Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

A High-Order Compact Boundary Value Method

for Solving One-Dimensional Heat Equations


Haiwei Sun, Jun Zhang
Laboratory for High-Performance Scientific Computing and Computer Simulation,
Department of Computer Science, University of Kentucky, 773 Anderson Hall,
Lexington, Kentucky 40506-0046

Received 22 April 2002; accepted 23 April 2003

DOI 10.1002/num.10076

We combine fourth-order boundary value methods (BVMs) for discretizing the temporal variable with
fourth-order compact difference scheme for discretizing the spatial variable to solve one-dimensional heat
equations. This class of new compact difference schemes achieve fourth-order accuracy in both temporal
and spatial variables and are unconditionally stable due to the favorable stability property of BVMs.
Numerical results are presented to demonstrate the accuracy and efficiency of the new compact difference
scheme, compared to the standard second-order Crank-Nicolson scheme. © 2003 Wiley Periodicals, Inc.
Numer Methods Partial Differential Eq 19: 846 – 857, 2003

Keywords: heat equation; compact difference scheme; BVMs; Crank-Nicolson scheme; unconditional
stability

I. INTRODUCTION

We consider a class of high-accuracy finite difference discretization schemes for solving the
one-dimensional heat equations of the form

⭸u ⭸ 2u
⫽ ␣2 2 , 0 ⬍ x ⬍ l, t ⬎ 0, (1)
⭸t ⭸x

subject to the initial condition:

Correspondence to: Jun Zhang (e-mail: jzhang@cs.uky.edu)


Contract grant sponsor: U.S. National Science Foundation; contract grant number: CCR-9988165 (to H.S.)
Contract grant sponsor: U.S. National Science Foundation; contract grant numbers: CCR-9902022, CCR-9988165,
CCR-0092532 (to J.Z.)
Contract grant sponsor: U.S. Department of Energy; contract grant number: DE-FG02-02ER45961 (to J.Z.)
Contract grant sponsor: Japanese Research Organization for Information Science and Technology (RIST) (to J.Z.)
Contract grant sponsor: University of Kentucky Research Committee (to J.Z.)
© 2003 Wiley Periodicals, Inc.
A HIGH-ORDER COMPACT BOUNDARY VALUE METHOD 847

u共x, 0兲 ⫽ f 共x兲, 0 ⱕ x ⱕ l,

and the Dirichlet boundary conditions:

u共0, t兲 ⫽ a共t兲, u共l, t兲 ⫽ b共t兲, t ⬎ 0.

