Computational Mechanics (2023) 72:1239–1261

https://doi.org/10.1007/s00466-023-02347-2

REVIEW PAPER

Derivation of third order Runge–Kutta methods (ELDIRK) by

embedding of lower order implicit time integration schemes for local
and global error estimation
Rolf Mahnken1

Received: 7 September 2022 / Accepted: 15 April 2023 / Published online: 17 June 2023
© The Author(s) 2023

Abstract
Three prominent low order implicit time integration schemes are the first order implicit Euler-method, the second order
trapezoidal rule and the second order Ellsiepen method. Its advantages are stability and comparatively low computational
cost, however, they require the solution of a nonlinear system of equations. This paper presents a general approach for the
construction of third order Runge–Kutta methods by embedding the above mentioned implicit schemes into the class of
ELDIRK-methods. These will be defined to have an Explicit Last stage in the general Butcher array of Diagonal Implicit
Runge–Kutta (DIRK) methods, with the consequence, that no additional system of equations must be solved. The main
results—valid also for non-linear ordinary differential equations—are as follows: Two extra function calculations are required
in order to embed the implicit Euler-method and one extra function calculation is required for the trapezoidal-rule and the
Ellsiepen method, in order to obtain the third order properties, respectively. Two numerical examples are concerned with a
parachute with viscous damping and a two-dimensional laser beam simulation. Here, we verify the higher order convergence
behaviours of the proposed new ELDIRK-methods, and its successful performances for asymptotically exact global error
estimation of so-called reversed embedded RK-method are shown.

Keywords Embedded Runge–Kutta methods · ELDIRK-methods · Implicit Euler-method · Trapezoidal-rule · Ellsiepen

method · Local error estimation · Global error estimation · Effectivity index · Laser beam simulation

Contents 5.1 Global error and effectivity index for standard embedded
RK(r)q . . . . . . . . . . . . . . . . . . . . . . . . . . . 1248
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 1240 5.2 Global error and effectivity index for reversed embedded
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . 1241 RK(q)r . . . . . . . . . . . . . . . . . . . . . . . . . . . 1249
2.1 Generalized Runge–Kutta scheme . . . . . . . . . . . . 1241 6 Representative numerical examples . . . . . . . . . . . . . . 1250
2.2 The class of ELDIRK-methods . . . . . . . . . . . . . . 1242
2.3 Order and stage conditions . . . . . . . . . . . . . . . . 1242 6.1 Parachute with viscous damping . . . . . . . . . . . . . 1250
2.4 Embedded Runge–Kutta methods . . . . . . . . . . . . . 1243 6.1.1 Initial value problem (IVP) . . . . . . . . . . . . . 1250
3 Second order ELDIRK for embedding of RK1 . . . . . . . . 1244 6.1.2 Global convergence investigations for equidistant
3.1 General setting . . . . . . . . . . . . . . . . . . . . . . 1244 step sizes . . . . . . . . . . . . . . . . . . . . . . 1250
3.2 Embedding of the explicit and implicit Euler method . . 1244
3.3 Evaluation of the new second order Butcher-Tableaus . . 1245 6.1.3 Adaptively refined step sizes . . . . . . . . . . . . 1252
4 Third order ELDIRK for embedding of RK2 . . . . . . . . . 1245 6.2 Laser beam simulation on a plate . . . . . . . . . . . . . 1255
4.1 General setting . . . . . . . . . . . . . . . . . . . . . . 1245 6.2.1 Initial boundary value problem . . . . . . . . . . . 1255
4.2 Embedding of the modified implicit Euler method . . . . 1246
4.3 Evaluation of the new third order Butcher-Tableaus . . . 1246 6.2.2 Global convergence investigations for equidistant
5 Revision of goal oriented global error measures applied to step sizes . . . . . . . . . . . . . . . . . . . . . . 1257
RK(r)q and RK(q)r . . . . . . . . . . . . . . . . . . . . . . . 1247 7 Summary and outlook . . . . . . . . . . . . . . . . . . . . . 1257
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1258
B Rolf Mahnken Appendix A: Derivation of higher order ELDIRK-schemes . . . 1258
[email protected] A.1. Embedding of the trapezoidal method . . . . . . . . . . 1258
1 Chair of Engineering Mechanics, University of Paderborn, A.2. Embedding of the Ellsiepen method . . . . . . . . . . . 1260
Warburger Str. 100, 33098 Paderborn, Germany References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1260

123
1240 Computational Mechanics (2023) 72:1239–1261

1 Introduction array

The mathematical modeling of several engineering prob- α α

lems requires the numerical solution of ordinary differential 1 1−α α 1√
equations. Typical examples for a first order ordinary differ- B(2)1 = , α=1− 2,
1−α α 2
ential equation are the resulting heat equation discretized
in space or inelastic constitutive equations, whilst struc- 1−β β
tural dynamic problems lead to examples of second order 5√
β=2− 2.
ordinary differential equations, see e.g. [16, 33]. Choos- 4
ing the appropriate integration method requires deciding on (1)
aspects such as the order of the integration scheme, sta-
bility properties and computational efficiency. A general This array defines an embedded RK-method, subsequently
class of algorithms is provided by so-called S-stage Runge– denoted as RK(2)1-Ell, combining a method RK2-Ell (using
Kutta methods (RK-methods), where explicit and implicit α below the horizontal line) of order 2 with embedded method
schemes are distinguished, [5, 10, 19]. Due to the prop- RK1-Ell of order 1 (using β) [7]. Both methods have the same
erties of A-stability, where the region of absolute stability stage and, different to Fehlbergs explicit schemes, both meth-
is unbounded, see e.g. [19], and relatively low computa- ods are implicit. Exploiting both choices below the horizontal
tional cost, the implicit Euler-method and the generalized line has the consequence, that even for nonlinear problems
trapezoidal-rule are most prominent integration schemes in no additional functional evaluation becomes necessary for
structural mechanics, see e.g. [16, 33]. Also the second local error estimation.
order Ellsiepen method [7], with the additional property of This paper intends to combine the advantage of Fehlsbergs
L-stability, proved to be very efficient for the laser beam embedding approach with only one additional functional
simulation in additive manufacturing processes, see e.g. [28, evaluation to obtain a local error estimate, with the advan-
29]. tages of A-stability and relatively low computational cost
A main focus of the paper is the derivation of reliable a of low order implicit methods. In order to avoid an addi-
posteriori local error estimates, which are essential for the tional solution of a system of equations we introduce a certain
two aims, class of RK-methods labeled as ELDIRK. These methods are
defined to have an explicit last stage in the general Butcher
1. to bound the local error by adaptive timestep control, array of Diagonal Implicit Runge–Kutta (DIRK) methods,
2. to contribute to an estimate of the global error. contrary to the class of EDIRK, where according to Kværnø
et al. [18] the RK-methods have an explicit first stage in the
Regarding the above mentioned first aim, to bound the general Butcher array
local error, two concepts are distinguished for the calcula- In addition to Mahnken [22], a general approach for the
tion of a posteriori local error estimates: firstly, by means construction of third order ELDIRK-methods is presented by
of higher order methods and secondly by means of extrap- embedding some well known low order implicit time inte-
olation methods, see e.g. [19] on general information and gration schemes. The proposed methodology is applied to
[23] on applications for creep analysis with the finite element embed the first order implicit Euler-method, the second order
method. In line with the first concept, Fehlberg suggested a trapezoidal rule and the second order Ellsiepen method by
strategy for stepsize selection by so-called embedded Runge– a third order scheme. As main results, two extra function
Kutta methods, [11–13]. The novelty of Fehlberg’s method is calculations are required in order to embed the implicit Euler-
that the applied integration scheme is embedded by a second method and one extra function calculation is required for
method with higher order, however, identical Runge–Kutta both, the trapezoidal-rule and Ellsiepen’s method, in order to
coefficients. As a consequence only one extra function calcu- obtain the third order properties. In this way, we will extend
lation is required to estimate the local error of the embedded the Butcher array in Eq. (1) to obtain a new method RK3-Ell
method. Thereupon, several embedded embedded Runge– of order 3 in order to embed RK2-Ell of order 2 with only
Kutta schemes have been developed. For example Bogacki one additional function evaluation. As verified in the numer-
and Shampine [3] present a Runge–Kutta method of order ical examples in Sect. 7, the higher order properties for all
three with four stages in order to embed a second-order new ELDIRK-methods with only one (and respectively two)
method. A recent paper [14] investigates strong stability additional function evaluation(s) are obtained for both linear
properties of embedded pairs of RK-methods, however, as well as for nonlinear ordinary differential equations.
restricted to explicit methods. Two different types of embedded RK-methods are distin-
A slightly different embedding compared to Fehlbergs guished in the literature. The original embedded RK-method
explicit schemes has been given by Ellsiepen [7] with Butcher explained above will subsequently be termed standard

