mechanics lagrangian and hamilitonian
TRANSCRIPT
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
1/56
Lecture notes
An introduction to Lagrangian and Hamiltonian mechanics
Simon J.A. Malham
Simon J.A. Malham (6th March 2014)Maxwell Institute for Mathematical Sciencesand School of Mathematical and Computer SciencesHeriot-Watt University, Edinburgh EH14 4AS, UKTel.: +44-131-4513200Fax: +44-131-4513249E-mail: [email protected]
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
2/56
2 Simon J.A. Malham
1 Introduction
Newtonian mechanics took the Apollo astronauts to the moon. It also took the voy-
ager spacecraft to the far reaches of the solar system. However Newtonian mechanics
is a consequence of a more general scheme. One that brought us quantum mechanics,
and thus the digital age. Indeed it has pointed us beyond that as well. The scheme is
Lagrangian and Hamiltonian mechanics. Its original prescription rested on two prin-
ciples. First that we should try to express the state of the mechanical system using
the minimum representation possible and which reflects the fact that the physics of
the problem is coordinate-invariant. Second, a mechanical system tries to optimize its
action from one split second to the next; often this corresponds to optimizing its to-
tal energy as it evolves from one state to the next. These notes are intended as an
elementary introduction into these ideas and the basic prescription of Lagrangian and
Hamiltonian mechanics. A pre-requisite is the thorough understanding of the calculus
of variations, which is where we begin.
2 Calculus of variations
Many physical problems involve the minimization (or maximization) of a quantity that
is expressed as an integral.
2.1 Example (Euclidean geodesic)
Consider the path that gives the shortest distance between two points in the plane, say
(x1, y1) and (x2, y2). Suppose that the general curve joining these two points is given
byy = y(x). Then our goal is to find the function y(x) that minimizes the arclength:
J(y) =
(x2,y2)(x1,y1)
ds
=
x2x1
1 + (yx)2 dx.
Here we have used that for a curve y = y(x), if we make a small increment in x,
say x, and the corresponding change in y is y, then by Pythagoras theorem the
corresponding change in length along the curve iss = +
(x)2 + (y)2. Hence we
see that
s= s
xx = 1 +
y
x2
x.
Note further that here, and hereafter, we use yx = yx(x) to denote the derivative ofy ,
i.e. yx(x) = y(x) for each x for which the derivative is defined.
2.2 Example (Brachistochrome problem; John and James Bernoulli 1697)
Suppose a particle/bead is allowed to slide freely along a wire under gravity (force
F = gk wherek is the unit upward vertical vector) from a point (x1, y1) to the origin
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
3/56
Lagrangian and Hamiltonian mechanics 3
1
2(x , y )
x , y )1
2
y=y(x)
Fig. 1 In the Euclidean geodesic problem, the goal is to find the path with minimum totallength between points (x1, y1) and (x2, y2).
(0,0)
(x , y )1 1
y=y(x)mgk
v
Fig. 2 In the Brachistochrome problem, a bead can slide freely under gravity along the wire.The goal is to find the shape of the wire that minimizes the time of descent of the bead.
(0, 0). Find the curve y = y(x) that minimizes the time of descent:
J(y) = (0,0)
(x1,y1)
1
vds
=
0x1
1 + (yx)2
2g(y1 y)dx.
Here we have used that the total energy, which is the sum of the kinetic and potential
energies,
E= 12mv2 + mgy,
is constant. Assume the initial condition isv = 0 when y = y1, i.e. the bead starts with
zero velocity at the top end of the wire. Since its total energy is constant, its energy at
any timet later, when its height is y and its velocity is v , is equal to its initial energy.
Hence we have
12mv
2 + mgy = 0 + mgy1 v = +
2g(y1 y).
3 EulerLagrange equation
We can see that the two examples above are special cases of a more general problem
scenario.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
4/56
4 Simon J.A. Malham
3.1 Classical variational problem
Suppose the given functionF(, , ) is twice continuously differentiable with respect toall of its arguments. Among all functions/pathsy = y(x), which are twice continuously
differentiable on the interval [a, b] with y(a) and y(b) specified, find the one which
extremizes thefunctional defined by
J(y):=
ba
F(x ,y,yx) dx.
Theorem 1 (EulerLagrange equations) The function u = u(x) that extremizes the
functionalJ necessarily satisfies the EulerLagrange equation
F
u d
dxF
ux = 0,on[a, b].
ProofConsider the family of functions on [a, b] given by
y(x; ):=u(x) + (x),
where the functions = (x) are twice continuously differentiable and satisfy (a) =
(b) = 0. Here is a small real parameter, and of course, the functionu = u(x) is our
candidate extremizing function. We set (see for example Evans [7, Section 3.3])
():=J(u + ).
If the functional J has a local maximum or minimum at u, then u is a stationary
function forJ, and for all we must have
(0) = 0.
To evaluate this condition for the functional given in integral form above, we need to
first determine(). By direct calculation,
() = d
dJ(u + )
= d
d
ba
F(x, u + ,ux+ x) dx
=
ba
F(x, y(x; ), yx(x; )) dx
= b
a
Fy
y
+ Fyx
yx
dx
=
ba
F
y(x) +
F
yx(x)
dx.
The integration by parts formula tells us that
ba
F
yx(x) dx=
F
yx(x)
x=bx=a
ba
d
dx
F
yx
(x) dx.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
5/56
Lagrangian and Hamiltonian mechanics 5
Recall that (a) = (b) = 0, so the boundary term (first term on the right) vanishes
in this last formula, and we see that
() =
ba
F
y d
dx
F
yx
(x) dx.
If we now set = 0, the condition for u to be a critical point ofJ is
ba
F
u d
dx
F
ux
(x) dx= 0.
Now we note that the functions = (x) are arbitrary, using Lemma 1 below, we
can deduce that pointwise, i.e. for all x [a, b], necessarily u must satisfy the EulerLagrange equation shown.
3.2 Useful lemma
Lemma 1 (Useful lemma) If(x) is continuous in[a, b] and
ba
(x) (x) dx= 0
for all continuously differentiable functions (x) which satisfy (a) = (b) = 0, then
(x) 0 in[a, b].
Proof Assume(z)> 0, say, at some pointa < z < b. Then since is continuous, we
must have that(x)> 0 in some open neighbourhood ofz, say in a < z < z < z < b.The choice
(x) =
(x z)2(z x)2, for x [z, z],0, otherwise,
which is a continuously differentiable function, implies
ba
(x) (x) dx=
zz
(x) (x z)2(z x)2 dx > 0,
a contradiction.
3.3 Remarks
Some important theoretical and practical points to keep in mind are as follows.
1. The EulerLagrange equation is a necessary condition: if such a u = u(x) exists
that extremizesJ, thenu satisfies the EulerLagrange equation. Such au is known
as a stationary functionof the functional J.
2. The EulerLagrange equation above is an ordinary differential equation foru; this
will be clearer once we consider some examples presently.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
6/56
6 Simon J.A. Malham
3. Note that the extremal solution u is independent of the coordinate system you
choose to represent it (see Arnold [3, Page 59]). For example, in the Euclideangeodesic problem, we could have used polar coordinates (r, ), instead of Cartesian
coordinates (x, y), to express the total arclength J. Formulating the EulerLagrange
equations in these coordinates and then solving them will tell us that the extrem-
izing solution is a straight line (only it will be expressed in polar coordinates).
4. Let Y denote a function space; in the context above Y was the space of twice
continuously differentiable functions on [a, b] which are fixed at x = a and x = b.
A functionalis a real-valued map and here J: Y R.5. We define thefirst variation J(u, ) of the functional J, at u in the direction ,
to be J(u, ):=(0).
6. Is u a maximum, minimum or saddle point for J? The physical context should
hint towards what to expect. Higher order variations will give you the appropriate
mathematical determination.
7. The functional J has a local minimum at u iff there is an open neighbourhoodU Y of u such that J(y) J(u) for all y U. The functional J has a localmaximum atu when this inequality is reversed.
8. We will generalize all these notions to multidimensions and systems later.
3.4 Solution (Euclidean geodesic)
Recall, this variational problem concerns finding the shortest distance between the two
points (x1, y1) and (x2, y2) in the plane. This is equivalent to minimizing the total
arclength functional
J(y) =
x2x1
1 + (yx)2 dx.
Hence in this case, theintegrand is F(x ,y,yx) =
1 + (yx)2. From the general theory
outlined above, we know that the extremizing solution satisfies the EulerLagrange
equationF
y d
dx
F
yx
= 0.
We substitute the actual form forFwe have in this case which gives
ddx
yx
1 + (yx)
2 12 = 0 d
dx
yx
1 + (yx)2
1
2
= 0
yxx1 + (yx)2
12
(yx)2
yxx1 + (yx)2
32
= 0
1 + (yx)2
yxx1 + (yx)2
32
(yx)2yxx
1 + (yx)2 32
= 0
yxx1 + (yx)2
32
= 0
yxx = 0.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
7/56
Lagrangian and Hamiltonian mechanics 7
Hence y(x) = c1 + c2x for some constants c1 and c2. Using the initial and starting
point data we see that the solution is the straight line function (as we should expect)
y=
y2 y1x2 x1
(x x1) + y1.
Note that this calculation might have been a bit shorter if we had recognised that this
example corresponds to the third special case in the next section.
4 Alternative form and special cases
4.1 Alternative form
The EulerLagrange equations given by
F
y d
dx F
yx
= 0
are equivalent to the following alternative formulation of the EulerLagrange equations
F
x d
dx
F yx F
yx
= 0.
4.2 Special cases
When the functional Fdoes not explicitly depend on one or more variables, then the
EulerLagrange equations simplify considerably. We have the following three notable
cases:
1. IfF = F(y, yx) only, i.e. it does not explicitly depend on x, then the alternative
form for the EulerLagrange equation impliesd
dx
F yx F
yx
= 0 F yx F
yx=c,
for some arbitrary constant c.
