Fast Evaluation of the Caputo Fractional Derivative ...jiang/Papers/2017CiCP.pdf... approximation...
-
Upload
nguyendiep -
Category
Documents
-
view
215 -
download
1
Transcript of Fast Evaluation of the Caputo Fractional Derivative ...jiang/Papers/2017CiCP.pdf... approximation...
Commun Comput Physdoi 104208cicpOA-2016-0136
Vol 21 No 3 pp 650-678March 2017
Fast Evaluation of the Caputo Fractional Derivative
and its Applications to Fractional Diffusion Equations
Shidong Jiang1 Jiwei Zhang2lowast Qian Zhang2 and Zhimin Zhang23
1 Department of Mathematical Sciences New Jersey Institute of Technology NewarkNJ 07102 USA2 Beijing Computational Science Research Center Beijing 100093 China3 Department of Mathematics Wayne State University Detroit MI 48202 USA
Communicated by Tao Zhou
Received 19 August 2016 Accepted (in revised version) 23 November 2016
Abstract The computational work and storage of numerically solving the time frac-tional PDEs are generally huge for the traditional direct methods since they requiretotal O(NSNT) memory and O(NSN2
T) work where NT and NS represent the totalnumber of time steps and grid points in space respectively To overcome this difficultywe present an efficient algorithm for the evaluation of the Caputo fractional derivativeC0 Dα
t f (t) of order αisin (01) The algorithm is based on an efficient sum-of-exponentials
(SOE) approximation for the kernel tminus1minusα on the interval [∆tT] with a uniform abso-lute error ε We give the theoretical analysis to show that the number of exponentials
Nexp needed is of order O(logNT) for T≫1 or O(log2 NT) for Tasymp1 for fixed accuracyε The resulting algorithm requires only O(NSNexp) storage and O(NSNT Nexp) workwhen numerically solving the time fractional PDEs Furthermore we also give the sta-bility and error analysis of the new scheme and present several numerical examplesto demonstrate the performance of our scheme
AMS subject classifications 33C10 33F05 35Q40 35Q55 34A08 35R11 26A33
Key words Fractional derivative Caputo derivative sum-of-exponentials approximation frac-tional diffusion equation fast convolution algorithm stability analysis
1 Introduction
Over the last few decades the fractional calculus has received much attention of bothphysical scientists and mathematicians since they can faithfully capture the dynamics of
lowastCorresponding author Email addresses shidongjiangnjitedu (S Jiang) jwzhangcsrcaccn (JZhang) qianzhangcsrcaccn (Q Zhang) zmzhangcsrcaccn (Z Zhang)
httpwwwglobal-scicom 650 ccopy2017 Global-Science Press
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 651
physical process in many applied sciences including biology ecology and control systemThe anomalous diffusion also referred to as the non-Gaussian process has been observedand validated in many phenomena with accurate physical measurement [192829464851] The mathematical and numerical analysis of the factional calculus became a subjectof intensive investigations
In this paper we consider a fast evaluation of the following fractional partial differ-ential equation
C0 Dα
t u(xt)=∆u(xt)+F(uxt) 0ltαlt1 (11)
where the Caputo fractional derivative C0 Dα
t u(xt) is defined by the formula
C0 Dα
t u(xt)=1
Γ(mminusα)
int t
0
u(m)(xτ)
(tminusτ)α+1minusmdτ mminus1ltαltm misinZ (12)
The existing schemes for solving (11) require the storage of the solution at all previoustime steps and the computational complexity of these schemes is O(N2
T NS) with NT thetotal number of time steps and NS the number of grid points in space This is in darkcontrast with the usual diffusion equations where one only needs to store the solution ata fixed number of time steps and the computational complexity is linear with respect toNT
It is easy to see that the difficulty is caused by the Caputo factional derivative ap-peared in (11) Indeed one of the popular schemes of discretizing the Caputo fractionalderivative is the so-called L1 approximation [15 16 23 30 32 36 39ndash41 52] which is sim-ply based on the piecewise linear interpolation of u on each subinterval For 0ltαlt1 theorder of accuracy of the L1 approximation is 2minusα There are also high-order discretiza-tion schemes by using piecewise high-order polynomial interpolation of u [10 17 33 47]and [9 34 37] For each spatial point x these methods require the storage of all previousfunction values u(0)u(t1)middotmiddotmiddot u(tn) and O(n) flops at the nth step Thus it requires onaverage O(NT) storage and the total computational cost is O(N2
T) which forms a bot-tleneck for long time simulations especially when one tries to solve the time fractionalpartial differential equations (PDEs)
Here we present an efficient scheme for solving the fractional PDEs (11) Our key ob-servation is that the Caputo derivative can be evaluated almost as efficient as the usuallyderivatives (besides some logarithmic factors) We first split the convolution integral in(12) into two parts - a local part containing the integral from tminus∆t to t and a history partcontaining the integral from 0 to tminus∆t The local part is approximated using the stan-dard L1 approximation For the history part integration by parts leads to a convolutionintegral of u with the kernel tminus1minusα We show that tminus1minusα (0 lt α lt 1) admits an efficientsum-of-exponentials (SOE) approximation on the interval [δT] with δ = ∆t a uniformabsolute error ε and the number of exponentials needed is of the order
Nexp =O(
log1
ε
(
loglog1
ε+log
T
δ
)
+log1
δ
(
loglog1
ε+log
1
δ
))
(13)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
652 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
That is for fixed precision ε we have Nexp =O(logNT) for T ≫ 1 or Nexp =O(log2 NT)
for T asymp 1 assuming that NT = T∆t The approximation can be used to accelerate the
evaluation of the convolution via the standard recurrence relation The resulting al-gorithm has nearly optimal complexity ndash O(NT Nexp) work and O(Nexp) storage Wewould like to remark here that the SOE approximations have been applied to speedup the evaluation of the convolution integrals in many applications (see for example[1 6 20 21 24 26 31 38 45 50 53 54 58 59] and references therein)
When we incorporate the fast evaluation scheme of the Caputo fractional derivativeto solve the fractional diffusion equations (both linear and nonlinear) the resulting algo-rithm for solving the fractional PDEs is both efficient and stable The computational costof the new algorithm is O(NS NT Nexp) as compared with O(NSN2
T) for direct methodsand the storage requirement is only O(NSNexp) as compared with O(NS NT) for directmethods since one needs to store the solution in the whole computational spatial domainat all times Furthermore we have carried out a rigorous and detailed analysis to provethat our scheme is unconditionally stable with respect to arbitrary step sizes With thesetwo properties our scheme provides an efficient and reliable tool for long time large scalesimulation of fractional PDEs In literatures there are many other efforts to speed up theevaluation of weakly singular kernel Lubich and Schadle [44] present a fast convolutionfor non-reflecting boundary conditions Baffet and Hesthaven [3] compressed the kernelin the Laplace domain and obtained a sum-of-poles approximation for the Laplace trans-form of the kernel And other fast approximation to time-fractional derivative is based onthe block triangular Toeplitz matrix or block triangular Toeplitz-like matrix see [27 43]
The paper is organized as follows In Section 2 we describe the fast algorithm for theevaluation of the Caputo fractional derivative and provide rigorous error analysis of ourdiscretization scheme In Section 3 we apply our fast algorithm to solve the linear frac-tional diffusion PDEs and present the stability and error analysis for the overall schemeIn Section 4 we study the nonlinear fractional diffusion PDEs and demonstrate that ourfast algorithm has the same order of convergence as the direct method in this case Fi-nally we conclude our paper with a brief discussion on the extension and generalizationof our scheme
2 Fast evaluation of the Caputo fractional derivative
In this section we consider the fast evaluation of the Caputo fractional derivative for0ltαlt1 Suppose that we would like to evaluate the Caputo fractional derivative on theinterval [0T] over a set of grid points Ωt =tn n=01middotmiddotmiddot NT with t0 =0 tNT
=T and∆tn = tnminustnminus1
We first split the convolution integral in (12) into a sum of local part and history partthat is
C0 Dα
t u(t)|t=tn =1
Γ(1minusα)
int tn
0
uprime(s)ds
(tnminuss)α
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 653
=1
Γ(1minusα)
int tn
tnminus1
uprime(s)ds
(tnminuss)α+
1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=Cl(tn)+Ch(tn) (21)
where the last equality defines the local part and the history part respectively For thelocal part we apply the standard L1 approximation which approximates u(s) on [tnminus1tn]by a linear polynomial (with u(tnminus1) and u(tn) as the interpolation nodes) or uprime(s) by a
constant u(tn)minusu(tnminus1)∆tn
We have
Cl(tn)asympu(tn)minusu(tnminus1)
∆tnΓ(1minusα)
int tn
tnminus1
1
(tnminuss)αds=
u(tn)minusu(tnminus1)
∆tαnΓ(2minusα)
(22)
For the history part we apply the integration by parts to eliminate uprime(s) and have
Ch(tn)=1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαint tnminus1
0
u(s)ds
(tnminuss)1+α
]
(23)
21 Efficient sum-of-exponentials approximation for the power function
We now show that the kernel tminusβ (0ltβlt2) can be approximated via a sum-of-exponentialsapproximation efficiently on the interval [δT] with δ=min1lenleN ∆tn and the absolute er-ror ε That is there exist positive real numbers si and wi (i = 1middotmiddotmiddot Nexp) such that for0ltβlt2
∣
∣
∣
∣
∣
1
tβminus
Nexp
sumi=1
ωieminussit
∣
∣
∣
∣
∣
le ε tisin [δT] (24)
where Nexp is given by (13) The evaluation of the history part via (23) only needsthe case 1 lt β lt 2 However one may also evaluate the history part directly withoutintegration by parts and the case 0 lt β lt 1 is then needed Meanwhile one can ob-tain the code for si and ωi in (24) from the homepage httpswebnjitedu~jiang
or httpwwwcsrcaccn~jwzhangWe start our proof from the following integral representation of the power function
Lemma 21 For any βgt01
tβ=
1
Γ(β)
int infin
0eminustmiddotssβminus1ds (25)
Proof This follows from a change of variable x= tmiddots and the integral definition of the Γ
function [49]
Eq (25) can be viewed as a representation of tminusβ using an infinitely many (continu-ous) exponentials In order to obtain an efficient SOE approximation we first truncate the
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
654 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
integral to a finite interval then subdivide the finite interval into a set of dyadic intervalsand discretize the integral on each dyadic interval with proper quadratures
We now assume 0ltβlt2
Lemma 22 For tgeδgt0
∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
le
eminusδp pβminus1
Γ(β)1δ 0ltβle1
eminusδp2βminus1(
pβ
Γ(β)+ 1
δβ
)
1ltβlt2(26)
Proof For the case of 0ltβle1∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp pβminus1
Γ(β)
int infin
0eminusδxdx
le eminusδp pβminus1
Γ(β)
1
δ (27)
For the case of 1ltβlt2∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp 1
Γ(β)
(
int p
0(2p)βminus1dx+
int infin
peminusδx(2x)βminus1dx
)
le eminusδp 2βminus1
Γ(β)
(
pβ+int infin
0eminusδxxβminus1dx
)
le eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
(28)
This completes the proof
Remark 21 The truncation error can be made arbitrarily small for fixed δ by choosingsufficiently large p Usually we have δlt1 and if one would like to bound the truncationerror by εlt1e then δpgt1 or pgt1δ When 0ltβle1
eminusδp pβminus1
Γ(β)
1
δlt eminusδp pβ
Γ(β)pδlt eminusδp p (29)
when 1ltβlt2
eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
lt eminusδp2βminus1
(
pβ
Γ(β)+pβ
)
lt5eminusδp p2 (210)
Thus p=O(
log(
1εδ
)
δ)
is sufficient to bound the truncation error by ε
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 655
We now proceed to discuss the discretization error for the integral on the interval[0p] On each dyadic interval we will discretize the integral via standard Gauss-Legendrequadrature and analyze the discretization error The error can be controlled more or lessuniformly since each dyadic interval is separated from the singular point ie the originby its own length
Lemma 23 Consider a dyadic interval [ab]= [2j 2j+1] and let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn be thenodes and weights for n-point Gauss-Legendre quadrature on the interval Then for βisin(02) andngt1
