Vous êtes sur la page 1sur 10

Calculus of Variations

Com S 477/577
Nov 16, 2010
1 Introduction
A functional assigns a real number to each function (or curve) in some class. One might say that a
functional is a function of another function (or curve). Functionals have an important role in many
problems arising in analysis, optimization, mechanics, geometry, etc.
Example 1. Consider all possible paths joining two points p and q in the plane, and a particle moving
along any of these paths with velocity v(x, y) at the point (x.y). Then we obtain a functional that assigns
to each path the time of traversal by the particle.
Example 2. Let y(x) be an arbitrary continuously dierentiable function dened on the interval [a, b]. Then
the integral
J[y] =
_
b
a
y
2
(x) dx
is a functional on the set of all such functions y(x).
Example 3. Let F(x, y, z) be a continuous function of three variables. Then the integral
J[y] =
_
b
a
F(x, y(x), y

(x)) dx,
where y(x) ranges over the set of all continuously dierentiable functions dened over the interval [a, b], is a
functional. Dierent choices of the function F(x, y, z) give dierent functionals. For example, if
F(x, y, z) =
_
1 + z
2
,
then J[y] is the length of the curve y(x). If
F(x, y, z) = z
2
,
J[y] reduces to the case considered in the previous example.
Instances of problems involving functionals were considered more than three hundred years
ago, and the rst important results are due to Euler (1707-1783). Though still not having methods
comparable to those of classical analysis in generality, the calculus of functionals has had the
most development in nding the extrema of functionals, a branch referred to as the calculus of
variations. Listed below are a few typical variational problems.

The material is adapted from the book Calculus of Variations by I. M. Gelfand and S. V. Fomin, Prentice Hall
Inc., 1963; Dover, 2000.
1
1. Find the shortest plane curve connecting two points p and q, that is, nd the curve y = y(x)
for which the functional
_
b
a
_
1 + y
2
dx
achieves its minimum. The shortest curve turns out to be the line segment pq.
2. Let p and q be two points in the vertical plane
1
. Consider a particle sliding under gravity
along some path (curve) joining p and q. The time it takes the particle to reach q from p
depends on the curve, and hence is a functional. The curve that results in the least time
is called the brachistochrone. The brachistochrone problem was posed by John Bernoulli in
1696, and solved by himself, James Bernoulli, Newton, and LHospital. The brachistochrone
turns out to be a cycloid, as we will discuss later.
3. The isoperimetric problem was solved by Euler: Among all closed plane curves of a given
length, nd the one that encloses the greatest area. The answer turns out, not surprisingly,
to be a circle.
All of the above problems can be written in the form
_
b
a
F(x, y, y

) dx.
Such a function has a localization property that the value of the functional equals the sum of its
values over segments generated by dividing the curve y = y(x). Below is a functional example that
does not have this property:
_
b
a
x
_
1 + y
2
dx
_
b
a
_
1 + y
2
dx
.
It is important to see how problems in calculus of variations are related to those of classical
analysis, especially, to the study of functions of n variables. Consider a functional of the form
J[y] =
_
b
a
F(x, y, y

) dx, y(a) = y
a
, y(b) = y
b
.
To see the connection, we divide the interval [a, b] into n + 1 equal parts:
a = x
0
, x
1
, . . . , x
n
, x
n+1
= b.
Then, replace the curve y = y(x) by the polygonal line with vertices
(x
0
, y
a
), (x
1
, y(x
1
), . . . , (x
n
, y(x
n
)), (x
n+1
, y
b
).
Essentially, we are approximating the functional J[y] by the sum

J(y
1
, . . . , y
n
) =
n+1

i=1
F
_
x
i
, y
i
,
y
i
y
i1
h
_
h,
where y
i
= y(x
i
) and h =
ba
n+1
. Thus, we can regard the variational problem as that of nding the
extrema of the function J(y
1
, . . . , y
n
), with n . In this sense, functionals can be regarded as
functions of innitely many variables, and the calculus of variations as the corresponding analog
of dierential calculus.
1
The two points are assumed to have distinct x-coordinate values.
2
2 The Variation of a Functional
In a variational problem, we treat each function belonging to some class as a point in some space
referred as the function space. For instance, if we are dealing with a functional of the form
_
b
a
F(x, y, y

) dx,
the space includes all functions with a continuous rst derivative. In the case of a functional
_
b
a
F(x, y, y

