Multibody System Dynamics 1: 3–21, 1997.

3
c 1997 Kluwer Academic Publishers. Printed in the Netherlands.

A Geometric Unification of Constrained

System Dynamics

WOJCIECH BLAJER
Department of Mechanics, Technical University of Radom, ul. Malczewskiego 29,
PL-26-600 Radom, Poland

(Received: 2 September 1996; accepted in revised form: 8 January 1997)

Abstract. A unified geometric formulation of the methods used for solving constrained system
problems is given. Both holonomic and nonholonomic systems are treated in like manner, and the
dynamic equations are expressible in either generalized velocities or quasi-velocities. Moreover, a
wide range of unconstrained systems are uniformly regarded as generalized particles in the multi-
dimensional metric spaces relating to their configuration. The derivation is grounded on the tensor
calculus formalism and appropriate geometric interpretations are reported. In its useful matrix form,
the formulation turns out short, elementary and general. This unified geometric approach to con-
strained system dynamics may deserve to become a generally accepted method in academic and
engineering applications.
Key words: analytical dynamics, nonholonomic systems, differential geometry.

1. Introduction
During the last two decades a renewed interest has been observed in the theo-
ry of constrained mechanical systems, stemmed from the rapid and simultaneous
development of many technological disciplines (robotics, spacecraft, machine and
vehicle dynamics, biomechanics, etc.), and stimulated by advances in computa-
tional techniques. These two factors – the needs for analysing of more and more
complex machineries and the powerful investigation tool offered by numerical
methods – resulted in a new “philosophy” of formulating and solving problems
of analytical mechanics. Many historical approaches consequent to Lagrangian
mechanics turned out to be laborious for large-scale systems and difficult to algo-
rithmize in computer codes. This was mainly due to the arduousness of generating
and differentiating the kinetic energy functions and the entanglements arising when
nonco-ordinate velocity components (quasi-velocities) are involved. Sheer impen-
etrability of many of the classical methods, their usual derivation for systems of
particles, and their strict assignment for either holonomic (H) or nonholonomic
(NH) systems might have also led to some confusion. These reasons have caused
intensive investigations aimed at more general, effective and suited for digital
implementation approaches for handling constrained motion.
Modern methods for the dynamic analysis of constrained multibody systems
fall into two main categories: the differential-algebraic equation (DAE) and ordi-
4 W. BLAJER

nary differential equation (ODE) formulations. The former employ a maximal set of
variables to describe the constrained motion – the co-ordinates and velocities of the
unconstrained system and the multipliers that model the constraining forces. The
motion equations that follow are DAEs, often referred to as Lagrange’s equations
of the first kind. Many researches have preferred the DAE formulation due to its
simplicity and ease of manipulation [1–4]. The approach is commonly recognized
as computationally inefficient, however, and special procedures must be applied
to avoid the constraint violation phenomenon. Another approach is thus, prior to
numerical integration, to reduce the motion equations to a smaller (possibly mini-
mal) set of ODEs. For open-loop (tree structure) systems a popular method of this
type is the joint co-ordinate method; see e.g. [2, 4, 5] (for systems with closed
loops, the method leads to reduced-dimension DAEs). A variety of other codes,
applicable for systems of any structure, exploit the concept of an orthogonal com-
plement matrix to the constraint Jacobian matrix, introduced in [6]. Premultiplying
the constraint reaction-induced dynamic equations by the orthogonal complement
results in the governing equations as ODEs, which can then be further reduced in
dimension by introducing a minimal set of (independent) velocity variables. The
co-ordinate partitioning (LU factorization) [3, 7], the natural/point co-ordinates
[8, 9], the zero eigenvalues theorem [10] or singular value decomposition [11, 12],
the QR or Householder decomposition [13, 14], and the Gramm–Schmidt orthog-
onalization [15, 16] are only a few examples of the computer-oriented techniques
using the concepts more or less directly. Useful applications of classical Gibbs–
Appell [17, 18] and Maggi’s [19] equations as well as (pretended to be new) Kane’s
method [17] have also been demonstrated. A good review of many of the mentioned
methods is provided in [20].
The prevailing direct/matrix notation used in constrained (multibody) dynamics
has done some damage to our understanding of the physics of concerned problems,
occasionally leading to inconsistencies in the mathematical modelling. Reconsider-
ing the problems from the geometric point of view may be stimulative in amending
the omissions and providing a uniform approach to constrained systems. Most of
the above cited references as well as [21–23] appeal to the geometry of constrained
motion as a direct generalization of the projection methods used in particle dynam-
ics. However, there are only few attempts, e.g. [24–27], that set the multibody
dynamics formulation precisely according to the differential geometry formalism
– a powerful tool of analysis. This paper is another contribution in this field.
Important advantages of the proposed formulation are its compactness, ver-
satility in applications, and the geometrical insight that it gives. A wide range of
uconstrained systems can be studied, including the systems composed of unbound-
ed particles and rigid bodies, and Lagrangian (internally constrained multi-rigid-
body) systems whose equations of motion in terms of independent state variables
are known. The constraints on the systems can then be either H or NH, and no
attention can be paid on whether the velocity components are generalized veloc-
ities or quasi-velocities. By using the unified geometric approach, the schemes
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 5

for obtaining several types of constraint reaction-free equations of motion and the
determination of associated constraint reactions are reconsidered. The other impor-
tant contributions are a precise definition of virtual (constraint-admissible) speeds,
and a novel approach to constraint violation stabilization.
The main body of the paper has been written in standard matrix notation which
is the current fashion of a large segment of the multibody dynamics “commu-
nity”. However, most of the fundamental formulae are simultaneously written
in indicial notation of tensor calculus, important for understanding the covari-
ance/contravariance aspects and the physical (geometrical) meaning of matrix
transformations. The present formulation is thus placed somewhere between these
two poles apart (but virtually equivalent) formalisms, combining compactness of
the former and the mathematical pertinence of the latter. For the convenience of the
reader, the pivotal principles of the notation used are summarized in the Appendix.