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
lt252 πaβ
(
e1e
4
)2n
(211)
Proof For any interval [ab] the standard estimate for n-point Gauss-Legendre quadra-ture [49] yields
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le (bminusa)2n+1
2n+1
(n)4
[(2n)]3max
sisin(ab)
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣ (212)
where Ds denotes the derivative with respect to s Applying Stirlingrsquos approximation [49]radic
2πnn+12eminusnltnlt2
radicπnn+12eminusn (213)
we obtain(n)4
[(2n)]3lt2
radicπ( e
8
)2nradic
n
n2n (214)
Observe now |(βminus1)middotmiddotmiddot(βminusk)|le k for kgt1 and thus∣
∣
∣Dk
s sβminus1∣
∣
∣= sβminus1le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=0∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)sβminuskminus1
∣
∣
∣le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=1∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)middotmiddotmiddot(βminusk)sβminuskminus1
∣
∣
∣le ksβminuskminus1
le2radic
πkk+12eminusksβminuskminus1le2radic
πradic
2n(2n)keminusksβminuskminus1 for kgt1
(215)
We also haveD2nminusk
s eminusst=(minust)2nminuskeminusst (216)
Combining (215) (216) with the Leibniz rule we obtain
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣=
∣
∣
∣
∣
∣
2n
sumk=0
(
2nk
)
(
D2nminusks eminusst
)(
Dks sβminus1
)
∣
∣
∣
∣
∣
le2radic
πradic
2nsβminus1eminusst2n
sumk=0
(
2nk
)
t2nminusk(2n)keminusksminusk
=2radic
πradic
2nsβminus1eminusst
(
t+2n
es
)2n
(217)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
656 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Combining (212) (214) (217) and bminusa= a we have
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le maxalesleb
2radic
2π(bminusa)sβminus1eminusst
(
e(bminusa)t
8n+
bminusa
4s
)2n
le252 πaβeminusat
(
eat
8n+
1
4
)2n
(218)
And (211) follows from the fact
maxxgt0
eminusx
(
ex
8n+
1
4
)2n
=
(
e1e
4
)2n
nge2 (219)
This completes the proof
We now consider the end interval [0a] We will use Gauss-Jacobi quadrature with theweight function sβminus1 to discretize the integral Under this weight the function to be in-tegrated is a simple exponential function Thus the discretization error can be controlledto be arbitrarily small with properly chosen parameters
Lemma 24 Let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn (nge 2) be the nodes and weights for n-point Gauss-Jacobi quadrature with the weight function sβminus1 on the interval Then for 0lt tltT βisin(02) andngt1
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
lt4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(220)
Proof The standard estimate for n-point Gauss-Jacobi quadrature [49] yields
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le a2n+β
2n+β
(n)2[Γ(n+β)]2
(2n)[Γ(2n+β)]2max
sisin(0a)
∣
∣D2ns eminusst
∣
∣ (221)
For ngt 1 we have Γ(n+β)lt Γ(n+lceilβrceil)= (n+lceilβrceilminus1) Γ(2n+β)gt Γ(2n+lfloorβrfloor)= (2n+lfloorβrfloorminus1) 2n+βgt2n+lfloorβrfloor Thus
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le max0leslea
a2n+β
2n+lfloorβrfloor(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2t2neminusst
le aβ
2n
(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2(aT)2n
le4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(222)
where the last inequality follows from Stirlingrsquos approximation (213) and the fact thatlceilβrceilminuslfloorβrfloorle1
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 651
physical process in many applied sciences including biology ecology and control systemThe anomalous diffusion also referred to as the non-Gaussian process has been observedand validated in many phenomena with accurate physical measurement [192829464851] The mathematical and numerical analysis of the factional calculus became a subjectof intensive investigations
In this paper we consider a fast evaluation of the following fractional partial differ-ential equation
C0 Dα
t u(xt)=∆u(xt)+F(uxt) 0ltαlt1 (11)
where the Caputo fractional derivative C0 Dα
t u(xt) is defined by the formula
C0 Dα
t u(xt)=1
Γ(mminusα)
int t
0
u(m)(xτ)
(tminusτ)α+1minusmdτ mminus1ltαltm misinZ (12)
The existing schemes for solving (11) require the storage of the solution at all previoustime steps and the computational complexity of these schemes is O(N2
T NS) with NT thetotal number of time steps and NS the number of grid points in space This is in darkcontrast with the usual diffusion equations where one only needs to store the solution ata fixed number of time steps and the computational complexity is linear with respect toNT
It is easy to see that the difficulty is caused by the Caputo factional derivative ap-peared in (11) Indeed one of the popular schemes of discretizing the Caputo fractionalderivative is the so-called L1 approximation [15 16 23 30 32 36 39ndash41 52] which is sim-ply based on the piecewise linear interpolation of u on each subinterval For 0ltαlt1 theorder of accuracy of the L1 approximation is 2minusα There are also high-order discretiza-tion schemes by using piecewise high-order polynomial interpolation of u [10 17 33 47]and [9 34 37] For each spatial point x these methods require the storage of all previousfunction values u(0)u(t1)middotmiddotmiddot u(tn) and O(n) flops at the nth step Thus it requires onaverage O(NT) storage and the total computational cost is O(N2
T) which forms a bot-tleneck for long time simulations especially when one tries to solve the time fractionalpartial differential equations (PDEs)
Here we present an efficient scheme for solving the fractional PDEs (11) Our key ob-servation is that the Caputo derivative can be evaluated almost as efficient as the usuallyderivatives (besides some logarithmic factors) We first split the convolution integral in(12) into two parts - a local part containing the integral from tminus∆t to t and a history partcontaining the integral from 0 to tminus∆t The local part is approximated using the stan-dard L1 approximation For the history part integration by parts leads to a convolutionintegral of u with the kernel tminus1minusα We show that tminus1minusα (0 lt α lt 1) admits an efficientsum-of-exponentials (SOE) approximation on the interval [δT] with δ = ∆t a uniformabsolute error ε and the number of exponentials needed is of the order
Nexp =O(
log1
ε
(
loglog1
ε+log
T
δ
)
+log1
δ
(
loglog1
ε+log
1
δ
))
(13)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
652 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
That is for fixed precision ε we have Nexp =O(logNT) for T ≫ 1 or Nexp =O(log2 NT)
for T asymp 1 assuming that NT = T∆t The approximation can be used to accelerate the
evaluation of the convolution via the standard recurrence relation The resulting al-gorithm has nearly optimal complexity ndash O(NT Nexp) work and O(Nexp) storage Wewould like to remark here that the SOE approximations have been applied to speedup the evaluation of the convolution integrals in many applications (see for example[1 6 20 21 24 26 31 38 45 50 53 54 58 59] and references therein)
When we incorporate the fast evaluation scheme of the Caputo fractional derivativeto solve the fractional diffusion equations (both linear and nonlinear) the resulting algo-rithm for solving the fractional PDEs is both efficient and stable The computational costof the new algorithm is O(NS NT Nexp) as compared with O(NSN2
T) for direct methodsand the storage requirement is only O(NSNexp) as compared with O(NS NT) for directmethods since one needs to store the solution in the whole computational spatial domainat all times Furthermore we have carried out a rigorous and detailed analysis to provethat our scheme is unconditionally stable with respect to arbitrary step sizes With thesetwo properties our scheme provides an efficient and reliable tool for long time large scalesimulation of fractional PDEs In literatures there are many other efforts to speed up theevaluation of weakly singular kernel Lubich and Schadle [44] present a fast convolutionfor non-reflecting boundary conditions Baffet and Hesthaven [3] compressed the kernelin the Laplace domain and obtained a sum-of-poles approximation for the Laplace trans-form of the kernel And other fast approximation to time-fractional derivative is based onthe block triangular Toeplitz matrix or block triangular Toeplitz-like matrix see [27 43]
The paper is organized as follows In Section 2 we describe the fast algorithm for theevaluation of the Caputo fractional derivative and provide rigorous error analysis of ourdiscretization scheme In Section 3 we apply our fast algorithm to solve the linear frac-tional diffusion PDEs and present the stability and error analysis for the overall schemeIn Section 4 we study the nonlinear fractional diffusion PDEs and demonstrate that ourfast algorithm has the same order of convergence as the direct method in this case Fi-nally we conclude our paper with a brief discussion on the extension and generalizationof our scheme
2 Fast evaluation of the Caputo fractional derivative
In this section we consider the fast evaluation of the Caputo fractional derivative for0ltαlt1 Suppose that we would like to evaluate the Caputo fractional derivative on theinterval [0T] over a set of grid points Ωt =tn n=01middotmiddotmiddot NT with t0 =0 tNT
=T and∆tn = tnminustnminus1
We first split the convolution integral in (12) into a sum of local part and history partthat is
C0 Dα
t u(t)|t=tn =1
Γ(1minusα)
int tn
0
uprime(s)ds
(tnminuss)α
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 653
=1
Γ(1minusα)
int tn
tnminus1
uprime(s)ds
(tnminuss)α+
1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=Cl(tn)+Ch(tn) (21)
where the last equality defines the local part and the history part respectively For thelocal part we apply the standard L1 approximation which approximates u(s) on [tnminus1tn]by a linear polynomial (with u(tnminus1) and u(tn) as the interpolation nodes) or uprime(s) by a
constant u(tn)minusu(tnminus1)∆tn
We have
Cl(tn)asympu(tn)minusu(tnminus1)
∆tnΓ(1minusα)
int tn
tnminus1
1
(tnminuss)αds=
u(tn)minusu(tnminus1)
∆tαnΓ(2minusα)
(22)
For the history part we apply the integration by parts to eliminate uprime(s) and have
Ch(tn)=1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαint tnminus1
0
u(s)ds
(tnminuss)1+α
]
(23)
21 Efficient sum-of-exponentials approximation for the power function
We now show that the kernel tminusβ (0ltβlt2) can be approximated via a sum-of-exponentialsapproximation efficiently on the interval [δT] with δ=min1lenleN ∆tn and the absolute er-ror ε That is there exist positive real numbers si and wi (i = 1middotmiddotmiddot Nexp) such that for0ltβlt2
∣
∣
∣
∣
∣
1
tβminus
Nexp
sumi=1
ωieminussit
∣
∣
∣
∣
∣
le ε tisin [δT] (24)
where Nexp is given by (13) The evaluation of the history part via (23) only needsthe case 1 lt β lt 2 However one may also evaluate the history part directly withoutintegration by parts and the case 0 lt β lt 1 is then needed Meanwhile one can ob-tain the code for si and ωi in (24) from the homepage httpswebnjitedu~jiang
or httpwwwcsrcaccn~jwzhangWe start our proof from the following integral representation of the power function
Lemma 21 For any βgt01
tβ=
1
Γ(β)
int infin
0eminustmiddotssβminus1ds (25)
Proof This follows from a change of variable x= tmiddots and the integral definition of the Γ
function [49]
Eq (25) can be viewed as a representation of tminusβ using an infinitely many (continu-ous) exponentials In order to obtain an efficient SOE approximation we first truncate the
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
654 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
integral to a finite interval then subdivide the finite interval into a set of dyadic intervalsand discretize the integral on each dyadic interval with proper quadratures
We now assume 0ltβlt2
Lemma 22 For tgeδgt0
∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
le
eminusδp pβminus1
Γ(β)1δ 0ltβle1
eminusδp2βminus1(
pβ
Γ(β)+ 1
δβ
)
1ltβlt2(26)
Proof For the case of 0ltβle1∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp pβminus1
Γ(β)
int infin
0eminusδxdx
le eminusδp pβminus1
Γ(β)
1
δ (27)
For the case of 1ltβlt2∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp 1
Γ(β)
(
int p
0(2p)βminus1dx+
int infin
peminusδx(2x)βminus1dx
)
le eminusδp 2βminus1
Γ(β)
(
pβ+int infin
0eminusδxxβminus1dx
)
le eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
(28)
This completes the proof
Remark 21 The truncation error can be made arbitrarily small for fixed δ by choosingsufficiently large p Usually we have δlt1 and if one would like to bound the truncationerror by εlt1e then δpgt1 or pgt1δ When 0ltβle1
eminusδp pβminus1
Γ(β)
1
δlt eminusδp pβ
Γ(β)pδlt eminusδp p (29)
