Sei sulla pagina 1di 23

CHAPTER 7

The Hamiltonian Approach to the Calculus of


Variations
In this chapter we will study a different approach to the calculus of variations. We will
begin in first two sections by studying the Legendre transformation, which is at the heart
of this approach.

7.1. The Legendre transformation for functions of one variable


Suppose that f 2 C k .I /; where I D .a; b/, and k  2. We will assume that
f .x/ > 0 for all x 2 I . This means that f is a convex function. Set .x/ D f 0 .x/. We
00

will refer to  as the derivative map. Since  0 .x/ D f 00 .x/ > 0 for all x,  is an increasing
function on I , and maps I one-to-one and onto another interval I  D .˛; ˇ/. The function
 has an inverse, which we will denote by . Since f 2 C k .I /,  D f 0 2 C k 1 .I /, and
2 C k 1 .I  /:
Next we introduce a change of variables by setting  D .x/ D f 0 .x/. Define
(7.1.1) f  ./ D x f .x/; where x D ./.

Definition 7.1. The function f ./ defined in (7.1.1) is called the Legendre transform of
f.
If we set x D ./, then
f  ./ D  ./ f . .//:
If we compute the differential of f  in (7.1.1) we get
df  D x d  C  dx f 0 .x/ dx D x d ;
since  D f 0 .x/: Therefore,
df 
(7.1.2) DxD ./:
d
Since f  0 D 2 Ck 1
.I  / we see that f  2 C k .I  /.
Remark 7.2. Computing the Legendre transform of a function f is a two step process.
(1) Solve the equation  D f 0 .x/ for x D ./.
(2) Set f  ./ D x f .x/:
1 ˛
Example 7.3. f .x/ D x on I D .0; 1/, where ˛ > 1.
˛
First solve  D f 0 .x/ D x ˛ 1
to get x D  1=.˛ 1/
: Then
˛  
x 1 ˛=.˛ 1

f ./ D x D    1=.˛ 1/
 1/
D 1  .˛ 1/=˛
:
˛ ˛ ˛
81
82 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

Set ˇ D ˛=.˛ 1/. Then


1 ˇ 1 1
f  ./ D  ; where C D 1:
ˇ ˛ ˇ
Let’s denote .f  / by f  :

Proposition 7.4. f  D f:

Remark 7.5. We sum up this proposition by saying that the Legendre transform is an
involution.

P ROOF. We start by solving the equation y D f  0 ./. By (7.1.2), y D ./; so


 D .y/ D f 0 .y/. Then,

f  .y/ D y f  ./; where  D f 0 .y/.

By (7.1.1), f  .y/ D f .y/:

Another way of characterizing f  ./ is as the solution to a maximum problem.

Proposition 7.6. f  ./ D maxy2I .y f .y// :

P ROOF. With  fixed, set g.y/ D y f .y/. Then g 2 C k .I / satisfies g 00 .y/ D


00
f .y/ < 0 for all y. Thus g is a concave function and therefore can have at most one
maximum x, at which point g 0 .x/ D 0. However, g 0 .x/ D  f 0 .x/ D 0 if and only if
f 0 .x/ D , or x D ./. Therefore g.x/ D x f .x/ D f  ./.

If we apply this proposition to f  , and use Proposition 7.4, we get

Corollary 7.7. minx2I f .x/ D minx2I max2I  .x f  .// :

By Proposition 7.6, f  ./  y f .y/ for all y 2 I . Replacing y by x, we get the


next result.

Corollary 7.8. x  f .x/ C f  ./, for all x 2 I and  2 I  .

Remark 7.9. The inequality in Corollary 7.8 is called Young’s inequality.


1
Example 7.10. Applying Young’s inequality to the case f .x/ D x ˛ in Example 7.3, we
˛
get
1 1
x  x ˛ C  ˇ ; for all x > 0 and  > 0,
˛ ˇ
1 1
provided ˛ > 1 and C D 1:
˛ ˇ
Example 7.11. Suppose that  W Œ0; 1/ ! Œ0; 1/ is a strictly increasing
Rx function satis-
fying .0/ D 0. We will apply Young’s inequality to f .x/ D 0 .y/ dy: A graphical
R
exploration shows us that f  ./ D 0  1 .t/ dt: Young’s inequality becomes
Z x Z 
1
x  .y/ dy C  .t/ dt; for x > 0 and  > 0.
0 0
7.2. THE LEGENDRE TRANSFORMATION FOR FUNCTIONS OF SEVERAL VARIABLES 83

7.2. The Legendre transformation for functions of several variables


The several variable theory of the Legendre transform is different from the one variable
theory only in that it is necessary to be a little more careful about the hypotheses needed.
Let   RN be an open set, and let f 2 C k ./, where k  2. Let .x/ D rf .x/.  is
called the gradient mapping. Clearly,  2 C k 1 .; RN /.

General assumption: The gradient mapping .x/ D rf .x/ maps  one-to-one and onto
an open set  , and  1 2 C k 1 . ; /.

Unlike in one variable, where we needed only to assume that f 00 .x/ > 0, there is not
one condition that is easy to verify, meets all of our needs, and ensures that the general
assumption is satisfied. However, let’s pursue the issue a little. The Hessian of f is the
N  N matrix  2 
@ f
Hf .x/ D .x/ :
@xi @xj
Since  D rf , we see that
 
