Numerical Algorithms (2022) 91:227–250

https://doi.org/10.1007/s11075-022-01260-8
ORIGINAL PAPER

Generalized second derivative linear multistep

methods for ordinary differential equations

Tahereh Majidi1 · Ali Abdi1 · Gholamreza Hojjati1

Received: 6 December 2021 / Accepted: 8 January 2022 / Published online: 14 April 2022
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2022

Abstract
This paper is devoted to investigate the modified extended second derivative back-
ward differentiation formulae from second derivative general linear methods point
of view. This makes it possible to open some maneuver rooms in developing the
methods with superior features by perturbing the abscissa vector of the methods. The
proposed methods are constructed to have better accuracy and stability properties in
comparison with the original ones. These improvements are verified by giving some
numerical experiments.

Keywords Stiff initial value problems · Extended second derivative multistep

methods · Second derivative general linear methods · A- and A(α)-stability ·
Error constant

Mathematics Subject Classification (2010) 65L05

1 Introduction

In designing the efficient numerical methods for solving stiff initial value problems
(IVPs) of ordinary differential equations (ODEs), one should consider the accuracy
and stability properties together with the modest computational cost. Usually, there
is a conflict between these aims: although linear multistep methods (LMMs) has low

 Ali Abdi
a [email protected]

Tahereh Majidi
[email protected]

Gholamreza Hojjati
[email protected]

1 Faculty of Mathematical Sciences, University of Tabriz, Tabriz, Iran

228 Numerical Algorithms (2022) 91:227–250

computational cost, they suffer from the accuracy and stability properties because
of their structure and existing of some barriers on the order of methods with desir-
able stability properties [19]; In contrast, Runge–Kutta (RK) methods usually do not
have such drawbacks whereas their computational cost grows considerably with the
number of stages and so with the order. To achieve a good balance between these fea-
tures and circumvent the barriers, various modifications on these traditional methods
have been done in some directions. One of the most famous and successful directions
for LMMs is using of super-future point technique based on backward differentia-
tion formula (BDF) which led to introduce extended BDF (EBDF) [16], modified
EBDF (MEBDF) [17], and matrix-free MEBDF [25]. The state-of-the-art discovery
of general linear methods (GLMs) by Butcher [9] as a middle ground between the
traditional LMMs, RK methods, and their modifications make possibility of deriv-
ing methods which are neither RK method nor LMMs and nor slight modifications
of these methods. In the last decades, GLMs have been investigated, for instance,
in [12–15, 27] and the efficient codes dim18 [10], dim13s [28], and irks14
[8] based on these methods have been developed for nonstiff and stiff ODEs. Con-
sidering MEBDF from GLMs point of view, perturbations of these methods in two
classes referred to as perturbed MEBDF (PMEBDF) and fully perturbed MEBDF
(FPMEBDF), which improve their stability properties while preserve the order, have
been introduced in [20].
One of the successful directions to circumvent the barriers for traditional methods
is developing methods incorporating the second derivative of the solution. Various
methods have been introduced in this way in the class of RK and multistep meth-
ods, for instance, see [18, 21, 23]. Furthermore, GLMs have been also extended
to second derivative GLMs (SGLMs) including second derivative of the solution in
the formula. For more details on SGLMs, see [3–5, 7, 11]. Second derivative BDF
methods (SDBDFs) [23] as a popular class of second derivative multistep methods
have been equipped to super-future point technique which the resulting methods have
been referred to as extended SDBDF (ESDBDF) and modified ESDBDF (MES-
DBDF) [24]. The k-step ESDBDF methods for the numerical solution of the initial
value problem
 
y (t) = f (y(t)), t ∈ [t0 , T ],
(1)
y(t0 ) = y0 ,

with f : Rm → Rm and m as the dimension of the system, have the form


k
 k fn+k + h2 (
αj yn+j = hβ γk gn+k − 
γk+1 gn+k+1 ), (2)
j =0

where g := y  = fy f , fn+k = f (yn+k ), gn+k = g(yn+k ) and gn+k+1 = g(yn+k+1 ).

In formula (2) the coefficients  k , 
αj , j = 0, . . . , k, β γk , 
γk+1 are normalized by
making αk = 1 and chosen so that the order of the methods is equal to k + 2. Due to
using of the solution at super-future point tn+k+1 , implementation of these methods
Numerical Algorithms (2022) 91:227–250 229

requires a separate predictor to compute yn+k+1 , and then, (2) is used as a corrector.
The used predictor in [24] is the k-step SDBDF method defined by

k−1
yn+k + αj yn+j = hβk fn+k + h2 γk gn+k , (3)
j =0

where the coefficients are computed so that the formula has order k + 1. Indeed,
to implement ESDBDF method (2), the following three-stage k-step scheme was
proposed in which the first two stages are k-step SDBDF method (3):
Stage 1: Compute y n+k as the solution of the k-step SDBDF


k−1
y n+k + αj yn+j = hβk f n+k + h2 γk g n+k , (4)
j =0

where f n+k = f (y n+k ) and g n+k = g(y n+k ).

Stage 2: Compute y n+k+1 as the solution of the k-step SDBDF


k−2
y n+k+1 + αk−1 y n+k + αj yn+j +1 = hβk f n+k+1 + h2 γk g n+k+1 , (5)
j =0

where f n+k+1 = f (y n+k+1 ) and g n+k+1 = g(y n+k+1 ).

Stage 3: Compute yn+k as the solution of the k-step ESDBDF

k−1
yn+k +  k fn+k + h2 (
αj yn+j = hβ γk gn+k − 
γk+1 g n+k+1 ).
j =0

In computational point of view, a modified version of ESDBDF scheme, refereed

to as MESDBDF, was introduced in which the last stage is replaced by the k-step
formula
Stage 3*: Compute yn+k from (2) as the solution of

k−1
yn+k +  k −betak )f n+k +hβk fn+k +h2 (
αj yn+j = h(β γk − γk )g n+k
j =0

−h2 
γk+1 g n+k+1 + h2 γk gn+k . (6)
This modification makes the scheme to have the same Jacobian matrix I − hβk ∂f
∂y −
h2 γk ∂g
∂y with Im as the identity matrix of dimension m, for the all stages which
reduces the computational cost in practical implementation.
In this paper, considering MESDBDF methods as SGLMs, we analyze the local
truncation error of the methods, and then, we introduce a more general class of
these methods which their stability and accuracy properties are improved while their
structure and computational complexity are preserved.
The organization of the paper is as follows. In Section 2, we review SGLMs
and their basic features together with the representation of MESDBDF methods as
230 Numerical Algorithms (2022) 91:227–250

SGLMs. Section 3 is devoted to analyzing the local truncation error of SGLMs in

the case the sage order is one unit less than the order of the methods. In Section 4, a
perturbation of MESDBDF to derive methods with smaller error constants and more
desirable stability properties is described. Some examples of such methods of various
orders are constructed in Section 5. The numerical results of applying the proposed
methods to some stiff problems are reported in Section 6. The paper is closed in
Section 7 by giving some concluding remarks.

