Anda di halaman 1dari 19

AIAA 2012-4880

AIAA/AAS Astrodynamics Specialist Conference


13 - 16 August 2012, Minneapolis, Minnesota

Implicit Runge-Kutta Methods for Orbit Propagation


Jerey M. Aristo and Aubrey B. Poore

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

Numerica Corporation, 4850 Hahns Peak Drive, Suite 200, Loveland, Colorado, 80538, USA
Accurate and efficient orbital propagators are critical for space situational awareness
because they drive uncertainty propagation which is necessary for tracking, conjunction
analysis, and maneuver detection. We have developed an adaptive, implicit Runge-Kuttabased method for orbit propagation that is superior to existing explicit methods, even
before the algorithm is potentially parallelized. Specifically, we demonstrate a significant
reduction in the computational cost of propagating objects in low-Earth orbit, geosynchronous orbit, and highly elliptic orbit. The new propagator is applicable to all regimes
of space, and additional features include its ability to estimate and control the truncation
error, exploit analytic and semi-analytic methods, and provide accurate ephemeris data
via built-in interpolation. Finally, we point out the relationship between collocation-based
implicit Runge-Kutta and Modified Chebyshev-Picard Iteration.

I.

Introduction

Accurate and timely surveillance of objects in the near-Earth space environment is becoming increasingly
important to national security and presents a unique and formidable challenge. Efficiently and accurately
modeling trajectories of the vast number of objects in orbit around the Earth is difficult because the equations
of motion are non-linear, the infrequency of observations may require modeling trajectories over long periods
of time, and the number of objects to be modeled is on the order of 105 and growing owing to object breakups
and improved sensors. While existing numerical methods for orbit propagation appear to meet the current
needs of satellite operators, advances in numerical analysis, together with improved speed, memory, and
architecture of modern computers (e.g., the availability of parallel processors), necessitate new and improved
algorithms for orbit propagation, especially when future needs in space situational awareness are taken into
consideration.
In this paper, we demonstrate the construction, implementation, and performance of collocation-based
implicit Runge-Kutta (IRK) methods for orbit propagation (i.e., the numerical solution of initial value problems arising in orbital mechanics). This includes collocation-based IRK methods such as Gauss-Legendre,
Gauss-Chebyshev, and their Radau and Lobatto variants,14 as well as band-limited collocation-based IRK
methods.5, 6 Particular attention is given to a class of collocation-based IRK methods known as GaussLegendre IRK (GL-IRK). Collocation-based IRK methods are ideally suited for orbit propagation for several
reasons. First, unlike multistep methods, GL-IRK and other collocation-based IRK methods are A-stable
at all orders, thus allowing for larger (and fewer) time steps to be taken without sacrificing stability.14
As a result, less error is accumulated and the computational cost is reduced. Second, and perhaps more
importantly, collocation-based IRK methods can exploit parallel computing architectures, which makes them
well-poised to meet the future demands of space surveillance. It is worth noting that current state-of-the-art
numerical integrators such as Dormand-Prince 8(7), Runge-Kutta-Nystrom 12(10), Gauss-Jackson (GJ), and
Adams-Bashforth-Moulton (ABM) are explicit methods that are not A-stable nor parallelizable.14, 7 The
properties of these methods are summarized in Table 1. A survey of numerical integration methods used for
orbit propagation can be found in Montenbruck & Gill8 and Jones & Anderson.9
Our implementation of IRK overcomes the disadvantages often associated with using implicit numerical
methods: the need to solve a system of nonlinear equations at time step. We mitigate this problem by making
use of available analytic and semi-analytic methods in astrodynamics.8, 10 Such approximate solutions are
Research
Chief

Scientist, Numerica Corporation.


Scientist, Numerica Corporation.

1 of 19

Copyright 2012 by the American Institute of Aeronautics and Astronautics, Inc. The U.S. Government has a royalty-free license to exercise all rights under the copyright claimed herein for Go

American Institute of Aeronautics and Astronautics

Type of method
Error control
Non-conservative forces
A-stable
Parallelizable

DP8(7)

RKN12(10)

GJ & ABM

GL-IRK

Explicit RK
Yes
Yes
No
No

Explicit RK
Yes
No
No
No

Multistep
No/Yes
Yes
No
No

Implicit RK
Yes
Yes
Yes
Yes

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

Table 1. Summary of the properties of numerical integration methods used for orbit propagation. While it is
possible to implement error control with Gauss-Jackson (GJ), its standard implementation lacks this feature.7
RKN12(10) is unable to handle cases wherein the force acting on the object depends on its velocity (e.g., when
non-conservative forces are present).

used to warm-start the iterations arising on each time step of IRK, thereby reducing the computational
cost. Second, we have developed an efficient error-control mechanism that allows for the propagation to be
entirely adaptivea . Time-adaptivity is achieved through the use of an error estimator which makes use of
the propagated solution in order to select the optimal time step for a given tolerance, thereby controlling
the accumulation of local error, and minimizing the cost of propagation per time step. This feature, among
others, distinguishes our work from that of Bai & Junkins,11 Bradley et al.,5 and Beylkin & Sandberg,6 who
studied the use of fixed-step collocation-based methods for orbit propagation.
By virtue of being adaptive, our implementation of IRK is not restricted to the propagation of nearlycircular orbits. Highly elliptic orbits can also be efficiently propagated. Fixed-step propagators, on the
other hand, are forced to take extremely small time steps when propagating highly elliptic orbits in order to
maintain high accuracy. Alternatively, the equations of motion can be transformed (e.g., using the Sundman
transform) so that fixed steps can be taken in an orbital anomaly, rather than in time, in an eort to
distribute the numerical error evenly across the steps. Unfortunately, this approach results in the need to
solve an additional ordinary dierential equation,12 and hence both options incur increased computational
costs.
This paper focuses on the propagation of objects in low-Earth orbit (LEO), geosynchronous orbit (GEO),
and highly elliptic orbit (HEO). The new propagator is found to be superior to existing propagators when
medium to high accuracy is required. Specifically, we propagate objects in LEO, GEO, and HEO for multiple
orbital periods. The performance is compared to that of DP8(7) and ABM, the latter being a method that
is similar to GJ, albeit ABM has error control, whereas the standard implementation of GJ does not have
error control.7 In a serial computing environment, we demonstrate that the new propagator outperforms
DP8(7) and ABM by 60% to 85% in LEO and in GEO, and by 30% to 55% in HEO. In a parallel computing
environment, the new propagator outperforms DP8(7) and ABM by 92% to 99% in LEO and in GEO, and
by 86% to 97% in HEO. Hence, the real impact of the new propagator is its ability to significantly reduce the
computational cost of orbit propagation, even before the algorithm is potentially parallelized.
The layout of this paper is the following. In Section II, we introduce implicit Runge-Kutta methods and
describe the construction of collocation-based IRK methods. In Section III, we describe the implementation
of adaptive IRK methods for efficient orbit propagation. In Section IV, we compare the performance of our
adaptive collocation-based IRK propagator to that of DP8(7) and ABM. In Section V, we summarize our
results.