@j
Hf .x/ D .x/ D D.x/:
@xi
In order that  be one-to-one, it is necessary that D.x/ D Hf .x/ be invertible for all
x. On the other hand, if D.x0 / is invertible, then the inverse function theorem (see
Theorem A.17) tells us that there is an open set U with x0 2 U such that  maps U one-
to-one and onto an open set U  , with  1 2 C k 1 .U  ; U /. Thus the general assumption
is satisfied locally if the Hessian is invertible. In addition there is the following global
result.
Theorem 7.12. Suppose that   RN is a convex open set, and f 2 C k ./, where
k  2. If the Hessian matrix Hf .x/ is positive definite for all x 2 , then the gradient
map .x/ D rf .x/ satisfies the general assumption.
This result prompts a definition.
Definition 7.13. The function f is uniformly convex if , the domain of f , is convex
and the Hessian matrix Hf .x/ is positive definite for all x 2 . (A matrix A is positive
definite if .Ax/  x > 0 for all x ¤ 0:)
Consequently, if f is uniformly convex, the general assumption is satisfied. However,
there are other situations where we want to use the Legendre transformation and f is not
uniformly convex, so we will simply assume that f satisfies the general assumption.
1
Definition 7.14. Suppose that f satisfies the general assumption and set D . The
Legendre transform of f is defined by
(1) solving  D .x/ for x D ./, and then
(2) setting f  ./ D   x f .x/:
We will refer to the coordinates x D .x1 ; x2 ; : : : ; xN / as the original coordinates,
while  D .1 ; 2 ; : : : ; N / will be the conjugate coordinates. The symbol x   is the
standard dot product, so
f  ./ D   x f .x/ D x1 1 C x2 2 C    C xN N f .x/:
Proposition 7.15. f  2 C k . / and rf  ./ D ./ D x:
84 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

P ROOF. The second of these results follows by calculating the differential of f  .


N N N
X X X @f
df  D xj d j C j dxj dxj
@xj
j D1 j D1 j D1
N
X
D xj d j ;
j D1

since rf .x/ D . Hence rf  ./ D x D ./. Since 2 Ck 1


. /, it follows that
f  2 C k . /.
It is natural to ask if f  is uniformly convex when f is. Here is the start of an answer.
Proposition 7.16. If Hf .x/ is positive definite, and  D .x/, the Hf  ./ is positive
definite.
P ROOF. We have x D ..x//, or in terms of the components of x, xi D i ..x//:
If we differentiate this formula and use the Kronecker delta
(
1 if i D j ,
ıij D
0 otherwise,
we get
N
@ X @ i @k
ıij D i ..x// D
@xj @k @xj
kD1
N 2  2
X @ f @ f
D :
@k @i @xj @xk
kD1

This means that if  D .x/ then the matrix product Hf  ./Hf .x/ D I . Thus Hf  ./
is the matrix inverse of Hf .x/. Since Hf .x/ is positive definite, so is its inverse Hf  ./.
To see this, suppose the matrix A is positive definite, and y ¤ 0. Since A is invertible,
there is an x such that Ax D y. Then .A 1 y/  y D .A 1 Ax/  .Ax/ D x  .Ax/ > 0.
Now we have,
Corollary 7.17. If f is uniformly convex, and  D ./ is convex, then f  is uniformly
convex.
Remark 7.18. The additional assumption that  is convex is necessary. We will present
an example in the exercises.
We will leave the next result as an exercise.
Proposition 7.19. The Legendre transform is an involution.
Because of this fact, it is good to write
(7.2.1) f .x/ C f  ./ D   x; where rf .x/ D  and rf  ./ D x,
to illustrate the involutive nature of the Legendre transform.
We will also leave the next two results as exercises.
Proposition 7.20. Suppose that f 2 C 2 ./ satisfies Hf .x/ > 0 for all x 2 , and that
 is convex. Then f  ./ D maxy2 .  y f .y// for all  2  .
7.3. THE HAMILTONIAN APPROACH 85

Corollary 7.21. Suppose that both f and f  are uniformly convex. Then x    f .x/ C
f  ./ for all x 2  and  2  .
Again, this is called Young’s inequality.

7.3. The Hamiltonian approach


Consider the functional
Z b
F .u/ D F .x; u.x/; u0 .x// dx;
a

where I D .a; b/ and the Lagrangian F 2 .I  /, where   RN  RN D R2N . We


will refer to I  RN as the configuration space, while I   is called phase space. Then
a mapping x ! .x; u.x// is a curve in configuration space, while x ! .x; u.x/; u0 .x// is
a curve in phase space.
An extremal for the functional F is a function u 2 C 2 .I ; RN / which satisfies the
Euler-Lagrange equations
d
rp F D ru F:
dx
When written out in terms of the coordinates of u this equation becomes
d
Fp D Fuj for j D 1; 2; : : : ; N .
dx j
Thus the Euler-Lagrange equations are a system of N second order differential equations
for u. This system has the same solutions as the first order system we get by introducing
functions pj .x/ and requiring that
duj
.x/ D pj .x/
(7.3.1) dx for j D 1; 2; : : : ; N .
d
Fpj .x; u.x/; p.x// D Fuj .x; u.x/; p.x//
dx
Written thus, the Euler-Lagrange equations are a first order system of 2N equations for
the 2N unknowns in the vectors u.x/ and p.x/. Of course, we can write this system more
succinctly as the vector equations
du
D p; and
(7.3.2) dx
d
rp F D ru F:
dx
The Hamiltonian approach to the calculus of variations begins by applying the Le-
gendre transform to the Lagrangian F in the p variables only. We have to adapt out general
assumption to this new situation.

General assumption: The mapping


.x; u; p/ ! .x; u; rp F .x; u; p//
is a C 1 one-to-one map of phase space onto an open set in I  RN ; RN and has a C 1
inverse.

