# CHAPTER 1

Introduction and Preliminaries
Introduction
The calculus of variations is almost as old as calculus itself. In Principia Mathematica
Isaac Newton posed the problem of ﬁnding the shape of a body moving along the x-axis
which minimized air resistance. This was the ﬁrst problem in the calculus of variations.
He solved his formulation of this problem in 1694. However, his formulation was not to
the liking of other scientists, and the basic problem remains of interest to this day. Other
problems soon followed, posed and solved by the Bernoulli brothers, Euler and others.
Here are some examples:
v Geodesics. Find the shortest path connecting two points. If the two points are in
the plane you know the answer. If the two points and the path connecting them
are on a sphere you might well know the answer. In both cases can you prove
that your answer is correct? What if the points and the curve lie on a different
surface, such as the torus?
v The brachistochrone. Aparticle slides under the force of gravity along a smooth
curve joining the points a and b in the plane. What is the curve that minimizes
the time of transit?
v The path of a ray of light. A light ray proceeds from point a to point b. Fermat
proposed that the path it takes minimizes the transit time. If the velocity of light
is constant in the region containing a and b, the path is the straight line. What is
the path if the velocity of light is not constant? For example, what if the point a
is in open air, but the point b is under water?
v The shape of a heavy chain. Suppose a heavy chain has its ends ﬁxed in space
and is attracted downward by the force of gravity. What is the shape of the chain?
v The shape of an elastic beam. Consider an elastic beam with its ends clamped
at two points. The beam is bent downward by the force of gravity. What is its
shape?
v Minimal surfaces. Given a closed curve in space what is the shape of the surface
with minimal area which has this curve as its boundary? You can think of the
curve as made of wire. The surface of minimal area is a soap ﬁlm achieved by
dipping the wire in a soap solution. This is an extremely hard problem. Consider
the special case when the boundary curve consists of two circles centered on the
.-axis lying in planes perpendicular to that axis. Then the surface is a surface of
revolution about the .-axis. What is the curve that generates that surface?
v The isoperimetric problem. Find the region in the plane with ﬁxed perimeter
1 having the largest area. Many of you will realize that it is a circle, but can you
prove it?
In this book we will solve most of the problems listed above, as well as many others.
One of our goals is to show how important the calculus of variations is to research that is
1
2 1. INTRODUCTION AND PRELIMINARIES
being pursued today. We will occasionally use ad hoc methods, but we will spend most of
our time developing the indirect methods. Hopefully at the end of the course we will ﬁnd
time to look at some direct methods as well.
Although we will spend most of the time solving problems with one spatial dimension,
we will also want to discuss the situation in several variables, such as the minimal surface
problem. This will allow us to consider Hamilton’s approach to mechanics.
Let’s start by examining some of the above problems in more detail.
1.1. Examples of variational problems
1.1.1. Geodesics
Let’s look at a couple of instances of the problem of ﬁnding the shortest path be tween
two points.
1.1.1.1. The shortest graph. The easiest problem of ﬁnding geodesics is ﬁnding shortest
graph connecting two points in the plane. Suppose that u is a continuously differentiable
function deﬁned on the interval Œa. b|. Let c = u(a) and J = u(b). Then the graph of u is
a curve ,
u
in R
2
connecting the points (a. c) and (b. J). Among all such functions which
has the shortest graph?
Consider a short segment of length ^s of the graph centered at the point (.. u(.).
Suppose it lies above a segment of length ^. in the .- axis. The corresponding segment in
. has length ^. ~ u
0
(.)^., where ~means approximately equal to. By the Pythagorean
theorem
^s
2
~ ^.
2
÷^.
2
~ (1 ÷u
0
(.)
2
) ^.
2
.
Thus, if we deﬁne the element of arclength by Js
2
= J.
2
÷J.
2
. then
Js =
p
J.
2
÷J.
2
=
p
1 ÷u
0
(.)
2
J.. (1.1.1)
and the length of the graph of u is given by
L(u) =
Z