II.

Mathematical Preliminaries

The first-order ordinary dierential equation (ODE)


y 0 (t) = f (t, y(t)) ,

(1)

y(t0 ) = y 0 ,

(2)

together with the initial condition


a An

adaptive numerical method for solving initial value problems estimates the local (truncation) error arising after each
time step and compares it to a user-specified tolerance to determine whether or not to accept the step, as well as to adjust the
size of the subsequent time step.

2 of 19
American Institute of Aeronautics and Astronautics

define an initial value problem (IVP). Many standard problems in dynamics can be converted to this form.2
The independent variable t (time) is permitted to take on any real value, and the dependent variable y 2 Rd
(the state of the object) is a vector-valued function y : R ! Rd , where d is the dimension of the state,
and f : R Rd ! Rd . Henceforth, we assume existence and uniqueness of the solution on some interval
[t0 , t0 + h].
Runge-Kutta methods may be used to solve the IVP given by (1)-(2), that is, to find the state of the
object at time t = t0 + h.14 An s-stage Runge-Kutta method is defined by its weights b = (b1 , b2 , . . . bs ),
nodes c = (c1 , c2 , . . . , cs ), and the s by s integration matrix A whose elements are aij . It is used to solve
the IVP over a given time stepb t0 to t0 + h by writing
s

X
i = f t0 + ci h, y 0 + h
aij j

i = 1, 2, . . . , s,

(3)

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

j=1

where i are the functionP


evaluations at the intermediate times t0 + ci h for i = 1, 2, . . . , s, and where the
s
internal stage z i = y 0 + h j=1 aij j is the approximation of y at time t0 + ci h. Once obtained, the internal
stages are used to advance the solution from t0 to t0 + h according to
y1 = y0 + h

s
X

bi i .

(4)

i=1

If aij = 0 for i < j, then (3) is an explicit expression for the internal stages. Otherwise, as with implicit
Runge-Kutta (IRK) methods, iterative techniques must be used to solve the nonlinear system of equations
(3), and an initial guess is required. In fact, as we shall demonstrate in Section III, it is the availability
of approximate analytical solutions to (1)-(2) that make the iterative solution of (4) tractable in orbital
mechanics. Examples of fixed-point methods for solving (3) include the Jacobi method, the Gauss-Seidel
method, and the Steensen method. Newton and quasi-Newton methods can also be used.13 Note that
the Runge-Kutta method (A, b, c) can be reformulated so that the s evaluations of f in (4) need not be
performed.4
A.

Runge-Kutta Methods for Second-Order Systems

Of particular importance in the modeling of the motion of objects subject to forces are equations of the form
Y 00 (t) = g t, Y (t), Y 0 (t) ,

(5)

which can be written as a first-order ODE (1) where


Y (t)
Y 0 (t)

y(t) =

(6)

and
f (t, y(t)) =

Y 0 (t)
g t, Y (t), Y 0 (t)

Let the initial condition (2) be


Y0
Y 00

y0 =

(7)

(8)

In this case, the system of equations defining the internal stages of the IRK method (3) becomes
Ps
i = Y 00 + h j=1 aij 0j

i = 1, 2, . . . , s,
Ps
Ps
0i = g t0 + ci h, Y 0 + h j=1 aij j , Y 00 + h j=1 aij 0j
and (4) becomes

Y1 =Y0+h

s
X

bi i ,

0
0

Y =Y +h

i=1

b Typically,

0
1

s
X

bi 0i .

i=1

multiple time-steps must be taken to solve an IVP using Runge-Kutta methods.

3 of 19
American Institute of Aeronautics and Astronautics

(9)

(10)

Substituting the first formula of (9) into the second yields


s
s

X
X
0i = g t0 + ci h, Y 0 + ci hY 00 + h2
a
ij 0j , Y 00 + h
aij 0j ,
j=1

(11)

j=1

and substituting the first formula of (9) into (10) yields


0
0

Y 1 = Y 0 + hY + h

s
X

bi 0 ,
i

0
1

0
0

Y =Y +h

i=1

where
Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

a
ij =

s
X

bi =

aik akj ,

s
X

(12)

i=1

s
X

bj aji ,

and

ci =

j=1

k=1

bi 0i .

s
X

aij .

(13)

j=1

Provided that a
ij , bi , and ci are pre-computed, these substitutions reduce the storage required by the implicit
Runge-Kutta methods as well as the number of iterations.3 Similar substitutions can be done for higher-order
ODEs to reformulate the associated IRK method.
B.

Construction of Collocation-Based Runge-Kutta Methods

There is an extensive literature devoted to the construction of IRK methods.14 Here, we describe the
construction of a class of IRK methods known as collocation-based IRK methods, which have particularly
strong stability properties. This framework allows us to readily construct collocation-based IRK methods
such as Gauss-Legendre, Gauss-Chebyshev, and their Radau and Lobatto variants.
Without loss of generalityc , a collocation-based, s-stage IRK method consists of choosing s collocation
points c1 , c2 , . . . , cs in [ 1, 1], and seeking a vector function u that obeys the initial condition (2) and satisfies
the dierential equation (1) at the collocation points. In other words, one seeks a function u such that
u(t0 ) = y 0
u0 (t0 + ci h) = f (t0 + ci h, u(t0 + ci h))

i = 1, 2, . . . , s.

(14)

A collocation method consists of finding such a u and setting y 1 = u(t0 + h). The collocation method for
c1 , c2 , . . . , cs is equivalent to the s-stage IRK method with coefficients
Z ci
aij =
Lj ( )d, i, j = 1, 2, . . . , s,
(15)
1

and
bi =

1
1

Li ( )d,

where
Li ( ) =

i = 1, 2, . . . , s,

s
Y

`=1, `6=i

ci

c`
c`

(16)

(17)

is the Lagrange interpolating polynomial.


Band-limited collocation-based IRK (BLC-IRK) methods,5 and methods wherein the interpolating function need not be polynomial, may be constructed by choosing a family of orthogonal basis functions and
computing
Z ci
s
X
1
aij =
(18)
` ( )d, i, j = 1, 2, . . . , s,
`j
1

`=1

and

bj =

s
X
`=1

c The

1
`j

1
` ( )d,

j = 1, 2, . . . , s,

interval for construction of the IRK method need not be [ 1, 1].