We assume that the functions f(x), a(t), and b(t) are sufficiently smooth and their required
high-order derivatives exist.
For a positive integer n, we denote hx ⫽ l/n, the step size of the spatial variable x, and ht, the
step size of the temporal variable t. We consider a rectangular grid (xi, tj), xi ⫽ ihx, tj ⫽ jht,
where i ⫽ 0, . . . , n and j ⫽ 1, . . . , N. Here N is the number of time steps and the final time
is T ⫽ Nht.
For the numerical integration of (1), a computational scheme that has enjoyed great popu-
larity is that of Crank and Nicolson [1]. The Crank-Nicolson (CN) scheme is unconditionally
stable. The CN scheme is based on the classical trapezoidal formula for integration in time and
the central difference formula for spatial discretization. It is well known that the accuracy of the
CN scheme is of order O(h2x ⫹ h2t ). To improve the spatial discretization accuracy of the CN
scheme, Douglas [2, p. 44] gives a discretization scheme that is based on the well known
Numerov discretization of the spatial variable. Its accuracy is of order O(h4x ⫹ h2t ). Recently,
Chawia et al. [3] propose to use a one-parameter family of L-stable Simpson rules to obtain a
family of finite difference schemes that are unconditionally stable and are of fourth-order
accuracy in time. That means the accuracy of their schemes is formally of order O(h4x ⫹ h4t ).
However, their schemes are parameter dependent, and the authors do not give a rule to choose
a suitable parameter in their method. Further, the numerical results reported in their article do
not show that the accuracy in time domain reaches O(h4t ).
Boundary value methods (BVMs) are unconditionally stable and are high-accuracy schemes
for solving initial value problems (IVPs) based on the linear multistep formulas [4 – 8]. Unlike
Runge-Kutta or other initial value methods (IVM), BVMs achieve the advantage of both good
stability and high-order accuracy [4 – 6, 9]. Furthermore, fourth-order compact difference
schemes for solving steady-state partial differential equations are well studied [10 –13]. These
fourth-order compact difference schemes have been shown to be highly accurate and compu-
tationally efficient. Recently, the fourth-order compact methodology has been extended to the
time-dependent problems using Crank-Nicolson-like schemes [14, 15]. These approaches result
in some compact difference schemes that are of order O(h4x ⫹ h2t ). Because the BVM schemes
are easily used in arbitrary block cases, we propose to combine BVMs with a fourth-order
compact difference scheme to solve the one-dimensional heat equations. This class of high-order
compact BVMs achieve high-order accuracy in both temporal and spatial variables and need not
choose any parameters.
This article is arranged as follows. In §2, we introduce BVMs for ordinary differential
equations. The fourth-order compact difference schemes are discussed in §3. In §4, we combine
BVMs with fourth-order compact difference schemes and derive a class of new finite difference
schemes with fourth-order accuracy for both temporal and spatial variables. In §5, we give some
numerical results to demonstrate the accuracy and efficiency of the new compact difference
scheme and compare it with the standard Crank-Nicolson scheme. Some concluding remarks are
given in §6.
848 SUN AND ZHANG

II. BOUNDARY VALUE METHODS

We first give a brief introduction to the BVMs for solving a general differential equation of the
form

dy共t兲
⫽ g共t, y兲, y共0兲 ⫽ y 0 , t ⱖ 0. (2)
dt

We will consider higher-order linear multistep methods (LMMs). For more detailed discussions
on BVMs, a reader is urged to consult [4].
Let yj ⫽ y(tj) ⫽ y( jht) and gj ⫽ g( yj, tj), j ⫽ 0, 1, . . . , N, the k-step formula for approximating
the Equation (2) can be written as

冘 ␣y 冘 ␤g
k k

i i⫹j ⫽ h t i i⫹j , j ⫽ 0, . . . , N ⫺ k, (3)


i⫽0 i⫽0

where ␣i and ␤i will be defined by different BVMs. The LMMs in (3) must be used with ␥ initial
conditions, and (k ⫺ ␥) final conditions. That is, we need the values y0, . . . , y␥⫺1 and
yN⫺k⫹␥⫹1, . . . , yN. The initial condition in (2) only provides us with one value ( y0). In order to
obtain the remaining initial and final values, we have to provide additional (k ⫺ 1) equations:

冘␣ 冘␤
k k
共 j兲 共 j兲
i y ⫽ ht
i i ig, j ⫽ 1, . . . , ␥ ⫺ 1, (4)
i⫽0 i⫽0

and

冘␣ 冘␤
k k
共 j兲 共 j兲
i y N⫺i ⫽ ht i g
N⫺i , j ⫽ N ⫺ k ⫹ ␥ ⫹ 1, . . . , N. (5)
i⫽0 i⫽0

The coefficients ␣( j) and ␤( j) in (4) and (5) should be chosen such that the truncation errors for
these initial and final conditions are of the same order as that in (3) (see [4, p. 132]). The
advantage of using BVMs is that they have much better stability properties than traditional
initial value methods do (see [4, p. 79]).
BVMs are numerical methods based on the LMMs for solving ordinary differential equa-
tions. By applying the LMMs, the solution to (2) is given by the solution of a system of linear
equations,