123
Computational Mechanics (2023) 72:1239–1261 1241

embedded RK-method. For its variation, the roles of embed- 2 Preliminaries

ded and improved solutions are simply reversed, see [19,
p. 108], and therefore this method will subsequently be 2.1 Generalized Runge–Kutta scheme
denoted as reversed embedded RK-method.
Regarding the above mentioned second aim of local error Consider the initial value problem (IVP)
estimates, to contribute to an estimate of the global error, we
follow closely the procedure recently presented in [22]. Here, 1. ẏ(t) + A(y(t)) = f(t), t ∈ I = (α, β],
we consider an estimate for the global error measure obtained 2. y(α) = yα , (2)
from a quantity of interest, for both, the standard embedded
and the reversed embedded RK-method. This strategy avoids where the solution y(t) ∈ Rn , n > 0 typically represents a
the solution of a dual problem as performed e.g. in [25]. physical process dependent on time t. Moreover, A is a non-
The dual problem runs backwards in time [2, 17, 21, 31], linear operator defined in D(A) ⊂ V , where V is a suitable
which due to additional storage requirements might become Hilbert space with corresponding norm induced by an inner
a serious drawback when applied e.g. to FEM or respec- product  · V = (·, ·)1/2 (omitting mathematical details for
tively to FDM simulations. Moreover, in [22], we perform a brevity of presentation). The source (or respectively loading)
linearisation of the quantity of interest, related to both embed- term f and the initial value yα are given data. We note, that
ded RK-methods, in order to investigate the corresponding the representation (2) can be generalized to parabolic equa-
effectivity indices with the following main results: The error tions, see e.g. [31]. An example for the heat equation taking
estimate for the standard RK-method approaches Ieff = κ into account an additional spatial dependence will be given
in the limit. Here κ is a constant, which can be above or in Subsection 6.2 of this work. For brevity of presentation, in
below one, meaning, the error can be over- or underesti- the remainder of this section we restrict ourselves to temporal
mated. The effectivity index for the reversed RK-method discretization.
approaches Ieff = 1 in the limit, meaning that this error We begin with a partition of the time interval [α, β] as
estimate is asymptotically exact. In this work, these results α = t0 < t1 < · · · ti < · · · < t N = β and use the notations
are verified numerically, for all new RK-schemes developed
N
in this paper. 1. Iˆ = [α, β] = {α} ∪ i=1 Ii , 2. Ii = (ti−1 , ti ],
The structure of the paper is as follows: Sect. 2 sum-
marises some preliminaries on numerical time integration 3. τi =ti − ti−1 , i = 1, .., N , 4. τ = {max τi }i=1
N
,
of systems of ordinary equations with RK-methods includ- (3)
ing a posteriori error estimation. Here also, the class of q
ELDIRK-methods is introduced. Moreover, we will recall and yi , i = 1, . . . , N denotes the numerical approximations
the order conditions from [4, 6] to guarantee a certain con- of a q-order RK-method for the initial value problem (2) at
vergence order of an RK-method. In Sects. 3 and 4 we times ti . Then, starting with y0 = yα at t0 = α the time
present a general approach for the construction of second stepping algorithm of the RK-method is generally written in
order embedding ELDIRK-methods and third order embed- terms of a slope function (or increment function, see [10]),
ding ELDIRK-methods, respectively. The methodology is of the method φ q as
applied to three prominent low order implicit integration q q q q
schemes, that is the implicit Euler-method, the generalized yi = yi−1 + τi φ q (ti , yi−1 , yi , τi ), i = 1, . . . , N . (4)
trapezoidal-rule and Ellsiepen’s method. In this way, for
each of the methods we obtain an extended Butcher tableau. For an S-stage RK-method the algorithm (4) specifies as
Section 5 provides a brief revision of goal oriented error mea-
q q  S
sures introduced in [22], and the asymptotic behaviour of 1. yi = yi−1 + τi b j k j , where
error estimates for the standard embedded and the reversed  j=1 
embedded RK-methods. Two numerical examples in Sect. 6 q S
2. k j = fi j − A yi−1 + τi a jl kl , j = 1, . . . , S
are concerned with a parachute with viscous damping and a   l=1
two-dimensional laser beam simulation. Here, we illustrate 3. fi j = f ti−1 + c j τi (source terms),
the higher order convergence behaviour for equidistant step (5)
sizes of the proposed new Runge–Kutta methods. Finally, a
summary of the main results of this paper and an outlook for
future work are given in Sect. 7.

123
1242 Computational Mechanics (2023) 72:1239–1261

where the weights b j , the knots c j and the RK-coefficients This is in contrast to the class of EDIRK methods, which
a jl , respectively, satisfy the restrictions according to Kværnø et al. [18] have an explicit first stage in
the general Butcher array in Eq. (7), that is (at least) a11 = 0.

S The practical implementation of ELDIRK can be split into
1. b j = 1, 2. 0≤c j ≤1, two steps: Firstly, the unknowns k j , j = 1, . . . , Q < R
j=1 of the related implicit Q-stage method are determined by a

S nonlinear Newton algorithm as explained e.g. in [22]. Then,
3. a jl =c j , j=1, . . . , S. (6) in a second (post-processing) step, the additional slopes are
l=1 obtained in an explicit manner by exploiting definition (8)
from Eq. (5.2) as
Then, the slope function in Eq. (4) becomes a weighted
q q
sum of the slopes in Eq. (5.2) as φ q (ti , yi−1 , yi , τi ) =
S q 
Q
j=1 b j k j . The coefficients a jl , b j , c j , with j, l ∈ {1, . . . , k j = fi j − A yi−1 + τi a jl kl , Q < j ≤ R. (9)
S} are the ones corresponding to the Butcher tableau [5] l=1

c1 a11 ... ... ... a1S In this way, there is no need to solve another nonlinear system
c2 a21 a2,S at all, which holds for linear as well as for nonlinear ordinary
.. .. .. . . differential equations of the form (2.1).
. . . .
B(r )q = . (7)
cS a S1 a S2 . . . a S,S−1 a S,S
2.3 Order and stage conditions
b̂1 b̂2 . . . b̂ S−1 b̂ S
b1 b2 . . . b S−1 bS Following standard text books on odes, see e.g. [19], a local
error is introduced,
The first row of coefficients b̂ j below the horizontal line
in (7) will be explained later. For a jl = 0, l ≥ j the slopes ei := y(ti ) − y(ti−1 ) − τi φ q (ti−1 , y(ti−1 ), y(ti ), τi ). (10)
k j in Eq. (5.2) are obtained explicitly such that the number
of stages S is equal to the number of function evaluations per As shown in several classical textbooks, see e.g. [19], the
time-step. Otherwise the method becomes implicit, where order of accuracy can be obtained analytically, Q by inserting
for non-linear ordinary differential equations generally the the slope φ q (ti−1 , y(ti−1 ), y(ti ), τi ) = j=1 b j k j for the
number of function evaluations per time-step exceeds the Runge–Kutta method under investigation into the definition
number of stages S due to some iterative procedure. An iter- for the local error in Eq. (10) followed by a Taylor-series
ative Newton scheme for determination of the slopes k j valid expansion.
for implicit RK-methods is described e.g. in [22]. For more complex RK-methods this procedure may ensue
a tedious analysis, which can completely be avoided by
2.2 The class of ELDIRK-methods means of the general results on RK-methods in [6]. Here it is
extensively shown, that if the RK-method is of order q, then
A classification of RK-methods is introduced in [18]. Here, the RK-coefficients ck , bk , akl of the Butcher array in Eq. (5)
in particular Diagonal Implicit Runge–Kutta (DIRK)-type have to satisfy certain order conditions for the RKq formula.
methods are of particular interest, where in the Butcher Butcher [4] has listed the related conditions up to order 8.
tableau in Eq. (7) one has ai j = 0 for all j > i and aii = 0 Table 1 contains the complete set of required equations for
for at least one i. We denote by Q and R > Q the stages orders up to q = 4. Note, that the particular case q = 1
for the pair of a q-order method embedded by an r −order accounts for consistency, see e.g. [19]. Additionally, in this
RK-method, where r > q. Such an embedding of implicit table we introduce the number of elementary differentials
DIRK methods has been considered for some prominent low nr thus resulting into the total number of order conditions
q
order methods in [22]. There, it is shown, that an additional mq = r =1 n r , see [6]. Also, the stage conditions up to
solution of a system of equations for the R-stage method can stage Q are recalled from Eq. (6.3) in Table 1.
be avoided, if the diagonal elements at the last stage(s) in the Example. Application of the order and stage conditions of
general Butcher array in Eq. (7) are zero. For a systematic Table 1 to the RK(2)1-Ell method of Ellsiepen [7] with
generalization of this convenient property we introduce in Butcher tableau in Eq. (1) reveals:
this paper a class of RK-methods defined as
(1) (2)
RK1-Ell : a1 = 0, a1 = 0
ELDIRK: Explicit Last stage DIRK, RK2-Ell : a1(1) = 0; a1(2) = 0 (11)
where a j j = 0, Q < j ≤ R. (8) RK(2)1-Ell : s (1) = 0, s (2) = 0

123
Computational Mechanics (2023) 72:1239–1261 1243
⎧
⎨ 1. RKr : yi =yi−1 +τi φ (ti , yi−1 , yi , τi )
r r r r r
Table 1 Order conditions (up to q = 4) and stage conditions ⎪
• Numbers of elementary differentials nr and order conditions m q
RK(q)r 2. RKq,r : yq,r =yr +τi φ q (ti , yr , yq,r , τi ),
⎪
⎩ i i−1 i−1 i
Orders r , q 1 2 3 4 i=1, . . . , N .
nr 1 1 2 4 (13)

q
mq = nr 1 2 4 8
r =1
Note, that the lower order method RKq,r, q ≤ r −1 and stage
q,r
• Order conditions for orders q = 1, . . . , 4 Q uses the initial value yi−1 = yri−1 at each time step, which
(1)  explains the upper index q, r .
q = 1 1. a1 = bj − 1 = 0
j According to Fehlberg’s suggestion the RK-formulae of

q = 2 2. a1
(2)
= bjcj − 1
2 =0 orders q and r > q both in (12) and in (13) (usually r =
j

q + 1) share the same function evaluations, i.e. they have
q = 3 3. a1(3) = 1
2 b j c2j − 1
6 =0 the same RK-coefficients a jl and the same knots c j , j =
j
(3)  1, ..., Q. In addition to b j , j = 1, ..., Q weighting factors b̂ j ,
4. a2 = b j a jl cl − 1
6 =0
jl j = 1, ..., R are introduced in the Butcher array in Eq. (7),

q = 4 5. a1
(4)
= 1
6 b j c3j − 1
24 =0 where the symbol ˆ refers to the higher r -order method with

j numerical solutions yri according to Eq. (5).
6. a2(4) = b j c j a jl cl − 18 = 0
jl
A reliable procedure for the numerical solution of initial
(4) 1  value problems requires adaptive time-step selection. Let us
7. a3 = 2 b j a jl cl2 − 241
=0
jl assume the following conditions [19]:

8. a4(4) = b j a jl alm cm − 24 =
1
0
jlm
q
• Stage conditions for stages j = 1, . . . , S 1. The value yi−1 at the beginning of the time interval is
 q
exact, i.e. yi−1 = y(ti−1 ).
s ( j) = l a jl − c j = 0, j = 1, 2, . . . , S (14)
q q+1
2. yri−1 − yi−1 = O(τi ), where r ≥ q + 1,
3. Terms of order O(τ q+2 ) are neglected.
which confirms, that RK1-Ell (using β) is indeed only a first
Then, from the above embedding
order method whilst RK2-Ell (using α) is a second order
method. Both RK-schemes satisfy the consistency condition.
r ,q r ,q q 
R
q+1
1. η̂i := yi − yi = τi (b̂ j − b j )k j = O(τi )
j=1
2.4 Embedded Runge–Kutta methods R
q,r q,r q+1
2. η̂i := yi − yri = τi (b j − b̂ j )k j = O(τi )
An efficient computation of error estimates is achieved by j=1

so-called embedded Runge–Kutta methods, where two con- (15)

ceptions are distinguished:
The standard embedded RK-method in its original form become a posteriori estimates for the local error of ei for both
introduced in [11–13], and subsequently denoted as RK(r)q, RK(r)q and RK(q)r. These are used to control the step-size
is given as a pair of time stepping algorithms according to for both, RK(r)q in (12) and RK(q)r in (13), respectively,
(4) as, see e.g. [19], according to the formulae