2. IfF =F(x, yx) only, i.e. it does not explicitlydepend ony, then the EulerLagrange
equation impliesd
dx
F
yx
= 0 F
yx=c,
for some arbitrary constant c.
3. IfF =F(yx) only, then the EulerLagrange equation implies
0 = F
y d
dxF
yx=
F
y
2F
xyx+ yx
2F
yyx+ yxx
2F
yxyx
= yxx 2F
yxyx.
Hence yxx = 0, i.e. y = y(x) is a linear function ofx and has the form
y= c1+ c2x,
for some constants c1 and c2.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
8/56
8 Simon J.A. Malham
4.3 Solution (Brachistochrome problem)
Recall, this variational problem concerns a particle/bead which can freely slide along
a wire under the force of gravity. The wire is represented by a curve y = y(x) from
(x1, y1) to the origin (0, 0). The goal is to find the shape of the wire, i.e. y = y(x),
which minimizes the time of descent of the bead, which is given by the functional
J(y) =
0x1
1 + (yx)2
2g(y1 y)dx= 1
2g
0x1
1 + (yx)2
(y1 y) dx.
Hence in this case, the integrand is
F(x ,y,yx) = 1 + (yx)2
(y1
y) .
From the general theory, we know that the extremizing solution satisfies the Euler
Lagrange equation. Note that the multiplicative constant factor 1/
2gshould not affect
the extremizing solution path; indeed it divides out of the EulerLagrange equations.
Noting that the integrand Fdoes not explicitly depend onx, then the alternative form
for the EulerLagrange equation may be easier:
F
x d
dx
F yx F
yx
= 0.
This immediately implies that for some constantc, the EulerLagrange equation is
F
yx
F
yx=c.
Now substituting the form for Finto this gives
1 + (yx)
2 12
(y1 y) 12 yx
yx
1 + (yx)
2 12
(y1 y) 12
=c
1 + (yx)2 12
(y1 y) 12 yx
12 2 yx
(y1 y) 12
1 + (yx)2 12
=c
1 + (yx)2 12
(y1 y) 12 (yx)
2
(y1 y) 12
1 + (yx)2
1
2
=c
1 + (yx)2
(y1 y) 12
1 + (yx)2 12
(yx)2
(y1 y) 12
1 + (yx)2 12
=c
1(y1 y)
1 + (yx)2
=c2.We can now rearrange this equation so that
(yx)2 =
1
c2(y1 y) 1 (yx)2 =
1 c2y1 c2yc2y1 c2y .
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
9/56
Lagrangian and Hamiltonian mechanics 9
If we seta = y1 and b = 1c2
y1, then this equation becomes
yx =
b + y
a y
12
.
To find the solution to this ordinary differential equation we make the change of variable
y= 12 (a b) 12 (a + b)cos .
If we substitute this into the ordinary differential equation above and use the chain
rule we get
12 (a + b)sin
d
dx =
1 cos 1 + cos
12
.
Now we use that 1/(d/dx) = dx/d, and that sin = 1 cos2
, to get
dx
d = 12 (a + b)sin
1 + cos
1 cos 1
2
dxd
= 12 (a + b) (1 cos2 )1
2
1 + cos
1 cos 1
2
dxd
= 12 (a + b) (1 + cos )1
2 (1 cos ) 12
1 + cos
1 cos 1
2
dxd
= 12 (a + b)(1 + cos ).
We can directly integrate this last equation to findx as a function of. In other words
we can find the solution to the ordinary differential equation for y = y(x) above inparametric form, which with some minor rearrangement, can expressed as (here d is
an arbitrary constant of integration)
x + d= 12 (a + b)(+ sin ),
y+ b= 12 (a + b)(1 cos ).
This is the parametric representation of a cycloid.
5 Multivariable systems
We consider possible generalizations of functionals to be extremized. For more details
see for example Keener [9, Chapter 5].
5.1 Higher derivatives
Suppose we are asked to find the curve y = y(x) R that extremizes the functional
J(y):=
ba
F(x ,y,yx, yxx) dx,
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
10/56
10 Simon J.A. Malham
subject to y (a), y (b), yx(a) and yx(b) being fixed. Here the functional quantity to be
extremized depends on the curvatureyxxof the path. Necessarily the extremizing curvey satisfies the EulerLagrange equation (which is an ordinary differential equation)
F
y d
dx
F
yx
+
d2
dx2
F
yxx
= 0.
Note this follows by analogous arguments to those used for the classical variational
problem above.
5.2 Multiple dependent variables
Suppose we are asked to find the multidimensional curve y = y(x) RN that extrem-izes the functional
J(y):=
b
aF(x,y,yx) dx,
subject to y(a) and y(b) being fixed. Note x [a, b] but here y = y(x) is a curve inN-dimensional space and is thus a vector so that (here we use the notation d/dx)
y =
y1...
yN
and yx =
y1...
yN
.
Necessarily the extremizing curvey satisfies a set of EulerLagrange equations, which
are equivalent to a system of ordinary differential equations, given fori = 1, . . . , N by:
Fyi
ddx
Fy i
= 0.
5.3 Multiple independent variables
Suppose we are asked to find the field y = y(x) that, for x Rn, extremizes thefunctional
J(y):=
F(x, y, y) dx,
subject to y being fixed at the boundary of the domain . Note herey xyis simply the gradient ofy , i.e. it is the vector of partial derivatives ofy with respect
to each of the components xi (i= 1, . . . , n) ofx:
y =
y/x1...
y/xn
.
Necessarily y satisfies an EulerLagrange equation which is a partial differential equa-
tion given byF
y yxF = 0.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
11/56
Lagrangian and Hamiltonian mechanics 11
Here x is the usual divergence gradient operator with respect tox, and
yxF =
F/yx1...
F/yxn
,
where to keep the formula readable, withy the usual gradient of y, we have setyx y so thatyxi = (y)i = y/xi for i = 1, . . . , n.
Example (Laplaces equation)
The variational problem here is to find the field = (x1, x2, x3), for x= (x1, x2, x3)T
R3, that extremizes the mean-square gradient average
J():=
||2 dx.
In this case the integrand of the functionalJ is
F(x, , ) = ||2 (x1)2 + (x2)2 + (x3)2.
Note that the integrand F depends on the partial derivatives of only. Using the form
for the EulerLagrange equation above we get
x xF = 0
/x1/x2/x3
F/x1F/x2F/x3
= 0
x1
2x1
+
x2
2x2
+
x3
2x3
= 0
x1x1+ x2x2+ x3x3 = 0 2 = 0.
This is Laplaces equation for in the domain ; the solutions are called harmonicfunctions. Note that implicit in writing down the EulerLagrange partial differential
equation above, we assumed that was fixed at the boundary, i.e. Dirichlet bound-
ary conditions were specified.
Example (stretched vibrating string)
Suppose a string is tied between the two fixed points x = 0 and x = . Let y = y(x, t)
be the small displacement of the string at position x[0, ] and time t > 0 from theequilibrium positiony = 0. If is the uniform mass per unit length of the string which
is stretched to a tensionK, the kinetic and potential energy of the string are given by
T := 12
0
y2t dx, and V :=K
0
1 + (yx)
21/2 dx
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
12/56
12 Simon J.A. Malham
respectively, where subscripts indicate partial derivatives and the effect of gravity is
neglected. If the oscillations of the string are quite small, we can replace the expressionfor V , using the binomial expansion
1 + (yx)
21/2 1 + (yx)2, by
V = K
2
0
y2xdx.
We seek a solution y = y(x, t) that extremizes the functional (this is the action func-
tional as we see later)
J(y):=
t2t1
(T V) dt,
where t1 and t2 are two arbitrary times. In this case, with
x=
x
t
and y =
y/x
y/t
the integrand is
F(x, y, y) 12(yt)2 12K(yx)2,
i.e. F =F(y) only. The EulerLagrange equation is thus
x yxF = 0
/x
/t
F/yxF/yt
= 0
x
Kyx t yt = 0
c2 yxx ytt = 0,
where c2 = K/. This is the partial differential equation ytt = c2 yxx known as the
wave equation. It admits travelling wave solutions of speedc.
6 Lagrange multipliers
For the moment, let us temporarily put aside variational calculus and consider a prob-
lem in standard multivariable calculus.
6.1 Constrained optimization problem
Consider the following problem.
Find the stationary points of the scalar functionf(x) where x = (x1, . . . , xN)T
subject to the constraintsgk(x) = 0, wherek = 1, . . . , m, with m < N.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
13/56
Lagrangian and Hamiltonian mechanics 13
Note that the graph y = f(x) of the function f represents a hyper-surface in
(N + 1)-dimensional space. The constraints are given implicitly; each one also rep-resents a hyper-surface in (N+ 1)-dimensional space. In principle we could solve the
system ofmconstraint equations, say forx1, . . . , xmin terms of the remaining variables
xm+1, . . . , xN. We could then substitute these into f, which would now be a function
of (xm+1, . . . , xN) only. (We could solve the constraints for any subset ofm variables
xi and substitute those in if we wished or this was easier, or avoided singularities, and
so forth.) We would then proceed in the usual way to find the stationary points off
by considering the partial derivative off with respect to all the remaining variables
xm+1, . . . , xN, setting those partial derivatives equal to zero, and then solving that
system of equations. However solving the constraint equations may be very difficult,
and the method ofLagrange multipliersprovides an elegant alternative (see McCallumet. al. [14, Section 14.3]).
