Forward and non-forward symplectic integrators in solving classical dynamics problems
Siu A. Chin

TL;DR
This paper compares forward and non-forward symplectic integrators for classical dynamics, highlighting their advantages in solving time-irreversible and time-reversible equations, with detailed analysis on various integrators and the 2D Kepler orbit.
Contribution
It provides a detailed comparison of forward and non-forward fourth-order symplectic integrators across different force evaluations, emphasizing optimization strategies.
Findings
Forward integrators are essential for time-irreversible equations.
Non-forward integrators can be optimized by minimizing backward time steps.
All non-forward integrators were tested on the 2D Kepler orbit.
Abstract
Forward time step integrators are splitting algorithms with only positive splitting coefficients. When used in solving physical evolution equations, these positive coefficients correspond to positive time steps. Forward algorithms are essential for solving time-irreversible equations that cannot be evolved using backward time steps. However, forward integrators are also better in solving time-reversible equations of classical dynamics by tracking as closely as possible the physical trajectory. This work compares in detail various forward and non-forward fourth-order integrators using three, fourth, five and six force evaluations. In the case of solving the 2D Kepler orbit, all non-forward integrators are optimized by simply minimizing the size of their backward time steps.
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsNumerical methods for differential equations · Superconducting Materials and Applications · Computational Fluid Dynamics and Aerodynamics
Forward and non-forward symplectic
integrators
in solving classical dynamics problems
Siu A. Chin
Department of Physics, Texas A&M University, College Station, TX 77843, USA
Abstract
Forward time step integrators are splitting algorithms with only positive splitting coefficients. When used in solving physical evolution equations, these positive coefficients correspond to positive time steps. Forward algorithms are essential for solving time-irreversible equations that cannot be evolved using backward time steps. However, forward integrators are also better in solving time-reversible equations of classical dynamics by tracking as closely as possible the physical trajectory. This work compares in detail various forward and non-forward fourth-order integrators using three, fourth, five and six force evaluations. In the case of solving the 2D Kepler orbit, all non-forward integrators are optimized by simply minimizing the size of their backward time steps
I Introduction
Many physical evolution equations are of the form
[TABLE]
where and are noncommuting operators. Important examples include the imaginary time Schrödinger equation
[TABLE]
and the Fokker-Planck equation
[TABLE]
Because the diffusion kernel cannot be evolved backward in time, both of these are time-irreversible evolution equations. Aside from these obvious examples of (1), any pair of equations of the form
[TABLE]
can also be casted into the form (1). This is because the evolution of a general function (including and themselves) can be formulated as
[TABLE]
from which one can identify
[TABLE]
Classical Hamiltonian dynamics corresponds to
[TABLE]
and the resulting evolution (4) is time-reversible.
The generic evolution equation (1) can be solved iteratively
[TABLE]
by approximating the short time evolution operator to any order in via
[TABLE]
assuming that the effect of and can be computed exactly. The set of coefficients are determined by the order condition. For classical dynamics (6), every factorizations of the form (9) produces a symplectic integratorruth83 ; neri87 ; cre89 ; fr90 ; yos90 ; wis91 ; yos93 ; hairer02 ; mcl02 ; lr04 which is an ordered sequence of alternating displacements of and preserving Poincaré invariants. For periodic motion, their energy errors are bounded and periodic, in contrast to explicit Runge-Kutta type algorithms whose energy error grows linearly with the number of periodsshita ; gladman . However, for solving time-irreversible evolution equations such as (2) or (3) with , the Green’s function
[TABLE]
is the diffusion kernel only if is positive. If were negative, then the kernel is unbound and there is no way of simulating the diffusion process backward in time.
Historically, symplectic integrators were developed extensively for use in classical dynamicsruth83 ; neri87 ; cre89 ; fr90 ; yos90 ; wis91 ; yos93 ; hairer02 ; mcl02 ; lr04 . Since classical dynamics is time-reversible, there was no impetus for demanding that all be positive. Moreover, Shengsheng89 and Suzukisuzuki91 have proved that all factorizations of the form (9) beyond second order must contain some negative coefficients in the set . Goldman and Kapergoldman96 later proved that for factorizations of the form (9) beyond second order, both operators must have at least one negative coefficient. Thus all conventional splitting schemes beyond second order must contain some negative coefficients and none can be used to solve time-irreversible problems. Because of the Sheng-Suzuki theorem, it is also difficult to see how one can devise all positive coefficients, forward time-step algorithms.
