Chapter 5

Calculus of Variations
In this chapter we will:
introduce the basic concepts in the calculus of variations
study how conservation laws in physics are connected to symmetries of the
Universe
apply this mathematical toolkit to several problems in mathematics and
physics
As we have seen in the previous chapters, Newton revolutionalized the study of
physical phenomena by introducing a number of simple laws expressed mathemati-
cally in the new language of calculus. These laws were regarded to be universal in
their validity and to be applicable for the motions of all objects, from the smallest
particle on Earth to the largest planet in the Solar System.
Contrary to earlier empirical attempts that aimed to describe the trajectories
of objects in terms of simple geometric curves, such as the circles and epicycles of
Ptolemys or the ellipses of Keplers, Newtons laws of motion are not concerned
with the long-term trajectories of objects. Instead, the second law is very local and
simply connects the instantaneous acceleration of an object to the net force acting
on it at that same point in time. Calculating a trajectory requires the integration
of the resulting dierential equation with some appropriate initial conditions. The
fact that a projectile on Earth follows a parabolic path or that a planet orbits the
Sun in an elliptical orbit are results of calculations and not a priori assumptions in
Newtonian mechanics.
Remarkably, however, it turns out that objects follow trajectories along which
quantities such as the linear momentum, the angular momentum, and often the me-
chanical energy are conserved in a global sense. This property allows us to calculate
the trajectories of objects in many cases simply using integrals of motion, without
ever having to integrate Newtons second law of motion. Of all possible trajectories
in space, we simply need to choose the one along which the three integrals of motion
are conserved. In other words, even within Newtonian mechanics, we can infer the
path of an object using global arguments and conservation laws instead of local
141
142 CHAPTER 5. CALCULUS OF VARIATIONS
considerations of forces and accelerations.
Starting in the years immediately after the publication of the Principia, sev-
Leonhard Euler
(1707-1783)
Joseph-Louis Lagrange
(1736-1813)
eral mathematicians and natural scientists attempted to reformulate Newtonian
mechanics in terms of global considerations. The German mathematician Leonhard
Euler (17071783) developed the most successful formalism that is widely used to-
day, based on the principle of least action. His work was perhaps inuenced by the
ideas of the French philosopher Pierre-Louis Moreau de Maupertuis (1698-1759),
who used a similar principle to derive Snells law for the refraction of light. Fol-
lowing on their footsteps, Joseph-Louis Lagrange (17361813), Carl Gustav Jacob
Jacobi (18041851), William Rowan Hamilton (1805-1865) and others further de-
veloped the formalism to its present form.
The principle of least action states that an object follows a trajectory along
which a particular integral of its kinetic and potential energy, called the action,
is at an extremum. Performing calculations with this principle requires that we
rst develop the mathematical tools to calculate the extrema of integral quantities.
The eld of mathematics that deals with such calculations is called the Calculus
of Variations. In this chapter, we will develop the basic ideas in the calculus of
variations and return to the principle of least action as an alternative formulation
of Newtonian mechanics in the following chapter. A more detailed study of the
subject can be found in several textbooks
[1],[2]
.
5.1 The Fundamental Problem in the Calculus of
Variations
The development of the calculus of variations started with a challenge posed by
Johann Bernoulli (16671748) to several of his contemporaries, including Newton.
The goal was simple: calculate the trajectory of an object in a uniform gravitational
eld such that, in the absence of friction, the time it takes for the object to travel
from point A to point B is the least possible. Clearly, the shortest path from point
A to point B is the straight line that connects the two points. However, along the
straight line, the acceleration is constant and not necessarily optimal. Wouldnt it
be better if the path close to point A is more vertical so that the velocity of the
object increases rapidly and then attens or even turn upwards towards point B?
The optimal curve is called the brachistochrone (from the Greek words for shortest
and time).
In order to calculate the optimal curve we set up a two-dimensional Cartesian
coordinate system on the vertical plane that contains the two points A and B. Our
! " #$%&' ( ) * +, - . / 0123456789: ; <=>?@ABCDEFGHI J KLMNOPQRSTUVWXYZ[ \ ] ^ _` abc def ghi j k l mnopqr s t uv wx y z { | } ~
!
"
#
$
#
%
&
goal is to nd the path x
2
= x
2
(x
1
) that minimizes the time it takes for an object
to move from point A to point B,
t
AB
=
_
B
A
dt =
_
B
A
ds
v
(5.1)
In this integral, ds is the dierential path length covered by the object and v is its
instantaneous speed. We can calculate the former using the Pythagorian theorem
5.1. THE FUNDAMENTAL PROBLEMIN THE CALCULUS OF VARIATIONS143
as
ds =
_
dx
2
1
+ dx
2
2
_
1/2
=
_
1 + x
2
2
_
1/2
dx
1
,
where prime denotes the derivative of x
2
with respect to x
1
. We can also evaluate
the velocity using the conservation of energy
1
2
mv
2
+ mgx
2
= mgx
A
2
v =
_
2g(x
A
2
x
2
)

,
where m is the mass of the object, x
A
2
is the vertical coordinate of point A, and we
assumed that the object has zero velocity initially. Inserting these expressions into
equation (5.1), we obtain
t
AB
=
_
x
B
1
x
A
1
_
1 + x
2
2
2g(x
A
2
x
2
)
_
1/2
dx
1
. (5.2)
The right-hand side of equation (5.2) depends on the unknown path x
2
(x
1
) as
well as on its derivative. Such an expression, the argument of which is a function,
is called a functional. The solution to our problem will be the path x
2
(x
1
) that
minimizes this functional. Because we are looking for a path that extremizes a
functional, we cannot simply use normal calculus to nd the solution. Indeed, the
brachistochrone is a particular case of the fundamental problem in the calculus of
variations. For this reason, we will rst develop the general mathematical tools
to deal with such problems and then return to the brachistochrone solution in the
second example below.
We begin by considering a general function y(x) that is twice dierentiable in
a domain and goes through the xed values y
A
and y
B
at its boundaries x
A
and
x
B
. The problem we aim to solve is to nd the function y(x) that extremizes the
integral between the boundaries
J =
_
xB
xA
f [y(x), y

(x); x] dx . (5.3)
The integrand in this expression is a function f that may depend on y(x), on
its derivative y

(x), as well as on the independent variable x. Note that we are

considering here the particular case of one independent variable (x), one function
y(x), and an integrand that depends at most on the rst derivative of the function.
We will consider more general cases later in this chapter.
We denote the solution to the problem by y
0
(x) and consider a second arbitrary,
! " #$%&' ( ) * +, - . / 0123456789: ; <=>?@ABCDEFGHI J KLMNOPQRSTUVWXYZ[ \ ] ^ _` abc def ghi j k l mnopqr s t uv wx y z { | } ~
! "
#
$
# #
!
$
"
$
$ &#'
(
$ &#)'
twice dierentiable function n(x) that vanishes at the boundaries of the domain,
i.e.,
n(x
A
) = n(x
B
) = 0 . (5.4)
Clearly, the linear combination
y(x, ) = y
0
(x) + n(x)
also satises y(x
A
, ) = y
A
and y(x
B
, ) = y
B
, for all values of the parameter .
The new parametric function y(x, ) represents dierent paths between the two
xed boundaries, one for each value of the parameter , which do not extremize
the integral (5.3) unless = 0.
144 CHAPTER 5. CALCULUS OF VARIATIONS
Given that we have xed the solution y
0
(x) and the arbitrary function n(x), the
integral (5.3) is not a functional any longer but now becomes a simple function of
the parameter , i.e.,
J() =
_
xB
xA
f [y(x, ), y

