Date post: | 03-Dec-2016 |
Category: |
Documents |
Upload: | srinivasan |
View: | 213 times |
Download: | 1 times |
This article was downloaded by: [University of Stellenbosch]On: 24 February 2013, At: 03:12Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of ComputerMathematicsPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/gcom20
Higher-order time accurate numericalmethods for singularly perturbedparabolic partial differential equationsRajdeep Deb a & Srinivasan Natesan ba Department of Chemical Engineering, Indian Institute ofTechnology Guwahati, Guwahati, Indiab Department of Mathematics, Indian Institute of TechnologyGuwahati, Guwahati, Assam, IndiaVersion of record first published: 17 Jun 2009.
To cite this article: Rajdeep Deb & Srinivasan Natesan (2009): Higher-order time accuratenumerical methods for singularly perturbed parabolic partial differential equations, InternationalJournal of Computer Mathematics, 86:7, 1204-1214
To link to this article: http://dx.doi.org/10.1080/00207160701798764
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions
This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representationthat the contents will be complete or accurate or up to date. The accuracy of anyinstructions, formulae, and drug doses should be independently verified with primarysources. The publisher shall not be liable for any loss, actions, claims, proceedings,demand, or costs or damages whatsoever or howsoever caused arising directly orindirectly in connection with or arising out of the use of this material.
International Journal of Computer MathematicsVol. 86, No. 7, July 2009, 1204–1214
Higher-order time accurate numerical methods for singularlyperturbed parabolic partial differential equations
Rajdeep Deba and Srinivasan Natesanb*
aDepartment of Chemical Engineering, Indian Institute of Technology Guwahati, Guwahati, India;bDepartment of Mathematics, Indian Institute of Technology Guwahati, Guwahati, Assam, India
(Received 26 February 2007; revised version received 26 May 2007; second revision received 01 August 2007;accepted 01 November 2007 )
This article presents two numerical methods for singularly perturbed time-dependent reaction-diffusioninitial–boundary-value problems. The spatial derivative is replaced by a hybrid scheme, which is a com-bination of the cubic spline and the classical central difference scheme in both the methods. In the firstmethod, the time derivative is replaced by the Crank–Nicolson scheme, whereas in the second methodthe time derivative is replaced by the extended-trapezoidal scheme. These schemes are applied on thelayer resolving piecewise-uniform Shishkin mesh. Some numerical examples are carried out to show theaccuracy and efficiency of these methods.
Keywords: singular perturbed parabolic problem; cubic spline; piecewise-uniform Shishkin mesh;Crank–Nicolson scheme; extended-trapezoidal scheme
2000 AMS Subject Classification: 65M06; 65M12
CCS Category: G1.8
1. Introduction
In this article, we consider the following singularly perturbed initial–boundary-value problem(IBVP):
∂u
∂t− ε
∂2u
∂x2+ b(x, t)u(x, t) = f (x, t), (x, t) ∈ � = (0, 1) × (0, T ],
u(x, 0) = s(x), on Sx = {(x, 0) : 0 ≤ x ≤ 1},u(0, t) = a0(t), on S0 = {(0, t) : 0 ≤ t ≤ T },u(1, t) = a1(t), on S1 = {(1, t) : 0 ≤ t ≤ T },
(1)
where 0 < ε � 1 is a small parameter, and the functions b, f are sufficiently smooth functionssuch that b(x, t) ≥ β > 0 on �. Under suitable continuity and compatibility conditions on the
*Corresponding author. Email: [email protected]
ISSN 0020-7160 print/ISSN 1029-0265 online© 2009 Taylor & FrancisDOI: 10.1080/00207160701798764http://www.informaworld.com
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
International Journal of Computer Mathematics 1205
data, a unique solution u(x, t) of Equation (1) exists. Parabolic boundary layers occur in thesesolutions when ε → 0. Classical numerical methods such as finite difference methods and finiteelement methods fail to yield satisfactory numerical results on uniform mesh, because of the pres-ence of boundary layers. To overcome these difficulties one has to use specially fitted nonuniformmesh. More details on special meshes for singularly perturbed parabolic problems can be seen inthe books of Farrell et al. [2], Miller et al. [7], and Roos et al. [11].
