VOL. 13, NO. 1, JAN.-FEB. 1990 J.

GUIDANCE 113

Role of Maggi's Equations in Computational Methods for

Constrained Multibody Systems
Andrew Kurdila,* John G. Papastavridis,t and Manohar P. KamatJ
Georgia Institute of Technology, Atlanta, Georgia

This paper presents a unified theoretical basis for a class of methods that generate the governing equations of
constrained dynamical systems by eliminating the constraints. By using Maggi's equations in conjunction with a
common projective theory from numerical analysis, it is shown that members of the class are precisely characterized
by the basis they choose for the null-space of the variational form of the constraints. For each method considered,
the specific basis chosen for the null-space of the variational constraints is derived, as well as a dual basis for the
orthogonal complement. The latter basis is of particular interest since it is shown that its knowledge theoretically
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

enables one to generalize certain methods of the class to calculate constraint forces and torques. Practical
approaches based on orthogonal transformations to effect this strategy are also outlined. In addition, since the
theory presented herein stresses a common, fundamental structure to the various methods, it is especially useful as
a means of comparing and evaluating individual numerical algorithms. The theory presented makes clear the
relationship between certain numerical instabilities that have been noted in some methods that eliminate a priori
constraint contributions to the governing equations by selecting an independent subset of unknowns. It is also briefly
indicated how this formalism can be extended, in principle, to the wider class of nonlinear nonholonomic constraints.

Introduction A close examination of the formalism of Wittenburg15 can

leave little doubt as to the complexity of some minimal
T HE need for general and efficient numerical techniques
for constrained mechanical systems is well established.
Studies concerning computational methods for constrained
methods, Likewise, Kane's method as described in Ref. 25,
although minimal, does not provide a means of constraint
dynamical systems have arisen in the modeling of mechanism force evaluation that is as systematic as Lagrangian ap-
dynamics1"3 and in the simulation of spacecraft.4"6 Further- proaches.
more, extensive growth is expected in techniques for the Within the past few years, several related methods have
optimization of mechanism dynamics7 and for large space arisen that first pose the constrained dynamics problems in
structure simulation and control.8 terms of a redundant set of unknowns and subsequently
One means of classifying the diverse methods that have reduce the system to minimal order for solution. Thus, these
been developed in this field categorizes the techniques as to methods seek to combine the simplicity of redundant formula-
whether they employ a minimal or redundant coordinate set. tions with the numerical efficiency of minimal techniques. The
This issue has been the focus of much discussion in the past, first such techniques that have appeared in the literature
particularly in comparing the relative merits of various tech- attempt to circumvent the numerical difficulties sometimes
niques.9"12 Lagrangian methods in which the constraint forces noted when implicit integration schemes are applied to mixed
are explicitly represented by multipliers13'14 and Newton-Euler systems of algebraic and differential equations. The method of
methods that retain interbody forces and torques are typical coordinate partitioning suggested by Wehage and Haug17 uses
of the redundant formulations. On the other hand, Kane's a Lagrangian formulation and Gaussian elimination to select
technique11 and Newton-Euler methods that analytically elim- a subset of the original unknowns, consequently eliminating
inate the constraint forces via graph theory15 are examples of the multipliers. As a result of poor conditioning in some
formalisms that can be employed to obtain a "minimal" submatrices required in the method, Mani18 employs singular-
number of unknowns. value decomposition instead of Gaussian elimination to select
Although it would seem desirable to use the least number of an independent subset of unknowns. The resulting technique
unknowns possible in formulating any particular problem, the is shown to have better stability characteristics. Kim and
choice regarding minimal or redundant coordinates in dynam- Vanderploeg19 further improve on the numerical efficiency of
ics is not always clear-cut. As most succinctly put by Schwer- Mani's work by introducing "null-space updating" based on
tassek and Roberson,16 QR decomposition.
As opposed to these Lagrangian-based methods, another
Minimal sets of equations are more difficult to derive, more set of algorithms has been developed using Kane's equations
complicated in form, possibly faster to solve, but might not as the theoretical foundation. Singh9 and Huston10 suggest
yield directly all the desired information about the systems modifications to Kane's method based on singular-value de-
(e.g., constraint forces and torques). Descriptions with redun- composition and eigenvector decomposition, respectively.
dant equations are more easily derived, generally simpler in Both approaches pose the governing equations in terms of
form, slower to solve, but provide additional information. Kane's equations without considering the constraints, and
then reduce the system numerically to a smaller set for
solution. More recently, Wang and Huston20 derived a formu-
lation based on Kane's method that includes "undetermined
Received Nov. 16, 1987; revision received May 30, 1988. Copyright multipliers," such as in Lagrange's equations. The multiplica-
© 1989 American Institute of Aeronautics and Astronautics, Inc. All tive terms are eliminated numerically through the use of
rights reserved. "orthogonal complement" arrays. In Ref. 31, Wampler et al.
* Associate Professor, Engineering Science and Mechanics Group;
currently, Department of Aerospace Engineering, Texas A&M Uni-
derives an "explicit" formulation, based on Kane's method,
versity. that is appropriate for symbolically, or numerically, generat-
tAssociate Professor, School of Mechanical Engineering. ing a constraint-free system of equations.
^Associate Professor, School of Aerospace Engineering. Associate One potential difficulty in assimilating and evaluating this
Fellow AIAA. class of methods as a whole may be attributed to the fact that
 114 KURDILA, PAPASTAVRIDIS, AND KAMAT J. GUIDANCE