when 1ltβlt2
eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
lt eminusδp2βminus1
(
pβ
Γ(β)+pβ
)
lt5eminusδp p2 (210)
Thus p=O(
log(
1εδ
)
δ)
is sufficient to bound the truncation error by ε
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 655
We now proceed to discuss the discretization error for the integral on the interval[0p] On each dyadic interval we will discretize the integral via standard Gauss-Legendrequadrature and analyze the discretization error The error can be controlled more or lessuniformly since each dyadic interval is separated from the singular point ie the originby its own length
Lemma 23 Consider a dyadic interval [ab]= [2j 2j+1] and let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn be thenodes and weights for n-point Gauss-Legendre quadrature on the interval Then for βisin(02) andngt1
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
lt252 πaβ
(
e1e
4
)2n
(211)
Proof For any interval [ab] the standard estimate for n-point Gauss-Legendre quadra-ture [49] yields
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le (bminusa)2n+1
2n+1
(n)4
[(2n)]3max
sisin(ab)
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣ (212)
where Ds denotes the derivative with respect to s Applying Stirlingrsquos approximation [49]radic
2πnn+12eminusnltnlt2
radicπnn+12eminusn (213)
we obtain(n)4
[(2n)]3lt2
radicπ( e
8
)2nradic
n
n2n (214)
Observe now |(βminus1)middotmiddotmiddot(βminusk)|le k for kgt1 and thus∣
∣
∣Dk
s sβminus1∣
∣
∣= sβminus1le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=0∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)sβminuskminus1
∣
∣
∣le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=1∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)middotmiddotmiddot(βminusk)sβminuskminus1
∣
∣
∣le ksβminuskminus1
le2radic
πkk+12eminusksβminuskminus1le2radic
πradic
2n(2n)keminusksβminuskminus1 for kgt1
(215)
We also haveD2nminusk
s eminusst=(minust)2nminuskeminusst (216)
Combining (215) (216) with the Leibniz rule we obtain
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣=
∣
∣
∣
∣
∣
2n
sumk=0
(
2nk
)
(
D2nminusks eminusst
)(
Dks sβminus1
)
∣
∣
∣
∣
∣
le2radic
πradic
2nsβminus1eminusst2n
sumk=0
(
2nk
)
t2nminusk(2n)keminusksminusk
=2radic
πradic
2nsβminus1eminusst
(
t+2n
es
)2n
(217)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
656 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Combining (212) (214) (217) and bminusa= a we have
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le maxalesleb
2radic
2π(bminusa)sβminus1eminusst
(
e(bminusa)t
8n+
bminusa
4s
)2n
le252 πaβeminusat
(
eat
8n+
1
4
)2n
(218)
And (211) follows from the fact
maxxgt0
eminusx
(
ex
8n+
1
4
)2n
=
(
e1e
4
)2n
nge2 (219)
This completes the proof
We now consider the end interval [0a] We will use Gauss-Jacobi quadrature with theweight function sβminus1 to discretize the integral Under this weight the function to be in-tegrated is a simple exponential function Thus the discretization error can be controlledto be arbitrarily small with properly chosen parameters
Lemma 24 Let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn (nge 2) be the nodes and weights for n-point Gauss-Jacobi quadrature with the weight function sβminus1 on the interval Then for 0lt tltT βisin(02) andngt1
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
lt4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(220)
Proof The standard estimate for n-point Gauss-Jacobi quadrature [49] yields
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le a2n+β
2n+β
(n)2[Γ(n+β)]2
(2n)[Γ(2n+β)]2max
sisin(0a)
∣
∣D2ns eminusst
∣
∣ (221)
For ngt 1 we have Γ(n+β)lt Γ(n+lceilβrceil)= (n+lceilβrceilminus1) Γ(2n+β)gt Γ(2n+lfloorβrfloor)= (2n+lfloorβrfloorminus1) 2n+βgt2n+lfloorβrfloor Thus
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le max0leslea
a2n+β
2n+lfloorβrfloor(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2t2neminusst
le aβ
2n
(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2(aT)2n
le4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(222)
where the last inequality follows from Stirlingrsquos approximation (213) and the fact thatlceilβrceilminuslfloorβrfloorle1
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
652 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
That is for fixed precision ε we have Nexp =O(logNT) for T ≫ 1 or Nexp =O(log2 NT)
for T asymp 1 assuming that NT = T∆t The approximation can be used to accelerate the
evaluation of the convolution via the standard recurrence relation The resulting al-gorithm has nearly optimal complexity ndash O(NT Nexp) work and O(Nexp) storage Wewould like to remark here that the SOE approximations have been applied to speedup the evaluation of the convolution integrals in many applications (see for example[1 6 20 21 24 26 31 38 45 50 53 54 58 59] and references therein)
When we incorporate the fast evaluation scheme of the Caputo fractional derivativeto solve the fractional diffusion equations (both linear and nonlinear) the resulting algo-rithm for solving the fractional PDEs is both efficient and stable The computational costof the new algorithm is O(NS NT Nexp) as compared with O(NSN2
T) for direct methodsand the storage requirement is only O(NSNexp) as compared with O(NS NT) for directmethods since one needs to store the solution in the whole computational spatial domainat all times Furthermore we have carried out a rigorous and detailed analysis to provethat our scheme is unconditionally stable with respect to arbitrary step sizes With thesetwo properties our scheme provides an efficient and reliable tool for long time large scalesimulation of fractional PDEs In literatures there are many other efforts to speed up theevaluation of weakly singular kernel Lubich and Schadle [44] present a fast convolutionfor non-reflecting boundary conditions Baffet and Hesthaven [3] compressed the kernelin the Laplace domain and obtained a sum-of-poles approximation for the Laplace trans-form of the kernel And other fast approximation to time-fractional derivative is based onthe block triangular Toeplitz matrix or block triangular Toeplitz-like matrix see [27 43]
The paper is organized as follows In Section 2 we describe the fast algorithm for theevaluation of the Caputo fractional derivative and provide rigorous error analysis of ourdiscretization scheme In Section 3 we apply our fast algorithm to solve the linear frac-tional diffusion PDEs and present the stability and error analysis for the overall schemeIn Section 4 we study the nonlinear fractional diffusion PDEs and demonstrate that ourfast algorithm has the same order of convergence as the direct method in this case Fi-nally we conclude our paper with a brief discussion on the extension and generalizationof our scheme
2 Fast evaluation of the Caputo fractional derivative
In this section we consider the fast evaluation of the Caputo fractional derivative for0ltαlt1 Suppose that we would like to evaluate the Caputo fractional derivative on theinterval [0T] over a set of grid points Ωt =tn n=01middotmiddotmiddot NT with t0 =0 tNT
=T and∆tn = tnminustnminus1
We first split the convolution integral in (12) into a sum of local part and history partthat is
C0 Dα
t u(t)|t=tn =1
Γ(1minusα)
int tn
0
uprime(s)ds
(tnminuss)α
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 653
=1
Γ(1minusα)
int tn
tnminus1
uprime(s)ds
(tnminuss)α+
1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=Cl(tn)+Ch(tn) (21)
where the last equality defines the local part and the history part respectively For thelocal part we apply the standard L1 approximation which approximates u(s) on [tnminus1tn]by a linear polynomial (with u(tnminus1) and u(tn) as the interpolation nodes) or uprime(s) by a
constant u(tn)minusu(tnminus1)∆tn
We have
Cl(tn)asympu(tn)minusu(tnminus1)
∆tnΓ(1minusα)
int tn
tnminus1
1
(tnminuss)αds=
u(tn)minusu(tnminus1)
∆tαnΓ(2minusα)
(22)
For the history part we apply the integration by parts to eliminate uprime(s) and have
Ch(tn)=1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαint tnminus1
0
u(s)ds
(tnminuss)1+α
]
(23)
21 Efficient sum-of-exponentials approximation for the power function
We now show that the kernel tminusβ (0ltβlt2) can be approximated via a sum-of-exponentialsapproximation efficiently on the interval [δT] with δ=min1lenleN ∆tn and the absolute er-ror ε That is there exist positive real numbers si and wi (i = 1middotmiddotmiddot Nexp) such that for0ltβlt2
∣
∣
∣
∣
∣
1
tβminus
Nexp
sumi=1
ωieminussit
∣
∣
∣
∣
∣
le ε tisin [δT] (24)
where Nexp is given by (13) The evaluation of the history part via (23) only needsthe case 1 lt β lt 2 However one may also evaluate the history part directly withoutintegration by parts and the case 0 lt β lt 1 is then needed Meanwhile one can ob-tain the code for si and ωi in (24) from the homepage httpswebnjitedu~jiang
or httpwwwcsrcaccn~jwzhangWe start our proof from the following integral representation of the power function
Lemma 21 For any βgt01
tβ=
1
Γ(β)
int infin
0eminustmiddotssβminus1ds (25)
Proof This follows from a change of variable x= tmiddots and the integral definition of the Γ
function [49]
Eq (25) can be viewed as a representation of tminusβ using an infinitely many (continu-ous) exponentials In order to obtain an efficient SOE approximation we first truncate the
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
654 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
integral to a finite interval then subdivide the finite interval into a set of dyadic intervalsand discretize the integral on each dyadic interval with proper quadratures
We now assume 0ltβlt2
Lemma 22 For tgeδgt0
∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
le
eminusδp pβminus1
Γ(β)1δ 0ltβle1
eminusδp2βminus1(
pβ
Γ(β)+ 1
δβ
)
1ltβlt2(26)
Proof For the case of 0ltβle1∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp pβminus1
Γ(β)
int infin
0eminusδxdx
le eminusδp pβminus1
Γ(β)
1
δ (27)
For the case of 1ltβlt2∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp 1
Γ(β)
(
int p
0(2p)βminus1dx+
int infin
peminusδx(2x)βminus1dx
)
le eminusδp 2βminus1
Γ(β)
(
pβ+int infin
0eminusδxxβminus1dx
)
le eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
(28)
This completes the proof
Remark 21 The truncation error can be made arbitrarily small for fixed δ by choosingsufficiently large p Usually we have δlt1 and if one would like to bound the truncationerror by εlt1e then δpgt1 or pgt1δ When 0ltβle1
eminusδp pβminus1
Γ(β)
1
δlt eminusδp pβ
Γ(β)pδlt eminusδp p (29)
when 1ltβlt2
eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
lt eminusδp2βminus1
(
pβ
Γ(β)+pβ
)
lt5eminusδp p2 (210)
Thus p=O(
log(
1εδ
)
δ)
is sufficient to bound the truncation error by ε
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 655
We now proceed to discuss the discretization error for the integral on the interval[0p] On each dyadic interval we will discretize the integral via standard Gauss-Legendrequadrature and analyze the discretization error The error can be controlled more or lessuniformly since each dyadic interval is separated from the singular point ie the originby its own length
Lemma 23 Consider a dyadic interval [ab]= [2j 2j+1] and let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn be thenodes and weights for n-point Gauss-Legendre quadrature on the interval Then for βisin(02) andngt1
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
lt252 πaβ
(
e1e
4
)2n
(211)
Proof For any interval [ab] the standard estimate for n-point Gauss-Legendre quadra-ture [49] yields
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le (bminusa)2n+1
2n+1
(n)4
[(2n)]3max
sisin(ab)
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣ (212)
where Ds denotes the derivative with respect to s Applying Stirlingrsquos approximation [49]radic
2πnn+12eminusnltnlt2
radicπnn+12eminusn (213)
we obtain(n)4
[(2n)]3lt2
radicπ( e
8
)2nradic
n
n2n (214)
Observe now |(βminus1)middotmiddotmiddot(βminusk)|le k for kgt1 and thus∣
∣
∣Dk
s sβminus1∣
∣
∣= sβminus1le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=0∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)sβminuskminus1
∣
∣
∣le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=1∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)middotmiddotmiddot(βminusk)sβminuskminus1
∣
∣
∣le ksβminuskminus1
le2radic
πkk+12eminusksβminuskminus1le2radic
πradic
2n(2n)keminusksβminuskminus1 for kgt1
(215)
We also haveD2nminusk
s eminusst=(minust)2nminuskeminusst (216)
Combining (215) (216) with the Leibniz rule we obtain
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣=
∣
∣
∣
∣
∣
2n
sumk=0
(
2nk
)
(
D2nminusks eminusst
)(
Dks sβminus1
)
∣
∣
∣
∣
∣
