)
= x L 0n (x) x L (A,)
n (x), (3.9)
and hence
(A,) (A,)
x L 0n (x) = (n + 1)L n+1 (x) (A + (n + 1 x)I )L (A,)
n (x), n 1. (3.10)
Taking n = n + 1 in (3.3), and using (3.8) and (3.6), one gets
(A,) (A,)
x L 0 n+1 (x) = xL (A+I,)
n (x) = (n + 1)L n+1 (x) (A + (n + 1)I )L (A,)
n (x)
(A+I,) (A,)
= xL n+1 (x) + xL n+1 (x), n 1, (3.11)
and, from (3.11), it follows that
1 (A,)
 n  (A,)
x L (A+I,) (x) = (A + n I )L n1 (x) + x L n (x), n 0, (3.12)
n

1 1 (A,)
= (A + (n + 1)I )L (A,) (x) (n + 1)L n+1 (x), n 1. (3.13)
n

Finally, from (3.6), one gets
n
L (A,) (x) = L (A+I,)
X
m n (x). (3.14)
m=0

4. Asymptotic expressions for positive large x and n

The next result provides an asymptotic expression for normalized Laguerre polynomials for positive large x and n.

Theorem 4.1. Let A Cr r , > 0, n N, x > 0, and s = 0 if x < (4 )n, otherwise s = 1/6, for any R
with 0 < < 4. Then the following asymptotic expressions hold as x , n , n > (A) + 1:
  A k0 +s
(A,) x n n
n (x) = O(1, A)e 2 1
(4.1)
x x k0 + 2
x 0
e 2 n 2 +k0 +s x 1
 
= O(1, A) log , (4.2)
x 0 +k0 + 2
1 n
where k0 = 0 if (A) >

1
2 and, if (A) 21 , k0 is an integer such that 12 k0 < (A) 1
2 k0 , and, for (4.2), if
n/x = o(1) then
(i) 0 = (A) and 1 = 0, or
(ii) 0 = (A) and 1 = r 1,

otherwise, if x/ n = o(1), then
(iii) 0 = (A) and 1 = 0, or
1030 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

(iv) 0 = (A), 1 = r 1,

otherwise, if x/ n = M, M > 0, then
(v) 0 = 1 = 0.
Proof. Taking u = cos2 in (2.4), by Definition 2.1 and (3.32) of [3], if n > (A) + 1, > 0 and (A) > 12 , one
gets
 A
[x(1 u)] A
Z 1 
(A,) n n! 
n (x) = O(1, A) Kn , Kn = H2n xu du. (4.3)
x (2n)! 0 u(1 u)
Hence, using (2.5) and Definition 2.1, it follows that
kK n k n s K n 1 + K n 2 , K n = n s O(1, A)K n 1 + O(1, A)K n 2 ,
   
(4.4)
where s = 0 if x (4 )n or s = 1/6 if x > (4 )n, 0 < < 4, and
Z 1 1 A
Z 1
[x(1 u)] A xu
x [x(1 u)] xu
K n1 = e 2 du, K n 2 = e 2 du. (4.5)
0 u(1 u) 1 x1 u(1 u)
Note that x(1 u) 1 in the first integral and 0 x(1 u) 1 in the second integral. Taking v = x(1 u) in
integral K n 2 of (4.5) and using (1.6) and (1.8), for x > G > 0, one gets
x Z x r 1 k Z 1 (A) 1 x
e2 1 kv A kv 1/2
v e2 X kAk r v 2 v e2
K n2 = q e 2 dv q ( log(v))k e 2 dv = O(1) ,
x 0 1 vx x k=0 k! 0 1 G1 x

