The Hamiltonian and Lagrangian Approaches To The Dynamics of Nonholonomic Systems
The Hamiltonian and Lagrangian Approaches To The Dynamics of Nonholonomic Systems
The Hamiltonian and Lagrangian Approaches To The Dynamics of Nonholonomic Systems
Nonholonomic Systems
Wang Sang Koon
Department of Mathematics
University of California
Berkeley, CA 94720-3840
[email protected]
Jerrold E. Marsden
, s
a
) where = 1, . . . n k, in which
a
has the form
a
(q) = ds
a
+A
a
(r, s)dr
(2.1.3)
where the summation convention is in force. In other words, we are locally writing the distribution
as
T = (r, s, r, s) TQ [ s +A
a
= 0.
The equations of motion (2.1.1) may be rewritten by noting that the allowed variations q
i
=
(r
, s
a
) satisfy s
a
+A
a
L
r
_
= A
a
_
d
dt
L
s
a
L
s
a
_
. (2.1.4)
Equation (2.1.4) combined with the constraint equations
s
a
= A
a
(2.1.5)
gives a complete description of the equations of motion of the system; this procedure may be viewed
as one way of eliminating the Lagrange multipliers. Using this notation, one nds that =
a
a
,
where
a
=
d
dt
L
s
a
L
s
a
.
Equations (2.1.4) can be written in the following way:
d
dt
L
c
r
L
c
r
+A
a
L
c
s
a
=
L
s
b
B
b
, (2.1.6)
where
L
c
(r
, s
a
, r
) = L(r
, s
a
, r
, A
a
(r, s) r
).
is the coordinate expression of the constrained Lagrangian dened by L
c
= L[T and where
B
b
=
_
A
b
A
b
+A
a
A
b
s
a
A
a
A
b
s
a
_
. (2.1.7)
Letting d
b
be the exterior derivative of
b
, a computation shows that
d
b
( q, ) = B
b
dr
4
and hence the equations of motion have the form
L
c
=
_
d
dt
L
c
r
L
c
r
+A
a
L
c
s
a
_
r
=
L
s
b
d
b
( q, r).
This form of the equations isolates the eects of the constraints, and shows, in particular, that in
the case where the constraints are integrable (i.e., d = 0), the equations of motion are obtained
by substituting the constraints into the Lagrangian and then setting the variation of L
c
to zero.
However in the non-integrable case the constraints generate extra (curvature) terms, which must
be taken into account.
The above coordinate results can be put into an interesting and useful intrinsic geometric
framework. The intrinsically given information is the distribution and the Lagrangian. Assume
that there is a bundle structure
Q,R
: Q R for our space Q, where R is the base manifold and
Q,R
is a submersion and the kernel of T
q
Q,R
at any point q Q is called the vertical space V
q
.
One can always do this locally. An Ehresmann connection A is a vertical valued one form on Q
such that
1. A
q
: T
q
Q V
q
is a linear map and
2. A is a projection: A(v
q
) = v
q
for all v
q
V
q
.
Hence, T
q
Q = V
q
H
q
where H
q
= kerA
q
is the horizontal space at q, sometimes denoted hor
q
.
Thus, an Ehresmann connection gives us a way to split the tangent space to Q at each point into
a horizontal and vertical part.
If the Ehresmann connection is chosen in such a way that the given constraint distribution
T is the horizontal space of the connection; that is, H
q
= T
q
, then in the bundle coordinates
q
i
= (r
, s
a
), the map
Q,R
is just projection onto the factor r and the connection A can be
represented locally by a vector valued dierential form
a
:
A =
a
s
a
,
a
(q) = ds
a
+A
a
(r, s)dr
,
and the horizontal projection is the map
( r
, s
a
) ( r
, A
a
(r, s) r
). (2.1.8)
The curvature of an Ehresmann connection A is the vertical valued two form dened by its
action on two vector elds X and Y on Q as B(X, Y ) = A([hor X, hor Y ]) where the bracket
on the right hand side is the Jacobi-Lie bracket of vector elds obtained by extending the stated
vectors to vector elds. This denition shows the sense in which the curvature measures the failure
of the constraint distribution to be integrable.
In coordinates, one can evaluate the curvature B of the connection A by the following formula:
B(X, Y ) = d
a
(hor X, hor Y )
s
a
,
so that the local expression for curvature is given by
B(X, Y )
a
= B
a
(2.1.9)
5
where the coecients B
a
,
L
c
r
d
dt
L
c
r
A
a
L
c
s
a
_
.
As shown in [BKMM], when there is a symmetry group G present, there is a natural bundle
one can work with and put a connection on, namely the bundle Q Q/G. In the generality of
the preceding discussion, one can get away with just the distribution itself and can introduce the
corresponding Ehresmann connection locally. In fact, the bundle structure Q R is really a red
Herring. The notion of curvature as a T
q
Q/T
q
valued form makes good sense and is given locally
by the same expressions as above. However, keeping in mind that we eventually want to deal with
symmetries and in that case there is a natural bundle, the Ehresmann assumption is nevertheless
a reasonable bridge to the more interesting case with symmetries.
2.2 Review of the Hamiltonian Formulation
The approach of [BS] starts on the Lagrangian side with a conguration space Q and a Lagrangian
L of the form kinetic energy minus potential energy, i.e.,
L(q, q) =
1
2
q, q)) V (q),
where , )) is a metric on Q dening the kinetic energy and V is a potential energy function. We
do not restrict ourselves to Lagrangians of this form.
As above, our nonholonomic constraints are given by a distribution T TQ. We also let
T
o
T
a
i
,
where i = 1, . . . , n, together with the constraint equations.
6
The preceding Hamiltonian equations can be rewritten as
X = dH +
a
a
, (2.2.1)
where X is the vector eld on T
Q, and
Q
: T
q
i + p
i
p
i
.
On Lagrangian side, we saw that one can get rid of the Lagrangian multipliers. On the Hamilto-
nian side, it is also desirable to model the Hamiltonian equations without the Lagrange multipliers
by a vector eld on a submanifold of T
Q.
Using a basis
a
of the annihilator T
o
, we can write these spaces as
/ = p T
Q [
a
((FL)
1
(p)) = 0, (2.2.2)
and
T = u TT
Q [
a
, u
_
= 0. (2.2.3)
Finally, we dene
H = T T/.
Using natural coordinates (q
i
, p
i
, q
i
, p
i
) on TT
Q berwise to the
distribution H of the constraint submanifold /, is nondegenerate. Note that
H
is not a true two
form on a manifold, so it does not make sense to speak about it being closed. We speak of it as a
ber-restricted two form to avoid any confusion. Of course it still makes sense to talk about it being
nondegenerate; it just means nondegenerate as a bilinear form on each ber of H. The dynamics
is then given by the vector eld X
H
on / which takes values in the constraint distribution H and
is determined by the condition
X
H
H
= dH
H
(2.2.4)
where dH
H
is the restriction of dH
M
to H. We will be exploring the coordinate meaning of this
condition and its comparison with the Lagrangian formulation in the subsequent sections.
2.3 Lagrangian Side
We now construct the geometric structures on the tangent bundle TQ corresponding to those on
the Hamiltonian side from the preceding subsection and formulate a similar procedure for obtaining
7
the equations of motion. By doing this, it will be easier to made comparison with the geometric
constructions and analytic formulations in [BKMM].
First of all, we can dene the energy function E simply as E = H FL and pull back to TQ
the canonical two-form on T
Q and denote it by
L
.
We dene the distribution ( = (T
Q
)
1
(T) TTQ, where
Q
: TQ Q. In coordinates, the
distribution ( consists of vectors annihilated by the form
a
:
( = u TTQ [
a
, u
_
= 0. (2.3.1)
When ( is restricted to the constraint submanifold T TQ, we obtain the constraint distribution
/:
/ = ( TT. (2.3.2)
Clearly / = FL(T) and H = TFL(/).
The dynamics is given by a vector eld X
K
on the manifold T which takes values in / and
satises the equation
X
K
K
= dE
K
, (2.3.3)
where dE
K
and
K
are the restrictions of dE
D
and
D
respectively to the distribution / and where
E
D
and
D
are the restrictions of E and
L
to T.
2.4 The Equivalence of the Hamiltonian and the Lagrange-dAlembert Formu-
lations
The Lagrangian procedure on TQ formulated in the preceding subsection acts as a bridge between
[BS] and [BKMM]. We can show the correctness of the Lagrangian procedure given above by
(carefully) invoking the results of [BS] (generalized to arbitrary Lagrangians and with some gaps
lled in), or by checking the methods against the results of [BKMM]. We choose the latter method.
Theorem 2.1 Consider a conguration space Q, a hyperregular Lagrangian L and a distribution
T that describes the kinematic nonholonomic constraints. The /-valued vector eld X
K
on T given
by the equation
X
K
K
= dE
K
(2.4.1)
denes dynamics that are equivalent to the Lagrange-dAlembert equations together with the con-
straints.
Proof Consider the following form of the equations: X
H
M
= dH
M
on H; that is,
X
H
M
, u) = dH
M
, u) ,
for all u H. If we rewrite this in the form dH
M
X
H
M
, u) = 0, then on the Lagrangian
side, this is nothing but
dE
D
X
K
(
L
)
D
, v) = 0,
where v /. With appropriate interpretations, this is equivalent to Lagrange-dAlembert principle:
_
d
dt
L
q
i
L
q
i
_
(q
i
) = 0
a
( q) = 0
where (q) = 0.
