RESEARCH PAPER

A FINITE ELEMENT METHOD FOR TIME FRACTIONAL

PARTIAL DIFFERENTIAL EQUATIONS

Neville J. Ford 1 , Jingyu Xiao 2 , Yubin Yan 3

Abstract
In this paper, we consider the finite element method for time fractional
partial differential equations. The existence and uniqueness of the solutions
are proved by using the Lax-Milgram Lemma. A time stepping method is
introduced based on a quadrature formula approach. The fully discrete
scheme is considered by using a finite element method and optimal con-
vergence error estimates are obtained. The numerical examples at the end
of the paper show that the experimental results are consistent with our
theoretical results.
MSC 2010 : Primary 65M12; Secondary 65M06, 65M60, 65M70, 35S10
Key Words and Phrases: fractional partial differential equations, finite
element method, error estimates, numerical examples

1. Introduction
In this paper, we will consider the finite element method for the time
fractional partial differential equation
R α
0 Dt u(t, x) − Δu(t, x) = f (t, x), t ∈ [0, T ], x ∈ Ω, (1.1)
u(0, x) = 0, x ∈ Ω, (1.2)
u(t, x) = 0, t ∈ [0, T ], x ∈ ∂Ω, (1.3)


c 2011 Diogenes Co., Sofia
pp. 454–474 , DOI: 10.2478/s13540-011-0028-2
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 455

where 0 < α < 1 and Ω is the bounded open domain in Rd , d = 1, 2, 3 and

∂2 ∂2 ∂2
∂Ω is the boundary of Ω. Here Δ = ∂x 2 + ∂x2 + ∂x2 denotes the Lapla-
1 2 3
cian operator with respect to the x variable, R α
0 Dt u(t, x) denotes the left
Riemann-Liouville fractional derivative with respect to the time variable t
defined by
 t
R α 1 ∂ u(τ, x)
0 Dt u(t, x) = dτ, 0 < α < 1, (1.4)
Γ(1 − α) ∂t 0 (t − τ )α
where Γ denotes the Gamma function.
Time fractional partial differential equations have many applications in
areas such as diffusion processes, electromagnetics, electrochemistry, mate-
rial science, turbulent flow, chaotic dynamics, etc. [3], [4], [14], [15], [24],
[25], [28], [29]. Analytical solutions of time fractional partial differential
equations have been studied using Green’s functions or Fourier-Laplace
transforms [26], [23], [30], [33].
Numerical methods for fractional ordinary differential equations were
studied in, for example, [6], [7], [8], [9] [12], [13]. Numerical methods for
fractional partial differential equations were also studied by some authors.
Liu et al. [22] employed the finite difference method in both space and time
and analyzed the stability condition. Sun and Wu [31] proposed a finite
difference method for the fractional diffusion-wave equation. Langlands and
Henry [18] considered an implicit numerical scheme for fractional diffusion
equation. Lin and Xu [20] proposed a finite difference method in time and
Legendre spectral method in space. Li and Xu [19] proposed a time-space
spectral method for time-space fractional partial differential equation based
on a weak formulation and a detailed error analysis was carried out.
Recently, Ervin and Roop [10], [11] used finite element methods to find
the variational solution of the fractional advection dispersion equation, in
which the fractional derivative depends on the space, related to the nonlo-
cal operator, but the time derivative term is of first order, related to the
local operator. Adolfsson et al. [1], [2] considered an efficient numerical
method to integrate the constitutive response of fractional order viscoelas-
ticity based on the finite element method. Li et al. [16] considered a time
fractional partial differential equation by using the finite element method
and obtained error estimates in both semidiscrete and fully discrete cases.
Jiang et al. [17] considered a high-order finite element method for the time
fractional partial differential equations and proved the optimal order error
estimates.
In this paper, we will use the framework in Li and Xu [19] in which the
authors introduced suitable spaces and norms in which the time fractional
differential problem can be formulated into an elliptic problem. Using these
456 N.J. Ford, J. Xiao, Y. Yan

spaces, we introduce a finite element method for time fractional partial

differential equation and obtain the optimal order error estimates both in
semidiscrete and fully discrete cases.
The paper is organized as follows. In Section 2, we consider the exis-
tence and uniqueness of the solution of the time fractional partial differ-
ential equation. In Section 3, we introduce a time discretization scheme
and prove the error estimate. In Section 4, we consider the finite element
method and obtain the optimal order error estimates in space discretiza-
tion. Finally in Section 5, we give two numerical examples and show that
the numerical results are consistent with the theoretical results.

2. Existence and uniqueness

Let C ∞ (0, T ) denote the space of infinitely differentiable functions on
(0, T ) and C0∞ (0, T ) denote the space of infinitely differentiable functions
on (0, T ) with compact support in (0, T ). Let 0 C ∞ (0, T ) denote the space of
infinitely differentiable functions on (0, T ) with compact support in (0, T ].
Then we introduce the following Sobolev space 0 H α (0, T ), 0 < α < 1
which is the closure of 0 C ∞ (0, T ) with respect to the norm ·H α (0,T ) , where
 · H α (0,T ) denotes the norm in the usual fractional Sobolev space H α (0, T )
[21]. Further, let L2 (Ω), H 1 (Ω), H 2 (Ω) denote the usual Sobolev spaces
with corresponding norms  · L2 (Ω) ,  · H 1 (Ω) and  · H 2 (Ω) , respectively.
Denote H01 (Ω) = {v ∈ H 1 (Ω), v|∂Ω = 0} with norm  · H 1 (Ω) .
Define the space, with 0 < α < 1,
     
B α (0, T ) × Ω = 0 H α (0, T ), L2 (Ω) ∩ L2 (0, T ), H01 (Ω) .
 
