Nguyen & Jang (2017)

RESEARCH PAPER
A HIGH-ORDER PREDICTOR-CORRECTOR METHOD

FOR SOLVING NONLINEAR DIFFERENTIAL
EQUATIONS OF FRACTIONAL ORDER
Thien Binh Nguyen 1,2 , Bongsoo Jang 2
Abstract
An accurate and ecient new class of predictor-corrector schemes are

proposed for solving nonlinear dierential equations of fractional order. By
introducing a new prediction method which is explicit and of the same
accuracy order as that of the correction stage, the new schemes achieve a
uniform accuracy order regardless of the values of fractional order . In
cases of 0 < 1, the new schemes signicantly improve the numerical
accuracy comparing with other predictor-corrector methods whose accuracy
depends on . Furthermore, by computing the memory term just once for
both the prediction and correction stages, the new schemes reduce the
computational cost of the so-called memory eect, which make numerical
schemes for fractional dierential equations expensive in general. Both
2nd-order scheme with linear interpolation and the high-order 3rd-order
one with quadratic interpolation are developed and show their advantages
over other comparing schemes via various numerical tests.
MSC 2010 : Primary 26A33; Secondary 34A08, 34K28, 34K37, 35R11
Key Words and Phrases: Caputo fractional derivative, fractional dier-
ential equations, predictor-corrector methods, explicit schemes, linear and
quadratic interpolation

c 2017 Diogenes Co., Soa
pp. 447476 , DOI: 10.1515/fca-2017-0023
Brought to you by | University of Toronto-Ocul

Authenticated
Download Date | 6/4/17 1:02 PM
448 T.B. Nguyen, B. Jang
1. Introduction
Recently, fractional dierential equations have drawn much attention to
researchers, thanks to their various applications in physics and engineering.
For example, fractional dierential equations can be used as mathematical
models for problems in anomalous diusion transport ([14]), viscoelastic
materials [20]), porous media ([1]) to name but a few. Other applications
in chemistry, chaos, nance, signal processing, etc. are discussed in the
books, e.g., [18], [20], [21], or [22], and the references therein.
In this work, we consider the following ordinary dierential equation
(ODE) with fractional order R+ ,

D0 y(t) = f (t, y(t)), t [0, T ],
(1.1)
y (k) (0) = yk , k = 0, 1, . . . , m 1,
where m 1 < m Z+ , and the fractional derivative is dened in the
Caputo sense ([2]), i.e.,
t
1
D0 y(t) = (t )m1 y (m) ( )d. (1.2)
(m ) 0
Here, () is the conventional Gamma-function.
We note here that other than the Caputo denition (1.2), there ex-
ist dierent approaches to dene fractional derivatives, for example, the
Riemann-Liouville approach, the Grunwald-Letnikov denition. It is well-
known that these approaches, under appropriate assumptions, are all re-
lated one another. We refer to [8], [21], or [18] for detailed analysis and
discussions. In this work, we have chosen the Caputo derivative because it
allows us to readily impose the initial problem (1.1) with inhomogeneous
conditions.
It is well-known that a continuous function y(t) is the solution of Eq.
(1.1) if and only it is the solution of the Volterra integral equation ([8], [21])
t
1
y(t) = g(t) + (t )1 f (, y( ))d, (1.3)
() 0
tk
where g(t) = m k=0 yk k! is related to the initial condition data. Theoretical
analysis and numerical approaches for the Volterra integral equations can
be found in [4], and [5], etc.
In terms of numerical approaches, the main diculty one has to tackle
is the non-locality property of the solution y(t) due to the kernel (t )1
under the integral sign on the right-hand side of Eq. (1.3). In order to
illustrate this, let us rst discretize the grid to be
N := {tj : 0 = t0 < t1 < . . . < tj < . . . < tn < tn+1 < . . . < tN = T }.
(1.4)

Authenticated
A HIGH-ORDER PREDICTOR-CORRECTOR METHOD . . . 449
For simplicity, we assume that the grid is uniform, i.e., h = tj+1 tj ,

j = 0, . . . , N 1. Eq. (1.3) can be rewritten at time tn+1 as follows ([5]),
y(tn+1 ) = g(tn+1 ) + y (tn+1 ) + Y (tn+1 ), (1.5)
where
tn
1
y (tn+1 ) = (tn+1 )1 f (, y( ))d, (1.6)
() 0
is called the lag term which takes y(t) from the initial time up to t = tn into
consideration in the computation of y(tn+1 ). This is the so-called memory
eect of the solution (1.3). It is noted that this memory eect is unique
for the dierential equations with fractional order, and is preferable for
engineering applications since it indicates the dependence of y(tn+1 ) on the
solution at all prior time points.
The increment term
tn+1
1
Y (tn+1 ) = (tn+1 )1 f (, y( ))d (1.7)
() tn
determines the advancing of the solution from tn to tn+1 .
For numerical approximations of y(t), a typical approach is to inter-
polate the right-hand side f (t, y(t)), given sucient continuity, over each
interval Ij = [tj , tj+1 ], j = 0, ..., N 1, by some types of interpolating
polynomials. For this purpose, the Lagrange polynomials are often taken
into consideration. For example, the celebrated Adams-Bashforth-Moulton
(ABM) by Diethelm et al. ([9], [10]) uses linear interpolation of the right-
hand side f (t, y(t)), or the recently developed high-order schemes ([19],
[25]) in which f (t) is interpolated by a quadratic interpolating polynomial.
For all these schemes, a predicted value yn+1 P of y(tn+1 ) has to be sought
for the approximation of the increment term Y (tn+1 ). For this purpose,
a lower-order interpolation of f (t) is facilitated in order to obtain yn+1 P .
Then the approximation yn+1 y(tn+1 ), which is called a correction, can

be carried out. We observe here that since yn+1 P and yn+1 are obtained at
dierent accuracy, more precisely, the former is of lower order than the lat-
ter, the lag term (1.6) has to be approximated separately in each stage. In
other words, the memory takes eects in both the prediction and correction
stages. This makes the schemes expensive. It is even more costly for high-
order schemes in which more accurate predicted yn+1 P or the approximation
of the solution at half-points tj+ 1 s are required ([19]).
2
It is worth mentioning here other numerical approaches to the dieren-
tial equation with fractional order (1.1). These include the implicit block-
by-block schemes by Kumar and Agrawal ([17]) and later improved by Cao
and Xu ([6]), the L1 2 scheme ([15]) in which the discretization is carried
out on Eq. (1.1) directly instead of on the Volterra integral solution (1.3),

Authenticated
or the Grunwald-based method ([16]). Fractional numerical methods de-

signed for specically PDE-modeled applications can be found in [12], [24].
Finite element, spectral or collocation methods are developed in the works,
e.g., [3], [11], [13] and the references therein.
Another issue one faces when developing a prediction-correction scheme
for the dierential equations with fractional order is the degeneration of the
accuracy order in comparison with that of a scheme for classical ODEs, that
is, when Z+ . We take the ABM method for instance. It is shown in [10]
that this scheme is of p-order where p = min(1 + , 2). Thus the scheme is
less accurate than the two-step Adams-Bashforth method which is of order
2, in case of 0 < < 1. It is indicated in the below section that this loss of
accuracy is due to the usage of a low-order prediction for the interpolation
of the increment term.
In this work, we aim to develop a new class of predictor-corrector
schemes which overcomes the drawbacks aforementioned. The novelty of
our proposed schemes are two-fold. Firstly, the new schemes achieve a uni-
form accuracy of order regardless of the values of . This can be achieved
by a new prediction method which is explicit and of the same accuracy
with that of the correction stage. Secondly, the new designed schemes are
more ecient than other corresponding prediction-correction ones in terms
of computational costs. By a reduction in the approximation of the lag
term y (tn+1 ) by applying this for both prediction and correction, we will
show that the new scheme with linear interpolation is only about half the
cost of that of the ABM. We also discuss schemes of higher order with a
quadratic interpolation rule.
We organize our paper as follows. In Section 2, we propose a new
ecient scheme with linear interpolation. An error analysis then follows
to prove that our new scheme has a uniform order 2 for both cases of
0 < < 1 and 1. A detailed check of computational costs show that
our new scheme is about of half of the cost of that of the corresponding
ABM method. In Section 3, we further improve the scheme to uniformly 3rd
order by employing a quadratic interpolation of f (t, y(t)) over each interval
Ij . Case by case numerical tests are presented in Section 4 to illustrate for
our analysis in the previous sections. Finally, we draw our conclusions in
the last Section 5.
2. Second-order Predictor-Corrector Scheme

with Linear Interpolation
2.1. Description of the Second-order Predictor-Corrector Scheme
with Linear Interpolation. Before proceeding, we rst dene some no-
tations. We denote yj = y(tj ) the restriction of the exact solution at time

