See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/347855839

Finite Difference Method on Non-Uniform Meshes for Time Fractional

Diffusion Problem

Article in Computational Methods in Applied Mathematics · December 2020

DOI: 10.1515/cmam-2020-0077

CITATIONS READS

3 381

2 authors:

Haili Qiao Aijie Cheng

Liaocheng Universtiy Shandong University
9 PUBLICATIONS 33 CITATIONS 102 PUBLICATIONS 611 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Haili Qiao on 06 December 2021.

The user has requested enhancement of the downloaded file.

 Comput. Methods Appl. Math. 2021; 21(4): 899–911

Research Article

Haili Qiao and Aijie Cheng*

Finite Difference Method on Non-Uniform

Meshes for Time Fractional Diffusion Problem
https://doi.org/10.1515/cmam-2020-0077
Received May 12, 2020; revised October 22, 2020; accepted November 4, 2020

Abstract: In this paper, we consider the time fractional diffusion equation with Caputo fractional derivative.
Due to the singularity of the solution at the initial moment, it is difficult to achieve an ideal convergence rate
when the time discretization is performed on uniform meshes. Therefore, in order to improve the convergence
order, the Caputo time fractional derivative term is discretized by the L2 − 1σ format on non-uniform meshes,
with σ = 1 − 2α , while the spatial derivative term is approximated by the classical central difference scheme on
uniform meshes. According to the summation formula of positive integer k power, and considering k = 3, 4, 5,
we propose three non-uniform meshes for time discretization. Through theoretical analysis, different time
convergence orders O(N − min{kα,2} ) can be obtained, where N denotes the number of time splits. Finally, the
theoretical analysis is verified by several numerical examples.

Keywords: Fractional Differential Equation, Finite Difference, Weak Singularity, Non-Uniform Meshes,
Alikhanov Scheme

MSC 2010: 35R11, 65M06, 65N12

1 Introduction
Consider the following time fractional diffusion problem:

{ C α ∂2 u
{
{ 0 D t u(x, t) − 2 (x, t) = f(x, t), (x, t) ∈ (0, L) × (0, T],
{
{ ∂x
{ u(0, t) = u(L, t) = 0, t ∈ (0, T], (1.1)
{
{
{
{
{ u(x, 0) = u0 (x), x ∈ (0, L).

Here 0 < α < 1, 0C D αt u(x,t) denotes the Caputo fractional derivative for the time variable t, defined as
t
C α 1 ∂u(x, s)
0 D t u(x, t) = ∫(t − s)−α ds.
Γ(1 − α) ∂s
0

The model is similar to the normal diffusion of the Brownian motion, which is derived from the continuous-
time random walks [10, 15], and the fractional derivative term is introduced to represent the degree of mem-
ory in the diffusion material. Problem (1.1) can be used to characterize the problems of ground environment
problem [11], highly heterogeneous aquifers [1] and thermal diffusion in fractal geometry media [16].
Since equation (1.1) can describe many practical application problems, it has attracted the attention of
many scholars over several decades of years, and various effective numerical solutions have been proposed.
For spatial discretization, the finite difference method, the spectral method with high convergence order and

*Corresponding author: Aijie Cheng, School of Mathematics, Shandong University, Jinan, Shandong 250100, P. R. China,
e-mail: [email protected]
Haili Qiao, School of Mathematics, Shandong University, Jinan, Shandong 250100, P. R. China, e-mail: [email protected]
900 | H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes

the finite element method are used. For the time Caputo fractional derivative, various effective discrete for-
mats have also been studied. For example, the classical L1 scheme proposed by [12], which is divided on
uniform meshes. On each subinterval, the piecewise linear interpolation function is used to approximate the
integrable function, and the time convergence order of the format is 2 − α. The L1 format has been widely
used in practice, see [3, 8, 24, 27]. Considering that the convergence order of the L1 format is relatively low,
some scholars have proposed some high-order formats. In [25], the piecewise linear interpolation function is
used to approximate the integrable function on non-uniform meshes, and the time second-order convergence
scheme is obtained. Gao [9] derived the so-called L1 − 2 format, where the convergence order in time direc-
tion is 3 − α. The properties of coefficients in L1 − 2 format and the local truncation error are analyzed. Lv [13]
provided comprehensive analysis of the L1 − 2 format proposed in [9], including stability analysis and global
error estimates. Alikhanov [2] constructed the L2 − 1σ scheme (also known as fractional C-N scheme) with
local truncation error O(τ3−α ) in the time direction. This scheme is discretized on uniform meshes, and an
offset parameter σ = 1 − 2α is introduced, and the integral function is approximated by linear interpolation
function in the integral interval [t k , t k+σ ]. In the integral interval [t k−1 , t k ], k ≥ 1, quadratic interpolation
function with (t k−1 , u(t k−1 )), (t k , u(t k )) and (t k+1 , u(t k+1 )) are used to approximate integral function. When
the format is applied to problem (1.1), and the diffusion term is discretized at the non-integral time node t k+σ ,
which is approximated by two nodes t k and t k+1 , the time convergence order is O(τ2 ).
For the numerical schemes proposed above, in order to achieve their own ideal convergence order, the
equation solution u is usually required to possess regularity of C2 [0, T] or C3 [0, T] for the time variable. How-
ever, this assumption is not always satisfied. Usually, the singularity occurs at the initial moment [17, 18], that
is, when the time t approaches 0, the partial derivative ∂u ∂t tends to ∞. In this case, it is not possible to achieve
the ideal convergence order by using the numerical scheme proposed in the above literature on uniform
meshes. There are several alternative approaches to obtain the optimal convergence order when the solu-
tions are not sufficiently smooth. For instance, Approach 1 is to use the non-uniform meshes which capture
the singularity of the solution near the origin [4, 5, 19]; Approach 2 is to correct some initial steps to capture
the singularity of the solution near the origin [7, 21, 22]; Approach 3 is use some nonpolynomial (or singular)
basis functions or collocation spectral methods to capture the singularity of the solution [6, 14, 23, 26]. In
this paper, we mainly focus on the non-uniform meshes method, the following is a detailed introduction to
the non-uniform meshes method related literature. In [19], the central difference formula is used to approxi-
mate the diffusion term, and the L1 format is used to approximate the time Caputo fractional derivative term
on graded meshes. The discrete format of problem (1.1) converges with rate O(N − min{rα,2−α} + h2 ), where r
is the parameter of the graded meshes. In [5], the spatial derivative is discretized by the Legendre spec-
tral method, and the time derivative is discretized on fitted meshes by L2 − 1σ format, while the diffusion
term is discretized at the non-integral time node t k+σ , where t k and t k+1 are still used to approximate t k+σ ,
then the convergence order of the time direction is prompted to O(N − min{rα,2} ). In [4], the spatial derivative
is discretized by the spectral method, the time derivative term is discretized by the L2 − 1σ format on graded
meshes, and the diffusion term is discretized at the non-integral time node t k+σ , with three-points approxi-
mation regarding t k−1 , t k and t k+1 , the time convergence order is improved to O(N − min{rα,3−α} ) by numerical
investigation without theoretical proof. There are other methods, we will not list them one by one.
In this paper, we notice the singularity of the solution. In order to obtain a higher convergence order
scheme, we will use the L2 − 1σ scheme to discretization the time fraction derivative term on non-uniform
meshes, and the diffusion term is discretized on uniform grid by second-order central difference quotient.
We propose three non-uniform meshes based on the k (positive integer) power summation formula, with
different k = 3, 4, 5 respectively. By means of theoretical analysis, we can obtain different time convergence
orders O(N − min{kα,2} ) for different non-uniform meshes, where N denotes the number of time components.
When the diffusion term is discretized at t k+σ , we use the value of t k and t k+1 to approximate, and the stability
analysis and error estimates are conducted.
The structure of this paper is organized as follows. In the second section, we give the numerical solu-
tion format for problem (1.1). In the third section, we perform stability analysis and error estimates. In the
fourth section, numerical examples are given to verify the theoretical analysis. A conclusion is made in the
last section.
 H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes | 901

