You are on page 1of 11

# Neural, Parallel, and Scientific Computations 17 (2009) 123-134

UNIFORM CONVERGENCE OF MONOTONE DIFFERENCE SCHEME FOR A SINGULARLY PERTURBED THIRD-ORDER CONVECTION-DIFFUSION EQUATION
ZHONGDI CEN Institute of Mathematics, Zhejiang Wanli University, Ningbo 315100, China ABSTRACT. A singularly perturbed third-order convection-diﬀusion problem is considered. The third-order boundary value problem is transformed into a system of weakly coupled system of two diﬀerential equations. Then the problem is solved numerically using a monotone diﬀerence scheme on a Bakhvalov-Shishkin mesh. We use appropriate estimates of the discrete Green’s function to obtain error estimates of the monotone diﬀerence scheme. Numerical experiments support our theoretical results. Key Words. convection-diﬀusion problem, singularly perturbed, Bakhvalov-Shishkin mesh, ﬁnite diﬀerence scheme, Green’s function.

1. INTRODUCTION Singular perturbation problems appear in many branches of applied mathematics, and for more than two decades quite a good number of research works on the qualitative and quantitative analysis of these problems for both ordinary diﬀerential equations and partial diﬀerential equations have been reported in the literature. Most of the papers connected with computational aspects are conﬁned to second order equation. But only few authors have developed numerical methods for singularly perturbed higher-order diﬀerential equations. In this paper, we treat the following third-order singularly perturbed ordinary diﬀerential equations (1.1) (1.2) −εy ′′′ (x) − a(x)y ′′ (x) + b(x)y ′ (x) − c(x)y (x) = f (x), x ∈ D, y (0) = p, y ′(0) = q, y ′(1) = r,

where 0 < ε ≪ 1 is a small positive parameter, a(x), b(x), c(x) and f (x) are suﬃciently smooth functions satisfying the following conditions: (1.3) (1.4) (1.5)

a(x) ≥ α > 0, b(x) ≥ 0, 0 ≥ c(x) ≥ −γ, γ > 0,
1061-5369 \$15.00 c Dynamic Publishers, Inc.

Lemma 1.Assume that P1 u ≥ 0. Notation. x ∈ D. to analyze the accuracy of ﬁnite diﬀerence schemes in the discrete maximum norm. ¯. 1] and y ∈ C (3) (D ) ∩ C (1) (D ¯ ).   y1 (0) = p. Note that C is not necessarily the same at each occurence. We analyze the convergence properties of the scheme in section 4. The present work is an extension of L1 norm stability inequality technique for second-order singularly perturbed boundary value problem to third-order singularly perturbed problem subject to a particular type of boundary conditions. Semper [4] and Roos et al. THE CONTINUOUS PROBLEM In this section we will present a priori bounds on the solution of (1.6) α − γ (1 + 3η ) ≥ η ′ > 0 for some η and η ′ . C will denote a generic positive constant that is independent of ε and of the mesh. Recently. where y = (y1 . u1 (0) ≥ 0. with D = (0.2). D In [1].In the next section we derive bounds for the derivatives of the solution of (1. y2). (Maximum principle [3]) Consider the boundary value problem (2. y2 (0) = q. The works of Gartland [2] are connected with exponentially ﬁtted higher order diﬀerence scheme with identity expansion method. D0 = (0.124 ZHONGDI CEN (1.1)-(1. The aim of this work is to illustrate an application of a priori estimates of the solutions of discrete problems. 1]. Niederdrenk and Yserentant considered higher order convection-diﬀusion type problems and derived conditions for the uniform stability of discrete and continuous problems.1)-(1.  ′′ ′ Ay = F ⇐⇒ P2 y ≡ −εy2 (x) − a(x)y2 (x) + b(x)y2 (x) + c(x)y1 (x) = f (x). Then u(x) ≥ 0 in [0.1). Finally numerical results are presented in section 5. Throughout the paper. 2. Also. Valarmathi and Ramanujam [3] proposed a method. P2 u ≥ 0 in D . ¯ = [0. 1]. u2 (0) ≥ 0.2) and its derivatives. 1). Assumption 1.1). to ﬁnd a numerical solution for the third-order problem (1.The singularly perturbed boundary value problem (1. In section 3 we introduce a upwind diﬀerence scheme and a Bakhvalov-Shishkin mesh. y2 (1) = r. and u2 (1) ≥ 0.2) can be transformed into an equivalent problem of the form (2. [5] considered fourth order equations and applied a standard ﬁnite element method. u2 (x)) for all x ∈ D . which are obtained using Green’s function. An outline of the paper is as follows.1)  ′  P1 y ≡ y 1 (x) − y2 (x) = 0. ’boundary value technique’. Here u(x) = (u1 (x).1)-(1. Throughout the paper we shall also assume that ε ≤ CN −1 as is generally the case.These bounds will be used in the error analysis in later sections.

