18th World IMACS / MODSIM Congress, Cairns, Australia 13-17 July 2009

http://mssanz.org.au/modsim09

Finite Difference/Spectral Approximations for the Fractional

Cable Equation ∗
Yumin Lin† Xianjuan Li‡ Chuanju Xu§

Abstract
In this paper, we consider the numerical solution of the fractional Cable equation,
which is a generalization of the classical Cable equation by taking into account the
anomalous diffusion in the movement of the ions in neuronal system. A schema com-
bining a finite difference approach in the time direction and a spectral method in the
space direction is proposed and analyzed. The main contribution of this work is three-
fold: 1) We construct a finite difference/Legendre spectral schema for discretization of
the fractional Cable equation. 2) We give a detailed analysis of the proposed schema
by providing some stability and error estimates. Based on this analysis, the conver-
gence of the method is rigourously established. We prove that the overall schema is
unconditionally stable, and the numerical solution converges to the exact one with
order O(4t2−max{α,β} ), where 4t is the time step size, α and β are two different
exponents between 0 and 1 involved in the fractional derivatives. 3) Finally, some
numerical experiments are carried out to support the theoretical claims.

AMS Mathematics Subject Classification: 65M12, 65M06, 65M70, 35S10.

Key words: Fractional Cable equation, numerical solution, stability, convergence.

1 Introduction
Due to its significant deviation from the dynamics of Brownian motion, the anomalous
diffusion in biological systems can not be adequately described by the traditional Nernst-
Planck equation or its simplification, the Cable equation. Very recently, a modified Cable
equation was introduced for modeling the anomalous diffusion in spiny neuronal dendrites
[2]. The resulting governing equation, the so-called fractional Cable equation, is similar
to the traditional Cable equation except that the order of derivative with respect to the
space and/or time is fractional.
∗
This research was partially supported by National NSF of China under Grant 10531080, and 973 High
Performance Scientific Computation Research Program 2005CB321703.
†
School of Mathematical Sciences, Xiamen University, 361005 Xiamen, China. The research of this
author was partially supported by Fujian NSF under Grant S0750017.
‡
School of Mathematical Sciences, Fuzhou University, 350200 Fuzhou, China.
§
School of Mathematical Sciences, Xiamen University, 361005 Xiamen, China. http://lsec.cc.ac.cn/˜
cjxu.

455
 FD/Spectral methods for fractional Cable equation 2

The goal of this paper is to address such an equation, and to design efficient numerical
schemes for its numerical solution. Due to the memory feature of the fractional equation,
it is very desirable to use high order methods for efficient computations of the numerical
solution. In this work, we aim at developing and analyzing a finite difference schema in
time and Legendre spectral methods in space for the fractional Cable equation.
Note that some similar investigations have been made for the time fractional diffusion
equation. For example, Langlands and Henry [3] considered an implicit numerical schema
for fractional diffusion equation in which the backward Euler approximation is used to
discretize the first order time derivative and the L1 schema is used to approximate the
fractional order time derivative. Lin and Xu [4] proposed a finite difference schema in
time and Legendre collocation spectral method in space for the time fractional diffusion
equation.
This work follows the idea proposed in [4] in an attempt to generalize the mixed finite
difference/Legendre spectral method in [4] to the fractional Cable equation. In particular,
an improved technique, as compared to the one used in [4], for the proof of the time
error estimate is provided. This new technique allows to obtain a detailed dependence of
the constant appeared in front of the convergence rate 4t2−max{α,β} . The outline of this
paper is as follows. In the next section we construct our finite difference method for the
Cable equation. A detailed error analysis is carried out to derive the error estimate for
the numerical solution. Some numerical results are presented in section 3 which support
the theoretical statement.