u
Js =
Z
b
a
p
1 ÷u
0
(.)
2
J.. (1.1.2)
We want to ﬁnd u which minimizes L(u) subject to the endpoint constraints
u(a) = c and u(b) = J. (1.1.3)
1.1.1.2. Geodesics in R
2
. Not all curves in the plane are graphs. How would we pose the
problem if we wanted to consider all curves instead of just graphs? Let a = (a
1
. a
2
) and
b = (b
1
. b
2
) be the two points. If , is a piecewise continuously differentiable curve joining
a to b, according to Example A.8 in the Appendix, it has a parametrization x : Œ0. 1| ÷R
2
satisfying x(0) = a and x(1) = b. The length of , is given by
l(,) =
Z
1
0
[x
0
(t )[ Jt.
We will indicate the dependence on the parametrization x by setting
L(x) =
Z
1
0
[x
0
(t )[ Jt. (1.1.4)
Then
l(,) = L(x) =
Z
1
0
[x
0
(t )[ Jt.
1.1. EXAMPLES OF VARIATIONAL PROBLEMS 3
The geodesic joining the points a and b is the curve joining them that has the shortest
length. The geometric problem of ﬁnding the geodesic joining a and b can now be stated
analytically as follows:
Problem 1. Find a continuously differentiable curve parametrization x deﬁned on the
interval Œ0. 1| satisfying the end-point conditions
x(0) = a and x(1) = b. (1.1.5)
such that
L(x) _ L(y) (1.1.6)
for all continuously differentiable curve parametrizations y : Œ0. 1| ÷R
2
which satisfy the
end-point conditions in (1.1.5).
In other words, we want to ﬁnd a curve parametrization which minimizes L(.) as
deﬁned in equation (1.1.4), subject to the end-point conditions in (1.1.5). Notice Ldepends
on the curve parametrization x. It is a function of a function. We will call such an object a
functional. When a functional is deﬁned by an integral, as is L in (1.1.4), we will call it a
variational integral. The functional L deﬁned in (1.1.4) is called the length functional.
One difference between the problem in Section 1.1.1.1 and this one is that the func-
tional in this section depends on a curve parametrization, which is a vector valued function.
In Section 1.1.1.1 the functional depended on a real valued function.
1.1.1.3. Geodesics in R
N
. There is nothing in Section 1.1.1.2 that requires the dimension
to be 2. If we consider the problem of ﬁnding geodesics in R
N
, we end up with the exactly
the same formulation.
1.1.1.4. Geodesics in the sphere. Of course we all know that a solution to the geodesics
problem is a straight line, parametrized by x(t ) = a ÷t (b ÷a). We are interested in more
serious problems. An example is to look for geodesics in a sphere. The set
S
2
R
=
˚
x = (.. .. :) ÷ R
3
.
2
÷.
2
÷:
2
= 1
2
«
is called the sphere of radius 1 centered at the origin in R
3
. Suppose that our end points a
and b are in S
2
R
, and we want the curve that connects them to lie completely in the sphere
as well. Stated analytically the problem becomes:
Problem 2. Find a continuously differentiable curve parametrization x deﬁned on the
interval Œ0. 1| satisfying the end-point conditions
x(0) = a and x(1) = b. (1.1.7)
[x(t )[ =
p
.(t )
2
÷.(t )
2
÷:(t )
2
= 1
2
for 0 _ t _ 1. (1.1.8)
such that
L(x) _ L(y) (1.1.9)
for all continuously differentiable curve parametrizations y : Œ0. 1| ÷ R
N
which satisfy
the end-point conditions in (1.1.7) as well as the additional constraint in (1.1.8).
This is the same minimization problemwe sawin Problem1, except that our parametriza-
tions must satisfy the additional constraint in (1.1.8), which requires the curve to lie on the
sphere.
Perhaps you know that a geodesic on a sphere is a great circle, deﬁned to be the
intersection of a plane though the center of the sphere with the sphere itself. However,
proving this is no easy matter.
4 1. INTRODUCTION AND PRELIMINARIES
1.1.2. The brachistochrone
This is one of the oldest problems in the calculus of variations. It was posed by Galileo
in 1638, but could not be solved until later, after the invention of calculus and the calculus
of variations.
Suppose we have a continuously differentiable function u deﬁned on the interval Œa. b|.
Let c = u(a) and J = u(b), where we will assume that J _ c, so that the point (b. J)
is lower in the plane than the point (a. c). The graph of u is a curve ,
u
in R
2
connecting
the points (a. c) and (b. J). Suppose there is a wire connecting the two points having the
shape of ,
u
. Suppose, in addition, we have a mass m which is constrained to follow the
wire, and moves along it without friction under the force of gravity. If we release it at
the point (a. c), the mass will follow the wire and end up at (b. J). How long will it take
to complete the passage from (a. c) to (b. J)? What function u will result in the shortest
passage?
The mass has kinetic energy T =
1
2