le2radic
πradic
2nsβminus1eminusst2n
sumk=0
(
2nk
)
t2nminusk(2n)keminusksminusk
=2radic
πradic
2nsβminus1eminusst
(
t+2n
es
)2n
(217)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
656 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Combining (212) (214) (217) and bminusa= a we have
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le maxalesleb
2radic
2π(bminusa)sβminus1eminusst
(
e(bminusa)t
8n+
bminusa
4s
)2n
le252 πaβeminusat
(
eat
8n+
1
4
)2n
(218)
And (211) follows from the fact
maxxgt0
eminusx
(
ex
8n+
1
4
)2n
=
(
e1e
4
)2n
nge2 (219)
This completes the proof
We now consider the end interval [0a] We will use Gauss-Jacobi quadrature with theweight function sβminus1 to discretize the integral Under this weight the function to be in-tegrated is a simple exponential function Thus the discretization error can be controlledto be arbitrarily small with properly chosen parameters
Lemma 24 Let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn (nge 2) be the nodes and weights for n-point Gauss-Jacobi quadrature with the weight function sβminus1 on the interval Then for 0lt tltT βisin(02) andngt1
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
lt4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(220)
Proof The standard estimate for n-point Gauss-Jacobi quadrature [49] yields
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le a2n+β
2n+β
(n)2[Γ(n+β)]2
(2n)[Γ(2n+β)]2max
sisin(0a)
∣
∣D2ns eminusst
∣
∣ (221)
For ngt 1 we have Γ(n+β)lt Γ(n+lceilβrceil)= (n+lceilβrceilminus1) Γ(2n+β)gt Γ(2n+lfloorβrfloor)= (2n+lfloorβrfloorminus1) 2n+βgt2n+lfloorβrfloor Thus
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le max0leslea
a2n+β
2n+lfloorβrfloor(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2t2neminusst
le aβ
2n
(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2(aT)2n
le4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(222)
where the last inequality follows from Stirlingrsquos approximation (213) and the fact thatlceilβrceilminuslfloorβrfloorle1
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 653
=1
Γ(1minusα)
int tn
tnminus1
uprime(s)ds
(tnminuss)α+
1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=Cl(tn)+Ch(tn) (21)
where the last equality defines the local part and the history part respectively For thelocal part we apply the standard L1 approximation which approximates u(s) on [tnminus1tn]by a linear polynomial (with u(tnminus1) and u(tn) as the interpolation nodes) or uprime(s) by a
constant u(tn)minusu(tnminus1)∆tn
We have
Cl(tn)asympu(tn)minusu(tnminus1)
∆tnΓ(1minusα)
int tn
tnminus1
1
(tnminuss)αds=
u(tn)minusu(tnminus1)
∆tαnΓ(2minusα)
(22)
For the history part we apply the integration by parts to eliminate uprime(s) and have
Ch(tn)=1
Γ(1minusα)
int tnminus1
0
uprime(s)ds
(tnminuss)α
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαint tnminus1
0
u(s)ds
(tnminuss)1+α
]
(23)
21 Efficient sum-of-exponentials approximation for the power function
We now show that the kernel tminusβ (0ltβlt2) can be approximated via a sum-of-exponentialsapproximation efficiently on the interval [δT] with δ=min1lenleN ∆tn and the absolute er-ror ε That is there exist positive real numbers si and wi (i = 1middotmiddotmiddot Nexp) such that for0ltβlt2
∣
∣
∣
∣
∣
1
tβminus
Nexp
sumi=1
ωieminussit
∣
∣
∣
∣
∣
le ε tisin [δT] (24)
where Nexp is given by (13) The evaluation of the history part via (23) only needsthe case 1 lt β lt 2 However one may also evaluate the history part directly withoutintegration by parts and the case 0 lt β lt 1 is then needed Meanwhile one can ob-tain the code for si and ωi in (24) from the homepage httpswebnjitedu~jiang
or httpwwwcsrcaccn~jwzhangWe start our proof from the following integral representation of the power function
Lemma 21 For any βgt01
tβ=
1
Γ(β)
int infin
0eminustmiddotssβminus1ds (25)
Proof This follows from a change of variable x= tmiddots and the integral definition of the Γ
function [49]
Eq (25) can be viewed as a representation of tminusβ using an infinitely many (continu-ous) exponentials In order to obtain an efficient SOE approximation we first truncate the
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
654 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
integral to a finite interval then subdivide the finite interval into a set of dyadic intervalsand discretize the integral on each dyadic interval with proper quadratures
We now assume 0ltβlt2
Lemma 22 For tgeδgt0
∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
le
eminusδp pβminus1
Γ(β)1δ 0ltβle1
eminusδp2βminus1(
pβ
Γ(β)+ 1
δβ
)
1ltβlt2(26)
Proof For the case of 0ltβle1∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp pβminus1
Γ(β)
int infin
0eminusδxdx
le eminusδp pβminus1
Γ(β)
1
δ (27)
For the case of 1ltβlt2∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp 1
Γ(β)
(
int p
0(2p)βminus1dx+
int infin
peminusδx(2x)βminus1dx
)
le eminusδp 2βminus1
Γ(β)
(
pβ+int infin
0eminusδxxβminus1dx
)
le eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
(28)
This completes the proof
Remark 21 The truncation error can be made arbitrarily small for fixed δ by choosingsufficiently large p Usually we have δlt1 and if one would like to bound the truncationerror by εlt1e then δpgt1 or pgt1δ When 0ltβle1
eminusδp pβminus1
Γ(β)
1
δlt eminusδp pβ
Γ(β)pδlt eminusδp p (29)
when 1ltβlt2
eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
lt eminusδp2βminus1
(
pβ
Γ(β)+pβ
)
lt5eminusδp p2 (210)
Thus p=O(
log(
1εδ
)
δ)
is sufficient to bound the truncation error by ε
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 655
We now proceed to discuss the discretization error for the integral on the interval[0p] On each dyadic interval we will discretize the integral via standard Gauss-Legendrequadrature and analyze the discretization error The error can be controlled more or lessuniformly since each dyadic interval is separated from the singular point ie the originby its own length
Lemma 23 Consider a dyadic interval [ab]= [2j 2j+1] and let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn be thenodes and weights for n-point Gauss-Legendre quadrature on the interval Then for βisin(02) andngt1
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
lt252 πaβ
(
e1e
4
)2n
(211)
Proof For any interval [ab] the standard estimate for n-point Gauss-Legendre quadra-ture [49] yields
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le (bminusa)2n+1
2n+1
(n)4
[(2n)]3max
sisin(ab)
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣ (212)
where Ds denotes the derivative with respect to s Applying Stirlingrsquos approximation [49]radic
2πnn+12eminusnltnlt2
radicπnn+12eminusn (213)
we obtain(n)4
[(2n)]3lt2
radicπ( e
8
)2nradic
n
n2n (214)
Observe now |(βminus1)middotmiddotmiddot(βminusk)|le k for kgt1 and thus∣
∣
∣Dk
s sβminus1∣
∣
∣= sβminus1le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=0∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)sβminuskminus1
∣
∣
∣le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=1∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)middotmiddotmiddot(βminusk)sβminuskminus1
∣
∣
∣le ksβminuskminus1
le2radic
πkk+12eminusksβminuskminus1le2radic
πradic
2n(2n)keminusksβminuskminus1 for kgt1
(215)
We also haveD2nminusk
s eminusst=(minust)2nminuskeminusst (216)
Combining (215) (216) with the Leibniz rule we obtain
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣=
∣
∣
∣
∣
∣
2n
sumk=0
(
2nk
)
(
D2nminusks eminusst
)(
Dks sβminus1
)
∣
∣
∣
∣
∣
le2radic
πradic
2nsβminus1eminusst2n
sumk=0
(
2nk
)
t2nminusk(2n)keminusksminusk
=2radic
πradic
2nsβminus1eminusst
(
t+2n
es
)2n
(217)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
656 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Combining (212) (214) (217) and bminusa= a we have
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le maxalesleb
2radic
2π(bminusa)sβminus1eminusst
(
e(bminusa)t
8n+
bminusa
4s
)2n
le252 πaβeminusat
(
eat
8n+
1
4
)2n
(218)
And (211) follows from the fact
maxxgt0
eminusx
(
ex
8n+
1
4
)2n
=
(
e1e
4
)2n
nge2 (219)
This completes the proof
We now consider the end interval [0a] We will use Gauss-Jacobi quadrature with theweight function sβminus1 to discretize the integral Under this weight the function to be in-tegrated is a simple exponential function Thus the discretization error can be controlledto be arbitrarily small with properly chosen parameters
Lemma 24 Let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn (nge 2) be the nodes and weights for n-point Gauss-Jacobi quadrature with the weight function sβminus1 on the interval Then for 0lt tltT βisin(02) andngt1
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
lt4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(220)
Proof The standard estimate for n-point Gauss-Jacobi quadrature [49] yields
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le a2n+β
2n+β
(n)2[Γ(n+β)]2
(2n)[Γ(2n+β)]2max
sisin(0a)
∣
∣D2ns eminusst
∣
∣ (221)
For ngt 1 we have Γ(n+β)lt Γ(n+lceilβrceil)= (n+lceilβrceilminus1) Γ(2n+β)gt Γ(2n+lfloorβrfloor)= (2n+lfloorβrfloorminus1) 2n+βgt2n+lfloorβrfloor Thus
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le max0leslea
a2n+β
2n+lfloorβrfloor(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2t2neminusst
le aβ
2n
(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2(aT)2n
le4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(222)
where the last inequality follows from Stirlingrsquos approximation (213) and the fact thatlceilβrceilminuslfloorβrfloorle1
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
654 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
integral to a finite interval then subdivide the finite interval into a set of dyadic intervalsand discretize the integral on each dyadic interval with proper quadratures
We now assume 0ltβlt2
Lemma 22 For tgeδgt0
∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
le
eminusδp pβminus1
Γ(β)1δ 0ltβle1
eminusδp2βminus1(
pβ
Γ(β)+ 1
δβ
)
1ltβlt2(26)
Proof For the case of 0ltβle1∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp pβminus1
Γ(β)
int infin
0eminusδxdx
le eminusδp pβminus1
Γ(β)
1
δ (27)
For the case of 1ltβlt2∣
∣
∣
∣
1
Γ(β)
int infin
peminustssβminus1ds
∣
∣
∣
∣
=
∣
∣
∣
∣
1
Γ(β)eminustp
int infin
0eminustx(x+p)βminus1dx
∣
∣
∣
∣
le eminusδp 1
Γ(β)
(
int p
0(2p)βminus1dx+
int infin
peminusδx(2x)βminus1dx
)
le eminusδp 2βminus1
Γ(β)
(
pβ+int infin
0eminusδxxβminus1dx
)
le eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
(28)
This completes the proof
Remark 21 The truncation error can be made arbitrarily small for fixed δ by choosingsufficiently large p Usually we have δlt1 and if one would like to bound the truncationerror by εlt1e then δpgt1 or pgt1δ When 0ltβle1
eminusδp pβminus1
Γ(β)
1
δlt eminusδp pβ
Γ(β)pδlt eminusδp p (29)
when 1ltβlt2
eminusδp2βminus1
(
pβ
Γ(β)+
1
δβ
)
lt eminusδp2βminus1
(
pβ
Γ(β)+pβ
)
lt5eminusδp p2 (210)
Thus p=O(
log(
1εδ
)
δ)
is sufficient to bound the truncation error by ε
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 655
We now proceed to discuss the discretization error for the integral on the interval[0p] On each dyadic interval we will discretize the integral via standard Gauss-Legendrequadrature and analyze the discretization error The error can be controlled more or lessuniformly since each dyadic interval is separated from the singular point ie the originby its own length
Lemma 23 Consider a dyadic interval [ab]= [2j 2j+1] and let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn be thenodes and weights for n-point Gauss-Legendre quadrature on the interval Then for βisin(02) andngt1
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
lt252 πaβ
(
e1e
4
)2n
(211)
Proof For any interval [ab] the standard estimate for n-point Gauss-Legendre quadra-ture [49] yields
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le (bminusa)2n+1
2n+1
(n)4
[(2n)]3max
sisin(ab)
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣ (212)
where Ds denotes the derivative with respect to s Applying Stirlingrsquos approximation [49]radic
2πnn+12eminusnltnlt2
radicπnn+12eminusn (213)
we obtain(n)4
[(2n)]3lt2
radicπ( e
8
)2nradic
n
n2n (214)
Observe now |(βminus1)middotmiddotmiddot(βminusk)|le k for kgt1 and thus∣
∣
∣Dk
s sβminus1∣
∣
∣= sβminus1le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=0∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)sβminuskminus1
∣
∣
∣le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=1∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)middotmiddotmiddot(βminusk)sβminuskminus1
∣
∣
∣le ksβminuskminus1
le2radic
πkk+12eminusksβminuskminus1le2radic
πradic
2n(2n)keminusksβminuskminus1 for kgt1
(215)
We also haveD2nminusk