2 SGLMs and general frame for MESDBDFs

In this section, we briefly review the main features of SGLMs and the representation
of MESDBDF methods as SGLMs.
SGLMs with s internal stages and r external stages for the numerical solution of
(1) are defined by
Y [n] = h(A ⊗ Im )f (Y [n] ) + h2 (A ⊗ Im )g(Y [n] ) + (U ⊗ Im )y [n−1] ,
y [n] = h(B ⊗ Im )f (Y [n] ) + h2 (B ⊗ Im )g(Y [n] ) + (V ⊗ Im )y [n−1] , (7)
n = 1, 2, · · · , N, where Nh = T − t0 , h is the stepsize, and ⊗ is the Kronecker
product of two matrices. Here, the vector Y [n] = [Yi[n] ]si=1 denotes approximations
of stage order q to the vector y(tn−1 + ch) = [y(tn−1 + ci h)]si=1 , i.e.,
p  k
 
c
Y [n] = ⊗ hk y (k) (tn−1 ) + O(hq+1 ), (8)
k!
k=0

where c = [c1 c2 . . . cs ]T is the abscissa vector and ck stands for component-wise

powers of the vector c. Also, the vectors f (Y [n] ) = [f (Yi[n] )]si=1 and g(Y [n] ) =
[g(Yi[n] )]si=1 are the stage first and second derivative values. The input and out-
put vectors at the step number n respectively denoted by y [n−1] = [yi[n−1] ]ri=1 and
y [n] = [yi[n] ]ri=1 are approximations of order p to the linear combinations of scaled
derivatives of y(tn ), i.e.,
p 
 
y [n−1] = qk ⊗ hk y (k) (tn−1 ) + O(hp+1 ), (9)
k=0

and
p 
 
y [n] = qk ⊗ hk y (k) (tn ) + O(hp+1 ), (10)
k=0

for some real r-dimensional vectors qk , k = 0, 1, . . . , p. Indeed, an SGLM is char-

acterized by for integers (p, q, r, s) and its coefficients matrices A, A ∈ Rs×s , U ∈
Rs×r , B, B ∈ Rr×s , and V ∈ Rr×r .
Numerical Algorithms (2022) 91:227–250 231

In order to derive the order and stage conditions for SGLMs of order p and stage
order q = p, or p − 1, introducing the Vandermonde matrix C, the shifting matrix
K, the matrix W , and the upper triangular Toeplitz matrix

c2 cp
C = 1 c ··· , K = 0 e1 e2 · · · ep ,
2! p!
W = q0 q1 q2 · · · qp ,
⎡ ⎤
1 1 1
1 ...
⎢ 1! 2! p! ⎥
⎢ ⎥
⎢ 1 1 ⎥
⎢0 ⎥
⎢ 1 ...
(p − 1)! ⎥
⎢ 1! ⎥
⎢ .. ⎥
E = exp(K) = ⎢ 0 0 1 .
.. ⎥,
⎢ . ⎥
⎢ ⎥
⎢. .. .. . . 1 ⎥
⎢ .. . ⎥
⎢ . . ⎥
⎣ 1! ⎦
0 0 0 ... 1

 as the p
with ej as the j th unit standard vector in Rp+1 , and defining the matrix X
first columns of a given matrix X, the following results have been proved.

Theorem 1 [4, 7] Assume that y [n−1] satisfies (9). Then, the SGLM (7) of order p
and stage order q = p satisfies (8) and (10) if and only if

U W = C − ACK − ACK 2 ,
V W = W E − BCK − BCK 2 .

Theorem 2 [2] Assume that y [n−1] satisfies (9). Then, the SGLM (7) of order p and
stage order q = p − 1 satisfies (8) and (10) if and only if

UW =C − AC K 
 − ACK K, (11)
V W = W E − BCK − BCK 2 . (12)

The linear stability properties of SGLMs are related to the stability function

p(w, z) = det(wIr − M(z)),

in which the stability matrix M(z) is given by

M(z) = V + (zB + z2 B)(I − zA − z2 A)−1 U .

Construction and implementation of these methods in different classes have been

studied and investigated in many papers such as [1–4, 6, 11].
Traditional numerical methods incorporating second derivative of the solution
into the formula can be represented as SGLMs. The representation of MESDBDF
232 Numerical Algorithms (2022) 91:227–250

scheme (4)–(6) as SGLM has been given in [22] as follows. By substituting (4) into
(5), we get

k−1
y n+k+1 = αk−1 α0 yn + (αk−1 αj − αj −1 )yn+j +1 − hαk−1 βk f n+k + hβk f n+k+1
j =0

−h2 αk−1 γk g n+k + h2 γk g n+k+1 . (13)

Then, MESDBDF scheme based on relations (4), (13) and (6) can be written as
SGLM (7) with s = 3, r = k, the abscissae vector c = [k +1k +2k +1]T , the vectors
of internal stages Y [n] , f (Y [n] ), g(Y [n] ) and the vector of external stages y [n] as
⎡ ⎤
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ yn+k
y n+k f n+k g n+k ⎢ yn+k−1 ⎥
⎢ ⎥
Y [n] = ⎣ y n+k+1 ⎦ , f (Y [n] ) = ⎣ f n+k+1 ⎦ , g(Y [n] ) = ⎣ g n+k+1 ⎦ , y [n] = ⎢ . ⎥ ,
yn+k gn+k ⎣ .. ⎦
fn+k
yn+1
and the coefficients matrices given by
⎡ ⎤ ⎡ ⎤
βk 0 0 γk 0 0
A = ⎣ −αk−1 βk βk 0 ⎦ , A = ⎣ −αk−1 γk γk 0 ⎦,
k − βk 0 βk
β γk − γk −
 γk+1 γk
⎡ ⎤
−αk−1 −αk−2 ··· −α1 −α0
U = ⎣ αk−1 αk−1 − αk−2 αk−1 αk−2 − αk−3 · · · αk−1 α1 − α0 αk−1 α0 ⎦ ,
−αk−1 −αk−2 ··· − α1 −α0
⎡ ⎤ ⎡ ⎤
k − βk
β 0 βk γk − γk −
 γk+1 γk
⎢ 0 0 0 ⎥ ⎢ 0 0 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ .. .. .. ⎥ , ⎢ .. .. .. ⎥ ,
B =⎢ . . . ⎥ B=⎢ . . . ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0 0 ⎦ ⎣ 0 0 0 ⎦
0 0 0 0 0 0
⎡ ⎤
−
αk−1 −
αk−2 · · · −
α1 −
α0
⎢ 1 0 ··· 0 0 ⎥
⎢ ⎥
⎢ .. .. .. ⎥
V =⎢ . . . 0 0 ⎥ .
⎢ ⎥
⎣ 0 0 ··· 0 0 ⎦
0 0 ··· 1 0
It is easy to show that in this representation, the vectors qj , j = 0, 1, . . . , p,
appearing in the input and output vectors take the form
(k − i + 1)j
qj = , j = 0, 1, . . . , k + 1. (14)
j! i=1,2,...,k
Considering traditional second derivative methods from SGLMs point of view, makes
the possibility of obtaining some modifications of the methods with desirable sta-
bility and/or accuracy. In this fashion, MESDBDFs have been already modified by
Numerical Algorithms (2022) 91:227–250 233

perturbations of the coefficients matrices B and B in [22] to derive methods with

better stability properties. In this paper, we are going to modify MESDBDFs by per-
turbing the abscissa vector c with the aim of improving the stability and accuracy
properties. The similar idea has been already used to improve MEBDF schemes in
[26] and the derived methods have been referred to as “generalized linear multistep
methods” (GLMMs).

