A PARAMETER-UNIFORM TAILORED FINITE POINT METHOD FOR

SINGULARLY PERTURBED LINEAR ODE SYSTEMS

Houde Han
Department of Mathematical Sciences, Tsinghua University, Beijing 100084, China
Email: hhan@math.tsinghua.edu.cn.
J. J. H. Miller
Department of Mathematics, Trinity College, Dublin 2, Ireland
Email: jmiller@tcd.ie.
Min Tang
Key Laboratory of Scientific and Engineering Computing,
Department of Mathematics and Institute of Natural Sciences, Shanghai Jiao Tong University,
Shanghai, China, 200240.
Email: tangmin@sjtu.edu.cn
Abstract
In scientific applications from plasma to chemical kinetics, a wide range of temporal scales can present
in a system of differential equations. A major difficulty is encountered due to the stiffness of the system and
it is required to develop fast numerical schemes that are able to access previously unattainable parameter
regimes
We consider an initial-final value problem for a multi-scale singularly perturbed system of linear ordinary
differential equations with discontinuous coefficients. We construct a tailored finite point method, which
yields approximate solutions that converge in the maximum norm, uniformly with respect to the singular
perturbation parameters, to the exact solution. A parameter-uniform error estimate in the maximum norm
is also proved. The results of numerical experiments, that support the theoretical results, are reported.
Mathematics subject classification: 37M05, 65G99
Key words: Tailored finite point method, parameter uniform, singular perturbation, ODE system

1. Introduction
We consider the following initial-final value problem for a system of linear ordinary differential equations with
discontinuous coefficients
Eu0 (t) + A(t)u(t)

= f (t),

u(pk + 0) u(pk 0)

B u(0) + (I B ) u(1)

0,

t (pk , pk+1 ),
k = 1, . . . , K,

= d,

k = 0, . . . , K,

(1.1)
(1.2)
(1.3)

where E, B are n n matrices, d is a vector, A(t) is an n n matrix function and f (t) is a vector function on
the interval [0, 1] such that


A(t) = ai,j (t) nn ,
(1.4)
f (t)

(f1 (t), f2 (t), . . . , fn (t))T

(1.5)

and {pk }K+1

are some numbers which satisfy
0
0 = p0 < p1 < < pK < pK+1 = 1.
The given functions ai,j (t), fi (t) (1 i, j n) may not be continuous on the whole interval [0, 1]. Here,
we consider the case when they are piecewise continuous. More precisely, we assume that the functions
http://www.global-sci.org/jcm

Global Science Preprint

2
ai,j (t), fi (t) (1 i, j n) have K points of discontinuity of the first kind, t = pk , (0 < pk < 1; k = 1, . . . , K), so
that on each subinterval (pk , pk+1 ), (k = 0, . . . , K; p0 = 0, pK+1 = 1) the functions are smooth and satisfy the
conditions

ai,i (t)

n
X

|ai,j (t)|

> 0,

t [0, 1].

(1.6)

j=1,j6=i

Furthermore, we assume that the matrices E = diag(1 , 2 , . . . , n ) and B = diag(b1 , b2 , . . . , bn ) are diagonal
and satisfy the conditions
|i | > 0, 1 i n
and

(
bi =

1,
0,

(1.7)

i > 0,
i < 0.

(1.8)

We also suppose that there exists at least one j (1 j n) such that

0 < |j |  1.

(1.9)

Problem (1.1)-(1.3) is then an initial-final value problem for a multi-scale singularly perturbed system of linear
ordinary differential equations with discontinuous coefficients. The solution of problem (1.1)-(1.3) may contain
initial, final and interior layers at any of the points pk (k = 1, . . . , K). The main goal in this paper is to develop
a class of numerical methods, which yield approximate solutions that converge in the maximum norm, uniformly
with respect to the singular perturbation parameters, to the exact solution of this problem.
When all of the parameters (j , j = 1, . . . , n) are positive, problem (1.1)-(1.3) reduces to the initial value
singularly perturbed problem
Lu(t) Eu0 (t) + A(t)u(t)
u(pk + 0) u(pk 0)
u(0)

= f (t),
=

0,

t (pk , pk+1 ),

k = 0, . . . , K,

k = 1, . . . , K,

= d,

(1.10)
(1.11)
(1.12)

which has been studied in [23].They proposed a Shishkin piecewise uniform mesh with a classical finite difference
scheme to obtain numerical solutions of this problem; a parameter-uniform error estimate was also given.
To motivate the study of the more general initial-final value problem in the paper, it should be noted that a
semi-discretization, with respect to variable x, of the following forward-backward parabolic problem

sign(x)|x|p

u(x, t)
t
ux=1

ut=0

2 u(x, t)
(x)u(x) + q(x, t),
1 < x < 2, 0 < t < T,
x2

= f (t),
ux=2 = g(t), 0 < t < T,

= s(x), 0 < x < 2,
ut=1 = (x), 1 < x < 0. ,

(1.13)
(1.14)
(1.15)

with p > 1, (x) > 0 can lead to the initial-final singularly perturbed problem (1.1)-(1.3). This is because
|x|p (0, 2p ) ranges from very small to 2p .
In this paper we construct a parameter-uniform scheme for the initial-final multi-scale singularly perturbed
problem(1.1)-(1.3) using the tailored finite point method (TFPM). The TFPM was proposed by Han, Huang
and Kellogg in [9] for the numerical solution of singular perturbation problems. The basic idea of the TFPM is