2. Unconstrained System Dynamics

The methods of analytical mechanics are usually introduced for systems composed
of a finite number of material points under the presence of ideal constraints. Such
a procedure excludes an important segment of finite-dimensional systems with an
infinite number of particles, like rigid bodies and multi-rigid-body systems with
kinematic tree structure, unless some limiting procedures are carried out. From the
point of view of differential geometry, however, there are no obstacles to regarding
all these systems as primitive concepts like mass points and treating them as such.
In order to meet the broader sense of unconstrained system, consider an n-
degree-of-freedom autonomous system characterized by n generalized co-ordinates
p =[ ]
p1 : : : pn T . The governing equations of the system can be written in the
following general matrix form:
_ = A(p)v
p (1)

M(p)_v = h (p; v; t); (2)

where A is an n  n invertible matrix of transformation between the n velocity

components v =[ ]
 1 : : :  n T (possibly quasi-velocities) and the generalized veloc-
_
ities p, M is the n  n symmetric positive definite mass matrix, h =[ ]
h1 : : : hn T
represents the sum of applied forces, gyroscopic terms, centrifugal forces and
Coriolis effects, t is the time, and the superscript ( ) is introduced after [28] to
distinguish between the associated vector representations and reciprocal bases in
matrix notation (see Appendix for details). The said unconstrained system can be
either:
_=
 A collection of unconstrained particles. In this case p v; M is a constant
diagonal matrix, and h is the representation of only the applied forces.
 A collection of unconstrained rigid bodies. Equation (1) relates the time deriv-
atives of translational and rotational co-ordinates of particular bodies and the
6 W. BLAJER

components of their linear and angular velocity (the latter, when expressed in
a body-fixed reference frame, are quasi-velocities). Typically, (2) is an aggre-
gation of Newton–Euler dynamic equations for individual bodies, and M is
constant. A and M are block-diagonal matrices.
 A Lagrangian system – an internally constrained autonomous holonomic sys-
tem whose equations of motion M q q ( ) = ( _ )
h q; q; t have been derived in
_
independent co-ordinates q (q  p, q  v). Equations of this type arise
among others in the first step modelling of closed-loop systems, after cutting
the closed loops and applying the joint co-ordinate method to the produced
open-loop system. The constraints on the systems are then the closing condi-
tions, see e.g. [2, 4, 5, 7, 20].
 A combination of systems as above.
The motion of a system described in (1) and (2) can be treated as the motion
of point p =[ ]
p1 : : : pn T representing the system in its configuration space. The
kinetic energy

T = 1 T
2
v Mv = _
1 T T
2
p A MA 1 p or _
T = 1
2
M ij  =2 k l
_ _
i j 1 B i B j Mij pk pl ; (3)

=
where B A 1 is a positive definite quadratic form on the tangent space to the
configuration space at point p. So endowed with the metric at each point p, the
configuration space becomes a Riemannian manifold N. The space En tangent to N
at p is a local n-dimensional Euclidean (linear vector) space [25]. The mass matrix
M is then a metric tensor matrix of a covariant basis ev =[ ]
~ev1 : : :~evn T of En ,
= =
M ev ev (Mij ~evi  ~evi ), in which the velocity (contravariant) components of
T
= =
~ are expressed, ~ eTv v (~  i~evi ). The vector ~ has thus a unified meaning
of the system velocity and momentum. Namely, by introducing a reciprocal basis
to ev , the contravariant basis ev =[ ] = =
~ev1 : : :~evn T M 1ev (~evi Gij~evj ), where
=
G M , ~ can be variantly written as
1

~ = eTv v = evT v or ~ =  i~evi = j~evj ; (4)

where v = [1 : : : n ]T = Mv (i = Mij  j ) is the momentum (covariant) repre-

sentation of ~ (see also Appendix). To realize the dual meaning of vectors in metric

_= _= _ _=_
spaces is of paramount importance for understanding the further analysis.
Following the interpretation, the vector ~ eTv v evT v (~  i~evi j~evj ) =_
_
represents both the system acceleration components v in ev and the system effective
_
force components v in ev , v _ = _ _ = _
Mv (i Mij  j ). The dynamic equation (2),
which in indicial notation reads as Mij  j _ =
hi, is then the representation of the
following vector formula
~h ~_ = evT (h _) = 0
Mv or ~h ~_ = (hi Mij _ j )~evi = 0 (5)
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 7

which directly appeals to D’Alembert’s form of Newton’s formula F ~ m~a 0.=

The use of base vectors can also be seen as a substitute for the virtual displacements
formalism.
The kinematic equation (1) (in indicial notation pi _ =
Aij  j ) denotes an inter-
dependence between the system velocity components expressed in ev and ep =
[ ]
~ep1 : : :~epn T bases of En . The transformation formula between these two bases is
ev = ( =
AT ep ~evj ) _= _= = =
Aij~epi , and we can write ~p eTp p eTp Av eTv v ~ . The
_=
equivalence ~p ~ expressed in ep leads to (1), while represented in ev is v Bp = _
= _ =
( i Bji pj ), where B A 1 .

3. Constraints on the System

Let the system as introduced in Section 2 be subjected to m independent ideal
constraints, mH H (position) and mNH NH (nonintegrable velocity) ones, mH +
=
mNH m. Assuming the NH constraints are linear in velocities, the constraint
equations are:

H(p; t) = 0 (6a)

NH(p; v; t)  CNH(p; t)v NH(p; t) = 0; (6b)

where H = [  ]
1 : : : mH T , = [ ]
1 : : : mNH T , C
NH is the mNH  n
=[ ]
1 : : :  mNH T vanishes for scleronomic con-
NH
NH constraint matrix, and  NH
H H NH NH
NH NH
straints. By differentiating with respect to time the H constraints, (6) can first be
unified to the velocity form = [_
HT TNH T , and then, by one more differen-
]
tiation of the obtained velocity constraints can be transformed to the acceleration
form _ = [ _ ]
TH TNH T , i.e.:
   
 C(p; t)v (p; t)  ( @  H =@ p)A @  H =@t
CNH
v
NH = 0 (7)

_  C(p; t)_v (p; v; t) = 0; (8)

