Вы находитесь на странице: 1из 16

Computational Mechanics (1993) 12, 81-96

Computational Mechanics
9 Springer-Verlag 1993

Sufficient conditions for penalty formulation methods in analytical dynamics


A. J. Kurdila t and Francis J. Narcowich
1 Department of Aerospace Engineering, Texas A&M University, College Station, TX 77843-3141, USA 2 Department of Mathematics, Texas A&M University, College Station, TX 77843-3368, USA

Abstract. The axiomatic derivation of constraints in analytical mechanics as the limiting case of motion in a potential field that grows asymptotically in strength has been known for years. However, with the emergence of the need for robust simulation methods for complex, multi-degree-of-freedom, nonlinear mechanical systems, researchers have shown renewed interest in these "penalty formulations" as practical computational schemes. While much empirical evidence has been collected regarding the efficiency of these methods, relatively few convergence results are available for a wide class of the nonlinear simulation methods. This paper derives sufficient conditions for the convergence of a class of penalty methods by extending the Rubin-Ungar theorem. One advantage of the approach taken in this paper is that considerable simplification of the original Rubin-Ungar derivation is achieved for the convergence of transverse constraint velocities. This paper also emphasizes the importance of maintaining a rank condition on the Jacobian of the constraint matrix. This is of particular importance in that one claimed benefit of certain penalty methods is that they are effective in cases in which the constraint Jacobian loses rank. For the class of penalty methods considered in this paper, if the Jacobian does not meet the specified rank conditions, a diverse collection of spurious, pathological responses can be obtained using this method. In one sense, this type of pathological response is worse than encountering a "configuration-singular" generalized mass matrix and having a simulation diverge; indeed, the regularized solution procedure can proceed along some incorrect trajectory with little to no indication that something is amiss.

Introduction

Classical texts [1, pp. 75-77; 5, pp. 168-171] have shown that an effective means of deriving the equations of analytical mechanics for systems subject to constraints is to consider a sequence of unconstrained systems under the action of an asymptotically increasing potential field. Some research in the field of quantum mechanics has used this axiomatic approach as a means of approximating specific problems. Prototypical of these efforts are those illustrated in [26], in which the constraints are approximately enforced by considering sequences of unconstrained systems subject to potentials of the form

V(x; 2,)
with the property that the potential grows for motions outside the constraint manifold as 2 -4 0o. One disadvantage of this approach is illustrated in [11], and is perhaps best summarized by Gallavotti in [5, p. 170]. In a simple example, he demonstrates that the constraint violation "... Y~ < 1/x/~ for 2 large. At the same time, however, the coordinate Ya(t), which simply represents the constraint's violation, oscillates more and more rapidly: in fact, its vibrations have a frequency:
(2 71- (D2) 1/2 ,,
tg;~ -- -

2re

As noted in [11], this behavior can be derived analytically in multi-degree-of-freedom classical linear and nonlinear systems. From a pragmatic viewpoint, it renders numerical integration of the penalty equations of this form via conditionally stable integration schemes infeasible: a large penalty value is required for convergence, but includes oscillations requiring a prohibitively small time step.

82

Computational Mechanics 12 (1993)

