You are on page 1of 20

Applied Numerical Mathematics 59 (2009) 2970–2989

Contents lists available at ScienceDirect

Applied Numerical Mathematics

www.elsevier.com/locate/apnum

A Nyström method for a class of Fredholm integral equations


of the third kind on unbounded domains ✩
Luisa Fermo
Department of Mathematics and Computer Science, University of Basilicata, v.le dell’Ateneo Lucano 10, 85100 Potenza, Italy

a r t i c l e i n f o a b s t r a c t

Article history: The author proposes a numerical procedure in order to approximate the solution of
Received 27 October 2008 a class of Fredholm integral equations of the third kind on unbounded domains. The
Received in revised form 10 July 2009 given equation is transformed in a Fredholm integral equation of the second kind. Hence,
Accepted 13 July 2009
according to the integration interval, the equation is regularized by means of a suitable
Available online 15 July 2009
one-to-one map or is transformed in a system of two Fredholm integral equations that
MSC: are subsequently regularized. In both cases a Nyström method is applied, the convergence
65R20 and the stability of which are proved in spaces of weighted continuous functions. Error
45E05 estimates and numerical tests are also included.
 2009 IMACS. Published by Elsevier B.V. All rights reserved.
Keywords:
Fredholm integral equations
Nyström method

1. Introduction

This paper deals with the numerical treatment of the following class of Fredholm integral equations of the third kind on
unbounded domains,

b
p( y) f ( y) − μ k(x, y ) f (x) w (x) dx = g ( y ), y ∈ (a, b), (1.1)
a
β
where (a, b) = (0, ∞) or (a, b) = (−∞, +∞), f is the unknown, k, g and p are given functions, μ ∈ R, w (x) = |x|α e −|x| ,
x ∈ (a, b), α > −1, β > 12 if (a, b) = (0, ∞) and β > 1 if (a, b) = (−∞, +∞). Moreover we will assume that the right-hand
side g is a smooth function, the coefficient p has only one zero at the origin of the type y ι , 0 < ι < 1 and that the kernel k
behaves like y λ logℓ y, 0 < λ < 1, ℓ  0 at the origin.
The case when w ≡ 1 and the integral is defined on finite interval was extensively investigated (see, for instance, [24,22])
and there exists a wide literature about numerical methods, e.g. collocation and spline methods (see, for instance, [9,8,25,
26] and the bibliography therein) in order to approximate the solution of these equations in different spaces of functions.
In this paper we examine the case when the given functions are defined on unbounded domains, have singularities at
the origin and increase exponentially for x → ∞. Hence in virtue of the smoothness of the given functions, we consider the
above equation in a suitable weighted space equipped with the uniform norm.
In order to approximate its solution (if it exists) we propose a strategy mainly consisting in three steps.


Work supported by the research project PRIN 2006 “Numerical methods for structured algebra and applications” of the Italian Ministry for the University
and Research.
E-mail address: luisa.fermo@unibas.it.

0168-9274/$30.00  2009 IMACS. Published by Elsevier B.V. All rights reserved.


doi:10.1016/j.apnum.2009.07.002
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2971

The first step transforms the given equation in a Fredholm integral equation of the second kind. Then, we could apply
a projection or a Nyström method based on the orthogonal polynomials with respect to the weight w appearing into the
integral (see, for instance, [27,13,17]). Nevertheless it is possible to see that, in virtue of the low smoothness of the kernel,
these methods led to very poor numerical results. From here the necessity to introduce a regularizing procedure that allow
us to improve the smoothness of the given functions and then to approximate the solution with a satisfactory order of
convergence.
This is the second step that is different according to the interval (a, b). If (a, b) = (0, ∞) we regularize the equation
by means of a suitable regularization, while if (a, b) = (−∞, +∞), proceeding as in [7], we first reduce the equation in a
system of two Fredholm integral equations on (0, ∞) and then we regularize them. Thus, we consider the new equation (or
system) in a suitable weighted space equipped with the uniform norm and we prove that it has a unique solution in this
space if and only if the original equation is unisolvent in a suitable space equipped with weighted uniform metric.
The last step consists in approximating this solution by using a Nyström type method (see, for instance, [13,27,4]) based
on a suitable truncated Gaussian rule (see, for instance, [14,15,6] and [16] (see also [5])). We prove the stability of the
method and that the matrix of the linear system we solve is well conditioned. Moreover we prove the convergence and give
an a priori error estimate.
Finally we show, by means of some numerical experiments, that the proposed procedure is also useful in order to
approximate the solution of Fredholm integral equations of the second kind in which the right-hand sides and/or the
kernels are singular at the origin.
The paper is structured as follows. In the first section we consider the equation on the real semiaxis giving some basic
facts in Subsection 2.1 and showing the main results in Subsection 2.2. In Section 3 we consider the equation on the real
line while in Section 4 we give an application with some numerical tests. Finally Section 5 is devoted to the proofs of the
main results.

2. The case (0, ∞)

In this section we consider Eq. (1.1) with (a, b) = (0, ∞). For the sake of simplicity, but without loss of the generality,
we set

p( y) = yδ , k(x, y ) = y λ logℓ y ko (x, y ), 0 < λ, δ < 1, ℓ  0,

where ko (x, y ) is a smooth function.


With this notation Eq. (1.1) becomes

∞
δ λ ℓ
y f ( y ) − μ y log y ko (x, y ) f (x) w (x) dx = g ( y ), (2.2)
0

i.e. a Fredholm Integral equation of the third kind in which the kernel has a low smoothness w.r.t. the variable y at the
origin and is singular in x = 0 if the parameter α of the weight w is negative.
The goal of this section is to introduce a numerical approach that allow us to approximate the solution (if it exists)
of this equation. Nevertheless, we have first to introduce the space in which we will consider it. This is the aim of the
following subsection.

2.1. Basic facts

β 1
Let u (x) = (1 + x)ρ xγ e −x /2 , x, ρ , γ  0, β > 2
and we define the space
   
Cu = f ∈ C (0, ∞) : lim ( f u )(x) = 0 , (2.3)
x→∞
x→0

where C ( J ) denotes the collection of all continuous functions on J ⊆ [0, ∞).


If γ = 0, the space C u consists of all continuous functions on [0, ∞) such that limx→∞ ( f u )(x) = 0. This space equipped
with the following norm
 
 f C u =  f u ∞ = sup( f u )(x)
x0

is a Banach space.
In order to introduce a subspace of C u we proceed as in [21] defining the following main part of the modulus of
smoothness

Ωϕr ( f , τ )u = sup  Δrhϕ f u  I ,


  
rh
0<hτ
2972 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989


where r  1, ϕ (x) = x,  ·  I rh denotes the uniform norm on the interval I rh = [8r 2 h2 , C h∗ ], h∗ = h−2/(2β−1) , C is a fixed
constant and
r



r r
Δrhϕ f (x) = (−1)i f x+ − i hϕ (x) .
i 2
i =0

We recall that the introduced modulus Ωϕr ( f , τ )u , can be estimated by means of the following estimate [21]

Ωϕr ( f , τ )u  C sup hr  f (r ) ϕ r u  I ,
 
C
= C ( f , τ ) (2.4)
rh
0<hτ

if the norm at the right-hand side is bounded.


Using Ωϕr ( f , τ )u we introduce the Zygmund-type spaces

Ωϕr ( f , τ )u

Z s,r (u ) = f ∈ C u : sup < ∞, r > s > 0 (2.5)
τ >0 τs
equipped with the norm

Ωϕr ( f , τ )u
 f  Z s,r (u ) =  f u ∞ + sup .
τ >0 τs
For simplicity of notation we set Z s (u ) := Z s,r (u ).
Now let Pm the set of all algebraic polynomials of degree at most m and we denote by
 
E m ( f )u = inf ( f − P )u ∞
P ∈Pm

the error of best approximation of f ∈ C u . It is well known [21] that for all f ∈ C u the following weak Jackson and Salem–
Stechkin inequalities hold true

am /m
Ωϕr ( f , t )u
E m ( f )u  C dt , (2.6)
t
0


√ m

r r
am am k+1 E k ( f )u
Ωϕr f , C √ , (2.7)
m u m ak k+1
k =0

where in both cases r < m, C


= C (m, f ) and

1/β
1/β
Ŵ(β) 2γ + 1
am := am (u ) = 4 1+ m1/β (2.8)
Ŵ(2β) 4m
is the Mhaskar–Rahmanov–Saff number related to the weight u (see, for instance, [20,23]). Here and in the sequel C denotes
a positive constant which may be distinct in different formulas. We write C
= C (a, b, . . .) to say that C is a constant inde-
pendent of the parameters a, b, . . . and we write C = C (a, b, . . .) to say that C depends on a, b, . . . , . Moreover if A , B > 0
are quantities depending on some parameters, we will write A ∼ B if there exists a positive constant C independent of the
parameters of A and B such that C −1  BA  C .
Finally we note that if f ∈ Z s (u ) by (2.6) we deduce that

√ s
am
E m ( f )u  C  f  Z s (u ) , m > s, (2.9)
m
where C
= C (m, f ).