8
Remark: A proof done in coordinates can be found in Koon [1997]. It provides a concrete
coordinate based procedure for nding the equations of motion on the Hamiltonian side and may
make the examples easier to follow.
2.5 Example: The Snakeboard
The snakeboard is a modied version of a skateboard in which the front and back pairs of wheels
are independently actuated. The extra degree of freedom enables the rider to generate forward
motion by twisting their body back and forth, while simultaneously moving the wheels with the
proper phase relationship. For details, see [BKMM] and the references listed there. Here we shall
include some of the computations shown in that paper both for completeness as well as to make
concrete the nonholonomic theory.
The snakeboard is modeled as a rigid body (the board) with two sets of independently actuated
wheels, one on each end of the board. The human rider is modeled as a momentum wheel which sits
in the middle of the board and is allowed to spin about the vertical axis. Spinning the momentum
wheel causes a counter-torque to be exerted on the board. The conguration of the board is given
by the position and orientation of the board in the plane, the angle of the momentum wheel, and
the angles of the back and front wheels. Let (x, y, ) represent the position and orientation of the
center of the board, the angle of the momentum wheel relative to the board, and
1
and
2
the angles of the back and front wheels, also relative to the board. Take the distance between the
center of the board and the wheels to be r. See gure 2.5.1.
(x, y)
r
Figure 2.5.1: The geometry of the snakeboard.
In [BKMM], a simplication is made which we shall also assume in this paper, namely
1
=
2
,
J
1
= J
2
. The parameters are also chosen such that J + J
0
+ J
1
+ J
2
= mr
2
, where m is the total
mass of the board, J is the inertia of the board, J
0
is the inertia of the rotor and J
1
, J
2
are the
inertia of the wheels. This simplication eliminates some terms in the derivation but does not
aect the essential geometry of the problem. Setting =
1
=
2
, then the conguration space
becomes Q = SE(2) S
1
S
1
and the Lagrangian L : TQ R is the total kinetic energy of the
system and is given by
L =
1
2
m( x
2
+ y
2
) +
1
2
mr
2
2
+
1
2
J
0
2
+J
0
+J
1
2
.
The Constraints. The rolling of the front and rear wheels of the snakeboard is modeled using
nonholonomic constraints which allow the wheels to spin about the vertical axis and roll in the
9
direction that they are pointing. The wheels are not allowed to slide in the sideways direction. The
constraints are dened by
sin( +) x + cos( +) y r cos
= 0 (2.5.1)
sin( ) x + cos( ) y +r cos
= 0 (2.5.2)
and can be simplied as
x = r cot cos
y = r cot sin
.
Since the constrained Legendre transform FL[T on the constraint submanifold T and its inverse
are given by
p
x
= mr cot cos
p
y
= mr cot sin
= mr
2
+J
0
= J
0
+J
0
= 2J
1
x =
r
mr
2
J
0
cot cos (p
)
y =
r
mr
2
J
0
cot sin (p
=
p
mr
2
J
0
=
mr
2
p
J
0
p
J
0
(mr
2
J
0
)
=
p
2J
1
,
the constraint submanifold / is dened by
/ = (x, y, , , , p
x
, p
y
, p
, p
, p
) [
p
x
=
mr
mr
2
J
0
cot cos (p
), p
y
=
mr
mr
2
J
0
cot sin (p
).
Notice that / may be thought of as a graph in T
, p
, p
,
p
,
p
,
p
.
The Hamiltonian. The corresponding Hamiltonian is given via the Legendre transform by
H =
1
2m
(p
2
x
+p
2
y
) +
1
2J
0
p
2
+
1
2(mr
2
J
0
)
(p
)
2
+
1
4J
1
p
2
.
10
Now if we restrict the Hamiltonian H to the submanifold /, we get
H
M
=
mr
2
2(mr
2
J
0
)
2
cot
2
(p
)
2
+
1
2J
0
p
2
+
1
2(mr
2
J
0
)
(p
)
2
+
1
4J
1
p
2
.
After computing its dierential dH
M
and restricting it to H, we have
dH
H
=
mr
2
(mr
2
J
0
)
2
cot csc
2
(p
)
2
d +
mr
2
(mr
2
J
0
)
2
cot
2
(p
)(dp
dp
)
+
1
J
0
p
dp
+
1
(mr
2
J
0
)
(p
)(dp
dp
) +
1
2J
1
p
dp
.
The Two Form. After pulling back the canonical two-form of T
Q to /, we have
M
= dx dp
x
+dy dp
y
+d dp
+d dp
+d dp
= kdx [csc
2
cos (p
)d + cot sin (p
dp
)]
+kdy [csc
2
sin (p
)d cot cos (p
dp
)]
+d dp
+d dp
+d dp
,
where k = mr/(mr
2
J
0
). If we restrict
M
to the distribution H, we get
H
= kr cot cos d
[csc
2
cos (p
)d + cot sin (p
dp
)]
kr cot sin d
[csc
2
sin (p
)d cot cos (p
dp
)]
+d dp
+d dp
+d dp
)d +kr cot
2
(dp
dp
) +dp
] +d dp
+d dp
.
The Equations of Motion. Notice that any vector eld X
M
is of the form
X
M
= x
x
+ y
y
+
+ p
+ p
+ p
.
But X
H
also lies in H = kerdx+r cot cos d, dy +r cot sin d and hence must be of the form
X
H
=
(r cot cos
x
r cot sin
y
+
) +
+ p
+ p
+ p
,
which gives us the rst set of relationships
x = r cot cos
y = r cot sin
.
Moreover,
X
H
H
= kr cot csc
2
(p
d +kr cot
2
(dp
dp
) +
dp
+
dp
)
d +
dp
kr cot
2
p
d p
d
+ kr cot
2
p
d p
d p
d,
11
and if equated with dH
H
and after simplication, we have
p
=
cot
2J
1
(1
J
0
mr
2
sin
2
)
p
(p
) (2.5.3)
p
= 0 (2.5.4)
p
= 0 (2.5.5)
=
p
mr
2
J
0
(2.5.6)
=
mr
2
p
J
0
p
J
0
(mr
2
J
0
)
(2.5.7)
=
p
2J
1
. (2.5.8)
Notice that the last 3 equations are nothing but the inverse of the constrained Legendre trans-
formation FL[T written in local coordinates. The rst equation is equivalent to the momentum
equation (discussed below and in [BKMM]) written in Hamiltonian form and the 2nd and 3rd
equations are the reduced equations on the shape space, again in their Hamiltonian forms.
Moreover, the corresponding Lagrangian procedure gives the equations of the motion on the
Lagrangian side as
cot
+
J
0
mr
2
sin
2
= 0 (2.5.9)
J
0
+J
0
= 0 (2.5.10)
2J
1
= 0 (2.5.11)
and it can be shown that both systems of equations are equivalent via the Legendre transform
FL[T.
3 Nonholonomic Mechanical Systems with Symmetry
Now we add the hypothesis of symmetry to the preceding development. Assume that we have a
conguration manifold Q, a Lagrangian of the form kinetic minus potential, and a distribution T
that describes the kinematic nonholonomic constraints. We also assume there is a symmetry group
G (a Lie group) that leaves the Lagrangian invariant, and that acts on Q (by isometries) and also
leaves the distribution invariant, i.e., the tangent of the group action maps T
q
to T
gq
(for more
details, see [BKMM].) Later, we shall refer this as a simple nonholonomic mechanical system.
3.1 Review of Lagrangian Reduction
We rst recall how [BKMM] explains in general terms how one constructs reduced systems by
eliminating the group variables.
Proposition 3.1 Under the assumptions that both the Lagrangian L and the distribution T are
G-invariant, we can form the reduced velocity phase space TQ/G and the constrained re-
duced velocity phase space T/G. The Lagrangian L induces well dened functions, the reduced
Lagrangian
l : TQ/G R
12
satisfying L = l
TQ
where
TQ
: TQ TQ/G is the projection, and the constrained reduced
Lagrangian
l
c
: T/G R,
which satises L[T = l
c
D
where
D
: T T/G is the projection. Also, the Lagrange-dAlembert
equations induce well dened reduced Lagrange-dAlembert equations on T/G. That is, the
vector eld on the manifold T determined by the Lagrange-dAlembert equations (including the
constraints) is G-invariant, and so denes a reduced vector eld on the quotient manifold T/G.
This proposition follows from general symmetry considerations, but to compute the associated
reduced equations explicitly and to reconstruct the group variables, one denes the nonholonomic
momentum map J
nh
, and extends the NoetherTheorem to nonholonomic system and synthesizes,
out of the mechanical connection and the Ehresmann connection, a nonholonomic connection /
nh
which is a connection on the principal bundle Q Q/G.
The Nonholonomic Momentum Map. Let the intersection of the tangent to the group orbit
and the distribution at a point q Q be denoted
o
q
= T
q
T
q
(Orb(q)).
Dene, for each q Q, the vector subspace g
q
to be the set of Lie algebra elements in g whose
innitesimal generators evaluated at q lie in S
q
:
g
q
= g [
Q
(q) o
q
.
