CALCULUS OF VARIATIONS

MA 4311 LECTURE NOTES

I. B. Russak
Department of Mathematics
Naval Postgraduate School
Code MA/Ru
Monterey, California 93943
July 9, 2002

c 1996 - Professor I. B. Russak

1
Contents
1 Functions of n Variables 1
1.1 Unconstrained Minimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Constrained Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Examples, Notation 10
2.1 Notation & Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Shortest Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 First Results 21
3.1 Two Important Auxiliary Formulas: . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Two Important Auxiliary Formulas in the General Case . . . . . . . . . . . . 26

4 Variable End-Point Problems 36

4.1 The General Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

5 Higher Dimensional Problems and Another Proof of the Second Euler

Equation 46
5.1 Variational Problems with Constraints . . . . . . . . . . . . . . . . . . . . . 47
5.1.1 Isoparametric Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1.2 Point Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

6 Integrals Involving More Than One Independent Variable 59

7 Examples of Numerical Techniques 63

7.1 Indirect Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.1.1 Fixed End Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.1.2 Variable End Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.2 Direct Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

8 The Rayleigh-Ritz Method 82

8.1 Euler’s Method of Finite Differences . . . . . . . . . . . . . . . . . . . . . . . 84

9 Hamilton’s Principle 90

10 Degrees of Freedom - Generalized Coordinates 97

11 Integrals Involving Higher Derivatives 104

12 Piecewise Smooth Arcs and Additional Results 110

13 Field Theory Jacobi’s Neccesary Condition and Sufficiency 116

i
List of Figures
1 Neighborhood S of X0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Neighborhood S of X0 and a particular direction H . . . . . . . . . . . . . . 2
3 Two dimensional neighborhood of X0 showing tangent at that point . . . . . 5
4 The constraint φ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
5 The surface of revolution for the soap example . . . . . . . . . . . . . . . . . 11
6 Brachistochrone problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
7 An arc connecting X1 and X2 . . . . . . . . . . . . . . . . . . . . . . . . . . 15
8 Admissible function η vanishing at end points (bottom) and various admissible
functions (top) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
9 Families of arcs y0 + η . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
10 Line segment of variable length with endpoints on the curves C, D . . . . . . 22
11 Curves described by endpoints of the family y(x, b) . . . . . . . . . . . . . . 27
12 Cycloid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
13 A particle falling from point 1 to point 2 . . . . . . . . . . . . . . . . . . . . 29
14 Cycloid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
15 Curves C, D described by the endpoints of segment y34 . . . . . . . . . . . . 33
16 Shortest arc from a fixed point 1 to a curve N. G is the evolute . . . . . . . 36
17 Path of quickest descent, y12 , from point 1 to the curve N . . . . . . . . . . 40
18 Intersection of a plane with a sphere . . . . . . . . . . . . . . . . . . . . . . 56
19 Domain R with outward normal making an angle ν with x axis . . . . . . . . 61
20 Solution of example given by (14) . . . . . . . . . . . . . . . . . . . . . . . . 71
21 The exact solution (solid line) is compared with φ0 (dash dot), y1 (dot) and
y2 (dash) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
22 Piecewise linear function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23 The exact solution (solid line) is compared with y1 (dot), y2 (dash dot), y3
(dash) and y4 (dot) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
24 Paths made by the vectors R and R + δR . . . . . . . . . . . . . . . . . . . 90
25 Unit vectors er , eθ , and eλ . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
26 A simple pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
27 A compound pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
28 Two nearby points 3,4 on the minimizing arc . . . . . . . . . . . . . . . . . 112
29 Line segment of variable length with endpoints on the curves C, D . . . . . . 116
30 Shortest arc from a fixed point 1 to a curve N. G is the evolute . . . . . . . 118
31 Line segment of variable length with endpoints on the curves C, D . . . . . . 120
32 Conjugate point at the right end of an extremal arc . . . . . . . . . . . . . . 121
33 Line segment of variable length with endpoints on the curves C, D . . . . . . 123
34 The path of quickest descent from point 1 to a cuve N . . . . . . . . . . . . 127

ii
 Credits

Much of the material in these notes was taken from the following texts:

1. Bliss - Calculus of Variations, Carus monograph - Open Court Publishing Co. - 1924

2. Gelfand & Fomin - Calculus of Variations - Prentice Hall 1963

3. Forray - Variational Calculus - McGraw Hill 1968

4. Weinstock - Calculus of Variations - Dover 1974

5. J. D. Logan - Applied Mathematics, Second Edition -John Wiley 1997

The figures are plotted by Lt. Thomas A. Hamrick, USN and Lt. Gerald N. Miranda,
USN using Matlab. They also revamped the numerical examples chapter to include Matlab
software and problems for the reader.

iii
CHAPTER 1

1 Functions of n Variables
The first topic is that of finding maxima or minima (optimizing) functions of n variables.
Thus suppose that we have a function f (x1 , x2 · · · , xn ) = f (X) (where X denotes the n-
tuple (x1 , x2 , · · · , xn )) defined in some subset of n dimensional space Rn and that we wish
to optimize f , i.e. to find a point X0 such that
f (X0 ) ≤ f (X) or f (X0 ) ≥ f (X) (1)
The first inequality states a problem in minimizing f while the latter states a problem
in maximizing f .
Mathematically, there is little difference between the two problems, for maximizing f is
equivalent to minimizing the function G = −f . Because of this, we shall tend to discuss
only minimization problems, it being understood that corresponding results carry over to
the other type of problem.
We shall generally (unless otherwise stated) take f to have sufficient continuous differ-
entiability to justify our operations. The notation to discuss differentiability will be that f
is of class C i which means that f has continuous derivatives up through the ith order.

1.1 Unconstrained Minimum

As a first specific optimization problem suppose that we have a function f defined on some
open set in Rn . Then f is said to have an unconstrained relative minimum at X0 if
f (X0 ) ≤ f (X) (2)
for all points X in some neighborhood S of X0 . X0 is called a relative minimizing point.
We make some comments: Firstly the word relative used above means that X0 is a
minimizing point for f in comparison to nearby points, rather than also in comparison to
distant points. Our results will generally be of this “relative” nature.
Secondly, the word unconstrained means essentially that in doing the above discussed
comparison we can proceed in any direction from the minimizing point. Thus in Figure 1,
we may proceed in any direction from X0 to any point in some neighborhood S to make this
comparison.
In order for (2) to be true, then we must have that


n
fxi hi = 0 ⇒ fxi = 0 i = 1, · · · , n (3a)
i=1

and


n
fxi xj hi hj ≥ 0 (3b)
i,j=1

1
 S

Xo

Figure 1: Neighborhood S of X0
 
for all vectors H = (h1 , h2 , · · · , hn ) where fxi and fxi xj are respectively the first and
second order partials at X0 .

∂f ∂2f
fxi ≡ , fxi xj ≡ ,
∂xi ∂xi ∂xj
The implication in (3a), follows since the first part of (3a) holds for all vectors H.
Condition (3a) says that the first derivative in the direction specified by the vector H
must be zero and (3b) says that the second derivative in that direction must be non-negative,
these statements being true for all vectors H.
In order to prove these statements, consider a particular direction H and the points
X() = X0 + H for small numbers  (so that X() is in S). The picture is given in Figure 2.

X(ε )=Xo+ε H

Xo r

Figure 2: Neighborhood S of X0 and a particular direction H

2
 Define the function
g() = f (X0 + H) 0≤≤δ (4)
where δ is small enough so that X0 + H is in S.
Since X0 is a relative minimizing point, then

g() − g(0) = f (X0 + H) − f (X0 ) ≥ 0 0≤≤δ (5a)

Since −H is also a direction in which we may find points X to compare with, then we may
also define g for negative  and extend (5a) to read

g() − g(0) = f (X0 + H) − f (X0 ) ≥ 0 −δ ≤≤δ (5b)

Thus  = 0 is a relative minimizing point for g and we know (from results for a function
in one variable) that
dg(0) d2 g(0)
= 0 and ≥0 (6)
d d2
Now f is a function of the point X = (x1 , · · · , xn ) where the components of X() are
specified by
xi () = x0,i + hi −δ ≤≤δ i = 1, · · · , n (7)
so that differentiating by the chain rule yields

dg(0)
n
dxi
n (which ⇒fxi =0)
0= = fxi = fxi hi i = 1, · · · , n (8a)
d i=1 d i=1

and
d2 g(0)
n
dxi dxj
n
0≤ = fxi xj = fxi xj hi hj (8b)
d i,j=1 d d i,j=1

in which (8b) has used (8a). In (8) all derivatives of f are at X0 and the derivatives of x are
at  = 0.
This proves (3a) and (3b) which are known as the first and second order necessary
conditions for a relative minimum to exist at X0 . The term necessary means that they are
required in order that X0 be a relative minimizing point. The terms first and second order
refer to (3a) being a condition on the first derivative and (3b) being a condition on the
second derivative of f .
In this course we will be primarily concerned with necessary conditions for minimization,
however for completeness we state the following:
As a sufficient condition for X0 to be relative minimizing point one has that if


n

n
fxi hi = 0 and fxi xj hi hj ≥ 0 (9)
i=1 i,j=1

for all vectors H = (h1 , · · · , hn ), with all derivatives computed at X0 , then X0 is an uncon-
strained relative minimizing point for f .

3
Theorem 1 If f  (x) exists in a neighborhood of x0 and is continuous at x0 , then
1 
f (x0 + h) − f (x0 ) = f  (x0 )h + f (x0 )h2 + (h) ∀|h| < δ (10)
2
(h)
where lim = 0.
h→0 h2

Proof By Taylor’s formula

1 
f (x0 + h) − f (x0 ) = f  (x0 )h + f (x0 + Θh)h2
2
1  1
f (x0 + h) − f (x0 ) = f  (x0 )h +
f (x0 )h2 + [f  (x0 + Θh) − f  (x0 )] h2 (11)
2 2

The term in brackets tends to 0 as h → 0 since f is continuous. Hence

(h) 1
2
= [f  (x0 + Θh) − f  (x0 )] → 0 as h → 0. (12)
h 2
This proves (10).
Now suppose f C 2 [a, b] and f has a relative minimum at x = x0 . Then clearly

f (x0 + h) − f (x0 ) ≥ 0 (13)

and
f  (x0 ) = 0. (14)
Using (10) and (13) we have
1 
f (x0 + h) − f (x0 ) = f (x0 )h2 + (h) ≥ 0 (15)
2
(h)
with lim = 0. Now pick h0 so that |h0 | < δ, then
h→0 h2

1 
f (x0 + λh0 ) − f (x0 ) = f (x0 )λ2 h20 + (λh0 ) ≥ 0 ∀|λ| ≤ 1 (16)
2
Since  
1  1 (λh0 )
f (x0 )λ2 h20 + (λh0 ) = λ2 h20 f  (x0 ) + 2 2 2
2 2 λ h0
and since
(λh0 )
lim = 0
h→0 λ2 h2
0
we have by necessity
f  (x0 ) ≥ 0.

4
1.2 Constrained Minimization

As an introduction to constrained optimization problems consider the situation of seeking a

minimizing point for the function f (X) among points which satisfy a condition
φ(X) = 0 (17)
Such a problem is called a constrained optimization problem and the function φ is
called a constraint.
If X0 is a solution to this problem, then we say that X0 is a relative minimizing point for
f subject to the constraint φ = 0.
In this case, because of the constraint φ = 0 all directions are no longer available to
get comparison points. Our comparison points must satisfy (17). Thus if X() is a curve of
comparison points in a neighborhood S of X0 and if X() passes through X0 (say at  = 0),
then since X() must satisfy (17) we have
φ(X()) − φ(X(0)) = 0 (18)
so that also
d φ(X()) − φ(X(0))
n
dxi (0)
φ(0) = lim = φx i =0 (19)
d →0  i=1 d
In two dimensions (i.e. for N = 2) the picture is

Tangent at X0 −−−>
(has components
dx1(0)/d ε ,dx2(0)/d ε )

X0

<−−− Points X(ε )

(for which φ = 0)

Figure 3: Two dimensional neighborhood of X0 showing tangent at that point

dxi (0)
Thus these tangent vectors, i.e. vectors H which satisfy (19), become (with re-
d
placed by hi )


n
φxi hi = 0 (20)
i=1

5
and are the only possible directions in which we find comparison points.
Because of this, the condition here which corresponds to the first order condition (3a) in
the unconstrained problem is


n
fxi hi = 0 (21)
i=1

for all vectors H satisfying (19) instead of for all vectors H.

This condition is not in usable form, i.e. it does not lead to the implications in (3a) which
is really the condition used in solving unconstrained problems. In order to get a usable
condition for the constrained problem, we depart from the geometric approach (although
one could pursue it to get a condition).
As an example of a constrained optimization problem let us consider the problem of
finding the minimum distance from the origin to the surface x2 − z 2 = 1. This can be stated
as the problem of
minimize f = x2 + y 2 + z 2
subject to φ = x2 − z 2 − 1 = 0
and is the problem of finding the point(s) on the hyperbola x2 − z 2 = 1 closest to the origin.

z−axis
15

10

0
0
5
y−axis
10
−2
−1.5

15
−1
−0.5
0
0.5
1
1.5
2

x−axis

Figure 4: The constraint φ

A common technique to try is substitution i.e. using φ to solve for one variable in terms of
the other(s).

6
 Solving for z gives z 2 = x2 − 1 and then

f = 2x2 + y 2 − 1

and then solving this as the unconstrained problem

min f = 2x2 + y 2 − 1

gives the conditions

0 = fx = 4x and 0 = fy = 2y
which implies x = y = 0 at the minimizing point. But at this point z 2 = −1 which means
that there is no real solution point. But this is nonsense as the physical picture shows.
A surer way to solve constrained optimization problems comes from the following: For
the problem of
minimize f
subject to φ = 0
then if X0 is a relative minimum, then there is a constant λ such that with the function F
defined by
F = f + λφ (22)
then


n
Fxi hi = 0 for all vectors H (23)
i=1

This constitutes the first order condition for this problem and it is in usable form since it’s
true for all vectors H and so implies the equations

Fxi = 0 i = 1, · · · , n (24)

This is called the method of Lagrange Multiplers and with the n equations (24) together
with the constraint equation, provides n + 1 equations for the n + 1 unknowns x1 , · · · , xn , λ.
Solving the previous problem by this method, we form the function

F = x2 + y 2 + z 2 + λ(x2 − z 2 − 1) (25)

The system (24) together with the constraint give equations

0 = Fx = 2x + 2λx = 2x(1 + λ) (26a)

0 = Fy = 2y (26b)
0 = Fz = 2z − 2λz = 2z(1 − λ) (26c)
φ = x2 − z 2 − 1 = 0 (26d)

Now (26b) ⇒ y = 0 and (26a) ⇒ x = 0 or λ = −1. For the case x = 0 and y = 0 we have
from (26d) that z 2 = −1 which gives no real solution. Trying the other possibility, y = 0
and λ = −1 then (26c) gives z = 0 and then (26d) gives x2 = 1 or x = ±1. Thus the only
possible points are (±1, 0, 0, ).

7
 The method covers the case of more than one constraint, say k constraints.

φi = 0 i = 1, · · · , k < n (27)
and in this situation there are k constants (one for each constraint) and the function

k
F =f+ λi φ i (28)
i=1

satisfying (24). Thus here there are k+n unknowns λ1 , · · · , λk , x1 , · · · , xn and k+n equations
to determine them, namely the n equations (24) together with the k constraints (27).

Problems

1. Use the method of Lagrange Multipliers to solve the problem

minimize f = x2 + y 2 + z 2
subject to φ = xy + 1 − z = 0

2. Show that  
 λ  λ0
max  =
λ cosh λ cosh λ0
where λ0 is the positive root of

cosh λ − λ sinh λ = 0.

Sketch to show λ0 .

3. Of all rectangular parallelepipeds which have sides parallel to the coordinate planes, and
which are inscribed in the ellipsoid
x2 y2 z2
+ + = 1
a2 b2 c2
determine the dimensions of that one which has the largest volume.

4. Of all parabolas which pass through the points (0,0) and (1,1), determine that one
which, when rotated about the x-axis, generates a solid of revolution with least possible
volume between x = 0 and x = 1. [Notice that the equation may be taken in the form
y = x + cx(1 − x), when c is to be determined.

5. a. If x = (x1 , x2 , · · · , xn ) is a real vector, and A is a real symmetric matrix of order n,

show that the requirement that

F ≡ xT Ax − λxT x

be stationary, for a prescibed A, takes the form

Ax = λx.

8
Deduce that the requirement that the quadratic form

α ≡ xT Ax

be stationary, subject to the constraint

β ≡ xT x = constant,

leads to the requirement

Ax = λx,
where λ is a constant to be determined. [Notice that the same is true of the requirement
that β is stationary, subject to the constraint that α = constant, with a suitable definition
of λ.]
b. Show that, if we write
xT Ax α
λ = ≡ ,
xT x β
the requirement that λ be stationary leads again to the matrix equation

Ax = λx.

[Notice that the requirement dλ = 0 can be written as

βdα − αdβ
= 0
β2
or
dα − λdβ
= 0]
β
Deduce that stationary values of the ratio

xT Ax
xT x
are characteristic numbers of the symmetric matrix A.

9
CHAPTER 2

2 Examples, Notation
In the last chapter we were concerned with problems of optimization for functions of a finite
number of variables.
Thus we had to select values of n variables

x1 , · · · , xn

in order to solve for a minimum of the function

f (x1 , · · · , xn ) .

Now we can also consider problems of an infinite number of variables such as selecting
the value of y at each point x in some interval [a, b] of the x axis in order to minimize (or
maximize) the integral  x2
F (x, y, y )dx .
x1
 x2
Again as in the finite dimensional case, maximizing F dx is the same as minimizing
 x2 x1

−F dx so that we shall concentrate on minimization problems, it being understood that

x1
these include maximization problems.
Also as in the finite dimensional case we can speak of relative minima. An arc y0 is said
to provide a relative minimum for the above integral if it provides a minimum of the integral
over those arcs which (satisfy all conditions of the problem and) are in a neighborhood of
y0 . A neighborhood of y0 means a neighborhood of the points (x, y0 (x), y0 (x)) x1 ≤ x ≤ x2
so that an arc y is in this neighborhood if

max |y(x) − y0 (x)| < γ

x1 ≤x≤x2

and
max |y (x) − y0 (x)| < γ
x1 ≤x≤x2

for some γ > 0. ∗

Thus a relative minimum is in contrast to a global minimum where the integral is mini-
mized over all arcs (which satisfy the conditions of the problem). Our results will generally
be of this relative nature, of course any global minimizing arc is also a relative minimizing
arc so that the necessary conditions which we prove for the relative case will also hold for
the global case.
The simplest of all the problems of the calculus of variations is doubtless that of deter-
mining the shortest arc joining two given points. The co-ordinates of these points will be
∗
We shall later speak of a different type of relative minimum and a different type of neighborhood of y0 .

10
denoted by (x1 , y1 ) and (x2 , y2 ) and we may designate the points themselves when convenient
simply by the numerals 1 and 2. If the equation of an arc is taken in the form

y : y(x) (x1 ≤ x ≤ x2 ) (1)

then the conditions that it shall pass through the two given points are

y(x1 ) = y1 , y(x2 ) = y2 (2)

and we know from the calculus that the length of the arc is given by the integral
 x2 
I= 1 + y  2 dx ,
x1

where in the evaluation of the integral, y  is to be replaced by the derivative y  (x) of the
function y(x) defining the arc. There is an infinite number of curves y = y(x) joining the
points 1 and 2. The problem of finding the shortest one is equivalent analytically to that
of finding in the class of functions y(x) satisfying the conditions (2) one which makes the
integral I a minimum.

Y
1

0 X

Figure 5: The surface of revolution for the soap example

There is a second problem of the calculus of variations, of a geometrical-mechanical type,

which the principles of the calculus readily enable us to express also in analytic form. When
a wire circle is dipped in a soap solution and withdrawn, a circular disk of soap film bounded
by the circle is formed. If a second smaller circle is made to touch this disk and then moved
away the two circles will be joined by a surface of film which is a surface of revolution (in
the particular case when the circles are parallel and have their centers on the same axis
perpendicular to their planes.) The form of this surface is shown in Figure 5. It is provable
by the principles of mechanics, as one may surmise intuitively from the elastic properties of
a soap film, that the surface of revolution so formed must be one of minimum area, and the
problem of determining the shape of the film is equivalent therefore to that of determining

11
such a minimum surface of revolution passing through two circles whose relative positions
are supposed to be given as indicated in the figure.
In order to phrase this problem analytically let the common axis of the two circles be
taken as the x-axis, and let the points where the circles intersect an xy-plane through that
axis be 1 and 2. If the meridian curve of the surface in the xy-plane has an equation y = y(x)
then the calculus formula for the area of the surface is 2π times the value of the integral
 x2 
I= y 1 + y  2 dx .
x1

The problem of determining the form of the soap film surface between the two circles is
analytically that of finding in the class of arcs y = y(x) whose ends are at the points 1 and
2 one which minimizes the last-written integral I.
As a third example of problems of the calculus of variations consider the problem of the
brachistochrone (shortest time) i.e. of determining a path down which a particle will fall
from one given point to another in the shortest time. Let the y-axis for convenience be taken
vertically downward, as in Figure 6, the two fixed points being 1 and 2.

0 X
1

Figure 6: Brachistochrone problem

The initial velocity v1 at the point 1 is supposed to be given. Later we shall see that for
1
an arc defined by an equation of the form y = y(x) the time of descent from 1 to 2 is √
2g
times the value of the integral
 x2
1 + y 2
I= dx ,
x1 y−α
v12
where g is the gravitational constant and α has the constant value α = y1 − . The problem
2g
of the brachistochrone is then to find, among the arcs y : y(x) which pass through two points
1 and 2, one which minimizes the integral I.
As a last example, consider the boundary value problem

12
 −u (x) = r(x), 0<x<1
subject to
u(0) = 0, u(1) = 1.
The Rayleigh-Ritz method for this differential equation uses the solution of the following
minimization problem:
Find u that minimizes the integral
 1

1
I(u) = (u)2 − r(x)u dx
0 2
where uV = {vC 2 [0, 1], v(0) = 0, v(1) = 0} . The function r(x) can be viewed as force per
unit mass.

2.1 Notation & Conventions

The above problems are included in the general problem of minimizing an integral of the
form  x2
I= F (x, y, y ) dx (3)
x1
within the class of arcs which are continuously differentiable and also satisfy the end-
point conditions
y(x1 ) = y1 y(x2 ) = y2 (4)
√
where y1 , y2 are constants.
√ In the previous three problems F was respectively F = 1 + y  2 ,
 1+y  2
F = y 1 + y 2 , F = √ and y1 , y2 were the y coordinates associated with the points
y−α
1 and 2.
It should be noted that in (3) the symbols x, y, y  denote free variables and are not directly
related to arcs. For example, we can differentiate with respect to these variables to get in
the case of our last example
−1

Fx = 0 Fy = (y − α)−3/2 (1 + y 2 )1/2 , Fy
= y (y − α)−1/2 (1 + y  2 )−1/2 (5a)
2
It is when these functions are to be evaluated along an arc that we substitute y(x) for y and
y (x) for y  .
The above considered only the two dimensional case. In the n + 1 (n > 1) dimensional
case our arcs are represented by
y: yi (x) x1 ≤ x ≤ x2 i = 1, · · · , n (5b)
(the distinction between yi (x) and y1 , y2 of (4) should be clear from the context) and the
integral (3) is  x2
I= F (x, y1, · · · , yn , y1 , · · · , yn )dx (6)
x1

13
so that the integrals are functions of 2n + 1 variables and similar conventions to those for
the two dimensional case hold for the n + 1 dimensional case. Thus for example we will
be interested in minimizing an integral of the form (6) among the class of continuously
differentiable arcs (5b) which satisfy the end-point conditions
yi (x1 ) = yi,1 yi (x2 ) = yi,2 i = 1, · · · , n (7)
where yi,1 , yi,2 are constants. For now, continuously differentiable arcs for which (6) is
well-defined are called admissible arcs. Our problem in general will be to minimize the
integral (6) over some sub-class of admissible arcs. In the type of problems where the end-
points of the arcs are certain fixed values (as the problems thus far considered) the term fixed
end point problem applies. In problems where the end points can vary, the term variable
end point applies.

2.2 Shortest Distances

The shortest arc joining two points. Problems of determining shortest distances furnish a
useful introduction to the theory of the calculus of variations because the properties char-
acterizing their solutions are familiar ones which illustrate very well many of the general
principles common to all of the problems suggested above. If we can for the moment erad-
icate from our minds all that we know about straight lines and shortest distances we shall
have the pleasure of rediscovering well-known theorems by methods which will be helpful in
solving more complicated problems.
Let us begin with the simplest case of all, the problem of determining the shortest arc
joining two given points. The integral to be minimized, which we have already seen may be
written in the form  x2
I= F (y )dx (8)
x1
  2 1
if we use the notation F (y ) = (1 + y ) , and the arcs y : y(x) (x1 ≤ x ≤ x2 ) whose
2

lengths are to be compared with each other will always be understood to be continuous with
a tangent turning continuously, as indicated in Figure 7.
Analytically this means that on the interval x1 ≤ x ≤ x2 the function y(x) is continuous,
and has a continuous derivative. As stated before, we agree to call such functions admissible
functions and the arcs which they define, admissible arcs. Our problem is then to find
among all admissible arcs joining two given points 1 and 2 one which makes the integral I a
minimum.
A first necessary condition. Let it be granted that a particular admissible arc
y0 : y0 (x) (x1 ≤ x ≤ x2 )
furnishes the solution of our problem, and let us then seek to find the properties which
distinguish it from the other admissible arcs joining points 1 and 2. If we select arbitarily
an admissible function η(x) satisfying the conditions η(x1 ) = η(x2 ) = 0, the form
y0 (x) + η(x) (x1 ≤ x ≤ x2 ) , (9)

14
 f(X1) f(X2)

[ ]
X1 X2

Figure 7: An arc connecting X1 and X2

involving the arbitrary constant a, represents a one-parameter family of arcs (see Figure 8)
which includes the arc y0 for the special value  = 0, and all of the arcs of the family pass
through the end-points 1 and 2 of y0 (since η = 0 at endpoints).

y0

[ ]
x1 x2

η (x)

[ ]
x1 x2

Figure 8: Admissible function η vanishing at end points (bottom) and various admissible
functions (top)

The value of the integral I taken along an arc of the family depends upon the value of 
and may be represented by the symbol
 x2
I() = F (y0 + η  )dx . (10)
x1

Along the initial arc y0 the integral has the value I(0), and if this is to be a minimum when
compared with the values of the integral along all other admissible arcs joining 1 with 2 it

15
must, in particular, be a minimum when compared with the values I() along the arcs of the
family (9). Hence according to the criterion for a minimum of a function given previously
we must have I  (0) = 0.
It should perhaps be emphasized here that the method of the calculus of variations, as
it has been developed in the past, consists essentially of three parts; first, the deduction
of necessary conditions which characterize a minimizing arc; second, the proof that these
conditions, or others obtained from them by slight modifications, are sufficient to insure the
minimum sought; and third, the search for an arc which satisfies the sufficient conditions.
For the deduction of necessary conditions the value of the integral I along the minimizing arc
can be compared with its values along any special admissible arcs which may be convenient
for the purposes of the proof in question, for example along those of the family (9) described
above, but the sufficiency proofs must be made with respect to all admissible arcs joining
the points 1 and 2. The third part of the problem, the determination of an arc satisfying the
sufficient conditions, is frequently the most difficult of all, and is the part for which fewest
methods of a general character are known. For shortest-distance problems fortunately this
determination is usually easy.
By differentiating the expression (10) with respect to  and then setting  = 0 the value
of I  (0) is seen to be  x2
I  (0) = Fy
η  dx , (11)
x1

where for convenience we use the notation Fy
for the derivative of the integrand F (y ) with
respect to y . It will always be understood that the argument in F and its derivatives is the
function y0 (x) belonging to the arc y0 unless some other is expressly indicated.
We now generalize somewhat on what we have just done for the shortest distance problem.
Recall that in the finite dimensional optimization problem, a point X0 which is a relative
(unconstrained) minimizing point for the function f has the property that


n

n
fxi hi = 0 and fxi xj hi hj ≥ 0 (12)
i=1 i,j=1

for all vectors H = (h1 , · · · , hn ) (where all derivatives of f are at X0 ). These were called the
first and second order necessary conditions.
We now try to establish analogous conditions for the two dimensional fixed end-point
problem  x2
minimize I = F (x, y, y )dx (13)
x1

among arcs which are continuously differentiable

y: y(x) x1 ≤ x ≤ x2 (14)

and which satisfy the end-point conditions

y(x1 ) = y1 y(x2) = y2 (15)

with y1 , y2 constants.