(x, ); x] dx . (5.5)
By construction, this function has an extremum at = 0 and, therefore,
dJ
d

= 0
d
d
_
xB
xA
f [y(x, ), y

(x, ); x] dx = 0

_
xB
xA
_
f
y
y(x, )

+
f
y

(x, )

_
dx = 0 , (5.6)
where, in the last expression, we used the chain rule to calculate the total derivative
of the integrand f with respect to the parameter . By construction,
y(x, )

= n(x)
and
y

(x, )

_
dy(x, )
dx
_
=
d
dx
_
y(x, )

_
=
dn(x)
dx
.
Inserting these expressions into the integral (5.6), we obtain
_
xB
xA
_
f
y
n(x) +
f
y

dn(x)
dx
_
dx = 0 , (5.7)
In order to make further progress, we will split this last integral into two and
evaluate the second of them using integration by parts as
_
xB
xA
f
y

dn(x)
dx
dx =
_
xB
xA
f
y

dn(x)
=
f
y

n(x)

xB
xA

_
xB
xA
n(x)
d
dx
_
f
y

_
=
_
xB
xA
n(x)
d
dx
_
f
y

_
,
where we set the boundary term to zero because of equation (5.4). Inserting this
last expression into equation (5.7) we nal obtain the result
_
xB
xA
_
f
y

d
dx
_
f
y

__
n(x)dx = 0 , (5.8)
which is a neccesary condition for the function J() to have an extremum at = 0,
for a given function n(x).
Our initial aim was to nd the path y
0
(x) the extremizes the functional J,
independent of the particular function n(x). As a result, the condition (5.8) has to
be satised for any arbitrary function n(x). This last requirement can be met only
if the term in the square brackets vanishes. In other words, a necessary condition
for a path y(x) to extremize the functional J given by equation (5.3) is for the
integrand f to satisfy the dierential equation
f
y

d
dx
_
f
y

_
= 0 . (5.9)
Euler-Lagrange
Equation
5.1. THE FUNDAMENTAL PROBLEMIN THE CALCULUS OF VARIATIONS145
We call this the Euler-Lagrange equation of the calculus of variations to honor Euler,
who rst derived it using geometric considerations, and Lagrange, who reached
the same result using calculus. Before discussing some important aspects of this
equation, we will rst apply it to solve a simple but illustrative example.
Example 5.1: A Straight Line is the Shortest Path Between Two
Points on a Plane
In this example, we will prove using Calculus of Variations the well known result
that the straight line is the shortest path between two points on a plane.
As in our discussion of the Brachistochrone, we will set up a Cartesian coordinate
system on the plane and consider two points A and B with coordinates (x
A
1
, x
A
2
)
and (x
B
1
, x
B
2
), respectively. If we follow an arbitrary path x
2
= x
2
(x
1
) between the
two points, the total path length will be
S
AB
=
_
B
A
ds =
_
x
B
1
x
A
1
_
1 + x
2
2
_
1/2
dx
1
.
Our goal is to nd the path that extremizes this integral. To achieve this, we can
use the Euler-Lagrange equation with the identications
x x
1
y(x) x
2
(x
1
)
f[y(x), y

(x); x]
_
1 + x
2
2
_
1/2
.
The Euler-Lagrange equation involves two derivatives, which we evaluate for the
present problem as
f
x
2
= 0
and
f
x

2
=
x

2
(1 + x
2
2
)
1/2
.
Inserting these expressions into equation (5.9) we nd that the path that minimizes
the length between points A and B must satisfy the dierential equation
d
dx
1
_
x

2
(1 + x
2
2
)
1/2
_
= 0 .
This dierential equation, which is of second order in the independent variable x
1
has the simple solution
x

2
(1 + x
2
2
)
1/2
= constant C ,
which we can convert into an expression for x

2
as
x

2
=
C
(1 C
2
)
1/2
C
1
.
This is now a rst-order dierential equation with respect to x
1
that also has the
simple linear solution
x
2
= C
1
x
1
+ C
2
. (5.10)
146 CHAPTER 5. CALCULUS OF VARIATIONS
In order to complete our solution, we need to specify the constants of integration
C
1
and C
2
by requiring that the general solution passes through the points A and
B, i.e.,
x
A
2
= C
1
x
A
1
+ C
2
and
x
B
2
= C
1
x
B
1
+ C
2
.
Solving these two algebraic equations for C
1
and C
2
and inserting the result in the
general expression (5.10) we obtain
x
2
=
_
x
B
2
x
A
2
x
B
1
x
a
1
_
_
x
1
x
A
1
_
+ x
A
2
.
This is, of course, nothing but the equation for a straight line that connects points
A and B on the plane.
Note that, formally, we have only shown that the straight line is an extremum to
the path length between the two points and not necessarily a minimum. In this and
in the following examples we will refrain from proving explicitly that the various
extrema are indeed minima or maxima, depending on the problem, because these
proofs are often very tedious and beyond the scope of our exploration of the calculus
of variations in the context of mechanics.
There are two important aspects of the derivation of the Euler-Lagrange equa-
tion that deserve close inspection. First, this equation provides a necessary condi-
tion for the path y(x) to extremize the functional J but not a sucient one. Indeed,
the Euler-Lagrange equation (under the conditions considered here) is, in general,
a second-order dierential equation in x with a solution that is required to satisfy
two conditions at the boundaries of the domain of solution. Such boundary value
problems may have no solution, one unique solution, or multiple solutions depend-
ing on the situation. A case with multiple solutions will imply that more than one
paths from point A to point B satisfy the Euler-Lagrange equation. However, not
all of these paths will necessarily minimize the functional J. We will explore such
situations in the following application as well as in Example 5.3 below.
A second important aspect of the Euler-Lagrange equation is related to our
assumption that the path that extremizes the functional J is described by a twice
dierentiable function y(x). Indeed, our considerations focused only on such smooth
functions. However, the actual path that extremizes an integral might be one with
a corner or a kink. Such paths are not relevant for the use of the Euler-Lagrange
equation in Newtonian mechanics. However, they are often the true solutions in
other problems in the calculus of variations, as we will see in the next section, when
we study the physics of soap lms.
5.2 Application: The Physics of Soap Films
An instructive application of the calculus of variations, which is often demonstrated
in many science museums around the world, is the determination of the shape of
5.2. APPLICATION: THE PHYSICS OF SOAP FILMS 147
a thin soap lm suspended in space within a wire frame. We will use this ap-
plication here in order to explore the various ne points of the solutions to the
Euler-Lagrange equation. The properties of thin soap lms (and of soap bubbles)
were identied experimentally by the Belgian physicist Joseph Antoine Ferdinand
Plateau (18011883), who determined a number of empirical laws for the formation
of their surfaces. To honor his work, this application is called the Plateau problem
and the corresponding laws are called Plateaus laws.
In order to make the problem tractable, we will rst consider the case of a thin
soap lm suspended between two concentric wire rings of radius
A
and
B
. The
conguration is axisymmetric around the axis that connects the two centers of the
rings. For this reason, we will set a cylindrical coordinate system with the zaxis
oriented along the axis of symmetry and its origin at the mid-distance between the
two rings. In order to describe the surface of the soap lm in this setup, we only
need to specify a function (z) that measures the distance of the surface from the
zaxis.
! " #$%&' ( ) * +, - . / 0123456789: ; <=>?@ABCDEFGHI J KLMNOPQRSTUVWXYZ[ \ ] ^ _` abc def ghi j k l mnopqr s t uv wx y z { | } ~
!

