Mathl. Comput. Modelling Vol. 18, No. 9, pp. 55-62, Printed in Great Britain. All rights reserved
1993 Copyright@
08957177/93 $6.00 + 0.00 1993 Pergamon Press Ltd
Some Recent Numerical Methods for Solving Nonlinear Hammerstein Integral Equations B.SOME Facultb
des Sciences
et Techniques (FA. S. T.) DBpartement de Mathhmatiques et Informatique 03 B.P. 7021 Ouagadougou 03, Universit6 de Ouagadougou, Burkina Faso
(Received
April
1993;
accepted
May 1993)
survey is given of some recent numerical methods for fixed points of nonlinear Hammerstein integral operators. These methods include projection methods (Galerkin and collocation) and Adomian’s decomposition method. Proof of convergence of Adomian’s method applied to Hammerstein nonlinear integral equation is given. A discussion is given of some recent works on iteration methods for solving these nonlinear equations.
Abstract-A
1. INTRODUCTION In the following survey, we consider applied to a wide variety of nonlinear be of the second kind,
numerical methods integral equations.
of a general The integral
nature, those that can be equations are restricted to
2 = H(z),
(1.1)
where H is a nonlinear integral operator. An important equation is considered here as the following:
special
case, the Hammerstein
integral
R z(t) = ?m
A well-known
example
+
s
with R > 0.
_R K(4s)f(s,4s))ds,
is the Chandrasekhar
(1.2)
H-equation
’ t H(t)
H(t) = ; /
H(s) ds t-k.9
0
*
(1.3)
It can be rewritten in the form (1.2) by letting z(t) = l/H(t) and rearranging the equation. For an exact solution and some recent works on (1.3) and related equations, see [l]. Another rich source of Hammerstein integral equations is the reformulation of boundary value problems for both ordinary and partial differential equations. The numerical solution of nonlinear equations has two major aspects. First, the equation z = H(z) is discretized, generally by replacing it with a sequence of finite-dimensional approximating problems (projection methods) z,, = H,(z,), with n --) +oo, and n is some discretization parameter. Second (Adomian’s method), the solution z of the equation z = H(z) is replaced by a series x = E z,,, and H(z) = E A,,, where the A, are polynomials of ~0,. . . , zn (Adomian’s n=O n=O polynomials). Projection methods are discussed in Section 2, Adomian’s method in Section 3, and a discussion follows in Section 4.
55
B.
56
SOME
2. PROJECTION
METHODS
General Description We give only the most important features of the analysis of projection methods and then discuss some applications. General theoretical frameworks for projection methods have been given by a number of researchers, among these are Atkinson [2,3] and Potra [3], Golberg [4], Kumar [5], and Saramen [6]. Let E be a Banach space, usually C”( [-R, R]) or L’([-R,
RI); and let E,,, n 2 1 be a sequence
of finite-dimensional subspaces being used to approximate z* (solution of (1.1)). Let P, : E z E, be a bounded projection, n 2 1 and, for simplicity, let E, have dimension n. It is usually assumed that P*x-+x,
CS?l++CXl,
x E E,
(2.1)
for all x in some dense subspace of E containing the range of H. In abstract form, the projection method amounts to solve 2, = P, H(xn). (2.2) For motivation about connecting this with a more concrete integral equation, see the discussion in [7, pp. 54-711 for a linear integral equation. Assume P, can be written
pnx= ~4(XMj,x
E E,
(2.3)
j=l
with (4j)l
a set of bounded linear functionals that are inde-
det[lj(4j)l
# 6.
(2.4)
To reduce (2.2) to a finite nonlinear system, let (2.5)
Xn=kOZj4j j=l
be solved for {ej},
j = 1,. . . , n from the nonlinear system
(2.6) The choice of (41,. . . , &} and (11, . . . , In} determine the particular method. framework for the error analysis of projection methods was given in [3].
A very general
THEOREM. Let H : R c E + E be completely continuous, with E, Banach and R open. Assume
that the sequence of bounded projections {P,,}
on E satisfies
SUP IIU - Pn) Wx)I/ + 0,
ZEB
asn--,+m,
(2.7)
for all bounded sets B c R. Then, for any bounded open set B, with B c R, there is an N such that I - H and I - P,, H have the same rotation on the boundary of B, for n 1 N. In the particular case that x* is an isolated fixed point of nonzero index of H, there is a neighborhood & = {x I lb - x*(1 5 E}, with P,, H h aving fixed points in B, that are convergent to x*. (The index of x* is defined as the rotation of I - H over the surface S, of any Be which contains only the one fixed point x*. In case [I - H’(x)] - 1 exists on E into E, the index of x* is fl.) For the proof of the above theorem, see [3]. Generally, the projections are assumed to be pointwise convergent on E, as in (2.1) and in that case (2.7) follows in straightforward way. However, only equation (2.7) is actually needed.
