Professional Documents
Culture Documents
1 School of Mathematical Sciences, Queensland University of Technology, GPO BOX 2434, Brisbane, QLD. 4001, Australia
Abstract: We present a finite volume method to solve the time-space two-sided fractional advection-dispersion equa-
tion on a one-dimensional domain. The spatial discretisation employs fractionally-shifted Grünwald formu-
las to discretise the Riemann-Liouville fractional derivatives at control volume faces in terms of function
values at the nodes. We demonstrate how the finite volume formulation provides a natural, convenient and
accurate means of discretising this equation in conservative form, compared to using a conventional finite
difference approach. Results of numerical experiments are presented to demonstrate the effectiveness of
the approach.
PACS (2008): 01.30.-y, 01.30.Xx, 01.30.Tt
Keywords: two-sided time-space fractional advection-dispersion • fractional Fick’s law • finite volume • finite difference
• shifted Grünwald
© Versita sp. z o.o.
1275
on the interval [a, b], subject to homogeneous Dirichlet rally to discretisation in a finite difference sense. A crucial
boundary conditions. In the context of particle transport, observation was the necessity of using shifted formulas to
C (x, t) represents the concentration of particles at position ensure stability of these numerical schemes [7].
γ
x and time t. The symbol Dt C (x, t) represents the Caputo
Many of these methods have since been generalised and
fractional derivative of order γ.
extended in various ways, leading to non-standard finite
difference schemes [10–12], finite difference schemes for
Definition 1 (Caputo fractional derivative on problems of variable fractional order [13–16] and “fast”
[0, ∞) [3]). finite difference schemes [17–21] to name a few.
Other numerical methods have received less attention in
Rt ∂C (x,η)
the literature to date. The literature on finite volume
1
∂η
(t − η)−γ dη, 0 < γ < 1,
γ
Γ(1−γ) 0 methods for fractional partial differential equations in par-
Dt C (x, t) =
∂
ticular is still in early stages of development. A finite vol-
∂t
C (x, t), γ = 1.
ume method for the space fractional advection-dispersion
equation was proposed by Zhang et al. [22, 23], who chose
The symbols ∂α C /∂x α and ∂α C /∂(−x)α represent the left to discretise the Riemann-Liouville derivatives directly,
and right Riemann-Liouville fractional derivatives of or- rather than use the Grünwald-Letnikov definition.
der α.
Previously we have considered a finite volume method
for the two-sided space-fractional advection-dispersion
Definition 2 (Riemann-Liouville fractional equation with constant coefficients [24], based on the
derivatives on [a, b] [3]). Grünwald-Letnikov definition, where we proved the sta-
bility and convergence of the method. In this paper, we
x
extend the method to solve the two-sided time-space frac-
∂α C (x, t) ∂n
Z
1 C (ξ, t)
α
dξ, tional advection-dispersion equation with variable coeffi-
Γ(n − α) ∂x n (x − ξ)α+1−n
= (2)
∂x a cients. Additionally we derive a new result concerning the
b
∂α C (x, t) (−1)n ∂n
Z
C (ξ, t)
α
dξ, relationship between the finite difference method and the
Γ(n − α) ∂x n (ξ − x)α+1−n
= (3)
∂(−x) x finite volume method in the constant coefficient case. We
also demonstrate the improvement in accuracy provided by
where n is the smallest integer greater than or equal to α. the finite volume method over the finite difference method
for a given test problem.
In this paper we consider regimes where 0 < γ ≤ 1 and The remainder of the paper is organised as follows. In
0 < α ≤ 1. The inclusion of the skewness β ∈ [0, 1] in Section 2 we derive equation (1) from conservation princi-
equation (1) allows for modelling of regimes where the ples. In Section 3 we derive the new finite volume method
forward and backward jump probabilities are different [4]. for (1) and also make comparisons between it and the fi-
The remaining components of equation (1) are the velocity nite difference method in the constant coefficient case. In
V (x, t), the anomalous dispersion coefficient K (x, t) > 0 Section 4 we illustrate the method’s performance on test
and the source term S(x, t). problems. Finally, we draw our conclusions in Section 5.