"

#
$%
$!
&
The dominant force that allows the soap lm to retain its shape is surface tension.
The energy related to this process is proportional to the magnitude of the surface
tension and the total surface area of the lm. The lm will, therefore, adjust its
shape in order for its surface area to be minimal, subject to the requirement that
it will be attached to the two wire rings. The surface area of the lm is
S =
_
2
0
d
_
dl ,
where
dl
_
dz
2
+ d
2
_
1/2
=
_
1 +
d
dz
_
1/2
dz
is the dierential length along the curve (z). Our goal is to nd the curve that
minimizes the integral
S = 2
_
L/2
L/2

_
1 +
d
dz
_
1/2
dz ,
where L is the distance between the two rings.
Setting
f
_
1 +
d
dz
_
1/2
dz
we can write the Euler-Lagrange equation for this problem as
f

d
dz
_
f

_
= 0

_
1 +
2
_
1/2

d
dz
_

(1 +
2
)
1/2
_
= 0
and, after a few lines of algebra, bring it to the more compact form
1
(1 +
2
)
1/2

(1 +
2
)
3/2
= 0 .
148 CHAPTER 5. CALCULUS OF VARIATIONS
This is a second-order dierential equation in z that is, in fact, equivalent to
d
dz
_

(1 +
2
)
1/2
_
= 0 , (5.11)
as a quick expansion of the derivative in the left-hand side of the latter expres-
sion demonstrates. Equation (5.11) implies that, along the surface of the lm, the
expression within the square brackets is a constant, or

(1 +
2
)
1/2
= C
1
.

2
C
2
1
1 .
It is easy to verify by substitution that the dierential equation is satised by
the function
(z) = C
1
cosh
_
z + C
2
C
1
_
,
which is called the catenary. Here c
2
is a second integration constant with a value
to be determined together with C
1
by the requirement that the soap lm is attached
to the two wire rings, i.e., z(L) =
A
and z(L) =
B
.
We will now simplify the problem even further and consider the particular case
of two rings with the same radius, i.e.,
A
=
B
=
0
. The mirror symmetry of the
boundary conditions with respect to the z = 0 plane suggests that the solution will
also be symmetric in z, i.e., (z) = (z). In order for the hyperbolic cosine to be
symmetric in z, the constant of integration C
2
has to vanish, i.e., C
2
= 0.
In order to obtain the value of C
1
, we need to nd the solution to the algebraic
equation

0
= C
1
cosh
_
L
2C
1
_
, (5.12)
which is not known in an analytic form. We can, nevertheless, investigate graph-
ically the number and types of solutions to this equation by rewriting it in the
form

0
C
1
= cosh
_
1
2
L

0
_

0
C
1
__
,
setting
y
1
=

0
C
1
(5.13)
y
2
= cosh
_
1
2
L

0
_

0
C
1
__
. (5.14)
and plotting the two functions y
1
and y
2
as a function of x
0
/C
1
. This is
shown in the margin gure, where equation (5.13) is plotted as a dashed line and

equation (5.13) is plotted as a series of solid lines for dierent values of the parameter
L/
0
. The points of intersection of the dashed line with each solid curve correspond
to the solutions to the algebraic equation (5.12). A quick glance at this gure reveals
two interesting properties of the solutions.
First, when the distance between the two rings is smaller than or comparable
to their radius, i.e., when L/
0
1.33, there are two values of the parameter C
1
5.3. THE BELTRAMI IDENTITY 149
that satisfy equation (5.12). For example, when L =
0
, the two solutions are
C
1
= 0.235
0
and C
1
= 0.85
0
. Each value corresponds to a dierent equation
for the surface, as shown in the second margin gure. However, only one of these

surfaces minimizes the surface area and, in this case, is the one that corresponds to
the smaller value of C
1
. This is a clear demonstration of the fact that the Euler-
Lagrange equation is a necessary but not sucient condition for the solution to the
original variational problem.
Second, when the distance between the two rings is larger than 1.33
0
, the
algebraic equation (5.12) has no real solutions. This occurs because the area of a
continuous surface between the two wire rings never attains a stationary value. It
does not mean, however, that we cannot nd a surface attached to the two rings
with a minimal area; it just has to be discontinuous. In this highly symmetric
conguration, the surface of minimal area consists of the two disconnected disks
that are surrounded by the wire rings. In other situations, as in the case of the
cubic wire frame shown in the gure, the minimal surface is not smooth, but is
characterized by intersections with discontinuous derivatives.
The surface of a thin soap lm
suspended from a wire cube
credit: sciencephoto.com
Plateau discovered experimentally that soap lms (or walls of soap bubbles) can
meet only at angles of 120

(when three walls intersect) and at angle of cos

1
(1/3)
110

(when four walls intersect). A theoretical proof of Plateaus laws did not be-
come possible until the mid-20th century, with the invention of advanced concepts
and techniques in geometry
[3]
.
5.3 The Beltrami Identity
The Euler-Lagrange equation (5.9) is, in general, a second-order dierential equa-
tion in the independent variable x. If, however, the integrand f does not depend
explicitly on x, then it is equivalent to a rst-order dierential equation rst derived
by the Italian mathematician Eugenio Beltrami (18351900). Because of its lower
order, the Beltrami identity is very useful in the solution of problems that satisfy
the requirements for its validity.
In order to prove the Beltrami identity, we start by evaluating two particular
derivatives, for reasons that will become obvious very quickly. First, is the total
derivative of the integrand f, i.e.,
df
dx
=
f
x
+
f
y
y