We let g
D
denote the corresponding bundle over Q whose ber at the point q is given by g
q
. The
nonholonomic momentum map J
nh
is the bundle map taking TQ to the bundle (g
D
)
(whose ber
over the point q is the dual of the vector space g
q
) that is dened by
J
nh
(v
q
), ) =
L
q
i
(
Q
)
i
, (3.1.1)
where g
q
. Notice that the nonholonomic momentum map may be viewed as encoding some
of the components of the ordinary momentum map, namely the projection along those symmetry
directions that are consistent with the constraints.
[BKMM] extends the Noether Theorem to nonholonomic systems by deriving the equation for
the momentum map that replace the usual conservation law. It is proven that if the Lagrangian L
is invariant under the group action and that
q
is a section of the bundle g
D
, then any solution q(t)
of the Lagrange-dAlembert equations must satisfy, in addition to the given kinematic constraints,
the momentum equation:
d
dt
_
J
nh
(
q(t)
)
_
=
L
q
i
_
d
dt
(
q(t)
)
_
i
Q
. (3.1.2)
When the momentum map is paired with a section in this way, we will just refer to it as the
momentum. Examples show that the nonholonomic momentum map may or may not be conserved.
13
The Momentum Equation in Body Representation Let a local trivialization (r, g) be chosen
on the principal bundle : Q Q/G. Let g
q
and = g
1
g. Since L is G-invariant, we can
dene a new function l by writing L(r, g, r, g) = l(r, r, ). Dene J
nh
loc
: TQ/G (g
D
)
by
_
J
nh
loc
(r, r, ),
_
=
_
l
,
_
.
As with connections, J
nh
and its version in a local trivialization are related by the Ad map; i.e.,
J
nh
(r, g, r, g) = Ad
g
1
J
nh
loc
(r, r, ).
Choose a q-dependent basis e
a
(q) for the Lie algebra such that the rst m elements span the
subspace g
q
. In a local trivialization, one chooses, for each r, such a basis at the identity element,
say
e
1
(r), e
2
(r), . . . , e
m
(r), e
m+1
(r), . . . , e
k
(r).
Dene the body xed basis by e
a
(r, g) = Ad
g
e
a
(r); thus, by G invariance, the rst m elements
span the subspace g
q
. In this basis, we have
_
J
nh
(r, g, r, g), e
b
(r, g)
_
=
_
l
, e
b
(r)
_
:= p
b
, (3.1.3)
which denes p
b
, a function of r, r and . Note that in this body representation, the functions p
b
are invariant rather than equivariant, as is usually the case with the momentum map. It is shown
in [BKMM] that in this body representation, the momentum equation is given by
d
dt
p
i
=
_
l
, [, e
i
] +
e
i
r
_
, (3.1.4)
where the range of i is 1 to m. Moreover, the momentum equation in this representation is
independent of, that is, decouples from, the group variables g.
The Nonholonomic Connection Recall that in the case of simple holonomic mechanical sys-
tem, the mechanical connection / is dened by /(v
q
) = I(q)
1
J(v
q
) where J is the associated
momentum map and I(q) is the locked inertia tensor of the system. Equivalently the mechanical
connection can also be dened by the fact that its horizontal space at q is orthogonal to the group
orbit at q with respect to the kinetic energy metric. For more information, see for example, Marsden
[1992] and Marsden and Ratiu [1994].
As [BKMM] points out, in the principal case where the constraints and the orbit directions
span the entire tangent space to the conguration space, that is,
T
q
+T
q
(Orb(q)) = T
q
Q, (3.1.5)
the denition of the momentum map can be used to augment the constraints and provide a con-
nection on Q Q/G. Let J
nh
be the nonholonomic momentum map and dene similarly as above
a map A
sym
q
: T
q
Q o
q
given by
A
sym
(v
q
) = (I
nh
(q)
1
J
nh
(v
q
))
Q
14
(this denes the momentum constraints) where I
nh
: g
D
(g
D
)
Q
is used to dene the constraint submanifold / T
Q where
/ = FL(T). (3.2.1)
On this manifold, there is a distribution H
H = T T/, (3.2.2)
where
T = (T)
1
(T), (3.2.3)
and : T
Q to the
distribution H of the constraint submanifold /, is nondegenerate and that the dynamics is given
by a vector eld X
H
on / taking values in H and satises the equation
X
H
H
= dH
H
(3.2.4)
15
where dH
H
is the (berwise) restriction of dH
M
to H.
Now let G be the symmetry group of this system and assume that the quotient space / = //G
of the G-orbit in /is a quotient manifold with projection map : / /. Since G is a symmetric
group, all intrinsically dened vector elds and distributions push down to /. In particular, the
vector eld X
M
on / pushes down to a vector eld X
M
=
X
M
, and the distribution H pushes
down to a distribution
H on /.
However,
H
need not push down to a two-form dened on
M
H
,= 0, To eliminate this diculty, [BS] restricts
H
to a subdistribution | of H dened by
| = u H [
H
(u, v) = 0 for all v 1 H = H (1 H)
, (3.2.5)
where 1 is the distribution on /tangent to the orbits of G in /and is spanned by the innitesimal
symmetries and (1 H)
is the
H
-orthogonal complement of (1 H). Clearly, | and 1 are both
G-invariant, project down to / and
1 = 0. Dene H by
H =
|. (3.2.6)
It is proven in [BS] that
1. The vector eld X
H
which satises the above Hamiltonian equation of motion (3.2.4) lies in
the distribution |.
2. The restriction
U
of to the distribution | pushes down to a nondegenerate 2-form
H
=
U
on H, which is modeled by the symplectic space (1 H)
/(1 H) (1 H)
.
3. Furthermore,
X
H
H
= dh
H
, (3.2.7)
where h
M
=
H
M
is the pushdown of the restriction to / of the Hamiltonian H and dh
H
is the restriction of dh
M
to H. This is because the equation X
H
H
= dH
H
, restricted to
| H, vanishes on vectors in 1, and is G-invariant. Hence both sides push down to H.
Note that the original equations of motion are
X
H
H
= dH
H
(3.2.8)
where H is a distribution in the constraint manifold /. After the reduction of symmetry we obtain
equations of the same type
X
H
H
= dh
H
, (3.2.9)
where H is a distribution in the reduced space / = //G.
3.3 Lagrangian Side
By using the Legendre transformation FL, we can construct dual geometric structures on the
tangent bundle TQ and formulate a similar Lagrangian reduction procedure. This allows us to
better compare with the geometric constructions and analytic formulations on the manifold Q in
[BKMM], and in the course of doing this, we realize that the requirement (see point (1) of last
16
subsection) that the vector eld X
H
lies in the subdistribution | is equivalent to the extended
Noether Theorem; that is, that any solution of the Lagrange-dAlembert equations must satisfy the
momentum equation.
Recall from Section 2. We consider T as a constraint submanifold of TQ and then construct
the distribution
/ = ( TT, (3.3.1)
on TTQ, where
( = (T
Q
)
1
(T), (3.3.2)
and
Q
: TQ Q. Clearly T = (FL)
1
(/), / = (TFL)
1
(H). The motion is then given by a
vector eld X
K
on the manifold T which takes values in / and satises the equation
X
K
K
= dE
K
, (3.3.3)
where dE
K
and
K
are the restrictions of dE
D
and
D
respectively to the distribution /.
Now let G be the symmetry group of this system and assume that the quotient space T = T/G
of the G-orbit in T is a smooth quotient manifold with projection map : T T. Since G
is a symmetric group, all intrinsically dened vector elds and distributions push down to T. In
particular, the vector eld X
D
on T pushes down to a vector eld X
D
=
X
D
, and the distribution
/ pushes down to a distribution
to mean that
the vector elds are -related.
For the same reason as the Hamiltonian side,
K
need not push down to a two-form dened on
, (3.3.4)
where T is the distribution on T tangent to the orbits of G in T and is spanned by the innitesimal
symmetries. Clearly, J and T are both G-invariant, J projects down to T and
T = 0. Dene
/ by
/ =
J. (3.3.5)
Since the above constructions are dual to those in the Hamiltonian side, we also have
1. The vector eld X
K
which satises the above equation (3.3.3) takes values in the distribution
J.
2. The restriction
W
of
L
to the distribution J, pushes down to a nondegenerate 2-form
K
=
W
on /, which is modeled by the symplectic space (T /)
/(T /) (T /)
.
3. The reduced equations of motion are given by
X
K
K
= dE
K
, (3.3.6)
where E
D
=
E
D
is the pushdown of the restriction to T of the energy function E. This
is because the equation X
K
K
= dE
K
, restricted to J /, vanishes on vectors in T ,
and is G-invariant. Hence both sides push down to /. All these will become clearer in the
subsequent computations.
17
3.4 The Equivalence of Hamiltonian and Lagrangian Reductions
Theorem 3.2 Consider a simple nonholonomic mechanical system with symmetry and assume
that it is in the principal case. Then the reduction procedure on TQ described in the preceding
section gives the same set of equations as in [BKMM].
