See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/267587766

The Newmark Integration Method for Simulation of Multibody Systems:

Analytical Considerations

Conference Paper · January 2005

DOI: 10.1115/IMECE2005-81770

CITATIONS READS

11 2,737

3 authors:

Bogdan Gavrea Dan Negrut

Universitatea Tehnica Cluj-Napoca University of Wisconsin–Madison
32 PUBLICATIONS   216 CITATIONS    209 PUBLICATIONS   1,673 CITATIONS   

SEE PROFILE SEE PROFILE

Florian Potra
University of Maryland, Baltimore County
188 PUBLICATIONS   5,621 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

ProjectChrono View project

projectchrono View project

All content following this page was uploaded by Dan Negrut on 29 August 2018.

The user has requested enhancement of the downloaded file.

 Proceedings of IMECE 2005
ASME International Mechanical Engineering Congress and Exposition 2005
November 5-11, 2005, Orlando, USA

IMECE2005-81770

THE NEWMARK INTEGRATION METHOD FOR SIMULATION OF MULTIBODY

SYSTEMS: ANALYTICAL CONSIDERATIONS

B. Gavrea D. Negrut F.A. Potra

Department of Mathematics, Mathematics and Computer Science Division Department of Mathematics
University of Maryland, Baltimore County Argonne National Laboratory University of Maryland, Baltimore County
Baltimore, Maryland 21250 Argonne, Illinois 60439 Baltimore, Maryland 21250
Email: [email protected] Email: [email protected] Email: [email protected]

ABSTRACT package are forthcoming. The new method has been imple-
When simulating the behavior of a mechanical system, the mented successfully for industrial strength Dynamic Analysis
time evolution of the generalized coordinates used to represent simulations in the 2005 version of the MSC.ADAMS software
the configuration of the model is computed as the solution of and used very effectively for the simulation of systems with more
a combined set of ordinary differential and algebraic equations than 15,000 differential-algebraic equations.
(DAEs). There are several ways in which the numerical solu-
tion of the resulting index 3 DAE problem can be approached.
The most well-known and time-honored algorithms are the di- 1 GENERAL CONSIDERATIONS ABOUT THE NEW-
rect discretization approach, and the state-space reduction ap- MARK METHOD
proach, respectively. In the latter, the problem is reduced to a The Newmark method [1] is by far one of the most widely
minimal set of potentially new generalized coordinates in which used integration method in the structural dynamics community
the problem assumes the form of a pure second order set of Ordi- for the numerical integration of a linear set of second Order Dif-
nary Differential Equations (ODE). This approach is very accu- ferential Equations (ODE). This problem is obtained at the end
rate, but computationally intensive, especially when dealing with of a finite element discretization. Provided the finite element ap-
large mechanical systems that contain flexible parts, stiff compo- proach is linear, the equations of motion assume the form
nents, and contact/impact. The direct discretization approach is
less but nevertheless sufficiently accurate yet significantly faster,
and it is the approach that is considered in this paper.
Mq̈ + Cq̇ + Kq = F(t) (1)
In the context of direct discretization methods, approaches
based on the Backward Differentiation Formulas (BDF) have
been the traditional choice for more than 20 years. This paper The p × p mass, damping, and stiffness matrices, M, C, and K,
proposes a new approach in which BDF methods are replaced respectively, are constant, the force F ∈ R p depends on time t,
by the Newmark formulas. Local convergence analysis is carried and q ∈ R p is the set of generalized coordinates used to represent
out for the proposed method, and step-size control, error estima- the configuration of the mechanical system. The attractive at-
tion, and nonlinear system solution related issues are discussed tributes associated with the Newmark method are: (a) the result-
in detail. A series of two simple models are used to validate the ing second order ODEs that govern the time evolution of the sys-
method. The global convergence analysis and a computational- tem do not have to be reduced to first order, which leads to sim-
efficiency comparison with the most widely used numerical in- pler implementation and a smaller dimension problem; (b) good
tegrator available in the MSC.ADAMS commercial simulation stability properties and ability to adjust the amount of damping

1 Copyright °
c 2005 by ASME
introduced into the system; (c) the method has been tested and the proposed algorithm are posed. The questions: (a) how to esti-
validated in a vast array of applications spanning many engineer- mate the local integration error; (b) how to select the integration
ing fields. The Newmark family of integration formulas depends step-size; and (c) how to define a corrector stopping criteria are
on two parameters β and γ: answered in a unitary framework. Section 5 presents the conver-
gence proof for the proposed method, and discusses the order of
the Newmark formulas when used in conjunction with the index
h2 3 DAE that arises in Multi-body Dynamics. Section 6 presents
qn+1 = qn + hq̇n + [(1 − 2β) q̈n + 2βq̈n+1 ] (2a) two simple numerical experiments carried out with the proposed
2
algorithm, aimed at validating the theoretical results derived in
conjunction with the method. The paper concludes with a series
of observations and directions of future work.
q̇n+1 = q̇n + h [(1 − γ) q̈n + γq̈n+1 ] (2b)

These formulas are used to discretize at time tn+1 the equations 2 THE INDEX 3 DAE OF MULTI-BODY DYNAMICS
of motion (1) The state of a multi-body system at the position level is rep-
resented in this paper by an array q = [q1 , . . . , qn ]T of generalized
coordinates. The velocity of the system is described by the array
Mq̈n+1 + Cq̇n+1 + Kqn+1 = Fn+1 (2c) of generalized velocities q̇ = [q̇1 , . . . , q̇n ]T . There is a multitude
of ways in which the set of generalized coordinates and velocities
Note that based on Eqs. (2a) and (2b), qn+1 and q̇n+1 are func- can be selected [5–8]. The generalized coordinates used here are
tions of the acceleration q̈n+1 , which in Eq. (2c) remains the sole Cartesian coordinates for position, and Euler angles for orienta-
unknown quantity that is eventually computed as the solution of tion of body centroidal reference frames. Thus, for each body i
a linear system. This method is implicit and A-stable (stable in its position is described by the vector ri = [xi , yi , zi ]T , while its
the whole left-hand plane) [2] provided [3] orientation is given by the array of local 3-1-3 Euler angles [9],
ei = [ψi , θi , φi ]T . Consequently, for a mechanical system contain-
£ ¤T
ing nb bodies, q = rT1 eT1 . . . rTnb eTnb ∈ R p , p = 6nb . When
¡ ¢2 compared with the alternative of using a set of relative gen-
γ + 12 eralized coordinates, the coordinates considered are convenient
γ ≥ 1/2 β≥ (3)
4 because of the rather complex formalism employed to obtain
the Jacobian information required for implicit integration. Note
The only combination of β and γ that leads to second order accu- that the set of position generalized coordinates q is augmented
racy is γ = 12 , and β = 14 . The method obtained is the trapezoidal with deformation modes when flexible bodies are present in the
method, which is therefore both A-stable and second order. The model. For notational simplicity the assumption is that there are
major drawback of the Newmark family is that it can not pro- no flexible bodies in the model, although as observed in practice
vide a formula that is A-stable, second order, and at the same this does not pose a problem with the proposed algorithm.
time displays a desirable level of numerical damping. The com- In any constrained mechanical system, joints connecting
promise for the Newmark family is to keep the A-stability and bodies restrict their relative motion and impose constraints on the
have numerical damping at the expense of the integration order, generalized coordinates. Kinematic constraints are then formu-
which thus go down to one. A method that alleviates this lim- lated as algebraic expressions involving generalized coordinates,
itation at the price of a slightly more convoluted formulation is
the so called α-method, also known as the Hilber-Hughes-Taylor £ ¤T
Φ(q,t) = Φ1 (q,t) . . . Φm (q,t) = 0 (4a)
(HHT) method [3]. A generalization of this method is provided
in [4].
The remaining of this paper is organized as follows. Section where m is the total number of constraint equations that must be
2 introduces the set of Differential Algebraic Equations (DAE) satisfied by the generalized coordinates throughout the simula-
that govern the time evolution of mechanical systems (the con- tion. It is assumed here that the m constraint equations are inde-
strained equations of motion), and discusses briefly approaches pendent. Although the implementation of the proposed method
for their numerical solution. Section 3 presents the framework handles non-holonomic constraints, to keep the presentation sim-
in which the Newmark method is applied for the solution of the pler the case of holonomic constraints is assumed in what fol-
index 3 DAE that arises in Multi-body Dynamics. In section 4, a lows.
set of three questions relevant to the robustness and efficiency of Differentiating Eq. (4a) with respect to time leads to the velocity