There are several articles dealt with the numerical methods for time-dependent convection-diffusion and reaction-diffusion problems. To cite a few: Stynes and O’Riordan [12] presented auniformly convergent finite element method for these types of problems using exponential basisfunctions. In [6], finite element method of exponentially fitted lumped schemes were given. Farrellet al. [3–5] proposed numerical methods for IBVPs of the form (1). Recently, Natesan and Deb[10] developed a robust numerical scheme for the parabolic reaction-diffusion IBVP (1), andobtained ε-uniform error estimates of order O(N−2
x ln2 Nx + N−1t ).
In this article, we propose two numerical schemes for the parabolic IBVP (1), which are ofhigher-order convergent in the time variable. More precisely, first we derive the schemes for thespatial derivative on variable meshes, and then we apply it on a piecewise-uniform Shishkin mesh.In both the methods the spatial derivatives are replaced by a hybrid scheme, in which the cubicspline scheme (for the boundary layer regions) and the classical finite difference scheme (forthe regular region) are combined suitably. In the first difference scheme (denoted as differencescheme-I), the time derivative is replaced by a second-order O(N−2
t ) accurate Crank–Nicolsonscheme, whereas in the second difference scheme (denoted as difference scheme-II) the timederivative is replaced by a third-order O(N−3
t ) accurate extended-trapezoidal scheme used byChawla and Al-Zanaidi [1]. These schemes provide higher-order accuracy for the time part, andthe hybrid scheme for the spatial derivative takes care of the boundary layer effect. By this waythe proposed schemes resolve the boundary layer with higher-order approximation for the timecomponent.
Section 2 presents the two numerical schemes for the IBVP (1). Some numerical experimentsare carried out in Section 3. Basically, the numerical examples are tested to show the accu-racies for both the spatial and temporal components separately, and also we have calculatedthe normalized flux by these schemes, which will be very useful to calculate the errors in theH 1-energy norm.
2. The numerical schemes
In this section, first, we present the piecewise-uniform Shishkin mesh for the spatial discretizationof the domain. And then, we derive the difference schemes on the Shishkin mesh. In both theseschemes, we discretize the spatial derivative by a hybrid difference scheme, which is a combinationof the cubic spline scheme and the classical central difference scheme. In the first method, the timederivative is replaced by the Crank–Nicolson scheme, whereas the extended-trapezoidal schemeis used in the second method for the time derivative.
2.1 Discretization of the domain
Consider the domain � = (0, 1) × (0, T ]. First, we present the Shishkin mesh for the spatial part:let D = [0, 1] be the spatial domain, which is divided into three subintervals as D = [0, σ ) ∪[σ, 1 − σ ] ∪ (1 − σ, 1] for some σ such that 0 < σ ≤ 1/4. On the subintervals [0, σ ), and (1 −σ, 1] a uniform mesh with Nx/4 mesh-intervals are placed, where [σ, 1 − σ ] has a uniform meshwith Nx/2 mesh-intervals. It is obvious that the mesh is uniform when σ = 1/4, and it is fitted to
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
1206 R. Deb and S. Natesan
the problem by choosing σ be the following function of Nx, ε
σ = min
{1
4, σ0
√ε ln Nx
},
where σ0 ≥ 2 is a constant.Here, we use the multi-index notation N = (Nx, Nt), where Nt denotes the number of mesh
elements in the t-direction. We shall introduce the meshes in the spatial and temporal variables,respectively, as ωNx : 0 = x0 < x1 < . . . < xNx
= 1, and ωNt : 0 = t0 < t1 < . . . < tNt= T . Let
the meshes in � be the tensor product of the one-dimensional meshes ωNx and ωNt , and denote
it by �Nx,Nt
ε . Further, let hi = xi+1 − xi be the mesh diameter in the spatial dimension, and τj =tj+1 − tj , with �i = (hi−1 + hi)/2, h = 4σ/Nx, H = 2(1 − 2σ)/Nx , and k = maxj=1,...,Nt
τj .The domain is discretized in the x-direction with Shishkin mesh and uniform mesh in thet-direction.