2 The Discrete Problem

In this section, we discretize problem (1.1). For the time fractional derivative term, we use the L2 − 1σ format
discretization on non-uniform meshes, for the spatial diffusion term, the standard center difference format is
used on uniform meshes, thus leads to the numerical solution format.
We take a positive integer M, let h = ML , and get the discrete point set in the spatial direction

{x m | x m = mh, m = 0, 1, 2, . . . , M}.
4T
Take a positive integer N, define μ = N 2 (N+1)2
, and get the discrete point set in the time direction
󵄨󵄨 n2 (n + 1)2
{t n 󵄨󵄨󵄨󵄨 t n = μ, n = 0, 1, 2, . . . , N},
󵄨 4
2 2
then τ n = t n − t n−1 = n3 μ. We propose this non-uniform mesh based on 13 + 23 + ⋅ ⋅ ⋅ + k3 = k (k+1) 4 . Let u nm
and U m n denote the true and approximate solutions of the function u(x, t) at the grid point (x , t ), respec-
m n
tively.
Problem (1.1) will be discretized at the point (x m , t n+σ ), m = 0, 1, 2, . . . , M, n = 0, 1, 2, . . . , N − 1, where
σ = 1 − 2α , and t n+σ = σt n+1 + (1 − σ)t n . Define
v(t j+1 ) − v(t j )
δ t v(t j ) = , j = 0, 1, 2, . . . , N − 1.
τ j+1
Caputo fractional derivative is discretized by the classic L2 − 1σ format on non-uniform meshes, then we
obtain the following discrete format:
n
C α j
0 D t u(x m , t n+σ ) ≈ δ αtn+σ u m := g n,n u n+1
m − ∑ (g n,j − g n,j−1 )u m . (2.1)
j=0

Here g n,−1 := 0, g0,0 = τ−1

1 a 0,0 , n = 1, 2, . . . , N − 1, and g n,j are defined as follows:

{
{ τ−1
j+1 (a n,0 − b n,0 ) for j = 0,
{ −1
g n,j = {τ j+1 (a n,j + b n,j−1 − b n,j ) for j = 1, 2, . . . , n − 1,
{
{ −1
{τ j+1 (a n,n + b n,n−1 ) for j = n,

and
σ1−α 1−α
a n,n = τ for 0 ≤ n ≤ N − 1,
Γ(2 − α) n+1
t j+1
1
a n,j = ∫ (t n+σ − s)−α ds for 1 ≤ n ≤ N − 1, 0 ≤ j ≤ n − 1,
Γ(1 − α)
tj
t j+1
1 2
b n,j = ∫ (t n+σ − s)−α (s − t j+ 21 ) ds for 1 ≤ n ≤ N − 1, 0 ≤ j ≤ n − 1.
Γ(1 − α) t j+1 − t j
tj
∂2 u
For the diffusion term ∂x2
, we use the classical second-order central difference quotient formula approxi-
mation, that is,
∂2 u n,σ
(x m , t n+σ ) ≈ δ2x u m := σδ2x u n+1 2 n
m + (1 − σ)δ x u m , (2.2)
∂x2
where
u km+1 − 2u km + u km+1
δ2x u km = , k = 0, 1, 2, . . . , N.
h2
Combining formulas (2.1) and (2.2), we get the approximate discrete format of problem (1.1):
n,σ
{ δ αt U m − δ2x U m = f mn+σ , 1 ≤ m ≤ M − 1, 1 ≤ n ≤ N − 1,
{
{ n+σ
{ U0n = U M
n
= 0, 0 ≤ n ≤ N, (2.3)
{
{ 0
{ U m = u0 (x m ), 0 ≤ m ≤ M.
902 | H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes

3 Stability and Error Estimates

In order to carry out theoretical analysis, we first give the inner products, norms, notations and related lem-
mas.
For two gird functions u, v ∈ V h0 = {v | v = (v0 , v1 , . . . , v M ), v0 = v M = 0}, let
u j − u j−1
δ x u j− 12 = , j = 1, 2, . . . , M,
h
and the discrete inner products and norms are defined as follows:
M−1 M
(u, v) = h ∑ u j v j , (δ x u, δ x v) = h ∑ δ x u j− 12 δ x v j− 12 ,
j=1 j=1

and
‖u‖2 = (u, u), ‖δ x u‖2 = (δ x u, δ x v), ‖u‖∞ = max |u m |.
1≤m≤M−1

It is easy to get that

(δ2x u, v) = −(δ x u, δ x v) = −‖δ x u‖2 = (u, δ2x v).