2
, where · is its speed along the wire. If we
specify that the potential energy U is 0 at the starting point where . = c, then U =
mg(. ÷c). The total energy of the particle is
1 = T ÷U =
1
2

2
÷mg(. ÷c).
Since energy is conserved under the force of gravity, we know that 1 is a constant. We
will set the mass free at (a. c) with speed ·
0
, so T =
1
2

2
0
there and we know that U = 0
there. Hence
1
2

2
0
= 1 =
1
2

2
÷mg(. ÷ c).
Solving this equation for the velocity · we get
· =
q
·
2
0
÷2g(c ÷.). (1.1.10)
Suppose that we have a small segment ^s of the graph centered at the point (.. u(.).
If ^t is the time needed to traverse ^s, then ^s ~ ·^t . Hence Jt =
1
v
Js, and using
(1.1.10), we see that the total time needed to traverse ,
u
is
T (u) =
Z

u
Js
·
=
Z
b
a
p
1 ÷u
0
(.)
2
q
·
2
0
÷2g(c ÷ u(.))
J. =
1

2g
Z
b
a
s
1 ÷u
0
(.)
2
C ÷ u(.)
. J..
(1.1.11)
where C = c ÷·
2
0
¡2g.
The constant 1¡

2g in T (u) does not affect things, so our problem is to ﬁnd the
function u which minimizes
F (u) =
Z
b
a
s
1 ÷u
0
(.)
2
C ÷ u(.)
J.. with C = c ÷·
2
0
¡2g. (1.1.12)
subject to the end conditions
u(a) = c and u(b) = J.
1.1. EXAMPLES OF VARIATIONAL PROBLEMS 5
1.1.3. Minimal surface of revolution
Suppose that once more we have a continuously differentiable function u deﬁned on
the interval Œa. b|, and that the graph of u is a curve ,
u
in R
2
connecting the points (a. c)
and (b. J). if we rotate ,
u
about the .-axis, we get a surface of revolution, which we will
denote by S
u
. We learn in calculus that the area of S
u
is given by
·(S
u
) = A(u) =
Z

u
2¬. Js =
Z
b
a
2¬u(.)
p
1 ÷u
0
(.)
2
J.. (1.1.13)
We want to ﬁnd the function u for which S
u
has minimal area. The 2¬ in (1.1.13) is
not important, so we want to minimize
F (u) =
Z
b
a
u(.)
p
1 ÷u
0
(.)
2
J.. (1.1.14)
subject to the end conditions in (1.1.3).
1.1.4. The shape of a heavy chain
Suppose that once more we have a continuously differentiable function u deﬁned on
the interval Œa. b|, and that the graph of u is a curve ,
u
in R
2
connecting the points (a. c)
and (b. J). The graph is a possible shape of a heavy chain of length l supported at the two
points if
L(u) =
Z
b
a
p
1 ÷u
0
(.)
2
J. = l. (1.1.15)
If we assume that the chain is stable under the force of gravity, which function u satisfying
(1.1.15) gives us the actual shape of the chain?
Physics tells us unless the potential energy of the chain is at a minimum, the chain will
not be stable, so we want to ﬁnd the chain of length l which minimizes potential energy.
Let j denote the density of the chain, which we will assume is constant. A small segment
^s of the chain at the point (.. u(.)) will have mass approximately equal to gj^s, so the
potential energy of that segment will be approximately gj.^s. Hence the total potential
energy of the chain is
U(u) =
Z