3
to choose, at each mesh point, suitable basis functions based on the local properties of the solutions of the given
problem; then to approximate the solution using these basis functions. At each point, the numerical scheme
is tailored to the given problem. The TFPM was successfully applied by Han, Huang ,and Kellogg to solve
the Hemker problem [9, 11]; they won the Hemker prize at the international conference BAIL 2008. Later the
TFPM was developed to solve the second-order elliptic singular pertubation problem [6, 21], the first order wave
equation [13], the one-dimension Helmholtz equation with high wave number [5], second order elliptic equations
with rough or highly oscillatory coefficients [10] and so on [7, 8]. For the one dimensional singular perturbation
problem the TFPM is close to the method of exponential fitting discussed in [1, 3, 16, 20]. The TFPM is also
applied to the one-dimensional discrete-ordinate transport equations in [17].
The TFPM uses the functions that exactly satisfy the PDE as the bases. For the linear ordinary differential
equations system under consideration, the bases are exponential functions. The exponential integrator appears
for a long time, see the review paper [19]. The new idea in TFPM is to approximate all the coefficients A(t)
and f (t) by piecewise contants. Compared with previous magnus integrators [12] and adiabatic integrators [18]
developed for the stiff problem or highly oscillatory problems, our approach is simple and proved to possess
parameter uniform convergence. More precisely, we can use time steps much larger than the parameters i in
(1.7) and achieve stable and accurate results.
The main contribution of this paper is that we construct a tailored finite point method for (1.1)-(1.3), which
yields approximate solutions that converge in the maximum norm, uniformly with respect to the singular
perturbation parameters, to the exact solution. The parameter-uniform error estimate in the maximum norm is
proved analytically as stated in Theorem 4.1. Some numerical experiments that support the theoretical results
are presented in section 5.

2. Existence and uniqueness of solutions to the problem (1.1)-(1.3)

In this section we discuss the existence and uniqueness of solutions to the problem (1.1)-(1.3). On [0, 1] we
introduce the following function spaces
(0)


= {v(t)v|(pk ,pk+1 ) C (0) (pk , pk+1 ), k = 0, 1, . . . , K},



(0)
(1)
C [0, 1], =
v(t)v(t) C (0) [0, 1], and v 0 (t) C [0, 1] ,



(1)
(0)
C, [0, 1] =
v(t)v(t), v 0 (t) C [0, 1] .
C, [0, 1]

(0)

(2.1)
(2.2)
(2.3)

(0)

Note that the space C (0) [0, 1] is a subspace of C [0, 1] and that, for any v(t) C, [0, 1], v(t) may have
discontinuity points of the first kind t = pk (k = 1, . . . , K). On each subinterval [pk , pk+1 ] we adopt the
(1)
conventions that v(pk ) = limtp+ v(pk ), v(pk+1 ) = limtp v(pk+1 ). The space C [0, 1] is a subspace of
k

k+1

(1)

C, [0, 1].
(0)
For any v(t) C, [0, 1], we define the norm

kvk

max

0kK

max

pk tpk+1

|v(t)|

(2.4)

(1)

and for any v(t) C, [0, 1], we define the norms

|v|1,

kvk1,


max
max |v 0 (t)| ,
pk tpk+1


max kvk , |v|1, .

0k,K

Furthermore, we introduce the following function vector and matrix spaces

(2.5)
(2.6)

(1)

C [0, 1]
(1)
C, [0, 1]
(1)

C, [0, 1]
(1)

(1)

n

C [0, 1]

(2.7)

n
(1)
C, [0, 1] ,

(2.8)

(1)

C, [0, 1]

(1)

nn

(2.9)

(1)

Note that C [0, 1] is a subspace of C, [0, 1]. For any v(t) C, [0, 1], we define the norms kk , ||1, , kk1,
kvk

|v|1,

kvk1,

max

max

pk tPk+1

0kK

kv(t)k ,

(2.10)

kv0 k ,

(2.11)
0

max kvk , kv k ,

(2.12)

where the norm kv(t)k is the norm in the vector space Rn for fixed t [0, 1]. Similarly we define the norms
(1)
k k , | |1, , k k1, in the space C, [0, 1].
For problem (1.1)-(1.3), we then have the following theorem.
(0)

(0)

Theorem 2.1. Suppose that A(t) C, [0, 1], f (t) C, [0, 1] and A(t), f (t) satisfy the conditions (1.4)-(1.9).
(1)
Then, for problem (1.1)-(1.3) there exists a unique solution u (t), u (t) C [0, 1], and the following estimate
holds

kf k
ku k max kdk ,
.

(2.13)
(1)

Proof. We first establish the estimate (2.13). Suppose that u (t) = (u1 (t), . . . , un (t))T C [0, 1] is a
solution of problem (1.1)-(1.3). Let
M = ku k .

(2.14)

By the continuity of u (t) on the interval [0, 1], we can find a point tm [0, 1], such that
M = |ui (tm )|,
for some integer i , 1 i n. If M = 0, the estimate follows directly.
Otherwise, we consider the case ui (tm ) > 0 ( for the case ui (tm ) < 0, the proof is similar ). Then we know that
|uj (t)|
(i) If tm 6= pk ,

ui (tm ) = M,

t [0, 1],

1 j n.