Lemma 1 ([20]). The following inequality holds for any gird function v ∈ V h0 :
2
‖v‖ ≤ ‖δ x v‖.
L
For any function v ∈ C[0, T] ∩ C3 (0, T], define the following notations (see [5]):

ϕ σv = τ1α sup (η1−α |δ t v(t0 ) − v󸀠 (η)|),

η∈(0,t1 )

ϕ k+σ
v = τ3−α α
k+1 t k+σ sup |v󸀠󸀠󸀠 (η)| for 1 ≤ k ≤ N − 1,
η∈(t k ,t k+1 )
k,1
ϕv = τ1α sup (η1−α |(I2,1 v(η))󸀠 − v󸀠 (η)|) for 1 ≤ k ≤ N − 1,
η∈(0,t1 )
k,j α 2
ϕ v = τ−α
k+1 t j τ j (τ j + τ j+1 ) sup |v󸀠󸀠󸀠 (η)| for 2 ≤ j ≤ k ≤ N − 1.
η∈(t j−1 ,t j+1 )

Here I2,j v(t) denotes the quadratic interpolation function of the function v(t) regarding points t j−1 , t j and t j+1 ,
i.e.
δ t v(t j ) − δ t v(t j−1 )
I2,j v(t) = v(t j−1 ) + δ t v(t j−1 )(t − t j−1 ) + (t − t j−1 )(t − t j ).
t j+1 − t j−1
Thus, it is easy to obtain
δ t v(t j ) − δ t v(t j−1 )
(I2,j v(t))󸀠 = δ t v(t j−1 ) + (2t − t j−1 − t j ). (3.1)
t j+1 − t j−1
τ n+1
For any 1 ≤ n ≤ N − 1, define ρ n := τn , and it is easy to prove 1 ≤ ρ n+1 ≤ ρ n . Thus, we can get the fol-
lowing lemmas.

Lemma 2 ([5]). The coefficient g n,k in the discrete format of the fractional derivative 0C D αt u satisfies the following
relationship:
t−α
n+σ
g n,0 > >0 for n ≥ 0,
Γ(1 − α)
(2σ − 1)g n,n − σg n,n−1 > 0 for n > 0,

g n,k > g n,k−1 for n > 0.

Lemma 3 ([5]). For any function v(t) ∈ C3 (0, T], there is a normal number C, which makes
n,k
|δ αn+σ V − 0C D αt v(t n+σ )| ≤ Ct−α n+σ
n+σ (ϕ v + max ϕ v )
1≤k≤n

for any n ∈ {0, 1, 2, . . . , N − 1} is established.

 H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes | 903

Lemma 4 ([5]). For any grid function {V n }Nn=0 , we have

|V n+1 | ≤ |V 0 | + Γ(1 − α) max t αk+σ δ αtk+σ |V|
0≤k≤n

for n = 0, 1, 2, . . . , N − 1.

Lemma 5 ([5]). For any grid function {V n }Nn=0 , we obtain

1 α
(σV n+1 + (1 − σ)V n , δ αtα V) ≥ δ ‖V‖2
n+σ 2 t n+σ
for n = 0, 1, 2, . . . , N − 1.

Lemma 6 ([5]). For any function v(t) ∈ C2 (0, T], by Taylor expansion, we get
1 2
|σv(t n+1 ) + (1 − σ)v(t n ) − v(t n+σ )| ≤ τ max |v󸀠󸀠 (t)|
8 n+1 t∈(t n ,t n+1 )
for n = 0, 1, 2, . . . , N − 1.

Lemma 7. Suppose there exists a constant C1 such that |v(l) | ≤ C(1 + t α−l ), l = 0, 1, 2, 3. Then there exists a con-
stant C2 so that
ϕ n+σ
v ≤ C2 N − min{4α,3−α} for 0 ≤ n ≤ N − 1, (3.2a)
n,k
ϕv ≤ C2 N − min{4α,3−α} for 1 ≤ k ≤ n ≤ N − 1. (3.2b)
Proof. When n = 0, by the assumption, we have
t1 t1

τ1α sup (η 1−α

|δ t v(t0 )|) ≤ ∫ |v (s)| ds ≤ C ∫ s α−1 ds ≤ Ct1α ≤ CN −4α
󸀠
(3.3)
η∈(0,t1 )
t0 t0

and
τ1α sup (η1−α |v󸀠 (η)|) ≤ Cτ1α ≤ CN −4α . (3.4)
η∈(t0 ,t1 )

Combining equations (3.3) and (3.4), we get ϕ σv ≤ CN −4α , n = 1, 2, . . . , N − 1. According to the assumption,
we obtain
ϕ n+σ
v = τ3−α α
n+1 t n+σ sup |u󸀠󸀠󸀠 (η)|
η∈(t n ,t n+1 )

(n + 1)2 (n + 2)2 α n2 (n + 1)2 α−3

≤ Cτ3−α α α−3
n+1 t n+σ t n ≤ C[(n + 1)3 μ]3−α [ μ] [ μ]
4 4
n 5α−3 −(3−α) (3.5)
≤ Cn5α−3 μ α ≤ Cn5α−3 N −4α ≤ C( ) N
N
{N −4α , α < 0.6,
≤{
N −(3−α) , α ≥ 0.6.
{
Combining (3.3), (3.4) and (3.5), we get (3.2a).
n,1
For ϕ v , using formula (3.1) and hypothetical conditions, we can imitate the proof of ϕ σv to obtain
n,1
ϕ v ≤ CN −4α . For 2 ≤ k ≤ n ≤ N − 1, we get
n,k α 2
ϕu = τ−α
n+1 t k τ k (τ k + τ k+1 ) sup |u󸀠󸀠󸀠 (η)|
η∈(t k−1 ,t k+1 )
α 2 α−3
≤ Cτ−α
n+1 t k τ k (τ k + τ k+1 )t k−1
k2 (k + 1)2 α k2 (k − 1)2 α−3
≤ C(n + 1)−3α μ−α k6 μ2 [k3 μ + (k + 1)3 μ][ μ] [ μ]
4 4
k 5α−3 −(3−α)
≤ Ck5α−3 μ α ≤ C( ) N
N
{N −4α , α < 0.6,
≤{
N −(3−α) , α ≥ 0.6.
{
Based on the above analysis, we complete the proof of (3.2b).
904 | H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes

In the following two theorems, we will give the stability analysis and error estimates of the format (2.3),
respectively.
n be the solution of equation (2.3). Then we have
Theorem 1. Let U m
Γ(1 − α)LT α
‖U n+1 ‖2 ≤ ‖U 0 ‖2 + max ‖f k+σ ‖2 for 0 ≤ n ≤ N − 1.
8 0≤k≤N−1
n,σ
Proof. Multiplying both sides of equation (2.3) by U m , and making summation from 1 to M − 1 with respect
to m, we obtain
(δ αtn+σ U, U n,σ ) + (δ x U n,σ , δ x U n,σ ) = (f n+σ , U n,σ ).
According to Lemma 5, we have
1 α
δ ‖U‖2 + ‖δ x U n,σ ‖2 ≤ (f n+σ , U n,σ ). (3.6)
2 t n+σ
According to the Cauchy–Schwarz inequality and Young’s inequality, we have
2 n,σ 2 L n+σ 2
(f n+σ , U n,σ ) ≤ ‖f n+σ ‖ ⋅ ‖U n,σ ‖ ≤
‖U ‖ + ‖f ‖ .
L 16
n+σ 2
Thus, by Lemma 1, inequality (3.6) becomes δ αtn+σ ‖U‖2 ≤ L‖f 8 ‖ . Then, according to Lemma 4, we get
‖U n+1 ‖2 ≤ ‖U 0 ‖2 + Γ(1 − α) max {t αk+σ δ αtk+σ ‖U‖2 }
0≤k≤n
Γ(1 − α)LT α
0 2
≤ ‖U ‖ + max ‖f k+σ ‖2 .
8 1≤k≤N−1

Theorem 2. Let u nm and U m

n be solutions of equations (1.1) and (2.3), respectively. Assuming

‖∂ lt u(t)‖ ≤ C(1 + t α−l ), l = 0, 1, 2, 3,

we have
max ‖u k − U k ‖ ≤ C{N − min{4α,2} + M −2 }.
1≤k≤N

Proof. Let e nm = u nm − U m n denote the error of the true solution and the approximate solution at the point

(x m , t n ). It is clear that e0n = 0. From equations (1.1) and (2.3), we get

n,σ n,σ ∂2 u
δ αtn+σ e m −δ2x e m = (δ αtn+σ u − 0C D αt u(x m , t n+σ )) + (δ2x u m
n+σ
− δ2x u m ) + ( n+σ
(x m , t n+σ ) − δ2x u m )
∂x2
=: r n+σ n+σ n+σ
m + Pm + Rm .
n,σ
Multiplying both sides of the above formula by e m = σe n+1 n
m + (1 − σ)e m , and making summation from m = 1
to m = M − 1, we obtain
(δ αtn+σ e, e n,σ ) + (δ x e n,σ , δ x e n,σ ) = (r n+σ + P n+σ + R n+σ , e n,σ ).
According to the Cauchy–Schwarz inequality and Lemma 5, we have
1 α
δ ‖e‖2 + ‖δ x e n,σ ‖2 ≤ (‖r n+σ ‖ + ‖P n+σ ‖ + ‖R n+σ ‖)‖e n,σ ‖,
2 t n+σ
thus
δ αtn+σ ‖e‖2 ≤ 2(‖r n+σ ‖ + ‖P n+σ ‖ + ‖R n+σ ‖) max ‖e j ‖
1≤j≤N

According to Lemma 4, we obtain

‖e k+1 ‖2 ≤ C max {t αn+σ δ αtn+σ ‖e‖2 } ≤ C max {t αn+σ (‖r n+σ ‖ + ‖P n+σ ‖ + ‖R n+σ ‖)} max ‖e j ‖.
0≤n≤k 0≤n≤N−1 1≤j≤N

Then we use the above formula by recurrence to get

max ‖e k ‖ ≤ C max {t αn+σ (‖r n+σ ‖ + ‖P n+σ ‖ + ‖R n+σ ‖)}. (3.7)
1≤k≤N 0≤n≤N−1

Similar to the derivation of [5, equation (29)], we can prove

max {t αn+σ ‖r n+σ ‖} ≤ CN − min{4α,3−α} . (3.8)
0≤n≤N−1
 H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes | 905

According to Lemma 6, we obtain

max {t αn+σ ‖P n+σ ‖} ≤ Ct αn+1 τ2n+1 t α−2
n
0≤n≤N−1
(n + 1)2 (n + 2)2 α n2 (n + 1)2 α−2
≤ C[ μ] (n + 1)6 μ2 [ μ]
4 4
n 8α−2 −2 (3.9)
≤ Cn8α−2 N −8α ≤ C( ) N
N
{CN −8α for α < 0.25,
≤{
CN −2 for α ≤ 0.25.
{
On the other hand, we have
max {t αn+σ ‖R n+σ ‖} ≤ CT α M −2 . (3.10)
0≤n≤N−1
Combining inequalities (3.7), (3.8), (3.9) and (3.10), we get

max ‖u k − U k ‖ ≤ C{N − min{4α,2} + M −2 }.

1≤k≤N

Remark. For different non-uniform meshes, there will be different convergence results. We call the mesh
mentioned above as Mesh1. Next, we present two other non-uniform meshes, called Mesh2 and Mesh3,
respectively.
6T
∙ Mesh2: Let μ = N(N+1)(2N+1) , t n = n(n+1)(2n+1)
6 μ, τ n = t n − t n−1 = n2 μ. Then the following convergence
result can be obtained for discrete format (2.3) on Mesh2:

max ‖u k − U k ‖ ≤ C(M −2 + N − min{3α,2} ).

1≤k≤N

This non-uniform mesh is proposed according to the formula 12 + 22 + ⋅ ⋅ ⋅ + k2 = k(k+1)(2k+1) 6 .

30T n(n+1)(2n+1)(3n2 +3n−1) 4 μ. Then the follow-
∙ Mesh3: Let μ = N(N+1)(2N+1)(3N 2 +3N−1) , t n = 30 μ, τ n = t n − t n−1 = n
ing convergence result can be obtained for discrete format (2.3) on Mesh3:

max ‖u k − U k ‖ ≤ C(M −2 + N − min{5α,2} ).

1≤k≤N
k(k+1)(2k+1)(3k2 +3k−1)
This non-uniform mesh is proposed according to the formula 14 + 24 + ⋅ ⋅ ⋅ + k4 = 30 .

The stability analysis and error estimates for the discrete scheme (2.3) with Mesh2 and Mesh3 can be proved
analogously. Here, we no longer proved in detail, but we will verify in numerical simulation.

