You are on page 1of 9

Classical Mechanics Revisited

S. H. Lam∗
Department of Mechanical and Aerospace Engineering
Princeton University, Princeton, NJ 08544, U. S. A.
October, 2009

Abstract
Two linear algebra based strategies are presented to deal with gen-
eral classical mechanics problems with constraints. Classical concepts
such as holonomic and nonholonomic constraints, virtual displace-
ments, principle of virtual work, Hamilton’s principle and method
of Lagrange multipliers are critiqued.

1 Introduction
Let x, a N -dimensional column vector, describe the configuration of a general
classical mechanics system. For example, a n-particles system in 3 dimensions
has N = 3n. Newton’s Law of Motion for x(t) in vector form is:
ṗ = F + f , p ≡ M ẋ, (1)
where F(x, ẋ; t) is the (known) applied force, f is the (unknown) constraint
force, ṗ is momentum—all are N -dimensional column vectors—and M is a
N × N (nonsingular and symmetric) mass matrix. The operator is the
N -dimensional inner product operator.
In addition to Eq.(1), the solution x(t) is required to honor K equations
of constraint (K ≤ N ). We assume that these equations of constraint are
first order differential equation (ODE) involving ẋ, x and t:
Gk (ẋ, x, t) = 0, k = 1, 2, . . . K. (2)

Professor emeritus. Email: lam@princeton.edu

1
Revised August 21, 2013 2

Some of them may be integrable. See §4.1 later. If a constraint is originally


given by an algebraic equation, it can always be differentiated with respect
to time to obtain an equation of constraint which Gk (ẋ, x, t) depends on ẋ
linearly.
These K constraints are being enforced by the constraint force f . We
assume that f can be represented by the sum of K linearly independent
column vectors:
K
λk dk ,
X
f= (3)
k=1

where the dk (x; t)’s are the given directions of the k-th contributing N -
dimensional constraint force vectors and the scalars λk ’s are the unknown
amplitudes of the k-th constraint force vectors.
Hence, a general N -dimensional mechanics problem with K constraints
has N +K scalar unknowns. They are determined by the N -dimensional vec-
tor equation Eq.(1) plus the K scalar equations Eq.(2). This completes the
mathematical formulation of general mechanics problems with constraints.
The two crucial and noteworthy assumptions of this formulation are:

• the resultant constraint force f can be represented by Eq.(3);

• the directions dk (x, t)’s are given and are therefore known, while the
amplitudes λk ’s are unknown and are to be found.

Neither received the attention they deserve in the classical mechanics litera-
ture.
The following concepts are important to the classical approach: holo-
nomic versus nonholonomic constraints, virtual displacements, virtual work,
Hamilton’s principle, method of Lagrange multipliers, etc. [1]. Most of the
elegant classical results are known to have “restricted” validity. Are these
concepts and their restrictions necessary?

2 The monkey problem


The following “monkey problem” is simple enough to be solved exactly so
that the strengths and weaknesses of various approaches can be compared.
Consider the two-dimensional motion of a small monkey of mass m cling-
ing to a massless rigid rod hinged to a frictionless hinge at the ceiling. The
Revised August 21, 2013 3

monkey is instructed to honor one of the following constraints by climbing


up and down the rigid rod:
1. Keep his distance to the hinge fixed.
2. Keep his elevation fixed.
3. His vertical velocity is some given function of his position and time.
For this problem, we have N = 2, K = 1, F is gravity and f is the constraint
force by the rigid rod acting on the monkey as he climbs up and down. Thus
d1 is in the radial direction. Constraint #1—the classical simple pendulum
problem—can be correctly solved by all strategies. Surprisingly, the other
two constraints will be found to be problematic for the classical approach.

3 Two linear algebra based strategies


The two linear algebra based strategies presented below are conceptually
distinct from the classical approach. Unlike the classical approach, these
“new” strategies have unrestricted validity. They are called new because
they are not considered sufficiently worthy to be included among respectable
strategies in the classical mechanics literature.
Can linear algebra alone deal with Eq.(1) and Eq.(2) without the help of
any of the above-mentioned classical concepts?