⎧ ⎡ ⎧ ⎫⎤
q q q q ⎨ 1/(1+q) ⎬
⎨ 1. RKq : yi =yi−1 + τi φ (ti , yi−1 , yi , τi )
q
⎪ ε
τi = τi−1 · min ⎣2, max 0.2, 0.9 ⎦,
RK(r)q 2. RKr,q : yr ,q =yq + τi φ r (ti , yq , yr ,q , τi ), ⎩ η̂ia,b  ⎭
⎪
⎩ i i−1 i−1 i
i = 1, . . . , N . (16)
(12)
where a, b = r , q or a, b = q, r , respectively, and where
Here RKq has order q and stage Q. The higher order method  ·  denotes some norm of the related quantity, e.g. the l2-
RKr,q, r ≥ q + 1 and stage R > Q, uses the initial value norm or the maximum-norm. Also, ε is a tolerance for the
r ,q q
yi−1 = yi−1 at each time step, which explains the upper index estimated error, 0.9 is a safety factor and 2 and 0.2 are values
r , q. Alternatively, a reversed embedded RK-method RK(q)r for upper and lower bounds of the step-length. Since the error
see e.g. [6], also known as local extrapolation, see e.g. [19], estimate η̂ia,b is dependent on the (unknown) final time-step
is given as a pair of time stepping algorithms according to τi , in practice τi is obtained in a trial and error procedure.
(4), We point out, that the procedure for local error estimation by

123
1244 Computational Mechanics (2023) 72:1239–1261

means of the extended Butcher tableau in Eq. (7) is applicable Taking into account one degree of freedom in the above con-
to both, explicit and implicit embedded RK-methods. ditions a reformulation is performed next, with the aim to
obtain explicit expressions for â21 , b̂1 , b̂2 in terms of a cho-
sen value for ĉ2 . To this end, from (20.1) we deduce
3 Second order ELDIRK for embedding of
RK1 b̂2 = 1 − b̂1 . (21)

3.1 General setting

Inserting Eq. (21) into Eq. (20.2) renders
This section intends to develop a general procedure to embed  
a Runge–Kutta method of order 1, subsequently denoted 1 1
b̂1 = − ĉ2 . (22)
as RK1 with Butcher array B1, by an ELDIRK-method of (c1 − ĉ2 ) 2
order 2, subsequently denoted as RK2, thus resulting into an
embedded RK-method RK(2)1 with extended Butcher array Finally, from Eq. (20.3) we obtain
B(2)1:

â21 = ĉ2 . (23)

c1 a11
c1 a11 ĉ2 â21
B1 = , B(2)1 = . (17) For given value of ĉ2 the unknowns â21 , b̂1 , b̂2 in Eqs. (21)–
b1 b̂1 b̂2 (23) can be evaluated, subject to the condition
b1

In these tableaus RK-coefficients without a hat are assumed ĉ2 = c1 . (24)

to be known, and RK-coefficients with a hat are assumed to be
unknown. Note, that in order to become an ELDIRK-method 3.2 Embedding of the explicit and implicit Euler
according to definition (8), we require â22 = 0. method
We also assume, that the stage condition
As a first example, an embedding of the explicit Euler-scheme
in Eq. (19) can be obtained as

j = 1 : s (1) = l a1l − c1 = 0 (18)
ĉ2 = 1/2 ⇒ b̂1 = 0, b̂2 = 1, â21 = 1/2, (25)
is a priori satisfied for RK1. Two well known examples for
B1 are the explicit and the implicit Euler-schemes
which is the classical modified Euler-method, see e.g. [19].
An embedding of the implicit Euler-scheme in Eq. (19) can
0 0 1 1
B1 Eul,ex p = , B1 Eul,imp = . (19) be obtained with the same choice for ĉ2 as
1 1

ĉ2 = 1/2 ⇒ b̂1 = 0, b̂2 = 1, â21 = 1/2. (26)

In the tableau for B(2)1 in Eq. (17) the higher order
ELDIRK-method has four unknowns ĉ2 , â21 , b̂1 , b̂2 . To
obtain order r = 2 with stage R = 2 reference to Table 1 For convenience, the resulting Butcher-Tableaus for the sec-
shows n 1 = 1, n 2 = 1, such that according to Table 1 m̂ 2 = 2 ond order embedded Euler methods RK(2)1-Eul-exp and
order conditions and R = 2 stage conditions have to be satis- RK(2)1-Eul-imp are given in Table 2.
fied. Since the stage condition (18) is a priori satisfied, with
3 conditions for 4 unknowns, there is one degree of free- Remark 1 The choices ĉ2 = 21 in Eq. (25) and Eq. (26) are
dom in the selection of the constants. The resulting 2 order indeed rather arbitrary. A different choice would only result
conditions and 1 stage condition are taken from Table 1 as: into a different Butcher–Tableau for the embedded RK(2)1-
Eul-exp and RK(2)1-Eul-imp in Table 2. The procedure can
(1) 
1. a1 = bj − 1 = b̂1 + b̂2 − 1 = 0 be compared to the classical choice of constants for deriva-
j tion of the modified Euler-method, the Heun-method and the
(2) 
2. a1 = bjcj − 1
2 = b̂1 · c1 + b̂2 · ĉ2 − 1
2 =0 (20) Ralston-method, see e.g. [19]. All of these are second order

j
explicit RK-methods where none of them has an advantage
3. s (2) = a2l − ĉ2 = â21 − ĉ2 = 0.
l over the other two.

123
Computational Mechanics (2023) 72:1239–1261 1245

Table 2 Summary of
Butcher-tableaus for second Embedded explicit Euler-method RK(2)1-Eul-exp: (R, S) = (2, 1), (r , q) = (2, 1)
order explicit and implicit Euler 0 0
embedded 1 1
Runge–Kutta-methods B(2)1-Eul-exp = 2 2
0 1
1
Embedded implicit Euler-method RK(2)1-Eul-imp: (R, S) = (2, 1), (r , q) = (2, 1)
1 1
1 1
B(2)1-Eul-imp = 2 2
0 1
1

3.3 Evaluation of the new second order c1 a11

Butcher-Tableaus c1 a11 c2 a21 a22
B2 = c2 a21 a22 , B(3)2 = ĉ3 â31 â32 . (29)
Comparing the new Butcher-Tableaus in Table 2 reveals the
b1 b2 b̂1 b̂2 b̂3
following stages R, Q for the embedded R K (r ) p schemes:
b1 b2

RK(2)1-Eul-exp : R = 2, Q = 1 As before, RK-coefficients without a hat are assumed to be

(27)
RK(2)1-Eul-imp : R = 2, Q = 1. known, and RK-coefficients with a hat are assumed to be
unknown. Note, that in order to become an ELDIRK-method
According to Sect. 2.2 on ELDIRK-methods, the practical according to definition (8), we require â33 = 0.
implementation of both embedded methods in Table 2 can We also assume that the stage conditions
be split into two steps: Firstly, the unknown slope k1 of the 
related 1-stage Euler-method is calculated, for RK1-Eul-exp j = 1 : s (1) = l a1l − c1 = 0
 (30)
as an explicit update, for RK1-Eul-imp by a nonlinear Newton j =2: s (2) = l a2l − c2 = 0
algorithm as explained e.g. in [22]. Then, according to Eq. (9)
the additional slope is obtained in a second (post-processing) are a priori satisfied for RK2.
step, in an explicit manner as (observe a22 = 0) Some examples for B2 are the embedded implicit modi-
fied Euler-method and the implicit Euler-scheme of Table 2
 q  as well as the trapezoidal method
k2 = fi2 − A yi−1 + τi a21 k1 . (28)
0 0
In this way, both two-stage methods, the forward scheme 1 1
B= 1 2 2 , (31)
RK2-Eul-exp and the backward scheme RK2-Eul-imp, require 1 1
only one extra function evaluation to increase the order from 2 2
q = 1 to r = 2, without solving another nonlinear system
at all. Note, that this feature holds for linear as well as for or RK2-Ell of Ellsiepen [7] according to Eq. (1)
nonlinear ordinary differential equations of the form (2.1).
α α
1
B= 1 1−α α, α =1− √ . (32)
2
4 Third order ELDIRK for embedding of RK2 1−α α

4.1 General setting In the tableau for B(3)2 in Eq. (29) the higher order
ELDIRK-method has six unknowns ĉ3 , â31 , â32 , b̂1 , b̂2 , b̂3 .
This section intends to develop a general procedure to embed To obtain order r = 3 with stage R = 3 reference to Table 1
a Runge–Kutta method of order 2, subsequently denoted shows n 1 = 1, n 2 = 1, n 3 = 2, such that according to Table 1
as RK2 with Butcher array B2, by an ELDIRK-method of m̂ 3 = 4 order conditions and R = 3 stage conditions have
order 3, subsequently denoted as RK3, thus resulting into an to be satisfied. Since the stage conditions (30) are a priori
embedded RK-method RK(3)2 with extended Butcher array satisfied, there is one degree of freedom in the selection of
B(3)2: the constants. The resulting 4 order conditions and 1 stage

123
1246 Computational Mechanics (2023) 72:1239–1261

condition are taken from Table 1 as: For given value ĉ3 the unknowns â31 , â32 , b̂1 , b̂2 , b̂3 in
 Eq. (29) can be evaluated by means of Eqs. (34)–(38) with
(1)
1. a1 = b̂ j − 1 the conditions
j

= b̂1 + b̂2 + b̂3 − 1 = 0 1. ĉ3 = c1 , 2. ĉ3 = c2 . (39)

(2)
 1
2. a1 = b̂ j c j −
2 4.2 Embedding of the modified implicit Euler
j
method
1
= b̂1 · c1 + b̂2 · c2 + b̂3 ĉ3 − =0
2 For the second order modified implicit Euler method the cor-
(3) 1 1
3. a1 = b̂ j c2j − responding RK-coefficients are obtained from Table 2 as
2 6
j
1  1 c1 = 1, c2 = 1/2, a11 = 1, a21 = 1/2, a22 = 0. (40)
= b̂1 · c12 + b̂2 · c22 + b̂3 ĉ32 − = 0
2 6
 1
4. a2(3) = b̂ j a jl cl − Consequently, with the choice
6
jl
ĉ3 = 1/4 (41)
= b̂1 · a11 c1 + b̂2 (a21 c1 + a22 c2 )
  1
+b̂3 â31 c1 + â32 c2 − = 0 Equations (36), (35) and (33.1) render
6

