Вы находитесь на странице: 1из 13

Chapter 5

Calculus of Variations
The calculus of variations is a powerful mathematical technique that is fundamental to the toolbox of any physicist and can be used to solve a great
variety of problems not only in classical mechanics but also in optics and
quantum mechanics. The calculus of variations can be seen as an extension to what you already know about the minimization of functions. To
minimize a function one takes derivatives, possibly with multiple variables
and searches for the point where the derivative vanishes. In the calculus
of variations, we formulate physical problems in the minimization of functionals, which are mappings from a set of functions to real numbers. Rather
than varying a set of variables, we vary entire functionals and the stationary
functions provide the minimum value of the functional.

5.1

Path minimization

Let us begin with a simple problem. You are standing at a point (x1 , y1 ) on
the beach and someone in the water, at the point (x2 , y2 ), in the sea needs
your help. For simplicity, lets say that the shore lies at x = 0. What path
between the two points gives the shortest travelling time?
Although you may be a good swimmer, your speed v1 running on the sand
is faster than your speed v2 swimming in the water. For this reason, the
path with shortest travelling time is not a straight line. The shortest path
is rather made up of two line segments. Let the point at which you cross
the shore be (0, y), then the travelling time T is a function of y:

1
1
2
2
T (y) =
x1 + (y y1 ) +
x22 + (y2 y)
(5.1)
v1
v2

39

WATER

SAND
v2

(x2 ,y2)

(o,y)
2

x
v1

y-y1
2
1+

x (y-y1)

(x1 ,y1)

x1

Figure 5.1: Finding the shortest path between a point on the beach and a
point in the water.
The minimum time is given by setting the derivative dT /dy = 0:
0=

dT (y)
y y1
y2 y
= 2
2
2
dy
v1 x1 + (y y1 )
v2 x1 + (y2 y)2
sin 1 sin 2
=

v1
v2

(5.2)
(5.3)

That is, the optimum path satises the Snell-Descartes law familiar from
optics in the form:
n1 sin 1 = n2 sin 2
(5.4)
where the only dierence is that one works with refractive indices, n1,2 =
c/v1,2 instead of velocities (here c is the vacuum speed of light).
Let us now consider what happens if your speed is not uniform on the beach.
There may be strips on the beach of dry sand, wet sand, or strips with stones
or washed up corals. Snells law will continue to hold at any interface:
sin i
sin i+1
=
vi
vi+1

(5.5)

In the limit of an innite number of strips of innitesimal width, we can


describe your speed with a continuous function, v(x), and write:
y (x)
sin (x)

=
=C
v(x)
v(x) 1 + y 2 (x)
40

(5.6)


where C is a constant and we used sin (x) = y (x)/ 1 + y 2 (this can
be seen from drawing an innitesimal triangle with sides dx, dy and dl =

dx2 + dy 2 ).
The travel time is now a functional, depending on the choice of function
y(x) representing your path:
x2

dl
1 + y 2
=
dx
(5.7)
T [y(x)] =
v
v(x)
x1
where we used:

( )2

dy
dl = dx2 + dy 2 = dx 1 +
= dx 1 + y 2
dx

(5.8)

If we knew v(x), then the shortest time is given by minimizing the functional
T [y(x)] over all possible functions y(x).

5.1.1

Functions and Functionals

Functionals of the form of Eq. 5.7 do not only apply to optical systems or
life guards at the beach. Similar functionals can be written for a great range
of physical problems. Through the minimization of functionals we will later
reformulate Newtons classical mechanics enabling a powerful approach to
the set up equations of motion for complicated scenarios.
First let us consider the mathematical foundation that we will use to formulate problems. We will generally formulate problems in the form:
b
(
)
L x, y(x), y (x) dx
(5.9)
S[y(x)] =
a

with the specic L depending on the problem considered. Note that Eq. 5.7
has this form.
In dierential calculus, we learn a great deal about functions by changing
the independent variables in small increments. Let us attempt a similar
thing here for functionals by changing whole paths by small amounts:
S = S[y + y] S[y]
x2

(
)

=
L x, y(x) + y(x), y (x) + y (x)
x1

x2

(5.10)
(
)
L x, y(x), y (x)

x1

(5.11)
The small function, y(x) is the vertical displacement of the path in our
chosen coordinate system. Note that we have considered the case where the
41

y
y(x)
(x2 ,y2)
y(x) + y(x)
(x1,y1)