Set .x; u; p/ D rp F .x; u; p/ D . The new variables  are the variables conjugate
to p. They are sometimes referred to as the canonical momenta. Then we have the map
.x; u; p/ ! .x; u; .x; u; p// D .x; u; /;
86 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

which is assumed to have an inverse


.x; u; / ! .x; u; p/ D .x; u; .x; u; //:
The maps  and are assumed to be C 1 , and we have
 D .x; u; p/; and p D .x; u; /:
To use the Legendre transform, we introduce the conjugate variable  D .x; u; p/ D
rp F .x; u; p/. Then p D .x; u; /. Next we introduce the partial Legendre transform of
F,
F  .x; u; / D   p F .x; u; p/; where p D .x; u; /:
We will denote the partial Legendre transform by H , and call it the Hamiltonian. Thus,
H.x; u; / D F  .x; u; / D   p F .x; u; p/:
Notice that
N
X N
X N
X N
X
dH D pj d j C j dpj Fx dx Fuj duj Fpj dpj
j D1 j D1 j D1 j D1
N
X N
X
D Fx dx Fuj duj C pj d j ;
j D1 j D1

since by definition j D Fpj : On the other hand,


N
X N
X
dH D Hx dx C Huj duj C Hj d j :
j D1 j D1

We conclude that
(7.3.3) Hx D Fx ; Huj D Fuj ; and Hj D pj :
These equations are referred to as the Hamiltonian partials.
Next we look at the Euler-Lagrange equations in the Hamiltonian formulation. Com-
bining (7.3.1) and (7.3.3) we see that u0j D pj D Hj . Next, since by definition
Fpj D j , j0 D Fuj D Huj : To sum up, the Euler-Lagrange equations become
duj
D Hj
(7.3.4) dx for j D 1; 2; : : : ; N .
d j
D Huj
dx
The equations in (7.3.4) are called the Hamiltonian form of the Euler-Lagrange equations,
or simply the Hamiltonian equations. The Hamiltonian equations in vector form are
du
D r H
(7.3.5) dt
d
D ru H:
dt
Remark 7.22. Suppose we have a broken extremal for the Lagrangian F . In the Hamil-
tonian formulation, we see that the Erdmann corner conditions require that the generalized
momentum  D rp F and the Hamiltonian H D p  rp F F must be continuous.
p
Example 7.23. F .x; u; p/ D !.x; u/ 1 C p 2 , N D 1:
7.3. THE HAMILTONIAN APPROACH 87

The canonical momentum is


!.x; u/p
 D .x; v; p/ D Fp .x; u; p/ D p :
1 C p2
Solving for p we find that
!.x; u/
pD .x; u; / D p :
!.x; u/2  2
The Hamiltonian is
p
H.x; u; / D p F .x; u; p/ D !.x; u/2  2:

The Hamiltonian system is


 !.x; u/!u .x; u/
u0 D H D p and  0 D Hu D p :
!.x; u/2 2 !.x; u/2  2
Example 7.24. Consider the motion of a large number of particles under a conservative
force. We studied this in Section 6.3, and we will use the notation developed there. The
j th particle has mass mj , position xj D .xj ; yj ; zj / and velocity vj D .xP j ; yPj ; zPj /. We
will accumulate the positions into one vector with 3N components,

x D .x1 ; x2 ; : : : ; xN / D .x1 ; y1 ; z1 ; x2 ; y2 ; z2 ; : : : ; xN ; yN ; zN /:

Similarly we will write v D .v1 ; v2 ; : : : ; vN /.

The kinetic energy of the system is


N N
X 1 1X
T D mj jvj j2 D mj .xP j2 C yPj2 C zPj2 :
2 2
j D1 j D1

The particles are under the control of a conservative force, so there is a potential function
U.t; x/, such that the force on the j th particle is
 
@U @U @U
Fj D rxj U D ; ; :
@xj @yj @zj
The Lagrangian for the system is
N
X 1
L.t; x; v/ D T U D mj jvj j2 U.t; x/:
2
j D1

Hamilton’s principle tells us that allowable motions are extrema for the Lagrangian L.
Let’s see what the Hamiltonian formulation looks like. First we introduce the canonical
momenta  D rv F . If we look at the j th particle we get

 j D m j vj :

Thus the canonical momenta are simply the ordinary momenta. Consequently, in physical
situations we will use p to denote the momenta, and we have pj D mj vj : The inverse is
vj D m1j pj :
88 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

The Hamiltonian is
H Dpv L
0 1
N N
@1
X X
D pj  vj mj jvj j2 U.t; x/A
2
j D1 j D1
N N
(7.3.6) X 1 1X jpj j2
D jpj j2 mj 2 C U.t; x/
mj 2 mj
j D1 j D1
N
1 X 1
D jpj j2 C U.t; x/:
2 mj
j D1

In terms of the original variables we see that


N
X 1
H D mj jvj j2 C U.t; x/ D T C U:
2
j D1

Thus the Hamiltonian is the total energy of the system.


From (7.3.6) we get the Hamiltonian equations
1
x0j D rpj H D pj ;
mj
p0j D rxj H D rxj U D Fj :

7.4. Conserved quantities


Suppose that x ! .x; u.x/; u0 .x// is an extremal for the functional
Z b
F .u/ D F .x; u; u0 / dx;
a
and let H.x/ D H.x; u.x/; .x//. Then, using the Hamiltonian equations we see that
N N
d X duj X d j
H D Hx C Huj C Hj
dx dx dx
j D1 j D1
N
X
D Hx C
 
Huj Hj Hj Huj
j D1
D Hx :
If F does not depend explicitly on x, it is easy to see that the same is true for H . Thus we
have one more proof of the following
Proposition 7.25. If the Lagrangian is independent of x, then the Hamiltonian H is a
conserved quantity.
p
Example 7.26. For a Lagrangian of the form F .x; u; p/ D !.x; u/ 1 C p 2 , we saw in
Example 7.23 that the Hamiltonian is H D ! 2  2 . Thus if ! is independent of x,
! 2  2 is a conserved quantity.
Example 7.27. Suppose we have a system of N particles moving under a conservative
force field from Example 7.24. This time we will assume that the potential U depends
7.4. CONSERVED QUANTITIES 89

only on the distances between the particles jxj xk j for 1  j < k  N . A computation
shows that the total force on the system satisfies
N
X N
X
Fj D rxj U D 0:
j D1 j D1
PN
Then if p D j D1 pj is the total momentum, the Hamiltonian equations imply that
N N
dp X d pj X
D D Fj D 0:
dt dt
j D1 j D1

Consequently total momentum is a conserved quantity.