2 Copyright °
c 2005 by ASME
kinematic constraint equation defined function that depends on other system states as indi-
cated in Eq. (5a).
2. External force definition, F, which allow the user to more
Φq (q,t) q̇ + Φt (q,t) = 0 (4b)
conveniently define the set of n f applied forces F ∈ Rn f that
act on the system. This is the mechanism through which a
where the over-dot denotes differentiation with respect h toitime complex tire model can be hooked up with a vehicle model,
and the subscript denotes partial differentiation, Φq = ∂Φ i
∂q j , for
or the avenue through which the user can define his/her own
bushing elements, custom non-linear dampers, friction, etc.
1 ≤ i ≤ m, and 1 ≤ j ≤ n. Finally, the acceleration kinematic
constraint equations are obtained by differentiating Eq. (4b) with
respect to time,
F − f(q, q̇, q̈, X, Ẋ, λ, V, F,t) = 0n f (5b)

¡ ¢
Φq (q,t) q̈ + Φq (q,t)q̇ q q̇ + 2Φqt (q,t) q̇ + Φtt (q,t) = 0 (4c) Equations (4a)–(4d) comprise a system of index 3 DAE [10].
It is known that differential-algebraic equations are not ordinary
differential equations [11]. Analytical solutions of Eqs. (4a) and
Equations (4a)–(4c) characterize the admissible motion of the (4d) automatically satisfy Eqs. (4b) and (4c), but this is no longer
mechanical system. true for numerical solutions. In general, the task of obtaining a
The state of the mechanical system changes in time under numerical solution of the DAE of Eqs. (4a)–(4d) is substantially
the effect of applied forces. The time evolution of the system more difficult and prone to intense numerical computation than
is governed by the Lagrange multiplier form of the constrained that of solving ordinary differential equations (ODE). For an ac-
equations of motion [7], count of relevant work in the area of numerical integration meth-
ods for the DAE of Multi-body Dynamics the reader is referred
to [2, 10, 12–15] and references therein.
M(q)q̈ + ΦTq (q)λ = Q (q̇, q,t) (4d) The theory and attractive features associated with the New-
mark method have been derived in conjunction with a linear sec-
ond order ODE. The only similarity between Eqs. (1) and (4d) is
where M(q) ∈ R p×p is the generalized mass, and Q (q̇, q,t) ∈ that they are both second order, and qualitatively obtained from
R p is the action (as opposed to the reaction ΦTq (q)λ) force act- Newton’s second law. In [16] and more recently [17], for the pur-
ing on the generalized coordinates q ∈ R p . These equations are pose of stability and convergence analysis the constrained equa-
neither linear, nor ordinary differential as is the case in Eq. (1), tions of motion are tackled in a stabilized index 2 DAE frame-
first and foremost because the solution q(t) must also satisfy work. The HHT and Newmark methods are also discussed in [8]
the kinematic constraint equations in Eq. (4a). These constraint and more recently in [18], where the proposed implementation is
equations lead, in Eq. (4d), to the presence of the reaction force based on a technique that accounts for violations in the position
ΦTq (q)λ, where λ ∈ Rm is the Lagrange multiplier associated and velocity constraints in a stabilization framework similar to
with the kinematic constraints. the one proposed in [19].
In addition to the equations of motion and kinematic con- There are also several Runge-Kutta based approaches for highly
straint equations, there are several classes of equations that need oscillatory mechanical system simulation that, like the Newmark
to be considered in a general purpose mechanical simulation family of methods, display the attractive attribute of selectively
package: damping frequency at the high end of the spectrum. In [20], a
Singly Diagonal Runge-Kutta (SDIRK) based method allows the
1. User defined variables, which can technically be regarded user to choose, within certain bounds, the diagonal value in the
as aliases or definition equations. A set of nv user defined formula, and thus control the amount of numerical damping as-
variables V ∈ Rnv is typically specified through an equation sociated with the algorithm. The role of the diagonal element in
of the form the formula becomes very similar to the role of the α parameter
in the α-method [3]. An approach based on additive Runge-Kutta
methods that has the potential to accurately handle highly oscilla-
V − v(q, q̇, q̈, X, λ, V, F,t) = 0nv (5a)
tory multi-body dynamics simulation was introduced in [21], and
further discussed in [22]. These novel Runge-Kutta based algo-
and which during the solution sequence are solved (or rather rithms are mathematically very sound, but more time is required
evaluated) simultaneously with the equations of motion and for them to achieve, vis-a-vis industrial strength applications, the
the kinematic constraint equations. Here v ∈ Rnv is a user level of acceptance and trust currently associated with the well

