Numerical Solutions of Linear and Nonlinear Singular Perturbation Problems PDF

Computers and Mathematics with Applications 55 (2008) 25742592
www.elsevier.com/locate/camwa
Numerical solutions of linear and nonlinear singular
perturbation problems
Tzu-Chu Lin
a
, David H. Schultz
a
, Weiqun Zhang
b,
a
Department of Mathematical Sciences, University of Wisconsin-Milwaukee, PO Box 413, Milwaukee, WI53201, United States
b
Department of Mathematics, Adrian College, 110 S Madison St, MI49221, United States
Received 15 September 2007; accepted 26 September 2007
Abstract
A new method is developed by detecting the boundary layer of the solution of a singular perturbation problem. On the non-
boundary layer domain, the singular perturbation problem is dominated by the reduced equation which is solved with standard
techniques for initial value problems. While on the boundary layer domain, it is controlled by the singular perturbation. Its
numerical solution is provided with nite difference methods, of which up to sixth order methods are developed. The numerical
error is maintained at the same level with a constant number of mesh points for a family of singular perturbation problems.
Numerical experiments support the analytical results.
c 2007 Elsevier Ltd. All rights reserved.
Keywords: Singular perturbation; Reduced equation; Differential equations; Numerical solutions; Stability
0. Introduction
We consider the singular perturbation problem (SPP) in the quasilinear form
u
= f (x, u)u
+ g(x, u) for x (a, b) and f (x, u) = 0,

u(a) = v
a
and u(b) = v
b
, (1)
where the positive constant 1 is the singular perturbation parameter, and f (x, u) and g(x, u) are C
1
in
the domain considered. Assume that the right hand side (x, u, u
) f (x, u)u
+ g(x, u) of Eq. (1) satises a

generalized Nagumo condition, Chang and Howes [2]. This simply means that every solution u = u(x) satisfying
(x) u(x) (x) on a subinterval J [a, b] has a bounded derivative, that is, there exists a constant N = N(, )
such that |u
(x)| N on J. (x) and (x) are functions of C

2
([a, b]).
In the early twentieth century, Prandtl gave the light of theoretical understanding of the singular perturbation
phenomenon of hydrodynamics to the Third International Congress of Mathematics. Since then, a great deal of effort
has been made to conquer this anomaly. Analytically, the asymptotic expansion of OMalley [2] and the a priori
Corresponding author.
E-mail addresses: lin@uwm.edu (T.-C. Lin), schultz@uwm.edu (D.H. Schultz), wzhang@adrian.edu (W. Zhang).
0898-1221/$ - see front matter c 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.camwa.2007.09.011
T.-C. Lin et al. / Computers and Mathematics with Applications 55 (2008) 25742592 2575
bound theorem of Chang and Howes [1] were among the prominent approaches. With the advance of computing
technology, there is a ow of literature on numerical solutions from the 1980s. Schultz and his students successfully
developed the stabilized high order nite difference methods, Schultz and Choo [3], and Schultz and Ilicasu [4].
Miller, ORiordan and Shishkin [5] constructed the Shishkin-type mesh to gain the error estimation independent of
the singular perturbation parameter. Segal [6] used a piecewise uniform mesh to overcome the drawback of uniform
mesh methods.
Our research is an analytically guided numerical approach. We applied the Nagumo differential inequalities to
establish a sharp bound for the solution of a singular perturbation problem. Consequently, the domain of a singular
perturbation problem is divided into two: the boundary layer and the non-boundary layer. On the non-boundary layer
domain, the singular perturbation problem is dominated by the reduced differential equation which is solved with
standard techniques for initial value problems. While on the boundary layer domain, it is controlled by the singular
perturbation. The differential equation on the boundary layer domain is solved with nite difference methods, of which
up to sixth order methods are developed. The stability and convergence is established theoretically and numerically.
As an important feature of the new method, the numerical error is maintained at the same level for a family of singular
perturbation problems.
1. Properties of the exact solution
When the singular perturbation parameter vanishes, we consider the reduced equations
f (x, u)u
+ g(x, u) = 0,
u(a) = v
a
,
(1a)
and
f (x, u)u
+ g(x, u) = 0,
u(b) = v
b
.
(1b)
Let R
a
and R
b
be the solutions of (1a) and (1b) respectively. We study the solutions of the reduced problems which
are strongly stable in a sense to be dened momentarily. In the following denitions, > 0 is a small constant, Chang
and Howes [1].
If a solution R
a
= R
a
(x) of (1a) exists in [a, b], we dene
(R
a
) = {(x, u) : a x b, |u R
a
(x)| d
a
(x)},
where the positive continuous function d
a
(x) satises d
a
(x) |v
b
R
a
(b)| + for b

2
x b and d
a
(x)
for a x b .
Similarly, if a solution R
b
= R
b
(x) of (1b) exists in [a, b], we dene
(R
b
) = {(x, u) : a x b, |u R
b
(x)| d
b
(x)},
where the positive continuous function d
b
(x) satises d
b
(x) |v
a
R
b
(a)| + for a x a +

2
and d
b
(x)
for a + x b.
A solution R
a
= R
a
(x) of the reduced equation (1a) is said to be strongly stable in [a, b] if there exists a positive
constant k such that
f (x, u) k > 0 in (R
a
).
A solution R
b
= R
b
(x) of the reduced equation (1b) is said to be strongly stable in [a, b] if there exists a positive
constant k such that
f (x, u) k < 0 in (R
b
).
The Reduced Problem Theorem (Part A). Let the reduced equation (1b) have a strongly stable solution R
b
(x) of
class C
2
([a, b]). Let c be a positive constant and =
k
k
2
4m
2
be a solution of the equation
2
+k +m = 0,
2576 T.-C. Lin et al. / Computers and Mathematics with Applications 55 (2008) 25742592
where k is the positive constant in the strongly stable condition and m is a positive constant such that | f
u
R
b
+g
u
| m
in (R
b
). Then there exists an
0
> 0 such that for 0 < <
0
, the SPP (1) has a solution u which satises
|u(x) R
b
(x)| < C for x [a +w, b],
where w is a constant for a family of small values.
Proof. By Theorem 4.1 of Chang and Howes [1], there exists an
0
0
, the SPP (1) has a
solution u satisfying
|u(x) R
b
(x)| s(x) +c for x [a, b], (2)
where s(x) = |v
a
R
b
(a)|e
(xa)
, =
k
k
2
4m
2
and c is a positive constant. To ensure as a real number, let
0
=
k
2
4m
.
We consider the positive function s(x) = |v
a
R
b
(a)|e
(xa)
. Choose a positive constant w, such that w
2 ln
k
for sufciently small values of . For x [a +w, b], it implies
x a +w
x a +
_
2 ln
k
_

