You are on page 1of 12

Nonholonomic systems as restricted

Euler-Lagrange systems
T. Mestdag and M. Crampin
Abstract. We recall the notion of a nonholonomic system by means of
an example of classical mechanics, namely the vertical rolling disk. For
a general mechanical system with nonholonomic constraints, we present
a Lagrangian formulation of the nonholonomic and vakonomic dynamics
using the method of anholonomic frames. We use this approach to deal
with the issue of when a nonholonomic system can be interpreted as the
restriction of a special type of Euler-Lagrange system.
M.S.C. 2000: 34A26, 37J60, 70G45, 70G75, 70H03.
Key words: Nonholonomic mechanics; vakonomic dynamics; anholonomic frames;
Lagrangian equations.
1 The vertical rolling disk
In this introductory section, we first recall how the notion of a nonholonomic system
appears in classical mechanics. We do so by means of a typical problem from rigid
body dynamics, namely that of a homogeneous disk, such as a coin, rolling on a
horizontal plane while remaining vertical. Let the notations be as in the figure. If C
is the centre of mass of the disk, the equations of motion are given by Euler’s laws:

M¨r
C
= F,
˙
L
C
= M
C
.
There are three forces working on the disk: gravity Mg, a reaction force R
1
due to
the disk’s contact with the horizontal floor in a point A and a reaction force R
2
in C
which ensures that the disk remains in a vertical position during the motion.
We can choose the coordinates of the system as follows. Let (x, y) be the Cartesian
coordinates of the centre of mass C. Since we assume the coin to remain vertical
during the motion, the z-component of the centre of mass is equal to R, the radius
of the disk. In fact, the condition z = R is an example of a holonomic constraint,
but we will not go deeper into that matter here. Further, let ϕ be the angle of
the disk with the (x, z)-plane, and θ be the angle of a fixed line on the disk with a

Balkan Journal of Geometry and Its Applications, Vol.15, No.2, 2010, pp. 70-81.
c Balkan Society of Geometers, Geometry Balkan Press 2010.
Nonholonomic systems as restricted Euler-Lagrange systems 71
vertical line. To completely determine the motion of the disk one needs to know at
each instant the position of the centre of mass r
C
(t) and the amount the disk has
rotated from its initial position. This last quantity is completely determined by the
angular velocity ωωω, which is here just a superposition of two elementary rotations
ωωω =
˙
θe
φ
+ ˙ ϕe
z
. Here, e
φ
is a unit vector that lies in the direction perpendicular to
the vertical plane of the disk. Also the reaction force R
2
lies in that direction, i.e.
R
2
= ρe
ϕ
. With that, and the first of Euler’s laws, the first reaction force must be of
the form R
1
= ν
1
e
x
+ ν
2
e
y
+ Mge
z
. We can now write Euler’s laws in the following
equivalent fashion:

M¨r
C
= Mg +R
1
+R
2
,
d
dt
(I
C
(ωωω)) = (−Re
z
) ×R
1
.
When projected to the coordinate axes, these equations become
M¨ x = ν
1
−ρ sin(ϕ)
˙
θ,
M¨ y = ν
2
+ ρ cos(ϕ)
˙
θ,
I
¨
θ = −Rν
1
cos(ϕ) −Rν
2
sin(ϕ),
J ¨ ϕ = 0,
I
˙
θ ˙ ϕ = −Rν
2
cos(ϕ) + Rν
1
sin(ϕ).
From these equations, one wishes to determine (x(t), y(t), θ(t), ϕ(t)). However, the
dynamical evolution of the reaction forces ρ(t) and (ν
1
(t), ν
2
(t)) is unknown. We can
use the equation in
˙
θ ˙ ϕ to eliminate ρ from the picture: if we put λ
1
= ν
1
−ρ sin(ϕ)
˙
θ
and λ
2
= ν
2
+ ρ cos(ϕ)
˙
θ, the first equations become
(1.1) M¨ x = λ
1
, M¨ y = λ
2
, I
¨
θ = −Rλ
1
cos(ϕ) −Rλ
2
sin(ϕ), J ¨ ϕ = 0.
Once all variables have been determined, we can determine ρ from the equation ρR
˙
θ =
I
˙
θ ˙ ϕ + RM cos(ϕ)¨ y −RM sin(ϕ)¨ x.
Obviously we cannot solve equations (1.1) unless we assume an extra hypothesis
in the model. One typical type of extra assumption is the one where one assumes
that the disk rolls fast enough on the plane to prevent slipping: that is, one assumes
72 T. Mestdag and M. Crampin
that during the motion the velocity of the instantaneous contact point A vanishes,
˙ r
A
= 0, or, equivalently, ˙ r
C
= ωωω × AC, which, in the chosen coordinates amounts
to ˙ x = Rcos(ϕ)
˙
θ and ˙ y = Rsin(ϕ)
˙
θ. The assumption ‘rolling without slipping’ is
a typical example of a nonholonomic constraint. This means that it is a velocity-
dependent constraint which cannot be integrated to a constraint that depends only
on the position of the body.
With the extra assumption, we can easily eliminate the reaction forces (λ
1
, λ
2
)
from the equations. In the end, the equations one needs to solve are simply
(1.2)
¨
θ = 0, ¨ ϕ = 0, ˙ x = Rcos(ϕ)
˙
θ, ˙ y = Rsin(ϕ)
˙
θ.
This is a mixed set of first- and second-order ordinary differential equations. One
easily verifies that its solution set is given by θ(t) = u
θ
t +θ
0
and ϕ(t) = u
ϕ
t +ϕ
0
. If
u
ϕ
= 0, then the disk follows a circular path:
x(t) =

u
θ
u
ϕ

Rsin(ϕ(t)) + x
0
and y(t) = −

u
θ
u
ϕ

Rcos(ϕ(t)) + y
0
.
On the other hand, if u
ϕ
= 0, the disk evolves on a fixed line:
x(t) = Rcos(ϕ
0
)u
θ
t + x
0
and y(t) = Rsin(ϕ
0
)u
θ
t + y
0
.
Nonholonomic constraints arise naturally in the context of mechanical systems
with rigid bodies rolling without slipping over a surface. Another typical example is
the Chaplygin sleigh. This is a rigid body where one of the contact points with the
surface forms a knife edge. The nonholonomic constraint assumed is that there is
no motion perpendicular to the knife edge, or that the velocity of the contact point
remains in the direction of a fixed axis of the body. Typical engineering problems that
involve such constraints arise for example in robotics, where the wheels of a mobile
robot are often required to roll without slipping, or where one is interested in guiding
the motion of a cutting tool. Basic reference books on nonholonomic systems are
[1, 3, 6].
2 The vertical rolling disk as the restriction of a
Lagrangian system
The main question we wish to address in this section is the following. Can the
solutions of the nonholonomic problem of the vertical rolling disk be interpreted as
(part of the) solutions of the Euler-Lagrange equations,
d
dt


˜
L
∂ ˙ q
j



˜
L
∂q
j
= 0,
of a regular Lagrangian
˜
L in (q
i
) = (x, y, ϕ, θ)?
A Lagrangian
˜
L is regular if the matrix of functions


2
˜
L
∂ ˙ q
i
∂ ˙ q
j

Nonholonomic systems as restricted Euler-Lagrange systems 73
is everywhere non-singular. In that case, the Euler-Lagrange equations can be writ-
ten explicitly in the normal form of a system of second-order ordinary differential
equations
¨ q
i
= f
i
(q, ˙ q).
In what follows, we will always interpret the solutions of the Euler-Lagrange equations
as defining (base) integral curves of the second-order differential equations field
˜
Γ,
given by
˜
Γ = ˙ q
i