5. s (3) = a3l − ĉ3
l
b̂1 = 29 , b̂3 = 49 , b̂2 = 13 . (42)
= â31 + â32 − ĉ3 = 0. (33)
Next, inserting (42) into Eq. (38) and using Eq. (37) gives
Taking into account one degree of freedom in the above con-
ditions a reformulation is performed next, with the aim to â31 = − 45 , â32 = 23 . (43)
obtain explicit expressions for â31 , â32 , b̂1 , b̂2 , b̂3 in terms
of a chosen value for ĉ3 . For convenience, the resulting new Butcher-Tableau for the
From (33.2) we deduce third order embedded implicit Euler method RK(3)2-Eul is
  given in Table 3.
1 1
b̂2 = − b̂3 ĉ3 − b̂1 c1 . (34)
c2 2 4.3 Evaluation of the new third order
Butcher-Tableaus
Inserting (34) into Eq. (33.3) multiplied with 2 renders
  The embedding schemes for the trapezoidal-rule RK(3)2-
1 1 1
b̂3 = − b̂1 (c12 − c1 c2 ) + c2 − . (35) Trap and the Ellsiepen method RK(3)2-Ell are derived in
ĉ3 (ĉ3 − c2 ) 2 3
a similar fashion as to the implicit Euler method RK(3)2-Eul
With the results (34) and (35), Eq. (33.1) yields in Appendix A. Comparing the new Butcher-Tableaus in
Table 3 reveals the following stages R, Q for the embedded
3c2 − 2 + 3ĉ3 (1 − 2c2 ) R K (r )q schemes:
b̂1 = − . (36)
6(c1 − ĉ3 )(c1 − c2 )
RK(3)2-Eul : R = 3, Q = 2
By means of Eq. (33.5) and RK(3)2-Trap : R = 3, Q = 2 (44)
RK(3)2-Ell : R = 3, Q = 2.
â32 = ĉ3 − â31 (37)
According to Sect. 2.2 on ELDIRK-methods, the practical
we exploit Eq. (33.4) to obtain implementation of both embedded methods in Table 3 can be
split into two steps: Firstly, the unknowns k1 , k2 of the related
−1 
â31 = b̂1 · a11 c1 2-stage method are calculated: For RK2-Eul as explained
b̂3 (c1 − c2 ) in Sect. 3.3 including Eq. (28), for RK2-Trap and RK2-Ell,

1 respectively, by a nonlinear Newton algorithm as explained
+b̂2 (a21 c1 + a22 c2 ) + b̂3 c2 ĉ3 − . (38) e.g. in [22]. Then, according to Eq. (9) the additional slopes
6

123
Computational Mechanics (2023) 72:1239–1261 1247

Table 3 Summary of new

Butcher-tableaus for third order Embedded implicit Euler-method RK(3)2-Eul: (R, S) = (3, 2), (r , q) = (3, 2)
implicit embedded 1 1
Runge–Kutta-methods 1 1
2 2
1
4 − 45 3
2
B(3)2-Eul =

2 1 4
9 3 9
0 1
Embedded trapezoidal method RK(3)2-Trap: (R, S) = (3, 2), (r , p) = (3, 2)
0 0
1 1
1 2 2
1 3 1
B(3)2-Trap = 2 8 8

1 1 2
6 6 3
1 1
2 2
Embedded Ellsiepen-method RK(3)2-Ell: (R, S) = (3, 2), (r , q) = (3, 2)
1
α = 1− √
α α 2
1
1 1−α α b̂1 =
6(α − α 2 )
B(3)2-Ell = 0 α−1 1−α
2 − 3α
b̂2 =
b̂1 b̂2 b̂3 6(1 − α)
1−α α 1  
b̂3 = 4α − 3α 2 − 1
6(α − α )
2

are obtained in a second (post-processing) step, in an explicit 5 Revision of goal oriented global error
manner as (observe a33 = 0) measures applied to RK(r)q and RK(q)r
 q  
k3 = fi3 − A yi−1 + τ a31 k1 + a32 k2 . (45) This section intends to provide a brief summary of goal ori-
ented global temporal error measures introduced in [22], and
In this way, RK3-Eul being a three-stage ELDIRK-method its asymptotic behaviour for standard embedded RK(r)q in
requires one extra function evaluation to increase the order Eq. (12) and for reversed embedded RK(q)r in Eq. (13)
from q = 2 for RK2-Eul to r = 3. In other words, RK3- Global temporal error estimation can be achieved within
Eul requires two extra function evaluations to increase the a mathematically sound way by means of two functionals
order from q = 1 for the classical Euler methods (backward as originally proposed for partial differential equations, see
and forward) to r = 3. RK3-Trap has already been derived e.g. [2, 8, 27, 30, 36]: Firstly, a weak form for temporal dis-
in [22], however from a somewhat different viewpoint. Both, cretization of the underlying initial value problem (IVP), and,
RK3-Trap and RK3-Ell are three stage ELDIRK-methods and secondly, a quantity of interest (QoI) or goal function, respec-
require one extra function evaluation to increase the order tively, which enables to measure the related global error. For
from q = 2 to r = 3, without solving another nonlinear the temporal discretization, the former can be obtained as a
system at all. As mentioned before, these features hold for weak Petrov-Galerkin formulation. Here the concept of the
linear as well as for nonlinear ordinary differential equations finite element method for space discretizations is adapted
of the form (2.1). to the time domain, originally introduced in [1, 15, 26]. To
this end the underlying differential equation (2) is reformu-
Remark 2 In [18] stability is investigated for certain sub-
lated into a corresponding weak form by a premultiplication
classes of DIRK, such as EDIRK. However, due to the differ-
with a test function and an integration over the time domain
ent structure of ELDIRK compared to EDIRK, the results of
I . The solution functions as well as the test functions are
[18] cannot directly be applied. Therefore, additional stabil-
then approximated by means of appropriate interpolations.
ity investigations are required for the new ELDIRK-methods
Conventionally, discontinuous Galerkin (dG) and continu-
in Tables 2 and 3, respectively. This non-trivial task, how-
ous Galerkin (cG) methods are distinguished, see e.g. [8].
ever, is not in the scope of the current paper and therefore
Eventually, the weak form enables to derive RK-schemes
will be a task of future work.

123
1248 Computational Mechanics (2023) 72:1239–1261

which result into the Butcher array (7). A simple example is measures coincide. For simplicity, if there is no danger of
e.g. given in [35] for the backward (implicit) Euler scheme. confusion, the upper index eq or respectively qe is dropped
Regarding the mathematical details for derivation of a gen- from now on. Note, also that both functionals E (·, ·) share
eral explicit S-stage RK-method on the basis of a weak form the property
the interested reader is referred e.g. to Muñoz-Matute et al.
[25]. The QoI approach offers a great flexibility on global E (y, y) = 0, ∀y ∈ V. (49)
error estimation, since the goal function is a user specified
q q
quantity, see e.g. [17, 21]. A Taylor series of E (y, yτ ) around yτ renders
In this section, a global error estimate is obtained directly  
q q q q q
from a quantity of interest, for both, the standard embedded E (y, yτ ) = E (yτ , yτ )+Dyqτ E (yτ , yτ ) · Eq +O Eq 2V
and the reversed embedded RK-method by means of the func-  
q q
tional representation for the weak form. This strategy avoids = ET (yτ , yτ ; Eq ) + O Eq 2V ,
the solution of a dual problem as performed e.g. in [25]. The (50)
dual problem runs backwards in time [2, 17, 21, 31], which
q q q
due to additional storage requirements might become a seri- where Eq. (49) has been used. ET (yτ , yτ ; Eq ) := Dyqτ E (yτ ,
q
ous drawback when applied e.g. to FEM or respectively to yτ ) · Eq is the Gateaux-derivative (or tangent, respectively)
q
FDM simulations. of E (·, yτ ) with respect to the first argument. Note, that the
tangent is linear in the third argument, which is indicated by
5.1 Global error and effectivity index for standard a semicolon.
embedded RK(r)q As a concrete example for an QoI, in this work we will
consider the time point error function
We let y ∈ V be the solution of a related weak form of the
q
IVP (2) where V is a suitable functional space. Also, yτ and 1. Q(y) := y(t N )V ,
r ,q
yτ are two solutions related to standard embedded RK(r)q 2. E (y, yqτ ) = Q(Eq ) = E N V .
q
(51)
of Eq. (12). Note, that these functions are continuous for
the cG-method, whilst for the dG-method they are allowed For an alternative, a time integrated goal function is consid-
to be discontinuous at the nodal points, but are taken to be ered in [22].
continuous within the time intervals Ii , see e.g. [8, 35]. As The following relations are introduced in [22] for general
mentioned above, the resulting weak forms can be exploited goal functions in Eq. (51.1) with related error measures in
to derive RK-schemes in the form of the Butcher array (7), Eq. (52.2) valid for “small” time steps τ defined in Eq. (3.4),
or to formulate dual problems, which however is not in the
q q q q
scope of this paper, see e.g. [25]. 1. Q(Eq )= E (y, yτ ) ≈ ET (yτ , yτ ; Eq ) ≈ Cr ,q τ P
In the sequel, the above functions are used to define two r ,q r ,q r ,q r ,q
2. Q(Er ,q )=, E (y, yτ ) ≈ ET (yτ , yτ ; Er ,q ) ≈ Cr ,q τ P
global semi-discretization (temporal) errors and a global r ,q η r ,q
3. Q(ηr ,q )=, E (yτ , yτ ) ≈ ET (yτ , yτ ; ηr ,q ) ≈ Cr ,q τ Z
q q q
error estimate as .
(52)
1. Er ,q (t) := y(t) − yrτ,q (t), 2. Eq (t) := y(t) − yqτ (t),
η
3. ηr ,q (t) := yrτ,q (t) − yqτ (t), t ∈ I . (46) Here, Cr ,q , Cr ,q and Cr ,q are some constants, and P q , P r ,q ,
Z r ,q are three convergence orders which satisfy, [19, 22],
Next, following [9] and [30], we introduce a goal quantity
of interest P q = P r ,q = Z r ,q = q. (53)

Q(.) : V → R (47) Remark 3 1. In Lambert, p. 58, it has been pointed out,

while the local truncation error ei for an RKq-method in
q+1
such that two global error measures E (., .) : V × V → R can Eq. (52.1) is O(τi ), the global truncation error Q(Eq )
be distinguished, an error of the quantity and respectively a in Eq. (53) is O(τ q ) with τ according to Eq. (3.4). That
quantity of the error is, one power of τ has been lost owing to the process of
accumulation and which leads to P q = q in Eq. (53).
1. E eq (y, yqτ ) = Q(y) − Q(yqτ ), This property is also verified for the classical methods
2. E qe (y, yqτ ) = Q(y − yqτ ) = Q(Eq ), (48) RK1-Eul, RK2-Trap and RK2-Ell by the numerical results
in the forthcoming Sect. 6, see e.g. Table 4.
where for the second case the definition for Eq in (46.2) 2. As mentioned in Sect. 2.3 the local error estimates η̂i
q+1
has been used. Clearly, for linear goal functions, both error in Eq. (15) are order O(τi ), provided the conditions