Proof The rst diculty is how to represent the constraint submanifold T TQ in a way that
is both intrinsic and ready for reduction. The comparison with the geometric constructions in
[BKMM] and the desire to have the dynamics break up in a way that are ready for reconstruction
give hints that we should use the tools like nonholonomic momentum p and the nonholonomic
connection / in [BKMM] to describe the constraint submanifold T
Recall that in [BKMM], the nonholonomic constraints together with the basic identity of the
nonholonomic momentum map are used to synthesis a nonholonomic connection / and the non-
holonomic constraints are then written in the form
g
1
g = A(r) r + (r)p, (3.4.1)
where p is G-invariant. Hence, the constraint manifold is nothing but
T = (g, r, g, r) [ g = g(A(r) r + (r))p). (3.4.2)
It is a submanifold in TQ and we can use (g, r, r, p) as its induced local coordinates. Then, clearly,
the corresponding coordinates for T = T/G are (r, r, p). From now on, we will use A(r) to
abbreviate /
nh
loc
(r).
The next diculty is to nd the corresponding representations for the distribution /, the
subdistribution T / and its annihilator distribution J where
J = / (T /)
. (3.4.3)
Recall that in [BKMM], a body xed basis e
b
(g, r) = Ad
g
e
b
(r) has been constructed such that the
innitesimal generators (e
i
(g, r))
Q
of its rst m elements at a point q span o
q
= T
q
T
q
(Orb(q)).
Assume that G is a matrix group and e
d
i
is the component of e
i
(r) with respect to a xed basis
b
a
of the Lie algebra g where (b
a
)
Q
=
g
a, then
(e
i
(g, r))
Q
= g
a
d
e
d
i
g
a.
Since / = (T)
1
(T) where T
q
is the direct sum of o
q
and the horizontal space of the nonholonomic
connection /
nh
, it can be represented in the induced coordinates by
/ = spang
a
d
e
d
i
g
a, g
a
b
A
b
g
a +
r
,
r
,
p
. (3.4.4)
Also, we have
T / = spang
a
d
e
d
i
g
a. (3.4.5)
To nd the distribution J, we have to compute g
a
d
e
d
i
g
a
D
, for all i = 1, . . . , m. Since L is
G-invariant, we have
D
= dg
a
d
_
L
g
a
_
+dr
d
_
L
r
_
= dg
a
d
_
(g
1
)
b
a
l
b
_
+dr
d
_
l
r
_
=
(g
1
)
b
a
g
c
l
b
dg
a
dg
c
+ (g
1
)
b
a
dg
a
d
_
l
b
_
+dr
d
_
l
r
_
18
Hence
g
a
f
e
f
i
g
a
D
= g
a
f
e
f
i
(g
1
)
b
a
g
c
l
b
dg
c
g
c
f
e
f
i
(g
1
)
b
a
g
c
l
b
dg
a
+e
b
i
d
_
l
b
_
= e
f
i
__
g
c
f
(g
1
)
b
c
g
a
(g
1
)
b
a
g
c
g
c
f
_
l
b
dg
a
+d
_
l
f
__
= e
f
i
_
(g
1
)
b
f
g
a
+
g
a
g
f
_
l
b
(g
1
)
a
e
dg
e
+d
_
l
f
__
= e
f
i
_
C
b
af
l
b
(g
1
)
a
e
dg
e
+d
_
l
f
__
= dp
i
l
f
d(e
f
i
) C
b
af
l
b
e
f
i
(g
1
)
a
e
dg
e
.
Here, C
b
af
is the structural constants for the Lie algebra g and p
i
=
l
f
e
f
i
as dened in (3.1.3).
Therefore, the subdistribution J / is
J = ker
_
dp
i
l
f
d(e
f
i
) C
b
af
l
b
e
f
i
(g
1
)
a
e
dg
e
_
. (3.4.6)
Since the constraint manifold T has the induced local coordinates (g, r, r, p), any vector eld
X
D
on the manifold T is of the form
X
D
= g
a
g
a + r
r
+ r
r
+ p
i
p
i
.
If X
D
lies in the distribution /, then we have g = g(A r + p). Moreover, if X
D
lies in the
distribution J, then for each j, we have
p
j
l
d
e
d
j
r
C
b
ad
l
a
e
d
j
= 0, i.e., p
j
=
_
l
, [, e
j
] + e
j
_
, (3.4.7)
which gives exactly the momentum equation (3.1.4). Therefore, any vector eld X
W
taking values
in J must be of the form
X
W
= g
a
b
g
a + r
r
+ r
r
+ p
i
p
i
, (3.4.8)
where
= A r + p p
j
=
_
l
, [, e
j
] + e
j
_
, (3.4.9)
Now we are ready to do the reduction. But before that, we need to compute all the ingredients
of the equation
X
K
K
= dE
K
. (3.4.10)
Notice rst that since E is G-invariant, we have
E =
L
q
i
q
i
L =
L
g
a
g
a
+
L
r
L
=
l
a
+
l
r
l
19
After restricting it to the submanifold T, we have
E
D
=
l
a
(A
a
+
ai
p
i
) +
_
l
c
r
+A
a
a
_
r
l
c
=
l
ai
p
i
+
l
c
r
l
c
Therefore,
dE
D
=
l
a
_
ai
r
p
i
dr
+
ai
dp
i
_
+
ai
p
i
_
2
l
r
a
dr
+
2
l
r
a
d r
+
2
l
p
j
a
dp
j
_
+ r
_
2
l
c
r
dr
+
2
l
c
r
d r
2
l
c
p
i
r
dp
i
_
l
c
r
dr
l
c
p
i
dp
i
.
Furthermore,
X
K
D
= g
a
f
f
(g
1
)
b
a
g
c
l
b
dg
c
g
c
f
f
(g
1
)
b
a
g
c
l
b
dg
a
+g
a
f
f
(g
1
)
b
a
d
_
l
b
_
_
r
_
l
b
_
r
+
r
_
l
b
_
r
+
p
i
_
l
b
_
p
i
_
(g
1
)
b
a
dg
a
+ ( r
r
+ r
r
+ p
i
p
i
)
_
dr
d
_
l
r
__
=
f
d
_
l
f
_
+
_
C
b
fa
l
d
dt
_
l
a
__
(g
1
)
a
e
dg
e
+ ( r
r
+ r
r
+ p
i
p
i
)
_
dr
d
_
l
r
__
. (3.4.11)
Clearly, both sides of the equation
X
K
K
= dE
K
(3.4.12)
are G-invariant, and when restricted to subdistribution J /, they vanish on the distribution
T /. This can be shown to be true either by invoking how J has been constructed or by direct
calculation, noticing that when
_
C
b
fa
l
d
dt
_
l
a
__
(g
1
)
a
e
dg
e
(3.4.13)
is paired with g
f
c
e
c
i
in T /, it is equal to zero on J. Hence both sides push down to / where
X
K
= r
r
+ r
r
+ p
i
p
i
, (3.4.14)
20
with
p
i
=
_
l
, [, e
i
] + e
i
_
. (3.4.15)
To nd the remaining reduced equations, notice that the restriction of (3.4.13) to the subdis-
tribution spanned by g
a
b
A
b
g
a +
r
,
r
,
p
i
is equivalent to
_
C
b
fa
l
d
dt
_
l
a
__
A
a
dr
. (3.4.16)
If we compute
_
C
b
fa
l
d
dt
_
l
a
__
A
a
dr
+
a
d
_
l
a
_
+ ( r
r
+ r
r
+ p
i
p
i
)
_
dr
d
_
l
r
__
and equate its terms with the corresponding terms of dE
K
which is the same as dE
K
, we have the
following equations after some computations
d
dt
_
l
c
r
l
c
r
= C
b
da
l
d
A
a
a
_
A
a
A
a
+
ai
p
i
r
_
.
After plugging in the constraint = A r+p and simplify, we get the desired reduced equations
d
dt
_
l
c
r
l
c
r
=
l
b
(B
b
+F
bi
p
i
), (3.4.17)
where
B
b
=
A
b
A
b
C
b
ac
A
a
A
c
(3.4.18)
F
bi
=
bi
r
C
b
ad
A
a
di
. (3.4.19)
In an orthogonal body frame where we choose our moving basis e
b
(g, r) to be orthogonal, that is,
the corresponding generators [e
b
(g, r)]
Q
are orthogonal in the given kinetic energy metric (actually,
all that is needed is that the vectors in the set of basis vectors corresponding to the subspace o
q
be orthogonal to the remaining basis vectors), the momentum equation (3.4.7) can be written as
(see [BKMM])
d
dt
p
i
= C
j
hi
I
hl
p
j
p
l
+T
j
i
r
p
j
+T
i
r
, (3.4.20)
where
T
j
i
= C
j
ai
A
a
+
j
i
+
a
C
a
li
I
lj
(3.4.21)
T
i
=
a
(C
a
ai
A
a
+
a
i
). (3.4.22)
Here
c
b
and
a
are dened by
e
b
r
=
c
b
e
c
(3.4.23)
=
l
b
A
b
. (3.4.24)
21
Notice that while the summation range of a, b, c, d... are over all Lie algebra element (1 to k). those
over i, j, l, ... are the restricted (constrained) range (1 to m) and those over a
, b
(A
a
e
a
) =
A
b
e
b
+A
a
b
a
e
b
(
bi
p
i
e
b
) =
bi
r
p
i
e
b
+
ai
b
a
p
i
e
b
,
we have
d
dt
_
l
c
r
l
c
r
=
l
b
_
A
b
A
b
C
b
ac
A
a
A
c
+A
c
b
c
A
c
b
c
_
r
b
_
bi
p
i
r
C
b
ad
A
a
di
p
i
+
ci
b
c
p
i
_
Now applying Proposition 7.1 of [BKMM] to the above reduced equations and notice that in
the orthogonal body basis,
bi
= 0 for any b > m (recall
ji
= I
ji
), we can rewrite the reduced
equations in the following form
d
dt
_
l
c
r
l
c
r
= (/
jl
p
j
p
l
+/
j
p
j
+/
) (3.4.25)
where
/
jl
=
I
jl
r
C
j
bh
A
b
I
hl
+
j
h
I
hl
(3.4.26)
/
j
=
a
(C
a
bh
A
b
I
hj
+
a
h
I
hj
) +B
j
(3.4.27)
/
=
a
B
a
. (3.4.28)
Here
B
b
=
A
b
A
b
C
b
ac
A
a
A
c
+A
c
b
c
A
c
b
c
. (3.4.29)
Remarks
1. A careful reading of the proof of Theorem 3.2 and the subsections 3.2 and 3.3 shows that the
Hamiltonian reduction procedure still works as long as the constrained Legendre transform
FL[T is invertible. This is important because in some examples like the bicycle the Legendre
transform FL is singular, but its restriction to the constraint submanifold T is invertible and
the Hamiltonian reduction procedure is also applicable.