Nonlinear Hammerstein Integral Equations
Iterated Projection
57
Methods
Given the projection
method solution x,,, define
(2.8)
P, = H(x& then using (2.2) we have
(2.9) and f,
satisfies f,
with H differentiable
(2.10)
= H(P,&&),
in a neightborhood of x*,
- CalI I CllX811x*
xnll,
n L N,
(2.11)
c > 0.
Thus, Z,, ---) x* at least as rapidly ss x,, + x*. For many methods, especially Galerkin methods, the convergence & + Z* can be shown to be more rapid than that of x,, + x*. For the nonlinear iterated projection method, see [3]. Galerkin’s Method Let E be C”((-R, R]) with the uniform norm (( - lloo or L2([-R, RI), and let En be a finite dimensional subspace of E. Define P, x to be the orthogonal projection of x onto E,,, based on using the inner product of L2((-R,R])
and regarding E,, as a subspace of L2([-R,
RI); thus (2.12)
y E En, (PnX>Y> = (XVY)? a.11 with (a, .) the inner product in L2([-R,R]). products are also used in some applications. Let xn = ~~=‘=, cq &, with (&)lsjgn,
Other Hilbert spaces (e.g., H’([-R,R])
a basis of En. Solve for {&},
and inner
using
(2.13)
Generally, the integrals must be evaluated numerically, thus introducting this is done, it is called the discrete Gale&in method. In general, for any projection
(lx* -&II
new errors.
When
method,
I c, IIV - pn)~*Il~
c, = c max{]](l
- Pn) x8]],
with
llH’(~*)(I- Pn)llI-
(2.14)
If E is a Hilbert space, then c, converges to zero, because ]]H’(z*) (I - Pn)ll ---, 0, as n + 00. This uses the fact that H/(x*) and H’(x*)* are compact linear operators. Thus, &., converges to Z’ more rapidly than xn does. In any Hilbert space, 5, + x* more rapidly than z,, + x’ does. Similar results can be shown for Galerkin’s method in C’([-R, R]) with the uniform norm, see [8]. For numerical examples of application of the Galerkin’s method to nonlinear integral equations, see [2].
B. SOME
58 Collocation
Method RI), let ti, . . . , t, E [-R, R] be such that
Let E = C”([-R,
deQj(ti)]
(2.15)
= O. x at the nodes ti, . . . , t,.
Define P,, x to be the element in En that interpolates of En, solve the nonlinear system
( )
&Wd=H &j4j
j=l
(t&
i=l,...,
To find x, element
(2.16)
n;
j=l
the integrals must usually be evaluated numerically, introducing a new error, this is called the discrete collocation method. In using the general error result [2], ]]x* - P, x* ]loo is simply an interpolation
error.
For the iterated
collocation
method,
with 11x* - xnll I cn11x* -G&II, c, = c max{]]x* - hll, ll(I - pJ*LlI, en)
(2.17) and (2.18)
To show superconvegence
of f,
to x*, one must show lim e, = 0. n+cG
Studies of this and various tion are given in [3]. Case of Hammerstein Consider
types of collocation
(2.19)
methods
applied
to the Urysohn
integral
equa-
Equation
using the projection
method
to solve
R x(t) = Y(t)+
s -R
with R > 0.
Wt,s)f(s,x(s))ds,
(2.20)
For nonlinear integral equations, the Galerkin and collocation methods can be quite expensive to implement. But, for this equation, there is an alternative formulation which can lead to a less expensive projection method. We consider the problem when using a collocation method. Let xn(t) = Cy=“=, aj 4j(t) and solve for {oj}, using
2
aj $j(ti) =
Y(h)
j=l
+
/” K(t, S)f -R
(sy$oj#j(&))
ds,
fori=I,...,n.
(2.21)
In the iterative solution of this system, many integrals will need to be computed, which usually becomes quite expensive. In particular, the integral on the right side will need to be re-evaluated with each new iterate. Kumar [5] and Kumar and Sloan [9] recommend the following variant approach. Define R 6 Y(t)+
z(t)= f
and obtain
W&s) s -R
(
z(t) from x(t) = y(t) + s_“, K(t, s) z(s) ds; the collocation
Z(t) =
k@j 4j(t) = f
j=l
h,
Y(h) +
(2.22)
f(s,x(s))ds , ) method
k/Jj1-1 K(ti, s> bj(s) ds
for (2.22) is
*
(2.23)
j=l
The integrals on the right hand side need to be evaluated only once, since they are dependent only on the basis, not on the unknowns {aj}. Many fewer integrals need to be calculated to solve this system. For the results on Hammerstein’s integral equations, see [6].