The operator product (9) has the general Campbell-Baker-Hausdorff expansion,
[TABLE]
where all the error coefficients , , , etc., are calculable functions of , in particular,
[TABLE]
In order for the product to be consistent with the original evolution operator, the coefficients must satisfy the above constraints with and . Forcing the remaining error coefficients to vanish results in order conditions that must satisfy. It is easy to force . Any left-right symmetric product will do. For example,
[TABLE]
produces the following second order symplectic algorithm according to (6):
[TABLE]
where the last numbered variables are the updated variables. Thus any symmetric splitting with will result in at least a second order algorithm. The surprise is that, as first shown by Shengsheng89 , beyond second order a general sum of products of the form (11) is incompatible with having positive coefficients. More specifically, Suzukisuzuki91 shown that the two error coefficients and cannot both be forced to zero for positive coefficients . Since Takahashi and Imadati84 have shown that is a local potential function when solving the imaginary time Schrödinger equation, Suzuki suggestedsuzu95 that this error commutator be kept and ways be found to eliminate .
Following up on Suzuki’s suggestion, this author derived three simple fourth-order forward algorithmschin97 in 1997 and demonstrated their efficiency in solving Kepler’s orbit. Interestingly, it was found that classically produces a force (also with potential ) first derived by Ruthruth83 via canonical transformations. The two forward schemes and derived in Ref.chin97 were also known to Suzukisuzu96 based on McLachan’s resultmcl952 on slightly perturbed Hamiltonians. However, Suzuki did not implement them to do any calculation.
Since 1997, fourth-order forward algorithms have been widely applied to time-irreversible systems such as the Fokker-Planck equation in deriving the first fourth-order Langevin algorithmfor01 , the Kramers equationfor012 for describing stochastic dynamics, the Diffusion Monte Carlo algorithmfor012 ; lest05 for solving quantum many-body ground states, the grid based imaginary time Schrödingerchin013 equation in doing density functional calculations, Path-Integral-Monte Carlo methodsjang01 ; chin04 ; sak04 ; cuer05 ; bon06 ; mezz06 for computing the quantum trace at finite temperature, short time evolved wave functionschincif03 for doing variational quantum many-body calculations, the Gross-Pitaevskii equation for describing a Bose-Einstein condensate in a rotating anisotropic trapchin057 and electrons in a magnetic field confined by quantum dotschin056 and ringschin064 . These forward fourth-order algorithms are far more accurate than second order algorithms and allow very large step sizes to be used.
While forward algorithms are indispensable for solving time-irreversible equations, from their inceptionchin97 they have also been shown to be efficient in solving time-reversible equations. In comparison with explicit Runge-Kutta algorithms and conventional non-forward symplectic integrators, forward algorithms have been shown to be superior in solving the Kepler problemchin97 ; chin2000 ; chin054 , gravitational few-body problemschin052 , the real time Schrödinger equationchinc01 , the real time Schrödinger equation in a laser fieldchinc02 ; bay03 ; gold04 and specially the radial Schrödinger equationchin061 . The incorporation of the commutator in forward algorithms has also inspired a new class of higher order gradient symplectic integratorschinc02 ; ome02 ; ome03 ; ome06 . These new algorithms, through not forward beyond fourth-order, have less backward steps at higher orders.