16
 In the process of establishing the above analogy, we first establish the concepts of the first
and second derivatives of an integral (13) about a general admissible arc. These concepts
are analagous to the first and second derivatives of a function f (X) about a general point
X.
Let y0 : y0 (x), x1 ≤ x ≤ x2 be any continuously differentiable arc and let η(x) be
another such arc (nothing is required of the end-point values of y0 (x) or η(x)). Form the
family of arcs
y0 (x) + η(x) x1 ≤ x ≤ x2 (16)

y0

[ ]

x1 x2

Figure 9: Families of arcs y0 + η

Then for sufficiently small values of  say −δ ≤  ≤ δ with δ small, these arcs will all be in
a neighborhood of y0 and will be admissible arcs for the integral (13). Form the function
 x2
I() = F (x, y0 (x) + η(x), y0 (x) + η  (x))dx, −δ <  < δ (17)
x1

The derivative I  () of this function is

 x2
I  () = [Fy (x, y0 (x) + η(x), y0 (x) + η  (x))η(x) + (18)
x1
+Fy
(x, y0 (x) + η(x), y0 (x) + η  (x))η  (x)]dx

Setting  = 0 we obtain the first derivative of the integral I along y0

 x2

I (0) = [Fy (x, y0 (x), y0 (x))η(x) + Fy
(x, y0 (x), y0 (x))η  (x)]dx (19)
x1

Remark: The first derivative of an integral I about an admissible arc y0 is given by (19).
Thus the first derivative of an integral I about an admissible arc y0 is obtained by
evaluating I across a family of arcs containing y0 (see Figure 9) and differentiating that

17
function at y0 . Note how analagous this is to the first derivative of a function f at a point
X0 in the finite dimensional case. There one evaluates f across a family of points containing
the point X0 and differentiates the function.
We will often write (19) as
 x2

I (0) = [Fy η + Fy
η  ]dx (20)
x1

where it is understood that the arguments are along the arc y0 .

Returning now to the function I() we see that the second derivative of I() is
 x2

I () = [Fyy (x, y0 (x) + η(x), y0 (x) + η  (x))η 2 (x) + (21)
x1
+2Fyy
(x, y0 (x) + η(x), y0 (x) + η  (x))η(x)η  (x) +
+Fy
y
(x, y0 (x) + η(x), y0 (x) + η  (x))η  2 (x)]dx

Setting  = 0 we obtain the second derivative of I along y0 . The second derivative of I about
y0 corresponds to the second derivative of f about a point X0 in finite dimensional problems.
 x2
I  (0) = [Fyy (x, y0 (x), y0 (x))η 2 (x) + 2Fyy
(x, y0 (x), y0 (x))η(x)η  (x) + (22)
x1
Fy
y
(x, y0 (x), y0 (x))η  2 (x)]dx

or more concisely  x2
I  (0) = [Fyy η 2 + 2Fyy
ηη  + Fy
y
η  2 ]dx (23)
x1

where it is understood that all arguments are along the arc y0 .

As an illustration, consider the integral
 x2
I= y(1 + y  2 )1/2 dx (24)
x1

In this case we have

F = y(1 + y  2 )1/2 Fy = (1 + y  2 )1/2 Fy
= yy (1 + y  2 )− 2

1
(25)

so that the first derivative is

 x2
I  (0) = [(1 + y  2 )1/2 η + yy (1 + y  2 )−1/2 η  ]dx (26)
x1

Similarly
Fyy = 0 Fyy
= y (1 + y  2 )−1/2 Fy
y
= y(1 + y  2 )−3/2 (27)
and the second derivative is
 x2

I (0) = [2y  (1 + y  2 )−1/2 ηη  + y(1 + y  2 )−3/2 η  2 ]dx . (28)
x1

18
The functions η(x) appearing in the first and second derivatives of I along the arc y0 corre-
spond to the directions H in which the family of points X() was formed in chapter 1.
Suppose now that an admissible arc y0 gives a relative minimum to I in the class of
admissible arcs satisfying y(x1 ) = y1 , y(x2 ) = y2 where y1 , y2 , x1 , x2 are constants defined
in the problem. Denote this class of arcs by B. Then there is a neighborhood R0 of the
points (x, y0 (x), y0 (x)) on the arc y0 such that

Iy0 ≤ Iy (29)

(where Iy0 , Iy means I evaluated along y0 and I evaluated along y respectively) for all arcs
in B whose points lie in R0 . Next, select an arbitrary admissible arc η(x) having η(x1 ) = 0
and η(x2 ) = 0. For all real numbers  the arc y0 (x) + η(x) satisfies

y0 (x1 ) + η(x1 ) = y1 , y0 (x2 ) + η(x2 ) = y2 (30)

since the arc y0 satisfies (30) and η(x1 ) = 0, η(x2 ) = 0. Moreover, if  is restricted to a
sufficiently small interval −δ <  < δ, with δ small, then the arc y0 (x) + η(x) will be an
admissible arc whose points be in R0 . Hence

Iy0 +η ≥ Iy0 −δ << δ (31)

The function
I() = Iy0 +η
therefore has a relative minimum at  = 0. Therefore from what we know about functions
of one variable (i.e. I()), we must have that

I  (0) = 0 I  (0) ≥ 0 (32)

where I  (0) and I  (0) are respectively the first and second derivatives of I along y0 . Since
η(x) was an arbitrary arc satisfying η(x1 ) = 0 , η(x2 ) = 0, we have:

Theorem 2 If an admissible arc y0 gives a relative minimum to I in the class of admissible

arcs with the same endpoints as y0 then

I  (0) = 0 I  (0) ≥ 0 (33)

(where I  (0) , I  (0) are the first and second derivatives of I along y0 ) for all admissible arcs
η(x), with η(x1 ) = 0 and η(x2 ) = 0.
The above was done with all arcs y(x) having just one component, i.e. the n dimensional
case with n = 1. Those results extend to n(n > 1) dimensional arcs
y : yi(x) x1 ≤ x ≤ x2 i = 1, · · · n).
In this case using our notational conventions the formula for the first and second deriva-
tives of I take the form  nx2

I  (0) = [Fyi ηi + Fyi
ηi ]dx (34a)
x1 i=1

19
  x2

n
I  (0) = [Fyi yj ηi ηj + 2Fyi yj
ηi ηj + Fyi
yj
ηi ηj ]dx (34b)
x1 i,j=1

dη
where η  = .
dx
Problems

1. For the integral  x2

I = f (x, y, y ) dx
x1

with 
f = y 1/2 1 + y 2
write the first and second variations I  (0), and I  (0).

2. Consider the functional  1

J(y) = (1 + x)(y  )2 dx
0

where y is twice continuously differentiable and y(0) = 0 and y(1) = 1. Of all functions of
the form
y(x) = x + c1 x(1 − x) + c2 x2 (1 − x),
where c1 and c2 are constants, find the one that minimizes J.

20
CHAPTER 3

3 First Results
Fundamental Lemma. Let M(x) be a piecewise continuous function on the interval x1 ≤
x ≤ x2 . If the integral  x2
M(x)η  (x)dx
x1

vanishes for every function η(x) with η (x) having at least the same order of continuity as
does M(x) † and also satisfying η(x1 ) = η(x2 ) = 0, then M(x) is necessarily a constant.
To see that this is so we note first that the vanishing of the integral of the lemma implies
also the equation  x2
[M(x) − C]η  (x)dx = 0 (1)
x1
for every constant C, since all the functions η(x) to be considered have η(x1 ) = η(x2 ) = 0.
The particular function η(x) defined by the equation
 x
η(x) = M(x)dx − C(x − x1 ) (2)
x1
evidently has the value zero at x = x1 , and it will vanish again at x = x2 if, as we shall
suppose, C is the constant value satisfying the condition
 x2
0= M(x)dx − C(x2 − x1 ) .
x1

The function η(x) defined by (2) with this value of C inserted is now one of those which must
satisfy (1). Its derivative is η  (x) = M(x) − C except at points where M(x) is discontinuous,
since the derivative of an integral with respect to its upper limit is the value of the integrand
at that limit whenever the integrand is continuous at the limit. For the special function
η(x), therefore, (1) takes the form
 x2
[M(x) − C]2 dx = 0
x1

and our lemma is an immediate consequence since this equation can be true only if M(x) ≡ C.
With this result we return to the shortest distance problem introduced earlier. In (9)
of the last chapter, y = y0 (x) + η(x) of the family of curves passing through the points 1
and 2, the function η(x) was entirely arbitrary except for the restrictions that it should be
admissible and satisfy the relations η(x1 ) = η(x2 ) = 0, and we have seen that the expression
for (11) of that chapter for I  (0) must vanish for every such family. The lemma just proven
is therefore applicable and it tells us that along the minimizing arc y0 an equation
y
Fy
=  =C
1 + y
2
†
Thus if M (x) is continuous (piecewise continuous), then η  (x) should be continuous (at least piecewise
continuous)

21
must hold, where C is a constant. If we solve this equation for y  we see that y  is also a
constant along y0 and that the only possible minimizing arc is therefore a single straight-line
joining the point 1 with the point 2.
The property just deduced for the shortest arc has so far only been proven to be necessary
for a minimum. We have not yet demonstrated conclusively that the straight-line segment
y0 joining 1 and 2 is actually shorter than every other admissible arc joining these points.
This will be done later.

3.1 Two Important Auxiliary Formulas:

At this point we shall develop two special cases of more general formulas which are frequently
applied in succeeding pages. Let y34 be a straight-line segment of variable length which moves
so that its end-points describe simultaneously the two curves C and D shown in Figure 10,
and let the equations of these curves in parametric form be

(C) : x = x1 (t), y = y1 (t) ,

(D) : x = x2 (t), y = y2 (t) .

4
y34 6

y
3
y56 D
θ
(x,y)

Figure 10: Line segment of variable length with endpoints on the curves C, D

For example, the point 3 in Figure 10 is described by an (x, y) pair at time t1 as x3 =

x1 (t1 ), y3 = y1 (t1 ). The other points are similarly given, (x4 , y4 ) = (x2 (t1 ), y2(t1 )), (x5 , y5 ) =
(x1 (t2 ), y1 (t2 )), and (x6 , y6) = (x2 (t2 ), y2 (t2 )).
The length  
 = (x4 − x3 )2 + (y4 − y3 )2
of the segment y34 has the differential
(x4 − x3 )(dx4 − dx3 ) + (y4 − y3 )(dy4 − dy3 )
d =  .
(x4 − x3 )2 + (y4 − y3 )2

22
Note that since y34 is a straight line, then (y4 − y3 )/(x4 − x3 ) is the constant slope of the
line . This slope is denoted by p. This result may be expressed in the convenient formula of
the following theorem:
Theorem 3 If a straight-line segment y34 moves so that its end-points 3 and 4 describe
simultaneously two curves C and D, as shown in Figure 10, then the length  of y34 has the
differential 
dx + pdy 4
d(y34 ) = √ (3)
1 + p2  3
where the vertical bar indicates that the value of the preceding expression at the point 3 is
to be subtracted from its value at the point 4. In this formula the differentials dx, dy at the
points 3 and 4 are those belonging to C and D, while p is the constant slope of the segment
y34 .
We shall need frequently to integrate the right hand side of (3) along curves such as C and
D. This is evidently justifiable along C, for example, since the slope p = (y4 − y3 )/(x4 − x3 )
is a function of t and since the differentials dx, dy can be calculated in terms of t and dt from
the equations of C, so that the expression takes the form of a function of t. The integral I ∗
defined by the formula 
∗ dx + pdy
I = √
1 + p2
will also be well defined along an arbitrary curve C when p is a function of x and y (and
no longer a constant), provided that we agree to calculate the value of I ∗ by substituting
for x, y, dx, dy the expressions for these variables in terms of t and dt obtained from the
parametric equations of C.
It is important to note that I ∗ is parametrically defined, i.e. we integrate with respect
to t. Before we state the next theorem, let’s go back to Figure 10 to get the geometric
interpretation of the integrand in I ∗ .
The integrand of I ∗ has a geometric interpretation at the points of C along which it is
evaluated. At the point (x, y) on C, we can define two tangent vectors, one along the curve
C (see Figure 10) and one along the line y.
The tangent vector along C is given by
x i + √ y

v1 = √ j
x2 + y 2 x2 + y 2
and the tangent vector along y is
1 p
v2 = √ 2
i+ √ j.
1+p 1 + p2
The angle θ between these two vectors v1 and v2 is given by the dot product (since the
vectors are of unit length),
cos θ = v1 · v2
or
x + py 
cos θ =  . (4)
(1 + p2 )(x2 + y 2)

23
The element of arc length, ds, along C can be written as

ds = x2 + y 2 dt

From (4) it follows that the integral I ∗ can also be expressed in the convenient form
 
dx + pdy
I∗ = √ = cos θ ds . (5)
1 + p2
Let t3 and t5 be two parameter values which define points 3 and 5 on C, and which at
the same time define two corresponding points 4 and 6 on D, as in Figure 10. If we integrate
the formula (3) with respect to t from t3 to t5 and use the notation I ∗ just introduced, we
find as a further result:

Theorem 4 The difference of the lengths (y34 ) and (y56 ) of the moving segment in two
positions y56 and y34 is given by the formula

(y56 ) − (y34 ) = I ∗ (D46 ) − I ∗ (C35 ) . (6)

This and the formula (3) are the two important ones which we have been seeking. It
is evident that they will still hold in even simpler form when one of the curves C or D
degenerates into a point, since along such a degenerate curve the differentials dx and dy are
zero.

We now do a similar investigation of a necessary condition for the general problem defined
in (13) and (15) of the last chapter: Minimize an integral
 x2
I= F (x, y, y )dx (7)
x1

on the class β of admissible arcs joining two fixed points (x1 , y1 ) and (x2 , y2) in the xy plane
(i.e. in 2 dimensional space). Suppose we are given an arc y0 that gives a relative minimum
to I on the class β. Then by the previous chapter, the first derivative I  (0) of I about y0
has the property that  x2
I  (0) = [Fy η + Fy
η  ]dx = 0 (8)
x1

for all admissible arcs η with η(x1 ) = 0 and η(x2 ) = 0 where the arguments in the derivatives
of F are along y0 .
If we make use of the formula
d x 
 x
ηFy (x) = (η Fy ds) − η Fy ds (9)
dx x1 x1

and the fact that η(x1 ) = η(x2 ) = 0 then (8) becomes

 x2  x

I (0) = [Fy
− Fy ds]η  dx (10)
x1 x1

24
Then by use of the fundamental lemma we find that
 x
F (x) =
y
Fy ds + C x1 ≤ x ≤ x2 (11)
x1

holds at every point along y0 . Since we are only thus far considering arcs on which y (x) is
continuous, then we may differentiate (11) to obtain
d
Fy
(x) = Fy (x) x1 ≤ x ≤ x2 (12)
dx
along y0 (i.e. the arguments in Fy
and Fy are those of the arc y0 ).
This is the famous Euler equation.
There is a second less well-known Euler equation, namely:
d
(F − y  Fy
) = Fx (13)
dx
which is true along y0 .
For now, we prove this result only in the case that y0 is of class C 2 (i. e. has continuous
second derivative y0). It is however true when y0 is of class C 1 (i.e. has continuous tangent)
except at most at a finite number of points. Beginning with the left hand side of (13)
d d
[F − y Fy
] = Fx + Fy y  + Fy
y  − y  Fy
−y  Fy
(14)
dx    dx
=0

Thus, factoring y  from last terms, we have

 
d d
[F − y  Fy
] = Fx + y  Fy − Fy
(15)
dx dx
  
=0 by (12)

Thus we end up with the right hand of (13). This proves:

Theorem 5 The Euler equations (12) and (13) are satisfied by an admissible arc y0 which
provides a relative minimum to I in the class of admissible arcs joining its endpoints.

Definition: An admissible arc y0 of class C 2 that satisfies the Euler equations on all of [x1 , x2 ]
is called an extremal.
We note that the proof of (13) relied on the fact that (12) was true. Thus on arcs of class
2
C , then (13) is not an independent result from (12). However (13) is valid on much more
general arcs and on many of these constitutes an independent result from (12).

We call (12)-(13) the complete set of Euler equations.

Euler’s equations are in general second order differential equations (when the 2nd deriva-
tive y0 exists on the minimizing arc). There are however some special cases where these
equations can be reduced to first order equations or algebraic equations. For example:

25
Case 1 Suppose that the integrand F does not depend on y, i. e. the integral to be minimized
is  x2
F (x, y ) dx (16)
x1

where F does not contain y explicitly. In this case the first Euler’s equation (12) becomes
along an extremal
d
Fy
= 0 (17)
dx
or
Fy
= C (18)
where C is a constant. This is a first order differential equation which does not contain y.
This was the case in the shortest distance problem done before.

Case 2 If the integrand does not depend on the independent variable x, i. e. if we have to
minimize  x2
F (y, y ) dx (19)
x1

then the second Euler equation (13) becomes

d
(F − y Fy
) = 0 (20)
dx
or
F − y  Fy
= C (21)
(where C is a constant) a first order equation.

Case 3 If F does not depend on y , then the first Euler equation becomes

0 = Fy (x, y) (22)

which is not a differential equation, but rather an algebraic equation.

We next develop for our general problem the general version of the two auxiliary formulas
(3) and (4) which were developed for the shortest distance problem.

3.2 Two Important Auxiliary Formulas in the General Case

For the purpose of developing our new equations let us consider a one-parameter family of
extremal arcs
y : y(x, b) (x3 ≤ x ≤ x4 ) (23)
satisfying the Euler differential equation
∂
Fy
= Fy . (24)
∂x

26
The partial derivative symbol is now used because there are always the two variables x and
b in our equations. If x3 , x4 and b are all regarded as variables the value of the integral I
along an arc of the family is a function of the form
 x4
I(x3 , x4 , b) = F (x, y(x, b), y (x, b))dx .
x3

With the help of Euler’s equation (24), we see that along an extremal
 
∂F ∂y ∂y  ∂y ∂ ∂y  ∂ ∂y
= Fy + Fy
= Fy
+ Fy
= Fy
.
∂b  ∂b ∂b ∂b ∂x ∂b ∂x ∂b
use (24)
and the three partial derivatives of the function I(x3 , x4 , b) have therefore the values
 
∂I 3 ∂I 4
= −F  , = F  ,
∂x3 ∂x4
 x4   
∂I ∂ ∂y ∂y 4
= Fy
dx = Fy
 ,
∂b x3 ∂x ∂b ∂b 3
in which the arguments of F and its derivatives are understood to be the values y, y  belonging
to the family (23).
Suppose now that the variables x3 , x4 , b are functions x3 (t), x4 (t), b(t) of a variable t so
that the end-points 3 and 4 of the extremals of the family (23) describe simultaneously two
curves C and D in Figure 11 whose equations are
x = x1 (t) , y = y(x1 (t), b(t)) = y1 (t) , (25)
x = x2 (t) , y = y(x2 (t), b(t)) = y2 (t) .

4
y34 6

3
y56 D

Figure 11: Curves described by endpoints of the family y(x, b)

The differentials dx3 , dy3 and dx4 , dy4 along these curves are found by attaching suitable
subscripts 3 and 4 to dx, and dy in the equations
dx = x (t)dt , dy = yx dx + yb db . (26)

27
From the formulas for the derivatives of I we now find the differential
 
∂I ∂I ∂I 4 4
dI = dx3 + dx4 + db = (F dx + Fy
yb db)  = (F dx + Fy
(dy − pdx)) 
∂x3 ∂x4 ∂b 3 3

where the vertical bar indicates the difference between the values at the points 4 and 3. With
the help of the second of (26) this gives the following important result:

Theorem 6 The value of the integral I taken along a one-parameter family of extremal
arcs y34 (x, b) whose end-points describe the two curves C and D shown in Figure 11 has the
differential 
4
dI = [F (x, y, p)dx + (dy − p dx)Fy
(x, y, p)]  , (27)
3
where at the points 3 and 4 the differentials dx, dy are those belonging to C and D, while y
and p are the ordinate and slope of y34 (x, b).
We may denote by I ∗ the integral

I∗ = {F (x, y, p)dx + (dy − p dx)Fy
(x, y, p)} .

If we integrate the formula (27) between the two values of t defining the points 3 and 5 in
Figure 11 we find the following useful relation between values of this integral and the original
integral I.

COROLLARY: For two arcs y34 (x, b) and y56 (x, b) of the family of extremals shown in
Figure 11 the difference of the values of the integral I is given by the formula

I(y56 (x, b)) − I(y34 (x, b)) = I ∗ (D46 ) − I ∗ (C35 ) . (28)

Let us now use the results just obtained in order to attack the Brachistochrone problem
introduced in chapter 2. That problem is to find the path joining points 1 and 2 such that a
particle starting at point 1 with velocity v1 and acted upon only by gravity will reach point
2 in minimum time.
It is natural at first sight to suppose that a straight line is the path down which a particle
will fall in the shortest time from a given point 1 to a second given point 2, because a straight
line is the shortest distance between the two points, but a little contemplation soon convinces
one that this is not the case. John Bernoulli explicitly warned his readers against such a
supposition when he formally proposed the brachistochrone problem in 1696. The surmise,
suggested by Galileo’s remarks on the brachistochrone problem, that the curve of quickest
descent is an arc of a circle, is a more reasonable one, since there seems intuitively some
justification for thinking that steepness and high velocity at the beginning of a fall will
conduce to shortness in the time of descent over the whole path. It turns out, however, that
this characteristic can also be overdone; the precise degree of steepness required at the start
can in fact only be determined by a suitable mathematical investigation.
The first step which will be undertaken in the discussion of the problem in the following
pages is the proof that a brachistochrone curve joining two given points must be a cycloid.

28
A cycloid is the arched locus of a point on the rim of a wheel which rolls on a horizontal
line, as shown in Figure 12. It turns out that the brachistochrone must consist of a portion
of one of the arches turned upside down, and the one on the underside of which the circle
rolls must be located at just the proper height above the given initial point of fall.

Figure 12: Cycloid

When these facts have been established we are then faced with the problem of determining
whether or not such a cycloid exists joining two arbitrarily given points. Fortunately we will
be able to prove that two points can always be joined by one and only one cycloid of the
type desired.
The analytic formulation of the problem. In order to discuss intelligently the problem of
the brachistochrone we should first obtain the integral which represents the time required
by a particle to fall under the action of gravity down an arbitrarily chosen curve joining two
fixed points 1 and 2. Assume that the initial velocity v1 at the point 1 is given, and that
the particle is to fall without friction on the curve and without resistance in the surrounding
medium. If the effects of friction or a resisting medium are to be taken into account the
brachistochrone problem becomes a much more complicated one.

0
1 x
y=α

P
τ

2
mg
τ

Figure 13: A particle falling from point 1 to point 2

Let m be the mass of the moving particle P in Figure 13 and s the distance through
which it has fallen from the point 1 along the curve of descent C in the time t. In order to
make our analysis more convenient we may take the positive y-axis vertically downward, as
shown in the figure. The vertical force of gravity acting upon P is the product of the mass
m by the gravitational acceleration g, and the only force acting upon P in the direction of

29
the tangent line to the curve is the projection mg sin τ of this vertical gravitational force
d2 s
upon that line. But the force along the tangent may also be computed as the product m 2
dt
of the mass of the particle by its acceleration along the curve. Equating these two values we
find the equation
d2 s dy
2
= g sin τ = g
dt ds
in which a common factor m has been cancelled and use has been made of the formula
dy
sin τ = .
ds
ds
To integrate this equation we multiply each side by 2 . The antiderivatives of the two
dt
sides are then found, and since they can differ only by a constant we have
 2
ds
= 2gy + c . (29)
dt

The value of the constant c can be determined if we remember that the values of y and
ds
v= at the initial point 1 of the fall are y1 and v1 , respectively, so that for t = 0 the last
dt
equation gives
v12 = 2gy1 + c .
With the help of the value of c from this equation, and the notation

v12
α = y1 − , (30)
2g

equation (29) becomes

 2  
ds v2
= 2gy + v12 − 2gy1 = 2gy + 2g 1 − y1 = 2g(y − α) . (31)
dt 2g

An integration now gives the following result The time T required by a particle starting with
the initial velocity v1 to fall from a point 1 to a point 2 along a curve is given by the integrals
   x2
1 ds 1 1 + y
2
T =√ √ =√ dx (32)
2g 0 y−α 2g x1 y−α

v12
where  is the length of the curve and α = y1 − .
2g
An arc which minimizes one of the integrals (32) expressing T will also minimize that
1
integral when the factor √ is omitted, and vice versa. Let us therefore use the notations
2g
 x2
1 + y
2
I= F (y, y )dx , F (y, y ) = (33)
x1 y−α

30
for our integral which we seek to minimize and its integrand. Since the value of the function
F (y, y ) is infinite when y = α and imaginary when y < α we must confine our curves to
the portion of the plane which lies below the liney =α in figure 13. This is not really a
2
2 ds
restriction of the problem since the equation v = = 2g(y − α) deduced above shows
dt
that a particle started on a curve with the velocity v1 at the point 1 will always come to rest
if it reaches the altitude y = α on the curve, and it can never rise above that altitude. For
the present we shall restrict our curves to lie in the half-plane y > α.
In our study of the shortest distance problems the arcs to be considered were taken in
the form y : y(x) (x1 ≤ x ≤ x2 ) with y(x) and y (x) continuous on the interval x1 ≤ x ≤ x2 ,
An admissible arc for the brachistochrone problem will always be understood to have these
properties besides the additional one that it lies entirely in the half-plane y > α. The
integrand F (y, y ) and its partial derivatives are:


1+y
2
−1 
 1 + y
2 y
F = , Fy = , Fy
=  (34)
y−α 2 (y − α)3 (y − α)(1 + y
2 )

Since our integrand in (33) is independent of x we may use the case 2 special result (21)
of the Euler equations.
When the values of F and its derivative Fy
for the brachistochrone problem are substi-
tuted from (34) this equation becomes
1 1
F − y Fy
=  = √ , (35)
(y − α)(1 + y
2 ) 2b

1
the value of the constant being chosen for convenience in the form √ .
2b
The curves which satisfy the differential equation (35) may be found by introducing a
new variable u defined by the equation
u sin u
y  = − tan =− . (36)
2 1 + cos u
From the differential equation (35) it follows then, with the help of some trigonometry, that
along a minimizing arc y0 we must have
2b 2 u
y−α =
2 = 2b cos = b(1 + cos u)
1+y 2
Thus
dy
= −b sin u.
du
Now
dx dx dy 1 + cos u
= =− (−b sin u) = b(1 + cos u)
du dy du sin u

31
Integrating, we get x
x = a + b(u + sin u)
where a is the new constant of integration. It will soon be shown that curves which satisfy
the first and third of these equations are the cycloids described in the following theorem:
Theorem 7 A curve down which a particle, started with the initial velocity v1 at the point
1, will fall in the shortest time to a second point 2 is necessarily an arc having equations of
the form
x − a = b(u + sin u) , y − α = b(1 + cos u) . (37)
These represent the locus of a point fixed on the circumference of a circle of radius b as the
v2
circle rolls on the lower side of the line y = α = y1 − 1 . Such a curve is called a cycloid.
2g
Cycloids. The fact that (37) represent a cycloid of the kind described in the theorem is
proved as follows: Let a circle of radius b begin to roll on the line y = α at the point whose
co-ordinates are (a, α), as shown in Figure 14. After a turn through an angle of u radians
the point of tangency is at a distance bu from (a, α) and the point which was the lowest in
the circle has rotated to the point (x, y). The values of x and y may now be calculated in
terms of u from the figure, and they are found to be those given by (37).

x
b u

(a,α)

Figure 14: Cycloid

The fact that the curve of quickest descent must be a cycloid is the famous result discov-
ered by James and John Bernoulli in 1697 and announced at approximately the same time
by a number of other mathematicians.
We next continue using the general theory results to develop two auxiliary formulas for
the Brachistochrone problem which are the analogues of (3), (4) for the shortest distance
problem.
Two Important Auxiliary Formulas If a segment y34 of a cycloid varies so that its end-
points describe two curves C and D, as shown in Figure 15 then it is possible to find a
formula for the differential of the value of the integral I taken along the moving segment,
and a formula expressing the difference of the values of I at two positions of the segment.
The equations
x = a(t) + b(t)(u + sin u) , y = α + b(t)(1 + cos u)

32
 (u3 (t) ≤ u ≤ u4 (t)) (38)
define a one-parameter family of cycloid segments y34 when a, b, u3 , u4 are functions of a
parameter t as indicated in the equations. If t varies, the end-points 3 and 4 of this segment
describe the two curves C and D whose equations in parametric form with t as independent
variable are found by substituting u3 (t) and u4 (t), respectively, in (38). These curves and
two of the cycloid segments joining them are shown in Figure 15.

y
5

D
C y
4

Figure 15: Curves C, D described by the endpoints of segment y34

Now applying (27) of the general theory to this problem, regrouping (27), then the integral
in (33) has the differential
d = (F − pFy
)dx + Fy
dy (39)
where (recalling (27)) the differentials dx, dy in (39) are those of C and D while p is the slope
of y34 . Then by (35) and the last part of (34) substituted into (39) the following important
result is obtained.
Theorem 8 If a cycloid segment y34 varies so that its end-points 3 and 4 describe simul-
taneously two curves C and D, as shown in Figure 15, then the value of the integral I taken
along y34 has the differential
4
dx + pdy 
d = √ √  (40)
y − α 1 + p 32

At the points 3 and 4 the differentials dx, dy in this expression are those belonging to C and
D, while p is the slope of the segment y34 .
If the symbol I ∗ is now used to denote the integral

∗ dx + p dy
I = √ √ (41)
y − α 1 + p2

then by an integration of the formula (39) with respect to t from t3 to t5 we find the further
result that
Theorem 9 The difference between the values of  at two different positions y34 and y56
of the variable cycloid segment, shown in Figure 15, is given by the formula

33
 (y56 ) − (y34 ) = I ∗ (D46 ) − I ∗ (C35 ) . (42)
The formulas (40) and (42) are the analogues for cycloids of the formulas (3) and (4) for
the shortest distance problems. We shall see that they have many applications in the theory
of brachistochrone curves.