22
2. In many examples like the snakeboard and the bicycle, the constraints satisfy a special condi-
tion, namely, they involve only the velocities of the group variables g and are independent of
the velocities of the shape variables r (see equations (2.5.1) and (2.5.2)). Under this special
condition, the distribution / in equation (3.4.4) can be represented by
/ = spang
a
d
e
d
i
g
a,
r
,
r
,
p
. (3.4.30)
This representation simplies the computation for nding the reduced equations because the
restriction of the one form (3.4.13) to the subdistribution / spanned by
r
,
r
,
p
will equal
to zero. Hence in pushing down X
K
D
in (3.4.11) to /, we can simply omit the one form
(3.4.13). In the following subsections, we will use this simplied procedure for the examples
of the snakeboard and the bicycle.
3. Since the momentum equation is central to the theory of nonholonomic mechanical systems
with symmetry, we make a few additional remarks about it. Before that, we state the following
proposition, the result of which is implicit in both [BKMM] and Ostrowski [1996].
Proposition 3.3 For a nonholonomic mechanical system with symmetry, we have
_
d
dt
_
L
q
i
_
L
q
i
_
(
q
Q
)
i
=
d
dt
__
L
q
i
_
(
q
Q
)
i
_
L
q
i
_
d
dt
q
_
i
Q
(3.4.31)
where
q
g
q
Proof: Choose a section of g
D
and apply the chain rule to give
d
dt
_
L
q
i
(
q
Q
)
i
_
=
d
dt
_
L
q
i
_
(
q
Q
)
i
+
L
q
i
_
(T
q
Q
q)
i
+
_
d
dt
q
_
i
Q
_
.
Invariance of the Lagrangian implies that
L(exp(s
q
) q, exp(s
q
) q) = L(q, q).
Dierentiating this expression and evaluating it at s = 0, we get
L
q
i
(
q
Q
)
i
+
L
q
i
(T
q
Q
q)
i
= 0
After eliminating the term
L
q
i
(T
q
Q
q)
i
from the above two equations, we arrive at the desired
result.
The above equation can be rewritten as
_
(dE X
L
)[
D
, (
q
Q
)
_
=
d
dt
__
L
q
i
_
(
q
Q
)
i
_
L
q
i
_
d
dt
q
_
i
Q
, (3.4.32)
where (
q
Q
)
T / and T
Q
((
q
Q
)
) =
q
Q
. Since both the energy function E and the submanifold
T are G-invariant, the left hand of the above equation reduces to
D
(X
D
, (
q
Q
)
L
q
i
_
d
dt
q
_
i
Q
= 0, (3.4.33)
as we have already seen in the proof of Theorem 3.2.
In showing that the vector eld X
H
, which satises the equation
X
H
H
= dH
H
,
must lie in the subdistribution |, one might think that any vector eld Y 1 H can be expressed
as a linear combination of innitesimal generators (generated by xed Lie algebra elements). But
this is not the case, as we have pointed out earlier in the Lagrangian side, in general (
q
Q
)
is the
(vertical) lift of a section of the bundle o (generated by a section of the bundle g
D
). This is also
true on the Hamiltonian side.
3.5 Example: The Snakeboard Revisited
Now we return to the snakeboard and discuss the role of the symmetry group G = SE(2). Recall
from our earlier discussion that the Lagrangian is
L(q, q) =
1
2
m( x
2
+ y
2
) +
1
2
mr
2
2
+ +
1
2
J
0
2
+J
0
+J
1
2
1
, (3.5.1)
which is independent of the conguration of the board and hence it is invariant to all possible group
actions.
The Constraint Submanifold. The condition of rolling without slipping gives rise to the con-
straint one forms
1
(q) = sin( +)dx + cos( +)dy r cos d
2
(q) = sin( )dx + cos( )dy +r cos d,
which are invariant under the SE(2) action. The constraints determine the kinematic distribution
T
q
:
T
q
= span
, a
x
+b
y
+c
,
where a = 2r cos
2
cos , b = 2r cos
2
sin , c = sin2. The tangent space to the orbits of the
SE(2) action is given by
T
q
(Orb(q)) = span
x
,
y
,
The intersection between the tangent space to the group orbits and the constraint distribution is
thus given by
o
q
= T
q
T
q
(Orb(q)) = spana
x
+b
y
+c
.
The momentum can be constructed by choosing a section of o = TTOrb regarded as a bundle
over Q. Since T
q
T
q
Orb(q) is one-dimensional, the section can be chosen to be
q
Q
= a
x
+b
y
+c
,
24
which is invariant under the action of SE(2) on Q. The nonholonomic momentum is thus given by
p =
L
q
i
(
q
Q
)
i
= ma x +mb y +mr
2
c
+J
0
c
.
The kinematic constraints plus the momentum are given by
0 = sin( +) x + cos( +) y r cos
p = 2mr cos
2
cos x 2mr cos
2
sin y
+mr
2
sin 2
+J
0
sin 2
.
Adding, subtracting, and scaling these equations, we can write (away from the point = /2),
_
_
cos x + sin y
sin x + cos y
_ +
_
J
0
2mr
sin 2
0
J
0
mr
2
sin
2
_
=
_
_
1
2mr
p
0
tan
2mr
2
p
_
_
. (3.5.2)
These equations have the form
g
1
g +A(r) r = (r)p
where
A(r) =
J
0
2mr
sin 2e
x
d +
J
0
mr
2
sin
2
e
d
(r) =
1
2mr
e
x
+
1
2mr
2
tan e
.
These are precisely the terms which appear in the nonholonomic connection relative to the (global)
trivialization (r, g).
After applying the constrained Legendre transformation and its inverse to the constraint equa-
tions (3.5.2), we have
_
_
cos p
x
+ sin p
y
sinp
x
+ cos p
y
p
_ +
_
mr sin cos
(mr
2
J
0
sin
2
)
p
mr
2
cos
2
(mr
2
J
0
sin
2
)
p
_
=
_
_
mr
2(mr
2
J
0
sin
2
)
p
0
(mr
2
J
0
) tan
2(mr
2
J
0
sin
2
)
p
_
_
, (3.5.3)
where
p = 2r cos
2
cos p
x
2r cos
2
sin p
y
+ sin2p
and is SE(2)-invariant.