4 of 19
American Institute of Aeronautics and Astronautics

(19)

where

=[

ij

j (ci )]

i, j = 1, 2, . . . , s

(20)

1
are the components of the interpolation matrix
.
Collocation-based IRK methods are amenable to interpolation (and extrapolation). The solution of the
1
IVP at the collocation points, together with the interpolation matrix
, the elements of which are given
by (20), can also be used to determine the state of the object at points that do not necessarily correspond
1
to the collocation points.5 Since
can be pre-computed, interpolation can be done efficiently.

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

C.

Other Collocation Methods (e.g., Modified Chebyshev-Picard Iteration)

An alternative, but mathematically equivalent (in the sense that it produces the same discrete approximation)
approach to collocation-based IRK is the following.14, 15 Rather than working in terms of the values of the
function f , one works in terms of the coefficients of a chosen family of basis functions Rj for j = 1, 2, . . . , s
(e.g., Chebyshev or Legendre polynomials) by expressing the approximate solution u(t) as
u(t) =

s
X

aj Rj (t).

(21)

j=1

The vector coefficients aj are obtained by satisfying the ODE at s chosen collocation points and by satisfying
the initial condition; the resulting system of nonlinear equations relating the coefficients aj and the function f
is solved using iterative techniques. This approach is sometimes referred to as a collocation method, and has
been used in astrodynamics by several authors.11, 1618 We emphasize that it is mathematically equivalent
to an s-stage collocation-based IRK method having the same collocation points and basis functions,14, 15
since this fact does not appear to be well known. Since collocation methods are a subset of IRK methods,
we prefer to describe the new propagator in the broader context of IRK. This choice also allows us to easily
make use of a number of results from numerical analysis, most notably, the theory of order conditions.
D.

Order Conditions

be the exact solution to the initial value problem (1) and (2) at t = t0 +h, and let y 1 be an approximate
Let y
numerical solution at t = t0 + h. The approximate numerical solution is said to have order p if the local
error satisfies
|| = O(hp+1 )
||y 1 y
(22)
as h ! 0. For an s-stage Runge-Kutta method to obtain order p, it must satisfy the order conditions B(p),
C(), and D() with p + + 1 and p 2 + 2, where
B(p) :
C() :
D() :

s
X
i=1
s
X
j=1
s
X

bi cqi

aij cqj
bi cqi

=
1

1
q

cqi
q

aij =

i=1

Note that for p = 1, B(p) reduces to

bj
(1
q

q = 1, . . . , p;

(23)

q = 1, . . . , ;

(24)

j = 1, . . . , s q = 1, . . . , .

(25)

i = 1, . . . , s,
cqj )

s
X

bi = 1,

(26)

i=1

which is the condition required for any Runge-Kutta method to be consistent and convergent,1, 2 a prerequisite for its use. Gauss-Legendre IRK methods achieve the highest possible order (p = 2s) and are
deemed super-convergent. Their Radau IA and IIA variants achieve order p = 2s 1, and their Lobatto
IIIA, IIIB, and IIIC variants achieve order p = 2s 2.2, 3 Gauss-Chebyshev IRK methods are order p = s.
Their Radau IA and IIA variants achieve order p = s 1, and their Lobatto IIIA, IIIB, and IIIC variants
achieve order p = s 2.2, 3 The order of a BLC-IRK method is between zero and 2s depending upon the
choice of band-limit.
5 of 19
American Institute of Aeronautics and Astronautics

III.

Implementation of Implicit Runge-Kutta for Orbit Propagation

One of the main difficulties with the implementation of implicit Runge-Kutta (IRK) methods is the need
to solve a nonlinear system of equations at each time step. These must be solved iteratively and in an efficient
manner (see Subsection A), and a stopping criterion for the iterations must be established (see Subsection
B). Moreover, for an IRK method to be practical, approaches to limit the number of iterations should be
used that exploit approximate solutions to the initial value problem (see Subsection C). Another difficulty
with the implementation of IRK arises from the requirement that the accumulation of error is monitored
and controlled so that the accuracy of the numerical solution can be established (see Subsection D). The
present implementation of IRK for orbit propagation addresses and circumvents each of these difficulties.

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

A.

Solution of the Nonlinear System via Fixed-Point Iteration

Newtons method or one of its many variants can be used for solving the nonlinear system of equations, but
this requires calculation (or approximation) of the Jacobian. In orbit propagation, evaluation of a single
component of the Jacobian is more expensive than evaluation of a single component of the force, and there
are more components to evaluate. Hence, this approach can be costly. Simplified Newton methods avoid
having to calculate the Jacobian at each of the intermediate times, but reduce the convergence from quadratic
to linear.4
Alternatively, fixed-point methods can be used for solving the nonlinear system. Although fixed-point
methods converge more slowly than Newtons method (but at the same rate as simplified Newton methods),
fixed-point methods do not require the computation of a Jacobian, and are thus well-suited for orbit propagation. The downside of using fixed-point methods is that they generally degrade the superior stability
properties of IRK methods. In particular, provided that the initial guess is sufficiently close to the solution,
there is no limit on the size of the time step that can be taken when a method is A-stable and Newtons
method is used. If a fixed-point, rather than Newtons method, is used, then there is a maximum step size
above which the iterations diverge. While this may appear to limit the applicability of fixed-point methods
to orbit propagation, we have studied the convergence of fixed-point methods in this context, and determined
a bound on the maximum step size, below which the iterations will converge. For the case of unperturbed
Keplerian dynamics, our proof of convergence indicates that steps sizes up to three-quarters of an orbital
period are acceptable. Practically speaking, the propagator we develop monitors the convergence of the
iterations during each step of the propagation, and reduces the step size if necessary.
There are a number of dierent fixed-point iteration methods that may be used to solve the nonlinear
system (11), the simplest being the Jacobi method (see Algorithm 1). Note that the s evaluations of the force
model and the s internal stage updates can each be done in parallel. An efficient and robust convergence
criteria for stopping the iterations will be presented in Subsection B.
Algorithm 1: Jacobi Method
Jacobi method for solving the nonlinear system (11) arising in orbit propagation.
inputs : Initial guess for the reformulated internal stages z i and z 0i , second-order integration matrices and nodes
(aij , a
ij , ci , ci ) for i, j = 1, 2, . . . , s, time step h, current state (y 0 , y 00 ), and iteration tolerance.
outputs: Solution for the reformulated internal stages and corresponding force-model evaluations.
begin
while stopping criterion for the iterations has not been met (see Subsection B) do
for i
1 to s do
Evaluate the force model at node i: g(t0 + ci h, y 0 + z i , y 00 + z 0i ) ! g i .

for i
1 to s do
P
P
Update the ith internal stage: ci hy 00 + h2 sj=1 a
ij g j ! z i and h sj=1 aij g j ! z 0i .