Nonlinear Hammerstein
Integral Equations
3. ADOMIAN’S
59
METHOD
General Description Consider,
for example,
the general
functional
equation
Ax=y, where
A is an operator
from an Hilbert
for x E V satisfying
are looking
(3.3).
(3-I)
V into V, y is a given
space
We assume
that
function
(3.1) has a unique
in V, and we
solution
for z E V.
If the operator A has linear and nonlinear terms, the linear term is decomposed into L + R, where L is easily invertible and R is the remainder of the linear operator. Then, the operator A is decomposed as following: A = L + R + N, where N represents the nonlinear term. With these considerations,
the equation
(3.1) becomes Lx+Rx+Nx=y.
Further,
if L-’
is the inverse of the linear operator
(3.2) L, then the solution
x of (3.1) or (3.2) verifies
z=L-‘~-L-‘Rx-L-~Nx. The initial
Adomian’s
method
[lO,ll]
consists
x=
The nonlinear
operator
is decomposed
(3.3)
of representing
x as a series,
-&k.
(3.4)
as
Nx
= FAn,
(3.5)
where the A, are polynomials z = C:“=, Xi xi, N(z) having
that of xc, . . ,x, (Adomian plynomials), = CT=“=, X, A,, with X being a parameter introduced
,!A,=&=
[N(~Yxi)l,_,;
n=0,1,2
we obtain
,....
It allows to determine the An’s by formula (3.6). Generally it is possible to obtain functions of xc, xi, . . . , x, from the nonlinearity N. With these decompositions, can be written 5
n=L-'y-L-lR
km 18:9-E
the other terms
of the above series
Rx0 - L-l Ao,
x2 = -L-l
Rx1 - L-’ Al,
x, = -L-=
Rx+~
=
(3.6) exactly A, as formula (3.3)
-L-1(zA+
-L-l
x1
and
F (nEoxq
n=O
Taking x0 = L-l y, we can identify algorithm
by writing
for convenience,
- L-l A,_1.
E x, by the following n=O
B. SOME
60
Adomian’s
Method
We consider
Applied to Hammerstein
a Hammerstein
integral
equation
Integral Equation
of the form
R
44 = Y(t) + From (3.7), we obtain
directly
Ax=y,
J KG, s>f(% x(s)) -R
Adornian’s
fundamental
where Ax(t)
= x(t) -
with R > 0.
6
functional
J
equation
(3.7) by writing
R
WC s) f(s, 4s))
ds.
(3.8)
-R
Here, the operator
A can be decomposed
as: A = L + N; where Lx = x is the linear
is no remainder linear term) and Nz = - J_“, K(t, s) f(s, z(s)) ds is the nonlinear nonlinear. Then (3.8) can be written Lx+Nx=y
or
According to Adomian’s technique the solution are calculated by the following algorithm:
term (there term,
with f
x+Nx=y.
(3.9)
x of (3.9) verifies x = 5 z,, n=O
where the terms x,,
x0 =Y, 21 =
Ao,
22
-41,
=
x, = A,_1. We obtain
the Adomian’s
polynomials
N @A%)
= -[IK(r,s)f
where X is a parameter introduced From (3.10) we obtain
Then
A, from the nonlinear
(+%(a))
term by writing
da = ~,‘X,
(3.10)
for convenience.
we have A, = ;
(3.11)
REMARK. Formula (3.11) allows us to calculate z,, and then the solution x of the Hammerstein integral equation. Numerically (3.11) is not expensive to implement [12]. Convergence of Adomian’s
Method
Applied to Hammerstein
Let us come back to our nonlinear integral equation technique, equation (3.9) can be replaced by
(3.9) x + N x = y. According
to Adomian’s
(3.12)
&+gAn=y,
n=O
Integral Equations
n=O
Nonlinear
but the convergence What
hypothesis
problem
will ensure
Hammerstein
remains
Integral
in suspense.
the convergence
61
Equations
We have to answer to the following zn and En”=-, A,?
of the series C,“=,
question: In a recent
paper [13], Y. Cherruault has given an answer to the above question of convergence. In a general case, he has given a proof of the convergence of Adomian’s method by using the fixed point theorem. proving
He introduced that
a new formulation
S, is a solution
of the method
N (x0 + S) = In the
following,
series substituted G. Saccomandi We consider
we will give a new convergence in another
S, = xi + x2 +. . . + x, and
by setting
of the fixed point equation
s.
theorem
using
the properties
series, and recent results of convergence
obtained
of the entire
by Y. Cherruault,
and B. Some [12]. a Hammerstein
nonlinear
integral
equation
of the form
R
x(t) = Y(t) + calling
J wt, s>f(s, x(s))
with R > 0;
ds,
(3.13)
-R
Na: = - J_“, K(t, s) f(s, X(S)) ds, equation
(3.13) becomes
x+Nx=y.