6.2 Method of Lagrange multipliers
The idea is to convert the constrained optimization problem to an unconstrained one
as follows. Form theLagrangian function
L(x,):=f(x) +mk=1
kgk(x),
where the parameter variables = (1, . . . , m)T are known as the Lagrange multipli-
ers. Note thatL is a function of both x and , i.e. ofN+ m variables in total. Thepartial derivatives ofL with respect to all of its dependent variables are:
Lxj
= f
xj+
mk=1
kgkxj
,
Lk
=gk,
wherej = 1, . . . , N and k = 1, . . . , m. At the stationary points ofL(x,), necessarily allthese partial derivatives must be zero, and we must solve the following unconstrained
problem:
f
xj+
mk=1
kgkxj
= 0,
gk = 0,
where j = 1, . . . , N and k = 1, . . . , m. Note we have N +m equations in N + m
unknowns. Assuming that we can solve this system to find a stationary point (x,)
ofL (there could be none, one, or more) then x is also a stationary point of theoriginal constrainedproblem. Recall: what is important about the formulation of theLagrangian functionL we introduced above, is that the given constraints mean that(on the constraint manifold) we haveL =f+ 0 and therefore the stationary points off andL coincide.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
14/56
14 Simon J.A. Malham
g=0
v
v
f=5
f=4
f=3
f=2
f=1
g
f
f
g
*(x , y )*
Fig. 3 At the constrained extremum f and g are parallel. (This is a rough reproductionof the figure on page 199 of McCallum et. al. [14])
6.3 Geometric intuition
Suppose we wish to extremize (find a local maximum or minimum) the objective func-
tion f(x, y) subject to the constraint g(x, y) = 0. We can think of this as follows. The
graph z = f(x, y) represents a surface in three dimensional (x ,y,z ) space, while the
constraint represents a curve in the x-y plane to which our movements are restricted.
Constrained extrema occur at points where the contours off are tangent to the
contours ofg(and can also occur at the endpoints of the constraint). This can be seen as
follows. At any point (x, y) in the plane fpoints in the direction of maximum increaseof f and thus perpendicular to the level contours of f. Suppose that the vector v is
tangent to the constraining curveg(x, y) = 0. If the directional derivativefv = f vis positive at some point, then moving in the direction ofv means thatfincreases (if
the directional derivative is negative then fdecreases in that direction). Thus at the
point (x, y) where fhas a constrained extremum we must havef v = 0 and sobothf andg are perpendicular to v and therefore parallel. Hence for some scalarparameter (the Lagrange multiplier) we have at the constrained extremum:
f = g and g = 0.Notice that here we have three equations in three unknowns x,y, .
7 Constrained variational problems
A common optimization problem is to extremize a functional Jwith respect to paths
y which are constrained in some way.
7.1 Constrained variational problem
We consider the following formulation.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
15/56
Lagrangian and Hamiltonian mechanics 15
Find the extrema of the functional
J(y):=
ba
F(x,y,yx) dx,
where y = (y1, . . . , yN)T RN subject to the set of m constraints, for k =
1, . . . , m < N :
Gk(x,y) = 0.
To solve this constrained variational problem we generalize the method of Lagrange
multipliers as follows. Note that them constraint equations above imply
b
a
k(x) Gk(x,y) dx= 0,
for each k = 1, . . . , m. Note that thek s are the Lagrange multipliers, which with the
constraints expressed in this integral form, can in general be functions ofx. We now
form the equivalent of the Lagrangian function which here is the functional
J(y,):=
ba
F(x,y,yx) dx +
mk=1
ba
k(x) Gk(x,y) dx
J(y,) = ba
F(x,y,yx) +
mk=1
k(x) Gk(x,y)
dx.
The integrand of this functional is
F(x,y,yx,):=F(x,y,yx) +
mk=1
k(x) Gk(x,y),
where = (1, . . . , m)T. We know from the classical variational problem that if (y,)
extremize Jthen necessarily they must satisfy the EulerLagrange equations:
F
yi d
dx
F
y i
= 0,
F
k d
dx
F
k
= 0,
which simplify to
F
yi d
dx
F
y i
= 0,
Gk(x,y) = 0,
for i = 1, . . . , N and k = 1, . . . , m. This is a system ofdifferential-algebraic equations:
the first set of relations are ordinary differential equations, while the constraints are
algebraic relations.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
16/56
16 Simon J.A. Malham
(x,y)=(x(t),y(t))c:
Fig. 4 For the isoperimetrical problem, the closed curve C has a fixed length , and the goalis to choose the shape that maximizes the area it encloses.
7.2 Integral constraints
If the constraints are (already) in integral form so that we have ba
Gk(x,y) dx= 0,
for k = 1, . . . , m, then set
J(y):=J(y) +
mk=1
k
ba
Gk(x,y) dx
=
ba
F(x,y,yx) +
mk=1
kGk(x,y)
dx.
Note we now have a variational problem with respect to the curve y = y(x) but the
constraints are classical in the sense that the Lagrangian multipliers here are simplyvariables. Proceeding as above, with this hybrid variational constraint problem, we
generate the EulerLagrange equations as above, together with the integral constraint
equations.
7.3 Example (Didos/isoperimetrical problem)
The goal of this classical constrained variational problem is to find the shape of the
closed curve, of a given fixed length , that encloses the maximum possible area.
Suppose that the curve is given in parametric coordinates
x(), y()
where the pa-
rameter [0, 2]. By Stokes theorem, the area enclosed by a closed contourC is
12C x dy y dx.
Hence our goal is to extremize the functional
J(x, y):= 12
20
(xy yx) d,
subject to the constraint 20
x2 + y2 d =.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
17/56
Lagrangian and Hamiltonian mechanics 17
Note that the constraint can be expressed in standard integral form as follows. Since
is fixed we have 20
12 d=.
Hence the constraint can be expressed in the form 20
x2 + y2 12 d= 0.
To solve this variational constraint problem we use the method of Lagrange multipliers
and form the functional
J(x, y):=J(x, y) +
20
x2 + y2 12 d
= 20
F(x,y, x, y, ) d .
where the integrand F is given by
F(x,y, x, y, ) = 12 (xy yx) + (x2 + y2)1
2 12 .
Note there are two dependent variables x and y (here the parameter is the inde-
pendent variable). By the theory above we know that the extremizing solution (x, y)
necessarily satisfies an EulerLagrange system of equations, which are the pair of or-
dinary differential equations
F
x d
d
F
x
= 0,
Fy
dd
Fy
= 0,
together with the integral constraint condition. Substituting the form for Fabove, the
pair of ordinary differential equations are
12y
d
d
12y+
x
(x2 + y2)1
2
= 0,
12x d
d
12x +
y
(x2 + y2)1
2
= 0.
Integrating both these equations with respect towe get
y x
(x2 + y2)12 =c2 and x
y
(x2 + y2)12 =c1,
for arbitrary constants c1 and c2. Combining these last two equations reveals
(x c1)2 + (y c2)2 = 2y2
x2 + y2 +
2x2
x2 + y2 =2.
Hence the solution curve is given by
(x c1)2 + (y c2)2 =2,
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
18/56
18 Simon J.A. Malham
which is the equation for a circle with radius and centre (c1, c2). The constraint
condition implies = /2 and c1 and c2 can be determined from the initial or endpoints of the closed contour/path.
The isoperimetrical problem has quite a history. It was formulated in Virgils poem
the Aeneid, one account of the beginnings of Rome; see Wikipedia [18] or Mont-
gomery [16]. Quoting from Wikipedia (Dido was also known as Elissa):
Eventually Elissa and her followers arrived on the coast of North Africa where
Elissa asked the local inhabitants for a small bit of land for a temporary refuge
until she could continue her journeying, only as much land as could be en-
compassed by an oxhide. They agreed. Elissa cut the oxhide into fine strips
so that she had enough to encircle an entire nearby hill, which was therefore
afterwards named Byrsa hide. (This event is commemorated in modern math-
ematics: The isoperimetric problem of enclosing the maximum area within
a fixed boundary is often called the Dido Problem in modern calculus ofvariations.)
Dido found the solutionin her case a half-circleand the semi-circular city of
Carthage was founded.
Example (Helmholtzs equation)
This is a constrained variational version of the problem that generated the Laplace
equation. The goal is to find the field = (x) that extremizes the functional
J():=
||2 dx,
subject to the constraint
2 dx= 1.
This constraint corresponds to saying that the total energy is bounded and in fact
renormalized to unity. We assume zero boundary conditions, (x) = 0 for x Rn. Using the method of Lagrange multipliers (for integral constraints) we form the
functional
J():=
||2 dx +
2 dx 1
.
We can re-write this in the form
J():= |
|2 + 2
1
|| dx,where|| is the volume of the domain . Hence the integrand in this case is
F(x, , , ):= ||2 +
2 1||
.
The extremzing solution satisfies the EulerLagrange partial differential equation
F
xF = 0,
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
19/56
Lagrangian and Hamiltonian mechanics 19
together with the constraint equation. Note, directly computing, we have
xF = 2 and F
= 2.
Substituting these two results into the EulerLagrange partial differential equation we
find
2 (2) = 0 2 = .
This is Helmholtzs equation on. The Lagrange multiplier also represents an eigen-
value parameter.
8 Optimal linear-quadratic control
We can use calculus of variations techniques to derive the solution to an important
problem in control theory. Suppose that a system state at any time t 0 is recorded
in the vector q = q(t). Suppose further that the state evolves according to a linear
system of differential equations and we can control the system via a set of inputs orcontrolsu= u(t), i.e. the system evolution is
dq
dt =Aq + Bu.
Here A is a matrix, which for convenience we will assume is constant, and B = B(t) is
another matrix mediating the controls.
Goal: The goal is, starting in the state q(0) = q0, to bring this initial state to a final
state q(T) at time t = T >0, as expediently as possible.
There are many criteria as to what constitutes expediency. Here we will measure
the cost on the system of our actions over [0, T] by the quadratic utility
J(u):=
T0
qT(t)C(t)q(t) + uT(t)D(t)u(t) dt + qT(T)Eq(T).