The following example shows use how making a change of variables can lead us to
conserved quantities.
Example 7.28. Consider a single particle moving in the plane under a central force field.
This means that the potential U.x/ depends only on the distance from the origin jxj. Thus
there is a function g such that U.x/ D g.jxj/.
It is natural to analyze this situation using polar coordinates
x D r cos  y D r sin :
Then
xP D rP cos  r P sin  and
yP D rP sin  C r P cos :
Therefore the kinetic energy is
m 2 m 2 
T D xP C yP 2 D rP C r 2 P 2 :

2 2
The Lagrangian in polar coordinates is
m 2 
LDT U D rP C r 2 P 2 g.r/:
2
We will put this into its Hamiltonian formulation. The generalized momenta are
pr D LrP D mrP and p D L P D mr 2 ; P

and the Hamiltonian is
1 2 1
H D pr C p 2 C g.r/:
2m 2mr 2 
Finally, the Hamiltonian equations are
dr pr
D Hpr D
dt m
d p
D Hp D
dt mr 2
dpr
D Hr D g 0 .r/
dt
dp
D H D 0:
dt
The last of these equations shows us that the angular momentum p D mr 2 P is a
conserved quantity. It comes about because the change of variables to polar coordinates led
90 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

to a Hamiltonian which does not depend explicitly on . We will look for more conserved
quantities by changing variables.

7.5. Canonical transformations


We will start by looking for transformations of phase space that preserve the Hamil-
tonian structure. What this means is covered by the following definition.
Definition 7.29. A canonical transformation is a mapping
(7.5.1) x  D x; u D u .x; u; /;   D   .x; u; /
of phase space that sends a Hamiltonian system
(7.5.2) u0 D r H; 0 D ru H
into another Hamiltonian system,
0 0
u D r  H  ;  D ru H 
for a suitable Hamiltonian H  .
7.5.1. How do we recognize a canonical transformation?
We should state immediately that not every transformation of phase space is a canon-
ical transformation. Consider the Hamiltonian H.x; p/ D xp C p 2 . The Hamiltonian
system is
(7.5.3) x 0 D Hp D x C 2p; p 0 D Hx D p:
If we make the change of variables
(7.5.4) y D p; q D x;
then the system in (7.5.3) becomes
dy dp
D D p D y; and
dt dt
dq dx
D D x 2p D q 2y:
dt dt
We want to know if this is a Hamiltonian system. This means that we need to know if there
is a function H  .y; q/ such that
Hy D 2y q; and Hq D y:
If we integrate the second equation with respect to q we see that H  .y; q/ D yq CC.y/:
Differentiating with respect to y, we get Hy D q C C 0 .y/, and this is equal to 2y q
only if C 0 .y/ D 2y. Hence, H  .y; q/ D y 2 yq is a Hamiltonian for the transformed
system and the transform (7.5.4) is a canonical transformation.
Now let’s make the change of variables
(7.5.5) y D ex ; q D p:
The system in (7.5.3) becomes
dy dx
D ex D e x .x C 2p/ D y.log y C 2q/; and
dt dt
dq dp
D D p D q:
dt dt
If H  .y; q/ is a Hamiltonian for this system, then
Hq D y.log y C 2q/; and Hy D q:
7.5. CANONICAL TRANSFORMATIONS 91

However, mixed second partials must be equal, so


@  @2 H  @2 H  @ 
Hq D D D H :
@y @q@y @y@q @q y
In this case that is not true. Hence there is no Hamiltonian, and therefore the transformation
in (7.5.5) is not a canonical transformation.
When we have two variables, the mixed second derivative test provides a method
to decide if a system is Hamiltonian. If a system passes this test, then the integration
procedure outlined above provides a way to find a Hamiltonian. In more variables things
are more complicated and we will not pursue the point.
7.5.2. Finding canonical transformations
Finding canonical transformations is easier than it looks at first glance, but we have to
do some work.
Suppose we have a Hamiltonian system like that in (7.5.2). Consider the Lagrangian
(7.5.6) FH .x; u; ; p; q/ D   p H.x; u; /:
The associated variational integral
Z b
FH .u; / D FH .x; u; ; p; q/ dx
a
(7.5.7) Z b
D .x/  u0 .x/ H.x; u.x/; .x// dx
 
a
is called the Poincaré-Cartan integral associated with H . Notice that FH is linear in p
and does not depend on q at all. It is highly degenerate. Nevertheless it is important, as the
following proposition makes clear.
Proposition 7.30. The Euler-Lagrange equations for the Poincaré-Cartan integral are the
Hamiltonian equations for H in (7.5.2).
P ROOF. We compute that ru F D ru H , and rp F D ; so the Euler-Lagrange
d
equation ru F dx rp F D 0 becomes  0 D ru H: Next we notice that r F D p r H
d
and rq F D 0, so r F dx rq F becomes p D r H . However, along an extremal of F
0 0
we have p D u . Hence, u D r H .
Remark 7.31. The import of Proposition 7.30 is that a function x ! .u.x/; .x// satisfies
the Hamiltonian system (7.5.2) if and only if it is an extremal for the Poincaré-Cartan
integral FH .
The functional FH has 2N variables .u; /. Therefore we would expect the Euler-
Lagrange equations to be 2N second order ordinary differential equations. Instead they
turn out to be 2N first order ordinary differential equations. This is a reflection of the
degeneracy of FH .
Rb
It is sometimes convenient to consider a functional F .u/ D a F .x; u; u0 / dx as a
line integral of the differential 1-form F .x; u; p/ dx over the curve in phase space with
the parametrization
x ! .x; u.x/; u0 .x//
determined by the function u.x/. Thus
Z Z b
F .u/ D F .x; u; p/ dx D F .x; u.x/; u0 .x// dx:
a
92 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

If F .x; u; p/ is any Lagrangian, consider the form


F dx C dG
where G is any function of .x; u/. If the curve is parametrized by x ! .x; u.x/; u0 .x//,
then along
N
X
dG D Gx dx C Guj duj
j D1
2 3
N
X
D 4Gx C Guj u0j 5 dx
j D1

D Gx C ru G  u0 dx:
 