3 The local truncation error of SGLMs for q = p − 1

The local truncation error of SGLMs in the case q = p have been already analyzed in
[4]. In this section, we are going to study it for the case q = p −1 which is applicable
for MESDBDF schemes where the order of two first stages is one unit less than the
order of overall method.

Definition 1 [4] The local truncation error, lte(tn ), of SGLM (7) at the point tn is
given by
[n] )
lte(tn ) = (W ⊗ Im )z(tn , h) − h(B ⊗ Im )f (Y
−h (B ⊗ Im )g(Y
2  ) − (V W ⊗ Im )z(tn−1 , h),
[n]
(15)
[n] is defined by
where Y
[n] = h(A ⊗ Im )f (Y
Y [n] ) + h2 (A ⊗ Im )g(Y
[n] ) + (U W ⊗ Im )z(tn−1 , h), (16)
and the vector z(t, h) stands for the Nordsieck vector defined by
⎡ ⎤
y(t)
⎢ hy  (t) ⎥
⎢ ⎥
z(t, h) := ⎢ .. ⎥.
⎣ . ⎦
hp y (p) (t)

The following theorem gives a specific form for lte(tn ) of SGLMs at the point tn
in the case q = p − 1.

Theorem 3 The local truncation error lte(tn ) of the method (7) with q = p − 1 at
the point tn is given by
lte(tn ) = (ϕp ⊗ Im )hp+1 y (p+1) (tn−1 )
∂f
+(ψp ⊗ Im )hp+1 (tn−1 , y(tn−1 ))y (p) (tn−1 ) + O(hp+2 ), (17)
∂y
with
ϕp = W Ep+1 − BCp − BCp−1 , (18)
and
ψp = B(Cp − ACp−1 − ACp−2 − U qp ), (19)
234 Numerical Algorithms (2022) 91:227–250

where the vector Ep+1 is given by

T
1 1 1
Ep+1 = ... ,
(p + 1)! p! 1!
cp cp−1 cp−2
and the vectors Cp , Cp−1 , and Cp−2 are given by , , and ,
p! (p − 1)! (p − 2)!
respectively.

Proof Suppose that the method (7) has order p and stage order q = p − 1, we have
[n] = y(etn−1 + ch) + γ (tn−1 )hp + O(hp+1 ),
Y (20)
where e is the all-ones s-dimensional vector, y(etn−1 + ch) = [y(tn−1 + ci h)]si=1 ,
[n] . Substituting
and γ (tn−1 ) stands for the principal part of the error in stage values Y
the last relation into (16) leads to
y(etn−1 + ch) + γ (tn−1 )hp = h(A ⊗ Im )y  (etn−1 + ch)
+h2 (A ⊗ Im )y  (etn−1 + ch)
+(U W ⊗ Im )z(tn−1 , h) + O(hp+1 ), (21)
where y  (etn−1 + ch) = [y  (t
n−1 + ci h)]si=1
and y  (et
n−1 + ch) = [y  (t
n−1 +
ci h)]si=1 . We expand the components of the vectors y(etn−1 + ch), y  (etn−1 + ch)
and y  (etn−1 + ch) in Taylor series around tn−1 and use the order conditions (11)
appearing in Theorem 2; then, by comparing the coefficients of hp , we get
γ (tn−1 ) = (−(Cp − ACp−1 − ACp−2 − U qp ) ⊗ Im )y (p) (tn−1 ). (22)
Also, using (20) and Taylor series, we can write
[n] ) = hf (y(etn−1 + ch) + γ (tn−1 )hp + O(hp+1 ))
hf (Y
∂f
= hf (y(etn−1 + ch)) + (y(etn−1 + ch))γ (tn−1 )hp+1 + O(hp+2 )
∂y
∂f
= hy  (etn−1 + ch) + (y(etn−1 + ch))γ (tn−1 )hp+1 + O(hp+2 )
∂y
= (CK ⊗ Im )z(tn−1 , h) + (Cp ⊗ Im )hp+1 y (p+1) (tn−1 )
∂f
+ (y(etn−1 ))γ (tn−1 )hp+1 + O(hp+2 ), (23)
∂y
and
[n] ) =
h2 g(Y h2 g(y(etn−1 + ch) + γ (tn−1 )hp + O(hp+1 ))
= h2 g(y(etn−1 + ch)) + O(hp+2 )
= h2 y  (etn−1 + ch) + O(hp+2 )
= (CK 2 ⊗ Im )z(tn−1 , h) + (Cp−1 ⊗ Im )hp+1 y (p+1) (tn−1 ) + O(hp+2 ),
(24)
and also
z(tn , h) = (E ⊗ Im )z(tn−1 , h) + (Ep+1 ⊗ Im )hp+1 y (p+1) (tn−1 ) + O(hp+2 ). (25)
Numerical Algorithms (2022) 91:227–250 235

Now, substituting the formulae (23)–(25) into (15), we obtain

lte(tn ) = ((W E − BCK − BCK 2 − V W ) ⊗ Im )z(tn−1 , h)
+((W Ep+1 − BCp − BCp−1 ) ⊗ Im )hp+1 y (p+1) (tn−1 )
∂f
−(B ⊗ Im )hp+1 (y(tn−1 ))γ (tn−1 ). (26)
∂y
Then, using the order conditions (12) appearing in Theorem 2 together with (22),
gives
lte(tn ) = ((W Ep+1 − BCp − BCp−1 ) ⊗ Im )hp+1 y (p+1) (tn−1 )
+(B(Cp − ACp−1 − ACp−2 − U qp )
∂f
⊗Im )hp+1 (y(tn−1 ))y (p) (tn−1 ) + O(hp+2 ), (27)
∂y
which completes the proof.

Since in MESDBDF schemes with their representation as SGLMs given in previ-

ous section, the stage order q is one unit less than the order p of the overall method,
one can apply the results of Theorem 3 to get the error coefficients vectors ϕp and
ψp . It should be noted that the values of ψp,j , j = 2, 3, . . . , k are always zero for
MESDBDF schemes because of the corresponding rows in the matrix B are zero.
Moreover, by the special structure of the matrix V and the fact that only the first
rows of the matrices B and B are nonzero, we understand that the error coefficients
ϕp,j , j = 2, 3, . . . , k do not any affect in propagating the error to succeeding steps.
The error coefficients ψp,1 and ϕp,1 for MESDBDF schemes have been reported in
Table 1.