(x a)
_
2 ln
k
_

k(x a)
2
ln
e
k(xa)
2
e
ln
= since e
x
is increasing.
Note that w can be chosen as a constant for a family of small values of since (ln ) increases rather slowly as
vanishes. The values of (ln ) are listed in the following table. We call w as the boundary layer parameter.
Values of (ln )
ln
10
5
11.513
10
6
13.816
10
7
16.118
10
8
18.421
10
9
20.723
10
10
23.026
10
12
27.631
10
15
34.539
10
20
46.052
Note that
=
k
k
2
4m
2
<
k
2
.
Let c
1
= |v
a
R
b
(a)|, then we have
s(x) = c
1
e
(xa)
< c
1
e
k(xa)
2
< c
1
for x [a +w, b].
Combining the bounds of (2) and taking C = c
1
+c, we get
|u(x) R
b
(x)| s(x) +c
< c
1
+c
= C,
for x [a +w, b].
Similarly, we give the conclusion about the reduced equation (1a).
The Reduced Problem Theorem (Part B). Let the reduced equation (1a) have a strongly stable solution R
a
(x) of
class C
2
([a, b]). Let c be a positive constant and =
k
k
2
4m
2
be a solution of the equation
2
+k +m = 0,
where k is the positive constant in the strongly stable condition and m is a positive constant such that | f
u
R
a
+g
u
| m
in (R
a
). Then there exists an
0
0
the SPP (1) has a solution u = u(x) which satises
|u(x) R
a
(x)| < C for x [a, b w],
where w is a constant for a family of small values.
Proof. The proof follows analogously to that of part A. By Theorem 1.1.2 of Zhang [7] which is equivalent to
Corollary 4.1 of Chang and Howes [1], there exists an
0
0
, the SPP has a solution
u which satises
|u(x) R
a
(x)| s(x) +c for x [a, b],
where s(x) = |v
b
R
a
(b)|e
(bx)
, =
k
k
2
4m
2
and c is a positive constant.
We now consider the positive function s(x) = |v
b
R
a
(b)|e
(bx)
. Choose a positive constant w, such that
w
2 ln
k
for sufciently small values of . For x [a, b w],
x b w
b x w
b x
_
2 ln
k
_

k(b x)
2
ln
e
k(bx)
2
e
ln
= since e
x
is increasing.
Also note
=
k
k
2
4m
2
<
k
2
.
Let c
1
= |v
b
R
a
(b)|. Then
s(x) = c
1
e
(bx)
< c
1
e
k(bx)
2
< c
1
for x [a, b w].
Note that w can be chosen as a constant for sufciently small values of according to the table in the proof of part A.
Combining the above bounds and taking C = c
1
+c, we get
|u(x) R
b
(x)| s(x) +t (x)
< c
1
+c
= C,
for x [a, b w].
By the reduced problem theorem, we can detect the boundary layer. Specically, if f (x, u) k < 0 in (R
b
)
and if there exists a boundary layer for the SPP (1), then it occurs at the left boundary x = a. If f (x, u) k > 0 in
(R
a
) and if there exists a boundary layer for the SPP (1), then it occurs at the right boundary x = b. Furthermore,
the length of the boundary layer is at most w, where w is the boundary parameter dened in the reduced-problem
theorem.
2. The new method
By the reduced-problem theorem, the SPP (1) is divided into two cases:
(a) A solution R
b
= R
b
(x) of the reduced equation (1b) is strongly stable. The SPP (1) can be analytically
approximated by the following two differential equations on the boundary layer and non-boundary layer domain
respectively,
f (x, u)u
+ g(x, u) = 0 for x (t, b),

u(b) = v
b
,
(3)
and
u
= f (x, u)u
+ g(x, u) for x (a, t ),

u(a) = v
a
and u(t ) = v
t
,
(4)
where the transition point is t = a +w and w is a constant for a family of values of .
(b) A solution R
a
= R
a
(x) of the reduced equation (1a) is strongly stable. The SPP (1) can be analytically
approximated by the following two differential equations on the boundary layer and non-boundary layer domain
respectively:
f (x, u)u
+ g(x, u) = 0 for x (a, t ),

u(a) = v
a
,
(5)
and
u
= f (x, u)u
+ g(x, u) for x (t, b),

u(t ) = v
t
and u(b) = v
b
,
(6)
where the transition point is t = b w and w is a constant for a family of values of .
Note that the boundary value v
t
at the transition point t of the Eqs. (4) and (6) is not known. To handle this, we
substitute R
b
(t ) for v
t
into Eq. (4) and substitute R
a
(t ) for v
t
into Eq. (6) respectively. In summary,
(a) when a solution R
b
= R
b
(x) of the reduced equation (1b) is strongly stable, the SPP (1) is approximated by
f (x, u)u
+ g(x, u) = 0 for x (t, b),

u(b) = v
b
,
(3)
and
u
= f (x, u)u
+ g(x, u) for x (a, t ),

u(a) = v
a
and u(t ) = R
b
(t ),
(4p)
where the transition point is t = a +w and w is a constant for a family of values of .
(b) when a solution R
a
= R
a
(x) of the reduced equation (1a) is strongly stable, the SPP (1) is approximated by
f (x, u)u
+ g(x, u) = 0 for x (a, t ),