3 Copyright °
c 2005 by ASME
established Newmark method. Everywhere in Eq. (6), the Newmark integration formulas of
Eq. (2) are used to express q and q̇ as a function of q̈. A Newton-
like algorithm [23] is used to solve the resulting system of non-
3 THE PROPOSED ALGORITHM linear equations for the set of unknowns (in this order) q̈, λ, V,
and F. The iterative method requires at each iteration (k) the
The index 3 DAE problem of multi-body dynamics is nei-
solutions of the linear system
ther linear, nor ordinary differential in nature and the Newmark
method is thus applied for a different class of problems than what
   (k)  (k)
it was originally designed for. Rather than approaching the solu- Ψ ΦTq 0 −Π ∆q̈ −e1
 0     
Φq   ∆λ  = −e2 
tion within an index 2 framework [16,17] or using a stabilization
 ¡ ¢ 0 0
approach [8, 18], the proposed algorithm uses the implicit New- − vq̈ + γhvq̇ + βh vq −vλ I − vV −vF  ∆V
2 −e3 
¡ ¢
mark formulas to discretize the equations of motion and requires − fq̈ + γhfq̇ + βh2 fq −fλ −fV I − fF ∆F −e4
that the position-level kinematic constraint equations be satisfied (7)
at the end of each time step. This is a direct index 3 approach where ei are the residuals in satisfying the set of discretized equa-
and it requires at each integration time step the solution of a non- tions of motion, constraint equations, variable definition equa-
linear system of equations. The theoretical foundation of this tions, and applied force definition equations, respectively.
method is provided by the convergence results presented in sec-
tion 5. ¡ ¢
By means of the Newmark formulas, the collection of dif- e1 = (Mq̈)n+1 + ΦTq λ − Q n+1
ferential and algebraic equations are discretized to obtain an al- 1
e2 = Φ (q,t)
gebraic non-linear system that is solved by means of a modified- βh2
Newton algorithm. In the most general case, the unknowns of in- e3 = V − v (q, q̇, q̈, X, λ, V,t)
terest are the generalized positions, velocities, and accelerations ¡ ¢
q, q̇, and q̈, respectively, the Lagrange multipliers λ, the applied e4 = F − f q, q̇, q̈, V, X, Ẋ,t
force states F, the user-defined variables (aliases) V. Note that
the generalized force Q is obtained by projecting (via a linear Likewise, the matrix Ψ in Eq. (7) is defined as
transformation) the force states F along the generalized coordi-
nates q; i.e., Q = Π F, where the projection operator Π = Π(q)
depends on the choice of generalized coordinates. h i
∂e1 ¡ ¢
For notational simplicity, when obvious, the dependency of some Ψ= = M + (Mq̈)q + ΦTq λ q − Πq F βh2 (8)
quantities on q and/or q̇ and/or time t will be omitted. Likewise, ∂q̈
throughout this paper, Is denotes the identity matrix of dimension
s. All the quantities in e1 through e4 above are evaluated at time
tn+1 . Finally, note that the non-linear equations associated with
the position kinematic constraints are scaled by βh12 in order to
¡ ¢ improve the conditioning of the coefficient matrix in Eq. (7).
(M(q)q̈)n+1 + ΦTq λ n+1 − (Q)n+1 = 0 (6a)
This is a compromise reached after considering the following al-
ternatives: (a) have the level-zero positions, q be the unknowns
(replacing q̈), but then some entries in the Jacobian matrix in
Eq. (7) will have to be divided by βh2 ; (b) have q̈ be the un-
knowns, but then the second row in the Jacobian matrix comes
Φ (qn+1 ,tn+1 ) = 0 (6b)
multiplied by βh2 ; (c) do as in (b), except that the set of posi-
tions kinematic constraint equations are scaled by βh12 . Alterna-
tive (a) is what the default GSTIFF integrator currently uses in
the MSC.ADAMS simulation package [24] (here entries get di-
Vn+1 − v (qn+1 , q̇n+1 , q̈n+1 , Xn+1 , λn+1 , Vn+1 , Fn+1 ,tn+1 ) = 0 vided by a factor β0 h rather then βh2 , as the second order equa-
(6c) tions of motion are reduced to an equivalent first order system of
differential equations). On numerous occasions this has been ob-
served to be the cause of numerical problems once the step-size
¡ ¢ becomes very small and consequently some entries in the Jaco-
Fn+1 −f qn+1 , q̇n+1 , q̈n+1 , Xn+1 , Ẋn+1 , λn+1 , Vn+1 , Fn+1 ,tn+1 = 0 bian become extremely large. A bad Jacobian condition num-
(6d) ber ensues, and the quality of the Newton corrections becomes

4 Copyright °
c 2005 by ASME
poor. The alternative (b) was not embraced due to the fact that as well the first time derivative of the equations of motion at time
the problem at (a) plagues in a more subtle way this approach tn , that is,
as well. If h becomes very small, the second row of the Jaco-
bian matrix is scaled by βh2 , which practically makes all the en-
tries in this row very small, from where the same large condition ...
number situation ensues. Alternative (c) proved a good solution M q n + Cq̈n + Kq̇n = Ḟn (11)
since typically the type of error that one sees in satisfying the
position kinematic constraint equations is very small. It is never ...
where q n formally represents the time derivative of the accelera-
that these constraint equations are problematic in a simulation tion at time tn . The fact that this quantity is typically not available
but rather some discontinuity in the model that causes the step- is set aside for the time being and will be revisited later.
size h to assume small values. But if h is small, when advancing
the simulation the position constraint violation stays very small,
and the value of e2 in Eq. (8) always remains very reasonable. The Newmark integration formula of Eq. (2) is rewritten in
With the corrections computed as the solution of the lin- the equivalent form
ear system of Eq. (7), the numerical solution is improved at
each iteration as q̈(k+1) = q̈(k) + ∆q̈(k) , λ(k+1) = λ(k) + ∆λ(k) ,
V(k+1) = V(k) + ∆V(k) , F(k+1) = F(k) + ∆F(k) . The following sec-
h2
tions present in detail the answer to three key questions; (a) when qn+1 = qn + hq̇n + q̈n + βh2 x (12a)
is the computed solution accurate enough, (b) how to select the 2
integration step-size h, and (c) when to stop the Newton-like iter-
ative process that computes at each integration step the unknowns
q̈, λ, F, and V. Recall that once q̈ is available, Eqs. (2a) and (2b) q̇n+1 = q̇n + hq̈n + hγx (12b)
are used to evaluate q, and q̇, respectively.

q̈n+1 = q̈n + x (12c)

4 IMPLEMENTATION DETAILS
4.1 Estimating the local integration error The quantity of interest is the unknown x, which represents the
The strategy for evaluating the local integration error is change in the value of acceleration from time tn to tn+1 . Once this
based on a linearization of the equations of motion in Eq. (4d) quantity is available Eq. (12) is used to compute the configuration
along with an asymptotic expansion of the solution q. This ap- of the system at tn+1 . Note that qn+1 , q̇n+1 and q̈n+1 are required
proach is similar to the one proposed in [25]. The discussion is to satisfy Eq. (9). The goal is then to compute an estimation of
going to focus on Eq. (1), since locally a linearization of Eq. (4d) the error at the end of one integration step (the local integration
leads to the previous form. Thus, Eq. (1) is rewritten as error)

Mq̈n+1 + Cq̇n+1 + Kqn+1 = Fn+1 (9) δn+1 = qn+1 − q̃n+1 (13)

For the purpose of computing the local integration error, the where q̃n+1 is the exact solution of the initial value problem
usual assumption is that the configuration at time tn ; i.e., (qn , q̇n , (IVP)
q̈n ), is perfectly consistent. That is, it satisfies the equations of
motion, along with the time derivatives of the equations of mo-
Mq̈ + Cq̇ + Kq = F (14)
tion. The focus is exclusively on computing the error associated
with advancing the simulation from tn to tn+1 using the Newmark
method. Since the configuration is considered to be consistent at that starts in the configuration (qn , q̇n , q̈n ) at t = tn .
time tn , it will satisfy the equations of motion, Using Taylor’s Theorem, q̃n+1 is obtained as

h2 h3 ... ¡ ¢
Mq̈n + Cq̇n + Kqn = Fn (10) q̃n+1 = qn + hq̇n + q̈n + q n + O h4 (15)
2 6