Authenticated
P , y
tj , j = 0, . . . , N . Let yn+1 n+1 yn+1 be the predicted and corrected, re-
spectively, approximations of yn+1 . Similarly, we also denote fj = f (tj , yj ),
fj = f (tj , yj ), and fjP = f (tj , yjP ). The lag and increment terms y (tn+1 ),
Y (tn+1 ) are approximated by yn+1 and Yn+1 , respectively.
For clarication, we re-derive the ABM ([9]) as follows. The exact
solution (1.3) is written as
n1
1 tj+1
y(tn+1 ) = g(tn+1 ) + (tn+1 )1 f (, y( ))d (2.1)
()
j=0 tj
tn+1
1
+ (tn+1 )1 f (, y( ))d.
() tn
On each interval Ij = [tj , tj+1 ], we interpolate f (t) by a linear Lagrange
polynomial
f (t) fj Lj (t) + fj+1 Lj+1 (t), t Ij , (2.2)
where
t tl
Lk (t) = , (k, l) = (j, j + 1), (j + 1, j).
tk tl
Substituting the approximation (2.2) into Eq. (2.1) with taking into
accounts the approximated values yj and that the grid is uniform, we obtain
that
n1
1 0,j 1,j 1
yn+1 = gn+1 + [Bn+1 fj + Bn+1 fj+1 ] + [B 0,n fn + Bn+1
1,n P
fn+1 ],
() () n+1
j=0
(2.3)
where the coecients are
tj+1 tj+1
0,j 1,j
Bn+1 = (tn+1 )1 Lj ( )d, Bn+1 = (tn+1 )1 Lj+1 ( )d,
tj tj
(2.4)
which can be evaluated explicitly.
Comparing Eq. (2.3) with Eq. (1.5), we notice that the second term on
the right-hand side of (2.3) is the approximation of the lag term y (tn+1 ),
and the last term approximates the increment Y (tn+1 ). Expanding the
summation, and collecting fj , one can easily obtain the formula for aj,n+1
given in [9].
P , we invoke a lower-order ap-
In order to compute the predicted yn+1
proximation of f (t), that is, a constant interpolation over each interval
t Ij ,
f (t) f (tj )Ij (t),
where the indicator function

Authenticated

1, if t Ij ,
Ij (t) =
0, otherwise,
and substituting into Eq. (2.1), we obtain that
n
P 1 j
yn+1 = gn+1 + Cn+1 fj , (2.5)
()
j=0
where
tj+1
j h
Cn+1 = (tn+1 )1 d = [(n + 1 j) (n j) ],
tj
h
which is the coecient bj,n+1 in [9] multiplying with .

Remark 2.1.
For the ABM method, the lag term y (tn+1 ) needs re-computing
P , i.e., the
twice for the corrected yn+1 and the predicted yn+1
terms in Eqs. (2.3) and (2.5), which is costly. Indeed, one can

reduce the cost by reducing the computation of yn+1 just once for
both prediction and correction steps (see the below section).
It is shown in [9], [10] that the order of the ABM scheme is
max |yn+1 yn+1 | = O(hp ),
0jN 1
where p = min(1 + , 2).
One can check that the order reduction for cases of 0 < < 1
P .
comes from the implementation of the low-order prediction yn+1
In the below section, we propose a new scheme which overcomes
this shortage, and obtain a uniform order O(h2 ) for any .
We now proceed to develop our new scheme. For this purpose, we need
the following lemma.
Lemma 2.1. Assume that (t) P1 [0, T ] where P1 is the space of all
polynomials of degree less than or equal to one. Let n , n = 0, . . . , N be
the restricted value of (t) on grid N dened in Eq. (1.4). Then, there
exist coecients b0n+1 and b1n+1 such that the following equality
tn+1
h
(tn+1 )1 ( )d = (b0 n1 + b1n+1 n ), (2.6)
tn ( + 1) n+1
is exact. More precisely,
b0n+1 = 1, b1n+1 = + 2. (2.7)

Authenticated
P r o o f. Since P1 [0, T ], there exist 0 , 1 R such that

(t) = 0 + 1 t. (2.8)
Substituting Eq. (2.8) into Eq. (2.6) and evaluating the integral on the
left-hand side, we obtain that

h 1 0
0 + 1 + tn = bn+1 (0 + 1 tn1 ) + b1n+1 (0 + 1 tn ) .
+1 ( + 1)
Equating the above terms by terms with respect to k , k = 0, 1, one
obtains a non-singular linear system with unknowns b0n+1 and b1n+1 whose
values are given in Eq. (2.7). This proves the lemma. 2
We then propose our new scheme with linear interpolation as follows.

We write our prediction and correction stages of the new scheme in terms
of the lag and increment term (1.6), (1.7) as below:

yn+1 = gn+1 + yn+1 + Yn+1 , (2.9)
where
n1
1 0,j 1,j
yn+1 = [Bn+1 fj + Bn+1 fj+1 ],
()
j=0
1
Yn+1 = [B 0,n fn + Bn+1
1,n P
fn+1 ],
() n+1
where the prediction is as follows,
P h
yn+1 = gn+1 + yn+1 + [b0 fn1 + b1n+1 fn ], (2.10)
( + 2) n+1
with b0n+1 and b1n+1 given in Eq. (2.7).
Remark 2.2. We notice that the prediction and correction of yn+1
are dierent from each other only by the approximations of the increment
Y (tn+1 ), whereas the lag term y (tn+1 ) is computed only one time which
is similar to [7]. This thus reduces the overall cost of the new scheme.
(Comparing Eqs. (2.9), (2.10) with Eqs. (2.3), (2.5) of the ABM scheme.)
We now show that the new scheme achieves a uniform accuracy order
regardless of .
2.2. Error Analysis.
2.2.1. Truncation and Global Error Analysis for Predictor. From

herein, we denote by C a generic constant which is independent of all grid
parameters and may change case by case.
We need the following lemmas.

Authenticated
Lemma 2.2. (Interpolation Errors) Let f C n+1 [a, b] and pn Pn [a, b]

interpolate the function f at the grid n in (1.4) with a = t0 and b = tn ,
then there exists (a, b) such that, for any t [a, b],
n
f (n+1) ()
f (t) pn (t) = (t tj ).
(n + 1)!
j=0
Lemma 2.3. For R+ ,

tj+1
n1
T
(tn+1 )1 d .
tj
j=0
P r o o f.
tj+1
n1 n
tj+1
1
(tn+1 ) d (tn+1 )1 d
j=0 tj j=0 tj
tn+1 tn+1 T
= (tn+1 )1 d = .
0
2
Lemma 2.4. (Discrete Gronwalls Inequality, [6]) Let {an }N N

n=0 , {bn }n=0
be non-negative sequences with bn s monotonic increasing, and satisfy that
n1

an bn + M h (n j)1 aj , 0 n N,
j=0
where M > 0 is bounded and independent of h, and 0 < 1. Then,
an bn E (M ()(nh) ).
In the above, E () denotes the (one-parameter) Mittag-Leer function,

which for = 1 becomes the exponential one, for details see e.g. [8], [18],
[20], [21].
Theorem 2.1. (Truncation Error) Suppose that f (, u()) C 2 [0, T ].
Let the truncation error at time step tn+1 be
tn+1
1
P
rn+1 = (tn+1 )1 f (, y( ))d (2.11)
() 0

1 0,j
n1
h 0
1,j
Bn+1 fj + Bn+1 fj+1 bn+1 fn1 + b1n+1 fn .
() ( + 2)
j=0
Then, there exists a constant C independent of all grid parameters such
that

Authenticated
P
rn+1 Ch2 . (2.12)
P r o o f. In view of the interpolation (2.2), Eq. (2.11) is written as