u(a) = v
a
,
(5)
and
u
= f (x, u)u
+ g(x, u) for x (t, b),

u(t ) = R
a
(t ) and u(b) = v
b
,
(6p)
where the transition point is t = b w and w is a constant for a family of values of .
Since the value v
t
at x = t of Eqs. (4) and (6) is replaced with the approximation from the reduced problems, it
is perturbed by . We show the stability of this type of -perturbation. Let = |v
t
R
b
(t )| for Eqs. (4) and (4p) or
= |v
t
R
a
(t )| for Eqs. (6) and (6p).
If a solution p = p(x) of (4p) exists in [a, t ], we dene
( p) = {(x, y) : a x t, | p(x) y| C},
where C is a positive constant. Similarly, if a solution p = p(x) of (6p) exists in [t, b],
( p) = {(x, y) : t x b, | p(x) y| C},
where C is a positive constant.
The Stability Theorem (Part A). Let p(x) be a solution of (4p). Let k and m be positive constants such that
f (x, u) k and | f
u
p
+ g
u
| m in ( p). Then there exists a positive constant C and the Eq. (4) has a solution
u = u(x) such that
|u(x) p(x)| C for x [a, a +w].
Proof. Let the transition point be t = a+w and the exact value of the SPP (1) at x = t be v
t
. We prove the inequality
on the boundary layer domain with Theorem 2.1 of Chang and Howes [2], by constructing a bounding pair satisfying
the following inequalities,
_
_
_
(x) (x),
(a) v
a
(a) and (t ) v
t
(t ),
(x, ,
) and
(x, ,
).
(7)
The proof follows these steps:
Step P1: Dene the bounding functions (x) and (x) such that (x) (x);
Step P2.1: Proof of (a) v
a
(a);
Step P2.2: Proof of (t ) v
b
(t );
Step P3.1: Proof of
(x, ,
);
Step P3.2: Proof of
(x, ,
) and
Step P4: Conclusion.
We consider the positive function
s(x) = e
(xt )
where =
k +
k
2
4m
2
.
Differentiating s(x), we get
s
(x) = e
(xt )
< 0 and s
(x) =
2
e
(xt )
> 0.
Thus we have
s
+ks
+ms = (
2
+k +m)e
(xt )
= 0,
since
2
+k +m = 0 by our choice of . Therefore, s(x) is the solution of the homogeneous differential equation
s
+ks
+ms = 0 for x (a, t )

s(t ) = , s
(t ) = .
Step P1: Now, we dene the bounding pair
(x) = p(x) s(x) and (x) = p(x) +s(x),
and proceed to show that it satises the inequalities of (7). Clearly, we have (x) (x) for x [a, t ] since s(x) is
positive and
(x) = p(x) s(x) p(x),
and
(x) = p(x) +s(x) p(x).
Step P2.1: At the boundary x = a, we have
(a) = p(a) s(a) = v
a
s(a) v
a
,
since s(a) is positive. Meanwhile, we have
(a) = p(a) +s(a) v
a
.
Thus we get (a) v
a
(a).
Step P2.2: While at the boundary x = t , we have
(t ) = p(t ) s(t ) = (v
t
) v
t
.
Meanwhile, we have
(t ) = p(t ) +s(t ) = (v
t
) + v
t
.
Thus we get (t ) v
t
(t ).
Step P3.1: Next we verify the differential inequality
(x, ,
).
By denition, we have
= p s p = s < 0,
= s
> 0,
= p
.
By the mean value theorem (MVT), we have
[ f (x, ) p
+ g(x, )] [ f (x, p) p
+ g(x, p)]
p
= f
u
(x, ) p
+ g
u
(x, )
[ f (x, p) p
+ g(x, p)] = [ f (x, ) p
+ g(x, )] +[ f
u
(x, ) p
+ g
u
(x, )]( p),
where (x, ) is between (x, ) and (x, p). Now, we consider
(x, ,
) = p
(x, ,
)
= s
[(x, ,
) p
]
= s
{[ f (x, )
+ g(x, )] [ f (x, p) p
+ g(x, p)]}
= s
{[ f (x, )
+ g(x, )] [ f (x, ) p
+ g(x, )]
+[ f
u
(x, ) p
+ g
u
(x, )]( p)} by MVT
= s
f (x, )(
) h(x, )( p),
where h(x, ) = f
u
(x, ) p
+ g
u
(x, ). Thus we obtain
(x, ,
) = s
f (x, )(s
) h(x, )(s)
= s
+ f (x, )s
+h(x, )s
s
ks
+h(x, )s, since f (x, ) k < 0 and s
< 0 f (x, )s
ks
ks
ms, since |h(x, )| = | f

u
(x, ) p
+ g
u
(x, )| m
h(x, ) m and ds > 0
h(x, )s ms
= 0 since s
+ks
+ms = 0.
Therefore, we get
(x, ,
).
Step P3.2: Similarly we verify the differential inequality
(x, ,
). We consider
(x, ,
) = p
+s
(x, ,
)
= s
+p
(x, ,
)
= s
+p
[ f (x, )
+ g(x, )]
= s
+[ f (x, p) p
+ g(x, p)] [ f (x, )
+ g(x, )]
= s
{[ f (x, )
+ g(x, )] [ f (x, p) p
+ g(x, p)]}
= s
f (x, )(
) h(x, )( p),
where (x, ) is between (x, p) and (x, ), and h(x, ) = f
u
(x, ) p
+ g
u
(x, ). We obtain
(x, ,
) s
+ks
h(x, )( p), since f (x, ) k < 0 f (x, ) k

and s
0 f (x, )s
ks
+ks
+ms, since |h(x, )| m h(x, ) m

and s > 0 h(x, )s ms
= 0 since s
+ks
+ms = 0.
Step P4: Now, it follows from Theorem 2.1 of Chang [1], that the Eq. (4) has a solution u satisfying (x) u(x)
(x), which is equivalent to,
p(x) s(x) u(x) p(x) +s(x)
s(x) u(x) p(x) s(x)
|u(x) p(x)| s(x).
Finally, we show that s(x) C for x [a, a +w].
For x [a, a +w], s(x) is decreasing since s
(x) is negative. Therefore, we have