Here C = C(t), D = D(t) and Eare non-negative definite symmetric matrices. The
final term represents a final state achievement cost. Note that we can in principle solve
the system of linear differential equations above for the stateq = q(t) in terms of the
controlu= u(t) so that J=J(u) is a functional ofu only (we see this presently).
Thus our goal is to find the control u = u(t) that minimizes the cost J =J(u)
whilst respecting the constraint which is the linear evolution of the system state. We
proceed as before. Suppose u = u(t) exists. Consider perturbations to u on [0, T]of the form
u= u+ u.
Changing the control/input changes the stateq= q(t) of the system so that
q= q+ q.
where we suppose here q = q(t) to be the system evolution corresponding to the
optimal control u = u(t). Note that linear system perturbations q are linear in .
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
20/56
20 Simon J.A. Malham
Substituting these last two perturbation expressions into the differential system for the
state evolution we find dq
dt =Aq + Bu,
where we have used that (d/dt)q = Aq+ Bu. The initial condition is q(0) = 0.
We can solve this system of differential equations forq in terms ofu by the variation
of constants formula (using an integrating factor) as follows. Since the matrix A is
constant we observe
d
dt
exp(At)q(t) = exp(At)B(t)u(t)
exp(At)q(t) = t0
exp(As)B(s)u(s) ds
q(t) =
t
0
expA(t s)B(s)u(s) ds.By the calculus of variations, if we set
():=J(u+ u),
then if the functional Jhas a minimum we have, for all u,
(0) = 0.
Note that we can substitute our expression forq(t) in terms ofuintoJ(u+u). Since
u= u+ uand q = q+ q are linear in and the functional J= J(u) is quadratic
in u and q , then J(u+ u) must be quadratic in so that
() = 0+ 1+ 22,
for some functionals0 = 0(u), 1 = 1(u) and 2 = 2(u) independent of. Since
() = 1 + 22, we see (0) = 1. This term in () = J(u + u) by direct
computation is thus
(0) = 2
T0
qT(t)C(t)q(t) +u
T(t)D(t)u(t) dt + 2qT(T)Eq(T),
where we used thatC, D and Eare symmetric. We now substitute our expression for
q in terms ofu above into this formula for(0), this gives
12
(0) =
T0
qT(t)C(t)q(t) dt +
T0
uT(s)D(s)u(s) ds + q
T(T)Eq(T)
= T0
t0
exp
A(t s)B(s)u(s)TC(t)q(t) +uT(s)D(s)u(s) ds dt+
T0
exp
A(T s)B(s)u(s)T ds Eq(T)=
T0
Ts
exp
A(t s)B(s)u(s)TC(t)q(t) +uT(s)D(s)u(s) dt ds+
T0
exp
A(T s)B(s)u(s)T ds Eq(T),
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
21/56
Lagrangian and Hamiltonian mechanics 21
where we have swapped the order of integration in the first term. Now we use the
transpose of the product of two matrices is the product of their transposes in reverseorder to get
12
(0) =
T0
uT(s)BT(s)
Ts
exp
AT(t s)C(t)q(t) dt +uT(s)D(s)u(s)+u(s)TBT(s)exp
AT(T s)Eq(T)
ds
=
T0
uT(s)BT(s)p(s) +uT(s)D(s)u(s) ds,
where for all s [0, T] we set
p(s):= T
s
expAT(t s)C(t)q(t) dt + expA
T(T s)Eq(T).We see the condition for a minimum, (0) = 0, is equivalent to T
0uT(s)
BT(s)p(s) + D(s)u(s)
ds= 0,
for all u. Hence a necessary condition for the minimum is that for all t [0, T] we haveu(t) = D1(t)BT(t)p(t).
Note that p depends solely on the optimal state q. To elucidate this relationship
further, note by definition p(T) = Eq(T) and differentiating p with respect to t we
finddp
dt = Cq ATp.
Using the expression for the optimal control u in terms of p(t) we derived, we seethat q and p satisfy the system of differential equations
d
dt
qp
=
A BD1(t)BT
C(t) AT
qp
.
Define S= S(t) to be the map S: q p, i.e. so that p(t) =S(t)q(t). Thenu(t) = D1(t)BT(t)S(t)q(t),
and S(t) we see characterizes the optimal current state feedback control, it tells how
to choose the current optimal controlu(t) in terms of the current state q(t). Finally
we observe that sincep(t) =S(t)q(t), we have
dp
dt =
dS
dtq+ S
dqdt
.
Thus we see that
dS
dtq =
dp
dt Sdq
dt
= (Cq ATSq) S(Aq BD1ATSq).Hence S= S(t) satisfies theRiccati equation
dS
dt = C ATS SA SBD1ATS.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
22/56
22 Simon J.A. Malham
Remark 1 We can easily generalize the argument above to the case when the coefficient
matrix A = A(t) is not constant. This is achieved by carefully replacing the flow mapexp
A(t s) for the linear constant coefficient system (d/dt)q= Aq, by the flow mapfor the corresponding linear nonautonomous system withA = A(t), and carrying that
through the rest of the computation.
9 Lagrangian dynamics
9.1 Newtons equations
Consider a system ofN particles in three dimensional space, each with position vector
ri(t) for i = 1, . . . , N . Note that each ri(t) R3 is a 3-vector. We thus need 3Ncoordinates to specify the system, this is the configuration space. Newtons 2nd law
tells us that the equation of motion for the ith particle is
pi = Fexti + F
coni ,
for i = 1, . . . , N . Here pi = mivi is the linear momentum of the ith particle and
vi = ri is its velocity. We decompose the total force on theith particle into an external
force Fexti and a constraint force Fconi . By external forces we imagine forces due to
gravitational attraction or an electro-magnetic field, and so forth.
9.2 Holonomic constraints
By a constraint on a particles we imagine that the particles motion is limited in some
rigid way. For example the particle/bead may be constrained to move along a wire or
its motion is constrained to a given surface. If the system ofNparticles constitute a
rigid body, then the distances between all the particles are rigidly fixed and we have
the constraint
|ri(t) rj(t)| =cij ,
for some constants cij , for all i, j = 1, . . . , N . All of these are examples ofholonomicconstraints.
Definition 1 (Holonomic constraints) For a system of particles with positions given
byri(t) for i = 1, . . . , N , constraints that can be expressed in the form
g(r1, . . . , rN, t) = 0,
are said to be holonomic. Note they only involve the configuration coordinates.
We willonlyconsider systems for which the constraints are holonomic. Systems withconstraints that are non-holonomic are: gas molecules in a container (the constraint
is only expressible as an inequality); or a sphere rolling on a rough surface without
slipping (the constraint condition is one of matched velocities).
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
23/56
Lagrangian and Hamiltonian mechanics 23
9.3 Degrees of freedom
Let us suppose that for the Nparticles there are m holonomic constraints given by
gk(r1, . . . , rN, t) = 0,
for k = 1, . . . , m. The positions ri(t) of all Nparticles are determined by 3N coordi-
nates. However due to the constraints, the positions ri(t) are not all independent. In
principle, we can use themholonomic constraints to eliminatem of the 3N coordinates
and we would be left with 3N m independent coordinates, i.e. the dimension of theconfiguration space is actually 3N m.
Definition 2 (Degrees of freedom) The dimension of the configuration space is called
the number of degrees of freedom, see Arnold [3, Page 76].
Thus we can transform from the old coordinates r1, . . . , rN to new generalized
coordinates q1, . . . , q n where n = 3N m:
r1 = r1(q1, . . . , q n, t),
...
rN = rN(q1, . . . , q n, t).
9.4 DAlemberts principle
We will restrict ourselves to systems for which the net work of the constraint forces is
zero, i.e. we suppose
Ni=1
Fconi dri = 0,
for every small change dri of the configuration of the system (fortfixed). If we combine
this with Newtons 2nd law we see that we get
Ni=1
( pi Fexti ) dri = 0.
This isDAlemberts principle. In particular, no forces of constraint are present.
The assumption that the constraint force does no net work is quite general. It is
true in particular for holonomic constraints. For example, for the case of a rigid body,
the internal forces of constraint do no work as the distances |rirj | between particlesis fixed, then d(ri rj) is perpendicular to ri rj and hence perpendicular to theforce between them which is parallel to ri rj . Similarly for the case of the bead ona wire or particle constrained to move on a surfacethe normal reaction forces are
perpendicular to dri.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
24/56
24 Simon J.A. Malham
9.5 Lagranges equations (from DAlemberts principle)
In his Mecanique Analytique [1788], Lagrange sought a coordinate-invariant expres-
sion for mass times acceleration, see Marsden and Ratiu [15, Page 231]. This lead to
Lagranges equations of motion. Consider the transformation to generalized coordinates
ri = ri(q1, . . . , q n, t),
for i = 1, . . . , N . If we consider a small increment in the displacements dri then the
corresponding increment in the work done by the external forces is
Ni=1
Fexti dri =
N,ni,j=1
Fexti riqj dqj =
nj=1
Qjdqj .
Here we have set for j = 1, . . . , n,
Qj =Ni=1
Fexti riqj ,
and think of these as generalized forces.
Further, the change in kinetic energy mediated through the momentum (the first
term in DAlemberts principle), due to the increment in the displacements dri, is given
byNi=1
pi dri =Ni=1
mivi dri =N,ni,j=1
mivi riqj dqj .
From the product rule we know that
d
dtvi
ri
qj vi
ri
qj+ vi
d
dtri
qj
vi riqj + vi viqj
.
Also, by differentiating the transformation to generalized coordinates (keeping t fixed),
we see that
vin
j=1
riqj
qj viqj riqj
.
