8th

8th Vienna
Vienna International
International Conference
Conference on
on Mathematical
Mathematical Modelling
Modelling
8th Vienna
8th Vienna International
February International Conference
Conference on Mathematical
on Mathematical Modelling
Modelling
February 18
18 -- 20,
20, 2015.
2015. Vienna
Vienna University
University of
of Technology,
Technology,
Available online
Vienna,
at Vienna,
www.sciencedirect.com
February
Austria 18 - 20, 2015. Vienna University of Technology, Vienna,
February 18 - 20, 2015. Vienna University of Technology, Vienna,
Austria
Austria
Austria

ScienceDirect
IFAC-PapersOnLine 48-1 (2015) 693–694
Taylor
Taylor Series Based Solution of Linear
Taylor Series
Series Based
Based Solution
Solution of
of Linear
Linear
ODE
ODE Systems
Systems and
and MATLAB
MATLAB Solvers
Solvers
ODE Systems and MATLAB Solvers
Comparison
Comparison
∗,∗∗ ∗∗ ∗∗ ∗∗∗
V. Šátek
V. Šátek ∗,∗∗ F. Kocina
∗,∗∗ F. Kocina ∗∗ J. Kunovský
∗∗ J. Kunovský ∗∗ A. Schirrer
∗∗ A. Schirrer ∗∗∗
∗∗∗
V.
V. Šátek
Šátek ∗,∗∗ F.
F. Kocina
Kocina ∗∗ J. J. Kunovský
Kunovský ∗∗ A.A. Schirrer
Schirrer ∗∗∗
∗
∗ IT4Innovations, VŠB Technical University of Ostrava,
∗ IT4Innovations,
∗
IT4Innovations, VŠB Technical University of Ostrava,
17.
17. listopadu
listopadu 15/2172,V
IT4Innovations,
15/2172,
ŠB
ŠB Technical
V708
708 Technical University
33 Ostrava-Poruba,
33 Ostrava-Poruba, of
of Ostrava,
UniversityCzech
CzechOstrava,
Republic,
17. listopadu
17. listopadu 15/2172,
15/2172, 708vaclav.satek@vsb.cz)
708
(e-mail: 33 Ostrava-Poruba,
33 Czech Republic,
Ostrava-Poruba, Czech Republic,
Republic,
(e-mail: vaclav.satek@vsb.cz)
∗∗
∗∗ Brno University
University ofof(e-mail:
(e-mail: vaclav.satek@vsb.cz)
vaclav.satek@vsb.cz)
Technology, Faculty of
of Information
Information Technology,
Technology,
∗∗ Brno
∗∗
Brno University of Technology,
Technology, Faculty
Faculty of
Brno University of
Božetěchova
BožetěchovaTechnology,
2,
2, 612
612 66
66Faculty
Brno,
Brno, of Information
Information
Czech
Czech Republic
Republic
Technology,
Technology,
∗∗∗ Božetěchova
Božetěchova
∗∗∗ Institute 2,
2, 612
612 66
66 Brno,
Brno, Czech
Czech Republic
Republic
∗∗∗ Institute of Mechanics and Mechatronics, Vienna University of
of Mechanics and Mechatronics, Vienna University
∗∗∗ Institute
Institute of Mechanics
of Mechanics
Technology, and Mechatronics,
and
GetreidemarktMechatronics,
9, 1060 Vienna
Vienna
1060 Vienna,
Vienna, University of
University
Austria of
of
Technology,
Technology, Getreidemarkt
Getreidemarkt 9,
9, 1060 Vienna, Austria
Austria
Technology, Getreidemarkt 9, 1060 Vienna, Austria
Abstract: The
Abstract: The Modern
Modern Taylor
Taylor Series
Series Method
Method (MTSM)
(MTSM) is is employed
employed here
here toto solve
solve initial
initial value
value
Abstract:
Abstract:
problems ofofThe
The Modern
Modern
linear ordinaryTaylor
Taylor Series
Series Method
differentialMethod (MTSM)
(MTSM)
equations. An is
is employed
employed
automatic here
here to
computationto solve
solve
of initial value
initialTaylor
higher value
problems
problems of linear ordinary differential equations. An automatic computation of higher Taylor
problems of linear
series terms
series terms anordinary
linear
and an
and efficient,differential
ordinary
efficient, vectorized equations.