An alternative to the Jacobi fixed-point method, known as the Gauss-Seidel method (see Algorithm 2),
combines the two for-loops of the Jacobi method into a single step. In practice, the Gauss-Seidel method
converges more quickly because the newly calculated function evaluations are immediately incorporated into
the internal stage updates. However, neither the s evaluations of the force model nor the s internal stage
updates can be done in parallel, so the computational advantages are only obtained in a serial computing
environment or in a parallel computing environment in which the number of processors does not exceed
the number of orbits to be propagated multiplied by the number of internal stages. In the latter situation,
6 of 19
American Institute of Aeronautics and Astronautics

parallelization at the level of the orbits can still take place, but finer parallelization, at a the level of the
internal stages, cannot.
Algorithm 2: Gauss-Seidel Method
Gauss-Seidel method for solving the nonlinear system (11) arising in orbit propagation.
inputs : Initial guess for the reformulated internal stages z i and z 0i and corresponding force-model evaluations,
second-order integration matrices and nodes (aij , a
ij , ci , ci ) for i, j = 1, 2, . . . , s, time step h, current state
(y 0 , y 00 ), and iteration tolerance.
outputs: Solution for the reformulated internal stages and corresponding force-model evaluations.

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

begin
while stopping criterion for the iterations has not been met (see Subsection B) do
for i
1 to s do
P
P
Update the ith internal stage: ci hy 00 + h2 sj=1 a
ij g j ! z i and h sj=1 aij g j ! z 0i .
0
0
Evaluate the force model at node i: g(t0 + ci h, y 0 + z i , y 0 + z i ) ! g i .

B.

Stopping Criterion for the Iterations

Rather than performing a fixed number of iterations, we seek a stopping criterion that minimizes the number
of iterations necessary for convergence to a given accuracy. Following Hairer & Wanner,4 denote the set of
internal stages after m iterations as
!
z1 z2 . . . zs
m
V =
,
(27)
z 01 z 02 . . . z 0s
and let

Vm = Vm+1

Vm denote the iteration error. Since convergence is linear, we have


|| Vm+1 || || Vm ||,

(28)

where || || is a suitable matrix norm that is consistent with the way in which the local error is estimated
(see Subsection D). Additionally, since we are dealing with a second order system, one must decide whether
to monitor the error Z or Z0 (or some combinationd ).
Provided that the iterations are converging (see Subsection A), the convergence rate < 1. Let V be
the exact solution to (11). Applying the triangle inequality to
Vm+1

V = (Vm+1

Vm+2 ) + (Vm+2

yields
||Vm+1

V ||

1
An estimate of the convergence rate can be made using

Vm+3 ) + . . .

|| Vm ||.

= || Vm ||/|| Vm

||,

(29)

(30)

(31)

or if Vm 1 is unavailable (as is true after only one iteration), from the previous integration step can be
used, provided that the previous and current time steps are comparable. We therefore stop the iterations
when

|| Vm || tol,
(32)
1
and accept Vm+1 as an approximation to V , where tol is the local error tolerance and is a numerical
pre-factor. If 1 or the number of iterations reaches the maximum allowable number of iterations, we
halt the iterations and reduce the time step by a factor of two.
d Note

that the physical dimensions of these matrices dier.

7 of 19
American Institute of Aeronautics and Astronautics

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

C.

Solution of the Nonlinear System Using Multiple Force Models

Physical systems are typically described by a hierarchy of mathematical models of increasing complexity.
In orbital mechanics, for example, the complexity of the force model depends upon the degree and order
of the gravity model, and how drag, solar radiation, and third-body eects and other perturbations are
accounted for. It is therefore reasonable to ask whether low- and medium-fidelity force models can be used
in conjunction with high-fidelity force models to efficiently solve dierential equations of the form (11). The
answer depends upon the underlying numerical methods. Implicit numerical methods can exploit multiple
force models, whereas explicit methods cannot.
In particular, the iterative technique used for solving the nonlinear system of equations that arise on
each step of an implicit Runge-Kutta (11), whether it be fixed-point or Newton-step, requires a guess for
the internal stages to start the iterations. Rather than using a single, high-fidelity force model, we use a
tiered approach wherein multiple force models are evaluated sequentially (see Algorithm 3). This approach
significantly reduces the number of high-fidelity force-model evaluations, and is similar to the approach used
by Beylkin & Sandberg6 and Bradley et. al.5
Algorithm 3: Sequentially Evaluate Force Models
Sequentially evaluate force models in succession together with IRK to solve the nonlinear system (11) arising in
orbit propagation.
inputs : Initial guess for the reformulated internal stages z i and z 0i and corresponding force-model evaluations,
second-order integration matrices and nodes (aij , a
ij , ci , ci ) for i, j = 1, 2, . . . , s, time step h, current state
(y 0 , y 00 ), and iteration tolerance.
outputs: Solution for the reformulated internal stages and corresponding force-model evaluations.
begin
Use Algorithm 1 or 2 with g = g (low) where g (low) is a low-fidelity force model, and pass the result as input to
the following step.
Use Algorithm 1 or 2 with g = g (med) where g (med) is a medium-fidelity force model, and pass the result as
input to the following step.
Use Algorithm 1 or 2 with g = g (high) where g (high) is a high-fidelity force model.

D.

Error Estimation and Control

Numerical error is intrinsic to orbit propagation (unless the propagation is done analytically). It arises
due to the finite precision of computations involving floating-point values (i.e., round-o error), and due to
the dierence between the exact mathematical solution and the approximation to the exact mathematical
solution (i.e., truncation error). Owing to round-o error, the final few digits of a floating-point computation
cannot be considered to be accurate. Round-o error can be mitigated to some extent,4 but it is truncation
error that is typically the dominant contributor to numerical error in orbit propagation. Recall that the
local (truncation) error is related to the order of the numerical method (see Section II). Fortunately, the
local error can be estimated and controlled. Note that the orbital propagators presented by Bradley et al.5
and Bai & Junkins11 do not estimate nor control the error.
Modern methods for solving initial value problems (IVPs) estimate and control the error by adjusting
the step size accordingly.19 For higher order methods such as those needed for orbit propagation,8, 9 rigorous
error bounds become unpractical, and it is therefore necessary to consider the first non-zero term in the
Taylor expansion of the error. Even so, one cannot simply compute the global error (without access to
the true solution), that is, the error of the computed solution after several steps. Rather, the local error
committed at each step of the method can be estimated and used to adjust the step size, keeping in mind
that local errors are accumulated and transported to the final state. Ideally, local error control results in
global error control, and the method takes the largest time steps possible without committing too much
error. Hence, the method does just the right amount of work solving the IVP.
Let tol = atol + rtol ||y 0 || be the local error tolerance, where atol is a user-specified absolute tolerance,
and rtol is a user-specified relative tolerance. On each step of the propagation, we estimate the local error
and require that it fall below tol. This approach is known as error per step (EPS) controle . The local error
e Conversely,

error per unit step (EPUS) control requires that the local error fall below tol h for the step to be accepted.