+
f
y

dy

dx
. (5.15)
Second is the second derivative
d
dx
_
y

f
y

_
=
dy

dx
f
y

+ y

d
dx
_
f
y

_
. (5.16)
Combining equations (5.15) and (5.16) under the assumption that the integrand f
does not depend on x, i.e.,
f
x
= 0 ,
we nd, after a few lines of algebra,
d
dx
_
f y

f
y

_
= y

_
f
y

d
dx
_
f
y

__
.
150 CHAPTER 5. CALCULUS OF VARIATIONS
If the path y(x) minimizes the functional J, then the Euler-Lagrange equation leads
to the conclusion that the right-hand side of the above equation vanishes. In other
words,
d
dx
_
f y

f
y

_
= 0 ,
or, equivalently, the term in the parenthesis is constant, independent of x,
f y

f
y

= C . (5.17)
Beltrami Identity
Having developed the necessary tools in the calculus of variations, we are now in
the position to solve the Brachistochrone problem introduced in the beginning of
this section.
Example 5.2: The Brachistochrone
In this example, we will nd the path x
2
(x
1
) between two xed points A and B
in a uniform vertical gravitational eld g such that the time it takes for an object
initially at rest at point A to reach point B is minimized.
Comparing the expression (5.2) we derived for the time it takes an object to com-
plete the motion to the general form of the functional (5.3), we see that we can
solve this problem using the Euler-Lagrange equation with the identications
x x
1
y(x) x
2
(x
1
)
f[y(x), y

(x); x]
_
1 + x
2
2
2g(x
A
2
x
2
)
_
1/2
.
The integrand f does not depend on the variable x
1
. As a result, we can use the
Beltrami identity to calculate the Brachistochrone curve. We only need to evaluate
the derivative
f
x

2
=

x

2
_
1 + x
2
2
2g(x
A
2
x
2
)
_
1/2
=
_
1
2g(x
A
2
x
2
)
_
1/2
x

2
(1 + x
2
2
)
1/2
so that the identity becomes
f x

2
f
x

2
= C

_
1 + x
2
2
2g(x
A
2
x
2
)
_
1/2
_
1
x
2
2
1 + x
2
2
_
= C

_
1
2g(x
A
2
x
2
)(1 + x
2
2
)
_
1/2
= C .
At this point, we make a change of variables to
z = x
A
2
x
2
,
which measures the vertical displacement from point A. In this case,
z

= x

2
5.4. VARIATIONAL PROBLEMS WITH TWOORMORE DEPENDENT VARIABLES151
and the Beltrami identity becomes, after squaring it and rearranging the terms,
z(1 + z
2
) =
1
2gC
2
. (5.18)
This is a famous dierential equation with a solution that is called the cycloid
and is given in the parametric form (expressed again in terms of the original vari-
ables)
x
1
() =
1
4gC
2
( sin) + x
A
1
x
2
() = x
A
2

1
4gC
2
(1 cos ) .
The coordinates of point A correspond to = 0, whereas the coordinates of point B
correspond to some other value
0
> 0. Requiring that the cycloid passes through
point B allows us to specify both the value of the parameter
0
as well as the
parameter C.
An example of a cycloid between two points A and B is shown in the margin
gure. As expected, the path is characterized by an initial steep downward slope
that allows the object to reach quickly a high velocity. Remarkably, depending on
the location of point B, it is often optimal for the brachistochrone curve to overshoot
the vertical location of point B for the object to attain a high enough velocity, before
turning upwards towards the xed end point. It is often suggested that this curve is
the optimal cross section of the half-pipe used in skate boarding and snow boarding
events. It is also worth pointing out that the cycloid is the curve traced out by a
point on the rim of a circle that is rolling on a horizontal plane. This and several
other properties of the cycloid are the subject of exercise 5.1 at the end of this
chapter.
5.4 Variational Problems with Two or More De-
pendent Variables
All the problems we have considered so far involve nding a path y(x) in a two-
dimensional space such that the integral (5.3) is extremized. We can now easily
generalize the same procedure in the case of two or more dependent variables,
y
i
(x), i = 1, ..., N, which correspond to nding the path in an N-dimensional space
that extremizes the integral
J =
_
xB
xA
f [y
1
(x), y

1
(x), ..., y
N
(x), y

N
(x); x] dx .
As before, we will denote by y
0
i
(x) the functions that extremize the integral and
consider N independent functions n
i
(x) that vanish at the boundaries, i.e.,
n
i
(x
A
) = n
i
(x
B
) = 0 , i = 1, ..., N .
We will then construct a parametric deviation of each function y
i
(x) from its optimal
form along the deviation described by the functions n
i
(x) as
y
i
(x, ) = y
0
i
(x) + n
i
(x) .
152 CHAPTER 5. CALCULUS OF VARIATIONS
Because we have specied both the solutions y
0
i
(x) and the functional deviations
n
i
(x), the functional J becomes a simple function of the parameter ,
J() =
_
xB
xA
f [y
1
(x, ), y

1
(x, ), ..., y
N
(x, ), y

N
(x, ); x] dx ,
with an extremum at = 0 or
dJ
d

= 0
_
xB
xA
N

i=1
_
f
y
i
y
i
(x, )

+
f
y

i
y

i
(x, )

_
dx = 0 ,
Following a similar derivation to that in section 5.1.1, we nd
y
i
(x, )

= n
i
(x)
and
y

i
(x, )

=
dn
i
(x)
dx
.
Inserting these expression into the above integral and performing integration by
parts we nally obtain
_
xB
xA
N

i=1
_
f
y
i

d
dx
_
f
y

i
__
n
i
(x)dx = 0 , (5.19)
which is a necessary condition for the function J() to have an extremum at = 0,
given the functions n
i
(x). Because the functions n
i
(x) are arbitrary and indepen-
dent of each other, this integral can vanish only if the coecient of each function
in the sum vanishes, i.e., if
f
y
i

d
dx
_
f
y

i
_
= 0 . (5.20)
Euler-Lagrange Equations with
Many Dependent Variables
In other words, each independent function y
i
(x) in the solution that extremizes the
integral between points A and B satises an independent Euler-Lagrange equation
of the form (5.20).
5.5 Constraints and Lagrange Multipliers
In our discussion so far, we have considered the problem of nding a path through
a space along which a particular integral is extremized, with no other constraints
imposed on the properties of the path itself. This is not general, however, as, in
several situations the setup of a problem forces us to consider only a particular set
of paths that satisfy dierent conditions.
In dierent problems, constraints may be imposed in a variety of ways. For
example, they may be given in terms of algebraic relations between the functions
y
i
of the form
g
j
[y
1
(x), ..., y
N
(x); x] = 0 , j = 1, ..., K . (5.21)
We call such constraints holonomic. Alternatively, the algebraic relations may also
involve the derivatives of the functions, i.e.,
g
j
[y
1
(x), y

1
(x), ..., y
N
(x), y

N
(x); x] = 0 , j = 1, ..., K . (5.22)
5.5. CONSTRAINTS AND LAGRANGE MULTIPLIERS 153
When these dierential equations cannot be integrated and converted into algebraic
equations, we call them non-holonomic constraints. In other situations, constraints
may also be given as integrals of the form
_
xB
xA
g
j
[y
1
(x), y