_ +_
where   Cv  . When handling the constraint equations in the accelera-
tion form (8), the initial values p0 and v0 must satisfy the lower-order constraint
( )=
conditions, H p0 ; t0 (
0 and p0 ; v0 ; t0 )=
0.
Assumed mH > 0, the configuration space of the constrained system, denoted
=
KH (kH n mH ), is restricted to the H constraint manifold (6a) embedded in N
= =
(if there are no H constraints on the system, mH 0 and KH N). The first mH
constraint vectors, represented in C as rows, are then the H constraint gradients,
and as such are orthogonal to KH . The other constraint vectors, corresponding to
NH constraints, may have arbitrary directions with respect to KH . Rewriting (7) as
8 W. BLAJER

j = Cij  j = 0 for j = 1; : : : ; m and i = 1; : : : ; n, it can be deduced that the j th

constraint vector is ~c j = Cij eiv = Cij Gik~evk , which in matrix notation reads

ec = [~c : : :~c m ]T = Cev = CM 1 ev : (9)

If rank(C) = m = max, the constraint vectors ~c j (j = 1; : : : ; m), are independent

in En , and as such define an m-dimensional subspace of En , called the constrained
subspace Cm . The metric tensor matrix of (contravariant) basis ec of Cm is Mc 1 =
e ec T = CM 1 CT (Gij c = ~c ~c = Ck~ev ~ev Cl = Ck G Cl ), where Gc = Mc
i j i k l j i kl j 1

and G = M . 1

For ideal constraints, the reactions of individual constraints are collinear to the
corresponding constraint vectors, ~rJ = J~c J (do not sum for J ). The generalized
reaction force on the system (the total of the individual constraint reactions) can
then be written as
m
X m
X
~r = ~rJ = J~c J = ecT  = evT (CT  ) = evT r ; (10)
J =1 J =1
where r = [r1 : : : rn]T =
CT  and  =[ ]
1 : : : m T (Lagrange multipliers)
are the representations of ~r in the bases ev and ec , respectively. Note that ~r is totally

sunk in Cm . The governing equations of the constrained system can then be written
in the following form:
_ = A(p)v or p_i = aij  j
p (11a)

M(p)_v = h (p; v; t) + CT (p; t) or Mij _ j = hi + Cik k (11b)

C(p; t)_v =  (p; v; t) or Cik _ i =  k (11c)

which are often referred to as Lagrange’s equations of the first kind [2, 4, 25]. With
reference to (5), the geometric interpretation of (11b) is
~h + ~r ~_ = evT (h + CT  Mv_ ) = 0 (12)

or in indicial notation ~h + ~r ~_ = (hi + Cik k Mij _ j )~evi = 0, which directly

appeals to the geometry of simple dynamics problems (see Figure 1b). Following
the geometric interpretation, the constraint velocity equation (7) can be viewed as
= = = = = =
 ecT ~ Cev eTv v Cv (i ~c i  ~ CJi ~evj  ~evk  k Cji  j ). The term 
is thus the projection of the system velocity ~ into Cm , expressed in the basis ec ,
= =
~ eTc  (~ i~ci ). As such,  represents only that part of “constraint velocity” at
p which is contained in Cm , and the eventual tangent component of the constraint
velocity is not “seen” in the explicit constraint equations (7) (see Figure 1a).
_ +_
Seemingly, the constituents of   Cv  in (8) and (11c) denote: Cv – the _
_
acceleration due to the constraint “curvature” at p, and  – the acceleration due
to the constraint motion at p (again only those parts which are contained in Cm ).
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 9

Figure 1. The geometry of a constrained system.

= _= _= _ = _=
Namely, one can write  vc T ~ Cev eTv v Cv ( i ~c i  ~ Cji~evj  ~evk  k_ =
Cji kj  k Cji  j ) and ~ eTc  (~  i~ci ). Note that  and  are representations
_ = _ = =
of ~ and ~ in the basis ec , reciprocal to ec in which the constraint reactions are
represented by  .
The DAE formulation (11) is valid for both H and NH systems, and the com-
ponents of v can be either generalized velocities, quasi-velocities, or both. The
equations can be solved directly by using a range of DAE solvers [29], or the
Lagrange multiplier elimination methods can be used. A popular technique of
latter type is to rewrite (11) as
p_ = Av (13a)
" #
M CT _  =  h 
v
C 0   (13b)

and, since the coefficient matrix in (13b) is by assumption invertible, solve the
() ()
equations as implicit ODEs in v t and p t , and simultaneously determine  t ()
in terms of the current state values v and p.
The analysis of the constrained motion evolution by using the above govern-
ing equations is commonly evaluated as computationally inefficient, mainly due
to possible large dimension of the sets of equations to handle. Moreover, when
handling the constraint equations in the acceleration form, the solution may suffer
from the problem of violation of the lower-order constraint equations, and special
procedures (e.g. Baumgarte’s method [30]) must be applied to avoid the phenom-
enon. The shortcomings of the dependent variable formulations have stimulated the
development of methods for obtaining equations of motion in terms of a minimum
number of (independent) variables. The first step of the approach is usually the pro-
jection of the dynamic equations into the null space of Cm (the space complemetary
to Cm in En ) [3, 6, 9–16, 21–23, 31].
10 W. BLAJER

4. The Projection Method

As Cm is an m-subspace of the n-space En , a k -dimensional (k n m) subspace =
Dk that complements Cm in En can be defined, Dk [ Cm En and Dn \ Cm 0. = =
Note that Dk is the tangent space to the configuration manifold KH at p 2 KH only
=
for a H system (mNH 0) – the tangent space defined as a complement of the space
spanned by H constraint gradients [25]. For a NH system such that 0 < mNH < m
(a system subject to both H and NH constraints), the tangent space is of dimension
=
kH n mH > k. Finally, in the extreme case of a NH system (a system subject
to only NH constraints, mNH = =
m), we have KH N, and the tangent space is
En . To meet the general case of a system subjected to H and/or NH constraints, Dk
will be called a virtual (velocity admissible) subspace, as it is formed by a set of
unconstrained (admissible by constraints) velocity directions.
By introducing k independent vectors ~d1 ; : : : ; ~dk that are totally sunk in Dk ,
represented in ev basis as columns of an n  n matrix D p; t of maximal column- ( )
rank, the (contravariant) basis of Dk can be defined as
= [~d1; : : : ~dk ]T = DT ev
ed (14)