Here B α (0, T ) × Ω is a Banach space with respect to the norm
 1/2
v((0,T )×Ω) = vH α ((0,T ),L2 (Ω)) + vL2 ((0,T ),H01 (Ω)) .

We have the following existence and uniqueness theorem.

Theorem 2.1. Assume that 0 < α < 1 and f ∈L2 ((0, T ) ×Ω). Then
the system (1.1) - (1.3) has a unique solution in B α (0, T ) × Ω . Further
the following stability result holds:
uB α/2 ((0,T )×Ω ≤ Cf L2 ((0,T )×Ω) . (2.1)

The proof of Theorem 2.1 can be found in [19]. For completeness, and
because we use the approach later, we will give the ideas of the proof of
this theorem here.
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 457

Recall that the right Riemann-Liouville fractional integral is defined as

 T
R α 1 ∂ v(τ )
t DT v(t) = dτ, 0 < α < 1. (2.2)
Γ(1 − α) ∂t t (t − τ )α

Definition 2.1. Define Hrα (0, T ) as the closure of C0∞ (0, T ) with
respect to the norm  · Hrα (0,T ) , that is,
  

Hrα (0, T ) = v ∈ L2 (0, T ) ∃ vn ∈ C0∞ (0, T ), such that vn −vHrα (0,T ) → 0 .

Here the norm  · Hrα (0,T ) is defined by

 1/2
vHrα (0,T ) = v2L2 (0,T ) + |v|2Hrα (0,T ) ,
where the seminorm | · |Hrα (0,T ) is defined by
|v|Hrα (0,T ) = R α
t DT vL2 (0,T ) .

Definition 2.2. Define Hcα (0, T ) as the closure of C0∞ (0, T ) with
respect to the norm  · Hcα (0,T ) , that is,
  

Hcα (0, T ) = v ∈ L2 (0, T ) ∃ vn ∈ C0∞ (0, T ), such that vn −vHcα (0,T ) → 0 .

Here the norm  · Hcα (0,T ) is defined by

 1/2
vHcα (0,T ) = v2L2 (0,T ) + |v|2Hcα (0,T ) ,
where the seminorm | · |Hcα (0,T ) is defined by
  1/2
|v|Hcα (0,T ) =  0 Dtα v, R α
t DT v L2 (0,T )
 ,

where (·, ·)L2 (0,T ) denotes the inner product in L2 (0, T ).

Lemma 2.1. Let 0 < α < 1. We have

Hrα (0, T ) = Hcα (0, T ) = H0α (0, T ),
and the norms  · Hrα (0,T ) ,  · Hcα (0,T ) and  · H0α (0,T ) are equivalent.

P r o o f. We first prove Hrα (0, T ) = Hcα (0, T ). In fact, ∀ v ∈ Hrα (0, T ),

there exists a sequence vn ∈ C0∞ (0, T ) such that
vn − vHrα (0,T ) → 0, n → ∞,
which implies that vn − vm Hrα (0,T ) → 0, n → ∞, m → ∞, that is, {vn }
is a Cauchy sequence in Hrα (0, T ). We will show that the norm  · Hrα (0,T )
is equivalent to the norm  · Hcα (0,T ) in C0∞ (0, T ). Assuming this for the
458 N.J. Ford, J. Xiao, Y. Yan

moment, we see that {vn } is also a Cauchy sequence in  · Hcα (0,T ) . Thus
there exists v  ∈ Hcα (0, T ) such that
vn − v  Hcα (0,T ) → 0, n → ∞.
Hence, we have
v − v  L2 (0,T ) ≤ v − vn L2 (0,T ) + vn − v  L2 (0,T )
≤ v − vn Hrα (0,T ) + vn − v  Hcα (0,T ) → 0, n → ∞,
(2.3)
which implies that v = v  and therefore Hrα (0, T ) ⊂ Hcα (0, T ). Similarly,
we can prove Hcα (0, T ) ⊂ Hrα (0, T ). Hence we get
Hcα (0, T ) = Hrα (0, T ).
Denote the norm equivalence by  · Hrα (0,T ) ∼ =  · Hcα (0,T ) . We now need
to prove that the norm  · Hrα (0,T ) is equivalent to the norm  · Hcα (0,T )
in C0∞ (0, T ). In fact, ∀ v ∈ C0∞ (0, T ), let ṽ be the extension of v by zero
outside (0, T ). Then we have
|v|H α (0,T ) = |ṽ|H α (R) ∼
r r
= |ṽ|H α (R) = |v|H α (0,T ) ,
c c

where we use the fact that

|v|Hcα (0,T ) = |ṽ|Hcα (R) ,
which follows from
 1/2  1/2
|ṽ|Hcα (R) =  −∞ Dtα ṽ, t D+∞
α
ṽ  =  0 Dtα ṽ, t DTα ṽ  = |v|Hcα (0,T ) .
Here we also use the fact that Hrα (R) = Hcα (R) and the norm  · Hrα (R) = ∼
 · Hcα (R) which can be found in [11].
Next we prove that the norm  · Hrα (0,T ) is equivalent to the norm
 · Hcα (0,T ) in space Hrα (0, T ) = Hcα (0, T ). In fact, following the ideas of
the proof above, ∀ v ∈ Hrα (0, T ), there exists a sequence vn ∈ C0∞ (0, T )
such that
vn − vHrα (0,T ) → 0, n → ∞,
and
vn − vHcα (0,T ) → 0, n → ∞.
Thus
vHrα (0,T ) ≤ v − vn Hrα (0,T ) + vn Hrα (0,T )
≤ v − vn Hrα (0,T ) + Cvn Hcα (0,T )
≤ v − vn Hrα (0,T ) + Cvn − vHcα (0,T ) + CvHcα (0,T ) .
Let n → ∞, we get vHrα (0,T ) ≤ CvHcα (0,T ) for any v ∈ Hcα (0, T ). Sim-
ilarly we can prove vHcα (0,T ) ≤ CvHrα (0,T ) for any v ∈ Hrα (0, T ). Thus
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 459