, y

) dx,
the appropriate function space is the set of all functions whose second derivatives are continuous.
For a variational problem, we consider the linear space of feasible functions.
We need to measure the closeness between two elements in a function space. This is done by
introducing the concept of the norm of a function.
2
x
y
b a
( )
y
x
2
( ) x
y*
Figure 1: Distance between two functions is the
maximum absolute dierence of their values at any
point.
Let the space D, or more precisely D(a, b), con-
sist of all continuous functions dened on an inter-
val [a, b]. Addition and multiplication by real num-
bers are in the usual sense. The norm is dened
as
y
0
= max
axb
|y(x)|.
As illustrated in Figure 1, the distance between the
functions y

(x) and y(x) does not exceed if the


graph of the function y(x) lies inside a band of
width 2 centered at the graph of y

(x).
Similarly, the space D
n
, or more precisely
D
n
(a, b), consists of all functions dened on [a, b]
that are n times continuously dierentiable. The
norm now is dened as
y
n
=
n

i=0
max
axb

y
(i)
(x)

.
A functional J[y] is said to be continuous at a point (i.e., a function) y
0
in a normed function
space F if for any > 0, there exists a > 0 such that
|J[y] J[y
0
]| < ,
for all y y
0
< . A continuous functional J is said to be linear if
1. J[h] = J[h] for any function h F and R;
2. J[h
1
+ h
2
] = J[h
1
] +J[h
2
] for any h
1
, h
2
F;
2
We have indeed used the concept of the norm in dening the best approximation of a function using a family of
functions.
3
Example 4. The integral
J[h] =
_
b
a
h(x) dx
denes a linear functional on D(a, b).
Example 5. The integral
J[h] =
_
b
a
_

0
(x)h(x) +
1
(x)h

(x) + +
n
(x)h
(n)
(x)
_
dx,
where
i
(x) are functions in D(a, b), denes a linear functional on D
n
(a, b).
Let J[y] be a functional dened on some normed linear space, and let
J[h] = J[y + h] J[y]
be its change due to an increment h(x) of the variable y = y(x). When y(x) is xed, J[h] is a
functional of h. Suppose
J[h] = (h) + h,
where (h) is a linear functional and 0 as h 0. Then the functional J[y] is said to be
dierentiable. The linear functional [h] which diers from J[h] by a higher order innitesimal
term, is called the variation (or dierential) of J[y] and denoted by J[h]. This dierential is unique
when the functional is dierentiable.
Theorem 1 The dierentiable functional J[y] has an extremum at y = y

only if its variation


vanishes for y = y

, that is,
J[h] = 0
for all admissible functions h.
In the theorem, admissible functions refer to those that satisfy the constraints of a given varia-
tional problem. We refer to [1, p. 13] for a proof of Theorem 1.
The function J[y] has an extremum for y = y

if J[y] J[y

] does not change its sign in some


neighborhood of the curve y

(x).
3 Eulers Equation
Now, we consider what might be called the simplest variational problem. Let F(x, y, z) be
a function which is twice continuously dierentiable with respect to all arguments. Among all
functions y(x) that are continuously dierentiable over [a, b] and satisfy the boundary conditions
y(a) = y
a
and y(b) = y
b
, (1)
nd the one for which the functional
J[y] =
_
b
a
F(x, y, y

) dx (2)
has an extremum.
4
Suppose we increase y(x) by h(x). For y(x) + h(x) to continue to satisfy the boundary condi-
tions (1), we must have
h(a) = h(b) = 0. (3)
The corresponding increment in the functional is
J = J[y + h] J[y]
=
_
b
a
F(x, y + h, y

+ h

) dx
_
b
a
F(x, y, y

) dx
=
_
b
a
_
F(x, y + h, y

+ h

) F(x, y, y

)
_
dx
=
_
b
a
_
F
y
(x, y, y

)h + F
y
(x, y, y

)h

_
dx + ,
after applying Taylors expansion. The integral in the right hand side of the last equation above
represents the principal linear part of the increment J, hence the variation of J[y] is
J =
_
b
a
_
F
y
(x, y, y

)h + F
y
(x, y, y

)h

_
dx.
According to Theorem 1, J[y] has an extremum only if J = 0. Thus we have
0 =
_
b
a
(F
y
(x, y, y