P
rn+1

n1

1 tj+1

(tn+1 )1 {f (, y( )) [fj Lj ( ) + fj+1 Lj+1 ( )]} d
() tj
j=0
tn+1
1 h 0
+ (tn+1 )1 f (, y( ))d bn+1 fn1 + b1n+1 fn
() tn ( + 2)
= I + II.
Since f (, u()) C 2 [0, T ], by Lemma 2.2,

f (t, y(t)) [fj Lj (t) + fj+1 Lj+1 (t)] 1 |f (j )|h2 ,
2
for some j (tj , tj+1 ). Let M := max0jN |f (j )|, by Lemma 2.3, we
have that
n1
M h2 tj+1 MT
I (tn+1 )1 d h2 .
2() tj 2( + 1)
j=0
By the Taylor expansion of f (, u()) around tn , we have that

1
f (t) = p1 (t) + f (n )(t tn )2 , n (tn , t),
2
where
p1 (t) = fn + f (tn )(t tn ), thus p1 (t) P1 [0, T ],
then by Lemma 2.1, with a note that p1 (tn )fn = 0 and |p1 (tn1 )fn1 |
(M/2)h2 ,
tn+1
1 1

II = (tn+1 ) 1 2
p1 ( ) + f (n )( tn ) d
() tn 2

h
b0 fn1 + b1n+1 fn
( + 2) n+1

1 tn+1
(tn+1 ) f (n )( tn ) d
1 2
2() tn
h 0 M M
+ bn+1 |p1 (tn1 ) fn1 | h2+ + h2+ .
( + 2) 2( + 1) 2( + 1)
Combining I and II, we conclude that

Authenticated
P MT M
rn+1 h2 + h2+ Ch2 . (2.13)
2( + 1) ( + 1)
2
Lemma 2.5. For k = 0, 1 and j = 0, 1, , n 1,

k,j (n j)1 h , 0 < < 1,
|Bn+1 |
T 1 h, 1.
P r o o f.

0,j 1 tj+1 1

tj+1
|Bn+1 | (tn+1 ) ( tj+1 )d (tn+1 )1 d.
h tj tj
1:
tj+1
(tn+1 )1 d (tn+1 tj )1 h T 1 h.
tj
0 < < 1:
tj+1
(tn+1 )1 d (tn+1 tj+1 )1 h = (n j)1 h .
tj
1,j
We then deduce the result. The proof for Bn+1 follows similarly. 2
For 1, we have
tn+1
h
= (tn+1 )1 d T 1 h,
tn
which gives the following lemma.
Lemma 2.6. For 1,

h
T 1 h.

Theorem 2.2. (Global Error)Suppose that f (, u()) C 2 [0, T ], and

furthermore is Lipschitz continuous in the second argument, i.e.,
|f (t, u1 ) f (t, u2 )| L|u1 u2 |, u1 , u2 R, (2.14)

P
then the global error En+1 P | is
= |yn+1 yn+1

Authenticated
n

P LT 1 h

rn+1 +C Ej , 1,
()
P j=0
En+1 n

Lh

P
rn+1 + C (n + 1 j)1 Ej , 0 < < 1.
()
j=1
P r o o f.

1 tn+1
P
En+1 = (tn+1 )1 f (, y( ))d
() 0
1 0,j
n1 h
1,j
Bn+1 fj + Bn+1 fj+1 b0
fn1 + bn+1 fn
1
() ( + 1)() n+1
j=0
1 0,j
n1
P 1,j
rn+1 + |Bn+1 ||fj fj | + |Bn+1 ||fj+1 fj+1 |
()
j=0
h
+ |b0n+1 ||fn1 fn1 | + |b1n+1 ||fn fn | .
( + 1)()
1: By Lemmas 2.4 and 2.6, and that f (t) is Lipschitz, we deduce
that
n1
P P LT 1 h Lh
En+1 rn+1 + [Ej + Ej+1 ] + (En1 + ( + 2)En )
() ( + 2)
j=0
n
P LT 1 h
rn+1 +C Ej .
()
j=0
0 < < 1: Similarly, we have that
1 0,j
n1
P P 1,j
En+1 rn+1 + |Bn+1 ||fj fj | + |Bn+1 ||fj+1 fj+1 |
()
j=0
h
+ |b0n+1 ||fn1 fn1 | + |b1n+1 ||fn fn | rn+1
P
+ Lh Rn ,
( + 2)
where
n1
1 1 ( + 2)
Rn = (n j)1 (Ej + Ej+1 ) + En1 + En .
() ( + 2) ( + 2)
j=0
Since 1 (n + 1 k)/(n k) 2 for k = 1, , n 1, (n k)1
21 (n + 1 k)1 ,
(n + 1 k)1 + (n k)1 (1 + 21 )(n + 1 k)1 3(n + 1 k)1 .
(2.15)
Then we have

Authenticated
3 1
Rn n E1 + (n 1)1 E2 + + (n (n 1))1 En1
()

1 1 +2
+ En1 + + En
( + 2) () ( + 2)
n
3 1 +2
(n + 1 j)1 Ej + En1 + En
() ( + 2) ( + 2)
j=1
n
C
(n + 1 j)1 Ej .
()
j=1
(2.16)
This completes the proof. 2
2.2.2. Truncation and Global Error Analysis for Corrector. Carry-

ing a similar procedure, we obtain the following results.
Theorem 2.3. (Truncation Error)With the same assumptions as those

of Theorem 2.1, we have that
C Lh En+1
P
rn+1 Ch2 + .
( + 1)
P r o o f. The proof follows that of Theorem 2.1:

n1
1 tj+1
1

C
rn+1 (tn+1 ) f (, y( )) [fj Lj ( ) + fj+1 Lj+1 ( )] d
()
j=0 tj
tn+1
1
1

(tn+1 ) P
+
() tn f (, y( )) fn Ln ( ) + fn+1 Ln+1 ( ) d
n
1 tj+1 1

(tn+1 ) f (, y( )) [fj Lj ( ) + fj+1 Lj+1 ( )] d
() tj
j=0
1
+ |B 1,n ||f P fn+1 |
() n+1 n+1
n
M h2 tj+1 Lh En+1
P
(tn+1 )1 d +
2() tj ( + 1)
j=0
MT Lh En+1
P
h2 + .
2( + 1) ( + 1)
2

Authenticated
For the global error of the corrector it can be easily checked by the
Lemma 2.4 and Theorems 2.2, 2.3.
Theorem 2.4. (Global Error) With the same assumptions as those of

Theorem 2.2, we have
En+1 = |yn+1 yn+1 | Ch2 ,
given that the starting error E1 Ch2 .

P r o o f.
1 0,j
n1
C 1,j
En+1 rn+1
+ |Bn+1 ||fj fj | + |Bn+1 ||fj+1 fj+1 |
()
j=0
1 0,n 1,n P P

+ |Bn+1 ||fn fn | + |Bn+1 ||fn+1 fn+1 | .
()
j,n
Let us note that |Bn+1 | h /, j = 0, 1 by the denition on (2.4).
1: By Theorem 2.2, Lemma 2.5, and that f (t) is Lipschitz, we
deduce that
n1
C LT 1 h Lh P
En+1 rn+1 + (Ej + Ej+1 ) + (En + En+1 )
() ( + 1)
j=0

n n

MT h 2 2Lh
P LT 1 h LT 1 h
+ rn+1 + C Ej + C Ej
2( + 1) ( + 1) () ()
j=0 j=0

n
M T h2 2Lh P LT 1 h 2Lh
+ rn+1 + C 1+ Ej .
2( + 1) ( + 1) () ( + 1)
j=0
By the Discrete Gronwalls Inequality in Lemma 2.4 we have

M T h2 2Lh P LT 2Lh
En+1 + rn+1 exp C 1+ .
2( + 1) ( + 1) () ( + 1)
0 < < 1: Carrying similar procedure for Rn in (2.16) combined

with the Theorem 2.2, Lemma 2.5 we have
n1
C Lh Lh
En+1 rn+1 + (n j)1 (Ej + Ej+1 ) + P
(En + En+1 )
() ( + 1)
j=0