4 Generalized second derivative linear multistep methods

In this section, we introduce a class of methods by perturbing of the abscissa vector

c of the MESDBDF methods which have smaller error coefficients and larger angles

Table 1 Error coefficients ψp,1

and ϕp,1 of MESDBDF methods k p ψp,1 ϕp,1
with k = 1, 2, . . . , 9
1 3 0 − 0.1250
2 4 0.0019 0.0006
3 5 0.0011 0.1693
4 6 0.0007 0.7561
5 7 0.0004 2.7004
6 8 0.0003 8.7589
7 9 0.0002 26.9124
8 10 0.0001 79.9392
9 11 0.0001 232.1533
236 Numerical Algorithms (2022) 91:227–250

α of A(α)-stability. The methods will be referred to as generalized second deriva-

tive linear multistep methods (GSLMMs). To do this, we consider SGLMs with the
coefficients
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
λ 0 0 μ 0 0 u11 u12 · · · u1k
A = ⎣ a21 λ 0 ⎦ , A = ⎣ a 21 μ 0 ⎦ , U = ⎣ u21 u22 · · · u2k ⎦ ,
a31 a32 λ a 31 a 32 μ u31 u32 · · · u3k
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
a31 a32 λ a 31 a 32 μ u31 u32 . . . u3,k−1 u3k
⎢ 0 0 0⎥ ⎢ 0 0 0⎥ ⎢ 1 0 ... 0 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ .. .. .. ⎥ ⎢ .. . . ⎥ ⎢ . . . . .. ⎥ ,
B = ⎢ . . . ⎥, B = ⎢ . .. .. ⎥ , V = ⎢ .. .. . . .. . ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ 0 0 0⎦ ⎣ 0 0 0⎦ ⎣ 0 0 ... 0 0 ⎦
0 0 0 0 0 0 0 0 ... 1 0
(28)
and the external stages
y [n] = [yn+k
T T
yn+k−1 T
. . . yn+1 ]T ,
as an approximation of order p = k + 2 to the vector
[y(tn+k )T y(tn+k−1 )T . . . y(tn+1 )T ]T ,
which is the same as for MESDBDF methods. Also, we assume that the internal
stages Yj[n] are approximations to y(tn−1 + cj h) of order k + 1 for j = 1, 2, and k + 2
for j = 3 with the abscissa vector
c = [k + 1 + η1 k + 1 + η2 k + 1]T , (29)
where η1 and η2 are some constants. After satisfying the order conditions, η1 and η2
together with one of the coefficients a31 , a32 , a 31 , and a 32 are free parameters. In our
construction, we set a32 = 0 which corresponds to MESDBDF in the case η1 = 0
and η2 = 1. Indeed, this class of SGLMs, referred to as GSLMMs, is a perturbation
of MESDBDF schemes in which the free parameters η1 and η2 are used to achieve
methods with desirable stability and accuracy properties.
The stability function of GSLMMs takes the form

1 
k
p(w, z) = aj (z)w j , (30)
(1 − λz − μz2 )3
j =0

where aj (z), j = 0, 1, . . . , k, are polynomials of degree at most six in z, whose

coefficients depend on the two parameters η1 and η2 . Since one of our aims in our
construction is deriving methods with larger α of A(α)-stability of the methods, we
are going to define the objective function
 π
fn : (η1 , η2 ) → 0, ,
2
approximating the value of the angle α for special choices of parameters η1 and η2 .
To do this, using boundary locus method [29], we compute n points z = z(ϑν ), for
ϑν ∈ [0, 2π ] given by
2π ν
ϑν = , ν = 1, . . . , n,
n
Numerical Algorithms (2022) 91:227–250 237

on the boundary of the stability region of the methods as the solutions of the equations

k
aj (z)eij ϑν = 0, ν = 1, 2, . . . , n,
j =0

with i as the imaginary unit. Then, introducing the set


n π 
S= Iν ∪ ,
2
ν=1

with
⎧ ⎫
⎨   ⎬
 I m(z)  
k
Iν = arctan   : Re(z) < 0, a (z)e ij ϑν
= 0 , ν = 1, . . . , n,
Re(z) 
j
⎩ ⎭
j =0

the objective function fn is defined as fn (η1 , η2 ) = − min(S). Moreover, our other

aim is deriving methods with smaller error constants, so we define the vector
ecp (η1 , η2 ) = [ψp,1 , ϕp,1 ]T .
in which its entries can be computed by the formulas (18) and (19) in terms of
η1 and η2 . Searching for an appropriate balance between the angle α of A(α)-
stability and error coefficients, we minimize the objective function fn (for a sufficient
large n) by using the fmincon command of MATLAB, subject to the constraint
ecp (η1 , η2 )1 ≤ e, for some values of e, e.g., e = 10−1 ecp (0, 1)1 . In the next
section, we describe the construction of GSLMMs of orders p = 3, 4, . . . , 11.

5 Examples of GSLMMs

In this section, we present the coefficients for the constructed GSLMMs of orders
p = 3, 4, . . . , 11. Here, we only give the optimum values for η1 and η2 , and the
coefficients matrices A, A, and U ; we refrain from giving the matrices B, B and
V and the abscissa vector c which are in the form (28) and (29). For the methods
of orders p = 3, 4, 5, 6, taking into account the original MESDBDF methods is A-
stable, we search for the pairs of (η1 , η2 ) for which the derived GSLMMs is also
A-stable with a value of ecp (η1 , η2 )1 very close to its minimum.

5.1 Coefficients of GSLMM with k = 1 and p = 3

388 262
η1 = − , η2 = − ,
⎡ 593 1113 ⎤
0.345699831365936 0 0
A = ⎣ 0.418900348328583 0.345699831365936 0 ⎦,
0.654300168634064 0 0.345699831365936
238 Numerical Algorithms (2022) 91:227–250
⎡ ⎤
−0.0597541867032183 0 0
A = ⎣ −0.0570750288614188 −0.0597541867032183 0 ⎦,
−0.0567051189885751 0.0445680163663581 −0.0597541867032183
⎡ ⎤
1
U = ⎣1⎦.
1

This method is A-stable with

ec3 (η1 , η2 )1 ≈ 0.0047 ≤ ec3 (0, 1)1 ≈ 0.1250.