3.1 Strategy A: solving for x(t) first


In classical mechanics, the main interest is on x(t). Often there is no interest
in the λk ’s at all. So a preferred strategy is to ignore the λk ’s and focus on
x(t).
Eq.(3) says the constraint force f resides in the K-dimensional column
subspace spanned by the K known column vectors dk ’s. This subspace shall
be referred to as the constrained column subspace. In general, the dual of
a N -dimensional column vector space is a N -dimensional row vector space.
Let bn (x, t) denote any complete set of N linearly independent row basis
vectors. One can easily use linear algebra to find a special set of bn (x; t)
which has the following properties:
bn dk = 0, n = 1, 2, . . . , N − K, k = 1, 2, . . . , K, (4a)
bN −K+m dk = Λmk , m, k = 1, 2, . . . , K. (4b)
Revised August 21, 2013 4

where Λm k can be any nonsingular K × K matrix. In other words, the first


N − K members of the set are orthogonal to the constrained subspace while
the last K members of the set are in the constrained subspace. Therefore the
b1 , . . . , bN −K row vectors span the (N − K)-dimensional unconstrained row
subspace. In general, these bn ’s as defined by the dk ’s and are conceptually
independent of the specifics in the equations of constraint.
Taking the inner product of bn with Eq.(1), the resulting N scalar equa-
tions can be divided into two groups:
bn [ṗ − F] = 0, n = 1, 2, . . . , N − K, (5a)
K
bN −K+m [ṗ − F] = λk Λm
X
k ,, m = 1, 2 . . . , K. (5b)
k=1

The unknowns λk ’s and the unknown x vector are now completely decoupled.
The solution x(t) can then be obtained by solving the initial-value problem of
the N ODEs provided by Eq.(5a) and Eq.(2)—provided that a solution exists.
If the λk ’s are of interest, they can be computed from Eq.(5b) afterwards.
For the monkey problem, b1 is in the azimuthal direction for any externally-
imposed constraints.

3.2 Strategy B: solving for the λk (t)’s first


It is also possible to use linear algebra to first solve for the λk ’s in terms of
x, ẋ and t.
Assuming that M is a nonsingular constant matrix, one can solve for ẍ
from the Newtonian Eq.(1):
K
0
ẍ = M−1 [F + λk dk0 ].
X
(6)
k0 =1

Regardless of whether any of the equations in Eq.(2) has ẋ dependence,


one can differentiate each Gk (ẋ, x.t) with respect to t (more than once if
necessary) until ẍ appears to yield K scalar equations:
dGk ∂Gk ∂Gk ∂Gk
= ẍ + ẋ + = 0, k = 1, 2, . . . , K. (7)
dt ∂ ẋ ∂x ∂t
An appropriate name for Eq.(7) would be super equations of constraint. Using
Eq.(6) to eliminate ẍ from Eq.(7), one obtains:
K
∂Gk 0 ∂Gk ∂Gk
M−1 [F + λk dk0 ] +
X
ẋ + = 0, k = 1, 2, . . . , K. (8)
∂ ẋ k0 =1
∂x ∂t
Revised August 21, 2013 5

This is now a system of linear algebraic equations for the K unknown λk ’s.
The solution is:
K
∂Gk” ∂Gk” ∂Gk”
" #
k
Qkk” M−1 F +
X
λ (ẋ, x, t) = − ẋ + ,(9a)
k”=1 ∂ ẋ ∂x ∂t
#−1
∂Gk”
"
Qkk” (ẋ, x, t) ≡ M−1 dk , k”, k = 1, 2, . . . , K. (9b)
∂ ẋ

All entities on the right hand side of Eq.(9a) and Eq.(9b) are known functions
of x, ẋ and t. Thus so long as Qkk” exists, the constraint force f by virtue of
Eq.(9a) is now a known function of x, ẋ and t. Hence x(t) can now be rou-
tinely generated by solving the initial-value problem of Eq.(6) supplemented
by Eq.(9a). It is guaranteed to automatically satisfy the original K equations
of constraints in Eq.(2). The Qkk” matrix depends on both the specifics in
the equations of constraint and the dk (x, t)’s. If Qkk” does not exist, then the
problem does not have a solution.