5 Copyright °
c 2005 by ASME
The local integration error δn+1 becomes available as soon the 4.2 The Accuracy Test
¢ q ∈ R , using for 1 ≤ i ≤ p the 3notation ci =
acceleration correction x is available. In order to obtain an esti- p
¡ With
1 ...
mation for x, from Eqs. (12) and (1), β − 6 q i,n , and dropping the terms of order h or higher in
Eq. (20) yields
M (q̈n + x) + C (q̇n + hq̈n + hγx) (16)
µ ¶
h2
+K qn + hq̇n + q̈n + βh2 x δi,n+1 = ci · h3 (22)
2
¡ ¢
= Fn + hḞn + O h2
Since ci is not known, the integration error at the end of the one
time step is actually approximated as
where Taylor’s theorem was used to expand F (tn + h). Using
¡ ¢ 2
Eqs. (10) and (11), and ignoring the O h2 term K h2 q̈n leads to µ ¶
1 2
δi,n+1 = β − h · xi 1≤i≤ p (23)
6
£ ¤ ... ¡ ¢
M + hγC + βh2 K x = M q n h + O h2 (17)
Based on this value of the local integration error, a composite
error first needs to be defined. The proposed form is
Denoting D = M + γhC + βh2 K, since D−1 = M−1 + O (h) · I p
the equation v "
u
u 1 µ δ1,n+1 ¶2 µ ¶2 #
δ
e≡t
p,n+1
q +...+ q (24)
... ¡ ¢ p Y1 Yp
D x = M q n h + O h2
q
where Yi = max(1, max j=1,...,n+1 |qi, j |) , and δi,n+1 , 1 ≤ i ≤ p,
leads to is the ith component of the quantity computed in Eq. (23). The
composite error is compared with the user prescribed error de-
... ¡ ¢
x = q n h + O h2 (18) noted here by ε. If e ≤ ε the integration time step is accepted,
otherwise it is rejected. Using the notation

Using Eq. (12a),

pε2
ψ≡ ¡ ¢2
h2 ... ¡ ¢ β − 61
qn+1 = qn + hq̇n + q̈n + βh3 q n + O h4 (19)
2
error test is equivalently expressed as
Based on Eqs. (19) and (15)

µ ¶ ψ
1 ... ¡ ¢ kxk2q ≤ (25)
δn+1 = qn+1 − q̃n+1 = h β −3
· q n + O h4 (20) h4
6

... where k · kq represents a weighted norm [26] defined as kxkq ≡

Substituting for q n from Eq. (18) and dropping the higher order · ³ ´2 ¸ 12
p x
terms leads to ∑i=1 Y qi . If the condition of Eq. (25) is satisfied at the
i

µ ¶ end of the corrector stage, the integration step is accepted, other-

1 2 wise it is rejected. In either case, the next step-size is selected as
δn+1 = β − h x (21)
6 indicated in subsection 4.3.

which provides an effective way of computing the local integra- 4.3 The Step-Size Selection
tion error since all the quantities that enter this equation are avail- Step-size selection plays a central role in the numerical in-
able at the end of the corrector stage. tegration algorithm. If enew ¿ ε, effort is wasted in computing a

6 Copyright °
c 2005 by ASME
solution that unnecessarily exceeds the user demanded accuracy. 4.4 The Correction Stage
In this case, a more aggressive step-size selection would result in The goal in this section is to determine a sensible stopping
larger values for h with potentially big CPU savings at the end of criteria for the corrector iterative process. The issue is how ac-
the simulation. At the other end of the spectrum, a step-size se- curate should the quantity x of Eqs. (12) be computed. This
lection mechanism that is too aggressive leads to a large number quantity is obtained as the solution of an iterative Newton-like
of integration time steps at the end of which the user accuracy algorithm that requires at least one evaluation of the residuals in
demands are not met. The effort to perform such an integration Eq. (8), followed by a forward/backward substitution to retrieve
step is wasted, as the integration step is discarded for a new at- the corrections in the unknowns. However, one corrector itera-
tempt with a more conservative step-size h. To strike the right tion might be as expensive as doing all of the above but preceded
note, every time the integration step-size is chosen the goal is for by a full blown evaluation and factorization of the coefficient
the error at the end of the integration step to be precisely equal to matrix of the linear system of Eq. (7). These operations are ex-
the one deemed acceptable by the user, and quantitatively defined pensive and should be kept to a minimum.
by ε. It is important to get an accurate solution of the non-linear
Key in the selection of a new step-size is Eq. (22), which discretization system because a sloppy solution would adversely
indicates that the composite error is proportional to the cube of impact the quality and stability of the numerical solution. In this
the step-size h. Ideally, the new step-size hnew is selected such context, suppose that x is approximated by x(k) , i.e., the value
that obtained after k corrector iterations. Equation (23) indicates that
the integration error e is computed based on the value x, and
therefore x(k) will lead to a value of the error e(k) . It is there-
" fore important to have a good approximation x(k) for x, if the
µ ¶ # 12 algorithm is to produce a meaningful approximation e(k) of the
1 p ci 2
ε = h3new ∑ Yi
p i=1 composite integration error e.
A second reason for having an accurate solution is that the stabil-
ity and convergence results associated with a numerical integra-
h3 tor are derived under the assumption that the numerical solution
Therefore, eε = h3new
, from where
is computed to the specifications of the integration formula; i.e.,
there is no room left for errors in finding the numerical solution at
the end of one integration step. Finding an approximate solution
ε3
1 translates into solving a different initial value problem, which
hnew = h ( ) 31 can be close or far from the original problem based on how ac-
· ³ ´2 ¸ 12
p δi,n+1 curate the non-linear system of Eq. (6) is solved, and the nature
1
p ∑i=1 Yi of the original initial value problem itself.
The corrector stopping criteria adopted here is that the rela-
tive difference between e and e(k) will be smaller than a threshold
Based on Eq. (21), value denoted by c. A typical value recommended in the litera-
ture is c = 0.001 [12].
The local integration error at the end of one time step was shown
£ ¤ 2
1 to be e = β − 16 √h p kxkq . After iteration k, the approxima-
s·ε3 £ ¤ 2
hnew = h h i1 tion obtained is e(k) = β − 61 √h p kx(k) kq . The question is what
¡ ¢ 3
√1 β − 1 h2 · kxkq
p 6 should k be such that e(k) is close to e within 0.1% (c = 0.001);
i.e.,
where a safety factor s = 0.9 was used to scale the value of the
new step-size [12]. ¯ e − e(k) ¯
¯ ¯≤c (27)
kxk2q ·h4 e
Defining Θ = ψ , the previous equation finally leads to
Since e is not available, the test above is replaced by

sh
hnew = (26) ¯ e − e(k) ¯
1 ¯ ¯≤c
Θ 6
ε
(28)

7 Copyright °
c 2005 by ASME
where ε is the user prescribed error. Note that the goal of the back to Eq. (29),
step-size control is to keep e as close as possible to ε, so replacing
Eq. (27) with Eq. (28) is acceptable. Therefore, ¯ ¯ ¯ ¯ h2
¯e − e(k) ¯ ≤ ¯β − 1 ¯ √ ξ
k4x(k) kq · (34)
6 p 1−ξ
¯ ¯ ¯ ¯ h2
¯e − e(k) ¯ ≤ ¯β − 1 ¯ √ kx − x(k) kq (29)
6 p The condition of Eq. (28) is then satisfied as soon as

1 ¯¯ 1 ¯ h2 ξ
and an approximation for kx − x(k) k is needed. Since for the β − ¯ √ k4x(k) kq ≤c
Newton-like method employed the convergence is linear, there is ε 6 p 1−ξ
a constant ξ that for convergence must satisfy 0 ≤ ξ < 1, such
that [23] Equivalently,

k4x(k+1) kq ≤ ξ · k4x(k) kq µ ¶2
(30) ξ ψ
k4x(k) k2q ≤ c2 · (35)
1−ξ h4
where 4x(k) represents the correction at iteration k, x(k+1) =
x(k) + 4x(k) . Then, Note that at the right of the inequality sign are quantities that re-
main constant during the corrector iterative process, while at the
h i h i
left are quantities that change at each iteration. Likewise, note
x(k) − x = x(k) − x(k+1) + x(k+1) − x(k+2) + . . .
that the stopping criteria of Eq. (35) can only be used at the end
of the second iteration since it is only then that an approximation
of the convergence rate ξ can be produced. In other words, the
proposed approach will not be able to stop the iterative process
⇒ kx(k) − xkq ≤ k4x(k) kq + k4x(k+1) kq + . . . after the first iteration. This is not a matter of great concern since
¡ ¢ models as simple as a one body pendulum are already non-linear.
≤ k4x(k) kq 1 + ξ + ξ2 + . . .