My ⫽ g,

where y ⫽ ( y0, y1, . . . , yN)T and g ⫽ ( g0, g1, . . . , gN)T, the coefficient matrix M depends on
which type of LMM being used.
Although IVMs (where the system of equations can be solved easily by forward recursions)
are computationally more efficient than BVMs, the major advantage of using BVMs over IVMs
comes from their stability properties, because BVMs are unconditionally stable [4 – 8].
A HIGH-ORDER COMPACT BOUNDARY VALUE METHOD 849

In the following, we choose a particular fourth-order BVM to generate a finite difference


scheme. For the detailed procedure, please see [4, p. 169].
A fourth-order BVM approximation of (2) can be written as (with k ⫽ 3 and ␥ ⫽ 2)

1 ht
共y j⫹3 ⫹ 9y j⫹2 ⫺ 9y j⫹1 ⫺ y j 兲 ⫽ 共g j⫹2 ⫹ g j⫹1 兲 ⫹ O共h 4t 兲, j ⫽ 0, . . . , N ⫺ 3.
12 2

The additional equations associated with the initial and final conditions are

1 ht
共⫺y3 ⫹ 9y2 ⫹ 9y1 ⫺ 17y0兲 ⫽ 共3g1 ⫹ g0兲 ⫹ O共h4t 兲,
24 4

and

1 ht
共y N⫺3 ⫺ 9y N⫺2 ⫺ 9y N⫺1 ⫹ 17y N 兲 ⫽ 共g N ⫹ 3g N⫺1 兲 ⫹ O共h 4t 兲.
24 4

For convenience, we rewrite the above equations in a matrix form as

A t y ⫽ Btg ⫹ e1y0 ⫹ O共h4t 兲,

where

冤 冥
1 0 0 0
⫺17/24 9/24 9/24 ⫺1/24
1/12 9/12 ⫺9/12 ⫺1/12
At ⫽ ·· ·· ·· ·· , (6)
· · · ·
1/12 9/12 ⫺9/12 ⫺1/12
1/24 ⫺9/24 ⫺9/24 17/24

and

冤 冥
0 0
1/4 3/4
1/2 1/2
B t ⫽ h tⴱ ·· ·· , (7)
· ·
1/2 1/2
3/4 1/4

and e1 ⫽ [1, 0, . . . , 0]T. We emphasize that the formula (3) can be applied to block cases (see
[4, p. 279]).
Now we consider a system of linear ordinary differential equations:

dy共t兲
⫽ Bxy共t兲 ⫹ g共t兲, y共0兲 ⫽ y0, t ⱖ 0, (8)
dt
850 SUN AND ZHANG

where Bx is an n-by-n matrix, y(t) ⫽ [ y1(t), . . . , yn(t)]T and g(t) ⫽ [ g1(t), . . . , gn(t)]T are n-by-1
vector functions, and y0 is an n-by-1 vector. To simplify our discussions on the block cases, we
first give a definition of Kronecker product of two matrices.
Let A ⫽ (aij)m⫻n and B be arbitrary matrices; then the matrix

冤 冥
a 11 B a 12 B · · · a 1n B
a 21 B a 22 B · · · a 2n B
A 丢 B⫽ · · ·· ·
· · · ·
· · ·
a m1 B a m2 B · · · a mn B

is called the Kronecker product of A and B [4, p. 354].

Theorem 1. Let A and B be two nonsingular matrices, and C and D be arbitrary matrices of
appropriate dimension. The Kronecker product posses the following properties [4, p. 354 –356]:

1. ( A R B)(C R D) ⫽ ( AC) R (BD);


2. ( A R B)⫺1 ⫽ A⫺1 R B ⫺1;
3. There exists a permutation matrix Q, such that Q( A R B)QT ⫽ B R A.