Using these last two identities we see that
Ni=1
pi dri =n
j=1
Ni=1
mivi riqj
dqj
=nj=1
Ni=1
ddt
mivi ri
qj
mivi vi
qj
dqj
=
nj=1
Ni=1
d
dt
mivi vi
qj
mivi vi
qj
dqj
=
nj=1
d
dt
qj
Ni=1
12mi|vi|2
qj
Ni=1
12mi|vi|2
dqj .
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
25/56
Lagrangian and Hamiltonian mechanics 25
Hence if the kinetic energy is defined to be
T :=
Ni=1
12mi|vi|2,
then we see that DAlemberts principle is equivalent to
nj=1
d
dt
T
qj
T
qj Qj
dqj = 0.
Since the qj for j = 1, . . . , n, where n = 3N m, are all independent, we have
d
dt
T
qj
T
qj Qj = 0,
for j = 1, . . . , n. If we now assume that the work done depends on the initial and final
configurations only and not on the path between them, then there exists a potential
function V(q1, . . . , q n) such that
Qj = Vqj
for j = 1, . . . , n (such forces are said to be conservative). If we define the Lagrange
function or Lagrangian to be
L= T V,then we see that DAlemberts principle is equivalent to
d
dt
L
qj
L
qj= 0,
for j = 1, . . . , n. These are known as Lagranges equations. As already noted the n-
dimensional subsurface of 3N-dimensional space, on which the solutions to Lagranges
equations lie (wheren = 3N m), is called the configuration space. It is parameterizedby the n generalized coordinates q1, . . . , q n.
Remark 2 (Non-conservative forces) If the system has non-conservative forces it may
still be possible to find a generalized potential function V such that
Qj = Vqj
+ d
dt
V
qj
,
for j = 1, . . . , n. From such potentials we can still deduce Lagranges equations of
motion. Examples of such generalized potentials are velocity dependent potentials due
to electro-magnetic fields, for example the Lorentz force on a charged particle.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
26/56
26 Simon J.A. Malham
10 Hamiltons principle
10.1 Action
We consider mechanical systems with holonomic constraints and with all other forces
conservative. Recall, we define theLagrange function or Lagrangian to be
L= T V,where
T =
Ni=1
12mi|vi|2
is the total kinetic energy for the system, andVis its potential energy.
Definition 3 (Action)If the LagrangianL is the difference of the kinetic and potential
energies for a system, i.e.L = T V, we define the action A from time t1 to t2 to bethe functional
A(q):=
t2t1
L(q, q, t) dt,
where q= (q1, . . . , q n)T.
10.2 Hamiltons principle of least action
Hamilton [1834] realized that Lagranges equations of motion were equivalent to a
variational principle (see Marsden and Ratiu [15, Page 231]).
Theorem 2 (Hamiltons principle of least action) The correct path of motion of a
mechanical system with holonomic constraints and conservative external forces, from
time t1 to t2, is a stationary solution of the action. Indeed, the correct path of mo-tion q = q(t), with q = (q1, . . . , q n)
T, necessarily and sufficiently satisfies Lagranges
equations of motiond
dt
L
qj
L
qj= 0,
forj = 1, . . . , n.
Quoting from Arnold [3, Page 60], it is Hamiltons form of the principle of least
action because in many cases the action ofq= q(t) is not only an extremal but also
a minimum value of the action functional.
10.3 Example (Simple harmonic motion)
Consider a particle of massm moving in a one dimensional Hookeian force fieldkx,where k is a constant. The potential function V = V(x) corresponding to this force
field satisfies
Vx
= kx
V(x) V(0) = x0
k d
V(x) = 12x2.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
27/56
Lagrangian and Hamiltonian mechanics 27
The Lagrangian L = T V is thus given by
L(x, x) = 12mx2 12kx2.
From Hamiltons principle the equations of motion are given by Lagranges equations.
Here, taking the generalized coordinate to be q= x, the single Lagrange equation is
d
dt
L
x
L
x = 0.
Substituting the form for the Lagrangian above this Lagrange equation becomes
mx + kx = 0.
10.4 Example (Kepler problem)
Consider a particle of mass m moving in an inverse square law force field,m/r2,such as a small planet or asteroid in the gravitational field of a star or larger planet.
Hence the corresponding potential function satisfies
Vr
= mr2
V() V(r) = r
m
2 d
V(r) =m
r .
The Lagrangian L = T V is thus given by
L(r,r,, ) = 12m(r2 + r22) +
m
r .
From Hamiltons principle the equations of motion are given by Lagranges equations,
which here, taking the generalized coordinates to beq1 = r and q2= , are the pair of
ordinary differential equations
d
dt
L
r
L
r = 0,
ddt
L
L
= 0,
which on substituting the form for the Lagrangian above, become
mr mr2 + mr2
= 0,
d
dt
mr2
= 0.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
28/56
28 Simon J.A. Malham
10.5 Non-uniqueness of the Lagrangian
Two LagrangiansL1 and L2 that differ by the total time derivative of any function of
q= (q1, . . . , q n)T and t generate the same equations of motion. In fact if
L2(q, q, t) = L1(q,q, t) + d
dt
f(q, t)
,
then for j = 1, . . . , n direct calculation reveals that
d
dt
L2qj
L2
qj=
d
dt
L1qj
L1
qj.
11 Constraints
Given a Lagrangian L(q, q, t) for a system, suppose we realize the system has some
constraints (so theqj are not all independent).
11.1 Holonomic constraints
Suppose we havem holonomic constraints of the form
Gk(q1, . . . , q n, t) = 0,
for k = 1, . . . , m < n. We can now use the method of Lagrange multipliers with
Hamiltons principle to deduce that the equations of motion are given by
ddt
Lqj
L
qj=
mk=1
k(t) Gk
qj,
Gk(q1, . . . , q n, t) = 0,
for j = 1, . . . , n and k = 1, . . . , m. We call the quantities on the right above
mk=1
k(t)Gkqj
,
the generalized forces of constraint.
11.2 Example (simple pendulum)
Consider the motion of a simple pendulum bob of mass m that swings at the end of
a light rod of length a. The other end is attached so that the rod and bob can swing
freely in a plane. Ifg is the acceleration due to gravity, then the Lagrangian L = TVis given by
L(r,r,, ) = 12m(r2 + r22) + mgr cos ,
together with the constraint
r a= 0.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
29/56
Lagrangian and Hamiltonian mechanics 29
r
Fig. 5 The mechanical problem for the simple pendulum, can be thought of as a particle ofmass m moving in a vertical plane, that is constrained to always be a distance a from a fixedpoint. In polar coordinates, the position of the mass is (r, ) and the constraint is r= a.
We could just substituter = ainto the Lagrangian, obtaining a system with one degree
of freedom, and proceed from there. However, we will consider the system as one with
two degrees of freedom,q1 = r and q2 = , together with a constraintG(r) = 0, where
G(r) =r a. Hamiltons principle and the method of Lagrange multipliers imply thatthe system evolves according to the pair of ordinary differential equations together
with the algebraic constraint given by
d
dt
L
r
L
r =
G
r,
d
dt
L
L
=
G
,
G= 0.
Substituting the form for the Lagrangian above, the two ordinary differential equations
together with the algebraic constraint become
mr mr2 mg cos = ,d
dt
mr2
+ mgr sin = 0,
r a= 0.
Note that the constraint is of courser = a, which implies r = 0. Using this, the system
of differential algebraic equations thus reduces to
ma2+ mga sin = 0,
which comes from the second equation above. The first equation tells us that the
Lagrange multiplier is given by
(t) = ma2 mg cos .
The Lagrange multiplier has a physical interpretation, it is the normal reaction force,
which here is the tension in the rod.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
30/56
30 Simon J.A. Malham
Remark 3 (Non-holonomic constraints)Mechanical systems with certain types of non-
holonomic constraints can also be treated, in particular constraints of the formn
j=1
A(q, t)kjqj + bk(q, t) = 0,
for k = 1, . . . , m, where q = (q1, . . . , q n)T. Note the assumption is that these equa-
tions are not integrable, in particular not exact, otherwise the constraints would be
holonomic.
12 Hamiltonian dynamics
12.1 Hamiltonian
We consider mechanical systems that are holonomic and and conservative (or for which
the applied forces have a generalized potential). For such a system we can construct
a Lagrangian L(q, q, t), where q = (q1, . . . , q n)T, which is the difference of the total
kinetic T and potential Venergies. These mechanical systems evolve according to the
n Lagrange equationsd
dt
L
qj
L
qj= 0,
forj = 1, . . . , n. These are each second order ordinary differential equations and so the
system is determined for all time once 2n initial conditionsq(t0),q(t0)
are specified
(or n conditions at two different times). The state of the system is represented by a
point q= (q1, . . . , q n)T in configuration space.
Definition 4 (Generalized momenta) We define the generalized momenta for a La-grangian mechanical system to be
pj = L
qj,
for j = 1, . . . , n. Note that we have pj =pj(q,q, t) in general, where q = (q1, . . . , q n)T
and q= (q1, . . . , qn)T.
In terms of the generalized momenta, Lagranges equations become, forj = 1, . . . , n:
pj = L
qj.
Further, in principle, we can solve the relations above which define the generalized
momenta, to find functional expressions for the qj in terms ofqi,piandt. In other wordswe can solve the relations defining the generalized momenta to find qj = qj(q,p, t)
where q= (q1, . . . , q n)T and p= (p1, . . . , pn)
T.
Definition 5 (Hamiltonian)We define theHamiltonian functionas theLegendre trans-
formof the Lagrangian function, i.e. the Hamiltonian is defined by
H(q,p, t):= q p L(q,q, t),
where q= (q1, . . . , q n)T and p= (p1, . . . , pn)
T and we suppose q= q(q,p, t).
