Sie sind auf Seite 1von 3

The Convergence of Semi-Implicit Numerical

Methods
Aleksandra V. Tutueva#, Ekaterina A. Rodionova#, Mariia P. Baidina#, Anastasiia V. Kavunskaia#, Maria N. Kozak*
#
Department of Computer Aided Design, *Youth Research Institute
Saint-Petersburg Electrotechnical University "LETI"
St. Petersburg, Russia
avtutueva@etu.ru

Abstract—Rapid development of semi-implicit and semi- affect the performance of the ODE solver. The main aim of this
explicit integration techniques allowed to create relatively stable research is to investigate the convergence of the semi-implicit
and efficient extrapolation and composition ODE solvers. numerical integration methods through the study of linear
However, there are several shortcomings in semi-implicit systems of differential equations with different properties.
approach that should be taken into consideration while solving
non-Hamiltonian systems. One of the most disturbing features of The paper is organized as follows. In Section II we describe
semi-implicit integration methods is their low convergence, the considered linear systems and methods under investigation.
which, in theory, can significantly affect the performance of the Section III presents the results of convergence estimation. In
solver. In this paper we study the convergence of ODE solvers Section IV the technique for improving the convergence of the
based on of semi-implicit integrators. The linear differential semi-implicit numerical integration methods is proposed.
equations of different order are considered as a test systems. The Section V gives conclusions and discussion.
dependence between method convergence and system order is
revealed. The comparison with traditional ODE solvers is given. II. INVESTIGATED SYSTEMS AND METHODS
We experimentally show that the semi-implicit algorithms may The semi-implicit CD-methods were proposed in [6–8]
exhibit a low convergence for a certain systems. We also propose
based on early ideas of Cromer and Verlet [10–12]. The class
a technique to reduce this effect – the introduction of correction
coefficient and give an experimental evaluation of this approach.
of such algorithms can be described using so-called
commutation matrices [8]. In this study we consider the
Keywords—numerical integration; semi-implicit method; initial common version of CD methods with the direct order of
value problem; convergence; ODE solvers; linear system calculating equations in the first part of a composition scheme,
and the reverse in the second. We simulate linear systems of
I. INTRODUCTION different dimension with the behavior of damped and
Computer simulation is one of the most important and undamped harmonic oscillators. For the matrix dimension of 3
widely used approach in scientific research. It implies analysis the coefficients of such systems can be choose as follows
of a mathematical model using numerical methods. The
dynamics of real processes or phenomena is often described by  −1 − 3 − 3
systems of ordinary differential equations (ODE) [1–3]. Most  
 − 0. 5 − 2 .5 − 3  (1)
of such systems comprise nonlinear equations, which are  2
difficult or practically impossible to solve analytically in  3 − 1 
general form. Therefore, simulation is performed using
numerical integration methods. The choice of a discrete and
operator significantly affects the accuracy of the resulting
solution and the simulation speed [4]. Over the last two
 0 − 1 − 1
decades, many new approaches to the numerical integration of  
ODEs have been proposed [5]. One of the classes of the newly  1 0 − 1. (2)
developed algorithms is semi-implicit numerical integration 1 1 0 

methods. So-called geometric integrators, they are common as
a simulation tool for Hamiltonian systems [4]. However, recent We investigate the convergence of the semi-implicit CD-
studies in the field of numerical integration have shown that method in comparison with the well-studied explicit (EMP)
semi-implicit algorithms effectively solve nonlinear ODE and implicit (IMP) midpoint methods. The simulation time t in
systems, including systems with chaotic dynamics [6–8], both all computational experiments was set 5 s. and the integration
conservative and dissipative. step h was set 0.01 s.
Semi-implicit methods offer high computation speed and
accuracy in comparison with implicit and explicit methods. III. EXPERIMENTAL COMPARISON OF INTEGRATION METHODS
However, one of the most important features of these methods Since CD-methods have already proven their effectiveness
is their questionable convergence [7, 9] that can significantly in chaotic system simulation, the study of their applying to

366 978-1-7281-0339-6/19/$31.00 ©2019 IEEE


linear systems and their convergence when solving dissipative
systems is of interest. We consider systems with different
dimension matrices and calculate the absolute truncation error
in the last point. The results of the experiment are shown in
Fig. 1. As it was expected, the best convergence in Hamiltonian
problems simulation can observed for the solution obtained by
the semi-implicit algorithm.

Fig. 3. The convergence of different ODE solvers when the damping factor is
varied

One can see that the convergence of the CD-method


improves with increasing of damping factor value. A decrease
in values of coefficients located along the main diagonal of the
matrix precise the additional members of the Taylor series.
IV. IMPROVING CONVERGENCE OF SEMI-IMPLICIT METHODS
Fig. 1. The dependence of truncation error of different ODE solvers from the
dimension of conservative linear systems The convergence and accuracy of geometrical numerical
methods can be improved by introducing a correction
Then we investigate the systems of different order with coefficient into the integration scheme at each integration
energy dissipation (Fig. 2). One should note that the step [13]:
convergence of the CD-method is much worse than the
convergence of explicit and implicit integration methods when x n +1 = c(x n + hf (x n , x n +1 )), (3)
solving dissipative ODE systems with a full matrix of
coefficients.
where x is a vector of state variables, c is a correction
coefficient.
The choice of the value c in (3) can be carried out using
optimization methods. The optimization criterion here is the
minimization of the truncation error. For matrices with fixed
parameters obtained in the previous experiment the correction
coefficients are presented in Table 1.