s(x) s(a)
= e
[a(a+w)]
= e
w
.
Meanwhile, we have
w = w
(k +
k
2
4m)
2
= w
(k
k
2
4m)
2
= w
(k
k
2
4m)(k +
k
2
4m)
2(k +
k
2
4m)
=
4wm
2(k +
k
2
4m)
=
2wm
k +
k
2
4m
2wm
k
<
2m
k
since 0 < 1 and w is much less than 1 for a family of small values.
Taking the positive constant C = e
2m
k
, we obtain
s(x) s(a) < C for x [a, a +w].
Next, we state a similar stability conclusion about the perturbation at the right boundary on the boundary layer
domain.
The Stability Theorem (Part B). Let p(x) be a solution of (6p) . Let k and m be positive constants such that
f (x, u) k and | f
u
p
+ g
u
| m in ( p). Then there exists a positive constant C and the Eq. (6) has a solution
u = u(x) such that |u(x) p(x)| C for x [b w, b].
Proof. The proof follows analogously to that of the Stability Theorem (part A).
Finally, we consider the choice for the boundary layer parameter w in the separation schemes (3) and (4p), and
(5) and (6p). It is understood that smaller values of w will reduce computing on the boundary domain with a dense
mesh. From the proof of the reduced-problem theorem, we need w
2 ln
k
. As noted, w can be chosen as a
constant for a family of small values of the singular perturbation parameter . Indeed, for the singular perturbation
parameter down to the 10
12
level, w = 2 28/k is sufcient; for the singular perturbation parameter down to the
10
15
level, w = 2 35/k is sufcient; and for the singular perturbation down to the 10
20
level, w = 2 47/k is
sufcient. For extreme small values of , =
k
k
2
4m
2

k
, we can choose w
ln
k
such that the above
s(x) = c
1
e
(xa)
c
1
e
k(xa)
< c
1
. Practically, the choice of w can be further improved.
3. The error analysis
The numerical error of the new method has two sources: one from the analytical approximation and the other from
the numerical approximation. Let h
n
and h
b
be the mesh spacing on the non-boundary layer and on the boundary layer
domain respectively.
3.1. Error on the non-boundary layer domain
Let u be the exact solution, R be the exact solution of the reduced problem, and R
N
be the numerical solution of
the reduced problem. Assume R
N
is obtained from the fourth- order Runge Kutta method. On the non-boundary layer
domain, the error is
u R
N
= max
i =1...n1
{|u(x
i
) R
N
(x
i
)|}.
By the triangle inequality, we conclude
u R
N
u R +R R
N
= O() + O(h
4
n
).
3.2. Error on the boundary layer domain
On the boundary layer domain, the analytical approximation error is generated from the -perturbation on the
boundary value and the numerical approximation error from the numerical methods. By the stability theorem, the
analytical-approximation error is O(), where is the perturbation on the boundary value. By the reduced-problem
theorem, = O() + O(h
4
n
). To simplify the discussion, we consider the central nite difference scheme with mesh
spacing h
b
. Then the error of numerical approximation is O(h
2
b
). Let u be the exact solution of the SPP (1). For the
SPP with -perturbation on a boundary value, let u
be the exact solution and u

N
be the numerical solution. On the
boundary layer domain, the error is u u
N
. By the triangle inequality, we conclude
u u
N
u u
+u
u
N
= O() + O(h
4
n
) + O(h
2
b
).
The new method works well for singular perturbation problems since the singular perturbation parameter is
extremely small.
4. Numerical experiments
Example 1. We consider the linear SPP
u
= u
for x (0, 1),

u(0) = 0 and u(1) = 1,
(8)
with the exact solution u(x) =
e
x1
e
1
1e
1
. We have f (x, u) = 1 > 0, which satisfy the assumptions for the reduced
problem theorem. Thus, the boundary layer occurs about the right boundary x = 1, which is shown in Fig. 1.
Fig. 1. Graphs of the solutions of the SPP (8) and its reduced equation.
Table 1.1
Maximal error comparison among methods with N = 60
Method Maximal error
= 10
1
= 10
2
= 10
3
= 10
4
= 10
5
= 10
8
= 10
10
Uniform mesh with the central
differences
7.2 10
4
1.9 10
2
N.A. N.A. N.A. N.A. N.A.
The 4th order
Schultz and Choo [3]
8.2 10
6
2.5 10
3
The order with mapping of
1.5 10
7
4.0 10
6
3.9 10
6
1.1 10
5
2.5 10
5
N.A. N.A.
The new method with central
differences
7.2 10
4
6.7 10
4
6.7 10
4
6.7 10
4
6.7 10
4
6.7 10
4
6.7 10
4
The new method with the 4th
order
8.2 10
6
1.1 10
6
1.1 10
6
1.1 10
6
1.1 10
6
1.1 10
6
1.2 10
6
Applying the new method, we get two separated differential equations,
u
= 0 for 0 < x < t and u(0) = 0,

and
u
= u
for x (t, 1),

u(t ) = R
a
(t ) and u(1) = 1,
where t = 1w and w is the boundary layer parameter. The numerical results are signicantly improved comparing
to that of the global uniform-mesh methods. With N = 60 mesh points, the uniform methods provide numerical results
when the singular perturbation parameter is great than 10
3
. The new method with the central nite difference or
fourth-order nite difference provides accurate numerical results for much smaller values of , Table 1.1.
With N = 1000 mesh points, the uniform methods provide a numerical result when the singular perturbation
parameter is great than 10
3
. The new method with central nite differences or fourth-order nite differences
provides better numerical results with fewer mesh points (N = 120). With the same number N = 120 of mesh points,
it generates numerical results for much smaller values of while the error is maintained at the same level, O(10
5
)
and O(10
7
) for the central differences and fourth order differences respectively, Table 1.2. For the new method, there
are 100 points on the boundary layer and 20 points on the non-boundary layer domain.
Table 1.2
Maximal error comparison among different methods
Method N Maximal error
= 10
3
= 10
5
= 10
8
Uniform with the central differences 1000 1.28 10
2
N.A. N.A.
Uniform with the 4th order (Schultz and Choo [3]) 1000 7.45 10
4
N.A. N.A.
The new method with the central differences 120 5.66 10
5
5.66 10
5
5.66 10
5
The new method with the fourth order (Schultz and Choo [3]) 120 3.22 10
7
3.22 10
7
3.22 10
7
Table 1.3
The robustness of the new method
Points New method = 10
8
= 10
9
= 10
10
N
b
= 50 With 2nd order differences 6.66 10
4
6.66 10
4
6.66 10
4
With 4th order differences 1.09 10
6
1.09 10
6
1.09 10
6
N
b
= 100 With 2nd order differences 5.66 10
5
5.66 10
5
5.66 10
5
With 4th order differences 3.22 10
7
3.22 10
7
3.22 10
7
N
b
is the number of points in the boundary layer.
The new method demonstrates a great advantage for singular perturbation problems with extremely small values
of the singular perturbation parameter. With a constant number of mesh points, the numerical error is maintained at
the same level for a family of singular perturbation problems, Table 1.3.
For the following nonlinear examples, the numerical results on the boundary domain are obtained with the
generalized Newtons method. Note for the methods of this paper, we need only iterate on the boundary layer
equations.
Example 2.
u
+e
u
u