4 Numerical Experiments
In this section we use several numerical examples to verify the previous theoretical analysis. It is well known
that for the spatial second-order central difference scheme, the convergence order is O(M −2 ), so we do not
verify it. The following examples only verify the time convergence rates.
Define
Max_err := max ‖u n − U n ‖∞ and L2_err := max ‖u n − U n ‖.
1≤n≤N 1≤n≤N

Example 1. Consider

{ C α ∂2 u
{
{ 0 D t u(x, t) − 2 (x, t) = f(x, t), (x, t) ∈ (0, π) × (0, 1],
{
{ ∂x
{
{ u(0, t) = u(π, t) = 0, t ∈ (0, 1],
{
{
{
{ u(x, 0) = u0 (x), x ∈ [0, π].
α 1+α
t t
We assume that the exact solution of the problem is u(x, t) = (1 + Γ(1+α) + Γ(2+α) ) sin(x), according to the
equation, the corresponding f(x, t) and u0 (x) can be defined. Obviously, u(x, t) satisfies the assumption in
the theorem, that is, at the initial moment, the solution has singularity.
906 | H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes

For Example 1, we use the discrete format (2.3) to solve the problem on non-uniform Mesh1. Considering that
the spatial convergence order is dominant, we take M = N to verify the time convergence order for different α.
In Tables 1 and 2, the maximum node error and its convergence order, and the maximum L2 error and its
convergence order are given, respectively. The datas demonstrate that the experimental simulation results
are consistent with the theoretical analysis.
Next, for comparison with other methods, we use the discrete formula (2.3) to solve Example 1 on the
standard graded mesh. The set of time direction grid nodes is defined as {t n | t n = T( Nn )r }, with r ≥ 1. We set
M = N, and select r = 2, 1α , and 2α , respectively, to implement numerical tests for different α. The results are
listed in Tables 3–5. These information show that the time convergence order is about min{rα, 2} for graded
meshes. When r = 2, as the mesh is refined, the time convergence order approaches min{2α, 2}, which is less
than the convergence order min{4α, 2} for Mesh1. For smaller α, the calculation errors are slightly larger than
that for Mesh1. When r = 1α , the convergence order tends to 1 for graded mesh. For small α (e.g., α = 0.2), the
convergence rate and calculation accuracy are better than those for Mesh1. However, for larger α, the con-
vergence rates with graded meshes (i.e. 1) are inferior to that for Mesh1 (i.e. min{4α, 2}). When r = 2α , for any
value of α, the calculation based on graded meshes can always get the optimal convergence of second order,
which only can be obtained for α ≥ 0.5 for Mesh1. For small α, the calculation accuracy for graded meshes
is better than the result for Mesh1 because of the difference of convergence order. But for larger α ≥ 0.5, the
numerical errors of Mesh1 are obviously better than the standard graded mesh even though the convergence
rates for both of Mesh1 and the graded mesh are of second order.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
Max_err rate Max_err rate Max_err rate Max_err rate

25 5.1443e−3 1.0960e−3 8.6190e−4 5.6617e−4

26 3.2276e−3 0.6725 2.7515e−4 1.9939 2.1584e−4 1.9976 1.3982e−4 2.0177
27 1.9429e−3 0.7323 7.7468e−5 1.8286 5.4263e−5 1.9919 3.4912e−5 2.0017
28 1.1453e−3 0.7625 2.5670e−5 1.5935 1.3651e−5 1.9909 8.7595e−6 1.9948
29 6.6749e−4 0.7789 8.4847e−6 1.5971 3.4323e−6 1.9918 2.2016e−6 1.9923
210 3.8657e−4 0.7880 2.8015e−6 1.5987 8.6214e−7 1.9932 5.5356e−7 1.9918

Table 1: Take M = N, the max L∞ error and convergence rate for Example 1 on non-uniform Mesh1.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
L2_err rate L2_err rate L2_err rate L2_err rate

25 6.4474e−3 1.2858e−3 1.0085e−3 6.7686e−4

26 4.0452e−3 0.6725 3.3389e−4 1.9453 2.6189e−4 1.9452 1.7182e−4 1.9779
27 2.4350e−3 0.7323 9.7091e−5 1.7820 6.6964e−5 1.9675 4.3382e−5 1.9857
28 1.4354e−3 0.7625 3.2172e−5 1.5935 1.6982e−5 1.9794 1.0936e−5 1.9880
29 8.3658e−4 0.7789 1.0634e−5 1.5971 4.2861e−6 1.9863 2.7543e−6 1.9893
210 4.8450e−4 0.7880 3.5111e−6 1.5987 1.0786e−6 1.9905 6.9319e−7 1.9904

Table 2: Take M = N, the max L2 error and convergence rate for Example 1 on non-uniform Mesh1.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
Max_err rate Max_err rate Max_err rate Max_err rate

25 1.0788e−2 5.7334e−3 1.4383e−3 8.4002e−4

26 9.5385e−3 0.1776 3.5079e−3 0.7088 6.4165e−4 1.1645 2.0836e−4 2.0114
27 8.0281e−3 0.2487 2.0836e−3 0.7515 2.8192e−4 1.1865 5.1739e−5 2.0097
28 6.5647e−3 0.2903 1.2193e−3 0.7731 1.2318e−4 1.1946 1.2859e−5 2.0084
29 5.2636e−3 0.3187 7.0772e−4 0.7848 5.3699e−5 1.1978 3.4030e−6 1.9179
210 4.1611e−3 0.3391 4.0893e−4 0.7913 2.3389e−5 1.1990 1.1235e−6 1.5988

Table 3: Take M = N and r = 2, the max L∞ error and convergence rate for Example 1 on graded mesh.
 H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes | 907

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
Max_err rate Max_err rate Max_err rate Max_err rate
25 2.1957e−3 3.0646e−3 2.7853e−3 1.6429e−3
26 1.1440e−3 0.9406 1.5932e−3 0.9437 1.4480e−3 0.9438 8.5693e−4 0.9390
27 5.8256e−4 0.9736 8.1088e−4 0.9744 7.3676e−4 0.9748 4.3626e−4 0.9740
28 2.9379e−4 0.9876 4.0889e−4 0.9878 3.7144e−4 0.9881 2.1993e−4 0.9881
29 1.4751e−4 0.9940 2.0529e−4 0.9940 1.8647e−4 0.9942 1.1040e−4 0.9943
210 7.3905e−5 0.9970 1.0286e−4 0.9971 9.3420e−5 0.9971 5.5305e−5 0.9972