which leads to 4.5 The Prediction Stage

When the Newton-like algorithm is used to find the solu-
tion of Eq. (8), a good starting point is essential both for conver-
1 gence, and reducing the effort to find the approximation of the
kx − x(k) kq ≤ k4x(k) kq · (31) solution at time tn+1 . In [27], the generalized accelerations pre-
1−ξ
diction is obtained by taking q̈n+1 = q̈n , which is equivalent to
setting x = 0 in Eq. (12). This strategy is replaced by one based
The value ξ is going to be approximated by on polynomial extrapolation, in which currently a polynomial of
order up to three is used to produce an initial guess for all the
k4x(k) kq unknowns. The approach used in [27] is thus obtained by setting
ξ ≈ ξk = (32) the degree of the interpolation polynomial to zero. The polyno-
k4x(k−1) kq
mial extrapolation is based on Newton divided differences and
it uses Horner’s scheme for evaluation of the interpolant at time
Based on Eq. (31) tn+1 [26]. The degree of the interpolant can be adjusted on the fly,
and the mechanism to control this takes into account the smooth-
ξ ness of the solution.
kx − x(k+1) kq ≤ k4x(k) kq · (33)
1−ξ
4.6 Summary of key formulas
which indicates that the distance between the true solution x, and The answers to the questions (a) what is the stopping cri-
the approximation of the solution obtained after applying k cor- teria for the non-linear discretization algebraic system, (b) how
rections; i.e. x(k+1) , is bounded from above by a quantity propor- is the integration error computed, and (c) how is the step-size
tional to the norm of the most recent correction k4x(k) kq . Going controlled, is summarized below.

8 Copyright °
c 2005 by ASME
Notation: A(h) the matrix of (37). Using the smoothness of the problem
data and Taylor’s theorem we obtain:
pε2 kxk2q · h4
ψ≡ £ ¤2 Θ= (36a) · ¸
β − 16 ψ Ψ(q̄n+1 ) ΦTq (tn+1 , q̄n+1 )
A(h) =
Φq (tn+1 , q̄n+1 ) 0
· ¸ (38)
Prediction: Performed based on divided differences (Newton M(qn ) + O(h) ΦTq (tn , qn ) + O(h)
interpolation and Horner’s scheme for extrapolation at tn+1 ). = .
Φq (tn , qn ) + O(h) 0
Correction: Linear convergence rate allows for computation of
ξ (Eq. (32)). Stopping criteria:
We also have that
µ ¶2
ξ ψ
k4x(k) k2q ≤ c2 , (c = 0.001) (36b) e1 = M(qn )an + ΦTq (qn )λn − Q(tn , vn , qn ) + O(h),
1−ξ h4

Accuracy Check: Performed after corrector converged, and

h2
Θ≤1 (36c) Φ(tn+1 , qn + hvn + an ) = Φ(tn+1 , q̃(tn+1 )) + O(h3 ) = O(h3 ).
2

Step-Size Selection: With a safety factor s = 0.9, Therefore the right-hand side of (37), e(h) := −(eT1 eT2 )T , sat-
isfies the asymptotic expansion: e(h) = e0 + O(h), where e0 is
(q) sh obtained from the above linearizations, and does not depend on
hnew = (36d)
Θ6
1 the integration step h. For h sufficiently small we have

Dense output: Based on third order Hermite interpolation [26], x(h) = (∆aT ∆λT )T = A(h)−1 e(h),
for q̇, and q
and therefore,

5 LOCAL CONVERGENCE ANALYSIS

||x(h)|| ≤ ||A(h)−1 ||||e(h)||. (39)
In what follows we perform a local error analysis of the
scheme when applied to the index 3 DAE (4a– 4d). In the cor-
rection stage we perform one Newton iteration, with predictor Taking h → 0 above gives that
ān+1 = an (we will use a := q̈ and v := q̇), so that the acceler-
ation an+1 and the Lagrange multipliers λn+1 are computed by ||x(h)|| ≤ C
solving the linear system:

· ¸· ¸ · ¸ for all h sufficiently small, where C is a constant independent

Ψ ΦT an+1 − an −e1
= . (37) of h. From the above estimate we conclude that the corrections
Φ 0 λn+1 − λn −e2 ∆a, ∆λ, obtained after the first Newton iteration are bounded as
h → 0, which proves our claim.
The linear system matrix as well as the right-hand side are com-
In the rest of the section we analyze the local truncation error
puted using the predicted values ān+1 = an , v̄n+1 = vn + han ,
2 when one Newton iteration is performed. We first look at the
q̄n+1 = qn + hvn + h2 an and λ̄n+1 = λn . Let (q̃(t), λ(t)) ˜ denote
level of accuracy that results when the method is applied to the
the true solution of (4a– 4d). For the truncation error analysis we true solution. Using Taylor’s theorem, it is easy to see that
take qn = q̃(tn ), vn = q̃(tn ), an = ã(tn ) and λn = λ̃(tn ).
We first show that locally, as h → 0, the stopping criteria 2
given by (36b) is satisfied after one iteration. For this to hold, it is q̃(tn+1 ) = qn + hvn + h2 ((1 − 2β)an + 2βã(tn+1 )) + O(h3 )
2
sufficient to prove that the corrections (∆a, ∆λ) remain bounded = qn + hvn + h2 an + h2 β(ã(tn+1 ) − an ) + O(h3 )
after one Newton iteration when h → 0. Let us denote by A := = q̄n+1 + h2 β(ã(tn+1 ) − an ) + O(h3 ).

9 Copyright °
c 2005 by ASME
Using the fact that Φ(q̃(tn+1 ),tn+1 ) = 0, we obtain together with (43–45) gives:

0 = Φ(q̃(tn+1 ),tn+1 ) M(q̄n+1 )ã(tn+1 ) + ΦTq (q̄n+1 )λ̃(tn+1 ) − Q(tn+1 , v̄n+1 , q̄n+1 )
= Φ(q̄n+1 ,tn+1 ) + h βΦq (q̄n+1 ,tn+1 )(ã(tn+1 ) − an ) + O(h3 ),
2
+hG(an − ã(tn+1 )) = hG(an − ã(tn+1 )) + O(h2 ),

and therefore which implies that equation (41) is satisfied up to O(h2 ), i.e.,

1
Φq (q̄n+1 ,tn+1 )(ã (tn+1 ) − an ) = − Φ(q̄n+1 ,tn+1 ) + O(h) M(q̄n+1 )ã(tn+1 ) + ΦTq (q̄n+1 )λ̃(tn+1 )
βh2
(40) −Q(tn+1 , v̄n+1 , q̄n+1 ) + hG(an − ã(tn+1 )) = O(h2 ), (46)
From the first set of equations in (37) we have that the numerical
solution (qn+1 , vn+1 , an+1 , λn+1 ) satisfies
The analysis above can be summarized by the following result

Ψ(an+1 − an ) + ΦTq (q̄n+1 ,tn+1 )(λn+1 − λn ) Theorem 1. The true solution of the DAE (4a–4d) satisfies the
following estimates:
= −M(q̄n+1 )an − ΦTq (q̄n+1 ,tn+1 )λn + Q(tn+1 , v̄n+1 , q̄n+1 ).