|w2 (x)| ≤ Cε−k exp(−αx/ε) for all k ≤ j . MESH AND SCHEME Our mesh is a modiﬁcation of the Shishkin mesh:Bakhvalov-Shishkin mesh combined the Bakhvalov mesh in the boundary layer with the Shishkin-type transition point. 2. N }. x ∈ D. P2 w(x) = 0. ¯ ). |y2(0)|. max |P2 y|} for all x ∈ D. b(x). The bounds can be obtained using results of reference [3]. The step sizes of the mesh D N satisfy hi ≤ 8ε ≤ CN −1 . Lemma 4. N − 1}. · · · . |y2(x)|}. · · · . ¯ N = {xi |i = 0. max |P1 y|. our discretization parameter. |y2(1)|.THIRD-ORDER CONVECTION-DIFFUSION EQUATION 125 Lemma 2.2) Let D N = {xi |i = 1. If a(x). then ¯ y(x) ≤ C max{|y1 (0)|. w(0) ≤ C. N − 1}. Proof. xi = . −2 α 2(N −i) 2ε 1 − (1 − α ln N ) N for i = N/2 + 1. x ∈ D. then the solution y(x) of (1. 3. while the layer part w satisﬁes P1 w(x) = 0. 1]. ¯ x∈D ¯ x∈D where y(x) = max{|y1 (x)|.1)Lemma 3. The construction of layer-adapted meshes and the analysis of numerical methods for singularly perturbed problems require precise knowledge about the behavior of the derivatives of the exact solution.2) has the representation y = v + w on [0. If y is a smooth function. (Stability result [3]) Consider the boundary value problem (2. Let N .1) Then our mesh is (3. 2 α i ε ln[1 − 2(1 − N −1 ) N ] for i = 0. 1. for all k ≤ j . The following lemma provides that information. We choose the transition point σ as Shishkin does: (3. D The following lemma gives some estimates of the mesh sizes that will be used later. for i = 1. w(1) ≤ C exp(−α/ε) and (k ) (k ) ¯ |w1 (x)| ≤ Cε1−k exp(−αx/ε) . N/2. α(N/2 − i + 1) N D0 = {xi |i = 0. · · · . ln N }. N.1). 1. N/2. · · · . · · · . 1 2ε σ = min{ . 2. c(x) and f (x) ∈ C (j ) (D (1. P2v(x) = f (x) and ¯ v(k) (x) ≤ C. hN/2+i ≤ 2N −1 . · · · . where the smooth part v satisﬁes P1 v(x) = 0. be an even positive integer. 1.