1
Table 4: Take M = N and r = α, the max L∞ error and convergence rate for Example 1 on graded mesh.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
Max_err rate Max_err rate Max_err rate Max_err rate
25 1.1425e−3 1.0462e−3 9.2440e−4 8.0003e−4
26 2.8558e−4 2.0002 2.6300e−4 1.9919 2.3258e−4 1.9908 2.0019e−4 1.9987
27 7.1712e−5 1.9936 6.6170e−5 1.9908 5.8550e−5 1.9900 5.0185e−5 1.9960
28 1.8015e−5 1.9930 1.6631e−5 1.9923 1.4729e−5 1.9910 1.2592e−5 1.9947
29 4.5211e−6 1.9944 4.1748e−6 1.9941 3.7011e−6 1.9926 3.1606e−6 1.9942
210 1.1334e−6 1.9961 1.0467e−6 1.9958 9.2906e−7 1.9941 7.9335e−7 1.9942

2
Table 5: Take M = N and r = α, the max L∞ error and convergence rate for Example 1 on graded mesh.

Example 2. Consider
{ C α ∂2 u
{
{ 0 D t u(x, t) − 2 (x, t) = f(x, t), (x, t) ∈ (0, π) × (0, 1],
{
{ ∂x
{
{ u(0, t) = u(π, t) = 0, t ∈ (0, 1],
{
{
{
{ u(x, 0) = u0 (x), x ∈ [0, π].
The function f(x, t) is chosen such that the exact solution of the problem is u(x, t) = (1 + t α + t2α + t1+α )sin(x).
It is clear that the u(x, t) satisfies the hypothetical condition in the theorem, that is, at the initial time the
solution is singularity.

We solve Example 2 on non-uniform Mesh1 by the format (2.3). In order to verify the time convergence order,
we select a larger spatial partition numbers M = 10000, so that the error caused by spatial discretization is
negligible. Taking different α, Tables 6 and 7 provide the maximum node error and its convergence order,
and the maximum L2 error and their convergence order, respectively. Numerical simulation results support
theoretical analysis.

Example 3. Consider
{ C α ∂2 u
{
{ 0 D t u(x, t) − 2 (x, t) = f(x, t), (x, t) ∈ (0, π) × (0, 1],
{
{ ∂x
{
{ u(0, t) = u(π, t) = 0, t ∈ (0, 1],
{
{
{
{ u(x, 0) = 0, x ∈ [0, π],
x
The function f(x, t) is chosen such that the exact solution of the problem is u(x, t) = E α,1 (−1.25t α )e 2 sin(x).
Here E α,β (z) represents the Mittag-Leffler function, defined as follows:
∞
zk
E α,β (z) = ∑ , α, β > 0.
k=0
Γ(αk + β)
Obviously, u(x, t) satisfies the assumption in the theorem.

For Example 3, we use the format (2.3) on three different non-uniform meshes. In Tables 8, 9 and 10, we
set M = N and select different α to give the maximum L2 errors and convergence orders on meshes Mesh1,
Mesh2 and Mesh3, respectively. We find that the time convergence orders are O(N − min{4α,2} ) for Mesh1,
O(N − min{3α,2} ) for Mesh2 and O(N − min{5α,2} ) for Mesh3, which are consistent with theoretical analysis.
908 | H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes

α = 0.2 α = 0.4 α = 0.6 α = 0.8

N
Max_err rate Max_err rate Max_err rate Max_err rate
25 5.0090e−3 6.1373e−4 5.9132e−4 1.4094e−3
26 3.0539e−3 0.7139 2.0664e−4 1.5705 1.5229e−4 1.9571 3.6283e−4 1.9577
27 1.8136e−3 0.7518 6.8778e−5 1.5871 3.8598e−5 1.9802 9.2176e−5 1.9768
28 1.0614e−3 0.7729 2.2780e−5 1.5942 9.7020e−6 1.9922 2.3258e−5 1.9866
29 6.1614e−4 0.7846 7.5285e−6 1.5973 2.4216e−6 2.0023 5.8421e−6 1.9932
210 3.5602e−4 0.7913 2.4857e−6 1.5987 5.9528e−7 2.0243 1.4575e−6 2.0030

Table 6: Take M = 10000, the max L∞ error and convergence rate for Example 2 on non-uniform Mesh1.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

N
L2_err rate L2_err rate L2_err rate L2_err rate
25 6.2778e−3 7.6919e−4 6.0605e−4 1.4008e−3
26 3.8275e−3 0.7139 2.5899e−4 1.5705 1.7240e−4 1.8137 4.0506e−4 1.7901
27 2.2730e−3 0.7518 8.6201e−5 1.5871 4.5949e−5 1.9076 1.0904e−4 1.8933
28 1.3303e−3 0.7729 2.8550e−5 1.5942 1.1848e−5 1.9554 2.8320e−5 1.9449
29 7.7222e−4 0.7846 9.4356e−6 1.5973 2.9956e−6 1.9837 7.2166e−6 1.9724
210 4.4621e−4 0.7913 3.1153e−6 1.5987 7.4113e−7 2.0151 1.8134e−6 1.9926

Table 7: Take M = 10000, the max L2 error and convergence rate for Example 2 on non-uniform Mesh1.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
L2_err rate L2_err rate L2_err rate L2_err rate
25 2.7181e−2 8.7493e−3 1.2985e−3 3.1866e−4
26 2.0984e−2 0.3733 3.9868e−3 1.1339 3.8032e−4 1.7716 8.5216e−5 1.9028
27 1.5399e−2 0.4465 1.7722e−3 1.1697 1.1017e−4 1.7875 2.3647e−5 1.8495
28 1.0916e−2 0.4964 7.7883e−4 1.1862 3.1766e−5 1.7942 6.4323e−6 1.8783
29 7.5579e−3 0.5304 3.4048e−4 1.1937 9.1396e−6 1.7973 1.7234e−6 1.9001
210 5.1502e−3 0.5534 1.4850e−4 1.1972 2.6271e−6 1.7987 4.5632e−7 1.9171

