Chapter 13

A New Numerical Approximation

of Fractional Differentiation: Upwind
Discretization for Riemann-Liouville
and Caputo Derivatives

Abdon Atangana

13.1 Introduction

Many real-world problems can be modeled with partial differential equations called
hyperbolic equations with integer- and non-integer-order derivatives. In general,
hyperbolic partial differential equation with order m is a partial differential equation
with a well-posed initial value problem in particular for the first m − 1 derivatives
in local concept [1–6]. In mechanic, many mathematical models are hyperbolics.
Thus the study of hyperbolic equation is of substantial contemporary interest.
There are approaches to solve these equations: on one hand, one can use analytical
methods to provide their exact solutions. On the other hand, one can rely on
numerical methods. We shall point out that the upwind numerical scheme has been
recognized as a powerful numerical method to solve more accurately hyperbolic
equations [6–8]. However, this numerical scheme has not yet been developed for
the concept of fractional differentiation; although there exist many fractional hyper-
bolic equations that model real-world problem, there is no upwind approximation
for fractional differentiation. In computational methods for physical problem, in
particular hyperbolic partial differential equations, the upwind numerical scheme
denotes a class of numerical discretization approaches for solving hyperbolic PDE.
Upwind schemes employ a modified- or solution-sensitive finite difference stencil
to numerically replicate the directive of spray of information in the flow field.
This numerical approximation approach attempt to approximate hyperbolic partial
differential equations by employing differencing biased in the directive determined
by the sign of the characteristic speeds. The available information in the literature
revealed that the genesis of upwind methods can be traced back to the work done

A. Atangana ()
Faculty of Natural and Agricultural Sciences, Institute for Groundwater Studies, University of the
Free State, Bloemfontein, South Africa

© Springer International Publishing AG, part of Springer Nature 2019 193

K. Taş et al. (eds.), Mathematical Methods in Engineering, Nonlinear Systems
and Complexity 24, https://doi.org/10.1007/978-3-319-90972-1_13
194 A. Atangana

by Courant, Isaacson, and Rees who suggested the CIR method [1–4]. Considering
the flow direction, the upwind approximation differencing scheme overcomes that
inability of the central differencing scheme. Reputed analysis like the well-known
level set that uses the level set equations that can be handled numerically, however,
requires sophisticated techniques. Simple finite difference methods fail quickly.
Upwinding methods, such as the Godunov method, fare better [9–15]. In this
paper, we provided using the upwind approach the Riemann-Liouville and Caputo
fractional derivative were discretized for first and second approximation. Combining
the Crank-Nicolson approach and the upwind, a new numerical approximation was
developed.

13.2 Upwind for Riemann-Liouville Derivative

In this section, we derive the numerical approximation of Riemann-Liouville

derivative using the upwind numerical approach. Without loss of generality, we
consider the backword. By definition, the Riemann-Liouville fractional derivative
is given by

1 ∂
RL α
0 Dx f (x, t) = F (x, t) (13.1)
(1 − α) ∂x
x
1 ∂
= f (τ , t) (x − τ )−α dτ .
(1 − α) ∂x
0

Therefore at the point (xn , ti ), we have the following:

1 F (xn+1 , ti ) − F (xn , ti )
RL α
0 Dx f (xn , ti ) = . (13.2)
(1 − α) x

where

xn
Fα (xn , ti ) = f (τ , ti ) (xn − τ )−α dτ (13.3)
0

n−1 k+1
x

= f (xk , ti ) (xn − τ )−α dτ
k=0 xk

 (x)1−α  
n−1
= f (xk , ti ) (n − k)1−α − (n − k − 1)1−α ,
1−α
k=0
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 195

(x)1−α 
n−1  
Fα (xn , ti ) = f (xk , ti ) (n − k)1−α − (n − k − 1)1−α (13.4)
(1 − α)
k=0

1−α 
n−1
(x)
= f (xk , ti )δ αn,k .
(1 − α)
k=0

Thus
 n 
(x)−α  
n−1
RL α
0 Dx f (xn , ti ) = f (xk , ti )δ n+1,k −
α α
f (xk , ti )δ n,k . (13.5)
(2 − α)
k=0 k=0