2
sin
x
2
e
2u
= 0 for x (0, 1),
u(0) = 0 and u(1) = 0,
(9)
with the asymptotic expansion
u(x, ) = ln
_
_
1 +cos
x
2
_
_
1
1
2
e
x
2
__
+ O(),
OMalley [2]. Note that since u is bounded, we can nd a positive constant k, such that f (x, u) = e
u
k < 0.
By the reduced problem theorem, we expect a left boundary layer as shown in Fig. 2.
We apply the new method with the separation scheme (3) and (4) to get two differential equations
e
u
u

2
sin
t
2
e
2u
= 0 for x (t, 1),
u(1) = 0,
and
u
+e
u
u

2
sin
x
2
e
2u
= 0 for x (0, t ),
u(0) = 0 and u(t ) = R
b
(t ),
where t = w and R
b
is the solution of above reduced equation. We compare the numerical results of the new method
with different numbers of mesh points. As expected, the new method is convergent and stable. Moreover, we use a
constant number of mesh points for a family of singular perturbation problems to control the numerical error at the
same level, Table 2.1.
Fig. 2. Graphs of the solutions of the SPP (9) and its reduced equation.
Table 2.1
The new method Maximal error
= 10
3
= 10
4
= 10
5
= 10
6
= 10
7
= 10
8
= 10
9
= 10
10
N
b
= 300 4.3610
5
3.7810
5
3.8210
5
3.8210
5
3.8210
5
3.8210
5
3.8210
5
3.82 10
5
N
b
= 600 4.7710
5
9.2610
6
9.5610
6
9.5710
6
9.5710
6
9.5710
6
9.5710
6
9.57 10
6
N
b
= 1500 4.9910
5
1.5210
6
1.8010
6
1.8310
6
1.8310
6
1.8410
6
1.8410
6
1.84 10
6
N
b
is the number of points on the boundary layer, N
n
= 100 is the number of points on the non-boundary layer.
Table 2.2
Maximal error comparison when = 10
10
Method Number of points Max error
Linss, Roos and Vulanovic [8] Standard Shishkin mesh 1024 1.21 10
2
Linss, Roos and Vulanovic [8] BakhvalovShishkin mesh 1024 2.63 10
3
The new method of the central differences 400 Non-boundary layer 100 3.82 10
5
Boundary layer 300
6
Boundary layer 600
6
Boundary layer 900
We compare the new method with the methods using Shishkin type meshes from Linss, Roos and Vulanovic [8].
As shown in Table 2.2, the numerical results are improved using fewer mesh points, compared to that of the Shishkin
mesh methods.
Example 3. We consider the SPP,
u
uu
= 0 x (1, 1),
u(1) = 0 and u(1) = 1.
(10)
The exact solution in [1, 1] is given as
u(x, ) =
1 e
(x+1)
1 +e
(x+1)
,
Table 3.1
Method Points Maximal error
= 10
2
= 10
3
= 10
4
= 10
5
= 10
6
= 10
7
= 10
8
Uniform N = 1000 3.38 10
4
3.84 10
2
New method N
b
= 50 1.36 10
2
1.36 10
2
1.36 10
2
1.36 10
2
1.36 10
2
1.36 10
2
1.36 10
2
New method N
b
= 100 3.39 10
3
3.39 10
3
3.39 10
3
3.39 10
3
3.39 10
3
3.39 10
3
3.39 10
3
New method N
b
= 200 8.41 10
4
8.41 10
4
8.41 10
4
8.41 10
4
8.41 10
4
8.41 10
4
8.41 10
4
New method N
b
= 300 3.77 10
4
3.77 10
4
3.77 10
4
3.77 10
4
3.77 10
4
3.77 10
4
3.77 10
4
The uniform method and the new method are applied with central nite differences.
OMalley [2]. Note that the reduced-problem theorem cannot be applied since 1 f (x, u) = u 0. But the
reduced problem is a good approximation for the SPP on the non-boundary layer domain, which is veried by our
numerical experiment. We apply the new method with the separation scheme to get two differential equations
uu
= 0 for x (t, 1),

u(1) = 1,
and
u
= uu
for x (1, t ),
u(1) = 0 and u(t ) = R
b
where t = 1 +w.
Choose the boundary layer parameter w = 30 for 10
12
. The boundary layer is at the left boundary x = 1. The
transitional point is t = 1 +w.
In addition to the comparison between the uniform-mesh method and the newmethod, the stability and convergence
is veried as the number of mesh points increases on the boundary layer domain, Table 3.1. As expected, the new
method works well for all 10
3
.
4.1. An improvement on the fourth order method
In Schultz and Ilicasu [4], the problem (10) was solved with the fourth-order method for the singular perturbation
parameter = 10
2
and = 10
3
. The maximum error was 1.0 10
5
and 1.30 10
2
respectively. For simplicity,
let =
1
where is the singular perturbation parameter in the problem (10). Then