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
31/56
Lagrangian and Hamiltonian mechanics 31
Note that in this definition we used the notation for the dot product
q p=n
j=1
qjpj .
The Legendre transform is nicely explained in Arnold [3, Page 61] and Evans [7,
Page 121]. In practice, as far as solving example problems herein, it simply involves the
task of solving the equations for the generalized momenta above, to find qj in terms of
qi, pi and t.
12.2 Hamiltons equations of motion
From Lagranges equation of motion, we can, using the definitions for the generalized
momenta
pj = L
qj,
and Hamiltonian
H=
nj=1
qjpj L,
above, deduce Hamiltons equations of motion.
Theorem 3 (Hamiltons equations of motion) With the Hamiltonian defined as the
Legendre transform of the Lagrangian, Lagranges equations of motion imply
qi = H
pi,
pi = Hqi ,
fori = 1, . . . , n. These are Hamiltons canonical equations, consisting of2n first orderequations of motion.
ProofUsing the chain rule and the definition for the generalized momenta we have
H
pi=
nj=1
qjpi
pj+ qi n
j=1
L
qj
qjpi
qi,
and
Hqi=
nj=1
qjqipj Lqi
nj=1
Lqjqjqi
Lqi .
Now using Lagranges equations, in the form pj =L/qj , the last relation reveals
pi = Hqi
,
for i = 1, . . . , n. Collecting these relations together, we see that Lagranges equations
of motion imply Hamiltons canonical equations as shown.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
32/56
32 Simon J.A. Malham
Two further observations are also useful. First, if the Lagrangian L = L(q, q) is
independent of explicitt, then when we solve the equations that define the generalizedmomenta we find q= q(q,p). Hence we see that
H= q(q,p) p Lq, q(q,p),i.e. the Hamiltonian H= H(q,p) is also independent oft explicitly. Second, using the
chain rule and Hamiltons equations we see that
dH
dt =
ni=1
H
qiqi+
ni=1
H
pipi+
H
t
dHdt
= H
t .
Hence ifHdoes notexplicitly depend on t then
H is a
constant of the motion,
conserved quantity,
integral of the motion.
Hence the absence of explicit t dependence in the Hamiltonian H could serve as a
more general definition of a conservative system, though in general H may not be
the total energy. However for simple mechanical systems for which the kinetic energyT = T(q,q) is a homogeneous quadratic function in q, and the potential V = V(q),
then the Hamiltonian H will bethe total energy. To see this, suppose
T =
n
i,j=1cij(q) qiqj ,
i.e. a homogeneous quadratic function in q; necessarilycij =cji . Then we have
T
qk=
nj=1
ckj(q) qj+
ni=1
cik(q) qi = 2
ni=1
cik(q) qi
which impliesn
k=1
qkT
qk= 2T.
Thus the HamiltonianH= 2T (T V) = T+ V, i.e. the total energy.
13 Hamiltonian formulation
13.1 Summary
To construct Hamiltons canonical equations for a mechanical system proceed as fol-
lows:
1. Choose your generalized coordinates q= (q1, . . . , q n)T and construct
L(q, q, t) = T V.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
33/56
Lagrangian and Hamiltonian mechanics 33
2. Define and compute the generalized momenta
pi = L
qi,
for i = 1, . . . , n. Solve these relations to find qi = qi(q,p, t).
3. Construct and compute the Hamiltonian function
H=
nj=1
qjpj L,
4. Write down Hamiltons equations of motion
qi = H
pi,
pi = H
qi ,
for i = 1, . . . , n, and evaluate the partial derivatives of the Hamiltonian on the
right.
13.2 Example (simple harmonic oscillator)
The Lagrangian for the simple harmonic oscillator, which consists of a mass m moving
in a quadratic potential field with characteristic coefficient k , is
L(x, x) = 12mx2 12kx2.
The corresponding generalized momentum is
p= L
x =mx
which is the usual momentum. This implies x= p/m and so the Hamiltonian is given
by
H(x, p) = x p L(x, x)
= p
mp
12mx
2 12kx2
= p2
m
12m
p
m
2 12kx2
=
1
2
p2
m +
1
2kx
2
.
Note that is last expression is just the sum of the kinetic and potential energies and so
His the total energy. Hamiltons equations of motion are thus given by
x= H/p,
p= H/x, x= p/m,
p= kx.Note that combining these two equations, we get the usual equation for a harmonic
oscillator: mx= kx.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
34/56
34 Simon J.A. Malham
Fig. 6 The mechanical problem for the simple harmonic oscillator consists of a particle movingin a quadratic potential field. As shown, we can think of a ball of mass m sliding freely back andforth in a vertical plane, without energy loss, inside a parabolic shaped bowl. The horizontalposition x(t) is its displacement.
13.3 Example (Kepler problem)
Recall the Kepler problem for a mass m moving in an inverse-square central force field
with characteristic coefficient. The Lagrangian L = T V isL(r,r,, ) = 12m(r
2 + r22) +m
r .
Hence the generalized momenta are
pr = L
r =mr and p =
L
=mr2.
These imply r= pr/m and = p/mr2 and so the Hamiltonian is given by
H(r,,pr , p) = r pr+ p L(r,r,, )
= 1
m
p2r+
p2r2
12m
p2rm2
+ r2 p2m2r4
+
m
r
= 1
2m
p2r+
p2r2
m
r ,
which in this case is also the total energy. Hamiltons equations of motion are
r= H/pr ,
= H/p,
pr = H/r,
p = H/,
r= pr/m,
= p/mr2,
pr =p2/mr
3 m/r2,
p = 0.Note that p = 0, i.e. we have that p is constant for the motion. This property
corresponds to theconservation of angular momentum.
Remark 4 The LagrangianL= TVmay change its functional form if we use differentvariables (Q, Q) instead of (q,q), but its magnitude will not change. However, the
functional form and magnitude of the Hamiltonian both depend on the generalized
coordinates chosen. In particular, the HamiltonianHmay be conserved for one set of
coordinates, but not for another.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
35/56
Lagrangian and Hamiltonian mechanics 35
m
0QUt
q
Fig. 7 Mass-spring system on a massless cart.
13.4 Example (harmonic oscillator on a moving platform)
Consider a mass-spring system, massmand spring stiffnessk, contained within a mass-
less cart which is translating horizontally with a fixed velocity Usee Figure 7. The
constant velocity Uof the cart is maintained by an external agency. The Lagrangian
L= T V for this system is
L(q, q, t) = 12mq2 12k(q U t)2.
The resulting equation of motion of the mass is
mq= k(q Ut).
If we setQ = q
U t, then the equation of motion is
m Q= kQ.
Let us now consider the Hamiltonian formulation using two different sets of coordinates.
First, using the generalized coordinateq, the corresponding generalized momentum
is p = mqand the Hamiltonian is
H(q ,p ,t) = p2
2m+
k
2(q U t)2.
Here the HamiltonianHisthe total energy, but it isnotconserved (there is an external
energy input maintaining U constant).
Second, using the generalized coordinateQ, the Lagrangian L= T V is
L(Q, Q, t) = 12mQ2 + m QU+ 12mU
2 12kQ2.
Here, the generalized momentum is P =m Q + mUand the Hamiltonian is
H(Q, P) = (P mU)2
2m +
k
2Q2 m
2U2.
Note that Hdoes not explicitly depend on t . Hence H isconserved, but it is not the
total energy.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
36/56
36 Simon J.A. Malham
14 Symmetries and conservation laws
14.1 Cyclic coordinates
We have already seen that
dH
dt =
H
t .
Hence if the Hamiltonian does not depend explicitly on t, then it is aconstantor integral
of the motion; sometimes called Jacobis integral. Itmaybe the total energy. Further
from the definition of the generalized momenta pi = L/qi, Lagranges equations,
and Hamiltons equations for the generalized momenta, we have
pi =
H
qi
= L
qi
.
From these relations we can see that ifqi is explicitly absent from the Lagrangian L,
then it is explicitly absent from the Hamiltonian H, and
pi = 0.
Hence pi is a conserved quantity, i.e. constant of the motion. Such aqi is calledcyclic
or ignorable. Note that for such coordinates qi, the transformation
t t + t,qi qi+ qi,
leave the Lagrangian/Hamiltonian unchanged. This invariance signifies a fundamental
symmetryof the system.
14.2 Example (Kepler problem)
Recall, the LagrangianL = T V for the Kepler problem is
L(r,r,, ) = 12m(r2 + r22) +
m
r ,
and the Hamiltonian is
H(r,,pr , p) = 1
2m
p2r+
p2r2
m
r .
Note that L, Hare independent oft and therefore H is conserved (and here it is the
total energy). Further,L,Hare independent of and thereforep = mr2is conserved
also. The original problem has two degrees of freedom (r, ). However, we have just
established two integrals of the motion, namelyH and p. This system is said to be
integrable.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
37/56
Lagrangian and Hamiltonian mechanics 37
mg
a
Fig. 8 Simple axisymmetric top: this consists of a body of massm that spins around its axis of
symmetry moving under gravity. It has a fixed point on its axis of symmetryhere this is thepivot at the narrow pointed end that touches the ground. The centre of mass is a distance afrom the fixed point. The configuration of the top is given in terms of the Euler angles ( ,,)shown. Typically the top also precesses around the vertical axis = 0.
14.3 Example (axisymmetric top)
Consider a simple axisymmetric top of massm with a fixed point on its axis of sym-
metry; see Figure 8. Suppose the centre of mass is a distance a from the fixed point,
and the principle moments of inertia are A = B=C. We assume there are no torquesabout the symmetry or vertical axes. The configuration of the top is given in terms of
the Euler angles (,,) as shown in Figure 8. The LagrangianL = T V is
L= 12A(2 + 2 sin2 ) + 12C(+ cos )2 mga cos .The generalized momenta are
p = A,
p = A sin2 + Ccos (+ cos ),
p = C(+ cos ).