2.2. Main results

The aim of this subsection is to approximate the solution of Eq. (2.2) (if it exists) in the space C u introduced in the
previous section.
To this end we multiply and divide in the integral by the function y δ . Then setting f¯ ( y ) = f ( y ) y δ we obtain

∞
ko (x, y ) β
f¯ ( y ) − μ y log y
λ ℓ
f¯ (x)xα e −x dx = g ( y ), δ < α + 1, (2.10)

0
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2973

i.e. we get a Fredholm integral equation of the second kind in which the kernel has a low smoothness with respect to the
variable y.
Now, we could apply a Nyström method based on the zeros of the orthogonal polynomials with to respect the generalized
β
Laguerre weight xα −δ e −x (see, for instance, [13]). Nevertheless, as already noted, the kernel has a low smoothness with
respect to the external variable, and this produces negative effects on the order of convergence. Therefore, we propose to
regularize the equation before applying a Nyström method to the obtained equation.
To this end we introduce the following one-to-one map γq : [0, ∞) → [0, ∞) defined as

γq (t ) = t q/λ , (2.11)
where 1  q ∈ N and λ is the parameter that appears in Eq. (2.10).
Then we make the change of the variable x = γq (t ) and y = γq (s) in Eq. (2.10). In this way we obtain

∞
F (s) − μ h(t , s) F (t ) w o (t ) dt = G (s) (2.12)
0

where

F (s) = f¯
   
γq (s) , G (s) = g γq (s) , (2.13)
q qβ/λ
h(t , s) = ko γq (t ), γq (s) t [η] sq logℓ sq/λ , w o (t ) = t η−[η] e −t
   
(2.14)
λ
q
with η = λ (1 + α − δ) − 1, δ < α + 1, β > 12 and [η] denotes the integer part of η .
We immediately note that the right-hand side and the kernel of the new equation are smooth functions.
Now we go to consider Eq. (2.12) in the space C v with
ρ q qβ/λ
γq (s) s−qδ/λ = 1 + sq/λ s λ (γ −δ) e−s /2
  
v (s) = u , (2.15)
β
where u (s) = (1 + s)ρ sγ e −s /2 , β > 12 , ρ > 0, δ < γ .
The advantage of studying Eq. (2.12) in this space is that it is possible to prove that (2.12) has a unique solution in C v if
and only if the original equation (1.1) is unisolvent in C u .
Nevertheless in order to prove this, we need two additional results. The first one concerns with the smoothness prop-
erties of the given functions of the regularized equation in the space C v , while the second one assures the compactness of
the operator H defined as

∞
( H F )(s) = μ h(t , s) F (t ) w o (t ) dt . (2.16)
0

In the sequel we denote by kx (respectively k y ) the function k(x, y ) as a function of the only variable y (respectively x).

β
Proposition 2.1. Let u (s) = (1 + s)ρ sγ e −s /2 and v as in (2.15) with β > 21 , ρ > 0 and γ > δ . If the known functions of the original
equation are such that
g
∈ Z r (u ) < ∞, (2.17)
p
 
 kx 
sup u (x)  < ∞, (2.18)
x0
  p Z r (u )
 
ky 
sup u ( y )  <∞ (2.19)
y 0
 p Z r (u )

then the known functions of the regularized equation (2.12) satisfy

G ∈ Z qr ( v ) < ∞, (2.20)
λ
sup v (t )ht  Z qr ( v ) < ∞, (2.21)
t 0 λ

sup v (s)h s  Z qr ( v ) < ∞, (2.22)


s0 λ

where r > 0 and q and λ are the parameters appearing in the transformation γq defined in (2.11).
2974 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989

β
Proposition 2.2. Let u (s) = (1 + s)ρ sγ e −s /2 , β > 21 , v as in (2.15) with ρ and γ such that

1 δ+α α+1
ρ> , max δ, <γ < . (2.23)
2 2 2
Then if the kernel kx satisfies (2.18) the operator H : C v → C v defined in (2.16) is compact and for (2.12) the Fredholm alternative
holds true in C v .

We remark that, if α < δ , in order to find a value of γ , is necessary that 0 < δ < α +1 .
2
Now we are able to state the following.

Proposition 2.3. Let u and v be as in Proposition 2.2 and let (2.18) satisfied. Then the original equation (1.1) has a unique solution
f ∗ ∈ C u for each given right-hand side g ∈ C u if and only if the regularized equation (2.12) has a unique solution F ∗ ∈ C v for each
given right-hand side G ∈ C v . Moreover the following relation holds true

( f ∗ u )(t ) = ( F ∗ v ) γq−1 (t )
 
(2.24)
for each point t ∈ [0, ∞).

2.2.1. The numerical method


In order to approximate the solution of (2.12) we apply a Nyström method. To this end we approximate the integral that
appears in the equation by the following truncated Gaussian rule (see, for instance, [14,15,6])
∞ j


h(t , s) F (t ) w o (t ) dt = λk ( w o )h(xk , s) F (xk ) + em (h y F ), (2.25)
0 k =1

where xk = xm,k ( w o ), k = 1, . . . , j, are the zeros of the polynomial pm ( w o ) which is orthonormal with respect to the weight
w o , λk , k = 1, . . . , j, are the Christoffel numbers corresponding to w o ,
 
x j = min xk : xk  θ am ( w o ) , 0 < θ < 1, (2.26)
1km
∗ (h F ) is the remainder term.
am ( w o ) denotes the Mhaskar–Rahmanov–Saff number related to the weight w o and em y
∞ w o (t )
We recall that, under the assumption 0 v 2 (t )
dt < ∞, the following estimate holds true [13]

+ e − Am h y F v 2 ∞ ,
 ∗     
e (h y F )  C E M (h y F )
m v2 (2.27)
θ β
where the constants C and A are independent of m and f and M = [( 1+θ ) m].
Using (2.25), Eq. (2.12) becomes
j
v (s)
( F m v )(s) − μ λk ( w o )h(xk , s) ( F m v )(xk ) = (G v )(s), (2.28)
v (xk )
k =1

where the weight v is chosen as in Proposition 2.2.


Now collocating the equations on the zeros xi , i = 1, . . . , j, we obtain the following linear system
j  
v ( xi )
δi ,k − μλk ( w o ) h(xk , xi ) bk = (G v )(xi ), i = 1, . . . , j , (2.29)
v (xk )
k =1

where δi ,k is the Kronecker symbol and bk = F m (xk ) v (xk ), k = 1, . . . , j, are the unknowns. Now if the above system has a
unique solution [b∗1 , . . . , b∗j ] T , then we can construct the following weighted Nyström interpolant

j

v (s)
Fm (s) v (s) = μ λk ( w o ) h(xk , s)bk∗ + G (s) v (s). (2.30)
v (xk )
k =1

Hence, in order to obtain the approximate solution of (2.12) we have to solve a linear system of j equations in j
unknowns rather than a system of m equations in m unknowns and this implies a significant economy in the computations.
Moreover we remark that system (2.29) can be easily constructed because it only requires the computation of the zeros
xk , k = 1, . . . , j, and of the Christoffel Numbers λk ( w 0 ), k = 1, . . . , j. To this end one can use, in the Laguerre case, the
routine gaussq (see [10]) or routines recur and gauss (see [11] and [12]), and in the general case, the Mathematica Package
“OrthogonalPolynomials” (see [3]).
Next theorem states the conditions on the parameters of the weight v and on the given functions of Eq. (2.12) ensuring
that system (2.29) is unisolvent and that, in the norm of C v , the sequence { F m ∗ } converges to the exact solution F ∗ .
m
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2975

Theorem 2.1. Assume that the hypothesis of Proposition 2.1 are satisfied, and that Eq. (1.1) is uniquely solvable in C u for each given
right-hand side g ∈ C u with u as in (2.23). Then Eq. (2.12) has a unique solution F ∗ ∈ C v with v defined as in (2.15).
Moreover, for m sufficiently large, system (2.29) is unisolvent and its matrix B j is well conditioned holding

cond(B j )  C , (2.31)
1
where C does not depend on m and cond(B j ) = B j ∞ B−
j
∞ .
Finally it results

√ qr /λ
 ∗ am
 F − F ∗ v  F ∗  Z qr ( v ) sup v (s) h s v ∞ + ks (γq ) Z
     
m  CA , (2.32)
∞ m q/λ ( w (γq ))
λ s0

β
where w (x) = (1 + x)ρ x[q/λ] e −x /2 , F m
∗ is the Nyström interpolant defined in (2.30), C
= C (m, F ∗ ), A = A(q, λ), A
= A(m, F ∗ ) and

the norms at the right-hand sides are bounded.