Problems

1. Find the extremals of  x2

I = F (x, y, y ) dx
x1

for each case

2
a. F = (y ) − k 2 y 2 (k constant)
2
b. F = (y ) + 2y
2
c. F = (y  ) + 4xy 
2
d. F = (y ) + yy  + y 2
2
e. F = x (y  ) − yy  + y
2
f. F = a(x) (y ) − b(x)y 2
2
g. F = (y ) + k 2 cos y
 b  
2
2. Solve the problem minimize I = (y  ) − y 2 dx
a
with
y(a) = ya , y(b) = yb .
What happens if b − a = nπ?

3. Show that if F = y 2 + 2xyy , then I has the same value for all curves joining the
endpoints.

4. A geodesic on a given surface is a curve, lying on that surface, along which distance
between two points is as small as possible. On a plane, a geodesic is a straight line. Determine
equations of geodesics on the following surfaces:

   2
 dz
2 2 2 2  2
a. Right circular cylinder. [Take ds = a dθ + dz and minimize a + dθ
dθ

   2

 2 dθ
or a + 1 dz]
dz
b. Right circular cone. [Use spherical coordinates with ds2 = dr 2 + r 2 sin2 αdθ2 .]
c. Sphere. [Use spherical coordinates with ds2 = a2 sin2 φdθ2 + a2 dφ2 .]
d. Surface of revolution. [Write x = r cos θ, y = r sin θ, z = f (r). Express the desired
relation between r and θ in terms of an integral.]

34
5. Determine the stationary function associated with the integral
 1
2
I = (y  ) f (x) ds
0

when y(0) = 0 and y(1) = 1, where

 1

 −1 0 ≤ x < 4


f (x) = 1


 1 4
<x≤1


6. Find the extremals

 1
a. J(y) = y dx, y(0) = 0, y(1) = 1.
0
 1
b. J(y) = yy dx, y(0) = 0, y(1) = 1.
0
 1
c. J(y) = xyy dx, y(0) = 0, y(1) = 1.
0

7. Find extremals for

 1 2
y
a. J(y) = dx,
0 x3
 1
b. J(y) = y 2 + (y  )2 + 2yex dx.
0

8. Obtain the necessary condition for a function y to be a local minimum of the functional
   b  b
J(y) = K(s, t)y(s)y(t)dsdt + y 2dt − 2 y(t)f (t)dt
R a a

where K(s, t) is a given continuous function of s and t on the square R, for which a ≤ s, t ≤ b,
K(s, t) is symmetric and f (t) is continuous.
Hint: the answer is a Fredholm integral equation.

9. Find the extremal for

 1
J(y) = (1 + x)(y  )2 dx, y(0) = 0, y(1) = 1.
0

What is the extremal if the boundary condition at x = 1 is changed to y  (1) = 0?

10. Find the extremals  b

J(y) = x2 (y )2 + y 2 dx.
a

35
CHAPTER 4

4 Variable End-Point Problems

We next consider problems in which one or both end-points are not fixed.
For illustration we again consider the shortest arc problem. However now we investigate
the shortest arc from a fixed point to a curve.
If a fixed point 1 and a fixed curve N are given instead of two fixed points then the
shortest arc joining them must again be a straight-line segment, but this property alone is
not sufficient to insure a minimum length. There are two further conditions on the shortest
line from a point to a curve for which we shall find very interesting analogues in connection
with the problems considered in later chapters.
Let the equations of the curve N in Figure 16 be written in terms of a parameter τ in
the form
x = x(τ ) , y = y(τ ) ,
Let y12 be the solution to the problem of finding the shortest arc from point 1 to curve N.

y12
−4

3 L 5 y56
1 6

Figure 16: Shortest arc from a fixed point 1 to a curve N. G is the evolute

Let τ2 be the parameter value defining the intersection point 2 of N. Clearly the arc y12
is a straight-line segment. The length of the straight-line segment joining the point 1 with
an arbitrary point (x(τ ) , y(τ )) of N is a function I(τ ) which must have a minimum at the
value τ2 defining the particular line y12 . The formula (3) of chapter 3 is applicable to the
one-parameter family of straight lines joining 1 with N when in that formula we replace C
by the point 1 and D by N. Since along C (now degenerated to a point) the differentials

36
dx, dy are then zero it follows that the differential of the function I(τ ) along the arc y12 is

dx + pdy 2
dI = √
1 + p2 
where the bar indicates that the value of the preceding expression is to be taken at the point
2. Since for a minimum the differential dI must vanish it follows that at the point 2 the
differentials dx, dy of N and the slope p of y12 satisfy the condition dx + pdy = 0, and hence
that these two curves must intersect at right angles (see (5) of chapter 3).
Even a straight-line segment through 1 and intersecting N at right angles may not be a
shortest arc joining 1 with N, as may be seen with the help of the familiar string property of
the evolute of N ‡ . The segments of the straight lines perpendicular to N cut off by N and its
evolute G in Figure 16 form a family to which the formula (6) of chapter 3 is applicable. If
in that formula we replace the curve C by G and D by N then (note that the points 2,3,5,6
are vertices of a quadrilateral similar to figure 11)

(y56 ) − (y32 ) = I ∗ (N26 ) − I ∗ (G35 ) .

But by using (5) of chapter 3 the integrals on the right hand side of this formula are seen to
have the values  s2
∗
I (N26 ) = cos θds = 0 , I ∗ (G35 ) = I(G35 )
s1

since cos θ = 0 along N (the straight lines of the family meet N at right angles), and cos θ = 1
along the envelope G (to which these lines are tangent). Hence from the next to last equation
we have the formula
(y32 ) = I(G35 ) + (y56 ) .
This is the string property of the evolute, for it implies that the lengths of the arcs y32 (x)
and G35 + y56 are the same, and hence that the free end 6 of the string fastened at 3 and
allowed to wrap itself around the evolute G will describe the curve N.
It is evident now that the segment y12 cannot be a shortest line from 1 to N if it has on
it a contact point 3 with the evolute G of N. For the composite arc y13 + G35 + y56 would in
that case have the same length as y12 and the arc y13 + L35 + y56 formed with the straight
line segment L35 , would be shorter than y12 . It follows then that:
If an arc y12 intersecting the curve N at the point 2 is to be the shortest joining 1 with
N it must be a straight line perpendicular to N at the point 2 and having on it no contact
point with the evolute G of N.
Our main purpose in this section was to obtain the straight line condition and also the
perpendicularity condition at N for the minimizing arc as we have done above. This last
result concerning the evolute G, is a hint of something that we shall see more of later on.

‡
The evolute of a curve is the locus of the centers of curvature of the given curve. The family of straight
lines normal to a given curve are tangent to the evolute of this curve, and the changes in length of the
radius of curvature is equal to the change in length of arc of the evolute as the point on the curve moves
continuously in one direction along the curve.

37
4.1 The General Problem
We now consider the general problem:
Minimize the integral  x2
I= F (x, y, y )dx (1)
x1

on the class of arcs joining fixed point 1 with coordinates (x, y) with the curve N. Note that
now point 2 with coordinates (x2 , y2 ) is not fixed since it is as yet an undetermined point on
N.

Necessary conditions when one end-point is variable. A minimizing arc y12 for this
problem, meeting the curve N at the point 2, must evidently be a minimizing arc for the
problem with end-points fixed at 1 and 2, and hence must satisfy at least the necessary
conditions (12), (13) of chapter 3.
For the problem with one variable end point there is a new necessary condition for a
minimum, involving the directions of the curves y12 and N at their intersection point 2,
which is called the transversality condition. This condition may be proved with the help of
the formula (27) of the last chapter. Let the points of N be joined to the point 1 of y12 by
a one-parameter family of arcs containing y12 as one member of the family. If the curve C
of the formula just cited is replaced by the fixed point 1, and the curve D by N, then this
formula shows that the value of I taken along the arcs of the one-parameter family has at
the particular arc y12 the differential
2

dI = [F (x, y, y )dx + (dy − y  dx)Fy
(x, y, y )]  ,

where at the point 2 the differentials dx, dy are those of N and the element (x, y, y ) belongs
to y12 . If the values of I along the arcs of the family are to have I(y12 ) as a minimum then
the differential dI must vanish along y12 and we have the following result:

THE TRANSVERSALITY CONDITION. If for an admissible arc y12 joining a fixed

point 1 to a fixed curve N the value I(y12 ) is a minimum with respect to the values of I on
neighboring admissible arcs joining 1 with N, then at the intersection point 2 of y12 and N
the direction dx : dy of N and the element (x, y, y ) of y12 must satisfy the relation

F (x, y, y )dx + (dy − y dx)Fy
(x, y, y ) = 0 . (2)

If this condition is satisfied the arc N is said to cut y12 transversally at the point 2. When
the arc N is the vertical line x = x1 or x = x2 , this condition is called a natural boundary
condition. For many problems the transversality condition implies that y12 and N must meet
at right angles. Indeed (2) when applied to the shortest distance problem gives the condition
of perpendicularity obtained there. However (2) does not in general imply perpendicularity
as one may verify in many special cases.
By a slight modification of the above reasoning we may treat the problem of minimizing
the integral (1) on the class of arcs joining two given curves C and D as in Figure 11. Let

38
y12 be a minimizing arc meeting curves C and D at points 1 and 2 respectively. Then y12
must also be a minimizing arc for the problem with fixed endpoints 1 and 2 and hence must
sastisfy the necessary conditions (12) and (13) of the last chapter. Furthermore, y12 is also a
minimizing arc for the problem of joining point 1 with the curve D so that the transversality
condition just deduced for the problem with one end-point varying must hold at point 2.
By a similar argument, with point 2 fixed for arcs joining point 2 with C, we see that the
transversality condition must also hold at point 1. Thus we have:

THE TRANSVERSALITY CONDITION (when both end-points vary). If for an admis-

sible arc y12 joining two fixed curves C and D, the value I(y12 ) is a minimum with respect
to the values of I on neighboring admissible arcs joining C and D, then at the intersection
points 1 and 2 of y12 with C and D respectively, the directions dx : dy of C and the element
(x, y, y ) of y12 at points 1 and 2 must satisfy the separate relations

  

[F (x, y, y )dx + (dy − y dx)Fy
(x, y, y )] =0 i = 1, 2 (3)
i

We now use the results just developed for the general theory by applying them to the
brachistochrone problem.
The path of quickest descent from a point to a curve. First necessary conditions. At
the conclusion of his now famous solution of the brachistochrone problem, published in
1697, James Bernoulli proposed to other mathematicians, but to his brother in particular,
a number of further questions. One of them was the problem of determining the arc down
which a particle, starting with a given initial velocity, will fall in the shortest time from a
fixed point to a fixed vertical straight line. This is a special case of the more genreral problem
of determining the brachistochrone arc joining a fixed point 1 to an arbitrarily chosen fixed
curve N.
Let the point 1, the curve N, and the path y12 of quickest descent be those shown in
Figure 17, (where α has the significance described in the previous chapter), and let the given
initial velocity at the point 1 again be v1 . Since by our general theory just developed, we
know that Euler’s equations (12) and (13) of the previous chapter apply, then by what has
been shown in chapter 3, the minimizing arc y12 must be a cycloid joining point 1 to some
as yet undetermined point 2 on the curve N. This constitutes a first necessary condition for
this problem.

Applying (2) to the present problem and using (33) of chapter 3 gives at point 2

1 + y
2 y
√ dx + (dy − y  dx)  √ =0 (4)
y−α 1 + y
2 y − α

where y , y are values on the minimizing arc y12 at point 2 and dy, dx are values of the curve
N at point 2. 
After multiplying and dividing by 1 + y
2 one obtains the condition

dx + y dy = 0 (5)

39
 y=α
1
N

y12 2

Figure 17: Path of quickest descent, y12 , from point 1 to the curve N

which is the transversality condition for this problem. This condition means that y12 must be
perpendicular to curve N at point 2. So the transversality condition here as in the shortest
distance problem, is one of perpendicularity, but as already noted, this is not true for all
problems.
Then for the brachistochrone problem from a point to a curve N, we have the result:
For a particle starting at point 1 with initial velocity v1 , the path of quickest descent from
1 to a curve N, is necessarily an arc y12 of a cycloid, generated by a point fixed on the
circumference of a circle, rolling on the lower side of the line y = y1 − v12 /2g. The path y12
must furthermore be cut at right angles by the curve N at their intersection point 2.

Example: Minimize the integral

 π/4  
I = y 2 − (y  )2 dx
0

with left end point fixed

y(0) = 1
and the right end point is along the curve
π
x = .
4

Since F = y 2 − (y )2 , then the Euler equation becomes

y  + y = 0.

The solution is
y(x) = A cos x + B sin x
Using the condition at x = 0, we get

y = cos x + B sin x

40
Now for the transversality condition

 

F + (φ − y )Fy
 = 0
x= π
4

where φ is the curve on the right end. Since the curve is a vertical line, the slope is infinite,
thus we have to rewrite the condition after dividing by φ . This will become (noting that
1/φ = 0) 

Fy
 = 0
x= π4

In our case
π
y( ) = 0
4
This implies B = 1, and thus the solution is

y = cos x + sin x.

4.2 Appendix

Let’s derive the transversality condition obtained before by a different method. Thus consider
the problem
 x2
min I = F (x, y, y )dx (1)
x1
among arcs
y: y(x) x1 ≤ x ≤ x2
(where x1 , x2 can vary with the arc) satisfying

y(xi ) = Yi (xi ) i = 1, 2 (2)

This is a variable end-point problem with Y1 (x) as the left boundary curve and Y2 (x) as the
right boundary curve.
Assume
y0 : y(x) x01 ≤ x ≤ x02
is a solution to this problem. Let η(x) be an arc and create the family of arcs

y() : y0 (x) + η(x) x1 () ≤ x ≤ x2 () −δ << δ (3)

for some δ > 0, where η(x), x1 (), x2 () are as yet arbitrary functions. In order that each arc
in this family satisfies (2) we must have

y0 (xi ()) + η(xi ()) = Yi (xi ()) i = 1, 2 (4)

41
Differentiating (4) with respect to  at  = 0 gives (recall that η  (xi ) term has a factor of )

dxi (0) dYi

y0 (xi (0)) + η(xi (0)) = (xi (0)) (5)
d d
dxi (0)
Equation (5) gives a relation between η, at the end-points of the solution arc and
d
dYi (0)
of the boundary curves. Namely, (with x0i = xi (0), the end-points of y0 )
d
dYi (0) dxi (0)
η(x0i ) = − y0 (x0i ) i = 1, 2 (6)
d d
These are the only η(x) arcs that can be used in this problem since they are the ones
which create families of arcs satisfying (2). We call these, admissible η(x). For such an
admissible η(x), evaluate I on the resultant family to get
 x2 ()
I() = F (x, y0 + η, y0 + η  )dx (7)
x1 ()

Differentiating with respect to  at  = 0 gives

 x2 (0) 
dxi (0) 2
I  (0) = [Fy η + Fy
η  ]dx + F (x0i ) =0 (8)
x1 (0) d 1

where F (x0i ) means F (xi (0), y0(xi (0)), y0 (xi (0))) i.e. F evaluated on the arc y0 at the ith
end-point and all terms in F or its derivatives are on y0 and the last term in the right side
means to difference the value at the left end point from its value at the right end-point and
where we have set I  (0) = 0 (why?). By doing the usual integration by parts we get
 x02  x  x02  x 
  d dxi (0) 2
0 = I (0) = [Fy
− Fy ds]η dx + [η Fy ds]dx + F (x0i ) (9)
x01 x01 x01 dx x01 d 1
Evaluating the second integral on the right side gives
 x02  x  x0i 
  dxi (0) 2
0 = I (0) = [Fy
− Fy ds]η dx + [η(x0i ) Fy ds + F (x0i ) ] (10)
x01 x01 x01 d 1
 x01
and noting that Fy ds = 0, gives
x01

 x02  x  x02 
  dxi (0) 2
0 = I (0) = [Fy
− Fy ds]η dx + η(x02 ) Fy ds + F (x0i ) (11)
x01 x01 x01 d 1
Now a particular class of admissible η(x) are those for which

dxi (0)
η(x02 ) = 0 =0 i = 1, 2 (12)
d

42
 For such η(x), all terms after the first integral on the right side in (11) are zero so that
for such η(x) we have   x02 x
0= [Fy
− Fy ds]η dx (13)
x01 x01

then by the fundamental lemma we have that

 x
Fy
(x) − Fy ds = c (14)
x01

holds along the solution arc y0 . This is the same as the Euler equation for the fixed end-point
problem. Furthermore by (14)
c = Fy
(x01 ) (15)
Now let η(x) be any arc satisfying (6), i.e. we are returning to the full class of admissible
η(x). Then by (14) and (15) we get that the first integral on the right side in (11) is
 x02  x  x02

[Fy
− Fy ds]η dx = cη  dx = c(η(x02 ) − η(x01 )) (16)
x01 x01 x01

= Fy
(x01 )[η(x02 ) − η(x01 )]

Then by (16), (15) and (14) the equation (11) becomes

 x02 
dxi (0) 2
0 = Fy
(x01 )[η(x02 ) − η(x01 )] + η(x02 ) · Fy ds +F (x0i ) (17)

x01
 
d 1
Fy
(x02 )−c by (14)
  
Fy
(x02 )−Fy
(x01 ) by (15)

Simplifying gives 
dxi (0) 2
[Fy
(x0i )η(x0i ) + F (x0i ) ] =0 (18)
d 1
Then by (6), for all admissible η(x), (18) becomes
 
dxi (0) dYi(0) 2
[F (x0i ) − y0 (x0i )Fy
(x0i )] + Fy
(x0i )  =0 (19)
d d 1

When (19) is multiplied by d, this is the transversality condition obtained previously.
Next, for future work, we’ll need an alternate form of the fundamental lemma which
we’ve been using.
Alternate fundamental lemma If α(x) is continuous on [x1 , x2 ] and if
 x2
α(x)η(x)dx = 0
x1

for every arc η(x) of class C 1 satisfying η(x1 ) = η(x2 ) = 0 then α(x) ≡ 0 for all x on [x1 , x2 ].

43
Problems
 x1  
2
1. Solve the problem minimize I = y 2 − (y ) dx
0

with left end point fixed and y(x1 ) is along the curve
π
x1 = .
4

2. Find the extremals for

 1

1 2
I = (y ) + yy  + y  + y dx
0 2
where end values of y are free.

3. Solve the Euler-Lagrange equation for

 b 
I = y 1 + (y  )2 dx
a

where
y(a) = A, y(b) = B.

b. Investigate the special case when

a = −b, A=B

and show that depending upon the relative size of b, B there may be none, one or two
candidate curves that satisfy the requisite endpoints conditions.

4. Solve the Euler-Lagrange equation associated with

 b  
2
I = y 2 − yy  + (y  ) dx
a

5. What is the relevant Euler-Lagrange equation associated with

 1  
2
I = y 2 + 2xy + (y ) dx
0

6. Investigate all possibilities with regard to tranversality for the problem

 b 
min 1 − (y )2 dx
a

7. Determine the stationary functions associated with the integral

44
  1  
2
I = (y ) − 2αyy  − 2βy  dx
0
where α and β are constants, in each of the following situations:
a. The end conditions y(0) = 0 and y(1) = 1 are preassigned.
b. Only the end conditions y(0) = 0 is preassigned.
c. Only the end conditions y(1) = 1 is preassigned.
d. No end conditions are preassigned.

8. Determine the natural boundary conditions associated with the determination of ex-
tremals in each of the cases considered in Problem 1 of Chapter 3.

9. Find the curves for which the functional

 x1 √
1 + y 2
I = dx
0 y

with y(0) = 0 can have extrema, if

a. The point (x1 , y1 ) can vary along the line y = x − 5.
b. The point (x1 , y1) can vary along the circle (x − 9)2 + y 2 = 9.

10. If F depends upon x2 , show that the transversality condition must be replaced by
   x2
∂F 
  ∂F
F + (φ − y )   + dx = 0.
∂y x=x2 x1 ∂x2

11. Find an extremal for

 e 
1 1
J(y) = x (y ) − y 2 dx,
2  2
y(1) = 1, y(e) is unspecified.
1 2 8

12. Find an extremal for

 1
J(y) = (y  )2 dx + y(1)2 , y(0) = 1, y(1) is unspecified.
0

45
CHAPTER 5

5 Higher Dimensional Problems and Another Proof of

the Second Euler Equation
Up to now our problems have been two-dimensional i.e. our arcs have been described by
two variables, namely x, y. A natural generalization is to consider problems involving arcs in
three-space with coordinates x, y, z or in even higher dimensional space, say N +1 dimensional
space with coordinates x, y1 , · · · , yN . The problem to be considered then involves an integral
of the form  x2
I= F (x, y1 , · · · , yN , y1 , · · · , yN

)dx . (1)
x1
and a class of admissible arcs y where superscript bar designates a vector arc, with compo-
nents
y : yi(x) x1 ≤ x ≤ x2 i = 1, · · · , N (2)
on which the integral (1) has a well defined value. As a fixed end-point version of this
problem one would want to minimize (1) on a subclass of arcs (2) that join two fixed points
y 1 and y 2 , i.e. arcs that satisfy
yi(x1 ) = yi1 yi(x2 ) = yi2 (3)
where y 1 has coordinates (y11 , · · · , yN 1 ) and y 2 has coordinates (y12 , · · · , yN 2).
Analogous to the proof used in obtaining the first Euler equation in chapter 3 for the
two-dimensional problem we could obtain the following condition along a minimizing arc
 x
Fyi
(x) = Fyi dx + ci i = 1, · · · , N (4)
x1

where ci are constants. And then by differentiation

d
Fy
(x) = Fyi (x) x1 ≤ x ≤ x2 (5)
dx i
Using this result we can now prove the second Euler equation for the two-dimensional
problem with the same generality as for the first Euler equation. We previously stated this
result in chapter 3 but only really proved it for arcs of class C 2 (i.e. having a continuous
second derivative y ). Thus let us now consider the two-dimensional problem of chapter 3,
defined by (7) and the remarks following it, without assuming even the existence of y  on
our arcs. We now write our arcs in parametric form. In particular our minimizing arc y0
(from chapter 3) is now written as
x = t y = y0 (t) x1 ≤ t ≤ x2 (6)
where t is the parameter and x(t) = t is the first component and y(t) = y0 (t) is the second
component. Being a minimizing arc, then this arc must minimize the integral (7) of chapter
3 on the class of parametric arcs
x = (t) y = ρ(t) t1 ≤ t ≤ t2 (7)

46
(where we have set t1 = x1 and t2 = x2 ), which join the fixed points (x1 , y1) and (x2 , y2 ) of
that problem and have  (t) > 0 on [t1 , t2 ]. This is true since each non-parametric arc of the
originally stated problem can be written (as in (6)) as a parametric vector arc of the class
just described and vice versa.
In terms of these parametric arcs, the integral (7) of chapter 3 takes the form
 t2
ρ 
I= F (, ρ, 
) dt (8)
t1    
F in (1)

where the primes now mean derivatives with respect to t § . This is an integral like (1) (i.e. in
three-dimensional t, x, y space). By (4) applied to the variable  (use i = 2 in (4)), there
results  t
ρ
F −  Fy
= Fx  dt + c (9)
 t1

 ρ
When we write y for  and use (6) we get along y0

 x
F − y Fy
= Fx dx + c (10)
x1

and by differentiation
d
[F − y  Fy
] = Fx (11)
dx
which is the result listed in chapter 3.

5.1 Variational Problems with Constraints

5.1.1 Isoparametric Problems

In the general problem considered thus far, the class of admissible arcs was specified (apart
from certain continuity conditions) by conditions imposed on the end-points. However many
applications of the calculus of variations lead to problems in which not only boundary con-
ditions, but also conditions of quite a different type, known as subsidiary conditions (or also
side conditions or constraints) are imposed on the admissible arcs. As an example of this, we
consider the isoparametric problem. This problem is one of finding an arc y passing through
the points (−x1 , 0) and (x1 , 0) of given length L which together with the interval [−x1 , x1 ] on
the x-axis, encloses the largest area. In general form, this problem can be stated as finding
the arc y for which the integral
 x2
I[y] = F (x, y, y )dx (12)
x1

§ ρ dy dy dx
The argument replaces y  in the original integral. This follows since by calculus = / .
 dx dt dt

47
is minimized (or maximized) where the admissible arcs satisfy the end-point conditions

y(xi) = yi i = 1, 2 (13)

and also are such that another integral

 x2
K[y] = G(x, y, y )dx (14)
x1

has a fixed value L.