s eminusst=(minust)2nminuskeminusst (216)
Combining (215) (216) with the Leibniz rule we obtain
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣=
∣
∣
∣
∣
∣
2n
sumk=0
(
2nk
)
(
D2nminusks eminusst
)(
Dks sβminus1
)
∣
∣
∣
∣
∣
le2radic
πradic
2nsβminus1eminusst2n
sumk=0
(
2nk
)
t2nminusk(2n)keminusksminusk
=2radic
πradic
2nsβminus1eminusst
(
t+2n
es
)2n
(217)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
656 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Combining (212) (214) (217) and bminusa= a we have
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le maxalesleb
2radic
2π(bminusa)sβminus1eminusst
(
e(bminusa)t
8n+
bminusa
4s
)2n
le252 πaβeminusat
(
eat
8n+
1
4
)2n
(218)
And (211) follows from the fact
maxxgt0
eminusx
(
ex
8n+
1
4
)2n
=
(
e1e
4
)2n
nge2 (219)
This completes the proof
We now consider the end interval [0a] We will use Gauss-Jacobi quadrature with theweight function sβminus1 to discretize the integral Under this weight the function to be in-tegrated is a simple exponential function Thus the discretization error can be controlledto be arbitrarily small with properly chosen parameters
Lemma 24 Let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn (nge 2) be the nodes and weights for n-point Gauss-Jacobi quadrature with the weight function sβminus1 on the interval Then for 0lt tltT βisin(02) andngt1
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
lt4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(220)
Proof The standard estimate for n-point Gauss-Jacobi quadrature [49] yields
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le a2n+β
2n+β
(n)2[Γ(n+β)]2
(2n)[Γ(2n+β)]2max
sisin(0a)
∣
∣D2ns eminusst
∣
∣ (221)
For ngt 1 we have Γ(n+β)lt Γ(n+lceilβrceil)= (n+lceilβrceilminus1) Γ(2n+β)gt Γ(2n+lfloorβrfloor)= (2n+lfloorβrfloorminus1) 2n+βgt2n+lfloorβrfloor Thus
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le max0leslea
a2n+β
2n+lfloorβrfloor(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2t2neminusst
le aβ
2n
(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2(aT)2n
le4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(222)
where the last inequality follows from Stirlingrsquos approximation (213) and the fact thatlceilβrceilminuslfloorβrfloorle1
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 655
We now proceed to discuss the discretization error for the integral on the interval[0p] On each dyadic interval we will discretize the integral via standard Gauss-Legendrequadrature and analyze the discretization error The error can be controlled more or lessuniformly since each dyadic interval is separated from the singular point ie the originby its own length
Lemma 23 Consider a dyadic interval [ab]= [2j 2j+1] and let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn be thenodes and weights for n-point Gauss-Legendre quadrature on the interval Then for βisin(02) andngt1
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
lt252 πaβ
(
e1e
4
)2n
(211)
Proof For any interval [ab] the standard estimate for n-point Gauss-Legendre quadra-ture [49] yields
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le (bminusa)2n+1
2n+1
(n)4
[(2n)]3max
sisin(ab)
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣ (212)
where Ds denotes the derivative with respect to s Applying Stirlingrsquos approximation [49]radic
2πnn+12eminusnltnlt2
radicπnn+12eminusn (213)
we obtain(n)4
[(2n)]3lt2
radicπ( e
8
)2nradic
n
n2n (214)
Observe now |(βminus1)middotmiddotmiddot(βminusk)|le k for kgt1 and thus∣
∣
∣Dk
s sβminus1∣
∣
∣= sβminus1le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=0∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)sβminuskminus1
∣
∣
∣le2
radicπradic
2n(2n)keminusksβminuskminus1 for k=1∣
∣
∣Dk
s sβminus1∣
∣
∣=∣
∣
∣(βminus1)middotmiddotmiddot(βminusk)sβminuskminus1
∣
∣
∣le ksβminuskminus1
le2radic
πkk+12eminusksβminuskminus1le2radic
πradic
2n(2n)keminusksβminuskminus1 for kgt1
(215)
We also haveD2nminusk
s eminusst=(minust)2nminuskeminusst (216)
Combining (215) (216) with the Leibniz rule we obtain
∣
∣
∣D2n
s (eminusstsβminus1)∣
∣
∣=
∣
∣
∣
∣
∣
2n
sumk=0
(
2nk
)
(
D2nminusks eminusst
)(
Dks sβminus1
)
∣
∣
∣
∣
∣
le2radic
πradic
2nsβminus1eminusst2n
sumk=0
(
2nk
)
t2nminusk(2n)keminusksminusk
=2radic
πradic
2nsβminus1eminusst
(
t+2n
es
)2n
(217)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
656 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Combining (212) (214) (217) and bminusa= a we have
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le maxalesleb
2radic
2π(bminusa)sβminus1eminusst
(
e(bminusa)t
8n+
bminusa
4s
)2n
le252 πaβeminusat
(
eat
8n+
1
4
)2n
(218)
And (211) follows from the fact
maxxgt0
eminusx
(
ex
8n+
1
4
)2n
=
(
e1e
4
)2n
nge2 (219)
This completes the proof
We now consider the end interval [0a] We will use Gauss-Jacobi quadrature with theweight function sβminus1 to discretize the integral Under this weight the function to be in-tegrated is a simple exponential function Thus the discretization error can be controlledto be arbitrarily small with properly chosen parameters
Lemma 24 Let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn (nge 2) be the nodes and weights for n-point Gauss-Jacobi quadrature with the weight function sβminus1 on the interval Then for 0lt tltT βisin(02) andngt1
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
lt4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(220)
Proof The standard estimate for n-point Gauss-Jacobi quadrature [49] yields
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le a2n+β
2n+β
(n)2[Γ(n+β)]2
(2n)[Γ(2n+β)]2max
sisin(0a)
∣
∣D2ns eminusst
∣
∣ (221)
For ngt 1 we have Γ(n+β)lt Γ(n+lceilβrceil)= (n+lceilβrceilminus1) Γ(2n+β)gt Γ(2n+lfloorβrfloor)= (2n+lfloorβrfloorminus1) 2n+βgt2n+lfloorβrfloor Thus
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le max0leslea
a2n+β
2n+lfloorβrfloor(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2t2neminusst
le aβ
2n
(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2(aT)2n
le4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(222)
where the last inequality follows from Stirlingrsquos approximation (213) and the fact thatlceilβrceilminuslfloorβrfloorle1
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
656 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Combining (212) (214) (217) and bminusa= a we have
∣
∣
∣
∣
∣
int b
aeminustssβminus1dsminus
n
sumk=1
wksβminus1k eminusskt
∣
∣
∣
∣
∣
le maxalesleb
2radic
2π(bminusa)sβminus1eminusst
(
e(bminusa)t
8n+
bminusa
4s
)2n
le252 πaβeminusat
(
eat
8n+
1
4
)2n
(218)
And (211) follows from the fact
maxxgt0
eminusx
(
ex
8n+
1
4
)2n
=
(
e1e
4
)2n
nge2 (219)
This completes the proof
We now consider the end interval [0a] We will use Gauss-Jacobi quadrature with theweight function sβminus1 to discretize the integral Under this weight the function to be in-tegrated is a simple exponential function Thus the discretization error can be controlledto be arbitrarily small with properly chosen parameters
Lemma 24 Let s1middotmiddotmiddot sn and w1middotmiddotmiddot wn (nge 2) be the nodes and weights for n-point Gauss-Jacobi quadrature with the weight function sβminus1 on the interval Then for 0lt tltT βisin(02) andngt1
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
lt4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(220)
Proof The standard estimate for n-point Gauss-Jacobi quadrature [49] yields
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le a2n+β
2n+β
(n)2[Γ(n+β)]2
(2n)[Γ(2n+β)]2max
sisin(0a)
∣
∣D2ns eminusst
∣
∣ (221)
For ngt 1 we have Γ(n+β)lt Γ(n+lceilβrceil)= (n+lceilβrceilminus1) Γ(2n+β)gt Γ(2n+lfloorβrfloor)= (2n+lfloorβrfloorminus1) 2n+βgt2n+lfloorβrfloor Thus
∣
∣
∣
∣
∣
int a
0eminustssβminus1dsminus
n
sumk=1
wkeminusskt
∣
∣
∣
∣
∣
le max0leslea
a2n+β
2n+lfloorβrfloor(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2t2neminusst
le aβ
2n
(n)2[(n+lceilβrceilminus1)]2
(2n)[(2n+lfloorβrfloorminus1)]2(aT)2n
le4radic
πaβn32( e
8
)2n(
aT
nminus1
)2n
(222)
where the last inequality follows from Stirlingrsquos approximation (213) and the fact thatlceilβrceilminuslfloorβrfloorle1
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 657
We are now in a position to combine the last three lemmas to give an efficient SOEapproximation for tminusβ on [δT] for βisin (02) as follows
Theorem 21 Let 0lt δ le t le T (δ le 1 and T ge 1) let εgt 0 be the desired precision let no =O(log 1
ε ) let M=O(logT) and let N=O(loglog 1ε +log 1
δ ) Furthermore let so1middotmiddotmiddot sono andwo1middotmiddotmiddot wono be the nodes and weights for the no-point Gauss-Jacobi quadrature on the interval[02minusM] let sj1middotmiddotmiddot sjns
and wj1middotmiddotmiddot wjnsbe the nodes and weights for ns-point Gauss-Legendre
quadrature on small intervals [2j2j+1] j=minusMmiddotmiddotmiddot minus1 where ns=O(
log 1ε
)
and let sj1middotmiddotmiddot sjnl
and wj1middotmiddotmiddot wjnsl be the nodes and weights for nl-point Gauss-Legendre quadrature on large
intervals [2j2j+1] j=0middotmiddotmiddot N where nl =O(
log 1ε +log 1
δ
)
Then for tisin [δT] and βisin (02)
∣
∣
∣
∣
∣
1
tβminus(
no
sumk=1
eminussoktwok+minus1
sumj=minusM
ns
sumk=1
eminuss jktsβminus1jk wjk+
N
sumj=0
nl
sumk=1
eminuss jktsβminus1jk wjk
)∣
∣
∣
∣
∣
le ε (223)
Remark 22 The important fact which emerges from this theorem is that the total numberof exponentials needed to approximate tminusβ for 0 lt δ le t le T with an absolute error ε isgiven by the formula (13) which gives the right complexity estimate for Nexp Howevereven though the proof of Theorem 21 is explicitly constructive and one may even obtainthe concrete number of exponentials needed by more detailed calculation the resultingnumber of exponentials following the construction in Theorem 21 is unnecessarily largeThis is because the construction is based on locally optimal quadrature on each dyadicinterval leading to a large prefactor in the O complexity
As a global after-processing optimization one may apply modified Pronyrsquos methodin [5] to reduce the number of exponentials for nodes on the interval (01) while standardmodel reduction method in [55] can be applied to reduce the number of exponentials fornodes on the interval [1p] Our numerical experiments indicate that there is about a fac-tor of ten in the reduction of the number of exponentials after these further optimizationsteps Thus we state Theorem 21 using the O notation
Table 1 and Table 2 list the actual number of exponentials needed to approximatetminus1minusα with various ǫ and NT =T∆t after applying the reduction algorithms in Remark22 We observe that the number of exponentials needed is very modest even for highaccuracy approximations Indeed one needs less than 90 terms in order to march onemillion steps with 9-digit or 8-digit accuracy
22 Fast evaluation of the history part
We replace the convolution kernel tminus1minusα by its SOE approximation in (24) to approximatethe history part defined in (23) as follows
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
658 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 1 Number of exponentials needed to approximate tminus1minusα with fixed ∆t=10minus3
∆t=10minus3 α=02 α=05
ǫ
NexpT∆t 103 104 105 106 103 104 105 106
10minus3 25 29 33 37 30 35 40 44
10minus6 34 40 45 51 39 46 52 59
10minus9 43 51 58 66 49 57 65 74
Table 2 Number of exponentials needed to approximate tminus1minusα with fixed T=1
T=1 α=02 α=05
ǫ
Nexp ∆t10minus3 10minus4 10minus5 10minus6 10minus3 10minus4 10minus5 10minus6
10minus3 25 34 43 55 30 41 52 66
10minus6 34 45 58 69 39 51 65 81
10minus8 40 51 64 79 43 57 72 84
Ch(tn)asymp1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωi
int tnminus1
0eminus(tnminusτ)siu(τ)dτ
]
=1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
(224)
Define
Uhisti(tn)=int tnminus1
0eminus(tnminusτ)siu(τ)dτ
To evaluate Uhisti(tn) for n=12middotmiddotmiddot NT we observe the following simple recurrence rela-tion
Uhisti(tn)= eminussi∆tnUhisti(tnminus1)+int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτ (225)
At each time step we only need O(1) work to compute Uhisti(tn) since Uhisti(tnminus1) isknown at that point Thus the total work is reduced from O(N2
T) to O(NT Nexp) and thetotal memory requirement is reduced from O(NT) to O(Nexp)
One may compute the integral on the right hand side of (225) by interpolating u viaa linear function and then evaluating the resulting approximation analytically We have
int tnminus1
tnminus2
eminussi(tnminusτ)u(τ)dτasymp eminussi∆tn
s2i ∆tnminus1
[
(eminussi∆tnminus1minus1+si∆tnminus1)u(tnminus1)
+ (1minuseminussi∆tnminus1minuseminussi∆tnminus1si∆tnminus1)u(tnminus2)]
(226)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 659