Using the notation of Kronecker product, the fourth-order BVM of (8) can be written as [4,
7] follows:

共A t 丢 I n ⫺ B t 丢 B x 兲y ⫽ e1 丢 y0 ⫹ 共Bt 丢 In兲g, (9)

where In is the n-by-n identity matrix, and

y ⫽ 关y1共t0兲, y2共t0兲, . . . , yn共t0兲, y1共t1兲, . . . , yn共t1兲, . . . , y1共tN兲, . . . , yn共tN兲兴T


g ⫽ 关g1共t0兲, g2共t0兲, . . . , gn共t0兲, g1共t1兲, . . . , gn共t1兲, . . . , g1共tN兲, . . . , gn共tN兲兴T.

Remark. If a linear ordinary differential equation is of the form:

dy共t兲
Ax ⫽ Bxy共t兲 ⫹ g共t兲, y共0兲 ⫽ y0, t ⱖ 0,
dt

where Ax is an n-by-n nonsingular matrix, we can rewrite it in the form of (8)

dy共t兲
⫽ Ax⫺1Bxy共t兲 ⫹ Ax⫺1g共t兲, y共0兲 ⫽ y0, t ⱖ 0.
dt

According to (9), its fourth-order BVM can be written as

共A t 丢 I n ⫺ B t 丢 共A x⫺1 B x 兲兲y ⫽ e1 丢 共A⫺1 ⫺1


x y0兲 ⫹ 共Bt 丢 In兲共Ax g兲.

By multiplying both sides with IN⫹1 R Ax, from Theorem 1, we have

共A t 丢 A x ⫺ B t 丢 B x 兲y ⫽ e1 丢 y0 ⫹ 共Bt 丢 In兲g. (10)


A HIGH-ORDER COMPACT BOUNDARY VALUE METHOD 851

III. A FOURTH-ORDER COMPACT DIFFERENCE SCHEME

We now discuss a fourth-order compact difference scheme for a second-order differential


equation,

d 2 z共x兲
⫽ p共x兲, 0 ⬍ x ⬍ l, (11)
dx 2

with the two-point boundary condition:

z共0兲 ⫽ z 0 , z共l兲 ⫽ z l .

We can design a fourth-order accuracy compact difference scheme for such partial differential
equations, following the approaches of [10, 16].
First, we study the truncation error for discretizing the spatial derivative. We denote the
central difference operator as

z i⫹1 ⫹ z i⫺1 ⫺ 2z i
␦ xx z i ⫽ ,
h x2

where zi ⫽ z(ihx). The central difference discretization of Equation (11) can be written as

1
␦ xx z i ⫽ z xx,i ⫹ z h 2 ⫹ O共h x4 兲,
12 xxxx,i x

where zxx,i and zxxxx,i denote the second and fourth derivative of z(x) at xi, respectively. To
achieve O(h4x ) accuracy, the term with h2x must be further approximated. For approximating the
zxxxx,i term, we have from (11) (after double differentiation with respect to x)

z xxxx,i ⫽ p xx,i ⫽ ␦ xx p i ⫹ O共h x2 兲,

where pi ⫽ p(ihx). Notice that

p i⫹1 ⫹ p i⫺1 ⫺ 2p i
␦ xx p i ⫽ ,
h x2

we have

1
p i ⫽ z xx,i ⫽ ␦ xx z i ⫺ ␦ p h 2 ⫹ O共h x4兲,
12 xx i x

or

10p i ⫹ p i⫹1 ⫹ p i⫺1 z i⫹1 ⫹ z i⫺1 ⫺ 2z i


⫽ ⫹ O共h x4 兲. (12)
12 h 2x
852 SUN AND ZHANG

We point out that the fourth-order compact methodology have been extended to more general
partial differential equations [10 –13].