In the specific application noted above (assuming that y is in the positive half-plane),
the integral I giving the area is  x1
I= y(x)dx (15)
−x1

while the integral K is the length integral

 x1 
K= 1 + y
2 dx (16)
−x1

and is required to have fixed length.

Returning now to the general version of this problem stated in (12)-(14), we will follow
the reasoning employed originally in solving the shortest distance problem. Thus assume
that y0 is a solution to this problem. Let η1 (x) and η2 (x) be two functions satisfying

η1 (xi ) = 0 , η2 (xi ) = 0 i = 1, 2 (17)

¶
Create the two-parameter family of arcs

y(α1 , α2 ) : y0 (x) + α1 η1 (x) + α2 η2 (x) (18)

By (17) this family satisfies the end-point conditions of our problem. Consider the integrals
(12) and (14) evaluated on this family. For example
 x2
I(y(α1 , α2 )) = F (x, y0 (x) + α1 η1 (x) + α2 η2 (x), y0 (x) + α1 η1 (x) + α2 η2 (x))dx (19)
x1

and similarly for K(y(α1 , α2 )). On this family of arcs our problem can be stated as

minimize I(y(α1 , α2 )) subject to K(y(α1 , α2 )) = L (20)

Now noting that on this family, these integrals can be considered as functions of two variables
(α1 , α2 ) (instead of arcs), then, when considering this family, our problem can be stated as

min I(α1 , α2 ) subject to K(α1 , α2 ) = L (21)

where in somewhat loose notation we have written I(α1 , α2 ) for I(y(α1 , α2 )) and similarly
for K. This is a finite (actually, two) dimensional problem of the type described in chapter 1.
¶
Note that up to now our families of arcs have been one parameter families.

48
By the results there and also noting that our minimizing arc y 0 = y(0, 0) solves this problem
we must have that
dI dK
0= (0, 0) + λ (0, 0) i = 1, 2 (22)
dαi dαi
where λ is a Lagrange multiplier.
Writing the integrals I and K out in terms of the family (18) and differentiating separately
with respect to α1 , α2 under the integral sign gives the two equations
 x2  x2
0= [Fy ηi + Fy
ηi ]dx + λ [Gy ηi + Gy
ηi ]dx i = 1, 2 (23)
x1 x1

where the partial derivatives of F and G are at (α1 , α2 ) = (0, 0) i.e. along the arc y0 .
Writing this as one integral, gives
 x2  x2
0= [(F + λG)y ηi + (F + λG)y
ηi ]dx = [F y ηi + F y
ηi ]dx i = 1, 2 (24)
x1 x1

where F ≡ F + λG and where this is true for all functions ηi (x) satisfying (17).
Making use of the integration by parts, formula (9) of chapter 3, but with F , we get as
there  x2  x
0= [F y
− F y ds]ηi dx (25)
x1 x1

Then by the fundamental lemma we obtain

 x
F y
(x) = F y ds + c x1 ≤ x ≤ x2 (26)
x1

(where c is a constant) which holds at every point along the arc y0 and then also by differ-
entiating
d
F y
(x) = F y (x) x1 ≤ x ≤ x2 (27)
dx
along y0 . In terms of the functions F and G, this is

d
(F + λG)y
= (F + λG)y x1 ≤ x ≤ x2 (28)
dx
This is the first Euler equation for the isoparametric problem.
In a manner similar to that used in the beginning of this chapter, it can be shown that
the second Euler equation
 x

(F + λG) − y (F + λG) = y
(F + λG)x dx x1 ≤ x ≤ x2 (29)
x1

or in differentiated form
d
[(F + λG) − y  (F + λG)y
] = (F + λG)x x1 ≤ x ≤ x2 (30)
dx
also holds along y0 .

49
 These results are summarized in
Theorem 10. For the problem stated in (12) - (14) let y0 be a solution, then there exists
a constant λ such that (26), (28), (29), and (30) are true along y0 .
Note that if our problem did not have a fixed right end point but instead was required
to intersect some curve N then η(x) would not have to satisfy (17) for i = 2 and then a line
of reasoning similar to that used in chapter 4, would give

(F + λG)dx + (dy − y dx)(F + λG)y
= 0 (31)

as the transversality condition at intersection with N, where the direction dy : dx comes

from N, and the arguments of F, G are from y0 . For this problem a corresponding condition
at left end point would hold if the left end point were not prescribed.
Let’s go through an application: Consider the problem of determining the curve of length
L with end-points (0,0) and (1,0) which encloses the largest area between it and the x-axis.
Thus we need to  1
maximize I = ydx (32)
0
subject to fixed end points
y(0) = y(1) = 0 (33)
and fixed length constraint  1
K= 1 + y
2 dx = L (34)
0
Setting 
F = y + λ 1 + y
2 (35)
the first Euler equation (27) is
 
d y
λ  −1 =0 (36)
dx 1 + y
2

Direct integration gives

λy 
 = x − c1 (37)
1 + y
2
Now make the substitution
tan θ = y  (38)
then (37) gives (recall that 1 + tan2 θ = sec2 θ)
x − c1
sin θ = (39)
λ
sin θ ± sin θ
Now since tan θ = =√ , then (38), (39) give
cos θ 1 − sin2 θ
±(x − c1 ) ±(x − c1 )
y =  = (40)
(x−c1 )2
λ 1 − λ2 λ2 − (x − c1 )2

50
 dy
or when using y  =
dx
±(x − c1 )
dy =  dx. (41)
λ2 − (x − c1 )2
Integration gives 
y = ∓ λ2 − (x − c1 )2 + c2 (42)
or then
(y − c2 )2 + (x − c1 )2 = λ2 (43)
This is part of a circle with center (c1 , c2 ), and radius λ.
The three constants c1 , c2 , λ are determined to satisfy the two end-point conditions and
the fixed length constraint this completes the problem solution. (see problem 5)

5.1.2 Point Constraints

We next discuss problems in which minimization is to be done subject to a different type of

constraint than the one just discussed. Thus consider the problem
 x2
minimize I = F (x, y, z, y , z  )dx (44)
x1

subject to fixed endpoints

y(xi) = yi z(xi ) = zi i = 1, 2 (45)

and also subject to a constraint

φ(x, y, z, y , z  ) = 0 (46)
Assume (as for previous problems) that y 0 is a solution to this problem. The notation y0
denotes a vector arc with components y0 (x), z0 (x). All arcs considered in this problem have
two components. Also assume that φy
, φz
do not equal zero simultaneously at any point on
y0 .
Next, let ψ(x), η(x) be functions which satisfy

ψ(xi ) = 0 η(xi ) = 0 i = 1, 2 . (47)

As in previous chapters, create the one-parameter family of arcs (but note that now our arcs
are vector arcs)

y() : y0 (x) + ψ(x) , z0 (x) + η(x) x1 ≤ x ≤ x2 (48)

We assume also that for some δ > 0, and for || < δ, the functions ψ(x), η(x) satisfy

φ(x, y0 (x) + ψ(x), z0 (x) + η(x), y0 (x) + ψ  (x), z0 (x) + η  (x)) = 0 x1 ≤ x ≤ x2 . (49)

51
 Again, similar to previous chapters, evaluate the integral in (44) on our family and define
 x2
I() = F (x, y0 (x) + ψ(x), z0 (x) + η(x), y0 (x) + ψ  (x), z0 (x) + η  (x))dx (50)
x1

Differentiating this with respect to  at  = 0 gives

 x2
0 = I  (0) = [Fy ψ + Fz η + Fy
ψ  + Fz
η  ]dx (51)
x1

where the partials of F are taken at points along y0 . Next, differentiate (49) with respect to
 at  = 0 to get
φy ψ + φz η + φy
ψ  + φz
η  = 0 x1 ≤ x ≤ x2 (52)
where the partials of φ are at points along y 0 . Equation (52) reveals that the ψ, η functions
are not independent of each other but are related. Multiplying (52) by an as yet unspecified
function λ(x) and adding the result to the integrand of (51) yields.
 x2
[(Fy + λφy )ψ + (Fy
+ λφy
)ψ  + (Fz + λφz )η + (Fz
+ λφz
)η  ]dx = 0 (53a)
x1

Setting F̂ = F + λφ gives (53a) in the form

 x2
[F̂y ψ + F̂y
ψ  + F̂z η + F̂z
η  ]dx = 0 (53b)
x1

Using the now familiar integration by parts formula on the 1st and 3rd terms in the integrand
of (53b) gives:
 x  x
d
ψ F̂y = (ψ F̂y dx) − ψ  F̂y dx (54)
dx x1 x1

and similarly for η F̂z . Using these and (47) yields

 x2  x  x
 
0 = I (0) = ([F̂y
− F̂y ds]ψ + [F̂z
− F̂z ds]η )dx (55)
x1 x1 x1

However we cannot take the step that we essentially did in developing the Euler equation
in the unconstrained case at the start of this chapter and say that the ψ, η functions, are
independent since as noted above (see (52)), they are not. Now, assuming that φy

=
0 (consistent with our assumption either φy
or φz

= 0) we can choose λ such that the
d
coefficient of ψ  is constant (i.e. choose λ such that (Fy
+ λφy
) − (Fy + λφy ) = 0 or then
dx
d d
λ̇ = (Fy + λφy − Fy
− λ φy
)/φy
and integrate this result). Next choose η arbitrarily
dx dx
(consistent with (47)) and ψ consistent with (49) and (47). By (47) and the fundamental
lemma, the coefficient of η  must also be constant. This results in
 x
F̂y
(x) − F̂y ds = c1 (56a)
x1

52
  x
F̂z
(x) − F̂z ds = c2 (56b)
x1

where c1 , c2 are constants. In differentiated form this is

d
F̂y − F̂y
= 0 (56c)
dx

d
F̂z − F̂z
= 0 (56d)
dx
Substituting for F̂ , then (56c), (56d) become

d
(Fy + λφy ) − (Fy
+ λφy
) = 0 (57a)
dx

d
(Fz + λφz ) − (Fz
+ λφz
) = 0 (57b)
dx
This result is actually contained in a larger result as follows. If the constraint (46) does not
depend on y  , z  i.e. if the constraint is

φ(x, y, z) = 0 (58)

and if φy and φz are not simultaneously zero at any point of y0 then the analogous equations
for (57a) and (57b) are
d
Fy + λφy − Fy
= 0 (59a)
dx
d
Fz + λφz − Fz
= 0 (59b)
dx
These results are summarized in the following:
Theorem: Given the problem
 x2
min I = F (x, y, z, y , z  )dx (60)
x1

subject to fixed end points and the constraint

φ(x, y, z, y , z  ) = 0 (61)

then if φy
, φz
(or in case φ does not depend on y  , z  , then if φy , φz ) do not simultaneously
equal zero at any point of a solution y0 , then there is a function λ(x) such that with F̂ ≡
F + λφ, then (56a) and (56b) or in differentiated form, (56c) and (56d) are satisfied along
y0 .
The three equations (56a,b) or (56c,d) and (61) are used to determine the three functions
y(x), z(x), λ(x) for the solution.

53
 In more general cases if our integrand has k dependent variables
 x2
I= F (x, y1 , y2 , · · · yk , y1 , · · · yk )dx (62)
x1

and we have (N < k) constraints

φi (x, y1 , · · · yk , y1 , · · · yk ) = 0 , i = 1, · · · , N (63)

∂φi ∂φi
such that the matrix 
(or in case the φ are independent of y1 · · · yk , then assume )
∂yj ∂yj
i = 1, · · · , N j = 1, · · · k has maximal rank along a solution curve, y0 then with

N
F̂ = F + λi (x)φi (64)
i=1

we have  
d
F̂yj − F̂yj
= 0 j = 1, · · · , k (65)
dx
holding on y0 where the λi (x) are N multiplier functions.
As an application, consider the problem of finding the curve of minimum length between
two points (x1 , y1 , z1 ) and (x2 , y2 , z2 ) on a surface

φ(x, y, z) = 0 (66)

Doing this in parametric form our curves will be written as x = x(t), y = y(t), z = z(t)
and with arc length as 
ds = ẋ2 + ẏ 2 + ż 2 dt (67)
where “·” denotes differentiation with respect to t. Then our problem is
 t2 
minimize I = ẋ2 + ẏ 2 + ż 2 dt (68)
t1

with fixed end points

x(ti ) = xi y(ti) = yi z(ti ) = zi i = 1, 2 (69)

subject to (66).
For this problem, with 
F = ẋ2 + ẏ 2 + ż 2 (70)
the Euler-Lagrange equations (65) are
     
d ẋ d ẏ d ż
λφx − = 0 λφy − = 0 λφz − =0 (71)
dt F dt F dt F
Now noting that
ds
F = (72)
dt

54
where s is arc length then e.g.
       
d ẋ d dx ds d dx dt d dx
= / = · = (73)
dt F dt dt dt dt dt ds dt ds
dt
and if we multiply this by we get
ds

 
 
d ẋ dt d dx dt d2 x
= = 2 (74a)
dt F ds dt ds ds ds
and similarly


d ẏ dt d2 y
= 2 (74b)
dt F ds ds


d ż dt d2 z
= 2 (74c)
dt F ds ds
dt
Thus, multiplying each of the equations of (71) by give as shown above
ds
d2 x dt d2 y dt d2 z dt
2
= λφx 2
= λφy 2
= λφz (75)
ds ds ds ds ds ds
or then
d2 x d2 y d2 z
: : = φx : φy : φz (76)
ds2 ds2 ds2
which has the geometric interpretation that the principal normal to the curve is parallel to
the gradient to the surface (i.e. it’s perpendicular to the surface).
If we do this in particular for geodesics on a sphere so that (66) is

φ(x, y, z) = x2 + y 2 + z 2 − R2 = 0 (77)

where R is the radius of sphere, then (71) becomes (after solving for λ)

F ẍ − ẋḞ F ÿ − ẏ Ḟ F z̈ − ż Ḟ
2
= 2
= (78)
2xF 2yF 2zF 2
2F 2
Multiplying by gives
F
ẍ − ẋ ḞF ÿ − ẏ ḞF z̈ − ż ḞF
= = (79)
x y z
which after cross multiplying gives

Ḟ Ḟ Ḟ Ḟ
y ẍ − y ẋ = xÿ − xẏ and yz̈ − y ż = z ÿ − z ẏ (80)
F F F F
and then
Ḟ Ḟ
y ẍ − xÿ = (y ẋ − xẏ) and yz̈ − z ÿ = (y ż − z ẏ) (81)
F F

55
or
y ẍ − xÿ yz̈ − z ÿ Ḟ
= = (82)
y ẋ − xẏ y ż − z ẏ F
The first equality can be restated as
d
dt
(y ẋ− xẏ) d
dt
(y ż− z ẏ)
= (83)
y ẋ − xẏ y ż − z ẏ

du
and integration using = ln |u| + c gives
u
y ẋ − xẏ = A(y ż − z ẏ) (84)

where A is a constant of integration. This gives

y(ẋ − Aż) = ẏ(x − Az) (85)

or then
ẋ − Aż ẏ
= (86)
x − Az y
so that another integration gives
x − Az = By (87)
where B is a constant. This is the equation of a plane through the center of sphere and
containing the two end points of the problem. The intersection of this plane with the two
points and passing through center of sphere is a great circle. This completes the problem
solution.

Figure 18: Intersection of a plane with a sphere

Note that to cover all possible pairs of points we really have to do this problem in
parametric form since for example if we tried to express solutions in terms of x as x, y(x), z(x),
then any two points given in yz plane would not have a great circle path expressible in x.

56
Problem

1. A particle moves on the surface φ(x, y, z) = 0 from the point (x1 , y1 , z1 ) to the point
(x2 , y2 , z2 ) in the time T . Show that if it moves in such a way that the integral of its kinetic
energy over that time is a minimum, its coordinates must also satisfy the equations
ẍ ÿ z̈
= = .
φx φy φz

2. Specialize problem 2 in the case when the particle moves on the unit sphere, from (0, 0, 1)
to (0, 0, −1), in time T .

3. Determine the equation of the shortest arc in the first quadrant, which passes through
π
the points (0, 0) and (1, 0) and encloses a prescribed area A with the x-axis, where A ≤ .
8
π
4. Finish the example on page 51. What if L = ?
2
5. Solve the following variational problem by finding extremals satisfying the conditions
 π 
4y12 + y22 + y1 y2 dx
4
J(y1 , y2) =
0
   
π π
y1 (0) = 1, y1 = 0, y2 (0) = 0, y2 = 1.
4 4

6. Solve the isoparametric problem

 1
J(y) = (y )2 + x2 dx, y(0) = y(1) = 0,
0

and  1
y 2 dx = 2.
0

7. Derive a necessary condition for the isoparametric problem

Minimize  b
I(y1 , y2 ) = L(x, y1 , y2, y1 , y2 )dx
a
subject to  b
G(x, y1 , y2, y1 , y2 )dx = C
a
and
y1 (a) = A1 , y2 (a) = A2 , y1 (b) = B1 , y2 (b) = B2
where C, A1 , A2 , B1 , and B2 are constants.

8. Use the results of the previous problem to maximize

57
  t1
I(x, y) = (xẏ − y ẋ)dt
t0

subject to  t1 
ẋ2 + ẏ 2dt = 1.
t0

Show that I represents the area enclosed by a curve with parametric equations x = x(t),
y = y(y) and the contraint fixes the length of the curve.

9. Find extremals of the isoparametric problem

 π
I(y) = (y )2 dx, y(0) = y(π) = 0,
0
subject to  π
y 2dx = 1.
0

58
CHAPTER 6

6 Integrals Involving More Than One Independent Vari-

able

Up to now our integrals have been single integrals, i.e. integrals involving only one indepen-
dent variable which we have usually called x.
There are problems in the calculus of variations where the integral involves more than
one independent variable. For example, given some contour C in xyz space, then find the
surface z = z(x, y) contained within C that has minimum surface area. In this case we’d
minimize the surface area integral
  
S= 1 + zx2 + zy2 dy dx (1)
R

where R is the region in the xy plane enclosed by the projection of C in the xy plane. In
this problem there are two independent variables, x, y and one dependent variable, z.
In order to see what conditions for a minimum hold when the integrand involves more
than one independent variable, i.e. the Euler Lagrange equations in this more general
case, let I be defined by  
I= F (x, y, z, zx , zy )dydx (2a)
R
where x, y are the independent variables and z is a continuously differentiable function of
x, y and is to be determined, subject to

z = g(s) (2b)

on the boundary of R where s is arc length, R is some closed region in the xy plane, and F
has continuous first and second partial derivatives with respect to its arguments.
Doing the analogous steps that we did in the single integral problems, assume that
z0 : z0 (x, y) is a solution to this problem and that η(x, y) is a surface which is continuous
with continuous first partials defined over R and satisfies

η(x, y) = 0 on boundary of R . (3)

Create the family of surfaces

z() = z0 (x, y) + η(x, y) (4)
and evaluate I on this family to obtain
 
I() = F [x, y, z0 (x, y) + η(x, y), z0x (x, y) + ηx (x, y), z0y (x, y) + ηy (x, y)]dxdy (5)
R

Differentiating I() with respect to  at  = 0 and setting it to zero (why?) gives

 

0 = I (0) = [Fz η + Fzx ηx + Fzy ηy ]dydx (6)
R

59
At this point, let’s recall (from an earlier chapter) the line of reasoning followed for the single
integral case. The expression corresponding to (6) was
 x2
0 = I  (0) = [Fy η + Fy
η  ]dx
x1

We then rewrote this integrand (by using integration by parts) to involve only η  terms
instead of η  and η and used the fundamental lemma to get the Euler-Lagrange equation.
As an alternate to this procedure we could have used a variant of the integration by parts
formula used above and then written the integrand above in terms of η, with no η  terms.
Our next step would have been to use a modified form of the fundamental lemma introduced
in chapter 4, involving η but not η  terms.
As a generalization to two variables of that modified form of the fundamental lemma we
have
Lemma 1. If α(x, y) is continuous over a region R in the xy plane and if
 
α(x, y)η(x, y)dydx = 0
R

for every continuous function η(x, y) defined over R and satisfying η = 0 on the boundary
of R, then α(x, y) ≡ 0 for all (x, y) in R.
We will not prove this lemma since it is not pertinent to the discussion.
Returning now to our double integral and equation (6), then the second term in the
integrand there can be written
∂ ∂Fzx
Fzx ηx = [Fzx η] − η (7)
∂x ∂x
This is analogous to the integration by parts formula used in the single integral problems.
Now recalling Green’s theorem
  
(Qx + Py )dydx = (Q cos ν + P sin ν)ds (8)
R
boundary
of R

where P, Q are functions of x, y; ν is the angle between the outward normal of the boundary
curve of R and the positive x-axis (see figure 19); ds is the differential of arc length and the
boundary integral is taken in a direction to keep R on the left (positive).
Integrating (7) over R and using (8) with Q as Fzx η and P ≡ 0 gives:
    
∂
Fzx ηx dydx = Fzx η cos νds − (Fz )ηdydx (9)
R
boundary
R ∂x x
of R

By performing a similar line of reasoning on the third term in the integrand of (6), then (6)
becomes
  
∂ ∂
0 = I  (0) = [Fzx cos ν − Fzy sin ν]ηds + [Fz − Fzx − Fz ]ηdydx (10)
boundary
R ∂x ∂y y
of R

60
 y

Figure 19: Domain R with outward normal making an angle ν with x axis

Thus in the expression for the derivative of I with respect to , (at  = 0), we have
written all terms involving η and eliminated ηx and ηy . This is entirely analogous to the
single integral case outlined above.
Since (10) is true for all η(x, y) which satisfy (3) then the first integral on the right side
of (10) is zero for such η and then by lemma 1, the coefficient of η in the second integral of
(10) must be zero over R. That is

∂ ∂
Fzx + Fz − Fz = 0 (11)
∂x ∂y y
which constitutes the Euler-Lagrange equation for this problem.
As an application of the above results, consider the minimal surface problem started
before. Thus minimize   
S= 1 + zx2 + zy2 dydx (12)
R

where the surface is assumed representable in the form z = z(x, y) with z(x, y) specified on
C, the given contour and R is the region in the xy plane, enclosed by the projection of C.
Then (11) gives    
∂ zx ∂ zy
 +  =0 (13)
∂x 1 + zx2 + zy2 ∂y 1 + zx2 + zy2
which by algebra can be reduced to

(1 + zy2 )zxx − 2zx zy zxy + (1 + zx2 )zyy = 0 (14)

Next, by setting
p = zx q = zy r = zxx u = zxy t = zyy (15)
then (14) becomes
(1 + q 2 )r − 2pqu + (1 + p2 )t = 0 (16)

61
Now from differential geometry the mean curvature, M, of the surface is
Eg − 2F f + Ge
M≡ (17)
2(EG − F 2 )

where E, F, G and e, f, g are the coefficients of the first and second fundamental forms of the
surface. For surfaces given by z = z(x, y) then one can show that

E = 1 + p2 F = pq G = 1 + q2 (18a)

and
r u t
e= √ f=√ g=√ (18b)
1 + p2 + q 2 1 + p2 + q 2 1 + p2 + q 2
so that
(1 + p2 )t − 2upq + (1 + q 2 )r
M= (19)
2(1 + p2 + q 2 )3/2
So the numerator is the same as the left side of Euler’s equation (16). Thus (16) says that
the mean curvature of the minimal surface must be zero.

Problems

1. Find all minimal surfaces whose equations have the form z = φ(x) + ψ(y).

2. Derive the Euler equation and obtain the natural boundary conditions of the problem
   
δ α(x, y)u2x + β(x, y)u2y − γ(x, y)u2 dxdy = 0.
R

In particular, show that if β(x, y) = α(x, y) the natural boundary condition takes the form

∂u
α δu = 0
∂n
∂u
where is the normal derivative of u.
∂n
3. Determine the natural boundary condition for the multiple integral problem

 
I(u) = L(x, y, u, ux, uy )dxdy, uC 2 (R), u unspecified on the boundary of R
R

4. Find the Euler equations corresponding to the following functionals

 
a. I(u) = (x2 u2x + y 2u2y )dxdy
R
 
b. I(u) = (u2t − c2 u2x )dxdt, c is constant
R

62
CHAPTER 7

7 Examples of Numerical Techniques

Now that we’ve seen some of the results of the Calculus of Variations, we can study the
solution of some problems by numerical techniques.
All of the numerical techniques used in variational problems are iterative in nature, that
is, they do not solve the problem in one step but rather proceed from an initial estimate
(usually input by the user) and generate a sequence of succeeding estimates which converges
to the answer.
The iterative procedures used, are based upon a search from the present estimate to
obtain a next estimate which has certain characteristics. The types of search procedures fall
into two main classes called “Indirect Methods” and “Direct Methods.” We will also look
at a computer program for the variable end point case using indirect methods.

7.1 Indirect Methods

Indirect methods are those which seek a next estimate satisfying certain of the necessary
conditions for a minimizing arc, established previously. Thus these methods for example
seek arcs that satisfy the Euler equations. An example of an indirect method is Newton’s
method for variational problems. We will now discuss this method and provide a sample
computer program written in Matlab for students to try on their computer. First we discuss
the fixed end points case.

7.1.1 Fixed End Points

Consider the fixed endpoint problem of minimizing the integral
 x2
I= f (x, y, y )dx (1)
x1

among arcs satisfying

y(x1 ) = Y1 , y(x2 ) = Y2 . (2)
The indirect method seeks to find an arc y0 which satisfies the Euler equations and also
satisfies the endpoint conditions (2).
Writing the Euler equation
d
fy
= fy (3)
dx
and then differentiating, gives

fy
x + fy
y y  + fy
y
y  = fy (4)

(Note that we assumed that our solution will have a second derivative y  at each point).

63
 In this procedure, the selection of y(x) and y (x) for x1 < x ≤ x2 is dictated by (4) as
soon as
y(x1 ) and y  (x1 ) (5)
are selected.
Thus each time we alter the initial conditions (5), we will get a different solution of
(4). Since by the first part of (2), the value of y(x1 ) is fixed, then the only variable left
to satisfy the second part of (2) is y (x1 ). Calling the initial estimate of the minimizing
arc y1 with value y1 (x1 ) and denoting the value of left end-point slope for any other arc
y (x1 , c) = y1 (x1 ) + c, then the solutions to (4) are a family of arcs

y(c) : y(x, c) x1 ≤ x ≤ x2 (6)

so that
y (x1 , c) = y1 (x1 ) + c and y (x1 , 0) = y1 (x1 ) (7)

∂y  (x1 , c) 
Differentiating the family (6) with respect to c at c = 0 we obtain (since  = 1)
∂c c=0
   
∂y(x, c)  ∂y(x, c) 
η(x) ≡  =  x1 ≤ x ≤ x2 (8)
∂c c=0 ∂y (x1 , c) c=0

∂y(x, c) 
where we have assigned the name η(x) to
∂c c=0
∂y(x2 , c) 
In particular at x = x2 we get (= η(x2 )) as the change in the value of
∂y  (x1 , c) c=0
y(x2 , 0) to a solution to (4) with each unit change in value of its left end-point slope
y1 (x1 ) (= y (x1 , 0)). Thus knowing η(x2 ), we can form the differential correction to y1 (x1 ) as

Y2 − y1 (x2 )
∆y1 (x1 ) = (9)
η(x2 )

and use this to iterate on y1 (x1 ) to satisfy the second part of (2). In order to obtain η(x) we
note that for any arc

y(c) : y(x, c) y  (x, c) x1 ≤ x ≤ x2 (10)

in our family (6) then by (4) we must have

fy
x (x, y(x, c), y (x, c)) + fy
y (x, y(x, c), y (x, c)) y (x, c)
(11)
+fy
y
(x, y(x, c), y (x, c)) y (x, c) = fy (x, y(x, c), y (x, c))

Differentiating (11) with respect to c at c = 0 and assuming that in our family, y(x, c) is
continuously differentiable up through third order in x, c so that order of differentiation is
immaterial and
  
 ∂ 2 y(x, c)  ∂ 2 y(x, c)  ∂y  (x, c) 
η (x) = = =
∂x∂c c=0 ∂c∂x c=0 ∂c c=0

64
and
 
 ∂ 3 y(x, c)  ∂ 3 y(x, c) 
η (x) = =
∂x∂x∂c c=0 ∂c∂x∂x c=0
which results in,

fy
xy η + fy
xy
η  + (fy
yy η + fy
yy
η  )y1 + fy
y η 

(12)
+(fy
y
y η + fy
y
y
η  )y1 + fy
y
η  − fyy η − fyy
η  = 0

where in (12) all arguments of the derivatives of f are x, y(x), y (x) i. e. along the arc y1 .
Equation (12) represents a second order linear differential equation for η. The initial
conditions for solution are obtained by differentiating (10) with respect to c at c = 0. Thus
  
∂y(x1 , c)  ∂y  (x1 , c)  ∂y  (x1 , c) 
= η(x1 ) , 1= = = η  (x1 ) (13)
∂c c=0 ∂y  (x1 , c) c=0 ∂c 
c=0

where in the second equation in (13) we have recalled the definition of c. Then by the second
equation of (13) we get that η  (x1 ) = 1. Furthermore, by the first part of (2) we see that for
any c, y(x1 , c) = Y1 = y1 (x1 ) so that η(x1 ) = 0.
Thus we solve for η(x) on x1 ≤ x ≤ x2 by solving the second order differential equation
(12) with initial conditions
η(x1 ) = 0 η  (x1 ) = 1.