2.2 Difference scheme-I
Difference scheme-I consists of the Crank–Nicolson scheme for the time derivative, and the hybriddifference scheme for the spatial derivatives. Consider the Crank–Nicolson scheme(
uj+1i − u
j
i
k
)= 1
2
(g
j
i + gj+1i
), i = 1, . . . , Nx − 1, j = 0, . . . , Nt − 1, (2)
where
gj
i = εδ2hbdu
j
i − bj
i uj
i + fj
i ,
gj+1i = εδ2
hbduj+1i − b
j+1i u
j+1i + f
j+1i .
Let Mj
i = δ2hbdu
j
i , where δ2hbd is such that for 0 < i < N/4, 3N/4 < i < N :
δ2hbd =
((D−/hi−1) − (2�i/hihi−1) + (D+/hi)
)(((hi−1D−)/6) + (2�i/3) + ((hiD+)/6))
,
and for N/4 ≤ i ≤ 3N/4:
δ2hbd =
(D−
hi−1− 2�i
hihi−1+ D+
hi
),
here D+uj
i = uj
i+1, D−u
j
i = uj
i−1. Then, we have for 0 < i < N/4, 3N/4 < i < N :
hi−1
6M
j
i−1 +(
hi−1 + hi
3
)M
j
i + hi
6M
j
i+1 = uj
i−1
hi−1− 2�iu
j
i
hi−1hi
+ uj
i+1
hi
,
and for N/4 ≤ i ≤ 3N/4:
Mj
i = uj
i−1
�ihi−1− 2u
j
i
hi−1hi
+ uj
i+1
�ihi
.
The above system of equations can be represented in the following compact form
r−i−1M
j+1i−1 + rc
i Mj+1i + r+
i+1Mj+1i+1 = q−
i−1uj+1i−1 + qc
i uj+1i + q+
i+1uj+1i+1 , (3)
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
International Journal of Computer Mathematics 1207
where for 0 < i < N/4, 3N/4 < i < N ,
r−i−1 = hi−1
6, rc
i = 2�i
3, r+
i+1 = hi
6,
q−i−1 = 1
hi−1, qc
i = −2�i
hi−1hi
, q+i+1 = 1
hi
,
and for N/4 ≤ i ≤ 3N/4,
r−i−1 = 0, rc
i = 1, r+i+1 = 0,
q−i−1 = 1
hi−1�i
, qci = −2
hi−1hi
, q+i+1 = 1
hi�i
.
The system (3) can be written in the matrix form as
AMj+1 = BUj+1 + Cj+1, j = 0, . . . , Nt − 1,
where
A =
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
rc1 r+
2
r−1 rc
2 r+3
· · ·· · ·
· · ·r−Nx−2 rc
Nx−1
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦Nx−1×Nx−1
B =
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
qc1 q+
2
q−1 qc
2 q+3
· · ·· · ·
· · ·q−
Nx−2 qcNx−1
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦Nx−1×Nx−1
Mj+1 =
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎣
Mj+11
Mj+12
...
Mj+1Nx−2
Mj+1Nx−1
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎦Nx−1×1
Uj+1 =
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎣
uj+11
uj+12
...
uj+1Nx−2
uj+1Nx−1
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎦Nx−1×1
Cj+1 =
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎣
q−0 u
j+10 − r−
0 Mj+10
0
...
0
q+Nx
uj+1Nx
− r+Nx
Mj+1Nx
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎦Nx−1×1
.
In a similar way, we obtain
AMj = BUj + Cj ,
and, so
gj = εMj − Bj
1 Uj + F j ,
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
1208 R. Deb and S. Natesan
where
Bj
1 =
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎣
bj
1
bj
2
. . .
bj
Nx−2
bj
Nx−1
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎦Nx−1×Nx−1
F j =
⎡⎢⎢⎢⎢⎢⎢⎣f
j
1
fj
2
...
fj
Nx−1
⎤⎥⎥⎥⎥⎥⎥⎦Nx−1×1
Thus,
gj = ε(A−1BUj + A−1Cj
) − Bj
1 Uj + F j . (4)
Similarly, we have
gj+1 = ε(A−1BUj+1 + A−1Cj+1
) − Bj+11 Uj+1 + F j+1. (5)
Using the values of Equations (4) and (5) in Equation (2), we have the following relation
Uj+1 − Uj =(
k
2
) [ε(A−1BUj + A−1Cj
) − Bj
1 Uj + F j]+
+(
k
2
) [ε(A−1BUj+1 + A−1Cj+1
) − Bj+11 Uj+1 + F j+1
].