the norm  · Hrα (0,T ) is equivalent to the norm  · Hcα (0,T ) in Hrα (0, T ) =
Hcα (0, T ).
Finally we turn to the proof of Hcα (0, T ) = H0α (0, T ) and the norm
·Hcα (0,T ) is equivalent to the norm ·H0α (0,T ) in Hcα (0, T ) = H0α (0, T ). The
arguments of the proof are the same as the proof for Hrα (0, T ) = Hcα (0, T )
above. Therefore it suffices to prove that the norm  · Hcα (0,T ) is equivalent
to the norm  · H0α (0,T ) in C0∞ (0, T ) which follows from
 
|v|Hcα (0,T ) = |ṽ|Hcα (R) ∼
= |ṽ|Hcα (R) = F t DTα ṽ L2 (R)
= (iw)α F(ṽ)L (R) ∼2 = |ṽ|H α (R) = |v|H α (0,T ) .
Here we used the Plancherel Theorem [26] and remark that F(ṽ) denotes
the Fourier transform of ṽ. 2

Lemma 2.2. [26] We have:

(1) If 0 < p < 1, 0 < q < 1, v(0) = 0, t > 0, then
p+q      
0 Dt v(t) = 0 Dtp 0 Dtq v(t) = 0 Dtq 0 Dtp v(t), ∀ w ∈ H p+q (0, T ).
(2) Let 0 < α < 1. Then we have
   
Dα w, v L2 (0,T ) = w, t DTα v L2 (0,T ) ,
0 t
∀ w ∈ H α (0, T ), v ∈ C0∞ (0, T ).

Lemma 2.3. Let 0 < α < 1. Then for any w ∈ 0 H α (0, T ), v ∈

α/2 (0, T ), we have
0H
 α   α/2 α/2 
D w, v L2 (0,T ) = 0 Dt w, t DT v L2 (0,T ) .
0 t

P r o o f. Since 0 < α < 1, we have [21]

α/2 α/2
0H (0, T ) = H0 (0, T ).
Thus, ∀ v ∈ 0 H α/2 (0, T ), there exists a sequence vn ∈ C0∞ (0, T ) such that
v − vn H α/2 (0,T ) → 0, n → ∞.
By Lemma 2.2, we have, for any w ∈ 0 H α (0, T )
with w(0) = 0,
 α    
α/2 α/2
D w, vn L2 (0,T ) =
0 t
D
0 t
D
0 t
w, vn
L2 (0,T )
 α/2 α/2 
= D w, t DT vn L2 (0,T ) .
0 t
We now prove that
 α   
D w, vn L2 (0,T ) → 0 Dtα w, v L2 (0,T ) , n → ∞,
0 t
(2.4)
 α/2 α/2   α/2 α/2 
D w, t DT vn L2 (0,T ) → 0 Dt w, t DT v L2 (0,T ) ,
0 t
n → ∞, (2.5)
460 N.J. Ford, J. Xiao, Y. Yan

It is easy to prove (2.4). For (2.5), we have

 
 α/2 α/2   α/2 α/2  
 0 t
D w, D
t T vn L2 (0,T ) − D
0 t
w, D
t T v L2 (0,T ) 
α/2 α/2 α/2
≤ 0 Dt wL2 (0,T ) · t
DT vn − t DT v L2 (0,T )
α/2
≤ 0 Dt wL2 (0,T ) · |vn − v|H α/2 (0,T )
r
α/2
≤ C0 Dt wL2 (0,T ) · vn − vH α/2 (0,T ) → 0, n → ∞,
where we used Lemma 2.1 in the last inequality.
Together, these estimates complete the proof of the Lemma 2.3. 2

P r o o f o f T h e o r e m 2.1. The weak formulation of (1.1)-(1.3) is to

find u ∈ B α/2 (0, T ) × Ω such that
 
A(u, v) = F(v), ∀ v ∈ B α/2 (0, T ) × Ω , (2.6)
where the bilinear forms A(·, ·) and F(v) are defined by using Lemma 2.3,
 
A(u, v) = 0 Dtα u, v L2 ((0,T )×Ω) + (∇u, ∇v)L2 ((0,T )×Ω)
 α/2 α/2 
= 0 Dt u, t DT v L2 ((0,T )×Ω) + (∇u, ∇v)L2 ((0,T )×Ω) ,
and
F(v) = (f, v)L2 ((0,T )×Ω) .
It is easy to prove the continuities of the bilinear form A(·, ·) and the right
hand functional F(v), that is, there exists a constant C > 0, such that
|A(u, v)| ≤ CuB α/2 ((0,T )×Ω) vB α/2 ((0,T )×Ω) ,
and
|F(v)| ≤ f L2 ((0,T )×Ω) vB α/2 ((0,T )×Ω) . (2.7)

 next prove the coercivity of the bilinear operator A(·, ·) on B
We α/2 (0, T )×
Ω . [20]
Note that
 α/2 α/2   α/2 α/2