Authenticated

2 n
MT h 2Lh rn+1 Lh
+ P
+C (n + 1 j)1 Ej
2( + 1) ( + 1) ()
j=1
n

Lh
+C (n + 1 j)1 Ej
()
j=1

n
M T h2 2Lh P Lh 2Lh
+ rn+1 + C 1+ (n+1j)1 Ej
2(+1) (+1) () (+1)
j=1
By the Discrete Gronwalls Inequality we have

M T h2 2Lh P 2Lh
En+1 + r E CLT 1+ .
2( + 1) ( + 1) n+1 ( + 1)
2
3. Third-order Predictor-Corrector Scheme

with with Quadratic Interpolation
3.1. Description of Third-order Predictor-Corrector Scheme with
Quadratic Interpolation. In this section, we further improve our scheme
by employing a quadratic interpolation of the right-hand side f (t, y(t)) over
each interval Ij . For this, we observe the following lemma.
Lemma 3.1. Assume that (t) P2 [0, T ] where P2 is the space of all
polynomials of degree less than or equal to two. Let n , n = 0, . . . , N be
the restricted value of (t) on grid N dened in Eq. (1.4). Then, there
exist coecients a0n+1 , a1n+1 , and a2n+1 such that the following equality
tn+1
(tn+1 )1 ( )d = A(a0n+1 n2 + a1n+1 n1 + a2n+1 n ),
tn
h +4 1
is exact, where A = , a0n+1 = , an+1 = 2( + 3) and
( + 1)( + 2) 2
22 + 9 + 12
a2n+1 = .
2
The proof follows that of Lemma 2.1, hence omitted. Over [tj1 , tj+1 ],
j 1, we interpolate f (t) by a quadratic Lagrange polynomial
j+1

f (t) fk Qjk (t), (3.1)
k=j1
where

Authenticated
j+1

t tm
Qjk (t) = ,
tk tm
m=j1
m=k
On [t0 , t1 ], f (t) can be interpolated by f (t) f0 Q00 (t) + f1/2 Q01/2 (t) +
(t t1/2 )(t t1 ) (t t0 )(t t1 )
f1 Q01 (t), where Q00 (t) = , Q0 (t) =
(t0 t1/2 )(t0 t1 ) 1/2 (t1/2 t1 )(t1/2 t1 )
(t t0 )(t t1/2 )
and Q01 (t) = . Substituting Eq. (3.1) into Eq. (2.1), tak-
(t1 t0 )(t1 t1/2 )
ing the uniform property of the grid, we obtain that

yn+1 = gn+1 + yn+1 + Yn+1 . (3.2)
Here, the lag term is approximated as follows,
1 0,0

yn+1 = An+1 f0 + A1,0
n+1
1/2 + A2,0 f1
f n+1
()
1 0,j
n1
+ An+1 fj1 + A1,j 2,j
n+1 fj + An+1 fj+1 ,
()
j=1
where for j = 0, i = i/2 and for 1 j n, i = i + j 1,

tj+1
Ai,j
n+1 = (tn+1 )1 Qji ( )d, i = 0, 1, 2.
tj
The increment term is

1 0,n
Yn+1 = An+1 fn1 + A1,n
n+1
n + A2,n fP
f n+1 n+1 ,
()
P
where the predicted yn+1 is approximated as follows,
2
P h
yn+1 = gn+1 + yn+1 + akn+1 fn+k2 , (3.3)
( + 3)
k=0
where a0n+1 , a1n+1 , a2n+1 can be computed by Lemma 3.1.
3.2. Error Analysis.
3.2.1. Truncation and Global Error Analysis for Predictor. By a

similar procedure, we can obtain the analysis for the accuracy order of the
new scheme using a quadratic interpolation. Let rn+1 P be the truncation
error of prediction at tn+1 dened by

Authenticated
2
P
1

tn+1
1 k,0
rn+1 = (tn+1 )1 f (, y( ))d An+1 fk/2
() 0 ()
k=0
n1
2 2
1 h
Ak,j
n+1 fj+k1 akn+1 fn+k2 .
() ( + 3)
j=1 k=0 k=0
Theorem 3.1. Suppose that f (, u()) C 3 [0, T ]. Then, there exists a

constant C independent of all grid parameters such that
P
rn+1 Ch3 .
P r o o f. From the quadratic interpolation of f (t) in (3.1), we have

t1
2
P 1
rn+1 (tn+1 )1 f (, y( )) fk/2 Q0k/2 ( ) d
() t0
k=0
n1

1 tj+1 2
1 j
+ (tn+1 ) f (, y( )) fj+k1 Qj+k1 ( )] d
()
j=1 tj k=0

1 tn+1 h 2
1 k
+ (tn+1 ) f (, y( ))d an+1 fn+k2
() tn ( + 3)
k=0
= I + II + III.
Since f (, u()) C 3 [0, T ], by Lemma 2.2, we can easily check that
I, II Ch3 .
The Taylor expansion f around tn gives
1
f (t) = p2 (t) + f (n )(t tn )3 , n (tn , t), (3.4)
3
where
p2 (t) = fn + f (tn )(t tn ) + f (tn )(t tn )2 /2, thus p2 (t) P2 [0, T ].
Then by Lemma 3.1 with (3.4), we have
h
III (|a0 |p2 (tn2 ) fn2 | + |a1n+1 ||p2 (tn1 ) fn1 |
( + 3) n+1
C
+ h3+ Ch3+ .
6( + 1)
Combining I, II and III, the proof is complete. 2

Authenticated
Lemma 3.2. For k = 0, 1, 2 and j = 0, 1, , n 1,

k,j (n j)1 h , 0 < < 1,
|An+1 |
T 1 h, 1.
Theorem 3.2. (Global Predictor Error) Suppose that f (, u())

C 3 [0, T ], and is Lipschitz continuous in the second argument in (2.14), then
the global predictor error is
n

P LT 1 h LT 1 h

r + E1/2 + C Ej , 1,
n+1 () ()
P j=0
En+1 = n

Lh Lh

r P
+ E + C (n + 1 j)1 Ej , 0 < < 1.
n+1 () 1/2 ()
j=1
P r o o f.
2
P P 1 k,0
En+1 rn+1 + |An+1 ||fk/2 fk/2 |
()
k=0
n1
2
1 k,j
+ |A ||fj+k1 fj+k1 |
n+1
()
j=1 k=0
2
h
+ |akn+1 ||fn+k2 fn+k2 |.
( + 3)
k=0
1: By Lemmas 2.6, 3.2, and that f (t) is Lipschitz, we have

n
P P LT 1 h LT 1 h
En+1 rn+1 + E1/2 + C Ej .
() ()
j=0
0 < < 1: Similarly, we have that

n1 2
P P Lh 1 Lh
En+1 rn+1 + {n (E1/2 + E1 )} + (n j)1 Ej+k1
() ()
j=1 k=0
2
Lh k
+ |an+1 |En+k2 .
( + 3)
k=0
Carrying a similar step in (2.15) we have for k = 1, 2, , n 2,

(n + 1 k)1 + (n k)1 + (n k 1)1 6(n + 1 k)1 . (3.5)
We have thus

Authenticated
n
P P Lh Lh
En+1 rn+1 + E1/2 + C (n + 1 j)1 Ej .
() ()
j=1
2
3.2.2. Truncation and Global Error Analysis for Corrector. Now
we have the following truncation error of the corrector with quadratic in-
terpolation in the similar manner.
Theorem 3.3. (Truncation Error) With the same assumptions as

those of Theorem 3.2, we have
C Lh
rn+1 Ch3 + EP .
( + 1) n+1
P r o o f.
t1
2
C 1 1 0
rn+1 (tn+1 ) f (, y( )) fk/2 Qk/2 ( ) d
() t0
k=0
n1

1 tj+1 2

+ (tn+1 )1 f (, y( )) fj+k1 Qjj+k1 ( ) d
()
j=1 tj k=0
tn+1
1
+ (tn+1 )1 f (, y( )) fn1 Qnn1 ( ) + fn Qnn ( )
() tn
MT Lh
P
+fn+1 Qnn+1 ( ) d h3 + EP .
6( + 1) ( + 1) n+1
2
Theorem 3.4. (Global Error) With the same assumptions as those of