(2.15)

k = 0, 1, . . . , K + 1, the inequality (2.15) yields

ui 0 (tm ) = 0.

(2.16)

From system (1.1), we obtain

ai,i (tm )ui (tm ) +

ai,j (tm )uj (tm ) = fi (tm )

(2.17)

1jn,j6=i

and the estimate follows immediately from (1.6).

(ii) If tm = pk , k = 0, 1, . . . , K + 1, suppose that i > 0 ( for the case i < 0 , the proof is similar ), If k = 0,
then tm = 0, and so, from the initial- final condition (1.3 ), we know that M = ui (tm ) = di , which yields the

5
estimate (2.13). Otherwise, if 0 < k K + 1, note that, on the interval [pk1 , pk ], the function ui (t) satisfies
system (1.1) at the the point t = pk , and so
X

ai,i (tm 0)ui (tm ) +

ai,j (tm 0)uj (tm ) fj (tm 0).

(2.18)

1jn,j6=i

This completes the proof of the estimate (2.13).

Using estimate (2.13), we can now show that Problem (1.1)-(1.3) has a unique solution. To see this we note
that the following homogeneous initial-final value problem
Eu0 (t) + A(t)u(t) =


u(pk + 0) u(pk 0) =

B u(0) + (I B ) u(1)

0,

t (pk , pk+1 ),

0,

k = 1, . . . , K,

k = 0, . . . , K

(2.19)
(2.20)

= 0,

(2.21)

has a unique solution u (t) = 0.

To prove the existence of a solution to problem (1.1)-(1.3), we introduce the following auxiliary initial value
problems
Eu0 (t) + A(t)u(t)
u(pk + 0) u(pk 0)
u(0)

= f (t),
=

0,

t (pk , pk+1 ),

k = 0, . . . , K

(2.22)

k = 1, . . . , K,

(2.23)

= 0

(2.24)

and
Eu0 (t) + A(t)u(t)
u(pk + 0) u(pk 0)
u(0)

= 0,

t (pk , pk+1 ),

k = 1, . . . , K,

0,

k = 0, . . . , K

(2.25)
(2.26)

= ej ,

(2.27)

for each integer j (1 j n).

By the existence theorem for the initial value problem for the linear ODE system, we know that for problem
(1)

(2.22)-(2.24) there exists a unique solution vf (t) = (v1,f

(t), . . . , vn,f
(t))T
C [0, 1] , and for each integer

1 j n, for problem (2.25)-(2.27) there exists a unique solution vj (t) = (vj,1

(t), . . . , vj,n
(t))T
The general solutions of system (1.1) are given by

u (t) =

cj vj (t) + vf (t),

(1)

C [0, 1] .

(2.28)

j=1,...,n

for arbitrary constants cj , (j = 1, . . . , n).

Let c = (c1 , . . . , cn )T and
V (t) = (v1 (t), . . . , vn (t)),

(2.29)

where V (t) is an n n matrix function. Then equality (2.28) can be rewritten as follows
u (t) = V (t)c + vf (t).

(2.30)

The vector-valued function u (t) given by (2.30) satisfies system (1.1) and the continuity condition (1.2) for
any constant vector c. Thus, if we can find a vector c Rn , such that u (t) satisfies the initial-final condition
(1.3), the existence of a solution to problem (1.1)-(1.3) has been proved. The initial-final condition (1.3) yields



B V (0) + (I B ) V (1) c =



d B vf (0) + (I B ) Vf (1) .

(2.31)

The uniqueness of the solution of problem (1.1)-(1.3) leads to





DetB V (0) + (I B ) V (1) 6= 0.

(2.32)

This implies that for system (2.31) there exists a unique solution c Rn . Then, from (2.30), we obtain a
solution of problem (1.1)-(1.3).
This completes the proof.

3. An approximation to problem (1.1)-(1.3)

In this section we construct an approximation to problem (1.1)-(1.3). The interval [0, 1] is divided into
subintervals by
0 = t0 < t1 < < tL = 1,

(3.1)

such that, for each pk [0, 1], we can find a tl such that pk = tl . Furthermore, let
hl = tl tl1 , l = 1, . . . , L;

h = max (hl ).

(3.2)

l=1,...,L

On each subinterval (tl1 , tl ) the functions ai,j (t), fi (t) are approximated by the constants ai,j (tl ), fi (tl ) with
tl1 tl tl . Then we introduce the matrix and vector
Al

fl


ai,j (tl )

(3.3)

nn
(f1 (tl ), f2 (tl ), . . . , fn (tl ))T .

(3.4)

We also define the approximating matrix and vector functions

Ah (t)
fh (t)

= Al ,
l

= f ,

t (tl1 , tl ),

(3.5)

t (tl1 , tl ).

(3.6)

It is easy to see that for these approximating functions, the following estimates hold
kfh k

kf fh k

kA Ah k

kf k ,

(3.7)
0

max

max

kf (t) fh (t)k kf k h,

(3.8)

max

max

kA(t) Ah (t)k ckA0 k h,

(3.9)

l=0,1,...,L tl1 ttl

l=0,1,...,L tl1 ttl

where c is a constant, independent of h.