(~di = Dij~evj , i = 1; : : : ; k , j = 1; : : : ; n). The condition Dk \ Cm = 0, in matrix

notation, is
DT CT =0 , CD = 0; (15)
i.e. D is an orthogonal complement matrix to the constraint matrix C, and (15)
denotes the mutual othogonality of vectors ~di and ~c j (i 1; : : : ; k , j 1; : : : m).
= =
Using (9) and (14), this can be justified by ~di  ~c j Dik~evk  ~evl Clj Dik kl Clj
= = =
= =
Dik Ckj 0, or in matrix form ed ecT DT ev evT CT DT CT 0. As it is well = =
known, the formulation of D for a given C is not unique – different sets of k
independent vectors ~di can span the same subspace Dk .
Being independent, the n vectors ~d1 ; : : : ; ~dk ;~c 1 ; : : : ;~c m form a basis in En .
Using (9) and (14), the transformation formula between the hybrid (covariant-
[ ]
contravariant) basis eTd ec T T and the basis ev can be formulated as
  " #
DT
ed
ec
= CM 1 ev = HT (p; t)ev ; (16)

where H is the nn matrix of transformation. The metric tensor matrix of eTd ec T T [ ]
basis is
" T #  
HT MH = = Md 0
D MD
0 CM
0
1
CT 0 Mc 1
; (17)

where Md p; t ( )=
ed eTd DT MD (Mdij ~di ~dj Dik~evk ~evl Djl Dik Mkl Djl )
= = = =
is the metric tensor matrix of the basis ed of Dk , and Gc p; t Mc 1 ec ec T ( )= = =
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 11

CM 1 CT (Gij
 m
= k v v l =k l =
i j C i~e k  ~e l C j C i Gkl C j ) is the metric tensor matrix
c ~c  ~c
of ec of C . From (19) it comes also (see Appendix) that
 
ev =H ed
ec
= Ded + M 1
CT ec (18)

(~evi = Dji~d j + Gik Ckl ~cl ), where ed = Md 1ed (~d i = Gijd~dj ), ec = Mcec (~ci =
Mcij~c l ).
Substituting (18) into (12), and considering (15), one obtains
(
edT DT h DT Mv _ ) + eTc (CM 1
h + CM 1
CT  _) = 0
Cv (19)
which expresses the projections of the dynamic equation (12) into Dk and Cm ,
respectively. By virtue of the complementarity of the subspaces in En , (19) means
that:
DT Mv _ = DT h (20a)

Cv _ = CM 1
h + CM 1
CT  (20b)

which can also be obtained by premultiplying (11b) by HT . The geometric inter-

pretation of (20) is illustrated in Figure 1b, and its vector representation is:
~hd ~_ d = 0 (21a)
~hc + ~r ~_ c = 0; (21b)

where the subscripts d and c denote the projections of appropriate vectors into Dk
and Cm .
Since the constraint equation (11c) can be interpreted as
~ ~_ c = eTc ( Cv _) = 0 (22)
the constrained subspace projections (20b) and (21b) can be manipulated to:
CM 1 h + CM 1
CT  =0 (23)
~hc + ~r ~ = 0: (24)

Then, the virtual subspace projections (20a) and (21a), supplied with (11c), leads
to the constraint reaction-free governing equations [5, 6, 9–15, 20, 21, 31, 32]:
_ = Av
p (25a)
" T # " T #
D M
C
_=
v
D h
 (25b)
12 W. BLAJER

()
which are 2n ODEs in v and p. Using the solution v t and p t to the ODEs, the ()
()
Lagrange multipliers  t can then be synthesized from the following algebraic
formula

(p; v; t) = (CM 1
CT )( CM 1 h ): (26)

The coefficient matrix on the left-hand side of (25b) is HT M. Then it comes

from (16) and (17) that HT M 1( ) =[
DMd 1 MCT Md and I MDMd 1 DT ] = +
CT Mc CM 1 , where I is the identity matrix. Using this, (25b) can be manipulated
_= +
to the following equivalent form Mv h CT Mc  MM 1 h , where Mc ( ) =
( )
CM 1 CT 1 . The result can also be obtained directly by substituting (26) into
(11b), while (26) is obtained after substituting (11b) into (11c), see e.g. [2].
Compared to DAEs (11) or (13), the dimension of ODEs (25) is reduced from
+
2n m to 2n. The benefit can, however, be fictitious since the n  n coefficient
( + ) ( + )
matrix in (25b) is a general matrix while the n m  n m coefficient matrix
in (13b) is a sparse matrix. As special procedures can be applied to sparse matrices,
the inversion of the two matrices in the process of numerical integration may thus be
equally “expensive”. Moreover, ODEs (25) still suffer from the constraint violation
problem. Therefore, further reduction in dimension by introduction a minimal set
of (independent) velocity components is usually undertaken.