At this point it is worth pointing out that there exists a vast collection of literature pertaining to penalty approximations of bounded linear operators that exhibit coercivity properties [7, pp. 45-121; 17, pp. 82-89]. In some cases, the convergence rate for this class of systems can be accelerated using "augmented Lagrangian" methods. There exist specific time integration schemes for first order, dual evolution equations. However, the nonlinear equations considered in [2, 3, 12, 21] do not, in general, fall within the class of systems considered in [7] and [17] owing to their 1) class of non-linearities, 2) integration schemes, 3) and lack of coercivity (although they may contain a coercive spatial operator). Only recently in [2, 3, 11, 12, 19] has this axiomatic approach been employed in practical simulations of complex non-linear systems arising in classical mechanics. Bayo in [2, 3] introduces variations of the penalty method that are more general than the approaches taken by van Kampen [26], for example. A distinguishing feature of these methods is that penalty contributions related to transverse constraint velocities are added and iterative corrective approaches based upon Augmented Lagrangian methods are used. Further work carried out in [3] extends the classes of multibody systems to include flexible constituent bodies. A distinct but related tack is taken in [19], in which the difficulty associated with conditionally stable integration schemes is avoided by using an implicit--staggered integration method. Kurdila in [11] introduces simple, yet effective, frequency domain estimates for spectral errors in the method of linear substructure synthesis. Some limited convergence results for nonlinear systems are presented in [12]. Even at this early state of development, the diverse directions taken in [2, 3, 11, 12, 19] preclude the derivation of a comprehensive computational/convergence analysis for all of the aforementioned approaches. Still, a significant step in this direction is made in this paper for a subclass of the approaches. In particular, this paper extends the Rubin-Ungar theorem to include some of the approaches taken by [2, 3] and [12] that employ "constraint velocity penalty terms". Two additional important conclusions can be drawn from this paper: (i) The derivation of the Rubin-Ungar theorem including constraint-velocity penalty terms is significantly simpler than the original derivation suggested by Rubin and Ungar. (ii) It is not a "sterile exercise" to derive the sufficient conditions for convergence. This paper demonstrates that if, in particular, the condition stipulating that the constraint Jacobian remain full rank is violated, the calculated response may not converge to the actual response. Even worse, the computed solution can exhibit pathological response in the form of one or more of the following: - - bifurcations - - quasiperiodic response - - chaotic response. These regimes may not exist in the true solution! It should be kept in mind that the cause of this pathological response is well-known and welldocumented [14, 24...] What is unique with respect to the penalty methods is the manner in which the pathological response becomes manifest. As opposed to some catastrophic failure, or divergence, on the part of the simulation method, the regularized solutions obtained usin9 the penalty formulation may be indistinguishable from correct trajectories. This latter conclusion suggests additional research in these approaches, particularly in view of the fact that some related methods claim improved performance precisely for systems that exhibit rank deficient Jacobians [19].

G o v e r n i n g equations

The derivations that follow specify sufficient conditions for the convergence of the penalty approximations for nonlinear multibody systems. Inasmuch as these results rigorously substantiate results in [12], and are an extension of results in [21], as much of the notation from these papers as possible has been adopted herein. The equations of motion are expressed in terms of n generalized coordinates q = { q l . . . q n } r where q(t)6~R" [Euclidian n-dimensional space].

A. J. Kurdila and F. J. Narcowich: Sufficient conditions for penalty formulation methods in analytical dynamics

83

Only natural systems [15, pg. 77-1 are considered, that is, systems for which the kinetic energy, potential energy, and Lagrangian have one form
T* = l d l i m i j ( q ) q J ,
V* =

V*(q),

L* = T* - V*.

In addition, the system is subject to holonomic and stationary (scleronomic) constraints r where
d<n.

i=l...d

The following assumptions hold throughout the remainder of this paper: (A1) The potential energy is continuously differentiable in a closed bounded region D of g~":
V*(q)eCl(D).

Consequently, V* must be bounded on D; indeed, we shall assume V* is nonnegative, so


O < V*(q) <- C

if qeD.

(A2) The elements of the generalized mass matrix [mlj(q) ] are twice continuously differentiable in the region D:
mij(q)e C2(D).

Moreover, we assume that [mij(q)] is symmetric and positive definite for all qeD. Obviously, these assumptions imply certain restrictions on the spectrum of m(q); namely, there are constants C O and C o, such that for all q contained in D,
/

C O _--<O'min(m(q)) < O'max(m(q)) =< C o.

(A3) The constraints are twice continuously differentiable in D;


O'(q)eC2(O).

(A4) There exists a qo in the region D such that


r =- O.

(A5) There exists a constant virtual velocity vector


(t o e 9 t "

such that r -k ~qk (qo)qo = 0"

(A6) By some appropriate numbering of the coordinates,


ql # 0 i=l...d

a j x j minor of the dacobian satisfies

j = 1...d

for all q~D.

(A7) The d d penalty matrices ~, fl, I~ are symmetric, positive definite, and have the property that their smallest eigenvalues become unbounded as e approaches zero. O'min(N(/~))---~(:x3 as O'min(fl(~))'---~ O(3 as O'min(~(/~)) -4. (30 as
e~O e~O e ~ O.

The class of penalty formulations considered in this paper are the "inertial penalty methods" discussed in [2, 3-1 and [12]. In these formulations, the holonomic constraints are enforced only

84 approximately by considering solutions to the equations

Computational Mechanics12 (1993)

-dt \ ~0 k

a (

dqk

71--- {~ij(S)~ j -t- #ij(S)~ j -~- O~ij(~)~j} dqk

= 0

(1)

as s approaches zero. One can view the constraint functions as depending upon the generalized coordinates as shown below:

dt \30 k 2 ) -- Oq----~ + --Oq T* k

-4- C3 (q) {fllj(s)t~J(0, q, q) + ].,tij(s)~)J(O, q) + (xij(8)~oJ(q)} = O.