Now consider the approximate problem
Euh 0 (t) + Ah (t)uh (t)

uh (tl + 0) uh (tl 0)
B uh (0) + (I B ) uh (1)

fh (t),

0,

= d.

t (tl1 , tl ),

l = 1, . . . , L 1,

l = 1, . . . , K,

(3.10)
(3.11)
(3.12)

7
By Theorem 2.1, we know that for problem (3.10)-(3.12) there exists a unique solution uh (t), and that the
following estimate holds

kf k
kuh k max kdk ,
.

(3.13)

We introduce the error function rh (t) = u (t) uh (t). Then rh (t) satisfies
Erh 0 (t) + Ah (t)rh (t)

dh (t),

rh (tl + 0) rh (tl 0)
B rh (0) + (I B ) rh (1)

0,

t (tl1 , tl ),

l = 1, . . . , K,

l = 1, . . . , L 1,

(3.14)
(3.15)

= 0,

(3.16)

with


dh (t) = f (t) fh (t) + Ah (t) A(t) uh (t).

(3.17)

Combining equality (3.17) and inequalities (3.7)-(3.9),(3.13), we arrive at

n
kf k
o
kdh k kf 0 k + ckA0 k kdk +
h,

(3.18)

where c is a constant independent of h and .

Applying Theorem 2.1 to problem (3.14)-(3.16), we see that the following error bound holds
Theorem 3.1. The error rh satisfies the estimate
krh k

1n 0
kf k
o
kf k + ckA0 k kdk +
h,

(3.19)

where c is a constant independent of h and E.

This shows that the approximate solution uh (t) converges E-uniformly to the solution u (t) of problem (1.1)(1.3) on the interval [0, 1].

4. A tailored finite point method for problem (1.1)(1.3)

Using uh (t) for the solution of Problem (3.9)(3.11), we now introduce the tailored finite point method for
problem (1.1)(1.3). Let
uh (tl ) = ul ,

l = 0, 1, . . . , L.

(4.1)

Then on each subinterval [tl1 , tl ],the approximate solution uh (t) satisfies

Euh 0 (t) + Al uh (t)
B uh (tl1 )

+ (I B

) uh (tl )

= f l,

t (tl1 , tl ),

= B ul1 + (I B ) ul .

(4.2)
(4.3)

It is easy to see that the ODE system (4.2) with constant coefficients has the particular solution
vfl = Al

1

f l.

We now construct the general solution vh of the homogeneous ODE system corresponding to (4.2)

(4.4)

Evh 0 (t) + Al vh (t)

0,

t (tl1 , tl ).

(4.5)

Let
vh (t) = et

(4.6)

be the solution of system (4.5). Then the vector and number are the solutions of the eigenvalue problem
Al = E.

(4.7)

Solving the eigenvalue problem (4.7), we obtain the eigenvalues l1 , l2 , . . . , l2r , l2r+1 , ln , corresponding to
l
l
, 2r+1
, . . . , nl . Assume that the eigenvalues l2r+1 , . . . , ln and the corresponding
the eigenvectors 1l , 2l , . . . , 2r
eigenvectors are real. The remaining eigenvalues occur in complex conjugate pairs, so that
l2j

= l2j1 ,

j = 1, . . . , r,

(4.8)

l
2j

l
2j1
,

j = 1, . . . , r,

(4.9)

These conjugate pairs of eigenvalues l2j1 , l2j and

l
2j1

l
2j
l

2j1

l
2j

l
l
eigenvectors 2j1
, 2j
, can be written in the form

= jl ijl ,
= jl + ijl ,
= jl ijl ,
= jl + ijl .

Then, for each 1 j r, there are two independent solutions of the system (4.5) of the form
l
v2j1
(t)

l
v2j
(t)

l
jl cos(jl t) jl sin(jl t) ej (ttl ) ,

l
jl cos(jl t) + jl sin(jl t) ej (ttl ) .

(4.10)
(4.11)

Furthermore, for each j (2r < j n), there is one solution of the system (4.5) given by
vjl (t)

jl ej (ttl ) .

(4.12)

Let



V l (t) v1l (t), v2l (t), . . . , vnl (t) ,

(4.13)

where V l (t) is a n n matrix function. On each interval [tl1 , tl ], uh (t) is given by

uh (t) = V l (t)cl + vfl ,

(4.14)

with cl = (cl1 , . . . , cln )T , which is determined by the initial-final condition (4.3). Then we have
cl

1 





B V l (tl1 ) + I B V l (tl )
B ul1 + I B ul vfl




Dl B ul1 + I B ul vfl ,

(4.15)

9
with

1


Dl = B V l (tl1 ) + I B V l (tl )
.
Furthermore, on each interval [tl1 , tl ], the expression (4.14) can be written as




uh (t) = V l (t)Dl B ul1 + I B ul vfl + vfl .

(4.16)

The continuity condition (3.12) yields

I B ul1
B ul

o





n
I B V l (tl1 )Dl B ul1 + I B ul vfl + vfl ,
o


n


= B V l (tl )Dl B ul1 + I B ul vfl + vfl .

(4.17)
(4.18)

We then obtain the following tailored finite point scheme

ul

=
+

u0

uL



o

n


I B V l+1 (tl )Dl+1 B ul + I B ul+1 vfl+1 + vfl+1
n


o


B V l (tl )Dl B ul1 + I B ul vfl + vfl , l = 1, 2, . . . , L 1,


o

n


B d + I B V 1 (t0 )D1 B d + I B u1 vf1 + vf1 ,
n


o




I B d + B V L (tL )DL B uL1 + I B d vfL + vfL .