h2
Simplifying the similar terms above leads to q̃(tn+1 ) = q̃(tn )+hṽ(tn )+ ((1 − 2β)ã(tn ) + 2βã(tn+1 ))+O(h3 )
2
(47a)
M(q̄n+1 )an+1 + ΦTq (q̄n+1 )λn+1
−Q(tn+1 , v̄n+1 , q̄n+1 ) + hG(an − an+1 ) = 0, (41) ṽ(tn+1 ) = ṽ(tn ) + h ((1 − γ)ã(tn ) + γã(tn+1 )) + O(h2 ) (47b)

where
(M(q̄n+1 ) − hG) ã(tn+1 ) + ΦTq (q̄n+1 )λ̃(tn+1 )
∂Q
G=γ
(t, q, v) −Q(tn+1 , v̄n+1 , q̄n+1 ) + hGan = O(h2 )(47c)
· ∂v ¸
∂Q
−βh (M(q)a)q (q, a) + (Φq λ)q (q,t, λ) −
T
(t, q, v) (42)
∂q
1
Φq (q̄n+1 ,tn+1 )(ã (tn+1 ) − an ) = − Φ(q̄n+1 ,tn+1 ) + O(h)
with all the quantities above evaluated at t = tn+1 , q = q̄n+1 , v = βh2
v̄n+1 , a = an and λ = λn . (47d)
For the true solution we have the following asymptotic estimates: where G is defined by (42).

In order to give the estimates for the local truncation error,

M(q̄n+1 )ã(tn+1 ) = M(q̃(tn+1 ))ã(tn+1 ) + O(h3 )(43) we assume that the generalized mass matrix M(q) is uniformly
ΦTq (q̄n+1 ,tn+1 )λ̃(tn+1 ) = ΦTq (q̃(tn+1 ),tn+1 )λ̃(tn+1 ) + O(h3 )(44) positive definite on the nullspace of the Jacobian of constraints
(one may assume the uniform positive definiteness of M on the
Q(tn+1 , v̄n+1 , q̄n+1 ) = Q(tn+1 , ṽ(tn+1 ), q̃(tn+1 )) + O(h2 )(45)
whole space, but the condition above is obviously less restric-
tive). As mentioned before, we also assume that the Jacobian Φq
Using that is full rank. To be more precise we assume that there is a matrix
valued mapping κ : Rn × R → Rn×(n−m) such that
M(q̃(tn+1 ))ã(tn+1 )
+Φq (q̃(tn+1 ),tn+1 )λ̃(tn+1 ) − Q(tn+1 , ṽ(tn+1 ), q̃(tn+1 )) = 0
T
¡ Φq (q,t)κ(q,t)¢ = 0, κ(q,t)¡T κ(q,t) = I, ¢
infq, t σ Φq (q,t)ΦTq (q,t) > 0, infq, t σ κ(q,t)T M(q)κ(q,t) > 0,
(48)
where σ(A), denotes the spectrum of the matrix A. In other
words the columns of the matrix κ(q,t) form an orthonormal ba-
sis of Φq (q,t)T and the eigenvalues of the symmetric positive

10 Copyright °
c 2005 by ASME
definite matrices Φq (q,t)M(q)ΦTq (q,t) and κT (q,t)M(q)κ(q,t) cessively
are bounded below by some positive constant.This implies that
the operator norms of their inverses are bounded, i.e.,
u = O(h), w = O(h), λ̃(tn+1 ) − λn+1 = O(h),
ṽ(tn+1 ) − vn+1 = O(h2 ), q̃(tn+1 ) − qn+1 = O(h3 ),
¡ ¢−1 ¡ ¢−1
sup k Φq (q,t)ΦTq (q,t) k < ∞, sup k κ(q,t)T M(q)κ(q,t) k < ∞.
q,t q,t
(49) which completes the proof of our corollary.

Corollary 2. Assume that conditions (48) are satisfied. Let

qn+1 , vn+1 , an+1 and λn+1 be the quantities obtained after one 6 NUMERICAL EXPERIMENTS
Newton iteration, with qn = q̃(tn ), vn = ṽ(tn ), an = ã(tn ) and 6.1 The simple pendulum
λn = λ̃(tn ). Then the following relations hold: The first example consists of a simple planar pendulum. A
point mass m (m = 1) is attached to a massless rod of length
q̃(tn+1 ) = qn+1 + O(h3 ), ṽ(tn+1 ) = vn+1 + O(h2 ), L = 1; all units are SI. The equations of motion are formulated in
terms of the Cartesian coordinates (x, y) of m, the motion being
ã(tn+1 ) = an+1 + O(h), λ̃(tn+1 ) = λn+1 + O(h). (50)
constrained by Φ(q) := x2 + y2 − L2 = 0. The initial position of
the point mass is given by
Proof. Using Theorem 1 and the fact that the numerical solu-
tion (qn+1 , vn+1 , an+1 , λn+1 ) satisfies exactly the estimates given ³ π π ´
in this theorem we obtain that q0 = (x0 , y0 ) = sin( ), − cos( ) ,
3 3

q̃(tn+1 ) − qn+1 = βh2 (ã(tn+1 ) − an+1 ) + O(h3 ) (51a) and the initial velocity is taken to be zero. The system was sim-
ulated for an interval of T = 4 (s). To obtain a reference so-
lution (qre f , vre f ), the equations of motion are reformulated as
a system of ordinary differential equations which are numeri-
ṽ(tn+1 ) − vn+1 = γh (ã(tn+1 ) − an+1 ) + O(h2 ) (51b) cally solved with the Matlab ODE solver ode45. To enforce
a high accuracy for the reference solution the relative and ab-
solute tolerances of the Matlab solver are set to RelTol= 10−8
and AbsTol= 10−16 , respectively.
(M(q̄n+1 ) − hG) (ã(tn+1 ) − an+1 ) For different values of the time-step h, we compute
³ ´ the error in position and velocity at the final time T = 4.
+ΦTq (q̄n+1 ) λ̃(tn+1 ) − λn+1 = O(h2 ) (51c) To be more precise, in what follows we look at the errors
∆hq = ||qh (T ) − qre f (T )||2 and ∆hv = ||vh (T ) − vre f (T )||2 , where
qh and vh are the numerical positions and velocities obtained
with the Newmark scheme when applied to the corresponding
Φq (q̄n+1 ,tn+1 )(ã (tn+1 ) − an ) = O(h) (51d) DAE. The two tables below list the above errors as well as the
rate of decrease in these errors as the time-step h is successively
∆hq
Now let us denote A = ΦTq (q̄n+1 ,tn+1 ), B = κ(q̄n+1 ,tn+1 ) and halved. More precisely this rates are defined by Rhq = (h/2)
∆q
Mb = M(q̄n+1 ) − hG. Then there are unique vectors u and w such ∆hv
that ã(tn+1 ) − an+1 = Au + Bw, and from (51a–51d), we obtain and Rhv = (h/2) . As it can be seen from Table 1, for γ = 1/2
∆v
and β = 1/4, i.e., parameters that correspond to the trapezoidal
method, the order of convergence both in positions and velocities
q̃(tn+1 ) − qn+1 = βh (Au + Bw) + O(h )
2 3
is 2.
ṽ(tn+1 ) − vn+1 = γh(Au + Bw) + O(h2 )
³ ´
b
AT MAu b
+ AT MBw + AT A λ̃(tn+1 ) − λn+1 = O(h2 )
b
BT MAu b
+ BT MBw = O(h2 )
6.2 Double Pendulum
AT Au = O(h)
Validation of the convergence order is also carried out us-
ing the double pendulum mechanism shown in Fig.1. This is
Now by using the above relations in reverse order, we get suc- the same model that has been recently used in [28, 29]; a large