123
Computational Mechanics (2023) 72:1239–1261 1249

Table 4 Parachute with viscous damping for goal function in Eq. (60): 5.2 Global error and effectivity index for reversed
results for four standard embedded RK-methods and its reversed meth- embedded RK(q)r
ods with equidistant step sizes
RK-Method Analogously to the relations Eq. (46) two global errors and
RK(r)q Z r ,q P r ,q Pq = q an error estimate are defined
RK(2)1-Ell 1.00e+00 9.70e−01 1.05e+00
RK(3)2-Trap 2.01e+00 2.01e+00 2.00e+00 1. Eq,r (t) := y(t) − yq,r
τ (t), 2. E (t) := y(t) − yτ (t),
r r

RK(3)2-Ell 2.01e+00 2.01e+00 2.01e+00 3. ηq,r (t) := yq,r

τ (t) − yτ (t), t ∈ I .
r
(55)
RK(3)2-Eul 1.99e+00 1.99e+00 1.99e+00

RK(q)r Z q,r P q,r Pr = r A relation between these quantities is

RK(1)2-Ell 1.01e+00 9.89e−01 2.01e+00
q,r q,r
RK(2)3-Trap 2.01e+00 2.02e+00 3.01e+00 Er = y − yrτ = y − yτ + yτ − yrτ = Eq,r + ηq,r . (56)
RK(2)3-Ell 2.01e+00 2.01e+00 3.02e+00
RK(2)3-Eul 1.99e+00 2.00e+00 3.00e+00 Then, analogously to (52) the following relations are intro-
Global convergence orders, top, for standard embedded RK(r)q duced in [22] for reversed embedded RK(q)r in in (13), valid
Z r ,q , P r ,q , P q = q according to Eq. (52) and, bottom, for reversed for “small” time steps τ defined in Eq. (3.4), as
embedded RK(q)r Z q,r , P q,r , P r = r according to Eq. (57)
r
1. Q(Er ) = E (y, yrτ ) ≈ ET (yrτ , yrτ ; Er ) ≈ Cr τ P
q,r q,r q,r q,r
2. Q(Eq,r ) = E (y, yτ ) ≈ ET (yτ , yτ ; Eq,r ) ≈ Cr τ P
(14) hold. Regarding the global error estimate Q(ηr ,q ) in q,r η q,r
3. Q(ηq,r ) = E (yτ , yrτ ) ≈ ET (yrτ , yrτ ; ηq,r ) ≈ Cq,r τ Z
Eq. (52.2), the order condition (14.1) for the initial value
does not hold, such that the global error estimate Q(ηr ,q ) (57)
q
in Eq. (52.2) is only order O(τi ), and which leads to
Z r ,q = q in Eq. (53), see also e.g. [32] and references η
where Cr , Cr and Cq,r are some constants. For the global
therein. convergence order constants it holds, [22],

1. P r = r = q + 1, 2. P q,r = Z q,r = q = r − 1.
The above relations allow to evaluate the quality of an error (58)
estimate. To this end, we borrow from Szabo and Babuska
[34] the definition for the effectivity index originally for-
mulated for finite element error estimation as the ratio of These properties are verified for the new ELDIRK methods
r ,q RK3-Eul, RK3-Trap and RK3-Ell by the numerical results in
the estimated global error E (yτ , yrτ ) and the global error
q
E (y, yτ ). Then, the effectivity index becomes in the limit the forthcoming Sect. 6, see e.g. Table 4.
τ →0 Then, by means of Eq. (56) and exploiting the linearity of
the tangent form in the third argument a related effectivity
index becomes in the limit τ → 0

r ,q
Q(ηr ,q )
q
q E (yτ , yτ ) q,r
E (yτ , yrτ ) Q(ηq,r )
lim Ieff = lim q = lim q,r
lim Ieff = lim = lim
τ →0 τ →0 E (y, yτ ) τ →0 Q(Eq ) τ →0
q,r
τ →0 E (y, yτ ) τ →0 Q(Eq,r )
η η
Cr ,q τ q Cr ,q ET (yrτ , yrτ ; ηq,r )
= lim = =: κ. (54) = lim
τ →0 Cr ,q τ q
q,r q,r q,r q,r
Cr ,q τ →0 |ET (yτ , yτ ; Er ) − ET (yτ , yτ ; ηq,r )|
 η  η
 Cq,r τ q  Cq,r
= lim   = η = 1. (59)
τ →0 Cr τ q+1 − C q,r τ q 
η
Cq,r
Consequently, the error estimate Q(ηr ,q ) is a κ-multiple of
q
the error Q(Eq ). The effectivity index Ieff approaches some The interpretation of Eq. (59) is, that the error estimate
constant κ, which depends on the specific RK-algorithm and Q(ηq,r ) is asymptotically exact for the global error Q(Eq,r )
the individual problem to be solved. In general, the effectivity of RKq,r. This is a main advantage of the reversed embed-
index cannot be one (even in the limit), which is a main ded method RK(q)r compared to standard embedded RK(r)q,
limitation of standard embedded RK-methods. [22].

123
1250 Computational Mechanics (2023) 72:1239–1261

6 Representative numerical examples order ODE typically attains the form m ẍ + d ẋ + cx = F(t),
where m is the mass, x is the position, d is a damping constant,
In the following, numerical results of the algorithms sum- c is an elastic constant and F(t) is some applied force, see
marized in Table 3, the embedded implicit Euler method e.g. [20]. As a particular problem we consider the motion of a
RK(3)2-Eul, the embedded trapezoidal rule RK(3)2-Trap and parachute with viscous damping d ẋ (alternatively a quadratic
the embedded Ellsiepen method RK(3)2-Ell are presented for ansatz could be applied for the damping term, see e.g [20].)
two test problems: Firstly, a parachute with viscous damp- Moreover we set c = 0 and F = mg, where g is the grav-
ing and, secondly, a laser beam simulation with the finite itational acceleration. Using the definition v := ẋ for the
difference method. Results are presented for equidistant step velocity, and assuming zero velocity at time t = 0, renders a
lengths as well as for non-equidistant step lengths based on first order IVP of the form (2) as:
adaptive time-step control. We will also verify the expected
order properties of the new third order ELDIRK methods d
1. v̇(t) = g − v, 2. v(t = 0) = 0. (62)
RK3-Eul, RK3-Trap and RK3-Ell, and we will present results m
for the reversed embedded Runge–Kutta methods, denoted In [20] its analytical solution is derived as
as RK(2)3-Eul, RK(2)3-Trap and RK(2)3-Ell. Thereby, the
  
nonlinear equations resulting from the implicit RK-schemes d mg
are solved with a Newton method, as explained in [22]. v(t) = vst 1 − exp − t , vst = . (63)
m d
For comparison, we present also results for the embedded
RK-method of Ellsiepen according to Eq. (1) [7], denoted as Here vst is the stationary velocity, obtained from the con-
RK(2)1-Ell. As explained before, both RK1-Ell and RK2- dition v̇(t = ∞) = 0 = g − dvst /m. In the subsequent
Ell are implicit, however, the Newton method iterates on calculations the constants are chosen as: m = 70 kg, d =
the same slopes k j , which satisfy a residual obtained from 20.5 kg/s, g = 9.81 m/s2 . The initial time is t0 = α = 0 and
Eq. (5.2). In analogy to previous notations, RK(1)2-Ell the total time is T = β − α = 10 s.
denotes the corresponding reversed RK-method.
For both, standard embedded RK(r)q and reversed embed- 6.1.2 Global convergence investigations for equidistant
ded RK(q)r the error measure in Eq. (51.2) is investigated. step sizes
Thus, by means of the notation · N := ·(t N ) and a 2-norm, for
standard embedded RK(r)q the following three global error Figure 1 shows results for all four standard embedded RK-
functions are introduced: methods on the left hand side and its reversed embedded
RK-methods on the right hand side for N = 10 equidistant
r ,q
1. Q(Eq ) = E N 2 , 2. Q(Er ,q ) = E N 2 , time steps. For the velocity versus time v(t) in Fig. 1a) a
q

r ,q
3. Q(ηr ,q ) = η N 2 (60) difference to the exact solution is hardly visible for all RK-
methods. The same holds for all reversed methods in Fig. 1b).
r ,q Table 4 summarizes the global convergence orders Z r ,q ,
with global error estimate η N according to Eq. (46.3). We ,q
q P , P q defined in Eq. (52) for all four embedded RK-
r
will verify the result limτ →0 Ieff = κ in Eq. (54). For reversed
methods and Z q,r , P q,r , P r defined in Eq. (57) for its
embedded RK(q)r the following three global error functions
reversed methods with equidistant step sizes. For standard
are introduced:
embedded RK(r)q the approximation P q ≈ P r ,q ≈ Z r ,q ≈
q,r q according to Eq. (53) is obtained, whereas for reversed
1. Q(Er ) = ErN 2 , 2. Q(Eq,r ) = E N 2 ,
q,r
embedded RK(q)r P r ≈ r = q + 1 > P q,r ≈ Z q,r ≈ q
3. Q(ηq,r ) = η N 2 , (61) according to Eq. (58) is obtained. In this way, all RK-methods
reach (and occasionally even surpass) its theoretical conver-
q,r
with global error estimate η N according to Eq. (55.3). We
q,r 
gence orders.

will verify the result limτ →0 Ieff = 1 in Eq. (59). As an Regarding the new ELDIRK methods in Table 3 the fol-
alternative to the above time point goal function, results for lowing is pointed out from the results in Table 4: The order for
a time integrated goal function are presented in [22]. RK3-Ell and RK3-Trap is increased by one compared to the
implicit methods RK2-Ell and RK2-Trap, respectively, with
6.1 Parachute with viscous damping only one additional function evaluation according to Eq. (45),
whilst the order of RK3-Eul is increased by two compared
6.1.1 Initial value problem (IVP) to the implicit method RK1-Eul with two additional function
evaluations according to Eq. (28) and Eq. (45), respectively.
The first numerical example is concerned with a single- There is no need to solve another nonlinear system, although
degree-of-freedom system in dynamics. Here the second the IVP (62) is nonlinear.