2 Discretization in time: a finite difference schema

2.1 Cable equation

We consider the initial boundary value problem of the fractional Cable equation. Let
Λ = (−1, 1) be the space domain, I = (0, T ] be the time domain. We consider the
fractional Cable equation:

(2.1) ∂t u = RDtβ ∂x2 u − µRDtα u, ∀(x, t) ∈ Λ × I,

subject to the initial condition:

(2.2) u(x, 0) = u0 (x), ∀x ∈ Λ,

and the boundary condition:

(2.3) u(−1, t) = u(1, t) = 0, ∀t ∈ I,

where 0 < α < 1, 0 < β < 1. RDtγ , with γ = α or β, denotes Riemann-Liouville fractional
derivative of γ-order with respect to variable t, defined by
Z t
R β 1 v(x, τ )
(2.4) Dt v(x, t) = ∂t β
dτ, 0 < β < 1.
Γ(1 − β) 0 (t − τ )

In order to follow the construction idea used in our previous paper [4], we will use the
Caputo fractional derivative instead of the Riemann definition. To this end, we recall the

456
 FD/Spectral methods for fractional Cable equation 3

following well-known relation, see e.g. [5]: For 0 < γ < 1, if v(t) has the integrable first
order derivative in [0, T ], then

v(0)t−γ
(2.5) R
Dtγ v(t) = Dtγ v(t) + ,
Γ(1 − γ)

where Dtγ denotes the Caputo fractional derivative of γ-order, defined by:
Z t
γ 1 ∂τ v(τ )
Dt v(t) = dτ.
Γ(1 − γ) 0 (t − τ )γ
Then the fractional Cable equation (2.1) can be transformed under the form of Caputo
definition:
1 µ
(2.6) ∂t u = Dtβ ∂x2 u − µDtα u + β
∂x2 u(x, 0) − u(x, 0).
Γ(1 − β)t Γ(1 − α)tα
In the next section, we are going to construct and analyze a finite difference schema for
the time discretization of the above equation.

2.2 Construction of the schema

First, we introduce a finite difference schema to discretize the time fractional derivative.
T
For a given integer K > 0, let tk = k4t, k = 0, 1, · · · , K, where 4t = K is the time step.
By using the Taylor formula with the intergral remainder:
Z t
f (t) = f (s) + ∂t f (s)(t − s) + ∂τ2 f (τ )(t − τ )dτ, ∀t, s ∈ I
s

to the function u(·, t) at t = tj and t = tj+1 respectively, we obtain

u(x, tj+1 ) − u(x, tj )

∂s u(x, s) =
∆t
Z tj+1 Z tj
1 2 1
− ∂τ u(x, τ )(tj+1 − τ )dτ + ∂ 2 u(x, τ )(tj − τ )dτ.
∆t s ∆t s τ
Thus for all 0 ≤ k ≤ K − 1, we have

X k Z tj+1
1 ds
(2.7) Dtα u(x, tk+1 ) = ∂s u(x, s)
Γ(1 − α) tj (tk+1 − s)α
j=0
k
X u(x, tk+1−j ) − u(x, tk−j )
1
= aj + rαk+1 ,
Γ(2 − α) 4tα
j=0

where

(2.8) aj = (j + 1)1−α − j 1−α ,

and
X k Z tj+1
1
(2.9) rαk+1 = ∂τ2 u(x, τ )Rjk+1 (τ )dτ.
Γ(2 − α)∆t tj
j=0

457
 FD/Spectral methods for fractional Cable equation 4

In the last equality of the above derivation we have used the notation Rjk+1 (τ ) to denote

Rjk+1 (τ ) := (tk+1 − τ )1−α ∆t − (tj+1 − τ )(tk+1 − tj )1−α + (tj − τ )(tk+1 − tj+1 )1−α .

It can be proved that

Rjk+1 (τ ) ≥ 0, for all τ ∈ [tj , tj+1 ].

Thus from equation (2.9), we obtain

X k Z tj+1
M
rαk+1 ≤ Rjk+1 (τ )dτ,
Γ(2 − α)∆t tj
j=0

where M = max ∂τ2 u(x, τ ). Furthermore, a careful analysis shows that

x∈Λ,τ ∈I

(2.10) rαk+1 ≤ c4t2−α ,

where c depends only on M , a constant measuring ∂t2 u.

We can derive an expression similar to (2.7) for the fractional derivative term of order
β in (2.6):
k
X
1 bj ¡ 2 ¢
(2.11) Dtβ ∂x2 u(x, tk+1 ) = β
∂x u(x, tk+1−j ) − ∂x2 u(x, tk−j ) + rβk+1 ,
Γ(2 − β) 4t
j=0

where

(2.12) bj = (j + 1)1−β − j 1−β , rβk+1 ≤ c∆t2−β .