Theorem 3.2, we have
En+1 Ch3 ,
given E1 , E2 Ch3 and E1/2 Ch3 (0 < < 1), Ch2 ( > 1).
P r o o f.
2 n1
2
1 k,0 1 k,j
C
En+1 rn+1 + |An+1 ||fk/2 fk/2 | + |An+1 ||fk fk |
() ()
k=0 j=1 k=0
1
+ |A0,n
n+1 |f n2 fn2 | + |A1,n ||fn1 fn1 | + |A2,n ||fP f P | .
n+1 n+1 n n
()
MT CLT 1 h
1: Let D1 = h3 + E1/2 . Theorems 3.2, 3.3 give
6( + 1) ()

Authenticated

n
2Lh P CLT 1 h 2Lh
En+1 D1 + rn+1 + 1+ Ej .
( + 1) () ( + 1)
j=0
By the Discrete Gronwalls Inequality we have

2Lh P LT 2Lh
En+1 D1 + r exp C 1+ .
( + 1) n+1 () ( + 1)
MT CLh
0 < < 1: Let D2 = h3 + E . Theorems 3.2, 3.3 yield
6( + 1) () 1/2
n
C Lh Lh P Lh
En+1 rn+1 + E1/2 + En+1 +C (n + 1 j)1 Ej
() ( + 1) ()
j=1
n
2Lh P CLh 2Lh
D2 + r + 1+ (n + 1 j)1 Ej .
( + 1) n+1 () ( + 1)
j=1
The Discrete Gronwalls Inequality yields

2Lh P 2Lh
En+1 D2 + r E CLT 1+ .
( + 1) n+1 ( + 1)
2
Remark 3.1. As indicated in Theorem 2.2 and 3.2, the global error of
the scheme depends on that of the start-up, which is a common diculty for
multi-step methods. For a classical ODE where Z+ , one could instead
use a multi-stage method with corresponding accuracy, e.g., Runge-Kutta,
for the rst few steps then switch back to the multi-step one, when appro-
priate. For fractional ODEs, in order to preserve the order of accuracy,
one could invoke an iteration, say Newtons method, for approximating the
solution at the rst few time steps. Here, since we would use just explicit
schemes, we obtain approximate solutions with ner step size at the rst
several steps. We propose the start-up for our new scheme with quadratic
interpolation in the Appendix.
4. Numerical Results
In this section, we illustrate the accuracy and eciency of our new
schemes, both with linear and quadratic interpolation. For all below tests,
we use the following error estimates.
Pointwise error at nal time tN = T :
Ept (T ) = |yN yN |.
L2 error over [0, T ]:

Authenticated
1/2
N

EL2 = h |yj yj |2 .
j=0
We choose the following examples for our numerical experiments. For

schemes with linear interpolation, we compare the results obtained from our
new scheme Eqs. (2.9) and (2.10) with those from the ABM method Eqs.
(2.3) and (2.5). For high-order schemes with quadratic interpolation, we
compare the results of our scheme Eqs. (3.2) and (3.3) with the Improved
Algorithm II scheme developed by Li et al. ([19]). We denote MF-PCL
and MF-PCQ by our new schemes with linear and quadratic interpolation,
respectively. Similarly, the ABM scheme is denoted as ABM, and Alg II
for the improved algorithm II scheme.
Example 1.

D0 = f (t, y(t)),
y(0) = 0, y (0) = 0,
whose exact solution is
9
y(t) = t8 3t4+ 2 + t .
4
and f (t, y(t)) is given by

40320 8 (5+ 2 ) 4 9 3 4 3 3
f (t, y(t)) = t 3 t 2 + (+1) + t t
2 y(t) 2 .
(9) (5 2 ) 4 2
Example 2.

(4 + ) 3
D0 y(t) = t + t3+ y(t),
6
y(0) = 0, y (0) = 0,
y(t) = t3+ .
Example 3.

(5 + ) 4
D0 y(t) = t + t8+2 y 2 (t),
24
y(0) = 0, y (0) = 0,
y(t) = t4+ .
For all these examples, we choose the nal time T = 1.

Authenticated
1.8 1
ABM ABM
MFPCL MFPCL
Exact 0.9 Exact
1.6
0.8
1.4
0.7
1.2
0.6
1
y(t)
y(t)
0.5
0.8
0.4
0.6
0.3
0.4
0.2
0.2 0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t t
Figure 1. Numerical solutions with grid N=20 in Example

1(Left: = 0.25; Right: = 1.25.).
4.1. Linear Interpolation. Numerical errors and orders of accuracy for

Example 1 obtained from schemes with linear interpolation are listed in
Table 1 for = 0.25, = 0.5 and = 1.25. For both cases < 1 and
> 1, we observe that our MF-PCL scheme achieves a 2nd order, which
well agrees with our error analysis in the above section. We also note the
dependence of the accuracy order on of the ABM scheme. It shows much
better results of the MF-PCL over the ABM in case = 0.25, in both error
magnitudes and orders. For the case where > 1, both schemes behave
similarly and achieve 2nd order. Numerical solutions with grid N = 20 for
Example 1 are plotted in Fig. 1 for both values of .
We remark that in case of < 1, as shown in Table 1 (see also Tables
2 and 3), the accuracy of the MF-PCL is even better than 2nd-order. We
explain this behavior by the competence between the errors in the lag and
increment term of the truncation errors (see Eq. (2.13)) when is small.
Numerical errors, and accuracy orders obtained from the ABM and
MF-PCL for Example 2 with dierent values of are listed in Table. It
is shown that the MF-PCL outperforms the ABM in terms of both error
magnitudes and orders for < 1.
Similarly, for Example 3, we list and plot the errors in Table 3. Again,
we observe signicantly better results of the MF-PCL over those of the
ABM scheme.

Authenticated
= 0.25
ABM MF-PCL
N Ept roc EL2 roc Ept roc EL2 roc
10 2.50E-1 - 3.14E-1 - 1.74E-1 - 2.10E-1 -
20 1.81E-2 3.79 8.69E-2 1.85 1.46E-2 3.58 1.44E-2 3.87
40 3.61E-3 2.33 2.48E-2 1.81 2.64E-3 2.46 1.64E-3 3.14
80 1.45E-3 1.31 8.05E-3 1.63 5.07E-4 2.38 2.47E-4 2.73
160 6.58E-4 1.14 2.82E-3 1.52 9.95E-5 2.35 4.13E-5 2.58
320 2.97E-4 1.15 1.03E-3 1.45 2.02E-5 2.30 7.46E-6 2.47
= 0.5
10 1.79E-2 - 4.94E-2 - 2.66E-2 - 1.41E-2 -
20 1.81E-3 3.30 1.38E-2 1.84 5.30E-3 2.33 2.09E-3 2.75
40 4.17E-4 2.12 4.15E-3 1.73 1.07E-3 2.30 3.60E-4 2.54
80 1.76E-4 1.24 1.32E-3 1.65 2.31E-4 2.22 7.10E-5 2.35
160 7.98E-5 1.15 4.33E-4 1.61 5.31E-5 2.12 1.56E-5 2.19
320 3.39E-5 1.24 1.46E-4 1.57 1.27E-5 2.06 3.65E-6 2.10
= 1.25
10 5.53E-3 - 8.14E-3 - 1.02E-2 - 6.04E-3 -
20 1.59E-3 1.80 1.88E-3 2.11 2.34E-3 2.12 1.38E-3 2.13
40 4.33E-4 1.88 4.43E-4 2.09 5.70E-4 2.04 3.34E-4 2.04
80 1.14E-4 1.92 1.06E-4 2.07 1.41E-4 2.01 8.26E-5 2.02
160 2.97E-5 1.94 2.54E-5 2.06 3.52E-5 2.00 2.05E-5 2.01
320 7.66E-6 1.96 6.13E-6 2.05 8.80E-6 2.00 5.12E-6 2.00
Table 1. Numerical comparisons by linear interpolation in
Ex. 1.
4.2. Quadratic Interpolation. We now turn to check the cases with our
new high-order scheme with quadratic interpolation. We denote our new
scheme MF-PCQ.
In Tables 4 and 5, we show the numerical errors and orders obtained
from the MF-PCQ for Example 2 and Example 3 with various values of
, respectively. We also present the results obtained from the Improved
Algorithm II (abbreviated by Alg. II) by Li et al. ([19]) for comparison
purpose. For all these cases, the MF-PCQ scheme achieves the designed
3rd order for both cases of < 1 and 1, and are much better than the
comparing results of the Alg. II scheme. In case of = 1.5, we notice that
although the errors of the MF-PCQ are less than those of the Alg. II, the
former remains 3rd-order whereas the latter order is around 2.5.