(4.6)
which holds for (A) > 1/2 and x .
Taking v = xu in integral K n 1 , using (1.6), (2.14) and Lemma 2.1 it follows that
r 1 k Z x1
kAk r v
(x v)(A) 2 v 2 (log(x v))k e 2 dv
X 1 1
K n1
k=0
k! 0
x
Z x1 1 v e2
(x v)(A)+ 2 v 2 e 2 dv = O(1) .
1
O(1) (4.7)
0 x
Using (4.6) and (4.7), if (A) > 12 and x , one gets
 A x
n se
2
(A,)
n (x) = O(1, A) n .
x x
If 32 < (A) 12 , taking into account that using Taylors expansion with (1 + t) A as t 0, one gets
 A  
n1
n = I + O n1 , A and, by (2.4) of [6], it follows that

A + (n + 1)I (A+I,)
r r
(A,) n (A+I,)
n (x) = n (x) (x)
n1
x
e 2 AI h A+I A+I
i
= x n O(1, A)n 2 n s O(1, A)(n 1) 2 +s I
x
 A s
n n n x
= e 2 O(1, A). (4.8)
x xx
Similarly, if 52 < (A) 32 , one gets
 A s  
(A,) n n n 2 x
n (x) = e 2 O(1, A),
x x x
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1031

and, by an induction argument, if 21 k0 < (A) 1


2 k0 , where k0 is a positive integer and (A) 21 , (4.1)

n
follows. Using (1.5) and (1.6) with t = x in (4.1), (4.2) follows with conditions (i)(v). 
(A,) (A,)
The next result provides asymptotic expressions for n+1 (x) n (x).

Corollary 4.1. Let A Cr r , > 0, n N, x > 0, and s = 0 if x < (4 )n, otherwise s = 1/6, for any R
with 0 < < 4. Then the following asymptotic expressions hold as x , n , n > (A) + 1:
  A k0 +s
(A,) (A,) x n n
n+1 (x) n (x) = O(1, A)e 2 1
(4.9)
x x k0 + 2
x 0
e 2 n 2 +k0 +s x 1
 
= O(1, A) log n , (4.10)
x 0 +k0 + 2
1

k0 = 1 if (A) > and, if (A) 12 , k0 is the integer such that 21 k0 < (A)
1 1
where 2 2 k0 , and, for (4.10),
if n/x = o(1), then
(i) 0 = (A) and 1 = 0, or
(ii) 0 = (A) and 1 = r 1,

otherwise, if x/ n = o(1), then
(iii) 0 = (A) and 1 = 0, or
(iv) 0 = (A), 1 = r 1,

otherwise, if x/ n = M, M > 0, then
(v) 0 = 1 = 0.
Proof. By (2.5) of [6] and (4.1), if (A) > 21 , one gets
x x
(A,) e2 e2
n+1 (x) n(A,) (x)
AI A
/(n + 1)O(1, A)(n + 1)
p +s I A+I
= 2 x + O(1, A)(n + 1) 2 +s I I x A
x x
  A s1
n n x
= O(1, A) 1
e2. (4.11)
x x 2 1
Working in a similar way, if 21 < (A) 12 , it follows that
  x
(A,) (A,) A
 n 1 A e2
n+1 (x) n (x) = O(1, A)(n + 1) 2 I +s I
x + O(1, A)(n + 1) 2 I +s I
x A
x x
 A s
n n x
= O(1, A) 1
e2. (4.12)
x x2
If 32 < (A) 12 , one gets
  n 2  n 1  e x2
(A,)
n+1 (x) n(A,) (x)
A A
= O(1, A)(n + 1) 2 I +s I
x + O(1, A)(n + 1) 2 I +s I x A
x x x
  n 1   n  x
A
= O(1, A)x + O(1, A) n 2 I +s I x A e2
x x
  A s+1
n n x
= O(1, A) 1
e2. (4.13)
x x 2 +1
(A,) (A,)
Using an induction argument together with (1.5) and (1.6), the asymptotic expressions for n+1 (x) n (x)
follow. 
1032 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