differential
vectorized equations.
coding of
coding
An
An automatic
automatic
of explicit
explicit computation
and implicit
and implicit schemesof
computation
schemes
higher
higher Taylor
ofenables
enables Taylor
aa very
very
series
series terms
terms and
and
fast computation
computation of an
an efficient,
efficient,
of the vectorized
vectorized
the solution
solution to coding
coding
to specified of
of explicit
explicit
specified accuracy.
accuracy. For and
and
For aa setimplicit
implicit
set of schemes
schemes
of benchmark problemsaa from
enables
enables
benchmark problems very
very
from
fast
fast computation of the
the solution to
tooutperforms
specified accuracy.
accuracy. Forsolvers.
fast computation
literature,
literature, the MTSM
the of
MTSM solution
significantly
significantly specified
outperforms standard
standard For aa set
set of
solvers.
of benchmark
benchmark
Finally,
Finally,
problems
problems
ideas of
ideas of from
from
parallelizing
parallelizing
literature,
literature,
the MTSM the
the MTSM
MTSM
computations significantly
significantly
are outperforms
outperforms
discussed. standard
standard solvers.
solvers. Finally,
Finally, ideas
ideas of
of parallelizing
parallelizing
the
the MTSM computations are discussed.
the MTSM
MTSM computations
computations are are discussed.
discussed.
© 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords: Ordinary
Keywords: Ordinary differential
differential equations,
equations, Initial
Initial value
value problems,
problems, Taylor
Taylor series,
series, MATLAB
MATLAB
Keywords:
Keywords: Ordinary
Ordinary differential
differential equations,
equations, Initial
Initial value
value problems,
problems, Taylor
Taylor series,
series, MATLAB
MATLAB
1. INTRODUCTION
1. INTRODUCTION The n−th
The n−th order order method method uses uses n n Taylor
Taylor series series terms terms in in the
the
1.
1. INTRODUCTION
INTRODUCTION The
The
explicitn−th
n−th formorder
order method method uses uses n n Taylor
Taylor series series terms terms in in the
the
explicit
explicit form
The “Modern
The “Modern Taylor Taylor Series
Series Method” (MTSM) (MTSM) is is used
used for explicit form form
Series Method” used for
2
h22 [1]
The
The “Modern
“Modern
numerical Taylor
Taylor
solution Series Method”
Method”
of differential
differential (MTSM)
(MTSM)
equations. isMTSM
Theis used for
for
is yyyi+1
i+1 =
= yyii + + hf hf (t (tii ,, yyii )) ++h h2 ff [1] + ·· ·· ··
(tii ,, yyii )) +
numerical solution of equations. The MTSM is h [1] (t
ff [1]
numerical
numerical
based on
based
solution
solution
on aa recurrent of
of differential
differential
recurrent calculation
calculation of equations.
equations.
of the
the Taylor The
The
Taylor series MTSM
MTSM
series terms
terms is
is yi+1 = y
h
+
n
hf (t
i+1 = yii +n hf (tii , yii ) + 2! , y ) + 2!
2!
(t + ·· ·· ··
(tii ,, yyii )) + (1)
(1)
based on a h 2! (1)
based
for each
for each ontimea recurrent
time interval.calculation
recurrent
interval. calculation
An important
An important of
of the
the Taylor
Taylor
part
part
series
series
of the
of the MTSM
MTSM
terms
terms is
is +h
+
n
nf f [n−1]
[n−1] (t
(t
hn! f [n−1] (ti , yi ).
[n−1] i ,
, y
y i ).