5. The Equations of Motion in Virtual Speeds

Further developments in the description of constrained system dynamics can be
achieved by introducing k independent virtual speeds u =[
u1 : : : uk T being the ]
components in the basis ed of the projection of ~ into Dk , ~u eTd u (~u ui~di ,
= =
=
i 1; : : : ; k). According to the illustration in Figure 1a, we can write
~u + ~ = ~ : (27)

To find a matrix representation of this vector formula, we first write ~ evT Mv, =
and then using (18) we have ~ =( + )
edT DT ec CM 1 Mv edT DT Mv ec Cv = + =
+ = =
~u ~. On the other hand, ~u edT u edT Md u and ~ eTc . Comparing the
 =
two formulations, we arrive at
" T # " T #
D MDu

D Mv
= Cv
 HT Mv: (28)

=
For scleronimic systems ( 0), (28) denotes that ~ is entirely contained in Dk
=
or, more specifically, the system momentum v Mv (the representation of ~ in
ev basis of En ) projects only into Dk and is represented in ed basis of the subspace
as u =Md u. For rheonomic systems ( =
0), ~ is deflected from Dk due to
the constraint “motion” perceived in C (~ is the component of ~ in Cm ), see
m
Figure 1a.
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 13

From (28) it comes directly that

u = (DT MD) 1
( ) ( ) ()
DT Mv  Md 1 p; t DT p; t M p v: (29)
As the components of u may have no physical meaning, (29) can serve to obtain the
consistent initial values u0 for given p0 and v0 , which is of paramount importance
for the succcessful initialization of the integration process of motion equations that
will follow. Then, by inverting (28) we obtain
v = Du + M 1CT (CM 1CT ) 1  D(p; t)u + (p; t); (30)
where = M 1 CMc  denotes the representation of ~ in ev and vanishes from
scleronomic constraints (note that  is the representation of ~ in ec , i.e. ~ = eTv =
eTc  ). The formulae (29) and (30) are the direct/matrix representations of the vector
equations ~ = ~u + ~ and ~u = ~ ~ , illustrated in Figure 1a.
The virtual speeds u are notionally equivalent to the independent kinematic
parameters introduced in Maggi’s [19, 32, 33] and Gibbs–Appell [17, 18, 32]
methods, and to the generalized speeds used in Kane’s method [7, 15, 17, 22].
The present formula states the definition more precisely from the mathematical
(geometrical) point of view. Also note that by using (28) the virtual speeds u need
not be introduced a priori – they may result from the constraint equations (7),
the introduced orthogonal complement matrix D to the constraint matrix C, and
the system inertial properties represented in the metric of configuration space
(excluding the latter from the analysis is erroneous; see the forthcoming discussion).
Moreover, none of the previous methods supplies the user with an explicit general
formula, such as (28) and (29), to determine explicitly the independent velocity
components.
Using (30) and its differentiated form v _= _+_ +_
Du Du , it is easy to check
that the velocity and acceleration constraint equations (7) and (8) are satisfied by
identity, and the governing equations (25) transform to the following n k ODEs +
in p and u:
_ = A(Du + )  p_ = Ad(p; t)u + ad(p; t)
p (31a)

DT MDu_ = DT h DT M(Du _ + _ )  Md(p; t)_u = hd(p; u; t) (31b)

Given p0 and v0 , from (29) the initial values u0 = Md 1 (p0 ; t0 )DT (p0 ; t0 )M(p0 )v0 .
Since it comes from (25) that CD_ = CD,_ and the  = Cv _ +  = C(Du _ + ),
the relation (26) can eventually be manipulated to
(p; u; t) = (CM 1
CT( _ + _ M 1h)
) 1
C Du (32)
from which, using the solution p(t) and u(t) to ODEs (40),  can be synthesized.
ODEs (31) provide a general procedure for obtaining the minimal-dimension
equations of motion of a constrained system. The formulation requires first to
determine an orthogonal complement matrix D to the constraint matrix C. For
14 W. BLAJER

small systems this can often be done by inspection or simply guessed. For large-
scale systems the determination is usually performed numerically, and numerous
computer-oriented codes of this type have been proposed (see e.g. [3, 4, 6, 7,
_ +_
10–16, 20]). A more complex task is to determine Du used in the dynamic
equation (31). In the co-ordinate partitioning method [3] the term is found directly
from the partitioned acceleration form of constraint equations (8). Namely, from the
= + + + =
partitioned velocity form of contraint equations, Cv   Uu Ww  0,
where C =[ ] =[
U W ,v ]
uT wT T , and the m dependent velocities w are chosen so
( )=
that det W 6 0, (30) is obtained as v =[ (
I ) ] +[
W 1 U T T u 0T ( )] W 1 T T 
+
Du , and I and 0 are the k  k and k  1 identity and zero matrices, respectively.
_ = _+
Then, from the partitioned constraint acceleration equations, Cv  
_+ _+ =
Uu Ww  0, it follows that v _ =[ (
I W U 1 T_
) ] +[ (
T u 0 )]
T W 1 T T 
_+_
Du ; see also [7, 34]. In this way, the dynamic equations (31b) in the chosen
k independent velocities u can be obtained. A recursive scheme for obtaining D _
_
numerically, based on C and C, have also been proposed in [15] and then followed
in [35]. Finally, in [26] an ingenious differentiation-free scheme (based on pure
differential geometry formalism) for obtaining the dynamic equations in terms of
independent variables is proposed. General and computationally cheap new codes
_ +_
for computing Du are still desirable, however,
The relations (27–30) and the consequent motion equations (31) have been
derived based on the explicit formulations (7) and (8) of constraint equations. For
H systems, however, the H constraints are commonly introduced implicitly by a
priori choice of independent co-ordinates q =[ ] =
q1 : : : qk T (mH m, k n m)=
that define the system position on the H constraint manifold (the configuration
=
space of the H constrained system, KH K. The dependence

p = g(q; t) (33)

[ ( ); t] 
stands for the m H constraints on the n dependent co-ordinates p, i.e. H g p; t
0. The differentiated form of (33), after considering (1), is

v = D0(q; t)_q + 0(q; t); (34)

where D0 = A 1 (@ g=@ q) and 0 = A 1 (@ g=@t), stands for the m constraint

conditions on the dependent velocities v. Finally, using (33), (34) and v_ = D0 q +
_ 0q_ + _ 0, the governing equations (31) transform to the following 2k ODEs in q
D
and q_ [2, 4, 5, 8, 9, 20, 26]

D0T MD0 q  = D0T h D0T M(D_ 0q_ + _ 0)  M0d(q; t)q = hd (q; q_ ; t): (35)
For nonautonomous systems, the virtual speeds u (quasi-velocities) defined in
_
(37) and the generalized velocities q as above are the representations of different
(but collinear) vectors in En (see Figure 2), and the relations (30) and (34), though
= +
similar, are not equivalent. Namely, (30) expresses ~ ~u ~ , while (34) stands
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 15

Figure 2. The virtual speeds u (quasi-velocities) and the generalized velocities q.