Therefore if we define a new functional


F  .x; u; p/ D F .x; u; p/ C Gx C ru G  p;
then along we have
F  dx D F dx C dG:
Proposition 7.32. The functionals F and F  have the same extremals.
P ROOF. The curve has end points Pa D .a; u.a/; u0 .a// and Pb D .b; u.b/; u0 .b//,
so
2 3
Z Z N
X
dG D 4Gx dx C Guj duj 5
j D1
2 3
b N
duj 5
Z X
D 4Gx C Guj dx
a dx
j D1
b
d
Z
D G.x; u.x// dx
a dx
D G.Pb / G.Pa /:
R
Thus dG depends only on the end points of . We have
Z
F .u/ D F  dx


Z
D F dx C dG

D F .u/ C G.Pb / G.Pa /:
If we vary u by u C t, where  2 C01 .I; RN / then all of the corresponding curves have
the same endpoints. Hence
F  .u C t/ D F .u C t/ C G.Pb / G.Pa /
for small t. Since the term involving the end points does not depend on t, when we differ-
entiate with respect to t, we see that ıF  .u; / D ıF .u; / for all  2 C01 .I; RN /. It
follows that F and F  have the same extremals.
Remark 7.33. The import of Proposition 7.32 is that a function u.x/ satisfies the Euler-
Lagrange equations for F  if and only if it satisfies those for F .
7.5. CANONICAL TRANSFORMATIONS 93

We now apply Propositions 7.5.6 and 7.32 to the Poincaré-Cartan integral defined
in (7.5.7). Notice that
N
X
(7.5.8) FH .x; u; ; p; q/ dx D j duj H dx:
j D1

We will call this the Poincaré-Cartan form associated with H . We will be interested in
functions G such that the modified form
N
X
(7.5.9) FH dx D FH dx C dG D j duj H dx C dG
j D1

is the Poincaré-Cartan form for another Hamiltonian H  in a new coordinate system re-
lated to the old by a transformation of phase space of the form in (7.5.1) . If this is so,
then by Proposition 7.32 FH and FH  have the same extremals. Since the Euler-Lagrange
equations for FH and FH  are the Hamiltonian systems associated with H and H  re-
spectively, we see that these two systems have the same solutions.
In symbols we want
N
X
(7.5.10) FH D FH  D j duj H  dx:
j D1

Comparing (7.5.9) and (7.5.10) we need


N
X N
X
(7.5.11) dG D .H H  / dx C j duj j duj :
j D1 j D1

We can use (7.5.11) to generate the transformation in (7.5.1), and therefore we will
call G a generating function for the canonical transformation. G is only required to be a
function in C 3 .I  RN  RN /. It can be expressed in terms of any combination of the
original variables and the starred variables. We will examine each case individually.

Case 1. G D G.x; u; u /
We have
N
X N
X
(7.5.12) dG D Gx dx C Guj duj C Guj duj :
j D1 j D1

Comparing (7.5.12) and (7.5.11), we see that we must have


(7.5.13) Gx .x; u; u / D H H ;
(7.5.14) ru G.x; u; u / D ; and
 
(7.5.15) ru G.x; u; u / D  :
To find the canonical transformation generated by G we follow these steps:
(1) Solve (7.5.14) for u D u .x; u; /:
(2) Substitute this into (7.5.15) to get   D ru G.x; u; u .x; u; //.
(3) Substitute these results into (7.5.13) to get H  H dG:
Let’s look at an example.
94 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

Example 7.34. Our first example shows that the separation between the variables u and 
is largely superficial when we allow transformations of phase space. let
N
X
G.u; u / D uj  uj :
j D1

In the first step we solve  D ru G D u for u D 


. In the second step we set
  D ru G D u: Thus the canonical transformation is
u D ;   D u:
Finally, the new Hamiltonian is H  .x; u ;   / D H.x;   ; u /:
Case 2. G D G.u;   /
This time
N
X N
X
dG D Gx dx C Guj duj C Gj d j ;
j D1 j D1

which does match up with (7.5.11). To get around this we introduce GQ D G


PN
j D1 j uj :
Then assuming that G satisfies (7.5.11), we have
N N
d GQ D dG
X X
j duj uj d j
j D1 j D1
(7.5.16)
N
X N
X
D .H H  / dx j duj C uj d j :
j D1 j D1

On the other hand, assuming that GQ D G.x;


Q u;   /,
N N
d GQ D GQ x dx C GQ uj duj C GQ j d j ;
X X
(7.5.17)
j D1 j D1

Comparing (7.5.16) and (7.5.17) we see that we need


(7.5.18) GQ x .x; u;   / D H H ;
(7.5.19) Q
ru G.x; u;   / D ; and
(7.5.20) Q
r  G.x; u;   / D u :
To find the canonical transformation, we again have three steps. First solve (7.5.19) for
Q
  D   .x; u; /: Next use (7.5.20) to set u D r  G.x; u;   .x; u; //: Finally set

H D H d G;. Q

Example 7.35. Let GQ D


PN 
j D1 uj j : The canonical transformation is the identity map
on phase space.
Example 7.36. Suppose that .x; u/ ! .x; u / is a transformation of of configuration
Q
space. Can we extend it to a canonical map of phase space? To do so, set G.x; u;   / D
PN  
kD1 k uk .x; u/: Then we must solve the equations
N
@uk 
j D GQ uj D
X
 ; 1j N
@uj k
kD1
7.6. HAMILTON-JACOBI THEORY 95

for j , 1  j  N . In vector form this equation reads  D Du u .x; u/    . This linear


system has the solution   D .Du u .x; u// 1  : Therefore the canonical transformation
is
u D u .x; u/   D .Du u .x; u// 1  :
Thus any mapping of configuration space can be extended to a canonical transformation of
phase space.
7.5.3. Case 3. G D G.x; ; u /
In this situation we use GQ D G C N
P
j D1 j uj . Proceeding as before we see that we
must have
(7.5.21) GQ x .x; ; u / D H H ;
(7.5.22) Q
r G.x; ; u / D u; and
(7.5.23) Q
ru G.x; 
; u / D  : 