11 Copyright °
c 2005 by ASME
 h ∆hq Rhq ∆hv Rhv

2−4 7.72E − 2 N/A 7.46E − 1 N/A

2−5 1.49E − 2 5.20E + 0 2.94E − 2 2.54E + 1
2−6 4.48E − 3 3.31E + 0 1.10E − 2 2.68E + 0
2−7 1.13E − 3 3.97E + 0 3.42E − 3 3.21E + 0
2−8 2.82E − 4 4.00E + 0 9.02E − 4 3.79E + 0
2−9 7.05E − 5 4.00E + 0 2.29E − 4 3.94E + 0
2−10 1.76E − 5 4.00E + 0 5.73E − 5 3.99E + 0
2−11 4.41E − 6 4.00E + 0 1.44E − 5 3.99E + 0

Table 1. Single Pendulum. Error results for γ = 1/2, β = 1/4.

h ∆hq Rhq ∆hv Rhv

Figure 1. Double pendulum problem
2−4 1.56E − 1 N/A 1.13E + 0 N/A
2−5 6.21E − 2 2.52E + 0 7.38E − 1 1.53E + 0
2−6 2.26E − 2 2.75E + 0 4.27E − 1 1.73E + 0
around an axis perpendicular on the plane of motion. For body
2−7 1 = ẏ1 = θ̇1 = 0, while for body 2, ẋ2 = 3.8822857,
1, ẋCM
8.19E − 3 2.76E + 0 2.31E − 1 1.85E + 0 CM CM CM

ẏ2 = 14.4888887, and θ̇2 = 10. All initial conditions are in SI

CM CM
2−8 3.15E − 3 2.60E + 0 1.20E − 1 1.92E + 0
units, and are consistent with the kinematic constraint equations
2−9 1.31E − 3 2.40E + 0 6.12E − 2 1.96E + 0 at position and velocity levels (Eqs.(4a) and (4b)).
2−10 5.88E − 4 2.23E + 0 3.09E − 2 1.98E + 0 The differential equations that govern the time evolution of
the six generalized coordinates used to model the double pendu-
2−11 2.77E − 4 2.13E + 0 1.55E − 2 1.99E + 0 lum are
(γ+1/2)2
Table 2. Single Pendulum. Error results for γ = 3/4, β = 4 .

m1 ẍ1 = −λ1 − λ3 (52a)

amount of stiffness is induced by means of two rotational spring- m1 ÿ1 = −λ2 − λ4 − m1 g (52b)
damper-actuators (RSDA). The masses of the two pendulums m1 L12
are m1 = 3 and m2 = 0.3, the dimension of the pendulums are θ̈1 = L1 sin(θ1 )(λ3 − λ1 )
3
L1 = 1 and L2 = 1.5, the stiffness coefficients are k1 = 400
and k2 = 3.E5, and the damping coefficients are c1 = 15 and +L1 cos(θ1 )(λ2 − λ4 ) + Q(1) (52c)
c2 = 5.E4. The zero-tension angles for the two RSDA elements m2 ẍ2 = λ3 (52d)
are α01 = 3π/2 and α02 = 0. All units are SI. m2 ÿ2 = λ4 − m2 g (52e)
In its initial configuration, the two degree of freedom dynamic m2 L22
system has a dominant eigenvalue with a small imaginary part θ̈2 = L2 sin(θ2 )λ3 − L2 cos(θ2 )λ4 + Q(2) (52f)
3
and a real part of the order -10E5. Since the two pendulums
are connected through two parallel revolute joints the problem
They are subjected to the algebraic constraints
is planar. In terms of initial conditions, the centers of mass
(CM) of bodies 1 and 2 are located at xCM CM
1 = 1, y1 = 0, and  
CM CM
x2 = 3.4488887, y2 = −0.388228. In the initial configura- x1 − L1 cos(θ1 )
tion, the centroidal principal reference frame of body 1 is paral-  y1 − L1 sin(θ1 ) 
Φ(q) =  
 x1 + L1 cos(θ1 ) + L2 cos(θ2 ) − x2  = 0. (52g)
lel with the global reference frame, while the centroidal prin-
cipal reference frame of body 2 is rotated with θ2 = 23π/12 y1 + L1 sin(θ1 ) + L2 sin(θ2 ) − y2

12 Copyright °
c 2005 by ASME
 6.5 8

6
6
4

5.5 2

[radians]
θ [radians]

0
5

θdot
−2
1

1
4.5 −4

−6
4
−8

3.5 −10
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
TIME [seconds] TIME [seconds]

(γ+1/2)2 (γ+1/2)2
Figure 2. Time variation of orientation θ1 , for γ = 34 , β = 4 and Figure 3. Time variation of angular velocity θ̇1 , for γ = 34 , β = 4
h = 0.001 and h = 0.001

Here Q(1) = 3π 2 k1 − (k1 + k2 )θ1 − (c1 + c2 )θ̇1 + k2 θ2 + c2 θ̇2 , h ∆hq Rhq ∆hv Rhv
Q = k2 (θ1 − θ2 ) + c2 (θ̇1 − θ̇2 ), and g = 9.81 is the gravita-
(2)

tional acceleration. To obtain a reference solution later used in 2−9 7.76E − 2 N/A 1.59E − 1 N/A
calculating the error, the above DAE is reduced to a state space 2−10 1.26E − 2 6.17E + 0 9.39E − 2 1.69E + 0
ODE in the two rotation angles θ1 and θ2 . The resulting system
of second order ordinary differential equations assumes the form 2−11 5.85E − 3 2.15E + 0 4.82E − 2 1.95E + 0
2−12 2.82E − 3 2.07E + 0 2.44E − 2 1.98E + 0
m1
4L12 ( + m2 )θ̈1 + 2m2 L1 L2 cos(θ1 − θ2 )θ̈2 = Q(1) 2−13 1.39E − 3 2.04E + 0 1.23E − 2 1.99E + 0
3
−2m2 L1 L2 sin(θ1 − θ2 )θ̇22 − (m1 + 2m2 )gL1 cos(θ1 ) (53) 2−14 6.85E − 4 2.02E + 0 6.16E − 3 1.99E + 0
4
2m2 L1 L2 cos(θ1 − θ2 )θ̈1 + m2 L22 θ̈2 = Q(2) Table 3. Double Pendulum. Error results for γ = 3/4, β =
(γ+1/2)2
.
3 4
+2m2 L1 L2 sin(θ1 − θ2 )θ̇21 − m2 gL2 cos(θ2 ). (54)
7 CONCLUSIONS
The reference solution (qre f , vre f ) is obtained by passing the The Newmark method used in structural dynamics was
above ODE to the Matlab solver ode23s. To enforce a high adapted in this paper for the numerical solution of index three
accuracy for the reference solution the relative and absolute tol- differential algebraic equations of multi-body dynamics. Strate-
erances of the Matlab solver are set to RelTol= 10−8 and gies for corrector stopping criteria, and error and step-size con-
AbsTol= 10−16 , respectively. trol were introduced, and a formal proof for local convergence of
The time evolution of this model was simulated for T = 2 the method was presented in detail. A set of numerical experi-
(s), and an error analysis similar to the one done for the simple ments confirmed the expected convergence order. A forthcoming
pendulum example is provided in Table 3. paper [30] will present an analysis of the global convergence be-
The error analysis results clearly indicate that as predicted havior of the method, and compare it with a BDF type integrator.
by the global convergence analysis [30], Newmark with a choice The proposed method has been implemented in the 2005 version
2
of parameters γ = 3/4, β = (γ+1/2)
4 is a first order method even of the commercial simulation package MSC.ADAMS and led to
when used for the integration of the index 3 DAEs associated a significant reduction in the simulation time for large models
with constrained mechanical systems. In [30], a set of real-life containing flexible bodies and/or contact elements.
mechanical systems (engine models) are used to show improved
efficiency for the Newmark method when compared to the BDF
type integrators currently implemented in MSC.ADAMS [24]. 8 ACKNOWLEDGMENT
Typically, it has been noticed that for engine models that lead to The work of the first and third author was supported in part
index 3 DAEs in excess of 16,000 equations, the proposed New- by the National Science Foundation, Grant No. 0139701. The
mark method reduced the simulation lengths by a factor from 2 second author would like to thank Nicolae Orlandea whose pas-
to 4 times. sion and enthusiasm is a continuous source of motivation.