). (1)
for each time interval. An important part of the MTSM is +
+ n! f (t ii , yii ).
for
an each
automatic
an automatic time interval.
integration
automatic integration An
integration order important
order setting,
order setting, part
setting, i.e. of
i.e. the
using MTSM
as many is n!
an
an automatic
Taylor integration
series terms
terms the order
as the defined setting,
accuracy i.e. using
i.e. using
using
as
as many
as
requires. many
many
Thus Equation (1) (1) for
n!
for linear
linear systems systems of of ODEs
ODEs in in the
the form form
Taylor
Taylor series
series terms as
as the defined
defined accuracy
accuracy requires.
requires. Thus
Thus Equation
Equation (1) for linear
linear systems of of ODEs
ODEs in in the
the form form
Taylor
it is
it is usual series
usual that terms
that the as the
the computation defined
computation uses accuracy
uses different requires.
different numbers
numbers Thusof
of Equation
y  =
y = Ay
Ay + +(1) b could
b couldfor be rewritten
be systems
rewritten
it
it is
is usual
Taylor usual seriesthat
that the
the computation
terms different uses
computation
for different uses different
stepsdifferent
of constantnumbers
numbers
constant length.of
of y =
y = Ay
Ay + +b b could
could be be rewritten
rewritten h22
Taylor series terms for steps of length.
Taylor
Taylor
The MTSM
The MTSM series
serieshas terms
terms
has been
been
for different
different steps
forimplemented
implemented steps of
of constant
in the
in the constant length.
length.
TKSL software
TKSL software y i+1
y = y
= y ii + h(A ·· y
+ h(A y ii + + b) b) + +h h22 A(Ay
h A(Ay ii + + b)+
b)+
The
The MTSM
MTSM has
has been
been implemented
implemented in
in the
the TKSL
TKSL software
software y
y i+1
i+1 =
= y
y i +
+ h(A
h(A ·· y
y i +
+ b)
b) +
+ 2! A(Ay
2! A(Ay ii + + b)+
b)+ (2)
(2)
package (Kunovský,
package (Kunovský, 1994). 1994). i+1 i h n
n i 2! (2)
package
package (Kunovský,
(Kunovský, 1994). 1994). +h
··· ··· ··· + h
h
n
nAA (n−1)
(n−1)
(n−1) (Ay
(Ay i +
+
2! b) ,
b) , (2)
Several
Several papers
papers focus on
focus on computer
computer implementations
implementations of of the
the ··· + + n!
n! A A (n−1) (Ay i +
(Ay
i + b) b) ,,
Several papers focus
focus on n! i
Several
Taylor series
Taylor papers
series method
method in aacomputer
on
in computer
variable-order
variable-order
implementations
implementations
and
of
of the
and variable-step
variable-stepthe where
where
where
A
A
A
is
is
is
the
the
the
n!
constant
constant
constant
Jacobian
Jacobian
Jacobian
matrix
matrix
matrix
and
and
and
b
b
b
is
is
is
the
the
the
Taylor
contextseries
Taylor (see, method
series method
for instance,in
in a
instance, a variable-order
variable-order
Barrio et et al. and
and variable-step
al. (2005),
(2005), variable-step
the TIDES
TIDES where
constant
constant A is the
right-hand
right-hand constant
side.
side. Jacobian matrix and b is the
context
context (see,
(see, for
for instance, Barrio
Barrio et al. (2005), the
the TIDES constant right-hand
right-hand side. side.
context
software (see, for
implemented instance,in Barrio
Wolfram et al.
(2014),(2005),
or inthe TIDES
Jorba and constant
software
software implemented
implemented in
in Wolfram
Wolfram (2014),
(2014), or
or in
in Jorba
Jorba and
and Vectorized
Vectorized MATLAB
MATLAB code
code of
of explicit
explicit Taylor
Taylor series
series exp-
exp-
software
Zou (2005)).
Zou (2005)). implemented
(2005)). The The in
reduction
The reduction
reduction ofWolfram
of rounding
of rounding(2014),
rounding errors or
errorsin Jorba and
(Rodrı́guez
errors (Rodrı́guez
(Rodrı́guez Vectorized
Vectorized MATLAB
MATLAB code
code of
of explicit
explicit Taylor
Taylor series
series exp-
exp-
Zou Tay
Tay with
with a
a variable
variable order
order and
and variable
variable step
step size
size scheme
scheme
Zou
and (2005)).
and Barrio, 2012)
Barrio, The reduction
2012) of rounding
and utilization
and utilization errors arithmetics
of multiple
of multiple (Rodrı́guez
arithmetics Tay
Tay with
with a
a variable
variable order
order and
and variable
variable step
step size
size scheme
scheme
and Barrio, 2012) and utilization of multiple arithmetics for linear
for linear systems
systems of of ODEs
ODEs (2) (2) has has been been implemented.