pertinent research in the spacecraft and mechanism dynamics recently, the method has not been amenable to actual calcula-
fields has been, to some degree, mutually isolated. This obser- tions and primarily has been of academic interest only. How-
vation has been noted by other authors.3 The lack of an ever, as shall be shown in this section, the advent of digital
overall viewpoint can be particularly disconcerting to those computers now permits a wide variety of practical numerical
interested in a "global" perspective of computational tech- implementations of the approach. Because of the apparent
niques for constrained dynamical systems. For example, the lack of familiarity with the technique, this section briefly
authors are interested in the development of computational develops the method.
algorithms appropriate for concurrent multiprocessors. In this Let S be a system of v particles, each of whose position r
context, the details involved in deriving the governing equa- relative to an inertial frame of reference is a function of N
tions via Newton's equations, Lagrange's equations, or generalized coordinates qk9 K = 1, .. . 9N and the time t. Sim-
Kane's technique are not nearly as important as determining ply stated, the principle of virtual work requires that the
the fundamental structure and numerical processes underlying difference in the virtual work of the applied forces and the
a wide variety of specific methods. virtual work of the inertial forces must vanish for all virtual
In this paper, it is shown that the class of computational displacements consistent with the constraints. Likewise, as a
methods described earlier9"11'17"20'31 may be characterized consequence of Newton's second law of motion for each
concisely in terms of Maggi's equations. The theoretical unifi- particle, the virtual work of the constraint forces must vanish
cation of these distinct techniques is accomplished by re- independently.
examining Maggi's equations in terms of modern linear oper-
ator theory. It is demonstrated that each method performs the
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

dma - dr — \ d f - d r (la)
following fundamental numerical steps:
1) The equations governing each of the constraint-free, or
"relaxed," substructures are formed and then assembled into (Ib)
a set of system equations.
2) The equations of the "released" system are required to
be orthogonal to a specific subspace for solution, particularly In the preceding expressions, df is the resultant of all forces
the null-space of the variational constraints. This step is acting on a particle /?, excluding constraint forces dR, and all
equivalent to requiring that the projection of the system integrations are carried out over the entire system of particles.
equations onto the null-space of the variational constraints is The quantities dm, a, and dr are the mass, acceleration, and
zero. virtual displacement of a particle p. Because one can express
3) Constraint forces may be evaluated, if desired, by pro- the contemporaneous variation of r as a linear combination of
jecting the system equations onto the orthogonal complement the variations in the generalized coordinates 6qk,
of the null-space of the constraints.
It is shown, in fact, that the algebraic manipulation carried dr = ekdqk
out in each technique to produce a reduced system of equa-
tions simply amounts to selecting a different basis for the where
null-space of the constraints and its orthogonal complement.
Several practical benefits are derived from the theory pre- dr
(2)
sented. Foremost among the advantages of the representation
is that the theory emphasizes the common structure of these
problems. For example, although the methods based on La- Equation (la) can be rewritten in the following familiar form:
grange's equations may all perform the constraint force calcu-
lations outlined in step 3 above, Kane's technique and the (dma - d f ) - i'ekdqk=0 (3a)
"Kane-based methods" in Refs. 9, 10, and 31 do not explicitly
represent the constraint force contributions in their minimal or
formulations and, consequently, cannot present a general
strategy for their solution. In fact, Singh9 advocates using a (3b)
Lagrangian formulation if constraint forces are desired. On
the other hand, Wampler et al.31 suggest that a redundant where
formulation is questionable unless constraint forces are re-
quired. Although Wang and Huston20 remedy this situation dT^
by bringing the constraint force contributions into evidence, Ek(T)^(^r}-^
' dt \dqk) dqk
the relationship of the undetermined multipliers to those
appearing in a Lagrangian approach is not made clear. Fur-
thermore, no general plan of attack is suggested for their T= -dmv - v
calculation. It is shown herein that these methods all may be
generalized to calculate constraint forces in a systematic fash-
ion using a unified theory that exploits the fundamental
structure of the problem.
The remainder of this paper is organized into three sections.
4
and the summation convention has been used in Eqs. (2) and
The next section reviews Maggi's equations and their relation (3). For unconstrained problems, the variations bqk are inde-
to projective methods in operator theory. The following sec- pendent and their respective coefficients in Eqs. (3) must be
tion briefly outlines the relationship of the methods in Refs. 9, identically zero for k = 1,...,7V. With respect to constrained
10, and 17-19 to Maggi's equations. Finally, the last section problems, Eqs. (3) have limited direct application since dqk
summarizes some important benefits derived from the unified are not independent.
theory. However, it is possible to create independent virtual dis-
placements, thus resulting in a useful set of equations similar
to Eqs. (3). For holonomic systems, and a large class of
Maggi's Equations nonholonomic systems, constraints on the system can be
Maggi's equations were first presented in 1896, and then expressed in terms of D equations that are linear in the
again in 1901. Although the method is conspicuously absent differentials of the generalized coordinates as
from most texts in the English language, Neimark and Fu-
faev21 summarize the essentials of the theory. Until quite atk dqk + bi dt = 0, / = !,... ,D (4)
 JAN.-FEB. 1990 COMPUTATIONAL METHODS FOR MULTIBODY SYSTEMS 115