The reason why forward integrators are also better in solving time-reversible, classical dynamics problems is not well understood. From the perspective of the operator product approximation (9), as long as the second-order error terms are zero, any symmetric factorization scheme will be fourth-order; it should not matters whether these error terms are forced to zero with positive or negative coefficients. However, just as in the discussion of time-reversible and time-irreversible algorithms, one must move beyond the purely algebraic discussion of factorization schemes to examine how the resulting algorithms produce the solution of any particular equation. This work uncovers a crucial difference between forward and non-forward schemes when they are implemented as integrators for solving classical dynamics problems. For classical dynamics, if the trajectory is the exact solution to (4), then the force is evaluated only along the trajectory. As will be shown below, non-forward algorithms, when compared to forward algorithms at the same finite step size , evaluate the force at intermediate positions far from the trajectory. As one reduces the size of the negative time steps, one also reduces the distance of these force evaluation points from the trajectory. In all cases examined, non-forward algorithms are improved by simply reducing the size of their backward time steps, allowing the force to be evaluated more closely along the trajectory. One can argue that these intermediate force evaluation points are not the trajectory outputs of the integrator and their placements are not required to be on the trajectory. That is correct. However, since the exact trajectory is determined only by forces evaluated on the trajectory, any unnecessary force evaluation off the trajectory is just “wasteful”. One then must reduce the time step to bring the force evaluation points closer to the trajectory. This is exactly what is observed in the following comparisons. To achieve the same accuracy, non-forward integrators must use smaller time steps. The question here is not about correctness; it is about efficiency. In the following sections we will compare in detail both types of integrators with three, four, five and six force evaluations.
II Integrators with three-force evaluations
We will begin by comparing the efficiency of integrators in solving the 2D Kepler problem defined by the Hamiltonian
[TABLE]
(Henceforth, we will always normalize the Hamiltonian with kinetic energy and .) This problem is an excellent benchmark because one can gauge an integrator’s performance not just by its energy error but also by its orbital precession errorgladman ; chin2000 ; chin054 . The latter is a more direct measure of the accuracy of the computed orbit.
There are basically three fourth-order integrators that required only three-force evaluations: 1) the non-forward Forest-Ruth (FR) integratorcre89 ; fr90 ; yos90 ,
[TABLE]
where
[TABLE]
[TABLE]
with defined by
[TABLE]
corresponding to an effective force
[TABLE]
Transcribing each operator in (18) yields the integrator
[TABLE]
Starting with initial values and , the updated variables are and . Algorithm A only requires two evaluations of the force and one evaluation of the force gradient. Recently, Omelyanome06 has suggested that the effective force (20) can be evaluated by extrapolation
[TABLE]
The resulting integrator remains angular momentum and phase-volume conserving and is nearly indistinguishable from a fully symplectic integrator when solving the Kepler problem. We will denote this use of an extrapolated effective force (22) as algorithm A′. Algorithm A′ only requires three force evaluations. 3) The Runge-Kutta-Nystrom (RKN) integratorbat99 . If one expresses and directly in terms of and , then (21) reduces to
[TABLE]
This is the form of the RKN algorithm, with , , but with . The conventional RKN algorithm is defined by , and , where
[TABLE]
are the estimated midpoint and final position respectively.
Fig.1 shows the fourth-order energy error coefficients of these four algorithms at step size , where is the period of an highly eccentric orbit with initial values , and eccentricity . The error coefficient is obtained by dividing the energy error by at smaller and smaller until a convergent curve emerges independent of . The curve is further normalized by the initial energy. Thus each algorithm has a characteristic error coefficient, its error “fingerprint”, in solving the Kepler orbit. For symplectic algorithms, this convergence is already set in when . For the RKN algorithm, the energy error curve after the mid period keep on lowering with decreasing step size, showing no sign of convergence at finite . The error only spikes at mid period near the pericenter point. Non-symplectic integrator such RKN are characterized by an irreversible increase in the energy error after each period. The error spike of the FR algorithm is nearly ten times as large as that of algorithm A. The extrapolated gradient algorithm A′ closely matches that of A.
Since all symplectic integrators have periodic energy errors, the energy error is not the most critical benchmark. The orbital precession error, as measured by the rotation of Laplace-Runge-Lenz (LRL) vectorchin2000 ; chin054 ; chin067 , is more discriminating. Fig.2 shows the rotation angle of the LRL vector along the trajectory of the particle. The corresponding error coefficient is again extracted by dividing by . In this case, the RKN integrator shows the same convergence as the symplectic integrators. Thus the precession error coefficient is well-defined and irreversible for all algorithms. The FR integrator’s error is three times as large as that of RKN and 10 times as large as integrator A and A′. When the force gradient is extrapolated, the energy error remains periodic, but the precession error can differs substantially. Here A′’s error is smaller, but in other cases it may not be. We will revisit this point later.