(2)

Our aim is to show that if q,(t) solves (1) (or, equivalently (2)) subject to q,(O) = qo, 0~(0) = qo (see A4 and A5), then there is an interval [0, T] on which

(q,(t), (h(t)) ~ (q(t), O(t)) as e ~ 0 uniformly, where q(t) is the unique solution of the constrained problem described earlier.
For the following derivations it is assumed, for convenience, that p = 0. All of the conclusions derived can be extended to the case in which # is not zero by utilizing the Lyapunov function for the system as in [12]. By employing standard kinematic identities [15, p. 74] d ~(~-~

(lidp'iflij( F.)~"j) --D(l~)J~ij(e)~


1 d "i "i, ok ~

j ) } = ~i~ij(C,)~Jtgq-~

(3)

one can derive a work-energy principal, or Jacobi integral, for the system.

q ~-~_kPijq = -~-~t(qb fliidP )

~.k 4) ~ ",'j

cqJ(aJ= ~ t (q~ cqjq~J)

ld

(4, 5) (6)

ok r d ,/ ,~ T , ) \

0 (T~) + - -

V* =

(T~ + V*).

This conservation statement takes the form: !,~q~ d~J "q- 1-,6i~.'~J "nt- T* + V* = constant. 2"r r'ij'r 2"r - - t j r But by evaluating at the initial conditions, we have: constant = T~(qo, 40) + V*(qo) = Eo (independent of 0. Suppose now that one solves the governing penalized equations for a sequence of values of s tending to zero. From the bounds above, one can easily conclude that

!5im..h j < E o. 2"-~e. zj-xe


One also has II~ II <

(
\ O'min (fl(8)) / /

--, 0

as

s ~0

(7)

2Eo ~ 1/~
~0 as s~0 (8)

I[~ll<\%i.~-~,(s)i/..... -

so that from (A2) it follows that

II0~11 \-~o/

<(2Eo~ 1/2
" (9)

This latter bound is immediately useful in that it defines a common domain of definition for the

A. J. Kurdila and F. J. Narcowich: Sufficient conditions for penalty formulation methods in analytical dynamics

85

family

Indeed, by integrating and employing the initial conditions, one finds that

[[q~(r)-q(O)l[ = ! 4.dz < ~ [I4.l[dz<


o

=\C0/

r=-cxT.

(10)

Consequently, if one chooses T so that c~ T is the radius of a ball B with center at qo that is contained in D, one is guaranteed that {q,(t)} c B c D for all time 0 < t < T, and for all e > 0. (See Wiggins [25], p. 37.) While it is by no means necessary, the derivations that follow are simplified considerably by employing Routh's equations, as described in [21]. To this end, introduce a change of coordinates in terms of the given holonomic constraints. ,i=(~b'(.q) i = 1...d ~. (11)

\ q'

i=d+l...nJ

Assumption (A6) and the implicit Function Theorem [4] permit the coordinate change in (11), at least locally. It may be necessary to adjust D so that the map ~: D ~ v(D) has an inverse in ?(D). One should note the similarity of this philosophy to the canonical choice of coordinates used in Maggi's equations [13]. With this change of coordinates, the potential energy can be written as V(7) -= V*(q(7)). Similarly, the kinetic energy can be transformed into
T2(7,

= ~7 Mkj(7)7 - ' .k

where the matrix M(?) is related to re(q) via

M(7) - J(7)m(q(~'))J- 1(?) Oqi


J = -=. O~, J (12) As noted in Rubin and Ungar [21], the assumption (A6) is sufficient to guarantee that J is nonsingular. Clearly, one also has that mkj(7) are

C~(),(D)).

The final step in setting up Routh's equations is showing that the system of equations

Pk- ~T2 k = d + l , . . . , n O~k


can be solved for the in terms of

~l,...,~,d

Pn+l,"',P,"

This system is, in fact, linear and being able to solve it is equivalent to the invertibility of a submatrix of M(7); namely, the block of elements M~k with j, k = d + 1,..., n. It is easy to check that, because M is positive definite, the submatrix is also positive definite. Hence, it is invertible. For future reference, note that the result of solving the linear system is

~k = F~(?)~' + Gkjpj

86

Computational Mechanics12 (1993)

where i = 1,..., d and j, k = d + 1,..., n. The coefficients on the right above are easily shown to be continuous on the closed region 7(D). They are therefore uniformly continuous on ?(D). The Routh function can now be introduced as

R(7,~l...~d, pa+l...pn)=pi~)i-- T 2 4" V i = d + 1...n = aij(7)~i~ j + 2a~(?)pi~ i + alJ(7)plp j q- V.