5.2 Coefficients of GSLMM with k = 2 and p = 4

249 1087
η1 = , η2 = ,
5000
⎡ 1066 ⎤
0.894620349779437 0 0
A = ⎣ 0.943766813757910 0.894620349779437 0 ⎦,
0.000013507606583 0 0.894620349779437
⎡ ⎤
−0.310542598088171 0 0
A = ⎣ −0.356838365106060 −0.310542598088171 0 ⎦,
−0.0781184735770454 0.046709612907379 −0.310542598088171
⎡ ⎤
1.155179650220563 −0.1551796502205635
U = ⎣ 1.181312648845392 −0.1813126488453925 ⎦ .
1.105366142613980 −0.1053661426139802

This method is A-stable with

ec4 (η1 , η2 )1 ≈ 0.000081 ≤ ec4 (0, 1)1 ≈ 0.0025.

5.3 Coefficients of GSLMM with k = 3 and p = 5

559 826
η1 = , η2 = − ,
1250
⎡ 1013 ⎤
1.04493723752626 0 0
A = ⎣ −0.09792722027844 1.04493723752626 0 ⎦,
−0.04304115422279 0 1.04493723752626
⎡ ⎤
−0.375955066021513 0 0
A = ⎣ 0.0992465090613176 −0.375955066021513 0 ⎦,
0.0733392030974382 −0.187046389359499 −0.375955066021513
⎡ ⎤
1.51430995958412 −0.626357156694508 0.112047197110383
U = ⎣ 0.09895709651888 1.03967598714780 −0.138633083666673 ⎦ .
1.00417004819333 −0.01023617969013 0.006066131496801
Numerical Algorithms (2022) 91:227–250 239

This method is A-stable with

ec5 (η1 , η2 )1 ≈ 0.0033 ≤ ec5 (0, 1)1 ≈ 0.1705.

5.4 Coefficients of GSLMM with k = 4 and p = 6

729 739
η1 = , η2 = − ,
994
⎡ 658 ⎤
1.09175537776366 0 0
A = ⎣ 0.0371625192846510 1.09175537776366 0 ⎦,
−0.0371433717354413 0 1.09175537776366
⎡ ⎤
−0.383917386720813 0 0
A = ⎣ −0.0172342496730442 −0.383917386720813 0 ⎦,
0.0460856740122321 −0.272839434295113 −0.383917386720813
⎡ ⎤
2.05731033248579 −0.419754094936734 1.01351493633442
⎢ −1.57401178271781 1.60792315408183 −0.0937747193205140 ⎥
UT = ⎢ ⎥
⎣ 0.617737592629104 −0.208602224353128 0.0923926236095538 ⎦ .
−0.101036142397075 0.0204331652080347 −0.0121328406234596
This method is A-stable with
ec6 (η1 , η2 )1 ≈ 0.0034 ≤ ec6 (0, 1)1 ≈ 0.7568.

5.5 Coefficients of GSLMM with k = 5 and p = 7

223 3689
η1 = , η2 = − ,
2500
⎡ 2094 ⎤
0.726882299008286 0 0
A = ⎣ 0.050932833410381 0.726882299008286 0 ⎦,
−0.073087869282132 0 0.726882299008286
⎡ ⎤
−0.168142504383280 0 0
A = ⎣ −0.024811668011971 −0.168142504383280 0 ⎦,
0.038580253381901 0.529256740672978 −0.168142504383280
⎡ ⎤
1.586335028857337 −0.2647399059918332 0.9094349903363600
⎢ −0.8991616013627598 0.9156607214238014 0.5974111911303670 ⎥
⎢ ⎥
U T =⎢
⎢ 0.4171872087187099 0.4343919385010570 −0.5840935009085844⎥ .
⎥
⎣ −0.1199120277867760 −0.0963216452358359 0.0844190373544743 ⎦
0.0155513915734884 0.0110088913028108 −0.0071717179126169
This method is A-stable with
ec7 (η1 , η2 )1 ≈ 0.0014 ≤ ec7 (0, 1)1 ≈ 2.7009.
240 Numerical Algorithms (2022) 91:227–250

5.6 Coefficients of GSLMM with k = 6 and p = 8

795 2653
η1 = , η2 = ,
898
⎡ 369 ⎤
1.00376386286450 0 0
A = ⎣ −383.843160941161 1.00376386286450 0 ⎦,
−0.090298614832670 0 1.00376386286450
⎡ ⎤
−0.303392797830922 0 0
A = ⎣ 368.189424432527 −0.303392797830922 0 ⎦,
0.062283265416413 −0.000019812804409 −0.303392797830922
⎡ ⎤
3.15712186021681 2106.26512124348 0.947076133927057
⎢ −4.40382431527382 −5713.03643169425 0.329989469072597 ⎥
⎢ ⎥
⎢ 3.60710937581011 6435.29315894129 −0.471047631906984 ⎥
U =⎢
T
⎢ −1.81450979299187
⎥.
⎢ −3939.59233578207 0.259456647103144 ⎥⎥
⎣ 0.518916223851599 1290.15441427517 −0.074574192442845 ⎦
−0.064813351612829 −178.083926983629 0.009099574247030
This method is A-stable with
ec8 (η1 , η2 )1 ≈ 0.0081 ≤ ec8 (0, 1)1 ≈ 8.7592.

5.7 Coefficients of GSLMM with k = 7 and p = 9

599 894
η1 = , η2 = − ,
998
⎡ 2107 ⎤
0.854665353891539 0 0
A = ⎣ −0.044094794206219 0.854665353891539 0 ⎦,
0.133545384315309 0 0.854665353891539
⎡ ⎤
−0.217328023674293 0 0
A = ⎣ 0.025768746703558 −0.217328023674293 0 ⎦,
−0.0250865040464106 −0.311199417769920 −0.217328023674293
⎡ ⎤
2.85207185421918 0.373589765611886 0.993168943541146
⎢ −3.98854388676492 1.37105727258033 0.040557875205721 ⎥
⎢ ⎥
⎢ 3.75659882783086 −1.29483689299705 −0.055811465556413 ⎥
⎢ ⎥
UT = ⎢
⎢ −2.40095117561742 0.809769808136031 0.030917257293390 ⎥ .
⎥
⎢ 0.998077215422595 −0.330690921784529 −0.010905729761232 ⎥
⎢ ⎥
⎣ −0.243872859896778 0.079724522351708 0.002291774385616 ⎦
0.026620024806484 −0.008613553898379 −0.000218655108228
This method is A(89.01◦ )-stable with
ec9 (η1 , η2 )1 ≈ 0.0059 ≤ ec9 (0, 1)1 ≈ 26.9126.
Numerical Algorithms (2022) 91:227–250 241

In Fig. 1, the region of absolute stability of the GSLMM of order p = 9 has been
plotted and compared with that for MESDBDF of the same order.