For example, suppose we wish to find the minimum of

 1  
I = (y )2 + y 2 dx (14)
0

y(0) = 0
y(1) = 1.
The function odeinput.m supplies the user with the boundary conditions, a guess for the
initial slope, tolerance for convergence. All the derivatives of f required in (4) are supplied
in rhs2f.m.

function [fy1y1,fy1y,fy,fy1x,t0,tf,y1,y2,rhs2,sg,tol] = odeinput

% Defines the problem for solving the ode:
% (f_{y’y’} )y" + (f_{y’y})y’ = f_y - f_{y’x}

% t0 - start time
% tf - end time
% y1 - left hand side boundary value
% y2 - right hand side boundary value
% sg - initial guess for the slope
% tol - tolerance e.g. 1e-4

65
t0 = 0;
tf = 1;
y1 = 0;
y2 = 1;
sg = 1;
tol = 1e-4;

%rhs2f.m
%
function [rhs2]=rhs2f(t,x)
%
%input
% t is the time
% x is the solution vector (y,y’)
%
% fy1fy1 - fy’y’ (2nd partial wrt y’ y’)
% fy1y - fy’y (2nd partial wrt y’ y)
% fy - fy (1st partial wrt y)
% fy1x - fy’x (2nd partial wrt y’ x)
%
fy1y1 = 2;
fy1y = 0;
fy = 2*x(1);
fy1x = 0;

rhs2=[-fy1y/fy1y1,(fy-fy1x)/fy1y1];

The main program is ode1.m which uses a modified version of ode23 from matlab. This
modified version is called ode23m.m. Since we have to solve a second order ordinary differ-
ential equation, we have to transform it to a system of first order to be able to use ode23. To
solve the η equation, the ode23 is used without any modifications. We also need the right
hand side of the 2 equations to be solved (one for y and one for η). These are called odef.m
and feta.m, respectively. All these programs (except the original ode23.m) are given here

% ode1.m
% This program requires an edited version of ode23 called ode23m.m
% Also required is odef.m, feta.m & odeinput.m
% All changes to a problem should ONLY be entered in odeinput.m

66
[fy1y1,fy1y,fy,fy1x,t0,tf,y1,y2,rhs2,sg,tol] = odeinput;
correct = 100;
while abs(correct) > tol
%solve the initial value with the slope guessed
x0=[y1,sg]’;
[t,x]=ode23m(’odef’,t0,tf,x0,y2,’rhs2f’,tol,0);
n1=size(x,1);
yy(1:n1)=x(1:n1,1);
plot(t,yy)

% check the value at tf

% change the value of the slope to match the solution
eta0=[0,1]’;
[tt,eta]=ode23(’feta’,t0,tf,eta0);
[nn1,nn2]=size(eta);
correct=(y2-yy(n1))/eta(nn1);
sg=sg+correct;
end

% msode23m.m
%
% This code is a modified version of MATLAB’s ODE23 to find a numerically integrated
% solution to the input system of ODEs.
%
% This code is currently defined for the variable right hand endpoint defined by the
% following boundary conditions:
% y(0) = 1, y(x1) = Y1 = x2 - 1
%
% Lines which require modification by the user when solving different problems
% (different boundary function) are identified by (user defined) at the right margin.
%
%
function [tout, yout] = msode23m(ypfun, t0, tfinal, y0, rhs2f, tol, trace)

%ODE23 Solve differential equations, low order method.

% ODE23 integrates a system of ordinary differential equations using
% 2nd and 3rd order Runge-Kutta formulas.
% [T,Y] = ODE23(’yprime’, T0, Tfinal, Y0, Y2, rhs2) integrates the system
% of ordinary differential equations described by the M-file YPRIME.M,

67
% over the interval T0 to Tfinal, with initial conditions Y0.
% [T, Y] = ODE23(F, T0, Tfinal, Y0, y2, rhs2, TOL, 1) uses tolerance TOL
% and displays status while the integration proceeds.
%
% INPUT:
% F - String containing name of user-supplied problem description.
% Call: yprime = fun(t,y) where F = ’fun’.
% t - Time (scalar).
% y - Solution column-vector.
% yprime - Returned derivative column-vector;
% yprime(i) = dy(i)/dt.
% t0 - Initial value of t.
% tfinal- Final value of t.
% y0 - Initial value column-vector.
% tol - The desired accuracy. (Default: tol = 1.e-3).
% trace - If nonzero, each step is printed. (Default: trace = 0).
%
% OUTPUT:
% T - Returned integration time points (column-vector).
% Y - Returned solution, one solution column-vector per tout-value.
%
% The result can be displayed by: plot(tout, yout).
%
% See also ODE45, ODEDEMO.
% C.B. Moler, 3-25-87, 8-26-91, 9-08-92.
% Copyright (c) 1984-93 by The MathWorks, Inc.
%
% Initialization
pow = 1/3;
if nargin < 7, tol = 1.e-3; end
if nargin < 8, trace = 0; end
t = t0;
hmax = (tfinal - t)/256; %(user defined)
%the denominator of this expression may
%require adjustment to
%refine the number of subintervals over
%which to numerically
%integrate - consider adjustment if infinite
%loops are encountered
%within this routine and keep the value as
%a power of 2
h = hmax/8;
y = y0(:);
chunk = 128;

68
tout = zeros(chunk,1);
yout = zeros(chunk,length(y));
k = 1;
tout(k) = t;
yout(k,:) = y.’;
if trace
clc, t, h, y
end
% The main loop
while (t < tfinal) & (t + h > t)
if t + h > tfinal, h = tfinal - t; end
% Compute the slopes
rhs2=feval(rhs2f,t,y); rhs2=rhs2(:);
s1 = feval(ypfun, t, y,rhs2); s1 = s1(:);
rhs2=feval(rhs2f,t+h,y+h*s1); rhs2=rhs2(:);
s2 = feval(ypfun, t+h, y+h*s1,rhs2); s2 = s2(:);
rhs2=feval(rhs2f,t+h/2,y+h*(s1+s2)/4); rhs2=rhs2(:);
s3 = feval(ypfun, t+h/2, y+h*(s1+s2)/4,rhs2); s3 = s3(:);
% Estimate the error and the acceptable error
delta = norm(h*(s1 - 2*s3 + s2)/3,’inf’);
tau = tol*max(norm(y,’inf’),1.0);
% Update the solution only if the error is acceptable
if delta <= tau
t = t + h;
y = y + h*(s1 + 4*s3 + s2)/6;
k = k+1;
if k > length(tout)
tout = [tout; zeros(chunk,1)];
yout = [yout; zeros(chunk,length(y))];
end
tout(k) = t;
yout(k,:) = y.’;
end
if trace
home, t, h, y
end
% Update the step size
if delta ~= 0.0
h = min(hmax, 0.9*h*(tau/delta)^pow);
end

varendpt = t - 1; %(user defined)

tolbnd = 1e-2; %(user defined)
%varendpt is the equation of the variable

69
 %endpoint as defined by
%the right hand side boundary curve where
%t is the independent variable
%tolbnd is the desired tolerance for meeting
%the variable right
%endpoint condition and may require some
%experimentation

if abs(y(1) - varendpt) < tolbnd

%this checks to see if the endpoint of the solution
disp(’hit boundary in msode23m’);
break; %curve comes within a user specified
end %tolerance of the right hand side
%boundary curve

end

if (t < tfinal)
disp(’Singularity likely.’)
t
end

tout = tout(1:k);
yout = yout(1:k,:);

% feta.m

function xdot=feta(t,x)
xdot=[x(2),0]’;

% odef.m
function xdot=odef(t,x,rhs2)
xdot=[x(2),rhs2(1)*x(2)+rhs2(2)]’;

The solution obtained via matlab is plotted in figure 20.

70
 1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 20: Solution of example given by (14)

7.1.2 Variable End Points

We have previously obtained necessary conditions that a solution arc to the variable end
point problem had to satisfy.
We consider now a computer program for a particular variable end point problem. We
will use Newton’s method to solve the problem:
Thus consider the problem of minimizing the integral
 x2  x2
I= f (x, y, y )dx = [(y )2 + y 2]dx (15)
x1 x1

among arcs satisfying

y(x1 ) = 1, y(x2 ) = Y2 = x2 − 1, x1 = 0 (16)

(where we use Y2 (x) for the right hand boundary, which is a straight line).
Our procedure now will be much the same as for the fixed end point problem done by
Newton’s method in that we’ll try to find a solution to the Euler equation. Also as before,
all of our estimate arcs y of solutions to this problem will have

y(x1 ) = 1 x1 = 0 (17)

so that these items are fixed. However we note that in general this will not be the case, and
in other problems we may be allowed to vary these quantities in our iterative procedure but
will then be required to satisfy a transversality condition involving them.
Returning now to the problem at hand, we start with an initial estimate y 1 , satisfying
the left end point condition
y1 (x1 ) = 1 x1 = 0 (18)
and the Euler equations
d
fy
= fy or fy
x + fy
y y  + fy
y
y  = fy . (19)
dx

71
As for the fixed endpoint case, only y (x1 ) is free to iterate with, so that setting

y  (x1 , c) = y1 (x1 ) + c with y(x1 , c) = y1 (x1 ) = 1 (20a)

and integrating the Euler equation we get the family

y(c) : y(x, c) x1 ≤ x ≤ x2 (c) −δ ≤c≤δ (20b)

(where only the right end value of x varies with c since the left end value is fixed and) which
safisfies the Euler equation and

y(x1 , c) = 1 x1 = 0 (21)

Thus we have on this family

fy
x (x, y(x, c), y (x, c)) + fy
y (x, y(x, c), y (x, c))y (x, c) + (22)

fy
y
(x, y(x, c), y (x, c))y (x, c) = fy (x, y(x, c), y (x, c))

Proceeding as we did in the fixed endpoint case we differentiate (22) with respect to c at
c = 0. Thus

fy
xy η + fy
xy
η  + (fy
yy η + fy
yy
η  )y1 + fy
y η  (23)

+(fy
y
y η + fy
y
y
η  )y1 + fy
y
η  = fyy η + fyy
η 

which is the same equation for η that we got in the fixed endpoint case.
The initial conditions for η, η , are obtained from (20a) by differentiation (at c = 0). In
particular, differentiating the second part of (20a) yields

∂y(x1 , c)
η(x1 ) = =0 (24)
∂c
and differentiating the first part of (20a) gives

∂y  (x1 , c)
η  (x1 ) = =1 (25)
∂c
We have two conditions that our estimates have to satisfy at the right hand end, namely,
(with subscript F denoting final values, e.g. yF (c) ≡ y(x2 (c), c)).

yF = Y2 = x2 − 1 (26a)

and the transversality condition (3) of chapter 4 which applied to this problem yields

2yF − (yF )2 + yF2 = 0 (26b)

72
Since x2 is unrestricted we choose to stop integration for each estimate when (26a) is satisfied
and there to evaluate the expression (26b) which we call TERM

TERM = 2yF − (yF )2 + yF2 (27)

Then if TERM differs from 0 we compute as before how much to change c by in order to
reduce this value
−T ERM
c = d(T ERM ) (28)
dc

Next, differentiating (27) yields

d(T ERM) dy  dy  dyF dy  dyF

= 2 F − 2yF F + 2yF = 2[(1 − yF ) F + yF ] (29a)
dc dc dc dc dc dc
where all arguments are along the arc y 1 .
Now concentrating on yF which is a function of c

yF (c) ≡ y(x2 (c), c) (29b)

and differentiating with respect to c at c = 0 (i.e. at y 1 ) yields

dyF ∂y(x2 , c) dx2

= +yF (30a)
dc  ∂c
  dc
ηF

Doing analogous operations for yF (c) yields after differentiation with respect to c at c = 0.

dyF dx2
= ηF + yF (30b)
dc dc
Also by differentiating the middle constraint in (16) i.e. the equation yF (c) = Y2 =
x2 (c) − 1 yields
dyF dx2
= (30c)
dc dc
so that putting together (30a) and (30c) gives

dx2 dyF dx2

= = ηF + yF (30d)
dc dc dc
or then
dx2
(1 − yF ) = ηF (30e)
dc
(compare to equation (6) of the appendix 4.2 but with  = c and with x0i = x2 and Yi = x2 −1
dYi dx2
so that = ) or then
dc dc
dx2 ηF
= (30f )
dc 1 − yF

73
and then by (29a), (30b), (30f), (30a) we get

d(T ERM) ηF ηF
= 2[(1 − yF )(ηF + yF 
) + yF (ηF + yF )] (31)
dc 1 − yF 1 − yF

From the Euler equation we get yF = yF so that after collecting terms

d(T ERM) yF ηF yF yF ηF

= 2[(1 − yF )(ηF + ) + y η
F F + ]= (32)
dc 1 − yF 1 − yF
yF ηF
= 2[ηF − yF ηF + yF ηF + ]
1 − yF

We have thus obtained all of the quantities necessary to compute the correction to c.

The program for the present problem is then:

a) start with an initial estimate y 1 with y1 (x1 ) = 1, y1 (x1 ) = y  (x1 ), x1 = 0

∂y
b) integrate the Euler equation for y and the equation for η = stopping the integration
∂c
when the end point condition y(x2 ) = Y2 is met

c) determine the error in the transversality condition and the correction in y (x1 ) needed
 
−T ERM d(T ERM)
to correct it d(T ERM ) = c, where is computed using ηF .
dc
dc

d) re-enter (b) with initial conditions y(x1) = 1, x1 = 0, y (x1 ) = y1 (x1 ) + c and continue
through the steps (b) and (c)

e) stop when the error is smaller than some arbitrary number .

7.2 Direct Methods

Direct methods are those which seek a next estimate by working directly to reduce the
functional value of the problem. Thus these methods search in directions which most quickly
tend to reduce I. This is done by representing I to various terms in its Taylor series and
reducing the functions represented.
The direct method we will work with is the gradient method (also called method of
steepest descent). This method is based on representing the integral to be minimized as
a linear functional of the arcs y over which it is evaluated. The gradient method has an
analogue for finite dimensional optimization problems and we will first describe this method
in the finite dimensional case.
Thus suppose that we wish to minimize a function of the two dimensional vector y =
(y1 , y2 )
f (y) (= (y1 )2 + (y2 )2 as an example) (33)

74
subject to the constraint

(y) = 0 (y1 + y2 − 1 = 0 for our example) . (34)

The gradient method says that starting with an initial estimate y 1 = (y1,1 , y1,2), we first
linearize f as a function of the change vector η = (η1 , η2 ).
Expanding f to first order at the point y 1 , gives
2 2
f (y 1 + η) ∼ f (y 1 ) + fy1 η1 + fy2 η2 (= y1,1 + y1,2 + 2(y1,1 η1 + y1,2 η2 )) (35)

if |η| is small. Since f (y 1 ) is constant, then this allows us to consider f as a function F of

only the change vector η = (η1 , η2 )

F (η) ≡ f (y 1 ) + fy1 η1 + fy2 η2 (36)

where we don’t list  as an argument of F since it will be determined independently of η and

we wish to concentrate on the determination of η first.
We can similarly linearize the constraint  and approximate  by the function L which
depends on η
L(η) (37)
Now we wish to choose η in order that F (η) is as small as possible and also L(η) = 0
for a given step size length, (|η| = ST ). Recall from calculus, that the maximum negative
(positive) change in a function occurs if the change vector is opposite (in the same direction)
to the gradient of the function. Now, the gradient of the function (36) considered as a
function of η is:

∇F = (Fη1 , Fη2 ) = (fy1 , fy2 ) (= (2y1,1 , 2y1,2) for our example) (38)

so that fastest way to reduce F requires that η be oriented in the direction

η = (η1 , η2 ) = −(fy1 , fy2 ) (= (−2y1,1 , −2y1,2 ) for our example) (39)

(note that this choice is independent of ). However since we have a constrained problem,
then our change should be restricted so that our new point y 1 + η satisfies

(y 1 + η) = 0 (40a)

or by our approximation
L(η) = 0 (40b)
according to the way we defined L. Thus we modify η from (39) so that it satisfies (40b).
These conditions establish the direction of η and then the value of  is established by the
requirement that |η| = ST , i. e. the change is equal to the step size.
The gradient procedure then computes the function f (y 1 + η) which should be smaller
than f (y1 ) and repeats the above procedure at the point y 2 = y 1 + η.

75
 In the infinite dimensional case, the idea is the same, except that we are now dealing
with a function of an infinite number of variables, namely arcs

y: y(x) x1 ≤ x ≤ x2

and our change vector will have direction defined by the arc

η: η(x) x1 ≤ x ≤ x2

Thus consider the case of minimizing the integral

 x2
I= f (x, y, y )dx (41)
x1

subject to the fixed endpoint conditions (the constraint on the problem)

y(x1 ) = a y(x2 ) = b (42)

Following the procedure used in the finite dimensional case, we start with an initial arc y 1
and first linearize the integral I by computing the first variation of I  .
 x2
I = [fy η + fy
η  ]dx (43)
x1

Integrating (by parts) the first term in the integrand gives,

 x2  x  x2  x

fy η(x)dx = [η(x) fy ds]xx21 − [η (x) fy ds]dx (44)
x1 x1 x1 x1

Since the variations η(x) must (why?) satisfy

η(x1 ) = η(x2 ) = 0 (45)

then the first term on the right in (44) vanishes and plugging back into (43) gives
 x2  x

I = [f −
y
fy ds]η  (x)dx . (46)
x1 x1

Corresponding to (36) we then approximate I(y 1 + η) by

Iˆ = I(y1 ) + I  (47)

where the second term is represented by (46). Analogous to the finite dimensional case, we
desire to select η or equivalently η(x1 ) and η  (x), x1 ≤ x ≤ x2 so that subject to a step size
constraint, we have that Iˆ (and also approximately I) has minimum value at y1 + η. The
stepsize constraint in this case looks like

Alternatively we can think of this as the derivative at  = 0 of I evaluated on the family y(): created
with the arc η(x) but we don’t set it = 0 (why?)

76
 max |η  (x)|
(48)
x1 ≤ x ≤ x2
(which represents the maximum change from y1 (x) along our arcs) and where  will be
selected according to the stepsize we wish. It can be shown formally that the best selection
of η  (x) at each x is  x

η (x) = −[fy
− fy ds] x1 ≤ x ≤ x2 (49)
x1

This hueristically can be considered the direction opposite to the gradient of Iˆ with respect
to η  (x) for each x. However, as in the finite dimensional case, we must modify this change
in order to satisfy the constraint (45). Defining the integral of η  (x) of (49) from x1 to x as
 x  ξ
M(x) = − [fy
− fy ds]dξ (50)
x1 x1

and defining the average of this as

 x2  x
M(x2 ) 1
Mavg = =− [f −
y
fy ds]dx (51)
x2 − x1 x2 − x1 x1 x1

(note that M(x1 ) = 0) then with η  (x) defined as

 x

η (x) = − [fy
− fy ds] − Mavg =
x1
 x  x2  x (52)
1
−[fy
− fy ds − [fy
− fy ds]dx]
x1 x2 − x1 x1 x1

we get  x2
η(x2 ) = η  (x)dx + η(x1 ) = η(x1 ) (53)
x1

which together with η(x1 ) = 0 (which we can easily choose) yields η which satisfies our
constraint (45). Integrate (52) from x1 to x

η(x) = M(x) − (x − x1 )Mavg .

While this is not the only way to create η satisfying (45), it can be formally shown that
subject to (45), this η(x) is the best selection to reduce I.

We now give a matlab program that uses direct method to minimize the integral I. This
program requires the user to supply the functions f, fy , fy
. These functions are supplied in
the finput.m file that follows.

% This program solves problems of the form

% ___x2
% |
% Minimize I = | f(x,y,y’) dx

77
% ___|x1
%
% using the direct method. The user must supply the functions
% F(x,y,y’), Fy(x,y,y’)
% and Fy’(x,y,y’) in a file called finput.m
%
% See finput.m

% By Jerry Miranda, 12/10/96

% WARNING: Early termination may occur if N is TOO large or if epsilon is

% TOO small. The count parameter is set at 50 and can be adjusted
% below. Count is required to prevent runaway in the while loop or
% excessive computation until this version is modified.
clear
C = 50;
% set the count paramater

% Here we solve the problem min [ int(0->1) {2y’ + y^2} dx]

% s.t. y(0)=0, y(1)=1

% setup boundary conditions (User define)

x1 = 0; y1 = 0; % y(x1) = y1
x2 = 1; y2 = 1; % y(x2) = y2

% choose an epsilon and the number of points to iterate (User define)

epsilon = .01; N = 25;

if x2-x1 == 0, error(’x2 and x1 are the same’), break, end

deltax = (x2-x1)/N; x = [x1:deltax:x2]’; % x is a col vector

% make an initial guess for the solution arc: (User define)

% this is a function satisfying the boundary conditions
ybar = (y2-y1)/(x2-x1)*(x-x1)+y1;

% this is the derivative of a linear function ybar

% if ybar is NOT linear,
% we should use finite difference to approximate yprime

yprime = ones(size(x))*(y2-y1)/(x2-x1);

% calculate M(x2) and Mavg

78
sum1=0;
MM(1)=0;
for i = 2:N+1
sum2=0;
for jj=1:i-1
sum2= deltax*finput(x(jj),ybar(jj),yprime(jj),2)+sum2;
end

sum1 = deltax*(finput(x(i),ybar(i),yprime(i),3)-sum2)+sum1;
MM(i)= - sum1;
end
Mx2 = - sum1;
Mavg = Mx2/(x2-x1);
% Calculate eta(x) for each x(i)
for i = 1:N+1
eta(i,1) = MM(i) - Mavg*(x(i)-x1);
end
% Calculate eta’(x) for each x(i)
for i = 1:N+1
sum2=0;
for jj=1:i-1
sum2= deltax*finput(x(jj),ybar(jj),yprime(jj),2)+sum2;
end
etaprm(i,1)= - finput(x(i),ybar(i),yprime(i),3)-sum2 -Mavg;
end
% The main loop
% We now compute Ihat = I(ybar1) + epsilon*I’ and check to minimize Ihat

% First Ihat
sum1=0;

for i = 1:N+1
F = finput(x(i),ybar(i),yprime(i),1);
sum1 = deltax*F+sum1;
end
Ihatnew = sum1; Ihatold = Ihatnew+1;
count = 0; %set counter to prevent runaway

while (Ihatnew <= Ihatold) & (count <= C)

count = count + 1;

% Integrate to get I’
sum1=0;

79
 for i = 1:N+1
sum2 =0;
for j = 1:i-1
Fy = finput(x(j),ybar(j),yprime(i),2);
sum2 = deltax*Fy+sum2; % what delta is used
end
Fyp = finput(x(i),ybar(i),yprime(i),3);
sum1 = deltax*(Fyp+sum1-sum2)*etaprm(i);
end
Iprm = sum1;

% Integrate to get I

sum1=0;

for i = 1:N+1
F = finput(x(i),ybar(i),yprime(i),1);
sum1 = deltax*F+sum1;
end
I = sum1;
Ihatnew = I + epsilon*Iprm;
if Ihatnew < Ihatold

ybar = ybar + epsilon*eta;

Ihatold = Ihatnew;
for ij=2:N+1
yprime(ij)=(ybar(ij)-ybar(ij-1))/(x(ij)-x(ij-1));
end
end
end

% we now have our solution arc ybar

plot(x,ybar), grid, xlabel(’x’), ylabel(’y’)

title(’Solution y(x) using the direct method’)

function value = finput(x,y,yp,num)

% function VALUE = FINPUT(x,y,yprime,num) returns the value of the
% functions F(x,y,y’), Fy(x,y,y’), Fy’(x,y,y’) at a given
% x,y,yprime
% for a given num,

80
% num defines which function you want to evaluate.
% 1 for F, 2 for Fy, 3 for Fy’.

if nargin < 4, error(’Four arguments are required’), break, end

if (num < 1) | (num > 3)
error(’num must be between 1 and 3’), break
end

if num == 1, value = yp^2 + y^2; end % F

if num == 2, value = 2*y; end % Fy
if num == 3, value = 2*yp; end % Fy’

Problems
 x1  
2
1. Find the minimal arc y(x) that solves, minimize I = y 2 − (y ) dx
0

a. Using the indirect (fixed end point) method when x1 = 1.

π
b. Using the indirect (variable end point) method with y(0)=1 and y(x1 ) = Y1 = x2 − .
4
 1

1 2
2. Find the minimal arc y(x) that solves, minimize I = (y  ) + yy  + y  + y dx
0 2
where y(0) = 1 and y(1) = 2.
 x1  
2
3. Solve the problem, minimze I = y 2 − yy  + (y  ) dx
0
a. Using the indirect (fixed end point) method when x1 = 1.
b. Using the indirect (variable end point) method with y(0)=1 and y(x1 ) = Y1 = x2 − 1.

4. Solve for the minimal arc y(x) :

 1  
I = y 2 + 2xy + 2y  dx
0

where y(0) = 0 and y(1) = 1.