D
0 t
ϕ, t DT ϕ L2 (0,T ) = ∞ Dt ϕ̃, t D∞ ϕ̃ L2 (R)
α 
∀ ϕ ∈ C0∞ (0, T ),
α/2
= cos π · ∞ Dt ϕ̃L2 (R) ,
2
where ϕ̃ is the extension of ϕ by zero outside of (0, T ). Thus we find that
 α/2 α/2 
D v, t DT v L2 (0,T ) is nonnegative for v ∈ H α/2 (0, T ), 0 < α < 1 since
0 t 
cos α2 π is nonnegative for 0 < α < 1.
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 461

Combining this with Lemma 2.1, we have

 α/2 α/2 
A(v, v) = 0 Dt v, t DT v L2 ((0,T )×Ω) + (∇v, ∇v)L2 ((0,T )×Ω)
 α/2 α/2 
≥ C 0 Dt v, 0 Dt v L2 ((0,T )×Ω) + (∇v, ∇v)L2 ((0,T )×Ω)
≥ CvB α/2 ((0,T )×Ω) , (2.8)
where we applied the Poincaré inequalities in the last inequality, that is
∇ϕL2 (Ω) ∼
= ϕH 1 (Ω) , ∀ ϕ ∈ H01 (Ω),
and
Dt
α/2
ψ ∼
= ψH α/2 (0,T ) , ∀ ψ ∈0 H α/2 (0, T ).
0 L2 (0,T )
By using the well-known Lax-Milgram Lemma, there exists a unique
solution u ∈ B α/2 ((0, T ) × Ω) such that (2.6) holds.
To prove the stability estimate (2.1), we take v = u in (2.6) to get, by
using (2.8) and (2.7),
CuB α/2 ((0,T )×Ω) ≤ A(u, u) = F(u) ≤ Cf L2 ((0,T )×Ω) uB α/2 ((0,T )×Ω) ,
which implies that
uB α/2 ((0,T )×Ω) ≤ Cf L2 ((0,T )×Ω) .
The proof is complete. 2

3. Time discretization
In this section, we will consider the time discretization of (1.1)- (1.3).
Define A = −Δ, D(A) = H01 (Ω) ∩ H 2 (Ω). Then the system (1.1)-(1.3)
can be written in the abstract form
R α
0 Dt u(t) + Au(t) = f (t), 0 < t < T, 0 < α < 1, (3.1)
u(0) = u0 = 0, (3.2)
or, equivalently,
R α
0 Dt [u − u0 ](t) + Au(t) = f (t), 0 < t < T, 0 < α < 1. (3.3)
Note that
 t  t
R α 1 d u(τ ) 1 u(τ )
0 Dt u(t) = dτ = dτ,
Γ(−α) dt 0 (t − τ )α Γ(−α) 0 (t − τ )α+1
where the integral must be interpreted as a Hadamard finite-part integral,
see [6].
Let 0 = t0 < t1 < · · · < tn = T be a partition of [0, T ]. Then, for fixed
tj , j = 1, 2, . . . n, we have
462 N.J. Ford, J. Xiao, Y. Yan


1 u(τ ) − u(0)
tj
R α
0 Dt [u − u0 ](tj ) = dτ
Γ(−α) 0 (t − τ )1+α
  1
t−α 1
u(tj − tj w) − u(0) t−α
g(w)w−1−α dw,
j j
= dw =
Γ(−α) 0 w1+α Γ(−α) 0
where g(w) = u(tj − tj w) − u(0).
Now, for every j, we replace the integral by a first-degree compound
quadrature formula with the equispaced nodes 0, 1j , 2j , . . . , jj and obtain
 1 j
−1−α
g(w)w dw = αkj g(k/j) + Rj (g),
0 k=0

where the weights αkj satisfy that ([6])

⎧
⎪
⎨ −1, for k = 0,
−α 1−α 1−α
α(1−α)j αkj = 2k −(k−1) −(k+1)1−α , for k = 1, 2, . . . , j −1,
⎪
⎩
(α−1)k−α −(k−1)1−α +k1−α , for k = j,
and the remainder term Rj (g) satisfies
Rj (g) ≤ Cj α−2 sup g (t).
0≤t≤T

Thus we have
t−α  j
 
j
R α
0 Dt [u − u0 ](tj ) = αkj u(tj − tk ) − u(0) + Rj (g)
Γ(−α)
k=0
j
  t−α
= Δt−α
j
wkj u(tj − tk ) − u(0) + Rj (g),
Γ(−α)
k=0
where
⎧
⎪
⎨ 1, for k = 0,
Γ(2 − α)wkj = −2k 1−α +(k−1)1−α +(k + 1)1−α , for k = 1, 2, . . . , j −1,
⎪
⎩
−(α−1)k +(k−1)1−α −k1−α , for k = j.
−α

Let t = tj . We can write (3.3) as

j
  t−α
Δt−α
j
wkj u(tj−tk )−u(0) +Au(tj ) = f (tj )− Rj (g), j = 1, 2, 3, . . . .
Γ(−α)
k=0
(3.4)
Denote U j ≈ u(tj ) as the approximation of u(tj ). We can define the
following time stepping method
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 463

j
−α
 
Δt U j−k − U 0 + AU j = fj , fj = f (tj ). (3.5)
k=0
Let ej = U j − u(tj ) denote the error. Then we have the following error
estimate:

Theorem 3.1. Let U n and u(tn ) be the solutions of (3.5) and (3.1),
respectively. Then we have
U n − u(tn ) ≤ CΔt2−α .