4 Classical approach revisited


The highly respected textbook Classical Mechanics by Herbert Goldstein [1]
shall be the main reference for this section.
The classical approach recognizes different kinds of constraints. When
∂Gk /∂ ẋ is a null row vector (i.e. the original Gk does not depend on ẋ at
all), the k-th constraint is said to be holonomic. Otherwise, it is said to
be nonholonomic. When J of the K equations of constraint are holonomic
(J ≤ K), the J algebraic equations of constraint enable all N components
of x to be expressed in terms of N − J generalized coordinates [1, see §1-3,
p. 13] which are denoted by qj here:

x = x(q1 , q2 , . . . , qN −J ; t). (10)

For the monkey problem, constraints #1 and #2 are holonomic, while con-
straint #3 is nonholonomic. Problems with all holonomic constraints are
considered easier by the classical approach because all the λk ’s are decoupled
(under certain restriction). If the holonomic λk ’s are of interest, they can be
found by the method of Lagrange multipliers afterwards. The λk ’s associated
with nonholonomic constraints are not decoupled. Goldstein firmly asserts
that “. . . there is no general way of attacking nonholonomic examples . . . ”
Revised August 21, 2013 6

and “. . . the more vicious cases of nonholonomic constraint must be tackled


individually . . . ” [1, see §1-3, pp. 15, 16]. When only K −J of the constraints
are nonholonomic, the classical approach brings the K −J nonholonomic λk ’s
into the picture as unknowns via the method of Lagrange multipliers [1, see
§2-4, p. 45]. The nonholonomic equations of constraint are directly involved
in solving for the desired solutions.

4.1 Holonomic constraints


The classical approach is essentially a variant of strategy A (with restric-
tion). Instead of the special bn row vectors defined by Eq.(4a) and Eq.(4b),
the classical approach introduces the concept of virtual displacement, a N -
dimensional row vector. Its definition is [1, see §1-4, p. 16]:
A virtual (infinitesimal) displacement of a system refers to a
change in the configuration of the system as the result of any
arbitrary infinitesimal change of the coordinates “δx,” consistent
with the forces and constraints imposed on the system at the given
instant t.
For problems with J holonomic constraints (i.e. Eq.(4.1)), any of the gener-
alized coordinates qj can be varied to generate N − J linearly independent
virtual displacement (row) vectors:
" #T ranspose
j ∂x
δx = , j = 1, 2, . . . , N − J. (11)
∂qj

The classical approach adopts these δxj ’s to be the special (unconstrained)


bj ’s of strategy A. Goldstein said the methodology is restricted “... to sys-
tems for which the virtual work of the forces of constraint vanishes ...” [1, see
§1-4, p. 18]. In other words, the δxj ’s are assumed to be completely inside
the unconstrained row subspace as defined by Eq.(4a) of strategy A. This
is the pervasive “restriction” in the classical approach. No advice is offered
when the virtual work of the forces of constraint is not zero.
In contrast, strategy A directly finds the special bn ’s using known knowl-
edge of the given dk (x, t)’s. The bn ’s so found are valid without restrictions
for any f expressible by Eq.(3) because bj · f = 0 is honored by definition.
The λk ’s are always decoupled from x regardless of whether the original con-
straints are holonomic or nonholonomic. If the λk ’s are not of interest, then
Revised August 21, 2013 7

only the first N − J unconstrained bj ’s are needed. For the monkey prob-
lem, d1 is deduced by knowing how the monkey enforces the constraint. The
non-zero virtual work done the virtual displacements (as defined by Eq.(11))
by constraint #2 and constraint #3 can readily be computed.
When the (N −J)-th constraint is holonomic, i.e. GN −j = H N −j (x, t), the
row vector ∂H N −j /∂x is parallel to the normal vector of the H N −j (x, t) =
0 surface in N -dimensional space. The vectors ∂H N −j /∂x and dN −j are
conceptually distinct. Since H N −j (x, t) = 0 is a surface of physical constraint,
dN −j on this surface is parallel to ∂H N −j /∂x only if the surface is frictionless.
Sliding friction on this surface would rotate dN −j away from ∂H N −j /∂x.