In order that the virtual displacements dr are consistent with where

the constraints, it is necessary that the variations in the
generalized coordinates dqk satisfy
= \dR - ek
aikdqk = 0 (5)
and dR = total constraint force on particle P, so that
Equation (5) can be alternately represented in terms of a dma =df+dR. Moreover, Eq. (Ib) requires that the TV vector
D x N matrix a and an N vector dq = [dql7...,dqN]T. (g(r) be orthogonal to dq. In other words, the TV vector g(r)
must be a member of the range (a *), the orthogonal comple-
adq = 0 (6) ment of the range (A). It follows that, for some D vector y,

Equations (5) and (6) now can be interpreted as requiring that g (r) =/>y (13)
the TV vector dq lie in the null-space of a, or that dq be
orthogonal to the range of aT. Making use of the identity in Eq. (13), it is possible to express
Eqs. (12) in either of two ways.
dqe. null-space (a) (7a)
:0 (14a)
dq J_ range (a *) (7b)
) (14b)
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

The range of aT is equal to the span of the columns of ar, and

is consequently of dimension D if the constraints are indepen- These two equations are clearly equivalent, as the columns of
dent. Since null-space (a) and range (a7) are orthogonal p are any basis for the range of aT. The two representations
subspaces, the dimension of null-space (a) is / = N — D. De- coincide when [piv,Pz>] are chosen to be the columns of aT.
ferring, for the moment, the discussion of their calculation, let Derivations that follow are carried out in the generic basis
[Al9...9A^ be any basis for the null-space of a and let [pi,.-.»pj {pD}, a description well suited to numerical techniques that
be any basis for the range of aT. The vector dq may lie automatically generate a basis for the range (a)T that does not
anywhere within the span of [Al9...9A^. In other words, it is coincide with the columns of (a7). Assuming that Eq. (11) has
true that been solved and that the original unknowns q and their
derivatives q and q have been reconstructed, the unknown
(8a) constraint contributions are simply the unique minimizers of
the TV-vector norm
dq = AdO (8b)
minimizer
Xe range (p) E(T)-Q-X\ (15a)
where dOt are arbitrary coefficients and A = [Al,...,AI] — [Akl]
is an TV x (N — D) matrix. or
Substituting Eq. (8) into Eq. (3b), a form of virtual work
emerges that is well suited for constrained systems. Using the (15b)
fact that each d9t is arbitrary, one obtains Maggi's equations.
Equations (11) and (15) suggest a simple "geometric"
(9a) interpretation to the problem of solving constrained dynamics
{Ek(T) - Qk}Aki = 0, i = 1,...,TV - D (9b) problems that has been widely used in applied linear algebra.
Equation (11) may be interpreted as a statement of orthogo-
In effect, Eqs. (9) reflect the fact that the virtual work due to nality. It requires that the TV vector E(T) — Q be perpendicu-
each of the independent, admissible virtual displacements drt, lar to the null-space (a). In other words, the projection of
i = 1,...,TV - D, obtained from Eqs. (8) and (2), E(T) — Q onto the null space of the variational constraints is
zero. Likewise, when Eqs. (15) require that g(r) is the vector
dr = ekdqk that lies in the range (a7) that is closest to E(T) - Q, it
implies that Q(r) is the orthogonal projection of E(T) — Q
Mi (10) onto range (a7), the complementary space to null-space (a):
summed over the system, be zero.
For ease of comparison to other methods, as in Refs. 9 and (16)
17, it is useful to express Eq. (10) in operator form as
These relationships are depicted symbolically in Fig. 1. Care-
T
A {E(T)-Q} = fully note that the orthogonality statement in Eq. (11) re-
(11) quires that the "abstract angle" 9 in Fig. 1 must be zero
The above TV — D equations, together with the D constraints
in Eq. (4), comprise a system of TV equations in terms of TV (17)
unknowns, and are sufficient to solve for the displacements,
velocities, and accelerations of the system. An illustrative
account of the actual details of implementing Eqs. (4) and where PA and PaT are the orthogonal projections onto the
(11) in a numerical simulation can be found in Ref. 9. null-space of a and the range of 0r, respectively.
Supposing that Eq. (11) has been solved either analytically The solution of Eqs. (15) and (16) is given, in theory, by the
or numerically at a particular time t, the decomposition of RN normal equations
into span {Af} and span {pD} is also useful because it is
possible to solve for constraint forces from the base vectors (18)
IPiv-PzJ. Because the constraint forces perform zero virtual
work, the formulations until now have not explicitly con- A strategy similar to this approach in the literature on
tained the constraint forces dR. One can always insist that constrained dynamics is derived via algebraic considerations
in Ref. 28. However, it is neither recommended nor usually
(12a) required to form the "generalized inverse" depicted in Eq.
(18).22 Instead, orthogonal factorizations are employed in
(12b) solving the minimization problem in Eqs. (15). If 9 is an
 116 KURDILA, PAPASTAVRIDIS, AND KAMAT J. GUIDANCE