∂q
i
+ f
i
(q, ˙ q)

∂ ˙ q
i
.
Remark that this vector field is completely determined by the assumption that it is
a second-order differential equations field (i.e. that its coefficients along ∂/∂q
i
are
exactly the velocities ˙ q
i
) and by the equations
(2.1)
˜
Γ


˜
L
∂ ˙ q
i



˜
L
∂q
i
= 0.
The fact that, after eliminating the reaction forces from (1.1), we end up with a
mixed set of first-and second-order differential equations indicates that the equations
of motion of a nonholonomic system cannot be viewed an sich as the Euler-Lagrange
equations of some regular Lagrangian. In fact, the principle that governs nonholo-
nomic systems is rather an extended version of Hamilton’s principle. Consider a
mechanical system, with n generalized coordinates (q
i
), subject to forces that can
be derived from a potential V (q). If T(q, ˙ q) stands for the kinetic energy of the sys-
tem, the function L(q, ˙ q) = T(q, ˙ q) − V (q) is called the Lagrangian of the system.
Suppose that the system is subject to m additional nonholonomic constraints of the
form a
b
j
(q) ˙ q
j
= 0, b = 1, . . . , m < n. Then, the (extended) principle of Hamilton (see
e.g. [1]) postulates that the trajectory q(t) between times t
1
and t
2
is such that the
constraints are satisfied and that
δ

t
2
t
1
L(q(t), ˙ q(t))dt = 0,
for all variations satisfying a
b
j
δq
j
= 0. One easily demonstrates that these trajectories
are exactly the solutions of the equations

a
b
j
(q) ˙ q
j
= 0,
d
dt

∂L
∂ ˙ q
j


∂L
∂q
j
= λ
b
a
b
j
.
These equations define a system of n + m differential equations that can be solved
for the n + m unknown functions q
j
(t) and λ
a
(t). The terms λ
b
a
b
j
in the right-hand
side are related to the reaction forces. As before, the multipliers λ
a
can easily be
eliminated from the picture.
In case of the vertical rolling disk, the Lagrangian is (up to a constant)
(2.2) L = T =
1
2
M˙ r
2
C
+
1
2
I
C
(ωωω, ωωω) =
1
2
M( ˙ x
2
+ ˙ y
2
) +
1
2
I
˙
θ
2
+
1
2
J ˙ ϕ
2
.
74 T. Mestdag and M. Crampin
Next to the constraints, the equations of motion are therefore

d
dt

∂T
∂ ˙ x


∂T
∂x
= λ
1
,
d
dt

∂T
∂ ˙ y


∂T
∂y
= λ
2
,
d
dt

∂T

˙
θ


∂T
∂θ
= −λ
1
Rcos(ϕ) −λ
2
Rsin(ϕ),
d
dt

∂T
∂ ˙ ϕ


∂T
∂ϕ
= 0,
which is obviously equivalent to the system (1.1).
After the elimination of the unknown λ
a
we arrive at the mixed set of coupled
first- and second-order equations (1.2). There are, however, infinitely many systems
of second-order equations (only), whose solution set contains the solutions of the
nonholonomic equations. For example, the second-order system
(2.3)
¨
θ = 0, ¨ ϕ = 0, ¨ x = −Rsin(ϕ)
˙
θ ˙ ϕ, ¨ y = Rcos(ϕ)
˙
θ ˙ ϕ
has, for u
ϕ
= 0, the solutions θ(t) = u
θ
t + θ
0
, ϕ(t) = u
ϕ
t + ϕ
0
and
x(t) =

u
θ
u
ϕ

Rsin(ϕ(t)) + u
x
t + x
0
and y(t) = −

u
θ
u
ϕ

Rcos(ϕ(t)) + u
y
t + y
0
.
By restricting our attention only to those solutions for which ˙ x = cos(ϕ)
˙
θ and ˙ y =
sin(ϕ)
˙
θ (i.e. u
x
= u
y
= 0), we get back the solutions of the nonholonomic equations
(and similarly for solutions with u
ϕ
= 0). Some other examples of second-order
systems with a similar property are the systems
(2.4)
¨
θ = 0, ¨ ϕ = 0, ¨ x = −
sin(ϕ)
cos(ϕ)
˙ x ˙ ϕ, ¨ y =
cos(ϕ)
sin(ϕ)
˙ y ˙ ϕ
and
(2.5)
¨
θ = 0, ¨ ϕ = 0, ¨ x = −˙ y ˙ ϕ, ¨ y = ˙ x ˙ ϕ.
One can, of course, think of many more systems which show that behavior.
The question whether any of the above second-order systems is equivalent to a
variational system is an example to the so-called ‘inverse problem of the calculus of
variations’ (see e.g. [7]). From [2] we know that there is no regular Lagrangian for
the system (2.3) and that
˜
L =
1
2
˙ ϕ
2
+

I + MR
2
2

˙
θ
2
˙ ϕ
+
˙ x
2
cos(ϕ) ˙ ϕ
+
˙ y
2
sin(ϕ) ˙ ϕ

and
˜
L =
1
2
˙ ϕ
2
+
1
2
˙
θ
2


I + MR
2
2

˙ x
2
cos(ϕ) ˙ ϕ
+
˙ y
2
sin(ϕ) ˙ ϕ

Nonholonomic systems as restricted Euler-Lagrange systems 75
are both (independent) Lagrangians for the system (2.4). A result from [8] shows that
˜
L =
1
2
˙ ϕ
2
+
1
2
˙
θ
2
+
1
2 ˙ ϕ

( ˙ x
2
− ˙ y
2
) cos(ϕ) + 2 ˙ x ˙ y sin(ϕ)

is a regular Lagrangian for the third system.
This way of looking for Lagrangians has some serious disadvantages. First of all,
there are infinitely many of those ‘associated’ second-order systems. If we use the
methods of the inverse problem to decide whether one of them is not variational,
there is no guarantee that there will not be another one which is. Secondly, it is
extremely difficult to solve the inverse problem, even for a particular case. In most
cases, the success of finding a Lagrangian relies on making a number of educated
guesses.
It would therefore be better if there were a direct way to construct a Lagrangian
for a given nonholonomic system. It seems that such a construction method is at the
basis of an observation from Fernandez and Bloch in [5]. Among other things, they
show that the solution set of the Euler-Lagrange equations of the regular Lagrangian
˜
L = −
1
2
M( ˙ x
2
+ ˙ y
2
) +
1
2
I
˙
θ
2
+
1
2
J ˙ ϕ
2
+ MR
˙
θ(cos(ϕ) ˙ x + sin(ϕ) ˙ y)
is such that, when restricted to the constraints, it is the solution set of the nonholo-
nomic equations (1.1). This is easy to see. The Euler-Lagrange equations of
˜
L are
equivalent with
J ¨ ϕ = −MR[sin(ϕ) ˙ x −cos(ϕ) ˙ y]
˙
θ,
(I + MR
2
)
¨
θ = MR[sin(ϕ) ˙ x −cos(ϕ) ˙ y] ˙ ϕ,
(I + MR
2
)¨ x = −R(I + MR
2
) sin(ϕ)
˙
θ ˙ ϕ + MR
2
cos(ϕ)[sin(ϕ) ˙ x −cos(ϕ) ˙ y] ˙ ϕ,
(I + MR
2
)¨ y = R(I + MR
2
) cos(ϕ)
˙
θ ˙ ϕ + MR
2
sin(ϕ)[sin(ϕ) ˙ x −cos(ϕ) ˙ y] ˙ ϕ.
Given that sin(ϕ) ˙ x −cos(ϕ) ˙ y = 0 on the constraints, these equations become on the
constraints:
¨
θ = 0, ¨ ϕ = 0, ¨ x = −Rsin(ϕ)
˙
θ ˙ ϕ, ¨ y = Rcos(ϕ)
˙
θ ˙ ϕ,
which is the system (2.3) again. We had already shown that those solutions of (2.3)
which satisfy the constraints, are also solutions of the mixed system (1.2).
Contrary to the Lagrangians for the systems (2.4) and (2.5), the Lagrangian of
Fernandez and Bloch has a very suggestive form. If we set v
1
= ˙ x − Rcos(ϕ)
˙
θ and
v
2
= ˙ y −Rsin(ϕ)
˙
θ, the Lagrangian
˜
L becomes
(2.6)
˜
L = L −
∂L
∂ ˙ s
a
v
a
, s
a
= (x, y),
where L stands here for the nonholonomic Lagrangian (2.2) of the disk. This brings
some immediate questions to mind. How general is this phenomenon? And, what are
the conditions for it to occur for an arbitrary given nonholonomic system? The above
expression of the Lagrangian looks a bit like the Lagrangian one uses for vakonomic
systems. Therefore, it will be of interest, in one of the next sections, to see the relation
of this phenomenon with the theory of vakonomic systems. A basic reference for the
theory of vakonomic systems is [9].
76 T. Mestdag and M. Crampin
3 A formulation of the nonholonomic dynamics
using anholonomic frames
We denote by Q the configuration space and TQ, its tangent bundle, the velocity
phase space. We wish to interpret the solutions of the nonholonomic equations (1.1)
as the integral curves of a vector field Γ on TQ. Moreover, we will need an expression
of that vector field in terms of an anholonomic frame.
Let us first remind the reader that there are two canonical ways to lift a vector
field X = X
i
∂/∂q
i
on Q to one on TQ. The first lift is the complete lift
X
C
= X
i