4.2 Hamilton’s principle and Lagrange multipliers


Hamilton’s principle, which can derive the Newtonian equations of motion
via calculus of variation (under appropriate conditions), is one of the major
gems of classical mechanics. In calculus of variation, constraints are routinely
accounted for by the method of Lagrange multipliers. Goldstein says “It
appears that a reasonably straightforward treatment of nonholonomic system
by a variational principle is possible only when the equations of constraint
can be put in the form . . . ” in which Gk (ẋ, x, t) depends linearly on ẋ [1,
see §2-4, p. 46]. Thus, constraint #3 is in the form the method of Lagrange
multipliers is supposed to work.
The success of Hamilton’s principle/Lagrange multipliers methodology
for some holonomic systems (e.g. constraint #1) is generally accepted. It
is easy to explicitly verify that the procedures outlined in Goldstein does
not work for the monkey problem neither for (holonomic) constraint #2
nor for (nonholonomic) constraint #3 [1, see §2-4 and its eq.(2-29), p. 47].
From the vantage point of the present paper, solutions generated by this
methodology are correct only when the system’s constraint force f vector is
spanned by certain dk ’s (not the ones specified for the given problem). For
constraint #2 and #3 of the monkey problem, this methodology generates
correct solutions only when d1 is in the vertical direction. It does not find
solutions for the given radial d1 . The “enigma” of this methodology for
nonholonomic constraints was discussed by Flannery [2, see its eq.(8)].
To be perfectly blunt, the monkey problem clearly demonstrated that:

• The use of Lagrange Equations of the First Kind to enforce constraints


(holonomic or non-holonomic) for classical mechanics is simply wrong.
Revised August 21, 2013 8

5 Concluding remarks
Classical mechanics has a glorious history. Virtual displacement defined by
Eq.(11) is indeed the correct choice for bj for many interesting and impor-
tant problems—including the monkey problem for constraint #1. But it
is not always the correct choice even when the constraint is holonomic—as
demonstrated by constraint #2. The admonition that the virtual work of a
holonomic constraint force must be zero is a restriction too often overlooked.
For “non-vicious” nonholonomic Gk (ẋ, x, t)’s—those that depend linearly on
ẋ such as constraint #3—the derivation of “constraint equations valid for the
virtual displacements” given in Goldstein is non sequitur from the present
vantage point [1, see §2-3, p. 46;1 see also 3, Chapter 5]. The present paper
proclaims that a classical mechanics problem with constraints is not ready
for solution until how the constraints are being enforced by the actuators are
clearly described. The specification of the equations of constraint alone is
not sufficient.
The weaknesses of the classical approach are clearly demonstrated by
constraint #2 (holonomic) and constraint #3 (nonholonomic) of the monkey
problem. The main flaw is that inadequate emphasis has been given to the
fact that the dk (x, t)’s of the actuators are known and are given. The classical
approach restricts itself to the subset of problems for which the actuators’
dj (x, t)’s are not arbitrary given directions. Virtual displacements are indeed
qualified to be the special bn ’s when the dj ’s are normal to the surfaces of
the holonomic constraints. Many interesting and important problems are
in this category (e.g. rigid body dynamics). However, when the actual
dk (x, t)’s of the system are known, arbitrary and given, virtual displacement
is conceptually irrelevant even for holonomic constraints—its decoupling role
can be achieved by the special bn ’s defined by Eq.(4a) and Eq.(4b). The
classical restriction that the virtual work of the force of holonomic constraints
be zero is totally unnecessary and is easily relaxed.
The linear algebra based strategies A and B presented here are exact and
their validity has no restrictions. Both exploit, as they must and should,
the given known knowledge of the constraint directions dk (x, t)’s. Whether
the equations of constraint are holonomic or nonholonomic and whether the
virtual work of the constraint force is zero are totally irrelevant.
1
The the logic that led its eq.(2-20) to eq.(2-21) made no sense at all.
Revised August 21, 2013 9

Revision
These notes were first written in 2009. Some minor editing were done in
2013.

References
[1] Goldstein, H., Classical Mechanics, Addison-Wesley Pub. Co., 1980.

[2] Flannery, M. R., The Enigma of Non-holonomic Constraints, Am. J.


Phys. 73 (3), March, 2005.

[3] Rund, H., The Hamilton-Jacobi Theory for the calculus of variations.
Van Nostrand, New York, 1966.

You might also like