u
= uu
,
u
= (u
)
2
+uu
, and
u
(4)
= 2u
+u
+uu
= 3u
+
2
u(u
)
2
+
2
u
2
u
=
2
u(u
)
2
+(3u
+
2
u
2
)u
.
By allowing A
3
= u
i
, B
3
= u
i
, and A
4
=
2
u
i
u
i
, B
4
= 3u
i
+
2
u
2
i
, we get
u
(3)
i
= A
3
u
i
+ B
3
u
i
,
u
(4)
i
= A
4
u
i
+ B
4
u
i
.
Dene the following:
A
=
h
4
A
4
24
,
B
= h +
h
3
A
3
6
,
C
=
h
2
2
+
h
4
B
4
24
,
D
=
h
3
B
3
6
.
Following the approach of Schultz and Ilicasu [4], we use the fourth-order nite differences to approximate the SPP
(10).
u
i
u
i
u
i
c
3
u
i +1
+c
2
u
i
+c
1
u
i 1
where
c
3
=
u
i
D
+A
+u
i
C
2(A
)
,
c
1
=
u
i
D
(A
+u
i
C
)
2(A
)
,
c
2
= (c
3
+c
1
) =
u
i
D
+B
.
According to Schultz and Ilicasu [4], the derivative u
i
in A
3
, A
4
and B
4
is replaced with u
i
=
u
i +1
u
i 1
2h
. We now use
fourth-order nite differences to develop a higher-order approximation to the derivative u
i
. For i = 1, 2, . . . , N 1,
let the rst derivative be
u
i
= c
3
u
i +1
+c
2
u
i
+c
1
u
i 1
where c
3
, c
2
andc
1
are constants.
Using Taylor series expansion,
u
i
= c
3
_
u
i
+hu
i
+
h
2
2
u
i
+
h
3
6
u
i
+
h
4
24
u
(4)
i
+
_
+c
2
u
i
+c
1
_
u
i
hu
i
+
h
2
2
u
i

h
3
6
u
i
+
h
4
24
u
(4)
i
+
_
c
3
_
u
i
+hu
i
+
h
2
2
u
i
+
h
3
6
(u
2
i
+u
i
u
i
) +
h
4
24
_
2
u
i
u
2
i
+(3u
i
+
2
u
2
i
)u
i
_
_
+c
2
u
i
+c
1
_
u
i
hu
i
+
h
2
2
u
i

h
3
6
(u
2
i
+u
i
u
i
) +
h
4
24
_
2
u
i
u
2
i
+(3u
i
+
2
u
2
i
)u
i
_
_
= (c
3
+c
2
+c
1
)u
i
+
_
(c
3
c
1
)
_
h +
h
3
6
u
i
_
+(c
3
+c
1
)
h
4
24
2
u
i
u
i
_
u
i
+
_
(c
3
+c
1
)
_
h
2
2
+
h
4
24
3u
i
+
h
4
24
2
u
2
i
_
+(c
3
c
1
)
h
3
6
u
i
_
u
i
.
Setting the corresponding coefcients equal, we have the following system of equations
c
3
+c
2
+c
1
= 0,
(c
3
c
1
)
_
h
3
6
u
i
_
(c
3
+c
1
)
h
4
24
2
u
i
u
i
= 1,
(c
3
+c
1
)
_
h
2
2
+
h
4
24
3u
i
+
h
4
24
2
u
2
i
_
+(c
3
c
1
)
h
3
6
u
i
= 0.
By letting
A =
h
4
24
2
u
i
u
i
,
B = h +
h
3
6
u
i
,
C =
h
2
2
+
h
4
24
3u
i
+
h
4
24
2
u
2
i
,
D =
h
3
6
u
i
,
the system of equations is equivalent to
c
3
+c
2
+c
1
= 0,
(c
3
+c
1
)A +(c
3
c
1
)B = 1,
(c
3
+c
1
)C +(c
3
c
1
)D = 0,
of which, c
3
+c
1
and c
3
c
1
are,
c
3
+c
1
=
D
AD BC
,
c
3
c
1
=
C
AD BC
.
Therefore, we get
c
3
=
D C
2(AD BC)
,
c
1
=
D +C
2(AD BC)
,
c
2
= (c
3
+c
1
) =
D
AD BC
.
The error term is
h
5
120
(c
3
u
(5)
(
1
) +c
1
u
(5)
(
2
)) where
1
,
2
[x
i
h, x
i
+h].
In comparison to the approach of Schultz and Ilicasu [4], c
3
and c
1
are updated to the fourth-order accuracy. The
improvement of the method is veried by numerical experiments, Tables 3.2 and 3.3.
4.2. The fth- and sixth order methods
The second improvement is to add more terms from the Taylor series to approximate the nonlinear problem (10).
We expand the u
i +1
and u
i 1
up to the sixth order derivatives. The following is a development of the sixth order
method. The fth-order method is developed by dropping sixth order derivative terms. First, we consider the fth-
order and sixth order derivatives:
u
(5)
= 3u
2
+3u
+
2
u
2
+2
2
uu
+2
2
uu
+
2
u
2
u
= 3u
2
+3u
(u
2
+uu
) +
2
u
3
+4
2
uu
+
2
u
2
(u
2
+uu
)
= 3u
2
+3
2
u
3
+3
2
uu
+
2
u
3
+4
2
uu
+
3
u
2
u
2
+
3
u
3
u
= 3u
2
+4
2
u
3
+7
2
uu
+
3
u
2
u
2
+
3
u
3
u
= (4
2
u
2
+
3
u
2
u
)u
+(3u
+7
2
uu
+
3
u
3
)u
and
u
(6)
= 6u
+12
2
u
2
u
+7
2
[(u
2
+uu
)u
+uu
]
+2
3
uu
3
+2
3
u
2
u
+3
3
u
2
u
+
3
u
3
u
= 6u
(u
2
+uu
) +12
2
u
2
u
+7
2
[u
2
u
+uu
2
+uu
(u
2
+uu
)]
+2
3
uu
3
+5
3
u
2
u
+
3
u
3
(u
2
+uu
)
= 6
2
u
2
u
+6
2
uu
2
+12
2
u
2
u
+7
2
u
2
u
+7
2
uu
2
+7
3
uu
3
+7
3
u
2
u
+2
3
uu
3
+5
3
u
2
u
+
4
u
3
u
2
+
4
u
4
u
= 25
2
u
2
u
+13
2
uu
2
+7
3
uu
3
+2
3
uu
3
+12
3
u
2
u
+
4
u
3
u
2
+
4
u
4
u
= (9
3
uu
2
+
4
u
3
u
)u
+(25
2
u
2
+13
2
uu
+12
3
u
2
u
+
4
u
4
)u
.
For simplicity, we rewrite the derivatives
u
(3)
i
= A
3
u
i
+ B
3
u
i
where A
3
= u
i
, B
3
= u
i
,
u
(4)
i
= A
4
u
i
+ B
4
u
i
where A
4
=
2
u
i
u
i
, B
4
= 3u
i
+
2
u
2
i
,
u
(5)
i
= A
5
u
i
+ B
5
u
i
where A
5
= 4
2
u
2
i
+
3
u
2
i
u
i
, B
5
= 3u
i
+7
2
u
i
u
i
+
3
u
3
i
,
and
u
(6)
i
= A
6
u
i
+ B
6
u
i
where A
6
= 9
3
u
i
u
2
i
+
4
u
3
i
u
i
, and
B
6
= 25
2
u
2
i
+13
2
u
i
u
i
+12
3
u
2
i
u
i
+
4
u
4
i
.
We write
u
i
u
i
u
i
= c
3
u
i +1
+c
2
u
i
+c
1
u
i 1
,
where c
3
, c
2
and c
1
are constants. By Taylor series expansion, we obtain
u
i
u
i
u
i
= c
3
u
i +1
+c
2
u
i
+c
1
u
i 1
c
3
_
u
i
+hu
i
+
h
2
2
u
i
+
h
3
6
u
i
+
h
4
24
u
(4)
i
+
h
5
120
u
(5)
i
+
h
6
720
u
(6)
i
_
+c
2
u
i
+c
1
_
u
i
hu
i
+
h
2
2
u
i