∂q
i
+
∂X
i
∂q
j
˙ q
j

∂ ˙ q
i
,
which is a vector field whose flow consists of the tangent maps of the flow of X. The
second is the vertical lift
X
V
= X
i

∂ ˙ q
i
.
This vector field is tangent to the fibres of τ : TQ → Q and on a particular fibre T
q
Q
its value is constant and coincides with X(q).
If {X
i
} is a frame, that is a (possibly locally defined) basis of vector fields on Q,
then an equivalent expression for the equations (2.1) determining the Euler-Lagrange
field Γ is
Γ(X
V
i
(L)) −X
C
i
(L) = 0.
The frame {X
i
} is called anholonomic if the Lie brackets [X
i
, X
j
] = R
k
ij
X
k
do not
all vanish. Each frame {X
i
} defines a set of quasi-velocities v
i
for a tangent vector
v
q
∈ T
q
Q. They are the coefficients of the tangent vector with respect to the frame,
i.e. v
q
= v
i
X
i
(q). Later on, we will need expressions for the derivatives of the v
i
along X
C
i
and X
V
i
. In general, if [X
i
, X
j
] = R
k
ij
X
k
then
X
C
i
(v
j
) = −R
j
ik
v
k
and X
V
i
(v
j
) = δ
j
i
(see [4] for details).
Let us now come back to the situation of a mechanical system that is subject
to some nonholonomic constraints. The constraints define a distribution D on Q or,
equivalently, a submanifold C of TQ. Let us choose a frame {X
i
} = {X
α
, X
a
} of
vector fields on Q whose first m members {X
α
} span D. If we decompose a general
tangent vector v
q
as v
q
= v
α
X
α
(q) + v
a
X
a
(q) (so that the quasi-velocities are now
(v
α
, v
a
)), then the condition for it to lie in C is simply v
a
= 0. Moreover, a vector
field Γ on C is tangent to C if and only if Γ(v
a
) = 0. A vector field Γ on C is then of
second-order type (i.e. satisfies τ
∗(q,u)
Γ = u, for all (q, u) ∈ C) and is tangent to C if
it is of the form Γ = v
α
X
C
α
+ Γ
α
X
V
α
.
We say that L is regular with respect to D if the matrix of functions

X
V
α
(X
V
β
(L))

is nonsingular on C. The following statement now easily follows (see also [4]).
Proposition 1. If L is regular with respect to D there is a unique vector field Γ on
C which is of second-order type, is tangent to C, and is such that on C
(3.1) Γ(X
V
α
(L)) −X
C
α
(L) = 0.
Nonholonomic systems as restricted Euler-Lagrange systems 77
The vector field Γ obtained in this way defines the nonholonomic dynamics of the
constrained system. The non-zero functions
(3.2) λ
a
= Γ(X
V
a
(L)) −X
C
a
(L)
can be interpreted as the multipliers in this framework.
4 Nonholonomic systems as restricted
Euler-Lagrange systems
We now wish to come to an explication of why the construction (2.6) gives a La-
grangian for the example of the vertical rolling disk. As a first step we will establish
a set of conditions for the existence of a Lagrangian of a general form, of which the
one in (2.6) is a particular case, which has the required property. We continue to use
the notation of the previous section.
Proposition 2. If there are functions Φ
a
defined on a neighborhood of C in TQ such
that on C
Φ
a
R
a
αβ
v
β
= 0 and Γ(Φ
a
) + Φ
b
R
b

v
α
= λ
a
,
where the λ
a
are the multipliers, as given by the expression (3.2), then the nonholo-
nomic field Γ is the restriction to C of an Euler-Lagrange field of the Lagrangian
˜
L = L −Φ
a
v
a
.
Proof. We derive the Euler-Lagrange expressions Γ(X
V
i
(
˜
L))−X
C
i
(
˜
L) for Γ with respect
to the Lagrangian
˜
L = L − Φ
a
v
a
. Recall that X
C
i
(v
j
) = −R
j
ik
v
k
and X
V
i
(v
j
) = δ
j
i
.
We have
X
V
i
(
˜
L) = X
V
i
(L) −X
V
i

a
)v
a
−Φ
a
δ
a
i
,
X
C
i
(
˜
L) = X
C
i
(L) −X
C
i

a
)v
a
+ Φ
a
R
a
ij
v
j
,
whence on C (where v
a
= 0 and Γ(v
a
) = 0)
Γ(X
V
α
(
˜
L)) −X
C
α
(
˜
L) = Γ(X
V
α
(L)) −X
C
α
(L) −Φ
a
R
a
αβ
v
β
= −Φ
a
R
a
αβ
v
β
,
while
Γ(X
V
a
(
˜
L)) −X
C
a
(
˜
L) = Γ(X
V
a
(L)) −X
C
a
(L) −Γ(Φ
a
) −Φ
b
R
b

v
α
= λ
a
−Γ(Φ
a
) −Φ
b
R
b

v
α
.
Thus the necessary and sufficient conditions for Γ to satisfy the Euler-Lagrange equa-
tions of
˜
L on C are that the equations
Φ
a
R
a
αβ
v
β
= 0 and Γ(Φ
a
) + Φ
b
R
b

v
α
= λ
a
hold on C.
78 T. Mestdag and M. Crampin
Notice that since Γ is tangent to C these conditions depend only on the values of
the Φ
a
on C. Let us set Φ
a
|
C
= φ
a
. Then we could rewrite the conditions as
(4.1) φ
a
R
a
αβ
v
β
= 0 and Γ(φ
a
) + φ
b
R
b

v
α
= λ
a
,
and when they are satisfied the conclusion of the proposition holds for any extensions
Φ
a
of the φ
a
off C.
These conditions turn out to have an important role to play in the context of an-
other interesting problem concerning constrained systems, namely determining when
the dynamics defined in (3.1) agrees with that obtained from the so-called vakonomic
formulation of systems with nonholonomic constraints.
One way of introducing the vakonomic approach is to regard the multipliers as
additional variables. The multipliers may be regarded as the components of a 1-
form (along a certain projection) with values in D
0
⊂ T