The choice of the parameter q: By estimate (2.32), it follows that, for any constant C and A independent of m, the error

tends to zero as ( am /m)qr /λ , since theoretically we can choose m sufficiently large. Moreover it would appear that this rate
of convergence increases as q does. Unfortunately the constant A graves on the speed of convergence. In fact A depends
on q and if A becomes large as q does, the numerical convergence is compromised because we need a number of points
m very large in order to realize the required convergence. But numerically we don’t take m very large because we have to
calculate the zeros of orthogonal polynomials and the Christoffel numbers. Now the following estimate can be given.

q q q q
Proposition 2.4. Let q  1, 0 < λ < 1. Then, by denoting with [ λ ] the integer part of λ and with W ([ λ ]) the [ λ ]th Bell number, the
following estimate

  [q/λ]
 
q q
A +1 W
λ λ
holds true.
n
We recall that in general the nth Bell number is defined as (see, for instance, [2, p. 210]) W (n) = k=1 S (k, n) where
S (k, n) are the Stirling numbers. In order to compute it, one can use the following recursion formula (see again [2, p. 210])
n

n
W (n + 1) = W (k), W (0) := 1
k
k =0

or the tables available on the main part of the books of combinatorial.


Hence, by estimate (2.32) and by Proposition 2.4 it follows that

  [q/λ]
 
√ qr /λ
 ∗ q q am
 F − F ∗ v  F ∗  Z qr ( v ) sup v (s) h s v ∞ + ks (γq ) Z
     
m C +1 W .
∞ λ λ m q/λ ( w (γq ))
λ s0

Therefore, if we want approximate the solution with a correct digits by using m0 points, than q have to be the smallest
natural number such that the following relation
[q/λ]
 
√ qr /λ
am0

 
q q 1
+1 W < (2.33)
λ λ m0 10a+1
holds true.
In summing up, given an integral equation, we choose as optimal parameter q the smallest number such that inequality
(2.33) is satisfied, we fix the weight v and consequently the space C v in which we consider the regularized equation, and
we apply our numerical method in order to approximate the solution of the original equation.
Finally, we underline that the influence that a parameter has on the speed of convergence, and then the relative choice of
this, is a problem that frequently appears in the theory of polynomial approximation. For instance in [−1, 1] if we consider
the function f (x) = log (1 + x) in the weighted space C v , v (x) = (1 − x)γ (1 + x)δ , γ , δ  0 equipped with the uniform norm,
then there exists a polynomial P (see, for instance, [19]) such that

[ f − P ] v γ ,δ   C (r − 1)! log m ,
 
∞ r
δ > r > 0.
m
Therefore, also in this case, a constant depending on a parameter (that increases as it does) appears in the error estimate
and influences the speed of convergence. And also here the optimal choice is the smallest natural number r such that by
using m0 points we get the precision that we want.
2976 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989

Table 1

m j ( f m∗ u )(0.3) ( f m∗ u )(2) cond


8 7 3.5 14.8 1.426648083706660
16 13 3.553 14.82 1.404869049442987
32 26 3.55381 14.827 1.441612079056785
64 50 3.55381853 14.8272989 1.442444526265630
100 78 3.553818531972006 14.8272989504997 1.444806928716694
256 199 3.553818531972006 14.8272989504997 1.445997982173338

∗ u )( y ).
Fig. 1. ( f 100

Example 1. Consider the equation


∞
2/9 y 2/9
y f ( y) − sin (x + y ) f (x)x−1/3 e −x dx = y 2 + 4 y + 2
13
0

in the weighted space C u with u (x) = (1 + x)0.9 x0.3 e −x/2 .


In virtue of the numerical procedure showed in Section 2, the given equation is equivalent to the following
∞
9qsq 9 9 9
t 2q−1 sin t 2 q + s 2 q F (t ) w (t ) dt = s9q + 4s 2 q + 2
 
F (s) −
26
0
9q
with w (t ) = e −t . 2

Assume that with m0 = 100 points we want to approximate the solution with 15 correct digits. Then taking into account
that the given functions belong to the Zygmund type space Z q r ( v ) with r = 2.15 and by using (2.8), we can immediately
λ
note that if q = 1 relation (2.33) is satisfied.
Hence we fix q = 1, and by using system (2.29) we construct the sequence { F m ∗ v }, v (s) = u (s9/2 )s−1 and then the original

weighted solution f m∗ u according to (2.24).


Table 1 shows the weighted approximate solution in the points y = 0.3 and y = 2 with θ = 0.7. We think as exact the
solution obtained with m = 256 points and we only show the digits which are correct according to them.
Moreover if we perform the values f m∗ u in 20 equispaced points { y i }20 i =1
of the interval [0, 10] we found that the absolute
∗ u )( y ) − ( f ∗ u )( y )| is of the order 3.55 × 10−15 .
error maxi |( f 256 i 100 i
The graph of the weighted approximated solution ( f 100 ∗ u )( y ) is given in Fig. 1.

3. The case (−∞, +∞)

We consider the following integral equation

+∞
β
p( y) f ( y) − μ k(x, y ) f (x)|x|α e −|x| dx = g ( y ), y ∈ R, (3.34)
−∞
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2977

where μ ∈ R, α > −1 and β > 1. As in the previous section, without loss of the generality, we will assume that
p ( y ) = | y |δ , k(x, y ) = | y |λ logℓ | y |ko (x, y ), 0 < λ, δ < 1, ℓ  0
where ko (x, y ) is a smooth function.
We would like approximate the solution of this equation in the Banach weighted space
  
Cν = f ∈ C (R)  lim ( f ν )(x) = 0 ,

|x|→∞

equipped with the norm


 
 f C ν =  f ν ∞ = max( f ν )(x),
x∈R
β
where ν (x) = (1 + |x|)σ |x|ζ e −|x| /2 ,
σ , ζ  0, x ∈ R.
In order to approximate the solution of (3.34) we first multiply and divide by | y |δ inside the integral. In this way, we
obtain
+∞
β
f¯ ( y ) − μ| y |λ logℓ | y | ko (x, y ) f¯ (x)|x|α −δ e −|x| dx = g ( y ), δ < α + 1,
−∞

i.e. a Fredholm integral equation of the second kind in which we have denoted by f¯ ( y ) = f ( y )| y |δ the new unknown
function.
Now we could apply a Nyström method (see, for instance, [17]) based on the zeros of the orthogonal polynomials w.r.t.
β
the generalized Freud weight |x|α −δ e −|x| but the low smoothness of the kernel induces a slow order of convergence. Then
to overcome this problem, following an idea in [7], we propose to proceed in the following way.
We rewrite the previous equation as

∞ ∞
β β
f¯ ( y ) − μ| y |λ logℓ | y | ko (x, y ) f¯ (x)xα −δ e −x dx − μ| y |λ logℓ | y | ko (−x, y ) f¯ (−x)xα −δ e −x dx = g ( y ), (3.35)
0 0

and we consider it separately for y ∈ [0, ∞) and for y ∈ (−∞, 0]. Thus splitting the functions f¯ and g by means of the new
functions

f¯ 1 ( z) = f¯ ( z),

f¯ 1 , f¯ 2 : [0, ∞) → R defined by z ∈ [0, ∞),
f¯ 2 ( z) = f¯ (− z),
ḡ 1 ( z) = g ( z),

ḡ 1 , ḡ 2 : [0, ∞) → R defined by z ∈ [0, ∞),
ḡ 2 ( z) = g (− z),
Eq. (3.35) turns to be equivalent to the following system of two Fredholm integral equations of the second kind

∞ ∞


α −δ −xβ β
¯ λ ℓ ¯ dx − μ z log z ko (−x, z) f¯ 2 (x)xα −δ e −x dx = ḡ 1 ( z),
λ ℓ

f 1 ( z) − μ z log z ko (x, z) f 1 (x)x e






0 0
(3.36)