(6)
The difference scheme (6) can be written in the compact form as
Pj+11 Uj+1 = Q
j
1Uj + R
j+11 , j = 0, . . . , Nt − 1, (7)
where
Pj+11 = I − kε
2A−1B + k
2B
j+11 ,
Qj
1 = I + kε
2A−1B − k
2B
j
1 ,
Rj+11 = k
2
(F j + F j+1 + εA−1Cj + εA−1Cj+1
).
One can see from the truncation errors using Taylor’s expansion that the Crank–Nicolsonscheme is of O(N−2
t ) accurate for the time derivative. In [9], the hybrid scheme has been shownof O(N−2
x ln2 Nx) for the spatial derivatives. The error estimates of this difference will be carriedout in our forthcoming article [10].
2.3 Difference scheme-II
Difference scheme-II uses the extended-trapezoidal formula for the time derivative, and the hybridscheme for the spatial derivative.
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
International Journal of Computer Mathematics 1209
Here, we use the extended-trapezoidal formula to discretize the time derivative, as
uj+1i − u
j
i
k=
(5g
j
i + 8gj+1i − g
j+2i
12
), i = 1, . . . , Nx − 1, j = 0, . . . , Nt − 1, (8)
where
gj
i = εMj
i − bj
i uj
i + fj
i ,
gj+1i = εM
j+1i − b
j+1i u
j+1i + f
j+1i ,
gj+2i = εM
j+2i − b
j+2i u
j+2i + f
j+2i ,
uj+2i = 5u
j
i − 4uj+1i + 2k
(g
j
i + 2gj+1i
).
In vector form uj+2i can be written as
U j+2 = 5Uj − 4Uj+1 + 2k[(
εMj − Bj
1 Uj + F j)
+ 2(εMj+1 − B
j+11 Uj+1 + F j+1
)].
After rearranging the terms, and using Equations (4) and (5), we have
U j+2 =(DjUj + Ej+1Uj+1 + F
j+11
),
where
Dj =(
5I + 2kεA−1B − 2kBj
1
),
Ej+1 =(−4I + 4kεA−1B − 4kB
j+11
),
Fj+11 = 2k
(F j + 2F j+1 + εA−1Cj + 2εA−1Cj+1
).
Now, substituting all these values in Equation (8), and rearranging it, we obtain
(Pj+12 − R
j+22 Ej+1)Uj+1 = (Q
j
2 + Rj+22 Dj)Uj + (R
j+22 F
j+11 + F
j+22 ), j = 0, . . . , Nt − 1
(9)where
Pj+12 = 1
3
(3I − 2kεA−1B + 2kB
j+11
), R
j+22 = k
12
(−εA−1B + B
j+21
),
Qj
2 = 1
12
(12I + 5kεA−1B − 5kB
j
1
), F
j+11 = 2k
(F j + 2F j+1 + εA−1Cj + 2εA−1Cj+1
),
Fj+22 = k
12
(5F j + 8F j+1 − F j+2 + 5εA−1Cj + 8εA−1Cj+1 − εA−1Cj+2
).
The truncation error of order O(N−3t ) and unconditional stability for the extended-trapezoidal
formula for parabolic reaction-diffusion problems on uniform mesh are derived in [1].We postponethe error analysis of this scheme to our future article [10].
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
1210 R. Deb and S. Natesan
3. Numerical results
In this section, we implement the proposed schemes to two test problems studied by variousresearchers. The numerical results have been given in terms of maximum point-wise errors, andrate of convergence. The maximum point-wise error and the ε-uniform errors are obtained to showseparately the spatial effect and the temporal effect, mainly because the difference scheme-I isof O(N−2
t ) accurate and Difference scheme-II is of O(N−3t ) accurate. Further, we have obtained
approximations for the normalized flux, which will play a crucial role in the H 1-energy normerror estimates of finite element analysis.