(4.19)

(4.20)
(4.21)

In the special case when the {j , j = 1. . . . , n} are all positive, we have B = I and the method (4.19)-(4.21),
for finding the numerical solution of problem (1.10)-(1.12), is reduced to the following method

ul

V l (tl )Dl ul1 vfl + vfl , l = 1, 2, . . . , L,

(4.22)

u0

d.

(4.23)

This is a one step explicit scheme, which is unconditionally stable.

Since the discrete scheme (4.19)-(4.21) is equivalent to the approximate problem (3.9)-(3.11), we attain the
following theorem;
(1)

(1)

Theorem 4.1. Suppose that A(t) C, [0, 1], f (t) C, [0, 1] and that A(t), f (t) satisfy the conditions (1.3)(1.9). Then there exists a unique solution {ul l = 0, 1, . . . , L} to problem (4.19)-(4.21) and the following
parameter-uniform error estimate holds

max

l=0,1,...,L

ku(tl ) ul k

1n 0
kf k
o
kf k + ckA0 k kdk +
h.

(4.24)

This shows that the tailored finite point method yields approximate solutions that are parameter-uniformly
convergent with respect to {0 < |j | 1, j = 1. . . . , n} to the solution of the multi-scale singularly perturbed
problem (1.1)-(1.3).
Solving the algebraic equation (4.19)-(4.21) gives the numerical solution {ul , l = 0, 1, . . . , L}. Combining
this with the expressions (4.15), (4.14) then yields the required continuous approximate solution uh (t).

10

5. Numerical examples
In this section three problems are solved numerically using a tailored finite point method. The first is an
initial value problem and the second an initial-final value problem for a singularly perturbed system of ordinary
differential equations. The third problem arises from a semi-discretization of a forward-backward parabolic
problem. The errors in the numerical solutions for the first two problems, as well as numerical estimates of the
parameter-uniform rates of convergence and the parameter-uniform error constants, are given in five tables. In
Tables 1, 3, 5 the errors of the numerical solutions are first calculated by comparing the numerical solutions with
the exact solutions that are obtained with the finest mesh t = 1/4096. On the other hand, in Tables 2, 4
parameter-uniform estimates of the convergence rates and error constants are computed using the methodology
described in [4]. Graphs of the solutions of all three problems are presented in the figures.
Example 1. For t (0, 1), we solve
1 u01 (t) + 4u1 (t) u2 (t) u3 (t) = t
2 u02 (t) u1 (t) + 4u2 (t) u3 (t) = 1
3 u03 u1 (t) u2 (t) + 4u3 (t) = 1 + t2
u1 (0) = 0,

u2 (0) = 0,

u3 (0) = 0.

for different values of 1 , 2 and 3 . This example appears in [23] and we compare our tailored finite point
method with the classical finite difference discretization using a Shishkin piecewise uniform mesh. Both methods
are seen to be first order parameter-uniform. Similarly to [23], let 1 = r/16, 2 = r/4, 3 = r. The numerical
solutions of u1 (t) for different r are displayed in Figure 5.1.

0.7

0.6

0.5

u1

0.4

0.3

0.2
r=1/128
r=1/32
r=1/8
r=1/2

0.1

0
0

0.1

0.2

0.3

0.4

0.5
t

0.6

0.7

0.8

0.9

Fig. 5.1. Example 1. u1 (t) for r = 1/2, 1/8, 1/32, 1/128.

The L norm of the numerical errors for different r and time step are given in Table 5.1, where we have used
= tl1 in the formulas (3.3) and (3.4). Parameter-uniform first order convergence for all r can be observed
in this table,
The parameter-uniform error parameters p and Cp are given by the well-established two-mesh procedure
for numerically finding a parameter-uniform error bound of the form

tl

kUt u k Cp tp ,

11
r \ t

1/128

1/256

1/512

1/1024

1/2048

21

2.311 103

1.115 103

5.194 104

2.224 104

7.412 105

22

2.507 103

1.205 103

5.605 104

2.398 104

7.987 105

23

2.638 103

1.258 103

5.829 104

2.489 104

8.283 105

24

2.768 103

1.300 103

5.979 104

2.544 104

8.449 105

25

2.967 103

1.354 103

6.132 104

2.589 104

8.568 105

26

3.315 103

1.448 103

6.373 104

2.649 104

8.700 105

27

3.877 103

1.619 103

6.820 104

2.757 104

8.917 105

210

5.139 103

2.418 103

1.057 103

4.010 104

1.174 104

215

5.280 103

2.559 103

1.195 103

5.124 104

1.708 104

216

5.280 103

2.559 103

1.195 103

5.124 104

1.708 104

217

5.280 103

2.559 103

1.195 103

5.124 104

1.708 104

Table 5.1: Example 1. L norm of the numerical errors, for different r and time steps, when tl = tl1 in (3.3) and (3.4),
1 = r/16, 2 = r/4, 3 = r

which is described in [4]. The technique uses the two-mesh method, which involves the quantities
Dt = kUt Ut/2 k ,
pt = log2