25
Therefore, the constraint submanifold / T
Q is dened by
p
x
=
mr sin cos
(mr
2
J
0
sin
2
)
p
cos
mr
2(mr
2
J
0
sin
2
)
p cos
p
y
=
mr sin cos
(mr
2
J
0
sin
2
)
p
sin
mr
2(mr
2
J
0
sin
2
)
p sin
p
=
mr
2
cos
2
(mr
2
J
0
sin
2
)
p
+
(mr
2
J
0
) tan
2(mr
2
J
0
sin
2
)
p
It is a submanifold in T
, p
,
p
,
p
,
p
(3.5.4)
and the subdistribution 1 H is
1 H = span2r cos
2
cos
x
2r cos
2
sin
y
+ sin 2
. (3.5.5)
As for the subdistribution |, we rst calculate the two form
M
. After pulling back the
canonical two-form of T
Q to /, we have
M
= dx dp
x
+dy dp
y
+d dp
+d dp
+d dp
mr
2(mr
2
J
0
sin
2
)
dp
_
+(cos dx + sin dy)
_
mr(mr
2
cos 2 +J
0
sin
2
)
(mr
2
J
0
sin
2
)
2
p
d
mrJ
0
sin 2
2(mr
2
J
0
sin
2
)
2
pd
_
+d
_
mr
2
cos
2
(mr
2
J
0
sin
2
)
dp
+
(mr
2
J
0
) tan
2(mr
2
J
0
sin
2
)
dp
_
+d
_
mr
2
(J
0
mr
2
) sin 2
(mr
2
J
0
sin
2
)
2
p
d +
(mr
2
J
0
)(mr
2
sec
2
+J
0
tan
2
cos 2)
2(mr
2
J
0
sin
2
)
2
pd
_
+(sindx + cos dy)
_
mr sin 2
2(mr
2
J
0
sin
2
)
p
mr
2(mr
2
J
0
sin
2
)
p
_
d
+d dp
+d dp
Since | = (1 H)
= ker(1 H)
H
, we need to calculate (1 H)
M
, and restrict it
26
to H:
(1 H)
H
=
2r cos
2
_
mr sin2
2(mr
2
J
0
sin
2
)
dp
mr
2(mr
2
J
0
sin
2
)
dp
_
2r cos
2
_
mr(mr
2
cos 2 +J
0
sin
2
)
(mr
2
J
0
sin
2
)
2
p
d
mrJ
0
sin 2
2(mr
2
J
0
sin
2
)
2
pd
_
+ sin 2
_
mr
2
cos
2
(mr
2
J
0
sin
2
)
dp
+
(mr
2
J
0
) tan
2(mr
2
J
0
sin
2
)
dp
_
+ sin 2
_
mr
2
(J
0
mr
2
) sin 2
(mr
2
J
0
sin
2
)
2
p
d +
(mr
2
J
0
)(mr
2
sec
2
+J
0
tan
2
cos 2)
2(mr
2
J
0
sin
2
)
2
pd
_
= dp
2mr
2
cos
2
mr
2
J
0
sin
2
d +
(mr
2
+J
0
cos 2) tan
mr
2
J
0
sin
2
pd
Hence,
| = ker
_
dp
2mr
2
cos
2
mr
2
J
0
sin
2
d +
(mr
2
+J
0
cos 2) tan
mr
2
J
0
sin
2
pd
_
. (3.5.6)
The Reconstruction and Momentum Equations A vector eld X
U
taking values in | must
be of the form
X
U
= x
x
+ y
y
+
+ p
+ p
+ p
p
(3.5.7)
where
x =
J
0
2mr
sin 2
cos
1
2mr
p cos
y =
J
0
2mr
sin 2
sin
1
2mr
p sin
=
J
0
mr
2
sin
2
+
tan
2mr
2
p
and
p =
2mr
2
cos
2
mr
2
J
0
sin
2
(mr
2
+J
0
cos 2) tan
mr
2
J
0
sin
2
p
(3.5.8)
The equations for x, y and
are the same reconstruction equations as equations (3.5.2) and the last
one for p is the momentum elution on the Hamiltonian side. As noted in [BKMM], the momentum
p is the angular momentum of the system about the point P shown in gure 3.5.1.
It can be checked that the momentum equation (3.5.8) is equivalent to the equation (2.5.3) via
a change of variables with
p = 2r cos
2
cos p
x
2r cos
2
sin p
y
+ sin2p
=
2(mr
2
J
0
sin
2
) cot
mr
2
J
0
p
2mr
2
cos
2
cot
mr
2
J
0
p
as the key link. Similarly the two full sets of equations of motion in both section 2.5 and this
section are also related in the same way.
27
P
Figure 3.5.1: The momentum p is the angular momentum of the snakeboard system about the
point P.
The Reduced Hamilton Equations. To nd the remaining reduced equations, we need to
compute
X
H
M
= dH
M
, (3.5.9)
restrict it to the subdistribution | and then push it down to the reduced constraint submanifold
/. Let us rst compute X
H
M
X
H
M
=
( xcos + y sin)
_
mr sin2
2(mr
2
J
0
sin
2
)
dp
mr
2(mr
2
J
0
sin
2
)
dp
_
+( xcos + y sin )
_
mr(mr
2
cos 2 +J
0
sin
2
)
(mr
2
J
0
sin
2
)
2
p
d
mrJ
0
sin 2
2(mr
2
J
0
sin
2
)
2
pd
_
+
_
mr
2
cos
2
(mr
2
J
0
sin
2
)
dp
+
(mr
2
J
0
) tan
2(mr
2
J
0
sin
2
)
dp
_
+
_
mr
2
(J
0
mr
2
) sin 2
(mr
2
J
0
sin
2
)
2
p
d +
(mr
2
J
0
)(mr
2
sec
2
+J
0
tan
2
cos 2)
2(mr
2
J
0
sin
2
)
2
pd
_
+
dp
+
dp
d p
_
mr sin2
2(mr
2
J
0
sin
2
)
p
mr
2(mr
2
J
0
sin
2
)
p
_
(sin dx + cos dy)
mr
_
mr
2
cos 2 +J
0
sin
2
(mr
2
J
0
sin
2
)
2
p
J
0
sin 2
2(mr
2
J
0
sin
2
)
2
p
_
(cos dx + sin dy)
_
mr
2
(J
0
mr
2
) sin 2
(mr
2
J
0
sin
2
)
2
p
+
(mr
2
J
0
)(mr
2
sec
2
+J
0
tan
2
cos 2)
2(mr
2
J
0
sin
2
)
2
p
_
d
mr sin 2
2(mr
2
J
0
sin
2
)
p
(mr
2
J
0
sin
2
)
p
d
+
mr
2(mr
2
J
0
sin
2
)
(cos dx + sin dy)
(mr
2
J
0
) tan
2(mr
2
J
0
sin
2
)
pd.
28
As for dH
H
, recall that the constrained Hamiltonian H
M
is
H
M
=
mr
2
2(mr
2
J
0
)
2
cot
2
(p
)
2
+
1
2J
0
p
2
+
1
2(mr
2
J
0
)
(p
)
2
+
1
4J
1
p
2
.
Notice that H
M
is SE(2)-invariant and hence H
M
= h
M
where
h
M
=
mr
2
2
_
1
2(mr
2
J
0
sin
2
)
p
sin
2
2(mr
2
J
0
sin
2
)
p
_
2
+
1
2J
0
p
2
+
mr
2
J
0
2
_
tan
2(mr
2
J
0
sin
2
)
p
sin
2
mr
2
J
0
sin
2
_
2
+
1
4J
1
p
2
.
Compute dH
M
= dh
M
and we have
dh
M
=
mr
2
(p sin2p
)
2(mr
2
J
0
sin
2
)
_
1
2(mr
2
J
0
sin
2
)
dp
sin 2
2(mr
2
J
0
sin
2
)
dp
_
+
mr
2
(p sin 2p
)
2(mr
2
J
0
sin
2
)
_
pd
_
1
2(mr
2
J
0
sin
2
)
_
p
d
_
sin 2
2(mr
2
J
0
sin
2
)
__
+
(mr
2
J
0
)(tan p 2 sin
2
p
)
2(mr
2
J
0
sin
2
)
_
tan
2(mr
2
J
0
sin
2
)
dp
sin
2
(mr
2
J
0
sin
2
)
dp
_
+
(mr
2
J
0
)(tan p 2 sin
2
p
)
2(mr
2
J
0
sin
2
)
_
pd
_
tan
2(mr
2
J
0
sin
2
)
_
p
d
_
sin
2
(mr
2
J
0
sin
2
)
__
+
1
J
0
p
dp
+ +
1
2J
1
p
dp
.
It is easy to check that X
H
M
= dH
M
is SE(2)-invariant, and vanishes on 1 H when
restricted to |. Hence both sides push down to H. The push down of X
H
M
is given by
X
H
H
=
_
J
0
2mr
sin(2)
1
2mr
p
__
mr sin 2
2(mr
2
J
0
sin
2
)
dp
mr
2(mr
2
J
0
sin
2
)
dp
_
+
_
J
0
2mr
sin(2)
1
2mr
p
__
mr(mr
2
cos 2 +J
0
sin
2
)
(mr
2
J
0
sin
2
)
2
p
d
mrJ
0
sin2
2(mr
2
J
0
sin
2
)
2
pd
_
+
_
J
0
mr
2
sin
2
()
+
tan
2mr
2
p
__
mr
2
cos
2
(mr
2
J
0
sin
2
)
dp
+
(mr
2
J
0
) tan
2(mr
2
J
0
sin
2
)
dp
_
+
_
J
0
mr
2
sin
2
()
+
tan
2mr
2
p
_
mr
2
(J
0
mr
2
) sin 2
(mr
2
J
0
sin
2
)
2
p
d
+
_
J
0
mr
2
sin
2
()
+
tan
2mr
2
p
_
(mr
2
J
0
)(mr
2
sec
2
+J
0
tan
2
cos 2)
2(mr
2
J
0
sin
2
)
2
pd
+
dp
+
dp
d p
d.
Equating the terms of dh
H
= dh
M
with those of the push down of X
H
M
gives the remaining
29
reduced Hamilton equations:
=
tan
2(mr
2
J
0
sin
2
)
p +
mr
2
J
0
(mr
2
J
0
sin
2
)
p
(3.5.10)
=
p
2J
1
(3.5.11)
p
= 0 (3.5.12)
p
= 0. (3.5.13)
Notice that both the momentum equation (3.5.8) and the above set of reduced equations are inde-
pendent of the group elements of the symmetry group SE(2). If we add in the set of reconstruction
equations (3.5.2), we recover the full dynamics of the system, and in a form that is suitable for
control theoretical purposes.
Finding the Reduced Equations on the Lagrangian Side As shown in the proof of Theorem
3.2, we can derive the reduced Lagrange-dAlembert equations in two ways. Here we will rst use
the equations (3.4.17).
d
dt
_
l
c
r
l
c
r
=
l
b
(B
b
+F
bi
p
i
), (3.5.14)
where
B
b
=
A
b
A
b
C
b
ac
A
a
A
c
and F
bi
=
bi
r
C
b
ad
A
a
di
.