Therefore the numerical approximation of Riemann-Liouville fractional derivative

with upwind backward is given by
 n 
(x) −α  
n−1

0 Dx f (xn , ti ) = f (xk , ti )δ αn+1,k −

RL α
f (xk , ti )δ αn,k . (13.6)
(2 − α)
k=0 k=0

The Riemann-Liouville approximation with forward upwind is thus given as

follows:
 n 
(x)−α  
n−1

0 Dx f (xn , ti ) = f (xk , ti )δ n+1,k −

RL α α α
f (xk , ti )δ n,k . (13.7)
(2 − α)
k=0 k=0

13.3 The Second-Order Upwind for Riemann-Liouville

It is important to mention that the first presented earlier can be improved by

including three-point data instead of just two. Therefore without loss of generality,
we consider first the backward upwind approach as below:
1 3F (xn , ti ) − 4F (xn−1 , ti ) + F (xn−2 , ti )
RL α
0 Dx f (xn , ti ) = . (13.8)
(1 − α) 2 (x)

where
(x)1−α 
n−1
F (xn , ti ) = f (xk , ti )δ αn,k , (13.9)
(1 − α)
k=0

1−α 
n−3
(x)
F (xn−2 , ti ) = f (xk , ti )δ αn−2,k ,
(1 − α)
k=0

(x)1−α 
n−2
F (xn−1 , ti ) = f (xk , ti )δ αn−1,k .
(1 − α)
k=0
196 A. Atangana

Therefore the Riemann-Liouville backward upwind scheme is given as

⎧ ⎫
⎪
⎪ 
n−1 ⎪
⎪ 3
⎪ f (xk , ti )δ n,k ⎪
α ⎪
⎪
⎪
⎪ ⎪
⎪
−α ⎪ ⎪
k=0
(x) ⎨ 
n−2 ⎬
RL α
D f (xn , t i ) = −4 f (x , t )δ
k i n−1,k
α . (13.10)
0 x
2(2 − α) ⎪
⎪ ⎪
⎪
⎪
⎪
k=0 ⎪
⎪
⎪
⎪ 
n−3 ⎪
⎪
⎪
⎩ + f (xk , ti )δ αn−2,k ⎪
⎭
k=0

Also the Riemann-Liouville forward upwind numerical scheme is given as

⎧ ⎫
⎪
⎪ 
n+2 ⎪
⎪
⎪ − f (xk , ti )δ n+2,k ⎪
α ⎪
⎪
⎪
⎪ ⎪
⎪
−α ⎨⎪ k=0 ⎪
⎬
(x) 
n+1
RL α
D f (x ,
n it ) = +4 α
f (xk , ti )δ n+1,k . (13.11)
0 x
2(2 − α) ⎪
⎪ ⎪
⎪
⎪
⎪
k=0 ⎪
⎪
⎪
⎪ n ⎪
⎪ −3
⎩ f (xk , ti )δ αn,k ⎪
⎪
⎭
k=0

To obtain an approximation that is less diffusive than the second-order accurate

scheme, we introduce the third-order upwind scheme Riemann-Liouville derivative.
The backward will be
 
1 2F (xn , ti ) + 3F (xn−1 , ti )
0 Dx f (xn , ti ) =
RL α
. (13.12)
6 (x) (1 − α) −6F (xn−2 , ti ) + F (xn−3 , ti )

So we have
⎧ ⎫
⎪ 
n−1
⎪
⎪
⎪ 2 f (xk , ti )δ n,k ⎪
α ⎪
⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪
k=0 ⎪
⎪
⎪
⎪ 
n−2 ⎪
⎪
⎪
⎨ +3 f (xk i n−1,k ⎪
, t )δ α
⎬
(x)−α
RL α
0 Dx f (xn , ti ) = k=0 . (13.13)
6(2 − α) ⎪
⎪ 
n−3 ⎪
⎪
⎪
⎪ −6 f (x , t )δ α ⎪
n−2,k ⎪
⎪
⎪
k i ⎪
⎪
⎪
⎪ k=0 ⎪
⎪
⎪
⎪ 
n−4 ⎪
⎪
⎪
⎩ + f (xk , ti )δ α ⎪
⎭
n−3,k
k=0

The forward third-order for Riemann-Liouville is given as below:

⎧ ⎫
⎪ 
n+1
⎪
⎪
⎪ −2 f (xk , ti )δ n,k ⎪
α ⎪
⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪
k=0 ⎪
⎪
⎪
⎪ 
n+2 ⎪
1−2α ⎪
⎨ −3 f (xk , ti )δ n+1,k ⎪
α ⎪
⎬
(x)
0 Dx f (xn , ti ) =
RL α k=0 . (13.14)
6(2 − α)x ⎪
⎪ 
n+3 ⎪
⎪
⎪
⎪ +6 f (x , t )δ α ⎪
⎪
⎪
⎪
k i n+2,k ⎪
⎪
⎪
⎪ n+4 k=0 ⎪
⎪
⎪
⎪  ⎪
⎪
⎪
⎩ − f (xk , ti )δ n+3,k ⎪
α ⎭
k=0
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 197

Theorem 1 Let f be nonnecessary differentiable on an open interval [0, x) such

that the Riemann-Liouville fractional derivative exists, thus
 n 
(x)−α  
n−1

0 Dx f (xn , ti ) = f (xk , ti )δn+1,k −

RL α α α
f (xk , ti )δn,k (13.15)
(2 − α)
k=0 k=0
 
 
where Rα  < ∞.