In order to prove this theorem, we need the following lemma.

Lemma 3.1. ([6]) For 0 < α < 1, let the sequence {dj } j = 1, 2, . . . be
given by d1 = 1 and
j−1
−α
dj = 1 + α(1 − α)j αkj dj−k , j = 2, 3, · · · ,
k=1
where αkj is as in (3.4). Then,
sin πα α
1 ≤ dj ≤ j , j = 1, 2, . . . .
πα(1 − α)

P r o o f o f T h e o r e m 3.1. Subtracting (3.5) from (3.4), we get the

error equation
j
  t−α
Δt−α wkj ej−k − e0 + Aej = −
j
Rj (g),
Γ(−α)
k=0
or
 −1  −α j t−α
ej = Δt−α w0j + A
j
Δt wkj ej−k − Rj (g)
Γ(−α)
k=1
 j
= (α0j + Γ(−α)tαj A)−1 αkj ej−k − Rj (g) .
k=1
Thus,
 j
ej  ≤ (α0j + Γ(−α)tαj A)−1 αkj ej−k  + Rj (g) .
k=1
Note that A is a positive definite elliptic operator, we have, since α0j < 0
and Γ(−α) < 0,
 
 
(α0j + Γ(−α)tαj A)−1 = sup (α0j + Γ(−α)tαj λ)−1  < −α−1 0j .
λ>0
464 N.J. Ford, J. Xiao, Y. Yan

Hence,
 j
e  ≤
j
−α−1
0j αkj ej−k  + Cj α−2 n−2 sup u (t)
k=1 0≤t≤T
j
−α
= α(1 − α)j αkj ej−k  + α(1 − α)Cn−2 sup u (t).
k=1 0≤t≤T

Note that e0 = u(0) − U 0 = 0. Denote

d1 = 1,
j−1
dj = 1 + α(1 − α)j −α αkj dj−k , j = 2, 3, . . . , n.
k=1

Then we have by induction,

ej  ≤ Cα(1 − α)n−2 sup u (t) · dj .
0≤t≤T

By Lemma 3.1, we get

sin πα
ej  ≤ C sup u (t)j −α n−2 ≤ CΔt2−α .
π 0≤t≤T
The proof is complete. 2

4. Space discretization
In this section, we will consider the space discretization of (1.1) - (1.3).

The variational form of (1.1) - (1.3) is to find u(t) ∈ H01 (Ω), such that,
 α
D u(t), v)L2 (Ω) + (∇u(t), ∇v)L2 (Ω) = (f (t), v)L2 (Ω) , ∀ v ∈ H01 (Ω).
0 t
(4.1)
Let T denote a partition of Ω into disjoint triangles such that no vertex
of any triangle lies on the interior of a side of another triangle and such
that the union of the triangles determines a polygonal domain Ωh ⊂ Ω with
boundary vertices on ∂Ω. Let h denote the maximal length of the sides of
the triangulation Th . We assume that the triangulations are quasiuniform
in the sense that the triangles of Th are of essentially the same size.
Let r be any nonnegative integer. We denote by  · H r (Ω) the norm in
H (Ω). Let Sh ⊂ H01 be a family of finite element space with the accuracy
r

of order r ≥ 2, i.e., Sh consists of continuous functions on the closure Ω̄ of

Ω which are polynomials of degree at most r − 1 in each triangle of Th and
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 465

which vanish outside Ωh , such that, for small h, with v ∈ H s (Ω) ∩ H01 (Ω)
[32]

infχ∈Sh v − χL2 (Ω) + h∇(v − χ)L2 (Ω) ≤ Chs vL2 (Ω) , for 1 ≤ s ≤ r.
(4.2)
The semidiscrete problem of (4.1) is to find the approximate solution uh (t) =
uh (·, t) ∈ Sh for each t such that
 α
D u (t), χ)L2 (Ω) + (∇uh (t), ∇χ)L2 (Ω) = (f (t), χ)L2 (Ω) , ∀ χ ∈ Sh . (4.3)
0 t h

Let Rh : H01 (Ω) → Sh be the elliptic projection, or Ritz projection,

defined by
(∇(Rh u), ∇χ) = (∇u, ∇χ), ∀ χ ∈ Sh . (4.4)
We then have,

Lemma 4.1. ([32]) Assume that (4.2) holds. Then, with Rh defined
by (4.4) we have, with v ∈ H s (Ω) ∩ H01 (Ω),

Rh v − vL2 (Ω) + h∇(Rh v − v)L2 (Ω) ≤ Chs vH s (Ω) , for 1 ≤ s ≤ r.

Lemma 4.2. Let 0 < α < 1 and assume that w ∈ H α ((0, T ), L2 (Ω)).
We have
 T  T
 α   α/2 α/2 
D
0 t
w(t), w(t) L2 (Ω)
dt = D w(t), t DT w(t) L2 (Ω) dt
0 t
0 0
 T  α/2 α/2 
= 0
Dt w(t), 0 Dt w(t) L2 (Ω) dt.
0

P r o o f. We have, by Lemmas 2.3, 2.1,

 T
 α 
D w(t), w(t) L2 (Ω) dt
0 t
(4.5)
0
 T
 α 
= D w(x, t) w(x, t) dxdt
0 t
0 Ω
 α 
= D w(x, ·), w(x, ·) L2 (0,T ) dx
0 t
(4.6)
Ω
 α/2 α/2 
= D w(x, ·), t DT w(x, ·) L2 (0,T ) dx
0 t
Ω
T  α/2 α/2 
= 0
Dt w(t), t DT w(t) L2 (Ω) dt.
0
466 N.J. Ford, J. Xiao, Y. Yan

Similarly, we can prove

 
T   T  α/2 α/2 
Dα w(t), w(t) L2 (Ω) dt =
0 t 0
Dt w(t), 0 Dt w(t) L2 (Ω) dt.
0 0

The proof is complete. 2

We now come to the main theorem in this section.