The literature on numerical methods for solving fractional
differential equations is not as well-developed as for
integer-order equations. Finite difference methods have 2. Derivation
dominated the literature. The L1 scheme [5] and varia-
tions thereof have formed the basis for many finite differ-
ence discretisations of Caputo time fractional derivatives, The derivation of the advection-dispersion equation,
such as the widely used method of Lin and Xu [6]. whether standard or fractional, begins with the law of mass
In space too, finite difference methods have proved very conservation which, in conservative form, is
popular. The pioneering work of Meerschaert et al. [4, 7–
∂C ∂Q
9] laid the groundwork for many of the popular finite =− + S(x, t) (4)
difference discretisations for space-fractional derivatives. ∂t ∂x
These are based on the equivalence of Riemann-Liouville
and Grünwald-Letnikov fractional derivatives (for suffi- where Q is the flux. The flux comprises two components:
ciently smooth functions). The Grünwald-Letnikov defini-
tion is preferred numerically because it lends itself natu- Q(x, t) = V (x, t)C (x, t) + q(x, t), (5)
1276
the advective component is V (x, t)C (x, t) and the disper- Applying Definition 2, the first term on the right of (8)
sive component is denoted q(x, t). For standard diffusion may be written as
or dispersion, where Fick’s law applies, the form of the
dispersive flux q(x, t) is
" Z x
∂ 1 ∂ C (ξ, t)
K β dξ
∂x Γ(1 − α) ∂x a (x − ξ)α
∂C
q(x, t) = −K (x, t) (−1) ∂
Z b
C (ξ, t)
#
∂x − (1 − β) dξ .
Γ(1 − α) ∂x x (ξ − x)α
where K is the dispersion coefficient. Substituting this
form of q into (5) and returning to (4), the standard Combining the derivatives and letting ν = α +1, we obtain
advection-diffusion equation
" Z x
1 ∂2 C (ξ, t)
∂C ∂ ∂ ∂C K β dξ
+ (V (x, t)C (x, t)) = K (x, t) + S(x, t) (6) Γ(2 − ν) ∂x 2 (x − ξ)ν−1
a
∂t ∂x ∂x ∂x #
Z b
(−1)2 ∂2 C (ξ, t)
+ (1 − β) dξ ,
is derived. In the special case of constant velocity and Γ(2 − ν) ∂x 2 x (ξ − x)ν−1
dispersion coefficient, we may write (6) as
which, again with the help of Definition 2, is simply
∂C ∂C ∂2 C
+V = K 2 + S(x, t) . (7)
∂t ∂x ∂x ν
∂ C ∂ν C
K β ν + (1 − β) .
∂x ∂(−x)ν
2.1. Space-fractional derivatives
Hence we have the space-fractional advection-dispersion
equation with constant coefficients
Fick’s law is equivalent to the assumption of Brownian
ν
∂ν C
motion at the particle scale [1]. In complex heterogeneous
∂C ∂C ∂ C
media this assumption may not be reasonable, and instead +V = K β ν + (1 − β) + S(x, t) . (9)
∂t ∂x ∂x ∂(−x)ν
an alternative law that allows for diverging jump length
variances (“long jumps”) and unequal forward and back-
ward jump probabilities at the particle scale is required. In the integer limit ν = 2 (α = 1) both fractional
The fractional Fick’s law [25] derivatives recover the standard second-order derivative
∂2 C /∂x 2 , and hence we see that (for any value of β) (9)
α
∂α C (x, t)
∂ C (x, t)
does indeed reduce to (7) in this case.
q(x, t) = −K (x, t) β − − β)
∂x α ∂(−x)α
(1
1277
α
∂ C (x, t) ∂α C (x, t)
For sufficiently smooth functions C , the Riemann-Liouville
q(x, t) = −K (x, t) β − − β) .
∂x α ∂(−x)α
(1 derivatives (2), (3) and the Grünwald-Letnikov derivatives
(11) (15), (16) coincide [3, p.199]. The standard Grünwald for-
Performing the integration on the first term on the right, mulas are obtained from (15) and (16) by replacing ∆x
and interchanging the order of integration and differenti- with the spatial step h, yielding finite sum approxima-
ation on the other terms, we have tions.
dγ xi+1/2
Z
C dx = − Q|xi+1/2 − Q|xi−1/2
Definition 4 (Grünwald formulas on [a, b] [3]).
dt γ xi−1/2
Z xi+1/2 [ x−a
h ]
∂α C (x, t)
S(x, t) dx . 1 X j α
+ (12) ≈ C (x − jh, t) ,
∂x α hα j=0
(−1) (17)
xi−1/2 j
[ b−x
h ]
∂α C (x, t)
j α
This leads to the finite volume discretisation 1 X
≈ C (x + jh, t) .