orthogonal transformation, i.e., 9T9 = 7, then 9 is an isome- concise in its representation of the solution. A knowledge of a
try. In other words, the transformation 9 is norm-preserving and A conveys both the specific basis chosen to reduce the
so that constrained system to minimal order, as well as the projection
operator selected to calculate constraint forces. It will be
I E(T)-Q- fi« I = || 9{E(T) -Q- (19) shown presently that the theory is sufficiently detailed to
provide an excellent means of tabulating computational alter-
A judicious choice of 9 can greatly simplify the form of the natives that have appeared in recent literature.
norm shown above, and lead to a straightforward solution
procedure. Typical choices of 9 are derived from the QR and Related Methods
SV decompositions.22
Equations (15), (16), and (19) provide a complete descrip- Kane's Method
tion of the solution for the constraint forces. When the Kane's method is a collection of steps, which, when fol-
calculations are performed via orthogonal transformations, lowed, can yield the smallest possible set of governing equa-
and not the normal equations, the solution exhibits well- tions for a constrained dynamical system. It will now be
known stability characteristics. However, Eqs. (15) and (16) shown that Kane's method can be considered as an analytical
are too general to permit a detailed comparison of methods procedure for choosing the null-space basis [Al9...,Af] in Eq.
that choose different bases in A and p. An alternative descrip- (11). The following steps summarize the approach:
tion of Maggi's equations given in Ref. 23, although not 1) Choose TV generalized coordinates qk, k = 19...9N9 suffi-
computationally efficient, depicts information regarding the cient to represent every configuration of the system. These
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

bases selected for p and A in a compact form. In addition, the coordinates may be redundant through choice, or will exceed
formalism displays enough information to reconstruct the the number of degrees of freedom of the system when the
projection operator PaT explicitly. In fact, when mutually system is constrained.
orthogonal bases are chosen for A and p, Maggi's equations 2) Choose N quantities Uk9 k = !,...,#, the "generalized
generate the same expressions obtained in the least-squares speeds," defined in terms of the generalized holonomic veloc-
solution of Eq. (19). ities q according to
Papastavridis23 generates the governing relations via Mag-
gi's equations by first supplementing the constraint matrix a (27a)
to form a full-rank TV x N matrix a.
or as

U=Yq (27b)

Next, the inverse A = a ~ l can be calculated and partitioned This definition is valid only if one can solve for each qk
in the following manner: uniquely in terms of Uk. That is, the matrix Y must be
invertible. If F"1 =y, one has
A] (21a)
q^ynUk+Z! (28a)
(21b) or
dA = l (22) (28b)
T
By inspection, it is clear that the columns of a and A' are 3) Express all constraints on the system in terms of D
reciprocal bases for the range (a7). Likewise, if the columns of equations of the form
a'T and A span the same space, they are dual bases for the
null-space (a). By multiplying Eq. (12a) by AT, one obtains
the governing relations

(A')T{E(T)-Q} =1 (23a)
(29)
(A)T{E(T)-Q} = 0 (23b)
These constraints arise either because the system is nonholo-
Upon comparing Eqs. (23) and (16), one can see that the nomic or by differentiating holonomic constraints due to
projection onto the range (a7) is easily constructed from the redundant coordinates qk. Another useful representation of
dual basis (A')T when aT and A' span the same space. these constraints is achieved by partioning the N vector U into
an / = (N — D) vector of "independent" generalized speeds U1
PaT=aT(A') (24) and a D vector of "dependent" generalized speeds UD.
The precise relationship defining the dual basis [Ai,...,A'D] of
[af,...,a£] can be extracted from Eq. (21), and is simply

A'TaT=aA' = (25) = C£7' + 0 (30)

T T
But if we choose p = a in Eq. (18), it is clear that (A ') is The matrix W is D x 7, the matrix = C is N x 7, and / is an
equal to the coefficient matrix (aaT)~la since there is exactly 7 x 7 identity matrix. Note that cC = 0, just as aA = 0 in Eqs.
one dual basis corresponding to [0ir,...,0#] and (21).
4) By virtue of the definitions in Eqs. (28) and (30), the
(26) velocity of every particle in the system can be expressed as
Thus, in this case, the normal equation's solution reduces to v= + »„ / = !,.. .,7 (31)
Eqs. (23).
In cases where the dual bases comprising the matrices a and The preceding summation only includes elements of the inde-
A can be derived analytically, the notation in Ref. 23 is pendent generalized speeds U1. In the absence of constraints,
 JAN.-FEB. 1990 COMPUTATIONAL METHODS FOR MULTIBODY SYSTEMS 117

the summation runs from 1 to N. The quantities vt are defined corresponding constraints in the notation of the preceding
to be the "partial velocities" of particle p. section are
5) Lastly, Kane postulates the governing dynamical equa-
tions cU + 0 = 0 (37a)