13 Copyright °
c 2005 by ASME
 View publication stats

REFERENCES and J. A. C. Ambrosio, eds., Kluwer Academic Publishers.

[1] Newmark, N. M., 1959. “A method of computation for [17] Yen, J., Petzold, L., and Raha, S., 1996. A time integra-
structural dynamics”. Journal of the Engineering Mechan- tion algorithm for flexible mechanism dynamics: the DAE
ics Division, ASCE, pp. 67–94. α-method. Tech. Rep. TR96-024, Dept. of Comp. Sci., Uni-
[2] Ascher, U. M., and Petzold, L. R., 1998. Computer Meth- versity of Minnesota.
ods for Ordinary Differential Equations and Differential- [18] Cuadrado, J., Dopico, D., Naya, M. N., and Gonzalez, M.,
Algebraic Equations. SIAM, Philadelphia, PA. 2004. “Penatly, semi-recursive and hybrid methods for
[3] Hilber, H. M., Hughes, T. J. R., and Taylor, R. L., 1977. MBS Real-Time dynamics in the context of structural in-
“Improved numerical dissipation for time integration al- tegrators”. Multibody Systems Dynamics(12), pp. 117–132.
gorithms in structural dynamics”. Earthquake Eng. and [19] Baumgarte, J., 1972. “Stabilization of constraints and inte-
Struct. Dynamics, 5, pp. 283–292. grals of motion in dynamical systems”. Computer Methods
[4] Chung, J., and Hulbert, G. M., 1993. “A time integration in Applied Mechanics and Engineering, 1, pp. 1–16.
algorithm for structural dynamics with improved numeri- [20] Owren, B., and Simonsen, H., 1995. “Alternative integra-
cal dissipation: the generalized-a method”. Transactions of tion methods for problems in structural dynamics”. Struc-
ASME, Journal of Applied Mechanics, 60(2), pp. 371–375. tural Dynamics. Comput. Meth. in Appl. Mech. and Eng.,
[5] Featherstone, R., 1983. “The calculation of robot dynamics 122, pp. 1–10.
using articulated-body inertias”. International Journal of [21] Jay, L. O., 1998. “Structure preservation for constrained
Robotics Research(1), pp. 13–30. dynamics with super partitioned additive runge-kutta meth-
[6] de Jalon, J. G., Unda, J., and Avello, A. “Natural coor- ods”. SIAM J. Sci. Comput., 20, pp. 416–446.
dinates for the computer analysis of multibody systems”. [22] Schaub, M., and Simeon, B., 2002. “Automatic h-scaling
Computer Methods in Applied Mechanics and Engineer- for the efficient time integration of stiff mechanical sys-
ing(56), pp. 309–327. tems”. Multibody System Dynamics, 8, pp. 329–345.
[7] Haug, E. J., 1989. Computer-Aided Kinematics and Dy- [23] Dennis, J., and Schnabel, R., 1983. Numerical Methods
namics of Mechanical Systems. Prenctice-Hall, Englewood for Unconstrained Optimization and Nonlinear Equations.
Cliffs, New Jersey. Prentice-Hall, Englewood Cliffs, NJ.
[8] de Jalon, J. G., and Bayo, E., 11994. Kinematic and [24] MSCsoftware, 2005. ADAMS User Manual.
Dynamic Simulation of Multi-body Systems. The real time [25] Zienkiewicz, O., and Xie, Y., 1991. “A simple error es-
challenge. Springer-Verlag, Berlin. timator and adaptive time-stepping procedure for dynamic
[9] Shabana, A. A., 1994. Computational Dynamics. John analysis”. Earthquake Engineering and Structural Dynam-
Wiley & Sons. ics, 20, pp. 871–887.
[10] Brenan, K. E., Campbell, S. L., and Petzold, L. R., [26] Atkinson, K. E., 1989. An introduction to numerical analy-
1989. Numerical Solution of Initial-Value Problems in sis, second ed. John Wiley & Sons Inc., New York.
Differential-Algebraic Equations. North-Holland, New [27] Hughes, T. J. R., 1987. Finite Element Method - Linear
York. Static and Dynamic Finite Element Analysis. Prentice-Hall,
[11] Petzold, L. R., 1982. “Differential-algebraic equations are Englewood Cliffs, New Jersey.
not ODE’s”. SIAM J. Sci., Stat. Comput., 3(3). [28] Negrut, D., Haug, E. J., and German, H. C., 2003. “An im-
[12] Hairer, E., and Wanner, G., 1991. Solving Ordinary Dif- plicit Runge-Kutta method for integration of Differential-
ferential Equations, Vol. II of Computational Mathematics. Algebraic Equations of Multibody Dynamics”. Multibody
Springer-Verlag. System Dynamics, 9(2), pp. 121–142.
[13] Potra, F. A., 1993. “Implementation of linear multi- [29] Oghbaei, M., and Anderson, K., 2005. “A new implicit inte-
step methods for solving constrained equations of motion”. gration scheme for the simulation of stiff-complex mechan-
SIAM. Numer. Anal., 30(3). ical systems-DETC2005-84471”. In Proceedings to the
[14] Lubich, C., Nowak, U., Pohle, U., and Engstler, C., 1995. Fifth ASME International Conference on Multibody Sys-
“MEXX - numerical software for the integration of con- tems, Nonlinear Dynamics and Control, A. Shabana and
strained mechanical multibody systems”. Mechanics of K. Anderson, eds., ASME.
Structures and Machines, 23, pp. 473–495. [30] Negrut, D., Gavrea, B., and Potra, F., 2005. “Multibody dy-
[15] Eich-Sollner, E., and Fuhrer, C., 1998. Numerical Methods namics simulation with Newmark integrators”. In Prepara-
in Multibody Dynamics. Teubner-Verlag, Stuttgart. tion.
[16] Cardona, A., and Geradin, M., 1994. “Numerical integra-
tion of second order differential-algebraic systems in flexi-
ble mechanics dynamics”. In Computer-Aided Analysis of
Rigid and Flexible Mechanical Systems, M. F. O. S. Pereira

14 Copyright °
c 2005 by ASME