Example 3.1 [8] Consider the following parabolic IBVP:
ut (x, t) − εuxx(x, t) + u(x, t) = f (x, t), (x, t) ∈ (0, 1) × (0, 1]. (10)
The right-hand side source term, initial and boundary conditions have been calculated from theexact solution
u(x, t) =(
t + x2
2ε
)erfc
(x
2√
εt
)−
√t
πεxe−x2/4εt .
The exact solution is used to calculate the maximum nodal error. More precisely, we determinethe maximum error and ε-uniform error by
ENx,Nt
ε = max�
Nx ,Ntε
|u(xi, tj ) − uj
i |, and ENx,Nt = maxε
ENx,Nt
ε ,
where u(x, t) denotes the exact solution, and uj
i stands for the numerical solution obtained by using
Nx, Nt mesh-intervals in the domain �Nx,Nt
ε . In addition, the rate of convergence is calculated by
p = log2
(ENx
ε
E2Nxε
), and p
uni= log2
(ENx
E2Nx
).
Also, we have calculated the normalized flux
Fε(x, t) = √ε
∂u(x, t)
∂x,
and its numerical approximation
FNx
ε (x, t) = √εD+
x u(x, t).
The maximum error and ε-uniform error in the normalized fluxes have been calculated as
QNx
ε = max0≤t≤T
|Fε(0, t) − FNx
ε (0, t)|, and QNx = maxε
QNx
ε .
The rates of convergence for the normalized flux are determined by
q = log2
(QNx
ε
Q2Nxε
), and q
uni= log2
(QNx
Q2Nx
).
The numerical experiments are carried out to show two important points: (1) the spatial effecton the error of the solution and the normalized flux; and (2) the temporal effect in the solutionand the normalized flux.
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
International Journal of Computer Mathematics 1211
Table 1. ε-Uniform error ENx,Nt and rate of convergence puni for Example 3.1 by difference scheme-I(spatial effect).
Nx 16 32 64 128 256 512ENx,Nt 4.8208e-2 1.8812e-2 7.4719e-3 2.5722e-3 8.4513e-4 2.7074e-4puni 1.3577 1.3321 1.5385 1.6058 1.6423
Table 2. ε-Uniform error QNx and rate of convergence quni for Example 3.2 by difference scheme-I(spatial effect).
Nx 16 32 64 128 256 512QNx 6.4800e-1 4.8285e-1 3.2692e-1 2.0564e-1 1.2282e-1 7.0844e-2quni 0.4244 0.5626 0.6688 0.7436 0.7938
Table 3. ε-Uniform error ENx,Nt and rate of convergence puni for Example 3.1 by differencescheme-II (spatial effect).
Nx 16 32 64 128 256 512ENx,Nt 5.3210e-2 2.1149e-2 7.7576e-3 2.6409e-3 8.5688e-4 2.7257e-4puni 1.3311 1.4469 1.5546 1.6239 1.6525
Table 4. ε-Uniform error QNx and rate of convergence quni for Example 3.1 by differencescheme-II (spatial effect).
Nx 16 32 64 128 256 512QNx 6.4559e-1 4.8105e-1 3.2594e-1 2.0518e-1 1.2264e-1 7.0780e-2quni 0.4244 0.5616 0.6677 0.7425 0.7930
The ε-uniform error of the solution and the normalized flux, and the rate of convergence forthe difference scheme-I, which show the spatial effect are presented in Tables 1 and 2. The samecan be shown for the difference scheme-II in Tables 3 and 4.
The ε-uniform error of the solution and the normalized flux showing the temporal effect fordifference schemes I and II are, respectively given in Tables 5–8. From the results of Table 5, onecan notice that the difference scheme-I provides O(N−2
t ) accuracy for the temporal variable. Thishigher order accuracy can also be reflected in the normalized flux results shown in Table 6.
Table 7 highlights the O(N−3t ) accurate results for time variable in the solution by the difference
scheme-II, and Table 8 shows the same for the corresponding normalized flux. These results reveal
Table 5. ε-Uniform error ENx,Nt and rate of convergence puni for Example 3.1 bydifference scheme-I (temporal effect).