T = 0, (32) (37b)

where F{ and Ff are defined via integration over all of the so that
particles.
= cY = [I -W]Y (38)
(33a) But the columns of the matrix A =yC clearly constitute a
basis for the null-space of the variational constraints since
Ff = -\ dma • vt (33b)
aA = (39)

Although Kane's terminology is quite different, his method Alternatively, the analytic nature of this formulation enables
can be derived in terms of the more common virtual work the solution procedure to be represented by the method
expressions of the preceding section. If Eqs. (28) and (30) are described in Ref. 23. In this case, by combining Eqs. (29) and
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

substituted into (36), one obtains the dual bases that may be used to calculate
the reduced equations as well as the constraint forces in Eqs.
dr dr (23). These bases are depicted in Table 1. It should be pointed
out, however, that it is not meant to suggest in Table 1 that
the matrices y and Y need to be formed in a practical
(34) implementation of Kane's method or its related variations. As
will be illustrated, the dual bases can be utilized in this case
then the partial velocities may be obtained by inspection as without forming y or Y.

(35) Coordinate Partitioning

Generalized coordinate partitioning17 is a computational
Upon inserting the preceding partial velocities into Eq. (32) method that uses Gaussian elimination with full pivoting on
and rearranging the summations, Kane's governing equations the D x N constraint matrix a to select a subset of / = N — D
are just independent generalized velocities q1 and D dependent veloc-
ities qD. If a has full row rank, then it is always possible to
(36) partition a into [aD a1} such that aD is a D x D nonsingular
matrix.
or
aq + b = [aD a1} (40)
{Ek(T)-Qk}yklCn =
or A= (41)

With this choice of independent velocities q1, the columns of

the matrix A are a basis for the null-space of the constraints.
The similiarity between Eqs. (36) and (9) is immediately Furthermore, premultiplication of Lagrange's equations by
apparent. To prove that Kane's method may be considered a AT as in Eq. (11) yields precisely the governing equations
null-space generating method, it only remains to show that obtained by Wehage and Haug in Eq. (26) of Ref. 17. As in
the range (yC) is indeed the null-space of the constraints. If Kane's method, the analytical nature of this strategy permits
the constraints are given in Kane's notation as in Eq. (29), the a closed-form expression of the dual bases in Eqs. (21), which

Table 1 Example bases

Reference Decomposition

Gaussian
elimination

Mani 1984 Singular value T

Q* decomposition \ *
Kirn 1986

Singh 1985
SV decomposition

-W W
Kane/Huston 1987
 118 KURDILA, PAPASTAVRIDIS, AND KAMAT J. GUIDANCE

are displayed in Table 1. It should also be noted that premul- One can now change basis from y to A by the normal
tiplication by (A')T for the method by Wehage and Haug in equations A = (oaT)~1aQly. Making this replacement above
Table 1 generates identical expressions for the Lagrange gives the equations used to calculate the Lagrange multipliers
multipliers in Eq. (24) of Ref. 17. in Ref. 19.
Mani's Method (47)
Although Gaussian elimination has been used quite effec^
tively in some simulations utilizing coordinate partitioning,27 The dual bases appropriate for this method are shown in
one deficiency in the method has been noted in Ref. 18. The Table 1. Again, Eq. (47) can be derived in short order by
choice of a fixed set of independent coordinates may lead to simply extracting the dual basis A' from Table 1 for Kirn's
poorly conditioned matrices that must be factored. On in- method.
specting the null-space matrix A in Eq. (41), one can observe
that the columns of A span the null-space of a only as long as Singh/Huston's Method
(aD)~l exists. Since the matrix a = [aD a1] is time varying, in As another example of computational applications, Singh
general, one cannot guarantee that this inverse is well defined and Likins9 present a modification of Kane's method based
for the entire time span of interest. on singular-value decomposition that is similar to Mani's
Mani addresses this problem by utilizing the singular-value work. In this approach, Kane's equations are first generated
decomposition of ar, for the unconstrained system, i.e., for C — /in Eq. (25). As in
Mani's method, the singular-value decomposition is
aT=[Ul
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

(42)
(48)
where
where C, Z, and V are defined in the manner analogous to
U = [Ul U2] is an N x N orthogonal matrix (43a) that shown in Eq. (42), is then used to obtain a basis [U2] for
the null-space of the coefficient matrix c in Eq. (29). It is
= I M i s an N x D diagonal matrix (43b) subsequently shown that the reduced governing equations are