Authenticated
= 0.25
ABM MF-PCL
10 1.20E-1 - 5.84E-2 - 2.17E-2 - 1.32E-2 -
20 4.46E-2 1.43 2.07E-2 1.49 3.71E-3 2.54 2.22E-3 2.57
40 1.65E-2 1.43 7.48E-3 1.47 6.64E-4 2.48 3.92E-4 2.50
80 6.20E-3 1.41 2.77E-3 1.44 1.25E-4 2.41 7.26E-5 2.43
160 2.36E-3 1.39 1.05E-3 1.40 2.46E-5 2.34 1.42E-5 2.36
320 9.15E-4 1.37 4.03E-4 1.38 5.09E-6 2.27 2.91E-6 2.29
= 0.5
10 4.39E-2 - 2.14E-2 - 8.33E-3 - 4.98E-3 -
20 1.38E-2 1.67 6.32E-3 1.76 1.53E-3 2.44 8.76E-4 2.51
40 4.41E-3 1.64 1.96E-3 1.69 3.20E-4 2.26 1.76E-4 2.31
80 1.44E-3 1.61 6.28E-4 1.64 7.28E-5 2.14 3.92E-5 2.17
160 4.82E-4 1.58 2.07E-4 1.60 1.74E-5 2.07 9.24E-6 2.08
320 1.63E-4 1.56 6.97E-5 1.57 4.25E-6 2.03 2.25E-6 2.04
= 1.25
10 1.33E-2 - 6.41E-3 - 9.46E-3 - 4.75E-3 -
20 3.17E-3 2.07 1.44E-3 2.16 2.34E-3 2.02 1.10E-3 2.10
40 7.60E-4 2.06 3.35E-4 2.10 5.83E-4 2.00 2.67E-4 2.05
80 1.83E-4 2.05 7.97E-5 2.07 1.46E-4 2.00 6.56E-5 2.02
160 4.43E-5 2.05 1.92E-5 2.05 3.64E-5 2.00 1.63E-5 2.01
320 1.08E-5 2.04 4.66E-6 2.04 9.10E-6 2.00 4.05E-6 2.01
Ex. 2.
We next compare the performance of our MF-PCQ and the Alg. II in

case the right-hand side is just C 0 continuous. For this case, we choose the
following test as suggested in [19].
Example 4.
2 1

t2 t1 y(t) + t2 t, 1,

D0 y(t) = (3 ) (2 )

2
t2 y(t) + t2 t, 1 < < 2.
(3 )
The exact solution is available as follows,
y(t) = t2 t.
Hence, the initial conditions are
y(0) = 0, y (0) = 1.

Authenticated
= 0.25
ABM MF-PCL
10 2.85E-1 - 9.86E-2 - 1.10E-1 - 3.80E-2 -
20 1.41E-1 1.02 4.00E-2 1.30 2.49E-2 2.15 6.99E-3 2.44
40 5.87E-2 1.26 1.49E-2 1.43 4.10E-3 2.60 1.07E-3 2.70
80 2.17E-2 1.43 5.32E-3 1.48 6.10E-4 2.75 1.68E-4 2.67
160 7.75E-3 1.49 1.91E-3 1.48 9.58E-5 2.67 2.92E-5 2.52
320 2.79E-3 1.47 7.03E-4 1.44 1.63E-5 2.56 5.56E-6 2.40
= 0.5
10 1.10E-1 - 3.92E-2 - 3.05E-2 - 1.18E-2 -
20 3.79E-2 1.54 1.15E-2 1.77 4.73E-3 2.69 1.80E-3 2.70
40 1.23E-2 1.62 3.47E-3 1.73 7.84E-4 2.59 3.29E-4 2.45
80 3.97E-3 1.63 1.08E-3 1.68 1.50E-4 2.38 6.97E-5 2.24
160 1.30E-3 1.61 3.51E-4 1.63 3.24E-5 2.21 1.61E-5 2.11
320 4.34E-4 1.58 1.16E-4 1.59 7.52E-6 2.11 3.89E-6 2.05
Ex. 3.
In Table 6, we list the numerical errors and orders for Example 4 with
= 0.3, 0.5 and 1.25. We note that our obtained results from the Alg.
II here are dierent in terms of error magnitudes comparing with those
given in Table 13 and 14 in [19]. Due to the violation of the conditions
given in Theorem 3.2, it is shown that both schemes cannot obtained the
designed convergence order. Nevertheless, we conclude that our new scheme
outperforms the Alg. II, especially for cases of < 1.
4.3. Computational issues. The proposed scheme does not repeat the
memory eect in approximating the predicted value y P which gives a benet
to reduce computational cost. Comparing with the ABM scheme it reduces
a half of computational cost roughly. For the quadratic interpolation we
do not analyze the computational cost because Alg II scheme uses the
intermediate grids. In Table 7, we show the computational time (in seconds)
for both linear and quadratic interpolation schemes for Example 2. We
observe a much cheaper cost of the proposed schemes comparing with those
of the ABM or Alg. II ones.
5. Conclusion
In this work, we have proposed a new class of accurate, ecient, and
explicit schemes using linear and quadratic interpolation. By implementing

Authenticated
= 0.2
Alg. II MF-PCQ
10 1.13E-1 - 5.47E-2 - 3.25E-3 - 2.93E-3 -
20 4.62E-2 1.29 2.14E-2 1.35 2.13E-4 3.93 2.32E-4 3.66
40 1.84E-2 1.33 8.33E-3 1.36 1.89E-5 3.49 2.02E-5 3.52
80 7.27E-3 1.34 3.25E-3 1.36 1.72E-6 3.46 1.83E-6 3.46
160 2.89E-3 1.33 1.28E-3 1.34 1.63E-7 3.40 1.71E-7 3.42
320 1.15E-3 1.32 5.12E-4 1.33 1.59E-8 3.36 1.65E-8 3.38
= 0.5
10 2.12E-2 - 1.02E-2 - 5.81E-4 - 5.63E-4 -
20 6.89E-3 1.62 3.14E-3 1.70 5.09E-5 3.51 4.69E-5 3.58
40 2.26E-3 1.61 9.96E-4 1.66 5.24E-6 3.28 4.56E-6 3.36
80 7.48E-4 1.59 3.25E-4 1.61 5.93E-7 3.14 4.96E-7 3.20
160 2.52E-4 1.57 1.08E-4 1.58 7.06E-8 3.07 5.77E-8 3.10
320 8.59E-5 1.55 3.68E-5 1.56 8.63E-9 3.03 6.97E-9 3.05
= 1.5
10 2.17E-4 - 9.28E-5 - 1.16E-3 - 6.37E-4 -
20 4.05E-5 2.42 1.61E-5 2.53 1.55E-4 2.91 8.35E-5 2.93
40 7.34E-6 2.46 2.81E-6 2.52 1.99E-5 2.95 1.07E-5 2.97
80 1.31E-6 2.48 4.94E-7 2.51 2.53E-6 2.98 1.35E-6 2.98
160 2.34E-7 2.49 8.69E-8 2.51 3.19E-7 2.99 1.70E-7 2.99
320 4.15E-8 2.50 1.53E-8 2.50 4.00E-8 2.99 2.13E-8 3.00
Table 4. Numerical comparisons by quad interpolation in
Ex. 2.
the same discretization method for the lag term in both the prediction and
correction stages, we have greatly reduced the overall computational costs.
This is a main diculty issue for the development of numerical schemes for
fractional dierential equations due to the so-called memory eect. Our
new schemes also improve the accuracy to achieve uniform orders for both
cases of 0 < < 1 and > 1, especially for the former case where, to
our knowledge, other schemes accuracy depends on this . This property
is obtained by a new prediction method, in which we did not employ a
low-order interpolation of the right-hand side f (t) for the prediction step.
Instead, we have introduced a new explicit method, which is of the same
order with that of the correction stage.
A variety of numerical tests show that our schemes are better than
other comparing ones, especially for cases with < 1 in which our new
schemes signicantly improve both error magnitudes and orders.