IV. COMPACT DIFFERENCE SCHEME WITH BVMs

The high-order compact difference schemes and the BVMs can be combined to form a class of
high-order compact difference BVMs for solving the heat equations. In this section, we will
combine BVMs with the fourth-order compact difference scheme given in the previous section
to design a particular high-order finite difference scheme, a compact BVM scheme (CBVM), for
solving Equation (1).
For simplicity we write the approximate solution of u and its temporal derivative at the spatial
grid point xi as

⭸u
u i 共t兲 ⫽ u共x i , t兲, and u⬘i共t兲 ⫽ 共x , t兲, i ⫽ 0, 1, . . . , n.
⭸t i

We can use a semi-discretization scheme to approximate the spatial variable. Denote

u i⫹1 共t兲 ⫹ u i⫺1 共t兲 ⫺ 2u i 共t兲


␦ xx u i共t兲 ⫽ ,
h x2

we have

1
␦ xx u i共t兲 ⫽ u xx,i共t兲 ⫹ u 共t兲h x2 ⫹ O共h x4 兲,
12 xxxx,i

and

1 1
u xxxx,i 共t兲 ⫽ u⬘ 共t兲 ⫽ 2 ␦ xx u⬘i 共t兲 ⫹ O共h x2 兲.
␣ 2 x x,i ␣

The derivative in this equation is with respect to the time t.


Following the approach outlined in the previous section, we have

u⬘i ⫹1 共t兲 ⫹ u⬘i ⫺1 共t兲 ⫺ 2u⬘i 共t兲


␦ xx u⬘i 共t兲 ⫽ ,
h x2

and

1
u⬘i 共t兲 ⫽ ␣ 2 u xx,i 共t兲 ⫽ ␣ 2 ␦ xx u i 共t兲 ⫺ ␦ u⬘共t兲h 2x ⫹ O共h x4兲.
12 xx i

We then, like in (12), obtain a semi-discretization scheme


A HIGH-ORDER COMPACT BOUNDARY VALUE METHOD 853

1 ␣2
共10u i 共t兲 ⫹ u i⫹1 共t兲 ⫹ u i⫺1 共t兲兲⬘ ⫽ 2 共u i⫹1 共t兲 ⫹ u i⫺1 共t兲 ⫺ 2u i 共t兲兲. (13)
12 hx

We remark that the boundary conditions u0(t) ⫽ a(t) and un(t) ⫽ b(t) can be incooperated. We
can give the formula (13) in a matrix form as

共A x u共t兲 ⫹ c共t兲兲⬘ ⫽ Bxu共t兲 ⫹ d共t兲 ⫹ O共h4x 兲, t ⱖ 0, (14)

in which

冤 冥 冤 冥
10 1 ⫺2 1
1 1 10 1 ␣2 1 ⫺2 1
·· ·· ·· ·· ·· ··
Ax ⫽ · · · , Bx ⫽ · · · , (15)
12 1 10 1 h x2 1 ⫺2 1
1 10 1 ⫺2

and

冤 冥 冤 冥 冤冥 冤 冥 冤冥
u1共t兲 c1共t兲 a共t兲 d1共t兲 a共t兲
u2共t兲 c2共t兲 ␣2 0 d2共t兲 1 0
u共t兲 ⫽ ·· , d共t兲 ⫽ ·· ⫽ 2 ·· , c共t兲 ⫽ ·· ⫽ · .
· · hx · · 12 ··
un⫺1共t兲 cn⫺1共t兲 b共t兲 dn⫺1共t兲 b共t兲

It is easy to see that the formula (13) is a semi-discrete ODE with a fourth-order accuracy in
space.
Now we consider the discretization of the temporal variable. From (9) and (10) (see [4]), we
can discretize (1) as

共A t 丢 A x 兲u ⫹ 共At 丢 In⫺1兲c ⫽ 共Bt 丢 Bx兲u ⫹ 共Bt 丢 In⫺1兲d ⫹ e1 丢 f ⫹ O共hx4 ⫹ h4t 兲, (16)

where the matrices At and Bt are defined in (6) and (7), and

u ⫽ 关u1共t0兲, . . . , un⫺1共t0兲, u1共t1兲, . . . , un⫺1共t1兲, . . . , u1共tN兲, . . . , un⫺1共tN兲兴T,


c ⫽ 关c1共t0兲, . . . , cn⫺1共t0兲, c1共t1兲, . . . , cn⫺1共t1兲, . . . , c1共tN兲, . . . , cn⫺1共tN兲兴T,
d ⫽ 关d1共t0兲, . . . , dn⫺1共t0兲, d1共t1兲, . . . , dn⫺1共t1兲, . . . , d1共tN兲, . . . , dn⫺1共tN兲兴T,
f ⫽ 关f 共t0兲, f共t1兲, . . . , f共tN兲兴T.