For the discretization of the first order time derivative ∂t u, we use the following devel-
opment: for k ≥ 1,

3u(x, tk+1 ) − 4u(x, tk ) + u(x, tk−1 )

(2.13) ∂t u(x, tk+1 ) = + O(∆t2 ),
2∆t
For a mesh function {f k }K α
k=0 , we define the fractional difference operators Lt by

k
X f k+1−j − f k−j
1
(2.14) Lαt f k+1 = aj , k ≥ 0,
Γ(2 − α) 4tα
j=0

and Lβt by
k
X f k+1−j − f k−j
1
(2.15) Lβt f k+1 = bj , k ≥ 0,
Γ(2 − α) 4tβ
j=0

We also define the difference operators L1t by

3f k+1 − 4f k + f k−1
(2.16) L1t f k+1 = , k ≥ 1.
2∆t

458
 FD/Spectral methods for fractional Cable equation 5

Then by combining (2.6), (2.7), (2.11), and (2.13), we have

(2.17) L1t u(x, tk+1 ) − rk+1

= −µLαt u(x, tk+1 ) + Lβt ∂x2 u(x, tk+1 ) − µrαk+1 + rβk+1
1 µu(x, 0)
+ ∂ 2 u(x, 0) − , k ≥ 1,
Γ(1 − β)(k + 1)β ∆tβ x Γ(1 − α)(k + 1)α ∆tα

where, according to (2.13), rk+1 = O(∆t2 ).

The above expression motivates us to consider the following finite difference schema
for the time discretization of (2.13):

(2.18) L1t uk+1 = −µLαt uk+1 + Lβt ∂x2 uk+1

1 2 0 µu0
+ ∂x u − , k ≥ 1.
Γ(1 − β)(k + 1)β ∆tβ Γ(1 − α)(k + 1)α ∆tα

In (2.18), uk , a simplified notation of uk (x), is an approximation of u(x, tk ). Formally,

(2.18) is a schema with the truncation error rk+1 +rαk+1 +rβk+1 , stemming respectively from
the discretizations of the first order time derivative and the time fractional derivatives of
orders α and β.
In details, the schema (2.18) reads

3uk+1 − 4uk + uk−1

(2.19)
2∆t  
k−1
X
−µ uk+1 −
= (aj − aj+1 )uk−j − ak u0 
Γ(2 − α)∆tα
j=0
 
k−1
X
1 ∂x2 uk+1 −
+ (bj − bj+1 )∂x2 uk−j − bk ∂x2 u0 
Γ(2 − β)∆tβ
j=0

1 2 0 µu0
+ ∂x u − , k ≥ 1.
Γ(1 − β)(k + 1)β ∆tβ Γ(1 − α)(k + 1)α ∆tα

2.3 Time error analysis

In this subsection, we aim at carrying out a rigorous error analysis for the time schema
(2.19). The error analysis is based on the weak formulation of the related problems. The
inner products of L2 (Λ) and H 1 (Λ) are defined respectively by
Z
α̃ β̃
(2.20) (u, v) = uv dx, (u, v)1 = (u, v) + (u, v) + (∂x u, ∂x v),
Λ 2 2
where
4µ∆t 4∆t
(2.21) α̃ = , β̃ = .
Γ(2 − α)∆tα Γ(2 − β)∆tβ

The norms of L2 (Λ) and H 1 (Λ) are defined by

1/2
(2.22) kvk0 = (v, v)1/2 , kvk1 = (v, v)1 .

459
 FD/Spectral methods for fractional Cable equation 6

Here we have used a H 1 -norm different from the standard one. We will see that the
H 1 -norm defined here is more convenient than the standard norm for the error analysis,
although the two norms are equivalent for fixed µ, ∆t, α, and β.
We now consider the weak formulation of the equation (2.19) subject to the homoge-
neous boundary condition: find uk+1 ∈ H01 (Λ), such that for all v ∈ H01 (Λ),
µ k+1 ¶
3u − 4uk + uk−1
(2.23) ,v
2∆t
µ ³ k−1
X ´
k+1 k−j 0
= − (u , v) − (aj − aj+1 )(u , v) − ak (u , v)
Γ(2 − α)∆tα
j=0