Proof Assume that RL α

0 Dx f (xn , ti ) exists then
RL α
0 Dx f (xn , ti ) (13.16)
⎧ ⎫
⎪ 
n
⎪
⎪
⎪ f (xk , ti )δ αn+1,k ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ k=0 ⎪
⎪
(x)−α ⎨ 
n−1 ⎬
= − f (xk , ti )δ αn,k .
(2 − α) ⎪
⎪   ⎪
⎪
⎪
⎪ x
k=0
⎪
⎪
⎪
⎪
n+1 ⎪
(xn+1 − y)−α dy ⎪
f (x, y)
⎪
⎩ + (1−α)
1 ⎪
⎭
−f (x , t )
0 n+1 i
xn+1
1
= {f (x, y) − f (xn+1 , ti )} (xn − y)−α dy
(1 − α)
0
xn+1 
1 f (x, y) − f (xn+1 , ti )
= (x − xn+1 ) (xn − y)−α dy,
(1 − α) (x − xn+1 )
0
xn+1
1 ∂
= f (ξ , xi ) (x − xn+1 ) (xn − y)−α dy
(1 − α) ∂x
0
xn+1
1 ∂
= f (ξ , xi ) (α − n − 1) t (xn − y)−α dy
(1 − α) ∂x
0

   (α − n − 1) t   (xn − y)1−α
 
Rα  ≤ max f  (ξ , xi )
(1 − α) 0<x<xn+1 1−α
(n + 1 − α) t    
≤ max f  (ξ , xi ) xn1−α − (xn − xn+1 )1−α
(2 − α) 0<x<xn+1
(n + 1 − α) (t)2−α   
≤ max f  (ξ , xi ) n1−α − 1
(2 − α) 0≤x<xn+1

(t)2−α
≤ M1 ,
(2 − α)
198 A. Atangana

thus

 n 2−α
R  ≤ (t) (M1 − M2 ) , (13.17)
(2 − α)

 n 2−α
R  ≤ (t) M.
(2 − α)

13.4 Compact Form of Upwind for Riemann-Liouville

Derivative

Let RL α
0 Dx − u be the numerical approximation of Riemann-Liouville using the
0 Dx + u is the forward. If β ∈ [0, 1], then the
upwind backward approach and RL α

compact form of upwind for Riemann-Liouville is given as

RL α
0 Dx u = β RL
0 Dx − u + (1 − β)0 Dx + u
α RL α
(13.18)

Remark If β = 1, we obtain the forward, and if β = 0, we obtain the backward.

13.4.1 Upwind for Riemann-Liouville Derivative 1 < α ≤ 2

The space Riemann-Liouville fractional derivative of order 1 < α ≤ 2 can also be

well approximated using the upwind approach.
By definition

x
1 d2
RL α
0 Dx f (x, t) = f (τ , t) (x − τ )1−α dτ (13.19)
(2 − α) dx 2
0

at (xn , ti ).

xn
1 d2
RL α
0 Dx f (xn , ti ) = f (τ , ti ) (xn − τ )1−α dτ (13.20)
(2 − α) dx 2
0
 
1 −Fn+2
i + 16Fn+1
i
= .
12(2 − α) (x)2 −30Fni + 16Fn−1
i − Fn−2
i
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 199