TABLE I. CORRECTION COEFFICIENTS

Coefficient value Damping factor


0.999998621 0.1

0.99999655 0.2

1.000006019 0.3

1.000001478 0.4
Fig. 2. The dependence of truncation error of different ODE solvers from the
dimension of disspative linear systems 1.000000751 0.5

1.000000416 0.6
It can be assumed that the source of the low convergence is
additional members of the Taylor series that are generated by 1.00000023 0.7
the semi-implicit principle of calculations. To prove this
0.999999985 0.8
hypothesis we carried out a computational experiment that we
iteratively changed the values of coefficients of matrix (1) to 0.999999866 0.9
the values of the coefficients of matrix (2). We multiplied the
0.999999802 1
difference of two matrices by the damping factor varied from 0
to 1. Then we simulated the resulting systems and calculated Using the calculated values of the correction coefficients,
the last point error. Figure 3 represents the experimental we repeated the abovementioned experiments and studied the
results.

367
methods’ convergence depending on the damping factor. The errors of explicit and implicit midpoint methods. In this case,
obtained results are shown in Fig. 4. semi-implicit calculations preserve their advantages in stability
and efficiency.
In this paper, we discussed the most common version of the
CD-methods. The dependence of the solution convergence on
commutation matrix will be the topic of further research.
REFERENCES
[1] E.N. Lorenz, "Deterministic nonperiodic flow," Journal of the
atmospheric sciences, vol. 20.2, 1963, pp.130-141.
[2] L.I. Dublin and A.J. Lotka, "On the true rate of natural increase: As
exemplified by the population of the United States, 1920," Journal of the
American statistical association, vol. 20.151, 1925, pp. 305-339.
[3] D.N. Butusov, V.Y. Ostrovskii, and D. O. Pesterev, "Numerical analysis
of memristor-based circuits with semi-implicit methods," Young
Researchers in Electrical and Electronic Engineering, IEEE Conference
of Russian, IEEE, 2017.
[4] E. Hairer, C. Lubich, and G. Wanner, Geometric numerical integration:
Fig. 4. The convergence of different methods when the damping factor is structure-preserving algorithms for ordinary differential equations,
varied. CD method is applied with correction coefficients from table I. vol. 31. Springer Science & Business Media, 2006.
[5] J.C. Butcher, Numerical methods for ordinary differential equations,
One can see that the proposed approach significantly John Wiley & Sons, 2016.
improves the convergence of semi-implicit method when [6] D.N. Butusov, V. S. Andreev, and D. O. Pesterev, "Composition semi-
simulating the dissipative linear systems. For nearly all implicit methods for chaotic problems simulation," Soft Computing and
considered systems, the last point truncation error of solution Measurements (SCM), 2016 XIX IEEE International Conference, IEEE,
obtained by the semi-implicit solver was less than the error of 2016.
the explicit and implicit integration methods. The main [7] D.N. Butusov, T. I. Karimov, and V. Y. Ostrovskii, "Semi-implicit ODE
solver for matrix Riccati equation," Young Researchers in Electrical and
disadvantage of this approach is that for systems with identical Electronic Engineering Conference (EIConRusNW), 2016 IEEE NW
matrices the values of correction coefficients are different for Russia, IEEE, 2016.
various integration steps. However, proposed technique of [8] D.N. Butusov, A. V. Tutueva, and E. S. Homitskaya, "Extrapolation
increasing convergence can be successfully applied in digital Semi-implicit ODE solvers with adaptive timestep," Soft Computing and
signal processing or control systems based on semi-implicit Measurements (SCM), 2016 XIX IEEE International Conference, IEEE,
discrete models. For such systems the timestep and other 2016.
parameters are usually known and do not change during the [9] M. Liu, W. Cao, and Z. Fan, "Convergence and stability of the semi-
time. implicit Euler method for a linear stochastic differential delay equation,"
Journal of Computational and Applied Mathematics, vol. 170.2, 2004,
pp. 255-268.
V. CONCLUSION
[10] E. Hairer, C. Lubich, and G. Wanner, "Geometric numerical integration
The main purpose of this study was to investigate the illustrated by the Störmer–Verlet method," Journal of Acta numerica,
convergence of semi-implicit CD-methods when various linear vol. 12, 2003, pp. 399-450.
systems are simulated. The truncation errors of the solutions of [11] A.Cromer, "Stable solutions using the Euler approximation," American
damped and undamped linear ODE systems were obtained and Journal of Physics, vol. 49.5, 1981, pp. 455-459.
analyzed. For conservative linear systems the semi-implicit [12] L. Verlet, "Computer" experiments" on classical fluids. I.
Thermodynamical properties of Lennard-Jones molecules," Journal of
method possesses the best convergence. Due to the extra Physical review, vol. 159.1, 1967, p. 98.
members of the Taylor series appearing in semi-implicit
[13] K. G. Zhukov and D. N. Butusov. "Korrekciya pogreshnosti resheniya
calculations, the efficiency of semi-implicit integration differencial'nyh uravnenij vtorogo poryadka." Nauchno-tekhnicheskie
decreases when systems with energy dissipation are simulated. vedomosti Sankt-Peterburgskogo gosudarstvennogo politekhnicheskogo
To reduce this effect we proposed a technique based on the universiteta. Informatika. Telekommunikacii. Upravlenie, 2010,
correction coefficient. Experiments show, that truncation error vol. 6.113, in Russian.
of the corrected semi-implicit method is comparable with the

368

Das könnte Ihnen auch gefallen