123
Computational Mechanics (2023) 72:1239–1261 1251

q
Fig. 1 Parachute with viscous damping: Results for four standard effectivity index Ieff in Eq. (54). Right for reversed embedded RK(q)r:
embedded RK-methods and its reversed methods with equidistant step b Velocity versus time, d exact global error E q,r in (55.1) and estimated
q,r
sizes. Left for standard embedded RK(r)q: a velocity versus time, c exact global error ηq,r in Eq. (55.3), f effectivity index Ieff in Eq. (59)
global error E q in (46.2) and estimated global error ηr ,q in Eq. (46.3), e

123
1252 Computational Mechanics (2023) 72:1239–1261

Table 5 Parachute with viscous damping for goal function in Eq. (60): results for four standard embedded RK-methods and its reversed methods
with equidistant step sizes
RK-Method τi = 1.000e+00 s τi = 1.000e−03 s
Q(Er ,q ) Q(ηr ,q ) Q(Er ,q ) Q(ηr ,q )
q q
RK(r)q Q(Eq ) Ieff → κ Q(Eq ) Ieff → κ

RK(2)1-Ell 1.63e+00 1.40e+00 2.95e+00 1.81e+00 1.14e−03 1.71e−03 2.85e−03 2.51e+00

RK(3)2-Trap 6.63e−01 1.07e+00 1.73e+00 2.61e+00 6.46e−07 9.75e−07 1.62e−06 2.51e+00
RK(3)2-Ell 3.29e−01 4.90e−01 8.19e−01 2.49e+00 3.13e−07 4.73e−07 7.86e−07 2.51e+00
RK(3)2-Eul 2.41e+00 3.68e+00 6.09e+00 2.52e+00 2.58e−06 3.90e−06 6.48e−06 2.51e+00
q,r q,r
RK(q)r Q(Er ) Q(Eq,r ) Q(ηq,r ) Ieff → 1 Q(Er ) Q(Eq,r ) Q(ηq,r ) Ieff → 1

RK(1)2-Ell 3.29e−01 2.64e+00 2.97e+00 1.12e+00 3.13e−07 2.85e−03 2.85e−03 1.00e+00

RK(2)3-Trap 1.03e−01 1.84e+00 1.74e+00 9.44e−01 9.47e−11 1.62e−06 1.62e−06 1.00e+00
RK(2)3-Ell 3.30e−02 8.54e−01 8.21e−01 9.61e−01 2.97e−11 7.86e−07 7.86e−07 1.00e+00
RK(2)3-Eul 2.78e−01 6.45e+00 6.18e+00 9.57e−01 2.84e−10 6.48e−06 6.48e−06 1.00e+00

Top standard embedded RK(r)q: exact global errors Q(Eq ), Q(Er ,q ), estimated global error Q(ηr ,q ) in Eq. (60), and effectivity index Ieff in Eq. (54).
q
q,r
Bottom reversed embedded RK(q)r: exact global errors Q(Er ), Q(Eq,r ), estimated global error Q(ηq,r ) in Eq. (61) and effectivity index Ieff in
Eq. (59)

A detailed documentation of further results on the the for τi = 1.0 s, whilst for τi = 1.0 × 10−3 s the relation
global convergence analysis for all eight methods with Q(ηr ,q ) ≈ Q(ηq,r ) is achieved.
equidistant step sizes is provided in Table 5. Regarding stan- Figure 2 depicts global error measures vs. time step of
dard embedded RK(r)q we present values for the exact global standard embedded RK-methods, and its reversed methods
errors Q(Eq ), Q(Er ,q ), the estimated global error Q(ηr ,q ) in with equidistant step sizes. The subfigures a, c, e on the
q
Eq. (60) and the effectivity index Ieff in Eq. (54). Regard- left refer to standard embedded RK(r)q with exact global
ing reversed embedded RK(q)r we present the exact global errors Q(Er ,q ), Q(Eq ) and estimated global error Q(ηr ,q ).
errors Q(Er ), Q(Eq,r ), the estimated global error Q(ηq,r ) The subfigures b, d, f on the right refer to reversed embedded
q,r
Eq. (61) and the effectivity index Ieff in Eq. (59). Figure 1 RK-method RK(q)r with exact global errors Q(Eq,r ), Q(Er )
and Figure 2 provide some corresponding graphical illus- and estimated global error Q(ηq,r ). Figure 2d compared to
trations for the goal function in Eq. (60). Figure 1.c) shows Fig. 2c demonstrates the superior convergence results of
the exact global error E q in (46.2) and the estimated global the new methods RK(1)2-Ell, RK(2)3-Eul, RK(2)3-Trap and
error ηr ,q in Eq. (46.3). The corresponding effectivity index RK(2)3-Ell which is documented by its higher slopes.
q
Ieff in Eq. (54) is displayed in Fig. 1.e). Figure 1.d) shows
the exact global error E q,r in (55.1) and the estimated global 6.1.3 Adaptively refined step sizes
error ηq,r in Eq. (55.3). The corresponding effectivity index
q,r
Ieff in Eq. (59) is displayed in Fig. 1.f). Next, results of adaptive step size selection according to
The results in Table 5 and in Fig. 1.e) reveal, that the effec- Eq. (16) based on a posteriori error estimation with embedded
tivity index for standard embedded RK(r)q tends to some RK-methods are presented. The tolerance for the estimated
q
constant, that is Ieff → κ according to Eq. (54). However, error in Eq. (16) is ε = 0.1. Figure 3 exhibits results for all
q
this constant is quite a distance away from 1.0, where Ieff four standard embedded RK-methods on the left hand side
may be as bad as κ ≈ 2.51, that is, only overestimation and its reversed embedded RK-methods on the right hand
occurs in this example. This however is not a general trend side. For the velocity versus time v(t), both in Fig. 3a and in
of standard RK-methods, see [22], where both under- and Fig. 3b, a difference to the exact solution is hardly visible for
overestimation occurs even for the same example. Conse- all RK-methods. Figure 3c shows the exact global error E q
quently, the error estimates for standard embedded RK(r)q in (46.2) and the estimated global error ηr ,q in Eq. (46.3).
are not very accurate or robust. Contrary, the effectivity Figure 3d shows the exact global error E q,r in (55.1) and
q,r
index Ieff for reversed embedded RK(q)r in Table 5 and in the estimated global error ηq,r in Eq. (55.3). Moreover, the
Fig. 1.f) approaches one, meaning it is asymptotically exact. Fig. 3e, f depict the time step lengths τi resulting from the
We point out, that this result holds for all reversed embed- adaptive selection according to Eq. (16).
ded RK-methods, RK(1)2-Ell, RK(2)3-Trap, RK(2)3-Ell and Table 6 summarizes the values for the exact global errors
RK(2)3-Eul. Table 5 further reveals, that the error estimates Q(Er ,q ), Q(Eq ) and the estimated global error Q(ηr ,q ) in
q
Q(ηr ,q ) and Q(ηq,r ) are some distance away from each other Eq. (60) with corresponding effectivity index Ieff in Eq. (54)
for RK(r)q; as well as the exact global errors Q(Er ), Q(Eq,r )

123
Computational Mechanics (2023) 72:1239–1261 1253

Fig. 2 Parachute with viscous damping: results for four standard Q(Er ,q ), Q(Er ) and e estimated global error Q(ηr ,q ) in Eq. (60). Right
embedded RK-methods and its reversed methods with equidistant step for reversed embedded RK(q)r: b, d exact global errors Q(Eq,r ), Q(Er )
sizes. Left for standard embedded RK(r)q: a, c exact global errors and f estimated global error Q(ηq,r ) in Eq. (61)

123
1254 Computational Mechanics (2023) 72:1239–1261

Fig. 3 Parachute with viscous damping: Results for four standard in Eq. (46.3). e time steps τi . Right for reversed embedded RK(q)r: b
embedded RK-methods and its reversed methods with adaptively velocity versus time, d exact global error E q,r in (55.1) and estimated
refined step sizes: left for standard embedded RK(r)q: a velocity versus global error ηq,r in Eq. (46.3). f Time steps τi
time, c exact global error E q in (46.2) and estimated global error ηr ,q

123
Computational Mechanics (2023) 72:1239–1261 1255

Table 6 Parachute with viscous damping: results for four standard occurring in selective laser melting (SLM) or respectively
embedded RK-methods and its reversed methods with adaptively selective electron beam melting (SEBM) in additive manu-
refined step sizes
facturing processes, [24, 29, 38]. Since our focus lies on the
Q(Er ,q ) Q(ηr ,q )
q
RK(r)q Q(Eq ) Ieff performance investigation of different Runge–Kutta methods
the numerical simulations are restricted to a solid phase of a
RK(2)1-Ell 1.67e+00 1.29e+00 2.91e+00 1.74e+00
two-dimensional area as illustrated schematically in Fig. 4.
RK(3)2-Trap 9.63e−01 1.44e+00 2.40e+00 2.49e+00
Within a rectangular sheet, the laser beam follows a sequence
RK(3)2-Ell 6.09e−01 7.98e−01 1.41e+00 2.31e+00
of straight lines representing the letters LTM. The underlying
RK(3)2-Eul 1.54e+00 2.17e+00 3.71e+00 2.41e+00
initial boundary value problem (IBVP) for the temperature θ
q,r
RK(q)r Q(Er ) Q(Eq,r ) Q(ηq,r ) Ieff reads
RK(1)2-Ell 3.58e−01 2.56e+00 2.92e+00 1.14e+00 1. ρcs θ̇(t, x)−λ θ (t, x)=r (t, x), x ∈ , 0<t ≤T
RK(2)3-Trap 1.99e−01 2.61e+00 2.41e+00 9.24e−01 2. θ |∂ = θ R on ∂ (64)
RK(2)3-Ell 9.39e−02 1.51e+00 1.41e+00 9.38e−01 3. θ (t = 0) = θ0 in .
RK(2)3-Eul 1.46e−01 3.88e+00 3.74e+00 9.62e−01
Top, standard embedded RK(r)q: exact global errors Q(Eq ), Q(Er ,q ),
In Eq. (64.1) ρ is the density of the material, cs is the
estimated global error Q(ηr ,q ) in Eq. (60) and effectivity index Ieff
q specific heat capacity, λ is the conductivity and r is the
in Eq. (54). Bottom, reversed embedded RK(q)r: exact global errors heat source. Moreover, = [a × b], a = 400 mm,
Q(Er ), Q(Eq,r ) and estimated global error Q(ηq,r ) in Eq. (61), effec- b = 200 mm, and ∂ describe the spatial domain and its
q,r
tivity index Ieff in Eq. (59)
boundary, respectively, of the rectangular sheet, and is the
Laplace operator, which in Cartesian coordinates x, y reads
and estimated global error Q(ηq,r ) in Eq. (61), with cor- (·) = ∂(·)/∂ x 2 + ∂(·)/∂ y 2 . The Dirichlet boundary con-
q,r
responding effectivity index Ieff for RK(q)r in Eq. (59) dition Eq. (64.2) prescribes a constant temperature θ R at the
for RK(q)r. The first order standard RK-method RK1-Ell boundary. In Eq. (64.3) initial conditions θ0 are prescribed.
has a comparatively large global error Q(Eq ) = 1.67 for In our simulations the temperatures for boundary and initial
standard RK-method. The third order reversed RK-method conditions in (64) are θ R = 172o C and θ0 = 172o C, respec-
RK3-Ell achieves the lowest values for the exact global error tively. The initial time is t0 = α = 0 and the total time is
Q(Er ) = 1.46 · 10−1 . T = β − α = 25 s. The material parameters for the sim-
q
The effectivity indices Ieff of the standard embedded ulation are given in Table 7. In order to model the energy
methods RK(r)q range from 1.74 to 2.49, meaning that in input of the electron beam in the SEBM process, we let b
this example all errors are overestimated with a range that represent the total beam path as illustrated in Fig. 4 and
q,r
is not very robust. Contrary, the range of Ieff is from 0.938

to 1.14, that is, these values have a close range around one,
q
which compared to the range of Ieff appears to be much more xb (t) = x0 + dx ∈ , sb (t) ∈ b , (65)
sb (t)
robust.
represents the actual position of the laser beam along the
Remark 4 Of course, the results in this subsection on adap- path. Here x0 ∈ is the initial position of the laser beam,
tively refined step sizes depend strongly on the maximum related to the point P0 in Fig. 4. Then the energy input of
allowable error ε in Eq. (16). Thereby, the approximation for the electron beam in the SEBM process is modeled as a time
“small” time step sizes as e.g. in (52) or (57), respectively, is dependent volumetric heat source r (t, x) moving along b
not satisfied anymore compared to the previous subsection with power density
on equidistant time steps in the limit τ → 0. As discussed
in [22], adaptive time step control by means of the scheme 1. r (t, x) = r̂ δ(x − xb (t)), where
(16) only controls the local error. Bounding the global error
by means of adaptive strategies requires further techniques, 41Ab Pb 2(x 2 + y 2 ) (66)
2. r̂ = exp − − βpz .
which are not within the scope of this paper, see e.g. [21]. 3.375rb π
3 2.25rb2

