Advances in Engineering Research, volume 170

7th International Conference on Energy and Environmental Protection (ICEEP 2018)

Implicit Finite Difference Method for Fractional Wave Equation with

Dirichlet and Fractional Boundary Conditions
Shaomei Fang1,2, a, Wenjie Huang1,b, Changping Xie1,c,*, Yingshu Zhang1,d ,
Jinyan Li1,e and Zhenfu Cai1,f
1
College of Mathematics and Informatics, South China Agricultural University, Guangzhou 510642,
People's Republic of China
2
School of Mathematics and statistics, Yangtze Normal University, Chongqing 408100, People's
Republic of China
a
dz90@scau.edu.cn, b2753751501@qq.com, cchangping@stu.scau.edu.cn, d469867110@qq.com,
e
duduwoxing@126.com, f 229430694@qq.com
*Corresponding author

Keywords: Fractional wave equation; Riemann-Liouville fractional derivative; Fractional boundary

conditions; Finite difference method; Consistency and stability

Abstract. In this paper, the implicit finite difference method is developed for the fractional wave
equation with Dirichlet and fractional boundary conditions. The consistency and stability of the
method are strictly proved by the Gerschgorin theorem and mathematical induction. Numerical
examples show the accuracy and efficiency of the scheme and coincide with the theoretical analysis.

Introduction
Fractional differential equations are generalizations of integer-order differential equations, which
have been used in many models, such as fractional diffusion equations describing anomalous
diffusion [1,2] and continuous-time finance problems [3], fractional Fokker-Planck equations
describing Lévy stable processes [4,5] and porous media [6], and fractional wave equations
describing chiral media [7-9], waveguides [10], reflection and scattering problems [11]. As most of
fractional differential equations do not have explicit analytical solutions, research on numerical
methods of fractional differential equations become critical.
In this paper, we consider the fractional wave equation with the the left-sided fractional spatial
derivative as follows

∂ 2u ( x , t ) ∂α u ( x , t )
= c ( x , t ) + d ( x, t ), 0 < x < L,0 < t < T , (1)
∂t 2 ∂ + xα

subject to the boundary and initial conditions

 ∂α −1u( x, t ) 
u(0, t ) = 0, γ u ( L, t ) +  c ( x , t )  = v (t ), 0 < t < T , (2)
 ∂ + xα −1  x = L

∂u( x, t )
u( x, 0) = p 0 ( x ), = p1 ( x ), 0 < x < L, (3)
∂t t =0

where the parameter α describes the fractional order of spatial derivatives with 1 < α < 2 .
Coefficient function c( x, t ) is positive, d ( x, t ) refers to a source term, γ = 0 corresponds to a
fractional Neumann boundary condition, and γ > 0 corresponds to a fractional Robin boundary
condition.

Copyright © 2018, the Authors. Published by Atlantis Press.

This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). 1
 Advances in Engineering Research, volume 170

The space fractional derivatives are defined as the Riemann-Liouville fractional derivative [12-14]
∂ β u( x ) 1 dn x u( s)
∂+ xβ
=
Γ( n − β ) dx n ∫
0 ( x − s) β +1−n
ds, (4)

where n is an integer, and n − 1 < β ≤ n . When β is an integer, (4) gives the standard integer
derivative.
Recently, Jia and Wang [15] developed a fast finite difference method for space-fractional
diffusion equations with fractional derivative boundary conditions in one space dimension. Guo et al.
[16] processed an implicit finite difference method for a one-dimensional fractional percolation
equation with the Dirichlet and fractional boundary conditions. To our knowledge, implict finite
difference method for fractional wave equation with the Dirichlet and fractional boundary conditions
is still limited. This motivate us to examine a numerical approach for it.
The rest of the paper is organized as follows. In Section 2, we construct an implicit finite difference
scheme and study its consistency. The stabilty of the scheme is proved in Section 3. In Section 4, we
carry out numerical examples to check the accuracy and efficiency of the proposed scheme. Finally,
we draw our conclusions in Section 5.