(14)

It should be noted that we follow Rubin and Ungar [21] in using the Hamiltonian-like Ronth function rather than the Lagrangian-like version used by some authors, e.g. [15, p. 86]. By employing the identities dR

Oyi

d~ i

(V-

T2)

i = 1...n;

dR

d~ i

d~ i

(V-

T2)

1...d;

--

aR

dpl

= ~'

i = d + 1 . . . n (15-17)

the equations of motion are

\~f

- ~yk + ~yk + fl'Jfi + #iJ~j + ~iJyj = 0 k=d+l...n.

k = 1...d;
(18)

d(~T2~-dT2+d--V=o dt \ d~k ,/ ~7k d~)k

Alternatively, one can write the equations in the form

( d R ] + d R + ( f l i j ~ i + llii~ j +~ijyJ)=o dt \-ff~ / ~?--i


-

i= 1...d

(19)
(20)

dR

~=

t~R

i=d+

1...n.

This completes the derivation of the equations governing the dynamics of the constrained system. Next, the convergence of the solutions as e approaches zero is discussed.
The limiting process

Our approach to showing that q~ and 4~ converge uniformly to a solution of the original constrained problem is to first apply the famous Arzela-Ascoli Theorem [-5, p. 540; 8, p. 270] to obtain uniformly convergent subsequences q~ and ~,~ that converge to continuous functions q(t) and ~(t) for which 0 = 4. We then employ a variational argument to demonstrate that q and ~ satisfy the original constrained problem. We conclude by showing that the restriction to a subsequence is unnecessary, and that q~ and ~ converge uniformly as e ~0 to q and t~, respectively. The Arzela-Ascoli Theorem essentially states that every sequence selected from a uniformly bounded and equicontinuous family of (continuous) functions defined on a closed interval has a uniformly convergent subsequence. The term uniformly bounded means not only that each function in the family is bounded, which is always the case for a continuous function defined on a closed interval, but also that the same bound works for every member of the family. A clear example of this is found in (9), where we have

=\C0/
Every member of the family {t~} is bounded by the same constant, and so the family is uniformly bounded. We will now use this fact to show that the family {q,} is equicontinuous and uniformly bounded. We start by integrating ~ to obtain
t

q~(t) -- q~(z) = S ~h(S)ds,

A. J. Kurdita and F. J. Narcowich:Sufficientconditions for penalty formulation methods in analytical dynamics which, when combined with the previous inequality, yields

87

]tq~(t) - q~(z) ]] < i ]l(h(s) Hds < [t - z] ( 2E~ ~ 1/2.


, \ C o /

This inequality shows that {q,} is equicontinuous: Given an arbitrary positive n u m b e r t/, take

( Co ~ 1/2
6=
~ \2-E~o/ "

Notice that this choice of 6 does not depend on the label e, and that, from (21),
ff q~(t) - q~(z)II < tl

holds whenever It - ~] < 6. Because 6 can be chosen independently of the label, the family is equicontinuous. (Had we been unable to choose 6 independently of the label e, each m e m b e r of the family would be continuous, but the family would not be equicontinuous.) The uniform boundedness {q~} is also a consequence of(21). In (21), take 9= 0, use t < T and q~(0) = q0 and finally employ standard inequalities to obtain IIq~(t)Jl < IIqo tl +

T(2Eo~l/2
\~-07

from which it follows that {q~} is uniformly b o u n d e d on the interval [0, T]. At this point, it is tempting to apply the Arzela-Ascoli Theorem and obtain a uniformly convergent sequence from the family {q~(t)}. There are technical difficulties with such a direct approach, however. For example, we might get a sequence {e,} and corresponding functions {q~k} for which ek Vr O! Consequently, we first pick a sequence of {ek} such that ek --" 0 as k tends to infinity. Let qk = q~*, to simplify the notation that follows. By the Arzela-Ascoli Theorem, we m a y extract a sequence {qk'} that converges uniformly to a continuous function q(t), t e [0, T]. Of course, since ek ~ 0 as k ~ 0o, one also has that ek, -* 0 as k' -~ 0o. F o r our purposes, we may assume the original sequence {ek} gave a uniformly convergent {q~k}" The continuity of the constraint functions implies that the limit function q(t) satisfies the constraints in the original system. To see this, recall that in (8) we found that lI~b~]l ~-ll~b(q~)II < --->0 ~ 0 .