h
3
6
u
i
+
h
4
24
u
(4)
i

h
5
120
u
(5)
i
+
h
6
720
u
(6)
i
_
= c
3
_
u
i
+hu
i
+
h
2
2
u
i
+
h
3
6
(A
3
u
i
+ B
3
u
i
) +
h
4
24
(A
4
u
i
+ B
4
u
i
) +
h
5
120
(A
5
u
i
+ B
5
u
i
)
+
h
6
720
(A
6
u
i
+ B
6
u
i
)
_
+c
2
u
i
+c
1
_
u
i
hu
i
+
h
2
2
u
i

h
3
6
(A
3
u
i
+ B
3
u
i
)
+
h
4
24
(A
4
u
i
+ B
4
u
i
)
h
5
120
(A
5
u
i
+ B
5
u
i
) +
h
6
720
(A
6
u
i
+ B
6
u
i
)
_
= c
3
_
u
i
+
_
h +
h
3
A
3
6
+
h
4
A
4
24
+
h
5
A
5
120
+
h
6
A
6
720
_
u
i
+
_
h
2
2
+
h
3
B
3
6
+
h
4
B
4
24
+
h
5
B
5
120
+
h
6
B
6
720
_
u
i
_
+c
2
u
i
+c
1
_
u
i
+
_
h
h
3
A
3
6
+
h
4
A
4
24