5. Application to Laguerre matrix polynomial series expansion

Using the asymptotic expressions of Section 4, the expansion theorem of matrix functions in a series of
Laguerre matrix polynomials given by Theorem 3.1 of [5] is extended to a more general situation; see [7]
(pp. 246247, 266268). Concretely the existence of integrals (3.55) and (3.56), the conditions (3.57) and (3.58)
of [5], and the condition (A) > 1/2 are improved and the following result holds:
(A,)
Theorem 5.1. Let > 0 and A Cr r . Let L n (x) be the Laguerre matrix polynomial defined by (2.1). Let F(x)
0 0
be a matrix function that takes values in Cr r or Cr r , bounded in a certain interval [x , x + ] with 0 < < x,
and locally absolutely integrable in ]0, [. Assume that the integrals
Z H
y (A)/21/4 ( log y)r 1 kF(y)k dy, 0 < H e1 , (A) > 1/2, (5.1)
0
or
Z H
y (A)/21/4 kF(y)k dy, 0 < H 1, (A) > 1/2, (5.2)
0
Z
ey/2 y (A)/23/4 (log y)r 1 kF(y)k dy, G e, (5.3)
G
or
Z
ey/2 y (A)/23/4 kF(y)k dy, G 1, (5.4)
G
and
Z
y 3/2 ey(1)/2 kF(y)k dy, G 1, (A) > 1/2, (5.5)
G
exist for certain values of H , G and with 0 < < 1, and let us consider the sums of the right and left series given
by
(A,) (r ) (A,) (r )
Sn(r ) = L 0 (x)c0 + L 1 (x)c1 + + L (A,)
n (x)cn(r ) , (5.6)
with
Z
(r ) (A,) 0
ck = A+I k! 1 (A + (k + 1)I ) y A ey L k (y)F(y)dy, F(y) : [0, [ Cr r , (5.7)
0
and by
(l) (A,) (l) (A,)
Sn(l) = c0 L 0 (x) + c1 L 1 (x) + + cn(l) L (A,)
n (x), (5.8)
with
Z 
(l) (A,) 0
ck = A y
F(y)y e Lk (y)dy A+I k! 1 (A + (k + 1)I ), F(y) : [0, [ Cr r . (5.9)
0
Then, for x > 0, one gets

Z x+ sin 2 n(x )

1
lim Sn(r ) F( 2
)d = 0, (5.10)
n x x

Z x+ sin 2 n(x )

1
lim Sn(l) F( 2
)d = 0. (5.11)
n x x

The existence of integral (5.5) can be replaced by the following more general conditions:
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1033

(i) if (A) > 1/2,


Z 4 n (A) 3
(

3/2 y/2 o(n 2 4 (log n)r +1 ) or
y e kF(y)k dy = (A) 3 (5.12)
nd o(n 2 4 ),
(A) 11
Z (
3/2 y/2 o(n 2 12 (log n)r +1 ) or
y e kF(y)k dy = (A) 11 (5.13)
4
n o(n 2 12 ),
(ii) if (A) > 1/2,
Z 4 n (A) 1
(

3/2 y/2 o(n 2 4 (log n)r +1 ) or
y e kF(y)k dy = (A) 1 (5.14)
nd o(n 2 4 ),
Z 4 n (A) 1
(

1/2 y/2 o(n 2 + 4 (log n)r +1 ) or
y e kF(y)k dy = (A) 1 (5.15)
nd o(n 2 + 4 ),
(A)
Z ( 1
1/2 y/2 o(n 2 + 12 (log n)r +1 ) or
y e kF(y)k dy = (A) 1 (5.16)
4
n o(n 2 + 12 ),
for any R with 0 < < 4 and d > 0 such that
 
2
d = min , 1/3 , with 0 < < 1/3, (5.17)
3l + 3
    
1 13 1 4
l + , 0 < < , l N, (5.18)
3 2 3 3
max{|2(A) + 1|, |2(A) 1|, |2(A) 3|}, R, if (A) > 1/2, (5.19)
max{|2(A) 1|, |2(A) 3|}, R, if (A) > 1/2. (5.20)
Proof. The proof is analogous to that given for Theorem 3.1 in Section 3 of [5], improving the bound for integral
J1 appearing in expression (3.13) of Section 3.2 of [5]. In Section 3 of [5], the partial sum of the Laguerre matrix
polynomial series of the matrix function F(x) is split into three integrals, I H + I HG + IG (see (3.1) [5]) and, if integrals
(5.1) or (5.2) exist, then it is shown that kI H k can be made less than any given positive number, taking H sufficiently
small, and