Q (the annihilator of D).
Therefore, we take D
0
as state space for the vakonomic system. Once {X
α
, X
a
} have
been chosen we can identify D
0
locally with Q × R
n−m
. This is the same as saying
that we fix fibre coordinates µ
a
on D
0
.
The vakonomic Lagrangian
ˆ
L is the function on TD
0
= T(Q×R
n−m
) given by
(4.2)
ˆ
L = L −µ
a
v
a
.
This is in fact a singular Lagrangian, so there is no unique Euler-Lagrange field
ˆ
Γ.
One can easily verify that such a
ˆ
Γ can only exist on C ×TR
n−m
⊂ T(Q×R
n−m
). It
is therefore natural to decompose
ˆ
Γ according to this product structure as Γ
C
+ Γ
µ
.
With that, the Euler-Lagrange equations of
ˆ
L are of the form
Γ
C
(X
V
α
(L)) −X
C
α
(L) = µ
a
R
a
αβ
v
β
,
Γ
C
(X
V
a
(L)) −X
C
a
(L) = µ
b
R
b

v
α
+ Γ
µ

a
).
These equations will not be enough to determine both Γ
C
and Γ
µ
. We will regard the
Γ
µ

a
) (which are just the ∂/∂µ
a
components of Γ
µ
) as at our disposal: then once a
choice is made for Γ
µ
, in favorable circumstances the equations will determine Γ
C
.
Comparison of the Lagrangian (2.6) with the vakonomic Lagrangian (4.2) suggests
that the µ
a
should be thought of as functions on C. This will lead to an attempt to
fix Γ
C
in a natural way.
From now on, we assume that a section φ of C ×R
n−m
→ C is given, in the form
µ
a
= φ
a
for functions φ
a
on C, and we restrict things to im(φ). The Euler-Lagrange
equations above, when restricted to im(φ), become
Γ
C
(X
V
α
(L)) −X
C
α
(L) = φ
a
R
a
αβ
v
β
,
Γ
C
(X
V
a
(L)) −X
C
a
(L) = φ
b
R
b
aαβ
v
α
+ A
a
,
where we have written A
a
for the restriction of Γ
µ

a
) to im(φ). We have
ˆ
Γ|
im(φ)
=
Γ
C
+ Γ
µ
where now all coefficients are functions on C. Let us set
Γ
C
= v
α
X
C
α
+ Γ
α
X
V
α
+ Γ
a
X
V
a
.
Since Γ
C
is not necessarily tangent to C, the functions Γ
a
are not necessarily zero.
However, one can show that our freedom to choose A
a
can be used to ensure that the
Nonholonomic systems as restricted Euler-Lagrange systems 79
corresponding Γ
C
has Γ
a
= 0, provided a certain non-degeneracy condition holds. In
fact, if we set X
V
α
(X
V
β
(L)) = g
αβ
and so on, then there is a unique choice of A
a
such
that Γ
a
= 0 if (g
ab
− g
αβ
g

g

) is nonsingular on C. If that is the case, the vector
field
Γ
C
= v
α
X
C
α
+ Γ
α
X
V
α
can be determined from the equations
(4.3) Γ
C
(X
V
α
(L)) −X
C
α
(L) = φ
a
R
a
αβ
v
β
.
We may set
Γ
C
(X
V
a
(L)) −X
C
a
(L) = Λ
a
= φ
b
R
b
aαβ
v
α
+ A
a
.
Notice that
ˆ
Γ(µ
a
−φ
a
) = A
a
−Γ
C

a
) = Λ
a
−Γ
C

a
) −φ
b
R
b
aαβ
v
α
.
When we put these results together with Proposition 2 we obtain the following theo-
rem.
Theorem. Suppose that functions φ
a
on C can be found to satisfy
φ
a
R
a
αβ
v
β
= 0 and Γ(φ
a
) + φ
b
R
b

v
α
= λ
a
.
Then
1. Γ is the restriction to C of an Euler-Lagrange field of the Lagrangian
˜
L = L −
Φ
a
v
a
defined on a neighborhood of C in TQ for any functions Φ
a
such that
Φ
a
|
C
= φ
a
;
2. with an appropriate choice of Γ
µ
, the vector field
ˆ
Γ = Γ + Γ
µ
is an Euler-
Lagrange field of the vakonomic problem with Lagrangian L − µ
a
v
a
, which is
tangent to the section φ : µ
a
= φ
a
.
One kind of system for which this theory works in a particularly straightforward
way is the so-called Chaplygin system. For such systems, the Lagrangian is invariant
under the action of a Lie group G and the constraint distribution is a horizontal
distribution for the principal bundle Q → Q/G. We may then take the vector fields
{X
a
} of the frame to be the fundamental vector fields of the action, in which case
X
C
a
(L) = 0, and X
V
a
(L) = p
a
, the component of momentum corresponding to X
a
for
the unconstrained problem. The multiplier equation is just Γ(p
a
) = λ
a
. If in addition
we take the remaining vector fields {X
α
} of the frame to be G-invariant then R
i

= 0.
We then have a natural choice for φ, namely φ
a
= p
a
; with this choice the condition
Γ(φ
a
) +φ
b
R
b

v
α
= λ
a
is automatically satisfied, so in order for the conditions of the
theorem to be satisfied it is enough that p
a
R
a
αβ
v
β
= 0.
For more details, and some deeper analysis of the issues concerning the consistency
of nonholonomic and vakonomic dynamics, we refer to [4].
80 T. Mestdag and M. Crampin
5 The vertical rolling disk again
We conclude by showing explicitly how this theory applies to the example with which
we began.
In the special case of the vertical rolling disk, we can use the anholonomic frame
given by
{X
α
} =

X
1
=

∂ϕ
, X
2
=

∂θ
+ Rcos(ϕ)

∂x
+ Rsin(ϕ)

∂y

and {X
a
} =


∂x
,

∂y

.
The only non-vanishing bracket is then
[X
1
, X
2
] = −sin(ϕ)

∂x
+ cos(ϕ)

∂y
.
The quantities v
1
= ˙ x − Rcos(ϕ)
˙
θ and v
2
= ˙ y − Rsin(ϕ)
˙
θ introduced at the end
of Section 2 are the quasi-velocities v
a
for the given frame; the constraints are just
v
1
= v
2
= 0.
The vertical rolling disk is a typical example of a Chaplygin system. The Lie
group is simply R
2
and the action is given by translations in the direction of the
(x, y)-coordinates. As a consequence, ∂L/∂x = 0 and ∂L/∂y = 0. As we pointed
out above, it follows that λ
x
= Γ(∂L/∂ ˙ x) and λ
y
= Γ(∂L/∂ ˙ y). Since moreover all
R
b

= 0, a perfect candidate for a section φ is therefore simply
φ
x
=
∂L
∂ ˙ x
and φ
y
=
∂L
∂ ˙ y
.
With this section the conditions φ
a
R
a
αβ
v
β
= 0 in the theorem become