Table 8: Take M = N, the max L2 error and convergence rate for Example 3 on non-uniform Mesh2.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
L2_err rate L2_err rate L2_err rate L2_err rate
25 1.7789e−2 2.5571e−3 4.8096e−4 5.2366e−4
26 1.1419e−2 0.6396 8.6453e−4 1.5645 1.1031e−4 2.1243 1.5009e−4 1.8028
27 7.0041e−3 0.7051 2.8815e−4 1.5851 2.9560e−5 1.8999 4.1461e−5 1.8560
28 4.1808e−3 0.7444 9.5484e−5 1.5935 8.0359e−6 1.8791 1.1195e−5 1.8889
29 2.4556e−3 0.7677 3.1562e−5 1.5971 2.1346e−6 1.9125 2.9766e−6 1.9112
210 1.4287e−3 0.7813 1.0421e−5 1.5987 5.5814e−7 1.9352 7.8258e−7 1.9273

Table 9: Take M = N, the max L2 error and convergence rate for Example 3 on non-uniform Mesh1.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

M, N
L2_err rate L2_err rate L2_err rate L2_err rate
25 1.3776e−1 6.2594e−3 6.7512e−4 8.7905e−4
26 7.5708e−2 0.8636 1.6495e−3 1.9240 1.7568e−4 1.9422 2.3844e−4 1.8823
27 4.0043e−2 0.9189 4.2664e−4 1.9510 4.6576e−5 1.9153 6.4152e−5 1.8941
28 2.0713e−2 0.9511 1.0978e−4 1.9584 1.2323e−5 1.9183 1.7095e−5 1.9079
29 1.0565e−2 0.9713 2.8120e−5 1.9649 3.2314e−6 1.9311 4.5140e−6 1.9211
210 5.3461e−3 0.9827 1.1564e−5 1.2820 8.3916e−7 1.9451 1.1823e−6 1.9329

Table 10: Take M = N, the max L2 error and convergence rate for Example 3 on non-uniform Mesh3.
 H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes | 909

In addition, although Mesh3 has higher convergence rate for smaller α, most of the numerical results
based on Mesh1 are more accurate than Mesh3. Maybe this is because most of the nodes are concentrated at
t = 0 in Mesh3, and the mesh is coarser away from t = 0, resulting in larger truncation errors.

Example 4. Consider

{ C α ∂2 u
{
{ 0 D t u(x, t) − 2 (x, t) = f(x, t), (x, t) ∈ (0, π) × (0, 1],
{
{ ∂x
{
{ u(0, t) = u(π, t) = 0, t ∈ (0, 1],
{
{
{
{ u(x, 0) = 0, x ∈ [0, π],

The function f(x, t) is chosen such that the exact solution of the problem is u(x, t) = t α sin(x). Obviously,
u(x, t) satisfies the assumption in the theorem.

For Example 4, we use the format (2.3) on three different non-uniform meshes. In Tables 11, 12 and 13, we
set M = 10000 and select different α and provide the maximum L2 errors and convergence orders for Mesh1,
Mesh2 and Mesh3, respectively. We find that the time convergence orders are O(N − min{4α,2} ) for Mesh1,
O(N − min{3α,2} ) for Mesh2, and O(N − min{5α,2} ) for Mesh3, which are consistent with theoretical analysis.
Considering the calculation accuracy, the performance of Mesh1 is slightly better than Mesh2 and Mesh3.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

N
L2_err rate L2_err rate L2_err rate L2_err rate

25 5.9961e−3 7.6815e−4 1.5547e−4 7.3769e−5

26 3.7252e−3 0.6867 2.5887e−4 1.5692 4.6788e−5 1.7324 2.3896e−5 1.6263
27 2.2374e−3 0.7355 8.6188e−5 1.5867 1.3146e−5 1.8315 6.9942e−6 1.7725
28 1.3182e−3 0.7633 2.8549e−5 1.5940 3.5551e−6 1.8867 1.9491e−6 1.8434
29 7.6815e−4 0.7791 9.4355e−6 1.5973 9.4134e−7 1.9171 5.3015e−7 1.8783
210 4.4485e−4 0.7881 3.1153e−6 1.5987 2.4796e−7 1.9245 1.4384e−7 1.8820

Table 11: Take M = 10000, the max L2 error and convergence rate for Example 4 on non-uniform Mesh1.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

N
L2_err rate L2_err rate L2_err rate L2_err rate
25 9.7052e−3 2.6587e−3 3.9129e−4 1.0009e−4
26 7.2008e−3 0.4306 1.20093-3 1.1466 1.1453e−4 1.7726 2.4457e−5 2.0331
27 5.1252e−3 0.4905 5.3159e−4 1.1757 3.3169e−5 1.7878 5.7188e−6 2.0964
28 3.5525e−3 0.5288 2.3317e−4 1.1889 9.5630e−6 1.7943 1.2995e−6 2.1378
29 2.4206e−3 0.5535 1.0185e−4 1.1950 2.7514e−6 1.7973 2.9282e−7 2.1499
210 1.6311e−3 0.5695 4.4403e−5 1.1977 7.9086e−7 1.7987 8.2207e−8 1.8327

Table 12: Take M = 10000, the max L2 error and convergence rate for Example 4 on non-uniform Mesh2.

α = 0.2 α = 0.4 α = 0.6 α = 0.8

N
L2_err rate L2_err rate L2_err rate L2_err rate

25 4.5634e−2 1.8847e−3 3.0322e−4 1.4985e−4

26 2.4350e−2 0.9062 4.9494e−4 1.9290 7.9255e−5 1.9358 4.1460e−5 1.8537
27 1.2661e−2 0.9435 1.2795e−4 1.9517 2.0893e−5 1.9235 1.1329e−5 1.8717
28 6.4852e−3 0.9652 3.2917e−5 1.9586 5.4860e−6 1.9292 3.0547e−6 1.8909
29 3.2907e−3 0.9788 8.4291e−6 1.9654 1.4309e−6 1.9388 8.1562e−7 1.9050
210 1.6603e−3 0.9869 2.1479e−6 1.9725 3.7276e−7 1.9406 2.1770e−7 1.9055

Table 13: Take M = 10000, the max L2 error and convergence rate for Example 4 on non-uniform Mesh3.
910 | H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes

5 Conclusion
In this paper, we consider the time fractional diffusion equation with Caputo fractional derivative. Consid-
ering that the solution of the class equation has the singularity at the initial time, it is difficult to obtain
the ideal convergence order on the uniform grid. Therefore, according to the k power formula, we consider
k = 3, 4, 5 and propose three kinds of non-uniform meshes. In these non-uniform meshes, the time fractional
derivative is discretized by the L2 − 1σ format, and the diffusion term is discretized by the center difference
quotient formula on uniform meshes. By the theoretical analysis, we can promote time convergence orders
to O(N − min{kα,2} ) for different non-uniform meshes, where N represents the number of time splits, and the sta-
bility of the scheme is analyzed. Finally, the theoretical analysis is verified by several numerical examples.
Among the three meshes Mesh1, Mesh2 and Mesh3, it seems Mesh1 has better performance due to its smaller
numerical errors. For α ≥ 0.5, Mesh1 possesses the convergence rate of second order, which is optimal for the
graded mesh with tunable parameter r. For the situation α ≥ 0.5, the numerical errors for Mesh1 are obviously
better than that for standard graded mesh.
The new meshes developed in this paper can be seen as variants of the graded meshes. They have the
same optimal order of convergence as the graded meshes, but for some particular situation, the numerical
errors based on the new meshes are better than that for the graded meshes.