⎪ ∞ ∞
α −δ −xβ β
¯ λ ℓ ¯ dx − μ z log z ko (−x, − z) f¯ 2 (x)xα −δ e −x dx = ḡ 2 ( z).
λ ℓ

f 2 ( z) − μ z log z ko (x, − z) f 1 (x)x e





0 0
β
Also here we could apply a Nyström method based on the orthogonal polynomials with respect to the weight xα −δ e −x
(see, for instance, [4]). Nevertheless since the kernels have still a low smoothness with respect to the variable y, in order to
improve the order of convergence, we first regularize the kernels.
To this end we use the transformation (2.11) and we make the change of the variable x = γq (t ) and y = γq (s). Then we
obtain the following system

∞ ∞


1,1
h1,2 (t , s) F 2 (t ) w o (t ) dt = G 1 (s),




⎪ F 1 ( s ) − μ h (t , s ) F 1 (t ) w o (t ) dt − μ


0 0
(3.37)

⎪ ∞ ∞
⎪ 2 ,1 2,2
⎪ F 2 (s) − μ h (t , s) F 1 (t ) w o (t ) dt − μ h (t , s) F 2 (t ) w o (t ) dt = G 2 (s),




0 0
2978 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989

where
qβ/λ q
w o (t ) = t η−[η] e −t , η = (1 + α − δ) − 1,
λ
G i (s) = ḡ i (−1)i +1 γq (s) , F i = f¯ i (−1)i +1 γq (s) ∀i = 1, 2
   

q
h i ,n (t , s) = sq logℓ sq/λ t [η] ko (−1)i +1 γq (t ), (−1)n+1 γq (s) ∀i , n = 1, 2.
   
λ
We observe that the kernels, the right-hand sides and the weight function w o have the same shape of those of Eq. (2.12)
considered in the previous section. Consequently, we will consider this system in a product weighted space in which the
corresponding weight function will be similar to the weight v defined in (2.15). In fact, in this way, the given functions will
have the same smoothness properties of those of Eq. (2.12), each integral operator will be compact and the system will be
unisolvent if and only if Eq. (3.34) has a unique solution in C ν .
But before going on we introduce some notations.
Setting

F 1 (s) h1,1 (t , s) h1,2 (t , s)





F( s ) = , h(t , s) = , (3.38)
F 2 (s) h2,1 (t , s) h2,2 (t , s)
G 1 (s) w o (t ) 0



G( s ) = , wo (t ) = , (3.39)
G 2 (s) 0 w o (t )
we can write system (3.37) in vectorial form as
∞
F( s ) − μ h(t , s)F(t )wo (t ) dt = G(s), s ∈ [0, ∞).
0

Moreover, defining the operator matrices


H 1,1 H 1,2 I [0,∞) →[0,∞) 0



H= , I= ,
H 2,1 H 2,2 0 I [0,∞) →[0,∞)

where I [0,∞)→[0,∞) is the identity operator and the operators H i , j are defined as

∞
H i , j F j (s) = μ h i , j (t , s) F j (t ) w o (t ) dt ,
 
i , j = 1, 2,
0

we can rewrite system (3.37) as

(I − H)F = G. (3.40)

Now we go to consider (3.40) or equivalently (3.37) in the product space C v o × C v o equipped with the norm
 
FC v o ×C v o = max  F 1 C v o ,  F 2 C v o
where

γq (s) s−qδ/λ , s ∈ (0, ∞).


 
v o (s) = ν (3.41)

The following proposition holds true.

β
Proposition 3.5. Let ν (x) = (1 + |x|)σ |x|ζ e −|x| /2 , v o as in (3.41) with β > 1, σ > 0 and ζ > δ . If the given functions of the original
equation are such that
g
∈ Z r (ν ), (3.42)
p
 
 kx 
sup ν (x) p
 < ∞, (3.43)
x∈R Z r (ν )
 
ky 
sup ν ( y )  <∞ (3.44)
y ∈R
  p Z r (ν )

then ∀i , n = 1, 2 the known functions of system (3.37) are such that


L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2979

G i ∈ Z qr ( v o ), (3.45)
λ

sup v o (s)h i ,n 
 
s Z qr ( v )
< ∞, (3.46)
s0 λ o
i ,n
 
sup v o (t )ht  Z qr < ∞, (3.47)
t 0 λ (vo )

where r > 0 and q and λ are the parameters defining γq in (2.11).

As we have already mentioned, the advantage in studying system (3.37) in the space C v o × C v o , with v o as in (3.41) is
that (3.37) has a unique solution in this space if and only if the original equation (3.34) is unisolvent in C ν . In fact we can
state the following.

β
Proposition 3.6. Let ν (x) = (1 + |x|)σ |x|ζ e −|x| /2 with

1 α+δ α+1
σ> , max δ, <ζ < , (3.48)
2 2 2
v o as in (3.41) and let (3.43) satisfied. Then

(1) The operator H : C v o × C v o → C v o × C v o is compact and for (3.37) the Fredholm alternative holds true in C v o × C v o .
(2) System (3.37) has a unique solution
F 1∗ (s)


F∗ ( s ) =
F 2∗ (s)
in C v o × C v o for each given right-hand side G ∈ C v o × C v o if and only if Eq. (3.34) is unisolvent in C ν for each given right-hand
side g ∈ C ν . Moreover the following relation holds true

( F 1∗ v o ) γq−1 ( y ) , y ∈ [0, ∞),


  

( f ν )( y ) = (3.49)
( F 2∗ v o ) γq−1 ( y ) , y ∈ (−∞, 0],
 

where f ∗ is the unique solution of Eq. (3.34).

Note that, if α < δ , in order to find a value for the parameter ζ , is necessary that δ < α +2
1
.
To approximate the solution of (3.37) we apply a Nyström method. So, proceeding as in [4], we approximate each integral
of system (3.37) by means of the truncated Gaussian rule defined in (2.25). Then we multiply both equations by the weight
v o chosen as in Proposition 3.6 and collocate the equations on the zeros xi , i = 1, . . . , j, of the polynomial pm ( w o ) where
x j is defined in (2.26).
Hence we obtain the following system of order 2 j
⎧ j   j
⎪ v o (xi ) 1,1 v o (xi ) 1,2
δ − λ ( w ) h ( x , x ) c − λk ( w o ) h (xk , xi )dk = (G 1 v o )(xi ),

i ,k k o k i k

v o (xk ) v o (xk )



k =1 k =1
(3.50)
j j  
v o (xi ) 2,1 v o (xi ) 2,2


i = 1, . . . , j ,


⎪ − λk ( w o ) h ( x , x
k i k)c + δ i ,k − λk ( w o ) h ( x ,
k ix ) dk = (G 2 v o )(xi ),
v (x ) v (x )

o k o k
k =1 k =1

where δi ,k is the Kronecker symbol, λk ( w o ) are the Christoffel numbers, ck = ( F m,1 v o )(xk ) and dk = ( F m,2 v o )(xk ), k =
1, . . . , j.
Now if for a sufficiently large m (say m > mo ) system (3.50) admits the unique solution (c 1∗ , . . . , c ∗j , d∗1 , . . . , d∗j ), then we
can construct the sequence

∗ (s)
Fm

∗ ,1
Fm (s) = , (3.51)
F m,2 (s)
with
j  
v o (s) v o (s)

h1,1 (xk , s)ck∗ + λk ( w o ) h1,2 (xk , s) dk∗ + (G 1 v o )(s),
 
Fm ,1 v o (s) = μ λk ( w o )
v o (xk ) v o (xk )
k =1
j  
v o (s) v o (s)

h2,1 (xk , s)ck∗ + λk ( w o ) h2,2 (xk , s) dk∗ + (G 2 v o )(s).
 
Fm ,2 v o (s) = μ λk ( w o ) (3.52)
v o (xk ) v o (xk )
k =1
2980 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989

Next theorem concerns with the convergence and the stability of the proposed procedure. In order to state it we need some
notations.
We will denote by C j the matrix of coefficients of system (3.50) and by cond(C j ) its condition number in infinity norm.
Moreover we will set

+∞
β
( K f )( y ) = μ k(x, y ) f (x)|x|α e −|x| dx, y∈R (3.53)
−∞

and we will denote by F Z s ( v o )× Z s ( v o ) the norm of the product space Z s ( v o ) × Z s ( v o ) defined as
 
F Z s ( v o )× Z s ( v o ) = max  F 1  Z s ( v o ) ,  F 2  Z s ( v o ) .