)h + F
y
(x, y, y

)h

) dx
=
_
b
a
h
_
F
y

d
dx
F
y

_
dx + F
y
h|
b
a
=
_
b
a
h
_
F
y

d
dx
F
y

_
dx,
under the conditions (3). Since the function h is arbitrary except h(a) = h(b) = 0, it is not dicult
to show that the last equation above implies
F
y

d
dx
F
y
= 0. (4)
Equation (4) is called Eulers equation. Thus, the functional (2) has an extremum for a given
function y(x) only if y(x) satises Eulers equation. Since the equation is second order, its solu-
tion depends on two constants, which can be determined from the two boundary conditions (1).
Nevertheless, in some special cases, Eulers equation can be reduced to a rst-order dierential
equation.
Case 1 The integrand does not depend on y. So the functional has the form
_
b
a
F(x, y

) dx.
Eulers equation reduces to
d
dx
F
y
= 0,
5
which yields the rst-order equation:
F
y
= C, for some constant C.
Solving this equation for y

, we obtain an equation of the form


y

= f(x, C),
which is integrated for the curve:
y =
_
f(x, C) dx + D, for some constant D.
Case 2 The integrand does not depend on x. So the functional has the form
_
b
a
F(y, y

) dx.
Eulers equation becomes
F
y

d
dx
F
y
(y, y

) = 0.
Take the partial derivative:
F
y
F
y

y
y

F
y

y
y

= 0.
Multiplying the above equation by y

, we obtain
F
y
y

F
y

y
y
2
F
y

y
y

=
d
dx
(F y

F
y
) = 0.
Thus, we end up with a rst order dierential equation:
F y

F
y
= C, for some constant C.
Case 3 The integrand does not depend on y

. Eulers equation takes the form


F
y
(x, y) = 0,
which is not a dierential equation.
Case 4 In a variety of problems, the functional is the integral of a function f(x, y) with respect
to the arc length
_
1 + y
2
dx. Eulers equation can be transformed as follows:
0 = F
y

d
dx
F
y

= f
y
(x, y)
_
1 + y
2

d
dx
_
f(x, y)
y

_
1 + y
2
_
= f
y
_
1 + y
2
f
x
y

_
1 + y
2
f
y
y
2
_
1 + y
2
f
y

(1 +y
2
)
3/2
=
1
_
1 + y
2
_
f
y
f
x
y

f
y

1 + y
2
_
,
that is,
f
y
f
x
y

f
y

1 + y
2
= 0.
6
Example 6. Consider the functional
J[y] =
_
2
1
_
1 + y
2
x
dx, y(1) = 0, y(2) = 1.
The integrand does not contain y. This is Case 1, and Eulers equation has the form F
y
= C, for some C.
Thus,
y

x
_
1 + y
2
= C.
so that
y
2
(1 C
2
x
2
) = C
2
x
2
.
Since C can take on either a positive or a negative value, we obtain the derivative from the above:
y

=
Cx

1 C
2
x
2
.
Integrating the above gives us
y =
_
Cx

1 C
2
x
2
dx
=
1
C
_
1 C
2
x
2
+ D,
or, equivalently,
(y D)
2
+ x
2
=
1
C
2
.
Namely, the solution is a circle centered on the y-axis. From the boundary conditions, we nd that
C =
1

5
and D = 2.
Thus, the nal solution is
(y 2)
2
+ x
2
= 5.
Example 7. Among all the curves joining two points (x
0
, y
0
) and (x
1
, y
1
), nd the one which generates the
surface of minimum area when rotated about the x-axis. In this problem, the functional is the area of the
surface of revolution as a result of rotating the curve y = y(x) about the x-axis:
2
_
x1
x0
y
_
1 + y
2
dx.
We let F(y, y

) = y
_
1 + y
2
and minimize the integral. This is Case 2 where Eulers equation yields the
rst order equation:
F y

F
y
= C.
Thus, we have
y
_
1 + y
2
y
y
2
_
1 + y
2
= C,
or
y = C
_
1 + y
2
.
7
So,
y