In order to understand the poor performance of the FR integrator, we track its approach toward the pericenter point 3 at two time steps earlier at point 1, as shown in Fig.3. The time step used here is . The FR integrator consists of three applications of , resulting in three overlapping triangles. The first application of begins at position 1, evaluates the force at F1, and lands at 1a. This is the triangle 1-F1-1a. The application of the backward substep begins at 1a, evaluates the force at F2 and brings trajectory back past the starting point to . This is the backward triangle 1a-F2-1b. The final application of begins at 1b, evaluates the force at F3 and lands the trajectory at position 2. This is the final triangle 1b-F3-2. Starting at position 2, the algorithm repeats its three overlapping triangles and zigzags its way to point 3. It is remarkable that FR can achieve fourth-order accuracy by such a tremendous zigzagging. Notice that as the FR algorithm tries to turn the “corner” near the pericenter 3, all of its force evaluation points are far off the trajectory. In Fig.4, the positions where each algorithm calculates the force are plotted. The backward loops executed by FR far off the trajectory is conspicuous. This is the fundamental reason by all non-forward algorithms perform poorly. By comparison, forward algorithm always evaluate the force and the force-gradient close to the exact trajectory. Since RKN is similar in form to , it strays from the exact trajectory only near the pericenter point 3.
III Integrators with four-force evaluations
Because the error of the FR integrator is uncomfortably large, there is an ongoing effort to construct better non-forward algorithms by use of more force evaluations. A non-forward fourth-order algorithm can be obtain by generalizing (16) to
[TABLE]
provided that the coefficients are left-right symmetric satisfyingcre89 ; yos90 ; suzu90
[TABLE]
Unfortunately, for , there are no real solutions to the above equations. We will examine algorithms of the general form
[TABLE]
previously studied by McLachlan’smcl95 . Since the algorithm is left-right symmetric, only operators from the center to the right are indicated. For a fourth-order integrator, the order condition requires that
[TABLE]
[TABLE]
and that the free parameter . There are four solution branches for . The choice of
[TABLE]
with
[TABLE]
reproduces McLachlan’smcl95 recommended algorithm. By simply reducing the size of the negative time steps , we obtain results as shown in Fig.5 and Fig.6. The choice of yielded simple analytical coefficients
[TABLE]
Also shown are results of forward integrator Cchin97 :
[TABLE]
where is as defined in (19) with the same interpretation (20). Algorithm C uses three force and one force gradient evaluations. The force gradient can again be extrapolated by another force evaluation. The results are similar and will be omitted in this comparison. To understand the poor performance of non-forward algorithms, we again plot their force evaluation points in Fig.7. Starting at position 2, McLachlan’s algorithm evaluates the force at F1, back tracks and evaluates the force the second time at F2, takes a giant leap to F3, then back track again to F4. By reducing to , the back tracking steps F2 and F3 are reduced to f2 and f3. However, it is not possible to move any force evaluation points closer to the midpoint of the trajectory. The force evaluation points of algorithm C are indicated by circles. Its first and last force evaluation points nearly coincide with F1 and F4, however, it evaluates the force and the force gradient right at the midpoint of the trajectory as shown by the unobstructed circle.
Better algorithms are obtained by interchanging in (27)
[TABLE]
so that the momentum is updated first with the choice
[TABLE]
Now the force is evaluated initially, at two intermediate points, and at the midpoint. The results, as shown in Fig.8 and Fig.9, are much improved over the previous case. However, the locations of the forces evaluation points remain unusual. As shown in Fig.10, for , the algorithm first evaluates the force at the starting point 2, backtracks past 2 to evaluate the force at F2, leaps forward to evaluate the force near the midpoint at F3, and shoots past the final point 3 to evaluate the force at F4. Tuning the parameter more negative to reduces the back tracking points from F2 to f2 and F4 to f4 and improves the algorithm. However, as becomes even more negative, such as or , those back tracking points bunch up very close to the initial and final points and do not sample the trajectory evenly as algorithm C.