Theorem 3.2. Let Ker{ I − K } = {0} in C ν with ν as in (3.48) and let f ∗ be the unique solution of Eq. (3.34) for each given g ∈ C ν .
Then under the assumptions of Proposition 3.5, system (3.37) has a unique solution F∗ ∈ C v o × C v o .
Moreover, for m sufficiently large, system (3.50) is unisolvent and

cond(C j )  C (3.54)

where C does not depend on m.


Finally the following order estimate holds true

√ qr /λ
 ∗ am
 F − F∗  F∗  Z qr ( v o )× Z qr ( v o ) ,

m C v o ×C v o
 CA (3.55)
m λ λ

∗ is the array defined in (3.51), C


= C (m, F∗ ) and A is a constant that depends on q, λ and is independent of m and F∗ .
where Fm

We recall that the constant A was estimated in Proposition 2.4. Consequently also in this case we have to choose q as
suggested in the previous section.

4. Some applications

The numerical treatment of the Fredholm integral equations of the third kind considered in the previous sections is
useful in order to solve some Fredholm integral equations of the second kind in which the right-hand side and/or the
kernel are singular at the origin.
In fact, we consider the following Fredholm integral equation

∞
f ( y) − μ k(x, y ) f (x) w (x) dx = g ( y ), y ∈ I, (4.56)
0
β
where f is the unknown, k and g are given functions, μ ∈ R and w (x) = xα e −x , α > −1, β > 12 . Moreover we assume that
the right-hand side (and could happen also for the kernel) is singular in zero.
The case when the kernel and the right hand-side are smooth functions was recently investigated in [13] where the
authors proposed a projection and a Nyström method approximating the solution of Eq. (4.56) in the space C u defined
in (2.3).
In particular the authors in [13] choose the parameters of the weight u as in (2.23) and solve the following system

j  
u( yi )
δi ,k − μλk ( w ) k( yk , y i ) ak = ( gu )( y i ), (4.57)
u ( yk )
k =1

where δi ,k is the Kronecker symbol, yk , ∀k = 1, . . . , j, are the zeros of pm ( w ), λk ( w ) are the corresponding Christoffel
numbers and j is chosen as in (2.26) with xk = yk and w o = w.
Moreover they prove that if the original equation is unisolvent in C u and if the given functions are such that

g ∈ Z r (u ), (4.58)
sup u (x)kx  Z r (u ) < ∞, (4.59)
x0

sup u ( y )k y  Z r (u ) < ∞ (4.60)


y 0
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2981

Table 2

m j ( f m∗ u )(0.1) ( f m∗ u )(1) cond(A j )


8 7 3.83 6.5 1.866243873297617
16 14 3.83 6.5 1.934678975648094
32 27 3.83 6.5 1.972649626291241
64 53 3.835 6.54 1.992337024184316
128 106 3.8351 6.54 2.014916131940892
256 210 3.8351 6.54 2.024463588874378

then system (4.57) has a unique solution (a∗1 , . . . , a∗j ) and the Nyström interpolant

j
u( y)
f m∗ u ( y ) = μ k( yk , y )ak∗ + ( gu )( y ),
 
λk ( w )
u ( yk )
k =1

converges to the exact solution f ∗ with order



√ r
 ∗
 f − f ∗ u  C am
 f ∗  Z r (u ) ,
 
m ∞
C
= C (m, f ). (4.61)
m
Thus we can see that if the value of the Zygmund parameter r is small (i.e. when the kernel and/or the right-hand side
have a low smoothness) the approximate solution converges to the exact solution with a very slow of convergence.
Take for instance the following equation

∞
μ g ∗( y)
f ( y) − k∗ (x, y ) f (x) w (x) dx = , ǫ < δ, (4.62)
yǫ yδ
0

where k∗ and g ∗ are smooth functions. Now since by (2.4) it results

Ωϕr ( g , τ )u ∼ t 2(γ −δ) , Ωϕr (kx , τ )u ∼ t 2(γ −ǫ ) , γ >δ


we have that g ∈ Z 2(γ −δ) (u ), k ∈ Z 2(γ −ǫ ) (u ) and by (4.61) we get

√ 2(γ −δ)
 ∗ am
 f − f ∗ u  C  f ∗  Z 2(γ −δ) (u ) ,
 
m ∞
(4.63)
m
where C
= C (m, f ∗ ).
So we deduce that if δ ≃ γ the rate of convergence is very slow and then the numerical results are very poor.
Nevertheless we note that if we multiply both sides of Eq. (4.62) by y δ we obtain
∞
δ
y f ( y) − μ y λ
k∗ (x, y ) f (x) w (x) dx = g ∗ ( y ), λ=δ−ǫ
0

and i.e. Eq. (2.2) with ℓ = 0, ko = k∗ and g = g ∗ .


Therefore we can apply the numerical procedure showed in Section 2, getting in this way a satisfactory order of conver-
gence, according to (2.32).
The following numerical test confirms our theoretical expectations.
In the sequel we denote by A j the matrix of coefficients of system (4.57) and by cond(A j ) its condition number in infinity
norm. Moreover we think as exact the approximate solutions obtained for m = 256 and in all the tables we only show the
digits which are correct according to them. All the computations were performed in 16-digits arithmetic.

Example 2. Consider the equation


∞
1 ( y + 4)
f ( y) − y (x + 1) f (x)x1/5 e −x dx = √ ,
32 y
0

in C u with u (x) = (1 + x)x0.55 e −x/2 .


Using system (4.57), we can construct the polynomial sequence { f m∗ u } that converges to exact solution in C u with order
O( m11/20 ) according to (4.63). Table 2 shows the value of the approximate solution in the points y = 0.1 and y = 1 with
θ = 0.5.
2982 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989

Table 3

m j ( f m∗ u )(0.1) ( f m∗ u )(1) cond(B j )


4 4 3.835188951819742 6.545862279718702 1.920593003414122
256 183 3.835188951819742 6.545862279718702 2.040294868159992

Fig. 2. ( f 4∗ u )( y ).

Nevertheless by applying the procedure described in Section 2 we obtain the following equation
∞
2q 3q 2q
t 2q + 1 F (t )t 7q/5−1 e −t dt = s2q + 4
 
F (s) − s
32
0

that, in virtue of Proposition 2.3, is unisolvent in C v with v (s) = u (s2q )s−q .


Hence, using system (2.29) we can construct the sequence { F m ∗ v } and then the weighted solution of the original equation

according to (2.24).
Taking q = 1 and θ = 0.5 we obtain the results presented in Table 3.
Note that both the right-hand side and the kernel are polynomials and then the convergence is very fast. Indeed we get
the machine precision solving a system of order j = 4. We mention that in this case A  8.
Moreover if we perform the values f m∗ u in 20 equispaced points { y i }20 i =1
of the interval [0, 10] we found that the absolute
∗ u )( y ) − ( f ∗ u )( y )| is of the order 1.77 × 10−15 . The graph of the weighted approximated solution ( f ∗ u )( y )
error maxi |( f 256 i 4 i 4
is given in Fig. 2.

Of course the showed procedure turns to be useful also in the case when we have a Fredholm integral equation of the
second kind on the real line
+∞
β
f ( y) − μ k(x, y ) f (x)|x|α e −|x| dx = g ( y ) (4.64)
−∞
in which the given functions are singular at the origin.
In fact, a Nyström method applied to this equation gives satisfactory theoretical and numerical results when the kernel
and the right-hand side are smooth functions. This case was recently investigated in [17].
If we want apply the same procedure, proceeding as in [17], we have to solve the following system
 ν ( zi )

δi ,k − μλk k( z , zi ) ãk = ( g ν )( zi ), (4.65)
ν (zk ) k
|k| j
β /2
where ν (x) = (1 + |x|)σ |x|ζ e −|x| , zk , k = 1, . . . , [ m2 ], are the positive zeros of the mth polynomial which is orthonormal
β
w.r.t. the weight |x|α e −|x| , z−k := zk are the negative ones, λk are the corresponding Christoffel numbers and j is the index
such that
xj = min { zk : zk  θ B m },
1k[ m
2
]
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2983

Table 4

m j ( f m∗ ν )(2) ( f m∗ ν )(−2) cond(Ã j )


8 6 1.305 −0.435 1.052035892107371
16 12 1.305 −0.435 1.058235116682020
32 24 1.3054 −0.4351 1.059827729207075
64 46 1.3054 −0.4351 1.061797844552008
128 92 1.3054 −0.43514 1.062756887022350
256 184 1.3054 −0.43514 1.063375329287274

Table 5

m j ( f m∗ ν )(2) ( f m∗ ν )(−2) cond(C j )


4 4 1.305438886060509 −0.4351462953535030 1.050952442839758
256 180 1.305438886060509 −0.4351462953535030 1.064470315284681

with m even, θ ∈ (0, 1), and B m the Mhaskar–Rahmanov–Saff number related to the generalized Freud weight.
Under the assumptions (3.48) and (4.58)–(4.60) with u := ν , this system has a unique solution (ã∗1 , . . . , ãk∗ ) and we can
construct the Nyström interpolant
ν ( y)
( f˜ ν )( y ) = μ λk k( z , y )ãk∗ + ( g ν )( y )
ν (zk ) k
|k| j

that converges to the exact solution with an order that depends on the smoothness of the given functions. Therefore we can
deduce that if the original given functions have some singularities this order of convergence leads to very poor numerical
results.
Nevertheless if we apply the procedure showed in Section 3, we can improve the smoothness of the given functions and
subsequently the order of convergence.
The following example confirms our theoretical expectations.