81
CHAPTER 8

8 The Rayleigh-Ritz Method

We now discuss another numerical method. In this technique, we approximate the variational
problem and end up with a finite dimensional problem. So let us start with the problem of
seeking a function y = y0 (x) that extremizes an integral I(y). Assume that we are able to
approximate y(x) by a linear combination of certain linearly independent functions of the
type:

y(x) ≈ φ0 (x) + c1 φ1 (x) + c2 φ2 (x) + · · · + cN φN (x) (1)

where we will need to determine the constant coefficients c1 , · · · cN .
The selection of which approximating function φ(x) to use is arbitrary except for the
following considerations:
a) If the problem has boundary conditions such as fixed end points, then φ0 (x) is chosen
to satisfy the problem’s boundary conditions, and all other φi vanish at the boundary. (This
should remind the reader of the method of eigenfunction expansion for inhomogeneous partial
differential equation. The functions φi are not necessarily eigenfunctions though.)
b) In those problems where one knows something about the form of the solution then
the functions φi(x) can be chosen so that the expression (1) will have that form.
By using (1) we essentially replace the variational problem of finding an arc y(x) that
extremizes I to finding a set of constants c1 , · · · , cN that extremizes I(c1 , c2 , · · · , cN ). We
solve this problem as a finite dimensional one i. e. by solving
∂I
= 0, i = 1, · · · , N (2)
∂ci
The procedure is to first determine an initial estimate of c1 by the approximation y ≈
φ0 +c1 φ1 . Next, the approximation y ≈ φ0 +c1 φ1 +c2 φ2 is used (with c1 being redetermined).
The process continues with y ≈ φ0 + c1 φ1 + c2 φ2 + c3 φ3 as the third approximation and so
on. At each stage the following two items are true:
a) At the ith stage, the terms c1 · · · , ci−1 that have been previously determined are re-
determined
b) The approximation at the ith stage
y ≈ φ 0 + c1 φ 1 + · · · + ci φ i (3)
will be better or at least no worse than the approximation at the i − 1st stage
y ≈ φ0 + c1 φ1 + · · · + ci−1 φi−1 (4)
Convergence of the procedure means that
 


N
lim φ0 + ci φi = y0 (x) (5)
N →∞
i=1

82
where y0 (x) is the extremizing function. In many cases one uses a complete set of functions
e. g. polynomials or sines and cosines. A set of functions φi (x) (i = 1, 2, · · ·) is called
complete over [a, b] if for each Riemann integrable∗∗ function f (x), there is a number N
(depending on , c1 , · · · , cN ) such that

N
max[f − ci φi ]2 <  (6)
[a,b]
i=1

The above outlined procedure can be extended in a number of ways. For example, more
than one independent variable may be involved. So for the problem of
 
min I = F (x, y, w, wx, wy )dydx (7)
R

subject to
w = h(s) on the boundary Γ of R (8)
where h(s) is some prescribed function and s is the arc length along Γ. Analogous to (1) we
write
w(x, y) = φ0 (x, y) + c1 φ1 (x, y) + · · · + cN φN (x, y) (9)
and φ0 satisfies (8) and φi (x) i = 1, 2, 3 · · · are zero on Γ. We could also extend the procedure
to functions involving higher derivatives, more independent variables, etc.
Example 1: Apply the procedure to the problem of
 1
min I = [(y  )2 − y 2 − 2xy]dx (10)
0

with the boundary conditions

y(0) = 1 y(1) = 2 (11)

Solution: Since the boundary conditions are NOT homogeneous, we have to take φ0
to satisfy the boundary conditions, i.e. φ0 = 1 + x. We choose φ1 (x) = x(1 − x) since it
should satisfy zero boundary conditions. Setting

y1 (x) = 1 + x + c1 x(1 − x) . (12)

Substituting (12) into (10) and performing the integration gives

3 2 2
I=1 + c1 − c1 (13)
10 3
∂I 10
Solving = 0 implies that c1 = or
∂c1 9
10
y1 (x) = 1 + x + x(1 − x) (14)
9
∗∗
A function f is Riemann integrable over [a, b] if all points of discontinuity of f can be enclosed in a set
of subintervals, the sum of whose lengths can be made as small as desired

83
 d2 I 1
as the first approximate solution. We note that 2
is positive at c1 = ; thus we have
dc1 9
minimized I on the class of functions defined by (12).
Continuing we try

yn (x) = 1 + x + x(1 − x)[c1 + c2 x + c3 x2 + · · · + cn xn−1 ] (15)

where n = 2, 3, 4, · · ·. The boundary conditions are satisfied by yn for all values of ci that is
for n = 1, 2, 3, · · ·
yn (0) = 1 yn (1) = 2 . (16)
For n = 2, when
y2 (x) = 1 + x + x(1 − x)[c1 + c2 x] (17)
is used and the integration carried out, we get 2 equations for the two parameters c1 and c2
∂I ∂I
when solving = = 0. This gives
∂c1 ∂c2
c1 = 0.9404 c2 = 0.3415 (18)

so that
y2 (x) = 1 + x + x(1 − x)[0.9404 + 0.3415x] (19)
Comparing the two approximations y1 (x) and y2 (x) with the exact solution††
3 − cos 1
y = cos x + sin x − x (20)
sin 1
In the next figure we plot the exact solution and y1 (x) and y2 (x).
It can be seen that y1 is already reasonably close.

8.1 Euler’s Method of Finite Differences

Euler solved many variational problems by the method of finite differences. Suppose we want
to extremize the integral  xn+1
I(y) = F (x, y, y )dx (21)
x0

where x0 and xn+1 are given and the function y is subject to the boundary conditions
y(x0 ) = y0 and y(xn+1 ) = yn+1 . Dividing the interval [x0 , xn+1 ] into n + 1 equal parts, the
width of each piece is
xn+1 − x0
∆x =
n+1
††
The exact solution is obtained in this problem, by noting that the variational problem (10) and (11) has
the same solution as the boundary-value problem

y  + y = −x y(0) = 1 y(1) = 2

84
 2

1.8

1.6

1.4

1.2

1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 21: The exact solution (solid line) is compared with φ0 (dash dot), y1 (dot) and y2
(dash)

(See Figure 22.) Next, let y1 , y2 , · · · , yn be the values of y corresponding to

x1 = x0 + ∆x, x2 = x0 + 2∆x, · · · , xn = x0 + n∆x
respectively. The associated values y1 , y2 , · · · , yn are unknowns because the function which
solves the problem is unknown as yet. The integral (21) (by definition) is the limit of
a summation, and thus we may approximate the integral by a function of n variables
φ(y1 , y2, · · · , yn ).

 
n
yi+1 − yi
φ(y1 , y2 , · · · , yn ) = F xi , yi , ∆x (22)
i=0 ∆x
In this way the derivative is replaced by a difference quotient and the integral by a finite
sum. The quantities y1 , y2 , · · · , yn are determined so that φ solves
∂φ
=0 i = 1, 2, · · · , n (23)
∂yi
The terms of (22) which involve yi are
   
yi+1 − yi yi − yi−1
F xi , yi , ∆x and F xi−1 , yi−1, ∆x (24)
∆x ∆x
so that
   
∂φ yi+1 − yi yi+1 − yi
= Fyi xi , yi, ∆x − Fyi
xi , yi, (25)
∂yi ∆x ∆x
 
yi − yi−1
+Fyi−1 xi−1 , yi−1,

=0
∆x

85
 y

y0 y1 y2 yN yN+1

x
0 x0 x0 + ∆ x x0 + 2 ∆ x x0 + N ∆ x xN+1

Figure 22: Piecewise linear function

where in (25)
yi+1 − yi
yi = .
∆x
With ∆yi = yi+1 − yi , (25) is
  [Fyi
(xi , yi, ∆y i
) − Fyi−1

(xi−1 , yi−1 , ∆y∆x
i−1
)]
∆yi ∆x
Fyi xi , yi, − = 0. (26)
∆x ∆x
This yields the following system of n equations in n unknowns:
  ∆Fyi−1
∆yi

Fyi xi , yi , − =0 i = 1, 2, · · · , n (27)
∆x ∆x
Equation (27) is the finite difference version of the Euler equation. As n → ∞, ∆x → 0
and (27) becomes the Euler equation.
Example: Find a polygonal line which approximates the extremizing curve for
 2
[(y )2 + 6x2 y]dx y(0) = 0, y(2) = 4 (28)
0

Solution: With n = 1, ∆x = 1, x0 = 0, x1 = 1, x2 = 2, y0 = 0, y2 = 4, and

y1 = y(x1 ) = y(1) is unknown.
We form (in accordance with (22))
1

 2   2  
yi+1 − yi y1 − y0 y2 − y1 2
φ(y1 ) = 6x2i yi + ∆x = 6x20 y0 + 6x21 y1 + + (29)
i=0 ∆x ∆x ∆x
= 0 + 6y1 + y12 + (4 − y1 )2 = 2y12 − 2y1 + 16

Now
dφ
= 4y1 − 2 = 0 ⇒ y1 = 1/2 . (30)
dy1

86
 2 2 4
With n = 2, ∆x = , x0 = 0, x1 = , x2 = , x3 = 2, y0 = 0, and y3 = 4. The
3 3 3
2 4
variables are y1 = y( ) and y2 = y( ). And then
3 3
9 8 9 9 22 2
φ(y1, y2 ) = [ y12 + y1 + y22 − y1 y2 − y2 + 36] (31)
2 3 2 2 3 3
So then
9 8
∂φ/∂y1 = 9y1 − y2 + = 0 (32)
2 3
9 22
∂φ/∂y2 = − y1 + 9y2 − =0
2 3
4 24
giving y1 = and y2 = .
27 27
1
With n = 3 and ∆x = , we have
2
1 3 37
φ(y1 , y2, y3 ) = [8(y12 + y22 + y32 − y1 y2 − y2 y3 ) + y1 + 6y2 − y3 + 64]
2 2 2
and the partial derivatives with respect to yi give
3
16y1 − 8y2 + = 0
2
16y2 − 8(y1 + y3 ) + 6 = 0
37
16y3 − 8y2 − = 0.
2
Solving gives
1 5 21
y1 = y2 = y3 = .
16 16 16

With n = 4 and ∆x = 0.4, we get

y1 = y(0.4) = 0.032 y2 = y(0.8) = 0.1408

y3 = y(1.2) = 0.5568 y4 = y(1.6) = 1.664

The Euler equation for this problem is

3x2 − y  = 0

which when solved with the boundary conditions gives y = x4 /4. If we compare the ap-
proximate values for n = 1, 2, 3, 4 with the exact result, the results are consistently more
accurate for the larger values of the independent variable (i.e., closer to x = 2). But the
relative errors are large for x close to zero. These results are summarized in the table below
and the figure.

87
 x y for n = 1 y for n = 2 y for n = 3 y for n = 4 yexact
0.0 0.0 0.0000 0.0000 0 0.0000
0.4 0.0320 0.0064
0.5 0.0625 0.0156
2/3 0.1481 0.0494
0.8 0.1408 0.1024
1.0 0.5 0.3125 0.2500
1.2 0.5568 0.5184
4/3 0.8889 0.7901
1.5 1.3125 1.2656
1.6 1.6640 1.6384
2.0 4.0 4.0000 4.0000 4.0000 4.0000

3.5

2.5

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Figure 23: The exact solution (solid line) is compared with y1 (dot), y2 (dash dot), y3 (dash)
and y4 (dot)

Problems

1. Write a MAPLE program for the Rayleigh-Ritz approximation to minimize the integral
 1 
I = (y )2 − y 2 − 2xy dx
0

88
 y(0) = 1
y(1) = 2.

Plot the graph of y0 , y1, y2 and the exact solution.

2. Solve the same problem using finite differences.

89
CHAPTER 9

9 Hamilton’s Principle
Variational principles enter into many physical real world problems and can be shown in
certain systems to derive equations which are equivalent to Newton’s equations of motion.
Such a case is Hamilton’s principle, the development is as follows: First let’s assume Newton’s
equations of motion hold for a particle of mass m with vector position R acted on by force
F . Thus
mR̈ − F = 0 (1)
(where “·” denotes time differentiation) is the differential equation which defines the motion
of the particle. Consider the resulting path in time R(t) t1 ≤ t ≤ t2 and let
δR(t) t1 ≤ t ≤ t2 (2)
be a curve satisfying
δR(t1 ) = 0, δR(t2 ) = 0 (3)
(this is η in our previous chapters) and consider (see Figure 24) the varied path R(t)+δR(t) ∗ .
When using the δ notation, it’s often called the variation. Thus δR is the variation in R. The
variation is likened to the differential. So e.g. for a function g(x, y) then δg = gx δx + gy δy,
is the variation in g due to variations δx in x and δy in y.
Varied path
z R(t) + δ R(t)

δ R(t)
t = t2

True path
R(t)

Figure 24: Paths made by the vectors R and R + δR

Now take the dot product between (1) and δR and integrate from t1 to t2
 t2
(mR̈ · δR − F · δR)dt = 0 (4)
t1

If the first term is integrated by parts using

d
(Ṙ · δR) = R̈ · δR + Ṙ · δ Ṙ (5)
dt
∗
In the notation of previous chapters, δR would be called η, (and η would have three components, one
each for x, y, z) however the δ notation is widely used in presenting Hamilton’s principle and has some
advantage over the other.

90
 d
where we’ve used δR = δ Ṙ. Then by (3) this gives
dt
 t2  t2
mR̈ · δRdt = [mṘ · δR
 ]tt21 − (mṘ · δ Ṙ)dt (6)
t1 t1
=0 at both ends
 t2
=− (mṘ · δ Ṙ)dt
t1

Now consider the variation (change) in the term Ṙ2 , due to δR

δ Ṙ2 = δ(Ṙ · Ṙ) = 2Ṙ · δ Ṙ (7)

so that (6) becomes

 t2  t2    t2
1 2
mR̈ · δRdt = − δ mṘ dt = − δT dt (8)
t1 t1 2 t1

1
where T = mṘ2 is the kinetic energy of the particle. Thus using (8) in (4) gives
2
 t2
(δT + F · δR)dt = 0 (9)
t1

This is the most general form of Hamilton’s Principle for a single particle under a general
force field and says that the path of motion is such that along it, the integral of the variation
δT of the kinetic energy T plus F · δR must be zero for variations in the path satisfying
δR(t1 ) = 0, δR(t2 ) = 0.
Conversely, from Hamilton’s Principle we may deduce Newton’s law as follows: From (9),
the definition of T and (7) comes
 t2
(mṘ · δ Ṙ + F · δR)dt = 0 (10)
t1

Now by (5) and integration by parts, using (3) we get

 t2
(−mR̈ + F ) · δRdt = 0 (11)
t1

And since this holds for all δR(t) satisfying (3) we get (by a modified form of the fundamental
lemma presented in chapter 4)
mR̈ − F = 0 (12)
which is Newton’s law of motion.
If the force field is conservative, then there is a function φ of position say φ(x, y, z) for
motion in 3-space such that

φx = F1 , φy = F2 , φz = F3 or then F = ∇φ (13)

91
where F1 , F2 , F3 are the components of force F along x, y, z axes respectively and ∇φ is the
gradient of φ. Then

δφ = φx δx + φy δy + φz δz = F1 δx + F2 δy + F3 δz = F · δR (14)

where δx, δy, δz are the components of δR. The function φ is called the force potential. The
function V defined by
V ≡ −φ (15a)
satisfies (by (14))
F · δR = −δV · (15b)
This function is called the potential energy. For example in gravitational motion in a spher-
ically symmetric field centered at the origin, the force on a particle is
−µ R −µR
F = = (16)
|R| |R|
2 |R|3

(where µ is the gravitational constant). In this case

−µ
V = (17)
|R|

and one can check that (13) and (15) hold.

For conservative fields, by (15b), then (9) becomes
 t2
δ(T − V )dt = 0 (18)
t1

and this is Hamilton’s principle for a conservative force field. Thus, Hamilton’s principle
for a conservative system states that the motion is such that the integral of the
difference between kinetic and potential energies has zero variation.
The difference T − V is often called the Lagrangian L

L≡T −V (19)

and in these terms, Hamilton’s principle for a conservative system says that the motion is
such that  t2  t2
δLdt = δ Ldt = 0 (20)
t1 t1
 t2
(where δ Ldt means the variation in the integral). Then in the usual way we can show
t1
that this means that the motion is such that
dLẋ dLẏ dLż
Lx − = 0 Ly − = 0 Lz − =0 (21)
dt dt dt
i.e. the Euler equations hold for L.

92
 For one dimensional x(t), Euler’s equation is
d
Lx − Lẋ = 0.
dt
Let’s define a canonical momentum, p, by Lẋ , then if Lẋẋ
= 0, then we can solve for ẋ in
terms of t, x, p,
ẋ = φ(t, x, p).
Define the Hamiltonian H by

H(t, x, p) = −L(t, x, φ(t, x, p)) + pφ(t, x, p).

In many systems H is the total energy.

Note that
∂
H = φ = ẋ
∂p
∂
H = −ṗ
∂x
These are known as Hamilton’s equations.
Let’s continue with our example of motion of a particle in a spherically symmetric grav-
itational force field. This is the situation for a satellite travelling about a spherical earth.
We only assume that the force is along a line directed through the center of the Earth,
where we put our origin. We don’t assume any special form for F , such that as we know
it depends only on distance to center of the Earth. Let t0 be some instant and let P be
the plane containing the position and velocity vectors at t0 . For ease of presentation let P
be the horizontal x, y plane pictured below (we can always orient our coordinates so that
this is true). Form spherical coordinates (r, θ, λ) with λ measured in P and θ, measured
perpendicular to P (θ = 0 for the current case). Then in these spherical coordinates

δR = er δr + eθ rδθ + eλ r cos θδλ (22)

where r = |R| is distance from origin to particle and er , eθ , eλ are the unit direction vectors
(see Figure 25) that R changes due to respective changes in the spherical coordinates (r, θ, λ).
Then by (22)

F · δR = F · er δr + F · eθ rδθ + F · eλ r cos θδλ = F · er δr (23)

where the last equality follows since F is along er according to our assumption. Now using
(15), (23) and second part of (13) results in
∂V ∂V ∂V
F · er δr = F · δR = −δV = −[ δr + δθ + δλ] (24)
∂r ∂θ ∂λ
and since δr, δθ, δλ are independent, then this gives
∂V ∂V
= =0 (25)
∂θ ∂λ

93
 z

eθ
λ

er
eλ

Figure 25: Unit vectors er , eθ , and eλ

−µ
i.e. V is only a functions of r (actually we know V = where µ is a constant),
r
V = V (r) (26)

Now for our particle in the xy plane in spherical coordinates, the velocity, of our particle at
“t0 ” has components along er , eθ , eλ respectively of

ṙ, 0, r λ̇ (27)

the second value being due to the velocity vector being in P and eθ being perpendicular to
P . Then the kinetic energy T is
1
T = m(ṙ 2 + r 2 λ̇2 ) (28)
2
So that
1
L=T −V = m(ṙ 2 + r 2 λ̇2 ) − V (r) (29)
2
and the Euler equations (21) given in spherical coordinates:
For r
d dV d
Lr − Lṙ = 0 ⇒ − + mr λ̇2 − [mṙ] = 0 (30a)
dt dr dt
or
dV
mr̈ = − + mr λ̇2
dr
While for θ we see that since θ does not enter in the problem at any time then the motion
stays in same plane.
For λ
d d
Lλ − Lλ̇ = 0 ⇒ (mr 2 λ̇) = 0 (30b)
dt dt

94
 −dV
Equation (30a) says that (since is the force in the r direction and that is the total
dr
force here then) the acceleration in that direction is the sum of that due to the force and
the centrifugal acceleration. Equation (30b) gives a first integral of the motion saying that

mr 2 λ̇ = constant (31)

which says that the angular momentum is constant. This is actually a first integral of the
motion resulting in a first order differential equation instead of a second order differential
equation as in (30a).

Problems

1. If  is not preassigned, show that the stationary functions corresponding to the problem
 1
δ y 2 dx = 0
0

subject to
y(0) = 2, y() = sin 
are of the form y = 2 + 2x cos , where  satisfies the transcendental equation

2 + 2 cos  − sin  = 0.
π 3π
Also verify that the smallest positive value of  is between and .
2 4
2. If  is not preassigned, show that the stationary functions corresponding to the problem
 1  
δ y 2 + 4(y − ) dx = 0
0

subject to
y(0) = 2, y() = 2
x
are of the form y = x2 − 2 + 2, where  is one of the two real roots of the quartic equation

24 − 3 − 1 = 0.

3. A particle of mass m is falling vertically, under the action of gravity. If y is distance

measured downward and no resistive forces are present.
a. Show that the Lagrangian function is
 
1
L = T − V = m ẏ 2 + gy + constant
2
and verify that the Euler equation of the problem
 t2
δ L dt = 0
t1

95
is the proper equation of motion of the particle.
b. Use the momentum p = mẏ to write the Hamiltonian of the system.
c. Show that
∂
H = φ = ẏ
∂p
∂
H = −ṗ
∂y

4. A particle of mass m is moving vertically, under the action of gravity and a resistive
force numerically equal to k times the displacement y from an equilibrium position. Show
that the equation of Hamilton’s principle is of the form
 t2  
1 1
δ mẏ + mgy − ky 2 dt = 0,
2
t1 2 2
and obtain the Euler equation.

5. A particle of mass m is moving vertically, under the action of gravity and a resistive force
numerically equal to c times its velocity ẏ. Show that the equation of Hamilton’s principle
is of the form  t2    t2
1 2
δ mẏ + mgy dt − cẏδy dt = 0.
t1 2 t1

6. Three masses are connected in series to a fixed support, by linear springs. Assuming
that only the spring forces are present, show that the Lagrangian function of the system is
1 
L = m1 ẋ21 + m2 ẋ22 + m3 ẋ23 − k1 x21 − k2 (x2 − x1 )2 − k3 (x3 − x2 )2 + constant,
2
where the xi represent displacements from equilibrium and ki are the spring constants.

96
CHAPTER 10

10 Degrees of Freedom - Generalized Coordinates

If we have a system of particles whose configuration we’re trying to describe, then usually,
owing to the presence of constraints on the system, it is not required to give the actual
coordinates of every particle. Suppose, for instance, that a rigid rod is moving in a plane,
then it’s sufficient to specify the (x, y) coordinates of mass center and the angle that the rod
makes with the x-axis. From these, the position of all points of the rod may be found.
In order to describe the configuration of a system, we choose the smallest possible number
of variables. For example, the configuration of a flywheel is specified by a single variable,
namely the angle through which the wheel has rotated from its initial position. The in-
dependent variables needed to completely specify the configuration of a system are called
generalized coordinates. The generalized coordinates are such that they can be varied ar-
bitrarily and independently without violating the constraints. The number of generalized
coordinates is called the number of degrees of freedom of a system. In the case of the flywheel
the number of degrees of freedom is one while the rigid bar in the plane has three generalized
coordinates. A deformable body does not possess a finite number of generalized coordinates.
Consider a system of N particles with masses m1 , · · · , mN and position vectors R1 , · · · , RN
and with Fi the resultant force on the ith particle. Then for each particle we have

mi R̈i − Fi = 0 i = 1, · · · , N (1)

For this system Hamilton’s principle gives

 t2

N
[δT + Fi · δRi ]dt = 0 (2)
t1 i=1

where T is the kinetic energy of the system of N particles, and

N
1
T = mi Ṙi2 . (3a)
i=1 2

As before, if there is a potential energy V (R1 , · · · RN ) then (2) becomes (see (15b) of Chapter
9)  t2
δ (T − V )dt = 0 (3b)
t1

Now each position vector consists of a triple of numbers so that the system configuration
is determined by 3N numbers. Generally, the system is subject to constraints which implies
that not all of the 3N coordinates are independent. Suppose that there are K constraints of
the type
φi (R1 , · · · , RN , t) = 0 i = 1, · · · k (4)
which must be satisfied by the coordinates. Then there are only 3N − k = p independent
coordinates so that we can select a set of p independent “generalized” coordinates q1 , · · · qp

97
which define the configuration of the system. Therefore, the position vectors Ri can be
written
Ri = Ri (q1 , · · · , qp , t) i = 1, · · · , N (5)
and similarly for the velocities

Ṙi = Ṙi (q1 , · · · qp , q̇1 , · · · q̇p , t) (6)

so that the kinetic energy is a function of q1 · · · qp , q̇1 , · · · q̇p , t

T = T (q1 · · · qp , q̇1 · · · q̇p , t) (7)

and also if there is a potential energy V = V (R1 · · · RN ) then

V = V (q1 , · · · qp , t) (8)

so that
L = T − V = L(q1 · · · qp , q̇1 , · · · q̇p , t) (9)
Then when using (3b), the independent variations are the δqi and not the δRi and the
resultant Euler equations are
 
d ∂ ∂
(T − V ) − (T − V ) = 0 (10)
dt ∂ q̇i ∂qi
Before we do examples let’s review some material on the potential energy V of a conservative
system. We know from a previous chapter that with F as the force then F · δR is the
infinitesimal amount of work done by moving through the displacement δR and also that

F · δR = −δV (11)

i.e. this infinitesimal work is equal to the negative of the infinitesimal change in the potential
energy. For non-infinitesimal changes, then we integrate (thinking of δR as dR and similarly
for δV ) and
 R
F · δR = −[V (R) − V (R1 )] (12)
R1

and get the change in the potential energy between R = R1 and R. For example if a particle
moves along the y-axis (y positive down) in a constant gravity field from y = y1 (R = y here
as the variable defining the system configuration) to y then the change in potential energy
is  R  y
F · δR = mgδy = mgy − mgy1 = −[V (y) − V (y1 )] = −V (y) + c (13)
R1 y1

(thinking of V (y1 ) as a fixed reference value) giving the potential energy

V (y) = −mgy (14)

(often the constant is included in V (y)).

98
 Of course, if the components of R are not all independent, and instead the q variables
are the independent ones we could express everything in terms of those variables and have
 q
F (q) · δq = −[V (q) − V (q 1 )] = −V (q) + c (15)
q1

(where q is the vector with components as the independent variables q).

Example:
A simple pendulum consisting of a point mass is suspended by an inextensible string
of length . The configuration of this system is completely specified by the single angle θ
between the deflected position and some reference position, say equilibrium position where
it’s hanging vertically (Figure 26).

Figure 26: A simple pendulum

Using (14) we see that the potential energy V = −mgy. Here y is determined by θ as

y =  cos θ (16)

so that
V = −mg cos θ (17)
Since the velocity is θ̇, then the kinetic energy T is
1
T = m(θ̇)2 (18)
2

99
so that with q = θ, then (10) becomes here


d ∂ 1 ∂ 1
( m(θ̇)2 + mg cos θ) − ( m(θ̇)2 + mg cos θ) = 0 (19)
dt ∂ θ̇ 2 ∂θ 2
or then
mθ̈ + mg sin θ = 0 (20)
the equation of motion for the pendulum.

Example:
Consider a compound pendulum as pictured in Figure 27.

θ1

l1

m1

l2
θ2

m2

Figure 27: A compound pendulum

In this problem we can’t go directly to circular coordinates since there are a different set
of these for motions about the two pivot points and the motion of m2 is the sum of these
motions so that we must add these two vectorially. We use rectangular coordinates with
(x1 , y1 ) and (x2 , y2 ) as the coordinates of the two masses m1 and m2 respectively. Then in
terms of the independent (generalized) coordinates θ1 , θ2 we have (choosing y negative down)

x1 = 1 sin θ1 y1 = −1 cos θ1 (21)

x2 = 2 sin θ2 + 1 sin θ1 y2 = −2 cos θ2 − 1 cos θ1

The potential and kinetic energies analogously as done before are

1 1
T1 = m1 (ẋ21 + ẏ12 ) T2 = m2 (ẋ22 + ẏ22) (22a)
2 2

100
 V1 = m1 gy1 V2 = m2 gy2 (22b)
Plugging (21) into (22) and writing the Lagrangian gives
1 1
L = T − V = (m1 + m2 )21 θ̇12 + m2 1 2 θ̇1 θ̇2 cos(θ1 − θ2 ) + m2 22 θ̇22 +
2 2 (23)
g(m1 + m2 )1 cos θ1 + gm2 2 cos θ2

Then
d d
0= [Lθ̇1 ] − Lθ1 = [(m1 + m2 )21 θ̇1 + m2 1 2 θ̇2 cos(θ1 − θ2 )] + m2 1 2 θ̇1 θ̇2 sin(θ1 − θ2 )
dt dt
+g(m1 + m2 )1 sin θ1
(24)
and
d d
0 = [Lθ̇2 ] − Lθ2 = [m2 1 2 θ̇1 cos(θ1 − θ2 ) + m2 22 θ̇2 ] + gm2 2 sin θ2 (25)
dt dt
− m2 1 2 θ̇1 θ̇2 sin(θ1 − θ2 )

Example:
Consider the harmonic oscillator whose Lagrangian is given by
1 2 1 2
L(t, y, ẏ) = mẏ − ky .
2 2
The canonical momentum is
p = Lẏ = mẏ.
Solving for ẏ gives
p
ẏ = ,
m
p
i.e. φ(t, y, p) = . Therefore the Hamiltonian is
m
1 p2 1 2
H = −L + ẏLẏ = + ky
2m 2
which is the sum of the kinetic and potential energy of the system. Differentiating
∂H
= ky
∂y
∂H p
=
∂p m
so Hamilton’s equations are
p
ẏ = , ṗ = −ky.
m

101
To solve these equations in the yp plane (the so called phase plane) we divide them to get
∂p −ky
=
∂y p/m
or
pdy + kmydy = 0
After integration, we have

p2 + kmy 2 = c, c is constant

which is a family of ellipses in the py plane. These represent trajectories that the system
evolves along in the position-momentum space. Fixing an initial value at time t0 selects out
the particular trajectory that the system takes.