From the Lagrangian L, we nd the reduced Lagrangian
l(r, r, ) =
1
2
m((
1
)
2
+ (
2
)
2
) +
1
2
mr
2
(
3
)
2
+
1
2
J
0
2
+ +J
0
(
3
) +J
1
2
, (3.5.15)
where = g
1
g. After plugging in the constraints (3.5.2), we have the constrained reduced La-
grangian
l
c
(r, r, p) =
J
2
0
2mr
2
sin
2
2
+
1
8mr
2
sec
2
p
2
+
1
2
J
0
2
+ +J
1
2
. (3.5.16)
Let us nd all the ingredients of the above equations:
l
1
= m
1
= m
_
J
0
2mr
sin 2
1
2mr
p
_
l
2
= m
2
= 0
l
3
= mr
2
_
J
0
mr
2
sin
2
+
tan
2mr
2
p
_
+J
0
;
since
l
2
= 0, we do not need to compute B
2
and F
2
J
0
2mr
sin2
_
=
J
0
mr
cos 2
B
3
12
=
_
J
0
mr
sin
2
_
=
J
0
mr
sin 2
F
3
2
=
_
tan
2mr
2
_
=
sec
2
2mr
2
,
30
and F
1
1
= F
3
1
= F
1
2
= 0. Substituting into (3.5.14), we get the reduced equations after some
computations
_
1
J
0
mr
2
sin
2
_
=
J
0
2mr
2
sin2
J
0
2mr
2
p (3.5.17)
J
1
= 0 (3.5.18)
It is easy to check that these two equations are equivalent to the set of reduced equations (3.5.10)-
(3.5.13) on the Hamiltonian side through the constrained Legendre transformation FL[T.
Next we will nd the reduced equations use the equations (3.4.25)
d
dt
_
l
c
r
l
c
r
= (/
jl
p
j
p
l
+/
j
p
j
+/
) (3.5.19)
where
/
jl
=
I
jl
r
C
j
bh
A
b
I
hl
+
j
h
I
hl
/
j
=
a
(C
a
bh
A
b
I
hj
+
a
h
I
hj
) +B
j
=
a
B
a
.
Here
B
b
=
A
b
A
b
C
b
ac
A
a
A
c
+A
c
b
c
A
c
b
c
.
First we need to construct the orthogonal body frame. Recall that
(e
1
(g, r))
Q
= g
a
d
e
d
1
g
a = 2r cos
2
cos
x
2r cos
2
sin
y
+ sin 2
.
Hence
e
1
= 2r cos
2
e
x
+ sin 2e
,
where e
x
, e
y
, e
e
3
=
1
m
sin e
x
+
1
m
cos e
y
+
1
mr
cos e
=
a
e
a
, then
1
= 2r cos
2
1
m
sin
2
+
1
m
sin
3
2
=
1
m
cos
2
+
1
m
cos
3
1
= sin 2
1
1
mr
cos
2
+
1
mr
cos
3
,
and
l(r, r,
a
) = l(r, r, T
b
a
a
) where T
a
b
is dened as above by
b
= T
b
a
a
.
31
Notice that in the new basis, the constraints (3.5.2) become
_
3
_
_ =
_
_
J
0
2mr
2
tan
0
0
_
_
+
_
_
1
4mr
2
sec
2
p
0
0
_
_
, (3.5.20)
but the constrained reduced equation
l
c
(r, r, p) remains the same and is equal to l
c
(r, r, p).
Let us nd all the ingredients of equations (3.5.19). After nding from (3.5.20) that
A
1
1
=
J
0
2mr
2
tan and I
11
=
1
4mr
2
sec
2
2
,
2
). Front and rear wheel contacts are constrained to
have velocities parallel to the lines of intersection of their respective wheel planes and the ground
plane, but free to turn about an axis through the wheel/ground contact and parallel to the z-axis.
Let (
2
,
2
) be the steering angle between the front wheel plane/ground plane intersection and
the contact line. With we associate a moment of inertia J which depends both on and . We
will parametrize the steering angle by := tan /b. For more details, see Getz and Marsden [1995]
and Getz [1996]. See gure 3.6.1.
32
x
y
m
(x, y)
z
a
b
c
Figure 3.6.1: Notation for the bike.
The conguration space is Q = SE(2) S
1
S
1
and the Lagrangian L : TQ R is the total
kinetic energy minus potential energy of the system and is given by
L = mga cos +
1
2
J(, )
2
+
m
2
_
(cos x + sin y +a sin
)
2
+ (sin x + cos y a cos
+c
)
2
+ (a sin
)
2
_
where m is the mass of the bicycle, considered for simplicity to be a point mass, and J(, ) is the
moment of inertia associated with the steering action. The nonholonomic constraints associated
with the front and rear wheels, assumed to roll without slipping, are expressed by
(cos x + sin y) = 0
sin x + cos y = 0.
Clearly both the Lagrangian and the constraints are invariant under the SE(2) action.
Notice that the Legendre transform FL is singular but by the remark following Theorem 3.2 the
Hamiltonian procedure still works because the constrained Legendre transform FL[T is invertible.
The Constraint Submanifold The constraints above give rise to the constraint one forms
1
(q) = d cos dx sin dy
2
(q) = sin dx + cos dy
which determine the kinematic distribution T
q
:
T
q
= span
, cos
x
+ sin
y
+
.
The tangent space to the orbits of the SE(2) action is given by
T
q
(Orb(q)) = span
x
,
y
,
,
33
and the intersection between the tangent space to the group orbits and the constraint distribution
is thus given by
o
q
= T
q
T
q
(Orb(q)) = spancos
x
+ sin
y
+
.
The momentum can be constructed by choosing a section of o = TTOrb regarded as a bundle
over Q. Since T
q
T
q
Orb(q) is one-dimensional, the section can be chosen to be
q
Q
= cos
x
+ sin
y
+
,
which is invariant under the action of SE(2) on Q. The nonholonomic momentum map is thus
given by
p =
L
q
i
(
q
Q
)
i
= m( x +a sin cos
+a cos sin
c sin
) cos
+m( y +a sin sin
a cos cos
+c cos
) sin
+m(cos x + sin y +a sin
)asin
+m(sin x + cos y a cos
+c
)c.
The kinematic constraints plus the momentum are given by
0 =
3
1
0 =
2
p = m(
1
+a sin
3
) +masin (
1
+a sin
3
)
m(c
2
ca cos
+c
2
3
)
where
1
= cos x + sin y
2
= sin x + cos y
3
=
3
_
_ +
_
cacos
K
ca
2
cos
K
_
=
_
_
1
mK
p
0
mK
p
_
_
(3.6.1)
where
K = (1 +asin )
2
+c
2
2
. (3.6.2)
These equations have the form
g
1
g +A(r) r = (r)p.
34
Next nd the Legendre transform FL and restrict it to the constraint submanifold T TQ, we
get
p
x
= m(1 +asin )
1
cos m(c
1
a cos
) sin
p
y
= m(1 +asin )
1
sin +m(c
1
a cos
) cos
p
= ma sin (1 +asin )
1
+m(c
2
1
ca cos
)
p
= ma
2
mac cos
1
p
= J(, )
.
After applying the constrained Legendre transformation FL[T and its inverse to the constraint
equations (3.6.1), we have
_
3
_
_ +
_
ccos (1 +asin )
F
p
a
(1 +asin )
2
cos
F
p
a
c cos (1 +asin )
F
p
a
_
_
=
_
_
1 +asin
F
p
csin
2
F
p
(1 +asin )a sin +c
2
sin
2
F
p
_
_
, (3.6.3)
where
1
= cos p
x
+ sin p
y
2
= sin p
x
+ cos p
y
3
= p
and
F = (1 +asin )
2
+c
2
2
sin
2
(3.6.4)
p = p
x
cos +p
y
sin +p
. (3.6.5)
Therefore, the constraint submanifold / T
Q is dened by
p
x
=
1
cos
2
sin
p
y
=
1
sin +
2
cos
p
=
3
.
It is a submanifold in T
, p
,
p
,
p
,
p
(3.6.6)
and the subdistribution 1 H is
1 H = spancos
x
+ sin
y
+
. (3.6.7)
Notice that in the case of the bicycle, the constraints are independent of the velocities of the
shape variables and hence the simplied procedure employed in the snakeboard is also used here.
35
As for the subdistribution |, we rst calculate the two form
M
. After pulling back the
canonical two-form of T
Q to /, we have
M
= dx dp
x
+dy dp
y
+d dp
+d dp
+d dp
+d dp
Since | = (1 H)
= ker(1 H)
H
, we need to calculate (1 H)
M
, and restrict it
to H:
(1 H)
H
= d
1
1
(sin dx + cos dy)
2
d +
2
(cos dx + sindy) +d
3
= d
1
+d
3
= dp +
c cos (1 +asin )
F
p
a
d
a sin(1 +asin ) +c
2
sin
2
F
pd.