_

=_+
for ~ ~q ~ 0 , where ~0 eTv 0 is the total effect of the constraint “motion” at
=
q 2 KH while ~ is the projection of ~ 0 in Cm . From inverting (34) we obtain

q_ = (D0MD0) 1
(
D0T M v 0 ): (36)
_
Evidently, for autonomous systems, u and q are the representations of the same
vector, contained in the subspace Dk of En , ~ ~ 0 0 and ~q ~u ~ .
= = _= =
6. Elimination of Constraint Violations
The problem of constraint violation occurs when the motion equations used for
simulation involve the differential forms of constraint equations. By assumption, the
exact realization of only those differentiated constraint equations is assured while
the lower-order (original) constraint equations may be violated by the solution
(burdened with a numerical error of integration) even though the initial values
of the state variables satisfy the constraint equations. The case relates mainly to
DAEs (11–13) and ODEs (25), where the constraint equations are considered in the
^( ) ^( )
acceleration form (16). The numerical solutions p t and v t to these equations
may violate the H constraint equation (6a) and its differentiated form as well as
(^ ) = (^ ^ ) =
the NH constraint equation (6b), H p; t 6 0 and p; v; t 6 0. The situation
is a little different when dealing with ODEs (31). By virtue of the definition of
independent speeds u, the used relation (30) stands for the m velocity constraint
equations (7), = [_
TH TNH 0, which are by assumption satisfied by the
]=
^( ) ^( )
solution p t and u t . The formulation does not, however, protect the H constraint
^( ) (^ ) =
equation (6a) from being violated by the solution p t , H p; t 6 0, see also [15].
Finally, ODEs (35), derived for H systems, are released from the constraint violation
problem – the used relations (33) and (34) stand for H =
0 and H  _ 0, =
and assure the exact realization of the constraint equations.
The conceptually simplest method for eliminating the H constraint violation
or, in other words, for correcting the system position so that H =
0 with a
^( )
required numerical accuracy, is to treat p t as a trial root of H p; t ( )=
0 for a
16 W. BLAJER

given t, and then to solve the constraint equations for a numerically exact root. As
( )=
dim H ( )=
mH < dim p n, n mH supplementary equations must be added
( )=
to H p; t 0 in order to complete the process successfully, and it is essential
to construct the supplementary equations so that to ensure the position correction
only in the subspace defined by the mH H constraint gradients (which is a part
of Cm for a system subject to H and NH constraints). Such a formulation can be
_ + + ^_ =
deduced from (28) applying CH A 1 p  H H 0 instead of the velocity form
of constraint equations. Then, the Newton–Raphson formula for transforming the
^ = (^ )
H constraint violation H H p; t to the position correction p p t
= ( ) ^( )
pt
is as follows
2 T 3 " #
^ ^
4 DH M 5 A^
^
1

p = ^ H ;
0
(37)
CH
where DH is an n  kH orthogonal complement matrix to the mH  n H con-
=(
straint matrix CH ) =
@ H=@ p A, kH n mH, and the superscript ( ) denotes ^
^
dependence on p. By inverting (37) we obtain a much more useful formula

p = A^ M^ 1C^ HM 1
CTH ) 1 ^ H (38)
released from the necessity of determining DH . The elimination of violation of
^ = (^ ^ )
velocity constraint equations, p; v; t , can be done by using formulae very
similar to (37) and (38). Namely, denoting v v t
= ( ) ^( )
v t , it can be deduced
from (35b) that
" T # " #
^ M^
v = 0
^
D
^
C
(39)

or in a resolved form

v = M^ 1 C^ T (C^ M 1
CT ) 1 ^ : (40)
The formulae (38) and (40) assure that the position and velocity correction are
^ ^
performed in Cm (the system position and velocity in Dk are not changed). The
geometric interpretation of the correcting terms (for scleronomic constraints) is
given in Figure 3.
The position and velocity corrections according to (38) and (40) can be applied
after each step of integration or after a sequence of steps when the constraint
violations surpass the accepted values. The formulae can also serve to modify
Baumgarte’s constraint stabilization method [30]. From that angle, ODEs (31)
should be changed to:
 Z 
_ = A(Du + )
p M CTH (CH M 1 CTH )
1 1
K1 H + K0 H dt (41a)

DT MDu _ = DT H DT M Du ( _ + _ ); (41b)
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 17

Figure 3. Position and velocity corrections.

while ODEs (25) will take the form:

 Z 
_ = Av
p (
AM 1 CTH CH M 1
CTH ) 1
K1 H + K0 H dt (42a)
" T # " #
DT h R
D M
C
_=  (
v
dt) ;
L1 L0 + (42b)

where K1 and K0 , and L1 and L0 are the mH  mH and m  m diagonal matrices

of appropriate gain values [30]. Seemingly one can use the stabilized kinematic
equation (42a) Rinstead of (13a) and substitute the stabilized constraint enforcement
( + )
 L1 L0 dt instead of  in (13b) to transform DAEs (13) into a stabilized
form.
The present scheme of separate position and velocity corrections for constraint
violation stabilization differs from that proposed by Baumgarte [30]. Here the
position correcting terms are placed in the kinematic equations and the velocity
correcting terms are situated in the dynamic equations. The position constraint
violation are thus directly converted into the appropriate position corrections while
the velocity corrections result from only the velocity constraint violation. In Baum-
garte’s method both the correcting terms are placed in the dynamic equations, and
the position constraint violation is controlled indirectly by inducing an additional
velocity correction – the effect in position is thus “delayed” in time.
The experience of the author is that the present method is very effective, accu-
rate and computationally cheap. It is preferred to use the schemes (38) and (40)
after each step of integration (or a sequence of steps), rather than to apply the
stabilized governing equations (41) and (42). It is also important not to allow large
^
constraint violations – the correction is performed in Cm defined for p 2 KH ^ ^
^
(see Figure 3), and the closer KH to KH the better accuracy. For larger constraint
violation the formulae (38) and (40) be used recursively. In order to justify the
effectiveness/accuracy of the method, it should be tested on examples and com-
pared against other approaches. This will be done in future works. The objective of
this presentation was only to show the geometric involvement of the problem and
the impact it gives on the dynamic formulations.
18 W. BLAJER