implemented.
and Barrio, 2012) andimprove
utilizationthe of multiple arithmetics for linear systems of ODEs
ODEs (2)
(Barrio
(Barrio
(Barrio
et al.,
et
et
al.,
al.,
2011)
2011)
2011)
improve
improve the
the
applicability
applicability
applicability
of Taylor
of
of
Taylor
Taylor for
This
This linear
algorithm
algorithm systems was
was of
compared
compared (2)
on
on aahas
has
set
set
been
been
of
of
implemented.
implemented.
“non-stiff”
“non-stiff” linear
linear
(Barrio
series based et al., 2011)
algorithms.
series based algorithms.
algorithms. improve the applicability of Taylor This
This algorithm
algorithm was
was compared
compared on
on a
a set
set of
of “non-stiff”
“non-stiff” linear
linear
series systems
systems (see
(see Enright
Enright and
and Pryce
Pryce (1987))
(1987)) with
with vectorized
vectorized
series based
based algorithms. systems
systems (see
(see Enright
Enright and
and Pryce
Pryce (1987))
(1987)) with
with vectorized
vectorized
This paper
This paper demonstrates
paper demonstrates
demonstrates that that the
that the MTSM,
MTSM, specialized
specialized to to MATLAB
MATLAB explicit
explicit odeNN
odeNN solvers.
solvers. Benchmarking
Benchmarking results
results
This the MTSM, specialized to MATLAB
MATLAB
are shown
shown in explicit
explicit
in Table odeNN odeNN solvers. runtime
solvers. Benchmarking
Benchmarking results
results
This
directlypaper
directly solving demonstrates
solving linear
solving linear
linear ODE ODE that
ODE systems,the
systems, MTSM,
systems, solvessolves specialized
non-stiff
solves non-stiff and
non-stiff and
andto are
are shown in Table
Table 111 (each
(each reported
(each
reported
reported runtime
runtime
is the
is
is
the median
the
median
median
directly
directly
stiff systems
systems solving very linear
fast ODE systems,
(in explicit
explicit andsolves non-stiff
implicit and
formula- are
value
value shown
of
of 100
100in Table
computations).
computations). 1 (each reported
Ratios
Ratios runtime
of
of computation
computation is the median
times
times
stiff
stiff systems very
very fast
fast (in
(in explicit and
and implicit
implicit formula-
formula- value of 100 computations). Ratios of computation times
stiff systems
tions, respectively) very
respectively) and fast (in explicit
and outperforms and
outperforms standard implicit
standard solversformula-
solvers in in value
ratio
ratio of= 100 computations).
ode23/expT ay
e = ode23/expT ay > 1 indicate faster compu- Ratios
> 1 of
indicate computationfaster times
compu-
tions,
tions, respectively) and ratio e = ode23/expT ay > 1 indicate faster compu-
tions, respectively)
the considered
the considered and outperforms
benchmark
benchmark problems. standard
outperforms
problems. standard solvers
solvers in in ratio
tation
e
tatione of = the
of ode23/expT
the MTSM in
MTSM inayall all>test test1 indicate
cases. Exact
cases. fastersolutions
Exact compu-
solutions
the tation of thethe by MTSM in all all software
test cases.cases. Exact solutions
the considered
considered benchmarkbenchmark problems.problems. tation
were of
were obtained
obtained MTSM
by the Maple
the in
Maple test
software package
package Exact solutions
(Maplesoft,
(Maplesoft,
2. EXPLICIT SCHEME OF OF TAYLOR
TAYLOR SERIES SERIES were
were
2014). obtained
obtained
All solvers’ by
by the
the Maple
Maple
tolerances software
software
were setpackage
package
to obtain (Maplesoft,
(Maplesoft,
relative
2.
2. EXPLICIT SCHEME 2014). All solvers’ tolerances were set to obtain relative
2. EXPLICIT
EXPLICIT SCHEME SCHEME OF OF TAYLOR
TAYLOR SERIES SERIES 2014).