where
xn±k
i
Fn±k = f (τ , ti ) (xn±k − τ )1−α dτ , k ∈ {0, 1, 2} (13.21)
0

n±k j +1
x

= f (xj , ti ) (xn±k − τ )1−α dτ
j =0 xj

xj +1

n±k
= f (xj , ti ) (xn±k − τ )1−α dτ
j =0 xj


n±k  −xj
xn±k  −xj
xn±k

= f (xj , ti ) Y 1−α dτ
j =0 xn±k −xj +1 xn±k −xj +1

 f (xj , ti )  2−α 
n±k
2−α 
= xn±k − xj − xn±k − xj +1
(2 − α)
j =0

 f (xj , ti ) (x)2  
n±k
= (n ± k − j )2−α − (n ± k − j − 1)2−α .
(2 − α)
j =0

Therefore
⎡ ⎤

n+2
⎢ − f (x , t )δ
j i n+2,j ⎥
α
⎢ j =0 ⎥
⎢ ⎥
⎢ 
n+1 ⎥
⎢ +16 f (xj , ti )δ n+1,j ⎥
α
⎢ ⎥
⎢ j =0 ⎥
−α ⎢ 
n+1 ⎥
(x) ⎢ −30 f (xj , ti )δ n,j ⎥
α
0 Dx f (xn , ti ) =
RL α
⎢ ⎥. (13.22)
12(3 − α) ⎢ =0 ⎥
⎢ j ⎥
⎢ 
n−1 ⎥
⎢ +16 f (x , t )δ α ⎥
⎢ j i n−1,j ⎥
⎢ j =0 ⎥
⎢ ⎥
⎣ 
n−2 ⎦
− f (xj , ti )δ αn−2,j
j =0
200 A. Atangana

An alternative approximation can be given as

⎡ ⎤

n+3
⎢ − f (xj , ti )δ αn+3,j
⎥
⎢ j =0 ⎥
⎢ ⎥
⎢ 
n+2 ⎥
⎢ +4 f (xj , ti )δ n+2,j ⎥
α
−α ⎢ ⎥
(x) ⎢ j =0 ⎥
0 Dx f (xn , ti ) =
RL α
⎢ ⎥. (13.23)
(3 − α) ⎢ 
n+1 ⎥
⎢ −5 f (xj , ti )δ n+1,j ⎥
α
⎢ ⎥
⎢ j =0 ⎥
⎢  ⎥
⎣ n
⎦
+2 α
f (xj , ti )δ n,j
j =0

For 1 < α ≤ 2 , the compact for m can be given as

0 Dx f (xn , ti ) + (2 − β)0 Dx + f (xn , ti ).

β RL α RL α
(13.24)

13.5 New Crank-Nicolson with Upwind

The Crank-Nicolson has been recognized as a powerful mathematical tool to solve

numerical ordinary and partial differential equations with integer and non-integer
orders. The Crank-Nicolson approach can be established using the upwind, and this
will be presented in this section. If f is two times differentiable, then the Crank-
Nicolson upwind approach is given by

i+1 i+1 i+1

∂U (xn , ti ) −Un+3 + 4Un+2 − 5Un+1 + 2Uni+1
= + (13.25)
∂x 2 2 (x)2

−Un+3
i + 4Un+2
i − 5Un+1
i + 2Uni
2 (x)2
or
⎧  ⎫
1 ⎨− −Un+2 + 16Un+1 − 30Un + 16Un−1 − Un−2 ⎬
i+1 i+1 i+1 i+1 i+1
∂U (xn , ti )
= .
∂x 2 2 (x)2 ⎩ + −U i + 16U i − 30U i + 16U i − U i  ⎭
n+2 n+1 n n−1 n−2
(13.26)

Using the above, we can find the Upwind-Crank-Nicolson scheme for Riemann-
Liouville derivative with 1 < α ≤ 2.
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 201

⎧ ⎫
1 ⎨ −U i+1 + 4U i+1 − 5U i+1 + 2Uni+1 ⎬
n+3 n+2 n+1
RL α
0 Dx U (xn , ti ) = (13.27)
2 (x)2 ⎩ + − U i + 4U i − 5U i + 2U i ⎭
n+3 n+2 n+1 n

where
xn±k
1
i+1
Un±k = u(x, ti+1 ) (xn±k − τ )−α dτ (13.28)
(2 − α)
0

(x)1−α  α
n±k
= δ n±k,j U (xj, ti ).
(3 − α)
j =0

Also if U is two times differentiable, then Riemann-Liouville approximation using

upwind-Crank-Nicolson scheme is given as
⎡ ⎤

n+2 
n+1
⎢− U (xj , ti+1 )δ αn+2,j + 16 U (xj , ti+1 )δ αn+1,j ⎥
⎢ j =0 j =0 ⎥
⎢ ⎥
−α−1 ⎢ ⎥
RL D α U (x , t ) = (x) ⎢ −30 n
U (xj , ti+1 )δ αn,j + 16