To find the canonical transformation we first solve (7.5.22) for u , and then substitute
into (7.5.23) to find   .
Example 7.37. This is similar to Example 7.36, but now we assume that we have a trans-
formation of configuration space in the form
xDx u D u.x; u /:
We set GQ D
PN  
kD1 uk .x; u /k : To find the canonical transformation, we solve u D
r GQ D u.x; u / for u , and then set   D ru GQ D Du :
 

Example 7.38. Let’s illustrate the preceding example by looking at polar coordinates.
In that case the independent variable is t, u D .x; y/, and u D .r; /. Then u D
.r cos ; r sin /. We will denote the momenta by corresponding subscripts. Then accord-
ing to the preceding example, we have
Q ; x ; y / D r cos   x C r sin   y :
G.r;
From (7.5.22) we get x D r cos  and y D r sin , as expected. we solve for
p y
r D x 2 C y 2 and tan  D :
x
Then from (7.5.23) we get
xx C yy
r D and  D xy yx :
r
7.5.4. Case 4. G D G.x; ;   /
We use GQ D G C N
P PN  
j D1 j uj j D1 j uj , and proceed as before.

7.6. Hamilton-Jacobi theory


We want to use canonical transformations to transform a given Hamiltonian system
duj
D Hj for 1  j  N
(7.6.1) dx
d j
D Huj for 1  j  N
dx
96 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

into one with Hamiltonian H  which is easier to solve. The simplest such system is when
the Hamiltonian is H  .x; u ;   / D 0; in which case the system is
duj
D0 for 1  j  N
dx
d j
D 0 for 1  j  N
dx
for which the solutions are uj D ˛j and j D ˇj for 1  j  N , where ˛j and ˇj are
constants. This is a complete solution to the problem.

7.6.1. The Hamilton-Jacobi equation


While this may seem overly ambitious, let’s try to do it. According to the previous
section we need a generating function S.x; u; u / which leads to the Hamiltonian H  D 0.
By (7.5.11), S must satisfy
N
X
j duj H dx
j D1
N
X
(7.6.2) D j duj H  dx C dS
j D1
N
X N
X N
X
D j duj C .Sx 
H / dx C Suj duj C SuJ duj :
j D1 j D1 j D1

For this to be true we need


(7.6.3) Sx D H  H
(7.6.4) Suj D j for 1  j  N
(7.6.5) Suj D j for 1  j  N
If H  D 0, then by (7.6.3) we must have Sx C H.x; u; / D 0. From (7.6.4),  D ru S ,
so we need
(7.6.6) Sx C H.x; u; ru S / D 0:
Equation (7.6.6) is called the Hamilton-Jacobi equation for S . It is a first order partial
differential equation for the unknown function S . S D S.x; u/ is a function of the N C 1
variables x 2 R and u 2 RN . The theory of partial differential equations informs us that
there is a complete solution to the equation which depends on N C 1 parameters, each of
which is a real number. However, if S is a solution to (7.6.6), then so is S C C where
C is a constant. Thus one of those parameters amounts to adding a constant, which does
not give us any information, so we will be looking for solutions to the Hamilton-Jacobi
equation which depends on N non-additive parameters.
Suppose that we do have a solution S.x; u; ˛/ to the Hamilton-Jacobi equation which
depends on the N parameters ˛ D .˛1 ; ˛2 ; : : : ; ˛N /: Then, replacing ˛ with u , we can
use S.x; u; u / to generate a canonical transformation. By (7.6.3), we have H  D Sx C
H.x; u; /. By (7.6.4),  D ru S , so H  D Sx C H.x; u; ru S /, and since S satisfies the
Hamilton-Jacobi equation, we conclude that H  D 0. Thus our new Hamiltonian system
simplifies and in the new variables the solutions are constants. In particular, from (7.6.5),
the quantity Suj D j is a constant. If we now replace u with ˛ and   with ˇ,
7.6. HAMILTON-JACOBI THEORY 97

(7.6.5) becomes r˛ S.x; u; ˛/ D ˇ. Assuming that we can solve this system, we have the
solutions to our original Hamiltonian system.
Let’s restate this as a theorem.
Theorem 7.39. Suppose that S.x; u; ˛/ is a complete solution to the Hamilton-Jacobi
equation
(7.6.7) Sx C H.x; u; ru S / D 0:
Then if ˇ D .ˇ1 ; ˇ2 ; : : : ; ˇN / is a constant vector, we get a complete solution to the
Hamiltonian system in (7.6.1) by solving the equation
(7.6.8) r˛ S.x; u; ˛/ D ˇ
for u D u.x; ˛; ˇ/; and then setting
(7.6.9)  D .x; ˛; ˇ/ D ru S.x; u.x; ˛; ˇ/; ˛/:
Thus we have a three step method of solving a Hamiltonian system. First we solve
the Hamilton-Jacobi equation (7.6.7). Next we solve equation (7.6.8), and finally we use
equation (7.6.9).
It may seem that we have substituting the hard problem of solving a Hamiltonian sys-
tem with the problem of solving the Hamilton-Jacobi equation, which is at least equally
hard. However, there are many cases where solving the Hamilton-Jacobi equation is possi-
ble and leads to solutions of Hamiltonian systems where other methods are not applicable.
We will illustrate this in the following sections.