7. Discussion and Conclusion

The proposed geometric formulation of constrained system dynamics appeals to
intuition as a generalization of methods used in simple dynamics problems (New-
tonian mechanics). The mathematical language used is that of vectors and tensors
in metric spaces, a powerful and precise tool of analysis. In particular, the velocity
(acceleration) and momentum (effective force) components of a system are treated
as associated representations of the same vector. This implies a special meaning
of the metric of system configuration space, involving the inertial properties of the
system. Neglecting the inertial attributes may occasionally lead to physical incon-
sistencies in mathematical formulation. In many contributions, for instance, the
= = + +
velocity “norm” is introduced as k~ k2 vT v 12


n2 . This can eventually
be accepted for systems of particles, due to the very specific (diagonal) metric
tensor matrix of the configuration space. For a system composed of rigid bodies v
usually consists of linear and angular velocities, however, and by using the above
“norm” we sum addends characterized by different units (what we tell students
never to do!). After setting the norm appropriately, k~ k2 = vT Mv, the inconsis-
tencies are clarified. The other example can be the matrix CT C introduced in the
zero-eigenvalue theorem [10], the singular value decomposition [11, 12], and the
QR/Householder decomposition [13, 14] methods. The entries of the matrix may
also be formed by summing addends of different units. From the geometrical point
( )
of view the matrix should rather be changed to CT CM 1 CT 1 C. The remarks
do not, however, condemn the computer-oriented codes involving the physically
inconsistent formulations – most of them are powerful tools of computer-aided
analysis of multibody systems, and playing with numbers the units are not “seen”.
Nevertheless, one can never entirely shake off the feeling that something is missed
in the formulations.
The proposed projective/geometric approach to solving problems of constrained
system dynamics can be recommended for many reasons. It offers a comparatively
simple, general and effective algorithm for obtaining constraint reaction-free equa-
tions of motion and, based on the solution to these equations, for synthesizing the
associated constraint reactions. The method applies for both H and NH systems
and no attention can be paid to distinguishing between generalized velocities and
quasi-velocities (the entanglements arising when introducing quasi-velocities are
the “Achilles heel” of Lagrangian mechanics [32]). The recent paper [36] presents
several examples of application of the method. Finally, the notion of unconstrained
system is extended to any autonomous Lagrangian (holonomic) system whose equa-
tions of motion are given in the form (1) and (2). All the systems can be regarded
as primitive concepts like particles and treated as such (in a unified way). There
are also no obstacles to analyse rheonomic and/or nonholonomic “unconstrained”
systems, such as those described in (35) and (31), respectively, on which additional
constraints are imposed. The configuration spaces of such systems become pseudo-
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 19

Figure A1. Geometrical illustration of (A1–A3).

Riemannian, however, and the geometric interpretation as well as the mathematical

description become more complex.

Appendix
In an n-dimensional vector space En , a vector ~a can be represented either by
its contravariant components a =[ ]
a1 : : : an T with respect to a covariant basis
e=[ ] =[ ]
~e1 : : :~en of E , or by covariant components a a1 : : : an T with respect
T n
to the contravariant basis e =[ ]
~e 1 : : :~e n T [24, 28, 37]:
~a = eT a  a1~e1 +


+ an~en
= eT a  a1~e 1 +


+ an~e n (A1)
or using indicial notation ~a = ai~ei = aj~e j . Using the metric tensor matrix M of
the basis e,
2 ~e  ~e : : : ~e  ~e 3
1 1 1 n
6
M = eeT = 4 .. .. 75 (A2)
. .
~en  ~e1 : : : ~en  ~en
(Mij = ~ei  ~ej ), the relationships between the associated vector representations
and the reciprocal bases are:
a= Ma (ai = Mij aj ); a = M 1a (ai = Gij aj );
e = Me (~ei = Mij~e j ); e = M 1 e (~e i = Gij~ej ); (A3)

where G = M 1 = e eT (Gij = ~e i  ~e j , Gij Mjk = ki ). A simple illustration of

the formalae (A1–A3) is shown in Figure A1.
A dot product of two vectors ~a and ~b in En can be written in four different ways:
~a  ~b = aT Mb = aT b = aT b = aT M 1
b (A4)
20 W. BLAJER

(~a  ~b ai Mij bj
= = = =
aibi aj bj aj Gjk bk ), and the orthogonalitypcondition of
the two vectors is ~a  ~b 0. Then, the “length” of a vector ~a is a
= =
~a  ~a.
The change from the basis e (or e ) to a new basis e0 (or e0 ) of En is associated
with the following transformation formulae:

a = Aa0 (ai = Aij a0j ); a0 = AT a (a0i = Aji aj );

e = Ae0 (~e i = Aij~e 0j ); e0 = AT e (~ei0 = Aji~ej ); (A5)

where A is the (inevitable) transformation matrix. The metric tensor matrix of e0 is

M0 = e0e0T = AT MA = (e0e0T ) 1
(Mij0 = ~ej0  ~ej0 = Aki MklAlj ): (A6)