Example 3. We consider the equation


∞ √
1 2 y+1 π
xy f (x) |x| e −x dx = √

f ( y) − − y,
48 | y| 96
−∞

whose exact solution is f ∗ ( y ) = ( y + 1)/ | y |.
2
Solving system (4.65) we can construct the sequence { f˜ m ν } with ν (x) = (1 + |x|)|x|3/5 e −x /2 , according to (3.48).
Table 4 shows the value of the approximated solution in the points y = ±2 with θ = 0.6 and the condition number in
infinity norm of the matrix of coefficients à j of system (4.65).

Multiplying both sides of equation by | y | we obtain the following
√ +∞ 
| y| 2 π 
xy |x|e −x f (x) dx = ( y + 1) −

| y| f ( y) − y | y |.
48 96
−∞

Therefore applying the procedure showed in Section 3, solving system (3.50), we can construct the polynomial sequences
{ F m,1 v o } and { F m,2 v o } defined in (3.52) with v o = ν (s2q )s−q and then the weighted solution of the original equation ac-
cording to (3.49).
The given functions of the regularized system are polynomials and then the convergence is very fast. The results (Table 5)
are obtained with q = 1 and θ = 0.6. We mention that in this case A  8.
Moreover if we perform the values f m∗ ν in 40 equispaced points { y i }40 i =1
of the interval [−10, 10] we found that the
∗ ν )( y ) − ( f ∗ ν )( y )| is of the order 2.08 × 10−17 .
absolute error maxi |( f 256 i 4 i

5. Proofs

−x β √ 2
Proof of Proposition 2.1. Let p ( y ) = y δ , u (x) = (1 + x)ρ xγ e 2 , ρ > 0, δ < γ , β > 21 , ϕ (x) = x and I rh = [8r 2 h2 , C h− 2β−1 ]
with C a fixed constant.
Since

′ 
 g
 C h2(γ −δ)−1 ,

 p ϕ u
 
I rh
2984 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989

g
we can deduce by (2.4) that p
∈ Z r (u ) with r = 2(γ − δ). Now for the sake of the simplicity we assume that the function g
q q
has n  [ λ ] derivatives. If n < [ λ ] it is possible to proceed in the same way. By the Faà di Bruno formula we have
( j )
G ( j ) (s) = g
 
γq (s)
j
 γq′ (s) k1 γq( j−k+1) (s)


k j−k+1
j!
g (k)

= γq (s) ···
k1 ! · · · k j −k+1 ! 1! ( j − k + 1)!
k =1 k1 ,k2 ,...∈{0}∪N
k1 +k2 +···=k
k1 +2k2 +···= j

j j −k+1
 n
kn
j!  1  q q
g (k) γq (s) s λ k− j ,
 
= −i+1
k1 ! · · · k j −k+1 ! n! λ
k=1 k1 ,k2 ,...∈{0}∪N n =1 i =1
k1 +k2 +···=k
k1 +2k2 +···= j

q
with 1  j  [ λ ].
Thus by estimate (2.4) we have
j
Ωϕ (G , τ ) v  sup h j G ( j ) (t )ϕ j (t ) v (t ) I
 
hj
0<hτ

j j −k+1
 n
kn
qr /λ
j!  1  q
sup  g (k) (γq )ϕ k (γq ) w (γq ) I
 
 Cτ −i+1
k1 ! · · · k j −k+1 ! n! λ 0<hτ
hj
k=1 k1 ,k2 ,...∈{0}∪N n =1 i =1
k1 +k2 +···=k
k1 +2k2 +···= j

β
with w (t ) = (1 + t )ρ t [q/λ]/2 e −t /2 .
Now since for every Banach space equipped with norm  ·  it results
m
 (s) s 
 g ϕ u  ∼  gu  +  g (m) ϕ m u ,
 

s =1

where u is an arbitrary weight function, we can write

j j j −k+1
 n kn
Ωϕ ( G , τ ) v

 
j!  1  q
sup C −i+2
τ >0 τ qr /λ k1 ! · · · k j −k+1 !
n =1
n!
i =1
λ
k=1 k1 ,k2 ,...∈{0}∪N
k1 +k2 +···=k
k1 +2k2 +···= j
 
× sup sup  g w  I hj + sup sup  g [q/λ] (γq )ϕ [q/λ] (γq ) w (γq ) I .
 
hj
τ >0 0<hτ τ >0 0<hτ

Thus setting

j j −k+1
 n
  kn
j!  1  q
Aj = −i+2
k1 ! · · · k j −k+1 ! n! λ
k =1 k1 ,k2 ,...∈{0}∪N n =1 i =1
k1 +k2 +···=k
k1 +2k2 +···= j
j j −k+1
kn
j!  [q/λ] + 1
= (5.66)
k1 ! · · · k j −k+1 ! n
k=1 k1 ,k2 ,...∈{0}∪N n =1
k1 +k2 +···=k
k1 +2k2 +···= j

and being (see, for instance, [18])

sup sup  g w  I hj + sup sup  g ([q/λ]) (γq )ϕ [q/λ] (γq ) w (γq ) I   g (γq ) Z q ( w (γ ))
   
hj q
τ >0 0<hτ τ >0 0<hτ λ

we have
j
Ωϕ ( G , τ ) v  
sup  CA j  g (γq ) Z q ( w (γ )) .
τ >0 τ qr /λ λ
q
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2985

Moreover since [13, p. 169]

Ωϕs (G , τ ) v  C Ωϕs−1 (G , τ ) v
we have
[q/λ]
Ωϕℓ (G , τ ) v Ωϕ (G , τ ) v   qr
sup qr /λ
 C sup qr /λ
 CA[q/λ]  g (γq ) Z q ( w (γ )) , ℓ> , (5.67)
τ >0 τ τ >0 τ λ
q λ
from which (2.20) follows.
In the same way it is possible to prove (2.21) and (2.22). ✷

Proof of Proposition 2.2. In order to prove the compactness of the operator H it is sufficient to prove (see, for instance,
[28, p. 44]) that

lim sup E m ( H F ) v = 0. (5.68)
m  F v ∞ =1

To this end we proceed as in [4]. Then we first observe that


∞
Δ ht (s) v (s) v (t ) w o (t ) dt ,
 r    r 
 Δ H F (s) v (s)   F v ∞
hϕ hϕ
v 2 (t )
0

from which taking the supremum on s ∈ I rh = [8r 2 h2 , C h−2/(2β−1) ] we get


∞
Δ ht v  v (t ) w o (t ) dt .
 r    r 
 Δ H F v   C  F v ∞
hϕ I hϕ I rh
rh v 2 (t )
0

Hence taking the supremum on 0 < h  τ , by (2.23) we obtain

Ωϕr ( H F , τ ) v  C  F v ∞ sup v (t )Ωϕr (ht , τ ) v .


t 0

Therefore by (2.6) and by applying Proposition 2.1 we deduce



am /m
Ωϕr (ht , τ ) v
E m ( H F ) v  C  F v ∞ sup v (t ) dτ
t 0 τ
0

and then (5.68).