Nt 8 16 32 64 128ENx,Nt 6.1977e-3 1.6928e-3 4.4301e-4 1.1244e-4 2.8503e-5puni 1.8511 1.9040 1.9369 1.9682
Table 6. ε-Uniform error QNx and rate of convergence quni for Example 3.1 bydifference scheme-I (temporal effect).
Nt 8 16 32 64 128QNx 3.1731e-1 1.7280e-1 9.0350e-2 4.5924e-2 2.3223e-2quni 0.8768 0.9355 0.9763 0.9837
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
1212 R. Deb and S. Natesan
Table 7. ε-Uniform error ENx,Nt andrate of convergence puni for Example 3.1by difference scheme-II (temporal effect).
Nt 8 16ENx,Nt 8.9952e-4 1.1378e-4puni 2.9829
Table 8. ε-Uniform error QNx and rateof convergence quni for Example 3.1 bydifference scheme-II (temporal effect).
Nt 8 16QNx 1.2410e-1 4.5813e-2quni 1.4377
our claim of higher-order accuracy for the temporal part in these newly proposed differenceschemes.
Also, one can observe that the proposed method produces ε-uniform numerical results, this isthe outcome of the proposed hybrid scheme on the Shishkin mesh. Thus, we have resolved theboundary layers as well as obtained higher-order time accurate results by these schemes.
Further, we consider the following variable coefficient parabolic PDE to examine the differenceschemes.
Example 3.2 Consider the following parabolic IBVP:
ut (x, t) − εuxx(x, t) + √x + 1u(x, t) = 1, (x, t) ∈ (0, 1) × (0, 1]
u(0, t) = 0, u(1, t) = 0, for t ∈ (0, 1]u(x, 0) = 0, for x ∈ [0, 1] (11)
The exact solution of the IBVP (3.2) is not available, therefore, we determine the accuracy ofthe numerical solution by comparing it with the numerical solution computed on a fine mesh. Wedetermine the maximum point-wise error, and ε-uniform error respectively, by
ENx,Nt
ε = max�
Nx ,Ntε
||uNx,Nt − u4Nx,4Nt ||, and ENx,Nt = maxε
ENx,Nt
ε ,
where uNx,Nt denotes numerical solution obtained by using Nx and Nt mesh-intervals, respectively,
on the x- and t-axes, i.e., on the discretized domain �Nx,Nt
ε . Then bisect twice each interval ωNx andωNt and take their tensor product to obtain a mesh �4Nx,4Nt
ε that has 16 times as many subintervals
as �Nx,Nt
ε (by this way, one can obtain the same transition points as in �Nx,Nt
ε ). Let u4Nx,4Nt denotesthe numerical solution computed on this mesh. In addition, the rate of convergence is calculated by
p = log2
(ENx
ε
E2Nxε
), and p
uni= log2
(ENx
E2Nx
).
The maximum point-wise error and ε-uniform error in the normalized flux have been calculated as
QNx
ε = max�
Nx ,Ntε
||F Nx
ε (xi, tj ) − F 4Nx
ε (xi, tj )||, and QNx = maxε
QNx
ε ,
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
International Journal of Computer Mathematics 1213
Table 9. ε-Uniform error ENx,Nt and rate of convergence puni for Example 3.2 by difference scheme-I(spatial effect).
Nx 16 32 64 128 256 512ENx,Nt 7.1535e-3 2.8045e-3 1.0236e-3 3.4895e-4 1.1398e-4 3.6035e-5puni 1.3509 1.4540 1.5526 1.6142 1.6613
Table 10. ε-Uniform error QNx and rate of convergence quni for Example 3.2 by difference scheme-I(spatial effect).
Nx 16 32 64 128 256 512QNx 1.7713e-1 1.2658e-1 8.3911e-2 5.3076e-2 3.0833e-2 1.7737e-2quni 0.4848 0.5931 0.6608 0.7836 0.7977
Table 11. ε-Uniform error ENx,Nt and rate of convergence puni for Example 3.2 bydifference scheme-II (spatial effect).