u
gj. Js =
Z
b
a
gju(.)
p
1 ÷u
0
(.)
2
J.. (1.1.16)
Since g and j are constant, we again want to ﬁnd the function u which minimizes
F (u) as deﬁned in (1.1.14), but now in addition to the endpoint constraints in (1.1.3), u
must satisfy the additional length constraint in (1.1.15).
1.1.5. Summary
In previous sections we have seen several examples of problems in the calculus of
variations. Let’s try to generalize what it is that we want to minimize or maximize. The
examples are in equations (1.1.2), (1.1.11), (1.1.12), (1.1.14), and (1.1.16). These quanti-
ties have the form
F (u) =
Z
b
a
J(.. u(.). u
0
(.)) J.. (1.1.17)
where u is a function deﬁned on the interval Œa. b|, and J is a function of three variables,
which we will denote as ., u, and ]. Thus F is a function that depends on the function u.
We will call such a thing a functional. An integral of the form we seen in (1.1.17) is called
a variational integral.
6 1. INTRODUCTION AND PRELIMINARIES
The problemis to ﬁnd functions u which minimize or maximize the functional in (1.1.17)
subject to the boundary conditions
u(a) = c and u(b) = J. (1.1.18)
1.2. Methods in the Calculus of Variations
All of these problems involve ﬁnding extrema where the object being varied is a curve
or a surface and not a point in Euclidean space, as is the case in calculus. It is this difference
that deﬁnes the calculus of variations. It is clear that the ordinary tools of calculus do not
directly apply. Over time three methods of solution have been developed:
v Ad hoc methods. Methods depending on the nature of the speciﬁc problems.
v Direct methods. Start with a sequence of functions which come closer and closer
to the actual minimum. We have to ﬁnd out if the sequence converges or not. If
it does converge, does it converges to a minimum?
v Indirect methods. These methods extend the methods of the ordinary calculus.
As is the case in calculus, we will ﬁnd necessary conditions that are not sufﬁcient,
and sufﬁcient conditions that are not necessary.
Let’s look brieﬂy at each of these.
These methods involve using everything we knowabout the subject at hand, and avoid-
ing the special mathematics we will develop in this book. An example is the problem of
ﬁnding the shortest curve joining two points in R
2
or in R
N
. Of course we knowthe answer
is a straight line, but we want to prove it.
We will give the proof in R
2
. The same proof works in R
N
, but it is easier to under-
stand in R
2
. Let · = (a. c) and T = (b. J) be two points in R
2
. Let
^ = [T ÷ ·[ =
p
(b ÷a)
2
÷(J ÷ c)
2
denote the distance between them.
Let , be any curve that joins · to T. We know that it has a parametrization deﬁned
on the unit interval. Thus there is function u : Œ0. 1| ÷ R
2
which parametrizes ,. u is a
vector valued function and we will set u(t ) = (.(t ). .(t )), where . and . are continuously
differentiable functions on Œ0. 1|. The length of , is
l(,) =
Z
1
0
[u
0
(t )[ Jt =
Z
1
0
p
.
0
(t )
2
÷.
0
(t )
2
Jt. (1.2.1)
Now we use what we know about geometry.
We may assume that A = (0. 0): We achieve this by translating the whole prob-
lem by the vector T ÷·. Since a translation is a rigid motion, it does not change
the distance between points, and therefore does not change the formula for the
length of , in (1.2.1).
We may assume that B = (l. 0): The line segment ·T makes an angle of 0 with
the positive .-axis. If we rotate around the origin by the angle ÷0 we send T
to a point on the .-axis. Again a rotation is a rigid motion, so it does not affect