N.N = r. i = 1. we use we deﬁne a discrete L1 norm by w N 1 · ∞ for the standard maximum norm. we can get results for the discrete problem.5).1 Discrete Green’s function and its properties For any mesh function w N . then N |y1.0|. The proof is analogous with that of Lemma 2. ξj ) = 0 .i − y2.3)-(3. then yi ≥ 0 for i = 0. N − 1.4) (3.i = 0.N |. w N ∈ RN −1 . max |P2 yi |}. (Stability result ) If yi is any mesh function. |y2. · · · .i| ≤ C max{|y2. xi ∈ D0 . ξj ) = δ N (xi . ξj ) = G2 (0.N ≥ 0. D − vi = . If y1.126 ZHONGDI CEN Proof. ξj ∈ D N . See [7]. N N N y1 . for i = 0.3) (3.4). 2 0 = h1 .i − ai Dy2. 1≤i≤N −1 Proof.i + ci y1. ξj ∈ D N . Dvi = and hi+1 hi i i N N N N P1 yi ≡ Dy1 . ξj ) = 0 . 4. N − 1. ξj ) = G2 (1. 1.1). N .3)-(3. 1. y2. · · · . · · · .0|. 1≤i≤N −1 N |y2. ξj ) .i = fi . y2. N N P2 G (xi . ξj ∈ D . Consider the Green’s function GN (xi . These can be easily calculated. · · · . w N ) = j =1 N N vj wj j. . max |P1 yi |}. · · · . 2. i = 0. N N N GN 1 (0. 4.0 ≥ 0. for i = 1.0 = p. As a function of xi for ﬁxed ξj this function is deﬁned by the relations (4. ANALYSIS OF THE METHOD Analogous to the continuous problem (2. N − 1. ξj ) of problem (3.3) N N N P1 G (xi .2) (4. 1. y2. = hi + hi+1 . (Discrete maximum principle [3]) Consider the discrete problem N (3. and N −1 = i=1 N i |wi |. 1. xi ∈ D N . y2.1) (4. Lemma 6. and N P2 yi ≥ 0 for i = 1. · · · . N.5) where D + vi = vi − vi−1 vi+1 − vi vi+1 − vi . ∀v N . We also deﬁne the scalar product in RN +1 by N −1 (v N .1 in [3].0 ≥ 0. Now we consider the upwind diﬀerence scheme (3. Lemma 5. 2. P1 yi ≥ 0 for i = 0.0 = q.i| ≤ C max{|y1.i + bi y2. N − 1. N N N N N N P2 yi ≡ −εD + D − y2 . · · · .

7). ξj ∈ D N .∗ N =⇒ P2 G (xi .4). ξj ) = δ N (xi . and the fact N.(3.j ) N + − N N N N = (δ N (xi . This arises from the following arguments: using (4.j + bj y2. y2 .5). The Green’s function GN (xi .6)-(3.THIRD-ORDER CONVECTION-DIFFUSION EQUATION 127 where δ N (xi . ξj ). ξj )P2 yj j N −1 = j =1 j N.i = j =1 GN 2 (xi . ξj ).j ) − (G1 (xi . ξj ) = 0 .j ) N N + − N N N N N = y2 . y2 .j + cj y1.j ) − (G1 (xi .we have . ξj ∈ D0 . ξj ). ξj ).∗ N N is adjoint to P1 .∗ N P2 G (xi . ξj ) .3)-(3.j ) − (G1 (xi . we can give the following formula for the solution of problem (3. ξj ). ξj ) ≤ . ξj ).j ) N N N N N N N = (P2 G . ξj ) = −1 i 0 for xi = ξj .4) N y1 . 1) = 0 . xi ∈ D N .5) N y2 . xi ∈ D N . 0) = G2 (xi . xi ∈ D N . ξj ). It is easy to see that using Green’s function. cj y2. fj ) = (G2 (xi .j ) + (D D G1 (xi .3) and (3.i .7).0 .∗ N G (xi . The Green’s function GN (xi .j ) = y2. P2 that P1 N. ξj ). cj y2. cj y1.i = j =1 N −1 N N y2 . ξj ) is nonnegative and bounded uniformly in ε: 1 0 ≤ GN (xi . ξj )fj j .6) (4. N. ξj ) = 0 P1 and N −1 N y2 . ξj ).∗ N. (4.8) N.i + y1.4) i−1 (4. taking into account (3.j ) + (G2 (xi . P2 respectively. α−γ . Lemma 7.j = j =1 N N GN 2 (xi .7) (4. for xi = ξj .j − aj Dy2. ξj ) = δ N (xi . −εD D y2.6)-(4.i + (G1 (xi . xi ∈ D .i N −1 = j =1 GN 2 (xi . ξj ). ξj )y2 . ξj )fj j N.∗ N N P1 G (xi . Indeed. cj y1. N N N GN 1 (xi . for xi ∈ D . we obtain N + − N N N N (G N 2 (xi . ξj ) as the function of a variable ξj for ﬁxed xi is the solution of the adjoint problem: (4. cj D D y1. 0) = G2 (xi . where we have used (4. cj y2.∗ N N P2 G (xi .