Using these the Hamiltonian is given by
H= p22A
+(p pcos )2
2A sin2 +
p22C
+ mga cos .
We see that L, H are both independent of t, and . Hence H, p and p areconserved. Respectively, they represent the total energy, the angular momentum about
the symmetry axis and the angular momentum about the vertical.
Remark 5 (Noethers Theorem) To accept only those symmetries which leave the
Lagrangian unchanged is needlessly restrictive. When searching for conservation laws
(integrals of the motion), we can in general consider transformations that leave the
action integralinvariant enough so that we get the same equations of motion. This is
the idea underlying (Emmy) Noethers theorem; see Arnold [3, Page 88].
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
38/56
38 Simon J.A. Malham
15 Poisson brackets
15.1 Definition and properties
Definition 6 (Poisson brackets) For two functions u = u(q,p) and v = v(q,p), where
q= (q1, . . . , q n)T and p= (p1, . . . , pn)
T, we define their Poisson bracket to be
[u, v]:=
ni=1
u
qi
v
pi u
pi
v
qi
.
The Poisson bracket satisfies some simple properties that can be checked directly. For
example, for any function u = u(q,p) and constant c we have [u, c] = 0. We also
observe that if v = v(q,p) and w = w(q,p) then [uv,w] = u[v, w] +v[u, w]. Three
crucial properties are summarized in the following lemma.
Lemma 2 (Lie algebra properties)The bracket satisfies the following properties for al lfunctionsu = u(q,p), v = v(q,p) andw= w(q,p) and scalars and:
1. Skew-symmetry: [v, u] = [u, v];2. Bilinearity:[u + v,w] =[u, w] + [v, w];
3. Jacobis identity:[u, [v, w]] + [v, [w, u]] + [w, [u, v]] = 0.
These three properties define a non-associative algebra known as aLie algebra.
Two further simple examples of Lie algebras are: the vector space of vectors equipped
with the wedge or vector product [u,v] = u v and the vector space of matricesequipped with the matrix commutator product [A, B] =AB BA.Corollary 1 (Properties from the definition)Using that all the coordinates(q1, . . . , q n)and(p1, . . . , pn)are independent, we immediately deduce by direct substitution into the
definition, the following results for anyu = u(q,p) and alli, j = 1, . . . , n:
u
qi= [u, pi],
u
pi= [u, qi], [qi, qj ] = 0, [pi, pj ] = 0 and [qi, pj ] =ij .
Hereij is the Kronecker delta which is zero wheni =j and unity wheni = j .Corollary 2 (Poisson bracket for canonical variables) If q = (q1, . . . , q n)
T and p =
(p1, . . . , pn)T are canonical Hamilton variables, i.e. they satisfy Hamiltons equations
for some HamiltonianH, then for any functionf=f(q,p) we have
df
dt =
f
t + [f, H].
ProofUsing Hamiltons equations of motion (in shorter vector notation) we know
dq
dt = pH and dp
dt = qH.
Thus by the chain rule and then Hamiltons equations we find
df
dt =
dq
dt qH+ dp
dt pH= pH qH+ (qH) pH
which is zero by the commutative property of the scalar (dot) product. All the properties above are essential for the next two main results of this section.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
39/56
Lagrangian and Hamiltonian mechanics 39
15.2 Invariance under canonical transformations
Remarkably the the Poisson bracket is invariant under canonical transformations. By
this we mean the following. Suppose we make a transformation of the canonical co-
ordinates (q,p), satisfying Hamiltons equations with respect to a HamiltonianH, to
new canonical coordinates (Q,P), satisfying Hamiltons equations with respect to a
Hamiltonian K, where
Q= Q(q,p) and P = P(q,p).
For two functions u = u(q,p) and v = v(q,p) define U =U(Q,P) and V =V(Q,P)
by the identities
U(Q,P) =u(q,p) and V(Q,P) =v(q,p).
Then we have the following resultwhose proof we leave as an exercise in the chain
rule! (Hint: start with [u, v]q,p and immediately substitute the definitions for U andV; also see Arnold [3, Page 216].)
Lemma 3 (Invariance under canonical transformations) The Poisson bracket is in-variant under a canonical transformations, i.e. we have
[U, V]Q,P = [u, v]q,p.
15.3 Generation of constants of the motion
Using Corollary 2 on the Poisson bracket of canonical variables we see Hamiltons
equations of motion are equivalent to the system of equations
dq
dt = [q, H] and
dp
dt = [p, H].
Here, by [q, H] we mean the vector with components [qi, H], and observe we have
simply substituted the components ofq and p for f in Corollary 2. Ifu = u(q,p, t) is
a constant of the motion thendu
dt = 0,
and by Corollary 2 we see thatu
t = [H, u].
If u = u(q,p) only and does not depend explicitly on t then it must must Poisson
commute with the HamiltonianH, i.e. it must satisfy
[H, u] = 0.
Now suppose we have two constants of the motionu = u(q,p) andv = v(q,p), so that[H, u] = 0 and [H, v] = 0. Then by Jacobis identity we find
[H, [u, v]] = [u, [v, H]] [v, [H, u]] = 0.In other words it appears as though [u, v] is another constant of the motion. (This result
also extends to the case when u and v explicitly depend on t.) Indeed, in principle, we
can generate a sequence of constants of the motionu,v, [u, v], [u, [u, v]], .. . . Sometimes
we generate new constants of the motion by this procedure, i.e. we get new information,
but often we generate a constant of the motion we already know about.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
40/56
40 Simon J.A. Malham
15.4 Example (Kepler problem)
Consider the Kepler problem of a particle of mass m in a three-dimensional central
force field with Hamiltonian
H= 1
2m(p21+p
22+p
23) + V(r)
where V = V(r) is the potential with r =
q21+ q22+ q
23. Known constants of the
motion are u = q2p3 q3p2 and v = q3p1 q1p3 and their Poisson bracket [u, v] =q1p2 q2p1 is another constant of the motion (not too surprisingly in this case).
16 Geodesic flow
We have already seen the simple example of the curve that minimizes the distance
between two points in Euclidean spaceunsurprisingly a straight line. What if the
two points lie on a surface or manifold? Here we wish to determine the characterizing
properties of curves that minimize the distance between two such points. To begin to
answer this question, we need to set out the essential components we require. Indeed,
the concepts we have discussed so far and our examples above have hinted at the need
to consider the notion of a manifold and its concomitant components. We assume in
this section the reader is familiar with the notions of Hausdorff topology, manifolds,
tangent spaces and tangent bundles. A comprehensive introduction to these can be
found in Abraham and Marsden [1, Chapter 1] or Marsden and Ratiu [15]. Thorough
treatments of the overall material in this section can, for example, be found in Abraham
and Marsden [1], Marsden and Ratiu [15] and Jost [8].
16.1 Riemannian metric and manifold
We now introduce a mechanism to measure the distance between two points on a
manifold. To achieve this we need to be able to measure the length of, and angles
between, tangent vectorssee for example Jost [8, Section 1.4] or Tao [17]. The mech-
anism for this in Euclidean space is the scalar product between vectors.
Definition 7 (Riemannian metric and manifold) On a smooth manifoldMwe assign,to any point q M and pair of vectors u,v TqM, an inner product
u,vg(q).
This assignment is assumed to be smooth with respect to the base pointq M, andg= g(q) is known as theRiemannian metric. The length of a tangent vectoru TqMis then
ug(q):= u,u1/2g(q).
ARiemannian manifoldis a smooth manifold M equipped with a Riemannian metric.
Remark 6 Every smooth manifold can be equipped with a Riemannian metric; see
Jost [8, Theorem 1.4.1].
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
41/56
Lagrangian and Hamiltonian mechanics 41
In local coordinates with q = (q1, . . . , q n)T M, u = (u1, . . . , un)T TqM and
v = (v1, . . . , vn)T
TqM, the Riemannian metric g = g(q) is a real symmetricinvertible positive definite matrix and the inner product above is given by
u,vg(q):=n
i,j=1
gij(q)uivj .
16.2 Length and energy
We are interested in minimizing the length of a smooth curve onM, however as wewill see, computations based on the energy of a smooth curve are simpler.
Definition 8 (Length and energy of a curve) Let q : [a, b] M be a smooth curve.Then we define the lengthof this curve by
(q):=
ba
q(t)g(q(t))
dt.
We define the energyof the curve to be
E(q):= 12
ba
q(t)2g(q(t))
dt.
Remark 7 The energy E(q) is the actionassociated with the curve q = q(t) on [a, b].
Remark 8 (Distance function)Thedistancebetween two points qa, qb M, assumingq(a) = qa and q(b) = qb, can be defined as d(a, b) := infq
(q)
. This distance
function satisfies the usual axioms (positive definiteness, symmetry in its arguments
and the triangle inequality); see Jost [8, pp. 1516].
By the CauchySchwarz inequality we know that
ba
q(t)g(q(t))
dt |b a| 12 b
a
q(t)2g(q(t))
dt
12
.
Using the definitions of(q) and E(q) this implies
(q)2
2 |b a| E(q).
We have equality in this last statement, i.e. we have
(q)2
= 2 |b a| E(q), if andonly ifq is constant. The length (q) of a smooth curve q = q(t) is invariant to
reparameterization; see for example Jost [8, p. 17]. Hence we can always parameterize
a curve so as to arrange for q to be constant (this is also known as parameterization
proportional to arclength). After such a parameterization, minimizing the energy is
equivalent to minimizing the length, and this is how we proceed henceforth.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
42/56
42 Simon J.A. Malham
16.3 Geodesic equations
We are now in a position to derive the equations that characterize the curves that
minimize the length/energy between two arbitrary points on a smooth manifoldM.For more background and further reading see Abraham and Marsden [1, p, 2245],
Marsden and Ratiu [15, Section 7.5], Montgomery [16, pp. 610] and Tao [17]. Our
goal is thus to minimize the total energy of a path q = q(t) where q = (q1, . . . , q n)T.