2014).
and
and
All
All
absolute
absolute
solvers’
solvers’
tolerances
tolerances
tolerances
tolerances
of
of 10
10
−4
−4were
werewith set
set to
to
respect obtain
obtainto the relative
relative
exact
and absolute tolerances of 10 −4 with respect to the exact
−4 with respect to the exact
In this
In this article,
article, we we have
have focused
focused on on effective
effective solution
solution ofof lin-
lin- and absolute
solutions.
solutions. tolerances of 10 with respect to the exact
In
earthis
In this
systems article,
article, we
we have
of ODEs
ODEs have focused
focused
using Tayloron effective
on series
effective
series solution
Theof
solution
scheme. lin-
ofbest-
lin-
best- solutions.
solutions.
ear
ear systems
systems of using Taylor scheme. The
ear
known
knownsystems and of
and of
most
most
ODEs
ODEs using
using Taylor
accurate
accurate Taylor
method
method
series
series scheme.
scheme. The
of calculating
of calculating Theaa best-
best-
new
new 3.
3. IMPLICIT
IMPLICIT SCHEME SCHEME OF OF TAYLORTAYLOR SERIES SERIES
known
known
value of and
and most
most
of aa numerical accurate
accurate
numerical solution
solution ofmethod
method
of ordinary of
of calculating
differentialaaequa-
calculating
ordinary differential new
new
equa- 3.
3. IMPLICIT
IMPLICIT SCHEME SCHEME OF OF TAYLORTAYLOR SERIES SERIES
value
value
tion yyof
value
tion  =
of a
a numerical
 = f
numerical
f (t,
(t, y),
y), y(0) solution
y(0) = yy00of
solution
= isordinary
of is
ordinary
to construct
to
differential
differential
construct equa-
equa-
the Taylor
the Taylor The implicit
The implicit Taylor Taylor series series scheme scheme for for linear
linear systems systems of of
tion y = f (t,
= f (t,et y),
y),al., y(0) = y is to construct the Taylor The implicit Taylor
Taylor series scheme
scheme for for linear
linear systems systems of
tion y(Hairer al., y(0) = y0 is to construct the Taylor
series 1987). 0 The
ODEs implicit
are constructed series
as follows: of
series
series (Hairer et 1987). ODEs are constructed as follows:
series (Hairer
(Hairer et et al.,
al., 1987).
1987). ODEs
ODEs are are constructed
constructed as as follows:
follows:
2405-8963 © 2015,
Copyright IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Copyright © © 2015,
2015, IFAC
IFAC 693
693
Copyright
Peer review©
Copyright 2015,
©under IFAC
2015,responsibility
IFAC 693
of International Federation of Automatic
693Control.
10.1016/j.ifacol.2015.05.210
MATHMOD 2015
694
February 18 - 20, 2015. Vienna, Austria V. Šátek et al. / IFAC-PapersOnLine 48-1 (2015) 693–694

    
Table 1. Median computation time: ex- m
 m
plicit Taylor expTay and MATLAB explicit y i+1 =  Aj  A + I y i +  Aj  b (5)
odeNN solver comparison j=1 j=1

ode23 ode45 ode113 expTay where I is the identity matrix.

[s] [s] [s] [s] ratioe
A1 0.00497 0.00537 0.00751 0.000831 5.98 5. CONCLUSION
B2 0.00633 0.00758 0.0128 0.00218 2.9
C1 0.00653 0.00574 0.0111 0.00114 5.72
C2 0.01 0.0147 0.0277 0.00651 1.54 The Taylor series scheme is highly efficient in solving linear
C3 0.00636 0.00805 0.0156 0.003 2.11 ODEs. It significantly outperforms standard solvers on
C4 0.00679 0.00836 0.0166 0.00359 1.89 the considered benchmark problems. Results for double
precision arithmetics and a maximum Taylor series order
h2 of 90 have been shown. Multiple arithmetics is needed for
y i+1 = y i + h(Ay i+1 + b) − A(Ay i+1 + b)−
2! (3) higher orders. Future studies will address the efficiency
(−h)n (n−1) and scalability of MTSM ODE solvers in different paral-
··· − A (Ay i+1 + b) .
n! lelization architectures.