Dt
,
Dt/2

Dt = max{Dt },


p = min pt
t

and
Cpt
=

Dt
,
t (1 2p )
p

Cp = max Cpt

For this example the results are given in Table 2, where it is seen that p = 0.996 and Cp = 0.685.
If we use
Al

fl

 Z tl

1
ai,j (t) dt
,
tl tl1 tl1
nn
Z tl
Z tl
Z tl
1
f1 (t) dt,
f2 (t) dt, . . . ,
fn (t) dt)T ,
(
tl tl1 tl1
tl1
tl1

(5.1)
(5.2)

instead of tl = tl1 , the corresponding errors are displayed in Table 5.3. It can be seen that parameteruniform first order convergence is achieved. However, second order convergence occurs when r is large, while
when r decreases, only first order convergence is attained. This is different from the one dimensional neutron
transport equation, where with cell averaging of the coefficients, parameter-uniform second order convergence
is obtained [17]. The second order parameter-uniform convergence for the neutron transport equation is due to
the specific scales of its coefficients. For the general form of problem (1.1)-(1.3), we can achieve only first order
parameter-uniform convergence by using piecewise constant approximations of the coefficients. The computed
parameter-uniform error parameters using the two-mesh method are given in Table 3.

12
r \ t

1/128

1/256

1/512

1/1024

1/2048

21

1.196 103

5.953 104

2.970 104

1.483 104

7.412 105

22

1.302 103

6.446 104

3.207 104

1.599 104

7.987 105

23

1.380 103

6.751 104

3.340 104

1.6610 104

8.283 105

24

1.468 103

7.024 104

3.435 104

1.699 104

8.449 105

25

1.612 103

7.411 104

3.543 104

1.733 104

8.568 105

26

1.867 103

8.104 104

3.723 104

1.779 104

8.700 105

27

2.258 103

9.366 104

4.063 104

1.866 104

8.917 105

210

2.721 103

1.361 103

6.556 104

2.836 104

1.174 104

215

2.721 103

1.364 103

6.827 104

3.416 104

1.708 104

216

2.721 103

1.364 103

6.827 104

3.416 104

1.708 104

217

2.721 103

1.364 103

6.827 104

3.416 104

1.708 104

Dt

2.721 103

1.364 103

6.827 104

3.416 104

1.708 104

pt

0.996

0.998

0.999

1.000

p = 0.996

Cpt

0.685

0.685

0.684

0.682

0.680

Table 5.2: Example 1. Values of Dt , Dt , pt and Cpt

with 1 = r/16, 2 = r/4, 3 = r for various r and t. We

find p = mint pt = 0.996 and Cp = maxt Cpt

= 0.685. Here tl = tl1 in (3.3) and (3.4),

Example 2. For a negative i , a final value must be imposed in order to preserve the maximum principle.
In this example, we solve a system with discontinuous coefficients such that for t [0, 0.5],
1 u01 (t) + (5 + et )u1 (t) tu2 (t) u3 (t) u4 (t) = t,
2 u02 (t) u1 (t) + (4 + t2 )u2 (t) u3 (t) u4 (t) = 1,
3 u03 u1 (t) u2 (t) + 5u3 (t) (1 + t)u4 (t) = 1 + t,
3 u03 u1 (t) tu2 (t) u3 (t) + 5u4 (t) = 1 t2 ,
and for t [(0.5, 1],
1 u01 (t) + (4 + et )u1 (t) tu2 (t) u3 (t) u4 (t) = 1,
2 u02 (t) u1 (t) + (4 + t2 )u2 (t) u3 (t) u4 (t) = 1 t,
3 u03 u1 (t) u2 (t) + (5 + t2 )u3 (t) (2 + t)u4 (t) = 1 t2 ,
3 u03 tu1 (t) (1 + t)u2 (t) u3 (t) + (4 + et )u4 (t) = 1 + t,
with the boundary conditions
u1 (1) = 0,

u2 (0) = 0,

u3 (1) = 0,

u4 (0) = 0.

This system exhibits initial, final and interface layers, when the |i | are small. Similarly to [23], let 1 = r/64,
2 = r/16, 3 = r/4, 4 = r. The numerical solutions of u1 (t) for different r are shown in Figure 5.2. The
results are consistent with first order parameter-uniform convergence, as expected from the theory.
In Table 5.4, we present the L norm of the numerical errors for different r and time steps, where we have
used tl = tl1 in (3.3) and (3.4). Uniform first order convergence can be observed.

13
r \ t

1/128

1/256

1/512

1/1024

1/2048

order

21

1.590 104

4.011 105

9.941 106

2.369 106

4.738 107

2.09

22

3.047 104

7.950 105

1.988 105

4.748 106

9.504 107

2.07

23

5.261 104

1.521 104

3.934 105

9.480 106

1.902 106

2.02

24

7.520 104

2.624 104

7.517 105

1.873 105

3.792 106

1.91

25

9.316 104

3.744 104

1.293 104

3.570 105

7.473 106

1.73

26

1.116 103

4.623 104

1.835 104

6.095 105

1.412 105

1.55

27

1.352 103

5.512 104

2.242 104

8.470 105

2.342 105

1.44

210

2.502 103

1.139 103

4.593 104

1.449 104

4.091 105

1.48

215

2.643 103

1.280 103

5.978 104

2.563 104

8.543 105

1.22

16

2.563 10

8.543 10

1.22

2.563 104

8.543 105

1.22

2.643 10

217

2.643 103

1.280 10

1.280 103

5.978 10

5.978 104

Table 5.3: Example 1. L norm of the numerical errors for different r and time steps, when Al and f l are as in (5.1)
and (5.2), and 1 = r/16, 2 = r/4, 3 = r. The last column displays the numerical convergence order by fitting the
log-log plot of the errors.