5.8 Coefficients of GSLMM with k = 8 and p = 10

1021 999
η1 = , η2 = − ,
⎡612 1000 ⎤
1.13103276675478 0 0
A = ⎣ 0.008473424615545 1.13103276675478 0 ⎦,
0.016749071756286 0 1.13103276675478
⎡ ⎤
−0.367894383836064 0 0
A = ⎣ −0.004729018694588 −0.367894383836064 0 ⎦,
−0.005500558260371 −0.497000817990068 −0.367894383836064
⎡ ⎤
8.52958947240140 −0.385211585082337 1.09236462944105
⎢ −22.9792413935374 1.59917126544019 −0.524048993916518 ⎥
⎢ ⎥
⎢ 33.4483929738593 −0.065341971394488 0.724479797961413 ⎥
⎢ ⎥
⎢ −31.1416649735185 −0.344046741390260 −0.430189094080153 ⎥
U =⎢
T ⎢ ⎥.
⎥
⎢ 19.1273561024877 0.301346895947004 0.181355616070281 ⎥
⎢ −7.54665185579959 −0.135873239011399 −0.052563088963023 ⎥
⎢ ⎥
⎣ 1.74152037016807 0.033549924280162 0.009376435129413 ⎦
−0.179300696060988 −0.003594548788872 −0.000775301642467
This method is A(85.76◦ )-stable with
ec10 (η1 , η2 )1 ≈ 0.0097 ≤ ec10 (0, 1)1 ≈ 79.9393.
In Fig. 2, the region of absolute stability of the GSLMM of order p = 10 has been
plotted and compared with that for MESDBDF of the same order.

4.5

3.5

3
Im(z)

2.5

1.5

0.5

-1 0 1 2 3 4 5 6 7
Re(z)

Fig. 1 Regions of absolute stability of GSLMM (solid line) and MESDBDF (dashed line) for k = 7
242 Numerical Algorithms (2022) 91:227–250

6
5.5
5
4.5
4
3.5
Im(z)

3
2.5
2
1.5
1
0.5

-1 0 1 2 3 4 5 6
Re(z)

Fig. 2 Regions of absolute stability of GSLMM (solid line) and MESDBDF (dashed line) for k = 8

5.9 Coefficients of GSLMM with k = 9 and p = 11

2447 1733
η1 = , η2 = − ,
⎡1542 2000 ⎤
1.058056544651493 0 0
A = ⎣ 0.0058915880426507 1.058056544651493 0 ⎦,
0.0156549147344995 0 1.058056544651493
⎡ ⎤
−0.3184611259077115 0 0
A = ⎣ −0.0028358925963986 −0.3184611259077115 0 ⎦,
−0.0050152548558462 −0.3709175051135338 −0.3184611259077115
⎡ ⎤
9.45028547028813 −0.406201798902460 1.05220117801065
⎢ −28.3606413769783 2.16183816440191 −0.290273764690985 ⎥
⎢ ⎥
⎢ 47.4661594414867 −1.16181684318959 0.416868129988613 ⎥
⎢ ⎥
⎢ −52.5362679459510 0.566078898898886 −0.273655690753791 ⎥
⎢ ⎥
UT = ⎢
⎢ 40.0717878368428 −0.195076404035980 0.131216353347548 ⎥ .
⎥
⎢ −20.9787129560454 0.035035822214096 −0.045914123193849 ⎥
⎢ ⎥
⎢ 7.23490297448344 0.002219182972755 0.011098976291083 ⎥
⎢ ⎥
⎣ −1.48538588577429 −0.002434544088596 −0.001655853187798 ⎦
0.137872441647813 0.000357521728979 0.000114794188534
This method is A(80.72◦ )-stable with
ec10 (η1 , η2 )1 ≈ 0.0053 ≤ ec10 (0, 1)1 ≈ 232.1534.
In Fig. 3, the region of absolute stability of the GSLMM of order p = 11 has been
plotted and compared with that for MESDBDF of the same order.
For k = 1, 2, . . . , 6, the resulting methods are A-stable and for k = 7, 8, 9,
the methods are A(α)-stable with larger angles α in comparison with those for
Numerical Algorithms (2022) 91:227–250 243

6
5.5
5
4.5
4
3.5
Im(z)

3
2.5
2
1.5
1
0.5

-1 0 1 2 3 4 5 6 7
Re(z)

Fig. 3 Regions of absolute stability of GSLMM (solid line) and MESDBDF (dashed line) for k = 9

MESDBDF. Furthermore, in all cases, GSLMMs have noticeably smaller error con-
stants than MESDBDFs. These results are reported in Table 2. For the sake of
comparison, the values of angle α of A(α)-stability of GSLMMs together with those
of MESDBDFs, SDBDFs, and GLMMs [26] are reported in Table 3.

6 Numerical experiments

In this section, we give some numerical experiments explaining the improvement

in the accuracy and stability of GSLMMs compared with MESDBDF methods.

Table 2 Error coefficients and angles of A(α)-stability for GSLMMs described in Section 4

GSLMM MESDBDF

k ψp,1 ϕp,1 η1 η2 ecp (η1 , η2 )1 α ecp (0, 1)1 α

388 262
1 0.0045 − 0.00022 − − 0.0047 90◦ 0.1250 90◦
593 1113
249 1087
2 10−7 0.000081 0.000081 90◦ 0.0025 90◦
5000 1066
559 826
3 − 0.0033 0.000023 − 0.0033 90◦ 0.1705 90◦
1250 1013
729 739
4 − 0.00332 − 0.00008 − 0.0034 90◦ 0.7568 90◦
994 658
223 3689
5 − 0.00071 0.00064 − 0.0014 90◦ 2.7009 89.86◦
2500 2094
795 2653
6 − 0.0057 − 0.0024 0.0081 90◦ 8.7592 88.49◦
898 369
599 894
7 0.0031 0.0028 − 0.0059 89.01◦ 26.9126 85.43◦
998 2107
1021 999
8 0.0038 − 0.0059 − 0.0097 85.76◦ 79.9393 81.81◦
612 1000
2447 1733
9 0.0026 0.0027 − 0.0053 80.72◦ 232.1534 76.34◦
1542 2000
244 Numerical Algorithms (2022) 91:227–250

Table 3 Angles α of A(α)-stability for GSLMMs, MESDBDFs, SDBDFs, and GLMMs for k =
1, 2, . . . , 8

GSLMM MESDBDF SDBDF GLMM

k p α p α p α p α

1 3 90◦ 3 90◦ 2 90◦ 2 90◦

2 4 90◦ 4 90◦ 3 90◦ 3 90◦
3 5 90◦ 5 90◦ 4 90◦ 4 90◦
4 6 90◦ 6 90◦ 5 89.36◦ 5 88.25◦
5 7 90◦ 7 89.86◦ 6 86.35◦ 6 83.41◦
6 8 90◦ 8 88.49◦ 7 80.82◦ 7 76.46◦
7 9 89.01◦ 9 85.43◦ 8 72.53◦ 8 67.23◦
8 10 85.76◦ 10 81.81◦ 9 60.71◦ 9 55.13◦