Here we have used: the absorption factor Ab , the absorp-

6.2 Laser beam simulation on a plate tion length β p , the laser beam power Pb and the beam radius
rb , see e.g. [28, 29]. For the two-dimensional simplification in
6.2.1 Initial boundary value problem our numerical example we neglect the penetration depth z of
the beam. Moreover, in Eq. (66) the δ-function is being used
The second example is concerned with the temperature dis- to describe a unit point source positioned at the point x, see
tribution on a plate due to a laser beam movement typically e.g. [37]. The total arc-length of the beam scan path illustrated

123
1256 Computational Mechanics (2023) 72:1239–1261

Table 7 Laser beam simulation: parameters for the simulation ning becomes tb = lb /vb = 15.986 s. The total number for
kg J W m the spatial discretization visualized in Fig. 4.b is n h = 1891,
ρ/ cs / λ/ vb / Pb /W Ab /[−] rb /μm see [22] on more details. Details on a reduced total beam
m3 kg K mK s
path length and the corresponding step size selections for a
700 65 0.078 0.05 0.055 0.95 200
convergence analysis are given in Sect. 6.2.2. Contrary to the
previous example, results are presented only for equidistant
step lengths. Results for non-equidistant step lengths based
on adaptive time-step control would give no new insight for
in Fig. 4 is lb = |b | = 799.3 mm, such that with the scan this example compared to those presented recently in [22].
velocity vb in Table 7 the resulting heating time for laser scan-

Fig. 4 Laser beam simulation: general settings

Fig. 5 Complete laser beam simulation: Temperature distribution at four time steps. Note, that the heating time for laser scanning is tb = 15.986 s.
The total time is T = 25 s

123
Computational Mechanics (2023) 72:1239–1261 1257

Table 8 Laser beam simulation: convergence orders Z r ,q , P r ,q , P q = P0 P1 = vb TR = 10.5 mm. A highly accurate solution, sub-
q according to Eq. (52) and Z q,r , P q,r , P r = r according to Eq. (57) sequently denoted as “reference solution”, is generated using
for four standard embedded RK-methods and its reversed methods with
equidistant step sizes
the classical 4th-order Runge–Kutta method in Eq. (67) in
N = 200 equidistant time steps.
RK-Method Table 8 shows that all RK-methods attain (and in most
RK(r)q Z r ,q P r ,q Pq = q
cases even surpass) its theoretical convergence orders accord-
RK(2)1-Ell 1.05e+00 2.15e+00 1.06e+00 ing to Eq. (53) for standard RK(r)q and according to Eq. (58)
RK(3)2-Trap 2.05e+00 2.15e+00 2.05e+00 for reversed RK(q)r. As in the previous example, the new
RK(3)2-Ell 2.04e+00 2.95e+00 2.15e+00 ELDIRK methods RK3-Eul, RK3-Trap and RK3-Ell attain
RK(3)2-Eul 2.06e+00 3.02e+00 2.05e+00 the third order properties, with no need to solve another non-
linear system.
RK(q)r Z q,r P q,r Pr = r
Table 9 provides the exact global errors Q(Eq ), Q(Er ,q ),
RK(1)2-Ell 1.05e+00 1.04e+00 2.15e+00 the estimated global error Q(ηr ,q ) in Eq. (60), and the
q
RK(2)3-Trap 2.05e+00 2.05e+00 3.42e+00 corresponding effectivity index Ieff in Eq. (54). Regard-
RK(2)3-Ell 2.04e+00 2.14e+00 2.99e+00 ing reversed embedded RK(q)r we present the exact global
RK(2)3-Eul 2.06e+00 2.07e+00 3.02e+00 errors Q(Er ), Q(Eq,r ), the estimated global error Q(ηq,r ) in
q,r
Eq. (61) and the effectivity index Ieff in Eq. (59).
All RK-methods (standard and reversed) attain their theoretical conver-
gence orders The results in Table 9 reveal, that the effectivity index for
standard embedded RK(r)q tends to some constant, that is
q
Ieff → κ = 1 according to Eq. (54). This effect, different to
The spatial discretization and the time integration for the the previous example, has not been investigated so far.
IVBP (64) are extensively described in [22]. Thereby, we It might be attributed to the fact, that the matrix A in heat
also account for the effect of order reduction by means of an transfer equation (64) is constant. As expected, the effec-
effective temporal “node-to-node” discretization technique. q,r
tivity index Ieff for reversed embedded RK(q)r in Table 9
Figure 5 provides the resulting temperature distribution at approaches one, meaning it is asymptotically exact. We
four time steps, here generated with the classical 4th-order point out, that this result holds for all reversed embed-
Runge–Kutta method with Butcher-Tableau ded RK-methods, RK(1)2-Ell, RK(2)3-Trap, RK(2)3-Ell and
0 RK(2)3-Eul.
1 1
2 2 Figure 6 depicts errors vs. time step of standard embed-
1 1 ded RK-methods, and its reversed methods with equidistant
B= 2 0 2 . (67)
1 0 0 1 step sizes. The subfigures a, c, e on the left refer to standard
embedded RK(r)q with exact global errors Q(Er ,q ), Q(Eq )
1
6
1
3
1
3
1
6 and estimated global error Q(ηr ,q ). The subfigures b, d, f
on the right refer to reversed embedded RK-method RK(q)r
All other RK-methods introduced above did not show any with exact global errors Q(Eq,r ), Q(Er ) and estimated global
visible difference for this illustrative simulation. The laser error Q(ηq,r ). Figure 6d compared to Fig. 6c demonstrates
beam starts at the upper point P0 of the letter L, introduced the superior convergence results of the new methods RK(2)3-
in Fig. 4, such that the current position of the beam spot can Eul, RK(2)3-Trap and RK(2)3-Ell documented by its higher
be seen from the current status of the letters LTM. The final slopes.
phase from laser scanning at tb = 15.986 s to the total time
T = 25 s shows some redistribution of the temperature field 7 Summary and outlook
as a result of stationary effects.
In this paper we have proposed a general approach to embed
6.2.2 Global convergence investigations for equidistant low order implicit time integration methods into third order
step sizes Runge–Kutta schemes. In order to obtain a systematic rep-
resentation for the higher order methods, we introduce the
This subsection provides convergence investigations for notion of ELDIRK-methods. These are defined to have an
equidistant step sizes analogously to the previous example. explicit last stage in the general Butcher tableau, contrary
Since an analytical solution is not available and in order to to the class of EDIRK methods, which according to Kværnø
reduce the numerical effort, only the initial phase of the et al. [18] have an explicit first stage in the general Butcher
laser beam movement is considered. The total time is set array. The consequence is, that there is no need to solve an
to TR = 0.21 s, such that the distance from the initial additional nonlinear system of equations for the ELDIRK-
node P0 to the final point P1 , illustrated in Fig. 4, becomes method compared to the embedded schemes. This feature

123
1258 Computational Mechanics (2023) 72:1239–1261

Table 9 Laser beam simulation: results for four standard embedded RK-methods and its reversed methods with equidistant step sizes
RK-Method τi = 5.250e−02 s τi = 2.100e−03 s
Q(Er ,q ) Q(ηr ,q ) Q(Er ,q ) Q(ηr ,q )
q q
RK(r)q Q(Eq ) Ieff →κ Q(Eq ) Ieff → κ

RK(2)1-Ell 2.13e−01 5.46e−02 1.97e−01 9.21e−01 6.44e−03 9.17e−05 6.52e−03 1.01e+00

RK(3)2-Trap 2.97e−01 4.73e−03 2.99e−01 1.01e+00 3.82e−04 4.27e−06 3.83e−04 1.00e+00
RK(3)2-Ell 5.65e−02 4.93e−02 3.44e−02 6.09e−01 4.64e−05 3.29e−06 4.61e−05 9.93e−01
RK(3)2-Eul 1.56e−01 2.87e−02 1.63e−01 1.04e+00 1.96e−04 1.63e−06 1.97e−04 1.00e+00
q,r q,r
RK(q)r Q(Er ) Q(Eq,r ) Q(ηq,r ) Ieff → 1 Q(Er ) Q(Eq,r ) Q(ηq,r ) Ieff → 1

RK(1)2-Ell 5.65e−02 1.94e−01 1.97e−01 1.01e+00 4.64e−05 6.54e−03 6.52e−03 9.98e−01

RK(2)3-Trap 5.21e−03 3.02e−01 2.99e−01 9.91e−01 1.17e−07 3.83e−04 3.83e−04 1.00e+00
RK(2)3-Ell 4.91e−02 5.07e−02 3.44e−02 6.78e−01 2.92e−06 4.58e−05 4.61e−05 1.01e+00
RK(2)3-Eul 2.93e−02 1.70e−01 1.63e−01 9.59e−01 1.71e−06 1.97e−04 1.97e−04 9.99e−01

Top, standard embedded RK(r)q: exact global errors Q(Eq ), Q(Er ,q ), estimated global error Q(ηr ,q ) in Eq. (60), and effectivity index Ieff in
q