Implicit finite difference method and its consistency

In this section, we construct an implicit finite difference scheme and study the consistency for (1-3).
For the numerical approximation scheme, Let tm = m∆t ( m = 0,1,2,L , M ) , xi = ih(i = 0,1, 2,L , N )

to be the temporal partition and the spatial partition, respectively, where M and N are positive
integers. Then time grid size is ∆t = T / M and space grid size is h = L / N . Denote the exact and
numerical solution at the mesh point ( xi , tm ) by U im and uim , respectively. Define

cim = c ( xi , tm ) , d im = d ( xi , tm ) , v m = v ( tm ) , pi0 = p0 ( xi ) , pi1 = p1 ( xi ) ,

and initial and the Dirichlet boundary conditions

ui0 = pi0 , u0m = 0.

According to [17], the left-sided fractional spatial derivative can be approximated by the shifted left
Grünwald-Letnikov derivative

∂α u ( x, t ) 1 i

∂ + xα ( x ,t
=
hα
∑g
j =0
(α )
j U im−+j1+1 + O ( h ), (5)
i m +1 )

α 
where 1 ≤ i ≤ N − 1,0 ≤ m ≤ M − 1 , and g (jα ) = ( −1) j   are the fractional binomial coefficients.
 j
To approximate (1) we use the backward Euler difference scheme for the second order time derivative

∂ 2u ( x , t ) U im +1 − 2U im + U im −1
= + O ( ∆t ), (6)
∂t 2 ( xi ,tm +1 ) （∆t） 2

where 1 ≤ i ≤ N − 1,1 ≤ m ≤ M − 1 .
And the left-sided fractional spatial derivative of the order α − 1 in (2) can be discretized by the
standard left Grünwald-Letnikov fractional derivative [17] as

2
 Advances in Engineering Research, volume 170

∂α −1u ( x, t ) 1 N −1

∂ + xα −1 ( L,t
=
h α −1 ∑g
j =0
(α −1)
j U Nm−+1j + O ( h ), (7)
m +1 )

where 0 ≤ m ≤ M − 1 .
The first order time derivative in (3) can be approximated by the forward Euler difference
scheme
∂u( x, t ) U 1 − U i0
= i + O ( ∆t ), (8)
∂t ( xi ,0) ∆t

where 1 ≤ i ≤ N − 1 .
Therefore, an implicit finite difference scheme can be expressed as follows:
For m = 0 ,
ui1 − ui0
pi1 = ,1 ≤ i ≤ N − 1, (9)
∆t
c1N N −1 (α −1) 1
α −1 ∑ j
u01 = 0, γ u1N + g u N − j = v1 , (10)
h j =0

For 1 ≤ m ≤ M − 1 ,
uim +1 − 2uim + uim −1 cim +1 i (α ) m +1
= α ∑ g j ui − j +1 + d im +1 ,1 ≤ i ≤ N − 1, (11)
（∆t） 2
h j =0

cNm +1 N −1 (α −1) m+1

α −1 ∑ j
u0m +1 = 0, γ u Nm +1 + g uN − j = v m+1 , (12)
h j =0

Denote the local truncation error by Rim for 1 ≤ i ≤ N . It follows from (9)-(12) that
U i1 − U i0
Ri1 = pi1 − = O ( ∆t ),1 ≤ i ≤ N − 1, (13)
∆t
U m +1 − 2U im + U im −1 cim +1 i (α ) m +1
Rim +1 = i − α ∑ g j U i − j +1 − d im +1 = O ( ∆t ) + O ( h ),1 ≤ i ≤ N − 1,1 ≤ m ≤ M − 1,
（∆t）2 h j =0
(14)
m +1 N −1
c
RNm +1 = γ U Nm +1 + Nα −1 ∑ g (jα −1)U Nm−+1j − v m+1 − O ( h ),0 ≤ m ≤ M − 1. (15)
h j =0
This implies the consistency of the implicit finite difference scheme difined by (9)-(12).