−M ˙ xsin(ϕ)
˙
θ + M ˙ y cos(ϕ)
˙
θ = 0,
M ˙ xsin(ϕ) ˙ ϕ −M ˙ y cos(ϕ) ˙ ϕ = 0.
They are clearly always satisfied on C. We can continue to use ∂L/∂ ˙ x and ∂L/∂ ˙ y for
the Φ
a
, and so obtain the Lagrangian (2.6).
This explains the observation about the vertical rolling disk in [5].
Acknowledgements. The first author is a Postdoctoral Fellow of the Research
Foundation – Flanders (FWO). This paper is the transcript of his talk at the Inter-
national Conference of Differential Geometry and Dynamical Systems (October 2009,
Bucharest). He would like to thank the organizers for the invitation. The second
author is a Guest Professor at Ghent University: he is grateful to the Department of
Mathematics at Ghent for its hospitality.
References
[1] A. M. Bloch, J. Baillieul, P. Crouch and J. E. Marsden, Nonholonomic Mechanics
and Control , Springer Verlag 2003.
[2] A. M. Bloch, O. E. Fernandez and T. Mestdag, Hamiltonization of nonholonomic
systems and the inverse problem of the calculus of variations, Rep. Math. Phys.
63 (2009), 225–249.
Nonholonomic systems as restricted Euler-Lagrange systems 81
[3] J. Cort´es Monforte, Geometric, Control and Numerical Aspects of Nonholonomic
Systems, Lecture Notes in Mathematics 1793, Springer Verlag 2002.
[4] M. Crampin and T. Mestdag, Anholonomic frames in constrained dynamics, to
appear in Dynamical Systems (2009). DOI: 10.1080/14689360903360888.
[5] O. E. Fernandez and A. M. Bloch, Equivalence of the dynamics of nonholonomic
and variational nonholonomic systems for certain initial conditions, J. Phys. A:
Math. Theor. 41 (2008), 344005.
[6] J. I. Ne˘ımark and N. A. Fufaev, Dynamics of Nonholonomic Systems, Transl. of
Math. Monographs 33, AMS 1972.
[7] R. M. Santilli, Foundations of Theoretical Mechanics I. The Inverse Problem in
Newtonian Mechanics, Springer Verlag 1978.
[8] G. Thompson, Variational connections on Lie groups, Diff. Geom. Appl. 18 (2003)
255–270.
[9] A.M. Vershik and V.Y. Gerschkovich, Nonholonomic dynamical systems,
geometry of distributions and variational problems, in: V.I. Arnold and S.P.
Novikov (eds.), Dynamical Systems VII , Encyclopedia of Mathematical Sciences
16, Springer Verlag 1994.
Authors’ address:
T. Mestdag and M. Crampin
Department of Mathematics, Ghent University,
Krijgslaan 281, B-9000 Ghent, Belgium.
E-mail: tom.mestdag@ugent.be, crampin@btinternet.com

However. This last quantity is completely determined by the angular velocity ω . Iθ ˙ y x Obviously we cannot solve equations (1. ¨ ¨ I θ = −Rν1 cos(ϕ) − Rν2 sin(ϕ). the first reaction force must be of the form R1 = ν1 ex + ν2 ey + M gez .Nonholonomic systems as restricted Euler-Lagrange systems 71 vertical line. which is here just a superposition of two elementary rotations ˙ ω = θeφ + ϕez . Iθ ˙ From these equations. Here. ν2 (t)) is unknown. One typical type of extra assumption is the one where one assumes that the disk rolls fast enough on the plane to prevent slipping: that is. one wishes to determine (x(t). one assumes . ϕ(t)). J ϕ = 0. dt When projected to the coordinate axes. R2 = ρeϕ . d ω (IC (ω )) = (−Rez ) × R1 . ¨ ¨ I θ = −Rλ1 cos(ϕ) − Rλ2 sin(ϕ). (1. eφ is a unit vector that lies in the direction perpendicular to ˙ the vertical plane of the disk. With that. Also the reaction force R2 lies in that direction. θ(t). To completely determine the motion of the disk one needs to know at each instant the position of the centre of mass rC (t) and the amount the disk has rotated from its initial position. We can ˙˙ ˙ use the equation in θϕ to eliminate ρ from the picture: if we put λ1 = ν1 − ρ sin(ϕ)θ ˙ the first equations become and λ2 = ν2 + ρ cos(ϕ)θ. ¨ ˙ M y = ν2 + ρ cos(ϕ)θ. J ϕ = 0. We can now write Euler’s laws in the following equivalent fashion: M¨C r = M g + R1 + R2 . ¨ M y = λ2 .1) M x = λ1 . ¨ ˙ Once all variables have been determined. y(t). i. the dynamical evolution of the reaction forces ρ(t) and (ν1 (t). these equations become ˙ M x = ν1 − ρ sin(ϕ)θ. we can determine ρ from the equation ρRθ = ˙ϕ + RM cos(ϕ)¨ − RM sin(ϕ)¨. ¨ ˙ϕ = −Rν2 cos(ϕ) + Rν1 sin(ϕ).1) unless we assume an extra hypothesis in the model.e. and the first of Euler’s laws.

The nonholonomic constraint assumed is that there is no motion perpendicular to the knife edge. ϕ = 0. which.72 T. or. d dt ˜ ∂L ∂ qj ˙ − ˜ ∂L = 0. θ)? ˜ A Lagrangian L is regular if the matrix of functions ˜ ∂2L ∂ qi ∂ qj ˙ ˙ . if uϕ = 0. equivalently. ¨ ˙ x = R cos(ϕ)θ. in the chosen coordinates amounts ˙ ˙ to x = R cos(ϕ)θ and y = R sin(ϕ)θ. or where one is interested in guiding the motion of a cutting tool. ∂q j ˜ of a regular Lagrangian L in (q i ) = (x. 2 The vertical rolling disk as the restriction of a Lagrangian system The main question we wish to address in this section is the following. In the end. This means that it is a velocitydependent constraint which cannot be integrated to a constraint that depends only on the position of the body. Crampin that during the motion the velocity of the instantaneous contact point A vanishes. Basic reference books on nonholonomic systems are [1. 6]. On the other hand.and second-order ordinary differential equations. we can easily eliminate the reaction forces (λ1 . Another typical example is the Chaplygin sleigh. or that the velocity of the contact point remains in the direction of a fixed axis of the body. 3. Typical engineering problems that involve such constraints arise for example in robotics. λ2 ) from the equations. Can the solutions of the nonholonomic problem of the vertical rolling disk be interpreted as (part of the) solutions of the Euler-Lagrange equations. ϕ. ˙ ˙ y = R sin(ϕ)θ. With the extra assumption. ˙ ˙ rA = 0. the equations one needs to solve are simply (1. Mestdag and M. Nonholonomic constraints arise naturally in the context of mechanical systems with rigid bodies rolling without slipping over a surface. y. where the wheels of a mobile robot are often required to roll without slipping. rC = ω × AC. One easily verifies that its solution set is given by θ(t) = uθ t + θ0 and ϕ(t) = uϕ t + ϕ0 . The assumption ‘rolling without slipping’ is ˙ ˙ a typical example of a nonholonomic constraint. ˙ This is a mixed set of first. If uϕ = 0. the disk evolves on a fixed line: x(t) = R cos(ϕ0 )uθ t + x0 and y(t) = R sin(ϕ0 )uθ t + y0 .2) ¨ θ = 0. then the disk follows a circular path: x(t) = uθ uϕ R sin(ϕ(t)) + x0 and y(t) = − uθ uϕ R cos(ϕ(t)) + y0 . This is a rigid body where one of the contact points with the surface forms a knife edge.