(49)
and V is a D x D orthogonal matrix. The diagonal terms a
are the singular values of the constraint matrix. In terms of Comparing Eqs. (9), (12), and (26) with the preceding, one
the theory presented earlier, the integration process carried can alternatively express the equations derived by Singh and
out by Mani is equivalent to choosing the dual bases as Likins as
depicted in Table 1. It should be noted that the actual
algorithm outlined in Ref. 18 is quite complicated in that the (50)
singular-value decomposition is carried out on an "as needed"
basis at certain time steps, and the constraint violation is These equations are identical to those derived in Ref. 20 for a
corrected by employing Newton-Raphson iteration. very specific choice of basis for the null-space of the con-
Null-space Updating straints. Since the columns^of U2 are a basis for the null-space
A more recent technique is based on the QR decomposition of c, it is clear that A = yU2 is a basis for the null-space of a.
of the constraint matrix a.19 In this decomposition, the N x D Consequently, it is apparent that this method simply chooses
matrix aT is expressed as the product of an N x N orthogonal a different basis for the null-space of a in the method of
matrix Q and a D x N upper-triangular matrix R. It is a virtual work. As with all the previous methods, this approach
characteristic of the decomposition that the matrix Q can be also may be succinctly represented in terms of the dual bases
partitioned into two matrices Q1 and Q2, where the columns it uses for the null-space of the variational constraints and its
of Ql are a basis of the range of aT and the columns of Q2 are orthogonal complement. These choices are found in Table 1.
a basis for the null-space of a,
Advantages of the Unified Theory
Several benefits can be derived from the theory discussed in
Q2] (44) the section regarding Maggi's equations. In one sense, all of
the advantages of the approach stem from the fact that the
where Q is an TV x N orthogonal matrix and R an N x D description given enhances one's ability to compare and con-
upper-triangular matrix. trast specific techniques. For instance, in this section, it is
In Ref. 19, Ql and Q2 are used to create two other matrices shown that the features of Lagrange's equations that make
^! and ()2 which, although not orthogonal, still satisfy constraint force calculation possible may be found in other
Maggi's equations. reduced formulations, if they are perceived in a "unified"
context. It is also demonstrated that certain ill-conditioning
Qi = Q\(aQ\)~l (45a) noted in coordinate partitioning methods is predicted by the
unified theory, and has been found, in some implementations
a lfl
62 = V — Qi( Qi) ~ ](?2 (45b) of Kane's method. As a final example, it is illustrated that the
method provides a foundation for numerical methods suitable
Choosing Q2 as the matrix A and premultiplying Lagrange's for a wider class of constraints.
equations by A T again produces the reduced governing equa-
tion in Eq. (33) of Ref. 19. Constraint Force Calculation
The calculation of Lagrange multipliers in Ref. 19 is also a One benefit of the general null-space decomposition theory
special case of the general method in the preceding section of presented earlier is that it enables calculation of Lagrange
this paper. Choosing the columns of Ql as the basis of the multipliers from Kane's equations after the reduced system of
range of a r , Eq. (18) allows one to calculate the Lagrange equations has been solved. Neither Kane's method25 nor its
multipliers referred to this basis. recent modifications9'10'31 have suggested an equivalent means
of calculating the eliminated constraint forces. Furthermore,
the development that follows extends the work in Ref. 20 in
two important respects. Although the derivation in Ref. 20 is
(46) limited to rigid bodies, no such assumption is made in this
 JAN.-FEB. 1990 COMPUTATIONAL METHODS FOR MULTIBODY SYSTEMS 119

paper. Second, although Ref. 20 establishes a form for Kane's (57b)

equations that retains multiplicative "undetermined multipli-
ers," the relationship between these terms and the multipliers (57c)
in a Lagrangian formulation is not explored.
From Eqs. (14) and (36), Kane's equations for the "re- Consequently, W — —(CD)~ICI and the basis of the null-space
leased" system, i.e., with C = /, are of the constraints is just

e-fl r A} = 0 (51a) ^Kane=J ~ ^ ^ * (58)

T T
-y a l=Q (51b)
Recalling for a moment the basis utilized in Ref. 17,
As a consequence, the constraint force contribution can be
calculated from V
^Weha8e = | " ,' " I (59)

= PyTa7{F + F*) (52)

the remarkable similarity in form in Eqs. (58) and (59) is
In this equation, F and F* are the generalized active forces immediately apparent. As noted earlier, Ref. 18 explains the
and generalized inertia forces for the unconstrained system, numerical difficulties associated with Wehage's method in
i.e., with C = I. However, if one is given the constraints in terms of the poor conditioning of (aD)~l. The preceding
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

terms of the generalized speeds, as shown in Eq. (29), the clarifies that a general formulation of Kane's method may
solution can be expressed exclusively in terms of matrices suffer the same fate if the condition of (cD)~l becomes poor.
available in Kane's formulation as The fact has been documented elsewhere (see, for example,
Ref. 31).
A = min* (53a)
Nonlinear Nonholonomic Constraints
= Pc7(F + F*) (53b)
A third advantage derives from one's ability to extend the
As noted earlier, specific forms for the projection theory easily to (first-order) nonlinear nonholonomic con-
P c r= projection onto the range of CTcan be derived from the straints, i.e., nonintegrable constraints of the form
normal equations, or from the dual basis for Kane's method
in Table 1. (60)
Following derivations along the same lines, Lagrange multi-
pliers also are readily calculated in Singh's approach. It is still Carefully note that these constraints differ from those shown
true that in Eqs. (4) and (29), in that they are not required to be linear
in the generalized velocities.
(54) The physical and analytical aspects of these constraints have
already been examined over the past 75 years (see, for exam-
However, in this method, the singular-value decomposition of ple, Refs. 29 and 30 and references cited therein). As of yet,
c has been calculated already. Using the fact that an orthogo- however, very little research on their computational treatment
nal transformation is an isometry, the solution of Eq. (54) can has appeared on the open literature. Such an extension is,
be calculated via standard techniques in numerical methods.22 however, straightforward, in principle, within the theory de-
scribed herein. Following the derivations carried out earlier in
(55) this paper and the developments outlined in Refs. 29 and 30,
one need only replace the constraint matrix a from Eq. (4) by
It is important to note that all of the information required to the Jacobian matrix of partial derivatives
evaluate the constraint forces is readily available and no
further matrix decomposition need be performed. This same (61)
result can be derived simply by inspecting the dual basis for
Singh's method in Table 1. Equations (53) and (55) illustrate
the fact mentioned earlier that the dual bases shown in Table With this new definition of the constraint matrix a, one can
1 can be utilized in practical implementations without expli- follow the same computational steps outlined in Eqs. (5-16)
citly forming the matrices y and y. to obtain the following governing equations:

AT{E(T)-Q}=$ (62a)
Explanation of Singularities
A second advantage of the unified theory may be attributed (62b)
to the ease with which the class of methods can be compared.
In this example, Singh9 noted certain "numerical instabilities" where aA = 0.
in a general implementation of Kane's method, but offer no The authors hope to examine in detail this nontrivial area
quantitative explanation for their origin. By comparing the in future publications.
basis implicitly chosen in Kane's method to that selected in
the method of Ref. 17, however, the nature of the singularities Hybrid Methods
is brought to light.
Finally, the theory presented herein suggests interesting
In Kane's method, it is assumed that the constraint matrix
c has the form "hybrid" methods that can be obtained from combining
methods shown in Table 1. For example, the method of
= [7 -W\ (56) null-space updating described in Ref. 19 need not be limited
only to Lagrangian formulations. By inspection of Table 1, it
This requirement is equivalent to stipulating that one can is clear that a null-space updating formulation of Kane's
partition a constraint matrix c = [CD \ c1] so that CD is invert- equations is obtained by choosing the dual bases
ible, whence
(63)
cU + <£' = 0 (57a)
 120 KURDILA, PAPASTAVRIDIS, AND KAMAT J. GUIDANCE

and Computer Aided Analysis and Optimization of Mechanical System

Dynamics, edited by E. J. Haug, Springer-Verlag, Berlin, 1984, pp.
A = y[Qi(c&) ~ l(I- &(<?&) - lc)Q2] (64)555-602.
8
Gluck, R., "Hope for Simulating Flexible Spacecraft," Aerospace
America, Vol. 24, Nov. 1986, 40-44.
where 9
Singh, R. P. and Likins, P. W., "Singular Value Decomposition
for Constrained Dynamical Systems," Journal of Applied Mechanics,
(65) Vol. 52, Dec. 1985, pp. 943-948. Also, Singh, R. P. and VanderVoort,
R. J., "Dynamics of Flexible Bodies in Tree Topology—A Computer-
Oriented Approach," Journal of Guidance, Vol. 8, Oct. 1985.
is the QR decomposition of the matrix c. Of course, as 10
Huston, R. L., "Useful Procedures In Multibody Dynamics,"
mentioned earlier, a practical implementation of Eqs. (63-65) Dynamics of Multibody Systems, edited by G. Bianchi and W. Schieh-
does not require actual formation and calculation of Y and y. len, Springer-Verlag, Berlin, 1986, pp. 69-78.
As another example, it seems that Kim19 is well aware that n
Kane, T. R. and Levinson, D. A., "The Use of Kane's Dynamical
evaluating the bases at each time step Equations in Robotics," The International Journal of Robotics
Research, Vol. 2, No. 3, Fall 1983, 3-21.
I2
Likins, P. W., "Quasicoordinate Equations for Flexible Space-
and ^= (66) craft," AIAA Journal , Vol. 13, April 1975, pp. 524-526.
13
Song, J. O. and Haugh, E. J., "Dynamic Analysis of Planar
Downloaded by UNIVERSITY OF OKLAHOMA on February 3, 2015 | http://arc.aiaa.org | DOI: 10.2514/3.20524