We note that the weights in front of u(tnminus1) and u(tnminus2) in (226) are subject to significantcancellation error when si∆tnminus1 is small In that case we can compute the weights by aTaylor expansion of exponentials with a small number of terms
Remark 23 From above analysis the fast evaluation of the Caputo fractional derivativeis valid for non-uniform mesh For simplicity we will use uniform mesh in time with∆tn =∆t in the remainder of the section
23 Error analysis
Define mesh functionsun=u(tn) 1lenleNT
It is straightforward to verify that our scheme of evaluating the Caputo fractional deriva-tive is equivalent to the following formula
C0 Dα
t u(t)|t=tn asympunminusunminus1
∆tαΓ(2minusα)+
1
Γ(1minusα)
[
unminus1
∆tαminus u0
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
FC0 D
αt un for ngt2 (227)
Noting that Uhisti(tn)=0 when n=1 we have
FC0 D
αt u1=
∆tminusα
Γ(2minusα)(u1minusu0) (228)
Recall that the L1-approximation (based on the linear interpolation of the function u) ofthe Caputo derivative C
0 Dαt u (see for example [42 48]) is defined by the formula
C0 D
αt un=
∆tminusα
Γ(2minusα)
[
a(α)0 unminus
nminus1
sumk=1
(a(α)nminuskminus1minusa
(α)nminusk)u
kminusa(α)nminus1u0
]
(229)
where a(α)k =(k+1)1minusαminusk1minusα The following lemma which can be found in [52] estab-
lishes an error bound for the L1-approximation (229)
Lemma 25 (see [40 52]) Suppose that f (t)isinC2[0tn] and
Rnu =C0 Dα
t u(t)∣
∣
t=tnminusC
0 Dαt un (230)
where 0ltαlt1 Then
|Rnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)| (231)
The following lemma provides an error bound for our fast evaluation scheme de-noted by FC
0 Dαt un in (227) and (228)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
660 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 26 Suppose that u(t)isinC2[0tn] and let
FRnu =C0 Dα
t u(t)∣
∣
t=tnminusFC
0 Dαt un (232)
where 0ltαlt1 Then
|FRnu|le ∆t2minusα
Γ(2minusα)
(
1minusα
12+
22minusα
2minusαminus(1+2minusα)
)
max0letletn
|uprimeprime(t)|+ αεtnminus1
Γ(1minusα)max
0letletnminus1
|u(t)| (233)
Proof Obviously the only difference between our approximation FC0 D
αt un and the L1-
approximation C0 D
αt un is that the convolution kernel admits an absolute error bounded
by ε in its sum-of-exponentials approximation (24) namely
∣
∣
∣
FC0 D
αt unminusC
0 Dαt un∣
∣
∣le αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (234)
where Π1lu(t)=u(tlminus1)tlminust∆t +u(tl)
tminustlminus1
∆t And the triangle inequality leads to
|FRnu|le |CRnu|+ αε
Γ(1minusα)
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|ds (235)
where
nminus1
suml=1
int tl
tlminus1
|Π1lu(s)|dsle max0letletnminus1
|u(t)|tnminus1 (236)
Combining Lemma 25 and (236) we obtain Lemma 26
We also have the following useful inequality The proof is given in Appendix A
Lemma 27 For any mesh functions g=gk |0lekleN defined on Ωt the following inequalityholds
∆tn
sumk=1
(FC0 D
αt gk)gk ge tminusα
n minus2αεtnminus1
2Γ(1minusα)∆t
n
sumk=1
(gk)2minus t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)(g0)2 (237)
The results in Lemmas 26 and 27 are useful in proving the stability and convergenceproperties of the overall numerical scheme for solving time-fractional PDEs when ourfast evaluation scheme is applied to compute the Caputo fractional derivative
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 661
3 Application I Linear fractional diffusion equation
We now consider the computation of the initial value (IV) problem of the linear fractionaldiffusion equation given by
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinR tgt0 (31)
u(x0)=u0(x) xisinR (32)
u(xt)rarr0 when |x|rarrinfin (33)
where the initial data u0 and the source term f (xt) are assumed to be compactly sup-ported in the interval Ωi =x|xl ltxltxr To solve this problem using a finite differencescheme one needs to truncate the computational domain to a finite interval and imposesome boundary conditions at the end points see [24714ndash161822] The exact nonreflect-ing boundary conditions for the above problem have been derived in [15] via standardLaplace transform method and it is shown in [15] that the above problem is equivalent tothe following initial-boundary value problem (IBVP)
C0 Dα
t u(xt)=uxx(xt)+ f (xt) xisinΩi tgt0 (34)
u(x0)=u0(x) xisinΩi (35)
partu(xt)
partx=
1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =C0 D
α2
t u(xt) x= xl (36)
partu(xt)
partx=minus 1
Γ(1minus α2 )
int t
0
us(xs)
(tminuss)α2
ds =minusC0 D
α2
t u(xt) x= xr (37)
We now incorporate the fast evaluation of the Caputo fractional derivative into the exist-ing finite difference (FD) scheme introduced in [15] to numerically solve the IBVP (34)-(37) establish the corresponding stability and convergence properties and present anexample to demonstrate the performance of our scheme by comparing it with the directscheme given in [15]
31 Construction of the finite difference scheme with fast evaluation
We first introduce some standard notations For two given positive integers NT and NS
let tnNTn=0 be a equidistant partition of [0T] with tn=n∆t and ∆t=TNT and let xjNS
j=0
be a partition of (xlxr) with xi = xl+ih and h=(xrminusxl)NS Denote
δtuni =
uni minusunminus1
i
∆t δxun
i+ 12=
uni+1minusun
i
h
δ2xun
i =δxui+ 1
2minusδxuiminus 1
2
h δxun
i =un
i+1minusuniminus1
2h
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
662 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Lemma 31 (see [52]) Suppose that uisinC3[xl xr] Then
uxx(x0)minus2
h
[
δxu 12minusux(x0)
]
=minush
3uxxx(x0+θ1h) θ1 isin (01) (38)
uxx(xNS)minus 2
h
[
ux(xNS)minusδxuNSminus 1
2
]
=h
3uxxx(xNS
minusθ2h) θ2isin (01) (39)
The FD scheme in [15] for the problem (34)-(37) can be written by
C0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (310)
C0 D
αt un
0 =2
h
[
δxun12minusC
0 Dα2t un
0
]
+ f n0 1lenleNT (311)
C0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (312)
u0i =u0(xi) 0le ileNS (313)
Replacing the standard L1-approximation C0 D for the Caputo derivative by our fast eval-
uation scheme FC0 D we obtain a fast FD scheme of the following form
FC0 D
αt un
i =δ2xun
i + f ni 1le ileNSminus1 1lenleNT (314)
FC0 D
αt un
0 =2
h
[
δxun12minusFC
0 Dα2t un
0
]
+ f n0 1lenleNT (315)
FC0 D
αt un
NS=
2
h
[
minusδxunNSminus 1
2minusFC
0 Dα2t un
NS
]
+ f nNS
1lenleNT (316)
u0i =u0(xi) 0le ileNS (317)
32 Stability and error analysis of the new scheme
The stability and error analysis of the scheme (310)-(313) have been given in [15] In thissection we study the corresponding theoretical analysis for our new scheme (314)-(317)We first denote Sh =u|u=(u0u1middotmiddotmiddot uNs)
Lemma 32 ( [15]) For any mesh function uisinSh the following inequality holds
u2infin le θδxu2+
(1
θ+
1
L
)
u2 forallθgt0
where L is the length of the computational domain and here L= xrminusxl
We now show the prior estimate for the solution of the scheme (314)-(317)
Theorem 31 (Prior Estimate) Suppose uki |0le ileNS 0lekleNT is the solution of the finite
difference scheme (314)-(317) Then for any 1lenleNT
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle2(
1+radic
1+L2micro)
Lmicro
(
ρ∥
∥u0∥
∥
2+[
(u00)
2+(u0NS)2]
+∆t
8ν
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micro
n
sumk=1
hNSminus1
sumi=1
( f ki )
2
)
(318)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 663
where
ρ=T1minusα
2Γ(2minusα) micro=
Tminusαminus2αεT
Γ(1minusα)
=T1minus α
2
2Γ(2minus α2 )
ν=Tminus α
2 minusαεT
Γ(1minus α2 )
(319)
Proof Multiplying huki on both sides of (314) and summing up for i from 1 to NSminus1 we
have
hNSminus1
sumi=1
(FC0 D
αt uk
i )uki minush
NSminus1
sumi=1
(δ2xuk
i )uki =h
NSminus1
sumi=1
f ki uk
i
Multiplying h2 uk
0 and h2 uk
NSon both sides of (315) and (316) respectively then adding the
results with the above identity we obtain
(FC0 D
αt ukuk)+
[
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS
]
+(FC0 D
α2t uk
0)uk0+(FC
0 D
α2t uk
NS)uk
NS=
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS (320)
Applying the summation by parts we have
minus(δxvk12)uk
0minushNSminus1
sumi=1
(δ2xuk
i )uki +(δxuk
NSminus 12)uk
NS=∥
∥
∥δxuk
∥
∥
∥
2 (321)
Substituting (321) into (320) and multiplying ∆t on both sides of the resulting identityand summing up for k from 1 to n it follows from Lemma 27 that
∆ttminusαn minus2αεtnminus1
2Γ(1minusα)
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
n
sumk=1
[
(uk0)
2+(ukNS)2]
+∆tn
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2
le t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα)
∥
∥u0∥
∥
2+
t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
[
(u00)
2+(u0NS)2]
+∆tn
sumk=1
[
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
]
(322)
Applying the Cauchy-Schwarz inequality we obtain
1
2(h f k
0 )uk0+h
NSminus1
sumi=1
f ki uk
i +1
2(h f k
NS)uk
NS
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
664 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+hNSminus1
sumi=1
[
tminusαn minus2αεtnminus1
4Γ(1minusα)(uk
i )2+
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2
]
le tminus α
2n minusαεtnminus1
2Γ(1minus α2 )
[
(uk0)
2+(ukNS)2]
+Γ(1minus α
2 )
8(tminus α
2n minusαεtnminus1)
[
(h f k0 )
2+(h f kNS)2]
+tminusαn minus2αεtnminus1
4Γ(1minusα)
∥
∥
∥uk∥
∥
∥
2+h
NSminus1
sumi=1
Γ(1minusα)
tminusαn minus2αεtnminus1
( f ki )
2 (323)
The substitution of (323) into (322) produces
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2leρn
∥
∥u0∥
∥
2+n
[
(u00)
2+(u0NS)2]
+∆t
8νn
n
sumk=1
[
(h f k0 )
2+(h f kNS)2]
+∆t
micron
n
sumk=1
hNSminus1
sumi=1
( f ki )
2 (324)
where
ρn =t1minusαn +α(1minusα)εtnminus1∆t
2Γ(2minusα) micron =
tminusαn minus2αεtnminus1
Γ(1minusα)
n =t1minus α
2n + α
2 (1minus α2 )εtnminus1∆t
2Γ(2minus α2 )
νn =tminus α
2n minusαεtnminus1
Γ(1minus α2 )
(325)
Taking θngt 0 such that 1θn+1L
θn = micron
4
(
ie θn = 2(
1+radic
1+L2micron)
(Lmicron))
and followingfrom Lemma 32 we have
∆tn
sumk=1
∥
∥
∥uk∥
∥
∥
2
infinle 2(
1+radic
1+L2micron)
Lmicron
(
micron
4∆t
n
sumk=1
∥
∥
∥uk∥
∥
∥
2+∆t
n
sumk=1
∥
∥
∥δxuk
∥
∥
∥
2)
(326)
Combining (324) with (326) and notice
ρn leρ n le micron gemicro νn geν
we obtain the inequality (318)
The priori estimate leads to the stability of the scheme (314)-(317)
Theorem 32 (Stability) The scheme (314)-(317) is unconditionally stable for any given com-pactly supported initial data and source term
We now present the error analysis of the scheme (314)-(317) whose proof can befound in Appendix B
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 665
Theorem 33 (Error Analysis) Suppose u(xt)isinC42xt ([xl xr]times[0T]) and uk
i |0leileNS 0lekleNT are solutions of the problem (34)-(37) and the difference scheme (314)-(317) respectivelyLet ek
i =uki minusu(xitk) Then there exists a positive constant c2 such that
radic
∆tn
sumk=1
ek2infin le c2(h
2+∆t2minusα+ε) 1lenleNT (327)
where c22 =
4c21T(
1+radic
1+L2micro)
Lmicro
(
1ν +
Lmicro
)
with c1 is a positive constant (see (B5)-(B7)) and micro ν are
defined in (319)
Remark 31 The convergence rate established in [15] for the scheme (310)-(313) is C(h2+∆t2minusα) which differs from the error estimate in (327) by a tolerance error ε see Fig 1 forexample This implies that the fast scheme (314)-(317) has almost the same accuracy andconvergence rate as the direct scheme (310)-(313) when ε is chosen to be smaller thanh2+∆t2minusα but is much faster (O(NS NT Nexp) vs O(NSN2
T) work) and needs less storage(O(NS Nexp) vs O(NSNT) memory) when the total time step number NT is moderate andlarge
33 Numerical results
In this section we illustrate the performance of our scheme (314)-(317) by comparing itwith the scheme (310)-(313) from various aspects ndash convergence rate and computationalcost We observe that the scheme (314)-(317) has the same convergence rate O(h2+∆t2minusα) but is much faster than the scheme (310)-(313) even when NT is of moderatesize
Example 31 Now we use the example in [15] with the exact solution of the problem(31)-(33)
u(xt)= x4(πminusx)4[
exp(minusx)t3+α+1]
(xt)isinΩitimes[0T] (328)
and the source term f (xt) in the form of
f (xt)=
Γ(4+α)x4(πminusx)4exp(minusx)t36minusx2(πminusx)2t3+α exp(minusx)[
x2(56minus16x+x2)minus2πx(28minus12x+x2)+π2(12minus8x+x2)]
+4(3π2minus14πx+14x2) (xt)isinΩitimes[0T]
0 (xt) isinΩitimes[0T]
We define the maximum norm of the error and the convergence rates with respect totemporal and spatial sizes respectively by the formulas
E(h∆t)=
radic
radic
radic
radic∆tNT
sumk=1
ek2infin rt = log2
E(h∆t)
E(h∆t2) rs = log2
E(h∆t)
E(h2∆t)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
666 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
Table 3 Errors and convergence orders in time with fixed spatial size h=π20000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rt E(hτ) rt E(hτ) rt E(hτ) rt