Eq. (54). Bottom, reversed embedded RK(q)r: exact global errors Q(Er ), Q(Eq,r ), estimated global error Q(ηq,r ) in Eq. (61) and effectivity index
q,r
Ieff in Eq. (59)

holds for linear as well as for nonlinear ordinary differential Open Access This article is licensed under a Creative Commons
equations. Attribution 4.0 International License, which permits use, sharing, adap-
tation, distribution and reproduction in any medium or format, as
The approach has been applied to three prominent low
long as you give appropriate credit to the original author(s) and the
order implicit methods, that is, the implicit Euler-method, source, provide a link to the Creative Commons licence, and indi-
the trapezoidal-rule and the Ellsiepen method. In this way, for cate if changes were made. The images or other third party material
each of the methods we obtain an extended Butcher tableau. in this article are included in the article’s Creative Commons licence,
unless indicated otherwise in a credit line to the material. If material
The new embedded standard RK-methods, RK(3)2-Eul,
is not included in the article’s Creative Commons licence and your
RK(3)2-Trap and RK(3)2-Ell and the reversed methods intended use is not permitted by statutory regulation or exceeds the
RK(2)3-Eul, RK(2)3-Trap and RK(2)3-Ell have been tested permitted use, you will need to obtain permission directly from the copy-
extensively in two numerical examples: Firstly, a parachute right holder. To view a copy of this licence, visit http://creativecomm
ons.org/licenses/by/4.0/.
with viscous damping and, secondly, two-dimensional tem-
perature evolution due to laser beam movement, and they
have been compared to the RK(2)1 method of Ellsiepen [7],
Eq. (1) in this work. For both examples it has been verified, Appendix
that with only two additional function evaluations for RK1-
Eul, the ELDIRK-method RK3-Eul attains the third order Appendix A: Derivation of higher order ELDIRK-
properties. Similarly, third order properties are obtained for schemes
the new ELDIRK-methods RK3-Trap and respectively for
RK3-Ell with only one additional function evaluation for This appendix provides the derivation of higher order
RK2-Trap and respectively for RK2-Ell. We point out, that ELDIRK-schemes to obtain the embedded trapezoidal rule,
this behaviour holds also for nonlinear IVPs. RK(3)2-Trap, and the embedded Ellsiepen method RK(3)2-
Concerning further extensions, being only half implicit, Ell summarized in Table 3.
a stability analysis of all new ELDIRK-methods is of inter-
est. Also, strategies to control the global error, see e.g. [21], A.1. Embedding of the trapezoidal method
should be considered for embedded RK-methods, based on
the proposed global error measures of [22] and recalled in For the second order trapezoidal method the corresponding
this work. RK-coefficients are obtained from Eq. (31) as
Acknowledgements This paper is based on investigations which are
financially supported by the German Research Foundation (Deutsche
c1 =0, c2 =1, a11 =0, a21 =1/2, a22 =1/2. (A.1)
Forschungsgemeinschaft, DFG) under Grant Number MA 1979/30-2.

Funding Open Access funding enabled and organized by Projekt

DEAL. Consequently, with the choice

ĉ3 = 1/2 (A.2)

123
Computational Mechanics (2023) 72:1239–1261 1259

Fig. 6 Laser beam simulation: errors vs. time step for four standard Q(Er ) and e estimated global error Q(ηr ,q ) in Eq. (60). Right reversed
embedded RK-methods and its reversed methods with equidistant step embedded RK(q)r: b, d exact global errors Q(Eq,r ), Q(Er ) and f esti-
sizes. Left standard embedded RK(r)q: a, c exact global errors Q(Er ,q ), mated global error Q(ηq,r ) in Eq. (61)

123
1260 Computational Mechanics (2023) 72:1239–1261

the Eqs. (36), (35) and (33.1) render 5. Butcher JC (2008) Numerical methods for ordinary differential
equations, 3rd edn. Wiley, New York
1 2 1 6. Dormand JR, Prince PJ (1980) A family of embedded Runge-Kutta
b̂1 = , b̂3 = , b̂2 = . (A.3) formulae, J Comput Appl Math 6(1)
6 3 6
7. Ellsiepen P (1999) Zeit- und ortsadaptive Verfahren angewandt auf
With these results and (A.3), Eqs. (38) and (37) render Mehrphasenprobleme poröser Medien. University Stuttgart
8. Eriksson K, Estep D, Hansbo P, Johnson C (1996) Computational
3 1 differential equations. Cambridge University Press, Cambridge
â31 = , â32 = . (A.4) 9. Eriksson K, Johnson C (1991) Adaptive finite element methods for
8 8 parabolic problems: I. A linear model problem. SIAM J Numer
Anal 28:43–77
For convenience, the resulting new Butcher-Tableau for 10. Ernst Hairer SPNa, Wanner Gerhard (1993) Solving ordinary
the embedded trapezoidal method RK(3)2-Trap is given in differential equations I: nonstiff problems. Springer series in
Table 3. computational mathematics, vol 8. 2nd edn. Springer, Berlin, Hei-
delberg
11. Fehlberg E (1968) Classical fifth, sixth, seventh and eighth order
A.2. Embedding of the Ellsiepen method Runge-Kutta formulas with stepsize control. George C. Marshall
Space Flight Center, Huntsville, Alabama, NASA Technical Report
For the second order Ellsiepen method the corresponding R-287
12. Fehlberg E (1969) Low-order classical Runge–Kutta formulas with
RK-coefficients are obtained from Eq. (32) as
step size control and their application to some heat transfer prob-
lems. NASA Technical Report 315
c1 =α, c2 =1, a11 =α, a21 =1−α, a22 =α. (A.5)
13. Fehlberg E (1970) Klassische Runge–Kutta-formeln vierter und
niedrigerer Ordnung mit Schrittweiten–Kontrolle und ihre Anwen-
Consequently, with the choice dung auf Wärmeleitungsprobleme. Computing (Arch Elektron
Rechnen) 6:61–71
ĉ3 = 0 (A.6) 14. Fekete I, Conde S, Shadid JN (2022) Embedded pairs for optimal
explicit strong stability preserving Runge–Kutta methods. J Com-
put Appl Mathe 412(2022):114325
the Eqs. (36) and (35) render 15. Fried I (1969) Finite element analysis of time-dependent phenom-
ena. Am Inst Aeronaut Astronaut J 6:1170–1173
1 2 − 3α 16. Hughes T (1987) The finite element method: linear static and
b̂1 = , b̂2 = . (A.7)
6(α − α 2 ) 6(1 − α) dynamic finite element analysis. Dover
17. Ju X, Mahnken R (2019) Goal-oriented h-type adaptive finite ele-
ments for micromorphic elastoplasticity. Comput Methods Appl
Equation (35) is not defined for ĉ3 = 0. Therefore, to com- Mech Eng 351:297–329
pute b̂3 we use Eq. (33.1) 18. Kværnø A, Nørsett S, Owren B (1996) Runge–Kutta research in
1   Trondheim. Appl Numer Math 22(1):263–277
b̂3 = 4α − 3α 2 − 1 . (A.8) 19. Lambert J (1991) Numerical methods for ordinary differential sys-
6(α − α )
2 tems. The initial value problem. Wiley
20. Mahnken R (2011) Lehrbuch der technischen mechanik, dynamik,
Next, inserting the relations (A.7), (A.8) into Eq. (38) and 2nd edn. Springer, Berlin, Heidelberg
21. Mahnken R (2013) Goal-oriented adaptive refinement for phase
using Eq. (36) renders
field modeling with finite elements. Int J Numer Methods Eng
94:418–440
â31 = α − 1, â32 = 1 − α. (A.9) 22. Mahnken R (2022) New low order Runge-Kutta schemes for
asymptotically exact global error estimation of embedded meth-
For convenience, the resulting new Butcher-Tableau for the ods without order reduction. Comput Methods Appl Mech Eng
third order embedded Ellsiepen method RK(3)2-Ell is given 401:115553
23. Mahnken R, Stein E (1989) Adaptive time-step control in creep
in Table 3. For the choice ĉ3 = 0 in Eq. (A.6), the matrix analysis. Int J Numer Methods Eng 28:1619–1633
coefficients ai j show an amazing simple structure, whilst the 24. Mellor S, Hao L, Zhang D (2014) Additive manufacturing: a frame-
weighting factors are somewhat more involved. work for implementation. Int J Prod Econ 149:194–201
25. Muñoz-Matute J, Pardo D, Calo VM, Alberdi E (2019) Variational
formulations for explicit Runge–Kutta methods. Finite Elem Anal
References Des 165:77–93
26. Oden JT (1969) A general theory of finite elements. II applications.
1. Argyris JH, Scharpf DW (1969) Finite elements in time and space. Int J Numer Methods Eng 1:247–259
Nucl Eng Des 10:456–464 27. Rannacher R (2001) Adaptive Galerkin finite element methods for
2. Bangerth W, Rannacher R (2001) Adaptive finite element tech- partial differential equations. J Comput Appl Math 128(1–2):205–
niques for the acoustic wave equation. J Comput Acoust 9(2):575– 233
591 28. Riedlbauer M, Mergheim J, Steinmann P (2014) Simulation of the
3. Bogacki P, Shampine LF (1989) A 3(2) pair of Runge–Kutta for- temperature distribution in the selective beam melting process for
mulas. Appl Math Lett 2(4):321–325 polymer material. In: 29th International conference of the polymer
4. Butcher JC (1963) Coefficients for the study of Runge–Kutta inte- processing society, PPS 2013, Nuremberg, Germany, 15–19 July
gration processes. J Aust Math Soc 3:185–201 2013. https://doi.org/10.1063/1.4873876

123
Computational Mechanics (2023) 72:1239–1261 1261

29. Riedlbauer D-R (2018) Thermal and thermomechanical modeling 36. Bangerth MGW, Rannacher R (2010) Adaptive Galerkin finite ele-
and simulation of selective beam melting processes. Doctoral The- ment methods for the wave equation. Comput Methods Appl Math
sis. Friedrich-Alexander University Erlangen 10(1):3–48
30. Rüter M (2006) Error-controlled adaptive finite element methods 37. Wheeler N (1997) Simplified production of Dirac delta functions
in large strain hyperelasticity and fracture mechanics. Dissertation, identities. Reed College Physics Department
University of Hannover 38. Wong KV, Hernandez A (2012) A review of additive manufactur-
31. Schmich M, Vexler B (2008) Adaptivity with dynamic meshes for ing. ISRN Mechanical Engineering, vol 2012 (2012), Article ID
space-time finite element discretizations of parabolic equations. 208760
SIAM J Sci Comput 30:369–393
32. Schröder B (2019) Consistent higher order accurate time discretiza-
tion methods for inelastic material models. Dissertation, Kassel
Publisher’s Note Springer Nature remains neutral with regard to juris-
University
dictional claims in published maps and institutional affiliations.
33. Simo JC, Hughes TJ (2006) Computational inelasticity, vol 7.
Springer
34. Szabo B, Babuska I (1991) Finite element analysis. Wiley, New
York
35. Thomée V (2006) Galerkin finite element methods for parabolic
problems, 2nd edn. Springer, Berlin, Heidelberg

123