Funding: This work was supported in part by the National Natural Science Foundation of China under Grants
91630207, 11971272.

References
[1] E. E. Adams and L. W. Gelhar, Field study of dispersion in a heterogeneous aquifer: 2. Spatial moments analysis,
Water Resources Res. 28 (1992), no. 12, 3293–3307.
[2] A. A. Alikhanov, A new difference scheme for the time fractional diffusion equation, J. Comput. Phys. 280 (2015), 424–438.
[3] H. Brunner, L. Ling and M. Yamamoto, Numerical simulations of 2D fractional subdiffusion problems, J. Comput. Phys. 229
(2010), no. 18, 6613–6622.
[4] H. Chen and M. Stynes, A high order method on graded meshes for a time-fractional diffusion problem, in: Finite Difference
Methods, Lecture Notes in Comput. Sci. 11386, Springer, Cham (2019), 15–27.
[5] H. Chen and M. Stynes, Error analysis of a second-order method on fitted meshes for a time-fractional diffusion problem,
J. Sci. Comput. 79 (2019), no. 1, 624–647.
[6] N. J. Ford, M. L. Morgado and M. Rebelo, Nonpolynomial collocation approximation of solutions to fractional differential
equations, Fract. Calc. Appl. Anal. 16 (2013), no. 4, 874–891.
[7] N. J. Ford and Y. Yan, An approach to construct higher order time discretisation schemes for time fractional partial
differential equations with nonsmooth data, Fract. Calc. Appl. Anal. 20 (2017), no. 5, 1076–1105.
[8] G.-H. Gao and Z.-Z. Sun, A compact finite difference scheme for the fractional sub-diffusion equations, J. Comput.
Phys. 230 (2011), no. 3, 586–595.
[9] G.-H. Gao, Z.-Z. Sun and H.-W. Zhang, A new fractional numerical differentiation formula to approximate the Caputo
fractional derivative and its applications, J. Comput. Phys. 259 (2014), 33–50.
[10] R. Gorenflo, F. Mainardi, D. Moretti and P. Paradisi, Time fractional diffusion: A discrete random walk approach, Nonlinear
Dynam. 29 (2002), 129–143.
[11] Y. Hatano and N. Hatano, Dispersive transport of ions in column experiments: An explanation of long-tailed profiles,
Water Resources Res. 34 (1998), no. 5, 1027–1033.
[12] Y. Lin and C. Xu, Finite difference/spectral approximations for the time-fractional diffusion equation, J. Comput. Phys. 225
(2007), no. 2, 1533–1552.
[13] C. Lv and C. Xu, Error analysis of a high order method for time-fractional diffusion equations, SIAM J. Sci. Comput. 38
(2016), no. 5, A2699–A2724.
[14] Z. Mao and J. Shen, Efficient spectral-Galerkin methods for fractional partial differential equations with variable
coefficients, J. Comput. Phys. 307 (2016), 243–261.
[15] E. W. Montroll and G. H. Weiss, Random walks on lattices. II, J. Math. Phys. 6 (1965), 167–181.
[16] R. R. Nigmatullin, The realization of the generalized transfer equation in a medium with fractal geometry, Phys. Status
Solidi B 133 (1986), no. 1, 425–430.
 H. Qiao and A. Cheng, Finite Difference Method on Non-Uniform Meshes | 911

[17] K. Sakamoto and M. Yamamoto, Initial value/boundary value problems for fractional diffusion-wave equations and
applications to some inverse problems, J. Math. Anal. Appl. 382 (2011), no. 1, 426–447.
[18] M. Stynes, Too much regularity may force too much uniqueness, Fract. Calc. Appl. Anal. 19 (2016), no. 6, 1554–1562.
[19] M. Stynes, E. O’Riordan and J. L. Gracia, Error analysis of a finite difference method on graded meshes for a time-fractional
diffusion equation, SIAM J. Numer. Anal. 55 (2017), no. 2, 1057–1079.
[20] Z.-Z. Sun, Numerical Methods for Partial Differential Equations (in Chinese), 2nd ed., Science Press, New York, 2012.
[21] Y. Xing and Y. Yan, A higher order numerical method for time fractional partial differential equations with nonsmooth data,
J. Comput. Phys. 357 (2018), 305–323.
[22] Y. Yang, Y. Yan and N. J. Ford, Some time stepping methods for fractional diffusion problems with nonsmooth data,
Comput. Methods Appl. Math. 18 (2018), no. 1, 129–146.
[23] F. Zeng, Z. Zhang and G. E. Karniadakis, A generalized spectral collocation method with tunable accuracy for variable-order
fractional differential equations, SIAM J. Sci. Comput. 37 (2015), no. 6, A2710–A2732.
[24] Y.-N. Zhang and Z.-Z. Sun, Alternating direction implicit schemes for the two-dimensional fractional sub-diffusion
equation, J. Comput. Phys. 230 (2011), no. 24, 8713–8728.
[25] Y.-N. Zhang, Z.-Z. Sun and H.-L. Liao, Finite difference methods for the time fractional diffusion equation on non-uniform
meshes, J. Comput. Phys. 265 (2014), 195–210.
[26] Z. Zhang, F. Zeng and G. E. Karniadakis, Optimal error estimates of spectral Petrov-Galerkin and collocation methods for
initial value problems of fractional differential equations, SIAM J. Numer. Anal. 53 (2015), no. 4, 2074–2096.
[27] P. Zhuang and F. Liu, Implicit difference approximation for the time fractional diffusion equation, J. Appl. Math. Comput. 22
(2006), no. 3, 87–99.

View publication stats