n−1
U (xj , ti+1 )δ αn−1,j ⎥
x n i ⎢ ⎥
0 2(3 − α) ⎢ =0 =0 ⎥
⎢ j j ⎥
⎢ 
n−2 ⎥
⎣ ⎦
− U (xj , ti+1 )δ αn−2,j
j =0
⎡ ⎤ (13.29)

n+2 
n+1
⎢ − + 16
U (xj , ti )δ αn+2,j U (xj , ti+1 )δ αn+1,j
⎥
⎢ j =0 j =0 ⎥
⎢ ⎥
(x)−α−1 ⎢⎢
n 
n−1 ⎥
+ −30 U (xj , ti+1 )δ n,j + 16
α U (xj , ti+1 )δ n−1,j ⎥
α
⎥.
2(3 − α) ⎢
⎢ j =0 j =0 ⎥
⎢ ⎥
⎣ 
n−2 ⎦
− U (xj , ti+1 )δ αn−2,j
j =0

13.6 Upwind for Caputo Fractional Derivative

We devote this section to numerical approximation of Caputo derivative using the

upwind approach. We shall start with 0 < α ≤ 1.
By definition

x
1
C α
0 Dx f (x, t) = fτ (τ , t) (x − τ )−α dτ . (13.30)
(1 − α)
0
202 A. Atangana

At (xn , ti ), we have the following:

xn
1
C α
0 Dx f (xn , ti ) = fτ (τ , ti ) (xn − τ )−α dτ (13.31)
(1 − α)
0
xj +1
 n j +1 j
1 fi − fi
= (xn − τ )−α dτ
(1 − α) x
j =0 xj

xj +1
f n j +1 j
1 − fi
= i
(xn − τ )−α dτ
(1 − α) x
j =0 xj
⎧ ⎫
⎪
⎪ (n − j )1−α ⎪
⎪
⎪ ⎪
− fi ⎨ − (n − j − 1)1−α ⎬
fn j +1 j
1
= i
(x)1−α
(1 − α) x ⎪
⎪ (1 − α) ⎪
⎪
j =0 ⎪
⎩ ⎪
⎭
 
(x)−α   j +1
n
j (n − j )1−α
= fi − fi
(2 − α) − (n − j − 1)1−α
j =0

(x)−α   j +1
n
j
= fi − fi δ αn,j .
(2 − α)
j =0

With the above derivation, the second-order upwind for Caputo derivative is given
as in backward approach.
 j j −1 j −2

(x)−α  3fi − 4fi
n
+ fi
C α
0 Dx + f (xn , ti ) = δ αn,j . (13.32)
(2 − α) 2
j =0

The forward second-order upwind for Caputo fraction derivative is given as

 j +1 j +2

(x)−α  −3fi + 4fi
n j
− fi
C α
0 Dx − f (xn , ti ) = δ αn,j . (13.33)
(2 − α) 2
j =0

With the above numerical approximation, we suggest the following compact form
of second-order upwind for space-fractional Caputo derivation that is given as
C α
0 Dx U (xn , ti ) = βC
0 Dx − U (xn , ti ) + (1 − β)0 Dx + U (xn , ti ).
α C α
(13.34)

This leads us to establish the second-order upwind-Crank-Nicolson scheme for

Caputo derivative in backward approach. If f is differentiable, then the Crank-
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 203

Nicolson upwind second-order approximation is given as for backward

⎧ ⎫
⎨ 3Uni+1 − 4U i+1 + U i+1 ⎬
n−1 n−2
⎩  ⎭
∂ + 3Uni − 4Un−1
i + Un−2
i
U (xn , ti ) = , (13.35)
∂x 4 (x)

for forward approach we have

⎧ ⎫
⎨ −3Uni+1 + 4U i+1 − U i+1 ⎬
n+1 n+2
⎩  ⎭
∂ + −3Uni + 4Un+1
i − Un+2
i
U (xn , ti ) = . (13.36)
∂x 4 (x)

With both approximations in hand, we construct the Caputo approximation for

Crank-Nicolson upwind backward as
⎧ ⎫
n ⎪⎨ 3Uji+1 − 4Uji+1 i+1 ⎪
(x) 
1−α −1 + Uj −2 ⎬
C α
D f (x ,
n it ) =  δ αn,j . (13.37)
0 x
4 (2 − α) ⎪ ⎪
j =0 ⎩ + 3Uj − 4Uj −1 + Uj −2 ⎭
i i i

Also the Caputo approximation for Crank-Nicolson upwind forward is given as :