q) − V (q) is called the Lagrangian of the system. . Consider a mechanical system.2) L=T = 1 1 1 ˙ 1 1 ω ˙C M r2 + IC (ω . q). ω ) = M (x2 + y 2 ) + I θ2 + J ϕ2 . we will always interpret the solutions of the Euler-Lagrange equations ˜ as defining (base) integral curves of the second-order differential equations field Γ. . In fact. . b = 1. the principle that governs nonholonomic systems is rather an extended version of Hamilton’s principle. q) = T (q.  j j dt ∂ q ˙ ∂q These equations define a system of n + m differential equations that can be solved for the n + m unknown functions q j (t) and λa (t). [1]) postulates that the trajectory q(t) between times t1 and t2 is such that the constraints are satisfied and that t2 δ t1 L(q(t). .Nonholonomic systems as restricted Euler-Lagrange systems 73 is everywhere non-singular. ¨ ˙ In what follows. q) i . If T (q. with n generalized coordinates (q i ).g. ˙ ˙ Suppose that the system is subject to m additional nonholonomic constraints of the form ab (q)q j = 0. Then.1) ˜ Γ ˜ ∂L ∂ qi ˙ − ˜ ∂L = 0. the function L(q. The terms λb ab in the right-hand j side are related to the reaction forces. q) stands for the kinetic energy of the sys˙ tem. that its coefficients along ∂/∂q i are exactly the velocities q i ) and by the equations ˙ (2. the Lagrangian is (up to a constant) (2. subject to forces that can be derived from a potential V (q). given by ∂ ∂ ˜ Γ = q i i + f i (q. In that case.e. after eliminating the reaction forces from (1. the multipliers λa can easily be eliminated from the picture. ∂q i The fact that. we end up with a mixed set of first-and second-order differential equations indicates that the equations of motion of a nonholonomic system cannot be viewed an sich as the Euler-Lagrange equations of some regular Lagrangian. m < n. ˙ ˙ ˙ 2 2 2 2 2 . As before. d ∂L ∂L − j = λb ab .1). the Euler-Lagrange equations can be written explicitly in the normal form of a system of second-order ordinary differential equations q i = f i (q. q(t))dt = 0. the (extended) principle of Hamilton (see ˙ j e. ˙ ˙ ∂q ∂q ˙ Remark that this vector field is completely determined by the assumption that it is a second-order differential equations field (i. ˙ for all variations satisfying ab δq j = 0. In case of the vertical rolling disk. One easily demonstrates that these trajectories j are exactly the solutions of the equations  b ˙  aj (q) q j = 0.

infinitely many systems of second-order equations (only). From [2] we know that there is no regular Lagrangian for the system (2. For example. After the elimination of the unknown λa we arrive at the mixed set of coupled first. however. ˙       d ∂T  ∂T     dt ∂ y − ∂y = λ2 . ¨ x=− ¨ sin(ϕ) xϕ. ˙˙ cos(ϕ) y= ¨ cos(ϕ) yϕ ˙˙ sin(ϕ) One can. think of many more systems which show that behavior. ux = uy = 0). ˙  d     dt        d    dt ∂T ˙ ∂θ ∂T ∂ϕ ˙ − − ∂T = −λ1 R cos(ϕ) − λ2 R sin(ϕ).2). the equations of motion are therefore  ∂T  d ∂T    dt ∂ x − ∂x = λ1 . ¨ ˙˙ y = R cos(ϕ)θϕ ¨ has. ϕ(t) = uϕ t + ϕ0 and x(t) = uθ uϕ R sin(ϕ(t)) + ux t + x0 and y(t) = − uθ uϕ R cos(ϕ(t)) + uy t + y0 .e. ˙ By restricting our attention only to those solutions for which x = cos(ϕ)θ and y = ˙ ˙ ˙ (i. we get back the solutions of the nonholonomic equations sin(ϕ)θ (and similarly for solutions with uϕ = 0). The question whether any of the above second-order systems is equivalent to a variational system is an example to the so-called ‘inverse problem of the calculus of variations’ (see e.74 T. Some other examples of second-order systems with a similar property are the systems (2. for uϕ = 0. the solutions θ(t) = uθ t + θ0 . ¨ ˙˙ x = −R sin(ϕ)θϕ. y = xϕ. ϕ = 0. the second-order system (2. Crampin Next to the constraints. of course.3) and that √ ˜ L = 1 ϕ2 + 2 ˙ and ˙ ˜ L = 1 ϕ2 + 1 θ2 − 2 ˙ 2 I + M R2 2 √ ˙ θ2 x2 ˙ y2 ˙ + + ϕ ˙ cos(ϕ)ϕ sin(ϕ)ϕ ˙ ˙ x2 ˙ y2 ˙ + cos(ϕ)ϕ sin(ϕ)ϕ ˙ ˙ I + M R2 2 . There are. ϕ = 0. x = −y ϕ.5) ¨ θ = 0.g.and second-order equations (1. Mestdag and M. ∂ϕ which is obviously equivalent to the system (1. [7]).4) and (2. ϕ = 0. ¨ ¨ ˙˙ ¨ ˙˙ ¨ θ = 0. whose solution set contains the solutions of the nonholonomic equations.1).3) ¨ θ = 0. ∂θ ∂T = 0.

How general is this phenomenon? And. in one of the next sections.3) which satisfy the constraints. In most cases. It would therefore be better if there were a direct way to construct a Lagrangian for a given nonholonomic system. to see the relation of this phenomenon with the theory of vakonomic systems.2).4). ϕ = 0. ¨ which is the system (2. If we set v 1 = x − R cos(ϕ) θ and ˙ ˙ ˜ v 2 = y − R sin(ϕ) θ. what are the conditions for it to occur for an arbitrary given nonholonomic system? The above expression of the Lagrangian looks a bit like the Lagrangian one uses for vakonomic systems. there are infinitely many of those ‘associated’ second-order systems. the success of finding a Lagrangian relies on making a number of educated guesses. the Lagrangian of ˙ Fernandez and Bloch has a very suggestive form. the Lagrangian L becomes ˙ (2. Secondly. even for a particular case.5). Contrary to the Lagrangians for the systems (2. y). This is easy to see.6) ∂L ˜ L = L − a va . (I + M R )¨ = −R(I + M R ) sin(ϕ)θ ˙ x ˙ ˙ ˙ ˙˙ (I + M R2 )¨ = R(I + M R2 ) cos(ϕ)θϕ + M R2 sin(ϕ)[sin(ϕ)x − cos(ϕ)y]ϕ. y ˙ ˙ ˙ Given that sin(ϕ)x − cos(ϕ)y = 0 on the constraints. The Euler-Lagrange equations of L are equivalent with J ϕ = −M R[sin(ϕ)x − cos(ϕ)y]θ. it will be of interest.Nonholonomic systems as restricted Euler-Lagrange systems 75 are both (independent) Lagrangians for the system (2. It seems that such a construction method is at the basis of an observation from Fernandez and Bloch in [5].2) of the disk. where L stands here for the nonholonomic Lagrangian (2. ∂s ˙ sa = (x. ˙ ˙ ˙ 2 2 ˙ϕ + M R2 cos(ϕ)[sin(ϕ)x − cos(ϕ)y]ϕ. This way of looking for Lagrangians has some serious disadvantages. A basic reference for the theory of vakonomic systems is [9]. ¨ ˙ ˙ ˙ ¨ (I + M R2 )θ = M R[sin(ϕ)x − cos(ϕ)y]ϕ. Therefore. This brings some immediate questions to mind.3) again. First of all. . If we use the methods of the inverse problem to decide whether one of them is not variational. We had already shown that those solutions of (2. these equations become on the ˙ ˙ constraints: ¨ θ = 0. there is no guarantee that there will not be another one which is. when restricted to the constraints. it is extremely difficult to solve the inverse problem. ¨ ˙˙ y = R cos(ϕ)θϕ. ¨ ˙˙ x = −R sin(ϕ)θϕ. Among other things.1).4) and (2. A result from [8] shows that 1 1 ˙ ˜ L = 2 ϕ2 + 1 θ 2 + ˙ (x2 − y 2 ) cos(ϕ) + 2xy sin(ϕ) ˙ ˙ ˙˙ 2 2ϕ ˙ is a regular Lagrangian for the third system. are also solutions of the mixed system (1. they show that the solution set of the Euler-Lagrange equations of the regular Lagrangian 1 1 ˙ ˙ ˜ L = − 2 M (x2 + y 2 ) + 2 I θ2 + 1 J ϕ2 + M Rθ(cos(ϕ)x + sin(ϕ)y) ˙ ˙ ˙ ˙ 2 ˙ is such that. it is the solution set of the nonholo˜ nomic equations (1.