Z x+ sin 2 n(x )

1
lim I HG F( 2
)d = 0. (5.21)
n x x

In the following, we show that, if integrals (5.3) or (5.4) exist, and integral (5.5) exists or the more general conditions
of (i) and (ii) are satisfied, then kIG k can be made less than any given positive number, taking G sufficiently large. In
a similar way to Section 3 of [5], we take x such that 0 < H < a x b < G for certain positive numbers a, b, H
and G. Using (3.13), (3.3) and (3.5) of [5], one gets

kIG k O(1)[n J0 (G, n d ) + J1 ], (5.22)


where
J1 = n J0 (n d , ), (5.23)
(A,) (A,) (A,)

n+1 (x) n (x) y A n (y)
Z b1
J0 (b0 , b1 ) = ey
b0 |y x|
(A,) (A,) (A,)
h i
y n+1 (y) n (y) n (x)
A
+ kF(y)k dy, (5.24)
|y x|
1034 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

and d > 0. Using Theorem 4.2 of [4], (5.24), (2.15), (1.5) and (1.6), taking into account that 0 < H < a x b < G
and assuming that G e, it follows that
Z nd A/2
y/2 y
4y
J0 (G, n ) O(1/n)
d
e kF(y)k dy, (5.25)
G (1 x/y)y
y (A)/23/4 (log y)r 1
Z nd
J0 (G, n ) O(1/n)
d
ey/2 kF(y)k dy,
G (1 b/G)
y (A)/23/4
Z nd
J0 (G, n d ) O(1/n) ey/2 kF(y)k dy,
G (1 b/G)
where, if (A) > 1/2, d is given by expressions (5.17)(5.19), and, if (A) > 1/2, d takes the form (5.17), (5.18)
and (5.20).
Taking n sufficiently large, then G < n d and, using the asymptotic behavior of Laguerre matrix polynomials given
by (4.1) of Theorem 4.1 for the polynomials in y, and (4.57), (4.59) of [4] for the polynomials in x, if (A) > 1/2
from (5.23) and (5.24) it follows that
J1 O(1)I1 + O(1)I2 , (5.26)
where
A A  k 0
2 2 s y2 y
A A
2 2 n
Z x n n e Z x n y ns e 2
I1 = n ey 3 3
kF(y)k dy, I2 = n 1
  3 kF(y)k dy, (5.27)
nd n4 (1 xy )y 2 nd n 4 1 xy y 2

with k0 = 1 if (A) > 1/2, k0 = 0 if 1/2 < (A) 1/2, and


(4 )n

0, if y <
,0 < < 4
s= 1 (5.28)
(4 )n
, if y
, 0 < < 4.
6
For the first integral I1 of (5.27) using (1.5) and (2.15), one gets
 
y
Z 4 n e 2 Z y
1 A 1 e 2
I1 O(1)n 4 n 2 3
kF(y)k dy + n 6   3
kF(y)k dy , (5.29)
4
nd y2 n y 2

and
y
3
A Z e 2
I2 O(1)n 4 +k0 n s
kF(y)k dy. (5.30)
2
n 3
nd y 2 +k0
We consider the following cases:
(i) (A) > 12 . In this case, k0 = 1 and, using (5.30), one gets
 
y
Z 4 n e 2 Z y
1 A 1 e 2
I2 = O(1)n 4 n 2 kF(y)k dy + n 6  4  kF(y)k dy . (5.31)
nd y n y