1
(x), ..., y
N
(x), y

N
(x); x] = 0 , j = 1, ..., K , (5.23)
or even as inequalities,
g
j
[y
1
(x), y

1
(x), ..., y
N
(x), y

N
(x); x] 0 , j = 1, ..., K . (5.24)
In this section, we will rst explore in detail variational problems with holonomic
constraints and then discuss strategies of addressing non-holonomic and integral
constraints.
One way of addressing problems with holonomic constraints is by substituting
the constraints into the integrand that we try to minimize and, hence, reducing the
dimensionality of the problem. In the following example, we explore this approach
in nding the path of shortest distance between two points on a sphere.
Example 5.3: Geodesics on a Sphere
In this example, we will prove the well known fact, that the path of shortest distance
between two points on the surface of a sphere lies along the great circle that connects
the two points. Such curves of minimum path length between two points are called,
in general, geodesics. They are generalizations of the denition of a straight line in
a curved space.
We start by setting a spherical-polar coordinate system with its origin at the center
of the sphere. The path length from point A to point B is then given by the integral
(see Example 1.11)
S
AB
=
_
B
A
ds =
_
B
A
_
dr
2
+ r
2
d
2
+ r
2
sin
2
d
2
_
1/2
.
Our goal is to nd the path of least length subject to the constraint
r = r
0
,
where r
0
is the radius of the sphere. This is a holonomic constraint, as it can be
expressed by the algebraic relation
g(r, , ) = r r
0
= 0 .
One way to attack this problem is by using the constraint to reduce the dimen-
sionality of the domain of solution. The constraint implies that, along the required
path, r = r
0
=constant and dr = 0. The path length, therefore, becomes
S
AB
= r
0
_
B
A
_
d
2
+ sin
2
d
2
_
1/2
= r
0
_
B
A
_
d
2
d
2
+ sin
2

_
1/2
d
and our goal reduces to nding the unconstrained path () that minimizes this
integral.
154 CHAPTER 5. CALCULUS OF VARIATIONS
The integrand in the above expression is independent of the coordinate (which
is why we chose to write the path as () instead of the equivalent ()) and,
therefore, we can use the Beltrami identity to nd the extreme of the integrand.
Setting
f =
_

2
+ sin
2

_
1/2
we write the identity as
f

= C
1

2
+ sin
2

_
1/2

2
+ sin
2

_
1/2
= C
1

sin
2

2
+ sin
2

_
1/2
= C
1
.
This last equation is a rst order dierential equation, which we can bring into a
more familiar form by writing it as
1

d
d
=
C
1
sin
_
sin
2
C
2
1
_
1/2
. (5.25)
Note that the sign of the right-hand side can be absorbed into the constant C
1
,
which is still unspecied. The solution to this equation is
= sin
1
_
C
1
1
cot
_
+ C
2
,
where C
2
is the constant of integration. We can then obtain the values of the two
constants C
1
and C
2
by the requirement that the path passes through the points A
and B.
In order to see that our solution describes the great circle between points A and
B, we rst solve it in terms of cot , i.e.,
cot = C
1
sin( C
2
)
cot = C
1
sin cos C
2
C
1
cos sin C
2
,
and then multiply both sides of this equation by r
0
sin to obtain
r
0
cos = (C
1
cos C
2
)r
0
sin sin (C
1
sin C
2
)r
0
sin cos .
We now make a change of variables to the Cartesian coordinates
x
1
= r
0
sin cos
x
2
= r
0
sin sin
x
3
= r
0
cos
and set
= C
1
cos C
2
and
= C
1
sin C
2
5.5. CONSTRAINTS AND LAGRANGE MULTIPLIERS 155
to write our solution as
x
1
x
2
+ x
3
= 0 .
This last expression describes the intersection of a plane in the three-dimensional
space that goes through the origin of the coordinate system with the surface of the
sphere at r = r
0
. This intersection is nothing but a great circle on the surface of
the sphere!
The example we just solved also oers a second demonstration of the fact that a
! " #$%&' ( ) * +, - . / 0123456789: ; <=>?@ABCDEFGHI J KLMNOPQRSTUVWXYZ[ \ ] ^ _` abc def ghi j k l mnopqr s t uv wx y z { | } ~
!
"
#
$
solution to the Euler-Lagrange equation (or equivalent the Beltrami identity) oers
only a necessary but not a sucient conditions for the existence of an extremum.
Indeed, a path that connects the point A and B along the same great circle but
along the back side of the sphere (shown as long-dashed line in the margin gure) is
also a solution to the dierential equation (5.25) with the same boundary conditions.
However, unless points A and B lie on a diameter, the length of this second path
is longer than the one shown as a solid curve and, therefore, does not minimize the
path between the two points.
In many situations, it is very dicult or even impossible to nd an appropri-
ate coordinate transformation that allows us to convert a variation problem with
constraints into an unconstrained problem of lower dimensionality. In this case, we
need to follow the method of the unspecied Lagrange multipliers, which we will
now explore.
For brevity, we will consider the simple variational problem with two dependent
variables, y
1
(x) and y
2
. Our goal is to nd the path in the two-dimensional space
that extremizes the integral
J =
_
xB
xA
f [y
1
(x), y

1
(x), y
2
(x), y

2
(x); x] dx ,
subject to a holonomic constraint of the form
g [y
1
(x), y
2
(x); x] = 0 . (5.26)
As before, we will denote by y
0
1
(x) and y
0
2
(x) the two functions that extremize
the integral. We will then introduce variations away from these optimal functions
of the form
y
1
(x, ) = y
1
0
(x) + n
1
(x)
and
y
2
(x, ) = y
2
0
(x) + n
2
(x) ,
where is a parameter and n
1
(x) and n
2
(x) are two functions that satisfy the
boundary conditions
n
1
(x
A
) = n
1
(x
B
) = n
2
(x
A
) = n
2
(x
B
) = 0 .
Under these assumptions, the functional J becomes a simple function of the param-
eter and is extremized at = 0 if (see eq. [5.19])
dJ
d

= 0
_
xB
xA
2

i=1
_
f
y
i

d
dx
_
f
y

i
__
n
i
(x)dx = 0 , (5.27)
156 CHAPTER 5. CALCULUS OF VARIATIONS
In contrast to the previous derivations, when constraints are imposed, the func-
tions n
1
(x) and n
2
(x) are no longer independent. They should be chosen in such a
way that the holonomic constraints are satised for all values of the parameter ,
i.e.,
g[y
1
(x, ), y
2
(x, )] = g[y
0
1
(x), y
0
2
(x)] = 0 .
Performing a Taylor expansion of the left-hand side of the above relation away from
= 0 we obtain
g[y
1
(x, ), y
2
(x, )] = g[y
0
1
(x), y
0
2
(x)] +
_
g
y
1
y
1
a
+
g
y
2
y
2
a
_

g
y
1
y
1
a
+
g
y
2
y
2
a
= 0

g
y
1
n
1
(x) +
g
y
2
n
2
(x) = 0 .
As a result, if we choose an arbitrary function n
1
(x), then the above relation allows
us to pick the second function n
2
(x) such that the condition for the constraint is
always satised.
The function g[y
1
(x), y
2
(x)] cannot be independent of both y
1
(x) and y
2
(x),
since it is meant to represent a constraint on these two functions. Without loss of
generality, we can assume that the function y
2
(x) appears explicitly in the expression
for the constraint and, therefore, g/y
2
= 0 and
n
2
(x) =
_
g
y
2
_
1
_
g
y
1
_
n
1
(x) .
Inserting this relation into the integral (5.27) we obtain
_
xB
xA
_
f
y
1