13) (α − γ )G N 2 (xi . ξj0 ). Multiply (4.5) and Lemma 7. Combining (4.6) we can get j −1 GN 2 (xi . ξj ) is nonnegative then (4. ξj0 ) ≤ j =1 δ N (xi . Because of the choice of ξj0 .∗ N P2 G (xi . On the other hand. The operator P2 satisﬁes N y2 ∞ ≤ 1 P N yN α−γ 2 1. ξj0 ) + j =1 (bj GN 2 (xi .14) we can obtain the desired results. Taking into account that GN 2 (xi . we obtain j0 N.7) by j and sum with respect to j from 1 to j0 . ξj ) j ≤ 1. So (4. .11) − N −1 Dξ G2 (xi . (4. ξk ) k =0 k = GN 1 (xi . ξj0 ) + εDξ G2 (xi .9) we have j −1 (4.13) and (4. ξj ) = k =1 G2 (xi . Also. From (4.12) N GN 1 (xi .9) +aj0 GN 2 (xi . ξj0 ) . we can easily get the nonnegativity of the Green’s function. we obtain j0 (4. ξ1 ) (4. ξj ) ≤ G2 (xi . ξj ).10) + N N Dξ G2 (xi . ξj ) = G2 (xi . and as GN 2 (xi . xi ∈ D . from (4. The proof follows directly from the representation of the solution in (4. from (4.12).(4.6). ξj ) j + cj G N 1 (xi .128 ZHONGDI CEN Proof.9)-(4. ξ1 )h1 ≥ 0. ξ1 ) = GN 2 (xi . ξj0 ) = (GN 2 (xi . ξj ) k . From Lemma 5. ξj0 ))hj0 +1 ≤ 0. ξj0 +1 ) − G2 (xi . ξj ) j ).14) GN 1 (xi . We now wish to prove the upper bound. Let the point ξj0 ∈ D N be such that ξj ∈D N N N max GN 2 (xi . 0) = 0. Proof. ξj ) j =1 j0 j + N − N = −εDξ G2 (xi . N Lemma 8.

we have one more representation ′ ψ1.15) (4. we have (4. · · · .i ≡ ψ1.i = ε(D + D − y2. For the smooth part.i .i − y1 . 2 where we have used hi exp(−αxi /ε) ≤ CN −1 for i = 1.i ).19) N N N |ψ2.20) |P2 vi − N P2 vi | xi+1 xi+1 ′′′ |v2 (t)|dt ≤ 2ε xi−1 + ai xi ′′ |v2 (t)|dt ≤ Chi+1 for i = 1.1). 2.21) ≤ Cε−2 exp(−αt/ε)dt for i = 1.N = 0. · · · .3)-(3.0 = z2.17) N N P1 zi = −P1 yi = −D + y1.5) and yi be the values ¯ N . · · · .i ). z1. N/2.18) 1 ′′ |ψ1.i | = |fi − P2 yi | ≤ |P2 vi − P2 vi | + |P2 wi − P2 wi | for i = 1. · · · . ε Recalling the decomposition of Lemma 3.i + ai D + y2.i − y2. 2. Using (2.4). · · · .2 Truncation error on a Bakhvalov-Shishkin mesh N Let yi be the solution of the discrete problem (3.i − bi y2.THIRD-ORDER CONVECTION-DIFFUSION EQUATION 129 4.0 = z2. N − 1. ξi ∈ (xi . The following estimates for the truncation error hold true: |ψ1 (xi )| = Chi+1 ε−1 exp(−αxi /ε) ≤ CN −1 for i = 0. N − 1. we have (4.3)-(3. ′′ + ′ ψ2. 1. For the truncation error of the method with respect to the layer part w we have N | P 2 wi − P 2 wi | ≤ 2 ε xi+1 xi−1 ′′′ |w2 (t)|dt + ai xi+1 xi+1 xi ′′ |w2 (t)|dt (4. For i = 0. 2.i + y2. N N P2 zi = fi − P2 yi = fi + εD + D − y2. · · · . · · · . Lemma 9. N/2 − 1. 2.i ≡ ψ2. 2. Then of the solution of the original continuous problem at the nodes of mesh D N N zi = yi − yi is the accuracy of the solution.   i C (hi+1 + hi exp(− αx ) + 1) for i = N/2.  αxi −1 −1  for i = 1. N − 1.Substituting yi = zi + yi into (3.i − y2 .i | = hi+1 |y1 (ξi )| ≤ Chi+1 ε−1 exp(−αxi /ε) ≤ CN −1 .i . We see that zi is the solution of the following problem (4.16) (4.i ) + ai (D y2. N − 1. N/2 + 1. 1. xi+1 ).i − ci y1.i = −(D + y1. We now estimate the truncation error ψi on the Bakhvalov-Shishkin mesh. · · · . N − 1 we use a Taylor expansion for x = xi to get (4. ε Proof.  C (hi+1 + N ε exp(− 2ε ) αxi−1 − 2 |ψ2 (xi )| = C (hi+1 + ε (hi + hi+1 ) exp(− ε )) for i = N/2 + 1. N − 1. xi−1 .