⎧ ⎫
n ⎪⎨ −3fji+1 + 4fji+1 i+1 ⎪
(x) 1−α +1 − fj +2 ⎬
0 Dx f (xn , ti ) =
C α
 δ αn,j . (13.38)
4 (2 − α) ⎪ ⎪
j =0 ⎩ + −3fj + 4fj +1 − fj +2 ⎭
i i i

13.7 Second-Order Approximation for Caputo

The second-order approximation is used to approximate the acceleration component

of the mathematical model in case of time derivative. In the section, the Caputo
fractional derivative will be approximated using the upwind approximation. By
definition
x
1 d2
C α
0 Dx f (x, t) = f (τ , t) (x − τ )1−α dτ , 1 < α ≤ 2. (13.39)
(2 − α) dτ 2
0
204 A. Atangana

Then at (xn , ti ), we have

xn
1 d2
C α
0 Dx f (xn , ti ) = f (τ , ti ) (xn − τ )1−α dτ (13.40)
(2 − α) dτ 2
0
1
= .
12(2 − α) (x)2
x  
n j +1
−fn+2
i + 16fn+1
i
(xn − τ )1−α dτ
−30fni + 16fn−1
i − fn−2
i
j =0 xj

1
= .
12(2 − α) (x)2
  xj +1
n
−fn+2
i + 16fn+1
i
(xn − τ )1−α dτ
−30fni + 16fn−1
i − fn−2
i
j =0 xj

2−α
(x)
= .
12(2 − α)(2 − α) (x)2
 
n
−fn+2
i + 16fn+1
i
δ αn,j
−30fni + 16fn−1
i − fn−2
i
j =0

where

δ αn,j = (n − j )2−α − (n − j − 1)2−α . (13.41)

Or we have the following:

 
(x)−α 
n
−fn+3
i + 4fn+2
i
C α
0 Dx f (xn , ti ) = δ αn,j . (13.42)
2(3 − α) −5fn+1 + 2fni
i
j =0

Using the Crank-Nicolson upwind approach, we obtain the following:

⎧   ⎫
⎪
⎪ n −f i+1
+ 4f i+1 ⎪
⎪
⎪ δ n,j ⎪
α
−α ⎪ ⎪
n+3 n+2
(x) ⎨ −5f i+1
+ 2f i+1 ⎬
j =0  n+1 n 
C α
D f (xn , t i ) = . (13.43)
0 x
4(3 − α) ⎪
⎪ n −fn+3
i + 4fn+2
i ⎪
⎪
⎪ ⎪
⎩+
α
⎪ δ n,j ⎪
⎭
j =0 −5fn+1i + 2fni

Using the other upwind numerical scheme for second-order derivative, we propose
the following approximation of Caputo derivative for 1 < α ≤ 2,:
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 205

⎧ ⎛ ⎞ ⎫
⎪
⎪ −f i+1
+ 16f i+1 ⎪
⎪
⎪
⎪ n
⎜ n+2 n+2
⎟ α ⎪ ⎪
⎪
⎪ ⎪
⎪
⎪
⎝ ⎠ δ n,j ⎪⎪
⎪
(x) −α ⎨ j =0 i+1 i+1 i+1
−30fn+1 + 16fn−1 − fn−2 ⎬
C α
D f (x ,
n it ) = ⎛ ⎞ .
0 x
24(3 − α) ⎪
⎪ −fn+2
i + 16fn+2
i ⎪
⎪
⎪
⎪ ⎜
n
⎟ α ⎪ ⎪
⎪
⎪ + ⎝ ⎠ δ n,j ⎪
⎪
⎪
⎪ ⎪
⎪
⎩ j =0 −30f i + 16f i − f i ⎭
n+1 n−1 n−2
(13.44)

13.8 Application to Advection Equation

In this section, we apply the proposed numerical scheme to solve the linear
advection model which describes a wave propagating along the x−axis with a
velocity. In this section the model is generalized by replacing the space derivative
by the space-fractional derivative in Riemann-Liouville sense then in Caputo sense.
Therefore in Caputo sense, the model has the following form:

x
∂U a dU (τ , t)
+ (x − τ )−α dτ = 0. (13.45)
∂t (1 − α) dτ
0

With the simplest first-order upwind scheme, we have the following:

a (x)−α   n
n
Uin+1 − Uin
+ Uj − Ujn−1 δ αn,j = 0forα > 0 (13.46)
t (2 − α)
j =0

and

a (x)−α   n
n
Uin+1 − Uin
+ Uj +1 − Ujn δ αn,j = 0ifα < 0. (13.47)
t (2 − α)
j =0