Figure 5.2: Paths y(x) and y(x) + y between points (x1 , y1 ) and (x2 , y2 ).
end-points of the paths, a and b, are xed.
The power series expansion of L is:
(
)
L x, y(x) + y(x), y (x) + y (x)
(
) L
(
)
L
= L x, y(x), y (x) +
y + y + O (y)2
y
y
so, the functional variation becomes:
]
x2 [
L
L
S =
y + y dx
y
y
x1

(
) ]
x2 [
L
d L
L x2
y
=
y dx +
y
y
dx y
y x1
x1

(5.12)

(5.13)
(5.14)

Now, since we are considering xed end points for the paths, there can be no
variation in the paths at the end points such that y(x1 ) = 0 and y(x2 ) = 0.
Thus, the last term in the above equation vanishes, and we have:
(
)]
x2 [
L
d L
S =

dxy
(5.15)
y
dx y
x1
We say that the functional derivative of S with respect to y(x) is:
(
)
S
L
d L
=

y
y
dx y
42

(5.16)

5.2

The Euler-Lagrange Equation

The functional S[y(x)] in Eq. 5.9 is extremized when the functional derivative given by Eq. 5.16 vanishes, giving a dierential equation for y(x):
(
)
L
d L

=0
(5.17)
y
dx y
This equation is known as the Euler-Lagrange equation.
There are two scenarios which frequently occur in which the Euler-Lagrange
equation can be further simplied. Let us consider them below:
L(y, y , x) independent of y
In this case the Euler-Lagrange equations give:
L
P = ,
(5.18)
y
where P is a constant, that is, a conserved quantity. In classical mechanics
problems,
this will turn out to be the generalized momentum. For L =

1
2 (corresponding to Eq. 5.7), we have:
1
+
y
v
y
P =
,
v 1 + y 2

(5.19)

which corresponds to Eq. 5.6, with P = C. Solving for y :


v(x)
y = 2
v0 v 2 (x)

(5.20)

where v0 = 1/P .
L(y, y , x) independent of x
In this case, consider the quantity:
H = y
Then:

L
L
y

[
]
dH
d
L
=
y
L
dx
dx
y
(
)
L
L
L L
L
d
=y
+y
y
y

y
dx y
y
y
x
[ (
)
]
d L
L
L
= y

dx y
y
x

(5.21)

(5.22)
(5.23)
(5.24)

The quantity in square brackets vanishes, according to the Euler-Lagrange


equations. Therefore, if L/x = 0, we have dH/dx = 0, that is, H is a
conserved quantity.
43

5.3

Example 1: Brachistochrone

Which of all paths between xed initial and nal points provides the path
of least time for a particle sliding along it under gravity? Let us try to solve
this problem using the calculus of variations. Assuming that the particle
starts from rest at the point (x, y) = (0, 0), the potential energy of the
particle at x is E = mgy(x) (measure y in the downward direction). From
(0,0)
x

dl
mg

(xf , y)
f

Figure 5.3: Brachistochrone the path of fastest descent between two points
for a particle in a uniform gravitational eld and no friction.
conservation of energy the kinetic energy is then 12 mv(x)2 = mgy(x), giving

v(x) = 2gy(x). Thus, the travel time of the particle along the path y(x)
is:
xf
xf
(
)
1 + y 2

S[y(x)] =
dx =
L x, y(x), y (x) dx
(5.25)
2gy(x)
0
0
where we have identied the function L as:

1 + y 2
L=
2gy(x)

(5.26)

Using the Euler-Lagrange equation:

1 + y 2
d

3/2
dx
2g2y

L
d

y
dx

L
y

2gy 1 + y 2

44

)
=0

(5.27)

=0

(5.28)

This doesnt look like a particularly straightforward route to follow. Fortunately, there are two other routes that we can follow:
Method 1) Note that the quantity L is independent of x in this problem.
Hence the quantity H (given by Eq. 5.21) is a constant:
H = y

L
L
y

(5.29)

y 2

1
1 + y 2


2gy
1 + y 2 2gy
( 2
)
1
1

=
y 1 y 2
2gy 1 + y 2
=

(5.30)
(5.31)