Stability analysis
In this section, we discuss the stability and convergence of the numerical method (9)-(12).
（∆t）2
Let r = α . One can rewrite (11) in the form:
h
i
uim +1 − rcim +1 ∑ gi(−αj)+1u mj +1 = 2uim − uim −1 +（∆t）2d im +1 ,1 ≤ i ≤ N − 1, (16)
j =0

where 1 ≤ m ≤ M − 1 . Then the implicit finite difference schemes (9)-(12) can be presented as the
matrix form as follows:
A(1)U 1 = P1 + P 0 , (17)

3
 Advances in Engineering Research, volume 170

A( 2)U m +1 = 2Q m − Q m −1 + D m +1 ,1 ≤ m ≤ M − 1, (18)
where
U m = ( u1m , u1m ,L , uNm ) ,
T

P1 = ( ∆tp11 , ∆tp11 ,L , ∆tp1N −1 ,0 ) ,

T

P 0 = ( p10 , p10 ,L , pN0 −1 , hα −1v1 ) ,

T

Q m = ( u1m , u1m ,L , uNm −1,0 ) ,

T

D m = ( ( ∆t )2 d1m ,( ∆t )2 d1m ,L ,( ∆t )2 d Nm−1 , hα −1v m ) ,

T

and the coefficient matrix A(1) , A( 2) and their entries are

 1, 1 ≤ i = j ≤ N − 1,

0, i ≠ j ,1 ≤ j ≤ N ,1 ≤ i ≤ N − 1,
A(1) = ( Ai(1)
, j )i , j =1
N
=  1 (α −1) (19)
 cN g N − j , 1 ≤ j ≤ N − 1, i = N ,
 hα −1γ + c1N g0(α −1) , i = j = N .

 − rcim +1 gi(−αj)+1 , 1 ≤ j < i ,1 ≤ i ≤ N − 1,

 m +1 (α )
1 − rci g1 , 1 ≤ j = i ≤ N − 1,
 − rc m +1 g (α ) , j = i + 1,1 ≤ i ≤ N − 1,
= ( Ai , j )i , j =1 = 
(2) N i 0
A( 2) (20)
0, i + 2 ≤ j ≤ N ,1 ≤ i ≤ N − 2,
 c m +1 g (α −1) , 1 ≤ j ≤ N − 1, i = N ,
 αN−1 N − j m+1 (α −1)
 h γ + cN g0 , i = i = N .

In the following, we present some lemmas which are used later in this paper. First, we list some
properties of the alternating fractional binomial coefficients [14].

Lemma 1. Let β , β1 , and β 2 be positive real numbers, and the integer n ≥ 1 . We have
 β + 1  (β )
(i) g 0( β ) = 1, g (jβ ) =  1 − g j −1 for j ≥ 1 ;
 j 
n
(ii) g1( β ) < g 2( β ) < L < 0, ∑ g (jβ ) > 0 for 0 < β < 1 ;
j =0
n
(iii) g 2( β ) > g3( β ) > L > 0, ∑ g (jβ ) < 0 for 1 < β < 2 ;
j =0

 β − 1
n
n
(iv) ∑g (β )
j = ( −1)   ;
j =0  n 
n
(v) ∑g
j =0
( β1 )
j g n( β−2j) = g n( β1 + β2 ) .

Before giving the Lemma 2, we consider a equivalent form of the the finite difference scheme (18).
Let us express the coefficient matrix A( 2) defined by (20) in a block form as follows
 AN −1, N −1 AN −1, N 
A( 2) =  T
AN(2), N 
, (21)
 AN , N −1

4
 Advances in Engineering Research, volume 170

where

AN −1, N −1 = ( A(2) )i , j =1 , AN −1, N = ( Ai(2)

, N )i =1 , AN , N −1 = ( AN , j ) j =1 .
N −1 N −1T (2) N −1

rcNm +−11
Denote ωN = . Multiply the N th equation by ωN and add to the N − 1 th equation of
cNm +1 + hα −1γ

the linear system (18). Then we get the following equivalent form of (18)
A(3)U m +1 = 2Q m − Q m −1 + D m +1 ,1 ≤ m ≤ M − 1, (22)
where