Hence, noting that


y y
e 2 e 2
Z Z
1
  3
kF(y)k dy O(1)   kF(y)k dy, (5.32)
4
n y2 n 4
n y
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1035

and using (5.29), it follows that


J1 I1 + I2
y
 
4
A Z 1 n e 2
O(1)n n 2 kF(y)k dy
4
3
nd y2

y
 
4
n e 2
Z

A
41
+n O(1) kF(y)k dy
2
n
nd y

y
e 2
Z
1
+O(1)n 6   kF(y)k dy , (A) > 1/2. (5.33)
4
n y

(ii) 1/2 < (A) 1/2. In this case, k0 = 0 and, using (5.30), one gets
 
y
Z 4 n e 2 Z y
3 A 1 e 2
I2 O(1)n 4 n 2 3
kF(y)k dy + n 6   3
kF(y)k dy . (5.34)
4
nd y2 n y2

Starting from (3.14) of [5] and using the asymptotic expressions (4.1) for the normalized Laguerre matrix
polynomials in y and (4.57) of [4] for the same polynomials in x, it is easy to show that (5.34) is also true
for (A) > 1/2. Hence, using (5.29), it follows that
J1 I1 + I2

y y
 
4
A Z n e 2 e 2
Z
3 1
O(1)n n kF(y)k dy + n kF(y)k dy , (A) > 1/2. (5.35)
2 4 6
3 
4
 3
nd y2 n y2
Using (5.22), (5.25), (5.33), (5.35), (1.5) and (1.6) one gets under conditions (5.12) and (5.13) if (A) > 1/2, or
(5.14)(5.16) if (A) > 1/2, then, for large enough values of G, kIG k can be bounded for any positive given number,
independent of n. Note that, if (A) > 1/2 and integral (5.5) exists, taking 0 < < 1, one gets
Z Z
n h 1 /2
y 3/2 y/2
e kF(y)k dy e y 3/2 ey(1)/2 kF(y)k dy = o(n h 2 (log n)h 3 ), (5.36)
nh1 nh1

for any h 1 > 0, h 2 R, h 3 N, and then, if integral (5.5) exists, conditions (5.12) and (5.13) are satisfied.
Finally, using (2.1) and (3.59)(3.62) of Theorem 3.1 of [5], one gets expressions (5.6)(5.9) for the right and left
series expansions. 

Note that, from (1.7), condition (A) > 1/2 is more general than (A) > 1/2 and integrals of (i), (ii) and
(5.5) avoid the differences (A) (A) and (A) (A), which can be large in the matrix case, and the terms in
logr 1 y of expressions (3.55)(3.58) of [5]. Condition (A) > 1/2 is only needed if the existence of integral (5.2)
instead of (5.1) is required, and the terms in o() in expressions (5.12) and (5.14)(5.16) are more general than those
on (3.57) and (3.58) of [5]. If (A) > 1/2, then the conditions of (ii), more general than those of (i), can be used.
From (5.10) and (5.11), the behavior of series
(5.6) and (5.8) at point x is the same as the behavior of the Fourier
trigonometric series of F(2 ) at the point = x; see [14, p. 385], (1.6.4) [7, p. 12]. Hence, for example, if
F(y) is of bounded variation in [x , x + ], (5.37)
from (5.10) and (5.11), by the convergence properties of Dirichlets integrals [15], one gets
F(x 0) + F(x + 0)
lim S (r )
0
= , F(y) : [0, [ Cr r , (5.38)
n n 2
F(x 0) + F(x + 0)
lim S (l)
0
= , F(y) : [0, [ Cr r . (5.39)
n n 2
1036 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