110 1570e-02 170 1570e-02 170 8151e-02 148 8151e-02 148
120 4846e-03 170 4844e-03 170 2922e-02 147 2923e-02 147
140 1489e-03 172 1488e-03 171 1052e-02 148 1052e-02 148
180 4524e-04 170 4541e-04 172 3781e-03 148 3784e-03 148
1160 1395e-04 1375e-04 1357e-03 1357e-03
Table 4 The errors convergence orders in space and CPU time in seconds with fixed temporal step size∆t=130000 for the fast scheme (314)-(317) and the direct scheme (310)-(313)
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
E(hτ) rs E(hτ) rs E(hτ) rs E(hτ) rs
π10 8862e-01 206 8632e-01 203 8652e-01 205 8035e-01 199
π20 2122e-01 201 2107e-01 201 2087e-01 201 2016e-01 199
π40 5258e-02 200 5244e-02 200 5177e-02 200 5064e-02 199
π80 1312e-02 200 1311e-02 200 1292e-02 200 1272e-02 199
π160 3280e-03 3277e-03 3229e-03 3195e-03
CPU(s) 4337 330466 4365 222606
where the error ek is measured against the exact solution (328)
In our calculation we set the computational domain Ωi=[0π] T=1 and the toleranceprecision ε=10minus7 for the SOE approximation Table 3 shows the errors in maximum normand temporal convergence rate O(∆t2minusα) for α=0205 computed by the schemes (314)-(317) and (310)-(313) respectively where the spatial mesh size is fixed at h= π
20000 AndTable 4 shows the relative Linfin errors spatial convergence rate O(h2) and CPU time forα=02 and α=05 where the temporal mesh size is fixed at ∆t= 1
30000
From Tables 3 and 4 it is not surprising to see that both schemes have the same errorsand convergence rate O(h2+∆t2minusα) The reason is that as shown in Theorem 33 theinfluence of the SOE approximation error ε is negligible when it is chosen to be verysmall To further investigate the conclusion of Theorem 33 we plot in Fig 1 the evolutionof the difference of the solutions to the schemes (314)-(317) and (310)-(313) using thesame mesh sizes One can see that the difference of the solutions is smaller than the SOEapproximation precision
In Fig 2 we plot out the CPU time (in seconds) of the two schemes by fixing all otherparameters and only varying the total number of time steps NT We observe that whilethe direct scheme scales like O(N2
T) the CPU time increases almost linearly with the totalnumber of time steps NT for the fast scheme
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 667
Figure 1 Example 31 The evolution of the difference (ie Error=FUnh minusDUn
h ) of the solutions to the schemes
(314)-(317) (denote by DUnh ) and (310)-(313) (denote by FUn
h ) for ε= 1e-71e-81e-91e-10 respectively Here
we set T=2 Ωi =[0π] M=80 N=2000 and α=05
42 44 46 48 5 5205
1
15
2
25
3
35
4
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 2 Example 31 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number oftime steps NT for Application I Here NS =30 and α=05
4 Application II Nonlinear fractional diffusion equation
In this section we apply the fast evaluation of the Caputo derivative to solve the follow-ing nonlinear fractional diffusion equation
C0 Dα
t u(xt)=uxx+ f (u)
u(x0)=u0(x)(41)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
668 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
If f (u) = minusu(1minusu) or f (u) = minus01u(1minusu)(uminus0001) (41) is called the time fractionalFisher equation or Huxley equation respectively
Suppose that the initial value u0(x) is compactly supported on Ωi = [xl xr] Efficientnonlinear absorbing boundary conditions (ABCs) have been derived in [8355657] Withthese ABCs the IVP (41) on unbounded domain can be reformulated into an IBVP on abounded domain given by
C0 Dα
t u(xt)=uxx+ f (u) 0lt tleT xisinΩi (42)
u(x0)=u0(x) xisinΩi (43)
(partx+3sα20 )
C0 Dα
t u+(3sα0partx+s
3α2
0 )u=(partx+3sα20 ) f (u) x= xr (44)
(partxminus3sα20 )
C0 Dα
t u+(3sα0partxminuss
3α2
0 )u=(partxminus3sα20 ) f (u) x= xl (45)
where s0 is the Pade expansion point that is chosen in such a way that it is close to thefrequency of the wave touching on the artificial boundaries
The discretization of the problem (42)-(45) is given as
C0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (46)
(δx+3sα20 )
C0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (47)
(δxminus3sα20 )
C0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (48)
Under the assumption that f (u) isin C2([0T]) it has been shown in [35] that the finitedifferent scheme (46)-(48) has the convergence rate of O(h2+∆t) in Linfin norm wherethe Linfin norm is defined by einfin =max0leileM |ei| Here we focus on the speed-up of theevaluation of the time fractional PDEs Thus we replace the approximation C
0 Dαt by our
fast evaluation scheme FC0 D
αt and have a new discretized scheme given by
FC0 D
αt Un
i =δ2xUn
i + f (Unminus1i ) 1le ileMminus1 1lenleNT (49)
(δx+3sα20 )
FC0 D
αt Un
Mminus1+(3sα0 δx+s
3α2
0 )UnMminus1=(δx+3s
α20 ) f (Unminus1
Mminus1) 1lenleNT (410)
(δxminus3sα20 )
FC0 D
αt Unminus1
1 +(3sα0 δxminuss
3α2
0 )Unminus11 =(δxminus3s
α20 ) f (Unminus1
1 ) 1lenleNT (411)
Using the similar process of numerical analysis as those in [35] we can obtain the conver-gence rate of O(h2+∆t+ε) in Linfin norm for the fast scheme (49)-(411) In fact one onlyneeds to replace the error estimate of C
0 Dαt in Lemma 25 by the one in Lemma 26
41 Numerical examples
We will give two examples ndash the Fisher equation and the Huxley equation to illustratethe performance of our scheme For both examples in order to investigate the con-vergence orders of our scheme the reference solution is computed over a large intervalΩr=[minus1212] with very small mesh sizes h=2minus10 and ∆t=2minus14 We then set Ωi=[minus66]
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 669
and the precision for the sum-of-exponentials approximation of the convolution kernelto ε=10minus9 and T=1 The temporal step size is fixed at ∆t=2minus14 when testing the order ofconvergence in space and the spatial step size is fixed at h=2minus10 when testing the orderof convergence in time
Example 41 We consider the time fractional Fisher equation (ie f (u) =minusu(1minusu) in(41)) with the double Gaussian initial value
u(x0)=exp(minus10(xminus05)2)+exp(minus10(x+05)2) (412)
Tables 5 and 6 show again the same accuracy and convergence order O(h2+∆t) in Linfin
norm for both schemes (46)-(48) and (49)-(411) with α=02 05 We plot out in Fig 4 theevolution of the difference of two solutions by taking ε= 1e-71e-81e-91e-10 respectivelyWe observe that the difference decreases as ε decreases Fig 3 shows the CPU time inseconds for both schemes versus NT Again one can see that the fast scheme has almostlinear complexity in NT and is much faster than the direct scheme
Table 5 Errors and convergence orders in time for the fractional Fisher equation by fixing h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 8601e-04 102 8601e-04 102 2194e-03 109 2194e-03 109
120 4243e-04 101 4243e-04 101 1030e-03 106 1030e-03 106
140 2104e-04 101 2104e-04 101 4934e-04 104 4934e-04 104
180 1046e-04 1046e-04 2392e-04 2392e-04
Table 6 Errors convergence orders in space and CPU time for fractional Fisher equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 8196e-04 196 8196e-04 196 6045e-04 197 6045e-04 197
1160 2112e-04 201 2112e-04 201 1547e-04 201 1547e-04 201
1320 5246e-05 200 5246e-05 200 3842e-05 200 3842e-05 200
1640 1316e-05 1316e-05 9649e-06 9649e-06
CPU(s) 3880 107191 4022 85058
Example 42 We consider the fractional Huxley equation (ie f (u) =minus01u(1minusu)(uminus0001) in (41)) with the double Gaussian initial value see Eq (412) Again Tables 7 and8 present the numerical results for α=02 05 which show that our fast scheme (49)-(411)has the same convergence order O(h2+∆t) in Linfin norm as the direct scheme (46)-(48)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
670 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
4 42 44 46 480
05
1
15
2
25
3
Log10
(NT)
Log 10
(cpu
time)
Direct EvaluationFast Evaluation
Figure 3 The log-log (in base 10) plot of the CPU time (in seconds) versus the total number of time steps NTfor two schemes Here NS = 80 and α= 05 The left panel shows the results for the Fisher equation and theright panel shows the results for the Huxley equation
Figure 4 Example 41 (Fisher equation) The evolution of the difference (ie Error= FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
but takes much less computational time in Fig 3 The dependence of ε is plotted in Fig 5to show the difference of the numerical solutions to both schemes Again the abovenumerical tests demonstrate our theoretical analysis
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 671
Table 7 Errors and convergence orders in time for fractional Huxley equation with fixed spatial step size h=2minus10
∆t
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rt eninfin rt eninfin rt eninfin rt
110 1089e-03 104 1089e-03 104 2961e-03 106 2961e-03 106
120 5304e-04 102 5304-04 102 1418e-03 104 1418e-03 104
140 2614e-04 101 2614e-04 101 6896e-04 103 6896e-04 103
180 1294e-04 1294e-04 3379e-04 3379e-04
Table 8 Errors convergence orders in space and CPU time for fractional Huxley equation by fixing ∆t=2minus14
h
α=02 α=05
Fast scheme Direct scheme Fast scheme Direct scheme
eninfin rs eninfin rs eninfin rs eninfin rs
180 9051e-04 196 9051e-04 196 6882e-04 198 6882e-04 198
1160 2319e-04 201 2319e-04 201 1746e-04 201 1746e-04 201
1320 5762e-05 200 5762e-05 200 4344e-05 200 4344e-05 200
1640 1447e-05 1447e-05 1091e-05 1091e-05
CPU(s) 3869 110658 4014 85587
Figure 5 Example 42 (Huxley equation) The evolution of the difference (ie Error=FUnh minusDUn
h ) between the
solution to the scheme (49)-(411) (denote by DUnh ) and the solution to the scheme (46)-(48) (denote by FUn
h )
for ε= 1e-71e-81e-91e-10 respectively In the calculation we take T=5[xl xr]= [minus55] M=160N=2000and α=05
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
672 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
5 Conclusions
We have presented an efficient algorithm for solving fractional partial differential equa-tions The algorithm combines the fast evaluation scheme for Caputo fractional deriva-tive C
0 Dαt f (t) and standard finite difference scheme We have proved a prior estimate and
error bound about the solution of our new FD scheme which leads to the stability of thenew scheme The numerical results on linear and nonlinear fraction PDEs show that ournew scheme has the same order of convergence as the existing standard FD schemes butwith nearly optimal complexity in CPU time and storage
Our work can be extended along several directions First it is straightforward to de-sign high order schemes for the evaluation of fractional derivatives Second one maydevelop fast high-order algorithms for solving fractional PDEs which contains fractionalderivatives in both time and space when the current scheme is combined with other ex-isting schemes [11ndash13 25] Third the scheme can be easily incorporated with other dis-cretization schemes for the spatial variable such as finite element methods Fourth effi-cient and stable artificial boundary conditions can be designed using similar techniquesin [24] for solving fractional PDEs in high dimensions These issues are currently underinvestigation and the results will be reported on a later date
Acknowledgments
Shidong Jiang is supported by NSF under grant DMS-1418918 Jiwei Zhang is partiallysupported by NSFC under grants 91430216 and U1530401 Zhimin Zhang is partiallysupported by NSFC grants 11471031 91430216 U1530401 and by NSF under grant DMS-1419040 The authors would like to thank the referees for helpful comments
A The proof of Lemma 27
Proof We rewrite the definition of (227) by
FC0 D
αt un=
u(tn)minusu(tnminus1)
(1minusα)∆tαnΓ(1minusα)
+1
Γ(1minusα)
[
u(tnminus1)
∆tαn
minus u(t0)
tαn
minusαNexp
sumi=1
ωiUhisti(tn)
]
=∆tminusα
n
Γ(1minusα)
(
un
1minusαminus(
α
1minusα+a0)u
nminus1minusnminus2
suml=1
(anminuslminus1+bnminuslminus2)ulminus(bnminus2+
1
nα)u0
)
(A1)
where
an =α∆tαNexp
sumj=1
ωjeminusns j∆tλ1
j bn =α∆tαNexp
sumj=1
ωjeminusns j∆tλ2
j
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 673
λ1j =
eminuss j∆t
s2j ∆t
(eminuss j∆tminus1+sj∆t) λ2j =
eminuss j∆t
s2j ∆t
(1minuseminuss j∆tminuseminuss j∆tsj∆t)
Applying the definition (A1) and the Cauchy-Schwarz inequality we have
(FC0 D
αt gk)gk =
1
∆tαΓ(1minusα)
(
1
1minusα(gk)2minus
( α
1minusα+a0
)
gkminus1gk
minuskminus2
suml=1
(akminuslminus1+bkminuslminus2)gl gkminus(
bkminus2+1
kα
)
g0gk
)
ge 1
∆tαΓ(1minusα)
[
(
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minus1
2
( α
1minusα+a0
)
(gkminus1)2minus 1
2
kminus2
suml=1
(akminuslminus1+bkminuslminus2)(gl)2
minus1
2
(
bkminus2+1
kα
)
(g0)2
)]
(A2)
Summing the above inequality from k=1 to n we obtain
∆tn
sumk=1
(FC0 D
αt gk)gk ge ∆t1minusα