Authenticated
= 0.2
Alg. II MF-PCQ
10 2.36E-1 - 8.26E-2 - 3.88E-2 - 1.31E-2 -
20 1.27E-1 0.90 3.68E-2 1.17 4.80E-3 3.01 1.26E-3 3.37
40 5.93E-2 1.10 1.53E-2 1.27 3.97E-4 3.60 9.26E-5 3.77
80 2.45E-2 1.28 6.01E-3 1.35 2.74E-5 3.86 6.68E-6 3.79
160 9.45E-3 1.37 2.31E-3 1.38 1.99E-6 3.78 5.41E-7 3.63
320 3.58E-3 1.40 8.88E-4 1.38 1.59E-7 3.65 4.79E-8 3.50
= 0.5
10 5.20E-2 - 1.84E-2 - 4.92E-3 - 1.98E-3 -
20 1.82E-2 1.51 5.58E-3 1.72 3.45E-4 3.84 1.45E-4 3.77
40 6.08E-3 1.58 1.74E-3 1.68 2.69E-5 3.68 1.32E-5 3.46
80 2.01E-3 1.60 5.59E-4 1.64 2.46E-6 3.45 1.41E-6 3.23
160 6.69E-4 1.59 1.84E-4 1.60 2.56E-7 3.27 1.63E-7 3.11
320 2.26E-4 1.57 6.20E-5 1.57 2.90E-8 3.14 1.96E-8 3.05
= 1.5
10 5.25E-4 - 1.75E-4 - 3.06E-3 - 1.45E-3 -
20 1.07E-4 2.30 2.96E-5 2.57 4.00E-4 2.93 1.80E-4 3.01
40 2.03E-5 2.40 5.06E-6 2.55 5.12E-5 2.97 2.24E-5 3.01
80 3.72E-6 2.45 8.76E-7 2.53 6.48E-6 2.98 2.79E-6 3.00
160 6.69E-7 2.47 1.53E-7 2.52 8.15E-7 2.99 3.49E-7 3.00
320 1.19E-7 2.49 2.69E-8 2.51 1.02E-7 3.00 4.36E-8 3.00
Ex. 3.
The low cost and high accuracy of our new schemes promise various ap-
plications, especially in dynamical problems or time derivative discretiza-
tion for PDE-based problems.
Appendix A. Start-up of the Scheme

To nd a desired accuracy for y1 , y2 we approximate the predicted values
at t1/4 , t1/2 by using the constant, linear and quadratic interpolation as
seen in Fig. 2. Let L0 (fa ) be the constant interpolation f at t = a.
That is L0 (fa ) = f (a). Let L1 (fa , fb ) and L2 (fa , fb , fc ) be the linear and
quadratic interpolation of f with grids (a, b) and (a, b, c), respectively. In
the algorithm below we describe the lag term with a detail interpolation
and the last intervals of predictor and corrector with detail interpolations.

Authenticated
= 0.3
Alg. II MF-PCQ
10 2.04E-2 - 1.41E-2 - 8.44E-4 - 4.77E-3 -
20 7.45E-3 1.45 5.05E-3 1.48 4.34E-6 7.60 1.21E-3 1.97
40 2.72E-3 1.46 1.81E-3 1.48 4.71E-7 3.20 3.23E-4 1.91
80 1.00E-3 1.44 6.55E-4 1.46 6.40E-8 2.88 8.88E-5 1.87
160 3.73E-4 1.42 2.41E-4 1.44 7.05E-8 0.14 2.51E-5 1.82
320 1.41E-4 1.41 9.00E-5 1.42 3.38E-8 1.06 7.30E-6 1.78
= 0.5
10 6.98E-3 - 5.38E-3 - 1.03E-4 - 3.25E-3 -
20 2.22E-3 1.65 1.62E-3 1.73 5.78E-6 4.15 8.29E-4 1.97
40 7.23E-4 1.62 4.93E-4 1.72 3.05E-6 0.92 2.30E-4 1.85
80 2.40E-4 1.59 1.53E-4 1.69 1.88E-6 0.70 6.84E-5 1.75
160 8.09E-5 1.57 4.90E-5 1.65 8.12E-7 1.21 2.16E-5 1.67
320 2.76E-5 1.55 1.61E-5 1.61 3.16E-7 1.36 7.10E-6 1.60
= 1.25
10 5.88E-5 - 7.93E-5 - 3.02E-4 - 2.81E-4 -
20 4.40E-6 3.74 3.04E-5 1.39 8.62E-5 1.81 8.72E-5 1.69
40 5.64E-6 0.36 1.15E-5 1.40 2.48E-5 1.80 2.68E-5 1.70
80 2.56E-6 1.14 4.01E-6 1.53 7.21E-6 1.78 8.11E-6 1.72
160 9.43E-7 1.44 1.32E-6 1.61 2.12E-6 1.77 2.44E-6 1.74
320 3.18E-7 1.57 4.18E-7 1.65 6.25E-7 1.76 7.28E-7 1.74

Ex. 4.
Figure 2. Grid for the start-up of the new quadratic scheme.

Authenticated
= 0.25
N 100 200 400 600 800 1000
ABM 0.002 0.010 0.040 0.090 0.160 0.262
MF-PCL 0.001 0.007 0.025 0.050 0.096 0.150
= 0.5
N 100 200 400 600 800 1000
Alg. II 0.210 0.491 1.102 2.127 2.643 3.224
MF-PCQ 0.123 0.231 0.622 0.981 1.490 1.471
Table 7. Computing time with linear and quadratic inter-
polations in Ex. 2.
Algorithm 1. (Start-up Procedure): y1/4 y1/2 y1 y2 .

(1) L0 (f0 ) y1/4
P L1 (f , fP ) y
0 1/4 1/4
1 0 P1 P1 P2

(2) L (f0 , f1/4 ) y1/2 ; L (f1/4 ) y1/2 L1 (f1/4 , f1/2 ) y1/2
2 P
L (f0 , f1/4 , f ) y1/2
2
1/2
(3) L (f0 , f1/2 ) y1 ; L0 (f1/2 ) y1P1 L1 (f1/2 , f1P1 )
1 y1P2
L2 (f0 , f1/2 , f1P2 ) y1
(4) L2 (f0 , f1/2 , f1 ) y2 ; L0 (f1 ) y2P1 L1 (f1 , f2P1 ) y2P2
L2 (f0 , f1 , f2P2 ) y2
Acknowledgments
This work was supported by Basic Science Research Program through
the National Research Foundation of Korea (NRF) funded by the Ministry
of Education (NRF-2014R1A1A2A16051147). We would like to address our
thanks to professor Changpin Li for the correction of the typos of the Alg
II scheme in [19].
References
[1] D.A. Benson, S.W. Wheatcraft, and M.M. Meerschaert, Application of
a fractional advection-dispersion equation. Water Resour. Res. 36, No
6 (2000), 14031412.
[2] M. Caputo, Linear models of dissipation whose Q is almost frquency
independent, II. The Geophysical J. of the Royal Astronomical Society
13 (1967), 529539; Re-published in: Fract. Calc. Appl. Anal. 11, No
1 (2008), 314.