Hence,
| = ker
_
dp +
c cos (1 +asin )
F
p
a
d
a sin(1 +asin ) +c
2
sin
2
F
pd
_
. (3.6.8)
The Reconstruction and Momentum Equations A vector eld X
U
taking values in | must
be of the form
X
U
= x
x
+ y
y
+
+ p
+ p
+ p
p
(3.6.9)
where
x =
1
cos
2
sin =
_
cacos
K
+
1
mK
p
_
cos
y =
1
sin +
2
cos =
_
cacos
K
+
1
mK
p
_
sin
=
1
=
_
ca
2
cos
K
+
mK
p
_
and
p =
c cos (1 +asin )
F
p
+
a sin (1 +asin ) +c
2
sin
2
F
p
. (3.6.10)
The equations for x, y and
are the same reconstruction equations as equations (3.6.1) and the
last one for p is the momentum equation on the Hamiltonian side. Similar to the example of the
snakeboard, the momentum p equals the angular momentum of the system about a xed point P
that can be determined in the same way as in the case of the snakeboard. Notice also that the last
equation can be written simply as p =
3
.
36
The Reduced Hamilton Equations. To nd the remaining reduced equations, we need to
compute
X
H
M
= dH
M
, (3.6.11)
restrict it to the subdistribution | and then push it down to the reduced constraint submanifold
/. Let us rst compute X
H
M
X
H
M
=
(cos x + sin y)d
1
+
1
(sin x + cos y)d
1
(sin dx + cos dy)
+(sin x + cos y)d
2
2
(cos x + sin y)d +
2
(cos dx + sin dy)
+
d
3
+
dp
+
dp
d p
d
((
+ p
+ p
+ p
p
) d
1
)(cos dx + sin dy)
((
+ p
+ p
+ p
p
) d
2
)(sin dx + cos dy)
As for dH
H
, we can nd the constrained Hamiltonian H
M
via the constrained Legendre trans-
form and have
H
M
= mga cos +
1
2J
p
2
+
1
2m
_
2
1
+
2
2
+
_
K sin
F
p
a
+
csin cos
F
p
_
2
_
.
Notice that H
M
is SE(2)-invariant and hence H
M
= h
M
. Compute dH
M
= dh
M
and we have
dh
M
=
mga sin d +
1
J
p
dp
1
2J
2
p
2
_
J
d +
J
d
_
+
1
m
_
1
d
1
+
2
d
2
+
_
K sin
F
p
a
+
csin cos
F
p
_
d
_
K sin
F
p
a
+
csin cos
F
p
__
.
It can be checked that X
H
M
= dH
M
is SE(2)-invariant, and vanishes on 1 H when
restricted to |. Hence both sides push down to H. The push down of X
H
M
is given by
X
H
H
= (cos x + sin y)d
1
+
d
3
+
dp
+
dp
d p
d
=
1
d
1
+
3
d
3
+
dp
+
dp
d p
d
Equating the terms of dh
H
= dh
M
with those of the push down of X
H
M
gives the remaining
reduced Hamilton equations:
=
1
ma
_
K
F
p
a
+
ccos
F
p
_
(3.6.12)
=
p
J
(3.6.13)
p
= mga sin +
1
2J
2
p
2
=
1
2J
2
J
p
2
, (3.6.15)
37
where
1
=
ccos
K
+
1
mK
p =
ccos
mF
p
a
+
1
mF
p
as dened earlier in (3.6.1). The rst two equations are nothing but the inverse of the constrained
Legendre transform. Notice that both the momentum equation (3.6.10) and the above set of reduced
equations are independent of the group elements of the symmetry group SE(2). If we add in the
set of reconstruction equations (3.6.1), we recover the full dynamics of the system, and in a form
that is suitable for control theoretical purposes. Methods developed in Koon and Marsden [1995]
and Ostrowski, Desai and Kumar [1996] will be used to study the optimal control of the bicycle
whose equations of motion have been found in this section.
Conclusions.
In this paper we have analyzed the relation between the Lagrangian and Hamiltonian approaches
to problems in nonholonomic mechanics. In the course of doing this, we have claried each of the
pictures. For example, we have shown how the momentum equation rst found on the Lagrangian
side ts into the Hamiltonian approach. We have also explored the reduced Lagrange-dAlembert
equations in greater detail than was known previously. An example, a simplied model of the
bicycle is used to illustrate the ideas.
This paper concentrates in comparing dierent but equivalent formulations of mechanics with
nonholonomic constraints from the intrinsic point of view. While a further comparison with the
extrinsic point of view taken in Dezord [1994] and Marle [1995] would be interesting, we will leave
it to the future.
One aspect we do not address in this paper is the point of view of Poisson geometry and the
Dirac theory of constraints. It is known that the obvious Poisson structures for nonholonomic
systems do not satisfy the Jacobi identity (this is already mentioned in [BS] and van der Schaft
and Maschke [1994]). Thus, any discussion in this direction should take this into account. We hope
to address some of these issues in the future. Another item for future work is the extension of the
theory of geometric phases (as in Marsden, Montgomery and Ratiu [1990]) to the nonholonomic
case.
References
Arnold, V.I. [1988] Dynamical Systems III. Encyclopedia of Mathematics 3, Springer-Verlag.
Bates, L., H.Graumann, and C.MacDonnell [1996] Examples of gauge conservation laws in non-
holonomic systems. Rep. Math. Phys., 37, 295308.
Bates, L. and J. Sniatycki [1993] Nonholonomic reduction, Reports on Math. Phys. 32, 99115.
Bloch, A.M. and P. Crouch [1992] On the dynamics and control of nonholonomic systems on
Riemannian Manifolds. Proceedings of NOLCOS 92, Bordeaux, 368372.
Bloch, A.M. and P. Crouch [1994] Nonholonomic control systems on Riemannian manifolds. SIAM
J. on Control (to appear).
38
Bloch, A.M. and P.E. Crouch [1994] Reduction of Euler Lagrange problems for constrained vari-
ational problems and relation with optimal control problems. Proc. CDC, IEEE 33, 2584
2590.
Bloch, A.M. and P. Crouch [1995b] Reduction of Euler Lagrange problems for constrained varia-
tional problems and relation with optimal control problems. 33rd CDC, to appear.
Bloch, A.M., P.S. Krishnaprasad, J.E. Marsden, and R. Murray [1996] Nonholonomic mechanical
systems with symmetry. Arch. Rat. Mech. An., 136, 2199.
Cushman, R., J. Hermans, and D. Kemppainen [1995] The rolling disc. in Nonlinear dynamical
systems and chaos (Groningen, 1995), Progr. Nonlinear Dierential Equations Appl., 19,
Birkhduser, Basel, 2160.
Dazord, P [1994] Mecanique hamiltonienne en presence de constraintes. Illinois J. of Math.38:148-
175.
Getz, N.H. [1994] Control of balance for a nonlinear nonholonomic non-minimum phase model of
a bicycle. ACC Conf., Baltimore, June, 1994.
Getz, N.H. and J. E. Marsden [1995] Control for an autonomous bicycle, International Conference
on Robotics and Automation, IEEE, Nagoya, Japan, May, 1995.
Hermans, J. [1995] A symmetric sphere rolling on a surface. Nonlinearity 8, 4:493-515.
Koiller, J. [1992] Reduction of some classical nonholonomic systems with symmetry. Arch. Rat.
Mech. An. 118, 113148.
Koon, W.S. [1997] Reduction, Reconstruction and Optimal Control for Nonholonomic Mechanical
Systems with Symmetry. Thesis, Department of Mathematics, UC Berkeley.
Koon, W.S. and J.E. Marsden [1997] Optimal control for holonomic and nonholonomic mechanical
systems with symmetry and Lagrangian reduction. SIAM J. Control and Optim. 35, 901929.
Lewis, A. [1996] Ane connections and distributions, preprint.
Marle, C.-M. [1995] Reduction of constrained mechanical systems and stability of relative equilib-
ria. Comm. in Math. Phys. 174:295-318.
Marsden, J.E. [1992], Lectures on Mechanics London Mathematical Society Lecture Note Series.
174, Cambridge University Press.
Marsden, J.E., R. Montgomery, and T.S. Ratiu [1990] Reduction, symmetry, and phases in me-
chanics. Memoirs AMS 436.
Marsden, J.E. and T.S. Ratiu [1992] An Introduction to Mechanics and Symmetry. Texts in Appl.
Math. 17, Springer-Verlag.
Marsden, J.E. and J. Scheurle [1993a] Lagrangian reduction and the double spherical pendulum,
ZAMP 44, 1743.
39
Marsden, J.E. and J. Scheurle [1993b] The reduced Euler-Lagrange equations, Fields Institute
Comm. 1, 139164.
Ostrowski, J. [1996] The Mechanics and Control of Undulatory Robotic Locomotion, PhD thesis,
California Institute of Technology.
van der Schaft, A.J. and B.M. Maschke [1994] On the Hamiltonian formulation of nonholonomic
mechanical systems, Rep. on Math. Phys. 34, 225233.
Vershik, A.M. and V.Ya. Gershkovich [1994] Nonholonomic dynamical systems, geometry of
distributions and variational problems, Dynamical Systems VII, V. Arnold and S.P. Novikov,
eds., 181. Springer-Verlag, New York.
Weber, R.W. [1986] Hamiltonian systems with constraints and their meaning in mechanics. Arch.
Rat. Mech. An. 91, 309335.
40