N/2 )] + C αxN/2−1 ) + 1).i | i + |ψ2.N/2+1 − w2. 1.i of the solution in terms of the truncation error ψ2.i − y2.i 1 = i=1 |ψ2.N/2 − bN/2 w2. we obtain N/2−1 N −1 ψ2. 1.N/2+1 | N/2+1 + i=N/2+2 |ψ2. N N |P2 wN/2 − P2 wN/2 | = |P2 wN/2 | = |εD + D − w2. ≤ CN −1 ε−1 exp(− 2ε ≤ Cε−1 N | P 2 wi − P 2 wi | ≤ Cε−2 (hi + hi+1 ) exp(−αxi−1 /ε) for i = N/2 + 1. The error of the diﬀerence scheme on the Bakhvalov-Shishkin mesh satisﬁes N yi − yi ≤ CN −1 for i = 0.26) N |y2.N/2 | 1 ≤ |ε(D + w2. we have the following a priori estimate for the accuracy N z2. where yi = max{|y1.i | i . We can now derive our main result. Theorem 1. 2. By (4. ε Using similar reasoning.N/2 ) + aN/2 (w2.i | ≤ C ψ2. Proof. ε Combining (4.5) and Lemma 8.19)-(4. 2. N/2 − 1.23) αxi−1 αxi−1 )dt ≤ CN −1 ε−1 exp(− ) 2ε 2ε ti−1 αxi ) for i = 1. |y2.i = y2 . · · · .N/2+1 − w2. N . · · · .N/2 | N/2 + |ψ2.i (4. Using Lemma 9. · · · .25) we can complete the local estimate of ψ2.N/2 − cN/2 w1. we obtain the following estimate ≤ C( −1 N/2 exp(− (4. · · · . N − 1. N − 1.N/2 + aN/2 D + w2.24) 1 N/2 ′ ′ [ε(w2 (ξN/2 ) − w2 (ξN/2−1 )) + aN/2 (w2.130 ZHONGDI CEN 2ε ϕ(t) α ε = −2 ln[1 − 2(1 − N −1 )t] and ti = ϕ−1 (xi ) for i = 1.N/2 )| + C N/2 = (4. 2.25) N 1 |P2 wN/2+1 − P2 wN/2+1 | ≤ C (1 + h− N/2+1 exp(− αxN/2 )).i 1 for i = 1. N/2.N/2 − D − w2. N Next we estimate |P2 wN/2 − P2 wN/2 |.i − y2 .i . · · · . α ti+1 ti−1 ti+1 ti+1 Let xi = Then N N | P2 ( wi − w i )| ≤ Cε−1 exp(−2ϕ(t))ϕ′ (t)dt ≤ Cε−1 exp(− exp(ϕ(t))dt ti−1 (4. · · · . N.i|.22) and (4.i|} for i = 0.