Authenticated
[3] A.R. Carella, and C.A. Dorao, Least-squares spectral method for the
solution of a fractional advection-dispersion equation. J. Comput. Phys.
232 (2013), 3345.
[4] G. Capobianco, D. Conte, I. Del Prete, and E. Russo, Fast Runge-
Kutta methods for nonlinear convolution systems of Voterra integral
equations. BIT Numerical Mathematics 47, No 2 (2007), 259275.
[5] D. Conte, and I.D. Prete, Fast collocation methods for Volterra integral
equations of convolution type. J. Comput. Appl. Math. 196 (2006),
652663.
[6] J. Cao, and C. Xu, A high order schema for the numerical solution
of the fractional ordinary dierential equations. J. Comput. Phys. 238
(2013), 154168.
[7] W. Deng, Numerical algorithm for the time fractional Fokker-Planck
equation. J. Comput. Phys. 227 (2007), 15101522.
[8] K. Diethelm, The Analysis of Fractional Dierential Equations.
Springer, Berlin, 2010.
[9] K. Diethelm, N.J. Ford, and A.D. Freed, A predictor-corrector ap-
proach for the numerical solution of fractional dierential equations.
Nonlinear Dynamics 29 (2002), 322.
[10] K. Diethelm, N.J. Ford, and A.D. Freed, Detailed error analysis for a
fractional Adams method. Numerical Algorithms 36 (2004), 3152.
[11] V.J. Ervin, N. Heuer, and J.P. Roop, Numerical approximation of a
time dependent, nonlinear, space-fractional diusion equation. SIAM
J. Numer. Anal. 45 (2007), 572591.
[12] F. Fu, H. Wang, A preconditioned fast nite dierence method for
space-time fractional partial dierential equations. Fract. Calc. Appl.
Anal. 20, No 1 (2017), 88116; DOI: 10.1515/fca-2017-0005;
https://www.degruyter.com/view/j/fca.2017.20.issue-1/
issue-les/fca.2017.20.issue-1.xml.
[13] N.J. Ford, M.L. Morgado, M. Rebelo, Nonpolynomial collocation ap-
proximation of solutions to fractional dierential equations. Fract. Calc.
Appl. Anal. 16, No 4 (2013), 874891; DOI: 10.2478/s13540-013-0054-
3; https://www.degruyter.com/view/j/fca.2013.16.issue-4/
issue-les/fca.2013.16.issue-4.xml.
[14] I. Goychuk, E. Heinsalu, M. Patriarca, G. Schmid, and P. Hanggi,
Current and universal scaling in anomalous transport. Phys. Rev. E.
73, No 2 (2006), ID # 020101.
[15] G.-H. Gao, Z.-Z. Sun, and H.-W. Zhang, A new fractional numerical
dierentiation formula to approximate the Caputo fractional derivative
and its applications. J. Comput. Phys. 259 (2014), 3350.

Authenticated
[16] Z.-P. Hao, Z.-Z. Sun, and W.-R. Cao, A fourth-order approximation
of fractional derivatives with its application. J. Comput. Phys. 281
(2015), 787805.
[17] P. Kumar, and O.P. Agrawal,An approximate method for numerical
solution of fractional dierential equations. Signal. Process. 86 No 10
(2006), 26022610.
[18] A. Kilbas, H. Srivastava, and J. Trujillo, Theory and Applications of
Fractional Dierential Equations. Elsevier, Boston, 2006.
[19] C. Li, A. Chen, and J. Ye, Numerical approaches to fractional calculus
and fractional ordinary dierential equation. J. Comput. Phys. 230
(2011), 33523368.
[20] F. Mainardi, Fractional Calculus and Waves in Linear Viscoelasticity.
Imperial College Press, London, 2010.
[21] I. Podlubny, Fractional Dierential Equations. Acad. Press, New York,
1999.
[22] J. Sabatier, O.P. Agrawal, and J.A.T. Machado, Advances in Frac-
tional Calculus: Theoretical Developments and Applications in Physics
and Engineering. Springer, Dordrecht, 2007.
[23] Y. Yan, K. Pal, N.J. Ford, Higher order numerical methods for solving
fractional dierential equations. BIT Numer. Math.54 (2014), 555584.
[24] Z.-Z. Sun, and Z. Wu, A fully discrete dierence scheme for a diusion-
wave system. Appl. Num. Math. 56 (2006), 193209.
[25] Y. Yan, K. Pal, N.J. Ford, Higher order numerical methods for solving
fractional dierential equations. BIT Numer. Math. 54 (2014), 555
584.
[26] G.M. Zaslavsky, Chaos, fractional kinetics, and anomalous transport.
Phys. Rep. 371 No 6 (2002), 461580.
1 School of Mathematical Sciences, Monash University
9 Rainforest Walk, Clayton, Melbourne, VIC 3800, AUSTRALIA
e-mail: binh.nguyen@monash.edu
2Dept. of Mathematical Sciences
Ulsan National Institute of Science and Technology (UNIST)
Ulsan 44919, Republic of KOREA Received: May 23, 2016
e-mail: bsjang@unist.ac.kr/ thienbinh84@unist.ac.kr Revised: January 9,
2017
Please cite to this paper as published in:

Fract. Calc. Appl. Anal., Vol. 20, No 2 (2017), pp. 447476,
DOI: 10.1515/fca-2017-0023

Authenticated

Nguyen &amp; Jang (2017)

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Nguyen &amp; Jang (2017)

Încărcat de

Drepturi de autor:

Formate disponibile

RESEARCH PAPER

A HIGH-ORDER PREDICTOR-CORRECTOR METHOD

Thien Binh Nguyen 1,2 , Bongsoo Jang 2

An accurate and ecient new class of predictor-corrector schemes are

Brought to you by | University of Toronto-Ocul

Brought to you by | University of Toronto-Ocul

For simplicity, we assume that the grid is uniform, i.e., h = tj+1 tj ,

Then the approximation yn+1 y(tn+1 ), which is called a correction, can

Brought to you by | University of Toronto-Ocul

or the Grunwald-based method ([16]). Fractional numerical methods de-

2. Second-order Predictor-Corrector Scheme

Brought to you by | University of Toronto-Ocul

Brought to you by | University of Toronto-Ocul

Brought to you by | University of Toronto-Ocul

P r o o f. Since P1 [0, T ], there exist 0 , 1 R such that

We then propose our new scheme with linear interpolation as follows.

2.2.1. Truncation and Global Error Analysis for Predictor. From

Brought to you by | University of Toronto-Ocul

Lemma 2.2. (Interpolation Errors) Let f C n+1 [a, b] and pn Pn [a, b]

Lemma 2.3. For R+ ,

Lemma 2.4. (Discrete Gronwalls Inequality, [6]) Let {an }N N

In the above, E () denotes the (one-parameter) Mittag-Leer function,

Brought to you by | University of Toronto-Ocul

P r o o f. In view of the interpolation (2.2), Eq. (2.11) is written as

By the Taylor expansion of f (, u()) around tn , we have that

Brought to you by | University of Toronto-Ocul

Lemma 2.5. For k = 0, 1 and j = 0, 1, , n 1,

Lemma 2.6. For 1,

Theorem 2.2. (Global Error)Suppose that f (, u()) C 2 [0, T ], and

|f (t, u1 ) f (t, u2 )| L|u1 u2 |, u1 , u2 R, (2.14)

Brought to you by | University of Toronto-Ocul

Brought to you by | University of Toronto-Ocul

2.2.2. Truncation and Global Error Analysis for Corrector. Carry-

Theorem 2.3. (Truncation Error)With the same assumptions as those

P r o o f. The proof follows that of Theorem 2.1:

Brought to you by | University of Toronto-Ocul

Theorem 2.4. (Global Error) With the same assumptions as those of

En+1 = |yn+1 yn+1 | Ch2 ,

given that the starting error E1 Ch2 .

By the Discrete Gronwalls Inequality in Lemma 2.4 we have

0 < < 1: Carrying similar procedure for Rn in (2.16) combined

Brought to you by | University of Toronto-Ocul

3. Third-order Predictor-Corrector Scheme

Brought to you by | University of Toronto-Ocul

where for j = 0, i = i/2 and for 1 j n, i = i + j 1,

The increment term is

where a0n+1 , a1n+1 , a2n+1 can be computed by Lemma 3.1.

3.2. Error Analysis.

3.2.1. Truncation and Global Error Analysis for Predictor. By a

Brought to you by | University of Toronto-Ocul

Theorem 3.1. Suppose that f (, u()) C 3 [0, T ]. Then, there exists a

P r o o f. From the quadratic interpolation of f (t) in (3.1), we have

Brought to you by | University of Toronto-Ocul

Lemma 3.2. For k = 0, 1, 2 and j = 0, 1, , n 1,

Theorem 3.2. (Global Predictor Error) Suppose that f (, u())

1: By Lemmas 2.6, 3.2, and that f (t) is Lipschitz, we have

0 < < 1: Similarly, we have that

Carrying a similar step in (2.15) we have for k = 1, 2, , n 2,

Brought to you by | University of Toronto-Ocul

Theorem 3.3. (Truncation Error) With the same assumptions as

Theorem 3.4. (Global Error) With the same assumptions as those of

Brought to you by | University of Toronto-Ocul

By the Discrete Gronwalls Inequality we have

Brought to you by | University of Toronto-Ocul

Nguyen & Jang (2017)

Nguyen & Jang (2017)