d
dx
_
f
y

1
_

_
g
y
2
_
1
_
g
y
1
__
f
y
2

d
dx
_
f
y

2
__
_
n
1
(x)dx = 0 .
We have assumed that the function n
1
(x) is arbitrary, subject only to the condition
that it vanishes at x
A
and x
B
. As a result, the above integral vanishes only if the
term in the curly brackets is equal to zero, i.e.,
_
g
y
1
_
1
_
f
y
1

d
dx
_
f
y

1
__
=
_
g
y
2
_
1
_
f
y
2

d
dx
_
f
y

2
__
. (5.28)
Both sides of this equation are, therefore, equal to the same function of the inde-
pendent variable x, which we will denote by (x) and call the unspecied Lagrange
multiplier. In other words, the solutions to the constrained variational problem that
we are considering here satisfy the modied Euler-Lagrange equations
f
y
1

d
dx
_
f
y

1
_
+ (x)
g
y
1
= 0
f
y
2

d
dx
_
f
y

2
_
+ (x)
g
y
2
= 0
These are, of course, two dierential equations for three unknown functions, y
1
(x),
y
2
(x), and (x). The third (algebraic) equation is provided by the constraint (5.26).
It is straight forward to prove that, in the general case of N dependent vari-
ables, y
i
(x), i = 1, ..., N with K holonomic constraints of the form (5.21), we need
5.5. CONSTRAINTS AND LAGRANGE MULTIPLIERS 157
to introduce K Lagrange multipliers,
j
(x), j = 1, ..., K and the Euler-Lagrange
equations become
f
y
i

d
dx
_
f
y

i
_
+
K

j=1

j
(x)
g
j
y
i
= 0 i = 1, ..., N . (5.29)
Euler-Lagrange Equations
with Holonomic Constraints
Following a similar procedure, we can also show that for non-holonomic constraints
of the form (5.21) or even for integral constraints of the form (5.23) the method of
Lagrange multipliers leads to the set of modied Euler-Lagrange equations
f
y
i

d
dx
_
f
y

i
_
+
K

j=1

j
(x)
_
g
j
y
i

d
dx
_
g
j
y

i
__
= 0 i = 1, ..., N . (5.30) Euler-Lagrange Equations
with non-Holonomic or with
Integral Constraints
Note that equations (5.29) and (5.30) are equivalent with dening a new function
F = f +
K

j=1

j
(x)g
j
(5.31)
and seeking for the N functions y
i
(x) and the K multipliers
j
(x) for which the
integral
J =
_
xB
xA
Fdx
is extremized. The variational procedure with respect to the functions y
i
(x) lead to
equations (5.29) or (5.30), depending on whether the constraints are holonomic or
non-holonomic. On the other hand, the variational procedure with respect to the
multipliers
j
(x) leads to K dierential equations of the form
L

d
dx
_
L

j
_
= 0
g
j
[y
1
(x), y

1
(x), ..., y
N
(x), y

N
(x); x] = 0 , j = 1, ..., K , (5.32)
which are nothing but the equations of the constraints.
The Lagrange multipliers are, in general, functions of the independent vari-
able, which we have denoted here by x. However, as it can be easily seen from
equation (5.28), when the integrand f and the constraint g are independent of x
themselves, then each Lagrange multiplier is a constant.
The method of Lagrange multipliers is similar in some aspects to the substitu-
tion method we discussed earlier in this section and explored in Example (5.3): they
both allow us to convert a constrained variational problem to an unconstrained one.
There is, however, a big dierence between the two approaches. In the substitution
method, we use the equations of the constraints to reduce the dimensionality of
the problem. On the other hand, in the method of Lagrange multipliers, we in-
crease the dimensionality of the problem and are required, in general, to nd the
multipliers themselves as part of the solution. Although this appears to introduce
complications, it is often the only approach to solve a problem with non-holonomic
158 CHAPTER 5. CALCULUS OF VARIATIONS
or integral constraints. Moreover, when we will introduce, in the next chapter, La-
grangian dynamics in terms of a variational principles, the Lagrange multipliers of
constraint systems will be related directly to and allow us to calculate the forces of
the constraints on the particles.
In the following example, we explore in detail a variational problem with the
method of Lagrange multipliers.
5.6 Application: Suspension Bridges and Arches
In this application, we will explore the shape of simple suspension bridges, such as
the ones used by the Incas to traverse the deep canyons of the Andes. In order to
make the problem tractable, we will assume that the bridge is very thin, with a
constant surface density , and is made of an elastic but not extendible material.
We will denote by D the width of the bridge, by L its length, and by L
0
< L the
distance between the two sides of the canyon that the bridge spans. We will also
assume that the bridge has no twists and that the two rims of the canyon are at
the same altitude.
The suspension bridge on the
Apurimac River in Peru
We will set up a Cartesian coordinate system, with the origin at the center
of the canyon at the altitude of the two rims, and with the x
1
axis along the
vertical direction, the x
2
axis perpendicular to the canyon, and the x
3
axis along
the canyon. The bridge will settle to the shape that minimizes its potential energy
in the gravitational eld of the Earth. The potential energy of an extended object,
such as the bridge, in a uniform gravitational eld is proportional to the height of
its center of mass with respect to the ducial plane of zero potential. As a result,
the bridge will settle to the shape for which the center of mass is the lowest possible,
given the constraint imposed by the length of the bridge.
! " #$%&' ( ) * +, - . / 0123456789: ; <=>?@ABCDEFGHI J KLMNOPQRSTUVWXYZ[ \ ] ^ _` abc def ghi j k l mnopqr s t uv wx y z { | } ~
!
"#
$ &'
'
!
(
)
*$ &'
)
In order to calculate the position of the center of mass of the bridge, we rst
observe that the bridge will be mirror symmetric with respect to the x
1
axis. As a
result, the center of mass will lie along the x
1
axis and we can calculate its location
with respect to the rims of the canyon using equation (4.2), i.e.,
x
1,CM
=
D
_
L0/2
L0/2
x
1
dl
D
_
L0/2
L0/2
dl
=
1
L
_
L0/2
L0/2
x
1
dl .
In this equation,
dl =
_
1 +
_
dx
1
dx
2
_
2
_
1/2
is the dierential length along the long dimension of the bridge. Our goal is to nd
the shape of the bridge x
1
(x
2
) that minimizes the integral
x
1,CM
=
1
L
_
L0/2
L0/2
x
1
_
1 +
_
dx
1
dx
2
_
2
_
1/2
dx
2
subject to the constraint that the length of the bridge is xed, i.e.,
_
L0/2
L0/2
_
1 +
_
dx
1
dx
2
_
2
_
1/2
dx
2
= L
_
L0/2
L0/2
_
_
_
_
1 +
_
dx
1
dx
2
_
2
_
1/2