= \ami.(~(e)))
This implies that ~b(q~(t))~ 0 as e ~ 0 on [0, T]. If we take e ~ 0 along the subsequence constructed above, then, since ~b is continuous on D, we also have that lira ~b(q~(t)) = ~b( l i m q~(t)) = (D(q(t)) =- O. ~-~o \~-~o / (22)

Note that the uniform b o u n d on [[~b~[] in (8) shows the convergence of q~(q~(t)) ~ c~(q(t)) -- 0 is uniform in [0, T]. The uniform convergence of

4)(q~(t)) -~ ~p(q(t))
coupled with the definition of~ coordinates given in (11) implies that the curve qdt) in ~,coordinates, ?~(t) = ~o q~(t), also converges uniformly to ~ o q(t) - 7(t). We could have derived this uniform convergence of L(t) in another way. The m a p 7: D-~ ~N is continuous in a closed, compact subset of ~sv, and it is uniformly continuous in D [4]. It is quite easy to show that the composition of ? o q~(t) is uniformly convergent on [0, T] as e ~ 0: First note

88

C o m p u t a t i o n a l Mechanics 12 (1993)

that the uniform continuity of 7 implies that for every choice of 3 we may find A so that II7 ( q ) - 7(q')II < whenever
IIq -- q' II < A,

no matter where q, q' are in D. Because q,(t)~ q(t) uniformly in t, we may find a k 0 so large that for all ek in our sequence, with k > ko, IIq~(t) -- q(t)II < A Thus, II~(q~(t)) - ~(q(t))II < whenever k > k0. This implies that 7oq,(t)~ ~oq(t) uniformly on [0, T]. The reason that we included this argument is that it can be used to establish the uniform convergence of M(7~(t)), J(~(t)) and a number of other functions that appear as compositions with 7~ or q,. This fact will prove useful later. We now turn to a discussion of the velocities and momenta involved in our problem. By writing the appropriate Jacobi integral in the transformed coordinate system, one can conclude (as in the case of the original generalized coordinates) that supt~to,rll~(t)[ < C2. In addition, from the definition of the conjugate momenta (23)

3T
Pi = g~i

i=d+l

.... ,n

one may obtain the identity [Rubin]


= p,.,j=d+l

2
j=l

i= d + 1...n.

(24)

One can conclude that the e-family of conjugate momenta is bounded uniformly. supt~to.rI ]pi,,(t)l < C3. Finally, from the equations of motion (20) and the definition of the Routhian function it follows that

[~i,, is uniformly bounded, i = d + 1...n.


Similarly, as derived earlier for the generalized coordinates, the uniform boundedness of the derivatives of the conjugate momenta implies that

Pi,~ is equicontinuous.
Now, (24) and (25) imply by the Arzela-Ascoli theorem that

(25)

pi,~---,pi~C~

T]

i = d + 1...n

(26)

uniformly for some subsequence, still indexed by 5. To show uniform convergence of the normal velocity components ~
9 _ j

j=l ..... d
"j

one need only observe that

~ -d?~ j = l , . . . , d
in this case. Hence, by (7), these components converge uniformly to zero on 0 < t < T as e approaches zero. To get uniform convergence for the remaining velocity components, recall that (13) gives

~J j = d +

l,...,n

as uniformly continuous functions of y and a linear function of the remaining velocity components

A. J. Kurdila and F. J. Narcowich:Sufficientconditions for penalty formulation methods in analytical dynamics and momenta; it therefore is uniformly continuous in all of the variables. Because

89

)'~,Pi,~,~ j = l , . . . , d

i=d+l,...,n

are uniformly convergent, we m a y employ the second m e t h o d that was used to derive the uniform convergence of ~ on [0, T] to show that

~ ~

j=d + l,...,n rt~C~ T]

is uniformly convergent on [0, T] as e approaches zero. Thus, we have that

uniformly in [0, T], which implies that

uniformly on [0, T], by the continuity of the change of variables. We now wish to show that
4(t) =

But this result follows easily since


t

q,(t) = qo + ~ dl~(z)dz for each


0

and the uniform convergence of

enables one to interchange limit and integral, so lira q~(t) = q(t) = lira qo + gt~(z)dz hence,
t

q(t) = qo + S r o

dz.