D m = ( ( ∆t )2 d1m ,( ∆t )2 d1m ,L ,( ∆t )2 d Nm−1 + hα −1v m , hα −1v m ) ,

T
(23)

A 0 
A(3) =  NT−1, N −1 ( 2) 
, (24)
 AN , N −1 AN , N 
AN −1, N −1 − I − rB , (25)
 cim +1 gi(−αj)+1 , 1 ≤ j < i ,1 ≤ i ≤ N − 2,
 m +1 (α )
 ci g1 , 1 ≤ j = i ≤ N − 2,
 m +1 (α )
B = ( Bi , j )i , j =1 =  ci g0 , j = i + 1,1 ≤ i ≤ N − 2,
N −1
(26)
 0, i + 2 ≤ j ≤ N ,1 ≤ i ≤ N − 3,
 m +1 m +1 (α −1)
 cN −1cN ( g N − j − g N(α−)j ) + hα −1γ g N(α−)j , 1 ≤ j ≤ N − 1, i = N − 1,

Lemma 2. The spectral radius of the matrix ( AN −1, N −1 )

−1
is less than 1.
Proof Our first goal is to show that the eigenvalues of the matrix B have negative real parts. As
cim +1 > 0 , and then by Lemma 1, we have

Bi , j > 0, j ≠ i , (27)

Bi , j < 0, j = i , (28)

where 1 ≤ i ≤ N − 1 .
When 1 ≤ i ≤ N − 2 , we have
N −1
 i −1
 i
Bi ,i + ∑ Bi , j = cim +1  g1(α ) + ∑ gi(−α j)+1 + g0(α )  = cim +1 ∑ g (jα ) < 0, (29)
j =1, j ≠i  j =1  j =0

Similarly, when i = N − 1 , we obtain

N −2
BN −1, N −1 + ∑ BN −1, j < 0, (30)
j =1

Then by Gerschgorin theorem [18], we have that all of the Greschgorin disks of the matrix B are
within the left half of the complex plane.
Next, from (27), we can clear know that the eigenvalues of the matrix AN −1, N −1 has a magnitude
larger than 1. Therefore, the spectral radius of the matrix ( AN −1, N −1 )
−1
is less than 1. This completes
the proof.
To discuss the stability of the numerical method, we denote

5
 Advances in Engineering Research, volume 170

ε im = uim − u%im ,1 ≤ i ≤ N ,1 ≤ m ≤ M , (31)

and

ε m = (ε1m , ε 2m ,L , ε Nm ) , ε m
T
= max ε im , (32)
∞ 1≤i ≤ N

where u%im is the approximate solution of the difference scheme with the initial condition u%i0 .
From the definition of the finite difference scheme, we have:
For 1 ≤ i ≤ N − 1 ,
ε i1 = ε i0 (33)
i
ε im +1 − rcim +1 ∑ gi(−α j)+1ε mj +1 = 2ε im − ε im −1 ,1 ≤ m ≤ M − 1. (34)
j =0

For i = N ,
N −1
( hα −1γ + cNm+1g0(α −1) ) ε Nm+1 + cNm+1 ∑ g (jα −1)ε Nm−+1j = 0,0 ≤ m ≤ M − 1, (35)
j =1

where ε i0 = 0(m = 0,1,L M ) .

For the stability of the numerical method, we have the following theorem.
Proposition 1. There exists a positive constant independent of ∆t and h such that

εm ≤ C ε0 ,1 ≤ m ≤ M . (36)
∞ ∞

Proof Let

ζ m = ( ε1m , ε 2m ,L, ε Nm−1 ) ,0 ≤ m ≤ M .