8 of 19
American Institute of Aeronautics and Astronautics

estimate err is used to adjust the subsequent step according to


hnext = h max(0.5, min(2, (0.35 tol/err)1/(p+1) )),

(33)

where p is the order of the IRK method (see Section II). The parameters 0.5 and 2 are chosen so that the
ratio hnext /h is neither too large nor too small. The parameter 0.35 is chosen to increase the probability
that the next step is accepted, with the probability depending on the order of the IRK method. We note
that at the culmination of each step, we ensure that the step size to be taken for the following step does
not exceed the final time. If so, the step size is reduced. Our approach to error control is summarized in
Algorithm 4.
Algorithm 4: Error Estimation and Control
Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

Estimate and control the local error.


inputs : Initial guess for the reformulated internal stages z i and z 0i and corresponding force-model evaluations,
second-order integration matrices and nodes (aij , a
ij , ci , ci ) for i, j = 1, 2, . . . , s, time step h, current state
(y 0 , y 00 ), iteration tolerance, and local error tolerance.
outputs: Solution for the reformulated internal stages and corresponding force-model evaluations accurate to
within the local error tolerance.
begin
while step has not been accepted do
Use Algorithm 3 to compute the s-stage IRK step and the error estimate err.
if err tol then
Accept the step. Use (33) to increase or decrease the time step.
else
Reject the step. Use (33) to decrease the time step.

We choose to estimate and control the error in propagating an orbit by using Earth-centered-inertial
coordinates (ECI). The reason for this choice is two-fold. First, the components of the internal stages
naturally separate into positions and velocities in ECI. Then, one simply decides whether to control the
error in position, the error in velocity, or both. Second is the fact that the geopotential calculation and its
derivatives (the component of the force model evaluation that is most expensive) are naturally expressed
in Earth-centered-Earth-fixed coordinates, which convert to ECI through a trivial rotation.8 Use of an
alternative coordinate system for propagation would incur an additional cost due to the necessary change
of variable at each node. Furthermore, one would not be able to exploit the second-order formulation of
implicit Runge-Kutta to reduce the computational cost of integration.

IV.

Performance of the New Orbital Propagator

In this section, we test the performance of the new IRK-based propagator in six realistic orbit propagation
scenarios. Both the efficiency and the accuracy of the propagations will be quantified. Although any IRK
scheme can be used within the new propagator (including those not based on collocation) we have chosen
to use the Gauss-Legendre IRK (GL-IRK) family of methods in the present study because:
GL-IRK methods are A-stable (and B-stable) and can therefore be applied to sti problems (e.g.,
satellite re-entry, highly elliptic orbits),
GL-IRK methods are parallelizable and can therefore exploit advanced computing architectures (each
evaluation of the force model can be done independently and on a dierent processor),
GL-IRK methods are super-convergent, they achieve the highest order of convergence possible for a
Runge-Kutta method (methods based on Chebyshev collocation are not super-convergent),
GL-IRK methods are symmetric and thus preserve time-reversibility of a dynamical system (if applicable), and
GL-IRK methods are symplectic and thus preserve the first integral of a Hamiltonian system (if applicable).

9 of 19
American Institute of Aeronautics and Astronautics

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

BLC-IRK methods share many of the same properties as GL-IRK (and reduce to GL-IRK in a suitable
limit). Our computational experience is that BLC-IRK methods can be up to 25% more efficient if large
(fixed) time steps are taken and the optimal step-size is known a-priori so that the error need not be
estimated. However, the challenge with using BLC-IRK methods is to develop an adaptive step-size control
that is comparable in performance to GL-IRK methods. This remains an open research problem.
We compare the performance of the new propagator to that of Dormand-Prince 8(7) (DP8(7)) and
Adams-Bashforth-Moulton (ABM). DP8(7) is an explicit Runge-Kutta method, and ABM is a multistep
method. Both methods are adaptive-step, high-order methods, and, like the new propagator, have the
ability to estimate and control the local error. Both methods were designed with the goal of reducing the
number of function calls, and, like the new propagator, both methods adaptively select the initial time stepf .
Hence, DP8(7) and ABM are good candidates to which we can compare the performance of IRK. Moreover,
ABM is similar to Gauss-Jackson, what is used by Space Command, and we have a robust implementation
of ABM (courtesy of MATLAB). An implementation of DP8(7) is available on the MATLAB file exchange.
Review of the code reveals it to be a robust implementation as well, though not to the high standard achieved
by ABM.
In order to make a fair comparison between the performance of the three methods, DP8(7) and ABM
are modified so that the same criterion for step acceptance is used as with the new propagator: on each
step of a given method, we require that the relative error in the 2-norm of the position of the object fall
below a user-specified relative toleranceg . Hence, we are using error per step (EPS) control. To quantify
the computational cost of orbit propagation, we count the number of evaluations of the high-fidelity force
model required to achieve a given global accuracy. We compute truth by propagating the initial orbital state
using a high-accuracy 50-stage GL-IRK method. Recall that evaluation of the high-fidelity force-model is
the dominant cost of orbit propagation, and that explicit methods can only use a single (high-fidelity) force
model.
The initial conditions of the six orbit propagation scenarios are listed in Table 2, along with the duration
of the propagations. A degree and order 32 gravity model (EGM2008) is used in LEO and in HEO, and
a degree and order 8 gravity model is used in GEO. For the sake of presentation, we do not consider drag
or solar radiation in the first five scenarios. In the sixth scenario (satellite re-entry), a Harris-Priester drag
model is used to demonstrate that the new propagator can handle cases wherein there is significant orbital
decay. Unperturbed Keplerian dynamics is used for the low-fidelity force model, and J2 gravity is used for
the medium-fidelity force model. Since the computational cost of evaluating the gravitational force using a
degree and order 32 gravity model is roughly 500 times greater than that using J2 -gravityh , we can safely
neglect the cost incurred during evaluation of the low- and medium-fidelity force models in LEO and in
HEO. J2 -gravity is roughly 32 times faster to evaluate than a degree and order 8 gravity model. Hence, we
can safely neglect the cost incurred during evaluation of the low- and medium-fidelity force models in GEO.
Scenario #
1
2
3
4
5
6