Therefore:
1
2gH 2 y

dy
1
1 2gH 2 y

1
=
y =
dx
2gH 2 y
2gH 2 y

1 + y 2 =

(5.32)
(5.33)

Method 2) Alternatively, we can note that there is no reason why x has


to be the independent variable. We could just as well use y, in which case
nothing changes in our derivation of Eq. 5.17 except that the symbols y and
x should be interchanged. In this case:
yf
yf
(
)
1 + x2

S[x(y)] =
dy =
L y, x(y), x (y) dy
(5.34)
2gy
0
0
with the function L now given by:

1 + x2
L=
2gy
The Euler-Lagrange equation now gives:
(
)
L
d L

=0
x dy x

(5.35)

(5.36)

This simplies the problem since the rst term vanishes, giving:
L
=P
x

(5.37)

where P is a constant (as in the case of Eq. 5.18, with x and y inerchanged).
Therefore:
1
x

P =
2gy 1 + x2
45

(5.38)

Solving for x :
x2
2gyP 2
(
)
2
2gyP = x2 1 2gyP 2

dx
2gP 2 y

x =
=
dy
1 2gP 2 y

1 + x2 =

(5.39)
(5.40)
(5.41)

which is no dierent to Eq. 5.33 with P = H.


Now, let us integrate Eq. 5.41. Let 2a = 1/(2gP 2 ), then:

2gP 2 y
x(y) =
dy
1 2gP 2 y

=
dy
2ay y 2

dy
=
2
2
a a + 2ay y 2

(y a) + a

=
dy
a2 (y a)2

(5.42)
(5.43)
(5.44)
(5.45)

For the rst part of the integral, use the substitution (y a)2 = z. In the
second part of the integral, use the substitution (y a) = a sin :

1
a2 cos
1

x(y) =
dz +
d
(5.46)
2
a2 z
a2 a2 sin2

= a2 z + a + b
(5.47)
(
)

ya
= a2 (y a)2 + a sin1
+b
(5.48)
a
Given that the starting point of the curve has x(0) = 0, the constant b =
a/2 (since sin1 (1) = /2). This allows the curve to be re-written as:
(
)

1 a y
2
x(y) = 2ay y + a cos
(5.49)
a
The end-point of the curve can be used to determine the constant a. This
curve is known as a cycloid.

46

5.4

Example 2: Minimal Surface of Revolution

Imagine you have two circular wire hoops, put them in a soap solution and
draw them out, keeping their planes parallel. In this case a soap lm can
form connecting the two hoops, and the minimum energy solution is the one
with the smallest surface area, called a catenoid. Let us try to calculate
the shape of this surface. Let us take the x-axis as the axis connecting the

x1

x2

Figure 5.4: Minimal area surface connecting two hoops.


centers of the two hoops. The surface area of the soap lm is then:

( )2
x2
dy
A[y(x)] =
2y 1 +
dx
(5.50)
dx
x1
where we assume cylindrical symmetry such that the surface is formed by
rotating the function y(x) about the x-axis. Let us identify the function,
L (x, y(x), y (x)) as:

(
)
L x, y(x), y (x) = 2y 1 + y 2
(5.51)
Note that L (x, y(x), y (x)) is independent of x, so we can take the quantity
H given by Eq. 5.21 as constant:
L
L
y

y 2
= 2y
2y 1 + y 2
1 + y 2
2y
=
1 + y 2

H = y

47

(5.52)
(5.53)
(5.54)

Solving for y :

(
)
dy
2y 2
1
=
dx
H

H
To integrate, we use the substitution y = 2
cosh u:

dy
x = ( )
2
2y
1
H

H
sinh u

=
du
2
cosh2 u 1
(
)
H
=
u+a
2
H
where a is a constant of integration. Let b = 2
, then:
)
(
xa
y(x) = b cosh
b

(5.55)

(5.56)

(5.57)
(5.58)

(5.59)

The constants a and b are determined by the boundary conditions of the


problem, given by the positions of the two hoops, x0 and x0 , and their
radii, y(x0 ) = y0 and y(x0 ) = y0 . Since we have chosen x = 0 as the
mid-point between the two hoops, a = 0. Then:
y0 = b cosh

x0
b

(5.60)

is used to determine b.
Let us characterize the system in terms of the ratio of the hoop radius to
half the distance between the hoops, = y0 /x0 . In this case:
=