(27)

The desired result 0 = ~ now follows from the fundamental theorem of calculus. Let us briefly summarize what we have shown thus far. If we can choose a sequence ek ~ 0 as k ~ ~ , then there is a subsequence for which

q~kt(t)~q(t), c~(q(t)) =--0,

dl,kz(t)~dl(t),
t~[0, T]

as/~

with convergence being uniform on [0, T]. Moreover, we also have that that is, q(t) is on the constraint surface, and, of course, q(t) is tangent to the constraint surface. What remains to be shown is that q(t) satisfies the constrained dynamical problem with which we started. Showing that q(t) satisfies the constrained dynamical problem can be done using a "variational" argument similar to the one in [21]. To begin, note that the dynamical equations in (1) [with p(~) = 0] are equivalent to setting

6 I(L* +89189
o

=0,

where the only constraint in 6q* is that

6qk(O) = 6qk( T) = O,
and that fiqk(t) be differentiable. (Of course, we are doing the variation about the penalized solution

90

Computational Mechanics 12 (1993)

q.(t).) Carrying out the variation above leads to


T

This formula can be rewritten as T[ dL* c~L*6qkldt+l~[0 : !L-~qkfZlk+ Oq~ J

-if~dPi~ j ~JoL~U(pk-~qkOq k'~ )+o:,/p , ( ~C3dPf q qk


9j

)]dt
.,

+,-oL
TVt~L*

(~qk6q)+fliJd?(~qk6q)]+[fl'J~)(~q-~q6q)+fl'J~(~q6q)]
.,

dr"

Using ~j = ctji, fiij = flji and relabeling summation indices, we find that the last equation becomes
~L*

k-1

t/t~rhi

d ~ i

Integrating by parts above and using 6qk(O) = 6qk(T) = O, we obtain the following:

o \ aq k

dt

8~IkJ

( t~qk
o \ cq
S (L*)

At this point, we require that


aq k 6q k -- 0,

l = i...d,

which gives us

dt ~ k

(L*)

6qkdt =

O.

From this point on, the proof that q(t) satisfies the unconstrained dynamical problem is identical with that used in [21], and we omit it. We conclude our derivation by dealing with a technical point. Thus far we have established uniform convergence of q~(t) and q~(t) when the e's used form a subsequence {F, } of the sequence kl {ek}, which apart from the condition that ek ~ 0, is arbitrary. Our original goal, however, was to show that q~(t) and 0~(t) converged uniformly to q(t) and 0(t). In fact, we have accomplished this, because we have shown that every sequence ek ~ 0 has a subsequence ekt --' 0 such that q~kt(t) ~ q(t) and q~u(t) ~ q(t), the same limits in all cases. If the original sequence q~k(t) did not converge to q(t), then we could find a number 60 and a subsequence ek~~ 0 such that for all Skit
t~[O, T]

sup Iq~(t)-- q(t)l > 60.

This is impossible, however. Simply start with ek -~ eki as the original sequence, and we would have convergence for yet another subsequence. Thus both q~k(t) and ~l~k(t) converge uniformly to q(t) and ~(t). Since ek--}O was arbitrary, standard arguments from analysis [8] then give us that q~(t)-~q(t) and ~h(t)--}gt(t), uniformly on [0, T] as e-~0 +.

Numerical examples
Since considerable empirical evidence has been supplied for cases in which the convergence of the penalty approximations are quite good [2,3, 11, 19], this section seeks to emphasize that the quality of the approximation relies upon the "regularity" of the penalized system. The model for simulation has been selected precisely because it is a trivial example of a multibody system which has a constraint Jacobian matrix whose rank depends upon the configuration. Practical simulations of complex systems that exhibit the same character can be found in [14, 10, 24]. The mechanical system selected for simulation is a slider-crank mechanism. As opposed to the usual (and in this case highly advisable) method of expressing the motion of the system in terms of a minimal coordinate system, the redundant coordinate system shown in Fig. 1 has been employed.

A, J. Kurdila and F. J. Narcowich: Sufficient conditions for penalty formulation methods in analytical dynamics

91

Fig. 1. Coordinates for slider-crank system

With this choice of coordinates, the potential and kinetic energy of the (open loop) system are, for a horizontal static equilibrium state,
I_M L 5 c o s ( O 5 - O1) ML 5

o l, '

V = g1 I O 2 + 89 K 1

_ 01)2

while the constraint, and its Jacobian matrix with respect to the generalized coordinates, are q ~ = L s i n @ l + L s i n O 5,

L 0qk_l

[~b'l=[cos@lcos@2]

respectively. Finally, the approximate penalized equations can be written as


2ML2 + (cos 191)2 8 ML2 c~ (O2 -- 191) + cos 191 cos 192 ML2 cos (O2 - 191) -~
8 01