Moreover, the results of the proposed methods are compared with those of GLMMs
derived in [26].
To verify the theoretical stability improvements, we consider the linear stiff system
 
y1 = −αy1 − βy2 + (α + β − 1)e−t , y1 (0) = 1,
(31)
y2 = βy1 − αy2 + (α − β − 1)e−t , y2 (0) = 0,

with the exact solution y1 (t) = y2 (t) = e−t . In our numerical experiments, we
consider α = 10, β = 290. Since eigenvalues of the Jacobian of the system are
−α ± iβ, in implementing of the methods, the stepsize h must be chosen such that
the points z := −hα ± ihβ lie inside the absolute stability region of the methods.
As it has been presented in Fig. 4, these points lie outside (or inside) of the stability
region of MESDBDFs and GLMMs with k = 7 and k = 8 for the stepsize h = 0.01
(or h = 0.005). These points, however, lie inside the absolute stability region of
GSLMMs for both stepsizes. As it is expected, MESDBDFs and GLMMs diverge for
the stepsize h = 0.01 while this is not the case for GSLMMs. These phenomenons
have been illustrated in Figs. 5 and 6.
To show the theoretical accuracy improvements, we consider the HIRES problem
[23, 30]
⎧ 
⎪
⎪ y1 = −1.71y1 + 0.43y2 + 8.32y3 + 0.0007, y1 (0) = 1,
⎪
⎪
⎪
⎪

y2 = 1.71y1 − 8.75y2 , y2 (0) = 0,
⎪
⎪
⎪
⎪ 
y3 = −10.03y3 + 0.43y4 + 0.035y5 , y3 (0) = 0,
⎪
⎪
⎪
⎨ y  = 8.32y + 1.71y − 1.12y ,
4 2 3 4 y 4 (0) = 0,
 (32)
⎪
⎪ y5 = −1.745y2 + 0.43(y6 + y7 ), y5 (0) = 0,
⎪
⎪
⎪
⎪ y6 = −280y6 y8 + 0.69y4 + 1.71y5 − 0.43y6 + 0.69y7 , y6 (0) = 0,
⎪
⎪
⎪
⎪
⎪
⎪ y7 = 280y6 y8 − 1.81y7 , y7 (0) = 0,
⎪
⎩ 
y8 = −280y6 y8 + 1.81y7 , y8 (0) = 0.0057,
Numerical Algorithms (2022) 91:227–250 245

3.5

2.5

2
Im(z)

1.5 MESDBDF, k = 7
GSLMM, k = 7
MESDBDF, k = 8
1 GSLMM, k = 8
GLMM, k = 7
0.5 GLMM, k = 8

0
-2 -1.8 -1.6 -1.4 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2
Re(z)

Fig. 4 Boundaries of the stability regions near the origin of GSLMMs, MESDBDFs, and GLMMs with
k = 7 and k = 8 together with the points z = −10h + 290hi for h = 0.01 and h = 0.005

with t ∈ [0, 321.8122]. The errors of MESDBDF and GSLMMs for k = 1 (as
A-stable methods) and k = 7 (as A(α)-stable methods) applied to the HIRES prob-
lem for various values of N = 210+i , i = 1, 2, 3, with Nh = 321.8122 have
been represented in Table 4. To compute the error of the methods, we use the refer-
ence solution obtained by the MATLAB function ode15s with very tight tolerances

100
1010
MESDBDF, k = 7 MESDBDF, k = 7
GSLMM, k = 7 GSLMM, k = 7
GLMM, k = 7 GLMM, k = 7
105

10-5

100
Error

Error

10-5
10-10

10-10

10-15 10-15
0 10 20 30 0 10 20 30
t t

Fig. 5 Errors versus t for GSLMMs, MESDBDFs, and GLMMs with k = 7 applied to problem (31) for
the fixed stepsizes h = 0.01 (left) and h = 0.005 (right). Every fiftieth (left) and hundredth (right) point
is plotted
246 Numerical Algorithms (2022) 91:227–250

10150 100

10100

10-5
Error

Error
1050

10-10
100

MESDBDF, k = 8 MESDBDF, k = 8
GSLMM, k = 8 GSLMM, k = 8
GLMM, k = 8 GLMM, k = 8
10-50 10-15
0 10 20 30 0 10 20 30
t t

Fig. 6 Errors versus t for GSLMMs, MESDBDFs, and GLMMs with k = 8 applied to problem (31) for
the fixed stepsizes h = 0.01 (left) and h = 0.005 (right). Every fiftieth (left) and hundredth (right) point
is plotted

atol = rtol = 10−14 . In this table, “Ratio” stands for the ratio of the error of MES-
BDF methods to that of GSLMMs. This quantity shows how much better the latter is
than the former in the accuracy point of view.
To show capability of the proposed methods in solving stiff problems in higher
dimensions, we consider the CUSP problem [23]
⎧ σ
⎪
⎪ yi = −ε−1 (yi3 + ai yi + bi ) + (yi−1 − 2yi + yi+1 ),
⎪
⎪ (Δx)2
⎪
⎨ σ
ai = bi + 0.07vi + (ai−1 − 2ai + ai+1 ), (33)
⎪
⎪ (Δx)2
⎪
⎪ σ
⎪
⎩ bi = (1 − ai2 )bi − ai − 0.4yi + 0.035vi + (bi−1 − 2bi + bi+1 ),
(Δx)2

Table 4 Numerical results of MESDBDF and GSLMM for k = 1 and k = 7 applied to problem (32)

N 211 212 213

k=1 MESDBDF 2.27 × 10−05 2.80 × 10−06 3.08 × 10−07

GSLMM 1.52 × 10−07 1.69 × 10−08 2.02 × 10−09
Ratio 1.49 × 102 1.66 × 102 1.52 × 102

k=7 MESDBDF 3.30 × 10−07 7.28 × 10−09 4.25 × 10−11

GSLMM 3.56 × 10−09 2.88 × 10−09 3.42 × 10−12
Ratio 9.27 × 101 2.53 × 100 1.24 × 101
Numerical Algorithms (2022) 91:227–250 247

for i = 1, 2, . . . , N and
ui
vi = , ui = (yi − 0.7)(yi − 1.3),
ui + 0.1

10-1 GSLMM, k = 1
MESDBDF, k = 1
GLMM, k = 2
10-2

10-3
Error

10-4

10-5

10-6
1 1.5 2 2.5 3 3.5 4
nfe 104

10-5
GSLMM, k = 7
MESDBDF, k = 7
GLMM, k = 8
Error

10-6

10-7
2 2.5 3 3.5 4 4.5
nfe 104

Fig. 7 Errors versus nfe for GSLMMs, MESDBDFs, and GLMMs with k = 1 (top) and k = 7 (bottom)
applied to problem (33)
248 Numerical Algorithms (2022) 91:227–250

resulting from the discretization of the diffusion terms in the PDEs

⎧
⎪
⎪ ∂y 1 3 ∂ 2y
⎪
⎪ = − (y + ay + b) + σ ,
⎪
⎪ ∂t ε ∂x 2
⎪
⎨
∂a ∂ 2a
⎪ = b + 0.07v + σ ,
⎪
⎪ ∂t ∂x 2
⎪
⎪
⎪
⎪
2
⎩ ∂b = (1 − a 2 )b − a − 0.4y + 0.035v + σ ∂ b ,
∂t ∂x 2
with 0 ≤ x ≤ 1 and
u
v= , u = (y − 0.7)(y − 1.3),
u + 0.1
by the method of lines. We take σ = 1/144, ε = 10−3 , N = 32, the initial values
   