Problems

1. Consider the functional

 b 
I(y) = r(t)ẏ 2 + q(t)y 2 dt.
a

Find the Hamiltonian and write the canonical equations for the problem.

2. Give Hamilton’s equations for

 b
I(y) = (t2 + y 2)(1 + ẏ 2 )dt.
a

Solve these equations and plot the solution curves in the yp plane.

3. A particle of unit mass moves along the y axis under the influence of a potential

f (y) = −ω 2 y + ay 2

where ω and a are positive constants.

a. What is the potential energy V (y)? Determine the Lagrangian and write down the
equations of motion.
b. Find the Hamiltonian H(y, p) and show it coincides with the total energy. Write
down Hamilton’s equations. Is energy conserved? Is momentum conserved?
ω2
c. If the total energy E is , and y(0) = 0, what is the initial velocity?
10
d. Sketch the possible phase trajectories in phase space when the total energy in the
ω6
system is given by E = 2
.
12a
√ 
Hint: Note that p = ± 2 E − V (y).
What is the value of E above which oscillatory solution is not possible?

102
4. A particle of mass m moves in one dimension under the influence of the force F (y, t) =
ky −2 et , where y(t) is the position at time t, and k is a constant. Formulate Hamilton’s
principle for this system, and derive the equations of motion. Determine the Hamiltonian
and compare it with the total energy.

5. A Lagrangian has the form

a2  4
L(x, y, y ) = (y ) + a(y )2 G(y) − G(y)2 ,
12
where G is a given differentaible function. Find Euler’s equation and a first integral.

6. If the Lagrangian L does not depend explicitly on time t, prove that H = constant, and
if L doesn’t depend explicitly on a generalized coordinate y, prove that p = constant.

7. Consider the differential equations

k −2
r 2 θ̇ = C, r̈ − r θ̇2 + r = 0
m
governing the motion of a mass in an inversely square central force field.
a. Show by the chain rule that
 2
2
−2 dr 2 −4 dr dr
ṙ = Cr , r̈ = C r 2
− 2C 2 r −5
dθ dθ dθ

and therefore the differential equations may be written

 2
d2 r dr k 2
2
− 2r −1 −r+ r = 0
dθ dθ C 2m

b. Let r = u−1 and show that

d2 u k
+ u = .
dθ2 C 2m

c. Solve the differential equation in part b to obtain

k
u = r −1 = (1 +  cos(θ − θ0 ))
C 2m
where  and θ0 are constants of integration.
d. Show that elliptical orbits are obtained when  < 1.

103
CHAPTER 11

11 Integrals Involving Higher Derivatives

Consider the problem  x2
min I = F (x, y, y , y )dx (1)
x1
among arcs
y: y(x) x1 ≤ x ≤ x2
with continuous derivatives up to and including the second derivative, that satisfy the bound-
ary conditions

y(x1 ) = A0 , y  (x1 ) = A1 , y(x2 ) = B0 , y  (x2 ) = B1 . (2)

(Notice now that we also have conditions on y  at the end-points.)

Let y0 (x) be a solution to this problem. Let η(x) be an arc on the interval [x1 , x2 ] with
continuous derivatives up through the second order and satisfying

η(x1 ) = 0, η(x2 ) = 0, η  (x1 ) = 0, η  (x2 ) = 0 (3)

Create the family of arcs

y() : y0 (x) + η(x) x1 ≤ x ≤ x2 −δ <<δ (4)

for some δ > 0.

Evaluate I on this family to get
 x2
I() = F (x, y0 + η, y0 + η  , y0 + η  ]dx (5)
x1

Differentiating at  = 0 and setting the derivative equal to zero gives

 x2
0 = I  (0) = [Fy η + Fy
η  + Fy

η  ]dx (6)
x1

As done previously we can write the second term in the integrand of (6) as
d d
Fy
η  = [Fy
η] − η [Fy
] (7)
dx dx
and the third term in the integrand can similarly be written
d d
Fy

η  = [Fy

η  ] − η  [Fy

] (8a)
dx dx
and then also
d d d d2
η [Fy

] = [η Fy

] − η 2 Fy

(8b)
dx dx dx dx

104
Then using (7) and (8) in (6) gives
 x2  x2
 d d2 d d
0 = I (0) = [Fy − Fy
+ 2 Fy

]ηdx + [Fy
η + Fy

η  + η Fy

]dx (9)
x1 dx dx x1 dx dx
Evaluating the last integral of (9) gives
 x2 
 d d2 d x2
0 = I (0) = [Fy − Fy
+ 2 Fy

]ηdx + [η(Fy
+ Fy

) + η  Fy

] (10)
x1 dx dx dx x1

By (3), the last integral is zero, leaving

 x2
d d2
0 = I  (0) = [Fy − Fy
+ 2 Fy

]ηdx (11)
x1 dx dx
which must be true for the full class of η arcs described above. Then by a slight extension†
of the modified form of the fundamental lemma we get that the integrand of (11) must be
zero giving
d d2
Fy − Fy
+ 2 Fy

= 0 (12)
dx dx
as the Euler equation for this problem.
Note that the integrated form of the Euler equation (12) is
 x s
[ Fy ds − Fy
]dx + c1 x + c2 = Fy

(13)
x1 x1

where c1 , c2 are constants.

As a generalization of this, it can be shown by a directly analagous argument that if the
integrand involves the first N derivatives of y so that the corresponding problem would be
 x2
min I = F (x, y, y , y , · · · , y (N ) )dx (14)
x1

among arcs with continuous N th derivatives on [x1 , x2 ] and satisfying

y(x1 ) = A0 , y  (x1 ) = A1 , · · · , y (N −1) (x1 ) = AN −1 (15)

y(x2 ) = B0 , y  (x2 ) = B1 , · · · , y (N −1) (x2 ) = BN −1

then the Euler equation is

d d2 dN
Fy − Fy
+ 2 Fy

− · · · + (−1)N N Fy(N) = 0 (16)
dx dx dx
a differential equation of order 2N. This result is summarized in
Theorem 1 Consider the problem defined by (14), (15) and the accompanying remarks.
Then a solution arc must satisfy (16).
†
To take into consideration that η  exists and is continuous

105
 As an application consider the problem
 π/4
min I = [16y 2 − (y )2 + φ(x)]dx (17)
0

(where φ is an arbitrary continuous function of x), among arcs possessing continuous deriva-
tives through second order and satisfying

y(0) = 0, y(π/4) = 1, y  (0) = 1 y (π/4) = 0 (18)

Applying the Euler equation gives

d d2
0 = Fy − Fy
+ 2 Fy

= 32y − 2y (4) = 0 (19)
dx dx
or
y (4) − 16y = 0 (20)
The roots of the characteristic equation

D 4 − 16 = 0 (21)

are
D = ±2, ±2i (22)
so that the solution is

y = c1 e2x + c2 e−2x + c3 cos 2x + c4 sin 2x (23)

and then
y  = 2c1 e2x − 2c2 e−2x − 2c3 sin 2x + 2c4 cos 2x (24)
Applying the conditions (18) gives

0 = y(0) = c1 + c2 + c3 (25a)

1 = y(π/4) = c1 eπ/2 + c2 e−π/2 + c4 (25b)

1 = y  (0) = 2c1 − 2c2 + 2c4 (25c)

0 = y  (π/4) = 2c1 eπ/2 − 2c2 e−π/2 − 2c3 (25d)

Solving this system will yield solution of the problem.

We discuss now the Newton method to solve this problem. Analagous to our procedure
for Newton method applications to problems involving terms in x, y, y  we start with an
initial estimate y1 satisfying the left hand conditions of (2) and the Euler equation
d d2
Fy − Fy
+ 2 Fy

= 0 (26)
dx dx

106
which after differentiating out (with respect to x) gives a fourth order equation in y.

Noting that
d
Fy
= Fy
x + Fy
y y  + Fy
y
y  + Fy
y

y  (27a)
dx
and
d
Fy

= Fy

x + Fy

y y  + Fy

y
y  + Fy

y

y  (27b)
dx
then the Euler equation for this problem is

Fy − Fy
x − Fy
y y  − Fy
y
y  − Fy
y

y 

+Fy

xx + Fy

xy y  + Fy

xy
y  + Fy

xy

y 
 
  
+ Fy

yx + Fy

yy y + Fy

yy
y + Fy

yy

y y  + Fy

y y  + (28)
 
+ Fy

y
x + Fy

y
y y  + Fy

y
y
y  + Fy

y
y

y  y  + Fy

y
y  +
 
  
+ Fy

y

x + Fy

y

y y + Fy

y

y
y + Fy

y

y

y y  + Fy

y

y (4) = 0

where all but the first five terms come from differentiating (27b) with respect to x and where
these have been grouped so that the second line of (28) comes from differentiating the first
term on the right in (27b) and each succeeding line comes from differentiating another term
on the right of (27b).

Calling this equation E4 (fourth order Euler equation) we define a two parameter family
of curves y(c1 , c2 ) which satisfies E4 and the left hand conditions of (2) and also has initial
values of y (x0 ) and y  (x0 ) as

y  (x0 ) = y1(x0 ) + c1 y  (x0 ) = y1(x0 ) + c2 (29)

Notice that we have two conditions to satisfy (namely the two right hand conditions of (2))
and we have two parameters to do it with.
As before, we set
∂y(x)
ηi (x) ≡ x0 ≤ x ≤ x1 (30)
∂ci
which now this means
∂y(x) ∂y(x) ∂y(x) ∂y(x)
η1 (x) = =  and η2 (x) = =  (31)
∂c1 ∂y (x0 ) ∂c2 ∂y (x0 )
We obtain the differential equation that ηi (x) has to satisfy by differentiating (28) (evaluated
on the curve y(c)) with respect to ci . By examining (28) we see that the differential equation
for ηi will be fourth order (since we only have terms up through fourth order in y in (28)
and differentiating term by term and using

∂y ∂y  ∂y  ∂y  (4) ∂y (4)

ηi = ηi = ηi = ηi = , ηi = (32)
∂ci ∂ci ∂ci ∂ci ∂ci

107
we get a fourth order equation for ηi , i = 1, 2).
The remainder of the program consists in determining initial conditions for
ηi , ηi , ηi , i = 1, 2 and setting up an iteration scheme to achieve the right–hand end
point conditions of (2).

Problems

1. Derive the Euler equation of the problem

 x2
δ F (x, y, y , y ) dx = 0
x1

in the form    
d2 ∂F d ∂F ∂F
− + = 0,
dx2 ∂y  dx ∂y  ∂y
and show that the associated natural boundary conditions are
  
d ∂F ∂F x2
− δy  = 0
dx ∂y  ∂y  x1

and  
∂F  x2
δy  = 0.
∂y  x1

2. Derive the Euler equation of the problem

 x 2  y2
δ F (x, y, u, ux, uy , uxx , uxy , uyy ) dxdy = 0,
x1 y1

where x1 , x2 , y1 , and y2 are constants, in the form

         
∂2 ∂F ∂2 ∂F ∂2 ∂F ∂ ∂F ∂ ∂F ∂F
+ + − − + = 0,
∂x2 ∂uxx ∂x∂y ∂uxy ∂y 2 ∂uyy ∂x ∂ux ∂y ∂uy ∂u

and show that the associated natural boundary conditions are then
  
∂ ∂F ∂ ∂F ∂F x2
+ − δu  = 0
∂x ∂uxx ∂y ∂uxy ∂ux x1
 
∂F x2
δux  = 0,
∂uxx x1

and   
∂ ∂F ∂ ∂F ∂F  y2
+ − δu  = 0
∂y ∂uyy ∂x ∂uxy ∂uy y1
 
∂F  y2
δuy  = 0.
∂uyy y1

108
3. Specialize the results of problem 2 in the case of the problem
 x 2  y2

1 1 2
δ u2xx + u + αuxx uyy + (1 − α)u2xy dxdy = 0,
x1 y1 2 2 yy
where α is a constant.
Hint: Show that the Euler equation is ∇4 u = 0, regardless of the value of α, but the
natural boundary conditions depend on α.

4. Specialize the results of problem 1 in the case

F = a(x)(y  )2 − b(x)(y  )2 + c(x)y 2 .

5. Find the extremals

 1
a. I(y) = (yy  + (y  )2 )dx, y(0) = 0, y  (0) = 1, y(1) = 2, y (1) = 4
0
 ∞
b. I(y) = (y 2 + (y  )2 + (y  + y )2 )dx, y(0) = 1, y (0) = 2, y(∞) = 0, y (∞) = 0.
0

6. Find the extremals for the functional

 b
I(y) = (y 2 + 2ẏ 2 + ÿ 2 )dt.
a

7. Solve the following variational problem by finding extremals satisfying the given condi-
tions
 1
I(y) = (1 + (y  )2 )dx, y(0) = 0, y (0) = 1, y(1) = 1, y  (1) = 1.
0

109
CHAPTER 12

12 Piecewise Smooth Arcs and Additional Results

Thus far we’ve only considered problems defined over a class of smooth arcs and hence have
only permitted smooth solutions, i.e. solutions with a continuous derivative y  (x). However
one can find examples of variational problems which have no solution in the class of smooth
arcs, but which do have solutions if we extend the class of admissible arcs to include piecewise
smooth arcs‡ . For example, consider the problem
 1
minimize y 2 (1 − y  )2 dx y(−1) = 0 y(1) = 1 . (1)
−1

The greatest lower bound of this integral is clearly (non negative integrand) zero, but it does
not achieve this value for any smooth arc. The minimum is achieved for the arc

0 −1 ≤ x ≤ 0
y(x) = (2)
x 0<x≤1

which is piecewise smooth and thus has a discontinuity in y  (i.e. a “corner”) at the point
x = 0.
In order to include such problems into our theory and to discuss results to follow we
consider again the term admissible arcs

y: y(x) x1 ≤ x ≤ x2 (3)

We will need to refer to the general integral (defined many times before) which we seek to
minimize  x2
I= F (x, y, y )dx (4)
x1

The definition of the particular class of admissible arcs may be made in many ways, each
of which gives rise to a distinct problem of the calculus of variations. For a special problem
the properties defining the class will in general be in part necessitated by the geometrical
or mechanical character of the problem itself, and in part free to be chosen with a large
degree of arbitrariness. An example of a property of the former type is the restriction for
the brachistochrone problem that the curves considered shall all lie below the line y = α,
since on arcs above that line the integral expressing the time of descent has no meaning. On
the other hand we frequently find it convenient to make the arbitrary restriction that our
curves shall all lie in a small neighborhood of a particular one whose minimizing properties
we are investigating, always remembering that on each of the arcs of our class the integral I
must have a well-defined value.
‡
An arc y : y(x) x1 ≤ x ≤ x2 is piecewise smooth if there are at most a finite number of points
x = xi i = 1, · · · , L in the interval [x1 , x2 ] where y  (x) is discontinuous. The points xi at which y  (x) is
discontinuous are called corners.

110
 In order to make a definition of a class of admissible arcs, which will be generally ap-
plicable, let us first assume that there is a region R of sets of values (x, y, y ) in which the
integrand F (x, y, y ) is continuous and has continuous derivatives of as many orders as may
be needed in our theory. The sets of values (x, y, y ) interior to the region R may be desig-
nated as admissible sets. An arc (3) will now be called an admissible arc if it is continuous
and has a continuously turning tangent except possibly at a finite number of corners, and
if the sets of values (x, y(x), y  (x)) on it are all admissible according to the definition just
given. For an admissible arc the interval [x1 , x2 ] can always be subdivided into a number of
subintervals on each of which y(x) is continuous and has a continuous derivative. At a value
x where the curve has a corner the derivative y (x) has two values which we may denote
by y  (x − 0) and y (x + 0), corresponding to the backward and forward slopes of the curve,
respectively.
With the above considerations in mind, then the problem with which we are concerned
is to minimize the integral (4) on the class of admissible arcs (3) joining two fixed points.
The Euler equations which we’ve seen in differentiated and integrated forms, e.g. the first
Euler equation  x
d
Fy
− Fy = 0 Fy
− Fy ds = c (5)
dx
(where c is a constant) were proven by explicitly considering only smooth arcs (i.e. arcs
without corners).
Recall our proof of the integrated form of the first Euler equation (the second equa-
tion of (5)) which we originally did for the shortest distance problem. There we used the
fundamental lemma involving the integral

M(x)η  (x)dx (6)

where in that lemma M(x) was allowed to be piecewise continuous § and η  (x) was required
to have at least the continuity properties of M(x). The term M(x) turned out to be
 x
M(x) = Fy
(x) − Fy ds . (7)

When we allowed only smooth arcs, then Fy
(x) (i.e. Fy
(x, y(x), y (x))) and Fy (x) (i.e.
Fy (x, y(x), y (x))) were continuous (since y(x) and y (x) were so) and the piecewise con-
tinuity provision of the fundamental lemma was not used. This is the procedure that was
followed in chapter 3 and proved that when considering only smooth arcs, the first Euler
equation held in integrated and in differentiated form on a minimizing arc. However, now in
allowing piecewise smooth arcs, then Fy
(x), and Fy (x) may be discontinuous at the corners
of the arc and then by (7) this will also be true for M(x).
Since the fundamental lemma allowed for this, then the proof of the integrated form is
still valid when permitting piecewise smooth arcs. The differentiated form also still holds
in between the corners of the arc but may not hold at the corners themselves. A similar
§
A function M (x) is said to be piecewise continuous on an interval [x1 , x2 ] if there at most a finite number
of points xi i = 1, · · · , L in [x1 , x2 ] where M (x) is discontinuous

111
statement is true for the other Euler equation. With this in mind the theorem concerning
the Euler equations for the general problem stated above, is: For the problem
 x2
minimize I = F (x, y, y )dx (8)
x1

on the class of admissible (where the term admissible is consistent with our above discussions)
arcs (3) joining two fixed points, let the arc

y0 : y0 (x) x1 ≤ x ≤ x2 (9)

be a solution. Then the Euler equations hold in integrated form

 x  x

Fy
(x) − Fy ds = c1 F (x) − y (x)Fy
(x) − Fy ds = c2 (10a)

(where c1 and c2 are constants) along y 0 while the differentiated forms

d d
Fy
− Fy = 0 (F − y Fy
) = Fx (10b)
dx dx
hold between the corners of y 0
Remark: These same modification to the Euler equations hold for the other types of
problems considered. All other results such as the transversality condition remain unchanged
when allowing piecewise smooth arcs.
Because we allow piecewise smooth arcs, then there are two additional necessary condi-
tions to be established and one of these will imply a third additional necessary condition.
Finally we will present one other necessary condition that has nothing to do with corners.
For our discussions to follow, assume that the arc y 0 of (9) is a solution to our problem.
The necessary conditions of Weierstrass and Legendre. In order to prove Weierstrass’
necessary condition, let us select arbitrarily a point 3 on our minimizing arc y 0 , and a second
point 4 of this arc so near to 3 that there is no corner of y 0 between them. Through the
point 3 we may pass an arbitrary curve C with an equation y = Y (x), and the fixed point
4 can be joined to a movable point 5 on C by a one parameter family of arcs y 54 containing
the arc y 34 as a member when the point 5 is in the position 3.

C
5

y0 4
3

Figure 28: Two nearby points 3,4 on the minimizing arc

112
We shall soon see that such a family can easily be constructed. If the integral I(y 0 ) is to be
a minimum then it is clear that as the point 5 moves along C from the point 3 the integral
 x5
I(C35 + y 54 ) = F (x, Y, Y  )dx + I(y 54 ) (11)
x3

must not decrease from the initial value I(y 34 ) which it has when 5 is at the point 3. Then
at the point 3 the differential of this integral with respect to x5 must not be negative.
The differential of the term I(y 54 ) in the expression (11), at the position y 34 , is given by
the expression derived in chapter 4 which we now repeat here

dI(y 54 ) = F (x, y, y )dx + (dy − y  dx)Fy
(x, y, y )|43

where the point 4 is fixed so that dx4 = dy4 = 0. For that formula holds along every arc of
the family in question which satisfies the Euler equations and we know that our minimizing
arc must satisfy these equations. Since the differential of the first integral in the expression
(11) with respect to its upper limit is the value of its integrand at that limit, it follows that
when 5 is at 3 we have for the differential of I(C35 + y 54 ) the value at the point 3 of the
quantity
F (x, Y, Y  )dx − F (x, y, y )dx − (dy − y dx)Fy
(x, y, y ) .
The differentials in this expression belong to the arc C and satisfy the equation dy = Y  dx,
and at the point 3 the ordinates of C and y are equal, so that the differential of (11) is also
expressible in the form

[F (x, y, Y  ) − F (x, y, y ) − (Y  − y )Fy
(x, y, y )]dx|3 (12)

Since this differential must be positive or zero for an arbitrarily selected point 3 and arc C
through it, i.e., for every element (x, y, y ) on y 0 and every admissible element (x, y, Y  ), we
have justified the necessary condition of Weierstrass.
Theorem The Necessary Condition of Weierstrass.
At every element (x, y, y ) of a minimizing arc y 0 , the condition

F (x, y, Y  ) − F (x, y, y ) − (Y  − y  )Fy
(x, y, y ) ≥ 0 (13)

is satisfied for every admissible point (x, y, Y  ) different from (x, y, y ).
The expression on the left side of (13) is usually called the Weierstrass E-function

E(x, y, y , Y  ) ≡ F (x, y, Y  ) − F (x, y, y ) − (Y  − y )Fy
(x, y, y ). (14)

Thus in terms of this quantity, the necessary condition of Weierstrass may be stated as

E(x, y, y , Y  ) ≥ 0 (15)

where (x, y, y ) and (x, y, Y  ) are as noted above.

With the help of Taylor’s formula, the Weierstrass E-function may be expressed in the
form
1
E(x, y, y , Y  ) = (Y  − y  )2 Fy
y
(x, y, y  + θ(Y  − y  )) (16)
2

113
where 0 < θ < 1. If we let Y  approach y  we find from this formula the necessary condition
of Legendre, as an immediate corollary of the condition of Weierstrass.
Theorem The Necessary Condition of Legendre
At every element (x, y, y ) of a minimizing arc y 0 , the condition

Fy
y
(x, y, y ) ≥ 0 (17)

must be satisfied.
In order now to demonstrate the consturction of a family of arcs y 54 of the type used in
the foregoing proof of Weierstrass’ condition, consider the equation

Y (a) − y(a)
y = y(x) + (x4 − x) = y(x, a) . (18)
x4 − a
For x = x4 these arcs all pass through the point 4, and for x = a they intersect the curve C.
For a = x3 the family contains the arc y 34 since at the intersection point 3 of y 34 and C we
have Y (x3 ) − y(x3 ) = 0 and the equation of the family reduces to the equation y = y(x) of
the arc y 34 .
For an element (x, y, y (x − 0)) at a corner of a minimizing arc the proof just given for
Weierstrass’ necessary condition does not apply, since there is always a corner between this
element and a point 4 following it on y 0 . But one can readily modify the proof so that it
makes use of a point 4 preceding the corner and attains the result stated in the condition
for the element in question.
There are two other necessary conditions that result from satisfaction of the Euler equa-
tions. One condition involves corners.
Consider the equation  x
Fy
= Fy dx + c . (19)
x1

The right hand side of this equation is a continuous function of x at every point of the arc
y 0 and the left hand side must therefore also be continuous, so that we have
Corollary 1. The Weierstrass-Erdmann Corner Condition. At a corner (x, y) of a
minimizing arc y 0 the condition

Fy
(x, y, y (x − 0)) = Fy
(x, y, y (x + 0)) (20)

must hold.
This condition at a point (x, y) frequently requires y (x − 0) and y  (x + 0) to be identical
so that at such a point a minimizing arc can have no corners. It will always require this
identity if the sets (x, y, y ) with y  between y  (x − 0) and y (x + 0) are all admissible and
the derivative Fy
y
is everywhere different from zero, since then the first derivative Fy
varies
monotonically with y  and cannot take the same value twice. The criterion of the corollary
has an interesting application in a second proof of Jacobi’s condition which will be given
later.

114
 We have so far made no assumption concerning the existence of a second derivative y  (x)
along our minimizing arc. If an arc has a continuous second derivative then Euler’s equation
along it can be expressed in the form

Fy
x + Fy
y y  + Fy
y
y  − Fy = 0 . (21)

The following corollary contains a criterion which for many problems enables us to prove
that a minimizing arc must have a continuous second derivative and hence satisfy the last
equation.
Corolary 2. Hilbert’s Differentiability condition. Near a point on a minimizing arc y 0
where Fy
y
is different from zero, the arc always has a continuous second derivative y (x).
To prove this let (x, y, y ) be a set of values on y 0 at which Fy
y
is different from zero,
and suppose further that (x + ∆x, y + ∆y, y  + ∆y ) is also on y 0 and with no corner between
it and the former set. If we denote the values of Fy
corresponding to these two sets by Fy

and Fy
+ ∆Fy
then with the help of Taylor’s formula we find

∆Fy
1
= {Fy
(x + ∆x, y + ∆y, y  + ∆y  ) − Fy
(x, y, y )} (22)
∆x ∆x
= Fy
x (x + θ∆x, y + θ∆y, y  + θ∆y  )
∆y
+ Fy
y (x + θ∆x, y + θ∆y, y  + θ∆y  )
∆x
∆y 
+ Fy
y
(x + θ∆x, y + θ∆y, y  + θ∆y  )
∆x
where 0 < θ < 1. In this expression the left hand side ∆Fy
/∆x has the definite limit Fy
as ∆x approaches zero, because of the definition of the derivative and the differentiated
form of the first Euler equation which holds on intervals that have no corners. Also, the
first two terms on the right hand side of (22) have well-defined limits. It follows that the
last term must have a unique limiting value, and since Fy
y

= 0 this can be true only if
y  = lim ∆y /∆x exists. The derivative Fy
y
remains different from zero near the element
(x, y, y ) on the sub-arc of y 0 on which this element lies. Consequently Euler’s equation in
the form given in (21) can be solved for y , and it follows that y  must be continuous near
every element (x, y, y ) of the kind described in the corollary.

115
CHAPTER 13

13 Field Theory Jacobi’s Neccesary Condition and Suf-

ficiency
In this chapter, we discuss the notion of a field and a sufficiency proof for the shortest
distance from a point to a curve.
Recall figure 10 of chapter 3 in which a straight line segement of variable length moved
so that its ends described two curves C and D. These curves written in parametric form are

C: x = x3 (t) y = y3 (t)
D: x = x4 (t) y = y4 (t) . (1)

4
y34
6

D
y56
3

(x,y)
5
Γ

Figure 29: Line segment of variable length with endpoints on the curves C, D

Points 5, 6 are two other points on these curves.