1) y (0) = 1 . ε N −1 i=N/2+2 Combining (4.27) +Cε −2 αxN/2 αxN/2−1 ) + exp(− )) ε ε (hi + hi+1 ) i exp(− αxi−1 ) ≤ CN −1 .28) and (4. We present two examples to illustrate the method described in this paper. NUMERICAL RESULTS In this section we verify experimentally the theoretically results obtained in the preceding section. Example 1. 1 − exp(−2/ε) 2(1 − exp(−2/ε)) For our tests we take ε = 10−8 which is a suﬃcient small choice to bring out the singularly perturbed nature of the problem.i | ≤ CN By (4. y ′(0) = 1 . 5. N.1) (5.2) −εy ′′′ (x) − 2y ′′ (x) = 1 . . . .i | ≤ CN From Lemma 6 we have (4.26) and (4. N.28) N −1 |y2.i | ≤ C |ψ1. · · · .THIRD-ORDER CONVECTION-DIFFUSION EQUATION N/2−1 N −1 131 ≤ C( i=1 hi+1 i + N/2 + N/2+1 + i=N/2+2 hi+1 i ) N/2−1 +CN −1 −1 ε exp(− i=1 αxi ) 2ε i +C (exp(− (4. y ′(1) = 1. · · · .29) we get the desired results. 1. for x ∈ (0.5) y (x) = ε x+ 2 exp(−2x/ε) ε + x − x2 /2 + 1 − . Consider the boundary value problem (5.27) we get (4.4) −εy ′′′(x) − 2(1 + x)y ′′ (x) + 4xy ′(x) − y (x) = f (x). 1) y (0) = 1 . y ′(1) = 1. Consider the boundary value problem (5.i − y1 for i = 0. where f (x) is chosen such that (5.29) N −1 |y1.3) (5. 1. y ′(0) = 1 . We measure the accuracy in the discrete maximum norm eN = y − yN ∞. Example 2. x ∈ (0.i − y2 for i = 0.

Valarmathi. [2] E. Appl.3275e-3 2.995 2. [3] S. From Tables 1 and 2 we see that eN /e2N is close to 2.K error 6..500 Table 2. which supports the convergence estimate of Theorem 1. Numer. Niederdrenk.6480e-2 2.6317e-3 4.496 0. Comput. Monotone diﬀerence scheme for Example 2 N 32 64 128 256 512 1024 the convergence rate eN r = log2 ( 2N ) e and the constants in the error estimate N.998 1. 41: 223-253.132 ZHONGDI CEN Table 1. 2002. They indicate that the theoretical results are fairly sharp.. 129: 345-373. C.160 0.959 2.8520e-3 2. A computational methods for third order singularly perturbed ordinary diﬀerntial equations. N.000 1.4644e-3 rate constant 0.487 0. 51: 631-657.1690e-2 5. Yserentant.498 1. Graded-mesh diﬀerence schemes for singularly perturbed two-point boundary value problems. Math.995 1.216 2.100 0. Comput. Math.978 2.996 2.493 0. REFERENCES [1] K. The uniform stability of singularly perturbed discrete and continuous boundary value problems.989 2.210 0. H. Y607504) and Ningbo Natural Science Foundation (2009A610082) of China.5625e-2 3.1717e-3 rate constant 0.Ramanujam. Monotone diﬀerence scheme for Example 1 N 32 64 128 256 512 1024 error 4.7129e-2 8. Gartlend. . I would like to thank the anonymous referees for several suggestions for the improvement of this paper. 1983.224 C N = eN /N −1 .3328e-2 1.3750e-2 1.999 1.. 1988.193 0.999 1.499 1. Math. Acknowledgement. The work was supported by Zhejiang Province Natural Science Foundation of China (Grant No.9278e-3 1.

G. L. Shishkin.1996. E. 1994. Semper.. Springer. [7] H. J. Roos. World Scientiﬁc. O’Riordan. T. I. Stynes.-G. [5] H. Miller.THIRD-ORDER CONVECTION-DIFFUSION EQUATION 133 [4] B. Numer. Linß. Suﬃcient conditions for uniform convergence on layer-adapted grids . [6] J. Berlin. H. M. 14: 97-109. Locking in ﬁnite element approximation of long thin extensible beams. Computing.Error estimates in the maximum norm for linear problems in one and two dimensions. Roos. Singapore.-G. Tobiska. Fitted numerical methods for singular perturbation problems.1996. Anal. 63: 27-45. 1999. . IMA J. Numerical methods for singularly perturbed diﬀerntial equations convection-diﬀusion and ﬂow problems.