Γ(1minusα)
[
n
sumk=2
(
1
1minusαminus α
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα
)
(gk)2
minusn
sumk=2
(
α
2(1minusα)+
a0
2
)
(gkminus1)2minus 1
2
n
sumk=2
kminus2
suml=1
(
akminuslminus1+bkminuslminus2
)
(gl)2
+1
2(1minusα)(g1)2minus 1
2(1minusα)(g0)2minus 1
2
n
sumk=2
(
bkminus2+1
kα
)
(g0)2
]
=∆t1minusα
Γ(1minusα)
n
sumk=1
(
Ck(gk)2minusC0(g0)2
)
(A3)
where the coefficients Ck (k=01middotmiddotmiddot n) are given by the formula
Cnk =
1
2(1minusα)+
1
2
n
sumk=2
(
bkminus2+1
kα
)
k=0
1
2minus 1
2
nminus2
suml=0
(al+bl)+1
2bnminus2 k=1
1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1 2le kltn
2minusα
2(1minusα)minus 1
2
kminus2
suml=0
(al+bl)minus1
2kα k=n
(A4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
674 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
From (24) we have the estimate
1
t1+αminusεle
Nexp
sumj=1
ωjeminuss jtle 1
t1+α+ε (A5)
It is also straightforward to verify that
kminus2
suml=0
(al+bl)=α∆tαint k∆t
∆t
Nexp
sumj=1
ωjeminuss jtdt (A6)
Combining (A5) and (A6) we obtain
(
1minus 1
kα
)
minusα∆tαtkminus1εlekminus2
suml=0
(al+bl)le(
1minus 1
kα
)
+α∆tαtkminus1ε (A7)
Substituting (A7) into (A4) yields the following estimates
Cn0 le
n1minusα
2(1minusα)+
α∆tαtnminus1ε
2
Cn1 ge
1
2minus 1
2
nminus2
suml=0
(al+bl)ge1
2nαminusα∆tαtnminus1ε
Cnk =1minus 1
2
kminus2
suml=0
(al+bl)minus1
2kαminus 1
2
nminuskminus1
suml=0
(al+bl)+1
2bnminuskminus1
ge 12nα minusα∆tαtnminus1ε 2le klenminus1
Cnn ge
2minusα
2(1minusα)minus
nminus2
suml=0
(al+bl)minus1
2nαge 1
2nαminusα∆tαtnminus1ε
(A8)
Combining (A3) and (A8) we obtain Lemma 27
B The proof of Theorem 33
Proof We observe that the error eki satisfies the following FD scheme
FC0 D
αt ek
i =δ2xek
i +Tki 1le ileNSminus1 1le kleNT (B1)
FC0 D
αt ek
0=2
h
[
δxek12minusFC
0 Dα2t ek
0
]
+Tk0 (B2)
FC0 D
αt ek
NS=
2
h
[
minusδxenNSminus 1
2minusFC
0 Dα2t ek
NS
]
+TkNS
(B3)
e0i =0 0le ileNS (B4)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 675
where the truncation terms Tk at the interior and boundary points are given by the for-mulas
Tki =minus
[
C0 Dα
t u(xitk)minusFC0 D
αt Uk
i
]
+[
uxx(xitk)minusδ2xUk
i
]
1le ileNSminus1 1le kleNT
Tk0 =
uxx(x0tk)minus2
h
[
δxUk12minusux(x0tk)
]
minus 2
h
[
C0 D
α2t u(x0tk)minusFC
0 D
α2t Uk
0
]
minus[
C0 Dα
t u(x0tk)minusFC0 D
αt Uk
0
]
1le kleNT
TkNS
=
uxx(xNStk)minus
2
h
[
ux(xNStk)minusδxUk
NSminus 12
]
minus 2
h
[
C0 D
α2t u(xNS
tk)minusFC0 D
α2t Uk
NS
]
minus[
C0 Dα
t u(xNStk)minusFC
0 Dαt Uk
NS
]
1le kleNT
Using Lemma 31 and Taylor expansion it is easy to show that the truncation terms Tk
satisfy the following error bounds
∣
∣
∣Tk
i
∣
∣
∣le c1(∆t2minusα+h2+ε) 1le ileNSminus1 1le kleNT (B5)
∣
∣
∣Tk
0
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B6)
∣
∣
∣Tk
NS
∣
∣
∣le c1(∆t2minusα+h+
∆t2minusα2
h+
ε
h) 1le kleNT (B7)
with c1 some positive constant Thus for hle1 and ∆tle1 we have
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
le c21
2ν
(
h∆t2minusα+∆t2minus α2 +h2+ε
)2+
2c21L
micro
(
∆t2minusα+h2+ε)2
le2c21
ν
(
∆t2minusα+h2+ε)2
+2c2
1L
micro
(
∆t2minusα+h2+ε)2
le4c21
(1
ν+
L
micro
)
(∆t2minusα+h2)+4c21
(1
ν+
L
micro
)
ε2 (B8)
A direct application of Theorem 31 to the system (B1)-(B4) produces
∆tn
sumk=1
ek2infin le ∆t(1+
radic
1+L2micro)
Lmicro
n
sumk=1
(
1
4ν
[
(hTk0 )
2+(hTkNS)2]
+2
microh
NSminus1
sumi=1
(Tki )
2
)
(B9)
Substituting (B8) into (B9) simplifying the resulting expressions and taking the squareroot for both sides we obtain (327)
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
676 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
References
[1] B Alpert L Greengard and T Hagstrom Rapid evaluation of nonreflecting boundary ker-nels for time-domain wave propagation SIAM J Numer Anal 37 (2000) 1138ndash1164
[2] A A Awotunde R A Ghanam and N Tatar Artificial boundary condition for a modifiedfractional diffusion problem Bound Value Probl 1 (2015) 1ndash17
[3] D Baffet and Jan S Hesthaven A Laplace transform based kernel reduction scheme forfractional differential equations httpsinfoscienceepflchrecord212955 2016
[4] H Brunner H Han and D Yin Artificial boundary conditions and finite difference approx-imations for a time-fractional diffusion-wave equation on a two-dimensional unboundedspatial domain J Comput Phys 276 (2014) 541ndash562
[5] G Beylkin and L Monzon On approximation of functions by exponential sums Appl Com-put Harmon Anal 19 (2005) 17ndash48
[6] G Beylkin and L Monzon Approximation by exponential sums revisited Appl ComputHarmon Anal 28 (2010) 131ndash149
[7] H Brunner H Han and D Yin The maximum principle for time-fractional diffusion equa-tions and its application Numer Funct Anal Optim 36 (2015) 1307ndash1321
[8] H Brunner X Wu and J Zhang Computational Solution of Blow-Up Problems for Semilin-ear Parabolic PDEs on Unbounded Domains SIAM J Sci Comput 31(2010) 4478ndash4496
[9] JX Cao CP Li and YQ Chen High-order approximation to Caputo derivatives andCaputo-type advection-diffusion equations (II) Frac Calc and Appl Anal 183 (2015) 735ndash761
[10] J Cao and C Xu A high order scheme for the numerical solution of the fractional ordinarydifferential equations J Comput Phys 238 (2013) 154ndash168
[11] C Chen F Liu V Anh and I Turner Numerical methods for solving a two-dimensionalvariable-order anomalous subdiffusion equation Math Comp 81 (2012) 345ndash366
[12] M Cui Compact finite difference method for the fractional diffusion equation J ComputPhys 228 (2009) 7792ndash7804
[13] M Cui Compact alternating direction implicit method for two-dimensional time fractionaldiffusion equation J Comput Phys 231 (2012) 2621ndash2633
[14] J R Dea Absorbing boundary conditions for the fractional wave equation Appl MathComput 219 (2013) 9810ndash9820
[15] G H Gao Z Z Sun and Y N Zhang A finite difference scheme for fractional sub-diffusionequations on an unbounded domain using artificial boundary conditions J Comput Phys231 (2012) 2865ndash2879
[16] G H Gao and Z Z Sun The finite difference approximation for a class of fractional sub-diffusion equations on a space unbounded domain J Comput Phys 236 (2013) 443ndash460
[17] G H Gao and Z Z Sun A new fractional numerical differentiation formula to approximatethe Caputo fractional derivative and its applications J Comput Phys 259 (2014) 33ndash50
[18] R Ghaffari and S M Hosseini Obtaining artificial boundary conditions for fractional sub-diffusion equation on space two-dimensional unbounded domains Comput Math Appl68 (2014) 13ndash26
[19] B L Guo Q Xu and A L Zhu A second-order finite difference method for two-dimensionalfractional percolation equations Commun Comput Phys 19 (2016) 733-757
[20] L Greengard and P Lin Spectral approximation of the free-space heat kernel Appl Com-put Harmon Anal 9 (2000) 83ndash97
[21] L Greengard and J Strain A fast algorithm for the evaluation of heat potentials Comm
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
S Jiang et al Commun Comput Phys 21 (2017) pp 650-678 677
Pure Appl Math 43 (1990) 949ndash963[22] H Han and X Wu Artificial Boundary Method Tsinghua Univ Press 2013[23] X Huang and D Yin Artificial boundary conditions and finite difference approximations for
a time-fractional diffusion-wave equation on high dimensional unbounded spatial domainJ Comput Phys to appear
[24] S Jiang and L Greengard Efficient representation of nonreflecting boundary conditions forthe time-dependent Schrodinger equation in two dimensions Comm Pure Appl Math 61(2008) 261ndash288
[25] S Jiang L Greengard and W Bao Fast and accurate evaluation of nonlocal Coulomb anddipole-dipole interactions via the nonuniform FFT SIAM J Sci Comput 365 (2014) B777-B794
[26] S Jiang L Greengard and S Wang Efficient sum-of-exponentials approximations for theheat kernel and their applications Adv Comput Math 413 (2015) 529ndash551
[27] R Ke M Ng and H Sun A fast direct method for block triangular Toeplitz-like with tridi-agonal block systems from time-fractional partial differential equations J Comput Phys303 (2015) 203-211
[28] A Kilbas H Srivastava and J Trujillo Theory and Applications of Fractional DifferentialEquations Elesvier Science and Technology Boston 2006
[29] T Langlands and B Henry Fractional chemotaxis diffusion equations Phys Rev E 81(2010) 051102
[30] T Langlands and B Henry The accuracy and stability of an implicit solution method for thefractional diffusion equation J Comput Phys 205 (2005) 719ndash736
[31] S Lei and H Sun A circulant preconditioner for fractional diffusion equations J ComputPhys 242 (2013) 715ndash725
[32] C Li W Deng and Y Wu Numerical analysis and physical simulations for the time frac-tional radial diffusion equation Comput Math Appl 62 (2011) 1024ndash1037
[33] C Li A Chen and J Ye Numerical approaches to fractional calculus and fractional ordinarydifferential equation J Comput Phys 230 (2011) 3352ndash3368
[34] C Li R Wu and H Ding High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations Commun Appl Indu Math 62 e-536 (2014) 1-32
[35] D Li and J Zhang Efficient implementation to numerically solve the nonlinear time frac-tional parabolic problems on unbounded spatial domain J Comput Phys 322 (2016) 415-428
[36] D Li C Zhang M Ran A linear finite difference scheme for generalized time fractionalBurgers equation Appl Math Model 40 (2016) 6069-6081
[37] H Li J Cao C Li High-order approximation to Caputo derivatives and Caputo-typeadvection-diffusion equations (III) J Comput Appl Math 299 (2016) 159-175
[38] J Li A fast time stepping method for evaluating fractional integrals SIAM J Sci Comput31 (2010) 4696ndash4714
[39] LM Li D Xu Alternating direction implicit Galerkin finite element method for the two-dimensional time fractional evolution equation Numer Math Theor Meth Appl 7 (2014)41-57
[40] Y Lin and C Xu Finite differencespectral approximations for the time-fractional diffusionequation J Comput Phys 225 (2007) 1533ndash1552
[41] Y Lin X Li and C Xu Finite differencespectral approximations for the fractional cableequation Math Comp 80 (2011) 1369ndash1396
[42] Y Lin and C Xu Finite differencespectral approximations for time-fractional diffusion
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at
678 S Jiang et al Commun Comput Phys 21 (2017) pp 650-678
equation J Comput Phys 225 (2007) 1533ndash1552[43] X Lu H Sun and H Pang Fast approximate inversion of a block triangular Toeplitz matrix
with applications to fractional subdiffusion equations J Numer Lin Alg Appl 22 (2015)866-882
[44] C Lubich and A Schadle Fast convolution for nonreflecting boundary conditions SIAM JSci Comput 24 (2002) 161ndash182
[45] W McLean Fast summation by interval clustering for an evolution equation with memorySIAM J Numer Anal 346 (2012) 3039ndash3056
[46] R Metzler and J Klafter The random walks guide to anomalous diffusion a fractionaldynamics approach Phys Rep 339 (2000)
[47] Z Odibat Approximations of fractional integrals and Caputo fractional derivatives ApplMath Comput 178 (2006) 527ndash533
[48] K B Oldham and J Spanier The Fractional Calculus Academic Press New York 1974[49] F W J Olver D W Lozier R F Boisvert and C W Clark editors NIST Handbook of
Mathematical Functions Cambridge University Press New York NY 2010[50] H Pang and H Sun Multigrid method for fractional diffusion equations J Comput Phys
231 (2012) 693ndash703[51] I Podlubny Fractional Differential Equations Academic Press San Diego 1999[52] Z Sun and X Wu A fully discrete difference scheme for a diffusion-wave system Appl
Numer Math 56 (2006) 193ndash209[53] H Wang K Wang and T Sircar A direct O(Nlog2N) finite difference method for fractional
diffusion equations J Comput Phys 229 (2010) 8095ndash8104[54] H Wang and S B Treena A fast finite difference method for two-dimensional space-
fractional diffusion equations SIAM J Sci Comput 34 (2012) 2444ndash2458[55] K Xu and S Jiang A bootstrap method for sum-of-poles approximations J Sci Comput
55 (2013) 16ndash39[56] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Phys Rev E 78 (2008) 026709[57] J Zhang Z Xu X Wu Unified approach to split absorbing boundary conditions for nonlin-
ear Schrodinger equations Two dimensional case Phys Rev E 79 (2009) 046711[58] Q Zhang J Zhang S Jiang Z Zhang Numerical solution to a linearized time fractional
KdV equation on unbounded domains accepted by Math Comput 2016[59] C Zheng Approximation stability and fast evaluation of exact artificial boundary condition
for one-dimensional heat equation J Comput Math 25 (2007) 730ndash745
httpswwwcambridgeorgcoreterms httpsdoiorg104208cicpOA-2016-0136Downloaded from httpswwwcambridgeorgcore New Jersey Institute of Technology on 16 Mar 2017 at 130234 subject to the Cambridge Core terms of use available at