7.6.2. The harmonic oscillator in R


Assume we have a mass which is attracted to the origin x D 0 subject to the restoring
force F D kx. This force has potential U D 12 kx 2 . The kinetic energy is T D 12 mv 2 .
We will use p instead of  for the momenta. Then in the conjugate coordinates .x; p/ we
1
have T D 2m p 2 . Therefore the total energy is
1 2 1 2
E DT CU D p C kx :
2m 2
Since the problem is invariant under translation in the independent variable t, we know
that energy is conserved and that the Hamiltonian is
1 2 1 2
H.u; p/ D E D
p C kx :
2m 2
The Hamilton-Jacobi equation is St C H.x; Sx / D 0 or
1 2 1 2
St C
S C kx D 0:
2m x 2
Our method of solution is called separation of variables, which suggests that we look for
solutions of the form1
S.t; x/ D f .t/ C g.x/:
The Hamilton-Jacobi equation becomes
1 0 2 1 2
f 0 .t/ C g .x/ C kx D 0:
2m 2
1We are assuming additive separation of variables. Separation of variables is also used in solving ODEs and
PDEs by assuming a solution which is the product of functions of one variable. That is multiplicative separation¿
98 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

This sum of a function of t and a function of x can be constant only if each is constant.
Hence there is a constant ˛ such that
1 0 2 1 2
(7.6.10) f 0 .t/ D ˛ and g .x/ C kx D ˛:
2m 2
The first of these equations has solution f .t/ D ˛t. On the other hand, b the
Hamilton-Jacobi equation
˛ D f 0 .t/ D St D H.t; x.t// D E:
Thus ˛ D E, which we know is constant by conservation of energy, and f .t/ D Et.
If we solve the second equation in (7.6.10) for g 0 we get
p
r
0 2˛
g .x/ D mk x2 :
k
Hence Z p r

g.x/ D mk x 2 dx;
k
and our solution to the Hamilton-Jacobi equation is2
p Z r

S.t; x/ D mk x 2 dx ˛t:
k
The next step in the process is to introduce a new parameter ˇ and to solve the equation
ˇ D S˛ .t; x/ for u. We get
ˇ
r
m dx
ˇ D S˛ D q t
k 2˛
x 2
k

Now we integrate and get


r r !
m 1 k
t Cˇ D cos x
k 2˛
p p
If we solve this equation for x and make the replacements A D 2˛=k and ! D k=m,
our solution is
x.t/ D A cos.!.t C ˇ//;
where A and ˇ are arbitrary constants. We could easily solve for the momentum in a
variety of ways, but it is not necessary in this case.
7.6.3. The reduced Hamilton-Jacobi equation
Suppose we have a Hamiltonian of the form H D H.x; /, where we are assum-
ing that the independent variable is t. In situations like this, where the Hamiltonian is
independent of t, the Hamiltonian itself is a conserved quantity, and it can be one of the
parameters. We separate the variable t and x by assuming that S.t; x/ D f .t/ C W .x/.
Then St D f 0 .t/ and rx S D rW so the Hamilton-Jacobi equation becomes
f 0 .t/ C H.x; rW .x// D 0:
This means that f 0 .t/ is a constant. Since f 0 .t/ D St D H , f .t/ D H t, and we have
S.t; x/ D W .x/ H t:
2We could evaluate the integral here, but it is usually not a good idea to do so at this point in the Hamilton-
Jacobi process.
7.6. HAMILTON-JACOBI THEORY 99

This means that the Hamilton-Jacobi equation becomes


(7.6.11) rW .x/ D H;
where H is a constant. This equation is called the reduced Hamilton-Jacobi equation.
7.6.4. The harmonic oscillator in R2
Suppose we have a mass m in R2 which is attracted to the origin by the force F.x; y/ D
k.x; y/. This means that the force has potential U.x; y/ D k2 .x 2 C y 2 /. The kinetic en-
ergy is T D 12 mjvj2 . In the conjugate variables T D 2m 1
jpj2 where p D .px ; py /: The
total energy is
1  k 2
ED px2 C py2 C x C y2 :

2m 2
Since the problem is invariant with respect to translations in t, E is constant and E D H ,
the Hamiltonian. Therefore we have S.t; x/ D W .x/ Et, and the reduced Hamilton-
Jacobi equation is
1  k 2
Wx2 C Wy2 C x C y 2 D E:

2m 2
We will attempt to solve this equation by separation of variables by assuming that
W .x; y/ D g.x/ C h.y/:
After multiplying by 2m the reduced Hamilton-Jacobi equation becomes
g 0 .x/2 C mkx 2 C h0 .y/2 C mky 2 D 2mE:
The functions of x and y must both be constant, so there exist constants ˛1 and ˛2 such
that
g 0 .x/2 C mkx 2 D ˛12 m and h0 .y/2 C mky 2 D ˛22 m where ˛12 C ˛22 D 2E:
p q
For the first equation we have g 0 .x/ D m ˛12 kx 2 , so
p
Z q
g.x/ D m ˛12 kx 2 dx:

Similarly we can solve for h.y/ and finally


p p 1 2
Z q Z q
S.t; x; y/ D m ˛12 kx 2 dx C m ˛22 ky 2 dy .˛ C ˛22 /t:
2 1
Our next task is to solve the equations ˇj D S˛j for j D 1 and j D 2. For j D 1,
we have
ˇ
p ˛1 dx
ˇ1 D S˛1 D m q ˛1 t
˛12 kx 2
r p !
m 1 k
D ˛1 cos x ˛1 t:
k ˛1
Solving for x, we get
r  !
˛1 k ˇ1
x.t/ D p cos tC :
k m ˛1
p p
If we set ! D k=m, A1 D ˛1 = k, and 1 D ˇ1 =˛1 , this becomes
x.t/ D A1 cos .! .t C 1 // ;
100 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

where A1 and 1 are arbitrary constants.


In exactly the same way, the equation ˇ2 D S˛1 leads to
y.t/ D A2 cos .! .t C 2 // ;
where A2 and 2 are arbitrary constants.