References
1. Orlandea, N., Chace, M.A. and Calahan, D.A., ‘A sparsity-oriented approach to the dynamic
analysis and design of mechanical systems, Parts I and II’, Journal of Engineering for Industry
99, 1977, 773–784.
2. Wittenburg, J., Dynamics of Systems of Rigid Bodies, Teubner, Stuttgart, 1977.
3. Wehage, R.A. and Haug, E.J., ‘Generalized co-ordinate partitioning for dimension reduction in
analysis of constrained dynamic systems’, Journal of Mechanical Design 104, 1982, 247–255.
4. Nikravesh, P.E., Computer-Aided Analysis of Mechanical Systems, Prentice Hall, Englewood
Cliffs, 1988.
5. Kim, S.S. and Vanderploeg, M.J., ‘A general and efficient method for dynamic analysis of
mechanical systems using velocity transformations’, Journal of Mechanisms Transmissions, and
Automation in Design 108, 1986, 177–182.
6. Hemami, H. and Weimer, F.C., ‘Modelling of nonholonomic dynamic systems with applications’,
Journal of Applied Mechanics 48, 1981, 177–182.
7. Wampler, C., Buffinton, K. and Shu-hui, J., ‘Formulation of equations of motion for systems
subject to constraints’, Journal of Applied Mechanics 52, 1985, 465–470.
8. Garcı́a de Jalón, J., Unda, J. and Avello, A., ‘Natural co-ordinates for the computer analysis
of multibody systems’, Computer Methods in Applied Mechanics and Engineering 56, 1986,
309–327.
9. Garcı́a de Jalón, J., Unda, J., Avello, A. and Jiménez, J.M., ‘Dynamic analysis of three-
dimensional mechanisms in “natural” co-ordinates’, Journal of Mechanisms, Transmissions,
and Automation in Design 109, 1987, 460–465.
10. Walton, Jr., W.C. and Steeves, E.C., ‘A new matrix theorem and its application for establishing
independent co-ordinates for complex dynamical systems with constraints’, NASA Technical
Report TR R-326, 1969.
11. Mani, N.K., Haug, E.J. and Atkinson, K.E., ‘Application of singular value decomposition for
analysis of mechanical system dynamics’, Journal of Mechanisms, Transmissions, and Automa-
tion in Design 107, 1985, 82–87.
12. Singh, C.L. and Likins, P.W., ‘Singular value decomposition for constrained dynamical systems’,
Journal of Applied Mechanics 52, 1985, 943–948.
13. Kim, S.S. and Vanderploeg, M.J., ‘QR decomposition for state space representation of constrained
mechanical dynamic systems’, Journal of Mechanisms, Transmissions, and Automation in Design
108, 1986, 183–188.
14. Amirouche, F.M.L., Jia, T. and Ider, S.K., ‘A recursive Householder transformation for complex
dynamical systems with constraints’, Journal of Applied Mechanics 55, 1988, 729–734.
15. Liang, C.G. and Lance, G.M., ‘A differentiable null space method for constrained dynamic
analysis’, Journal of Mechanisms, Transmissions, and Automation in Design 109, 1987, 405–
411.
A GEOMETRIC UNIFICATION OF CONSTRAINED SYSTEM DYNAMICS 21

16. Agrawal, O.P. and Saigal, S., ‘Dynamic analysis of multi-body systems using tangent co-
ordinates’, Computers & Structures 31, 1989, 349–355.
17. Kane, T.R. and Levinson, D.A., ‘Formulation of equations of motion for complex spacecraft’,
Journal of Guidance and Control 3, 1980, 99–112.
18. Desloge, D.A., ‘The Gibbs–Appell equations of motion’, American Journal of Physics 56, 1988,
841–846.
19. Kurdila, A., Papastavridis, J.G. and Kamat, M.P., ‘Role of Maggi’s equations in computational
methods for constrained multibody systems’, Journal of Guidance, Control, and Dynamics 13,
1990, 113–120.
20. Schiehlen, W. (ed.), Multibody System Handbook, Springer-Verlag, Berlin, 1990.
21. Scott, D., ‘Can a projection method of obtaining equations of motion compete with Lagrange’s
equations?’, American Journal of Physics 56, 1988, 451–456.
22. Lesser, M., ‘A geometrical interpretation of Kane’s equations’, Proceedings of the Royal Society,
London A436, 1992, 69–87.
23. Essén, H., ‘On the geometry of nonholonomic dynamics’, Journal of Applied Mechanics 61,
1994, 689–694.
24. Synge, J.L. and Schild, A., Tensor Calculus, University of Toronto Press, Toronto, 1949.
25. Arnold, V.I., Mathematical Methods of Classical Mechanics, Springer-Verlag, New York, 1989.
26. Maisser, P., ‘Analytische Dynamik von Mehrkörpersystemen’, ZAMM (Zeitschrift für Ange-
wandte Mathematik und Mechanik) 68, 1988, 463–481.
27. Cardin, F. and Zanzotto, G., ‘On constrained mechanical systems: D’Alembert’s and Gauss’
principles’, Journal of Mathematical Physics 30, 1989, 1473–1479.
28. Pobedrya, B.E., Lectures on Tensor Analysis, Moscow University Publishers, Moscow, 1974.
29. Petzold, L.R., ‘Methods and software for differential-algebraic systems’, in Real-Time Integration
Methods for Mechanical System Simulations, E.J. Haug and R.C. Deyo (eds.), NATO ASI Series
F, Vol. 69, Springer-Verlag, Berlin, 1990, 127–140.
30. Baumgarte, J., ‘Stabilization of constraints and integrals of motion in dynamical systems’,
Computer Methods in Applied Mechanics and Engineering 1, 1972, 1–16.
31. Blajer, W., ‘A projection method approach to constrained dynamic analysis’, Journal of Applied
Mechanics 59, 1992, 643–649.
32. Neimark, J.I. and Fufaev, N.A., Dynamics of Nonholonomic Systems [in Russian], Moscow
University Publishers, Moscow, 1967.
33. Blajer, W., ‘Projective formulation of Maggi’s method for nonholonomic systems analysis’,
Journal of Guidance, Control, and Dynamics 15, 1992, 522–525.
34. Blajer, W., Schiehlen, W. and Schirm, W., ‘Projective criterion to the co-ordinate partitioning
method for multibody dynamics’, Archive of Applied Mechanics 64, 1994, 86–98.
35. Blajer, W., ‘An orthonormal tangent space method for constrained multibody systems’, Computer
Methods in Applied Mechanics and Engineering 121, 1995, 45–57.
36. Arczewski, K. and Blajer, W., ‘A unified approach to the modelling of holonomic and nonholo-
nomic mechanical systems’, Mathematical Modelling of Systems 2, 1996, 157–174.
37. De Boer, R., Vector- und Tensorrechnung für Ingenieure, Springer-Verlag, Berlin, 1982.