M L 2 + (cos 191)2

95

+[L
e k cos

.cos
O,+ OXsinO,,=o

191

'Cc~176

Completely analooous equations can be written for the case in which the vertical configuration is the static equilibrium state. Figures 2a-d illustrate how well the method works when the derived conditions for its use have been met. In Figs. 2a-d, the parameter e has been selected to be 1.0E-5. Figures 2a and b depict the Poincar6 sections corresponding to O1 and 195, respectively. In Fig. 2c, the "constraint coordinate"

q~= L sin 191 + L sin

192

and its derivative are plotted in the phase plane. As one can see, the constraint violation is of the order of 10- 7, while the error Error = 191 + 192 is of the order of 10-s (Fig. 2d). In simulations shown in Figs. 3 and 4, the constraint violation

1
II

m
0 0

II
cq ,-~ ~
0

~lO~lh'H
o0

~10~1~t~
o

o. H

"7,

6
I I I I 0,1 0 Cq
I I I I ~ Q I I I I

?
0o 6
r e~

~0

C',1

t"l

o
(clvu) Zo
en

(uvu) ~o
e~

((]Wl) ~e ~0 ~-

Cq

~l

00 ~D e,l

(avu) ~.o

((]V~l) tO

(avu) i?

g
i I I

g
he~

II

c~
E

rr

;'nmumum~ui~m

"~

q
~OHSH 1r

q
80888
i i r

r~ o.
r

Jl

I I ! t

~J
c-,I 0

I !

l
iI

/
I
0 0

c0

II

Z %. I
eO ,~

~
i p

o
~,
i

~. o

o
t~

'7~
<

],q.~
i

n~
o

'7

0 0

'7

,.~

~ 9

~.'

o
(GVa) ~o

o,
e~

((IV~ Zo
I

(aw)

c~

~=~

,-i

n~
c,4

,..~<

M*
r~ ~=~

c3 0

r~
)4
I I

(ow) ~?

(cIV~) ze

( o v a ) Ze

g
i i i

g
C~

Cq

.... ..._____.__

o
o

II

c'q

o aoaaa

cq

o
s

uOU~

~'

o,"
c~

Cq

"~

In o~

o.
0

9 "'.'L ..-'..., .

i::

H
i, . .... ~..

..:....--

.-

..

-......

:.:.

II
e

."!~!"
I I [

0
9 : '.:., . ~:'...';;.-:!:. ". 9 : ~.i" . . "~.~.
" .L .~' ; ~ : ~ "::~,"', ".

~.:';."....y..:...'.,.:.,,.: ~L'.'. ;':: ~,,. ~,"


.,..,... 9 %- 9 ....~

9 .:/4, . ::..... ,,.,,:


: >: " . , ~ . "t '~:

I n

o.

e~

o"

o,"
7

o,

~c

......:.......
,

,~
,

..' ..:::~,~,;,~,,;;:. ..
:;.-:.,, :'-~,.'-.

'

9: : "
-, : -.:"

~;:..
..:

..

,~::
~ :

9 ~
~r "

~z~

..:.. I.r162

j... o

iii ! "
I
eq ~

iii !
t

:'.:

.:

; .:'r

"~

,"',

c',f'T
C~

~-7

((IV}l) ro

(aV}l) ~0

'

n
n

~'~

0 :..,~ : ". :'." ~;~,'..; " 9 9 ::~.r<"

9: ; ~ ' ~ . -

..

./ o 0

. . . . . }# :.

I
0

w~

t0

~'

" ~ r,O

o.
tt~ co e,I

II

,'0

,'o
C~ ~O~R 0

C~

o~
i I Wi i i

t'~

o.
0

II

('-,I tt~

0 eq

q~ v~ .!-. ~L

C~ ....~'.... tt~ . ,~.... C 0

,,q.

oO

oO

Ox

t~

oo

oo

t"-

oO

oo

,:?
((]v~I) ~0 ((lWl) ~0 ((IV~l) ~0

"t,;

~
0

tt~ < " 9 . . . ~ . ,." "

'9,

__

I ~.

((Ira) t0

((IVY)

io

96

Computational Mechanics 12 (1993)

and error converge to zero as ~ approaches zero. Figures 5c-d depict exactly the same data, but in a configuration that renders the constraint Jacobian singular. In this case the simulations do not converge to the true solution. Instead, a limit cycle behaviour can be observed in the second degree of freedom that is not even present in the first degree of freedom. Figures 6-13 depict the forced response of the same system. As in the previous case, convergence is clearly illustrated when the Jacobian retains full rank. In the singular case, the system response is apparently chaotic.
Conclusions