Let us choose a frame {Xi } = {Xα . we will need expressions for the derivatives of the v i C V k along Xi and Xi . a submanifold C of T Q. Mestdag and M. and is such that on C (3. then an equivalent expression for the equations (2. i. if [Xi . u) ∈ C) and is tangent to C if C V it is of the form Γ = v α Xα + Γα Xα .1) determining the Euler-Lagrange field Γ is V C Γ(Xi (L)) − Xi (L) = 0. Moreover. If we decompose a general tangent vector vq as vq = v α Xα (q) + v a Xa (q) (so that the quasi-velocities are now (v α . a vector field Γ on C is tangent to C if and only if Γ(v a ) = 0. We wish to interpret the solutions of the nonholonomic equations (1. Later on. The first lift is the complete lift XC = Xi ∂ ∂X i j ∂ + q ˙ .76 T. v a )).e. that is a (possibly locally defined) basis of vector fields on Q. is tangent to C. V V We say that L is regular with respect to D if the matrix of functions Xα (Xβ (L)) is nonsingular on C. we will need an expression of that vector field in terms of an anholonomic frame. for all (q. The second is the vertical lift ∂ XV = Xi i . its tangent bundle. vq = v i Xi (q). .e. then the condition for it to lie in C is simply v a = 0. ∂q ˙ This vector field is tangent to the fibres of τ : T Q → Q and on a particular fibre Tq Q its value is constant and coincides with X(q). Let us first remind the reader that there are two canonical ways to lift a vector field X = X i ∂/∂q i on Q to one on T Q. The following statement now easily follows (see also [4]). equivalently.1) as the integral curves of a vector field Γ on T Q. k The frame {Xi } is called anholonomic if the Lie brackets [Xi . Moreover. If L is regular with respect to D there is a unique vector field Γ on C which is of second-order type. If {Xi } is a frame. Crampin 3 A formulation of the nonholonomic dynamics using anholonomic frames We denote by Q the configuration space and T Q. Each frame {Xi } defines a set of quasi-velocities v for a tangent vector vq ∈ Tq Q. Let us now come back to the situation of a mechanical system that is subject to some nonholonomic constraints.1) V C Γ(Xα (L)) − Xα (L) = 0. Xj ] = Rij Xk do not i all vanish. satisfies τ∗(q. Xj ] = Rij Xk then j C Xi (v j ) = −Rik v k and j V Xi (v j ) = δi (see [4] for details). The constraints define a distribution D on Q or. They are the coefficients of the tangent vector with respect to the frame. the velocity phase space. Proposition 1. In general. Xa } of vector fields on Q whose first m members {Xα } span D. ∂q i ∂q j ∂ q i ˙ which is a vector field whose flow consists of the tangent maps of the flow of X. A vector field Γ on C is then of second-order type (i.u) Γ = u.

Thus the necessary and sufficient conditions for Γ to satisfy the Euler-Lagrange equa˜ tions of L on C are that the equations a b Φa Rαβ v β = 0 and Γ(Φa ) + Φb Raα v α = λa hold on C. while V ˜ C ˜ V C b Γ(Xa (L)) − Xa (L) = Γ(Xa (L)) − Xa (L) − Γ(Φa ) − Φb Raα v α b = λa − Γ(Φa ) − Φb Raα v α . C ˜ C C a Xi (L) = Xi (L) − Xi (Φa )v a + Φa Rij v j . If there are functions Φa defined on a neighborhood of C in T Q such that on C b a Φa Rαβ v β = 0 and Γ(Φa ) + Φb Raα v α = λa . As a first step we will establish a set of conditions for the existence of a Lagrangian of a general form. . which has the required property.6) is a particular case. where the λa are the multipliers.6) gives a Lagrangian for the example of the vertical rolling disk. We have V ˜ V V a Xi (L) = Xi (L) − Xi (Φa )v a − Φa δi . whence on C (where v a = 0 and Γ(v a ) = 0) V ˜ C ˜ V C a a Γ(Xα (L)) − Xα (L) = Γ(Xα (L)) − Xα (L) − Φa Rαβ v β = −Φa Rαβ v β . Proposition 2. of which the one in (2. as given by the expression (3.Nonholonomic systems as restricted Euler-Lagrange systems 77 The vector field Γ obtained in this way defines the nonholonomic dynamics of the constrained system. The non-zero functions (3.2). then the nonholonomic field Γ is the restriction to C of an Euler-Lagrange field of the Lagrangian ˜ L = L − Φa v a . We derive the Euler-Lagrange expressions Γ(Xi (L))−Xi (L) for Γ with respect j k j a C j V ˜ to the Lagrangian L = L − Φa v . 4 Nonholonomic systems as restricted Euler-Lagrange systems We now wish to come to an explication of why the construction (2.2) V C λa = Γ(Xa (L)) − Xa (L) can be interpreted as the multipliers in this framework. Recall that Xi (v ) = −Rik v and Xi (v j ) = δi . We continue to use the notation of the previous section. V ˜ C ˜ Proof.

78 T.2) ˆ L = L − µa v a . Xa } have been chosen we can identify D0 locally with Q × Rn−m . ˆ With that. so there is no unique Euler-Lagrange field Γ. This is the same as saying that we fix fibre coordinates µa on D0 . namely determining when the dynamics defined in (3. Therefore. ˆ where we have written Aa for the restriction of Γµ (µa ) to im(φ).2) suggests that the µa should be thought of as functions on C. The Euler-Lagrange equations above. Mestdag and M. and we restrict things to im(φ). ˆ can only exist on C × T Rn−m ⊂ T (Q × Rn−m ). we assume that a section φ of C × Rn−m → C is given. Comparison of the Lagrangian (2. and when they are satisfied the conclusion of the proposition holds for any extensions Φa of the φa off C. in favorable circumstances the equations will determine ΓC . From now on. We have Γ|im(φ) = ΓC + Γµ where now all coefficients are functions on C. These conditions turn out to have an important role to play in the context of another interesting problem concerning constrained systems. the Euler-Lagrange equations of L are of the form V C a ΓC (Xα (L)) − Xα (L) = µa Rαβ v β . Let us set C V V ΓC = v α Xα + Γα Xα + Γa Xa . V C b ΓC (Xa (L)) − Xa (L) = φb Raαβ v α + Aa . in the form µa = φa for functions φa on C. ˆ This is in fact a singular Lagrangian.6) with the vakonomic Lagrangian (4. It One can easily verify that such a Γ ˆ is therefore natural to decompose Γ according to this product structure as ΓC + Γµ . Since ΓC is not necessarily tangent to C. The multipliers may be regarded as the components of a 1form (along a certain projection) with values in D0 ⊂ T ∗ Q (the annihilator of D). One way of introducing the vakonomic approach is to regard the multipliers as additional variables. However. This will lead to an attempt to fix ΓC in a natural way. Crampin Notice that since Γ is tangent to C these conditions depend only on the values of the Φa on C. we take D0 as state space for the vakonomic system. We will regard the Γµ (µa ) (which are just the ∂/∂µa components of Γµ ) as at our disposal: then once a choice is made for Γµ . Then we could rewrite the conditions as (4. when restricted to im(φ). These equations will not be enough to determine both ΓC and Γµ . Let us set Φa |C = φa . b C V ΓC (Xa (L)) − Xa (L) = µb Raα v α + Γµ (µa ). the functions Γa are not necessarily zero. one can show that our freedom to choose Aa can be used to ensure that the .1) agrees with that obtained from the so-called vakonomic formulation of systems with nonholonomic constraints. ˆ The vakonomic Lagrangian L is the function on T D0 = T (Q × Rn−m ) given by (4. Once {Xα .1) a φa Rαβ v β = 0 b and Γ(φa ) + φb Raα v α = λa . become V C a ΓC (Xα (L)) − Xα (L) = φa Rαβ v β .