∂(−x)α hα j=0
(−1) (18)
j
γ
d C̄i 1
γ
= Q|xi−1/2 − Q|xi+1/2 + S̄i (13)
dt h
The shifted Grünwald formulas modify these equations by
where barred quantities denote control volume averages introducing a shift value.
R xi+1/2
f¯i = 1/h xi−1/2 f dx.
dγ Ci 1 [ x−a
h +p]
∂α C (x, t)
Q|xi−1/2 − Q|xi+1/2 + S(xi , t) X α
dt γ (−1)j
= (14) 1
h ≈ α C (x − (j − p)h, t) , (19)
∂x α h j=0
j
where the introduction of this approximation means that
[ b−x
h +p]
∂α C (x, t)
this is now an equation for the approximate numerical X α
(−1)j
1
solution Ci (t) ≈ C (xi , t) at node xi . ≈ α C (x + (j − p)h, t) , (20)
∂(−x)α h j=0
j
To complete the spatial discretisation we require approx-
imations for the total flux Q|xi±1/2 at the control volume where p is the shift value.
faces. We derive these approximations by utilising the
link between the Riemann-Liouville fractional derivative
If we define weights
and the Grünwald-Letnikov fractional derivatives. Simi-
larly to the finite difference method of Meerschaert et al.
α(α − 1) . . . (α − j + 1)
[7], the Grünwald-Letnikov derivatives must be shifted. w0α = 1, wα,j = (−1)j , j = 1, 2, . . .
j!
We first introduce the standard (non-shifted) definition for (21)
the Grünwald-Letnikov fractional derivatives. then we may write (19) and (20) more simply as
[ x−a
h +p]
Definition 3 (Grünwald-Letnikov fractional ∂α C (x, t) X
wjα C (x − (j − p)h, t) ,
1
≈ α
∂x α
(22)
derivatives on [a, b] [3]). h j=0
1278
[ b−x
h +p]
∂α C (x, t) X Previously [24] we have shown this discretisation to be
wjα C (x + (j − p)h, t) .
1
≈ α
∂(−x)α
(23)
h
of first order spatial accuracy for the constant coefficient
j=0 case.
The inclusion of the shift p in the formulas was origi- For the advective flux V (x, t)C (x, t) we use a standard
nally motivated by the fact that the standard (non-shifted) averaging scheme
Grünwald formulas lead to unstable methods for certain
V (xi±1/2 , t)
V (xi±1/2 , t)C (xi±1/2 , t) ≈ [C (xi , t) + C (xi±1 , t) ] .
finite difference discretisations [7, 9, 26, 27], whereas the
shifted Grünwald formulas lead to stable methods. 2
In the present method, the fractional shift p = 1/2 is used, (30)
allowing us to build approximations of fractional deriva-
tives at control volume faces xi±1/2 in terms of function val-
ues at the nodes xj [24]. Hence we derive the fractionally- This completes the spatial discretisation. For the temporal
shifted Grünwald formulas discretisation, we define a temporal partition tn = nτ for
n = 0, 1, . . . where τ is the timestep, and discretise the
i
∂α C (xi−1/2 , t) 1 X α Caputo time fractional derivative using the L1-algorithm
≈ α w C (xi−j , t) ,
α h j=0 j
(24)
∂x (see [6])
N−i+1
∂α C (xi−1/2 , t) 1 X α dγ Ci (tn+1 )
≈ w C (xi+j−1 , t)
∂(−x)α hα j=0 j
(25)
dt γ
n
τ −γ X γ
b [Ci (tn+1−j ) − Ci (tn−j )] + O(τ 2−γ ), (31)
Γ(2 − γ) j=0 j
=
at the face xi−1/2 , and
i+1 γ
∂α C (xi+1/2 , t) 1 X α where bj = (j + 1)1−γ − j 1−γ , j = 0, 1, . . . , n.