h
5
A
5
120
+
h
6
A
6
720
_
u
i
+
_
h
2
2

h
3
B
3
6
+
h
4
B
4
24

h
5
B
5
120
+
h
6
B
6
720
_
u
i
_
= (c
3
+c
2
+c
1
)u
i
+
_
(c
3
+c
1
)
_
h
4
A
4
24
+
h
6
A
6
720
_
+(c
3
c
1
)
_
h +
h
3
A
3
6
+
h
5
A
5
120
__
u
i
+
_
(c
3
+c
1
)
_
h
2
2
+
h
4
B
4
24
+
h
6
B
6
720
_
+(c
3
c
1
)
_
h
3
B
3
6
+
h
5
B
5
120
__
u
i
.
Equating the coefcients of both sides, we get the following system of equation in terms of c
3
, c
2
and c
1
:
c
3
+c
2
+c
1
= 0,
(c
3
+c
1
)A
+(c
3
c
1
)B
= u
i
,
(c
3
+c
1
)C
+(c
3
c
1
)D
= ,
where
A
=
h
4
A
4
24
+
h
6
A
6
720
,
B
= h +
h
3
A
3
6
+
h
5
A
5
120
,
C
=
h
2
2
+
h
4
B
4
24
+
h
6
B
6
720
,
D
=
h
3
B
3
6
+
h
5
B
5
120
.
Note that the derivatives contained in A
3
, A
4
, A
5
and B
3
, B
4
, B
5
are replaced with the following:
u
i
=
u
i +1
u
i 1
2h
, u
i
=
u
i +1
2u
i
+u
i 1
h
2
.
Thus, we get
c
3
+c
1
=
u
i
D
,
c
3
c
1
=
A
+u
i
C
.
Therefore, the solution is
c
3
=
u
i
D
+A
+u
i
C
2(A
)
,
c
1
=
u
i
D
(A
+u
i
C
)
2(A
)
,
c
2
= (c
3
+c
1
) =
u
i
D
+B
.
The error term is
h
7
5040
(c
3
u
(7)
(
3
) +c
1
u
(7)
(
4
)) where
3
,
4
[x
i
h, x
i
+h].
The results are signicantly improved as shown by numerical experiments, Tables 3.23.5.
For N
n
= 170 mesh points on the non-boundary layer, the maximal error is 10
8
from the new method. On
the boundary layer domain, the numerical results of the improved fourth-order method, fth-order method and sixth
order method are signicantly improved compared to that of the second-order central differences. Table 3.1 shows the
numerical comparison of the second-order central differences for the uniform method and the new method. The new
method is convergent and stable.
Table 3.2 shows the numerical comparison between the different methods. For the methods of Choudhury [9] and
Schultz and Ilicasu [4], a uniform mesh is used with N = 2000 mesh points. For the improved fourth-order, fth-
order and sixth order methods, the number N
n
of mesh points on the non-boundary layer domain is 170 and the
number N
b
of mesh points on the boundary layer is 300. The total number of mesh points for our separation approach
is N = N
n
+ N
b
= 470, which is much less than that of the uniform methods. As shown in the tables, the numerical
results are signicantly improved. In addition, the computing cost is reduced signicantly too, which is reected from
the number of mesh points and iterations.We compare our method with the method of Choudhury [9] and the method
of Schultz and Ilicasu [4]. For 10
3
, the method of Choudhury and the second-order method did not provide
accurate results. When = 0.01, the error from Choudhurys method is about 10
2
. The error of the fourth-order
method of Schultz and Ilicasu [4] is about 10
5
. The numerical results of the new method are much better. The
maximal error of the new method with only 470 points is 10
5
for the improved fourth-order method, 10
7
for the
fth-order method and 10
8
for the sixth order method. Choudhury only gave results for = 0.01. For 10
4
,
Table 3.2
Maximal error comparison among different methods with = 0.01
Method Number of points Number of iterations Max error
Choudhurys Method [9] 2000 Not known 2.91 10
2
Second order of Schultz and Ilicasu [4] 2000 3201 2.61 10
4
Fourth order of Schultz and Ilicasu [4] 2000 3152 1.00 10
5
Improved fourth-order method of this paper 470 697 8.40 10
5
Fifth order-method of this paper 470 697 1.34 10
7
Sixth order method of this paper 470 697 7.71 10
8
For the methods of this thesis, the tolerance of iteration is set at 10
10
.
Table 3.3
Maximal error comparison among different methods with = 0.001
Method Number of points Number of iterations Max error
Fourth order of Schultz and Ilicasu [4] 2000 5102 1.30 10
2
Improved fourth-order method of this paper 470 697 8.41 10
5
Fifth- order method of this paper 470 697 1.54 10
7
Sixth order method of this paper 470 697 7.56 10
8
For the methods of this thesis, the tolerance of iteration is set at 10
10
.
Table 3.4
The convergence of the new method with the improved fourth-order differences
Number of Points Maximal error
= 10
5
= 10
10
= 10
12
N = 350 (N
n
= 200, N
b
= 150) 3.53 10
4
3.53 10
4
3.56 10
4
N = 400 (N
n
= 200, N
b
= 200) 1.91 10
4
1.94 10
4
1.94 10
4
N = 450 (N
n
= 200, N
b
= 250) 1.20 10
4
1.22 10
4
1.35 10
4
N = 500 (N
n
= 200, N
b
= 300) 8.39 10
5
8.40 10
5
9.71 10
5
Table 3.5
The convergence of the new method with the sixth order differences
Number of points Maximal error
= 10
5
= 10
10
= 10
12
N = 350 (N
n
= 200, N
b
= 150) 1.19 10
6
1.16 10
6
2.79 10
5
N = 400 (N
n
= 200, N
b
= 200) 4.51 10
7
4.18 10
7
2.64 10
5
N = 450 (N
n
= 200, N
b
= 250) 1.29 10
7
2.88 10
7
2.42 10
5
N = 500 (N
n
= 200, N
b
= 300) 7.49 10
8
7.49 10
8
2.39 10
5
the method of Schultz [4] did not provide accurate results. The error of the fourth-order method of Schultz [4] is
about 10
2
when = 0.001. The numerical results of the new methods are much better. The maximal error of the new
method is achieved at 10
5
for the improved fourth order nite differences, 10
7
for the fth-order method and 10
8
for the sixth order nite differences.
The power of the new method is that it works for extreme small values of the singular perturbation parameter.
The convergence of the improved fourth order method and 6th order method is shown in Tables 3.4 and 3.5 with the
smallest values of singular perturbation parameter = 10
12
. For small values of , as the number of mesh points
increases, the error is getting smaller.
We compare the numerical results of the new method with different nite differences, Table 3.6. The numerical
error decreases when the higher-order nite difference methods are applied.
Table 3.6
Maximal error comparison among the different methods with 300 points on the boundary layer
The new method Maximal error
= 10
5
= 10
10
With central difference 3.77 10
4
3.77 10
4
With the 4th order difference 1.02 10
4
1.02 10
4
With the improved 4th order 8.40 10
5
8.40 10
5
With the 6th order difference 7.49 10
8
7.49 10
8
References
[1] K.W. Chang, F.A. Howes, Nonlinear Singular Perturbation Phenomena: Theory and Application, Springer-Verlag, 1984.
[2] R.E. OMalley Jr., Introduction to Singular Perturbations, Academic Press, New York, 1974.
[3] D.H. Schultz, J.Y. Choo, High order methods for differential equations with small coefcients for the second order terms, Comput. Math. Appl.
25 (1) (1993) 105123.
[4] D.H. Schultz, F.O. Ilicasu, High order methods for singular perturbation problems, Comput. Math. Appl. 47 (2002) 391417.
[5] J.J.H. Miller, E. ORiordan, G.I. Shishkin, Fitted Numerical Methods for Singular Perturbation Problems: Error Estimates in the Maximum
Norm for Linear Problems in One and Two Dimensions, World Scientic, 1996.
[6] A. Segal, Aspects of numerical methods for elliptic singular perturbation problems, SIAM J. Sci. Comput. 3 (3) (1982) 327349.
[7] W. Zhang, Numerical solutions of linear and nonlinear singular perturbation problems, Ph. D. Dissertation, University of Wisconsin Milwaukee,
May 2006.
[8] T. Linss, H. Roos, R. Vulanovic, Uniform pointwise convergence on Shishkin type meshes for Quasi-linear convectiondiffusion problems,
SIAM J. Numer. Anal. 38 (3) (2000) 897912.
[9] S.R. Choudhury, Nonstandard difference schemes of nonlinear singular perturabtion problems, Internat. J. Appl. Sci. Comput. 2 (3) (1996)
375392.

Numerical Solutions of Linear and Nonlinear Singular Perturbation Problems PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Numerical Solutions of Linear and Nonlinear Singular Perturbation Problems PDF

Uploaded by

Copyright:

Available Formats

Computers and Mathematics with Applications 55 (2008) 25742592

+ g(x, u) for x (a, b) and f (x, u) = 0,

+ g(x, u) of Eq. (1) satises a

(x)| N on J. (x) and (x) are functions of C

+ g(x, u) = 0 for x (t, b),

+ g(x, u) for x (a, t ),

+ g(x, u) = 0 for x (a, t ),

+ g(x, u) for x (t, b),

+ g(x, u) = 0 for x (t, b),

+ g(x, u) for x (a, t ),

+ g(x, u) = 0 for x (a, t ),

+ g(x, u) for x (t, b),

+ms = 0 for x (a, t )

+ g(x, p)] = [ f (x, ) p

+h(x, )s, since f (x, ) k < 0 and s

ms, since |h(x, )| = | f

+ g(x, p)] [ f (x, )

h(x, )( p), since f (x, ) k < 0 f (x, ) k

+ms, since |h(x, )| m h(x, ) m

(x) is negative. Therefore, we have

be the exact solution and u

for x (0, 1),

= 0 for 0 < x < t and u(0) = 0,

for x (t, 1),

= 0 for x (t, 1),

where is the singular perturbation parameter in the problem (10). Then

You might also like