This paper has derived sufficient conditions for the convergence of a class of penalty approximation methods suitable for multibody simulation. The derivation is based upon the work of Rubin and Ungar, but includes penalty terms that are quadratic in the derivatives of the constraint coordinates. This approach not only simplifies the original Rubin-Ungar derivation, but also enables one to control the prohibitively large frequency of oscillation described in Gallavotti [5] that is associated with the constraint degrees of freedom. Perhaps most importantly, this paper illustrates the importance of retaining a full-rank constraint Jacobian matrix throughout the simulation when using the class of penalty formulations described herein.
References
1. Arnold, V. I. (1978): Mathematical methods of classical mechanics. Berlin, Heidelberg, New York: Springer 2. Bayo, E.; Serna, M.A. (1989): Penalty formulation for the dynamic analysis of elastic mechanisms. J. Mech. Trans. Automat. Design 111,321-327 3. Bayo, E. et al. (1988): A modified Lagrangian formulation for the dynamic analysis of constrained mechanical systems. Comput. Meth. Appl. Mech. Eng. 71, 183-195 4. Fulks, W. (1978): Advanced calculus, 3rd Ed. New York: Wiley 5. Gallavotti, G. (1983): The elements of mechanics. Berlin, Heidelberg, New York: Springer 6. Gear, C. W. (1971): Simultaneous numerical solution of differential-algebraic equations. IEEE Transact. Circuit Theory 18,1 7. Glowinski, R.; Le Tallec, P. (1989): Augmented Lagrangian and operator-splitting methods in nonlinear mechanics. SIAM Studies in Appl. Mathem. Philadelphia 8. Goldberg, R. R. (1964): Methods of real analysis. Waltham, Mau., Blaisdell 9. Joshi, S. M. (1989): Control of large flexible space structures. Berlin, Heidelberg, New York: Springer 10. Kim, S. S.; Vandoerploeg, M. J. (1986): QR decomposition for state space representation of constrained mechanical dynamic systems. Journal of Mechanisms, Transmissions, and Automation in Design 108, 183-188, June 11. Kurdila, A. J.; Junkins, J. L.; Menon, R. G. (1991): Linear substructure synthesis via Lyapunov stable penalty methods. Finite Elements in Analysis and Design (accepted for publication) 12. Kurdila, A. J.; Junkins, J. L.; Hsu, S. (1990): On the existence of limit cycles for a class of penalty formulations of dissipative multibody systems. Southeastern Congress on Theoretical and Applied Mechanics, Georgia Institute of Technology, Atlanta, Georgia, March 22-23 13. Kurdila, A. J.; Papastavridis, J. G.; Kamat, M. P. (1990): Role of Maggi's equations in computational methods for constrained multibody systems. AIAA Journal of Guidance, Control and Dynamics 13, 1, 113-120 14. Liang, C. G.; Lance, G. M. (1987): A differentiable null space method for constrained dynamic analysis. Journal of Mechanisms, Transmissions, and Automation in Design 109, 405-411, Septembei" 15. Meirovitch, L. (1970): Methods of analytical dynamics. New York: McGraw-Hill 16. Nayfeh, A. H. (1990) Perturbation methods. Wiley-Interscience, New York 17. Oden, J. T. (1990): Qualitative methods in nonlinear mechanics. John Wiley & Sons, Inc., New York 18. Papastavridis, J. G. (1990): The Maggi, or canonical form of Lagrange's equations of motion of holonomic mechanical systems. J. AppL Mechanics 57, 4, 1004-1010 19. Park, K. C. (1990): Stabilization methods for simulation of constrained multibody dynamics. To appear in the proceedings published by Computational Mechanics Publications 20. Rasband, S. N. (1990): Chaotic dynamics of nonlinear systems. John Wiley & Sons, Inc., New York 21. Rubin, H.; Ungar, P. (1957): Motion under a strong constraining force. Communications on Pure Appl. Mathem. 10, 65-87 22. Serna, M. A.; Bayo, E. (1988): Numerical implementation of penalty methods for the analysis of elastic mechanisms. Trends and Developments in Mechanisms, Machines and Robotics--1988, pp. 449-456, ASME, New York 23. Walker, J. A. (1980): Dynamical systems and evolution equations. Plenum Press, New York 24. Wehage, R. A.; Haug, E. J. (1982): Generalized coordinate partitioning for dimension reduction in analysis of constrained dynamic systems. J. Mech. Design 104, 247-255, January 25. Wiggins, S. (1990): Introduction to applied nonlinear dynamical systems and chaos. Springer Verlag, New York 26. van Kampen, N. G. (1985): Elimination of fast variables. Physics Reports 124, 69-160 27. Zeidler, E. (1984): Nonlinear functional analysis and its applications IlL Springer-Verlag, New York
Communicated by S. N. Atturi, October I, 1992

Вам также может понравиться