We have seen in chapter 3 that necessary conditions on the shortest arc problem may be
deduced by comparing it with other admissible arcs of special types. It is also true that for
sufficiency in that problem, then a particular arc can be proved to be actually the shortest,
only by comparing it with all of the admissible arcs satisfying the conditions of the problem.
The sufficiency proof of this chapter is valid not only for the arcs which we have named
admissible but also for arcs with equations in the parametric form

x = x(t), y = y(t) (t3 ≤ t ≤ t5 ) . (2)

We suppose always that the functions x(t) and y(t) are continuous, and that the interval
[t3 , t5 ] can be subdivided into one or more parts on each of which x(t) and y(t) have continuous

116



derivatives such that x 2 + y 2
= 0. The curve represented is then continuous and has a
continuously turning tangent except possibly at a finite number of corners. A much larger
variety of curves can be represented by such parametric equations than by an equation of
the form y = y(x) because the parametric representation lays no restriction upon the slope
of the curve or the number of points of the curve which may lie upon a single ordinate. On
the other hand for an admissible arc of the form y = y(x) the slope must always be finite
and the number of points on each ordinate must at most be one.
The mathematician who first made satisfactory sufficiency proofs in the calculus of vari-
ations was Weierstrass, and the ingenious device which he used in his proofs is called a field.
We describe first a generic field for shortest distance problems in general and after giving
some other examples of fields, we introduce the particular field which will be used for the
shortest distance problem from a point to a curve.
For the shortest distance problems, a field Γ is a region of the xy-plane with which there
is associated a one-parameter family of straight-line segments all of which intersect a fixed
curve D, and which have the further property that through each point (x, y) of Γ there passes
one and only one of the segments. The curve D may be either inside the field, or outside as
illustrated in Figure 29, and as a special case it may degenerate into a single fixed point.
The whole plane is a field when covered by a system of parallel lines, the curve D being
in this case any straight line or curve which intersects all of the parallels. The plane with
the exception of a single point 0 is a field when covered by the rays through 0, and 0 is a
degenerate curve D. The tangents to a circle do not cover a field since through each point
outside of the circle there pass two tangents, and through a point inside the circle there is
none. If, however, we cut off half of each tangent at its contact point with the circle, leaving
only a one parameter family of half-rays all pointing in the same direction around the circle,
then the exterior of the circle is a field simply covered by the family of half-rays.
At every point (x, y) of a field Γ the straight line of the field has a slope p(x, y), the
function so defined being called the slope-function of the field. The integral I ∗ introduced
in chapter 4 
∗ dx + pdy
I = √ (3)
1 + p2
with this slope function used for p and with dx, dy coming from the arc C of figure 29, has
a definite value along every arc C35 in the field having equations of the form (2), as we have
seen before. We can prove with the help of the formulas of chapter 3 that the integral I ∗
associated in this way with a field has the two following useful properties:
The values of I ∗ are the same along all curves C35 in the field Γ having the same end-
points 3 and 5. Furthermore along each segment of one of the straight lines of the field the
value of I ∗ is equal to the length of the segment.
To prove the first of these statements we may consider the curve C35 shown in the field Γ
of Figure 29. Through every point (x, y) of this curve there passes, by hypothesis, a straight
line of the field Γ intersecting D, and (4) of chapter 3, applied to the one parameter family
of straight-line segments so determined by the points of C35 , gives

I ∗ (C35 ) = I ∗ (D46 ) − I(y 56 ) + I(y 34 ) . (4)

117
The values of the terms on the right are completely determined when the points 3 and 5 in
the field are given, and are entirely independent of the form of the curve C35 joining these
two points. This shows that the value I ∗ (C35 ) is the same for all arcs (C35 ) in the field
joining the same two end-points, as stated in the theorem.
The second property of the theorem follows from the fact that along a straight-line
segment of the field the differentials
 dx and dy satisfy the equation dy = p dx, and the
∗ 2
integrand of I reduces to 1 + p dx which is the integrand of the length integral.
We now have the mechanism necessary for the sufficiency proof for the problem of shortest
distance from a fixed point 1 to a fixed curve N (introduced in chapter 4).
We recall figure 16 (chapter 4) which is repeated below in which the curve N, its evolute
G and two positions of normals to N, one of them containing the point 1, are shown.

y12
−4

3 L 5 y56
1 6

Figure 30: Shortest arc from a fixed point 1 to a curve N. G is the evolute

We infer by inspection from the Figure, that when the end-point 1 lies between 3 and
2, there is adjoining y 12 a region Γ of the plane which is simply covered by the normals
to N which are near to y 12 . An analytic proof of this statement for a more general case
will be given if time permits. For the present we shall rely on our inference of it from the
figure. The region Γ so covered by the normals to N forms a field such as was described
above. The integral I ∗ formed with the slope function p(x, y) of the field in its integrand,
is independent of the path and has the same value as I along the straight-line segment y 12
of the field. It also has the value zero on every arc of N since the straight lines of the field
are all perpendicular to N and its integrand therefore vanishes identically along that curve.
Hence for an arbitrarily selected arc C14 in F joining 1 with N, as shown in Figure 30, we
have
I(y 12 ) = I ∗ (y 12 ) = I ∗ (C14 + N42 ) = I ∗ (C14 ) , (5a)

118
and the difference between the lengths of C14 and y 12 is
 s2
I(C14 ) − I(y 12 ) = I(C14 ) − I ∗ (C14 ) = (1 − cos θ)ds ≥ 0 (5b)
s1

with the last equality following from (6) of chapter 3.

The difference between the values of I along C14 and y 12 is therefore
 s2
I(C14 ) − I(y12 ) = (1 − cos θ)ds ≥ 0 . (6)
s1

The equality sign can hold only if C14 coincides with y 12 . For when the integral in the last
equation is zero we must have cos θ = 1 at every point of C14 , from which it follows that C14
is tangent at every point to a straight line of the field and satisfies the equation dy = p dx.
Such a differential equation can have but one solution through the initial point 1 and that
solution is y 12 . We have proved therefore that the length I(C14 ) of C14 is always greater
than that of y 12 unless C14 is coincident with y 12 .
For a straight-line segment y 12 perpendicular to the curve N at the point 2 and not
touching the evolute G of N there exists a neighborhood Γ in which y 12 is shorter than every
other arc joining 1 with N.
We now prove a sufficiency theorem for the general problem of chapters 12 and 3 which
we repeat here for completeness.
We wish to minimize the integral
 x2
I= F (x, y, y )dx (7)
x1

on a class of admissible arcs

y : y(x) x1 ≤ x ≤ x2 (8)
joining two given points and lying in some region R of admissible points.
We will often refer to a class of extremals. Recalling the definition from previous chapters
an extremal y is an arc which is a solution to the Euler equations on [x1 , x2 ] and which has
continuous first and second derivatives (y  (x) and y  (x)).
We also define a field for this general problem. A region Γ of the plane is called a field if
it has associated with it a one-parameter family of extremals all intersecting a curve D and
furthermore such that through each point (x, y) of Γ there passes one and but one extremal
of the family. Figure 31 is a picture suggesting such a field.
The function p(x, y) defining the slope of the extremal of the field at a point (x, y) is called
the slope-function of the field. With this slope-function substituted, then the integrand of
the integral I ∗ of chapter 3

I∗ = [F (x, y, p)dx + (dy − pdx)Fy
(x, y, p)] (9)

depends only upon x, y, dx, dy, and the integral itself will have a well-defined value on every
arc C35 in Γ having equations

x = x(t) , y = y(t) (t3 ≤ t ≤ t5 ) (10)

119
 D

y
3

Γ 5

Figure 31: Line segment of variable length with endpoints on the curves C, D

of the type described in (2). Furthermore the endpoints of C35 determine two extremal arcs
y 34 and y 56 of the field, and a corresponding arc D46 , which are related to it like those in
equation (28) of chapter 3, which we repeat here
I(y 56 ) − I(y 34 ) = I ∗ (D46 ) − I ∗ (C35 ) . (11)
It is clear then that the value I ∗ (C35 ) depends only upon the points 3 and 5, and not at all
upon the form of the arc C35 joining them, since the other three terms in equation (11) have
this property.
The importance of the integral I ∗ in the calculus of variations was first emphasized by
Hilbert and it is usually called Hilbert’s invariant integral. Its two most useful properties
are described in the following corollary:
Corollary: For a field Γ simply covered by a one parameter family of extremals all of
which intersect a fixed curve D, the Hilbert integral I ∗ formed with the slope-function p(x,y)
of the field has the same value on all arcs C35 in Γ with the same end-points 3 and 5.
Furthermore on an extremal arc of the field, I ∗ has the same value as I.
The last statement follows, since along an extremal of the field we have dy = p dx and
the integrand of I ∗ reduces to F (x, y, p)dx.
The formula (27) of chapter 3 which we also repeat
4

dI = F (x, y, p) + (dy − pdx)Fy
(x, y, p) (12)
3

and (11) of this chapter are the two important auxiliary formulas developed in chapter 3.
They remain valid in simpler forms if one of the curves C35 or D46 degenerates into a point,
since then the differentials dx, dy along that curve are zero.

120
 We shall see that through a fixed point 1 there passes in general a one-parameter family
of extremals. If such a family has an envelope G as shown in figure 32, then the contact
point 3 of an extremal arc y 12 of the family with the envelope, is called conjugate to point
1 on y 12 .

y
4
y
1
3

Figure 32: Conjugate point at the right end of an extremal arc

We next prove two results which are required for the sufficiency theorem.
The envelope theorem and Jacobi’s condition. The formula (11) enables us to prove the
envelope theorem which is a generalization of the string property of the evolute noted in the
shortest distance problem of chapter 4. Let y 14 and y 13 be two extremals of a one-parameter
family through the point 1, touching an envelope G of the family at their end-points 4 and
3, as shown in Figure 32. When we replace the arc C35 of the formula (11) above by the
fixed point 1, and the arc D46 by G43 , we find the equation

I(y 13 ) − I(y 14 ) = I ∗ (G43 ) . (13)

Furthermore the differentials dx, dy at a point of the envelope G satisfy the equation dy =
p dx with the slope p of the extremal tangent to G at that point, and it follows that the
value of the (Hilbert) integral (9) along G43 is the same as that of I. Hence we have:
The Envelope Theorem. Let y 14 and y 13 be two members of a one-parameter family of
extremals through the point 1, touching an envelope G of the family at their end-points 4
and 3, as shown in Figure 32. Then the values of the integral I along the arcs y 14 , y 13 , G43
satisfy the relation
I(y13 ) + I(G43 ) = I(y 14 ) (14)
for every position of the point 4 preceding 3 on G.
We next prove a condition which was hinted at in chapter 4. This is Jacobi’s condition.
Theorem (Jacobi). On a minimizing arc y 12 which is an extremal with Fy
y

= 0 ev-
erywhere on y 12 , there can be no point 3 conjugate to 1 between 1 and 2. We notice that
according to the envelope theorem, the value of I along the composite arc y 14 + G43 + y 32 in
Figure 32 is always the same as its value along y 12 . But G43 is not an extremal and can be
replaced therefore by an arc C43 giving I a smaller value. In every neighborhood of y 12 there
is consequently an arc y 14 + C43 + y 32 giving I a smaller value than y 12 and I(y 12 ) cannot
be a minimum.

121
 To insure that G43 is not an extremal arc we make use of a well-known property of
(Euler’s) second order differential equation expanded out:
d
Fy
− Fy = Fy
x + Fy
y y  + Fy
y
y  − Fy = 0 (15)
dx
which is satisfied by all extremals. That property states that when such an equation can be
solved for the derivative y  there is one and only one solution of it through an arbitrarily
selected initial point and direction (x3 , y3, y3 ). But we know that equation (15) is solvable
for y  near the arc y 12 since the hypothesis of Jacobi’s condition requires Fy
y
to be different
from zero along that arc. Hence if G43 were an extremal it would necessarily coincide with
y 13 , in which case all of the extremal arcs of the family through the point 1 would by the
same property be tangent to and coincide with y 13 . There would then be no one-parameter
family such as the theorem supposes.
The fundamental sufficiency theorem. The conditions for a minimum which have so
far been deduced for our problem have been only necessary conditions, but we shall see in
the following that they can be made over with moderate changes into conditions which are
also sufficient to insure an extreme value for our integral. Since the comparison of necessary
with sufficient conditions is one of the more delicate parts of the theory of the calculus of
variations, it’s a good idea before undertaking it to consider a sufficiency theorem which in
special cases frequently gives information so complete that after using it one does not need
to use farther the general theory.
Using the general field described above, we as usual designate the function p(x, y) defining
the slope of the extremal of the field at a point (x, y) as the slope-function of the field. With
E as the Weierstrass E− function of chapter 14

E(x, y, y , Y  ) = F (x, y, Y  ) − F (x, y, y ) − (Y  − y )Fy
(x, y, y ) (16)

we have the following theorem, which is fundamental for all of the sufficiency proofs:
The Fundamental Sufficiency Theorem. Let y 12 be an extremal arc of a field Γ such that
at each point (x, y) of Γ the inequality

E(x, y, p(x, y), y ) ≥ 0 (17)

holds for every admissible set (x, y, y ) different from (x, y, p). Then I(y12 ) is a minimum in
Γ, or, more explicitly, the inequality I(y 12 ) ≤ I(C12 ) is satisfied for every admissible arc C12
in Γ joining the points 1 and 2. If the equality sign is excluded in the hypothesis (17) then
I(y 12 ) < I(C12 ) unless C12 coincides with y 12 , and the minimum is a so-called proper one.
In order to accomplish the analysis involved in the proof of the above sufficiency theorem
we now list the properties of the family of extremal arcs covering the field Γ. It is supposed
that the family has an equation of the form

y = y(x, a) (a1 ≤ a ≤ a2 ; x1 (a) ≤ x ≤ x2 (a)) (18)

in which the functions y(x, a), y (x, a) and their partial derivatives up to and including
those of the second order, as well as the functions x1 (a) and x2 (a) defining the end-points

122
of the extremal arcs, are continuous. It is understood that the point of the curve D on each
extremal is defined by a function x = ξ(a) which with its first derivative is continuous on the
interval [a1 , a2 ], and furthermore that the derivative ya is everywhere different from zero on
the extremal arcs. To each point (x, y) in Γ there corresponds a value a(x, y) which defines
the unique extremal of the field through that point, and as a result of the hypothesis that ya
is different from zero we can prove that a(x, y) and its first partial derivatives are continuous
in Γ. The same is then true of the slope-function p(x, y) = y (x, a(x, y)) of the field. These
properties form the analytical basis of the theory of the field, and we assume them always.
The Hilbert integral (9) formed with the slope-function p(x, y) in place of p has now a
definite value on every admissible arc C12 in the field. Furthermore as shown above its values
are the same on all such arcs C12 which have the same end-points, and if the points 1 and 2
are the end-points of an extremal arc y 12 of the field, this value is that of the original integral
I. Hence we find for the pair of arcs C12 and y 12 shown in figure 33,

I(C12 ) − I(y 12 ) = I(C12 ) − I ∗ (y 12 ) = I(C12 ) − I ∗ (C12 ) , (19)

2
C
y

1
Γ

Figure 33: Line segment of variable length with endpoints on the curves C, D

and when we substitute for I and I ∗ their values as integrals, it follows that
 x2
I(C12 ) − I(y 12 ) = E(x, y, p(x, y), y )dx . (20)
x1

In the integral on the right, y and its derivative y  are functions of x obtained from the
equation y = y(x) of the admissible arc C12 .
The sufficiency theorem is an immediate consequence of this formula. For the hypothesis
(17) that the E−function is greater than or equal to zero in the field implies at once that
I(y 12 ) ≤ I(C12 ). If the E−function vanishes in the field only when y  = p then the equality
I(y 12 ) = I(C12 ) can hold only if the equation y  = p(x, y) is satisfied at every point of C12 .
But in that case C12 must coincide with y 12 since the differential equation y  = p(x, y) has
one and but one solution through the initial point 1, and that one is y 12 .

123
 The sufficiency proof of the shortest distance problem was an application of a special
case of the formula (20) and this theorem. For that special problem the second derivative
Fy
y
is positive for all admissible sets (x, y, y ) and the formula (16) of chapter 12 which we
repeat here
1
E(x, y, p, y ) = (y  − p)2 Fy
y
(x, y, p + θ(y  − p)) (0 < θ < 1) (21)
2
shows that the E−function is positive whenever y 
= p, as presupposed in the last sentence
of the sufficiency theorem.
In order to efficiently discuss further sufficiency results it is convenient now to collect
together all of the necessary conditions which have been obtained thus far for our general
problem.
I. For every minimizing arc y 12 there exists a constant c such that the equation
 x

Fy
(x, y(x), y (x)) = Fy (x, y(x), y (x))dx + c (22)
x1

holds identically on y 12 . An immediate consequence of this equation is that on each arc of

y 12 having a continuously turning tangent Euler’s differential equation

d
Fy
− Fy = 0 (23)
dx
must also be satisied.
II. (Weierstrass). At every element (x, y, y ) of a minimizing arc y 12 the condition

E(x, y, y , Y  ) ≥ 0 (24)

must be satisfied foe every admissible set (x, y, Y  ) different from (x, y, y ).
III. (Legendre). At every element (x, y, y ) of a minimizing arc y 12 the condition

Fy
y
(x, y, y ) ≥ 0 (25)

must be satisfied.
IV. (Jacobi). On a minimizing arc y 12 which is an extremal with Fy
y

= 0 everywhere
on it, there can be no point 3 conjugate to 1 between 1 and 2.
The fundamental sufficiency theorem (f.s.t.) proven above refers to a set of admissible
points (of which the admissible arcs are made up) which according to our discussion in
chapter 12 will be contained in some region R. The results to follow will each be closely
associated with a specific R. Also the selection of R will depend in part on the field Γ (also
referred to in the f.s.t) that we are able to construct.
Next, using a notation introduced by Bolza let us designate by II’, III’ the conditions
II, III with the equality sign excluded, and by IV’ the condition IV when strengthened to
exclude the possibility of a conjugate point at the end-point 2 as well as between 1 and 2 on
y 12 . If time permits it will be proven later that for an extremal arc y 12 which satisfies the

124
conditions I, III’, IV’ there is always some neighborhood Γ which is a field simply covered
by a one-parameter family of extremals having y 12 as a member of the family.
The value I(y 12 ) is said to be a weak relative minimum if there is a neighborhood R of
the values (x, y, y ) on y 12 such that the inequality I(y12 ) ≤ I(C12 ) is true, not necessarily for
all admissible arcs C12 , but at least for all those whose elements (x, y, y ) lie in R . With the
help of the sufficiency theorem stated above and the field described in the last paragraph we
shall be able to prove that an arc y 12 which satisfies the conditions I, III’, IV’ will make the
value I(y 12 ) at least a weak relative minimum. This result will be established by replacing
the original region R by R and choosing R so small that every admissible arc with respect
to it is necessarily in the field Γ, and furthermore so small that the condition 17) of the
theorem holds in Γ in its stronger form with respect to all of the sets (x, y, y ) in R .
Following Bolza again let us denote by IIb the condition II strengthened to hold not only
for elements (x, y, y ) on y 12 but also for all such elements in a neighborhood of those on
y 12 . It will be proved that for an arc which satisfies the conditions I, IIb , III’, IV’ the field
Γ about y 12 , existent as a result of the conditions I, III’, IV’, can be so constructed that
the stronger condition 17) holds in it with respect to the sets (x, y, y ) in the region R itself.
The value I(y 12 ) will therefore again be a minimum in Γ, and it is called a strong relative
minimum because it is effective with respect to all admissible comparison curves C whose
elements (x, y, y ) have their points (x, y) in a small neighborhood Γ of those on y 12 . No
restrictions are in this case imposed upon the slopes y  except those due to the definition of
the original region R.
Sufficient Condition for Relative Minima
For our immediate purposes we state now and will prove if time permits a result referred
to above.
Lemma: Every extremal arc y 12 having Fy
y

= 0 along it and containing no point conju-
gate to 1 is interior to a field Γ of which it itself is an extremal arc.
We now discuss the important sets of sufficient conditions which insure for an arc y 12
the property of furnishing a relative minimum. We have seen in chapter 12 that there is
a considerable degree of arbitrariness in the choice of the region R in which the minimum
problem may be studied. Relative minima are really minima in certain types of sub-regions
of the region R originally selected, and their existence is assured by the conditions described
in the following two theorems.
Sufficient conditions for a weak relative minimum. Let y 12 be an arc having the properties:
1) it is an extremal,
2) Fy
y
> 0 at every set of values (x, y, y ) on it,
3) it contains no point 3 conjugate to 1.
Then I(y 12 ) is a weak relative minimum, or, in other words, the inequality I(y 12 ) < I(C12 )
holds for every admissible arc C12 distinct from y 12 , joining 1 with 2, and having its elements
(x, y, y ) all in a sufficiently small neighborhood R of those on y 12 .
To prove this we note in the first place that the conditions 1, 2, 3 of the theorem imply
the conditions I, III’, IV’. Furthermore the same three properties insure the existence of a
field Γ having the arc y 12 as one of its extremals, as indicated in the lemma just stated
above. Let us now choose a neighborhood R of the values (x, y, y ) on y 12 so small that all
elements (x, y, y ) in R have their points (x, y) in Γ, and so small that for the slope-function

125
p = p(x, y) of Γ, the elements x, y, p + θ(y  − p) having 0 ≤ θ ≤ 1 are all admissible and
make Fy
y

= 0. Then the function

1
E(x, y, p(x, y), y ) = (y  − p)2 Fy
y
(x, y, p + θ(y  − p)) , (26)
2
is positive for all elements (x, y, y ) in R with y 
= p, and the fundamental sufficiency
theorem proven earlier in this chapter, with R replaced by R in the definition of admissible
sets, justifies the theorem stated above for a weak relative minimum.
Sufficient Conditions for a Strong Relative Minimum. Let y 12 be an arc having the
properties of the preceding theorem and the further property
4) at every element (x, y, y ) in a neighborhood R of those on y 12 the condition

E(x, y, y , Y  ) > 0

is satisfied for every admissible set (x, y, Y  ) with Y 

= y  .
Then y 12 satisfies the conditions I, IIb , III  , IV  and I(y 12 ) is a strong relative minimum. In
other words, the inequality I(y 12 ) < I(C12 ) holds for every admissible arc C12 distinct from
y 12 , joining 1 with 2, and having its points (x, y) all in a sufficiently small neighborhood Γ
of those on y 12 .
The properties 1), 2), 3) insure again in this case the existence of a field Γ having y 12
as one of its extremal arcs, and we may denote the slope-function of the field as usual by
p(x, y). If we take the field so small that all of the elements (x, y, p(x, y)) belonging to it are
in the neighborhood R of the property 4), then according to that property the inequality

E(x, y, p(x, y), y ) > 0

holds for every admissible element (x, y, y ) in Γ distinct from (x, y, p(x, y)), and the funda-
mental sufficiency theorem gives the desired conclusion of the theorem.
We now use the results just developed for the general theory by applying them to the
brachistochrone problem of finding the curve of quickest descent for a particle to slide from
a given point 1 with coordinates (x1 , y1 ) to a given curve N with a given initial velocity v1 .
This is the same problem we saw in chapter 4 where first necessary conditions were obtained.
Let the point 1, the curve N and the path y 12 of quickest descent be those shown in figure
34. The constant α has the same meaning as in chapter 4, namely α = y1 − v12 /2g where y1
is the value of y at point 1, and g is the gravitational acceleration.

We recall the integral I to be minimized from chapter 4

 x2
1 + y 2
I= dx . (27)
x1 y−α
By chapters 3 and 4 we already know that a minimizing arc y 12 for this problem must
consist of a cycloid lying below the line y = α. We also know by Jacobi’s condition that
y 12 cannot contain a conjugate point between its end-points. We now prove that with the

126
 y=α

N
2
y
3 1

C 4
5

6
y
G

Figure 34: The path of quickest descent from point 1 to a cuve N

assumption of a slight strengthening of Jacobi’s condition, this cycloid provides a strong min-
imizing arc for the problem at hand, (i.e. it satisfies the conditions of the strong sufficiency
theorem).
With F as the integrand of (27) we first compute

y
Fy
= √ √ (28)
y − α 1 + y 2

Next, by the Weierstrass-Erdmann Corner Condition (chapter 12) one sees that the expres-
sion on the right-hand side of (28) is continuous on y 12 . We now show that this implies that
y
y  must also be continuous on y 12 . With the substitution y  = tan α, then √ = sin α
1 + y 2
and the continuity of (28) implies that sin α and hence also α and tan α = y  must be contin-
1
uous along y 12 . Thus y 12 contains no coners. Next, note that Fy
y
= √ >0
y − α(1 − y 2 )3/2
for all admissible (see chapter 3) points (x, y, y ) with y > α, and so certainly also on y 12
then Hilbert’s Differentiability Condition (chapter 12) shows that y  is continuous on y 12 .
Now let R be any neighborhood of y 12 that is contained within the admissible set of
points. Let x, y, y  and x, y, Y  be any points in R (with the same x, y). Then by (26) and
the positivity of Fy
y
for all admissible points, we have condition 4) of the strong sufficiency
theorem. Finally, if we assume that y 12 does not contain a conjugate point at its right end-
point, then all of the conditions of the strong sufficiency theorem are met and y 12 provides
a strong relative minimum for our problem as stated in that theorem.

127
Index
C i, 1 fixed end point, 14
Rn , 1 fixed end points, 63
force potential, 92
admissible functions, 14 fundamental lemma, 21, 25, 43, 60, 91
admissible arc, 42 generalized coordinates, 97
admissible arc, 14 gradient method, 74
admissible arcs, 47 gravitational constant, 92
approximating function, 82 Green’s theorem, 60
Auxiliary Formulas, 22, 26, 32
Hamilton’s equations, 93
both end-points vary, 39 Hamilton’s Principle, 91
brachistochrone, 12, 28, 31, 39 Hamilton’s principle, 90, 92, 97
Hamiltonian, 93
canonical momentum, 93, 101
harmonic oscillator, 101
complete, 83
complete set of Euler equations, 25 indirect method, 63
compound pendulum, 100 infinite number of variables, 10
conservative field, 91 initial estimate, 64
conservative force field, 92 isoparametric, 47, 49
constrained optimization, 5 iterative, 63
constraint, 5
constraints, 47, 97 kinetic energy, 91
cycloid, 28, 29, 32
Lagrange multiplers, 7
degrees of freedom, 97 Lagrange multiplier, 49
difference quotient, 85 Lagrangian, 92, 101
differential correction, 64 maxima, 1
direct method, 63, 74 mean curvature, 62
Euler equation, 25 method of eigenfunction expansion, 82
Euler equations, 63 minima, 1
Euler Lagrange equations, 59 minimal surface problem, 61
Euler’s method, 86 modified version of ode23, 66
evolute, 37 natural boundary condition, 38
extremal, 25 Necessary condition, 38
feta.m, 66, 70 Newton’s equations of motion, 90
finite differences, 84 Newton’s law, 91
finput.m, 77 Newton’s method, 63, 71
first Euler equation, 46, 49 numerical techniques, 63
first necessary condition, 14 ode1.m, 66
first order necessary condition, 3 ode23m.m, 66, 67

128
odef.m, 66, 70
odeinput.m, 65
optimize, 1

phase plane, 102

potential energ, 92

Rayleigh-Ritz, 13
Rayleigh-Ritz method, 82
relative minimum, 10, 19
rhs2f.m, 65, 66
Riemann integrable, 83

second Euler equation, 46, 49

second order necessary condition, 3
shortest arc, 10, 14, 36
shortest distance, 21, 31
shortest time, 12
side conditions, 47
simple pendulum, 99
steepest descent, 74
subsidiary conditions, 47
sufficient condition, 3
surface of revolution, 11

Taylor series, 74
transversality condition, 38–41, 50, 71
two independent variables, 59
two-dimensional problem, 46

unconstrained, 1
unconstrained relative minimum, 1

variable end point, 14, 36, 38, 71

129