≈ α w C (xi−j+1 , t) ,
α h j=0 j
(26)
∂x
Letting Cin ≈ Ci (tn ) denote the approximate numerical so-
N−i
∂α C (xi+1/2 , t) 1 X α lution at node i at time tn , and using the discretisations
≈ w C (xi+j , t)
∂(−x)α hα j=0 j
(27)
just derived, we obtain the fully implicit scheme
n
at the face xi+1/2 . τ −γ X γ n+1−j n−j
b [C − Ci ]
Altogether then, the dispersive flux (11) is approximated Γ(2 − γ) j=0 j i
at the face xi−1/2 by n+1 n+1
Vi−1/2 Vi+1/2
Cin+1 + Ci−1
n+1
Cin+1 + Ci+1
n+1
= −
2h 2h
q(xi−1/2 , t) ≈ −K (xi−1/2 , t) × n+1
Ki−1/2
i N−i+1
β X α n+1 (1 − β) X α n+1
i N−i+1
! w C − wj Ci+j−1
hα j=0 j i−j hα
+
β X α (1 − β) X α h j=0
w j C i−j , t) − wj C (xi+j−1 , t)
hα j=0 hα
(x
j=0 n+1 i+1 N−i
Ki+1/2 β X α n+1 (1 − β) X α n+1
− w j Ci−j+1 − w j Ci+j +Sin+1 .
(28) h hα j=0 hα j=0
q(xi+1/2 , t) ≈ −K (xi+1/2 , t) ×
Collecting like terms, we may write the scheme in the form
i+1 N−i
!
β X α (1 − β) X α n N
w C i−j+1 , t) − wj C (xi+j , t) . τ −γ X γ n+1−j n−j 1X
hα j=0 j hα bj [Ci − Ci ] = gij Cjn+1 + Sin+1
(x
j=0 Γ(2 − γ) j=0 h j=0
(29) (33)
1279
n+1 n+1
Ki+1/2 β w α Ki−1/2 β α
i−j+1 − wi−j , j < i − 1;
h α h α
n+1 n+1
n+1 β
Ki+1/2
α
Ki−1/2 β α Ki−1/2 (1 − β) α n+1
hα w − α
w + α
w0 + Vi−1/2 /2, j = i − 1;
h h
2 1
n+1 n+1 n+1 n+1 n+1 n+1
Ki+1/2 β + Ki−1/2 (1 − β) α Ki−1/2 β + Ki+1/2 (1 − β) α Vi−1/2 − Vi+1/2
gij = w1 − w0 + , j = i; (34)
h α h α 2
n+1 n+1 n+1
Ki−1/2 (1 − β) α Ki+1/2 (1 − β) α Ki+1/2 β α
n+1
α
w2 − α
w1 + α
w0 − Vi+1/2 /2, j = i + 1;
h h h
n+1 n+1
Ki−1/2 (1 − β) w α Ki+1/2 (1 − β) α
− wj−i , j > i + 1.
j−i+1
hα hα
As our problem has homogeneous Dirichlet boundary con- For the finite volume method, the coefficients (34) simplify
ditions, we solve for just the internal nodes, and set to
C0n = CNn = 0 for n = 0, 1, . . .. Denoting the numeri-
cal solution vector Cn = [C1n , . . . , CN−1n
]T and source vector Kβ α α
hα
(wi−j+1 − wi−j ), j < i − 1;
n+1
S = [S(x1 , tn+1 ), . . . , S(xN−1 , tn+1 )] , we have the vector
T
K β K
α α α
α (w2 − w1 ) + α w0 + V /2, j = i − 1;
(1−β)
equation h
h
K
gij = hα
(w1α − w0α ), j = i;
K (1−β) α α Kβ α
τ γ Γ(2 − γ)
hα
(w − w ) + hα
w − V /2, j = i + 1;
A Cn+1
2 1 0
I+
K (1−β) α
α
h hα
(wj−i+1 − wj−i ), j > i + 1.
n−1
X (37)
γ γ
= bγn C0 + (bj − bj+1 )Cn−j + τ γ Γ(2 − γ)Sn+1 (35)
j=0
The finite difference discretisation of [4] applied to (36),
combined with a standard second order central difference
to solve at each timestep, where the matrix A has elements approximation for the advection term, yields a scheme with
aij = −gij . coefficients
K β α+1
hα
wi−j+1 , j < i − 1;
3.2. Comparison with finite difference
K β α+1 K (1−β) α+1
hα w 2 + w0 + V /2, j = i − 1;
hα
scheme
K
g̃ij = w α+1 , j = i;
hα 1
(38)
K (1−β) α+1
w2 + Khαβ w0α+1
hα
− V /2, j = i + 1;
It is instructive to compare the coefficients gij with those
K (1−β) α+1
hα
wj−i+1 , j > i + 1.