2iπ 2iπ
yi (0) = 0, ai (0) = −2 cos , bi (0) = 2 sin , i = 1, 2, . . . , N ,
N N
and periodic boundary conditions
y0 = yN , a0 = aN , b0 = bN ,

yN+1 = y1 , aN+1 = a1 , bN +1 = b1 .
with tout = 1.1. In our implementation, the second derivative function is computed as
g(·) = fy (·)f (·); also we use (fy (·))2 as a piecewise constant approximation to the
Jacobian gy (·). Indeed, we implement the second derivative methods without addi-
tional computational cost. Errors versus the number of function evaluations, nfe, for
GSLMMs, MESDBDFs, and GLMMs with k = 1 and k = 7 applied to this prob-
lem have been represented in Fig. 7. These results illustrate that GSLMMs are more
cost-effective than two other mentioned methods.

7 Conclusion

In this paper, we investigated MESDBDFs as SGLMs with q = p − 1 by perturb-

ing the abscissas of the internal stages. This idea makes it possible to derive a new
class of methods, referred to as GSLMMs. By discussing the local truncation error of
SGLMs for the case q = p−1, we found formulae for error coefficients of GSLMMs
and constructed methods with better stability and accuracy properties compared with
MESDBDF. The future work contains the implementation of the proposed methods
in a variable-stepsize/variable-order environment.

Data availability Data sharing not applicable to this article as no datasets were generated or analyzed
during the current study.

Declarations

Conflict of interest The authors declare no competing interests.

Numerical Algorithms (2022) 91:227–250 249

References

1. Abdi, A.: Construction of high-order quadratically stable second-derivative general linear methods for
the numerical integration of stiff ODEs. J. Comput. Appl. Math. 303, 218–228 (2016)
2. Abdi, A., Behzad, B.: Efficient Nordsieck second derivative general linear methods: construction and
implementation. Calcolo 55(28), 1–16 (2018)
3. Abdi, A., Braś, M., Hojjati, G.: On the construction of second derivative diagonally implicit multistage
integration methods for ODEs. Appl. Numer. Math. 76, 1–18 (2014)
4. Abdi, A., Conte, D.: Implementation of second derivative general linear methods. Calcolo 57, 1–29
(2020)
5. Abdi, A., Hojjati, G.: An extension of general linear methods. Numer. Algorithms 57, 149–167 (2011)
6. Abdi, A., Hojjati, G.: Implementation of Nordsieck second derivative methods for stiff ODEs. Appl.
Numer. Math. 94, 241–253 (2015)
7. Abdi, A., Hojjati, G.: Maximal order for second derivative general linear methods with Runge–Kutta
stability. Appl. Numer. Math. 61, 1046–1058 (2011)
8. Abdi, A., Jackiewicz, Z.: Towards a code for nonstiff differential systems based on general linear
methods with inherent Runge–Kutta stability. Appl. Numer. Math. 136, 103–121 (2019)
9. Butcher, J.C.: On the convergence of numerical solutions to ordinary differential equations. Math.
Comp. 20, 1–10 (1966)
10. Butcher, J.C., Chartier, P., Jackiewicz, Z.: Experiments with a variable-order type 1 DIMSIM code.
Numer. Algorithms 22, 237–261 (1999)
11. Butcher, J.C., Hojjati, G.: Second derivative methods with RK stability. Numer. Algorithms 40, 415–
429 (2005)
12. Butcher, J.C., Jackiewicz, Z.: Construction of diagonally implicit general linear methods of type 1 and
2 for ordinary differential equations. Appl. Numer. Math. 21, 385–415 (1996)
13. Butcher, J.C., Jackiewicz, Z.: Construction of high order DIMSIMs for ordinary differential equations.
Appl. Numer. Math. 27, 1–12 (1998)
14. Butcher, J.C., Jackiewicz, Z.: Diagonally implicit general linear methods for ordinary differential
equations. BIT 33, 452–472 (1993)
15. Butcher, J.C., Jackiewicz, Z.: Implementation of diagonally implicit multistage integration methods
for ordinary differential equations. SIAM J. Numer. Anal. 34, 2119–2141 (1997)
16. Cash, J.R.: On the integration of stiff systems of ODEs using extended backward differentiation
formulae. Numer. Math. 34, 235–246 (1980)
17. Cash, J.R.: The integration of stiff initial value problems in ODEs using modified extended backward
differentiation formulae. Comput. Math. App. 9, 645–657 (1983)
18. Chan, R.P.K., Tsai, A.Y.J.: On explicit two-derivative Runge–Kutta methods. Numer. Algorithms 53,
171–194 (2010)
19. Dahlquist, G.: A special stability problem for linear multistep methods. BIT 3, 27–43 (1963)
20. D’Ambrosio, R., Izzo, G., Jackiewicz, Z.: Perturbed MEBDF methods. Comput. Math. Appl. 63,
851–861 (2012)
21. Enright, W.H.: Second derivative multistep methods for stiff ordinary differential equations. SIAM J.
Numer. Anal. 11, 321–331 (1974)
22. Ezzeddine, A.K., Hojjati, G., Abdi, A.: Perturbed second derivative multistep methods. J. Numer.
Math. 23, 235–245 (2015)
23. Hairer, E., Wanner, G.: Solving Ordinary Differential Equations II: Stiff and Differential-Algebraic
Problems. Springer, Berlin (2010)
24. Hojjati, G., Rahimi Ardabili, M.Y., Hosseini, S.M.: New second derivative multistep methods for stiff
systems. Appl. Math. Model. 30, 466–476 (2006)
25. Hosseini, S.M., Hojjati, G.: Matrix free MEBDF method for the solution of stiff systems of ODEs.
Math. Comput. Modell. 29, 67–77 (1999)
26. Izzo, G., Jackiewicz, Z.: Generalized linear multistep methods for ordinary differential equations.
Appl. Numer. Math. 114, 165–178 (2017)
27. Jackiewicz, Z.: General Linear Methods for Ordinary Differential Equations. Wiley, New Jersey
(2009)
28. Jackiewicz, Z.: Implementation of DIMSIMs for stiff differential systems. Appl. Numer. Math. 42,
251–267 (2002)
250 Numerical Algorithms (2022) 91:227–250

29. Lambert, J.D.: Numerical Methods for Ordinary Differential Systems. Wiley, New York (1991)
30. Mazzia, F., Iavernaro, F., Magherini, C.: Test set for initial value problem solvers. University of Bari
(2003)

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published
maps and institutional affiliations.