Orbit Type
LEO
LEO
GEO
GEO
HEO
Re-Entry

a (km)
6640
6640
42164
42164
26628
6518

e
0.009500
0.009500
0
0
0.7416
0.0003875

i( )
72.9
72.9
0
0
63.4
53.0

( )
116
116
0
0
120
145

!( )
57.7
57.7
0
0
0
267

M ( )
105
105
250
250
144
94.0

orbital periods
3
30
3
30
3
3

Table 2. Initial orbital elements for the orbit propagation scenarios. The first five are taken from Vinti,10 the
sixth from Peat.20

For each scenario, we consider the relationship between the user-specified local accuracy, the global
accuracy, and the computational cost. The relationship between the local and global error is shown first,
followed by the relationship between the computational cost and the global accuracy. A quadratic-leastsquares-fit to each of the data sets is shown. As the local accuracy is progressively increased (i.e. the local
f The

new propagator also adaptively determines the number of IRK stages to use for a given propagation.
error refers to the relative error in the 2-norm of the position of the object.
h The cost of evaluating a degree and order N gravity model using spherical harmonics is O(N 2 ), and J -gravity is approxi2
mately twice as fast as a degree and order 2 gravity model.
g Henceforth,

10 of 19
American Institute of Aeronautics and Astronautics

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

error tolerance is decreased), the global accuracy increases (i.e. the global error decreases), as does the
computational cost (i.e. the number of function calls). Eective error control is evidenced by the quadraticleast-squares-fit through the global vs. local accuracy data being nearly linear and the correlation being
positive. Note that the global accuracy is always less than the local accuracy, in some cases by an order
of magnitude or more. For a short and a long propagation to achieve the same global accuracy, the long
propagation requires a greater local accuracy than does the short propagation. This fact is inherent to
numerical propagation. The results shown pertain to a serial computing environment. Since 5-15 IRK stages
are used in the scenarios, an additional speedup of 5-15 times would be observed in a parallel computing
environmenti . For reference, 7 digits of accuracy correspond to approximately one-meter accuracy for the
object in LEO, and 7.5 digits of accuracy correspond to approximately one-meter accuracy for the object
in GEO. For the object in HEO, 7 digits of accuracy correspond to approximately one-meter accuracy at
perigee.
Scenario 1. An object in low-altitude LEO is propagated for 3 orbital periods, or approximately
4.5 hours, using adaptive step-size control. The performance of the new IRK-based propagator is
summarized in Figure 1, and compared to that of DP8(7) and ABM. All three methods eectively
control the error, and IRK is 70-80% more efficient than DP8(7) and ABM over the range of accuracies
considered.
Scenario 2. An object in low-altitude LEO is propagated for 30 orbital periods, or approximately
45 hours, using adaptive step-size control. The performance of the new IRK-based propagator is
summarized in Figure 2, and compared to that of DP8(7) and ABM. All three methods eectively
control the error, and IRK is 60-70% more efficient than DP8(7) and ABM over the range of accuracies
considered.
Scenario 3. A object in GEO is propagated for 3 orbital periods, or approximately 3 days, using
adaptive step-size control. The performance of the new IRK-based propagator is summarized in Figure
3, and compared to that of DP8(7) and ABM. All three methods eectively control the error, though
the IRK solution is ten times more accurate than that of DP8(7) and ABM for the same local accuracy.
Moreover, IRK is 80-90% more efficient than DP8(7) and 60-65% more efficient than ABM over the
range of accuracies considered.
Scenario 4. An object in GEO is propagated for 30 orbital periods, or approximately 30 days, using
adaptive step-size control. The performance of the new IRK-based propagator is summarized in Figure
4, and compared to that of DP8(7) and ABM. All three methods eectively control the error. Again,
IRK is much more accurate (globally) for a given local accuracy, and 90-95% more efficient than DP8(7)
and 75-85% more efficient than ABM over the range of accuracies considered.
Scenario 5. An object in HEO is propagated for 3 orbital periods, or approximately 36 hours, using
adaptive step-size control. The performance of the new IRK-based propagator is summarized in Figure
5, and compared to that of DP8(7) and ABM. All three methods adequately control the error, and
IRK is 50-70% more efficient than DP8(7) and 30-55% more efficient than ABM over the range of
accuracies considered.
Scenario 6. A satellite that re-entered the atmosphere, known as ROSAT,20 is propagated for 3
orbital periods, or approximately 4.5 hours, using adaptive step-size control. The performance of the
new IRK-based propagator is summarized in Figure 6, and compared to that of DP8(7) and ABM.
Using the published weight of ROSAT (2400 kg) we estimate the the area-to-mass ratio to be 4.710 4
m2 kg 1 . This results in a nearly 1 km decay in the altitude of the spacecraft per orbital period. All
three methods adequately control the error, and IRK is 65-75% more efficient than DP8(7) and ABM
over the range of accuracies considered.

V.

Conclusion

The use of implicit numerical methods for orbit propagation represents a paradigm shift in astrodynamics.
Standard algorithms are based on explicit and multistep numerical methods.7, 8 As such, they solve an initial
i This

estimate assumes that the communication time between the processors is negligible.

11 of 19
American Institute of Aeronautics and Astronautics

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

value problem by calculating the state of a system at a later time from the state of the system at the current
time. Implicit numerical methods, on the other hand, solve an initial value problem by calculating the state
of a system at a later time from the state of the system at the current time, together with the state of the
system at future times. Hence, an initial approximation for the solution is required, and the resulting system
of nonlinear equations must be solved iteratively.
What makes implicit Runge-Kutta (IRK) methods practical for orbit propagation is the availability of
analytic and semi-analytic approximations to the solution that can be computed efficiently and used to
warm-start the iterations, thereby lowering the computational cost of orbit propagation. IRK methods
can also be super-convergent, meaning that larger (and fewer) time steps can be taken than their explicit
counterparts. What is more, IRK methods are parallelizable. Explicit Runge-Kutta and multistep methods
are not. Even before parallelization, the new adaptive-step IRK-based orbital propagator is found to be
significantly more efficient in our test scenarios than adaptive-step explicit and multistep methods often
used for orbital propagation, specifically, Dormand-Prince 8(7) and Adams-Bashforth-Moulton. Table 3 lists
the computational savings obtained in LEO (Scenarios 12), GEO (Scenarios 34), HEO (Scenario 5), and
Satellite Re-Entry (Scenario 6) when medium- to high-accuracy propagations are performed.