L
L
0
_
_
_
dx
2
= 0 .
5.6. APPLICATION: SUSPENSION BRIDGES AND ARCHES 159
Note that we are looking for a minimum of x
1,CM
since, along the length of the
bridge, x
1
< 0.
We will use the method of the Lagrange multipliers and convert this constrained
problem into an unconstrained one by searching for the shape which minimizes the
integral
J =
1
L
_
L0/2
L0/2
(f + g)dx
2
,
where
f x
1
_
1 +
_
dx
1
dx
2
_
2
_
1/2
and
g
_
1 +
_
dx
1
dx
2
_
2
_
1/2

L
L
0
.
Both the integrand f and the constraint g are independent of the variable x
2
and, therefore, the Lagrange multiplier is a constant. For the same reason, we
can use the Beltrami identity to write
(f + g) x

1
(f + g) = C
1
( + x
1
)
_
1 + x
2
1
_
_
1
x
2
1
1 + x
2
1
_

L
L
0
= C
1

+ x
1
(1 + x
2
1
)
1/2
= C

1
,
where we have dened
C

1
= C
1
+
L
L
0
.
Making a change of variables to y + x
1
, such that y

= x

1
, we can square both
sides of the above equation and rearrange its terms to obtain
y

=
_
y
2
C
2
1
1
_
.
This dierential equation should be familiar from our discussion in section 5.2. Its
solution is the catenary
y = + x
1
= C

1
cosh
_
x
2
+ C
2
C

1
_
.
We will now use the boundary conditions as well as the constraint of the original
variational problem in order to specify the values of the two constants of integration,
C

1
and C
2
, and of the Lagrange multiplier. By construction, we have set the origin of
the coordinate system in the geometric center of the suspension bridge and assumed
that the two rims of the canyon are at the same altitude. These two conditions imply
that the function form of the catenary has to be symmetric across the x
1
axis, i.e.,
C
2
= 0 and
x
1
= C

1
cosh
_
x
2
C

1
_
.
160 CHAPTER 5. CALCULUS OF VARIATIONS
The Catenary
The catenary is a function that appears in many distinct applications in science
and engineering, which is a fact that attracted the attention of many early
mathematicians. In this chapter, we showed that the catenary describes the
cross section of the surface of a thin soap lm subtended by two rings as well as
the shape of a simple suspension bridge.
In 1675, Robert Hooke announced in the Description of Helioscopes that he had
found the true mathematical and mechanical form of all manner of arches for
building, with the true butment necessary to each of them; a problem which no
architectonic writer hath ever yet attempted, much less performed. He then
proceeded to give the solution to the problem in the form of the anagram
abcccddeeeeeefggiiiiiiii-illmmmmnnnnnooprrsssttttttuuuuuuuux
that was later found to stand for Ut pendet continuum exile, sic stabit con-
tiguum rigidum inversum, i.e., As a exible cable hangs, so inverted stand the
touching pieces of an arch
[4]
. It is, indeed, rather straightforward to show that
the stresses in an arch that has the shape of an inverted catenary are tangen-
tial to its curve, greatly reducing the tendency for the arch to bulge. Several
historic and modern arches, such as the Gateway Arch in St. Louis, Missouri,
and those in Gauds Sagrada Famlia in Barcelona follow closely the shapes of
slightly modied catenaries.
A succession of inverted catenaries can also be used to construct a road on which
a bicycle with wheels in the shape of a polygon can roll without problems, keeping
the height of its axle constant. A tricycle with square wheels was constructed
by Stan Wagon, a mathematician at Macalester College, and is shown in the
picture below. The applicability of such a construction in any real-life situation
is highly doubtful because of the inherent diculty to steer it!
The Gateway Arch in St. Louis,
Missouri
A tricycle with square wheels
on a catenary road.
In the setup of the problem, we placed the zero point of altitude at the vertical
location of the two canyon rims, i.e.,
x
1
_

L
0
2
_
= x
1
_
L
0
2
_
= 0.
These boundary conditions lead to a joint constraint between the remaining param-
eters,
= C

1
cosh
_
L
0
2C

1
_
. (5.33)
Moreover, the length of the bridge is given by
L =
_
L0/2
L0/2
_
1 +
_
dx
1
dx
2
_
2
_
1/2
dx
2
= 2
_
L0/2
0
_
1 + sinh
2
_
x
2
C

1
_
2
_
1/2
dx
2
5.7. ADVANCED TOPIC: GENERALIZATIONS OF THE EULER-LAGRANGE EQUATIONS161
= 2C

1
sinh
_
L
0
2C

1
_
. (5.34)
At this point, we need to solve numerically equation (5.34) in order to obtain the
value of the parameter C

1
, which we can then insert into equation (5.33) to obtain
the value of the Lagrange multiplier . As long as L > L
0
, i.e., the bridge ts
between the canyon rim, equation (5.34) has two solutions for C

1
of equal magnitude
but of opposite sign. They correspond to two symmetric congurations, one in the
form of a hanging bridge and one in the form of an arch, as shown in the margin
gure. Clearly, these two solutions correspond to a minimum and a maximum of the
potential energy, respectively, and only one of them is the solution to our original
variational problem.
5.7 Advanced Topic: Generalizations of the Euler-
Lagrange Equations
Setting the foundations of Newtonian mechanics in terms of a variational principle,
as we will explore in the following chapters, requires the solution to problems that
involve at most one independent variable and functionals that incorporate at most
rst order derivatives. In several other situations in physics, however, we are often
faced with variational problems with many independent variables or with high-order
derivatives. We will explore one such problem at the end of the following chapter,
where we derive Poissons equation for the gravitational eld from a variational
principle. In this section, we will generalize the Euler-Lagrange equations to render
them applicable to such situations.
5.7.1 Variational Problems with High-order Derivatives
We will rst consider the problem of nding the path y(x) that extremizes the
integral
J =
_
xB
xA
f [y(x), y

(x), y

(x); x] dx .
This the situation in which the integrand f involves second-order derivatives of
the dependent variable y. For reasons that will become apparent very soon, we
will consider only paths for which the rst four derivatives with respect to the
independent variable are continuous along the domain of solution. Moreover, we
will consider paths for which not only the values of the functions at the initial and
nal points, y(x
A
) and y(x
B
), are specied, but also the rst derivatives of the
paths at those points, i.e., the values of y

(x
A
) and y

(x
B
).
As before, we will consider a specic yet arbitrary function n(x), for which both
the value and the rst derivative vanish at the boundaries x
A
and x
B
, i.e.,
n(x
A
) = n(x
B
) = n

(x
A
) = n

(x
B
) = 0 .
If we denote by y
0
(x) the optimal path, then we will consider variations away from
it of the form
y(x, ) = y
0
(x) + n(x) ,
162 CHAPTER 5. CALCULUS OF VARIATIONS
where is a real parameter. Under these assumptions, the integral J becomes a
simple function of the parameter , and
J

=0
= 0
_
xB
xA
_
f
y
y

+
f
y

+
f
y

_
dx = 0 .
The rst two terms in the last integral are identical to the two terms in equa-
tion (5.7) and we can, therefore, convert them to the form of the left-hand side of
equation (5.8). The last term in the integral is simply equal to
_
xB
xA
f
y