To look at the uniform convergence order, we present in Table 5.5 the values of Dt , Dt , pt and Cpt
.
Example 3. In this example, we consider the following semi-discretization of the forward-backward parabolic
problem
sign(xi )|xi |p

ui+1 (t) 2ui (t) + ui1 (t)

ui (t) + qi (t),
x2
i = 1, 2, , N 1,

ui (t)
t

u0 (t)

ui (1)

= (x), 1 < xi < 0.

f (t),

uN (t) = g(t),

0 < t < T,

0 < t < T,

(5.3)
(5.4)

ui (0) = s(xi ), 0 < xi < 2,

(5.5)

with p > 0 and

xi = 1 + 3 i/N 6= 0,

i = 1, 2, , N 1.

In Figure 5.3 we show the numerical results for

T = 0.5,

qi (t) = 1,

(x) = sin(x),

s(x) = cos2 ( x),

2

f (t) = 0,

g(t) = 1.

Initial and final layers in time occur for both p = 1 and p = 10. The layers become more significant and obvious
when p is large. We can capture the layers without resolving them (using a lot of nodes in the layer).

6. Conclusion
A tailored finite point method for a multi-scale singularly perturbed system of linear ordinary differential
equations is proposed in this paper. We can give either initial or final values for the ODE system as well as use
discontinuous coefficients.
The tailored finite point method yields approximate solutions that converge in the maximum norm, uniformly
with respect to the singular perturbation parameters. We prove a parameter-uniform error estimate in the

14

0.8
r=1/64
r=1/16
r=1/4
r=1

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Fig. 5.2. Example 2. u1 (t) for r = 1, 1/4, 1/16, 1/64.

maximum norm and verify our analytical results numerically. To show the performance of our proposed scheme,
three numerical examples that exhibit initial and final layers are considered.
In particular, the third example shows the existence of initial and final layers of the solution of the forwardbackward parabolic equation. To investigate and understand these layers will be the subject of our future
work.
Acknowledgments. H. Han was supported by the NSFC Project No. 10971116. M. Tang is supported by
Natural Science Foundation of Shanghai under Grant No. 12ZR1445400.

References
[1] A. E. Berger, H. Han and R.B. Kellogg, A prior estimate and analysis of a numerical method for a turning point
problem, Math. Comp. 42, 465-492, 1984.
[2] A. K. Aziz, D. A. French, S. Jensen and R. B. Kellogg, Origins, analysis, numerical analysis, and numerical
approximation of a forward-backward parabolic problem, Modelisation mathmatique et analyse numerique, 33, 5,
895-922, 1999.
[3] E.P. Doolan, J. J. H. Miller and W. H. A. Schilders Uniform numerical methods for problems with initial and
boundary layers, Boole Press, Dublin, 1980.
[4] P.A. Farrell, A. Hegarty, J. J. H. Miller, E. ORiordan, G. I. Shishkin, Robust Computational Techniques for
Boundary Layers, Applied Mathematics & Mathematical Computation (Eds. R. J. Knops & K. W. Morton),
Chapman & Hall/CRC Press. 2000.
[5] H. Han and Z. Huang, A tailored finite point method for the Helmholtz equation with high wave numbers in heterogeneous medium, J. Comp. Math, 26, 728-739, 2008.
[6] H. Han and Z. Huang, Tailored finite point method for a singular perturbation problem with variable coefficients in
two dimensions, J. Sci. Comp., 41, 200-220, 2009.
[7] H. Han and Z. Huang, Tailored finite point method for steady-state reaction-diffusion equation, Comm. Math. Sci.,
8, 887-899, 2010.
[8] H. Han and Z. Huang, An equation decomposition method for the numerical solution of a fourth-order elliptic
singular perturbation problem, Numerical Methods for P. D. E., 28, 942-953, 2012.
[9] H. Han, Z. Huang and B. Kellogg, A tailored finite point method for a singular perturbation problem on an unbounded
domain, J. Sci. Comp. 36, 243-261, 2008.
[10] H. Han, Z. Zhang, Multiscale tailored finite point method for second order elliptic equations with rough or highly
osillatory coefficients, 10, 945-976, 2012.