The formula (16) results in a system of linear equations

Mu ⫽ q, (17)

where the coefficient matrix is

M ⫽ A t 丢 A x ⫺ B t 丢 B x, (18)

and the right-hand side is q ⫽ (Bt R In⫺1)d ⫹ e1 R f ⫺ ( At R In⫺1)c.


854 SUN AND ZHANG
TABLE I. Maximum absolute errors of the CBVM and CN schemes for solving the Problem 1.
CN CBVM
hx ⫽ ht Error Order CPU Error Order CPU

1/5 1.1e ⫺ 1 0.01 2.8e ⫺ 2 0.01


1/10 3.0e ⫺ 2 1.87 0.01 3.8e ⫺ 3 2.87 0.01
1/20 6.9e ⫺ 3 2.10 0.01 2.7e ⫺ 4 3.82 0.03
1/40 1.7e ⫺ 3 2.02 0.03 1.3e ⫺ 5 4.37 0.10
1/80 4.2e ⫺ 4 2.02 0.09 5.1e ⫺ 7 4.67 0.52
1/160 1.1e ⫺ 4 1.93 0.29 3.6e ⫺ 8 3.81 4.06

V. NUMERICAL RESULTS

We consider two problems to experimentally illustrate the high-order accuracy and stability of
the derived CBVM scheme and compare its performance with the Crank-Nicolson (CN) scheme.
We remark that in those two test problems, both Ax and Bx in (15) are symmetric tridiagonal
Toeplitz matrices, which can be diagonalized by sine transform (see, e.g., [17]). Thus, for
solving (17), we can use fast sine transform (FST) to diagonalize Ax and Bx of the matrix M in
(18) first. The resulting matrix will be a block band matrix with each block matrix being a
diagonal matrix. According to Theorem 1, permuting this matrix can result in a block diagonal
matrix with each block matrix being a band matrix. Hence we can solve the coefficient matrix
(18) block by block with a fast band matrix solver for solving each block. All numerical results
are computed with MATLAB 5.3 on a Sun Ultra 5 machine.

Problem 1.

⭸u ⭸ 2 u
⫽ 2, 0 ⬍ x ⬍ 1, t ⬎ 0,
⭸t ⭸x

with boundary and initial conditions

u共x, 0兲 ⫽ sin共␲x兲, 0 ⬍ x ⬍ 1, u共0, t兲 ⫽ u共1, t兲 ⫽ 0, t ⱖ 0.

The exact solution of this problem is u(x, t) ⫽ e⫺␲ tsin(␲x).


2

We solve the Problem 1 with hx ⫽ ht ⫽ 1/5, 1/10, . . . , 1/160. The final time is chosen as T ⫽
1. The maximum absolute errors, the accuracy order, and the CPU time in seconds of the CBVM
and CN schemes are shown in Table I. Clearly, the CBVM scheme provides much more accurate
solution and demonstrates fourth-order accuracy in both temporal and spatial variables. The CN
scheme is seen to be of second-order accuracy, because the error is reduced approximately by
a factor of 4 when the mesh is refined by one half. The CBVM scheme demonstrates the
fourth-order accuracy, because the error is reduced approximately by a factor of 16 when the
mesh is refined by one half.
Figure 1 depicts the maximum absolute errors of the CBVM and CN schemes as a function
of the temporal variable t for solving the Problem 1 with hx ⫽ ht ⫽ 1/20. It can be seen that the
computed errors of both schemes do not magnify with the time steps. This experiment
demonstrates the stability of both the CBVM and CN schemes. It also shows that the errors of
the CBVM scheme are much smaller than that of the CN scheme at every time step.
A HIGH-ORDER COMPACT BOUNDARY VALUE METHOD 855

FIG. 1. Maximum absolute errors of the CBVM and CN schemes at each time step for the Problem 1.