SCE
PCE

Scenario 1
LEO

Scenario 2
LEO

Scenario 3
GEO

Scenario 4
GEO

Scenario 5
HEO

Scenario 6
Re-Entry

70-80%
94-99%

60-70%
92-98%

60-65%
92-98%

75-85%
95-99%

30-55%
86-97%

65-75%
93-98%

Table 3. Summary of computational savings in a serial computing environment (SCE) or a parallel computing
environment (PCE). The computational savings tend to increase as the accuracy of the propagation increases.

Not all orbital propagators estimate and control the truncation error by adapting the step size. Instead,
fixed steps (in time or in an orbital anomaly) are taken.5, 7, 8, 11 By taking fixed steps, one is assuming that
(1) the nonlinearity in the underlying dynamics is approximately uniform over each step and (2) the step
size yields a truncation error that is less than the error in the force models (but not too much less or else
unnecessary work is done to propagate the orbit). Unfortunately, one or both of these assumptions can be
violated. The first assumption breaks down when propagating low-altitude orbits, highly-elliptic orbits, or
orbits over long enough time intervals. The second assumption requires a-priori knowledge of the truncation
error, which would need to be tabulated oine for a given set of force models and a large number of orbit
propagation scenarios. Our approach is to use an adaptive-step orbital propagator that uses online error
estimation and control wherein the acceptable level of truncation error can be tuned to the error intrinsic to
the force models. Therefore, the error estimates are more accurate, and the propagator does just the right
amount of work for a given propagation.
A large class of methods for propagating an orbital state and its uncertainty (e.g. the unscented Kalman
filter,21, 22 particle filters,23, 24 Gaussian sum filters2528 ) require the propagation of an ensemble of particles or
states through the nonlinear dynamics. Hence, orbit propagation is a prerequisite for uncertainty propagation
using these methods. In this paper, we demonstrated the use of implicit Runge-Kutta methods for accurate
and efficient orbit propagation. In a companion paper,29 we demonstrate the use of implicit Runge-Kutta
methods for accurate and efficient uncertainty propagation.

Acknowledgments
The authors thank G. Beylkin and J. T. Horwood for helpful comments on earlier versions of this paper.
This work was funded, in part, by a Phase II STTR from the Air Force Office of Scientific Research (FA955012-C-0034) and a grant from the Air Force Office of Scientific Research (FA9550-11-1-0248).

References
1 Iserles, A., A First Course in the Numerical Analysis of Dierential Equations, Cambridge University Press, Cambridge,
UK, 2004.
2 Butcher, J. C., Numerical Methods for Ordinary Dierential Equations, John Wiley & Sons, West Sussex, England,
2008.

12 of 19
American Institute of Aeronautics and Astronautics

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

3 Hairer, E., Norsett, S. P., and Wanner, G., Solving Ordinary Dierential Equations I: Nonsti Problems, Springer
Series in Computational Mathematics, Springer, 2nd ed., 2009.
4 Hairer, E. and Wanner, G., Solving Ordinary Dierential Equations II: Sti and Dierential-Algebraic Problems,
Springer Series in Computational Mathematics, Springer, 2nd ed., 2010.
5 Bradley, B. K., Jones, B. A., Beylkin, G., and Axelrad, P., A new numerical integration technique in astrodynamics,
Proceedings of the 22nd Annual AAS/AIAA Spaceflight Mechanics Meeting, AAS 12-216, Charleston, SC, Jan. 30 - Feb. 2
2012, pp. 120.
6 Beylkin, G. and Sandberg, K., ODE Solvers Using Bandlimited Approximations, arXiv:1208.3285v1 [math.NA], 2012.
7 Berry, M. M. and Healy, L. M., Implementation of Gauss-Jackson integration for orbit propagation, Journal of the
Astronautical Sciences, Vol. 52, No. 3, 2004, pp. 331357.
8 Montenbruck, O. and Gill, E., Satellite Orbits: Models, Methods, and Applications, Springer, Berlin, 2000.
9 Jones, B. A. and Anderson, R. L., A survey of symplectic and collocation integration methods for orbit propagation,
Proceedings of the 22nd Annual AAS/AIAA Spaceflight Mechanics Meeting, AAS 12-214, Charleston, SC, Jan. 30 - Feb. 2
2012, pp. 120.
10 Vinti, J. P., Orbital and Celestial Mechanics, Progress in Astronautics and Aeronautics, edited by G. J. Der and N. L.
Bonavito, Vol. 177, American Institute of Aeronautics and Astronautics, Cambridge, MA, 1998.
11 Bai, X. and Junkins, J., Modified Chebyshev-Picard iteration methods for orbit propagation, J. Astronautical Sci.,
Vol. 3, 2011, pp. 127.
12 Berry, M. and Healy, L., The generalized Sundman transformation for propagation of high-eccentricity elliptical orbits,
Proceedings of the 12th AAS/AIAA Space Flight Mechanics Meeting, San Antonio, TX, January 2002, pp. 120, Paper AAS02-109.
13 Kelley, C. T., Iterative methods for linear and nonlinear equations, Frontiers in Applied Mathematics, Vol. 16, SIAM,
1995, pp. 1180.
14 Wright, K., Some relationships between implicit Runge-Kutta, collocation and Lanczos methods and their stability
properties, BIT , Vol. 10, 1970, pp. 217227.
15 Hulme, B. L., One-step piecewise polynomial Galerkin methods for initial value problems, Mathematics of Computation,
Vol. 26, No. 118, 1972, pp. 415426.
16 Feagin, T. and Nacozy, P., Matrix formulation of the Picard method for parallel computation, Celestial Mechanics and
Dynamical Astronomy, Vol. 29, 1983, pp. 107115.
17 Fukushima, T., Vector integration of the dynamical motions by the Picard-Chebyshev method, The Astronomical
Journal, Vol. 113, 1997, pp. 23252328.
18 Barrio, R., Palacios, M., and Elipe, A., Chebyshev collocation methods for fast orbit determination, Applied Mathematics and Computation, Vol. 99, 1999, pp. 195207.
19 Shampine, L. F., Error estimation and control for ODEs, J. Sci. Comp., Vol. 25, 2005, pp. 316.
20 Peat, C., Heavens Above, July 2012.
21 Julier, S. J. and Uhlmann, J. K., Method and Apparatus for Fusing Signals with Partially Known Independent Error
Components, U.S. Patent Number 6,829,568 B2, Issued on 7 December 2004.
22 Julier, S. J. and Uhlmann, J. K., Unscented filtering and nonlinear estimation, Proceedings of the IEEE , Vol. 92, 2004,
pp. 401422.
23 A. Doucet, N.F. Freitas, N. G. and Smith, A., Sequential Monte Carlo Methods in Practice, Statistics for Engineering
and Information Sciences, Springer, 2001.
24 Ristic, B., Arulampalam, S., and Gordon, N., Beyond the Kalman Filter: Particle Filters for Tracking Applications,
Artech House, Boston, 2004.
25 Ito, K. and Xiong, K., Gaussian filters for nonlinear filtering problems, IEEE Transactions on Automatic Control,
Vol. 45, No. 5, 2000, pp. 910927.
26 Horwood, J. T. and Poore, A. B., Adaptive Gaussian sum filters for space surveillance, IEEE Transactions on Automatic
Control, Vol. 56, No. 8, 2011, pp. 17771790.
27 Horwood, J. T., Aragon, N. D., and Poore, A. B., Gaussian sum filters for space surveillance: theory and simulations,
Journal of Guidance, Control, and Dynamics, Vol. 34, No. 6, 2011, pp. 18391851.
28 DeMars, K. J., Jah, M. K., Cheng, Y., and Bishop, R. H., Methods for splitting Gaussian distributions and applications
within the AEGIS filter, Proceedings of the 22nd AAS/AIAA Space Flight Mechanics Meeting, Charleston, SC, February
2012, Paper AAS-12-261.
29 Aristo, J. M., Horwood, J. T., and Poore, A. B., Implicit Runge-Kutta methods for uncertainty propagation, Proceedings of the 2012 AMOS Conference, September 2012.