so in order for the conditions of the a theorem to be satisfied it is enough that pa Rαβ v β = 0. For such systems. We then have a natural choice for φ. Γ is the restriction to C of an Euler-Lagrange field of the Lagrangian L = L − Φa v a defined on a neighborhood of C in T Q for any functions Φa such that Φa |C = φa . and some deeper analysis of the issues concerning the consistency of nonholonomic and vakonomic dynamics.Nonholonomic systems as restricted Euler-Lagrange systems 79 corresponding ΓC has Γa = 0. Then ˜ 1. In V V fact. the component of momentum corresponding to Xa for the unconstrained problem. V C a ΓC (Xα (L)) − Xα (L) = φa Rαβ v β . with this choice the condition b Γ(φa ) + φb Raα v α = λa is automatically satisfied.3) We may set V C b ΓC (Xa (L)) − Xa (L) = Λa = φb Raαβ v α + Aa . we refer to [4]. For more details. the vector field C V ΓC = v α Xα + Γα Xα can be determined from the equations (4. Notice that b ˆ Γ(µa − φa ) = Aa − ΓC (φa ) = Λa − ΓC (φa ) − φb Raαβ v α . Suppose that functions φa on C can be found to satisfy a φa Rαβ v β = 0 and b Γ(φa ) + φb Raα v α = λa . in which case C V Xa (L) = 0. If that is the case. The multiplier equation is just Γ(pa ) = λa . ˆ 2. and Xa (L) = pa . with an appropriate choice of Γµ . namely φa = pa . which is tangent to the section φ : µa = φa . We may then take the vector fields {Xa } of the frame to be the fundamental vector fields of the action. Theorem. One kind of system for which this theory works in a particularly straightforward way is the so-called Chaplygin system. provided a certain non-degeneracy condition holds. If in addition i we take the remaining vector fields {Xα } of the frame to be G-invariant then Raα = 0. then there is a unique choice of Aa such that Γa = 0 if (gab − g αβ gaα gbβ ) is nonsingular on C. . the Lagrangian is invariant under the action of a Lie group G and the constraint distribution is a horizontal distribution for the principal bundle Q → Q/G. if we set Xα (Xβ (L)) = gαβ and so on. When we put these results together with Proposition 2 we obtain the following theorem. the vector field Γ = Γ + Γµ is an EulerLagrange field of the vakonomic problem with Lagrangian L − µa v a .

63 (2009). E. ∂y ˙ a With this section the conditions φa Rαβ v β = 0 in the theorem become ˙ ˙ −M x sin(ϕ)θ + M y cos(ϕ)θ = 0. We can continue to use ∂L/∂ x and ∂L/∂ y for ˙ ˙ the Φa . ∂L/∂x = 0 and ∂L/∂y = 0. E. ˙ ˙ M x sin(ϕ)ϕ − M y cos(ϕ)ϕ = 0. Nonholonomic Mechanics and Control . As we pointed out above. the constraints are just v 1 = v 2 = 0. Phys. Acknowledgements. J. ˙ ˙ ˙ ˙ They are clearly always satisfied on C. The vertical rolling disk is a typical example of a Chaplygin system. M. Rep. Mestdag. Baillieul. The second author is a Guest Professor at Ghent University: he is grateful to the Department of Mathematics at Ghent for its hospitality. Mestdag and M. y)-coordinates. M. ∂x ∂y ˙ ˙ The quantities v 1 = x − R cos(ϕ) θ and v 2 = y − R sin(ϕ) θ introduced at the end ˙ ˙ a of Section 2 are the quasi-velocities v for the given frame. Springer Verlag 2003. we can use the anholonomic frame given by {Xα } = X1 = ∂ ∂ ∂ ∂ . Bloch. Math. it follows that λx = Γ(∂L/∂ x) and λy = Γ(∂L/∂ y). X2 ] = − sin(ϕ) ∂ ∂ + cos(ϕ) . This paper is the transcript of his talk at the International Conference of Differential Geometry and Dynamical Systems (October 2009. He would like to thank the organizers for the invitation.6). [2] A. The only non-vanishing bracket is then [X1 . Bloch. Crouch and J. Bucharest). This explains the observation about the vertical rolling disk in [5]. and so obtain the Lagrangian (2. Marsden. .80 T. The Lie group is simply R2 and the action is given by translations in the direction of the (x. In the special case of the vertical rolling disk. ∂x ∂y . The first author is a Postdoctoral Fellow of the Research Foundation – Flanders (FWO). References [1] A. Since moreover all ˙ ˙ b Raα = 0. As a consequence. Fernandez and T. O. 225–249. a perfect candidate for a section φ is therefore simply φx = ∂L ∂x ˙ and φy = ∂L . Hamiltonization of nonholonomic systems and the inverse problem of the calculus of variations. Crampin 5 The vertical rolling disk again We conclude by showing explicitly how this theory applies to the example with which we began. P. X2 = + R cos(ϕ) + R sin(ϕ) ∂ϕ ∂θ ∂x ∂y and {Xa } = ∂ ∂ .

Mestdag. A: Math. [6] J. The Inverse Problem in Newtonian Mechanics. Dynamics of Nonholonomic Systems. A. Novikov (eds. of Math. Krijgslaan 281. Ghent University. Arnold and S. E. Encyclopedia of Mathematical Sciences 16. M. DOI: 10.com . Monographs 33.I. [8] G.P. Springer Verlag 2002. [9] A. Theor. Ne˘ ımark and N. Belgium. Springer Verlag 1994. Authors’ address: T.). J.Y. B-9000 Ghent. crampin@btinternet. Springer Verlag 1978. [7] R. Fernandez and A. 18 (2003) 255–270. Anholonomic frames in constrained dynamics. Lecture Notes in Mathematics 1793. Geom. to appear in Dynamical Systems (2009). Foundations of Theoretical Mechanics I. 344005. Crampin Department of Mathematics.Nonholonomic systems as restricted Euler-Lagrange systems 81 [3] J. [4] M. Thompson. Dynamical Systems VII . Diff. geometry of distributions and variational problems. in: V. Fufaev. Crampin and T. M.M. Control and Numerical Aspects of Nonholonomic e Systems. Bloch. [5] O. AMS 1972. Cort´s Monforte.mestdag@ugent. Transl. 41 (2008). Appl. Phys.1080/14689360903360888. Equivalence of the dynamics of nonholonomic and variational nonholonomic systems for certain initial conditions.be. Mestdag and M. Geometric. Variational connections on Lie groups. I. Nonholonomic dynamical systems. E-mail: tom. Gerschkovich. Santilli. Vershik and V.