1
cosh

(5.61)

where we dened = x0 /b. Note that for > 1.5089 there are two solutions
for () for a given , which can be found numerically.
Recall that when minimizing functions, setting the derivative to zero may
give not only the minimum points, but also maximum and inection points.
It is the same case when minimizing functionals. There may be multiple
solutions and in some cases one needs to evaluate the functional explicitly
for these solutions to determine which is the minimum. In our example, the

48

area is given by:

(x)

(x)
1 + sinh2
dx
2b cosh
b
b
x0
x0
(x)
= 2b
dx
cosh2
b
x0
x0

A[y(x)] =

= b (2x0 + b sinh (2))

(5.62)
(5.63)
(5.64)

Let us normalize the area by the quantity A0 = 2y02 . Then:


A[y(x)]
b
= 2 2 (2x0 + b sinh (2))
A0
2x0
(
)
1
sinh (2)
=
2
+
2 2

(5.65)
(5.66)

The area of these solutions is shown in Fig. 5.5 as a function of .

1.4
1.2

AA0

1.0
0.8
0.6
0.4
0.2
0.0
1.0

1.2

1.4

1.6

1.8
=y0x0

2.0

2.2

2.4

Figure 5.5: Areas of Catenoid solutions. The curves show the two solutions
arising from Eq. 5.66. The dotted line corresponds to the end-point solution
A = A0 .
While minimizing functions between nite intervals, we also know that there
may be minima at the end-points of the interval. These minima may not
necessarily have zero gradient, so we dont nd them by setting the gradient
to zero. The same is true with functionals. We may need to evaluate the
functional for functions at the boundary of function space as it may correspond to the minimum. In our current example, a function at the edge of
the space of all available functions is the discontinuous solution:

y0 if x = x0
0 if |x| < x0
y(x) =
(5.67)

y0 if x = x0
49

This corresponds to a surface of two discs of radii y0 joined by an innitesimally thin thread. The area of this solution is A = 2y02 = A0 .

5.5

Example 3: Ocean Waves

The dispersion relation, (k), of waves propagating on the surface of water


is:
{

ghk if kh 1 (shallow)
(k) = gk tanh(kh)
(5.68)
gk if kh 1
(deep)
where h is the water depth and g is a constant. The group velocity of the
waves is given by v(k) = d/dk.
Let us consider the shallow case, where the wavelength of the waves, =
2/k, is
much greater than the water depth. The wave velocity is then
v(h) = gh. The waves propagate slower as they approach the shore.
Let x represent the distance parallel to the shoreline and y the distance

Figure 5.6: Shallow water waves emanating from a point source to the shore.
perpendicular to the shore (at
on y and h(y) = 0. Consider
propagating until it reaches the
is:

T [y(x)] =

y = 0). Assume that h(y) depends only


a wave generated at the position (x2 ,y2 ),
shore at (x1 ,y = 0). The propagation time
x2
ds
1 + y 2
=
dx
(5.69)
v
gh(y)
x1

50

To minimize the propagation time, let us identify the function:

1 + y 2

L(x, y, y ) =
gh(y)

(5.70)

Note that L(x, y, y ) is not directly dependent of x, hence the quantity H,


given by Eq. 5.21, is constant:
H = y

L
L
y

gh(y) 1 + y 2

(5.71)
(5.72)

Solving for y :
1
h(y)(1 + y 2 )
1
(1 + y 2 ) =
2
gH h(y)

dy
a
=
1
dx
h(y)

gH 2 =

(5.73)
(5.74)
(5.75)

with a = 1/(gH 2 ). We can conclude that near the shore, where h(y) tends
to zero, dy/dx 7 , and the waves come in parallel to the shoreline.
If we assume that the water depth has a linear prole, h(y) = y, then
Eq. 5.75 has the same form as Eq. 5.33 in the Brachistochrone problem. In
this case the waves propagate along a cycloid curve and, taking x(y = 0) = 0,
the solution can again be written:
(
)

1 a y
2
(5.76)
x(y) = 2ay y + a cos
a
this time with a = 1/(2gH 2 ).

51

Вам также может понравиться