Implicit Taylor series method with recurrent calculation ACKNOWLEDGEMENTS
of Taylor series terms and Newton method (impTay)
based on (3) was implemented in MATLAB using vector- This paper has been elaborated in the framework of the
ization. The Jacobian matrix is computed using Broyden’s project New creative teams in priorities of scientific re-
method. search, reg. no. CZ.1.07/2.3.00/30.0055 (as well as the
A benchmark problem set of “stiff” linear ODEs from the IT4Innovations Centre of Excellence CZ.1.05/1.1.00/
Enright and Pryce (1987) was used for tests. Comparisons 02.0070), supported by Operational Programme Educa-
of the problems A1, A3, A4, and B1-B5 whose analytic tion for Competitiveness and co-financed by the European
solutions are known (from the Maple software package Social Fund and the state budget of the Czech Republic.
(Maplesoft, 2014)) have been completed. The simulated The paper includes the solution results of the international
intervals were adopted from Enright and Pryce (1987), and AKTION research project Number 69p22 and the internal
the integration time step was set to the entire time interval BUT projects FIT-S-12-1 and FIT-S-14-2486.
(just 1 integration step was needed). Relative and absolute
tolerances for the computations were again set to 10−4 . REFERENCES
Comparisons of MATLAB “stiff” odeNNs solvers with
Barrio, R., Blesa, F., and Lara, M. (2005). VSVO
impTay are shown in Table 2. High ratios of computation
Formulation of the Taylor Method for the Numerical
times ratioi = ode15s/impT ay show that the MTSM
Solution of ODEs. In Computers and Mathematics with
method significantly outperforms the standard solvers.
Applications, volume 50, 93–111.
Table 2. Time of solutions: implicit Taylor Barrio, R., Rodrı́guez, M., Abad, A., and Blesa, F. (2011).
impTay and MATLAB implicit odeNNs Breaking the limits: The taylor series method. Applied
solvers comparisons Mathematics and Computation, 217, 7940–7954.
Enright, W.H. and Pryce, J.D. (1987). Two fortran
ode15s ode23s ode23tb impTay packages for assessing initial value methods. In ACM
[s] [s] [s] [s] ratioi Trans. Math. Softw., volume 13, 1–27. ACM.
A1 0.0605 0.169 0.101 0.0003 194.6 Hairer, E., Nørsett, S.P., and Wanner, G. (1987). Solving
A3 0.085 0.243 0.144 0.00001 263 Ordinary Differential Equations I. vol. Nonstiff Prob-
A4 0.111 0.478 0.192 0.0003 294.8
lems. Springer-Verlag Berlin Heidelberg. ISBN 3-540-
B1 0.268 1.473 0.8 0.0003 244.4
B2 0.069 0.285 0.134 0.00003 172.4
56670-8.
B3 0.073 0.308 0.146 0.00003 211.4 Jorba, A. and Zou, M. (2005). A software package for the
B4 0.117 0.549 0.242 0.00003 348.4 numerical integration of ODE by means of high-order
B5 1.155 1.529 0.664 0.00003 3306.7 Taylor methods. In Exp. Math., volume 14, 99–117.
Kunovský, J. (1994). Modern Taylor Series Method. FEI-
VUT Brno. Habilitation work.
4. PARALLEL IMPLEMENTATION Maplesoft (2014). MAPLE software. URL
http://www.maplesoft.com [online].
As can be seen from (2), each term of Taylor Series for MathWorks, T. (2014). MATLAB and Simulink software.
a linear system can be computed independently. So their URL http://www.mathworks.com [online].
computation can be distributed into multiple computation Rodrı́guez, M. and Barrio, R. (2012). Reducing rounding
units (utilizing a distributed memory architecture). Hence errors and achieving brouwers law with taylor series
thread j ∈ {1 . . . m} evaluates method. Applied Numerical Mathematics, 62, 1014–
n
1024.
m −1
 Wolfram (2014). MATHEMATICA software. URL
hmk+j
Aj = Amk+j−1 (4) http://www.wolfram.com/mathematica/ [online].
(mk + j)!
k=0
and the final sum is computed afterwards. Therefore
expression (2) can be transformed to

694