T

Then (33)-(35) can be rewritten as

ζ i1 = ζ i0 (37)

AN −1, N −1ζ m +1 = 2ζ m − ζ m −1 ,1 ≤ m ≤ M − 1. (38)

For i = N ,
ANT , N −1ζ m+1 + AN( 2), N ε Nm+1 = 0,0 ≤ m ≤ M − 1. (39)

It follows from (39) that

N −1 N −1
cNm +1 ∑ g (jα −1)ε Nm−+1j cNm +1 ∑ g (jα −1) ε Nm−+1j
j =1 j =1
ε Nm +1 = α −1 m +1 (α −1)
≤ α −1 m +1 (α −1)
< max ε im +1 . (40)
h γ +c N g 0 h γ +c N g 0
1≤i ≤ N −1

This means that

ε m +1 = max ε im +1 = ζ m +1 . (41)
∞ 1≤i ≤ N ∞

Then (37) implies that

ε1 = ζ1 = ζ0 = ε0 . (42)
∞ ∞ ∞ ∞

6
 Advances in Engineering Research, volume 170

Suppose that ε k ≤ C ε0 , k = 2,3,L, m, let ε lm +1 = max ε im +1 .

∞ ∞ 1≤i ≤ N

Then by lemma 2, we have

ε m +1 = ε lm +1 ≤ 2 ζ m + ζ m −1 ≤ 2 ζ m + ζ m −1 ≤ 3C ζ 0 = 3C ε 0 = C' ε0 . (43)
∞ ∞ ∞ ∞ ∞ ∞

Using mathematical induction, we complete the proof.

The following theorem can be obtained by above proposition [19-20].
Theorem 1. The implicit finite difference (17)-(18) is unconditional stable.

Numerical example
In this section, we consider the numerical solution of the fractional wave equation using the
proposed scheme (9)-(12).
Example. We consider the following fractional wave equation

∂ 2u ( x , t ) ∂α u ( x, t )
= c ( x , t ) + d ( x, t ), 0 < x < 1,0 < t < 1,
∂t 2 ∂ + xα

with the fractional boundary conditions

 ∂α −1u( x, t )   6 
u(0, t ) = 0, γ u (1, t ) +  c ( x , t ) α −1  = v (t ) = exp(−t )  3γ +  ,
 ∂+ x  x =1  3−α 

and initial conditions

∂u( x, t )
u( x, 0) = p 0 ( x ) = 3x 2 , = p1 ( x ) = −3x 2 ,
∂t t =0

where c( x, t ) = Τ(3 − α ) xα , d ( x, t ) = −3x 2 exp(−t ).

The exact solution of this problem is given by
u( x, t ) = 3x 2 exp( −t ).
A comparison between the exact and the numerical solutions using the proposed method with
∆t = h = 1 / 40 at T = 1 is presented in Fig. 1. It can be seen that our numerical results are in excellent
agreement with the exact solution both for the 145 fractional Neumann boundary condition ( γ = 0 )
and Robin boundary condition ( γ = 1 ).
Table 1 shows the maximum errors at the time T = 1 , where h = 10 −4 , a value small enough such
that the space discretization errors are negligible while comparing with the time errors. It illustrates
that the numerical method is temporally first-order accurate both for the fractional Neumann and
Robin boundary condition. In Tables 2, with sufficiently small time step size ∆t = 10 −4 , the
convergence order in space of our schemes is also checked as first-order accurate. This implys the
stability of the implicit finite difference scheme (9)-(12) proved in Theorem 1 and shows that the
scheme are temporally and spatially first-order accurate.

7
 Advances in Engineering Research, volume 170

(a) (b)
Fig.1 The comparisons of numerical solution and exact solution for α = 1.9, ∆t = h = 1 / 40 at the
time T = 1 (a) γ = 0 (b) γ = 1 .