_
d
2
y
dx
2
_
dx =
_
xB
xA
f
y

d
2
dx
2
_
y

_
dx =
_
xB
xA
f
y

d
2
n
dx
2
dx .
and we can modify it using successive integrations by parts as
_
xB
xA
f
y

dx =
f
y

xB
xA

_
xB
xA
n

d
dx
_
f
y

_
dx
= 0 n
d
dx
_
f
y

xB
xA
+
_
xB
xA
n
d
2
dx
2
_
f
y

_
dx
= 0 +
_
xB
xA
n
d
2
dx
2
_
f
y

_
dx .
Putting all these terms together, we nd
J

=0
= 0
_
xB
xA
_
f
y

d
dx
_
f
y

_
+
d
2
dx
2
_
f
y

__
n(x)dx = 0 .
In order for the last integral to vanish for an arbitrary choice of the function n(x),
the term in the square brackets needs to vanish, i.e.,
f
y

d
dx
_
f
y

_
+
d
2
dx
2
_
f
y

_
= 0 .
This is the Euler-Lagrange equation when the integrand f depends explicitly on
derivatives of up to second order of the function y. Solving this dierential equa-
tion requires imposing four boundary conditions, which are the values and rst
derivatives of the function y at the two boundaries x
A
and x
B
.
It is straightforward now to generalize this result to the case where the in-
tegrand f depends on derivatives of the function y of up to order n, i.e., when
f = f(y, y

, ..., y
(n)
). Following the above procedure, we obtain
f
y

d
dx
_
f
y

_
+ ... + (1)
n
d
n
dx
n
_
f
y
(n)
_
= 0 . (5.35)
Euler-Lagrange equation
with high-order derivatives
Note that, in general, when the integrand f contains derivatives of the function y
of up to order n, the Euler-Lagrange equation is a dierential equation of order 2n.
This will become important in the following chapter, where we will derive Newtons
equations of motion from variational principles.
5.7.2 Variational Problems with Many Independent Vari-
ables
The second generalization we will consider, involves nding the Euler-Lagrange
equation for the path y(x
1
, x
2
, ..., x
n
) in an ndimensional space that extremizes
5.7. ADVANCED TOPIC: GENERALIZATIONS OF THE EULER-LAGRANGE EQUATIONS163
the functional
J =
_
...
_
f(y, y

; x
1
, x
2
, ..., x
n
)dx
1
dx
2
...dx
n
.
As before, we will assume that the function y is twice dierentiable and has xed
values along the boundaries of a closed domain. We will denote by y
0
(x
1
, x
2
, ..., x
n
)
the solution to this problem.
We will consider a specied yet arbitrary function n(x
1
, x
2
, ..., x
n
) that is also
twice dierentiable throughout the domain of solution but vanishes at its boundary
and write the variation
y(x
1
, x
2
, ..., x
n
, ) = y
0
(x
1
, x
2
, ..., x
n
) + n(x
1
, x
2
, ...x
n
) .
Under these assumptions, the functional J becomes a simple function of the param-
eter that has an extremum at = 0, i.e.,
J

=0
= 0
_
...
_ _
f
y
y

+
f
y
x1
y
x1

+ ... +
f
y
xn
y
xn

_
dx = 0 .
In this expression, we denoted by y
xn
the partial derivative of y with respect to x
n
.
Using the fact that
y
xn

_
y
x
n
_
=

x
n
_
y

_
=
n
x
n
n
xn
.
the condition for the function J() to have an extremum becomes
J

=0
= 0
_
...
_ _
f
y
n +
f
y
x1
n
x1
+ ... +
f
y
xn
n
xn
_
dx = 0 .
Integrating the terms in parenthesis by parts, as before, we nally obtain
_
...
_ _
f
y


x
1
_
f
y
x1
_
...

x
n
_
f
y
xn
__
n(x
1
, ..., x
n
)dx = 0 .
In order for this integral to vanish for any arbitrary function n(x
1
, ..., x
n
), the term
in square brackets needs to vanish. This requirement leads to the Euler-Lagrange
equation with many independent variables, i.e.,
f
y


x
1
_
f
y
x1
_
...

x
n
_
f
y
xn
_
= 0 . (5.36)
Euler-Lagrange equation
with many independent variables
Equation (5.36) is a second-order partial dierential equation with Dirichlet bound-
ary conditions provided by the values of the function y along the closed curve that
surrounds the domain of solution.
Further Reading
1. Mathematical Methods for Physicists, by G. B. Arfken, H. J. Weber, and F. Harris,
6th edition (Academic Press)
2. Calculus of variations with Applications to Physics and Engineering, by R. We-
instock, 1st edition (Dover)
3. The geometry of soap lms and soap bubbles by F. J. Almgren and J. E. Taylor,
Scientic American, 235, 82 (July 1976)
164 CHAPTER 5. CALCULUS OF VARIATIONS
4. Restless Genius: Robert Hooke and His Earthly Thoughts by E. T. Drake, 1996
(Oxford University Press)
Suggested Problems
1. Calculate the area of a soap lm suspended between two concentric rings of
unit radius, as a function of the separation of the rings. Compare your result
to the sum of the areas of the two rings and address what you nd in terms of
the assumptions inherent in the derivation of the Euler-Lagrange equations.
2. Properties of the Cycloid. Show that: (a) the cycloid is the curve traced out by
a point on the rim of a circle that is rolling on a horizontal plane; (b) the area
under one arc of the cycloid is three times the area of the generating circle. (c)
the length of one arc of the cycloid is equal to four times the circumference of
the generating circle.
3. A lifeguard is sitting on a beach (point A in the gure) and wants to save a
distressed swimmer in the ocean (point B) in the gure. The lifeguard can run
on the beach at a speed V
r
and swim at a speed V
s
. What is the optimal path
for the lifeguard to take in order to reach the swimmer at the shortest amount
of time?
4. Fermats Principle: The path that light takes between two points A and B is
the one for which the time of ight
t
AB
=
_
B
A
n
ds
c
is an extremum. In this expression, ds is the dierential path length, c is the
speed of light in vacuum, and n is the index of refraction. Consider an observer
standing on a road in a hot day. If the index of refraction increases linearly
with vertical distance from the road, i.e., if
n(z) = n
0
+ n

z ,
with n

<< n
0
, show that a light path that leaves the observers eyes pointing
initially downwards may bend As a result, the observer may see the blue sky
even if they are looking downwards.
5. Calculate the geodesic curve, i.e., the curve of least length, that connects two
points A and B on the surface of a cylinder of radius r
0
.
6. Isoperimetric Problem: Show that among all closed curves of constant length
on a plane, the circle encloses the maximum surface area.
7. Chaplygins Problem: Consider an airplane performing a search by owing
horizontally in a closed curve with a constant airspeed V
a
, in a region where
strong uniform winds blow at a speed V
w
. Show that the path that extremizes
the enclosed area of the search region for a given length of the closed curve is
an ellipse with its major axis perpendicular to the wind velocity and with an
eccentricity equal to
e =
V
w
V
a
.
8. Properties of the catenary: Show that for an arch in the shape of an inverted
catenary, the tension in equilibrium is tangent to the arc. What is the magni-
tude of the tension at the topmost point of the arch?