Therefore to take the full axis into account, we have the following compact form. If
we consider

a + = max(a, 0), a − = min(a, 0) (13.48)

and

Uin − Ui−1
n n − Un
Ui+1 i
Ux − = , Ux + = (13.49)
x x
206 A. Atangana

then

(x)−α   n
n
Uxα− = Uj − Ujn−1 δ αn,j , (13.50)
(2 − α)
j =0

(x)−α  n
Ux − = Uj +1 − Ujn δ αn,j .
(2 − α)

Then the numerical solution is given by

 
Uin+1 = Uin − t a − Uxα− + a + Uxα+ . (13.51)

To obtain a more accurate first-order fractional upwind scheme for the advection
model in space, the following numerical solution is given by using the second-order
fractional upwind scheme:

(x)−α   n
n
Uin+1 − Uin
+a 3Uj − 4Ujn−1 + Ujn−2 δ αn,j = 0ifa < 0, (13.52)
t 2(2 − α)
j =0

(x)−α  
n
Uin+1 − Uin
+a −3Ujn + 4Ujn+1 − Ujn+2 δ αn,j = 0ifa > 0.
t 2(2 − α)
j =0

Again

(x)−α   n
n
Uxα,1
− = 3Uj − 4Ujn−1 + Ujn−2 δ αn,j , (13.53)
2(2 − α)
j =0

(x)−α  
n
Uxα,1
+ = −3Ujn + 4Ujn+1 − Ujn+2 δ αn,j .
2(2 − α)
j =0

Then the compact form is given as


Uin+1 = Uin − t a − Uxα,1 + α,1
− + a Ux + . (13.54)

Using the new numerical scheme proposed in this paper (Crank-Nicholsan upwind
backward scheme) we obtain for second-order approach
⎛ ⎞
(x)−α  ⎝ 3U
n+1
Uin+1 − Uin
n − 4Ujn+1
−1 + Ujn+1
−2
+a  j ⎠ δ αn,j = 0ifa < 0,
t 4(2 − α) + 3U n − 4U n + U n
j =0 j j −1 j −2

(13.55)
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 207

⎛ ⎞
(x)−α ⎝ −3U
n+1 n+1 n+1
Uin+1 − Uin
n + 4Uj +1 − U j +2 ⎠ α
+a  j δ n,j = 0ifa > 0.
t 4(2 − α) + −3U n + 4U n − U n
j =0 j j +1 j +2

Then the compact form is given as


Uin+1 = Uin − t a − Uxα,2 + α,2
− + a Ux + , (13.56)

where
⎛ ⎞
(x)−α  ⎝ 3U
n+1
n − 4Ujn+1
−1 + Ujn+1
−2
Uxα,2
− =  j ⎠ δ αn,j , (13.57)
4(2 − α) + 3Uj − 4Uj −1 + Uj −2
n n n
j =0
⎛ ⎞
(x)−α  ⎝ −3U
n+1 n+1 n+1
n + 4U j +1 − U j +2 ⎠ α
Uxα,2
+ =  j δ n,j .
4(2 − α) + −3U n + 4U n − U n
j =0 j j +1 j +2

A more accurate approach would be applied in space to obtain a more accurate

approximation of the model. To do this, we used the third-order approach developed
previously to get
 
(x)−α 
n
Uin+1 − Uin 2Ujn + 3Ujn−1
= −a δ αn,j = 0ifa < 0 (13.58)
t 6(2 − α) −6Ujn−2 + Ujn−3
j =0

or
 
(x)−α 
n
Uin+1 − Uin −2Ujn − 3Ujn+1
= −a δ αn,j = 0ifa > 0. (13.59)
t 6(2 − α) +6Ujn+2 − Ujn+3
j =0

13.9 Numerical Simulation of Model with Riemann-Liouville

In this section we solved numerically the following equation:

∂u(x, t) ∂u(x, t) ∂ α (x, t)

= −v(x) +d . (13.60)
∂t ∂x ∂x α
The equation used the Riemann-Liouville fractional derivative. The numerical
solutions are depicted for fixed x and t for different values of alpha. Here we
consider 1 < α ≤ 2,; the solution are depicted in figures below (Figs. 13.1, 13.2,
13.3, 13.4, 13.5, 13.6, 13.7, 13.8, 13.9, and 13.10).
208 A. Atangana

Fig. 13.1 Numerical

simulation for fixed x for
alpha = 1

Fig. 13.2 Numerical

simulation for fixed x for
alpha = 1.25

Fig. 13.3 Numerical

simulation for fixed x for
alpha = 1.5
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 209