1 ³ k−1
X ´
k+1 k−j 0
− (∂x u , ∂ x v) − (b j − b j+1 )(∂x u , ∂ x v) − b k (∂x u , ∂ x v)
Γ(2 − β)∆tβ
j=0
1 µ
− (∂x u0 , ∂x v) − (u0 , v), k ≥ 1.
Γ(1 − β)(k + 1)β ∆tβ Γ(1 − α)(k + 1)α ∆tα
For the sake of simplification, let’s introduce the notations:
4µ∆t 4∆t
(2.24) α̃k+1 = , β̃k+1 = .
Γ(1 − α)(k + 1)α ∆tα Γ(1 − β)(k + 1)β ∆tβ
By using the notations (2.21) and (2.24), the schema (2.23) becomes

(2.25) 2(3uk+1 − 4uk + uk−1 , v)

³ k−1
X ´
k+1
= −α̃ (u , v) − (aj − aj+1 )(uk−j , v) − ak (u0 , v)
j=0
³ k−1
X ´
k+1
−β̃ (∂x u , ∂x v) − (bj − bj+1 )(∂x uk−j , ∂x v) − bk (∂x u0 , ∂x v)
j=0

−β̃k+1 (∂x u , ∂x v) − α̃k+1 (u0 , v), k ≥ 1.

0

The stability and error estimate are given in the following theorems.
Theorem 2.1. The semi-discretized problem (2.25) is unconditionally stable in the sense
that for all 4t > 0, it holds

(2.26) E k+1 ≤ E k , k = 1, · · · , K − 1,
P Pk
where E k = kuk k20 + k2uk − uk−1 k20 + α̃2 kj=0 aj kuk−j k20 + β̃
2 j=0 bj k∂x u
k−j k2 ,
0 k ≥ 1.

Proof. Omitted.

Theorem 2.2. Let u be the solution of the continuous problem (2.1)-(2.3), {uk }K
k=0 be the
time-discrete solution of (2.18). Then

(2.27) ku(x, tk ) − uk k1 ≤ cT α ∆tmin(2−α,2−β) , k ≥ 1,

where c is independent of T and 4t.

Proof. Omitted.

460
 FD/Spectral methods for fractional Cable equation 7

3 Numerical validation
The semi-discretized problem (2.25) is furthermore discretized in space by using a standard
spectral method as described in [4]. We omit the details of the description due to the length
limit of the paper. We consider the Cable equation with the exact solution u(x, t) =
t2 sin(2πx) for a suitable forcing term. All the numerical results reported in the figures
below have been evaluated at T = 1.
We fix N = 16, a value large enough such that the space discretization error is negligible
as compared with the time error. In figures 1, we plot the errors in the L2 and H 1 semi
norms as a function of the time step sizes for two different sets of α, β . A logarithmic scale
has been used for both 4t−axis and error-axis in these figures. As expected, the finite
difference schema yields a fractional temporal approximation order min{2 − α, 2 − β},
that is, the slopes of the error curves in these log − log plots are respectively 1.4 for
α = 0.1, β = 0.6 and 1.1 for α = 0.9, β = 0.1.
alpha=0.1,beta=0.6 alpha=0.9,beta=0.1
0.01 1e-04

0.001

1e-05

1e-04
errors

errors

1e-05 1e-06

1e-06

1e-07

1e-07
1
L2 norm L2 norm
H seminorm H1 seminorm
Line of slope 1.4 Line of slope 1.1

1e-08 1e-08
1e-04 0.001 0.01 0.1 1e-05 1e-04 0.001 0.01
time step time step

Figure 1: Errors as functions of 4t for α = 0.1, β = 0.6 (left) and α = 0.9, β = 0.1 (right).

References
[1] C. Bernardi and Y. Maday. Approximations spectrales de problemès aux limites ellip-
tiques. Springer-Verlag, 1992.

[2] B.I. Henry, T.A.M. Langlands, and S.L. Wearne. Fractional cable models for spiny
neuronal dendrites. Phys. Rev. Lett., 100(12):128103, 2008.

[3] T. A. M. Langlands and B. I. Henry. The accuracy and stability of an implicit solution
method for the fractional diffusion equation. J. Comput. Phys., 205(2):719–736, 2005.

[4] Y. M. Lin and C. J. Xu. Finite difference/spectral approximation for the time fractional
diffusion equations. J. Comput. Phys., 2(3):1533–1552, 2007.

[5] I. Podlubny. Fractional Differential Equations. Academic Press, 1999.

461