obtained from a finite difference discretisation of the same
problem. To do this, we consider the case of constant
velocity V and dispersion coefficient K . The problem (1)
can then be written For the two schemes to be equivalent, we require the iden-
tity wjα −wj−1
α
= wjα+1 for j = 1, 2, . . .. To prove this result,
γ ∂C we will use formula (21) for the Grünwald weights, as well
Dt C (x, t) + V
∂x as the recursive formula
α+1
∂α+1 C
∂ C
= K β α+1 + (1 − β) + S(x, t)
∂(−x)α+1 α +1
(36)
∂x
w0α = 1, wjα = 1− α
wj−1 , j = 1, 2, . . . (39)
j
and both the finite difference and finite volume methods
can be used to discretise this problem. which is easily seen to be equivalent to (21).
1280
Proposition 1.
0.25
numerical solution
The Grünwald weights wjα
satisfy wjα − α
wj−1 = wjα+1 for exact solution
j = 1, 2, . . .. 0.2
Proof: From (39) we have, for j ≥ 1,
Concentration
0.15
α +1
wjα − wj−1
α
= 1− α
wj−1 α
− wj−1
j 0.1
α +1 α
=− wj−1 T=2
j 0.05
α +1
wjα − wj−1
α
=− (−1)j−1 ×
j
α(α − 1) . . . (α − (j − 1) + 1) for (x, t) ∈ [0, 1] × [0, T ] together with following boundary
(j − 1)! and initial conditions
− . . . (α + 1 − j + 1)
= (−1)j
(α + 1)(α + 1 1)
j!
C (0, t) = C (1, t) = 0, 0 ≤ t ≤ T ,
= wjα+1 C (x, 0) = 0, 0 ≤ x ≤ 1
(41)
1281
Table 1. Maximum error with temporal and spatial step reduction at T = 1 for γ = 0.8, α = 0.85, β = 0.5.
Concentration
the values of h and τ, the third and fifth columns show 0.6
0.035
Example 2.
Concentration
0.03
0.02
advection-dispersion equation
0.015
α
∂α C
∂C (x, t) ∂ ∂ C 0.01
K (x, t) β α − (1 − β)
∂(−x)α
=
∂t ∂x ∂x 0.005
(42)
0
for (x, t) ∈ [−1, 1]×[0, T ] together with following boundary −1 −0.5 0 0.5 1
x
and initial conditions
( Figure 3. Comparison between fine mesh solution, finite volume and
C (−1, t) = C (1, t) = 0, 0 ≤ t ≤ T , finite difference methods for example 2 with α = 0.7 and
β = 0.5, using h = 0.02 and τ = 0.0001 at final time T = 1.
(43)
C (x, 0) = δ(x), −1 ≤ x ≤ 1
with dispersion coefficient K (x, t) = 2 + tanh(10x). This We use this test problem to highlight the advantage of our
function effectively partitions the domain [−1, 1] into two finite volume formulation compared to a standard finite
regions with dispersion coefficients approximately 1 and difference method. Using finite differences, the product
3, with a rapid transition between the two around the rule is applied to (42), leading to a term involving the
origin. Figure 2 illustrates the solution to this problem derivative ∂K /∂x. The finite volume method deals directly
over space and time, illustrating the asymmetric dispersion with (42) in its conservative form, and hence involves no
of the initial pulse. such term.
1282
Firstly, we solve (42) using a very fine mesh (h = 0.0002) [2] Y. Zhang, D. A. Benson, D. M. Reeves, Adv. Water
and small time steps (τ = 0.0001) to get a benchmark Resour. 32, (2009)
fine mesh solution. Then we compare finite volume and fi- [3] I. Podlubny, Fractional Differential Equations (Aca-
nite difference numerical solutions against that benchmark demic Press, 1999)
solution. [4] M. M. Meerschaert, C. Tadjeran, Appl. Numer. Math.