Fig. 13.4 Numerical

simulation for fixed x for
alpha = 1.75

Fig. 13.5 Numerical

simulation for fixed x for
alpha = 2

Fig. 13.6 Numerical

simulation for fixed t for
alpha = 1
210 A. Atangana

Fig. 13.7 Numerical

simulation for fixed t for
alpha = 1.25

Fig. 13.8 Numerical

simulation for fixed t for
alpha = 1.5

Fig. 13.9 Numerical

simulation for fixed t for
alpha = 1.75
13 A New Numerical Approximation of Fractional Differentiation: Upwind. . . 211

Fig. 13.10 Numerical

simulation for fixed t for
alpha = 2

13.10 Conclusion

Anonymous diffusions and advection have been recognized as complex problems

that describe very important physical problems. Mathematical equations describing
these physical problems are classified under hyperbolic equations. They are often
found in the field of mechanics, hydrology, geohydrology, thermal science, and
others, for instance, the spray of heat within a homogeneous or heterogeneous
matters, the movement of water within a nonhomogeneous geological formation
called aquifers, and the movement of contamination and pollution within subsurface
water. Based on the literature review, it is revealed that the concept of fractional
differentiation is more suitable for describing such natural occurrence rather than
local differentiation, because of their ability of including into mathematical formu-
lation the effect of heterogeneity and also memory. Nevertheless, some of these
models can only be solved via numerical schemes for differential equations with
integer-order derivative in particular, the upwind was accepted as accurate numerical
scheme able to handle these problems, and this is also true for those versions with
fractional differentiation. However, there is no sign of upwind numerical scheme for
fractional differentiation. In this paper, we provided using the upwind approach the
Riemann-Liouville and Caputo fractional derivative were discretized for first and
second approximation. Combining the Crank-Nicolson approach and the upwind, a
new numerical approximation was developed. The approach was tested by solving
a simple advection model.
212 A. Atangana

References

1. Osher, S., Sethian, J.A.: Fronts propagating with curvature-dependent speed: algorithms based
on Hamilton–Jacobi formulations (PDF). J. Comput. Phys. 79, 12–49 (1988)
2. Versteeg, H.K., Malalasekera, W.: An Introduction to Computational Fluid Dynamics, Chapter
5, p. 104. Wiley, New York (1995)
3. Courant, R., Isaacson, E., Rees, M.: On the solution of nonlinear hyperbolic differential
equations by finite differences. Commun. Pure Appl. Math. 5, 243–255 (1952)
4. Patankar, S.V.: Numerical Heat Transfer and Fluid Flow. Taylor & Francis (1980). ISBN: 978-
0-89116-522-4
5. Hirsch, C.: Numerical Computation of Internal and External Flows. Wiley (1990). ISBN: 978-
0-471-92452-4
6. Godunov, S.K.: A difference scheme for numerical solution of discontinuous solution of
hydrodynamic equations. Math. Sbornik 47, 271–306 (1959)
7. Liou, M.-S., Steffen, C.: A new flux splitting scheme. J. Comput. Phys. 107, 23–39 (1993)
8. Liou, M.-S.: A sequel to AUSM: AUSM+. J. Comput. Phys. 129, 364–382 (1996)
9. Wada, Y., Liou, M.-S.: An accurate and robust flux splitting scheme for shock and contact
discontinuities. SIAM J. Sci. Comput. 18, 633–657 (1997)
10. Liou, M.-S.: A sequel to AUSM, part II: AUSM+-up. J. Comput. Phys. 214, 137–170 (2006)
11. Edwards, J.R., Franklin, R., Liou, M.-S.: Low-diffusion flux-splitting methods for real fluid
flows with phase transitions. AIAA J. 38, 1624–1633 (2000)
12. Chang, C.-H., Liou, M.-S.: A new approach to the simulation of compressible multifluid flows
with AUSM+ scheme. In: 16th AIAA CFD Conference on AIAA Paper 2003-4107, Orlando,
23–26 June 2003
13. Edwards, J.R., Liou, M.-S.: Low-diffusion flux-splitting methods for flows at all speeds. AIAA
J. 36, 1610–1617 (1998)
14. Kim, K.H., Kim, C., Rho, O.: Methods for the accurate computations of hypersonic flows I.
AUSMPW+ scheme. J. Comput. Phys. 174, 38–80 (2001)
15. Mary, I., Sagaut, P.: Large Eddy simulation of flow around an airfoil near stall. AIAA J. 40,
1139–1145 (2002)