Theorem 4.1. Let uh and u be the solutions of (4.1) and (4.3). Then
 T  T
2 2r
0 Dt u(t)2H r (Ω) dt.
α/2 α/2
0 Dt (uh (t) − u(t))L2 (Ω) dt ≤ Ch
0 0

P r o o f. We write
uh − u = θ + ρ, where θ = uh − Rh u, ρ = Rh u − u.

The second term is easily bounded by Lemma 3.1 and has the obvious
estimates
 T  T
2 2r
0 Dt u(t)2H r (Ω) dt.
α/2 α/2
0 Dt ρ(t)L2 (Ω) dt ≤ Ch (4.7)
0 0

In order to estimate θ, we note that by our definitions,

(0 Dtα θ(t), χ)L2 (Ω) + (∇θ(t), ∇χ)L2 (Ω)

= (0 Dtα (uh (t) − Rh u(t)), χ)L2 (Ω) + (∇(uh (t) − Rh u(t)), ∇χ)L2 (Ω)
= (0 Dtα uh (t), χ)L2 (Ω) + (∇uh (t), ∇χ)L2 (Ω)
− (0 Dtα Rh u(t), χ)L2 (Ω) + (∇Rh u(t), ∇χ)L2 (Ω)
= (f (t), χ)L2 (Ω) − (0 Dtα Rh u(t), χ)L2 (Ω) + (∇Rh u(t), ∇χ)L2 (Ω)
= (0 Dtα u(t), χ)L2 (Ω) − (0 Dtα Rh u(t), χ)L2 (Ω)
= ((I − Rh )0 Dtα Rh u(t), χ)L2 (Ω) = (0 Dtα ρ(t), χ)L2 (Ω) , ∀ χ ∈ Sh .

Choose χ = θ(t) and integrating on both sides with respect to t on

[0, T ], we obtain
 
T   T  
Dα θ(t), θ(t) L2 (Ω)
0 t
dt+ ∇θ(t), ∇χ L2 (Ω) dt
0 0
 T  
= 0
Dtα ρ(t), θ(t) L2 (Ω) dt.
0
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 467

By Lemma 4.2, we have, for any small > 0,

 T  T
2
∇θ(t)2L2 (Ω) dt
α/2
0 Dt θ(t)L2 (Ω) dt +
0 0
 T  T
α/2 α/2
= (0 Dtα ρ(t), θ(t))L2 (Ω) dt = (0 Dt ρ(t), 0 Dt θ(t))L2 (Ω) dt
0 0
 T 
C 0 Dt ρ(t)2L2 (Ω) dt + 0 Dt θ(t)2L2 (Ω) dt.
α/2 α/2
≤
0

For sufficiently small > 0, we get, by (4.7)

 T  T
2
∇θ(t)2L2 (Ω) dt
α/2
0 Dt θ(t)L2 (Ω) dt +
0 0
 T  T
0 Dt ρ(t)2L2 (Ω) dt 2r
u(t)2H r (Ω) dt.
α/2 α/2
≤C ≤ Ch 0 Dt (4.8)
0 0

Combining (4.7) with (4.8), we complete the proof of the theorem. 2

Corollary 4.1. Let uh and u be the solutions of (4.1) and (4.3).

Then
 T  T
2 2r
0 Dt u(t)2H r (Ω) dt.
α/2
uh (t) − u(t)L2 (Ω) dt ≤ Ch (4.9)
0 0

P r o o f. Note that, by Theorem 4.1, we have

 T  T
uh (t) − u(t)2L2 (Ω) dt = |uh (x, t) − u(x, t)|2 dxdt
0 0 Ω
  T   T
2
= |uh (x, t) − u(x, t)| dtdx ≤ |0 Dtα (uh (x, t) − u(x, t))|2 dtdx
Ω 0 Ω 0
 T  T
u(t)2L2 (Ω) dt 2r
u(t)2H r (Ω) dt,
α/2
≤ 0 Dtα (uh (t) − ≤ Ch 0 Dt
0 0

which is (4.9). The proof of the lemma is complete. 2

5. Numerical simulations
In this section, we present some numerical results by using the finite
element method for solving the time-fractional partial differential equation
(1.1) – (1.3). The numerical results are consistent with our theoretical
results. We can see that convergence rate of numerical solutions is of order
2 − α as the time stepsize tends to zero and of order 2 as the space step
size tends to zero, on condition that the exact solution is smooth.
468 N.J. Ford, J. Xiao, Y. Yan

Example 5.1. Consider

d2
R α
D
0 t u(x, t) − u(x, t) = f (x, t), t ∈ [0, T ], 0 < x < 1, (5.1)
dx2
u(x, 0) = 0, 0 < x < 1, (5.2)
u(0, t) = u(1, t) = 0, t ∈ [0, T ], (5.3)
where
2
f (x, t) = t2−α sin(2πx) + 4π 2 sin(2πx)t2 .
Γ(3 − α)
The exact solution is u(x, t) = t2 sin 2πx.