Flexible Mechanisms," Computer Methods in Applied Mechanics and

are perfectly reasonable choices of dual bases, although per- Engineering, Vol. 24, 1980, pp. 359-381.
haps not as computationally efficient as null-space updating. 14
Boland, P., Samin, J. C., and Willems, P. Y., "Stability Analysis
The corresponding bases for a Kane-based algorithm can then of Interconnected Deformable Bodies in a Topological Tree," AIAA
be extracted from Journal, Vol. 12, Aug. 1974, pp. 1025-1030.
15
Wittenburg, J., Dynamics of Systems of Rigid Bodies, B. G.
l Teubner, Stuttgart, 1977.
« = r^~|r and A=y[Ql(cQl)- Q2] (67) 16
Schwertassek, R. and Roberson, R. E., "A Perspective on Com-
puter-Oriented Multi-body Dynamical Formalisms and Their Imple-
mentations," Dynamics of Multibody Systems, edited by G. Bianchi
Conclusions and W. Schiehlen, Springer-Verlag, Berlin, 1986, pp. 261-274.
I7
The difficulty in systematically accommodating a wide vari- Wehage, R. A. and Haug, E. J., "Generalized Coordinate Parti-
ety of constraints in simulations of system dynamics has given tioning for Dimension Reduction in Analysis of Constrained Dy-
rise to a variety of numerical methods that eliminate the namic Systems," Journal of Mechanical Design, Vol. 104, Jan. 1982,
pp. 247-582.
constraints automatically. A thorough review of these meth- 18
Mani, N. K., "Application of Singular Value Decomposition for
ods has been presented in this paper. By using Maggi's Analysis of Mechanical System Dynamics," American Society of
equations and a projective interpretation from numerical Mechanical Engineers, New York, Vol. 84-DET-89, pp. 1-6.
analysis, the equivalence of these methods has been proven. In 19
Kim, S. S. and Vanderploeg, M. J., "QR Decomposition for State
fact, it is shown that each method differs only in its selection Space Representation of Constrained Mechanical Dynamic Systems,"
of a basis for the null-space of the variational constraints and Journal of Mechanisms, Transmissions, and Automation in Design, Vol.
its orthogonal complement. For each method considered, the 108, June 1986, pp. 183-188.
20
appropriate bases have been derived explicitly, and provide a Wang, J. T. and Huston, R. L., "Kane's Equations with Undeter-
convenient means of evaluating and comparing the individual mined Multipliers—Application to Constrained Multibody Systems,"
Transactions of the ASME, Vol. 54, June 1987, pp. 424^429.
methods. 21
Neimark, J. I. and Fufaev, N. A., Dynamics on Nonholonomic
The theory presented also suggests generalizations to some Systems, American Mathematical Society, Providence, RI, 1972.
of the methods under consideration. It has been shown that 22
Golub, G. H. and Loan, C. F., Matrix Computations, The Johns
those methods that currently do not calculate constraint Hopkins University Press, Baltimore, MD, 1983.
forces may do so—often with little additional work—by 23
Papastavridis, J. G., On the Foundations of Analytical Dynamics,
utilizing the projection onto the orthogonal complement of with Special Emphasis on Nonholonomic Systems, School of Mech-
the null-space of the variational constraints. anical Engineering, Georgia Inst. of Technology, Class Notes, Winter
1987.
24
Kane, T. R. and Levinson, D. A., "The Use of Kane's Dynamical
References Equations in Robotics," The International Journal of Robotics Re-
^infrey, R. C., "Elastic Link Mechanism Dynamics," Transac- search, Vol. 2, No. 3, 1983, pp. 3-21.
25
tions of the ASME, Feb. 1971, pp. 268-272. Kane, T. R. and Levinson, D. A., Dynamics: Theory and Applica-
2
Erdman, A. G., Sandor, G. N., and Oakberg, R. G., "A General tions, McGraw-Hill, New York, 1985.
26
Method for Kineto-Elastodynamic Analysis and Synthesis of Mecha- Desloge, E. A., "A Comparison of Kane's Equations of Motion
nisms," Journal of Engineering for Industry, Nov. 1972, pp. 1193- and the Gibbs-Appell Equations of Motion," American Journal of
1205. Physics, Vol. 54, No. 5, May 1986, pp. 470-471.
3 27
Haug, E. J., "Elements and Methods of Computational Dynam- Shabana, A. A. and Wehage, R. A., "A Coordinate Reduction
ics," Computer Aided Analysis and Optimization of Mechanical System Technique for Dynamic Analysis of Spatial Substructures with Large
Dynamics, edited by E. J. Haug, Springer-Verlag, Berlin, 1984, pp. Angular Rotations," Journal of Structural Mechanics, Vol. 11, March
3-37. 1983, pp. 401-431.
4 28
Hooker, W. W., "Equations of Motion for Interconnected Rigid Schiehlen, W. O., "Dynamics of Complex Multibody Systems,"
and Elastic Bodies: A Derivation Independent of Angular Momen- SM Archives 9, 1984, pp. 159-195.
29
tum," Celestial Mechanics, Vol. 11, 1975, pp. 337-359. Kirgetov, "The Motion of Controlled Mechanical Systems with
5
Ho, J. Y. L., "Direct Path Method for Flexible Multibody Space- Prescribed Constraints (Servoconstraints)," Journal of Applied Math-
craft Dynamics," Journal of Spacecraft, Vol. 14, Feb. 1977, pp. ematics and Mechanics, Vol. 31, 1967, pp. 433-446.
30
102-110. Papastavridis, J. A., " On the Nonlinear Appell's Equations and
6
Bodley, C., Devers, A., Park, A., and Frisch, H., "A Digital the Determination of Generalized Reaction Forces," International
Computer Program for Dynamic Interaction and Simulation of Con- Journal of Engineering Science (to be published).
31
trols and Structures (DISCOS)," NASA TP 1219, May 1978. Wampler, C., Buffington, K., and Shu-hui, J., "Formulation of
7
Haug, E. J., Mani, N. K., and Krishnaswami, P., "Design Senstiv- Equations of Motion for Systems Subject to Constraints," Journal of
itiy Analysis and Optimization of Dynamically Driven Systems," Applied Mechanics, Vol. 52, June 1985, pp. 465^70.