IV Integrators with five and six force evaluations
For , (25) can be solved to give
[TABLE]
with free parameter and coefficients
[TABLE]
[TABLE]
The FR integrator is reproduced with . By introducing a non-vanishing , one is able to reduce the negative step size . This is shown in Fig.11. Fig.12 and Fig.13 show the energy and the precession error as a function of . While the energy error height is lowest for , the procession error is the smallest at . The latter is related to the fact that the backward step size is minimized at . The resulting algorithm with
[TABLE]
has long been advocated by Creutz and Gockschcre89 , Suzukisuzu90 and McLachlanmcl022 .
Recently, a fundamental theoremchin063 has allowed fourth order forward algorithms to be derived for any number of operatorschin062 . In particular, one can generalize algorithm A to force plus one force-gradient evaluations (or force evaluations using extrapolation). This is the class of algorithm with uniform splitting coefficients
[TABLE]
and where the algorithm begins and ends with a momentum updating step:
[TABLE]
We will denote this class of algorithm as AN. The energy and precession errors for A5 are as indicated in Fig.12 and Fig.13. Algorithm A5’s precession error is more than 4 times smaller than that of algorithm C and 200 times smaller than non-forward algorithm (35) at . As shown in Fig.14, algorithms (35) again characteristically evaluates the force off the trajectory. As the negative time step is reduced by increasing to , the off-trajectory force-evaluation triangle is reduced from F2-F3-F4 to f2-f3-f4.
In Fig.15 and Fig.16, we compare the energy and precession error of A6 with that of Blanes and Moanbm02 (BM), a widely cited fourth-order integrator with six force evaluations. BM’s energy error is comparable to that of algorithm C, but its maximum error height is four times that of A6. For the precession error, BM’s error is two orders of magnitude larger than A6 and twenty times larger than C. Included in the comparison is algorithm A6′, in which the force gradient is computed via extrapolation. Its energy error is nearly indistinguishable from that of A6, however, its precession error is much larger. Blanes and Moan’s integrator is superior among non-forward algorithms because it has only two very small backward time steps, as shown in Fig.17. (The momentum updating step first version of the BM integrator is not considered because it has much large errors than the position-first version discussed above.)
V Conclusions
All approximation methods for solving any evolution equation should emulate its exact solution as much as possible. The efficiency of an algorithm cannot be decided on the basis of factorization schemes, in which only the error coefficient of the error commutators are known. In the past, symplectic integrators have been prized for their excellent conservation properties. However, because of the perceived difficulty in circumventing the Shang-Suzuki theorem, forward integrators were not developed until this decade. In this work, we showed that forward integrators are more attuned to the exact solution by evaluating the force closely on the trajectory. By comparison, non-forward integrators, because of their backward time steps, are constrained to evaluate the force off the trajectory, resulting in the loss of efficiency. In all cases studied, non-forward integrators are improved by simply reducing the size of their negative time steps.
Acknowledgements.
This work is supported, in part, by a National Science Foundation grant, No. DMS-0310580.
REFERENCES
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1(1) R. Ruth, “A canonical integration technique”, IEEE Trans. Nucl. Science, 30 , 2669 (1983).
- 2(2) F. Neri, “Lie Algebra and Canonical Integration”, Department of Physics, Univeristy of Maryland print,1987.
- 3(3) M. Creutz and A. Gocksch, “Higher-order hydrid Monte-Carlo algorithms”, Phys. Rev. Letts. 63 , 9 (1989).
- 4(4) E. Forest and R. D. Ruth, “4th-order symplectic integration”, Physica D 43 , 105 (1990).
- 5(5) H. Yoshida, “Construction of higher order symplectic integrators”, Phys. Lett. A 150 , 262-268.
- 6(6) J. Wisdom, M. Holman, “Symplectic maps for the N-body problem”, Astron. J. 102 , 1528 (1991)
- 7(7) H. Yoshida, “Recent progress in the theory and application of symplectic integrators”, Celest. Mech. Dyn. Astron. 56 , 27 (1993).
- 8(8) E. Hairer, C. Lubich, and G. Wanner, Geometric Numerical Integration , Springer-Verlag, Berlin-New York, 2002.