Nx 16 32 64 128 256ENx,Nt 8.5433e-2 4.4195e-2 1.8666e-2 7.3712e-3 2.7092e-3puni 0.9509 1.2435 1.3405 1.4440
Table 12. ε-Uniform error QNx and rate of convergence quni for Example 3.2 bydifference scheme-II (spatial effect).
Nx 16 32 64 128 256QNx 1.0829e-1 6.8780e-2 4.5416e-2 3.1807e-2 2.0389e-2quni 0.6548 0.5988 0.5139 0.6415
where
F Nx
ε (xi, t) = √εD+
x u(x, t), for i = 1, . . . , Nx − 1, and F Nx
ε (xNx, t) = √
εD−x u(xNx
, t).
The numerical results obtained by Difference scheme-I are given in Tables 9 and 10 for thesolution and the normalized flux, respectively. Tables 11 and 12 show the same for the differencescheme-II.
Although we have not supplied the numerical results supporting the temporal effect of thesedifference schemes (mainly, to reduce the number of tables), we have experimented and observedthe same higher-order accuracy as like in the constant coefficient problem given in Example 3.1.Further the independency of ε can be observed from the numerical results.
4. Conclusions
This paper presents two numerical schemes for singularly perturbed parabolic reaction-diffusionIBVPs of the type (1). In both the schemes the spatial derivatives are replaced by a hybrid scheme,which combines the cubic spline and classical finite difference scheme in an appropriate manner.The time derivative is replaced by the O(N−2
t ) accurate Crank–Nicolson scheme, and the O(N−3t )
accurate extended-trapezoidal scheme for the first and second difference schemes, respectively.The numerical experiments reveal that the proposed schemes are capable of resolving the boundarylayer and producing higher-order time accurate approximations.
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3
1214 R. Deb and S. Natesan
Acknowledgement
This work is supported by the Department of Science and Technology, Government of India under research grantSR/S4/MS:318/06.
References
[1] M.M. Chawla and M.A. Al-Zanaidi, An extended trapezoidal formula for diffusion equation, Comput. Math. Appl.38 (1999), pp. 51–59.
[2] P.A. Farrell et al., Robust Computational Techniques for Boundary Layers, Chapman & Hall/CRC Press, BocaRaton, 2000.
[3] P.A. Farrell, P.W. Hemker, and G.I. Shishkin, Discrete approximations for singularly perturbed boundary valueproblems with parabolic layers, I, J. Comp. Math. 14(1) (1996), pp. 71–97.
[4] ———, Discrete approximations for singularly perturbed boundary value problems with parabolic layers. II,J. Comp. Math. 14(2) (1996), pp. 183–194.
[5] ———, Discrete approximations for singularly perturbed boundary value problems with parabolic layers. III,J. Comp. Math. 14(3) (1996), pp. 273–290.
[6] W. Guo and M. Stynes, Finite element analysis of exponentially fitted lumped schemes for time-dependent convection-diffusion problems, Numer. Math. 66 (1993), pp. 347–371.
[7] J.J.H. Miller, E. O’Riordan, and G.I. Shishkin, Fitted Numerical Methods for Singular Perturbation Problems, WorldScientific, Singapore, 1996.
[8] J.J.H. Miller et al., Fitted mesh methods for problems with parabolic boundary layers, Math. Proc. R. Ir. Acad.98A(2) (1993), pp. 173–190.
[9] S. Natesan and R. Deb, A robust numerical scheme for singularly perturbed parabolic reaction-diffusion problems,submitted for publication, 2007.
[10] ———, Error analysis for higher-order schemes for parabolic reaction-diffusion problems, Working paper, 2007.[11] H.-G. Roos, M. Stynes, and L.Tobiska, Numerical Methods for Singularly Perturbed Differential Equations, Springer,
Berlin, 1996.[12] M. Stynes and E. O’Riordan, Uniformly convergent difference schemes for singularly perturbed parabolic diffusion-
convection problems without turning points, Numer. Math. 55 (1989), pp. 521–544.
Dow
nloa
ded
by [
Uni
vers
ity o
f St
elle
nbos
ch]
at 0
3:12
24
Febr
uary
201
3