Taking F(y) : [0, [ R, A Rr r , r = 1, and = 1 (scalar real case), the conditions of existence of the
integrals (5.1)(5.4), expressions (5.10) and (5.11), and the series coefficient expressions (5.7) and (5.9) are equivalent
to those given by Uspenskys and Szegos theorems; see the second integral of (2) of [16, p. 619], and the integral on
the right of (9.1.11) and the first integral of (9.1.14) of Theorem 9.1.5 [7, p. 246247]. Conditions (5.12)(5.16) are of
the same type as Szegos (9.1.12) and (9.1.14) [7, p. 247]. In fact, in the scalar real case (A R, (A) = (A) = A,
= 1), if condition (5.13) is satisfied, taking = 3 it follows that
Z
A 5
n 2 + 12 y 3/2 ey/2 |F(y)|dy = o(n 1/2 ), (5.40)
n
and taking into account that, from (1.7) and the hypotheses of (i), one gets A > 1/2, it follows that
Z Z
A 5
+ 12 3/2 y/2
n 2 y e |F(y)|dy y A/213/12 ey/2 |F(y)|dy. (5.41)
n n
Using (5.40), if the condition (9.1.12) [7, p. 247] of Szegos theorem is verified, then condition (5.40) is satisfied, but
the reciprocal is not true. Hence, (5.13) is more general than Szegos. If condition (5.13) is verified, then (5.16) is also
satisfied, and it follows that this condition is also more general than (9.1.12) of [7, p. 247].
In a similar way, if condition of (5.12) is satisfied in the scalar real case, it follows that
Z n
A 3
n 2 +4 y 3/2 ey/2 |F(y)|dy = o(1), (5.42)
nd
and hence
A 3
Z n Z n
n 2 +4 y 3/2 ey/2 |F(y)|dy y A/23/4 ey/2 |F(y)|dy = o(1), (5.43)
nd nd
if the first condition of (9.1.14) of [7, p. 247] is satisfied, then
Z
y A/23/4 ey/2 |F(y)|dy = O(1), (5.44)
1
and
Z n
y A/23/4 ey/2 |F(y)|dy = o(1). (5.45)
nd
However, (5.44) does not imply (5.12), and this condition is more restrictive than (9.1.14) of [7, p. 247]. Analogously,
it is easy to prove that the same holds for condition (5.14) and (5.15).
It is important to note that, in the scalar case, Theorem 5.1 holds for complex functions F(y) and complex
parameter A, imposing restrictions only on the real part of A. Szego and Uspenskys theorems are valid only for
A R.

6. Application to differential systems

In many engineering, physical, biological and economic phenomena, the processes are described by systems of
linear first-order ordinary differential equations with constant coefficients. So, the system

Y 0 (t) = AY (t), Y (0) = y0 , (6.1)


where Y (t) is a matrix function Y : R Cr , A is a matrix in Cr r , and y0 is a vector in Cr , arises in circuit theory,
mechanics, control theory [17, p. 1828], etc. In this section, we present a method to solve a differential linear system
using series of Laguerre matrix polynomials and their recurrence relations. Let us assume for the moment that the
solution of system (6.1) can be expanded in series of Laguerre matrix polynomials such as the following:
" #
(D,)
X
Y (t) = ck L k (t) y0 , t > 0, (6.2)
k0
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1037

(a) N = 3. (b) N = 4.

(c) N = 5.

Fig. 1. Absolute error of the truncated approximation series of Y1 (t = 1).

where > 0, D Cr r , (D) > 1/2, ck are the matrix coefficients in Cr r (see Theorem 5.1), and let us impose
the condition
(D,)
X
ck L k (0) = I. (6.3)
k0

Substituting (6.2) in the system (6.1) and using (3.3) in the left member and (3.6) in the right member, it follows
that
" # " #
(D+I,) (D+I,)
X X
ck+1 L k (t) y0 = A (ck ck+1 )L k (t) y0 , (6.4)
k0 k0

and hence one possible solution for the coefficients is


(A I )ck+1 = Ack . (6.5)
If (A I )1 exists for certain value of > 0, then one gets

ck+1 = (A I )1 Ack , k = 0, 1, . . . . (6.6)


1038 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

(a) Y1 (t = 1), N = 9. (b) Y2 (t = 1), N = 9.

(c) Y3 (t = 1), N = 8.

Fig. 2. Absolute error of the truncated approximation series of Yi (t = 1), i = 1, 2, 3.