Figure 3 shows the comparison between fine mesh solu- 56, (2006)
tion, finite volume and finite difference methods for solving [5] K. Oldham, J. Spanier, The fractional calculus: theory
(42) with α = 0.7 and β = 0.5 at final time T = 1, us- and applications of differentiation and integration to
ing h = 0.02 and τ = 0.0001. The small timestep τ has arbitrary order (Academic Press, 1974)
been used to ensure that any differences in the solutions [6] Y. Lin, C. Xu, J. Comput. Phys. 225, 1533 (2007)
are purely due to the spatial discretisation. From the fig- [7] M. M. Meerschaert, C. Tadjeran, J. Comput. Appl.
ure it is evident that the finite volume solution with 100 Math. 172, (2004)
nodes more accurately represents the correct solution as [8] M. M. Meerschaert, H.-P. Scheffler, C. Tadjeran, J.
compared to the finite difference solution with 100 nodes. Comput. Phys. 211, (2006)
We conclude that the finite volume method offers a gen- [9] C. Tadjeran, M. M. Meerschaert, H. P. Scheffler, J.
uine advantage when solving problems such as this with Comput. Phys. 213, (2006)
variable dispersion coefficients. [10] K. Moaddy, I. Hashim, S. Momani, Comput. Math.
Appl. 62, 1068 (2011)
[11] K. Moaddy, S. Momani, I. Hashim, Comput. Math.
5. Conclusion Appl. 61, 1209 (2011)
[12] S. Momani, A. Rqayiq, D. Baleanu, Int. J. Bifurcat.
In this paper, we presented a finite volume method Chaos 22, 1250079 (2012)
to solve the time-space two-sided fractional advection- [13] P. Zhuang, F. Liu, V. Anh, I. Turner, SIAM J. Numer.
dispersion equation with variable coefficients on a one- Anal. 47, (2012)
dimensional finite domain with homogeneous Dirichlet [14] H. Sun, W. Chen, Y. Chen, Physica A: Statistical Me-
boundary conditions. The novel spatial discretisation em- chanics and its Applications 388, 4586 (2009)
ploys fractionally-shifted Grünwald formulas to discretise [15] S. Shen, F. Liu, J. Chen, I. Turner, V. Anh, Appl. Math.
the Riemann-Liouville fractional derivatives at control vol- Comput. 218, 10861 (2012)
ume faces in terms of function values at the nodes, while [16] H. Sun, W. Chen, C. LI, Y. Chen, Int. J. Bifurcat. Chaos
the L1-algorithm is used to discretise the Caputo time 22, 1250085 (2012)
fractional derivative. [17] H. Wang, K. Wang, T. Sircar, J. Comput. Phys. 229,
We showed that in the constant coefficient case, the 8095 (2010)
scheme is equivalent to the finite difference scheme de- [18] K. Wang, H. Wang, Adv. Water Resour. 34, 810 (2011)
scribed by [4]. For the variable coefficient case, the finite [19] H.-K. Pang, H.-W. Sun, J. Comput. Phys. 231, 693
volume scheme presented in this work is preferred for solv- (2012)
ing equation (1), since it deals directly with the equation [20] T. Moroney, Q. Yang, Comput. Math. Appl. 66, 659
in conservative form. (2013)
Numerical experiments confirm that the scheme recovers [21] T. Moroney, Q. Yang, J. Comput. Phys. 246, 304
the correct solution for test problems with fractional orders (2013)
in both time and space, as well as time- and space-varying [22] X. Zhang, J. W. Crawford, L. K. Deeks, M. I. Stutter,
velocity and dispersion coefficients. A. G. Bengough, I. M. Young, Water Resour. Res. 41,
Tadjeran et al. [9] have shown how to increase the order of (2005)
their finite difference scheme by Richardson extrapolation. [23] X. Zhang, M. Lv, J. W. Crawford, I. M. Young, Adv.
This too should be possible for the finite volume scheme, Water Resour. 30, 1205 (2007)
and will be the subject of future research. [24] H. Hejazi, T. Moroney, F. Liu, J. Comput. Appl. Math.
255, 684 (2014)
[25] S. Kim, M. L. Kavvas, J. Hydro. Eng. 1, (2006)
References [26] Q. Yang, F. Liu, I. Turner, Appl. Math. Model. 34,
(2010)
[1] R. Metzler, J. Klafter, Phys. Rep. 339, (2000) [27] S. Shen, F. Liu, V. Anh, Numer. Algorithms 56, (2011)
1283