7.6.5. The brachistochrone p


We will examine Lagrangians of the form F .x; u; p/ D !.x; u/ 1 C p 2 : We have
seen several examples of this type. In Example 7.23 we showed that the associated Hamil-
tonian is p
H.x; u; / D ! 2 .x; u/  2 :
The Hamilton-Jacobi equation is
q
Sx D ! 2 .x; u/ Su2 :
Many problems of this type can be solved by separation of variables. In particular, if
! D !.u/, then setting S.x; u/ D f .x/ C g.u/ leads to
p
f 0 .x/ D !.u/2 g 0 .u/2 :
2 2
There is a constant C such that f 0 .x/ D C , and 0
p g .u/ D !.u/ C 2:
For pthe brachistochrone, !.x; u/ D 1= c u. Let’s take the constant in the form
C D 1= 2˛. Then
x
f .x/ D p ;

and r
0 1 1
g .u/ D :
c u 2˛
Hence
Z r
x 1 1
S.x; u/ D p C du:
2˛ c u 2˛
Next we have
" Z   1=2 #
3=2 1 1 1
S˛ D .2˛/ x p du :
2˛ c u 2˛
To find u we must solve the equation S˛ D constant. It is convenient to take the
constant of the form .2˛/ 3=2 ˇ: Then the equation becomes
Z   1=2
1 1 1
(7.6.12) x p du D ˇ:
2˛ c u 2˛
To evaluate the integral we substitute
uDc 2˛ sin2 .=2/ D c ˛.1 cos /:
Then
du D 2˛ sin.=2/ cos.=2/ d;
and
1 1 1 cos2 .=2/
D :
c u 2˛ 2˛ sin2 .=2/
7.6. HAMILTON-JACOBI THEORY 101

Therefore, equation (7.6.12) becomes


Z
ˇ D x C 2˛ sin2 .=2/ d
Z
D x C ˛ .1 cos / d
D x C ˛. sin /:
Thus our solution is given parametrically by
xDˇ ˛. sin /;
uDc ˛.1 cos /:
This agrees with the result we found in Section 2.5.3.
7.6.6. The pendulum
We considered the pendulum in Section 4.4.3. Here we will analyze it again using
Hamilton-Jacobi theory.
Suppose that the pendulum rod is of length R and fixed at the origin. The potential
energy depends only on the height of the bob, so if  is the angle of displacement,
U D mgy D mg cos :
Let s.t/ be the distance of the bob from equilibrium along the arc of the pendulum at time
t. Then s.t/ D R.t/. Since the bob is moving only in the  direction, the speed of the
bob is js 0 .t/j D jR 0 .t/j: Therefore the kinetic energy is
1
T D mR2 P 2 :
2
The Lagrangian is
1
L.; P / D T U D mR2 P 2 C mg cos :
2
We will use the Hamiltonian formulation. The (angular) momentum is
p D L P D mR2 ; P

and the Hamiltonian is
p2
H.; p/ D mgR cos :
2mR2
Since L does not depend on t, H D E, the total energy, and this is a constant.
The Hamilton-Jacobi equation is
S2
St C mgR cos  D 0:
2mR2
We assume that S.t; / D g./ Et, and use the reduced Hamilton-Jacobi equation
from (7.6.11),
S2
mgR cos  D E:
2mR2
0
We solve for S D g ./,
p p
g 0 ./ D 2mR2 E C mgR cos 
Therefore
p Z p
S.t; / D 2mR 2 E C mgR cos  d Et:
102 7. THE HAMILTONIAN APPROACH TO THE CALCULUS OF VARIATIONS

To find .t/, we must solve the equation SE D ˇ, where ˇ is a constant. This is


r l
mR2 d
p t D ˇ:
2 E C mgR cos 
This can be written
r l
2 d
.t C ˇ/ D :
mR2
p
E C mgR cos 
The integral on the right can not be evaluated in terms of elementary functions. However,
it is worthwhile noticing that the Hamilton-Jacobi method leads to an expression for the
solution as an integral. In this case it is an elliptic integral, which is well understood and
easily computed.

7.6.7. A mass in R2 under a central force field


This generalizes the problem we
p examined in Section 7.6.4, Now we assume only that
the potential has the form U D U. x 2 C y 2 /. To take advantage of this we will use polar
coordinates
x D r cos  y D r sin ;
Then
xP D rP cos  r P sin  and yP D rP sin  C r P cos ;
so the velocity satisfies
jvj2 D xP 2 C yP 2 D rP 2 C r 2 P 2 :
Therefore the kinetic energy is
1 m
T D mjvj2 D .Pr 2 C r 2 P 2 /:
2 2
The Lagrangian is
P DT m 2
L.r; ; rP ; / .Pr C r 2 P 2 / U.r/:
U D
2
We will use the Hamiltonian formulation. The momenta are
pr D LrP D mrP and p D L P D mr 2 ; P

and the Hamiltonian is
 
1 1
H.r; ; pr ; p / D pr2 C 2 p2 C U.r/:
2m r
Since H is independent of t, H is conserved and is equal to the total energy E. Hence, we
can set
S.t; r; / D W .r; / Et;
where W satisfies the reduced Hamilton-Jacobi equation
 
1 1
Wr2 C 2 W2 C U.r/ D E:
2m r
To solve this equation we assume that W .r; / D g.r/Ch./: Then the reduced Hamilton-
Jacobi equation becomes
 
1 0 2 1 0 2
g .r/ C 2 h ./ C U.r/ D E:
2m r
7.6. HAMILTON-JACOBI THEORY 103

Therefore there is a constant ˛ such that


h0 ./ D ˛ and
2
1 0 2 ˛
g .r/ C C U.r/ D E:
2m 2mr 2
The first equation is solved by h./ D ˛: We solve the second equation for
r
0 ˛2
g .r/ D 2m .E U.r// :
r2
Therefore Z r
˛2
g.r/ D 2m .E U.r// dr;
r2
and
S.t; r; / D W .r; / Et D g.r/ C h./ Et
Z r
˛2
D 2m .E U.r// dr C ˛ Et
r2
where E and ˛ are constants.
To solve for r.t/ we solve the equation
l
mdr
ˇ1 D SE D q t:
2
2m .E U.r// ˛r 2
Evaluating the integral gives us r.t/:
Finally, we get .t/ by solving the equation
dr
Z
ˇ2 D S˛ D ˛ q :
˛2
r 2 2m .E U.r// r2
Evaluating the integral gives  as a function of r and therefore of t.

Potrebbero piacerti anche