Taking into account (6.6) and (6.3), it follows that

ck = [(A I )1 A]k c0 , (6.7)

and
(D,)
X
[(A I )1 A]k c0 L k (0) = I. (6.8)
k0

Hence, noting that, by (2.1), one gets

(D,) (D + I )k
Lk (0) = , (6.9)
k!
if we choose D Cr r commuting with A, then it follows that
" #1
X [(A I )1 A]k (D + I )k
c0 = , (6.10)
k0
k!
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1039

(a) N = 3. (b) N = 4.

(c) N = 5.

Fig. 3. Absolute error of the truncated approximation series of Y1 (t = 1) for matrix A11 of [20, p. 379].

" #1 " #
X [(A I )1 A]k (D + I )k X (D,)
Y (t) = [(A I ) 1
A] k
Lk (t) y0 . (6.11)
k0
k! k0

Truncating both series of (6.11) from k = 0 to k = N , N N, it is possible to approximate the solution of (6.1) using
only N + 1 terms of the Laguerre matrix polynomial series. From [17, p. 32], the exact solution of the system is

Y (t) = e At y0 . (6.12)

Using (6.12) together with (1.3), it follows that

r 1
k
(A)t
X kAkr 1/2 t
kY (t)k ke kky0 k ky0 ke
At
, t 0, (6.13)
k=0
k!

and hence, from (6.13), it is easy to show that function Y (t) verifies the existence of integrals (5.1)(5.5) of
Theorem 5.1 taking

> 2(A). (6.14)


1040 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

(a) Y1 (t = 1), N = 9. (b) Y5 (t = 1), N = 8.

(c) Y10 (t = 1), N = 8.

Fig. 4. Absolute error of the truncated approximation series of Yi (t = 1), i = 1, 5, 10 for matrix A11 of [20, p. 379].

Let us deal with some numerical examples. All calculations have been performed with the mathematical package
M ATLAB v. 6.5, in a Pentium IV 3.4 GHz with 1 GB of RAM memory. Consider the differential system (6.1) with
dates, see Example 3.1 of [18, p. 112],

3 1 1 1
A = 2 0 1 , y0 = 3 , (6.15)
1 1 2 2
where (A) = {1, 2} and A is nondiagonalizable, and let us approximate its solution Y (t) for t = 1. From [17, p. 32]
and using the minimal theorem [9, p. 571], one gets the exact value

e2

7.38905609893065
Y (1) = 2e + e2 = 12.82561975584874 . (6.16)
2e 5.43656365691809

Figs. 1 and 2 present the absolute errors



Yi (t) Yi (t) , i = 1, 2, 3, (6.17)

J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1041

(a) N = 3. (b) N = 4.

(c) N = 5.

Fig. 5. Absolute error of the truncated approximation series of Y1 (t = 1) for matrix A11 of [20, p. 379], size 50 50.

where Yi (t) is the exact solution and Yi (t) is the truncated series approximation for t = 1, varying the parameter ,
for different values of the number of terms of the series (N + 1), using scaling of the matrix A as described in [19,
p. 42] and taking D equal to the scaled A in (6.2):
D = A/2s , s = 4. (6.18)
As it can be seen in Fig. 1, the error has some minimum below = 20, and for > 30 it tends to a constant, which
is greatly improved when one more term is added in the truncated series. Fig. 2 shows that, for N 8, the limit of
machine precision is reached for many of the values of presented.
Now consider the differential system (6.1) with dates, see Example 11 of [20, p. 379], A Rr r , y0 Rr ,

i i > j,

y0 (i) = 1, A(i, j) = 1 i = j, 1 i, j r, (6.19)


j i < j,

with r = 20. Using function expm2 of M ATLAB (see [19, p. 43]), it is easy to test that the exponential of this matrix
does not converge using Taylor series directly. Figs. 3 and 4 present the absolute errors (6.17) with respect to the
reference value obtained using the Pade method with the appropriate scaling described in [19, p. 42], function expm
of M ATLAB, in the same conditions as in the previous example: t = 1, varying the parameter , for different values
1042 J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043

(a) Y1 (t = 1), N = 9. (b) Y5 (t = 1), N = 9.