13 of 19
American Institute of Aeronautics and Astronautics

14

Global Accuracy (in digits)

12

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

11
10
9
8
7
6
5
7

10

11

12

Local Accuracy (in digits)

2500

2000

Function Calls

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

13

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

1500

1000

500

10

11

Global Accuracy (in digits)

Figure 1. Performance of Gauss-Legendre IRK vs. DP8(7) and ABM in Scenario 1 (LEO, 3 orbital periods).
(a) Digits of accuracy in the propagated state versus the local accuracy. (b) Computational cost of orbit
propagation, measured by the number of high-fidelity force-model evaluations, versus the global accuracy in
a serial computing environment (an additional speedup of 5-15 times would be observed for IRK in a parallel
computing environment). Note that 7 digits of accuracy corresponds to approximately one-meter accuracy.
14 of 19
American Institute of Aeronautics and Astronautics

11

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

Global Accuracy (in digits)

9
8
7
6
5
4
3
7

10

11

12

Local Accuracy (in digits)

x 10

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

1.8
1.6
1.4

Function Calls

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

10

1.2
1
0.8
0.6
0.4
0.2
0

4.5

5.5

6.5

7.5

8.5

Global Accuracy (in digits)

Figure 2. Performance of Gauss-Legendre IRK vs. DP8(7) and ABM in Scenario 2 (LEO, 30 orbital periods).
(a) Digits of accuracy in the propagated state versus the local accuracy. (b) Computational cost of orbit
propagation, measured by the number of high-fidelity force-model evaluations, versus the global accuracy in
a serial computing environment (an additional speedup of 5-15 times would be observed for IRK in a parallel
computing environment). Note that 7 digits of accuracy corresponds to approximately one-meter accuracy.
15 of 19
American Institute of Aeronautics and Astronautics

12

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

Global Accuracy (in digits)

10
9
8
7
6
5
4
3
6

10

11

Local Accuracy (in digits)

2000

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

1800
1600
1400

Function Calls

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

11

1200
1000
800
600
400
200
0

6.5

7.5

8.5

9.5

10

10.5

Global Accuracy (in digits)

Figure 3. Performance of Gauss-Legendre IRK vs. DP8(7) and ABM in Scenario 3 (GEO, 3 orbital periods).
(a) Digits of accuracy in the propagated state versus the local accuracy. (b) Computational cost of orbit
propagation, measured by the number of high-fidelity force-model evaluations, versus the global accuracy in
a serial computing environment (an additional speedup of 5-15 times would be observed for IRK in a parallel
computing environment). Note that 7.5 digits of accuracy corresponds to approximately one-meter accuracy.
16 of 19
American Institute of Aeronautics and Astronautics

10

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

Global Accuracy (in digits)

8
7
6
5
4
3
2
6

10

11

Local Accuracy (in digits)

2.5

x 10

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

Function Calls

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

1.5

0.5

5.5

6.5

7.5

8.5

Global Accuracy (in digits)

Figure 4. Performance of Gauss-Legendre IRK vs. DP8(7) and ABM in Scenario 4 (GEO, 30 orbital periods).
(a) Digits of accuracy in the propagated state versus the local accuracy. (b) Computational cost of orbit
propagation, measured by the number of high-fidelity force-model evaluations, versus the global accuracy in
a serial computing environment (an additional speedup of 5-15 times would be observed for IRK in a parallel
computing environment). Note that 7.5 digits of accuracy corresponds to approximately one-meter accuracy.
17 of 19
American Institute of Aeronautics and Astronautics

11

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

Global Accuracy (in digits)

9
8
7
6
5
4
8

8.5

9.5

10

10.5

11

Local Accuracy (in digits)

3500

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

3000

Function Calls

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

10

2500

2000

1500

1000

500

5.5

6.5

7.5

8.5

9.5

Global Accuracy (in digits)

Figure 5. Performance of Gauss-Legendre IRK vs. DP8(7) and ABM in Scenario 5 (HEO, 3 orbital periods).
(a) Digits of accuracy in the propagated state versus the local accuracy. (b) Computational cost of orbit
propagation, measured by the number of high-fidelity force-model evaluations, versus the global accuracy in
a serial computing environment (an additional speedup of 5-15 times would be observed for IRK in a parallel
computing environment). Note that 7.5 digits of accuracy corresponds to approximately one-meter accuracy
at perigee.
18 of 19
American Institute of Aeronautics and Astronautics

8.5

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

Global Accuracy (in digits)

7.5
7
6.5
6
5.5
5
4.5
4
3.5
6

6.5

7.5

8.5

Local Accuracy (in digits)

1600

DormandPrince
AdamsBashforthMoulton
Implicit RungeKutta

1400
1200

Function Calls

Downloaded by INST NAC DE TECNICA AEROESPACIAL on March 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/6.2012-4880

1000
800
600
400
200
0

5.5

6.5

7.5

Global Accuracy (in digits)

Figure 6. Performance of Gauss-Legendre IRK vs. DP8(7) and ABM in Scenario 6 (Satellite Re-Entry, 3
orbital periods). (a) Digits of accuracy in the propagated state versus the local accuracy. (b) Computational
cost of orbit propagation, measured by the number of high-fidelity force-model evaluations, versus the global
accuracy in a serial computing environment (an additional speedup of 5-15 times would be observed for IRK
in a parallel computing environment). Note that 7 digits of accuracy corresponds to approximately one-meter
accuracy.
19 of 19
American Institute of Aeronautics and Astronautics