length. Therefore the image of T must be (^. 0). Again formula (1.2.1) is not
affected by rotations.
1.2. METHODS IN THE CALCULUS OF VARIATIONS 7
The curve , nowconnects (0. 0) to (l. 0) so the component .(t ) of the parametrization
u satisﬁes .(0) = 0, and .(1) = l. Thus we have the equality/inequality chain
1(,) =
Z
1
0
p
.
0
(t )
2
÷.
0
(t )
2
Jt
_
Z
1
0
p
.
0
(t )
2
Jt (1.2.2)
_
Z
1
0
.
0
(t ) Jt (1.2.3)
= .(1) ÷.(0)
= ^.
Thus 1(,) _ ^ = [T ÷·[ for every curve joining · to T. The length of the straight
line joining · to T is equal to ^ = [T ÷ ·[, so we can conclude that the straight line is a
geodesic.
However, this does not prove that every geodesic is a straight line. Suppose, therefore,
that , is a geodesic. Then 1(,) = ^ = [T ÷ ·[, so we must have equality at every step
of the equality/inequality chain. To have equality in (1.2.2) we must have .
0
(t ) = 0 for all
t . Thus .(t ) is a constant, and since .(0) = 0 we must have .(t ) = 0 for all t . To have
equality in (1.2.3), we must have .
0
(t ) _ 0 for all t . Thus the parametrization for , has
the form u(t ) = (.(t ). 0), where .
0
(t ) _ 0. This is a parametrization for the line segment
·T. Thus any geodesic is a straight line.
The only difference in the proof in R
3
or in general in R
N
is that there are more com-
ponents in ·, T, and the parametrization u(t ). However, we still have the two important
geometric simpliﬁcations, and we still have the same equality/inequality chain, so the proof
goes through in R
N
.
1.2.2. Direct methods
Consider the functional
F (u) =
Z
b
a
J(.. u(.). u
0
(.)) J.. (1.2.4)
The calculus of variations studies how to ﬁnd a function u which is a minimum of func-
tionals like this. The minimum of F (u) over all available functions u may be ÷o, but for
this section let’s suppose that
m = inf
u
F (u) > ÷o.
Then for each integer n there is a function u
n
such that
m _ F (u
n
) _ m÷
1
n
.
Thus we have a sequence {u
n
] of functions such that F (u
n
) ÷m as n ÷o. This is the
beginning of the direct methods in the calculus of variations.
There are many questions that arise:
v Does the sequence {u
n
] converge? If not, perhaps it has a subsequence that
converges.
v If the sequence {u
n
] (or a subsequence) converges to u, what can we say about
u? Assuming that each u
n
is continuous or differentiable, can we say the same
thing about u? Is F (u) deﬁned?
v Can we conclude that F (u) = m?
8 1. INTRODUCTION AND PRELIMINARIES
The answer to each of these questions is “not always”. We have to put conditions on
the Lagrangian J to get a “yes”. Finding the answers involves some mathematics at a very
high level. This approach has been very fruitful and continues to lead to important results.
However, since we are not assuming much in the way of mathematical background, we
will not say any more about the direct methods of the calculus of variations.
1.2.3. Indirect methods
Indirect methods do not require too much advanced mathematics. We start by assum-
ing that a function u is a minimizer for a functional like (1.2.4). We then use the methods
of calculus to derive properties of such a minimizer. These amount to necessary condi-
tions for a minimizer to exist. Often these conditions are enough to isolate one or a few
candidates for a minimizer. We are then left to show that a candidate is actually a mini-
mizer. This can frequently be done using ad hoc methods, but we will also derive general
conditions that will ensure that a function is a minimizer.
This is the subject matter of this book, and we will begin the study in the next chapter.