Problem 2.

⭸u 1 ⭸ 2u
⫽ 2 2, 0 ⬍ x ⬍ 1, t ⬎ 0,
⭸t ␲ ⭸x

with boundary and initial conditions

u共x, 0兲 ⫽ sin共␲x兲, 0 ⬍ x ⬍ 1, u共0, t兲 ⫽ u共1, t兲 ⫽ 0, t ⱖ 0.

The exact solution of this problem is u(x, t) ⫽ e⫺tsin(␲x).


We solve the Problem 2 with ht ⫽ hx ⫽ 1/5, 1/10, . . . , 1/160. The final time is again chosen
as T ⫽ 1. The computational data analogous to those reported in Table I are listed in Table II.
We see again that the CBVM scheme yields much more accurate solution than the CN scheme
does. Although for the same values of hx and ht, the CBVM scheme is more expensive to
implement than the CN scheme, the better computational efficiency of the CBVM scheme is
obvious. For instance, if we want to compute a solution with the maximum absolute error at the
final time no greater than 1.5 ⫻ 10⫺5, the CN scheme needs to use hx ⫽ ht ⫽ 1/160 for about

TABLE II. Maximum absolute errors of the CBVM and CN schemes for solving the Problem 2.
CN CBVM
hx ⫽ ht Error Order CPU Error Order CPU

1/5 1.1e ⫺ 2 0.01 2.3e ⫺ 4 0.01


1/10 2.7e ⫺ 3 2.03 0.01 1.5e ⫺ 5 3.94 0.01
1/20 6.8e ⫺ 4 1.99 0.01 9.5e ⫺ 7 3.98 0.03
1/40 1.7e ⫺ 4 2.00 0.03 5.9e ⫺ 8 4.01 0.10
1/80 4.2e ⫺ 5 2.02 0.09 3.7e ⫺ 9 4.00 0.52
1/160 1.1e ⫺ 5 1.93 0.30 2.3e ⫺ 10 4.01 4.21
856 SUN AND ZHANG

FIG. 2. The computed solution of the Problem 2 using CBVM with hx ⫽ ht ⫽ 1/20.

1.93 seconds. The CBVM scheme needs only 0.01 seconds by using hx ⫽ ht ⫽ 1/10. In this
sense, the CBVM scheme is seen to be much faster than the CN scheme.
Figure 2 is the solution profile of the Problem 2, computed using the CBVM scheme with
hx ⫽ ht ⫽ 1/20. We can see that the computed solution is not oscillatory. This again supports
the claim that the CBVM scheme is stable.

VI. CONCLUDING REMARKS

In this article, we propose a class of new finite difference schemes, CBVM, for solving the
one-dimension heat equations. The CBVM schemes are of high-order accuracy and uncondi-
tionally stable. For the fourth-order compact CBVM that we implemented, it computes much
more accurate solution than the CN scheme does. CBVM can be extended to variable coefficient
cases with fourth-order accuracy both in time and space. However, the matrices Ax and Bx will
no longer be tridiagonal Toeplitz matrices. Thus we cannot apply DST to solve them. But the
CN scheme can be solved step by step directly. Bertaccini [18] and Chan et al. [7] provide a fast
algorithm for solving this type of discretized system when the coefficients are independent of t.
How to develop a fast algorithm for the general case is an interesting topic that we intend to
investigate in the future.
To some extend, the CBVM schemes use similar ideas as in the high-order collocation
methods [19 –21]. The high-accuracy solution of both methods is computed by solving larger
sparse linear systems, compared to that of the CN scheme. Nevertheless, as we remarked in the
numerical result section, to compute numerical solution of comparable accuracy, the CBVM
scheme is much faster than the CN scheme. Like the collocation methods, the CBVM scheme
can also be viewed as accommodating potential parallel computation across time, as we can use
larger time steps to compute solution of comparable accuracy [19, 20].
A HIGH-ORDER COMPACT BOUNDARY VALUE METHOD 857