The main purpose of these experiments is to check the convergence

rate of the numerical solutions with respect to the fractional order α. We
use the linear finite element method and therefore the convergence order is
O(Δt2−α + Δx2 ).
In the first test, we fix T = 1, α = 0.5 and Δx = 0.001 which is small
enough such that the space discretization errors are negligible as compared
with the time errors. We choose stepsize Δt = 1/2i (i = 1, · · · , 5), then we
obtain Table 1 with the estimated convergence rate when α = 0.5, tending
to a limit close to 1.5. In the same way, we can plot the errors in the
logscale as functions of the log(Δt−1 ) for α = 0.2, 0.5, 0.9 in Figures 1, 2,
3 and obtain the convergence rates. Here we investigate both the L2 -norm
and the H 1 -norm in space.

Δx Δt H 1 -norm 2-norm estimated cvgce. rates

0.001 0.5000 0.01822017 0.00250060
0.001 0.2500 0.00675406 0.00092695 1.4317
0.001 0.1250 0.00245740 0.00033726 1.4586
0.001 0.0625 0.00087822 0.00012053 1.4845
0.001 0.03125 0.00030493 4.18492564e-05 1.5261

Table 1. Example 5.1, fix α = 0.5, Δx = 0.001

In Figures 1-3 , one can observe that the error curves are all nearly
straight lines. The convergence orders are the slopes of the lines respec-
tively. By Example 5.1, we can observe that the convergence order of the
method with respect to the time step is 2 − α, which have been shown in
Table 2.
On the other hand, if we fix Δt small enough, then the convergence
rate of space discretization errors can be shown clearly (see Table 3 and
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 469

α 0.1 0.2 0.3 0.4 0.5

est. cvgce. rate 1.8805 1.8343 1.6786 1.5712 1.4752
α 0.6 0.7 0.8 0.9
est. cvgce. rate 1.3818 1.2882 1.1935 1.0976

Table 2. Convergence rates in Example 5.1

Figure 1. Example 5.1, H 1 norm and L2 norm of errors

with α = 0.2, Δx = 0.001

Figure 4). In this case α = 0.5 , the limiting vale of the convergence rate
is consistent with the value of 2 that is expected from the theory.

Δx Δt H 1 -norm 2-norm est. cvge. rates

0.2500 0.001 0.79733222 0.10885742
0.1250 0.001 0.23998057 0.03202152 1.7653
0.0625 0.001 0.06266713 0.00842864 1.9257
0.03125 0.001 0.01584965 0.00215046 1.9707
0.015625 0.001 0.00397465 0.00054228 1.9875

Table 3. Example 5.1, fix α = 0.5, Δt = 0.001

470 N.J. Ford, J. Xiao, Y. Yan

Figure 2. Example 5.1, H 1 norm and L2 norm of errors

with α = 0.5, Δx = 0.001

Figure 3. Example 5.1, H 1 norm and L2 norm of errors

with α = 0.9, Δx = 0.001

Example 5.2. Consider

d2
R α
0 Dt u(x, t) − u(x, t) = f (x, t), t ∈ [0, T ], 0 < x < 1, (5.4)
dx2
u(x, 0) = 0, x ∈ Ω, (5.5)
u(0, t) = u(1, t) = 0, t ∈ [0, T ]. (5.6)

The exact solution is u(x, t) = sin πt sin πx.

 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 471

Figure 4. Example 5.1, H 1 norm and L2 norm of errors

with α = 0.5, Δt = 0.001

In this example, we fix T = 1, α = 0.5 and Δx = 0.0002. We obtain

the convergence rate in Table 4. In Figure 5, we plot the errors in logscale
as functions of time stepsize log(Δt−1 ) for α = 0.5. Here we can observe
much better convergence than predicted by the theory, and this is worthy
of further investigation.

Δx Δt H 1 -norm 2-norm est. cvge. rates

0.0002 0.5000 0.10734868 0.02591770
0.0002 0.2500 0.02757097 0.00665659 1.9611
0.0002 0.1250 0.00541156 0.00130654 2.3490
0.0002 0.0625 0.00057297 0.00013833 3.2395

Table 4. Example 5.2, fix α = 0.5, Δx = 0.002

Acknowledgements
The work of the second author was carried out during her stay at the
University of Chester, which is supported financially by China Scholarship
Council (CSC[2010]3006, No. 2010612215), P. R. China.

References
[1] K. Adolfsson, M. Enelund, and S. Larsson, Adaptive discretization of
integro-differential equation with a weakly singular convolution kernel.
Comput. Methods Appl. Mech. Engrg. 192 (2003), 5285–5304.
472 N.J. Ford, J. Xiao, Y. Yan