If x can be developed as an infinite parameterized by introducing X
The series (3.15) is absolutely
series, we can write
convergent
(3.14)
x =
E x,. n=O
This expression
00 xx = cx,x,.
can be
(3.15)
for JX] 5 1, and for such values of X the series
(3.16)
where m is the upper limit of the series’ elements and 0 < p < 1, is a majorant On the other hand, if f is an analytical function with respect to x in the interval
series for (3.15). [-R, R] we have
Nx=&,x,.
(3.17)
n=O
Then,
subtituting
(3.15) into (3.17) we obtain
the following
array
~o+a~~o+a2s~+~~~+.~x~+'.~+u~x~X+202x~x~X+...+71,~,~;;L-~~~~ +... +alx2X2
+ a2(x? +2x0x2)+....
(3.18)
The i-row of this array converges to the Ai defined as in (3.11) (for details, see [12]), when we set X = 1 because Nx can be developed in a Taylor series. Our problem is now to prove the convergence of the double serie in the array (3.18) for X = 1. It is necessary to have X = 1 inside the radius of convergence: otherwise, we are not able to recover the solution (3.14). In our situation, the classical theorems of elementary calculus on double series cannot ensure that 1 will be inside the radius of convergence of (3.18). The following theorem states a sufficient condition to ensure the convergence of (3.18) for X = 1. Obviously, we must require some stronger hypothesis than usual but generally the equations coming from physics satisfy such hypotheses.
B. SOME
62 THEOREM.
If f is an analytical
function
series x = Cr=, x,, the parameterization and the series x can be majored by
of x in (-R,
R), x can be decomposed
xx = 5 x, X, is absolutely n=O
convergent
as an infinite for X E [-1, 11,
(3.19) where m’ 2 m (m the upper
converges
This theorem recent
limit of the x,),
E > m/R,
and p I 1, then the double
series (3.18)
way as the proof of our theorem
given in the
for X = 1.
can be demonstrated
in a similar
work [12, pp. 88-891
4. DISCUSSION In practice, Adornian’s method is a numerical elegant method that can solve wide classes of nonlinear equations (algebraic differential, intergro-differential, partial differential equations, etc.) It avoids the cumbersome integrations of the projection methods and can solve physical problems which cannot be done by other numerical methods (iterative collocation methods, etc.) Adomian’s method is especially well-adapted to Hammerstein integral equations because of the form of the nonlinear functional equation chosen by Adomian. Convergence hypotheses are easy to verify and we obtain rapidly an approximation of the solution. The solution is given by a function, and not only at some grid points as is in the collocation method [5]. This method is powerful, more efficient, and faster than projection methods (Galerkin’s method, collocation method, etc.) [2]. The only difficulties, could arise from the calculations of Adomian’s polynomials applied to nonlinear integral equations. But, in most pratical cases, the nonlinearities to simple A,‘s, which can be easily determined.
A,
when give rise
REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13.
Dans. Theory Stat. Physics 18, 185-204 C.T. Kelley, A fast two-grid method for matrix H-equations, (1989). K.E. Atkinson, A survey of numerical methods for solving nonlinear integral equations, J. of Integral Equations 4 (l), 15-46 (1992). K. Atkinson and F. Potra, Projection and iterated methods for nonlinear integral equations, SIAM. J. Num. Anal. 24, 1352-1373 (1989). M. Golberg, Editor, Perturbed projection methods for various classes of operators and integral equations in numerical solution of integral equations, Plenum Press, New York, (1990). S. Kumar, A discrete collocation type method for Hammerstein equations, SIAM J. Num. Anal. 25, 328-341 (1988). J. Saranen, Projection methods for a class of Hammerstein equations, SIAM J. Num. Anal. (to appear). K. Atkinson, A survey of numerical methods for Fredholm integral equations of the second kind, SIAM J. Num. Anal. 24, 248-262 (1976). K. Atkinson and G. Chandler, BIE methods for solving Laplace’s equation with nonlinear boundary conditions: The smooth boundary case, Math. Comp. 55, 451-472 (1990). S. Kumar and I. Sloan, A new collocation type method for Hammerstein integral equations, Math. Comp. 48, 585-593 (1987). G. Adomian, Convergent series solution of nonlinear equations, J. of Comp. and App. Maths. 11, 225-230 (1984). G. Adomian, A review of the decomposition method and some recent results for nonlinear equations, Mathl. Comput. Modelling 13 (7), 17-43 (1990). Y. Cherruault, G. Saccomandi and B. Some, News results for convergence of Adomian’s method applied to integral equations, Mathl. Comput. Modelling 16 (2), 85-93 (1992). Y. Cherruault, Convergence of Adomian’s method, Kybernetes 18 (2), 31-38 (1989).