References
1. J. Crank and P. Nicolson, A practical method for numerical evaluation of solutions of partial
differential equations of the heat-conduction type, Proc Cambridge Philos Soc 43 (1947), 50 – 67.
2. G. Smith, Numerical solution of partial differential equations: finite difference methods, 2nd Ed.,
Oxford University Press, Oxford, 1978.
3. M. M. Chawia, M. A. Al-Zanaidi, and A. Z. Al-Shammeri, High-accuracy finite-difference schemes
for the diffusion equation, Neural Parallel Sci Comput 6 (1998), 523–535.
4. L. Brugnano and D. Trigiante, Solving differential problems by multistep initial and boundary value
methods, Gordon and Beach Science Publishers, Amsterdam, 1998.
5. L. Brugnano and D. Trigiante, Stability properties of some BVM methods, Appl Numer Math 13
(1993), 201–304.
6. L. Brugnano and D. Trigiante, Boundary value methods: the third way between linear multistep and
Runge-Kutta methods, Computers Math Applic 36(10 –12) (1998), 269 –284.
7. R. Chan, M. Ng, and X. Jin, Strang-type preconditioners for systems of LMF-based ODE codes, IMA
J Numer Anal 21 (2001), 451– 462.
8. P. Ghelardoni and P. Marzulli, Stability of some boundary value methods for IVPs, Appl Num Math
18 (1995), 141–153.
9. P. Amodio, F. Mazzia, and D. Trigiante, Stability of some boundary value methods for the solution
of initial value problems, BIT 33 (1993), 434 – 451.
10. M. M. Gupta, R. P. Manohar, and J. W. Stephenson, High-order difference schemes for two-
dimensional elliptic equations, Numer Methods Partial Differential Eq 1 (1985), 71– 80.
11. W. F. Spotz and G. F. Carey, High-order compact scheme for the steady stream-function vorticity
equations, Int J Numer Methods Engrg 38 (1995), 3497–3512.
12. J. Zhang, An explicit fourth-order compact finite difference scheme for three dimensional convection-
diffusion equation, Commun Numer Methods Engrg 14 (1998), 209 –218.
13. J. Zhang, L. Ge, and J. Kouatchou, A two colorable fourth-order compact difference scheme and
parallel iterative solution of the 3D convection diffusion equation, Math Comput Simulation 54(1–3)
(2000), 65– 80.
14. W. F. Spotz and G. F. Carey, Extension of high order compact schemes to time dependent problems,
Numer Methods Partial Differential Eq 17(6) (2001), 657– 672.
15. J. Zhang and J. J. Zhao, High accuracy stable numerical solution of 1D microscale heat transport
equation, Commun Numer Methods Engrg 17 (2001), 821– 832.
16. W. F. Spotz, High-order compact finite difference schemes for computational mechanics, Ph.D. thesis,
University of Texas at Austin, Austin, TX, 1995.
17. R. Chan, M. Ng, and C. Wong, Sine transform based preconditioners for symmetric Toeplitz systems,
Linear Algebra Appl 232 (1996), 237–260.
18. D. Bertaccini, A circulant preconditioner for the systems of LMF-based ODE codes, SIAM J Sci
Comput 22(3) (2000), 767–786.
19. F. Jèzèquel, A validated parallel across time and space solution of the heat transfer equation, Appl
Numer Math 31 (1999), 65–79.
20. J. Kouatchou, Parallel implementation for a high-order implicit collocation method for the heat
equation, Math Comput Simulation 54 (2001), 509 –519.
21. J. Kouatchou, Finite differences and collocation methods for the solution of the two-dimensional heat
equation, Numer Methods Partial Differential Eq 17 (2001), 54 – 63.

You might also like