15
r \ t

1/128

1/256

1/512

1/1024

1/2048

1.418 103

6.920 104

3.242 104

1.392 104

4.645 105

21

1.627 103

7.813 104

3.633 104

1.554 104

5.176 105

22

1.933 103

9.222 104

4.273 104

1.824 104

6.070 105

23

2.192 103

1.036 103

4.775 104

2.033 104

6.756 105

24

2.400 103

1.118 103

5.106 104

2.163 104

7.165 105

25

2.627 103

1.194 103

5.372 104

2.254 104

7.428 105

26

2.960 103

1.293 103

5.671 104

2.345 104

7.658 105

27

3.468 103

1.451 103

6.120 104

2.469 104

7.949 105

210

4.609 103

2.166 103

9.453 104

3.592 104

1.056 104

215

4.736 103

2.293 103

1.070 103

4.587 104

1.529 104

216

4.736 103

2.293 103

1.070 103

4.587 104

1.529 104

217

4.736 103

2.293 104

1.070 103

4.587 104

1.529 104

Table 5.4: Example 2. L norm of the numerical errors for different r and time steps, when tl = tl1 in (3.3) and (3.4),
1 = r/64, 2 = r/16, 3 = r/4, 4 = r.
[11] P. W. Hemker, A singularly perturbed modal problem for numerical computation, J. Comp. Appl. Math, 76, 277-285,
1996.
[12] M. Hochbruck , C. Lubich, On Magnus Integrators for Time-Dependent Schrodinger Equations, SIAM J. Numerical
Analysis, 945-963, 2003.
[13] Z. Huang and X. Yang, Tailored Finite Point Method for First Order Wave Equation, Journal of Scientific Computing
49, 351-366, 2011.
[14] H. Han and D. S. Yin, A non-overlap domain decomposition method for the forward-backward heat equation, Journal
of Computational and Applied Mathematics, 159, 35-44.
[15] Z. Huang, 2009, Tailored finite point method for the interface problem, Networks and Heterogeneous Media, 4,
91-106, 2003.
[16] A. M. Ilin, Difference scheme for a differential equation with a small parameter affecting the highest derivatve,
Math. Notes, 6, 596-602, 1969.
[17] S. Jin, M. Tang and H. Han, A uniformly second order numerical method for the one-dimensional discrete transport
equation and its diffusion limit with interface, Networks and Heterogeneous Media, 4, 35-65, 2009.
[18] K. Lorenz, T. Jahnke and C. Lubich, Adiabatic integrators for highly oscillatory second-order linear differential
equations with time-varying eigendecomposition, BIT Numerical Mathematics, 45, 91-115, 2005.
[19] C.B. Moler and C.F. Van Loan, Nineteen dubious ways to compute the exponential of a matrix, twenty-five years
later, SIAM Rev., 45:1, pp. 349, 2003.
[20] J.J.H. Miller, On the convergence ,unifomly in , of difference schemes for a two-point boundary singular perturbation problem, Numerical analysis of singular perturbation problems, eds, by P.W. Hemker and J.J.H. Miller,
Academic Press, 467-474, 1979.
[21] Y. Shih, R.B. Kellogg and Y.Y. Chang, Characteristic tailored finite point method for convection-dominated
convection-diffusion-reaction problems, Journal of Scientific Computing, 47, 351-366, 2011.
[22] M. Tokman, Efficient integration of large stiff systems of ODEs with exponential propagation iterative (EPI) methods,
Journal of Computational Physics, 213, 2, 748-776, 2006.
[23] S. Valarmathi and John J.H. Miller, A parameter-uniform finite difference method for a singularly perturbed linear
dynamical system, International Journal of Numerical, Analysis and Modeling, 7, 535-548, 2010.

16

r \ t

1/128

1/256

1/512

1/1024

1/2048

7.503 104

3.678 104

1.850 104

9.277 105

4.645 105

21

8.452 104

4.180 104

2.079 104

1.037 104

5.176 105

22

1.011 103

4.950 104

2.448 104

1.217 104

6.070 105

23

1.207 103

5.588 104

2.742 104

1.358 104

6.756 105

24

1.286 103

6.140 104

2.944 104

1.446 104

7.165 105

25

1.434 103

6.569 104

3.118 104

1.511 104

7.428 105

26

1.668 103

7.256 104

3.326 104

1.578 104

7.658 105

27

2.017 103

8.394 104

3.652 104

1.674 104

7.949 105

210

2.446 103

1.221 103

5.861 104

2.536 104

1.056 104

215

2.446 103

1.223 103

6.116 104

3.058 104

1.529 104

216

2.446 103

1.223 103

6.116 104

3.058 104

1.529 104

217

2.446 103

1.223 103

6.116 104

3.058 104

1.529 104

Dt

2.446 103

1.223 103

6.116 104

3.058 104

1.529 104

pt

1.000

1.000

1.000

1.000

p = 1.000

Cpt

0.626

0.626

0.626

0.626

0.626

Table 5.5: Example 2. Values of D , D, p, p, Cp and Cp with 1 = r/64, 2 = r/16, 3 = r/4, 4 = r, for various r and

t. We find p = mint pt = 1.000 and Cp = maxt Cpt

= 0.626. Here tl = tl1 in (3.3) and (3.4),

17

0.9

0.8

0.7

0.8

0.6

0.6
u

0.5
0.4
0.4
0.2
0.3
0
0.5
2
0.4

0.2

1.5
1

0.3
0.5

0.2

0.1

0
0.1

0.5
0

0.05

0.1

0.15

0.2

0.25
t

0.3

0.35

0.4

0.45

0.05

0.1

0.15

0.2

0.25
t

0.3

0.35

0.4

0.45

a)
0.9

0.8
1
0.7
0.8

0.6

0.6
u

0.5

0.4

0.4

0.2

0.3

0.2

0
0.5

2
1.5

0.4
1

0.3

0.1

0.5
0.2

0
0.1

0.5
0

1
x

0.5

b)
Fig. 5.3. Example 3. The numerical solution of the forward-backward parabolic problem. Left: ui (t) for i = 1, 2, , N ;
Right: uM where M satisfies xM < 0 and xM +1 > 0. a) p = 1; b) p = 10. Here x = 1/40, t = 1/256.