Table 1: Numerical errors and convergence orders in time direction with α = 1.9, h = 10−4
γ =0 γ =1
∆t Error Order Error Order
−4 −2 −2
2 8.4209e - 4.7898e -
−5 −2 −2
2 4.2197e 0.9978 2.4132e 0.9924
−6 −2 −2
2 2.1129e 0.9986 1.2134e 0.9944
−7 −2 −3
2 1.0525e 1.0037 6.0529e 1.0023

Table 2: Numerical errors and convergence orders in time direction with α = 1.9, ∆t = 10−4
γ =0 γ =1
h Error Order Error Order
−4 −2 −2
2 6.1262e - 4.7052e -
−5 −2 −2
2 3.1962e 0.9583 2.4737e 0.9510
−2
2 −6
1.6288e 0.9811 1.2673e −2 0.9759
2 −7 8.1761e −2 0.9961 6.3904e −3 0.9916

Conclusion
In this paper, an implicit finite difference method is developed for the one-dimensional space
fractional wave equation with the Dirichlet and fractional boundary conditions. The unconditional
stability of the method are proved with the Gerschgorin theorem and mathematical induction. The
numerical experiment confirms the theoretical analysis and illustrates the practicability of the
numerical scheme. Higher-order methods for the fractional wave equation will be considered in our
future work.

Acknowledgements
This work was financially supported by the National Natural Science Foundation of China
(11271141).

References
[1] R. Metzler and J. Klafter: Phys. Rep., Vol.339(1)(2000),p.1-77.

8
 Advances in Engineering Research, volume 170

[2] R. Metzler and J. Klafter: J. Phys. A: Math. Gen., Vol.37(31)(2004),p.R161-R208.

[3] F. Mainardi, M. Raberto, R. Gorenflo and E. Scalas: Physica A., Vol.287(2000),p.468-481.
[4] B.J. West, P. Grigolini and R. Metzler, and T.F. Nonnenmacher: Phys. Rev. E.,
Vol.55(1)(1997),p.99-106.
[5] A.V. Chechkin, V.Y. Gonchar, J. Klafter and R. Metzler: Advances in Chemical Physics, Vol.133
(B)(2006),p.439-496.
[6] A. De Pablo, F. Quirós, A. Rodriguez and J.L.Vázquez: Adv. Math., Vol.226(2)(2011),p.
1378-1409.
[7] A.Hussain and Q.A. Naqvi : Progress In Electromagnetic Research, Vol. 59(2006),p.199-213.
[8] Q.A. Naqvi, G. Murtaza and A.A. Rizvi: Optics Communications, Vol.178 (2000),p.27-30.
[9] A. Lakhtakia: Microwave Opt. Tech. Lett., 28(2001),p.385-386.
[10] A. Hussain, S.Ishfaq and Q.A.: Progress In Electro magnetics Research,
Vol.63(2006),p.319-335.
[11] M.V. Ivakhnychenko, E.I. Veliev and T.V.: Journal of Electromagnetic Waves and Applications,
Vol.21(13)(2007),p.1787-1802.
[12] K. Miller and B. Ross: An Introduction to the Fractional Calculus and Fractional Differential
Equations( Wiley, New York, 1993).
[13] I. Podlubny: Fractional Differential Equations(Academic, 1999).
[14] S. Samko, A. Kilbas and O. Marichev: Fractional Integrals and Derivatives: Theory and
Applications(Gordon and Breach, 1993).
[15] J. Jia, H.Wang: Journal of Computational Physics, Vol.293(2015),p.359-369.
[16] B.L. Guo, Q. Xu and Z. Yin: Applied Mathematics and Mechanics(English Edition),
Vol.37(3)(2016),p.403-416.
[17] M.M. Meerschaert and C. Tadjeran: Journal of Computational and Applied Mathematics,
Vol.172(2004),p.65-77.
[18] SG Samko, AA Kilbas, OI Marichev: Fractional Integrals and Derivatives: Theory and
Applications (Gordon Breach, Yverdon, 1993).
[19] E. Isaacson and H.B. Keller: Analysis of Numerical Methods(Wiley, New York,1966).
[20] G.D. Smith: Numerical solution of partial differetial equations: Finite difference
methods(Clarendon press, Oxford, 1985).
[21] D. Yu and H.Tan: Numerical methods of differetial equations(Science Publisher, Beijing, 2003.)
You and M. Kaveh: IEEE Trans. Image Process. Vol. 9(2000), p. 1723-1730.