=
_
y
2
C
2
C
2
,
which becomes
dx =
C
_
y
2
C
2
dy,
with solution
x + D = C ln
y +
_
y
2
C
2
C
,
or equivalently,
y = C cosh
x + D
C
.
The solution curve in Example 7 is a catenary, and the surface generated by rotation of the
catenary is called a catenoid. The constants C and D are determined using the conditions
y(x
0
) = y
0
and y(x
1
) = y
1
.
Figure 2 plots the solution curve y = 2 cosh(
x+3
2
) joining two points (1, 2 cosh(2)) and (2, cosh(
5
2
)).
0
Figure 2: A catenary y = 2 cosh(
x+3
2
) over [1, 2].
Depending on the locations (x
0
, y
0
) and (x
1
, y
1
), the solution surface may not exist. Consider
the case where the distance between the two points is suciently large compared to y
0
and y
1
.
The area generated by rotating the polygonal line p
1
p
2
p
3
p
4
, where p
1
= (x
0
, y
0
), p
2
= (x
0
, 0),
p
3
= (x
1
, 0), and p
4
= (x
1
, y
1
) will be less than the area generated by rotating a smooth curve
passing through the points.
Example 8. For the functional
J[y] =
_
b
a
(x y)
2
dx,
8
Eulers equation reduces to the equation
x y = 0.
So the solution is the line y = x, along which the integral vanishes.
4 The Variational Derivative
In this section we introduce the concept of the variational (or functional) derivative, which plays
the same role for functionals as derivatives for functions. Consider a functional
J[y] =
_
b
a
F(x, y, y

) dx, y(a) = y
a
, y(b) = y
b
.
The approach is to approximate the problem with an n-dimensional problem, and then take the
limit n .
Divide the interval [a, b] into n + 1 equal subintervals at the points
x
0
= a, x
1
, . . . , x
n
, x
n+1
= b,
with x
i+1
x
i
= x =
ba
n+1
. Now, we replace the function y(x) with a polyline with vertices:
(x
0
, y
0
), (x
1
, y
1
), . . . , (x
n+1
, y
n+1
),
where y
i
= y(x
i
). The original functional is approximated by the sum
J(y
1
, . . . , y
n
) =
n

i=0
F
_
x
i
, y
i
,
y
i+1
y
i
x
_
x, (5)
which is a function of n variables y
1
, . . . , y
n
.
Next, we calculate the partial derivatives and consider what happens as the number of subdi-
vision points goes to innity. Note that in the sum (5) each y
k
, 1 k n, appears in only two
terms. Thus, we have
J
y
k
= F
y
_
x
k
, y
k
,
y
k+1
y
k
x
_
x + F
y

_
x
k1
, y
k1
,
y
k
y
k1
x
_
F
y

_
x
k1
, y
k1
,
y
k+1
y
k
x
_
.
(6)
As x 0, i.e., as the number of subdivision points goes to innity, the right-hand side of the
above equation goes to zero as well (its last two terms together is a quantity of order x). To
obtain a limit, we divide (6) by x:
J
y
k
x
= F
y
_
x
k
, y
k
,
y
k+1
y
k
x
_

1
x
_
F
y

_
x
k1
, y
k1
,
y
k+1
y
k
x
_
F
y

_
x
k1
, y
k1
,
y
k
y
k1
x
__
.
(7)
The expression y
k
x appearing in the denominator on the left is the area of the region over
[x
k1
, x
k+1
] between the original polyline and the new polyline formed by moving the vertex (x
k
, y
k
)
to (x
k
, y
k
+ y
k
) while keeping all other vertices xed. See Figure 3. As x 0, the expression
converges to the limit
J
y
= F
y
(x, y, y

)
d
dx
F
y
(x, y, y

),
9
x x
0
x
k n+1
k

y
y
k
Figure 3: Variational derivative at y = y
k
is taken over the shaded region as its area goes to zero and the
number n + 1 of subdivision points goes to innity.
which is called the variational derivative of the functional J[y]. We see that
J
y
is just the left
side of Eulers equation (4). Thus, Eulers equation states that the variational derivative of the
function must vanish at an extremum. This is analogous to the result in analysis where a necessary
condition for a function with one variable to have an extremum is that its derivative vanishes.
In general, we dene the variational derivative of a functional J[y] as follows. Suppose we
increase y(x) by an amount h(x) which is non-zero only in the neighborhood of a point x
0
. Divide
the corresponding increment J[y + h] J[y] of the functional by the area between the curve
y = h(x) and the x-axis:
J[y + h] J[y]

. (8)
Let the area goes to zero in a way that both the maximum absolute value of h(x) and the
length of the interval surrounding x
0
in which h(x) does not vanish goes to zero. If the ratio (8)
converges to a limit, the limit is called the variational derivative of J[y] at the point x
0
, and is
denoted by
J
y

x=x
0
.
References
[1] I. M. Gelfand and S. V. Fomin. Calculus of Variations Prentice-Hall, Inc., 1963; Dover Publi-
cations, Inc., 2000.
[2] R. Weinstock. Calculus of Variations: With Applications to Physics and Engineering. Dover
Publications, Inc., 1974.
10

Vous aimerez peut-être aussi