Figure 5. Example 5.2, H 1 norm and L2 norm of errors

with α = 0.5, Δx = 0.0002

[2] K. Adolfsson, M. Enelund, and S. Larsson, Adaptive discretization of

fractional order viscoelasticity using sparse time history. Comput. Meth-
ods Appl. Mech. Engrg. 193 (2004), 4567–4590.
[3] F. Amblard, A. C. Maggs, B. Yurke, A. N. Pargellis, and S. Leibler,
Subdiffusion and anomalous local viscoelasticity in actin networks.
Phys. Rev. Lett. 77 (1996), 4470.
[4] E. Barkai, R. Metzler, and J. Klafter, Form continuous time random
walks to the fractional Fokker-Planck equation. Phys. Rev. E 61 (2000),
132–138.
[5] M. Caputo, The Green function of the diffusion of fluids in porous media
with memory. Rend. Fis. Acc. Lincei (Ser. 9) 7 (1996), 243–250.
[6] K. Diethelm, Generalized compound quadrature formulae for finite-part
integrals. IMA J. Numer. Anal. 17 (1997), 479–493.
[7] K. Diethelm, An algorithm for the numerical solution of differential
equations of fractional order. Electronic Trans. on Numerical Analysis
5 (1997), 1–6.
[8] K. Diethelm, N.J. Ford, A.D. Freed, Yu. Luchko, Algorithms for the
fractional calculus: A selection of numerical methods. Computer Meth-
ods in Appl. Mechanics and Engineering 194 (2005), 743–773.
[9] N.J. Ford and A.C. Simpson, The numerical solution of fractional differ-
ential equations: speed versus accuracy. Numer. Algorithms 26 (2001),
333–346.
[10] V.J. Ervin and J.P. Roop, Variational formulation for the stationary
fractional advection dispersion equation. Numerical Meth. P.D.E. 22
(2006), 558–576.
 A FINITE ELEMENT METHOD FOR TIME FRACTIONAL 473

[11] V.J. Ervin and J.P. Roop, Variational solution of fractional advec-
tion dispersion equations on bounded domain in Rd . Numerical. Meth.
P.D.E. 23 (2007), 256–281.
[12] N.J. Ford and J.A. Connolly, Comparison of numerical methods for
fractional differential equations. Comm. Pure Appl. Anal. 5 (2006),
289–307.
[13] N.J. Ford and J.A. Connolly, Systems-based decomposition schemes for
the approximate solution of multi-term fractional differential equations.
J. Comput. Appl. Math. 229 (2009), 382–391.
[14] B. Henry and S. Wearne, Fractional reaction-diffusion. Physica A 276
(2000), 448–455.
[15] B. Henry and S. Wearne, Existence of turring instabilities in a two-
species fractional reaction-diffusion system. SIAM J. Appl. Math. 62
(2002), 870–887.
[16] C. Li, Z. Zhao, and Y. Chen, Numerical approximation and error
estimates of a time fractional order diffusion equation. In: Proc. of
the ASME 2009 Intern. Design Engineering Technical Conf. and Com-
puters and Information in Engineering Conf. IDETC/CIE 2009, San
Diego, California, USA.
[17] Y. Jiang and J. Ma, High-order finite element methods for time-
fractional partial differential equations. J. Comput. Appl. Math. 235
(2011), 3285–3290.
[18] T.A.M. Langlands and B.I. Henry, The accuracy and stability of an im-
plicit solution method for the fractional diffusion equation. J. Comput.
Phys. 205 (2005), 719–736.
[19] X.J. Li and C.J. Xu, Existence and uniqeness of the weak solution
of the space-time fractional diffusion equation and a spectral method
approximation. Commun. Comput. Phys. 8 (2010), 1016–1051.
[20] Y.M. Lin and C.J. Xu, Finite difference/spectral approximation for the
time fractional diffusion equations. J. Comput. Phys. 2 (2007), 1533–
1552.
[21] J.L. Lions and E. Magenes, Non-homogeneous Boundary Value Prob-
lems and Applications, Volume 1. Springer-Verlag, 1972.
[22] F. Liu, S. Shen, V. Anh and I. Turner, Analysis of a discrete non-
markovian random walk approximation for the time fractional diffusion
equation. ANZIAMJ 46 (2005), 488–504.
[23] F. Liu, V. Anh, I. Turner, and P. Zhuang, Time fractional advection
dispersion equation. J. Comput. Appl. Math. 13 (2003), 233–245.
474 N.J. Ford, J. Xiao, Y. Yan

[24] H.P. Müller, R. Kimmich, and J. Weis, NMR flow velocity mapping
in random percolation model objects: Evidence for a power-law de-
pendence of the volume-averaged velocity on the probe-volume radius.
Phys. Rev. E 54 (1996), 5278–5285.
[25] R.R. Nigmatullin, Realization of the generalized transfer equation in
a medium with fractal geometry. Physica B 133 (1986), 425–430.
[26] I. Podlubny, Fractional Differential Equations. Academic Press, San
Diego, 1999.
[27] J.P. Roop, Computational aspects of FEM approximation of fractional
advection dispersion equations on bounded domain in R2 . J. Comp.
Appl. Math. 193 (2006), 243–268.
[28] H. Scher and M. Lax, Stochastic transport in a disordered solid. Phys.
Rev. B 7 (1973), 4491–4502.
[29] H. Scher and E. Montroll, Anomalous transit-time dispersion in amor-
phous solids. Phys. Rev. B 12 (1975), 2455–2477.
[30] W.R. Schneider and W. Wyss, Fractional diffusion and wave equations.
J. Math. Phys. 30 (1989), 134-144.
[31] Z.Z. Sun and X.N. Wu, A fully discrete difference scheme for a
diffusion-wave system. Appl. Numer. Math. 56 (2006), 193–209.
[32] V. Thomée, Galerkin Finite Element Methods for Parabolic Problems.
Springer-Verlag, Berlin, 2007.
[33] W. Wyss, The fractional diffusion equation. J. Math. Phys. 27 (1989),
2782–2785.

1,3Department of Mathematics
University of Chester
Parkgate Road
Chester, CH1 4BJ, UK
1 e-mail: [email protected] Received: March 30, 2011
3 e-mail: [email protected]
2 Department of Mathematics
Harbin Institute of Technology
Harbin, 150001, P.R. CHINA
e-mail: [email protected]

Please cite to this paper as published in:

Fract. Calc. Appl. Anal., Vol. 14, No 3 (2011), pp. 454–474;
DOI: 10.2478/s13540-011-0028-2