In definitive the operator H : C v → C v is compact and we can assume that Eq. (2.12) is uniquely solvable in C v for a
fixed right-hand side G ∈ C v . ✷

Proof of Proposition 2.3. We first observe that proceeding as in the proof of Proposition 2.2 (see also [4]) it is possible
to prove that the integral operator of Eq. (1.1) is compact. Therefore for this equation the Fredholm alternative holds true
in C u .
Now we assume that the original equation (1.1) has a unique solution f ∗ ∈ C u and then that the homogeneous equation
has only the trivial solution f ∗ ≡ 0 in C u .
We consider the regularized homogeneous equation

( I − H ) F = 0. (5.69)
In virtue of Proposition 2.2 to obtain the thesis we have to prove that (5.69) has only the trivial solution F ∗ = 0 in C v . To
this end we assume that

F ∗ (s) = f¯ ∗ γq (s) = f ∗ γq (s) sqδ/λ


= 0
   

satisfies Eq. (5.69) in C v and show that this leads to a contradiction. It is easily to show that

f ∗ (x) = F ∗ γq−1 (x) x−δ


= 0
 

is the solution of the following integral equation


∞
β
δ ∗
y f ( y) − μ k(x, y ) f (x)xα e −x dx = 0.
0
2986 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989

Moreover since F ∗ ∈ C v we have that

lim ( f u )(x) = lim F ∗ γq−1 (x) x−δ u (x) = lim F ∗ v (s) = 0.


    
(5.70)
x→0 x→0 s→0
x→∞ x→∞ s→∞

In other words we have found a function f ∗


= 0 ∈ C u satisfying the homogeneous equation associated to Eq. (1.1) and this
leads to a contradiction.
The other implication can be proved in the same way. Moreover relation (2.24) follows by the adopted procedure. ✷

Proof of Theorem 2.1. The first part of the theorem follows by Proposition 2.3.
Now we set
j
v (s)
( H m F )(s) = μ λk ( w o )h(xk , s) ( F v )(xk ),
v (xk )
k =1

and for the sake of simplicity we will omit the subscript C v → C v in the norms of the operators, since all the operators
involved in the discussion are maps in these spaces.
In order to prove the convergence and the stability of the Nyström method it is sufficient to prove that

(1) [ H F − H m F ]u ∞ tends to zero ∀ F ∈ C v ;


(2) limm [ H − H m ] H m  = 0.

In fact by step (1), in virtue of the principle of uniform boundedness, we can deduce

sup  H m C v →C v < ∞. (5.71)


m

Then using a well known-result (see, for instance, [1]) under the assumptions (5.71) and (1) and (2), for m sufficiently large,
the operator ( I − H m )−1 exists and is uniformly bounded since

1 + ( I − H )−1   H m 
( I − H m )−1  
 
.
1 − ( I − H )−1  ( H − H m ) H m 
Moreover the following relation
   
[ F − F m ] v  ∼ [ H F − H m F ] v 
∞ ∞
(5.72)
holds true. Therefore now we prove (1) and (2). It results
∞ j

   
[ H F − H m F ] v  = sup v (s) h(x, y ) F (x) w o (x) dx −

λk ( w o )h(xk , s) F (xk )


s0  
0 k =1
 ∗ 
= sup v (s)e M (h s F ),
s0

where e ∗M (h s F ) is the remainder term of the Gaussian rule (2.25).


Now by applying (2.27) with M = am, 0 < a < 1, and taking into account that for all f , g ∈ C u it results
 
E m ( f g )u  C  f u  E m ( g ) + 2 g ∞ E m ( f )u (5.73)
we obtain

[ H F − H m F ] v   C sup v (s) E M (h s F ) 2 + e − Am h s F v 2 
     
∞ v ,∞ ∞
s0
! "
 C  F v ∞ sup v (s) E [ M ] (h s ) v + sup v (s)h s v ∞ E [ M ] ( F ) v +  F v ∞ sup v (s)h s v ∞ e − Am
2 2
s0 s0 s0

 C  F v ∞ sup v (s)h s  Z qr ( v ) (5.74)


s0 λ

and then applying (2.22) we have (1). Now we prove (2). By using (5.74) with H m F instead of F we obtain
  !
[ H − H m ] H m F v 

 C  H m F v ∞ sup v (s) E [ M ] (h s ) v + sup v (s)h s v ∞ E [ M ] ( H m F ) v
2 2
s0 s0
"
+  H m F v ∞ sup v (s)h s v ∞ e − Am . (5.75)
s0
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2987

Now, we observe that


∞
  w o (t )  
v (s)Δhϕ ( H m F )(s)  C  F v  v (t ) v (s)Δhϕ h(t , s) dt
v 2 (t )
0

from which we deduce

Ωϕ ( H m F , τ ) v  C  F v ∞ sup v (t )Ωϕ (ht , τ ) v .


t 0

Then by applying (2.21) and (2.6) we have that



√ qr /λ
am
E m ( H m F )v  C  F v ∞
m
and therefore by (5.75) and taking into account (2.9) and (2.22)

√ qr /λ
  am
[ H − H m ] H m F v  C  F v ∞
∞ m
from which the thesis follows.
About estimate (2.32) by applying (5.72) and (5.74) we have
! "
 C  F ∗ v ∞ sup v (s) E [ M ] (h s ) v + sup v (s)h s v ∞ E [ M ] ( F ∗ ) v +  F ∗ v ∞ sup v (s)h s v ∞ e − Am .
 ∗
 F − F ∗ v
 
m ∞ 2 2
s0 s0 s0

Since h, g ∈ Z qr (v ) then F ∗ ∈ Z ( v ) and by using (2.9) we get


qr
λ λ


√ qr /λ
am
[ F − F ∗ ] v   C  F ∗  Z qr (v ) sup v (s)h s  Z qr ( v )
 
∞ m λ s0 λ

√ qr /λ
Ω ℓ (h s , τ ) v
 
am ∗
C  F  Z qr (v ) sup v (s)h s v ∞ + sup v (s) sup
m λ s0 s0 τ >0 τ qr /λ
qr
with ℓ > λ . Now by estimating the last term at the right-hand side as in the proof of Proposition 2.1 with h s in place of
G, k s in place of g (see estimate (5.67)) and setting A := A[q/λ] we deduce

√ qr /λ
 ∗ am 
 F − F ∗ v  F ∗  Z qr (v ) sup v (s)h s v ∞ + sup v (s)ks (γq ) Z
   
m  CA .
∞ m q/λ ( w (γq ))
λ s0 s0

Finally about the well conditioning of the system it is sufficient to prove that

cond(B j )  cond( I − H m ) = ( I − H m )( I − H m )−1 .


  

To this end we can use the same arguments in [1, p. 113] only by replacing the usual infinity norm with the weighted
uniform norm of the space C v . ✷

Proof of Proposition 2.4. By definition (5.66) we have


q/λ]
[ [q/λ]! k+1

[q/λ]−
[q/λ] + 1
kn
A := A[q/λ] = .
k1 ! · · · k[q/λ]−k+1 ! n
k =1 k1 ,k2 ,...∈{0}∪N n =1
k1 +k2 +···=k
k1 +2k2 +···=[q/λ]

Now since in general


pq


p

q q!
we have

q/λ]
[ [ λq ]−k+1
kn
[q/λ]!  ([q/λ] + 1)n
A
k1 ! · · · k[q/λ]−k+1 ! n!
k =1 k1 ,k2 ,...∈{0}∪N n =1
k1 +k2 +···=k
k1 +2k2 +···=[q/λ]
2988 L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989


 
q
q/λ]
[q/λ] [ [q/λ]! k+1

[q/λ]−
1
kn
= +1
λ k1 ! · · · k[q/λ]−k+1 ! n!
k =1 k1 ,k2 ,...∈{0}∪N n =1
k1 +k2 +···=k
k1 +2k2 +···=[q/λ]

  q/λ]
[q/λ] [
q
= +1 B[q/λ],k (1, 1, . . . , 1),
λ
k =1

where in general Bn,k are the partial Bell polynomials (see, for instance, [2]). Then, the thesis follows taking into account
that (see again [2, p. 135 and p. 210]
q/λ]
[ q/λ]
[
 
 
q q
B[ q (1, 1, . . . , 1) = S ,k = W ,
λ ],k λ λ
k =1 k =1

where in general S (n; k) and W (n) denote the Stirling numbers and the nth Bell number, respectively. ✷

Proof of Proposition 3.5. By the assumptions we have

g ( y) g (− y )
g1 ( y ) = ∈ Z r ( v̄ ), g2 ( y ) = ∈ Z r ( v̄ ),
p( y) p (− y )
 i ,n   i ,n 
ky   kx 
sup v̄ ( y )  < ∞, sup v̄ (x)  <∞
y 0
  p Z r ( v̄ ) x0
 p
Z r ( v̄ )
β
with ki ,n (x, y ) = k((−1)i +1 x, (−1)n+1 y ), ∀i , n = 1, 2, y ∈ (0, ∞) and v̄ (x) = (1 + x)σ xζ e −x /2 . Therefore since the kernels h i ,n
and the right-hand sides G i have the same expressions of those of Eq. (2.12) ∀i , n = 1, 2, proceeding as in the proof of
Proposition 2.1 with u = v̄ and v = v o we obtain the assertion. ✷

Proof of Proposition 3.6. In order to prove that the operator H is compact it is sufficient to prove the compactness of
the operators H i ,n ∀i , n = 1, 2. To this end we can proceed exactly as in the proof of Proposition 2.2 with v := v o and
i ,n
H m := H m .
Finally statement (2) can be proved proceeding as in the proof of Proposition 2.3. ✷

Proof of Theorem 3.2. The first statement follows by Proposition 3.6.