The total energy of a path q= q(t), betweent = a and t = b, as we have seen, can be
defined in terms of a Lagrangian function L = L(q, q) as follows
E(q):=
t2t1
Lq(t),q(t)
dt.
The path q = q(t) that minimizes the total energy E= E(q) necessarily satisfies the
EulerLagrange equations. Here these take the form of Lagranges equations of motion
d
dt
L
qj
L
qj= 0,
for each j = 1, . . . , n. In the following, we use gij to denote the inverse matrix ofgij(where i, j = 1, . . . , n) so that
nk=1
gikgkj =ij ,
where ij is the Kronecker delta function that is 1 wheni = j and zero otherwise.
Lemma 4 (Geodesic equations) Lagranges equations of motion for the Lagrangian
L(q,q):= 12
ni,k=1
gik(q) qiqk,
are given in local coordinates by the system of ordinary differential equations
qi+
nj,k=1
ijk qjqk = 0,
where the quantitiesijk are known as the Christoffel symbolsand fori, j, k= 1, . . . , nare given by
ijk :=
n
=1
12g
igjqk
+ gk
qj gjk
q.
ProofWe complete the proof in three steps as follows.
Step 1. For the Lagrangian L = L(q, q) defined in the statement of the lemma, using
the product and chain rules, we find that
L
qj= 12
ni,k=1
gikqj
qiqk
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
43/56
Lagrangian and Hamiltonian mechanics 43
and
L
qj= 12
nk=1
gjk qk+ 12
ni=1
gij qi
= 12
nk=1
gjk qk+ 12
nk=1
gkj qk
=
nk=1
gjk qk,
where in the last step we utilized the symmetry ofg . Using this last expression we see
d
dt
L
qj
=
nk=1
d
dt
gjk qk
=
nk=1
dgjkdt
qk+
nk=1
gjk qk
=
nk,=1
gjkdq
qqk+
nk=1
gjk qk
.
Substituting the expressions above into Lagranges equations of motion, using that the
summation indicesi and k can be relabelled and that g is symmetric, we find
nk=1
gjk qk+
nk,=1
gjkq
12gkqj
qkq = 0,
for each j = 1, . . . , n.
Step 2. We multiply the last equations from Step 1 by g ij and sum over the index j .
This gives
n
j,k=1 gijgjk qk+
n
j,k,=1 gij
gjkq
12gkqj qkq = 0
qi+n
j,k,=1
gijgjk
q 12
gkqj
qkq = 0
qi+n
j,k,=1
gigj
qk 12
gjkq
qjqk = 0,
where in the last step we relabelled summation indices.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
44/56
44 Simon J.A. Malham
Step 3. By using the symmetry ofg and performing some further relabelling of sum-
mation indices, we seen
j,k,=1
gi
gjqk
12gjkq
qjqk =
nj,k,=1
12g
i
gjqk
+ gj
qk gjk
q
qjqk
=
nj,k,=1
12g
i
gjqk
+ gk
qj gjk
q
qjqk
=
nj,k=1
iij qjqk,
where the iij are identified with their definition in the statement of the lemma.
Remark 9 On a smooth Riemannian manifold there is unique torsion free connection or covariant derivative defined on the tangent bundle known as the LeviCivitaconnection. A curve q : [a, b] M is autoparallel or geodesicwith respect to if
q q 0,
i.e. the tangent field ofq = q(t) is parallel alongq = q(t); see Jost [8, Definition 3.1.6].
In local coordinates, the coordinates ofq q are given by
(q q)i = qi+n
j,k=1
ijkqjqk,
for i = 1, . . . , n, where the quantities ijk are equivalent to the Christoffel symbols
above.
Now suppose thatM is an embedded submanifold in a higher dimensional Euclideanspace RN. Then M is a Riemannian manifold. (Nashs Theorem says that any Rieman-nian manifold can be embedded in a higher dimensional Euclidean space RN.) For an
embedded submanifoldM, its Riemannian metric is that naturally induced from theembedding Euclidean space using the map from intrinsic to extrinsic coordinates. For
example, in our discussion of DAlemberts principle, we saw that the dynamics of the
system evolved in the Euclidean space R3N, parameterized by the extrinsic coordinates
r1, . . . , rN, though subject to m
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
45/56
Lagrangian and Hamiltonian mechanics 45
equations of motion for this Lagrangian, we can trace the direction of implication back
through to DAlemberts principle and thus deduce that
q TqM.
As we argued in the lead up to DAlemberts principle, this makes sense on a intuitive
and physical level. The system is free to evolve on the manifold of constraint and,
assuming no external forces, the forces of constraint must act normally to the manifold
(normal to the tangent plane at every pointq M) in order to maintain the evolutionon the manifold. In other words, as Tao [17] eminently puts it,
... the normal quantity q then corresponds to the centripetal force necessaryto keep the particle lying inM (otherwise it would fly off along a tangent linetoM, as per Newtons first law).
16.4 Example: Euler equations for incompressible flow
We formally show here how the Euler equations for incompressible flow represent a
geodesic submanifold flow. The argument is an extension of that above for finite di-
mensional geodesic submanifold flow to an infinite dimensional context and is originally
due to Arnold. Our arguments essentially combine those in Daneri and Figalli [6] and
Tao [17] and we cite Chorin and Marsden [5] and Malham [13] for further background
material on incompressible fluid mechanics.
The Euler equations for incompressible flow in a bounded Lipschitz domain D Rdand on an arbitrary finite time interval [0, T] are prescribed as follows. The velocity
field u = u(x, t) for (x, t) D [0, T] evolves according to the system of nonlinearpartial differential equations
u
t + (u )u= p,
u= 0.
Here the field p = p(x, t) is the pressure field. We need to augment this system of
equations with the zero flux boundary condition for all (x, t) D [0, T], i.e. we haveu n = 0, where n is the unit outer normal to D. We also need to prescribe initialdata u(x, 0) =u0(x) for some given function u0 = u0(x) onD.
From the Lagrangian viewpoint we can describe the flow through the swarm of
particle trajectories prescribed by integral curves of the local velocity fieldu = u(x, t).
Let q = q(x, t) Rd represent the position at time t 0 of the particle that startedat position x D at time t = 0. The flow q= q(x, t) satisfies the system of equations
q(x, t) =uq(x, t), t
,
q(x, 0) =x,
for all (x, t) D [0, T]. A straightforward though lengthy calculation (see Chorinand Marsden [5] or Malham [13]) reveals that the Jacobian of the flow det q =det x q(x, t) satisfies the system of equations
d
dt
det q = u(q, t) det q.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
46/56
46 Simon J.A. Malham
Since we assume the flow is incompressible so u= 0, andq(x, 0) = x, we deduce forallt [0, T] we have det q 1. This means that for allt [0, T], q( , , t) SDiff(D),the group of measure preserving diffeomorphisms onD . If we now differentiate theevolution equations for q = q(x, t) above with respect to time t and compare this
to the Euler equations, we obtain the following equivalent formulation for the Euler
equations of incompressible flow in terms ofqt(x):= q(x, t) for all t [0, T]:qt = p(qt, t),
with the constraint
qt SDiff(D).We now demonstrate formally how these equations can be realized as a geodesic
flow on SDiff(D) equipped with the L2(D;Rd) metric induced from the inner product
v,L2
(D;Rd
)
:= D
v
dx.
We consider SDiff(D) as an infinite dimensional submanifold of Diff(D), the group ofdiffeomorphisms on D. By analogy with the finite dimensional case, the curve q : t qton SDiff(D) is geodesic if
q TqSDiff(D).Remark 10 Suppose the flow is not constrained and evolves on Diff(D). Then a geodesicwould be characterized by q = 0, i.e. Newtons Second Law, as the metric is flat. For
the intrinsic velocity field (see below) we would have u/t +(u )u= 0; see Tao [17].Our goal now is to characterize the vector space orthogonal to TqSDiff(D). Using thedefinition of the flow above, we define the intrinsic velocity field ut(x) = u(x, t) by
ut := qt q1t . Using the result for the Jacobian detq above we note that the innerproduct
,
above is invariant to transformations x
qt(x)
SDiff(
D). Further,
we can thus characterize
TqSDiff(D) =u(qt, t) : u(qt, t) = 0 and u n(qt, t) = 0 on D
.
Equivalently, since it is the Lie algebra of SDiff(D) which is the vector space of allsmooth divergence-free vector fields parallel to the b oundary, we have
TqSDiff(D) =ut : ut = 0 and ut n= 0 on D
.
The HelmholtzHodge Decomposition Theorem tells us that any vector L2(D,Rd)can be orthogonally decomposed into a divergence-free vector u parallel to D and asolenoidal vector p. In other words we can write
= u + p,
where u = 0 and p =p for some scalar field p. See for example Chorin andMarsden [5, p. 37]. In terms of tangent spaces, we have
TqDiff(D) = TqSDiff(D)
TqSDiff(D)
where TqSDiff(D)
=p L2(D,Rd) : p = p.
Hence we deduce thatqt = p(qt, t) for qt SDiff(D), which are the Euler equationsfor incompressible flow.
-
8/9/2019 Mechanics Lagrangian and Hamilitonian
47/56
Lagrangian and Hamiltonian mechanics 47
xx
y=y(x)
x0 0x
ds
0
surface of
revolution
Fig. 9 Soap film stretched between two concentric rings. The radius of the surface of revolutionis given by y = y(x) for x [x0, x0].
17 Exercises
Exercise (EulerLagrange alternative form)
Show that the EulerLagrange equation is equivalent to
F
x d
dx
F yx F