(c) Y10 (t = 1), N = 9.

Fig. 6. Absolute error of the truncated approximation series of Yi (t = 1), i = 1, 5, 10 for matrix A11 of [20, p. 379], size 50 50.

of the number of terms of the series (N + 1), and using scaling of the matrix A and D [19, p. 42]:
D = A/2s , s = 10. (6.20)
Figs. 5 and 6 present the absolute errors (6.17) for the same example with r = 50, with the same conditions.
As can be seen, the behavior of the error is similar for a wide range of values of on the examples of sizes 20 20
and 50 50.

Acknowledgements

We want to acknowledge engineer Jorge Sanchez Quilis for his help in developing the software to test the method.
This work was partially supported by the Generalitat Valenciana.

References

[1] L. Jodar, R. Company, E. Navarro, Laguerre matrix polynomials and systems of second order differential equations, Appl. Numer. Math. 15
(1994) 5363.
[2] L. Jodar, J. Sastre, The growth of Laguerre matrix polynomials on bounded intervals, Appl. Math. Lett. 14 (2001) 2126.
[3] L. Jodar, J. Sastre, Asymptotic expressions of normalized Laguerre matrix polynomials on bounded intervals, Util. Math. 65 (2004) 331.
J. Sastre et al. / Mathematical and Computer Modelling 44 (2006) 10251043 1043

[4] J. Sastre, L. Jodar, On the asymptotics of Laguerre matrix polynomials, Util. Math. 70 (2006) (in press).
[5] J. Sastre, L. Jodar, On Laguerre matrix polynomials series, Util. Math. 71 (2006) (in press).
[6] J. Sastre, E. Defez, On the asymptotics of Laguerre matrix polynomials for large x and n, Appl. Math. Lett. 19 (8) (2006) (in press).
doi:1016/j.aml.2005.10.003.
[7] G. Szego, Orthogonal Polynomials, 4th ed., in: Amer. Math. Soc. Colloquium Publications, vol. 23, Amer. Math. Soc., Providence, Rhode
Island, 1975.
[8] G. Golub, C.F. Van Loan, Matrix Computations, Johns Hopkins Univ. Press, Baltimore, MA, 1989.
[9] N. Dunford, J. Schwartz, Linear Operators, Part I, Interscience, New York, 1955.
[10] E. Hille, Lectures on Ordinary Differential Equations, Addison-Wesley, New York, 1969.
[11] I. Higueras, B. Garca-Celayeta, Logarithmic norms for matrix pencils, SIAM J. Matrix Anal. 20 (3) (1999) 646666.
[12] J. Stoer, R. Bulirsch, Introduction to Numerical Analysis, Springer, New York, 1980.
[13] I.S. Gradshteyn, I.M. Ryzhik, Tables of Integrals, Series and Products, 5th ed., Academic Press, Inc., 1980.
[14] G. Sansone, Orthogonal Functions, Dover, 1991.
[15] T.M. Apostol, Mathematical Analysis, Addison-Wesley, New York, 1957.
[16] J.V. Uspensky, On the development of arbitrary functions in series of Hermites and Laguerres polynomials, Ann. of Math. 28 (1927)
593619.
[17] S.L. Campbell, Singular Systems of Differential Equations, in: Research Notes in Mathematics, vol. 40, Pitman Advanced Publishing Program,
1980.
[18] E. Defez, L. Jodar, Some Applications of Hermite matrix polynomials series expansions, J. Comput. Appl. Math. 99 (1998) 105117.
[19] C.B. Moler, C.F. Van Loan, Nineteen dubious ways to compute the exponential of a matrix, twenty-five years later, SIAM Rev. 45 (1) (2003)
349.
[20] D. Westreich, A practical method for computing the exponential of a matrix and its integral, Commun. Appl. Numer. Methods 6 (1990)
375380.