Now in order to prove the second part we proceed as in the proof of Theorem 2.1 by proving steps (1) and (2) with Hm in
place of H m . Moreover in the following, in order to simplify the notation, we will omit the subscript C v o × C v o → C v o × C v o
in the norm of the involved operator. To prove (1) we observe that by the definition of the norm of the product spaces we
can write

[H − Hm ]F  max  H 1,1 − H 1,1  F 1 v o ∞ +  H 1,2 − H 1,2  F 2 v o ∞ ,


     
m m
 2,1
 H − H 2,2  F 1 v o ∞ +  H 2,2 − H 2,2  F 2 v o ∞ .
   
m m
i ,n
Now we can estimate [ H i ,n − H m ] F n v o ∞ proceeding as in step (1) of Theorem 2.1 obtaining that [H − Hm ]F tends to
zero ∀F ∈ C v o × C v o . Moreover we have

(H − Hm )Hm F  max  H 1,1 − H 1,1 H 1,1 F 1 v o  +  H 1,2 − H 1,2 H 2,1 F 1 v o 


       
m m ∞ m m ∞
+  H 1,1 − H m
1,1
 1,2
H m F 2 v o ∞ +  H 1,2 − H m
1,2
    2,2 
H m F 2 v o ∞ ,
 2,1
 H − H 2,1 H 1,1 F 1 v o  +  H 2,2 − H 2,2 H 2,1 F 1 v o 
    
m m ∞ m m ∞
+  H 2,1 − H 2,1 H 1,2 F 2 v o  +  H 2,2 − H 2,2 H 2,2 F 2 v o  .
      
m m ∞ m m ∞
i ,n i ,n n,ℓ
Therefore considering [ H − ∀i , n, ℓ = 1, 2, and proceeding as in step (2) of Theorem 2.1, we obtain that
H m ] H m F ℓ v o ∞
(H − Hm )Hm  tends to zero. Then the operator (I − Hm ) is invertible and uniformly bounded since

(I − Hm )−1  
  (I − H)−1 
.
1 − (I − H)−1 H − Hm 
In addition we have that

√ qr /λ
 ∗ am
F∗  Z qr ( v o )× Z qr ( v o ) ,
  
 F − Fm  ∼ [HF − Hm F]∞  C A
∞ m λ λ
L. Fermo / Applied Numerical Mathematics 59 (2009) 2970–2989 2989

i.e. (3.55).
About the estimate of the condition number of the system (3.50) we observe that denoting by
# 1,1 1,2 $
Cj Cj
Cj =
2,1 2,2
Cj Cj
i ,n
the matrix of coefficient of system (3.50) where C j are the blocks of the matrix related to the discretization of the corre-
i ,n
sponding operator H , we have that (see, for instance, [1]) ∀i , n = 1, 2
i ,n  i ,n i ,n  i ,n −1 
      
cond C j  cond I − H m = I − Hm I − Hm

and then the theorem is completed. ✷

Acknowledgements

The author is very grateful to Professor Giuseppe Mastroianni for his helpful suggestions and remarks and thanks also
the referees for their contribution in improving the paper.

References

[1] K.E. Atkinson, The Numerical Solution of Integral Equations of the Second Kind, Cambridge Monographs on Applied and Computational Mathematics,
University Press, Cambridge, 1997.
[2] L. Comtet, Advanced Combinatorics. The Art of Finite and Infnite Expansions, revised and enlarged edition, D. Reidel Publishing Co., Dordrecht, 1974.
[3] A.S. Cvetković, G.V. Milovanović, The Mathematica package “OrthogonalPolynomials”, Facta Univ. Ser. Math. Inform. 19 (2004) 17–36.
[4] M.C. De Bonis, G. Mastroianni, Nyström method for systems of integral equations on the real semiaxis, IMA J. Numer. Anal. 29 (July 2009) 632–650.
[5] M.C. De Bonis, G. Mastroianni, Some simple Quadrature Rules to evaluate the Hilbert Transform on the real axis, Arch. Inequal. Appl. 1 (3–4) (2003)
475–494.
[6] B. Della Vecchia, G. Mastroianni, Gaussian rules on unbounded intervals, J. Complexity 19 (3) (2003) 247–258.
[7] L. Fermo, M.G. Russo, A Nyström method for Fredholm integral equations with right-hand sides having isolated singularities, Calcolo 46 (2009) 61–93.
[8] N.S. Gabbasov, A special version of the collocation method for integral equations of the third kind, Differential Equations 41 (2006) 1768–1774.
[9] N.S. Gabbasov, S.A. Solov’eva, Generalized moment method for a class of integral equations of the third kind, Differential Equations 42 (2006) 1490–
1498.
[10] W. Gauschy, Algorithm 726:ORTHPOL – a package of routines for generating orthogonal polynomials and Gauss-type quadrature rules, ACM Trans.
Math. Softw. 20 (1994) 21–62.
[11] G.H. Golub, Some modified matrix eigenvalue problems, SIAM Rev. 15 (1973) 318–334.
[12] G.H. Golub, J.H. Welsch, Calculation of Gaussian quadrature rules, Math. Comput. 23 (1969) 221–230.
[13] G. Mastroianni, G.V. Milovanovic, Some numerical methods for second kind Fredholm integral equation on the real semiaxis, IMA J. Numer. Anal., in
press, doi:10.1093/imanum/drn056.
[14] G. Mastroianni, G. Monegato, Truncated Gauss–Laguerre quadrature rules, in: Recent Trends in Numerical Analysis, in: Adv. Theory Comput. Math.,
vol. 3, Nova Sci. Publ., Huntington, NY, 2001, pp. 213–221.
[15] G. Mastroianni, G. Monegato, Truncated quadrature rules over (0, ∞) and Nyström type methods, SIAM J. Numer. Anal. 41 (2003) 1870–1892.
[16] G. Mastroianni, I. Notarangelo, A Lagrange-type projector on the real line, Math. Comp., in press, Posted on July 7 (2009), doi:10.1090/S0025-5718-09-
02278-9.
[17] G. Mastroianni, I. Notarangelo, A Nyström method for Fredholm integral equations on the real line, J. Integral Equations Appl., in press.
[18] G. Mastroianni, M.G. Russo, Lagrange Interpolation in weighted Besov spaces, Constr. Approx. 15 (1990) 257–289.
[19] G. Mastroianni, M.G. Russo, Lagrange interpolation in some weighted uniform spaces, Facta Univ. Ser. Math. Inform. 12 (1997) 185–201. Dedicated to
Professor Dragoslav S. Mitrinović (1908–1995), Niš, 1996.
[20] G. Mastroianni, J. Szabados, Polynomial Approximation on infinite intervals with weights having inner zeros, Acta Math. Hungar. 96 (2002) 221–258.
[21] G. Mastroianni, J. Szabados, Polynomial approximation on the real semiaxis with generalized Laguerre weights, Stud. Univ. Babeş-Bolyai Math. 52 (4)
(2007) 105–128.
[22] S.V. Pereverzev, E. Schock, S.G. Solodky, On the efficient discretization of integral equations of the third kind, J. Integral Equations Appl. 11 (4) (1999)
501–514.
[23] E.B. Saff, V. Totik, Logarithmic Potentials with External Fields, Grundlehren der Mathematischen Wissenschaften, Springer-Verlag, Berlin, 1997.
[24] E. Schock, Integral equations of the third kind, Studia Math. 81 (1) (1985) 1–11.
[25] D. Shulaia, On one Fredholm integral equation of third kind, Georgian Math. J. 4 (1997) 461–476.
[26] D. Shulaia, Solution of a linear integral equation of third kind, Georgian Math. J. 9 (2002) 179–196.
[27] I. Sloan, Quadrature methods for integral equations of the second kind over infinite intervals, Math. Comp. 36 (154) (1981) 511–523.
[28] A.F. Timan, Theory of Approximation of Functions of a Real Variable, Dover Publications, Inc., New York, 1994.

You might also like